diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_0/index_tensor([7254, 7254])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_0/index_tensor([7254, 7254])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..55f3852cd507b8bc6772a735017498720072c9da --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_0/index_tensor([7254, 7254])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, low-angle, dolly shot moves across a rain-soaked, muddy clearing, approaching the massive base of an ancient oak tree. The camera begins its upward, heavy lift, navigating the storm. The tree's massive, gnarled branches, as thick as a human trunk, twist and strain violently under the relentless assault of unseen gales, groaning audibly under the immense pressure. The entire tree sways, its dark, damp bark deeply furrowed like ancient leather, speckled with shimmering green moss, catching the dull, steel-gray glow of the storm-choked sky. The camera weaves between the flapping lower limbs, capturing how smaller branches whip back and forth frantically, while tufts of dark green leaves are torn off and flung into a swirling, bruised, purple cloud. As torrential rain slants across the frame, the atmosphere is one of primal, violent struggle, a testament to the enduring power against overwhelming force. The tree's powerful, agonized movements convey a desperate fight for survival.", + "seed": 11671, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_0/index_7254_seed_11671_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1653535535559058, + "vlm_logprob": -0.1653535535559058, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.1653535535559058, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013917187228798866 + ], + "opposite_logprob_tokens": [ + -4.513916969299316 + ], + "logprob_reward": -0.013917187228798866, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013917187228798866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17163392901420593, + 0.08965570479631424, + 0.2202436476945877, + 0.059858668595552444, + 0.11304651200771332, + 0.10784584283828735, + 0.08453850448131561, + 0.04450786113739014, + 0.07878616452217102, + 0.029883157461881638 + ], + "credit_map": [ + 0.011858237907290459, + 0.01609622873365879, + 0.032739829272031784, + 0.022516829892992973, + 0.019735142588615417, + 0.020221497863531113, + 0.024919047951698303, + 0.023547116667032242, + 0.0058340029790997505, + 0.007350812666118145, + 0.014365521259605885, + 0.014914813451468945, + 0.007882863283157349, + 0.00862621609121561, + 0.01588008552789688, + 0.01480139046907425, + 0.007353269029408693, + 0.055712535977363586, + 0.009476277977228165, + 0.015767395496368408, + 0.0104870880022645, + 0.014311254024505615, + 0.08751697838306427, + 0.01961885765194893, + 0.003354110522195697, + 0.004987925756722689, + 0.007406384218484163, + 0.00513755576685071, + 0.004629506263881922, + 0.008316718973219395, + 0.014860466122627258, + 0.011166002601385117, + 0.005773867946118116, + 0.01177303772419691, + 0.012082395143806934, + 0.014884043484926224, + 0.007638104725629091, + 0.010904178954660892, + 0.03483126312494278, + 0.015159624628722668, + 0.006947179324924946, + 0.007589519023895264, + 0.016443517059087753, + 0.009623727761209011, + 0.008630800060927868, + 0.021910469979047775, + 0.017696337774395943, + 0.019004283472895622, + 0.012631207704544067, + 0.010027979500591755, + 0.00836972426623106, + 0.010664265602827072, + 0.008245641365647316, + 0.007891549728810787, + 0.01143765076994896, + 0.015270489268004894, + 0.0028490438126027584, + 0.0032494228798896074, + 0.00448089512065053, + 0.004911121446639299, + 0.005606171675026417, + 0.007815788500010967, + 0.00952097587287426, + 0.006074442993849516, + 0.004040630999952555, + 0.005156927276402712, + 0.008266731165349483, + 0.010279088281095028, + 0.006198738235980272, + 0.0056750900112092495, + 0.02260354347527027, + 0.016565417870879173, + 0.006433580070734024, + 0.0035855071619153023, + 0.004633944481611252, + 0.002842334797605872, + 0.0018926251213997602, + 0.002746135462075472, + 0.0040460871532559395, + 0.0037029418163001537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.013917187228798866 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004648474045097828 + ], + "opposite_logprob_tokens": [ + -5.379648685455322 + ], + "logprob_reward": -0.004648474045097828, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004648474045097828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21249479055404663, + 0.07020674645900726, + 0.2871605157852173, + 0.11106596142053604, + 0.059881437569856644, + 0.059781212359666824, + 0.051816731691360474, + 0.04223761707544327, + 0.05416179448366165, + 0.05119318887591362 + ], + "credit_map": [ + 0.012405812740325928, + 0.01859874278306961, + 0.04450574144721031, + 0.04233114421367645, + 0.013963266275823116, + 0.027967559173703194, + 0.026966368779540062, + 0.025756150484085083, + 0.005548052955418825, + 0.013128524646162987, + 0.010732670314610004, + 0.009974252432584763, + 0.005608355160802603, + 0.008708188310265541, + 0.00910170003771782, + 0.007404995616525412, + 0.008733708411455154, + 0.07463199645280838, + 0.010665034875273705, + 0.018995169550180435, + 0.012858258560299873, + 0.019037630409002304, + 0.11833242326974869, + 0.023906296119093895, + 0.009294691495597363, + 0.012261775322258472, + 0.013394486159086227, + 0.01219960767775774, + 0.009563841857016087, + 0.012053036130964756, + 0.026638129726052284, + 0.015660390257835388, + 0.0026392776053398848, + 0.006708938162773848, + 0.00546501949429512, + 0.010422782972455025, + 0.004417645279318094, + 0.005249552428722382, + 0.016006343066692352, + 0.00897187739610672, + 0.003931887447834015, + 0.005932209547609091, + 0.0073760696686804295, + 0.010297208093106747, + 0.0029486685525625944, + 0.009659925475716591, + 0.010062789544463158, + 0.009572451002895832, + 0.00562346400693059, + 0.0060920692048966885, + 0.004404670558869839, + 0.007628175429999828, + 0.004299202002584934, + 0.00843413919210434, + 0.007342095952481031, + 0.007992910221219063, + 0.002633383497595787, + 0.0029932609759271145, + 0.004858943168073893, + 0.005010224413126707, + 0.004287680611014366, + 0.007585426792502403, + 0.008565610274672508, + 0.006303085014224052, + 0.0037306523881852627, + 0.0037487621884793043, + 0.0052862754091620445, + 0.007452027406543493, + 0.0041844481602311134, + 0.005053746979683638, + 0.0143295768648386, + 0.01037630159407854, + 0.00974668376147747, + 0.0035147855523973703, + 0.0068161459639668465, + 0.005058348644524813, + 0.0034829978831112385, + 0.004532217979431152, + 0.011977407149970531, + 0.006064596585929394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.004648474045097828 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.757826566696167 + ], + "opposite_logprob_tokens": [ + -0.632826566696167 + ], + "logprob_reward": -0.757826566696167, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.757826566696167, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16746972501277924, + 0.10050665587186813, + 0.19511140882968903, + 0.20975148677825928, + 0.07787090539932251, + 0.06480665504932404, + 0.05057404190301895, + 0.0701022669672966, + 0.027565626427531242, + 0.03624119609594345 + ], + "credit_map": [ + 0.011799506843090057, + 0.016268545761704445, + 0.028268029913306236, + 0.02402605675160885, + 0.01969757117331028, + 0.027071652933955193, + 0.024653974920511246, + 0.015684407204389572, + 0.005924850236624479, + 0.008390679024159908, + 0.015719076618552208, + 0.01802353747189045, + 0.008610104210674763, + 0.009764960035681725, + 0.020107880234718323, + 0.013965579681098461, + 0.016198869794607162, + 0.04695820435881615, + 0.009978088550269604, + 0.015608065761625767, + 0.010547180660068989, + 0.015325181186199188, + 0.06584981828927994, + 0.014646019786596298, + 0.012020070105791092, + 0.024956420063972473, + 0.03011525422334671, + 0.02730134129524231, + 0.016732342541217804, + 0.034760575741529465, + 0.032728999853134155, + 0.031136518344283104, + 0.0038340154569596052, + 0.00869827251881361, + 0.006458495743572712, + 0.01325914915651083, + 0.004746574908494949, + 0.004962248727679253, + 0.022064367309212685, + 0.013847787864506245, + 0.0057824840769171715, + 0.008597366511821747, + 0.00884971022605896, + 0.008263373747467995, + 0.0054589565843343735, + 0.0069770170375704765, + 0.010710758157074451, + 0.010166994296014309, + 0.007434953469783068, + 0.005946903955191374, + 0.006013999227434397, + 0.005883580073714256, + 0.0038367146626114845, + 0.0071090878918766975, + 0.006350918207317591, + 0.007997891865670681, + 0.007475250866264105, + 0.005958877969533205, + 0.007048177532851696, + 0.006623389665037394, + 0.008447830565273762, + 0.010448929853737354, + 0.013452793471515179, + 0.01064702495932579, + 0.0023926072753965855, + 0.0025607363786548376, + 0.003324262099340558, + 0.004002817440778017, + 0.0024684639647603035, + 0.0032244850881397724, + 0.006278680637478828, + 0.0033135784324258566, + 0.0046893698163330555, + 0.002932910807430744, + 0.005499105900526047, + 0.00479351170361042, + 0.0027885327581316233, + 0.0042846836149692535, + 0.006494327448308468, + 0.004758759401738644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.757826566696167 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020582539960741997 + ], + "opposite_logprob_tokens": [ + -3.895582437515259 + ], + "logprob_reward": -0.020582539960741997, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020582539960741997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08750534802675247, + 0.13647280633449554, + 0.1384720355272293, + 0.2184429168701172, + 0.04202425852417946, + 0.05914667993783951, + 0.11647000908851624, + 0.11954426020383835, + 0.04203203320503235, + 0.03988964483141899 + ], + "credit_map": [ + 0.004904633853584528, + 0.008434468880295753, + 0.02006087452173233, + 0.010221357457339764, + 0.006297222338616848, + 0.011842293664813042, + 0.013743297196924686, + 0.012001194059848785, + 0.00597106758505106, + 0.02143009379506111, + 0.014543889090418816, + 0.021950338035821915, + 0.014767162501811981, + 0.008889982476830482, + 0.03324891999363899, + 0.01567136123776436, + 0.010611288249492645, + 0.03477248549461365, + 0.010566671378910542, + 0.014719760976731777, + 0.010280923917889595, + 0.013532296754419804, + 0.03313213959336281, + 0.010856470093131065, + 0.013695720583200455, + 0.0204241331666708, + 0.024657225236296654, + 0.02244744263589382, + 0.024200374260544777, + 0.03137539327144623, + 0.04745179042220116, + 0.034190837293863297, + 0.0021938171703368425, + 0.00452180951833725, + 0.005102424882352352, + 0.006753617897629738, + 0.0028022676706314087, + 0.003420637920498848, + 0.010960077866911888, + 0.006269602570682764, + 0.005312616005539894, + 0.006298813968896866, + 0.006871507037431002, + 0.006505225319415331, + 0.0043090819381177425, + 0.007482957560569048, + 0.007420747075229883, + 0.014945728704333305, + 0.012016606517136097, + 0.008917280472815037, + 0.01378745399415493, + 0.013391203247010708, + 0.007762941066175699, + 0.01538277417421341, + 0.01790975034236908, + 0.027301998808979988, + 0.011865997686982155, + 0.01314876414835453, + 0.013845568522810936, + 0.011990649625658989, + 0.011580641381442547, + 0.016546158120036125, + 0.02235567755997181, + 0.018210800364613533, + 0.002696797950193286, + 0.0023505017161369324, + 0.004004555754363537, + 0.005487574264407158, + 0.003578563453629613, + 0.003950205631554127, + 0.01246738713234663, + 0.007496445905417204, + 0.008147843182086945, + 0.0025033075362443924, + 0.005686980672180653, + 0.004442601930350065, + 0.0024892918299883604, + 0.005078364163637161, + 0.00649271858856082, + 0.005048538092523813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.020582539960741997 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979299984872341 + ], + "opposite_logprob_tokens": [ + -3.5297930240631104 + ], + "logprob_reward": -0.02979299984872341, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979299984872341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11109454929828644, + 0.11033023148775101, + 0.22238627076148987, + 0.23769131302833557, + 0.04796649515628815, + 0.072822205722332, + 0.07550938427448273, + 0.06488630175590515, + 0.026881396770477295, + 0.030431831255555153 + ], + "credit_map": [ + 0.008699835278093815, + 0.013102065771818161, + 0.022098438814282417, + 0.01243507768958807, + 0.011275183409452438, + 0.014281434938311577, + 0.015963159501552582, + 0.01323934830725193, + 0.0061474815011024475, + 0.016591722145676613, + 0.014043780043721199, + 0.016235267743468285, + 0.012057241052389145, + 0.008447932079434395, + 0.018755024299025536, + 0.018051782622933388, + 0.007983813062310219, + 0.05772019550204277, + 0.011109028942883015, + 0.017459416761994362, + 0.011018719524145126, + 0.015832047909498215, + 0.08289530873298645, + 0.018367737531661987, + 0.017163079231977463, + 0.023795394226908684, + 0.02388583868741989, + 0.029135150834918022, + 0.011318432167172432, + 0.019949989393353462, + 0.06313639879226685, + 0.049307022243738174, + 0.004126550164073706, + 0.005275804549455643, + 0.006060176063328981, + 0.008156382478773594, + 0.003643621690571308, + 0.005876163952052593, + 0.008811482228338718, + 0.006016314961016178, + 0.005286218132823706, + 0.007925492711365223, + 0.006857846863567829, + 0.008180555887520313, + 0.004257704596966505, + 0.011642755940556526, + 0.01245096791535616, + 0.016220662742853165, + 0.014060206711292267, + 0.00823395699262619, + 0.007826183922588825, + 0.010732563212513924, + 0.005524374544620514, + 0.010380174964666367, + 0.009117628447711468, + 0.009634298272430897, + 0.008817032910883427, + 0.007211611140519381, + 0.013357683084905148, + 0.0069991955533623695, + 0.005687205120921135, + 0.006794999353587627, + 0.007908733561635017, + 0.00810984056442976, + 0.0014840422663837671, + 0.0015500168083235621, + 0.002413818845525384, + 0.003354880725964904, + 0.0016661122208461165, + 0.002122778445482254, + 0.008835842832922935, + 0.005453903693705797, + 0.004621082451194525, + 0.0027666636742651463, + 0.003460629377514124, + 0.0035355272702872753, + 0.0022301902063190937, + 0.003102091606706381, + 0.005264964886009693, + 0.005450680386275053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.02979299984872341 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle, dolly shot moves across a rain-soaked, muddy clearing, approaching the massive base of an ancient oak tree. The camera begins its upward, heavy lift, navigating the storm. The tree's massive, gnarled branches, as thick as a human trunk, twist and strain violently under the relentless assault of unseen gales, groaning audibly under the immense pressure. The entire tree sways, its dark, damp bark deeply furrowed like ancient leather, speckled with shimmering green moss, catching the dull, steel-gray glow of the storm-choked sky. The camera weaves between the flapping lower limbs, capturing how smaller branches whip back and forth frantically, while tufts of dark green leaves are torn off and flung into a swirling, bruised, purple cloud. As torrential rain slants across the frame, the atmosphere is one of primal, violent struggle, a testament to the enduring power against overwhelming force. The tree's powerful, agonized movements convey a desperate fight for survival.", + "seed": 11672, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_0/index_7254_seed_11672_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10411301786080003, + "vlm_logprob": -0.10411301786080003, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.10411301786080003, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02210315316915512 + ], + "opposite_logprob_tokens": [ + -3.8971030712127686 + ], + "logprob_reward": -0.02210315316915512, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02210315316915512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055293936282396317, + 0.12450321763753891, + 0.01597941666841507, + 0.021839827299118042, + 0.28600960969924927, + 0.1091136559844017, + 0.08280228078365326, + 0.28658318519592285, + 0.004853243939578533, + 0.013021611608564854 + ], + "credit_map": [ + 0.003636095905676484, + 0.007753506302833557, + 0.005704184528440237, + 0.006539565976709127, + 0.00730215385556221, + 0.008030090481042862, + 0.008811797015368938, + 0.007516541518270969, + 0.007234100718051195, + 0.009002124890685081, + 0.008052216842770576, + 0.01088701281696558, + 0.033422742038965225, + 0.006951048504561186, + 0.03267402946949005, + 0.016279948875308037, + 0.0009188716649077833, + 0.001551422057673335, + 0.0010575499618425965, + 0.0024828799068927765, + 0.003195322584360838, + 0.0014884636038914323, + 0.0029559931717813015, + 0.0023289136588573456, + 0.0016105396207422018, + 0.003542194375768304, + 0.0013083943631500006, + 0.002034537959843874, + 0.0038697889540344477, + 0.0019532502628862858, + 0.003476145211607218, + 0.004044977482408285, + 0.013655086047947407, + 0.009680858813226223, + 0.011187155731022358, + 0.014877556823194027, + 0.10002966225147247, + 0.008198300376534462, + 0.11113870143890381, + 0.01724228821694851, + 0.008582120761275291, + 0.01006076019257307, + 0.007163249421864748, + 0.012782929465174675, + 0.014285706914961338, + 0.011188868433237076, + 0.02538478560745716, + 0.019665228202939034, + 0.005965516436845064, + 0.0063064140267670155, + 0.005098626017570496, + 0.009558480232954025, + 0.008015371859073639, + 0.00836512167006731, + 0.021542945876717567, + 0.017949802801012993, + 0.019835028797388077, + 0.022829893976449966, + 0.06913317739963531, + 0.02634619176387787, + 0.026560576632618904, + 0.021341348066926003, + 0.03164631500840187, + 0.06889064610004425, + 0.0005112090148031712, + 0.0007815745775587857, + 0.000611023569945246, + 0.0008640633895993233, + 0.0004933083546347916, + 0.00044636501115746796, + 0.0006795141962356865, + 0.00046618535998277366, + 0.0011340902419760823, + 0.0014114102814346552, + 0.0021076034754514694, + 0.0016571525484323502, + 0.0014921127585694194, + 0.0009664594545029104, + 0.0024484398309141397, + 0.0018043422605842352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.02210315316915512 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032035005278885365 + ], + "opposite_logprob_tokens": [ + -5.753203392028809 + ], + "logprob_reward": -0.0032035005278885365, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032035005278885365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04390232264995575, + 0.1328854262828827, + 0.01182547677308321, + 0.03047354705631733, + 0.24919718503952026, + 0.12095239013433456, + 0.061868537217378616, + 0.3349258005619049, + 0.008715428411960602, + 0.005253847222775221 + ], + "credit_map": [ + 0.0031572463922202587, + 0.005270866677165031, + 0.0029481316450983286, + 0.004771267529577017, + 0.008457552641630173, + 0.005105124786496162, + 0.007299128454178572, + 0.006893006153404713, + 0.005036733578890562, + 0.009343033656477928, + 0.007225146982818842, + 0.020493045449256897, + 0.021578975021839142, + 0.013298784382641315, + 0.019906772300601006, + 0.03600294515490532, + 0.0007594507187604904, + 0.0015935576520860195, + 0.0010449192486703396, + 0.002075283555313945, + 0.0018408969044685364, + 0.001146425842307508, + 0.0018925380427390337, + 0.0014724056236445904, + 0.0019945017993450165, + 0.0048250420950353146, + 0.0020626881159842014, + 0.003103929804638028, + 0.005854950752109289, + 0.0026175910606980324, + 0.004885078873485327, + 0.005129766650497913, + 0.012215514667332172, + 0.008798345923423767, + 0.010144342668354511, + 0.014559799805283546, + 0.08578188717365265, + 0.007130349986255169, + 0.0956326276063919, + 0.014934325590729713, + 0.009372111409902573, + 0.010578369721770287, + 0.00799641851335764, + 0.013116701506078243, + 0.015028241090476513, + 0.01230532769113779, + 0.029342789202928543, + 0.023212440311908722, + 0.004361084662377834, + 0.005929694976657629, + 0.004320099484175444, + 0.007994725368916988, + 0.007260718382894993, + 0.00737584987655282, + 0.014267663471400738, + 0.010358704254031181, + 0.027132591232657433, + 0.02201985754072666, + 0.08097705245018005, + 0.03238777816295624, + 0.031065985560417175, + 0.025344135239720345, + 0.03740948811173439, + 0.07858894765377045, + 0.0010178751545026898, + 0.0014535534428432584, + 0.0015220929635688663, + 0.0009534964337944984, + 0.0007292164373211563, + 0.0008605570183135569, + 0.0010453653521835804, + 0.001133272424340248, + 0.0005403444520197809, + 0.0007453561993315816, + 0.0005543518927879632, + 0.0006009098724462092, + 0.0007380369934253395, + 0.0005071955965831876, + 0.0006412701332010329, + 0.000926382839679718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.0032035005278885365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289168119430542 + ], + "opposite_logprob_tokens": [ + -1.0539168119430542 + ], + "logprob_reward": -0.4289168119430542, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289168119430542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06424707174301147, + 0.17639772593975067, + 0.042022280395030975, + 0.06669413298368454, + 0.41597992181777954, + 0.05762038379907608, + 0.06812981516122818, + 0.04963817447423935, + 0.03194187209010124, + 0.027328625321388245 + ], + "credit_map": [ + 0.006174192298203707, + 0.007751269731670618, + 0.005310920532792807, + 0.007326058112084866, + 0.010447836481034756, + 0.006272369530051947, + 0.011300862766802311, + 0.00966355949640274, + 0.011631066910922527, + 0.022029420360922813, + 0.013974333181977272, + 0.030228573828935623, + 0.02825307846069336, + 0.018028227612376213, + 0.029964270070195198, + 0.022288739681243896, + 0.0031737268436700106, + 0.0028065412770956755, + 0.0024013477377593517, + 0.006108996924012899, + 0.009744871407747269, + 0.0030825624708086252, + 0.0075829410925507545, + 0.007121291011571884, + 0.00371501874178648, + 0.0071600028313696384, + 0.004948172252625227, + 0.009161113761365414, + 0.010895195417106152, + 0.006154214031994343, + 0.011659283190965652, + 0.01300112996250391, + 0.019665945321321487, + 0.011390372179448605, + 0.011494831182062626, + 0.015957973897457123, + 0.1533404439687729, + 0.01069882232695818, + 0.17037536203861237, + 0.0230561550706625, + 0.005079062655568123, + 0.006704031955450773, + 0.004099491983652115, + 0.006910910364240408, + 0.010294212028384209, + 0.005921124015003443, + 0.010368096642196178, + 0.008243447169661522, + 0.006295985542237759, + 0.0045988187193870544, + 0.004125817213207483, + 0.006632966920733452, + 0.005312428809702396, + 0.006361618638038635, + 0.01584392413496971, + 0.018958253785967827, + 0.003764499444514513, + 0.004956139251589775, + 0.009512464515864849, + 0.004708398599177599, + 0.003247625194489956, + 0.0046891761012375355, + 0.006670215167105198, + 0.012089655734598637, + 0.0029554995708167553, + 0.00435103802010417, + 0.006041424814611673, + 0.0032472065649926662, + 0.004840400069952011, + 0.002645860891789198, + 0.003865632927045226, + 0.003994806203991175, + 0.0028760815039277077, + 0.004369437228888273, + 0.003778425743803382, + 0.0030957816634327173, + 0.003772282274439931, + 0.0018695055041462183, + 0.003239803947508335, + 0.0043273065239191055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.4289168119430542 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306088760495186 + ], + "opposite_logprob_tokens": [ + -3.168060779571533 + ], + "logprob_reward": -0.04306088760495186, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306088760495186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024668097496032715, + 0.049865078181028366, + 0.015833446756005287, + 0.03193666785955429, + 0.24123993515968323, + 0.11813289672136307, + 0.06431825459003448, + 0.43369296193122864, + 0.014452951028943062, + 0.00585970189422369 + ], + "credit_map": [ + 0.0014603992458432913, + 0.0035329938400536776, + 0.001886459649540484, + 0.0020630366634577513, + 0.003933771513402462, + 0.0025046358350664377, + 0.0052590942941606045, + 0.0040277051739394665, + 0.002946779364719987, + 0.005378118250519037, + 0.0044278334826231, + 0.008786676451563835, + 0.007503299508243799, + 0.00416589854285121, + 0.009020552039146423, + 0.007635919842869043, + 0.0009939022129401565, + 0.002119923708960414, + 0.0015672134468331933, + 0.002714417641982436, + 0.0022149148862808943, + 0.00135724653955549, + 0.00225677527487278, + 0.0026090515311807394, + 0.002054950688034296, + 0.0033375686034560204, + 0.0026507575530558825, + 0.00315105146728456, + 0.0059896172024309635, + 0.002943874103948474, + 0.005287969950586557, + 0.006520875263959169, + 0.011561124585568905, + 0.007978824898600578, + 0.009019438177347183, + 0.011978582479059696, + 0.08516687154769897, + 0.006845101714134216, + 0.09455449134111404, + 0.014135509729385376, + 0.008933132514357567, + 0.010063553228974342, + 0.007828570902347565, + 0.01172116119414568, + 0.014232203364372253, + 0.012216594070196152, + 0.030719319358468056, + 0.022418363019824028, + 0.004763209726661444, + 0.005767563823610544, + 0.004254773259162903, + 0.007907498627901077, + 0.00658669788390398, + 0.006812254432588816, + 0.015566643327474594, + 0.012659617699682713, + 0.03343610465526581, + 0.029334595426917076, + 0.10510360449552536, + 0.04154505580663681, + 0.04048577696084976, + 0.03278541937470436, + 0.04826858267188072, + 0.10273382812738419, + 0.0011095888912677765, + 0.0019791352096945047, + 0.0019183466210961342, + 0.002112877555191517, + 0.0022481149062514305, + 0.0014332225546240807, + 0.0023622994776815176, + 0.0012893660459667444, + 0.0007687549223192036, + 0.0009653132292442024, + 0.0006923119071871042, + 0.0007630267064087093, + 0.0006074964185245335, + 0.0004294700629543513, + 0.0008983865845948458, + 0.0007349415100179613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.04306088760495186 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280736058950424 + ], + "opposite_logprob_tokens": [ + -3.773280620574951 + ], + "logprob_reward": -0.023280736058950424, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280736058950424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09223435819149017, + 0.10987639427185059, + 0.05278145521879196, + 0.06962303072214127, + 0.11553580313920975, + 0.13475069403648376, + 0.03868117928504944, + 0.3567640483379364, + 0.014488698914647102, + 0.01526428759098053 + ], + "credit_map": [ + 0.005809377878904343, + 0.013338829390704632, + 0.008637712337076664, + 0.011534616351127625, + 0.01317650731652975, + 0.012454558163881302, + 0.014405864290893078, + 0.012876888737082481, + 0.006029426120221615, + 0.00746768107637763, + 0.007514627650380135, + 0.007326293736696243, + 0.030253689736127853, + 0.006571871694177389, + 0.03588614985346794, + 0.008826653473079205, + 0.0037404175382107496, + 0.005912299267947674, + 0.005562842823565006, + 0.010150696150958538, + 0.009737717919051647, + 0.003817033488303423, + 0.008475721813738346, + 0.005384728312492371, + 0.004322079475969076, + 0.01004602201282978, + 0.0059676882810890675, + 0.006535186432301998, + 0.012298536486923695, + 0.0053237443789839745, + 0.012115449644625187, + 0.013014325872063637, + 0.005961880553513765, + 0.0063290405087172985, + 0.007427744567394257, + 0.008752704598009586, + 0.034606583416461945, + 0.0041898940689861774, + 0.03932856768369675, + 0.008939393796026707, + 0.010578270070254803, + 0.0116664357483387, + 0.009141852147877216, + 0.01335114799439907, + 0.01596287451684475, + 0.014197919517755508, + 0.03314363583922386, + 0.02670856937766075, + 0.0027583311311900616, + 0.0047178990207612514, + 0.00337313418276608, + 0.005515784025192261, + 0.005017081741243601, + 0.00546601228415966, + 0.008091820403933525, + 0.003741115564480424, + 0.028796836733818054, + 0.026116622611880302, + 0.08262623846530914, + 0.03292808681726456, + 0.03338263928890228, + 0.027564577758312225, + 0.04278262332081795, + 0.08256643265485764, + 0.0010958045022562146, + 0.0013143649557605386, + 0.002756359986960888, + 0.002296459162607789, + 0.0013334399554878473, + 0.0014014674816280603, + 0.0025328302290290594, + 0.0017579731065779924, + 0.001382717746309936, + 0.0016996454214677215, + 0.002416520845144987, + 0.001959420507773757, + 0.0012814762303605676, + 0.0012781624682247639, + 0.002500829752534628, + 0.0027455142699182034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.023280736058950424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle, dolly shot moves across a rain-soaked, muddy clearing, approaching the massive base of an ancient oak tree. The camera begins its upward, heavy lift, navigating the storm. The tree's massive, gnarled branches, as thick as a human trunk, twist and strain violently under the relentless assault of unseen gales, groaning audibly under the immense pressure. The entire tree sways, its dark, damp bark deeply furrowed like ancient leather, speckled with shimmering green moss, catching the dull, steel-gray glow of the storm-choked sky. The camera weaves between the flapping lower limbs, capturing how smaller branches whip back and forth frantically, while tufts of dark green leaves are torn off and flung into a swirling, bruised, purple cloud. As torrential rain slants across the frame, the atmosphere is one of primal, violent struggle, a testament to the enduring power against overwhelming force. The tree's powerful, agonized movements convey a desperate fight for survival.", + "seed": 11673, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_0/index_7254_seed_11673_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1415589855052531, + "vlm_logprob": -0.1415589855052531, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.1415589855052531, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015915267169475555 + ], + "opposite_logprob_tokens": [ + -4.265915393829346 + ], + "logprob_reward": -0.015915267169475555, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015915267169475555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16332308948040009, + 0.08210250735282898, + 0.0740169808268547, + 0.26800966262817383, + 0.048144083470106125, + 0.10212786495685577, + 0.17360149323940277, + 0.040210358798503876, + 0.02007213607430458, + 0.02839188277721405 + ], + "credit_map": [ + 0.014036775566637516, + 0.023563135415315628, + 0.01684352569282055, + 0.012531117536127567, + 0.03695514798164368, + 0.010627204552292824, + 0.033053234219551086, + 0.015712937340140343, + 0.006586861331015825, + 0.006622453220188618, + 0.006297329906374216, + 0.008643856272101402, + 0.01717127487063408, + 0.006621069740504026, + 0.02063622884452343, + 0.009523422457277775, + 0.010397656820714474, + 0.006011960096657276, + 0.010896234773099422, + 0.009896693751215935, + 0.007701970636844635, + 0.005884082987904549, + 0.013230259530246258, + 0.00999811664223671, + 0.09007308632135391, + 0.014417522586882114, + 0.01750141568481922, + 0.014794522896409035, + 0.017181921750307083, + 0.013154291547834873, + 0.023197287693619728, + 0.07768958806991577, + 0.006944092456251383, + 0.007152896840125322, + 0.007879120297729969, + 0.007167380303144455, + 0.003046548692509532, + 0.003325073979794979, + 0.005975665990263224, + 0.006653298158198595, + 0.008150284178555012, + 0.008981089107692242, + 0.0318484753370285, + 0.018011130392551422, + 0.005223480053246021, + 0.005854068323969841, + 0.008076886646449566, + 0.015982437878847122, + 0.013802994042634964, + 0.026896152645349503, + 0.033763233572244644, + 0.021192187443375587, + 0.007420843001455069, + 0.010544397868216038, + 0.015806082636117935, + 0.044175587594509125, + 0.004197091329842806, + 0.005658112466335297, + 0.004847478121519089, + 0.00611330009996891, + 0.005225300323218107, + 0.003987612668424845, + 0.004929219838231802, + 0.0052522397600114346, + 0.003680508118122816, + 0.0019911075942218304, + 0.00220950273796916, + 0.0038922419771552086, + 0.002262907335534692, + 0.0012034589890390635, + 0.002316955244168639, + 0.0025154531467705965, + 0.003937905188649893, + 0.0034691907931119204, + 0.0024035561364144087, + 0.003935565706342459, + 0.0025635012425482273, + 0.003778889775276184, + 0.005381870549172163, + 0.002921400824561715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.015915267169475555 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006748266983777285 + ], + "opposite_logprob_tokens": [ + -5.006748199462891 + ], + "logprob_reward": -0.006748266983777285, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748266983777285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1479589194059372, + 0.211601123213768, + 0.039226312190294266, + 0.12349238246679306, + 0.06365424394607544, + 0.034657981246709824, + 0.14436431229114532, + 0.038607578724622726, + 0.07389182597398758, + 0.12254530936479568 + ], + "credit_map": [ + 0.008558868430554867, + 0.019234642386436462, + 0.01240323856472969, + 0.010943259112536907, + 0.044030360877513885, + 0.010564801283180714, + 0.03072248585522175, + 0.011501283384859562, + 0.020012525841593742, + 0.00631356704980135, + 0.005975416861474514, + 0.014329648576676846, + 0.05965368449687958, + 0.006408706773072481, + 0.08047128468751907, + 0.018436308950185776, + 0.0034310524351894855, + 0.00421534851193428, + 0.004319459665566683, + 0.004464538302272558, + 0.005661196541041136, + 0.004117157310247421, + 0.007740224711596966, + 0.005277338903397322, + 0.01980186253786087, + 0.012845034711062908, + 0.011985340155661106, + 0.017578957602381706, + 0.015741772949695587, + 0.015234984457492828, + 0.014715863391757011, + 0.015588576905429363, + 0.0064153955318033695, + 0.007122018840163946, + 0.013399608433246613, + 0.008620352484285831, + 0.0032945701386779547, + 0.003666760167106986, + 0.007262486033141613, + 0.013873058371245861, + 0.0030875862576067448, + 0.003687486285343766, + 0.008412111550569534, + 0.007735831197351217, + 0.002317242557182908, + 0.0023296657018363476, + 0.0027818214148283005, + 0.004306239541620016, + 0.011212220415472984, + 0.016077393665909767, + 0.03118165209889412, + 0.019384505227208138, + 0.007028345484286547, + 0.00896421167999506, + 0.011785570532083511, + 0.038730427622795105, + 0.004415667150169611, + 0.004042543470859528, + 0.005353201180696487, + 0.007379576098173857, + 0.003818153403699398, + 0.004054732155054808, + 0.0043860794976353645, + 0.005157629493623972, + 0.0073835114017128944, + 0.007729134988039732, + 0.010769673623144627, + 0.015353045426309109, + 0.008781826123595238, + 0.004872024990618229, + 0.008314153179526329, + 0.010688460431993008, + 0.019198136404156685, + 0.021423636004328728, + 0.014416182413697243, + 0.014171246439218521, + 0.012091134674847126, + 0.013252521865069866, + 0.01555297989398241, + 0.012439479120075703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.006748266983777285 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328380107879639 + ], + "opposite_logprob_tokens": [ + -0.7578380107879639 + ], + "logprob_reward": -0.6328380107879639, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328380107879639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1719299852848053, + 0.08660825341939926, + 0.041966311633586884, + 0.0720660388469696, + 0.08811476081609726, + 0.37804755568504333, + 0.0852769985795021, + 0.024338901042938232, + 0.02256394922733307, + 0.029087213799357414 + ], + "credit_map": [ + 0.011121844872832298, + 0.023556413128972054, + 0.012701104395091534, + 0.009960012510418892, + 0.0537797249853611, + 0.01069654244929552, + 0.037518929690122604, + 0.01259540393948555, + 0.007511894218623638, + 0.003267261665314436, + 0.002717783907428384, + 0.006611799821257591, + 0.024094048887491226, + 0.0032457122579216957, + 0.03139524534344673, + 0.007764510810375214, + 0.004660772159695625, + 0.00409351522102952, + 0.004815342370420694, + 0.006274781189858913, + 0.004613849800080061, + 0.005289762280881405, + 0.006900981534272432, + 0.00531730754300952, + 0.011323180049657822, + 0.006338142324239016, + 0.007202078588306904, + 0.008475858718156815, + 0.009548981674015522, + 0.009525570087134838, + 0.009176459163427353, + 0.0104757659137249, + 0.008131048642098904, + 0.007929173298180103, + 0.025539901107549667, + 0.010390936397016048, + 0.004405038896948099, + 0.004286639858037233, + 0.008051113225519657, + 0.019380908459424973, + 0.023703668266534805, + 0.021318553015589714, + 0.12247909605503082, + 0.04492635279893875, + 0.016779504716396332, + 0.04617413133382797, + 0.02126299776136875, + 0.08140324801206589, + 0.00679405452683568, + 0.009485950693488121, + 0.014143800362944603, + 0.01344976481050253, + 0.007343444041907787, + 0.005860614124685526, + 0.011466610245406628, + 0.016732752323150635, + 0.0026396254543215036, + 0.002556371036916971, + 0.0033319753129035234, + 0.0047920336946845055, + 0.0022000225726515055, + 0.0018199371406808496, + 0.003293644869700074, + 0.0037052910774946213, + 0.002366922562941909, + 0.002595755271613598, + 0.0031560370698571205, + 0.006015917286276817, + 0.001997252693399787, + 0.0012868617195636034, + 0.0022795042023062706, + 0.002865696558728814, + 0.003569840220734477, + 0.0038624187000095844, + 0.003367391414940357, + 0.005806543864309788, + 0.003056739456951618, + 0.002276831306517124, + 0.0040395003743469715, + 0.003107947064563632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.6328380107879639 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810384124517441 + ], + "opposite_logprob_tokens": [ + -3.2881038188934326 + ], + "logprob_reward": -0.03810384124517441, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810384124517441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49005603790283203, + 0.0384720042347908, + 0.039958491921424866, + 0.09749657660722733, + 0.09771157056093216, + 0.05895640328526497, + 0.06792750209569931, + 0.05738074705004692, + 0.01899975910782814, + 0.033040910959243774 + ], + "credit_map": [ + 0.039546869695186615, + 0.05731517821550369, + 0.030279681086540222, + 0.02931467816233635, + 0.16307663917541504, + 0.030879607424139977, + 0.10563276708126068, + 0.034010618925094604, + 0.004414498805999756, + 0.002968026790767908, + 0.0022063045762479305, + 0.0044449856504797935, + 0.007024375721812248, + 0.003078349167481065, + 0.010045952163636684, + 0.004289510194212198, + 0.004893267992883921, + 0.0038530135061591864, + 0.00527943205088377, + 0.004060069564729929, + 0.005861217156052589, + 0.004310152027755976, + 0.0069150375202298164, + 0.004786301404237747, + 0.024806242436170578, + 0.006788173690438271, + 0.008437512442469597, + 0.009733582846820354, + 0.008703477680683136, + 0.006829033140093088, + 0.00936585571616888, + 0.02283269725739956, + 0.009414702653884888, + 0.00978738721460104, + 0.024692893028259277, + 0.011799711734056473, + 0.004915608558803797, + 0.00531786261126399, + 0.010187874548137188, + 0.021595533937215805, + 0.004484447650611401, + 0.00562400883063674, + 0.008403408341109753, + 0.010313793085515499, + 0.004609805066138506, + 0.01059682760387659, + 0.006580783054232597, + 0.008343327790498734, + 0.007403258699923754, + 0.00713649857789278, + 0.01480057556182146, + 0.01346524152904749, + 0.005554886534810066, + 0.005800905637443066, + 0.006583248730748892, + 0.007182886358350515, + 0.007051852997392416, + 0.0062236180528998375, + 0.009100017137825489, + 0.012414258904755116, + 0.0044652242213487625, + 0.005005436018109322, + 0.006358674727380276, + 0.006761664059013128, + 0.0017057913355529308, + 0.001337180146947503, + 0.003108946606516838, + 0.004346105270087719, + 0.0021142836194485426, + 0.001307235099375248, + 0.0018791374750435352, + 0.003201080020517111, + 0.004884845111519098, + 0.006342492997646332, + 0.002771747065708041, + 0.004665685351938009, + 0.003388968762010336, + 0.002670841757208109, + 0.005434890277683735, + 0.002881438937038183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.03810384124517441 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189541339874268 + ], + "opposite_logprob_tokens": [ + -4.264189720153809 + ], + "logprob_reward": -0.014189541339874268, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189541339874268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11890556663274765, + 0.07507943361997604, + 0.023742206394672394, + 0.11325278878211975, + 0.0733601301908493, + 0.22090454399585724, + 0.19097644090652466, + 0.02235369011759758, + 0.055145327001810074, + 0.10627979040145874 + ], + "credit_map": [ + 0.008654708042740822, + 0.014551757834851742, + 0.006515459157526493, + 0.006494678556919098, + 0.03901570662856102, + 0.005940366070717573, + 0.02901592291891575, + 0.008716970682144165, + 0.00542060611769557, + 0.007482865359634161, + 0.004662205930799246, + 0.016970915719866753, + 0.011897635646164417, + 0.006267308723181486, + 0.013887845911085606, + 0.00849005114287138, + 0.0032077161595225334, + 0.00229746475815773, + 0.0034073239658027887, + 0.00344871380366385, + 0.002572520636022091, + 0.002160544041544199, + 0.0034057763405144215, + 0.003242149716243148, + 0.029052235186100006, + 0.008136366493999958, + 0.00833207368850708, + 0.01228383183479309, + 0.009429188445210457, + 0.010739368386566639, + 0.011075782589614391, + 0.02420394867658615, + 0.0056236847303807735, + 0.00637327553704381, + 0.02138720639050007, + 0.008324744179844856, + 0.0037121695932000875, + 0.0032586432062089443, + 0.00722698587924242, + 0.017453426495194435, + 0.013744420371949673, + 0.012786228209733963, + 0.053711503744125366, + 0.02964061312377453, + 0.015101770870387554, + 0.03795268386602402, + 0.018575048074126244, + 0.03939228504896164, + 0.016706304624676704, + 0.01875610463321209, + 0.04078539088368416, + 0.029254503548145294, + 0.011336952447891235, + 0.011399916373193264, + 0.017508303746581078, + 0.0452289804816246, + 0.0023534316569566727, + 0.002986552892252803, + 0.0028350537177175283, + 0.0033550795633345842, + 0.0023373235017061234, + 0.002275630598887801, + 0.0028940022457391024, + 0.003316618502140045, + 0.005154700018465519, + 0.00382238975726068, + 0.009917747229337692, + 0.007824013940989971, + 0.005093814805150032, + 0.003509683068841696, + 0.008237010799348354, + 0.011585974134504795, + 0.01764548569917679, + 0.015909088775515556, + 0.009880919009447098, + 0.013154969550669193, + 0.013306664302945137, + 0.009352822788059711, + 0.016196655109524727, + 0.010833192616701126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.014189541339874268 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle, dolly shot moves across a rain-soaked, muddy clearing, approaching the massive base of an ancient oak tree. The camera begins its upward, heavy lift, navigating the storm. The tree's massive, gnarled branches, as thick as a human trunk, twist and strain violently under the relentless assault of unseen gales, groaning audibly under the immense pressure. The entire tree sways, its dark, damp bark deeply furrowed like ancient leather, speckled with shimmering green moss, catching the dull, steel-gray glow of the storm-choked sky. The camera weaves between the flapping lower limbs, capturing how smaller branches whip back and forth frantically, while tufts of dark green leaves are torn off and flung into a swirling, bruised, purple cloud. As torrential rain slants across the frame, the atmosphere is one of primal, violent struggle, a testament to the enduring power against overwhelming force. The tree's powerful, agonized movements convey a desperate fight for survival.", + "seed": 11674, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_0/index_7254_seed_11674_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.100122738070786, + "vlm_logprob": -0.100122738070786, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.100122738070786, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016649113968014717 + ], + "opposite_logprob_tokens": [ + -4.39164924621582 + ], + "logprob_reward": -0.016649113968014717, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.016649113968014717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4514627754688263, + 0.07171507924795151, + 0.08461321145296097, + 0.04321976751089096, + 0.05732331424951553, + 0.07372517883777618, + 0.03872082382440567, + 0.07412748038768768, + 0.03220574930310249, + 0.07288654893636703 + ], + "credit_map": [ + 0.03821999207139015, + 0.040289416909217834, + 0.12065684050321579, + 0.07244504243135452, + 0.03480290621519089, + 0.036430805921554565, + 0.045529305934906006, + 0.06308853626251221, + 0.006060989107936621, + 0.004051702097058296, + 0.009198456071317196, + 0.02108915150165558, + 0.004141463432461023, + 0.004338746890425682, + 0.013292785733938217, + 0.009541792795062065, + 0.0075018322095274925, + 0.007040213327854872, + 0.017423778772354126, + 0.011642674915492535, + 0.008058194071054459, + 0.006233801133930683, + 0.015399636700749397, + 0.011313095688819885, + 0.003177883569151163, + 0.00463877385482192, + 0.008114706724882126, + 0.006661334540694952, + 0.0030112946406006813, + 0.004923999309539795, + 0.0063312160782516, + 0.006360567640513182, + 0.005716024432331324, + 0.003934574779123068, + 0.01058386079967022, + 0.013542139902710915, + 0.00239929324015975, + 0.0027602571062743664, + 0.012409351766109467, + 0.00597782339900732, + 0.0064399754628539085, + 0.008083458989858627, + 0.009749268181622028, + 0.014764482155442238, + 0.008189712651073933, + 0.007424939423799515, + 0.0105880843475461, + 0.008485266007483006, + 0.0052704354748129845, + 0.00696422578766942, + 0.0039078062400221825, + 0.006486853584647179, + 0.0032280839513987303, + 0.0036795521154999733, + 0.004489541053771973, + 0.004694334696978331, + 0.009713506326079369, + 0.007893471978604794, + 0.011330940760672092, + 0.006544717121869326, + 0.007430460769683123, + 0.0058125960640609264, + 0.009007556363940239, + 0.016394246369600296, + 0.004500271286815405, + 0.005158655345439911, + 0.004091206938028336, + 0.00619334215298295, + 0.0024259095080196857, + 0.0034411910455673933, + 0.003709055483341217, + 0.002686123363673687, + 0.007283987011760473, + 0.007228815462440252, + 0.018695365637540817, + 0.007369175553321838, + 0.003639677306637168, + 0.008722485043108463, + 0.004840170498937368, + 0.01510688103735447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.016649113968014717 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004653339274227619 + ], + "opposite_logprob_tokens": [ + -5.379653453826904 + ], + "logprob_reward": -0.004653339274227619, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004653339274227619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1254117488861084, + 0.07956189662218094, + 0.0984458476305008, + 0.08580957353115082, + 0.10852091014385223, + 0.048673730343580246, + 0.04226422309875488, + 0.17230172455310822, + 0.05614645034074783, + 0.18286390602588654 + ], + "credit_map": [ + 0.008602948859333992, + 0.013829305768013, + 0.026634573936462402, + 0.02714988961815834, + 0.01030062884092331, + 0.011530568823218346, + 0.013979136012494564, + 0.013384690508246422, + 0.007201694883406162, + 0.007336498238146305, + 0.0125223807990551, + 0.013560201041400433, + 0.008984038606286049, + 0.008908490650355816, + 0.010388705879449844, + 0.01065988838672638, + 0.006550292018800974, + 0.00868171826004982, + 0.014322422444820404, + 0.02290201745927334, + 0.010685940273106098, + 0.007512924261391163, + 0.019020967185497284, + 0.008769560605287552, + 0.00423470651730895, + 0.008601674810051918, + 0.022594954818487167, + 0.01105718407779932, + 0.00492525938898325, + 0.0053556994535028934, + 0.01804768294095993, + 0.010992413386702538, + 0.009139569476246834, + 0.008466081693768501, + 0.023082349449396133, + 0.016898151487112045, + 0.007309314329177141, + 0.004853497724980116, + 0.030337130650877953, + 0.00843480508774519, + 0.006610441952943802, + 0.005343616474419832, + 0.005106641445308924, + 0.006558343302458525, + 0.008263763040304184, + 0.005757336039096117, + 0.0060861497186124325, + 0.004947434179484844, + 0.00716195534914732, + 0.00973909255117178, + 0.004764278419315815, + 0.00489765452221036, + 0.0033704794477671385, + 0.002937054494395852, + 0.005084290634840727, + 0.00430941442027688, + 0.023142380639910698, + 0.015855705365538597, + 0.025607986375689507, + 0.014049850404262543, + 0.02051614783704281, + 0.013239325024187565, + 0.02062518708407879, + 0.039265140891075134, + 0.007959889248013496, + 0.011401399970054626, + 0.007449130993336439, + 0.006519263610243797, + 0.004874793346971273, + 0.005739358253777027, + 0.008439499884843826, + 0.0037631154991686344, + 0.019046848639845848, + 0.02513829618692398, + 0.04321718588471413, + 0.015618895180523396, + 0.01093385647982359, + 0.023030024021863937, + 0.010541961528360844, + 0.035336825996637344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.004653339274227619 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42893123626708984 + ], + "opposite_logprob_tokens": [ + -1.0539312362670898 + ], + "logprob_reward": -0.42893123626708984, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42893123626708984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26356303691864014, + 0.23528172075748444, + 0.163059800863266, + 0.08882958441972733, + 0.03363325819373131, + 0.03873066604137421, + 0.06227056682109833, + 0.02456960827112198, + 0.017685458064079285, + 0.07237628847360611 + ], + "credit_map": [ + 0.01891474425792694, + 0.02700832299888134, + 0.06515945494174957, + 0.04951008781790733, + 0.02362547256052494, + 0.017852334305644035, + 0.03338250890374184, + 0.028110120445489883, + 0.015893302857875824, + 0.014331890270113945, + 0.03236890584230423, + 0.057119619101285934, + 0.022067053243517876, + 0.022560816258192062, + 0.0410335510969162, + 0.02990657649934292, + 0.012251734733581543, + 0.010446514002978802, + 0.027890753000974655, + 0.029187476262450218, + 0.020613545551896095, + 0.012798693962395191, + 0.030789677053689957, + 0.019081396982073784, + 0.006910358555614948, + 0.013468841090798378, + 0.014088024385273457, + 0.012401700951159, + 0.00943406205624342, + 0.01020564790815115, + 0.012180810794234276, + 0.010140142403542995, + 0.00465176859870553, + 0.004581830929964781, + 0.005665142554789782, + 0.006654292345046997, + 0.001699104905128479, + 0.002637425670400262, + 0.0049306838773190975, + 0.00281300931237638, + 0.005286056082695723, + 0.004838890861719847, + 0.003732837038114667, + 0.005757092032581568, + 0.00620714807882905, + 0.002977056661620736, + 0.00620223069563508, + 0.003729352029040456, + 0.010440493933856487, + 0.014088382944464684, + 0.007516882382333279, + 0.006706165615469217, + 0.005355899687856436, + 0.004576520528644323, + 0.007388448342680931, + 0.00619776826351881, + 0.004259317182004452, + 0.0035825225058943033, + 0.0035223611630499363, + 0.0028466852381825447, + 0.0017768261022865772, + 0.002042905893176794, + 0.002582770772278309, + 0.003956218250095844, + 0.0022114196326583624, + 0.003540745237842202, + 0.0024788675364106894, + 0.0029075047932565212, + 0.0014900770038366318, + 0.0016745132161304355, + 0.001680290326476097, + 0.0017020406667143106, + 0.006632483098655939, + 0.014654615893959999, + 0.010864721611142159, + 0.007952813059091568, + 0.004895019344985485, + 0.012931866571307182, + 0.0051575833931565285, + 0.009287182241678238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.42893123626708984 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803065583109856 + ], + "opposite_logprob_tokens": [ + -3.5298030376434326 + ], + "logprob_reward": -0.029803065583109856, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803065583109856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1823536902666092, + 0.03212682157754898, + 0.044507864862680435, + 0.0595594085752964, + 0.15059517323970795, + 0.08538642525672913, + 0.07837453484535217, + 0.29424574971199036, + 0.03141878545284271, + 0.04143156111240387 + ], + "credit_map": [ + 0.015819072723388672, + 0.017374521121382713, + 0.062240343540906906, + 0.03814966604113579, + 0.009705569595098495, + 0.010320217348635197, + 0.01467245165258646, + 0.014071856625378132, + 0.0028107045218348503, + 0.0026609485503286123, + 0.0055426121689379215, + 0.005212605465203524, + 0.004397754557430744, + 0.00361357512883842, + 0.004250325728207827, + 0.0036382940597832203, + 0.00386734283529222, + 0.0032801153138279915, + 0.009060605429112911, + 0.005877508316189051, + 0.004722746089100838, + 0.0035062066745013, + 0.008800022304058075, + 0.005393317900598049, + 0.003203746397048235, + 0.0038975076749920845, + 0.021672841161489487, + 0.005871964618563652, + 0.0030212239362299442, + 0.0024396206717938185, + 0.010558359324932098, + 0.008894141763448715, + 0.011939656920731068, + 0.011853448115289211, + 0.034683242440223694, + 0.01482994295656681, + 0.007613606750965118, + 0.0038615341763943434, + 0.057308342307806015, + 0.008505402132868767, + 0.009205886162817478, + 0.007513437885791063, + 0.011014489457011223, + 0.016251305118203163, + 0.01113396231085062, + 0.005733562167733908, + 0.015691032633185387, + 0.008842752315104008, + 0.010785076767206192, + 0.010189204476773739, + 0.01276241522282362, + 0.018036944791674614, + 0.005289926193654537, + 0.005856324452906847, + 0.006655439734458923, + 0.008799202740192413, + 0.03841770812869072, + 0.023260965943336487, + 0.04354412481188774, + 0.02226298674941063, + 0.03579328581690788, + 0.025008326396346092, + 0.034055497497320175, + 0.07190287113189697, + 0.003734828671440482, + 0.005195934325456619, + 0.004294600337743759, + 0.0050501953810453415, + 0.0032347284723073244, + 0.0035025898832827806, + 0.0036067324690520763, + 0.0027991782408207655, + 0.003941680770367384, + 0.00748674338683486, + 0.009193718433380127, + 0.007510368712246418, + 0.0030575923155993223, + 0.004340642131865025, + 0.00292678433470428, + 0.002974032424390316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.029803065583109856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a rain-soaked, muddy clearing with an ancient oak tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree have massive, gnarled branches that are dark, damp, and speckled with green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tree swaying violently with smaller branches whipping back and forth due to strong winds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a dull, steel-gray color with a swirling, bruised purple cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle dolly shot approaching the tree base and then perform an upward heavy lift?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057693526148796 + ], + "opposite_logprob_tokens": [ + -3.8955769538879395 + ], + "logprob_reward": -0.02057693526148796, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057693526148796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21813616156578064, + 0.06913800537586212, + 0.051737237721681595, + 0.08264562487602234, + 0.07986889034509659, + 0.05198235437273979, + 0.10321547091007233, + 0.10176677256822586, + 0.14453190565109253, + 0.09697757661342621 + ], + "credit_map": [ + 0.018485745415091515, + 0.021846260875463486, + 0.04031335189938545, + 0.038373131304979324, + 0.02382347546517849, + 0.017424220219254494, + 0.02636488899588585, + 0.031505074352025986, + 0.004400373436510563, + 0.006894588936120272, + 0.012352127581834793, + 0.009492681361734867, + 0.008953608572483063, + 0.007683048024773598, + 0.010707513429224491, + 0.00865405984222889, + 0.0036844355054199696, + 0.004137291572988033, + 0.007890348322689533, + 0.008201600052416325, + 0.00727815181016922, + 0.00437599141150713, + 0.009255655109882355, + 0.006913770455867052, + 0.005578026175498962, + 0.010985100641846657, + 0.014561647549271584, + 0.011776967905461788, + 0.005154861602932215, + 0.005858851131051779, + 0.016477268189191818, + 0.01225290261209011, + 0.008499749936163425, + 0.007490108255296946, + 0.012255902402102947, + 0.019503887742757797, + 0.003759883577004075, + 0.005041619297116995, + 0.016369327902793884, + 0.006948411464691162, + 0.006054552737623453, + 0.006311065051704645, + 0.007191300857812166, + 0.009494271129369736, + 0.004336167126893997, + 0.004351636860519648, + 0.008067510090768337, + 0.006175856571644545, + 0.015934858471155167, + 0.01634475588798523, + 0.01732451468706131, + 0.015163584612309933, + 0.008546045981347561, + 0.007295261602848768, + 0.011754848062992096, + 0.010851605795323849, + 0.013446562923491001, + 0.00714462623000145, + 0.016836561262607574, + 0.008827137760818005, + 0.01108500175178051, + 0.008865376003086567, + 0.01092453021556139, + 0.024636978283524513, + 0.016281772404909134, + 0.019669560715556145, + 0.020137453451752663, + 0.03334566205739975, + 0.009264202788472176, + 0.013342658057808876, + 0.015375806018710136, + 0.01711479388177395, + 0.009693566709756851, + 0.018005119636654854, + 0.011360225267708302, + 0.013793700374662876, + 0.009174531325697899, + 0.015483878552913666, + 0.009435911662876606, + 0.010030639357864857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.02057693526148796 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_1/index_tensor([383, 383])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_1/index_tensor([383, 383])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7e573a37c1072aaa5af931574ca194e6ea950ba8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_1/index_tensor([383, 383])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a slightly overhead angle, the camera held still, focused on a man in a flowing white silk robe. His long, dark hair was meticulously tied into a high bun, secured with an antique wooden hairpin. His profile was sharply defined, his eyes slightly downcast, his expression focused and serene. He sat at a rustic wooden table, upon which lay a dark brown guqin, its lacquer surface delicately cracked by age. His long, bony fingers delicately stroked the taut silk strings, his fingertips gently gliding across them. The movement was steady and deliberate, more a reverent exchange with the instrument than a performance. A soft beam of afternoon sunlight slanted through the blurred carved wooden window in the background, skewing the air with floating dust particles and casting mottled shadows on the lacquered surface of the guqin. The main colors of the entire picture are quiet wood color and pure white, illuminated by warm sunlight, creating an atmosphere of extreme tranquility, full of Zen and classical elegance.", + "seed": 6482, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_1/index_383_seed_6482_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.014203900936990976, + "vlm_logprob": -0.014203900936990976, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.014203900936990976, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011986489407718182 + ], + "opposite_logprob_tokens": [ + -6.136986255645752 + ], + "logprob_reward": -0.011986489407718182, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.011986489407718182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11858358234167099, + 0.03944027051329613, + 0.09668590873479843, + 0.07182660698890686, + 0.04536538943648338, + 0.02997704967856407, + 0.035075314342975616, + 0.051496900618076324, + 0.042439885437488556, + 0.46910905838012695 + ], + "credit_map": [ + 0.019340740516781807, + 0.013757356442511082, + 0.016147209331393242, + 0.011705692857503891, + 0.011976313777267933, + 0.01378504280000925, + 0.01162985060364008, + 0.0202413909137249, + 0.004898367915302515, + 0.005994182080030441, + 0.005778594873845577, + 0.006180488038808107, + 0.004242832772433758, + 0.0036021799314767122, + 0.004498062189668417, + 0.00424556527286768, + 0.012357181869447231, + 0.02456986904144287, + 0.008551870472729206, + 0.007353075779974461, + 0.01722661405801773, + 0.00883671548217535, + 0.006837234832346439, + 0.010953359305858612, + 0.00812008697539568, + 0.012043825350701809, + 0.010192617774009705, + 0.008371071889996529, + 0.011879636906087399, + 0.010525061748921871, + 0.005093944258987904, + 0.005600367207080126, + 0.006725342944264412, + 0.005084960255771875, + 0.005142990965396166, + 0.004150079563260078, + 0.007115037180483341, + 0.008034853264689445, + 0.003329087048768997, + 0.005783043801784515, + 0.004516863729804754, + 0.005334235727787018, + 0.0027284135576337576, + 0.002913043834269047, + 0.0033896341919898987, + 0.004359399899840355, + 0.0036533079110085964, + 0.003082155017182231, + 0.004888276569545269, + 0.010806815698742867, + 0.0033504203893244267, + 0.004225166514515877, + 0.003306162776425481, + 0.0030408750753849745, + 0.0023120814003050327, + 0.003145515685901046, + 0.007216401398181915, + 0.011220990680158138, + 0.007881791330873966, + 0.005248439498245716, + 0.004991080611944199, + 0.005903763696551323, + 0.0034683605190366507, + 0.005566077306866646, + 0.006568357348442078, + 0.008378512226045132, + 0.005412867292761803, + 0.005000900477170944, + 0.0055253757163882256, + 0.004943348001688719, + 0.002793697640299797, + 0.0038168311584740877, + 0.056437574326992035, + 0.029542356729507446, + 0.18607157468795776, + 0.023799803107976913, + 0.017239192500710487, + 0.027865927666425705, + 0.020819183439016342, + 0.1073334738612175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.011986489407718182 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812633082270622 + ], + "opposite_logprob_tokens": [ + -3.288126230239868 + ], + "logprob_reward": -0.03812633082270622, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812633082270622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27644121646881104, + 0.07111414521932602, + 0.11732613295316696, + 0.0891248807311058, + 0.04550357162952423, + 0.05396810546517372, + 0.018175175413489342, + 0.055676016956567764, + 0.04558888450264931, + 0.2270818054676056 + ], + "credit_map": [ + 0.04622844234108925, + 0.040638603270053864, + 0.03222162276506424, + 0.02633567340672016, + 0.03423214703798294, + 0.028941549360752106, + 0.022842008620500565, + 0.045001205056905746, + 0.007627503946423531, + 0.008106584660708904, + 0.010542649775743484, + 0.010469714179635048, + 0.009531673975288868, + 0.010383965447545052, + 0.007861156947910786, + 0.006590908858925104, + 0.013892932794988155, + 0.0205196775496006, + 0.014104022644460201, + 0.011088159866631031, + 0.019150467589497566, + 0.02083229273557663, + 0.00907581951469183, + 0.008662777952849865, + 0.008164364844560623, + 0.012426027096807957, + 0.009591354988515377, + 0.00906235072761774, + 0.01950211636722088, + 0.012062110006809235, + 0.010268318466842175, + 0.008048243820667267, + 0.004874935373663902, + 0.005694741383194923, + 0.0054033356718719006, + 0.005662970710545778, + 0.009827311150729656, + 0.006140843033790588, + 0.004301715642213821, + 0.003597722854465246, + 0.008869647979736328, + 0.007661959156394005, + 0.0051757777109742165, + 0.004873214289546013, + 0.006189154926687479, + 0.008166534826159477, + 0.006041227839887142, + 0.0069905961863696575, + 0.002758557675406337, + 0.0018799083773046732, + 0.002243154216557741, + 0.0023974308278411627, + 0.0027338718064129353, + 0.003371758386492729, + 0.001367362798191607, + 0.0014231340028345585, + 0.007047981023788452, + 0.007882172241806984, + 0.00547716673463583, + 0.009492230601608753, + 0.0067721279338002205, + 0.005958600901067257, + 0.006173630245029926, + 0.006872113794088364, + 0.004837576299905777, + 0.005907893180847168, + 0.00836796686053276, + 0.006813609506934881, + 0.0035226349718868732, + 0.004980576690286398, + 0.0045334696769714355, + 0.006625157780945301, + 0.03347300738096237, + 0.01531905122101307, + 0.09105093032121658, + 0.012000194750726223, + 0.011404543183743954, + 0.011771914549171925, + 0.00940062664449215, + 0.04266155883669853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.03812633082270622 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631433662027121 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631433662027121, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631433662027121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15495172142982483, + 0.03480830416083336, + 0.0756789967417717, + 0.09424212574958801, + 0.036814745515584946, + 0.042675476521253586, + 0.022722573950886726, + 0.03956111520528793, + 0.039116982370615005, + 0.45942792296409607 + ], + "credit_map": [ + 0.031478069722652435, + 0.0229662973433733, + 0.02129891887307167, + 0.012644019909203053, + 0.009691768325865269, + 0.01562102884054184, + 0.010321618989109993, + 0.03092999756336212, + 0.0041811503469944, + 0.00835083145648241, + 0.0043447595089674, + 0.004915366414934397, + 0.0031133939046412706, + 0.0037350754719227552, + 0.0031656224746257067, + 0.0030021064449101686, + 0.010753124952316284, + 0.01619734987616539, + 0.005375997629016638, + 0.006029736250638962, + 0.013350039720535278, + 0.013430455699563026, + 0.0043574133887887, + 0.006184878759086132, + 0.010995317250490189, + 0.01971258968114853, + 0.009105934761464596, + 0.010855656117200851, + 0.01450551301240921, + 0.015059261582791805, + 0.006579095497727394, + 0.007428760174661875, + 0.004399917554110289, + 0.0037491992115974426, + 0.003962450195103884, + 0.0057091182097792625, + 0.005288402084261179, + 0.006985741667449474, + 0.002851522993296385, + 0.0038683926686644554, + 0.00605238601565361, + 0.00359478872269392, + 0.007206611335277557, + 0.0040396288968622684, + 0.0035184016451239586, + 0.006763184443116188, + 0.005301859229803085, + 0.006198612507432699, + 0.0027783093973994255, + 0.0031051887199282646, + 0.0037468785885721445, + 0.0036437353119254112, + 0.0015403121942654252, + 0.0019466717494651675, + 0.003055051900446415, + 0.0029064256232231855, + 0.005171075463294983, + 0.0057996115647256374, + 0.00700050825253129, + 0.005020269192755222, + 0.004507094621658325, + 0.004173454362899065, + 0.0035133056808263063, + 0.004375798162072897, + 0.006214721128344536, + 0.005102898459881544, + 0.009373744018375874, + 0.004246485885232687, + 0.0037418710999190807, + 0.00384728261269629, + 0.002723204903304577, + 0.0038667782209813595, + 0.06446163356304169, + 0.024832366034388542, + 0.19206085801124573, + 0.024236803874373436, + 0.015018727630376816, + 0.02175574190914631, + 0.01829288899898529, + 0.09876887500286102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.007631433662027121 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463933777064085 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.00463933777064085, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463933777064085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14537861943244934, + 0.09094823151826859, + 0.06054781377315521, + 0.05470574274659157, + 0.04576513171195984, + 0.02596246637403965, + 0.024260353296995163, + 0.057201556861400604, + 0.05033864080905914, + 0.44489148259162903 + ], + "credit_map": [ + 0.02473732829093933, + 0.01966513693332672, + 0.01827394776046276, + 0.013947593979537487, + 0.013539215549826622, + 0.017136408016085625, + 0.012408862821757793, + 0.025670133531093597, + 0.009862649254500866, + 0.013344592414796352, + 0.016339968889951706, + 0.011809336952865124, + 0.009863084182143211, + 0.011598574928939342, + 0.008685208857059479, + 0.00944481696933508, + 0.008322406560182571, + 0.010987591929733753, + 0.005494649522006512, + 0.007032826542854309, + 0.010430139489471912, + 0.010166201740503311, + 0.0036421550903469324, + 0.004471843596547842, + 0.005876755341887474, + 0.005800564307719469, + 0.005943589378148317, + 0.005646287929266691, + 0.011147969402372837, + 0.009385526180267334, + 0.005799098871648312, + 0.005105953197926283, + 0.006017184350639582, + 0.004319025203585625, + 0.0054051815532147884, + 0.006270976271480322, + 0.005207687150686979, + 0.006436911877244711, + 0.004358186852186918, + 0.007749979849904776, + 0.0024600683245807886, + 0.0033826064318418503, + 0.004265701863914728, + 0.0023322736378759146, + 0.0039784652180969715, + 0.002866692841053009, + 0.004298341926187277, + 0.0023783151991665363, + 0.0029438030906021595, + 0.0037311185151338577, + 0.003598814597353339, + 0.004395616706460714, + 0.0027001304551959038, + 0.0019506218377500772, + 0.0027438136748969555, + 0.0021964339539408684, + 0.006078027654439211, + 0.007037821691483259, + 0.009232172742486, + 0.011233537457883358, + 0.003733946941792965, + 0.005272828508168459, + 0.008122660219669342, + 0.0064905607141554356, + 0.005214362870901823, + 0.009514623321592808, + 0.009371732361614704, + 0.006504267454147339, + 0.0038898552302271128, + 0.004184870980679989, + 0.004831140395253897, + 0.0068277898244559765, + 0.06974568963050842, + 0.03269575536251068, + 0.17652373015880585, + 0.024148322641849518, + 0.018463091924786568, + 0.023224078118801117, + 0.018482103943824768, + 0.08160868287086487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.00463933777064085 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635913021862507 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.008635913021862507, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635913021862507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2196534276008606, + 0.055296096950769424, + 0.04767448082566261, + 0.04359817132353783, + 0.131690114736557, + 0.02839081548154354, + 0.017548205330967903, + 0.04251722991466522, + 0.21646465361118317, + 0.19716677069664001 + ], + "credit_map": [ + 0.025723109021782875, + 0.0573912151157856, + 0.020205220207571983, + 0.027682654559612274, + 0.026808757334947586, + 0.021269502118229866, + 0.01672091893851757, + 0.02385205216705799, + 0.006392196752130985, + 0.011252643540501595, + 0.007115685380995274, + 0.007360726594924927, + 0.008015246130526066, + 0.005893065594136715, + 0.004237092100083828, + 0.005029442720115185, + 0.007137803360819817, + 0.006939671467989683, + 0.005287587642669678, + 0.0059057679027318954, + 0.0070428866893053055, + 0.008104360662400723, + 0.003297441406175494, + 0.003958962392061949, + 0.005282572936266661, + 0.006012879312038422, + 0.005218386184424162, + 0.00469087203964591, + 0.0059738764539361, + 0.0057178642600774765, + 0.006246855948120356, + 0.004454862792044878, + 0.0176861472427845, + 0.026117615401744843, + 0.011323174461722374, + 0.02853294275701046, + 0.014277291484177113, + 0.008498800918459892, + 0.0101665323600173, + 0.01508761290460825, + 0.003636860754340887, + 0.006306121125817299, + 0.0025019224267452955, + 0.0029888912104070187, + 0.0033262595534324646, + 0.0038412956055253744, + 0.0028655827045440674, + 0.0029238821007311344, + 0.0015514951664954424, + 0.002431802684441209, + 0.002605390502139926, + 0.004060938488692045, + 0.0016066214302554727, + 0.001856744522228837, + 0.0016521827783435583, + 0.001783031621016562, + 0.005911828950047493, + 0.004717124160379171, + 0.008168760687112808, + 0.005254920572042465, + 0.004784273449331522, + 0.0042259772308170795, + 0.00275425030849874, + 0.006700093392282724, + 0.03504267334938049, + 0.04044633358716965, + 0.02322320267558098, + 0.030283233150839806, + 0.01783297210931778, + 0.020503759384155273, + 0.01755383051931858, + 0.031578630208969116, + 0.0430508516728878, + 0.04126788303256035, + 0.03925798088312149, + 0.012847743928432465, + 0.015172768384218216, + 0.014372536912560463, + 0.012993843294680119, + 0.01820315793156624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.008635913021862507 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, the camera held still, focused on a man in a flowing white silk robe. His long, dark hair was meticulously tied into a high bun, secured with an antique wooden hairpin. His profile was sharply defined, his eyes slightly downcast, his expression focused and serene. He sat at a rustic wooden table, upon which lay a dark brown guqin, its lacquer surface delicately cracked by age. His long, bony fingers delicately stroked the taut silk strings, his fingertips gently gliding across them. The movement was steady and deliberate, more a reverent exchange with the instrument than a performance. A soft beam of afternoon sunlight slanted through the blurred carved wooden window in the background, skewing the air with floating dust particles and casting mottled shadows on the lacquered surface of the guqin. The main colors of the entire picture are quiet wood color and pure white, illuminated by warm sunlight, creating an atmosphere of extreme tranquility, full of Zen and classical elegance.", + "seed": 6483, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_1/index_383_seed_6483_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020260970667004587, + "vlm_logprob": -0.020260970667004587, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.020260970667004587, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010846114717423916 + ], + "opposite_logprob_tokens": [ + -6.135846138000488 + ], + "logprob_reward": -0.010846114717423916, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.010846114717423916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26363304257392883, + 0.08970139175653458, + 0.052507005631923676, + 0.107132688164711, + 0.12112590670585632, + 0.12524646520614624, + 0.06653794646263123, + 0.06520435959100723, + 0.047538816928863525, + 0.0613723061978817 + ], + "credit_map": [ + 0.02676098421216011, + 0.04345238208770752, + 0.020801540464162827, + 0.029321214184165, + 0.03872797265648842, + 0.028483590111136436, + 0.028176112100481987, + 0.047909270972013474, + 0.007857054471969604, + 0.013912777416408062, + 0.00996676366776228, + 0.008553951047360897, + 0.016066189855337143, + 0.008012603968381882, + 0.01069293636828661, + 0.014639124274253845, + 0.006805947981774807, + 0.007091508712619543, + 0.0043213763274252415, + 0.006448410917073488, + 0.011733680963516235, + 0.005031166598200798, + 0.004687803331762552, + 0.0063871159218251705, + 0.008475936017930508, + 0.008362479507923126, + 0.014151595532894135, + 0.014301386661827564, + 0.01216575875878334, + 0.013075705617666245, + 0.011508934199810028, + 0.0250909011811018, + 0.00960585381835699, + 0.00895718578249216, + 0.009739676490426064, + 0.01580292172729969, + 0.016756732016801834, + 0.01744440197944641, + 0.015397980809211731, + 0.02742116153240204, + 0.010436373762786388, + 0.013478811830282211, + 0.013488298282027245, + 0.01140354759991169, + 0.01948755607008934, + 0.01752885803580284, + 0.016567125916481018, + 0.02285589836537838, + 0.007193387486040592, + 0.008385946974158287, + 0.0045133838430047035, + 0.006596704944968224, + 0.009200098924338818, + 0.007139492314308882, + 0.008808530867099762, + 0.014700409024953842, + 0.006254768464714289, + 0.008142664097249508, + 0.00740350317209959, + 0.00844606850296259, + 0.008215895853936672, + 0.006834062281996012, + 0.008107952773571014, + 0.011799453757703304, + 0.004198197275400162, + 0.004307394381612539, + 0.006511122919619083, + 0.008129431866109371, + 0.005244427360594273, + 0.005155164748430252, + 0.00616906862705946, + 0.007824013940989971, + 0.00550798187032342, + 0.008444949984550476, + 0.006758319679647684, + 0.009657683782279491, + 0.0064448025077581406, + 0.0044686817564070225, + 0.007981177419424057, + 0.012108714319765568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.010846114717423916 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868816211819649 + ], + "opposite_logprob_tokens": [ + -3.0486881732940674 + ], + "logprob_reward": -0.04868816211819649, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868816211819649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34056198596954346, + 0.0806020125746727, + 0.12117946147918701, + 0.06303086876869202, + 0.04041272774338722, + 0.12103600054979324, + 0.045006342232227325, + 0.036239124834537506, + 0.06062353402376175, + 0.09130793064832687 + ], + "credit_map": [ + 0.027771668508648872, + 0.06143796443939209, + 0.023280909284949303, + 0.03316935524344444, + 0.0450550876557827, + 0.040177807211875916, + 0.03716651722788811, + 0.07250268757343292, + 0.007763364817947149, + 0.01419828087091446, + 0.006803241092711687, + 0.008814569562673569, + 0.01418984867632389, + 0.007836691103875637, + 0.008815724402666092, + 0.012180289253592491, + 0.011497985571622849, + 0.011733810417354107, + 0.011151536367833614, + 0.015189556404948235, + 0.019305409863591194, + 0.013655883260071278, + 0.013028956018388271, + 0.025616319850087166, + 0.004856426734477282, + 0.005384041462093592, + 0.006610274314880371, + 0.01224818266928196, + 0.006925151217728853, + 0.0073089394718408585, + 0.006844471208751202, + 0.012853388674557209, + 0.003310436848551035, + 0.004150155000388622, + 0.004248702432960272, + 0.0051929824985563755, + 0.0051286653615534306, + 0.005988089833408594, + 0.004696388263255358, + 0.007697307504713535, + 0.010440909303724766, + 0.012271082028746605, + 0.010100185871124268, + 0.010306375101208687, + 0.019250117242336273, + 0.019842730835080147, + 0.014862731099128723, + 0.02396186999976635, + 0.005152360536158085, + 0.0052093337289988995, + 0.0033853144850581884, + 0.00661462964490056, + 0.004433377180248499, + 0.006102884653955698, + 0.0056836530566215515, + 0.008424787782132626, + 0.003914128988981247, + 0.006008506286889315, + 0.0031802579760551453, + 0.003856615861877799, + 0.006423307582736015, + 0.0036857726518064737, + 0.004188398830592632, + 0.004982133861631155, + 0.0037835901603102684, + 0.004886076785624027, + 0.005023522302508354, + 0.008429932408034801, + 0.010921237990260124, + 0.005290784407407045, + 0.007571734022349119, + 0.014716656878590584, + 0.006345625501126051, + 0.010882033035159111, + 0.008823636919260025, + 0.007675739470869303, + 0.011950254440307617, + 0.015282762236893177, + 0.009695635177195072, + 0.020652243867516518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.04868816211819649 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543593533337116 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543593533337116, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543593533337116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25044235587120056, + 0.08533331006765366, + 0.04259692505002022, + 0.2098025530576706, + 0.1401377171278, + 0.08921165764331818, + 0.044814880937337875, + 0.04416632652282715, + 0.059626951813697815, + 0.03386736661195755 + ], + "credit_map": [ + 0.02643844485282898, + 0.04994262009859085, + 0.01579599268734455, + 0.03398866206407547, + 0.047934845089912415, + 0.02357659861445427, + 0.02655983716249466, + 0.026205340400338173, + 0.009521051310002804, + 0.012816790491342545, + 0.00822104699909687, + 0.01339761447161436, + 0.006411612033843994, + 0.008431900292634964, + 0.009580853395164013, + 0.01695243827998638, + 0.0050705852918326855, + 0.006683518178761005, + 0.004382445011287928, + 0.006458987481892109, + 0.005879122763872147, + 0.0038523494731634855, + 0.004381026141345501, + 0.0058888886123895645, + 0.01793191023170948, + 0.01721682958304882, + 0.02389216050505638, + 0.028164418414235115, + 0.024266019463539124, + 0.023642966523766518, + 0.023406263440847397, + 0.051281969994306564, + 0.010967991314828396, + 0.016656294465065002, + 0.01259545050561428, + 0.015050185844302177, + 0.016590014100074768, + 0.019623443484306335, + 0.015399287454783916, + 0.033255044370889664, + 0.007445181254297495, + 0.008723653852939606, + 0.006645513232797384, + 0.009493224322795868, + 0.012925048358738422, + 0.017735103145241737, + 0.00964729767292738, + 0.016596641391515732, + 0.0042284163646399975, + 0.004760806914418936, + 0.0037701313849538565, + 0.0067619276233017445, + 0.00618346594274044, + 0.005358013790100813, + 0.005224708467721939, + 0.008527410216629505, + 0.00573701923713088, + 0.007062337826937437, + 0.005035509821027517, + 0.005847764201462269, + 0.004541031084954739, + 0.004776185378432274, + 0.004661315120756626, + 0.006505161989480257, + 0.0031027738004922867, + 0.0046978541649878025, + 0.005110292229801416, + 0.008809154853224754, + 0.009474387392401695, + 0.007029129192233086, + 0.00724996579810977, + 0.01415339671075344, + 0.0024076446425169706, + 0.0038913965690881014, + 0.002882415195927024, + 0.004868981894105673, + 0.0052735949866473675, + 0.0030748324934393167, + 0.004001772962510586, + 0.007466726005077362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.012543593533337116 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059071511030197 + ], + "opposite_logprob_tokens": [ + -3.8955907821655273 + ], + "logprob_reward": -0.02059071511030197, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059071511030197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19129084050655365, + 0.09445568174123764, + 0.047394853085279465, + 0.11332965642213821, + 0.23026156425476074, + 0.17629262804985046, + 0.030962124466896057, + 0.051772523671388626, + 0.03163495659828186, + 0.03260518237948418 + ], + "credit_map": [ + 0.01802881807088852, + 0.02339257299900055, + 0.01789107732474804, + 0.023086389526724815, + 0.034098803997039795, + 0.022043250501155853, + 0.021689390763640404, + 0.031060507521033287, + 0.010012989863753319, + 0.01130025926977396, + 0.007858514785766602, + 0.009781156666576862, + 0.019566478207707405, + 0.010649102739989758, + 0.010077090002596378, + 0.01521008089184761, + 0.005808011628687382, + 0.007439989130944014, + 0.0038700862787663937, + 0.004431817680597305, + 0.009098383598029613, + 0.005298139061778784, + 0.0038820866029709578, + 0.007566332817077637, + 0.007496653590351343, + 0.011624129489064217, + 0.008769496344029903, + 0.014982931315898895, + 0.011993195861577988, + 0.020353645086288452, + 0.012546570971608162, + 0.025563033297657967, + 0.021664639934897423, + 0.023696668446063995, + 0.02053169719874859, + 0.023527070879936218, + 0.02999241277575493, + 0.0335005447268486, + 0.028320180252194405, + 0.049028340727090836, + 0.015990935266017914, + 0.021176233887672424, + 0.019519135355949402, + 0.01761217601597309, + 0.0274965800344944, + 0.021008342504501343, + 0.023842647671699524, + 0.029646560549736023, + 0.00408787839114666, + 0.0030995351262390614, + 0.0032838855404406786, + 0.0029237864073365927, + 0.0030019336845725775, + 0.004036101512610912, + 0.005340521223843098, + 0.005188480485230684, + 0.0031621363013982773, + 0.005052391905337572, + 0.003928443882614374, + 0.006696867290884256, + 0.005917031317949295, + 0.008212030865252018, + 0.006160362157970667, + 0.012643259018659592, + 0.002626363653689623, + 0.0030776241328567266, + 0.0035366590600460768, + 0.005476105958223343, + 0.006029265467077494, + 0.0021982609760016203, + 0.002996063092723489, + 0.005694609135389328, + 0.002273204969242215, + 0.004102195147424936, + 0.003597216447815299, + 0.004494272172451019, + 0.0040149264968931675, + 0.003236248856410384, + 0.0038422290235757828, + 0.007044883444905281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.02059071511030197 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636267855763435 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636267855763435, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636267855763435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27567628026008606, + 0.11497388780117035, + 0.06419093906879425, + 0.08003170788288116, + 0.07152381539344788, + 0.10430201143026352, + 0.08581126481294632, + 0.10049314051866531, + 0.05787604674696922, + 0.04512093588709831 + ], + "credit_map": [ + 0.028791377320885658, + 0.0599367618560791, + 0.01982211321592331, + 0.049819376319646835, + 0.032928217202425, + 0.026930445805191994, + 0.02699907496571541, + 0.030448904260993004, + 0.014843294396996498, + 0.025978881865739822, + 0.009650953114032745, + 0.01681399531662464, + 0.012341247871518135, + 0.010183075442910194, + 0.011949396692216396, + 0.013213052414357662, + 0.006775491870939732, + 0.011638786643743515, + 0.006387706380337477, + 0.010857680812478065, + 0.008815561421215534, + 0.005230667535215616, + 0.005934575106948614, + 0.008550466038286686, + 0.0059852879494428635, + 0.009601141326129436, + 0.011372232809662819, + 0.016241302713751793, + 0.006353250704705715, + 0.00810743123292923, + 0.009401421993970871, + 0.012969646602869034, + 0.0057877544313669205, + 0.00719196442514658, + 0.009176052175462246, + 0.008887717500329018, + 0.008151659741997719, + 0.011060593649744987, + 0.008567416109144688, + 0.012700655497610569, + 0.01123439148068428, + 0.01379337813705206, + 0.014904164709150791, + 0.015811992809176445, + 0.012433178722858429, + 0.011625505983829498, + 0.010851339437067509, + 0.01364805642515421, + 0.011083277873694897, + 0.011929561384022236, + 0.013926234096288681, + 0.007689940743148327, + 0.007827808149158955, + 0.008556471206247807, + 0.012836020439863205, + 0.01196195650845766, + 0.008107233792543411, + 0.013833137229084969, + 0.008930755779147148, + 0.01319782342761755, + 0.010962910950183868, + 0.011726228520274162, + 0.011647356674075127, + 0.022087691351771355, + 0.004188471473753452, + 0.006435200572013855, + 0.0055290586315095425, + 0.009916923940181732, + 0.009747746400535107, + 0.005481541622430086, + 0.006349510513246059, + 0.010227590799331665, + 0.004801788367331028, + 0.00721578998491168, + 0.00553863076493144, + 0.005683972500264645, + 0.0044351983815431595, + 0.003922705538570881, + 0.005638959817588329, + 0.007883888669312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.008636267855763435 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, the camera held still, focused on a man in a flowing white silk robe. His long, dark hair was meticulously tied into a high bun, secured with an antique wooden hairpin. His profile was sharply defined, his eyes slightly downcast, his expression focused and serene. He sat at a rustic wooden table, upon which lay a dark brown guqin, its lacquer surface delicately cracked by age. His long, bony fingers delicately stroked the taut silk strings, his fingertips gently gliding across them. The movement was steady and deliberate, more a reverent exchange with the instrument than a performance. A soft beam of afternoon sunlight slanted through the blurred carved wooden window in the background, skewing the air with floating dust particles and casting mottled shadows on the lacquered surface of the guqin. The main colors of the entire picture are quiet wood color and pure white, illuminated by warm sunlight, creating an atmosphere of extreme tranquility, full of Zen and classical elegance.", + "seed": 6484, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_1/index_383_seed_6484_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.017263595946133137, + "vlm_logprob": -0.017263595946133137, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.017263595946133137, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011260410770773888 + ], + "opposite_logprob_tokens": [ + -5.636260509490967 + ], + "logprob_reward": -0.011260410770773888, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.011260410770773888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33182936906814575, + 0.0730820968747139, + 0.09348393231630325, + 0.08895149827003479, + 0.057515550404787064, + 0.07681822031736374, + 0.04870784282684326, + 0.07458200305700302, + 0.09405116736888885, + 0.06097831204533577 + ], + "credit_map": [ + 0.07931870222091675, + 0.027853695675730705, + 0.04222820699214935, + 0.02998674288392067, + 0.03550325334072113, + 0.03854987770318985, + 0.044207315891981125, + 0.034181613475084305, + 0.00819548312574625, + 0.006935505196452141, + 0.008554138243198395, + 0.01130112074315548, + 0.008337186649441719, + 0.010611786507070065, + 0.010490641929209232, + 0.008656234480440617, + 0.012539955787360668, + 0.010712175630033016, + 0.012428783811628819, + 0.01210357528179884, + 0.011750321835279465, + 0.009462784044444561, + 0.016242241486907005, + 0.00824410654604435, + 0.010586407966911793, + 0.015339069068431854, + 0.008701447397470474, + 0.017514007166028023, + 0.0069038718938827515, + 0.006745772901922464, + 0.011671134270727634, + 0.01148979365825653, + 0.009331680834293365, + 0.007288143504410982, + 0.006501690484583378, + 0.008510016836225986, + 0.003909542225301266, + 0.008536476641893387, + 0.005947401747107506, + 0.007490606978535652, + 0.014857674948871136, + 0.009277507662773132, + 0.007274752948433161, + 0.009377573616802692, + 0.006408262997865677, + 0.009387103840708733, + 0.0118112713098526, + 0.008424080908298492, + 0.0039294385351240635, + 0.005979351699352264, + 0.004941139370203018, + 0.00471210153773427, + 0.006927725858986378, + 0.009691385552287102, + 0.0065793064422905445, + 0.005947394762188196, + 0.005142137408256531, + 0.005814365576952696, + 0.009041217155754566, + 0.010622977279126644, + 0.007306463085114956, + 0.009092587977647781, + 0.013532514683902264, + 0.014029745943844318, + 0.0052420771680772305, + 0.012773375026881695, + 0.008927897550165653, + 0.01319004688411951, + 0.01807020977139473, + 0.019350362941622734, + 0.008434812538325787, + 0.008062386885285378, + 0.006318218540400267, + 0.007877958938479424, + 0.006627232767641544, + 0.0075915101915597916, + 0.006201483774930239, + 0.00910686980932951, + 0.009665843099355698, + 0.0075892009772360325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.011260410770773888 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370694860816002 + ], + "opposite_logprob_tokens": [ + -3.6513707637786865 + ], + "logprob_reward": -0.026370694860816002, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370694860816002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3968813419342041, + 0.09065835177898407, + 0.11957582086324692, + 0.07362685352563858, + 0.0495770238339901, + 0.029381530359387398, + 0.08979730308055878, + 0.04879284277558327, + 0.045484792441129684, + 0.05622426047921181 + ], + "credit_map": [ + 0.03419605270028114, + 0.08454018086194992, + 0.04468927159905434, + 0.08243461698293686, + 0.028699466958642006, + 0.02611229382455349, + 0.04978547617793083, + 0.04642396420240402, + 0.01615810953080654, + 0.011300420388579369, + 0.008284792304039001, + 0.012561192736029625, + 0.0052581368945539, + 0.008413988165557384, + 0.014472980983555317, + 0.014208728447556496, + 0.028782643377780914, + 0.011853442527353764, + 0.012912624515593052, + 0.0124086644500494, + 0.014018340036273003, + 0.010354108177125454, + 0.014338650740683079, + 0.0149073526263237, + 0.00728769414126873, + 0.014264024794101715, + 0.0067513491958379745, + 0.01511194184422493, + 0.00461988802999258, + 0.005492806900292635, + 0.010885026305913925, + 0.009214124642312527, + 0.009732336737215519, + 0.0057488395832479, + 0.004877073690295219, + 0.006085192319005728, + 0.004496581386774778, + 0.00349327246658504, + 0.006624646484851837, + 0.008519078604876995, + 0.00642832275480032, + 0.0028224752750247717, + 0.0030856109224259853, + 0.003112170612439513, + 0.0020968755707144737, + 0.0037170934956520796, + 0.004114390350878239, + 0.004004590213298798, + 0.007575174793601036, + 0.006678140256553888, + 0.00948933232575655, + 0.007032486144453287, + 0.01895630732178688, + 0.019723467528820038, + 0.00905641820281744, + 0.011285977438092232, + 0.004750627558678389, + 0.003351430408656597, + 0.0035220040008425713, + 0.005270278546959162, + 0.009853100404143333, + 0.010362057946622372, + 0.007246975786983967, + 0.004436363931745291, + 0.0032008520793169737, + 0.004207951948046684, + 0.003309684805572033, + 0.004496408626437187, + 0.00896348524838686, + 0.009497634135186672, + 0.006505638360977173, + 0.005303139798343182, + 0.004252088256180286, + 0.005667877849191427, + 0.0043079303577542305, + 0.005464432295411825, + 0.012876104563474655, + 0.012582640163600445, + 0.006510414648801088, + 0.0045627751387655735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.026370694860816002 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978530291467905 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.00978530291467905, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978530291467905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28668686747550964, + 0.1268312782049179, + 0.06121648848056793, + 0.17080438137054443, + 0.05663434788584709, + 0.05362194404006004, + 0.05406563729047775, + 0.060050178319215775, + 0.06937205791473389, + 0.06071680411696434 + ], + "credit_map": [ + 0.04484367370605469, + 0.036845311522483826, + 0.03840017318725586, + 0.04056952893733978, + 0.032369066029787064, + 0.029310602694749832, + 0.03352748602628708, + 0.030821040272712708, + 0.009549352340400219, + 0.015660427510738373, + 0.013667145743966103, + 0.028586184605956078, + 0.010342489928007126, + 0.011133696883916855, + 0.017299531027674675, + 0.020592445507645607, + 0.010609489865601063, + 0.008228021673858166, + 0.007031846325844526, + 0.010178295895457268, + 0.006431742571294308, + 0.005107392091304064, + 0.008442205376923084, + 0.005187493283301592, + 0.013994654640555382, + 0.033698488026857376, + 0.017418930307030678, + 0.04035383462905884, + 0.010651743039488792, + 0.013460862450301647, + 0.024484526365995407, + 0.016741354018449783, + 0.004301304928958416, + 0.007060589734464884, + 0.006252794526517391, + 0.01045307982712984, + 0.006242811214178801, + 0.007413799874484539, + 0.00716651463881135, + 0.007743454072624445, + 0.010477665811777115, + 0.006395943928509951, + 0.004827857483178377, + 0.006688126362860203, + 0.0036835945211350918, + 0.005453766789287329, + 0.007256294135004282, + 0.008838689886033535, + 0.0066195703111588955, + 0.005757300648838282, + 0.006011545658111572, + 0.0052739945240318775, + 0.008089126087725163, + 0.007266109809279442, + 0.006084011401981115, + 0.00896397978067398, + 0.004332216456532478, + 0.005903732497245073, + 0.00530244130641222, + 0.005924195516854525, + 0.012125708162784576, + 0.013745113275945187, + 0.006395322736352682, + 0.0063214474357664585, + 0.005204895976930857, + 0.011906181462109089, + 0.005949343089014292, + 0.009959355928003788, + 0.009389705024659634, + 0.010397182777523994, + 0.008269035257399082, + 0.008296363055706024, + 0.010278115049004555, + 0.005946311634033918, + 0.006747571751475334, + 0.010543035343289375, + 0.004462201613932848, + 0.006185729056596756, + 0.008592939004302025, + 0.007960902526974678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.00978530291467905 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026362450793385506 + ], + "opposite_logprob_tokens": [ + -3.651362419128418 + ], + "logprob_reward": -0.026362450793385506, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026362450793385506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35797131061553955, + 0.07808592170476913, + 0.05713207647204399, + 0.06209343671798706, + 0.06468845158815384, + 0.05170753598213196, + 0.1822768598794937, + 0.044864945113658905, + 0.049501191824674606, + 0.05167829245328903 + ], + "credit_map": [ + 0.044017624109983444, + 0.07156403362751007, + 0.0426170788705349, + 0.07306867837905884, + 0.026528047397732735, + 0.02622097171843052, + 0.040110789239406586, + 0.033844128251075745, + 0.00804396253079176, + 0.011016469448804855, + 0.009544400498270988, + 0.01347908191382885, + 0.006588196847587824, + 0.005942278541624546, + 0.012872741557657719, + 0.010598794557154179, + 0.013409744948148727, + 0.00642827432602644, + 0.007165214512497187, + 0.007668558973819017, + 0.0037602619268000126, + 0.0065346574410796165, + 0.007616066839545965, + 0.004549297038465738, + 0.005034552421420813, + 0.011098434217274189, + 0.007067346479743719, + 0.011313404887914658, + 0.004922010470181704, + 0.0067806304432451725, + 0.00869123823940754, + 0.007185819558799267, + 0.00544554740190506, + 0.008450600318610668, + 0.006507413927465677, + 0.00820003729313612, + 0.005824750289320946, + 0.007445035967975855, + 0.011769160628318787, + 0.011045902967453003, + 0.009370094165205956, + 0.0038965672720223665, + 0.005982432514429092, + 0.006600468885153532, + 0.0043540955521166325, + 0.006319081410765648, + 0.008416453376412392, + 0.006768343970179558, + 0.014200217090547085, + 0.015401676297187805, + 0.019260898232460022, + 0.016132861375808716, + 0.034383077174425125, + 0.03573691472411156, + 0.02134898491203785, + 0.02581222914159298, + 0.005947629921138287, + 0.004819522611796856, + 0.005472038872539997, + 0.004697312600910664, + 0.0045196497812867165, + 0.004723793361335993, + 0.00855417549610138, + 0.006130822002887726, + 0.010108305141329765, + 0.007514807395637035, + 0.004851335659623146, + 0.006551200523972511, + 0.0028624010737985373, + 0.005157276056706905, + 0.00718011986464262, + 0.005275745410472155, + 0.004289587493985891, + 0.0046168905682861805, + 0.004752027802169323, + 0.006721555721014738, + 0.008601822890341282, + 0.009356037713587284, + 0.006276522297412157, + 0.007063844706863165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.026362450793385506 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539120391011238 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.012539120391011238, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539120391011238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3628993332386017, + 0.0857652947306633, + 0.08189521729946136, + 0.09499756246805191, + 0.038326334208250046, + 0.03543494641780853, + 0.05370029807090759, + 0.10442031174898148, + 0.0944715216755867, + 0.048089202493429184 + ], + "credit_map": [ + 0.05328699201345444, + 0.05096397548913956, + 0.04454156756401062, + 0.052658066153526306, + 0.030955947935581207, + 0.030531590804457664, + 0.05304495245218277, + 0.04691624641418457, + 0.014261918142437935, + 0.008491628803312778, + 0.009012904949486256, + 0.010642175562679768, + 0.005477194208651781, + 0.009229795075953007, + 0.017700480297207832, + 0.010949197225272655, + 0.008518890477716923, + 0.010187558829784393, + 0.009213161654770374, + 0.012200871482491493, + 0.005950667429715395, + 0.009077205322682858, + 0.017645230516791344, + 0.009101632982492447, + 0.008829055353999138, + 0.009327251464128494, + 0.009951137006282806, + 0.015298536978662014, + 0.007856382057070732, + 0.007721635513007641, + 0.018763281404972076, + 0.017250284552574158, + 0.004023813642561436, + 0.004110442940145731, + 0.0040043918415904045, + 0.004555655177682638, + 0.0036895282100886106, + 0.006856183987110853, + 0.007051818538457155, + 0.004034499172121286, + 0.0049886275082826614, + 0.0037715353537350893, + 0.0032950297463685274, + 0.0051554422825574875, + 0.0027469568885862827, + 0.004687691107392311, + 0.005071428604424, + 0.0057182349264621735, + 0.0061244601383805275, + 0.005348592065274715, + 0.005226271692663431, + 0.005368356592953205, + 0.006728560198098421, + 0.009303642436861992, + 0.00950879231095314, + 0.00609162449836731, + 0.009294221177697182, + 0.00639728270471096, + 0.010893004946410656, + 0.008236924186348915, + 0.01915811561048031, + 0.022478191182017326, + 0.015156617388129234, + 0.012805958278477192, + 0.005730230826884508, + 0.007178506813943386, + 0.009218883700668812, + 0.00899655930697918, + 0.018464703112840652, + 0.021048683673143387, + 0.013783320784568787, + 0.010050642304122448, + 0.00748099759221077, + 0.004449860658496618, + 0.004810109734535217, + 0.008826183155179024, + 0.0038294652476906776, + 0.006540846545249224, + 0.006174364127218723, + 0.005977376829832792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.012539120391011238 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, the camera held still, focused on a man in a flowing white silk robe. His long, dark hair was meticulously tied into a high bun, secured with an antique wooden hairpin. His profile was sharply defined, his eyes slightly downcast, his expression focused and serene. He sat at a rustic wooden table, upon which lay a dark brown guqin, its lacquer surface delicately cracked by age. His long, bony fingers delicately stroked the taut silk strings, his fingertips gently gliding across them. The movement was steady and deliberate, more a reverent exchange with the instrument than a performance. A soft beam of afternoon sunlight slanted through the blurred carved wooden window in the background, skewing the air with floating dust particles and casting mottled shadows on the lacquered surface of the guqin. The main colors of the entire picture are quiet wood color and pure white, illuminated by warm sunlight, creating an atmosphere of extreme tranquility, full of Zen and classical elegance.", + "seed": 6485, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_1/index_383_seed_6485_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.015861373860388994, + "vlm_logprob": -0.015861373860388994, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.015861373860388994, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008693940006196499 + ], + "opposite_logprob_tokens": [ + -6.258694171905518 + ], + "logprob_reward": -0.008693940006196499, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.008693940006196499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23920674622058868, + 0.09854190051555634, + 0.07629216462373734, + 0.06888282299041748, + 0.10466541349887848, + 0.11602962762117386, + 0.1069115698337555, + 0.04311061277985573, + 0.0699923038482666, + 0.07636677473783493 + ], + "credit_map": [ + 0.022126900032162666, + 0.01564399152994156, + 0.03040436841547489, + 0.04585903137922287, + 0.0281699039041996, + 0.017463013529777527, + 0.04708535596728325, + 0.032454200088977814, + 0.007832812145352364, + 0.006577920168638229, + 0.012169533409178257, + 0.022214733064174652, + 0.015786336734890938, + 0.005995516665279865, + 0.01514887809753418, + 0.012816183269023895, + 0.0066301217302680016, + 0.00830103736370802, + 0.008823463693261147, + 0.015081431716680527, + 0.011405715718865395, + 0.004507184494286776, + 0.011745861731469631, + 0.009797351434826851, + 0.008207461796700954, + 0.00681985542178154, + 0.008081791922450066, + 0.010959791019558907, + 0.011399812996387482, + 0.007023858372122049, + 0.009134514257311821, + 0.007255744654685259, + 0.012964082881808281, + 0.013068674132227898, + 0.01178639754652977, + 0.01795356720685959, + 0.014147822745144367, + 0.008421272970736027, + 0.014450088143348694, + 0.011873521842062473, + 0.020916620269417763, + 0.014490912668406963, + 0.01027653831988573, + 0.016219573095440865, + 0.012651186436414719, + 0.009990215301513672, + 0.01701486110687256, + 0.014469729736447334, + 0.022927621379494667, + 0.011474627070128918, + 0.008659787476062775, + 0.016412705183029175, + 0.011446964927017689, + 0.0073532299138605595, + 0.015932245180010796, + 0.012704400345683098, + 0.005302237346768379, + 0.004057830199599266, + 0.003474474884569645, + 0.006079178303480148, + 0.007449178025126457, + 0.003249941859394312, + 0.007833226583898067, + 0.005664548370987177, + 0.006732181645929813, + 0.006630653515458107, + 0.007669568527489901, + 0.014746366068720818, + 0.009903554804623127, + 0.006207452155649662, + 0.009220273233950138, + 0.008882258087396622, + 0.01194609422236681, + 0.007438724394887686, + 0.007027721498161554, + 0.01051683072000742, + 0.010791604407131672, + 0.005517525598406792, + 0.012666040100157261, + 0.010462244041264057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.008693940006196499 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371884673833847 + ], + "opposite_logprob_tokens": [ + -3.4087188243865967 + ], + "logprob_reward": -0.03371884673833847, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371884673833847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25441282987594604, + 0.08847086876630783, + 0.09191868454217911, + 0.09796101599931717, + 0.04964928701519966, + 0.17392806708812714, + 0.09778207540512085, + 0.04539605602622032, + 0.04771784693002701, + 0.05276326462626457 + ], + "credit_map": [ + 0.035566046833992004, + 0.01970260962843895, + 0.02655291184782982, + 0.038270868360996246, + 0.04198624938726425, + 0.020112546160817146, + 0.0378868393599987, + 0.03433474898338318, + 0.009969831444323063, + 0.005558226723223925, + 0.013700803741812706, + 0.014104647561907768, + 0.013338177464902401, + 0.0074739097617566586, + 0.01325302105396986, + 0.011072253808379173, + 0.012300806120038033, + 0.007141441106796265, + 0.009540000930428505, + 0.010377052240073681, + 0.01499051135033369, + 0.00953199714422226, + 0.016530655324459076, + 0.011506225913763046, + 0.012443062849342823, + 0.010548953898251057, + 0.012254069559276104, + 0.013424184173345566, + 0.015478532761335373, + 0.010171067900955677, + 0.012668820098042488, + 0.010972314514219761, + 0.005551128648221493, + 0.005799829959869385, + 0.005570708308368921, + 0.0059073022566735744, + 0.0077305450104177, + 0.00416555954143405, + 0.008711409755051136, + 0.006212805397808552, + 0.028498390689492226, + 0.02128661796450615, + 0.01764313504099846, + 0.028358004987239838, + 0.022659974172711372, + 0.011331249959766865, + 0.022826556116342545, + 0.021324140951037407, + 0.019621724262833595, + 0.009527509100735188, + 0.00896438304334879, + 0.013129804283380508, + 0.01082414947450161, + 0.006931225303560495, + 0.01667369157075882, + 0.01210958231240511, + 0.003550778841599822, + 0.004418604541569948, + 0.004421926103532314, + 0.00581163028255105, + 0.007372538559138775, + 0.00443920586258173, + 0.009289175271987915, + 0.0060921949334442616, + 0.006690721958875656, + 0.004900124855339527, + 0.004622348118573427, + 0.0068741063587367535, + 0.007152765057981014, + 0.004116814583539963, + 0.007716681342571974, + 0.0056442818604409695, + 0.004718136042356491, + 0.0035587106831371784, + 0.0050628529861569405, + 0.007197197526693344, + 0.008318635635077953, + 0.005446283146739006, + 0.010264272801578045, + 0.008197176270186901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.03371884673833847 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076509952545166 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076509952545166, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076509952545166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2644590139389038, + 0.07399099320173264, + 0.13943630456924438, + 0.11855046451091766, + 0.043004199862480164, + 0.09382937103509903, + 0.08878526091575623, + 0.08888520300388336, + 0.05414345860481262, + 0.03491567447781563 + ], + "credit_map": [ + 0.026125593110919, + 0.018771227449178696, + 0.04019753634929657, + 0.044466935098171234, + 0.045024361461400986, + 0.021136226132512093, + 0.03525419533252716, + 0.03348295018076897, + 0.008757879957556725, + 0.006014232989400625, + 0.006854034960269928, + 0.013745579868555069, + 0.01241244561970234, + 0.00613904045894742, + 0.011039153672754765, + 0.009028630331158638, + 0.0095972353592515, + 0.009843855164945126, + 0.013415833003818989, + 0.029687518253922462, + 0.024089619517326355, + 0.008873850107192993, + 0.024553531780838966, + 0.019374866038560867, + 0.010064441710710526, + 0.01893230527639389, + 0.013055717572569847, + 0.017949989065527916, + 0.020637890323996544, + 0.006484254263341427, + 0.016095135360956192, + 0.015330730937421322, + 0.004754630848765373, + 0.0042029330506920815, + 0.0051123108714818954, + 0.007853995077311993, + 0.006567262578755617, + 0.003292313776910305, + 0.006223480682820082, + 0.004997271578758955, + 0.00988884549587965, + 0.011571032926440239, + 0.01004497054964304, + 0.015864264219999313, + 0.013282723724842072, + 0.006456135306507349, + 0.014980312436819077, + 0.01174108311533928, + 0.010525773279368877, + 0.010882308706641197, + 0.007050985470414162, + 0.013285692781209946, + 0.014646733179688454, + 0.005996987223625183, + 0.01537003181874752, + 0.01102675311267376, + 0.006033218931406736, + 0.009756259620189667, + 0.006912638898938894, + 0.012346071191132069, + 0.015791818499565125, + 0.004375736229121685, + 0.020677907392382622, + 0.01299155130982399, + 0.005064551252871752, + 0.00493937823921442, + 0.006124209612607956, + 0.006303354632109404, + 0.009911472909152508, + 0.007977067492902279, + 0.00740364845842123, + 0.006419780198484659, + 0.004230612888932228, + 0.002518112538382411, + 0.004278053529560566, + 0.005236595403403044, + 0.00547643331810832, + 0.0036870059557259083, + 0.004770733881741762, + 0.004718126263469458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.011076509952545166 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631906773895025 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631906773895025, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631906773895025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23116041719913483, + 0.1219901368021965, + 0.09699274599552155, + 0.08308454602956772, + 0.06862395256757736, + 0.07617464661598206, + 0.13521656394004822, + 0.04934488236904144, + 0.08723615109920502, + 0.050175990909338 + ], + "credit_map": [ + 0.0218926053494215, + 0.016705168411135674, + 0.029776260256767273, + 0.0349767841398716, + 0.045774199068546295, + 0.016097325831651688, + 0.037622954696416855, + 0.028315091505646706, + 0.008953201584517956, + 0.007965752854943275, + 0.015538286417722702, + 0.020057959482073784, + 0.02385036274790764, + 0.007110363803803921, + 0.023054344579577446, + 0.015459851361811161, + 0.010661829262971878, + 0.011041747406125069, + 0.011049198918044567, + 0.013663148507475853, + 0.015309371054172516, + 0.007539638318121433, + 0.015488971024751663, + 0.012238830327987671, + 0.005772481672465801, + 0.008542908355593681, + 0.008182813413441181, + 0.010717418976128101, + 0.017300551757216454, + 0.006311861798167229, + 0.015268797986209393, + 0.010987707413733006, + 0.006558868568390608, + 0.0075367521494627, + 0.008054831065237522, + 0.009227504022419453, + 0.009145073592662811, + 0.007640200201421976, + 0.01180892251431942, + 0.008651796728372574, + 0.008472259156405926, + 0.008396499790251255, + 0.006712147034704685, + 0.011095443740487099, + 0.00995182991027832, + 0.007296996656805277, + 0.014066841453313828, + 0.010182618163526058, + 0.013623283244669437, + 0.016332201659679413, + 0.014103628695011139, + 0.017164671793580055, + 0.016876041889190674, + 0.009517071768641472, + 0.028838330879807472, + 0.018761323764920235, + 0.007022060453891754, + 0.0058252993039786816, + 0.0033159819431602955, + 0.0073607973754405975, + 0.006706809159368277, + 0.0021280930377542973, + 0.010277219116687775, + 0.006708613596856594, + 0.011007115244865417, + 0.007828909903764725, + 0.010197178460657597, + 0.014405583031475544, + 0.013083147816359997, + 0.009162074886262417, + 0.011518341489136219, + 0.010033794678747654, + 0.004446143750101328, + 0.003489420050755143, + 0.005896987393498421, + 0.00741574028506875, + 0.008229485712945461, + 0.00388687988743186, + 0.009350107982754707, + 0.007461221423000097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.007631906773895025 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man wearing a flowing white silk robe with long dark hair tied in a high bun secured by an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting at a rustic wooden table playing a dark brown guqin with visible age cracks on its lacquer surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's long bony fingers shown stroking the guqin strings with steady, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a soft beam of afternoon sunlight passing through a blurred carved wooden window, creating floating dust particles and mottled shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818566583096981 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.01818566583096981, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818566583096981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20921999216079712, + 0.16863223910331726, + 0.05001235008239746, + 0.12055183947086334, + 0.09819353371858597, + 0.11252772063016891, + 0.10922641307115555, + 0.05065594241023064, + 0.043209899216890335, + 0.03777006268501282 + ], + "credit_map": [ + 0.0245189405977726, + 0.02029801532626152, + 0.023151863366365433, + 0.03255479410290718, + 0.03424813598394394, + 0.017887523397803307, + 0.030692599713802338, + 0.0258681271225214, + 0.020212631672620773, + 0.011273298412561417, + 0.02279111184179783, + 0.025448374450206757, + 0.029454341158270836, + 0.01309871394187212, + 0.02530848979949951, + 0.02104528620839119, + 0.006994529627263546, + 0.004085324238985777, + 0.0062420060858130455, + 0.006491187494248152, + 0.008587360382080078, + 0.005360298790037632, + 0.006367426831275225, + 0.005884222220629454, + 0.019766880199313164, + 0.011492838151752949, + 0.01459004357457161, + 0.018933584913611412, + 0.01546739786863327, + 0.009684654884040356, + 0.01551377959549427, + 0.015102661214768887, + 0.01716475561261177, + 0.011685673147439957, + 0.0082706268876791, + 0.015029423870146275, + 0.013484221883118153, + 0.006585063878446817, + 0.01433590054512024, + 0.011637881398200989, + 0.016685185953974724, + 0.010647663846611977, + 0.014913477934896946, + 0.016413381323218346, + 0.015362200327217579, + 0.008571590296924114, + 0.01653686724603176, + 0.013397359289228916, + 0.018154166638851166, + 0.011224896647036076, + 0.008677447214722633, + 0.011561883613467216, + 0.018120931461453438, + 0.008193537592887878, + 0.019011471420526505, + 0.014282090589404106, + 0.004799632355570793, + 0.005793891381472349, + 0.004633373115211725, + 0.007983755320310593, + 0.008747225627303123, + 0.0036660877522081137, + 0.008850233629345894, + 0.006181746255606413, + 0.0051968893967568874, + 0.0038717142306268215, + 0.0037887291982769966, + 0.006681105587631464, + 0.007566276006400585, + 0.0026589001063257456, + 0.007780272513628006, + 0.005666018463671207, + 0.004904455039650202, + 0.004111217334866524, + 0.0035182295832782984, + 0.0067874290980398655, + 0.007075577042996883, + 0.0024395561777055264, + 0.0040237559005618095, + 0.004909846931695938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.01818566583096981 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_10/index_tensor([3767, 3767])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_10/index_tensor([3767, 3767])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b650a17664c1dcba1351a0b2d314c93a72068e49 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_10/index_tensor([3767, 3767])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide-angle lens pans slowly and steadily to the right, surveying the murky depths of a freshwater river, creating an atmosphere of mystery and tranquility. Faint, aimless sunlight filters down from unseen surfaces, its feeble rays struggling through the murky water, illuminating the dance of suspended silt particles. The riverbed is a soft, undulating landscape of fine sediment and dark sand, scattered with decaying leaves and smooth, algae-dusted stones. A loose school of elephantnose fish glides above the bottom, their dark brown, laterally compressed bodies almost blending into the gloom. The true spectacle is their collective action: their distinctive, elongated, trunk-like jaw barbs twitch and probe the silty substrate with meticulous precision. This feeding is not chaotic, but unfolds in a synchronized, almost choreographed pattern, the entire school's sensory organs sweeping the riverbed in a rhythmic, balletic search for food, their nearly transparent fins constantly quivering, anchoring them to the gentle current.", + "seed": 26095, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_10/index_3767_seed_26095_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10846380870789289, + "vlm_logprob": -0.10846380870789289, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.10846380870789289, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08746425062417984 + ], + "opposite_logprob_tokens": [ + -2.5874643325805664 + ], + "logprob_reward": -0.08746425062417984, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08746425062417984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04186369478702545, + 0.028169546276330948, + 0.04868079721927643, + 0.02325310744345188, + 0.02232307940721512, + 0.1469298005104065, + 0.12470831722021103, + 0.18430620431900024, + 0.36220118403434753, + 0.017564328387379646 + ], + "credit_map": [ + 0.00709556695073843, + 0.00597199983894825, + 0.004542191978543997, + 0.004872883670032024, + 0.0052784280851483345, + 0.005761814769357443, + 0.004942023660987616, + 0.0033987793140113354, + 0.003938019275665283, + 0.0050493283197283745, + 0.0028323810547590256, + 0.0027060152497142553, + 0.004245638847351074, + 0.004048461560159922, + 0.002701474353671074, + 0.0026482234243303537, + 0.008826791308820248, + 0.00762599240988493, + 0.0029824168886989355, + 0.0042829737067222595, + 0.006929839961230755, + 0.006903445348143578, + 0.004677407909184694, + 0.006451922468841076, + 0.0027091982774436474, + 0.002792463405057788, + 0.0023171594366431236, + 0.0017295971047133207, + 0.004260736517608166, + 0.004916402976959944, + 0.0021691853180527687, + 0.0023583616130053997, + 0.001810387009754777, + 0.0027564933989197016, + 0.002381854923442006, + 0.0018183620413765311, + 0.00494791055098176, + 0.003825034247711301, + 0.002786623779684305, + 0.001996410544961691, + 0.008072346448898315, + 0.008005086332559586, + 0.0025689557660371065, + 0.0030836870428174734, + 0.06590571254491806, + 0.052413925528526306, + 0.002950107678771019, + 0.003929963335394859, + 0.014382039196789265, + 0.020285775884985924, + 0.010963775217533112, + 0.009668400511145592, + 0.011369669809937477, + 0.014688990078866482, + 0.02530619129538536, + 0.01804346777498722, + 0.03272518143057823, + 0.04593265429139137, + 0.008789315819740295, + 0.014587079174816608, + 0.022765636444091797, + 0.02495463937520981, + 0.020477257668972015, + 0.014074428007006645, + 0.06293849647045135, + 0.1531328558921814, + 0.011589010246098042, + 0.010267366655170918, + 0.04462171718478203, + 0.05773722007870674, + 0.008516659960150719, + 0.01339782029390335, + 0.0022668002638965845, + 0.0028704942669719458, + 0.0012624816736206412, + 0.0009973393753170967, + 0.003467654576525092, + 0.002881882479414344, + 0.0020358054898679256, + 0.0017818688647821546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.08746425062417984 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008664040826261044 + ], + "opposite_logprob_tokens": [ + -4.758664131164551 + ], + "logprob_reward": -0.008664040826261044, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008664040826261044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1097937598824501, + 0.08181485533714294, + 0.030375264585018158, + 0.04558929428458214, + 0.4143826365470886, + 0.0458855964243412, + 0.06654896587133408, + 0.0515110120177269, + 0.08851336687803268, + 0.06558525562286377 + ], + "credit_map": [ + 0.016837215051054955, + 0.01630215533077717, + 0.010701869614422321, + 0.014547242783010006, + 0.02013152465224266, + 0.011071313172578812, + 0.009754929691553116, + 0.010447514243423939, + 0.010855922475457191, + 0.01644972525537014, + 0.006669254973530769, + 0.009979533031582832, + 0.012488062493503094, + 0.01102464459836483, + 0.008406716398894787, + 0.005940988194197416, + 0.003333801869302988, + 0.0051900772377848625, + 0.0021172603592276573, + 0.0020013318862766027, + 0.007048800587654114, + 0.004568260628730059, + 0.003387636039406061, + 0.0027280983049422503, + 0.006060673389583826, + 0.005254513118416071, + 0.002995890099555254, + 0.0053571718744933605, + 0.008095312863588333, + 0.007455342914909124, + 0.005814387928694487, + 0.004556004423648119, + 0.028097931295633316, + 0.03021654114127159, + 0.022380033507943153, + 0.0213082917034626, + 0.13554947078227997, + 0.10766720771789551, + 0.04115103930234909, + 0.028012104332447052, + 0.004279866814613342, + 0.003708604723215103, + 0.0024978311266750097, + 0.0025321566499769688, + 0.01577129028737545, + 0.011049123480916023, + 0.0024727946147322655, + 0.0035739275626838207, + 0.007477193139493465, + 0.010038208216428757, + 0.004736525937914848, + 0.004210224375128746, + 0.00672117667272687, + 0.007876350544393063, + 0.017597800120711327, + 0.007891484536230564, + 0.00646588858217001, + 0.010147910565137863, + 0.0036001955159008503, + 0.0036348598077893257, + 0.007690053433179855, + 0.007041310425847769, + 0.00619095005095005, + 0.006739842239767313, + 0.014731302857398987, + 0.03278106078505516, + 0.004748085979372263, + 0.004263856913894415, + 0.01055684220045805, + 0.013487235642969608, + 0.003709872718900442, + 0.004235099535435438, + 0.008215237408876419, + 0.011604049243032932, + 0.005679727997630835, + 0.0037481384351849556, + 0.014418804086744785, + 0.010971631854772568, + 0.006037092301994562, + 0.004910574294626713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.008664040826261044 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649386465549469 + ], + "opposite_logprob_tokens": [ + -4.7586493492126465 + ], + "logprob_reward": -0.008649386465549469, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649386465549469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12985838949680328, + 0.09068384021520615, + 0.0465819276869297, + 0.046532876789569855, + 0.16274596750736237, + 0.10951752960681915, + 0.06990426778793335, + 0.17620083689689636, + 0.10072126984596252, + 0.06725306063890457 + ], + "credit_map": [ + 0.012894716113805771, + 0.016926582902669907, + 0.01707022823393345, + 0.01769726537168026, + 0.019863298162817955, + 0.015680287033319473, + 0.01371243316680193, + 0.016013581305742264, + 0.010032364167273045, + 0.015599527396261692, + 0.010108906775712967, + 0.01214832253754139, + 0.015324389562010765, + 0.009416051208972931, + 0.010529518127441406, + 0.007524773012846708, + 0.00467582605779171, + 0.006203868426382542, + 0.003901805728673935, + 0.0048081232234835625, + 0.0077842711471021175, + 0.004999869037419558, + 0.008178528398275375, + 0.006029642652720213, + 0.007281174883246422, + 0.004440131597220898, + 0.00511543545871973, + 0.004390890244394541, + 0.0053504048846662045, + 0.006023772526532412, + 0.007450725883245468, + 0.00648034643381834, + 0.013698285445570946, + 0.013112512417137623, + 0.015742460265755653, + 0.012626856565475464, + 0.038502633571624756, + 0.03387356176972389, + 0.020509354770183563, + 0.014680306427180767, + 0.004856612998992205, + 0.005085395649075508, + 0.004828586708754301, + 0.004886295180767775, + 0.04252491146326065, + 0.035868410021066666, + 0.005333231762051582, + 0.006134091876447201, + 0.008232946507632732, + 0.008798790164291859, + 0.00799950584769249, + 0.0072225346229970455, + 0.007724354043602943, + 0.007786117028445005, + 0.01095468457788229, + 0.01118534617125988, + 0.03382674232125282, + 0.046278081834316254, + 0.010107869282364845, + 0.014685886912047863, + 0.023067960515618324, + 0.024556271731853485, + 0.012357264757156372, + 0.011320776306092739, + 0.008014638908207417, + 0.016796045005321503, + 0.010348742827773094, + 0.008593441918492317, + 0.018337856978178024, + 0.01987910270690918, + 0.008784658275544643, + 0.009966781362891197, + 0.006587177515029907, + 0.009414232335984707, + 0.006027156487107277, + 0.0062942104414105415, + 0.014698758721351624, + 0.009892742149531841, + 0.006787068676203489, + 0.007551724556833506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.008649386465549469 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639931678771973 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639931678771973, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639931678771973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34858086705207825, + 0.042696114629507065, + 0.037844669073820114, + 0.014237659983336926, + 0.15873414278030396, + 0.1648312211036682, + 0.04062837362289429, + 0.07986200600862503, + 0.07682521641254425, + 0.035759735852479935 + ], + "credit_map": [ + 0.030106671154499054, + 0.04999632388353348, + 0.03644498437643051, + 0.048747483640909195, + 0.046021394431591034, + 0.05349958688020706, + 0.046175289899110794, + 0.03758913278579712, + 0.004348117392510176, + 0.008100522682070732, + 0.005008155945688486, + 0.004640782251954079, + 0.00640441058203578, + 0.005417108070105314, + 0.0043414076790213585, + 0.004435614217072725, + 0.004053104668855667, + 0.004923587199300528, + 0.002685771556571126, + 0.0031172202434390783, + 0.006365303881466389, + 0.006428166758269072, + 0.006011400371789932, + 0.004260120447725058, + 0.0021651186980307102, + 0.0012915137922391295, + 0.0012562219053506851, + 0.0009837915422394872, + 0.0023056543432176113, + 0.0019427454099059105, + 0.0016021287301555276, + 0.0026904866099357605, + 0.008855889551341534, + 0.014307481236755848, + 0.007185261696577072, + 0.008606730960309505, + 0.06000017374753952, + 0.04214219003915787, + 0.008972489275038242, + 0.008663923479616642, + 0.00834624469280243, + 0.008899480104446411, + 0.002807274227961898, + 0.003085172502323985, + 0.07442490011453629, + 0.059473052620887756, + 0.0040463898330926895, + 0.00374871538951993, + 0.0049591511487960815, + 0.005453536752611399, + 0.003179374150931835, + 0.0028898895252496004, + 0.004257539752870798, + 0.005415246821939945, + 0.009251739829778671, + 0.00522189773619175, + 0.013508426956832409, + 0.018845094367861748, + 0.0048506250604987144, + 0.005638115108013153, + 0.012402714230120182, + 0.011803695932030678, + 0.0054349107667803764, + 0.007378427777439356, + 0.012258841656148434, + 0.025731945410370827, + 0.00461609149351716, + 0.00470334617421031, + 0.00955381989479065, + 0.009743254631757736, + 0.0051579950377345085, + 0.005059918388724327, + 0.0036837300285696983, + 0.007047368213534355, + 0.0033225687220692635, + 0.002616096753627062, + 0.006202606949955225, + 0.005142371170222759, + 0.0030822877306491137, + 0.004662713501602411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.008639931678771973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42890143394470215 + ], + "opposite_logprob_tokens": [ + -1.0539014339447021 + ], + "logprob_reward": -0.42890143394470215, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42890143394470215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33173561096191406, + 0.05901726335287094, + 0.030236486345529556, + 0.023765409365296364, + 0.11439608782529831, + 0.07442808151245117, + 0.07539795339107513, + 0.12527312338352203, + 0.10648898780345917, + 0.059260960668325424 + ], + "credit_map": [ + 0.031100226566195488, + 0.04125775769352913, + 0.03522422909736633, + 0.05097410827875137, + 0.03767961263656616, + 0.04085944965481758, + 0.056015435606241226, + 0.038624804466962814, + 0.006178222130984068, + 0.012034657411277294, + 0.003521083854138851, + 0.003685101168230176, + 0.016689496114850044, + 0.009250348433852196, + 0.0031677919905632734, + 0.004490564577281475, + 0.003021730575710535, + 0.0042757573537528515, + 0.0020268179941922426, + 0.0039831772446632385, + 0.0050586615689098835, + 0.0039555891416966915, + 0.004837184678763151, + 0.0030775696504861116, + 0.0025138026103377342, + 0.0026514148339629173, + 0.0019462715135887265, + 0.002117612399160862, + 0.0032975589856505394, + 0.0042519038543105125, + 0.002523671602830291, + 0.004463172052055597, + 0.01241485308855772, + 0.01609966531395912, + 0.012133101001381874, + 0.01014813594520092, + 0.018780143931508064, + 0.016431525349617004, + 0.014112837612628937, + 0.014275822788476944, + 0.007406898308545351, + 0.0076524317264556885, + 0.004603797569870949, + 0.008323787711560726, + 0.02182079292833805, + 0.013805077411234379, + 0.004413297399878502, + 0.006401998922228813, + 0.00981721468269825, + 0.008131450042128563, + 0.007406198885291815, + 0.00654420955106616, + 0.006759564392268658, + 0.009240684099495411, + 0.01505391113460064, + 0.012444719672203064, + 0.022825585678219795, + 0.03119710274040699, + 0.0065167429856956005, + 0.008695494383573532, + 0.0159981120377779, + 0.016568249091506004, + 0.013176323845982552, + 0.010295513086020947, + 0.019839851185679436, + 0.044996701180934906, + 0.003618330927565694, + 0.0035755797289311886, + 0.01232984010130167, + 0.01574038900434971, + 0.002490648301318288, + 0.0038976455107331276, + 0.007376549299806356, + 0.009336719289422035, + 0.005556583870202303, + 0.0034109901171177626, + 0.012663225643336773, + 0.008814798668026924, + 0.006462474819272757, + 0.0056396182626485825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.42890143394470215 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle lens pans slowly and steadily to the right, surveying the murky depths of a freshwater river, creating an atmosphere of mystery and tranquility. Faint, aimless sunlight filters down from unseen surfaces, its feeble rays struggling through the murky water, illuminating the dance of suspended silt particles. The riverbed is a soft, undulating landscape of fine sediment and dark sand, scattered with decaying leaves and smooth, algae-dusted stones. A loose school of elephantnose fish glides above the bottom, their dark brown, laterally compressed bodies almost blending into the gloom. The true spectacle is their collective action: their distinctive, elongated, trunk-like jaw barbs twitch and probe the silty substrate with meticulous precision. This feeding is not chaotic, but unfolds in a synchronized, almost choreographed pattern, the entire school's sensory organs sweeping the riverbed in a rhythmic, balletic search for food, their nearly transparent fins constantly quivering, anchoring them to the gentle current.", + "seed": 26096, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_10/index_3767_seed_26096_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16408428233116865, + "vlm_logprob": -0.16408428233116865, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.16408428233116865, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16761673986911774 + ], + "opposite_logprob_tokens": [ + -1.9176167249679565 + ], + "logprob_reward": -0.16761673986911774, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16761673986911774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05555412173271179, + 0.524095892906189, + 0.033305127173662186, + 0.03897320479154587, + 0.057266637682914734, + 0.10745667666196823, + 0.0545496828854084, + 0.02978367544710636, + 0.05703391879796982, + 0.04198111593723297 + ], + "credit_map": [ + 0.006684755440801382, + 0.006649189628660679, + 0.004233426880091429, + 0.00401258235797286, + 0.009832796640694141, + 0.012493856251239777, + 0.005550790578126907, + 0.006096719764173031, + 0.01524286437779665, + 0.023025991395115852, + 0.003524478292092681, + 0.006500497926026583, + 0.24063430726528168, + 0.21968582272529602, + 0.008798261173069477, + 0.006683629006147385, + 0.0027847453020513058, + 0.004567483905702829, + 0.0027357882354408503, + 0.003184483852237463, + 0.006498292554169893, + 0.004789330530911684, + 0.005436781328171492, + 0.0033082221634685993, + 0.0024351272732019424, + 0.0025372710078954697, + 0.0027876694221049547, + 0.002286016708239913, + 0.009453927166759968, + 0.009071540087461472, + 0.003506353357806802, + 0.00689529487863183, + 0.006210345774888992, + 0.009634804911911488, + 0.0039499616250395775, + 0.0035623854491859674, + 0.012684307992458344, + 0.010945749469101429, + 0.0035293761175125837, + 0.006749700754880905, + 0.00450636213645339, + 0.005572729278355837, + 0.001689180382527411, + 0.0018423013389110565, + 0.048744622617959976, + 0.03911401703953743, + 0.0030230232514441013, + 0.002964429557323456, + 0.004571383353322744, + 0.010336020030081272, + 0.007022805046290159, + 0.0058212969452142715, + 0.007544400170445442, + 0.00765154929831624, + 0.006079433020204306, + 0.0055227940902113914, + 0.0037688070442527533, + 0.003861583536490798, + 0.00310573261231184, + 0.003221602411940694, + 0.003773519303649664, + 0.00412559462711215, + 0.004203528631478548, + 0.003723307978361845, + 0.006312191486358643, + 0.005185650195926428, + 0.00790487788617611, + 0.0057951826602220535, + 0.009661493822932243, + 0.009235830046236515, + 0.00644851615652442, + 0.006490173749625683, + 0.00316443108022213, + 0.008512193337082863, + 0.003936060704290867, + 0.0024286224506795406, + 0.014150477014482021, + 0.0057317581959068775, + 0.0024940003640949726, + 0.0015635680174455047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.16761673986911774 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867487773299217 + ], + "opposite_logprob_tokens": [ + -3.0486748218536377 + ], + "logprob_reward": -0.04867487773299217, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867487773299217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04476148262619972, + 0.701723575592041, + 0.05059497058391571, + 0.07819778472185135, + 0.0173453688621521, + 0.024009637534618378, + 0.0072882454842329025, + 0.014997165650129318, + 0.01717873103916645, + 0.04390304163098335 + ], + "credit_map": [ + 0.004730722401291132, + 0.005316364578902721, + 0.0029694766271859407, + 0.005167922470718622, + 0.008960429579019547, + 0.010030372068285942, + 0.00420326879248023, + 0.0033829256426542997, + 0.02019176445901394, + 0.0295922439545393, + 0.004098748788237572, + 0.007667298894375563, + 0.3238842487335205, + 0.2975551187992096, + 0.010810279287397861, + 0.007923845201730728, + 0.0049837250262498856, + 0.006651034113019705, + 0.0027343700639903545, + 0.003376439679414034, + 0.014961558394134045, + 0.008952923119068146, + 0.0055877105332911015, + 0.003347208024933934, + 0.006876011379063129, + 0.006784055382013321, + 0.003299376228824258, + 0.004057933576405048, + 0.025079429149627686, + 0.020670562982559204, + 0.005301146302372217, + 0.006129263434559107, + 0.0019198774825781584, + 0.0020039137452840805, + 0.0014359167544171214, + 0.0011903034755960107, + 0.0034783538430929184, + 0.003510101465508342, + 0.00159314158372581, + 0.0022137598134577274, + 0.0019736322574317455, + 0.0013049625558778644, + 0.0009480293956585228, + 0.0009744462440721691, + 0.00883563794195652, + 0.008361046202480793, + 0.0009236246696673334, + 0.000688257219735533, + 0.0006141268531791866, + 0.001698641455732286, + 0.0007852469570934772, + 0.0006287106079980731, + 0.0009530505631119013, + 0.0011289103422313929, + 0.0007352671818807721, + 0.0007442916394211352, + 0.0013850899413228035, + 0.0027533078100532293, + 0.0012229797430336475, + 0.0010892592836171389, + 0.0036026567686349154, + 0.002575803315266967, + 0.0012452055234462023, + 0.0011228632647544146, + 0.0017583759035915136, + 0.0020264950580894947, + 0.001590965548530221, + 0.001077441032975912, + 0.002971539506688714, + 0.004553570412099361, + 0.0017679224256426096, + 0.0014324195217341185, + 0.004109532106667757, + 0.0071948436088860035, + 0.002551390789449215, + 0.002036442281678319, + 0.015116523019969463, + 0.008458811789751053, + 0.0023322582710534334, + 0.0021032397635281086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.04867487773299217 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789316914975643 + ], + "opposite_logprob_tokens": [ + -4.63478946685791 + ], + "logprob_reward": -0.009789316914975643, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789316914975643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040998831391334534, + 0.7354938983917236, + 0.02269396372139454, + 0.049030184745788574, + 0.05067995563149452, + 0.055982232093811035, + 0.009444163180887699, + 0.011884997598826885, + 0.012528371065855026, + 0.011263354681432247 + ], + "credit_map": [ + 0.0035359435714781284, + 0.004036071244627237, + 0.005149818956851959, + 0.0038061661180108786, + 0.007293253671377897, + 0.0064294901676476, + 0.004965329542756081, + 0.005782764405012131, + 0.021258948370814323, + 0.03129473328590393, + 0.00442880392074585, + 0.008486450649797916, + 0.3394899368286133, + 0.3111838698387146, + 0.011250589974224567, + 0.008100642822682858, + 0.002608350245282054, + 0.0030094152316451073, + 0.0021500501316040754, + 0.002172463573515415, + 0.004202456679195166, + 0.0026239973958581686, + 0.002808760618790984, + 0.003118471009656787, + 0.003431828925386071, + 0.003775515593588352, + 0.004446608945727348, + 0.003275420982390642, + 0.012759258970618248, + 0.013550767675042152, + 0.003451199736446142, + 0.004339591600000858, + 0.00516098877415061, + 0.007079300936311483, + 0.005312452558428049, + 0.004423405975103378, + 0.00838977750390768, + 0.008001014590263367, + 0.0060296314768493176, + 0.006283388007432222, + 0.002260885899886489, + 0.0035066676791757345, + 0.0011648864019662142, + 0.001288213417865336, + 0.024804919958114624, + 0.01897028647363186, + 0.0020764365326613188, + 0.0019099428318440914, + 0.001050180522724986, + 0.0019380406010895967, + 0.0009728881996124983, + 0.0009636497125029564, + 0.0013681279961019754, + 0.0012631573481485248, + 0.0010323047172278166, + 0.0008558155968785286, + 0.001523636165075004, + 0.0018228194676339626, + 0.0010261893039569259, + 0.0011235733982175589, + 0.002146088518202305, + 0.0019195835338905454, + 0.0012194350128993392, + 0.001103673712350428, + 0.0012862267903983593, + 0.001219628844410181, + 0.0009515038109384477, + 0.000748736783862114, + 0.0025502850767225027, + 0.0034633364994078875, + 0.0011851259041577578, + 0.00112352822907269, + 0.001356982160359621, + 0.001900478731840849, + 0.0009264305699616671, + 0.000817478634417057, + 0.0026812164578586817, + 0.0017498492961749434, + 0.0009470031945966184, + 0.0008839163347147405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.009789316914975643 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186602741479874 + ], + "opposite_logprob_tokens": [ + -4.018186569213867 + ], + "logprob_reward": -0.018186602741479874, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186602741479874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050378911197185516, + 0.7973429560661316, + 0.024180473759770393, + 0.008870583027601242, + 0.010404292494058609, + 0.02156880311667919, + 0.0046686348505318165, + 0.028731267899274826, + 0.023357722908258438, + 0.030496278777718544 + ], + "credit_map": [ + 0.006192972417920828, + 0.0062721106223762035, + 0.0037235901691019535, + 0.005380635615438223, + 0.007424043957144022, + 0.010692854411900043, + 0.0056443349458277225, + 0.005048373714089394, + 0.023484261706471443, + 0.03426310047507286, + 0.007748874370008707, + 0.010997273959219456, + 0.36354637145996094, + 0.333156943321228, + 0.014033121056854725, + 0.010113056749105453, + 0.002001962624490261, + 0.003123375354334712, + 0.0018461786676198244, + 0.002625655848532915, + 0.005484758410602808, + 0.003021840238943696, + 0.003736649639904499, + 0.0023400557693094015, + 0.0006034595426172018, + 0.0007083525415509939, + 0.0005865651182830334, + 0.0005517929093912244, + 0.002343283500522375, + 0.002361379563808441, + 0.0005974868545308709, + 0.0011182647431269288, + 0.0009865568717941642, + 0.0014886262360960245, + 0.0010761572048068047, + 0.0009074002155102789, + 0.001534133218228817, + 0.0020846968982368708, + 0.0011202238965779543, + 0.001206498360261321, + 0.0011594693642109632, + 0.0008541256538592279, + 0.00046679668594151735, + 0.0003729296731762588, + 0.009164956398308277, + 0.008319025859236717, + 0.0006737430230714381, + 0.0005577581468969584, + 0.0004207710735499859, + 0.0008415049524046481, + 0.0005463490961119533, + 0.00046041529276408255, + 0.0009653431479819119, + 0.0007906776154413819, + 0.0003523074265103787, + 0.0002912663039751351, + 0.0023871236480772495, + 0.004948044195771217, + 0.001976600382477045, + 0.002108706859871745, + 0.006953112781047821, + 0.005289985798299313, + 0.002690665191039443, + 0.002377030672505498, + 0.0022932004649192095, + 0.001989910379052162, + 0.0033035180531442165, + 0.0020910995081067085, + 0.005062445532530546, + 0.004860044457018375, + 0.001986598828807473, + 0.0017709083622321486, + 0.0021713636815547943, + 0.005526955705136061, + 0.002666148357093334, + 0.001998492516577244, + 0.010231230407953262, + 0.004294943064451218, + 0.002054477808997035, + 0.001552669913507998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.018186602741479874 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761538743972778 + ], + "opposite_logprob_tokens": [ + -0.8261538743972778 + ], + "logprob_reward": -0.5761538743972778, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761538743972778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09812285751104355, + 0.4720252752304077, + 0.04722261428833008, + 0.022953137755393982, + 0.03685638681054115, + 0.04079483076930046, + 0.01771862432360649, + 0.08875861763954163, + 0.09672419726848602, + 0.07882342487573624 + ], + "credit_map": [ + 0.009181694127619267, + 0.009605580940842628, + 0.006471563130617142, + 0.011496013961732388, + 0.02178925834596157, + 0.021336574107408524, + 0.010106459259986877, + 0.008135713636875153, + 0.013986844569444656, + 0.021176796406507492, + 0.0044148387387394905, + 0.00670698843896389, + 0.21275290846824646, + 0.19474442303180695, + 0.011127280071377754, + 0.0071152071468532085, + 0.004912212956696749, + 0.00786635372787714, + 0.0032654323149472475, + 0.0036173255648463964, + 0.01077426876872778, + 0.006808890495449305, + 0.00523777399212122, + 0.004740356933325529, + 0.001760340528562665, + 0.0024642758071422577, + 0.00210746843367815, + 0.0018913629464805126, + 0.0053267753683030605, + 0.0053212703205645084, + 0.002035513985902071, + 0.0020461289677768946, + 0.003246912034228444, + 0.004659285303205252, + 0.004474671091884375, + 0.0038886875845491886, + 0.0051773926243186, + 0.005545663647353649, + 0.004169352352619171, + 0.005694422870874405, + 0.003164087887853384, + 0.0023247564677149057, + 0.0015541046159341931, + 0.0016305948374792933, + 0.015207394026219845, + 0.014493724331259727, + 0.0012389603070914745, + 0.0011812119046226144, + 0.0016634364146739244, + 0.0037526569794863462, + 0.0019737628754228354, + 0.0018482288578525186, + 0.00263458420522511, + 0.0023912834003567696, + 0.001979988068342209, + 0.0014746830565854907, + 0.006822917144745588, + 0.013152181170880795, + 0.0056603290140628815, + 0.005380002316087484, + 0.0252060629427433, + 0.02078746259212494, + 0.006342487875372171, + 0.005407170858234167, + 0.009967437013983727, + 0.009550516493618488, + 0.01393112726509571, + 0.007333694491535425, + 0.020364083349704742, + 0.020341431722044945, + 0.00781454611569643, + 0.007421355228871107, + 0.0049877590499818325, + 0.016747692599892616, + 0.004340830724686384, + 0.0032103636767715216, + 0.03157654404640198, + 0.011553100310266018, + 0.003915306646376848, + 0.002491828054189682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.5761538743972778 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle lens pans slowly and steadily to the right, surveying the murky depths of a freshwater river, creating an atmosphere of mystery and tranquility. Faint, aimless sunlight filters down from unseen surfaces, its feeble rays struggling through the murky water, illuminating the dance of suspended silt particles. The riverbed is a soft, undulating landscape of fine sediment and dark sand, scattered with decaying leaves and smooth, algae-dusted stones. A loose school of elephantnose fish glides above the bottom, their dark brown, laterally compressed bodies almost blending into the gloom. The true spectacle is their collective action: their distinctive, elongated, trunk-like jaw barbs twitch and probe the silty substrate with meticulous precision. This feeding is not chaotic, but unfolds in a synchronized, almost choreographed pattern, the entire school's sensory organs sweeping the riverbed in a rhythmic, balletic search for food, their nearly transparent fins constantly quivering, anchoring them to the gentle current.", + "seed": 26097, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_10/index_3767_seed_26097_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.18715136870741844, + "vlm_logprob": -0.18715136870741844, + "vlm_yes_no_margin": 2.05, + "vlm_token_credit": -0.18715136870741844, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2866765856742859 + ], + "opposite_logprob_tokens": [ + -1.4116766452789307 + ], + "logprob_reward": -0.2866765856742859, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2866765856742859, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13658301532268524, + 0.6743033528327942, + 0.04522561654448509, + 0.0067290375009179115, + 0.006489089224487543, + 0.08646350353956223, + 0.005969207733869553, + 0.015320513397455215, + 0.007824075408279896, + 0.015092570334672928 + ], + "credit_map": [ + 0.008199404925107956, + 0.0385466143488884, + 0.0068398029543459415, + 0.008048182353377342, + 0.04530883580446243, + 0.01573309302330017, + 0.004867687355726957, + 0.009039402939379215, + 0.013405744917690754, + 0.2740599513053894, + 0.006270538549870253, + 0.0063614072278141975, + 0.3161993622779846, + 0.04454892873764038, + 0.007885637693107128, + 0.005571769084781408, + 0.0018556376453489065, + 0.013460301794111729, + 0.0027254936285316944, + 0.004021399188786745, + 0.01479188259691, + 0.003446021117269993, + 0.002723540412262082, + 0.002201338764280081, + 0.0006671486189588904, + 0.0008362679509446025, + 0.0008688599336892366, + 0.0008352245786227286, + 0.0008345572859980166, + 0.0007740366854704916, + 0.0010918524349108338, + 0.0008210901287384331, + 0.0005464269779622555, + 0.0009148301323875785, + 0.0007764706388115883, + 0.0008067681919783354, + 0.0010098490165546536, + 0.001260198070667684, + 0.0005752251599915326, + 0.0005993208615109324, + 0.0062073105946183205, + 0.02040022984147072, + 0.005796937271952629, + 0.006949261762201786, + 0.01921760104596615, + 0.018134435638785362, + 0.005876477342098951, + 0.003881247015669942, + 0.0004303298192098737, + 0.0010849297977983952, + 0.0005149427452124655, + 0.0005352513398975134, + 0.0014641775051131845, + 0.000688195985276252, + 0.0005048852181062102, + 0.0007464950322173536, + 0.0009946374921128154, + 0.0022378277499228716, + 0.0014842173550277948, + 0.0018169842660427094, + 0.002588990842923522, + 0.002115772571414709, + 0.0016175276832655072, + 0.0024645559024065733, + 0.0006601400673389435, + 0.002490247832611203, + 0.0007563292165286839, + 0.0007315160473808646, + 0.0013520893407985568, + 0.0008332760189659894, + 0.0005721242632716894, + 0.0004283524467609823, + 0.0012247789418324828, + 0.0019816611893475056, + 0.0019647327717393637, + 0.002071054419502616, + 0.001973964273929596, + 0.00179017533082515, + 0.0020523720886558294, + 0.002033831086009741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.2866765856742859 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207394227385521 + ], + "opposite_logprob_tokens": [ + -2.8120739459991455 + ], + "logprob_reward": -0.06207394227385521, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207394227385521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0292962733656168, + 0.9213209748268127, + 0.016618916764855385, + 0.00197726022452116, + 0.00518819410353899, + 0.011320499703288078, + 0.0015637666219845414, + 0.004340775310993195, + 0.003991853911429644, + 0.004381546750664711 + ], + "credit_map": [ + 0.00217505544424057, + 0.006373124197125435, + 0.0020905763376504183, + 0.00211439561098814, + 0.006300691980868578, + 0.004449176602065563, + 0.002969898981973529, + 0.0028233530465513468, + 0.01796804554760456, + 0.373913049697876, + 0.008453737944364548, + 0.008035157807171345, + 0.43322932720184326, + 0.0627937987446785, + 0.009967322461307049, + 0.00696043623611331, + 0.000768990081269294, + 0.004845125135034323, + 0.0010618099477142096, + 0.0012632319703698158, + 0.0053155068308115005, + 0.001393871963955462, + 0.000992053421214223, + 0.0009783277055248618, + 0.0001878155890153721, + 0.0002843399706762284, + 0.00020741036860272288, + 0.00024268093693535775, + 0.0002974545059259981, + 0.0002900141989812255, + 0.00023512562620453537, + 0.00023241896997205913, + 0.0005360879586078227, + 0.0007976830238476396, + 0.0005887039587832987, + 0.0006681095692329109, + 0.0008078233804553747, + 0.0008292237180285156, + 0.00047655272646807134, + 0.00048400956438854337, + 0.0007457216270267963, + 0.0022546108812093735, + 0.0008208255167119205, + 0.0008348728297278285, + 0.0027796942740678787, + 0.002519893692806363, + 0.0008385727414861321, + 0.0005263076163828373, + 0.0001257025869563222, + 0.0003611641004681587, + 0.0001127188588725403, + 0.00011820333747891709, + 0.00040852336678653955, + 0.0001875763264251873, + 0.00011980388808296993, + 0.00013007396773900837, + 0.0002803556271828711, + 0.0007407563389278948, + 0.0004081267397850752, + 0.00043273635674268007, + 0.0007912991568446159, + 0.0006439904682338238, + 0.00042902168934233487, + 0.0006144887302070856, + 0.00033140217419713736, + 0.0008189082145690918, + 0.0004880518827121705, + 0.0005590249784290791, + 0.0006156956660561264, + 0.000544198730494827, + 0.00027728595887310803, + 0.0003572863060981035, + 0.0004209514008834958, + 0.0007771722157485783, + 0.0006388546316884458, + 0.0005005559651181102, + 0.0005996202817186713, + 0.0005933187203481793, + 0.0003866876650135964, + 0.00046438543358817697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.06207394227385521 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03374212980270386 + ], + "opposite_logprob_tokens": [ + -3.4087421894073486 + ], + "logprob_reward": -0.03374212980270386, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03374212980270386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023191291838884354, + 0.9376771450042725, + 0.008684269152581692, + 0.0028120537754148245, + 0.00827289093285799, + 0.005659562069922686, + 0.002313828095793724, + 0.005845199339091778, + 0.001375841093249619, + 0.004167971666902304 + ], + "credit_map": [ + 0.0020942152477800846, + 0.0033002113923430443, + 0.0020305460784584284, + 0.0018199350452050567, + 0.00466730585321784, + 0.0035835776943713427, + 0.0027943947352468967, + 0.002901105908676982, + 0.018216708675026894, + 0.37963995337486267, + 0.008649284951388836, + 0.007777514401823282, + 0.4414682388305664, + 0.06509417295455933, + 0.009659756906330585, + 0.00717151677235961, + 0.0004809601232409477, + 0.0023141743149608374, + 0.000737309455871582, + 0.0008118670666590333, + 0.0022078913170844316, + 0.0008582058944739401, + 0.0006760904216207564, + 0.0005977707915008068, + 0.00025494556757621467, + 0.0004121488018427044, + 0.00027976484852842987, + 0.0003150491102132946, + 0.0004804435302503407, + 0.00042607038631103933, + 0.0003246748819947243, + 0.0003189566486980766, + 0.0008493817294947803, + 0.0011762476060539484, + 0.0008693899144418538, + 0.0009580581681802869, + 0.0012373136123642325, + 0.0015357054071500897, + 0.0008543053409084678, + 0.00079248973634094, + 0.00039603799814358354, + 0.0009795649675652385, + 0.00045786635018885136, + 0.000545453280210495, + 0.001132404082454741, + 0.0010969731956720352, + 0.0005246041691862047, + 0.0005266582011245191, + 0.00018329777230974287, + 0.0004876654129475355, + 0.00020876081543974578, + 0.00022094539599493146, + 0.00033807253930717707, + 0.00039777904748916626, + 0.00018380139954388142, + 0.00029350578552111983, + 0.0004065193352289498, + 0.000847483694087714, + 0.0005567191983573139, + 0.0005693923449143767, + 0.0010729000205174088, + 0.0007441900088451803, + 0.0005545811145566404, + 0.0010934133315458894, + 0.00011452979379100725, + 0.0002587762719485909, + 0.00015675243048463017, + 0.00018160896433982998, + 0.0002300548949278891, + 0.00017449670122005045, + 0.00010967040725518018, + 0.0001499515783507377, + 0.0003381843853276223, + 0.000787166936788708, + 0.00044225924648344517, + 0.0006839096895419061, + 0.0005389199359342456, + 0.000492968421895057, + 0.00042906199814751744, + 0.0004555014311335981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.03374212980270386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899175584316254 + ], + "opposite_logprob_tokens": [ + -2.578991651535034 + ], + "logprob_reward": -0.07899175584316254, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899175584316254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06866907328367233, + 0.8882221579551697, + 0.012842957861721516, + 0.002547157695516944, + 0.0012158992467448115, + 0.012386228889226913, + 0.004871377255767584, + 0.004237135406583548, + 0.0011991868959739804, + 0.0038088627625256777 + ], + "credit_map": [ + 0.004364446736872196, + 0.019412798807024956, + 0.0035512931644916534, + 0.004031694959849119, + 0.02035325951874256, + 0.009342347271740437, + 0.003302281256765127, + 0.004310951568186283, + 0.017565248534083366, + 0.36203473806381226, + 0.008251190185546875, + 0.008282588794827461, + 0.41663411259651184, + 0.05865190550684929, + 0.010010742582380772, + 0.006791599094867706, + 0.0007128444849513471, + 0.0030274426098912954, + 0.0009212002041749656, + 0.0016329734353348613, + 0.0034044114872813225, + 0.0013014711439609528, + 0.0011223857291042805, + 0.0007202285341918468, + 0.0002258506865473464, + 0.00030440479167737067, + 0.0003138821048196405, + 0.00036203398485668004, + 0.00035872007720172405, + 0.0003132323909085244, + 0.0003546248481143266, + 0.0003144086804240942, + 0.00010534813918638974, + 0.0002268793759867549, + 0.0001419862819602713, + 0.00012056633568136021, + 0.0002069640177069232, + 0.00021048115740995854, + 9.885456529445946e-05, + 0.00010481938807060942, + 0.0009248250280506909, + 0.002940994454547763, + 0.0008233895641751587, + 0.0009978627786040306, + 0.0025895212311297655, + 0.002597868675366044, + 0.0009246965637430549, + 0.0005870694876648486, + 0.0003213500021956861, + 0.0012842887081205845, + 0.00030518515268340707, + 0.00039375934284180403, + 0.0012909932993352413, + 0.0007763463072478771, + 0.0002936817181762308, + 0.00020577307441271842, + 0.000309162016492337, + 0.0007510183495469391, + 0.0004234827065374702, + 0.0005053933709859848, + 0.0006351316696964204, + 0.0005089426413178444, + 0.00032046184060163796, + 0.0007835427531972528, + 0.00012610880366992205, + 0.0003073653788305819, + 0.00013973176828585565, + 0.0001374411949655041, + 0.00018431981152389199, + 0.00014646636554971337, + 7.868569809943438e-05, + 7.906787504907697e-05, + 0.0002548308693803847, + 0.0007893036236055195, + 0.0003551378904376179, + 0.0006151222041808069, + 0.0005555010284297168, + 0.0004863631329499185, + 0.0003840746940113604, + 0.0003685294941533357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.07899175584316254 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742724299430847 + ], + "opposite_logprob_tokens": [ + -0.9742724299430847 + ], + "logprob_reward": -0.4742724299430847, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742724299430847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052151065319776535, + 0.9148610830307007, + 0.013386441394686699, + 0.0017038939986377954, + 0.0017345743253827095, + 0.00765618309378624, + 0.0016987437848001719, + 0.0021142205223441124, + 0.0010355078848078847, + 0.0036582392640411854 + ], + "credit_map": [ + 0.0030132178217172623, + 0.015419426374137402, + 0.0022281999699771404, + 0.0026578677352517843, + 0.01830166019499302, + 0.006032843142747879, + 0.0016713518416509032, + 0.0028264913707971573, + 0.01802186667919159, + 0.3727196455001831, + 0.008503458462655544, + 0.008731952868402004, + 0.42895960807800293, + 0.0604420006275177, + 0.010205411352217197, + 0.007277068216353655, + 0.0004156534268986434, + 0.004563621710985899, + 0.0005764611414633691, + 0.000941951060667634, + 0.004867005627602339, + 0.0008711022674106061, + 0.000591942691244185, + 0.0005587005871348083, + 0.00015833525685593486, + 0.0002401326346443966, + 0.0002065685548586771, + 0.0002540147106628865, + 0.0002861008106265217, + 0.0001801564358174801, + 0.00019302814325783402, + 0.0001855571463238448, + 0.00018471444491297007, + 0.00034297307138331234, + 0.0001371486287098378, + 0.00017595983808860183, + 0.000353759853169322, + 0.00022263811843004078, + 0.00015030060603749007, + 0.00016707951726857573, + 0.000561388791538775, + 0.0017490548780187964, + 0.0005666132783517241, + 0.0006362319109030068, + 0.0016006960067898035, + 0.0015447946498170495, + 0.0006046642665751278, + 0.0003927381767425686, + 0.00016483348736073822, + 0.00031180097721517086, + 0.00013399016461335123, + 0.00017145593301393092, + 0.00032398998155258596, + 0.00022784991597291082, + 0.00016634778876323253, + 0.0001984752161661163, + 0.00015286882990039885, + 0.0003412002988625318, + 0.00023432073066942394, + 0.0003855180984828621, + 0.00029226060723885894, + 0.00021711198496632278, + 0.00020881900854874402, + 0.00028212059987708926, + 9.461611625738442e-05, + 0.0002725736703723669, + 9.543704800307751e-05, + 0.00010603363625705242, + 0.00020921335089951754, + 0.00010891099373111501, + 7.163685950217769e-05, + 7.70861079217866e-05, + 0.00030513585079461336, + 0.0005272693233564496, + 0.0004320040752645582, + 0.000488603487610817, + 0.0005601979210041463, + 0.00044696155237033963, + 0.0004312319797463715, + 0.000466834258986637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.4742724299430847 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle lens pans slowly and steadily to the right, surveying the murky depths of a freshwater river, creating an atmosphere of mystery and tranquility. Faint, aimless sunlight filters down from unseen surfaces, its feeble rays struggling through the murky water, illuminating the dance of suspended silt particles. The riverbed is a soft, undulating landscape of fine sediment and dark sand, scattered with decaying leaves and smooth, algae-dusted stones. A loose school of elephantnose fish glides above the bottom, their dark brown, laterally compressed bodies almost blending into the gloom. The true spectacle is their collective action: their distinctive, elongated, trunk-like jaw barbs twitch and probe the silty substrate with meticulous precision. This feeding is not chaotic, but unfolds in a synchronized, almost choreographed pattern, the entire school's sensory organs sweeping the riverbed in a rhythmic, balletic search for food, their nearly transparent fins constantly quivering, anchoring them to the gentle current.", + "seed": 26098, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_10/index_3767_seed_26098_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.21808922346681356, + "vlm_logprob": -0.21808922346681356, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.21808922346681356, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28804993629455566 + ], + "opposite_logprob_tokens": [ + -1.4130499362945557 + ], + "logprob_reward": -0.28804993629455566, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28804993629455566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2474534809589386, + 0.11638724058866501, + 0.06462089717388153, + 0.10792740434408188, + 0.156311497092247, + 0.06902250647544861, + 0.0463922955095768, + 0.048551883548498154, + 0.042687881737947464, + 0.10064490139484406 + ], + "credit_map": [ + 0.04090461879968643, + 0.03550132364034653, + 0.0365709513425827, + 0.026542939245700836, + 0.032128095626831055, + 0.03995349630713463, + 0.014849412254989147, + 0.021002646535634995, + 0.01758563704788685, + 0.014608892612159252, + 0.015911586582660675, + 0.012311430647969246, + 0.014769133180379868, + 0.016664227470755577, + 0.011163217946887016, + 0.013373119756579399, + 0.010596639476716518, + 0.008890082128345966, + 0.006780292838811874, + 0.008997215889394283, + 0.009759555570781231, + 0.005386741366237402, + 0.006422942504286766, + 0.007787428796291351, + 0.022286955267190933, + 0.014625963754951954, + 0.01005738414824009, + 0.009034814313054085, + 0.01359274610877037, + 0.010444806888699532, + 0.01329941675066948, + 0.014585316181182861, + 0.015071770176291466, + 0.02075636200606823, + 0.024980053305625916, + 0.012363742105662823, + 0.026207588613033295, + 0.01403854414820671, + 0.01979777216911316, + 0.023095646873116493, + 0.009719626046717167, + 0.00997356977313757, + 0.011458955705165863, + 0.006154986564069986, + 0.009289613924920559, + 0.005571790970861912, + 0.0077455779537558556, + 0.009108380414545536, + 0.006088003516197205, + 0.005441153887659311, + 0.004500869195908308, + 0.003925475291907787, + 0.005040224175900221, + 0.00415279995650053, + 0.007518203463405371, + 0.009725567884743214, + 0.009741831570863724, + 0.005911107640713453, + 0.003989080432802439, + 0.00630668131634593, + 0.0075622377917170525, + 0.004409101326018572, + 0.005925115197896957, + 0.004706728737801313, + 0.004110212437808514, + 0.005069596227258444, + 0.00662932638078928, + 0.005634883884340525, + 0.005154111422598362, + 0.0027757976204156876, + 0.007726695854216814, + 0.0055872597731649876, + 0.011739599518477917, + 0.011861298233270645, + 0.010334152728319168, + 0.01745850406587124, + 0.013427835889160633, + 0.008119245059788227, + 0.016323333606123924, + 0.011380932293832302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.28804993629455566 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421962771564722 + ], + "opposite_logprob_tokens": [ + -4.264219760894775 + ], + "logprob_reward": -0.01421962771564722, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421962771564722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14457327127456665, + 0.09132501482963562, + 0.07371145486831665, + 0.1514803171157837, + 0.07562891393899918, + 0.04406028985977173, + 0.10863944888114929, + 0.16353587806224823, + 0.11741560697555542, + 0.02962978184223175 + ], + "credit_map": [ + 0.017905311658978462, + 0.01696142368018627, + 0.018883269280195236, + 0.020009450614452362, + 0.01685650460422039, + 0.01699034310877323, + 0.016975317150354385, + 0.019991664215922356, + 0.011652547866106033, + 0.01132764108479023, + 0.008564529940485954, + 0.011787286028265953, + 0.013667671009898186, + 0.009497122839093208, + 0.011824809946119785, + 0.013003413565456867, + 0.014795301482081413, + 0.010095329023897648, + 0.0068795401602983475, + 0.007938024587929249, + 0.010966450907289982, + 0.008837954141199589, + 0.005306756589561701, + 0.008892102167010307, + 0.022208500653505325, + 0.019780656322836876, + 0.014106689020991325, + 0.014612183906137943, + 0.023429401218891144, + 0.01743609644472599, + 0.018292736262083054, + 0.0216140728443861, + 0.010763948783278465, + 0.009666146710515022, + 0.007656380534172058, + 0.007265905849635601, + 0.013159233145415783, + 0.007172939833253622, + 0.008321751840412617, + 0.011622614227235317, + 0.007133039645850658, + 0.007775055710226297, + 0.004957559052854776, + 0.0038641938008368015, + 0.006185502279549837, + 0.002953726565465331, + 0.0050851041451096535, + 0.006106110289692879, + 0.01399102434515953, + 0.01270593237131834, + 0.009787185117602348, + 0.01547111477702856, + 0.012540893629193306, + 0.007030121516436338, + 0.019181979820132256, + 0.017931198701262474, + 0.03645266965031624, + 0.02007584646344185, + 0.008979951031506062, + 0.010615176521241665, + 0.04135364666581154, + 0.012827824801206589, + 0.0174450371414423, + 0.01578574813902378, + 0.024503106251358986, + 0.014256197027862072, + 0.008932698518037796, + 0.006382604129612446, + 0.019422467797994614, + 0.014886670745909214, + 0.011730124242603779, + 0.01730174571275711, + 0.0032898171339184046, + 0.0031672080513089895, + 0.0028394002001732588, + 0.005478325765579939, + 0.0027574668638408184, + 0.0019479969050735235, + 0.004488951526582241, + 0.005660620052367449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.01421962771564722 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789552539587021 + ], + "opposite_logprob_tokens": [ + -4.63478946685791 + ], + "logprob_reward": -0.009789552539587021, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789552539587021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14248058199882507, + 0.046755243092775345, + 0.05339322239160538, + 0.033539123833179474, + 0.07184423506259918, + 0.0850914716720581, + 0.10547017306089401, + 0.24654699862003326, + 0.1368224173784256, + 0.07805655151605606 + ], + "credit_map": [ + 0.020274754613637924, + 0.018277479335665703, + 0.01491962093859911, + 0.016659386456012726, + 0.021315449848771095, + 0.023221319541335106, + 0.0117658581584692, + 0.016046710312366486, + 0.006129407789558172, + 0.005950053222477436, + 0.004976760596036911, + 0.006851581856608391, + 0.006766411010175943, + 0.005793638061732054, + 0.00526651693508029, + 0.0050208731554448605, + 0.008610381744801998, + 0.005739843472838402, + 0.006117558106780052, + 0.0069274199195206165, + 0.006658552214503288, + 0.0052049849182367325, + 0.005624652374535799, + 0.008509833365678787, + 0.004455452784895897, + 0.003724539652466774, + 0.00395840872079134, + 0.002883091103285551, + 0.003588346065953374, + 0.0037596747279167175, + 0.006339295767247677, + 0.004830314312130213, + 0.011101889424026012, + 0.010091077536344528, + 0.008393983356654644, + 0.005240337923169136, + 0.008724762126803398, + 0.00645734928548336, + 0.009106677025556564, + 0.012728164903819561, + 0.012971317395567894, + 0.012073757126927376, + 0.015093864873051643, + 0.008237314410507679, + 0.007981540635228157, + 0.00873700063675642, + 0.007404447998851538, + 0.012592228129506111, + 0.012647378258407116, + 0.013805180788040161, + 0.01637127995491028, + 0.011227201670408249, + 0.014417111873626709, + 0.007936782203614712, + 0.013400115072727203, + 0.015665123239159584, + 0.045327283442020416, + 0.030509423464536667, + 0.014435168355703354, + 0.01810331456363201, + 0.06198416277766228, + 0.020667558535933495, + 0.029429852962493896, + 0.026090234518051147, + 0.024243924766778946, + 0.015749741345643997, + 0.010039674118161201, + 0.008512010797858238, + 0.030319195240736008, + 0.019693147391080856, + 0.013249577023088932, + 0.01501514296978712, + 0.015525157563388348, + 0.009397710673511028, + 0.007540043443441391, + 0.010506125167012215, + 0.007842141203582287, + 0.0059145307168364525, + 0.010759582743048668, + 0.010571260005235672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.009789552539587021 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058218978345394 + ], + "opposite_logprob_tokens": [ + -3.8955821990966797 + ], + "logprob_reward": -0.02058218978345394, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058218978345394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17451679706573486, + 0.06837635487318039, + 0.08814668655395508, + 0.13170082867145538, + 0.02879124879837036, + 0.03980761021375656, + 0.04605649411678314, + 0.20930461585521698, + 0.13854219019412994, + 0.07475714385509491 + ], + "credit_map": [ + 0.014989502727985382, + 0.02079441025853157, + 0.023960018530488014, + 0.02938375249505043, + 0.018334880471229553, + 0.017750954255461693, + 0.027218583971261978, + 0.02208467572927475, + 0.011035598814487457, + 0.009054934605956078, + 0.008266124874353409, + 0.00795857422053814, + 0.009507526643574238, + 0.006473673041909933, + 0.007648597937077284, + 0.008431321009993553, + 0.01275460235774517, + 0.011444799602031708, + 0.009064688347280025, + 0.0109299561008811, + 0.015139675699174404, + 0.008905930444598198, + 0.008555332198739052, + 0.011351702734827995, + 0.01585681550204754, + 0.016942743211984634, + 0.01618051528930664, + 0.014942986890673637, + 0.01844698190689087, + 0.0137623380869627, + 0.018854232504963875, + 0.01671421341598034, + 0.005539261270314455, + 0.0038386937230825424, + 0.002979908138513565, + 0.0038688790518790483, + 0.003767890390008688, + 0.0028741625137627125, + 0.0028782659210264683, + 0.0030441884882748127, + 0.006935361307114363, + 0.0053445277735590935, + 0.0040974547155201435, + 0.0034402033779770136, + 0.004591915290802717, + 0.004534451756626368, + 0.005048554390668869, + 0.005815140902996063, + 0.006861527916043997, + 0.0060870167799293995, + 0.007249221205711365, + 0.005101905670017004, + 0.005293915048241615, + 0.0035583432763814926, + 0.005328298080712557, + 0.006576265674084425, + 0.03373347967863083, + 0.024506064131855965, + 0.015422157943248749, + 0.021600298583507538, + 0.049661923199892044, + 0.017741983756422997, + 0.025785205885767937, + 0.02085350826382637, + 0.02218795195221901, + 0.015592040494084358, + 0.01013944111764431, + 0.00969698466360569, + 0.03295207768678665, + 0.01946277730166912, + 0.014364762231707573, + 0.014146162196993828, + 0.010870294645428658, + 0.009394736029207706, + 0.005976615007966757, + 0.01169915683567524, + 0.010011374019086361, + 0.006630284246057272, + 0.012451003305613995, + 0.007723670452833176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.02058218978345394 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a slow, steady pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a murky freshwater river with suspended silt particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the riverbed composed of fine sediment, dark sand, decaying leaves, and algae-dusted stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dark brown elephantnose fish visible gliding above the bottom with their distinctive trunk-like jaw barbs twitching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the school of fish feeding in a synchronized, rhythmic pattern with their fins constantly quivering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.757804811000824 + ], + "opposite_logprob_tokens": [ + -0.632804811000824 + ], + "logprob_reward": -0.757804811000824, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.757804811000824, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.254936546087265, + 0.07458842545747757, + 0.11490577459335327, + 0.15864522755146027, + 0.044788770377635956, + 0.03559981286525726, + 0.046991340816020966, + 0.1453900933265686, + 0.07542189955711365, + 0.048732127994298935 + ], + "credit_map": [ + 0.0265364907681942, + 0.029837356880307198, + 0.03485504537820816, + 0.04317506030201912, + 0.027935346588492393, + 0.031122788786888123, + 0.028391553089022636, + 0.033082909882068634, + 0.012921320274472237, + 0.009205900132656097, + 0.008587469346821308, + 0.00853781308978796, + 0.009397194720804691, + 0.009628851898014545, + 0.006464495323598385, + 0.009845377877354622, + 0.01827174238860607, + 0.013630587607622147, + 0.013880257494747639, + 0.016095096245408058, + 0.016100849956274033, + 0.013177857734262943, + 0.009761713445186615, + 0.013987661339342594, + 0.025442717596888542, + 0.01969773881137371, + 0.01539033092558384, + 0.01613171026110649, + 0.018399210646748543, + 0.016763783991336823, + 0.014019008725881577, + 0.03280071169137955, + 0.008965442888438702, + 0.006429802160710096, + 0.005286207888275385, + 0.004207922611385584, + 0.006236336659640074, + 0.004493018612265587, + 0.003953577484935522, + 0.005216458812355995, + 0.0054714977741241455, + 0.005081328097730875, + 0.005119304172694683, + 0.0028421368915587664, + 0.004241554532200098, + 0.0026071767788380384, + 0.004139693919569254, + 0.006097116973251104, + 0.008999894373118877, + 0.006335325073450804, + 0.004904086701571941, + 0.004829057026654482, + 0.006856665480881929, + 0.002934250980615616, + 0.006540786940604448, + 0.005591277498751879, + 0.016316378489136696, + 0.01641036570072174, + 0.012483516708016396, + 0.020094169303774834, + 0.02973921038210392, + 0.011745884083211422, + 0.018961627036333084, + 0.01963893324136734, + 0.011346463114023209, + 0.008434123359620571, + 0.005539834499359131, + 0.007473255041986704, + 0.018201366066932678, + 0.010022707283496857, + 0.008576647378504276, + 0.005827501881867647, + 0.006052765063941479, + 0.005383925512433052, + 0.005302717909216881, + 0.007758826483041048, + 0.0061340974643826485, + 0.0038825219962745905, + 0.008881705813109875, + 0.005335567984730005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.757804811000824 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_11/index_tensor([6402, 6402])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_11/index_tensor([6402, 6402])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d3d473c78651117930978b53ff58f69f3ccd62c7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_11/index_tensor([6402, 6402])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static wide-angle shot, captured from a gentle overhead angle, captures a small wooden sightseeing boat anchored in the middle of a broad, recently receding waterway. The waters are an opaque, milk chocolate-slurry hue, barely rippled by a gentle undercurrent that causes the boat to pitch back and forth with a slow, rhythmic lurch. The boat, painted a slightly worn creamy white with faded sky-blue trim along the edges, is empty save for a few simple rows of wooden benches. The entire image is bathed in a distinct soft-focus effect and a warm, golden glow, as if the afternoon sun is filtering through a mist, blurring the sharp edges of all objects. From the water to the boat's hull to the waterlogged riverbank in the distance, every surface possesses a hand-painted brushstroke and visible texture, as if it were a living, breathing illustration from a classic fairytale, creating an atmosphere of profound tranquility, innocence, and a touch of dreamlike childlike innocence.", + "seed": 15147, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_11/index_6402_seed_15147_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01856316349003464, + "vlm_logprob": -0.01856316349003464, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.01856316349003464, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021781159564852715 + ], + "opposite_logprob_tokens": [ + -5.646780967712402 + ], + "logprob_reward": -0.021781159564852715, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.021781159564852715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13601015508174896, + 0.06198925897479057, + 0.06651018559932709, + 0.0920877456665039, + 0.05569462850689888, + 0.06327401101589203, + 0.0713813453912735, + 0.22806555032730103, + 0.145779088139534, + 0.07920800149440765 + ], + "credit_map": [ + 0.024569256231188774, + 0.010257460176944733, + 0.02711811661720276, + 0.013059969991445541, + 0.009038319811224937, + 0.01487824134528637, + 0.013139921240508556, + 0.023948874324560165, + 0.013281011953949928, + 0.0051578632555902, + 0.008333574049174786, + 0.0055186692625284195, + 0.004349295981228352, + 0.008739043958485126, + 0.006940981838852167, + 0.009668816812336445, + 0.013522356748580933, + 0.008643907494843006, + 0.011641141027212143, + 0.005230342037975788, + 0.004728678613901138, + 0.007590267341583967, + 0.004175922833383083, + 0.010977568104863167, + 0.020021818578243256, + 0.0068318950943648815, + 0.01477840356528759, + 0.008408221416175365, + 0.007245181128382683, + 0.016325481235980988, + 0.006488114129751921, + 0.011988634243607521, + 0.011032061651349068, + 0.006591372191905975, + 0.00983313750475645, + 0.00669165188446641, + 0.005577091593295336, + 0.008574851788580418, + 0.002888499991968274, + 0.0045059616677463055, + 0.008034485392272472, + 0.005991268437355757, + 0.010239759460091591, + 0.007856679148972034, + 0.0043276031501591206, + 0.015547131188213825, + 0.0035354909487068653, + 0.007741594221442938, + 0.011790863238275051, + 0.011070078238844872, + 0.011001266539096832, + 0.009425865486264229, + 0.006488163955509663, + 0.008271872065961361, + 0.006700277794152498, + 0.006632965058088303, + 0.03922732174396515, + 0.028890909627079964, + 0.035511817783117294, + 0.03339400887489319, + 0.013303238898515701, + 0.04072319343686104, + 0.015701130032539368, + 0.021313924342393875, + 0.030171707272529602, + 0.02004779316484928, + 0.019767653197050095, + 0.01830967143177986, + 0.011287792585790157, + 0.020209677517414093, + 0.01272402610629797, + 0.01326076965779066, + 0.018277538940310478, + 0.014258342795073986, + 0.01241401955485344, + 0.011961380019783974, + 0.005487250164151192, + 0.0070487381890416145, + 0.004643503110855818, + 0.005117221735417843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.021781159564852715 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076333265751600266 + ], + "opposite_logprob_tokens": [ + -4.882633209228516 + ], + "logprob_reward": -0.0076333265751600266, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076333265751600266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18009468913078308, + 0.1528797596693039, + 0.2224467545747757, + 0.07326783239841461, + 0.049918461591005325, + 0.04317976534366608, + 0.04799256846308708, + 0.05754712224006653, + 0.09106075018644333, + 0.08161233365535736 + ], + "credit_map": [ + 0.04583481326699257, + 0.015565186738967896, + 0.022008277475833893, + 0.010079987347126007, + 0.027509376406669617, + 0.019105033949017525, + 0.01135991420596838, + 0.02863207273185253, + 0.02981512062251568, + 0.014514953829348087, + 0.021638479083776474, + 0.012715895660221577, + 0.009530833922326565, + 0.01888769306242466, + 0.017414378002285957, + 0.02836238592863083, + 0.056291595101356506, + 0.03297831118106842, + 0.037663642317056656, + 0.014438518323004246, + 0.015963967889547348, + 0.019207486882805824, + 0.013237989507615566, + 0.03266523405909538, + 0.01556836161762476, + 0.006391411181539297, + 0.013180346228182316, + 0.0060151610523462296, + 0.005254033952951431, + 0.011477879248559475, + 0.00501913670450449, + 0.010361491702497005, + 0.011260977946221828, + 0.004502185620367527, + 0.009337526746094227, + 0.004367339890450239, + 0.0032751313410699368, + 0.005895288195461035, + 0.003157329512760043, + 0.00812267791479826, + 0.00820575188845396, + 0.004527931101620197, + 0.00790790468454361, + 0.003927356097847223, + 0.003634971333667636, + 0.006312352139502764, + 0.0030448411125689745, + 0.005618652328848839, + 0.00792128499597311, + 0.01030514296144247, + 0.007262234576046467, + 0.005490019917488098, + 0.0044745709747076035, + 0.004504980053752661, + 0.003334274748340249, + 0.004700056742876768, + 0.011715240776538849, + 0.0101347416639328, + 0.008621050044894218, + 0.006844533607363701, + 0.0046663484536111355, + 0.006651826202869415, + 0.0039713457226753235, + 0.004942029714584351, + 0.02035869099199772, + 0.016052257269620895, + 0.013748092576861382, + 0.01170345675200224, + 0.006319638807326555, + 0.008039072155952454, + 0.005395795684307814, + 0.0094437375664711, + 0.017555685713887215, + 0.016918353736400604, + 0.012199439108371735, + 0.011430801823735237, + 0.006535029038786888, + 0.00573272118344903, + 0.004317879676818848, + 0.0069224233739078045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0076333265751600266 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740333512425423 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.006740333512425423, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740333512425423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23755869269371033, + 0.17766448855400085, + 0.08262214809656143, + 0.05181729421019554, + 0.04411279782652855, + 0.03544902801513672, + 0.0591035895049572, + 0.09998618811368942, + 0.0894383043050766, + 0.12224743515253067 + ], + "credit_map": [ + 0.03609490022063255, + 0.01873658411204815, + 0.031274598091840744, + 0.030122127383947372, + 0.02613718993961811, + 0.04174112528562546, + 0.02042836882174015, + 0.03302379697561264, + 0.026360584422945976, + 0.017029395326972008, + 0.024400031194090843, + 0.021556910127401352, + 0.023568544536828995, + 0.02186661772429943, + 0.02012558840215206, + 0.022756820544600487, + 0.012127269059419632, + 0.006078494247049093, + 0.012407860718667507, + 0.01090333517640829, + 0.008802523836493492, + 0.014123090542852879, + 0.007309896871447563, + 0.010869676247239113, + 0.00904578436166048, + 0.004950362257659435, + 0.007829324342310429, + 0.0042735557071864605, + 0.004329079296439886, + 0.009877956472337246, + 0.004430601838976145, + 0.0070806266739964485, + 0.0065079047344625, + 0.004017780534923077, + 0.0076181055046617985, + 0.005671384744346142, + 0.005395868793129921, + 0.006622752640396357, + 0.002790727186948061, + 0.005488273221999407, + 0.005155094899237156, + 0.004113145638257265, + 0.00649001682177186, + 0.0032573179341852665, + 0.0028321368154138327, + 0.005899145267903805, + 0.0021528268698602915, + 0.0055493442341685295, + 0.01016524713486433, + 0.008934947662055492, + 0.008415631018579006, + 0.007917896844446659, + 0.005155768245458603, + 0.007977640256285667, + 0.004992533940821886, + 0.005543922074139118, + 0.016906525939702988, + 0.012621885165572166, + 0.015520608052611351, + 0.015795797109603882, + 0.006524647586047649, + 0.01596551015973091, + 0.007272631395608187, + 0.00937858410179615, + 0.014403257519006729, + 0.010064907371997833, + 0.013916502706706524, + 0.012816616334021091, + 0.006047472357749939, + 0.014957496896386147, + 0.007206825073808432, + 0.01002522837370634, + 0.016632549464702606, + 0.012240452691912651, + 0.0173597801476717, + 0.018966814503073692, + 0.010460197925567627, + 0.022453267127275467, + 0.010117360390722752, + 0.014017008244991302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.006740333512425423 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054947640746831894 + ], + "opposite_logprob_tokens": [ + -2.9299476146698 + ], + "logprob_reward": -0.054947640746831894, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054947640746831894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.188650444149971, + 0.12653189897537231, + 0.12662345170974731, + 0.08372264355421066, + 0.08447540551424026, + 0.10703393071889877, + 0.0922003760933876, + 0.10103930532932281, + 0.04206756874918938, + 0.047655005007982254 + ], + "credit_map": [ + 0.035432830452919006, + 0.010334660299122334, + 0.03423676639795303, + 0.019027093425393105, + 0.01683345064520836, + 0.025223374366760254, + 0.017630085349082947, + 0.0299321748316288, + 0.017467167228460312, + 0.013583295978605747, + 0.01483090128749609, + 0.016081932932138443, + 0.015094528906047344, + 0.0221561212092638, + 0.01308013778179884, + 0.01423780806362629, + 0.031204059720039368, + 0.01777142845094204, + 0.023334769532084465, + 0.008968290872871876, + 0.0074952528811991215, + 0.010485291481018066, + 0.008290366269648075, + 0.019073985517024994, + 0.023792419582605362, + 0.006883711088448763, + 0.017568569630384445, + 0.005748939234763384, + 0.004702359903603792, + 0.00813869759440422, + 0.005885033402591944, + 0.011002914048731327, + 0.01423108484596014, + 0.007341583725064993, + 0.013915257528424263, + 0.011846344918012619, + 0.008346576243638992, + 0.013931697234511375, + 0.0063666789792478085, + 0.008496182970702648, + 0.012991673313081264, + 0.007900932803750038, + 0.018066417425870895, + 0.017208002507686615, + 0.012372455559670925, + 0.015587631613016129, + 0.010662183165550232, + 0.012244638055562973, + 0.017400948330760002, + 0.011393604800105095, + 0.014191895723342896, + 0.012231807224452496, + 0.007496124133467674, + 0.011330625973641872, + 0.007660093251615763, + 0.010495278984308243, + 0.019531957805156708, + 0.017525017261505127, + 0.015312784351408482, + 0.012884660623967648, + 0.006419532001018524, + 0.013086696155369282, + 0.006366043817251921, + 0.009912610985338688, + 0.008220408111810684, + 0.005484089720994234, + 0.006316822022199631, + 0.005554677452892065, + 0.0029834602028131485, + 0.004712105728685856, + 0.0021890844218432903, + 0.006606918293982744, + 0.009028377011418343, + 0.008548499085009098, + 0.00625360244885087, + 0.0067125228233635426, + 0.003989032004028559, + 0.005837848410010338, + 0.002912482712417841, + 0.004372644703835249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.054947640746831894 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017133570509031415 + ], + "opposite_logprob_tokens": [ + -6.376713275909424 + ], + "logprob_reward": -0.0017133570509031415, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017133570509031415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26876044273376465, + 0.10604710131883621, + 0.08033278584480286, + 0.12720660865306854, + 0.0628933236002922, + 0.06158027425408363, + 0.054563093930482864, + 0.1095740869641304, + 0.08031624555587769, + 0.04872607812285423 + ], + "credit_map": [ + 0.05168013647198677, + 0.01585308462381363, + 0.04753851890563965, + 0.0230988971889019, + 0.022237831726670265, + 0.03821994736790657, + 0.023222029209136963, + 0.04690999537706375, + 0.016545644029974937, + 0.01288310345262289, + 0.012060610577464104, + 0.013936015777289867, + 0.010947879403829575, + 0.013367400504648685, + 0.011008398607373238, + 0.01529805175960064, + 0.01460699737071991, + 0.011463223956525326, + 0.013069297187030315, + 0.00852432381361723, + 0.006033509504050016, + 0.007722623646259308, + 0.004466106649488211, + 0.01444670744240284, + 0.026209402829408646, + 0.011453368701040745, + 0.0224843118339777, + 0.008762327954173088, + 0.009865055792033672, + 0.01685435138642788, + 0.00962523277848959, + 0.021952548995614052, + 0.010006380267441273, + 0.005303344689309597, + 0.01118761021643877, + 0.006965873762965202, + 0.006154134403914213, + 0.009127194061875343, + 0.005265836603939533, + 0.008882950060069561, + 0.014922426082193851, + 0.008208231069147587, + 0.012069632299244404, + 0.0060454122722148895, + 0.003367847530171275, + 0.006081266328692436, + 0.003640100359916687, + 0.00724536320194602, + 0.007568087428808212, + 0.011601037345826626, + 0.007767041213810444, + 0.006079789251089096, + 0.005156191531568766, + 0.006704259198158979, + 0.003592072520405054, + 0.006094614043831825, + 0.020372740924358368, + 0.01645496115088463, + 0.01703130081295967, + 0.015583798289299011, + 0.006092572119086981, + 0.01407595258206129, + 0.00742295291274786, + 0.012539803981781006, + 0.015941066667437553, + 0.01106009166687727, + 0.010745668783783913, + 0.00973241776227951, + 0.0065659936517477036, + 0.012718155048787594, + 0.005867266096174717, + 0.007685593795031309, + 0.01172590535134077, + 0.010080411098897457, + 0.006476931739598513, + 0.00453301751986146, + 0.0038643241859972477, + 0.005146983545273542, + 0.002087837317958474, + 0.004810667596757412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.0017133570509031415 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static wide-angle shot, captured from a gentle overhead angle, captures a small wooden sightseeing boat anchored in the middle of a broad, recently receding waterway. The waters are an opaque, milk chocolate-slurry hue, barely rippled by a gentle undercurrent that causes the boat to pitch back and forth with a slow, rhythmic lurch. The boat, painted a slightly worn creamy white with faded sky-blue trim along the edges, is empty save for a few simple rows of wooden benches. The entire image is bathed in a distinct soft-focus effect and a warm, golden glow, as if the afternoon sun is filtering through a mist, blurring the sharp edges of all objects. From the water to the boat's hull to the waterlogged riverbank in the distance, every surface possesses a hand-painted brushstroke and visible texture, as if it were a living, breathing illustration from a classic fairytale, creating an atmosphere of profound tranquility, innocence, and a touch of dreamlike childlike innocence.", + "seed": 15148, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_11/index_6402_seed_15148_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016628889250569046, + "vlm_logprob": -0.016628889250569046, + "vlm_yes_no_margin": 5.45, + "vlm_token_credit": -0.016628889250569046, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015714852139353752 + ], + "opposite_logprob_tokens": [ + -5.390714645385742 + ], + "logprob_reward": -0.015714852139353752, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.015714852139353752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20825061202049255, + 0.1364509016275406, + 0.0778982937335968, + 0.049749091267585754, + 0.06402947008609772, + 0.17122189700603485, + 0.12118586152791977, + 0.07890187948942184, + 0.0468512587249279, + 0.04546074941754341 + ], + "credit_map": [ + 0.023812592029571533, + 0.014554471708834171, + 0.043566491454839706, + 0.016571734100580215, + 0.03158698230981827, + 0.040664780884981155, + 0.014284149743616581, + 0.023209407925605774, + 0.016666732728481293, + 0.006848691496998072, + 0.03455424681305885, + 0.0101988660171628, + 0.015771783888339996, + 0.027279488742351532, + 0.008277717046439648, + 0.016853367909789085, + 0.007219193037599325, + 0.0052190073765814304, + 0.012400148436427116, + 0.004601431079208851, + 0.013674579560756683, + 0.02124886028468609, + 0.004905486013740301, + 0.008629589341580868, + 0.0074406336061656475, + 0.003712642239406705, + 0.008073464035987854, + 0.0029297953005880117, + 0.006683389190584421, + 0.010739188641309738, + 0.003813698422163725, + 0.0063562775030732155, + 0.010810072533786297, + 0.004040271509438753, + 0.013426811434328556, + 0.0043345182202756405, + 0.009673170745372772, + 0.010165993124246597, + 0.004414907656610012, + 0.007163725793361664, + 0.018409883603453636, + 0.01774761639535427, + 0.03314734250307083, + 0.013648941181600094, + 0.02869924157857895, + 0.030950890854001045, + 0.014983736909925938, + 0.013634247705340385, + 0.029540345072746277, + 0.009619948454201221, + 0.01620025373995304, + 0.007146271876990795, + 0.01861456222832203, + 0.02034299261868, + 0.008723481558263302, + 0.01099800318479538, + 0.008770933374762535, + 0.005268181208521128, + 0.013029645197093487, + 0.005773924291133881, + 0.015319675207138062, + 0.01828920841217041, + 0.00513896020129323, + 0.007311343681067228, + 0.005203118547797203, + 0.006206565536558628, + 0.006847224663943052, + 0.003912258893251419, + 0.005619902163743973, + 0.009331876412034035, + 0.004101324826478958, + 0.005628984421491623, + 0.005362001247704029, + 0.006381858605891466, + 0.006704611238092184, + 0.004037474747747183, + 0.005157384555786848, + 0.009681344032287598, + 0.004031877964735031, + 0.004104193300008774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.015714852139353752 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011857867939397693 + ], + "opposite_logprob_tokens": [ + -6.751185894012451 + ], + "logprob_reward": -0.0011857867939397693, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011857867939397693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22898000478744507, + 0.10290931910276413, + 0.07173986732959747, + 0.0722866803407669, + 0.021014539524912834, + 0.16695129871368408, + 0.10819579660892487, + 0.031229348853230476, + 0.05730617791414261, + 0.13938695192337036 + ], + "credit_map": [ + 0.025175461545586586, + 0.012600559741258621, + 0.0423266738653183, + 0.014016793109476566, + 0.032410115003585815, + 0.056573349982500076, + 0.021771874278783798, + 0.024105189368128777, + 0.013553435914218426, + 0.006322134286165237, + 0.024039776995778084, + 0.00874406285583973, + 0.015612021088600159, + 0.01912507601082325, + 0.006409581284970045, + 0.009103231132030487, + 0.00904153659939766, + 0.004636525642126799, + 0.011177782900631428, + 0.0037298952229321003, + 0.011347701773047447, + 0.017885921522974968, + 0.005633918102830648, + 0.008286583237349987, + 0.007469242438673973, + 0.005119585432112217, + 0.011749443598091602, + 0.005958868190646172, + 0.013516008853912354, + 0.015124455094337463, + 0.006074860692024231, + 0.007274223957210779, + 0.002406933344900608, + 0.0020286617800593376, + 0.0033593480475246906, + 0.002344150561839342, + 0.002805187599733472, + 0.0036323266103863716, + 0.002104972954839468, + 0.002332959556952119, + 0.016221538186073303, + 0.014624453149735928, + 0.03158184513449669, + 0.012679515406489372, + 0.02995966374874115, + 0.03323501721024513, + 0.014233051799237728, + 0.014416207559406757, + 0.018083373084664345, + 0.006836907938122749, + 0.02335636503994465, + 0.007467405404895544, + 0.015915147960186005, + 0.020574724301695824, + 0.00790497101843357, + 0.008056906051933765, + 0.00652870861813426, + 0.002200379967689514, + 0.00428729085251689, + 0.0019189096055924892, + 0.003802236169576645, + 0.006880562286823988, + 0.00252523017115891, + 0.0030860311817377806, + 0.009495554491877556, + 0.0036034800577908754, + 0.008847096003592014, + 0.0033202667254954576, + 0.008822141215205193, + 0.011995387263596058, + 0.004624300170689821, + 0.006597952917218208, + 0.017002485692501068, + 0.010312408208847046, + 0.024930810555815697, + 0.0088911522179842, + 0.02453753538429737, + 0.030356615781784058, + 0.009156807325780392, + 0.01419912837445736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.0011857867939397693 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002493488835170865 + ], + "opposite_logprob_tokens": [ + -6.002493381500244 + ], + "logprob_reward": -0.002493488835170865, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002493488835170865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1864337921142578, + 0.12047339230775833, + 0.04955907538533211, + 0.14867745339870453, + 0.10592257231473923, + 0.12181516736745834, + 0.10051269084215164, + 0.07928283512592316, + 0.0478217713534832, + 0.03950124979019165 + ], + "credit_map": [ + 0.0205535888671875, + 0.012742262333631516, + 0.04018665850162506, + 0.014129796996712685, + 0.025197917595505714, + 0.04403352364897728, + 0.011101375333964825, + 0.018488675355911255, + 0.011732451617717743, + 0.008402466773986816, + 0.02567024901509285, + 0.0111854774877429, + 0.017877360805869102, + 0.026044918224215508, + 0.007979830726981163, + 0.01158064417541027, + 0.004588827025145292, + 0.003867501625791192, + 0.006938421167433262, + 0.0038378769531846046, + 0.007836749777197838, + 0.011408079415559769, + 0.005412505939602852, + 0.005669116973876953, + 0.014848154038190842, + 0.008890838362276554, + 0.02489025518298149, + 0.01396345067769289, + 0.028304114937782288, + 0.03232160583138466, + 0.01236476469784975, + 0.013094276189804077, + 0.013517591170966625, + 0.006352303549647331, + 0.020230596885085106, + 0.008166163228452206, + 0.019119611009955406, + 0.02110750786960125, + 0.008523245342075825, + 0.008905550464987755, + 0.011542538180947304, + 0.011769459582865238, + 0.02379206381738186, + 0.009846899658441544, + 0.021253306418657303, + 0.022788600996136665, + 0.011090809479355812, + 0.009731498546898365, + 0.02914302982389927, + 0.008562879636883736, + 0.015856482088565826, + 0.004991564899682999, + 0.013626284897327423, + 0.01511453278362751, + 0.00591184850782156, + 0.007306083105504513, + 0.012861553579568863, + 0.004596720915287733, + 0.01147173997014761, + 0.0036846029106527567, + 0.012745559215545654, + 0.015642568469047546, + 0.006237587425857782, + 0.012042512185871601, + 0.006294942926615477, + 0.004566536284983158, + 0.006811908446252346, + 0.0038544260896742344, + 0.0062816995196044445, + 0.00951027125120163, + 0.004912242293357849, + 0.005589745007455349, + 0.004374790471047163, + 0.0037460075691342354, + 0.00798110943287611, + 0.002464448567479849, + 0.004860754124820232, + 0.009229574352502823, + 0.0032823006622493267, + 0.0035622664727270603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.002493488835170865 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203708052635193 + ], + "opposite_logprob_tokens": [ + -2.8120369911193848 + ], + "logprob_reward": -0.06203708052635193, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203708052635193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1810811311006546, + 0.1768331080675125, + 0.06874576956033707, + 0.02526056207716465, + 0.06121465191245079, + 0.3010486662387848, + 0.07924564927816391, + 0.04847842827439308, + 0.026601256802678108, + 0.0314907431602478 + ], + "credit_map": [ + 0.016936618834733963, + 0.010850421153008938, + 0.03892439976334572, + 0.012315968051552773, + 0.026466095820069313, + 0.046732157468795776, + 0.0123493866994977, + 0.016506073996424675, + 0.019217882305383682, + 0.010255279019474983, + 0.04060758277773857, + 0.015354673378169537, + 0.02803131751716137, + 0.03729956969618797, + 0.011761668138206005, + 0.014305139891803265, + 0.008737500756978989, + 0.004723528400063515, + 0.012060856446623802, + 0.006103078834712505, + 0.010671742260456085, + 0.015115033835172653, + 0.004900328814983368, + 0.0064337062649428844, + 0.002468475140631199, + 0.0022450468968600035, + 0.0032513791229575872, + 0.002122608944773674, + 0.004344433080404997, + 0.006770461797714233, + 0.001784426742233336, + 0.002273731166496873, + 0.008719225414097309, + 0.004218855407088995, + 0.013014290481805801, + 0.004812613595277071, + 0.009969783015549183, + 0.010453876107931137, + 0.00441691093146801, + 0.005609097424894571, + 0.05969636142253876, + 0.02384781837463379, + 0.07393424957990646, + 0.01874212734401226, + 0.04554755613207817, + 0.043849993497133255, + 0.018654678016901016, + 0.016775861382484436, + 0.02326437085866928, + 0.006817292887717485, + 0.013156138360500336, + 0.0039681484922766685, + 0.00976897869259119, + 0.01089703943580389, + 0.005142472684383392, + 0.006231209728866816, + 0.005739574320614338, + 0.0036040805280208588, + 0.009643069468438625, + 0.0034699507523328066, + 0.007951159030199051, + 0.009743358939886093, + 0.0037227082066237926, + 0.004604525864124298, + 0.005424140021204948, + 0.0026085060089826584, + 0.003844824153929949, + 0.0019324093591421843, + 0.003540990874171257, + 0.004387552849948406, + 0.0021650756243616343, + 0.002697758376598358, + 0.0034792751539498568, + 0.004044216126203537, + 0.0051286532543599606, + 0.002305809408426285, + 0.004379145801067352, + 0.0060382564552128315, + 0.002896175254136324, + 0.0032192119397222996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.06203708052635193 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017132379580289125 + ], + "opposite_logprob_tokens": [ + -6.376713275909424 + ], + "logprob_reward": -0.0017132379580289125, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017132379580289125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23014307022094727, + 0.06233265623450279, + 0.040034424513578415, + 0.028059417381882668, + 0.025922812521457672, + 0.23032929003238678, + 0.2506825029850006, + 0.04323264956474304, + 0.05831841006875038, + 0.030944768339395523 + ], + "credit_map": [ + 0.02270558662712574, + 0.012826662510633469, + 0.03700399398803711, + 0.017756231129169464, + 0.031981460750103, + 0.06459163874387741, + 0.014792680740356445, + 0.02848481759428978, + 0.009149698540568352, + 0.005107296165078878, + 0.013238674029707909, + 0.004879442043602467, + 0.00745224766433239, + 0.011873694136738777, + 0.004758632276207209, + 0.0058729681186378, + 0.004477903246879578, + 0.0037134012673050165, + 0.007619145326316357, + 0.003243905957788229, + 0.006545193027704954, + 0.00761669734492898, + 0.0027739843353629112, + 0.004044197499752045, + 0.0027144779451191425, + 0.0025123516097664833, + 0.004451208282262087, + 0.0024167350493371487, + 0.003974036313593388, + 0.005451187025755644, + 0.0023794001899659634, + 0.004160021431744099, + 0.002614813158288598, + 0.0031837199348956347, + 0.003394267987459898, + 0.0027884445153176785, + 0.003968060947954655, + 0.004622693173587322, + 0.002900195773690939, + 0.002450617030262947, + 0.021896298974752426, + 0.022275427356362343, + 0.03642259165644646, + 0.018102850764989853, + 0.04201154783368111, + 0.04604072868824005, + 0.021338196471333504, + 0.02224164642393589, + 0.06906212866306305, + 0.01989113539457321, + 0.039698753505945206, + 0.011951274238526821, + 0.036485932767391205, + 0.03739320859313011, + 0.014552915468811989, + 0.021647173911333084, + 0.008002226240932941, + 0.0029933778569102287, + 0.00908864289522171, + 0.0035467902198433876, + 0.006291622761636972, + 0.006943012587726116, + 0.0035068397410213947, + 0.0028601367957890034, + 0.006693198811262846, + 0.005444423761218786, + 0.009591285139322281, + 0.00444016233086586, + 0.00800845492631197, + 0.011496976017951965, + 0.005719918757677078, + 0.006923992186784744, + 0.002838113810867071, + 0.0028124633245170116, + 0.004779268521815538, + 0.001849263091571629, + 0.0040952288545668125, + 0.00764709385111928, + 0.0023813466541469097, + 0.0045419903472065926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.0017132379580289125 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static wide-angle shot, captured from a gentle overhead angle, captures a small wooden sightseeing boat anchored in the middle of a broad, recently receding waterway. The waters are an opaque, milk chocolate-slurry hue, barely rippled by a gentle undercurrent that causes the boat to pitch back and forth with a slow, rhythmic lurch. The boat, painted a slightly worn creamy white with faded sky-blue trim along the edges, is empty save for a few simple rows of wooden benches. The entire image is bathed in a distinct soft-focus effect and a warm, golden glow, as if the afternoon sun is filtering through a mist, blurring the sharp edges of all objects. From the water to the boat's hull to the waterlogged riverbank in the distance, every surface possesses a hand-painted brushstroke and visible texture, as if it were a living, breathing illustration from a classic fairytale, creating an atmosphere of profound tranquility, innocence, and a touch of dreamlike childlike innocence.", + "seed": 15149, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_11/index_6402_seed_15149_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01993310023099184, + "vlm_logprob": -0.01993310023099184, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.01993310023099184, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013260742649435997 + ], + "opposite_logprob_tokens": [ + -5.388260841369629 + ], + "logprob_reward": -0.013260742649435997, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.013260742649435997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1131918802857399, + 0.0503782294690609, + 0.04054554924368858, + 0.03869949281215668, + 0.10960663110017776, + 0.03866000846028328, + 0.05643925443291664, + 0.10537567734718323, + 0.40543147921562195, + 0.041671812534332275 + ], + "credit_map": [ + 0.012623702175915241, + 0.011461049318313599, + 0.0085300849750638, + 0.015005865134298801, + 0.02083457261323929, + 0.022762274369597435, + 0.00938470009714365, + 0.012589627876877785, + 0.004528694786131382, + 0.0027914901729673147, + 0.003367531346157193, + 0.002434684196487069, + 0.014359785243868828, + 0.015341314487159252, + 0.0034227962605655193, + 0.0041319336742162704, + 0.003941508010029793, + 0.003591186134144664, + 0.002621979219838977, + 0.0029871021397411823, + 0.0074455756694078445, + 0.010386629030108452, + 0.0044438778422772884, + 0.005127688869833946, + 0.006068514194339514, + 0.0036529963836073875, + 0.0037324009463191032, + 0.002825675066560507, + 0.00535878399387002, + 0.009059963747859001, + 0.0032464563846588135, + 0.0047547039575874805, + 0.005403981544077396, + 0.005514348391443491, + 0.007038621697574854, + 0.0062446813099086285, + 0.03225654363632202, + 0.03203820437192917, + 0.012169444002211094, + 0.008940810337662697, + 0.006384784821420908, + 0.0029206527397036552, + 0.005503901280462742, + 0.0025531852152198553, + 0.004506631754338741, + 0.008405527099967003, + 0.0033263328950852156, + 0.005058993119746447, + 0.006420377641916275, + 0.005238498095422983, + 0.006676145829260349, + 0.0035079794470220804, + 0.009593323804438114, + 0.011675736866891384, + 0.005254513584077358, + 0.008072679862380028, + 0.006331777665764093, + 0.005662436131387949, + 0.015356362797319889, + 0.006262617185711861, + 0.03494846075773239, + 0.01772761344909668, + 0.007340741343796253, + 0.011745668947696686, + 0.010405872948467731, + 0.011105790734291077, + 0.028120536357164383, + 0.0059511722065508366, + 0.16638733446598053, + 0.16637855768203735, + 0.009236692450940609, + 0.00784552562981844, + 0.0055528674274683, + 0.003204652341082692, + 0.004338952247053385, + 0.0026250937953591347, + 0.008586842566728592, + 0.008170307613909245, + 0.0051424540579319, + 0.004050643183290958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.013260742649435997 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541121803224087 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541121803224087, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541121803224087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18310819566249847, + 0.08334918320178986, + 0.09076504409313202, + 0.07606916129589081, + 0.07892107963562012, + 0.05899752303957939, + 0.03191915526986122, + 0.06160905957221985, + 0.30563169717788696, + 0.029629912227392197 + ], + "credit_map": [ + 0.021936561912298203, + 0.020525604486465454, + 0.01561070792376995, + 0.0215444453060627, + 0.03520316258072853, + 0.031415536999702454, + 0.015830250456929207, + 0.021041948348283768, + 0.0049820756539702415, + 0.004373075906187296, + 0.004127432592213154, + 0.00405878247693181, + 0.02757667750120163, + 0.027530621737241745, + 0.004789676051586866, + 0.0059108431451022625, + 0.007778285536915064, + 0.008690614253282547, + 0.006252932362258434, + 0.007125155534595251, + 0.01323946937918663, + 0.028023825958371162, + 0.009334016591310501, + 0.010320750065147877, + 0.00941664632409811, + 0.00595434382557869, + 0.007328707724809647, + 0.0044849892146885395, + 0.010391569696366787, + 0.02638266421854496, + 0.004936076235026121, + 0.007174167316406965, + 0.007802240084856749, + 0.005763830151408911, + 0.007389432284981012, + 0.004408211447298527, + 0.01615704409778118, + 0.020183607935905457, + 0.009100493043661118, + 0.008116237819194794, + 0.006166646722704172, + 0.006115909665822983, + 0.0057313512079417706, + 0.0041173286736011505, + 0.008579478599131107, + 0.010476009920239449, + 0.005102601367980242, + 0.012708203867077827, + 0.0030180523172020912, + 0.0031232209876179695, + 0.00360258505679667, + 0.0017187439370900393, + 0.009334900416433811, + 0.00497093889862299, + 0.0026334694121032953, + 0.0035172472707927227, + 0.003990317229181528, + 0.0030404669232666492, + 0.014418277889490128, + 0.0049319397658109665, + 0.0052207582630217075, + 0.018453005701303482, + 0.0038858260959386826, + 0.007668468169867992, + 0.009998385794460773, + 0.008051220327615738, + 0.02079901471734047, + 0.004487035796046257, + 0.12571963667869568, + 0.12377407401800156, + 0.006350350100547075, + 0.006451992783695459, + 0.00409315712749958, + 0.0028937270399183035, + 0.004007045645266771, + 0.0025390617083758116, + 0.006146855652332306, + 0.0034021318424493074, + 0.003853732021525502, + 0.002694206079468131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.012541121803224087 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181918703019619 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.0181918703019619, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181918703019619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14095565676689148, + 0.09126321971416473, + 0.04555889219045639, + 0.04665015637874603, + 0.04148845002055168, + 0.048402462154626846, + 0.01519038062542677, + 0.1419162005186081, + 0.404131680727005, + 0.024442927911877632 + ], + "credit_map": [ + 0.014684011228382587, + 0.018823057413101196, + 0.015452600084245205, + 0.020769832655787468, + 0.020258428528904915, + 0.02422344870865345, + 0.013175053521990776, + 0.013569220900535583, + 0.008877206593751907, + 0.007403820287436247, + 0.0076521229930222034, + 0.005829192698001862, + 0.020112255588173866, + 0.0230711717158556, + 0.007565722335129976, + 0.010751724243164062, + 0.004232882522046566, + 0.004227020312100649, + 0.003035495989024639, + 0.0036020579282194376, + 0.007667896803468466, + 0.011972950771450996, + 0.006155854556709528, + 0.00466473214328289, + 0.005120715592056513, + 0.005546504631638527, + 0.004058893769979477, + 0.005187701433897018, + 0.005894847214221954, + 0.008516842499375343, + 0.0058041722513735294, + 0.006520475260913372, + 0.005036941263824701, + 0.004141550976783037, + 0.004414803814142942, + 0.0031535921152681112, + 0.006854026112705469, + 0.008265468291938305, + 0.005436398088932037, + 0.004185669124126434, + 0.004282205365598202, + 0.005254063289612532, + 0.004873765632510185, + 0.003428028430789709, + 0.006952954921871424, + 0.009046416729688644, + 0.004459885880351067, + 0.010105138644576073, + 0.0012570470571517944, + 0.0015554444398730993, + 0.0016079188790172338, + 0.0012319618836045265, + 0.0023656520061194897, + 0.0022186737041920424, + 0.0016742226434871554, + 0.003279459895566106, + 0.005161282140761614, + 0.00582771236076951, + 0.014789278618991375, + 0.009595874696969986, + 0.040909431874752045, + 0.04959418997168541, + 0.005988602060824633, + 0.010049832984805107, + 0.014929350465536118, + 0.011386417783796787, + 0.02691798284649849, + 0.007712341379374266, + 0.16622428596019745, + 0.1549021601676941, + 0.013213247992098331, + 0.00884584803134203, + 0.002792938146740198, + 0.0017195057589560747, + 0.0034595320466905832, + 0.002552206162363291, + 0.003182466607540846, + 0.0041151405312120914, + 0.00315521820448339, + 0.0034659220837056637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0181918703019619 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495486035943031 + ], + "opposite_logprob_tokens": [ + -2.929954767227173 + ], + "logprob_reward": -0.05495486035943031, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495486035943031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03660723939538002, + 0.011622973717749119, + 0.011638844385743141, + 0.016786403954029083, + 0.01834103651344776, + 0.014122780412435532, + 0.01389674935489893, + 0.07927605509757996, + 0.7880390882492065, + 0.009668814018368721 + ], + "credit_map": [ + 0.004746152553707361, + 0.003937483765184879, + 0.002572806319221854, + 0.004065411631017923, + 0.00992229487746954, + 0.005734117701649666, + 0.0027197266463190317, + 0.002909248461946845, + 0.0020959360990673304, + 0.0010642935521900654, + 0.001105789327993989, + 0.0008837654604576528, + 0.0018855908419936895, + 0.0021507495548576117, + 0.0012220579665154219, + 0.0012147912057116628, + 0.0013124713441357017, + 0.0010064303642138839, + 0.0007828166126273572, + 0.001045303069986403, + 0.002330708084627986, + 0.0025959324557334185, + 0.001238813973031938, + 0.0013263686560094357, + 0.0015931858215481043, + 0.002077183686196804, + 0.0015158846508711576, + 0.002063048304989934, + 0.002088853856548667, + 0.0026741193141788244, + 0.002926073269918561, + 0.0018480545841157436, + 0.0018218039767816663, + 0.0018101928289979696, + 0.00195212266407907, + 0.0014843882527202368, + 0.003606432117521763, + 0.00382807245478034, + 0.0018636323511600494, + 0.001974391285330057, + 0.0015642417129129171, + 0.0011928791645914316, + 0.001381373731419444, + 0.0008060786058194935, + 0.002520643174648285, + 0.002329892013221979, + 0.001463656430132687, + 0.002864015055820346, + 0.0015426086029037833, + 0.0014100992120802402, + 0.0014121042331680655, + 0.0009015630348585546, + 0.0036166219506412745, + 0.00197470816783607, + 0.0016045955708250403, + 0.0014344480587169528, + 0.002219832967966795, + 0.0025904588401317596, + 0.007051266264170408, + 0.001735235913656652, + 0.03308326378464699, + 0.02780890092253685, + 0.002306208945810795, + 0.0024808901362121105, + 0.021372828632593155, + 0.01740279234945774, + 0.054534126073122025, + 0.0064040725119411945, + 0.3452325463294983, + 0.32114166021347046, + 0.011436732485890388, + 0.010514356195926666, + 0.0012344267452135682, + 0.0008320416673086584, + 0.001153067103587091, + 0.0007282623555511236, + 0.001775080687366426, + 0.0013914034934714437, + 0.0012792161433026195, + 0.0012753154151141644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.05495486035943031 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007169060409069061 + ], + "opposite_logprob_tokens": [ + -7.250716686248779 + ], + "logprob_reward": -0.0007169060409069061, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007169060409069061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15929049253463745, + 0.039706092327833176, + 0.046685222536325455, + 0.021574728190898895, + 0.04635181650519371, + 0.023390548303723335, + 0.030155064538121223, + 0.07116010040044785, + 0.5493849515914917, + 0.012301045469939709 + ], + "credit_map": [ + 0.02081095241010189, + 0.018913818523287773, + 0.01784827560186386, + 0.022561734542250633, + 0.026995180174708366, + 0.022439977154135704, + 0.013992932625114918, + 0.015727614983916283, + 0.0040320358239114285, + 0.003367415862157941, + 0.003960893955081701, + 0.003318934701383114, + 0.0063519906252622604, + 0.008096335455775261, + 0.0037014775443822145, + 0.006877005100250244, + 0.006149267312139273, + 0.003455061698332429, + 0.0048113674856722355, + 0.002406694460660219, + 0.007579564116895199, + 0.01287683192640543, + 0.003376976354047656, + 0.0060294573195278645, + 0.001927835401147604, + 0.0020683978218585253, + 0.0018808458698913455, + 0.0024684604723006487, + 0.0027890491764992476, + 0.00397188775241375, + 0.003232775256037712, + 0.0032354737631976604, + 0.0028606869745999575, + 0.0027936683036386967, + 0.003171922406181693, + 0.0020057163201272488, + 0.015871664509177208, + 0.013876466080546379, + 0.0027878256514668465, + 0.002983858808875084, + 0.0026229547802358866, + 0.002126730978488922, + 0.002801567316055298, + 0.0018970236415043473, + 0.004143822472542524, + 0.0029515009373426437, + 0.0024747424758970737, + 0.004372202791273594, + 0.003838764736428857, + 0.0030394038185477257, + 0.003563365899026394, + 0.0023939504753798246, + 0.004586738999933004, + 0.006115771364420652, + 0.003181392326951027, + 0.003435672726482153, + 0.0028311433270573616, + 0.003341044532135129, + 0.005975141189992428, + 0.004200618248432875, + 0.021719306707382202, + 0.02328086644411087, + 0.0037032661493867636, + 0.006108710076659918, + 0.016691192984580994, + 0.012630813755095005, + 0.037265632301568985, + 0.004745937883853912, + 0.23829354345798492, + 0.22366276383399963, + 0.008269350044429302, + 0.007825617678463459, + 0.0018601121846586466, + 0.0011637499555945396, + 0.0015925720799714327, + 0.0006983859348110855, + 0.002237537642940879, + 0.0014501382829621434, + 0.0014894533669576049, + 0.0018090953817591071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.0007169060409069061 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static wide-angle shot, captured from a gentle overhead angle, captures a small wooden sightseeing boat anchored in the middle of a broad, recently receding waterway. The waters are an opaque, milk chocolate-slurry hue, barely rippled by a gentle undercurrent that causes the boat to pitch back and forth with a slow, rhythmic lurch. The boat, painted a slightly worn creamy white with faded sky-blue trim along the edges, is empty save for a few simple rows of wooden benches. The entire image is bathed in a distinct soft-focus effect and a warm, golden glow, as if the afternoon sun is filtering through a mist, blurring the sharp edges of all objects. From the water to the boat's hull to the waterlogged riverbank in the distance, every surface possesses a hand-painted brushstroke and visible texture, as if it were a living, breathing illustration from a classic fairytale, creating an atmosphere of profound tranquility, innocence, and a touch of dreamlike childlike innocence.", + "seed": 15150, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_11/index_6402_seed_15150_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.019303228659555315, + "vlm_logprob": -0.019303228659555315, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.019303228659555315, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024101369082927704 + ], + "opposite_logprob_tokens": [ + -5.149101257324219 + ], + "logprob_reward": -0.024101369082927704, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.024101369082927704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13829419016838074, + 0.04032785817980766, + 0.039747923612594604, + 0.12411367893218994, + 0.032220933586359024, + 0.47003084421157837, + 0.03470246493816376, + 0.02496611699461937, + 0.049698516726493835, + 0.04589741677045822 + ], + "credit_map": [ + 0.02410793863236904, + 0.012705730274319649, + 0.03288310766220093, + 0.013010858558118343, + 0.018364980816841125, + 0.012068207375705242, + 0.007542831357568502, + 0.017610548064112663, + 0.0070325518026947975, + 0.002402891870588064, + 0.01234346441924572, + 0.0027521823067218065, + 0.0032025212422013283, + 0.0029580551199615, + 0.0022487398236989975, + 0.007387451361864805, + 0.009786290116608143, + 0.003184638684615493, + 0.006570966448634863, + 0.0028632173780351877, + 0.006176795810461044, + 0.004020486958324909, + 0.0021899184212088585, + 0.004955609329044819, + 0.02282698266208172, + 0.008069261908531189, + 0.03148190304636955, + 0.005956423003226519, + 0.012934645637869835, + 0.01064598374068737, + 0.005131227895617485, + 0.02706725522875786, + 0.004441488534212112, + 0.002443769946694374, + 0.006952175870537758, + 0.001592649961821735, + 0.006393732037395239, + 0.0030498208943754435, + 0.001585809513926506, + 0.005761486478149891, + 0.041200824081897736, + 0.012145843356847763, + 0.19712133705615997, + 0.012177621014416218, + 0.016052082180976868, + 0.01980898343026638, + 0.00838465616106987, + 0.16313950717449188, + 0.007017976604402065, + 0.003349914448335767, + 0.0065254224464297295, + 0.0031601404771208763, + 0.004100619349628687, + 0.003911806736141443, + 0.0020446989219635725, + 0.004591888748109341, + 0.005024613346904516, + 0.0019879951141774654, + 0.004899583291262388, + 0.0021478799171745777, + 0.004278970416635275, + 0.002179708331823349, + 0.0016054738080129027, + 0.002841891720890999, + 0.006369618233293295, + 0.008557879365980625, + 0.005432176403701305, + 0.004660570528358221, + 0.01041420642286539, + 0.005111620761454105, + 0.0035819062031805515, + 0.005570537876337767, + 0.004928547888994217, + 0.006874703802168369, + 0.006922959815710783, + 0.003340029390528798, + 0.006043784320354462, + 0.008395534940063953, + 0.00436409329995513, + 0.005027766339480877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.024101369082927704 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024931321386247873 + ], + "opposite_logprob_tokens": [ + -6.002492904663086 + ], + "logprob_reward": -0.0024931321386247873, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024931321386247873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2418530136346817, + 0.07088931649923325, + 0.0804433822631836, + 0.03922223672270775, + 0.15796996653079987, + 0.2926599383354187, + 0.03666844964027405, + 0.04109647125005722, + 0.021490519866347313, + 0.017706777900457382 + ], + "credit_map": [ + 0.05124993994832039, + 0.018641313537955284, + 0.04751889407634735, + 0.017245039343833923, + 0.01912456750869751, + 0.029190296307206154, + 0.013360250741243362, + 0.04552268236875534, + 0.01623324677348137, + 0.004948807880282402, + 0.013112870045006275, + 0.004522654693573713, + 0.00641675153747201, + 0.008875992149114609, + 0.00418804120272398, + 0.012590949423611164, + 0.013814177364110947, + 0.007130485028028488, + 0.012652760371565819, + 0.0064960685558617115, + 0.0076996637508273125, + 0.008540620096027851, + 0.005063536576926708, + 0.019046058878302574, + 0.006286307703703642, + 0.0025167674757540226, + 0.010024633258581161, + 0.002068150322884321, + 0.004908155649900436, + 0.004707569722086191, + 0.0016575874760746956, + 0.007053060922771692, + 0.037270307540893555, + 0.011407033540308475, + 0.040069133043289185, + 0.006473991088569164, + 0.010433420538902283, + 0.011705465614795685, + 0.005773349665105343, + 0.034837253391742706, + 0.026608549058437347, + 0.007597830146551132, + 0.1222815066576004, + 0.007885647006332874, + 0.010035037994384766, + 0.01266513578593731, + 0.005651248153299093, + 0.09993495792150497, + 0.01070798747241497, + 0.0033454212825745344, + 0.00701008178293705, + 0.0021717085037380457, + 0.0033536083064973354, + 0.003376896260306239, + 0.0017560248961672187, + 0.004946716129779816, + 0.00795498676598072, + 0.004405967425554991, + 0.006447603460401297, + 0.003072422696277499, + 0.008742029778659344, + 0.003581585129722953, + 0.0025977296754717827, + 0.00429414352402091, + 0.0029571703635156155, + 0.004511972889304161, + 0.0029218397103250027, + 0.0018622635398060083, + 0.0029773840215057135, + 0.002165477955713868, + 0.0016549803549423814, + 0.002439428586512804, + 0.0027119803708046675, + 0.0027065954636782408, + 0.0038090685848146677, + 0.000865078647620976, + 0.001915998524054885, + 0.002235738793388009, + 0.0011184564791619778, + 0.002343857428058982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0024931321386247873 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951303523033857 + ], + "opposite_logprob_tokens": [ + -5.130951404571533 + ], + "logprob_reward": -0.005951303523033857, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951303523033857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09066710621118546, + 0.062226563692092896, + 0.07251055538654327, + 0.12481257319450378, + 0.3136175870895386, + 0.16175512969493866, + 0.11118993163108826, + 0.02584921568632126, + 0.012477656826376915, + 0.024893756955862045 + ], + "credit_map": [ + 0.014972057193517685, + 0.00892029982060194, + 0.015122635290026665, + 0.007754581514745951, + 0.010890374891459942, + 0.015437108464539051, + 0.005135004408657551, + 0.012435038574039936, + 0.008615164086222649, + 0.005339063238352537, + 0.019379498437047005, + 0.0033611871767789125, + 0.006064544897526503, + 0.007061983924359083, + 0.0025338090490549803, + 0.009871304035186768, + 0.006731156259775162, + 0.005352786276489496, + 0.021143309772014618, + 0.005097517743706703, + 0.009734098799526691, + 0.0065847053192555904, + 0.003486274043098092, + 0.01438069622963667, + 0.02683338150382042, + 0.010325171053409576, + 0.021508216857910156, + 0.008615567348897457, + 0.01053640991449356, + 0.016618773341178894, + 0.005843059159815311, + 0.024531977251172066, + 0.0647963136434555, + 0.020930519327521324, + 0.08349812030792236, + 0.012246937490999699, + 0.020543204620480537, + 0.02113393321633339, + 0.011980964802205563, + 0.0784875750541687, + 0.014525827020406723, + 0.006879035849124193, + 0.060559503734111786, + 0.0048175412230193615, + 0.011103611439466476, + 0.009234638884663582, + 0.0034482788760215044, + 0.05118666961789131, + 0.011530832387506962, + 0.012765951454639435, + 0.01166810467839241, + 0.007902675308287144, + 0.03338097408413887, + 0.01467097271233797, + 0.007256247103214264, + 0.012014170177280903, + 0.003216883633285761, + 0.003537805052474141, + 0.003041463904082775, + 0.0024482705630362034, + 0.006260666996240616, + 0.003284392412751913, + 0.001498425379395485, + 0.0025613047182559967, + 0.0019593064207583666, + 0.001810201327316463, + 0.0021735904738307, + 0.0008595080580562353, + 0.0013022570637986064, + 0.0020692460238933563, + 0.0007916328613646328, + 0.0015119139570742846, + 0.0039226035587489605, + 0.0038098846562206745, + 0.004306950606405735, + 0.0013684496516361833, + 0.0030675886664539576, + 0.0036444137804210186, + 0.0014194867108017206, + 0.0033543757162988186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.005951303523033857 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062031254172325134 + ], + "opposite_logprob_tokens": [ + -2.8120312690734863 + ], + "logprob_reward": -0.062031254172325134, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062031254172325134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06291709095239639, + 0.0245197881013155, + 0.013566374778747559, + 0.13473181426525116, + 0.048049937933683395, + 0.6274356842041016, + 0.04203416407108307, + 0.010889388620853424, + 0.011409143917262554, + 0.024446576833724976 + ], + "credit_map": [ + 0.015082762576639652, + 0.0051035494543612, + 0.009554089047014713, + 0.005415330175310373, + 0.005258862394839525, + 0.007849286310374737, + 0.0036286201793700457, + 0.011024591512978077, + 0.005978694185614586, + 0.0017446603160351515, + 0.006139776203781366, + 0.0012761641992256045, + 0.0016900122864171863, + 0.002793594729155302, + 0.0012129490496590734, + 0.0036839372478425503, + 0.003311588428914547, + 0.0009429029887542129, + 0.002917712554335594, + 0.0009121124749071896, + 0.001348724472336471, + 0.0010734040988609195, + 0.0005653605330735445, + 0.0024945696350187063, + 0.028287498280405998, + 0.00845933798700571, + 0.0341498889029026, + 0.0070052617229521275, + 0.008202314376831055, + 0.012007022276520729, + 0.005390811711549759, + 0.03122968040406704, + 0.01073111779987812, + 0.00336502306163311, + 0.011999985203146935, + 0.0019410125678405166, + 0.0030048976186662912, + 0.0033481139689683914, + 0.001962649170309305, + 0.011697137728333473, + 0.05417827516794205, + 0.01583516225218773, + 0.2655413746833801, + 0.015983209013938904, + 0.021015023812651634, + 0.02569938823580742, + 0.01118802186101675, + 0.2179952710866928, + 0.012049858458340168, + 0.003408050164580345, + 0.008228217251598835, + 0.003166018519550562, + 0.0031325933523476124, + 0.003475069534033537, + 0.0021416288800537586, + 0.006432725582271814, + 0.0020378862973302603, + 0.0011677674483507872, + 0.0021462622098624706, + 0.0007070808205753565, + 0.001751138363033533, + 0.0009892835514619946, + 0.0006601568893529475, + 0.0014298133319243789, + 0.001451269257813692, + 0.002242081332951784, + 0.0017485077260062099, + 0.0009855468524619937, + 0.0012974180281162262, + 0.00180103350430727, + 0.0008589675999246538, + 0.0010243197903037071, + 0.0022880015894770622, + 0.006878741085529327, + 0.002448444487527013, + 0.0013593053445219994, + 0.004509731195867062, + 0.003414225997403264, + 0.0016186784487217665, + 0.0019294488010928035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.062031254172325134 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot captured from a gentle overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small wooden sightseeing boat anchored in the middle of a broad waterway with opaque, milk chocolate-colored water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat painted creamy white with faded sky-blue trim and contains only simple wooden benches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat exhibit a slow, rhythmic lurching motion caused by a gentle undercurrent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a soft-focus effect, warm golden glow, and visible hand-painted brushstroke textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001939084380865097 + ], + "opposite_logprob_tokens": [ + -6.251939296722412 + ], + "logprob_reward": -0.001939084380865097, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001939084380865097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12865562736988068, + 0.04057154059410095, + 0.03031875751912594, + 0.22498907148838043, + 0.05407039448618889, + 0.44470036029815674, + 0.024275610223412514, + 0.02022268809378147, + 0.012791702523827553, + 0.019404232501983643 + ], + "credit_map": [ + 0.02208804525434971, + 0.009452611207962036, + 0.030489606782794, + 0.009001842699944973, + 0.009840165264904499, + 0.016128916293382645, + 0.006967916619032621, + 0.024686522781848907, + 0.006927233189344406, + 0.0022141009103506804, + 0.015035203658044338, + 0.0021914246026426554, + 0.0025643170811235905, + 0.0028970339335501194, + 0.0012419942067936063, + 0.007500231731683016, + 0.003624929813668132, + 0.001604594523087144, + 0.011919267475605011, + 0.001413231948390603, + 0.001972024329006672, + 0.0016445752698928118, + 0.0008522741845808923, + 0.007287859451025724, + 0.04761902242898941, + 0.014261339791119099, + 0.05594133585691452, + 0.011711197905242443, + 0.013823457062244415, + 0.018675807863473892, + 0.009115413762629032, + 0.053841497749090195, + 0.010979396291077137, + 0.003565762657672167, + 0.013346402905881405, + 0.002472480060532689, + 0.00346948835067451, + 0.004082698840647936, + 0.002478876383975148, + 0.013675292022526264, + 0.037681519985198975, + 0.010831058956682682, + 0.18870516121387482, + 0.011377346701920033, + 0.014246419072151184, + 0.01783420704305172, + 0.007706387434154749, + 0.15631826221942902, + 0.003121865913271904, + 0.002600123407319188, + 0.004660508129745722, + 0.0018895927350968122, + 0.0052075632847845554, + 0.002499445341527462, + 0.0015268867136910558, + 0.0027696245815604925, + 0.003361474722623825, + 0.0023507289588451385, + 0.00407932884991169, + 0.0014777079923078418, + 0.003447372931987047, + 0.002221141941845417, + 0.0013104479294270277, + 0.001974485581740737, + 0.0021215018350631, + 0.0015000591520220041, + 0.0020216084085404873, + 0.0010385570349171758, + 0.0011441053356975317, + 0.0016740416176617146, + 0.001906914752908051, + 0.0013849146198481321, + 0.0021956837736070156, + 0.006015844643115997, + 0.0022849387023597956, + 0.001177229918539524, + 0.0021676344331353903, + 0.0025064204819500446, + 0.0009721340029500425, + 0.0020843471866101027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.001939084380865097 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_12/index_tensor([181, 181])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_12/index_tensor([181, 181])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..94babea2eb752b3ddf03c618ad1a63ac9ba3566d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_12/index_tensor([181, 181])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a dramatic low angle, the camera remains still, gazing into a dark, smoke-filled void, where elegant white lettering hangs suspended in absolute darkness. Crafted in a sharp, classic serif font with a clean, matte finish, the letters are illuminated by a single, harsh keylight from an unseen source, creating the stark, high-contrast black-and-white contrast characteristic of classic film noir. The typography executes a slow, hypnotic torsional pendulum oscillation, deliberately twisting back and forth on invisible vertical lines. As the letters rotate, harsh light hits their clean edges and flat planes, while their sharp, elongated shadows are cast on the slowly billowing smoke behind them, dancing and contorting ominously in a silent, menacing ballet, imbuing the image with a palpable tension and mystery.", + "seed": 28027, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_12/index_181_seed_28027_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08767803749069572, + "vlm_logprob": -0.08767803749069572, + "vlm_yes_no_margin": 3.424999988079071, + "vlm_token_credit": -0.08767803749069572, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0659177303314209 + ], + "opposite_logprob_tokens": [ + -3.815917730331421 + ], + "logprob_reward": -0.0659177303314209, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0659177303314209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40193599462509155, + 0.1018744707107544, + 0.06028595194220543, + 0.11525724083185196, + 0.12523441016674042, + 0.04942000284790993, + 0.023688428103923798, + 0.044452957808971405, + 0.01869327947497368, + 0.05915721878409386 + ], + "credit_map": [ + 0.014682433567941189, + 0.01224813424050808, + 0.08967258036136627, + 0.01045287773013115, + 0.01700998656451702, + 0.16628478467464447, + 0.01749259978532791, + 0.07409263402223587, + 0.009029385633766651, + 0.018395883962512016, + 0.009991656057536602, + 0.009039080701768398, + 0.019933365285396576, + 0.009328602813184261, + 0.009421546943485737, + 0.016734948381781578, + 0.004949780646711588, + 0.0070059048011898994, + 0.006591966841369867, + 0.006635869853198528, + 0.01466598641127348, + 0.010199089534580708, + 0.004173322580754757, + 0.006064033135771751, + 0.010529573075473309, + 0.00536209624260664, + 0.035540949553251266, + 0.004774722270667553, + 0.006370930932462215, + 0.03731587901711464, + 0.003826639149338007, + 0.011536456644535065, + 0.0223868228495121, + 0.008445492014288902, + 0.03190060332417488, + 0.006935775279998779, + 0.015423687174916267, + 0.01271863467991352, + 0.006492675747722387, + 0.02093072235584259, + 0.0034482532646507025, + 0.003898446448147297, + 0.005270019173622131, + 0.0033017133828252554, + 0.006037509068846703, + 0.01273330207914114, + 0.0031123000662773848, + 0.01161846425384283, + 0.003423194168135524, + 0.0023622428998351097, + 0.004309139680117369, + 0.0020221350714564323, + 0.004274145234376192, + 0.0028964753728359938, + 0.0017569975461810827, + 0.0026441002264618874, + 0.005732244346290827, + 0.0035067445132881403, + 0.007094248663634062, + 0.0050915018655359745, + 0.004892860539257526, + 0.0046685123816132545, + 0.0054149623028934, + 0.00805188249796629, + 0.0018364953575655818, + 0.002164404606446624, + 0.002444207202643156, + 0.0024244869127869606, + 0.0017021161038428545, + 0.002161082113161683, + 0.0024849087931215763, + 0.0034755803644657135, + 0.004812848288565874, + 0.004709263797849417, + 0.009687226265668869, + 0.004457260016351938, + 0.005419300869107246, + 0.01861155405640602, + 0.0033443940337747335, + 0.008115374483168125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0659177303314209 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005972159560769796 + ], + "opposite_logprob_tokens": [ + -5.130972385406494 + ], + "logprob_reward": -0.005972159560769796, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005972159560769796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.495138555765152, + 0.06084194779396057, + 0.04885674640536308, + 0.21791017055511475, + 0.07158646732568741, + 0.025335902348160744, + 0.01707589067518711, + 0.013254577293992043, + 0.020856071263551712, + 0.02914373204112053 + ], + "credit_map": [ + 0.020199887454509735, + 0.01224268227815628, + 0.09692185372114182, + 0.011399236507713795, + 0.015164838172495365, + 0.2056107372045517, + 0.02155083790421486, + 0.1120484471321106, + 0.0037372855003923178, + 0.008092707954347134, + 0.005210111383348703, + 0.01048226933926344, + 0.008687336929142475, + 0.005296819843351841, + 0.011332014575600624, + 0.008003400638699532, + 0.008503268472850323, + 0.0031595907639712095, + 0.005905047059059143, + 0.004227867815643549, + 0.004818591754883528, + 0.011132864281535149, + 0.0031272543128579855, + 0.007982264272868633, + 0.022115278989076614, + 0.004264630377292633, + 0.08108600229024887, + 0.007463005371391773, + 0.00503584137186408, + 0.07657447457313538, + 0.00665394589304924, + 0.014717002399265766, + 0.012419807724654675, + 0.005739517975598574, + 0.017900696024298668, + 0.004680894780904055, + 0.008629290387034416, + 0.006704272236675024, + 0.0038969230372458696, + 0.011615068651735783, + 0.0015069162473082542, + 0.0011826788540929556, + 0.0030192099511623383, + 0.0011089174076914787, + 0.0015645475359633565, + 0.008946577087044716, + 0.0012930555967614055, + 0.006713999435305595, + 0.0024268883280456066, + 0.0014102427521720529, + 0.003591390559449792, + 0.0015377092640846968, + 0.0024760013911873102, + 0.0022908158134669065, + 0.0012148454552516341, + 0.002127996413037181, + 0.0014925498981028795, + 0.0015714345499873161, + 0.0016775564290583134, + 0.0025567766278982162, + 0.0013143933610990644, + 0.0012512527173385024, + 0.0018198118777945638, + 0.001570802996866405, + 0.0018529159715399146, + 0.0031359046697616577, + 0.002378797624260187, + 0.0032196855172514915, + 0.002825529081746936, + 0.0020479667000472546, + 0.0030149214435368776, + 0.0023803510703146458, + 0.002459752606227994, + 0.0028865353669971228, + 0.005301093682646751, + 0.0027421058621257544, + 0.002846519695594907, + 0.008299047127366066, + 0.0022261308040469885, + 0.0023825480602681637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.005972159560769796 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134177327156067 + ], + "opposite_logprob_tokens": [ + -1.313417673110962 + ], + "logprob_reward": -0.3134177327156067, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.3134177327156067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6733931303024292, + 0.0522456169128418, + 0.03727938234806061, + 0.08389313519001007, + 0.05219266936182976, + 0.03257833793759346, + 0.013867156580090523, + 0.025434177368879318, + 0.00655567878857255, + 0.02256070077419281 + ], + "credit_map": [ + 0.02712280862033367, + 0.012415294535458088, + 0.12759442627429962, + 0.01282663457095623, + 0.01753070391714573, + 0.2841477692127228, + 0.030084388330578804, + 0.16167111694812775, + 0.004591358359903097, + 0.01027983333915472, + 0.005532150622457266, + 0.004964364692568779, + 0.010131051763892174, + 0.004625310655683279, + 0.004740734584629536, + 0.007380817085504532, + 0.0031663221307098866, + 0.003593314439058304, + 0.004414383787661791, + 0.0026905550621449947, + 0.0052482676692306995, + 0.00745152635499835, + 0.0025325510650873184, + 0.008182467892765999, + 0.00667888019233942, + 0.004758380353450775, + 0.023579338565468788, + 0.0031281206756830215, + 0.004815171007066965, + 0.027726221829652786, + 0.003093098057433963, + 0.010113934986293316, + 0.010767278261482716, + 0.004896633327007294, + 0.013928919099271297, + 0.0037795426324009895, + 0.00372455851174891, + 0.004192651715129614, + 0.002467538695782423, + 0.008435552939772606, + 0.001889183302409947, + 0.0014052053447812796, + 0.0033593534026294947, + 0.0023409686982631683, + 0.0058704763650894165, + 0.00844455137848854, + 0.0026280307210981846, + 0.006640569306910038, + 0.0017481670947745442, + 0.001538244541734457, + 0.0017510044854134321, + 0.0015808362513780594, + 0.0029978498350828886, + 0.0020132330246269703, + 0.0009915520204231143, + 0.001246271189302206, + 0.003466041525825858, + 0.0012737232027575374, + 0.007128357887268066, + 0.00239658635109663, + 0.0011907810112461448, + 0.001281911158002913, + 0.0020116937812417746, + 0.006685085128992796, + 0.0005688406527042389, + 0.0005231850082054734, + 0.0007190940668806434, + 0.000642957747913897, + 0.0006309182499535382, + 0.0007271560025401413, + 0.0013590168673545122, + 0.0013845106586813927, + 0.002006174996495247, + 0.002009705174714327, + 0.00521838990971446, + 0.0016166435088962317, + 0.0017271600663661957, + 0.006703871767967939, + 0.0010720901191234589, + 0.0022066677920520306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.3134177327156067 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023286210373044014 + ], + "opposite_logprob_tokens": [ + -3.7732861042022705 + ], + "logprob_reward": -0.023286210373044014, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023286210373044014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3993840217590332, + 0.11174708604812622, + 0.06584608554840088, + 0.048850562423467636, + 0.1323966085910797, + 0.08565264940261841, + 0.032923247665166855, + 0.052453361451625824, + 0.03113616444170475, + 0.03961021453142166 + ], + "credit_map": [ + 0.015858180820941925, + 0.015699181705713272, + 0.07325975596904755, + 0.012503067962825298, + 0.02325434982776642, + 0.15621547400951385, + 0.01911250315606594, + 0.08348152786493301, + 0.007720010820776224, + 0.01723564974963665, + 0.01055537536740303, + 0.01646890677511692, + 0.021116893738508224, + 0.010815316811203957, + 0.013574476353824139, + 0.01426045410335064, + 0.0063078925013542175, + 0.0066696470603346825, + 0.009001664817333221, + 0.0082131028175354, + 0.008753739297389984, + 0.010688550770282745, + 0.00517275370657444, + 0.011038733646273613, + 0.006162743084132671, + 0.003449616488069296, + 0.0068987407721579075, + 0.003196358447894454, + 0.00442341435700655, + 0.01432593073695898, + 0.003064631251618266, + 0.007329125888645649, + 0.022080089896917343, + 0.011707084253430367, + 0.03047635778784752, + 0.013005321845412254, + 0.014465712010860443, + 0.016946226358413696, + 0.009449639357626438, + 0.014266187325119972, + 0.007418166380375624, + 0.008780091069638729, + 0.01174941472709179, + 0.009267154149711132, + 0.018593424931168556, + 0.012178127653896809, + 0.007879157550632954, + 0.009787118062376976, + 0.004443145357072353, + 0.00302444351837039, + 0.004431350156664848, + 0.003315899521112442, + 0.007958626374602318, + 0.0037448254879564047, + 0.0028547639958560467, + 0.0031501930207014084, + 0.003780203405767679, + 0.004942069295793772, + 0.00870614405721426, + 0.007458868436515331, + 0.006172190885990858, + 0.004871800076216459, + 0.007426187861710787, + 0.009095896035432816, + 0.003622667631134391, + 0.003408057615160942, + 0.0035790305119007826, + 0.0030204399954527617, + 0.0037324903532862663, + 0.003973482176661491, + 0.004493134096264839, + 0.005306862760335207, + 0.004201738629490137, + 0.004530259873718023, + 0.004164623562246561, + 0.004183041863143444, + 0.0057779354974627495, + 0.0064628152176737785, + 0.003525746986269951, + 0.006764051970094442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.023286210373044014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796354472637177 + ], + "opposite_logprob_tokens": [ + -3.5297963619232178 + ], + "logprob_reward": -0.029796354472637177, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796354472637177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5468281507492065, + 0.034015294164419174, + 0.01399920042604208, + 0.28148984909057617, + 0.03311346471309662, + 0.021484093740582466, + 0.018539469689130783, + 0.01175245177000761, + 0.02322361432015896, + 0.015554423443973064 + ], + "credit_map": [ + 0.021814683452248573, + 0.01280659157782793, + 0.11580220609903336, + 0.012361698783934116, + 0.016119396314024925, + 0.2289235144853592, + 0.02318986877799034, + 0.11581020802259445, + 0.0018661043141037226, + 0.0048006451688706875, + 0.003316738409921527, + 0.00598848145455122, + 0.003927578683942556, + 0.0031759091652929783, + 0.005094836000353098, + 0.005844998639076948, + 0.0012243158416822553, + 0.0015898848650977015, + 0.0015395868103951216, + 0.002160390606150031, + 0.002059446880593896, + 0.0022798229474574327, + 0.0010883985087275505, + 0.002057353500276804, + 0.033515822142362595, + 0.00601428234949708, + 0.11019241064786911, + 0.009723042137920856, + 0.005519601982086897, + 0.0947619304060936, + 0.007407117635011673, + 0.014355653896927834, + 0.0051246630027890205, + 0.0028303642757236958, + 0.007827778346836567, + 0.002341307932510972, + 0.0046270983293652534, + 0.0030742089729756117, + 0.001890304614789784, + 0.0053977398201823235, + 0.0018083227332681417, + 0.0022032931447029114, + 0.0029823980294167995, + 0.0022759672719985247, + 0.0042301141656935215, + 0.003378208726644516, + 0.0019818430300801992, + 0.0026239464059472084, + 0.002229810692369938, + 0.0015207348624244332, + 0.0029699921142309904, + 0.0015743193216621876, + 0.0035072460304945707, + 0.0023471650201827288, + 0.001392650417983532, + 0.0029975513461977243, + 0.0014277346199378371, + 0.001164863584563136, + 0.0014048703014850616, + 0.002328373258933425, + 0.0015406834427267313, + 0.0013222926063463092, + 0.0011480264365673065, + 0.0014156088000163436, + 0.0024789690505713224, + 0.002930205315351486, + 0.0033824662677943707, + 0.0029756734147667885, + 0.002706629689782858, + 0.00294332392513752, + 0.0028588962741196156, + 0.002947452710941434, + 0.0015329403104260564, + 0.0017978677060455084, + 0.002054475713521242, + 0.0020001446828246117, + 0.001890338840894401, + 0.0028758232947438955, + 0.001820312230847776, + 0.0015825203154236078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.029796354472637177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a dramatic low angle, the camera remains still, gazing into a dark, smoke-filled void, where elegant white lettering hangs suspended in absolute darkness. Crafted in a sharp, classic serif font with a clean, matte finish, the letters are illuminated by a single, harsh keylight from an unseen source, creating the stark, high-contrast black-and-white contrast characteristic of classic film noir. The typography executes a slow, hypnotic torsional pendulum oscillation, deliberately twisting back and forth on invisible vertical lines. As the letters rotate, harsh light hits their clean edges and flat planes, while their sharp, elongated shadows are cast on the slowly billowing smoke behind them, dancing and contorting ominously in a silent, menacing ballet, imbuing the image with a palpable tension and mystery.", + "seed": 28028, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_12/index_181_seed_28028_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09646692546084523, + "vlm_logprob": -0.09646692546084523, + "vlm_yes_no_margin": 3.574999988079071, + "vlm_token_credit": -0.09646692546084523, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04613925516605377 + ], + "opposite_logprob_tokens": [ + -4.421139240264893 + ], + "logprob_reward": -0.04613925516605377, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.04613925516605377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06881099194288254, + 0.05419202521443367, + 0.08485857397317886, + 0.08990182727575302, + 0.21947532892227173, + 0.09406723082065582, + 0.09463294595479965, + 0.13396285474300385, + 0.11205077916383743, + 0.04804738610982895 + ], + "credit_map": [ + 0.01106636505573988, + 0.005283553618937731, + 0.009516109712421894, + 0.007139767054468393, + 0.006192748434841633, + 0.014930576086044312, + 0.00655791349709034, + 0.008123958483338356, + 0.005851568654179573, + 0.005434275139123201, + 0.006303325295448303, + 0.00934641808271408, + 0.006485643330961466, + 0.007322433404624462, + 0.007963832467794418, + 0.005484527908265591, + 0.01131616160273552, + 0.007823246531188488, + 0.008729986846446991, + 0.0106586255133152, + 0.007069945335388184, + 0.0156027153134346, + 0.011966380290687084, + 0.011691514402627945, + 0.013806103728711605, + 0.008568891324102879, + 0.009283609688282013, + 0.011564476415514946, + 0.007672374602407217, + 0.015666553750634193, + 0.009643745608627796, + 0.013696075417101383, + 0.03107934445142746, + 0.021572917699813843, + 0.012486775405704975, + 0.01621505804359913, + 0.02773936465382576, + 0.05442095175385475, + 0.015515021979808807, + 0.040445905178785324, + 0.011559197679162025, + 0.015433837659657001, + 0.012250718660652637, + 0.010724583640694618, + 0.011143357492983341, + 0.011403617449104786, + 0.007538152858614922, + 0.014013765379786491, + 0.017199497669935226, + 0.006224328186362982, + 0.008866998367011547, + 0.008297036401927471, + 0.010558677837252617, + 0.017290767282247543, + 0.010811174288392067, + 0.015384472906589508, + 0.014422902837395668, + 0.016148880124092102, + 0.025458425283432007, + 0.023172659799456596, + 0.009251493029296398, + 0.014158054254949093, + 0.013644305989146233, + 0.017706135287880898, + 0.021202504634857178, + 0.011004338040947914, + 0.010967038571834564, + 0.015044370666146278, + 0.013348076492547989, + 0.01990729197859764, + 0.00789589062333107, + 0.012681267224252224, + 0.004073185846209526, + 0.005052495282143354, + 0.005932359490543604, + 0.006705943029373884, + 0.007130368147045374, + 0.0054040467366576195, + 0.00784455519169569, + 0.005904434248805046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.04613925516605377 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016096286475658417 + ], + "opposite_logprob_tokens": [ + -4.141096115112305 + ], + "logprob_reward": -0.016096286475658417, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016096286475658417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1010613963007927, + 0.14930179715156555, + 0.05426108464598656, + 0.04965977370738983, + 0.09200194478034973, + 0.24093817174434662, + 0.04545645788311958, + 0.12794406712055206, + 0.07140297442674637, + 0.067972332239151 + ], + "credit_map": [ + 0.016603603959083557, + 0.010350325144827366, + 0.010979707352817059, + 0.009600856341421604, + 0.010217057541012764, + 0.019649554044008255, + 0.008802421391010284, + 0.014857862144708633, + 0.02157953940331936, + 0.016670232638716698, + 0.017412938177585602, + 0.012772348709404469, + 0.019071513786911964, + 0.020969852805137634, + 0.012945451773703098, + 0.027879934757947922, + 0.007975873537361622, + 0.004790929611772299, + 0.007769413758069277, + 0.008072060532867908, + 0.004763144534081221, + 0.008049963042140007, + 0.006229355465620756, + 0.006610347889363766, + 0.008041500113904476, + 0.004308151546865702, + 0.005631375592201948, + 0.008884265087544918, + 0.0034193331375718117, + 0.009638054296374321, + 0.004221734590828419, + 0.005515362601727247, + 0.013243259862065315, + 0.0089634470641613, + 0.006181838922202587, + 0.006639118306338787, + 0.011440571397542953, + 0.022475628182291985, + 0.00648579141125083, + 0.016572294756770134, + 0.033312320709228516, + 0.029106076806783676, + 0.029090164229273796, + 0.027274493128061295, + 0.02114572562277317, + 0.036700427532196045, + 0.021734194830060005, + 0.04257477447390556, + 0.006188786588609219, + 0.005831328220665455, + 0.00490317028015852, + 0.003328293561935425, + 0.004901510197669268, + 0.0077161891385912895, + 0.004137113690376282, + 0.008450065739452839, + 0.013795565813779831, + 0.024116164073348045, + 0.016446495428681374, + 0.019429707899689674, + 0.009694892913103104, + 0.01408290583640337, + 0.01022985391318798, + 0.020148487761616707, + 0.00898039061576128, + 0.005179323721677065, + 0.007483246270567179, + 0.014379429630935192, + 0.009731331840157509, + 0.011628272011876106, + 0.006371245253831148, + 0.0076497383415699005, + 0.006487751845270395, + 0.009261319413781166, + 0.006935281679034233, + 0.009235099889338017, + 0.009970656596124172, + 0.0074791028164327145, + 0.010933120734989643, + 0.007669998798519373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.016096286475658417 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38702112436294556 + ], + "opposite_logprob_tokens": [ + -1.1370210647583008 + ], + "logprob_reward": -0.38702112436294556, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.38702112436294556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06085798144340515, + 0.12221910059452057, + 0.06709287315607071, + 0.06643641740083694, + 0.10364249348640442, + 0.37288978695869446, + 0.03437134623527527, + 0.08205632120370865, + 0.041252363473176956, + 0.049181267619132996 + ], + "credit_map": [ + 0.007731356658041477, + 0.0057879104278981686, + 0.010832399129867554, + 0.006579853128641844, + 0.006792139727622271, + 0.0069969757460057735, + 0.007855352014303207, + 0.008282000198960304, + 0.015415074303746223, + 0.013943866826593876, + 0.015232938341796398, + 0.010886892676353455, + 0.016862941905856133, + 0.015395764261484146, + 0.014548041857779026, + 0.019933588802814484, + 0.010022205300629139, + 0.006068219896405935, + 0.006766798906028271, + 0.0073824492283165455, + 0.006111285649240017, + 0.014893650077283382, + 0.009279780089855194, + 0.00656849006190896, + 0.01101535465568304, + 0.005922670941799879, + 0.00679598655551672, + 0.009065941907465458, + 0.004262662027031183, + 0.013992921449244022, + 0.005949752405285835, + 0.009431137703359127, + 0.01613226719200611, + 0.010019941255450249, + 0.008578118868172169, + 0.007742899935692549, + 0.013578626327216625, + 0.023918205872178078, + 0.006611996795982122, + 0.017060445621609688, + 0.042371559888124466, + 0.04720763862133026, + 0.04869230464100838, + 0.04562479630112648, + 0.02808403968811035, + 0.0561186820268631, + 0.035097479820251465, + 0.06969331204891205, + 0.005721710156649351, + 0.00241909665055573, + 0.0037927019875496626, + 0.0024207066744565964, + 0.0044871303252875805, + 0.005306785926222801, + 0.00417970260605216, + 0.00604351656511426, + 0.009326300583779812, + 0.014747598208487034, + 0.012148965150117874, + 0.010256743989884853, + 0.006425417494028807, + 0.009400250390172005, + 0.006704471539705992, + 0.013046583160758018, + 0.004965255036950111, + 0.0049020834267139435, + 0.004066759720444679, + 0.0039116693660616875, + 0.009675304405391216, + 0.005783012602478266, + 0.0035959903616458178, + 0.004352292977273464, + 0.005193790420889854, + 0.007611782290041447, + 0.00553909270092845, + 0.006279448047280312, + 0.0064177741296589375, + 0.005701986141502857, + 0.007902615703642368, + 0.004534777719527483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.38702112436294556 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263423640280962 + ], + "opposite_logprob_tokens": [ + -3.6513423919677734 + ], + "logprob_reward": -0.0263423640280962, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263423640280962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07344114780426025, + 0.14228637516498566, + 0.04488225281238556, + 0.03811357915401459, + 0.037711601704359055, + 0.06973055750131607, + 0.04228590801358223, + 0.3703227639198303, + 0.14343640208244324, + 0.03778943419456482 + ], + "credit_map": [ + 0.012493912130594254, + 0.005990997422486544, + 0.00863994937390089, + 0.00868223700672388, + 0.004851287696510553, + 0.012146309949457645, + 0.00787640642374754, + 0.012760045006871223, + 0.016185007989406586, + 0.015271082520484924, + 0.01870538853108883, + 0.013632562011480331, + 0.019023386761546135, + 0.016314635053277016, + 0.02073763683438301, + 0.022416673600673676, + 0.007539643906056881, + 0.005906971171498299, + 0.004997656214982271, + 0.005155575927346945, + 0.004529241472482681, + 0.0056227222084999084, + 0.005743836518377066, + 0.005386603996157646, + 0.006341458763927221, + 0.0043592010624706745, + 0.00396416150033474, + 0.004135170951485634, + 0.00329004623927176, + 0.005628759041428566, + 0.0036866627633571625, + 0.006708120461553335, + 0.0064794267527759075, + 0.0026277850847691298, + 0.005144109949469566, + 0.003810645081102848, + 0.0032479313667863607, + 0.007997237145900726, + 0.002726187463849783, + 0.005678280256688595, + 0.014021673239767551, + 0.007928627543151379, + 0.007336962968111038, + 0.007918966934084892, + 0.0055261580273509026, + 0.011053464375436306, + 0.004289701581001282, + 0.011655003763735294, + 0.0056523908860981464, + 0.00395190017297864, + 0.0040144952945411205, + 0.0038338496815413237, + 0.005119460169225931, + 0.008671467192471027, + 0.004789619706571102, + 0.006252726074308157, + 0.04136652499437332, + 0.06938917189836502, + 0.05155850574374199, + 0.0511808842420578, + 0.028473736718297005, + 0.04251335561275482, + 0.029554041102528572, + 0.05628654733300209, + 0.015412420965731144, + 0.015267959795892239, + 0.01582457683980465, + 0.019579971209168434, + 0.028643695637583733, + 0.017218759283423424, + 0.016981173306703568, + 0.014507838524878025, + 0.0037106506060808897, + 0.005130299367010593, + 0.004544403869658709, + 0.005046232603490353, + 0.006061542313545942, + 0.0042821732349693775, + 0.005231729708611965, + 0.003782398533076048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.0263423640280962 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735597271472216 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735597271472216, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735597271472216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09334580600261688, + 0.44310781359672546, + 0.06783007085323334, + 0.03688400611281395, + 0.07086057960987091, + 0.07012662291526794, + 0.040544893592596054, + 0.07014285027980804, + 0.07645684480667114, + 0.030700549483299255 + ], + "credit_map": [ + 0.014582549221813679, + 0.008164573460817337, + 0.01026247814297676, + 0.01168295182287693, + 0.009343368001282215, + 0.01302774716168642, + 0.012814006768167019, + 0.013468124903738499, + 0.05131059139966965, + 0.05154498293995857, + 0.053928326815366745, + 0.04285486787557602, + 0.06389091908931732, + 0.05361657217144966, + 0.05925263836979866, + 0.06670887023210526, + 0.010901043191552162, + 0.004181307274848223, + 0.010454662144184113, + 0.008162780664861202, + 0.003297919873148203, + 0.013045165687799454, + 0.005857422947883606, + 0.011929762549698353, + 0.006069978699088097, + 0.0036921948194503784, + 0.004640218801796436, + 0.005006761755794287, + 0.0029826161917299032, + 0.006680876482278109, + 0.003827498760074377, + 0.003983858972787857, + 0.010262926109135151, + 0.006571270991116762, + 0.004664805717766285, + 0.0052704275585711, + 0.008746176958084106, + 0.01792260818183422, + 0.0048902821727097034, + 0.012532074004411697, + 0.009648664854466915, + 0.006606430280953646, + 0.009260108694434166, + 0.008322088979184628, + 0.006030902732163668, + 0.009541735053062439, + 0.007539247162640095, + 0.013177438639104366, + 0.005284471902996302, + 0.0030929294880479574, + 0.005124975461512804, + 0.0028121843934059143, + 0.004372949246317148, + 0.006378193385899067, + 0.004915446043014526, + 0.008563744835555553, + 0.008250294253230095, + 0.011571777984499931, + 0.011609956622123718, + 0.008504381403326988, + 0.005974273197352886, + 0.008620969951152802, + 0.00572505546733737, + 0.009886137209832668, + 0.00802201870828867, + 0.00845271348953247, + 0.006786540616303682, + 0.008361672051250935, + 0.0172199048101902, + 0.011236516758799553, + 0.006602191366255283, + 0.00977527815848589, + 0.003116794629022479, + 0.003288991516456008, + 0.0036915973760187626, + 0.004767002537846565, + 0.004570020362734795, + 0.003202534979209304, + 0.004882047884166241, + 0.003181557869538665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.006735597271472216 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a dramatic low angle, the camera remains still, gazing into a dark, smoke-filled void, where elegant white lettering hangs suspended in absolute darkness. Crafted in a sharp, classic serif font with a clean, matte finish, the letters are illuminated by a single, harsh keylight from an unseen source, creating the stark, high-contrast black-and-white contrast characteristic of classic film noir. The typography executes a slow, hypnotic torsional pendulum oscillation, deliberately twisting back and forth on invisible vertical lines. As the letters rotate, harsh light hits their clean edges and flat planes, while their sharp, elongated shadows are cast on the slowly billowing smoke behind them, dancing and contorting ominously in a silent, menacing ballet, imbuing the image with a palpable tension and mystery.", + "seed": 28029, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_12/index_181_seed_28029_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06602609874680639, + "vlm_logprob": -0.06602609874680639, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.06602609874680639, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07240331918001175 + ], + "opposite_logprob_tokens": [ + -3.5724034309387207 + ], + "logprob_reward": -0.07240331918001175, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.07240331918001175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22028741240501404, + 0.08070162683725357, + 0.1311115324497223, + 0.07758177816867828, + 0.10300567001104355, + 0.1364365816116333, + 0.03829703480005264, + 0.04569265618920326, + 0.07868913561105728, + 0.08819658309221268 + ], + "credit_map": [ + 0.028699027374386787, + 0.034932106733322144, + 0.025991709902882576, + 0.026980971917510033, + 0.030600160360336304, + 0.02050546370446682, + 0.033989135175943375, + 0.018588829785585403, + 0.007655633147805929, + 0.010501053184270859, + 0.010998592711985111, + 0.00983155146241188, + 0.011386184953153133, + 0.010168052278459072, + 0.010770631954073906, + 0.009389923885464668, + 0.01719285361468792, + 0.01992803066968918, + 0.018823828548192978, + 0.013505006209015846, + 0.014340112917125225, + 0.01765783317387104, + 0.015976443886756897, + 0.013687425293028355, + 0.007276867050677538, + 0.010175015777349472, + 0.007780984975397587, + 0.008809955790638924, + 0.009821867570281029, + 0.011879787780344486, + 0.01310126855969429, + 0.008736028335988522, + 0.011353663168847561, + 0.010400488041341305, + 0.012300143949687481, + 0.008885634131729603, + 0.019905492663383484, + 0.013740591704845428, + 0.014512112364172935, + 0.011907554231584072, + 0.014367037452757359, + 0.018454408273100853, + 0.019053997471928596, + 0.020148422569036484, + 0.015484243631362915, + 0.01579934172332287, + 0.019705265760421753, + 0.013423868454992771, + 0.0034593904856592417, + 0.005101181101053953, + 0.00407414510846138, + 0.0038619190454483032, + 0.005641864147037268, + 0.004903725814074278, + 0.004773367661982775, + 0.006481444463133812, + 0.00618696166202426, + 0.005768853239715099, + 0.0073515987023711205, + 0.0037454506382346153, + 0.0060521382838487625, + 0.006114482413977385, + 0.004543808754533529, + 0.005929360166192055, + 0.013232882134616375, + 0.008396132849156857, + 0.010519467294216156, + 0.01469010766595602, + 0.008225478231906891, + 0.007617126684635878, + 0.009330966509878635, + 0.0066769737750291824, + 0.008097372949123383, + 0.011220843531191349, + 0.0074401856400072575, + 0.015378031879663467, + 0.010259627364575863, + 0.01384074054658413, + 0.012295867316424847, + 0.009663909673690796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.07240331918001175 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002222926588729024 + ], + "opposite_logprob_tokens": [ + -6.127223014831543 + ], + "logprob_reward": -0.002222926588729024, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002222926588729024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4271557033061981, + 0.19285334646701813, + 0.07014912366867065, + 0.04448559507727623, + 0.08271628618240356, + 0.04222620278596878, + 0.03677959740161896, + 0.032144732773303986, + 0.027166223153471947, + 0.0443231575191021 + ], + "credit_map": [ + 0.04942803829908371, + 0.044586654752492905, + 0.055152833461761475, + 0.04912353307008743, + 0.05386650562286377, + 0.05275057628750801, + 0.07300987094640732, + 0.04923766478896141, + 0.019041800871491432, + 0.023831386119127274, + 0.029610738158226013, + 0.019482571631669998, + 0.021118827164173126, + 0.02598053403198719, + 0.02864507958292961, + 0.025142403319478035, + 0.00833197683095932, + 0.01231673825532198, + 0.008126330561935902, + 0.012156188488006592, + 0.004878736566752195, + 0.005188946612179279, + 0.01210601907223463, + 0.007044186815619469, + 0.0064604803919792175, + 0.004404766485095024, + 0.009939443320035934, + 0.005060082767158747, + 0.004496247041970491, + 0.004087680019438267, + 0.004753805231302977, + 0.005283091217279434, + 0.008201823569834232, + 0.010188981890678406, + 0.009916836395859718, + 0.00816943310201168, + 0.015880681574344635, + 0.009920952841639519, + 0.00914489384740591, + 0.011292688548564911, + 0.0058718896470963955, + 0.005411657504737377, + 0.006499806419014931, + 0.004976991564035416, + 0.004477500915527344, + 0.004844288341701031, + 0.005222311243414879, + 0.00492175854742527, + 0.004418911412358284, + 0.0033281217329204082, + 0.005327984225004911, + 0.003615261986851692, + 0.00446321303024888, + 0.005747481249272823, + 0.00479582604020834, + 0.00508279399946332, + 0.004514245316386223, + 0.005550723988562822, + 0.0041014207527041435, + 0.003518900601193309, + 0.0033703558146953583, + 0.002982380101457238, + 0.004343700595200062, + 0.00376300560310483, + 0.0038171657361090183, + 0.0033164387568831444, + 0.004262445494532585, + 0.0031141077633947134, + 0.003861890407279134, + 0.002642463194206357, + 0.003022607183083892, + 0.003129104385152459, + 0.004041832871735096, + 0.004816913977265358, + 0.0038567243609577417, + 0.006384750362485647, + 0.004893547855317593, + 0.004655023105442524, + 0.009626987390220165, + 0.006047375965863466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.002222926588729024 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22554677724838257 + ], + "opposite_logprob_tokens": [ + -1.6005468368530273 + ], + "logprob_reward": -0.22554677724838257, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22554677724838257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3207499384880066, + 0.21792320907115936, + 0.11300237476825714, + 0.062099773436784744, + 0.06002812460064888, + 0.06180303543806076, + 0.039672646671533585, + 0.03649390861392021, + 0.03553612530231476, + 0.05269084498286247 + ], + "credit_map": [ + 0.03703492134809494, + 0.033740125596523285, + 0.04308994114398956, + 0.03773610293865204, + 0.04236485809087753, + 0.03866280987858772, + 0.052784454077482224, + 0.03533673286437988, + 0.02389460802078247, + 0.025541454553604126, + 0.038508545607328415, + 0.019876673817634583, + 0.02249475009739399, + 0.028553497046232224, + 0.028884537518024445, + 0.030169161036610603, + 0.012724208645522594, + 0.016987528651952744, + 0.01447543129324913, + 0.014639712870121002, + 0.010577346198260784, + 0.01434379443526268, + 0.018754856660962105, + 0.01049951184540987, + 0.012464901432394981, + 0.007105323951691389, + 0.007382560521364212, + 0.0060265702195465565, + 0.007274561561644077, + 0.009349452331662178, + 0.0062229810282588005, + 0.00627342751249671, + 0.004690845031291246, + 0.006260266527533531, + 0.00591142475605011, + 0.005026159808039665, + 0.008825008757412434, + 0.012090273201465607, + 0.008201946504414082, + 0.009022204205393791, + 0.006554241292178631, + 0.009094958193600178, + 0.007627693470567465, + 0.007546430919319391, + 0.008761612698435783, + 0.007833764888346195, + 0.008851470425724983, + 0.0055328658781945705, + 0.004218483343720436, + 0.0048668161034584045, + 0.0052477228455245495, + 0.0037160112988203764, + 0.005862117744982243, + 0.006175266113132238, + 0.004319603089243174, + 0.005266627762466669, + 0.003387467935681343, + 0.004388675093650818, + 0.004406076390296221, + 0.005421073641628027, + 0.004823226947337389, + 0.005760557018220425, + 0.0048182508908212185, + 0.0034885816276073456, + 0.005362601950764656, + 0.004626732785254717, + 0.00482576247304678, + 0.005025174468755722, + 0.004863013979047537, + 0.003335393499583006, + 0.003653757506981492, + 0.0038436902686953545, + 0.004214990884065628, + 0.004759481176733971, + 0.0038867632392793894, + 0.013843358494341373, + 0.005060586612671614, + 0.005539336241781712, + 0.01023888774216175, + 0.005147441755980253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.22554677724838257 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026344453915953636 + ], + "opposite_logprob_tokens": [ + -3.6513445377349854 + ], + "logprob_reward": -0.026344453915953636, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026344453915953636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30356642603874207, + 0.12319360673427582, + 0.10891947895288467, + 0.0953051745891571, + 0.1463712602853775, + 0.05027569457888603, + 0.051013994961977005, + 0.01958879828453064, + 0.047504257410764694, + 0.054261233657598495 + ], + "credit_map": [ + 0.033290401101112366, + 0.0363270603120327, + 0.039753030985593796, + 0.03499572351574898, + 0.04473065584897995, + 0.034423161298036575, + 0.046975910663604736, + 0.03307052701711655, + 0.012201598845422268, + 0.01529780961573124, + 0.017572764307260513, + 0.012615104205906391, + 0.01409207098186016, + 0.01809483766555786, + 0.01730998419225216, + 0.01600944995880127, + 0.01694805733859539, + 0.02274753339588642, + 0.013320466503500938, + 0.013924072496592999, + 0.010788969695568085, + 0.010960311628878117, + 0.009701274335384369, + 0.010528797283768654, + 0.010680010542273521, + 0.014704538509249687, + 0.015367873944342136, + 0.009280433878302574, + 0.01197004597634077, + 0.012347034178674221, + 0.011313660070300102, + 0.009641592390835285, + 0.015438219532370567, + 0.019140198826789856, + 0.024491429328918457, + 0.014940688386559486, + 0.024633223190903664, + 0.01544701587408781, + 0.016691451892256737, + 0.015589045360684395, + 0.004224647302180529, + 0.005925199016928673, + 0.00574059272184968, + 0.006274755112826824, + 0.00900217890739441, + 0.0066817025654017925, + 0.0074271755293011665, + 0.004999447148293257, + 0.005453138146549463, + 0.00486821960657835, + 0.009575027041137218, + 0.007400551810860634, + 0.0058622513897717, + 0.005648646969348192, + 0.006471117492765188, + 0.005735049024224281, + 0.0023826363030821085, + 0.002434095134958625, + 0.002415949013084173, + 0.0017412477172911167, + 0.002312908647581935, + 0.0028124800883233547, + 0.0026244898326694965, + 0.002864993177354336, + 0.006657542195171118, + 0.005746073555201292, + 0.009416099637746811, + 0.005108333192765713, + 0.005597755778580904, + 0.00431457394734025, + 0.005376320332288742, + 0.005287563893944025, + 0.0048478832468390465, + 0.007740040775388479, + 0.005767387803643942, + 0.006803626660257578, + 0.0069120810367167, + 0.006395426578819752, + 0.008542406372725964, + 0.007252382580190897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.026344453915953636 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003613016800954938 + ], + "opposite_logprob_tokens": [ + -5.628612995147705 + ], + "logprob_reward": -0.003613016800954938, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613016800954938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4325389862060547, + 0.2018737494945526, + 0.04769293591380119, + 0.032702405005693436, + 0.030464354902505875, + 0.10387926548719406, + 0.02631244994699955, + 0.032617226243019104, + 0.04798373579978943, + 0.04393487423658371 + ], + "credit_map": [ + 0.05246318131685257, + 0.04682403430342674, + 0.05587121099233627, + 0.05174138396978378, + 0.05435764789581299, + 0.05254558101296425, + 0.07380181550979614, + 0.04493415728211403, + 0.01924164593219757, + 0.025427639484405518, + 0.02959408611059189, + 0.01925554871559143, + 0.023052768781781197, + 0.028057927265763283, + 0.030694767832756042, + 0.02654937468469143, + 0.005142337176948786, + 0.0091543635353446, + 0.005598640069365501, + 0.007730324752628803, + 0.004039016552269459, + 0.003950999118387699, + 0.007831187918782234, + 0.0042460705153644085, + 0.004281527362763882, + 0.0034599979408085346, + 0.006218885537236929, + 0.0034335171803832054, + 0.0038572540506720543, + 0.003574231406673789, + 0.004417112562805414, + 0.0034598764032125473, + 0.0033491537906229496, + 0.0035502398386597633, + 0.0032310227397829294, + 0.0030990159139037132, + 0.004429350141435862, + 0.005646122619509697, + 0.003543505212292075, + 0.003615943482145667, + 0.006906358525156975, + 0.00995558500289917, + 0.011220406740903854, + 0.019012851640582085, + 0.013310007750988007, + 0.007599890232086182, + 0.02194237895309925, + 0.013931789435446262, + 0.003531248774379492, + 0.0022629019804298878, + 0.0034484625793993473, + 0.002767878584563732, + 0.0036115790717303753, + 0.0030448725447058678, + 0.004432359244674444, + 0.003213148331269622, + 0.0023578971158713102, + 0.004822241608053446, + 0.0038787571247667074, + 0.003406996838748455, + 0.004519899375736713, + 0.004986576270312071, + 0.0037791584618389606, + 0.004865699913352728, + 0.007603617385029793, + 0.00580644141882658, + 0.006969460751861334, + 0.007591372821480036, + 0.006327548995614052, + 0.005004385486245155, + 0.004168478772044182, + 0.004512431100010872, + 0.00345971598289907, + 0.004720630124211311, + 0.0033743754029273987, + 0.007303002290427685, + 0.005295832175761461, + 0.004850720055401325, + 0.009529342874884605, + 0.005401257891207933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.003613016800954938 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a dramatic low angle, the camera remains still, gazing into a dark, smoke-filled void, where elegant white lettering hangs suspended in absolute darkness. Crafted in a sharp, classic serif font with a clean, matte finish, the letters are illuminated by a single, harsh keylight from an unseen source, creating the stark, high-contrast black-and-white contrast characteristic of classic film noir. The typography executes a slow, hypnotic torsional pendulum oscillation, deliberately twisting back and forth on invisible vertical lines. As the letters rotate, harsh light hits their clean edges and flat planes, while their sharp, elongated shadows are cast on the slowly billowing smoke behind them, dancing and contorting ominously in a silent, menacing ballet, imbuing the image with a palpable tension and mystery.", + "seed": 28030, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_12/index_181_seed_28030_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.18569052256643773, + "vlm_logprob": -0.18569052256643773, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.18569052256643773, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10383053123950958 + ], + "opposite_logprob_tokens": [ + -3.228830575942993 + ], + "logprob_reward": -0.10383053123950958, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.10383053123950958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1608966886997223, + 0.06862729787826538, + 0.13762009143829346, + 0.2116759866476059, + 0.04147348180413246, + 0.07636275887489319, + 0.0678129494190216, + 0.03839273378252983, + 0.036741532385349274, + 0.1603964865207672 + ], + "credit_map": [ + 0.022954894229769707, + 0.012795214541256428, + 0.022154338657855988, + 0.025247730314731598, + 0.019113689661026, + 0.015457834117114544, + 0.02299887128174305, + 0.020174123346805573, + 0.007267913315445185, + 0.006079991813749075, + 0.009564616717398167, + 0.009929695166647434, + 0.006355154328048229, + 0.008057343773543835, + 0.009014157578349113, + 0.012358427047729492, + 0.013735475018620491, + 0.016911054030060768, + 0.016162490472197533, + 0.030269324779510498, + 0.011262962594628334, + 0.009597072377800941, + 0.02222154103219509, + 0.0174601748585701, + 0.015993038192391396, + 0.02216215245425701, + 0.034675560891628265, + 0.048397600650787354, + 0.018297486007213593, + 0.02322573773562908, + 0.027073223143815994, + 0.021851172670722008, + 0.0032030099537223577, + 0.00471745477989316, + 0.004453687928617001, + 0.004652975592762232, + 0.007025539875030518, + 0.004378983750939369, + 0.007686594035476446, + 0.0053552365861833096, + 0.007063696160912514, + 0.007528898771852255, + 0.007385743781924248, + 0.009030134417116642, + 0.00778333842754364, + 0.010483581572771072, + 0.013633228838443756, + 0.013454136438667774, + 0.004753548186272383, + 0.007045931648463011, + 0.01001302246004343, + 0.009848889894783497, + 0.006743964739143848, + 0.0035023551899939775, + 0.01574845239520073, + 0.010156787000596523, + 0.0029278555884957314, + 0.003317339112982154, + 0.005062687676399946, + 0.006302446126937866, + 0.0035989084281027317, + 0.003503990126773715, + 0.008328857831656933, + 0.005350648425519466, + 0.004240596666932106, + 0.0028271167539060116, + 0.0051891254261136055, + 0.004640472587198019, + 0.004921524319797754, + 0.004194814711809158, + 0.0059443372301757336, + 0.00478354562073946, + 0.007541880942881107, + 0.011321071535348892, + 0.03775688260793686, + 0.04204821214079857, + 0.007974725216627121, + 0.007187907118350267, + 0.022761397063732147, + 0.023804420605301857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.10383053123950958 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009814225137233734 + ], + "opposite_logprob_tokens": [ + -4.634814262390137 + ], + "logprob_reward": -0.009814225137233734, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009814225137233734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2391010969877243, + 0.11104989051818848, + 0.0501745380461216, + 0.29605019092559814, + 0.03969667851924896, + 0.027856558561325073, + 0.0312613882124424, + 0.13264013826847076, + 0.020278314128518105, + 0.051891207695007324 + ], + "credit_map": [ + 0.03077797405421734, + 0.018353356048464775, + 0.03160446882247925, + 0.03911624476313591, + 0.02723780833184719, + 0.026699330657720566, + 0.03541410341858864, + 0.029897816479206085, + 0.014460466802120209, + 0.00617577088996768, + 0.018839171156287193, + 0.01125229150056839, + 0.0066109346225857735, + 0.010893278755247593, + 0.019345130771398544, + 0.023472853004932404, + 0.004744317382574081, + 0.0049171834252774715, + 0.006014702841639519, + 0.006799693685024977, + 0.004594212863594294, + 0.005610216408967972, + 0.009906359016895294, + 0.007587849628180265, + 0.025798115879297256, + 0.01403461117297411, + 0.08648901432752609, + 0.026330044493079185, + 0.013194269500672817, + 0.018978802487254143, + 0.018005860969424248, + 0.09321948140859604, + 0.0028367494232952595, + 0.0039581903256475925, + 0.0037736305966973305, + 0.005268523935228586, + 0.0024613423738628626, + 0.007437107618898153, + 0.00553726963698864, + 0.008423864841461182, + 0.0024454020895063877, + 0.0027071109507232904, + 0.002889110706746578, + 0.0029103741981089115, + 0.00280592730268836, + 0.0039526005275547504, + 0.0047998386435210705, + 0.0053461953066289425, + 0.0017464709235355258, + 0.0026907301507890224, + 0.004163593053817749, + 0.005565294995903969, + 0.0029974374920129776, + 0.0018992352997884154, + 0.008576160296797752, + 0.0036224666982889175, + 0.009703600779175758, + 0.011575774289667606, + 0.021147919818758965, + 0.016350114718079567, + 0.01358407735824585, + 0.011021200567483902, + 0.029179900884628296, + 0.020077554509043694, + 0.002134754089638591, + 0.0015724219847470522, + 0.002942451508715749, + 0.0022323185112327337, + 0.002743466291576624, + 0.0025854844134300947, + 0.0033781067468225956, + 0.002689309883862734, + 0.002308692317456007, + 0.00369256641715765, + 0.009177441708743572, + 0.008322680369019508, + 0.003757827915251255, + 0.004160978831350803, + 0.0055608381517231464, + 0.014910179190337658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.009814225137233734 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578008770942688 + ], + "opposite_logprob_tokens": [ + -0.6328008770942688 + ], + "logprob_reward": -0.7578008770942688, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578008770942688, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16154684126377106, + 0.13184791803359985, + 0.25489160418510437, + 0.14605548977851868, + 0.01476673036813736, + 0.01701781526207924, + 0.0119622265920043, + 0.1494055539369583, + 0.014304655604064465, + 0.09820110350847244 + ], + "credit_map": [ + 0.018437586724758148, + 0.013006831519305706, + 0.022868694737553596, + 0.02797120064496994, + 0.01982569694519043, + 0.015880266204476357, + 0.02237563766539097, + 0.02118094265460968, + 0.01577296294271946, + 0.0119610745459795, + 0.01506778970360756, + 0.01957821287214756, + 0.011504601687192917, + 0.017363078892230988, + 0.02294054627418518, + 0.017659654840826988, + 0.02258407324552536, + 0.04150490090250969, + 0.028482193127274513, + 0.040673550218343735, + 0.025971338152885437, + 0.028980521485209465, + 0.037304531782865524, + 0.029390517622232437, + 0.010088621638715267, + 0.007773445453494787, + 0.03528652712702751, + 0.010031615383923054, + 0.011425131931900978, + 0.014901145361363888, + 0.011530897580087185, + 0.045018114149570465, + 0.0013774639228358865, + 0.0017749828984960914, + 0.0014000978553667665, + 0.001897654146887362, + 0.0024220452178269625, + 0.0017105159349739552, + 0.0023737167939543724, + 0.0018102553440257907, + 0.0020604508463293314, + 0.0012607958633452654, + 0.0017234133556485176, + 0.0015748351579532027, + 0.0023183494340628386, + 0.002863500965759158, + 0.0021717501804232597, + 0.00304472167044878, + 0.0008462302503176033, + 0.0010755133116617799, + 0.0015744582051411271, + 0.002068450441583991, + 0.0012255270266905427, + 0.000949729816056788, + 0.0023165575694292784, + 0.0019057615427300334, + 0.010916263796389103, + 0.00831773690879345, + 0.023716913536190987, + 0.016733242198824883, + 0.007810261100530624, + 0.012399651110172272, + 0.047241173684597015, + 0.022270333021879196, + 0.0015711970627307892, + 0.001030292478390038, + 0.0021999201271682978, + 0.001874606474302709, + 0.0014119212282821536, + 0.0014926313888281584, + 0.002264687092974782, + 0.0024594017304480076, + 0.005500460509210825, + 0.007439275272190571, + 0.01571805402636528, + 0.01643049716949463, + 0.007080814801156521, + 0.008162104524672031, + 0.010712621733546257, + 0.027157284319400787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.7578008770942688 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033713314682245255 + ], + "opposite_logprob_tokens": [ + -3.4087133407592773 + ], + "logprob_reward": -0.033713314682245255, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033713314682245255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30212023854255676, + 0.09231387078762054, + 0.0573296919465065, + 0.21695592999458313, + 0.037013426423072815, + 0.040353503078222275, + 0.03222242370247841, + 0.06988749653100967, + 0.05881715193390846, + 0.09298627823591232 + ], + "credit_map": [ + 0.035808537155389786, + 0.02284955605864525, + 0.03964286670088768, + 0.050098616629838943, + 0.034920718520879745, + 0.03217236325144768, + 0.04931344836950302, + 0.03731413930654526, + 0.007141447160393, + 0.00622418150305748, + 0.013912259601056576, + 0.012286687269806862, + 0.007624596823006868, + 0.005913503933697939, + 0.022875970229506493, + 0.016335222870111465, + 0.004924883600324392, + 0.00856383703649044, + 0.006924813613295555, + 0.009618490934371948, + 0.005488349124789238, + 0.007216753903776407, + 0.008074895478785038, + 0.0065176719799637794, + 0.013123164884746075, + 0.008250182494521141, + 0.06306929886341095, + 0.013965443708002567, + 0.011499154381453991, + 0.01795644313097, + 0.016591545194387436, + 0.07250069826841354, + 0.004059535451233387, + 0.0039757913909852505, + 0.004292176570743322, + 0.004986450541764498, + 0.0035483629908412695, + 0.005085715092718601, + 0.005935164634138346, + 0.005130225792527199, + 0.006127310451120138, + 0.0037975686136633158, + 0.005010924767702818, + 0.00370965083129704, + 0.005599344614893198, + 0.004938654135912657, + 0.005468202754855156, + 0.005701844580471516, + 0.0022791035007685423, + 0.002714062575250864, + 0.004306379705667496, + 0.0047985706478357315, + 0.002981330268085003, + 0.0033446806482970715, + 0.007334599737077951, + 0.004463700111955404, + 0.0036098093260079622, + 0.00635853037238121, + 0.010032734833657742, + 0.012303047813475132, + 0.006488619837909937, + 0.005474487319588661, + 0.015082069672644138, + 0.010538198985159397, + 0.005245720036327839, + 0.0044766380451619625, + 0.00842906255275011, + 0.010084605775773525, + 0.005949882324784994, + 0.007022614125162363, + 0.008727571927011013, + 0.008881053887307644, + 0.004153579473495483, + 0.007002648431807756, + 0.015668142586946487, + 0.015383221209049225, + 0.007183121517300606, + 0.007763733621686697, + 0.010454855859279633, + 0.025376977398991585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.033713314682245255 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and remains still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are elegant white letters visible hanging suspended in a dark, smoke-filled void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the letters use a sharp, classic serif font with a clean, matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the letters execute a slow, hypnotic torsional pendulum oscillation (twisting back and forth)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene characterized by high-contrast black-and-white lighting with harsh shadows cast on billowing smoke?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023293664678931236 + ], + "opposite_logprob_tokens": [ + -3.7732937335968018 + ], + "logprob_reward": -0.023293664678931236, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023293664678931236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23644360899925232, + 0.09693935513496399, + 0.10362895578145981, + 0.2812054455280304, + 0.017409145832061768, + 0.05355888977646828, + 0.020092906430363655, + 0.10583653301000595, + 0.03695318102836609, + 0.047931965440511703 + ], + "credit_map": [ + 0.028066031634807587, + 0.01910831406712532, + 0.03059741109609604, + 0.04194573685526848, + 0.026821894571185112, + 0.023814231157302856, + 0.0377783477306366, + 0.02831164374947548, + 0.011353282257914543, + 0.007673810701817274, + 0.015338707715272903, + 0.010650181211531162, + 0.006889035925269127, + 0.0068486821837723255, + 0.018448922783136368, + 0.01973673328757286, + 0.008608668111264706, + 0.02095085196197033, + 0.011460109613835812, + 0.014767548069357872, + 0.011738166213035583, + 0.010403926484286785, + 0.014111358672380447, + 0.011588329449295998, + 0.03516703099012375, + 0.023238569498062134, + 0.0658302754163742, + 0.03911779448390007, + 0.017385587096214294, + 0.019681179895997047, + 0.023411879315972328, + 0.05737312510609627, + 0.0016723808366805315, + 0.0023710995446890593, + 0.0017441111849620938, + 0.0025857584550976753, + 0.0025697017554193735, + 0.0019505422096699476, + 0.0023015812039375305, + 0.002213971456512809, + 0.0060079023241996765, + 0.005275927018374205, + 0.0074051315896213055, + 0.005552648101001978, + 0.004730693064630032, + 0.007063612807542086, + 0.0077058435417711735, + 0.009817130863666534, + 0.0010379520244896412, + 0.0017090695910155773, + 0.002523839706555009, + 0.003250752342864871, + 0.0015439105918630958, + 0.0015922555467113853, + 0.004510330967605114, + 0.003924795426428318, + 0.007536535616964102, + 0.0069124712608754635, + 0.01695338822901249, + 0.014242327772080898, + 0.007572825066745281, + 0.011084107682108879, + 0.026492631062865257, + 0.015042249113321304, + 0.003365233074873686, + 0.002530933590605855, + 0.0057039447128772736, + 0.004494336433708668, + 0.003872078377753496, + 0.004408676642924547, + 0.006427315063774586, + 0.006150664296001196, + 0.0045443191193044186, + 0.004392901435494423, + 0.007287667132914066, + 0.009560966864228249, + 0.003084113821387291, + 0.004716111812740564, + 0.00712123466655612, + 0.007224648259580135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.023293664678931236 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_13/index_tensor([2163, 2163])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_13/index_tensor([2163, 2163])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..aa2404b615ac412d6e9bae44d74cf37dfffe86e3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_13/index_tensor([2163, 2163])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic medium shot, captured from a low, upward angle in Annie Leibovitz's signature style, imbues the image with intense dramatic tension. Amidst a pouring rain, a chiseled man calmly steps from the driver's seat of a classic black sedan. He wears a long, charcoal-gray trench coat, the collar neatly raised to protect against the chill. The damp fabric, illuminated by a cold, off-screen streetlight, reflects a deep, textured sheen, sculpting the lines of his shoulders. Dense raindrops are clearly visible in the light, and his dark hair, slightly damp, a few strands plastering to his forehead. His resolute gaze, a mixture of fatigue and determination, passes the lens, gazing into the unknown darkness ahead. One hand remains on the open car door for balance, the faint glow from the dashboard filtering through, casting a hint of warmth on his dark trousers. His other foot has just touched the ground, creating a subtle splash of water. The asphalt beneath our feet is completely soaked by rain, forming mirror-like puddles that reflect the blurred red and blue glow of the distant city's neon lights. The entire scene is dominated by cold tones, with a strong contrast between light and shadow, accurately capturing the characters' resolute gestures as they step into the rain, creating an atmosphere imbued with mystery, loneliness, and a sense of fate.", + "seed": 5756, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_13/index_2163_seed_5756_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.040113830380141734, + "vlm_logprob": -0.040113830380141734, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.040113830380141734, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14635851979255676 + ], + "opposite_logprob_tokens": [ + -2.7713584899902344 + ], + "logprob_reward": -0.14635851979255676, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.14635851979255676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10199425369501114, + 0.022733323276042938, + 0.058139823377132416, + 0.1268003135919571, + 0.3166280686855316, + 0.032174400985240936, + 0.07794667035341263, + 0.055929895490407944, + 0.1328670084476471, + 0.07478615641593933 + ], + "credit_map": [ + 0.009531531482934952, + 0.007374826353043318, + 0.009227833710610867, + 0.017088618129491806, + 0.007159994915127754, + 0.016594842076301575, + 0.01980864629149437, + 0.015207960270345211, + 0.0021041573490947485, + 0.0022062216885387897, + 0.002071656286716461, + 0.0054096924141049385, + 0.0019502780633047223, + 0.0021357196383178234, + 0.0038883830420672894, + 0.0029672146774828434, + 0.0030561843886971474, + 0.0020162996370345354, + 0.003010716987773776, + 0.004453121218830347, + 0.004989721346646547, + 0.002554169623181224, + 0.020014051347970963, + 0.01804555580019951, + 0.013169332407414913, + 0.043999556452035904, + 0.004873332101851702, + 0.008567551150918007, + 0.006926068104803562, + 0.004924322944134474, + 0.0381973460316658, + 0.006142809987068176, + 0.017313852906227112, + 0.010089773684740067, + 0.010516715236008167, + 0.009929783642292023, + 0.12424414604902267, + 0.016928302124142647, + 0.11220907419919968, + 0.01539640687406063, + 0.0035966273862868547, + 0.003158689010888338, + 0.003635494504123926, + 0.0035110448952764273, + 0.004309475421905518, + 0.0031425359193235636, + 0.006936934776604176, + 0.0038835983723402023, + 0.008729524910449982, + 0.007151077035814524, + 0.012152730487287045, + 0.009138445369899273, + 0.012126320041716099, + 0.007795417681336403, + 0.01048927940428257, + 0.010363876819610596, + 0.007600355893373489, + 0.007625111378729343, + 0.007834556512534618, + 0.005700191017240286, + 0.00827703159302473, + 0.005628024227917194, + 0.007196309044957161, + 0.006068314425647259, + 0.01998753659427166, + 0.019085310399532318, + 0.01258112397044897, + 0.011503959074616432, + 0.014125226065516472, + 0.03581233322620392, + 0.010625123977661133, + 0.009146394208073616, + 0.011776159517467022, + 0.004315126687288284, + 0.0035680143628269434, + 0.002048665191978216, + 0.004895575810223818, + 0.028982575982809067, + 0.014777489006519318, + 0.004422556143254042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.14635851979255676 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008673022501170635 + ], + "opposite_logprob_tokens": [ + -4.758673191070557 + ], + "logprob_reward": -0.008673022501170635, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008673022501170635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14508122205734253, + 0.04472455009818077, + 0.025535693392157555, + 0.09115371108055115, + 0.08603201806545258, + 0.04042542353272438, + 0.10398534685373306, + 0.034260641783475876, + 0.06400899589061737, + 0.3647924065589905 + ], + "credit_map": [ + 0.01358872838318348, + 0.010351425968110561, + 0.016397571191191673, + 0.031139669939875603, + 0.0112718241289258, + 0.01227338146418333, + 0.025147365406155586, + 0.024911241605877876, + 0.004584333393722773, + 0.003356548026204109, + 0.004996991716325283, + 0.008920127525925636, + 0.0036435702349990606, + 0.004821114707738161, + 0.0075211478397250175, + 0.006880716420710087, + 0.0025765197351574898, + 0.0013468723045662045, + 0.002571315737441182, + 0.0031371780205518007, + 0.004022607114166021, + 0.002795735839754343, + 0.005136631894856691, + 0.003948829602450132, + 0.012910851277410984, + 0.007111383602023125, + 0.008204462006688118, + 0.01056207250803709, + 0.007920836098492146, + 0.010570230893790722, + 0.026427455246448517, + 0.00744640314951539, + 0.005870386958122253, + 0.003631750587373972, + 0.003039832226932049, + 0.0035927467979490757, + 0.032557833939790726, + 0.005107141099870205, + 0.027763240039348602, + 0.00446908175945282, + 0.0038295534905046225, + 0.0032327440567314625, + 0.002987049985677004, + 0.003931192681193352, + 0.009823635220527649, + 0.005152062512934208, + 0.007624615449458361, + 0.003844561753794551, + 0.010191254317760468, + 0.0073790657334029675, + 0.01999608427286148, + 0.012315735220909119, + 0.016842041164636612, + 0.011980690993368626, + 0.012622229754924774, + 0.012658236548304558, + 0.003700971370562911, + 0.0038483752869069576, + 0.002292357850819826, + 0.0030462206341326237, + 0.005917530041188002, + 0.0057231164537370205, + 0.004795729648321867, + 0.004936335142701864, + 0.007327886763960123, + 0.009012269787490368, + 0.005020411219447851, + 0.006512913387268782, + 0.015786724165081978, + 0.007581004872918129, + 0.007695252075791359, + 0.005072526168078184, + 0.03987168148159981, + 0.01666407659649849, + 0.012934204190969467, + 0.0081437723711133, + 0.020436160266399384, + 0.15242329239845276, + 0.0956292673945427, + 0.01868990808725357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.008673022501170635 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954147316515446 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.005954147316515446, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954147316515446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05708054080605507, + 0.02588970586657524, + 0.04540003091096878, + 0.15091615915298462, + 0.016001848503947258, + 0.015756381675601006, + 0.027781415730714798, + 0.02049054205417633, + 0.11142051219940186, + 0.529262900352478 + ], + "credit_map": [ + 0.005097640212625265, + 0.0035925882402807474, + 0.004896416794508696, + 0.010158001445233822, + 0.004921643063426018, + 0.009719514288008213, + 0.008873865939676762, + 0.009820866398513317, + 0.0043893661350011826, + 0.001934281550347805, + 0.002029139082878828, + 0.004872568417340517, + 0.001991124590858817, + 0.0028850738890469074, + 0.003384791547432542, + 0.004403360653668642, + 0.002291530603542924, + 0.0011101998388767242, + 0.0017874783370643854, + 0.0027409486938267946, + 0.0028818543069064617, + 0.0019848623778671026, + 0.017571328207850456, + 0.01503183227032423, + 0.015091531910002232, + 0.05352390184998512, + 0.005141741130501032, + 0.010196470655500889, + 0.007954906672239304, + 0.004995884373784065, + 0.04755178466439247, + 0.006459933239966631, + 0.0014333507278934121, + 0.001056026783771813, + 0.0012379700783640146, + 0.0011519514955580235, + 0.0050711059011518955, + 0.0016169106820598245, + 0.002935014432296157, + 0.0014995179371908307, + 0.001477368758060038, + 0.0015302877873182297, + 0.001569796702824533, + 0.0019074429292231798, + 0.0034626484848558903, + 0.0015691019361838698, + 0.00274540763348341, + 0.0014943273272365332, + 0.002373799681663513, + 0.0022320060525089502, + 0.003632311476394534, + 0.003236742690205574, + 0.0057796332985162735, + 0.0030615373980253935, + 0.0038426173850893974, + 0.0036227682139724493, + 0.0027284591924399137, + 0.0023852053564041853, + 0.0016670345794409513, + 0.0020872321911156178, + 0.0030425027944147587, + 0.002427179366350174, + 0.0039068483747541904, + 0.002246081130579114, + 0.01373229082673788, + 0.009772057645022869, + 0.0062975347973406315, + 0.00986382644623518, + 0.01991019770503044, + 0.025110116228461266, + 0.015840495005249977, + 0.010893989354372025, + 0.05477979779243469, + 0.021301737055182457, + 0.017224324867129326, + 0.012448875233530998, + 0.029591483995318413, + 0.2210247665643692, + 0.14684873819351196, + 0.02604314126074314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.005954147316515446 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980630472302437 + ], + "opposite_logprob_tokens": [ + -3.52980637550354 + ], + "logprob_reward": -0.02980630472302437, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980630472302437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08087422698736191, + 0.04334869235754013, + 0.03700925409793854, + 0.5713101625442505, + 0.08587738126516342, + 0.03703304007649422, + 0.03717231750488281, + 0.03262083977460861, + 0.045139726251363754, + 0.0296144001185894 + ], + "credit_map": [ + 0.008361341431736946, + 0.004518590401858091, + 0.0072502251714468, + 0.01911262609064579, + 0.006666249595582485, + 0.010482560843229294, + 0.01302994042634964, + 0.01145268976688385, + 0.007501024287194014, + 0.003003650577738881, + 0.00441536633297801, + 0.007265209220349789, + 0.002805733121931553, + 0.00505087012425065, + 0.00605907803401351, + 0.00724776042625308, + 0.0028635964263230562, + 0.0015379342949017882, + 0.003048251150175929, + 0.004429936874657869, + 0.005687616765499115, + 0.0032722270116209984, + 0.009061804041266441, + 0.007107886020094156, + 0.04775192216038704, + 0.18260733783245087, + 0.019804203882813454, + 0.04311385378241539, + 0.02818545512855053, + 0.019304106011986732, + 0.20325376093387604, + 0.02728951722383499, + 0.0049756974913179874, + 0.002105207182466984, + 0.0016243441496044397, + 0.0027593665290623903, + 0.034418631345033646, + 0.004899203777313232, + 0.030601007863879204, + 0.004493920598179102, + 0.0033753477036952972, + 0.0022968947887420654, + 0.003782030427828431, + 0.0030019015539437532, + 0.008390845730900764, + 0.005874419119209051, + 0.006461708806455135, + 0.0038498914800584316, + 0.003994581289589405, + 0.0035802642814815044, + 0.0026524339336901903, + 0.003936075605452061, + 0.008112253621220589, + 0.004970381036400795, + 0.006087377201765776, + 0.003838949603959918, + 0.0033362996764481068, + 0.003545320127159357, + 0.0027425806038081646, + 0.0040335021913051605, + 0.0032083841506391764, + 0.0055752890184521675, + 0.0065071010030806065, + 0.003672362770885229, + 0.005765525158494711, + 0.0037505379877984524, + 0.003332948312163353, + 0.0030118206050246954, + 0.006216026376932859, + 0.015726987272500992, + 0.004482244607061148, + 0.0028536356985569, + 0.006285435985773802, + 0.00231486139819026, + 0.0014308637473732233, + 0.0016537425108253956, + 0.005380092654377222, + 0.003971036057919264, + 0.006366371177136898, + 0.0022119965869933367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.02980630472302437 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777157567441463 + ], + "opposite_logprob_tokens": [ + -4.634777069091797 + ], + "logprob_reward": -0.009777157567441463, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777157567441463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12266834825277328, + 0.05660075321793556, + 0.18428713083267212, + 0.35761559009552, + 0.028442101553082466, + 0.0722067579627037, + 0.02959958277642727, + 0.06399445235729218, + 0.06021145358681679, + 0.02437388151884079 + ], + "credit_map": [ + 0.010941346175968647, + 0.00623992457985878, + 0.009633773006498814, + 0.01726352609694004, + 0.00878871139138937, + 0.013859725557267666, + 0.022237302735447884, + 0.033704038709402084, + 0.007935881614685059, + 0.005932661704719067, + 0.005558797158300877, + 0.008327731862664223, + 0.005687706172466278, + 0.005534797441214323, + 0.008124095387756824, + 0.009499079547822475, + 0.010447829030454159, + 0.004169912543147802, + 0.0070665134117007256, + 0.010873681865632534, + 0.010999745689332485, + 0.009079012088477612, + 0.07054159790277481, + 0.06110882759094238, + 0.03128591552376747, + 0.12020306289196014, + 0.011870959773659706, + 0.02615608461201191, + 0.018358193337917328, + 0.009577549062669277, + 0.12339822947978973, + 0.016765547916293144, + 0.0029276248533278704, + 0.0018040603026747704, + 0.0026901611126959324, + 0.0025669708847999573, + 0.00641582952812314, + 0.0027221289929002523, + 0.006429894827306271, + 0.002885432681068778, + 0.007200151681900024, + 0.008305838331580162, + 0.009706630371510983, + 0.006440117955207825, + 0.0120312524959445, + 0.006905080284923315, + 0.0130221638828516, + 0.008595523424446583, + 0.0024645391386002302, + 0.0025706856977194548, + 0.003270077984780073, + 0.0035590450279414654, + 0.006872888654470444, + 0.003278805175796151, + 0.003911452367901802, + 0.0036720875650644302, + 0.006817528046667576, + 0.008237150497734547, + 0.007537081837654114, + 0.009762969799339771, + 0.006492148153483868, + 0.006577648222446442, + 0.010982170701026917, + 0.007587750442326069, + 0.0067086778581142426, + 0.01164794061332941, + 0.0057458579540252686, + 0.005498463287949562, + 0.005787929054349661, + 0.014449221082031727, + 0.005095789674669504, + 0.005277575459331274, + 0.0037222718819975853, + 0.0018651741556823254, + 0.002374372910708189, + 0.0017400850774720311, + 0.0032719774171710014, + 0.003372503910213709, + 0.0059600817039608955, + 0.0020674136467278004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.009777157567441463 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic medium shot, captured from a low, upward angle in Annie Leibovitz's signature style, imbues the image with intense dramatic tension. Amidst a pouring rain, a chiseled man calmly steps from the driver's seat of a classic black sedan. He wears a long, charcoal-gray trench coat, the collar neatly raised to protect against the chill. The damp fabric, illuminated by a cold, off-screen streetlight, reflects a deep, textured sheen, sculpting the lines of his shoulders. Dense raindrops are clearly visible in the light, and his dark hair, slightly damp, a few strands plastering to his forehead. His resolute gaze, a mixture of fatigue and determination, passes the lens, gazing into the unknown darkness ahead. One hand remains on the open car door for balance, the faint glow from the dashboard filtering through, casting a hint of warmth on his dark trousers. His other foot has just touched the ground, creating a subtle splash of water. The asphalt beneath our feet is completely soaked by rain, forming mirror-like puddles that reflect the blurred red and blue glow of the distant city's neon lights. The entire scene is dominated by cold tones, with a strong contrast between light and shadow, accurately capturing the characters' resolute gestures as they step into the rain, creating an atmosphere imbued with mystery, loneliness, and a sense of fate.", + "seed": 5757, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_13/index_2163_seed_5757_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.028037882503122093, + "vlm_logprob": -0.028037882503122093, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.028037882503122093, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09587552398443222 + ], + "opposite_logprob_tokens": [ + -2.7208755016326904 + ], + "logprob_reward": -0.09587552398443222, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09587552398443222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09786529093980789, + 0.3675905764102936, + 0.05325149744749069, + 0.09392423182725906, + 0.03137931227684021, + 0.1270277053117752, + 0.08118769526481628, + 0.07501254975795746, + 0.030350472778081894, + 0.04241067171096802 + ], + "credit_map": [ + 0.008640976622700691, + 0.009343484416604042, + 0.014587645418941975, + 0.01960977353155613, + 0.006775206886231899, + 0.01063609030097723, + 0.010992289520800114, + 0.01727982610464096, + 0.015337876044213772, + 0.012524902820587158, + 0.10685037821531296, + 0.16348852217197418, + 0.008447309024631977, + 0.008283333852887154, + 0.024046722799539566, + 0.02861151657998562, + 0.006426070351153612, + 0.00556257925927639, + 0.006336333230137825, + 0.006151551380753517, + 0.006052359938621521, + 0.004832414444535971, + 0.010151798836886883, + 0.0077383918687701225, + 0.012720927596092224, + 0.00834446121007204, + 0.004117720294743776, + 0.0039260126650333405, + 0.008424059487879276, + 0.025135161355137825, + 0.024858511984348297, + 0.006397382356226444, + 0.006961151957511902, + 0.0030546023044735193, + 0.0020275136921554804, + 0.0028752919752150774, + 0.003942443523555994, + 0.0033870514016598463, + 0.005335918627679348, + 0.0037953394930809736, + 0.012220172211527824, + 0.014597075060009956, + 0.01412589754909277, + 0.02050572820007801, + 0.012969001196324825, + 0.0113027673214674, + 0.021022986620664597, + 0.02028409019112587, + 0.009384599514305592, + 0.009838241152465343, + 0.005353064276278019, + 0.01909065432846546, + 0.005206335801631212, + 0.013281535357236862, + 0.01174522191286087, + 0.007288035936653614, + 0.016828952357172966, + 0.008321375586092472, + 0.006749094929546118, + 0.012878485955297947, + 0.006562324240803719, + 0.007136696018278599, + 0.008916780352592468, + 0.007618837058544159, + 0.004836059641093016, + 0.0034746776800602674, + 0.002765535842627287, + 0.003581041470170021, + 0.005231922492384911, + 0.0036565240006893873, + 0.003233250230550766, + 0.003571462817490101, + 0.005723391659557819, + 0.004615181125700474, + 0.004072541370987892, + 0.004357035271823406, + 0.007085489109158516, + 0.004063096363097429, + 0.007130520883947611, + 0.005363414064049721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.09587552398443222 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006761173252016306 + ], + "opposite_logprob_tokens": [ + -5.006761074066162 + ], + "logprob_reward": -0.006761173252016306, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006761173252016306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1300010085105896, + 0.21606428921222687, + 0.11223874241113663, + 0.1903677135705948, + 0.024713272228837013, + 0.040670983493328094, + 0.07830428332090378, + 0.09961682558059692, + 0.0324355885386467, + 0.0755872055888176 + ], + "credit_map": [ + 0.009045146405696869, + 0.007283905521035194, + 0.03230930492281914, + 0.016413100063800812, + 0.010765611194074154, + 0.00925149954855442, + 0.012383525259792805, + 0.032548937946558, + 0.012064715847373009, + 0.013835505582392216, + 0.07432373613119125, + 0.04939718171954155, + 0.010260069742798805, + 0.010991568677127361, + 0.01770189218223095, + 0.02748965658247471, + 0.011364880949258804, + 0.012425391003489494, + 0.01407493557780981, + 0.013372151181101799, + 0.011518539860844612, + 0.014535048976540565, + 0.024861274287104607, + 0.010086548514664173, + 0.020145833492279053, + 0.018188025802373886, + 0.007567533291876316, + 0.006667952053248882, + 0.017714418470859528, + 0.05476351082324982, + 0.05159909650683403, + 0.01372138224542141, + 0.0049069952219724655, + 0.0021341799292713404, + 0.0018183356150984764, + 0.002209592144936323, + 0.004220240283757448, + 0.002451377920806408, + 0.003976209554821253, + 0.0029963445849716663, + 0.00432168273255229, + 0.0042101796716451645, + 0.002309650182723999, + 0.006451154127717018, + 0.006531247403472662, + 0.005001357290893793, + 0.00868703331798315, + 0.0031586873810738325, + 0.008618256077170372, + 0.015076951123774052, + 0.005627795122563839, + 0.012496940791606903, + 0.007335149683058262, + 0.010663683526217937, + 0.01201712153851986, + 0.0064683957025408745, + 0.02336922287940979, + 0.010895187966525555, + 0.008271198719739914, + 0.01620226539671421, + 0.00782740581780672, + 0.009619367308914661, + 0.01113951113075018, + 0.012292684987187386, + 0.005250600632280111, + 0.0026798187755048275, + 0.002852494129911065, + 0.0034351381473243237, + 0.006146852392703295, + 0.0026745968498289585, + 0.005119717679917812, + 0.00427637342363596, + 0.006599455140531063, + 0.006389440968632698, + 0.003933989442884922, + 0.006751867476850748, + 0.006945337634533644, + 0.0153757743537426, + 0.023596184328198433, + 0.00599517161026597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.006761173252016306 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024946779012680054 + ], + "opposite_logprob_tokens": [ + -6.002494812011719 + ], + "logprob_reward": -0.0024946779012680054, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024946779012680054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09139171242713928, + 0.12029752135276794, + 0.0755186602473259, + 0.15448682010173798, + 0.03002922050654888, + 0.17023685574531555, + 0.08397237211465836, + 0.05391858145594597, + 0.07464820891618729, + 0.145500048995018 + ], + "credit_map": [ + 0.007426868658512831, + 0.009893222711980343, + 0.015351742506027222, + 0.016297316178679466, + 0.01095068920403719, + 0.007148989476263523, + 0.0104598980396986, + 0.01386298518627882, + 0.008867904543876648, + 0.006715650204569101, + 0.023456308990716934, + 0.04455683380365372, + 0.006457955576479435, + 0.005554449278861284, + 0.0102898133918643, + 0.0143986064940691, + 0.006809755694121122, + 0.008686540648341179, + 0.008445018902420998, + 0.008543216623365879, + 0.008062873035669327, + 0.013440033420920372, + 0.012846150435507298, + 0.008685074746608734, + 0.01700042188167572, + 0.014306116849184036, + 0.005722967907786369, + 0.005008827429264784, + 0.013679973781108856, + 0.04547971859574318, + 0.04303881525993347, + 0.010249972343444824, + 0.005095009692013264, + 0.0027937914710491896, + 0.0028794927056878805, + 0.003403813811019063, + 0.004338953644037247, + 0.002940164413303137, + 0.004616165068000555, + 0.003961830399930477, + 0.013684096746146679, + 0.015073129907250404, + 0.013063987717032433, + 0.039453305304050446, + 0.020922217518091202, + 0.02296029031276703, + 0.030465565621852875, + 0.014614266343414783, + 0.006402225699275732, + 0.010955609381198883, + 0.007158525288105011, + 0.016769802197813988, + 0.009999170899391174, + 0.00875673908740282, + 0.019224094226956367, + 0.004706206265836954, + 0.009758000262081623, + 0.008192261680960655, + 0.004610815551131964, + 0.00809154286980629, + 0.006023479625582695, + 0.0032102137338370085, + 0.008890265598893166, + 0.0051420037634670734, + 0.006263722665607929, + 0.007235275115817785, + 0.003997241612523794, + 0.008871637284755707, + 0.015303867869079113, + 0.009346483275294304, + 0.014721923507750034, + 0.008908062241971493, + 0.01322263479232788, + 0.011815613135695457, + 0.006670273840427399, + 0.0112450011074543, + 0.012368950061500072, + 0.03233815357089043, + 0.04755892977118492, + 0.010280493646860123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 9 + ], + "selected_reward": -0.0024946779012680054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029808156192302704 + ], + "opposite_logprob_tokens": [ + -3.5298080444335938 + ], + "logprob_reward": -0.029808156192302704, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029808156192302704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08690077811479568, + 0.17513282597064972, + 0.0950046256184578, + 0.08050960302352905, + 0.0320231169462204, + 0.19317467510700226, + 0.0783219262957573, + 0.1663242131471634, + 0.02925170212984085, + 0.06335647404193878 + ], + "credit_map": [ + 0.011465182527899742, + 0.007568905595690012, + 0.01472484041005373, + 0.014652058482170105, + 0.005886431317776442, + 0.008096861653029919, + 0.01180971134454012, + 0.01269678957760334, + 0.009284153580665588, + 0.009498550556600094, + 0.03149772062897682, + 0.06303538382053375, + 0.010386846959590912, + 0.007298886310309172, + 0.020362475886940956, + 0.023768825456500053, + 0.010249469429254532, + 0.008855694904923439, + 0.00973806157708168, + 0.012030869722366333, + 0.019046520814299583, + 0.0073414151556789875, + 0.017784254625439644, + 0.009958340786397457, + 0.009696346707642078, + 0.006796586327254772, + 0.007000035606324673, + 0.004904105793684721, + 0.008161080069839954, + 0.0146643640473485, + 0.017765389755368233, + 0.011521696113049984, + 0.006847284268587828, + 0.002776311943307519, + 0.002059138147160411, + 0.0026600484270602465, + 0.004798186477273703, + 0.003422030946239829, + 0.005669162608683109, + 0.003790957620367408, + 0.015612884424626827, + 0.015918897464871407, + 0.0151000851765275, + 0.04527810588479042, + 0.01973937824368477, + 0.02797083556652069, + 0.0360788032412529, + 0.017475709319114685, + 0.01673240214586258, + 0.0070140063762664795, + 0.00549189792945981, + 0.017955627292394638, + 0.005387939512729645, + 0.010552532970905304, + 0.010641324333846569, + 0.004546210169792175, + 0.046346500515937805, + 0.020049842074513435, + 0.010862111113965511, + 0.026486773043870926, + 0.013809526339173317, + 0.014015973545610905, + 0.019514016807079315, + 0.015239490196108818, + 0.005795832257717848, + 0.003824599552899599, + 0.0026063984259963036, + 0.0025177733041346073, + 0.005392186809331179, + 0.0025146743282675743, + 0.0034470364917069674, + 0.0031532018911093473, + 0.010103615000844002, + 0.00993689801543951, + 0.006545911077409983, + 0.007985948584973812, + 0.006120743229985237, + 0.006280702538788319, + 0.0078894617035985, + 0.008493197150528431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.029808156192302704 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249881185591221 + ], + "opposite_logprob_tokens": [ + -5.255249977111816 + ], + "logprob_reward": -0.005249881185591221, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249881185591221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08915546536445618, + 0.11615516245365143, + 0.05567843094468117, + 0.22732912003993988, + 0.01696142740547657, + 0.037011850625276566, + 0.11243069916963577, + 0.2873283624649048, + 0.018966492265462875, + 0.03898301348090172 + ], + "credit_map": [ + 0.007661065086722374, + 0.00727128004655242, + 0.015339541248977184, + 0.01560476515442133, + 0.007517009973526001, + 0.006307016592472792, + 0.008957164362072945, + 0.020497620105743408, + 0.006994002033025026, + 0.005689626093953848, + 0.022066893056035042, + 0.04502520337700844, + 0.005904414225369692, + 0.004335613455623388, + 0.01252311933785677, + 0.013616285286843777, + 0.0061330730095505714, + 0.009230360388755798, + 0.00721868546679616, + 0.007619393058121204, + 0.0046341316774487495, + 0.005581359378993511, + 0.010326103307306767, + 0.004935320932418108, + 0.023055383935570717, + 0.021032141521573067, + 0.009078409522771835, + 0.00837481115013361, + 0.01914685033261776, + 0.06645370274782181, + 0.06278642266988754, + 0.01740136928856373, + 0.004145276732742786, + 0.0015754271298646927, + 0.0012069841613993049, + 0.0016183260595425963, + 0.0021909920033067465, + 0.0018460049759596586, + 0.0028626315761357546, + 0.0015157840680330992, + 0.004593416582792997, + 0.0029598153196275234, + 0.0029478229116648436, + 0.008526751771569252, + 0.00315369525924325, + 0.004548217169940472, + 0.007003470323979855, + 0.0032786584924906492, + 0.021158186718821526, + 0.009284247644245625, + 0.006536767352372408, + 0.02802150323987007, + 0.0072699133306741714, + 0.01957022212445736, + 0.01466414425522089, + 0.005925709847360849, + 0.07416006177663803, + 0.03520350158214569, + 0.02547091618180275, + 0.043039679527282715, + 0.022028421983122826, + 0.0289436187595129, + 0.02800281159579754, + 0.03047931380569935, + 0.002843692200258374, + 0.0014721309999004006, + 0.0013190191239118576, + 0.0023567129392176867, + 0.0030841208063066006, + 0.0018769105663523078, + 0.0033936062827706337, + 0.0026202977169305086, + 0.004278381820768118, + 0.004710690584033728, + 0.0030151852406561375, + 0.0032263481989502907, + 0.0029559375252574682, + 0.008538677357137203, + 0.008961617015302181, + 0.0032961734104901552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.005249881185591221 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic medium shot, captured from a low, upward angle in Annie Leibovitz's signature style, imbues the image with intense dramatic tension. Amidst a pouring rain, a chiseled man calmly steps from the driver's seat of a classic black sedan. He wears a long, charcoal-gray trench coat, the collar neatly raised to protect against the chill. The damp fabric, illuminated by a cold, off-screen streetlight, reflects a deep, textured sheen, sculpting the lines of his shoulders. Dense raindrops are clearly visible in the light, and his dark hair, slightly damp, a few strands plastering to his forehead. His resolute gaze, a mixture of fatigue and determination, passes the lens, gazing into the unknown darkness ahead. One hand remains on the open car door for balance, the faint glow from the dashboard filtering through, casting a hint of warmth on his dark trousers. His other foot has just touched the ground, creating a subtle splash of water. The asphalt beneath our feet is completely soaked by rain, forming mirror-like puddles that reflect the blurred red and blue glow of the distant city's neon lights. The entire scene is dominated by cold tones, with a strong contrast between light and shadow, accurately capturing the characters' resolute gestures as they step into the rain, creating an atmosphere imbued with mystery, loneliness, and a sense of fate.", + "seed": 5758, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_13/index_2163_seed_5758_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05245730271562934, + "vlm_logprob": -0.05245730271562934, + "vlm_yes_no_margin": 3.950000023841858, + "vlm_token_credit": -0.05245730271562934, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20811931788921356 + ], + "opposite_logprob_tokens": [ + -2.0831193923950195 + ], + "logprob_reward": -0.20811931788921356, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.20811931788921356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1698828488588333, + 0.1296100914478302, + 0.169941708445549, + 0.0811714306473732, + 0.05449490621685982, + 0.06820407509803772, + 0.07639630138874054, + 0.09588351845741272, + 0.051037516444921494, + 0.10337767004966736 + ], + "credit_map": [ + 0.015394306741654873, + 0.015627749264240265, + 0.031588394194841385, + 0.04058131203055382, + 0.012969058007001877, + 0.013621700927615166, + 0.02544597163796425, + 0.014654349535703659, + 0.011603010818362236, + 0.008393339812755585, + 0.02055121585726738, + 0.02206224575638771, + 0.01060666237026453, + 0.009661387652158737, + 0.03748593479394913, + 0.00924629159271717, + 0.015003498643636703, + 0.01253960095345974, + 0.02623482421040535, + 0.025104105472564697, + 0.025201968848705292, + 0.019680781289935112, + 0.031634844839572906, + 0.014542083255946636, + 0.007755205500870943, + 0.005901300814002752, + 0.009787382557988167, + 0.005713494028896093, + 0.010990955866873264, + 0.006978014018386602, + 0.030010225251317024, + 0.004034846555441618, + 0.004884463269263506, + 0.0056410059332847595, + 0.004128163680434227, + 0.007486347574740648, + 0.011241409927606583, + 0.007145717274397612, + 0.008675462566316128, + 0.005292336922138929, + 0.007061566226184368, + 0.012959993444383144, + 0.007673953659832478, + 0.007306620012968779, + 0.007395484484732151, + 0.008010738529264927, + 0.00957574974745512, + 0.008219967596232891, + 0.00602654879912734, + 0.00865317601710558, + 0.00902826152741909, + 0.007883463054895401, + 0.010746613144874573, + 0.014671525917947292, + 0.010892216116189957, + 0.00849449448287487, + 0.007626831531524658, + 0.011497005820274353, + 0.01339977141469717, + 0.017412757501006126, + 0.01089728157967329, + 0.016036346554756165, + 0.007985444739460945, + 0.011028082109987736, + 0.006161798257380724, + 0.007610759697854519, + 0.0042648050002753735, + 0.005831640213727951, + 0.008788025006651878, + 0.006506060250103474, + 0.003994333557784557, + 0.007880092598497868, + 0.01594921387732029, + 0.009357905015349388, + 0.008253658190369606, + 0.008450743742287159, + 0.010131112299859524, + 0.022331465035676956, + 0.010304735042154789, + 0.01859883964061737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.20811931788921356 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007660772651433945 + ], + "opposite_logprob_tokens": [ + -4.882660865783691 + ], + "logprob_reward": -0.007660772651433945, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007660772651433945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1831279993057251, + 0.1576622575521469, + 0.22819742560386658, + 0.10724423825740814, + 0.0397786870598793, + 0.0415121428668499, + 0.0516875758767128, + 0.05293549597263336, + 0.05434505268931389, + 0.08350915461778641 + ], + "credit_map": [ + 0.024144602939486504, + 0.016080303117632866, + 0.029147600755095482, + 0.03776910528540611, + 0.01215131301432848, + 0.02192537859082222, + 0.024818604812026024, + 0.01709107868373394, + 0.014933180063962936, + 0.01274696085602045, + 0.02859438583254814, + 0.025876622647047043, + 0.012285439297556877, + 0.011792657896876335, + 0.03891630470752716, + 0.012516706250607967, + 0.01590345986187458, + 0.012440372258424759, + 0.03818754106760025, + 0.044275760650634766, + 0.028753718361258507, + 0.019394028931856155, + 0.0537230484187603, + 0.015519489534199238, + 0.011769207194447517, + 0.007832011207938194, + 0.013031745329499245, + 0.008980022743344307, + 0.012429245747625828, + 0.008590806275606155, + 0.039317596703767776, + 0.005293592810630798, + 0.002507175784558058, + 0.004478766117244959, + 0.0029825386591255665, + 0.0059603736735880375, + 0.008604662492871284, + 0.005985217168927193, + 0.006443454883992672, + 0.0028164999093860388, + 0.004841143731027842, + 0.00902883056551218, + 0.005427106283605099, + 0.004352134186774492, + 0.004003418143838644, + 0.005616901908069849, + 0.0043021151795983315, + 0.003940492868423462, + 0.004717299249023199, + 0.005896156188100576, + 0.006134480237960815, + 0.006174636073410511, + 0.006862678565084934, + 0.007685831282287836, + 0.006038620602339506, + 0.008177878335118294, + 0.003270740620791912, + 0.009090879932045937, + 0.006149831227958202, + 0.0071402983739972115, + 0.00967074278742075, + 0.006428558845072985, + 0.003927284851670265, + 0.007257158402353525, + 0.010270562954246998, + 0.004781767725944519, + 0.003609588136896491, + 0.004644095432013273, + 0.012514437548816204, + 0.010673095472157001, + 0.002689337357878685, + 0.005162166431546211, + 0.009497113525867462, + 0.006317138206213713, + 0.005945506971329451, + 0.004732904490083456, + 0.006817677989602089, + 0.0246571097522974, + 0.006666607689112425, + 0.018875082954764366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.007660772651433945 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595545070245862 + ], + "opposite_logprob_tokens": [ + -5.130955219268799 + ], + "logprob_reward": -0.00595545070245862, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595545070245862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17048092186450958, + 0.16947311162948608, + 0.15924440324306488, + 0.10666533559560776, + 0.06597089767456055, + 0.028756560757756233, + 0.039998795837163925, + 0.06344559788703918, + 0.03565794974565506, + 0.16030637919902802 + ], + "credit_map": [ + 0.02218937687575817, + 0.01557738333940506, + 0.027589136734604836, + 0.029804542660713196, + 0.009115652181208134, + 0.025020867586135864, + 0.02524849772453308, + 0.015935475006699562, + 0.008382349275052547, + 0.003869200823828578, + 0.05828333646059036, + 0.013063838705420494, + 0.006704026833176613, + 0.0057117571122944355, + 0.06710490584373474, + 0.0063536930829286575, + 0.012822756543755531, + 0.009130611084401608, + 0.024987025186419487, + 0.03209756314754486, + 0.028452781960368156, + 0.014498666860163212, + 0.022820167243480682, + 0.014434817247092724, + 0.008062136359512806, + 0.009247099980711937, + 0.01198433618992567, + 0.016206782311201096, + 0.014196337200701237, + 0.008451993577182293, + 0.03068755567073822, + 0.007829100824892521, + 0.003933373373001814, + 0.008426593616604805, + 0.007931625470519066, + 0.014115552417933941, + 0.0109021607786417, + 0.0075654457323253155, + 0.008940501138567924, + 0.004155649337917566, + 0.002493093488737941, + 0.0031976366881281137, + 0.0029132100753486156, + 0.0031245453283190727, + 0.005133099388331175, + 0.0035382327623665333, + 0.004572887439280748, + 0.0037838544230908155, + 0.003050616243854165, + 0.004502512980252504, + 0.00511866295710206, + 0.004249829798936844, + 0.0062642754055559635, + 0.005974682047963142, + 0.00509753730148077, + 0.005740679334849119, + 0.004760974086821079, + 0.010163861326873302, + 0.0070958975702524185, + 0.013273226097226143, + 0.0073790582828223705, + 0.008225399069488049, + 0.006905570160597563, + 0.005641615018248558, + 0.005687142722308636, + 0.003799937665462494, + 0.003082946641370654, + 0.004426662810146809, + 0.005248907022178173, + 0.005289563909173012, + 0.0023730683606117964, + 0.005749722011387348, + 0.024531427770853043, + 0.019275318831205368, + 0.017297711223363876, + 0.01618763990700245, + 0.02179531566798687, + 0.022387031465768814, + 0.019026998430490494, + 0.01980496384203434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.00595545070245862 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635931596159935 + ], + "opposite_logprob_tokens": [ + -3.6513593196868896 + ], + "logprob_reward": -0.02635931596159935, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635931596159935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1934773325920105, + 0.44512736797332764, + 0.07214426249265671, + 0.05716607719659805, + 0.021013768389821053, + 0.021026205271482468, + 0.042618103325366974, + 0.040359653532505035, + 0.04081547260284424, + 0.06625176966190338 + ], + "credit_map": [ + 0.024909354746341705, + 0.016588915139436722, + 0.03263983130455017, + 0.033570725470781326, + 0.01131718885153532, + 0.027578191831707954, + 0.030171236023306847, + 0.016701895743608475, + 0.020850682631134987, + 0.013654938898980618, + 0.13257057964801788, + 0.04281139373779297, + 0.025116588920354843, + 0.0254875048995018, + 0.1641828417778015, + 0.02045283280313015, + 0.008329548873007298, + 0.005589779932051897, + 0.012307601049542427, + 0.011998021975159645, + 0.010831179097294807, + 0.006133407354354858, + 0.011685313656926155, + 0.005269409157335758, + 0.004262269474565983, + 0.004163624253123999, + 0.007023159880191088, + 0.005284794140607119, + 0.007448745425790548, + 0.006209974177181721, + 0.018148835748434067, + 0.004624669905751944, + 0.0016361420275643468, + 0.0022484404034912586, + 0.0024924203753471375, + 0.003877859329804778, + 0.0036355287302285433, + 0.002959112636744976, + 0.002755513647571206, + 0.0014087531017139554, + 0.002408435568213463, + 0.0027787901926785707, + 0.001747102360241115, + 0.002047886373475194, + 0.002826583804562688, + 0.003219911130145192, + 0.0026945634745061398, + 0.003302930621430278, + 0.003059499664232135, + 0.00418735109269619, + 0.004482334479689598, + 0.004800626076757908, + 0.004642251413315535, + 0.008881399407982826, + 0.007002943195402622, + 0.0055616977624595165, + 0.003419003216549754, + 0.0077513414435088634, + 0.0032717857975512743, + 0.006456318311393261, + 0.004052710719406605, + 0.0058018239215016365, + 0.003619604976847768, + 0.00598707003518939, + 0.007394082378596067, + 0.004096896853297949, + 0.003259925404563546, + 0.0037756147794425488, + 0.00854155607521534, + 0.006704377010464668, + 0.002432416658848524, + 0.004610601346939802, + 0.008236018940806389, + 0.005649269558489323, + 0.004612750839442015, + 0.0034498912282288074, + 0.005837019067257643, + 0.019558962434530258, + 0.005316856782883406, + 0.013590999878942966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.02635931596159935 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419165637344122 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.01419165637344122, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419165637344122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049487195909023285, + 0.251451313495636, + 0.03361470624804497, + 0.030114024877548218, + 0.012464980594813824, + 0.014378371648490429, + 0.06155679374933243, + 0.0719439759850502, + 0.07131684571504593, + 0.40367183089256287 + ], + "credit_map": [ + 0.005186712834984064, + 0.004895759280771017, + 0.0070465630851686, + 0.008937124162912369, + 0.003537324722856283, + 0.005573721136897802, + 0.008262335322797298, + 0.0060476502403616905, + 0.011619549244642258, + 0.004160072188824415, + 0.08869440108537674, + 0.018256057053804398, + 0.009232818149030209, + 0.006955774966627359, + 0.10321607440710068, + 0.009316541254520416, + 0.0032723608892410994, + 0.0034974634181708097, + 0.006149783730506897, + 0.004417779855430126, + 0.004147904459387064, + 0.0030897893011569977, + 0.0061339121311903, + 0.0029057092033326626, + 0.0025355301331728697, + 0.00184122403152287, + 0.003568176180124283, + 0.002620410406962037, + 0.003669487312436104, + 0.0032048597931861877, + 0.010672029107809067, + 0.002002307213842869, + 0.0006819950067438185, + 0.0010890142293646932, + 0.0012687697308138013, + 0.002248680219054222, + 0.002443088684231043, + 0.0015372094931080937, + 0.0022903732024133205, + 0.0009058497962541878, + 0.0013046300737187266, + 0.0020922853145748377, + 0.0015812435885891318, + 0.0016960023203864694, + 0.0018341371323913336, + 0.0016425708308815956, + 0.0019764392636716366, + 0.002251063007861376, + 0.004077459219843149, + 0.005302042234688997, + 0.007758148480206728, + 0.006379002705216408, + 0.008703435771167278, + 0.010829243808984756, + 0.008165569975972176, + 0.010341886430978775, + 0.006920911837369204, + 0.006718468386679888, + 0.007206271402537823, + 0.005987631157040596, + 0.01951070874929428, + 0.009186375886201859, + 0.006073889322578907, + 0.010339714586734772, + 0.012803655117750168, + 0.005869073327630758, + 0.004702343139797449, + 0.004228204023092985, + 0.019764870405197144, + 0.010058542713522911, + 0.00345173804089427, + 0.010438413359224796, + 0.07299207150936127, + 0.04694948345422745, + 0.036077775061130524, + 0.03484757989645004, + 0.0537145622074604, + 0.05679944157600403, + 0.04444433003664017, + 0.0578465573489666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.01419165637344122 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic medium shot, captured from a low, upward angle in Annie Leibovitz's signature style, imbues the image with intense dramatic tension. Amidst a pouring rain, a chiseled man calmly steps from the driver's seat of a classic black sedan. He wears a long, charcoal-gray trench coat, the collar neatly raised to protect against the chill. The damp fabric, illuminated by a cold, off-screen streetlight, reflects a deep, textured sheen, sculpting the lines of his shoulders. Dense raindrops are clearly visible in the light, and his dark hair, slightly damp, a few strands plastering to his forehead. His resolute gaze, a mixture of fatigue and determination, passes the lens, gazing into the unknown darkness ahead. One hand remains on the open car door for balance, the faint glow from the dashboard filtering through, casting a hint of warmth on his dark trousers. His other foot has just touched the ground, creating a subtle splash of water. The asphalt beneath our feet is completely soaked by rain, forming mirror-like puddles that reflect the blurred red and blue glow of the distant city's neon lights. The entire scene is dominated by cold tones, with a strong contrast between light and shadow, accurately capturing the characters' resolute gestures as they step into the rain, creating an atmosphere imbued with mystery, loneliness, and a sense of fate.", + "seed": 5759, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_13/index_2163_seed_5759_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03671145644038916, + "vlm_logprob": -0.03671145644038916, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.03671145644038916, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14129574596881866 + ], + "opposite_logprob_tokens": [ + -2.3912956714630127 + ], + "logprob_reward": -0.14129574596881866, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.14129574596881866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19800372421741486, + 0.15845546126365662, + 0.07681310921907425, + 0.15138813853263855, + 0.04204167425632477, + 0.04255970939993858, + 0.09623999893665314, + 0.08285164088010788, + 0.10021103918552399, + 0.05143553391098976 + ], + "credit_map": [ + 0.02612525038421154, + 0.01367204263806343, + 0.023170260712504387, + 0.03648616373538971, + 0.022655436769127846, + 0.011726132594048977, + 0.034039489924907684, + 0.030128948390483856, + 0.013461161404848099, + 0.016435815021395683, + 0.014576447196304798, + 0.01880647987127304, + 0.02832660637795925, + 0.013894304633140564, + 0.02900349721312523, + 0.023951152339577675, + 0.013045942410826683, + 0.009661181829869747, + 0.006844208110123873, + 0.010201364755630493, + 0.007155298255383968, + 0.004230353981256485, + 0.013101061806082726, + 0.012573699466884136, + 0.012636112980544567, + 0.008691420778632164, + 0.008480455726385117, + 0.01367372740060091, + 0.02598884515464306, + 0.040521591901779175, + 0.02830415405333042, + 0.013091842643916607, + 0.005516002420336008, + 0.004475426860153675, + 0.0029778573662042618, + 0.004095788579434156, + 0.007129695266485214, + 0.007030161563307047, + 0.005647624377161264, + 0.005169113632291555, + 0.004488605540245771, + 0.006326357368379831, + 0.0033163155894726515, + 0.005228618159890175, + 0.00493807066231966, + 0.005173287354409695, + 0.005731185432523489, + 0.007357270922511816, + 0.011975624598562717, + 0.010639562271535397, + 0.0054746102541685104, + 0.013032537885010242, + 0.020794665440917015, + 0.009124130010604858, + 0.014463509432971478, + 0.010735353454947472, + 0.009304136037826538, + 0.010815304704010487, + 0.005582151003181934, + 0.014816220849752426, + 0.014773045666515827, + 0.006934476085007191, + 0.011504566296935081, + 0.009121731854975224, + 0.012661576271057129, + 0.01370025984942913, + 0.009589379653334618, + 0.013396764174103737, + 0.01830603927373886, + 0.008362543769180775, + 0.017202282324433327, + 0.006992194335907698, + 0.004781556781381369, + 0.007253681775182486, + 0.0036213556304574013, + 0.011881857179105282, + 0.00740386638790369, + 0.002712331246584654, + 0.008903958834707737, + 0.0048769270069897175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.14129574596881866 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004663780797272921 + ], + "opposite_logprob_tokens": [ + -5.379663944244385 + ], + "logprob_reward": -0.004663780797272921, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004663780797272921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29685384035110474, + 0.1391609162092209, + 0.050894733518362045, + 0.12898199260234833, + 0.08813268691301346, + 0.05158195272088051, + 0.04681282863020897, + 0.060652464628219604, + 0.06291983276605606, + 0.07400877773761749 + ], + "credit_map": [ + 0.025097565725445747, + 0.013219562359154224, + 0.01608123444020748, + 0.025166960433125496, + 0.018939314410090446, + 0.007119448389858007, + 0.117994524538517, + 0.07323525100946426, + 0.013162625953555107, + 0.013260339386761189, + 0.011371268890798092, + 0.016940193250775337, + 0.02894829586148262, + 0.008773893117904663, + 0.028788724914193153, + 0.01791556365787983, + 0.0060053495690226555, + 0.004527353215962648, + 0.006069141905754805, + 0.006763949058949947, + 0.004878512583673, + 0.004194552078843117, + 0.009973491542041302, + 0.008482384495437145, + 0.01214983407407999, + 0.009826766327023506, + 0.01577782817184925, + 0.015164264477789402, + 0.011340359225869179, + 0.011974590830504894, + 0.031179694458842278, + 0.02156866528093815, + 0.012157052755355835, + 0.004209205042570829, + 0.004948032554239035, + 0.008896456100046635, + 0.006973555311560631, + 0.008381803520023823, + 0.025620589032769203, + 0.016945993527770042, + 0.0071959407068789005, + 0.006096764001995325, + 0.004933563061058521, + 0.007330421824008226, + 0.008229325525462627, + 0.0075079514645040035, + 0.005993889644742012, + 0.00429409509524703, + 0.0050966753624379635, + 0.004407631233334541, + 0.0032511181198060513, + 0.007126035168766975, + 0.009936266578733921, + 0.003284717444330454, + 0.007671997416764498, + 0.006038387306034565, + 0.005800650455057621, + 0.007384638302028179, + 0.004842037800699472, + 0.011884658597409725, + 0.01009781938046217, + 0.0060667069628834724, + 0.009591760113835335, + 0.004984190687537193, + 0.008712474256753922, + 0.006964364089071751, + 0.006177830044180155, + 0.006769049447029829, + 0.008884362876415253, + 0.008435259573161602, + 0.010740363039076328, + 0.006236129906028509, + 0.012358661741018295, + 0.016412904486060143, + 0.004530853591859341, + 0.008901444263756275, + 0.008675934746861458, + 0.006403495091944933, + 0.011837268248200417, + 0.004888211376965046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.004663780797272921 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632262073457241 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632262073457241, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632262073457241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20526650547981262, + 0.02991361916065216, + 0.03174034133553505, + 0.0766083225607872, + 0.15590769052505493, + 0.03505000099539757, + 0.14619939029216766, + 0.11614018678665161, + 0.03000197559595108, + 0.1731719672679901 + ], + "credit_map": [ + 0.02150282822549343, + 0.010303004644811153, + 0.01280928310006857, + 0.02070111222565174, + 0.01668526791036129, + 0.007425636984407902, + 0.07231437414884567, + 0.04352499172091484, + 0.00316455215215683, + 0.0023221320006996393, + 0.0026142334099859, + 0.003750982228666544, + 0.003553735325112939, + 0.0019278781255707145, + 0.005688970442861319, + 0.006891133729368448, + 0.005162804387509823, + 0.004807920660823584, + 0.0030123707838356495, + 0.0035934781190007925, + 0.0032540049869567156, + 0.0019800160080194473, + 0.004077443387359381, + 0.005852303467690945, + 0.007838185876607895, + 0.004887543618679047, + 0.005286885425448418, + 0.01075312402099371, + 0.00680279778316617, + 0.00919482670724392, + 0.018137047067284584, + 0.013707908801734447, + 0.018905935809016228, + 0.006510165985673666, + 0.0068575358018279076, + 0.01563085988163948, + 0.013079971075057983, + 0.01952637918293476, + 0.0489778034389019, + 0.026419030502438545, + 0.005755975376814604, + 0.0036346311680972576, + 0.002894152421504259, + 0.004967188462615013, + 0.004461147356778383, + 0.004164672922343016, + 0.004189737141132355, + 0.004982495214790106, + 0.015415522269904613, + 0.015675343573093414, + 0.013668278232216835, + 0.03675617277622223, + 0.01869029924273491, + 0.00960699375718832, + 0.023474451154470444, + 0.012912320904433727, + 0.014708660542964935, + 0.010894899256527424, + 0.008359890431165695, + 0.024898286908864975, + 0.018211064860224724, + 0.010478781536221504, + 0.016320914030075073, + 0.012267690151929855, + 0.004882637411355972, + 0.004796938039362431, + 0.0036990027874708176, + 0.004049073439091444, + 0.0029595810920000076, + 0.0023451989982277155, + 0.004529545549303293, + 0.002739999210461974, + 0.025372935459017754, + 0.036043934524059296, + 0.010903635993599892, + 0.023431900888681412, + 0.0276784785091877, + 0.012525110505521297, + 0.023990310728549957, + 0.013225659728050232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.007632262073457241 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635200135409832 + ], + "opposite_logprob_tokens": [ + -3.6513519287109375 + ], + "logprob_reward": -0.02635200135409832, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635200135409832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20049183070659637, + 0.15159566700458527, + 0.06703212857246399, + 0.08996669948101044, + 0.10362327843904495, + 0.04811486229300499, + 0.06766685843467712, + 0.07643324136734009, + 0.08838720619678497, + 0.1066882386803627 + ], + "credit_map": [ + 0.024307575076818466, + 0.00978083536028862, + 0.014959671534597874, + 0.023713530972599983, + 0.012511465698480606, + 0.005721037741750479, + 0.06846242398023605, + 0.04103527590632439, + 0.013277874328196049, + 0.017104417085647583, + 0.013124988414347172, + 0.01944969780743122, + 0.032893769443035126, + 0.008608249947428703, + 0.027450647205114365, + 0.019686013460159302, + 0.01083350833505392, + 0.009209915064275265, + 0.0065017156302928925, + 0.009118974208831787, + 0.0073699867352843285, + 0.0028058206662535667, + 0.008324312046170235, + 0.012867896817624569, + 0.009786865673959255, + 0.00546842347830534, + 0.009606507606804371, + 0.011814349330961704, + 0.007135225925594568, + 0.01300770789384842, + 0.0238362904638052, + 0.009311328642070293, + 0.01238563098013401, + 0.005652768071740866, + 0.003940148279070854, + 0.01168788131326437, + 0.0074623082764446735, + 0.008751899003982544, + 0.03827469050884247, + 0.015467952005565166, + 0.006016515661031008, + 0.004412250593304634, + 0.0031206514686346054, + 0.0071707372553646564, + 0.00960061140358448, + 0.006150157656520605, + 0.005661213770508766, + 0.005982719827443361, + 0.008496429771184921, + 0.007453111000359058, + 0.004852583631873131, + 0.014787351712584496, + 0.011774517595767975, + 0.00426100380718708, + 0.009091340005397797, + 0.0069505153223872185, + 0.01068868301808834, + 0.0187781173735857, + 0.004102374892681837, + 0.010016081854701042, + 0.009997986257076263, + 0.006633588578552008, + 0.010968770831823349, + 0.0052476380951702595, + 0.014028675854206085, + 0.014650804921984673, + 0.007787537761032581, + 0.008255965076386929, + 0.01211678795516491, + 0.00983603484928608, + 0.014793705195188522, + 0.006917694117873907, + 0.016186153516173363, + 0.023918403312563896, + 0.006796350236982107, + 0.011384534649550915, + 0.011908206157386303, + 0.010626846924424171, + 0.01799054443836212, + 0.007877196185290813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.02635200135409832 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man stepping out of a classic black sedan into pouring rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a long charcoal-gray trench coat with the collar raised?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops visible in the light, creating a splash where the man's foot touches the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature asphalt with mirror-like puddles reflecting blurred red and blue neon lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036134920082986355 + ], + "opposite_logprob_tokens": [ + -5.628613471984863 + ], + "logprob_reward": -0.0036134920082986355, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036134920082986355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10312166064977646, + 0.07467029988765717, + 0.06368325650691986, + 0.04465194046497345, + 0.13725627958774567, + 0.06114018335938454, + 0.1424480378627777, + 0.1619793325662613, + 0.09223470836877823, + 0.11881431937217712 + ], + "credit_map": [ + 0.009491752833127975, + 0.009711180813610554, + 0.009428530000150204, + 0.012080245651304722, + 0.01237041037529707, + 0.004626931622624397, + 0.028090855106711388, + 0.017321757972240448, + 0.0063659511506557465, + 0.008127485401928425, + 0.006762758828699589, + 0.010390355251729488, + 0.0157536119222641, + 0.003965558484196663, + 0.012144234031438828, + 0.011160342954099178, + 0.009548991918563843, + 0.009207596071064472, + 0.008015057072043419, + 0.007492133881896734, + 0.007298782933503389, + 0.004468186292797327, + 0.008618742227554321, + 0.009033764712512493, + 0.007207837421447039, + 0.00348542514257133, + 0.004418866243213415, + 0.004769198130816221, + 0.0040830508805811405, + 0.0055597745813429356, + 0.010017513297498226, + 0.005110274534672499, + 0.01806892640888691, + 0.0139644555747509, + 0.007399419788271189, + 0.017298998311161995, + 0.017055684700608253, + 0.018069537356495857, + 0.02897627279162407, + 0.01642298884689808, + 0.008055062964558601, + 0.00686619384214282, + 0.005573100410401821, + 0.011358067393302917, + 0.005267034284770489, + 0.008225305005908012, + 0.008000549860298634, + 0.007794865872710943, + 0.01697603613138199, + 0.01804862730205059, + 0.01041189581155777, + 0.02842230536043644, + 0.020138457417488098, + 0.012643551453948021, + 0.021945839747786522, + 0.013861315324902534, + 0.02671830542385578, + 0.016876792535185814, + 0.010966965928673744, + 0.028581934049725533, + 0.025310959666967392, + 0.018308527767658234, + 0.021261930465698242, + 0.01395391020923853, + 0.0124415447935462, + 0.016787344589829445, + 0.008584531024098396, + 0.008234385401010513, + 0.010420364327728748, + 0.010162340477108955, + 0.01427567657083273, + 0.011328515596687794, + 0.02036847174167633, + 0.02152353711426258, + 0.008761691860854626, + 0.014980992302298546, + 0.016753006726503372, + 0.009235760197043419, + 0.017842929810285568, + 0.009347927756607533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.0036134920082986355 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_14/index_tensor([1013, 1013])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_14/index_tensor([1013, 1013])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7cd600c49a864cf2d68f294e24f6fc831ed87048 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_14/index_tensor([1013, 1013])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static wide shot frames a narrow, rain-slicked alley, creating an atmospheric, gritty, retro-cyberpunk scene. The entire scene is illuminated almost entirely by the vibrant, electric glow of unseen storefront neon signs in burgundy and cyan, their light casting long, distorted reflections across the dark, muddy asphalt. At the center of the frame sits a heavily modified hatchback, its compact form featuring a widebody kit and a matte gunmetal gray paint job that absorbs more light than it reflects. Thickly insulated cables and exposed wires snaking chaotically from where the front grille should be, hint at a powerful, unconventional engine. The car is fitted with rugged, oversized rally tires, their deep treads coated in dirt to grip the wet surface. In a sudden, violent move, the car surges forward with electric fury; its tires spin for a fraction of a second, violently churning up the accumulated water and sending forth a spectacular spray of luminous light that captures the burgundy and cyan glow. The vehicle pitches forward at explosive speed, motion blurring as it hurtles down the alley, its taillights leaving a faint red trail in the thick, humid air, perfectly capturing the high-octane dystopian moment.", + "seed": 27120, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_14/index_1013_seed_27120_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07316796416416764, + "vlm_logprob": -0.07316796416416764, + "vlm_yes_no_margin": 3.6499999523162843, + "vlm_token_credit": -0.07316796416416764, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.29659122228622437 + ], + "opposite_logprob_tokens": [ + -1.5465912818908691 + ], + "logprob_reward": -0.29659122228622437, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.29659122228622437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04692208021879196, + 0.052266694605350494, + 0.21707464754581451, + 0.03901378810405731, + 0.19247807562351227, + 0.05450363829731941, + 0.09136723726987839, + 0.22255340218544006, + 0.03305783122777939, + 0.050762586295604706 + ], + "credit_map": [ + 0.0049229576252400875, + 0.0038400867488235235, + 0.009198581799864769, + 0.005302170757204294, + 0.0038633327931165695, + 0.004354104865342379, + 0.009896233677864075, + 0.005544612649828196, + 0.007026736158877611, + 0.004243234638124704, + 0.006465232465416193, + 0.004055843222886324, + 0.004387499298900366, + 0.004219817463308573, + 0.016902126371860504, + 0.004966206382960081, + 0.013324354775249958, + 0.006477560382336378, + 0.0635545551776886, + 0.009888175874948502, + 0.013652308844029903, + 0.01660715974867344, + 0.08233639597892761, + 0.011234135366976261, + 0.0056154304184019566, + 0.004006766714155674, + 0.0056458283215761185, + 0.00845256820321083, + 0.003331024432554841, + 0.0031351775396615267, + 0.0038978338707238436, + 0.004929157439619303, + 0.014138361439108849, + 0.013904270716011524, + 0.05619024112820625, + 0.017497314140200615, + 0.014661785215139389, + 0.010225771926343441, + 0.05480315908789635, + 0.011057174764573574, + 0.004738750401884317, + 0.008225553669035435, + 0.007491970434784889, + 0.006466431077569723, + 0.003914313856512308, + 0.004654170013964176, + 0.01000894419848919, + 0.00900350697338581, + 0.011526975780725479, + 0.01040586270391941, + 0.010071467608213425, + 0.010240965522825718, + 0.006268444005399942, + 0.008577738888561726, + 0.014607291668653488, + 0.019668491557240486, + 0.01702493615448475, + 0.016880299896001816, + 0.016053782775998116, + 0.032313890755176544, + 0.0444699227809906, + 0.018251480534672737, + 0.0435844324529171, + 0.03397465497255325, + 0.002551028970628977, + 0.0033972300589084625, + 0.001959016313776374, + 0.003598719136789441, + 0.004121060948818922, + 0.0044597117230296135, + 0.005417371168732643, + 0.007553695235401392, + 0.004641598556190729, + 0.00538368197157979, + 0.004611247684806585, + 0.005862284917384386, + 0.003939373884350061, + 0.010249929502606392, + 0.007724933326244354, + 0.008349533192813396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.29659122228622437 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00413171062245965 + ], + "opposite_logprob_tokens": [ + -5.50413179397583 + ], + "logprob_reward": -0.00413171062245965, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00413171062245965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09438805282115936, + 0.1118892952799797, + 0.562100887298584, + 0.031469374895095825, + 0.03813604637980461, + 0.04590383544564247, + 0.020401643589138985, + 0.023601533845067024, + 0.030280975624918938, + 0.041828349232673645 + ], + "credit_map": [ + 0.01064029335975647, + 0.003566086757928133, + 0.016827965155243874, + 0.006200344767421484, + 0.00497754430398345, + 0.005502982996404171, + 0.028551729395985603, + 0.018121102824807167, + 0.014125148765742779, + 0.006950585637241602, + 0.011290027759969234, + 0.007996640168130398, + 0.00828398484736681, + 0.008345035836100578, + 0.038828738033771515, + 0.016069134697318077, + 0.035813964903354645, + 0.01760481484234333, + 0.16827915608882904, + 0.026910489425063133, + 0.03406484052538872, + 0.042898379266262054, + 0.20476745069026947, + 0.03176180273294449, + 0.007092362269759178, + 0.0036076768301427364, + 0.004963149782270193, + 0.002499388065189123, + 0.0029644949827343225, + 0.003692873287945986, + 0.0027291125152260065, + 0.003920315764844418, + 0.002926451852545142, + 0.003695237683132291, + 0.010586871765553951, + 0.003442531917244196, + 0.003123187692835927, + 0.003078628098592162, + 0.007965810596942902, + 0.003317330265417695, + 0.002353460295125842, + 0.003177965059876442, + 0.002845590468496084, + 0.004353923257440329, + 0.0030084976460784674, + 0.0028308522887527943, + 0.013125317171216011, + 0.014208232052624226, + 0.0023460332304239273, + 0.0024467932526022196, + 0.0021882911678403616, + 0.002896650694310665, + 0.0017671359237283468, + 0.0014721241313964128, + 0.0034509364049881697, + 0.0038336780853569508, + 0.002720058197155595, + 0.0019232339691370726, + 0.0025596865452826023, + 0.004531197249889374, + 0.00254798517562449, + 0.0019969625864177942, + 0.004021814558655024, + 0.003300595562905073, + 0.0022341273725032806, + 0.002220009919255972, + 0.0014324821531772614, + 0.0021002667490392923, + 0.0026904193218797445, + 0.00440201023593545, + 0.005383714102208614, + 0.009817947633564472, + 0.0047196620143949986, + 0.004867429379373789, + 0.0034248502925038338, + 0.0033850858453661203, + 0.004100240767002106, + 0.008722658269107342, + 0.005826845765113831, + 0.006781576666980982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.00413171062245965 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181919876486063 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.0181919876486063, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.0181919876486063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.252853661775589, + 0.03419741988182068, + 0.061381351202726364, + 0.07327891141176224, + 0.06463731825351715, + 0.25908830761909485, + 0.1078297570347786, + 0.07465291023254395, + 0.028875600546598434, + 0.043204739689826965 + ], + "credit_map": [ + 0.0235579963773489, + 0.010944835841655731, + 0.06815329939126968, + 0.014450552873313427, + 0.014039641246199608, + 0.00971401110291481, + 0.08970722556114197, + 0.022286107763648033, + 0.0037148059345781803, + 0.0027408734895288944, + 0.0042975167743861675, + 0.002195141278207302, + 0.002205796539783478, + 0.0026268719229847193, + 0.010806451551616192, + 0.0056099658831954, + 0.00570175563916564, + 0.0029597252141684294, + 0.012189007364213467, + 0.003485668683424592, + 0.004727360792458057, + 0.005660757422447205, + 0.019779793918132782, + 0.00687728775665164, + 0.006858768407255411, + 0.010118345730006695, + 0.011190393008291721, + 0.015130010433495045, + 0.007260170765221119, + 0.006087582092732191, + 0.009710397571325302, + 0.006923248991370201, + 0.0048155407421290874, + 0.006843190640211105, + 0.011755595915019512, + 0.0069617885164916515, + 0.006858084350824356, + 0.004501528572291136, + 0.017476830631494522, + 0.005424766801297665, + 0.009373513050377369, + 0.015554208308458328, + 0.011386976577341557, + 0.024221042171120644, + 0.012864460237324238, + 0.01190192624926567, + 0.07918102294206619, + 0.09460518509149551, + 0.014147193171083927, + 0.013210857287049294, + 0.009977075271308422, + 0.014418724924325943, + 0.011988408863544464, + 0.008247451856732368, + 0.01775648258626461, + 0.018083568662405014, + 0.0051896655932068825, + 0.006547186058014631, + 0.005965227726846933, + 0.015305797569453716, + 0.010058957152068615, + 0.010124748572707176, + 0.010823706164956093, + 0.010637630708515644, + 0.0028627102728933096, + 0.003394758328795433, + 0.001860457705333829, + 0.0031787073239684105, + 0.0035192472860217094, + 0.003815357806161046, + 0.004279583692550659, + 0.005964777898043394, + 0.0034658152144402266, + 0.007192668505012989, + 0.0023481377866119146, + 0.003494115313515067, + 0.004768651910126209, + 0.00971536710858345, + 0.00536168459802866, + 0.0068582999520003796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.0181919876486063 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581839606165886 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581839606165886, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581839606165886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03819699585437775, + 0.02976830117404461, + 0.0781550407409668, + 0.04017922654747963, + 0.11793983727693558, + 0.48455074429512024, + 0.05551210790872574, + 0.033575449138879776, + 0.04653162881731987, + 0.07559064775705338 + ], + "credit_map": [ + 0.005707995500415564, + 0.004507147707045078, + 0.005873601417988539, + 0.0042184945195913315, + 0.002544639864936471, + 0.004026192240417004, + 0.005654963664710522, + 0.005663965828716755, + 0.00313846324570477, + 0.0018902260344475508, + 0.0038608016911894083, + 0.0019524323288351297, + 0.0017927020089700818, + 0.0020959284156560898, + 0.00999085046350956, + 0.005046898499131203, + 0.004958260804414749, + 0.0022836916614323854, + 0.025815505534410477, + 0.0037597804330289364, + 0.004272870719432831, + 0.005684182979166508, + 0.026933301240205765, + 0.004447448067367077, + 0.007621001917868853, + 0.0031392329838126898, + 0.005176584701985121, + 0.004623171873390675, + 0.00397227518260479, + 0.005228498484939337, + 0.0040520839393138885, + 0.006366375368088484, + 0.009042945690453053, + 0.008288905024528503, + 0.03651639446616173, + 0.010807578451931477, + 0.0057134246453642845, + 0.00649371650069952, + 0.0335673987865448, + 0.007509483024477959, + 0.02487519010901451, + 0.02546698786318302, + 0.02826216071844101, + 0.045848917216062546, + 0.024174269288778305, + 0.024599049240350723, + 0.14762823283672333, + 0.16369594633579254, + 0.005034636706113815, + 0.007138535380363464, + 0.009527412243187428, + 0.007324226200580597, + 0.005929620936512947, + 0.0041311392560601234, + 0.008755745366215706, + 0.0076708015985786915, + 0.0025697157252579927, + 0.002665218897163868, + 0.0025247212033718824, + 0.005669500678777695, + 0.005288933869451284, + 0.003558572381734848, + 0.006215835455805063, + 0.005082952324301004, + 0.0035509776789695024, + 0.004942173603922129, + 0.002558235777541995, + 0.005875271745026112, + 0.005522547289729118, + 0.006048187613487244, + 0.007034530863165855, + 0.0109997084364295, + 0.006340892985463142, + 0.007362979929894209, + 0.00745152635499835, + 0.006404107436537743, + 0.00702629005536437, + 0.01786045916378498, + 0.01139696966856718, + 0.011747429147362709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.020581839606165886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634306065738201 + ], + "opposite_logprob_tokens": [ + -3.6513431072235107 + ], + "logprob_reward": -0.02634306065738201, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634306065738201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1742679476737976, + 0.04562436044216156, + 0.014999939128756523, + 0.036739859730005264, + 0.09507088363170624, + 0.05541510134935379, + 0.047320157289505005, + 0.07826241105794907, + 0.25061044096946716, + 0.2016889452934265 + ], + "credit_map": [ + 0.016846608370542526, + 0.008176952600479126, + 0.048305150121450424, + 0.01053975522518158, + 0.009985920041799545, + 0.007773966994136572, + 0.05958030745387077, + 0.013059291988611221, + 0.005872667301446199, + 0.0032315324060618877, + 0.004642306361347437, + 0.003596285590901971, + 0.0032466768752783537, + 0.003177190897986293, + 0.015394214540719986, + 0.0064634885638952255, + 0.0011619210708886385, + 0.0010171126341447234, + 0.004155173432081938, + 0.0011710578110069036, + 0.0008477208320982754, + 0.001506016356870532, + 0.0033147952053695917, + 0.0018261417280882597, + 0.004161572549492121, + 0.0036547654308378696, + 0.005583253223448992, + 0.007139612454921007, + 0.003363023977726698, + 0.00350019708275795, + 0.004111901391297579, + 0.005225533153861761, + 0.00737980380654335, + 0.006565786898136139, + 0.0294914860278368, + 0.008544279262423515, + 0.004630950279533863, + 0.005304997321218252, + 0.027369389310479164, + 0.005784190725535154, + 0.006283514201641083, + 0.00819262582808733, + 0.0074105653911828995, + 0.007252678740769625, + 0.004793686326593161, + 0.005716078449040651, + 0.008068625815212727, + 0.007697323337197304, + 0.005064597353339195, + 0.005302743520587683, + 0.004225449170917273, + 0.007411091588437557, + 0.0044185402803123, + 0.0035216871183365583, + 0.00824669934809208, + 0.00912934634834528, + 0.0066919270902872086, + 0.0064042615704238415, + 0.006586908362805843, + 0.017755717039108276, + 0.009029797278344631, + 0.007319955620914698, + 0.013275899924337864, + 0.01119794137775898, + 0.017793072387576103, + 0.014602115377783775, + 0.010121546685695648, + 0.016881871968507767, + 0.014414389617741108, + 0.043749261647462845, + 0.04486386105418205, + 0.08818431943655014, + 0.028680456802248955, + 0.021100711077451706, + 0.021358085796236992, + 0.014748042449355125, + 0.018108557909727097, + 0.029582751914858818, + 0.027535205706954002, + 0.040575120598077774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.02634306065738201 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static wide shot frames a narrow, rain-slicked alley, creating an atmospheric, gritty, retro-cyberpunk scene. The entire scene is illuminated almost entirely by the vibrant, electric glow of unseen storefront neon signs in burgundy and cyan, their light casting long, distorted reflections across the dark, muddy asphalt. At the center of the frame sits a heavily modified hatchback, its compact form featuring a widebody kit and a matte gunmetal gray paint job that absorbs more light than it reflects. Thickly insulated cables and exposed wires snaking chaotically from where the front grille should be, hint at a powerful, unconventional engine. The car is fitted with rugged, oversized rally tires, their deep treads coated in dirt to grip the wet surface. In a sudden, violent move, the car surges forward with electric fury; its tires spin for a fraction of a second, violently churning up the accumulated water and sending forth a spectacular spray of luminous light that captures the burgundy and cyan glow. The vehicle pitches forward at explosive speed, motion blurring as it hurtles down the alley, its taillights leaving a faint red trail in the thick, humid air, perfectly capturing the high-octane dystopian moment.", + "seed": 27121, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_14/index_1013_seed_27121_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03778726551681757, + "vlm_logprob": -0.03778726551681757, + "vlm_yes_no_margin": 4.049999952316284, + "vlm_token_credit": -0.03778726551681757, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12590231001377106 + ], + "opposite_logprob_tokens": [ + -2.5009024143218994 + ], + "logprob_reward": -0.12590231001377106, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.12590231001377106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11190330982208252, + 0.03129151090979576, + 0.1162591353058815, + 0.21607555449008942, + 0.0843249261379242, + 0.08264824002981186, + 0.06696180999279022, + 0.04589197039604187, + 0.20232003927230835, + 0.0423235148191452 + ], + "credit_map": [ + 0.013052094727754593, + 0.014544624835252762, + 0.01943054236471653, + 0.013458977453410625, + 0.012351004406809807, + 0.01576850190758705, + 0.012688945978879929, + 0.010608606040477753, + 0.005547735374420881, + 0.0030422958079725504, + 0.004090268164873123, + 0.00398149574175477, + 0.0020668176002800465, + 0.004753739573061466, + 0.003088832600042224, + 0.004720325116068125, + 0.01015506125986576, + 0.01353338360786438, + 0.011761202476918697, + 0.01704225316643715, + 0.015718065202236176, + 0.011539221741259098, + 0.021508987993001938, + 0.015000961720943451, + 0.01652538776397705, + 0.02216230519115925, + 0.03616892918944359, + 0.02663198858499527, + 0.03002166375517845, + 0.013372550718486309, + 0.03867831081151962, + 0.03251442685723305, + 0.006735417526215315, + 0.016292056068778038, + 0.007321425713598728, + 0.01788156107068062, + 0.008464859798550606, + 0.0039076232351362705, + 0.014353607781231403, + 0.009368374943733215, + 0.005137690342962742, + 0.00915562454611063, + 0.00917108729481697, + 0.015877876430749893, + 0.010266766883432865, + 0.00629321439191699, + 0.01629802957177162, + 0.010447953827679157, + 0.004943189211189747, + 0.009547452442348003, + 0.0033787174616008997, + 0.010137412697076797, + 0.016463544219732285, + 0.004622109699994326, + 0.012917891144752502, + 0.0049514975398778915, + 0.005334812216460705, + 0.005751557648181915, + 0.005580737255513668, + 0.004865466617047787, + 0.005555001087486744, + 0.005375622306019068, + 0.006187303923070431, + 0.007241470273584127, + 0.016935214400291443, + 0.02820325642824173, + 0.028914760798215866, + 0.020068634301424026, + 0.026727763935923576, + 0.026519307866692543, + 0.02046635001897812, + 0.034484751522541046, + 0.004335047677159309, + 0.005162439774721861, + 0.005894250702112913, + 0.004675616975873709, + 0.005159721244126558, + 0.0043237232603132725, + 0.00651176180690527, + 0.006260956637561321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.12590231001377106 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002515130676329136 + ], + "opposite_logprob_tokens": [ + -6.0025153160095215 + ], + "logprob_reward": -0.002515130676329136, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002515130676329136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1688106507062912, + 0.08879811316728592, + 0.15605968236923218, + 0.08334372192621231, + 0.24367541074752808, + 0.07038675248622894, + 0.02267208881676197, + 0.060426972806453705, + 0.04616023600101471, + 0.059666313230991364 + ], + "credit_map": [ + 0.01816175878047943, + 0.02054770290851593, + 0.02222505956888199, + 0.0279596708714962, + 0.016412224620580673, + 0.02571347914636135, + 0.025006458163261414, + 0.012784302234649658, + 0.011776098050177097, + 0.007832071743905544, + 0.012597233057022095, + 0.011778787709772587, + 0.006187421269714832, + 0.014682642184197903, + 0.01311109121888876, + 0.010832777246832848, + 0.011479944922029972, + 0.017293144017457962, + 0.014076822437345982, + 0.02464667148888111, + 0.02011614479124546, + 0.01616312563419342, + 0.02881394885480404, + 0.023469887673854828, + 0.008041450753808022, + 0.008450391702353954, + 0.01745588332414627, + 0.011858693324029446, + 0.00888480618596077, + 0.005729542579501867, + 0.01249273493885994, + 0.010430230759084225, + 0.027711171656847, + 0.022160355001688004, + 0.036434758454561234, + 0.054976098239421844, + 0.01379733718931675, + 0.016314193606376648, + 0.04792635887861252, + 0.024355148896574974, + 0.006048040464520454, + 0.005861910060048103, + 0.01012509036809206, + 0.014156910590827465, + 0.006794673856347799, + 0.005391124170273542, + 0.012592321261763573, + 0.009416691958904266, + 0.0024979335721582174, + 0.002912035444751382, + 0.0018296641064807773, + 0.004748568404465914, + 0.001979394815862179, + 0.0017397765768691897, + 0.005604187957942486, + 0.0013605317799374461, + 0.009171489626169205, + 0.0076681943610310555, + 0.008424867875874043, + 0.00732771772891283, + 0.007726061623543501, + 0.004896045662462711, + 0.007861201651394367, + 0.007351406384259462, + 0.004596078768372536, + 0.004654960241168737, + 0.007541596423834562, + 0.007565783336758614, + 0.0048639229498803616, + 0.0031927081290632486, + 0.007290774490684271, + 0.006454413756728172, + 0.006748152431100607, + 0.0049753268249332905, + 0.014964389614760876, + 0.0069023012183606625, + 0.006031372118741274, + 0.005053771194070578, + 0.007319197524338961, + 0.007671804167330265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.002515130676329136 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542534619569778 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.012542534619569778, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542534619569778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08780042082071304, + 0.04087047278881073, + 0.1220247745513916, + 0.2512805461883545, + 0.058272719383239746, + 0.2254345864057541, + 0.08103365451097488, + 0.09037879854440689, + 0.024844950065016747, + 0.018059058114886284 + ], + "credit_map": [ + 0.011345354840159416, + 0.0101395845413208, + 0.00903985183686018, + 0.010173231363296509, + 0.012211903929710388, + 0.01333634089678526, + 0.012703499756753445, + 0.008850659243762493, + 0.005765868816524744, + 0.0039763068780303, + 0.0044080885127186775, + 0.005063789896667004, + 0.004441355355083942, + 0.005690690595656633, + 0.006065674591809511, + 0.005458704195916653, + 0.012595407664775848, + 0.014460100792348385, + 0.010491183958947659, + 0.02119644358754158, + 0.018876001238822937, + 0.011135032400488853, + 0.021638691425323486, + 0.011631927452981472, + 0.02230309322476387, + 0.021920986473560333, + 0.06179950013756752, + 0.03367001935839653, + 0.028055978938937187, + 0.014801793731749058, + 0.03749235346913338, + 0.031236810609698296, + 0.004317367449402809, + 0.008688035421073437, + 0.004177417606115341, + 0.010816087014973164, + 0.008861461654305458, + 0.004750249907374382, + 0.011653880588710308, + 0.0050082216039299965, + 0.012891016900539398, + 0.02116112969815731, + 0.027633605524897575, + 0.04750433191657066, + 0.02238151989877224, + 0.011148312129080296, + 0.05358145385980606, + 0.029133237898349762, + 0.006364747881889343, + 0.011872584000229836, + 0.005982268135994673, + 0.013228102587163448, + 0.0196230486035347, + 0.0057438150979578495, + 0.012327923439443111, + 0.0058911750093102455, + 0.013734174892306328, + 0.011400814168155193, + 0.01287921890616417, + 0.01234534103423357, + 0.010868486016988754, + 0.0079644825309515, + 0.010841959156095982, + 0.010344329290091991, + 0.0019521984504535794, + 0.004414852242916822, + 0.0028830282390117645, + 0.003126337891444564, + 0.0030706936959177256, + 0.003935917746275663, + 0.0032342940103262663, + 0.0022276288364082575, + 0.0022907129023224115, + 0.0017399570206180215, + 0.0032913971226662397, + 0.0022124280221760273, + 0.0017732995329424739, + 0.00143200543243438, + 0.0036996169947087765, + 0.0016196422511711717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.012542534619569778 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185783177614212 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.018185783177614212, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018185783177614212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14993155002593994, + 0.020241975784301758, + 0.21078640222549438, + 0.07023168355226517, + 0.36003798246383667, + 0.02926558256149292, + 0.03930733725428581, + 0.02121996320784092, + 0.07368598878383636, + 0.025291509926319122 + ], + "credit_map": [ + 0.014833180233836174, + 0.017612580209970474, + 0.02043122984468937, + 0.024835966527462006, + 0.016478976234793663, + 0.022037673741579056, + 0.021313592791557312, + 0.012388354167342186, + 0.0027843203861266375, + 0.0014010189333930612, + 0.0036635291762650013, + 0.002339028287678957, + 0.0016438151942566037, + 0.002959909848868847, + 0.0031807105988264084, + 0.0022696428932249546, + 0.02135821245610714, + 0.025063036009669304, + 0.020026110112667084, + 0.035231489688158035, + 0.02898070402443409, + 0.02204124443233013, + 0.030318545177578926, + 0.027767060324549675, + 0.006866983603686094, + 0.00668964022770524, + 0.00952102243900299, + 0.015838218852877617, + 0.0064606377854943275, + 0.004898137412965298, + 0.011089080013334751, + 0.008867965079843998, + 0.042455755174160004, + 0.03130267560482025, + 0.054129183292388916, + 0.07924174517393112, + 0.021498853340744972, + 0.025718068704009056, + 0.06796658784151077, + 0.03772509843111038, + 0.001445070025511086, + 0.0025379913859069347, + 0.002464994555339217, + 0.00848781131207943, + 0.003189865732565522, + 0.0013644788414239883, + 0.005497487261891365, + 0.004277883097529411, + 0.0036050351336598396, + 0.005767227616161108, + 0.002157913986593485, + 0.006832018028944731, + 0.009043782949447632, + 0.002469514962285757, + 0.006625864654779434, + 0.0028059817850589752, + 0.002135159447789192, + 0.0026825512759387493, + 0.0019729454070329666, + 0.00166412687394768, + 0.0028562007937580347, + 0.0023110767360776663, + 0.0034469321835786104, + 0.004150970373302698, + 0.007604216691106558, + 0.009097547270357609, + 0.00937564019113779, + 0.008749308064579964, + 0.010728860273957253, + 0.0091650215908885, + 0.010173527523875237, + 0.008791867643594742, + 0.002542028436437249, + 0.0022453416604548693, + 0.00489624822512269, + 0.0031336054671555758, + 0.0029885717667639256, + 0.0020381337963044643, + 0.0038686911575496197, + 0.0035788894165307283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.018185783177614212 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029790569096803665 + ], + "opposite_logprob_tokens": [ + -3.5297906398773193 + ], + "logprob_reward": -0.029790569096803665, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029790569096803665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055677056312561035, + 0.03403516113758087, + 0.09298072755336761, + 0.10845869034528732, + 0.33098092675209045, + 0.027911435812711716, + 0.11562074720859528, + 0.08215580880641937, + 0.09594539552927017, + 0.056234102696180344 + ], + "credit_map": [ + 0.007584163919091225, + 0.005827540531754494, + 0.007064132951200008, + 0.009426303207874298, + 0.005306686274707317, + 0.005988048389554024, + 0.008459442295134068, + 0.006020733155310154, + 0.0044319708831608295, + 0.003005379345268011, + 0.004730948247015476, + 0.005199762526899576, + 0.003445379203185439, + 0.004843583796173334, + 0.005267186556011438, + 0.0031109468545764685, + 0.009107789024710655, + 0.007911399006843567, + 0.012511155568063259, + 0.014690347015857697, + 0.008117418736219406, + 0.008314360864460468, + 0.01904919557273388, + 0.013279058039188385, + 0.0101207559928298, + 0.012823781929910183, + 0.018172401934862137, + 0.01486667338758707, + 0.010710134170949459, + 0.007580758538097143, + 0.018427539616823196, + 0.0157566349953413, + 0.039956867694854736, + 0.0311594121158123, + 0.04937295243144035, + 0.0672009140253067, + 0.021223941817879677, + 0.023352721706032753, + 0.06379085779190063, + 0.03492322191596031, + 0.0016620710957795382, + 0.002451277105137706, + 0.002789314603433013, + 0.005240383557975292, + 0.003737089456990361, + 0.002073044655844569, + 0.0060642012394964695, + 0.0038940489757806063, + 0.010571355000138283, + 0.015247545205056667, + 0.00913894735276699, + 0.018644722178578377, + 0.023990711197257042, + 0.008333430625498295, + 0.021009482443332672, + 0.008684540167450905, + 0.011876561678946018, + 0.010051390156149864, + 0.012343949638307095, + 0.011821545660495758, + 0.008718493394553661, + 0.007430145516991615, + 0.010176034644246101, + 0.009737679734826088, + 0.007899485528469086, + 0.01818278431892395, + 0.011822332628071308, + 0.013078130781650543, + 0.011961054056882858, + 0.009445720352232456, + 0.013718648813664913, + 0.009837225079536438, + 0.003975199069827795, + 0.005593037698417902, + 0.008040446788072586, + 0.009103946387767792, + 0.009761114604771137, + 0.004258695058524609, + 0.009974631480872631, + 0.005527027416974306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.029790569096803665 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static wide shot frames a narrow, rain-slicked alley, creating an atmospheric, gritty, retro-cyberpunk scene. The entire scene is illuminated almost entirely by the vibrant, electric glow of unseen storefront neon signs in burgundy and cyan, their light casting long, distorted reflections across the dark, muddy asphalt. At the center of the frame sits a heavily modified hatchback, its compact form featuring a widebody kit and a matte gunmetal gray paint job that absorbs more light than it reflects. Thickly insulated cables and exposed wires snaking chaotically from where the front grille should be, hint at a powerful, unconventional engine. The car is fitted with rugged, oversized rally tires, their deep treads coated in dirt to grip the wet surface. In a sudden, violent move, the car surges forward with electric fury; its tires spin for a fraction of a second, violently churning up the accumulated water and sending forth a spectacular spray of luminous light that captures the burgundy and cyan glow. The vehicle pitches forward at explosive speed, motion blurring as it hurtles down the alley, its taillights leaving a faint red trail in the thick, humid air, perfectly capturing the high-octane dystopian moment.", + "seed": 27122, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_14/index_1013_seed_27122_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07587692458182574, + "vlm_logprob": -0.07587692458182574, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.07587692458182574, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.27041423320770264 + ], + "opposite_logprob_tokens": [ + -1.6454142332077026 + ], + "logprob_reward": -0.27041423320770264, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.27041423320770264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10116246342658997, + 0.066830575466156, + 0.08588734269142151, + 0.041715897619724274, + 0.10257153958082199, + 0.17046362161636353, + 0.0843823254108429, + 0.22654592990875244, + 0.08419235795736313, + 0.03624795377254486 + ], + "credit_map": [ + 0.016623878851532936, + 0.008667427115142345, + 0.01560104638338089, + 0.013538871891796589, + 0.012743262574076653, + 0.010414458811283112, + 0.011836333200335503, + 0.011737183667719364, + 0.0045541319996118546, + 0.007926001213490963, + 0.007534008007496595, + 0.018152635544538498, + 0.005257227923721075, + 0.006492974702268839, + 0.008557936176657677, + 0.008355655707418919, + 0.01254641730338335, + 0.007789205759763718, + 0.009665959514677525, + 0.013146202079951763, + 0.009892202913761139, + 0.011822051368653774, + 0.010075131431221962, + 0.010950172320008278, + 0.0045855967327952385, + 0.004379694350063801, + 0.006159659009426832, + 0.005985261872410774, + 0.0043624527752399445, + 0.0043854122050106525, + 0.0045981514267623425, + 0.007259666919708252, + 0.011153675615787506, + 0.008368764072656631, + 0.013396073132753372, + 0.01221601851284504, + 0.013159104622900486, + 0.01174804288893938, + 0.013873540796339512, + 0.01865631714463234, + 0.010022349655628204, + 0.017419790849089622, + 0.01851116307079792, + 0.02261763997375965, + 0.016686217859387398, + 0.02337312325835228, + 0.027175365015864372, + 0.03465797379612923, + 0.006668253801763058, + 0.008528130128979683, + 0.015373008325695992, + 0.017078854143619537, + 0.0064069186337292194, + 0.004837771411985159, + 0.012731327675282955, + 0.012758063152432442, + 0.015850262716412544, + 0.01154053583741188, + 0.07060014456510544, + 0.024433836340904236, + 0.011930759996175766, + 0.010914048179984093, + 0.05319942161440849, + 0.028076913207769394, + 0.007961460389196873, + 0.01104099489748478, + 0.010424739681184292, + 0.014615640975534916, + 0.011169574223458767, + 0.010902643203735352, + 0.00854643527418375, + 0.009530870243906975, + 0.0037377476692199707, + 0.00532687921077013, + 0.005884827580302954, + 0.005602821242064238, + 0.005313599947839975, + 0.0028843539766967297, + 0.0038959418889135122, + 0.0036017857491970062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.27041423320770264 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00866853166371584 + ], + "opposite_logprob_tokens": [ + -4.758668422698975 + ], + "logprob_reward": -0.00866853166371584, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00866853166371584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17933687567710876, + 0.07373709231615067, + 0.05243994668126106, + 0.09361003339290619, + 0.1848430037498474, + 0.06477384269237518, + 0.10758005082607269, + 0.10223318636417389, + 0.04974687099456787, + 0.09169908612966537 + ], + "credit_map": [ + 0.01949789933860302, + 0.02878698520362377, + 0.025190135464072227, + 0.019526850432157516, + 0.017163310199975967, + 0.02746570110321045, + 0.018082624301314354, + 0.023623371496796608, + 0.008080385625362396, + 0.006057788617908955, + 0.00721511198207736, + 0.011163451708853245, + 0.005034362897276878, + 0.012007332406938076, + 0.009565047919750214, + 0.01461360789835453, + 0.009256587363779545, + 0.004772929474711418, + 0.008104460313916206, + 0.005919433198869228, + 0.004382271319627762, + 0.004991157911717892, + 0.005883090663701296, + 0.009130015037953854, + 0.009431044571101665, + 0.007419207599014044, + 0.008390453644096851, + 0.026867680251598358, + 0.005333068780601025, + 0.004697637166827917, + 0.02422882243990898, + 0.007242117077112198, + 0.013585429638624191, + 0.017918454483151436, + 0.030114293098449707, + 0.024124279618263245, + 0.018481330946087837, + 0.017010707408189774, + 0.02697852998971939, + 0.03662997484207153, + 0.004845950286835432, + 0.0070631313137710094, + 0.006078851409256458, + 0.00861587468534708, + 0.011240227147936821, + 0.007326409220695496, + 0.011137193068861961, + 0.008466209284961224, + 0.007622142788022757, + 0.0056806025095283985, + 0.013944821432232857, + 0.018139271065592766, + 0.003939996939152479, + 0.0038620843552052975, + 0.022260131314396858, + 0.03213100880384445, + 0.011454293504357338, + 0.009721511043608189, + 0.009287606924772263, + 0.011599276214838028, + 0.006179408170282841, + 0.006688659079372883, + 0.019179124385118484, + 0.02812330797314644, + 0.006657635327428579, + 0.007543112151324749, + 0.012192773632705212, + 0.007312064524739981, + 0.003280092729255557, + 0.0034096429590135813, + 0.005968926940113306, + 0.003382622031494975, + 0.010474913753569126, + 0.012163480743765831, + 0.01191081665456295, + 0.012748056091368198, + 0.013732245191931725, + 0.007841161452233791, + 0.010497672483325005, + 0.012330739758908749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.00866853166371584 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198120683431625 + ], + "opposite_logprob_tokens": [ + -4.264198303222656 + ], + "logprob_reward": -0.014198120683431625, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198120683431625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13419802486896515, + 0.0930345356464386, + 0.04936164617538452, + 0.09699463099241257, + 0.08506636321544647, + 0.14618614315986633, + 0.0943547785282135, + 0.22431646287441254, + 0.04471411928534508, + 0.03177329897880554 + ], + "credit_map": [ + 0.021033376455307007, + 0.012102273292839527, + 0.015044149942696095, + 0.01996723748743534, + 0.014450146816670895, + 0.014692558906972408, + 0.020957546308636665, + 0.015950733795762062, + 0.008206898346543312, + 0.010832150466740131, + 0.0100247235968709, + 0.021716594696044922, + 0.007460250984877348, + 0.009559961035847664, + 0.01356523483991623, + 0.011668724939227104, + 0.005837768316268921, + 0.004435269627720118, + 0.004653445445001125, + 0.007080672774463892, + 0.004601323511451483, + 0.0058522578328847885, + 0.007256159093230963, + 0.009644744917750359, + 0.009654663503170013, + 0.007505904883146286, + 0.008689837530255318, + 0.024506572633981705, + 0.009460517205297947, + 0.006957062520086765, + 0.021326864138245583, + 0.008893204852938652, + 0.006604416761547327, + 0.01010988187044859, + 0.011022496037185192, + 0.01251531857997179, + 0.007819230668246746, + 0.007494908291846514, + 0.01235399954020977, + 0.017146116122603416, + 0.00870313961058855, + 0.019465111196041107, + 0.014955534599721432, + 0.03171151131391525, + 0.01233481615781784, + 0.014538112096488476, + 0.022415366023778915, + 0.022062545642256737, + 0.007372803054749966, + 0.010442642495036125, + 0.010992842726409435, + 0.018548032268881798, + 0.0073334756307303905, + 0.00880477111786604, + 0.015039199031889439, + 0.01582101173698902, + 0.01731988601386547, + 0.01556666474789381, + 0.014599754475057125, + 0.024707023054361343, + 0.030693819746375084, + 0.0283814687281847, + 0.0548284687101841, + 0.0382193922996521, + 0.00552082434296608, + 0.00560768973082304, + 0.005490550771355629, + 0.008728834800422192, + 0.004992929752916098, + 0.005022033117711544, + 0.004842348862439394, + 0.0045089093036949635, + 0.003060591174289584, + 0.004500469658523798, + 0.0044077117927372456, + 0.005021875724196434, + 0.004281526431441307, + 0.0025090405251830816, + 0.0039059328846633434, + 0.004086153116077185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.014198120683431625 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305415228009224 + ], + "opposite_logprob_tokens": [ + -3.1680541038513184 + ], + "logprob_reward": -0.04305415228009224, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305415228009224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0751958042383194, + 0.04113830626010895, + 0.07959196716547012, + 0.20588336884975433, + 0.1238565668463707, + 0.0808948278427124, + 0.12593293190002441, + 0.12239724397659302, + 0.08636739104986191, + 0.05874158442020416 + ], + "credit_map": [ + 0.009853771887719631, + 0.009717624634504318, + 0.011385458521544933, + 0.00913423951715231, + 0.00699659064412117, + 0.007985176518559456, + 0.009252057410776615, + 0.010870887897908688, + 0.0051603675819933414, + 0.0042276461608707905, + 0.004531217738986015, + 0.008697583340108395, + 0.0027191226836293936, + 0.004135794937610626, + 0.006340797059237957, + 0.005325776059180498, + 0.007711412385106087, + 0.007058818358927965, + 0.008600627072155476, + 0.011843843385577202, + 0.005886494182050228, + 0.008250427432358265, + 0.012427976354956627, + 0.017812371253967285, + 0.021530384197831154, + 0.015748871490359306, + 0.015326530672609806, + 0.052274275571107864, + 0.019777366891503334, + 0.01504433061927557, + 0.04745417460799217, + 0.01872742734849453, + 0.01143091730773449, + 0.013756579719483852, + 0.01988128013908863, + 0.0234182458370924, + 0.009542248211801052, + 0.008073060773313046, + 0.022622589021921158, + 0.015131643041968346, + 0.005672235507518053, + 0.009802564978599548, + 0.01072027999907732, + 0.014837386086583138, + 0.008897867985069752, + 0.007727607619017363, + 0.010716470889747143, + 0.012520416639745235, + 0.010392893105745316, + 0.007197893224656582, + 0.02121296152472496, + 0.02615082450211048, + 0.008256510831415653, + 0.012281057424843311, + 0.019025908783078194, + 0.021414875984191895, + 0.006906186696141958, + 0.00638769194483757, + 0.018096331506967545, + 0.013530551455914974, + 0.005528339650481939, + 0.004963720683008432, + 0.044306639581918716, + 0.02267778478562832, + 0.009377258829772472, + 0.011331490240991116, + 0.011538797989487648, + 0.014834546484053135, + 0.011302538216114044, + 0.01017285231500864, + 0.009541219100356102, + 0.008268686942756176, + 0.006268044468015432, + 0.010389537550508976, + 0.008270910941064358, + 0.009850851260125637, + 0.0076960413716733456, + 0.00317287421785295, + 0.0069695282727479935, + 0.006123797502368689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.04305415228009224 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043049585074186325 + ], + "opposite_logprob_tokens": [ + -3.1680495738983154 + ], + "logprob_reward": -0.043049585074186325, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043049585074186325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14219975471496582, + 0.19921760261058807, + 0.15371854603290558, + 0.10394888371229172, + 0.06712164729833603, + 0.054416269063949585, + 0.04722868278622627, + 0.0780884325504303, + 0.06486175209283829, + 0.08919842541217804 + ], + "credit_map": [ + 0.012138509191572666, + 0.013313292525708675, + 0.015072528272867203, + 0.020339541137218475, + 0.01849040761590004, + 0.015542523935437202, + 0.02687210962176323, + 0.020430848002433777, + 0.01772741973400116, + 0.018511736765503883, + 0.02394566498696804, + 0.03400581702589989, + 0.020064759999513626, + 0.020385228097438812, + 0.033634334802627563, + 0.03094264678657055, + 0.015966884791851044, + 0.011955403722822666, + 0.016093868762254715, + 0.026664871722459793, + 0.008728901855647564, + 0.010714635252952576, + 0.022770479321479797, + 0.04082350805401802, + 0.011279805563390255, + 0.007559656165540218, + 0.005706253927201033, + 0.030009035021066666, + 0.0069362567737698555, + 0.006596255116164684, + 0.027646329253911972, + 0.008215295150876045, + 0.004104590509086847, + 0.004702380858361721, + 0.007765259128063917, + 0.009594378061592579, + 0.005950392689555883, + 0.006370355375111103, + 0.014006390236318111, + 0.014627899043262005, + 0.0051438817754387856, + 0.00506580388173461, + 0.004739953670650721, + 0.006892439443618059, + 0.006728492211550474, + 0.006980078760534525, + 0.011162393726408482, + 0.007703224662691355, + 0.003506394801661372, + 0.004290123004466295, + 0.0050026290118694305, + 0.007159417495131493, + 0.003941033035516739, + 0.004978555720299482, + 0.008859654888510704, + 0.009490872733294964, + 0.008047432638704777, + 0.006944210268557072, + 0.017230452969670296, + 0.008282586932182312, + 0.007898896932601929, + 0.00872139073908329, + 0.011876247823238373, + 0.00908721424639225, + 0.007591279223561287, + 0.007822173647582531, + 0.010298407636582851, + 0.010088508948683739, + 0.008593657985329628, + 0.006711236666887999, + 0.007349047344177961, + 0.0064074392430484295, + 0.011013584211468697, + 0.01259433850646019, + 0.012307252734899521, + 0.012875123880803585, + 0.012292654253542423, + 0.0055817593820393085, + 0.010918491519987583, + 0.011615215800702572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.043049585074186325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static wide shot frames a narrow, rain-slicked alley, creating an atmospheric, gritty, retro-cyberpunk scene. The entire scene is illuminated almost entirely by the vibrant, electric glow of unseen storefront neon signs in burgundy and cyan, their light casting long, distorted reflections across the dark, muddy asphalt. At the center of the frame sits a heavily modified hatchback, its compact form featuring a widebody kit and a matte gunmetal gray paint job that absorbs more light than it reflects. Thickly insulated cables and exposed wires snaking chaotically from where the front grille should be, hint at a powerful, unconventional engine. The car is fitted with rugged, oversized rally tires, their deep treads coated in dirt to grip the wet surface. In a sudden, violent move, the car surges forward with electric fury; its tires spin for a fraction of a second, violently churning up the accumulated water and sending forth a spectacular spray of luminous light that captures the burgundy and cyan glow. The vehicle pitches forward at explosive speed, motion blurring as it hurtles down the alley, its taillights leaving a faint red trail in the thick, humid air, perfectly capturing the high-octane dystopian moment.", + "seed": 27123, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_14/index_1013_seed_27123_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13066107602789997, + "vlm_logprob": -0.13066107602789997, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.13066107602789997, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5617018938064575 + ], + "opposite_logprob_tokens": [ + -0.9367018938064575 + ], + "logprob_reward": -0.5617018938064575, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5617018938064575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6588951349258423, + 0.03302036598324776, + 0.05458502098917961, + 0.03422507271170616, + 0.026828257367014885, + 0.03508959710597992, + 0.026305334642529488, + 0.07046151906251907, + 0.03601706773042679, + 0.02457261271774769 + ], + "credit_map": [ + 0.014322744682431221, + 0.023050429299473763, + 0.013276939280331135, + 0.05185692012310028, + 0.02140342816710472, + 0.03742077201604843, + 0.2631419897079468, + 0.2344219982624054, + 0.004003581590950489, + 0.0026579215191304684, + 0.004456339869648218, + 0.007102374918758869, + 0.003338643116876483, + 0.0026428611017763615, + 0.004898097366094589, + 0.003920550923794508, + 0.004330881405621767, + 0.008909155614674091, + 0.005386557895690203, + 0.009391055442392826, + 0.00906720757484436, + 0.004003330133855343, + 0.007295440882444382, + 0.006201401352882385, + 0.0039544496685266495, + 0.0039011796470731497, + 0.005328908562660217, + 0.007203154731541872, + 0.0031430351082235575, + 0.003050524741411209, + 0.004302326589822769, + 0.0033414962235838175, + 0.002254936145618558, + 0.0036095024552196264, + 0.0026038705836981535, + 0.003914253786206245, + 0.00515799131244421, + 0.0033422147389501333, + 0.003568607848137617, + 0.0023768830578774214, + 0.002872062847018242, + 0.00433558551594615, + 0.0037698077503591776, + 0.004758829716593027, + 0.003349908161908388, + 0.007818473502993584, + 0.004450303036719561, + 0.0037346312310546637, + 0.004018973093479872, + 0.002596671925857663, + 0.003425884759053588, + 0.002867673756554723, + 0.0027515573892742395, + 0.0038461158983409405, + 0.0036430952604860067, + 0.003155365353450179, + 0.009026959538459778, + 0.010259278118610382, + 0.0069097899831831455, + 0.010534495115280151, + 0.006987110711634159, + 0.0034923760686069727, + 0.01362606044858694, + 0.009625454433262348, + 0.004074928816407919, + 0.004404339473694563, + 0.003488620975986123, + 0.004903735592961311, + 0.004029417876154184, + 0.004347702953964472, + 0.005489285569638014, + 0.005279040429741144, + 0.0027708872221410275, + 0.0026427495758980513, + 0.001940616755746305, + 0.0019897480960935354, + 0.0025012667756527662, + 0.004561895038932562, + 0.004520542919635773, + 0.003644909244030714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.5617018938064575 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046737478114664555 + ], + "opposite_logprob_tokens": [ + -5.379673957824707 + ], + "logprob_reward": -0.0046737478114664555, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046737478114664555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.617275595664978, + 0.05960075184702873, + 0.06726490706205368, + 0.022934045642614365, + 0.024022715166211128, + 0.013514423742890358, + 0.053531765937805176, + 0.045113783329725266, + 0.08325845748186111, + 0.01348357554525137 + ], + "credit_map": [ + 0.01370912417769432, + 0.019951138645410538, + 0.016039282083511353, + 0.048786815255880356, + 0.022670095786452293, + 0.03812665492296219, + 0.23852412402629852, + 0.21946832537651062, + 0.0037355010863393545, + 0.006436858791857958, + 0.007817496545612812, + 0.013906789012253284, + 0.004801855888217688, + 0.006310681346803904, + 0.00774590065702796, + 0.00884566642343998, + 0.007998945191502571, + 0.00751853222027421, + 0.012018710374832153, + 0.011864294297993183, + 0.005476418882608414, + 0.0055328491143882275, + 0.007815038785338402, + 0.009040115401148796, + 0.0018679405329748988, + 0.0031936790328472853, + 0.0027340068481862545, + 0.004388906992971897, + 0.002691651461645961, + 0.002682951744645834, + 0.0028396754059940577, + 0.0025352342054247856, + 0.0023112313356250525, + 0.003084518713876605, + 0.002037416212260723, + 0.0057646892964839935, + 0.0037624819669872522, + 0.0015958412550389767, + 0.003327606711536646, + 0.0021389287430793047, + 0.0012033688835799694, + 0.0021646772511303425, + 0.001907110563479364, + 0.001778140664100647, + 0.0016251979395747185, + 0.0018785683205351233, + 0.00160220789257437, + 0.0013551524607464671, + 0.012505467981100082, + 0.0038893723394721746, + 0.010464568622410297, + 0.003839127253741026, + 0.005306667182594538, + 0.00618351437151432, + 0.0075680045410990715, + 0.003775043645873666, + 0.005005420185625553, + 0.004325798247009516, + 0.003334350883960724, + 0.0035522952675819397, + 0.002577282255515456, + 0.003466436406597495, + 0.012211446650326252, + 0.010640750639140606, + 0.012977682054042816, + 0.0073326025158166885, + 0.0125502934679389, + 0.00952866580337286, + 0.006954969838261604, + 0.009501487948000431, + 0.0135847432538867, + 0.010828007943928242, + 0.00201640697196126, + 0.001802352606318891, + 0.0015635695308446884, + 0.001987000461667776, + 0.0012457809643819928, + 0.0014184375759214163, + 0.0016227476298809052, + 0.0018272798042744398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.0046737478114664555 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023293081670999527 + ], + "opposite_logprob_tokens": [ + -3.7732930183410645 + ], + "logprob_reward": -0.023293081670999527, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023293081670999527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3774206042289734, + 0.13940240442752838, + 0.03193425014615059, + 0.023975227028131485, + 0.03342372924089432, + 0.0232787374407053, + 0.05278599262237549, + 0.24300457537174225, + 0.03944145515561104, + 0.035332970321178436 + ], + "credit_map": [ + 0.007975657470524311, + 0.013384107500314713, + 0.007816744968295097, + 0.03011363372206688, + 0.012715250253677368, + 0.021984420716762543, + 0.1507004052400589, + 0.1327303946018219, + 0.011865799315273762, + 0.01509488932788372, + 0.016111329197883606, + 0.029613012447953224, + 0.013251066207885742, + 0.015905097126960754, + 0.020770471543073654, + 0.01679074950516224, + 0.003568180138245225, + 0.005332978442311287, + 0.004612323362380266, + 0.006054227706044912, + 0.0032982765696942806, + 0.0024703892413526773, + 0.0033689329866319895, + 0.0032289396040141582, + 0.002515140688046813, + 0.004514682572335005, + 0.003150585573166609, + 0.004731491208076477, + 0.0015119260642677546, + 0.0017320020124316216, + 0.0032229158096015453, + 0.0025964821688830853, + 0.003953561186790466, + 0.006074612960219383, + 0.004293735139071941, + 0.007789633236825466, + 0.0023788027465343475, + 0.0016761242877691984, + 0.004906798247247934, + 0.002350465627387166, + 0.002470087492838502, + 0.002952686743810773, + 0.0031337460968643427, + 0.0032710921950638294, + 0.002312659053131938, + 0.003923038952052593, + 0.002997864969074726, + 0.0022175628691911697, + 0.007552208844572306, + 0.005781198851764202, + 0.007781065069139004, + 0.004266608506441116, + 0.006116592325270176, + 0.009615997783839703, + 0.007112594321370125, + 0.004559728316962719, + 0.021340632811188698, + 0.01316304411739111, + 0.013856136240065098, + 0.020380880683660507, + 0.015500430017709732, + 0.022803742438554764, + 0.0735163614153862, + 0.06244335696101189, + 0.004820753820240498, + 0.0038460013456642628, + 0.004564159084111452, + 0.004119866993278265, + 0.003657556837424636, + 0.004842190537601709, + 0.007667238358408213, + 0.005923689343035221, + 0.00279712094925344, + 0.0035011060535907745, + 0.002288619987666607, + 0.0028061617631465197, + 0.003791885683313012, + 0.006451728288084269, + 0.008125764317810535, + 0.005570585839450359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.023293081670999527 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581021904945374 + ], + "opposite_logprob_tokens": [ + -3.895581007003784 + ], + "logprob_reward": -0.020581021904945374, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581021904945374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22064821422100067, + 0.080155149102211, + 0.15707725286483765, + 0.033290278166532516, + 0.19939327239990234, + 0.07405873388051987, + 0.08633609861135483, + 0.0391615554690361, + 0.0768243670463562, + 0.033055130392313004 + ], + "credit_map": [ + 0.0050969175063073635, + 0.013370536267757416, + 0.00803872849792242, + 0.01803635247051716, + 0.011649945750832558, + 0.014121522195637226, + 0.08110713958740234, + 0.06922705471515656, + 0.008428895846009254, + 0.008731086738407612, + 0.009199079126119614, + 0.01645887829363346, + 0.008262024261057377, + 0.008337154053151608, + 0.011885170824825764, + 0.008852855302393436, + 0.016932262107729912, + 0.015339111909270287, + 0.02485046721994877, + 0.02381710335612297, + 0.01490127295255661, + 0.013955241069197655, + 0.01811406761407852, + 0.02916770800948143, + 0.003203034633770585, + 0.0068448325619101524, + 0.004015855025500059, + 0.00687989080324769, + 0.0021040779538452625, + 0.003257524920627475, + 0.004327965900301933, + 0.0026570954360067844, + 0.019221393391489983, + 0.029253588989377022, + 0.01986945979297161, + 0.02572072483599186, + 0.04005168005824089, + 0.021427879109978676, + 0.024482443928718567, + 0.019366081804037094, + 0.0073906173929572105, + 0.0071009183302521706, + 0.014892946928739548, + 0.010252573527395725, + 0.0062011247500777245, + 0.011105419136583805, + 0.009865855798125267, + 0.007249266840517521, + 0.01731952093541622, + 0.007637538015842438, + 0.018300214782357216, + 0.01058892160654068, + 0.009283476509153843, + 0.009778549894690514, + 0.008106804452836514, + 0.005321070551872253, + 0.0037765090819448233, + 0.0025429537054151297, + 0.002515908796340227, + 0.005174555815756321, + 0.0025368505157530308, + 0.0025677334051579237, + 0.011523175984621048, + 0.008523866534233093, + 0.008678840473294258, + 0.009003783576190472, + 0.006783707998692989, + 0.0077726999297738075, + 0.009348610416054726, + 0.00960202794522047, + 0.016010751947760582, + 0.009623933583498001, + 0.0024267183616757393, + 0.0028081464115530252, + 0.0022581450175493956, + 0.0034131528809666634, + 0.003908493090420961, + 0.004512311425060034, + 0.008491117507219315, + 0.0052370429039001465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.020581021904945374 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a narrow, rain-slicked alley illuminated by burgundy and cyan neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a heavily modified hatchback with a matte gunmetal gray paint job and widebody kit visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car surge forward violently, spinning its tires and churning up water to create a spray of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the car's taillights leave a faint red trail in the humid air as it speeds down the alley?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305563494563103 + ], + "opposite_logprob_tokens": [ + -3.168055534362793 + ], + "logprob_reward": -0.04305563494563103, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305563494563103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46423062682151794, + 0.057322438806295395, + 0.07294201105833054, + 0.023004429414868355, + 0.08717695623636246, + 0.06956441700458527, + 0.07220716774463654, + 0.06352238357067108, + 0.048793207854032516, + 0.041236378252506256 + ], + "credit_map": [ + 0.00939152855426073, + 0.016812780871987343, + 0.009506885893642902, + 0.037181418389081955, + 0.014678196981549263, + 0.026984700933098793, + 0.18573376536369324, + 0.16394133865833282, + 0.006159917451441288, + 0.00637815659865737, + 0.007095223758369684, + 0.009760127402842045, + 0.007459575776010752, + 0.005298411939293146, + 0.008869527839124203, + 0.006301495246589184, + 0.00782044418156147, + 0.007742659654468298, + 0.013512486591935158, + 0.012433040887117386, + 0.008506286889314651, + 0.0044268397614359856, + 0.012221303768455982, + 0.00627894327044487, + 0.002777211833745241, + 0.003171258606016636, + 0.004153933841735125, + 0.003307642647996545, + 0.002526285592466593, + 0.0018983964109793305, + 0.0029712680261582136, + 0.002198432106524706, + 0.0077783046290278435, + 0.017844630405306816, + 0.009367946535348892, + 0.021379871293902397, + 0.006903852801769972, + 0.005245578475296497, + 0.012346001341938972, + 0.006310771219432354, + 0.006046686787158251, + 0.00531901977956295, + 0.011765922419726849, + 0.009482304565608501, + 0.0062169902957975864, + 0.01359148696064949, + 0.009768132120370865, + 0.00737387640401721, + 0.015391500666737556, + 0.0049877162091434, + 0.01815618760883808, + 0.013080351054668427, + 0.004113682080060244, + 0.004999097436666489, + 0.007336256094276905, + 0.004142371006309986, + 0.003623340046033263, + 0.004286114126443863, + 0.002694539027288556, + 0.005521944724023342, + 0.004992588888853788, + 0.0049241892993450165, + 0.021511418744921684, + 0.01596825011074543, + 0.007204053457826376, + 0.005140041001141071, + 0.006223381496965885, + 0.005665073171257973, + 0.004714437294751406, + 0.0052195661701262, + 0.008797915652394295, + 0.005828744266182184, + 0.004769156686961651, + 0.004146262537688017, + 0.002742953598499298, + 0.0036419315729290247, + 0.004366234876215458, + 0.006689917761832476, + 0.008406182751059532, + 0.006473737768828869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.04305563494563103 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_15/index_tensor([8510, 8510])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_15/index_tensor([8510, 8510])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d7f8a64251792a15c41ed90f344948940e5b83fb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_15/index_tensor([8510, 8510])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Rendered from a static, high-angle, top-down perspective, the scene is rendered in a hyper-detailed, gongbi-style painting, emphasizing every fine line and subtle texture with artistic precision. A pair of gleaming chrome hairdressing scissors and a magnificent pale yellow jade comb are meticulously arranged on a dark, high-gloss lacquered wooden tray. The tray itself rests on a bed of deep crimson silk, its soft, luxurious folds catching a soft, focused beam of overhead light, causing the polished chrome to gleam and the translucent jade to glow from within. As the first delicate, needle-like particles of ice begin to form on the cold metal of the scissors' blades, a slow, mesmerizing crystallization begins. Steadily and deliberately, intricate patterns of frost-like fragments bloom, spiderwebs spreading across the polished metal surface and the cool, richly textured stone of the comb. Growth is a silent, captivating dance, as the tools gradually become completely encased in a delicate, translucent crystal structure, transforming them into a single, frozen sculpture. Their original form remains visible beneath the shimmering, faceted ice, creating an atmosphere of tranquil and magical beauty.", + "seed": 13056, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_15/index_8510_seed_13056_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08074668250046671, + "vlm_logprob": -0.08074668250046671, + "vlm_yes_no_margin": 5.000000011920929, + "vlm_token_credit": -0.08074668250046671, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3893820643424988 + ], + "opposite_logprob_tokens": [ + -1.1393821239471436 + ], + "logprob_reward": -0.3893820643424988, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3893820643424988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11177254468202591, + 0.08191423118114471, + 0.04872194677591324, + 0.17492832243442535, + 0.0466906800866127, + 0.05492909625172615, + 0.08570456504821777, + 0.031893856823444366, + 0.028042960911989212, + 0.3354018032550812 + ], + "credit_map": [ + 0.012782800942659378, + 0.013428031466901302, + 0.016622401773929596, + 0.01574834994971752, + 0.010332942940294743, + 0.012683296576142311, + 0.016734939068555832, + 0.013439781963825226, + 0.011910920962691307, + 0.007852470502257347, + 0.010843549855053425, + 0.009623708203434944, + 0.008819155395030975, + 0.012504232116043568, + 0.010382852517068386, + 0.00997734535485506, + 0.006129333283752203, + 0.006440433673560619, + 0.005029518622905016, + 0.00887246709316969, + 0.004133058711886406, + 0.0064702765084803104, + 0.005467463284730911, + 0.006179393734782934, + 0.050102319568395615, + 0.012306127697229385, + 0.012182675302028656, + 0.008549188263714314, + 0.007491232827305794, + 0.013884758576750755, + 0.01711001992225647, + 0.05330199748277664, + 0.005462325643748045, + 0.004399549216032028, + 0.005826849024742842, + 0.006379426456987858, + 0.0048145088367164135, + 0.00990624912083149, + 0.005683908239006996, + 0.004217862151563168, + 0.007453319616615772, + 0.007005392573773861, + 0.00570839224383235, + 0.006895109079778194, + 0.00592486746609211, + 0.0071112606674432755, + 0.006227394565939903, + 0.008603357709944248, + 0.014827513135969639, + 0.007386660668998957, + 0.009053019806742668, + 0.007468279451131821, + 0.01197617407888174, + 0.010851321741938591, + 0.008346110582351685, + 0.015795482322573662, + 0.00622556684538722, + 0.0038266200572252274, + 0.0030101854354143143, + 0.0034192160237580538, + 0.004198792856186628, + 0.003499649465084076, + 0.002664038445800543, + 0.005049786064773798, + 0.002337288809940219, + 0.004407515749335289, + 0.0029183048754930496, + 0.004587664268910885, + 0.004830274265259504, + 0.002212810330092907, + 0.002118519274517894, + 0.004630584269762039, + 0.07164425402879715, + 0.022571897134184837, + 0.0231445524841547, + 0.12523475289344788, + 0.01775403879582882, + 0.01789128966629505, + 0.038183845579624176, + 0.018977176398038864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.3893820643424988 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013436584267765284 + ], + "opposite_logprob_tokens": [ + -6.626343727111816 + ], + "logprob_reward": -0.0013436584267765284, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013436584267765284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.402214914560318, + 0.050942305475473404, + 0.032503437250852585, + 0.07055714726448059, + 0.034567080438137054, + 0.04568759351968765, + 0.021973934024572372, + 0.06277622282505035, + 0.049576397985219955, + 0.22920091450214386 + ], + "credit_map": [ + 0.04748837649822235, + 0.04553764685988426, + 0.0439862497150898, + 0.04963603615760803, + 0.05477194860577583, + 0.046938639134168625, + 0.06311767548322678, + 0.0507383868098259, + 0.0075771985575556755, + 0.006404785439372063, + 0.005334360525012016, + 0.0055557964369654655, + 0.004843199159950018, + 0.008844426833093166, + 0.007353620138019323, + 0.005028923507779837, + 0.0044315108098089695, + 0.0020517276134341955, + 0.005220057908445597, + 0.003750414587557316, + 0.0024310133885592222, + 0.004589479882270098, + 0.005773888435214758, + 0.004255346488207579, + 0.011050505563616753, + 0.012989441864192486, + 0.007360949646681547, + 0.008753782138228416, + 0.008320172317326069, + 0.007454201113432646, + 0.006755330599844456, + 0.007872768677771091, + 0.0036404281854629517, + 0.0051881177350878716, + 0.0050744968466460705, + 0.0036120028235018253, + 0.004446056671440601, + 0.004477867390960455, + 0.0037345828022807837, + 0.00439353147521615, + 0.0047170319594442844, + 0.0066404021345078945, + 0.006368996575474739, + 0.005671569146215916, + 0.005280624143779278, + 0.005416963715106249, + 0.005141329951584339, + 0.00645067635923624, + 0.002631359500810504, + 0.002564236056059599, + 0.0022341336589306593, + 0.0027702650986611843, + 0.0027378485538065434, + 0.003220785642042756, + 0.0027411936316639185, + 0.0030741128139197826, + 0.01496290322393179, + 0.007942592725157738, + 0.005680777598172426, + 0.0055652279406785965, + 0.00475452421233058, + 0.005966287571936846, + 0.005170943681150675, + 0.012732968665659428, + 0.008294807747006416, + 0.006970918737351894, + 0.003989849239587784, + 0.005900617688894272, + 0.005618201103061438, + 0.0049224067479372025, + 0.0038960075471550226, + 0.00998359452933073, + 0.048097480088472366, + 0.018784288316965103, + 0.01875576749444008, + 0.07501791417598724, + 0.012041503563523293, + 0.01353692077100277, + 0.025290604680776596, + 0.01767643354833126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.0013436584267765284 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010387268848717213 + ], + "opposite_logprob_tokens": [ + -6.876038551330566 + ], + "logprob_reward": -0.0010387268848717213, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010387268848717213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3038424849510193, + 0.10838019847869873, + 0.06535767763853073, + 0.05754454433917999, + 0.03217148035764694, + 0.05006379261612892, + 0.036336470395326614, + 0.08949580788612366, + 0.024614939466118813, + 0.23219260573387146 + ], + "credit_map": [ + 0.05024882033467293, + 0.041469212621450424, + 0.03334207087755203, + 0.03305358439683914, + 0.03326570242643356, + 0.04589221999049187, + 0.03255704790353775, + 0.03401380777359009, + 0.01686328649520874, + 0.010302365757524967, + 0.010531717911362648, + 0.012170907109975815, + 0.010828508995473385, + 0.021252553910017014, + 0.013649995438754559, + 0.012780864723026752, + 0.010695184580981731, + 0.006726061925292015, + 0.007638194132596254, + 0.007233607582747936, + 0.005141664296388626, + 0.013700997456908226, + 0.006374815944582224, + 0.00784714613109827, + 0.009994901716709137, + 0.007051221095025539, + 0.00481806555762887, + 0.004860086366534233, + 0.008069262839853764, + 0.00466592563316226, + 0.006463592406362295, + 0.011621491052210331, + 0.0034974233713001013, + 0.00260551436804235, + 0.004716206807643175, + 0.003723863512277603, + 0.0031569721177220345, + 0.0073891314677894115, + 0.0035579779651015997, + 0.003524388186633587, + 0.008817778900265694, + 0.004254459869116545, + 0.006073899567127228, + 0.004329761955887079, + 0.004523806739598513, + 0.009224966168403625, + 0.004459964111447334, + 0.008379156701266766, + 0.0051481821574270725, + 0.004781180527061224, + 0.003918347880244255, + 0.0036288376431912184, + 0.0034937693271785975, + 0.004393015056848526, + 0.004473290406167507, + 0.006499848794192076, + 0.01929943449795246, + 0.008816287852823734, + 0.009045591577887535, + 0.006306186318397522, + 0.011348721571266651, + 0.008573791943490505, + 0.006817047018557787, + 0.01928875222802162, + 0.004325625021010637, + 0.0032327284570783377, + 0.0029441083315759897, + 0.0030979712028056383, + 0.002387615852057934, + 0.0039002830162644386, + 0.0019933683797717094, + 0.0027332401368767023, + 0.05012468248605728, + 0.02234521694481373, + 0.018617477267980576, + 0.06821875274181366, + 0.01435014232993126, + 0.01881398819386959, + 0.02239488624036312, + 0.017327452078461647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0010387268848717213 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002197353169322014 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.002197353169322014, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002197353169322014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06481027603149414, + 0.03979778289794922, + 0.024941598996520042, + 0.38636070489883423, + 0.025754185393452644, + 0.04195059463381767, + 0.050795577466487885, + 0.03530741482973099, + 0.017438096925616264, + 0.31284379959106445 + ], + "credit_map": [ + 0.008706117980182171, + 0.007100122049450874, + 0.007579015102237463, + 0.008258032612502575, + 0.008162311278283596, + 0.007015411276370287, + 0.010341677814722061, + 0.007647587917745113, + 0.006021022330969572, + 0.004143958445638418, + 0.005017850082367659, + 0.005143656861037016, + 0.0037720748223364353, + 0.005522072780877352, + 0.005373063497245312, + 0.004804084077477455, + 0.003770727664232254, + 0.00214138301089406, + 0.0032203898299485445, + 0.003177126869559288, + 0.0027570948004722595, + 0.0032342253252863884, + 0.003020338946953416, + 0.003620312549173832, + 0.12474323809146881, + 0.020670467987656593, + 0.027492716908454895, + 0.011659316718578339, + 0.011915095150470734, + 0.031557220965623856, + 0.03443014994263649, + 0.12389250099658966, + 0.00256274058483541, + 0.00309551227837801, + 0.0025075506418943405, + 0.0028233451303094625, + 0.005904046818614006, + 0.0033303266391158104, + 0.003200714709237218, + 0.0023299476597458124, + 0.005420863628387451, + 0.005502659827470779, + 0.004010188393294811, + 0.004070671740919352, + 0.007692607585340738, + 0.004061006009578705, + 0.005134598817676306, + 0.006058000959455967, + 0.006528542377054691, + 0.006185945123434067, + 0.005160681903362274, + 0.004734379705041647, + 0.006083877291530371, + 0.007185100112110376, + 0.005591406021267176, + 0.009325647726655006, + 0.004253686871379614, + 0.004997815005481243, + 0.004418555181473494, + 0.0033186788205057383, + 0.006005221977829933, + 0.005164330825209618, + 0.004052228294312954, + 0.0030969020444899797, + 0.004099178593605757, + 0.002100017387419939, + 0.0019176295027136803, + 0.0030855867080390453, + 0.002180033829063177, + 0.0015442215371876955, + 0.0012267393758520484, + 0.0012846901081502438, + 0.06126127764582634, + 0.02412320300936699, + 0.0214408989995718, + 0.11380176246166229, + 0.01728191040456295, + 0.016944844275712967, + 0.03807826340198517, + 0.0199116338044405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.002197353169322014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009771609678864479 + ], + "opposite_logprob_tokens": [ + -4.634771823883057 + ], + "logprob_reward": -0.009771609678864479, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009771609678864479, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15562088787555695, + 0.041201990097761154, + 0.03604568541049957, + 0.27308332920074463, + 0.04521084576845169, + 0.07550058513879776, + 0.04285021871328354, + 0.06503214687108994, + 0.016829008236527443, + 0.24862536787986755 + ], + "credit_map": [ + 0.02233235538005829, + 0.0156721081584692, + 0.01535661518573761, + 0.015686629340052605, + 0.018852978944778442, + 0.028251919895410538, + 0.022904423996806145, + 0.01656385324895382, + 0.004760160576552153, + 0.0037011606618762016, + 0.004689014051109552, + 0.004715723916888237, + 0.004132035654038191, + 0.008708573877811432, + 0.006342166103422642, + 0.004153154790401459, + 0.004292188212275505, + 0.004598721396178007, + 0.004293479491025209, + 0.004719804041087627, + 0.004065866116434336, + 0.005932917818427086, + 0.004381174687296152, + 0.0037615327164530754, + 0.08306844532489777, + 0.015720941126346588, + 0.019612617790699005, + 0.010487457737326622, + 0.00913669541478157, + 0.023553429171442986, + 0.02511685900390148, + 0.08638685941696167, + 0.004327619448304176, + 0.005502937361598015, + 0.007348423358052969, + 0.005187459290027618, + 0.005377527326345444, + 0.006338142789900303, + 0.004248914774507284, + 0.006879819557070732, + 0.011124104261398315, + 0.006712222006171942, + 0.007268225774168968, + 0.008680123835802078, + 0.008402817882597446, + 0.00811950396746397, + 0.007706811185926199, + 0.017486771568655968, + 0.004912763833999634, + 0.005272651091217995, + 0.004889754578471184, + 0.004378473851829767, + 0.007322391495108604, + 0.006662565283477306, + 0.004117797594517469, + 0.005293822847306728, + 0.01584722101688385, + 0.0052162338979542255, + 0.006689012981951237, + 0.004047431983053684, + 0.009115912020206451, + 0.005315999500453472, + 0.004319996107369661, + 0.014480335637927055, + 0.002006017602980137, + 0.0024482307489961386, + 0.0022934675216674805, + 0.0016361348098143935, + 0.003344598924741149, + 0.0016344484174624085, + 0.001126348040997982, + 0.0023397610057145357, + 0.05839519575238228, + 0.016633622348308563, + 0.018280694261193275, + 0.08993285149335861, + 0.012770788744091988, + 0.013986123725771904, + 0.02704707533121109, + 0.01157901156693697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.009771609678864479 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered from a static, high-angle, top-down perspective, the scene is rendered in a hyper-detailed, gongbi-style painting, emphasizing every fine line and subtle texture with artistic precision. A pair of gleaming chrome hairdressing scissors and a magnificent pale yellow jade comb are meticulously arranged on a dark, high-gloss lacquered wooden tray. The tray itself rests on a bed of deep crimson silk, its soft, luxurious folds catching a soft, focused beam of overhead light, causing the polished chrome to gleam and the translucent jade to glow from within. As the first delicate, needle-like particles of ice begin to form on the cold metal of the scissors' blades, a slow, mesmerizing crystallization begins. Steadily and deliberately, intricate patterns of frost-like fragments bloom, spiderwebs spreading across the polished metal surface and the cool, richly textured stone of the comb. Growth is a silent, captivating dance, as the tools gradually become completely encased in a delicate, translucent crystal structure, transforming them into a single, frozen sculpture. Their original form remains visible beneath the shimmering, faceted ice, creating an atmosphere of tranquil and magical beauty.", + "seed": 13057, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_15/index_8510_seed_13057_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.19381282567046582, + "vlm_logprob": -0.19381282567046582, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.19381282567046582, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9001051783561707 + ], + "opposite_logprob_tokens": [ + -0.5251051783561707 + ], + "logprob_reward": -0.9001051783561707, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9001051783561707, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15462420880794525, + 0.05425696074962616, + 0.06911633163690567, + 0.1526450216770172, + 0.11405587941408157, + 0.09491562098264694, + 0.08610313385725021, + 0.03222062066197395, + 0.20536655187606812, + 0.03669566288590431 + ], + "credit_map": [ + 0.018089983612298965, + 0.026080120354890823, + 0.02176392450928688, + 0.016328750178217888, + 0.02551187388598919, + 0.018797585740685463, + 0.013566263020038605, + 0.014485699124634266, + 0.008789760991930962, + 0.006013865116983652, + 0.0077202715910971165, + 0.006754973903298378, + 0.007444670889526606, + 0.007698023226112127, + 0.004059426486492157, + 0.005775968544185162, + 0.007622163742780685, + 0.003912923391908407, + 0.0084166768938303, + 0.005217313300818205, + 0.006179890129715204, + 0.022419562563300133, + 0.01005700696259737, + 0.005290800705552101, + 0.009335574693977833, + 0.011339614167809486, + 0.010565897449851036, + 0.018820224329829216, + 0.010245578363537788, + 0.00965904537588358, + 0.04724867641925812, + 0.035430409014225006, + 0.01306938286870718, + 0.017254259437322617, + 0.014817466959357262, + 0.017274364829063416, + 0.016325725242495537, + 0.012246406637132168, + 0.01170810591429472, + 0.011360159143805504, + 0.010331226512789726, + 0.01699788123369217, + 0.009845765307545662, + 0.015266389586031437, + 0.011532802134752274, + 0.011055906303226948, + 0.009376633912324905, + 0.010509020648896694, + 0.010532151907682419, + 0.00755893113091588, + 0.010663586668670177, + 0.014626266434788704, + 0.010024411603808403, + 0.013194529339671135, + 0.010456075891852379, + 0.009047174826264381, + 0.0031520246993750334, + 0.002799462992697954, + 0.003723561530932784, + 0.0047224597074091434, + 0.004956263583153486, + 0.0035343363415449858, + 0.005212469957768917, + 0.004120041616261005, + 0.04537028819322586, + 0.008675548247992992, + 0.00920451246201992, + 0.01566481590270996, + 0.07515086978673935, + 0.01587212271988392, + 0.022813094779849052, + 0.012615319341421127, + 0.00732825044542551, + 0.0025415390264242887, + 0.0036179213784635067, + 0.003405017778277397, + 0.007715433835983276, + 0.003830946749076247, + 0.0049155596643686295, + 0.0033409944735467434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.9001051783561707 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028299542609602213 + ], + "opposite_logprob_tokens": [ + -5.8778300285339355 + ], + "logprob_reward": -0.0028299542609602213, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028299542609602213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11031639575958252, + 0.053203921765089035, + 0.05168018490076065, + 0.16030992567539215, + 0.09648224711418152, + 0.10727966576814651, + 0.16969119012355804, + 0.025723116472363472, + 0.1875414103269577, + 0.03777196630835533 + ], + "credit_map": [ + 0.017290771007537842, + 0.016509082168340683, + 0.018589142709970474, + 0.010998251847922802, + 0.016289647668600082, + 0.011869700625538826, + 0.007667367812246084, + 0.011102425865828991, + 0.008164748549461365, + 0.005372909363359213, + 0.006640422623604536, + 0.006626126356422901, + 0.007964170537889004, + 0.007303230930119753, + 0.0051642050966620445, + 0.005968110635876656, + 0.004639617633074522, + 0.002351257484406233, + 0.0049823252484202385, + 0.004613990895450115, + 0.004333703778684139, + 0.01161892618983984, + 0.011466575786471367, + 0.007673785090446472, + 0.015364891849458218, + 0.018945250660181046, + 0.01785445027053356, + 0.021384544670581818, + 0.016653524711728096, + 0.020435694605112076, + 0.029377760365605354, + 0.02029380388557911, + 0.009846501983702183, + 0.01690748706459999, + 0.012616997584700584, + 0.014506609179079533, + 0.013953855261206627, + 0.009542460553348064, + 0.009838362224400043, + 0.009269967675209045, + 0.013801638968288898, + 0.019274326041340828, + 0.010651592165231705, + 0.014177007600665092, + 0.013557651080191135, + 0.01412682980298996, + 0.014142372645437717, + 0.007548247464001179, + 0.01906231790781021, + 0.014600533992052078, + 0.01960289292037487, + 0.026749717071652412, + 0.019903166219592094, + 0.02638734132051468, + 0.021837757900357246, + 0.0215474683791399, + 0.0030860635451972485, + 0.0022048265673220158, + 0.0024841325357556343, + 0.003819725476205349, + 0.0041283839382231236, + 0.002834067214280367, + 0.004297733772546053, + 0.002868184121325612, + 0.04095279425382614, + 0.007676499430090189, + 0.008208667859435081, + 0.01795835979282856, + 0.06797432899475098, + 0.014149727299809456, + 0.02003915049135685, + 0.010581872425973415, + 0.0055955806747078896, + 0.00262805400416255, + 0.002654493087902665, + 0.0039673857390880585, + 0.00781838409602642, + 0.006419999524950981, + 0.004795853979885578, + 0.0038922163657844067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.0028299542609602213 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005942060146480799 + ], + "opposite_logprob_tokens": [ + -5.130941867828369 + ], + "logprob_reward": -0.005942060146480799, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942060146480799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07675426453351974, + 0.04711296781897545, + 0.05667255446314812, + 0.42414847016334534, + 0.03734512999653816, + 0.09622316807508469, + 0.1343679577112198, + 0.014786886051297188, + 0.07074958831071854, + 0.041839033365249634 + ], + "credit_map": [ + 0.010305401869118214, + 0.012524344027042389, + 0.01314151007682085, + 0.00820084661245346, + 0.01124782394617796, + 0.008500464260578156, + 0.005415497813373804, + 0.007418379653245211, + 0.006329421419650316, + 0.0044980417005717754, + 0.007559550926089287, + 0.004344054032117128, + 0.005068853963166475, + 0.008955925703048706, + 0.005502046551555395, + 0.004855071660131216, + 0.007434704340994358, + 0.004509308375418186, + 0.006467768456786871, + 0.00715383468195796, + 0.005481036379933357, + 0.014866482466459274, + 0.006050141528248787, + 0.004709276836365461, + 0.047233305871486664, + 0.046339262276887894, + 0.06013020500540733, + 0.06534436345100403, + 0.05172372981905937, + 0.0683525800704956, + 0.03967515006661415, + 0.045349862426519394, + 0.006120091769844294, + 0.005220303311944008, + 0.005885603372007608, + 0.004940561950206757, + 0.004284245427697897, + 0.0033939366694539785, + 0.0035934727638959885, + 0.0039069135673344135, + 0.01190729346126318, + 0.017582468688488007, + 0.010437896475195885, + 0.013282516039907932, + 0.012386721558868885, + 0.010853521525859833, + 0.010074888356029987, + 0.009697859175503254, + 0.014893929474055767, + 0.011092024855315685, + 0.015598420053720474, + 0.02123820222914219, + 0.015918752178549767, + 0.02161516062915325, + 0.017219210043549538, + 0.01679224893450737, + 0.0016273414948955178, + 0.0014221264282241464, + 0.0022478734608739614, + 0.001758433529175818, + 0.0022597857750952244, + 0.0018867056351155043, + 0.002265171380713582, + 0.001319447997957468, + 0.013200835324823856, + 0.0033926069736480713, + 0.00370970880612731, + 0.010855955071747303, + 0.02184450440108776, + 0.005624745972454548, + 0.0072291456162929535, + 0.0048920828849077225, + 0.0056777154095470905, + 0.003157627070322633, + 0.0034310382325202227, + 0.004527653567492962, + 0.008134887553751469, + 0.007401815615594387, + 0.005449021700769663, + 0.004059275146573782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.005942060146480799 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525260902941227 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.00525260902941227, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525260902941227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06586931645870209, + 0.049410149455070496, + 0.025737568736076355, + 0.05224868655204773, + 0.09114207327365875, + 0.16004666686058044, + 0.16673465073108673, + 0.048228874802589417, + 0.31873011589050293, + 0.021851904690265656 + ], + "credit_map": [ + 0.010101979598402977, + 0.009675811044871807, + 0.010486896149814129, + 0.007194163277745247, + 0.008465769700706005, + 0.007505488581955433, + 0.005339961498975754, + 0.007099243346601725, + 0.005938708316534758, + 0.004444364923983812, + 0.006140911485999823, + 0.008853847160935402, + 0.007116167806088924, + 0.005735708400607109, + 0.0054099117405712605, + 0.00577052915468812, + 0.003104289760813117, + 0.002349867019802332, + 0.0023815413005650043, + 0.003143097972497344, + 0.002579758409410715, + 0.0026648074854165316, + 0.005417726002633572, + 0.004096480086445808, + 0.0037656519562005997, + 0.0059257070533931255, + 0.004666003864258528, + 0.005950259510427713, + 0.005002686288207769, + 0.004185251425951719, + 0.012689536437392235, + 0.010063592344522476, + 0.010016010142862797, + 0.014392971992492676, + 0.01202465407550335, + 0.015195930376648903, + 0.012898330576717854, + 0.008493755012750626, + 0.009422914125025272, + 0.008697504177689552, + 0.019009236246347427, + 0.03203941136598587, + 0.015533709898591042, + 0.022731773555278778, + 0.020565547049045563, + 0.02012273669242859, + 0.01823313534259796, + 0.01181110180914402, + 0.018922099843621254, + 0.013683313503861427, + 0.019352933391928673, + 0.025192709639668465, + 0.019524194300174713, + 0.02400011196732521, + 0.021816793829202652, + 0.024242499843239784, + 0.005671372637152672, + 0.004435369744896889, + 0.005652351770550013, + 0.006481826771050692, + 0.007917345501482487, + 0.00554838590323925, + 0.007608432788401842, + 0.004913792945444584, + 0.0743366926908493, + 0.01316970493644476, + 0.013156323693692684, + 0.02126239612698555, + 0.1232428029179573, + 0.023162977769970894, + 0.0345635712146759, + 0.015835653990507126, + 0.004042258020490408, + 0.0013300487771630287, + 0.0016480933409184217, + 0.001664346782490611, + 0.005870606284588575, + 0.0017344094812870026, + 0.0033567349892109632, + 0.0022054086439311504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.00525260902941227 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493432655930519 + ], + "opposite_logprob_tokens": [ + -2.92993426322937 + ], + "logprob_reward": -0.05493432655930519, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493432655930519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12346493452787399, + 0.05312252417206764, + 0.06660467386245728, + 0.08111666887998581, + 0.05451282113790512, + 0.0663561075925827, + 0.15753662586212158, + 0.07438955456018448, + 0.13204438984394073, + 0.19085170328617096 + ], + "credit_map": [ + 0.01517802570015192, + 0.017820416018366814, + 0.022783422842621803, + 0.01466693077236414, + 0.018733710050582886, + 0.015799270942807198, + 0.009331737644970417, + 0.00915142334997654, + 0.006541363894939423, + 0.005980541929602623, + 0.006937234662473202, + 0.00758868083357811, + 0.00868118368089199, + 0.006677946075797081, + 0.005007689818739891, + 0.005707885138690472, + 0.009678629226982594, + 0.006293899845331907, + 0.010943474248051643, + 0.0078045339323580265, + 0.007239169906824827, + 0.012078532949090004, + 0.006662909407168627, + 0.005903523415327072, + 0.008583968505263329, + 0.009690824896097183, + 0.009515558369457722, + 0.011310744099318981, + 0.008913254365324974, + 0.010481039993464947, + 0.010030847042798996, + 0.012590430676937103, + 0.005242191720753908, + 0.004868768621236086, + 0.007375503424555063, + 0.009459373541176319, + 0.004656217526644468, + 0.006344372406601906, + 0.009797480888664722, + 0.006768911611288786, + 0.006238176487386227, + 0.012613234110176563, + 0.0063662403263151646, + 0.009031934663653374, + 0.006084670778363943, + 0.00669954065233469, + 0.012088716961443424, + 0.007233600132167339, + 0.018928950652480125, + 0.013706495985388756, + 0.01866908371448517, + 0.026039326563477516, + 0.018757019191980362, + 0.024490956217050552, + 0.019204890355467796, + 0.01773991622030735, + 0.008345048874616623, + 0.005689284298568964, + 0.0094313258305192, + 0.010870693251490593, + 0.011489416472613811, + 0.006871731951832771, + 0.011506518349051476, + 0.010185537859797478, + 0.023901546373963356, + 0.008804194629192352, + 0.007082989905029535, + 0.021424464881420135, + 0.032883793115615845, + 0.01723003014922142, + 0.012505738995969296, + 0.008211633190512657, + 0.02346261963248253, + 0.016755780205130577, + 0.016130730509757996, + 0.024483073502779007, + 0.028526782989501953, + 0.03486543893814087, + 0.02830541506409645, + 0.018321875482797623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.05493432655930519 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered from a static, high-angle, top-down perspective, the scene is rendered in a hyper-detailed, gongbi-style painting, emphasizing every fine line and subtle texture with artistic precision. A pair of gleaming chrome hairdressing scissors and a magnificent pale yellow jade comb are meticulously arranged on a dark, high-gloss lacquered wooden tray. The tray itself rests on a bed of deep crimson silk, its soft, luxurious folds catching a soft, focused beam of overhead light, causing the polished chrome to gleam and the translucent jade to glow from within. As the first delicate, needle-like particles of ice begin to form on the cold metal of the scissors' blades, a slow, mesmerizing crystallization begins. Steadily and deliberately, intricate patterns of frost-like fragments bloom, spiderwebs spreading across the polished metal surface and the cool, richly textured stone of the comb. Growth is a silent, captivating dance, as the tools gradually become completely encased in a delicate, translucent crystal structure, transforming them into a single, frozen sculpture. Their original form remains visible beneath the shimmering, faceted ice, creating an atmosphere of tranquil and magical beauty.", + "seed": 13058, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_15/index_8510_seed_13058_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21879539894871414, + "vlm_logprob": -0.21879539894871414, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.21879539894871414, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0553570985794067 + ], + "opposite_logprob_tokens": [ + -0.43035706877708435 + ], + "logprob_reward": -1.0553570985794067, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0553570985794067, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06753502786159515, + 0.024511586874723434, + 0.020159659907221794, + 0.22583238780498505, + 0.050299640744924545, + 0.08127608895301819, + 0.09411665052175522, + 0.21516214311122894, + 0.03465158864855766, + 0.18645520508289337 + ], + "credit_map": [ + 0.006876994390040636, + 0.005953729152679443, + 0.00891798548400402, + 0.011686788871884346, + 0.00820132251828909, + 0.0095783406868577, + 0.008758201263844967, + 0.007561672478914261, + 0.0035260904114693403, + 0.0024652827996760607, + 0.00375297456048429, + 0.0036085715983062983, + 0.0036104072351008654, + 0.002448001876473427, + 0.0024094944819808006, + 0.0026907650753855705, + 0.0027711177244782448, + 0.0023571415804326534, + 0.002580559579655528, + 0.003375481581315398, + 0.0022508110851049423, + 0.002175395144149661, + 0.002331988187506795, + 0.002317166654393077, + 0.029868628829717636, + 0.020703207701444626, + 0.011392946355044842, + 0.009365339763462543, + 0.08444693684577942, + 0.028147656470537186, + 0.011967687867581844, + 0.02993999607861042, + 0.0059431167319417, + 0.0044166045263409615, + 0.003730079159140587, + 0.005157269537448883, + 0.01698438636958599, + 0.00672066630795598, + 0.003999763168394566, + 0.00334775447845459, + 0.0049591208808124065, + 0.005414330400526524, + 0.01482407096773386, + 0.029377728700637817, + 0.009480494074523449, + 0.0038033092860132456, + 0.009568246081471443, + 0.0038487922865897417, + 0.008326804265379906, + 0.0101506682112813, + 0.010361719876527786, + 0.013520289212465286, + 0.020728223025798798, + 0.010755340568721294, + 0.007752920035272837, + 0.012520691379904747, + 0.021357227116823196, + 0.023552972823381424, + 0.01587754301726818, + 0.006766107864677906, + 0.08820968121290207, + 0.04061970114707947, + 0.008614313788712025, + 0.010164606384932995, + 0.0056580062955617905, + 0.0035814011935144663, + 0.00377028901129961, + 0.0040627457201480865, + 0.005066187586635351, + 0.0033013508655130863, + 0.004871021490544081, + 0.004340585321187973, + 0.016482478007674217, + 0.03564680367708206, + 0.01224292442202568, + 0.00773809477686882, + 0.07018441706895828, + 0.023853754624724388, + 0.010600426234304905, + 0.009706314653158188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 9 + ], + "selected_reward": -1.0553570985794067 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004107254557311535 + ], + "opposite_logprob_tokens": [ + -5.504107475280762 + ], + "logprob_reward": -0.004107254557311535, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004107254557311535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3171156048774719, + 0.0665033832192421, + 0.022245090454816818, + 0.1325622797012329, + 0.023237833753228188, + 0.03225502371788025, + 0.04813981428742409, + 0.18629905581474304, + 0.04204878583550453, + 0.1295931190252304 + ], + "credit_map": [ + 0.08431115746498108, + 0.022486595436930656, + 0.021560026332736015, + 0.017189795151352882, + 0.0963970348238945, + 0.02920144982635975, + 0.022492019459605217, + 0.023477528244256973, + 0.011170990765094757, + 0.004725286271423101, + 0.004355371929705143, + 0.004123997874557972, + 0.022987084463238716, + 0.007910114713013172, + 0.004801640287041664, + 0.006428894586861134, + 0.0028155434411019087, + 0.002025346038863063, + 0.0019389215158298612, + 0.0022556863259524107, + 0.0067040883004665375, + 0.002726556034758687, + 0.0016740587307140231, + 0.002104890998452902, + 0.019142987206578255, + 0.012246828526258469, + 0.007278101984411478, + 0.005697998683899641, + 0.04871254041790962, + 0.015946470201015472, + 0.007183840498328209, + 0.016353510320186615, + 0.0019635676871985197, + 0.002145598642528057, + 0.0034271483309566975, + 0.0034378068521618843, + 0.00502390181645751, + 0.0029646106995642185, + 0.002284238114953041, + 0.0019909623079001904, + 0.0046027107164263725, + 0.0023132148198783398, + 0.002970564179122448, + 0.0045326207764446735, + 0.008844460360705853, + 0.0036345673725008965, + 0.0023342494387179613, + 0.0030226358212530613, + 0.006816653069108725, + 0.005101378075778484, + 0.007495936006307602, + 0.008791990578174591, + 0.004439208190888166, + 0.004051374737173319, + 0.004234166815876961, + 0.007209110073745251, + 0.018063781782984734, + 0.019923893734812737, + 0.014165883883833885, + 0.006248917896300554, + 0.07566943019628525, + 0.03535396605730057, + 0.007236725650727749, + 0.00963645614683628, + 0.007169798482209444, + 0.00469312584027648, + 0.004378533456474543, + 0.004565929062664509, + 0.0070537361316382885, + 0.003724653273820877, + 0.005182911176234484, + 0.005280099343508482, + 0.012827454134821892, + 0.020400507375597954, + 0.0098415557295084, + 0.005947359371930361, + 0.0470881387591362, + 0.017437687143683434, + 0.008273188024759293, + 0.0077772256918251514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.004107254557311535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003609453560784459 + ], + "opposite_logprob_tokens": [ + -5.628609657287598 + ], + "logprob_reward": -0.003609453560784459, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003609453560784459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11901858448982239, + 0.05981336534023285, + 0.02969956211745739, + 0.14016832411289215, + 0.036862850189208984, + 0.04722538962960243, + 0.10062742978334427, + 0.2904300391674042, + 0.08084806799888611, + 0.09530637413263321 + ], + "credit_map": [ + 0.014145546592772007, + 0.009623789228498936, + 0.01253517810255289, + 0.014242435805499554, + 0.026997679844498634, + 0.013724356889724731, + 0.013275649398565292, + 0.01447396818548441, + 0.009454379789531231, + 0.004423853475600481, + 0.00503412913531065, + 0.006630346179008484, + 0.018696827813982964, + 0.007295594085007906, + 0.004372486378997564, + 0.003905755700543523, + 0.004719128832221031, + 0.002197264228016138, + 0.0026184467133134604, + 0.003445488167926669, + 0.008775935508310795, + 0.003633046057075262, + 0.0021071957889944315, + 0.002203060546889901, + 0.02065589278936386, + 0.012954577803611755, + 0.008349663577973843, + 0.006314681842923164, + 0.05055404454469681, + 0.01643747091293335, + 0.007479426451027393, + 0.01742257922887802, + 0.0035854121670126915, + 0.003470765892416239, + 0.006173546425998211, + 0.006956030149012804, + 0.004738152492791414, + 0.0030018347315490246, + 0.003631159197539091, + 0.005305952858179808, + 0.00950949639081955, + 0.005655253771692514, + 0.005442650523036718, + 0.0072329058311879635, + 0.009047462604939938, + 0.003996542654931545, + 0.0038027458358556032, + 0.0025383366737514734, + 0.007613320369273424, + 0.00973839033395052, + 0.011391139589250088, + 0.016916785389184952, + 0.02983396127820015, + 0.011399340815842152, + 0.00701511837542057, + 0.0067193834111094475, + 0.025556528940796852, + 0.031192224472761154, + 0.02575533650815487, + 0.012892481870949268, + 0.11358118057250977, + 0.05022123083472252, + 0.014414102770388126, + 0.01681698113679886, + 0.01519132312387228, + 0.010178672149777412, + 0.008524630218744278, + 0.007169621530920267, + 0.012864382937550545, + 0.007875772193074226, + 0.00865800492465496, + 0.010385668836534023, + 0.019391940906643867, + 0.016454337164759636, + 0.011212139390408993, + 0.00805159192532301, + 0.015140477567911148, + 0.007291213143616915, + 0.008926418609917164, + 0.008838264271616936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.003609453560784459 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630487438291311 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630487438291311, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630487438291311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33868005871772766, + 0.025505604222416878, + 0.024963997304439545, + 0.06435863673686981, + 0.01545256283134222, + 0.12411997467279434, + 0.14263306558132172, + 0.16035377979278564, + 0.05028701201081276, + 0.0536452941596508 + ], + "credit_map": [ + 0.09014089405536652, + 0.02390461415052414, + 0.02303287945687771, + 0.017839284613728523, + 0.1035083681344986, + 0.030645236372947693, + 0.023848554119467735, + 0.025760238990187645, + 0.0019630135502666235, + 0.002345405053347349, + 0.0036565824411809444, + 0.0036328784190118313, + 0.005406599026173353, + 0.0025959424674510956, + 0.0029794133733958006, + 0.0029257680289447308, + 0.0036865267902612686, + 0.002468602266162634, + 0.00257531413808465, + 0.0036769940052181482, + 0.005834706127643585, + 0.0026062140241265297, + 0.0021673382725566626, + 0.0019483021460473537, + 0.009370201267302036, + 0.006051203235983849, + 0.0038233143277466297, + 0.002755471272394061, + 0.02318340726196766, + 0.0076243216171860695, + 0.0033753998577594757, + 0.008175319992005825, + 0.0015111285028979182, + 0.001521604135632515, + 0.002315102843567729, + 0.002367938868701458, + 0.002250264398753643, + 0.001396762439981103, + 0.002139301970601082, + 0.001950459904037416, + 0.022655455395579338, + 0.008092357777059078, + 0.012316771782934666, + 0.017698684707283974, + 0.03465865179896355, + 0.01180092804133892, + 0.0058422451838850975, + 0.011054875329136848, + 0.010352292098104954, + 0.011768837459385395, + 0.015218897722661495, + 0.025454968214035034, + 0.04607977718114853, + 0.016545403748750687, + 0.010285059921443462, + 0.006927829701453447, + 0.016686273738741875, + 0.017671197652816772, + 0.010058412328362465, + 0.005119186360388994, + 0.06602280586957932, + 0.031207183375954628, + 0.006263776682317257, + 0.0073249381966888905, + 0.005450150463730097, + 0.008408562280237675, + 0.004299483727663755, + 0.007022756151854992, + 0.007789531257003546, + 0.004040211904793978, + 0.007913410663604736, + 0.005362905561923981, + 0.007245880551636219, + 0.005224789027124643, + 0.0038853930309414864, + 0.003555343020707369, + 0.015319814905524254, + 0.0065847947262227535, + 0.006151323672384024, + 0.005677956622093916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.007630487438291311 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023272700607776642 + ], + "opposite_logprob_tokens": [ + -3.773272752761841 + ], + "logprob_reward": -0.023272700607776642, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023272700607776642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.144444078207016, + 0.0554780587553978, + 0.06550831347703934, + 0.058454353362321854, + 0.03353584185242653, + 0.2541884481906891, + 0.15508708357810974, + 0.1242368072271347, + 0.05518224090337753, + 0.053884804248809814 + ], + "credit_map": [ + 0.03664199262857437, + 0.010793698020279408, + 0.013285635970532894, + 0.011851372197270393, + 0.03428800776600838, + 0.014613717794418335, + 0.011974221095442772, + 0.010995428077876568, + 0.006060674320906401, + 0.005721692927181721, + 0.007605229038745165, + 0.008463832549750805, + 0.010484916158020496, + 0.005461791064590216, + 0.006031981203705072, + 0.005647940561175346, + 0.009010758250951767, + 0.005558725446462631, + 0.003222622675821185, + 0.002963270526379347, + 0.024841545149683952, + 0.009075885638594627, + 0.004131977446377277, + 0.006703528109937906, + 0.007062152959406376, + 0.005071435123682022, + 0.0038680024445056915, + 0.004036153666675091, + 0.019810054451227188, + 0.007791030220687389, + 0.003775149816647172, + 0.0070403749123215675, + 0.003068145364522934, + 0.003279728814959526, + 0.004256974905729294, + 0.004958643577992916, + 0.005683823022991419, + 0.0042774248868227005, + 0.0036841565743088722, + 0.004326947499066591, + 0.06117616593837738, + 0.023698754608631134, + 0.021104544401168823, + 0.012944722548127174, + 0.07409416139125824, + 0.026019534096121788, + 0.013023792766034603, + 0.02212676964700222, + 0.010020682588219643, + 0.01872158609330654, + 0.009862003847956657, + 0.0051755718886852264, + 0.06496430188417435, + 0.025330491364002228, + 0.005945681128650904, + 0.015066766180098057, + 0.012978781946003437, + 0.013392086140811443, + 0.009635698981583118, + 0.004629464820027351, + 0.049288950860500336, + 0.02203511819243431, + 0.0054319784976542, + 0.006844721268862486, + 0.008645232766866684, + 0.0062441290356218815, + 0.005917646922171116, + 0.005526703782379627, + 0.007725926116108894, + 0.006202885415405035, + 0.006587640848010778, + 0.00833207368850708, + 0.0074174716137349606, + 0.007279548794031143, + 0.004322585184127092, + 0.0028321808204054832, + 0.016369609162211418, + 0.006209359038621187, + 0.004810146056115627, + 0.0046439035795629025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.023272700607776642 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered from a static, high-angle, top-down perspective, the scene is rendered in a hyper-detailed, gongbi-style painting, emphasizing every fine line and subtle texture with artistic precision. A pair of gleaming chrome hairdressing scissors and a magnificent pale yellow jade comb are meticulously arranged on a dark, high-gloss lacquered wooden tray. The tray itself rests on a bed of deep crimson silk, its soft, luxurious folds catching a soft, focused beam of overhead light, causing the polished chrome to gleam and the translucent jade to glow from within. As the first delicate, needle-like particles of ice begin to form on the cold metal of the scissors' blades, a slow, mesmerizing crystallization begins. Steadily and deliberately, intricate patterns of frost-like fragments bloom, spiderwebs spreading across the polished metal surface and the cool, richly textured stone of the comb. Growth is a silent, captivating dance, as the tools gradually become completely encased in a delicate, translucent crystal structure, transforming them into a single, frozen sculpture. Their original form remains visible beneath the shimmering, faceted ice, creating an atmosphere of tranquil and magical beauty.", + "seed": 13059, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_15/index_8510_seed_13059_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.5250592410564423, + "vlm_logprob": -0.5250592410564423, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.5250592410564423, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.4644596576690674 + ], + "opposite_logprob_tokens": [ + -0.08945956081151962 + ], + "logprob_reward": -2.4644596576690674, + "yes_no_margin_reward": -2.375, + "token_credit_reward": -2.4644596576690674, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.46963056921958923, + 0.058565009385347366, + 0.039001211524009705, + 0.13012182712554932, + 0.049856942147016525, + 0.06876004487276077, + 0.0475681871175766, + 0.04227643832564354, + 0.015878718346357346, + 0.0783410295844078 + ], + "credit_map": [ + 0.09210888296365738, + 0.0347166508436203, + 0.030851494520902634, + 0.04397549107670784, + 0.037958599627017975, + 0.09102601557970047, + 0.08247899264097214, + 0.056514475494623184, + 0.009639454074203968, + 0.005755885969847441, + 0.005474924575537443, + 0.006487974431365728, + 0.005298225674778223, + 0.008047112263739109, + 0.009795140475034714, + 0.008066296577453613, + 0.0060548847541213036, + 0.0036981154698878527, + 0.0033468783367425203, + 0.004813170526176691, + 0.003325491212308407, + 0.006840139627456665, + 0.005734054837375879, + 0.005188478622585535, + 0.02037757635116577, + 0.012819230556488037, + 0.008432693779468536, + 0.01136084459722042, + 0.010857918299734592, + 0.027616724371910095, + 0.018726753070950508, + 0.019930096343159676, + 0.009105654433369637, + 0.004269409459084272, + 0.0034731572959572077, + 0.003653311636298895, + 0.003349650651216507, + 0.011776131577789783, + 0.006850103847682476, + 0.007379522081464529, + 0.009864771738648415, + 0.006459539756178856, + 0.007917333394289017, + 0.005861201789230108, + 0.007208535447716713, + 0.010330588556826115, + 0.010710285976529121, + 0.010407793335616589, + 0.00615124823525548, + 0.006439520511776209, + 0.004596015904098749, + 0.0039487360045313835, + 0.005439349450170994, + 0.007717002648860216, + 0.0073405844159424305, + 0.005935730878263712, + 0.005616183392703533, + 0.0039176298305392265, + 0.0039277751930058, + 0.005246412940323353, + 0.004985537845641375, + 0.007297592703253031, + 0.006222437601536512, + 0.00506287207826972, + 0.0026275431737303734, + 0.0021739762742072344, + 0.0016878328751772642, + 0.0018510717200115323, + 0.0013843115884810686, + 0.002836737083271146, + 0.0010085589019581676, + 0.002308688359335065, + 0.014111942611634731, + 0.008518043905496597, + 0.005980332847684622, + 0.010817388072609901, + 0.005267465021461248, + 0.015383944846689701, + 0.008709262125194073, + 0.009552654810249805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -2.4644596576690674 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07907227426767349 + ], + "opposite_logprob_tokens": [ + -2.5790722370147705 + ], + "logprob_reward": -0.07907227426767349, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07907227426767349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14362935721874237, + 0.1302807778120041, + 0.0887027159333229, + 0.24768607318401337, + 0.11011438816785812, + 0.06006155163049698, + 0.029334761202335358, + 0.030311686918139458, + 0.02496154047548771, + 0.13491712510585785 + ], + "credit_map": [ + 0.023765334859490395, + 0.011568672955036163, + 0.010721791535615921, + 0.01360316202044487, + 0.01534276269376278, + 0.024699080735445023, + 0.021833086386322975, + 0.022095469757914543, + 0.025037463754415512, + 0.006177518982440233, + 0.007081991992890835, + 0.0109022231772542, + 0.011308533139526844, + 0.02303963713347912, + 0.02966291643679142, + 0.017070500180125237, + 0.016321057453751564, + 0.004706972278654575, + 0.004783928394317627, + 0.007745081093162298, + 0.007741068489849567, + 0.016745736822485924, + 0.018137507140636444, + 0.012521356344223022, + 0.03965334966778755, + 0.025280006229877472, + 0.013503976166248322, + 0.022841712459921837, + 0.02143571712076664, + 0.04921505227684975, + 0.04222889989614487, + 0.03352736681699753, + 0.019133219495415688, + 0.009965155273675919, + 0.007053406909108162, + 0.009660225361585617, + 0.007901628501713276, + 0.023659875616431236, + 0.015900209546089172, + 0.01684066839516163, + 0.008013089187443256, + 0.005771229974925518, + 0.0060697137378156185, + 0.004582164343446493, + 0.0055795880034565926, + 0.0096531230956316, + 0.0099295973777771, + 0.010463044978678226, + 0.004613653291016817, + 0.002875699894502759, + 0.0032985853031277657, + 0.0026177000254392624, + 0.00262230122461915, + 0.006189943756908178, + 0.003492853371426463, + 0.0036240229383111, + 0.005071675404906273, + 0.0028575703036040068, + 0.002941989107057452, + 0.0033751281443983316, + 0.002821258967742324, + 0.005671295803040266, + 0.003881910815834999, + 0.0036908576730638742, + 0.004673972260206938, + 0.002078408608213067, + 0.0018959450535476208, + 0.0024610916152596474, + 0.0022623108234256506, + 0.0052717020735144615, + 0.003003302263095975, + 0.003314806381240487, + 0.020839236676692963, + 0.013419934548437595, + 0.012995865195989609, + 0.017642565071582794, + 0.01159107405692339, + 0.022473666816949844, + 0.014821241609752178, + 0.021133538335561752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.07907227426767349 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076509952545166 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076509952545166, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076509952545166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12861670553684235, + 0.05674886330962181, + 0.08655396848917007, + 0.38232553005218506, + 0.05379720777273178, + 0.043674830347299576, + 0.016864655539393425, + 0.057710595428943634, + 0.049017637968063354, + 0.12469002604484558 + ], + "credit_map": [ + 0.02345353551208973, + 0.010307660326361656, + 0.010316631756722927, + 0.011474058963358402, + 0.012177451513707638, + 0.026367388665676117, + 0.015363827347755432, + 0.019156143069267273, + 0.010140471160411835, + 0.002887042937800288, + 0.003395256819203496, + 0.004982808604836464, + 0.0048365239053964615, + 0.010264324955642223, + 0.012518775649368763, + 0.007723655551671982, + 0.017629986628890038, + 0.004410250578075647, + 0.004492027685046196, + 0.007578867953270674, + 0.006805980112403631, + 0.017821531742811203, + 0.016121232882142067, + 0.011694090440869331, + 0.06500651687383652, + 0.035502683371305466, + 0.020634323358535767, + 0.03501850366592407, + 0.029289131984114647, + 0.0863756611943245, + 0.059323135763406754, + 0.051175571978092194, + 0.008254208602011204, + 0.005500334315001965, + 0.004297043662518263, + 0.003988821990787983, + 0.003923467826098204, + 0.013581601902842522, + 0.006665410939604044, + 0.007586312480270863, + 0.0073890988714993, + 0.0035235299728810787, + 0.0032564157154411077, + 0.002874027704820037, + 0.00440965173766017, + 0.009141516871750355, + 0.006213259883224964, + 0.006867328193038702, + 0.002070722170174122, + 0.002276555635035038, + 0.0015951949171721935, + 0.0016541380900889635, + 0.002192845568060875, + 0.002558761043474078, + 0.0023154073860496283, + 0.002201029332354665, + 0.008557595312595367, + 0.006704553030431271, + 0.005033750087022781, + 0.006333657540380955, + 0.007800240535289049, + 0.009491940028965473, + 0.007354097906500101, + 0.0064347549341619015, + 0.007899582386016846, + 0.004613610915839672, + 0.003977741580456495, + 0.004609028343111277, + 0.00546124717220664, + 0.01035298127681017, + 0.00528492359444499, + 0.006818519439548254, + 0.025175675749778748, + 0.011203947477042675, + 0.009823629632592201, + 0.0091883335262537, + 0.007435759995132685, + 0.02978585474193096, + 0.008718655444681644, + 0.023358160629868507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.011076509952545166 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646313101053238 + ], + "opposite_logprob_tokens": [ + -4.758646488189697 + ], + "logprob_reward": -0.008646313101053238, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646313101053238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20502886176109314, + 0.035094983875751495, + 0.15979982912540436, + 0.10765588283538818, + 0.056605517864227295, + 0.09903206676244736, + 0.10906310379505157, + 0.06319553405046463, + 0.04567769542336464, + 0.11884652823209763 + ], + "credit_map": [ + 0.034092165529727936, + 0.014065491035580635, + 0.0144199812784791, + 0.017743447795510292, + 0.018403517082333565, + 0.04460930824279785, + 0.030830856412649155, + 0.030864093452692032, + 0.005026038736104965, + 0.00296092894859612, + 0.002574311336502433, + 0.0035125266294926405, + 0.003874580143019557, + 0.004509856458753347, + 0.007784772198647261, + 0.004851970821619034, + 0.0323810875415802, + 0.007871412672102451, + 0.00784920621663332, + 0.013276259414851665, + 0.01240128930658102, + 0.031001441180706024, + 0.03359895199537277, + 0.021420177072286606, + 0.019601386040449142, + 0.008102732710540295, + 0.006179744843393564, + 0.008997593075037003, + 0.007732061669230461, + 0.026629142463207245, + 0.014253247529268265, + 0.01615997776389122, + 0.007685928139835596, + 0.00811329297721386, + 0.005900042597204447, + 0.006246333476155996, + 0.004988417029380798, + 0.007259100675582886, + 0.009169057011604309, + 0.007243345491588116, + 0.013298406265676022, + 0.008328540250658989, + 0.009784344583749771, + 0.008401845581829548, + 0.008488430641591549, + 0.022218327969312668, + 0.014165445230901241, + 0.014346715994179249, + 0.015352332964539528, + 0.013241784647107124, + 0.011254948563873768, + 0.00999073963612318, + 0.01343671977519989, + 0.018742308020591736, + 0.013199723325669765, + 0.013844546861946583, + 0.00835607759654522, + 0.0063850972801446915, + 0.005123510491102934, + 0.007990473881363869, + 0.008015291765332222, + 0.01149784680455923, + 0.008371896110475063, + 0.007455340586602688, + 0.006909732241183519, + 0.004993529058992863, + 0.00394442118704319, + 0.0044598765671253204, + 0.0059310635551810265, + 0.006920638028532267, + 0.0073584578931331635, + 0.005159978289157152, + 0.02228393405675888, + 0.013241301290690899, + 0.009407684206962585, + 0.0154340211302042, + 0.008306127041578293, + 0.02081088162958622, + 0.015402808785438538, + 0.013959773816168308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.008646313101053238 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a hyper-detailed gongbi-style painting with fine lines and subtle textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are gleaming chrome hairdressing scissors and a pale yellow jade comb visible on a dark high-gloss lacquered wooden tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tray resting on a bed of deep crimson silk with soft, luxurious folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do intricate patterns of frost-like fragments bloom and spread across the scissors and comb, eventually encasing them in a translucent crystal structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene viewed from a static, high-angle, top-down perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062041450291872025 + ], + "opposite_logprob_tokens": [ + -2.8120415210723877 + ], + "logprob_reward": -0.062041450291872025, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062041450291872025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2181420773267746, + 0.15710051357746124, + 0.04671251028776169, + 0.22053785622119904, + 0.05552642419934273, + 0.07339942455291748, + 0.0332556813955307, + 0.052787281572818756, + 0.0489836186170578, + 0.09355458617210388 + ], + "credit_map": [ + 0.04181025177240372, + 0.011236759833991528, + 0.013694526627659798, + 0.020185520872473717, + 0.017578614875674248, + 0.04781295731663704, + 0.038620054721832275, + 0.027203382924199104, + 0.030815305188298225, + 0.008014671504497528, + 0.009708528406918049, + 0.015003000386059284, + 0.011116237379610538, + 0.03449977934360504, + 0.03001241944730282, + 0.01793055422604084, + 0.01023772731423378, + 0.0028017237782478333, + 0.0036038130056113005, + 0.0033562886528670788, + 0.003685273230075836, + 0.007980521768331528, + 0.00911482609808445, + 0.005932334344834089, + 0.040226783603429794, + 0.020115243270993233, + 0.012812528759241104, + 0.020613282918930054, + 0.016201147809624672, + 0.050898678600788116, + 0.03270205855369568, + 0.026968136429786682, + 0.009368114173412323, + 0.00638990942388773, + 0.004120724741369486, + 0.005074478220194578, + 0.0037697141524404287, + 0.011021551676094532, + 0.006443644408136606, + 0.009338288567960262, + 0.011887677945196629, + 0.007333260960876942, + 0.006392196752130985, + 0.006166514940559864, + 0.005984894931316376, + 0.01455870270729065, + 0.008625329472124577, + 0.012450837530195713, + 0.0044767120853066444, + 0.0027531462255865335, + 0.004241093061864376, + 0.004250590689480305, + 0.003547272179275751, + 0.005244027823209763, + 0.004079853650182486, + 0.004662984516471624, + 0.007599381729960442, + 0.006281012203544378, + 0.004910988733172417, + 0.0072392686270177364, + 0.0053231315687298775, + 0.00823823269456625, + 0.006074625998735428, + 0.00712064141407609, + 0.009905287995934486, + 0.004773941822350025, + 0.0046915956772863865, + 0.004721560515463352, + 0.005088600795716047, + 0.007742777466773987, + 0.005258546676486731, + 0.006801310461014509, + 0.020491372793912888, + 0.009565907530486584, + 0.007871938869357109, + 0.011220189742743969, + 0.006332542281597853, + 0.018011214211583138, + 0.006974873133003712, + 0.013086560182273388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.062041450291872025 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_2/index_tensor([2388, 2388])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_2/index_tensor([2388, 2388])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d119f5dc57074a7b8cd5ab3721181c4e9d01ebc9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_2/index_tensor([2388, 2388])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In an extreme close-up, the camera remains still, focused on a newly emerged butterfly clinging to a emerald green stem, next to the translucent, pale green shell it had escaped from. A warm, soft morning light streams in from behind, casting a captivating golden rim around the butterfly's furry black body and the edges of its wings, which have yet to fully unfold. Its wrinkled, damp wings appear translucent against the soft-focus background, their intricate veins of orange and black barely visible. The wings tremble slightly, almost imperceptibly, as if gathering strength. Then, steadily and deliberately, they unfold, bit by bit. As the fluids slowly flow in, the wrinkles gradually smooth out, and the pattern on the wings becomes clearer and more vivid. The entire picture is shrouded in a tranquil and hopeful warm-toned atmosphere. The green leaves dotted with morning dew in the background form dreamy light spots, together forming a poetic picture of rebirth and transformation.", + "seed": 21607, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_2/index_2388_seed_21607_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04412276158109307, + "vlm_logprob": -0.04412276158109307, + "vlm_yes_no_margin": 4.675000023841858, + "vlm_token_credit": -0.04412276158109307, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19361652433872223 + ], + "opposite_logprob_tokens": [ + -2.0686166286468506 + ], + "logprob_reward": -0.19361652433872223, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.19361652433872223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19260524213314056, + 0.16086816787719727, + 0.041916534304618835, + 0.08128996938467026, + 0.10803390294313431, + 0.1017439067363739, + 0.09434064477682114, + 0.08568306267261505, + 0.09454595297574997, + 0.038972631096839905 + ], + "credit_map": [ + 0.021174324676394463, + 0.029175935313105583, + 0.013883756473660469, + 0.015919452533125877, + 0.03384295105934143, + 0.02805786393582821, + 0.0175172071903944, + 0.03303373605012894, + 0.016651663929224014, + 0.015160276554524899, + 0.015227043069899082, + 0.018917381763458252, + 0.02964998595416546, + 0.022893154993653297, + 0.021464740857481956, + 0.020903928205370903, + 0.004206574056297541, + 0.006381482817232609, + 0.004857285413891077, + 0.0034253064077347517, + 0.006103845778852701, + 0.0069204713217914104, + 0.003888828679919243, + 0.006132740061730146, + 0.007873770780861378, + 0.01203036867082119, + 0.006703947205096483, + 0.006445372011512518, + 0.015869583934545517, + 0.012459199875593185, + 0.00901114847511053, + 0.010896575637161732, + 0.011465039104223251, + 0.01975121721625328, + 0.009577031247317791, + 0.010373526252806187, + 0.01792857050895691, + 0.012672762386500835, + 0.011686996556818485, + 0.014578756876289845, + 0.010190018452703953, + 0.015198394656181335, + 0.009721052832901478, + 0.01074009109288454, + 0.018242403864860535, + 0.011454559862613678, + 0.010217346251010895, + 0.015980029478669167, + 0.009396853856742382, + 0.010362094268202782, + 0.009529240429401398, + 0.008812973275780678, + 0.015513517893850803, + 0.013872000388801098, + 0.014391792938113213, + 0.012462174519896507, + 0.010912532918155193, + 0.01117675006389618, + 0.006925240624696016, + 0.006701877806335688, + 0.015156312845647335, + 0.009105457924306393, + 0.014347796328365803, + 0.011357094161212444, + 0.0094071626663208, + 0.011075596325099468, + 0.006300534587353468, + 0.005894804839044809, + 0.017258943989872932, + 0.014190269634127617, + 0.012667826376855373, + 0.0177508145570755, + 0.0036441804841160774, + 0.004164501558989286, + 0.0021628541871905327, + 0.0019380876328796148, + 0.007899198681116104, + 0.008018537424504757, + 0.0032228445634245872, + 0.007922424003481865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.19361652433872223 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00597832165658474 + ], + "opposite_logprob_tokens": [ + -5.130978107452393 + ], + "logprob_reward": -0.00597832165658474, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00597832165658474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.137116938829422, + 0.09540744125843048, + 0.07353518158197403, + 0.061207205057144165, + 0.053470343351364136, + 0.059292983263731, + 0.07723943889141083, + 0.3620467185974121, + 0.02338484115898609, + 0.05729891359806061 + ], + "credit_map": [ + 0.01579817570745945, + 0.025307737290859222, + 0.01532595045864582, + 0.012679290026426315, + 0.020951537415385246, + 0.012373006902635098, + 0.01595352403819561, + 0.018727721646428108, + 0.00793539173901081, + 0.017345456406474113, + 0.006807784084230661, + 0.010802743025124073, + 0.014698824845254421, + 0.01594894379377365, + 0.007494363468140364, + 0.014373932033777237, + 0.009368221275508404, + 0.016799187287688255, + 0.006163537502288818, + 0.005810914095491171, + 0.009741599671542645, + 0.00785022508352995, + 0.006186456419527531, + 0.011615035124123096, + 0.007248662877827883, + 0.01450182031840086, + 0.005254894029349089, + 0.007547910790890455, + 0.0073885442689061165, + 0.005888484418392181, + 0.005425657611340284, + 0.007951234467327595, + 0.005469708237797022, + 0.007157562300562859, + 0.004550195299088955, + 0.004931271076202393, + 0.009542501531541348, + 0.009208555333316326, + 0.005591563414782286, + 0.0070189847610890865, + 0.007112566847354174, + 0.00977662019431591, + 0.0057458835653960705, + 0.004305175039917231, + 0.00731204729527235, + 0.0098350141197443, + 0.0056628212332725525, + 0.009542858228087425, + 0.010807623155415058, + 0.01336764357984066, + 0.0064420211128890514, + 0.005887770559638739, + 0.012181377969682217, + 0.007578317075967789, + 0.009139631874859333, + 0.011835053563117981, + 0.05961006134748459, + 0.04332488402724266, + 0.028734587132930756, + 0.01703096739947796, + 0.02294583059847355, + 0.06444771587848663, + 0.01996746100485325, + 0.10598523169755936, + 0.0027566717471927404, + 0.004617892671376467, + 0.0017172981752082705, + 0.0025985552929341793, + 0.003751135431230068, + 0.0026842067018151283, + 0.0026488900184631348, + 0.002610190771520138, + 0.00429173419252038, + 0.002788946032524109, + 0.0021232564467936754, + 0.0024094453547149897, + 0.004791037179529667, + 0.017960915341973305, + 0.002778172492980957, + 0.0201554074883461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.00597832165658474 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036216876469552517 + ], + "opposite_logprob_tokens": [ + -5.628621578216553 + ], + "logprob_reward": -0.0036216876469552517, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036216876469552517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2367880493402481, + 0.09893552213907242, + 0.11633983999490738, + 0.12439551204442978, + 0.12813115119934082, + 0.10548354685306549, + 0.06729849427938461, + 0.050908397883176804, + 0.04641798138618469, + 0.02530151791870594 + ], + "credit_map": [ + 0.022952642291784286, + 0.0339127853512764, + 0.01740885153412819, + 0.017982615157961845, + 0.03791156783699989, + 0.029497316107153893, + 0.028343306854367256, + 0.0487789586186409, + 0.00908285565674305, + 0.01146054919809103, + 0.009044842794537544, + 0.008352205157279968, + 0.01578104868531227, + 0.01579166390001774, + 0.013148331083357334, + 0.016274025663733482, + 0.0120516587048769, + 0.015081798657774925, + 0.011867891065776348, + 0.009493262507021427, + 0.015788039192557335, + 0.021255651488900185, + 0.012981690466403961, + 0.017819857224822044, + 0.013807639479637146, + 0.01797289028763771, + 0.011935518123209476, + 0.008665816858410835, + 0.01982273906469345, + 0.021570591256022453, + 0.011430585756897926, + 0.01918973959982395, + 0.017251772806048393, + 0.02279655635356903, + 0.011023793369531631, + 0.010356495156884193, + 0.019403714686632156, + 0.018676050007343292, + 0.01184405479580164, + 0.01677870750427246, + 0.012201216071844101, + 0.017199136316776276, + 0.007749342825263739, + 0.010228261351585388, + 0.014404595829546452, + 0.01139684859663248, + 0.01564754918217659, + 0.01665659435093403, + 0.010141005739569664, + 0.013169384561479092, + 0.005424043163657188, + 0.006254192441701889, + 0.008280660025775433, + 0.007589076645672321, + 0.007487007882446051, + 0.008953118696808815, + 0.008924533613026142, + 0.005419314838945866, + 0.006401596590876579, + 0.004329124465584755, + 0.00506212143227458, + 0.006607157178223133, + 0.0053383051417768, + 0.0088262427598238, + 0.00386588042601943, + 0.0057486132718622684, + 0.0026962237898260355, + 0.002856129314750433, + 0.009628371335566044, + 0.006190368440002203, + 0.007299019023776054, + 0.008133377879858017, + 0.002635982818901539, + 0.003818019526079297, + 0.0018449699273332953, + 0.0018799270037561655, + 0.005886028055101633, + 0.0030872581992298365, + 0.0031927695963531733, + 0.002956564072519541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0036216876469552517 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013375869020819664 + ], + "opposite_logprob_tokens": [ + -6.62633752822876 + ], + "logprob_reward": -0.0013375869020819664, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013375869020819664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20084168016910553, + 0.05368760600686073, + 0.12098754197359085, + 0.16903120279312134, + 0.12070770561695099, + 0.12471073865890503, + 0.06171051412820816, + 0.05755774304270744, + 0.06202264502644539, + 0.02874263934791088 + ], + "credit_map": [ + 0.02017461508512497, + 0.028611624613404274, + 0.014078913256525993, + 0.014832560904324055, + 0.030868669971823692, + 0.025963084772229195, + 0.02455070987343788, + 0.04176151007413864, + 0.004959487821906805, + 0.0076527842320501804, + 0.004293940495699644, + 0.0048013711348176, + 0.011748421005904675, + 0.007203676272183657, + 0.006106652319431305, + 0.006921275518834591, + 0.011469081044197083, + 0.017407994717359543, + 0.011496119201183319, + 0.01062461081892252, + 0.016167841851711273, + 0.023367909714579582, + 0.012361827306449413, + 0.018092161044478416, + 0.017000164836645126, + 0.018638573586940765, + 0.015896160155534744, + 0.012584595009684563, + 0.02955346368253231, + 0.02997954934835434, + 0.017540594562888145, + 0.027838094159960747, + 0.01616775617003441, + 0.018635477870702744, + 0.011516417376697063, + 0.010400871746242046, + 0.018306005746126175, + 0.01899516023695469, + 0.012052983976900578, + 0.01463303342461586, + 0.01364738866686821, + 0.018585827201604843, + 0.008833778090775013, + 0.011217912659049034, + 0.0207337848842144, + 0.013824550434947014, + 0.022159000858664513, + 0.015708502382040024, + 0.0076375436037778854, + 0.009537318721413612, + 0.006831453647464514, + 0.006705126259475946, + 0.007553535513579845, + 0.006990203633904457, + 0.00858725793659687, + 0.007868080399930477, + 0.007567840628325939, + 0.008679941296577454, + 0.00796810258179903, + 0.0049178823828697205, + 0.008374915458261967, + 0.006195519119501114, + 0.006588187534362078, + 0.007265354972332716, + 0.005043117795139551, + 0.007992833852767944, + 0.003691325429826975, + 0.004037872422486544, + 0.011863283812999725, + 0.009330622851848602, + 0.008796023204922676, + 0.011267568916082382, + 0.0024723210372030735, + 0.0022901594638824463, + 0.0010346387280151248, + 0.001128883333876729, + 0.0020435217302292585, + 0.008478843607008457, + 0.0011647194623947144, + 0.010129551403224468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0013375869020819664 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016059687361121178 + ], + "opposite_logprob_tokens": [ + -4.141059875488281 + ], + "logprob_reward": -0.016059687361121178, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016059687361121178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06483618170022964, + 0.06612492352724075, + 0.0960877537727356, + 0.043063607066869736, + 0.3587561249732971, + 0.1641387641429901, + 0.036466293036937714, + 0.02706235460937023, + 0.06168169900774956, + 0.08178237080574036 + ], + "credit_map": [ + 0.0058874753303825855, + 0.012944159097969532, + 0.005589094944298267, + 0.005333392880856991, + 0.008154514245688915, + 0.007953522726893425, + 0.005968296434730291, + 0.01300572045147419, + 0.006485521327704191, + 0.008675520308315754, + 0.0053751361556351185, + 0.007933284156024456, + 0.011842617765069008, + 0.011125944554805756, + 0.0049413410015404224, + 0.009745554067194462, + 0.011237171478569508, + 0.017693424597382545, + 0.006960141938179731, + 0.00933898240327835, + 0.01401621475815773, + 0.015963740646839142, + 0.007805746514350176, + 0.013072315603494644, + 0.004462952259927988, + 0.0077111306600272655, + 0.00508185476064682, + 0.004176753107458353, + 0.00737501448020339, + 0.0034342051949352026, + 0.0054175457917153835, + 0.005404144991189241, + 0.04922692850232124, + 0.06355538219213486, + 0.02639087103307247, + 0.025038747116923332, + 0.051002807915210724, + 0.05795145779848099, + 0.032085176557302475, + 0.05350472405552864, + 0.01891859620809555, + 0.026517348363995552, + 0.011580937542021275, + 0.009470634162425995, + 0.022616155445575714, + 0.031427573412656784, + 0.014651546254754066, + 0.028955945745110512, + 0.004578334745019674, + 0.004814863204956055, + 0.004558291751891375, + 0.004393953364342451, + 0.005371399223804474, + 0.004061734303832054, + 0.004403241910040379, + 0.004284471273422241, + 0.004485059529542923, + 0.003258122829720378, + 0.0025928940158337355, + 0.0027403216809034348, + 0.002885078312829137, + 0.0029664316680282354, + 0.0034524037037044764, + 0.004682040773332119, + 0.005642910487949848, + 0.011534510180354118, + 0.0037210467271506786, + 0.005074575077742338, + 0.010959425941109657, + 0.007875115610659122, + 0.008261021226644516, + 0.00861309003084898, + 0.0063841999508440495, + 0.004809079226106405, + 0.0035007682163268328, + 0.0031790435314178467, + 0.00878758355975151, + 0.023041505366563797, + 0.0043874685652554035, + 0.027692709118127823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.016059687361121178 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, the camera remains still, focused on a newly emerged butterfly clinging to a emerald green stem, next to the translucent, pale green shell it had escaped from. A warm, soft morning light streams in from behind, casting a captivating golden rim around the butterfly's furry black body and the edges of its wings, which have yet to fully unfold. Its wrinkled, damp wings appear translucent against the soft-focus background, their intricate veins of orange and black barely visible. The wings tremble slightly, almost imperceptibly, as if gathering strength. Then, steadily and deliberately, they unfold, bit by bit. As the fluids slowly flow in, the wrinkles gradually smooth out, and the pattern on the wings becomes clearer and more vivid. The entire picture is shrouded in a tranquil and hopeful warm-toned atmosphere. The green leaves dotted with morning dew in the background form dreamy light spots, together forming a poetic picture of rebirth and transformation.", + "seed": 21608, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_2/index_2388_seed_21608_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.032200551871210334, + "vlm_logprob": -0.032200551871210334, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.032200551871210334, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11580602079629898 + ], + "opposite_logprob_tokens": [ + -2.6158061027526855 + ], + "logprob_reward": -0.11580602079629898, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.11580602079629898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1321822702884674, + 0.1488094925880432, + 0.056952957063913345, + 0.2739258110523224, + 0.057538267225027084, + 0.036490634083747864, + 0.0575302392244339, + 0.10164067894220352, + 0.10765571892261505, + 0.027273956686258316 + ], + "credit_map": [ + 0.023698754608631134, + 0.01664987951517105, + 0.00785668846219778, + 0.009428995661437511, + 0.008956708945333958, + 0.0392812117934227, + 0.008686457760632038, + 0.017623554915189743, + 0.018785981461405754, + 0.017899759113788605, + 0.010433482006192207, + 0.011813375167548656, + 0.02537512220442295, + 0.03449508547782898, + 0.013305729255080223, + 0.01670093648135662, + 0.00714460713788867, + 0.006012459751218557, + 0.0065881372429430485, + 0.0067015658132731915, + 0.007928293198347092, + 0.007494454272091389, + 0.008438407443463802, + 0.006645024288445711, + 0.030149729922413826, + 0.02773343026638031, + 0.02718925289809704, + 0.027164630591869354, + 0.033271174877882004, + 0.03357302397489548, + 0.047010671347379684, + 0.04783385246992111, + 0.006338552106171846, + 0.004819763358682394, + 0.005003920290619135, + 0.005328543484210968, + 0.007985347881913185, + 0.008738760836422443, + 0.008086666464805603, + 0.011236708611249924, + 0.0045717679895460606, + 0.0037206776905804873, + 0.002704052487388253, + 0.003572429995983839, + 0.007733648177236319, + 0.0053653777576982975, + 0.0032659978605806828, + 0.005556680262088776, + 0.007896958850324154, + 0.006365362089127302, + 0.00481094466522336, + 0.004203067626804113, + 0.013361676596105099, + 0.007704258430749178, + 0.0055307853035628796, + 0.007657179143279791, + 0.011678140610456467, + 0.010395718738436699, + 0.010547910816967487, + 0.01063323300331831, + 0.015559718012809753, + 0.015045750886201859, + 0.013868807815015316, + 0.013911393471062183, + 0.013296256773173809, + 0.01148122176527977, + 0.009566539898514748, + 0.009790427051484585, + 0.021747076883912086, + 0.01689167320728302, + 0.007146738935261965, + 0.01773577183485031, + 0.0035933025646954775, + 0.002497419947758317, + 0.0028760593850165606, + 0.002795221284031868, + 0.004594340920448303, + 0.0036173865664750338, + 0.003170775016769767, + 0.004129447508603334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.11580602079629898 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005281305871903896 + ], + "opposite_logprob_tokens": [ + -5.255281448364258 + ], + "logprob_reward": -0.005281305871903896, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005281305871903896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5119971632957458, + 0.09005360305309296, + 0.0867258831858635, + 0.03792745620012283, + 0.02169046550989151, + 0.03804687783122063, + 0.11404117941856384, + 0.06545251607894897, + 0.02130802907049656, + 0.012756812386214733 + ], + "credit_map": [ + 0.0722181424498558, + 0.09222178161144257, + 0.045593347400426865, + 0.02544696070253849, + 0.020834779366850853, + 0.19299735128879547, + 0.01838284358382225, + 0.04430193081498146, + 0.00840034894645214, + 0.012118693441152573, + 0.012525374069809914, + 0.00878388062119484, + 0.01641722209751606, + 0.011571467854082584, + 0.008481092751026154, + 0.011755521409213543, + 0.010221225209534168, + 0.010103312321007252, + 0.012559614144265652, + 0.008412779308855534, + 0.00970112718641758, + 0.012957872822880745, + 0.008933854289352894, + 0.013836098834872246, + 0.0038958063814789057, + 0.004262053407728672, + 0.003996153362095356, + 0.005150345154106617, + 0.004622164648026228, + 0.004886855371296406, + 0.0054865386337041855, + 0.005627536680549383, + 0.0032457176130264997, + 0.0021956698037683964, + 0.0019411436514928937, + 0.002013847464695573, + 0.0032065855339169502, + 0.003259865799918771, + 0.0021025759633630514, + 0.003725058864802122, + 0.0033673220314085484, + 0.004578329157084227, + 0.0032204766757786274, + 0.00390186021104455, + 0.007598865777254105, + 0.005559132434427738, + 0.004787410143762827, + 0.005033480003476143, + 0.01170131005346775, + 0.011143447831273079, + 0.009052862413227558, + 0.008235316723585129, + 0.028148695826530457, + 0.016274798661470413, + 0.011911533772945404, + 0.01757321134209633, + 0.006328949239104986, + 0.0049894521944224834, + 0.007397144101560116, + 0.005207687150686979, + 0.006319551263004541, + 0.009654631838202477, + 0.01193191111087799, + 0.01362319104373455, + 0.002665624488145113, + 0.0019967376720160246, + 0.0022383376490324736, + 0.002031539101153612, + 0.0038706909399479628, + 0.0034338333643972874, + 0.0019077215110883117, + 0.0031635467894375324, + 0.00152258959133178, + 0.0015319210942834616, + 0.0014038658700883389, + 0.0009818051476031542, + 0.0012113908305764198, + 0.0019249506294727325, + 0.0017447139834985137, + 0.0024355752393603325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.005281305871903896 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545241974294186 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012545241974294186, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545241974294186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2313774824142456, + 0.14233502745628357, + 0.055187564343214035, + 0.1457696408033371, + 0.09242613613605499, + 0.10300874710083008, + 0.11679707467556, + 0.04368441924452782, + 0.04363647848367691, + 0.025777464732527733 + ], + "credit_map": [ + 0.04242299497127533, + 0.023397235199809074, + 0.01822095550596714, + 0.017467301338911057, + 0.018090378493070602, + 0.06854993104934692, + 0.015454908832907677, + 0.027773773297667503, + 0.011537865735590458, + 0.016701197251677513, + 0.010950751602649689, + 0.01650729961693287, + 0.02844170480966568, + 0.02164124697446823, + 0.01565239019691944, + 0.020902570337057114, + 0.005969085730612278, + 0.0060798535123467445, + 0.011122636497020721, + 0.0065466309897601604, + 0.00612479355186224, + 0.009222172200679779, + 0.00395155930891633, + 0.006170828826725483, + 0.01859092339873314, + 0.010745320469141006, + 0.015523361973464489, + 0.009593887254595757, + 0.013561507686972618, + 0.021258821710944176, + 0.017655186355113983, + 0.03884062543511391, + 0.009712021797895432, + 0.00948462262749672, + 0.00988833513110876, + 0.011966019868850708, + 0.012315780855715275, + 0.012709340080618858, + 0.012001329101622105, + 0.014348688535392284, + 0.014298158697783947, + 0.0076157571747899055, + 0.008414436131715775, + 0.007075591012835503, + 0.011671938002109528, + 0.016810093075037003, + 0.0065986281260848045, + 0.03052414394915104, + 0.013552551157772541, + 0.01141856424510479, + 0.007242313586175442, + 0.008131948299705982, + 0.022838681936264038, + 0.019352329894900322, + 0.007432674057781696, + 0.026828009635210037, + 0.004882717505097389, + 0.004802096169441938, + 0.0037615224719047546, + 0.003407079027965665, + 0.008093207143247128, + 0.006072418298572302, + 0.004841194488108158, + 0.007824181579053402, + 0.006638210732489824, + 0.004740044008940458, + 0.002831771271303296, + 0.0033319464419037104, + 0.006114387419074774, + 0.007576189003884792, + 0.002880415413528681, + 0.009523516520857811, + 0.003060788381844759, + 0.0037292102351784706, + 0.003061094554141164, + 0.0023264724295586348, + 0.003216438926756382, + 0.00474554905667901, + 0.0014605856267735362, + 0.004177324939519167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.012545241974294186 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095975775271654 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004095975775271654, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095975775271654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21903076767921448, + 0.2000550776720047, + 0.06987965852022171, + 0.15666590631008148, + 0.0603938028216362, + 0.07684638351202011, + 0.09932591021060944, + 0.03716351091861725, + 0.05005247890949249, + 0.03058650530874729 + ], + "credit_map": [ + 0.032991815358400345, + 0.02199581079185009, + 0.016441339626908302, + 0.02598864585161209, + 0.021200723946094513, + 0.05264858156442642, + 0.018997665494680405, + 0.028766188770532608, + 0.019451197236776352, + 0.02372954972088337, + 0.015561827458441257, + 0.01868675835430622, + 0.04706097021698952, + 0.033312417566776276, + 0.024403566494584084, + 0.017848795279860497, + 0.007612451910972595, + 0.00946163572371006, + 0.011225190944969654, + 0.00862722285091877, + 0.00798877514898777, + 0.010197657160460949, + 0.005404147785156965, + 0.009362580254673958, + 0.019556041806936264, + 0.01240267138928175, + 0.015470702201128006, + 0.011840423569083214, + 0.01633528433740139, + 0.021455980837345123, + 0.02285168133676052, + 0.03675312548875809, + 0.007293189875781536, + 0.006776018068194389, + 0.005952621344476938, + 0.006514281500130892, + 0.009034546092152596, + 0.008496507070958614, + 0.005250104237347841, + 0.011076534166932106, + 0.00804529245942831, + 0.008722268976271152, + 0.006413913797587156, + 0.007501628715544939, + 0.01754811219871044, + 0.009262202307581902, + 0.00741015188395977, + 0.011942815035581589, + 0.010418645106256008, + 0.008648074232041836, + 0.007263246923685074, + 0.0072927894070744514, + 0.023865193128585815, + 0.015127243474125862, + 0.008563995361328125, + 0.018146729096770287, + 0.0050241658464074135, + 0.0037494690623134375, + 0.002865815768018365, + 0.0027280887588858604, + 0.006866373121738434, + 0.00574155431240797, + 0.00292233400978148, + 0.00726570887491107, + 0.007636392954736948, + 0.0047325375489890575, + 0.0057287090457975864, + 0.004984447732567787, + 0.006447548512369394, + 0.005425645969808102, + 0.007614077068865299, + 0.0074831219390034676, + 0.003927207551896572, + 0.003719582222402096, + 0.0028126086108386517, + 0.0026423807721585035, + 0.00416435906663537, + 0.005256254225969315, + 0.003026482416316867, + 0.005037631839513779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.004095975775271654 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023274214938282967 + ], + "opposite_logprob_tokens": [ + -3.7732741832733154 + ], + "logprob_reward": -0.023274214938282967, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023274214938282967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3320721685886383, + 0.11832736432552338, + 0.06273997575044632, + 0.058296408504247665, + 0.042620930820703506, + 0.05108500272035599, + 0.05348914861679077, + 0.12717215716838837, + 0.11929560452699661, + 0.0349012054502964 + ], + "credit_map": [ + 0.04279187694191933, + 0.053841471672058105, + 0.018244324252009392, + 0.016751982271671295, + 0.027402324602007866, + 0.11134765297174454, + 0.022952638566493988, + 0.03873991593718529, + 0.0157240629196167, + 0.012229877524077892, + 0.021454796195030212, + 0.008427645079791546, + 0.012244325131177902, + 0.021835332736372948, + 0.009492709301412106, + 0.016918610781431198, + 0.009660301730036736, + 0.005887400358915329, + 0.008471609093248844, + 0.004429982975125313, + 0.006014371756464243, + 0.009862798266112804, + 0.0040176440961658955, + 0.01439587026834488, + 0.006076600402593613, + 0.005616934038698673, + 0.005926302168518305, + 0.0056627849116921425, + 0.006525538396090269, + 0.008388683199882507, + 0.010910013690590858, + 0.009189550764858723, + 0.0039244224317371845, + 0.003685305593535304, + 0.005657725967466831, + 0.005425715819001198, + 0.005482935346662998, + 0.005171176977455616, + 0.005415966268628836, + 0.00785768311470747, + 0.004415125120431185, + 0.004327478352934122, + 0.008430378511548042, + 0.004240960814058781, + 0.009441139176487923, + 0.007039258256554604, + 0.004739740863442421, + 0.008450919762253761, + 0.006322432775050402, + 0.0050826529040932655, + 0.005079891066998243, + 0.0038264202885329723, + 0.006622595246881247, + 0.008412606082856655, + 0.004807238932698965, + 0.013335316441953182, + 0.012729796580970287, + 0.012246590107679367, + 0.011435762979090214, + 0.012455927208065987, + 0.022660253569483757, + 0.022160720080137253, + 0.01898457296192646, + 0.01449853926897049, + 0.013974989764392376, + 0.010680482722818851, + 0.008568494580686092, + 0.009251374751329422, + 0.021395178511738777, + 0.019738484174013138, + 0.008002196438610554, + 0.0276844073086977, + 0.004455845337361097, + 0.004073943942785263, + 0.003164357505738735, + 0.002703390084207058, + 0.003523588413372636, + 0.0065641626715660095, + 0.003434952814131975, + 0.006980963051319122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.023274214938282967 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, the camera remains still, focused on a newly emerged butterfly clinging to a emerald green stem, next to the translucent, pale green shell it had escaped from. A warm, soft morning light streams in from behind, casting a captivating golden rim around the butterfly's furry black body and the edges of its wings, which have yet to fully unfold. Its wrinkled, damp wings appear translucent against the soft-focus background, their intricate veins of orange and black barely visible. The wings tremble slightly, almost imperceptibly, as if gathering strength. Then, steadily and deliberately, they unfold, bit by bit. As the fluids slowly flow in, the wrinkles gradually smooth out, and the pattern on the wings becomes clearer and more vivid. The entire picture is shrouded in a tranquil and hopeful warm-toned atmosphere. The green leaves dotted with morning dew in the background form dreamy light spots, together forming a poetic picture of rebirth and transformation.", + "seed": 21609, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_2/index_2388_seed_21609_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.015653778100386262, + "vlm_logprob": -0.015653778100386262, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.015653778100386262, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.056105565279722214 + ], + "opposite_logprob_tokens": [ + -4.056105613708496 + ], + "logprob_reward": -0.056105565279722214, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.056105565279722214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055946942418813705, + 0.03775671124458313, + 0.11859017610549927, + 0.04618562012910843, + 0.1389777958393097, + 0.0322992205619812, + 0.18619132041931152, + 0.07738931477069855, + 0.07733003795146942, + 0.22933292388916016 + ], + "credit_map": [ + 0.004328212235122919, + 0.006787991151213646, + 0.005398325622081757, + 0.009497236460447311, + 0.007568659260869026, + 0.007200845051556826, + 0.007574961055070162, + 0.007590704597532749, + 0.0052766758017241955, + 0.0038423745427280664, + 0.004433768801391125, + 0.005431666504591703, + 0.005689946468919516, + 0.004052929580211639, + 0.005190274678170681, + 0.003839068580418825, + 0.01026768609881401, + 0.01793903112411499, + 0.008464256301522255, + 0.01226860098540783, + 0.020949510857462883, + 0.021714048460125923, + 0.014173988252878189, + 0.01281304657459259, + 0.003611633786931634, + 0.007056057918816805, + 0.005223884712904692, + 0.0041806260123848915, + 0.006978586316108704, + 0.009012140333652496, + 0.005029842723160982, + 0.005092840176075697, + 0.010521596297621727, + 0.034661538898944855, + 0.00983783695846796, + 0.008907542563974857, + 0.015773294493556023, + 0.024335982277989388, + 0.012189074419438839, + 0.022750917822122574, + 0.0036529721692204475, + 0.0037900677416473627, + 0.0029443746898323298, + 0.0030052638612687588, + 0.007041638717055321, + 0.004188525024801493, + 0.0036158945877104998, + 0.0040604835376143456, + 0.010510146617889404, + 0.062436264008283615, + 0.009224114939570427, + 0.0070649487897753716, + 0.012755150906741619, + 0.018060240894556046, + 0.006997151765972376, + 0.059143271297216415, + 0.008069055154919624, + 0.008871275931596756, + 0.006968156900256872, + 0.0069150496274232864, + 0.013202978298068047, + 0.0125004006549716, + 0.007016510237008333, + 0.01384588610380888, + 0.005623165518045425, + 0.024984903633594513, + 0.00478424783796072, + 0.004638398066163063, + 0.003653695108368993, + 0.005226231180131435, + 0.003122573485597968, + 0.025296814739704132, + 0.02994486503303051, + 0.05746376886963844, + 0.028401797637343407, + 0.013392343185842037, + 0.026302676647901535, + 0.02932354062795639, + 0.014671099372208118, + 0.02983279526233673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "selected_reward": -0.056105565279722214 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006764133460819721 + ], + "opposite_logprob_tokens": [ + -5.006763935089111 + ], + "logprob_reward": -0.006764133460819721, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006764133460819721, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058485303074121475, + 0.034260980784893036, + 0.06856752932071686, + 0.08774250000715256, + 0.030417228117585182, + 0.22968368232250214, + 0.22869153320789337, + 0.041095685213804245, + 0.12979687750339508, + 0.09125874936580658 + ], + "credit_map": [ + 0.006709225010126829, + 0.0072117461822927, + 0.005536742974072695, + 0.009492161683738232, + 0.011135885491967201, + 0.005039438605308533, + 0.008455018512904644, + 0.0049050780944526196, + 0.003498570527881384, + 0.00544239766895771, + 0.003794493619352579, + 0.006247552111744881, + 0.005133442580699921, + 0.002966411178931594, + 0.0044231838546693325, + 0.0027549259830266237, + 0.005389547441154718, + 0.008956126868724823, + 0.005537833087146282, + 0.006054218392819166, + 0.013174773193895817, + 0.0170900896191597, + 0.006828076671808958, + 0.00553686311468482, + 0.008198359981179237, + 0.012095604091882706, + 0.009022860787808895, + 0.008005974814295769, + 0.014655642211437225, + 0.018983041867613792, + 0.008374175988137722, + 0.00840682815760374, + 0.003461254760622978, + 0.0076891761273145676, + 0.0028020411264151335, + 0.0020492251496762037, + 0.003314022673293948, + 0.004416438285261393, + 0.0020789026748389006, + 0.004606163129210472, + 0.017785098403692245, + 0.05024771764874458, + 0.01317781861871481, + 0.011096509173512459, + 0.04520820826292038, + 0.018315771594643593, + 0.021283350884914398, + 0.05256916955113411, + 0.011986943893134594, + 0.07896140217781067, + 0.009537291713058949, + 0.0076691824942827225, + 0.02096857875585556, + 0.01234005019068718, + 0.011018064804375172, + 0.07620999962091446, + 0.0038856088649481535, + 0.00534401647746563, + 0.0033543144818395376, + 0.003223972860723734, + 0.00900237075984478, + 0.006522119045257568, + 0.005032520741224289, + 0.004730754531919956, + 0.008396586403250694, + 0.04802865907549858, + 0.005575197748839855, + 0.0049291132017970085, + 0.006090256385505199, + 0.006270616315305233, + 0.004782458767294884, + 0.04572395980358124, + 0.0095845190808177, + 0.018657125532627106, + 0.010439789853990078, + 0.013796682469546795, + 0.01017883699387312, + 0.008074404671788216, + 0.008693050593137741, + 0.011834328062832355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.006764133460819721 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595355499535799 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.00595355499535799, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595355499535799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10363828390836716, + 0.05553685128688812, + 0.05636884644627571, + 0.03553737699985504, + 0.1295776665210724, + 0.07319754362106323, + 0.09220663458108902, + 0.06694222241640091, + 0.0410531684756279, + 0.34594136476516724 + ], + "credit_map": [ + 0.011014699004590511, + 0.011936736293137074, + 0.011015641503036022, + 0.018680088222026825, + 0.015737030655145645, + 0.014747444540262222, + 0.010805017314851284, + 0.009701631031930447, + 0.0039529381319880486, + 0.007824028842151165, + 0.004262272268533707, + 0.006590451579540968, + 0.011843327432870865, + 0.010207911022007465, + 0.006277666427195072, + 0.00457826117053628, + 0.005743431858718395, + 0.007519392296671867, + 0.005164764355868101, + 0.00467925937846303, + 0.00933014415204525, + 0.013453440740704536, + 0.005055228713899851, + 0.005423189606517553, + 0.003465140936896205, + 0.005485887173563242, + 0.004792264197021723, + 0.003247785149142146, + 0.006648415233939886, + 0.003879913594573736, + 0.004279524087905884, + 0.0037384526804089546, + 0.009238081984221935, + 0.03881915658712387, + 0.006607194431126118, + 0.007290943525731564, + 0.012284435331821442, + 0.01852254569530487, + 0.008515101857483387, + 0.028300216421484947, + 0.0063146743923425674, + 0.015482628718018532, + 0.005206506233662367, + 0.0035095219500362873, + 0.009689909406006336, + 0.009456311352550983, + 0.006207619793713093, + 0.017330385744571686, + 0.006660060025751591, + 0.02405383065342903, + 0.004346775356680155, + 0.004205264616757631, + 0.013053101487457752, + 0.014410018920898438, + 0.006006898358464241, + 0.019470691680908203, + 0.0078117200173437595, + 0.008680202066898346, + 0.0056588780134916306, + 0.00580710219219327, + 0.015511862933635712, + 0.006791339721530676, + 0.010204795747995377, + 0.0064763291738927364, + 0.0025756810791790485, + 0.012301022186875343, + 0.0025311370845884085, + 0.0021708449348807335, + 0.0034690722823143005, + 0.0028304236475378275, + 0.002745565725490451, + 0.012429427355527878, + 0.03865917772054672, + 0.081730917096138, + 0.04372328147292137, + 0.020352357998490334, + 0.04639167711138725, + 0.042677707970142365, + 0.02281768061220646, + 0.04958861321210861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.00595355499535799 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008127961773425341 + ], + "opposite_logprob_tokens": [ + -7.125813007354736 + ], + "logprob_reward": -0.0008127961773425341, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008127961773425341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03235124424099922, + 0.036040838807821274, + 0.05121098458766937, + 0.028328942134976387, + 0.034218236804008484, + 0.041395530104637146, + 0.17164555191993713, + 0.027627207338809967, + 0.02880607172846794, + 0.5483754277229309 + ], + "credit_map": [ + 0.0033727092668414116, + 0.005103668197989464, + 0.002529926598072052, + 0.004291659686714411, + 0.0032616162206977606, + 0.007239209953695536, + 0.0028667845763266087, + 0.003685668110847473, + 0.0027049737982451916, + 0.004219444002956152, + 0.005399941001087427, + 0.004341890569776297, + 0.003634419059380889, + 0.0050009433180093765, + 0.005359846632927656, + 0.0053793760016560555, + 0.004353738389909267, + 0.006616482976824045, + 0.004261970520019531, + 0.0050316439010202885, + 0.008432401344180107, + 0.01300620473921299, + 0.004747413098812103, + 0.004761124961078167, + 0.003164321882650256, + 0.004016422200948, + 0.0031029481906443834, + 0.001755016390234232, + 0.0043738349340856075, + 0.00469902902841568, + 0.003948934376239777, + 0.0032684325706213713, + 0.002371944021433592, + 0.008023060858249664, + 0.0025268010795116425, + 0.0025958644691854715, + 0.0038821317721158266, + 0.005319834221154451, + 0.0029313869308680296, + 0.006567211356014013, + 0.003918029833585024, + 0.00948964525014162, + 0.002432250650599599, + 0.0024157576262950897, + 0.007333762012422085, + 0.004178013652563095, + 0.0038275287952274084, + 0.007800538558512926, + 0.009728183038532734, + 0.06090841069817543, + 0.006060629617422819, + 0.004701136611402035, + 0.01652630604803562, + 0.009053541347384453, + 0.007613958325237036, + 0.05705336853861809, + 0.003586376318708062, + 0.003891803789883852, + 0.0031304771546274424, + 0.0025847642682492733, + 0.00490305433049798, + 0.002467252081260085, + 0.0040243444964289665, + 0.0030391330365091562, + 0.002097010612487793, + 0.004432184621691704, + 0.0027920035645365715, + 0.0026365944650024176, + 0.003606460290029645, + 0.0061192926950752735, + 0.002211181214079261, + 0.004911341238766909, + 0.06251583993434906, + 0.13616177439689636, + 0.06556792557239532, + 0.03252102807164192, + 0.07125415652990341, + 0.06919369846582413, + 0.035082705318927765, + 0.07607825845479965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.0008127961773425341 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863284058868885 + ], + "opposite_logprob_tokens": [ + -4.758632659912109 + ], + "logprob_reward": -0.00863284058868885, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863284058868885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15564891695976257, + 0.06641827523708344, + 0.08938462287187576, + 0.03399897366762161, + 0.07180356234312057, + 0.1266634315252304, + 0.18564379215240479, + 0.031529154628515244, + 0.0458793118596077, + 0.1930299699306488 + ], + "credit_map": [ + 0.016907861456274986, + 0.02172272838652134, + 0.014528941363096237, + 0.014895740896463394, + 0.015788856893777847, + 0.04429372400045395, + 0.010402862913906574, + 0.01710820570588112, + 0.005134520586580038, + 0.00880715623497963, + 0.007732515223324299, + 0.006883392576128244, + 0.006017525214701891, + 0.01359010674059391, + 0.007906786166131496, + 0.010346271097660065, + 0.00765626598149538, + 0.012455442920327187, + 0.007833525538444519, + 0.01017152238637209, + 0.013698509894311428, + 0.018977388739585876, + 0.008754965849220753, + 0.009837007150053978, + 0.002420929493382573, + 0.004771583713591099, + 0.0037006374914199114, + 0.0029921026434749365, + 0.006213868502527475, + 0.007656059227883816, + 0.0032751900143921375, + 0.002968605374917388, + 0.0071067591197788715, + 0.02491738460958004, + 0.004132630303502083, + 0.0029719131998717785, + 0.005725424736738205, + 0.005634265020489693, + 0.004645680543035269, + 0.016669506207108498, + 0.007934429682791233, + 0.03213531896471977, + 0.0068305907770991325, + 0.0065404209308326244, + 0.01542754378169775, + 0.01507008820772171, + 0.008921674452722073, + 0.033803366124629974, + 0.010614152997732162, + 0.06993268430233002, + 0.006854836829006672, + 0.005273364018648863, + 0.013527865521609783, + 0.006914641708135605, + 0.006779992952942848, + 0.0657462552189827, + 0.0038857366889715195, + 0.0039034648798406124, + 0.0029875561594963074, + 0.003869165200740099, + 0.00535160256549716, + 0.004131385590881109, + 0.0033431509509682655, + 0.004057092592120171, + 0.00292376521974802, + 0.015194633044302464, + 0.002535561565309763, + 0.0026386985555291176, + 0.002687571570277214, + 0.0022512865252792835, + 0.0024548459332436323, + 0.015192953869700432, + 0.023652706295251846, + 0.0475451722741127, + 0.02266436442732811, + 0.011922628618776798, + 0.023173369467258453, + 0.027572238817811012, + 0.012640558183193207, + 0.02385893277823925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.00863284058868885 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, the camera remains still, focused on a newly emerged butterfly clinging to a emerald green stem, next to the translucent, pale green shell it had escaped from. A warm, soft morning light streams in from behind, casting a captivating golden rim around the butterfly's furry black body and the edges of its wings, which have yet to fully unfold. Its wrinkled, damp wings appear translucent against the soft-focus background, their intricate veins of orange and black barely visible. The wings tremble slightly, almost imperceptibly, as if gathering strength. Then, steadily and deliberately, they unfold, bit by bit. As the fluids slowly flow in, the wrinkles gradually smooth out, and the pattern on the wings becomes clearer and more vivid. The entire picture is shrouded in a tranquil and hopeful warm-toned atmosphere. The green leaves dotted with morning dew in the background form dreamy light spots, together forming a poetic picture of rebirth and transformation.", + "seed": 21610, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_2/index_2388_seed_21610_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03432965613901615, + "vlm_logprob": -0.03432965613901615, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.03432965613901615, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09701820462942123 + ], + "opposite_logprob_tokens": [ + -2.972018241882324 + ], + "logprob_reward": -0.09701820462942123, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.09701820462942123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0562133863568306, + 0.022670594975352287, + 0.026672346517443657, + 0.12094496935606003, + 0.05307946726679802, + 0.02407555654644966, + 0.040135618299245834, + 0.25258275866508484, + 0.37064987421035767, + 0.03297547250986099 + ], + "credit_map": [ + 0.0069664400070905685, + 0.006155907176434994, + 0.009547504596412182, + 0.01004585437476635, + 0.005557742901146412, + 0.0039232103154063225, + 0.007580321282148361, + 0.006436405703425407, + 0.002860303968191147, + 0.002768876263871789, + 0.0031630827579647303, + 0.005042016040533781, + 0.0020972215570509434, + 0.0013336067786440253, + 0.0027507555205374956, + 0.0026547329034656286, + 0.0030035017989575863, + 0.0027481631841510534, + 0.004697269760072231, + 0.004524962045252323, + 0.0026127914898097515, + 0.0019551210571080446, + 0.003557482734322548, + 0.0035730539821088314, + 0.01559841725975275, + 0.01629524677991867, + 0.010719073936343193, + 0.020635629072785378, + 0.011916655115783215, + 0.0070834290236234665, + 0.023002715781331062, + 0.01569380611181259, + 0.004699536599218845, + 0.009070680476725101, + 0.004582563880831003, + 0.01263162586838007, + 0.005803037900477648, + 0.0036298343911767006, + 0.0059282053261995316, + 0.006733981892466545, + 0.003946160431951284, + 0.0032450712751597166, + 0.003444437403231859, + 0.0041654533706605434, + 0.0020721210166811943, + 0.0019244329305365682, + 0.0024471115320920944, + 0.0028307675383985043, + 0.005015394184738398, + 0.00781598873436451, + 0.005859567318111658, + 0.005302507430315018, + 0.004847395233809948, + 0.003582759527489543, + 0.0034833471290767193, + 0.004228657111525536, + 0.029649188742041588, + 0.07282371073961258, + 0.030072391033172607, + 0.03258533403277397, + 0.020437337458133698, + 0.016495076939463615, + 0.017592376098036766, + 0.032927338033914566, + 0.042516350746154785, + 0.09411514550447464, + 0.04749056324362755, + 0.06248556450009346, + 0.024571023881435394, + 0.01746315509080887, + 0.03083610348403454, + 0.051171932369470596, + 0.0035525034181773663, + 0.006198224611580372, + 0.0046895756386220455, + 0.0053129056468605995, + 0.0026031569577753544, + 0.002682160120457411, + 0.0028177856002002954, + 0.0051191640086472034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.09701820462942123 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020605897530913353 + ], + "opposite_logprob_tokens": [ + -3.8956058025360107 + ], + "logprob_reward": -0.020605897530913353, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020605897530913353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04837515577673912, + 0.05130918696522713, + 0.055077120661735535, + 0.038935404270887375, + 0.015678388997912407, + 0.02062428742647171, + 0.045300181955099106, + 0.2671654224395752, + 0.4291246235370636, + 0.02841016836464405 + ], + "credit_map": [ + 0.0051375264301896095, + 0.004665043670684099, + 0.0057939221151173115, + 0.011199056170880795, + 0.005855878349393606, + 0.004320202395319939, + 0.006008294876664877, + 0.005395229905843735, + 0.005340177565813065, + 0.005894102621823549, + 0.007016830146312714, + 0.009332381188869476, + 0.006975297816097736, + 0.0046355631202459335, + 0.004248850978910923, + 0.007865985855460167, + 0.006331645883619785, + 0.007117618806660175, + 0.005213918164372444, + 0.011644766665995121, + 0.006610149517655373, + 0.003349514212459326, + 0.006476064678281546, + 0.008333444595336914, + 0.0036255461163818836, + 0.0047166673466563225, + 0.0035053794272243977, + 0.005043469835072756, + 0.004703648388385773, + 0.003264176892116666, + 0.009184076450765133, + 0.004892445635050535, + 0.001800387166440487, + 0.002235906897112727, + 0.0020075386855751276, + 0.003219645470380783, + 0.0014882787363603711, + 0.0014132724609225988, + 0.0012863503070548177, + 0.002227011602371931, + 0.0025501258205622435, + 0.0034507729578763247, + 0.0029553461354225874, + 0.003973556216806173, + 0.0016684994334354997, + 0.0015565856592729688, + 0.002128038788214326, + 0.0023413640446960926, + 0.005751193035393953, + 0.009042786434292793, + 0.006433822214603424, + 0.005272008944302797, + 0.005333288107067347, + 0.004235826898366213, + 0.004217908252030611, + 0.005013349931687117, + 0.02638975717127323, + 0.08588939160108566, + 0.028858697041869164, + 0.03665291517972946, + 0.019903019070625305, + 0.013783210888504982, + 0.019629063084721565, + 0.03605939447879791, + 0.051466137170791626, + 0.1035149022936821, + 0.058514304459095, + 0.07376932352781296, + 0.02763838693499565, + 0.020816000178456306, + 0.03488980606198311, + 0.058515764772892, + 0.005499920342117548, + 0.003838622011244297, + 0.003983613569289446, + 0.0026774066500365734, + 0.002893138909712434, + 0.002625867957249284, + 0.0020181548316031694, + 0.004873445723205805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.020605897530913353 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330170013010502 + ], + "opposite_logprob_tokens": [ + -3.773301601409912 + ], + "logprob_reward": -0.02330170013010502, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330170013010502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03346453607082367, + 0.042648814618587494, + 0.019043678417801857, + 0.018461275845766068, + 0.015688475221395493, + 0.030050761997699738, + 0.04630538076162338, + 0.3386324942111969, + 0.43485501408576965, + 0.0208495631814003 + ], + "credit_map": [ + 0.003939185757189989, + 0.003357631852850318, + 0.0038176709786057472, + 0.006633457262068987, + 0.003974526189267635, + 0.003092327853664756, + 0.004359146114438772, + 0.004290588200092316, + 0.0038153501227498055, + 0.004014128353446722, + 0.007729572709649801, + 0.0072999680414795876, + 0.003432049648836255, + 0.0034895879216492176, + 0.006189639214426279, + 0.006678519304841757, + 0.0013807895593345165, + 0.0020920594688504934, + 0.0018439999548718333, + 0.003882381599396467, + 0.0021772715263068676, + 0.002134209033101797, + 0.0030241627246141434, + 0.002508804202079773, + 0.0018238313496112823, + 0.0020193217787891626, + 0.0018521696329116821, + 0.0023760220501571894, + 0.0023555250372737646, + 0.002010512165725231, + 0.003590890672057867, + 0.0024330022279173136, + 0.0017924628918990493, + 0.0020478172227740288, + 0.0018245420651510358, + 0.0035577230155467987, + 0.0014575975947082043, + 0.0016881385818123817, + 0.0013046818785369396, + 0.0020155126694589853, + 0.004155347123742104, + 0.0038526307325810194, + 0.004054899327456951, + 0.004633874632418156, + 0.002451574429869652, + 0.0032004695385694504, + 0.0030299844220280647, + 0.0046719820238649845, + 0.006185838952660561, + 0.00922332238405943, + 0.0060630845837295055, + 0.006832183804363012, + 0.0046316878870129585, + 0.004006825387477875, + 0.004177813418209553, + 0.005184624809771776, + 0.03268011286854744, + 0.11250640451908112, + 0.03659228980541229, + 0.046280387789011, + 0.02341209352016449, + 0.01684846356511116, + 0.02427539974451065, + 0.046037331223487854, + 0.05073704943060875, + 0.10730365663766861, + 0.05899141728878021, + 0.07498500496149063, + 0.027891986072063446, + 0.020069817081093788, + 0.03536625951528549, + 0.05950985103845596, + 0.002360641723498702, + 0.004675648175179958, + 0.0036908667534589767, + 0.00246264785528183, + 0.0013796897837892175, + 0.0016002942575141788, + 0.001661596936173737, + 0.0030181780457496643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.02330170013010502 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545006349682808 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012545006349682808, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545006349682808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03579552844166756, + 0.0264336746186018, + 0.014536339789628983, + 0.01341908611357212, + 0.02140762284398079, + 0.045345187187194824, + 0.04618099331855774, + 0.33042052388191223, + 0.42208608984947205, + 0.04437495023012161 + ], + "credit_map": [ + 0.0034424574114382267, + 0.0034671721514314413, + 0.004685141611844301, + 0.007047266699373722, + 0.00349250016734004, + 0.003152522724121809, + 0.005298207048326731, + 0.005210263188928366, + 0.002082300605252385, + 0.002210956299677491, + 0.004226668272167444, + 0.0052147479727864265, + 0.0021692877635359764, + 0.002480153227224946, + 0.0036221486516296864, + 0.004427411127835512, + 0.001091359299607575, + 0.0015397059032693505, + 0.0013694613007828593, + 0.002740997588261962, + 0.001565363840200007, + 0.0017066613072529435, + 0.002722719684243202, + 0.0018000714480876923, + 0.001312187872827053, + 0.0013960005016997457, + 0.0013832454569637775, + 0.002117208205163479, + 0.0014369756681844592, + 0.0016967884730547667, + 0.0024331521708518267, + 0.0016435288125649095, + 0.001969186356291175, + 0.004087704699486494, + 0.001970540266484022, + 0.0055619170889258385, + 0.0018708488205447793, + 0.0014046438736841083, + 0.0018539938610047102, + 0.002688789740204811, + 0.005390291567891836, + 0.006599081680178642, + 0.005460231099277735, + 0.007765042129904032, + 0.0038042606320232153, + 0.005120553076267242, + 0.004152529872953892, + 0.00705319968983531, + 0.0061412956565618515, + 0.00911700539290905, + 0.005758296232670546, + 0.007140567060559988, + 0.0041667986661195755, + 0.004220020025968552, + 0.004286943003535271, + 0.00535007007420063, + 0.03229660913348198, + 0.10927504301071167, + 0.0362795926630497, + 0.04502084478735924, + 0.022286681458353996, + 0.01666703261435032, + 0.023408910259604454, + 0.045185837894678116, + 0.04997313767671585, + 0.10158834606409073, + 0.05673929676413536, + 0.0741313174366951, + 0.027592651546001434, + 0.02023419179022312, + 0.034350231289863586, + 0.057476937770843506, + 0.005406802520155907, + 0.010995958931744099, + 0.005717027001082897, + 0.0065483166836202145, + 0.0028530294075608253, + 0.0025508753024041653, + 0.0036327021662145853, + 0.006670242175459862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.012545006349682808 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly clinging to an emerald green stem visible next to a translucent pale green shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butterfly have a furry black body with wings that are initially wrinkled, damp, and translucent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the butterfly's wings unfold steadily and deliberately while the wrinkles smooth out and the pattern becomes clearer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a warm soft morning light from behind casting a golden rim on the butterfly and dew-dotted green leaves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018177472054958344 + ], + "opposite_logprob_tokens": [ + -4.018177509307861 + ], + "logprob_reward": -0.018177472054958344, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018177472054958344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05396534875035286, + 0.04437784105539322, + 0.02972414717078209, + 0.05158728361129761, + 0.03026721067726612, + 0.023085912689566612, + 0.047233738005161285, + 0.27793827652931213, + 0.4232204556465149, + 0.018599817529320717 + ], + "credit_map": [ + 0.004518555477261543, + 0.006485266610980034, + 0.009620584547519684, + 0.007281840778887272, + 0.005920626223087311, + 0.004313798155635595, + 0.007584779057651758, + 0.008239894174039364, + 0.0032162887509912252, + 0.0052349986508488655, + 0.004626092500984669, + 0.012704952619969845, + 0.004689209163188934, + 0.0032231996301561594, + 0.004859309643507004, + 0.005823786836117506, + 0.0032460293732583523, + 0.003484170651063323, + 0.0024261309299618006, + 0.007798054255545139, + 0.0034018647857010365, + 0.0019765933975577354, + 0.003540856298059225, + 0.003850444918498397, + 0.006900766398757696, + 0.006888285744935274, + 0.004767322912812233, + 0.009196636267006397, + 0.004794170148670673, + 0.0028777862899005413, + 0.009007309563457966, + 0.007155003026127815, + 0.004188681021332741, + 0.003551318310201168, + 0.003031598636880517, + 0.004797063767910004, + 0.004470017738640308, + 0.002416190691292286, + 0.0029085397254675627, + 0.004903800319880247, + 0.0025090582203119993, + 0.005048674531280994, + 0.0028058206662535667, + 0.0037504511419683695, + 0.0022848069202154875, + 0.0016976251499727368, + 0.002250340534374118, + 0.0027391326148062944, + 0.005021383054554462, + 0.009649737738072872, + 0.006658389698714018, + 0.008002471178770065, + 0.004209577105939388, + 0.003134696511551738, + 0.003970998339354992, + 0.006586481351405382, + 0.029255148023366928, + 0.08730804175138474, + 0.03131169080734253, + 0.03702373430132866, + 0.020816650241613388, + 0.016078976914286613, + 0.01920589804649353, + 0.03693810850381851, + 0.048619166016578674, + 0.10486418753862381, + 0.056895382702350616, + 0.07282232493162155, + 0.027784382924437523, + 0.019616862758994102, + 0.03481759503483772, + 0.0578005351126194, + 0.0024913907982409, + 0.003531556809321046, + 0.002904023975133896, + 0.00272164773195982, + 0.0014004296390339732, + 0.0014670796226710081, + 0.0015409501502290368, + 0.002542738104239106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.018177472054958344 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_3/index_tensor([5453, 5453])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_3/index_tensor([5453, 5453])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3ab4aac363d1fa41c339e3865ba15eb4ae54b346 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_3/index_tensor([5453, 5453])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This static, eye-level medium close-up, rendered in the dramatic Todd McFarlane art style, captures a venomous monitor lizard perched on a dark, gnarled branch in predatory stillness. The creature's head is a vibrant crown of blood-ruby scales, possessing a wet, lustrous sheen that contrasts sharply with its body, which is covered in overlapping obsidian-black scales carved with intricate webbed detail. A harsh key light pierces the oppressive darkness, sculpting the lizard's form and casting deep, moody shadows against a backdrop of twisted, blurred foliage. Its menacing, intelligent eyes gleam a malevolent yellow within the deep shadows of its forehead. Captured with surreal clarity at a high frame rate, the creature's head turns smoothly and deliberately, muscles rippling almost imperceptibly beneath its skin, creating an atmosphere of tension and anticipation. As its gaze locks on an unseen target, its slender, forked black tongue darts out in subtle, swift movements, savoring the heavy air with unsettling precision.", + "seed": 21627, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_3/index_5453_seed_21627_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02474297732114792, + "vlm_logprob": -0.02474297732114792, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.02474297732114792, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010830548591911793 + ], + "opposite_logprob_tokens": [ + -4.635830402374268 + ], + "logprob_reward": -0.010830548591911793, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010830548591911793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47670695185661316, + 0.06374022364616394, + 0.05329233035445213, + 0.05260317772626877, + 0.08083705604076385, + 0.05741790309548378, + 0.053791724145412445, + 0.05358631908893585, + 0.05158495530486107, + 0.05643937736749649 + ], + "credit_map": [ + 0.027719413861632347, + 0.030972512438893318, + 0.033007822930812836, + 0.12382768094539642, + 0.02559601701796055, + 0.1711929589509964, + 0.024434462189674377, + 0.03995605558156967, + 0.00515326764434576, + 0.009145091287791729, + 0.008576767519116402, + 0.008585235103964806, + 0.006916648708283901, + 0.00788681861013174, + 0.00896961148828268, + 0.008506777696311474, + 0.0048040952533483505, + 0.003819066099822521, + 0.006433034781366587, + 0.012339718639850616, + 0.002736762398853898, + 0.01117733959108591, + 0.004934387747198343, + 0.0070479256100952625, + 0.004580415319651365, + 0.00489802984520793, + 0.00921620149165392, + 0.00900472141802311, + 0.004348656628280878, + 0.006663855630904436, + 0.005922756623476744, + 0.00796853844076395, + 0.007550177630037069, + 0.007996893487870693, + 0.016929835081100464, + 0.011007832363247871, + 0.007633252069354057, + 0.008157883770763874, + 0.009835446253418922, + 0.011725728400051594, + 0.004776088520884514, + 0.00835090596228838, + 0.011014249175786972, + 0.0060532051138579845, + 0.008154100738465786, + 0.008520017378032207, + 0.005541147198528051, + 0.005008192732930183, + 0.006965364329516888, + 0.00506520364433527, + 0.005849565379321575, + 0.0059072342701256275, + 0.0071833813562989235, + 0.007745812647044659, + 0.0047913808375597, + 0.010283787734806538, + 0.0042947689071297646, + 0.005211883224546909, + 0.010117470286786556, + 0.006699624937027693, + 0.004821361042559147, + 0.011017690412700176, + 0.005485071334987879, + 0.005938447546213865, + 0.00653147604316473, + 0.003990039695054293, + 0.009962126612663269, + 0.007640273775905371, + 0.003139219479635358, + 0.006409490015357733, + 0.006125996820628643, + 0.0077863396145403385, + 0.0062841554172337055, + 0.006600601132959127, + 0.007869693450629711, + 0.0064713251776993275, + 0.004926883615553379, + 0.01022142730653286, + 0.006371188908815384, + 0.007694104220718145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.010830548591911793 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018181568011641502 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.018181568011641502, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018181568011641502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5090420842170715, + 0.07839503139257431, + 0.0532493069767952, + 0.028501097112894058, + 0.05561407655477524, + 0.0839853435754776, + 0.07095613330602646, + 0.05619892477989197, + 0.03415762633085251, + 0.02990036830306053 + ], + "credit_map": [ + 0.029483258724212646, + 0.04565751552581787, + 0.03394061326980591, + 0.12657423317432404, + 0.03100837580859661, + 0.17423497140407562, + 0.02928965911269188, + 0.03885343670845032, + 0.009536663070321083, + 0.0072972155176103115, + 0.011682356707751751, + 0.011273376643657684, + 0.006314652506262064, + 0.013984779827296734, + 0.008620857261121273, + 0.009685131721198559, + 0.004302481655031443, + 0.004967871122062206, + 0.011278868652880192, + 0.007089950144290924, + 0.006734662223607302, + 0.006198971997946501, + 0.0059701078571379185, + 0.0067063914611935616, + 0.0032687941566109657, + 0.002880606334656477, + 0.006071227136999369, + 0.004150701686739922, + 0.0018023870652541518, + 0.004219612572342157, + 0.00300062308087945, + 0.003107145195826888, + 0.005882351193577051, + 0.005192945245653391, + 0.006269556004554033, + 0.007358885835856199, + 0.005386017728596926, + 0.00813274085521698, + 0.006788493134081364, + 0.010603091679513454, + 0.007067054510116577, + 0.010827086865901947, + 0.01736833155155182, + 0.008137778379023075, + 0.012925599701702595, + 0.011864898726344109, + 0.008304542861878872, + 0.00749004865065217, + 0.00761769013479352, + 0.00751460948958993, + 0.008391990326344967, + 0.009010160341858864, + 0.010017389431595802, + 0.010225430130958557, + 0.010024629533290863, + 0.008154232054948807, + 0.006032161880284548, + 0.004154131282120943, + 0.011444585397839546, + 0.006322473753243685, + 0.004919872619211674, + 0.009735465049743652, + 0.006859245244413614, + 0.006730987690389156, + 0.003830089932307601, + 0.003118521301075816, + 0.0062876432202756405, + 0.003995764534920454, + 0.0037057516165077686, + 0.00407467782497406, + 0.00519535131752491, + 0.003949825651943684, + 0.0026022568345069885, + 0.003013042965903878, + 0.0040660337544977665, + 0.0034048715606331825, + 0.0037439267616719007, + 0.005380488000810146, + 0.003094165353104472, + 0.004595582373440266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.018181568011641502 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606132835149765 + ], + "opposite_logprob_tokens": [ + -4.141061305999756 + ], + "logprob_reward": -0.01606132835149765, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606132835149765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5970730185508728, + 0.08512222021818161, + 0.07577574998140335, + 0.04895488917827606, + 0.052041444927453995, + 0.027752241119742393, + 0.04174090176820755, + 0.012664048932492733, + 0.03646266832947731, + 0.022412851452827454 + ], + "credit_map": [ + 0.03436177596449852, + 0.05265923589468002, + 0.040565382689237595, + 0.14803576469421387, + 0.03910904377698898, + 0.2055453062057495, + 0.03374926745891571, + 0.043047238141298294, + 0.007856073789298534, + 0.009783619083464146, + 0.013562338426709175, + 0.009981852024793625, + 0.011704842559993267, + 0.011673679575324059, + 0.010601352900266647, + 0.009958463720977306, + 0.006991340778768063, + 0.005970822647213936, + 0.009112836793065071, + 0.013376327231526375, + 0.0057523213326931, + 0.016679860651493073, + 0.008740875869989395, + 0.009151363745331764, + 0.004080251324921846, + 0.004007895011454821, + 0.010008273646235466, + 0.008178113959729671, + 0.005562076345086098, + 0.004963191691786051, + 0.0047224778681993484, + 0.0074326093308627605, + 0.004105157218873501, + 0.004398240242153406, + 0.009845626540482044, + 0.00642816349864006, + 0.004091968294233084, + 0.0072840191423892975, + 0.0060884165577590466, + 0.009799853898584843, + 0.002907475456595421, + 0.0030386324506253004, + 0.006097286008298397, + 0.003009114181622863, + 0.0035766588989645243, + 0.0032003559172153473, + 0.003798352787271142, + 0.0021243656519800425, + 0.004241475369781256, + 0.0037258262746036053, + 0.009386750869452953, + 0.0048033869825303555, + 0.004759908653795719, + 0.006244965363293886, + 0.004827806726098061, + 0.0037507805973291397, + 0.0015636399621143937, + 0.0009836583631113172, + 0.002625391585752368, + 0.0016254816437140107, + 0.0008503194549120963, + 0.0018996496219187975, + 0.0012147465022280812, + 0.001901161507703364, + 0.004806807730346918, + 0.0027189748361706734, + 0.005708163604140282, + 0.005022251512855291, + 0.001955823739990592, + 0.005986494943499565, + 0.0038013209123164415, + 0.006462831981480122, + 0.0021311561577022076, + 0.0027719628997147083, + 0.0023654045071452856, + 0.002859807573258877, + 0.0024154342245310545, + 0.002818221924826503, + 0.0029965960420668125, + 0.004054269753396511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01606132835149765 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000470161437988 + ], + "opposite_logprob_tokens": [ + -2.69500470161438 + ], + "logprob_reward": -0.07000470161437988, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000470161437988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8395169377326965, + 0.05285252630710602, + 0.011697043664753437, + 0.012370857410132885, + 0.017734134569764137, + 0.024430330842733383, + 0.011267507448792458, + 0.010500588454306126, + 0.00679769366979599, + 0.012832400389015675 + ], + "credit_map": [ + 0.04690920561552048, + 0.07286839187145233, + 0.054389987140893936, + 0.20917023718357086, + 0.051835112273693085, + 0.2937425673007965, + 0.04613310843706131, + 0.06446834653615952, + 0.006062056869268417, + 0.00516542000696063, + 0.008565861731767654, + 0.006806815508753061, + 0.0055970135144889355, + 0.007247467990964651, + 0.005484716966748238, + 0.007923178374767303, + 0.0011530248448252678, + 0.0007701446884311736, + 0.0012037587584927678, + 0.0022848169319331646, + 0.0011880712117999792, + 0.0022275226656347513, + 0.0013589998707175255, + 0.0015107057988643646, + 0.001542065292596817, + 0.0009933165274560452, + 0.0017352020367980003, + 0.0018639563350006938, + 0.0013387695653364062, + 0.0016949649434536695, + 0.0012264265678822994, + 0.001976157771423459, + 0.0016912504797801375, + 0.0017514098435640335, + 0.003338887356221676, + 0.0020470204763114452, + 0.002337951445952058, + 0.0019453710410743952, + 0.0021752442698925734, + 0.0024470000062137842, + 0.0022224707063287497, + 0.002866304013878107, + 0.0033993071410804987, + 0.0026915345806628466, + 0.0027410651091486216, + 0.004586259834468365, + 0.0026679334696382284, + 0.003255459712818265, + 0.0010041348868981004, + 0.0012972045224159956, + 0.0017868026625365019, + 0.001297669019550085, + 0.0012583202915266156, + 0.0014393277233466506, + 0.0010144057450816035, + 0.0021696435287594795, + 0.0008467460284009576, + 0.0006339290412142873, + 0.002692839130759239, + 0.0014180407160893083, + 0.0009381011477671564, + 0.001053937361575663, + 0.0012351482873782516, + 0.0016818471485748887, + 0.0009773405035957694, + 0.0005084811127744615, + 0.0009479714790359139, + 0.0008483208948746324, + 0.0005840722005814314, + 0.0007746880874037743, + 0.0008618122665211558, + 0.0012950077652931213, + 0.0011308264220133424, + 0.0014314005384221673, + 0.0021510750520974398, + 0.0016341863665729761, + 0.0012698281789198518, + 0.0015968679217621684, + 0.0016742788720875978, + 0.0019439386669546366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.07000470161437988 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636740036308765 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636740036308765, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636740036308765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4923366904258728, + 0.12803179025650024, + 0.061534274369478226, + 0.09478921443223953, + 0.035977624356746674, + 0.03262351453304291, + 0.04367430880665779, + 0.03162970766425133, + 0.04236111417412758, + 0.03704174607992172 + ], + "credit_map": [ + 0.03464794531464577, + 0.05378665030002594, + 0.042474474757909775, + 0.09478513151407242, + 0.04085709527134895, + 0.14420798420906067, + 0.03468911722302437, + 0.04688828065991402, + 0.013631565496325493, + 0.017010588198900223, + 0.018613176420331, + 0.016517963260412216, + 0.016385655850172043, + 0.01829298958182335, + 0.01457756944000721, + 0.013002276420593262, + 0.0070174746215343475, + 0.005438573192805052, + 0.007241721730679274, + 0.009636134840548038, + 0.0039687068201601505, + 0.012511366978287697, + 0.006637171376496553, + 0.009083124808967113, + 0.007723863702267408, + 0.008214186877012253, + 0.01980629377067089, + 0.01621348038315773, + 0.010318429209291935, + 0.013376743532717228, + 0.008017282001674175, + 0.0111189354211092, + 0.004496782086789608, + 0.004357298836112022, + 0.0061031971126794815, + 0.0038894203025847673, + 0.003395088715478778, + 0.005825503263622522, + 0.0034951032139360905, + 0.004415231291204691, + 0.003501677652820945, + 0.004051766823977232, + 0.005366077180951834, + 0.003574045840650797, + 0.004388225730508566, + 0.004816825967282057, + 0.004180123098194599, + 0.002744773868471384, + 0.0050261481665074825, + 0.0039679016917943954, + 0.008861266076564789, + 0.005055887624621391, + 0.005307565443217754, + 0.00544664915651083, + 0.005918068811297417, + 0.004090824164450169, + 0.004786464385688305, + 0.002495670225471258, + 0.004471971187740564, + 0.0038785641081631184, + 0.0028126456309109926, + 0.004868894349783659, + 0.003958206158131361, + 0.004357293713837862, + 0.006514393724501133, + 0.003775323275476694, + 0.00693913409486413, + 0.005367799196392298, + 0.002624711487442255, + 0.005655086599290371, + 0.005163877736777067, + 0.006320788525044918, + 0.0031579516362398863, + 0.0033533566165715456, + 0.004124341998249292, + 0.004829172510653734, + 0.002422197023406625, + 0.0045149121433496475, + 0.004217661917209625, + 0.010422151535749435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.008636740036308765 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This static, eye-level medium close-up, rendered in the dramatic Todd McFarlane art style, captures a venomous monitor lizard perched on a dark, gnarled branch in predatory stillness. The creature's head is a vibrant crown of blood-ruby scales, possessing a wet, lustrous sheen that contrasts sharply with its body, which is covered in overlapping obsidian-black scales carved with intricate webbed detail. A harsh key light pierces the oppressive darkness, sculpting the lizard's form and casting deep, moody shadows against a backdrop of twisted, blurred foliage. Its menacing, intelligent eyes gleam a malevolent yellow within the deep shadows of its forehead. Captured with surreal clarity at a high frame rate, the creature's head turns smoothly and deliberately, muscles rippling almost imperceptibly beneath its skin, creating an atmosphere of tension and anticipation. As its gaze locks on an unseen target, its slender, forked black tongue darts out in subtle, swift movements, savoring the heavy air with unsettling precision.", + "seed": 21628, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_3/index_5453_seed_21628_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02453039325773716, + "vlm_logprob": -0.02453039325773716, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.02453039325773716, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020381074398756027 + ], + "opposite_logprob_tokens": [ + -4.020380973815918 + ], + "logprob_reward": -0.020381074398756027, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.020381074398756027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3277710974216461, + 0.07404915243387222, + 0.043163422495126724, + 0.17625603079795837, + 0.1329168677330017, + 0.07896135747432709, + 0.029267562553286552, + 0.02423100732266903, + 0.050348326563835144, + 0.06303517520427704 + ], + "credit_map": [ + 0.035201165825128555, + 0.04180577024817467, + 0.042376771569252014, + 0.0482293963432312, + 0.037784919142723083, + 0.04851715266704559, + 0.038586899638175964, + 0.03526901453733444, + 0.006134871859103441, + 0.005972808226943016, + 0.010654186829924583, + 0.014164619147777557, + 0.008878507651388645, + 0.010035854764282703, + 0.008020258508622646, + 0.010188051499426365, + 0.003767962334677577, + 0.004056531935930252, + 0.006108009722083807, + 0.006064288318157196, + 0.004829217679798603, + 0.0065450649708509445, + 0.006520786322653294, + 0.005271559581160545, + 0.013833655044436455, + 0.022719278931617737, + 0.02553747594356537, + 0.028148103505373, + 0.019718583673238754, + 0.018184257671236992, + 0.024280715733766556, + 0.02383396588265896, + 0.01377114374190569, + 0.015825603157281876, + 0.017947863787412643, + 0.02109840326011181, + 0.015552259981632233, + 0.01758771575987339, + 0.0179252028465271, + 0.013208676129579544, + 0.006498483009636402, + 0.010117480531334877, + 0.010809109546244144, + 0.01394642610102892, + 0.007147784810513258, + 0.012143022380769253, + 0.008360493928194046, + 0.009938552975654602, + 0.002961326390504837, + 0.0038877876941114664, + 0.003527949098497629, + 0.004846127703785896, + 0.0021514848340302706, + 0.0050514377653598785, + 0.003584402846172452, + 0.0032570459879934788, + 0.002264160430058837, + 0.002304836641997099, + 0.0041813659481704235, + 0.003865522565320134, + 0.0026082992553710938, + 0.0028167201671749353, + 0.003091485472396016, + 0.0030986180063337088, + 0.004487390164285898, + 0.00965852476656437, + 0.007024271879345179, + 0.007581057492643595, + 0.0045652082189917564, + 0.005360977724194527, + 0.0066413492895662785, + 0.005029548890888691, + 0.009343561716377735, + 0.006583332549780607, + 0.00850584078580141, + 0.007429768331348896, + 0.008022340945899487, + 0.006870106793940067, + 0.009039769880473614, + 0.00724045280367136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.020381074398756027 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542770244181156 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.012542770244181156, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542770244181156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33798959851264954, + 0.10795390605926514, + 0.055658724159002304, + 0.1494668424129486, + 0.09464230388402939, + 0.02710178680717945, + 0.026632599532604218, + 0.019455155357718468, + 0.035046715289354324, + 0.14605233073234558 + ], + "credit_map": [ + 0.03583839535713196, + 0.04940851032733917, + 0.037567418068647385, + 0.048028457909822464, + 0.03960106894373894, + 0.05193323269486427, + 0.038487017154693604, + 0.03712553158402443, + 0.01508957240730524, + 0.017832260578870773, + 0.014862488023936749, + 0.014828257262706757, + 0.010523738339543343, + 0.012291382066905499, + 0.011406872421503067, + 0.011119340546429157, + 0.0060538025572896, + 0.008040453307330608, + 0.00943909864872694, + 0.00886333454400301, + 0.004786142148077488, + 0.004190362989902496, + 0.007292789872735739, + 0.006992744281888008, + 0.018108416348695755, + 0.023148639127612114, + 0.020925071090459824, + 0.02012125961482525, + 0.01706491969525814, + 0.015017393976449966, + 0.015040453523397446, + 0.02004070393741131, + 0.01001564972102642, + 0.010871967300772667, + 0.0133215868845582, + 0.014085045084357262, + 0.012472888454794884, + 0.014663011766970158, + 0.008483829908072948, + 0.01072833500802517, + 0.0026830413844436407, + 0.0040803346782922745, + 0.0038967125583440065, + 0.003951750695705414, + 0.0028985152021050453, + 0.002987603424116969, + 0.003051125444471836, + 0.003552705282345414, + 0.0028048160020262003, + 0.0025000947061926126, + 0.00346104777418077, + 0.0037031006067991257, + 0.0029619382694363594, + 0.0046564205549657345, + 0.003340476891025901, + 0.0032047058921307325, + 0.0025820715818554163, + 0.0018200104823336005, + 0.0023567702155560255, + 0.0034019441809505224, + 0.002089989371597767, + 0.0026323599740862846, + 0.0027078900020569563, + 0.0018641228089109063, + 0.0037119542248547077, + 0.004610002972185612, + 0.004031599033623934, + 0.0045080664567649364, + 0.004715189337730408, + 0.0030108874198049307, + 0.005357160698622465, + 0.005101857241243124, + 0.014727705158293247, + 0.016274921596050262, + 0.021328743547201157, + 0.02049887366592884, + 0.016465218737721443, + 0.015451629646122456, + 0.022712277248501778, + 0.018592989072203636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.012542770244181156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577870309352875 + ], + "opposite_logprob_tokens": [ + -3.895577907562256 + ], + "logprob_reward": -0.020577870309352875, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577870309352875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2660882771015167, + 0.08464273810386658, + 0.06454673409461975, + 0.19406743347644806, + 0.0672890692949295, + 0.0753246322274208, + 0.03513587638735771, + 0.02723965235054493, + 0.10501208901405334, + 0.08065347373485565 + ], + "credit_map": [ + 0.035972971469163895, + 0.03271634876728058, + 0.03092311881482601, + 0.03919012472033501, + 0.025148864835500717, + 0.03712595999240875, + 0.03143174201250076, + 0.03357913717627525, + 0.005617037881165743, + 0.011705909855663776, + 0.012939965352416039, + 0.014670019969344139, + 0.007816031575202942, + 0.009604403749108315, + 0.010752670466899872, + 0.011536708101630211, + 0.005215190816670656, + 0.007007033564150333, + 0.00954988319426775, + 0.012352759949862957, + 0.005755494348704815, + 0.006014783401042223, + 0.011192990466952324, + 0.0074585964903235435, + 0.017928803339600563, + 0.024108855053782463, + 0.02753380313515663, + 0.024898869916796684, + 0.028612889349460602, + 0.028617141768336296, + 0.020298175513744354, + 0.022068874910473824, + 0.007957855239510536, + 0.008104156702756882, + 0.00843574944883585, + 0.011202752590179443, + 0.009526897221803665, + 0.006610756739974022, + 0.00855680089443922, + 0.006894100923091173, + 0.006589354481548071, + 0.01086149737238884, + 0.01015267800539732, + 0.01274210773408413, + 0.007052426692098379, + 0.011154764331877232, + 0.0078928517177701, + 0.008878951892256737, + 0.003055195091292262, + 0.003623383119702339, + 0.004383604507893324, + 0.005062949378043413, + 0.004929923918098211, + 0.006381590384989977, + 0.0037420408334583044, + 0.003957186825573444, + 0.0025355189573019743, + 0.003997716121375561, + 0.0038488961290568113, + 0.004381845705211163, + 0.0030378075316548347, + 0.0025624458212405443, + 0.0029275864362716675, + 0.003947836346924305, + 0.008008499629795551, + 0.020443551242351532, + 0.013673401437699795, + 0.01616358384490013, + 0.01012202724814415, + 0.01306221168488264, + 0.009440965950489044, + 0.014097853563725948, + 0.00673158373683691, + 0.007907774299383163, + 0.011648109182715416, + 0.010116969235241413, + 0.009357496164739132, + 0.01141721848398447, + 0.011566709727048874, + 0.01190761849284172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.020577870309352875 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549527183175087 + ], + "opposite_logprob_tokens": [ + -2.929952621459961 + ], + "logprob_reward": -0.0549527183175087, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549527183175087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.197406604886055, + 0.06412975490093231, + 0.0646679699420929, + 0.19531947374343872, + 0.04397260397672653, + 0.07374470680952072, + 0.05226314440369606, + 0.0525568425655365, + 0.10288330912590027, + 0.1530555784702301 + ], + "credit_map": [ + 0.014283553697168827, + 0.0254774522036314, + 0.025773992761969566, + 0.030037028715014458, + 0.01853271573781967, + 0.03186996281147003, + 0.026043666526675224, + 0.025388233363628387, + 0.0055785831063985825, + 0.011874099262058735, + 0.009306023828685284, + 0.009353461675345898, + 0.005260706413537264, + 0.008290658704936504, + 0.006750272121280432, + 0.007715947460383177, + 0.006645881570875645, + 0.00790609885007143, + 0.00876858364790678, + 0.009637372568249702, + 0.006576696410775185, + 0.008842648938298225, + 0.007305163890123367, + 0.008985523134469986, + 0.017195696011185646, + 0.028458857908844948, + 0.025904979556798935, + 0.028452979400753975, + 0.03242316097021103, + 0.019334577023983, + 0.024167053401470184, + 0.019382158294320107, + 0.004988734144717455, + 0.006150267086923122, + 0.005831584334373474, + 0.006895608268678188, + 0.006072456948459148, + 0.00450070109218359, + 0.00522100692614913, + 0.004312243312597275, + 0.007852103561162949, + 0.008638661354780197, + 0.010434980504214764, + 0.011464581824839115, + 0.008711122907698154, + 0.009471169672906399, + 0.010171713307499886, + 0.007000375539064407, + 0.006339680403470993, + 0.007144095841795206, + 0.0064596934244036674, + 0.006843701470643282, + 0.006440622732043266, + 0.005032182205468416, + 0.007158320397138596, + 0.006844846997410059, + 0.004653478506952524, + 0.007614036090672016, + 0.006563297472894192, + 0.008529450744390488, + 0.005320832598954439, + 0.007663680240511894, + 0.005450637545436621, + 0.0067614284344017506, + 0.00775177264586091, + 0.025021657347679138, + 0.0140133798122406, + 0.018038073554635048, + 0.010767103172838688, + 0.006892253644764423, + 0.010963023640215397, + 0.0094360476359725, + 0.016889387741684914, + 0.0156870037317276, + 0.023754790425300598, + 0.02104956842958927, + 0.01649419404566288, + 0.012575097382068634, + 0.025840383023023605, + 0.020765146240592003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.0549527183175087 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197533018887043 + ], + "opposite_logprob_tokens": [ + -4.26419734954834 + ], + "logprob_reward": -0.014197533018887043, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197533018887043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2964407503604889, + 0.10596995800733566, + 0.09214276820421219, + 0.07985477149486542, + 0.07885533571243286, + 0.05328669771552086, + 0.0753684937953949, + 0.06098364666104317, + 0.07074670493602753, + 0.08635088801383972 + ], + "credit_map": [ + 0.03710038214921951, + 0.036181047558784485, + 0.044161297380924225, + 0.057699788361787796, + 0.024563711136579514, + 0.03441673889756203, + 0.024430083110928535, + 0.03788772597908974, + 0.01154654286801815, + 0.016284320503473282, + 0.01586565561592579, + 0.01660866290330887, + 0.012777305208146572, + 0.010532275773584843, + 0.010461490601301193, + 0.01189370732754469, + 0.006813721265643835, + 0.011865660548210144, + 0.012066617608070374, + 0.01408400572836399, + 0.009275884367525578, + 0.009225851856172085, + 0.017379822209477425, + 0.011431203223764896, + 0.01124697644263506, + 0.009622779674828053, + 0.012817264534533024, + 0.011219161562621593, + 0.00652293348684907, + 0.008197052404284477, + 0.01102534681558609, + 0.009203257039189339, + 0.007878395728766918, + 0.010149520821869373, + 0.009508059360086918, + 0.013546790927648544, + 0.010517410933971405, + 0.007811026647686958, + 0.010909749194979668, + 0.0085343848913908, + 0.005391997285187244, + 0.007434119936078787, + 0.007163509726524353, + 0.008814929984509945, + 0.005140939261764288, + 0.007675221189856529, + 0.005591247230768204, + 0.0060747298412024975, + 0.005946018733084202, + 0.007374155335128307, + 0.008675416000187397, + 0.013084396719932556, + 0.008903072215616703, + 0.008972007781267166, + 0.012907072901725769, + 0.009506353177130222, + 0.008249295875430107, + 0.007420901209115982, + 0.008219439536333084, + 0.00839401874691248, + 0.006096788682043552, + 0.008009476587176323, + 0.008142273873090744, + 0.006451444700360298, + 0.00649391720071435, + 0.01717545837163925, + 0.009162010625004768, + 0.011420642957091331, + 0.006854074075818062, + 0.0055785938166081905, + 0.0074310507625341415, + 0.006630954798310995, + 0.011889615096151829, + 0.009161015041172504, + 0.012741570360958576, + 0.012269322760403156, + 0.011101157404482365, + 0.008533813059329987, + 0.011615660041570663, + 0.009038730524480343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014197533018887043 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This static, eye-level medium close-up, rendered in the dramatic Todd McFarlane art style, captures a venomous monitor lizard perched on a dark, gnarled branch in predatory stillness. The creature's head is a vibrant crown of blood-ruby scales, possessing a wet, lustrous sheen that contrasts sharply with its body, which is covered in overlapping obsidian-black scales carved with intricate webbed detail. A harsh key light pierces the oppressive darkness, sculpting the lizard's form and casting deep, moody shadows against a backdrop of twisted, blurred foliage. Its menacing, intelligent eyes gleam a malevolent yellow within the deep shadows of its forehead. Captured with surreal clarity at a high frame rate, the creature's head turns smoothly and deliberately, muscles rippling almost imperceptibly beneath its skin, creating an atmosphere of tension and anticipation. As its gaze locks on an unseen target, its slender, forked black tongue darts out in subtle, swift movements, savoring the heavy air with unsettling precision.", + "seed": 21629, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_3/index_5453_seed_21629_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.013128179032355547, + "vlm_logprob": -0.013128179032355547, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.013128179032355547, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009143739938735962 + ], + "opposite_logprob_tokens": [ + -4.884143829345703 + ], + "logprob_reward": -0.009143739938735962, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.009143739938735962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14146876335144043, + 0.14048391580581665, + 0.06670165061950684, + 0.20746147632598877, + 0.06235237792134285, + 0.20282793045043945, + 0.043325576931238174, + 0.033763058483600616, + 0.033759068697690964, + 0.06785621494054794 + ], + "credit_map": [ + 0.018447179347276688, + 0.011136846616864204, + 0.01245616190135479, + 0.01908758096396923, + 0.017713584005832672, + 0.019384942948818207, + 0.025050992146134377, + 0.018191486597061157, + 0.02160303108394146, + 0.01063263788819313, + 0.010410293005406857, + 0.008852235972881317, + 0.008269126527011395, + 0.0326857827603817, + 0.03350699320435524, + 0.014523809775710106, + 0.012450463138520718, + 0.004974985495209694, + 0.005334543064236641, + 0.005352499429136515, + 0.004506177734583616, + 0.013597953133285046, + 0.011658319272100925, + 0.008826712146401405, + 0.015245137736201286, + 0.015566096641123295, + 0.005907902494072914, + 0.004478019196540117, + 0.0056528435088694096, + 0.08361561596393585, + 0.016207223758101463, + 0.06078863888978958, + 0.008945646695792675, + 0.004915814381092787, + 0.007183882873505354, + 0.008001178503036499, + 0.005258720833808184, + 0.008723415434360504, + 0.00986344926059246, + 0.009460271336138248, + 0.03290921822190285, + 0.014519091695547104, + 0.028752189129590988, + 0.018928425386548042, + 0.020856022834777832, + 0.033151958137750626, + 0.03421718254685402, + 0.019493838772177696, + 0.005137065891176462, + 0.0037063525523990393, + 0.005234625190496445, + 0.0044837347231805325, + 0.004393583629280329, + 0.006961601320654154, + 0.007666009943932295, + 0.005742604378610849, + 0.003076625522226095, + 0.002614092780277133, + 0.002350354567170143, + 0.003086287062615156, + 0.0026045325212180614, + 0.008221897296607494, + 0.003987721633166075, + 0.007821545004844666, + 0.003557124175131321, + 0.0021141739562153816, + 0.0030861562117934227, + 0.003769251750782132, + 0.0022346253972500563, + 0.006553810555487871, + 0.005558345466852188, + 0.0068855807185173035, + 0.007044367026537657, + 0.00657569570466876, + 0.005613757763057947, + 0.00901114847511053, + 0.0057680136524140835, + 0.009033245965838432, + 0.015065737068653107, + 0.009744247421622276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.009143739938735962 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947037134319544 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005947037134319544, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947037134319544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14610977470874786, + 0.1047346442937851, + 0.08448245376348495, + 0.23129762709140778, + 0.09789656102657318, + 0.12366905808448792, + 0.0829564779996872, + 0.05132026970386505, + 0.033157140016555786, + 0.0443759523332119 + ], + "credit_map": [ + 0.018883511424064636, + 0.012679219245910645, + 0.01067129522562027, + 0.015460112132132053, + 0.015044299885630608, + 0.01820424199104309, + 0.033735573291778564, + 0.021431531757116318, + 0.017171118408441544, + 0.01327296532690525, + 0.011988439597189426, + 0.01060889009386301, + 0.006879322696477175, + 0.015075418166816235, + 0.017496734857559204, + 0.01224176399409771, + 0.010793659836053848, + 0.006218103691935539, + 0.00656677084043622, + 0.006459375843405724, + 0.008543303236365318, + 0.019309524446725845, + 0.018529126420617104, + 0.00806259922683239, + 0.017724597826600075, + 0.01673373021185398, + 0.00650817109271884, + 0.0065482608042657375, + 0.010330461896955967, + 0.0917876586318016, + 0.020526446402072906, + 0.0611383281648159, + 0.013313381001353264, + 0.00542449438944459, + 0.010582543909549713, + 0.008593374863266945, + 0.01315601821988821, + 0.0150448651984334, + 0.01410518866032362, + 0.0176767036318779, + 0.01876131072640419, + 0.007528671529144049, + 0.01832740195095539, + 0.013176413252949715, + 0.013879104517400265, + 0.021565154194831848, + 0.02056754007935524, + 0.009863468818366528, + 0.009606999345123768, + 0.005196142476052046, + 0.009039018303155899, + 0.009131114929914474, + 0.006305686663836241, + 0.01397059578448534, + 0.017795247957110405, + 0.011911677196621895, + 0.005023545119911432, + 0.0031071612611413, + 0.0034661050885915756, + 0.0037253485061228275, + 0.00276008783839643, + 0.014337224885821342, + 0.006859793793410063, + 0.012041004374623299, + 0.0030045972671359777, + 0.0019224180141463876, + 0.0026057560462504625, + 0.0032008294947445393, + 0.002159122144803405, + 0.005573162343353033, + 0.009502832777798176, + 0.005188428331166506, + 0.006586674600839615, + 0.005023788660764694, + 0.004473818466067314, + 0.0072158626280725, + 0.003716533537954092, + 0.0055946847423911095, + 0.006597703322768211, + 0.005166892893612385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.005947037134319544 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018175598233938217 + ], + "opposite_logprob_tokens": [ + -4.0181756019592285 + ], + "logprob_reward": -0.018175598233938217, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018175598233938217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12953461706638336, + 0.054999567568302155, + 0.10761741548776627, + 0.26178407669067383, + 0.12042339891195297, + 0.14141841232776642, + 0.0748586505651474, + 0.0457942858338356, + 0.03220846503973007, + 0.031361065804958344 + ], + "credit_map": [ + 0.017070982605218887, + 0.00930891651660204, + 0.015897022560238838, + 0.014859544113278389, + 0.014649415388703346, + 0.019736411049962044, + 0.02238544635474682, + 0.015626898035407066, + 0.011579832062125206, + 0.004286268725991249, + 0.006291485857218504, + 0.005131093319505453, + 0.004521504044532776, + 0.008130872622132301, + 0.006981799378991127, + 0.008076719008386135, + 0.013726835139095783, + 0.006490825209766626, + 0.0060109831392765045, + 0.006195564288645983, + 0.00740479864180088, + 0.0370418056845665, + 0.02290121279656887, + 0.007845394313335419, + 0.021038396283984184, + 0.017779022455215454, + 0.009788887575268745, + 0.008911389857530594, + 0.011994830332696438, + 0.10036584734916687, + 0.02776906080543995, + 0.06413665413856506, + 0.012166623026132584, + 0.01195022277534008, + 0.014617150649428368, + 0.013827950693666935, + 0.016898391768336296, + 0.01798471249639988, + 0.021731199696660042, + 0.011247161775827408, + 0.02374957874417305, + 0.00919297058135271, + 0.019315386191010475, + 0.013538846746087074, + 0.013940173201262951, + 0.024285048246383667, + 0.024742648005485535, + 0.012653767131268978, + 0.0091971131041646, + 0.005156452301889658, + 0.009485771879553795, + 0.006935009732842445, + 0.005659337155520916, + 0.012772738933563232, + 0.014800306409597397, + 0.010851928032934666, + 0.00519276037812233, + 0.0024440099950879812, + 0.002564702881500125, + 0.0028767341282218695, + 0.0027265590615570545, + 0.01394482608884573, + 0.004041280597448349, + 0.012003417126834393, + 0.0038310459349304438, + 0.0018930650549009442, + 0.0035384278744459152, + 0.0030105803161859512, + 0.0021334984339773655, + 0.006490753032267094, + 0.006101713050156832, + 0.005209381226450205, + 0.003411103505641222, + 0.002396468771621585, + 0.0025512585416436195, + 0.0042625791393220425, + 0.0028976064641028643, + 0.004196315072476864, + 0.00626307213678956, + 0.00538266496732831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.018175598233938217 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818133518099785 + ], + "opposite_logprob_tokens": [ + -4.018181324005127 + ], + "logprob_reward": -0.01818133518099785, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818133518099785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22823219001293182, + 0.1743592470884323, + 0.1154324859380722, + 0.1259039342403412, + 0.08107564598321915, + 0.12212715297937393, + 0.046649400144815445, + 0.037060271948575974, + 0.03923933207988739, + 0.029920358210802078 + ], + "credit_map": [ + 0.024105144664645195, + 0.019817546010017395, + 0.02246243692934513, + 0.020195255056023598, + 0.019603708758950233, + 0.03789114952087402, + 0.04179679974913597, + 0.042360156774520874, + 0.03190813213586807, + 0.01282207015901804, + 0.013817169703543186, + 0.013273281045258045, + 0.017470521852374077, + 0.03629358112812042, + 0.03120187483727932, + 0.01757262833416462, + 0.018595227971673012, + 0.00853611622005701, + 0.00834528636187315, + 0.011394983157515526, + 0.008634313009679317, + 0.027634702622890472, + 0.01798897795379162, + 0.014302878640592098, + 0.012244705110788345, + 0.008622497320175171, + 0.005779665429145098, + 0.006634365301579237, + 0.006526796147227287, + 0.04267874360084534, + 0.009511645883321762, + 0.0339055135846138, + 0.006576568819582462, + 0.006736115086823702, + 0.007105591241270304, + 0.00974455289542675, + 0.010473951697349548, + 0.011378007009625435, + 0.015492473728954792, + 0.013568385504186153, + 0.016553416848182678, + 0.009011120535433292, + 0.016450421884655952, + 0.0128104779869318, + 0.013539373874664307, + 0.018615951761603355, + 0.02103975974023342, + 0.014106634072959423, + 0.005393357947468758, + 0.002534580649808049, + 0.004457094240933657, + 0.006719795521348715, + 0.005444888491183519, + 0.00630735419690609, + 0.008699072524905205, + 0.007093258202075958, + 0.005628732964396477, + 0.002487373072654009, + 0.004032182972878218, + 0.004345161374658346, + 0.0030005283188074827, + 0.006982756778597832, + 0.004207039717584848, + 0.006376494187861681, + 0.005005854181945324, + 0.0035087077412754297, + 0.005320784170180559, + 0.0045926813036203384, + 0.0036451041232794523, + 0.006907823495566845, + 0.005138102453202009, + 0.0051202718168497086, + 0.0038140485994517803, + 0.0021321214735507965, + 0.0028658576775342226, + 0.0047453888691961765, + 0.0032291761599481106, + 0.004000172019004822, + 0.004599657375365496, + 0.004533935338258743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.01818133518099785 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193184673786163 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014193184673786163, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193184673786163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11405634880065918, + 0.10796792060136795, + 0.2840748727321625, + 0.0843714252114296, + 0.08165962249040604, + 0.13633733987808228, + 0.07074662297964096, + 0.05611851066350937, + 0.03433104231953621, + 0.030336366966366768 + ], + "credit_map": [ + 0.01567191816866398, + 0.012915102764964104, + 0.008887688629329205, + 0.013041440397500992, + 0.012127007357776165, + 0.014015567488968372, + 0.021930856630206108, + 0.015466758981347084, + 0.024512065574526787, + 0.010226739570498466, + 0.010200370103120804, + 0.00994111504405737, + 0.010328546166419983, + 0.015882374718785286, + 0.0121093625202775, + 0.014767344109714031, + 0.0731869712471962, + 0.021338164806365967, + 0.025778353214263916, + 0.01463480107486248, + 0.015194813720881939, + 0.06491384655237198, + 0.04343581572175026, + 0.025592084974050522, + 0.0073349373415112495, + 0.0070486292243003845, + 0.00412720488384366, + 0.0052146511152386665, + 0.006009811535477638, + 0.024639585986733437, + 0.011350508779287338, + 0.01864609308540821, + 0.010845434851944447, + 0.007558489218354225, + 0.008414322510361671, + 0.007152508944272995, + 0.008616681210696697, + 0.0123375803232193, + 0.01700054109096527, + 0.009734056890010834, + 0.021764634177088737, + 0.008957549929618835, + 0.017652755603194237, + 0.01143100205808878, + 0.012000488117337227, + 0.026760432869195938, + 0.022032322362065315, + 0.01573815755546093, + 0.00636174576357007, + 0.0034138860646635294, + 0.006605952046811581, + 0.007129888515919447, + 0.00553264282643795, + 0.011629651300609112, + 0.017114926129579544, + 0.012957927770912647, + 0.005589173175394535, + 0.0024431690108031034, + 0.00341928843408823, + 0.0033444783184677362, + 0.0026760846376419067, + 0.01734914816915989, + 0.005815927870571613, + 0.015481237322092056, + 0.004669301211833954, + 0.002400851808488369, + 0.00413076626136899, + 0.0032572420313954353, + 0.002802617382258177, + 0.006833518389612436, + 0.004662470426410437, + 0.005574275739490986, + 0.003698228392750025, + 0.0025334435049444437, + 0.00281900935806334, + 0.0034851578529924154, + 0.002452543005347252, + 0.004493748769164085, + 0.006422668229788542, + 0.004431563895195723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.014193184673786163 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This static, eye-level medium close-up, rendered in the dramatic Todd McFarlane art style, captures a venomous monitor lizard perched on a dark, gnarled branch in predatory stillness. The creature's head is a vibrant crown of blood-ruby scales, possessing a wet, lustrous sheen that contrasts sharply with its body, which is covered in overlapping obsidian-black scales carved with intricate webbed detail. A harsh key light pierces the oppressive darkness, sculpting the lizard's form and casting deep, moody shadows against a backdrop of twisted, blurred foliage. Its menacing, intelligent eyes gleam a malevolent yellow within the deep shadows of its forehead. Captured with surreal clarity at a high frame rate, the creature's head turns smoothly and deliberately, muscles rippling almost imperceptibly beneath its skin, creating an atmosphere of tension and anticipation. As its gaze locks on an unseen target, its slender, forked black tongue darts out in subtle, swift movements, savoring the heavy air with unsettling precision.", + "seed": 21630, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_3/index_5453_seed_21630_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01837383219972253, + "vlm_logprob": -0.01837383219972253, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.01837383219972253, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00956795271486044 + ], + "opposite_logprob_tokens": [ + -4.759567737579346 + ], + "logprob_reward": -0.00956795271486044, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00956795271486044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2593654692173004, + 0.070511594414711, + 0.03649701178073883, + 0.04305680841207504, + 0.07444258779287338, + 0.05953224375844002, + 0.06984548270702362, + 0.036353856325149536, + 0.0579688586294651, + 0.29242607951164246 + ], + "credit_map": [ + 0.0346110463142395, + 0.017340080812573433, + 0.040619563311338425, + 0.05020882189273834, + 0.017991455271840096, + 0.041414204984903336, + 0.02207981050014496, + 0.03510049358010292, + 0.008894065394997597, + 0.007407496217638254, + 0.0115773044526577, + 0.008931548334658146, + 0.006509342696517706, + 0.013433963060379028, + 0.0047641838900744915, + 0.00899368803948164, + 0.004403727129101753, + 0.0026433062739670277, + 0.0059344349429011345, + 0.005420624278485775, + 0.0024084080941975117, + 0.007582667283713818, + 0.0025928448885679245, + 0.005511000752449036, + 0.004875916056334972, + 0.005059152841567993, + 0.00644269771873951, + 0.006824786774814129, + 0.0031690048053860664, + 0.008320651017129421, + 0.0025119197089225054, + 0.005852679256349802, + 0.009206793271005154, + 0.004815308842808008, + 0.013896065764129162, + 0.007009183056652546, + 0.004296247847378254, + 0.01913420483469963, + 0.005815250799059868, + 0.01026953849941492, + 0.008472600020468235, + 0.003971376456320286, + 0.010504710488021374, + 0.008302493952214718, + 0.003930195700377226, + 0.012112593278288841, + 0.003448060480877757, + 0.008790217339992523, + 0.008371355943381786, + 0.004601954482495785, + 0.012850255705416203, + 0.011155042797327042, + 0.004759326111525297, + 0.013208180665969849, + 0.005927134770900011, + 0.008972239680588245, + 0.004240612033754587, + 0.0034528281539678574, + 0.007007374428212643, + 0.004575067665427923, + 0.002419690601527691, + 0.006578852888196707, + 0.0031654140911996365, + 0.004914018325507641, + 0.007033579982817173, + 0.00518630538135767, + 0.009005601517856121, + 0.008542102761566639, + 0.003545638406649232, + 0.010792995803058147, + 0.005994156468659639, + 0.007868479937314987, + 0.030038315802812576, + 0.023925073444843292, + 0.04633888974785805, + 0.055378254503011703, + 0.031151123344898224, + 0.03664291650056839, + 0.03969702869653702, + 0.02925449050962925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.00956795271486044 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979288250207901 + ], + "opposite_logprob_tokens": [ + -3.5297927856445312 + ], + "logprob_reward": -0.02979288250207901, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979288250207901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28100642561912537, + 0.14259140193462372, + 0.09037522971630096, + 0.06135892495512962, + 0.05143267288804054, + 0.06347216665744781, + 0.03294287994503975, + 0.05262623727321625, + 0.052985213696956635, + 0.171208918094635 + ], + "credit_map": [ + 0.0371389240026474, + 0.024159105494618416, + 0.041096609085798264, + 0.03962239995598793, + 0.01957741752266884, + 0.05794348195195198, + 0.021006012335419655, + 0.0404624305665493, + 0.020585276186466217, + 0.015963951125741005, + 0.023614123463630676, + 0.020747514441609383, + 0.009927622973918915, + 0.022542964667081833, + 0.011516615748405457, + 0.017693322151899338, + 0.01272212527692318, + 0.009342927485704422, + 0.016797564923763275, + 0.01205411460250616, + 0.005346214864403009, + 0.016089098528027534, + 0.006891900207847357, + 0.011131277307868004, + 0.00849140901118517, + 0.007359408773481846, + 0.01068073883652687, + 0.00796100590378046, + 0.004140570294111967, + 0.010912584140896797, + 0.004786161705851555, + 0.007027040235698223, + 0.00603856286033988, + 0.004386944230645895, + 0.008782203309237957, + 0.006321062333881855, + 0.0036326278932392597, + 0.010451198555529118, + 0.005180473439395428, + 0.006639596540480852, + 0.007306638639420271, + 0.00667570112273097, + 0.01001769583672285, + 0.010872590355575085, + 0.005075735040009022, + 0.009546259418129921, + 0.005325603764504194, + 0.008651934564113617, + 0.003645604941993952, + 0.003346948651596904, + 0.005735509563237429, + 0.006056363694369793, + 0.0017447860445827246, + 0.006233213469386101, + 0.002331041730940342, + 0.0038494106847792864, + 0.006442096084356308, + 0.004801962058991194, + 0.008287442848086357, + 0.008786260150372982, + 0.0035588282626122236, + 0.009832832962274551, + 0.0041449652053415775, + 0.006771848537027836, + 0.006220449693500996, + 0.007721601519733667, + 0.007199798244982958, + 0.008190318942070007, + 0.005324885714799166, + 0.007071528118103743, + 0.005915623623877764, + 0.005341005977243185, + 0.02089664153754711, + 0.014420890249311924, + 0.02611076645553112, + 0.029719248414039612, + 0.01791447214782238, + 0.027654599398374557, + 0.01627691462635994, + 0.018215373158454895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.02979288250207901 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777866303920746 + ], + "opposite_logprob_tokens": [ + -4.634778022766113 + ], + "logprob_reward": -0.009777866303920746, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777866303920746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26252812147140503, + 0.10363329946994781, + 0.07885147631168365, + 0.04586642235517502, + 0.08311689645051956, + 0.042178086936473846, + 0.0728970617055893, + 0.05043921619653702, + 0.06141629070043564, + 0.19907307624816895 + ], + "credit_map": [ + 0.03545833379030228, + 0.018117457628250122, + 0.04232848808169365, + 0.04104054346680641, + 0.01722860336303711, + 0.05121193453669548, + 0.02253221534192562, + 0.03461053967475891, + 0.016798127442598343, + 0.009534874930977821, + 0.015762990340590477, + 0.01185885351151228, + 0.006625954527407885, + 0.01704300381243229, + 0.009205775335431099, + 0.01680372655391693, + 0.011373110115528107, + 0.006973880808800459, + 0.01295204646885395, + 0.00951462797820568, + 0.0049621122889220715, + 0.015078065916895866, + 0.006128130946308374, + 0.011869499459862709, + 0.00535660469904542, + 0.0037705989088863134, + 0.00827041082084179, + 0.007896585389971733, + 0.0029485083650797606, + 0.00781713705509901, + 0.0039525520987808704, + 0.0058540222235023975, + 0.011129540391266346, + 0.005463274661451578, + 0.015438102185726166, + 0.007501308340579271, + 0.005741937551647425, + 0.019906479865312576, + 0.005621875170618296, + 0.0123143857344985, + 0.005935234017670155, + 0.0030429272446781397, + 0.008064445108175278, + 0.004661012440919876, + 0.0028691214974969625, + 0.008937220089137554, + 0.002530409023165703, + 0.0061377184465527534, + 0.007912114262580872, + 0.005705756600946188, + 0.01325842086225748, + 0.011945129372179508, + 0.004779350012540817, + 0.014787022024393082, + 0.005112869665026665, + 0.009396406821906567, + 0.0060935253277421, + 0.004381160717457533, + 0.009851562790572643, + 0.005008822772651911, + 0.003079682821407914, + 0.010382227599620819, + 0.003671448677778244, + 0.007970785722136497, + 0.006339130457490683, + 0.006949083413928747, + 0.009043307974934578, + 0.012215042486786842, + 0.0038158094976097345, + 0.008522790856659412, + 0.007276149932295084, + 0.0072549753822386265, + 0.017975397408008575, + 0.017264528200030327, + 0.030603419989347458, + 0.037216976284980774, + 0.021061468869447708, + 0.03229285776615143, + 0.02328203245997429, + 0.019376402720808983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.009777866303920746 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038094088435173035 + ], + "opposite_logprob_tokens": [ + -3.2880940437316895 + ], + "logprob_reward": -0.038094088435173035, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038094088435173035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19638755917549133, + 0.09164013713598251, + 0.039830584079027176, + 0.04477870091795921, + 0.059448741376399994, + 0.062003813683986664, + 0.11262856423854828, + 0.060535233467817307, + 0.053096845746040344, + 0.2796497941017151 + ], + "credit_map": [ + 0.025314640253782272, + 0.015353713184595108, + 0.03233140707015991, + 0.035267423838377, + 0.01463435310870409, + 0.033501170575618744, + 0.01740032620728016, + 0.022584522143006325, + 0.013387374579906464, + 0.009228549897670746, + 0.015230483375489712, + 0.009300202131271362, + 0.007511831820011139, + 0.01684846729040146, + 0.007415281143039465, + 0.01271794643253088, + 0.004964177031069994, + 0.005304930731654167, + 0.006971963681280613, + 0.006493556313216686, + 0.003459520637989044, + 0.004682677332311869, + 0.0033219114411622286, + 0.004631848074495792, + 0.00564325088635087, + 0.005311888642609119, + 0.007116056978702545, + 0.0064896224066615105, + 0.0031183045357465744, + 0.006430314853787422, + 0.0039004830177873373, + 0.0067687793634831905, + 0.007345259655267, + 0.004694913513958454, + 0.01007529441267252, + 0.00987295527011156, + 0.003737095510587096, + 0.011213588528335094, + 0.0038523266557604074, + 0.008657305501401424, + 0.00799358356744051, + 0.003604241879656911, + 0.010503637604415417, + 0.006765109486877918, + 0.0046790409833192825, + 0.014077631756663322, + 0.005421202164143324, + 0.008959367871284485, + 0.012790758162736893, + 0.011138015426695347, + 0.020676927641034126, + 0.01375677902251482, + 0.0076642888598144054, + 0.024307293817400932, + 0.008203226141631603, + 0.01409127190709114, + 0.007602796424180269, + 0.006407785229384899, + 0.009806515648961067, + 0.008270296268165112, + 0.004846792668104172, + 0.008804048411548138, + 0.00559991504997015, + 0.009197083301842213, + 0.006215190049260855, + 0.005089243408292532, + 0.008828645572066307, + 0.007176058366894722, + 0.003636928042396903, + 0.008822447620332241, + 0.006771926302462816, + 0.006556408945471048, + 0.0365186408162117, + 0.024991219863295555, + 0.03703100234270096, + 0.04316570237278938, + 0.041025999933481216, + 0.042358119040727615, + 0.02454490028321743, + 0.03001423180103302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.038094088435173035 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a venomous monitor lizard perched on a dark, gnarled branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lizard's head colored with vibrant blood-ruby scales and its body with obsidian-black scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lizard's eye a malevolent yellow color visible within the shadows of its forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lizard's head turn smoothly and deliberately while its black forked tongue darts out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the dramatic Todd McFarlane art style with a harsh key light and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046363710425794125 + ], + "opposite_logprob_tokens": [ + -5.379636287689209 + ], + "logprob_reward": -0.0046363710425794125, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046363710425794125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2860852777957916, + 0.09880831092596054, + 0.06765905022621155, + 0.0457860492169857, + 0.04716130718588829, + 0.08188599348068237, + 0.07984962314367294, + 0.04564320296049118, + 0.050772495567798615, + 0.19634868204593658 + ], + "credit_map": [ + 0.03183218091726303, + 0.023786135017871857, + 0.04418320953845978, + 0.05151898041367531, + 0.021250730380415916, + 0.05134033411741257, + 0.02452481910586357, + 0.03764887899160385, + 0.012698723934590816, + 0.009733094833791256, + 0.01626805029809475, + 0.01625552587211132, + 0.006465711165219545, + 0.016019141301512718, + 0.008761774748563766, + 0.012606296688318253, + 0.009284249506890774, + 0.005220967344939709, + 0.010293624363839626, + 0.010876206681132317, + 0.004486727993935347, + 0.01322621013969183, + 0.004583219066262245, + 0.009687845595180988, + 0.005933120846748352, + 0.003995493054389954, + 0.007366500329226255, + 0.006987783592194319, + 0.002806338481605053, + 0.009401318617165089, + 0.0030331516172736883, + 0.006262346636503935, + 0.005850133486092091, + 0.005278893746435642, + 0.008692029863595963, + 0.005548457149416208, + 0.0031482211779803038, + 0.009521737694740295, + 0.003433158388361335, + 0.005688678007572889, + 0.009600860066711903, + 0.004502629861235619, + 0.013871940784156322, + 0.010614143684506416, + 0.0056427824310958385, + 0.02043919637799263, + 0.006617097184062004, + 0.010597337037324905, + 0.01169576309621334, + 0.006194080226123333, + 0.01356667559593916, + 0.009722582064568996, + 0.004686379339545965, + 0.01648685336112976, + 0.0047967457212507725, + 0.01270054467022419, + 0.005365320947021246, + 0.005198757629841566, + 0.00784290861338377, + 0.00751533592119813, + 0.0027227262035012245, + 0.007207327056676149, + 0.0037524693179875612, + 0.006038358435034752, + 0.0053183273412287235, + 0.005978491622954607, + 0.007183889392763376, + 0.00934693869203329, + 0.0038982657715678215, + 0.006912202574312687, + 0.005802875384688377, + 0.006331503391265869, + 0.023749900981783867, + 0.017957696691155434, + 0.02613808959722519, + 0.03219624236226082, + 0.03324083983898163, + 0.02349049597978592, + 0.020074831321835518, + 0.019500574097037315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0046363710425794125 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_4/index_tensor([8762, 8762])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_4/index_tensor([8762, 8762])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..06fb378747524e768cd481c2c2a9129f9355f238 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_4/index_tensor([8762, 8762])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, tight close-up centers on the face of a young, anthropomorphic panda monk, his entire body radiating palpable frustration. His black and white fur, rendered in a complex 3D style reminiscent of the \"Two Cities\" series, boasts rich painterly textures, while sharp, expressive, glass-shaded outlines define his form. His forehead is deeply furrowed, casting harsh, graphic shadows over his eyes, which narrow to a furious gleam, trained on a failed potion off-screen. His jaw clenches hard, his teeth visibly set, and his lips draw back slightly in a silent roar of intense annoyance. The faint, sickly purple glow of the unseen potion casts a dramatic backlight on his nose and chin, contrasting starkly with the warm ambient light of the dimly lit stone chamber in the background. Every muscle in his furry face is tense, a mask of concentrated annoyance that captures the fleeting moment of alchemical failure with profound emotional weight and a unique, hand-painted aesthetic.", + "seed": 9998, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_4/index_8762_seed_9998_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02353017353452742, + "vlm_logprob": -0.02353017353452742, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.02353017353452742, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03234958276152611 + ], + "opposite_logprob_tokens": [ + -3.5323495864868164 + ], + "logprob_reward": -0.03234958276152611, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03234958276152611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08486314862966537, + 0.045154277235269547, + 0.11222229152917862, + 0.10380241274833679, + 0.1783847212791443, + 0.10006651282310486, + 0.09977981448173523, + 0.20182789862155914, + 0.03518744558095932, + 0.038711532950401306 + ], + "credit_map": [ + 0.007948060519993305, + 0.005018400959670544, + 0.0087078046053648, + 0.009793894365429878, + 0.00950693804770708, + 0.007133226376026869, + 0.025747232139110565, + 0.011007582768797874, + 0.004507936537265778, + 0.004113050177693367, + 0.007037302013486624, + 0.005772459786385298, + 0.00377242942340672, + 0.0035568310413509607, + 0.011740472167730331, + 0.004653793293982744, + 0.010688996873795986, + 0.013068588450551033, + 0.016582198441028595, + 0.011536271311342716, + 0.015552339144051075, + 0.014781362377107143, + 0.013240594416856766, + 0.016771940514445305, + 0.02449030615389347, + 0.01217866875231266, + 0.027409980073571205, + 0.010439818724989891, + 0.0061174738220870495, + 0.00929091963917017, + 0.0052529736422002316, + 0.008622267283499241, + 0.019110456109046936, + 0.07330308854579926, + 0.023642417043447495, + 0.014343337155878544, + 0.015524792484939098, + 0.015883788466453552, + 0.005902138538658619, + 0.010674691759049892, + 0.010168975219130516, + 0.030059782788157463, + 0.010936422273516655, + 0.007302012760192156, + 0.013299225829541683, + 0.010923483408987522, + 0.011064180172979832, + 0.006312420126050711, + 0.006002369802445173, + 0.007639138959348202, + 0.006072440650314093, + 0.010468604974448681, + 0.008364773355424404, + 0.01836344599723816, + 0.03220190480351448, + 0.010667127557098866, + 0.01843186654150486, + 0.015746811404824257, + 0.040800876915454865, + 0.0540272556245327, + 0.018075942993164062, + 0.013088052161037922, + 0.028480183333158493, + 0.013176897540688515, + 0.004610609728842974, + 0.004603907465934753, + 0.0059561701491475105, + 0.0034067677333950996, + 0.005408312659710646, + 0.0044425115920603275, + 0.004115962889045477, + 0.0026432019658386707, + 0.007005638442933559, + 0.0032144845463335514, + 0.00485825864598155, + 0.003679972607642412, + 0.0029134678188711405, + 0.002666223794221878, + 0.009897206909954548, + 0.004476275760680437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.03234958276152611 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00322548346593976 + ], + "opposite_logprob_tokens": [ + -5.753225326538086 + ], + "logprob_reward": -0.00322548346593976, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00322548346593976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14749087393283844, + 0.08887699246406555, + 0.08877386897802353, + 0.15984292328357697, + 0.15290583670139313, + 0.03724683076143265, + 0.18892797827720642, + 0.08175736665725708, + 0.03621905669569969, + 0.017958227545022964 + ], + "credit_map": [ + 0.00775934336706996, + 0.008871177211403847, + 0.02060423605144024, + 0.03147902712225914, + 0.013060685247182846, + 0.012180810794234276, + 0.039832089096307755, + 0.013703516684472561, + 0.012048549018800259, + 0.006119119934737682, + 0.02619258686900139, + 0.008090224117040634, + 0.006125432439148426, + 0.005947747267782688, + 0.014267034828662872, + 0.010086300782859325, + 0.013069333508610725, + 0.0094536067917943, + 0.0087687186896801, + 0.007008138112723827, + 0.010841509327292442, + 0.01912231184542179, + 0.007286266423761845, + 0.013223990797996521, + 0.04712584242224693, + 0.03183821588754654, + 0.028946178033947945, + 0.01364409551024437, + 0.011850130744278431, + 0.00960079301148653, + 0.007870514877140522, + 0.008967174217104912, + 0.02876306138932705, + 0.06044914573431015, + 0.012922775000333786, + 0.012005224823951721, + 0.012741969898343086, + 0.014277396723628044, + 0.005249100271612406, + 0.0064971731044352055, + 0.0035377712920308113, + 0.006106252782046795, + 0.0028363496530801058, + 0.0036845607683062553, + 0.006013399455696344, + 0.007482386659830809, + 0.004386590793728828, + 0.003199523314833641, + 0.010688761249184608, + 0.021052174270153046, + 0.017616864293813705, + 0.02334946021437645, + 0.019501574337482452, + 0.013226505368947983, + 0.05879780650138855, + 0.024694837629795074, + 0.005362056661397219, + 0.005655121989548206, + 0.011145965196192265, + 0.028317024931311607, + 0.004783059470355511, + 0.004503224045038223, + 0.017820531502366066, + 0.0041703879833221436, + 0.0038771205581724644, + 0.006544729229062796, + 0.003643406555056572, + 0.004718473181128502, + 0.004976850003004074, + 0.004492307547479868, + 0.00451006880030036, + 0.00345610361546278, + 0.0028858098667114973, + 0.0022062070202082396, + 0.003265578532591462, + 0.002140005584806204, + 0.0022162501700222492, + 0.001417732099071145, + 0.002203615615144372, + 0.0016230292385444045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.00322548346593976 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036193120758980513 + ], + "opposite_logprob_tokens": [ + -5.628619194030762 + ], + "logprob_reward": -0.0036193120758980513, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036193120758980513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15611481666564941, + 0.03560836240649223, + 0.3019406199455261, + 0.17777279019355774, + 0.08544173836708069, + 0.04557725787162781, + 0.05869690328836441, + 0.054635852575302124, + 0.04312708601355553, + 0.041084617376327515 + ], + "credit_map": [ + 0.008406045846641064, + 0.009192262776196003, + 0.01996280439198017, + 0.03785191848874092, + 0.010253574699163437, + 0.009791748598217964, + 0.04540274292230606, + 0.015253705903887749, + 0.004350957926362753, + 0.0034589667338877916, + 0.010230205953121185, + 0.004043285734951496, + 0.0031656043138355017, + 0.0034168201964348555, + 0.0031781974248588085, + 0.003764324588701129, + 0.03168347850441933, + 0.054705604910850525, + 0.028613194823265076, + 0.030962547287344933, + 0.05484494939446449, + 0.03373533487319946, + 0.03356628865003586, + 0.0338292121887207, + 0.03445412591099739, + 0.05802818760275841, + 0.015790322795510292, + 0.016638660803437233, + 0.015117263421416283, + 0.017469018697738647, + 0.00837628822773695, + 0.011898912489414215, + 0.007458571344614029, + 0.03234931454062462, + 0.014232856221497059, + 0.006743855774402618, + 0.00754048814997077, + 0.007251373026520014, + 0.0037954470608383417, + 0.006069835741072893, + 0.006494647823274136, + 0.005372136831283569, + 0.0037276886869221926, + 0.005250672809779644, + 0.006831199862062931, + 0.008185486309230328, + 0.0050142863765358925, + 0.004701136611402035, + 0.004692199174314737, + 0.0036876460071653128, + 0.003606328973546624, + 0.0033258122857660055, + 0.003433028468862176, + 0.020573481917381287, + 0.013182920403778553, + 0.006195487454533577, + 0.005173174664378166, + 0.005006683059036732, + 0.007286532316356897, + 0.013083263300359249, + 0.0071140495128929615, + 0.004308138974010944, + 0.008151649497449398, + 0.0045123593881726265, + 0.004226892255246639, + 0.0052927276119589806, + 0.00922232773154974, + 0.005497795529663563, + 0.006066875532269478, + 0.0033352829050272703, + 0.005329501815140247, + 0.004155680537223816, + 0.006062920205295086, + 0.004803636111319065, + 0.008561801165342331, + 0.004730276297777891, + 0.0051678880117833614, + 0.004207881633192301, + 0.0034897837322205305, + 0.004060428589582443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.0036193120758980513 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865897819399834 + ], + "opposite_logprob_tokens": [ + -3.048659086227417 + ], + "logprob_reward": -0.04865897819399834, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865897819399834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30383995175361633, + 0.07200851291418076, + 0.13263154029846191, + 0.03720151260495186, + 0.13608567416667938, + 0.03861960396170616, + 0.10704495757818222, + 0.09815850853919983, + 0.0451110303401947, + 0.02929869294166565 + ], + "credit_map": [ + 0.009402538649737835, + 0.01716473512351513, + 0.03629811480641365, + 0.09895429015159607, + 0.017419956624507904, + 0.01317036896944046, + 0.09504151344299316, + 0.016388434916734695, + 0.006167037878185511, + 0.005552589427679777, + 0.008468546904623508, + 0.011284824460744858, + 0.0072546047158539295, + 0.007585839368402958, + 0.016675882041454315, + 0.009019186720252037, + 0.014186418615281582, + 0.01971781812608242, + 0.015121273696422577, + 0.015039329417049885, + 0.016568103805184364, + 0.024822795763611794, + 0.012319039553403854, + 0.014856762252748013, + 0.003043291624635458, + 0.0053615025244653225, + 0.007662596646696329, + 0.005058612674474716, + 0.004481729585677385, + 0.003042582655325532, + 0.003752901451662183, + 0.004798297304660082, + 0.019683217629790306, + 0.056263335049152374, + 0.013465710915625095, + 0.010982352308928967, + 0.011919139884412289, + 0.012512559071183205, + 0.004562940448522568, + 0.006696412339806557, + 0.0038518481887876987, + 0.005364385433495045, + 0.0029373117722570896, + 0.004218669142574072, + 0.008592196740210056, + 0.005055374465882778, + 0.005144371651113033, + 0.003455445868894458, + 0.010977281257510185, + 0.006376674398779869, + 0.006077020429074764, + 0.012594313360750675, + 0.010577770881354809, + 0.010697225108742714, + 0.03192512318491936, + 0.01781955361366272, + 0.01789962686598301, + 0.0052657839842140675, + 0.01517957728356123, + 0.02380109764635563, + 0.0066415611654520035, + 0.006082281935960054, + 0.011816534213721752, + 0.011472043581306934, + 0.005682956427335739, + 0.005160896573215723, + 0.00739116407930851, + 0.0036158415023237467, + 0.007542181760072708, + 0.0049950131215155125, + 0.006502014584839344, + 0.004220963455736637, + 0.003683237126097083, + 0.0029740959871560335, + 0.002494335640221834, + 0.0033977890852838755, + 0.002349252114072442, + 0.002496151952072978, + 0.007097106426954269, + 0.004806723445653915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.04865897819399834 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979751117527485 + ], + "opposite_logprob_tokens": [ + -3.5297975540161133 + ], + "logprob_reward": -0.02979751117527485, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979751117527485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10581011325120926, + 0.03919145464897156, + 0.14376981556415558, + 0.15184786915779114, + 0.17994599044322968, + 0.03961537778377533, + 0.019267648458480835, + 0.1378760188817978, + 0.11796952039003372, + 0.06470617651939392 + ], + "credit_map": [ + 0.00766185624524951, + 0.00692497193813324, + 0.021060479804873466, + 0.012915000319480896, + 0.011429275386035442, + 0.008117259480059147, + 0.026519963517785072, + 0.011181303299963474, + 0.0060799382627010345, + 0.002945872023701668, + 0.012081083841621876, + 0.003469388000667095, + 0.0029476259369403124, + 0.0023138734977692366, + 0.005723997019231319, + 0.003629677463322878, + 0.02077755518257618, + 0.009905568324029446, + 0.012860758230090141, + 0.010739673860371113, + 0.011339494027197361, + 0.03851860761642456, + 0.01070654671639204, + 0.02892162650823593, + 0.022164592519402504, + 0.0453120656311512, + 0.013986818492412567, + 0.01477116160094738, + 0.013405125588178635, + 0.01809770055115223, + 0.008105607703328133, + 0.016004787757992744, + 0.017055215314030647, + 0.07266786694526672, + 0.026068557053804398, + 0.014354798942804337, + 0.01575218141078949, + 0.015938078984618187, + 0.006223161239176989, + 0.01188612636178732, + 0.004828508943319321, + 0.009772755205631256, + 0.0041137817315757275, + 0.003423375776037574, + 0.0052958265878260136, + 0.004254139959812164, + 0.00462484871968627, + 0.0033021445851773024, + 0.0018119236920028925, + 0.00259106932207942, + 0.0024898338597267866, + 0.0021773027256131172, + 0.002739417366683483, + 0.0027014072984457016, + 0.0029234401881694794, + 0.0018332520266994834, + 0.009293588809669018, + 0.006971781142055988, + 0.021096106618642807, + 0.049674347043037415, + 0.007401803974062204, + 0.006503818556666374, + 0.029085956513881683, + 0.007848605513572693, + 0.01429685764014721, + 0.01632349193096161, + 0.015010219998657703, + 0.010828227736055851, + 0.02190254256129265, + 0.015824919566512108, + 0.015050960704684258, + 0.008732309564948082, + 0.007410730235278606, + 0.0038822293281555176, + 0.012597417458891869, + 0.007757720537483692, + 0.004953667055815458, + 0.004330512601882219, + 0.013577288947999477, + 0.010196608491241932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.02979751117527485 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tight close-up centers on the face of a young, anthropomorphic panda monk, his entire body radiating palpable frustration. His black and white fur, rendered in a complex 3D style reminiscent of the \"Two Cities\" series, boasts rich painterly textures, while sharp, expressive, glass-shaded outlines define his form. His forehead is deeply furrowed, casting harsh, graphic shadows over his eyes, which narrow to a furious gleam, trained on a failed potion off-screen. His jaw clenches hard, his teeth visibly set, and his lips draw back slightly in a silent roar of intense annoyance. The faint, sickly purple glow of the unseen potion casts a dramatic backlight on his nose and chin, contrasting starkly with the warm ambient light of the dimly lit stone chamber in the background. Every muscle in his furry face is tense, a mask of concentrated annoyance that captures the fleeting moment of alchemical failure with profound emotional weight and a unique, hand-painted aesthetic.", + "seed": 9999, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_4/index_8762_seed_9999_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03370250056032091, + "vlm_logprob": -0.03370250056032091, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.03370250056032091, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07198619842529297 + ], + "opposite_logprob_tokens": [ + -2.696986198425293 + ], + "logprob_reward": -0.07198619842529297, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07198619842529297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18888507783412933, + 0.0358135811984539, + 0.03941531479358673, + 0.06727400422096252, + 0.1367405652999878, + 0.10513337701559067, + 0.17089694738388062, + 0.0288125891238451, + 0.07305537164211273, + 0.15397322177886963 + ], + "credit_map": [ + 0.025633657351136208, + 0.017673328518867493, + 0.019536461681127548, + 0.022731872275471687, + 0.028937317430973053, + 0.029112784191966057, + 0.02168414369225502, + 0.02357550896704197, + 0.0035810344852507114, + 0.0045002298429608345, + 0.003384581534191966, + 0.0063283867202699184, + 0.0030630233231931925, + 0.004550746642053127, + 0.006181757897138596, + 0.004223818890750408, + 0.008135554380714893, + 0.005147507879883051, + 0.004973266739398241, + 0.004415072966367006, + 0.0036336921621114016, + 0.003701384412124753, + 0.003579690819606185, + 0.005829146597534418, + 0.01411556638777256, + 0.0066354949958622456, + 0.009810689836740494, + 0.006774933077394962, + 0.008413457311689854, + 0.006585423834621906, + 0.0065074521116912365, + 0.008430983871221542, + 0.022156761959195137, + 0.011103667318820953, + 0.022631775587797165, + 0.018614565953612328, + 0.014361880719661713, + 0.011942804791033268, + 0.012869794853031635, + 0.023059308528900146, + 0.014638622291386127, + 0.010639740154147148, + 0.008786004967987537, + 0.020274648442864418, + 0.010964510031044483, + 0.012509256601333618, + 0.012454402633011341, + 0.014866200275719166, + 0.011881270445883274, + 0.005189971067011356, + 0.035429127514362335, + 0.011468816548585892, + 0.00749413575977087, + 0.011751674115657806, + 0.02260585129261017, + 0.06507610529661179, + 0.0026741339825093746, + 0.001993391662836075, + 0.004286253359168768, + 0.0034695893991738558, + 0.0035785913933068514, + 0.003269145730882883, + 0.0033785852137953043, + 0.006162900011986494, + 0.011974803172051907, + 0.008821824565529823, + 0.006841923110187054, + 0.008161568082869053, + 0.010020523332059383, + 0.010601218789815903, + 0.007366932462900877, + 0.009266573004424572, + 0.0051370784640312195, + 0.003385753370821476, + 0.006586016621440649, + 0.003926679491996765, + 0.034985706210136414, + 0.0650622546672821, + 0.006021140143275261, + 0.028868593275547028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.07198619842529297 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00152757263276726 + ], + "opposite_logprob_tokens": [ + -6.501527786254883 + ], + "logprob_reward": -0.00152757263276726, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.00152757263276726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10454502701759338, + 0.02225629612803459, + 0.05209392309188843, + 0.09676419943571091, + 0.06933409720659256, + 0.057812027633190155, + 0.26181045174598694, + 0.21050022542476654, + 0.04119586572051048, + 0.08368795365095139 + ], + "credit_map": [ + 0.01798139326274395, + 0.00895785540342331, + 0.018184322863817215, + 0.011946363374590874, + 0.010370714589953423, + 0.01138307061046362, + 0.010663546621799469, + 0.015057744458317757, + 0.0028542487416416407, + 0.0032472372986376286, + 0.002831165213137865, + 0.001969136530533433, + 0.0019169457955285907, + 0.0026732569094747305, + 0.002978963777422905, + 0.003785338718444109, + 0.012513302266597748, + 0.005769872572273016, + 0.007877402007579803, + 0.008453668095171452, + 0.003690351964905858, + 0.0034885017666965723, + 0.004189436789602041, + 0.006111384369432926, + 0.01811179518699646, + 0.006481256801635027, + 0.015482591465115547, + 0.01416723895817995, + 0.014841483905911446, + 0.0067171016708016396, + 0.007440929301083088, + 0.013521792367100716, + 0.01164085790514946, + 0.004344167187809944, + 0.00743699399754405, + 0.011862857267260551, + 0.007467486895620823, + 0.006416689604520798, + 0.008723457343876362, + 0.011441581882536411, + 0.010176340118050575, + 0.004262740723788738, + 0.005860867444425821, + 0.00917770341038704, + 0.0070670717395842075, + 0.0059572141617536545, + 0.006221019197255373, + 0.00908906664699316, + 0.01829303614795208, + 0.007073776330798864, + 0.06665384024381638, + 0.0181393064558506, + 0.012814355082809925, + 0.015209665521979332, + 0.021172957494854927, + 0.10245350748300552, + 0.019037898629903793, + 0.00847214087843895, + 0.04884561151266098, + 0.023821409791707993, + 0.010972053743898869, + 0.012527582235634327, + 0.015660712495446205, + 0.07116279006004333, + 0.004519545007497072, + 0.004209026228636503, + 0.004055040422827005, + 0.005041510332375765, + 0.005655535496771336, + 0.005231400951743126, + 0.00660334387794137, + 0.005880463402718306, + 0.007715190760791302, + 0.003174036042764783, + 0.018761025741696358, + 0.0062016830779612064, + 0.011648200452327728, + 0.02138247899711132, + 0.0040720137767493725, + 0.010733314789831638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.00152757263276726 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017131189815700054 + ], + "opposite_logprob_tokens": [ + -6.376713275909424 + ], + "logprob_reward": -0.0017131189815700054, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017131189815700054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12048860639333725, + 0.05831613019108772, + 0.0735870897769928, + 0.04889707639813423, + 0.08739063888788223, + 0.07521064579486847, + 0.21328243613243103, + 0.045282408595085144, + 0.03202517703175545, + 0.24551981687545776 + ], + "credit_map": [ + 0.023962784558534622, + 0.009864219464361668, + 0.020932890474796295, + 0.011226644739508629, + 0.011535529978573322, + 0.012333775870501995, + 0.01205776073038578, + 0.018575003370642662, + 0.007191562559455633, + 0.0074235303327441216, + 0.008141153492033482, + 0.006648722570389509, + 0.006724013946950436, + 0.005931429099291563, + 0.009310952387750149, + 0.006944769062101841, + 0.0130523182451725, + 0.00740360002964735, + 0.011266373097896576, + 0.007719415705651045, + 0.008848614059388638, + 0.007161581423133612, + 0.006566791329532862, + 0.011568400077521801, + 0.007348667364567518, + 0.00531376525759697, + 0.004978475626558065, + 0.00474644685164094, + 0.00817381776869297, + 0.004868739750236273, + 0.004881849978119135, + 0.008585312403738499, + 0.015709375962615013, + 0.00704193813726306, + 0.012202165089547634, + 0.011658145114779472, + 0.008587992750108242, + 0.007607501465827227, + 0.011086694896221161, + 0.013496823608875275, + 0.015464494936168194, + 0.0057257008738815784, + 0.009442312642931938, + 0.012064103968441486, + 0.008230009116232395, + 0.006023171357810497, + 0.008259641006588936, + 0.010001207701861858, + 0.013451701030135155, + 0.005171309690922499, + 0.051217272877693176, + 0.013432643376290798, + 0.010063414461910725, + 0.012281407602131367, + 0.02094120718538761, + 0.08672346919775009, + 0.0050553223118186, + 0.0026466213166713715, + 0.008012254722416401, + 0.007180540356785059, + 0.003939246293157339, + 0.003529550973325968, + 0.005114228930324316, + 0.009804640896618366, + 0.002316124504432082, + 0.0036632828414440155, + 0.0027040764689445496, + 0.0032964120618999004, + 0.004693587776273489, + 0.004654030315577984, + 0.0037866125348955393, + 0.006911048199981451, + 0.00945296511054039, + 0.005653675179928541, + 0.012074090540409088, + 0.0080108055844903, + 0.04758250713348389, + 0.10232747346162796, + 0.009635339491069317, + 0.05078296735882759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.0017131189815700054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023279689252376556 + ], + "opposite_logprob_tokens": [ + -3.7732796669006348 + ], + "logprob_reward": -0.023279689252376556, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023279689252376556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10740169882774353, + 0.031232232227921486, + 0.022126374766230583, + 0.15607328712940216, + 0.09484706073999405, + 0.04884197935461998, + 0.3070306181907654, + 0.05682392418384552, + 0.08014723658561707, + 0.09547560662031174 + ], + "credit_map": [ + 0.015634795650839806, + 0.010073232464492321, + 0.01773301512002945, + 0.010845731012523174, + 0.010808720253407955, + 0.01087815873324871, + 0.012661033309996128, + 0.018767019733786583, + 0.0044269924983382225, + 0.0024118302389979362, + 0.003634093329310417, + 0.0038352825213223696, + 0.0035637374967336655, + 0.005066493526101112, + 0.003936119843274355, + 0.004357682075351477, + 0.0026468122377991676, + 0.0022532260045409203, + 0.002812072401866317, + 0.003230195725336671, + 0.002119034295901656, + 0.0021374982316046953, + 0.0030318009667098522, + 0.003895736299455166, + 0.05270591378211975, + 0.01017508003860712, + 0.023494431748986244, + 0.018374867737293243, + 0.01573895290493965, + 0.009582613594830036, + 0.009240616112947464, + 0.016760803759098053, + 0.02917775698006153, + 0.005936114117503166, + 0.015519182197749615, + 0.013016177341341972, + 0.008532516658306122, + 0.0060530537739396095, + 0.005533228628337383, + 0.011079024523496628, + 0.008930874057114124, + 0.004389825742691755, + 0.005055386107414961, + 0.009359955787658691, + 0.005979962646961212, + 0.004492353647947311, + 0.005515682976692915, + 0.005117941182106733, + 0.017890458926558495, + 0.006761422846466303, + 0.07304094731807709, + 0.01751459203660488, + 0.014148528687655926, + 0.017955485731363297, + 0.03185732662677765, + 0.1278618723154068, + 0.004182637203484774, + 0.0016042198985815048, + 0.013443364761769772, + 0.0034199063666164875, + 0.002787624252960086, + 0.0037171293515712023, + 0.004541329573839903, + 0.023127714172005653, + 0.012830227613449097, + 0.005419370252639055, + 0.010982452891767025, + 0.005641216412186623, + 0.008511530235409737, + 0.0063360207714140415, + 0.0075540668331086636, + 0.022872351109981537, + 0.003462327178567648, + 0.0028002620674669743, + 0.003738241735845804, + 0.003986359108239412, + 0.02167496643960476, + 0.03653125837445259, + 0.003881160169839859, + 0.019401026889681816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.023279689252376556 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000592350959778 + ], + "opposite_logprob_tokens": [ + -2.6950058937072754 + ], + "logprob_reward": -0.07000592350959778, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000592350959778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13621734082698822, + 0.06886008381843567, + 0.07275785505771637, + 0.055398497730493546, + 0.2801086902618408, + 0.10690084844827652, + 0.0894874855875969, + 0.05277318134903908, + 0.07033684849739075, + 0.06715913861989975 + ], + "credit_map": [ + 0.025740135461091995, + 0.01263184193521738, + 0.018679648637771606, + 0.014107182621955872, + 0.015430580824613571, + 0.01371687650680542, + 0.01598551496863365, + 0.019925566390156746, + 0.007140819448977709, + 0.007027760148048401, + 0.010339384898543358, + 0.0060501014813780785, + 0.007764105219393969, + 0.012248408980667591, + 0.007263127248734236, + 0.011026372201740742, + 0.01172573957592249, + 0.006245478056371212, + 0.011828995309770107, + 0.010789528489112854, + 0.007057346403598785, + 0.005739031825214624, + 0.006807975471019745, + 0.012563755735754967, + 0.010550661012530327, + 0.005108370911329985, + 0.006762497127056122, + 0.007605532184243202, + 0.008203179575502872, + 0.004875882063060999, + 0.005916468799114227, + 0.006375907454639673, + 0.046332165598869324, + 0.022957105189561844, + 0.03355306386947632, + 0.05284380912780762, + 0.031163040548563004, + 0.022855767980217934, + 0.02828104794025421, + 0.042122695595026016, + 0.019188329577445984, + 0.007958482950925827, + 0.011931638233363628, + 0.017754999920725822, + 0.010840129107236862, + 0.009408457204699516, + 0.012004627846181393, + 0.01781417615711689, + 0.005495880264788866, + 0.002632939023897052, + 0.019540585577487946, + 0.005392080172896385, + 0.005055708345025778, + 0.0064212363213300705, + 0.009312577545642853, + 0.03563648462295532, + 0.003553106216713786, + 0.0028142977971583605, + 0.009184597991406918, + 0.004373165313154459, + 0.004522856790572405, + 0.004509457387030125, + 0.007255011238157749, + 0.016560690477490425, + 0.006401623133569956, + 0.005788461305201054, + 0.006664006970822811, + 0.008302118629217148, + 0.008751637302339077, + 0.008351381868124008, + 0.008599960245192051, + 0.01747766323387623, + 0.005195469129830599, + 0.0036570520605891943, + 0.005730218719691038, + 0.004373655188828707, + 0.004957485478371382, + 0.015058781020343304, + 0.004626829642802477, + 0.023559648543596268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.07000592350959778 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tight close-up centers on the face of a young, anthropomorphic panda monk, his entire body radiating palpable frustration. His black and white fur, rendered in a complex 3D style reminiscent of the \"Two Cities\" series, boasts rich painterly textures, while sharp, expressive, glass-shaded outlines define his form. His forehead is deeply furrowed, casting harsh, graphic shadows over his eyes, which narrow to a furious gleam, trained on a failed potion off-screen. His jaw clenches hard, his teeth visibly set, and his lips draw back slightly in a silent roar of intense annoyance. The faint, sickly purple glow of the unseen potion casts a dramatic backlight on his nose and chin, contrasting starkly with the warm ambient light of the dimly lit stone chamber in the background. Every muscle in his furry face is tense, a mask of concentrated annoyance that captures the fleeting moment of alchemical failure with profound emotional weight and a unique, hand-painted aesthetic.", + "seed": 10000, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_4/index_8762_seed_10000_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08419947465881705, + "vlm_logprob": -0.08419947465881705, + "vlm_yes_no_margin": 3.974999988079071, + "vlm_token_credit": -0.08419947465881705, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3518446087837219 + ], + "opposite_logprob_tokens": [ + -1.2268445491790771 + ], + "logprob_reward": -0.3518446087837219, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3518446087837219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35060223937034607, + 0.1783471703529358, + 0.037859439849853516, + 0.031280629336833954, + 0.04300064221024513, + 0.0386032909154892, + 0.18908122181892395, + 0.0251605361700058, + 0.029474567621946335, + 0.07659028470516205 + ], + "credit_map": [ + 0.02406342700123787, + 0.017726369202136993, + 0.025505661964416504, + 0.0196261927485466, + 0.032151445746421814, + 0.11454152315855026, + 0.03820931538939476, + 0.07877831161022186, + 0.014532026834785938, + 0.022361230105161667, + 0.010111532174050808, + 0.01595550961792469, + 0.017322096973657608, + 0.046090077608823776, + 0.019592106342315674, + 0.03238258883357048, + 0.005007485393434763, + 0.0027559695299714804, + 0.004719694145023823, + 0.0038655174430459738, + 0.003697873093187809, + 0.005004147067666054, + 0.005562702659517527, + 0.007246050052344799, + 0.005804198794066906, + 0.002649178495630622, + 0.0034925579093396664, + 0.0048410724848508835, + 0.0020168456248939037, + 0.004008327145129442, + 0.0037844586186110973, + 0.004683987703174353, + 0.00937844067811966, + 0.0027796742506325245, + 0.00538274273276329, + 0.0043927752412855625, + 0.005251467227935791, + 0.005492268595844507, + 0.005129487719386816, + 0.005193783901631832, + 0.006677049677819014, + 0.0028429534286260605, + 0.004411814268678427, + 0.004995056893676519, + 0.00293299020268023, + 0.005875394679605961, + 0.005109851714223623, + 0.005758182145655155, + 0.0318489596247673, + 0.013734852895140648, + 0.019113194197416306, + 0.028511973097920418, + 0.015063627623021603, + 0.02934461086988449, + 0.022790923714637756, + 0.028673088178038597, + 0.004423400852829218, + 0.0018497012788429856, + 0.002436501905322075, + 0.00400124303996563, + 0.002098013646900654, + 0.0032116356305778027, + 0.0040460689924657345, + 0.003093968378379941, + 0.005432957783341408, + 0.0023936249781399965, + 0.002671601716428995, + 0.003387944772839546, + 0.002008687471970916, + 0.005246582441031933, + 0.0034809696953743696, + 0.004852198995649815, + 0.008306670002639294, + 0.007183725479990244, + 0.009995724074542522, + 0.008928562514483929, + 0.006145624443888664, + 0.015087158419191837, + 0.011610133573412895, + 0.009332683868706226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.3518446087837219 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002230895683169365 + ], + "opposite_logprob_tokens": [ + -6.127231121063232 + ], + "logprob_reward": -0.002230895683169365, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002230895683169365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14904417097568512, + 0.10760315507650375, + 0.09937221556901932, + 0.06343203783035278, + 0.08327911049127579, + 0.13646450638771057, + 0.1486426293849945, + 0.06139729544520378, + 0.10106788575649261, + 0.04969688877463341 + ], + "credit_map": [ + 0.013585472479462624, + 0.012705579400062561, + 0.015314494259655476, + 0.01291261613368988, + 0.016085846349596977, + 0.03352924436330795, + 0.020063241943717003, + 0.02484768070280552, + 0.00743083143606782, + 0.009490550495684147, + 0.00654645636677742, + 0.012265502475202084, + 0.010316713713109493, + 0.028878269717097282, + 0.013767432421445847, + 0.018907394260168076, + 0.020133469253778458, + 0.0071309409104287624, + 0.018317032605409622, + 0.010935026220977306, + 0.009646434336900711, + 0.009161947295069695, + 0.010386092588305473, + 0.013661270029842854, + 0.006178244948387146, + 0.005946367979049683, + 0.006965347100049257, + 0.00631104176864028, + 0.006374991033226252, + 0.011665092781186104, + 0.010645070113241673, + 0.0093458853662014, + 0.017056060954928398, + 0.006614898331463337, + 0.00948327500373125, + 0.009191042743623257, + 0.006890749093145132, + 0.01341121643781662, + 0.010258521884679794, + 0.01037334743887186, + 0.0285430159419775, + 0.010023021139204502, + 0.016900671645998955, + 0.016081390902400017, + 0.008892878890037537, + 0.01956534944474697, + 0.014980854466557503, + 0.021477332338690758, + 0.030314454808831215, + 0.010054904036223888, + 0.015169021673500538, + 0.019567998126149178, + 0.010409456677734852, + 0.02325940504670143, + 0.015508491545915604, + 0.024358902126550674, + 0.013020869344472885, + 0.00537855364382267, + 0.007070150226354599, + 0.005160398315638304, + 0.0044763716869056225, + 0.010661587119102478, + 0.006083250045776367, + 0.009546119719743729, + 0.020869821310043335, + 0.007738019805401564, + 0.009663526900112629, + 0.012535049580037594, + 0.005623532459139824, + 0.01602163538336754, + 0.012075025588274002, + 0.016541272401809692, + 0.0050851451233029366, + 0.0035419180057942867, + 0.005573888774961233, + 0.006183004938066006, + 0.004373958799988031, + 0.009998765774071217, + 0.007478740531951189, + 0.007461466360837221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.002230895683169365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00248968368396163 + ], + "opposite_logprob_tokens": [ + -6.0024895668029785 + ], + "logprob_reward": -0.00248968368396163, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00248968368396163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21662750840187073, + 0.11390186101198196, + 0.11370783299207687, + 0.11841802299022675, + 0.05611129105091095, + 0.06007210910320282, + 0.10270951688289642, + 0.08199112862348557, + 0.05183015018701553, + 0.084630586206913 + ], + "credit_map": [ + 0.020679272711277008, + 0.016165824607014656, + 0.021224457770586014, + 0.016081472858786583, + 0.021603962406516075, + 0.05370193347334862, + 0.026619425043463707, + 0.04055115580558777, + 0.009634844958782196, + 0.01118040643632412, + 0.013570991344749928, + 0.015062973834574223, + 0.014787090010941029, + 0.011485411785542965, + 0.02471851371228695, + 0.013461630791425705, + 0.01950812339782715, + 0.00942174345254898, + 0.014928063377737999, + 0.012038921006023884, + 0.015359229408204556, + 0.012161701917648315, + 0.013690844178199768, + 0.016599204391241074, + 0.011354978196322918, + 0.010514318011701107, + 0.014020227827131748, + 0.012990101240575314, + 0.022253118455410004, + 0.014155865646898746, + 0.018869711086153984, + 0.014259705320000648, + 0.008203110657632351, + 0.004903199151158333, + 0.004308462142944336, + 0.004931677132844925, + 0.0062782904133200645, + 0.009966971352696419, + 0.009706197306513786, + 0.00781338196247816, + 0.01115494780242443, + 0.005552287213504314, + 0.006183062680065632, + 0.0057486919686198235, + 0.003519851714372635, + 0.00960375927388668, + 0.006395669188350439, + 0.011913840658962727, + 0.017098164185881615, + 0.008069136179983616, + 0.008948346599936485, + 0.012842471711337566, + 0.006818978115916252, + 0.01622794196009636, + 0.01413760706782341, + 0.018566874787211418, + 0.010964006185531616, + 0.006195470225065947, + 0.005157020874321461, + 0.012721638195216656, + 0.007422328926622868, + 0.016273323446512222, + 0.01183291245251894, + 0.011424426920711994, + 0.008667971938848495, + 0.003898652968928218, + 0.006026669405400753, + 0.0074814194813370705, + 0.0044837212190032005, + 0.008656698279082775, + 0.0063340310007333755, + 0.00628098426386714, + 0.011255839839577675, + 0.006457023788243532, + 0.00853704009205103, + 0.011780645698308945, + 0.008322280831634998, + 0.01244258787482977, + 0.013728205114603043, + 0.012106958776712418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.00248968368396163 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809133172035217 + ], + "opposite_logprob_tokens": [ + -3.2880914211273193 + ], + "logprob_reward": -0.03809133172035217, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809133172035217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3179793357849121, + 0.20040975511074066, + 0.05406307056546211, + 0.09130307286977768, + 0.045800574123859406, + 0.027543220669031143, + 0.03396305441856384, + 0.023692933842539787, + 0.07383637130260468, + 0.13140869140625 + ], + "credit_map": [ + 0.031846076250076294, + 0.023589933291077614, + 0.04121297597885132, + 0.030203603208065033, + 0.01955217495560646, + 0.06539895385503769, + 0.04390610009431839, + 0.06226952001452446, + 0.015454872511327267, + 0.02366640418767929, + 0.02124321088194847, + 0.0205927025526762, + 0.021669704467058182, + 0.042223669588565826, + 0.02822328358888626, + 0.027335911989212036, + 0.005501564126461744, + 0.004165382124483585, + 0.007415417116135359, + 0.008145110681653023, + 0.0060956720262765884, + 0.006662048399448395, + 0.007394964806735516, + 0.008682910352945328, + 0.012905928306281567, + 0.007482248358428478, + 0.012317347340285778, + 0.017043687403202057, + 0.009400779381394386, + 0.010018926113843918, + 0.009178778156638145, + 0.012955380603671074, + 0.005837068427354097, + 0.004206773824989796, + 0.0037473728880286217, + 0.004938556347042322, + 0.004211327061057091, + 0.009068118408322334, + 0.007551266811788082, + 0.006240088492631912, + 0.003794778836891055, + 0.0020721214823424816, + 0.00366037804633379, + 0.003494177246466279, + 0.0025034300051629543, + 0.00422360934317112, + 0.003600851632654667, + 0.004193873610347509, + 0.007671450264751911, + 0.0024751564487814903, + 0.003814875613898039, + 0.003066950710490346, + 0.0021164817735552788, + 0.005520268809050322, + 0.0034931693226099014, + 0.005804697051644325, + 0.003699206979945302, + 0.00216857367195189, + 0.003075846703723073, + 0.0032619305420666933, + 0.0018448146292939782, + 0.003110380144789815, + 0.0031072800047695637, + 0.003424903145059943, + 0.014817305840551853, + 0.005542473401874304, + 0.006892944220453501, + 0.008451377972960472, + 0.004337582271546125, + 0.01291665993630886, + 0.008531929925084114, + 0.012346095405519009, + 0.012925051152706146, + 0.011305023916065693, + 0.014049533754587173, + 0.018233440816402435, + 0.011347068473696709, + 0.02578565664589405, + 0.017264999449253082, + 0.020497918128967285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.03809133172035217 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340853422880173 + ], + "opposite_logprob_tokens": [ + -3.651340961456299 + ], + "logprob_reward": -0.026340853422880173, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340853422880173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1064714565873146, + 0.092312753200531, + 0.064759261906147, + 0.04511425644159317, + 0.03509881719946861, + 0.10618335753679276, + 0.04868340864777565, + 0.07573571801185608, + 0.2171989381313324, + 0.20844203233718872 + ], + "credit_map": [ + 0.011970906518399715, + 0.006870499812066555, + 0.013378547504544258, + 0.010778608731925488, + 0.008791403844952583, + 0.018196962773799896, + 0.01752633787691593, + 0.018958190456032753, + 0.010972815565764904, + 0.012302294373512268, + 0.008386349305510521, + 0.013360186479985714, + 0.012006114237010479, + 0.00826631486415863, + 0.01650957390666008, + 0.01050910446792841, + 0.013080005533993244, + 0.004666610155254602, + 0.013984293676912785, + 0.005765703972429037, + 0.0049160136841237545, + 0.006228972692042589, + 0.008236231282353401, + 0.007881433703005314, + 0.006589790806174278, + 0.0036853929050266743, + 0.004375631920993328, + 0.006071516778320074, + 0.0052603124640882015, + 0.0061306944116950035, + 0.0053572217002511024, + 0.007643692195415497, + 0.004587715957313776, + 0.002935409778729081, + 0.003642462892457843, + 0.004258093424141407, + 0.0036425816360861063, + 0.006837991066277027, + 0.003719245083630085, + 0.005475316196680069, + 0.01754528284072876, + 0.007445464842021465, + 0.014278935268521309, + 0.010578736662864685, + 0.005806423723697662, + 0.01990271918475628, + 0.009748389944434166, + 0.020877407863736153, + 0.00826271716505289, + 0.003810496535152197, + 0.004269829019904137, + 0.008096410892903805, + 0.004753821063786745, + 0.006150030996650457, + 0.006790871266275644, + 0.0065492321737110615, + 0.01610415428876877, + 0.006289967801421881, + 0.007315462455153465, + 0.007887357845902443, + 0.0048467665910720825, + 0.012321367859840393, + 0.009007224813103676, + 0.011963418684899807, + 0.03907899558544159, + 0.01724451221525669, + 0.019820908084511757, + 0.03017844446003437, + 0.014822456054389477, + 0.033354368060827255, + 0.02882450260221958, + 0.03387473151087761, + 0.01950938068330288, + 0.014510049484670162, + 0.01794961839914322, + 0.02661772631108761, + 0.016416827216744423, + 0.05153380706906319, + 0.036220256239175797, + 0.02568436786532402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.026340853422880173 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tight close-up centers on the face of a young, anthropomorphic panda monk, his entire body radiating palpable frustration. His black and white fur, rendered in a complex 3D style reminiscent of the \"Two Cities\" series, boasts rich painterly textures, while sharp, expressive, glass-shaded outlines define his form. His forehead is deeply furrowed, casting harsh, graphic shadows over his eyes, which narrow to a furious gleam, trained on a failed potion off-screen. His jaw clenches hard, his teeth visibly set, and his lips draw back slightly in a silent roar of intense annoyance. The faint, sickly purple glow of the unseen potion casts a dramatic backlight on his nose and chin, contrasting starkly with the warm ambient light of the dimly lit stone chamber in the background. Every muscle in his furry face is tense, a mask of concentrated annoyance that captures the fleeting moment of alchemical failure with profound emotional weight and a unique, hand-painted aesthetic.", + "seed": 10001, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_4/index_8762_seed_10001_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17049847096204757, + "vlm_logprob": -0.17049847096204757, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.17049847096204757, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8270085453987122 + ], + "opposite_logprob_tokens": [ + -0.5770085453987122 + ], + "logprob_reward": -0.8270085453987122, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8270085453987122, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2028389424085617, + 0.16593217849731445, + 0.26974475383758545, + 0.05508694052696228, + 0.05934182181954384, + 0.07672431319952011, + 0.031121041625738144, + 0.02614285796880722, + 0.05711207538843155, + 0.05595509707927704 + ], + "credit_map": [ + 0.030519934371113777, + 0.020169230177998543, + 0.015363700687885284, + 0.022054169327020645, + 0.026372093707323074, + 0.03585154563188553, + 0.02470739185810089, + 0.02780088037252426, + 0.02997679077088833, + 0.008626144379377365, + 0.007845032028853893, + 0.009942890144884586, + 0.007802045438438654, + 0.06631333380937576, + 0.00987202301621437, + 0.02555391564965248, + 0.03840120881795883, + 0.021100763231515884, + 0.026507355272769928, + 0.025614328682422638, + 0.03267120569944382, + 0.0477590374648571, + 0.04132016375660896, + 0.03637067601084709, + 0.0078000775538384914, + 0.006586663890630007, + 0.005742654670029879, + 0.005623225122690201, + 0.006664991844445467, + 0.006682934705168009, + 0.008979320526123047, + 0.007007074076682329, + 0.006433885544538498, + 0.007250932510942221, + 0.004812716506421566, + 0.0052907420322299, + 0.00889243558049202, + 0.00785794761031866, + 0.011554277502000332, + 0.007248884066939354, + 0.008878884837031364, + 0.008240674622356892, + 0.007449477445334196, + 0.009743035770952702, + 0.009239192120730877, + 0.011112443171441555, + 0.01163303293287754, + 0.010427572764456272, + 0.004627713467925787, + 0.0032305335626006126, + 0.0038883809465914965, + 0.004043896216899157, + 0.0027274766471236944, + 0.004596533253788948, + 0.0038799846079200506, + 0.004126523621380329, + 0.004904655739665031, + 0.0030440439004451036, + 0.003642277093604207, + 0.002559748711064458, + 0.002493613865226507, + 0.0029921478126198053, + 0.0037669248413294554, + 0.0027394460048526525, + 0.006745268125087023, + 0.007015840616077185, + 0.0053039551712572575, + 0.00871710292994976, + 0.0065211402252316475, + 0.00806359201669693, + 0.007241726852953434, + 0.0075034527108073235, + 0.00633884035050869, + 0.0058623082004487514, + 0.00544344587251544, + 0.005510575603693724, + 0.009246678091585636, + 0.010429199784994125, + 0.006763961166143417, + 0.006360087543725967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.8270085453987122 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017302555497735739 + ], + "opposite_logprob_tokens": [ + -6.376730442047119 + ], + "logprob_reward": -0.0017302555497735739, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017302555497735739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20157697796821594, + 0.1594565361738205, + 0.19076313078403473, + 0.07510935515165329, + 0.03763365000486374, + 0.02783036231994629, + 0.12880940735340118, + 0.06666766107082367, + 0.06057203561067581, + 0.05158087983727455 + ], + "credit_map": [ + 0.025085028260946274, + 0.021970832720398903, + 0.014977147802710533, + 0.023188691586256027, + 0.020604077726602554, + 0.03568171337246895, + 0.023183610290288925, + 0.03688587620854378, + 0.0375693142414093, + 0.009474116377532482, + 0.009057097136974335, + 0.009317991323769093, + 0.006823331117630005, + 0.05706825107336044, + 0.009216890670359135, + 0.020929541438817978, + 0.026024993509054184, + 0.01517315674573183, + 0.01913432776927948, + 0.01974557712674141, + 0.02512320689857006, + 0.028825005516409874, + 0.02836078219115734, + 0.02837609499692917, + 0.011225782334804535, + 0.008000904694199562, + 0.010515172965824604, + 0.008849882520735264, + 0.00898349191993475, + 0.008125507272779942, + 0.012453898787498474, + 0.006954720243811607, + 0.0035322026815265417, + 0.004234542138874531, + 0.0033703064545989037, + 0.005025604739785194, + 0.004859139211475849, + 0.004236132372170687, + 0.007452441845089197, + 0.004923281725496054, + 0.0029586805030703545, + 0.003019188065081835, + 0.0027113589458167553, + 0.0040603261440992355, + 0.0031648899894207716, + 0.003318948671221733, + 0.0055120643228292465, + 0.0030849066097289324, + 0.017552819103002548, + 0.012610248290002346, + 0.013357740826904774, + 0.015292958356440067, + 0.01732318475842476, + 0.016082003712654114, + 0.016260676085948944, + 0.020329780876636505, + 0.014152368530631065, + 0.0063921017572283745, + 0.007201461587101221, + 0.007136838510632515, + 0.006862160749733448, + 0.0072635202668607235, + 0.009780499152839184, + 0.007878710515797138, + 0.006481111980974674, + 0.007026241160929203, + 0.004689732100814581, + 0.007730355951935053, + 0.008122929371893406, + 0.0075816004537045956, + 0.010304315015673637, + 0.008635749109089375, + 0.006284219678491354, + 0.004637641832232475, + 0.004663658328354359, + 0.005067064892500639, + 0.007557214703410864, + 0.011455650441348553, + 0.005886588245630264, + 0.006028841715306044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0017302555497735739 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019394413102418184 + ], + "opposite_logprob_tokens": [ + -6.251939296722412 + ], + "logprob_reward": -0.0019394413102418184, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019394413102418184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21255044639110565, + 0.09609492123126984, + 0.16861407458782196, + 0.09708722680807114, + 0.045219484716653824, + 0.03478377312421799, + 0.0888812467455864, + 0.1578676402568817, + 0.048466749489307404, + 0.05043443292379379 + ], + "credit_map": [ + 0.0398147813975811, + 0.017041325569152832, + 0.015385167673230171, + 0.02336651273071766, + 0.019929926842451096, + 0.046192292124032974, + 0.02173333242535591, + 0.029087120667099953, + 0.013781536370515823, + 0.007996273227036, + 0.007643315941095352, + 0.01297682523727417, + 0.008926557376980782, + 0.02536929026246071, + 0.009835279546678066, + 0.009565845131874084, + 0.018566187471151352, + 0.015385369770228863, + 0.01656155474483967, + 0.01742413453757763, + 0.023953983560204506, + 0.02113715186715126, + 0.02857223153114319, + 0.027013450860977173, + 0.01587631180882454, + 0.011536430567502975, + 0.009377826936542988, + 0.009974732995033264, + 0.011436191387474537, + 0.01479349285364151, + 0.014167211949825287, + 0.009925030171871185, + 0.005334855057299137, + 0.0037348680198192596, + 0.003341123228892684, + 0.005085618235170841, + 0.004379442892968655, + 0.009922415018081665, + 0.007759014144539833, + 0.005662147421389818, + 0.004850235302001238, + 0.0034858642611652613, + 0.0033017967361956835, + 0.003739830106496811, + 0.00455642631277442, + 0.004139293916523457, + 0.006403186358511448, + 0.004307142458856106, + 0.009993297979235649, + 0.0076906816102564335, + 0.009782114066183567, + 0.009184998460114002, + 0.012356115505099297, + 0.012171264737844467, + 0.014200150035321712, + 0.013502617366611958, + 0.023010093718767166, + 0.01983995921909809, + 0.016089919954538345, + 0.020240316167473793, + 0.015237262472510338, + 0.01903330348432064, + 0.02440497651696205, + 0.020011814311146736, + 0.006479725241661072, + 0.004780951421707869, + 0.0042793783359229565, + 0.006523055490106344, + 0.006106143817305565, + 0.0073180426843464375, + 0.007782652508467436, + 0.005196800921112299, + 0.004988583270460367, + 0.005484634079039097, + 0.004759656731039286, + 0.006531488616019487, + 0.005115823354572058, + 0.008709397166967392, + 0.008845560252666473, + 0.005999291781336069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.0019394413102418184 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189658686518669 + ], + "opposite_logprob_tokens": [ + -4.264189720153809 + ], + "logprob_reward": -0.014189658686518669, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189658686518669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15244057774543762, + 0.034407682716846466, + 0.3793536126613617, + 0.1125413030385971, + 0.06035955995321274, + 0.04777466878294945, + 0.03205427899956703, + 0.06699579209089279, + 0.07413218170404434, + 0.03994035720825195 + ], + "credit_map": [ + 0.022492190822958946, + 0.010244500823318958, + 0.009480695240199566, + 0.015945719555020332, + 0.014820994809269905, + 0.04519062861800194, + 0.013038474135100842, + 0.02122737094759941, + 0.004569736309349537, + 0.002805654425173998, + 0.00254178699105978, + 0.0044982158578932285, + 0.003472903510555625, + 0.008091215044260025, + 0.004198353737592697, + 0.004229819402098656, + 0.07749627530574799, + 0.0347340926527977, + 0.04568594694137573, + 0.03016948327422142, + 0.029627369716763496, + 0.04871687665581703, + 0.0512387715280056, + 0.06168478727340698, + 0.022415390238165855, + 0.00859704241156578, + 0.014600526541471481, + 0.0107412775978446, + 0.013822520151734352, + 0.01427527517080307, + 0.01493013370782137, + 0.01315914187580347, + 0.009817901067435741, + 0.00668734684586525, + 0.007665061857551336, + 0.007419293280690908, + 0.006771189626306295, + 0.007355377543717623, + 0.007976998575031757, + 0.006666390225291252, + 0.006970353424549103, + 0.004677990451455116, + 0.006315215490758419, + 0.006021274719387293, + 0.005058529786765575, + 0.005799852777272463, + 0.006395197007805109, + 0.006536252796649933, + 0.0039026979357004166, + 0.0031751899514347315, + 0.003791533410549164, + 0.0041379439644515514, + 0.004295031540095806, + 0.004167427308857441, + 0.003855709917843342, + 0.004728745669126511, + 0.009459352120757103, + 0.008979663252830505, + 0.00814706739038229, + 0.00901804305613041, + 0.007239436265081167, + 0.009529783390462399, + 0.008035381324589252, + 0.006587067153304815, + 0.0137641541659832, + 0.007858316414058208, + 0.009160362184047699, + 0.005951144266873598, + 0.010547665879130363, + 0.008836098946630955, + 0.009477770887315273, + 0.00853666476905346, + 0.004247276112437248, + 0.004043595865368843, + 0.003631297964602709, + 0.00410480098798871, + 0.00637912517413497, + 0.007389497943222523, + 0.004979385063052177, + 0.005165379494428635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.014189658686518669 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young anthropomorphic panda monk with black and white fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a static tight close-up centered on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display specific facial expressions including a furrowed forehead, narrowed eyes, clenched jaw, and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style a complex 3D render with rich painterly textures and sharp glass-shaded outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint sickly purple backlight on the nose and chin contrasting with warm ambient light in a dimly lit stone chamber background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624453864991665 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624453864991665, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624453864991665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27689382433891296, + 0.35742148756980896, + 0.08349894732236862, + 0.038450319319963455, + 0.03013353794813156, + 0.028978703543543816, + 0.03134378045797348, + 0.07332246750593185, + 0.04293474555015564, + 0.0370221883058548 + ], + "credit_map": [ + 0.025660643354058266, + 0.025155238807201385, + 0.019071849063038826, + 0.03286794573068619, + 0.03317497670650482, + 0.06061621382832527, + 0.03236359730362892, + 0.04798336327075958, + 0.06845688819885254, + 0.01743577979505062, + 0.014415380544960499, + 0.023922208696603775, + 0.016281338408589363, + 0.1420559138059616, + 0.021618759259581566, + 0.0532352477312088, + 0.011446785181760788, + 0.0073103164322674274, + 0.009425808675587177, + 0.008229440078139305, + 0.011367171071469784, + 0.010205612517893314, + 0.013233383186161518, + 0.012280428782105446, + 0.00779857998713851, + 0.0034415756817907095, + 0.004143268335610628, + 0.00540538877248764, + 0.0027887732721865177, + 0.005991168785840273, + 0.003825768129900098, + 0.005055795889347792, + 0.004487238824367523, + 0.002672904636710882, + 0.0025541987270116806, + 0.002872084267437458, + 0.003590976819396019, + 0.004981662146747112, + 0.004933864809572697, + 0.004040607251226902, + 0.004294756334275007, + 0.0030983134638518095, + 0.0030488790944218636, + 0.004092171788215637, + 0.002855051541700959, + 0.0044630467891693115, + 0.0038186542224138975, + 0.003307831007987261, + 0.005966695956885815, + 0.0029824418015778065, + 0.0030631818808615208, + 0.00295082270167768, + 0.0034591772127896547, + 0.004060846753418446, + 0.005163820460438728, + 0.0036967932246625423, + 0.018953006714582443, + 0.005975230131298304, + 0.010117308236658573, + 0.005612461362034082, + 0.006958537735044956, + 0.007618569768965244, + 0.009510558098554611, + 0.008576791733503342, + 0.0042377058416605, + 0.0051893978379666805, + 0.003414343111217022, + 0.0057379999198019505, + 0.007150005083531141, + 0.005029202904552221, + 0.006404556334018707, + 0.005771534517407417, + 0.003781893290579319, + 0.0030887264292687178, + 0.0034241382963955402, + 0.004380779806524515, + 0.003918410744518042, + 0.008114269003272057, + 0.005796576384454966, + 0.004517395980656147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.007624453864991665 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_5/index_tensor([8820, 8820])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_5/index_tensor([8820, 8820])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f2b3d1b84861133c1676c59fb00c4bd03959ebce --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_5/index_tensor([8820, 8820])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, mid-shot shot, shot from a horizontal angle, captures a quiet corner of an ancient stone courtyard at midnight. Near a stone archway in the stone wall, a pale gray moth with intricate dark markings on its wings hovers softly and erratically, its wings fluttering gently in the air. Dense ivy climbs along the massive, weathered stone walls, its deep green leaves reflecting a faint, moist sheen in the pale, cold moonlight. The moth is evidently attracted to the faint, fading gold leaf embedded in the cracks of the stone wall. The warm, flickering yellow light of an antique lantern hanging in the courtyard catches the leaf, giving it a faint, alluring golden glow. The interplay of cold moonlight and warm lamplight creates a quiet, ancient, and slightly magical midnight atmosphere.", + "seed": 33038, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_5/index_8820_seed_33038_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1343029179610312, + "vlm_logprob": -0.1343029179610312, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.1343029179610312, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24376851320266724 + ], + "opposite_logprob_tokens": [ + -1.6187684535980225 + ], + "logprob_reward": -0.24376851320266724, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24376851320266724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30306974053382874, + 0.10679584741592407, + 0.10246963053941727, + 0.12294141948223114, + 0.092321015894413, + 0.06949972361326218, + 0.03455452248454094, + 0.03534728288650513, + 0.07316512614488602, + 0.059835709631443024 + ], + "credit_map": [ + 0.01457529328763485, + 0.01171684730798006, + 0.0889640748500824, + 0.023915918543934822, + 0.015924129635095596, + 0.10092472285032272, + 0.01676004007458687, + 0.030288707464933395, + 0.006766242440789938, + 0.017657263204455376, + 0.011490077711641788, + 0.01912335492670536, + 0.013974137604236603, + 0.01139125693589449, + 0.016758829355239868, + 0.009634692221879959, + 0.010588543489575386, + 0.009897392243146896, + 0.013285302557051182, + 0.014213628135621548, + 0.012492417357861996, + 0.01693928986787796, + 0.011623010039329529, + 0.013430053368210793, + 0.008513164706528187, + 0.0106598986312747, + 0.010198662057518959, + 0.019413841888308525, + 0.016705727204680443, + 0.015887262299656868, + 0.015950370579957962, + 0.025612493976950645, + 0.01193981897085905, + 0.00812888890504837, + 0.014084124937653542, + 0.010886461474001408, + 0.008765035308897495, + 0.009437733329832554, + 0.01521815825253725, + 0.013860802166163921, + 0.005398558918386698, + 0.007542487699538469, + 0.010204815305769444, + 0.013749359175562859, + 0.007909972220659256, + 0.007427752949297428, + 0.007722083944827318, + 0.009544696658849716, + 0.004179390147328377, + 0.0030900337733328342, + 0.0044222064316272736, + 0.004285804927349091, + 0.0038019323255866766, + 0.0033004514407366514, + 0.00524259964004159, + 0.006232104264199734, + 0.0025864916387945414, + 0.0025634269695729017, + 0.004029100760817528, + 0.005368915852159262, + 0.005072373431175947, + 0.0053878300823271275, + 0.006139521487057209, + 0.004199624061584473, + 0.005483664106577635, + 0.009710506536066532, + 0.007016456685960293, + 0.00837960746139288, + 0.011912756599485874, + 0.011113671585917473, + 0.009783822111785412, + 0.009764647111296654, + 0.004980451427400112, + 0.006163773592561483, + 0.008482817560434341, + 0.007910544984042645, + 0.005573954898864031, + 0.006524406373500824, + 0.00908777117729187, + 0.011111991479992867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.24376851320266724 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043089091777801514 + ], + "opposite_logprob_tokens": [ + -3.1680891513824463 + ], + "logprob_reward": -0.043089091777801514, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043089091777801514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21394310891628265, + 0.15198515355587006, + 0.18748919665813446, + 0.11828278005123138, + 0.06066763401031494, + 0.03078632988035679, + 0.028145238757133484, + 0.028445066884160042, + 0.052668020129203796, + 0.12758740782737732 + ], + "credit_map": [ + 0.012317627668380737, + 0.020826000720262527, + 0.04802076146006584, + 0.022479521110653877, + 0.022030798718333244, + 0.0466616153717041, + 0.018737733364105225, + 0.022869065403938293, + 0.013370121829211712, + 0.030253823846578598, + 0.016651323065161705, + 0.019380532205104828, + 0.024920709431171417, + 0.016326667740941048, + 0.016287511214613914, + 0.014794481918215752, + 0.0192575566470623, + 0.018837178125977516, + 0.018491201102733612, + 0.02587803080677986, + 0.029787449166178703, + 0.027149412781000137, + 0.02571473829448223, + 0.022373652085661888, + 0.013686890713870525, + 0.009780100546777248, + 0.011450598947703838, + 0.014266103506088257, + 0.013380746357142925, + 0.01645994372665882, + 0.019961144775152206, + 0.019297266378998756, + 0.003470273222774267, + 0.003605926176533103, + 0.0055465386249125, + 0.005140469875186682, + 0.007447949610650539, + 0.008797043934464455, + 0.014029795303940773, + 0.012629647739231586, + 0.002765210811048746, + 0.002740784315392375, + 0.0048612444661557674, + 0.0057859402149915695, + 0.003148677758872509, + 0.003704383270815015, + 0.003747050417587161, + 0.004033041652292013, + 0.001979847438633442, + 0.0018452858785167336, + 0.0022649867460131645, + 0.0043886941857635975, + 0.0039779371581971645, + 0.003278101561591029, + 0.0036710770800709724, + 0.006739311385899782, + 0.003877288429066539, + 0.0041478704661130905, + 0.0033575717825442553, + 0.0046393973752856255, + 0.0030754245817661285, + 0.003157244063913822, + 0.0031029568053781986, + 0.0030873187351971865, + 0.004035346210002899, + 0.004395943600684404, + 0.005929579958319664, + 0.0099590253084898, + 0.0063760848715901375, + 0.007957677356898785, + 0.007499517407268286, + 0.00651484914124012, + 0.008860602974891663, + 0.00873386301100254, + 0.027497122064232826, + 0.014998498372733593, + 0.01015597116202116, + 0.012841432355344296, + 0.01515559758991003, + 0.029344335198402405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.043089091777801514 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319886626675725 + ], + "opposite_logprob_tokens": [ + -5.753199100494385 + ], + "logprob_reward": -0.00319886626675725, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319886626675725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14034411311149597, + 0.10008041560649872, + 0.19382834434509277, + 0.0895206555724144, + 0.08421256393194199, + 0.07444749772548676, + 0.13311412930488586, + 0.03521748632192612, + 0.04857783019542694, + 0.10065698623657227 + ], + "credit_map": [ + 0.009281132370233536, + 0.016033148393034935, + 0.017898032441735268, + 0.020154904574155807, + 0.02004632167518139, + 0.02188357710838318, + 0.019499264657497406, + 0.015547736547887325, + 0.005534169729799032, + 0.015493970364332199, + 0.01265249028801918, + 0.01087271049618721, + 0.012116390280425549, + 0.019541308283805847, + 0.01245486456900835, + 0.011414509266614914, + 0.011742393486201763, + 0.02556600235402584, + 0.015177438966929913, + 0.028917649760842323, + 0.037874121218919754, + 0.026939641684293747, + 0.027027461677789688, + 0.02058364450931549, + 0.005641232710331678, + 0.005842812824994326, + 0.008515781722962856, + 0.011212487705051899, + 0.012714999727904797, + 0.01780228689312935, + 0.01597338728606701, + 0.011817666701972485, + 0.004931438714265823, + 0.005732163321226835, + 0.007927333936095238, + 0.007978696376085281, + 0.013229362666606903, + 0.010097604244947433, + 0.015902139246463776, + 0.018413826823234558, + 0.006865627598017454, + 0.006881513632833958, + 0.011098206043243408, + 0.01349340658634901, + 0.009938220493495464, + 0.007770431227982044, + 0.00811936054378748, + 0.010280736722052097, + 0.007392878644168377, + 0.01388205960392952, + 0.009741123765707016, + 0.022813567891716957, + 0.01595531776547432, + 0.018441272899508476, + 0.01974341832101345, + 0.025144489482045174, + 0.003563379868865013, + 0.0036957915872335434, + 0.003675894346088171, + 0.007753190118819475, + 0.0029990868642926216, + 0.003960196860134602, + 0.00459333136677742, + 0.004976614378392696, + 0.0038193687796592712, + 0.006045626010745764, + 0.003941763658076525, + 0.006984780076891184, + 0.006614274345338345, + 0.006784908473491669, + 0.007691039238125086, + 0.0066960714757442474, + 0.010283028706908226, + 0.01037086732685566, + 0.02009671926498413, + 0.011620033532381058, + 0.009764072485268116, + 0.011271098628640175, + 0.013540493324398994, + 0.013710676692426205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.00319886626675725 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157891511917114 + ], + "opposite_logprob_tokens": [ + -1.7015788555145264 + ], + "logprob_reward": -0.20157891511917114, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157891511917114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31670626997947693, + 0.10196997225284576, + 0.044554900377988815, + 0.06739308685064316, + 0.05897994339466095, + 0.0897483229637146, + 0.09423653036355972, + 0.044163964688777924, + 0.08225817233324051, + 0.09998887777328491 + ], + "credit_map": [ + 0.01552130188792944, + 0.029271170496940613, + 0.05211718752980232, + 0.02721734531223774, + 0.03639182448387146, + 0.09192291647195816, + 0.026337558403611183, + 0.03792697563767433, + 0.006539466325193644, + 0.014892352744936943, + 0.015011336654424667, + 0.012613669969141483, + 0.014891657046973705, + 0.009373024106025696, + 0.01680181547999382, + 0.011846650391817093, + 0.0030213443096727133, + 0.0043135336600244045, + 0.004303622990846634, + 0.0071051595732569695, + 0.005658134818077087, + 0.005381416995078325, + 0.010369398631155491, + 0.004402290564030409, + 0.007290403358638287, + 0.00923878513276577, + 0.006168076302856207, + 0.010590624995529652, + 0.006827980279922485, + 0.009592375718057156, + 0.009115192107856274, + 0.00856965035200119, + 0.0045722248032689095, + 0.005048103164881468, + 0.007108671125024557, + 0.009037774987518787, + 0.006152321118861437, + 0.005961436778306961, + 0.009914420545101166, + 0.011184992268681526, + 0.01081149186939001, + 0.00895635224878788, + 0.016759168356657028, + 0.013956011272966862, + 0.009663538075983524, + 0.006950172130018473, + 0.00994422659277916, + 0.012707370333373547, + 0.005117341876029968, + 0.011855543591082096, + 0.009374234825372696, + 0.020033566281199455, + 0.00857425108551979, + 0.009019737131893635, + 0.012904498726129532, + 0.017357351258397102, + 0.005937590729445219, + 0.005293646827340126, + 0.006171470042318106, + 0.007505503017455339, + 0.004345234017819166, + 0.004261449445039034, + 0.005119164939969778, + 0.005529905203729868, + 0.010890326462686062, + 0.013258134946227074, + 0.011236105114221573, + 0.013260950334370136, + 0.005229632835835218, + 0.009898058138787746, + 0.008960071951150894, + 0.009524902328848839, + 0.00936067383736372, + 0.02254054881632328, + 0.011062380857765675, + 0.018493540585041046, + 0.007014462258666754, + 0.008769774809479713, + 0.011038145050406456, + 0.011709345504641533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.20157891511917114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987920343875885 + ], + "opposite_logprob_tokens": [ + -1.8048791885375977 + ], + "logprob_reward": -0.17987920343875885, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987920343875885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18410633504390717, + 0.14359277486801147, + 0.08951312303543091, + 0.12850387394428253, + 0.06922346353530884, + 0.04342706874012947, + 0.043905820697546005, + 0.1654248684644699, + 0.03891628980636597, + 0.09338635206222534 + ], + "credit_map": [ + 0.01344036590307951, + 0.015047411434352398, + 0.05806879699230194, + 0.022258125245571136, + 0.011556570418179035, + 0.03951979801058769, + 0.011881195940077305, + 0.012334072962403297, + 0.006606248207390308, + 0.015720443800091743, + 0.01567837968468666, + 0.010845704935491085, + 0.011479939334094524, + 0.04482293501496315, + 0.01461564376950264, + 0.023823492228984833, + 0.005590622778981924, + 0.008744649589061737, + 0.007767509203404188, + 0.014845073223114014, + 0.014862815849483013, + 0.015939554199576378, + 0.013076932169497013, + 0.00868596788495779, + 0.011475877836346626, + 0.008738301694393158, + 0.014749820344150066, + 0.020269151777029037, + 0.01193673349916935, + 0.016082879155874252, + 0.019317079335451126, + 0.02593402750790119, + 0.004419992212206125, + 0.004120505414903164, + 0.007519901730120182, + 0.008735395967960358, + 0.008987311273813248, + 0.008095748722553253, + 0.012024972587823868, + 0.015319635160267353, + 0.005110019817948341, + 0.005644960794597864, + 0.00592009536921978, + 0.008135000243782997, + 0.004141011741012335, + 0.004513831809163094, + 0.004373166244477034, + 0.005588983651250601, + 0.0041430676355957985, + 0.004237390123307705, + 0.005794679746031761, + 0.005676050670444965, + 0.006911824457347393, + 0.00410276697948575, + 0.006625593174248934, + 0.006414447445422411, + 0.020229773595929146, + 0.0160013847053051, + 0.021176330745220184, + 0.03134189918637276, + 0.012621290050446987, + 0.014851506799459457, + 0.025397218763828278, + 0.023805461823940277, + 0.0034389887005090714, + 0.003733521094545722, + 0.005069808568805456, + 0.0054986379109323025, + 0.005953315179795027, + 0.005719564389437437, + 0.004836589563637972, + 0.004665867891162634, + 0.008545775897800922, + 0.011091400869190693, + 0.014998555183410645, + 0.014392760582268238, + 0.006542844232171774, + 0.007317994721233845, + 0.011464430019259453, + 0.019032588228583336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.17987920343875885 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-shot shot, shot from a horizontal angle, captures a quiet corner of an ancient stone courtyard at midnight. Near a stone archway in the stone wall, a pale gray moth with intricate dark markings on its wings hovers softly and erratically, its wings fluttering gently in the air. Dense ivy climbs along the massive, weathered stone walls, its deep green leaves reflecting a faint, moist sheen in the pale, cold moonlight. The moth is evidently attracted to the faint, fading gold leaf embedded in the cracks of the stone wall. The warm, flickering yellow light of an antique lantern hanging in the courtyard catches the leaf, giving it a faint, alluring golden glow. The interplay of cold moonlight and warm lamplight creates a quiet, ancient, and slightly magical midnight atmosphere.", + "seed": 33039, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_5/index_8820_seed_33039_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1327412474900484, + "vlm_logprob": -0.1327412474900484, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.1327412474900484, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.536510705947876 + ], + "opposite_logprob_tokens": [ + -0.911510705947876 + ], + "logprob_reward": -0.536510705947876, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.536510705947876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18575482070446014, + 0.06351221352815628, + 0.02403389848768711, + 0.17527250945568085, + 0.030363887548446655, + 0.06126902997493744, + 0.27778998017311096, + 0.10394418239593506, + 0.0397278256714344, + 0.03833168372511864 + ], + "credit_map": [ + 0.01648169569671154, + 0.012987102381885052, + 0.02032610960304737, + 0.015109793283045292, + 0.018377427011728287, + 0.05362926796078682, + 0.013943644240498543, + 0.03489978611469269, + 0.004944684915244579, + 0.005886259023100138, + 0.010026928968727589, + 0.008476912043988705, + 0.008128509856760502, + 0.010101565159857273, + 0.005896185524761677, + 0.010051172226667404, + 0.002599994884803891, + 0.003133589867502451, + 0.004548148717731237, + 0.0041180760599672794, + 0.0022277124226093292, + 0.0024181166663765907, + 0.002286295872181654, + 0.00270196539349854, + 0.011597225442528725, + 0.006441162433475256, + 0.03742755204439163, + 0.010415364056825638, + 0.01579715684056282, + 0.06454045325517654, + 0.006347822956740856, + 0.02270575799047947, + 0.0024030052591115236, + 0.002635831478983164, + 0.006969139911234379, + 0.0026279138401150703, + 0.003276390954852104, + 0.007725422736257315, + 0.0019777389243245125, + 0.0027484456077218056, + 0.0055498057045042515, + 0.006502461154013872, + 0.01506193820387125, + 0.0056441668421030045, + 0.005077446345239878, + 0.015594650991261005, + 0.0033534655813127756, + 0.004485095851123333, + 0.021516814827919006, + 0.013055765070021152, + 0.06569655239582062, + 0.01584162749350071, + 0.023100242018699646, + 0.09486320614814758, + 0.011000427417457104, + 0.0327153354883194, + 0.011310311034321785, + 0.006606330629438162, + 0.03188762441277504, + 0.008577888831496239, + 0.007764884270727634, + 0.01923353411257267, + 0.0050059654749929905, + 0.013557640835642815, + 0.003989482298493385, + 0.006242838688194752, + 0.00577920489013195, + 0.004236365668475628, + 0.005437010433524847, + 0.005621424410492182, + 0.002895464887842536, + 0.005526034161448479, + 0.0024381752591580153, + 0.0031602096278220415, + 0.004629224073141813, + 0.002906742738559842, + 0.002838818123564124, + 0.01016987580806017, + 0.001982620218768716, + 0.010206018574535847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.536510705947876 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370229199528694 + ], + "opposite_logprob_tokens": [ + -3.6513702869415283 + ], + "logprob_reward": -0.026370229199528694, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370229199528694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05329045653343201, + 0.07441273331642151, + 0.01734703779220581, + 0.06806094199419022, + 0.05807749554514885, + 0.019013863056898117, + 0.3351856470108032, + 0.2820197641849518, + 0.051573414355516434, + 0.04101859778165817 + ], + "credit_map": [ + 0.006535414140671492, + 0.005947557277977467, + 0.006761715281754732, + 0.005213758908212185, + 0.008012370206415653, + 0.008499441668391228, + 0.005455906502902508, + 0.006864293944090605, + 0.006919462233781815, + 0.009998287074267864, + 0.015167417004704475, + 0.010443794541060925, + 0.004687338136136532, + 0.006752321496605873, + 0.003651291597634554, + 0.016792818903923035, + 0.0013417996233329177, + 0.0023497268557548523, + 0.0017947426531463861, + 0.0019350909860804677, + 0.001401697052642703, + 0.003929333761334419, + 0.0012937183491885662, + 0.00330092734657228, + 0.006300900597125292, + 0.0038028445560485125, + 0.020639119669795036, + 0.004520843271166086, + 0.005114006344228983, + 0.01913435198366642, + 0.0033426242880523205, + 0.005206250585615635, + 0.004636998288333416, + 0.0028545239474624395, + 0.01620439440011978, + 0.003658291883766651, + 0.004910746589303017, + 0.018738167360424995, + 0.0025409336667507887, + 0.004533443134278059, + 0.001838142634369433, + 0.0012895867694169283, + 0.005779261700809002, + 0.001590108498930931, + 0.0014525618171319366, + 0.004897853825241327, + 0.0007530744187533855, + 0.0014132721116766334, + 0.024821769446134567, + 0.010722006671130657, + 0.07363396883010864, + 0.014928560703992844, + 0.02709144353866577, + 0.1226171925663948, + 0.00952989887446165, + 0.05184084177017212, + 0.023233098909258842, + 0.009412702172994614, + 0.06815856695175171, + 0.014769036322832108, + 0.023127323016524315, + 0.10067085176706314, + 0.008363514207303524, + 0.034284669905900955, + 0.0031468376982957125, + 0.004083531443029642, + 0.004257762338966131, + 0.004133753478527069, + 0.0038735128473490477, + 0.015372144989669323, + 0.002394924173131585, + 0.014310947619378567, + 0.0033073376398533583, + 0.004614897537976503, + 0.007367611862719059, + 0.004720624070614576, + 0.0033022845163941383, + 0.008230555802583694, + 0.0026350095868110657, + 0.00684027886018157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.026370229199528694 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607177034020424 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.01607177034020424, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607177034020424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24748055636882782, + 0.1177622377872467, + 0.052539266645908356, + 0.10195703059434891, + 0.02415613830089569, + 0.029112812131643295, + 0.1542864739894867, + 0.18003836274147034, + 0.0354820191860199, + 0.057185158133506775 + ], + "credit_map": [ + 0.020031407475471497, + 0.01911303587257862, + 0.01713687740266323, + 0.01499173417687416, + 0.042194947600364685, + 0.059735506772994995, + 0.03477137163281441, + 0.03950566425919533, + 0.009976918809115887, + 0.008177598938345909, + 0.02942296676337719, + 0.01368990819901228, + 0.011900425888597965, + 0.02505424991250038, + 0.009216787293553352, + 0.010323373600840569, + 0.003109466517344117, + 0.0023253613617271185, + 0.011469904333353043, + 0.0036951981019228697, + 0.004064517095685005, + 0.01778947189450264, + 0.002216401044279337, + 0.00786894466727972, + 0.005510103423148394, + 0.007302718237042427, + 0.012721236795186996, + 0.008270171470940113, + 0.009824633598327637, + 0.03169305995106697, + 0.009969948790967464, + 0.016665145754814148, + 0.0013918830081820488, + 0.0023599527776241302, + 0.0021914620883762836, + 0.0019094967283308506, + 0.003099462017416954, + 0.005106506869196892, + 0.0034950855188071728, + 0.004602289292961359, + 0.0017508750315755606, + 0.002533828606829047, + 0.005123650189489126, + 0.0020786940585821867, + 0.0037955085281282663, + 0.008507505059242249, + 0.001970870653167367, + 0.0033518779091537, + 0.015320483595132828, + 0.007693646475672722, + 0.04929549619555473, + 0.0111541748046875, + 0.0123473284766078, + 0.03985578194260597, + 0.006071805953979492, + 0.012547752819955349, + 0.018419113010168076, + 0.00833936594426632, + 0.06044093519449234, + 0.013555590994656086, + 0.01120077632367611, + 0.027896732091903687, + 0.007145618554204702, + 0.03304021805524826, + 0.0034349923953413963, + 0.00377213628962636, + 0.0067361206747591496, + 0.0027292263694107533, + 0.006823147181421518, + 0.004697197116911411, + 0.004444518126547337, + 0.0028446775395423174, + 0.0028481907211244106, + 0.002114305505529046, + 0.0031016229186207056, + 0.0016488931141793728, + 0.004718398675322533, + 0.021016083657741547, + 0.0025841023307293653, + 0.01915355958044529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.01607177034020424 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495937541127205 + ], + "opposite_logprob_tokens": [ + -2.929959297180176 + ], + "logprob_reward": -0.05495937541127205, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495937541127205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052603431046009064, + 0.07218262553215027, + 0.09167813509702682, + 0.08533848077058792, + 0.01721917651593685, + 0.03157626464962959, + 0.16312193870544434, + 0.25057268142700195, + 0.033360470086336136, + 0.2023468166589737 + ], + "credit_map": [ + 0.007135838735848665, + 0.005781857762485743, + 0.006502246018499136, + 0.004968519322574139, + 0.005956833250820637, + 0.007055624853819609, + 0.008554773405194283, + 0.006647738162428141, + 0.006161754485219717, + 0.007553097791969776, + 0.013156729750335217, + 0.007776639889925718, + 0.006847338750958443, + 0.018083712086081505, + 0.004961972124874592, + 0.007641382049769163, + 0.006766895763576031, + 0.004773366264998913, + 0.02539970353245735, + 0.008890511468052864, + 0.006506363861262798, + 0.026669781655073166, + 0.004062310792505741, + 0.00860920175909996, + 0.0058229835703969, + 0.0030571625102311373, + 0.018959498032927513, + 0.004354432225227356, + 0.007668481674045324, + 0.031074760481715202, + 0.004304296802729368, + 0.010096863843500614, + 0.0015451280632987618, + 0.002043540123850107, + 0.002546973992139101, + 0.0017486463766545057, + 0.0015823733992874622, + 0.0042022462002933025, + 0.0018149807583540678, + 0.0017352886497974396, + 0.002694022608920932, + 0.003480276558548212, + 0.009182109497487545, + 0.0030138681177049875, + 0.002294788369908929, + 0.007578867021948099, + 0.0011943094432353973, + 0.0021380227990448475, + 0.013349214568734169, + 0.0055909790098667145, + 0.040741708129644394, + 0.00854395143687725, + 0.013517783023416996, + 0.05702632665634155, + 0.004948670044541359, + 0.01940331980586052, + 0.02281247265636921, + 0.011923644691705704, + 0.07132366299629211, + 0.016994768753647804, + 0.010711876675486565, + 0.02596808783710003, + 0.010971908457577229, + 0.07986626774072647, + 0.0027902170550078154, + 0.003998167347162962, + 0.0069946469739079475, + 0.00400009099394083, + 0.004226972348988056, + 0.005795205477625132, + 0.002558233216404915, + 0.0029969378374516964, + 0.009723059833049774, + 0.008080477826297283, + 0.010999603196978569, + 0.006485734134912491, + 0.0158962681889534, + 0.07660558819770813, + 0.007944699376821518, + 0.06661137193441391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.05495937541127205 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794156551361084 + ], + "opposite_logprob_tokens": [ + -3.529794216156006 + ], + "logprob_reward": -0.029794156551361084, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794156551361084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18848258256912231, + 0.09387096017599106, + 0.09661123901605606, + 0.10655596107244492, + 0.16890299320220947, + 0.04459820315241814, + 0.15053170919418335, + 0.07862067222595215, + 0.04111155495047569, + 0.030714156106114388 + ], + "credit_map": [ + 0.016480131074786186, + 0.01517297700047493, + 0.01950083300471306, + 0.015144247561693192, + 0.026530105620622635, + 0.04502252861857414, + 0.01989901438355446, + 0.03073275275528431, + 0.008714033290743828, + 0.006374407093971968, + 0.026003824546933174, + 0.009453452192246914, + 0.006580119486898184, + 0.010605767369270325, + 0.008151928894221783, + 0.01798742264509201, + 0.006635628174990416, + 0.008982378989458084, + 0.014025709591805935, + 0.007096091751009226, + 0.007882844656705856, + 0.027498936280608177, + 0.006701904349029064, + 0.01778775081038475, + 0.008945394307374954, + 0.015678275376558304, + 0.01132467482239008, + 0.012848015874624252, + 0.011904451064765453, + 0.021285422146320343, + 0.011108954437077045, + 0.013460776768624783, + 0.013887906447052956, + 0.009063757956027985, + 0.04856027290225029, + 0.014212199486792088, + 0.015380852855741978, + 0.04855141043663025, + 0.008305065333843231, + 0.010941537097096443, + 0.004168517421931028, + 0.005486899055540562, + 0.00971338339149952, + 0.004717282019555569, + 0.0038423321675509214, + 0.004410053603351116, + 0.0033009706530719995, + 0.008958766236901283, + 0.013786795549094677, + 0.006575341802090406, + 0.04622301831841469, + 0.009903520345687866, + 0.011991810984909534, + 0.04639242962002754, + 0.005407941062003374, + 0.010250856168568134, + 0.0044418759644031525, + 0.003683190094307065, + 0.005595870316028595, + 0.002624843269586563, + 0.006106488406658173, + 0.02977626398205757, + 0.00373407662846148, + 0.02265807054936886, + 0.003259446006268263, + 0.004070566035807133, + 0.004275258164852858, + 0.0031342587899416685, + 0.004185736645013094, + 0.009728903882205486, + 0.003609335981309414, + 0.00884804967790842, + 0.0025865354109555483, + 0.0033617049921303988, + 0.003720661159604788, + 0.0023842938244342804, + 0.0019916475284844637, + 0.006805721204727888, + 0.0022671374026685953, + 0.007596455980092287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.029794156551361084 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-shot shot, shot from a horizontal angle, captures a quiet corner of an ancient stone courtyard at midnight. Near a stone archway in the stone wall, a pale gray moth with intricate dark markings on its wings hovers softly and erratically, its wings fluttering gently in the air. Dense ivy climbs along the massive, weathered stone walls, its deep green leaves reflecting a faint, moist sheen in the pale, cold moonlight. The moth is evidently attracted to the faint, fading gold leaf embedded in the cracks of the stone wall. The warm, flickering yellow light of an antique lantern hanging in the courtyard catches the leaf, giving it a faint, alluring golden glow. The interplay of cold moonlight and warm lamplight creates a quiet, ancient, and slightly magical midnight atmosphere.", + "seed": 33040, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_5/index_8820_seed_33040_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09229279961436987, + "vlm_logprob": -0.09229279961436987, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.09229279961436987, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.30986806750297546 + ], + "opposite_logprob_tokens": [ + -1.4348680973052979 + ], + "logprob_reward": -0.30986806750297546, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.30986806750297546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18652908504009247, + 0.06257946044206619, + 0.21794743835926056, + 0.034242723137140274, + 0.06014027073979378, + 0.03963801637291908, + 0.05883954092860222, + 0.1103319600224495, + 0.18103021383285522, + 0.0487213209271431 + ], + "credit_map": [ + 0.025510745123028755, + 0.025737633928656578, + 0.020532051101326942, + 0.021850384771823883, + 0.024356579408049583, + 0.02216220274567604, + 0.026149064302444458, + 0.02023041434586048, + 0.008407529443502426, + 0.005818225909024477, + 0.0076528145000338554, + 0.006724235136061907, + 0.00550304539501667, + 0.010126600041985512, + 0.008010173216462135, + 0.010336838662624359, + 0.03425000607967377, + 0.02378489449620247, + 0.022158635780215263, + 0.019920656457543373, + 0.026203956454992294, + 0.029184382408857346, + 0.03818989172577858, + 0.024255022406578064, + 0.002409526379778981, + 0.0045576333068311214, + 0.003042124444618821, + 0.003041717689484358, + 0.004582400433719158, + 0.0068378145806491375, + 0.0032092281617224216, + 0.006562276277691126, + 0.005404282361268997, + 0.006813778076320887, + 0.0060367826372385025, + 0.006105673499405384, + 0.00993911549448967, + 0.007418957073241472, + 0.007106403820216656, + 0.011315278708934784, + 0.0028964681550860405, + 0.002877680119127035, + 0.0051172710955142975, + 0.005069680977612734, + 0.0032184787560254335, + 0.00813240371644497, + 0.005700348410755396, + 0.00662568025290966, + 0.006758075673133135, + 0.005578743293881416, + 0.008607225492596626, + 0.009204060770571232, + 0.0048700300976634026, + 0.006435778457671404, + 0.008184967562556267, + 0.009200659580528736, + 0.010207140818238258, + 0.01371692679822445, + 0.015022841282188892, + 0.015260153450071812, + 0.01264029648154974, + 0.016324909403920174, + 0.010951525531709194, + 0.016208168119192123, + 0.017875174060463905, + 0.013582364656031132, + 0.014242695644497871, + 0.012940596789121628, + 0.010347990319132805, + 0.045952629297971725, + 0.008018271997570992, + 0.05807049572467804, + 0.0025380761362612247, + 0.006855817511677742, + 0.0031733685173094273, + 0.006366754416376352, + 0.006630025804042816, + 0.008246229961514473, + 0.007479928899556398, + 0.0074311220087111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.30986806750297546 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029822388663887978 + ], + "opposite_logprob_tokens": [ + -3.52982234954834 + ], + "logprob_reward": -0.029822388663887978, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029822388663887978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4366268217563629, + 0.046564649790525436, + 0.1915004551410675, + 0.015287584625184536, + 0.015195879153907299, + 0.05777941644191742, + 0.06595657020807266, + 0.07173322886228561, + 0.08382799476385117, + 0.015527427196502686 + ], + "credit_map": [ + 0.012527585960924625, + 0.023104088380932808, + 0.009727944619953632, + 0.011389763094484806, + 0.019986562430858612, + 0.1811201274394989, + 0.015455340966582298, + 0.16331535577774048, + 0.005237830337136984, + 0.0060266489163041115, + 0.003166325856000185, + 0.004009720869362354, + 0.005411186721175909, + 0.00954445730894804, + 0.003869911888614297, + 0.009298562072217464, + 0.03298027440905571, + 0.017892874777317047, + 0.020579375326633453, + 0.02036057412624359, + 0.01925472356379032, + 0.022036835551261902, + 0.03507717698812485, + 0.02331860549747944, + 0.001046702265739441, + 0.001919012633152306, + 0.0012505658669397235, + 0.0018107183277606964, + 0.002853659214451909, + 0.0030891422647982836, + 0.0015081360470503569, + 0.0018096452113240957, + 0.0012441411381587386, + 0.0018282296368852258, + 0.001289797481149435, + 0.001852384302765131, + 0.0018497457494959235, + 0.001916876994073391, + 0.0024617016315460205, + 0.0027529997751116753, + 0.0023739859461784363, + 0.0040273042395710945, + 0.0030727325938642025, + 0.005000295117497444, + 0.0041381181217730045, + 0.017339089885354042, + 0.006186941172927618, + 0.01564093865454197, + 0.007097627501934767, + 0.006695988122373819, + 0.0073135849088430405, + 0.008904831483960152, + 0.006598933134227991, + 0.008469902910292149, + 0.013242047280073166, + 0.0076336488127708435, + 0.004493973217904568, + 0.009194765239953995, + 0.007278953213244677, + 0.009379101917147636, + 0.009025136940181255, + 0.01068874727934599, + 0.009805120527744293, + 0.011867428198456764, + 0.012209098786115646, + 0.007974526844918728, + 0.008250171318650246, + 0.0067155552096664906, + 0.005251988768577576, + 0.016837527975440025, + 0.004040343686938286, + 0.022548779845237732, + 0.0010466029634699225, + 0.0026692566461861134, + 0.0019583816174417734, + 0.0022142897360026836, + 0.0018921184819191694, + 0.0015474216779693961, + 0.0014914058847352862, + 0.002707947976887226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.029822388663887978 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083112098276615 + ], + "opposite_logprob_tokens": [ + -4.511083126068115 + ], + "logprob_reward": -0.011083112098276615, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083112098276615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47173911333084106, + 0.038552865386009216, + 0.07559715211391449, + 0.029757028445601463, + 0.07640775293111801, + 0.06813425570726395, + 0.051899444311857224, + 0.12126591801643372, + 0.03105776570737362, + 0.03558872640132904 + ], + "credit_map": [ + 0.016680147498846054, + 0.02281656116247177, + 0.013074610382318497, + 0.0111722806468606, + 0.01949184760451317, + 0.19632408022880554, + 0.013955524191260338, + 0.17822404205799103, + 0.003559718606993556, + 0.004203754477202892, + 0.0036437243688851595, + 0.005761375185102224, + 0.005425183568149805, + 0.004902299493551254, + 0.005877992138266563, + 0.005178818013519049, + 0.011613051407039165, + 0.007381953299045563, + 0.006909865420311689, + 0.007335409987717867, + 0.008701014332473278, + 0.009227395057678223, + 0.014899170957505703, + 0.009529289789497852, + 0.0020023747347295284, + 0.0056283059529960155, + 0.002283487468957901, + 0.003744716290384531, + 0.00491356011480093, + 0.004221738316118717, + 0.0029033781029284, + 0.0040594665333628654, + 0.0036202550400048494, + 0.009294787421822548, + 0.0038442627992480993, + 0.008616101928055286, + 0.009855290874838829, + 0.011447388678789139, + 0.010480627417564392, + 0.01924903318285942, + 0.002795993350446224, + 0.004230694379657507, + 0.002983710030093789, + 0.005906125530600548, + 0.004750785883516073, + 0.020293300971388817, + 0.006946572568267584, + 0.02022707276046276, + 0.005794895347207785, + 0.0058600702323019505, + 0.005292880814522505, + 0.008422141894698143, + 0.005597150418907404, + 0.00645669549703598, + 0.007071724161505699, + 0.007403881289064884, + 0.007489674258977175, + 0.015074958093464375, + 0.008060595951974392, + 0.017542200163006783, + 0.015988918021321297, + 0.01900297775864601, + 0.020312566310167313, + 0.017794020473957062, + 0.00395516213029623, + 0.004362733568996191, + 0.003441722597926855, + 0.003364815842360258, + 0.0027101102750748396, + 0.007010715082287788, + 0.002164914971217513, + 0.004047588910907507, + 0.0023110867477953434, + 0.0037823577877134085, + 0.0026513803750276566, + 0.004799010697752237, + 0.005179208237677813, + 0.006521856877952814, + 0.004884791094809771, + 0.005459035746753216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.011083112098276615 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062042009085416794 + ], + "opposite_logprob_tokens": [ + -2.812041997909546 + ], + "logprob_reward": -0.062042009085416794, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062042009085416794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.269092321395874, + 0.0951676219701767, + 0.13315829634666443, + 0.0541338250041008, + 0.07525043189525604, + 0.043242260813713074, + 0.08247438073158264, + 0.06972703337669373, + 0.13948126137256622, + 0.03827260062098503 + ], + "credit_map": [ + 0.011598641984164715, + 0.015990737825632095, + 0.009820785373449326, + 0.01400477159768343, + 0.014929498545825481, + 0.09814564883708954, + 0.014233939349651337, + 0.09036829322576523, + 0.01126084104180336, + 0.016722559928894043, + 0.008927864953875542, + 0.009981337934732437, + 0.009964813478291035, + 0.014212332665920258, + 0.011099984869360924, + 0.012997884303331375, + 0.02618362382054329, + 0.014662851579487324, + 0.018827004358172417, + 0.013985207304358482, + 0.013865547254681587, + 0.01651727966964245, + 0.014001256786286831, + 0.015115523710846901, + 0.007887857034802437, + 0.005962049122899771, + 0.0065947845578193665, + 0.006941302679479122, + 0.006167104933410883, + 0.00697975093498826, + 0.005314357113093138, + 0.008286619558930397, + 0.00823234859853983, + 0.00821510050445795, + 0.012660720385611057, + 0.008130538277328014, + 0.00719836400821805, + 0.009425684809684753, + 0.008039461448788643, + 0.013348209671676159, + 0.0022172024473547935, + 0.003841473488137126, + 0.00493053300306201, + 0.004442058503627777, + 0.00415309751406312, + 0.008849937468767166, + 0.006734199356287718, + 0.008073756471276283, + 0.009304615668952465, + 0.008504577912390232, + 0.013883105479180813, + 0.013949613086879253, + 0.006586448755115271, + 0.008373680524528027, + 0.0109777906909585, + 0.010894548147916794, + 0.0065717571415007114, + 0.007104273419827223, + 0.012846803292632103, + 0.00748772406950593, + 0.007194627076387405, + 0.00694911228492856, + 0.01049910206347704, + 0.011073638685047626, + 0.02286941558122635, + 0.023409537971019745, + 0.01863567717373371, + 0.0117295878008008, + 0.010627204552292824, + 0.021412616595625877, + 0.01099727489054203, + 0.01979995332658291, + 0.0018928282661363482, + 0.0070337774232029915, + 0.0034716627560555935, + 0.00388703728094697, + 0.00487118773162365, + 0.004783282522112131, + 0.004662510938942432, + 0.0076703159138560295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.062042009085416794 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048648420721292496 + ], + "opposite_logprob_tokens": [ + -3.0486483573913574 + ], + "logprob_reward": -0.048648420721292496, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048648420721292496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29815566539764404, + 0.07729391753673553, + 0.3011136054992676, + 0.044165994971990585, + 0.030978359282016754, + 0.036433205008506775, + 0.08605457842350006, + 0.05807309225201607, + 0.044198084622621536, + 0.023533381521701813 + ], + "credit_map": [ + 0.012990457937121391, + 0.015595842152833939, + 0.010425561107695103, + 0.009877033531665802, + 0.013326839543879032, + 0.1208772212266922, + 0.009052186273038387, + 0.1060105636715889, + 0.005582891870290041, + 0.014005540870130062, + 0.005601424723863602, + 0.01001735869795084, + 0.009779328480362892, + 0.01175880990922451, + 0.011380129493772984, + 0.009168445132672787, + 0.058159563690423965, + 0.032353680580854416, + 0.03638535365462303, + 0.030095428228378296, + 0.03189142420887947, + 0.03687235713005066, + 0.036407340317964554, + 0.03894849494099617, + 0.004292563069611788, + 0.006018105428665876, + 0.00497916666790843, + 0.00447180075570941, + 0.005491144955158234, + 0.006679703947156668, + 0.003997684456408024, + 0.008235825225710869, + 0.002374754287302494, + 0.004852378740906715, + 0.00289948470890522, + 0.003730922471731901, + 0.003465555142611265, + 0.004266277886927128, + 0.004151711706072092, + 0.005237277131527662, + 0.002802640898153186, + 0.003753854427486658, + 0.0033193205017596483, + 0.005903241224586964, + 0.0037137989420443773, + 0.004484704229980707, + 0.007140250410884619, + 0.005315396934747696, + 0.008429406210780144, + 0.009067011065781116, + 0.00739939883351326, + 0.014497926458716393, + 0.009732276201248169, + 0.010651244781911373, + 0.013311837799847126, + 0.012965483590960503, + 0.0036781521048396826, + 0.0068704658187925816, + 0.00418957369402051, + 0.008912607096135616, + 0.007113123778253794, + 0.00791264045983553, + 0.010668784379959106, + 0.008727751672267914, + 0.004608989227563143, + 0.004847036674618721, + 0.005993302445858717, + 0.0069331033155322075, + 0.004031817894428968, + 0.00667296489700675, + 0.00453585060313344, + 0.006575022824108601, + 0.0014230908127501607, + 0.004605940543115139, + 0.00154632527846843, + 0.0037099323235452175, + 0.0022814315743744373, + 0.0025414307601749897, + 0.0028834633994847536, + 0.004541770555078983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.048648420721292496 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-shot shot, shot from a horizontal angle, captures a quiet corner of an ancient stone courtyard at midnight. Near a stone archway in the stone wall, a pale gray moth with intricate dark markings on its wings hovers softly and erratically, its wings fluttering gently in the air. Dense ivy climbs along the massive, weathered stone walls, its deep green leaves reflecting a faint, moist sheen in the pale, cold moonlight. The moth is evidently attracted to the faint, fading gold leaf embedded in the cracks of the stone wall. The warm, flickering yellow light of an antique lantern hanging in the courtyard catches the leaf, giving it a faint, alluring golden glow. The interplay of cold moonlight and warm lamplight creates a quiet, ancient, and slightly magical midnight atmosphere.", + "seed": 33041, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_5/index_8820_seed_33041_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2086266942322254, + "vlm_logprob": -0.2086266942322254, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.2086266942322254, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7089343070983887 + ], + "opposite_logprob_tokens": [ + -0.7089343070983887 + ], + "logprob_reward": -0.7089343070983887, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7089343070983887, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14087967574596405, + 0.05041917786002159, + 0.0588187575340271, + 0.14790715277194977, + 0.10677335411310196, + 0.049697697162628174, + 0.08695360273122787, + 0.2080247551202774, + 0.11987418681383133, + 0.030651630833745003 + ], + "credit_map": [ + 0.015823550522327423, + 0.013665665872395039, + 0.02878696657717228, + 0.01227097399532795, + 0.013769451528787613, + 0.015027950517833233, + 0.026605602353811264, + 0.014929518103599548, + 0.006321212276816368, + 0.006228136830031872, + 0.010217562317848206, + 0.003964060451835394, + 0.005485294386744499, + 0.004455187823623419, + 0.005311844404786825, + 0.008435879833996296, + 0.009433561936020851, + 0.004338597878813744, + 0.014047864824533463, + 0.004999714903533459, + 0.005657816305756569, + 0.0042229024693369865, + 0.008855358697474003, + 0.0072629451751708984, + 0.026208750903606415, + 0.013452843762934208, + 0.037273429334163666, + 0.008203989826142788, + 0.012790829874575138, + 0.010329270735383034, + 0.011106652207672596, + 0.028541380539536476, + 0.012609865516424179, + 0.010042808018624783, + 0.020135387778282166, + 0.00918495561927557, + 0.014488589018583298, + 0.008937079459428787, + 0.01660696044564247, + 0.014767716638743877, + 0.004901721142232418, + 0.006580413784831762, + 0.007104759570211172, + 0.006094388198107481, + 0.006559353321790695, + 0.005833932664245367, + 0.006887048948556185, + 0.00573608186095953, + 0.01013017725199461, + 0.00977584719657898, + 0.012250643223524094, + 0.007857058197259903, + 0.013042014092206955, + 0.007697773166000843, + 0.01490221731364727, + 0.011297868564724922, + 0.018833784386515617, + 0.015849772840738297, + 0.025512561202049255, + 0.018867261707782745, + 0.03395025432109833, + 0.014504853636026382, + 0.04923994839191437, + 0.03126631677150726, + 0.014877578243613243, + 0.015820564702153206, + 0.017426397651433945, + 0.013734487816691399, + 0.01747024618089199, + 0.010635792277753353, + 0.015929488465189934, + 0.013979632407426834, + 0.004529984667897224, + 0.005174985155463219, + 0.0038441510405391455, + 0.004075806587934494, + 0.0035527944564819336, + 0.002998161828145385, + 0.0030305087566375732, + 0.0034452383406460285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.7089343070983887 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157891511917114 + ], + "opposite_logprob_tokens": [ + -1.7015788555145264 + ], + "logprob_reward": -0.20157891511917114, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157891511917114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17931072413921356, + 0.19253356754779816, + 0.04510590434074402, + 0.058867618441581726, + 0.17911170423030853, + 0.06651758402585983, + 0.02356940507888794, + 0.181386798620224, + 0.04455409571528435, + 0.029042642563581467 + ], + "credit_map": [ + 0.03019554354250431, + 0.013571232557296753, + 0.05224686115980148, + 0.011898951604962349, + 0.018980827182531357, + 0.013789080083370209, + 0.016535216942429543, + 0.02209300920367241, + 0.03581489995121956, + 0.027902886271476746, + 0.04465717449784279, + 0.017815090715885162, + 0.01834360882639885, + 0.009293037466704845, + 0.013588366098701954, + 0.02511848881840706, + 0.0069072614423930645, + 0.0035473406314849854, + 0.012001815252006054, + 0.003807994769886136, + 0.0031012974213808775, + 0.0033126643393188715, + 0.006534948479384184, + 0.005892580840736628, + 0.006038763094693422, + 0.004973013419657946, + 0.011041945777833462, + 0.005291472189128399, + 0.008137212134897709, + 0.005898077040910721, + 0.0077308304607868195, + 0.009756302461028099, + 0.025596441701054573, + 0.017869316041469574, + 0.03847387060523033, + 0.016687579452991486, + 0.02229192666709423, + 0.01348805520683527, + 0.02146069146692753, + 0.023243816569447517, + 0.012922799214720726, + 0.012105843052268028, + 0.011063738726079464, + 0.010116741992533207, + 0.00558316009119153, + 0.004649150650948286, + 0.005680441856384277, + 0.004395715426653624, + 0.0028990369755774736, + 0.0027787811122834682, + 0.003676919499412179, + 0.0023986923042684793, + 0.00393965607509017, + 0.0020106141455471516, + 0.0033472077921032906, + 0.002518497873097658, + 0.019279899075627327, + 0.012952613644301891, + 0.023962682113051414, + 0.016803229227662086, + 0.029681768268346786, + 0.013178931549191475, + 0.04356309399008751, + 0.021964579820632935, + 0.004872316960245371, + 0.0053826551884412766, + 0.0070433360524475574, + 0.0050757951103150845, + 0.005997980944812298, + 0.004576269071549177, + 0.005683244206011295, + 0.005922500044107437, + 0.00505266897380352, + 0.003738813102245331, + 0.005581636913120747, + 0.0031643167603760958, + 0.0024956692941486835, + 0.001956833293661475, + 0.0035563893616199493, + 0.003496316960081458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.20157891511917114 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003619549795985222 + ], + "opposite_logprob_tokens": [ + -5.62861967086792 + ], + "logprob_reward": -0.003619549795985222, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003619549795985222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18439918756484985, + 0.15051542222499847, + 0.05653325840830803, + 0.0921296700835228, + 0.10637274384498596, + 0.029460858553647995, + 0.03317606821656227, + 0.18778371810913086, + 0.0872928649187088, + 0.07233627885580063 + ], + "credit_map": [ + 0.02226870320737362, + 0.020368343219161034, + 0.02322157844901085, + 0.021672194823622704, + 0.02421068586409092, + 0.021363483741879463, + 0.030761059373617172, + 0.02053312212228775, + 0.024230141192674637, + 0.019651059061288834, + 0.026661718264222145, + 0.01388288103044033, + 0.016170764341950417, + 0.013386022299528122, + 0.01804058812558651, + 0.018492233008146286, + 0.006590743083506823, + 0.0074735903181135654, + 0.008760886266827583, + 0.0057320380583405495, + 0.007369655184447765, + 0.0065738786943256855, + 0.007043804507702589, + 0.0069886548444628716, + 0.013361208140850067, + 0.00805006641894579, + 0.02603476494550705, + 0.00654665008187294, + 0.00839990098029375, + 0.006695827003568411, + 0.006637536454945803, + 0.01640371046960354, + 0.01200886256992817, + 0.013640635646879673, + 0.015079298056662083, + 0.011424867436289787, + 0.015148001722991467, + 0.008055181242525578, + 0.017095502465963364, + 0.013920383527874947, + 0.002450139494612813, + 0.0029577501118183136, + 0.004732239060103893, + 0.001956841442734003, + 0.005149392411112785, + 0.004039018880575895, + 0.004313694313168526, + 0.003861780045554042, + 0.003801517654210329, + 0.004220927599817514, + 0.004793941974639893, + 0.0037189272698014975, + 0.004204049706459045, + 0.003674016799777746, + 0.0053266799077391624, + 0.0034360066056251526, + 0.0172830019146204, + 0.01475573517382145, + 0.023349637165665627, + 0.01902584731578827, + 0.030342839658260345, + 0.012990758754312992, + 0.04544152319431305, + 0.024594368413090706, + 0.010726033709943295, + 0.012227790430188179, + 0.012014660984277725, + 0.011160347610712051, + 0.013184239156544209, + 0.008225565776228905, + 0.011033189482986927, + 0.008721026591956615, + 0.013242200948297977, + 0.008047838695347309, + 0.01713024079799652, + 0.006117192097008228, + 0.005606135819107294, + 0.005867012310773134, + 0.0051779127679765224, + 0.011147736571729183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.003619549795985222 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293596774339676 + ], + "opposite_logprob_tokens": [ + -2.237936019897461 + ], + "logprob_reward": -0.11293596774339676, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293596774339676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09133594483137131, + 0.04146290570497513, + 0.03977920487523079, + 0.08696182072162628, + 0.09147288650274277, + 0.05953333526849747, + 0.05339093133807182, + 0.47396621108055115, + 0.022040488198399544, + 0.040056247264146805 + ], + "credit_map": [ + 0.009345960803329945, + 0.00851140171289444, + 0.01618083193898201, + 0.0069144791923463345, + 0.011600710451602936, + 0.011650124564766884, + 0.015053702518343925, + 0.012078727595508099, + 0.0052157738246023655, + 0.004771094769239426, + 0.0076453667134046555, + 0.0037782075814902782, + 0.00489744171500206, + 0.0050430926494300365, + 0.0050310599617660046, + 0.005080866161733866, + 0.004376403521746397, + 0.0036179516464471817, + 0.00822298415005207, + 0.003100157016888261, + 0.0045529925264418125, + 0.005369389895349741, + 0.003569582011550665, + 0.006969737354665995, + 0.014743300154805183, + 0.007628392893821001, + 0.021413473412394524, + 0.004908588714897633, + 0.00720051396638155, + 0.007097424007952213, + 0.00806634221225977, + 0.01590377278625965, + 0.013124645687639713, + 0.006814535707235336, + 0.02393265627324581, + 0.006407422944903374, + 0.011947550810873508, + 0.008333588019013405, + 0.007690468803048134, + 0.013222008012235165, + 0.005890505388379097, + 0.005472523160278797, + 0.008793909102678299, + 0.004654139745980501, + 0.01101311668753624, + 0.008004097267985344, + 0.009385013952851295, + 0.006320025306195021, + 0.009197281673550606, + 0.004599114879965782, + 0.012462767772376537, + 0.0051239957101643085, + 0.005805948283523321, + 0.005443680565804243, + 0.006613636389374733, + 0.004144494887441397, + 0.040163516998291016, + 0.03593830019235611, + 0.055672284215688705, + 0.04746105894446373, + 0.07833281904459, + 0.032506491988897324, + 0.1215631440281868, + 0.062328554689884186, + 0.004171330481767654, + 0.0018010132480412722, + 0.006208169739693403, + 0.0011219293810427189, + 0.0028403911273926497, + 0.0018394183134660125, + 0.0018948160577565432, + 0.0021634171716868877, + 0.006906562019139528, + 0.003470177296549082, + 0.009777805767953396, + 0.0036033373326063156, + 0.003993851598352194, + 0.0034315306693315506, + 0.004665695130825043, + 0.004207281395792961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.11293596774339676 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient stone courtyard at midnight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pale gray moth with intricate dark markings on its wings visible hovering near a stone archway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dense deep green ivy leaves climbing the massive weathered stone walls with a moist sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a faint, fading gold leaf embedded in the cracks of the stone wall illuminated by a warm flickering yellow antique lantern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured as a static mid-shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064731404185295 + ], + "opposite_logprob_tokens": [ + -4.141064643859863 + ], + "logprob_reward": -0.016064731404185295, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064731404185295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21713696420192719, + 0.05136136710643768, + 0.041012972593307495, + 0.10942337661981583, + 0.11341829597949982, + 0.1047639325261116, + 0.048531513661146164, + 0.09810896962881088, + 0.04986269026994705, + 0.166379913687706 + ], + "credit_map": [ + 0.03977882117033005, + 0.020324762910604477, + 0.054055195301771164, + 0.017493603751063347, + 0.01937420666217804, + 0.022562969475984573, + 0.024324754253029823, + 0.019222650676965714, + 0.00799976009875536, + 0.004194192122668028, + 0.014322693459689617, + 0.003982113674283028, + 0.005311248358339071, + 0.003936349414288998, + 0.004836367908865213, + 0.006778644863516092, + 0.0055158259347081184, + 0.005152045749127865, + 0.00549029977992177, + 0.003783601801842451, + 0.0056001534685492516, + 0.005248758476227522, + 0.004547737538814545, + 0.0056745512410998344, + 0.020378705114126205, + 0.011122571304440498, + 0.028540436178445816, + 0.009191934950649738, + 0.00848705880343914, + 0.005850858520716429, + 0.009474446065723896, + 0.016377372667193413, + 0.015978902578353882, + 0.011901029385626316, + 0.023423558101058006, + 0.01143315527588129, + 0.015187895856797695, + 0.008820699527859688, + 0.013265702873468399, + 0.013407357968389988, + 0.013992786407470703, + 0.011413772590458393, + 0.019482316449284554, + 0.012742995284497738, + 0.012127288617193699, + 0.009443461894989014, + 0.013697905465960503, + 0.011863408610224724, + 0.007092229090631008, + 0.0045278118923306465, + 0.008345695212483406, + 0.0049079409800469875, + 0.008290492929518223, + 0.005568501073867083, + 0.005773372948169708, + 0.0040254732593894005, + 0.0112025560811162, + 0.007709880825132132, + 0.014697804115712643, + 0.009302127175033092, + 0.014272698201239109, + 0.009213285520672798, + 0.020358024165034294, + 0.0113525977358222, + 0.006257335655391216, + 0.007060611620545387, + 0.007018911652266979, + 0.005445001181215048, + 0.009179510176181793, + 0.004189569037407637, + 0.005792511627078056, + 0.00491924025118351, + 0.025537490844726562, + 0.021283315494656563, + 0.025740675628185272, + 0.015020143240690231, + 0.01765538565814495, + 0.017837580293416977, + 0.012579968199133873, + 0.03072535991668701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.016064731404185295 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_6/index_tensor([143, 143])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_6/index_tensor([143, 143])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b8b7a4ed8886566ffaa13e6e165726e9cead7222 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_6/index_tensor([143, 143])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "striking overhead close-up shot strips the image of all color, revealing a stark contrast of black and white. Against a stark black void, a sharp, piercing overhead light shines vertically, illuminating a massive white tiger with the precision of a stage spotlight. The high-contrast light sculpts its muscular features, its snow-white fur appearing almost blown out in the highlights, while its deep black stripes and shadows are submerged in pure darkness. Its icy blue eyes gleam with a fierce and cold brilliance in the light, its maw stretched to its limit, revealing dagger-sharp canines and a barbed tongue, as if unleashing a silent yet terrifying roar. In an instant, it exerted force with its hind legs, lowered its entire body like a white arrow shot from a bow, and with primitive wild power, it stretched out its fully opened deadly claws and rushed straight towards the camera, creating a thrilling and tense atmosphere full of primitive power and deadly threats.", + "seed": 28535, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_6/index_143_seed_28535_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13632746888324618, + "vlm_logprob": -0.13632746888324618, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.13632746888324618, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3868768513202667 + ], + "opposite_logprob_tokens": [ + -1.2618768215179443 + ], + "logprob_reward": -0.3868768513202667, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3868768513202667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06411108374595642, + 0.25208213925361633, + 0.035576291382312775, + 0.1304083913564682, + 0.10383439809083939, + 0.060007598251104355, + 0.11411339789628983, + 0.11595993489027023, + 0.08241641521453857, + 0.04149027541279793 + ], + "credit_map": [ + 0.004527967423200607, + 0.0071959164924919605, + 0.005785076878964901, + 0.007658699993044138, + 0.008181547746062279, + 0.00508689833804965, + 0.01733056642115116, + 0.008344410918653011, + 0.014564762823283672, + 0.0268477164208889, + 0.024333294481039047, + 0.038495566695928574, + 0.0368177704513073, + 0.02176610939204693, + 0.06542961299419403, + 0.0238273274153471, + 0.0038856726605445147, + 0.003634790424257517, + 0.005982804577797651, + 0.006804302334785461, + 0.0028109601698815823, + 0.0043095918372273445, + 0.0040536741726100445, + 0.004094491712749004, + 0.014303062111139297, + 0.01602821610867977, + 0.01494140736758709, + 0.018984897062182426, + 0.017570339143276215, + 0.013084415346384048, + 0.02118673361837864, + 0.014309320598840714, + 0.027515605092048645, + 0.0061759185045957565, + 0.014407480135560036, + 0.008520633913576603, + 0.00788673385977745, + 0.008918646723031998, + 0.005655187182128429, + 0.02475418522953987, + 0.008208762854337692, + 0.005939109716564417, + 0.009151487611234188, + 0.004591227043420076, + 0.00871539581567049, + 0.007009271997958422, + 0.006826198194175959, + 0.00956614688038826, + 0.013256910257041454, + 0.01570037007331848, + 0.0219305157661438, + 0.013949858956038952, + 0.012862498871982098, + 0.01464454922825098, + 0.011460901238024235, + 0.010307792574167252, + 0.006059290375560522, + 0.005307454615831375, + 0.007901445031166077, + 0.009735413827002048, + 0.005663808900862932, + 0.03859132155776024, + 0.0046092732809484005, + 0.038091931492090225, + 0.005416306667029858, + 0.010903913527727127, + 0.006921278778463602, + 0.013987149111926556, + 0.006623739842325449, + 0.012626461684703827, + 0.0098121901974082, + 0.016125373542308807, + 0.00369563908316195, + 0.007349388208240271, + 0.003641068935394287, + 0.008668730966746807, + 0.005689125508069992, + 0.005930568557232618, + 0.004014989826828241, + 0.00250076362863183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -0.3868768513202667 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03816591948270798 + ], + "opposite_logprob_tokens": [ + -3.288165807723999 + ], + "logprob_reward": -0.03816591948270798, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03816591948270798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14470286667346954, + 0.13066519796848297, + 0.02983149141073227, + 0.24916869401931763, + 0.03503917157649994, + 0.0915432870388031, + 0.08131292462348938, + 0.050773702561855316, + 0.04904802143573761, + 0.13791465759277344 + ], + "credit_map": [ + 0.013693362474441528, + 0.011046618223190308, + 0.02148289419710636, + 0.014858556911349297, + 0.016064662486314774, + 0.010956482961773872, + 0.04044131189584732, + 0.01615898683667183, + 0.008757088333368301, + 0.01678653620183468, + 0.02301821857690811, + 0.014254504814743996, + 0.02687039040029049, + 0.01155536063015461, + 0.01660335622727871, + 0.012819739989936352, + 0.0029680042061954737, + 0.003974419552832842, + 0.003470922354608774, + 0.005807347595691681, + 0.00362513680011034, + 0.0032129937317222357, + 0.0030243094079196453, + 0.003748357994481921, + 0.026869578287005424, + 0.02382112480700016, + 0.03256125748157501, + 0.04823474586009979, + 0.03147038817405701, + 0.01687166653573513, + 0.041772667318582535, + 0.027567271143198013, + 0.006459527648985386, + 0.0030493137892335653, + 0.0040597557090222836, + 0.0038235224783420563, + 0.0042703100480139256, + 0.004545285366475582, + 0.002594348043203354, + 0.006237108260393143, + 0.01613517664372921, + 0.005389702972024679, + 0.034826505929231644, + 0.005718388129025698, + 0.005886048544198275, + 0.006447860971093178, + 0.0055610425770282745, + 0.01157856173813343, + 0.008333627134561539, + 0.012265956029295921, + 0.01675933226943016, + 0.008512230589985847, + 0.008626691065728664, + 0.010377265512943268, + 0.007023421581834555, + 0.009414400905370712, + 0.004497692920267582, + 0.004237235523760319, + 0.004623610060662031, + 0.007802954409271479, + 0.00530356215313077, + 0.010194622911512852, + 0.0037187652196735144, + 0.010395259596407413, + 0.0030937595292925835, + 0.012486043386161327, + 0.00454016262665391, + 0.007701461669057608, + 0.006631563883274794, + 0.005974403582513332, + 0.004148054402321577, + 0.004472571890801191, + 0.007818996906280518, + 0.019069302827119827, + 0.014107239432632923, + 0.035755742341279984, + 0.019223598763346672, + 0.014568228274583817, + 0.01319397334009409, + 0.014177580364048481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.03816591948270798 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647999536246061 + ], + "opposite_logprob_tokens": [ + -5.379648208618164 + ], + "logprob_reward": -0.004647999536246061, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647999536246061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07443181425333023, + 0.1218879222869873, + 0.04608801752328873, + 0.1349698305130005, + 0.06421008706092834, + 0.0470522865653038, + 0.08387754112482071, + 0.09179271012544632, + 0.13124164938926697, + 0.20444810390472412 + ], + "credit_map": [ + 0.008303612470626831, + 0.00833681970834732, + 0.010285813361406326, + 0.010909896343946457, + 0.011489867232739925, + 0.006010036915540695, + 0.009867437183856964, + 0.009228331968188286, + 0.008682342246174812, + 0.009625295177102089, + 0.025870176032185555, + 0.008878730237483978, + 0.025686470791697502, + 0.010225383564829826, + 0.01700221374630928, + 0.01591731235384941, + 0.004413570277392864, + 0.005561805795878172, + 0.007568400353193283, + 0.008415808901190758, + 0.005022301338613033, + 0.0049148062244057655, + 0.004830934107303619, + 0.005360394716262817, + 0.013874625787138939, + 0.014088071882724762, + 0.01769564300775528, + 0.023390071466565132, + 0.01827899180352688, + 0.009924852289259434, + 0.024112675338983536, + 0.013604901731014252, + 0.009209557436406612, + 0.005769388750195503, + 0.007631425745785236, + 0.007052416447550058, + 0.01294842828065157, + 0.00898993294686079, + 0.0049789827316999435, + 0.007629963103681803, + 0.008154296316206455, + 0.0038161464035511017, + 0.01233487855643034, + 0.004271568730473518, + 0.004315861035138369, + 0.005843475926667452, + 0.003970768768340349, + 0.004345293622463942, + 0.011172300204634666, + 0.00992704089730978, + 0.008455566130578518, + 0.008555429987609386, + 0.008345814421772957, + 0.011214742437005043, + 0.009496754966676235, + 0.016709892079234123, + 0.007960853166878223, + 0.006810626480728388, + 0.014585684984922409, + 0.010179249569773674, + 0.010598455555737019, + 0.016448507085442543, + 0.00782643910497427, + 0.017382889986038208, + 0.008061579428613186, + 0.019620366394519806, + 0.01291085034608841, + 0.02477443404495716, + 0.012594998814165592, + 0.01803591288626194, + 0.012117166072130203, + 0.02312633767724037, + 0.016453133895993233, + 0.031050635501742363, + 0.02084898203611374, + 0.045002471655607224, + 0.02583998255431652, + 0.02069004438817501, + 0.02137029357254505, + 0.02319256030023098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.004647999536246061 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559596598148346 + ], + "opposite_logprob_tokens": [ + -1.6005959510803223 + ], + "logprob_reward": -0.22559596598148346, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559596598148346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062307823449373245, + 0.1376786231994629, + 0.07059112191200256, + 0.12991443276405334, + 0.09960506111383438, + 0.08716907352209091, + 0.05259973183274269, + 0.109891876578331, + 0.11821673065423965, + 0.13202552497386932 + ], + "credit_map": [ + 0.006737304851412773, + 0.006918949540704489, + 0.007868411019444466, + 0.007438130676746368, + 0.007063786964863539, + 0.007392969448119402, + 0.008731476031243801, + 0.010156798176467419, + 0.010263459756970406, + 0.012065714225172997, + 0.017987364903092384, + 0.015643686056137085, + 0.01422921009361744, + 0.015232794918119907, + 0.02908160723745823, + 0.02317478507757187, + 0.006579448003321886, + 0.007251959294080734, + 0.012804185040295124, + 0.012541357427835464, + 0.005911668296903372, + 0.006038289051502943, + 0.007687355857342482, + 0.011776861734688282, + 0.014280511066317558, + 0.01586627960205078, + 0.012966521084308624, + 0.02022615261375904, + 0.01933370903134346, + 0.011638378724455833, + 0.022672822698950768, + 0.01293005608022213, + 0.024274352937936783, + 0.005920848343521357, + 0.011989145539700985, + 0.007234733086079359, + 0.011551156640052795, + 0.010241048410534859, + 0.005232750438153744, + 0.0231610294431448, + 0.012005878612399101, + 0.007642380427569151, + 0.017982609570026398, + 0.007373561151325703, + 0.011905699968338013, + 0.009656662121415138, + 0.008165388368070126, + 0.012436894699931145, + 0.005490949377417564, + 0.008328882977366447, + 0.010402201674878597, + 0.004495099652558565, + 0.006890841759741306, + 0.006277008913457394, + 0.005780395586043596, + 0.004934350494295359, + 0.014382117427885532, + 0.006328405346721411, + 0.008098609745502472, + 0.007527472916990519, + 0.00595938041806221, + 0.030895989388227463, + 0.005958717782050371, + 0.030741184949874878, + 0.00837523490190506, + 0.01931193470954895, + 0.01101767923682928, + 0.022951625287532806, + 0.016241520643234253, + 0.015904974192380905, + 0.011421781033277512, + 0.012991971336305141, + 0.01036372222006321, + 0.020807407796382904, + 0.014787061139941216, + 0.029756562784314156, + 0.015980862081050873, + 0.013696587644517422, + 0.013278428465127945, + 0.013354893773794174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 3 + ], + "selected_reward": -0.22559596598148346 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350608095526695 + ], + "opposite_logprob_tokens": [ + -3.651350498199463 + ], + "logprob_reward": -0.026350608095526695, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350608095526695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051973193883895874, + 0.12594707310199738, + 0.03378979116678238, + 0.07591637969017029, + 0.03963080048561096, + 0.06509518623352051, + 0.06356191635131836, + 0.03187469020485878, + 0.06740304827690125, + 0.44480791687965393 + ], + "credit_map": [ + 0.004615066107362509, + 0.006026849616318941, + 0.00472192233428359, + 0.006309379357844591, + 0.006287544500082731, + 0.004645354114472866, + 0.011780760250985622, + 0.007586314342916012, + 0.008572109043598175, + 0.012432113289833069, + 0.01377079077064991, + 0.014806537888944149, + 0.013238364830613136, + 0.012485794723033905, + 0.030073752626776695, + 0.020567603409290314, + 0.0029767537489533424, + 0.0037451523821800947, + 0.004024633672088385, + 0.00777506222948432, + 0.003114788793027401, + 0.004484400153160095, + 0.004027525428682566, + 0.0036414763890206814, + 0.008660334162414074, + 0.009189638309180737, + 0.007285073399543762, + 0.013917829841375351, + 0.009800728410482407, + 0.00529316533356905, + 0.013584351167082787, + 0.008185258135199547, + 0.008821681141853333, + 0.002788456855341792, + 0.0036497614346444607, + 0.003957190550863743, + 0.0039077321998775005, + 0.005346210673451424, + 0.0031937193125486374, + 0.007966048084199429, + 0.009929658845067024, + 0.005730667617172003, + 0.014569061808288097, + 0.005214035511016846, + 0.007859596982598305, + 0.005975698120892048, + 0.007226322311908007, + 0.008590144105255604, + 0.0077868010848760605, + 0.013099435716867447, + 0.008089292794466019, + 0.006233456078916788, + 0.008601803332567215, + 0.007105661556124687, + 0.006938994862139225, + 0.0057064746506512165, + 0.002365540713071823, + 0.0019427182851359248, + 0.0024066392797976732, + 0.004781549796462059, + 0.0018744819099083543, + 0.008136841468513012, + 0.002364799380302429, + 0.00800212100148201, + 0.003807326313108206, + 0.010933630168437958, + 0.005906392354518175, + 0.0126034552231431, + 0.007954754866659641, + 0.009533119387924671, + 0.008017280139029026, + 0.008647091686725616, + 0.02867519110441208, + 0.05874714255332947, + 0.042647674679756165, + 0.11910884827375412, + 0.06105926260352135, + 0.046345245093107224, + 0.03915587067604065, + 0.049068666994571686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.026350608095526695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking overhead close-up shot strips the image of all color, revealing a stark contrast of black and white. Against a stark black void, a sharp, piercing overhead light shines vertically, illuminating a massive white tiger with the precision of a stage spotlight. The high-contrast light sculpts its muscular features, its snow-white fur appearing almost blown out in the highlights, while its deep black stripes and shadows are submerged in pure darkness. Its icy blue eyes gleam with a fierce and cold brilliance in the light, its maw stretched to its limit, revealing dagger-sharp canines and a barbed tongue, as if unleashing a silent yet terrifying roar. In an instant, it exerted force with its hind legs, lowered its entire body like a white arrow shot from a bow, and with primitive wild power, it stretched out its fully opened deadly claws and rushed straight towards the camera, creating a thrilling and tense atmosphere full of primitive power and deadly threats.", + "seed": 28536, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_6/index_143_seed_28536_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16327359732240437, + "vlm_logprob": -0.16327359732240437, + "vlm_yes_no_margin": 2.6499999940395353, + "vlm_token_credit": -0.16327359732240437, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23654696345329285 + ], + "opposite_logprob_tokens": [ + -1.7365469932556152 + ], + "logprob_reward": -0.23654696345329285, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.23654696345329285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1257190853357315, + 0.2794971764087677, + 0.2837095260620117, + 0.04513216391205788, + 0.05808400735259056, + 0.017786551266908646, + 0.03425879776477814, + 0.026571180671453476, + 0.05475868284702301, + 0.07448284327983856 + ], + "credit_map": [ + 0.01217926386743784, + 0.01520705409348011, + 0.014048120938241482, + 0.011311357840895653, + 0.016927286982536316, + 0.014504221267998219, + 0.0262192040681839, + 0.015322583727538586, + 0.0431453175842762, + 0.02409931644797325, + 0.04589739441871643, + 0.024248767644166946, + 0.027674587443470955, + 0.022423580288887024, + 0.04599440470337868, + 0.04601380601525307, + 0.03291779011487961, + 0.04134639725089073, + 0.03113522008061409, + 0.02235749550163746, + 0.014285292476415634, + 0.07045037299394608, + 0.012526288628578186, + 0.058690644800662994, + 0.0038415680173784494, + 0.0061436425894498825, + 0.006212251260876656, + 0.010462186299264431, + 0.0028525691013783216, + 0.003758698236197233, + 0.00648838933557272, + 0.005372856743633747, + 0.007946008816361427, + 0.004643486812710762, + 0.007900120690464973, + 0.005548362620174885, + 0.0037198031786829233, + 0.01631953939795494, + 0.004288968164473772, + 0.0077177188359200954, + 0.0023828474804759026, + 0.0014926554867997766, + 0.001700371503829956, + 0.002842053072527051, + 0.0014397158520296216, + 0.002813741099089384, + 0.0019318521954119205, + 0.0031833131797611713, + 0.003330813255161047, + 0.006317760795354843, + 0.0027257846668362617, + 0.004695512354373932, + 0.004440964665263891, + 0.004830481484532356, + 0.003288663923740387, + 0.004628818482160568, + 0.0023967395536601543, + 0.004782402887940407, + 0.0024394362699240446, + 0.006262524519115686, + 0.0026363658253103495, + 0.0025055292062461376, + 0.0025485530495643616, + 0.0029996284283697605, + 0.002668193308636546, + 0.015229631215333939, + 0.0046585300005972385, + 0.014330994337797165, + 0.005658998619765043, + 0.0029813507571816444, + 0.005454149097204208, + 0.0037768345791846514, + 0.006623515859246254, + 0.007352772634476423, + 0.013090969994664192, + 0.008579876273870468, + 0.01005968451499939, + 0.01534203253686428, + 0.0065721492283046246, + 0.006861842703074217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.23654696345329285 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06213343143463135 + ], + "opposite_logprob_tokens": [ + -2.812133312225342 + ], + "logprob_reward": -0.06213343143463135, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06213343143463135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08432565629482269, + 0.14545518159866333, + 0.17330984771251678, + 0.07731269299983978, + 0.019165029749274254, + 0.2720626890659332, + 0.032483723014593124, + 0.07313939183950424, + 0.04766148701310158, + 0.07508422434329987 + ], + "credit_map": [ + 0.008231737650930882, + 0.00862099789083004, + 0.014499276876449585, + 0.009637763723731041, + 0.008975991047918797, + 0.009388139471411705, + 0.01269893255084753, + 0.012272825464606285, + 0.029638443142175674, + 0.014524687081575394, + 0.022836703807115555, + 0.010014171712100506, + 0.015125231817364693, + 0.013139796443283558, + 0.022091709077358246, + 0.018084445968270302, + 0.02061217650771141, + 0.020819159224629402, + 0.010065303184092045, + 0.009088138118386269, + 0.009826133958995342, + 0.043856121599674225, + 0.008896589279174805, + 0.050146233290433884, + 0.009233511984348297, + 0.011356263421475887, + 0.00938148982822895, + 0.013736865483224392, + 0.006098705343902111, + 0.009115161374211311, + 0.010437430813908577, + 0.007953268475830555, + 0.002780392300337553, + 0.0015843764413148165, + 0.0024240324273705482, + 0.002203864511102438, + 0.0013520182110369205, + 0.00462854839861393, + 0.0020946913864463568, + 0.0020971072372049093, + 0.02431649900972843, + 0.02137233130633831, + 0.019304266199469566, + 0.03252607583999634, + 0.021838903427124023, + 0.07033567875623703, + 0.020062383264303207, + 0.06230657175183296, + 0.0025670998729765415, + 0.009477009065449238, + 0.0023976347874850035, + 0.00453812163323164, + 0.003981056623160839, + 0.004728294443339109, + 0.001834013033658266, + 0.002960493788123131, + 0.005537460558116436, + 0.004899540916085243, + 0.004170243628323078, + 0.024464232847094536, + 0.010399620048701763, + 0.011795718222856522, + 0.007629264611750841, + 0.004243313800543547, + 0.0016956928884610534, + 0.015054939314723015, + 0.003188603324815631, + 0.015004999935626984, + 0.0048504299484193325, + 0.0022826686035841703, + 0.003434506943449378, + 0.0021496464032679796, + 0.006544790230691433, + 0.008856745436787605, + 0.01047216635197401, + 0.007585663814097643, + 0.016353365033864975, + 0.01076458115130663, + 0.0067972661927342415, + 0.007709645200520754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.06213343143463135 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005259486846625805 + ], + "opposite_logprob_tokens": [ + -5.2552595138549805 + ], + "logprob_reward": -0.005259486846625805, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005259486846625805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08621867001056671, + 0.09452719986438751, + 0.3402808904647827, + 0.06789664179086685, + 0.05279620736837387, + 0.08320406079292297, + 0.13364708423614502, + 0.03284870833158493, + 0.047180622816085815, + 0.0613999105989933 + ], + "credit_map": [ + 0.009015745483338833, + 0.010732620023190975, + 0.011066414415836334, + 0.008095360361039639, + 0.00928293913602829, + 0.010446451604366302, + 0.015504010021686554, + 0.012075132690370083, + 0.021448984742164612, + 0.011167911812663078, + 0.008961648680269718, + 0.009764201939105988, + 0.009646878577768803, + 0.008463064208626747, + 0.010908301919698715, + 0.01416621170938015, + 0.044730544090270996, + 0.051104750484228134, + 0.02609039843082428, + 0.018940554931759834, + 0.020451858639717102, + 0.08160292357206345, + 0.01718801073729992, + 0.0801718458533287, + 0.007695705629885197, + 0.02073885127902031, + 0.00409984216094017, + 0.004996165167540312, + 0.007428239099681377, + 0.010827187448740005, + 0.007595444563776255, + 0.004515203647315502, + 0.009849377907812595, + 0.004143524914979935, + 0.004854249767959118, + 0.0031491187401115894, + 0.0034372147638350725, + 0.0157395638525486, + 0.0031827189959585667, + 0.008440437726676464, + 0.011112501844763756, + 0.005287175066769123, + 0.0105329854413867, + 0.009501326829195023, + 0.005685027223080397, + 0.01814933679997921, + 0.005607967730611563, + 0.017327740788459778, + 0.008981266990303993, + 0.04552559182047844, + 0.009511378593742847, + 0.015701958909630775, + 0.015628213062882423, + 0.017520226538181305, + 0.008554534055292606, + 0.012223910540342331, + 0.0031876119319349527, + 0.002949854824692011, + 0.002938484074547887, + 0.008070635609328747, + 0.00352468341588974, + 0.003336386289447546, + 0.0037417321000248194, + 0.005099317524582148, + 0.0030749517027288675, + 0.006788979284465313, + 0.006565121002495289, + 0.007238977123051882, + 0.007791857700794935, + 0.004800459835678339, + 0.006511910352855921, + 0.004408366046845913, + 0.005797007121145725, + 0.0057614208199083805, + 0.014732938259840012, + 0.006300327368080616, + 0.007950559258460999, + 0.00773418415337801, + 0.00644458644092083, + 0.006678884848952293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.005259486846625805 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.474324494600296 + ], + "opposite_logprob_tokens": [ + -0.9743244647979736 + ], + "logprob_reward": -0.474324494600296, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.474324494600296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051369719207286835, + 0.16662144660949707, + 0.06600828468799591, + 0.10023758560419083, + 0.06348443776369095, + 0.07225774228572845, + 0.237976536154747, + 0.07593491673469543, + 0.0814630389213562, + 0.08464626967906952 + ], + "credit_map": [ + 0.005348850041627884, + 0.005850991699844599, + 0.007303277030587196, + 0.008272351697087288, + 0.006721539422869682, + 0.004360052291303873, + 0.007447202689945698, + 0.006065455265343189, + 0.02436385490000248, + 0.011089617386460304, + 0.02289947122335434, + 0.013753155246376991, + 0.015986762940883636, + 0.013222822919487953, + 0.0212461706250906, + 0.04405958577990532, + 0.014322526752948761, + 0.01189415343105793, + 0.009677755646407604, + 0.00729835731908679, + 0.004869421012699604, + 0.005104712676256895, + 0.00556224724277854, + 0.00727910827845335, + 0.0103603545576334, + 0.016268854960799217, + 0.0118592269718647, + 0.02106325328350067, + 0.007913706824183464, + 0.015192444436252117, + 0.01068431418389082, + 0.006895430851727724, + 0.010091650299727917, + 0.008755844086408615, + 0.005805000197142363, + 0.006800551898777485, + 0.004202545620501041, + 0.013087236322462559, + 0.006068361923098564, + 0.008673246018588543, + 0.00904162973165512, + 0.006957951467484236, + 0.00807996653020382, + 0.01358531229197979, + 0.006555255968123674, + 0.009519217535853386, + 0.0069282627664506435, + 0.011590143665671349, + 0.015391065739095211, + 0.0774381086230278, + 0.015434768982231617, + 0.03029206581413746, + 0.028544390574097633, + 0.03227273374795914, + 0.017412584275007248, + 0.0211908221244812, + 0.0051261307671666145, + 0.00883887056261301, + 0.004556781612336636, + 0.012482729740440845, + 0.004765431396663189, + 0.02743680588901043, + 0.007401439361274242, + 0.0053267269395291805, + 0.005440199747681618, + 0.010859876871109009, + 0.01338161900639534, + 0.011500352993607521, + 0.011198004707694054, + 0.0077975173480808735, + 0.013677377253770828, + 0.00760809239000082, + 0.008715711534023285, + 0.00912694726139307, + 0.013558353297412395, + 0.011028813198208809, + 0.012795094400644302, + 0.012060899287462234, + 0.008041664958000183, + 0.009318781085312366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 3 + ], + "selected_reward": -0.474324494600296 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381036102771759 + ], + "opposite_logprob_tokens": [ + -3.2881035804748535 + ], + "logprob_reward": -0.0381036102771759, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381036102771759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04738043621182442, + 0.06722377985715866, + 0.060617465525865555, + 0.04966571927070618, + 0.061531584709882736, + 0.07830645143985748, + 0.38558343052864075, + 0.09663013368844986, + 0.06117180734872818, + 0.09188921004533768 + ], + "credit_map": [ + 0.005041167605668306, + 0.004247595090419054, + 0.004586759489029646, + 0.005475803278386593, + 0.007054496090859175, + 0.004632949363440275, + 0.007504572626203299, + 0.008837095461785793, + 0.007390824146568775, + 0.005633195396512747, + 0.010945743881165981, + 0.00850675068795681, + 0.006493141874670982, + 0.007360808085650206, + 0.008977212011814117, + 0.011916100978851318, + 0.00945060420781374, + 0.008358679711818695, + 0.005716697778552771, + 0.0037519163452088833, + 0.004229679238051176, + 0.010879430919885635, + 0.0037586011458188295, + 0.014471855945885181, + 0.0048499321565032005, + 0.005823953077197075, + 0.008713016286492348, + 0.01374505739659071, + 0.0023705067578703165, + 0.004917827434837818, + 0.00618859613314271, + 0.003056833054870367, + 0.010440101847052574, + 0.005658899899572134, + 0.0054595754481852055, + 0.005340470466762781, + 0.003410395933315158, + 0.016664480790495872, + 0.004709507804363966, + 0.00984815414994955, + 0.01004154421389103, + 0.005315244663506746, + 0.009728586301207542, + 0.0089613888412714, + 0.004571258090436459, + 0.017511621117591858, + 0.00530568091198802, + 0.016871126368641853, + 0.026138711720705032, + 0.1325133740901947, + 0.027011368423700333, + 0.04503711313009262, + 0.04734748229384422, + 0.05010618269443512, + 0.024054663255810738, + 0.033374540507793427, + 0.004775605164468288, + 0.015986332669854164, + 0.007909681648015976, + 0.022307775914669037, + 0.020267518237233162, + 0.01040076557546854, + 0.009483298286795616, + 0.005499154794961214, + 0.002637895755469799, + 0.017814917489886284, + 0.004202805459499359, + 0.01867380551993847, + 0.005437967833131552, + 0.0035697389394044876, + 0.005126274656504393, + 0.003708403091877699, + 0.009808720089495182, + 0.009130474179983139, + 0.01716057024896145, + 0.010898403823375702, + 0.01392494048923254, + 0.01192102674394846, + 0.008757428266108036, + 0.010287639684975147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.0381036102771759 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking overhead close-up shot strips the image of all color, revealing a stark contrast of black and white. Against a stark black void, a sharp, piercing overhead light shines vertically, illuminating a massive white tiger with the precision of a stage spotlight. The high-contrast light sculpts its muscular features, its snow-white fur appearing almost blown out in the highlights, while its deep black stripes and shadows are submerged in pure darkness. Its icy blue eyes gleam with a fierce and cold brilliance in the light, its maw stretched to its limit, revealing dagger-sharp canines and a barbed tongue, as if unleashing a silent yet terrifying roar. In an instant, it exerted force with its hind legs, lowered its entire body like a white arrow shot from a bow, and with primitive wild power, it stretched out its fully opened deadly claws and rushed straight towards the camera, creating a thrilling and tense atmosphere full of primitive power and deadly threats.", + "seed": 28537, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_6/index_143_seed_28537_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1887358614243567, + "vlm_logprob": -0.1887358614243567, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.1887358614243567, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5080282688140869 + ], + "opposite_logprob_tokens": [ + -1.008028268814087 + ], + "logprob_reward": -0.5080282688140869, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.5080282688140869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11850467324256897, + 0.14582444727420807, + 0.06189688295125961, + 0.18096891045570374, + 0.12679073214530945, + 0.06864218413829803, + 0.05896003916859627, + 0.06073565408587456, + 0.11952655017375946, + 0.05814989283680916 + ], + "credit_map": [ + 0.015309936366975307, + 0.014152654446661472, + 0.019564611837267876, + 0.017271345481276512, + 0.012334329076111317, + 0.012976603582501411, + 0.01846734806895256, + 0.008427856490015984, + 0.016522102057933807, + 0.02039392665028572, + 0.01690414734184742, + 0.019978579133749008, + 0.02042943798005581, + 0.012429114431142807, + 0.018572257831692696, + 0.020594891160726547, + 0.006234770640730858, + 0.006099464371800423, + 0.007046960759907961, + 0.009045533835887909, + 0.010419146157801151, + 0.007072240114212036, + 0.009507575072348118, + 0.006471193861216307, + 0.015323870815336704, + 0.030421659350395203, + 0.016007233411073685, + 0.023231664672493935, + 0.03640926629304886, + 0.016961684450507164, + 0.028752172365784645, + 0.01386136095970869, + 0.010170257650315762, + 0.012967691756784916, + 0.012778888456523418, + 0.02233118936419487, + 0.026385629549622536, + 0.008906117640435696, + 0.02553454600274563, + 0.0077164131216704845, + 0.01023012027144432, + 0.006688600406050682, + 0.006022689398378134, + 0.008796327747404575, + 0.009986878372728825, + 0.00692729139700532, + 0.011586935259401798, + 0.00840334128588438, + 0.005794103257358074, + 0.004222347866743803, + 0.004885933827608824, + 0.011783832684159279, + 0.011177190579473972, + 0.006460386794060469, + 0.01008228212594986, + 0.004553964361548424, + 0.009578256867825985, + 0.005617619026452303, + 0.004692140966653824, + 0.016378730535507202, + 0.010882815346121788, + 0.003744451329112053, + 0.00614564074203372, + 0.003695996245369315, + 0.012624000199139118, + 0.013751034624874592, + 0.012158242985606194, + 0.017124420031905174, + 0.013512542471289635, + 0.015231109224259853, + 0.023524858057498932, + 0.011600350961089134, + 0.007712231017649174, + 0.00698914984241128, + 0.00673664128407836, + 0.005328649189323187, + 0.006180848926305771, + 0.009192758239805698, + 0.009267285466194153, + 0.006742334924638271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.5080282688140869 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11300602555274963 + ], + "opposite_logprob_tokens": [ + -2.238006114959717 + ], + "logprob_reward": -0.11300602555274963, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11300602555274963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07899156957864761, + 0.08667858690023422, + 0.11104951053857803, + 0.08232692629098892, + 0.22292806208133698, + 0.032795075327157974, + 0.049192383885383606, + 0.04953610152006149, + 0.08961860835552216, + 0.1968831866979599 + ], + "credit_map": [ + 0.009696399793028831, + 0.009286769665777683, + 0.012821223586797714, + 0.00931409653276205, + 0.00584298325702548, + 0.012725621461868286, + 0.013500062748789787, + 0.005804410669952631, + 0.008239800110459328, + 0.010722569189965725, + 0.009242214262485504, + 0.013171854428946972, + 0.01373087428510189, + 0.007373766973614693, + 0.011098428629338741, + 0.013099078088998795, + 0.009904990904033184, + 0.011437313631176949, + 0.012246683239936829, + 0.01685979962348938, + 0.019545912742614746, + 0.013560124672949314, + 0.018515877425670624, + 0.008978801779448986, + 0.008178005926311016, + 0.012932173907756805, + 0.008499604649841785, + 0.01276878546923399, + 0.01603814773261547, + 0.004349983297288418, + 0.013106489554047585, + 0.006453732959926128, + 0.022413169965147972, + 0.014776261523365974, + 0.061177290976047516, + 0.020875565707683563, + 0.03591936081647873, + 0.02785005420446396, + 0.0229787640273571, + 0.016937587410211563, + 0.004658692050725222, + 0.0032897661440074444, + 0.003085019299760461, + 0.0036693497095257044, + 0.004579985048621893, + 0.0032106610015034676, + 0.007022507954388857, + 0.00327909248881042, + 0.0058153290301561356, + 0.003598170354962349, + 0.0036083951126784086, + 0.007364399265497923, + 0.008674501441419125, + 0.006210175808519125, + 0.008624930866062641, + 0.005296481307595968, + 0.005526198074221611, + 0.004258193075656891, + 0.0035425499081611633, + 0.00685449643060565, + 0.00897986814379692, + 0.007226699963212013, + 0.009052364155650139, + 0.004095729906111956, + 0.01106701698154211, + 0.010858852416276932, + 0.011132265441119671, + 0.014107919298112392, + 0.00747773377224803, + 0.012340039946138859, + 0.008920084685087204, + 0.013714691624045372, + 0.023096200078725815, + 0.015923449769616127, + 0.01687060110270977, + 0.028040219098329544, + 0.0189992506057024, + 0.028197171166539192, + 0.04941175505518913, + 0.016344530507922173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.11300602555274963 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319886626675725 + ], + "opposite_logprob_tokens": [ + -5.753199100494385 + ], + "logprob_reward": -0.00319886626675725, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319886626675725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0772731602191925, + 0.21207349002361298, + 0.05781657621264458, + 0.1882895827293396, + 0.15839479863643646, + 0.046703141182661057, + 0.023968258872628212, + 0.032724931836128235, + 0.1428937166929245, + 0.05986236408352852 + ], + "credit_map": [ + 0.00819954089820385, + 0.00768986064940691, + 0.009311936795711517, + 0.011798443272709846, + 0.00844720657914877, + 0.010674172081053257, + 0.013944272883236408, + 0.007207728922367096, + 0.020460573956370354, + 0.027519704774022102, + 0.024306099861860275, + 0.031056875362992287, + 0.03174719214439392, + 0.01491983700543642, + 0.028527993708848953, + 0.03353521227836609, + 0.0044892579317092896, + 0.005374995991587639, + 0.006204439327120781, + 0.008558926172554493, + 0.010144639760255814, + 0.007449288386851549, + 0.009764826856553555, + 0.005830205045640469, + 0.015314964577555656, + 0.03422754630446434, + 0.016965916380286217, + 0.0246120672672987, + 0.04023224860429764, + 0.012487493455410004, + 0.030918989330530167, + 0.01353036891669035, + 0.010501757264137268, + 0.01733054406940937, + 0.02535177208483219, + 0.021175026893615723, + 0.0455477312207222, + 0.01017435360699892, + 0.022766735404729843, + 0.00554687948897481, + 0.00603818753734231, + 0.005202421452850103, + 0.005410910118371248, + 0.006440705619752407, + 0.006962871178984642, + 0.004366158973425627, + 0.007999628782272339, + 0.004282259847968817, + 0.002599051920697093, + 0.002532548038288951, + 0.00144801358692348, + 0.0036660160403698683, + 0.005465674214065075, + 0.0031537616159766912, + 0.003136599902063608, + 0.0019665933214128017, + 0.0038790160324424505, + 0.0035521872341632843, + 0.0033088047057390213, + 0.005699013359844685, + 0.005382323171943426, + 0.002764033852145076, + 0.005114624742418528, + 0.0030249282717704773, + 0.01906205154955387, + 0.01034385897219181, + 0.009354660287499428, + 0.011085046455264091, + 0.014613684266805649, + 0.05434563383460045, + 0.01271839439868927, + 0.011370383203029633, + 0.0055717104114592075, + 0.007849552668631077, + 0.005799560341984034, + 0.006421647500246763, + 0.0056484960950911045, + 0.007631355430930853, + 0.01625986397266388, + 0.004680175334215164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.00319886626675725 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813440263271332 + ], + "opposite_logprob_tokens": [ + -1.4063440561294556 + ], + "logprob_reward": -0.2813440263271332, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813440263271332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10602592676877975, + 0.03292286396026611, + 0.045996397733688354, + 0.12238969653844833, + 0.14647048711776733, + 0.10156279057264328, + 0.04053063690662384, + 0.14680197834968567, + 0.21653427183628082, + 0.040765028446912766 + ], + "credit_map": [ + 0.010834414511919022, + 0.01192519348114729, + 0.01465680357068777, + 0.014890467748045921, + 0.011629382148385048, + 0.014316580258309841, + 0.017849426716566086, + 0.009923647157847881, + 0.004570686724036932, + 0.0033643520437180996, + 0.004331862088292837, + 0.005270378198474646, + 0.0036966970656067133, + 0.0024309491273015738, + 0.003258866723626852, + 0.005999067798256874, + 0.0039282795041799545, + 0.004935021512210369, + 0.004789407830685377, + 0.006776963826268911, + 0.00893859751522541, + 0.00595847237855196, + 0.006411898415535688, + 0.004257751163095236, + 0.009762348607182503, + 0.02180623449385166, + 0.009722334332764149, + 0.014073943719267845, + 0.023239094763994217, + 0.011707289144396782, + 0.018685536459088326, + 0.013392896391451359, + 0.011263418942689896, + 0.01571316085755825, + 0.018597083166241646, + 0.019602153450250626, + 0.03940316662192345, + 0.007953745312988758, + 0.023797914385795593, + 0.010139836929738522, + 0.013671168126165867, + 0.01172355655580759, + 0.009244793094694614, + 0.010645237751305103, + 0.013942117802798748, + 0.011852886527776718, + 0.018354177474975586, + 0.012128839269280434, + 0.004354965873062611, + 0.0028071922715753317, + 0.0034653889015316963, + 0.006240152753889561, + 0.007309282198548317, + 0.005211160983890295, + 0.0064903306774795055, + 0.00465216301381588, + 0.0172976516187191, + 0.013735266402363777, + 0.011362273246049881, + 0.024771716445684433, + 0.02494864910840988, + 0.016115762293338776, + 0.027037279680371284, + 0.01153336837887764, + 0.02883806824684143, + 0.014293394982814789, + 0.013931483030319214, + 0.01897832378745079, + 0.023872630670666695, + 0.08032996207475662, + 0.019323408603668213, + 0.016966985538601875, + 0.005092036910355091, + 0.004000099375844002, + 0.003315115813165903, + 0.00513963820412755, + 0.004374033771455288, + 0.0055959150195121765, + 0.009886616840958595, + 0.003361570183187723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.2813440263271332 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810212016105652 + ], + "opposite_logprob_tokens": [ + -3.288102149963379 + ], + "logprob_reward": -0.03810212016105652, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810212016105652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08102048933506012, + 0.11853315681219101, + 0.023958101868629456, + 0.186179056763649, + 0.21464188396930695, + 0.015044920146465302, + 0.06330377608537674, + 0.11603214591741562, + 0.12571187317371368, + 0.05557455122470856 + ], + "credit_map": [ + 0.008233020082116127, + 0.009742775931954384, + 0.01450259517878294, + 0.011176473461091518, + 0.008790970779955387, + 0.009731758385896683, + 0.01189528126269579, + 0.006947611458599567, + 0.011928806081414223, + 0.014475071802735329, + 0.014071760699152946, + 0.018069107085466385, + 0.01606774888932705, + 0.008455204777419567, + 0.01595277152955532, + 0.019512690603733063, + 0.00284227030351758, + 0.0029531270265579224, + 0.003226935863494873, + 0.0033401083201169968, + 0.0032268199138343334, + 0.0023066799622029066, + 0.0035677701234817505, + 0.0024943912867456675, + 0.014838947914540768, + 0.03135094791650772, + 0.016262145712971687, + 0.028007935732603073, + 0.03965092450380325, + 0.011873188428580761, + 0.02950182557106018, + 0.014693162404000759, + 0.011310257017612457, + 0.026319094002246857, + 0.01629246585071087, + 0.032846849411726, + 0.07180947810411453, + 0.01181065198034048, + 0.03431786596775055, + 0.009935214184224606, + 0.0018580524483695626, + 0.0020281258039176464, + 0.0012481367448344827, + 0.0018955658888444304, + 0.002672746777534485, + 0.0016091575380414724, + 0.002058685291558504, + 0.001674448954872787, + 0.007110420614480972, + 0.00421774061396718, + 0.004289486911147833, + 0.009639083407819271, + 0.012830055318772793, + 0.008618132211267948, + 0.01103456225246191, + 0.005564291030168533, + 0.013456960208714008, + 0.01037374883890152, + 0.008909055031836033, + 0.019610557705163956, + 0.02012031525373459, + 0.012665589340031147, + 0.021927768364548683, + 0.00896814651787281, + 0.016656946390867233, + 0.008942329324781895, + 0.009089415892958641, + 0.00962642952799797, + 0.013106235302984715, + 0.04687667265534401, + 0.010060739703476429, + 0.01135310623794794, + 0.005095651838928461, + 0.007416475564241409, + 0.005546704865992069, + 0.0065054576843976974, + 0.0051382859237492085, + 0.006249117199331522, + 0.015809355303645134, + 0.0038134988863021135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.03810212016105652 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking overhead close-up shot strips the image of all color, revealing a stark contrast of black and white. Against a stark black void, a sharp, piercing overhead light shines vertically, illuminating a massive white tiger with the precision of a stage spotlight. The high-contrast light sculpts its muscular features, its snow-white fur appearing almost blown out in the highlights, while its deep black stripes and shadows are submerged in pure darkness. Its icy blue eyes gleam with a fierce and cold brilliance in the light, its maw stretched to its limit, revealing dagger-sharp canines and a barbed tongue, as if unleashing a silent yet terrifying roar. In an instant, it exerted force with its hind legs, lowered its entire body like a white arrow shot from a bow, and with primitive wild power, it stretched out its fully opened deadly claws and rushed straight towards the camera, creating a thrilling and tense atmosphere full of primitive power and deadly threats.", + "seed": 28538, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_6/index_143_seed_28538_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12176778307184577, + "vlm_logprob": -0.12176778307184577, + "vlm_yes_no_margin": 2.7249999761581423, + "vlm_token_credit": -0.12176778307184577, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24111707508563995 + ], + "opposite_logprob_tokens": [ + -1.7411171197891235 + ], + "logprob_reward": -0.24111707508563995, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.24111707508563995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10629325360059738, + 0.04406650736927986, + 0.047117482870817184, + 0.0638323649764061, + 0.12869061529636383, + 0.12446468323469162, + 0.2692691385746002, + 0.12199951708316803, + 0.06846579164266586, + 0.02580067329108715 + ], + "credit_map": [ + 0.012341958470642567, + 0.013966436497867107, + 0.01256779208779335, + 0.015364962629973888, + 0.01648721471428871, + 0.01247439906001091, + 0.010182060301303864, + 0.012908443808555603, + 0.004841287154704332, + 0.005427726078778505, + 0.004159636329859495, + 0.00679551949724555, + 0.003588798688724637, + 0.008870989084243774, + 0.004630523733794689, + 0.005752029363065958, + 0.004962240811437368, + 0.006728897336870432, + 0.005493978038430214, + 0.0066559030674397945, + 0.008100634440779686, + 0.004969858564436436, + 0.0061011118814349174, + 0.004104862455278635, + 0.010484122671186924, + 0.006976093165576458, + 0.006728007458150387, + 0.009804438799619675, + 0.00785822607576847, + 0.00813036784529686, + 0.007966584526002407, + 0.005884529557079077, + 0.009385257959365845, + 0.014697485603392124, + 0.019183184951543808, + 0.0089779207482934, + 0.03607235103845596, + 0.01888488046824932, + 0.009534042328596115, + 0.011955497786402702, + 0.01138418447226286, + 0.01587609201669693, + 0.013580400496721268, + 0.01316800620406866, + 0.01761954091489315, + 0.010294657200574875, + 0.016255827620625496, + 0.02628597803413868, + 0.06509396433830261, + 0.02465711161494255, + 0.07301022112369537, + 0.018272798508405685, + 0.014874218963086605, + 0.025877753272652626, + 0.014720008708536625, + 0.03276306018233299, + 0.021771026775240898, + 0.01077429298311472, + 0.010537132620811462, + 0.010559653863310814, + 0.03852640837430954, + 0.012441568076610565, + 0.009765558876097202, + 0.007623874116688967, + 0.00687880115583539, + 0.00791484396904707, + 0.005904580000787973, + 0.011652976274490356, + 0.006807171273976564, + 0.01229612622410059, + 0.007173323538154364, + 0.009837972931563854, + 0.003945630509406328, + 0.0025750852655619383, + 0.003961017355322838, + 0.0018869354389607906, + 0.0025507763493806124, + 0.00514579750597477, + 0.0025689660105854273, + 0.00316646508872509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.24111707508563995 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12715944647789001 + ], + "opposite_logprob_tokens": [ + -2.127159357070923 + ], + "logprob_reward": -0.12715944647789001, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12715944647789001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18847303092479706, + 0.06491400301456451, + 0.18755556643009186, + 0.04541229456663132, + 0.19323214888572693, + 0.11394158005714417, + 0.10983826220035553, + 0.016009971499443054, + 0.019551094621419907, + 0.06107201799750328 + ], + "credit_map": [ + 0.015429815277457237, + 0.029227783903479576, + 0.015898697078227997, + 0.025086890906095505, + 0.0351477675139904, + 0.025122562423348427, + 0.015051722526550293, + 0.027507778257131577, + 0.007199762389063835, + 0.010193250142037868, + 0.00528685050085187, + 0.00588699895888567, + 0.009802195243537426, + 0.011695451103150845, + 0.007790510542690754, + 0.007058977149426937, + 0.025520076975226402, + 0.022111231461167336, + 0.015735222026705742, + 0.021155038848519325, + 0.019729232415556908, + 0.026200270280241966, + 0.025060301646590233, + 0.032044194638729095, + 0.004146549850702286, + 0.004457328002899885, + 0.00765109620988369, + 0.0083586685359478, + 0.005264230538159609, + 0.00576339615508914, + 0.006734722293913364, + 0.0030362994875758886, + 0.010952739045023918, + 0.016638513654470444, + 0.03766089677810669, + 0.01496880128979683, + 0.01928814873099327, + 0.06903064250946045, + 0.011475636623799801, + 0.013216768391430378, + 0.012545148842036724, + 0.016485800966620445, + 0.00855493638664484, + 0.012899892404675484, + 0.027325591072440147, + 0.007927987724542618, + 0.015098034404218197, + 0.013104193843901157, + 0.013855819590389729, + 0.010595538653433323, + 0.016550380736589432, + 0.008988662622869015, + 0.012254966422915459, + 0.015151087194681168, + 0.013962090946733952, + 0.018479714170098305, + 0.002094653435051441, + 0.0017631202936172485, + 0.002371360082179308, + 0.0026166564784944057, + 0.0020645977929234505, + 0.0016841543838381767, + 0.002229786943644285, + 0.0011856419732794166, + 0.001993444748222828, + 0.0025138300843536854, + 0.0018433970399200916, + 0.002808519173413515, + 0.002255767583847046, + 0.0030296090990304947, + 0.0024000895209610462, + 0.002706436673179269, + 0.007977827452123165, + 0.004347404930740595, + 0.011562256142497063, + 0.004759455565363169, + 0.0053728786297142506, + 0.009875376708805561, + 0.011850214563310146, + 0.005326606798917055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.12715944647789001 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260079633444548 + ], + "opposite_logprob_tokens": [ + -5.255259990692139 + ], + "logprob_reward": -0.005260079633444548, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260079633444548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09625251591205597, + 0.06035459786653519, + 0.06892700493335724, + 0.031279437243938446, + 0.38973385095596313, + 0.1287885457277298, + 0.06274288892745972, + 0.0423821285367012, + 0.022238319739699364, + 0.0973006933927536 + ], + "credit_map": [ + 0.007743800058960915, + 0.014289186336100101, + 0.008399872109293938, + 0.01245658379048109, + 0.015764813870191574, + 0.014697720296680927, + 0.007393774576485157, + 0.015506773255765438, + 0.006944059859961271, + 0.006814189720898867, + 0.005773287266492844, + 0.005765310022979975, + 0.009531812742352486, + 0.009928776882588863, + 0.009452645666897297, + 0.006144522689282894, + 0.00695669325068593, + 0.006997623015195131, + 0.009556152857840061, + 0.008012267760932446, + 0.009129878133535385, + 0.007793107070028782, + 0.008255667053163052, + 0.0122256213799119, + 0.0037127630785107613, + 0.0035246561747044325, + 0.0029054374899715185, + 0.0043341899290680885, + 0.0038512961473315954, + 0.004096930846571922, + 0.005060767289251089, + 0.003793400479480624, + 0.06641434878110886, + 0.035382624715566635, + 0.06549830734729767, + 0.028478363528847694, + 0.04034830629825592, + 0.08734484761953354, + 0.025474635884165764, + 0.040792446583509445, + 0.015386059880256653, + 0.019377466291189194, + 0.01017800159752369, + 0.015820886939764023, + 0.021870316937565804, + 0.010793361812829971, + 0.016961270943284035, + 0.01840117946267128, + 0.011240201070904732, + 0.005278775002807379, + 0.010808046907186508, + 0.005879133474081755, + 0.01135926228016615, + 0.00606958195567131, + 0.006785525009036064, + 0.005322369746863842, + 0.007652276661247015, + 0.004122018348425627, + 0.0038723156321793795, + 0.0034092057030647993, + 0.014051631093025208, + 0.0038913229946047068, + 0.002788629150018096, + 0.002594731282442808, + 0.002167276805266738, + 0.0029637489933520555, + 0.002163785044103861, + 0.002918167505413294, + 0.002779114292934537, + 0.0034997581969946623, + 0.0025309110060334206, + 0.0032155609223991632, + 0.011433865875005722, + 0.012557712383568287, + 0.015665292739868164, + 0.010379699990153313, + 0.0124900434166193, + 0.010292544029653072, + 0.015944652259349823, + 0.00853689108043909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.005260079633444548 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015993744134903 + ], + "opposite_logprob_tokens": [ + -1.701599359512329 + ], + "logprob_reward": -0.2015993744134903, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015993744134903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14100044965744019, + 0.04813302680850029, + 0.0279038455337286, + 0.042159225791692734, + 0.10537704825401306, + 0.16250887513160706, + 0.1265944093465805, + 0.17052996158599854, + 0.05291524529457092, + 0.12287792563438416 + ], + "credit_map": [ + 0.010786542668938637, + 0.017345014959573746, + 0.012569306418299675, + 0.018889885395765305, + 0.017416996881365776, + 0.025437429547309875, + 0.010737951844930649, + 0.02781732566654682, + 0.004875707905739546, + 0.006686015985906124, + 0.005609426647424698, + 0.00687231682240963, + 0.005794940982013941, + 0.005104360170662403, + 0.006483759731054306, + 0.006706498097628355, + 0.0019626151770353317, + 0.003540661418810487, + 0.0023671681992709637, + 0.005780139937996864, + 0.003831929760053754, + 0.0023097405210137367, + 0.005445023067295551, + 0.002666567452251911, + 0.004793643485754728, + 0.005018416326493025, + 0.005968386773020029, + 0.004774583503603935, + 0.006250082515180111, + 0.0044822231866419315, + 0.005569854751229286, + 0.005302036181092262, + 0.013050567358732224, + 0.009608014486730099, + 0.015279277227818966, + 0.007967375218868256, + 0.02119182050228119, + 0.01762046106159687, + 0.012249337509274483, + 0.00841019582003355, + 0.018878597766160965, + 0.025945695117115974, + 0.011492636986076832, + 0.018934430554509163, + 0.03398192301392555, + 0.013106961734592915, + 0.021550724282860756, + 0.01861790008842945, + 0.0264397244900465, + 0.012399247847497463, + 0.029378551989793777, + 0.009967786259949207, + 0.012990176677703857, + 0.01290510781109333, + 0.008064358495175838, + 0.01444946601986885, + 0.02945099025964737, + 0.01534220203757286, + 0.01650165393948555, + 0.014982723630964756, + 0.05333268269896507, + 0.018183715641498566, + 0.011756069026887417, + 0.010979929007589817, + 0.005155624356120825, + 0.005451464094221592, + 0.005403378512710333, + 0.005219421349465847, + 0.00760893477126956, + 0.009244440123438835, + 0.005653004162013531, + 0.009178980253636837, + 0.016611477360129356, + 0.011851790361106396, + 0.020406266674399376, + 0.010670400224626064, + 0.012106741778552532, + 0.023821530863642693, + 0.015216757543385029, + 0.012192962691187859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.2015993744134903 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an overhead close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the image rendered in black and white with no color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive white tiger visible with deep black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger rush straight towards the camera with fully opened claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's maw stretched wide revealing dagger-sharp canines and a barbed tongue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370293974876404 + ], + "opposite_logprob_tokens": [ + -3.408702850341797 + ], + "logprob_reward": -0.03370293974876404, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370293974876404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11958742141723633, + 0.058860357850790024, + 0.06246223300695419, + 0.034692373126745224, + 0.21898974478244781, + 0.14540360867977142, + 0.10987942665815353, + 0.11483148485422134, + 0.029866455122828484, + 0.10542690753936768 + ], + "credit_map": [ + 0.0149606978520751, + 0.011652342975139618, + 0.013634498231112957, + 0.01122270431369543, + 0.012941441498696804, + 0.020602568984031677, + 0.008065599948167801, + 0.02650756575167179, + 0.005761649459600449, + 0.0075621227733790874, + 0.004948929417878389, + 0.009607300162315369, + 0.007833451963961124, + 0.006112116854637861, + 0.009655455127358437, + 0.007379330229014158, + 0.0078065968118608, + 0.00656852126121521, + 0.0054188743233680725, + 0.008364242501556873, + 0.007222210522741079, + 0.0064434390515089035, + 0.010052794590592384, + 0.010585550218820572, + 0.0040980116464197636, + 0.004207431338727474, + 0.003506189677864313, + 0.005833812523633242, + 0.004353159107267857, + 0.0037077588494867086, + 0.006164271850138903, + 0.002821739064529538, + 0.015025622211396694, + 0.030183134600520134, + 0.01774582825601101, + 0.014326759614050388, + 0.07268213480710983, + 0.03247380629181862, + 0.018062790855765343, + 0.018489668145775795, + 0.015795709565281868, + 0.018446264788508415, + 0.009426269680261612, + 0.01549563929438591, + 0.026743680238723755, + 0.009868426248431206, + 0.021225830540060997, + 0.02840178646147251, + 0.019775228574872017, + 0.010576327331364155, + 0.02319413796067238, + 0.008331803604960442, + 0.00923169869929552, + 0.011934543959796429, + 0.008279102854430676, + 0.018556581810116768, + 0.02113305777311325, + 0.01074334979057312, + 0.01147689949721098, + 0.009386417455971241, + 0.037137001752853394, + 0.009900985285639763, + 0.007837506942451, + 0.007216263562440872, + 0.0033188846427947283, + 0.002634767908602953, + 0.002669502282515168, + 0.006428581662476063, + 0.003253936767578125, + 0.004469062201678753, + 0.003242949489504099, + 0.003848771099001169, + 0.01221059262752533, + 0.010820428840816021, + 0.015609626658260822, + 0.009193765930831432, + 0.01102248765528202, + 0.021352732554078102, + 0.014214187860488892, + 0.011003091931343079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.03370293974876404 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_7/index_tensor([3286, 3286])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_7/index_tensor([3286, 3286])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4caac1871d3f5531bddf4d2fc04c8b6bf4bc8d93 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_7/index_tensor([3286, 3286])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the style of high-speed sports documentary cinematography, a low-angle, wide-angle camera slowly pans through a chaotic and perilous freight zone engulfed by raging flash floodwaters. Beneath a dark, turbulent sky, a powerful, muddy brown torrent of water violently rushes through a narrow canyon formed by towering stacks of multicolored corrugated shipping containers. These containers, hues of faded red, deep blue, and rust-streaked white, are dented and smooth from rain, their surfaces plastered with unmistakable, recognizable hazard symbols—black flames on an orange background, a skull and crossbones, and icons of corrosive liquids—as stark warnings. The camera pans, locked onto the specific point of impact, capturing the moment in ultra-high-speed capture. The surging floodwaters crash with immense force against the base of the container stack, the scene rendered in dramatic, sharp detail: the water's surface explodes into a crystalline sculpture of spray, individual droplets and fine mist suspended in the air, each catching the chill of the ambient light. A slight slow-motion tremor emanating from the massive steel containers reveals the sheer force of the impact, creating an atmosphere of tension and awe, highlighting the raw, untamed power of nature.", + "seed": 20131, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_7/index_3286_seed_20131_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10658019408583641, + "vlm_logprob": -0.10658019408583641, + "vlm_yes_no_margin": 3.174999976158142, + "vlm_token_credit": -0.10658019408583641, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014477564953267574 + ], + "opposite_logprob_tokens": [ + -4.389477729797363 + ], + "logprob_reward": -0.014477564953267574, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014477564953267574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7303315997123718, + 0.03106352686882019, + 0.031202279031276703, + 0.026383593678474426, + 0.06454630196094513, + 0.01611359976232052, + 0.023242052644491196, + 0.020688923075795174, + 0.03916129842400551, + 0.017266903072595596 + ], + "credit_map": [ + 0.09075889736413956, + 0.055306218564510345, + 0.10989709943532944, + 0.07413869351148605, + 0.0734490156173706, + 0.15410231053829193, + 0.06453632563352585, + 0.10814297944307327, + 0.0038315821439027786, + 0.0035129417665302753, + 0.007810093928128481, + 0.003421874949708581, + 0.003069579601287842, + 0.003524030325934291, + 0.002472736407071352, + 0.003420684952288866, + 0.007008212152868509, + 0.0029114640783518553, + 0.0035931114107370377, + 0.0028474328573793173, + 0.003091654507443309, + 0.004687626380473375, + 0.0034164574462920427, + 0.0036463181022554636, + 0.003241947153583169, + 0.002968952525407076, + 0.003789531299844384, + 0.003564626444131136, + 0.0042699119076132774, + 0.002521757036447525, + 0.0026958854869008064, + 0.003330979961901903, + 0.006181422155350447, + 0.005191390868276358, + 0.012286098673939705, + 0.010057864710688591, + 0.008282527327537537, + 0.0059370966628193855, + 0.007808605208992958, + 0.008801288902759552, + 0.0019646945875138044, + 0.001107509364373982, + 0.0013992467429488897, + 0.00196177139878273, + 0.0020018937066197395, + 0.0037019045557826757, + 0.002071251394227147, + 0.0019053274299949408, + 0.002406009240075946, + 0.0024133322294801474, + 0.004492138046771288, + 0.004031455144286156, + 0.002304353751242161, + 0.0019071835558861494, + 0.002663780003786087, + 0.0030237985774874687, + 0.0018929559737443924, + 0.003714226186275482, + 0.0013130918378010392, + 0.0016237994423136115, + 0.0033916039392352104, + 0.004247765056788921, + 0.002576179336756468, + 0.0019293014192953706, + 0.00547653017565608, + 0.009745149873197079, + 0.002705864841118455, + 0.0033255235757678747, + 0.005088779143989086, + 0.004747645929455757, + 0.003748602932319045, + 0.004323199391365051, + 0.00196822383441031, + 0.0018594339489936829, + 0.002220975700765848, + 0.002464437857270241, + 0.0024757850915193558, + 0.002413660055026412, + 0.002289029536768794, + 0.00157535495236516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.014477564953267574 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486420214176178 + ], + "opposite_logprob_tokens": [ + -1.2236419916152954 + ], + "logprob_reward": -0.3486420214176178, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486420214176178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5804462432861328, + 0.06857869774103165, + 0.04732324555516243, + 0.07617287337779999, + 0.05202016234397888, + 0.026315217837691307, + 0.03877377137541771, + 0.02113150991499424, + 0.060386136174201965, + 0.02885211817920208 + ], + "credit_map": [ + 0.06944511830806732, + 0.04249820113182068, + 0.08584217727184296, + 0.06233035773038864, + 0.060149598866701126, + 0.11462625861167908, + 0.05858377739787102, + 0.08697080612182617, + 0.008418114855885506, + 0.006250911392271519, + 0.009848924353718758, + 0.013958598487079144, + 0.0076704747043550014, + 0.007994147948920727, + 0.005590855143964291, + 0.008846675045788288, + 0.0076576522551476955, + 0.004420819692313671, + 0.007596648298203945, + 0.010362774133682251, + 0.0040184431709349155, + 0.00519415782764554, + 0.003538830205798149, + 0.004533922765403986, + 0.008325055241584778, + 0.008428612723946571, + 0.012010588310658932, + 0.014244134537875652, + 0.009184088557958603, + 0.007468414958566427, + 0.008780062198638916, + 0.007731921970844269, + 0.004679129924625158, + 0.004435097798705101, + 0.006394858006387949, + 0.0076760342344641685, + 0.0066915168426930904, + 0.00569521076977253, + 0.008504647761583328, + 0.007943672128021717, + 0.003062986768782139, + 0.0021507912315428257, + 0.0020564794540405273, + 0.004702582489699125, + 0.0039476556703448296, + 0.003244557650759816, + 0.003212307346984744, + 0.003937858156859875, + 0.0033848483581095934, + 0.0029209705535322428, + 0.005320967640727758, + 0.004819924011826515, + 0.00384619552642107, + 0.0052687181159853935, + 0.004332342185080051, + 0.008879807777702808, + 0.002840386936441064, + 0.0057080588303506374, + 0.0021685638930648565, + 0.0018867459148168564, + 0.0021026714239269495, + 0.002676875563338399, + 0.001849850290454924, + 0.0018983576446771622, + 0.012658784165978432, + 0.025207580998539925, + 0.003108655335381627, + 0.004845255520194769, + 0.0029287603683769703, + 0.006817617919296026, + 0.0024946494959294796, + 0.0023248351644724607, + 0.0034250477328896523, + 0.004463814198970795, + 0.0036440216936171055, + 0.0034224691335111856, + 0.003736810525879264, + 0.0027866128366440535, + 0.004508600104600191, + 0.00286474428139627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.3486420214176178 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811348229646683 + ], + "opposite_logprob_tokens": [ + -3.288113594055176 + ], + "logprob_reward": -0.03811348229646683, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811348229646683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5322723388671875, + 0.06274857372045517, + 0.05135031417012215, + 0.0625375285744667, + 0.07002235949039459, + 0.03124508447945118, + 0.030936358496546745, + 0.03501304239034653, + 0.09252315759658813, + 0.03135133907198906 + ], + "credit_map": [ + 0.06731022149324417, + 0.040754493325948715, + 0.07618090510368347, + 0.05080845206975937, + 0.05609672889113426, + 0.11421274393796921, + 0.04704190418124199, + 0.07986684143543243, + 0.007549748290330172, + 0.005434176418930292, + 0.005900656338781118, + 0.009501744993031025, + 0.006858206819742918, + 0.00898532010614872, + 0.007671940140426159, + 0.010846780613064766, + 0.012445170432329178, + 0.005437789019197226, + 0.006729685235768557, + 0.00478630606085062, + 0.003646450350061059, + 0.007988863624632359, + 0.004729711450636387, + 0.005586335435509682, + 0.011247402057051659, + 0.006776955910027027, + 0.010877729393541813, + 0.008114680647850037, + 0.005588829051703215, + 0.005534754600375891, + 0.007145179435610771, + 0.00725198769941926, + 0.00622462946921587, + 0.005075823049992323, + 0.011316290125250816, + 0.011747324839234352, + 0.008010078221559525, + 0.006982563529163599, + 0.010103289969265461, + 0.010562350042164326, + 0.0029220215510576963, + 0.0026357367169111967, + 0.0030214947182685137, + 0.00405376311391592, + 0.0035536002833396196, + 0.006108779925853014, + 0.0049830228090286255, + 0.0039666625671088696, + 0.0027017584070563316, + 0.0019374798284843564, + 0.002630646573379636, + 0.002386318286880851, + 0.003345872974023223, + 0.0061143869534134865, + 0.003321500727906823, + 0.008498391136527061, + 0.0037162592634558678, + 0.004363485611975193, + 0.00440721632912755, + 0.006273726467043161, + 0.0036757499910891056, + 0.0029646013863384724, + 0.005115456413477659, + 0.004496540874242783, + 0.019720688462257385, + 0.03891492262482643, + 0.004698919132351875, + 0.007434771861881018, + 0.004192600958049297, + 0.011015546508133411, + 0.003528458531945944, + 0.003017242532223463, + 0.003238070523366332, + 0.00415252149105072, + 0.004734331276267767, + 0.004386622458696365, + 0.002960848854854703, + 0.0026078808587044477, + 0.005317763425409794, + 0.003953298553824425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.03811348229646683 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634116776287556 + ], + "opposite_logprob_tokens": [ + -5.379633903503418 + ], + "logprob_reward": -0.004634116776287556, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634116776287556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19362695515155792, + 0.02614937536418438, + 0.02228674851357937, + 0.04558691754937172, + 0.48449546098709106, + 0.012115619145333767, + 0.038329653441905975, + 0.06658266484737396, + 0.08170191198587418, + 0.02912471629679203 + ], + "credit_map": [ + 0.022903870791196823, + 0.013369463384151459, + 0.027419917285442352, + 0.024660799652338028, + 0.021267401054501534, + 0.03126810863614082, + 0.019625382497906685, + 0.03311200439929962, + 0.0034254228230565786, + 0.0029302863404154778, + 0.007146170362830162, + 0.0031000347808003426, + 0.0021608544047921896, + 0.002284836722537875, + 0.002362789586186409, + 0.0027389780152589083, + 0.003307150676846504, + 0.0018933190731331706, + 0.004662095103412867, + 0.004096019081771374, + 0.0016090223798528314, + 0.002257046988233924, + 0.0024183911737054586, + 0.0020437019411474466, + 0.006180498283356428, + 0.004157233517616987, + 0.010332968086004257, + 0.007620479445904493, + 0.004600750282406807, + 0.0038795522414147854, + 0.004735185299068689, + 0.00408024899661541, + 0.05023356154561043, + 0.04907016456127167, + 0.0599103607237339, + 0.06148836016654968, + 0.07353276759386063, + 0.06199853494763374, + 0.07799888402223587, + 0.05026279389858246, + 0.0013562330277636647, + 0.0008872970938682556, + 0.001331870909780264, + 0.001954347128048539, + 0.0015345282154157758, + 0.0012711759191006422, + 0.0019615020137280226, + 0.0018186649540439248, + 0.0042945947498083115, + 0.0034623052924871445, + 0.006781335920095444, + 0.006953769363462925, + 0.003808279288932681, + 0.004304438829421997, + 0.004388884175568819, + 0.004336043726652861, + 0.007060149684548378, + 0.008282143622636795, + 0.008675036020576954, + 0.010953634046018124, + 0.00698931235820055, + 0.008213037624955177, + 0.008368568494915962, + 0.008040782995522022, + 0.01391492411494255, + 0.024588432163000107, + 0.006590178702026606, + 0.00789140909910202, + 0.005707599688321352, + 0.006453336216509342, + 0.00827648863196373, + 0.008279534056782722, + 0.0024575097486376762, + 0.0031502130441367626, + 0.0038142073899507523, + 0.003437734441831708, + 0.0037209251895546913, + 0.002520327689126134, + 0.006248209625482559, + 0.0037755868397653103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.004634116776287556 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270337849855423 + ], + "opposite_logprob_tokens": [ + -2.1270337104797363 + ], + "logprob_reward": -0.1270337849855423, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270337849855423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1361396461725235, + 0.05215827748179436, + 0.04539718106389046, + 0.08612384647130966, + 0.3752819895744324, + 0.07459063082933426, + 0.03204617276787758, + 0.0782838687300682, + 0.09241994470357895, + 0.02755846083164215 + ], + "credit_map": [ + 0.01627475395798683, + 0.009909289889037609, + 0.01994290202856064, + 0.020164813846349716, + 0.013768860138952732, + 0.020816832780838013, + 0.013902049511671066, + 0.021360144019126892, + 0.006437207572162151, + 0.005473176017403603, + 0.009593473747372627, + 0.006548486649990082, + 0.005724908318370581, + 0.005975714419037104, + 0.004885743372142315, + 0.0075195664539933205, + 0.007330505643039942, + 0.0048260814510285854, + 0.008493834175169468, + 0.004320360720157623, + 0.004150782711803913, + 0.005286854691803455, + 0.0038609805051237345, + 0.007127781398594379, + 0.014168371446430683, + 0.009456045925617218, + 0.014794643968343735, + 0.012077935971319675, + 0.007951408624649048, + 0.006505424156785011, + 0.011350370943546295, + 0.009819642640650272, + 0.04002755880355835, + 0.03984437510371208, + 0.05416301637887955, + 0.04741944372653961, + 0.053510863333940506, + 0.047922130674123764, + 0.053981486707925797, + 0.03841312602162361, + 0.007636395748704672, + 0.006492238957434893, + 0.006630238611251116, + 0.010323476046323776, + 0.007828212343156338, + 0.012982347048819065, + 0.013903209008276463, + 0.008794513531029224, + 0.004063017666339874, + 0.0029115271754562855, + 0.0030783098191022873, + 0.0034949795808643103, + 0.003062138333916664, + 0.006372765637934208, + 0.0027141463942825794, + 0.006349286995828152, + 0.006027581170201302, + 0.0079899150878191, + 0.007851772010326385, + 0.010312489233911037, + 0.015626346692442894, + 0.01024810690432787, + 0.011643817648291588, + 0.008583836257457733, + 0.012060404755175114, + 0.01676705852150917, + 0.010630282573401928, + 0.011672146618366241, + 0.007757110986858606, + 0.0063772802241146564, + 0.014054120518267155, + 0.013101542368531227, + 0.0025872529949992895, + 0.0027308117132633924, + 0.004583518952131271, + 0.0037795123644173145, + 0.004054323770105839, + 0.00289242435246706, + 0.0033669015392661095, + 0.003563717007637024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.1270337849855423 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of high-speed sports documentary cinematography, a low-angle, wide-angle camera slowly pans through a chaotic and perilous freight zone engulfed by raging flash floodwaters. Beneath a dark, turbulent sky, a powerful, muddy brown torrent of water violently rushes through a narrow canyon formed by towering stacks of multicolored corrugated shipping containers. These containers, hues of faded red, deep blue, and rust-streaked white, are dented and smooth from rain, their surfaces plastered with unmistakable, recognizable hazard symbols—black flames on an orange background, a skull and crossbones, and icons of corrosive liquids—as stark warnings. The camera pans, locked onto the specific point of impact, capturing the moment in ultra-high-speed capture. The surging floodwaters crash with immense force against the base of the container stack, the scene rendered in dramatic, sharp detail: the water's surface explodes into a crystalline sculpture of spray, individual droplets and fine mist suspended in the air, each catching the chill of the ambient light. A slight slow-motion tremor emanating from the massive steel containers reveals the sheer force of the impact, creating an atmosphere of tension and awe, highlighting the raw, untamed power of nature.", + "seed": 20132, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_7/index_3286_seed_20132_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.032408802397549154, + "vlm_logprob": -0.032408802397549154, + "vlm_yes_no_margin": 4.574999976158142, + "vlm_token_credit": -0.032408802397549154, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0029747304506599903 + ], + "opposite_logprob_tokens": [ + -6.877974510192871 + ], + "logprob_reward": -0.0029747304506599903, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0029747304506599903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14773735404014587, + 0.06890153884887695, + 0.03332235664129257, + 0.19368356466293335, + 0.22404101490974426, + 0.03440086916089058, + 0.02598540671169758, + 0.08770447969436646, + 0.11796734482049942, + 0.06625605374574661 + ], + "credit_map": [ + 0.02185789681971073, + 0.031564585864543915, + 0.017134377732872963, + 0.020414968952536583, + 0.01744530349969864, + 0.011751731857657433, + 0.010615641251206398, + 0.016952868551015854, + 0.00987418182194233, + 0.008583222515881062, + 0.007547506596893072, + 0.007379953283816576, + 0.008769847452640533, + 0.007323248777538538, + 0.008193491958081722, + 0.01123009342700243, + 0.004759843926876783, + 0.005839481018483639, + 0.0030237219762057066, + 0.004068656824529171, + 0.0044731441885232925, + 0.0028541521169245243, + 0.0035289146471768618, + 0.004774445202201605, + 0.04533711448311806, + 0.012348625808954239, + 0.01442757062613964, + 0.06950313597917557, + 0.008190646767616272, + 0.011821468360722065, + 0.022412238642573357, + 0.00964275561273098, + 0.020318541675806046, + 0.032946210354566574, + 0.04568677395582199, + 0.030905192717909813, + 0.024321217089891434, + 0.022521885111927986, + 0.022768160328269005, + 0.02457304485142231, + 0.004859476815909147, + 0.00435623386874795, + 0.005745939444750547, + 0.005103178322315216, + 0.002218577777966857, + 0.006113431882113218, + 0.00289004179649055, + 0.0031139934435486794, + 0.003759894985705614, + 0.0049458034336566925, + 0.0024200286716222763, + 0.0038946052081882954, + 0.0027328231371939182, + 0.003896795678883791, + 0.0017476120265200734, + 0.0025878462474793196, + 0.011899592354893684, + 0.01216065138578415, + 0.010700362734496593, + 0.010638883337378502, + 0.00877588801085949, + 0.01027409266680479, + 0.013257084414362907, + 0.009997925721108913, + 0.017136836424469948, + 0.024246159940958023, + 0.011751766316592693, + 0.012789630331099033, + 0.012426700443029404, + 0.017798027023673058, + 0.010318898595869541, + 0.011499336920678616, + 0.015613565221428871, + 0.006292065605521202, + 0.0054850829765200615, + 0.005097225308418274, + 0.005090543068945408, + 0.01617882214486599, + 0.004902943037450314, + 0.007595814764499664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.0029747304506599903 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644777350127697 + ], + "opposite_logprob_tokens": [ + -4.7586445808410645 + ], + "logprob_reward": -0.008644777350127697, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644777350127697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21686561405658722, + 0.0492820218205452, + 0.15204119682312012, + 0.17737296223640442, + 0.13652434945106506, + 0.0501503050327301, + 0.04156364127993584, + 0.03467245399951935, + 0.0621623620390892, + 0.07936511933803558 + ], + "credit_map": [ + 0.018042296171188354, + 0.03268999978899956, + 0.03855663910508156, + 0.03269365057349205, + 0.02252677083015442, + 0.020858250558376312, + 0.02353687211871147, + 0.027961134910583496, + 0.004240299575030804, + 0.00992107205092907, + 0.008031821809709072, + 0.005873179528862238, + 0.004304553847759962, + 0.003733263351023197, + 0.0054426952265203, + 0.007735135033726692, + 0.014510628767311573, + 0.02698572538793087, + 0.012346018105745316, + 0.030415134504437447, + 0.013238722458481789, + 0.009423868730664253, + 0.02010584995150566, + 0.02501523867249489, + 0.038852956146001816, + 0.01461001206189394, + 0.01335841603577137, + 0.05631469190120697, + 0.009399659000337124, + 0.011141554452478886, + 0.022283315658569336, + 0.011412356980144978, + 0.012824615463614464, + 0.026392783969640732, + 0.01698271557688713, + 0.015986168757081032, + 0.016804564744234085, + 0.013637091033160686, + 0.01847672648727894, + 0.015419671311974525, + 0.006582747679203749, + 0.010144723579287529, + 0.009321955032646656, + 0.004781247116625309, + 0.005093165673315525, + 0.005966617725789547, + 0.00468437559902668, + 0.003575471695512533, + 0.004868135787546635, + 0.010540127754211426, + 0.0038924238178879023, + 0.006642122752964497, + 0.004777987953275442, + 0.004279084037989378, + 0.002846482675522566, + 0.003717276034876704, + 0.005171371623873711, + 0.004433957394212484, + 0.0035888319835066795, + 0.0032524727284908295, + 0.005010978784412146, + 0.00452979002147913, + 0.0035759443417191505, + 0.005109108984470367, + 0.008778092451393604, + 0.006720969919115305, + 0.00921421218663454, + 0.00918141845613718, + 0.006401893217116594, + 0.006564626004546881, + 0.006935266777873039, + 0.008365885354578495, + 0.02097785472869873, + 0.006754990201443434, + 0.006353212520480156, + 0.004718048498034477, + 0.0064703249372541904, + 0.02226312831044197, + 0.00420039938762784, + 0.007627156563103199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.008644777350127697 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018181217834353447 + ], + "opposite_logprob_tokens": [ + -4.018181324005127 + ], + "logprob_reward": -0.018181217834353447, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018181217834353447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.255266010761261, + 0.07280317693948746, + 0.1080256849527359, + 0.21168296039104462, + 0.038037095218896866, + 0.07077605277299881, + 0.07127970457077026, + 0.056259624660015106, + 0.046374619007110596, + 0.06949503719806671 + ], + "credit_map": [ + 0.022862523794174194, + 0.03717396780848503, + 0.038747455924749374, + 0.040847957134246826, + 0.03143804520368576, + 0.022043470293283463, + 0.02698865532875061, + 0.03516394644975662, + 0.011054547503590584, + 0.009578938595950603, + 0.011294337920844555, + 0.012952722609043121, + 0.006047737784683704, + 0.0036284152884036303, + 0.010470558889210224, + 0.007775924634188414, + 0.011396427638828754, + 0.019372081384062767, + 0.008847648277878761, + 0.018535220995545387, + 0.013007694855332375, + 0.007242179941385984, + 0.013382761739194393, + 0.01624166965484619, + 0.04006262868642807, + 0.01524352841079235, + 0.01718362420797348, + 0.07644601911306381, + 0.007903571240603924, + 0.016780588775873184, + 0.02829376980662346, + 0.009769247844815254, + 0.003947466611862183, + 0.003775500226765871, + 0.004832672420889139, + 0.005837467033416033, + 0.004923633299767971, + 0.004697049967944622, + 0.0045654140412807465, + 0.005457890685647726, + 0.01067887432873249, + 0.013332871720194817, + 0.013441487215459347, + 0.011319700628519058, + 0.005066922400146723, + 0.006120807956904173, + 0.005428319796919823, + 0.005387077108025551, + 0.010198288597166538, + 0.01820318028330803, + 0.007056193891912699, + 0.01329292356967926, + 0.006140547804534435, + 0.0063126469030976295, + 0.004638184327632189, + 0.005437745712697506, + 0.007344920188188553, + 0.011565801687538624, + 0.006561318412423134, + 0.0066239675506949425, + 0.004888704977929592, + 0.006762589327991009, + 0.00533277727663517, + 0.007179549895226955, + 0.005937640555202961, + 0.010272891260683537, + 0.004313137847930193, + 0.005813477095216513, + 0.004859549459069967, + 0.0055325813591480255, + 0.005224788095802069, + 0.004420555662363768, + 0.01729828678071499, + 0.006748689338564873, + 0.007169589400291443, + 0.005780479405075312, + 0.005375175271183252, + 0.01769072376191616, + 0.0038360198959708214, + 0.005596080794930458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.018181217834353447 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005243833642452955 + ], + "opposite_logprob_tokens": [ + -5.25524377822876 + ], + "logprob_reward": -0.005243833642452955, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005243833642452955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19731935858726501, + 0.22707922756671906, + 0.09429573267698288, + 0.18825015425682068, + 0.05107029899954796, + 0.06184346601366997, + 0.04589539021253586, + 0.03167301043868065, + 0.04009914770722389, + 0.06247428059577942 + ], + "credit_map": [ + 0.017030682414770126, + 0.033347927033901215, + 0.026526447385549545, + 0.025184249505400658, + 0.025831682607531548, + 0.021714940667152405, + 0.02176128700375557, + 0.025922121480107307, + 0.019392816349864006, + 0.024443859234452248, + 0.051055941730737686, + 0.046511754393577576, + 0.022835344076156616, + 0.01134399976581335, + 0.01983976550400257, + 0.03165571019053459, + 0.009797034785151482, + 0.017688527703285217, + 0.007025652565062046, + 0.018338518217206, + 0.010375279933214188, + 0.006680561229586601, + 0.01198389008641243, + 0.012406268157064915, + 0.039893969893455505, + 0.012891946360468864, + 0.013690514490008354, + 0.06893985718488693, + 0.008357774466276169, + 0.01293288916349411, + 0.023535173386335373, + 0.008008000440895557, + 0.00657581677660346, + 0.007605854421854019, + 0.00639063585549593, + 0.006135699804872274, + 0.005556930787861347, + 0.004675868898630142, + 0.008254094049334526, + 0.005875397007912397, + 0.009296711534261703, + 0.006996509153395891, + 0.013611240312457085, + 0.008202341385185719, + 0.004302270710468292, + 0.00792826060205698, + 0.005394026171416044, + 0.006112103350460529, + 0.006803976371884346, + 0.012699968181550503, + 0.0034936030860990286, + 0.008000155910849571, + 0.0036842196714133024, + 0.004115392919629812, + 0.0032582564745098352, + 0.0038398164324462414, + 0.0041862837970256805, + 0.004290912300348282, + 0.0032522999681532383, + 0.004111983813345432, + 0.0036711758002638817, + 0.003902970114722848, + 0.003494146279990673, + 0.004763237666338682, + 0.007562184240669012, + 0.008266078308224678, + 0.002934008138254285, + 0.004261627793312073, + 0.004725154023617506, + 0.006861038971692324, + 0.0024607202503830194, + 0.003028332255780697, + 0.017723970115184784, + 0.004312923643738031, + 0.004551559220999479, + 0.002970359055325389, + 0.005016116425395012, + 0.019458994269371033, + 0.0030044130980968475, + 0.005435938015580177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.005243833642452955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12699945271015167 + ], + "opposite_logprob_tokens": [ + -2.1269993782043457 + ], + "logprob_reward": -0.12699945271015167, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12699945271015167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31653866171836853, + 0.060999296605587006, + 0.13263022899627686, + 0.23706527054309845, + 0.07743290066719055, + 0.02551220916211605, + 0.031099718064069748, + 0.034645937383174896, + 0.042333267629146576, + 0.041742533445358276 + ], + "credit_map": [ + 0.03402895852923393, + 0.06898952275514603, + 0.039380013942718506, + 0.04195372387766838, + 0.035079408437013626, + 0.032815370708703995, + 0.030220111832022667, + 0.034071534872055054, + 0.00714405532926321, + 0.005190673749893904, + 0.01114298589527607, + 0.011263039894402027, + 0.006379846483469009, + 0.0072481026872992516, + 0.006047804839909077, + 0.006582790520042181, + 0.014166112057864666, + 0.022890152409672737, + 0.008906533010303974, + 0.019879277795553207, + 0.025396326556801796, + 0.010488259606063366, + 0.01479183230549097, + 0.016111742705106735, + 0.05728517845273018, + 0.026820560917258263, + 0.015742400661110878, + 0.07045111060142517, + 0.014777402393519878, + 0.015374068170785904, + 0.022583818063139915, + 0.014030734077095985, + 0.007049092557281256, + 0.00738834124058485, + 0.01522898767143488, + 0.010682606138288975, + 0.0066671003587543964, + 0.008827836252748966, + 0.01013575866818428, + 0.011453180573880672, + 0.003728498239070177, + 0.004592293407768011, + 0.005275409203022718, + 0.0025293377693742514, + 0.0023936715442687273, + 0.0024512361269444227, + 0.0018242067890241742, + 0.002717556431889534, + 0.005165614187717438, + 0.005674170330166817, + 0.003650601487606764, + 0.006963315419852734, + 0.002190816681832075, + 0.002741089090704918, + 0.002169084269553423, + 0.0025450277607887983, + 0.005309435073286295, + 0.0053777205757796764, + 0.004479978233575821, + 0.004130511078983545, + 0.0027920252177864313, + 0.004012940917164087, + 0.0045522963628172874, + 0.003991030156612396, + 0.005767136812210083, + 0.010599464178085327, + 0.0034210921730846167, + 0.005575384013354778, + 0.004685375839471817, + 0.004624494351446629, + 0.003358820453286171, + 0.004301499109715223, + 0.006890412885695696, + 0.004238272551447153, + 0.006095879711210728, + 0.005016757640987635, + 0.0032675149850547314, + 0.007038075942546129, + 0.00476504024118185, + 0.00443057669326663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.12699945271015167 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of high-speed sports documentary cinematography, a low-angle, wide-angle camera slowly pans through a chaotic and perilous freight zone engulfed by raging flash floodwaters. Beneath a dark, turbulent sky, a powerful, muddy brown torrent of water violently rushes through a narrow canyon formed by towering stacks of multicolored corrugated shipping containers. These containers, hues of faded red, deep blue, and rust-streaked white, are dented and smooth from rain, their surfaces plastered with unmistakable, recognizable hazard symbols—black flames on an orange background, a skull and crossbones, and icons of corrosive liquids—as stark warnings. The camera pans, locked onto the specific point of impact, capturing the moment in ultra-high-speed capture. The surging floodwaters crash with immense force against the base of the container stack, the scene rendered in dramatic, sharp detail: the water's surface explodes into a crystalline sculpture of spray, individual droplets and fine mist suspended in the air, each catching the chill of the ambient light. A slight slow-motion tremor emanating from the massive steel containers reveals the sheer force of the impact, creating an atmosphere of tension and awe, highlighting the raw, untamed power of nature.", + "seed": 20133, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_7/index_3286_seed_20133_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03226774726063013, + "vlm_logprob": -0.03226774726063013, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.03226774726063013, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007120470982044935 + ], + "opposite_logprob_tokens": [ + -6.757120609283447 + ], + "logprob_reward": -0.007120470982044935, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.007120470982044935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.587222695350647, + 0.0893540307879448, + 0.06668652594089508, + 0.04683288186788559, + 0.03585295379161835, + 0.03033529594540596, + 0.04336564242839813, + 0.02488848753273487, + 0.03793661668896675, + 0.03752490505576134 + ], + "credit_map": [ + 0.054356563836336136, + 0.1758040487766266, + 0.023254066705703735, + 0.03387588635087013, + 0.04749849811196327, + 0.15694496035575867, + 0.06505507975816727, + 0.03043351322412491, + 0.007921038195490837, + 0.009038064628839493, + 0.005057172849774361, + 0.010175599716603756, + 0.020306941121816635, + 0.013640044257044792, + 0.010086673311889172, + 0.013128483667969704, + 0.007301393896341324, + 0.009794967249035835, + 0.010558949783444405, + 0.010914687998592854, + 0.005578579846769571, + 0.0041132341139018536, + 0.010525799356400967, + 0.00789890717715025, + 0.005586667452007532, + 0.008422433398663998, + 0.003426734823733568, + 0.006015627179294825, + 0.005518173333257437, + 0.004015911370515823, + 0.008744797669351101, + 0.00510253245010972, + 0.003306060563772917, + 0.0033940852154046297, + 0.004036810249090195, + 0.0036684530787169933, + 0.00340022100135684, + 0.003917401190847158, + 0.005264520179480314, + 0.008865397423505783, + 0.0031050415709614754, + 0.0027878780383616686, + 0.0036805428098887205, + 0.0069074309431016445, + 0.003232107264921069, + 0.0038426266983151436, + 0.0039076050743460655, + 0.002872059354558587, + 0.004540883935987949, + 0.004758425056934357, + 0.0054479981772601604, + 0.0076603880152106285, + 0.004310396499931812, + 0.005439009517431259, + 0.005882772617042065, + 0.005325767211616039, + 0.0026335176080465317, + 0.0020777019672095776, + 0.00243878155015409, + 0.006211531814187765, + 0.0024881833232939243, + 0.0054033612832427025, + 0.001845076447352767, + 0.0017903306288644671, + 0.004560929257422686, + 0.0030413735657930374, + 0.003056400688365102, + 0.009093338623642921, + 0.0036998325958848, + 0.008597430773079395, + 0.0030128238722682, + 0.002874483820050955, + 0.005564655177295208, + 0.0031415338162332773, + 0.006397715769708157, + 0.006403552368283272, + 0.004138467833399773, + 0.004246370401233435, + 0.003018901916220784, + 0.00461370125412941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.007120470982044935 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608232781291008 + ], + "opposite_logprob_tokens": [ + -4.141082286834717 + ], + "logprob_reward": -0.01608232781291008, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608232781291008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3951433002948761, + 0.04801143705844879, + 0.03839701786637306, + 0.06884180754423141, + 0.052958350628614426, + 0.03590988367795944, + 0.07005394250154495, + 0.18156853318214417, + 0.06403195858001709, + 0.045083802193403244 + ], + "credit_map": [ + 0.04167589545249939, + 0.1315508931875229, + 0.012015591375529766, + 0.0204156506806612, + 0.03100227192044258, + 0.094013512134552, + 0.04785137623548508, + 0.01661807857453823, + 0.0053967759013175964, + 0.005090313032269478, + 0.00365247018635273, + 0.006722155027091503, + 0.008530061691999435, + 0.00686032185330987, + 0.005952023901045322, + 0.005807314999401569, + 0.003399619134142995, + 0.005648533347994089, + 0.0042344327084720135, + 0.005002932623028755, + 0.005083257798105478, + 0.0031972620636224747, + 0.006796037778258324, + 0.0050349365919828415, + 0.004628838039934635, + 0.007785415276885033, + 0.006773283705115318, + 0.01071277353912592, + 0.005331649444997311, + 0.005515926983207464, + 0.014703720808029175, + 0.013390195555984974, + 0.005233319941908121, + 0.004668783396482468, + 0.006094489712268114, + 0.005219359416514635, + 0.005376726854592562, + 0.0053066471591591835, + 0.007450167555361986, + 0.013608853332698345, + 0.003752796445041895, + 0.003046358237043023, + 0.00396018847823143, + 0.005691345315426588, + 0.0036954882089048624, + 0.004233209881931543, + 0.005206940695643425, + 0.006323553156107664, + 0.006593621335923672, + 0.006196100730448961, + 0.00836909282952547, + 0.015259991399943829, + 0.007009574212133884, + 0.009531400166451931, + 0.007607094012200832, + 0.009487064555287361, + 0.016427041962742805, + 0.01610790193080902, + 0.01444745808839798, + 0.05934169888496399, + 0.013145687989890575, + 0.04129576310515404, + 0.012025878764688969, + 0.00877709686756134, + 0.006566139403730631, + 0.007002390455454588, + 0.005998468492180109, + 0.009995102882385254, + 0.011555950157344341, + 0.010965164750814438, + 0.00662775244563818, + 0.005320984870195389, + 0.005140537396073341, + 0.0043905265629291534, + 0.00704991864040494, + 0.005089163314551115, + 0.006644907873123884, + 0.005979203153401613, + 0.004313334356993437, + 0.006476202979683876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.01608232781291008 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329086884856224 + ], + "opposite_logprob_tokens": [ + -3.7732908725738525 + ], + "logprob_reward": -0.02329086884856224, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329086884856224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37254032492637634, + 0.05803970992565155, + 0.044714391231536865, + 0.04898630082607269, + 0.05711975321173668, + 0.030216945335268974, + 0.04849792644381523, + 0.27500468492507935, + 0.04415727034211159, + 0.02072271704673767 + ], + "credit_map": [ + 0.04635930061340332, + 0.1332671344280243, + 0.009621742181479931, + 0.01643872819840908, + 0.026732102036476135, + 0.08071950078010559, + 0.04589100182056427, + 0.013510817661881447, + 0.005105264019221067, + 0.006254495121538639, + 0.006741445977240801, + 0.010444465093314648, + 0.006600257474929094, + 0.004558529704809189, + 0.00822159182280302, + 0.010113663040101528, + 0.004270550329238176, + 0.004963708575814962, + 0.004327854607254267, + 0.0049989246763288975, + 0.006773775443434715, + 0.0060244216583669186, + 0.006375118624418974, + 0.006980037782341242, + 0.0066748857498168945, + 0.008645639754831791, + 0.0037051329854875803, + 0.005615760572254658, + 0.005268638022243977, + 0.005367150064557791, + 0.007956634275615215, + 0.0057524582371115685, + 0.005743531044572592, + 0.005769955925643444, + 0.006791490130126476, + 0.005951674655079842, + 0.006852258928120136, + 0.0051761590875685215, + 0.009369763545691967, + 0.011464918032288551, + 0.0033045965246856213, + 0.0031701528932899237, + 0.005013341084122658, + 0.003419493790715933, + 0.0046386378817260265, + 0.003888913430273533, + 0.004310492426156998, + 0.0024713182356208563, + 0.004814078100025654, + 0.004555473569780588, + 0.0060732560232281685, + 0.009268772788345814, + 0.00453146081417799, + 0.0062119667418301105, + 0.005099670495837927, + 0.007943244650959969, + 0.024885933846235275, + 0.02480425499379635, + 0.020969323813915253, + 0.09271840751171112, + 0.019457625225186348, + 0.062381938099861145, + 0.017769161611795425, + 0.012018011882901192, + 0.005045521073043346, + 0.002968010026961565, + 0.003628690028563142, + 0.011666997335851192, + 0.003780422965064645, + 0.011301728896796703, + 0.002986862091347575, + 0.0027790372259914875, + 0.002984838094562292, + 0.0018046853365376592, + 0.002863265573978424, + 0.0027405600994825363, + 0.0022131623700261116, + 0.0024300948716700077, + 0.002259382512420416, + 0.003426728770136833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.02329086884856224 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019364668987691402 + ], + "opposite_logprob_tokens": [ + -6.251936435699463 + ], + "logprob_reward": -0.0019364668987691402, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019364668987691402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8618046641349792, + 0.024070004001259804, + 0.014261020347476006, + 0.01816001906991005, + 0.013396698981523514, + 0.009492888115346432, + 0.012141169048845768, + 0.023723255842924118, + 0.01887812651693821, + 0.004072175361216068 + ], + "credit_map": [ + 0.10440545529127121, + 0.3086197078227997, + 0.022436384111642838, + 0.040398404002189636, + 0.06201966106891632, + 0.18770523369312286, + 0.10408014059066772, + 0.03213969245553017, + 0.0018944642506539822, + 0.00233303289860487, + 0.0021772875916212797, + 0.0038845788221806288, + 0.0038627982139587402, + 0.0021501625888049603, + 0.0033147172071039677, + 0.004452962893992662, + 0.0012711380841210485, + 0.0016170874005183578, + 0.0023454551119357347, + 0.001979319378733635, + 0.0019400576129555702, + 0.0018155454890802503, + 0.0015368734020739794, + 0.001755544333718717, + 0.0016939081251621246, + 0.0025517093017697334, + 0.0016956233885139227, + 0.0022008204832673073, + 0.001857409835793078, + 0.0016690941993147135, + 0.003227038076147437, + 0.0032644164748489857, + 0.0013545112451538444, + 0.001314405701123178, + 0.001574072171933949, + 0.0014574702363461256, + 0.0009725245763547719, + 0.001308263512328267, + 0.0022528518456965685, + 0.0031625996343791485, + 0.0011098077520728111, + 0.0008516458328813314, + 0.001709944335743785, + 0.0013699379051104188, + 0.0010761960875242949, + 0.0010184937855228782, + 0.0014990429626777768, + 0.0008578195120207965, + 0.0010021295165643096, + 0.0011965358862653375, + 0.0010051570134237409, + 0.0029124829452484846, + 0.0012437786208465695, + 0.0016773083480075002, + 0.0014738679165020585, + 0.0016299092676490545, + 0.0021891645155847073, + 0.002121337689459324, + 0.001888651866465807, + 0.007785577792674303, + 0.0017997793620452285, + 0.005276705138385296, + 0.0015948988730087876, + 0.0010671414202079177, + 0.0021617577876895666, + 0.0013354368275031447, + 0.0012771367328241467, + 0.005262777674943209, + 0.0014220633311197162, + 0.004987460561096668, + 0.0011598055716603994, + 0.001271686633117497, + 0.0005849897861480713, + 0.00044082989916205406, + 0.0005461046239361167, + 0.0005564861348830163, + 0.000418695155531168, + 0.0004960735095664859, + 0.0003799932892434299, + 0.0006490032537840307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0019364668987691402 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290860176086426 + ], + "opposite_logprob_tokens": [ + -2.2379086017608643 + ], + "logprob_reward": -0.11290860176086426, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290860176086426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7107685804367065, + 0.05889081209897995, + 0.019729530438780785, + 0.04395407438278198, + 0.021189339458942413, + 0.0323219858109951, + 0.026806816458702087, + 0.05136539041996002, + 0.022340431809425354, + 0.012632944621145725 + ], + "credit_map": [ + 0.09172312915325165, + 0.2587946951389313, + 0.020946580916643143, + 0.0338086299598217, + 0.04851415008306503, + 0.14552317559719086, + 0.08566004782915115, + 0.025798261165618896, + 0.006531056948006153, + 0.006830866914242506, + 0.009350343607366085, + 0.009553337469696999, + 0.006816599518060684, + 0.0031978555489331484, + 0.009519794024527073, + 0.007090962491929531, + 0.0015521729364991188, + 0.002870452357456088, + 0.003634556196630001, + 0.0033430105540901423, + 0.0017400068463757634, + 0.0015695486217737198, + 0.002921530744060874, + 0.0020982532296329737, + 0.004159699659794569, + 0.006049534305930138, + 0.004162789322435856, + 0.005801207851618528, + 0.006426243111491203, + 0.004481479525566101, + 0.0076602608896791935, + 0.0052128625102341175, + 0.0028917528688907623, + 0.002427775179967284, + 0.0023645740002393723, + 0.002327871508896351, + 0.0023393225856125355, + 0.00212612422183156, + 0.0027354301419109106, + 0.003976491279900074, + 0.003359558992087841, + 0.00342537765391171, + 0.0038190805353224277, + 0.005420828703790903, + 0.004009220749139786, + 0.003540989011526108, + 0.003743447596207261, + 0.005003484431654215, + 0.003994132857769728, + 0.0025746170431375504, + 0.00529653113335371, + 0.003462541848421097, + 0.0022763246670365334, + 0.002995364833623171, + 0.00239554257132113, + 0.0038117635995149612, + 0.006454966962337494, + 0.0047364975325763226, + 0.00498374504968524, + 0.011729378253221512, + 0.005175493657588959, + 0.008726418949663639, + 0.00477247778326273, + 0.0047864182852208614, + 0.0031352632213383913, + 0.0026990645565092564, + 0.002036281395703554, + 0.0022613797336816788, + 0.003875370603054762, + 0.004000295419245958, + 0.0018917761044576764, + 0.002441002056002617, + 0.0017368552507832646, + 0.0014414734905585647, + 0.0016175195341929793, + 0.0021683175582438707, + 0.0014746737433597445, + 0.0015991185791790485, + 0.0012727476423606277, + 0.001322240219451487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.11290860176086426 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of high-speed sports documentary cinematography, a low-angle, wide-angle camera slowly pans through a chaotic and perilous freight zone engulfed by raging flash floodwaters. Beneath a dark, turbulent sky, a powerful, muddy brown torrent of water violently rushes through a narrow canyon formed by towering stacks of multicolored corrugated shipping containers. These containers, hues of faded red, deep blue, and rust-streaked white, are dented and smooth from rain, their surfaces plastered with unmistakable, recognizable hazard symbols—black flames on an orange background, a skull and crossbones, and icons of corrosive liquids—as stark warnings. The camera pans, locked onto the specific point of impact, capturing the moment in ultra-high-speed capture. The surging floodwaters crash with immense force against the base of the container stack, the scene rendered in dramatic, sharp detail: the water's surface explodes into a crystalline sculpture of spray, individual droplets and fine mist suspended in the air, each catching the chill of the ambient light. A slight slow-motion tremor emanating from the massive steel containers reveals the sheer force of the impact, creating an atmosphere of tension and awe, highlighting the raw, untamed power of nature.", + "seed": 20134, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_7/index_3286_seed_20134_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03757133269682526, + "vlm_logprob": -0.03757133269682526, + "vlm_yes_no_margin": 4.300000023841858, + "vlm_token_credit": -0.03757133269682526, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005550093948841095 + ], + "opposite_logprob_tokens": [ + -6.255549907684326 + ], + "logprob_reward": -0.005550093948841095, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.005550093948841095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1143784373998642, + 0.25896182656288147, + 0.03220114856958389, + 0.05155326798558235, + 0.051908303052186966, + 0.07171575725078583, + 0.07474026829004288, + 0.13957272469997406, + 0.11943958699703217, + 0.08552872389554977 + ], + "credit_map": [ + 0.014880201779305935, + 0.012259619310498238, + 0.014512649737298489, + 0.030211711302399635, + 0.009147007949650288, + 0.010930478572845459, + 0.013634703122079372, + 0.008802059106528759, + 0.028739092871546745, + 0.018881909549236298, + 0.023976918309926987, + 0.03677589073777199, + 0.025741850957274437, + 0.0334259457886219, + 0.06345214694738388, + 0.027968060225248337, + 0.0035090388264507055, + 0.004294474143534899, + 0.00341382110491395, + 0.004354859236627817, + 0.003289847867563367, + 0.003950560465455055, + 0.0052185263484716415, + 0.004170022439211607, + 0.006578362081199884, + 0.007726295385509729, + 0.0068292515352368355, + 0.009262277744710445, + 0.004391597583889961, + 0.007068897597491741, + 0.004942883737385273, + 0.004753698594868183, + 0.006457075010985136, + 0.00504717230796814, + 0.0074000610038638115, + 0.007639562711119652, + 0.005534683354198933, + 0.008915605954825878, + 0.005033491645008326, + 0.005880648735910654, + 0.0069205621257424355, + 0.006394919473677874, + 0.012339509092271328, + 0.011838975362479687, + 0.006782687269151211, + 0.010572077706456184, + 0.008595925755798817, + 0.008271095342934132, + 0.0042885891161859035, + 0.0044554187916219234, + 0.0172682274132967, + 0.012705390341579914, + 0.005470274016261101, + 0.008740824647247791, + 0.012580848298966885, + 0.009230688214302063, + 0.00544773880392313, + 0.005083276424556971, + 0.048436667770147324, + 0.04632670432329178, + 0.009376218542456627, + 0.009541359730064869, + 0.008826387114822865, + 0.006534367799758911, + 0.011160346679389477, + 0.014387980103492737, + 0.019664201885461807, + 0.018799450248479843, + 0.01762707158923149, + 0.015059961937367916, + 0.010809777304530144, + 0.01193078514188528, + 0.006966604385524988, + 0.008287235163152218, + 0.012206951156258583, + 0.014682818204164505, + 0.008353721350431442, + 0.010510455816984177, + 0.011777341365814209, + 0.012743591330945492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.005550093948841095 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608690246939659 + ], + "opposite_logprob_tokens": [ + -4.141087055206299 + ], + "logprob_reward": -0.01608690246939659, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608690246939659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2699527144432068, + 0.202530175447464, + 0.09057628363370895, + 0.0724768415093422, + 0.04439177364110947, + 0.038521185517311096, + 0.04416331276297569, + 0.07006165385246277, + 0.09614259004592896, + 0.07118339091539383 + ], + "credit_map": [ + 0.04852937534451485, + 0.026296693831682205, + 0.021698419004678726, + 0.04330890253186226, + 0.02943608909845352, + 0.044429223984479904, + 0.034107666462659836, + 0.022146383300423622, + 0.05294125899672508, + 0.011610998772084713, + 0.011397402733564377, + 0.013386048376560211, + 0.012679020874202251, + 0.06904296576976776, + 0.01602127030491829, + 0.015451209619641304, + 0.009346879087388515, + 0.007259404752403498, + 0.008983759209513664, + 0.009915621019899845, + 0.016167355701327324, + 0.018307069316506386, + 0.010720794089138508, + 0.009875400923192501, + 0.0081250024959445, + 0.006562596652656794, + 0.019515590742230415, + 0.009823312051594257, + 0.004226890858262777, + 0.013114186003804207, + 0.005263453349471092, + 0.005845814477652311, + 0.007420889101922512, + 0.003671864978969097, + 0.008918137289583683, + 0.007361908443272114, + 0.0035254438407719135, + 0.005870241671800613, + 0.004092548508197069, + 0.003530744230374694, + 0.00506017217412591, + 0.00492860795930028, + 0.00763141131028533, + 0.005437188781797886, + 0.0048784008249640465, + 0.0027866961900144815, + 0.004635242745280266, + 0.003163467161357403, + 0.003230867674574256, + 0.0024012357462197542, + 0.009810585528612137, + 0.008218982256948948, + 0.0035260829608887434, + 0.004212461411952972, + 0.008011211641132832, + 0.004751886706799269, + 0.004594601225107908, + 0.005441426299512386, + 0.011356396600604057, + 0.010493187233805656, + 0.010046997107565403, + 0.009461933746933937, + 0.0102218147367239, + 0.008445301093161106, + 0.008087636902928352, + 0.011721624992787838, + 0.014269628562033176, + 0.012441786006093025, + 0.01481782179325819, + 0.013635450974106789, + 0.009692049585282803, + 0.011476595886051655, + 0.006151909939944744, + 0.006155773531645536, + 0.009240190498530865, + 0.0155721977353096, + 0.005607992876321077, + 0.007245820481330156, + 0.011787518858909607, + 0.009421990253031254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.01608690246939659 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819467917084694 + ], + "opposite_logprob_tokens": [ + -4.018194675445557 + ], + "logprob_reward": -0.01819467917084694, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819467917084694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13877712190151215, + 0.18662129342556, + 0.02972911298274994, + 0.08951100707054138, + 0.07746025919914246, + 0.12743540108203888, + 0.06687036901712418, + 0.07881958782672882, + 0.12639930844306946, + 0.07837656140327454 + ], + "credit_map": [ + 0.022724658250808716, + 0.016264617443084717, + 0.01419900730252266, + 0.018449241295456886, + 0.011416513472795486, + 0.011767479591071606, + 0.030416537076234818, + 0.013539067469537258, + 0.04723881185054779, + 0.010126600041985512, + 0.011200129054486752, + 0.011040650308132172, + 0.014305857010185719, + 0.06849890947341919, + 0.011839940212666988, + 0.01237039640545845, + 0.004185639787465334, + 0.005238030105829239, + 0.0033900097478181124, + 0.0037548525724560022, + 0.003580448217689991, + 0.0035025461111217737, + 0.0030259026680141687, + 0.0030516849365085363, + 0.013172519393265247, + 0.014283915050327778, + 0.012140709906816483, + 0.014984294772148132, + 0.00819175224751234, + 0.009331375360488892, + 0.008135021664202213, + 0.009271414019167423, + 0.011492066085338593, + 0.00657670060172677, + 0.014888474717736244, + 0.010714818723499775, + 0.007286855485290289, + 0.011791892349720001, + 0.007042996119707823, + 0.007666454184800386, + 0.01420158613473177, + 0.016920287162065506, + 0.024607926607131958, + 0.01732753962278366, + 0.01550570223480463, + 0.014509488828480244, + 0.015444036573171616, + 0.008918843232095242, + 0.0043633258901536465, + 0.0038613707292824984, + 0.019019586965441704, + 0.013303160667419434, + 0.004668421112000942, + 0.006030903197824955, + 0.009350437670946121, + 0.006273161619901657, + 0.004682340193539858, + 0.005261817015707493, + 0.018683139234781265, + 0.0166696198284626, + 0.00936392042785883, + 0.007345856167376041, + 0.00892632920295, + 0.007886559702455997, + 0.009922080673277378, + 0.014555567875504494, + 0.01875140145421028, + 0.01914866641163826, + 0.01932583749294281, + 0.017534375190734863, + 0.013662978075444698, + 0.013498406857252121, + 0.006880401633679867, + 0.007510732859373093, + 0.010409926995635033, + 0.015314916148781776, + 0.007419332396239042, + 0.0069753252901136875, + 0.012831275351345539, + 0.011034657247364521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.01819467917084694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005246442276984453 + ], + "opposite_logprob_tokens": [ + -5.255246639251709 + ], + "logprob_reward": -0.005246442276984453, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005246442276984453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11507578194141388, + 0.1465553492307663, + 0.0336340107023716, + 0.12026254087686539, + 0.10952818393707275, + 0.07551318407058716, + 0.13939101994037628, + 0.12162179499864578, + 0.06045681610703468, + 0.07796128839254379 + ], + "credit_map": [ + 0.019092155620455742, + 0.01504526101052761, + 0.011730237863957882, + 0.023999158293008804, + 0.011736636981368065, + 0.011543571017682552, + 0.014121148735284805, + 0.007807603105902672, + 0.02900233305990696, + 0.01233369018882513, + 0.009323100559413433, + 0.012855972163379192, + 0.012458228506147861, + 0.04796897992491722, + 0.011342728510499, + 0.011270305141806602, + 0.005241078790277243, + 0.0048087541945278645, + 0.00488327257335186, + 0.004128049593418837, + 0.0028201104141771793, + 0.0041158259846270084, + 0.0038306291680783033, + 0.003806287655606866, + 0.016257233917713165, + 0.01670924760401249, + 0.020448051393032074, + 0.015520836226642132, + 0.010450436733663082, + 0.018877271562814713, + 0.009367910213768482, + 0.012631549499928951, + 0.01692681573331356, + 0.01122392900288105, + 0.01797846518456936, + 0.02145828865468502, + 0.007758290506899357, + 0.01481284573674202, + 0.008533170446753502, + 0.010836373083293438, + 0.007892784662544727, + 0.007686923258006573, + 0.015081902965903282, + 0.015479720197618008, + 0.010218831710517406, + 0.006621106993407011, + 0.007480080239474773, + 0.005051825195550919, + 0.0072114551439881325, + 0.007219020742923021, + 0.051348406821489334, + 0.03291789069771767, + 0.009900878183543682, + 0.007321185898035765, + 0.014997937716543674, + 0.00847423542290926, + 0.006273909471929073, + 0.005254077725112438, + 0.03672388941049576, + 0.03768850117921829, + 0.010096760466694832, + 0.007760421838611364, + 0.011342305690050125, + 0.006481919903308153, + 0.004880799446254969, + 0.0043083070777356625, + 0.011088431812822819, + 0.00977460015565157, + 0.006235212087631226, + 0.010579218156635761, + 0.00807222444564104, + 0.005518022924661636, + 0.006378254387527704, + 0.0066528767347335815, + 0.013006449677050114, + 0.01286013051867485, + 0.00879221223294735, + 0.00673525407910347, + 0.011805020272731781, + 0.01173108909279108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.005246442276984453 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a freight zone with towering stacks of multicolored corrugated shipping containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the containers colored in faded red, deep blue, and rust-streaked white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are recognizable hazard symbols visible on the containers, including black flames on orange, a skull and crossbones, and corrosive liquid icons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful, muddy brown torrent of water violently rushing through the canyon formed by the containers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan while locked onto the point of impact where the floodwaters crash against the containers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14277854561805725 + ], + "opposite_logprob_tokens": [ + -2.0177786350250244 + ], + "logprob_reward": -0.14277854561805725, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14277854561805725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06399524956941605, + 0.05845066159963608, + 0.07265995442867279, + 0.06460285931825638, + 0.06528365612030029, + 0.159919872879982, + 0.14863954484462738, + 0.1421077400445938, + 0.06380290538072586, + 0.16053761541843414 + ], + "credit_map": [ + 0.011940596625208855, + 0.0076813302002847195, + 0.007154266815632582, + 0.010037174448370934, + 0.005576743744313717, + 0.006877209525555372, + 0.00836488138884306, + 0.006363051477819681, + 0.008611870929598808, + 0.006227752659469843, + 0.0051418086513876915, + 0.0056414371356368065, + 0.006426644511520863, + 0.012644579634070396, + 0.008108172565698624, + 0.005648395977914333, + 0.005618032533675432, + 0.010215943679213524, + 0.007917112670838833, + 0.00904440600425005, + 0.007536837365478277, + 0.009215493686497211, + 0.013847707770764828, + 0.009264424443244934, + 0.006692367140203714, + 0.007400866132229567, + 0.009835680015385151, + 0.013366813771426678, + 0.0063636405393481255, + 0.00653870077803731, + 0.007743014022707939, + 0.0066617815755307674, + 0.01129025500267744, + 0.006326331757009029, + 0.011212495155632496, + 0.013204519636929035, + 0.005730406381189823, + 0.005506570916622877, + 0.007277100812643766, + 0.004735974594950676, + 0.01679166592657566, + 0.015064943581819534, + 0.03383360430598259, + 0.021684756502509117, + 0.017080657184123993, + 0.015615174546837807, + 0.019434960559010506, + 0.02041410468518734, + 0.009988580830395222, + 0.009105455130338669, + 0.0383717343211174, + 0.028275664895772934, + 0.010525343008339405, + 0.018512798473238945, + 0.01774788461625576, + 0.0161120742559433, + 0.00581207824870944, + 0.00677197752520442, + 0.04725748673081398, + 0.043370310217142105, + 0.009100529365241528, + 0.011216824874281883, + 0.010995942167937756, + 0.007582595106214285, + 0.00580281438305974, + 0.005744843743741512, + 0.006641591899096966, + 0.00720733217895031, + 0.006210379768162966, + 0.013666421175003052, + 0.008517223410308361, + 0.010012301616370678, + 0.012579689733684063, + 0.01242239959537983, + 0.020988088101148605, + 0.02500411495566368, + 0.014593340456485748, + 0.02222583070397377, + 0.02116253413259983, + 0.031561631709337234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.14277854561805725 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_8/index_tensor([9171, 9171])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_8/index_tensor([9171, 9171])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5d450b22c7ae109e581efa96832edd58faf52b51 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_8/index_tensor([9171, 9171])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level close-up centers on a half-eaten sandwich, placed on a rustic off-white ceramic plate, resting on the warm, grainy surface of a wooden cafe table. Soft golden light from a nearby window illuminates the scene, casting gentle shadows. A clean bite into the sandwich reveals its layers in crisp detail: toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey. A bokeh of warm fairy lights in the softly blurred background hints at the café's cozy atmosphere. Suddenly, unsettlingly, the sandwich's structure begins to fail; the hard bread begins to sag and soften, and the solid filling loses its shape and begins to melt downward. As the disparate elements dissolve into one another, the process accelerates, transforming into a gooey, swirling puddle of color on the plate. The green of the lettuce, the red of the tomato, and the pale tones of the turkey and bread bleed and marble together, forming a mesmerizing, kaleidoscopic liquid that slowly swirls, its surface catching the light with an iridescent, oily sheen, creating a surreal and dreamlike atmosphere that contrasts sharply with the initially cozy setting.", + "seed": 24569, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_8/index_9171_seed_24569_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.015656446386128665, + "vlm_logprob": -0.015656446386128665, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.015656446386128665, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005432841368019581 + ], + "opposite_logprob_tokens": [ + -6.630432605743408 + ], + "logprob_reward": -0.005432841368019581, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.005432841368019581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10384152084589005, + 0.030543716624379158, + 0.03327950835227966, + 0.05112726613879204, + 0.07202798873186111, + 0.3584309220314026, + 0.12406732887029648, + 0.07774695754051208, + 0.04094954580068588, + 0.10798521339893341 + ], + "credit_map": [ + 0.01161863747984171, + 0.01765788160264492, + 0.007812055759131908, + 0.008110694587230682, + 0.016438135877251625, + 0.014360967092216015, + 0.010534756816923618, + 0.017308395355939865, + 0.003957327455282211, + 0.003990114200860262, + 0.003183080581948161, + 0.003405693219974637, + 0.004740660544484854, + 0.004006593022495508, + 0.003003510879352689, + 0.00425673695281148, + 0.0035598413087427616, + 0.004847688600420952, + 0.003926684148609638, + 0.004398616496473551, + 0.004636223427951336, + 0.0042629726231098175, + 0.004201299510896206, + 0.003446183167397976, + 0.0050530508160591125, + 0.00416463753208518, + 0.004526031203567982, + 0.005952529143542051, + 0.010066376067698002, + 0.007782801054418087, + 0.006437254138290882, + 0.007144593633711338, + 0.007431893143802881, + 0.010459252633154392, + 0.008226905949413776, + 0.008035042323172092, + 0.01129137072712183, + 0.011235573329031467, + 0.009867978282272816, + 0.005479977000504732, + 0.03286006674170494, + 0.04270845279097557, + 0.022219838574528694, + 0.03161569684743881, + 0.10445353388786316, + 0.060139596462249756, + 0.03962511941790581, + 0.024808648973703384, + 0.009660448879003525, + 0.012099848128855228, + 0.007938547991216183, + 0.011182436719536781, + 0.03741854056715965, + 0.022696228697896004, + 0.013767805881798267, + 0.009303483180701733, + 0.007072508335113525, + 0.009239455685019493, + 0.007812491152435541, + 0.01150477584451437, + 0.01272265799343586, + 0.009229831397533417, + 0.012235655449330807, + 0.007929581217467785, + 0.005556242540478706, + 0.006347968243062496, + 0.0033691751305013895, + 0.003812571056187153, + 0.0077667091973125935, + 0.0058930786326527596, + 0.004958796314895153, + 0.0032450079452246428, + 0.008604506962001324, + 0.015486621297895908, + 0.009252354502677917, + 0.019149750471115112, + 0.0213418360799551, + 0.01544218324124813, + 0.012432899326086044, + 0.006275068502873182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.005432841368019581 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036467495374381542 + ], + "opposite_logprob_tokens": [ + -5.6286468505859375 + ], + "logprob_reward": -0.0036467495374381542, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036467495374381542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15475913882255554, + 0.0371258407831192, + 0.05821666121482849, + 0.06097075343132019, + 0.0787058174610138, + 0.1615862101316452, + 0.09508683532476425, + 0.1365242749452591, + 0.08066118508577347, + 0.13636331260204315 + ], + "credit_map": [ + 0.023010725155472755, + 0.0233631394803524, + 0.012324610725045204, + 0.01688159629702568, + 0.017853431403636932, + 0.02189280465245247, + 0.016245746985077858, + 0.023187076672911644, + 0.003075046930462122, + 0.003791382536292076, + 0.004211983177810907, + 0.004907970782369375, + 0.005781012121587992, + 0.004776105284690857, + 0.006823607720434666, + 0.00375872990116477, + 0.005709568504244089, + 0.0073098596185445786, + 0.006799541879445314, + 0.0067208874970674515, + 0.007705566938966513, + 0.00742708845064044, + 0.007021073251962662, + 0.009523073211312294, + 0.005902257282286882, + 0.006430360022932291, + 0.005373877473175526, + 0.008370271883904934, + 0.0125783272087574, + 0.009029443375766277, + 0.008646255359053612, + 0.0046399617567658424, + 0.00788826122879982, + 0.009149283170700073, + 0.009292155504226685, + 0.010177656076848507, + 0.012345959432423115, + 0.011827225796878338, + 0.011855576187372208, + 0.006169699598103762, + 0.015619310550391674, + 0.01729239709675312, + 0.009422896429896355, + 0.020014924928545952, + 0.04349380359053612, + 0.026541095227003098, + 0.017441481351852417, + 0.011760300025343895, + 0.020852966234087944, + 0.014091456308960915, + 0.005860702600330114, + 0.00798751413822174, + 0.01701916567981243, + 0.016819851472973824, + 0.005703231319785118, + 0.006751943379640579, + 0.018130211159586906, + 0.0225086472928524, + 0.016614610329270363, + 0.017978457733988762, + 0.018544694408774376, + 0.018212810158729553, + 0.014290392398834229, + 0.010244440287351608, + 0.00992122758179903, + 0.010239723138511181, + 0.00520976772531867, + 0.012560402974486351, + 0.012965165078639984, + 0.012725417502224445, + 0.010597399435937405, + 0.006442077457904816, + 0.013339501805603504, + 0.029366588220000267, + 0.009459408931434155, + 0.019853131845593452, + 0.02401283197104931, + 0.017191918566823006, + 0.01381127443164587, + 0.009328649379312992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.0036467495374381542 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309479892253876 + ], + "opposite_logprob_tokens": [ + -3.1680948734283447 + ], + "logprob_reward": -0.04309479892253876, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309479892253876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05698896571993828, + 0.03198562189936638, + 0.058886293321847916, + 0.11995544284582138, + 0.14248640835285187, + 0.14262551069259644, + 0.08871239423751831, + 0.22549809515476227, + 0.05511175096035004, + 0.07774951308965683 + ], + "credit_map": [ + 0.005807893350720406, + 0.008220646530389786, + 0.005253060720860958, + 0.008586577139794827, + 0.007455191109329462, + 0.007645912002772093, + 0.009337475523352623, + 0.004682208877056837, + 0.0044000763446092606, + 0.00420039426535368, + 0.002705724211409688, + 0.0038942787796258926, + 0.005235796794295311, + 0.004714420530945063, + 0.00421038456261158, + 0.002624547341838479, + 0.005134499166160822, + 0.009510040283203125, + 0.004861526191234589, + 0.005740925669670105, + 0.013296769000589848, + 0.008351695723831654, + 0.0056905061937868595, + 0.006300334352999926, + 0.011948722414672375, + 0.01273215003311634, + 0.009376775473356247, + 0.014147871173918247, + 0.02319415844976902, + 0.018083391711115837, + 0.016359413042664528, + 0.014112971723079681, + 0.017292730510234833, + 0.024800186976790428, + 0.011352542787790298, + 0.020819339901208878, + 0.016160277649760246, + 0.022544652223587036, + 0.018958680331707, + 0.010557996109127998, + 0.021366296336054802, + 0.021107472479343414, + 0.009944991208612919, + 0.015344768762588501, + 0.027711976319551468, + 0.021966971457004547, + 0.01697838492691517, + 0.008204656653106213, + 0.013185913674533367, + 0.009883532300591469, + 0.005392313469201326, + 0.009526345878839493, + 0.021257247775793076, + 0.012834915891289711, + 0.009919144213199615, + 0.006712982896715403, + 0.03612019121646881, + 0.0206436924636364, + 0.01513038668781519, + 0.031745828688144684, + 0.03429075703024864, + 0.0364849790930748, + 0.033652521669864655, + 0.017429737374186516, + 0.004774859640747309, + 0.009435398504137993, + 0.0029301627073436975, + 0.006085649132728577, + 0.010631748475134373, + 0.009435461834073067, + 0.00691588968038559, + 0.004902581684291363, + 0.00966173131018877, + 0.018054179847240448, + 0.008591516874730587, + 0.011708274483680725, + 0.00860598124563694, + 0.010586398653686047, + 0.005901934113353491, + 0.004639493301510811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.04309479892253876 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002821514382958412 + ], + "opposite_logprob_tokens": [ + -5.877821445465088 + ], + "logprob_reward": -0.002821514382958412, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002821514382958412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05687296390533447, + 0.01856115646660328, + 0.028775529935956, + 0.08790939301252365, + 0.05845768004655838, + 0.3262440860271454, + 0.11675607413053513, + 0.18095852434635162, + 0.04192424938082695, + 0.08354036509990692 + ], + "credit_map": [ + 0.005830564070492983, + 0.009846559725701809, + 0.0047417315654456615, + 0.007674700114876032, + 0.005450459662824869, + 0.008012992329895496, + 0.006180536933243275, + 0.009135422296822071, + 0.0023815359454602003, + 0.004198293201625347, + 0.001786386943422258, + 0.0022094517480582, + 0.0022166187409311533, + 0.0026774441357702017, + 0.0017022551037371159, + 0.0013891715789213777, + 0.0030512860976159573, + 0.004985302221029997, + 0.0028598727658391, + 0.0032050879672169685, + 0.004102976992726326, + 0.003963174764066935, + 0.002596821403130889, + 0.0040110074914991856, + 0.008881420828402042, + 0.008730803616344929, + 0.0047584944404661655, + 0.00888359360396862, + 0.02266712300479412, + 0.014995845966041088, + 0.01116105541586876, + 0.007831066846847534, + 0.005922360345721245, + 0.008024833165109158, + 0.0071967295370996, + 0.007610809523612261, + 0.007862132973968983, + 0.008943798951804638, + 0.007470795419067144, + 0.005426221992820501, + 0.030728649348020554, + 0.04016784206032753, + 0.020751245319843292, + 0.03348669409751892, + 0.09250196069478989, + 0.05461593717336655, + 0.033746387809515, + 0.020245343446731567, + 0.01237091701477766, + 0.014482502825558186, + 0.007654799148440361, + 0.012903501279652119, + 0.03226923570036888, + 0.01774129644036293, + 0.010970212519168854, + 0.008363607339560986, + 0.03953484445810318, + 0.01666180044412613, + 0.011855094693601131, + 0.020050130784511566, + 0.030287303030490875, + 0.02885487861931324, + 0.01947101578116417, + 0.014243442565202713, + 0.006347321905195713, + 0.007035401184111834, + 0.002860387321561575, + 0.0038880007341504097, + 0.007038627285510302, + 0.006314166821539402, + 0.004428387153893709, + 0.004011953715234995, + 0.010160564444959164, + 0.01878710649907589, + 0.007533317431807518, + 0.012456335127353668, + 0.011092976666986942, + 0.012598681263625622, + 0.0059716771356761456, + 0.004939706064760685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.002821514382958412 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023286327719688416 + ], + "opposite_logprob_tokens": [ + -3.7732863426208496 + ], + "logprob_reward": -0.023286327719688416, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023286327719688416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04165719076991081, + 0.017236467450857162, + 0.01983242854475975, + 0.029214218258857727, + 0.042135532945394516, + 0.20804919302463531, + 0.09034599363803864, + 0.4314824044704437, + 0.048410430550575256, + 0.07163603603839874 + ], + "credit_map": [ + 0.00497262179851532, + 0.005823464132845402, + 0.0032978246454149485, + 0.004503190517425537, + 0.005420200061053038, + 0.005582704674452543, + 0.006126494612544775, + 0.00593069801107049, + 0.0031250964384526014, + 0.003286374034360051, + 0.0014880956150591373, + 0.001604673103429377, + 0.0017236057901754975, + 0.002430510241538286, + 0.0017752388957887888, + 0.0018028768245130777, + 0.0018260100623592734, + 0.0034888547379523516, + 0.0019780918955802917, + 0.002365513239055872, + 0.0034964613150805235, + 0.002896546386182308, + 0.001977454172447324, + 0.001803501625545323, + 0.0030234227888286114, + 0.0028600136283785105, + 0.0019978820346295834, + 0.002774217864498496, + 0.007283484097570181, + 0.005109577439725399, + 0.0031331812497228384, + 0.003032443579286337, + 0.0046838559210300446, + 0.007407604716718197, + 0.004220908973366022, + 0.004708444699645042, + 0.006640538573265076, + 0.006710223387926817, + 0.004839587956666946, + 0.002924375468865037, + 0.019981808960437775, + 0.024940090253949165, + 0.012634264305233955, + 0.019229941070079803, + 0.05914047732949257, + 0.03474783897399902, + 0.02232932113111019, + 0.015045507811009884, + 0.020375236868858337, + 0.010974569246172905, + 0.006034805439412594, + 0.007187808398157358, + 0.01791537180542946, + 0.014996984042227268, + 0.005819143261760473, + 0.007042097859084606, + 0.08178737759590149, + 0.03811163827776909, + 0.029147103428840637, + 0.05394691228866577, + 0.0708397626876831, + 0.0629301369190216, + 0.05928049609065056, + 0.03543904796242714, + 0.006142891477793455, + 0.006732362788170576, + 0.002906864508986473, + 0.005992563907057047, + 0.011082549579441547, + 0.007358790375292301, + 0.005424319766461849, + 0.0027701023500412703, + 0.00531808752566576, + 0.010468884371221066, + 0.006163489073514938, + 0.011488387361168861, + 0.013147925958037376, + 0.010752702131867409, + 0.008589004166424274, + 0.005707571282982826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.023286327719688416 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up centers on a half-eaten sandwich, placed on a rustic off-white ceramic plate, resting on the warm, grainy surface of a wooden cafe table. Soft golden light from a nearby window illuminates the scene, casting gentle shadows. A clean bite into the sandwich reveals its layers in crisp detail: toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey. A bokeh of warm fairy lights in the softly blurred background hints at the café's cozy atmosphere. Suddenly, unsettlingly, the sandwich's structure begins to fail; the hard bread begins to sag and soften, and the solid filling loses its shape and begins to melt downward. As the disparate elements dissolve into one another, the process accelerates, transforming into a gooey, swirling puddle of color on the plate. The green of the lettuce, the red of the tomato, and the pale tones of the turkey and bread bleed and marble together, forming a mesmerizing, kaleidoscopic liquid that slowly swirls, its surface catching the light with an iridescent, oily sheen, creating a surreal and dreamlike atmosphere that contrasts sharply with the initially cozy setting.", + "seed": 24570, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_8/index_9171_seed_24570_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03415848091244698, + "vlm_logprob": -0.03415848091244698, + "vlm_yes_no_margin": 4.399999976158142, + "vlm_token_credit": -0.03415848091244698, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004433208145201206 + ], + "opposite_logprob_tokens": [ + -6.2544331550598145 + ], + "logprob_reward": -0.004433208145201206, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.004433208145201206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14554837346076965, + 0.18424636125564575, + 0.06651398539543152, + 0.08738753944635391, + 0.11984768509864807, + 0.06620528548955917, + 0.08216460794210434, + 0.1504317671060562, + 0.07401920855045319, + 0.0236351415514946 + ], + "credit_map": [ + 0.02734045684337616, + 0.017417073249816895, + 0.015287977643311024, + 0.012165031395852566, + 0.01058156881481409, + 0.01887059211730957, + 0.010352907702326775, + 0.033532772213220596, + 0.032502394169569016, + 0.008312969468533993, + 0.009582878090441227, + 0.009264333173632622, + 0.006432740483433008, + 0.04329603165388107, + 0.010840761475265026, + 0.0640142560005188, + 0.011603507213294506, + 0.006758670322597027, + 0.0036049969494342804, + 0.005594609770923853, + 0.007404749747365713, + 0.010827244259417057, + 0.007535726297646761, + 0.013184481300413609, + 0.013973057270050049, + 0.013073885813355446, + 0.006137806922197342, + 0.009200994856655598, + 0.010786843486130238, + 0.01424474362283945, + 0.012635595165193081, + 0.007334613241255283, + 0.0199452955275774, + 0.0132621293887496, + 0.008843081071972847, + 0.01907864771783352, + 0.014023849740624428, + 0.01795266941189766, + 0.01796378381550312, + 0.008778229355812073, + 0.008636287413537502, + 0.008946176618337631, + 0.00610686931759119, + 0.010065825656056404, + 0.00987754762172699, + 0.007140962406992912, + 0.009430055506527424, + 0.006001563277095556, + 0.012207603082060814, + 0.012372135184705257, + 0.005173676181584597, + 0.014256051741540432, + 0.007160975597798824, + 0.010650365613400936, + 0.012449891306459904, + 0.007893909700214863, + 0.014634541235864162, + 0.015801427885890007, + 0.011230066418647766, + 0.028343139216303825, + 0.010739972814917564, + 0.015609093010425568, + 0.03431108593940735, + 0.01976245455443859, + 0.018978282809257507, + 0.005951453931629658, + 0.0026417619083076715, + 0.0050186458975076675, + 0.0056863343343138695, + 0.024943092837929726, + 0.005888262763619423, + 0.004911371506750584, + 0.0023180737625807524, + 0.0033947969786822796, + 0.001383479917421937, + 0.003085830481722951, + 0.0023052520118653774, + 0.005102002061903477, + 0.0017850989243015647, + 0.004260607995092869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.004433208145201206 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007662428542971611 + ], + "opposite_logprob_tokens": [ + -4.882662296295166 + ], + "logprob_reward": -0.007662428542971611, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007662428542971611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16274666786193848, + 0.28418806195259094, + 0.23222126066684723, + 0.033548127859830856, + 0.04594951495528221, + 0.0590365044772625, + 0.03834567964076996, + 0.08887657523155212, + 0.037993282079696655, + 0.017094261944293976 + ], + "credit_map": [ + 0.03587348759174347, + 0.02527143619954586, + 0.007753417827188969, + 0.012270539999008179, + 0.01351149845868349, + 0.022095493972301483, + 0.0110618332400918, + 0.03490898013114929, + 0.06052487716078758, + 0.010625665076076984, + 0.010684793815016747, + 0.013740181922912598, + 0.006996826734393835, + 0.06123797968029976, + 0.01638690009713173, + 0.10399088263511658, + 0.024567726999521255, + 0.03174888715147972, + 0.014966581016778946, + 0.017220880836248398, + 0.01839020662009716, + 0.05672413110733032, + 0.020311083644628525, + 0.04829179868102074, + 0.004254990257322788, + 0.006346153561025858, + 0.0026716270949691534, + 0.004176463931798935, + 0.003779029008001089, + 0.004059680737555027, + 0.004785545170307159, + 0.003474641591310501, + 0.006059466395527124, + 0.004680611193180084, + 0.002993145026266575, + 0.006216884590685368, + 0.005729204975068569, + 0.004936929326504469, + 0.011276530101895332, + 0.004056749865412712, + 0.007332926150411367, + 0.004559068474918604, + 0.005502177402377129, + 0.007534399628639221, + 0.004820037167519331, + 0.007530525326728821, + 0.009179818443953991, + 0.012577561661601067, + 0.007662705145776272, + 0.0038210577331483364, + 0.0020725929643958807, + 0.004526855889707804, + 0.004483864642679691, + 0.009757468476891518, + 0.0035819467157125473, + 0.002439195057377219, + 0.010736632160842419, + 0.008123739622533321, + 0.005914011038839817, + 0.01376964058727026, + 0.008602438494563103, + 0.014036719687283039, + 0.01593673601746559, + 0.011756663210690022, + 0.009611517190933228, + 0.004271265584975481, + 0.0013709833147004247, + 0.002623169682919979, + 0.0027696306351572275, + 0.012524029240012169, + 0.002263585804030299, + 0.002559103537350893, + 0.003567216219380498, + 0.0029401269275695086, + 0.001027231803163886, + 0.0017312790732830763, + 0.0019122071098536253, + 0.003217538585886359, + 0.0010236359667032957, + 0.0016750271897763014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.007662428542971611 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12710201740264893 + ], + "opposite_logprob_tokens": [ + -2.1271018981933594 + ], + "logprob_reward": -0.12710201740264893, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12710201740264893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12318450212478638, + 0.0824427604675293, + 0.0768706202507019, + 0.03634339198470116, + 0.20038670301437378, + 0.17796331644058228, + 0.07233352214097977, + 0.14300556480884552, + 0.060542136430740356, + 0.02692752704024315 + ], + "credit_map": [ + 0.026851609349250793, + 0.01005995087325573, + 0.006228500045835972, + 0.008233697153627872, + 0.00775429792702198, + 0.01740514114499092, + 0.009169853292405605, + 0.03748144954442978, + 0.02113891765475273, + 0.008157032541930676, + 0.0036017943639308214, + 0.0066976603120565414, + 0.005352128762751818, + 0.009450199082493782, + 0.006708846427500248, + 0.021336182951927185, + 0.008959080092608929, + 0.005373458843678236, + 0.00594824505969882, + 0.007809632457792759, + 0.005607434082776308, + 0.01983322761952877, + 0.007394381798803806, + 0.015945157036185265, + 0.0048736026510596275, + 0.0049981968477368355, + 0.0029055846389383078, + 0.004231957718729973, + 0.005406178068369627, + 0.00392036372795701, + 0.00498151034116745, + 0.005025993101298809, + 0.03298571705818176, + 0.01468015369027853, + 0.011012597940862179, + 0.029641741886734962, + 0.02893058769404888, + 0.02526933141052723, + 0.0435483455657959, + 0.014318209141492844, + 0.03092418983578682, + 0.010865696705877781, + 0.010894394479691982, + 0.02044137939810753, + 0.011885309591889381, + 0.038653500378131866, + 0.023645451292395592, + 0.03065338172018528, + 0.008132842369377613, + 0.011487184092402458, + 0.004867637064307928, + 0.010013616643846035, + 0.01008332148194313, + 0.008434855379164219, + 0.013423187658190727, + 0.005890870466828346, + 0.018546808511018753, + 0.016316194087266922, + 0.01022016815841198, + 0.02364404685795307, + 0.009458750486373901, + 0.018992669880390167, + 0.02764960378408432, + 0.018177324905991554, + 0.01032045017927885, + 0.0066168890334665775, + 0.004253603518009186, + 0.007380020804703236, + 0.0066359867341816425, + 0.014429178088903427, + 0.005295506678521633, + 0.005610492546111345, + 0.0038613826036453247, + 0.005700793582946062, + 0.002300754887983203, + 0.003661851631477475, + 0.003311535343527794, + 0.0029093101620674133, + 0.0022667350713163614, + 0.0029151621274650097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.12710201740264893 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253083072602749 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253083072602749, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253083072602749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1523730307817459, + 0.11233700811862946, + 0.03096640110015869, + 0.0398741252720356, + 0.18152360618114471, + 0.1811286211013794, + 0.06523274630308151, + 0.11716299504041672, + 0.07909534871578217, + 0.040306173264980316 + ], + "credit_map": [ + 0.020013902336359024, + 0.03609136864542961, + 0.009094645269215107, + 0.014246473088860512, + 0.014527237974107265, + 0.025501085445284843, + 0.01190891582518816, + 0.020989403128623962, + 0.019593918696045876, + 0.00704954331740737, + 0.005405230913311243, + 0.005945112090557814, + 0.004441560711711645, + 0.02477302774786949, + 0.006690232083201408, + 0.038438379764556885, + 0.0035530112218111753, + 0.007547781337052584, + 0.002308002207428217, + 0.002905267057940364, + 0.002840719185769558, + 0.004870804958045483, + 0.003841332159936428, + 0.003099482972174883, + 0.004063514061272144, + 0.004923919681459665, + 0.005629024002701044, + 0.003728593699634075, + 0.003862272948026657, + 0.006259798537939787, + 0.004279725719243288, + 0.007127278484404087, + 0.023407455533742905, + 0.016368156298995018, + 0.013303782790899277, + 0.028413940221071243, + 0.021366478875279427, + 0.015631072223186493, + 0.04688107967376709, + 0.016151633113622665, + 0.036186203360557556, + 0.011555071920156479, + 0.011939825490117073, + 0.018080973997712135, + 0.010795866139233112, + 0.04487883672118187, + 0.02017982490360737, + 0.027512017637491226, + 0.013940908946096897, + 0.005783178843557835, + 0.004388457629829645, + 0.008225037716329098, + 0.005513753741979599, + 0.01787475310266018, + 0.005339278373867273, + 0.004167374223470688, + 0.024810275062918663, + 0.015190265141427517, + 0.0076901875436306, + 0.008816619403660297, + 0.011512387543916702, + 0.026770779862999916, + 0.011535699479281902, + 0.010836784727871418, + 0.01879807375371456, + 0.009099863469600677, + 0.0032702009193599224, + 0.0073247747495770454, + 0.005632083397358656, + 0.02459598332643509, + 0.00461540138348937, + 0.005758970510214567, + 0.007385095581412315, + 0.006248589605093002, + 0.0027966646011918783, + 0.00533059099689126, + 0.0037018912844359875, + 0.00694621866568923, + 0.003171369666233659, + 0.0047257523983716965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.005253083072602749 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026341667398810387 + ], + "opposite_logprob_tokens": [ + -3.651341676712036 + ], + "logprob_reward": -0.026341667398810387, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026341667398810387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12144052237272263, + 0.14890696108341217, + 0.09103493392467499, + 0.02577195130288601, + 0.03057057224214077, + 0.202678382396698, + 0.05010506510734558, + 0.21405912935733795, + 0.07616347074508667, + 0.039268944412469864 + ], + "credit_map": [ + 0.013479499146342278, + 0.019600098952651024, + 0.008603441528975964, + 0.01196927111595869, + 0.013415883295238018, + 0.021845458075404167, + 0.013899161480367184, + 0.018627719953656197, + 0.03018300049006939, + 0.006233826279640198, + 0.0054258136078715324, + 0.007461536210030317, + 0.005196684971451759, + 0.031107638031244278, + 0.009119049645960331, + 0.054179415106773376, + 0.008448575623333454, + 0.00932461116462946, + 0.00512663321569562, + 0.006615862250328064, + 0.007752504199743271, + 0.02354774810373783, + 0.008588716387748718, + 0.021630292758345604, + 0.00418164161965251, + 0.0024004795122891665, + 0.0024785140994936228, + 0.0029453521128743887, + 0.0024713606107980013, + 0.0036552443634718657, + 0.0026468161959201097, + 0.004992546513676643, + 0.004050596617162228, + 0.00405254727229476, + 0.00295612053014338, + 0.00334324617870152, + 0.0039373780600726604, + 0.003938527777791023, + 0.005406777840107679, + 0.0028853793628513813, + 0.05137943476438522, + 0.010700601153075695, + 0.01211515162140131, + 0.01205480471253395, + 0.010764938779175282, + 0.060734014958143234, + 0.01582479290664196, + 0.029104655608534813, + 0.012983474880456924, + 0.00516266655176878, + 0.0020217669662088156, + 0.004547055810689926, + 0.004103815648704767, + 0.014160745777189732, + 0.0036565063055604696, + 0.00346904038451612, + 0.05208570882678032, + 0.02660241536796093, + 0.011376945301890373, + 0.015277731232345104, + 0.017191415652632713, + 0.04904814437031746, + 0.02370358258485794, + 0.018773211166262627, + 0.01980588398873806, + 0.008259715512394905, + 0.0027207755483686924, + 0.00458915252238512, + 0.005463253706693649, + 0.026353873312473297, + 0.004219902213662863, + 0.0047509209252893925, + 0.007897218689322472, + 0.00749228335916996, + 0.0025088887196034193, + 0.004420641344040632, + 0.0039313677698373795, + 0.006450397428125143, + 0.0028735408559441566, + 0.0036946088075637817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.026341667398810387 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up centers on a half-eaten sandwich, placed on a rustic off-white ceramic plate, resting on the warm, grainy surface of a wooden cafe table. Soft golden light from a nearby window illuminates the scene, casting gentle shadows. A clean bite into the sandwich reveals its layers in crisp detail: toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey. A bokeh of warm fairy lights in the softly blurred background hints at the café's cozy atmosphere. Suddenly, unsettlingly, the sandwich's structure begins to fail; the hard bread begins to sag and soften, and the solid filling loses its shape and begins to melt downward. As the disparate elements dissolve into one another, the process accelerates, transforming into a gooey, swirling puddle of color on the plate. The green of the lettuce, the red of the tomato, and the pale tones of the turkey and bread bleed and marble together, forming a mesmerizing, kaleidoscopic liquid that slowly swirls, its surface catching the light with an iridescent, oily sheen, creating a surreal and dreamlike atmosphere that contrasts sharply with the initially cozy setting.", + "seed": 24571, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_8/index_9171_seed_24571_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04666739767417312, + "vlm_logprob": -0.04666739767417312, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.04666739767417312, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005817505531013012 + ], + "opposite_logprob_tokens": [ + -5.630817413330078 + ], + "logprob_reward": -0.005817505531013012, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.005817505531013012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.517738938331604, + 0.051927536725997925, + 0.09210977703332901, + 0.031860485672950745, + 0.06064503639936447, + 0.03712223097681999, + 0.14186279475688934, + 0.026065750047564507, + 0.028535893186926842, + 0.012131544761359692 + ], + "credit_map": [ + 0.10406874865293503, + 0.056818392127752304, + 0.018591925501823425, + 0.020632298663258553, + 0.0648263618350029, + 0.2188129872083664, + 0.019361332058906555, + 0.01462689321488142, + 0.012999067083001137, + 0.008024320937693119, + 0.00289433216676116, + 0.007610616739839315, + 0.005795749835669994, + 0.007971096783876419, + 0.0032685885671526194, + 0.0033637613523751497, + 0.014025158248841763, + 0.011530996300280094, + 0.007031403016299009, + 0.006843936163932085, + 0.033234432339668274, + 0.010282277129590511, + 0.005304200574755669, + 0.003857377916574478, + 0.004697615280747414, + 0.0039578876458108425, + 0.0035947845317423344, + 0.00362565228715539, + 0.0034143077209591866, + 0.005769912153482437, + 0.00263397628441453, + 0.0041663506999611855, + 0.007584813050925732, + 0.007741019129753113, + 0.006177342496812344, + 0.006338211242109537, + 0.0055098929442465305, + 0.012467464432120323, + 0.006174854934215546, + 0.008651437237858772, + 0.004105914384126663, + 0.004039449617266655, + 0.002544035902246833, + 0.007526285946369171, + 0.004711867310106754, + 0.005980366840958595, + 0.005322861485183239, + 0.0028914485592395067, + 0.027724627405405045, + 0.013817974366247654, + 0.007004294544458389, + 0.010991412214934826, + 0.009288345463573933, + 0.05122446268796921, + 0.008518969640135765, + 0.013292711228132248, + 0.0042405566200613976, + 0.004342813976109028, + 0.002161359414458275, + 0.002240569796413183, + 0.003902221331372857, + 0.004987077787518501, + 0.002011840930208564, + 0.0021793104242533445, + 0.003540408331900835, + 0.00397961400449276, + 0.00253176037222147, + 0.006434530485421419, + 0.0028668351005762815, + 0.0034865669440478086, + 0.0034223601687699556, + 0.0022738182451575994, + 0.0017654476687312126, + 0.0011602469021454453, + 0.0015285080298781395, + 0.0018082951428368688, + 0.0008265853975899518, + 0.0014881499810144305, + 0.0016569389263167977, + 0.0018973726546391845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.005817505531013012 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004120432306081057 + ], + "opposite_logprob_tokens": [ + -5.504120349884033 + ], + "logprob_reward": -0.004120432306081057, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004120432306081057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12159311771392822, + 0.05411691218614578, + 0.4742469787597656, + 0.021912874653935432, + 0.06365619599819183, + 0.07613062113523483, + 0.07547085732221603, + 0.030260538682341576, + 0.02354707010090351, + 0.059064798057079315 + ], + "credit_map": [ + 0.016096265986561775, + 0.01786240004003048, + 0.006925170309841633, + 0.017374426126480103, + 0.022853637114167213, + 0.018926285207271576, + 0.01333585660904646, + 0.008219078183174133, + 0.010996795259416103, + 0.0049565876834094524, + 0.0032693487592041492, + 0.004554489627480507, + 0.017253125086426735, + 0.006213306915014982, + 0.0038701300509274006, + 0.0030031276401132345, + 0.07559091597795486, + 0.06315459311008453, + 0.02423522062599659, + 0.02671208046376705, + 0.1909560114145279, + 0.055840566754341125, + 0.022064972668886185, + 0.01569262333214283, + 0.002867000177502632, + 0.003365236334502697, + 0.0020583877339959145, + 0.0029829288832843304, + 0.0023927483707666397, + 0.002948713256046176, + 0.002732778899371624, + 0.002565080299973488, + 0.009855872020125389, + 0.007510081399232149, + 0.005270138382911682, + 0.006497937720268965, + 0.005904552061110735, + 0.01989346742630005, + 0.004917710088193417, + 0.0038064434193074703, + 0.008998437784612179, + 0.006311462260782719, + 0.004230642691254616, + 0.012718931771814823, + 0.005970808211714029, + 0.02123270183801651, + 0.011601245030760765, + 0.0050663938745856285, + 0.016408266499638557, + 0.00757636409252882, + 0.003694657701998949, + 0.004886557813733816, + 0.004612710326910019, + 0.02961152046918869, + 0.0031726588495075703, + 0.005508126690983772, + 0.0040088011883199215, + 0.004469920881092548, + 0.0021102228201925755, + 0.0035485513508319855, + 0.004753399174660444, + 0.006345361936837435, + 0.0022511137649416924, + 0.002773164538666606, + 0.003289882093667984, + 0.003318064147606492, + 0.0016981553053483367, + 0.0042983354069292545, + 0.004513615742325783, + 0.003040054813027382, + 0.0017544376896694303, + 0.0016345242038369179, + 0.008098792284727097, + 0.007718228735029697, + 0.008715327829122543, + 0.00867328979074955, + 0.004790271632373333, + 0.005905982572585344, + 0.005933669395744801, + 0.009229238145053387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.004120432306081057 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992590367794037 + ], + "opposite_logprob_tokens": [ + -1.8049259185791016 + ], + "logprob_reward": -0.17992590367794037, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992590367794037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15439939498901367, + 0.0712219700217247, + 0.4920653700828552, + 0.01895112916827202, + 0.026882406324148178, + 0.1362704634666443, + 0.03183997794985771, + 0.025278223678469658, + 0.02801753580570221, + 0.015073505230247974 + ], + "credit_map": [ + 0.029911978170275688, + 0.014233291149139404, + 0.005671394523233175, + 0.009624386206269264, + 0.02272639237344265, + 0.05958132445812225, + 0.006974691059440374, + 0.005675942171365023, + 0.007284705061465502, + 0.008267981931567192, + 0.0032681727316230536, + 0.002521697198972106, + 0.03015611134469509, + 0.014009921811521053, + 0.0027298976201564074, + 0.002983482787385583, + 0.08079811185598373, + 0.06575553119182587, + 0.023916717618703842, + 0.02774954028427601, + 0.19861674308776855, + 0.057116344571113586, + 0.022404786199331284, + 0.015707608312368393, + 0.0023751507978886366, + 0.0019129220163449645, + 0.0014311640989035368, + 0.002546894596889615, + 0.004654861520975828, + 0.0030116222333163023, + 0.001660146634094417, + 0.0013583668041974306, + 0.0055858902633190155, + 0.002910048235207796, + 0.0026203147135674953, + 0.002822181908413768, + 0.0027490335050970316, + 0.004185250960290432, + 0.0020253534894436598, + 0.0039843348786234856, + 0.017974119633436203, + 0.011843417771160603, + 0.006485720165073872, + 0.02176184020936489, + 0.011643990874290466, + 0.039269767701625824, + 0.019194504246115685, + 0.008097104728221893, + 0.005826496984809637, + 0.003101059701293707, + 0.002570507815107703, + 0.003392714075744152, + 0.0022626162972301245, + 0.010148308239877224, + 0.0017390268621966243, + 0.0027992497198283672, + 0.0029304546769708395, + 0.0027929346542805433, + 0.002528366632759571, + 0.00495662959292531, + 0.003416357096284628, + 0.004354163073003292, + 0.002826557494699955, + 0.001472760341130197, + 0.003194202668964863, + 0.0032728936057537794, + 0.001985744806006551, + 0.0034074424766004086, + 0.0062961820513010025, + 0.003291902132332325, + 0.004107828252017498, + 0.0024613405112177134, + 0.0018606727244332433, + 0.00215534376911819, + 0.0017741736955940723, + 0.0022485200315713882, + 0.0009503313922323287, + 0.001327823381870985, + 0.002353828400373459, + 0.00240281131118536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.17992590367794037 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785421192646027 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.009785421192646027, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785421192646027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2692480683326721, + 0.06042526289820671, + 0.19517453014850616, + 0.018551142886281013, + 0.028948456048965454, + 0.2805026173591614, + 0.07538983970880508, + 0.02337920293211937, + 0.038675203919410706, + 0.009705668315291405 + ], + "credit_map": [ + 0.057232316583395004, + 0.028498074039816856, + 0.009427906014025211, + 0.011753987520933151, + 0.03185741603374481, + 0.11170103400945663, + 0.011349950917065144, + 0.0074273645877838135, + 0.008820509538054466, + 0.005417692009359598, + 0.003945447038859129, + 0.0056950394064188, + 0.017189666628837585, + 0.009777862578630447, + 0.004903093911707401, + 0.004675951320677996, + 0.02996809594333172, + 0.026410657912492752, + 0.009591297246515751, + 0.01608196645975113, + 0.07413828372955322, + 0.021999500691890717, + 0.00993544701486826, + 0.007049279753118753, + 0.0019776129629462957, + 0.002187990350648761, + 0.001137213665060699, + 0.0020257390569895506, + 0.005463176406919956, + 0.0035229187924414873, + 0.0013278580736368895, + 0.0009086327045224607, + 0.004388964734971523, + 0.004187894519418478, + 0.0029960128013044596, + 0.0033794238697737455, + 0.004262408707290888, + 0.003642541589215398, + 0.0023743172641843557, + 0.0037168937269598246, + 0.03583124279975891, + 0.024396490305662155, + 0.01343405432999134, + 0.045237887650728226, + 0.023440076038241386, + 0.08068865537643433, + 0.040470946580171585, + 0.017003251239657402, + 0.016005530953407288, + 0.007028757128864527, + 0.004016150254756212, + 0.004834793508052826, + 0.004625838249921799, + 0.028269274160265923, + 0.0038756425492465496, + 0.00673384964466095, + 0.0027431827038526535, + 0.0025853205006569624, + 0.0018351246835663915, + 0.004307032562792301, + 0.003239188576117158, + 0.0047046467661857605, + 0.002249910729005933, + 0.0017147967591881752, + 0.004648365546017885, + 0.004751762840896845, + 0.002951704664155841, + 0.005298747681081295, + 0.008323676884174347, + 0.004681353457272053, + 0.004860912915319204, + 0.0031586824916303158, + 0.0013679617550224066, + 0.0010129919974133372, + 0.0011234452249482274, + 0.0009053658577613533, + 0.0008008104050531983, + 0.0014786894898861647, + 0.0012362096458673477, + 0.001780194230377674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.009785421192646027 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368772566318512 + ], + "opposite_logprob_tokens": [ + -3.4086878299713135 + ], + "logprob_reward": -0.03368772566318512, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368772566318512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49654173851013184, + 0.021533261984586716, + 0.348163902759552, + 0.01081962138414383, + 0.021708697080612183, + 0.02337849885225296, + 0.03787562623620033, + 0.021502545103430748, + 0.01219335850328207, + 0.006282757502049208 + ], + "credit_map": [ + 0.10690023005008698, + 0.0507882721722126, + 0.013034326024353504, + 0.014936063438653946, + 0.06893675774335861, + 0.22027947008609772, + 0.01050499640405178, + 0.011161644011735916, + 0.0037772117648273706, + 0.00280890055000782, + 0.0019207344157621264, + 0.003152816090732813, + 0.0031283951830118895, + 0.0029465709812939167, + 0.0017935418291017413, + 0.0020050914026796818, + 0.05880860984325409, + 0.046549439430236816, + 0.017004607245326042, + 0.016250627115368843, + 0.1433715671300888, + 0.04091883823275566, + 0.014794771559536457, + 0.01046545896679163, + 0.0013073831796646118, + 0.0012554681161418557, + 0.0008454027120023966, + 0.0013030406553298235, + 0.0029466114938259125, + 0.0015606957022100687, + 0.0008257589070126414, + 0.0007752613164484501, + 0.004035329446196556, + 0.0024402146227657795, + 0.0014218937139958143, + 0.001875569112598896, + 0.0020233173854649067, + 0.006735224276781082, + 0.0019187724683433771, + 0.0012583773350343108, + 0.0031455550342798233, + 0.002429558429867029, + 0.0015425638994202018, + 0.003488261252641678, + 0.0017620637081563473, + 0.006554005201905966, + 0.0030432110652327538, + 0.0014132802607491612, + 0.008045085705816746, + 0.003706388408318162, + 0.0017052913317456841, + 0.0024531595408916473, + 0.002245957963168621, + 0.01445014402270317, + 0.002126191044226289, + 0.003143406705930829, + 0.0027689978014677763, + 0.0032696882262825966, + 0.0014646603958681226, + 0.0022637532092630863, + 0.0036701944191008806, + 0.0045948028564453125, + 0.0018049607751891017, + 0.0016654862556606531, + 0.0015537695726379752, + 0.001918513560667634, + 0.0009068507933989167, + 0.001758275437168777, + 0.0022760964930057526, + 0.001649032928980887, + 0.0011346281971782446, + 0.0009961910545825958, + 0.0007896503084339201, + 0.0009513274417258799, + 0.0007845574873499572, + 0.0010341565357521176, + 0.00036423924029804766, + 0.0005359708447940648, + 0.0008746579405851662, + 0.0009481976157985628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.03368772566318512 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up centers on a half-eaten sandwich, placed on a rustic off-white ceramic plate, resting on the warm, grainy surface of a wooden cafe table. Soft golden light from a nearby window illuminates the scene, casting gentle shadows. A clean bite into the sandwich reveals its layers in crisp detail: toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey. A bokeh of warm fairy lights in the softly blurred background hints at the café's cozy atmosphere. Suddenly, unsettlingly, the sandwich's structure begins to fail; the hard bread begins to sag and soften, and the solid filling loses its shape and begins to melt downward. As the disparate elements dissolve into one another, the process accelerates, transforming into a gooey, swirling puddle of color on the plate. The green of the lettuce, the red of the tomato, and the pale tones of the turkey and bread bleed and marble together, forming a mesmerizing, kaleidoscopic liquid that slowly swirls, its surface catching the light with an iridescent, oily sheen, creating a surreal and dreamlike atmosphere that contrasts sharply with the initially cozy setting.", + "seed": 24572, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_8/index_9171_seed_24572_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.028832093346863985, + "vlm_logprob": -0.028832093346863985, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.028832093346863985, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008467486128211021 + ], + "opposite_logprob_tokens": [ + -5.758467674255371 + ], + "logprob_reward": -0.008467486128211021, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.008467486128211021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15750540792942047, + 0.04702870175242424, + 0.04199301451444626, + 0.039202503859996796, + 0.019261276349425316, + 0.16528919339179993, + 0.10917097330093384, + 0.1914522647857666, + 0.06646735966205597, + 0.16262924671173096 + ], + "credit_map": [ + 0.028520435094833374, + 0.01421777531504631, + 0.033591143786907196, + 0.018000280484557152, + 0.01269288919866085, + 0.015195242129266262, + 0.01854143850505352, + 0.016746215522289276, + 0.006516042165458202, + 0.005640705581754446, + 0.009669087827205658, + 0.0042299749329686165, + 0.0038151859771460295, + 0.0059936814941465855, + 0.004748146049678326, + 0.00641588494181633, + 0.005735754035413265, + 0.0039479173719882965, + 0.008262965828180313, + 0.003885362297296524, + 0.003479973180219531, + 0.005064023192971945, + 0.003743889043107629, + 0.00787313561886549, + 0.00520078931003809, + 0.0032996460795402527, + 0.00842303317040205, + 0.0032796314917504787, + 0.002659331541508436, + 0.0037891126703470945, + 0.004038400948047638, + 0.008512563072144985, + 0.0030247343238443136, + 0.0017914111958816648, + 0.0040262131951749325, + 0.0018543278565630317, + 0.0012714258627966046, + 0.002195571083575487, + 0.002491628285497427, + 0.0026059672236442566, + 0.01789598912000656, + 0.011601954698562622, + 0.01302402000874281, + 0.0326961912214756, + 0.012662948109209538, + 0.014563228003680706, + 0.046176113188266754, + 0.01666877605021, + 0.025812609121203423, + 0.008269247598946095, + 0.014362107962369919, + 0.008746206760406494, + 0.005555319134145975, + 0.012258561328053474, + 0.00895142462104559, + 0.025215502828359604, + 0.027904683724045753, + 0.023679474368691444, + 0.02460847608745098, + 0.03050336055457592, + 0.013153956271708012, + 0.02122090943157673, + 0.03374672681093216, + 0.01663469709455967, + 0.010094720870256424, + 0.013736054301261902, + 0.00714509841054678, + 0.006725194398313761, + 0.006131995003670454, + 0.004926892463117838, + 0.010240831412374973, + 0.00746657932177186, + 0.011858037672936916, + 0.01620285026729107, + 0.021588489413261414, + 0.030979204922914505, + 0.01398558635264635, + 0.013212045654654503, + 0.02492467872798443, + 0.029878366738557816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.008467486128211021 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006782249081879854 + ], + "opposite_logprob_tokens": [ + -5.006782054901123 + ], + "logprob_reward": -0.006782249081879854, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006782249081879854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17116209864616394, + 0.10220880806446075, + 0.05976228043437004, + 0.08600468188524246, + 0.03900004178285599, + 0.12367729842662811, + 0.054541923105716705, + 0.1514064371585846, + 0.12357612699270248, + 0.08866026997566223 + ], + "credit_map": [ + 0.03287990018725395, + 0.015258210711181164, + 0.03241415321826935, + 0.020650384947657585, + 0.010922645218670368, + 0.014951792545616627, + 0.019217513501644135, + 0.024867506697773933, + 0.0233388002961874, + 0.011572163552045822, + 0.021558852866292, + 0.00799174141138792, + 0.006137060932815075, + 0.010740600526332855, + 0.007280273828655481, + 0.013589316979050636, + 0.010636020451784134, + 0.004728508647531271, + 0.008197883144021034, + 0.005988788791000843, + 0.004672554787248373, + 0.006893690209835768, + 0.006563141942024231, + 0.01208169013261795, + 0.01453476119786501, + 0.008080223575234413, + 0.01482883095741272, + 0.009962485171854496, + 0.005110324826091528, + 0.008492534048855305, + 0.008524824865162373, + 0.01647070236504078, + 0.006025557406246662, + 0.00296649313531816, + 0.006165239494293928, + 0.0059041730128228664, + 0.002670004963874817, + 0.0031757275573909283, + 0.006988111417740583, + 0.005104732234030962, + 0.020471712574362755, + 0.008410430513322353, + 0.014340524561703205, + 0.018120158463716507, + 0.008526064455509186, + 0.016196921467781067, + 0.02464902773499489, + 0.012962457723915577, + 0.010544324293732643, + 0.00399289233610034, + 0.006644992157816887, + 0.005045432131737471, + 0.0035629053600132465, + 0.006641115061938763, + 0.005596124567091465, + 0.012514137662947178, + 0.018520772457122803, + 0.01942303404211998, + 0.017017878592014313, + 0.02443397417664528, + 0.012250219471752644, + 0.013856249861419201, + 0.030835259705781937, + 0.015069049783051014, + 0.016118666157126427, + 0.020505664870142937, + 0.016022587195038795, + 0.015343020670115948, + 0.012411066330969334, + 0.012351412326097488, + 0.02016795054078102, + 0.010655767284333706, + 0.009548411704599857, + 0.009475656785070896, + 0.009728336706757545, + 0.013244704343378544, + 0.012645293027162552, + 0.00801005307585001, + 0.013238468207418919, + 0.012769347056746483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.006782249081879854 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035917162895203 + ], + "opposite_logprob_tokens": [ + -2.3503592014312744 + ], + "logprob_reward": -0.10035917162895203, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035917162895203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07186457514762878, + 0.09234501421451569, + 0.0389779731631279, + 0.04675000533461571, + 0.07351585477590561, + 0.18996639549732208, + 0.08160033822059631, + 0.17889642715454102, + 0.11263801902532578, + 0.11344540864229202 + ], + "credit_map": [ + 0.02114396169781685, + 0.004336742218583822, + 0.021979782730340958, + 0.004711227957159281, + 0.0036099685821682215, + 0.007591832894831896, + 0.004398237448185682, + 0.004092824645340443, + 0.026528257876634598, + 0.005098631605505943, + 0.02985480986535549, + 0.007354449480772018, + 0.0044227587059140205, + 0.008336339145898819, + 0.005337035283446312, + 0.005412729922682047, + 0.005451252683997154, + 0.0034535282757133245, + 0.00796355027705431, + 0.005111817270517349, + 0.003400139743462205, + 0.005243521183729172, + 0.004006972536444664, + 0.004347189329564571, + 0.007670288905501366, + 0.004434030968695879, + 0.0061331382021307945, + 0.0063719539903104305, + 0.004502982832491398, + 0.005179189145565033, + 0.006053089629858732, + 0.006405333522707224, + 0.01188918948173523, + 0.006708707194775343, + 0.011135082691907883, + 0.0099129443988204, + 0.006329003255814314, + 0.005673776846379042, + 0.012427165172994137, + 0.009439981542527676, + 0.023344263434410095, + 0.01567595824599266, + 0.015444569289684296, + 0.032796356827020645, + 0.016054607927799225, + 0.01885533146560192, + 0.04643486812710762, + 0.02136044204235077, + 0.015289315953850746, + 0.006604095455259085, + 0.008562575094401836, + 0.013795578852295876, + 0.0081302085891366, + 0.007969881407916546, + 0.01238824799656868, + 0.008860436268150806, + 0.02319275215268135, + 0.0265967920422554, + 0.02128174901008606, + 0.029406515881419182, + 0.013740046881139278, + 0.01699536293745041, + 0.032369665801525116, + 0.015313544310629368, + 0.0179942324757576, + 0.013860027305781841, + 0.015015443786978722, + 0.019887933507561684, + 0.008525482378900051, + 0.011042573489248753, + 0.017395421862602234, + 0.00891690980643034, + 0.01394589338451624, + 0.009989438578486443, + 0.021509388461709023, + 0.01515292376279831, + 0.015873951837420464, + 0.008129004389047623, + 0.012721733190119267, + 0.016123076900839806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.10035917162895203 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052617397159338 + ], + "opposite_logprob_tokens": [ + -5.2552618980407715 + ], + "logprob_reward": -0.0052617397159338, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052617397159338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07796692848205566, + 0.08006541430950165, + 0.02810252457857132, + 0.09334386885166168, + 0.04664875566959381, + 0.13325512409210205, + 0.05946026369929314, + 0.13439738750457764, + 0.1004328578710556, + 0.24632689356803894 + ], + "credit_map": [ + 0.014985696412622929, + 0.008286889642477036, + 0.01969265192747116, + 0.01033027470111847, + 0.0051932986825704575, + 0.007759090978652239, + 0.00604949239641428, + 0.0056695351377129555, + 0.012210455723106861, + 0.005770259071141481, + 0.017987819388508797, + 0.006910378113389015, + 0.005351281259208918, + 0.00810522772371769, + 0.0074632135219872, + 0.016266779974102974, + 0.005776719190180302, + 0.003002545330673456, + 0.004731521010398865, + 0.002553676487877965, + 0.002264635171741247, + 0.003061296185478568, + 0.0027066990733146667, + 0.004005433060228825, + 0.016584673896431923, + 0.009118366055190563, + 0.019510721787810326, + 0.008591135032474995, + 0.005951572209596634, + 0.009966044686734676, + 0.0076921372674405575, + 0.015929216518998146, + 0.0066958530806005, + 0.0050809383392333984, + 0.005575471092015505, + 0.005347693804651499, + 0.004054962191730738, + 0.004846499301493168, + 0.006567604374140501, + 0.00847973395138979, + 0.017265070229768753, + 0.010575618594884872, + 0.011870978400111198, + 0.022325659170746803, + 0.01161907147616148, + 0.014811118133366108, + 0.03093840554356575, + 0.01384920347481966, + 0.009465087205171585, + 0.004794045351445675, + 0.006331924349069595, + 0.009362767450511456, + 0.006174085661768913, + 0.006607560906559229, + 0.009399072267115116, + 0.007325716782361269, + 0.01726856268942356, + 0.02080860175192356, + 0.016901755705475807, + 0.023309286683797836, + 0.009347539395093918, + 0.013335377909243107, + 0.02203468233346939, + 0.011391575448215008, + 0.013124782592058182, + 0.015201706439256668, + 0.011698982678353786, + 0.014698532409965992, + 0.010525531135499477, + 0.010310529731214046, + 0.015455877408385277, + 0.009416919201612473, + 0.02217857353389263, + 0.024648824706673622, + 0.04012494906783104, + 0.04591915383934975, + 0.031957343220710754, + 0.020316829904913902, + 0.03390569984912872, + 0.027275502681732178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.0052617397159338 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a half-eaten sandwich on a rustic off-white ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sandwich clearly show layers of toasted sourdough bread, vibrant lettuce greens, juicy red tomato slices, and thick-cut roast turkey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sandwich's structure begin to fail with the bread sagging and the filling melting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ingredients dissolve into a gooey, swirling puddle where green, red, and pale tones bleed and marble together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final liquid surface exhibit an iridescent, oily sheen that catches the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289820179343224 + ], + "opposite_logprob_tokens": [ + -3.773289918899536 + ], + "logprob_reward": -0.023289820179343224, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289820179343224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1311865746974945, + 0.1793508678674698, + 0.10070934146642685, + 0.07152445614337921, + 0.03374645486474037, + 0.05421808734536171, + 0.04464288428425789, + 0.05156073719263077, + 0.06949714571237564, + 0.26356351375579834 + ], + "credit_map": [ + 0.03226744383573532, + 0.009942669421434402, + 0.02918492630124092, + 0.015216702595353127, + 0.010682156309485435, + 0.01249697059392929, + 0.011196929961442947, + 0.010198771022260189, + 0.04169049486517906, + 0.00927267037332058, + 0.03877450153231621, + 0.015314425341784954, + 0.011786025948822498, + 0.015532719902694225, + 0.017577748745679855, + 0.02940228395164013, + 0.01747141219675541, + 0.006345523986965418, + 0.020046591758728027, + 0.01100479532033205, + 0.009171457961201668, + 0.012172876857221127, + 0.012745894491672516, + 0.011750791221857071, + 0.012750282883644104, + 0.007553539704531431, + 0.015951354056596756, + 0.007059832569211721, + 0.005525222513824701, + 0.0075486162677407265, + 0.005335777532309294, + 0.009799831546843052, + 0.004721817560493946, + 0.004307388793677092, + 0.005149322561919689, + 0.003753877244889736, + 0.0033594744745641947, + 0.0032379606273025274, + 0.004618603736162186, + 0.004598012659698725, + 0.006441758945584297, + 0.004382234066724777, + 0.006586153991520405, + 0.00947653315961361, + 0.004104329273104668, + 0.004899189341813326, + 0.012349219061434269, + 0.005978670436888933, + 0.008071486838161945, + 0.002856732113286853, + 0.00611317902803421, + 0.002561915200203657, + 0.0028015871066600084, + 0.005973610561341047, + 0.003973830491304398, + 0.01229054480791092, + 0.00681022135540843, + 0.006498642731457949, + 0.004849663004279137, + 0.009563717991113663, + 0.0034010407980531454, + 0.004107998218387365, + 0.010563088580965996, + 0.005766363348811865, + 0.007984133437275887, + 0.009889543056488037, + 0.01098797656595707, + 0.008813677355647087, + 0.006251787766814232, + 0.006341488566249609, + 0.012153416872024536, + 0.007075123954564333, + 0.029448945075273514, + 0.030690010637044907, + 0.040857478976249695, + 0.046264298260211945, + 0.019359398633241653, + 0.02198989875614643, + 0.02972610667347908, + 0.04522735998034477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.023289820179343224 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_9/index_tensor([6630, 6630])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_9/index_tensor([6630, 6630])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2b7ae120945f4e6ec2c7211f7a9fb2241090c432 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_9/index_tensor([6630, 6630])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful close-up, shot handheld and from a low angle, almost level with the water's surface, the camera shakes violently with the intensity of the simulated scene, tracking the roaring torrent with urgency. Under a gloomy, oppressive sky, the murky, yellow-brown floodwaters, laden with mud, sand, broken branches, and surging white foam, race through the narrow channel at devastating speed. In the foreground, clusters of emerald green water chestnut leaves, previously serenely floating on the surface, their diamond-shaped, vibrant leaves a stark contrast to the impending catastrophe. Almost instantly, the raging torrent reaches these tranquil waters, unfazed, savagely devouring, engulfing, and tearing apart the leaves. The emerald green vitality, vulnerable to the turbid, yellow-brown force, quickly vanishes beneath the murky water, leaving behind only a few struggling whirlpools and scattered fragments. The continuous shaking of the camera and the rapid tracking movement perfectly create an immersive sense of chaos and the fear of being powerless in the face of the power of nature. The whole scene is full of primitive, tense and irresistible sense of disaster record.", + "seed": 31907, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_9/index_6630_seed_31907_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1537164645153098, + "vlm_logprob": -0.1537164645153098, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.1537164645153098, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0717146024107933 + ], + "opposite_logprob_tokens": [ + -2.8217146396636963 + ], + "logprob_reward": -0.0717146024107933, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0717146024107933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17382268607616425, + 0.10722598433494568, + 0.08637350797653198, + 0.13362480700016022, + 0.05016973987221718, + 0.04712160304188728, + 0.060919344425201416, + 0.19693885743618011, + 0.11689364165067673, + 0.026909803971648216 + ], + "credit_map": [ + 0.014711963012814522, + 0.019833426922559738, + 0.03362787887454033, + 0.017595481127500534, + 0.02087552845478058, + 0.011260831728577614, + 0.025812869891524315, + 0.030104707926511765, + 0.016721021384000778, + 0.011394300498068333, + 0.013219884596765041, + 0.021601572632789612, + 0.009821085259318352, + 0.009859297424554825, + 0.01308526936918497, + 0.011523554101586342, + 0.01726822555065155, + 0.008092335425317287, + 0.012756425887346268, + 0.015454450622200966, + 0.005571783985942602, + 0.006191645283252001, + 0.011247777380049229, + 0.009790856391191483, + 0.011942686513066292, + 0.00702438922598958, + 0.026742223650217056, + 0.020856650546193123, + 0.008103725500404835, + 0.00997406430542469, + 0.017609983682632446, + 0.031371086835861206, + 0.005266851279884577, + 0.0036257861647754908, + 0.007323224097490311, + 0.010625194758176804, + 0.0031517858151346445, + 0.0035596706438809633, + 0.008959094062447548, + 0.0076581318862736225, + 0.00493123522028327, + 0.0050862133502960205, + 0.007527078501880169, + 0.006124918814748526, + 0.003895455738529563, + 0.004420584067702293, + 0.006993259768933058, + 0.008142856881022453, + 0.0039590816013514996, + 0.005743511952459812, + 0.007550929207354784, + 0.00914018414914608, + 0.005482630338519812, + 0.010414252057671547, + 0.010519951581954956, + 0.00810881145298481, + 0.010551679879426956, + 0.01271016988903284, + 0.05515563488006592, + 0.06372924149036407, + 0.011796790175139904, + 0.01664622873067856, + 0.014106069691479206, + 0.012243042699992657, + 0.011655702255666256, + 0.011559315025806427, + 0.008943813852965832, + 0.014790263026952744, + 0.006809175480157137, + 0.014607349410653114, + 0.03214319422841072, + 0.016384828835725784, + 0.00390887213870883, + 0.0028576613403856754, + 0.003260370809584856, + 0.002806267933920026, + 0.0024627968668937683, + 0.0026941958349198103, + 0.004089920315891504, + 0.004829718731343746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.0717146024107933 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933772563934326 + ], + "opposite_logprob_tokens": [ + -0.6933772563934326 + ], + "logprob_reward": -0.6933772563934326, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933772563934326, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12029363214969635, + 0.04581926017999649, + 0.04583961144089699, + 0.04911648854613304, + 0.125847727060318, + 0.26605701446533203, + 0.07100150734186172, + 0.13533969223499298, + 0.08103285729885101, + 0.05965224653482437 + ], + "credit_map": [ + 0.013467400334775448, + 0.01658509485423565, + 0.014387042261660099, + 0.01727970503270626, + 0.01391808781772852, + 0.013969305902719498, + 0.019656866788864136, + 0.011030112393200397, + 0.007222286891192198, + 0.004523457959294319, + 0.004958561155945063, + 0.008077320642769337, + 0.003328269813209772, + 0.005764600355178118, + 0.005696361418813467, + 0.0062483977526426315, + 0.006214425433427095, + 0.004636288620531559, + 0.0056118411011993885, + 0.006374100223183632, + 0.004605776164680719, + 0.005396277643740177, + 0.007890763692557812, + 0.005110137164592743, + 0.005272317212074995, + 0.003923455718904734, + 0.009385536424815655, + 0.008815149776637554, + 0.003491688519716263, + 0.005644625518471003, + 0.0052183945663273335, + 0.007365318015217781, + 0.007529818452894688, + 0.010515699163079262, + 0.02295922487974167, + 0.028234148398041725, + 0.008851117454469204, + 0.00706714391708374, + 0.024643337354063988, + 0.016047218814492226, + 0.01893835887312889, + 0.020894333720207214, + 0.07181690633296967, + 0.047947078943252563, + 0.02398594282567501, + 0.017840955406427383, + 0.030892858281731606, + 0.0337405763566494, + 0.006490460131317377, + 0.006519075483083725, + 0.01157219149172306, + 0.012650614604353905, + 0.008361445739865303, + 0.007719275541603565, + 0.007800299674272537, + 0.009888135828077793, + 0.007677455898374319, + 0.00807472039014101, + 0.031048860400915146, + 0.035020411014556885, + 0.008684404194355011, + 0.014350241050124168, + 0.016863394528627396, + 0.013620191253721714, + 0.007572581525892019, + 0.006921437103301287, + 0.00923188030719757, + 0.011570844799280167, + 0.003969460725784302, + 0.009268049150705338, + 0.0218012984842062, + 0.010697300545871258, + 0.0053414939902722836, + 0.005248131696134806, + 0.010208295658230782, + 0.00795276090502739, + 0.005433718673884869, + 0.004909702576696873, + 0.00890903826802969, + 0.011649097315967083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.6933772563934326 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015168600948527455 + ], + "opposite_logprob_tokens": [ + -6.501516819000244 + ], + "logprob_reward": -0.0015168600948527455, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015168600948527455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15116052329540253, + 0.04554423689842224, + 0.06762469559907913, + 0.1334567368030548, + 0.09322219341993332, + 0.08928760886192322, + 0.18756242096424103, + 0.1060526967048645, + 0.04016543924808502, + 0.08592350035905838 + ], + "credit_map": [ + 0.014166959561407566, + 0.020014464855194092, + 0.017070956528186798, + 0.021169573068618774, + 0.027381202206015587, + 0.013759588822722435, + 0.025313353165984154, + 0.012284420430660248, + 0.006955746561288834, + 0.00474116625264287, + 0.005413114558905363, + 0.007930953055620193, + 0.004978927783668041, + 0.004588955547660589, + 0.006049414165318012, + 0.004885959438979626, + 0.005943078547716141, + 0.007740651722997427, + 0.011634022928774357, + 0.017863808199763298, + 0.004438706208020449, + 0.006123456172645092, + 0.008170868270099163, + 0.005710104946047068, + 0.010684572160243988, + 0.008414644747972488, + 0.03255273029208183, + 0.024960435926914215, + 0.009701011702418327, + 0.00943606998771429, + 0.017487658187747, + 0.02021961100399494, + 0.009297242388129234, + 0.009325392544269562, + 0.01316965650767088, + 0.01631157658994198, + 0.005402284674346447, + 0.010403464548289776, + 0.014674974605441093, + 0.014637601561844349, + 0.01327031385153532, + 0.011340648867189884, + 0.011656476184725761, + 0.012558870017528534, + 0.010309026576578617, + 0.007700697053223848, + 0.01016419380903244, + 0.01228738110512495, + 0.018123604357242584, + 0.023704692721366882, + 0.02448039874434471, + 0.028462031856179237, + 0.022901982069015503, + 0.01509388443082571, + 0.027058565989136696, + 0.027737263590097427, + 0.009073251858353615, + 0.006385619752109051, + 0.015866152942180634, + 0.023486685007810593, + 0.008702963590621948, + 0.014507542364299297, + 0.017525795847177505, + 0.010504677891731262, + 0.00691354600712657, + 0.0035482142120599747, + 0.004358019679784775, + 0.005437164567410946, + 0.0033098424319177866, + 0.0054832762107253075, + 0.0071993134915828705, + 0.003916061483323574, + 0.009520313702523708, + 0.012410370633006096, + 0.013017588295042515, + 0.00898357667028904, + 0.008138240315020084, + 0.008244947530329227, + 0.011140147224068642, + 0.01446831040084362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.0015168600948527455 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013388964580371976 + ], + "opposite_logprob_tokens": [ + -6.626338958740234 + ], + "logprob_reward": -0.0013388964580371976, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013388964580371976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11874047666788101, + 0.08423121273517609, + 0.09350112080574036, + 0.05211381986737251, + 0.16941896080970764, + 0.1759473979473114, + 0.08532305806875229, + 0.10150217264890671, + 0.08943953365087509, + 0.029782172292470932 + ], + "credit_map": [ + 0.012017624452710152, + 0.013985512778162956, + 0.016737869009375572, + 0.014894230291247368, + 0.01574595831334591, + 0.012800643220543861, + 0.01748577505350113, + 0.015072877518832684, + 0.010904689319431782, + 0.009559519588947296, + 0.009105412289500237, + 0.01653638295829296, + 0.009198949672281742, + 0.009028569795191288, + 0.009674648754298687, + 0.01022304780781269, + 0.010307829827070236, + 0.006729927845299244, + 0.01253910269588232, + 0.015392993576824665, + 0.009716867469251156, + 0.006137308664619923, + 0.022697696462273598, + 0.009979412890970707, + 0.00600635539740324, + 0.004560019820928574, + 0.00978647917509079, + 0.007645439822226763, + 0.004569368902593851, + 0.005930487532168627, + 0.007086031138896942, + 0.006529643665999174, + 0.018326034769415855, + 0.04345523938536644, + 0.01591784507036209, + 0.023997895419597626, + 0.024863947182893753, + 0.009882573038339615, + 0.014854718931019306, + 0.01812073215842247, + 0.01624714955687523, + 0.013542535714805126, + 0.038195379078388214, + 0.028483929112553596, + 0.01684289239346981, + 0.016274254769086838, + 0.022846948355436325, + 0.023514321073889732, + 0.005840203259140253, + 0.009703264571726322, + 0.012906128540635109, + 0.012373996898531914, + 0.011787831783294678, + 0.010793508030474186, + 0.00983242318034172, + 0.01208570972084999, + 0.004356042016297579, + 0.006078785751014948, + 0.027943430468440056, + 0.03477555885910988, + 0.005714252591133118, + 0.006709783338010311, + 0.0068320115096867085, + 0.009092307649552822, + 0.009210996329784393, + 0.007906786166131496, + 0.007001334801316261, + 0.012530765496194363, + 0.008309684693813324, + 0.015320906415581703, + 0.019159095361828804, + 0.009999960660934448, + 0.003682920476421714, + 0.004075405187904835, + 0.004726646933704615, + 0.004517654422670603, + 0.0022921671625226736, + 0.0024376297369599342, + 0.003649819176644087, + 0.0043999310582876205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.0013388964580371976 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006347072194330394 + ], + "opposite_logprob_tokens": [ + -7.375634670257568 + ], + "logprob_reward": -0.0006347072194330394, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006347072194330394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3108852803707123, + 0.09125278145074844, + 0.06901329755783081, + 0.07426310330629349, + 0.08418167382478714, + 0.05506610497832298, + 0.0827597975730896, + 0.12577985227108002, + 0.07018687576055527, + 0.03661126643419266 + ], + "credit_map": [ + 0.026280801743268967, + 0.03731939196586609, + 0.05270060524344444, + 0.03869396820664406, + 0.043404147028923035, + 0.024705560877919197, + 0.04679587855935097, + 0.040984947234392166, + 0.011817844584584236, + 0.010491953231394291, + 0.012106174603104591, + 0.01906602829694748, + 0.007646680809557438, + 0.009062950499355793, + 0.011161751113831997, + 0.009899395518004894, + 0.009384438395500183, + 0.00463155098259449, + 0.012129382230341434, + 0.017644783481955528, + 0.0053830803371965885, + 0.004828996025025845, + 0.008977798745036125, + 0.006033267825841904, + 0.00796008761972189, + 0.006684230174869299, + 0.014032503589987755, + 0.013296353630721569, + 0.0047902436926960945, + 0.006811757106333971, + 0.00795827992260456, + 0.012729647569358349, + 0.006429567467421293, + 0.011627715080976486, + 0.010695436969399452, + 0.0168837308883667, + 0.007560820784419775, + 0.006279692519456148, + 0.013053946197032928, + 0.011650766246020794, + 0.006341462954878807, + 0.006893079727888107, + 0.009314537048339844, + 0.008539613336324692, + 0.004403987433761358, + 0.004270584788173437, + 0.007782232481986284, + 0.007520606275647879, + 0.006594737991690636, + 0.006798996590077877, + 0.010682608932256699, + 0.014183339662849903, + 0.007476185914129019, + 0.011197805404663086, + 0.013656917959451675, + 0.01216920930892229, + 0.004757529590278864, + 0.007379543501883745, + 0.033514462411403656, + 0.040180545300245285, + 0.006630737334489822, + 0.008092930540442467, + 0.01071833074092865, + 0.014505764469504356, + 0.008078965358436108, + 0.005753411445766687, + 0.010947098024189472, + 0.011170998215675354, + 0.0061843120492994785, + 0.009548463858664036, + 0.010593540966510773, + 0.007910086773335934, + 0.006664174143224955, + 0.005125431343913078, + 0.0034706841688603163, + 0.003374579595401883, + 0.003723865607753396, + 0.0034053761046379805, + 0.005246850196272135, + 0.005600304808467627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0006347072194330394 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful close-up, shot handheld and from a low angle, almost level with the water's surface, the camera shakes violently with the intensity of the simulated scene, tracking the roaring torrent with urgency. Under a gloomy, oppressive sky, the murky, yellow-brown floodwaters, laden with mud, sand, broken branches, and surging white foam, race through the narrow channel at devastating speed. In the foreground, clusters of emerald green water chestnut leaves, previously serenely floating on the surface, their diamond-shaped, vibrant leaves a stark contrast to the impending catastrophe. Almost instantly, the raging torrent reaches these tranquil waters, unfazed, savagely devouring, engulfing, and tearing apart the leaves. The emerald green vitality, vulnerable to the turbid, yellow-brown force, quickly vanishes beneath the murky water, leaving behind only a few struggling whirlpools and scattered fragments. The continuous shaking of the camera and the rapid tracking movement perfectly create an immersive sense of chaos and the fear of being powerless in the face of the power of nature. The whole scene is full of primitive, tense and irresistible sense of disaster record.", + "seed": 31908, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_9/index_6630_seed_31908_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.18233391614630817, + "vlm_logprob": -0.18233391614630817, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.18233391614630817, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08238581568002701 + ], + "opposite_logprob_tokens": [ + -2.707385778427124 + ], + "logprob_reward": -0.08238581568002701, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08238581568002701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12284854054450989, + 0.21099457144737244, + 0.0767228826880455, + 0.04116037115454674, + 0.02589867077767849, + 0.060399651527404785, + 0.03518562391400337, + 0.20232413709163666, + 0.1346161663532257, + 0.08984938263893127 + ], + "credit_map": [ + 0.0109009500592947, + 0.01306746806949377, + 0.011150816455483437, + 0.011579655110836029, + 0.017504887655377388, + 0.012561384588479996, + 0.018998341634869576, + 0.027085036039352417, + 0.00983980018645525, + 0.008845798671245575, + 0.01181551069021225, + 0.01402271632105112, + 0.030122218653559685, + 0.06137809902429581, + 0.019641084596514702, + 0.0553293414413929, + 0.007311261259019375, + 0.023750945925712585, + 0.005030624568462372, + 0.004008295014500618, + 0.005417319014668465, + 0.022378018125891685, + 0.0035048273857682943, + 0.005321594420820475, + 0.00783303938806057, + 0.0062111420556902885, + 0.004766183439642191, + 0.003779627149924636, + 0.004316598176956177, + 0.0056285373866558075, + 0.004750984720885754, + 0.0038742583710700274, + 0.0016899743350222707, + 0.0034198968205600977, + 0.002263800473883748, + 0.0026217035483568907, + 0.005136747844517231, + 0.003791326889768243, + 0.002029081806540489, + 0.004946139641106129, + 0.006362892687320709, + 0.007632533088326454, + 0.006981722544878721, + 0.009397712536156178, + 0.006018467713147402, + 0.008145119063556194, + 0.009219491854310036, + 0.00664171390235424, + 0.0041606794111430645, + 0.004701017402112484, + 0.005047446116805077, + 0.004688671790063381, + 0.003185860114172101, + 0.004924111068248749, + 0.003138192929327488, + 0.0053396462462842464, + 0.018804216757416725, + 0.021265909075737, + 0.029097849503159523, + 0.03177792206406593, + 0.02172253094613552, + 0.02654213458299637, + 0.03637053444981575, + 0.016743043437600136, + 0.01180044375360012, + 0.012770022265613079, + 0.008847442455589771, + 0.011812585406005383, + 0.014498382806777954, + 0.023709725588560104, + 0.024043209850788116, + 0.02713436260819435, + 0.011512878350913525, + 0.009230422787368298, + 0.009010892361402512, + 0.010018031112849712, + 0.015982799232006073, + 0.013054571114480495, + 0.011239529587328434, + 0.009800254367291927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.08238581568002701 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261597156524658 + ], + "opposite_logprob_tokens": [ + -0.5761597156524658 + ], + "logprob_reward": -0.8261597156524658, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261597156524658, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10324221104383469, + 0.10118956863880157, + 0.044027213007211685, + 0.030613787472248077, + 0.02726568467915058, + 0.04290557652711868, + 0.10908203572034836, + 0.28264304995536804, + 0.17930705845355988, + 0.07972380518913269 + ], + "credit_map": [ + 0.009486396797001362, + 0.013589480891823769, + 0.014462738297879696, + 0.017324261367321014, + 0.015780558809638023, + 0.007954946719110012, + 0.013701001182198524, + 0.010942834429442883, + 0.004958913195878267, + 0.006495583802461624, + 0.00529534462839365, + 0.008791450411081314, + 0.014126813970506191, + 0.03022284246981144, + 0.00896499864757061, + 0.022333620116114616, + 0.0032673862297087908, + 0.009990130551159382, + 0.0033988195937126875, + 0.004503712523728609, + 0.009545297361910343, + 0.004997554235160351, + 0.003530343761667609, + 0.004793970845639706, + 0.005137915723025799, + 0.004857374355196953, + 0.0038996331859380007, + 0.0019362313905730844, + 0.005465763155370951, + 0.003897564485669136, + 0.0026332936249673367, + 0.002786015160381794, + 0.0024291151203215122, + 0.004928810987621546, + 0.001934591680765152, + 0.004285082221031189, + 0.0032592478673905134, + 0.0031272880733013153, + 0.0022822129540145397, + 0.005019336938858032, + 0.007540552411228418, + 0.0059820362366735935, + 0.005130230449140072, + 0.0044709425419569016, + 0.004745571408420801, + 0.005107222590595484, + 0.004920058883726597, + 0.005008961074054241, + 0.01153324544429779, + 0.015963329002261162, + 0.015484357252717018, + 0.014937725849449635, + 0.012501843273639679, + 0.01628723181784153, + 0.007369011174887419, + 0.01500528771430254, + 0.007238598540425301, + 0.017000779509544373, + 0.009614904411137104, + 0.008755004964768887, + 0.11745286732912064, + 0.0994018092751503, + 0.01291622407734394, + 0.010262872092425823, + 0.0164621751755476, + 0.018043775111436844, + 0.012103259563446045, + 0.01946219429373741, + 0.021798189729452133, + 0.03406454622745514, + 0.026934795081615448, + 0.030438125133514404, + 0.013664555735886097, + 0.012964282184839249, + 0.0073783970437943935, + 0.011106757447123528, + 0.006581947207450867, + 0.006574732717126608, + 0.011158897541463375, + 0.010294231586158276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.8261597156524658 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015184074873104692 + ], + "opposite_logprob_tokens": [ + -6.501518249511719 + ], + "logprob_reward": -0.0015184074873104692, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015184074873104692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20922569930553436, + 0.0937315970659256, + 0.07062352448701859, + 0.06821960210800171, + 0.07281874865293503, + 0.014968287199735641, + 0.06601681560277939, + 0.07830020785331726, + 0.2538698613643646, + 0.07222562283277512 + ], + "credit_map": [ + 0.016960520297288895, + 0.030584992840886116, + 0.018319174647331238, + 0.01818518340587616, + 0.0464225672185421, + 0.026104141026735306, + 0.024107273668050766, + 0.02854185923933983, + 0.008914620615541935, + 0.008341842330992222, + 0.009975375607609749, + 0.013704192824661732, + 0.011856154538691044, + 0.015566504560410976, + 0.013109803199768066, + 0.012263105250895023, + 0.0061373282223939896, + 0.022893639281392097, + 0.005576721858233213, + 0.0038855390157550573, + 0.010674698278307915, + 0.013460138812661171, + 0.00331482314504683, + 0.004680638201534748, + 0.005696696694940329, + 0.01099273283034563, + 0.006505015771836042, + 0.005275880452245474, + 0.013057801872491837, + 0.012957444414496422, + 0.005488981492817402, + 0.008245046250522137, + 0.004685088526457548, + 0.009935541078448296, + 0.006041575688868761, + 0.007209814619272947, + 0.016911786049604416, + 0.01253162045031786, + 0.005193868186324835, + 0.010309453122317791, + 0.002127488376572728, + 0.002002784749493003, + 0.0017970653716474771, + 0.00129553338047117, + 0.0023134544026106596, + 0.0020169406197965145, + 0.0017181970179080963, + 0.0016968237468972802, + 0.00663900887593627, + 0.00856391154229641, + 0.011279036290943623, + 0.010648579336702824, + 0.007704765070229769, + 0.007689767051488161, + 0.004879796877503395, + 0.008611954748630524, + 0.00829546432942152, + 0.012810689397156239, + 0.009688199497759342, + 0.012320679612457752, + 0.009221237152814865, + 0.006315058562904596, + 0.011933265253901482, + 0.007715614978224039, + 0.016926510259509087, + 0.02142694965004921, + 0.021258240565657616, + 0.03313169628381729, + 0.025871258229017258, + 0.04649640992283821, + 0.04355127364397049, + 0.04520753398537636, + 0.007179921492934227, + 0.010006752796471119, + 0.005547889042645693, + 0.007551243063062429, + 0.01550579909235239, + 0.009241855703294277, + 0.00872611254453659, + 0.008466050028800964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0015184074873104692 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010446811793372035 + ], + "opposite_logprob_tokens": [ + -6.876044750213623 + ], + "logprob_reward": -0.0010446811793372035, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010446811793372035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10397935658693314, + 0.45150521397590637, + 0.028724145144224167, + 0.03981674090027809, + 0.022866783663630486, + 0.06044701859354973, + 0.03480169177055359, + 0.10624996572732925, + 0.11564254760742188, + 0.03596656024456024 + ], + "credit_map": [ + 0.008503411896526814, + 0.014058892615139484, + 0.010217954404652119, + 0.015023636631667614, + 0.02160668931901455, + 0.010739877820014954, + 0.012587335892021656, + 0.011241557076573372, + 0.024115532636642456, + 0.016917860135436058, + 0.021543852984905243, + 0.022087695077061653, + 0.068948894739151, + 0.1372443288564682, + 0.0393378846347332, + 0.12130916118621826, + 0.0028343640733510256, + 0.005249228328466415, + 0.0032190752681344748, + 0.003218127181753516, + 0.0023069351445883512, + 0.004594744648784399, + 0.0034413812682032585, + 0.003860288765281439, + 0.007822049781680107, + 0.005885222461074591, + 0.0050567700527608395, + 0.0024021966382861137, + 0.0061624678783118725, + 0.004657851532101631, + 0.0038743624463677406, + 0.003955819644033909, + 0.0032540932297706604, + 0.002759075490757823, + 0.0018287466373294592, + 0.0031050320249050856, + 0.003775641554966569, + 0.003029698971658945, + 0.001983553171157837, + 0.0031309425830841064, + 0.009593036957085133, + 0.005182814318686724, + 0.006453482899814844, + 0.008619116619229317, + 0.005526846740394831, + 0.008479774929583073, + 0.007778169587254524, + 0.008813774213194847, + 0.004311813972890377, + 0.0045145717449486256, + 0.006901330780237913, + 0.005289299413561821, + 0.003002817276865244, + 0.00337495026178658, + 0.0027744099497795105, + 0.004632496740669012, + 0.005864682607352734, + 0.00890512764453888, + 0.0073326388373970985, + 0.009411226958036423, + 0.029308568686246872, + 0.029461173340678215, + 0.009606273844838142, + 0.006360273342579603, + 0.010477821342647076, + 0.011164195835590363, + 0.009144728071987629, + 0.013978998176753521, + 0.013863819651305676, + 0.02216068096458912, + 0.01653335802257061, + 0.01831895112991333, + 0.006207031197845936, + 0.00586398970335722, + 0.0030178213492035866, + 0.005217081401497126, + 0.0024705075193196535, + 0.003130757948383689, + 0.005755816586315632, + 0.004303553141653538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.0010446811793372035 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005609607324004173 + ], + "opposite_logprob_tokens": [ + -7.500560760498047 + ], + "logprob_reward": -0.0005609607324004173, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005609607324004173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1785862147808075, + 0.14798864722251892, + 0.03137318789958954, + 0.03790196776390076, + 0.05339896306395531, + 0.024771157652139664, + 0.08084014058113098, + 0.22877302765846252, + 0.15897856652736664, + 0.0573880635201931 + ], + "credit_map": [ + 0.017517702654004097, + 0.026582028716802597, + 0.01930222101509571, + 0.024842247366905212, + 0.036333076655864716, + 0.01724834367632866, + 0.02033122442662716, + 0.01642937771975994, + 0.008191365748643875, + 0.007990791462361813, + 0.008086046203970909, + 0.015176409855484962, + 0.020647883415222168, + 0.03621844947338104, + 0.01473954040557146, + 0.03693815693259239, + 0.0026702256873250008, + 0.004805870819836855, + 0.0035643773153424263, + 0.003280121134594083, + 0.006609814707189798, + 0.003836852265521884, + 0.003508510999381542, + 0.0030974121764302254, + 0.0034208118449896574, + 0.0053796544671058655, + 0.003916279878467321, + 0.0036808494478464127, + 0.007233686279505491, + 0.004811247810721397, + 0.004804540891200304, + 0.004654897376894951, + 0.0032498470973223448, + 0.009997131302952766, + 0.0024659018963575363, + 0.007401525508612394, + 0.007285690866410732, + 0.00745615316554904, + 0.004566759802401066, + 0.01097595039755106, + 0.0029842799995094538, + 0.0032249786891043186, + 0.002744668396189809, + 0.0023918920196592808, + 0.0034708110615611076, + 0.003787720575928688, + 0.0030003516003489494, + 0.00316645298153162, + 0.008841405622661114, + 0.010880415327847004, + 0.012348960153758526, + 0.014882944524288177, + 0.009325198829174042, + 0.006249518133699894, + 0.006959663704037666, + 0.011352034285664558, + 0.006561477202922106, + 0.016625558957457542, + 0.00967635028064251, + 0.009534255601465702, + 0.09120193123817444, + 0.07095301151275635, + 0.015274565666913986, + 0.008945871144533157, + 0.015270378440618515, + 0.014064673334360123, + 0.010821684263646603, + 0.014292844571173191, + 0.0182840283960104, + 0.030658092349767685, + 0.025777606293559074, + 0.029809264466166496, + 0.011575447395443916, + 0.0083335991948843, + 0.0048186639323830605, + 0.00751021783798933, + 0.004097516182810068, + 0.005155634600669146, + 0.008793138898909092, + 0.0071038450114429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.0005609607324004173 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful close-up, shot handheld and from a low angle, almost level with the water's surface, the camera shakes violently with the intensity of the simulated scene, tracking the roaring torrent with urgency. Under a gloomy, oppressive sky, the murky, yellow-brown floodwaters, laden with mud, sand, broken branches, and surging white foam, race through the narrow channel at devastating speed. In the foreground, clusters of emerald green water chestnut leaves, previously serenely floating on the surface, their diamond-shaped, vibrant leaves a stark contrast to the impending catastrophe. Almost instantly, the raging torrent reaches these tranquil waters, unfazed, savagely devouring, engulfing, and tearing apart the leaves. The emerald green vitality, vulnerable to the turbid, yellow-brown force, quickly vanishes beneath the murky water, leaving behind only a few struggling whirlpools and scattered fragments. The continuous shaking of the camera and the rapid tracking movement perfectly create an immersive sense of chaos and the fear of being powerless in the face of the power of nature. The whole scene is full of primitive, tense and irresistible sense of disaster record.", + "seed": 31909, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_9/index_6630_seed_31909_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.33932282784953716, + "vlm_logprob": -0.33932282784953716, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.33932282784953716, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35570719838142395 + ], + "opposite_logprob_tokens": [ + -1.2307071685791016 + ], + "logprob_reward": -0.35570719838142395, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35570719838142395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1008773148059845, + 0.23261553049087524, + 0.02716711349785328, + 0.04323739558458328, + 0.0785614550113678, + 0.03578300029039383, + 0.06665442883968353, + 0.09202297776937485, + 0.30291569232940674, + 0.02016507275402546 + ], + "credit_map": [ + 0.0064077009446918964, + 0.028117459267377853, + 0.008732066489756107, + 0.005175165366381407, + 0.009614207781851292, + 0.006106584332883358, + 0.005851580761373043, + 0.03087254799902439, + 0.016828685998916626, + 0.021055154502391815, + 0.06924106925725937, + 0.013897757977247238, + 0.018679045140743256, + 0.012339732609689236, + 0.01414161454886198, + 0.06643246859312057, + 0.0034456271678209305, + 0.003778811078518629, + 0.0032403829973191023, + 0.0029634488746523857, + 0.004171337466686964, + 0.00214482843875885, + 0.003655799198895693, + 0.0037668785080313683, + 0.0059896777383983135, + 0.007682844530791044, + 0.0054399496875703335, + 0.005676319822669029, + 0.0037744431756436825, + 0.003856471972540021, + 0.004590640310198069, + 0.006227051839232445, + 0.016580678522586823, + 0.012092959135770798, + 0.01674688793718815, + 0.0073590027168393135, + 0.004025159869343042, + 0.005245964042842388, + 0.006927432958036661, + 0.009583364240825176, + 0.007871557958424091, + 0.007263717707246542, + 0.004808556754142046, + 0.004069680813699961, + 0.004131610505282879, + 0.0026368338149040937, + 0.0028001449536532164, + 0.002200897317379713, + 0.009737370535731316, + 0.010920015163719654, + 0.0095929354429245, + 0.007728707045316696, + 0.006682218983769417, + 0.007189339026808739, + 0.007687900681048632, + 0.0071159470826387405, + 0.007364410441368818, + 0.01086126547306776, + 0.008904440328478813, + 0.00846276618540287, + 0.013913881033658981, + 0.008342858403921127, + 0.024610450491309166, + 0.009562908671796322, + 0.11965075880289078, + 0.07720983028411865, + 0.036496974527835846, + 0.010913937352597713, + 0.02074838988482952, + 0.01614919863641262, + 0.011214914731681347, + 0.010531711392104626, + 0.002543431706726551, + 0.0023269723169505596, + 0.002792706713080406, + 0.0020549700129777193, + 0.003884045174345374, + 0.0025170589797198772, + 0.0019074567826464772, + 0.0021384323481470346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.35570719838142395 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.31343412399292 + ], + "opposite_logprob_tokens": [ + -0.31343409419059753 + ], + "logprob_reward": -1.31343412399292, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.31343412399292, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17856380343437195, + 0.2207563817501068, + 0.05406726896762848, + 0.11733957380056381, + 0.07026685774326324, + 0.041189853101968765, + 0.1743542104959488, + 0.06662549823522568, + 0.050894420593976974, + 0.025942057371139526 + ], + "credit_map": [ + 0.015095447190105915, + 0.03948524221777916, + 0.013815789483487606, + 0.013100323267281055, + 0.039885398000478745, + 0.015773477032780647, + 0.011224893853068352, + 0.030183251947164536, + 0.017213791608810425, + 0.027718435972929, + 0.05754975229501724, + 0.014272262342274189, + 0.01839302107691765, + 0.015986686572432518, + 0.01749885082244873, + 0.05212360620498657, + 0.006888877600431442, + 0.01084750983864069, + 0.006642345804721117, + 0.0054549709893763065, + 0.006289258133620024, + 0.005317298229783773, + 0.007695848122239113, + 0.004931167233735323, + 0.014208790846168995, + 0.027099167928099632, + 0.017262153327465057, + 0.010291490703821182, + 0.010783200152218342, + 0.010251709260046482, + 0.009500800631940365, + 0.01794227585196495, + 0.014277758076786995, + 0.011926096864044666, + 0.011057970114052296, + 0.006795384921133518, + 0.004582623951137066, + 0.005671566352248192, + 0.006285257637500763, + 0.009670206345617771, + 0.007515319157391787, + 0.005330011248588562, + 0.007428833749145269, + 0.0036836115177720785, + 0.004288413096219301, + 0.002999083371832967, + 0.005538086406886578, + 0.004406499210745096, + 0.0217171311378479, + 0.025161784142255783, + 0.02977610193192959, + 0.017718268558382988, + 0.02125195972621441, + 0.020720621570944786, + 0.018160443753004074, + 0.01984790526330471, + 0.0077360691502690315, + 0.009720242582261562, + 0.009348961524665356, + 0.004985730163753033, + 0.011322722770273685, + 0.0064346506260335445, + 0.008516671136021614, + 0.008560453541576862, + 0.006423071958124638, + 0.005287650972604752, + 0.007411173079162836, + 0.005574995186179876, + 0.008942718617618084, + 0.00664762407541275, + 0.0052218507044017315, + 0.005385338328778744, + 0.007605177350342274, + 0.006233758293092251, + 0.002225506817921996, + 0.0016016964800655842, + 0.002702154917642474, + 0.0024883560836315155, + 0.001561085577122867, + 0.0015243228990584612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -1.31343412399292 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331276424229145 + ], + "opposite_logprob_tokens": [ + -3.77331280708313 + ], + "logprob_reward": -0.02331276424229145, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331276424229145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2312108278274536, + 0.06898375600576401, + 0.13917647302150726, + 0.07451019436120987, + 0.06022103875875473, + 0.02644519694149494, + 0.22535856068134308, + 0.05536121502518654, + 0.10017157346010208, + 0.018561162054538727 + ], + "credit_map": [ + 0.014189042150974274, + 0.07244496792554855, + 0.020345456898212433, + 0.010973473079502583, + 0.02745174989104271, + 0.01528205256909132, + 0.008167271502315998, + 0.06235681474208832, + 0.0049612331204116344, + 0.007704693358391523, + 0.017454175278544426, + 0.005562180653214455, + 0.008104429580271244, + 0.004485496319830418, + 0.005806377157568932, + 0.014905172400176525, + 0.015640554949641228, + 0.024174638092517853, + 0.01765516959130764, + 0.01609615981578827, + 0.013377027586102486, + 0.01161821000277996, + 0.019695904105901718, + 0.020918801426887512, + 0.0105977738276124, + 0.015069649554789066, + 0.010328438133001328, + 0.007090368308126926, + 0.00723250862210989, + 0.007433431223034859, + 0.005811834242194891, + 0.01094619370996952, + 0.011005905456840992, + 0.013417327776551247, + 0.013382304459810257, + 0.00422291737049818, + 0.004767447244375944, + 0.004309226293116808, + 0.0043106102384626865, + 0.004805301316082478, + 0.006081936415284872, + 0.005260549485683441, + 0.0035594699438661337, + 0.0019232651684433222, + 0.0033505428582429886, + 0.0023503329139202833, + 0.0012449671048671007, + 0.0026741325855255127, + 0.02725142613053322, + 0.03261387348175049, + 0.04544511064887047, + 0.020735686644911766, + 0.02957778051495552, + 0.03213387355208397, + 0.016439588740468025, + 0.021161213517189026, + 0.005762981250882149, + 0.0075738211162388325, + 0.010886527597904205, + 0.0055040717124938965, + 0.009233213029801846, + 0.005540664307773113, + 0.004026519600301981, + 0.006833413150161505, + 0.037220779806375504, + 0.021306894719600677, + 0.014139371924102306, + 0.00433146208524704, + 0.007939590141177177, + 0.006567423231899738, + 0.004620774649083614, + 0.004045281559228897, + 0.005200291983783245, + 0.004823246505111456, + 0.0015038533601909876, + 0.0010606641881167889, + 0.002003491623327136, + 0.0015855947276577353, + 0.0012129832757636905, + 0.0011710356920957565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.02331276424229145 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002822227543219924 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.002822227543219924, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002822227543219924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24766232073307037, + 0.11156928539276123, + 0.15029709041118622, + 0.09513585269451141, + 0.05081053078174591, + 0.03201479837298393, + 0.11109773069620132, + 0.034165333956480026, + 0.14502182602882385, + 0.022225238382816315 + ], + "credit_map": [ + 0.013956376351416111, + 0.07156941294670105, + 0.013715040870010853, + 0.009754241444170475, + 0.040148332715034485, + 0.017291443422436714, + 0.011658630333840847, + 0.06956885010004044, + 0.008127461187541485, + 0.008054213598370552, + 0.0343870110809803, + 0.008093219250440598, + 0.006875153630971909, + 0.006655079312622547, + 0.007129627745598555, + 0.032247528433799744, + 0.015992790460586548, + 0.02064705453813076, + 0.016725769266486168, + 0.01975567825138569, + 0.01803896576166153, + 0.012110649608075619, + 0.027145517989993095, + 0.019880658015608788, + 0.012891060672700405, + 0.012002565898001194, + 0.011370979249477386, + 0.012771103531122208, + 0.01050546858459711, + 0.008432187139987946, + 0.012953739613294601, + 0.014208750799298286, + 0.010267497971653938, + 0.008979106321930885, + 0.010652049444615841, + 0.004338493105024099, + 0.0032683745957911015, + 0.003906676545739174, + 0.00378673174418509, + 0.005611606407910585, + 0.003834341187030077, + 0.0036270306445658207, + 0.0051965257152915, + 0.004232470411807299, + 0.0029263326432555914, + 0.002769456012174487, + 0.006310597062110901, + 0.0031180432997643948, + 0.011479306034743786, + 0.012514797039330006, + 0.013956662267446518, + 0.014037946239113808, + 0.015413203276693821, + 0.009795227088034153, + 0.019800465553998947, + 0.014100120402872562, + 0.003750809235498309, + 0.004679099656641483, + 0.004780119750648737, + 0.0035521178506314754, + 0.0060006482526659966, + 0.003555602626875043, + 0.004484555218368769, + 0.0033623836934566498, + 0.05352077633142471, + 0.032366763800382614, + 0.018391961231827736, + 0.007058365270495415, + 0.011632418259978294, + 0.00956657063215971, + 0.006450099870562553, + 0.0060348715633153915, + 0.007104468531906605, + 0.006023122929036617, + 0.0018181252526119351, + 0.0013295658864080906, + 0.0017043761909008026, + 0.0012296605855226517, + 0.0016297181136906147, + 0.0013862010091543198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.002822227543219924 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013378250878304243 + ], + "opposite_logprob_tokens": [ + -6.626338005065918 + ], + "logprob_reward": -0.0013378250878304243, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013378250878304243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3388330340385437, + 0.08161059767007828, + 0.11431311070919037, + 0.053173456341028214, + 0.06051991879940033, + 0.027560781687498093, + 0.08496306091547012, + 0.0423009917140007, + 0.18488985300064087, + 0.011835174635052681 + ], + "credit_map": [ + 0.020304011180996895, + 0.11586611717939377, + 0.017058800905942917, + 0.013402964919805527, + 0.029623592272400856, + 0.01930537447333336, + 0.012665215879678726, + 0.11060699075460434, + 0.006382594350725412, + 0.010240982286632061, + 0.016720034182071686, + 0.008321401663124561, + 0.01122135017067194, + 0.009136554785072803, + 0.006470929831266403, + 0.013116751797497272, + 0.012489184737205505, + 0.017413314431905746, + 0.013491899706423283, + 0.01432854775339365, + 0.013343562372028828, + 0.009327100589871407, + 0.01862405613064766, + 0.015295454300940037, + 0.0069768899120390415, + 0.006942604202777147, + 0.007114978041499853, + 0.005799747537821531, + 0.005371213890612125, + 0.00656284810975194, + 0.00588723411783576, + 0.008517944253981113, + 0.013902496546506882, + 0.012000126764178276, + 0.011127147823572159, + 0.005147209390997887, + 0.005404530093073845, + 0.004334406461566687, + 0.004405447747558355, + 0.004198556765913963, + 0.002918847603723407, + 0.004255029838532209, + 0.004289234988391399, + 0.0027528139762580395, + 0.004180183634161949, + 0.002377329161390662, + 0.004265311639755964, + 0.002522032707929611, + 0.010767219588160515, + 0.011965562589466572, + 0.010398950427770615, + 0.01072871033102274, + 0.012562575750052929, + 0.009648657403886318, + 0.01192517951130867, + 0.006966210436075926, + 0.005021589808166027, + 0.006517090369015932, + 0.005807689856737852, + 0.0036235975567251444, + 0.0056947930715978146, + 0.0031818810384720564, + 0.008267765864729881, + 0.004186587408185005, + 0.07173477858304977, + 0.04375455528497696, + 0.023673759773373604, + 0.007454087026417255, + 0.014199272729456425, + 0.010270245373249054, + 0.006936086341738701, + 0.006867074407637119, + 0.001496348064392805, + 0.0018808669410645962, + 0.0015534735284745693, + 0.001036183093674481, + 0.002019908744841814, + 0.001597160124219954, + 0.0010977111523970962, + 0.0011535242665559053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.0013378250878304243 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful close-up, shot handheld and from a low angle, almost level with the water's surface, the camera shakes violently with the intensity of the simulated scene, tracking the roaring torrent with urgency. Under a gloomy, oppressive sky, the murky, yellow-brown floodwaters, laden with mud, sand, broken branches, and surging white foam, race through the narrow channel at devastating speed. In the foreground, clusters of emerald green water chestnut leaves, previously serenely floating on the surface, their diamond-shaped, vibrant leaves a stark contrast to the impending catastrophe. Almost instantly, the raging torrent reaches these tranquil waters, unfazed, savagely devouring, engulfing, and tearing apart the leaves. The emerald green vitality, vulnerable to the turbid, yellow-brown force, quickly vanishes beneath the murky water, leaving behind only a few struggling whirlpools and scattered fragments. The continuous shaking of the camera and the rapid tracking movement perfectly create an immersive sense of chaos and the fear of being powerless in the face of the power of nature. The whole scene is full of primitive, tense and irresistible sense of disaster record.", + "seed": 31910, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000005/group_9/index_6630_seed_31910_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11609173931647092, + "vlm_logprob": -0.11609173931647092, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.11609173931647092, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1391170471906662 + ], + "opposite_logprob_tokens": [ + -2.1391170024871826 + ], + "logprob_reward": -0.1391170471906662, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1391170471906662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18236428499221802, + 0.15410116314888, + 0.14530311524868011, + 0.052331842482089996, + 0.10300776362419128, + 0.04486430436372757, + 0.048489369451999664, + 0.08570145815610886, + 0.11475764960050583, + 0.06907904148101807 + ], + "credit_map": [ + 0.021000446751713753, + 0.014630833640694618, + 0.02867833524942398, + 0.02299048937857151, + 0.02090504579246044, + 0.01850748062133789, + 0.03542712330818176, + 0.02022453583776951, + 0.011053625494241714, + 0.009213675744831562, + 0.014664784073829651, + 0.010852622799575329, + 0.010409348644316196, + 0.029580725356936455, + 0.03083058074116707, + 0.03749581053853035, + 0.010117447935044765, + 0.044481515884399414, + 0.01838938519358635, + 0.012665703892707825, + 0.016859861090779305, + 0.015048702247440815, + 0.012553326785564423, + 0.015187188051640987, + 0.00863181333988905, + 0.005205551162362099, + 0.008163748309016228, + 0.006456769071519375, + 0.00609661964699626, + 0.006033797282725573, + 0.0063965232111513615, + 0.0053470260463654995, + 0.007515177596360445, + 0.03710091859102249, + 0.027754828333854675, + 0.0071907793171703815, + 0.010507920756936073, + 0.004034984391182661, + 0.002894621342420578, + 0.006008540745824575, + 0.00476829381659627, + 0.004410391673445702, + 0.005231185350567102, + 0.00794597715139389, + 0.005533111281692982, + 0.005712894722819328, + 0.00697407778352499, + 0.004288377705961466, + 0.0044979616068303585, + 0.003795529715716839, + 0.006861527916043997, + 0.00705602066591382, + 0.004963972140103579, + 0.006673993542790413, + 0.010253417305648327, + 0.0043869465589523315, + 0.005637792870402336, + 0.009798817336559296, + 0.017526855692267418, + 0.012964456342160702, + 0.01129399985074997, + 0.009112659841775894, + 0.012476884759962559, + 0.0068900021724402905, + 0.010950864292681217, + 0.01982463151216507, + 0.017095761373639107, + 0.014078276231884956, + 0.015106351114809513, + 0.011251973919570446, + 0.015233393758535385, + 0.01121640857309103, + 0.00485922209918499, + 0.005408263765275478, + 0.009064369834959507, + 0.010094569064676762, + 0.008806601166725159, + 0.009712428785860538, + 0.008790814317762852, + 0.012342780828475952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.1391170471906662 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.428915411233902 + ], + "opposite_logprob_tokens": [ + -1.0539153814315796 + ], + "logprob_reward": -0.428915411233902, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.428915411233902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060426242649555206, + 0.04338739812374115, + 0.16979481279850006, + 0.08222126960754395, + 0.03061957098543644, + 0.08312218636274338, + 0.053376954048871994, + 0.06883706897497177, + 0.35127705335617065, + 0.05693748965859413 + ], + "credit_map": [ + 0.008352616801857948, + 0.00474600400775671, + 0.007112482096999884, + 0.006138749420642853, + 0.005884900689125061, + 0.008589022792875767, + 0.00953009631484747, + 0.010072370059788227, + 0.00512425834313035, + 0.003461018670350313, + 0.006892917212098837, + 0.0036967876367270947, + 0.0035068602301180363, + 0.009894022718071938, + 0.004783895332366228, + 0.006027639843523502, + 0.011052286252379417, + 0.05368311330676079, + 0.02644059620797634, + 0.012354938313364983, + 0.016642659902572632, + 0.017322305589914322, + 0.01739584282040596, + 0.014903074130415916, + 0.015082350000739098, + 0.01128862239420414, + 0.01571853645145893, + 0.010336155071854591, + 0.008276106789708138, + 0.008465667255222797, + 0.007434090133756399, + 0.005619735922664404, + 0.0031724146101623774, + 0.005767603404819965, + 0.003959344234317541, + 0.0035814896691590548, + 0.003265518695116043, + 0.0037385073956102133, + 0.0036096132826060057, + 0.003525078296661377, + 0.010440120473504066, + 0.012518158182501793, + 0.01206508744508028, + 0.012027380988001823, + 0.009742870926856995, + 0.008815241977572441, + 0.008621587418019772, + 0.00889174360781908, + 0.004248236771672964, + 0.0039754570461809635, + 0.007661969866603613, + 0.007755928207188845, + 0.008045332506299019, + 0.005100830923765898, + 0.012222224846482277, + 0.004366968292742968, + 0.005049864761531353, + 0.008477982133626938, + 0.01002566795796156, + 0.010252837091684341, + 0.00709701469168067, + 0.009658007882535458, + 0.010423128493130207, + 0.00785256177186966, + 0.03425752744078636, + 0.07832469046115875, + 0.043586406856775284, + 0.043634675443172455, + 0.04301666095852852, + 0.03453758731484413, + 0.04609756916761398, + 0.02782195247709751, + 0.00517257209867239, + 0.0096108578145504, + 0.006105194799602032, + 0.006891867611557245, + 0.007345927879214287, + 0.005637536756694317, + 0.008312285877764225, + 0.007861245423555374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.428915411233902 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741991266608238 + ], + "opposite_logprob_tokens": [ + -5.006742000579834 + ], + "logprob_reward": -0.006741991266608238, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741991266608238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21541664004325867, + 0.10893462598323822, + 0.07864214479923248, + 0.04204942658543587, + 0.04955444112420082, + 0.07895217835903168, + 0.055224016308784485, + 0.12492205202579498, + 0.14803476631641388, + 0.09826966375112534 + ], + "credit_map": [ + 0.02433665655553341, + 0.017878040671348572, + 0.026386063545942307, + 0.021558431908488274, + 0.0218038409948349, + 0.022599153220653534, + 0.050959326326847076, + 0.029895160347223282, + 0.008782139047980309, + 0.00994256604462862, + 0.012086580507457256, + 0.010345910675823689, + 0.008893975988030434, + 0.021065644919872284, + 0.019374549388885498, + 0.018443269655108452, + 0.0037373965606093407, + 0.005313349887728691, + 0.003637877991423011, + 0.005201621446758509, + 0.014833059161901474, + 0.019703803583979607, + 0.014708300121128559, + 0.011506740003824234, + 0.008016611449420452, + 0.004632643889635801, + 0.005877128802239895, + 0.0044705686159431934, + 0.004123125225305557, + 0.00531498808413744, + 0.004942404106259346, + 0.004671961534768343, + 0.0036411022301763296, + 0.011955170892179012, + 0.013639030046761036, + 0.005828244611620903, + 0.004773876629769802, + 0.002733867382630706, + 0.003090218873694539, + 0.0038929374422878027, + 0.009655228815972805, + 0.009135075844824314, + 0.010687757283449173, + 0.008145512081682682, + 0.013658488169312477, + 0.008648409508168697, + 0.0126219242811203, + 0.006399788893759251, + 0.00370296323671937, + 0.006353912875056267, + 0.007013925816863775, + 0.008246797136962414, + 0.008678246289491653, + 0.00813326332718134, + 0.007248071487993002, + 0.005846840795129538, + 0.007079003378748894, + 0.011196338571608067, + 0.014524459838867188, + 0.02616320550441742, + 0.01346536260098219, + 0.03403004631400108, + 0.01019303873181343, + 0.00827061478048563, + 0.011745769530534744, + 0.0229310505092144, + 0.013793948106467724, + 0.02136414870619774, + 0.0326085165143013, + 0.01683441363275051, + 0.018110590055584908, + 0.010646343231201172, + 0.007890892215073109, + 0.012604809366166592, + 0.012119357474148273, + 0.009470725432038307, + 0.015600070357322693, + 0.012547132559120655, + 0.013691660016775131, + 0.014345027506351471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.006741991266608238 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004642066545784473 + ], + "opposite_logprob_tokens": [ + -5.379642009735107 + ], + "logprob_reward": -0.004642066545784473, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004642066545784473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16758692264556885, + 0.15793120861053467, + 0.2071177214384079, + 0.04310222342610359, + 0.0678272619843483, + 0.050702400505542755, + 0.06391184031963348, + 0.07384580373764038, + 0.12097130715847015, + 0.04700333997607231 + ], + "credit_map": [ + 0.01876121759414673, + 0.012547901831567287, + 0.02558731660246849, + 0.019712787121534348, + 0.023151449859142303, + 0.024991532787680626, + 0.025210246443748474, + 0.017624463886022568, + 0.0104157580062747, + 0.009270577691495419, + 0.024548392742872238, + 0.009129608049988747, + 0.015078087337315083, + 0.015021676197648048, + 0.028515150770545006, + 0.04595194011926651, + 0.018043480813503265, + 0.05039708688855171, + 0.021501552313566208, + 0.0146715696901083, + 0.024892309680581093, + 0.028868691995739937, + 0.020974716171622276, + 0.027768293395638466, + 0.003679638495668769, + 0.00418479461222887, + 0.00629093311727047, + 0.006077651400119066, + 0.006076973397284746, + 0.007474267855286598, + 0.004806779325008392, + 0.004511179868131876, + 0.008340415544807911, + 0.017667287960648537, + 0.015586097724735737, + 0.0051364959217607975, + 0.006391497794538736, + 0.004821683280169964, + 0.003649493446573615, + 0.006234284956008196, + 0.005416376516222954, + 0.005970842204988003, + 0.006536388769745827, + 0.007281076163053513, + 0.008520230650901794, + 0.005423287395387888, + 0.006481911055743694, + 0.005072284024208784, + 0.004536484368145466, + 0.006010336335748434, + 0.011011666618287563, + 0.009141641668975353, + 0.008245016448199749, + 0.006234660744667053, + 0.01282061543315649, + 0.0059114121831953526, + 0.004310414660722017, + 0.0075691756792366505, + 0.009035023860633373, + 0.010095309466123581, + 0.00877243373543024, + 0.0086539126932621, + 0.015289310365915298, + 0.010120217688381672, + 0.00841318815946579, + 0.023295920342206955, + 0.012391035445034504, + 0.017657261341810226, + 0.020159196108579636, + 0.013709601014852524, + 0.015580707229673862, + 0.009764391928911209, + 0.004090361297130585, + 0.007294835988432169, + 0.004789823666214943, + 0.00487879291176796, + 0.006639033555984497, + 0.005339245777577162, + 0.0060564144514501095, + 0.007914826273918152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.004642066545784473 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld from a low angle, almost level with the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera shake violently and track the water with rapid movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the floodwaters murky, yellow-brown, and laden with mud, sand, broken branches, and white foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green, diamond-shaped water chestnut leaves visible in the foreground before being destroyed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the torrent engulfing and tearing apart the leaves, leaving only whirlpools and fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010421803453937173 + ], + "opposite_logprob_tokens": [ + -6.876042366027832 + ], + "logprob_reward": -0.0010421803453937173, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010421803453937173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07084921002388, + 0.06365776062011719, + 0.18851043283939362, + 0.0662265419960022, + 0.08981718868017197, + 0.07447624951601028, + 0.07286826521158218, + 0.08101057261228561, + 0.24819479882717133, + 0.04438893869519234 + ], + "credit_map": [ + 0.007240769919008017, + 0.006158454809337854, + 0.012948859483003616, + 0.007981965318322182, + 0.010392858646810055, + 0.007599036209285259, + 0.011579552665352821, + 0.006947712041437626, + 0.006496580317616463, + 0.005892702378332615, + 0.0063069784082472324, + 0.005634567234665155, + 0.007168452721089125, + 0.015486147254705429, + 0.009485351853072643, + 0.007186990696936846, + 0.014949522912502289, + 0.053283631801605225, + 0.02287677675485611, + 0.014128709211945534, + 0.020759321749210358, + 0.015131233260035515, + 0.025192316621541977, + 0.022188927978277206, + 0.012968065217137337, + 0.00739675248041749, + 0.011080702766776085, + 0.007926012389361858, + 0.007159277331084013, + 0.007513116113841534, + 0.0060500651597976685, + 0.006132554728537798, + 0.006745707709342241, + 0.027147535234689713, + 0.02775372751057148, + 0.007819009944796562, + 0.006645734887570143, + 0.0037039639428257942, + 0.004127096850425005, + 0.0058744195848703384, + 0.008028862997889519, + 0.009850094094872475, + 0.00925181433558464, + 0.009701593779027462, + 0.011222034692764282, + 0.007870136760175228, + 0.009863371029496193, + 0.008688349276781082, + 0.007851102389395237, + 0.008009935729205608, + 0.008804027922451496, + 0.015362140722572803, + 0.007462304085493088, + 0.005060307681560516, + 0.013226067647337914, + 0.007092384155839682, + 0.004254150670021772, + 0.0064830719493329525, + 0.010080724023282528, + 0.010236811824142933, + 0.010514571331441402, + 0.012642198242247105, + 0.015004674904048443, + 0.011794373393058777, + 0.019460149109363556, + 0.056813716888427734, + 0.028772199526429176, + 0.031135844066739082, + 0.04084266722202301, + 0.0238677728921175, + 0.03295101970434189, + 0.014351433143019676, + 0.004278458189219236, + 0.004339088220149279, + 0.004618532955646515, + 0.005389522761106491, + 0.006343388464301825, + 0.0059964037500321865, + 0.005253736861050129, + 0.00816980842500925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.0010421803453937173 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_0/index_tensor([9364, 9364])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_0/index_tensor([9364, 9364])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a26598398d0a197d5e854801208c19df712d964e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_0/index_tensor([9364, 9364])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium, static shot, taken from a level angle, captures a woman with her long, dark brown hair tied into a neat ponytail. She wears an off-white silk blouse and high-waisted gray trousers. She leans forward slightly in her office chair, her body curved into an elegant arc. In her right hand, she clutches a ball of white A4 paper, wrinkled from the effort of her squeezing. She's in a modern office, where soft afternoon sunlight filters through the gaps in the blinds, casting dappled shadows on the floor and walls. In the corner of the frame, a cylindrical, brushed stainless steel trash can sits quietly, next to a vibrant plant. With focused concentration, she extends her arm, flicking her wrist gently, and tosses the ball of paper with a deft yet effortless gesture. The ball traces a gentle parabola through the air, finally landing with a precise, unerring \"thud\" into the trash can's opening. The whole scene presents a focused and peaceful daily office atmosphere, and the smoothness and precision of the movements reveal a hint of casual competence and calmness.", + "seed": 12316, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_0/index_9364_seed_12316_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04141123965382576, + "vlm_logprob": -0.04141123965382576, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.04141123965382576, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13741634786128998 + ], + "opposite_logprob_tokens": [ + -2.762416362762451 + ], + "logprob_reward": -0.13741634786128998, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.13741634786128998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39766815304756165, + 0.12007442116737366, + 0.058750517666339874, + 0.050717197358608246, + 0.08622264862060547, + 0.026649629697203636, + 0.0651625245809555, + 0.07831128686666489, + 0.06008812040090561, + 0.05635547637939453 + ], + "credit_map": [ + 0.031027374789118767, + 0.041531044989824295, + 0.0394781269133091, + 0.10733656585216522, + 0.022468505427241325, + 0.023705674335360527, + 0.05943502485752106, + 0.0726858526468277, + 0.017243800684809685, + 0.008783046156167984, + 0.012022781185805798, + 0.027336252853274345, + 0.009502705186605453, + 0.008896848186850548, + 0.022884123027324677, + 0.013404873199760914, + 0.010031444020569324, + 0.004052245523780584, + 0.009079144336283207, + 0.008130691014230251, + 0.006863098591566086, + 0.00660801911726594, + 0.009691117331385612, + 0.00429476099088788, + 0.006537330336868763, + 0.0033067117910832167, + 0.006544436328113079, + 0.007391897030174732, + 0.006966719403862953, + 0.007547164801508188, + 0.00802050530910492, + 0.004402437712997198, + 0.010204976424574852, + 0.005572652444243431, + 0.012574777938425541, + 0.011693711392581463, + 0.013533213175833225, + 0.0087257856503129, + 0.012817960232496262, + 0.011099588125944138, + 0.0034240391105413437, + 0.002005991991609335, + 0.0038936559576541185, + 0.003336310386657715, + 0.0047069936990737915, + 0.0022016470320522785, + 0.004437665455043316, + 0.002643329557031393, + 0.008141819387674332, + 0.002901372965425253, + 0.008891215547919273, + 0.008837646804749966, + 0.006864367984235287, + 0.007711561396718025, + 0.01520596630871296, + 0.006608576513826847, + 0.007785869762301445, + 0.006437663454562426, + 0.017102651298046112, + 0.009819303639233112, + 0.011835522949695587, + 0.006349405273795128, + 0.012226778082549572, + 0.006754096131771803, + 0.006636162754148245, + 0.003763745538890362, + 0.007114736828953028, + 0.008404169231653214, + 0.008639822714030743, + 0.003994838334619999, + 0.01156565174460411, + 0.009968999773263931, + 0.005835486110299826, + 0.007160637993365526, + 0.005022966302931309, + 0.008082784712314606, + 0.008936425670981407, + 0.005217721685767174, + 0.008853304199874401, + 0.007246153429150581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.13741634786128998 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03815042972564697 + ], + "opposite_logprob_tokens": [ + -3.2881503105163574 + ], + "logprob_reward": -0.03815042972564697, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03815042972564697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29740387201309204, + 0.13134002685546875, + 0.12987297773361206, + 0.0521225705742836, + 0.07907029986381531, + 0.0485873706638813, + 0.05070120468735695, + 0.10263457149267197, + 0.04230940341949463, + 0.0659576952457428 + ], + "credit_map": [ + 0.03472762554883957, + 0.017110278829932213, + 0.037389278411865234, + 0.10499178618192673, + 0.016228970140218735, + 0.017188090831041336, + 0.03414487466216087, + 0.03562299162149429, + 0.019682597368955612, + 0.009336922317743301, + 0.013104099780321121, + 0.038805264979600906, + 0.008496512658894062, + 0.009767245501279831, + 0.017667070031166077, + 0.014480317011475563, + 0.017772305756807327, + 0.011608307249844074, + 0.016195740550756454, + 0.014317533001303673, + 0.01109294593334198, + 0.012490863911807537, + 0.02350364439189434, + 0.022891642525792122, + 0.005956057459115982, + 0.003132142825052142, + 0.005464963614940643, + 0.0071712760254740715, + 0.009180394001305103, + 0.009562976658344269, + 0.006628736853599548, + 0.005026021972298622, + 0.010261937975883484, + 0.0049620396457612514, + 0.009479602798819542, + 0.010462433099746704, + 0.009552576579153538, + 0.01097989920526743, + 0.014484301209449768, + 0.008887508884072304, + 0.005940092261880636, + 0.0027440283447504044, + 0.007375477813184261, + 0.008265988901257515, + 0.0077990214340388775, + 0.00471006752923131, + 0.0072881304658949375, + 0.004464560188353062, + 0.00489346357062459, + 0.002164070727303624, + 0.006937493570148945, + 0.008255597203969955, + 0.003626532619819045, + 0.003713698126375675, + 0.014163765124976635, + 0.006946587935090065, + 0.010669062845408916, + 0.006755555514246225, + 0.013963681645691395, + 0.03061908483505249, + 0.009310704655945301, + 0.007210008334368467, + 0.01428949274122715, + 0.00981698278337717, + 0.0058475397527217865, + 0.0030954338144510984, + 0.0034407861530780792, + 0.004121973644942045, + 0.009150533936917782, + 0.0047185784205794334, + 0.007157611194998026, + 0.004776946268975735, + 0.006163147743791342, + 0.00806943979114294, + 0.005398466717451811, + 0.009936768561601639, + 0.01683823950588703, + 0.0049901679158210754, + 0.008444355800747871, + 0.0061171045526862144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03815042972564697 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644895628094673 + ], + "opposite_logprob_tokens": [ + -4.758645057678223 + ], + "logprob_reward": -0.008644895628094673, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644895628094673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2965919077396393, + 0.05884349346160889, + 0.07507222145795822, + 0.15929736196994781, + 0.08655279874801636, + 0.03451693430542946, + 0.05623219162225723, + 0.09632328152656555, + 0.06562721729278564, + 0.07094254344701767 + ], + "credit_map": [ + 0.022842837497591972, + 0.025206128135323524, + 0.03349744528532028, + 0.0938371866941452, + 0.01779758930206299, + 0.017182661220431328, + 0.03829791024327278, + 0.047930166125297546, + 0.006442877929657698, + 0.005438548978418112, + 0.007245458662509918, + 0.010011805221438408, + 0.005321664735674858, + 0.0050862254574894905, + 0.011385614052414894, + 0.007911299355328083, + 0.008659830316901207, + 0.007646345533430576, + 0.008517366833984852, + 0.008370775729417801, + 0.006329854018986225, + 0.0065895868465304375, + 0.01519909780472517, + 0.013759368099272251, + 0.01802699640393257, + 0.0072260405868291855, + 0.014255436137318611, + 0.013000890612602234, + 0.035159505903720856, + 0.03438682481646538, + 0.01866644248366356, + 0.018575221300125122, + 0.010038023814558983, + 0.004563051741570234, + 0.013183959759771824, + 0.010062934830784798, + 0.008846599608659744, + 0.011400088667869568, + 0.018339091911911964, + 0.010119050741195679, + 0.004029276315122843, + 0.0020397971384227276, + 0.00589488260447979, + 0.005680333357304335, + 0.003653655294328928, + 0.004089057445526123, + 0.005959124304354191, + 0.0031708073802292347, + 0.005756446160376072, + 0.0024323700927197933, + 0.005920756608247757, + 0.008797232061624527, + 0.006283624097704887, + 0.006769287399947643, + 0.014263036660850048, + 0.006009440869092941, + 0.009735963307321072, + 0.006389891263097525, + 0.015619004145264626, + 0.025255277752876282, + 0.009886730462312698, + 0.006680249702185392, + 0.014141235500574112, + 0.008614934049546719, + 0.00951311830431223, + 0.005320276133716106, + 0.01090833730995655, + 0.006613913457840681, + 0.008950789459049702, + 0.005433992948383093, + 0.011363927274942398, + 0.007522860076278448, + 0.006821894086897373, + 0.008857809007167816, + 0.005976574495434761, + 0.01085963100194931, + 0.010896099731326103, + 0.00480363704264164, + 0.012172729708254337, + 0.010554167442023754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.008644895628094673 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202939346432686 + ], + "opposite_logprob_tokens": [ + -4.264203071594238 + ], + "logprob_reward": -0.014202939346432686, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202939346432686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26741230487823486, + 0.16273203492164612, + 0.047159384936094284, + 0.07380979508161545, + 0.03954722359776497, + 0.028050830587744713, + 0.02551191858947277, + 0.10501527786254883, + 0.11015494912862778, + 0.14060623943805695 + ], + "credit_map": [ + 0.025556150823831558, + 0.016650492325425148, + 0.02933090552687645, + 0.10101915150880814, + 0.010442608967423439, + 0.014747998677194118, + 0.03511354327201843, + 0.034551460295915604, + 0.022454863414168358, + 0.01330475602298975, + 0.016569772735238075, + 0.01611768640577793, + 0.019730597734451294, + 0.02378033846616745, + 0.034644078463315964, + 0.01612994261085987, + 0.006898440420627594, + 0.0032679405994713306, + 0.005776162259280682, + 0.005081355106085539, + 0.006309371441602707, + 0.005340016447007656, + 0.00906316377222538, + 0.0054229325614869595, + 0.010034446604549885, + 0.0041582356207072735, + 0.006347035989165306, + 0.011434745974838734, + 0.015206770971417427, + 0.013451775535941124, + 0.008728577755391598, + 0.0044482056982815266, + 0.0062022740021348, + 0.0020257846917957067, + 0.007315571419894695, + 0.004816163331270218, + 0.003950999118387699, + 0.006757274270057678, + 0.005480199586600065, + 0.0029989539179950953, + 0.004140240605920553, + 0.0017307538073509932, + 0.003724995069205761, + 0.004866105504333973, + 0.004596269223839045, + 0.0027957421261817217, + 0.0038154851645231247, + 0.0023812383878976107, + 0.002351308474317193, + 0.0017160227289423347, + 0.003870408283546567, + 0.005091845523566008, + 0.003517648670822382, + 0.0026086396537721157, + 0.0044539617374539375, + 0.001902083051390946, + 0.008587286807596684, + 0.007021147757768631, + 0.01217569224536419, + 0.029163923114538193, + 0.01399977970868349, + 0.008785135112702847, + 0.015555708669126034, + 0.009726610966026783, + 0.01737995445728302, + 0.006195303983986378, + 0.014999018982052803, + 0.011154815554618835, + 0.02931307628750801, + 0.00924304872751236, + 0.013814706355333328, + 0.008055020123720169, + 0.021471180021762848, + 0.015317726880311966, + 0.018887609243392944, + 0.019682135432958603, + 0.02056126296520233, + 0.013484783470630646, + 0.016599297523498535, + 0.014602242968976498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.014202939346432686 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864158570766449 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.00864158570766449, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864158570766449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1264294534921646, + 0.0735115185379982, + 0.09440227597951889, + 0.10488919913768768, + 0.05882812663912773, + 0.0554020069539547, + 0.05557603761553764, + 0.17305675148963928, + 0.1336679458618164, + 0.12423675507307053 + ], + "credit_map": [ + 0.01816902868449688, + 0.007433097809553146, + 0.012857753783464432, + 0.03363185003399849, + 0.012061321176588535, + 0.010571794584393501, + 0.017791748046875, + 0.013912837952375412, + 0.010720490477979183, + 0.005724482238292694, + 0.006900510285049677, + 0.012827655300498009, + 0.008098544552922249, + 0.006996050477027893, + 0.015431292355060577, + 0.006812486797571182, + 0.012109513394534588, + 0.007617485709488392, + 0.0128003666177392, + 0.010629680939018726, + 0.009942738339304924, + 0.009625145234167576, + 0.016226595267653465, + 0.015450751408934593, + 0.012677122838795185, + 0.0058855353854596615, + 0.007630788721144199, + 0.00947453361004591, + 0.024218132719397545, + 0.021436024457216263, + 0.012069369666278362, + 0.01149768941104412, + 0.006320647895336151, + 0.003111495403572917, + 0.011753836646676064, + 0.008289007470011711, + 0.006863811518996954, + 0.007435408420860767, + 0.008394003845751286, + 0.006659911014139652, + 0.0069847614504396915, + 0.0027808675076812506, + 0.010866405442357063, + 0.007458380423486233, + 0.006533315870910883, + 0.007391277235001326, + 0.007397152949124575, + 0.005989843048155308, + 0.007058240473270416, + 0.0026282956823706627, + 0.006742013152688742, + 0.009857218712568283, + 0.006726658437401056, + 0.002930283546447754, + 0.013918657787144184, + 0.005714664235711098, + 0.019323047250509262, + 0.011465061455965042, + 0.029660673812031746, + 0.04291147366166115, + 0.018371401354670525, + 0.011465432122349739, + 0.024915380403399467, + 0.014944266527891159, + 0.015551581047475338, + 0.009902779012918472, + 0.02082890272140503, + 0.014099451713263988, + 0.019238971173763275, + 0.009982609190046787, + 0.02700820565223694, + 0.01705542765557766, + 0.014498522505164146, + 0.014371903613209724, + 0.010681268759071827, + 0.019517749547958374, + 0.025412103161215782, + 0.008612299337983131, + 0.016815153881907463, + 0.01432774867862463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.00864158570766449 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, static shot, taken from a level angle, captures a woman with her long, dark brown hair tied into a neat ponytail. She wears an off-white silk blouse and high-waisted gray trousers. She leans forward slightly in her office chair, her body curved into an elegant arc. In her right hand, she clutches a ball of white A4 paper, wrinkled from the effort of her squeezing. She's in a modern office, where soft afternoon sunlight filters through the gaps in the blinds, casting dappled shadows on the floor and walls. In the corner of the frame, a cylindrical, brushed stainless steel trash can sits quietly, next to a vibrant plant. With focused concentration, she extends her arm, flicking her wrist gently, and tosses the ball of paper with a deft yet effortless gesture. The ball traces a gentle parabola through the air, finally landing with a precise, unerring \"thud\" into the trash can's opening. The whole scene presents a focused and peaceful daily office atmosphere, and the smoothness and precision of the movements reveal a hint of casual competence and calmness.", + "seed": 12317, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_0/index_9364_seed_12317_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024474070966243745, + "vlm_logprob": -0.024474070966243745, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.024474070966243745, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07270409911870956 + ], + "opposite_logprob_tokens": [ + -4.072704315185547 + ], + "logprob_reward": -0.07270409911870956, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.07270409911870956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16149114072322845, + 0.05260774493217468, + 0.060173917561769485, + 0.051717691123485565, + 0.08947907388210297, + 0.14078567922115326, + 0.04943003132939339, + 0.06504878401756287, + 0.1454164683818817, + 0.18384955823421478 + ], + "credit_map": [ + 0.01605447195470333, + 0.016369102522730827, + 0.0188241396099329, + 0.022755630314350128, + 0.021238155663013458, + 0.01960069313645363, + 0.02341623418033123, + 0.02323269657790661, + 0.005204194225370884, + 0.0060865930281579494, + 0.006749407388269901, + 0.007125298026949167, + 0.007069525308907032, + 0.008200595155358315, + 0.006750521715730429, + 0.005421603564172983, + 0.004371544346213341, + 0.0076898206025362015, + 0.00780061911791563, + 0.007789498195052147, + 0.008060001768171787, + 0.007600329350680113, + 0.010170979425311089, + 0.006691124755889177, + 0.0036880094558000565, + 0.005100653972476721, + 0.005687198601663113, + 0.008321048691868782, + 0.006474814377725124, + 0.006223715841770172, + 0.01002704817801714, + 0.006195197347551584, + 0.007339471485465765, + 0.008015681058168411, + 0.010030626319348812, + 0.008722309954464436, + 0.012606409378349781, + 0.01491088978946209, + 0.018513932824134827, + 0.009339741431176662, + 0.00798109546303749, + 0.01439411286264658, + 0.011898910626769066, + 0.015522697009146214, + 0.017714517191052437, + 0.024844489991664886, + 0.029677588492631912, + 0.01875225454568863, + 0.003583749057725072, + 0.004703063052147627, + 0.0049848053604364395, + 0.0093192458152771, + 0.007017652038484812, + 0.00507547939196229, + 0.008913778699934483, + 0.005832252558320761, + 0.00560663640499115, + 0.006187391001731157, + 0.0076407198794186115, + 0.00663681048899889, + 0.008257808163762093, + 0.008582644164562225, + 0.015274565666913986, + 0.006862206384539604, + 0.012485365383327007, + 0.013341151177883148, + 0.012163960374891758, + 0.02514905482530594, + 0.026782596483826637, + 0.013088639825582504, + 0.03064783103764057, + 0.011757860891520977, + 0.01059234794229269, + 0.014019986614584923, + 0.012470485642552376, + 0.03158501163125038, + 0.029829135164618492, + 0.03256053477525711, + 0.038755763322114944, + 0.014036271721124649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.07270409911870956 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982667088508606 + ], + "opposite_logprob_tokens": [ + -3.5298266410827637 + ], + "logprob_reward": -0.02982667088508606, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982667088508606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29640957713127136, + 0.10190066695213318, + 0.052401892840862274, + 0.026093075051903725, + 0.09463196992874146, + 0.13356433808803558, + 0.04145561531186104, + 0.05256608501076698, + 0.10374166071414948, + 0.09723508358001709 + ], + "credit_map": [ + 0.022711655125021935, + 0.03639974817633629, + 0.033975593745708466, + 0.050787825137376785, + 0.0411568321287632, + 0.0360754132270813, + 0.045345768332481384, + 0.029956739395856857, + 0.006882957648485899, + 0.008074858225882053, + 0.014779496006667614, + 0.01154420804232359, + 0.01432355958968401, + 0.011151164770126343, + 0.02289148047566414, + 0.012252946384251118, + 0.003743859939277172, + 0.005074270535260439, + 0.007755981758236885, + 0.005146398209035397, + 0.00797660555690527, + 0.005428310018032789, + 0.010704617016017437, + 0.006571853067725897, + 0.0024275565519928932, + 0.0025281559210270643, + 0.0040710982866585255, + 0.0024823106359690428, + 0.004745780024677515, + 0.0030696429312229156, + 0.004140383563935757, + 0.0026281457394361496, + 0.004400525242090225, + 0.009240069426596165, + 0.008775699883699417, + 0.013110732659697533, + 0.013410312123596668, + 0.015229436568915844, + 0.018680067732930183, + 0.011785118840634823, + 0.007059590425342321, + 0.012723908759653568, + 0.011894564144313335, + 0.018391797319054604, + 0.015494193881750107, + 0.021185770630836487, + 0.028449052944779396, + 0.018365463241934776, + 0.0041716876439750195, + 0.005688115023076534, + 0.005822415929287672, + 0.007435975596308708, + 0.003912926651537418, + 0.006331883370876312, + 0.004619296174496412, + 0.003473312593996525, + 0.003596689784899354, + 0.004905535373836756, + 0.005717850755900145, + 0.009903262369334698, + 0.005538638681173325, + 0.006760351825505495, + 0.009566575288772583, + 0.006577182095497847, + 0.004900707863271236, + 0.010394726879894733, + 0.007514988072216511, + 0.019448013976216316, + 0.01785390079021454, + 0.018884100019931793, + 0.012829074636101723, + 0.011916144751012325, + 0.0041744522750377655, + 0.009215678088366985, + 0.006690972950309515, + 0.02817790023982525, + 0.010260568000376225, + 0.010039213113486767, + 0.020194973796606064, + 0.008481325581669807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.02982667088508606 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252490285784006 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252490285784006, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252490285784006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18155664205551147, + 0.08250994980335236, + 0.0647626742720604, + 0.03614091873168945, + 0.05603446811437607, + 0.06049014627933502, + 0.06889704614877701, + 0.04016038030385971, + 0.310137003660202, + 0.09931079298257828 + ], + "credit_map": [ + 0.01884894259274006, + 0.018962331116199493, + 0.02483300119638443, + 0.031042052432894707, + 0.022479720413684845, + 0.021640244871377945, + 0.02509874664247036, + 0.018651600927114487, + 0.006302548106759787, + 0.0069846780970692635, + 0.01067192293703556, + 0.012961534783244133, + 0.011997993104159832, + 0.009491927921772003, + 0.013296553865075111, + 0.01080278679728508, + 0.003940979018807411, + 0.007620962802320719, + 0.005994691047817469, + 0.01103141438215971, + 0.011320925317704678, + 0.007928012870252132, + 0.009622085839509964, + 0.0073036025278270245, + 0.002939516445621848, + 0.003289950080215931, + 0.00508129084482789, + 0.006321573629975319, + 0.005415451712906361, + 0.004106130450963974, + 0.005218526814132929, + 0.0037684820126742125, + 0.002406209707260132, + 0.005513775162398815, + 0.004530094098299742, + 0.007765349466353655, + 0.00822626706212759, + 0.008050366304814816, + 0.011039498262107372, + 0.008502906188368797, + 0.003737787948921323, + 0.005378835368901491, + 0.005463386420160532, + 0.007587111089378595, + 0.008215146139264107, + 0.007957086898386478, + 0.012817475013434887, + 0.009333313442766666, + 0.004068792797625065, + 0.00793825089931488, + 0.006184026133269072, + 0.012928935699164867, + 0.01033744402229786, + 0.008389909751713276, + 0.010271541774272919, + 0.008778146468102932, + 0.003408116288483143, + 0.003901788732036948, + 0.004732176661491394, + 0.007236482109874487, + 0.0057451194152235985, + 0.004847194533795118, + 0.006372775882482529, + 0.003916725516319275, + 0.013511843979358673, + 0.02545067109167576, + 0.02131735533475876, + 0.07011744379997253, + 0.06603299081325531, + 0.03928273171186447, + 0.04411761835217476, + 0.030306337401270866, + 0.0031559846829622984, + 0.009391040541231632, + 0.005267200991511345, + 0.016850728541612625, + 0.014695148915052414, + 0.013402773067355156, + 0.023798614740371704, + 0.012749302200973034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.005252490285784006 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059498813934624195 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.0059498813934624195, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059498813934624195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41497042775154114, + 0.05990143120288849, + 0.04021010547876358, + 0.0374043807387352, + 0.1064363643527031, + 0.06444363296031952, + 0.052569590508937836, + 0.046867936849594116, + 0.12648700177669525, + 0.050709161907434464 + ], + "credit_map": [ + 0.04260975122451782, + 0.05226680636405945, + 0.062077268958091736, + 0.08873841911554337, + 0.03885571286082268, + 0.04828539118170738, + 0.040901120752096176, + 0.04123596474528313, + 0.005578473675996065, + 0.006218211259692907, + 0.007588806562125683, + 0.01180187426507473, + 0.007094057276844978, + 0.007477311883121729, + 0.008405269123613834, + 0.0057374220341444016, + 0.005567767191678286, + 0.004557863809168339, + 0.005955157335847616, + 0.005426988936960697, + 0.004773521795868874, + 0.005985039286315441, + 0.004344308748841286, + 0.0035994541831314564, + 0.0044448766857385635, + 0.00510573573410511, + 0.005705199670046568, + 0.005676674656569958, + 0.0035912508610635996, + 0.0054282075725495815, + 0.004299356136471033, + 0.0031530740670859814, + 0.014142413623631, + 0.011663582175970078, + 0.016925014555454254, + 0.01728985644876957, + 0.014962071552872658, + 0.010162826627492905, + 0.012938410975039005, + 0.00835217535495758, + 0.009346974082291126, + 0.006398133002221584, + 0.01057084184139967, + 0.007691077422350645, + 0.007775500882416964, + 0.008830451406538486, + 0.008688286878168583, + 0.005142367444932461, + 0.005354258697479963, + 0.004061589017510414, + 0.006312613841146231, + 0.00619466183707118, + 0.005270757712423801, + 0.007282344624400139, + 0.010559371672570705, + 0.00753398984670639, + 0.006041926331818104, + 0.005018436349928379, + 0.007920467294752598, + 0.009814118035137653, + 0.004493024665862322, + 0.004892686847597361, + 0.004726083483546972, + 0.003961190115660429, + 0.007757569197565317, + 0.011622872203588486, + 0.0092483414337039, + 0.021559350192546844, + 0.027819478884339333, + 0.01484367623925209, + 0.02148628979921341, + 0.012149423360824585, + 0.005156615749001503, + 0.006209257058799267, + 0.007183653302490711, + 0.011312715709209442, + 0.004983367398381233, + 0.004567339085042477, + 0.007415200583636761, + 0.003881011391058564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.0059498813934624195 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863721314817667 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.00863721314817667, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863721314817667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2207404524087906, + 0.05929097905755043, + 0.038538914173841476, + 0.03737908974289894, + 0.03588508814573288, + 0.16698771715164185, + 0.07483498752117157, + 0.035550203174352646, + 0.210767462849617, + 0.12002509087324142 + ], + "credit_map": [ + 0.015465330332517624, + 0.026877732947468758, + 0.024086633697152138, + 0.04167986288666725, + 0.021262096241116524, + 0.028954505920410156, + 0.02992289513349533, + 0.032491397112607956, + 0.003766374895349145, + 0.005157746374607086, + 0.00843858066946268, + 0.009254736825823784, + 0.008298264816403389, + 0.007321768440306187, + 0.008616176433861256, + 0.008437331765890121, + 0.003667071694508195, + 0.004880283027887344, + 0.005108262877911329, + 0.0050857639871537685, + 0.005674409214407206, + 0.004073117394000292, + 0.005551485810428858, + 0.0044985185377299786, + 0.0033216848969459534, + 0.003652469255030155, + 0.005583939142525196, + 0.00599235063418746, + 0.005786977242678404, + 0.003882485441863537, + 0.005565710831433535, + 0.00359347159974277, + 0.0026015061885118484, + 0.004179155454039574, + 0.002975860610604286, + 0.0037713998463004827, + 0.005437865387648344, + 0.0053438530303537846, + 0.006462089717388153, + 0.005113356746733189, + 0.008505381643772125, + 0.015852058306336403, + 0.012596958316862583, + 0.0200041513890028, + 0.01937570609152317, + 0.02851022593677044, + 0.0366646908223629, + 0.025478556752204895, + 0.004143782425671816, + 0.007856586016714573, + 0.007174456492066383, + 0.01451782789081335, + 0.013167832978069782, + 0.009085840545594692, + 0.011348784901201725, + 0.007539876736700535, + 0.0025041692424565554, + 0.0038229329511523247, + 0.0041155763901770115, + 0.008121880702674389, + 0.0038699025753885508, + 0.003984177950769663, + 0.005141744390130043, + 0.003989816643297672, + 0.009969428181648254, + 0.019331645220518112, + 0.015088485553860664, + 0.04562544822692871, + 0.036364421248435974, + 0.037555992603302, + 0.030600760132074356, + 0.016231277957558632, + 0.004494708962738514, + 0.011793102137744427, + 0.008279086090624332, + 0.02999904192984104, + 0.015448041260242462, + 0.019935263320803642, + 0.021013086661696434, + 0.009062768891453743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.00863721314817667 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, static shot, taken from a level angle, captures a woman with her long, dark brown hair tied into a neat ponytail. She wears an off-white silk blouse and high-waisted gray trousers. She leans forward slightly in her office chair, her body curved into an elegant arc. In her right hand, she clutches a ball of white A4 paper, wrinkled from the effort of her squeezing. She's in a modern office, where soft afternoon sunlight filters through the gaps in the blinds, casting dappled shadows on the floor and walls. In the corner of the frame, a cylindrical, brushed stainless steel trash can sits quietly, next to a vibrant plant. With focused concentration, she extends her arm, flicking her wrist gently, and tosses the ball of paper with a deft yet effortless gesture. The ball traces a gentle parabola through the air, finally landing with a precise, unerring \"thud\" into the trash can's opening. The whole scene presents a focused and peaceful daily office atmosphere, and the smoothness and precision of the movements reveal a hint of casual competence and calmness.", + "seed": 12318, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_0/index_9364_seed_12318_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.029597092792391776, + "vlm_logprob": -0.029597092792391776, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.029597092792391776, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09203449636697769 + ], + "opposite_logprob_tokens": [ + -3.3420345783233643 + ], + "logprob_reward": -0.09203449636697769, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.09203449636697769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17285829782485962, + 0.08899585157632828, + 0.04105207696557045, + 0.041297297924757004, + 0.024144882336258888, + 0.057504504919052124, + 0.12237973511219025, + 0.19309267401695251, + 0.15218323469161987, + 0.10649140179157257 + ], + "credit_map": [ + 0.014647279866039753, + 0.019962387159466743, + 0.010468394495546818, + 0.025007005780935287, + 0.019805293530225754, + 0.03370654582977295, + 0.025946520268917084, + 0.02331487461924553, + 0.008808357641100883, + 0.010861234739422798, + 0.008277690038084984, + 0.014517001807689667, + 0.00997793860733509, + 0.014309480786323547, + 0.011775059625506401, + 0.010469086468219757, + 0.00351553107611835, + 0.004598583094775677, + 0.003592700231820345, + 0.005639968439936638, + 0.004197782836854458, + 0.008229766972362995, + 0.005881744436919689, + 0.005396001506596804, + 0.004814652260392904, + 0.0049723610281944275, + 0.00433429516851902, + 0.005998494569212198, + 0.0038686732295900583, + 0.007714627310633659, + 0.005341123323887587, + 0.004253068007528782, + 0.002303943270817399, + 0.002790626836940646, + 0.0018651402788236737, + 0.0035590571351349354, + 0.0029900583904236555, + 0.00536340894177556, + 0.0028694418724626303, + 0.002403206191956997, + 0.008269918151199818, + 0.008600900880992413, + 0.007452607620507479, + 0.009560281410813332, + 0.004803790245205164, + 0.007817864418029785, + 0.006503291893750429, + 0.004495848435908556, + 0.01363817322999239, + 0.011369569227099419, + 0.007126593496650457, + 0.012014271691441536, + 0.019226161763072014, + 0.02094392478466034, + 0.02221122570335865, + 0.01584981381893158, + 0.018715713173151016, + 0.02130482718348503, + 0.014865030534565449, + 0.028544830158352852, + 0.02077036164700985, + 0.02570630982518196, + 0.02551909349858761, + 0.03766651079058647, + 0.017970897257328033, + 0.03235970437526703, + 0.01646549440920353, + 0.02320156618952751, + 0.020196428522467613, + 0.013652696274220943, + 0.015422284603118896, + 0.012914170511066914, + 0.020068729296326637, + 0.016972605139017105, + 0.00887457188218832, + 0.00949180219322443, + 0.01690114103257656, + 0.011137112975120544, + 0.011056062765419483, + 0.011989372782409191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.09203449636697769 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026384279131889343 + ], + "opposite_logprob_tokens": [ + -3.6513843536376953 + ], + "logprob_reward": -0.026384279131889343, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026384279131889343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23360185325145721, + 0.11559803038835526, + 0.05493547394871712, + 0.02564476989209652, + 0.02915073186159134, + 0.04033594951033592, + 0.10767633467912674, + 0.12798497080802917, + 0.16443178057670593, + 0.10064011067152023 + ], + "credit_map": [ + 0.022180700674653053, + 0.02566785365343094, + 0.021423911675810814, + 0.044740356504917145, + 0.022681821137666702, + 0.04126179590821266, + 0.02666350081562996, + 0.028981929644942284, + 0.009717736393213272, + 0.010441131889820099, + 0.0076310099102556705, + 0.013250395655632019, + 0.013361766003072262, + 0.027680987492203712, + 0.014401246793568134, + 0.019113754853606224, + 0.005293500144034624, + 0.004758237395435572, + 0.0034909509122371674, + 0.0073092784732580185, + 0.006583083420991898, + 0.012390785850584507, + 0.006595853250473738, + 0.008513784036040306, + 0.0024529309011995792, + 0.0028479155153036118, + 0.0014893502229824662, + 0.003355829045176506, + 0.002622475614771247, + 0.005241983104497194, + 0.0038617420941591263, + 0.0037725421134382486, + 0.0027025118470191956, + 0.004230576101690531, + 0.0026162657886743546, + 0.0034907772205770016, + 0.003493309486657381, + 0.005833640228956938, + 0.0036269736010581255, + 0.0031566752586513758, + 0.005332141183316708, + 0.005211082752794027, + 0.004094298928976059, + 0.00542827881872654, + 0.0047216955572366714, + 0.006161839235574007, + 0.0049228910356760025, + 0.0044637215323746204, + 0.012946760281920433, + 0.015907300636172295, + 0.008486293256282806, + 0.01742677018046379, + 0.016278047114610672, + 0.012535281479358673, + 0.012918813154101372, + 0.011177070438861847, + 0.01460832729935646, + 0.011895127594470978, + 0.0076719848439097404, + 0.01419072225689888, + 0.02037927508354187, + 0.02233775332570076, + 0.014820721931755543, + 0.02208104357123375, + 0.018252551555633545, + 0.01931915432214737, + 0.010982650332152843, + 0.02270878292620182, + 0.02363249659538269, + 0.02918342687189579, + 0.01823289506137371, + 0.02211981825530529, + 0.009808131493628025, + 0.009642455726861954, + 0.00769359664991498, + 0.012531736865639687, + 0.013598968274891376, + 0.015556642785668373, + 0.018802925944328308, + 0.013005649670958519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.026384279131889343 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950236693024635 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950236693024635, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950236693024635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11444991827011108, + 0.07210363447666168, + 0.04551290348172188, + 0.0606849305331707, + 0.04521172493696213, + 0.08007441461086273, + 0.11264380067586899, + 0.17676162719726562, + 0.16230277717113495, + 0.13025419414043427 + ], + "credit_map": [ + 0.0165887288749218, + 0.014400622807443142, + 0.012130862101912498, + 0.016250984743237495, + 0.00993927102535963, + 0.015523373149335384, + 0.014003204181790352, + 0.015612888149917126, + 0.008141904138028622, + 0.006580437067896128, + 0.005345895886421204, + 0.0071654897183179855, + 0.008630783297121525, + 0.014628715813159943, + 0.011006016284227371, + 0.010604399256408215, + 0.004881684202700853, + 0.004837803076952696, + 0.003962199669331312, + 0.005454531405121088, + 0.0055228122510015965, + 0.008378672413527966, + 0.0071983980014920235, + 0.005276805721223354, + 0.00582126434892416, + 0.007021654862910509, + 0.0033007001038640738, + 0.005992095451802015, + 0.007778521627187729, + 0.013890919275581837, + 0.008950957097113132, + 0.007928825914859772, + 0.004781176336109638, + 0.004797649569809437, + 0.0026555212680250406, + 0.005945628508925438, + 0.006116440985351801, + 0.00873407069593668, + 0.006480108946561813, + 0.005701134912669659, + 0.009738270193338394, + 0.010238121263682842, + 0.006608679424971342, + 0.011066598817706108, + 0.008264957927167416, + 0.011814044788479805, + 0.014290557242929935, + 0.008053196594119072, + 0.009323560632765293, + 0.012868853285908699, + 0.0072568668983876705, + 0.014779740944504738, + 0.017502998933196068, + 0.01882852241396904, + 0.021237043663859367, + 0.010846227407455444, + 0.01571446657180786, + 0.01965961419045925, + 0.010920173488557339, + 0.019803760573267937, + 0.024129802361130714, + 0.025259090587496758, + 0.022433912381529808, + 0.03884083777666092, + 0.0186017993837595, + 0.017504164949059486, + 0.00956572312861681, + 0.02099337987601757, + 0.019503165036439896, + 0.021053073927760124, + 0.02297373302280903, + 0.03210774436593056, + 0.019247474148869514, + 0.017968181520700455, + 0.006265262141823769, + 0.015314294025301933, + 0.024746205657720566, + 0.013966246508061886, + 0.01893988437950611, + 0.013806663453578949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.005950236693024635 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541946023702621 + ], + "opposite_logprob_tokens": [ + -4.387541770935059 + ], + "logprob_reward": -0.012541946023702621, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541946023702621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2068747878074646, + 0.11751530319452286, + 0.052744876593351364, + 0.04127350449562073, + 0.11548202484846115, + 0.08872467279434204, + 0.06356212496757507, + 0.061962418258190155, + 0.2202460765838623, + 0.031614236533641815 + ], + "credit_map": [ + 0.023624001070857048, + 0.03312690928578377, + 0.0297839492559433, + 0.036454327404499054, + 0.0182030089199543, + 0.018082140013575554, + 0.026927750557661057, + 0.020672684535384178, + 0.016938799992203712, + 0.017841540277004242, + 0.017392149195075035, + 0.018893050029873848, + 0.009351286105811596, + 0.013083076104521751, + 0.014937976375222206, + 0.009077425114810467, + 0.006509171798825264, + 0.007285083178430796, + 0.00800124742090702, + 0.009351144544780254, + 0.004538139794021845, + 0.00590096227824688, + 0.007160842884331942, + 0.003998283762484789, + 0.004632129333913326, + 0.006398951169103384, + 0.006802479270845652, + 0.006587671115994453, + 0.004268764518201351, + 0.0045073796063661575, + 0.004745526239275932, + 0.0033306055702269077, + 0.013081393204629421, + 0.02501535415649414, + 0.00819622352719307, + 0.01814744807779789, + 0.013773021288216114, + 0.012942030094563961, + 0.015883997082710266, + 0.008442558348178864, + 0.01080637238919735, + 0.017559435218572617, + 0.007233453914523125, + 0.013424951583147049, + 0.008931487798690796, + 0.011656954884529114, + 0.012087749317288399, + 0.007024260703474283, + 0.006769569590687752, + 0.010057520121335983, + 0.004713436588644981, + 0.011322029866278172, + 0.007289299741387367, + 0.00897190161049366, + 0.008824409916996956, + 0.005613957531750202, + 0.007899172604084015, + 0.007254337426275015, + 0.0047224778681993484, + 0.007401017472147942, + 0.011891648173332214, + 0.0072829327546060085, + 0.008153789676725864, + 0.007357041817158461, + 0.020709572359919548, + 0.023831292986869812, + 0.023539798334240913, + 0.027993081137537956, + 0.03602653369307518, + 0.0247889906167984, + 0.03303683549165726, + 0.03031996451318264, + 0.0036913047078996897, + 0.0035401307977735996, + 0.00419516721740365, + 0.004677590914070606, + 0.0046555218286812305, + 0.0037719132378697395, + 0.003504096297547221, + 0.0035785117652267218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.012541946023702621 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074505746364594 + ], + "opposite_logprob_tokens": [ + -4.511074542999268 + ], + "logprob_reward": -0.011074505746364594, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074505746364594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.151530921459198, + 0.11857766658067703, + 0.0544348880648613, + 0.04423918202519417, + 0.03752415254712105, + 0.04273426532745361, + 0.07126209884881973, + 0.13502158224582672, + 0.17853765189647675, + 0.16613753139972687 + ], + "credit_map": [ + 0.022297777235507965, + 0.017750270664691925, + 0.016691986471414566, + 0.023155024275183678, + 0.014270155690610409, + 0.022474190220236778, + 0.015066475607454777, + 0.01982504315674305, + 0.011478028260171413, + 0.012857316993176937, + 0.009069708175957203, + 0.013524890877306461, + 0.01407675538212061, + 0.023491768166422844, + 0.016131427139043808, + 0.017947779968380928, + 0.0052812108770012856, + 0.006541476119309664, + 0.0030723821837455034, + 0.0060254428535699844, + 0.005979905836284161, + 0.01224321685731411, + 0.007496429141610861, + 0.007794826757162809, + 0.004064294975250959, + 0.0062042176723480225, + 0.0035096120554953814, + 0.005686901044100523, + 0.005032006651163101, + 0.008462982252240181, + 0.005771738942712545, + 0.005507428664714098, + 0.0032878578640520573, + 0.004025984089821577, + 0.003867196384817362, + 0.004887359216809273, + 0.0038355642464011908, + 0.007860778830945492, + 0.005250906106084585, + 0.004508506506681442, + 0.0035815341398119926, + 0.005990045145153999, + 0.0027031332720071077, + 0.005010971799492836, + 0.005914660170674324, + 0.006399852689355612, + 0.008945761248469353, + 0.004188308492302895, + 0.008448591455817223, + 0.00958668440580368, + 0.007738483604043722, + 0.011918541043996811, + 0.008088227361440659, + 0.009159915149211884, + 0.008789991959929466, + 0.007531662005931139, + 0.01764846220612526, + 0.01708696223795414, + 0.012808983214199543, + 0.018096700310707092, + 0.015089345164597034, + 0.015463586896657944, + 0.01605801284313202, + 0.02276952750980854, + 0.03312180936336517, + 0.020657172426581383, + 0.02064024843275547, + 0.023456351831555367, + 0.025596212595701218, + 0.017810847610235214, + 0.021531354635953903, + 0.015723655000329018, + 0.02601938508450985, + 0.02261420711874962, + 0.008890276774764061, + 0.021275455132126808, + 0.03160008415579796, + 0.018180763348937035, + 0.02130074054002762, + 0.016256604343652725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.011074505746364594 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, static shot, taken from a level angle, captures a woman with her long, dark brown hair tied into a neat ponytail. She wears an off-white silk blouse and high-waisted gray trousers. She leans forward slightly in her office chair, her body curved into an elegant arc. In her right hand, she clutches a ball of white A4 paper, wrinkled from the effort of her squeezing. She's in a modern office, where soft afternoon sunlight filters through the gaps in the blinds, casting dappled shadows on the floor and walls. In the corner of the frame, a cylindrical, brushed stainless steel trash can sits quietly, next to a vibrant plant. With focused concentration, she extends her arm, flicking her wrist gently, and tosses the ball of paper with a deft yet effortless gesture. The ball traces a gentle parabola through the air, finally landing with a precise, unerring \"thud\" into the trash can's opening. The whole scene presents a focused and peaceful daily office atmosphere, and the smoothness and precision of the movements reveal a hint of casual competence and calmness.", + "seed": 12319, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_0/index_9364_seed_12319_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.032633600477129224, + "vlm_logprob": -0.032633600477129224, + "vlm_yes_no_margin": 4.399999952316284, + "vlm_token_credit": -0.032633600477129224, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11768660694360733 + ], + "opposite_logprob_tokens": [ + -3.6176865100860596 + ], + "logprob_reward": -0.11768660694360733, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.11768660694360733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.174886554479599, + 0.040663618594408035, + 0.03736019879579544, + 0.030939601361751556, + 0.03582730144262314, + 0.1046103984117508, + 0.10566125810146332, + 0.1286960244178772, + 0.18056531250476837, + 0.16078972816467285 + ], + "credit_map": [ + 0.03310467675328255, + 0.0186255294829607, + 0.015858832746744156, + 0.019342515617609024, + 0.01808520220220089, + 0.03037530928850174, + 0.02511289343237877, + 0.01438160426914692, + 0.0038770048413425684, + 0.004968471359461546, + 0.004431223496794701, + 0.00796338077634573, + 0.003920856863260269, + 0.005082375835627317, + 0.006145166698843241, + 0.004275135695934296, + 0.002848639152944088, + 0.0031815925613045692, + 0.003120268229395151, + 0.007847035303711891, + 0.0032088016159832478, + 0.004712861962616444, + 0.008056405000388622, + 0.004384592175483704, + 0.003490692237392068, + 0.003134072059765458, + 0.0031158828642219305, + 0.004438419360667467, + 0.0032856666948646307, + 0.0037476830184459686, + 0.005919272545725107, + 0.003807911416515708, + 0.003122251946479082, + 0.003272358560934663, + 0.0030458748806267977, + 0.006792500149458647, + 0.0038591984193772078, + 0.0043281251564621925, + 0.007561429403722286, + 0.0038455643225461245, + 0.008468766696751118, + 0.01057069469243288, + 0.008233544416725636, + 0.011039353907108307, + 0.008600885979831219, + 0.031035050749778748, + 0.015727611258625984, + 0.010934495367109776, + 0.01224686298519373, + 0.015185028314590454, + 0.011520336382091045, + 0.016720540821552277, + 0.009524405002593994, + 0.009992888197302818, + 0.017691344022750854, + 0.012779852375388145, + 0.008560266345739365, + 0.015664931386709213, + 0.018866781145334244, + 0.01124904677271843, + 0.026796700432896614, + 0.016529593616724014, + 0.016373254358768463, + 0.014655450358986855, + 0.01601969078183174, + 0.023228181526064873, + 0.03580372408032417, + 0.018389878794550896, + 0.03227655217051506, + 0.02111620269715786, + 0.017050672322511673, + 0.016680404543876648, + 0.012890013866126537, + 0.03190900385379791, + 0.03626348450779915, + 0.015545157715678215, + 0.025183700025081635, + 0.012152240611612797, + 0.014529845677316189, + 0.01231627818197012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.11768660694360733 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821586675941944 + ], + "opposite_logprob_tokens": [ + -4.018215656280518 + ], + "logprob_reward": -0.01821586675941944, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01821586675941944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.189738929271698, + 0.050645921379327774, + 0.1254144310951233, + 0.10663238167762756, + 0.034945711493492126, + 0.04718135669827461, + 0.1422111839056015, + 0.17462709546089172, + 0.07130748778581619, + 0.05729549750685692 + ], + "credit_map": [ + 0.02315899357199669, + 0.020270293578505516, + 0.01507209986448288, + 0.026550468057394028, + 0.017647476866841316, + 0.03145476058125496, + 0.03465849161148071, + 0.020926345139741898, + 0.006911446340382099, + 0.0038877276238054037, + 0.0037835994735360146, + 0.006884844042360783, + 0.0034642659593373537, + 0.010103842243552208, + 0.009767472743988037, + 0.005842721089720726, + 0.009794977493584156, + 0.008674483746290207, + 0.009978724643588066, + 0.02674800157546997, + 0.00838533602654934, + 0.025558173656463623, + 0.023329515010118484, + 0.01294522825628519, + 0.011762449517846107, + 0.008680201135575771, + 0.008572558872401714, + 0.014232628047466278, + 0.010318218730390072, + 0.025978388264775276, + 0.017264779657125473, + 0.009823156520724297, + 0.003376292297616601, + 0.004213233944028616, + 0.0029989290051162243, + 0.004296672996133566, + 0.003978462889790535, + 0.0053766220808029175, + 0.005970505531877279, + 0.004734992049634457, + 0.004070238675922155, + 0.0056412434205412865, + 0.004551118239760399, + 0.005579381715506315, + 0.007386087905615568, + 0.006921061780303717, + 0.007282576989382505, + 0.005749647971242666, + 0.017153454944491386, + 0.016806015744805336, + 0.014622006565332413, + 0.014229499734938145, + 0.012319947592914104, + 0.03549961373209953, + 0.016246240586042404, + 0.015334393829107285, + 0.01858355849981308, + 0.03192342072725296, + 0.051326122134923935, + 0.01494709961116314, + 0.020777929574251175, + 0.010148986242711544, + 0.012274504639208317, + 0.014645475894212723, + 0.005882273893803358, + 0.0060561178252100945, + 0.006697250064462423, + 0.007579609751701355, + 0.01165649015456438, + 0.012143254280090332, + 0.007477087900042534, + 0.013815405778586864, + 0.005421913228929043, + 0.007136044558137655, + 0.0068176318891346455, + 0.005360042676329613, + 0.009447932243347168, + 0.008560697548091412, + 0.008265498094260693, + 0.006285735405981541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.01821586675941944 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251185968518257 + ], + "opposite_logprob_tokens": [ + -5.255251407623291 + ], + "logprob_reward": -0.005251185968518257, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251185968518257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.151773601770401, + 0.03375070542097092, + 0.021600523963570595, + 0.03181463107466698, + 0.05458804965019226, + 0.049173276871442795, + 0.10513115674257278, + 0.34370627999305725, + 0.16584353148937225, + 0.042618248611688614 + ], + "credit_map": [ + 0.020941561087965965, + 0.015300717204809189, + 0.01280411146581173, + 0.017626170068979263, + 0.01239095814526081, + 0.035521164536476135, + 0.02459874376654625, + 0.012590174563229084, + 0.0030944105237722397, + 0.003921070601791143, + 0.0027132986579090357, + 0.00868680514395237, + 0.002239266177639365, + 0.004024911671876907, + 0.005402653012424707, + 0.0036682910285890102, + 0.0016949319979175925, + 0.0018568661762401462, + 0.002134192269295454, + 0.003952877130359411, + 0.0023356133606284857, + 0.002717452822253108, + 0.004153725691139698, + 0.0027548628859221935, + 0.003958424553275108, + 0.0027887781616300344, + 0.0025196990463882685, + 0.004245762713253498, + 0.0026699884328991175, + 0.007681158371269703, + 0.00491371750831604, + 0.0030370987951755524, + 0.006003629881888628, + 0.005214245989918709, + 0.00704108364880085, + 0.009949260391294956, + 0.0031629104632884264, + 0.0051190792582929134, + 0.009253536351025105, + 0.00884429831057787, + 0.004703161306679249, + 0.00814706552773714, + 0.004215021152049303, + 0.005559283308684826, + 0.004361283499747515, + 0.008313460275530815, + 0.008730604313313961, + 0.00514339841902256, + 0.014860987663269043, + 0.014116734266281128, + 0.010271822102367878, + 0.012771341018378735, + 0.00932274293154478, + 0.020967816933989525, + 0.013103149831295013, + 0.009716566652059555, + 0.03911302983760834, + 0.06462831795215607, + 0.11355140805244446, + 0.031128806993365288, + 0.02126956731081009, + 0.01844850182533264, + 0.02559095062315464, + 0.02997569367289543, + 0.012561026960611343, + 0.020439235493540764, + 0.029491746798157692, + 0.019128043204545975, + 0.02769831009209156, + 0.02480427920818329, + 0.01656080223619938, + 0.015160087496042252, + 0.0038981768302619457, + 0.005453350953757763, + 0.0034750672057271004, + 0.009457181207835674, + 0.005522994324564934, + 0.003924294840544462, + 0.0069886124692857265, + 0.003898569382727146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.005251185968518257 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067897900938988 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.016067897900938988, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067897900938988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07992809265851974, + 0.02159785106778145, + 0.021594375371932983, + 0.06430206447839737, + 0.027670131996273994, + 0.10159782320261002, + 0.2011050134897232, + 0.347561776638031, + 0.09235330671072006, + 0.04228954017162323 + ], + "credit_map": [ + 0.00979283545166254, + 0.007978946901857853, + 0.00596682308241725, + 0.011994704604148865, + 0.006207275670021772, + 0.011283289641141891, + 0.017241887748241425, + 0.009462329559028149, + 0.0019449371611699462, + 0.0024731617886573076, + 0.0015859708655625582, + 0.004377962090075016, + 0.0021782314870506525, + 0.002532707527279854, + 0.004435714799910784, + 0.002069166162982583, + 0.001754583092406392, + 0.0014237700961530209, + 0.002137925708666444, + 0.004750610329210758, + 0.0014308842364698648, + 0.00405150605365634, + 0.00395988579839468, + 0.0020852137822657824, + 0.006837212946265936, + 0.005417433567345142, + 0.005774044897407293, + 0.008001454174518585, + 0.006056199315935373, + 0.016342520713806152, + 0.009608598425984383, + 0.0062646022997796535, + 0.003170396201312542, + 0.003460479201748967, + 0.0026757591404020786, + 0.006176780443638563, + 0.001699943095445633, + 0.0028956537134945393, + 0.004558186046779156, + 0.003032936481758952, + 0.00878999661654234, + 0.017455890774726868, + 0.006600141990929842, + 0.012116813100874424, + 0.007568534463644028, + 0.02687232382595539, + 0.011013876646757126, + 0.011180249974131584, + 0.026549486443400383, + 0.02147778496146202, + 0.02583407610654831, + 0.02723829261958599, + 0.015528024174273014, + 0.033335231244564056, + 0.0283649954944849, + 0.022777128964662552, + 0.03455193340778351, + 0.06531070172786713, + 0.1068798378109932, + 0.03170079365372658, + 0.03757312893867493, + 0.018462637439370155, + 0.024325991049408913, + 0.028756799176335335, + 0.009411509148776531, + 0.015466099604964256, + 0.025031747296452522, + 0.00914125144481659, + 0.010821791365742683, + 0.008397377096116543, + 0.0068999724462628365, + 0.007183566689491272, + 0.00419012364000082, + 0.008320952765643597, + 0.009129391051828861, + 0.0052383095026016235, + 0.005657856352627277, + 0.002523004077374935, + 0.003970677498728037, + 0.003259227378293872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.016067897900938988 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium, static shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing an off-white silk blouse and high-waisted gray trousers with her dark brown hair in a ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman lean forward in her chair and toss a ball of wrinkled white A4 paper into a cylindrical brushed stainless steel trash can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a modern office with soft afternoon sunlight filtering through blinds, casting dappled shadows, and featuring a vibrant plant in the corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper ball trace a gentle parabola through the air before landing in the trash can?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946444813162088 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.005946444813162088, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946444813162088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12679646909236908, + 0.03362641483545303, + 0.018511461094021797, + 0.049037009477615356, + 0.03847017511725426, + 0.03749392554163933, + 0.18711380660533905, + 0.3499799966812134, + 0.09287421405315399, + 0.06609653681516647 + ], + "credit_map": [ + 0.020316658541560173, + 0.014648892916738987, + 0.010688441805541515, + 0.01559520699083805, + 0.013474440202116966, + 0.024440094828605652, + 0.018245026469230652, + 0.009387712925672531, + 0.0034714853391051292, + 0.0034815852995961905, + 0.002831575693562627, + 0.007420981768518686, + 0.002646826673299074, + 0.00392468785867095, + 0.006265834905207157, + 0.0035834377631545067, + 0.001557989977300167, + 0.001828747452236712, + 0.0017156924586743116, + 0.003926029894500971, + 0.0015880507417023182, + 0.0021928902715444565, + 0.0036137434653937817, + 0.002088316483423114, + 0.004883927758783102, + 0.004304553847759962, + 0.004151853732764721, + 0.00661823246628046, + 0.005329820327460766, + 0.010432341136038303, + 0.008833633735775948, + 0.004482647404074669, + 0.004026867914944887, + 0.004268184304237366, + 0.004291458521038294, + 0.007049339823424816, + 0.002503808354958892, + 0.004064344335347414, + 0.006784763187170029, + 0.005481406580656767, + 0.002955458592623472, + 0.0036312479060143232, + 0.005828863941133022, + 0.006263934075832367, + 0.003510286333039403, + 0.003965859301388264, + 0.006693576462566853, + 0.00464470311999321, + 0.014088013209402561, + 0.025617798790335655, + 0.011594907380640507, + 0.019588148221373558, + 0.025191431865096092, + 0.040366679430007935, + 0.02987835370004177, + 0.02078845538198948, + 0.03685666620731354, + 0.058403488248586655, + 0.11231811344623566, + 0.03430507704615593, + 0.028479164466261864, + 0.020641829818487167, + 0.02709907293319702, + 0.031876590102910995, + 0.008546539582312107, + 0.009913243353366852, + 0.008337559178471565, + 0.010667804628610611, + 0.017747897654771805, + 0.01745951920747757, + 0.010209936648607254, + 0.009991711936891079, + 0.0056952121667563915, + 0.011629809625446796, + 0.007389320060610771, + 0.006320210173726082, + 0.011374331079423428, + 0.008281448855996132, + 0.009825628250837326, + 0.005580577999353409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.005946444813162088 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_1/index_tensor([7804, 7804])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_1/index_tensor([7804, 7804])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a88215e4ad7b899f2ee8042de2123d258769d57d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_1/index_tensor([7804, 7804])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady medium close-up places the viewer at the center of a fierce sandstorm, focusing on a weathered man. He defiantly leans sideways against the gusting wind, leaning slightly forward for balance. He wears a faded, yet still a deep, cyan-colored, coarse robe, its rough fabric caked with fine grains of sand. Every wrinkle on his face, carved by the scorching sun and wind, seems to tell a long story. His dark, cracked skin contrasts sharply with his gray, dust-stained beard. Quickly and instinctively, he raises his right arm, tightly shielding his mouth and nose with his muscular forearm to shield himself from the overwhelming, almost suffocating dust. However, his hand does not clench completely; instead, he carefully opens his fingers, creating a narrow gap. Through this gap, his deep eyes gaze with an unwavering, resolute gaze at the unknown world ahead, obscured by the wind and sand. There is no fear in his eyes, only focus and determination. The entire picture is completely immersed in Steve McCurry's iconic photography style: the air is filled with rich and saturated ochre and orange tones, forming a strong and harmonious color contrast with the blue of the man's robe; although the light is diffused by the sand and dust and is extremely soft, it depicts every texture and attached sand on his face with extremely high clarity, creating an epic atmosphere full of struggle and extremely human dignity and tenacious strength.", + "seed": 9784, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_1/index_7804_seed_9784_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02449608885217458, + "vlm_logprob": -0.02449608885217458, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.02449608885217458, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0178145132958889 + ], + "opposite_logprob_tokens": [ + -5.017814636230469 + ], + "logprob_reward": -0.0178145132958889, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0178145132958889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1876334846019745, + 0.11237247288227081, + 0.10866731405258179, + 0.11589886993169785, + 0.15577079355716705, + 0.04474814981222153, + 0.06039684638381004, + 0.10868071019649506, + 0.03773967921733856, + 0.0680917501449585 + ], + "credit_map": [ + 0.017469944432377815, + 0.01418947521597147, + 0.02556951902806759, + 0.03328658640384674, + 0.019170306622982025, + 0.02324306219816208, + 0.026307573541998863, + 0.028396986424922943, + 0.019074205309152603, + 0.013036923483014107, + 0.017704002559185028, + 0.019392425194382668, + 0.014171355403959751, + 0.006969323381781578, + 0.012844206765294075, + 0.009180022403597832, + 0.010838115587830544, + 0.012349934317171574, + 0.019004572182893753, + 0.0173221193253994, + 0.014035898260772228, + 0.010861706919968128, + 0.013764920644462109, + 0.010490045882761478, + 0.009722030721604824, + 0.017695773392915726, + 0.013162962161004543, + 0.020439209416508675, + 0.014259503223001957, + 0.01054496131837368, + 0.015061427839100361, + 0.015012997202575207, + 0.01628400757908821, + 0.016889754682779312, + 0.01928628608584404, + 0.020984936505556107, + 0.020534884184598923, + 0.020775537937879562, + 0.01941973716020584, + 0.021595632657408714, + 0.010474937036633492, + 0.004970510955899954, + 0.006822406779974699, + 0.004929400514811277, + 0.006213050801306963, + 0.0026584307197481394, + 0.005193205084651709, + 0.0034862009342759848, + 0.015922119840979576, + 0.005262219812721014, + 0.011731545440852642, + 0.008035959675908089, + 0.005443208385258913, + 0.004092948976904154, + 0.004932046867907047, + 0.004976792261004448, + 0.016565898433327675, + 0.012965470552444458, + 0.01337452046573162, + 0.016215352341532707, + 0.010416271165013313, + 0.007435612380504608, + 0.01486372109502554, + 0.016843853518366814, + 0.009487481787800789, + 0.003826815402135253, + 0.006633429322391748, + 0.004120124038308859, + 0.0036259249318391085, + 0.0025085657835006714, + 0.003880233271047473, + 0.0036571044474840164, + 0.012713742442429066, + 0.008488617837429047, + 0.00997784174978733, + 0.015163647010922432, + 0.0071371435187757015, + 0.00439122598618269, + 0.00506896385923028, + 0.0051505593582987785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0178145132958889 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025138226337730885 + ], + "opposite_logprob_tokens": [ + -6.002513885498047 + ], + "logprob_reward": -0.0025138226337730885, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025138226337730885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17839987576007843, + 0.1118449866771698, + 0.14729079604148865, + 0.11218230426311493, + 0.1413019895553589, + 0.05904097482562065, + 0.03592418134212494, + 0.07335641235113144, + 0.05973318591713905, + 0.08092531561851501 + ], + "credit_map": [ + 0.016020221635699272, + 0.019133754074573517, + 0.02440727688372135, + 0.0236409530043602, + 0.022944090887904167, + 0.022884881123900414, + 0.026069657877087593, + 0.023299019783735275, + 0.013937792740762234, + 0.020813893526792526, + 0.013936462812125683, + 0.01763223297894001, + 0.013131463900208473, + 0.006559999659657478, + 0.014769621193408966, + 0.011063518933951855, + 0.014031119644641876, + 0.016594935208559036, + 0.021583737805485725, + 0.030132891610264778, + 0.014207308180630207, + 0.014456149190664291, + 0.02068033255636692, + 0.015604332089424133, + 0.011584324762225151, + 0.014990677125751972, + 0.012309844605624676, + 0.01804460771381855, + 0.013288423418998718, + 0.012050800025463104, + 0.01553665567189455, + 0.014376972801983356, + 0.019562391564249992, + 0.01636563427746296, + 0.016566982492804527, + 0.01847873441874981, + 0.01781236194074154, + 0.013724196702241898, + 0.01536113303154707, + 0.023430559784173965, + 0.008691724389791489, + 0.00742181483656168, + 0.007032732013612986, + 0.00785340927541256, + 0.010738654993474483, + 0.004435393027961254, + 0.007723748683929443, + 0.005143499001860619, + 0.004830676596611738, + 0.003599554533138871, + 0.004580516368150711, + 0.005798359867185354, + 0.005293333902955055, + 0.0033678319305181503, + 0.004599492531269789, + 0.0038544144481420517, + 0.009641895070672035, + 0.014293552376329899, + 0.008217367343604565, + 0.0098623251542449, + 0.00935665424913168, + 0.0058122542686760426, + 0.008851421065628529, + 0.007320939563214779, + 0.01005125604569912, + 0.006695356220006943, + 0.007002346683293581, + 0.008533690124750137, + 0.006705652456730604, + 0.005522540304809809, + 0.006986792664974928, + 0.008235550485551357, + 0.00866714958101511, + 0.012785093858838081, + 0.009746076539158821, + 0.01710754819214344, + 0.01289012935012579, + 0.005897447466850281, + 0.006620371248573065, + 0.0072115021757781506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0025138226337730885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543240562081337 + ], + "opposite_logprob_tokens": [ + -4.387543201446533 + ], + "logprob_reward": -0.012543240562081337, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543240562081337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16643381118774414, + 0.08809205144643784, + 0.09981651604175568, + 0.16109494864940643, + 0.13297869265079498, + 0.038990698754787445, + 0.058180954307317734, + 0.2068430334329605, + 0.021906519308686256, + 0.02566269226372242 + ], + "credit_map": [ + 0.01827767677605152, + 0.013531814329326153, + 0.02222619764506817, + 0.020339813083410263, + 0.018625276163220406, + 0.02213405631482601, + 0.023143837228417397, + 0.02815515547990799, + 0.009534981101751328, + 0.014749468304216862, + 0.010712172836065292, + 0.014196861535310745, + 0.009031898342072964, + 0.005672474391758442, + 0.014602200128138065, + 0.009592002257704735, + 0.008601359091699123, + 0.01288237888365984, + 0.012686061672866344, + 0.017470328137278557, + 0.014062030240893364, + 0.009692386724054813, + 0.012429858557879925, + 0.011992122046649456, + 0.014485317282378674, + 0.018802903592586517, + 0.021626384928822517, + 0.022326761856675148, + 0.020311664789915085, + 0.014656062237918377, + 0.026266807690262794, + 0.022619057446718216, + 0.01749100722372532, + 0.014719106256961823, + 0.016124723479151726, + 0.018757609650492668, + 0.016121404245495796, + 0.013310354202985764, + 0.01543640997260809, + 0.02101808786392212, + 0.008463509380817413, + 0.004455126356333494, + 0.005677452310919762, + 0.006810145452618599, + 0.0032446517143398523, + 0.003342011012136936, + 0.003279980504885316, + 0.003717828541994095, + 0.01049372460693121, + 0.005790264345705509, + 0.008147213608026505, + 0.00843476876616478, + 0.008188002742826939, + 0.004635218530893326, + 0.00672893924638629, + 0.005762831307947636, + 0.0335560142993927, + 0.02748763933777809, + 0.026626529172062874, + 0.03165192902088165, + 0.0196478720754385, + 0.014058359898626804, + 0.027415169402956963, + 0.026399556547403336, + 0.005131445359438658, + 0.0029292432591319084, + 0.0037284179124981165, + 0.0024557614233344793, + 0.0019703840371221304, + 0.0013399675954133272, + 0.0019368469947949052, + 0.0024144556373357773, + 0.004132136702537537, + 0.003866942133754492, + 0.0035969314631074667, + 0.0030160744208842516, + 0.004982589744031429, + 0.0018642924260348082, + 0.002079635625705123, + 0.002124091377481818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.012543240562081337 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005640584276989102 + ], + "opposite_logprob_tokens": [ + -7.500564098358154 + ], + "logprob_reward": -0.0005640584276989102, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005640584276989102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14821013808250427, + 0.09202304482460022, + 0.2424302101135254, + 0.23513437807559967, + 0.0395558625459671, + 0.0655839741230011, + 0.0638507679104805, + 0.031151607632637024, + 0.01822707988321781, + 0.06383293122053146 + ], + "credit_map": [ + 0.018003983423113823, + 0.016268249601125717, + 0.01934242434799671, + 0.027097312733530998, + 0.012638824060559273, + 0.01640532910823822, + 0.01888878457248211, + 0.019565245136618614, + 0.00782494805753231, + 0.009232302196323872, + 0.011509817093610764, + 0.015863316133618355, + 0.013869727030396461, + 0.009218108840286732, + 0.014610184356570244, + 0.00989464670419693, + 0.02512592263519764, + 0.03187497332692146, + 0.03693104162812233, + 0.05338253825902939, + 0.0241851843893528, + 0.016056137159466743, + 0.03057309426367283, + 0.024301331490278244, + 0.01634584367275238, + 0.032990776002407074, + 0.028016138821840286, + 0.03045777603983879, + 0.03127758949995041, + 0.02266019396483898, + 0.037358373403549194, + 0.036027707159519196, + 0.005754115059971809, + 0.004133508540689945, + 0.005060666706413031, + 0.006891697179526091, + 0.004736008588224649, + 0.0031904997304081917, + 0.00380589603446424, + 0.005983476061373949, + 0.01788732409477234, + 0.006229082588106394, + 0.011736416257917881, + 0.0076735969632864, + 0.00836890283972025, + 0.0039993408136069775, + 0.005541607737541199, + 0.004147708881646395, + 0.014313689433038235, + 0.009389862418174744, + 0.011172972619533539, + 0.006592953577637672, + 0.006903745699673891, + 0.0046372064389288425, + 0.005830364301800728, + 0.005009977146983147, + 0.005409553647041321, + 0.005386337637901306, + 0.004251196049153805, + 0.003204186214134097, + 0.002890742151066661, + 0.002383979270234704, + 0.0039991154335439205, + 0.0036264993250370026, + 0.004127394407987595, + 0.002581998473033309, + 0.0029142163693904877, + 0.0015763748669996858, + 0.0018621602794155478, + 0.0013703544391319156, + 0.001653432846069336, + 0.0021411492489278316, + 0.011061822064220905, + 0.0076779890805482864, + 0.009652743116021156, + 0.013931233435869217, + 0.008576277643442154, + 0.003955623600631952, + 0.004823631606996059, + 0.004153616726398468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.0005640584276989102 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904480934143066 + ], + "opposite_logprob_tokens": [ + -2.4640448093414307 + ], + "logprob_reward": -0.08904480934143066, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904480934143066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2735845744609833, + 0.11456021666526794, + 0.26502400636672974, + 0.1355278342962265, + 0.039542365819215775, + 0.04944512993097305, + 0.023692570626735687, + 0.03305429965257645, + 0.014979096129536629, + 0.05058987811207771 + ], + "credit_map": [ + 0.043265655636787415, + 0.028797447681427002, + 0.044403690844774246, + 0.041693877428770065, + 0.026746943593025208, + 0.01868095062673092, + 0.03335901349782944, + 0.03663700073957443, + 0.009500576183199883, + 0.012414292432367802, + 0.01300364825874567, + 0.016728589311242104, + 0.017413491383194923, + 0.011206553317606449, + 0.020157430320978165, + 0.014135640114545822, + 0.020305337384343147, + 0.036017656326293945, + 0.031508490443229675, + 0.04061615839600563, + 0.03319768235087395, + 0.03433292731642723, + 0.037935979664325714, + 0.031109826639294624, + 0.010277742519974709, + 0.022402992472052574, + 0.018648555502295494, + 0.03109695389866829, + 0.015678828582167625, + 0.008923697285354137, + 0.016312088817358017, + 0.012186992913484573, + 0.006949084345251322, + 0.0044884951785206795, + 0.005903338082134724, + 0.006506098434329033, + 0.004154117777943611, + 0.0035144207067787647, + 0.003700614208355546, + 0.00432620570063591, + 0.012724725529551506, + 0.005497312173247337, + 0.00843694619834423, + 0.006275358609855175, + 0.0053385780192911625, + 0.0033246243838220835, + 0.00411659711971879, + 0.0037309967447072268, + 0.0060448055155575275, + 0.002422946272417903, + 0.004273267928510904, + 0.0020409729331731796, + 0.002855005906894803, + 0.0019010654650628567, + 0.0019771724473685026, + 0.0021773362532258034, + 0.004155799280852079, + 0.00486329710111022, + 0.003815239295363426, + 0.0035160903353244066, + 0.004214151296764612, + 0.0027325619012117386, + 0.004895445425063372, + 0.0048617152497172356, + 0.0033922074362635612, + 0.002014393452554941, + 0.002281774766743183, + 0.0017631734954193234, + 0.0016710245981812477, + 0.0008808934362605214, + 0.0013856717851012945, + 0.0015899587888270617, + 0.00982989277690649, + 0.008041645400226116, + 0.007132555358111858, + 0.007852978073060513, + 0.0077216168865561485, + 0.0028280853293836117, + 0.0034635611809790134, + 0.0037195493932813406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.08904480934143066 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium close-up places the viewer at the center of a fierce sandstorm, focusing on a weathered man. He defiantly leans sideways against the gusting wind, leaning slightly forward for balance. He wears a faded, yet still a deep, cyan-colored, coarse robe, its rough fabric caked with fine grains of sand. Every wrinkle on his face, carved by the scorching sun and wind, seems to tell a long story. His dark, cracked skin contrasts sharply with his gray, dust-stained beard. Quickly and instinctively, he raises his right arm, tightly shielding his mouth and nose with his muscular forearm to shield himself from the overwhelming, almost suffocating dust. However, his hand does not clench completely; instead, he carefully opens his fingers, creating a narrow gap. Through this gap, his deep eyes gaze with an unwavering, resolute gaze at the unknown world ahead, obscured by the wind and sand. There is no fear in his eyes, only focus and determination. The entire picture is completely immersed in Steve McCurry's iconic photography style: the air is filled with rich and saturated ochre and orange tones, forming a strong and harmonious color contrast with the blue of the man's robe; although the light is diffused by the sand and dust and is extremely soft, it depicts every texture and attached sand on his face with extremely high clarity, creating an epic atmosphere full of struggle and extremely human dignity and tenacious strength.", + "seed": 9785, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_1/index_7804_seed_9785_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.037611781369196254, + "vlm_logprob": -0.037611781369196254, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.037611781369196254, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02412790060043335 + ], + "opposite_logprob_tokens": [ + -5.149127960205078 + ], + "logprob_reward": -0.02412790060043335, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.02412790060043335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21769699454307556, + 0.08407614380121231, + 0.21289514005184174, + 0.17117011547088623, + 0.041596248745918274, + 0.08885148167610168, + 0.06073255464434624, + 0.03899359330534935, + 0.04181954637169838, + 0.04216817766427994 + ], + "credit_map": [ + 0.028023947030305862, + 0.027503356337547302, + 0.027534203603863716, + 0.03274138271808624, + 0.022550322115421295, + 0.023234346881508827, + 0.030526425689458847, + 0.02558300830423832, + 0.008543628267943859, + 0.010371362790465355, + 0.01152733527123928, + 0.013188239187002182, + 0.00954811368137598, + 0.00793334562331438, + 0.010808787308633327, + 0.012155328877270222, + 0.0164959654211998, + 0.04707854613661766, + 0.02180328220129013, + 0.024108894169330597, + 0.024760039523243904, + 0.02770773321390152, + 0.03061727248132229, + 0.020323406904935837, + 0.011775166727602482, + 0.023889966309070587, + 0.0209360234439373, + 0.041810303926467896, + 0.015021084807813168, + 0.015968231484293938, + 0.02020079828798771, + 0.0215685423463583, + 0.004293649457395077, + 0.005579608958214521, + 0.004355002194643021, + 0.006520116701722145, + 0.00382195389829576, + 0.004502585623413324, + 0.00662629259750247, + 0.005897035356611013, + 0.01874767430126667, + 0.0076352315954864025, + 0.02383272908627987, + 0.007168836425989866, + 0.004689077381044626, + 0.005609442014247179, + 0.00499738147482276, + 0.016171110793948174, + 0.011537578888237476, + 0.007114696782082319, + 0.013061030767858028, + 0.007434461265802383, + 0.005661184433847666, + 0.0034731493797153234, + 0.00641388026997447, + 0.006036574486643076, + 0.006564911920577288, + 0.007456288207322359, + 0.005876125302165747, + 0.006257577333599329, + 0.004022258799523115, + 0.0025617103092372417, + 0.0027232978027313948, + 0.00353141943924129, + 0.010807261802256107, + 0.005734584294259548, + 0.007731264922767878, + 0.0049744416028261185, + 0.0031675214413553476, + 0.0028443429619073868, + 0.0031628103461116552, + 0.003397321328520775, + 0.006029979325830936, + 0.005942117422819138, + 0.005265936255455017, + 0.005417944397777319, + 0.00455348938703537, + 0.005251233000308275, + 0.004147551022469997, + 0.005559926852583885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.02412790060043335 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004663543310016394 + ], + "opposite_logprob_tokens": [ + -5.379663467407227 + ], + "logprob_reward": -0.004663543310016394, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004663543310016394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12589679658412933, + 0.10278104245662689, + 0.11370065063238144, + 0.04564688727259636, + 0.06990273296833038, + 0.4435468316078186, + 0.030204858630895615, + 0.025547726079821587, + 0.023472987115383148, + 0.019299499690532684 + ], + "credit_map": [ + 0.015470776706933975, + 0.013147187419235706, + 0.013349656946957111, + 0.019285455346107483, + 0.014783673919737339, + 0.013982996344566345, + 0.019405188038945198, + 0.01647186651825905, + 0.010964928194880486, + 0.0091581204906106, + 0.013296435587108135, + 0.013788960874080658, + 0.013621985912322998, + 0.014108240604400635, + 0.010962258093059063, + 0.016880111768841743, + 0.011061456054449081, + 0.011033880524337292, + 0.013508661650121212, + 0.023373248055577278, + 0.008130945265293121, + 0.01375338714569807, + 0.014494108967483044, + 0.01834496296942234, + 0.0038396932650357485, + 0.005635630805045366, + 0.004909662529826164, + 0.008080837316811085, + 0.007325915619730949, + 0.00666331360116601, + 0.005107812583446503, + 0.0040840222500264645, + 0.006815411150455475, + 0.012735757976770401, + 0.007783128879964352, + 0.013037231750786304, + 0.007209411822259426, + 0.005824398715049028, + 0.007621611934155226, + 0.008875786326825619, + 0.1146988496184349, + 0.021366611123085022, + 0.14806681871414185, + 0.0229099802672863, + 0.011208000592887402, + 0.01361131388694048, + 0.01854722388088703, + 0.09313803166151047, + 0.004222105722874403, + 0.004241477698087692, + 0.0038703326135873795, + 0.00413230387493968, + 0.002772617619484663, + 0.004063083790242672, + 0.003710448043420911, + 0.003192492760717869, + 0.00411989726126194, + 0.0042136250995099545, + 0.003435816615819931, + 0.003087290795519948, + 0.0024987985379993916, + 0.0031365707982331514, + 0.002877999795600772, + 0.0021777283400297165, + 0.004625160712748766, + 0.0035276885610073805, + 0.003601173171773553, + 0.0033125663176178932, + 0.0021495542023330927, + 0.0019303682493045926, + 0.0020387033000588417, + 0.0022877720184624195, + 0.0027599846944212914, + 0.00283807166852057, + 0.002358527621254325, + 0.002222580835223198, + 0.0022303666919469833, + 0.0025411350652575493, + 0.0021600935142487288, + 0.0021887391339987516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.004663543310016394 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016068601980805397 + ], + "opposite_logprob_tokens": [ + -4.141068458557129 + ], + "logprob_reward": -0.016068601980805397, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016068601980805397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09967591613531113, + 0.042315445840358734, + 0.08550579100847244, + 0.06635008752346039, + 0.018461402505636215, + 0.6198403835296631, + 0.02469727396965027, + 0.017176339402794838, + 0.011518451385200024, + 0.014458895660936832 + ], + "credit_map": [ + 0.011908907443284988, + 0.01205470971763134, + 0.010691491886973381, + 0.01767108216881752, + 0.013679812662303448, + 0.010615694336593151, + 0.011137393303215504, + 0.01191682182252407, + 0.0038269406650215387, + 0.005232870113104582, + 0.005756521597504616, + 0.008354061283171177, + 0.003664546413347125, + 0.0024545900523662567, + 0.0041054110042750835, + 0.008920506574213505, + 0.006486806552857161, + 0.008874601684510708, + 0.011045347899198532, + 0.01650627888739109, + 0.005681570153683424, + 0.009653333574533463, + 0.009660767391324043, + 0.017597083002328873, + 0.0057185348123312, + 0.012419197708368301, + 0.007743621710687876, + 0.009376352652907372, + 0.010047024115920067, + 0.0069054849445819855, + 0.007441480178385973, + 0.006698393728584051, + 0.0024086281191557646, + 0.002748341066762805, + 0.0023757419548928738, + 0.002466142177581787, + 0.0018029639031738043, + 0.0020517809316515923, + 0.0022403921466320753, + 0.0023674126714468002, + 0.16095155477523804, + 0.034401293843984604, + 0.20369313657283783, + 0.04033849015831947, + 0.015997549518942833, + 0.021152816712856293, + 0.025869715958833694, + 0.11743583530187607, + 0.005594897083938122, + 0.0028269572649151087, + 0.0054941424168646336, + 0.003111571539193392, + 0.0017578573897480965, + 0.001377763575874269, + 0.002199699403718114, + 0.002334384946152568, + 0.003458770690485835, + 0.002657463774085045, + 0.0025386461056768894, + 0.002537427470088005, + 0.001516262418590486, + 0.0014693863922730088, + 0.001433563302271068, + 0.0015648200642317533, + 0.0029040740337222815, + 0.0017837767954915762, + 0.002005137735977769, + 0.0013384617632254958, + 0.0009052317473106086, + 0.0007799382437951863, + 0.0008237717556767166, + 0.0009780596010386944, + 0.0026761284098029137, + 0.00172773830126971, + 0.0024124281480908394, + 0.0016062599606812, + 0.0011579120764508843, + 0.0017265169881284237, + 0.001385401701554656, + 0.0017665114719420671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.016068601980805397 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003904534096363932 + ], + "opposite_logprob_tokens": [ + -7.875390529632568 + ], + "logprob_reward": -0.0003904534096363932, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.0003904534096363932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22901730239391327, + 0.09634039551019669, + 0.13934463262557983, + 0.16356074810028076, + 0.04041948541998863, + 0.2280048280954361, + 0.03065013326704502, + 0.02309359423816204, + 0.024034982547163963, + 0.02553383819758892 + ], + "credit_map": [ + 0.021447526291012764, + 0.03523401916027069, + 0.029735075309872627, + 0.040463365614414215, + 0.030928878113627434, + 0.02199256792664528, + 0.02724350616335869, + 0.021972358226776123, + 0.011325912550091743, + 0.009979121387004852, + 0.01188267394900322, + 0.012833851389586926, + 0.008786489255726337, + 0.011796212755143642, + 0.011052331887185574, + 0.01868380419909954, + 0.011674558743834496, + 0.019090872257947922, + 0.016534049063920975, + 0.020892435684800148, + 0.014644450508058071, + 0.014786834828555584, + 0.021347394213080406, + 0.02037404663860798, + 0.013320628553628922, + 0.015633344650268555, + 0.01855270192027092, + 0.03925810754299164, + 0.019905876368284225, + 0.018873615190386772, + 0.01656409725546837, + 0.021452367305755615, + 0.005657281260937452, + 0.004418801516294479, + 0.005282628815621138, + 0.0057133338414132595, + 0.003867542836815119, + 0.0037918142043054104, + 0.004747495520859957, + 0.006940588355064392, + 0.05365249142050743, + 0.013710593804717064, + 0.07184034585952759, + 0.013167107477784157, + 0.006729929242283106, + 0.008344419300556183, + 0.010190565139055252, + 0.05036938562989235, + 0.005932831205427647, + 0.005097971763461828, + 0.004937437828630209, + 0.004078712314367294, + 0.0029722359031438828, + 0.002071992726996541, + 0.003134547034278512, + 0.0024244042579084635, + 0.0037680347450077534, + 0.0035681412555277348, + 0.0030944463796913624, + 0.003075686749070883, + 0.0023592361249029636, + 0.0026978873647749424, + 0.0021493122912943363, + 0.0023808504920452833, + 0.005304454360157251, + 0.0034815792459994555, + 0.004067502915859222, + 0.0036879717372357845, + 0.0019515211461111903, + 0.0017134988447651267, + 0.0016219104873016477, + 0.0022065441589802504, + 0.005231034010648727, + 0.0028060697950422764, + 0.0042905258014798164, + 0.0038863597437739372, + 0.002157769165933132, + 0.0027283974923193455, + 0.0019721004646271467, + 0.002461582887917757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0003904534096363932 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280840754508972 + ], + "opposite_logprob_tokens": [ + -2.017808437347412 + ], + "logprob_reward": -0.14280840754508972, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14280840754508972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1790701150894165, + 0.12337549030780792, + 0.21865735948085785, + 0.11141074448823929, + 0.028242403641343117, + 0.21157024800777435, + 0.039191074669361115, + 0.04487849026918411, + 0.025532659143209457, + 0.018071362748742104 + ], + "credit_map": [ + 0.020575974136590958, + 0.026241902261972427, + 0.02153247594833374, + 0.032700199633836746, + 0.021135695278644562, + 0.018207240849733353, + 0.01773795485496521, + 0.020938677713274956, + 0.013593259267508984, + 0.014479837380349636, + 0.01573992520570755, + 0.022148581221699715, + 0.009096928872168064, + 0.01027207262814045, + 0.013999114744365215, + 0.024045774713158607, + 0.01765037700533867, + 0.03756019100546837, + 0.022312864661216736, + 0.029168255627155304, + 0.021146055310964584, + 0.027896685525774956, + 0.029777545481920242, + 0.03314539045095444, + 0.010074464604258537, + 0.013814284466207027, + 0.012397576123476028, + 0.01587541401386261, + 0.02063562162220478, + 0.012623735703527927, + 0.014637786895036697, + 0.011351857334375381, + 0.0032613142393529415, + 0.0038425687234848738, + 0.0031700334511697292, + 0.0047036223113536835, + 0.0022497489117085934, + 0.002729712752625346, + 0.0033453877549618483, + 0.004940017592161894, + 0.05441658943891525, + 0.011390101164579391, + 0.07014123350381851, + 0.012151218019425869, + 0.005339766386896372, + 0.007304180413484573, + 0.008271200582385063, + 0.04255596175789833, + 0.0054518552497029305, + 0.008639697916805744, + 0.005232950672507286, + 0.006116142030805349, + 0.0039396220818161964, + 0.0022629043087363243, + 0.0037204925902187824, + 0.003827407956123352, + 0.005718873348087072, + 0.007812002208083868, + 0.00553943170234561, + 0.006310106720775366, + 0.004897141829133034, + 0.004903776571154594, + 0.004636796656996012, + 0.005060361698269844, + 0.005341494455933571, + 0.0037868148647248745, + 0.004281177651137114, + 0.0029680209700018167, + 0.002236620057374239, + 0.002198034431785345, + 0.001968148397281766, + 0.002752348082140088, + 0.002493923297151923, + 0.0029137774836272, + 0.002011719858273864, + 0.0023269103839993477, + 0.001694812555797398, + 0.001965014496818185, + 0.0017694855341687799, + 0.0028957193717360497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.14280840754508972 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium close-up places the viewer at the center of a fierce sandstorm, focusing on a weathered man. He defiantly leans sideways against the gusting wind, leaning slightly forward for balance. He wears a faded, yet still a deep, cyan-colored, coarse robe, its rough fabric caked with fine grains of sand. Every wrinkle on his face, carved by the scorching sun and wind, seems to tell a long story. His dark, cracked skin contrasts sharply with his gray, dust-stained beard. Quickly and instinctively, he raises his right arm, tightly shielding his mouth and nose with his muscular forearm to shield himself from the overwhelming, almost suffocating dust. However, his hand does not clench completely; instead, he carefully opens his fingers, creating a narrow gap. Through this gap, his deep eyes gaze with an unwavering, resolute gaze at the unknown world ahead, obscured by the wind and sand. There is no fear in his eyes, only focus and determination. The entire picture is completely immersed in Steve McCurry's iconic photography style: the air is filled with rich and saturated ochre and orange tones, forming a strong and harmonious color contrast with the blue of the man's robe; although the light is diffused by the sand and dust and is extremely soft, it depicts every texture and attached sand on his face with extremely high clarity, creating an epic atmosphere full of struggle and extremely human dignity and tenacious strength.", + "seed": 9786, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_1/index_7804_seed_9786_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028122508677188306, + "vlm_logprob": -0.028122508677188306, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.028122508677188306, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014278624206781387 + ], + "opposite_logprob_tokens": [ + -5.764278411865234 + ], + "logprob_reward": -0.014278624206781387, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.014278624206781387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03343813121318817, + 0.015041515231132507, + 0.012812399305403233, + 0.012374961748719215, + 0.9005780816078186, + 0.0048624142073094845, + 0.004304971545934677, + 0.003652973799034953, + 0.007073617074638605, + 0.005860847886651754 + ], + "credit_map": [ + 0.0026533440686762333, + 0.004926093854010105, + 0.003344492521136999, + 0.004495587665587664, + 0.004742617718875408, + 0.004745081998407841, + 0.004838123917579651, + 0.003692793194204569, + 0.0018150802934542298, + 0.00213810452260077, + 0.0016481002094224095, + 0.001551978406496346, + 0.0020585861057043076, + 0.001943191047757864, + 0.0023574689403176308, + 0.0015290076844394207, + 0.002284458139911294, + 0.0015559432795271277, + 0.0021179281175136566, + 0.0015565165085718036, + 0.001341209514066577, + 0.001216118922457099, + 0.0018274810863658786, + 0.0009127446101047099, + 0.002058628713712096, + 0.0012272478779777884, + 0.0017519386019557714, + 0.00177620945032686, + 0.0009639866766519845, + 0.001626084907911718, + 0.0016896314918994904, + 0.0012812357163056731, + 0.3935002386569977, + 0.022169062867760658, + 0.4059949517250061, + 0.024182191118597984, + 0.006702023558318615, + 0.0179196335375309, + 0.013487176969647408, + 0.016622891649603844, + 0.0006420407444238663, + 0.0006758600939065218, + 0.000465399120002985, + 0.0005512465722858906, + 0.0004074022581335157, + 0.000648453482426703, + 0.0006841341964900494, + 0.0007878781761974096, + 0.0005725097726099193, + 0.0005623564356938004, + 0.0004066601977683604, + 0.0005541783175431192, + 0.00045275286538526416, + 0.0005435202037915587, + 0.0004955736803822219, + 0.0007174208876676857, + 0.0005160192376933992, + 0.0005511323106475174, + 0.0005304639926180243, + 0.0004047858528792858, + 0.00043156289029866457, + 0.0004697387630585581, + 0.00037623642128892243, + 0.00037303456338122487, + 0.0010390077950432897, + 0.0011813860619440675, + 0.0008692172705195844, + 0.0006615689489990473, + 0.0012719167862087488, + 0.0007945435354486108, + 0.0006429552449844778, + 0.0006130226538516581, + 0.001175678102299571, + 0.0004974452313035727, + 0.0015161436749622226, + 0.0007908067782409489, + 0.00047112072934396565, + 0.0005528636975213885, + 0.0003635030298028141, + 0.0004932871670462191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.014278624206781387 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004117939155548811 + ], + "opposite_logprob_tokens": [ + -5.504117965698242 + ], + "logprob_reward": -0.004117939155548811, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004117939155548811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02161991037428379, + 0.010645190253853798, + 0.00994542334228754, + 0.007650510407984257, + 0.9270779490470886, + 0.007648590486496687, + 0.0017241518944501877, + 0.004261511843651533, + 0.006906534545123577, + 0.0025203381665050983 + ], + "credit_map": [ + 0.0019691670313477516, + 0.0033852041233330965, + 0.0019547315314412117, + 0.0033054612576961517, + 0.0026455579791218042, + 0.0031284792348742485, + 0.0028779408894479275, + 0.002353368327021599, + 0.001423209672793746, + 0.0018801932455971837, + 0.0009963703341782093, + 0.0013831587275490165, + 0.0012893357779830694, + 0.001108755124732852, + 0.001490073511376977, + 0.001074094558134675, + 0.0012598895700648427, + 0.0018434547819197178, + 0.001167292008176446, + 0.0011428958969190717, + 0.0008072717464528978, + 0.0012376995291560888, + 0.0013218404492363334, + 0.0011650798842310905, + 0.0013706715544685721, + 0.001027771388180554, + 0.0011460784589871764, + 0.000965391518548131, + 0.0005911689950153232, + 0.0008986974135041237, + 0.0007987362914718688, + 0.0008519942639395595, + 0.4049064517021179, + 0.023331299424171448, + 0.4177691340446472, + 0.024943144991993904, + 0.007030380889773369, + 0.018211305141448975, + 0.013846904039382935, + 0.017039282247424126, + 0.001064501702785492, + 0.0010164271807298064, + 0.000992594170384109, + 0.0014444233383983374, + 0.0007664727163501084, + 0.0006950561073608696, + 0.0008297801250591874, + 0.0008393349125981331, + 0.00024311958986800164, + 0.00020338733156677336, + 0.0001975234627025202, + 0.00020484607375692576, + 0.00015357451047748327, + 0.0002867418806999922, + 0.00019716224051080644, + 0.0002377967321081087, + 0.0007603706908412278, + 0.0005344860255718231, + 0.0006307351868599653, + 0.00034649187000468373, + 0.0004430653643794358, + 0.000684016034938395, + 0.0003355459193699062, + 0.0005268006352707744, + 0.0009562713094055653, + 0.0010658829705789685, + 0.0007501759100705385, + 0.000722092401701957, + 0.0011412921594455838, + 0.0010137851350009441, + 0.0005992923397570848, + 0.0006577428430318832, + 0.0003838274278678, + 0.0003674031177069992, + 0.00026203252491541207, + 0.0003903720644302666, + 0.000323968444718048, + 0.00024387019220739603, + 0.00030122927273623645, + 0.00024763503461144865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.004117939155548811 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637804538011551 + ], + "opposite_logprob_tokens": [ + -4.75863790512085 + ], + "logprob_reward": -0.008637804538011551, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637804538011551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.00979267805814743, + 0.007832235656678677, + 0.005696739070117474, + 0.004195499233901501, + 0.9467242956161499, + 0.014627310447394848, + 0.006701605394482613, + 0.0012558680027723312, + 0.000913609575945884, + 0.002260129200294614 + ], + "credit_map": [ + 0.0008653329568915069, + 0.0013419011374935508, + 0.0011229229858145118, + 0.001464266562834382, + 0.0011226218193769455, + 0.0012333308113738894, + 0.0016173474723473191, + 0.001024953555315733, + 0.0009441803558729589, + 0.0010941920336335897, + 0.0008952061762101948, + 0.0008519984548911452, + 0.0007633729255758226, + 0.0010679545812308788, + 0.0010376309510320425, + 0.0011776997707784176, + 0.0006818678230047226, + 0.0009174774750135839, + 0.0005279788165353239, + 0.0008529944461770356, + 0.0005340539501048625, + 0.0005980777787044644, + 0.0010271893115714192, + 0.0005570992361754179, + 0.0006520127644762397, + 0.0005312457215040922, + 0.0006488401559181511, + 0.0005322404904291034, + 0.0004014499136246741, + 0.0004621882690116763, + 0.0005916632944718003, + 0.00037585836253128946, + 0.4137912094593048, + 0.02367459423840046, + 0.4264225363731384, + 0.025476641952991486, + 0.007200302556157112, + 0.01867632195353508, + 0.014094839803874493, + 0.017387909814715385, + 0.002046627923846245, + 0.002213619416579604, + 0.0010831319959834218, + 0.0036779700312763453, + 0.0013188262237235904, + 0.0010152820032089949, + 0.001668831566348672, + 0.0016030214028432965, + 0.0009634679299779236, + 0.0010001864284276962, + 0.0008215712150558829, + 0.0009415838867425919, + 0.0006636684411205351, + 0.000828618067316711, + 0.0006546948570758104, + 0.0008278142777271569, + 0.0001871985150501132, + 0.0002163166500395164, + 0.00018414613441564143, + 0.00015257428458426148, + 0.00012915265688207, + 0.0001440304913558066, + 0.00010708496847655624, + 0.0001353643456241116, + 0.0001416947488905862, + 0.00010363957699155435, + 0.00012712679745163769, + 0.00011203793837921694, + 0.00011691485997289419, + 0.00011905770224984735, + 0.00011304396321065724, + 8.00940251792781e-05, + 0.00035661691799759865, + 0.00023683891049586236, + 0.0003882247838191688, + 0.00034548764233477414, + 0.00019236875232309103, + 0.0003011799417436123, + 0.0001976555067813024, + 0.00024175670114345849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.008637804538011551 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006381620769388974 + ], + "opposite_logprob_tokens": [ + -7.375638008117676 + ], + "logprob_reward": -0.0006381620769388974, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006381620769388974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10122769325971603, + 0.06521923840045929, + 0.06129331886768341, + 0.02323593944311142, + 0.7070999145507812, + 0.009504320099949837, + 0.005824862513691187, + 0.008739267475903034, + 0.00688827084377408, + 0.010967206209897995 + ], + "credit_map": [ + 0.011101851239800453, + 0.014470258727669716, + 0.01080379355698824, + 0.01384379155933857, + 0.014806787483394146, + 0.012491828761994839, + 0.011061085388064384, + 0.012648279778659344, + 0.009532807394862175, + 0.010469251312315464, + 0.006791132502257824, + 0.008643892593681812, + 0.007217924576252699, + 0.00805842038244009, + 0.009172743186354637, + 0.00533306086435914, + 0.012032379396259785, + 0.007578949443995953, + 0.01101363729685545, + 0.006224918179214001, + 0.00562255596742034, + 0.005978863220661879, + 0.007790431845933199, + 0.005051582120358944, + 0.005904400255531073, + 0.0020170852076262236, + 0.004127520136535168, + 0.0023569592740386724, + 0.0013971644220873713, + 0.00279977940954268, + 0.0021656125318259, + 0.0024674171581864357, + 0.3085702359676361, + 0.01765422150492668, + 0.31821244955062866, + 0.019049199298024178, + 0.0056383851915597916, + 0.014177155680954456, + 0.010800366289913654, + 0.012997878715395927, + 0.001086115138605237, + 0.001317475805990398, + 0.0009304421255365014, + 0.0009835092350840569, + 0.0008434408227913082, + 0.0008176066912710667, + 0.0018594274297356606, + 0.0016663017449900508, + 0.0008380986982956529, + 0.0010108196875080466, + 0.0006816847599111497, + 0.0006971648545004427, + 0.0006659456412307918, + 0.0007388715166598558, + 0.0005536535754799843, + 0.0006386233726516366, + 0.001342227915301919, + 0.0013490724377334118, + 0.0009269873844459653, + 0.000819977605715394, + 0.0009139789035543799, + 0.0012361656408756971, + 0.0008198156137950718, + 0.0013310409849509597, + 0.000913557130843401, + 0.000989130581729114, + 0.0007956615881994367, + 0.0010942864464595914, + 0.0010651875054463744, + 0.0006049092626199126, + 0.0007714221719652414, + 0.0006541150505654514, + 0.0017703089397400618, + 0.001388023141771555, + 0.0016861679032444954, + 0.0015842136926949024, + 0.0014467625878751278, + 0.0008695904398337007, + 0.0011620126897469163, + 0.0010601251851767302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0006381620769388974 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294001340866089 + ], + "opposite_logprob_tokens": [ + -2.2379400730133057 + ], + "logprob_reward": -0.11294001340866089, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294001340866089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.013147108256816864, + 0.011539206840097904, + 0.006762273143976927, + 0.0027614273130893707, + 0.9503966569900513, + 0.0032799644395709038, + 0.0019746480975300074, + 0.0055738878436386585, + 0.0028045079670846462, + 0.0017602726584300399 + ], + "credit_map": [ + 0.0011751918355003, + 0.0019470051629468799, + 0.0012984187342226505, + 0.0019936133176088333, + 0.001776844379492104, + 0.002218390116468072, + 0.0013131462037563324, + 0.0014244989724829793, + 0.0009722422691993415, + 0.0017394255846738815, + 0.0012043308233842254, + 0.001518554170615971, + 0.0012952719116583467, + 0.0015576549340039492, + 0.0019510159036144614, + 0.0013007116504013538, + 0.0008624595357105136, + 0.0013418332673609257, + 0.0007847261731512845, + 0.0009307569707743824, + 0.0006085455534048378, + 0.0006083709886297584, + 0.0009121132898144424, + 0.0007134669576771557, + 0.0003405270108487457, + 0.0003844969323836267, + 0.0003262695390731096, + 0.0003809824993368238, + 0.0002945137384813279, + 0.00027429615147411823, + 0.000381576013751328, + 0.0003787653986364603, + 0.41519561409950256, + 0.023514114320278168, + 0.4284709692001343, + 0.02554512955248356, + 0.007170567754656076, + 0.01881657913327217, + 0.014201405458152294, + 0.017482319846749306, + 0.0004095611802767962, + 0.0005340538918972015, + 0.0002832397003658116, + 0.00041203643195331097, + 0.00036547164199873805, + 0.00034072634298354387, + 0.0004962319508194923, + 0.00043864338658750057, + 0.00025134251336567104, + 0.00034935461007989943, + 0.00017011433374136686, + 0.00019332613737788051, + 0.00018526753410696983, + 0.00023379523190669715, + 0.000212085826206021, + 0.0003793618525378406, + 0.0010075026657432318, + 0.0009347015875391662, + 0.0007914176094345748, + 0.0005848557921126485, + 0.0008126002503558993, + 0.0005343042430467904, + 0.0004646806337404996, + 0.00044382494525052607, + 0.00039989178185351193, + 0.0005175484693609178, + 0.00027955733821727335, + 0.00033060467103496194, + 0.00042966153705492616, + 0.00030498963315039873, + 0.00030716569744981825, + 0.00023508889717049897, + 0.0003400426357984543, + 0.00024257853510789573, + 0.0003976117295678705, + 0.00024675525492057204, + 0.000171226667589508, + 0.00011150859063491225, + 0.00013702429714612663, + 0.00011352482397342101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.11294001340866089 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium close-up places the viewer at the center of a fierce sandstorm, focusing on a weathered man. He defiantly leans sideways against the gusting wind, leaning slightly forward for balance. He wears a faded, yet still a deep, cyan-colored, coarse robe, its rough fabric caked with fine grains of sand. Every wrinkle on his face, carved by the scorching sun and wind, seems to tell a long story. His dark, cracked skin contrasts sharply with his gray, dust-stained beard. Quickly and instinctively, he raises his right arm, tightly shielding his mouth and nose with his muscular forearm to shield himself from the overwhelming, almost suffocating dust. However, his hand does not clench completely; instead, he carefully opens his fingers, creating a narrow gap. Through this gap, his deep eyes gaze with an unwavering, resolute gaze at the unknown world ahead, obscured by the wind and sand. There is no fear in his eyes, only focus and determination. The entire picture is completely immersed in Steve McCurry's iconic photography style: the air is filled with rich and saturated ochre and orange tones, forming a strong and harmonious color contrast with the blue of the man's robe; although the light is diffused by the sand and dust and is extremely soft, it depicts every texture and attached sand on his face with extremely high clarity, creating an epic atmosphere full of struggle and extremely human dignity and tenacious strength.", + "seed": 9787, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_1/index_7804_seed_9787_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03466351053211838, + "vlm_logprob": -0.03466351053211838, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.03466351053211838, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03304736688733101 + ], + "opposite_logprob_tokens": [ + -5.033047199249268 + ], + "logprob_reward": -0.03304736688733101, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.03304736688733101, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12095688283443451, + 0.03780967742204666, + 0.04877543821930885, + 0.0502694733440876, + 0.11407455056905746, + 0.05201352760195732, + 0.10898187756538391, + 0.38207748532295227, + 0.049277227371931076, + 0.03576378896832466 + ], + "credit_map": [ + 0.033825989812612534, + 0.016140077263116837, + 0.0180814266204834, + 0.010487263090908527, + 0.009523358196020126, + 0.012653880752623081, + 0.008347366005182266, + 0.011897540651261806, + 0.00585576519370079, + 0.0033858097158372402, + 0.00347653916105628, + 0.0036948437336832285, + 0.006242241710424423, + 0.0070961518213152885, + 0.003397877560928464, + 0.004660450387746096, + 0.0070780483074486256, + 0.004269130993634462, + 0.003932771272957325, + 0.004950109403580427, + 0.009758343920111656, + 0.007082892116159201, + 0.004960279446095228, + 0.006743867881596088, + 0.003362983465194702, + 0.005889840889722109, + 0.004278869368135929, + 0.005108961835503578, + 0.013875328935682774, + 0.007084940560162067, + 0.0049473256804049015, + 0.005721225868910551, + 0.007617153227329254, + 0.017191234976053238, + 0.00578342517837882, + 0.009015277028083801, + 0.03678295761346817, + 0.019170908257365227, + 0.008964036591351032, + 0.009549570269882679, + 0.005108554847538471, + 0.006973030976951122, + 0.004248914308845997, + 0.005476906429976225, + 0.01278176810592413, + 0.00644877552986145, + 0.006320093292742968, + 0.004655489232391119, + 0.012122269719839096, + 0.012255092151463032, + 0.018669767305254936, + 0.01648840680718422, + 0.011804201640188694, + 0.013676431961357594, + 0.015391194261610508, + 0.008574526757001877, + 0.04161352291703224, + 0.037945106625556946, + 0.0707504153251648, + 0.03177540376782417, + 0.040157098323106766, + 0.05250490456819534, + 0.04880546033382416, + 0.058525603264570236, + 0.005221505183726549, + 0.005831885151565075, + 0.005360174458473921, + 0.0063755870796740055, + 0.005396115127950907, + 0.006282859016209841, + 0.007965087890625, + 0.006844015326350927, + 0.0077288164757192135, + 0.0038812877610325813, + 0.002284393412992358, + 0.002473720582202077, + 0.006859323009848595, + 0.006243929732590914, + 0.003608859609812498, + 0.002683460246771574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.03304736688733101 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036513819359242916 + ], + "opposite_logprob_tokens": [ + -5.6286516189575195 + ], + "logprob_reward": -0.0036513819359242916, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036513819359242916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11489605158567429, + 0.03714557737112045, + 0.0392020083963871, + 0.0266177449375391, + 0.05525953695178032, + 0.05976226180791855, + 0.21636033058166504, + 0.06157539039850235, + 0.038047172129154205, + 0.35113394260406494 + ], + "credit_map": [ + 0.03555548936128616, + 0.014163616113364697, + 0.021032799035310745, + 0.008690983057022095, + 0.008539368398487568, + 0.01183023490011692, + 0.00641031377017498, + 0.008673243224620819, + 0.005387740675359964, + 0.004209407605230808, + 0.004865138791501522, + 0.003990589175373316, + 0.004375683143734932, + 0.006495289970189333, + 0.003615674329921603, + 0.004206052049994469, + 0.008551944978535175, + 0.003698477754369378, + 0.004930731374770403, + 0.006298231426626444, + 0.003743103938177228, + 0.004714507143944502, + 0.0040182252414524555, + 0.0032467851415276527, + 0.001909063314087689, + 0.0030024689622223377, + 0.0024380648974329233, + 0.002311976859346032, + 0.007417100016027689, + 0.0036383026745170355, + 0.0028189406730234623, + 0.0030818271916359663, + 0.0032316562719643116, + 0.008180166594684124, + 0.002586064161732793, + 0.0032608245965093374, + 0.020539501681923866, + 0.010148267261683941, + 0.0039535341784358025, + 0.0033595224376767874, + 0.005087403114885092, + 0.006956202443689108, + 0.005363724660128355, + 0.00483598792925477, + 0.012731155380606651, + 0.008110885508358479, + 0.005333258304744959, + 0.011343645863234997, + 0.030235981568694115, + 0.022377487272024155, + 0.03487670049071312, + 0.02619435451924801, + 0.027051838114857674, + 0.02939653769135475, + 0.028331216424703598, + 0.017896201461553574, + 0.009158573113381863, + 0.007101677358150482, + 0.008228502236306667, + 0.004680613987147808, + 0.008776246570050716, + 0.009233416989445686, + 0.00661584734916687, + 0.007780516054481268, + 0.002932819304987788, + 0.004343410488218069, + 0.003590682754293084, + 0.004487290512770414, + 0.0058556897565722466, + 0.004916283767670393, + 0.006174422800540924, + 0.005746574606746435, + 0.05680739879608154, + 0.0357515774667263, + 0.03762008249759674, + 0.03237033635377884, + 0.030300378799438477, + 0.05430197715759277, + 0.04750776290893555, + 0.05647442862391472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.0036513819359242916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644304238259792 + ], + "opposite_logprob_tokens": [ + -4.758644104003906 + ], + "logprob_reward": -0.008644304238259792, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644304238259792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05469294264912605, + 0.05350307375192642, + 0.0449734628200531, + 0.02820681408047676, + 0.07342954725027084, + 0.14065025746822357, + 0.3044355809688568, + 0.08952002972364426, + 0.06653162091970444, + 0.14405667781829834 + ], + "credit_map": [ + 0.010403821244835854, + 0.006364708300679922, + 0.005759991239756346, + 0.006072735879570246, + 0.006551598198711872, + 0.007942226715385914, + 0.004760896787047386, + 0.006836962420493364, + 0.007787002250552177, + 0.004191427957266569, + 0.006166075821965933, + 0.004301046021282673, + 0.009164448827505112, + 0.009795553050935268, + 0.004897345323115587, + 0.007200174033641815, + 0.009479224681854248, + 0.004362506326287985, + 0.0033895778469741344, + 0.003722528228536248, + 0.008290719240903854, + 0.007144232280552387, + 0.00387013447470963, + 0.004714542534202337, + 0.003415814135223627, + 0.0031372830271720886, + 0.003035562112927437, + 0.002818301785737276, + 0.00638654176145792, + 0.0036111241206526756, + 0.0020302741322666407, + 0.0037719132378697395, + 0.007222259882837534, + 0.012859482318162918, + 0.004962194245308638, + 0.006353416014462709, + 0.015667956322431564, + 0.009673280641436577, + 0.008648848161101341, + 0.008042111992835999, + 0.0104513606056571, + 0.0178520604968071, + 0.00953864399343729, + 0.011526918970048428, + 0.03884945437312126, + 0.022305898368358612, + 0.015857309103012085, + 0.014268612489104271, + 0.03980409726500511, + 0.03134576231241226, + 0.054318636655807495, + 0.03790836036205292, + 0.03628578037023544, + 0.040521007031202316, + 0.039022479206323624, + 0.0252294410020113, + 0.00816134549677372, + 0.007650978863239288, + 0.010711107403039932, + 0.010102191008627415, + 0.012052800506353378, + 0.010956818237900734, + 0.013722073286771774, + 0.01616271398961544, + 0.006798154208809137, + 0.006606097333133221, + 0.007126152515411377, + 0.009525060653686523, + 0.008104847744107246, + 0.0077040186151862144, + 0.01100704725831747, + 0.009660248644649982, + 0.022512398660182953, + 0.015084188431501389, + 0.016223208978772163, + 0.013354781083762646, + 0.012630644254386425, + 0.022004766389727592, + 0.01911841332912445, + 0.023128265514969826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.008644304238259792 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009248746791854501 + ], + "opposite_logprob_tokens": [ + -7.000925064086914 + ], + "logprob_reward": -0.0009248746791854501, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009248746791854501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04817543923854828, + 0.015127141959965229, + 0.023484740406274796, + 0.0134424502030015, + 0.02111910469830036, + 0.037008002400398254, + 0.054662082344293594, + 0.3949887454509735, + 0.016953567042946815, + 0.375038743019104 + ], + "credit_map": [ + 0.01386045478284359, + 0.003898652270436287, + 0.012318897061049938, + 0.003519118996337056, + 0.003998955711722374, + 0.004513699561357498, + 0.002548280404880643, + 0.003517378121614456, + 0.0020184465683996677, + 0.0011704126372933388, + 0.0019033228745684028, + 0.0017436088528484106, + 0.0018749560695141554, + 0.002559197600930929, + 0.0019754478707909584, + 0.0018817486707121134, + 0.004175707697868347, + 0.0023161566350609064, + 0.002324027009308338, + 0.0034987840335816145, + 0.002945698332041502, + 0.003117685206234455, + 0.0026923955883830786, + 0.002414288464933634, + 0.0013384276535362005, + 0.0014203840401023626, + 0.0011775469174608588, + 0.0014966203598305583, + 0.003456028876826167, + 0.0018846882740035653, + 0.0014158034464344382, + 0.0012529506348073483, + 0.0012362360721454024, + 0.0031907674856483936, + 0.0014091030461713672, + 0.0018160217441618443, + 0.006225074641406536, + 0.0034526775125414133, + 0.0021493067033588886, + 0.001639916212297976, + 0.003309649182483554, + 0.00483744079247117, + 0.004700648132711649, + 0.004710332956165075, + 0.006214938126504421, + 0.004718275740742683, + 0.004589330404996872, + 0.003927391022443771, + 0.007465539034456015, + 0.006041716318577528, + 0.010289324447512627, + 0.007393870502710342, + 0.005883851088583469, + 0.007339132484048605, + 0.0063713472336530685, + 0.0038773026317358017, + 0.04678216949105263, + 0.04050049930810928, + 0.06923744827508926, + 0.031153937801718712, + 0.04417388513684273, + 0.055533017963171005, + 0.04991057142615318, + 0.05769720673561096, + 0.002508767182007432, + 0.0018165509682148695, + 0.001961443806067109, + 0.0019101640209555626, + 0.0016414605779573321, + 0.0022459926549345255, + 0.003209809772670269, + 0.0016593781765550375, + 0.07094696164131165, + 0.03806925565004349, + 0.03621203452348709, + 0.03168337047100067, + 0.031793609261512756, + 0.06021731346845627, + 0.05025942251086235, + 0.055856797844171524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.0009248746791854501 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up focusing on a weathered man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a coarse robe that is faded but retains a deep cyan color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man raise his right arm to shield his mouth and nose, leaving a narrow gap between his fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by rich, saturated ochre and orange tones with a cyan robe providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with Steve McCurry's iconic photography style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704962491989136 + ], + "opposite_logprob_tokens": [ + -2.127049684524536 + ], + "logprob_reward": -0.12704962491989136, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704962491989136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10388360172510147, + 0.04993278905749321, + 0.0513143315911293, + 0.06622422486543655, + 0.08180893212556839, + 0.16300833225250244, + 0.2466021329164505, + 0.1382787823677063, + 0.042724814265966415, + 0.056221961975097656 + ], + "credit_map": [ + 0.019632870331406593, + 0.009240806102752686, + 0.018085375428199768, + 0.012217937968671322, + 0.010788156650960445, + 0.010612450540065765, + 0.010003537870943546, + 0.013302484527230263, + 0.00664832117035985, + 0.0045365761034190655, + 0.006553092505782843, + 0.004373730625957251, + 0.00790605042129755, + 0.007851374335587025, + 0.006234506610780954, + 0.00582914799451828, + 0.010818714275956154, + 0.004259276669472456, + 0.0058881002478301525, + 0.003464328357949853, + 0.0061212582513689995, + 0.008435539901256561, + 0.004075771197676659, + 0.008251352235674858, + 0.005863072816282511, + 0.007886208593845367, + 0.003991460427641869, + 0.006104243453592062, + 0.019081026315689087, + 0.009445973671972752, + 0.007323646452277899, + 0.006528604310005903, + 0.005695193540304899, + 0.012111004441976547, + 0.004349935799837112, + 0.008444733917713165, + 0.023571360856294632, + 0.01257211621850729, + 0.0073129283264279366, + 0.007751678582280874, + 0.017477184534072876, + 0.023984920233488083, + 0.012943025678396225, + 0.018997760489583015, + 0.03719330206513405, + 0.02349928393959999, + 0.015337157063186169, + 0.013575726188719273, + 0.02833348512649536, + 0.021843072026968002, + 0.023614611476659775, + 0.037893231958150864, + 0.047063715755939484, + 0.03535544127225876, + 0.030749723315238953, + 0.02174888178706169, + 0.03715335577726364, + 0.013439991511404514, + 0.012631789781153202, + 0.012300848960876465, + 0.014806525781750679, + 0.023321278393268585, + 0.011685441248118877, + 0.012939567677676678, + 0.00553205655887723, + 0.004989687819033861, + 0.0045851352624595165, + 0.005416125059127808, + 0.005636805202811956, + 0.00576189998537302, + 0.005557206459343433, + 0.005245906766504049, + 0.006704876199364662, + 0.006212420761585236, + 0.006818186026066542, + 0.006875771097838879, + 0.0064348322339355946, + 0.008268970996141434, + 0.0076641421765089035, + 0.007242776453495026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.12704962491989136 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_10/index_tensor([6542, 6542])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_10/index_tensor([6542, 6542])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6977e4f970a7a1494102aa58247a8120b247f6d4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_10/index_tensor([6542, 6542])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the distinctive aesthetic of traditional Japanese Ukiyo-e woodblock prints, the scene opens with an extreme close-up of an assault rifle, its metallic form rendered in flat black ink and strong outlines. The camera remains still, allowing the eye to linger on the intricate decoration: delicate pink cherry blossoms, their petals and stems depicted with elegant lines and soft, unshaded colors, artfully arranged across the weapon's butt and barrel. The camera then pans across the sleek, steady back of the doll, revealing the rifle clutched by a figure clad in a deep indigo wingsuit, the fabric displaying stylized folds and catching the unseen wind. As the camera continues to pan away, the towering figure's full form is revealed, their body maintaining a tense, horizontal glide. The doll's back accelerates, transitioning into a sweeping wide shot, revealing a stunning backdrop: the figure, a lone silhouette, silhouetted against a vast sky filled with swirling, stylized clouds rendered in planes of cream and pale gold. Below, jagged, blue-hued mountains rise from billowing white mist, and the entire composition, flattened in perspective, is defined by the bold lines and harmonious color palette characteristic of Ukiyo-e art, creating a surreal and contemplative atmosphere that juxtaposes serene natural beauty with modern military imagery.", + "seed": 9456, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_10/index_6542_seed_9456_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01163348841946572, + "vlm_logprob": -0.01163348841946572, + "vlm_yes_no_margin": 4.874999952316284, + "vlm_token_credit": -0.01163348841946572, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020675843581557274 + ], + "opposite_logprob_tokens": [ + -4.0206756591796875 + ], + "logprob_reward": -0.020675843581557274, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.020675843581557274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06087061017751694, + 0.3361644744873047, + 0.020027900114655495, + 0.07687275856733322, + 0.09695743769407272, + 0.1472540646791458, + 0.061741627752780914, + 0.13993316888809204, + 0.021444285288453102, + 0.038733746856451035 + ], + "credit_map": [ + 0.00659966142848134, + 0.006943140644580126, + 0.0065253241918981075, + 0.005549653433263302, + 0.00957934744656086, + 0.006875282619148493, + 0.012094822712242603, + 0.006703374441713095, + 0.03088526427745819, + 0.06222616881132126, + 0.03859764710068703, + 0.029715238139033318, + 0.052178580313920975, + 0.03260358050465584, + 0.047247447073459625, + 0.04271052032709122, + 0.002240063389763236, + 0.001661090413108468, + 0.003181080799549818, + 0.0023107798770070076, + 0.002510824939236045, + 0.0033326351549476385, + 0.002112595597282052, + 0.0026788306422531605, + 0.007078929804265499, + 0.006600527558475733, + 0.014482667669653893, + 0.006503929849714041, + 0.00822287704795599, + 0.01031514909118414, + 0.012111732736229897, + 0.011556941084563732, + 0.015707729384303093, + 0.013313519768416882, + 0.013099676929414272, + 0.013471197336912155, + 0.00936905201524496, + 0.010290122590959072, + 0.009101531468331814, + 0.012604599818587303, + 0.011086419224739075, + 0.011690784245729446, + 0.037328410893678665, + 0.03010975569486618, + 0.006616305094212294, + 0.014977394603192806, + 0.015563245862722397, + 0.019881736487150192, + 0.005315573886036873, + 0.004251232836395502, + 0.007021383382380009, + 0.006695881951600313, + 0.006414053961634636, + 0.013760985806584358, + 0.009116359055042267, + 0.009166152216494083, + 0.01681557111442089, + 0.00885170791298151, + 0.017333026975393295, + 0.02712920494377613, + 0.012639117427170277, + 0.01993248611688614, + 0.018396390601992607, + 0.01883564703166485, + 0.004149296786636114, + 0.001986265880987048, + 0.0036280308850109577, + 0.0022717781830579042, + 0.0009800115367397666, + 0.0029915517661720514, + 0.0024871695786714554, + 0.0029501779936254025, + 0.009432220831513405, + 0.003386753611266613, + 0.005210389383137226, + 0.0035626846365630627, + 0.002433415036648512, + 0.006765921600162983, + 0.00468468526378274, + 0.0032576769590377808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.020675843581557274 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596204325556755 + ], + "opposite_logprob_tokens": [ + -3.8955962657928467 + ], + "logprob_reward": -0.020596204325556755, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596204325556755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11338960379362106, + 0.12003940343856812, + 0.05266434699296951, + 0.24371227622032166, + 0.08239707350730896, + 0.06911581009626389, + 0.0682770237326622, + 0.1737060546875, + 0.023645170032978058, + 0.053053248673677444 + ], + "credit_map": [ + 0.011190056800842285, + 0.015801087021827698, + 0.010378711856901646, + 0.011216075159609318, + 0.016258254647254944, + 0.011606655083596706, + 0.01740160770714283, + 0.01953716017305851, + 0.010539259761571884, + 0.027190234512090683, + 0.01441450696438551, + 0.009854469448328018, + 0.014673843048512936, + 0.011741062626242638, + 0.01596236415207386, + 0.015663666650652885, + 0.005107577424496412, + 0.006210531108081341, + 0.00699763186275959, + 0.006789876613765955, + 0.007260454352945089, + 0.006728911306709051, + 0.005780297331511974, + 0.007789067458361387, + 0.022570842877030373, + 0.025187265127897263, + 0.0457720123231411, + 0.0178154818713665, + 0.030195798724889755, + 0.03591044619679451, + 0.03325781971216202, + 0.03300260379910469, + 0.01560905296355486, + 0.00907361600548029, + 0.012395061552524567, + 0.010816353373229504, + 0.008923600427806377, + 0.011239005252718925, + 0.005575083661824465, + 0.008765299804508686, + 0.007268075831234455, + 0.007952108979225159, + 0.01528488751500845, + 0.01223666686564684, + 0.003811310976743698, + 0.006838311906903982, + 0.007915080524981022, + 0.0078093623742461205, + 0.006823537405580282, + 0.005115463398396969, + 0.011718187481164932, + 0.010110316798090935, + 0.005499382968991995, + 0.010971183888614178, + 0.010197418741881847, + 0.007841535843908787, + 0.02260766737163067, + 0.01011437363922596, + 0.02133440226316452, + 0.029120266437530518, + 0.013162953779101372, + 0.027309143915772438, + 0.0216529443860054, + 0.02840430662035942, + 0.005268110893666744, + 0.0030631902627646923, + 0.0029161993879824877, + 0.002403462305665016, + 0.0019143992103636265, + 0.003589947707951069, + 0.002348869340494275, + 0.0021409899927675724, + 0.011460681445896626, + 0.006192339118570089, + 0.008025499060750008, + 0.004167702980339527, + 0.004123772028833628, + 0.007818253710865974, + 0.004454188514500856, + 0.0068108136765658855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "selected_reward": -0.020596204325556755 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011798333143815398 + ], + "opposite_logprob_tokens": [ + -6.7511796951293945 + ], + "logprob_reward": -0.0011798333143815398, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011798333143815398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05029774829745293, + 0.27004244923591614, + 0.025817805901169777, + 0.0340004600584507, + 0.1673918217420578, + 0.15814897418022156, + 0.07273412495851517, + 0.11052440106868744, + 0.03784332424402237, + 0.07319892942905426 + ], + "credit_map": [ + 0.004812592174857855, + 0.0035720381420105696, + 0.00493749650195241, + 0.0060670762322843075, + 0.008057028986513615, + 0.007544881198555231, + 0.008596856147050858, + 0.00670977495610714, + 0.02498917654156685, + 0.051982004195451736, + 0.026681480929255486, + 0.022821402177214622, + 0.04317796975374222, + 0.02198483981192112, + 0.038132622838020325, + 0.04027297720313072, + 0.0025944304652512074, + 0.002243561204522848, + 0.0027397458907216787, + 0.004366259556263685, + 0.004052616655826569, + 0.003137922612950206, + 0.0026699199806898832, + 0.004013348370790482, + 0.0035328518133610487, + 0.0031652457546442747, + 0.004525409080088139, + 0.003405551426112652, + 0.004025894217193127, + 0.004794897977262735, + 0.005398639943450689, + 0.005151968915015459, + 0.030465293675661087, + 0.023463770747184753, + 0.027762753888964653, + 0.021799210458993912, + 0.015363314189016819, + 0.016825636848807335, + 0.012386397458612919, + 0.01932545378804207, + 0.015844479203224182, + 0.01358780637383461, + 0.04496404528617859, + 0.03561386466026306, + 0.005194879602640867, + 0.011090865358710289, + 0.011383351869881153, + 0.02046968787908554, + 0.006665554828941822, + 0.004232399165630341, + 0.01293220091611147, + 0.01268787495791912, + 0.006983113940805197, + 0.010357571765780449, + 0.010893098078668118, + 0.007982316426932812, + 0.010500998236238956, + 0.00664144242182374, + 0.01587156020104885, + 0.02133374847471714, + 0.010007222183048725, + 0.016711879521608353, + 0.012897930108010769, + 0.016559619456529617, + 0.008264056406915188, + 0.002853234065696597, + 0.007195949088782072, + 0.0029846476390957832, + 0.0018753843614831567, + 0.006262083537876606, + 0.003926506731659174, + 0.004481458570808172, + 0.008488690480589867, + 0.01126947347074747, + 0.013685294426977634, + 0.006783842574805021, + 0.008925375528633595, + 0.009008944034576416, + 0.00638125604018569, + 0.008656053803861141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.0011798333143815398 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009772789664566517 + ], + "opposite_logprob_tokens": [ + -4.634772777557373 + ], + "logprob_reward": -0.009772789664566517, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009772789664566517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20008203387260437, + 0.1513654887676239, + 0.05504589527845383, + 0.14988994598388672, + 0.07934587448835373, + 0.12212143838405609, + 0.07280271500349045, + 0.056300435215234756, + 0.06259264051914215, + 0.05045358091592789 + ], + "credit_map": [ + 0.021978991106152534, + 0.025689804926514626, + 0.020029449835419655, + 0.01536524761468172, + 0.041189879179000854, + 0.027804216369986534, + 0.02493331953883171, + 0.023091131821274757, + 0.015286650508642197, + 0.020028037950396538, + 0.014457632787525654, + 0.013783484697341919, + 0.03397474065423012, + 0.016273485496640205, + 0.019970908761024475, + 0.017590545117855072, + 0.004470162559300661, + 0.004817907698452473, + 0.006699686869978905, + 0.007690977305173874, + 0.007042266894131899, + 0.010209184139966965, + 0.0058395168744027615, + 0.008276190608739853, + 0.013402792625129223, + 0.012525498867034912, + 0.02401570975780487, + 0.014025978744029999, + 0.016118517145514488, + 0.017587564885616302, + 0.032109227031469345, + 0.020104652270674706, + 0.013047050684690475, + 0.00945975724607706, + 0.01602794975042343, + 0.008701148442924023, + 0.008290688507258892, + 0.010025636292994022, + 0.004977399483323097, + 0.008816246874630451, + 0.014358661137521267, + 0.014300274662673473, + 0.027841471135616302, + 0.02216438204050064, + 0.006836604326963425, + 0.013390952721238136, + 0.009707932360470295, + 0.013521159999072552, + 0.008129339665174484, + 0.005112252663820982, + 0.01144474558532238, + 0.0077356197871267796, + 0.00590900843963027, + 0.012950676493346691, + 0.011660684831440449, + 0.00986038614064455, + 0.009252187795937061, + 0.0037067378871142864, + 0.005043399054557085, + 0.005420192144811153, + 0.005712379235774279, + 0.00895354151725769, + 0.010014082305133343, + 0.008197915740311146, + 0.012148115783929825, + 0.0058827269822359085, + 0.011247621849179268, + 0.005946981720626354, + 0.0045439680106937885, + 0.010362503118813038, + 0.006589583586901426, + 0.005871141795068979, + 0.00864977203309536, + 0.006134354509413242, + 0.0084255151450634, + 0.00477907108142972, + 0.004240740090608597, + 0.007398583460599184, + 0.004894325044006109, + 0.0059312209486961365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.009772789664566517 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005942771211266518 + ], + "opposite_logprob_tokens": [ + -5.1309428215026855 + ], + "logprob_reward": -0.005942771211266518, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942771211266518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05409767106175423, + 0.05639145150780678, + 0.017804760485887527, + 0.022235959768295288, + 0.052313681691884995, + 0.050065916031599045, + 0.049081120640039444, + 0.2363397777080536, + 0.0890144631266594, + 0.3726552426815033 + ], + "credit_map": [ + 0.005531091243028641, + 0.004982243292033672, + 0.00461608124896884, + 0.005617675371468067, + 0.008221650496125221, + 0.005448069889098406, + 0.010949214920401573, + 0.008731642737984657, + 0.004849247634410858, + 0.010679387487471104, + 0.00800204835832119, + 0.004960728343576193, + 0.005870234686881304, + 0.0059283580631017685, + 0.00845907349139452, + 0.0076423706486821175, + 0.0017301118932664394, + 0.0020326697267591953, + 0.0026450534351170063, + 0.002320094732567668, + 0.0018834350630640984, + 0.002423257101327181, + 0.0020325486548244953, + 0.0027375894133001566, + 0.0025789416395127773, + 0.0025410845410078764, + 0.0028989678248763084, + 0.002034547971561551, + 0.0025429055094718933, + 0.001960431458428502, + 0.004564355593174696, + 0.0031147259287536144, + 0.007280431222170591, + 0.006218498107045889, + 0.006865243427455425, + 0.005652772728353739, + 0.0072080218233168125, + 0.007165288552641869, + 0.005271464120596647, + 0.006651959847658873, + 0.004517658147960901, + 0.00637223944067955, + 0.010624830611050129, + 0.009490663185715675, + 0.002618903061375022, + 0.004832180216908455, + 0.005672048777341843, + 0.005937391892075539, + 0.00440993020310998, + 0.002682066522538662, + 0.008176587522029877, + 0.007017772179096937, + 0.004691597074270248, + 0.009506218135356903, + 0.006417587399482727, + 0.006179361138492823, + 0.03559492528438568, + 0.012087353505194187, + 0.02288706973195076, + 0.03594077378511429, + 0.01784566417336464, + 0.03753828629851341, + 0.04071193188428879, + 0.03373377397656441, + 0.014077256433665752, + 0.008095302619040012, + 0.01791566051542759, + 0.008485169149935246, + 0.006290305405855179, + 0.012934211641550064, + 0.011229028925299644, + 0.00998752936720848, + 0.05039634928107262, + 0.05239955708384514, + 0.06431704014539719, + 0.034431178122758865, + 0.04517504945397377, + 0.05060289055109024, + 0.03428740054368973, + 0.04104577377438545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.005942771211266518 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the distinctive aesthetic of traditional Japanese Ukiyo-e woodblock prints, the scene opens with an extreme close-up of an assault rifle, its metallic form rendered in flat black ink and strong outlines. The camera remains still, allowing the eye to linger on the intricate decoration: delicate pink cherry blossoms, their petals and stems depicted with elegant lines and soft, unshaded colors, artfully arranged across the weapon's butt and barrel. The camera then pans across the sleek, steady back of the doll, revealing the rifle clutched by a figure clad in a deep indigo wingsuit, the fabric displaying stylized folds and catching the unseen wind. As the camera continues to pan away, the towering figure's full form is revealed, their body maintaining a tense, horizontal glide. The doll's back accelerates, transitioning into a sweeping wide shot, revealing a stunning backdrop: the figure, a lone silhouette, silhouetted against a vast sky filled with swirling, stylized clouds rendered in planes of cream and pale gold. Below, jagged, blue-hued mountains rise from billowing white mist, and the entire composition, flattened in perspective, is defined by the bold lines and harmonious color palette characteristic of Ukiyo-e art, creating a surreal and contemplative atmosphere that juxtaposes serene natural beauty with modern military imagery.", + "seed": 9457, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_10/index_6542_seed_9457_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023755778232589363, + "vlm_logprob": -0.023755778232589363, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.023755778232589363, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026099780574440956 + ], + "opposite_logprob_tokens": [ + -3.776099681854248 + ], + "logprob_reward": -0.026099780574440956, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.026099780574440956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03760959208011627, + 0.022422736510634422, + 0.026788262650370598, + 0.03954068943858147, + 0.061080578714609146, + 0.07338455319404602, + 0.31096941232681274, + 0.3824026882648468, + 0.014857884496450424, + 0.030943650752305984 + ], + "credit_map": [ + 0.004018435254693031, + 0.004549164790660143, + 0.005819502752274275, + 0.004791625775396824, + 0.003277938347309828, + 0.006020893808454275, + 0.004747792147099972, + 0.004384241066873074, + 0.0036433034110814333, + 0.003372941166162491, + 0.0019250118639320135, + 0.0033644153736531734, + 0.0019308545161038637, + 0.0030961858574301004, + 0.0023894791956990957, + 0.0027005458250641823, + 0.0021156107541173697, + 0.0031213220208883286, + 0.0017191799124702811, + 0.0019373127724975348, + 0.0032727874349802732, + 0.006686022505164146, + 0.0030626505613327026, + 0.004873377736657858, + 0.004917938727885485, + 0.0032588071189820766, + 0.006150323431938887, + 0.005315576680004597, + 0.004169994033873081, + 0.0034212740138173103, + 0.006612877361476421, + 0.005693895276635885, + 0.00844205729663372, + 0.0076526920311152935, + 0.006973794661462307, + 0.006078038364648819, + 0.007683881558477879, + 0.006655346602201462, + 0.009103736840188503, + 0.008491030894219875, + 0.008920680731534958, + 0.008613036945462227, + 0.009515678510069847, + 0.00833282619714737, + 0.007775153499096632, + 0.00829251017421484, + 0.01344623975455761, + 0.0084884287789464, + 0.05586119368672371, + 0.06364979594945908, + 0.03730154037475586, + 0.037201449275016785, + 0.018476152792572975, + 0.05842994526028633, + 0.021872609853744507, + 0.018176691606640816, + 0.07284106314182281, + 0.06861622631549835, + 0.03223500773310661, + 0.04598274827003479, + 0.04099859669804573, + 0.05612416937947273, + 0.03404177352786064, + 0.03156309574842453, + 0.0017490229802206159, + 0.002552444115281105, + 0.0016132862074300647, + 0.0016540945507586002, + 0.0013854865683242679, + 0.003564739366993308, + 0.0008820585208013654, + 0.0014567525358870625, + 0.00488916365429759, + 0.006318311206996441, + 0.0022528921253979206, + 0.002160552190616727, + 0.0034651518799364567, + 0.007902903482317924, + 0.001522808917798102, + 0.0024318646173924208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.026099780574440956 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004015892744064 + ], + "opposite_logprob_tokens": [ + -2.695040225982666 + ], + "logprob_reward": -0.07004015892744064, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004015892744064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08953163772821426, + 0.05016025900840759, + 0.04522662237286568, + 0.06271057575941086, + 0.07006583362817764, + 0.06949403136968613, + 0.277487576007843, + 0.2741367518901825, + 0.03643425554037094, + 0.024752425029873848 + ], + "credit_map": [ + 0.0075845662504434586, + 0.0074813407845795155, + 0.013353283517062664, + 0.010323142632842064, + 0.01773073710501194, + 0.010095628909766674, + 0.009614352136850357, + 0.013348586857318878, + 0.005223877727985382, + 0.006293391343206167, + 0.004893499892205, + 0.008232501335442066, + 0.00699559785425663, + 0.006732533685863018, + 0.006394871044903994, + 0.005393986590206623, + 0.004252104554325342, + 0.004435447044670582, + 0.003679882735013962, + 0.0038615979719907045, + 0.004383341409265995, + 0.01055785734206438, + 0.005826030857861042, + 0.008230362087488174, + 0.009625234641134739, + 0.004486854188144207, + 0.009032166562974453, + 0.00906608160585165, + 0.004749802406877279, + 0.004714909940958023, + 0.007806925103068352, + 0.013228601776063442, + 0.009102119132876396, + 0.009482679888606071, + 0.0068555958569049835, + 0.008256146684288979, + 0.007142629940062761, + 0.004817325156182051, + 0.014822899363934994, + 0.009586443193256855, + 0.009012130089104176, + 0.008546593599021435, + 0.01004816684871912, + 0.006924543064087629, + 0.006744899321347475, + 0.007630567066371441, + 0.011799165047705173, + 0.008787970058619976, + 0.048611853271722794, + 0.059270262718200684, + 0.029859628528356552, + 0.03229028359055519, + 0.01795693300664425, + 0.05435286834836006, + 0.018923042342066765, + 0.016222719103097916, + 0.051516350358724594, + 0.04705236479640007, + 0.02396286465227604, + 0.03270420432090759, + 0.02910584770143032, + 0.04086412489414215, + 0.025701461359858513, + 0.023229533806443214, + 0.006701507139950991, + 0.006418876815587282, + 0.002504894742742181, + 0.0025744466111063957, + 0.0037490553222596645, + 0.008829092606902122, + 0.0023463519755750895, + 0.003310032654553652, + 0.0034778122790157795, + 0.00500568188726902, + 0.0016135983169078827, + 0.0016801763558760285, + 0.002822350012138486, + 0.006463600788265467, + 0.0015301280654966831, + 0.0021590746473520994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.07004015892744064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002490278100594878 + ], + "opposite_logprob_tokens": [ + -6.002490043640137 + ], + "logprob_reward": -0.002490278100594878, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002490278100594878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07913040369749069, + 0.08011777698993683, + 0.12490403652191162, + 0.11616590619087219, + 0.1148717924952507, + 0.07080253958702087, + 0.15040475130081177, + 0.17646954953670502, + 0.05399302393198013, + 0.033140186220407486 + ], + "credit_map": [ + 0.009282546117901802, + 0.009019462391734123, + 0.012997698038816452, + 0.009635765105485916, + 0.007369007915258408, + 0.009311674162745476, + 0.010823818854987621, + 0.010690436698496342, + 0.011304061859846115, + 0.010044260881841183, + 0.008697410114109516, + 0.0133523466065526, + 0.006553801707923412, + 0.008778071030974388, + 0.012123129330575466, + 0.009264703840017319, + 0.011225837282836437, + 0.012864825315773487, + 0.007721807342022657, + 0.009262698702514172, + 0.013049311935901642, + 0.0331832654774189, + 0.015460275113582611, + 0.02213602513074875, + 0.0132206492125988, + 0.009558410383760929, + 0.01328563317656517, + 0.012827331200242043, + 0.012173867784440517, + 0.013379299081861973, + 0.018126247450709343, + 0.02359447441995144, + 0.016347285360097885, + 0.012114773504436016, + 0.017441309988498688, + 0.011661149561405182, + 0.015469863079488277, + 0.009679787792265415, + 0.02034103497862816, + 0.011816604994237423, + 0.008135988377034664, + 0.005131990648806095, + 0.012930192984640598, + 0.008202943950891495, + 0.009351140819489956, + 0.005353223532438278, + 0.011707911267876625, + 0.009989148005843163, + 0.027746262028813362, + 0.022724438458681107, + 0.020996233448386192, + 0.01891801878809929, + 0.010627971962094307, + 0.02351274900138378, + 0.01563657633960247, + 0.010242507793009281, + 0.033050552010536194, + 0.029538320377469063, + 0.015728384256362915, + 0.023387320339679718, + 0.01936188153922558, + 0.025971917435526848, + 0.013912349008023739, + 0.015518845058977604, + 0.007375118788331747, + 0.008470194414258003, + 0.0039900802075862885, + 0.00473251985386014, + 0.007460240740329027, + 0.014111504890024662, + 0.003359382040798664, + 0.004493985790759325, + 0.004410493653267622, + 0.004622756969183683, + 0.003389096586033702, + 0.0021740137599408627, + 0.005833396688103676, + 0.007008494809269905, + 0.00284721702337265, + 0.0028547209221869707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.002490278100594878 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605839654803276 + ], + "opposite_logprob_tokens": [ + -4.141058444976807 + ], + "logprob_reward": -0.01605839654803276, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605839654803276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05242918059229851, + 0.05659259855747223, + 0.03321100026369095, + 0.07360391318798065, + 0.039629194885492325, + 0.069023996591568, + 0.2604939043521881, + 0.28243863582611084, + 0.055014435201883316, + 0.07756312936544418 + ], + "credit_map": [ + 0.005667936056852341, + 0.005187390837818384, + 0.008621586486697197, + 0.006378444842994213, + 0.0064417896792292595, + 0.006718046963214874, + 0.0072070700116455555, + 0.006206914782524109, + 0.006555636879056692, + 0.007082525175064802, + 0.005011275876313448, + 0.010069062002003193, + 0.006684355903416872, + 0.006692925468087196, + 0.00761581351980567, + 0.006881005130708218, + 0.0026429197750985622, + 0.004500188399106264, + 0.002469733590260148, + 0.0027230922132730484, + 0.004960070829838514, + 0.005748494062572718, + 0.003912010695785284, + 0.006254491861909628, + 0.006287938449531794, + 0.00783004704862833, + 0.00836441945284605, + 0.010303743183612823, + 0.009851128794252872, + 0.007773729972541332, + 0.015714364126324654, + 0.007478549610823393, + 0.0067770229652523994, + 0.005175946746021509, + 0.0046167559921741486, + 0.003999036271125078, + 0.004571638070046902, + 0.0036376884672790766, + 0.005938698537647724, + 0.004912406671792269, + 0.008623946458101273, + 0.007526643108576536, + 0.009753978811204433, + 0.008036088198423386, + 0.006751097273081541, + 0.007707979064434767, + 0.010979118756949902, + 0.009645137004554272, + 0.04771067574620247, + 0.049310922622680664, + 0.03296373784542084, + 0.03263522684574127, + 0.016741832718253136, + 0.04453682154417038, + 0.02099553681910038, + 0.015599124133586884, + 0.05326292663812637, + 0.048581723123788834, + 0.024235965684056282, + 0.03420909494161606, + 0.03022979386150837, + 0.040782488882541656, + 0.026625102385878563, + 0.024511532858014107, + 0.008109518326818943, + 0.010943197645246983, + 0.004158381372690201, + 0.003627472324296832, + 0.0068794358521699905, + 0.0125989755615592, + 0.004253019113093615, + 0.004444433841854334, + 0.012611749581992626, + 0.01443655788898468, + 0.0053030867129564285, + 0.004834635183215141, + 0.009805840440094471, + 0.01953948847949505, + 0.004611278418451548, + 0.006420500576496124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.01605839654803276 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004090277012437582 + ], + "opposite_logprob_tokens": [ + -5.504090309143066 + ], + "logprob_reward": -0.004090277012437582, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004090277012437582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06550586223602295, + 0.05173879489302635, + 0.06321728229522705, + 0.10682041943073273, + 0.0651116818189621, + 0.14886440336704254, + 0.13776347041130066, + 0.08194480836391449, + 0.18389037251472473, + 0.09514296054840088 + ], + "credit_map": [ + 0.007310297805815935, + 0.008388807997107506, + 0.009368842467665672, + 0.008690289221704006, + 0.008644761517643929, + 0.010512723587453365, + 0.005217718426138163, + 0.007372413761913776, + 0.006776807364076376, + 0.006229031831026077, + 0.005222698207944632, + 0.006293631624430418, + 0.007296876050531864, + 0.006762182805687189, + 0.007295113056898117, + 0.005862452555447817, + 0.007613229565322399, + 0.007814325392246246, + 0.00616313936188817, + 0.007291713263839483, + 0.006034980993717909, + 0.010357463732361794, + 0.009728288277983665, + 0.008214139379560947, + 0.017443960532546043, + 0.007979271002113819, + 0.014354857616126537, + 0.011707858182489872, + 0.00766074750572443, + 0.007548334077000618, + 0.01488487794995308, + 0.02524050883948803, + 0.01229399535804987, + 0.006390321999788284, + 0.009119467809796333, + 0.005265661980956793, + 0.0063127134926617146, + 0.00565686309710145, + 0.009634746238589287, + 0.010437906719744205, + 0.03374733030796051, + 0.015203988179564476, + 0.02736526168882847, + 0.015177546069025993, + 0.009652528911828995, + 0.01296087447553873, + 0.010426430962979794, + 0.024330437183380127, + 0.029281772673130035, + 0.02152291126549244, + 0.021228943020105362, + 0.016558147966861725, + 0.008686782792210579, + 0.02048567309975624, + 0.008522996678948402, + 0.011476240120828152, + 0.01229044422507286, + 0.013777799904346466, + 0.005483527667820454, + 0.007512614596635103, + 0.008676758036017418, + 0.02147217094898224, + 0.0061194696463644505, + 0.006612015422433615, + 0.02984681911766529, + 0.02882177010178566, + 0.015665313228964806, + 0.021933097392320633, + 0.016988009214401245, + 0.04188423231244087, + 0.013464069925248623, + 0.015287037938833237, + 0.014946844428777695, + 0.015552748925983906, + 0.009260841645300388, + 0.008058393374085426, + 0.01534091867506504, + 0.016965074464678764, + 0.006706009618937969, + 0.008312125690281391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.004090277012437582 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the distinctive aesthetic of traditional Japanese Ukiyo-e woodblock prints, the scene opens with an extreme close-up of an assault rifle, its metallic form rendered in flat black ink and strong outlines. The camera remains still, allowing the eye to linger on the intricate decoration: delicate pink cherry blossoms, their petals and stems depicted with elegant lines and soft, unshaded colors, artfully arranged across the weapon's butt and barrel. The camera then pans across the sleek, steady back of the doll, revealing the rifle clutched by a figure clad in a deep indigo wingsuit, the fabric displaying stylized folds and catching the unseen wind. As the camera continues to pan away, the towering figure's full form is revealed, their body maintaining a tense, horizontal glide. The doll's back accelerates, transitioning into a sweeping wide shot, revealing a stunning backdrop: the figure, a lone silhouette, silhouetted against a vast sky filled with swirling, stylized clouds rendered in planes of cream and pale gold. Below, jagged, blue-hued mountains rise from billowing white mist, and the entire composition, flattened in perspective, is defined by the bold lines and harmonious color palette characteristic of Ukiyo-e art, creating a surreal and contemplative atmosphere that juxtaposes serene natural beauty with modern military imagery.", + "seed": 9458, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_10/index_6542_seed_9458_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13071822170168163, + "vlm_logprob": -0.13071822170168163, + "vlm_yes_no_margin": 3.499999976158142, + "vlm_token_credit": -0.13071822170168163, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4805670380592346 + ], + "opposite_logprob_tokens": [ + -0.9805670380592346 + ], + "logprob_reward": -0.4805670380592346, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4805670380592346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21559865772724152, + 0.04654325544834137, + 0.04655616730451584, + 0.07537931948900223, + 0.04042121395468712, + 0.1095162183046341, + 0.12075142562389374, + 0.17097117006778717, + 0.12311854958534241, + 0.051144007593393326 + ], + "credit_map": [ + 0.025888463482260704, + 0.05078955367207527, + 0.021624909713864326, + 0.015936696901917458, + 0.01784374751150608, + 0.056089408695697784, + 0.01623349077999592, + 0.011192389763891697, + 0.009080078452825546, + 0.0081424405798316, + 0.0033867759630084038, + 0.003786505199968815, + 0.005498155485838652, + 0.00555498618632555, + 0.006034448277205229, + 0.005059863906353712, + 0.008614972233772278, + 0.006165742874145508, + 0.005317140370607376, + 0.005541949532926083, + 0.006844739429652691, + 0.0046760751865804195, + 0.005264220759272575, + 0.0041313255205750465, + 0.01715841516852379, + 0.007134435698390007, + 0.01193980872631073, + 0.010067245922982693, + 0.006454060319811106, + 0.012462673708796501, + 0.006241329945623875, + 0.003921352792531252, + 0.004781448282301426, + 0.009373863227665424, + 0.0035789222456514835, + 0.006648357026278973, + 0.0046844263561069965, + 0.004430669359862804, + 0.004053231328725815, + 0.002870295662432909, + 0.014964736998081207, + 0.015952449291944504, + 0.011274220421910286, + 0.011308985762298107, + 0.015207849442958832, + 0.015928953886032104, + 0.016752520576119423, + 0.008126499131321907, + 0.029168950393795967, + 0.010392903350293636, + 0.011565557681024075, + 0.012290614657104015, + 0.011290237307548523, + 0.01805267110466957, + 0.018918579444289207, + 0.009071915410459042, + 0.028946515172719955, + 0.018291421234607697, + 0.014423339627683163, + 0.020371088758111, + 0.017873167991638184, + 0.021703774109482765, + 0.03883669525384903, + 0.010525156743824482, + 0.01900801621377468, + 0.013387714512646198, + 0.012947642244398594, + 0.01864749751985073, + 0.009076792746782303, + 0.018083535134792328, + 0.022173510864377022, + 0.00979384034872055, + 0.009026800282299519, + 0.005758913699537516, + 0.006031948607414961, + 0.004407946020364761, + 0.0061633409932255745, + 0.008404720574617386, + 0.006532954052090645, + 0.004817384760826826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.4805670380592346 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283621311187744 + ], + "opposite_logprob_tokens": [ + -2.017836093902588 + ], + "logprob_reward": -0.14283621311187744, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14283621311187744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3095085918903351, + 0.05015839636325836, + 0.04739003628492355, + 0.11483306437730789, + 0.08105166256427765, + 0.040811579674482346, + 0.04586514085531235, + 0.10478956997394562, + 0.10289519280195236, + 0.1026967465877533 + ], + "credit_map": [ + 0.041374024003744125, + 0.06075895205140114, + 0.036024220287799835, + 0.029729342088103294, + 0.027491681277751923, + 0.05872219055891037, + 0.029236983507871628, + 0.026171201840043068, + 0.008578145876526833, + 0.009429424069821835, + 0.0034211722668260336, + 0.0035511902533471584, + 0.006917108781635761, + 0.007435281295329332, + 0.00571239972487092, + 0.005113671533763409, + 0.01167487446218729, + 0.006632902659475803, + 0.005099873524159193, + 0.004877953324466944, + 0.0057593053206801414, + 0.004915973171591759, + 0.004009963478893042, + 0.004419189412146807, + 0.016763856634497643, + 0.012117126025259495, + 0.017390547320246696, + 0.014972400851547718, + 0.008613070473074913, + 0.022809753194451332, + 0.010735191404819489, + 0.011431117542088032, + 0.010108016431331635, + 0.015744008123874664, + 0.005470456555485725, + 0.01328455563634634, + 0.009941210970282555, + 0.012178318575024605, + 0.0095028942450881, + 0.004822203423827887, + 0.0057305539958179, + 0.0042980569414794445, + 0.005607252474874258, + 0.004078048747032881, + 0.005439423490315676, + 0.006372366566210985, + 0.005847967229783535, + 0.0034379097633063793, + 0.009880064986646175, + 0.003869040869176388, + 0.005636224057525396, + 0.005304680205881596, + 0.004247267730534077, + 0.007348041050136089, + 0.00512143038213253, + 0.00445839436724782, + 0.026967881247401237, + 0.010089577175676823, + 0.010933893732726574, + 0.007122673559933901, + 0.00880314502865076, + 0.0176764614880085, + 0.01306107733398676, + 0.010134859010577202, + 0.014801456592977047, + 0.010175247676670551, + 0.009522700682282448, + 0.019221646711230278, + 0.007296858821064234, + 0.014392192475497723, + 0.02072303742170334, + 0.006762055214494467, + 0.02321668155491352, + 0.009981486015021801, + 0.009853456169366837, + 0.013424372300505638, + 0.008565164171159267, + 0.016157027333974838, + 0.014798905700445175, + 0.006699658464640379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.14283621311187744 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028206822462379932 + ], + "opposite_logprob_tokens": [ + -5.8778204917907715 + ], + "logprob_reward": -0.0028206822462379932, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028206822462379932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2499842643737793, + 0.1213771402835846, + 0.07324276864528656, + 0.22717012465000153, + 0.04572256654500961, + 0.07308496534824371, + 0.07502807676792145, + 0.03466827794909477, + 0.05164175480604172, + 0.04808005690574646 + ], + "credit_map": [ + 0.03911762684583664, + 0.04889592155814171, + 0.02031121961772442, + 0.018274227157235146, + 0.021239781752228737, + 0.06362512707710266, + 0.020830608904361725, + 0.017689773812890053, + 0.02768915519118309, + 0.023047201335430145, + 0.009273468516767025, + 0.012131244875490665, + 0.013699089176952839, + 0.012692457996308804, + 0.013236765749752522, + 0.009607761166989803, + 0.01906069554388523, + 0.011205647140741348, + 0.008587964810431004, + 0.0065696630626916885, + 0.0063665942288935184, + 0.008099289610981941, + 0.00584623496979475, + 0.007506688125431538, + 0.0389152355492115, + 0.032879285514354706, + 0.043171606957912445, + 0.04464675113558769, + 0.014160930179059505, + 0.019055383279919624, + 0.02296687290072441, + 0.011374074034392834, + 0.005299095064401627, + 0.010318130254745483, + 0.004542411305010319, + 0.006701501086354256, + 0.0049974615685641766, + 0.005006395746022463, + 0.005436664912849665, + 0.0034209119621664286, + 0.012580153532326221, + 0.0119929863139987, + 0.005184377543628216, + 0.009417387656867504, + 0.009952440857887268, + 0.012058550491929054, + 0.00827962625771761, + 0.0036194517742842436, + 0.016884565353393555, + 0.006721677724272013, + 0.00903084222227335, + 0.0067217000760138035, + 0.005837057717144489, + 0.012994502671062946, + 0.008559050969779491, + 0.008278684690594673, + 0.006636483129113913, + 0.0038408953696489334, + 0.0031593020539730787, + 0.0030720022041350603, + 0.003675861982628703, + 0.004945473279803991, + 0.006756267044693232, + 0.0025819968432188034, + 0.006773512810468674, + 0.005075158551335335, + 0.0050252461805939674, + 0.008769185282289982, + 0.003843173384666443, + 0.007502240128815174, + 0.010710138827562332, + 0.003943103831261396, + 0.008359285071492195, + 0.005210921633988619, + 0.006194583140313625, + 0.004684819839894772, + 0.004201528616249561, + 0.007689023856073618, + 0.0061479974538087845, + 0.005591900087893009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0028206822462379932 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023276660591363907 + ], + "opposite_logprob_tokens": [ + -3.7732765674591064 + ], + "logprob_reward": -0.023276660591363907, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023276660591363907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23802226781845093, + 0.07848703116178513, + 0.12009941786527634, + 0.20297829806804657, + 0.03902006521821022, + 0.043242234736680984, + 0.11520172655582428, + 0.08387310802936554, + 0.056114561855793, + 0.022961268201470375 + ], + "credit_map": [ + 0.041479796171188354, + 0.049255479127168655, + 0.017962146550416946, + 0.014192144386470318, + 0.02593894489109516, + 0.05800028145313263, + 0.015161402523517609, + 0.016032099723815918, + 0.012129547074437141, + 0.01348439697176218, + 0.007454127073287964, + 0.008233644999563694, + 0.010484668426215649, + 0.011863515712320805, + 0.00825552362948656, + 0.006581616122275591, + 0.0200579222291708, + 0.012539491057395935, + 0.015665203332901, + 0.011636939831078053, + 0.014750147238373756, + 0.024192489683628082, + 0.010971607640385628, + 0.010285625234246254, + 0.0413912869989872, + 0.020635344088077545, + 0.028499659150838852, + 0.022720562294125557, + 0.014195882715284824, + 0.048936646431684494, + 0.014828684739768505, + 0.01177025306969881, + 0.004418353084474802, + 0.006033741403371096, + 0.0031701941043138504, + 0.0038089470472186804, + 0.00863785482943058, + 0.0057882219552993774, + 0.004008154850453138, + 0.0031546037644147873, + 0.00652414932847023, + 0.007101698778569698, + 0.004020784981548786, + 0.004742887802422047, + 0.006793045438826084, + 0.006362226326018572, + 0.00507954228669405, + 0.0026179065462201834, + 0.023690588772296906, + 0.009800160303711891, + 0.011944969184696674, + 0.008569419384002686, + 0.008376783691346645, + 0.019280821084976196, + 0.018212152644991875, + 0.01532684825360775, + 0.01660909689962864, + 0.008197608403861523, + 0.008003412745893002, + 0.007472597062587738, + 0.007302179932594299, + 0.01287869457155466, + 0.01568608172237873, + 0.007723444607108831, + 0.009427291341125965, + 0.005293636117130518, + 0.0044016726315021515, + 0.009122122079133987, + 0.004889168310910463, + 0.009976448491215706, + 0.008201390504837036, + 0.00480283796787262, + 0.0027782598044723272, + 0.002757817506790161, + 0.003432545345276594, + 0.00199705408886075, + 0.002486234763637185, + 0.003208035370334983, + 0.0033778240904212, + 0.002923500258475542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.023276660591363907 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004090514499694109 + ], + "opposite_logprob_tokens": [ + -5.504090309143066 + ], + "logprob_reward": -0.004090514499694109, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004090514499694109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24889948964118958, + 0.09231483191251755, + 0.03067898377776146, + 0.07837013900279999, + 0.02919071353971958, + 0.05686197429895401, + 0.07184381783008575, + 0.07353327423334122, + 0.06276402622461319, + 0.2555427551269531 + ], + "credit_map": [ + 0.04052314534783363, + 0.05052199587225914, + 0.025728123262524605, + 0.02119676023721695, + 0.03267282620072365, + 0.03726392611861229, + 0.023211902007460594, + 0.01778082177042961, + 0.020329242572188377, + 0.015220973640680313, + 0.007827606052160263, + 0.009262592531740665, + 0.012046374380588531, + 0.008802470751106739, + 0.011825891211628914, + 0.006999682635068893, + 0.00478444853797555, + 0.0027551911771297455, + 0.004233784507960081, + 0.003358573419973254, + 0.005189081653952599, + 0.0032951212488114834, + 0.003982796799391508, + 0.003079984337091446, + 0.013173820450901985, + 0.00956853199750185, + 0.013636444695293903, + 0.010537877678871155, + 0.006439104676246643, + 0.011000988073647022, + 0.007759463042020798, + 0.006253907456994057, + 0.0027719640638679266, + 0.004726352635771036, + 0.002766046207398176, + 0.004029156640172005, + 0.004861653782427311, + 0.003337177447974682, + 0.0034994357265532017, + 0.0031989256385713816, + 0.007264193147420883, + 0.006772340275347233, + 0.006691761314868927, + 0.0049558887258172035, + 0.005911507178097963, + 0.008032048121094704, + 0.011871634051203728, + 0.0053625996224582195, + 0.014602378010749817, + 0.005506169982254505, + 0.007570784073323011, + 0.005971338599920273, + 0.005346635822206736, + 0.011021316051483154, + 0.011283071711659431, + 0.010542127303779125, + 0.015270779840648174, + 0.0077768415212631226, + 0.00908310804516077, + 0.0061705647967755795, + 0.005779847968369722, + 0.011744733899831772, + 0.010620883665978909, + 0.007086515426635742, + 0.011031418107450008, + 0.005641762167215347, + 0.006467576138675213, + 0.010609675198793411, + 0.003670834470540285, + 0.009295716881752014, + 0.011678379960358143, + 0.004368662368506193, + 0.06492390483617783, + 0.018368761986494064, + 0.03080456145107746, + 0.017174871638417244, + 0.018837742507457733, + 0.04921218380331993, + 0.031106583774089813, + 0.02511417493224144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.004090514499694109 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the distinctive aesthetic of traditional Japanese Ukiyo-e woodblock prints, the scene opens with an extreme close-up of an assault rifle, its metallic form rendered in flat black ink and strong outlines. The camera remains still, allowing the eye to linger on the intricate decoration: delicate pink cherry blossoms, their petals and stems depicted with elegant lines and soft, unshaded colors, artfully arranged across the weapon's butt and barrel. The camera then pans across the sleek, steady back of the doll, revealing the rifle clutched by a figure clad in a deep indigo wingsuit, the fabric displaying stylized folds and catching the unseen wind. As the camera continues to pan away, the towering figure's full form is revealed, their body maintaining a tense, horizontal glide. The doll's back accelerates, transitioning into a sweeping wide shot, revealing a stunning backdrop: the figure, a lone silhouette, silhouetted against a vast sky filled with swirling, stylized clouds rendered in planes of cream and pale gold. Below, jagged, blue-hued mountains rise from billowing white mist, and the entire composition, flattened in perspective, is defined by the bold lines and harmonious color palette characteristic of Ukiyo-e art, creating a surreal and contemplative atmosphere that juxtaposes serene natural beauty with modern military imagery.", + "seed": 9459, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_10/index_6542_seed_9459_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0599878387991339, + "vlm_logprob": -0.0599878387991339, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.0599878387991339, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.060124658048152924 + ], + "opposite_logprob_tokens": [ + -2.935124635696411 + ], + "logprob_reward": -0.060124658048152924, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.060124658048152924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39852532744407654, + 0.04790189489722252, + 0.05582311376929283, + 0.054898105561733246, + 0.1396237164735794, + 0.04640436917543411, + 0.07536948472261429, + 0.04164644703269005, + 0.08871795237064362, + 0.051089633256196976 + ], + "credit_map": [ + 0.02337227389216423, + 0.13051745295524597, + 0.01578889973461628, + 0.019760051742196083, + 0.02556878887116909, + 0.013565434142947197, + 0.13952647149562836, + 0.030425896868109703, + 0.005832106806337833, + 0.005375347100198269, + 0.007862292230129242, + 0.0050887162797153, + 0.005836716387420893, + 0.006017063278704882, + 0.006358304060995579, + 0.005531344097107649, + 0.006362590938806534, + 0.0048166620545089245, + 0.008728583343327045, + 0.009743034839630127, + 0.00575536023825407, + 0.008692472241818905, + 0.007438230328261852, + 0.004286174196749926, + 0.008881251327693462, + 0.004806821700185537, + 0.008053061552345753, + 0.011014911346137524, + 0.004187737591564655, + 0.0055852108635008335, + 0.006081047002226114, + 0.006288057193160057, + 0.010801441967487335, + 0.014272528700530529, + 0.022746559232473373, + 0.012091224081814289, + 0.01785964146256447, + 0.022552745416760445, + 0.021526489406824112, + 0.01777307130396366, + 0.008339173160493374, + 0.005888286512345076, + 0.005979661829769611, + 0.005116620566695929, + 0.0038780164904892445, + 0.005899672396481037, + 0.005375510081648827, + 0.005927424877882004, + 0.013716038316488266, + 0.011362036690115929, + 0.009590190835297108, + 0.007252463139593601, + 0.007017994299530983, + 0.011896957643330097, + 0.00845921691507101, + 0.006074579898267984, + 0.006095716264098883, + 0.004715397022664547, + 0.005091813858598471, + 0.003564988262951374, + 0.004817566368728876, + 0.006710952613502741, + 0.006522241979837418, + 0.00412776879966259, + 0.013946056365966797, + 0.010523993521928787, + 0.01729866862297058, + 0.010723314248025417, + 0.005146535113453865, + 0.0152259711176157, + 0.006203708704560995, + 0.009649695828557014, + 0.005725568626075983, + 0.007830096408724785, + 0.008173567242920399, + 0.011369957588613033, + 0.0036561093293130398, + 0.005896703340113163, + 0.004242635797709227, + 0.004194993991404772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.060124658048152924 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159177482128143 + ], + "opposite_logprob_tokens": [ + -1.7015917301177979 + ], + "logprob_reward": -0.20159177482128143, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159177482128143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48035162687301636, + 0.018411725759506226, + 0.03801741823554039, + 0.1607157289981842, + 0.0286287609487772, + 0.11898651719093323, + 0.025811119005084038, + 0.021537382155656815, + 0.02673172578215599, + 0.08080800622701645 + ], + "credit_map": [ + 0.02982148714363575, + 0.1686965972185135, + 0.017228182405233383, + 0.024873953312635422, + 0.031418804079294205, + 0.01616550050675869, + 0.15791450440883636, + 0.03423256427049637, + 0.002277607563883066, + 0.002281628083437681, + 0.0027212987188249826, + 0.0022852947004139423, + 0.0019911821000277996, + 0.0027193196583539248, + 0.002193916356191039, + 0.0019414806738495827, + 0.004656838718801737, + 0.0034979311749339104, + 0.006116664968430996, + 0.007311055436730385, + 0.0036530597135424614, + 0.005260145757347345, + 0.004355015233159065, + 0.0031667060684412718, + 0.02579399012029171, + 0.012582949362695217, + 0.021980810910463333, + 0.03831476718187332, + 0.014007965102791786, + 0.015477090142667294, + 0.021214624866843224, + 0.011343524791300297, + 0.0031832095701247454, + 0.0027229376137256622, + 0.004320448264479637, + 0.004074865020811558, + 0.0035713494289666414, + 0.0032053354661911726, + 0.004355797078460455, + 0.0031948185060173273, + 0.020067600533366203, + 0.01555375475436449, + 0.012519565410912037, + 0.01254830788820982, + 0.010667449794709682, + 0.014924242161214352, + 0.014484241604804993, + 0.018221355974674225, + 0.005178801715373993, + 0.0035701128654181957, + 0.00430658645927906, + 0.0023659055586904287, + 0.0015868135960772634, + 0.004662937484681606, + 0.00188282725866884, + 0.002257134998217225, + 0.005810865201056004, + 0.002355241449549794, + 0.003850093111395836, + 0.0014622630551457405, + 0.0009273405303247273, + 0.004664278123527765, + 0.0008719823090359569, + 0.0015953182009980083, + 0.004531440325081348, + 0.0030121004674583673, + 0.004268480930477381, + 0.0031837604474276304, + 0.001952980412170291, + 0.004619887564331293, + 0.0012315780622884631, + 0.0039314948953688145, + 0.011113922111690044, + 0.012863989919424057, + 0.011635526083409786, + 0.018448011949658394, + 0.004321894142776728, + 0.009203399531543255, + 0.0069419448263943195, + 0.006279319990426302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.20159177482128143 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031928059179335833 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.0031928059179335833, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031928059179335833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4806663691997528, + 0.013368945568799973, + 0.023654622957110405, + 0.18287166953086853, + 0.017378490418195724, + 0.02457599714398384, + 0.02250209078192711, + 0.040865734219551086, + 0.07036376744508743, + 0.12375228106975555 + ], + "credit_map": [ + 0.030062202364206314, + 0.16809357702732086, + 0.017333563417196274, + 0.025301141664385796, + 0.031996142119169235, + 0.017106477171182632, + 0.15668492019176483, + 0.03408833593130112, + 0.001313279732130468, + 0.0021520177833735943, + 0.0017226968193426728, + 0.0012211899738758802, + 0.0019183080876246095, + 0.0020138786640018225, + 0.0016326034674420953, + 0.0013949702261015773, + 0.002340303035452962, + 0.0017608103808015585, + 0.00300962314940989, + 0.0031060019973665476, + 0.0022974966559559107, + 0.0070012714713811874, + 0.0023296931758522987, + 0.0018094242550432682, + 0.029237207025289536, + 0.014435007236897945, + 0.02297794260084629, + 0.042702414095401764, + 0.016206488013267517, + 0.020515626296401024, + 0.022353237494826317, + 0.01444372907280922, + 0.0016012720298022032, + 0.0015675885370001197, + 0.0030776746571063995, + 0.0021031296346336603, + 0.0027789385057985783, + 0.0023580845445394516, + 0.0020279097370803356, + 0.00186389300506562, + 0.004105488769710064, + 0.003943906165659428, + 0.0036727760452777147, + 0.0030293993186205626, + 0.0018729162402451038, + 0.003427250310778618, + 0.0018510288791731, + 0.0026732305996119976, + 0.0036101648584008217, + 0.0025360072031617165, + 0.0035317775327712297, + 0.002562415786087513, + 0.00233343243598938, + 0.0036331401206552982, + 0.002300052437931299, + 0.0019951011054217815, + 0.004858565516769886, + 0.003453389275819063, + 0.008385018445551395, + 0.0036244564689695835, + 0.0034732401836663485, + 0.007579217664897442, + 0.004114182200282812, + 0.00537766981869936, + 0.011628219857811928, + 0.009320452809333801, + 0.010940182954072952, + 0.007314563728868961, + 0.007482251152396202, + 0.01110558956861496, + 0.0037270006723701954, + 0.008845506235957146, + 0.012561684474349022, + 0.013401107862591743, + 0.014332887716591358, + 0.017348015680909157, + 0.0156164038926363, + 0.014784777536988258, + 0.01337786577641964, + 0.0223295446485281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.0031928059179335833 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029783394187688828 + ], + "opposite_logprob_tokens": [ + -3.5297834873199463 + ], + "logprob_reward": -0.029783394187688828, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029783394187688828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39898911118507385, + 0.05092789977788925, + 0.04341129586100578, + 0.1522694081068039, + 0.04618626832962036, + 0.024029145017266273, + 0.04565349966287613, + 0.08519181609153748, + 0.10114694386720657, + 0.05219462141394615 + ], + "credit_map": [ + 0.023454977199435234, + 0.1288701295852661, + 0.01660813018679619, + 0.02131742052733898, + 0.025634929537773132, + 0.015136298723518848, + 0.13822659850120544, + 0.029740627855062485, + 0.006244372110813856, + 0.006347046699374914, + 0.007780331186950207, + 0.005081686191260815, + 0.005886739119887352, + 0.006694532465189695, + 0.006289173383265734, + 0.006604015827178955, + 0.004741096403449774, + 0.0033055911771953106, + 0.006624642293900251, + 0.007355721201747656, + 0.00409452011808753, + 0.008450988680124283, + 0.0057997386902570724, + 0.003038991242647171, + 0.024268964305520058, + 0.012663658708333969, + 0.01739727519452572, + 0.03590454161167145, + 0.013268755748867989, + 0.016492651775479317, + 0.020050445571541786, + 0.012223114259541035, + 0.004517556633800268, + 0.004220976959913969, + 0.011431610211730003, + 0.004925982560962439, + 0.006529127713292837, + 0.007512285839766264, + 0.0037054619751870632, + 0.0033432624768465757, + 0.0032050020527094603, + 0.00307171861641109, + 0.0035602105781435966, + 0.0026949348393827677, + 0.0021750570740550756, + 0.003314077155664563, + 0.0026387458201497793, + 0.00336939818225801, + 0.005971887614578009, + 0.008323739282786846, + 0.006654507480561733, + 0.005157994106411934, + 0.0039307367987930775, + 0.006615699268877506, + 0.004620445892214775, + 0.004378487356007099, + 0.014616595581173897, + 0.006990740541368723, + 0.015022855252027512, + 0.009520821273326874, + 0.004067262168973684, + 0.015200348570942879, + 0.004756771493703127, + 0.015016415156424046, + 0.018241673707962036, + 0.011292469687759876, + 0.02364172413945198, + 0.009748493321239948, + 0.005323437042534351, + 0.02029225043952465, + 0.004737394396215677, + 0.007869492284953594, + 0.006674334406852722, + 0.007154710125178099, + 0.006650266703218222, + 0.011120173148810863, + 0.004494468215852976, + 0.005968563724309206, + 0.005869005806744099, + 0.004263097885996103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.029783394187688828 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an assault rifle held by a figure in a wingsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are pink cherry blossoms depicted on the rifle with flat black ink and strong outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a deep indigo wingsuit with stylized folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and then pan to reveal a wide shot of the figure gliding horizontally?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast sky with cream and pale gold clouds above blue-hued mountains and white mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005246561020612717 + ], + "opposite_logprob_tokens": [ + -5.255246639251709 + ], + "logprob_reward": -0.005246561020612717, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005246561020612717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4642200171947479, + 0.0199056938290596, + 0.04401343688368797, + 0.09974255412817001, + 0.02153952419757843, + 0.03340530022978783, + 0.041899364441633224, + 0.12250811606645584, + 0.05647565796971321, + 0.09629040211439133 + ], + "credit_map": [ + 0.029046645388007164, + 0.161667138338089, + 0.01713145524263382, + 0.02442198432981968, + 0.03102945163846016, + 0.018015926703810692, + 0.15017476677894592, + 0.03273262828588486, + 0.0019438407616689801, + 0.0023623097222298384, + 0.003232494695112109, + 0.0021469746716320515, + 0.002475262386724353, + 0.0030544125474989414, + 0.002622135216370225, + 0.002068262780085206, + 0.004990444518625736, + 0.0035593726206570864, + 0.007555400487035513, + 0.009869363158941269, + 0.004192645195871592, + 0.005597032606601715, + 0.0048075285740196705, + 0.003441642737016082, + 0.01606331765651703, + 0.008064580149948597, + 0.0146948816254735, + 0.02233494259417057, + 0.008372530341148376, + 0.009996752254664898, + 0.012356624938547611, + 0.007858922705054283, + 0.002501319395378232, + 0.0018908916972577572, + 0.004263178911060095, + 0.0023702499456703663, + 0.0029902951791882515, + 0.0030025348532944918, + 0.0021366742439568043, + 0.002384378807619214, + 0.004751653876155615, + 0.00402166647836566, + 0.00471185939386487, + 0.0063214972615242004, + 0.0023567043244838715, + 0.004563221707940102, + 0.003025469370186329, + 0.003653226885944605, + 0.00543454522266984, + 0.008098803460597992, + 0.006256063003093004, + 0.0046957070007920265, + 0.0036579861771315336, + 0.005873833782970905, + 0.003370997728779912, + 0.004511422477662563, + 0.019612492993474007, + 0.009210233576595783, + 0.026170754805207253, + 0.012552419677376747, + 0.005229299422353506, + 0.02461567148566246, + 0.006840190384536982, + 0.01827705278992653, + 0.01226972509175539, + 0.01076355017721653, + 0.006952776573598385, + 0.0057570128701627254, + 0.0035199285484850407, + 0.010512687265872955, + 0.0032620178535580635, + 0.003437953069806099, + 0.011873717419803143, + 0.015995463356375694, + 0.012949930503964424, + 0.016273705288767815, + 0.011246773414313793, + 0.010686680674552917, + 0.010036597028374672, + 0.007227524649351835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.005246561020612717 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_11/index_tensor([5849, 5849])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_11/index_tensor([5849, 5849])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..490134f736ddf5ca24a6884bc4fe9abff5bf4a1f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_11/index_tensor([5849, 5849])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In an extreme close-up, first-person perspective shot, rendered in a bright, highly saturated, flat-block art style, a pair of precision stainless steel tweezers, gleaming with a cool silver sheen, are held steadily by a pair of hands whose fingertips tremble slightly with intense concentration. The tips of the tweezers delicately grasp a tiny silver screw, barely the size of a grain of sand. The cross-shaped grooves on its head are clearly visible under the even light from an overhead shadowless lamp. As the camera slowly and steadily moves downward, the tweezers approach the green circuit board beneath, a structure composed of crisp geometric shapes within the phone. Golden circuit paths crisscross like sharp lines, interspersed with square electronic components in pure black, bright blue, and vivid red. The image, devoid of complex gradients or shading, exudes a clean, graphic aesthetic. Finally, the tweezers hover at their target, precisely aligning the tiny screw with a nearly invisible screw hole on the circuit board. The entire scene is imbued with extreme focus and a tense, technologically aesthetic atmosphere suspended in the moment.", + "seed": 10874, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_11/index_5849_seed_10874_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14083052081987263, + "vlm_logprob": -0.14083052081987263, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.14083052081987263, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16197995841503143 + ], + "opposite_logprob_tokens": [ + -1.9119799137115479 + ], + "logprob_reward": -0.16197995841503143, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16197995841503143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26201826333999634, + 0.17948654294013977, + 0.025613892823457718, + 0.336074560880661, + 0.0717393159866333, + 0.05765025317668915, + 0.03346303477883339, + 0.018034569919109344, + 0.010822832584381104, + 0.005096727050840855 + ], + "credit_map": [ + 0.02425956353545189, + 0.04347895085811615, + 0.024131538346409798, + 0.019628064706921577, + 0.06584666669368744, + 0.04587377980351448, + 0.01992974989116192, + 0.018869968131184578, + 0.02495829574763775, + 0.03157516196370125, + 0.01761426031589508, + 0.01818752847611904, + 0.01903570257127285, + 0.029045339673757553, + 0.02382211573421955, + 0.015248141251504421, + 0.0030525398906320333, + 0.0035801539197564125, + 0.002623677719384432, + 0.0034603094682097435, + 0.0027615854050964117, + 0.004201349336653948, + 0.0033776049967855215, + 0.0025566734839230776, + 0.07511179149150848, + 0.05295168235898018, + 0.13205496966838837, + 0.017904872074723244, + 0.010845329612493515, + 0.009207134135067463, + 0.007053864188492298, + 0.030944932252168655, + 0.014408895745873451, + 0.006512120366096497, + 0.0061526820063591, + 0.00648069940507412, + 0.00502658961340785, + 0.018259448930621147, + 0.006877198349684477, + 0.008021683432161808, + 0.007850220426917076, + 0.009075969457626343, + 0.00753050809726119, + 0.007576292380690575, + 0.008451154455542564, + 0.004530998878180981, + 0.007889077067375183, + 0.004746029619127512, + 0.0036728198174387217, + 0.004283367656171322, + 0.00786413811147213, + 0.00430378969758749, + 0.004562980029731989, + 0.0025700374972075224, + 0.003396368119865656, + 0.0028095373418182135, + 0.003214446594938636, + 0.003144654678180814, + 0.0023219743743538857, + 0.0018102530157193542, + 0.001943982788361609, + 0.0017697809962555766, + 0.002348140114918351, + 0.0014813361922279, + 0.0015384586295112967, + 0.0008943295106291771, + 0.0021645661909133196, + 0.0013145867269486189, + 0.0008899927488528192, + 0.0015438409755006433, + 0.001256694900803268, + 0.0012203624937683344, + 0.0006052220123820007, + 0.0006360856932587922, + 0.0007946226396597922, + 0.0006437560659833252, + 0.000515779887791723, + 0.0007270573987625539, + 0.0005503201973624527, + 0.000623882922809571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.16197995841503143 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25215834379196167 + ], + "opposite_logprob_tokens": [ + -1.5021584033966064 + ], + "logprob_reward": -0.25215834379196167, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25215834379196167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2322024703025818, + 0.3601832687854767, + 0.03387557342648506, + 0.07386138290166855, + 0.19078727066516876, + 0.025288376957178116, + 0.03225571662187576, + 0.029298268258571625, + 0.011596529744565487, + 0.010651144199073315 + ], + "credit_map": [ + 0.020963624119758606, + 0.03847718983888626, + 0.023600945249199867, + 0.017852960154414177, + 0.061096303164958954, + 0.03910710662603378, + 0.017992109060287476, + 0.013112214393913746, + 0.053135406225919724, + 0.06057894229888916, + 0.03874438256025314, + 0.03439074754714966, + 0.03740337863564491, + 0.062036801129579544, + 0.04639487713575363, + 0.02749871090054512, + 0.003634704276919365, + 0.003955098334699869, + 0.004907845985144377, + 0.0037987451069056988, + 0.0038548384327441454, + 0.003422342473641038, + 0.0070190997794270515, + 0.0032828941475600004, + 0.016607733443379402, + 0.009318971075117588, + 0.01665559411048889, + 0.006103072315454483, + 0.005692056380212307, + 0.005759969353675842, + 0.006044686306267977, + 0.007679295260459185, + 0.03297262266278267, + 0.020213084295392036, + 0.01737462542951107, + 0.019718339666724205, + 0.013472653925418854, + 0.04793120548129082, + 0.01811782829463482, + 0.020986899733543396, + 0.0035970592871308327, + 0.0040122140198946, + 0.0028368395287543535, + 0.003088130382820964, + 0.0031189010478556156, + 0.002494565211236477, + 0.00422374252229929, + 0.001916922046802938, + 0.004434267990291119, + 0.004315402824431658, + 0.004903197754174471, + 0.002782488241791725, + 0.005545406136661768, + 0.004410095512866974, + 0.0034858346916735172, + 0.0023790188133716583, + 0.005341654643416405, + 0.004110944457352161, + 0.004343139939010143, + 0.0030466315802186728, + 0.0034756057430058718, + 0.0028569099958986044, + 0.0034810781944543123, + 0.0026423002127557993, + 0.0017198085552081466, + 0.0013130470179021358, + 0.002248275326564908, + 0.0011993256630375981, + 0.0011082073906436563, + 0.001574028399772942, + 0.0014711411204189062, + 0.0009626952232792974, + 0.0019889450632035732, + 0.0010342474561184645, + 0.0017783717485144734, + 0.0011880147503688931, + 0.0010578634683042765, + 0.0013219959801062942, + 0.0014092068886384368, + 0.000872497446835041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.25215834379196167 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813459038734436 + ], + "opposite_logprob_tokens": [ + -1.4063458442687988 + ], + "logprob_reward": -0.2813459038734436, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813459038734436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12067427486181259, + 0.13531969487667084, + 0.020541710779070854, + 0.44373345375061035, + 0.15524381399154663, + 0.027380218729376793, + 0.02439192309975624, + 0.04992135241627693, + 0.013976257294416428, + 0.008817262947559357 + ], + "credit_map": [ + 0.012758715078234673, + 0.017098456621170044, + 0.013609332032501698, + 0.015042038634419441, + 0.026958633214235306, + 0.01852557249367237, + 0.01046134065836668, + 0.006220203824341297, + 0.022404415532946587, + 0.021945158019661903, + 0.014052622951567173, + 0.011240532621741295, + 0.012707347050309181, + 0.033865924924612045, + 0.00959484837949276, + 0.009508853778243065, + 0.0023132069036364555, + 0.0024656946770846844, + 0.002522183582186699, + 0.003144474932923913, + 0.0017199866706505418, + 0.0033647920936346054, + 0.002109162276610732, + 0.002902211854234338, + 0.11698385328054428, + 0.04535334184765816, + 0.17721523344516754, + 0.024544652551412582, + 0.012820342555642128, + 0.013341203331947327, + 0.009356913156807423, + 0.044117942452430725, + 0.02955925650894642, + 0.01682290993630886, + 0.014043298549950123, + 0.016188183799386024, + 0.012398405000567436, + 0.03712999075651169, + 0.01570732146501541, + 0.013394464738667011, + 0.0029578283429145813, + 0.004243185743689537, + 0.003777634585276246, + 0.003216936020180583, + 0.004729443695396185, + 0.0025024537462741137, + 0.0036308295093476772, + 0.00232190964743495, + 0.0030062601435929537, + 0.0030909141059964895, + 0.004235065542161465, + 0.002509249374270439, + 0.0039421748369932175, + 0.00280402647331357, + 0.0027316000778228045, + 0.0020726355724036694, + 0.008994150906801224, + 0.006776511203497648, + 0.0047659543342888355, + 0.003331865416839719, + 0.006654099561274052, + 0.007442442234605551, + 0.0074905031360685825, + 0.004465829115360975, + 0.0018651641439646482, + 0.002623584819957614, + 0.0020821138750761747, + 0.0012231424916535616, + 0.0016217852244153619, + 0.0012966087087988853, + 0.0016961864894255996, + 0.0015676732873544097, + 0.0014921041438356042, + 0.0013320030411705375, + 0.0011402980890125036, + 0.0007757661514915526, + 0.0009778995299711823, + 0.0007295577088370919, + 0.0012806988088414073, + 0.0010889364639297128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.2813459038734436 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730979308485985 + ], + "opposite_logprob_tokens": [ + -5.006731033325195 + ], + "logprob_reward": -0.006730979308485985, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730979308485985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2102096974849701, + 0.12802553176879883, + 0.0405704565346241, + 0.13836726546287537, + 0.32298070192337036, + 0.07120709121227264, + 0.048893339931964874, + 0.017487648874521255, + 0.016566067934036255, + 0.0056922235526144505 + ], + "credit_map": [ + 0.021917147561907768, + 0.03446108475327492, + 0.02025522105395794, + 0.016003668308258057, + 0.04193764179944992, + 0.039115022867918015, + 0.01818338967859745, + 0.01833653450012207, + 0.02067222259938717, + 0.021404189988970757, + 0.01397741213440895, + 0.012183038517832756, + 0.012851743027567863, + 0.024697158485651016, + 0.013299034908413887, + 0.008940738625824451, + 0.004066536668688059, + 0.004593816585838795, + 0.004525998141616583, + 0.006284317467361689, + 0.003822463797405362, + 0.004990825429558754, + 0.007323043886572123, + 0.004963453859090805, + 0.024570507928729057, + 0.028311919420957565, + 0.04964791238307953, + 0.007472391705960035, + 0.006473146844655275, + 0.0055814278312027454, + 0.004442681558430195, + 0.011867273598909378, + 0.058071330189704895, + 0.03144066035747528, + 0.028263213112950325, + 0.03315609320998192, + 0.021311264485120773, + 0.08580674231052399, + 0.02895374223589897, + 0.035977646708488464, + 0.009623062796890736, + 0.011110719293355942, + 0.007254891097545624, + 0.006872972007840872, + 0.011133578605949879, + 0.007416481152176857, + 0.011177701875567436, + 0.006617681588977575, + 0.00811991561204195, + 0.004974075127393007, + 0.010149558074772358, + 0.006884569302201271, + 0.005634709727019072, + 0.004519443027675152, + 0.00570679921656847, + 0.00290427403524518, + 0.002368105109781027, + 0.003419452579692006, + 0.002726514358073473, + 0.0017574999947100878, + 0.0024016862735152245, + 0.001585107878781855, + 0.001963711343705654, + 0.0012655708706006408, + 0.0031594170723110437, + 0.0020356273744255304, + 0.0019372308161109686, + 0.0017252406105399132, + 0.0017475247150287032, + 0.0012974694836884737, + 0.002926727756857872, + 0.00173683010507375, + 0.0008137173135764897, + 0.0007320952718146145, + 0.0005872478941455483, + 0.000667692394927144, + 0.0006658837082795799, + 0.0007856799638830125, + 0.0007319936994463205, + 0.0007079131901264191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.006730979308485985 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019374187104403973 + ], + "opposite_logprob_tokens": [ + -6.251937389373779 + ], + "logprob_reward": -0.0019374187104403973, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019374187104403973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20986102521419525, + 0.09195829927921295, + 0.045708030462265015, + 0.2186167687177658, + 0.1398565173149109, + 0.09044525772333145, + 0.09915027767419815, + 0.052810654044151306, + 0.027554739266633987, + 0.024038447067141533 + ], + "credit_map": [ + 0.015686603263020515, + 0.047929536551237106, + 0.01696614921092987, + 0.013426552526652813, + 0.030933532863855362, + 0.0524991974234581, + 0.019514387473464012, + 0.012905078008770943, + 0.014608415775001049, + 0.010368638671934605, + 0.01737690344452858, + 0.011701120063662529, + 0.006651442032307386, + 0.013836764730513096, + 0.009734627790749073, + 0.0076803951524198055, + 0.0046677365899086, + 0.006191845517605543, + 0.004627571441233158, + 0.010327989235520363, + 0.0046767001040279865, + 0.005173082463443279, + 0.004901234060525894, + 0.005141871050000191, + 0.027659697458148003, + 0.061687029898166656, + 0.07179436832666397, + 0.013239692896604538, + 0.012235903181135654, + 0.008766891434788704, + 0.008344990201294422, + 0.014888186007738113, + 0.028535181656479836, + 0.013423149473965168, + 0.014006677083671093, + 0.013057980686426163, + 0.009236481040716171, + 0.03864920884370804, + 0.010118640959262848, + 0.012829192914068699, + 0.012569691054522991, + 0.011468173936009407, + 0.010988435707986355, + 0.008401959203183651, + 0.011766201816499233, + 0.00990275852382183, + 0.016322746872901917, + 0.009025291539728642, + 0.015291141346096992, + 0.011987624689936638, + 0.01757308468222618, + 0.01011059619486332, + 0.015154222026467323, + 0.010299966670572758, + 0.011080033145844936, + 0.007653603330254555, + 0.007667104247957468, + 0.0076354555785655975, + 0.009149263612926006, + 0.00565816555172205, + 0.006529216654598713, + 0.005556213669478893, + 0.005270328838378191, + 0.005344907753169537, + 0.004245582036674023, + 0.0029052128084003925, + 0.005725056864321232, + 0.002996060997247696, + 0.002762389136478305, + 0.0029895398765802383, + 0.00330744544044137, + 0.0026234504766762257, + 0.003985668998211622, + 0.0022512844298034906, + 0.0046949139796197414, + 0.0026046254206448793, + 0.002703708363696933, + 0.0026665558107197285, + 0.0032072451431304216, + 0.0019244439899921417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.0019374187104403973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, first-person perspective shot, rendered in a bright, highly saturated, flat-block art style, a pair of precision stainless steel tweezers, gleaming with a cool silver sheen, are held steadily by a pair of hands whose fingertips tremble slightly with intense concentration. The tips of the tweezers delicately grasp a tiny silver screw, barely the size of a grain of sand. The cross-shaped grooves on its head are clearly visible under the even light from an overhead shadowless lamp. As the camera slowly and steadily moves downward, the tweezers approach the green circuit board beneath, a structure composed of crisp geometric shapes within the phone. Golden circuit paths crisscross like sharp lines, interspersed with square electronic components in pure black, bright blue, and vivid red. The image, devoid of complex gradients or shading, exudes a clean, graphic aesthetic. Finally, the tweezers hover at their target, precisely aligning the tiny screw with a nearly invisible screw hole on the circuit board. The entire scene is imbued with extreme focus and a tense, technologically aesthetic atmosphere suspended in the moment.", + "seed": 10875, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_11/index_5849_seed_10875_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023544660210609435, + "vlm_logprob": -0.023544660210609435, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.023544660210609435, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013903550803661346 + ], + "opposite_logprob_tokens": [ + -4.638903617858887 + ], + "logprob_reward": -0.013903550803661346, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013903550803661346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10289350897073746, + 0.12566623091697693, + 0.028911838307976723, + 0.017741601914167404, + 0.48731106519699097, + 0.04947173222899437, + 0.06112657114863396, + 0.02733069472014904, + 0.03716791421175003, + 0.06237887591123581 + ], + "credit_map": [ + 0.009716126136481762, + 0.008501056581735611, + 0.015060463920235634, + 0.010670427232980728, + 0.00956436712294817, + 0.013222826644778252, + 0.01802249439060688, + 0.01813574507832527, + 0.007749704644083977, + 0.007715712767094374, + 0.036512043327093124, + 0.007709984667599201, + 0.0076215327717363834, + 0.010846300981938839, + 0.00891757570207119, + 0.03859337419271469, + 0.004100071731954813, + 0.00360017828643322, + 0.004684310406446457, + 0.002948105800896883, + 0.0030713235028088093, + 0.003581190248951316, + 0.0022846146021038294, + 0.004642043262720108, + 0.0019345029722899199, + 0.002770108636468649, + 0.002904255175963044, + 0.0025361906737089157, + 0.0021673685405403376, + 0.0013860028702765703, + 0.00166822480969131, + 0.002374947303906083, + 0.06265381723642349, + 0.0336068831384182, + 0.20639459788799286, + 0.1000702828168869, + 0.0148664191365242, + 0.01905289851129055, + 0.023746004328131676, + 0.026920123025774956, + 0.006061936728656292, + 0.005330653861165047, + 0.006308394018560648, + 0.005251748021692038, + 0.003725882153958082, + 0.010424342937767506, + 0.006136414594948292, + 0.006232359446585178, + 0.006546542048454285, + 0.008075997233390808, + 0.006237172521650791, + 0.007401387672871351, + 0.005139623768627644, + 0.010433383285999298, + 0.007016901858150959, + 0.010275562293827534, + 0.004454190377146006, + 0.004305626731365919, + 0.00375802768394351, + 0.0024183676578104496, + 0.0025543251540511847, + 0.003755114274099469, + 0.003573735011741519, + 0.0025113089941442013, + 0.004753526300191879, + 0.004711179528385401, + 0.0042269048281013966, + 0.0035792968701571226, + 0.003270189044997096, + 0.0032638078555464745, + 0.010446888394653797, + 0.0029161209240555763, + 0.007158684078603983, + 0.007987071759998798, + 0.003974286373704672, + 0.006596852093935013, + 0.006732420530170202, + 0.009566865861415863, + 0.014713727869093418, + 0.005648968741297722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.013903550803661346 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026378009468317032 + ], + "opposite_logprob_tokens": [ + -3.6513779163360596 + ], + "logprob_reward": -0.026378009468317032, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026378009468317032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1256973147392273, + 0.19506965577602386, + 0.05539293587207794, + 0.01806253381073475, + 0.24042722582817078, + 0.05500054731965065, + 0.10091451555490494, + 0.09726996719837189, + 0.04783819988369942, + 0.0643271654844284 + ], + "credit_map": [ + 0.014739749021828175, + 0.01356513611972332, + 0.014908472076058388, + 0.01164740789681673, + 0.017115183174610138, + 0.021551955491304398, + 0.020983513444662094, + 0.011185902170836926, + 0.009606773965060711, + 0.009084747172892094, + 0.05861528590321541, + 0.01184058003127575, + 0.012537572532892227, + 0.018572349101305008, + 0.014616874977946281, + 0.060195472091436386, + 0.005608442705124617, + 0.004053729120641947, + 0.016229161992669106, + 0.007109994068741798, + 0.004048179369419813, + 0.003511302173137665, + 0.008636170998215675, + 0.006195956375449896, + 0.0024802142288535833, + 0.002333912765607238, + 0.002887662732973695, + 0.0026187128387391567, + 0.002175732282921672, + 0.002143456367775798, + 0.001813477254472673, + 0.0016093662707135081, + 0.02797611989080906, + 0.019591454416513443, + 0.09350960701704025, + 0.04918592423200607, + 0.01045230496674776, + 0.012338037602603436, + 0.0128822922706604, + 0.014491495676338673, + 0.006583943963050842, + 0.00823118630796671, + 0.009131908416748047, + 0.005815142299979925, + 0.0036841961555182934, + 0.008584242314100266, + 0.006156557239592075, + 0.00681336922571063, + 0.010876980610191822, + 0.015479844994843006, + 0.009356500580906868, + 0.010620209388434887, + 0.007901829667389393, + 0.01627546176314354, + 0.013664179481565952, + 0.016739513725042343, + 0.016475148499011993, + 0.021996265277266502, + 0.009302343241870403, + 0.00903969444334507, + 0.009686070494353771, + 0.010347018018364906, + 0.009961448609828949, + 0.010461978614330292, + 0.006875503808259964, + 0.008153078146278858, + 0.004796414170414209, + 0.004981104750186205, + 0.004598995670676231, + 0.004583720583468676, + 0.00924896914511919, + 0.004600412677973509, + 0.010121922008693218, + 0.007493719924241304, + 0.006171831861138344, + 0.005877065006643534, + 0.006125583779066801, + 0.007983028888702393, + 0.010006856173276901, + 0.010547157377004623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.026378009468317032 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207685545086861 + ], + "opposite_logprob_tokens": [ + -2.8120768070220947 + ], + "logprob_reward": -0.06207685545086861, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207685545086861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0948706790804863, + 0.0387951135635376, + 0.050168726593256, + 0.0333746112883091, + 0.35466545820236206, + 0.05615464225411415, + 0.1496284008026123, + 0.06606761366128922, + 0.09802316874265671, + 0.05825159326195717 + ], + "credit_map": [ + 0.013996749185025692, + 0.014196169562637806, + 0.009766759350895882, + 0.007473291363567114, + 0.01320017222315073, + 0.013610064052045345, + 0.01189967431128025, + 0.010727795772254467, + 0.003398589324206114, + 0.003933748695999384, + 0.007103273179382086, + 0.005156995728611946, + 0.004753949586302042, + 0.0036968188360333443, + 0.0055915964767336845, + 0.005160143598914146, + 0.005739261396229267, + 0.0033742294181138277, + 0.019898204132914543, + 0.009218908846378326, + 0.002838422544300556, + 0.002882747445255518, + 0.0037505868822336197, + 0.0024663687217980623, + 0.005877734161913395, + 0.003619520692154765, + 0.004763856064528227, + 0.008421937935054302, + 0.0027767641004174948, + 0.002584360307082534, + 0.0024426837917417288, + 0.0028877544682472944, + 0.04225826635956764, + 0.026510892435908318, + 0.14500266313552856, + 0.07485410571098328, + 0.012142494320869446, + 0.01485206838697195, + 0.01872972957789898, + 0.020315252244472504, + 0.007617123890668154, + 0.008309750817716122, + 0.007953723892569542, + 0.005847449880093336, + 0.0035498670767992735, + 0.008383860811591148, + 0.006666015833616257, + 0.007826855406165123, + 0.01847047545015812, + 0.02384197898209095, + 0.015465140342712402, + 0.017525864765048027, + 0.009500890038907528, + 0.018281159922480583, + 0.0187531728297472, + 0.02778971940279007, + 0.011488506570458412, + 0.014510281383991241, + 0.0051809195429086685, + 0.0060581122525036335, + 0.006916362792253494, + 0.008317201398313046, + 0.006768436636775732, + 0.006827789358794689, + 0.014307240955531597, + 0.019686847925186157, + 0.00735373143106699, + 0.011009401641786098, + 0.00772165646776557, + 0.00660928338766098, + 0.02301459200680256, + 0.008320415392518044, + 0.007521328516304493, + 0.006425312254577875, + 0.004020247608423233, + 0.005446647293865681, + 0.005550716072320938, + 0.007976151071488857, + 0.01466026995331049, + 0.006650920026004314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -0.06207685545086861 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731926463544369 + ], + "opposite_logprob_tokens": [ + -5.006731986999512 + ], + "logprob_reward": -0.006731926463544369, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731926463544369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2632122337818146, + 0.20683109760284424, + 0.07190713286399841, + 0.09119704365730286, + 0.05514160171151161, + 0.03597238287329674, + 0.07494889944791794, + 0.06230252981185913, + 0.06078281253576279, + 0.07770423591136932 + ], + "credit_map": [ + 0.027498703449964523, + 0.02952035702764988, + 0.03889939561486244, + 0.02978670783340931, + 0.028118060901761055, + 0.020297039300203323, + 0.030235987156629562, + 0.05885598808526993, + 0.01370607316493988, + 0.01269625872373581, + 0.04643052816390991, + 0.01426980271935463, + 0.017612336203455925, + 0.027794495224952698, + 0.02094925008714199, + 0.0533723458647728, + 0.007181946188211441, + 0.005650578532367945, + 0.016159310936927795, + 0.007600776851177216, + 0.007246457505971193, + 0.007851386442780495, + 0.011421595700085163, + 0.008795076049864292, + 0.01302128192037344, + 0.011918937787413597, + 0.014265604317188263, + 0.01999846287071705, + 0.007550716865807772, + 0.0071356045082211494, + 0.008122027851641178, + 0.009184408001601696, + 0.005554946139454842, + 0.0062398724257946014, + 0.010212358087301254, + 0.008554380387067795, + 0.0070543717592954636, + 0.006281154230237007, + 0.005005594342947006, + 0.006238922476768494, + 0.003267007414251566, + 0.005872813519090414, + 0.008569014258682728, + 0.004178141709417105, + 0.0021028935443609953, + 0.0023828954435884953, + 0.0038516302593052387, + 0.005747988820075989, + 0.008612750098109245, + 0.012074610218405724, + 0.006246794015169144, + 0.009244500659406185, + 0.007772871293127537, + 0.006057580001652241, + 0.016670743003487587, + 0.008269046433269978, + 0.008177944459021091, + 0.010334798134863377, + 0.004861751105636358, + 0.005067923106253147, + 0.009265495464205742, + 0.007475655991584063, + 0.010945423506200314, + 0.00617353618144989, + 0.005117558408528566, + 0.006383832544088364, + 0.008542265743017197, + 0.007336887530982494, + 0.005881685763597488, + 0.006622435990720987, + 0.014320671558380127, + 0.006577478256076574, + 0.00905194878578186, + 0.01229123491793871, + 0.005864124745130539, + 0.006636008620262146, + 0.00983633752912283, + 0.00635473383590579, + 0.019500812515616417, + 0.008169034495949745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.006731926463544369 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008632958866655827 + ], + "opposite_logprob_tokens": [ + -4.758633136749268 + ], + "logprob_reward": -0.008632958866655827, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008632958866655827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22186969220638275, + 0.08897516876459122, + 0.02640162780880928, + 0.03883029520511627, + 0.25209537148475647, + 0.04922959953546524, + 0.052019644528627396, + 0.06559626758098602, + 0.09071247279644012, + 0.11426988244056702 + ], + "credit_map": [ + 0.029242146760225296, + 0.01928549073636532, + 0.03684329241514206, + 0.026774415746331215, + 0.01043902151286602, + 0.01638615131378174, + 0.018886206671595573, + 0.06401295959949493, + 0.005269587971270084, + 0.005110741592943668, + 0.029984550550580025, + 0.005938103888183832, + 0.003691030666232109, + 0.00709480419754982, + 0.003128273645415902, + 0.028758063912391663, + 0.002826994750648737, + 0.002250833436846733, + 0.0037635036278516054, + 0.003119352040812373, + 0.004725850187242031, + 0.0025871223770081997, + 0.00366470986045897, + 0.0034632612951099873, + 0.004523701500147581, + 0.006752514746040106, + 0.00583302928134799, + 0.00815938226878643, + 0.00384499947540462, + 0.002504986710846424, + 0.004086408764123917, + 0.0031252724584192038, + 0.03188975527882576, + 0.01700473017990589, + 0.10859359055757523, + 0.052462611347436905, + 0.007837645709514618, + 0.006439007353037596, + 0.012961608357727528, + 0.014906425960361958, + 0.0060601322911679745, + 0.010178633965551853, + 0.009827351197600365, + 0.005953241139650345, + 0.002776917302981019, + 0.004063270054757595, + 0.003826708998531103, + 0.006543343421071768, + 0.005142082925885916, + 0.007425624877214432, + 0.00497221015393734, + 0.0075166127644479275, + 0.004128393717110157, + 0.007988150231540203, + 0.005248165223747492, + 0.009598404169082642, + 0.010778804309666157, + 0.012305288575589657, + 0.006875301245599985, + 0.005369532387703657, + 0.007147963624447584, + 0.009471483528614044, + 0.008255137130618095, + 0.005392762366682291, + 0.01441762875765562, + 0.014790844172239304, + 0.008263196796178818, + 0.008669374510645866, + 0.009649144485592842, + 0.012011096812784672, + 0.012484916485846043, + 0.010426272638142109, + 0.016705838963389397, + 0.020637406036257744, + 0.010555163025856018, + 0.00827381107956171, + 0.013111136853694916, + 0.013742486014962196, + 0.017993932589888573, + 0.013250111602246761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.008632958866655827 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, first-person perspective shot, rendered in a bright, highly saturated, flat-block art style, a pair of precision stainless steel tweezers, gleaming with a cool silver sheen, are held steadily by a pair of hands whose fingertips tremble slightly with intense concentration. The tips of the tweezers delicately grasp a tiny silver screw, barely the size of a grain of sand. The cross-shaped grooves on its head are clearly visible under the even light from an overhead shadowless lamp. As the camera slowly and steadily moves downward, the tweezers approach the green circuit board beneath, a structure composed of crisp geometric shapes within the phone. Golden circuit paths crisscross like sharp lines, interspersed with square electronic components in pure black, bright blue, and vivid red. The image, devoid of complex gradients or shading, exudes a clean, graphic aesthetic. Finally, the tweezers hover at their target, precisely aligning the tiny screw with a nearly invisible screw hole on the circuit board. The entire scene is imbued with extreme focus and a tense, technologically aesthetic atmosphere suspended in the moment.", + "seed": 10876, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_11/index_5849_seed_10876_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.026237391494214536, + "vlm_logprob": -0.026237391494214536, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.026237391494214536, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012694385834038258 + ], + "opposite_logprob_tokens": [ + -5.137694358825684 + ], + "logprob_reward": -0.012694385834038258, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.012694385834038258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1692247986793518, + 0.2429991215467453, + 0.0804382935166359, + 0.07724076509475708, + 0.1416204273700714, + 0.0457223579287529, + 0.06718965619802475, + 0.06571315228939056, + 0.0535607673227787, + 0.0562906488776207 + ], + "credit_map": [ + 0.02485448680818081, + 0.03505370393395424, + 0.01254794280976057, + 0.023032600060105324, + 0.019269878044724464, + 0.013448204845190048, + 0.025933459401130676, + 0.01508451160043478, + 0.04435882344841957, + 0.0628974586725235, + 0.02605818212032318, + 0.01892932876944542, + 0.013975726440548897, + 0.017950382083654404, + 0.017849111929535866, + 0.04098011180758476, + 0.00954547617584467, + 0.01899445615708828, + 0.011044589802622795, + 0.008486345410346985, + 0.006116378586739302, + 0.006136962678283453, + 0.005734357051551342, + 0.014379728585481644, + 0.006009334232658148, + 0.009373676031827927, + 0.024728506803512573, + 0.00822966918349266, + 0.0046102856285870075, + 0.006142131052911282, + 0.004788613878190517, + 0.013358545489609241, + 0.018357783555984497, + 0.021496685221791267, + 0.015657734125852585, + 0.021564725786447525, + 0.014520577155053616, + 0.01797064021229744, + 0.017596546560525894, + 0.014455733820796013, + 0.00484822504222393, + 0.005289171356707811, + 0.005267038010060787, + 0.009074218571186066, + 0.0037627078127115965, + 0.008871549740433693, + 0.0031314147636294365, + 0.005478031001985073, + 0.01044183224439621, + 0.007112840656191111, + 0.007523598149418831, + 0.007713991217315197, + 0.005352330859750509, + 0.009519166313111782, + 0.006155975628644228, + 0.013369917869567871, + 0.006628650240600109, + 0.015751559287309647, + 0.005049339961260557, + 0.005184962414205074, + 0.00890279095619917, + 0.007352803833782673, + 0.007482338696718216, + 0.009360706433653831, + 0.0049992511048913, + 0.007829198613762856, + 0.0065191397443413734, + 0.00641095032915473, + 0.0061483862809836864, + 0.00890932884067297, + 0.005009192507714033, + 0.00773532222956419, + 0.004934041760861874, + 0.010557514615356922, + 0.0030714664608240128, + 0.012108894065022469, + 0.004629953298717737, + 0.0053605870343744755, + 0.0038675637915730476, + 0.011760623194277287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.012694385834038258 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016088075935840607 + ], + "opposite_logprob_tokens": [ + -4.141088008880615 + ], + "logprob_reward": -0.016088075935840607, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016088075935840607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14969131350517273, + 0.2417738437652588, + 0.033027105033397675, + 0.050988905131816864, + 0.2891260087490082, + 0.0448218509554863, + 0.0499093271791935, + 0.024577373638749123, + 0.06269411742687225, + 0.05339009687304497 + ], + "credit_map": [ + 0.02584296092391014, + 0.04123272746801376, + 0.010282046161592007, + 0.016369864344596863, + 0.012261862866580486, + 0.008658129721879959, + 0.0196240097284317, + 0.015419707633554935, + 0.05501550808548927, + 0.06772154569625854, + 0.023041801527142525, + 0.015959784388542175, + 0.013574311509728432, + 0.016856150701642036, + 0.016732512041926384, + 0.03287224471569061, + 0.005339872092008591, + 0.00719631090760231, + 0.004040084779262543, + 0.005369214341044426, + 0.0025569854769855738, + 0.0024744083639234304, + 0.0027137533761560917, + 0.003336477791890502, + 0.0050841933116316795, + 0.0113594401627779, + 0.005971884820610285, + 0.009365467354655266, + 0.00396511098369956, + 0.004475482273846865, + 0.005732815712690353, + 0.00503450958058238, + 0.0359061025083065, + 0.03853832930326462, + 0.03323465585708618, + 0.051764342933893204, + 0.03052772395312786, + 0.03552941605448723, + 0.034306734800338745, + 0.02931869775056839, + 0.0037320249248296022, + 0.0038930086884647608, + 0.005332909058779478, + 0.012542659416794777, + 0.002710264176130295, + 0.005731943063437939, + 0.0038673283997923136, + 0.007011716719716787, + 0.005999731365591288, + 0.0050282469019293785, + 0.005198120139539242, + 0.00687593687325716, + 0.0036230108235031366, + 0.006084633991122246, + 0.005820808466523886, + 0.011278839781880379, + 0.00271463836543262, + 0.006478695664554834, + 0.0018579603638499975, + 0.0020855723414570093, + 0.004001894034445286, + 0.0023520654067397118, + 0.002557662082836032, + 0.0025288849137723446, + 0.004597443155944347, + 0.005417690146714449, + 0.007107255049049854, + 0.01608472503721714, + 0.006699147168546915, + 0.008368328213691711, + 0.00773974834010005, + 0.006679781246930361, + 0.004786476492881775, + 0.010439841076731682, + 0.002980076242238283, + 0.011205008253455162, + 0.00411860179156065, + 0.00460855383425951, + 0.003925898112356663, + 0.01132564153522253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.016088075935840607 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003971189260483 + ], + "opposite_logprob_tokens": [ + -2.695039749145508 + ], + "logprob_reward": -0.07003971189260483, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003971189260483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13532011210918427, + 0.24111297726631165, + 0.050155919045209885, + 0.16202054917812347, + 0.13935457170009613, + 0.027790583670139313, + 0.0636707991361618, + 0.04725633189082146, + 0.05193791165947914, + 0.0813802182674408 + ], + "credit_map": [ + 0.01954057067632675, + 0.031380899250507355, + 0.011676447466015816, + 0.016767269000411034, + 0.011862829327583313, + 0.010689649730920792, + 0.020220281556248665, + 0.013182173483073711, + 0.0500820092856884, + 0.06443201750516891, + 0.022848034277558327, + 0.01721762865781784, + 0.014620793052017689, + 0.021347131580114365, + 0.02190270647406578, + 0.028662659227848053, + 0.008653505705296993, + 0.009428641758859158, + 0.006611440796405077, + 0.006868382450193167, + 0.0037063146010041237, + 0.0030912645161151886, + 0.005550867412239313, + 0.006245499011129141, + 0.01583622209727764, + 0.02539713680744171, + 0.03208460658788681, + 0.027380509302020073, + 0.009684612974524498, + 0.020041225478053093, + 0.01604515127837658, + 0.01555108092725277, + 0.015389356762170792, + 0.015962613746523857, + 0.0200071781873703, + 0.030739976093173027, + 0.011056914925575256, + 0.016595760360360146, + 0.015711406245827675, + 0.013891358859837055, + 0.0026361444033682346, + 0.003164742374792695, + 0.0029498557560145855, + 0.006038324907422066, + 0.002426787046715617, + 0.004187476355582476, + 0.002538766246289015, + 0.0038484851829707623, + 0.008299002423882484, + 0.005730911623686552, + 0.005304638762027025, + 0.0092350197955966, + 0.003439063671976328, + 0.006113869603723288, + 0.0058034504763782024, + 0.019744839519262314, + 0.004647589288651943, + 0.005574395414441824, + 0.005093993153423071, + 0.004995730705559254, + 0.009682711213827133, + 0.004779316484928131, + 0.005836352240294218, + 0.006646243389695883, + 0.006817343644797802, + 0.015010586939752102, + 0.005538172088563442, + 0.0037194218020886183, + 0.006042808759957552, + 0.006851637735962868, + 0.0036409967578947544, + 0.004316943231970072, + 0.005388225428760052, + 0.00785263255238533, + 0.004594812169671059, + 0.023168904706835747, + 0.006460770033299923, + 0.005451040808111429, + 0.0044859955087304115, + 0.023977836593985558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.07003971189260483 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817641779780388 + ], + "opposite_logprob_tokens": [ + -4.018176555633545 + ], + "logprob_reward": -0.01817641779780388, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817641779780388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11351438611745834, + 0.13370230793952942, + 0.09425857663154602, + 0.048569221049547195, + 0.3408564031124115, + 0.09397663176059723, + 0.12790584564208984, + 0.02376781590282917, + 0.009023002348840237, + 0.014425823464989662 + ], + "credit_map": [ + 0.026448071002960205, + 0.04532260820269585, + 0.004515801556408405, + 0.008381414227187634, + 0.0065618352964520454, + 0.006821082904934883, + 0.00976814515888691, + 0.005695431958884001, + 0.02088976837694645, + 0.027977416291832924, + 0.013744670897722244, + 0.015358530916273594, + 0.00660843588411808, + 0.012798987329006195, + 0.01362028531730175, + 0.022704219445586205, + 0.010876109823584557, + 0.0183852706104517, + 0.011752867139875889, + 0.009781035594642162, + 0.006254320498555899, + 0.008091219700872898, + 0.008029581047594547, + 0.021088173612952232, + 0.004015426617115736, + 0.005451568402349949, + 0.017598433420062065, + 0.004857373423874378, + 0.002921679522842169, + 0.004064519423991442, + 0.002470924286171794, + 0.007189295720309019, + 0.042251065373420715, + 0.04687783494591713, + 0.0363631397485733, + 0.06168677657842636, + 0.03584432601928711, + 0.0418110117316246, + 0.03981272503733635, + 0.03620953857898712, + 0.007768047042191029, + 0.010180304758250713, + 0.011349042877554893, + 0.02585902251303196, + 0.006099177990108728, + 0.01265136431902647, + 0.008417407982051373, + 0.011652269400656223, + 0.020312106236815453, + 0.011220956221222878, + 0.015003279782831669, + 0.019371062517166138, + 0.009343573823571205, + 0.015324447304010391, + 0.012692976742982864, + 0.024637453258037567, + 0.0029981741681694984, + 0.006443056743592024, + 0.0022705516312271357, + 0.00202995166182518, + 0.003368817502632737, + 0.002197558293119073, + 0.002125177765265107, + 0.002334528835490346, + 0.001168739516288042, + 0.0009058846044354141, + 0.0014716607984155416, + 0.0009446059120818973, + 0.0010923699010163546, + 0.0012765263672918081, + 0.0010540778748691082, + 0.0011091369669884443, + 0.0014133058721199632, + 0.0035021165385842323, + 0.0013195268111303449, + 0.0024106716737151146, + 0.0010459029581397772, + 0.0014849145663902164, + 0.0010650173062458634, + 0.0021843689028173685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.01817641779780388 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188366010785103 + ], + "opposite_logprob_tokens": [ + -4.264188289642334 + ], + "logprob_reward": -0.014188366010785103, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188366010785103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053498800843954086, + 0.2076580822467804, + 0.0747496709227562, + 0.15758270025253296, + 0.11650410294532776, + 0.07214895635843277, + 0.04326904937624931, + 0.11623885482549667, + 0.044616054743528366, + 0.11373375356197357 + ], + "credit_map": [ + 0.00840279646217823, + 0.011585643514990807, + 0.005170077085494995, + 0.006438559386879206, + 0.005315490532666445, + 0.004234333988279104, + 0.005613960791379213, + 0.006737936288118362, + 0.04770602658390999, + 0.06091945990920067, + 0.01901889778673649, + 0.015056888572871685, + 0.012516945600509644, + 0.013547374866902828, + 0.013459446839988232, + 0.025433029979467392, + 0.008100555278360844, + 0.015462253242731094, + 0.009930871427059174, + 0.00632104417309165, + 0.00564078101888299, + 0.005287153646349907, + 0.0057782987132668495, + 0.01822870597243309, + 0.012999230995774269, + 0.022519540041685104, + 0.038579463958740234, + 0.019691409543156624, + 0.012619436718523502, + 0.013282389380037785, + 0.008726727217435837, + 0.02916448377072811, + 0.011952865868806839, + 0.010813305154442787, + 0.01758989505469799, + 0.024122579023241997, + 0.0095596294850111, + 0.01700218766927719, + 0.012814214453101158, + 0.012649418786168098, + 0.00739605538547039, + 0.006359593942761421, + 0.010721665807068348, + 0.015209052711725235, + 0.003894846886396408, + 0.010296504944562912, + 0.005463823210448027, + 0.012807406485080719, + 0.005446346942335367, + 0.004816540516912937, + 0.007950611412525177, + 0.00540787260979414, + 0.002933529205620289, + 0.008329777978360653, + 0.0035471634473651648, + 0.004837206099182367, + 0.015072738751769066, + 0.013014535419642925, + 0.015296319499611855, + 0.008978847414255142, + 0.01528728287667036, + 0.014822968281805515, + 0.010902817361056805, + 0.022863339632749557, + 0.0059854560531675816, + 0.009319759905338287, + 0.005322312470525503, + 0.005197388119995594, + 0.00494906771928072, + 0.005175892263650894, + 0.003852794412523508, + 0.004813377745449543, + 0.0078105744905769825, + 0.007056214846670628, + 0.006013917736709118, + 0.042874667793512344, + 0.006731847301125526, + 0.005654287524521351, + 0.006324309855699539, + 0.0312679186463356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.014188366010785103 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, first-person perspective shot, rendered in a bright, highly saturated, flat-block art style, a pair of precision stainless steel tweezers, gleaming with a cool silver sheen, are held steadily by a pair of hands whose fingertips tremble slightly with intense concentration. The tips of the tweezers delicately grasp a tiny silver screw, barely the size of a grain of sand. The cross-shaped grooves on its head are clearly visible under the even light from an overhead shadowless lamp. As the camera slowly and steadily moves downward, the tweezers approach the green circuit board beneath, a structure composed of crisp geometric shapes within the phone. Golden circuit paths crisscross like sharp lines, interspersed with square electronic components in pure black, bright blue, and vivid red. The image, devoid of complex gradients or shading, exudes a clean, graphic aesthetic. Finally, the tweezers hover at their target, precisely aligning the tiny screw with a nearly invisible screw hole on the circuit board. The entire scene is imbued with extreme focus and a tense, technologically aesthetic atmosphere suspended in the moment.", + "seed": 10877, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_11/index_5849_seed_10877_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03180905757471919, + "vlm_logprob": -0.03180905757471919, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.03180905757471919, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01538868248462677 + ], + "opposite_logprob_tokens": [ + -4.765388488769531 + ], + "logprob_reward": -0.01538868248462677, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01538868248462677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046494171023368835, + 0.3991975486278534, + 0.027483943849802017, + 0.04011274874210358, + 0.032043710350990295, + 0.06958423554897308, + 0.03872103989124298, + 0.25636595487594604, + 0.04424120858311653, + 0.045755378901958466 + ], + "credit_map": [ + 0.0037261696998029947, + 0.005596964620053768, + 0.010885531082749367, + 0.007353822700679302, + 0.004903575871139765, + 0.004795634187757969, + 0.005555902142077684, + 0.003676577704027295, + 0.05508427321910858, + 0.07078629732131958, + 0.050573475658893585, + 0.06678614765405655, + 0.06401780992746353, + 0.023233478888869286, + 0.049429647624492645, + 0.019286468625068665, + 0.003654694650322199, + 0.004708605352789164, + 0.005339464638382196, + 0.003100964706391096, + 0.003619992407038808, + 0.0030263373628258705, + 0.0017799471970647573, + 0.0022539414931088686, + 0.004619840532541275, + 0.004626392386853695, + 0.01052304357290268, + 0.005199171137064695, + 0.0043708099983632565, + 0.0035543888807296753, + 0.003204063279554248, + 0.004015043843537569, + 0.0031835618428885937, + 0.0037583825178444386, + 0.004766539670526981, + 0.004778439644724131, + 0.0032770999241620302, + 0.0040056342259049416, + 0.00497414730489254, + 0.0032999112736433744, + 0.007470676675438881, + 0.008069196715950966, + 0.014009524136781693, + 0.00894186832010746, + 0.005498643033206463, + 0.010202953591942787, + 0.004837362561374903, + 0.010554017499089241, + 0.004460533149540424, + 0.0063543180003762245, + 0.00505988672375679, + 0.005649039521813393, + 0.0026434215251356363, + 0.006214241497218609, + 0.004009661730378866, + 0.00432994170114398, + 0.10779647529125214, + 0.06909948587417603, + 0.00720819691196084, + 0.013220755383372307, + 0.00857826229184866, + 0.026090851053595543, + 0.019630983471870422, + 0.004741006065160036, + 0.01520595047622919, + 0.004179155919700861, + 0.0035439222119748592, + 0.0052796220406889915, + 0.0017010937444865704, + 0.008773288689553738, + 0.002770389895886183, + 0.002787790959700942, + 0.008677002042531967, + 0.006094320211559534, + 0.0058512077666819096, + 0.005808601155877113, + 0.0028257451485842466, + 0.006188333034515381, + 0.005536757409572601, + 0.00477341515943408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 5 + ], + "selected_reward": -0.01538868248462677 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020603911951184273 + ], + "opposite_logprob_tokens": [ + -3.895603895187378 + ], + "logprob_reward": -0.020603911951184273, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020603911951184273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0654124841094017, + 0.17002727091312408, + 0.032426200807094574, + 0.049431271851062775, + 0.049681078642606735, + 0.08226616680622101, + 0.052392326295375824, + 0.304388165473938, + 0.153462216258049, + 0.04051275551319122 + ], + "credit_map": [ + 0.004206617828458548, + 0.007450487930327654, + 0.01104812603443861, + 0.011862475425004959, + 0.00879221223294735, + 0.006972832139581442, + 0.0096914516761899, + 0.005388279445469379, + 0.017108071595430374, + 0.02177688106894493, + 0.02284431643784046, + 0.033593498170375824, + 0.030918965116143227, + 0.011191283352673054, + 0.023283159360289574, + 0.009311092086136341, + 0.0039136759005486965, + 0.008156407624483109, + 0.004590987227857113, + 0.003415977116674185, + 0.005965553689748049, + 0.0023556570522487164, + 0.0018622158095240593, + 0.002165726851671934, + 0.006838085129857063, + 0.006364969536662102, + 0.008499384857714176, + 0.008703375235199928, + 0.00485954387113452, + 0.00393532682210207, + 0.0047964416444301605, + 0.005434143822640181, + 0.0044026137329638, + 0.004620500840246677, + 0.00712974276393652, + 0.009143667295575142, + 0.0035507723223417997, + 0.004771195352077484, + 0.009145393036305904, + 0.006917190738022327, + 0.008715989999473095, + 0.007293437607586384, + 0.012660525739192963, + 0.016987444832921028, + 0.004679467994719744, + 0.005616371985524893, + 0.008776228874921799, + 0.017536697909235954, + 0.005335988011211157, + 0.005026075057685375, + 0.009025155566632748, + 0.007068291772156954, + 0.004183290526270866, + 0.005070388782769442, + 0.007892535999417305, + 0.008790601044893265, + 0.13305354118347168, + 0.09782283753156662, + 0.008299623616039753, + 0.008986257947981358, + 0.012366732582449913, + 0.02656514197587967, + 0.011994658038020134, + 0.005299380514770746, + 0.04882374405860901, + 0.014012997038662434, + 0.008965437300503254, + 0.033768899738788605, + 0.007333012297749519, + 0.01876411959528923, + 0.015037592500448227, + 0.006756414659321308, + 0.010402529500424862, + 0.004913466051220894, + 0.003773247357457876, + 0.007043412886559963, + 0.003016113303601742, + 0.004435457289218903, + 0.004088124725967646, + 0.0028404048644006252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.020603911951184273 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790213868021965 + ], + "opposite_logprob_tokens": [ + -2.579021453857422 + ], + "logprob_reward": -0.0790213868021965, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790213868021965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045783355832099915, + 0.086904376745224, + 0.028753243386745453, + 0.028197329491376877, + 0.04171938821673393, + 0.10983086377382278, + 0.0667145699262619, + 0.5173675417900085, + 0.047906544059515, + 0.026822838932275772 + ], + "credit_map": [ + 0.003998131025582552, + 0.005528962705284357, + 0.010786590166389942, + 0.006848866119980812, + 0.005188696086406708, + 0.0037142811343073845, + 0.005289944354444742, + 0.0044278791174292564, + 0.009702997282147408, + 0.01342352107167244, + 0.011196253821253777, + 0.016120798885822296, + 0.01483870204538107, + 0.005995817482471466, + 0.010845335200428963, + 0.004780949559062719, + 0.003732529003173113, + 0.004635456018149853, + 0.004770760890096426, + 0.0030887629836797714, + 0.004398498684167862, + 0.0030431917402893305, + 0.001705441507510841, + 0.003378599649295211, + 0.003149941097944975, + 0.003756516147404909, + 0.009433414787054062, + 0.0029338118620216846, + 0.001745049492456019, + 0.0019136570626869798, + 0.0022759605199098587, + 0.0029889747966080904, + 0.004008623771369457, + 0.004374356474727392, + 0.0074087874963879585, + 0.007079254370182753, + 0.003549288958311081, + 0.004633627831935883, + 0.005949830636382103, + 0.004715615417808294, + 0.011815594509243965, + 0.009641531854867935, + 0.015054930932819843, + 0.01910487748682499, + 0.0061353049241006374, + 0.007578727789223194, + 0.013466575182974339, + 0.027033302932977676, + 0.0066711073741316795, + 0.007028254680335522, + 0.011460640467703342, + 0.00904815923422575, + 0.004579119849950075, + 0.005670220125466585, + 0.010236898437142372, + 0.012020166032016277, + 0.20085464417934418, + 0.20217770338058472, + 0.012917506508529186, + 0.015896141529083252, + 0.024050943553447723, + 0.0369686521589756, + 0.01698654517531395, + 0.007515366189181805, + 0.007002790458500385, + 0.0040376256220042706, + 0.005212161689996719, + 0.01180946733802557, + 0.002547742798924446, + 0.005860128905624151, + 0.007359817624092102, + 0.00407680869102478, + 0.0051110913045704365, + 0.0030209142714738846, + 0.0035575772635638714, + 0.002781756455078721, + 0.001887041493318975, + 0.003225465537980199, + 0.004065866582095623, + 0.0031731235794723034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.0790213868021965 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059450226835906506 + ], + "opposite_logprob_tokens": [ + -5.130945205688477 + ], + "logprob_reward": -0.0059450226835906506, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059450226835906506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1602054089307785, + 0.11632416397333145, + 0.07380955666303635, + 0.12984603643417358, + 0.05364178866147995, + 0.12730522453784943, + 0.0577690526843071, + 0.21717676520347595, + 0.03175891190767288, + 0.032163165509700775 + ], + "credit_map": [ + 0.012850831262767315, + 0.018900441005825996, + 0.038020290434360504, + 0.020646484568715096, + 0.017073705792427063, + 0.018708301708102226, + 0.019322054460644722, + 0.014683276414871216, + 0.015348316170275211, + 0.015107208862900734, + 0.021657947450876236, + 0.019502075389027596, + 0.016556451097130775, + 0.0070090400986373425, + 0.015180762857198715, + 0.0059623499400913715, + 0.007828525267541409, + 0.011436660774052143, + 0.011538898572325706, + 0.0070592439733445644, + 0.012180858291685581, + 0.009658011607825756, + 0.00688580609858036, + 0.007221540901809931, + 0.012551460415124893, + 0.01784754917025566, + 0.042195022106170654, + 0.014130028896033764, + 0.009297452867031097, + 0.010990682989358902, + 0.009781218133866787, + 0.013052614405751228, + 0.007132007274776697, + 0.006152348127216101, + 0.00943957082927227, + 0.009476925246417522, + 0.003448826726526022, + 0.006524667609483004, + 0.006347137037664652, + 0.005120298359543085, + 0.011791683733463287, + 0.011753951199352741, + 0.01704735867679119, + 0.02178594283759594, + 0.007877393625676632, + 0.013540655374526978, + 0.012911870144307613, + 0.03059634566307068, + 0.008074772544205189, + 0.008134345524013042, + 0.007174970582127571, + 0.008966579101979733, + 0.004269667435437441, + 0.0053390078246593475, + 0.009328476153314114, + 0.006481226533651352, + 0.08255904912948608, + 0.07082872092723846, + 0.007408453617244959, + 0.01210407167673111, + 0.009786006063222885, + 0.010808154009282589, + 0.017912276089191437, + 0.00577003788203001, + 0.006780763156712055, + 0.002300756052136421, + 0.0037266663275659084, + 0.002941344166174531, + 0.002105406718328595, + 0.005069192033261061, + 0.0038445356767624617, + 0.004990243352949619, + 0.004332097712904215, + 0.003424963681027293, + 0.005320030264556408, + 0.0038295097183436155, + 0.0023423878010362387, + 0.004073889460414648, + 0.004936544690281153, + 0.0039037419483065605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.0059450226835906506 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are precision stainless steel tweezers with a cool silver sheen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tweezers holding a tiny silver screw with visible cross-shaped grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move slowly and steadily downward toward a green circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a bright, highly saturated, flat-block art style without complex gradients?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808628395199776 + ], + "opposite_logprob_tokens": [ + -3.288086175918579 + ], + "logprob_reward": -0.03808628395199776, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808628395199776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04758076369762421, + 0.11583836376667023, + 0.019348176196217537, + 0.04504171013832092, + 0.025064511224627495, + 0.06420597434043884, + 0.027034854516386986, + 0.5931174755096436, + 0.030434919521212578, + 0.03233322873711586 + ], + "credit_map": [ + 0.004298071842640638, + 0.00549559760838747, + 0.010184906423091888, + 0.005934977903962135, + 0.005790112540125847, + 0.004276658408343792, + 0.0072144088335335255, + 0.004386031534522772, + 0.01658000610768795, + 0.023802518844604492, + 0.01285131648182869, + 0.018023988232016563, + 0.01902913674712181, + 0.007004203274846077, + 0.012436242774128914, + 0.006110954564064741, + 0.0027466900646686554, + 0.0025886620860546827, + 0.002767830854281783, + 0.002032961929216981, + 0.003106720047071576, + 0.0020771678537130356, + 0.002274360740557313, + 0.0017537830863147974, + 0.005265732761472464, + 0.005896976217627525, + 0.013228628784418106, + 0.005605665035545826, + 0.004453842528164387, + 0.003723636269569397, + 0.0032347433734685183, + 0.0036324821412563324, + 0.002702032681554556, + 0.0025032053235918283, + 0.004484924953430891, + 0.0040935007855296135, + 0.0018115206621587276, + 0.0021319612860679626, + 0.004039911087602377, + 0.0032974565401673317, + 0.004444760270416737, + 0.006903524044901133, + 0.009321746416389942, + 0.0172276608645916, + 0.0030180576723068953, + 0.006882945541292429, + 0.005775616969913244, + 0.01063165720552206, + 0.0032756426371634007, + 0.00425940565764904, + 0.003178667277097702, + 0.004178971983492374, + 0.0017576012760400772, + 0.0032455658074468374, + 0.0036214529536664486, + 0.00351754785515368, + 0.16651824116706848, + 0.24636445939540863, + 0.013740215450525284, + 0.016070948913693428, + 0.030292833223938942, + 0.09959246963262558, + 0.010922758840024471, + 0.009615568444132805, + 0.007141872309148312, + 0.0025161488447338343, + 0.004256007727235556, + 0.0031295588705688715, + 0.0015523156616836786, + 0.005539815407246351, + 0.0026456762570887804, + 0.00365352351218462, + 0.00760834151878953, + 0.0031556033063679934, + 0.0042741140350699425, + 0.004499808885157108, + 0.0014664932386949658, + 0.003929483704268932, + 0.00397413270547986, + 0.003425248898565769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.03808628395199776 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_12/index_tensor([4747, 4747])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_12/index_tensor([4747, 4747])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bb3e7404cd9196e460a938108a38a3a872a0849f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_12/index_tensor([4747, 4747])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, fixed wide shot frames a dark, storm-swept sky, swirling clouds occasionally streaked with purple lightning. A floating formation of countless luminous swords, like a giant dragon composed of blades, rages from the depths of the frame to the front right, swooping down at a speed that rips through the air. Each sword gleams with a cool metallic quality, its blade enveloped in a cyan-blue halo of energy, its core radiating a fiery golden glow. This vast formation shifts constantly in flight, sometimes converging into a massive drill, sometimes dispersing like a shower of flowers. Its collective trajectory carves countless sharp, defined lines and smooth arcs across the air, leaving magnificent, persistent streaks of light against the dim background. Each piercing blow unleashes a torrent of golden and blue particle effects, resembling stardust. These particles dance, collide, and slowly dissipate, leaving a spectacular energy trail. Illuminated by the light of the sword formation itself, the entire scene presents a highly dynamic contrast of light and shadow, illuminating the rugged black rock ground below and the rolling clouds in the sky, creating a tense and epic atmosphere full of destructive beauty, perfectly reproducing the ultimate visual impact and ultra-high quality of Ufotable animation.", + "seed": 25, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_12/index_4747_seed_25_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12637670529074968, + "vlm_logprob": -0.12637670529074968, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.12637670529074968, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.588451087474823 + ], + "opposite_logprob_tokens": [ + -0.838451087474823 + ], + "logprob_reward": -0.588451087474823, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.588451087474823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15484769642353058, + 0.08095701783895493, + 0.060057632625103, + 0.1362076997756958, + 0.05219876766204834, + 0.1402667760848999, + 0.14675745368003845, + 0.053560350090265274, + 0.09369385242462158, + 0.08145277947187424 + ], + "credit_map": [ + 0.0205661840736866, + 0.020738618448376656, + 0.022250549867749214, + 0.01878236047923565, + 0.011938275769352913, + 0.017801577225327492, + 0.01647745817899704, + 0.026292674243450165, + 0.012252471409738064, + 0.008948546834290028, + 0.006168243009597063, + 0.008312085643410683, + 0.010482831858098507, + 0.008803748525679111, + 0.018321149051189423, + 0.007667945232242346, + 0.006176292430609465, + 0.00898861326277256, + 0.009665549732744694, + 0.007335178088396788, + 0.008621260523796082, + 0.0076049091294407845, + 0.0060228873044252396, + 0.005642938893288374, + 0.011764061637222767, + 0.013935430906713009, + 0.02647424302995205, + 0.024573177099227905, + 0.013999064452946186, + 0.012301318347454071, + 0.015334272757172585, + 0.017826126888394356, + 0.005015847273170948, + 0.007498225197196007, + 0.004147730302065611, + 0.008275410160422325, + 0.008094562217593193, + 0.00488725071772933, + 0.006456900853663683, + 0.007822840474545956, + 0.012833654880523682, + 0.0385856069624424, + 0.01034566294401884, + 0.013236386701464653, + 0.024181745946407318, + 0.015437914989888668, + 0.01485070213675499, + 0.0107951071113348, + 0.01298728957772255, + 0.01601949892938137, + 0.01907953806221485, + 0.014437290839850903, + 0.02678685449063778, + 0.01978602260351181, + 0.02098260447382927, + 0.016678353771567345, + 0.010336391627788544, + 0.00593316275626421, + 0.00868002325296402, + 0.004693725612014532, + 0.007444052025675774, + 0.006960904225707054, + 0.004818258341401815, + 0.004693831317126751, + 0.0159237552434206, + 0.006558119785040617, + 0.020076271146535873, + 0.00849844515323639, + 0.010453274473547935, + 0.011970550753176212, + 0.008182714693248272, + 0.01203071977943182, + 0.011176157742738724, + 0.007540568709373474, + 0.008180019445717335, + 0.010416719131171703, + 0.006166033912450075, + 0.010889586061239243, + 0.010027647018432617, + 0.017056046053767204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.588451087474823 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636036090552807 + ], + "opposite_logprob_tokens": [ + -3.651360273361206 + ], + "logprob_reward": -0.02636036090552807, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636036090552807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06915063410997391, + 0.18647895753383636, + 0.01866079494357109, + 0.023252932354807854, + 0.05834142118692398, + 0.033683475106954575, + 0.13129489123821259, + 0.16116544604301453, + 0.16406868398189545, + 0.15390270948410034 + ], + "credit_map": [ + 0.00809297151863575, + 0.007458618842065334, + 0.006812335457652807, + 0.007469499949365854, + 0.00687347911298275, + 0.010378917679190636, + 0.008354555815458298, + 0.01371025200933218, + 0.02107677422463894, + 0.025592459365725517, + 0.022675305604934692, + 0.016749773174524307, + 0.027145633473992348, + 0.021808207035064697, + 0.02071135677397251, + 0.030719449743628502, + 0.00153651915024966, + 0.0027714697644114494, + 0.0018173422431573272, + 0.0021639710757881403, + 0.0033689539413899183, + 0.0021385876461863518, + 0.0020778225734829903, + 0.002786128781735897, + 0.0023817315232008696, + 0.002541126450523734, + 0.003116956213489175, + 0.003408156568184495, + 0.003421566216275096, + 0.002705766586586833, + 0.0029865799006074667, + 0.0026910495944321156, + 0.006127167958766222, + 0.008388693444430828, + 0.00622072909027338, + 0.005713167134672403, + 0.008316799066960812, + 0.005573834292590618, + 0.0063927206210792065, + 0.011608310043811798, + 0.003871277440339327, + 0.004470618907362223, + 0.0024203017819672823, + 0.0028576648328453302, + 0.005101809278130531, + 0.0037767295725643635, + 0.0042054252699017525, + 0.006979647558182478, + 0.01292562298476696, + 0.015129371546208858, + 0.012586544267833233, + 0.012970960699021816, + 0.027791233733296394, + 0.0164260882884264, + 0.021066386252641678, + 0.012398681603372097, + 0.02119624800980091, + 0.015657473355531693, + 0.018608221784234047, + 0.016599781811237335, + 0.02536975033581257, + 0.019323421642184258, + 0.021501949056982994, + 0.02290860004723072, + 0.025243127718567848, + 0.01138865016400814, + 0.029339518398046494, + 0.020636893808841705, + 0.016772035509347916, + 0.02123904414474964, + 0.018161684274673462, + 0.021287720650434494, + 0.022597791627049446, + 0.011515005491673946, + 0.021141739562153816, + 0.011900500394403934, + 0.0143894637003541, + 0.024744519963860512, + 0.01351074781268835, + 0.03410293161869049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.02636036090552807 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634140715003014 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.008634140715003014, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634140715003014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06842368841171265, + 0.09315793216228485, + 0.06146284192800522, + 0.04097127914428711, + 0.15296082198619843, + 0.056750621646642685, + 0.109528549015522, + 0.07790298014879227, + 0.15372341871261597, + 0.18511785566806793 + ], + "credit_map": [ + 0.007426389027386904, + 0.007841520942747593, + 0.008873417973518372, + 0.007116920780390501, + 0.00782753899693489, + 0.012218334712088108, + 0.006291184574365616, + 0.01082838699221611, + 0.011667470447719097, + 0.012400535866618156, + 0.01311470940709114, + 0.010975372977554798, + 0.009952126070857048, + 0.010752004571259022, + 0.012734860181808472, + 0.011560849845409393, + 0.004578666761517525, + 0.00935618206858635, + 0.006206026300787926, + 0.008165718987584114, + 0.01130302157253027, + 0.006927455775439739, + 0.009225302375853062, + 0.005700464826077223, + 0.004045700654387474, + 0.004540596157312393, + 0.007306290324777365, + 0.005539270583540201, + 0.004728373605757952, + 0.003928987309336662, + 0.005071394611150026, + 0.005810665898025036, + 0.014039980247616768, + 0.014945591799914837, + 0.01376995537430048, + 0.011788666248321533, + 0.019360162317752838, + 0.01403714157640934, + 0.019860494881868362, + 0.04515882954001427, + 0.005769215989857912, + 0.013005387969315052, + 0.006135500036180019, + 0.006300332024693489, + 0.006945902947336435, + 0.005988067016005516, + 0.006571184378117323, + 0.006035028025507927, + 0.010384303517639637, + 0.011837664991617203, + 0.01928265579044819, + 0.010985610075294971, + 0.018576601520180702, + 0.013428307138383389, + 0.01578076183795929, + 0.00925265159457922, + 0.01764748804271221, + 0.006805503740906715, + 0.0122603178024292, + 0.0058778803795576096, + 0.009693104773759842, + 0.01120749395340681, + 0.007079174742102623, + 0.007332023233175278, + 0.02117256075143814, + 0.011130469851195812, + 0.027935083955526352, + 0.015021992847323418, + 0.016224009916186333, + 0.015829863026738167, + 0.01706026867032051, + 0.029349176213145256, + 0.022023214027285576, + 0.030033431947231293, + 0.01573304273188114, + 0.02580985799431801, + 0.020645709708333015, + 0.023373935371637344, + 0.019332462921738625, + 0.02816620282828808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.008634140715003014 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031894787680357695 + ], + "opposite_logprob_tokens": [ + -5.753189563751221 + ], + "logprob_reward": -0.0031894787680357695, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031894787680357695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050876788794994354, + 0.04570343717932701, + 0.03971058875322342, + 0.046052150428295135, + 0.02802470326423645, + 0.2084597796201706, + 0.18334993720054626, + 0.11248838901519775, + 0.17672821879386902, + 0.10860603302717209 + ], + "credit_map": [ + 0.00496671674773097, + 0.005566297098994255, + 0.0062322067096829414, + 0.00611896812915802, + 0.007333680056035519, + 0.007132349070161581, + 0.0060401116497814655, + 0.00748645793646574, + 0.004757533315569162, + 0.006509165279567242, + 0.007048559375107288, + 0.005260583944618702, + 0.005610730033367872, + 0.00490548275411129, + 0.0052790045738220215, + 0.0063323769718408585, + 0.0030267962720245123, + 0.005411911755800247, + 0.004676090087741613, + 0.0037305511068552732, + 0.008539434522390366, + 0.0034431591629981995, + 0.004873879253864288, + 0.006008766125887632, + 0.004567711614072323, + 0.005015394184738398, + 0.004116331692785025, + 0.0061677973717451096, + 0.00777215464040637, + 0.00370394648052752, + 0.00615207152441144, + 0.008556739427149296, + 0.0027947709895670414, + 0.0043966928496956825, + 0.002724482910707593, + 0.004344284534454346, + 0.004269772209227085, + 0.002644261112436652, + 0.00383788556791842, + 0.0030125512275844812, + 0.01599060744047165, + 0.03790518641471863, + 0.017887510359287262, + 0.018194003030657768, + 0.03698713332414627, + 0.022142544388771057, + 0.02254156582057476, + 0.0368112176656723, + 0.018923230469226837, + 0.02126595750451088, + 0.017156312242150307, + 0.018529849126935005, + 0.03664059191942215, + 0.022911706939339638, + 0.02961484156548977, + 0.018307428807020187, + 0.011917361058294773, + 0.010374844074249268, + 0.015888411551713943, + 0.01043920312076807, + 0.018720436841249466, + 0.012933176010847092, + 0.015342514030635357, + 0.016872430220246315, + 0.026181858032941818, + 0.012286622077226639, + 0.033493440598249435, + 0.013690139167010784, + 0.018888788297772408, + 0.020889822393655777, + 0.014639373868703842, + 0.036658164113759995, + 0.010157558135688305, + 0.014241480268537998, + 0.010126473382115364, + 0.02066793292760849, + 0.010763251222670078, + 0.0118024917319417, + 0.016099002212285995, + 0.014747834764420986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.0031894787680357695 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248458590358496 + ], + "opposite_logprob_tokens": [ + -5.255248546600342 + ], + "logprob_reward": -0.005248458590358496, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248458590358496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1348649114370346, + 0.142963245511055, + 0.044368721544742584, + 0.03546162694692612, + 0.14447712898254395, + 0.08613718301057816, + 0.12673862278461456, + 0.08500184118747711, + 0.1292717009782791, + 0.07071500271558762 + ], + "credit_map": [ + 0.014154414646327496, + 0.012888598255813122, + 0.021960005164146423, + 0.020189087837934494, + 0.014122162014245987, + 0.014978070743381977, + 0.018243329599499702, + 0.018329262733459473, + 0.012047144584357738, + 0.019512558355927467, + 0.013241508975625038, + 0.015108698047697544, + 0.02346234954893589, + 0.01324540190398693, + 0.024594798684120178, + 0.021750807762145996, + 0.003416630672290921, + 0.005838241893798113, + 0.005532285198569298, + 0.006041578948497772, + 0.00740820961073041, + 0.0029051376041024923, + 0.0066612157970666885, + 0.00656542694196105, + 0.002939358353614807, + 0.004346159286797047, + 0.004579651169478893, + 0.0037144243251532316, + 0.005316391121596098, + 0.0038600375410169363, + 0.005284917540848255, + 0.005420692265033722, + 0.012781444936990738, + 0.019839072600007057, + 0.01599058136343956, + 0.02084314078092575, + 0.01971571147441864, + 0.013230660930275917, + 0.012587012723088264, + 0.02948951907455921, + 0.007968232966959476, + 0.008613495156168938, + 0.009391889907419682, + 0.008626355789601803, + 0.013428693637251854, + 0.009689796715974808, + 0.009460053406655788, + 0.01895865984261036, + 0.013993681408464909, + 0.01354854553937912, + 0.014375665225088596, + 0.01212343294173479, + 0.023104986175894737, + 0.01641177199780941, + 0.01967904344201088, + 0.013501517474651337, + 0.02401042729616165, + 0.007399988826364279, + 0.014297643676400185, + 0.006361375097185373, + 0.006326204631477594, + 0.013868318870663643, + 0.006511798594146967, + 0.006226089783012867, + 0.01768704503774643, + 0.009621907956898212, + 0.021842636168003082, + 0.016965683549642563, + 0.0170280858874321, + 0.018429601565003395, + 0.014100741595029831, + 0.013596005737781525, + 0.010776655748486519, + 0.007038197945803404, + 0.006899719592183828, + 0.00908729713410139, + 0.007856037467718124, + 0.01019138190895319, + 0.0071738953702151775, + 0.011691825464367867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.005248458590358496 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, fixed wide shot frames a dark, storm-swept sky, swirling clouds occasionally streaked with purple lightning. A floating formation of countless luminous swords, like a giant dragon composed of blades, rages from the depths of the frame to the front right, swooping down at a speed that rips through the air. Each sword gleams with a cool metallic quality, its blade enveloped in a cyan-blue halo of energy, its core radiating a fiery golden glow. This vast formation shifts constantly in flight, sometimes converging into a massive drill, sometimes dispersing like a shower of flowers. Its collective trajectory carves countless sharp, defined lines and smooth arcs across the air, leaving magnificent, persistent streaks of light against the dim background. Each piercing blow unleashes a torrent of golden and blue particle effects, resembling stardust. These particles dance, collide, and slowly dissipate, leaving a spectacular energy trail. Illuminated by the light of the sword formation itself, the entire scene presents a highly dynamic contrast of light and shadow, illuminating the rugged black rock ground below and the rolling clouds in the sky, creating a tense and epic atmosphere full of destructive beauty, perfectly reproducing the ultimate visual impact and ultra-high quality of Ufotable animation.", + "seed": 26, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_12/index_4747_seed_26_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05499596914742142, + "vlm_logprob": -0.05499596914742142, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.05499596914742142, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21302074193954468 + ], + "opposite_logprob_tokens": [ + -1.7130208015441895 + ], + "logprob_reward": -0.21302074193954468, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21302074193954468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6063361763954163, + 0.05585813149809837, + 0.02745797298848629, + 0.043490614742040634, + 0.04813990369439125, + 0.0726274624466896, + 0.036299098283052444, + 0.017184121534228325, + 0.06035887822508812, + 0.032247621566057205 + ], + "credit_map": [ + 0.023700837045907974, + 0.036775436252355576, + 0.1024416834115982, + 0.23092100024223328, + 0.025074997916817665, + 0.0951966941356659, + 0.05510912463068962, + 0.0371163934469223, + 0.004782121628522873, + 0.004741291981190443, + 0.010085268877446651, + 0.006443768739700317, + 0.006824431009590626, + 0.007625505328178406, + 0.006553776562213898, + 0.008801966905593872, + 0.0036570655647665262, + 0.004226800054311752, + 0.0030780541710555553, + 0.003775391262024641, + 0.0032427895348519087, + 0.0029472876340150833, + 0.003241019556298852, + 0.0032895656768232584, + 0.004451580811291933, + 0.00500372052192688, + 0.0037798339035362005, + 0.0044610584154725075, + 0.0044179377146065235, + 0.011457041837275028, + 0.004169974010437727, + 0.0057494682259857655, + 0.007858636789023876, + 0.0037977511528879404, + 0.008076038211584091, + 0.005454575642943382, + 0.005076015368103981, + 0.004204172641038895, + 0.006450849585235119, + 0.007221863605082035, + 0.010022834874689579, + 0.007611818611621857, + 0.014574320055544376, + 0.006243859883397818, + 0.005523318890482187, + 0.012103680521249771, + 0.006004350259900093, + 0.010543273761868477, + 0.00535458093509078, + 0.004706117790192366, + 0.00431792950257659, + 0.004198690410703421, + 0.003251563524827361, + 0.003789127804338932, + 0.007104182615876198, + 0.0035769056994467974, + 0.002695757197216153, + 0.001538338023237884, + 0.0032873668242245913, + 0.001685171970166266, + 0.0015724609838798642, + 0.001611025189049542, + 0.001926255994476378, + 0.0028677452355623245, + 0.007798381615430117, + 0.008614913560450077, + 0.008285867981612682, + 0.007263213396072388, + 0.00753859244287014, + 0.006264886353164911, + 0.005610974971204996, + 0.00898205116391182, + 0.003557066898792982, + 0.0026213419623672962, + 0.0037509046960622072, + 0.004075756296515465, + 0.004678779281675816, + 0.003369774203747511, + 0.006238078698515892, + 0.003955920226871967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.21302074193954468 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308018460869789 + ], + "opposite_logprob_tokens": [ + -3.1680800914764404 + ], + "logprob_reward": -0.04308018460869789, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308018460869789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1979082226753235, + 0.06574878096580505, + 0.019196592271327972, + 0.048875000327825546, + 0.10538560152053833, + 0.12535159289836884, + 0.07281491905450821, + 0.08481134474277496, + 0.14750973880290985, + 0.13239821791648865 + ], + "credit_map": [ + 0.018302040174603462, + 0.016740234568715096, + 0.02458314225077629, + 0.029438477009534836, + 0.01985715888440609, + 0.013737752102315426, + 0.026049897074699402, + 0.049199510365724564, + 0.004365133587270975, + 0.004754351917654276, + 0.007723631337285042, + 0.006507428828626871, + 0.006695539690554142, + 0.008630608208477497, + 0.010440677404403687, + 0.016631411388516426, + 0.0020871239248663187, + 0.0025799276772886515, + 0.001890100073069334, + 0.0025392218958586454, + 0.002637693425640464, + 0.002187316073104739, + 0.003139680717140436, + 0.0021355287171900272, + 0.00503990612924099, + 0.006617886945605278, + 0.006571496836841106, + 0.00536767253652215, + 0.004019498359411955, + 0.009440302848815918, + 0.004260104149580002, + 0.00755813205614686, + 0.02129705436527729, + 0.007794039789587259, + 0.017288168892264366, + 0.01004178449511528, + 0.008588263764977455, + 0.009812275879085064, + 0.00829740334302187, + 0.022266611456871033, + 0.01750340312719345, + 0.014245394617319107, + 0.02427092380821705, + 0.01084921695291996, + 0.009514683857560158, + 0.019393451511859894, + 0.011487149633467197, + 0.018087362870573997, + 0.00981274526566267, + 0.009954589419066906, + 0.010439482517540455, + 0.008421827107667923, + 0.005019559524953365, + 0.006972465198487043, + 0.01238314714282751, + 0.009811103343963623, + 0.01063590869307518, + 0.010720121674239635, + 0.013726680539548397, + 0.008761952631175518, + 0.0096168527379632, + 0.009768854826688766, + 0.008678993210196495, + 0.012901991605758667, + 0.020647501572966576, + 0.0226543340831995, + 0.021274523809552193, + 0.015637243166565895, + 0.017353925853967667, + 0.01689935475587845, + 0.011774354614317417, + 0.02126850187778473, + 0.025591682642698288, + 0.014603370800614357, + 0.016443287953734398, + 0.017067749053239822, + 0.00831450056284666, + 0.024681493639945984, + 0.008384731598198414, + 0.017311397939920425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.04308018460869789 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533704750239849 + ], + "opposite_logprob_tokens": [ + -4.387533664703369 + ], + "logprob_reward": -0.012533704750239849, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533704750239849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20531708002090454, + 0.08287889510393143, + 0.038455795496702194, + 0.09348995983600616, + 0.07317245751619339, + 0.27597576379776, + 0.05727270618081093, + 0.02449554204940796, + 0.07432466745376587, + 0.07461708784103394 + ], + "credit_map": [ + 0.009319252334535122, + 0.012487845495343208, + 0.0337502658367157, + 0.07610185444355011, + 0.010911643505096436, + 0.028814351186156273, + 0.02208069898188114, + 0.011851191520690918, + 0.00854106992483139, + 0.00985788181424141, + 0.00839379895478487, + 0.009074348956346512, + 0.011324253864586353, + 0.009951540268957615, + 0.010094347409904003, + 0.01564166508615017, + 0.005412846803665161, + 0.004883688874542713, + 0.0047040339559316635, + 0.004094249103218317, + 0.004578138701617718, + 0.005509425885975361, + 0.004801732022315264, + 0.004471685271710157, + 0.007432871498167515, + 0.010139457881450653, + 0.010135645046830177, + 0.010968859307467937, + 0.011685735546052456, + 0.01783227175474167, + 0.01398030761629343, + 0.011314816772937775, + 0.0063216062262654305, + 0.006639961153268814, + 0.006585813593119383, + 0.009862460196018219, + 0.011041978374123573, + 0.007122335955500603, + 0.014141617342829704, + 0.011456690728664398, + 0.04163447767496109, + 0.030709290876984596, + 0.05349736660718918, + 0.022710829973220825, + 0.019078930839896202, + 0.04426372051239014, + 0.023108094930648804, + 0.040973078459501266, + 0.006826540920883417, + 0.007513041142374277, + 0.005089606158435345, + 0.005974939092993736, + 0.009760620072484016, + 0.004528020042926073, + 0.011518439278006554, + 0.00606150459498167, + 0.0028565169777721167, + 0.0023370233830064535, + 0.0036576855927705765, + 0.002958003431558609, + 0.003465352114289999, + 0.0022235584910959005, + 0.003893497632816434, + 0.0031039067544043064, + 0.009947145357728004, + 0.010803056880831718, + 0.011935656890273094, + 0.006429337430745363, + 0.012386671267449856, + 0.007736979052424431, + 0.007346310652792454, + 0.007739511784166098, + 0.008893466554582119, + 0.007371561136096716, + 0.006508580408990383, + 0.009028968401253223, + 0.013373730704188347, + 0.008606127463281155, + 0.013253293000161648, + 0.007581368088722229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.012533704750239849 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017116909148171544 + ], + "opposite_logprob_tokens": [ + -6.376711845397949 + ], + "logprob_reward": -0.0017116909148171544, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017116909148171544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32976481318473816, + 0.06475905328989029, + 0.060114260762929916, + 0.1958325058221817, + 0.06750684976577759, + 0.058530740439891815, + 0.07503873854875565, + 0.04795381799340248, + 0.03153728321194649, + 0.0689619705080986 + ], + "credit_map": [ + 0.016397077590227127, + 0.026331758126616478, + 0.04700484871864319, + 0.1077822744846344, + 0.024713026359677315, + 0.044721703976392746, + 0.043223652988672256, + 0.019590474665164948, + 0.006410912144929171, + 0.008867268450558186, + 0.007916453294456005, + 0.007818850688636303, + 0.00927204079926014, + 0.007692219689488411, + 0.008217652328312397, + 0.008563653565943241, + 0.00789507757872343, + 0.007022376172244549, + 0.008572258986532688, + 0.009678555652499199, + 0.007320762611925602, + 0.006997328717261553, + 0.006226913537830114, + 0.006400985177606344, + 0.02183753438293934, + 0.022236596792936325, + 0.01810458116233349, + 0.023264320567250252, + 0.022247614338994026, + 0.04602742940187454, + 0.023452164605259895, + 0.01866227202117443, + 0.01569989137351513, + 0.00564454635605216, + 0.008151265792548656, + 0.005874055903404951, + 0.006286830175668001, + 0.006286135874688625, + 0.006329894065856934, + 0.01323422696441412, + 0.0075412713922560215, + 0.005873358342796564, + 0.011754224076867104, + 0.005009265150874853, + 0.004894762299954891, + 0.009527304209768772, + 0.005998876411467791, + 0.007931675761938095, + 0.008662664331495762, + 0.008256036788225174, + 0.005633693654090166, + 0.009401514194905758, + 0.012095835991203785, + 0.00592206371948123, + 0.018160711973905563, + 0.006906215567141771, + 0.005528628826141357, + 0.005411821883171797, + 0.00814441405236721, + 0.004755556583404541, + 0.006316713523119688, + 0.00500181270763278, + 0.006328251212835312, + 0.006466619670391083, + 0.004146905150264502, + 0.004442165605723858, + 0.004420327953994274, + 0.003225838067010045, + 0.0026072897017002106, + 0.004054336342960596, + 0.004711390472948551, + 0.003929028753191233, + 0.012070483528077602, + 0.0071436441503465176, + 0.0090651735663414, + 0.009476796723902225, + 0.006309780292212963, + 0.008331947959959507, + 0.006626366171985865, + 0.009937778115272522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.0017116909148171544 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633523523807526 + ], + "opposite_logprob_tokens": [ + -5.37963342666626 + ], + "logprob_reward": -0.004633523523807526, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633523523807526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4079647958278656, + 0.03740870580077171, + 0.04371924325823784, + 0.055399488657712936, + 0.053323034197092056, + 0.050892047584056854, + 0.05684072524309158, + 0.13634289801120758, + 0.0428595133125782, + 0.1152496188879013 + ], + "credit_map": [ + 0.031048445031046867, + 0.04249078035354614, + 0.05492482706904411, + 0.08293735235929489, + 0.042788777500391006, + 0.026963375508785248, + 0.08915714174509048, + 0.0376540832221508, + 0.004245184827595949, + 0.0045273806899785995, + 0.0036866115406155586, + 0.0034028179943561554, + 0.005426340736448765, + 0.004962600767612457, + 0.005612948909401894, + 0.005544815678149462, + 0.007268141023814678, + 0.006195272319018841, + 0.00530555984005332, + 0.005078981164842844, + 0.0041644275188446045, + 0.005210309289395809, + 0.004049122333526611, + 0.006447423715144396, + 0.006377301644533873, + 0.006109266541898251, + 0.0052395174279809, + 0.005670978222042322, + 0.005538629833608866, + 0.011872443370521069, + 0.0073763285763561726, + 0.007215017452836037, + 0.008212766610085964, + 0.0054824138060212135, + 0.008285225369036198, + 0.007028440944850445, + 0.005916875321418047, + 0.004780087154358625, + 0.0052032326348125935, + 0.00841398723423481, + 0.006008519325405359, + 0.005471888463944197, + 0.006649751681834459, + 0.004643649328500032, + 0.005335765890777111, + 0.007718938868492842, + 0.008232616819441319, + 0.006830913946032524, + 0.007065516896545887, + 0.008257953450083733, + 0.007018363568931818, + 0.00677085854113102, + 0.006780768744647503, + 0.00578284589573741, + 0.008501887321472168, + 0.006662521976977587, + 0.01501193456351757, + 0.018608571961522102, + 0.013293680734932423, + 0.017835814505815506, + 0.0222589373588562, + 0.01254492811858654, + 0.023350007832050323, + 0.013439012691378593, + 0.006566967349499464, + 0.004248335957527161, + 0.006620902568101883, + 0.005051529500633478, + 0.0046241856180131435, + 0.0033432405907660723, + 0.005425244104117155, + 0.00697910413146019, + 0.021328527480363846, + 0.015965884551405907, + 0.010955884121358395, + 0.015336180105805397, + 0.010398867540061474, + 0.017144568264484406, + 0.009268730878829956, + 0.014850970357656479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.004633523523807526 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, fixed wide shot frames a dark, storm-swept sky, swirling clouds occasionally streaked with purple lightning. A floating formation of countless luminous swords, like a giant dragon composed of blades, rages from the depths of the frame to the front right, swooping down at a speed that rips through the air. Each sword gleams with a cool metallic quality, its blade enveloped in a cyan-blue halo of energy, its core radiating a fiery golden glow. This vast formation shifts constantly in flight, sometimes converging into a massive drill, sometimes dispersing like a shower of flowers. Its collective trajectory carves countless sharp, defined lines and smooth arcs across the air, leaving magnificent, persistent streaks of light against the dim background. Each piercing blow unleashes a torrent of golden and blue particle effects, resembling stardust. These particles dance, collide, and slowly dissipate, leaving a spectacular energy trail. Illuminated by the light of the sword formation itself, the entire scene presents a highly dynamic contrast of light and shadow, illuminating the rugged black rock ground below and the rolling clouds in the sky, creating a tense and epic atmosphere full of destructive beauty, perfectly reproducing the ultimate visual impact and ultra-high quality of Ufotable animation.", + "seed": 27, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_12/index_4747_seed_27_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14803590294905006, + "vlm_logprob": -0.14803590294905006, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.14803590294905006, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7172973155975342 + ], + "opposite_logprob_tokens": [ + -0.7172973155975342 + ], + "logprob_reward": -0.7172973155975342, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7172973155975342, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.6126997470855713, + 0.020145850256085396, + 0.012605656869709492, + 0.04643763229250908, + 0.24186323583126068, + 0.016967589035630226, + 0.009307648055255413, + 0.008399569429457188, + 0.005093224812299013, + 0.02647988311946392 + ], + "credit_map": [ + 0.020704258233308792, + 0.03104740008711815, + 0.08613013476133347, + 0.011899800039827824, + 0.17675378918647766, + 0.24651432037353516, + 0.0162037406116724, + 0.023446308448910713, + 0.0016528433188796043, + 0.00447474280372262, + 0.001725271693430841, + 0.0025288723409175873, + 0.005096758250147104, + 0.0017737305024638772, + 0.001512214308604598, + 0.0013814172707498074, + 0.0012749157613143325, + 0.0017411779845133424, + 0.0011509231990203261, + 0.0014100249391049147, + 0.0018927452620118856, + 0.0020704378839582205, + 0.0012576774461194873, + 0.0018077543936669827, + 0.006766460835933685, + 0.007313018199056387, + 0.005316735245287418, + 0.0046069729141891, + 0.005923171527683735, + 0.005320461001247168, + 0.004561918321996927, + 0.006628892384469509, + 0.036045148968696594, + 0.030938999727368355, + 0.01808296889066696, + 0.024610858410596848, + 0.04277149960398674, + 0.03175264596939087, + 0.0362975038588047, + 0.02136361412703991, + 0.001461711130104959, + 0.00335478107444942, + 0.000798811495769769, + 0.0011115903034806252, + 0.0028190661687403917, + 0.002056883880868554, + 0.0024639195762574673, + 0.0029008251149207354, + 0.0010138876968994737, + 0.0012025032192468643, + 0.0007919573108665645, + 0.0012197012547403574, + 0.0009403181611560285, + 0.001101456000469625, + 0.0015966573264449835, + 0.0014411674346774817, + 0.0008216488058678806, + 0.001057296060025692, + 0.0007720513385720551, + 0.001176518271677196, + 0.0012436318211257458, + 0.001226966269314289, + 0.0008844581316225231, + 0.0012169982073828578, + 0.00034838737337850034, + 0.0005455997888930142, + 0.0005597954732365906, + 0.0006864583701826632, + 0.0009342756238766015, + 0.0006134166033007205, + 0.0006731990724802017, + 0.0007320926524698734, + 0.0024523173924535513, + 0.0026478904765099287, + 0.002923924010246992, + 0.001782803563401103, + 0.0050580683164298534, + 0.00310596008785069, + 0.004344166722148657, + 0.004164751619100571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.7172973155975342 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012566314078867435 + ], + "opposite_logprob_tokens": [ + -4.387566089630127 + ], + "logprob_reward": -0.012566314078867435, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012566314078867435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9120672941207886, + 0.009178034961223602, + 0.004825687035918236, + 0.007943494245409966, + 0.017928700894117355, + 0.011559330858290195, + 0.008704856038093567, + 0.009416714310646057, + 0.0029557072557508945, + 0.015420163981616497 + ], + "credit_map": [ + 0.030100839212536812, + 0.0469905249774456, + 0.12964335083961487, + 0.015977460891008377, + 0.2684418261051178, + 0.371090829372406, + 0.023354778066277504, + 0.026467693969607353, + 0.0008796710171736777, + 0.0015802853740751743, + 0.0009024112368933856, + 0.0012566044460982084, + 0.001600344548933208, + 0.0012306090211495757, + 0.000799923378508538, + 0.0009281857637688518, + 0.0003962481569033116, + 0.0007749897195026278, + 0.0006167968967929482, + 0.000729224004317075, + 0.0007753960089758039, + 0.0005963349249213934, + 0.0005131801008246839, + 0.00042351725278422236, + 0.00133583415299654, + 0.0015279814833775163, + 0.0006133719580247998, + 0.0007693571387790143, + 0.0009130114340223372, + 0.0011468189768493176, + 0.0007526371045969427, + 0.000884481705725193, + 0.002834250219166279, + 0.0021326078567653894, + 0.001409650780260563, + 0.0018698457861319184, + 0.0028983561787754297, + 0.0024606389924883842, + 0.0027027688920497894, + 0.001620583701878786, + 0.0011326681124046445, + 0.0017924566054716706, + 0.0007637593662366271, + 0.0008302335627377033, + 0.0017956699011847377, + 0.0015028357738628983, + 0.0015599927864968777, + 0.0021817146334797144, + 0.0005609391955658793, + 0.0011826492846012115, + 0.0004278919950593263, + 0.0006464034668169916, + 0.0011883678380399942, + 0.0020767883397638798, + 0.0013509498676285148, + 0.0012708669528365135, + 0.0008257355075329542, + 0.0013285221066325903, + 0.0007955673499964178, + 0.001198944402858615, + 0.0013610197929665446, + 0.0011983911972492933, + 0.0006949189119040966, + 0.0020136150997132063, + 0.00023175528622232378, + 0.00035960940294899046, + 0.0003305971622467041, + 0.0004153521149419248, + 0.00041739686275832355, + 0.0002329630369786173, + 0.0005713491118513048, + 0.0003966841904912144, + 0.0017260507447645068, + 0.001355862244963646, + 0.002429203363135457, + 0.0010257220128551126, + 0.0020281742326915264, + 0.0020704406779259443, + 0.002187843434512615, + 0.002596867736428976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.012566314078867435 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046386257745325565 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.0046386257745325565, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046386257745325565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9072594046592712, + 0.005658592097461224, + 0.011478460393846035, + 0.005148718599230051, + 0.016940169036388397, + 0.004565353970974684, + 0.014671054668724537, + 0.004999472759664059, + 0.02497001178562641, + 0.004308740142732859 + ], + "credit_map": [ + 0.029951728880405426, + 0.04679973050951958, + 0.12811490893363953, + 0.015398599207401276, + 0.267151802778244, + 0.36790022253990173, + 0.023156147450208664, + 0.028786268085241318, + 0.000446121470304206, + 0.0009777677478268743, + 0.0007579215453006327, + 0.0007375281420536339, + 0.0012710767332464457, + 0.0006036166450940073, + 0.00047075957991182804, + 0.0003938004083465785, + 0.0008428056607954204, + 0.0017595457611605525, + 0.0006731462199240923, + 0.0012684614630416036, + 0.0024435115046799183, + 0.001432795193977654, + 0.0019729179330170155, + 0.0010852768318727612, + 0.0006666952394880354, + 0.0007751231314614415, + 0.000487154524307698, + 0.0006050841766409576, + 0.0007457974134013057, + 0.0005697659798897803, + 0.0006839688285253942, + 0.0006151290726847947, + 0.0025185050908476114, + 0.002282812725752592, + 0.0012404703302308917, + 0.001754110911861062, + 0.003130650846287608, + 0.002221213188022375, + 0.0023209303617477417, + 0.0014714754652231932, + 0.00047881348291411996, + 0.0006459839642047882, + 0.00024721812224015594, + 0.00042913382640108466, + 0.0005234726704657078, + 0.000485454365843907, + 0.001173979020677507, + 0.0005812987219542265, + 0.0010056800674647093, + 0.0019270841730758548, + 0.000984790618531406, + 0.0017371759749948978, + 0.001390295336022973, + 0.003112263046205044, + 0.002474317792803049, + 0.0020394488237798214, + 0.0004688815970439464, + 0.0006545954965986311, + 0.00043859914876520634, + 0.0005167669150978327, + 0.0005967047763988376, + 0.0007343380129896104, + 0.00047758605796843767, + 0.0011120010167360306, + 0.0020627458579838276, + 0.0025743322912603617, + 0.003245164640247822, + 0.002913576550781727, + 0.004522471688687801, + 0.002917282050475478, + 0.0031282694544643164, + 0.003606169717386365, + 0.0005404087132774293, + 0.0004074160533491522, + 0.00048713403521105647, + 0.0003071960818488151, + 0.0006594525184482336, + 0.0005273952847346663, + 0.0005313572473824024, + 0.0008483800920657814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.0046386257745325565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024873053189367056 + ], + "opposite_logprob_tokens": [ + -6.0024871826171875 + ], + "logprob_reward": -0.0024873053189367056, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024873053189367056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7655392289161682, + 0.015126679092645645, + 0.01677689701318741, + 0.026410793885588646, + 0.03508398309350014, + 0.032361503690481186, + 0.07776103913784027, + 0.0122644929215312, + 0.007448784075677395, + 0.01122650783509016 + ], + "credit_map": [ + 0.025098761543631554, + 0.03974476829171181, + 0.1065119281411171, + 0.01363370567560196, + 0.2248784452676773, + 0.3078679144382477, + 0.019947608932852745, + 0.027856074273586273, + 0.001546933432109654, + 0.0017447508871555328, + 0.0023235827684402466, + 0.0021834177896380424, + 0.002563260495662689, + 0.0016814941773191094, + 0.0015192473074421287, + 0.0015639927005395293, + 0.0017259924206882715, + 0.002533606020733714, + 0.0014895376516506076, + 0.0013995053013786674, + 0.0034946619998663664, + 0.003061249852180481, + 0.001650524907745421, + 0.0014218180440366268, + 0.004915506113320589, + 0.003116502659395337, + 0.002466910518705845, + 0.0025683592539280653, + 0.002894927514716983, + 0.002652514958754182, + 0.00212629372254014, + 0.005669778678566217, + 0.005316554568707943, + 0.004750131629407406, + 0.002538844011723995, + 0.004090543370693922, + 0.005101718008518219, + 0.0048944479785859585, + 0.005157445557415485, + 0.003234298201277852, + 0.0026810921262949705, + 0.004880162887275219, + 0.0017536397790536284, + 0.0022684342693537474, + 0.005719834007322788, + 0.004152040462940931, + 0.005411224905401468, + 0.005495074205100536, + 0.005211882293224335, + 0.010851653292775154, + 0.003889726009219885, + 0.006262645591050386, + 0.008742055855691433, + 0.02384168468415737, + 0.010212082415819168, + 0.008749312721192837, + 0.0012122546322643757, + 0.001601000432856381, + 0.0013844543136656284, + 0.0019172105239704251, + 0.0016094498569145799, + 0.0014540957054123282, + 0.001562772085890174, + 0.00152325548697263, + 0.0005791013827547431, + 0.0009009068016894162, + 0.0007793722325004637, + 0.0007532399031333625, + 0.001016402617096901, + 0.0009440330904908478, + 0.001291984342969954, + 0.0011837441707029939, + 0.0012927487259730697, + 0.0009465148323215544, + 0.00157441187184304, + 0.0007969162543304265, + 0.0014517115196213126, + 0.0015433485386893153, + 0.0015577190788462758, + 0.002063137013465166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0024873053189367056 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003189953975379467 + ], + "opposite_logprob_tokens": [ + -5.753190040588379 + ], + "logprob_reward": -0.003189953975379467, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003189953975379467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4519297480583191, + 0.05359867960214615, + 0.03910285606980324, + 0.051456253975629807, + 0.1296399086713791, + 0.03953218087553978, + 0.08312994986772537, + 0.03765449300408363, + 0.042962878942489624, + 0.0709930956363678 + ], + "credit_map": [ + 0.015961719676852226, + 0.022387266159057617, + 0.06313769519329071, + 0.009913221932947636, + 0.12905022501945496, + 0.18057774007320404, + 0.012047350406646729, + 0.018854467198252678, + 0.004119174554944038, + 0.010589037090539932, + 0.007645845413208008, + 0.00592496944591403, + 0.011642594821751118, + 0.004698934033513069, + 0.004754258785396814, + 0.004223860800266266, + 0.004281378351151943, + 0.006758158560842276, + 0.003122056135907769, + 0.0030289278365671635, + 0.008365840651094913, + 0.0073503004387021065, + 0.0034335844684392214, + 0.0027626045048236847, + 0.0062752035446465015, + 0.006553678773343563, + 0.003930262289941311, + 0.004863181617110968, + 0.0075459713116288185, + 0.006951435934752226, + 0.006537377368658781, + 0.00879913941025734, + 0.020301856100559235, + 0.019381018355488777, + 0.00937406811863184, + 0.012906397692859173, + 0.023829149082303047, + 0.018114592880010605, + 0.014612390659749508, + 0.011120419949293137, + 0.003555091330781579, + 0.005513729993253946, + 0.002311095828190446, + 0.004364040680229664, + 0.006800858769565821, + 0.005606560502201319, + 0.006208919454365969, + 0.005171880125999451, + 0.005637820810079575, + 0.011437428183853626, + 0.0063467081636190414, + 0.009416736662387848, + 0.00864923931658268, + 0.009461374022066593, + 0.01821080408990383, + 0.013969823718070984, + 0.003754647681489587, + 0.004864852875471115, + 0.004032427910715342, + 0.004743242636322975, + 0.006049321498721838, + 0.005866548046469688, + 0.004351573996245861, + 0.003991876728832722, + 0.0031742644496262074, + 0.0044352151453495026, + 0.005115637090057135, + 0.0045494502410292625, + 0.008052971214056015, + 0.004663779400289059, + 0.00638906704261899, + 0.00658248970285058, + 0.008545939810574055, + 0.00653827702626586, + 0.009568260982632637, + 0.0046909828670322895, + 0.009833873249590397, + 0.008678135462105274, + 0.010795429348945618, + 0.012342186644673347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.003189953975379467 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, fixed wide shot frames a dark, storm-swept sky, swirling clouds occasionally streaked with purple lightning. A floating formation of countless luminous swords, like a giant dragon composed of blades, rages from the depths of the frame to the front right, swooping down at a speed that rips through the air. Each sword gleams with a cool metallic quality, its blade enveloped in a cyan-blue halo of energy, its core radiating a fiery golden glow. This vast formation shifts constantly in flight, sometimes converging into a massive drill, sometimes dispersing like a shower of flowers. Its collective trajectory carves countless sharp, defined lines and smooth arcs across the air, leaving magnificent, persistent streaks of light against the dim background. Each piercing blow unleashes a torrent of golden and blue particle effects, resembling stardust. These particles dance, collide, and slowly dissipate, leaving a spectacular energy trail. Illuminated by the light of the sword formation itself, the entire scene presents a highly dynamic contrast of light and shadow, illuminating the rugged black rock ground below and the rolling clouds in the sky, creating a tense and epic atmosphere full of destructive beauty, perfectly reproducing the ultimate visual impact and ultra-high quality of Ufotable animation.", + "seed": 28, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_12/index_4747_seed_28_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05558608383871615, + "vlm_logprob": -0.05558608383871615, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.05558608383871615, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16667743027210236 + ], + "opposite_logprob_tokens": [ + -1.916677474975586 + ], + "logprob_reward": -0.16667743027210236, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16667743027210236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08082719892263412, + 0.030710741877555847, + 0.04195639491081238, + 0.05686405673623085, + 0.11603055894374847, + 0.039590492844581604, + 0.036531317979097366, + 0.10669919848442078, + 0.14754721522331238, + 0.3432428538799286 + ], + "credit_map": [ + 0.006443116348236799, + 0.006867388263344765, + 0.011614084243774414, + 0.01969997026026249, + 0.006843669805675745, + 0.007014268543571234, + 0.012883572839200497, + 0.009461124427616596, + 0.004568631295114756, + 0.0031787105835974216, + 0.005886169150471687, + 0.004077796824276447, + 0.0034345074091106653, + 0.0024405259173363447, + 0.004278700333088636, + 0.0028457031585276127, + 0.004231334198266268, + 0.0038581471890211105, + 0.0076217930763959885, + 0.007112063001841307, + 0.005204963497817516, + 0.003257396165281534, + 0.004758111666887999, + 0.0059125893749296665, + 0.00351201300509274, + 0.0040539042092859745, + 0.010360130108892918, + 0.0032583200372755527, + 0.005826584063470364, + 0.005851855035871267, + 0.01949368417263031, + 0.004507563076913357, + 0.0074380142614245415, + 0.0042170328088104725, + 0.025604484602808952, + 0.005001982674002647, + 0.00721063744276762, + 0.006179925985634327, + 0.03912762552499771, + 0.021250851452350616, + 0.005910668522119522, + 0.0023346664384007454, + 0.007334509398788214, + 0.0040932297706604, + 0.0033637373708188534, + 0.0034456741996109486, + 0.008834577165544033, + 0.004273429047316313, + 0.005316630471497774, + 0.0036914923693984747, + 0.005349306855350733, + 0.00438065268099308, + 0.002939461264759302, + 0.0039250198751688, + 0.0065711322240531445, + 0.004357622936367989, + 0.011447431519627571, + 0.014615929685533047, + 0.024012112990021706, + 0.012738263234496117, + 0.008619870990514755, + 0.008299773558974266, + 0.014520280063152313, + 0.012445535510778427, + 0.021121501922607422, + 0.01389195118099451, + 0.03560014069080353, + 0.010017217136919498, + 0.010659069754183292, + 0.010640514083206654, + 0.030369659885764122, + 0.015247170813381672, + 0.05192283168435097, + 0.035930100828409195, + 0.0544469952583313, + 0.046229925006628036, + 0.04195347800850868, + 0.035503990948200226, + 0.03801774978637695, + 0.03923778235912323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.16667743027210236 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033608973026276 + ], + "opposite_logprob_tokens": [ + -2.3503360748291016 + ], + "logprob_reward": -0.10033608973026276, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033608973026276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33502548933029175, + 0.07210400700569153, + 0.19390170276165009, + 0.09810417145490646, + 0.03128657117486, + 0.07821597903966904, + 0.0342591218650341, + 0.03861679509282112, + 0.09247935563325882, + 0.026006853207945824 + ], + "credit_map": [ + 0.03257925435900688, + 0.025337541475892067, + 0.04820745438337326, + 0.09026427567005157, + 0.020600751042366028, + 0.025007417425513268, + 0.05212777480483055, + 0.04090103134512901, + 0.010369185358285904, + 0.008543399162590504, + 0.011525428853929043, + 0.009849277324974537, + 0.0062176831997931, + 0.007058861665427685, + 0.009216368198394775, + 0.00932379812002182, + 0.020297477021813393, + 0.01276586763560772, + 0.028606312349438667, + 0.03313879668712616, + 0.01642407290637493, + 0.013794753700494766, + 0.04955744370818138, + 0.019316980615258217, + 0.011809439398348331, + 0.007241670973598957, + 0.028572186827659607, + 0.00910921860486269, + 0.006901859305799007, + 0.00909406691789627, + 0.013015670701861382, + 0.012360055930912495, + 0.005145885515958071, + 0.002763401484116912, + 0.00522063160315156, + 0.003857484320178628, + 0.0031529769767075777, + 0.0028375990223139524, + 0.004592212848365307, + 0.0037163803353905678, + 0.011796264909207821, + 0.0047188811004161835, + 0.018710996955633163, + 0.007752371486276388, + 0.00584552763029933, + 0.0077970875427126884, + 0.010019063949584961, + 0.011575784534215927, + 0.004402305465191603, + 0.0031012780964374542, + 0.005530941765755415, + 0.005056560039520264, + 0.002521909773349762, + 0.004159119445830584, + 0.004759299103170633, + 0.004727710969746113, + 0.006003251299262047, + 0.00407074298709631, + 0.00714827561751008, + 0.003946887794882059, + 0.0035362071357667446, + 0.003405242459848523, + 0.006150452420115471, + 0.00435573561117053, + 0.015101132914423943, + 0.010796433314681053, + 0.0177510567009449, + 0.007868335582315922, + 0.007298056036233902, + 0.007730289362370968, + 0.012140156701207161, + 0.013793901540338993, + 0.00359334796667099, + 0.0026456224732100964, + 0.005349033046513796, + 0.0028281500563025475, + 0.0028105296660214663, + 0.002381471451371908, + 0.003147902898490429, + 0.003250795416533947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.10033608973026276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944785661995411 + ], + "opposite_logprob_tokens": [ + -5.130944728851318 + ], + "logprob_reward": -0.005944785661995411, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944785661995411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09506073594093323, + 0.07631904631853104, + 0.07001735270023346, + 0.06735217571258545, + 0.16991201043128967, + 0.08836452662944794, + 0.10566093772649765, + 0.08009457588195801, + 0.09382300823926926, + 0.15339556336402893 + ], + "credit_map": [ + 0.009684275835752487, + 0.007453898899257183, + 0.014361291192471981, + 0.026289453729987144, + 0.008494715206325054, + 0.007133869454264641, + 0.011749276891350746, + 0.009893963113427162, + 0.007968631573021412, + 0.007137826178222895, + 0.008945249021053314, + 0.009276973083615303, + 0.008631176315248013, + 0.007996611297130585, + 0.017471330240368843, + 0.008891256526112556, + 0.01043196115642786, + 0.00809349026530981, + 0.010483261197805405, + 0.011135878972709179, + 0.0068383910693228245, + 0.004644959699362516, + 0.009630725719034672, + 0.00875869207084179, + 0.004500555340200663, + 0.0040267230942845345, + 0.017986780032515526, + 0.004351986572146416, + 0.004173761233687401, + 0.0065511236898601055, + 0.020120063796639442, + 0.005641186609864235, + 0.008596602827310562, + 0.006258220877498388, + 0.022940289229154587, + 0.0055384826846420765, + 0.010534057393670082, + 0.007574101444333792, + 0.06669250875711441, + 0.04177776351571083, + 0.010284841060638428, + 0.005651477724313736, + 0.01160281803458929, + 0.009438050910830498, + 0.007743646390736103, + 0.011555483564734459, + 0.020966215059161186, + 0.011122005991637707, + 0.012720752507448196, + 0.010010426864027977, + 0.0145584000274539, + 0.011561844497919083, + 0.011745551601052284, + 0.012562289834022522, + 0.011842286214232445, + 0.020659398287534714, + 0.011969522573053837, + 0.011280643753707409, + 0.02137145958840847, + 0.007100276183336973, + 0.005991107318550348, + 0.006671725306659937, + 0.00832733791321516, + 0.007382514420896769, + 0.012169103138148785, + 0.011396963149309158, + 0.013905334286391735, + 0.01070188544690609, + 0.010988052934408188, + 0.008375750854611397, + 0.014421285130083561, + 0.01186464074999094, + 0.022205667570233345, + 0.01706286519765854, + 0.02220229059457779, + 0.01971249096095562, + 0.021625781431794167, + 0.016532940790057182, + 0.017335282638669014, + 0.016718251630663872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.005944785661995411 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024862352292984724 + ], + "opposite_logprob_tokens": [ + -6.002486228942871 + ], + "logprob_reward": -0.0024862352292984724, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024862352292984724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18212890625, + 0.0666598305106163, + 0.16919255256652832, + 0.04010453075170517, + 0.08728273957967758, + 0.06270874291658401, + 0.06663169711828232, + 0.13562238216400146, + 0.10509669035673141, + 0.08457181602716446 + ], + "credit_map": [ + 0.016872160136699677, + 0.015764929354190826, + 0.024872832000255585, + 0.04083458334207535, + 0.01628192886710167, + 0.01532828900963068, + 0.03245999664068222, + 0.01971421018242836, + 0.007420094218105078, + 0.007001784164458513, + 0.008057900704443455, + 0.008597728796303272, + 0.007612751796841621, + 0.006170507054775953, + 0.014789421111345291, + 0.007009648717939854, + 0.021582653746008873, + 0.02018646150827408, + 0.015056549571454525, + 0.018750024959445, + 0.016812771558761597, + 0.01203417032957077, + 0.050333552062511444, + 0.0144363883882761, + 0.0032460803631693125, + 0.0020913665648549795, + 0.00841809343546629, + 0.003286570543423295, + 0.002953980816528201, + 0.0036516531836241484, + 0.013372979126870632, + 0.0030838141683489084, + 0.007681125309318304, + 0.007125993724912405, + 0.00641867658123374, + 0.006022135727107525, + 0.008248013444244862, + 0.005917947739362717, + 0.024086734279990196, + 0.021782122552394867, + 0.00754364812746644, + 0.00523427315056324, + 0.009294914081692696, + 0.007520880084484816, + 0.008906583301723003, + 0.009275255724787712, + 0.0071152690798044205, + 0.00781792402267456, + 0.00744246831163764, + 0.006480412092059851, + 0.008785971440374851, + 0.006571006495505571, + 0.007436666171997786, + 0.009150529280304909, + 0.006825184915214777, + 0.013939461670815945, + 0.020168887451291084, + 0.022084781900048256, + 0.036943819373846054, + 0.014873604290187359, + 0.010646692477166653, + 0.009868909604847431, + 0.011965376324951649, + 0.009070322848856449, + 0.017011508345603943, + 0.011371598578989506, + 0.02361922338604927, + 0.007896438241004944, + 0.00784378219395876, + 0.008356789126992226, + 0.014638738706707954, + 0.01435862947255373, + 0.01208315510302782, + 0.009594098664820194, + 0.012252090498805046, + 0.012454849667847157, + 0.010214326903223991, + 0.009318618103861809, + 0.009300231002271175, + 0.009354451671242714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.0024862352292984724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a fixed wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, storm-swept sky with swirling clouds and purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a formation of luminous swords resembling a giant dragon moving from the depths to the front right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords exhibit a cool metallic quality with a cyan-blue energy halo and a fiery golden core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include golden and blue particle effects resembling stardust that dissipate into energy trails?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002485878299921751 + ], + "opposite_logprob_tokens": [ + -6.002485752105713 + ], + "logprob_reward": -0.002485878299921751, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002485878299921751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1609831601381302, + 0.10768546164035797, + 0.09059014916419983, + 0.08486289530992508, + 0.04317571222782135, + 0.06700435280799866, + 0.08157108724117279, + 0.0920359268784523, + 0.13793961703777313, + 0.1341516673564911 + ], + "credit_map": [ + 0.015288063324987888, + 0.011698788963258266, + 0.024611186236143112, + 0.04152864217758179, + 0.011783859692513943, + 0.011215909384191036, + 0.02781924046576023, + 0.017037471756339073, + 0.012302981689572334, + 0.0079368706792593, + 0.01917918026447296, + 0.01647811383008957, + 0.007899888791143894, + 0.008576501160860062, + 0.020905308425426483, + 0.014406617730855942, + 0.011819145642220974, + 0.007664672564715147, + 0.011787581257522106, + 0.009114116430282593, + 0.007452805060893297, + 0.005194821860641241, + 0.030367253348231316, + 0.007189755793660879, + 0.006117215380072594, + 0.00425981217995286, + 0.019803466275334358, + 0.006943997461348772, + 0.00492130545899272, + 0.008218117989599705, + 0.026915933936834335, + 0.007683041971176863, + 0.004883984103798866, + 0.0038613334763795137, + 0.005311904940754175, + 0.0036357578355818987, + 0.004209464881569147, + 0.0053095147013664246, + 0.007461220491677523, + 0.008502530865371227, + 0.00928216427564621, + 0.0062021128833293915, + 0.014101334847509861, + 0.009643759578466415, + 0.005536379758268595, + 0.006014574784785509, + 0.006869887001812458, + 0.009354138746857643, + 0.009261307306587696, + 0.005479545332491398, + 0.01981304958462715, + 0.011315698735415936, + 0.006471620406955481, + 0.008780061267316341, + 0.00749309454113245, + 0.012956710532307625, + 0.014905748888850212, + 0.015017589554190636, + 0.021307865157723427, + 0.010127795860171318, + 0.0064574056304991245, + 0.006760261487215757, + 0.00877256691455841, + 0.008686691522598267, + 0.024266518652439117, + 0.018762094900012016, + 0.021001113578677177, + 0.013984781689941883, + 0.011682434007525444, + 0.010508702136576176, + 0.016392473131418228, + 0.021341495215892792, + 0.013225588016211987, + 0.015728339552879333, + 0.02875007688999176, + 0.014516711235046387, + 0.013372549787163734, + 0.01173806469887495, + 0.02231081761419773, + 0.014509526081383228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.002485878299921751 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_13/index_tensor([3619, 3619])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_13/index_tensor([3619, 3619])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..87067341183c2c4c0fbf463313d9d0b9f92949f7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_13/index_tensor([3619, 3619])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sharp, dynamic swooping shot descends rapidly from a height, capturing the interior of a penthouse brimming with modern urban chic. An Asian woman, dressed in a boldly cut, dark green silk, reworked qipao, takes center stage. Her hair is coiffed, glossy black, and reaches down to her shoulders, her expression stern and resolute. The apartment's mirror-polished black marble floor reflects the expansive, bustling nightscape outside, awash in neon lights. On the floor before her, a group of crystal-clear art glass pieces casts a chilling light under the cool interior spotlights. As the camera descends to near ground level, she suddenly raises her right leg, clad in a black stiletto, the high slit of her qipao tracing a sharp arc through the air. With swift and decisive force, she delivers a fierce kick towards the glass pieces. At the moment of contact, time seems to be stretched. The camera captures the glass in an extreme slow-motion close-up, instantly bursting into countless glittering fragments under the huge impact force. These fragments, under the reflection of the cold spotlights in the apartment and the neon lights of the city outside the window, fly around violently like diamonds. The whole picture is full of strong visual impact and a cold and resolute atmosphere of catharsis.", + "seed": 3546, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_13/index_3619_seed_3546_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.056827271915972234, + "vlm_logprob": -0.056827271915972234, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.056827271915972234, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10974162071943283 + ], + "opposite_logprob_tokens": [ + -2.359741687774658 + ], + "logprob_reward": -0.10974162071943283, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10974162071943283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06880367547273636, + 0.05580538883805275, + 0.05695383995771408, + 0.05327068641781807, + 0.35624176263809204, + 0.03153618425130844, + 0.13289210200309753, + 0.04083739221096039, + 0.13487060368061066, + 0.06878837943077087 + ], + "credit_map": [ + 0.005727308802306652, + 0.00669135432690382, + 0.009948881343007088, + 0.012135868892073631, + 0.009251200594007969, + 0.010236041620373726, + 0.006925317458808422, + 0.007887703366577625, + 0.00623928289860487, + 0.0059828488156199455, + 0.007054844871163368, + 0.004372762516140938, + 0.006222638767212629, + 0.006540244910866022, + 0.008101996034383774, + 0.011290766298770905, + 0.00818681437522173, + 0.005365181248635054, + 0.005968487821519375, + 0.004643613006919622, + 0.008585607632994652, + 0.007440922316163778, + 0.008559035137295723, + 0.008204177021980286, + 0.008098972029983997, + 0.006376716773957014, + 0.006374555639922619, + 0.005040526390075684, + 0.008332823403179646, + 0.006273006089031696, + 0.006376184988766909, + 0.006397903896868229, + 0.029471680521965027, + 0.030950607731938362, + 0.08632459491491318, + 0.06788546591997147, + 0.029583141207695007, + 0.029195966199040413, + 0.039729949086904526, + 0.04310034215450287, + 0.004044738598167896, + 0.002170263323932886, + 0.003423056798055768, + 0.004462104756385088, + 0.005477071739733219, + 0.004869381431490183, + 0.00356026878580451, + 0.0035293016117066145, + 0.00873477105051279, + 0.006981381215155125, + 0.0173327773809433, + 0.04044898599386215, + 0.030834294855594635, + 0.008580401539802551, + 0.011386255733668804, + 0.008593237027525902, + 0.0036056172102689743, + 0.0035155722871422768, + 0.005915832240134478, + 0.007600205019116402, + 0.0035185213200747967, + 0.004854276776313782, + 0.004327575210481882, + 0.007499790284782648, + 0.017562804743647575, + 0.013311812654137611, + 0.03221283480525017, + 0.015399959869682789, + 0.009578227996826172, + 0.016036320477724075, + 0.008541681803762913, + 0.022226961329579353, + 0.006887944880872965, + 0.005613649729639292, + 0.010880978778004646, + 0.006864985451102257, + 0.006031452678143978, + 0.010751420632004738, + 0.007107130251824856, + 0.014650820754468441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.10974162071943283 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008662977255880833 + ], + "opposite_logprob_tokens": [ + -4.758663177490234 + ], + "logprob_reward": -0.008662977255880833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008662977255880833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12604452669620514, + 0.09665897488594055, + 0.05831119045615196, + 0.08913476020097733, + 0.08349664509296417, + 0.05345379188656807, + 0.1926962435245514, + 0.14919400215148926, + 0.09753281623125076, + 0.053476955741643906 + ], + "credit_map": [ + 0.012605488300323486, + 0.012440547347068787, + 0.015379124321043491, + 0.025949519127607346, + 0.020194265991449356, + 0.011041008867323399, + 0.014704429544508457, + 0.01373015996068716, + 0.012672067619860172, + 0.006578961852937937, + 0.007227998226881027, + 0.006530494894832373, + 0.011028734967112541, + 0.014134005643427372, + 0.009612083435058594, + 0.02887464314699173, + 0.006515467539429665, + 0.0061575924046337605, + 0.011952444911003113, + 0.006946359761059284, + 0.004984146449714899, + 0.007144551258534193, + 0.006605259142816067, + 0.008005375973880291, + 0.01189376786351204, + 0.013688343577086926, + 0.010875442065298557, + 0.00938285794109106, + 0.014734928496181965, + 0.008060465566813946, + 0.01196901872754097, + 0.008529944345355034, + 0.008715424686670303, + 0.009081853553652763, + 0.019448189064860344, + 0.01575850136578083, + 0.006179461721330881, + 0.005955380853265524, + 0.009876077063381672, + 0.00848176795989275, + 0.006287680473178625, + 0.004301856737583876, + 0.007951075211167336, + 0.00781339593231678, + 0.004543095361441374, + 0.005077163223177195, + 0.011442583054304123, + 0.006036953069269657, + 0.011968247592449188, + 0.010487737134099007, + 0.026895219460129738, + 0.0626961812376976, + 0.03924202546477318, + 0.011308240704238415, + 0.01569022797048092, + 0.014408375136554241, + 0.01964528113603592, + 0.015689242631196976, + 0.03733375295996666, + 0.0267608892172575, + 0.01110928412526846, + 0.012692305259406567, + 0.01221845205873251, + 0.013744818978011608, + 0.012423214502632618, + 0.010423639789223671, + 0.018553152680397034, + 0.010429199784994125, + 0.00746351620182395, + 0.008346068672835827, + 0.011582759208977222, + 0.01831127516925335, + 0.003254195209592581, + 0.0037429642397910357, + 0.00901333149522543, + 0.005840543657541275, + 0.0066030570305883884, + 0.006856601219624281, + 0.01021143700927496, + 0.007954832166433334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.008662977255880833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038100630044937134 + ], + "opposite_logprob_tokens": [ + -3.2881007194519043 + ], + "logprob_reward": -0.038100630044937134, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038100630044937134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08136438578367233, + 0.07444348931312561, + 0.04236631095409393, + 0.08717599511146545, + 0.13973954319953918, + 0.10143768042325974, + 0.1295943707227707, + 0.10903962701559067, + 0.18073563277721405, + 0.05410304665565491 + ], + "credit_map": [ + 0.010389497503638268, + 0.006636528298258781, + 0.010107932612299919, + 0.01150370854884386, + 0.007794929202646017, + 0.0094135832041502, + 0.007974334061145782, + 0.017543869093060493, + 0.010767487809062004, + 0.005530407186597586, + 0.0063727013766765594, + 0.006789220031350851, + 0.006202485412359238, + 0.010293891653418541, + 0.007958444766700268, + 0.020528847351670265, + 0.0063048554584383965, + 0.00402348255738616, + 0.0051856450736522675, + 0.005909887608140707, + 0.00578256556764245, + 0.004583618603646755, + 0.005967746488749981, + 0.004608504474163055, + 0.01301927212625742, + 0.01405647024512291, + 0.009793903678655624, + 0.008170507848262787, + 0.011359612457454205, + 0.007566824089735746, + 0.012197212316095829, + 0.011012183502316475, + 0.011862842366099358, + 0.012705354019999504, + 0.04056036099791527, + 0.02748003788292408, + 0.008307191543281078, + 0.00926476065069437, + 0.015656612813472748, + 0.013902360573410988, + 0.012115363031625748, + 0.008448859676718712, + 0.012650391086935997, + 0.017428958788514137, + 0.010305625386536121, + 0.013130382634699345, + 0.014724058099091053, + 0.0126340351998806, + 0.009830246679484844, + 0.006914157420396805, + 0.022835426032543182, + 0.0370650514960289, + 0.024558430537581444, + 0.00826703105121851, + 0.008811562322080135, + 0.011312456801533699, + 0.009758288972079754, + 0.013585773296654224, + 0.022985083982348442, + 0.020819202065467834, + 0.007314248941838741, + 0.007400491740554571, + 0.011644570156931877, + 0.015531966462731361, + 0.023215558379888535, + 0.01838942989706993, + 0.043634504079818726, + 0.02183546870946884, + 0.011948643252253532, + 0.015961842611432076, + 0.016732921823859215, + 0.029017243534326553, + 0.006030889227986336, + 0.004854528233408928, + 0.010428658686578274, + 0.005847263149917126, + 0.0050801183097064495, + 0.00685154041275382, + 0.006053768564015627, + 0.00895627774298191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.038100630044937134 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897930592298508 + ], + "opposite_logprob_tokens": [ + -2.578979253768921 + ], + "logprob_reward": -0.07897930592298508, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897930592298508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06428928673267365, + 0.057881228625774384, + 0.08604930341243744, + 0.09072050452232361, + 0.15744087100028992, + 0.0815010592341423, + 0.16407254338264465, + 0.08823031932115555, + 0.1617601364850998, + 0.048054736107587814 + ], + "credit_map": [ + 0.005046824924647808, + 0.005484830588102341, + 0.012812337838113308, + 0.008977332152426243, + 0.007594333495944738, + 0.010066119953989983, + 0.00641065277159214, + 0.007896851748228073, + 0.007494510151445866, + 0.005815739277750254, + 0.005995126441121101, + 0.006623887922614813, + 0.007385162636637688, + 0.008033549413084984, + 0.007121791131794453, + 0.00941146444529295, + 0.013515721075236797, + 0.011987343430519104, + 0.010982155799865723, + 0.010240246541798115, + 0.011135692708194256, + 0.008958070538938046, + 0.013405252248048782, + 0.0058248210698366165, + 0.012686855159699917, + 0.010651806369423866, + 0.01013067178428173, + 0.009193222038447857, + 0.016241546720266342, + 0.008307418785989285, + 0.011207425966858864, + 0.012301566079258919, + 0.013340123929083347, + 0.016649547964334488, + 0.03809402137994766, + 0.024278853088617325, + 0.015348944813013077, + 0.01604313589632511, + 0.01874573715031147, + 0.014940504916012287, + 0.011210296303033829, + 0.005946087650954723, + 0.01256602257490158, + 0.009207198396325111, + 0.009953216649591923, + 0.016632312908768654, + 0.006860404275357723, + 0.00912551861256361, + 0.011791429482400417, + 0.00898317527025938, + 0.027601977810263634, + 0.047659892588853836, + 0.033387310802936554, + 0.008601007983088493, + 0.012913195416331291, + 0.013134549371898174, + 0.009545008651912212, + 0.008572878316044807, + 0.020753826946020126, + 0.014037481509149075, + 0.006687566637992859, + 0.007619003299623728, + 0.008051779121160507, + 0.012962780892848969, + 0.02111770585179329, + 0.017035217955708504, + 0.04385146126151085, + 0.020804230123758316, + 0.008461923338472843, + 0.014183735474944115, + 0.010987105779349804, + 0.02531876228749752, + 0.004961715545505285, + 0.005079669877886772, + 0.008210529573261738, + 0.004963705316185951, + 0.0051992288790643215, + 0.007131661754101515, + 0.005099893547594547, + 0.007408329285681248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.07897930592298508 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865182563662529 + ], + "opposite_logprob_tokens": [ + -3.048651933670044 + ], + "logprob_reward": -0.04865182563662529, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865182563662529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030460413545370102, + 0.044869810342788696, + 0.02704288810491562, + 0.07384370267391205, + 0.06816470623016357, + 0.02681623585522175, + 0.06898674368858337, + 0.06349662691354752, + 0.5044889450073242, + 0.09182991087436676 + ], + "credit_map": [ + 0.0032790142577141523, + 0.0032691534142941236, + 0.004045594483613968, + 0.0033383413683623075, + 0.003725911257788539, + 0.0031874789856374264, + 0.0037315539084374905, + 0.005883367732167244, + 0.00741134537383914, + 0.003765192348510027, + 0.004318540450185537, + 0.003094237530604005, + 0.004108738619834185, + 0.007726151961833239, + 0.005583776161074638, + 0.008861826732754707, + 0.0029607985634356737, + 0.002886640140786767, + 0.0018333910265937448, + 0.0024287053383886814, + 0.004376148339360952, + 0.004103909246623516, + 0.003774259239435196, + 0.004679037258028984, + 0.009809788316488266, + 0.00950255524367094, + 0.007617407478392124, + 0.005741036031395197, + 0.016188139095902443, + 0.007681064773350954, + 0.010013668797910213, + 0.007290042005479336, + 0.006642321590334177, + 0.011971560306847095, + 0.013238269835710526, + 0.008822361007332802, + 0.00711066136136651, + 0.006415150128304958, + 0.007663902826607227, + 0.006300473585724831, + 0.0039137969724833965, + 0.00222501577809453, + 0.006264165509492159, + 0.00431074621155858, + 0.0022213051561266184, + 0.0032303258776664734, + 0.0025674053467810154, + 0.0020834750030189753, + 0.004395306576043367, + 0.003939751535654068, + 0.008857752196490765, + 0.022460877895355225, + 0.01439897995442152, + 0.004606334958225489, + 0.005274346098303795, + 0.005053394008427858, + 0.006001922767609358, + 0.006034236866980791, + 0.009959075599908829, + 0.012032701633870602, + 0.006911724805831909, + 0.007267408072948456, + 0.007662540301680565, + 0.007627017796039581, + 0.0665784552693367, + 0.0498441644012928, + 0.12066880613565445, + 0.058099377900362015, + 0.030987247824668884, + 0.054280418902635574, + 0.02999441884458065, + 0.0940360426902771, + 0.011281417682766914, + 0.0094876354560256, + 0.01377817615866661, + 0.00895703211426735, + 0.0073971133679151535, + 0.013299733400344849, + 0.008499076589941978, + 0.01912972889840603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.04865182563662529 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sharp, dynamic swooping shot descends rapidly from a height, capturing the interior of a penthouse brimming with modern urban chic. An Asian woman, dressed in a boldly cut, dark green silk, reworked qipao, takes center stage. Her hair is coiffed, glossy black, and reaches down to her shoulders, her expression stern and resolute. The apartment's mirror-polished black marble floor reflects the expansive, bustling nightscape outside, awash in neon lights. On the floor before her, a group of crystal-clear art glass pieces casts a chilling light under the cool interior spotlights. As the camera descends to near ground level, she suddenly raises her right leg, clad in a black stiletto, the high slit of her qipao tracing a sharp arc through the air. With swift and decisive force, she delivers a fierce kick towards the glass pieces. At the moment of contact, time seems to be stretched. The camera captures the glass in an extreme slow-motion close-up, instantly bursting into countless glittering fragments under the huge impact force. These fragments, under the reflection of the cold spotlights in the apartment and the neon lights of the city outside the window, fly around violently like diamonds. The whole picture is full of strong visual impact and a cold and resolute atmosphere of catharsis.", + "seed": 3547, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_13/index_3619_seed_3547_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.030692888796329497, + "vlm_logprob": -0.030692888796329497, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.030692888796329497, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0672486200928688 + ], + "opposite_logprob_tokens": [ + -2.942248582839966 + ], + "logprob_reward": -0.0672486200928688, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0672486200928688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39672666788101196, + 0.08396326750516891, + 0.045748259872198105, + 0.08651837706565857, + 0.03596958890557289, + 0.10935163497924805, + 0.11148534715175629, + 0.053187549114227295, + 0.02940651960670948, + 0.04764282703399658 + ], + "credit_map": [ + 0.022546641528606415, + 0.029275918379426003, + 0.023869993165135384, + 0.030479805544018745, + 0.02581196464598179, + 0.04944845288991928, + 0.09309599548578262, + 0.12219791114330292, + 0.009498192928731441, + 0.013688545674085617, + 0.011618590913712978, + 0.011934013105928898, + 0.007094771601259708, + 0.00946892611682415, + 0.009849237278103828, + 0.01081098336726427, + 0.004066207446157932, + 0.004169938620179892, + 0.00432329997420311, + 0.005394346080720425, + 0.006295203231275082, + 0.003783069783821702, + 0.013808411546051502, + 0.003907784353941679, + 0.006008408032357693, + 0.018269851803779602, + 0.005941805895417929, + 0.011272924020886421, + 0.014505875296890736, + 0.008494909852743149, + 0.015806786715984344, + 0.0062178149819374084, + 0.0037748932372778654, + 0.0070799184031784534, + 0.004226350225508213, + 0.004280072171241045, + 0.004201443865895271, + 0.004685623571276665, + 0.004683435894548893, + 0.0030378515366464853, + 0.013281850144267082, + 0.019041655585169792, + 0.011911546811461449, + 0.010099352337419987, + 0.015810610726475716, + 0.009713050909340382, + 0.01495749969035387, + 0.014536071568727493, + 0.010380207560956478, + 0.011064877733588219, + 0.022674284875392914, + 0.015405846759676933, + 0.010264601558446884, + 0.009243828244507313, + 0.014572042040526867, + 0.017879653722047806, + 0.004264906980097294, + 0.006607198156416416, + 0.009836224839091301, + 0.007681176532059908, + 0.0029441926162689924, + 0.005329746752977371, + 0.009316785261034966, + 0.007207317743450403, + 0.003435796592384577, + 0.005413282196968794, + 0.003109462559223175, + 0.004206172656267881, + 0.0027182085905224085, + 0.003142652101814747, + 0.004137029405683279, + 0.003243913408368826, + 0.003802137915045023, + 0.0054977769032120705, + 0.007124748080968857, + 0.01047236192971468, + 0.0022813291288912296, + 0.006188570987433195, + 0.006094169337302446, + 0.006181731820106506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0672486200928688 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008658131584525108 + ], + "opposite_logprob_tokens": [ + -4.758657932281494 + ], + "logprob_reward": -0.008658131584525108, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008658131584525108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2599520683288574, + 0.08099430054426193, + 0.11433883011341095, + 0.13997220993041992, + 0.06878770142793655, + 0.0464765727519989, + 0.05171409994363785, + 0.08397805690765381, + 0.1269749402999878, + 0.026811303570866585 + ], + "credit_map": [ + 0.016201747581362724, + 0.019526276737451553, + 0.019919518381357193, + 0.030373381450772285, + 0.0321609303355217, + 0.027957133948802948, + 0.058224618434906006, + 0.05558845028281212, + 0.0057079484686255455, + 0.009797940030694008, + 0.008480794727802277, + 0.00677531911060214, + 0.015369768254458904, + 0.01151232048869133, + 0.014728142879903316, + 0.008622055873274803, + 0.008993125520646572, + 0.013777265325188637, + 0.008690830320119858, + 0.008210095576941967, + 0.03228043019771576, + 0.011473363265395164, + 0.022514618933200836, + 0.008399080485105515, + 0.009986512362957, + 0.03258997201919556, + 0.009600561112165451, + 0.01865122839808464, + 0.024985438212752342, + 0.014252573251724243, + 0.020699238404631615, + 0.009206661023199558, + 0.006059995852410793, + 0.01154265459626913, + 0.005037588998675346, + 0.00608794717118144, + 0.01063306163996458, + 0.010457847267389297, + 0.01271811593323946, + 0.006250477861613035, + 0.004861889872699976, + 0.007230655290186405, + 0.005747533403337002, + 0.004481886513531208, + 0.006664325017482042, + 0.005000745877623558, + 0.004861720837652683, + 0.007627811282873154, + 0.007536549586802721, + 0.005581318400800228, + 0.007885811850428581, + 0.005415463354438543, + 0.005427912808954716, + 0.004496271256357431, + 0.00775741133838892, + 0.007613354362547398, + 0.014291550032794476, + 0.009318996220827103, + 0.015678778290748596, + 0.012895230203866959, + 0.005278562195599079, + 0.008095833472907543, + 0.007767929695546627, + 0.010651174001395702, + 0.00834622886031866, + 0.008758838288486004, + 0.013067088089883327, + 0.027167383581399918, + 0.011097974143922329, + 0.011388156563043594, + 0.02226714789867401, + 0.024882104247808456, + 0.00266791763715446, + 0.002209952101111412, + 0.0025889079552143812, + 0.0037951923441141844, + 0.003394980216398835, + 0.0030578519217669964, + 0.005549794528633356, + 0.0035467029083520174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.008658131584525108 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579388365149498 + ], + "opposite_logprob_tokens": [ + -3.8955793380737305 + ], + "logprob_reward": -0.020579388365149498, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579388365149498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29648682475090027, + 0.1181420162320137, + 0.03345688432455063, + 0.07471238076686859, + 0.18924802541732788, + 0.039537783712148666, + 0.04445333033800125, + 0.09004023671150208, + 0.06732247769832611, + 0.0466001071035862 + ], + "credit_map": [ + 0.01467535924166441, + 0.016227878630161285, + 0.016721753403544426, + 0.026405762881040573, + 0.016081852838397026, + 0.03482003137469292, + 0.07346757501363754, + 0.09808660298585892, + 0.005972553510218859, + 0.007563698571175337, + 0.0075531271286308765, + 0.007213945500552654, + 0.006257247179746628, + 0.04147789999842644, + 0.007344064302742481, + 0.03475948050618172, + 0.0037594034802168608, + 0.002872679615393281, + 0.0026314747519791126, + 0.002926083980128169, + 0.005119089037179947, + 0.005322169978171587, + 0.005500286817550659, + 0.005325698293745518, + 0.004379470832645893, + 0.009247398003935814, + 0.0066026439890265465, + 0.009262947365641594, + 0.013402183540165424, + 0.008898427709937096, + 0.01220149826258421, + 0.010717814788222313, + 0.015027711167931557, + 0.02294665016233921, + 0.012389065697789192, + 0.01741950586438179, + 0.030704911798238754, + 0.033149149268865585, + 0.03742588683962822, + 0.020185144618153572, + 0.0036099450662732124, + 0.004862875211983919, + 0.004526750184595585, + 0.004250595811754465, + 0.006258608773350716, + 0.0038816356100142, + 0.006424597464501858, + 0.005722775589674711, + 0.0027883993461728096, + 0.004673896823078394, + 0.006141505669802427, + 0.005445896647870541, + 0.005567569751292467, + 0.004124436993151903, + 0.007309025153517723, + 0.008402597159147263, + 0.0071993498131632805, + 0.013692331500351429, + 0.016507888212800026, + 0.013086123391985893, + 0.006075313780456781, + 0.01187576912343502, + 0.008848241530358791, + 0.012755216099321842, + 0.004038704559206963, + 0.005223645828664303, + 0.00709994463250041, + 0.015610164031386375, + 0.004659823141992092, + 0.005683738272637129, + 0.010105378925800323, + 0.01490108110010624, + 0.0040519810281693935, + 0.003543163649737835, + 0.0041003418155014515, + 0.0077397930435836315, + 0.005065688863396645, + 0.00536006735637784, + 0.009056339040398598, + 0.007682733703404665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.020579388365149498 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023285046219825745 + ], + "opposite_logprob_tokens": [ + -3.773285150527954 + ], + "logprob_reward": -0.023285046219825745, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023285046219825745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07582467794418335, + 0.12229839712381363, + 0.07364390790462494, + 0.09494424611330032, + 0.15260492265224457, + 0.07832930237054825, + 0.0841749906539917, + 0.2013074904680252, + 0.06157989427447319, + 0.05529211834073067 + ], + "credit_map": [ + 0.005371657200157642, + 0.006882813293486834, + 0.006480588112026453, + 0.012113712728023529, + 0.008419310674071312, + 0.007097860332578421, + 0.017390772700309753, + 0.012067971751093864, + 0.003989539574831724, + 0.007197792176157236, + 0.007029776927083731, + 0.0055212522856891155, + 0.008254443295300007, + 0.04494505748152733, + 0.006516729015856981, + 0.03884381055831909, + 0.005380940157920122, + 0.012335485778748989, + 0.007054643705487251, + 0.008811013773083687, + 0.011577368713915348, + 0.008132334798574448, + 0.013365313410758972, + 0.006986813619732857, + 0.007197122555226088, + 0.011122584342956543, + 0.007162062451243401, + 0.009276154451072216, + 0.015204006806015968, + 0.01031292974948883, + 0.022861303761601448, + 0.011808083392679691, + 0.01180669479072094, + 0.029794370755553246, + 0.02083556540310383, + 0.016681091859936714, + 0.02268502674996853, + 0.018756745383143425, + 0.02056289091706276, + 0.011482532136142254, + 0.010733801871538162, + 0.011709172278642654, + 0.006211775355041027, + 0.00632024509832263, + 0.015253509394824505, + 0.005791671574115753, + 0.01214184146374464, + 0.010167283937335014, + 0.009041433222591877, + 0.008684183470904827, + 0.011399077251553535, + 0.014592674560844898, + 0.008818073198199272, + 0.006090570241212845, + 0.01420686487108469, + 0.011342115700244904, + 0.016702016815543175, + 0.02176937833428383, + 0.03181649371981621, + 0.04796571657061577, + 0.015568340197205544, + 0.014642316848039627, + 0.02204059064388275, + 0.030802637338638306, + 0.004717754200100899, + 0.008762671612203121, + 0.005434280727058649, + 0.012185122817754745, + 0.006199058145284653, + 0.006277803797274828, + 0.006946983747184277, + 0.011056221090257168, + 0.0034997262991964817, + 0.004152528941631317, + 0.006434701383113861, + 0.010568168014287949, + 0.003811859292909503, + 0.007580135948956013, + 0.008526921272277832, + 0.01071807648986578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.023285046219825745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033693257719278336 + ], + "opposite_logprob_tokens": [ + -3.408693313598633 + ], + "logprob_reward": -0.033693257719278336, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033693257719278336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.215761199593544, + 0.06251905113458633, + 0.05469376966357231, + 0.06915152817964554, + 0.05416864529252052, + 0.06445367634296417, + 0.2025214433670044, + 0.2083844691514969, + 0.04942069575190544, + 0.01892552152276039 + ], + "credit_map": [ + 0.011251275427639484, + 0.00929903145879507, + 0.011791232973337173, + 0.018276605755090714, + 0.01178079005330801, + 0.02016846276819706, + 0.06089133024215698, + 0.07230247557163239, + 0.004253150895237923, + 0.0064763561822474, + 0.00545195396989584, + 0.014976940117776394, + 0.005815071519464254, + 0.010144637897610664, + 0.007981347851455212, + 0.00741959223523736, + 0.004121514968574047, + 0.006101811770349741, + 0.004731229040771723, + 0.009430235251784325, + 0.007933405227959156, + 0.004983747377991676, + 0.013373319990932941, + 0.00401850463822484, + 0.004737380426377058, + 0.014928940683603287, + 0.0051040551625192165, + 0.009112142026424408, + 0.01155669055879116, + 0.007317979820072651, + 0.0117406714707613, + 0.004653667099773884, + 0.009521318599581718, + 0.008278668858110905, + 0.005088903475552797, + 0.005039749201387167, + 0.008755478076636791, + 0.005895023234188557, + 0.007182335946708918, + 0.004407165572047234, + 0.009029743261635303, + 0.014888246543705463, + 0.004936512093991041, + 0.0066406745463609695, + 0.010321622714400291, + 0.005898411851376295, + 0.006781561300158501, + 0.005956907290965319, + 0.012413347139954567, + 0.020719265565276146, + 0.03457771614193916, + 0.027405673637986183, + 0.02034568227827549, + 0.014916644431650639, + 0.03359764441847801, + 0.03854546323418617, + 0.01992950588464737, + 0.025353869423270226, + 0.039232827723026276, + 0.02709379233419895, + 0.013057760894298553, + 0.019511498510837555, + 0.036059893667697906, + 0.028145311400294304, + 0.003583249868825078, + 0.005276711657643318, + 0.003963876981288195, + 0.01504252664744854, + 0.0022202287800610065, + 0.0035701023880392313, + 0.0034195284824818373, + 0.012344470247626305, + 0.001975556369870901, + 0.002508391859009862, + 0.002426609629765153, + 0.0029479179065674543, + 0.0015691656153649092, + 0.0024033391382545233, + 0.0022320260759443045, + 0.0028625151608139277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.033693257719278336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sharp, dynamic swooping shot descends rapidly from a height, capturing the interior of a penthouse brimming with modern urban chic. An Asian woman, dressed in a boldly cut, dark green silk, reworked qipao, takes center stage. Her hair is coiffed, glossy black, and reaches down to her shoulders, her expression stern and resolute. The apartment's mirror-polished black marble floor reflects the expansive, bustling nightscape outside, awash in neon lights. On the floor before her, a group of crystal-clear art glass pieces casts a chilling light under the cool interior spotlights. As the camera descends to near ground level, she suddenly raises her right leg, clad in a black stiletto, the high slit of her qipao tracing a sharp arc through the air. With swift and decisive force, she delivers a fierce kick towards the glass pieces. At the moment of contact, time seems to be stretched. The camera captures the glass in an extreme slow-motion close-up, instantly bursting into countless glittering fragments under the huge impact force. These fragments, under the reflection of the cold spotlights in the apartment and the neon lights of the city outside the window, fly around violently like diamonds. The whole picture is full of strong visual impact and a cold and resolute atmosphere of catharsis.", + "seed": 3548, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_13/index_3619_seed_3548_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05128549356013536, + "vlm_logprob": -0.05128549356013536, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.05128549356013536, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17205412685871124 + ], + "opposite_logprob_tokens": [ + -1.9220541715621948 + ], + "logprob_reward": -0.17205412685871124, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17205412685871124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12884241342544556, + 0.17826086282730103, + 0.09441035240888596, + 0.06900715082883835, + 0.16137488186359406, + 0.11521483212709427, + 0.0869588777422905, + 0.04391217604279518, + 0.06551429629325867, + 0.05650416389107704 + ], + "credit_map": [ + 0.017251618206501007, + 0.015605315566062927, + 0.010948330163955688, + 0.025406712666153908, + 0.010849328711628914, + 0.021960239857435226, + 0.011844600550830364, + 0.014976272359490395, + 0.015731539577245712, + 0.017261909320950508, + 0.019162330776453018, + 0.018900418654084206, + 0.035204458981752396, + 0.035921406000852585, + 0.01736595295369625, + 0.018712852150201797, + 0.009785679168999195, + 0.011776444502174854, + 0.009892625734210014, + 0.015316249802708626, + 0.014235378243029118, + 0.010244258679449558, + 0.011209898628294468, + 0.011949826031923294, + 0.01161225512623787, + 0.005888733547180891, + 0.008454812690615654, + 0.006469027604907751, + 0.009341337718069553, + 0.00935108121484518, + 0.006446794141083956, + 0.011443114839494228, + 0.024818556383252144, + 0.01930031180381775, + 0.013393556699156761, + 0.024905063211917877, + 0.02074435167014599, + 0.025163406506180763, + 0.01551053673028946, + 0.01753911003470421, + 0.012892547063529491, + 0.006926618050783873, + 0.01020348072052002, + 0.024754734709858894, + 0.015580925159156322, + 0.03473678603768349, + 0.005103088915348053, + 0.005016662180423737, + 0.010055230930447578, + 0.00935002788901329, + 0.007208420895040035, + 0.012609632685780525, + 0.01596931368112564, + 0.01216152124106884, + 0.010120542719960213, + 0.00948419515043497, + 0.0057998946867883205, + 0.006822049617767334, + 0.004071300849318504, + 0.00471627339720726, + 0.005115303210914135, + 0.006440077442675829, + 0.003813746850937605, + 0.007133533246815205, + 0.00985774677246809, + 0.013724343851208687, + 0.005443112924695015, + 0.005936082918196917, + 0.009277062490582466, + 0.012524577789008617, + 0.0040659490041434765, + 0.0046854265965521336, + 0.007389032281935215, + 0.004838918801397085, + 0.005222866777330637, + 0.008996154181659222, + 0.014356657862663269, + 0.00675795366987586, + 0.0054867686703801155, + 0.0034558118786662817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.17205412685871124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011098791845142841 + ], + "opposite_logprob_tokens": [ + -4.511098861694336 + ], + "logprob_reward": -0.011098791845142841, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011098791845142841, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1739736944437027, + 0.20623713731765747, + 0.03576740622520447, + 0.049387600272893906, + 0.04949765279889107, + 0.31372103095054626, + 0.05227910727262497, + 0.024825045838952065, + 0.03062915988266468, + 0.0636821985244751 + ], + "credit_map": [ + 0.016993584111332893, + 0.021264823153614998, + 0.016678379848599434, + 0.04698885977268219, + 0.01681152917444706, + 0.016138127073645592, + 0.017528420314192772, + 0.021569976583123207, + 0.01962585560977459, + 0.033852916210889816, + 0.017671560868620872, + 0.031309764832258224, + 0.023637698963284492, + 0.038551490753889084, + 0.019313182681798935, + 0.022274671122431755, + 0.004919364117085934, + 0.004917672369629145, + 0.003432092722505331, + 0.005163369234651327, + 0.005312839522957802, + 0.0037052377592772245, + 0.003922080155462027, + 0.004394751042127609, + 0.006863558664917946, + 0.005008751060813665, + 0.00659352308139205, + 0.0047400412149727345, + 0.008340678177773952, + 0.006109405774623156, + 0.006490495055913925, + 0.005241145379841328, + 0.005133644677698612, + 0.004514283034950495, + 0.0036828576121479273, + 0.003887651488184929, + 0.0063731474801898, + 0.007964658550918102, + 0.009890662506222725, + 0.00805074442178011, + 0.032811932265758514, + 0.017186567187309265, + 0.0341043695807457, + 0.09052655100822449, + 0.05212369188666344, + 0.05708184838294983, + 0.01659606397151947, + 0.013290015049278736, + 0.0053648315370082855, + 0.005228927358984947, + 0.005327174440026283, + 0.0089208222925663, + 0.009945168159902096, + 0.007148898672312498, + 0.006175067741423845, + 0.004168217536062002, + 0.0030799931846559048, + 0.0027556419372558594, + 0.002289190888404846, + 0.003215083386749029, + 0.003441781969740987, + 0.004737337585538626, + 0.0023185713216662407, + 0.002987443935126066, + 0.0036192205734550953, + 0.0030173391569405794, + 0.0019769880454987288, + 0.003419039770960808, + 0.0064683579839766026, + 0.006361652165651321, + 0.002317911945283413, + 0.003448650473728776, + 0.005914702545851469, + 0.0035458048805594444, + 0.008231095038354397, + 0.01600692979991436, + 0.015949396416544914, + 0.005778141785413027, + 0.0035274536348879337, + 0.004728679079562426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.011098791845142841 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346080005168915 + ], + "opposite_logprob_tokens": [ + -3.65134596824646 + ], + "logprob_reward": -0.026346080005168915, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346080005168915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08752833306789398, + 0.1332586407661438, + 0.08793231844902039, + 0.05365213006734848, + 0.07464411109685898, + 0.33504560589790344, + 0.07433678954839706, + 0.053741224110126495, + 0.03638998046517372, + 0.06347090750932693 + ], + "credit_map": [ + 0.011937068775296211, + 0.009292956441640854, + 0.009882078506052494, + 0.018656274303793907, + 0.0071593765169382095, + 0.010322773829102516, + 0.00859531294554472, + 0.011682488024234772, + 0.012637891806662083, + 0.015806736424565315, + 0.019694859161973, + 0.01675216481089592, + 0.018113600090146065, + 0.02376411110162735, + 0.01213888544589281, + 0.014350399374961853, + 0.01237501297146082, + 0.011186939664185047, + 0.00854380615055561, + 0.014682828448712826, + 0.015026001259684563, + 0.008726801723241806, + 0.007886180654168129, + 0.00950474850833416, + 0.007562450598925352, + 0.004423040896654129, + 0.007994181476533413, + 0.00638315174728632, + 0.007674179505556822, + 0.008813954889774323, + 0.005799355451017618, + 0.00500181969255209, + 0.008744505234062672, + 0.00949863251298666, + 0.005632046144455671, + 0.007613348308950663, + 0.00906057097017765, + 0.012945782393217087, + 0.008321513421833515, + 0.012827713042497635, + 0.034690529108047485, + 0.018898172304034233, + 0.03459428250789642, + 0.0924709141254425, + 0.07199312001466751, + 0.04993712902069092, + 0.018975554034113884, + 0.013485908508300781, + 0.007406106684356928, + 0.00739264627918601, + 0.006342735141515732, + 0.011048211716115475, + 0.014239571057260036, + 0.010526011697947979, + 0.01098189502954483, + 0.006399611011147499, + 0.007761897053569555, + 0.004289325792342424, + 0.004002623725682497, + 0.0053216987289488316, + 0.009013194590806961, + 0.009321337565779686, + 0.007763857487589121, + 0.006267289165407419, + 0.0056405095383524895, + 0.004872025456279516, + 0.0036794377956539392, + 0.003412810852751136, + 0.00573666300624609, + 0.005544034298509359, + 0.0031058990862220526, + 0.004398598801344633, + 0.006557118147611618, + 0.005953934974968433, + 0.007216328755021095, + 0.01320249680429697, + 0.013195975683629513, + 0.008285936899483204, + 0.004611643496900797, + 0.00444746995344758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.026346080005168915 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635188400745392 + ], + "opposite_logprob_tokens": [ + -3.6513519287109375 + ], + "logprob_reward": -0.02635188400745392, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635188400745392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05270780995488167, + 0.09084796160459518, + 0.09375842660665512, + 0.05058803781867027, + 0.03275253623723984, + 0.4285223186016083, + 0.0710277184844017, + 0.05707385018467903, + 0.026909079402685165, + 0.09581230580806732 + ], + "credit_map": [ + 0.005670893471688032, + 0.005594838876277208, + 0.006736183073371649, + 0.008935579098761082, + 0.006390407215803862, + 0.0073798587545752525, + 0.0058616274036467075, + 0.006138412281870842, + 0.008602827787399292, + 0.010811807587742805, + 0.012784631922841072, + 0.012312255799770355, + 0.008480462245643139, + 0.019442614167928696, + 0.00795002095401287, + 0.010463329963386059, + 0.008617517538368702, + 0.02068948745727539, + 0.013090894557535648, + 0.013116306625306606, + 0.011164290830492973, + 0.008796535432338715, + 0.008752294816076756, + 0.009531093761324883, + 0.006911423522979021, + 0.004564332775771618, + 0.004576436709612608, + 0.00703386077657342, + 0.008049125783145428, + 0.006088470108807087, + 0.0061216652393341064, + 0.007242713123559952, + 0.00592416524887085, + 0.002708864863961935, + 0.0031056334264576435, + 0.00502906646579504, + 0.00465168384835124, + 0.005063712131232023, + 0.0033023199066519737, + 0.002967085922136903, + 0.0449722558259964, + 0.02382436953485012, + 0.04566432163119316, + 0.12321354448795319, + 0.07477540522813797, + 0.07602626830339432, + 0.02178971841931343, + 0.018256403505802155, + 0.007905562408268452, + 0.006381440442055464, + 0.005585411097854376, + 0.008839510381221771, + 0.013781161978840828, + 0.010131166316568851, + 0.010226205922663212, + 0.008177247829735279, + 0.008571499027311802, + 0.006197019945830107, + 0.0044974228367209435, + 0.005492577329277992, + 0.006579329259693623, + 0.008103414438664913, + 0.006035189144313335, + 0.011597388423979282, + 0.0022118822671473026, + 0.0027971691451966763, + 0.002191432286053896, + 0.0021346481516957283, + 0.007110942620784044, + 0.005268535576760769, + 0.0028321503195911646, + 0.0023623155429959297, + 0.010486768558621407, + 0.005354702938348055, + 0.011796940118074417, + 0.022231757640838623, + 0.023157544434070587, + 0.008840911090373993, + 0.006860272027552128, + 0.007083401549607515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.02635188400745392 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020576585084199905 + ], + "opposite_logprob_tokens": [ + -3.8955764770507812 + ], + "logprob_reward": -0.020576585084199905, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020576585084199905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07079823315143585, + 0.08746130764484406, + 0.10545989871025085, + 0.05557027459144592, + 0.12461337447166443, + 0.23728418350219727, + 0.042146794497966766, + 0.08749447762966156, + 0.04555700719356537, + 0.14361447095870972 + ], + "credit_map": [ + 0.006190194748342037, + 0.007721351459622383, + 0.009064849466085434, + 0.011829915456473827, + 0.007844394072890282, + 0.01013723574578762, + 0.008808906190097332, + 0.009201377630233765, + 0.007890063337981701, + 0.012435294687747955, + 0.011871336959302425, + 0.011520649306476116, + 0.013878858648240566, + 0.011908959597349167, + 0.007428118493407965, + 0.01052801962941885, + 0.021131686866283417, + 0.008715059608221054, + 0.00889871921390295, + 0.009597851894795895, + 0.013341803103685379, + 0.012467415072023869, + 0.00980985164642334, + 0.02149750478565693, + 0.006819911766797304, + 0.005930519197136164, + 0.006066303234547377, + 0.007430578116327524, + 0.009740574285387993, + 0.0062273587100207806, + 0.007426723372191191, + 0.005928295198827982, + 0.015096460469067097, + 0.01676161214709282, + 0.010215549729764462, + 0.016260026022791862, + 0.017722034826874733, + 0.019622845575213432, + 0.013452847488224506, + 0.015481988899409771, + 0.030710775405168533, + 0.017284052446484566, + 0.012721206992864609, + 0.019786648452281952, + 0.05876494199037552, + 0.06419538706541061, + 0.013132084161043167, + 0.020689064636826515, + 0.005158144049346447, + 0.004829323850572109, + 0.003113087499514222, + 0.0052843098528683186, + 0.006876451428979635, + 0.005757199600338936, + 0.005668578203767538, + 0.005459694657474756, + 0.012811828404664993, + 0.015170032158493996, + 0.004839249420911074, + 0.0066046444699168205, + 0.008518052287399769, + 0.016698453575372696, + 0.006634569261223078, + 0.016217639669775963, + 0.0058332281187176704, + 0.005511967930942774, + 0.004886497277766466, + 0.005690440535545349, + 0.006344607099890709, + 0.00823941733688116, + 0.0032756037544459105, + 0.0057752421125769615, + 0.01333682518452406, + 0.010428005829453468, + 0.017615053802728653, + 0.034454312175512314, + 0.029460787773132324, + 0.013740315102040768, + 0.01215322781354189, + 0.012425933964550495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.020576585084199905 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sharp, dynamic swooping shot descends rapidly from a height, capturing the interior of a penthouse brimming with modern urban chic. An Asian woman, dressed in a boldly cut, dark green silk, reworked qipao, takes center stage. Her hair is coiffed, glossy black, and reaches down to her shoulders, her expression stern and resolute. The apartment's mirror-polished black marble floor reflects the expansive, bustling nightscape outside, awash in neon lights. On the floor before her, a group of crystal-clear art glass pieces casts a chilling light under the cool interior spotlights. As the camera descends to near ground level, she suddenly raises her right leg, clad in a black stiletto, the high slit of her qipao tracing a sharp arc through the air. With swift and decisive force, she delivers a fierce kick towards the glass pieces. At the moment of contact, time seems to be stretched. The camera captures the glass in an extreme slow-motion close-up, instantly bursting into countless glittering fragments under the huge impact force. These fragments, under the reflection of the cold spotlights in the apartment and the neon lights of the city outside the window, fly around violently like diamonds. The whole picture is full of strong visual impact and a cold and resolute atmosphere of catharsis.", + "seed": 3549, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_13/index_3619_seed_3549_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.039086837135255334, + "vlm_logprob": -0.039086837135255334, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.039086837135255334, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09462396800518036 + ], + "opposite_logprob_tokens": [ + -2.5946240425109863 + ], + "logprob_reward": -0.09462396800518036, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09462396800518036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04156644642353058, + 0.020980549976229668, + 0.01755276881158352, + 0.0137075399979949, + 0.082480788230896, + 0.015748275443911552, + 0.7591647505760193, + 0.023736365139484406, + 0.015244263224303722, + 0.009818272665143013 + ], + "credit_map": [ + 0.004824804607778788, + 0.00410292437300086, + 0.005148420110344887, + 0.006413477938622236, + 0.003957041539251804, + 0.0047521223314106464, + 0.005632804706692696, + 0.006734851747751236, + 0.0027819557581096888, + 0.0019587844144552946, + 0.004843412432819605, + 0.0025568322744220495, + 0.0014970568008720875, + 0.0023860465735197067, + 0.0029152699280530214, + 0.002041192026808858, + 0.0013628205051645637, + 0.002000755863264203, + 0.0017413083696737885, + 0.002173813758417964, + 0.003074158914387226, + 0.002141351578757167, + 0.002234686864539981, + 0.0028238745871931314, + 0.0009859076235443354, + 0.0027294931933283806, + 0.0009926549391821027, + 0.00246236240491271, + 0.0021953240502625704, + 0.0017311200499534607, + 0.001381194917485118, + 0.001229483401402831, + 0.004115425515919924, + 0.012919686734676361, + 0.0047010923735797405, + 0.006246122065931559, + 0.004646616522222757, + 0.02954963780939579, + 0.004213377367705107, + 0.016088828444480896, + 0.0015344062121585011, + 0.0016249361215159297, + 0.0018953698454424739, + 0.0015207664109766483, + 0.0019396246643736959, + 0.0016690962947905064, + 0.0030406282749027014, + 0.002523446921259165, + 0.03661293908953667, + 0.008327685296535492, + 0.03343309462070465, + 0.018262773752212524, + 0.05710120499134064, + 0.03099929541349411, + 0.2784373164176941, + 0.2959904074668884, + 0.0028140987269580364, + 0.0020906529389321804, + 0.00292115262709558, + 0.00210933736525476, + 0.002852271543815732, + 0.004419204778969288, + 0.0029732543043792248, + 0.003556392155587673, + 0.0018373916391283274, + 0.0025361713487654924, + 0.0015269456198439002, + 0.0025433828122913837, + 0.0019025210058316588, + 0.0009878334822133183, + 0.0025843211915344, + 0.0013256954262033105, + 0.0014498722739517689, + 0.0008375695324502885, + 0.0014935904182493687, + 0.001145611167885363, + 0.0011411227751523256, + 0.00110415939707309, + 0.0013425369979813695, + 0.0013038106262683868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.09462396800518036 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00866604968905449 + ], + "opposite_logprob_tokens": [ + -4.758666038513184 + ], + "logprob_reward": -0.00866604968905449, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00866604968905449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051862943917512894, + 0.051477350294589996, + 0.033385924994945526, + 0.04567592218518257, + 0.10268845409154892, + 0.025622213259339333, + 0.6072349548339844, + 0.04045518860220909, + 0.014628737233579159, + 0.026968354359269142 + ], + "credit_map": [ + 0.004554746672511101, + 0.007937947288155556, + 0.005198825150728226, + 0.007548500783741474, + 0.0063162934966385365, + 0.006239408627152443, + 0.0078117274679243565, + 0.00625549117103219, + 0.006895841099321842, + 0.006955598946660757, + 0.007939333096146584, + 0.009672984480857849, + 0.004067968111485243, + 0.006468959152698517, + 0.006058200728148222, + 0.003418466541916132, + 0.002568711992353201, + 0.00688167754560709, + 0.0017946743173524737, + 0.003872431581839919, + 0.00497072096914053, + 0.0053695677779614925, + 0.0048463186249136925, + 0.0030818204395473003, + 0.0055336426012218, + 0.00827870611101389, + 0.004012539517134428, + 0.007487105671316385, + 0.005263205152004957, + 0.004536793567240238, + 0.004520179238170385, + 0.006043754983693361, + 0.00646599754691124, + 0.013317011296749115, + 0.0065352050587534904, + 0.009324699640274048, + 0.005361703224480152, + 0.03221207112073898, + 0.006009950768202543, + 0.023461811244487762, + 0.0016494359588250518, + 0.0027012312784790993, + 0.0022468869574368, + 0.003060839371755719, + 0.004022486973553896, + 0.0026965299621224403, + 0.005566307809203863, + 0.0036784952972084284, + 0.028991421684622765, + 0.007156917359679937, + 0.02704593725502491, + 0.01476871408522129, + 0.045131973922252655, + 0.02568967081606388, + 0.22221605479717255, + 0.23623421788215637, + 0.0071511478163301945, + 0.0028190654702484608, + 0.00404707994312048, + 0.0037980428896844387, + 0.0062769134528934956, + 0.006781154312193394, + 0.005366698373109102, + 0.004215084947645664, + 0.00177628337405622, + 0.0017896468052640557, + 0.0023948585148900747, + 0.002086950931698084, + 0.0020816160831600428, + 0.0011958037503063679, + 0.0020732353441417217, + 0.0012303423136472702, + 0.0038793557323515415, + 0.002828464610502124, + 0.003810903988778591, + 0.003306084079667926, + 0.0034961095079779625, + 0.00392512883991003, + 0.0029270693194121122, + 0.0027952375821769238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.00866604968905449 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370893374085426 + ], + "opposite_logprob_tokens": [ + -3.4087090492248535 + ], + "logprob_reward": -0.03370893374085426, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370893374085426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06330931186676025, + 0.06212460622191429, + 0.04261638969182968, + 0.1292990893125534, + 0.16232754290103912, + 0.05697554349899292, + 0.29830402135849, + 0.06816244125366211, + 0.07129126042127609, + 0.045589759945869446 + ], + "credit_map": [ + 0.0052416143007576466, + 0.006272921338677406, + 0.00785435363650322, + 0.008295102044939995, + 0.006260054651647806, + 0.010853013955056667, + 0.011106149293482304, + 0.007426111958920956, + 0.010121526196599007, + 0.006511885207146406, + 0.013077744282782078, + 0.010081332176923752, + 0.004643380641937256, + 0.006478006020188332, + 0.0063478294759988785, + 0.0048629119992256165, + 0.006148453336209059, + 0.008391170762479305, + 0.0042576780542731285, + 0.005586202722042799, + 0.003652936080470681, + 0.005761345848441124, + 0.003896767273545265, + 0.004921837244182825, + 0.016402477398514748, + 0.021747607737779617, + 0.012604198418557644, + 0.020013323053717613, + 0.015003051608800888, + 0.013815900310873985, + 0.012994335032999516, + 0.01671821065247059, + 0.008543545380234718, + 0.024623289704322815, + 0.00789026077836752, + 0.016802610829472542, + 0.008686455897986889, + 0.0536779947578907, + 0.007663565222173929, + 0.03443983569741249, + 0.004352516960352659, + 0.0075361961498856544, + 0.006232009269297123, + 0.00955808162689209, + 0.006547783967107534, + 0.005715820472687483, + 0.009047053754329681, + 0.00798608735203743, + 0.014743045903742313, + 0.009864111430943012, + 0.01579204574227333, + 0.011606631800532341, + 0.020505361258983612, + 0.014323702082037926, + 0.10687380284070969, + 0.1045953556895256, + 0.0085862772539258, + 0.005996905267238617, + 0.0062453607097268105, + 0.007724510505795479, + 0.007680848240852356, + 0.010233314707875252, + 0.00962755922228098, + 0.01206766813993454, + 0.008952178061008453, + 0.012794246897101402, + 0.007003540173172951, + 0.01180721540004015, + 0.012767891399562359, + 0.006554696708917618, + 0.0060209897346794605, + 0.005390505772083998, + 0.006442104931920767, + 0.005206944886595011, + 0.006768818479031324, + 0.005078781861811876, + 0.008291088975965977, + 0.0052399407140910625, + 0.004290223587304354, + 0.004271858837455511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.03370893374085426 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048655007034540176 + ], + "opposite_logprob_tokens": [ + -3.0486550331115723 + ], + "logprob_reward": -0.048655007034540176, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048655007034540176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02128387615084648, + 0.019756432622671127, + 0.016720421612262726, + 0.02753552794456482, + 0.039584048092365265, + 0.03068893775343895, + 0.7329559922218323, + 0.0339030958712101, + 0.04659456014633179, + 0.030977148562669754 + ], + "credit_map": [ + 0.0017043681582435966, + 0.0033761358354240656, + 0.0018158858874812722, + 0.0038787699304521084, + 0.0022518965415656567, + 0.0035259155556559563, + 0.0028110758867114782, + 0.0019198275404050946, + 0.003165443195030093, + 0.001702016219496727, + 0.0034566675312817097, + 0.0028303232975304127, + 0.0018089318182319403, + 0.0026325900107622147, + 0.0024939230643212795, + 0.0016665381845086813, + 0.0011671489337459207, + 0.00256905285641551, + 0.0012569664977490902, + 0.0022668270394206047, + 0.0024173487909138203, + 0.00287665450014174, + 0.0020199900027364492, + 0.0021464317105710506, + 0.0032248205970972776, + 0.004449693486094475, + 0.002516979118809104, + 0.005756194703280926, + 0.0034242444671690464, + 0.002757460344582796, + 0.002535446546971798, + 0.0028706879820674658, + 0.002226688899099827, + 0.009110786952078342, + 0.0023507406003773212, + 0.0035116667859256268, + 0.0030336815398186445, + 0.012796472758054733, + 0.002952708164229989, + 0.0036013047210872173, + 0.00323228957131505, + 0.002551668556407094, + 0.003691911930218339, + 0.003082381561398506, + 0.0034184136893600225, + 0.0038662231527268887, + 0.006294235121458769, + 0.004551812540739775, + 0.034615084528923035, + 0.00799406785517931, + 0.03226544335484505, + 0.01751672849059105, + 0.05420980229973793, + 0.029389604926109314, + 0.27194690704345703, + 0.28501829504966736, + 0.0035466209519654512, + 0.0051506259478628635, + 0.0035922930110245943, + 0.0056152623146772385, + 0.004449356347322464, + 0.005114744417369366, + 0.0034669709857553244, + 0.002967222360894084, + 0.006645651068538427, + 0.0028315435629338026, + 0.007803143467754126, + 0.003285979153588414, + 0.00860217772424221, + 0.004734043963253498, + 0.005295370705425739, + 0.00739665050059557, + 0.0045497664250433445, + 0.0031138283666223288, + 0.004435434006154537, + 0.004095863085240126, + 0.003827533684670925, + 0.004532500635832548, + 0.00318894418887794, + 0.0032332795672118664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.048655007034540176 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a penthouse with modern urban chic decor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an Asian woman wearing a dark green silk reworked qipao visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman deliver a fierce kick with her right leg clad in a black stiletto towards crystal-clear art glass pieces on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera descend rapidly from a height to near ground level and capture an extreme slow-motion close-up of the glass bursting into fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor made of mirror-polished black marble reflecting a bustling nightscape with neon lights outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780227206647396 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780227206647396, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780227206647396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02438143454492092, + 0.011423608288168907, + 0.01336174551397562, + 0.01831885240972042, + 0.024252155795693398, + 0.026911118999123573, + 0.617520272731781, + 0.1460738182067871, + 0.07543531805276871, + 0.04232166334986687 + ], + "credit_map": [ + 0.002103425096720457, + 0.003266529878601432, + 0.002953706542029977, + 0.0029692482203245163, + 0.002832036465406418, + 0.0035297148860991, + 0.004157533403486013, + 0.0025692421477288008, + 0.0012556310975924134, + 0.0016369769582524896, + 0.001082247355952859, + 0.0019595841877162457, + 0.0010105076944455504, + 0.0016077051404863596, + 0.0016387575305998325, + 0.0012321986723691225, + 0.0011908219894394279, + 0.0015647703548893332, + 0.0012527784565463662, + 0.0013960943324491382, + 0.002136231167241931, + 0.002382498700171709, + 0.0017204248579218984, + 0.001718125189654529, + 0.003093152306973934, + 0.002685528015717864, + 0.0024727648124098778, + 0.0022462382912635803, + 0.002225134987384081, + 0.0015052692033350468, + 0.002096952637657523, + 0.0019938128534704447, + 0.0019381761085242033, + 0.004609489813446999, + 0.0017440697411075234, + 0.0031348583288490772, + 0.003388444660231471, + 0.0035658376291394234, + 0.002827266464009881, + 0.003044015495106578, + 0.0018734540790319443, + 0.003755474928766489, + 0.00252556218765676, + 0.003978463355451822, + 0.0033637320157140493, + 0.0037465321365743876, + 0.0039888243190944195, + 0.0036790769081562757, + 0.029986761510372162, + 0.006758894771337509, + 0.027332574129104614, + 0.015593593008816242, + 0.046369217336177826, + 0.026441317051649094, + 0.22436979413032532, + 0.24066811800003052, + 0.01611427031457424, + 0.006375111639499664, + 0.016599014401435852, + 0.008139780722558498, + 0.0172698013484478, + 0.023096797987818718, + 0.029079195111989975, + 0.029399843886494637, + 0.009214472025632858, + 0.013833594508469105, + 0.008053862489759922, + 0.011891715228557587, + 0.008744706399738789, + 0.006617954466491938, + 0.010538844391703606, + 0.006540161091834307, + 0.005774363875389099, + 0.0034514861181378365, + 0.006419417914003134, + 0.00490932073444128, + 0.005440987646579742, + 0.00628438126295805, + 0.004706636071205139, + 0.005335070658475161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.009780227206647396 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_14/index_tensor([7483, 7483])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_14/index_tensor([7483, 7483])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ed32a4d6cd79b91955957a84735bc3129a6603dc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_14/index_tensor([7483, 7483])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, shot from an extremely low angle, focuses squarely on a chubby groundhog. In the bright, warm afternoon sun, a dark earthen burrow stands out in the center of a vibrant, emerald meadow, its edges littered with fresh dirt. The groundhog's chubby body, covered in shaggy, coarse brown fur, casts a beautiful golden hue around the edges in the sunlight. Its head cautiously pokes out of the burrow, its small, obsidian eyes darting rapidly from side to side, its nose and whiskers twitching rapidly as it assesses its surroundings. After a brief pause, it swiftly pushes its hind legs and nimbly pulls its entire body completely out of the burrow, finally resting its forepaws on the lip, its upper body erect, maintaining a posture of high alert. The sunlight shines through the grass, casting mottled light and shadows on the ground. A few small white and yellow wildflowers dot the background. The whole picture is warm and bright in color, creating an atmosphere that is tranquil and natural yet full of alertness to life.", + "seed": 18478, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_14/index_7483_seed_18478_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05485405437648296, + "vlm_logprob": -0.05485405437648296, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.05485405437648296, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14439153671264648 + ], + "opposite_logprob_tokens": [ + -2.1443915367126465 + ], + "logprob_reward": -0.14439153671264648, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.14439153671264648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060409124940633774, + 0.032486479729413986, + 0.03248495236039162, + 0.09153624624013901, + 0.03448975831270218, + 0.38880592584609985, + 0.16482572257518768, + 0.048373494297266006, + 0.027796315029263496, + 0.11879194527864456 + ], + "credit_map": [ + 0.007428974378854036, + 0.006886533461511135, + 0.007965080440044403, + 0.007102932780981064, + 0.006476264912635088, + 0.009005101397633553, + 0.007571293506771326, + 0.007972950115799904, + 0.00416987668722868, + 0.0030500919092446566, + 0.00532125448808074, + 0.003178045153617859, + 0.003940877504646778, + 0.003714611055329442, + 0.005203843582421541, + 0.0039078835397958755, + 0.003708504606038332, + 0.0041848598048090935, + 0.0034437866415828466, + 0.0028197113424539566, + 0.006554123479872942, + 0.0040240478701889515, + 0.003554132767021656, + 0.004195788409560919, + 0.0074393306858837605, + 0.01855737343430519, + 0.014610949903726578, + 0.011074877344071865, + 0.005721293389797211, + 0.007529041729867458, + 0.009211569093167782, + 0.017391817644238472, + 0.0030761719681322575, + 0.0037984480150043964, + 0.0031882638577371836, + 0.007721337024122477, + 0.0037950703408569098, + 0.0031369735952466726, + 0.0055261412635445595, + 0.0042473552748560905, + 0.014681457541882992, + 0.019516369327902794, + 0.00918292161077261, + 0.018943393602967262, + 0.015465755946934223, + 0.017249122262001038, + 0.16215217113494873, + 0.13161475956439972, + 0.015157423913478851, + 0.042322009801864624, + 0.008210490457713604, + 0.013051353394985199, + 0.011239246465265751, + 0.009671647101640701, + 0.016693435609340668, + 0.048480138182640076, + 0.007032150402665138, + 0.006858918350189924, + 0.005360006354749203, + 0.0045128329657018185, + 0.006104948464781046, + 0.0040863286703825, + 0.004476345609873533, + 0.009941965341567993, + 0.0038069873116910458, + 0.004263964481651783, + 0.0031853222753852606, + 0.0033698114566504955, + 0.003535511204972863, + 0.0037934030406177044, + 0.002106956671923399, + 0.003734359983354807, + 0.01023494265973568, + 0.020904239267110825, + 0.009297412820160389, + 0.009525890462100506, + 0.015697292983531952, + 0.009332193061709404, + 0.03537614643573761, + 0.008423841558396816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.14439153671264648 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012560663744807243 + ], + "opposite_logprob_tokens": [ + -4.387560844421387 + ], + "logprob_reward": -0.012560663744807243, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012560663744807243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05942655727267265, + 0.03608876094222069, + 0.06111376732587814, + 0.03963904082775116, + 0.09907922893762589, + 0.41930973529815674, + 0.16500423848628998, + 0.05637703463435173, + 0.021030809730291367, + 0.042930860072374344 + ], + "credit_map": [ + 0.009514578618109226, + 0.007978904992341995, + 0.0067505743354558945, + 0.007040882483124733, + 0.006632828153669834, + 0.007862050086259842, + 0.00737346475943923, + 0.006273273378610611, + 0.005966458469629288, + 0.00319640408270061, + 0.006195935420691967, + 0.0035176370292901993, + 0.004025877919048071, + 0.004818846937268972, + 0.004309569485485554, + 0.004058032296597958, + 0.007688600569963455, + 0.011938096955418587, + 0.007701136637479067, + 0.004287843592464924, + 0.010493551380932331, + 0.00709459837526083, + 0.0054330648854374886, + 0.006476873066276312, + 0.0031483308412134647, + 0.005488700699061155, + 0.004883847199380398, + 0.008310055360198021, + 0.0036052304785698652, + 0.004060110542923212, + 0.005664493422955275, + 0.004478271119296551, + 0.0071827732026577, + 0.010430408641695976, + 0.00889329519122839, + 0.014885001815855503, + 0.008092142641544342, + 0.00655582919716835, + 0.017888618633151054, + 0.02515116147696972, + 0.018202224746346474, + 0.03269385173916817, + 0.0141935208812356, + 0.025468599051237106, + 0.022511346265673637, + 0.025869697332382202, + 0.15831027925014496, + 0.12206023931503296, + 0.02132718451321125, + 0.04631559178233147, + 0.010319889523088932, + 0.014796646311879158, + 0.01458559837192297, + 0.012490120716392994, + 0.011687660589814186, + 0.03348156809806824, + 0.008268754929304123, + 0.012444517575204372, + 0.004008166957646608, + 0.004706181585788727, + 0.008930694311857224, + 0.004250687547028065, + 0.005359863396733999, + 0.008408162742853165, + 0.0023998108226805925, + 0.0034138415940105915, + 0.002244436414912343, + 0.0027340385131537914, + 0.0028827032074332237, + 0.002830487210303545, + 0.001952214865013957, + 0.0025732768699526787, + 0.003456136677414179, + 0.006056876387447119, + 0.00416210014373064, + 0.007462764158844948, + 0.006873591803014278, + 0.006092953961342573, + 0.005911148153245449, + 0.002915291115641594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.012560663744807243 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955095402896404 + ], + "opposite_logprob_tokens": [ + -5.130955219268799 + ], + "logprob_reward": -0.005955095402896404, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955095402896404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11679747700691223, + 0.12330969423055649, + 0.1464824378490448, + 0.08151131123304367, + 0.05409737303853035, + 0.24905504286289215, + 0.08208604902029037, + 0.04941665381193161, + 0.04652104526758194, + 0.05072284862399101 + ], + "credit_map": [ + 0.013695922680199146, + 0.01854817382991314, + 0.012708032503724098, + 0.01455398928374052, + 0.012792992405593395, + 0.01755514368414879, + 0.015105953440070152, + 0.01183728314936161, + 0.01828324794769287, + 0.01410109456628561, + 0.013253740966320038, + 0.008657154627144337, + 0.02309253439307213, + 0.017283253371715546, + 0.01403572503477335, + 0.014602961018681526, + 0.021431513130664825, + 0.022169072180986404, + 0.02205760031938553, + 0.013242022134363651, + 0.02238386683166027, + 0.015855202451348305, + 0.013181624934077263, + 0.01616155169904232, + 0.005710602272301912, + 0.024235587567090988, + 0.0044283089227974415, + 0.005310130771249533, + 0.006999413948506117, + 0.010160298086702824, + 0.005238556303083897, + 0.019428417086601257, + 0.006368826609104872, + 0.008029010146856308, + 0.007879679091274738, + 0.00702892430126667, + 0.0054332418367266655, + 0.0052545624785125256, + 0.007066251710057259, + 0.007036881521344185, + 0.00917494110763073, + 0.013312268070876598, + 0.006180297117680311, + 0.011366987600922585, + 0.010174197144806385, + 0.01166574377566576, + 0.10122663527727127, + 0.08595399558544159, + 0.00827985629439354, + 0.02324274741113186, + 0.005906990263611078, + 0.008725718595087528, + 0.0072563411667943, + 0.006551267579197884, + 0.00908055528998375, + 0.013042581267654896, + 0.007292511407285929, + 0.017277533188462257, + 0.003393947146832943, + 0.004665410611778498, + 0.006576559506356716, + 0.0035121282562613487, + 0.0025798347778618336, + 0.0041187359020113945, + 0.00608131242915988, + 0.007912558503448963, + 0.004655804019421339, + 0.006685319822281599, + 0.006589923519641161, + 0.006262824870646, + 0.0036292881704866886, + 0.004704019520431757, + 0.004451028071343899, + 0.00910487212240696, + 0.005225236061960459, + 0.006148490589112043, + 0.0051970588974654675, + 0.002575031016021967, + 0.015087385661900043, + 0.00293374783359468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.005955095402896404 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078749783337116 + ], + "opposite_logprob_tokens": [ + -4.511078834533691 + ], + "logprob_reward": -0.011078749783337116, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078749783337116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09325231611728668, + 0.03921990841627121, + 0.028562434017658234, + 0.052030257880687714, + 0.09265482425689697, + 0.5471870303153992, + 0.07560383528470993, + 0.013128910213708878, + 0.015814876183867455, + 0.0425456278026104 + ], + "credit_map": [ + 0.010484224185347557, + 0.013036129996180534, + 0.01808932051062584, + 0.00896504521369934, + 0.013119014911353588, + 0.010843068361282349, + 0.009464092552661896, + 0.009251421317458153, + 0.005164228845387697, + 0.0039042120333760977, + 0.008807368576526642, + 0.003440170083194971, + 0.005220900289714336, + 0.004855541046708822, + 0.0037740403786301613, + 0.004053449723869562, + 0.002368590561673045, + 0.0035901446826756, + 0.002765342127531767, + 0.002018904546275735, + 0.007937682792544365, + 0.003603630233556032, + 0.0019699716940522194, + 0.0043081664480268955, + 0.004479266703128815, + 0.006585205905139446, + 0.005689381621778011, + 0.011212339624762535, + 0.004429281689226627, + 0.0035641661379486322, + 0.007807024288922548, + 0.008263591676950455, + 0.007332528010010719, + 0.01370678748935461, + 0.009587653912603855, + 0.013136658817529678, + 0.007167763076722622, + 0.006946065463125706, + 0.014404965564608574, + 0.020372403785586357, + 0.020444445312023163, + 0.025295250117778778, + 0.013652530498802662, + 0.02640860714018345, + 0.020445508882403374, + 0.024432774633169174, + 0.2277936488389969, + 0.18871426582336426, + 0.00713955145329237, + 0.020519543439149857, + 0.005947406869381666, + 0.00732793565839529, + 0.005366758443415165, + 0.004978276789188385, + 0.007051033899188042, + 0.01727333664894104, + 0.0017769180703908205, + 0.0034608948044478893, + 0.0011271985713392496, + 0.00114694784861058, + 0.0015150104882195592, + 0.0010647948365658522, + 0.0010970898438245058, + 0.0019400552846491337, + 0.00240247487090528, + 0.002436525421217084, + 0.0016288754995912313, + 0.00268879858776927, + 0.001715178252197802, + 0.001974254148080945, + 0.001373389852233231, + 0.0015953790862113237, + 0.003875260939821601, + 0.0073273153975605965, + 0.0034915704745799303, + 0.003718370571732521, + 0.0074916197918355465, + 0.005534157622605562, + 0.0074440957978367805, + 0.00366323534399271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.011078749783337116 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028422623872757 + ], + "opposite_logprob_tokens": [ + -2.3502843379974365 + ], + "logprob_reward": -0.10028422623872757, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028422623872757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16253662109375, + 0.055729806423187256, + 0.2717573940753937, + 0.03978251293301582, + 0.03674823045730591, + 0.15228362381458282, + 0.07931061089038849, + 0.10445977002382278, + 0.026546495035290718, + 0.07084490358829498 + ], + "credit_map": [ + 0.018541989848017693, + 0.014425849542021751, + 0.05112006515264511, + 0.014930037781596184, + 0.017933571711182594, + 0.015263376757502556, + 0.01442006602883339, + 0.015901658684015274, + 0.007064483128488064, + 0.004746821243315935, + 0.011516041122376919, + 0.006013320293277502, + 0.00758944870904088, + 0.004829857032746077, + 0.007704225834459066, + 0.006265613716095686, + 0.03728938102722168, + 0.04384364187717438, + 0.04102766141295433, + 0.02513887919485569, + 0.03556716442108154, + 0.032099075615406036, + 0.02691018395125866, + 0.02988140657544136, + 0.0042562345042824745, + 0.006348948925733566, + 0.006916832644492388, + 0.005408550146967173, + 0.0037094976287335157, + 0.004002207424491644, + 0.004261307418346405, + 0.004878934007138014, + 0.003296641167253256, + 0.005893270950764418, + 0.0042954375967383385, + 0.004846719093620777, + 0.004485802724957466, + 0.004156188108026981, + 0.00525103323161602, + 0.0045231361873447895, + 0.0093093765899539, + 0.012909937649965286, + 0.00727661931887269, + 0.012554972432553768, + 0.007607489358633757, + 0.010303817689418793, + 0.05282173678278923, + 0.0394996777176857, + 0.008236179128289223, + 0.022495847195386887, + 0.0048383246175944805, + 0.008983765728771687, + 0.007735466584563255, + 0.006311473902314901, + 0.006412046495825052, + 0.014297506771981716, + 0.021079158410429955, + 0.020081941038370132, + 0.00913416501134634, + 0.009797506965696812, + 0.016893593594431877, + 0.009103478863835335, + 0.008617909625172615, + 0.00975201278924942, + 0.00254321307875216, + 0.005522649269551039, + 0.0023327942471951246, + 0.005272034555673599, + 0.0038031728472560644, + 0.0022579918149858713, + 0.002399534685537219, + 0.002415106166154146, + 0.0072820852510631084, + 0.014154398813843727, + 0.0061867693439126015, + 0.00620837789028883, + 0.007434828672558069, + 0.0037385071627795696, + 0.02178560569882393, + 0.004054329823702574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.10028422623872757 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from an extremely low angle, focuses squarely on a chubby groundhog. In the bright, warm afternoon sun, a dark earthen burrow stands out in the center of a vibrant, emerald meadow, its edges littered with fresh dirt. The groundhog's chubby body, covered in shaggy, coarse brown fur, casts a beautiful golden hue around the edges in the sunlight. Its head cautiously pokes out of the burrow, its small, obsidian eyes darting rapidly from side to side, its nose and whiskers twitching rapidly as it assesses its surroundings. After a brief pause, it swiftly pushes its hind legs and nimbly pulls its entire body completely out of the burrow, finally resting its forepaws on the lip, its upper body erect, maintaining a posture of high alert. The sunlight shines through the grass, casting mottled light and shadows on the ground. A few small white and yellow wildflowers dot the background. The whole picture is warm and bright in color, creating an atmosphere that is tranquil and natural yet full of alertness to life.", + "seed": 18479, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_14/index_7483_seed_18479_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023728868924081326, + "vlm_logprob": -0.023728868924081326, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.023728868924081326, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04794928431510925 + ], + "opposite_logprob_tokens": [ + -3.5479493141174316 + ], + "logprob_reward": -0.04794928431510925, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.04794928431510925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17691351473331451, + 0.07947643846273422, + 0.07353205233812332, + 0.06680835038423538, + 0.03110208921134472, + 0.15499377250671387, + 0.13091595470905304, + 0.09745625406503677, + 0.08686704933643341, + 0.1019345298409462 + ], + "credit_map": [ + 0.018555117771029472, + 0.03886611387133598, + 0.01881904900074005, + 0.016868004575371742, + 0.02458822727203369, + 0.01608138531446457, + 0.02054714970290661, + 0.022588452324271202, + 0.004390163812786341, + 0.02626201882958412, + 0.005144075024873018, + 0.0038118381053209305, + 0.02292630635201931, + 0.00581204891204834, + 0.005017691757529974, + 0.006112297996878624, + 0.005324187688529491, + 0.008311181329190731, + 0.009320966899394989, + 0.01227872259914875, + 0.011559811420738697, + 0.007444523740559816, + 0.012830519117414951, + 0.0064621358178555965, + 0.00737693952396512, + 0.007526877801865339, + 0.007139582186937332, + 0.008671920746564865, + 0.007494120392948389, + 0.008611363358795643, + 0.010803277604281902, + 0.009184266440570354, + 0.005620194133371115, + 0.006212701555341482, + 0.0031415284611284733, + 0.0027978303842246532, + 0.0032159644179046154, + 0.0035022946540266275, + 0.0033192518167197704, + 0.0032923256512731314, + 0.03001539595425129, + 0.050011128187179565, + 0.010926301591098309, + 0.010797349736094475, + 0.011183282360434532, + 0.019269319251179695, + 0.010399332270026207, + 0.012391657568514347, + 0.03946669399738312, + 0.03731130436062813, + 0.010786435566842556, + 0.013613689690828323, + 0.00888759084045887, + 0.009336856193840504, + 0.005916155409067869, + 0.005597231909632683, + 0.014538966119289398, + 0.00837469007819891, + 0.010236076079308987, + 0.006981593556702137, + 0.008373807184398174, + 0.0302690751850605, + 0.011980519630014896, + 0.006701528560370207, + 0.006672461051493883, + 0.008242380805313587, + 0.011392689310014248, + 0.008238033391535282, + 0.009550251998007298, + 0.024267490953207016, + 0.012560475617647171, + 0.005943263880908489, + 0.020752880722284317, + 0.006227970123291016, + 0.01529922429472208, + 0.005666816141456366, + 0.00854724645614624, + 0.009379984810948372, + 0.027797622606158257, + 0.008262785151600838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.04794928431510925 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009793920442461967 + ], + "opposite_logprob_tokens": [ + -4.634793758392334 + ], + "logprob_reward": -0.009793920442461967, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009793920442461967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12682466208934784, + 0.10864215344190598, + 0.11340024322271347, + 0.06117180362343788, + 0.04024146869778633, + 0.07772819697856903, + 0.10315696895122528, + 0.08348830044269562, + 0.19095227122306824, + 0.09439393877983093 + ], + "credit_map": [ + 0.015003778971731663, + 0.02046336978673935, + 0.012742212042212486, + 0.009024696424603462, + 0.026349470019340515, + 0.015273775905370712, + 0.011024603620171547, + 0.01694275625050068, + 0.009113145992159843, + 0.011083617806434631, + 0.009114731103181839, + 0.006925405468791723, + 0.017108986154198647, + 0.03325182944536209, + 0.009107911959290504, + 0.012936513870954514, + 0.009543667547404766, + 0.014096924103796482, + 0.007623299024999142, + 0.011691461317241192, + 0.023436816409230232, + 0.012058920226991177, + 0.016371695324778557, + 0.0185774564743042, + 0.009009788744151592, + 0.011258065700531006, + 0.005748045165091753, + 0.005168019328266382, + 0.008255618624389172, + 0.008245738223195076, + 0.005770419724285603, + 0.00771610951051116, + 0.0038448679260909557, + 0.0037694028578698635, + 0.007893718779087067, + 0.004637022502720356, + 0.005663844756782055, + 0.0035412050783634186, + 0.0067693572491407394, + 0.0041220523416996, + 0.012376298196613789, + 0.01984417624771595, + 0.009861905127763748, + 0.006650046445429325, + 0.008305526338517666, + 0.0071921576745808125, + 0.00663111312314868, + 0.006866978015750647, + 0.010862349532544613, + 0.0175709817558527, + 0.009871500544250011, + 0.02287587709724903, + 0.009865464642643929, + 0.010271045379340649, + 0.011527970433235168, + 0.010311773046851158, + 0.018824202939867973, + 0.007789595518261194, + 0.01086517609655857, + 0.005922320298850536, + 0.009432494640350342, + 0.010915113613009453, + 0.013422028161585331, + 0.006317366845905781, + 0.04348692297935486, + 0.007675287313759327, + 0.021871106699109077, + 0.009245560504496098, + 0.013475933112204075, + 0.07006122171878815, + 0.014532935805618763, + 0.010603303089737892, + 0.013872955925762653, + 0.008079037070274353, + 0.009404070675373077, + 0.01023336872458458, + 0.010369010269641876, + 0.022436203435063362, + 0.008502792567014694, + 0.011496506631374359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.009793920442461967 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003621806390583515 + ], + "opposite_logprob_tokens": [ + -5.628621578216553 + ], + "logprob_reward": -0.003621806390583515, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003621806390583515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1380705088376999, + 0.3096249997615814, + 0.08731216192245483, + 0.018387190997600555, + 0.012430462054908276, + 0.0510571114718914, + 0.04801631718873978, + 0.03393420949578285, + 0.24917511641979218, + 0.05199188366532326 + ], + "credit_map": [ + 0.007759542670100927, + 0.04012194648385048, + 0.00975080393254757, + 0.010292461141943932, + 0.03602997958660126, + 0.00951753556728363, + 0.011323759332299232, + 0.013274472206830978, + 0.028064467012882233, + 0.05104316771030426, + 0.02347869612276554, + 0.019705379381775856, + 0.04459093138575554, + 0.07949616760015488, + 0.024319089949131012, + 0.0389271043241024, + 0.007575733587145805, + 0.010211247950792313, + 0.006639421451836824, + 0.012730976566672325, + 0.012925882823765278, + 0.01271098293364048, + 0.015820050612092018, + 0.008697867393493652, + 0.0018393599893897772, + 0.002083615632727742, + 0.0013135260669514537, + 0.0017570297932252288, + 0.002450494561344385, + 0.0023179634008556604, + 0.004164579324424267, + 0.002460621064528823, + 0.0011790088610723615, + 0.0012463292805477977, + 0.001212534960359335, + 0.0021811663173139095, + 0.001425307709723711, + 0.0011988027254119515, + 0.002251386409625411, + 0.0017359256744384766, + 0.006318570580333471, + 0.012403804808855057, + 0.0036689203698188066, + 0.005688723176717758, + 0.006019253749400377, + 0.005436017643660307, + 0.004436637740582228, + 0.0070851827040314674, + 0.009011628106236458, + 0.004034190904349089, + 0.003916924353688955, + 0.011584983207285404, + 0.008405265398323536, + 0.004182473290711641, + 0.0039814068004488945, + 0.002899447688832879, + 0.005448188632726669, + 0.004284312948584557, + 0.003711434081196785, + 0.003374060383066535, + 0.0048075770027935505, + 0.003395209787413478, + 0.004913009703159332, + 0.004000418819487095, + 0.04714270308613777, + 0.009800532832741737, + 0.0198932196944952, + 0.008144062012434006, + 0.015553037635982037, + 0.10119087994098663, + 0.03795401006937027, + 0.00949664693325758, + 0.0067399549297988415, + 0.0036835581995546818, + 0.006724006962031126, + 0.005090624559670687, + 0.0064760539680719376, + 0.006997647695243359, + 0.01060406118631363, + 0.005675976630300283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.003621806390583515 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864158570766449 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.00864158570766449, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864158570766449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12654508650302887, + 0.21472010016441345, + 0.09760650992393494, + 0.0204290971159935, + 0.026097798720002174, + 0.1262117475271225, + 0.06823506951332092, + 0.03809286653995514, + 0.23640315234661102, + 0.045658595860004425 + ], + "credit_map": [ + 0.006240051239728928, + 0.044200584292411804, + 0.006563706323504448, + 0.004797344096004963, + 0.04128256440162659, + 0.007108473218977451, + 0.007372342050075531, + 0.00898001715540886, + 0.02073616161942482, + 0.04086000472307205, + 0.014040575362741947, + 0.012441511265933514, + 0.03268805518746376, + 0.053204309195280075, + 0.015291593968868256, + 0.02545786090195179, + 0.007832891307771206, + 0.010440342128276825, + 0.00880249310284853, + 0.016572775319218636, + 0.012817274779081345, + 0.013398583047091961, + 0.016763774678111076, + 0.01097836997359991, + 0.0017602202715352178, + 0.0020287944935262203, + 0.0023148702457547188, + 0.0021624716464430094, + 0.0020731596741825342, + 0.0025592194870114326, + 0.0042041908018291, + 0.0033261689823120832, + 0.0024505495093762875, + 0.0038658934645354748, + 0.0023639374412596226, + 0.0034956170711666346, + 0.003946424461901188, + 0.003399516921490431, + 0.003208385780453682, + 0.003367472905665636, + 0.016390152275562286, + 0.029752597212791443, + 0.009190068580210209, + 0.014036267064511776, + 0.013640421442687511, + 0.013585025444626808, + 0.011114075779914856, + 0.018503140658140182, + 0.013655762188136578, + 0.02271850034594536, + 0.006926734000444412, + 0.0069000511430203915, + 0.004153718240559101, + 0.0039131599478423595, + 0.005309354048222303, + 0.004657784942537546, + 0.0037509670946747065, + 0.0038154558278620243, + 0.0035785194486379623, + 0.0023254479747265577, + 0.004500476643443108, + 0.008772570639848709, + 0.006617449223995209, + 0.00473197503015399, + 0.05311530455946922, + 0.00910472683608532, + 0.024369683116674423, + 0.006929344963282347, + 0.011618384160101414, + 0.09892084449529648, + 0.023458268493413925, + 0.008886593393981457, + 0.004679790697991848, + 0.004412153270095587, + 0.005632818676531315, + 0.007872349582612514, + 0.00822817999869585, + 0.006215987261384726, + 0.00437585636973381, + 0.004241457674652338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.00864158570766449 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486377477645874 + ], + "opposite_logprob_tokens": [ + -3.048637866973877 + ], + "logprob_reward": -0.0486377477645874, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486377477645874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1552727073431015, + 0.19156749546527863, + 0.07079526782035828, + 0.021762749180197716, + 0.07158517092466354, + 0.11661078035831451, + 0.12787510454654694, + 0.04763950780034065, + 0.12989208102226257, + 0.06699913740158081 + ], + "credit_map": [ + 0.019773706793785095, + 0.022746115922927856, + 0.016587214544415474, + 0.014904049225151539, + 0.029912319034337997, + 0.01947070099413395, + 0.014669740572571754, + 0.01720886491239071, + 0.034135352820158005, + 0.027696210891008377, + 0.012874122709035873, + 0.015231169760227203, + 0.028209907934069633, + 0.0338091179728508, + 0.01605050638318062, + 0.023561108857393265, + 0.014058507047593594, + 0.01049695536494255, + 0.004433002322912216, + 0.006317045073956251, + 0.00965921301394701, + 0.00990747008472681, + 0.007566001266241074, + 0.008357075043022633, + 0.002467625541612506, + 0.002799984533339739, + 0.001972513273358345, + 0.0022952586878091097, + 0.00345601886510849, + 0.002575797727331519, + 0.0033506513573229313, + 0.002844897797331214, + 0.007793993689119816, + 0.011947538703680038, + 0.009913178160786629, + 0.007670612074434757, + 0.008507479913532734, + 0.008592560887336731, + 0.008040383458137512, + 0.009119424968957901, + 0.011196763254702091, + 0.026380280032753944, + 0.016464054584503174, + 0.010730698704719543, + 0.011160661466419697, + 0.015287958085536957, + 0.011756409890949726, + 0.013633954338729382, + 0.017413577064871788, + 0.0485922247171402, + 0.008049500174820423, + 0.027901969850063324, + 0.008219254203140736, + 0.00672605587169528, + 0.005652605555951595, + 0.005319913383573294, + 0.005870733875781298, + 0.004215341527014971, + 0.004317088983952999, + 0.0032350593246519566, + 0.006782299838960171, + 0.013234036974608898, + 0.004921812564134598, + 0.005063134711235762, + 0.014419332146644592, + 0.006623815279453993, + 0.011085662990808487, + 0.007059626281261444, + 0.011968610808253288, + 0.043493058532476425, + 0.028936171904206276, + 0.006305800750851631, + 0.013922050595283508, + 0.007070706691592932, + 0.008668660186231136, + 0.0050674728117883205, + 0.0056905015371739864, + 0.007723110727965832, + 0.01330486498773098, + 0.005551768466830254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0486377477645874 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from an extremely low angle, focuses squarely on a chubby groundhog. In the bright, warm afternoon sun, a dark earthen burrow stands out in the center of a vibrant, emerald meadow, its edges littered with fresh dirt. The groundhog's chubby body, covered in shaggy, coarse brown fur, casts a beautiful golden hue around the edges in the sunlight. Its head cautiously pokes out of the burrow, its small, obsidian eyes darting rapidly from side to side, its nose and whiskers twitching rapidly as it assesses its surroundings. After a brief pause, it swiftly pushes its hind legs and nimbly pulls its entire body completely out of the burrow, finally resting its forepaws on the lip, its upper body erect, maintaining a posture of high alert. The sunlight shines through the grass, casting mottled light and shadows on the ground. A few small white and yellow wildflowers dot the background. The whole picture is warm and bright in color, creating an atmosphere that is tranquil and natural yet full of alertness to life.", + "seed": 18480, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_14/index_7483_seed_18480_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01878112480044365, + "vlm_logprob": -0.01878112480044365, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.01878112480044365, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03240278735756874 + ], + "opposite_logprob_tokens": [ + -4.032402992248535 + ], + "logprob_reward": -0.03240278735756874, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.03240278735756874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1921253204345703, + 0.03587125241756439, + 0.037730369716882706, + 0.05496256425976753, + 0.03499269112944603, + 0.055397454649209976, + 0.19505923986434937, + 0.18019472062587738, + 0.05634084716439247, + 0.15732555091381073 + ], + "credit_map": [ + 0.019986987113952637, + 0.022561727091670036, + 0.021433617919683456, + 0.012352827936410904, + 0.02263074740767479, + 0.042275577783584595, + 0.023569723591208458, + 0.027314122766256332, + 0.003975630272179842, + 0.005515106488019228, + 0.005247830878943205, + 0.0033657930325716734, + 0.003424027469009161, + 0.004508354235440493, + 0.003707880387082696, + 0.0061266301199793816, + 0.0035355137661099434, + 0.004461163189262152, + 0.007050419691950083, + 0.0036549807991832495, + 0.004490547347813845, + 0.005457374732941389, + 0.005351727362722158, + 0.003728643525391817, + 0.005278755910694599, + 0.005863194353878498, + 0.006524214055389166, + 0.005201797932386398, + 0.008425208739936352, + 0.01130901463329792, + 0.005501473322510719, + 0.006858904846012592, + 0.003955140244215727, + 0.00429630046710372, + 0.0028237467631697655, + 0.003394646570086479, + 0.005892812740057707, + 0.00646211439743638, + 0.004088358022272587, + 0.004079567268490791, + 0.010154502466320992, + 0.005181156564503908, + 0.006601089145988226, + 0.0050888666883111, + 0.009517588652670383, + 0.007492652162909508, + 0.0036496149841696024, + 0.007711987942457199, + 0.01281803660094738, + 0.060366299003362656, + 0.01156320609152317, + 0.053936783224344254, + 0.0252347644418478, + 0.011470219120383263, + 0.010173111222684383, + 0.00949682854115963, + 0.019435888156294823, + 0.009293805807828903, + 0.0770656168460846, + 0.019840259104967117, + 0.009748105891048908, + 0.02660323679447174, + 0.009643170982599258, + 0.008564641699194908, + 0.006043963599950075, + 0.004239057190716267, + 0.004630716051906347, + 0.002890105592086911, + 0.0053793867118656635, + 0.015740450471639633, + 0.004950531758368015, + 0.01246663462370634, + 0.015988711267709732, + 0.008653685450553894, + 0.06686810404062271, + 0.011829561553895473, + 0.010092842392623425, + 0.02167794667184353, + 0.014109860174357891, + 0.008104831911623478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.03240278735756874 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764184445142746 + ], + "opposite_logprob_tokens": [ + -4.882641792297363 + ], + "logprob_reward": -0.00764184445142746, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764184445142746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08076845109462738, + 0.027513064444065094, + 0.03405662253499031, + 0.05084241181612015, + 0.025859316810965538, + 0.04235493019223213, + 0.1233978345990181, + 0.19642110168933868, + 0.11933455616235733, + 0.2994517385959625 + ], + "credit_map": [ + 0.009850035421550274, + 0.015106702223420143, + 0.006816932465881109, + 0.00529540004208684, + 0.009153067134320736, + 0.01192230824381113, + 0.007656813599169254, + 0.014967188239097595, + 0.0031267227604985237, + 0.0037448701914399862, + 0.0026029713917523623, + 0.002281902125105262, + 0.0045264423824846745, + 0.0051382784731686115, + 0.002331063849851489, + 0.0037608109414577484, + 0.0033848381135612726, + 0.005040276329964399, + 0.005758310202509165, + 0.004570698365569115, + 0.0030482334550470114, + 0.004687309265136719, + 0.003910833969712257, + 0.003656117245554924, + 0.004944634158164263, + 0.00998075120151043, + 0.007145411800593138, + 0.00553837139159441, + 0.005753100384026766, + 0.006354193203151226, + 0.00531770521774888, + 0.0058082411997020245, + 0.0026205661706626415, + 0.003854453330859542, + 0.0025924895890057087, + 0.003230696078389883, + 0.0035784137435257435, + 0.004226697143167257, + 0.0021393517963588238, + 0.0036166454665362835, + 0.005065380595624447, + 0.004428877495229244, + 0.004336001351475716, + 0.005147953052073717, + 0.010511585511267185, + 0.005455770995467901, + 0.003533262526616454, + 0.0038760965690016747, + 0.0061603146605193615, + 0.04506424441933632, + 0.006558476481586695, + 0.037724677473306656, + 0.011663715355098248, + 0.00597338704392314, + 0.006514721550047398, + 0.0037382831797003746, + 0.019809609279036522, + 0.013210318051278591, + 0.08917123079299927, + 0.0223043505102396, + 0.010063917376101017, + 0.014208492822945118, + 0.015026864595711231, + 0.012626283802092075, + 0.011719903908669949, + 0.008194010704755783, + 0.05089223384857178, + 0.008445249870419502, + 0.006296106614172459, + 0.019222043454647064, + 0.007935239002108574, + 0.006629754323512316, + 0.02178160659968853, + 0.014939057640731335, + 0.10998869687318802, + 0.016720695421099663, + 0.014712697826325893, + 0.020669646561145782, + 0.07801283895969391, + 0.02262648195028305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.00764184445142746 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636639289557934 + ], + "opposite_logprob_tokens": [ + -4.882636547088623 + ], + "logprob_reward": -0.007636639289557934, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636639289557934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22644145786762238, + 0.13698768615722656, + 0.027327988296747208, + 0.061634715646505356, + 0.020097509026527405, + 0.0682370737195015, + 0.09954112023115158, + 0.05460483580827713, + 0.06065533682703972, + 0.24447226524353027 + ], + "credit_map": [ + 0.021339958533644676, + 0.03948977589607239, + 0.028105342760682106, + 0.01665193773806095, + 0.021878253668546677, + 0.019549619406461716, + 0.03926283121109009, + 0.040163733065128326, + 0.013991543091833591, + 0.02089765854179859, + 0.017045792192220688, + 0.010969621129333973, + 0.012784137390553951, + 0.01680573634803295, + 0.023975586518645287, + 0.020517608150839806, + 0.003964077215641737, + 0.0032517241779714823, + 0.0036349077709019184, + 0.0027642420027405024, + 0.004566021729260683, + 0.0037842304445803165, + 0.0024925032630562782, + 0.002870280994102359, + 0.00672261044383049, + 0.011307090520858765, + 0.007554216310381889, + 0.01027380395680666, + 0.00803804025053978, + 0.005836199969053268, + 0.005077091511338949, + 0.006825664546340704, + 0.0018951145466417074, + 0.003163592191413045, + 0.002303797984495759, + 0.00270929466933012, + 0.0028959631454199553, + 0.0032141748815774918, + 0.001743871602229774, + 0.0021716998890042305, + 0.006929881405085325, + 0.007592214271426201, + 0.00668563274666667, + 0.009751575067639351, + 0.014607907272875309, + 0.006657591089606285, + 0.008418973535299301, + 0.007593299262225628, + 0.0046250526793301105, + 0.03450856730341911, + 0.004616588819772005, + 0.0322713777422905, + 0.00890614278614521, + 0.004718335345387459, + 0.005158726125955582, + 0.004736325237900019, + 0.005870386026799679, + 0.0032543791458010674, + 0.021474454551935196, + 0.0063250064849853516, + 0.003707593074068427, + 0.004373057745397091, + 0.003462191903963685, + 0.006137765944004059, + 0.006675240118056536, + 0.0037888013757765293, + 0.019472777843475342, + 0.0034041479229927063, + 0.0036505998577922583, + 0.008025050163269043, + 0.005733399651944637, + 0.009905317798256874, + 0.02803122065961361, + 0.012069566175341606, + 0.1095600500702858, + 0.01586057059466839, + 0.011359116062521935, + 0.016597315669059753, + 0.026067130267620087, + 0.024927277117967606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.007636639289557934 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541004456579685 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012541004456579685, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541004456579685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16391749680042267, + 0.07381348311901093, + 0.02785751409828663, + 0.04209927096962929, + 0.02973242662847042, + 0.06721307337284088, + 0.15145036578178406, + 0.11792811006307602, + 0.029383832588791847, + 0.2966044545173645 + ], + "credit_map": [ + 0.017661964520812035, + 0.03150375559926033, + 0.0186248030513525, + 0.0112946517765522, + 0.01691766455769539, + 0.013758492656052113, + 0.022450745105743408, + 0.03170541301369667, + 0.007107125129550695, + 0.009301173500716686, + 0.00928092934191227, + 0.006240507587790489, + 0.00815787073224783, + 0.010389858856797218, + 0.013396847993135452, + 0.009939165785908699, + 0.003385583171620965, + 0.0033186785876750946, + 0.004694700241088867, + 0.003130692755803466, + 0.004207141697406769, + 0.003383871866390109, + 0.0027958620339632034, + 0.002940983744338155, + 0.004708415362983942, + 0.007124701514840126, + 0.005393230821937323, + 0.006010214798152447, + 0.005611454602330923, + 0.004356681369245052, + 0.0034608463756740093, + 0.005433720536530018, + 0.0028435394633561373, + 0.005277770571410656, + 0.0031293740030378103, + 0.0037518213503062725, + 0.003879664931446314, + 0.004719891585409641, + 0.0027730586007237434, + 0.0033573058899492025, + 0.006961926352232695, + 0.0077657983638346195, + 0.005942978896200657, + 0.00932944193482399, + 0.014359806664288044, + 0.007892931811511517, + 0.008752605877816677, + 0.006207580678164959, + 0.007625738624483347, + 0.05584915354847908, + 0.007914486341178417, + 0.047529563307762146, + 0.01465914212167263, + 0.0057955048978328705, + 0.007672358304262161, + 0.00440441956743598, + 0.011893080547451973, + 0.007815968245267868, + 0.051416363567113876, + 0.014183430932462215, + 0.007025591563433409, + 0.006924898829311132, + 0.010304750874638557, + 0.00836402177810669, + 0.0025816152337938547, + 0.0031206579878926277, + 0.009601622819900513, + 0.0027031777426600456, + 0.0025724133010953665, + 0.0041773016564548016, + 0.00252012861892581, + 0.0021069154608994722, + 0.033876221626996994, + 0.013847709633409977, + 0.1336134374141693, + 0.019481049850583076, + 0.013573378324508667, + 0.02283872477710247, + 0.03198834881186485, + 0.027385562658309937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.012541004456579685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368334844708443 + ], + "opposite_logprob_tokens": [ + -3.4086833000183105 + ], + "logprob_reward": -0.03368334844708443, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368334844708443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1293737143278122, + 0.24282251298427582, + 0.11042140424251556, + 0.06547172367572784, + 0.04104486480355263, + 0.05413166806101799, + 0.12945744395256042, + 0.126387819647789, + 0.03421448916196823, + 0.06667433679103851 + ], + "credit_map": [ + 0.014450560323894024, + 0.016553489491343498, + 0.013567942194640636, + 0.010596067644655704, + 0.023432094603776932, + 0.022581804543733597, + 0.01259590219706297, + 0.015595878474414349, + 0.03202521800994873, + 0.022421469911932945, + 0.02825295925140381, + 0.018251243978738785, + 0.03218218311667442, + 0.04624554142355919, + 0.03637361899018288, + 0.027070287615060806, + 0.012183641083538532, + 0.010500692762434483, + 0.020597146824002266, + 0.011382080614566803, + 0.014908064156770706, + 0.01626485027372837, + 0.012727226130664349, + 0.011857708916068077, + 0.0064187911339104176, + 0.013139983639121056, + 0.00927596166729927, + 0.0070280698128044605, + 0.0073620472103357315, + 0.008241912350058556, + 0.006489685270935297, + 0.007515273988246918, + 0.004394998773932457, + 0.0062875282019376755, + 0.004914573859423399, + 0.004915433470159769, + 0.005679999012500048, + 0.005439125932753086, + 0.0051381029188632965, + 0.004275104962289333, + 0.006608400959521532, + 0.005848810076713562, + 0.0058774445205926895, + 0.00501288240775466, + 0.013579401187598705, + 0.007026377599686384, + 0.0052969008684158325, + 0.004881454166024923, + 0.007465471979230642, + 0.03902802616357803, + 0.014954086393117905, + 0.03186594694852829, + 0.014087293297052383, + 0.008848079480230808, + 0.007489149458706379, + 0.005719396751374006, + 0.013845637440681458, + 0.006230694241821766, + 0.0557107999920845, + 0.013273792341351509, + 0.0063318004831671715, + 0.016792692244052887, + 0.006957171484827995, + 0.007245229557156563, + 0.003160372143611312, + 0.002408882835879922, + 0.007635599933564663, + 0.0028743185102939606, + 0.0027397116646170616, + 0.0075506106950342655, + 0.004304337315261364, + 0.0035406574606895447, + 0.0053755440749228, + 0.002759616356343031, + 0.02574320323765278, + 0.0033461805433034897, + 0.003747551469132304, + 0.0031115268357098103, + 0.018703432753682137, + 0.0038872924633324146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.03368334844708443 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from an extremely low angle, focuses squarely on a chubby groundhog. In the bright, warm afternoon sun, a dark earthen burrow stands out in the center of a vibrant, emerald meadow, its edges littered with fresh dirt. The groundhog's chubby body, covered in shaggy, coarse brown fur, casts a beautiful golden hue around the edges in the sunlight. Its head cautiously pokes out of the burrow, its small, obsidian eyes darting rapidly from side to side, its nose and whiskers twitching rapidly as it assesses its surroundings. After a brief pause, it swiftly pushes its hind legs and nimbly pulls its entire body completely out of the burrow, finally resting its forepaws on the lip, its upper body erect, maintaining a posture of high alert. The sunlight shines through the grass, casting mottled light and shadows on the ground. A few small white and yellow wildflowers dot the background. The whole picture is warm and bright in color, creating an atmosphere that is tranquil and natural yet full of alertness to life.", + "seed": 18481, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_14/index_7483_seed_18481_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02237780876457691, + "vlm_logprob": -0.02237780876457691, + "vlm_yes_no_margin": 4.299999952316284, + "vlm_token_credit": -0.02237780876457691, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03424324095249176 + ], + "opposite_logprob_tokens": [ + -4.034243106842041 + ], + "logprob_reward": -0.03424324095249176, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.03424324095249176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18064679205417633, + 0.022185612469911575, + 0.04764430224895477, + 0.05905386060476303, + 0.25205889344215393, + 0.08739005029201508, + 0.08602942526340485, + 0.04056224599480629, + 0.16991212964057922, + 0.0545167401432991 + ], + "credit_map": [ + 0.013754661194980145, + 0.015666980296373367, + 0.02131226845085621, + 0.04596675932407379, + 0.027934467419981956, + 0.016258440911769867, + 0.02643858641386032, + 0.013314601965248585, + 0.002496988046914339, + 0.0024866340681910515, + 0.0030890803318470716, + 0.003755736630409956, + 0.002926816465333104, + 0.0023426078259944916, + 0.0027913260273635387, + 0.0022964212112128735, + 0.008577086962759495, + 0.0057800523936748505, + 0.0046194917522370815, + 0.006627500988543034, + 0.004351514857262373, + 0.010264220647513866, + 0.0032938390504568815, + 0.004130592569708824, + 0.006948741152882576, + 0.007205937057733536, + 0.006862557493150234, + 0.006811263505369425, + 0.00943034514784813, + 0.0065903267823159695, + 0.008932465687394142, + 0.006272220052778721, + 0.08968822658061981, + 0.013744652271270752, + 0.012012538500130177, + 0.016074441373348236, + 0.011287473142147064, + 0.085858054459095, + 0.017025506123900414, + 0.006367990281432867, + 0.010816003195941448, + 0.015228373929858208, + 0.009687588550150394, + 0.008995303884148598, + 0.009022284299135208, + 0.017690198495984077, + 0.010124396532773972, + 0.005825900472700596, + 0.011107485741376877, + 0.015460572205483913, + 0.007334411144256592, + 0.007517698686569929, + 0.014092009514570236, + 0.013761389069259167, + 0.010775091126561165, + 0.00598076730966568, + 0.008306439965963364, + 0.005427136551588774, + 0.004503106232732534, + 0.00290053803473711, + 0.003903590142726898, + 0.008573641069233418, + 0.004323847591876984, + 0.0026239450089633465, + 0.0336507149040699, + 0.076004259288311, + 0.0072508882731199265, + 0.005815346725285053, + 0.007560202851891518, + 0.013759598135948181, + 0.021333802491426468, + 0.004537305794656277, + 0.006904050707817078, + 0.003794902004301548, + 0.010057207196950912, + 0.007543629035353661, + 0.008609671145677567, + 0.004378741141408682, + 0.009022045880556107, + 0.004206489771604538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.03424324095249176 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820111833512783 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.01820111833512783, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820111833512783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1014302521944046, + 0.0308240856975317, + 0.03806706517934799, + 0.09697570651769638, + 0.15615881979465485, + 0.1003568097949028, + 0.0994887426495552, + 0.07443356513977051, + 0.17540030181407928, + 0.12686464190483093 + ], + "credit_map": [ + 0.007214063312858343, + 0.009253423660993576, + 0.013347397558391094, + 0.024219935759902, + 0.01451780553907156, + 0.009826086461544037, + 0.012132401578128338, + 0.010919141583144665, + 0.0035106902942061424, + 0.003743655513972044, + 0.004202605690807104, + 0.004824023228138685, + 0.0033359385561197996, + 0.004953125491738319, + 0.002809221623465419, + 0.0034448252990841866, + 0.004488295875489712, + 0.0045234160497784615, + 0.004009212367236614, + 0.004639973398298025, + 0.005425974726676941, + 0.005862985271960497, + 0.0061230529099702835, + 0.002994153881445527, + 0.009740365669131279, + 0.012218156829476357, + 0.010250954888761044, + 0.012815291993319988, + 0.01786688342690468, + 0.010414288379251957, + 0.013024156913161278, + 0.010645602829754353, + 0.051846180111169815, + 0.008840985596179962, + 0.0071340627036988735, + 0.025383464992046356, + 0.00760888634249568, + 0.036449071019887924, + 0.012305282056331635, + 0.006590885575860739, + 0.009181302972137928, + 0.01608596183359623, + 0.007586962543427944, + 0.013181542977690697, + 0.012628159485757351, + 0.02354646660387516, + 0.011741053313016891, + 0.006405367515981197, + 0.012817568145692348, + 0.018412508070468903, + 0.01094973087310791, + 0.009506896138191223, + 0.010988465510308743, + 0.016266942024230957, + 0.01317777018994093, + 0.007368863094598055, + 0.011586195789277554, + 0.008113705553114414, + 0.00880763866007328, + 0.007543809711933136, + 0.011888944543898106, + 0.010095189325511456, + 0.009407094679772854, + 0.006990989204496145, + 0.03272981941699982, + 0.0686795562505722, + 0.011195790953934193, + 0.00710879173129797, + 0.012003343552350998, + 0.013761688023805618, + 0.021224578842520714, + 0.008696746081113815, + 0.011050219647586346, + 0.018444331362843513, + 0.012898189015686512, + 0.010865466669201851, + 0.025671491399407387, + 0.022587593644857407, + 0.016941377893090248, + 0.008405971340835094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.01820111833512783 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022031813859939575 + ], + "opposite_logprob_tokens": [ + -6.127202987670898 + ], + "logprob_reward": -0.0022031813859939575, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022031813859939575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18364964425563812, + 0.06702540069818497, + 0.08869766443967819, + 0.10798502713441849, + 0.2764113247394562, + 0.0851045697927475, + 0.06668649613857269, + 0.050240978598594666, + 0.03378896787762642, + 0.04040990769863129 + ], + "credit_map": [ + 0.02620661072432995, + 0.023220347240567207, + 0.023157672956585884, + 0.02833147719502449, + 0.021615080535411835, + 0.02149556204676628, + 0.019994422793388367, + 0.019628459587693214, + 0.00797539483755827, + 0.00682050921022892, + 0.008300491608679295, + 0.008631954900920391, + 0.013984734192490578, + 0.00732742715626955, + 0.008485807105898857, + 0.005499076098203659, + 0.012206630781292915, + 0.011319582350552082, + 0.011177022941410542, + 0.010887068696320057, + 0.010939166881144047, + 0.013357343152165413, + 0.008927370421588421, + 0.00988348014652729, + 0.012783195823431015, + 0.013072104193270206, + 0.013802128843963146, + 0.01287827081978321, + 0.014795242808759212, + 0.012881258502602577, + 0.017040520906448364, + 0.010732308961451054, + 0.08941967040300369, + 0.01561190839856863, + 0.01315957959741354, + 0.017080215737223625, + 0.010671346448361874, + 0.10654523223638535, + 0.015661705285310745, + 0.00826167594641447, + 0.014068957418203354, + 0.007291883230209351, + 0.007905377075076103, + 0.011507770977914333, + 0.010060694999992847, + 0.014120267704129219, + 0.009484993293881416, + 0.010664628818631172, + 0.00809858925640583, + 0.007692110259085894, + 0.008128042332828045, + 0.006319626700133085, + 0.011944413185119629, + 0.009097976610064507, + 0.008835060521960258, + 0.006570680998265743, + 0.006155091803520918, + 0.005290555767714977, + 0.0077560218051075935, + 0.00501624308526516, + 0.008195986971259117, + 0.006422101054340601, + 0.005827224347740412, + 0.0055777509696781635, + 0.003896541427820921, + 0.010049127042293549, + 0.003472353797405958, + 0.003428841009736061, + 0.0027500540018081665, + 0.004127981141209602, + 0.004254407715052366, + 0.0018096612766385078, + 0.005215451121330261, + 0.007364848628640175, + 0.004311140161007643, + 0.004110423382371664, + 0.005632858723402023, + 0.005232393275946379, + 0.004835946951061487, + 0.003706845687702298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0022031813859939575 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196828007698059 + ], + "opposite_logprob_tokens": [ + -4.264196872711182 + ], + "logprob_reward": -0.014196828007698059, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196828007698059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07930799573659897, + 0.06165437400341034, + 0.07327506691217422, + 0.04749145731329918, + 0.11931177973747253, + 0.12231409549713135, + 0.09304367750883102, + 0.1175672858953476, + 0.20829571783542633, + 0.07773852348327637 + ], + "credit_map": [ + 0.008245209231972694, + 0.01007726602256298, + 0.010540899820625782, + 0.007927271537482738, + 0.016093282029032707, + 0.010456139221787453, + 0.009131662547588348, + 0.006836270447820425, + 0.006564261391758919, + 0.006198090501129627, + 0.007154654711484909, + 0.009179276414215565, + 0.014717178419232368, + 0.007078897207975388, + 0.005278360564261675, + 0.0054836515337228775, + 0.010102842003107071, + 0.00883491337299347, + 0.005921483971178532, + 0.009740184061229229, + 0.009240747429430485, + 0.014295080676674843, + 0.009603627026081085, + 0.005536191631108522, + 0.0044732969254255295, + 0.005710644647479057, + 0.004501514136791229, + 0.008342894725501537, + 0.00594041682779789, + 0.006529022008180618, + 0.006039249245077372, + 0.005954419728368521, + 0.013049924746155739, + 0.009094348177313805, + 0.00844761822372675, + 0.027362920343875885, + 0.009855140000581741, + 0.03123490884900093, + 0.011740187183022499, + 0.008526739664375782, + 0.014930796809494495, + 0.011101788841187954, + 0.011088202707469463, + 0.021432384848594666, + 0.018662409856915474, + 0.012405975721776485, + 0.018516751006245613, + 0.0141757782548666, + 0.011400077491998672, + 0.011620208621025085, + 0.009123513475060463, + 0.009575868956744671, + 0.015756620094180107, + 0.01566949114203453, + 0.012376608327031136, + 0.007521290332078934, + 0.02413427084684372, + 0.011697603389620781, + 0.014378278516232967, + 0.01094811875373125, + 0.008220739662647247, + 0.02504948526620865, + 0.010890688747167587, + 0.01224810816347599, + 0.03718694671988487, + 0.0870765894651413, + 0.012546198442578316, + 0.007697528228163719, + 0.012065569870173931, + 0.017048534005880356, + 0.026693614199757576, + 0.007980729453265667, + 0.011560669168829918, + 0.005896307993680239, + 0.008268680423498154, + 0.01437491923570633, + 0.011503097601234913, + 0.006393429823219776, + 0.013276563957333565, + 0.0064648594707250595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.014196828007698059 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a chubby groundhog with shaggy, coarse brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the groundhog emerge from a dark earthen burrow in a vibrant emerald meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the groundhog push its hind legs to pull its entire body out of the burrow and rest its forepaws on the lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small white and yellow wildflowers visible dotting the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304467514157295 + ], + "opposite_logprob_tokens": [ + -3.1680445671081543 + ], + "logprob_reward": -0.04304467514157295, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304467514157295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37073254585266113, + 0.0471821129322052, + 0.07249652594327927, + 0.08276835083961487, + 0.08042646199464798, + 0.10088597983121872, + 0.05292733386158943, + 0.1041746586561203, + 0.06434071063995361, + 0.024065380915999413 + ], + "credit_map": [ + 0.05534876137971878, + 0.048131201416254044, + 0.03841692954301834, + 0.039962925016880035, + 0.06611796468496323, + 0.04737637937068939, + 0.03679349273443222, + 0.03858485817909241, + 0.007743179798126221, + 0.005481610540300608, + 0.006041582208126783, + 0.007071531377732754, + 0.004953993950039148, + 0.005707504227757454, + 0.006169517990201712, + 0.0040131863206624985, + 0.013820838183164597, + 0.008513709530234337, + 0.010177665390074253, + 0.007254600524902344, + 0.006448589731007814, + 0.01434379257261753, + 0.005129006691277027, + 0.0068083214573562145, + 0.011231450363993645, + 0.008249560371041298, + 0.014081304892897606, + 0.009787626564502716, + 0.00970167014747858, + 0.008162138983607292, + 0.011876238510012627, + 0.009678349830210209, + 0.008315651677548885, + 0.012386162765324116, + 0.005510480143129826, + 0.010269200429320335, + 0.005379308946430683, + 0.028225697576999664, + 0.005334288813173771, + 0.005005663726478815, + 0.016984783113002777, + 0.010263027623295784, + 0.009432082064449787, + 0.01244703121483326, + 0.010276872664690018, + 0.01925354078412056, + 0.010743877850472927, + 0.011484754271805286, + 0.007217018865048885, + 0.010416441597044468, + 0.005768453236669302, + 0.005118636880069971, + 0.006259738001972437, + 0.008349542506039143, + 0.0062217796221375465, + 0.003575716633349657, + 0.01538397092372179, + 0.012780056335031986, + 0.013498335145413876, + 0.009779155254364014, + 0.01233185175806284, + 0.016405543312430382, + 0.011660629883408546, + 0.012335102073848248, + 0.013531574979424477, + 0.025782939046621323, + 0.0044815149158239365, + 0.00266709690913558, + 0.0032810878474265337, + 0.005028084851801395, + 0.00698636332526803, + 0.0025820371229201555, + 0.0029020507354289293, + 0.0024830494076013565, + 0.0033419295214116573, + 0.0031821923330426216, + 0.0031182917300611734, + 0.0031051470432430506, + 0.004024735651910305, + 0.001907982979901135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.04304467514157295 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_15/index_tensor([2779, 2779])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_15/index_tensor([2779, 2779])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..14165c0c0470cc1b46ca482b8c9378f42368a0ad --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_15/index_tensor([2779, 2779])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, wide shot gazes from afar at a surreal scene suspended high above a cyberpunk city. Below, a swirling sea of electronic clouds, like a thick fog composed of low-resolution pixels, slowly churns, emitting a faint internal glow, blurring the city's skyscrapers and flickering holographic billboards into a fluid shimmer of light and shadow. A hanging garden of red-tongued bamboo floats above this digital mist, its base atop a mossy biomechanical platform, faintly visible beneath the gleaming tubes and circuitry. The bamboo itself, with its slender stems, possesses a deep, almost black sheen, while each leaf, seemingly meticulously polished, its dark surface reflecting the ever-shifting magenta and cyan light projected from the city's giant holographic billboards below. The light shimmers across the leaves in a shifting pattern. As the electronic sea of clouds slowly rolls, this quiet bamboo forest is sometimes swallowed up by the thick digital fog, becoming looming, and sometimes fully revealed. Against the backdrop of flashing neon lights and endless skyscrapers, it forms a futuristic, Zen-like, and breathtakingly tranquil picture.", + "seed": 26818, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_15/index_2779_seed_26818_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021588368341326714, + "vlm_logprob": -0.021588368341326714, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.021588368341326714, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007656986825168133 + ], + "opposite_logprob_tokens": [ + -5.132657051086426 + ], + "logprob_reward": -0.007656986825168133, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.007656986825168133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10423128306865692, + 0.08032462745904922, + 0.03990892320871353, + 0.040159400552511215, + 0.12365961819887161, + 0.27427706122398376, + 0.059628602117300034, + 0.17920030653476715, + 0.0710165873169899, + 0.02759357914328575 + ], + "credit_map": [ + 0.013197755441069603, + 0.01121354941278696, + 0.01337064802646637, + 0.011197006329894066, + 0.012170923873782158, + 0.013900995254516602, + 0.011220644228160381, + 0.017959756776690483, + 0.009205027483403683, + 0.0071964445523917675, + 0.01202996913343668, + 0.014932786114513874, + 0.0065422882325947285, + 0.007630548905581236, + 0.014933208003640175, + 0.007854357361793518, + 0.0033952794037759304, + 0.0026946766301989555, + 0.009896337054669857, + 0.009783295914530754, + 0.002115610521286726, + 0.002529511461034417, + 0.005402084439992905, + 0.00409212801605463, + 0.003891209838911891, + 0.00402879947796464, + 0.004927354399114847, + 0.007032753434032202, + 0.004521180875599384, + 0.003692005295306444, + 0.006272723898291588, + 0.005793374963104725, + 0.006685906555503607, + 0.006394138094037771, + 0.03158709779381752, + 0.04388995096087456, + 0.0054859560914337635, + 0.0036712412256747484, + 0.0172557532787323, + 0.008689573034644127, + 0.02163742482662201, + 0.012245235964655876, + 0.02663157321512699, + 0.030588580295443535, + 0.018251175060868263, + 0.009418243542313576, + 0.10413200408220291, + 0.051372818648815155, + 0.0052411966025829315, + 0.006057319231331348, + 0.00888479221612215, + 0.00577431358397007, + 0.012730198912322521, + 0.0064171613194048405, + 0.006288929842412472, + 0.008234692737460136, + 0.011011029593646526, + 0.009051880799233913, + 0.041821423918008804, + 0.036057524383068085, + 0.008197753690183163, + 0.00802676659077406, + 0.03472767397761345, + 0.030306242406368256, + 0.0063628279604017735, + 0.009353973902761936, + 0.012238781899213791, + 0.010884233750402927, + 0.006462134886533022, + 0.010988492518663406, + 0.008508418686687946, + 0.006217724643647671, + 0.0030582041945308447, + 0.0036085741594433784, + 0.0030758054926991463, + 0.004592792131006718, + 0.003784780390560627, + 0.0031027556397020817, + 0.003515847958624363, + 0.0028548201080411673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.007656986825168133 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207758009433746 + ], + "opposite_logprob_tokens": [ + -4.26420783996582 + ], + "logprob_reward": -0.014207758009433746, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207758009433746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09669268876314163, + 0.03291451930999756, + 0.08272489905357361, + 0.0426056906580925, + 0.16641966998577118, + 0.1339341402053833, + 0.09067372977733612, + 0.2616152763366699, + 0.0649612620472908, + 0.02745809778571129 + ], + "credit_map": [ + 0.011236906051635742, + 0.011819574050605297, + 0.014518612064421177, + 0.01753549836575985, + 0.010967953130602837, + 0.009764897637069225, + 0.01204367820173502, + 0.008805569261312485, + 0.003921791445463896, + 0.004005336668342352, + 0.005792046431452036, + 0.004884886555373669, + 0.0038948296569287777, + 0.0029661846347153187, + 0.004075672011822462, + 0.0033737733028829098, + 0.010661188513040543, + 0.009930583648383617, + 0.011885487474501133, + 0.014941795729100704, + 0.006477412302047014, + 0.009661738760769367, + 0.010583551600575447, + 0.00858313962817192, + 0.0028166421689093113, + 0.004300442058593035, + 0.006126751191914082, + 0.008721837773919106, + 0.0036679471377283335, + 0.0032829835545271635, + 0.0071690683253109455, + 0.006520019844174385, + 0.00811364408582449, + 0.008085018955171108, + 0.05310102179646492, + 0.061797551810741425, + 0.006800593342632055, + 0.004239246249198914, + 0.015858598053455353, + 0.008424009196460247, + 0.01262570358812809, + 0.0067152902483940125, + 0.009208539500832558, + 0.015306474640965462, + 0.008838375099003315, + 0.005514891818165779, + 0.048547811806201935, + 0.027177050709724426, + 0.00810515508055687, + 0.0069849309511482716, + 0.016907084733247757, + 0.01105712540447712, + 0.009898943826556206, + 0.007357441354542971, + 0.012941693887114525, + 0.017421359196305275, + 0.021286776289343834, + 0.01331287994980812, + 0.04521188884973526, + 0.06509244441986084, + 0.011392075568437576, + 0.014482577331364155, + 0.029517950490117073, + 0.061318691819906235, + 0.007306207437068224, + 0.008704555220901966, + 0.011033066548407078, + 0.010274708271026611, + 0.005063488148152828, + 0.01050511933863163, + 0.00644370261579752, + 0.0056304181925952435, + 0.0021903172601014376, + 0.003503690240904689, + 0.004495840985327959, + 0.004464182537049055, + 0.0038409652188420296, + 0.002416949486359954, + 0.0038921746890991926, + 0.0026539787650108337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.014207758009433746 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026354555040597916 + ], + "opposite_logprob_tokens": [ + -3.6513545513153076 + ], + "logprob_reward": -0.026354555040597916, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026354555040597916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07736115157604218, + 0.015368526801466942, + 0.032571613788604736, + 0.07272037863731384, + 0.4501933753490448, + 0.08345487713813782, + 0.057813894003629684, + 0.14786185324192047, + 0.0411447174847126, + 0.021509641781449318 + ], + "credit_map": [ + 0.015038925223052502, + 0.007741025183349848, + 0.010068252682685852, + 0.00818714126944542, + 0.009080909192562103, + 0.009751577861607075, + 0.007412431761622429, + 0.01008088979870081, + 0.0017998686525970697, + 0.0017783355433493853, + 0.0025430130772292614, + 0.002296881517395377, + 0.0017531089251860976, + 0.001903202966786921, + 0.0018805963918566704, + 0.0014135183300822973, + 0.006051885429769754, + 0.0038782956544309855, + 0.0046137417666614056, + 0.0040804981254041195, + 0.0025615880731493235, + 0.003394060768187046, + 0.00330609199590981, + 0.004685455933213234, + 0.008723810315132141, + 0.010238048620522022, + 0.007733566220849752, + 0.008999896235764027, + 0.008660007268190384, + 0.009795336052775383, + 0.007713998202234507, + 0.010855712927877903, + 0.023827478289604187, + 0.019674908369779587, + 0.13711632788181305, + 0.16323693096637726, + 0.01902579702436924, + 0.010499696247279644, + 0.0511479377746582, + 0.025664299726486206, + 0.013213247992098331, + 0.005426640622317791, + 0.014072781428694725, + 0.010817478410899639, + 0.004575153347104788, + 0.005480500403791666, + 0.015912752598524094, + 0.013956326059997082, + 0.006463709753006697, + 0.004861658904701471, + 0.0067815640941262245, + 0.00457742577418685, + 0.010493073612451553, + 0.004824044648557901, + 0.006269525270909071, + 0.013542886823415756, + 0.006082956213504076, + 0.0085524907335639, + 0.021673712879419327, + 0.060313574969768524, + 0.006491048261523247, + 0.005242728162556887, + 0.032162535935640335, + 0.007342814933508635, + 0.00428896676748991, + 0.0056555126793682575, + 0.007191662676632404, + 0.006220848765224218, + 0.003210371593013406, + 0.006666453089565039, + 0.00455790339037776, + 0.003353000385686755, + 0.0017767056124284863, + 0.003217830788344145, + 0.004063156899064779, + 0.003508305875584483, + 0.0022590928710997105, + 0.001869388041086495, + 0.0023950638715177774, + 0.002420097589492798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.026354555040597916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486522801220417 + ], + "opposite_logprob_tokens": [ + -3.048652172088623 + ], + "logprob_reward": -0.0486522801220417, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486522801220417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06798555701971054, + 0.029411837458610535, + 0.043073952198028564, + 0.04862726479768753, + 0.1275966465473175, + 0.1433352828025818, + 0.058183424174785614, + 0.38606494665145874, + 0.034693095833063126, + 0.06102795526385307 + ], + "credit_map": [ + 0.006074823439121246, + 0.011747822165489197, + 0.007581649348139763, + 0.006370726507157087, + 0.009624429978430271, + 0.007552583236247301, + 0.0084132831543684, + 0.010620244778692722, + 0.003896743291988969, + 0.003470187308266759, + 0.003697157371789217, + 0.004563094582408667, + 0.003540260251611471, + 0.0027343868277966976, + 0.0047452799044549465, + 0.0027647309470921755, + 0.0037787628825753927, + 0.0035405498929321766, + 0.01184875052422285, + 0.010167191736400127, + 0.00354590336792171, + 0.0027238193433731794, + 0.003839541459456086, + 0.0036294374149292707, + 0.004208022728562355, + 0.006470168009400368, + 0.006806494668126106, + 0.00648646941408515, + 0.004799715708941221, + 0.003995141480118036, + 0.007308720611035824, + 0.008552532643079758, + 0.006283673923462629, + 0.009433436207473278, + 0.03335095942020416, + 0.03946070373058319, + 0.013652043417096138, + 0.005047485698014498, + 0.010301326401531696, + 0.010067031718790531, + 0.021228346973657608, + 0.007727055344730616, + 0.03985625132918358, + 0.014096719212830067, + 0.007656760048121214, + 0.007371936459094286, + 0.03193291276693344, + 0.01346530020236969, + 0.005708266515284777, + 0.004625819157809019, + 0.007844353094696999, + 0.006104519125074148, + 0.005834321491420269, + 0.00446636276319623, + 0.00777528528124094, + 0.015824507921934128, + 0.014887194149196148, + 0.01752077415585518, + 0.1068197712302208, + 0.14315831661224365, + 0.011720551177859306, + 0.01151199359446764, + 0.05605624243617058, + 0.02439013123512268, + 0.003678281558677554, + 0.00454347487539053, + 0.0055128163658082485, + 0.004036588128656149, + 0.004160931799560785, + 0.006307027768343687, + 0.0035958008375018835, + 0.0028581791557371616, + 0.0049034901894629, + 0.009066056460142136, + 0.012736772187054157, + 0.009594547562301159, + 0.006170288659632206, + 0.004359122831374407, + 0.007445646449923515, + 0.006752036977559328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.0486522801220417 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107026170939207 + ], + "opposite_logprob_tokens": [ + -4.511070251464844 + ], + "logprob_reward": -0.01107026170939207, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107026170939207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04769885912537575, + 0.02228640392422676, + 0.04337625205516815, + 0.04203580692410469, + 0.03507380187511444, + 0.05947345495223999, + 0.04273221641778946, + 0.6483754515647888, + 0.045769087970256805, + 0.013178705237805843 + ], + "credit_map": [ + 0.008330185897648335, + 0.005020868964493275, + 0.005479373503476381, + 0.005811084993183613, + 0.005611953791230917, + 0.005846059415489435, + 0.0061911060474812984, + 0.00540822884067893, + 0.0024597058072686195, + 0.002642595674842596, + 0.0034714052453637123, + 0.003112985985353589, + 0.002524638082832098, + 0.002455521607771516, + 0.0030554616823792458, + 0.0025640889070928097, + 0.003986249677836895, + 0.003916991408914328, + 0.004909810144454241, + 0.005872661713510752, + 0.0031292468775063753, + 0.004166636150330305, + 0.006342344917356968, + 0.011052310466766357, + 0.003989462740719318, + 0.005856839474290609, + 0.004898107144981623, + 0.005444691050797701, + 0.0032041515223681927, + 0.004837122745811939, + 0.007308178581297398, + 0.006497258320450783, + 0.0020238624420017004, + 0.0018619922921061516, + 0.010444492101669312, + 0.012044914066791534, + 0.001654757303185761, + 0.0012531690299510956, + 0.00365639291703701, + 0.0021342209074646235, + 0.004990307614207268, + 0.002763697411864996, + 0.003830673173069954, + 0.007193081080913544, + 0.003321816911920905, + 0.002507945988327265, + 0.020776402205228806, + 0.01408953033387661, + 0.004861684050410986, + 0.0039110020734369755, + 0.005381020251661539, + 0.003744151210412383, + 0.008215410634875298, + 0.0033488988410681486, + 0.004633134230971336, + 0.008636916987597942, + 0.03714311867952347, + 0.03288109600543976, + 0.05490595102310181, + 0.27552884817123413, + 0.01795441284775734, + 0.02176276408135891, + 0.12162292748689651, + 0.08657632023096085, + 0.003983840811997652, + 0.005952983163297176, + 0.0075330352410674095, + 0.0070783235132694244, + 0.004902652930468321, + 0.007800535764545202, + 0.004937915597110987, + 0.0035798028111457825, + 0.0015116073191165924, + 0.002004910260438919, + 0.002053344389423728, + 0.0022296179085969925, + 0.001643533119931817, + 0.0010863854549825191, + 0.0012526334030553699, + 0.0013966728001832962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.01107026170939207 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, wide shot gazes from afar at a surreal scene suspended high above a cyberpunk city. Below, a swirling sea of electronic clouds, like a thick fog composed of low-resolution pixels, slowly churns, emitting a faint internal glow, blurring the city's skyscrapers and flickering holographic billboards into a fluid shimmer of light and shadow. A hanging garden of red-tongued bamboo floats above this digital mist, its base atop a mossy biomechanical platform, faintly visible beneath the gleaming tubes and circuitry. The bamboo itself, with its slender stems, possesses a deep, almost black sheen, while each leaf, seemingly meticulously polished, its dark surface reflecting the ever-shifting magenta and cyan light projected from the city's giant holographic billboards below. The light shimmers across the leaves in a shifting pattern. As the electronic sea of clouds slowly rolls, this quiet bamboo forest is sometimes swallowed up by the thick digital fog, becoming looming, and sometimes fully revealed. Against the backdrop of flashing neon lights and endless skyscrapers, it forms a futuristic, Zen-like, and breathtakingly tranquil picture.", + "seed": 26819, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_15/index_2779_seed_26819_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01868165675550699, + "vlm_logprob": -0.01868165675550699, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.01868165675550699, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012107805348932743 + ], + "opposite_logprob_tokens": [ + -4.512107849121094 + ], + "logprob_reward": -0.012107805348932743, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.012107805348932743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18198980391025543, + 0.06617410480976105, + 0.029401781037449837, + 0.02980954572558403, + 0.03337356448173523, + 0.040452830493450165, + 0.48496997356414795, + 0.03745456039905548, + 0.020062865689396858, + 0.07631096243858337 + ], + "credit_map": [ + 0.02614155039191246, + 0.01809198409318924, + 0.02438344806432724, + 0.04734169691801071, + 0.016346853226423264, + 0.015021286904811859, + 0.01655992865562439, + 0.01810305565595627, + 0.006285385228693485, + 0.008405153639614582, + 0.007149530574679375, + 0.008817007765173912, + 0.009148339740931988, + 0.006778133567422628, + 0.012554324232041836, + 0.007036229129880667, + 0.002833674196153879, + 0.004012543708086014, + 0.0031087726820260286, + 0.0033510804641991854, + 0.003572714515030384, + 0.003620622679591179, + 0.0037982023786753416, + 0.005104167852550745, + 0.0030164653435349464, + 0.004030702635645866, + 0.005393483676016331, + 0.0033218050375580788, + 0.003970538266003132, + 0.0031157764606177807, + 0.0024789138697087765, + 0.004481861367821693, + 0.002363614970818162, + 0.003679974004626274, + 0.006125304847955704, + 0.003963508177548647, + 0.002972302958369255, + 0.0031994024757295847, + 0.004775104578584433, + 0.006294353399425745, + 0.0030162709299474955, + 0.00373796746134758, + 0.007670999970287085, + 0.0056106457486748695, + 0.004913860000669956, + 0.003222969826310873, + 0.004023377317935228, + 0.008256737142801285, + 0.025541607290506363, + 0.025486456230282784, + 0.09076330810785294, + 0.10733180493116379, + 0.018272927030920982, + 0.019245367497205734, + 0.0641624853014946, + 0.13416606187820435, + 0.004499876406043768, + 0.002636419842019677, + 0.004761029966175556, + 0.005955673288553953, + 0.0021127904765307903, + 0.0028645277488976717, + 0.005395089276134968, + 0.009229151532053947, + 0.0013188504381105304, + 0.002766916062682867, + 0.002217700006440282, + 0.0025694617070257664, + 0.0038347430527210236, + 0.0023112772032618523, + 0.0024640753399580717, + 0.0025798429269343615, + 0.003402447560802102, + 0.0016017386224120855, + 0.002493753097951412, + 0.004219883121550083, + 0.0025837181601673365, + 0.002808439079672098, + 0.03136491775512695, + 0.027836063876748085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.012107805348932743 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330123446881771 + ], + "opposite_logprob_tokens": [ + -3.773301124572754 + ], + "logprob_reward": -0.02330123446881771, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330123446881771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13062186539173126, + 0.1403283029794693, + 0.031589530408382416, + 0.023103682324290276, + 0.010184542275965214, + 0.010679110884666443, + 0.46660345792770386, + 0.11893843114376068, + 0.03415420278906822, + 0.03379688411951065 + ], + "credit_map": [ + 0.011686623096466064, + 0.016629904508590698, + 0.017831526696681976, + 0.03635269030928612, + 0.014403438195586205, + 0.01005692407488823, + 0.013492354191839695, + 0.010168402455747128, + 0.010356666520237923, + 0.017912691459059715, + 0.018345845863223076, + 0.019778156653046608, + 0.021956078708171844, + 0.011953787878155708, + 0.02462526224553585, + 0.015399813652038574, + 0.002618329832330346, + 0.0035958108492195606, + 0.004019628278911114, + 0.0032618793193250895, + 0.003222003113478422, + 0.004117272328585386, + 0.004056175239384174, + 0.006698427256196737, + 0.0022385555785149336, + 0.003216722048819065, + 0.0026705842465162277, + 0.002243148162961006, + 0.002562507288530469, + 0.003247953951358795, + 0.0025655857753008604, + 0.004358625039458275, + 0.0014523710124194622, + 0.0011680069146677852, + 0.001166443806141615, + 0.0010341750457882881, + 0.0017197530250996351, + 0.0015269663417711854, + 0.000835841812659055, + 0.0012809830950573087, + 0.0010130670852959156, + 0.0012632720172405243, + 0.0017965588485822082, + 0.0012902694288641214, + 0.0019088593544438481, + 0.0010654478101059794, + 0.0009659301722422242, + 0.0013757051201537251, + 0.024691950529813766, + 0.025035429745912552, + 0.08728430420160294, + 0.10229787230491638, + 0.018464379012584686, + 0.018482938408851624, + 0.061679448932409286, + 0.12866713106632233, + 0.010435477830469608, + 0.009752191603183746, + 0.02131340280175209, + 0.03235546499490738, + 0.011560729704797268, + 0.006265194155275822, + 0.012754234485328197, + 0.014501707628369331, + 0.003272973233833909, + 0.004191193263977766, + 0.004325490444898605, + 0.004400460049510002, + 0.0049216956831514835, + 0.004804241470992565, + 0.004499552305787802, + 0.003738594939932227, + 0.0012053765822201967, + 0.0012206375831738114, + 0.0012457022676244378, + 0.0018139028688892722, + 0.0018706900300458074, + 0.0017475913045927882, + 0.012537251226603985, + 0.012155730277299881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.02330123446881771 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058872953057289 + ], + "opposite_logprob_tokens": [ + -3.8955886363983154 + ], + "logprob_reward": -0.02058872953057289, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058872953057289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19220323860645294, + 0.16200730204582214, + 0.061871934682130814, + 0.025444895029067993, + 0.01293193269520998, + 0.044809624552726746, + 0.4261474907398224, + 0.031927142292261124, + 0.03087339550256729, + 0.011783055029809475 + ], + "credit_map": [ + 0.022371187806129456, + 0.02348695881664753, + 0.02905185893177986, + 0.045153405517339706, + 0.018377382308244705, + 0.01880154199898243, + 0.017831509932875633, + 0.017129406332969666, + 0.015157263725996017, + 0.02405114285647869, + 0.01912478543817997, + 0.015193905681371689, + 0.03253714740276337, + 0.01532590389251709, + 0.022573400288820267, + 0.018043775111436844, + 0.004906482994556427, + 0.008032968267798424, + 0.007921356707811356, + 0.007802771404385567, + 0.006277310661971569, + 0.006294186692684889, + 0.007496611215174198, + 0.01314025279134512, + 0.0023986673913896084, + 0.0035572010092437267, + 0.0030488648917526007, + 0.00256614969111979, + 0.002760749077424407, + 0.0030513713136315346, + 0.0028257896192371845, + 0.005236103665083647, + 0.0016042640199884772, + 0.0013910550624132156, + 0.0015431165229529142, + 0.0012180197518318892, + 0.001798825105652213, + 0.0016644792631268501, + 0.0015991367399692535, + 0.0021130372770130634, + 0.004092605318874121, + 0.0038777291774749756, + 0.010080047883093357, + 0.0066987294703722, + 0.005884268786758184, + 0.0031384292524307966, + 0.004713347647339106, + 0.00632447050884366, + 0.021128224208950996, + 0.02191479690372944, + 0.08297760039567947, + 0.09661667793989182, + 0.01882912963628769, + 0.015882322564721107, + 0.0554170086979866, + 0.1133817508816719, + 0.003774180542677641, + 0.003389624645933509, + 0.004841750953346491, + 0.006347061600536108, + 0.0031372318044304848, + 0.0025753499940037727, + 0.003443719819188118, + 0.004418225958943367, + 0.0027364203706383705, + 0.004860432352870703, + 0.003702951595187187, + 0.002733085071668029, + 0.0049324496649205685, + 0.0031022096518427134, + 0.004021380562335253, + 0.004784469958394766, + 0.0009466026676818728, + 0.0015812126221135259, + 0.0010965666733682156, + 0.0012836019741371274, + 0.0017060496611520648, + 0.0015356334624812007, + 0.0019948198460042477, + 0.0016385696362704039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.02058872953057289 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026341667398810387 + ], + "opposite_logprob_tokens": [ + -3.651341676712036 + ], + "logprob_reward": -0.026341667398810387, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026341667398810387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21529705822467804, + 0.04133251681923866, + 0.04576338082551956, + 0.014307236298918724, + 0.015768680721521378, + 0.042675476521253586, + 0.34903019666671753, + 0.035068437457084656, + 0.0280297864228487, + 0.21272723376750946 + ], + "credit_map": [ + 0.025195710361003876, + 0.020425621420145035, + 0.03205261752009392, + 0.057393498718738556, + 0.018353838473558426, + 0.016767458990216255, + 0.02114679478108883, + 0.023961516097187996, + 0.006452473346143961, + 0.006098468787968159, + 0.005330999847501516, + 0.0034938582684844732, + 0.005981918890029192, + 0.004847878124564886, + 0.0031432921532541513, + 0.005983623210340738, + 0.004332990385591984, + 0.0066341920755803585, + 0.005162728019058704, + 0.005545998457819223, + 0.007191319949924946, + 0.005130802746862173, + 0.005294069182127714, + 0.006471280008554459, + 0.0013831782853230834, + 0.0021289943251758814, + 0.0020705319475382566, + 0.001290819956921041, + 0.0023089393507689238, + 0.0019087936962023377, + 0.0012611020356416702, + 0.001954877283424139, + 0.0017938838573172688, + 0.0021570760291069746, + 0.0019370733061805367, + 0.0012103597400709987, + 0.003710294608026743, + 0.0019625292625278234, + 0.001248542801477015, + 0.00174892065115273, + 0.004546528682112694, + 0.004542948212474585, + 0.005959557369351387, + 0.004252305254340172, + 0.004468279425054789, + 0.004383534658700228, + 0.005846298765391111, + 0.008676024153828621, + 0.016382532194256783, + 0.01754656247794628, + 0.06861100345849991, + 0.07808338850736618, + 0.013739725574851036, + 0.013233271427452564, + 0.045813873410224915, + 0.09561984241008759, + 0.004021561238914728, + 0.0034332871437072754, + 0.005464722868055105, + 0.007105962373316288, + 0.00314742187038064, + 0.0031541138887405396, + 0.004042140673846006, + 0.004699228797107935, + 0.004800904076546431, + 0.0040094926953315735, + 0.002759235678240657, + 0.0022029189858585596, + 0.005323474761098623, + 0.002639230340719223, + 0.0030544823966920376, + 0.0032400458585470915, + 0.006365817505866289, + 0.005190420895814896, + 0.007636683061718941, + 0.012083660811185837, + 0.006296026054769754, + 0.007695748005062342, + 0.08876024186611176, + 0.07869862765073776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.026341667398810387 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01106884703040123 + ], + "opposite_logprob_tokens": [ + -4.511068820953369 + ], + "logprob_reward": -0.01106884703040123, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01106884703040123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1608605533838272, + 0.019802752882242203, + 0.029190603643655777, + 0.01860576681792736, + 0.010477760806679726, + 0.026931870728731155, + 0.11218627542257309, + 0.13001832365989685, + 0.03474831581115723, + 0.4571778178215027 + ], + "credit_map": [ + 0.015069368295371532, + 0.016126977279782295, + 0.025416485965251923, + 0.0506562739610672, + 0.011496871709823608, + 0.012414613738656044, + 0.014895135536789894, + 0.014784814789891243, + 0.0033399632666260004, + 0.0027161978650838137, + 0.002153802430257201, + 0.002274660859256983, + 0.001996471779420972, + 0.002943824278190732, + 0.0018776506185531616, + 0.0025001794565469027, + 0.0034243904519826174, + 0.003703109687194228, + 0.003329205559566617, + 0.0034484961070120335, + 0.003473212942481041, + 0.0038531918544322252, + 0.003576993942260742, + 0.004381999839097261, + 0.0023391670547425747, + 0.002102185972034931, + 0.0025713967625051737, + 0.002522122347727418, + 0.0025209214072674513, + 0.0021839053370058537, + 0.0019801415037363768, + 0.0023859248030930758, + 0.0011743024224415421, + 0.0015056757256388664, + 0.0011862068204209208, + 0.0010962928645312786, + 0.0014404332032427192, + 0.0012077940627932549, + 0.0011397944763302803, + 0.0017272599507123232, + 0.001946144038811326, + 0.0019538335036486387, + 0.005973105784505606, + 0.004771186970174313, + 0.002064027823507786, + 0.0018436918035149574, + 0.004149856045842171, + 0.004230022430419922, + 0.006212923210114241, + 0.006139621138572693, + 0.02066143974661827, + 0.025125665590167046, + 0.004655541852116585, + 0.004639146383851767, + 0.014803510159254074, + 0.029948420822620392, + 0.01069773267954588, + 0.01531166024506092, + 0.024342704564332962, + 0.02724085934460163, + 0.015681400895118713, + 0.006819285452365875, + 0.013623042032122612, + 0.01630162075161934, + 0.003246833337470889, + 0.0046740928664803505, + 0.004989094100892544, + 0.003413568018004298, + 0.005903700832277536, + 0.0028044753707945347, + 0.0037121146451681852, + 0.006004433613270521, + 0.013800963759422302, + 0.010969324968755245, + 0.015564335510134697, + 0.025090008974075317, + 0.013271315023303032, + 0.017308630049228668, + 0.19051869213581085, + 0.17065446078777313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.01106884703040123 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, wide shot gazes from afar at a surreal scene suspended high above a cyberpunk city. Below, a swirling sea of electronic clouds, like a thick fog composed of low-resolution pixels, slowly churns, emitting a faint internal glow, blurring the city's skyscrapers and flickering holographic billboards into a fluid shimmer of light and shadow. A hanging garden of red-tongued bamboo floats above this digital mist, its base atop a mossy biomechanical platform, faintly visible beneath the gleaming tubes and circuitry. The bamboo itself, with its slender stems, possesses a deep, almost black sheen, while each leaf, seemingly meticulously polished, its dark surface reflecting the ever-shifting magenta and cyan light projected from the city's giant holographic billboards below. The light shimmers across the leaves in a shifting pattern. As the electronic sea of clouds slowly rolls, this quiet bamboo forest is sometimes swallowed up by the thick digital fog, becoming looming, and sometimes fully revealed. Against the backdrop of flashing neon lights and endless skyscrapers, it forms a futuristic, Zen-like, and breathtakingly tranquil picture.", + "seed": 26820, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_15/index_2779_seed_26820_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.023414222151041032, + "vlm_logprob": -0.023414222151041032, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.023414222151041032, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01534618716686964 + ], + "opposite_logprob_tokens": [ + -4.390346050262451 + ], + "logprob_reward": -0.01534618716686964, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01534618716686964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19890668988227844, + 0.07392442226409912, + 0.13018448650836945, + 0.18950608372688293, + 0.06809467077255249, + 0.06883308291435242, + 0.06700874865055084, + 0.03311452642083168, + 0.11893261969089508, + 0.05149468407034874 + ], + "credit_map": [ + 0.022855108603835106, + 0.026167377829551697, + 0.02785845845937729, + 0.021494878455996513, + 0.02517428807914257, + 0.01954852044582367, + 0.03151153400540352, + 0.02429652400314808, + 0.005763206165283918, + 0.010237128473818302, + 0.010927687399089336, + 0.012939411215484142, + 0.008601337671279907, + 0.0070121497847139835, + 0.008577962405979633, + 0.009865538217127323, + 0.008531635627150536, + 0.012256158515810966, + 0.019838333129882812, + 0.01843903586268425, + 0.015060038305819035, + 0.007623582612723112, + 0.02353072166442871, + 0.02490498498082161, + 0.024385850876569748, + 0.022599009796977043, + 0.022225474938750267, + 0.019127044826745987, + 0.026966938748955727, + 0.02003372460603714, + 0.01840304583311081, + 0.03576499968767166, + 0.0064405822195112705, + 0.006900702603161335, + 0.010651792399585247, + 0.008710284717381, + 0.006417336408048868, + 0.006447228603065014, + 0.012340245768427849, + 0.010186494328081608, + 0.011009898036718369, + 0.0119650699198246, + 0.008738065138459206, + 0.007143011782318354, + 0.008585742674767971, + 0.006517371628433466, + 0.0066219051368534565, + 0.008252015337347984, + 0.007689413148909807, + 0.009260451421141624, + 0.010498624294996262, + 0.008483907207846642, + 0.010307211428880692, + 0.00582375330850482, + 0.008080665022134781, + 0.006864720489829779, + 0.0031812230590730906, + 0.003364038420841098, + 0.004265376832336187, + 0.0032769585959613323, + 0.005565316416323185, + 0.004998371936380863, + 0.0035700686275959015, + 0.004893170669674873, + 0.014437577687203884, + 0.017151493579149246, + 0.024540916085243225, + 0.011149310506880283, + 0.010447571985423565, + 0.015181783586740494, + 0.009551383554935455, + 0.016472583636641502, + 0.00619924021884799, + 0.009322498925030231, + 0.009004714898765087, + 0.005303404293954372, + 0.006156856659799814, + 0.007943256758153439, + 0.003349444828927517, + 0.0042152684181928635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.01534618716686964 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083466000854969 + ], + "opposite_logprob_tokens": [ + -4.511083602905273 + ], + "logprob_reward": -0.011083466000854969, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083466000854969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14302697777748108, + 0.12241280823945999, + 0.08442520350217819, + 0.16932299733161926, + 0.05027547851204872, + 0.06767842173576355, + 0.04236945882439613, + 0.10832072794437408, + 0.06721605360507965, + 0.14495185017585754 + ], + "credit_map": [ + 0.011744320392608643, + 0.019613096490502357, + 0.020965728908777237, + 0.026225659996271133, + 0.01682516559958458, + 0.01148273702710867, + 0.0184345543384552, + 0.017735717818140984, + 0.010009889490902424, + 0.018238039687275887, + 0.016045885160565376, + 0.018184669315814972, + 0.019644036889076233, + 0.012987341731786728, + 0.0149140739813447, + 0.012388869188725948, + 0.005644496530294418, + 0.007646680809557438, + 0.014486466534435749, + 0.01226810459047556, + 0.008637148886919022, + 0.005997614469379187, + 0.015136401169002056, + 0.01460829097777605, + 0.023087890818715096, + 0.019279351457953453, + 0.019962284713983536, + 0.01575547643005848, + 0.02562415413558483, + 0.01942913606762886, + 0.016901252791285515, + 0.029283445328474045, + 0.0044870926067233086, + 0.005853666923940182, + 0.008879676461219788, + 0.005838742479681969, + 0.005446888506412506, + 0.005081939045339823, + 0.008172564208507538, + 0.006514909211546183, + 0.006346624810248613, + 0.007936174049973488, + 0.009913422167301178, + 0.008771762251853943, + 0.00999264232814312, + 0.007325720973312855, + 0.008287574164569378, + 0.009104500524699688, + 0.0049075717106461525, + 0.007516792044043541, + 0.005792992655187845, + 0.005935428664088249, + 0.004547033458948135, + 0.0035696220584213734, + 0.0040407064370810986, + 0.006059309467673302, + 0.011625326238572598, + 0.0157389547675848, + 0.01370896864682436, + 0.012768897227942944, + 0.015615730546414852, + 0.013128173537552357, + 0.012774760834872723, + 0.012959917075932026, + 0.008482634089887142, + 0.009956671856343746, + 0.012232332490384579, + 0.008241578936576843, + 0.006222382187843323, + 0.006212891545146704, + 0.006846053525805473, + 0.009021513164043427, + 0.011491783894598484, + 0.017529072239995003, + 0.025052540004253387, + 0.016148516908288002, + 0.023817233741283417, + 0.02411343902349472, + 0.012463397346436977, + 0.014335878193378448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.011083466000854969 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072005033493042 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.016072005033493042, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072005033493042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2711864113807678, + 0.07393147051334381, + 0.13821350038051605, + 0.03628508374094963, + 0.053285058587789536, + 0.09721541404724121, + 0.06642824411392212, + 0.03939562663435936, + 0.04812539368867874, + 0.17593379318714142 + ], + "credit_map": [ + 0.024413075298070908, + 0.03647801652550697, + 0.042099013924598694, + 0.045421380549669266, + 0.0302729532122612, + 0.02203991264104843, + 0.040004849433898926, + 0.030457213521003723, + 0.00891746673732996, + 0.01288946345448494, + 0.0077960253693163395, + 0.007584813982248306, + 0.010409749113023281, + 0.009866436943411827, + 0.008344968780875206, + 0.008122551254928112, + 0.007721377536654472, + 0.009393179789185524, + 0.023561470210552216, + 0.023386487737298012, + 0.01587025262415409, + 0.00918313954025507, + 0.026395777240395546, + 0.0227018054574728, + 0.0036053298972547054, + 0.004834148101508617, + 0.004460030235350132, + 0.003962791990488768, + 0.005225914064794779, + 0.003610607236623764, + 0.005281745456159115, + 0.005304516293108463, + 0.005810938309878111, + 0.005976728163659573, + 0.009056147187948227, + 0.005714287981390953, + 0.006565096788108349, + 0.005424347706139088, + 0.008795663714408875, + 0.005941849201917648, + 0.009753573685884476, + 0.017810191959142685, + 0.012891213409602642, + 0.010986502282321453, + 0.012146256864070892, + 0.007535826414823532, + 0.013481208123266697, + 0.012610637582838535, + 0.00918550044298172, + 0.014893907122313976, + 0.009316054172813892, + 0.007535854354500771, + 0.005963763687759638, + 0.004223256837576628, + 0.007249587681144476, + 0.008060324005782604, + 0.00700110150501132, + 0.0036704218946397305, + 0.007267689798027277, + 0.005242160055786371, + 0.00430333474650979, + 0.003683623392134905, + 0.0034913821145892143, + 0.004735911730676889, + 0.007489312440156937, + 0.007862886413931847, + 0.0064458539709448814, + 0.005740485154092312, + 0.005006168037652969, + 0.005076884292066097, + 0.003402876900509, + 0.007100926246494055, + 0.013805425725877285, + 0.015526164323091507, + 0.052765343338251114, + 0.033786311745643616, + 0.013804691843688488, + 0.0185355506837368, + 0.01051301509141922, + 0.01719728484749794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.016072005033493042 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062037866562604904 + ], + "opposite_logprob_tokens": [ + -2.812037944793701 + ], + "logprob_reward": -0.062037866562604904, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062037866562604904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12519916892051697, + 0.19942606985569, + 0.1395735740661621, + 0.07463914901018143, + 0.05696946755051613, + 0.06533233076334, + 0.0796399712562561, + 0.048515528440475464, + 0.1136593297123909, + 0.09704536944627762 + ], + "credit_map": [ + 0.010973723605275154, + 0.012722078710794449, + 0.01740131713449955, + 0.024378662928938866, + 0.00947171077132225, + 0.008846677839756012, + 0.023989612236618996, + 0.017415380105376244, + 0.016398245468735695, + 0.029658367857336998, + 0.025711379945278168, + 0.031963903456926346, + 0.02942882850766182, + 0.021062983199954033, + 0.023252032697200775, + 0.021950330585241318, + 0.0070917620323598385, + 0.010053858160972595, + 0.025236526504158974, + 0.021078674122691154, + 0.015382726676762104, + 0.010760674253106117, + 0.025450734421610832, + 0.024518616497516632, + 0.012634345330297947, + 0.008771567605435848, + 0.010054186917841434, + 0.007793070282787085, + 0.007096356712281704, + 0.0064514391124248505, + 0.009779269807040691, + 0.012058909051120281, + 0.004704220220446587, + 0.009199731051921844, + 0.007842929102480412, + 0.006091066636145115, + 0.008155614137649536, + 0.005540496204048395, + 0.008617815561592579, + 0.00681759137660265, + 0.006296159699559212, + 0.007387756370007992, + 0.00795665755867958, + 0.007923281751573086, + 0.010678053833544254, + 0.007423741277307272, + 0.010245831683278084, + 0.007420849986374378, + 0.011378671042621136, + 0.017194852232933044, + 0.008286651223897934, + 0.007422897964715958, + 0.01157620083540678, + 0.007484705653041601, + 0.006273953709751368, + 0.010022043250501156, + 0.008291307836771011, + 0.004197984002530575, + 0.006558516528457403, + 0.004909432493150234, + 0.006660316605120897, + 0.005933478940278292, + 0.005478035658597946, + 0.006486452650278807, + 0.014087566174566746, + 0.02156505547463894, + 0.021596401929855347, + 0.012000530026853085, + 0.008617528714239597, + 0.011504752561450005, + 0.01008125115185976, + 0.014206238090991974, + 0.0077841151505708694, + 0.008900090120732784, + 0.022637423127889633, + 0.011288605630397797, + 0.01142662949860096, + 0.015842651948332787, + 0.007860006764531136, + 0.01130584068596363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.062037866562604904 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531585991382599 + ], + "opposite_logprob_tokens": [ + -4.387531757354736 + ], + "logprob_reward": -0.012531585991382599, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531585991382599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08937772363424301, + 0.03132842481136322, + 0.040182895958423615, + 0.26772844791412354, + 0.025463439524173737, + 0.043498359620571136, + 0.04078102484345436, + 0.041229430586099625, + 0.2487127035856247, + 0.17169755697250366 + ], + "credit_map": [ + 0.007308179046958685, + 0.008375195786356926, + 0.014163141138851643, + 0.015738217160105705, + 0.010854458436369896, + 0.006013756152242422, + 0.01440061442553997, + 0.012524171732366085, + 0.003589375177398324, + 0.004668912850320339, + 0.003420341294258833, + 0.003698809538036585, + 0.0045760865323245525, + 0.003564617596566677, + 0.004213136155158281, + 0.0035971486940979958, + 0.0031223795376718044, + 0.004881764762103558, + 0.006005334202200174, + 0.005571131594479084, + 0.003777923062443733, + 0.0027790197636932135, + 0.006606399081647396, + 0.007438946515321732, + 0.036405306309461594, + 0.03198131546378136, + 0.03167657554149628, + 0.024174969643354416, + 0.0401822030544281, + 0.02918456494808197, + 0.02694975584745407, + 0.04717378690838814, + 0.00267192255705595, + 0.003625893499702215, + 0.0035237236879765987, + 0.003340861527249217, + 0.0024518948048353195, + 0.0022795600816607475, + 0.004353165160864592, + 0.0032164189033210278, + 0.0053984601981937885, + 0.007538843434303999, + 0.006267836317420006, + 0.004699642304331064, + 0.004683588165789843, + 0.003818206489086151, + 0.004836217034608126, + 0.0062555656768381596, + 0.005159256514161825, + 0.00794978253543377, + 0.006517553701996803, + 0.004265082534402609, + 0.0033639271277934313, + 0.003125839401036501, + 0.004174600355327129, + 0.006224987097084522, + 0.0060696834698319435, + 0.004925242159515619, + 0.005336908157914877, + 0.005651901010423899, + 0.005312320776283741, + 0.0038342871703207493, + 0.005449969787150621, + 0.004649121314287186, + 0.022263433784246445, + 0.03243664652109146, + 0.04867887124419212, + 0.031298231333494186, + 0.02734331227838993, + 0.03743940219283104, + 0.021006138995289803, + 0.0282466821372509, + 0.012167262844741344, + 0.01466898713260889, + 0.06284747272729874, + 0.04125577211380005, + 0.007578625809401274, + 0.010238245129585266, + 0.009296916425228119, + 0.01364427525550127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.012531585991382599 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, wide shot gazes from afar at a surreal scene suspended high above a cyberpunk city. Below, a swirling sea of electronic clouds, like a thick fog composed of low-resolution pixels, slowly churns, emitting a faint internal glow, blurring the city's skyscrapers and flickering holographic billboards into a fluid shimmer of light and shadow. A hanging garden of red-tongued bamboo floats above this digital mist, its base atop a mossy biomechanical platform, faintly visible beneath the gleaming tubes and circuitry. The bamboo itself, with its slender stems, possesses a deep, almost black sheen, while each leaf, seemingly meticulously polished, its dark surface reflecting the ever-shifting magenta and cyan light projected from the city's giant holographic billboards below. The light shimmers across the leaves in a shifting pattern. As the electronic sea of clouds slowly rolls, this quiet bamboo forest is sometimes swallowed up by the thick digital fog, becoming looming, and sometimes fully revealed. Against the backdrop of flashing neon lights and endless skyscrapers, it forms a futuristic, Zen-like, and breathtakingly tranquil picture.", + "seed": 26821, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_15/index_2779_seed_26821_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.031219997070729734, + "vlm_logprob": -0.031219997070729734, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.031219997070729734, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01011236198246479 + ], + "opposite_logprob_tokens": [ + -4.885112285614014 + ], + "logprob_reward": -0.01011236198246479, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01011236198246479, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24513524770736694, + 0.11846905946731567, + 0.04750822111964226, + 0.040314216166734695, + 0.20919737219810486, + 0.04230056330561638, + 0.04906073212623596, + 0.06521163880825043, + 0.03557698056101799, + 0.1472259759902954 + ], + "credit_map": [ + 0.018082408234477043, + 0.023249654099345207, + 0.04064508527517319, + 0.039029717445373535, + 0.03477957099676132, + 0.016797363758087158, + 0.04365307092666626, + 0.028898384422063828, + 0.00376055296510458, + 0.007390018552541733, + 0.009832508862018585, + 0.044979386031627655, + 0.006532734259963036, + 0.0044251433573663235, + 0.021647896617650986, + 0.019900815561413765, + 0.0046774037182331085, + 0.007612382993102074, + 0.005732723977416754, + 0.00438764737918973, + 0.009427611716091633, + 0.007355127949267626, + 0.0038539657834917307, + 0.004461356904357672, + 0.004870534874498844, + 0.004966162610799074, + 0.004085719585418701, + 0.004742270801216364, + 0.005108639597892761, + 0.005615278147161007, + 0.0062315440736711025, + 0.004694066941738129, + 0.010153418406844139, + 0.004833209794014692, + 0.07735420763492584, + 0.010942809283733368, + 0.007243304047733545, + 0.008945607580244541, + 0.019029967486858368, + 0.07069484889507294, + 0.004077581223100424, + 0.005604530218988657, + 0.004206812474876642, + 0.005373610649257898, + 0.004829037003219128, + 0.007370499894022942, + 0.005737836938351393, + 0.005100659094750881, + 0.005431423429399729, + 0.006137606222182512, + 0.006520372349768877, + 0.007844917476177216, + 0.005107077304273844, + 0.004658346530050039, + 0.0058126673102378845, + 0.0075483182445168495, + 0.0069917538203299046, + 0.00898651871830225, + 0.006664793007075787, + 0.00577967707067728, + 0.016837162896990776, + 0.007556980475783348, + 0.006712936796247959, + 0.005681813228875399, + 0.003149176947772503, + 0.005369978025555611, + 0.004179049748927355, + 0.005838512908667326, + 0.004092562478035688, + 0.004422807600349188, + 0.004433640278875828, + 0.0040912507101893425, + 0.005318731535226107, + 0.01145430002361536, + 0.006737764924764633, + 0.006562927737832069, + 0.05413872376084328, + 0.05069372430443764, + 0.006980297155678272, + 0.0053395177237689495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.01011236198246479 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607998088002205 + ], + "opposite_logprob_tokens": [ + -4.141079902648926 + ], + "logprob_reward": -0.01607998088002205, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607998088002205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.797642171382904, + 0.014941695146262646, + 0.010599298402667046, + 0.005876393057405949, + 0.03244808316230774, + 0.012533094733953476, + 0.010881362482905388, + 0.03075837902724743, + 0.017604297026991844, + 0.06671521067619324 + ], + "credit_map": [ + 0.017116393893957138, + 0.01970820315182209, + 0.33150967955589294, + 0.33180075883865356, + 0.0177235696464777, + 0.009710313752293587, + 0.04854405298829079, + 0.02152925357222557, + 0.000884360633790493, + 0.0015970307867974043, + 0.001588995335623622, + 0.003335787681862712, + 0.0023796383757144213, + 0.0010274109663441777, + 0.0021831016056239605, + 0.0019453690620139241, + 0.0009052766254171729, + 0.0013517758343368769, + 0.0015913861570879817, + 0.001087802927941084, + 0.0018782236147671938, + 0.001489462098106742, + 0.001062441966496408, + 0.0012329311575740576, + 0.0007160274544730783, + 0.000620964216068387, + 0.000735806068405509, + 0.0007900884374976158, + 0.0005708040553145111, + 0.0006354721263051033, + 0.000770326005294919, + 0.0010369051015004516, + 0.0014846357516944408, + 0.0008400592487305403, + 0.012287118472158909, + 0.0016806265339255333, + 0.0009112399420700967, + 0.0014340763445943594, + 0.002546105068176985, + 0.011264223605394363, + 0.0014140387065708637, + 0.0015988966915756464, + 0.0024193855933845043, + 0.0015038930578157306, + 0.001131564611569047, + 0.0010415171273052692, + 0.0018483700696378946, + 0.001575428992509842, + 0.0008849080768413842, + 0.00188635871745646, + 0.0010070021962746978, + 0.0014238354051485658, + 0.0014492871705442667, + 0.0016743890009820461, + 0.0012695027980953455, + 0.001286080339923501, + 0.0037122583016753197, + 0.0038357677403837442, + 0.004332939628511667, + 0.0027020862326025963, + 0.006704557687044144, + 0.003160377498716116, + 0.0030718904454261065, + 0.0032385040540248156, + 0.0020284445490688086, + 0.002938373014330864, + 0.001957090338692069, + 0.0026596221141517162, + 0.002122002188116312, + 0.00228480389341712, + 0.0018537837313488126, + 0.0017601791769266129, + 0.003908525686711073, + 0.005873985588550568, + 0.003439501393586397, + 0.003020782256498933, + 0.024058504030108452, + 0.020242994651198387, + 0.003341092262417078, + 0.0028298222459852695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.01607998088002205 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029812438413500786 + ], + "opposite_logprob_tokens": [ + -3.5298123359680176 + ], + "logprob_reward": -0.029812438413500786, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029812438413500786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8346778750419617, + 0.01466695312410593, + 0.005352343898266554, + 0.016316143795847893, + 0.055538855493068695, + 0.007965316995978355, + 0.012219995260238647, + 0.015882862731814384, + 0.010213241912424564, + 0.027166396379470825 + ], + "credit_map": [ + 0.01759670116007328, + 0.01968040131032467, + 0.3492019474506378, + 0.3527616858482361, + 0.015643512830138206, + 0.009600500576198101, + 0.04831103980541229, + 0.021882109344005585, + 0.0008445814601145685, + 0.0017407066188752651, + 0.0021399592515081167, + 0.002391480840742588, + 0.0026252097450196743, + 0.0012826130259782076, + 0.0018425595480948687, + 0.001799845602363348, + 0.0004267541225999594, + 0.0006840902497060597, + 0.000954285089392215, + 0.0006193867884576321, + 0.0008705976651981473, + 0.0005058295791968703, + 0.000630726688541472, + 0.0006606742972508073, + 0.0016651955666020513, + 0.0013819793239235878, + 0.003911981824785471, + 0.00129350321367383, + 0.0012714210897684097, + 0.0011060144752264023, + 0.0020199508871883154, + 0.0036660991609096527, + 0.0030678685288876295, + 0.001727258670143783, + 0.01980973593890667, + 0.0030677998438477516, + 0.0024259244091808796, + 0.0031814039684832096, + 0.004788221791386604, + 0.017470646649599075, + 0.0005711065023206174, + 0.0012497932184487581, + 0.0011940001277253032, + 0.0008713113493286073, + 0.0009300159872509539, + 0.0011955654481425881, + 0.0009518465958535671, + 0.0010016791056841612, + 0.0011751351412385702, + 0.0020842773374170065, + 0.0013190757017582655, + 0.0013460202608257532, + 0.0017454400658607483, + 0.0014468268491327763, + 0.0016694387886673212, + 0.0014337822794914246, + 0.00240287184715271, + 0.0022173067554831505, + 0.0012463926104828715, + 0.0014905795687809587, + 0.0036976009141653776, + 0.0017492539482191205, + 0.0016898319590836763, + 0.0013890264090150595, + 0.0009734508930705488, + 0.001785265514627099, + 0.0012280868832021952, + 0.0010729959467425942, + 0.001833306043408811, + 0.0012811222113668919, + 0.0010062332730740309, + 0.0010327826021239161, + 0.002506407443434, + 0.0022707530297338963, + 0.0026945893187075853, + 0.001702355919405818, + 0.006922948639839888, + 0.00719534233212471, + 0.0017788336845114827, + 0.0020951663609594107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.029812438413500786 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890280082821846 + ], + "opposite_logprob_tokens": [ + -2.4640281200408936 + ], + "logprob_reward": -0.0890280082821846, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890280082821846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3755348026752472, + 0.04595755785703659, + 0.05318199470639229, + 0.04431375861167908, + 0.10240106284618378, + 0.04933612421154976, + 0.07610715925693512, + 0.09193181991577148, + 0.045864615589380264, + 0.11537110805511475 + ], + "credit_map": [ + 0.008925579488277435, + 0.011265764944255352, + 0.1533670276403427, + 0.1631738841533661, + 0.007289054337888956, + 0.006148140411823988, + 0.013764863833785057, + 0.011600499041378498, + 0.0025751017965376377, + 0.005197981372475624, + 0.00591457961127162, + 0.010622347705066204, + 0.007151959463953972, + 0.003459727158769965, + 0.004974702373147011, + 0.006061159539967775, + 0.006011446006596088, + 0.00720270024612546, + 0.0073763532564044, + 0.003813399001955986, + 0.01202149223536253, + 0.006674400996416807, + 0.003961009904742241, + 0.0061211902648210526, + 0.002010694472119212, + 0.0035538137890398502, + 0.006149055436253548, + 0.008633880876004696, + 0.0028648532461375, + 0.003711194032803178, + 0.009403309784829617, + 0.00798695720732212, + 0.005351794417947531, + 0.003744791029021144, + 0.03422627970576286, + 0.007205140311270952, + 0.004916023928672075, + 0.005647636018693447, + 0.009252632968127728, + 0.03205676004290581, + 0.005480210762470961, + 0.0064940000884234905, + 0.005266163963824511, + 0.00742693105712533, + 0.005933139938861132, + 0.006548748817294836, + 0.005387652665376663, + 0.006799277849495411, + 0.0075480761006474495, + 0.011752963066101074, + 0.007230222225189209, + 0.009018928743898869, + 0.009742503054440022, + 0.015725774690508842, + 0.007313891313970089, + 0.007774800527840853, + 0.011068654246628284, + 0.012505259364843369, + 0.007446387317031622, + 0.010984106920659542, + 0.01634548045694828, + 0.01385599933564663, + 0.010591219179332256, + 0.009134712629020214, + 0.0035237756092101336, + 0.006910407450050116, + 0.003895132802426815, + 0.005102998577058315, + 0.008240156807005405, + 0.007705243304371834, + 0.0056390478275716305, + 0.004847852047532797, + 0.0041764890775084496, + 0.010274255648255348, + 0.006399942096322775, + 0.008200481534004211, + 0.0386834591627121, + 0.03557504713535309, + 0.007889466360211372, + 0.004171967506408691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.0890280082821846 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide shot gazing from afar at a surreal environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a hanging garden of red-tongued bamboo floating above a cyberpunk city?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bamboo base a mossy biomechanical platform with gleaming tubes and circuitry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bamboo dark with a black sheen and polished leaves reflecting magenta and cyan light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a swirling sea of low-resolution pixelated electronic clouds below the bamboo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011067195795476437 + ], + "opposite_logprob_tokens": [ + -4.5110673904418945 + ], + "logprob_reward": -0.011067195795476437, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011067195795476437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8459317684173584, + 0.00693345395848155, + 0.002648282330483198, + 0.003586592385545373, + 0.050527770072221756, + 0.0021143993362784386, + 0.005148336291313171, + 0.00988355278968811, + 0.005421869922429323, + 0.06780404597520828 + ], + "credit_map": [ + 0.017857398837804794, + 0.02033344842493534, + 0.3557189106941223, + 0.35974496603012085, + 0.015445047058165073, + 0.008735733106732368, + 0.046872351318597794, + 0.021223895251750946, + 0.00043793366057798266, + 0.0008497902890667319, + 0.001075663836672902, + 0.0010903693037107587, + 0.0012957132421433926, + 0.0005629412480629981, + 0.0007925081299617887, + 0.0008285344229079783, + 0.00022184067347552627, + 0.0002652715193107724, + 0.0005718053434975445, + 0.0003866732877213508, + 0.00036889893817715347, + 0.00027906420291401446, + 0.00027298019267618656, + 0.00028174815815873444, + 0.00023054640041664243, + 0.00029320476460270584, + 0.0006416206015273929, + 0.0005454238271340728, + 0.0002779968490358442, + 0.0003521097532939166, + 0.0005747085670009255, + 0.0006709813023917377, + 0.0023866069968789816, + 0.0010990884620696306, + 0.018560681492090225, + 0.0027078301645815372, + 0.0017454969929531217, + 0.0020409717690199614, + 0.004888365976512432, + 0.01709873229265213, + 0.00018814056238625199, + 0.0002527822507545352, + 0.00022495703888125718, + 0.0003154336882289499, + 0.00024834435316734016, + 0.0002958361874334514, + 0.0002917761739809066, + 0.0002971290668938309, + 0.000356982956873253, + 0.0006595043814741075, + 0.0004982681130059063, + 0.0004007458919659257, + 0.0008826471166685224, + 0.0007416630396619439, + 0.0006264045368880033, + 0.0009821199346333742, + 0.0011464031413197517, + 0.0012617837637662888, + 0.001090451143682003, + 0.00096946454141289, + 0.0024338215589523315, + 0.001011119456961751, + 0.0009754005586728454, + 0.000995108624920249, + 0.0005468136514537036, + 0.0007634111680090427, + 0.0006111354450695217, + 0.0006697667995467782, + 0.000973699614405632, + 0.0008903153939172626, + 0.0004737953713629395, + 0.0004929326241835952, + 0.0020297530572861433, + 0.0029034405015408993, + 0.002543930197134614, + 0.0014990699710324407, + 0.028389722108840942, + 0.02676583267748356, + 0.0023395090829581022, + 0.0013327877968549728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.011067195795476437 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_2/index_tensor([3612, 3612])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_2/index_tensor([3612, 3612])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..aed9dbbfa64880421f3eac1a93a0e3f719dd1ad6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_2/index_tensor([3612, 3612])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "first-person, subjective shot completely immerses the viewer in a magnificent Gothic hall. The camera tilts upward from a level angle, revealing the soaring ribbed vaults, their intricate stonework looming in the gloom. Countless magical paper airplanes, emitting a faint white glow, soar gracefully beneath the vast dome, like birds imbued with life, circling and weaving in a harmonious and ordered rhythm, leaving fleeting golden trails in their wake. The view slowly returns to the front, focusing on an elder at a distant podium. He possesses shoulder-length silver-white hair and a kind face lined with lines of wisdom. He wears a solemn, dark blue velvet robe, the edges of which are delicately embroidered with silver stars. He preaches from behind an ancient, dark oak lectern, his hands resting steadily on the edge, his body leaning slightly forward. His expression is focused and calm, as if his gentle yet resolute voice echoes silently throughout the room. Rich beams of golden sunlight penetrate the massive stained-glass windows overhead, casting the vibrant colors of the mythological scene onto the ancient stone floor. A soft halo of light forms behind the elder, framing him with a sacred and warm silhouette. The entire scene is enveloped in a tranquil, solemn, and hopeful atmosphere, creating a profound sense of warmth and healing.", + "seed": 21936, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_2/index_3612_seed_21936_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.16239012610167264, + "vlm_logprob": -0.16239012610167264, + "vlm_yes_no_margin": 3.1500000238418577, + "vlm_token_credit": -0.16239012610167264, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14680925011634827 + ], + "opposite_logprob_tokens": [ + -2.0218093395233154 + ], + "logprob_reward": -0.14680925011634827, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14680925011634827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13232707977294922, + 0.15755261480808258, + 0.08812282979488373, + 0.11500592529773712, + 0.22710074484348297, + 0.03147966042160988, + 0.038442254066467285, + 0.07169638574123383, + 0.08612176030874252, + 0.05215080827474594 + ], + "credit_map": [ + 0.01935962773859501, + 0.017231417819857597, + 0.009763156063854694, + 0.019643332809209824, + 0.01911056414246559, + 0.01091435644775629, + 0.020925231277942657, + 0.015379389747977257, + 0.022197546437382698, + 0.027282612398266792, + 0.018656080588698387, + 0.015996716916561127, + 0.020075859501957893, + 0.020366014912724495, + 0.0172980148345232, + 0.015679772943258286, + 0.006819752044975758, + 0.008970367722213268, + 0.008200500160455704, + 0.0092428307980299, + 0.017333168536424637, + 0.013966666534543037, + 0.009999428875744343, + 0.013590112328529358, + 0.01728629507124424, + 0.020245501771569252, + 0.014271833933889866, + 0.006735560018569231, + 0.019618673250079155, + 0.016103344038128853, + 0.015198715962469578, + 0.0055460017174482346, + 0.011628874577581882, + 0.016202565282583237, + 0.0157085619866848, + 0.022079303860664368, + 0.014011858962476254, + 0.07278528064489365, + 0.006877764593809843, + 0.06780652701854706, + 0.004485492128878832, + 0.005595811642706394, + 0.0030806746799498796, + 0.00461237970739603, + 0.003778893267735839, + 0.003611200489103794, + 0.0031014324631541967, + 0.0032137746457010508, + 0.0031395473051816225, + 0.008868250995874405, + 0.003916044719517231, + 0.0031384688336402178, + 0.006135237403213978, + 0.006398491561412811, + 0.00227896636351943, + 0.004567247815430164, + 0.009408263489603996, + 0.012597952969372272, + 0.005299397744238377, + 0.005685457028448582, + 0.014473349787294865, + 0.00818474031984806, + 0.005087125580757856, + 0.010960100218653679, + 0.013730764389038086, + 0.016339298337697983, + 0.009919899515807629, + 0.009830835275352001, + 0.010180731303989887, + 0.010477005504071712, + 0.009124010801315308, + 0.006519215647131205, + 0.004182983189821243, + 0.012003155425190926, + 0.00452455086633563, + 0.005807857494801283, + 0.010379632003605366, + 0.005180943291634321, + 0.005987329874187708, + 0.004084354732185602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.14680925011634827 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328150629997253 + ], + "opposite_logprob_tokens": [ + -0.7578150629997253 + ], + "logprob_reward": -0.6328150629997253, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328150629997253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09225403517484665, + 0.23344817757606506, + 0.15028595924377441, + 0.04022800549864769, + 0.375270277261734, + 0.017134716734290123, + 0.023401526734232903, + 0.012415428645908833, + 0.01650254987180233, + 0.039059314876794815 + ], + "credit_map": [ + 0.00759874377399683, + 0.011228131130337715, + 0.010312053374946117, + 0.015345823019742966, + 0.013826578855514526, + 0.00993198249489069, + 0.012785340659320354, + 0.011225386522710323, + 0.027531754225492477, + 0.03609419986605644, + 0.026856182143092155, + 0.022760184481739998, + 0.031796276569366455, + 0.035121046006679535, + 0.025776641443371773, + 0.027511902153491974, + 0.007059890311211348, + 0.01271789986640215, + 0.010997314006090164, + 0.016468778252601624, + 0.03768979012966156, + 0.02478989027440548, + 0.020961768925189972, + 0.01960062049329281, + 0.004659514874219894, + 0.0066635701805353165, + 0.004929642658680677, + 0.0023839243222028017, + 0.009038981050252914, + 0.006237777415663004, + 0.004841902293264866, + 0.001472694450058043, + 0.016574598848819733, + 0.019096078351140022, + 0.02483377791941166, + 0.034916166216135025, + 0.026384735479950905, + 0.1135987862944603, + 0.01729121059179306, + 0.12257492542266846, + 0.0013023552019149065, + 0.0035501006059348583, + 0.002707522129639983, + 0.002728088293224573, + 0.0012810983462259173, + 0.0016857088776305318, + 0.002025030320510268, + 0.001854812609963119, + 0.002898924285545945, + 0.0026886085979640484, + 0.0028717033565044403, + 0.0019718948751688004, + 0.00398975471034646, + 0.0035971065517514944, + 0.0013906965032219887, + 0.003992837853729725, + 0.000988281099125743, + 0.0023830432910472155, + 0.001230914145708084, + 0.0013154236366972327, + 0.0023303909692913294, + 0.0013635160867124796, + 0.0013719921698793769, + 0.0014318672474473715, + 0.002348502865061164, + 0.0038319833111017942, + 0.0011907576117664576, + 0.0015515738632529974, + 0.002311321208253503, + 0.002005004556849599, + 0.0014350810088217258, + 0.0018283259123563766, + 0.003503269050270319, + 0.008683690801262856, + 0.0033124079927802086, + 0.004180641379207373, + 0.007123927585780621, + 0.003779285354539752, + 0.004774872679263353, + 0.003701221663504839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.6328150629997253 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076249269768595695 + ], + "opposite_logprob_tokens": [ + -4.882625102996826 + ], + "logprob_reward": -0.0076249269768595695, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076249269768595695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1380559206008911, + 0.12403565645217896, + 0.04811464250087738, + 0.04150579497218132, + 0.26714223623275757, + 0.08353649079799652, + 0.08318254351615906, + 0.0646258220076561, + 0.05429377406835556, + 0.09550711512565613 + ], + "credit_map": [ + 0.023714806884527206, + 0.015110895037651062, + 0.01550755463540554, + 0.019585514441132545, + 0.01510677207261324, + 0.016278866678476334, + 0.017595594748854637, + 0.0151559142395854, + 0.01186484843492508, + 0.015355397947132587, + 0.012781824916601181, + 0.010769598186016083, + 0.017433391883969307, + 0.022373486310243607, + 0.012301638722419739, + 0.021155480295419693, + 0.003972801845520735, + 0.007035462651401758, + 0.005286873783916235, + 0.004716457333415747, + 0.01206495612859726, + 0.008548827841877937, + 0.0028883772902190685, + 0.00360089004971087, + 0.004664632026106119, + 0.005961528979241848, + 0.006252374034374952, + 0.003070696024224162, + 0.0067385220900177956, + 0.006998543161898851, + 0.0043168156407773495, + 0.003502686507999897, + 0.011913638561964035, + 0.014115422032773495, + 0.02002740651369095, + 0.027593454346060753, + 0.019718779250979424, + 0.0752917006611824, + 0.013750207610428333, + 0.08473162353038788, + 0.0069655198603868484, + 0.014294827356934547, + 0.013370722532272339, + 0.014140305109322071, + 0.007086047902703285, + 0.00814794097095728, + 0.0088889105245471, + 0.010642217472195625, + 0.005724396090954542, + 0.01705731265246868, + 0.005760107189416885, + 0.0084760170429945, + 0.01642303168773651, + 0.00854236539453268, + 0.0053022499196231365, + 0.01589706540107727, + 0.005315021146088839, + 0.015599753707647324, + 0.006887189578264952, + 0.007013093680143356, + 0.01011514849960804, + 0.00917765125632286, + 0.005686459131538868, + 0.004831511061638594, + 0.004800902213901281, + 0.01192496344447136, + 0.005731188226491213, + 0.007536186370998621, + 0.007242904510349035, + 0.005516924429684877, + 0.005916601978242397, + 0.005624108482152224, + 0.008160486817359924, + 0.022753840312361717, + 0.008174539543688297, + 0.009370334446430206, + 0.018602615222334862, + 0.009129459038376808, + 0.010531000792980194, + 0.008784841746091843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0076249269768595695 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864158570766449 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.00864158570766449, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864158570766449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09752721339464188, + 0.10767177492380142, + 0.144300639629364, + 0.0597248449921608, + 0.39430707693099976, + 0.014126484282314777, + 0.054570168256759644, + 0.07188932597637177, + 0.043629277497529984, + 0.012253228574991226 + ], + "credit_map": [ + 0.012919457629323006, + 0.011266031302511692, + 0.006573908030986786, + 0.013295082375407219, + 0.01353083923459053, + 0.009385732002556324, + 0.017305416986346245, + 0.013250738382339478, + 0.013748022727668285, + 0.017023956403136253, + 0.014544344507157803, + 0.011633732356131077, + 0.013950973749160767, + 0.013675120659172535, + 0.012893227860331535, + 0.010202376171946526, + 0.008089683949947357, + 0.013647463172674179, + 0.009410656057298183, + 0.01584504544734955, + 0.04261555150151253, + 0.026534244418144226, + 0.014634868130087852, + 0.013523105531930923, + 0.010424310341477394, + 0.005062095820903778, + 0.008664590306580067, + 0.003895557951182127, + 0.011551127769052982, + 0.010271675884723663, + 0.006856758147478104, + 0.0029987215530127287, + 0.016178743913769722, + 0.013376180082559586, + 0.019127771258354187, + 0.0349896214902401, + 0.02880401723086834, + 0.14274974167346954, + 0.011287569999694824, + 0.12779337167739868, + 0.0017456798814237118, + 0.0028002362232655287, + 0.001371922786347568, + 0.0019237572560086846, + 0.0015388394240289927, + 0.0019580726511776447, + 0.0014041949762031436, + 0.0013837796868756413, + 0.007463039830327034, + 0.009462864138185978, + 0.005286002531647682, + 0.005213434807956219, + 0.005139083601534367, + 0.010662376880645752, + 0.002811983460560441, + 0.008531376719474792, + 0.009188391268253326, + 0.018000371754169464, + 0.005275717005133629, + 0.0064764367416501045, + 0.01117083802819252, + 0.00820714607834816, + 0.004398820456117392, + 0.009171602316200733, + 0.006720159202814102, + 0.009781202301383018, + 0.004213329870253801, + 0.004718670155853033, + 0.006392980460077524, + 0.004547053016722202, + 0.0038127985317260027, + 0.0034430769737809896, + 0.0013273762306198478, + 0.002268188865855336, + 0.0012475287076085806, + 0.0016029149992391467, + 0.0013471732381731272, + 0.0018714589532464743, + 0.0012488594511523843, + 0.0013397267321124673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.00864158570766449 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605980470776558 + ], + "opposite_logprob_tokens": [ + -4.141059875488281 + ], + "logprob_reward": -0.01605980470776558, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605980470776558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16573673486709595, + 0.08668459951877594, + 0.17610464990139008, + 0.14999239146709442, + 0.13455744087696075, + 0.053146135061979294, + 0.09499432891607285, + 0.05044708773493767, + 0.049259211868047714, + 0.039077453315258026 + ], + "credit_map": [ + 0.02148408629000187, + 0.01754097454249859, + 0.01809065230190754, + 0.02736213244497776, + 0.01871408335864544, + 0.018077949061989784, + 0.0279448963701725, + 0.016521962359547615, + 0.0091484310105443, + 0.011035055853426456, + 0.007604353129863739, + 0.0085655078291893, + 0.012995269149541855, + 0.015224468894302845, + 0.007252214010804892, + 0.01485929824411869, + 0.0184914693236351, + 0.027998702600598335, + 0.029377281665802002, + 0.019455451518297195, + 0.030222471803426743, + 0.029147222638130188, + 0.01124950684607029, + 0.01016254536807537, + 0.020368967205286026, + 0.01942899078130722, + 0.019442088901996613, + 0.010925131849944592, + 0.03293783962726593, + 0.022707605734467506, + 0.01529434509575367, + 0.008887422271072865, + 0.00780452648177743, + 0.012399914674460888, + 0.013145893812179565, + 0.014782018959522247, + 0.01080826111137867, + 0.02813410386443138, + 0.010121239349246025, + 0.03736148774623871, + 0.0038736106362193823, + 0.009297596290707588, + 0.0067709069699049, + 0.008069618605077267, + 0.003947890363633633, + 0.007423908915370703, + 0.004012326709926128, + 0.009750278666615486, + 0.007896666415035725, + 0.01617657020688057, + 0.00751640647649765, + 0.00835991371423006, + 0.012345931492745876, + 0.012577779591083527, + 0.005350450519472361, + 0.02477060630917549, + 0.003754881676286459, + 0.015366381034255028, + 0.0037538150791078806, + 0.006628341041505337, + 0.006414821371436119, + 0.005637726280838251, + 0.004425991792231798, + 0.004465132486075163, + 0.004648488014936447, + 0.015393930487334728, + 0.0037493756972253323, + 0.004867600742727518, + 0.00733318692073226, + 0.005116065964102745, + 0.00423680804669857, + 0.003913756459951401, + 0.003461199114099145, + 0.011101722717285156, + 0.0031237900257110596, + 0.004870283417403698, + 0.00516600301489234, + 0.0038918417412787676, + 0.004609900526702404, + 0.0028527125250548124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.01605980470776558 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "first-person, subjective shot completely immerses the viewer in a magnificent Gothic hall. The camera tilts upward from a level angle, revealing the soaring ribbed vaults, their intricate stonework looming in the gloom. Countless magical paper airplanes, emitting a faint white glow, soar gracefully beneath the vast dome, like birds imbued with life, circling and weaving in a harmonious and ordered rhythm, leaving fleeting golden trails in their wake. The view slowly returns to the front, focusing on an elder at a distant podium. He possesses shoulder-length silver-white hair and a kind face lined with lines of wisdom. He wears a solemn, dark blue velvet robe, the edges of which are delicately embroidered with silver stars. He preaches from behind an ancient, dark oak lectern, his hands resting steadily on the edge, his body leaning slightly forward. His expression is focused and calm, as if his gentle yet resolute voice echoes silently throughout the room. Rich beams of golden sunlight penetrate the massive stained-glass windows overhead, casting the vibrant colors of the mythological scene onto the ancient stone floor. A soft halo of light forms behind the elder, framing him with a sacred and warm silhouette. The entire scene is enveloped in a tranquil, solemn, and hopeful atmosphere, creating a profound sense of warmth and healing.", + "seed": 21937, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_2/index_3612_seed_21937_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10176946390420198, + "vlm_logprob": -0.10176946390420198, + "vlm_yes_no_margin": 3.174999952316284, + "vlm_token_credit": -0.10176946390420198, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18647119402885437 + ], + "opposite_logprob_tokens": [ + -1.8114712238311768 + ], + "logprob_reward": -0.18647119402885437, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18647119402885437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5566198229789734, + 0.05166042223572731, + 0.14960505068302155, + 0.0559048131108284, + 0.06356249004602432, + 0.07266323268413544, + 0.020634539425373077, + 0.009235799312591553, + 0.013408854603767395, + 0.00670495443046093 + ], + "credit_map": [ + 0.040608569979667664, + 0.23539048433303833, + 0.026210859417915344, + 0.04868658632040024, + 0.0333491787314415, + 0.14280293881893158, + 0.015706058591604233, + 0.013865143992006779, + 0.0035816002637147903, + 0.0036789446603506804, + 0.007187026087194681, + 0.00468361284583807, + 0.007075269240885973, + 0.00804617628455162, + 0.00569671206176281, + 0.011711076833307743, + 0.0196712426841259, + 0.020277438685297966, + 0.017220834270119667, + 0.01575327292084694, + 0.0189453586935997, + 0.022648755460977554, + 0.015008634887635708, + 0.020079493522644043, + 0.004644410219043493, + 0.0033434780780225992, + 0.005770551506429911, + 0.005588885862380266, + 0.005496553145349026, + 0.009695492684841156, + 0.012719494290649891, + 0.008645947091281414, + 0.005444258451461792, + 0.004792809020727873, + 0.005601983983069658, + 0.011216380633413792, + 0.004425801802426577, + 0.009293212555348873, + 0.011662989854812622, + 0.0111250514164567, + 0.003223442006856203, + 0.0055235507898032665, + 0.0056066131219267845, + 0.00806700624525547, + 0.01854800619184971, + 0.010464249178767204, + 0.010660272091627121, + 0.010570090264081955, + 0.0015597034944221377, + 0.0016369613585993648, + 0.0022044084034860134, + 0.0024047112092375755, + 0.0037230050656944513, + 0.0026186855975538492, + 0.0033170930109918118, + 0.0031699687242507935, + 0.0005130807403475046, + 0.0007933331071399152, + 0.000692003988660872, + 0.0008909633615985513, + 0.0014841188676655293, + 0.0009522067848592997, + 0.0018911159131675959, + 0.0020189760252833366, + 0.0014620426809415221, + 0.0012120554456487298, + 0.0012805622536689043, + 0.000982566038146615, + 0.0011435472406446934, + 0.0016413838602602482, + 0.0026513829361647367, + 0.0030353122856467962, + 0.0006490551168099046, + 0.0006289576995186508, + 0.0006152858841232955, + 0.0005707516684196889, + 0.000970950408373028, + 0.0010188681771978736, + 0.001115419203415513, + 0.0011356663890182972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.18647119402885437 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28138667345046997 + ], + "opposite_logprob_tokens": [ + -1.4063866138458252 + ], + "logprob_reward": -0.28138667345046997, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28138667345046997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7505937218666077, + 0.0919158086180687, + 0.04591052234172821, + 0.025960518047213554, + 0.025628255680203438, + 0.015025808475911617, + 0.008912784978747368, + 0.007332048378884792, + 0.022888559848070145, + 0.005832029972225428 + ], + "credit_map": [ + 0.05300067737698555, + 0.3106800317764282, + 0.03560027852654457, + 0.04456789419054985, + 0.04532305896282196, + 0.2214929312467575, + 0.019204482436180115, + 0.020724283531308174, + 0.012788219377398491, + 0.016600387170910835, + 0.013037122786045074, + 0.007421381771564484, + 0.01001728419214487, + 0.012664544396102428, + 0.009528189897537231, + 0.009858674369752407, + 0.004391319118440151, + 0.007677761372178793, + 0.0040497975423932076, + 0.004562695976346731, + 0.006132885813713074, + 0.006665886379778385, + 0.006617332808673382, + 0.005812840070575476, + 0.002320784144103527, + 0.0027727277483791113, + 0.003328364109620452, + 0.0028886066284030676, + 0.003319798968732357, + 0.003553019603714347, + 0.004160627722740173, + 0.0036165856290608644, + 0.0028675321955233812, + 0.001984529895707965, + 0.0022798050194978714, + 0.0028784514870494604, + 0.002773116109892726, + 0.003535819472745061, + 0.005556768737733364, + 0.0037522295024245977, + 0.001306612160988152, + 0.0014752206625416875, + 0.0009887083433568478, + 0.0013019135221838951, + 0.002076277043670416, + 0.00283187092281878, + 0.0030581362079828978, + 0.001987067051231861, + 0.0008622665191069245, + 0.0010393523843958974, + 0.0007118879002518952, + 0.0006048398208804429, + 0.0016525066457688808, + 0.0015433654189109802, + 0.0017604657914489508, + 0.0007380995084531605, + 0.00038872702862136066, + 0.00043271089089103043, + 0.0004368931404314935, + 0.00051704520592466, + 0.0018413218203932047, + 0.000986209837719798, + 0.0015026975888758898, + 0.0012264418182894588, + 0.0023463794495910406, + 0.001726880087517202, + 0.0024481043219566345, + 0.001874182024039328, + 0.0021908567287027836, + 0.0022737900726497173, + 0.004603810142725706, + 0.00542455492541194, + 0.0003352425410412252, + 0.00032911438029259443, + 0.00035334733547642827, + 0.0003715851344168186, + 0.0011119904229417443, + 0.0007177190273068845, + 0.0012592985294759274, + 0.0013537319609895349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.28138667345046997 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183909356594086 + ], + "opposite_logprob_tokens": [ + -4.018183708190918 + ], + "logprob_reward": -0.018183909356594086, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018183909356594086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4897775948047638, + 0.059666458517313004, + 0.05223517492413521, + 0.044307366013526917, + 0.0741717591881752, + 0.06901770830154419, + 0.08404895663261414, + 0.04480079188942909, + 0.04897155985236168, + 0.0330025888979435 + ], + "credit_map": [ + 0.03606203943490982, + 0.20146732032299042, + 0.024970872327685356, + 0.04317343980073929, + 0.028955718502402306, + 0.12544696033000946, + 0.014237514697015285, + 0.015463776886463165, + 0.0044526890851557255, + 0.003648999845609069, + 0.009462529793381691, + 0.0052693868055939674, + 0.005578180775046349, + 0.009009641595184803, + 0.005654068663716316, + 0.016590965911746025, + 0.005723316688090563, + 0.004652019590139389, + 0.0064550056122243404, + 0.005483036860823631, + 0.006756707560271025, + 0.008208341896533966, + 0.005406477488577366, + 0.009550274349749088, + 0.0036127506755292416, + 0.0043285139836370945, + 0.004829370882362127, + 0.00399508373811841, + 0.00425285380333662, + 0.007090759929269552, + 0.009717031382024288, + 0.006481005810201168, + 0.005497531034052372, + 0.003958848770707846, + 0.005095543805509806, + 0.006860394962131977, + 0.005189207382500172, + 0.011408588849008083, + 0.016369592398405075, + 0.019792063161730766, + 0.006285167299211025, + 0.0068827904760837555, + 0.006303397007286549, + 0.008245605044066906, + 0.007386486046016216, + 0.010450142435729504, + 0.01098665501922369, + 0.012477468699216843, + 0.0040968069806694984, + 0.004963954444974661, + 0.008221224881708622, + 0.009401186369359493, + 0.016811130568385124, + 0.009075167588889599, + 0.015224906615912914, + 0.01625458337366581, + 0.0018810377223417163, + 0.002718516392633319, + 0.003913035616278648, + 0.005377104040235281, + 0.006493763532489538, + 0.005726702976971865, + 0.010358412750065327, + 0.008332225494086742, + 0.002571489429101348, + 0.0027903213631361723, + 0.0048542027361691, + 0.005181474145501852, + 0.00476047582924366, + 0.005955992732197046, + 0.011010820977389812, + 0.011846786364912987, + 0.0014898806111887097, + 0.002063424326479435, + 0.0033504078164696693, + 0.004030036274343729, + 0.005730655509978533, + 0.003916732035577297, + 0.007289050612598658, + 0.005132400896400213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.018183909356594086 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741280667483807 + ], + "opposite_logprob_tokens": [ + -5.006741046905518 + ], + "logprob_reward": -0.006741280667483807, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741280667483807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20223554968833923, + 0.5157228112220764, + 0.046245183795690536, + 0.05213361978530884, + 0.037810422480106354, + 0.057099081575870514, + 0.03479371219873428, + 0.023255614563822746, + 0.022534728050231934, + 0.008169317618012428 + ], + "credit_map": [ + 0.01940368302166462, + 0.07069117575883865, + 0.01935844123363495, + 0.017905890941619873, + 0.016875343397259712, + 0.02846849150955677, + 0.016578825190663338, + 0.012953699566423893, + 0.041914407163858414, + 0.03082394227385521, + 0.07711353898048401, + 0.04967913031578064, + 0.049442484974861145, + 0.07574710994958878, + 0.05174350365996361, + 0.1392587125301361, + 0.004879281390458345, + 0.006137663498520851, + 0.0067130629904568195, + 0.004161059390753508, + 0.0060977628454566, + 0.0060315257869660854, + 0.00454885745421052, + 0.00767596997320652, + 0.004890843294560909, + 0.004187355283647776, + 0.004869828466325998, + 0.009290087968111038, + 0.005527233239263296, + 0.0074824439361691475, + 0.011010588146746159, + 0.004875238984823227, + 0.0028193381149321795, + 0.0035705685149878263, + 0.0029734312556684017, + 0.004204757045954466, + 0.005171562545001507, + 0.005284138955175877, + 0.008036532439291477, + 0.005750092677772045, + 0.00571093987673521, + 0.006245236843824387, + 0.005238965153694153, + 0.006701677106320858, + 0.011447424069046974, + 0.007372586522251368, + 0.008652959950268269, + 0.005729293450713158, + 0.003338905982673168, + 0.0029824671801179647, + 0.005064613185822964, + 0.005324226804077625, + 0.003019944066181779, + 0.0032432314474135637, + 0.00605215597897768, + 0.00576816825196147, + 0.00247783400118351, + 0.002033048076555133, + 0.0031375745311379433, + 0.0026900500524789095, + 0.0021623033098876476, + 0.002224344527348876, + 0.004034708719700575, + 0.004495751578360796, + 0.0033757661003619432, + 0.0026308088563382626, + 0.003121904330328107, + 0.00211290642619133, + 0.0023247073404490948, + 0.0033347790595144033, + 0.0030249294359236956, + 0.0026089251041412354, + 0.0005333897424861789, + 0.0004890758427791297, + 0.0010030022822320461, + 0.0009311335743404925, + 0.0008430637535639107, + 0.0009678124915808439, + 0.0016801784513518214, + 0.0017216617707163095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.006741280667483807 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606426201760769 + ], + "opposite_logprob_tokens": [ + -4.141064167022705 + ], + "logprob_reward": -0.01606426201760769, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606426201760769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43519479036331177, + 0.05933735892176628, + 0.10738041251897812, + 0.057654622942209244, + 0.03819618001580238, + 0.1015700176358223, + 0.101095050573349, + 0.035237547010183334, + 0.037761300802230835, + 0.02657267265021801 + ], + "credit_map": [ + 0.027455469593405724, + 0.1700061708688736, + 0.024995870888233185, + 0.021353574469685555, + 0.027820557355880737, + 0.13254183530807495, + 0.015734175220131874, + 0.015287154354155064, + 0.005028534214943647, + 0.006343272514641285, + 0.00828919094055891, + 0.0057228319346904755, + 0.006137019023299217, + 0.007898757234215736, + 0.008107260800898075, + 0.011810492724180222, + 0.010328144766390324, + 0.012250750325620174, + 0.01163533702492714, + 0.009736157953739166, + 0.015261116437613964, + 0.016302596777677536, + 0.014990373514592648, + 0.016875937581062317, + 0.0036350858863443136, + 0.004120625089854002, + 0.0064617786556482315, + 0.007345366757363081, + 0.005227914545685053, + 0.010085132904350758, + 0.00946678314357996, + 0.01131193432956934, + 0.002752463100478053, + 0.0036590080708265305, + 0.0029786545783281326, + 0.0034001069143414497, + 0.002796462271362543, + 0.005410729441791773, + 0.0078731132671237, + 0.009325642138719559, + 0.012501022778451443, + 0.009945886209607124, + 0.008980250917375088, + 0.010382609441876411, + 0.016430990770459175, + 0.015384288504719734, + 0.014414530247449875, + 0.01353044155985117, + 0.004135081544518471, + 0.0053693936206400394, + 0.004861505702137947, + 0.006706844083964825, + 0.023140186443924904, + 0.012097693048417568, + 0.019770795479416847, + 0.025013556703925133, + 0.0018746984424069524, + 0.0021142130717635155, + 0.0022855426650494337, + 0.002145896665751934, + 0.005477309226989746, + 0.005026195198297501, + 0.007394468877464533, + 0.00891922228038311, + 0.0015579976607114077, + 0.0020000797230750322, + 0.0021801816765218973, + 0.002682100050151348, + 0.007380272261798382, + 0.0041823238134384155, + 0.007887904532253742, + 0.009890438988804817, + 0.0016191492322832346, + 0.0015855173114687204, + 0.0017174611566588283, + 0.0015314634656533599, + 0.0046102129854261875, + 0.004335361532866955, + 0.005062252748757601, + 0.006111252121627331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.01606426201760769 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "first-person, subjective shot completely immerses the viewer in a magnificent Gothic hall. The camera tilts upward from a level angle, revealing the soaring ribbed vaults, their intricate stonework looming in the gloom. Countless magical paper airplanes, emitting a faint white glow, soar gracefully beneath the vast dome, like birds imbued with life, circling and weaving in a harmonious and ordered rhythm, leaving fleeting golden trails in their wake. The view slowly returns to the front, focusing on an elder at a distant podium. He possesses shoulder-length silver-white hair and a kind face lined with lines of wisdom. He wears a solemn, dark blue velvet robe, the edges of which are delicately embroidered with silver stars. He preaches from behind an ancient, dark oak lectern, his hands resting steadily on the edge, his body leaning slightly forward. His expression is focused and calm, as if his gentle yet resolute voice echoes silently throughout the room. Rich beams of golden sunlight penetrate the massive stained-glass windows overhead, casting the vibrant colors of the mythological scene onto the ancient stone floor. A soft halo of light forms behind the elder, framing him with a sacred and warm silhouette. The entire scene is enveloped in a tranquil, solemn, and hopeful atmosphere, creating a profound sense of warmth and healing.", + "seed": 21938, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_2/index_3612_seed_21938_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09010822232812643, + "vlm_logprob": -0.09010822232812643, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.09010822232812643, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10242826491594315 + ], + "opposite_logprob_tokens": [ + -2.3524281978607178 + ], + "logprob_reward": -0.10242826491594315, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10242826491594315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12843671441078186, + 0.10086656361818314, + 0.06420725584030151, + 0.10662465542554855, + 0.24437884986400604, + 0.06511909514665604, + 0.06052856892347336, + 0.15608195960521698, + 0.03694477677345276, + 0.036811523139476776 + ], + "credit_map": [ + 0.019413450732827187, + 0.01478347834199667, + 0.011909092776477337, + 0.010848172940313816, + 0.019243545830249786, + 0.014721563085913658, + 0.022235924378037453, + 0.0152814956381917, + 0.007094475906342268, + 0.005680263973772526, + 0.031276725232601166, + 0.0209462009370327, + 0.004727223888039589, + 0.020458785817027092, + 0.005010110791772604, + 0.005672786384820938, + 0.005662452895194292, + 0.006659212056547403, + 0.010495157912373543, + 0.008094335906207561, + 0.008378668688237667, + 0.006200385745614767, + 0.010745851323008537, + 0.007971192710101604, + 0.009091897867619991, + 0.013158921152353287, + 0.018900340422987938, + 0.015427699312567711, + 0.0108634689822793, + 0.01132749393582344, + 0.013182840310037136, + 0.014672010205686092, + 0.011064390651881695, + 0.009123079478740692, + 0.014081636443734169, + 0.006485069170594215, + 0.008736993186175823, + 0.09500927478075027, + 0.006486056838184595, + 0.09339234977960587, + 0.00691224355250597, + 0.008750120177865028, + 0.004881818778812885, + 0.0069848704151809216, + 0.008374450728297234, + 0.012360815890133381, + 0.009510008618235588, + 0.007344777695834637, + 0.004137488082051277, + 0.006105723325163126, + 0.005824922118335962, + 0.010984290391206741, + 0.005317233502864838, + 0.011541611514985561, + 0.007681543473154306, + 0.00893576443195343, + 0.004242992028594017, + 0.0020275437273085117, + 0.0018987369257956743, + 0.003670070320367813, + 0.00482482323423028, + 0.06741858273744583, + 0.003142841160297394, + 0.06885640323162079, + 0.005491519346833229, + 0.0033682880457490683, + 0.004626499954611063, + 0.004602616187185049, + 0.003998856525868177, + 0.004991488065570593, + 0.004834794905036688, + 0.005030719097703695, + 0.003091066610068083, + 0.0055654277093708515, + 0.002689855173230171, + 0.0069534326903522015, + 0.004594001453369856, + 0.00524227786809206, + 0.003847033018246293, + 0.004828432574868202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.10242826491594315 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813405990600586 + ], + "opposite_logprob_tokens": [ + -1.4063405990600586 + ], + "logprob_reward": -0.2813405990600586, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813405990600586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1957787126302719, + 0.2009744942188263, + 0.07029858976602554, + 0.06630270928144455, + 0.09025340527296066, + 0.04675165191292763, + 0.073194719851017, + 0.19940124452114105, + 0.023048387840390205, + 0.033996060490608215 + ], + "credit_map": [ + 0.030101807788014412, + 0.0154061084613204, + 0.01255986001342535, + 0.01919221691787243, + 0.02598928101360798, + 0.019743148237466812, + 0.05237983539700508, + 0.020406443625688553, + 0.014583436772227287, + 0.007252135314047337, + 0.060641080141067505, + 0.0378153920173645, + 0.013034203089773655, + 0.0459236316382885, + 0.01088542677462101, + 0.010839203372597694, + 0.002943800063803792, + 0.004635884892195463, + 0.013703792355954647, + 0.014493932016193867, + 0.007498507387936115, + 0.005493765696883202, + 0.009294395335018635, + 0.012234518304467201, + 0.007287084124982357, + 0.008034964092075825, + 0.011591176502406597, + 0.011096090078353882, + 0.006031349301338196, + 0.006101888604462147, + 0.009030770510435104, + 0.007129384204745293, + 0.0077465595677495, + 0.006252647843211889, + 0.006286282557994127, + 0.00571701442822814, + 0.011048832908272743, + 0.020045755431056023, + 0.009353267028927803, + 0.023803045973181725, + 0.005680415779352188, + 0.005021413788199425, + 0.0045780399814248085, + 0.005631637293845415, + 0.0070632887072861195, + 0.005520532373338938, + 0.0073021394200623035, + 0.00595418456941843, + 0.006955085787922144, + 0.006045917049050331, + 0.007785249967128038, + 0.011460330337285995, + 0.005292707588523626, + 0.015502029098570347, + 0.008743870072066784, + 0.011409526690840721, + 0.00485204765573144, + 0.0029164603911340237, + 0.0024802060797810555, + 0.002919220831245184, + 0.006936228834092617, + 0.08571173250675201, + 0.0059841289184987545, + 0.08760121464729309, + 0.0034369106870144606, + 0.002723757643252611, + 0.0028576860204339027, + 0.0028830186929553747, + 0.0016804506303742528, + 0.003951943479478359, + 0.002523562638089061, + 0.002991057699546218, + 0.002949815709143877, + 0.0055108023807406425, + 0.0029557656962424517, + 0.004167202394455671, + 0.004236436448991299, + 0.004257195629179478, + 0.006105579901486635, + 0.0038132639601826668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.2813405990600586 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369441255927086 + ], + "opposite_logprob_tokens": [ + -3.4086945056915283 + ], + "logprob_reward": -0.03369441255927086, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369441255927086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15743954479694366, + 0.08008516579866409, + 0.052898749709129333, + 0.16103902459144592, + 0.11232798546552658, + 0.17568151652812958, + 0.050993192940950394, + 0.10972705483436584, + 0.03863760083913803, + 0.061170198023319244 + ], + "credit_map": [ + 0.030015526339411736, + 0.011269203387200832, + 0.009496105834841728, + 0.013953424990177155, + 0.0180678553879261, + 0.01353186834603548, + 0.045139964669942856, + 0.01596558280289173, + 0.009794129058718681, + 0.009307867847383022, + 0.01165015622973442, + 0.011877642944455147, + 0.006584088783711195, + 0.007158859632909298, + 0.01063960138708353, + 0.013072817586362362, + 0.004967196378856897, + 0.007217134814709425, + 0.009856452234089375, + 0.009433765895664692, + 0.006438642740249634, + 0.004618381150066853, + 0.005803485866636038, + 0.004563688766211271, + 0.018818633630871773, + 0.013964960351586342, + 0.024748744443058968, + 0.018906593322753906, + 0.01796584203839302, + 0.016492197290062904, + 0.025466615334153175, + 0.02467542327940464, + 0.010680009610950947, + 0.007619370240718126, + 0.0160642359405756, + 0.011403795331716537, + 0.011011609807610512, + 0.01931636966764927, + 0.0123280119150877, + 0.02390456385910511, + 0.020704852417111397, + 0.02978593111038208, + 0.01797131448984146, + 0.021430369466543198, + 0.01589789055287838, + 0.030720649287104607, + 0.01597917266190052, + 0.023191306740045547, + 0.004943952430039644, + 0.004767224658280611, + 0.0055409870110452175, + 0.008670631796121597, + 0.005177718121558428, + 0.008957200683653355, + 0.0053026205860078335, + 0.00763285206630826, + 0.0032433574087917805, + 0.002952944254502654, + 0.0022635033819824457, + 0.0047355168499052525, + 0.0036218625027686357, + 0.04393674433231354, + 0.0037967567332088947, + 0.04517635703086853, + 0.0033533794339746237, + 0.006042191758751869, + 0.003303670324385166, + 0.005426440387964249, + 0.003963045310229063, + 0.0062324898317456245, + 0.004826313816010952, + 0.005490069277584553, + 0.004916054662317038, + 0.007989042438566685, + 0.005242009647190571, + 0.009464363567531109, + 0.008566895499825478, + 0.008782844059169292, + 0.009578365832567215, + 0.006630620453506708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.03369441255927086 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741044111549854 + ], + "opposite_logprob_tokens": [ + -5.006741046905518 + ], + "logprob_reward": -0.006741044111549854, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741044111549854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10030929744243622, + 0.04381699487566948, + 0.06194979324936867, + 0.1376846730709076, + 0.15023323893547058, + 0.0743127167224884, + 0.04477168619632721, + 0.2655033767223358, + 0.030688075348734856, + 0.09073018282651901 + ], + "credit_map": [ + 0.012071018107235432, + 0.012613721191883087, + 0.010005712509155273, + 0.015486309304833412, + 0.00839953776448965, + 0.010234982706606388, + 0.018161317333579063, + 0.013336694799363613, + 0.0042173126712441444, + 0.0033552709501236677, + 0.008002933114767075, + 0.006293245125561953, + 0.0068259211257100105, + 0.00428336625918746, + 0.005544622894376516, + 0.005294322036206722, + 0.004601004999130964, + 0.006241310387849808, + 0.01096194889396429, + 0.009065022692084312, + 0.009279876947402954, + 0.006830369587987661, + 0.00921226479113102, + 0.005757997743785381, + 0.016469260677695274, + 0.009327900595963001, + 0.021994294598698616, + 0.015904083847999573, + 0.015328128822147846, + 0.01542205736041069, + 0.020139923319220543, + 0.02309902012348175, + 0.009933949448168278, + 0.006688401103019714, + 0.013579451479017735, + 0.00849170796573162, + 0.013136040419340134, + 0.04381047189235687, + 0.00857487041503191, + 0.046018343418836594, + 0.008812147192656994, + 0.007445933762937784, + 0.007106984034180641, + 0.007074419409036636, + 0.011373784393072128, + 0.01225611474364996, + 0.010519151575863361, + 0.00972417276352644, + 0.003838947042822838, + 0.004584391601383686, + 0.006501501891762018, + 0.0059079439379274845, + 0.0038836728781461716, + 0.0067590950056910515, + 0.007099529262632132, + 0.006196608301252127, + 0.007023727521300316, + 0.003295263508334756, + 0.0037157924380153418, + 0.00463059451431036, + 0.008675734512507915, + 0.11439494788646698, + 0.0052831438370049, + 0.11848417669534683, + 0.004255751613527536, + 0.002380538498982787, + 0.003743278095498681, + 0.005898284260183573, + 0.002100882586091757, + 0.0037965569645166397, + 0.004576919600367546, + 0.00393586466088891, + 0.00599696347489953, + 0.010062934830784798, + 0.006579769775271416, + 0.01690615713596344, + 0.01269312109798193, + 0.014225146733224392, + 0.013817977160215378, + 0.01044811587780714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.006741044111549854 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263367909938097 + ], + "opposite_logprob_tokens": [ + -3.651336908340454 + ], + "logprob_reward": -0.0263367909938097, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263367909938097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049867965281009674, + 0.08698871731758118, + 0.06972444802522659, + 0.03514399379491806, + 0.10139638185501099, + 0.09484212100505829, + 0.05459752306342125, + 0.4398588538169861, + 0.025439921766519547, + 0.04214000329375267 + ], + "credit_map": [ + 0.00696205347776413, + 0.0064127398654818535, + 0.0035533015616238117, + 0.005304873920977116, + 0.007913879118859768, + 0.004848099313676357, + 0.00895778276026249, + 0.005915239918977022, + 0.006598740350455046, + 0.0050505949184298515, + 0.024925611913204193, + 0.01651262305676937, + 0.0077537656761705875, + 0.016561930999159813, + 0.004899108316749334, + 0.0046863555908203125, + 0.0035740267485380173, + 0.004990449640899897, + 0.011708498932421207, + 0.010898652486503124, + 0.00973999872803688, + 0.007492422126233578, + 0.012992470525205135, + 0.008327936753630638, + 0.0025717278476804495, + 0.005490773357450962, + 0.005129953846335411, + 0.004501403775066137, + 0.0032785108778625727, + 0.0031202835962176323, + 0.006379785481840372, + 0.004671557806432247, + 0.006065079476684332, + 0.004839120898395777, + 0.005474514327943325, + 0.004581107292324305, + 0.005641275551170111, + 0.03469594568014145, + 0.004187458660453558, + 0.03591189160943031, + 0.012226441875100136, + 0.014275464229285717, + 0.010479499585926533, + 0.01225564256310463, + 0.008285630494356155, + 0.016768470406532288, + 0.009401161223649979, + 0.011149824596941471, + 0.004884596448391676, + 0.004572581499814987, + 0.006884842179715633, + 0.01239131297916174, + 0.004421274643391371, + 0.008304218761622906, + 0.0052431202493608, + 0.007895581424236298, + 0.01089301984757185, + 0.004780647810548544, + 0.002599279163405299, + 0.006176964845508337, + 0.01304176077246666, + 0.1948181688785553, + 0.006577952764928341, + 0.2009711116552353, + 0.0020050874445587397, + 0.0018998972373083234, + 0.0028536655008792877, + 0.004861489869654179, + 0.0028131252620369196, + 0.0037689553573727608, + 0.0032939149532467127, + 0.0039437878876924515, + 0.0027794018387794495, + 0.005226573906838894, + 0.0034712536726146936, + 0.006431311834603548, + 0.006251687183976173, + 0.006420152727514505, + 0.006190177518874407, + 0.0053694481030106544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.0263367909938097 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "first-person, subjective shot completely immerses the viewer in a magnificent Gothic hall. The camera tilts upward from a level angle, revealing the soaring ribbed vaults, their intricate stonework looming in the gloom. Countless magical paper airplanes, emitting a faint white glow, soar gracefully beneath the vast dome, like birds imbued with life, circling and weaving in a harmonious and ordered rhythm, leaving fleeting golden trails in their wake. The view slowly returns to the front, focusing on an elder at a distant podium. He possesses shoulder-length silver-white hair and a kind face lined with lines of wisdom. He wears a solemn, dark blue velvet robe, the edges of which are delicately embroidered with silver stars. He preaches from behind an ancient, dark oak lectern, his hands resting steadily on the edge, his body leaning slightly forward. His expression is focused and calm, as if his gentle yet resolute voice echoes silently throughout the room. Rich beams of golden sunlight penetrate the massive stained-glass windows overhead, casting the vibrant colors of the mythological scene onto the ancient stone floor. A soft halo of light forms behind the elder, framing him with a sacred and warm silhouette. The entire scene is enveloped in a tranquil, solemn, and hopeful atmosphere, creating a profound sense of warmth and healing.", + "seed": 21939, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_2/index_3612_seed_21939_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03073473908007145, + "vlm_logprob": -0.03073473908007145, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.03073473908007145, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.046698931604623795 + ], + "opposite_logprob_tokens": [ + -3.171699047088623 + ], + "logprob_reward": -0.046698931604623795, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.046698931604623795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14674340188503265, + 0.07070270925760269, + 0.09337355196475983, + 0.19387152791023254, + 0.05372673645615578, + 0.08025742322206497, + 0.047158870846033096, + 0.12858694791793823, + 0.16498050093650818, + 0.020598335191607475 + ], + "credit_map": [ + 0.008990917354822159, + 0.013611393049359322, + 0.014388653449714184, + 0.021489735692739487, + 0.0210459865629673, + 0.011874538846313953, + 0.033038731664419174, + 0.022303448989987373, + 0.005856356583535671, + 0.005955738481134176, + 0.005640449468046427, + 0.010173363611102104, + 0.010795149020850658, + 0.01355869509279728, + 0.008453563787043095, + 0.01026939321309328, + 0.013878564350306988, + 0.017899153754115105, + 0.008320586755871773, + 0.009474804624915123, + 0.010332977399230003, + 0.01126457192003727, + 0.009631005115807056, + 0.012571891769766808, + 0.016738226637244225, + 0.04047110304236412, + 0.012752954848110676, + 0.021338511258363724, + 0.04486208036541939, + 0.031122250482439995, + 0.011936771683394909, + 0.014649643562734127, + 0.0087742293253541, + 0.005821259692311287, + 0.007098667789250612, + 0.004905818495899439, + 0.007411190774291754, + 0.011710206046700478, + 0.003865887876600027, + 0.004139476455748081, + 0.00876355916261673, + 0.009187685325741768, + 0.01677430048584938, + 0.008929144591093063, + 0.010606334544718266, + 0.013561473228037357, + 0.007896571420133114, + 0.004538351204246283, + 0.0036213251296430826, + 0.0033760052174329758, + 0.00675353454425931, + 0.007257173303514719, + 0.008195631206035614, + 0.00694803474470973, + 0.004866921342909336, + 0.0061402455903589725, + 0.015858182683587074, + 0.01094767451286316, + 0.014213023707270622, + 0.01333085261285305, + 0.02541467361152172, + 0.022756081074476242, + 0.010412638075649738, + 0.015653828158974648, + 0.020450463518500328, + 0.01269964873790741, + 0.02034573443233967, + 0.01791568472981453, + 0.014777285978198051, + 0.02105100080370903, + 0.026968101039528847, + 0.030772585421800613, + 0.00363500346429646, + 0.0029541596304625273, + 0.0027349700685590506, + 0.002247224096208811, + 0.0020559225231409073, + 0.0033996631391346455, + 0.0018167565576732159, + 0.0017546360613778234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.046698931604623795 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370628505945206 + ], + "opposite_logprob_tokens": [ + -3.4087061882019043 + ], + "logprob_reward": -0.03370628505945206, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370628505945206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10467927902936935, + 0.06473249942064285, + 0.07648975402116776, + 0.19905847311019897, + 0.05907319858670235, + 0.11012120544910431, + 0.11855138838291168, + 0.09475991874933243, + 0.13080108165740967, + 0.04173323139548302 + ], + "credit_map": [ + 0.007167722098529339, + 0.00967479683458805, + 0.012158489786088467, + 0.021573392674326897, + 0.013916556723415852, + 0.0083975400775671, + 0.017217889428138733, + 0.01457288209348917, + 0.004691860172897577, + 0.007119572255760431, + 0.005179992411285639, + 0.007752751000225544, + 0.012542087584733963, + 0.008534246124327183, + 0.00954833347350359, + 0.009363655932247639, + 0.008855517953634262, + 0.007763401139527559, + 0.005259956698864698, + 0.005846730899065733, + 0.01132819801568985, + 0.009304990991950035, + 0.008282951079308987, + 0.01984800212085247, + 0.019584299996495247, + 0.02925577573478222, + 0.02412140741944313, + 0.017313752323389053, + 0.054251428693532944, + 0.029171092435717583, + 0.012128935195505619, + 0.013231774792075157, + 0.008524312637746334, + 0.00893537886440754, + 0.00814108643680811, + 0.004100878722965717, + 0.00873621553182602, + 0.011294511146843433, + 0.004068195354193449, + 0.00527261383831501, + 0.00706914346665144, + 0.008687745779752731, + 0.017731420695781708, + 0.011838216334581375, + 0.026501039043068886, + 0.02138395421206951, + 0.011241934262216091, + 0.005667736753821373, + 0.009145038202404976, + 0.011437551118433475, + 0.01663646474480629, + 0.016014287248253822, + 0.020002581179142, + 0.01689554564654827, + 0.013752646744251251, + 0.014667252078652382, + 0.009702898561954498, + 0.008511211723089218, + 0.01484408974647522, + 0.009326779283583164, + 0.01947878673672676, + 0.011146505363285542, + 0.010473283939063549, + 0.011276363395154476, + 0.01639336347579956, + 0.01049624104052782, + 0.01684981770813465, + 0.014101971872150898, + 0.012463672086596489, + 0.01627882942557335, + 0.021236304193735123, + 0.022980881854891777, + 0.008241306990385056, + 0.004255628678947687, + 0.005392803810536861, + 0.005285935010761023, + 0.003987508360296488, + 0.007820269092917442, + 0.003203006461262703, + 0.0035467687994241714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.03370628505945206 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340970769524574 + ], + "opposite_logprob_tokens": [ + -3.651340961456299 + ], + "logprob_reward": -0.026340970769524574, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340970769524574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0641091912984848, + 0.032953519374132156, + 0.024516034871339798, + 0.06416727602481842, + 0.09022170305252075, + 0.10227903723716736, + 0.06581982970237732, + 0.16772156953811646, + 0.2938481271266937, + 0.09436367452144623 + ], + "credit_map": [ + 0.006541816052049398, + 0.00572842825204134, + 0.008619503118097782, + 0.014427270740270615, + 0.006800081115216017, + 0.005123247858136892, + 0.007814088836312294, + 0.009054763242602348, + 0.0028588708955794573, + 0.002241338836029172, + 0.002942910883575678, + 0.003470534924417734, + 0.002898313570767641, + 0.005769970361143351, + 0.005110701080411673, + 0.0076608806848526, + 0.0029939976520836353, + 0.0031282762065529823, + 0.0021916222758591175, + 0.0027021639980375767, + 0.003133172169327736, + 0.0033253009896725416, + 0.002668555825948715, + 0.004372948780655861, + 0.007312018424272537, + 0.011496864259243011, + 0.003750946605578065, + 0.008581211790442467, + 0.01234311144798994, + 0.013240252621471882, + 0.003657618770375848, + 0.0037852551322430372, + 0.01737808808684349, + 0.010547537356615067, + 0.011384244076907635, + 0.008267856203019619, + 0.012601935304701328, + 0.01845669187605381, + 0.005713606718927622, + 0.00587175227701664, + 0.011882076971232891, + 0.011006799526512623, + 0.019236838445067406, + 0.01186916884034872, + 0.012633097358047962, + 0.013597440905869007, + 0.01528516598045826, + 0.006768450140953064, + 0.010888125747442245, + 0.005510479211807251, + 0.010290571488440037, + 0.009040342643857002, + 0.009534996002912521, + 0.007491440046578646, + 0.0072042313404381275, + 0.0058596464805305, + 0.01560793537646532, + 0.018303444609045982, + 0.02786899544298649, + 0.02040819451212883, + 0.02944125421345234, + 0.015833843499422073, + 0.020138941705226898, + 0.020118972286581993, + 0.03690679743885994, + 0.03134312108159065, + 0.02801572158932686, + 0.03320765867829323, + 0.03176237642765045, + 0.03733741492033005, + 0.04486222192645073, + 0.05041281506419182, + 0.01434675045311451, + 0.01226684171706438, + 0.010630124248564243, + 0.013720457442104816, + 0.012128869071602821, + 0.013838579878211021, + 0.008773097768425941, + 0.008658960461616516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.026340970769524574 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020591065287590027 + ], + "opposite_logprob_tokens": [ + -3.8955910205841064 + ], + "logprob_reward": -0.020591065287590027, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020591065287590027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08966069668531418, + 0.10074659436941147, + 0.05992772802710533, + 0.10079444944858551, + 0.0960865393280983, + 0.06442619860172272, + 0.07065410912036896, + 0.12310180068016052, + 0.27622315287590027, + 0.018378689885139465 + ], + "credit_map": [ + 0.010272522456943989, + 0.010736748576164246, + 0.0152235496789217, + 0.014164093881845474, + 0.009552428498864174, + 0.009434917941689491, + 0.011676700785756111, + 0.008599728345870972, + 0.008309445343911648, + 0.006874020677059889, + 0.010690675117075443, + 0.014558165334165096, + 0.013102306053042412, + 0.01570434309542179, + 0.016121933236718178, + 0.015385702252388, + 0.007942872121930122, + 0.009991807863116264, + 0.005566687788814306, + 0.00552833266556263, + 0.006604970898479223, + 0.007335110567510128, + 0.006089724600315094, + 0.010868225246667862, + 0.009971316903829575, + 0.015680860728025436, + 0.012380301021039486, + 0.01812940090894699, + 0.014432930387556553, + 0.01345007587224245, + 0.0070467861369252205, + 0.009702778421342373, + 0.015429694205522537, + 0.01121906191110611, + 0.010924615897238255, + 0.00906949769705534, + 0.012935602106153965, + 0.0211738720536232, + 0.008130486123263836, + 0.007203713059425354, + 0.004764548037201166, + 0.004546923562884331, + 0.010648432187736034, + 0.005940751638263464, + 0.01368577592074871, + 0.012516473419964314, + 0.008982975035905838, + 0.0033403197303414345, + 0.004185907542705536, + 0.005113077349960804, + 0.009843549691140652, + 0.008491900749504566, + 0.016809720546007156, + 0.008532836101949215, + 0.008331441320478916, + 0.00934568140655756, + 0.015152771025896072, + 0.010511977598071098, + 0.01911485567688942, + 0.013450280763208866, + 0.023279177024960518, + 0.014299362897872925, + 0.012279072776436806, + 0.015014302916824818, + 0.033493075519800186, + 0.02086557075381279, + 0.03584548085927963, + 0.03078281879425049, + 0.025768844410777092, + 0.03473157435655594, + 0.044169023633003235, + 0.05056678131222725, + 0.002704818267375231, + 0.00223374436609447, + 0.0019594726618379354, + 0.002496283035725355, + 0.0021468293853104115, + 0.0035771105904132128, + 0.00180080346763134, + 0.0014596275286749005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.020591065287590027 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a first-person perspective that tilts upward from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless magical paper airplanes emitting a faint white glow visible soaring beneath a vast dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elder with shoulder-length silver-white hair and a dark blue velvet robe embroidered with silver stars visible at a distant podium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rich beams of golden sunlight penetrating stained-glass windows and casting vibrant colors onto the ancient stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view slowly return to the front to focus on the elder after the initial upward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026336442679166794 + ], + "opposite_logprob_tokens": [ + -3.651336431503296 + ], + "logprob_reward": -0.026336442679166794, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026336442679166794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049594081938266754, + 0.08312676846981049, + 0.07599351555109024, + 0.111838199198246, + 0.0928800031542778, + 0.07505080848932266, + 0.05343731492757797, + 0.10977403074502945, + 0.2511877417564392, + 0.0971175879240036 + ], + "credit_map": [ + 0.0042357309721410275, + 0.006090054288506508, + 0.006640415173023939, + 0.009127392433583736, + 0.005706312600523233, + 0.0047660088166594505, + 0.006895077880471945, + 0.006133086979389191, + 0.005807519890367985, + 0.004579768981784582, + 0.008003517985343933, + 0.007872449234127998, + 0.013810033909976482, + 0.016132522374391556, + 0.011918029747903347, + 0.015002917498350143, + 0.014884427189826965, + 0.009355220012366772, + 0.007093987427651882, + 0.008256013505160809, + 0.006469768472015858, + 0.012253127992153168, + 0.0089194281026721, + 0.008761540055274963, + 0.011372526176273823, + 0.021145600825548172, + 0.011933008208870888, + 0.015545599162578583, + 0.018112502992153168, + 0.014459722675383091, + 0.007580944802612066, + 0.01168828271329403, + 0.017578039318323135, + 0.015549269504845142, + 0.013505501672625542, + 0.0074537391774356365, + 0.013431175611913204, + 0.014034975320100784, + 0.0058470312505960464, + 0.0054802619852125645, + 0.006613452918827534, + 0.008522105403244495, + 0.012945132330060005, + 0.011936021037399769, + 0.010911820456385612, + 0.009250927716493607, + 0.010378774255514145, + 0.004492565058171749, + 0.006188477389514446, + 0.006147404666990042, + 0.006428951397538185, + 0.0059391711838543415, + 0.009700274094939232, + 0.008139104582369328, + 0.00536749791353941, + 0.00552642997354269, + 0.015563069842755795, + 0.008665385656058788, + 0.014846366830170155, + 0.012671067379415035, + 0.018800852820277214, + 0.0134186502546072, + 0.011882263235747814, + 0.013926361687481403, + 0.03042171336710453, + 0.02252151072025299, + 0.02870061621069908, + 0.0280607920140028, + 0.02444336749613285, + 0.031785234808921814, + 0.0396432988345623, + 0.04561116173863411, + 0.017765654250979424, + 0.00813255738466978, + 0.008252395316958427, + 0.010823613964021206, + 0.011216887272894382, + 0.028804512694478035, + 0.005127552896738052, + 0.006994408089667559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.026336442679166794 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_3/index_tensor([9614, 9614])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_3/index_tensor([9614, 9614])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1af4d47bafadf17bc383fe0c2bc28e3f4cc12a0d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_3/index_tensor([9614, 9614])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a dramatic aerial drone shot, swooping down at breakneck speed through a sky shrouded in thick, dark clouds before stabilizing at a low, upward-looking angle, focusing on the center of a vast expanse of ancient battlefield ruins. The ground is cracked, dry earth, littered with broken stone pillars and crumbling walls. A knight clad in dark silver plate armor takes center stage. His armor is deeply scarred and battle-scarred, its pauldrons intricately carved with griffins, and its helmet adorned with faded red plumes. The metal reflects the cold light filtering through gaps in the clouds. In the center of the ruins, he tenses his legs and leaps from the ground, his heavy body arcing in a powerful arc as he reaches his highest point. Then, with all his might, he grasps a massive two-handed sword, nearly his height, engraved with ancient runes, and strikes down with thunderous force. As the tip of the giant sword slammed heavily on the ground, a powerful shock wave instantly erupted with the tip of the sword as the center, stirring up countless gravel, sand and thick dust, forming a circle of dust cloud that spread outward. The entire picture presented a highly detailed, top-level game CG style full of motion blur and particle effects, creating an epic and tension-filled battle atmosphere.", + "seed": 30110, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_3/index_9614_seed_30110_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.017421833891421556, + "vlm_logprob": -0.017421833891421556, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.017421833891421556, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04501086473464966 + ], + "opposite_logprob_tokens": [ + -3.170010805130005 + ], + "logprob_reward": -0.04501086473464966, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04501086473464966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7480418682098389, + 0.034009940922260284, + 0.0526154562830925, + 0.02649848721921444, + 0.03359962999820709, + 0.02767099440097809, + 0.01375527773052454, + 0.01610512286424637, + 0.030728943645954132, + 0.016974283382296562 + ], + "credit_map": [ + 0.03931393101811409, + 0.03782923147082329, + 0.062106672674417496, + 0.01990543305873871, + 0.022374730557203293, + 0.30849114060401917, + 0.019758757203817368, + 0.23826192319393158, + 0.005384311079978943, + 0.0048941271379590034, + 0.006122560240328312, + 0.004684345796704292, + 0.00514949532225728, + 0.0027330860029906034, + 0.003235897747799754, + 0.0018061107257381082, + 0.005964591167867184, + 0.007563322316855192, + 0.005761005450040102, + 0.005966137163341045, + 0.010202724486589432, + 0.005473775323480368, + 0.007445031311362982, + 0.004238867200911045, + 0.0027223725337535143, + 0.0037895231507718563, + 0.003827616572380066, + 0.002819629153236747, + 0.002763500204309821, + 0.0031684807036072016, + 0.003616407047957182, + 0.0037909545935690403, + 0.0057480898685753345, + 0.003135721432045102, + 0.00636383518576622, + 0.002362609375268221, + 0.0036033219657838345, + 0.003389340825378895, + 0.0018615410663187504, + 0.007135168183594942, + 0.002897344995290041, + 0.0014042418915778399, + 0.008227206766605377, + 0.004068007692694664, + 0.0018614010186865926, + 0.003389112651348114, + 0.002002461813390255, + 0.0038212148938328028, + 0.001854481641203165, + 0.0010487892432138324, + 0.00302827637642622, + 0.002046232344582677, + 0.0007802833570167422, + 0.0020338494796305895, + 0.0012404322624206543, + 0.0017229331424459815, + 0.0015923904720693827, + 0.0010280837304890156, + 0.0019853406120091677, + 0.0032609927002340555, + 0.0012681951047852635, + 0.0017608777852728963, + 0.001753882272168994, + 0.0034553592558950186, + 0.004177361726760864, + 0.002114837756380439, + 0.006710076238960028, + 0.0027913262601941824, + 0.004051554016768932, + 0.0032663613092154264, + 0.004743740428239107, + 0.002873681951314211, + 0.0013768058270215988, + 0.004441662225872278, + 0.0016184698324650526, + 0.0016134638572111726, + 0.0018092229729518294, + 0.002952695358544588, + 0.0016696093371137977, + 0.001492352457717061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04501086473464966 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636639289557934 + ], + "opposite_logprob_tokens": [ + -4.882636547088623 + ], + "logprob_reward": -0.007636639289557934, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636639289557934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4941738247871399, + 0.04261048883199692, + 0.20004700124263763, + 0.03100162371993065, + 0.02699706144630909, + 0.03768526762723923, + 0.04950344190001488, + 0.05702448636293411, + 0.03164231404662132, + 0.02931441366672516 + ], + "credit_map": [ + 0.024192193523049355, + 0.022724993526935577, + 0.03804723173379898, + 0.011792208068072796, + 0.014580868184566498, + 0.20690302550792694, + 0.01316776592284441, + 0.16276559233665466, + 0.007287819869816303, + 0.006544449366629124, + 0.007331806235015392, + 0.003890403313562274, + 0.007643424440175295, + 0.0037359304260462523, + 0.0037623639218509197, + 0.002414294518530369, + 0.017278755083680153, + 0.023211926221847534, + 0.015179253183305264, + 0.017311906442046165, + 0.062074072659015656, + 0.026307426393032074, + 0.0203764159232378, + 0.018307246267795563, + 0.0036773360334336758, + 0.0036600613966584206, + 0.0050977556966245174, + 0.003132200799882412, + 0.004976961761713028, + 0.0035333032719790936, + 0.0025645503774285316, + 0.00435945950448513, + 0.004876935388892889, + 0.0018499420257285237, + 0.004391519818454981, + 0.0026962768752127886, + 0.0033800355158746243, + 0.0023490660823881626, + 0.0016737767728045583, + 0.005779511760920286, + 0.007318131159991026, + 0.003039563773199916, + 0.0059082163497805595, + 0.0033623811323195696, + 0.004118388053029776, + 0.00619164202362299, + 0.003109656972810626, + 0.004637295380234718, + 0.003959991037845612, + 0.003327684011310339, + 0.007053594570606947, + 0.010390152223408222, + 0.006196336820721626, + 0.003230348229408264, + 0.00892519298940897, + 0.00642014667391777, + 0.005001607351005077, + 0.005040043033659458, + 0.007057730574160814, + 0.012743327766656876, + 0.00630686990916729, + 0.00473218597471714, + 0.00955174583941698, + 0.006590978242456913, + 0.002107863547280431, + 0.00270983949303627, + 0.004759581293910742, + 0.004069177899509668, + 0.004697796888649464, + 0.004372297320514917, + 0.005629442166537046, + 0.0032963163685053587, + 0.0028887626249343157, + 0.006191049236804247, + 0.003075882326811552, + 0.002731821034103632, + 0.0036894469521939754, + 0.0037659117951989174, + 0.0030549729708582163, + 0.003916569519788027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.007636639289557934 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057611756026745 + ], + "opposite_logprob_tokens": [ + -3.895576000213623 + ], + "logprob_reward": -0.02057611756026745, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057611756026745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4536762535572052, + 0.0676991418004036, + 0.05912483483552933, + 0.06187873333692551, + 0.07941939681768417, + 0.050506554543972015, + 0.05926619470119476, + 0.06036917492747307, + 0.05168634280562401, + 0.05637335032224655 + ], + "credit_map": [ + 0.03436071798205376, + 0.02363857813179493, + 0.030925936996936798, + 0.012247049249708652, + 0.017719702795147896, + 0.1850571483373642, + 0.015262342058122158, + 0.1344647854566574, + 0.007221129722893238, + 0.008534396067261696, + 0.012043113820254803, + 0.005758536048233509, + 0.018809836357831955, + 0.005903879180550575, + 0.006483487784862518, + 0.0029447602573782206, + 0.005302269943058491, + 0.007453343831002712, + 0.005726570263504982, + 0.005903956014662981, + 0.014579858630895615, + 0.007346574682742357, + 0.008362749591469765, + 0.004449513275176287, + 0.006009408738464117, + 0.007309694774448872, + 0.006606384180486202, + 0.0041864411905407906, + 0.008907309733331203, + 0.007806519512087107, + 0.007345974445343018, + 0.013707002624869347, + 0.010049714706838131, + 0.010644655674695969, + 0.010095254518091679, + 0.007524494081735611, + 0.013256169855594635, + 0.008162934333086014, + 0.008697755634784698, + 0.01098842453211546, + 0.005872875452041626, + 0.007127075921744108, + 0.006937059573829174, + 0.006072704680263996, + 0.00721888430416584, + 0.004692500457167625, + 0.006319976877421141, + 0.006265477742999792, + 0.004904433619230986, + 0.0037906712386757135, + 0.00853433832526207, + 0.010952322743833065, + 0.006399116478860378, + 0.004811840131878853, + 0.012565475888550282, + 0.007307996973395348, + 0.00716855563223362, + 0.0057808468118309975, + 0.007966096512973309, + 0.007998685352504253, + 0.00958204921334982, + 0.0052115339785814285, + 0.00900069996714592, + 0.00766070606186986, + 0.004890604875981808, + 0.004733762703835964, + 0.00486032897606492, + 0.006048336625099182, + 0.010798094794154167, + 0.007365622092038393, + 0.00770873948931694, + 0.005280853249132633, + 0.004798165522515774, + 0.008395770564675331, + 0.009020734578371048, + 0.007721705362200737, + 0.005324711091816425, + 0.007708132266998291, + 0.005052486434578896, + 0.008351645432412624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.02057611756026745 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00524667976424098 + ], + "opposite_logprob_tokens": [ + -5.255246639251709 + ], + "logprob_reward": -0.00524667976424098, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00524667976424098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3625486493110657, + 0.05194497108459473, + 0.0472700260579586, + 0.02772875875234604, + 0.05826060473918915, + 0.04464435577392578, + 0.05753384158015251, + 0.06596755236387253, + 0.1453186273574829, + 0.13878263533115387 + ], + "credit_map": [ + 0.027486490085721016, + 0.020431868731975555, + 0.02665903978049755, + 0.010458124801516533, + 0.011777241714298725, + 0.14848260581493378, + 0.010472521185874939, + 0.10678074508905411, + 0.007316414266824722, + 0.008347906172275543, + 0.015805521979928017, + 0.00413869833573699, + 0.005726941861212254, + 0.0036264804657548666, + 0.004398593213409185, + 0.0025844078045338392, + 0.005140714813023806, + 0.004725706297904253, + 0.004536123480647802, + 0.00434427335858345, + 0.013503916561603546, + 0.006458846852183342, + 0.004745415411889553, + 0.003815023461356759, + 0.002950281137600541, + 0.003285857615992427, + 0.00505068339407444, + 0.002066894667223096, + 0.004825083538889885, + 0.0036592455580830574, + 0.002825462957844138, + 0.0030652470886707306, + 0.009232263080775738, + 0.005831778049468994, + 0.01038448791950941, + 0.004553992301225662, + 0.006841730792075396, + 0.0056509096175432205, + 0.0043077305890619755, + 0.011457711458206177, + 0.008234200067818165, + 0.003810103517025709, + 0.00461380323395133, + 0.004845352377742529, + 0.004497801419347525, + 0.009378734976053238, + 0.004282742273062468, + 0.004981613252311945, + 0.004987176042050123, + 0.003607925958931446, + 0.009013686329126358, + 0.009242184460163116, + 0.00773084070533514, + 0.003712543984875083, + 0.011255957186222076, + 0.007983526214957237, + 0.008373070508241653, + 0.005291423294693232, + 0.005955772940069437, + 0.0113396430388093, + 0.004667566157877445, + 0.005985765717923641, + 0.009679855778813362, + 0.014674448408186436, + 0.02320326305925846, + 0.010269650258123875, + 0.03505044803023338, + 0.016574610024690628, + 0.011183299124240875, + 0.01593206264078617, + 0.019707180559635162, + 0.01339809875935316, + 0.01323380321264267, + 0.02781306952238083, + 0.015551329590380192, + 0.02031300775706768, + 0.012133298441767693, + 0.018725313246250153, + 0.014909415505826473, + 0.016103386878967285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.00524667976424098 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638868108391762 + ], + "opposite_logprob_tokens": [ + -4.758638858795166 + ], + "logprob_reward": -0.008638868108391762, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638868108391762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6596695184707642, + 0.040642641484737396, + 0.03314302861690521, + 0.07262314110994339, + 0.023681070655584335, + 0.0401882529258728, + 0.039426255971193314, + 0.025664938613772392, + 0.02856125868856907, + 0.036399926990270615 + ], + "credit_map": [ + 0.040318094193935394, + 0.02080588787794113, + 0.047359783202409744, + 0.0111645283177495, + 0.01870541274547577, + 0.2918052673339844, + 0.013004649430513382, + 0.21650587022304535, + 0.007054294925183058, + 0.006551897618919611, + 0.005806136876344681, + 0.0041834586299955845, + 0.006252684164792299, + 0.003753324970602989, + 0.0040372698567807674, + 0.0030035744421184063, + 0.0036326509434729815, + 0.003935310058295727, + 0.0026825263630598783, + 0.003179107327014208, + 0.007866812869906425, + 0.003936681430786848, + 0.0056627849116921425, + 0.0022471551783382893, + 0.006534476298838854, + 0.006810969673097134, + 0.007466029841452837, + 0.005093507003039122, + 0.01963617652654648, + 0.01020363625138998, + 0.00592677341774106, + 0.010951570235192776, + 0.003945763222873211, + 0.0018958853324875236, + 0.003166841808706522, + 0.0021023990120738745, + 0.0037609620485454798, + 0.0023522425908595324, + 0.0015164166688919067, + 0.004940559156239033, + 0.006759959738701582, + 0.0033582241740077734, + 0.005215133540332317, + 0.003564076730981469, + 0.004151724278926849, + 0.007492444012314081, + 0.0041496786288917065, + 0.00549701415002346, + 0.002936175325885415, + 0.0022557624615728855, + 0.007205785717815161, + 0.009844566695392132, + 0.0035005107056349516, + 0.002272088546305895, + 0.006177264265716076, + 0.0052341013215482235, + 0.002943785861134529, + 0.001957429340109229, + 0.0035379130858927965, + 0.0037326752208173275, + 0.0020404832903295755, + 0.0031920140609145164, + 0.004461234901100397, + 0.0037994026206433773, + 0.002357447287067771, + 0.002698898548260331, + 0.003597510978579521, + 0.0037378533743321896, + 0.0050569865852594376, + 0.003884596284478903, + 0.003961043432354927, + 0.0032669210340827703, + 0.0028947070240974426, + 0.004862907808274031, + 0.0038880598731338978, + 0.0062372940592467785, + 0.0037377062253654003, + 0.005121559835970402, + 0.0036252604331821203, + 0.006032432429492474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.008638868108391762 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a dramatic aerial drone shot, swooping down at breakneck speed through a sky shrouded in thick, dark clouds before stabilizing at a low, upward-looking angle, focusing on the center of a vast expanse of ancient battlefield ruins. The ground is cracked, dry earth, littered with broken stone pillars and crumbling walls. A knight clad in dark silver plate armor takes center stage. His armor is deeply scarred and battle-scarred, its pauldrons intricately carved with griffins, and its helmet adorned with faded red plumes. The metal reflects the cold light filtering through gaps in the clouds. In the center of the ruins, he tenses his legs and leaps from the ground, his heavy body arcing in a powerful arc as he reaches his highest point. Then, with all his might, he grasps a massive two-handed sword, nearly his height, engraved with ancient runes, and strikes down with thunderous force. As the tip of the giant sword slammed heavily on the ground, a powerful shock wave instantly erupted with the tip of the sword as the center, stirring up countless gravel, sand and thick dust, forming a circle of dust cloud that spread outward. The entire picture presented a highly detailed, top-level game CG style full of motion blur and particle effects, creating an epic and tension-filled battle atmosphere.", + "seed": 30111, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_3/index_9614_seed_30111_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09329651384614408, + "vlm_logprob": -0.09329651384614408, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.09329651384614408, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4302840232849121 + ], + "opposite_logprob_tokens": [ + -1.055284023284912 + ], + "logprob_reward": -0.4302840232849121, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4302840232849121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08323825150728226, + 0.03723937273025513, + 0.1100255623459816, + 0.15980465710163116, + 0.03650243207812309, + 0.07033596932888031, + 0.18941831588745117, + 0.08350351452827454, + 0.20946571230888367, + 0.020466230809688568 + ], + "credit_map": [ + 0.0073980060406029224, + 0.009573396295309067, + 0.005069092381745577, + 0.009860056452453136, + 0.012419033795595169, + 0.010578226298093796, + 0.011871805414557457, + 0.016468634828925133, + 0.0032177597749978304, + 0.0067327506840229034, + 0.003286263905465603, + 0.003831394948065281, + 0.004458222538232803, + 0.004898840561509132, + 0.0037015890702605247, + 0.007112550083547831, + 0.009393932297825813, + 0.012242551892995834, + 0.014203445054590702, + 0.023019133135676384, + 0.011111916974186897, + 0.014147663488984108, + 0.013406416401267052, + 0.012500504031777382, + 0.015661876648664474, + 0.01939254440367222, + 0.028291044756770134, + 0.007188024930655956, + 0.020973866805434227, + 0.02684418112039566, + 0.010280431248247623, + 0.03117268905043602, + 0.0034340566489845514, + 0.004599031526595354, + 0.004665216896682978, + 0.004405100829899311, + 0.006391808390617371, + 0.004794426262378693, + 0.0049536870792508125, + 0.0032591044437140226, + 0.006146976724267006, + 0.008333210833370686, + 0.011347792111337185, + 0.00859619677066803, + 0.008720450103282928, + 0.008700213395059109, + 0.008090109564363956, + 0.010401018895208836, + 0.02944282814860344, + 0.02039257436990738, + 0.024959072470664978, + 0.023306993767619133, + 0.031027112156152725, + 0.026663314551115036, + 0.017452174797654152, + 0.01617424190044403, + 0.00816653948277235, + 0.009440063498914242, + 0.010156148113310337, + 0.013252552598714828, + 0.01116117276251316, + 0.008642653934657574, + 0.0105157271027565, + 0.01216865237802267, + 0.030811043456196785, + 0.03363354131579399, + 0.028884314000606537, + 0.028225578367710114, + 0.020391840487718582, + 0.028367239981889725, + 0.019734423607587814, + 0.019417742267251015, + 0.0012877839617431164, + 0.0018516331911087036, + 0.003154376056045294, + 0.005618121940642595, + 0.0021752361208200455, + 0.0015673761954531074, + 0.0016897703753784299, + 0.0031219334341585636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.4302840232849121 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979179609566927 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.00979179609566927, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979179609566927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09914307296276093, + 0.08555977046489716, + 0.03276728466153145, + 0.14828598499298096, + 0.06479167193174362, + 0.05578352510929108, + 0.12555475533008575, + 0.079631008207798, + 0.16051186621189117, + 0.14797106385231018 + ], + "credit_map": [ + 0.010152860544621944, + 0.009685324504971504, + 0.0075168730691075325, + 0.01203150488436222, + 0.01654326356947422, + 0.013654930517077446, + 0.013200446963310242, + 0.016357863321900368, + 0.014171987771987915, + 0.008693710900843143, + 0.00976079422980547, + 0.007600524928420782, + 0.009262170642614365, + 0.018900876864790916, + 0.006942261476069689, + 0.010227448306977749, + 0.0023724811617285013, + 0.002929176203906536, + 0.005130243953317404, + 0.004591049626469612, + 0.00403731781989336, + 0.004889333620667458, + 0.0026853533927351236, + 0.0061323256231844425, + 0.016086267307400703, + 0.021137064322829247, + 0.021953662857413292, + 0.008649940602481365, + 0.024935787543654442, + 0.02177940309047699, + 0.01316085085272789, + 0.02058299444615841, + 0.0038780104368925095, + 0.006791005376726389, + 0.007911182940006256, + 0.007147230673581362, + 0.017551088705658913, + 0.009428441524505615, + 0.0043706586584448814, + 0.007714054547250271, + 0.0055636512115597725, + 0.006381949409842491, + 0.008593413047492504, + 0.0061560459434986115, + 0.007897812873125076, + 0.007589071057736874, + 0.006239952519536018, + 0.007361629977822304, + 0.016758615151047707, + 0.013067320920526981, + 0.01663394458591938, + 0.01658807136118412, + 0.019511304795742035, + 0.018466869369149208, + 0.01233535073697567, + 0.012193286791443825, + 0.007118035573512316, + 0.009667030535638332, + 0.00920356810092926, + 0.013199457898736, + 0.012887351214885712, + 0.006981487851589918, + 0.011628830805420876, + 0.008945243433117867, + 0.024990025907754898, + 0.026129446923732758, + 0.022503169253468513, + 0.021335631608963013, + 0.015556998550891876, + 0.021905183792114258, + 0.013091426342725754, + 0.014999978244304657, + 0.008542375639081001, + 0.007950165309011936, + 0.028988147154450417, + 0.0636061578989029, + 0.004077706020325422, + 0.00423092395067215, + 0.007042238488793373, + 0.023533357307314873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.00979179609566927 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606144569814205 + ], + "opposite_logprob_tokens": [ + -4.141061305999756 + ], + "logprob_reward": -0.01606144569814205, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606144569814205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17782579362392426, + 0.07920356094837189, + 0.07060728222131729, + 0.053279899060726166, + 0.04950713738799095, + 0.04448164999485016, + 0.10463912039995193, + 0.11213463544845581, + 0.12151055037975311, + 0.18681037425994873 + ], + "credit_map": [ + 0.02709757164120674, + 0.015293402597308159, + 0.013025972060859203, + 0.021522942930459976, + 0.017992179840803146, + 0.03668155148625374, + 0.021742790937423706, + 0.024469375610351562, + 0.008484571240842342, + 0.01359956432133913, + 0.00873852614313364, + 0.009674986824393272, + 0.009521433152258396, + 0.010215124115347862, + 0.008077790960669518, + 0.0108915651217103, + 0.005948169156908989, + 0.006620461121201515, + 0.010049574077129364, + 0.00928243063390255, + 0.010987360030412674, + 0.010065952315926552, + 0.009321614168584347, + 0.008331716991961002, + 0.008987979963421822, + 0.00990898348391056, + 0.0064393593929708, + 0.004041911568492651, + 0.008935592137277126, + 0.005998921114951372, + 0.004211439285427332, + 0.004755714908242226, + 0.0038191669154912233, + 0.00610726373270154, + 0.006551798433065414, + 0.005108521319925785, + 0.0109793646261096, + 0.006820569280534983, + 0.004152506124228239, + 0.0059679425321519375, + 0.005527683068066835, + 0.005901940632611513, + 0.008506379090249538, + 0.005155807826668024, + 0.004432540386915207, + 0.0061684539541602135, + 0.0045082345604896545, + 0.004280609544366598, + 0.013413538224995136, + 0.012099393643438816, + 0.014888542704284191, + 0.015554863028228283, + 0.016508663073182106, + 0.012933623977005482, + 0.010184995830059052, + 0.00905549991875887, + 0.02054521068930626, + 0.01572016067802906, + 0.015048001892864704, + 0.012383205816149712, + 0.00998708326369524, + 0.016240429133176804, + 0.01310320571064949, + 0.009107334539294243, + 0.019723908975720406, + 0.018824158236384392, + 0.01580211892724037, + 0.014827192761003971, + 0.01252335961908102, + 0.01486483495682478, + 0.011404159478843212, + 0.013540823012590408, + 0.009776424616575241, + 0.01016976498067379, + 0.03489839658141136, + 0.0798238068819046, + 0.005591046530753374, + 0.005177067592740059, + 0.009102736599743366, + 0.0322711206972599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.01606144569814205 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730505730956793 + ], + "opposite_logprob_tokens": [ + -5.006730556488037 + ], + "logprob_reward": -0.006730505730956793, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730505730956793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09054536372423172, + 0.05185714364051819, + 0.2392035275697708, + 0.04022364690899849, + 0.022966479882597923, + 0.03988727927207947, + 0.019011622294783592, + 0.07923528552055359, + 0.022174227982759476, + 0.39489540457725525 + ], + "credit_map": [ + 0.012206392362713814, + 0.009940593503415585, + 0.006557297892868519, + 0.010403755120933056, + 0.010829668492078781, + 0.015843244269490242, + 0.011028897017240524, + 0.013735524378716946, + 0.006623391527682543, + 0.0065606869757175446, + 0.005247137043625116, + 0.003992615733295679, + 0.005315719172358513, + 0.010471650399267673, + 0.003867030842229724, + 0.009778918698430061, + 0.020202338695526123, + 0.025750696659088135, + 0.03573542460799217, + 0.051398586481809616, + 0.021148061379790306, + 0.03371959552168846, + 0.027873283252120018, + 0.023375580087304115, + 0.0033905673772096634, + 0.005004940088838339, + 0.005891038570553064, + 0.00196203775703907, + 0.00630606384947896, + 0.006858484353870153, + 0.0026145214214920998, + 0.008195997215807438, + 0.001838532742112875, + 0.002733390312641859, + 0.003419472835958004, + 0.002510790480300784, + 0.0028044346254318953, + 0.0026696003042161465, + 0.0018529982771724463, + 0.005137263331562281, + 0.004150593653321266, + 0.005765071604400873, + 0.006972613278776407, + 0.004516538232564926, + 0.005503498949110508, + 0.0044053043238818645, + 0.004452765453606844, + 0.004120897501707077, + 0.0022452035918831825, + 0.0018071216763928533, + 0.003330688923597336, + 0.0033341911621391773, + 0.0026791850104928017, + 0.002280108630657196, + 0.0016420696629211307, + 0.0016930548008531332, + 0.008539787493646145, + 0.009138544090092182, + 0.010123137384653091, + 0.013849781826138496, + 0.008423643186688423, + 0.0076484899036586285, + 0.00673156650736928, + 0.014780350029468536, + 0.002586625050753355, + 0.0018769849557429552, + 0.0034367649350315332, + 0.0029795304872095585, + 0.0031791666988283396, + 0.0018331410828977823, + 0.0031368345953524113, + 0.003145183203741908, + 0.020610585808753967, + 0.023901639506220818, + 0.06716737896203995, + 0.16760095953941345, + 0.011419069021940231, + 0.01173428725451231, + 0.018312670290470123, + 0.07414884120225906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.006730505730956793 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036147984210401773 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.0036147984210401773, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036147984210401773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1508817821741104, + 0.05800756439566612, + 0.14895528554916382, + 0.14559105038642883, + 0.050296567380428314, + 0.05347210913896561, + 0.09282077103853226, + 0.09055205434560776, + 0.1550467610359192, + 0.05437607318162918 + ], + "credit_map": [ + 0.014632737264037132, + 0.012644859962165356, + 0.00883684866130352, + 0.01623176969587803, + 0.024371420964598656, + 0.02894139289855957, + 0.018432684242725372, + 0.02679005265235901, + 0.007653405424207449, + 0.00697871670126915, + 0.00653868168592453, + 0.006698449142277241, + 0.006300691980868578, + 0.00917905569076538, + 0.007396745961159468, + 0.00726181548088789, + 0.01255705300718546, + 0.014513458125293255, + 0.024959614500403404, + 0.028768904507160187, + 0.01608915440738201, + 0.023318827152252197, + 0.01701204478740692, + 0.011736216023564339, + 0.014908895827829838, + 0.016286585479974747, + 0.02729116566479206, + 0.007646560203284025, + 0.012000858783721924, + 0.02581927552819252, + 0.010551050305366516, + 0.03108665905892849, + 0.0034274475183337927, + 0.005866596009582281, + 0.005857943091541529, + 0.0054863011464476585, + 0.011438239365816116, + 0.006847611628472805, + 0.004898059647530317, + 0.00647437060251832, + 0.006151523906737566, + 0.006819515489041805, + 0.009535720571875572, + 0.006435589864850044, + 0.006215869914740324, + 0.007393793668597937, + 0.00470745237544179, + 0.006212641019374132, + 0.013649296946823597, + 0.008805655874311924, + 0.012666985392570496, + 0.01061701774597168, + 0.016101635992527008, + 0.012684421613812447, + 0.01021132618188858, + 0.008084423840045929, + 0.01450355164706707, + 0.007986926473677158, + 0.010376373305916786, + 0.010031710378825665, + 0.011128113605082035, + 0.015035581775009632, + 0.008867844007909298, + 0.012621943838894367, + 0.017906127497553825, + 0.025334151461720467, + 0.024012086912989616, + 0.02327125519514084, + 0.017010046169161797, + 0.018021631985902786, + 0.015035651624202728, + 0.014455815777182579, + 0.0035588599275797606, + 0.00451640272513032, + 0.005568416323512793, + 0.01732112653553486, + 0.005875835660845041, + 0.004655112512409687, + 0.003437548642978072, + 0.009442773647606373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.0036147984210401773 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a dramatic aerial drone shot, swooping down at breakneck speed through a sky shrouded in thick, dark clouds before stabilizing at a low, upward-looking angle, focusing on the center of a vast expanse of ancient battlefield ruins. The ground is cracked, dry earth, littered with broken stone pillars and crumbling walls. A knight clad in dark silver plate armor takes center stage. His armor is deeply scarred and battle-scarred, its pauldrons intricately carved with griffins, and its helmet adorned with faded red plumes. The metal reflects the cold light filtering through gaps in the clouds. In the center of the ruins, he tenses his legs and leaps from the ground, his heavy body arcing in a powerful arc as he reaches his highest point. Then, with all his might, he grasps a massive two-handed sword, nearly his height, engraved with ancient runes, and strikes down with thunderous force. As the tip of the giant sword slammed heavily on the ground, a powerful shock wave instantly erupted with the tip of the sword as the center, stirring up countless gravel, sand and thick dust, forming a circle of dust cloud that spread outward. The entire picture presented a highly detailed, top-level game CG style full of motion blur and particle effects, creating an epic and tension-filled battle atmosphere.", + "seed": 30112, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_3/index_9614_seed_30112_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010502344276756049, + "vlm_logprob": -0.010502344276756049, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.010502344276756049, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027682071551680565 + ], + "opposite_logprob_tokens": [ + -3.652682065963745 + ], + "logprob_reward": -0.027682071551680565, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.027682071551680565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3606387674808502, + 0.12198758870363235, + 0.05217164009809494, + 0.08965647965669632, + 0.047924675047397614, + 0.028854595497250557, + 0.10323312878608704, + 0.045139677822589874, + 0.11187351495027542, + 0.038519930094480515 + ], + "credit_map": [ + 0.0206095390021801, + 0.06135474890470505, + 0.04976958408951759, + 0.04419777914881706, + 0.048870354890823364, + 0.046296536922454834, + 0.055480215698480606, + 0.03405999392271042, + 0.01282793190330267, + 0.00996509287506342, + 0.01493447832763195, + 0.00981687381863594, + 0.02711140178143978, + 0.016034746542572975, + 0.00847755465656519, + 0.022819507867097855, + 0.011569222435355186, + 0.007763689383864403, + 0.006360258441418409, + 0.00617863517254591, + 0.0051714167930185795, + 0.006103704683482647, + 0.0039542973972857, + 0.005070418585091829, + 0.013207967393100262, + 0.012700260616838932, + 0.010188796557486057, + 0.008981899358332157, + 0.01337017584592104, + 0.014720999635756016, + 0.009918145835399628, + 0.0065682390704751015, + 0.006151206325739622, + 0.010241340845823288, + 0.005211197305470705, + 0.0035769580863416195, + 0.00792483426630497, + 0.00642902497202158, + 0.0039965142495930195, + 0.004393601790070534, + 0.0035988353192806244, + 0.002983508864417672, + 0.004361978732049465, + 0.004558923188596964, + 0.003580929711461067, + 0.0029310949612408876, + 0.0024655931629240513, + 0.0043737334199249744, + 0.016296356916427612, + 0.012292946688830853, + 0.019558526575565338, + 0.01866041123867035, + 0.00548541359603405, + 0.010382680222392082, + 0.005783650558441877, + 0.014773149974644184, + 0.006140195764601231, + 0.007310456596314907, + 0.0041815186850726604, + 0.004488070961087942, + 0.00544372433796525, + 0.006467740051448345, + 0.00570430513471365, + 0.005403667222708464, + 0.005246960557997227, + 0.005189922638237476, + 0.005741747561842203, + 0.007634938694536686, + 0.005776879843324423, + 0.040590550750494, + 0.009217827580869198, + 0.032474685460329056, + 0.004013890866190195, + 0.0036425222642719746, + 0.0033098862040787935, + 0.007968848571181297, + 0.003663112176582217, + 0.003912176936864853, + 0.0051721432246267796, + 0.0068373484537005424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.027682071551680565 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549126520752907 + ], + "opposite_logprob_tokens": [ + -4.387548923492432 + ], + "logprob_reward": -0.012549126520752907, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549126520752907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14502881467342377, + 0.08616591989994049, + 0.1461125910282135, + 0.13091091811656952, + 0.07892534136772156, + 0.07082847505807877, + 0.05299247428774834, + 0.032013989984989166, + 0.13916213810443878, + 0.11785931885242462 + ], + "credit_map": [ + 0.0089564835652709, + 0.016322415322065353, + 0.027883196249604225, + 0.01674402318894863, + 0.017532911151647568, + 0.022022301331162453, + 0.021684471517801285, + 0.013883024454116821, + 0.01105574332177639, + 0.0064315409399569035, + 0.010035129263997078, + 0.00871648732572794, + 0.01282585971057415, + 0.011517993174493313, + 0.008977196179330349, + 0.016605978831648827, + 0.02050633542239666, + 0.0188608281314373, + 0.01750856637954712, + 0.016991060227155685, + 0.021352943032979965, + 0.02148362062871456, + 0.014614078216254711, + 0.014795168302953243, + 0.012909134849905968, + 0.013972348533570766, + 0.012736620381474495, + 0.014905152842402458, + 0.02417105995118618, + 0.018222389742732048, + 0.0247059129178524, + 0.009288310073316097, + 0.010354486294090748, + 0.0167705025523901, + 0.008159471675753593, + 0.009523165412247181, + 0.011101125739514828, + 0.010373624041676521, + 0.006088052876293659, + 0.006554918363690376, + 0.007354159839451313, + 0.007474671583622694, + 0.013572328723967075, + 0.011990560218691826, + 0.009271372109651566, + 0.006551612168550491, + 0.007043395657092333, + 0.007570378948003054, + 0.008743871003389359, + 0.004652251955121756, + 0.005795808508992195, + 0.006499127019196749, + 0.008773175068199635, + 0.005089350510388613, + 0.006864745169878006, + 0.006574148777872324, + 0.0023534377105534077, + 0.003366334130987525, + 0.0041008200496435165, + 0.004157214891165495, + 0.004981957841664553, + 0.0035882231313735247, + 0.004448712803423405, + 0.005017291754484177, + 0.005895598325878382, + 0.00801650807261467, + 0.008398915641009808, + 0.008488559164106846, + 0.008028686046600342, + 0.04890652745962143, + 0.015735214576125145, + 0.03569214046001434, + 0.008741041645407677, + 0.007744139060378075, + 0.009889770299196243, + 0.018717819824814796, + 0.010906572453677654, + 0.021307218819856644, + 0.02835123799741268, + 0.012201525270938873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.012549126520752907 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008632604032754898 + ], + "opposite_logprob_tokens": [ + -4.758632659912109 + ], + "logprob_reward": -0.008632604032754898, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008632604032754898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10458050668239594, + 0.07181695848703384, + 0.0985228419303894, + 0.0714382529258728, + 0.06200965866446495, + 0.04935625195503235, + 0.0995655506849289, + 0.09123271703720093, + 0.13925129175186157, + 0.21222597360610962 + ], + "credit_map": [ + 0.008344143629074097, + 0.010509434156119823, + 0.01363045908510685, + 0.016156824305653572, + 0.01329069770872593, + 0.01758165843784809, + 0.01224957499653101, + 0.012817714363336563, + 0.012590069323778152, + 0.008208474144339561, + 0.01011478528380394, + 0.00874992087483406, + 0.007699429988861084, + 0.010473642498254776, + 0.004098060075193644, + 0.009882581420242786, + 0.010335457511246204, + 0.01350827980786562, + 0.012937765568494797, + 0.015835294499993324, + 0.010721245780587196, + 0.014557138085365295, + 0.009642892517149448, + 0.010984768159687519, + 0.00812519621104002, + 0.01153628807514906, + 0.008073295466601849, + 0.00847499631345272, + 0.009621159173548222, + 0.010619401931762695, + 0.009791450574994087, + 0.0051964581944048405, + 0.007571674883365631, + 0.010846751742064953, + 0.00627626059576869, + 0.00759799825027585, + 0.0099776741117239, + 0.0067350296303629875, + 0.00595526909455657, + 0.00704900361597538, + 0.0067174131982028484, + 0.004632155876606703, + 0.006090199109166861, + 0.0056069823913276196, + 0.009579654783010483, + 0.005501114763319492, + 0.005293767899274826, + 0.0059349630028009415, + 0.010903893038630486, + 0.009840765036642551, + 0.019998818635940552, + 0.011160404421389103, + 0.014759081415832043, + 0.010237101465463638, + 0.007484591566026211, + 0.015180894173681736, + 0.010760488919913769, + 0.009524453431367874, + 0.011968020349740982, + 0.01150402333587408, + 0.010548560880124569, + 0.011882228776812553, + 0.011739084497094154, + 0.013305849395692348, + 0.007791559211909771, + 0.008597740903496742, + 0.006531909108161926, + 0.00972395483404398, + 0.010157025419175625, + 0.04113030061125755, + 0.016581762582063675, + 0.038737036287784576, + 0.011895609088242054, + 0.013857359066605568, + 0.01453295536339283, + 0.03021598793566227, + 0.019030679017305374, + 0.04524094611406326, + 0.058591246604919434, + 0.018861183896660805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.008632604032754898 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017094297800213099 + ], + "opposite_logprob_tokens": [ + -6.376709461212158 + ], + "logprob_reward": -0.0017094297800213099, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017094297800213099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061130937188863754, + 0.0436001792550087, + 0.05968907102942467, + 0.04008355364203453, + 0.04264606907963753, + 0.030300354585051537, + 0.03654955327510834, + 0.038393449038267136, + 0.21406987309455872, + 0.43353700637817383 + ], + "credit_map": [ + 0.004451563116163015, + 0.0074647036381065845, + 0.009554313495755196, + 0.006815435830503702, + 0.008046261966228485, + 0.00969701912254095, + 0.007759970147162676, + 0.007341674063354731, + 0.005432118196040392, + 0.005001481622457504, + 0.006181133445352316, + 0.0045869005843997, + 0.0061978972516953945, + 0.00639266287907958, + 0.0026827005203813314, + 0.007125279400497675, + 0.006403557490557432, + 0.00698531698435545, + 0.006155197508633137, + 0.009702158160507679, + 0.008863349445164204, + 0.010197431780397892, + 0.005881757475435734, + 0.00550030916929245, + 0.00729574728757143, + 0.005852560978382826, + 0.0053064231760799885, + 0.0039050441700965166, + 0.005576439667493105, + 0.005491680931299925, + 0.0032211756333708763, + 0.0034344824962317944, + 0.004385098814964294, + 0.0075789643451571465, + 0.004506317432969809, + 0.007733254227787256, + 0.004687163047492504, + 0.004690181463956833, + 0.00481776800006628, + 0.004247324541211128, + 0.0036919598933309317, + 0.0028353219386190176, + 0.007044043391942978, + 0.004108476918190718, + 0.0025577726773917675, + 0.003529701381921768, + 0.0015683203237131238, + 0.004964756313711405, + 0.004724455531686544, + 0.0042039682157337666, + 0.005776888225227594, + 0.004265820607542992, + 0.004361088387668133, + 0.003849372733384371, + 0.0027438474353402853, + 0.006624113768339157, + 0.0032157190144062042, + 0.003657417604699731, + 0.004861588589847088, + 0.004643608350306749, + 0.006217031739652157, + 0.00352148967795074, + 0.007029715925455093, + 0.005246877204626799, + 0.013056333176791668, + 0.015351898968219757, + 0.0120118148624897, + 0.017605438828468323, + 0.011735660955309868, + 0.06275574117898941, + 0.02649061754345894, + 0.05506236106157303, + 0.02259019948542118, + 0.024564046412706375, + 0.026966717094182968, + 0.061204057186841965, + 0.03573141247034073, + 0.09919240325689316, + 0.12014570832252502, + 0.043142449110746384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0017094297800213099 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019384894985705614 + ], + "opposite_logprob_tokens": [ + -6.251938343048096 + ], + "logprob_reward": -0.0019384894985705614, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019384894985705614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1719018667936325, + 0.15296484529972076, + 0.154715433716774, + 0.06840702891349792, + 0.05533924326300621, + 0.06482026726007462, + 0.0694761574268341, + 0.041162099689245224, + 0.10866721719503403, + 0.11254587769508362 + ], + "credit_map": [ + 0.014045367948710918, + 0.01815451681613922, + 0.03316202014684677, + 0.01771840825676918, + 0.018936850130558014, + 0.02604641392827034, + 0.020365767180919647, + 0.023472515866160393, + 0.013821687549352646, + 0.01619431935250759, + 0.026614785194396973, + 0.01709476299583912, + 0.01914294809103012, + 0.02381492778658867, + 0.00785469263792038, + 0.028426723554730415, + 0.03223472461104393, + 0.021136097609996796, + 0.015422158874571323, + 0.016880139708518982, + 0.01387492660433054, + 0.021636812016367912, + 0.016552822664380074, + 0.016977759078145027, + 0.00896939355880022, + 0.008798101916909218, + 0.007717390079051256, + 0.007845171727240086, + 0.011245139874517918, + 0.011326749809086323, + 0.007112265098839998, + 0.00539281964302063, + 0.007807496003806591, + 0.011895251460373402, + 0.007864880375564098, + 0.005853666923940182, + 0.005112538114190102, + 0.0058495295234024525, + 0.005712973419576883, + 0.005242906976491213, + 0.008481940254569054, + 0.0067276665940880775, + 0.010751302354037762, + 0.006982667837291956, + 0.0070844292640686035, + 0.010241258889436722, + 0.004164066165685654, + 0.010386936366558075, + 0.009749103337526321, + 0.007530790753662586, + 0.011519275605678558, + 0.013302396051585674, + 0.005960281938314438, + 0.006184304133057594, + 0.004639900755137205, + 0.010590102523565292, + 0.0038015067111700773, + 0.0038326019421219826, + 0.005765868816524744, + 0.004928850568830967, + 0.006713750772178173, + 0.005262366496026516, + 0.0054467287845909595, + 0.005410424433648586, + 0.007346655707806349, + 0.016653921455144882, + 0.010007588192820549, + 0.013008241541683674, + 0.008840507827699184, + 0.020908579230308533, + 0.0165242962539196, + 0.015377424657344818, + 0.011651300825178623, + 0.013808739371597767, + 0.011464012786746025, + 0.01986781507730484, + 0.016331389546394348, + 0.012125820852816105, + 0.014666945673525333, + 0.012629847973585129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0019384894985705614 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a dramatic aerial drone shot, swooping down at breakneck speed through a sky shrouded in thick, dark clouds before stabilizing at a low, upward-looking angle, focusing on the center of a vast expanse of ancient battlefield ruins. The ground is cracked, dry earth, littered with broken stone pillars and crumbling walls. A knight clad in dark silver plate armor takes center stage. His armor is deeply scarred and battle-scarred, its pauldrons intricately carved with griffins, and its helmet adorned with faded red plumes. The metal reflects the cold light filtering through gaps in the clouds. In the center of the ruins, he tenses his legs and leaps from the ground, his heavy body arcing in a powerful arc as he reaches his highest point. Then, with all his might, he grasps a massive two-handed sword, nearly his height, engraved with ancient runes, and strikes down with thunderous force. As the tip of the giant sword slammed heavily on the ground, a powerful shock wave instantly erupted with the tip of the sword as the center, stirring up countless gravel, sand and thick dust, forming a circle of dust cloud that spread outward. The entire picture presented a highly detailed, top-level game CG style full of motion blur and particle effects, creating an epic and tension-filled battle atmosphere.", + "seed": 30113, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_3/index_9614_seed_30113_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.052693761279806496, + "vlm_logprob": -0.052693761279806496, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.052693761279806496, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2281564474105835 + ], + "opposite_logprob_tokens": [ + -1.6031564474105835 + ], + "logprob_reward": -0.2281564474105835, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2281564474105835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3960345685482025, + 0.07290700823068619, + 0.1605781465768814, + 0.08338318020105362, + 0.06912392377853394, + 0.08878380060195923, + 0.03382303938269615, + 0.02737373858690262, + 0.04383057728409767, + 0.02416197583079338 + ], + "credit_map": [ + 0.056078508496284485, + 0.09917167574167252, + 0.021285125985741615, + 0.020948775112628937, + 0.04877125471830368, + 0.06797310709953308, + 0.038785167038440704, + 0.04302092641592026, + 0.006239317357540131, + 0.006331148091703653, + 0.006845458876341581, + 0.010615675710141659, + 0.00825596135109663, + 0.016631457954645157, + 0.008779791183769703, + 0.0092081967741251, + 0.01883399672806263, + 0.024167029187083244, + 0.01435905508697033, + 0.02725795842707157, + 0.016626877710223198, + 0.020637763664126396, + 0.013125961646437645, + 0.025569504126906395, + 0.010001321323215961, + 0.012131470255553722, + 0.006700756028294563, + 0.007853750139474869, + 0.01923673413693905, + 0.010355084203183651, + 0.008317314088344574, + 0.00878674816340208, + 0.009563324972987175, + 0.009396901354193687, + 0.014156718738377094, + 0.00896032527089119, + 0.0070127262733876705, + 0.01065691839903593, + 0.005022765137255192, + 0.00435424642637372, + 0.01609458029270172, + 0.008820879273116589, + 0.009887464344501495, + 0.00895035732537508, + 0.012081452645361423, + 0.014522521756589413, + 0.011534121818840504, + 0.006892428267747164, + 0.003174582263454795, + 0.004233493469655514, + 0.00546906515955925, + 0.003766501322388649, + 0.003707536496222019, + 0.005630085710436106, + 0.0033911268692463636, + 0.00445064902305603, + 0.0025961212813854218, + 0.0038509387522935867, + 0.004169345833361149, + 0.002912539290264249, + 0.0025524890515953302, + 0.0028838913422077894, + 0.0038460500072687864, + 0.004562361165881157, + 0.005977393593639135, + 0.005754548124969006, + 0.008606074377894402, + 0.005697163753211498, + 0.003431698540225625, + 0.004752416163682938, + 0.004187153652310371, + 0.00542412931099534, + 0.0021636495366692543, + 0.003199222031980753, + 0.003091285703703761, + 0.0030963802710175514, + 0.0033201121259480715, + 0.0031057195737957954, + 0.0033687972463667393, + 0.00281680840998888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.2281564474105835 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210578985512257 + ], + "opposite_logprob_tokens": [ + -4.2642107009887695 + ], + "logprob_reward": -0.014210578985512257, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210578985512257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2851613461971283, + 0.0784207209944725, + 0.14890486001968384, + 0.09219103306531906, + 0.03888631612062454, + 0.09407760202884674, + 0.11308322101831436, + 0.0374300479888916, + 0.06478094309568405, + 0.047063980251550674 + ], + "credit_map": [ + 0.03572586923837662, + 0.0860370397567749, + 0.011411341838538647, + 0.013646568171679974, + 0.037747517228126526, + 0.03617330640554428, + 0.02229410596191883, + 0.04212556406855583, + 0.008378541097044945, + 0.005688579753041267, + 0.00637019844725728, + 0.008779079653322697, + 0.006643630098551512, + 0.021883418783545494, + 0.009750794619321823, + 0.010926472954452038, + 0.013675425201654434, + 0.035643529146909714, + 0.012839742004871368, + 0.01876978762447834, + 0.019010081887245178, + 0.022000497207045555, + 0.011775695718824863, + 0.01519008632749319, + 0.009914325550198555, + 0.011823708191514015, + 0.0171117652207613, + 0.01444228645414114, + 0.011232987977564335, + 0.011743185110390186, + 0.007673808839172125, + 0.008248956874012947, + 0.006033800076693296, + 0.006067358888685703, + 0.006415593437850475, + 0.004561902489513159, + 0.0032144614960998297, + 0.007300279103219509, + 0.0021918562706559896, + 0.0031010606326162815, + 0.016930673271417618, + 0.00912137608975172, + 0.009671520441770554, + 0.00933786015957594, + 0.012255378998816013, + 0.016747144982218742, + 0.012889821082353592, + 0.007123812101781368, + 0.010085039772093296, + 0.025040052831172943, + 0.018113410100340843, + 0.017743593081831932, + 0.007421033922582865, + 0.014011465944349766, + 0.00729290721938014, + 0.01337569858878851, + 0.004630313720554113, + 0.005147892981767654, + 0.004968799185007811, + 0.004447244573384523, + 0.004342529457062483, + 0.005990008357912302, + 0.00380186946131289, + 0.004101385362446308, + 0.008361796848475933, + 0.007878031581640244, + 0.014018182642757893, + 0.009282619692385197, + 0.004410429857671261, + 0.006148773245513439, + 0.006491788197308779, + 0.008189314976334572, + 0.005362844094634056, + 0.006967118475586176, + 0.007937174290418625, + 0.005760383326560259, + 0.004513418301939964, + 0.00611685449257493, + 0.005664558615535498, + 0.0047416239976882935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.014210578985512257 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191421680152416 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.014191421680152416, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191421680152416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15827250480651855, + 0.07887271791696548, + 0.12028876692056656, + 0.09290069341659546, + 0.059872426092624664, + 0.12116793543100357, + 0.10829496383666992, + 0.07376934587955475, + 0.09775254875421524, + 0.08880812674760818 + ], + "credit_map": [ + 0.020422544330358505, + 0.014942298643290997, + 0.024002421647310257, + 0.017514407634735107, + 0.01742769405245781, + 0.024150656536221504, + 0.023568961769342422, + 0.016243524849414825, + 0.007799929939210415, + 0.005427337717264891, + 0.006917812395840883, + 0.006989998277276754, + 0.00782744213938713, + 0.024555014446377754, + 0.008095216937363148, + 0.011259960941970348, + 0.008183049969375134, + 0.022408394142985344, + 0.01107280608266592, + 0.02096494659781456, + 0.013370038010179996, + 0.01570908911526203, + 0.0102346520870924, + 0.018345799297094345, + 0.011123036034405231, + 0.011849213391542435, + 0.009174549020826817, + 0.010868996381759644, + 0.019371384754776955, + 0.012155873700976372, + 0.0074300579726696014, + 0.010927586816251278, + 0.007968518882989883, + 0.008333852514624596, + 0.008887436240911484, + 0.008035955019295216, + 0.0046220687218010426, + 0.010233577340841293, + 0.0075936890207231045, + 0.004197330214083195, + 0.013453487306833267, + 0.02307119593024254, + 0.014280859380960464, + 0.012896519154310226, + 0.014189968816936016, + 0.014546835795044899, + 0.018674353137612343, + 0.010054714977741241, + 0.011353380046784878, + 0.017913952469825745, + 0.021359456703066826, + 0.012733839452266693, + 0.007272255141288042, + 0.0159283559769392, + 0.00977956410497427, + 0.011954161338508129, + 0.005043033044785261, + 0.005590276792645454, + 0.007488815579563379, + 0.006558396387845278, + 0.005206099711358547, + 0.0046829963102936745, + 0.023477058857679367, + 0.015722671523690224, + 0.009560630656778812, + 0.013704128563404083, + 0.016194000840187073, + 0.017850350588560104, + 0.007281220052391291, + 0.009040826931595802, + 0.011876029893755913, + 0.012245367281138897, + 0.008583547547459602, + 0.010762715712189674, + 0.01144015695899725, + 0.01179330050945282, + 0.01257809717208147, + 0.008230604231357574, + 0.014582954347133636, + 0.010836745612323284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.014191421680152416 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002814976265653968 + ], + "opposite_logprob_tokens": [ + -5.877814769744873 + ], + "logprob_reward": -0.002814976265653968, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002814976265653968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15738680958747864, + 0.05112677812576294, + 0.053830988705158234, + 0.051483966410160065, + 0.03952041268348694, + 0.05388571694493294, + 0.15826688706874847, + 0.08333170413970947, + 0.13806352019309998, + 0.21310317516326904 + ], + "credit_map": [ + 0.01589278317987919, + 0.03706063702702522, + 0.009865214116871357, + 0.008985275402665138, + 0.021027464419603348, + 0.020012035965919495, + 0.01637279987335205, + 0.028170600533485413, + 0.005666729994118214, + 0.0046414039097726345, + 0.003744829446077347, + 0.0033404051791876554, + 0.006759245414286852, + 0.015287326648831367, + 0.005435804836452007, + 0.006251031998544931, + 0.0075436923652887344, + 0.006081167608499527, + 0.004527585115283728, + 0.005728514865040779, + 0.007404521107673645, + 0.007279852405190468, + 0.005239862948656082, + 0.01002578902989626, + 0.005332948639988899, + 0.006728227250277996, + 0.00480519188567996, + 0.004909001290798187, + 0.009607849642634392, + 0.007002926431596279, + 0.004590692464262247, + 0.00850712601095438, + 0.004790079779922962, + 0.004342515487223864, + 0.00484183244407177, + 0.004916264209896326, + 0.004208196420222521, + 0.008595890365540981, + 0.004580374341458082, + 0.0032452568411827087, + 0.008915732614696026, + 0.007401022594422102, + 0.006169837899506092, + 0.005608243402093649, + 0.005962785799056292, + 0.0113924415782094, + 0.004574093502014875, + 0.0038615597877651453, + 0.013655520044267178, + 0.028006471693515778, + 0.024936417117714882, + 0.022315626963973045, + 0.016553206369280815, + 0.01967807114124298, + 0.014360055327415466, + 0.0187615267932415, + 0.007389450445771217, + 0.009849902242422104, + 0.010470596142113209, + 0.007527846843004227, + 0.0038937076460570097, + 0.006956127006560564, + 0.020902495831251144, + 0.01634157821536064, + 0.019457770511507988, + 0.014491603709757328, + 0.01669284701347351, + 0.01660585030913353, + 0.010760386474430561, + 0.01864996738731861, + 0.02406683750450611, + 0.017338264733552933, + 0.03306566923856735, + 0.020057402551174164, + 0.035615239292383194, + 0.029623743146657944, + 0.022439537569880486, + 0.022176798433065414, + 0.026342207565903664, + 0.023782581090927124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.002814976265653968 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an aerial drone shot swooping down through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low and looking upward at a knight in dark silver plate armor with griffin-carved pauldrons and a helmet with faded red plumes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the knight leap from the cracked, dry earth ruins and strike down with a massive two-handed sword engraved with ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sword strike create a shock wave that stirs up gravel, sand, and dust forming an expanding circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a highly detailed game CG style featuring motion blur and particle effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095382057130337 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095382057130337, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095382057130337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15648280084133148, + 0.07216869294643402, + 0.37885594367980957, + 0.03957776725292206, + 0.04505740478634834, + 0.04487955942749977, + 0.09475862234830856, + 0.09112279862165451, + 0.03463916480541229, + 0.042457226663827896 + ], + "credit_map": [ + 0.027551213279366493, + 0.02594384364783764, + 0.01591082476079464, + 0.01109479833394289, + 0.016254186630249023, + 0.02943846955895424, + 0.014457848854362965, + 0.015831610187888145, + 0.00816193874925375, + 0.008461100980639458, + 0.006655234843492508, + 0.007357403170317411, + 0.007240749895572662, + 0.01729467138648033, + 0.008405259810388088, + 0.0085923345759511, + 0.05093800276517868, + 0.040997982025146484, + 0.023376382887363434, + 0.033661603927612305, + 0.05471518635749817, + 0.0552188903093338, + 0.03499269858002663, + 0.08495521545410156, + 0.004101916681975126, + 0.005759711377322674, + 0.0032842354848980904, + 0.0034694073256105185, + 0.007006688509136438, + 0.0050645796582102776, + 0.0037373362574726343, + 0.007153893820941448, + 0.005689405836164951, + 0.005626744125038385, + 0.006783790420740843, + 0.007077141664922237, + 0.004295710474252701, + 0.009301912039518356, + 0.003026125254109502, + 0.003256575670093298, + 0.0050018057227134705, + 0.007589935325086117, + 0.005710366647690535, + 0.006040958222001791, + 0.004596233367919922, + 0.00746136624366045, + 0.004521828144788742, + 0.003957065287977457, + 0.009362734854221344, + 0.01044390257447958, + 0.015578770078718662, + 0.012110540643334389, + 0.007747817784547806, + 0.01875418983399868, + 0.010102847591042519, + 0.010657815262675285, + 0.009202885441482067, + 0.008830380626022816, + 0.008748991414904594, + 0.005524527747184038, + 0.007922153919935226, + 0.007646229118108749, + 0.027143679559230804, + 0.0161039549857378, + 0.004403539467602968, + 0.004846180323511362, + 0.00407221307978034, + 0.003806506050750613, + 0.00400247098878026, + 0.005036008078604937, + 0.004281943663954735, + 0.004190301988273859, + 0.004786035977303982, + 0.007656693458557129, + 0.006820457521826029, + 0.0044623627327382565, + 0.00573399942368269, + 0.003623076481744647, + 0.005375439766794443, + 0.003999160602688789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.004095382057130337 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_4/index_tensor([9984, 9984])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_4/index_tensor([9984, 9984])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e2b32f6fb7d1c729cee3f594b7dbc644dcbd7666 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_4/index_tensor([9984, 9984])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Viewed from a static, high-angle perspective, the scene captures a sleek black smartphone lying motionless, face-up, on a polished wooden table. The phone's glossy, dark screen vaguely reflects soft, diffuse light from an unseen window, while its matte black exterior contrasts with the warm tones of the wood desk, revealing a prominent, detailed grain. In a shallow depth of field, the out-of-focus corner of a notebook and a simple ceramic mug rest in the background, establishing an atmosphere of quiet concentration. This tranquility is abruptly shattered when an internal mechanism activates, causing the device to vibrate with an intense, humming energy, its form slightly blurred by its rapid oscillations. Driven by this force, the smartphone slides and rotates erratically across the wooden surface, its hard edges rattling against the table as it moves in an unpredictable, short arc. Just as suddenly as it began, the vibration ceases, and the phone instantly and completely stops, now resting at a new, tilted angle, allowing the scene to return to a deep, expectant silence.", + "seed": 10539, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_4/index_9984_seed_10539_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12728359308093787, + "vlm_logprob": -0.12728359308093787, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.12728359308093787, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06155082583427429 + ], + "opposite_logprob_tokens": [ + -5.061550617218018 + ], + "logprob_reward": -0.06155082583427429, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.06155082583427429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.111542247235775, + 0.08431565761566162, + 0.43994560837745667, + 0.021720068529248238, + 0.02123126946389675, + 0.05246396362781525, + 0.08212929219007492, + 0.054868027567863464, + 0.04455387964844704, + 0.08722996711730957 + ], + "credit_map": [ + 0.012020297348499298, + 0.018900742754340172, + 0.017907464876770973, + 0.010724354535341263, + 0.011495575308799744, + 0.014065521769225597, + 0.007222298998385668, + 0.019206000491976738, + 0.005984223447740078, + 0.02690267190337181, + 0.02320653758943081, + 0.0065827080979943275, + 0.006448688451200724, + 0.0059861717745661736, + 0.003361219773069024, + 0.005843434017151594, + 0.014356876723468304, + 0.18325471878051758, + 0.18454109132289886, + 0.013407024554908276, + 0.017692606896162033, + 0.010043734684586525, + 0.007664520759135485, + 0.008985008113086224, + 0.00171499268617481, + 0.0028098176699131727, + 0.00452858442440629, + 0.0017072896007448435, + 0.003079047193750739, + 0.0020607567857950926, + 0.0016035287408158183, + 0.0042160507291555405, + 0.0037801058497279882, + 0.0028812354430556297, + 0.0028351149521768093, + 0.0025285505689680576, + 0.0022832336835563183, + 0.002063981257379055, + 0.0018741204403340816, + 0.0029849272686988115, + 0.00896404404193163, + 0.00543327908962965, + 0.005095845088362694, + 0.007520332001149654, + 0.005024945829063654, + 0.005028265994042158, + 0.008252972736954689, + 0.007144278381019831, + 0.00871992576867342, + 0.011018612422049046, + 0.01863659732043743, + 0.010483148507773876, + 0.008818274363875389, + 0.004768947139382362, + 0.009866239503026009, + 0.009817552752792835, + 0.004991649184376001, + 0.006425959523767233, + 0.013511664234101772, + 0.006314439233392477, + 0.005536213517189026, + 0.004390248097479343, + 0.005570235196501017, + 0.008127618581056595, + 0.005378893576562405, + 0.005285231862217188, + 0.009350820444524288, + 0.005245025735348463, + 0.0051702396012842655, + 0.005051602143794298, + 0.0035010487772524357, + 0.005571017973124981, + 0.010383423417806625, + 0.009906225837767124, + 0.023400355130434036, + 0.00833928119391203, + 0.010567350313067436, + 0.005351539235562086, + 0.006190524436533451, + 0.013091268949210644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.06155082583427429 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007651426829397678 + ], + "opposite_logprob_tokens": [ + -4.882651329040527 + ], + "logprob_reward": -0.007651426829397678, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007651426829397678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22461721301078796, + 0.2384522408246994, + 0.08049099147319794, + 0.22025902569293976, + 0.030372917652130127, + 0.0984865203499794, + 0.022491557523608208, + 0.02654111012816429, + 0.02384544350206852, + 0.03444289043545723 + ], + "credit_map": [ + 0.020577799528837204, + 0.0713357999920845, + 0.06060086563229561, + 0.016078349202871323, + 0.017678773030638695, + 0.011088109575212002, + 0.00820531789213419, + 0.019052227959036827, + 0.018082285299897194, + 0.0814041867852211, + 0.06897243112325668, + 0.019873814657330513, + 0.016789093613624573, + 0.01393282413482666, + 0.008229710161685944, + 0.011167911812663078, + 0.00694719422608614, + 0.0179895106703043, + 0.014017242938280106, + 0.005957309156656265, + 0.008169724605977535, + 0.009813744574785233, + 0.005229020491242409, + 0.012367253191769123, + 0.025542035698890686, + 0.036459580063819885, + 0.03271564468741417, + 0.02033442072570324, + 0.03158015385270119, + 0.019606076180934906, + 0.027965035289525986, + 0.02605610340833664, + 0.0030477677937597036, + 0.003066920442506671, + 0.002880572574213147, + 0.004668010398745537, + 0.0038428539410233498, + 0.005155670456588268, + 0.0035809441469609737, + 0.004130180925130844, + 0.013741271570324898, + 0.0077988808043301105, + 0.012889414094388485, + 0.01600191555917263, + 0.007965918630361557, + 0.009647448547184467, + 0.016290077939629555, + 0.014151599258184433, + 0.002292918972671032, + 0.0032147408928722143, + 0.003874581540003419, + 0.0030275690369307995, + 0.002624379238113761, + 0.0016817506402730942, + 0.002749839797616005, + 0.003025779500603676, + 0.003940329886972904, + 0.00250918953679502, + 0.0038225671742111444, + 0.003657999448478222, + 0.0028713957872241735, + 0.002830191981047392, + 0.003783224383369088, + 0.003126216121017933, + 0.003003440098837018, + 0.002701678080484271, + 0.0048288460820913315, + 0.0024392236955463886, + 0.002402447396889329, + 0.0029582069255411625, + 0.0026695888955146074, + 0.0028420144226402044, + 0.004301887936890125, + 0.004510856699198484, + 0.007608382496982813, + 0.00450159190222621, + 0.003513240721076727, + 0.0029808590188622475, + 0.002836589701473713, + 0.0041894870810210705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.007651426829397678 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059001475572586 + ], + "opposite_logprob_tokens": [ + -3.89559006690979 + ], + "logprob_reward": -0.02059001475572586, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059001475572586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040066804736852646, + 0.08763255178928375, + 0.21200844645500183, + 0.06640678644180298, + 0.04587697237730026, + 0.43449607491493225, + 0.012439485639333725, + 0.0331963486969471, + 0.023184338584542274, + 0.04469215124845505 + ], + "credit_map": [ + 0.003752920078113675, + 0.00539228692650795, + 0.0062225922010838985, + 0.0036105019971728325, + 0.0042925868183374405, + 0.004148128442466259, + 0.0024694614112377167, + 0.010178325697779655, + 0.005096031352877617, + 0.028658155351877213, + 0.024010036140680313, + 0.00835232250392437, + 0.006702573969960213, + 0.006593231577426195, + 0.003834783798083663, + 0.0043854243122041225, + 0.011414065025746822, + 0.08008480072021484, + 0.08069440722465515, + 0.006914124824106693, + 0.010518042370676994, + 0.0060909283347427845, + 0.004853846505284309, + 0.011438229121267796, + 0.003554865252226591, + 0.01689913123846054, + 0.021518535912036896, + 0.0037894481793045998, + 0.004045425448566675, + 0.0024506007321178913, + 0.005464188754558563, + 0.008684592321515083, + 0.007520909886807203, + 0.005092855542898178, + 0.007886790670454502, + 0.004137847106903791, + 0.004454964306205511, + 0.00526244193315506, + 0.0036077306140214205, + 0.007913434877991676, + 0.0610104463994503, + 0.0335882306098938, + 0.056954555213451385, + 0.06928106397390366, + 0.0351397879421711, + 0.04703392833471298, + 0.07035231590270996, + 0.061135753989219666, + 0.0012628939002752304, + 0.001428206218406558, + 0.002022570464760065, + 0.001824885723181069, + 0.0018427014583721757, + 0.0014912100741639733, + 0.001207429450005293, + 0.001359588117338717, + 0.003911115229129791, + 0.0038494167383760214, + 0.009250088594853878, + 0.0035108120646327734, + 0.0028925552032887936, + 0.002206679666414857, + 0.0027466833125799894, + 0.00482900207862258, + 0.0030537100974470377, + 0.002697515767067671, + 0.0033122303429991007, + 0.0026082429103553295, + 0.0027136807329952717, + 0.003094074549153447, + 0.002686580875888467, + 0.0030183044727891684, + 0.004157165996730328, + 0.004854057915508747, + 0.010026896372437477, + 0.00491702975705266, + 0.00581812160089612, + 0.004545519128441811, + 0.0037411709781736135, + 0.006632189266383648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.02059001475572586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233461260795593 + ], + "opposite_logprob_tokens": [ + -0.8983461260795593 + ], + "logprob_reward": -0.5233461260795593, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233461260795593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18572528660297394, + 0.08314644545316696, + 0.07844605296850204, + 0.04505929350852966, + 0.04998718202114105, + 0.31731846928596497, + 0.03657429292798042, + 0.12731385231018066, + 0.05644749104976654, + 0.019981641322374344 + ], + "credit_map": [ + 0.019358782097697258, + 0.0571180023252964, + 0.04977892339229584, + 0.012376406230032444, + 0.012496255338191986, + 0.009911475703120232, + 0.007168509066104889, + 0.01751692406833172, + 0.005180239677429199, + 0.028548182919621468, + 0.02397090196609497, + 0.006945842411369085, + 0.005736819002777338, + 0.005409720819443464, + 0.0030456881504505873, + 0.0043090470135211945, + 0.004168686922639608, + 0.022586774080991745, + 0.02807467244565487, + 0.005600818898528814, + 0.005405985750257969, + 0.004400529898703098, + 0.0031203667167574167, + 0.005088213365525007, + 0.004415519069880247, + 0.008250563405454159, + 0.007243455853313208, + 0.004404445178806782, + 0.005862720776349306, + 0.0025987017434090376, + 0.003644251963123679, + 0.008639631792902946, + 0.008648612536489964, + 0.0059573170728981495, + 0.006885898765176535, + 0.005608047358691692, + 0.005677147768437862, + 0.004914527293294668, + 0.005008357111364603, + 0.007287270855158567, + 0.04370483011007309, + 0.028256181627511978, + 0.04148544371128082, + 0.055002521723508835, + 0.02644455060362816, + 0.022311417385935783, + 0.054279815405607224, + 0.045833710581064224, + 0.0031118213664740324, + 0.004527836572378874, + 0.010104364715516567, + 0.00443563237786293, + 0.0030778711661696434, + 0.002316355472430587, + 0.0033925147727131844, + 0.005607896018773317, + 0.018116410821676254, + 0.012635317631065845, + 0.025464361533522606, + 0.015899471938610077, + 0.012272884137928486, + 0.011436615139245987, + 0.015646465122699738, + 0.015842324122786522, + 0.00800463929772377, + 0.00565456273034215, + 0.009553967043757439, + 0.006875744089484215, + 0.0059395902790129185, + 0.00783650390803814, + 0.005909511353820562, + 0.006672971416264772, + 0.0024869402404874563, + 0.0020224216859787703, + 0.003184752305969596, + 0.0026808169204741716, + 0.0025824957992881536, + 0.00263019441626966, + 0.0021337412763386965, + 0.0022602786775678396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.5233461260795593 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023279571905732155 + ], + "opposite_logprob_tokens": [ + -3.7732796669006348 + ], + "logprob_reward": -0.023279571905732155, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023279571905732155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35664331912994385, + 0.12228384613990784, + 0.10065566748380661, + 0.03736194968223572, + 0.027403593063354492, + 0.21532435715198517, + 0.014304695650935173, + 0.051596906036138535, + 0.021414687857031822, + 0.05301099643111229 + ], + "credit_map": [ + 0.03251880034804344, + 0.12502700090408325, + 0.10619688779115677, + 0.024278922006487846, + 0.024548597633838654, + 0.014930279925465584, + 0.011335181072354317, + 0.017807668074965477, + 0.008229674771428108, + 0.043775785714387894, + 0.0366562195122242, + 0.009307355619966984, + 0.008078119717538357, + 0.006669233087450266, + 0.0038985111750662327, + 0.005668956320732832, + 0.006924476940184832, + 0.03102976456284523, + 0.03348023071885109, + 0.005348749924451113, + 0.007176410872489214, + 0.005505641456693411, + 0.0031876876018941402, + 0.008002709597349167, + 0.002767120022326708, + 0.008197189308702946, + 0.008397913537919521, + 0.004420164041221142, + 0.004100843798369169, + 0.00210655783303082, + 0.002876564161852002, + 0.004495595116168261, + 0.003912167623639107, + 0.003550096182152629, + 0.00379051617346704, + 0.003942668903619051, + 0.0032873768359422684, + 0.002684278879314661, + 0.0030220116022974253, + 0.003214475931599736, + 0.029845474287867546, + 0.018547115847468376, + 0.02871754951775074, + 0.036358289420604706, + 0.017716500908136368, + 0.016594888642430305, + 0.03619693964719772, + 0.031347598880529404, + 0.0015601812629029155, + 0.0018329048762097955, + 0.0038505529519170523, + 0.0016191682079806924, + 0.0015371715417131782, + 0.0010381125612184405, + 0.00126068782992661, + 0.0016059159534052014, + 0.006864215712994337, + 0.00536541361361742, + 0.01225392147898674, + 0.005870695225894451, + 0.004697391297668219, + 0.004149344749748707, + 0.005409396719187498, + 0.006986529566347599, + 0.002663547405973077, + 0.002001322340220213, + 0.00446282047778368, + 0.002247927477583289, + 0.0023236905690282583, + 0.0030025348532944918, + 0.0021410309709608555, + 0.002571812132373452, + 0.008400049060583115, + 0.005931891035288572, + 0.011225966736674309, + 0.004780980292707682, + 0.007305331062525511, + 0.0038940308149904013, + 0.004657716955989599, + 0.006815027911216021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.023279571905732155 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Viewed from a static, high-angle perspective, the scene captures a sleek black smartphone lying motionless, face-up, on a polished wooden table. The phone's glossy, dark screen vaguely reflects soft, diffuse light from an unseen window, while its matte black exterior contrasts with the warm tones of the wood desk, revealing a prominent, detailed grain. In a shallow depth of field, the out-of-focus corner of a notebook and a simple ceramic mug rest in the background, establishing an atmosphere of quiet concentration. This tranquility is abruptly shattered when an internal mechanism activates, causing the device to vibrate with an intense, humming energy, its form slightly blurred by its rapid oscillations. Driven by this force, the smartphone slides and rotates erratically across the wooden surface, its hard edges rattling against the table as it moves in an unpredictable, short arc. Just as suddenly as it began, the vibration ceases, and the phone instantly and completely stops, now resting at a new, tilted angle, allowing the scene to return to a deep, expectant silence.", + "seed": 10540, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_4/index_9984_seed_10540_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.028316983301192523, + "vlm_logprob": -0.028316983301192523, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.028316983301192523, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06879287958145142 + ], + "opposite_logprob_tokens": [ + -3.9437928199768066 + ], + "logprob_reward": -0.06879287958145142, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.06879287958145142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16675320267677307, + 0.060908615589141846, + 0.09633401781320572, + 0.02664809860289097, + 0.14019833505153656, + 0.10201234370470047, + 0.12010432779788971, + 0.04605644568800926, + 0.1906888782978058, + 0.050295718014240265 + ], + "credit_map": [ + 0.01641685888171196, + 0.022146787494421005, + 0.02472764253616333, + 0.015435167588293552, + 0.019230006262660027, + 0.03511525318026543, + 0.012065134942531586, + 0.021616358309984207, + 0.010684018023312092, + 0.008800378069281578, + 0.010896088555455208, + 0.0070044309832155704, + 0.005625195801258087, + 0.006020750850439072, + 0.006330372765660286, + 0.0055473847314715385, + 0.017084062099456787, + 0.012619714252650738, + 0.018248364329338074, + 0.010932628996670246, + 0.012667912989854813, + 0.008087229914963245, + 0.009712763130664825, + 0.006981342565268278, + 0.002534659346565604, + 0.004029601812362671, + 0.003932440187782049, + 0.003794747870415449, + 0.0031095182057470083, + 0.0025659562088549137, + 0.0033403094857931137, + 0.00334086618386209, + 0.026026804000139236, + 0.015054944902658463, + 0.021795742213726044, + 0.019939865916967392, + 0.015193079598248005, + 0.013970065861940384, + 0.01646418310701847, + 0.011753654107451439, + 0.011726398952305317, + 0.011242333799600601, + 0.01546643115580082, + 0.005733845289796591, + 0.019836032763123512, + 0.014133068732917309, + 0.008441139943897724, + 0.015433091670274734, + 0.01654723845422268, + 0.01963420771062374, + 0.019051803275942802, + 0.014449892565608025, + 0.01541842520236969, + 0.012009992264211178, + 0.014754596166312695, + 0.008238174952566624, + 0.006639336235821247, + 0.008463404141366482, + 0.007484071422368288, + 0.004319680854678154, + 0.005223438609391451, + 0.0051721008494496346, + 0.004647932946681976, + 0.004106481093913317, + 0.04068419709801674, + 0.0235995315015316, + 0.030668670311570168, + 0.01775459758937359, + 0.016029799357056618, + 0.030373839661478996, + 0.010544574819505215, + 0.021033667027950287, + 0.00823819451034069, + 0.008514746092259884, + 0.009385643526911736, + 0.005585630889981985, + 0.00420443806797266, + 0.00711916433647275, + 0.0026857720222324133, + 0.004562129732221365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.06879287958145142 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023311832919716835 + ], + "opposite_logprob_tokens": [ + -3.7733118534088135 + ], + "logprob_reward": -0.023311832919716835, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023311832919716835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044809941202402115, + 0.06039584055542946, + 0.021581687033176422, + 0.01948530785739422, + 0.0615912489593029, + 0.141274556517601, + 0.05838453769683838, + 0.2760518491268158, + 0.3091471493244171, + 0.00727787334471941 + ], + "credit_map": [ + 0.0060319663025438786, + 0.005986434407532215, + 0.007691800128668547, + 0.006177169736474752, + 0.004372882191091776, + 0.004934409167617559, + 0.005691757425665855, + 0.003923519514501095, + 0.006659471895545721, + 0.0071053700521588326, + 0.00980395544320345, + 0.009274818003177643, + 0.009153422899544239, + 0.004601113498210907, + 0.00827400479465723, + 0.005523686297237873, + 0.002374570583924651, + 0.0024754584301263094, + 0.0034195291809737682, + 0.0025797360576689243, + 0.0037533179856836796, + 0.0020195276010781527, + 0.003328320337459445, + 0.001631226041354239, + 0.00464641535654664, + 0.003025349462404847, + 0.003996223211288452, + 0.0016455466393381357, + 0.001367591554298997, + 0.0020946995355188847, + 0.0011452487669885159, + 0.001564233098179102, + 0.008350590243935585, + 0.005771326832473278, + 0.009604518301784992, + 0.007652514614164829, + 0.005768349394202232, + 0.005825425963848829, + 0.01095551811158657, + 0.007663006894290447, + 0.01640634424984455, + 0.014064790681004524, + 0.01919931173324585, + 0.011752664111554623, + 0.023190991953015327, + 0.017836781218647957, + 0.01720849797129631, + 0.02161516807973385, + 0.008102710358798504, + 0.008475201204419136, + 0.008909033611416817, + 0.005553045775741339, + 0.008344058878719807, + 0.005986048374325037, + 0.008415536023676395, + 0.0045989020727574825, + 0.025064898654818535, + 0.044180408120155334, + 0.032704293727874756, + 0.03561956435441971, + 0.03702852874994278, + 0.03232394903898239, + 0.040295880287885666, + 0.028834309428930283, + 0.07645123451948166, + 0.036125630140304565, + 0.05619277432560921, + 0.020965617150068283, + 0.02649054303765297, + 0.03929964825510979, + 0.020953811705112457, + 0.032667890191078186, + 0.0011153204832226038, + 0.0008005799609236419, + 0.0011925151338800788, + 0.0011093883076682687, + 0.0007048612460494041, + 0.0008041151450015604, + 0.0009331629262305796, + 0.0006179296178743243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.023311832919716835 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645507782697678 + ], + "opposite_logprob_tokens": [ + -5.379645347595215 + ], + "logprob_reward": -0.004645507782697678, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645507782697678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20415326952934265, + 0.046757232397794724, + 0.10145597159862518, + 0.035667747259140015, + 0.04887593910098076, + 0.07604942470788956, + 0.16210883855819702, + 0.09767360985279083, + 0.2087322473526001, + 0.018525760620832443 + ], + "credit_map": [ + 0.03677047789096832, + 0.029852159321308136, + 0.03753463551402092, + 0.02168016880750656, + 0.02728242613375187, + 0.012116452679038048, + 0.02460186742246151, + 0.01431505847722292, + 0.007696247193962336, + 0.006087779998779297, + 0.008581569418311119, + 0.005593126639723778, + 0.005338679999113083, + 0.003995566163212061, + 0.005342896096408367, + 0.004121367819607258, + 0.019327431917190552, + 0.01245150901377201, + 0.01995173841714859, + 0.010833305306732655, + 0.012774656526744366, + 0.008043062873184681, + 0.010310690850019455, + 0.007763582281768322, + 0.006445035338401794, + 0.004792752210050821, + 0.006681700237095356, + 0.002984199672937393, + 0.005501693580299616, + 0.002978289732709527, + 0.0035638343542814255, + 0.0027202412020415068, + 0.004028115887194872, + 0.007491342257708311, + 0.00677008181810379, + 0.006111980881541967, + 0.006137534510344267, + 0.007170886266976595, + 0.006033660843968391, + 0.00513233570381999, + 0.009256096556782722, + 0.00821830052882433, + 0.011909090913832188, + 0.004821746610105038, + 0.01749507151544094, + 0.006716025527566671, + 0.00912570022046566, + 0.00850739423185587, + 0.02695309929549694, + 0.027814533561468124, + 0.02627325803041458, + 0.014557631686329842, + 0.020882340148091316, + 0.019378257915377617, + 0.014351586811244488, + 0.011898132972419262, + 0.013561978936195374, + 0.022944146767258644, + 0.01781287044286728, + 0.009399307891726494, + 0.007745473179966211, + 0.010754234157502651, + 0.007200379390269518, + 0.00825522094964981, + 0.051387347280979156, + 0.024691905826330185, + 0.03797808662056923, + 0.015584863722324371, + 0.017897672951221466, + 0.023361003026366234, + 0.016041088849306107, + 0.0217902734875679, + 0.0021691210567951202, + 0.0017729750834405422, + 0.002684961771592498, + 0.0019513400038704276, + 0.0017267650691792369, + 0.0047127725556492805, + 0.0012365581933408976, + 0.0022712666541337967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.004645507782697678 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810407221317291 + ], + "opposite_logprob_tokens": [ + -3.2881040573120117 + ], + "logprob_reward": -0.03810407221317291, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810407221317291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12248168140649796, + 0.028438184410333633, + 0.06000887230038643, + 0.08290662616491318, + 0.07967222481966019, + 0.05187125504016876, + 0.21090742945671082, + 0.12358690053224564, + 0.23263655602931976, + 0.007490337826311588 + ], + "credit_map": [ + 0.024362314492464066, + 0.018549377098679543, + 0.022730756551027298, + 0.012227043509483337, + 0.013913819566369057, + 0.00961315631866455, + 0.012055505067110062, + 0.009029692970216274, + 0.0029631690122187138, + 0.002614773577079177, + 0.004029415547847748, + 0.003438113024458289, + 0.0036311675794422626, + 0.004102389328181744, + 0.003963280934840441, + 0.003695872612297535, + 0.00749233877286315, + 0.007188216783106327, + 0.010430550202727318, + 0.00728790694847703, + 0.010687313042581081, + 0.00391144584864378, + 0.008307846263051033, + 0.004703242797404528, + 0.014124041423201561, + 0.01137497927993536, + 0.015791237354278564, + 0.009871135465800762, + 0.011833765543997288, + 0.005024605896323919, + 0.0084710493683815, + 0.006415802054107189, + 0.013098731637001038, + 0.010759404860436916, + 0.0141981802880764, + 0.008077729493379593, + 0.012933282181620598, + 0.007052565459161997, + 0.008040993474423885, + 0.005511328112334013, + 0.00849943608045578, + 0.005123913288116455, + 0.007870594039559364, + 0.0038928056601434946, + 0.006667132955044508, + 0.0053643109276890755, + 0.004289823584258556, + 0.010163232684135437, + 0.034337807446718216, + 0.03582232818007469, + 0.034473221749067307, + 0.019267050549387932, + 0.025722762569785118, + 0.02413230575621128, + 0.02153346687555313, + 0.015618470497429371, + 0.015928907319903374, + 0.021609531715512276, + 0.019359607249498367, + 0.011457148008048534, + 0.013581895269453526, + 0.01200312003493309, + 0.01602676324546337, + 0.013619912788271904, + 0.053784869611263275, + 0.02953949384391308, + 0.040180373936891556, + 0.018599282950162888, + 0.01859649084508419, + 0.03235563635826111, + 0.014282923191785812, + 0.025297464802861214, + 0.0014002384850755334, + 0.0009478426654823124, + 0.0013095365138724446, + 0.0005503107095137239, + 0.0006632919539697468, + 0.0010966265108436346, + 0.0007265336462296546, + 0.0007959569920785725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.03810407221317291 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730624008923769 + ], + "opposite_logprob_tokens": [ + -5.006730556488037 + ], + "logprob_reward": -0.006730624008923769, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730624008923769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06565504521131516, + 0.08189557492733002, + 0.05897311866283417, + 0.040373291820287704, + 0.11541055887937546, + 0.07004974782466888, + 0.24359852075576782, + 0.09476189315319061, + 0.21318402886390686, + 0.016098110005259514 + ], + "credit_map": [ + 0.011585868895053864, + 0.00828759279102087, + 0.010994627140462399, + 0.005355242174118757, + 0.007436058018356562, + 0.007790530100464821, + 0.006111482158303261, + 0.008093652315437794, + 0.007964292541146278, + 0.009133948013186455, + 0.01247656811028719, + 0.010178690776228905, + 0.013323015533387661, + 0.008803484961390495, + 0.010389082133769989, + 0.009626506827771664, + 0.006526785437017679, + 0.00732795987278223, + 0.00968597549945116, + 0.007307994645088911, + 0.01116890273988247, + 0.004193897824734449, + 0.007779714185744524, + 0.00498189078643918, + 0.005013545509427786, + 0.005351712461560965, + 0.006646404508501291, + 0.0055481246672570705, + 0.006527040619403124, + 0.0030910319183021784, + 0.004562709014862776, + 0.00363272987306118, + 0.016275009140372276, + 0.012508576735854149, + 0.02138982154428959, + 0.012408504262566566, + 0.015994204208254814, + 0.009092719294130802, + 0.016236430034041405, + 0.011505312286317348, + 0.008617102168500423, + 0.008602500893175602, + 0.010026209056377411, + 0.008279379457235336, + 0.0067076534032821655, + 0.009030255489051342, + 0.006103777326643467, + 0.01268287654966116, + 0.03776669502258301, + 0.041475556790828705, + 0.03976045176386833, + 0.02577916532754898, + 0.03367895260453224, + 0.02334856428205967, + 0.024435024708509445, + 0.017354140058159828, + 0.01061967946588993, + 0.020176570862531662, + 0.013695276342332363, + 0.011948947794735432, + 0.010838210582733154, + 0.009159618057310581, + 0.010141825303435326, + 0.008181769400835037, + 0.0478699654340744, + 0.027041660621762276, + 0.036382175981998444, + 0.016454273834824562, + 0.017335033044219017, + 0.030399758368730545, + 0.01308458298444748, + 0.024616610258817673, + 0.003463881555944681, + 0.002011440694332123, + 0.003004673635587096, + 0.0019786544144153595, + 0.0014832083834335208, + 0.00130113132763654, + 0.00172720558475703, + 0.0011279186001047492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.006730624008923769 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Viewed from a static, high-angle perspective, the scene captures a sleek black smartphone lying motionless, face-up, on a polished wooden table. The phone's glossy, dark screen vaguely reflects soft, diffuse light from an unseen window, while its matte black exterior contrasts with the warm tones of the wood desk, revealing a prominent, detailed grain. In a shallow depth of field, the out-of-focus corner of a notebook and a simple ceramic mug rest in the background, establishing an atmosphere of quiet concentration. This tranquility is abruptly shattered when an internal mechanism activates, causing the device to vibrate with an intense, humming energy, its form slightly blurred by its rapid oscillations. Driven by this force, the smartphone slides and rotates erratically across the wooden surface, its hard edges rattling against the table as it moves in an unpredictable, short arc. Just as suddenly as it began, the vibration ceases, and the phone instantly and completely stops, now resting at a new, tilted angle, allowing the scene to return to a deep, expectant silence.", + "seed": 10541, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_4/index_9984_seed_10541_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02594359335489571, + "vlm_logprob": -0.02594359335489571, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.02594359335489571, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09258559346199036 + ], + "opposite_logprob_tokens": [ + -3.842585563659668 + ], + "logprob_reward": -0.09258559346199036, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.09258559346199036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19714665412902832, + 0.03966965898871422, + 0.12359289079904556, + 0.04543871805071831, + 0.08885958045721054, + 0.03596387431025505, + 0.05133019760251045, + 0.11449011415243149, + 0.22514910995960236, + 0.0783592090010643 + ], + "credit_map": [ + 0.02875625155866146, + 0.02377835288643837, + 0.025258339941501617, + 0.013048598542809486, + 0.02675856463611126, + 0.03939823806285858, + 0.015741921961307526, + 0.024406399577856064, + 0.003049073740839958, + 0.005572228226810694, + 0.004148183390498161, + 0.0036176249850541353, + 0.004602467641234398, + 0.007744766306132078, + 0.005546537693589926, + 0.005388778634369373, + 0.0169343464076519, + 0.016143282875418663, + 0.014125554822385311, + 0.013896636664867401, + 0.02010422572493553, + 0.013949405401945114, + 0.01350395753979683, + 0.014935486018657684, + 0.005938323680311441, + 0.006187870632857084, + 0.005323414225131273, + 0.004369215574115515, + 0.006189119536429644, + 0.007866954430937767, + 0.004363304004073143, + 0.005200516432523727, + 0.006767673417925835, + 0.011922786943614483, + 0.02173616923391819, + 0.004938140511512756, + 0.028345195576548576, + 0.005082216579467058, + 0.004618353676050901, + 0.005449047312140465, + 0.003581432858482003, + 0.004577845335006714, + 0.005527973175048828, + 0.0029531412292271852, + 0.007729835342615843, + 0.006168384104967117, + 0.0023218831047415733, + 0.0031033779960125685, + 0.005563557613641024, + 0.006692500784993172, + 0.007568184286355972, + 0.004577996209263802, + 0.0064477212727069855, + 0.006544226780533791, + 0.008916218765079975, + 0.005019792355597019, + 0.009070051833987236, + 0.01890711858868599, + 0.010432168841362, + 0.01681581512093544, + 0.01684706099331379, + 0.017590006813406944, + 0.012078036554157734, + 0.012749847956001759, + 0.013614879921078682, + 0.023827271535992622, + 0.023049332201480865, + 0.040011268109083176, + 0.027490997686982155, + 0.03765307366847992, + 0.029464177787303925, + 0.030038120225071907, + 0.005059616174548864, + 0.007615959737449884, + 0.008641539141535759, + 0.020731475204229355, + 0.010998810641467571, + 0.006083804648369551, + 0.00813296064734459, + 0.011095038615167141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.09258559346199036 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032151455525308847 + ], + "opposite_logprob_tokens": [ + -5.753215312957764 + ], + "logprob_reward": -0.0032151455525308847, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032151455525308847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09421542286872864, + 0.0646253302693367, + 0.4115580916404724, + 0.01906718499958515, + 0.059732090681791306, + 0.04930293932557106, + 0.09991299360990524, + 0.08639393746852875, + 0.08937855809926987, + 0.02581351064145565 + ], + "credit_map": [ + 0.013997710309922695, + 0.012748575769364834, + 0.013644144870340824, + 0.007581081707030535, + 0.0140461390838027, + 0.010959615930914879, + 0.008350908756256104, + 0.012887229211628437, + 0.008704476058483124, + 0.009079535491764545, + 0.006979742087423801, + 0.004780410788953304, + 0.009686339646577835, + 0.00896323099732399, + 0.006323748733848333, + 0.010107842274010181, + 0.05203619226813316, + 0.05623511224985123, + 0.047074031084775925, + 0.04124928638339043, + 0.07211915403604507, + 0.050764378160238266, + 0.03899681195616722, + 0.05308310315012932, + 0.0033064025919884443, + 0.002655903110280633, + 0.0022396373096853495, + 0.0016351680969819427, + 0.0026230502407997847, + 0.0030620356556028128, + 0.0018541966564953327, + 0.001690788776613772, + 0.0046910508535802364, + 0.006445862352848053, + 0.015629585832357407, + 0.0027209294494241476, + 0.021040767431259155, + 0.0038423361256718636, + 0.0021897312253713608, + 0.0031718185637146235, + 0.004248528741300106, + 0.006034674122929573, + 0.010241948999464512, + 0.003022348741069436, + 0.013633101247251034, + 0.0060905152931809425, + 0.0020714483689516783, + 0.003960368689149618, + 0.005990668665617704, + 0.013287649489939213, + 0.012029878795146942, + 0.011333534494042397, + 0.010854162275791168, + 0.01565372571349144, + 0.017349211499094963, + 0.013414155691862106, + 0.009368731640279293, + 0.012400699779391289, + 0.010757958516478539, + 0.012127158232033253, + 0.009833082556724548, + 0.011689373292028904, + 0.009094435721635818, + 0.011122492142021656, + 0.006255393382161856, + 0.010441464371979237, + 0.009799192659556866, + 0.01535072736442089, + 0.01167371217161417, + 0.012668793089687824, + 0.010925167240202427, + 0.012264098040759563, + 0.001355196931399405, + 0.0021313675679266453, + 0.0031751750502735376, + 0.0057106357999145985, + 0.0036274499725550413, + 0.0031263381242752075, + 0.003065472235903144, + 0.003621872514486313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.0032151455525308847 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086530983448029 + ], + "opposite_logprob_tokens": [ + -4.511086463928223 + ], + "logprob_reward": -0.011086530983448029, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086530983448029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2712061107158661, + 0.13613717257976532, + 0.05803511664271355, + 0.03651420772075653, + 0.05118851363658905, + 0.20348891615867615, + 0.08135120570659637, + 0.06557177007198334, + 0.054306164383888245, + 0.04220088943839073 + ], + "credit_map": [ + 0.03872817009687424, + 0.0337846577167511, + 0.0294356532394886, + 0.019987421110272408, + 0.03787190839648247, + 0.05825773999094963, + 0.02380291186273098, + 0.029337629675865173, + 0.01631062850356102, + 0.01763155870139599, + 0.015245622955262661, + 0.009127949364483356, + 0.016989436000585556, + 0.03212961181998253, + 0.01352104265242815, + 0.015181321650743484, + 0.004138413351029158, + 0.006818616762757301, + 0.006238474976271391, + 0.005430206656455994, + 0.00997229479253292, + 0.008194142021238804, + 0.011171812191605568, + 0.00607115775346756, + 0.003085200907662511, + 0.004487868398427963, + 0.0029347643721848726, + 0.002235322492197156, + 0.004489583894610405, + 0.01215461827814579, + 0.0024468384217470884, + 0.004680013284087181, + 0.002634843112900853, + 0.0064033749513328075, + 0.00807973276823759, + 0.0037514062132686377, + 0.012837553396821022, + 0.010248061269521713, + 0.0031007598154246807, + 0.00413278304040432, + 0.011933787725865841, + 0.024335185065865517, + 0.020375851541757584, + 0.013884924352169037, + 0.03040406107902527, + 0.06510970741510391, + 0.014587749727070332, + 0.02285763993859291, + 0.005403671879321337, + 0.013690674677491188, + 0.010394832119345665, + 0.009290468879044056, + 0.009166153147816658, + 0.012524157762527466, + 0.011463340371847153, + 0.009417903609573841, + 0.005436987150460482, + 0.009837093763053417, + 0.0070562297478318214, + 0.009231655858457088, + 0.0076382337138056755, + 0.011148486286401749, + 0.00678992411121726, + 0.008433162234723568, + 0.004109248984605074, + 0.004293007776141167, + 0.006105771288275719, + 0.009703302755951881, + 0.0059766690246760845, + 0.009442545473575592, + 0.00753835029900074, + 0.0071372720412909985, + 0.0039616115391254425, + 0.003737082239240408, + 0.004167148843407631, + 0.004905147012323141, + 0.006062192376703024, + 0.008432880975306034, + 0.004596036393195391, + 0.006338789127767086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.011086530983448029 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018194444477558136 + ], + "opposite_logprob_tokens": [ + -4.018194675445557 + ], + "logprob_reward": -0.018194444477558136, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018194444477558136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13721345365047455, + 0.035422854125499725, + 0.3000694811344147, + 0.044755496084690094, + 0.047320589423179626, + 0.044242970645427704, + 0.1474808007478714, + 0.1598677933216095, + 0.0551699660718441, + 0.02845657616853714 + ], + "credit_map": [ + 0.017649149522185326, + 0.017123473808169365, + 0.01451657060533762, + 0.0099265705794096, + 0.01766560785472393, + 0.03180280327796936, + 0.01081968005746603, + 0.01770959421992302, + 0.0032492680475115776, + 0.0042682527564466, + 0.004492527339607477, + 0.002772135194391012, + 0.00405997596681118, + 0.007819388061761856, + 0.0038151389453560114, + 0.0049461680464446545, + 0.037689533084630966, + 0.03955736756324768, + 0.03397059068083763, + 0.03385050594806671, + 0.04644238203763962, + 0.0358361154794693, + 0.03289307281374931, + 0.03982989490032196, + 0.005101947113871574, + 0.0053706387989223, + 0.00447287131100893, + 0.003770295297726989, + 0.005304042715579271, + 0.009483512490987778, + 0.00501228729262948, + 0.006239902228116989, + 0.00235569616779685, + 0.0050739324651658535, + 0.011395815759897232, + 0.004525231663137674, + 0.014864112250506878, + 0.0031405643094331026, + 0.003538592718541622, + 0.002426647115498781, + 0.0027611262630671263, + 0.004400988575071096, + 0.009835501201450825, + 0.002269769785925746, + 0.014420093037188053, + 0.00554325757548213, + 0.002061747945845127, + 0.002950486494228244, + 0.012583810836076736, + 0.020794982090592384, + 0.02051055245101452, + 0.01661122590303421, + 0.016543950885534286, + 0.021526873111724854, + 0.018382620066404343, + 0.02052677981555462, + 0.01471965666860342, + 0.025868244469165802, + 0.014449318870902061, + 0.021489350125193596, + 0.024853529408574104, + 0.020252034068107605, + 0.01865077205002308, + 0.01958489790558815, + 0.0038039747159928083, + 0.005024746060371399, + 0.005789243616163731, + 0.008936700411140919, + 0.005859744269400835, + 0.008714606054127216, + 0.008845099247992039, + 0.008195851929485798, + 0.0017547707539051771, + 0.0026734345592558384, + 0.002949667163193226, + 0.006927731446921825, + 0.004205680452287197, + 0.0026814488228410482, + 0.0027209704276174307, + 0.004542873241007328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -0.018194444477558136 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636252298951149 + ], + "opposite_logprob_tokens": [ + -5.379636287689209 + ], + "logprob_reward": -0.004636252298951149, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636252298951149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07349763065576553, + 0.019611798226833344, + 0.382980614900589, + 0.041100941598415375, + 0.07251712679862976, + 0.0492551252245903, + 0.08757507055997849, + 0.11649184674024582, + 0.11565372347831726, + 0.04131612554192543 + ], + "credit_map": [ + 0.006155813578516245, + 0.010385042056441307, + 0.008865418843925, + 0.004907478112727404, + 0.010384091176092625, + 0.01800040900707245, + 0.005375902634114027, + 0.009423483163118362, + 0.0021341850515455008, + 0.002661339472979307, + 0.003180324099957943, + 0.0015285025583580136, + 0.0027387330774217844, + 0.0033527372870594263, + 0.0021680130157619715, + 0.0018479624995961785, + 0.045753709971904755, + 0.052504826337099075, + 0.04383398965001106, + 0.037260279059410095, + 0.06839506328105927, + 0.04858545958995819, + 0.035537924617528915, + 0.05110934004187584, + 0.00495400233194232, + 0.004704978317022324, + 0.0056405942887067795, + 0.002549889963120222, + 0.005230455659329891, + 0.008689748123288155, + 0.0037698904052376747, + 0.005561383906751871, + 0.004312256816774607, + 0.007594924885779619, + 0.018787724897265434, + 0.0048768771812319756, + 0.025216560810804367, + 0.004222364164888859, + 0.0045341928489506245, + 0.0029722205363214016, + 0.004138352815061808, + 0.005674326326698065, + 0.011517532169818878, + 0.0025866739451885223, + 0.01624368689954281, + 0.0038238989654928446, + 0.002482752315700054, + 0.0027879048138856888, + 0.010944143868982792, + 0.010954015888273716, + 0.013661430217325687, + 0.00852771382778883, + 0.010500825941562653, + 0.01045567449182272, + 0.010768050327897072, + 0.011763211339712143, + 0.00771935423836112, + 0.01485382579267025, + 0.015697238966822624, + 0.015169587917625904, + 0.014185101725161076, + 0.02419036068022251, + 0.009970899671316147, + 0.014705476351082325, + 0.00767477136105299, + 0.011978224851191044, + 0.01101560890674591, + 0.022207623347640038, + 0.012963655404746532, + 0.01763898693025112, + 0.012184246443212032, + 0.019990604370832443, + 0.004359550308436155, + 0.004982292652130127, + 0.005333543289452791, + 0.004528180696070194, + 0.005357628688216209, + 0.00610788818448782, + 0.005404639523476362, + 0.005242402199655771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.004636252298951149 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Viewed from a static, high-angle perspective, the scene captures a sleek black smartphone lying motionless, face-up, on a polished wooden table. The phone's glossy, dark screen vaguely reflects soft, diffuse light from an unseen window, while its matte black exterior contrasts with the warm tones of the wood desk, revealing a prominent, detailed grain. In a shallow depth of field, the out-of-focus corner of a notebook and a simple ceramic mug rest in the background, establishing an atmosphere of quiet concentration. This tranquility is abruptly shattered when an internal mechanism activates, causing the device to vibrate with an intense, humming energy, its form slightly blurred by its rapid oscillations. Driven by this force, the smartphone slides and rotates erratically across the wooden surface, its hard edges rattling against the table as it moves in an unpredictable, short arc. Just as suddenly as it began, the vibration ceases, and the phone instantly and completely stops, now resting at a new, tilted angle, allowing the scene to return to a deep, expectant silence.", + "seed": 10542, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_4/index_9984_seed_10542_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.028698648698627947, + "vlm_logprob": -0.028698648698627947, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.028698648698627947, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04777359962463379 + ], + "opposite_logprob_tokens": [ + -4.672773361206055 + ], + "logprob_reward": -0.04777359962463379, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.04777359962463379, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19176559150218964, + 0.2049284130334854, + 0.13875968754291534, + 0.0898854061961174, + 0.06037646532058716, + 0.07969629764556885, + 0.05821910873055458, + 0.05736817792057991, + 0.04376188665628433, + 0.07523898780345917 + ], + "credit_map": [ + 0.025207169353961945, + 0.021639348939061165, + 0.02612736076116562, + 0.01581083983182907, + 0.01883251965045929, + 0.034915316849946976, + 0.025163304060697556, + 0.024069737643003464, + 0.012074303813278675, + 0.045587822794914246, + 0.02970825508236885, + 0.018623996526002884, + 0.03210093453526497, + 0.02962448075413704, + 0.015272680670022964, + 0.02193593420088291, + 0.010133825242519379, + 0.015560206025838852, + 0.017587952315807343, + 0.015200459398329258, + 0.03398771584033966, + 0.019989443942904472, + 0.0120353102684021, + 0.014264770783483982, + 0.009716400876641273, + 0.012133847922086716, + 0.012273378670215607, + 0.00894441269338131, + 0.010695836506783962, + 0.017428068444132805, + 0.007579553406685591, + 0.011113908141851425, + 0.004866820760071278, + 0.005600718781352043, + 0.007433302700519562, + 0.008994636125862598, + 0.010031921789050102, + 0.0075402772054076195, + 0.01032122503966093, + 0.0055875652469694614, + 0.006810726597905159, + 0.00924330297857523, + 0.00846871081739664, + 0.00961032509803772, + 0.011344785802066326, + 0.01251001376658678, + 0.012413534335792065, + 0.009294896386563778, + 0.004331768024712801, + 0.0077961706556379795, + 0.007490126416087151, + 0.005990372039377689, + 0.01047089695930481, + 0.009761468507349491, + 0.005605674814432859, + 0.006772635038942099, + 0.004548689816147089, + 0.0063885170966386795, + 0.006617279723286629, + 0.007563482504338026, + 0.007669512182474136, + 0.011889415793120861, + 0.005000466946512461, + 0.007690813858062029, + 0.002899480750784278, + 0.005521007813513279, + 0.005692789331078529, + 0.007926879450678825, + 0.005821662023663521, + 0.005592599511146545, + 0.0058656735345721245, + 0.004441795405000448, + 0.004138040356338024, + 0.011481082066893578, + 0.011411139741539955, + 0.008739752694964409, + 0.010024622082710266, + 0.012543381191790104, + 0.008036279119551182, + 0.008864694274961948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.04777359962463379 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029840901494026184 + ], + "opposite_logprob_tokens": [ + -3.5298409461975098 + ], + "logprob_reward": -0.029840901494026184, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029840901494026184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18572776019573212, + 0.20480374991893768, + 0.042055144906044006, + 0.12514616549015045, + 0.06615555286407471, + 0.20947062969207764, + 0.0777868777513504, + 0.030167238786816597, + 0.01541698258370161, + 0.04326992109417915 + ], + "credit_map": [ + 0.02276216633617878, + 0.01766887493431568, + 0.020305076614022255, + 0.020799534395337105, + 0.02409633994102478, + 0.031098315492272377, + 0.02460668236017227, + 0.02439075894653797, + 0.011073633097112179, + 0.04309188574552536, + 0.02924717776477337, + 0.020961932837963104, + 0.03499987721443176, + 0.0280239786952734, + 0.016691341996192932, + 0.02071390300989151, + 0.003182745538651943, + 0.00566180981695652, + 0.005362438503652811, + 0.004993545822799206, + 0.006397175136953592, + 0.00542822340503335, + 0.005948053672909737, + 0.005081151612102985, + 0.013332320377230644, + 0.016621608287096024, + 0.017076265066862106, + 0.01677156798541546, + 0.020928438752889633, + 0.015231257304549217, + 0.01243455521762371, + 0.012750139459967613, + 0.006287869997322559, + 0.007190047297626734, + 0.008669106289744377, + 0.005740249063819647, + 0.0062738750129938126, + 0.014606212265789509, + 0.006589426193386316, + 0.010798764415085316, + 0.021277034655213356, + 0.024403849616646767, + 0.026770619675517082, + 0.021335596218705177, + 0.03165082260966301, + 0.03449372574687004, + 0.018434148281812668, + 0.031104810535907745, + 0.0038645316381007433, + 0.00939998496323824, + 0.01011328399181366, + 0.006367025896906853, + 0.019091518595814705, + 0.01003810204565525, + 0.010793021880090237, + 0.008119406178593636, + 0.0015684760874137282, + 0.003076831577345729, + 0.003284990321844816, + 0.003803521627560258, + 0.005314480513334274, + 0.004846197087317705, + 0.004571639932692051, + 0.003701100591570139, + 0.0011914123315364122, + 0.0013908238615840673, + 0.0015848620096221566, + 0.0026535112410783768, + 0.0021547034848481417, + 0.0021998181473463774, + 0.0022194217890501022, + 0.0020224284380674362, + 0.002682909369468689, + 0.005719990469515324, + 0.005694702733308077, + 0.005099884234368801, + 0.010170316323637962, + 0.004834584426134825, + 0.004339332226663828, + 0.0047282008454203606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.029840901494026184 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078924760222435 + ], + "opposite_logprob_tokens": [ + -4.141078948974609 + ], + "logprob_reward": -0.016078924760222435, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078924760222435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15371328592300415, + 0.11671643704175949, + 0.3930108845233917, + 0.030728688463568687, + 0.05629460886120796, + 0.06464879214763641, + 0.06743679195642471, + 0.017453519627451897, + 0.025181017816066742, + 0.0748160183429718 + ], + "credit_map": [ + 0.018740713596343994, + 0.018548646941781044, + 0.017391864210367203, + 0.017425596714019775, + 0.018913520500063896, + 0.023806799203157425, + 0.016702570021152496, + 0.02218356914818287, + 0.0082966648042202, + 0.022880075499415398, + 0.016863496974110603, + 0.010741405189037323, + 0.014656752347946167, + 0.02332060970366001, + 0.006563798524439335, + 0.013393616303801537, + 0.04765373095870018, + 0.06933780759572983, + 0.06379879266023636, + 0.02642848528921604, + 0.04553307592868805, + 0.06518223881721497, + 0.02574104815721512, + 0.04933563619852066, + 0.0032675534021109343, + 0.0045937346294522285, + 0.0041928766295313835, + 0.0033693041186779737, + 0.0043068681843578815, + 0.004573815502226353, + 0.003079758258536458, + 0.0033447775058448315, + 0.005823843646794558, + 0.005897175054997206, + 0.009609916247427464, + 0.004942667670547962, + 0.006082445848733187, + 0.00848055724054575, + 0.006563806440681219, + 0.00889419112354517, + 0.005720624700188637, + 0.0052946764044463634, + 0.007131711579859257, + 0.009313001297414303, + 0.010891268029808998, + 0.01005134079605341, + 0.006992854177951813, + 0.009253308176994324, + 0.0034356536343693733, + 0.008100192062556744, + 0.008600208908319473, + 0.005417817737907171, + 0.01615092344582081, + 0.008904809132218361, + 0.009495406411588192, + 0.007331774570047855, + 0.0013462037313729525, + 0.0017585376044735312, + 0.0023249483201652765, + 0.002811271231621504, + 0.002489760285243392, + 0.002505697077140212, + 0.0022337951231747866, + 0.00198330357670784, + 0.0020278391893953085, + 0.0019624873530119658, + 0.00283197988756001, + 0.005927813705056906, + 0.00306149129755795, + 0.0034403109457343817, + 0.0030037746764719486, + 0.0029253188986331224, + 0.004924112465232611, + 0.00844484195113182, + 0.011236203834414482, + 0.005942767485976219, + 0.016083840280771255, + 0.012217769399285316, + 0.005888159386813641, + 0.010078313760459423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.016078924760222435 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306682571768761 + ], + "opposite_logprob_tokens": [ + -3.1680667400360107 + ], + "logprob_reward": -0.04306682571768761, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306682571768761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12923100590705872, + 0.24689413607120514, + 0.07276584208011627, + 0.04721985012292862, + 0.06091424077749252, + 0.06003988906741142, + 0.03572513908147812, + 0.1474701166152954, + 0.09895608574151993, + 0.10078368335962296 + ], + "credit_map": [ + 0.011361689306795597, + 0.016155127435922623, + 0.012583677656948566, + 0.01860477216541767, + 0.016665663570165634, + 0.018404515460133553, + 0.02059965766966343, + 0.014855902642011642, + 0.01368977315723896, + 0.0643862932920456, + 0.038693904876708984, + 0.01986510306596756, + 0.04884080961346626, + 0.02379588410258293, + 0.013864021748304367, + 0.02375834248960018, + 0.004260883666574955, + 0.010515227913856506, + 0.008361134678125381, + 0.011013912037014961, + 0.009099220857024193, + 0.008085723035037518, + 0.013265898451209068, + 0.008163842372596264, + 0.00426137913018465, + 0.005865892861038446, + 0.006828142795711756, + 0.0060298559255898, + 0.007082429248839617, + 0.006811165250837803, + 0.005144582130014896, + 0.005196400918066502, + 0.00399578083306551, + 0.006521467585116625, + 0.007227093446999788, + 0.005660682916641235, + 0.011802725493907928, + 0.008974789641797543, + 0.009414254687726498, + 0.0073174466378986835, + 0.005174505524337292, + 0.004321571905165911, + 0.005594467744231224, + 0.00904041901230812, + 0.009933941066265106, + 0.010156318545341492, + 0.009552341885864735, + 0.006266324315220118, + 0.0020779711194336414, + 0.005733349826186895, + 0.003966351505368948, + 0.004702682141214609, + 0.0066740186884999275, + 0.004619363229721785, + 0.0035444162786006927, + 0.004406983498483896, + 0.010377977974712849, + 0.01315391156822443, + 0.015509477816522121, + 0.025208404287695885, + 0.014197444543242455, + 0.034199561923742294, + 0.012326333671808243, + 0.02249700203537941, + 0.006835984531790018, + 0.0088467076420784, + 0.013728044927120209, + 0.016816118732094765, + 0.010877514258027077, + 0.01658773049712181, + 0.012228148989379406, + 0.013035837560892105, + 0.0062847132794559, + 0.013659796677529812, + 0.014679597690701485, + 0.008310176432132721, + 0.021721292287111282, + 0.015799537301063538, + 0.008603072725236416, + 0.01172549370676279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.04306682571768761 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sleek black smartphone lying face-up on a polished wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are an out-of-focus notebook and a ceramic mug visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the smartphone vibrate, slide, and rotate erratically across the table before stopping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the smartphone end up resting at a new, tilted angle after the motion stops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732991896569729 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.006732991896569729, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732991896569729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1695685088634491, + 0.30317816138267517, + 0.11145098507404327, + 0.05485376715660095, + 0.07661654055118561, + 0.054135046899318695, + 0.033240243792533875, + 0.07520526647567749, + 0.028159236535429955, + 0.09359223395586014 + ], + "credit_map": [ + 0.010204418562352657, + 0.019828656688332558, + 0.014803249388933182, + 0.025430623441934586, + 0.027362147346138954, + 0.0239359550178051, + 0.030605224892497063, + 0.017398236319422722, + 0.01622515544295311, + 0.08061067014932632, + 0.04920347034931183, + 0.02498535066843033, + 0.0616433285176754, + 0.025363294407725334, + 0.016781968995928764, + 0.02836493030190468, + 0.0060122013092041016, + 0.013158202171325684, + 0.013027679175138474, + 0.018154771998524666, + 0.021612180396914482, + 0.012580730020999908, + 0.015497240237891674, + 0.01140797883272171, + 0.004932047799229622, + 0.009359385818243027, + 0.007868091575801373, + 0.005357039161026478, + 0.009675496257841587, + 0.007419858127832413, + 0.004692983813583851, + 0.005548868328332901, + 0.004743080586194992, + 0.007011495064944029, + 0.008515876717865467, + 0.011955875903367996, + 0.011379790492355824, + 0.014080715365707874, + 0.010768841952085495, + 0.00816087145358324, + 0.004282166715711355, + 0.007040439639240503, + 0.00588946370407939, + 0.006609508767724037, + 0.008942061103880405, + 0.008053875528275967, + 0.007816361263394356, + 0.005501174367964268, + 0.0016420106403529644, + 0.0036648306995630264, + 0.004406683146953583, + 0.0038600165862590075, + 0.007364256773144007, + 0.005015500355511904, + 0.00349624315276742, + 0.00379070732742548, + 0.002995818154886365, + 0.006215570494532585, + 0.00826507993042469, + 0.010395071469247341, + 0.014058737084269524, + 0.01247663889080286, + 0.011359114199876785, + 0.009439234621822834, + 0.0017462905962020159, + 0.0030606111977249384, + 0.003932766150683165, + 0.005347272381186485, + 0.0034276589285582304, + 0.003153208177536726, + 0.004318310413509607, + 0.0031731200870126486, + 0.006689824163913727, + 0.014087182469666004, + 0.013521905057132244, + 0.008744779974222183, + 0.01681193895637989, + 0.013162490911781788, + 0.01019677147269249, + 0.010377353988587856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.006732991896569729 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_5/index_tensor([1683, 1683])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_5/index_tensor([1683, 1683])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..46d5ef5347a4fc8e4104f1a01578443df04c9a8e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_5/index_tensor([1683, 1683])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot from a static, slightly elevated perspective, the extreme close-up renders the scene in a delicate watercolor and pen-and-ink style. Fine ink lines define the subject's elegant contours, while soft, translucent colors create a dreamlike quality. The photograph focuses on a pair of slender, elegant hands, their skin smooth and pale, their nails immaculately manicured and painted with a pure, neutral sheen. One hand gently cups the chin, fingers resting lightly on it, while the other carefully applies lipstick from a classic, polished black tube with a shimmering gold ribbon. The lipstick itself is a vibrant, creamy crimson, and as the bullet glides smoothly across the lower lip, the camera captures deliberate, steady movement, depositing layers of rich, saturated color. The background disappears into a beautiful, soft-focus blur of a sunlit, luxurious room; warm golden light streams from an unseen source, creating a brilliant bokeh effect and suggesting the grandeur of the mansion, the blurred gleam of gilded furniture, and the vast, airy space of a soaring ceiling. The atmosphere is one of tranquility, focused intimacy and quiet luxury, a personal ritual transformed into a living work of art.", + "seed": 33162, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_5/index_1683_seed_33162_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03683581268414855, + "vlm_logprob": -0.03683581268414855, + "vlm_yes_no_margin": 4.0749999523162845, + "vlm_token_credit": -0.03683581268414855, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09036282449960709 + ], + "opposite_logprob_tokens": [ + -2.465362787246704 + ], + "logprob_reward": -0.09036282449960709, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09036282449960709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3731353282928467, + 0.0928494930267334, + 0.15310357511043549, + 0.09857113659381866, + 0.055604226887226105, + 0.06674861907958984, + 0.03216881677508354, + 0.040403060615062714, + 0.035326529294252396, + 0.05208921432495117 + ], + "credit_map": [ + 0.034238189458847046, + 0.04846448823809624, + 0.0529988668859005, + 0.028662296012043953, + 0.03568177670240402, + 0.09302578866481781, + 0.030813543125987053, + 0.04925035312771797, + 0.011346502229571342, + 0.012701570056378841, + 0.017226673662662506, + 0.007167675532400608, + 0.008944019675254822, + 0.016996070742607117, + 0.007901351898908615, + 0.010565629228949547, + 0.013538180850446224, + 0.017327168956398964, + 0.01507358718663454, + 0.011096479371190071, + 0.023483091965317726, + 0.04333975538611412, + 0.010082762688398361, + 0.01916254498064518, + 0.007022966630756855, + 0.010751599445939064, + 0.008937213569879532, + 0.008798102848231792, + 0.020303165540099144, + 0.018631458282470703, + 0.007768955081701279, + 0.016357671469449997, + 0.005984833929687738, + 0.005729214753955603, + 0.008476641029119492, + 0.006798384711146355, + 0.006982989609241486, + 0.008986098691821098, + 0.005073591601103544, + 0.007572473026812077, + 0.006813857704401016, + 0.00773937813937664, + 0.008351931348443031, + 0.010879439301788807, + 0.006781178526580334, + 0.009116011671721935, + 0.006565416231751442, + 0.010501404292881489, + 0.0030132669489830732, + 0.0036549873184412718, + 0.004206688143312931, + 0.00409485399723053, + 0.003176153404638171, + 0.005427135154604912, + 0.0038143908604979515, + 0.0047813402488827705, + 0.0037126343231648207, + 0.004287703428417444, + 0.004648847039788961, + 0.005895290523767471, + 0.004750641994178295, + 0.007529395166784525, + 0.004330413416028023, + 0.005248136352747679, + 0.003121134825050831, + 0.004231391940265894, + 0.00346366292797029, + 0.0045548416674137115, + 0.0038935088086873293, + 0.006518636364489794, + 0.0036349676083773375, + 0.005908386316150427, + 0.004879993852227926, + 0.006146782543510199, + 0.006407449953258038, + 0.006849441211670637, + 0.007053045555949211, + 0.008046639151871204, + 0.005226440727710724, + 0.007479419931769371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.09036282449960709 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012557603418827057 + ], + "opposite_logprob_tokens": [ + -4.387557506561279 + ], + "logprob_reward": -0.012557603418827057, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012557603418827057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13674403727054596, + 0.12647758424282074, + 0.13171133399009705, + 0.1491849571466446, + 0.021536368876695633, + 0.08777960389852524, + 0.08857481926679611, + 0.09421137720346451, + 0.051984433084726334, + 0.111795574426651 + ], + "credit_map": [ + 0.010354092344641685, + 0.01297217421233654, + 0.02329397387802601, + 0.017522791400551796, + 0.0183656495064497, + 0.019889747723937035, + 0.014739598147571087, + 0.019606010988354683, + 0.009865464642643929, + 0.013141104951500893, + 0.012694393284618855, + 0.010461093857884407, + 0.014539957977831364, + 0.03353540971875191, + 0.010775470174849033, + 0.02146468497812748, + 0.009657874703407288, + 0.014661937952041626, + 0.012534238398075104, + 0.011797178536653519, + 0.0259699709713459, + 0.0318499431014061, + 0.010024412535130978, + 0.015215761959552765, + 0.012057124637067318, + 0.015273626893758774, + 0.009904845617711544, + 0.013770236633718014, + 0.03863786906003952, + 0.032330576330423355, + 0.012240367941558361, + 0.014970296993851662, + 0.0021325533743947744, + 0.0025951010175049305, + 0.002852991921827197, + 0.002618792001157999, + 0.0028806524351239204, + 0.002728336025029421, + 0.002103813923895359, + 0.003624128643423319, + 0.006636741571128368, + 0.009544800966978073, + 0.009309636428952217, + 0.01462244801223278, + 0.010545494966208935, + 0.012575911357998848, + 0.009434117935597897, + 0.015110457316040993, + 0.007265567779541016, + 0.008802843280136585, + 0.007286610547453165, + 0.014139414764940739, + 0.018289798870682716, + 0.011857767589390278, + 0.00889919325709343, + 0.012033623643219471, + 0.007648369297385216, + 0.008755221962928772, + 0.011085724458098412, + 0.009958301670849323, + 0.019433332607150078, + 0.016294755041599274, + 0.008072308264672756, + 0.012963362038135529, + 0.004520479589700699, + 0.004459585528820753, + 0.005407626274973154, + 0.004387744702398777, + 0.009940332733094692, + 0.012088966555893421, + 0.005067471880465746, + 0.006112225353717804, + 0.008853351697325706, + 0.011378784663975239, + 0.014139438048005104, + 0.01801038160920143, + 0.01304685603827238, + 0.019351767376065254, + 0.011383170261979103, + 0.01563183218240738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.012557603418827057 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818625070154667 + ], + "opposite_logprob_tokens": [ + -4.018186092376709 + ], + "logprob_reward": -0.01818625070154667, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818625070154667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15716241300106049, + 0.1168040856719017, + 0.22424490749835968, + 0.05657971650362015, + 0.023606935515999794, + 0.027199216187000275, + 0.059233158826828, + 0.20023490488529205, + 0.057440802454948425, + 0.07749387621879578 + ], + "credit_map": [ + 0.012799546122550964, + 0.018088649958372116, + 0.023526329547166824, + 0.023460201919078827, + 0.024661835283041, + 0.01636403612792492, + 0.020291117951273918, + 0.017970699816942215, + 0.010176328010857105, + 0.012649456970393658, + 0.010745331645011902, + 0.010430928319692612, + 0.02843516133725643, + 0.02192750945687294, + 0.01319890096783638, + 0.009240466170012951, + 0.017908675596117973, + 0.02237922139465809, + 0.017306972295045853, + 0.021473510190844536, + 0.05587552487850189, + 0.049056921154260635, + 0.017278511077165604, + 0.022965559735894203, + 0.00455196388065815, + 0.006785186473280191, + 0.0065346830524504185, + 0.00794532522559166, + 0.007587263826280832, + 0.010161307640373707, + 0.006860973313450813, + 0.006153014022856951, + 0.002629606518894434, + 0.0026113528292626143, + 0.0033580600284039974, + 0.0031790288630872965, + 0.0033064992167055607, + 0.0035454847384244204, + 0.002172281267121434, + 0.0028046234510838985, + 0.002132570603862405, + 0.0026694631669670343, + 0.002765239682048559, + 0.002883083187043667, + 0.0035749454982578754, + 0.004531355574727058, + 0.003240093355998397, + 0.005402464885264635, + 0.00539743946865201, + 0.005672753322869539, + 0.007024460006505251, + 0.007196708582341671, + 0.012068947777152061, + 0.00964258797466755, + 0.0043100579641759396, + 0.007920201867818832, + 0.014420723542571068, + 0.018724223598837852, + 0.02264685556292534, + 0.03454829379916191, + 0.037766680121421814, + 0.026859072968363762, + 0.02052718959748745, + 0.02474186196923256, + 0.004779178649187088, + 0.005616513080894947, + 0.006910355295985937, + 0.00528867868706584, + 0.009366227313876152, + 0.01367668341845274, + 0.00430142879486084, + 0.007501739542931318, + 0.0058672078885138035, + 0.006798040587455034, + 0.00813508965075016, + 0.009585341438651085, + 0.011346852406859398, + 0.018552200868725777, + 0.007380761671811342, + 0.009828384034335613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.01818625070154667 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203091889619827 + ], + "opposite_logprob_tokens": [ + -2.8120310306549072 + ], + "logprob_reward": -0.06203091889619827, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203091889619827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4317668676376343, + 0.13080976903438568, + 0.12905243039131165, + 0.05577386915683746, + 0.036758892238140106, + 0.06387774646282196, + 0.033361971378326416, + 0.04856715351343155, + 0.026484506204724312, + 0.04354678466916084 + ], + "credit_map": [ + 0.038363296538591385, + 0.04429765045642853, + 0.04163052514195442, + 0.03701199218630791, + 0.04859992489218712, + 0.12158846110105515, + 0.03449802100658417, + 0.0657769963145256, + 0.01104592066258192, + 0.012638545595109463, + 0.01510071475058794, + 0.011331990361213684, + 0.013583263382315636, + 0.032465994358062744, + 0.011960016563534737, + 0.022683322429656982, + 0.01104268990457058, + 0.012373149394989014, + 0.014784423634409904, + 0.014711471274495125, + 0.026799432933330536, + 0.020949391648173332, + 0.011835290119051933, + 0.01655658893287182, + 0.004380811937153339, + 0.00566614605486393, + 0.004304199479520321, + 0.006472989916801453, + 0.012193428352475166, + 0.012051133438944817, + 0.005037556402385235, + 0.005667603574693203, + 0.0030278703197836876, + 0.0034127174876630306, + 0.0038854852318763733, + 0.00382209406234324, + 0.008348215371370316, + 0.007607648614794016, + 0.0030505280010402203, + 0.003604333847761154, + 0.005147687625139952, + 0.005792001727968454, + 0.006035516504198313, + 0.009078855626285076, + 0.010826479643583298, + 0.01109342835843563, + 0.005630632396787405, + 0.01027314830571413, + 0.003525023814290762, + 0.002771935425698757, + 0.003733046818524599, + 0.003322629490867257, + 0.006455881521105766, + 0.0064162337221205235, + 0.002498147776350379, + 0.004639074206352234, + 0.003365727374330163, + 0.0043852911330759525, + 0.0044174217619001865, + 0.004826040472835302, + 0.010592848993837833, + 0.00943093840032816, + 0.005252305883914232, + 0.006296580657362938, + 0.002729708794504404, + 0.0027488181367516518, + 0.004227766767144203, + 0.0039028904866427183, + 0.0027578752487897873, + 0.0037354156374931335, + 0.002654942451044917, + 0.003727089613676071, + 0.0034853103570640087, + 0.0045305220410227776, + 0.005542750004678965, + 0.007525223307311535, + 0.0041928403079509735, + 0.007344591431319714, + 0.005241573788225651, + 0.005683973431587219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.06203091889619827 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010414659045636654 + ], + "opposite_logprob_tokens": [ + -6.876041412353516 + ], + "logprob_reward": -0.0010414659045636654, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010414659045636654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3021019399166107, + 0.10277998447418213, + 0.15563543140888214, + 0.1054973378777504, + 0.039271749556064606, + 0.1234779879450798, + 0.04231530427932739, + 0.05768371745944023, + 0.03141826391220093, + 0.03981823846697807 + ], + "credit_map": [ + 0.02617482654750347, + 0.03498227149248123, + 0.03686081990599632, + 0.023634351789951324, + 0.033815231174230576, + 0.07897868752479553, + 0.026210958138108253, + 0.041444774717092514, + 0.008711269125342369, + 0.011580812744796276, + 0.01175870280712843, + 0.00982288084924221, + 0.012628340162336826, + 0.02231615222990513, + 0.011521364562213421, + 0.014440467581152916, + 0.013266936875879765, + 0.01791619509458542, + 0.015738055109977722, + 0.011916769668459892, + 0.025112155824899673, + 0.044079456478357315, + 0.009890886954963207, + 0.017714980989694595, + 0.007847984321415424, + 0.010693933814764023, + 0.007579421158879995, + 0.00979819055646658, + 0.025041038170456886, + 0.022115208208560944, + 0.007822874002158642, + 0.014598684385418892, + 0.003896834561601281, + 0.0044409530237317085, + 0.005705021787434816, + 0.0043667396530508995, + 0.005649342201650143, + 0.006000884808599949, + 0.0032862629741430283, + 0.005925712175667286, + 0.012908128090202808, + 0.013590230606496334, + 0.016625510528683662, + 0.020843030884861946, + 0.013837028294801712, + 0.014859743416309357, + 0.011368164792656898, + 0.019446149468421936, + 0.0035081866662949324, + 0.004794870037585497, + 0.00500758271664381, + 0.005133942700922489, + 0.007996019907295704, + 0.0065219770185649395, + 0.003938385751098394, + 0.0054143392480909824, + 0.005862684454768896, + 0.0057999915443360806, + 0.007955239154398441, + 0.008681020699441433, + 0.008168261498212814, + 0.00861231330782175, + 0.005279609002172947, + 0.00732459919527173, + 0.002495408756658435, + 0.003560279496014118, + 0.0032592311035841703, + 0.0040334174409508705, + 0.0031397955026477575, + 0.005818931385874748, + 0.0035705300979316235, + 0.005540669895708561, + 0.004346459172666073, + 0.00493658147752285, + 0.0075201615691185, + 0.004056049510836601, + 0.003922607284039259, + 0.005936195142567158, + 0.0038246773183345795, + 0.005275510251522064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.0010414659045636654 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a static, slightly elevated perspective, the extreme close-up renders the scene in a delicate watercolor and pen-and-ink style. Fine ink lines define the subject's elegant contours, while soft, translucent colors create a dreamlike quality. The photograph focuses on a pair of slender, elegant hands, their skin smooth and pale, their nails immaculately manicured and painted with a pure, neutral sheen. One hand gently cups the chin, fingers resting lightly on it, while the other carefully applies lipstick from a classic, polished black tube with a shimmering gold ribbon. The lipstick itself is a vibrant, creamy crimson, and as the bullet glides smoothly across the lower lip, the camera captures deliberate, steady movement, depositing layers of rich, saturated color. The background disappears into a beautiful, soft-focus blur of a sunlit, luxurious room; warm golden light streams from an unseen source, creating a brilliant bokeh effect and suggesting the grandeur of the mansion, the blurred gleam of gilded furniture, and the vast, airy space of a soaring ceiling. The atmosphere is one of tranquility, focused intimacy and quiet luxury, a personal ritual transformed into a living work of art.", + "seed": 33163, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_5/index_1683_seed_33163_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3250442951917648, + "vlm_logprob": -0.3250442951917648, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.3250442951917648, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3149627447128296 + ], + "opposite_logprob_tokens": [ + -0.3149627447128296 + ], + "logprob_reward": -1.3149627447128296, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3149627447128296, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.32914865016937256, + 0.07018814980983734, + 0.1102859303355217, + 0.05023447051644325, + 0.09014277160167694, + 0.056802865117788315, + 0.0737142339348793, + 0.04325234889984131, + 0.04686953127384186, + 0.12936092913150787 + ], + "credit_map": [ + 0.060903504490852356, + 0.04673156887292862, + 0.07141789048910141, + 0.024437859654426575, + 0.0352497436106205, + 0.036365851759910583, + 0.016747906804084778, + 0.037294358015060425, + 0.009131109341979027, + 0.0086022038012743, + 0.014540165662765503, + 0.008358282037079334, + 0.009182061068713665, + 0.00881930161267519, + 0.004612520337104797, + 0.006942512933164835, + 0.015137271024286747, + 0.011982499621808529, + 0.02536766231060028, + 0.013527925126254559, + 0.014324411749839783, + 0.012782839126884937, + 0.007342631928622723, + 0.009820699691772461, + 0.010374880395829678, + 0.006693914532661438, + 0.012883586809039116, + 0.004559327382594347, + 0.004345067776739597, + 0.004494188819080591, + 0.0027580035384744406, + 0.004125507082790136, + 0.014571737498044968, + 0.015738610178232193, + 0.020901944488286972, + 0.005946069955825806, + 0.01474034134298563, + 0.009261563420295715, + 0.0038913164753466845, + 0.005091200582683086, + 0.007997415959835052, + 0.011394447647035122, + 0.008270421996712685, + 0.004856240935623646, + 0.007680597249418497, + 0.006752551067620516, + 0.005355267319828272, + 0.004495932720601559, + 0.009469164535403252, + 0.010624147951602936, + 0.012282118201255798, + 0.007776107173413038, + 0.011527015827596188, + 0.0071019274182617664, + 0.006278723943978548, + 0.008655034936964512, + 0.005861847661435604, + 0.004684204701334238, + 0.007329890970140696, + 0.004794749431312084, + 0.0064564854837954044, + 0.005105055868625641, + 0.0037424834445118904, + 0.005277634132653475, + 0.0066130938939750195, + 0.004912212491035461, + 0.008892910555005074, + 0.004978763870894909, + 0.008461454883217812, + 0.005186864174902439, + 0.0037522874772548676, + 0.0040719518437981606, + 0.02332933060824871, + 0.013481363654136658, + 0.027964264154434204, + 0.0163680762052536, + 0.017200807109475136, + 0.011254286393523216, + 0.009406151250004768, + 0.010356658138334751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -1.3149627447128296 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22568272054195404 + ], + "opposite_logprob_tokens": [ + -1.6006827354431152 + ], + "logprob_reward": -0.22568272054195404, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22568272054195404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24899114668369293, + 0.069150909781456, + 0.15131455659866333, + 0.10597776621580124, + 0.12450472265481949, + 0.09478288888931274, + 0.07038885354995728, + 0.03523838147521019, + 0.03269173949956894, + 0.06695909053087234 + ], + "credit_map": [ + 0.03277403488755226, + 0.03625670075416565, + 0.0381627194583416, + 0.03182360902428627, + 0.039023496210575104, + 0.020560406148433685, + 0.020471539348363876, + 0.02991863340139389, + 0.007498018443584442, + 0.01187641266733408, + 0.008777718991041183, + 0.013114400207996368, + 0.006360730156302452, + 0.004932136740535498, + 0.006511296611279249, + 0.010080195032060146, + 0.023208338767290115, + 0.015392809174954891, + 0.037913765758275986, + 0.016086457297205925, + 0.020994538441300392, + 0.016777118667960167, + 0.008901836350560188, + 0.012039699591696262, + 0.018188752233982086, + 0.01388123631477356, + 0.01215468067675829, + 0.013651089742779732, + 0.016715120524168015, + 0.013926518149673939, + 0.006055065430700779, + 0.01140530314296484, + 0.016497034579515457, + 0.02199002355337143, + 0.017407942563295364, + 0.01127759087830782, + 0.02688387967646122, + 0.009309589862823486, + 0.009164274670183659, + 0.011974385008215904, + 0.010575155727565289, + 0.012678690254688263, + 0.0242475513368845, + 0.010221891105175018, + 0.010181069374084473, + 0.015074235387146473, + 0.003481341991573572, + 0.008322952315211296, + 0.01001020148396492, + 0.012095049023628235, + 0.008355073630809784, + 0.0069146002642810345, + 0.013319727033376694, + 0.005781969986855984, + 0.0063980938866734505, + 0.007514137774705887, + 0.00474499212577939, + 0.0037196672055870295, + 0.004190573003143072, + 0.00509355403482914, + 0.007770052179694176, + 0.0032337140291929245, + 0.0031099210027605295, + 0.00337590673007071, + 0.004796361085027456, + 0.0037688727024942636, + 0.004136073403060436, + 0.004398874007165432, + 0.004829619079828262, + 0.004109002649784088, + 0.0030793987680226564, + 0.0035735396668314934, + 0.008570351637899876, + 0.007838641293346882, + 0.010269837453961372, + 0.015643959864974022, + 0.008556116372346878, + 0.005189018789678812, + 0.005710406694561243, + 0.005180754698812962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.22568272054195404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308497905731201 + ], + "opposite_logprob_tokens": [ + -3.1680850982666016 + ], + "logprob_reward": -0.04308497905731201, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308497905731201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19318093359470367, + 0.11454316973686218, + 0.28244221210479736, + 0.04514921456575394, + 0.0848054587841034, + 0.08187219500541687, + 0.046832431107759476, + 0.07199709862470627, + 0.024594349786639214, + 0.05458291247487068 + ], + "credit_map": [ + 0.028799792751669884, + 0.03270233795046806, + 0.03014288656413555, + 0.017555970698595047, + 0.03201524168252945, + 0.019362105056643486, + 0.011558516882359982, + 0.021044079214334488, + 0.013769383542239666, + 0.019087113440036774, + 0.01858002133667469, + 0.0140046002343297, + 0.013526861555874348, + 0.012403890490531921, + 0.007646908052265644, + 0.015524392016232014, + 0.036774277687072754, + 0.031105758622288704, + 0.05444133281707764, + 0.036328159272670746, + 0.040510065853595734, + 0.027294345200061798, + 0.029218681156635284, + 0.026769598945975304, + 0.007283459417521954, + 0.005970574449747801, + 0.005988326855003834, + 0.007949833758175373, + 0.005463737994432449, + 0.0041417283937335014, + 0.003030583029612899, + 0.005320970434695482, + 0.010488328523933887, + 0.01701623760163784, + 0.014714821241796017, + 0.010268882848322392, + 0.01065064501017332, + 0.007471517194062471, + 0.005137724801898003, + 0.009057295508682728, + 0.010900449007749557, + 0.013980789110064507, + 0.011141953058540821, + 0.008501660078763962, + 0.014269337058067322, + 0.007163644768297672, + 0.009428507648408413, + 0.006485850550234318, + 0.005060737021267414, + 0.00800296850502491, + 0.006048052106052637, + 0.005058815702795982, + 0.00532893231138587, + 0.005722424015402794, + 0.006474992725998163, + 0.005135505460202694, + 0.00892886333167553, + 0.0071081784553825855, + 0.011061816476285458, + 0.0069616069085896015, + 0.015426275320351124, + 0.007710590958595276, + 0.006722580175846815, + 0.008077187463641167, + 0.0038699961733072996, + 0.0026323911733925343, + 0.003430167445912957, + 0.0032328939996659756, + 0.003931157756596804, + 0.0024550259113311768, + 0.0023758928291499615, + 0.0026668233331292868, + 0.0064261252991855145, + 0.0071387211792171, + 0.006905946880578995, + 0.013265935704112053, + 0.006544073577970266, + 0.0044913748279213905, + 0.004487418569624424, + 0.005323314107954502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.04308497905731201 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018196435645222664 + ], + "opposite_logprob_tokens": [ + -4.0181965827941895 + ], + "logprob_reward": -0.018196435645222664, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018196435645222664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2815486490726471, + 0.08325327932834625, + 0.05676349624991417, + 0.06309774518013, + 0.11914242058992386, + 0.06107278913259506, + 0.056194718927145004, + 0.10404187440872192, + 0.08526916056871414, + 0.08961591869592667 + ], + "credit_map": [ + 0.03726363554596901, + 0.04776680842041969, + 0.038957882672548294, + 0.025699520483613014, + 0.05412427335977554, + 0.02569129317998886, + 0.021395288407802582, + 0.03064991533756256, + 0.008633902296423912, + 0.011960591189563274, + 0.00932453665882349, + 0.011474906466901302, + 0.01519444677978754, + 0.008525470271706581, + 0.009326819330453873, + 0.008812599815428257, + 0.00966748408973217, + 0.008946823887526989, + 0.009448523633182049, + 0.0070550222881138325, + 0.00598911615088582, + 0.005785117857158184, + 0.004077086690813303, + 0.005794314667582512, + 0.00941876694560051, + 0.00909271277487278, + 0.006472229957580566, + 0.006797580514103174, + 0.00950226653367281, + 0.009456534869968891, + 0.005139394197613001, + 0.007218253333121538, + 0.01577068120241165, + 0.01820448599755764, + 0.016383567824959755, + 0.008455770090222359, + 0.03273727372288704, + 0.01064144354313612, + 0.007718079723417759, + 0.00923110917210579, + 0.00972391664981842, + 0.010551565326750278, + 0.009382187388837337, + 0.004975572228431702, + 0.01225504744797945, + 0.004563868977129459, + 0.004392427392303944, + 0.0052282013930380344, + 0.0072354767471551895, + 0.009782802313566208, + 0.005842549726366997, + 0.005537251941859722, + 0.011294643394649029, + 0.004545426927506924, + 0.0068140896037220955, + 0.005142476409673691, + 0.012156162410974503, + 0.01568296179175377, + 0.017294414341449738, + 0.017425185069441795, + 0.016060341149568558, + 0.009824734181165695, + 0.007442292291671038, + 0.008155778050422668, + 0.010720103979110718, + 0.011588173918426037, + 0.011402977630496025, + 0.009196065366268158, + 0.009689395315945148, + 0.011107084341347218, + 0.010943499393761158, + 0.010621847584843636, + 0.008996858261525631, + 0.010081551037728786, + 0.01285114698112011, + 0.01437765546143055, + 0.012202625162899494, + 0.00811343640089035, + 0.010860716924071312, + 0.012131919153034687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.018196435645222664 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294596001505852 + ], + "opposite_logprob_tokens": [ + -3.773294687271118 + ], + "logprob_reward": -0.023294596001505852, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294596001505852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1849672645330429, + 0.05842244625091553, + 0.09055738151073456, + 0.04478197917342186, + 0.12244316935539246, + 0.12017524242401123, + 0.04811825603246689, + 0.048191316425800323, + 0.04301099106669426, + 0.23933196067810059 + ], + "credit_map": [ + 0.036619286984205246, + 0.025311443954706192, + 0.0354628823697567, + 0.01470355037599802, + 0.020325420424342155, + 0.01811066083610058, + 0.011135385371744633, + 0.023298613727092743, + 0.008342553861439228, + 0.008549229241907597, + 0.00938535574823618, + 0.007453364320099354, + 0.00863999966531992, + 0.006089947186410427, + 0.004006373696029186, + 0.005955612286925316, + 0.014534931629896164, + 0.008539699018001556, + 0.02581089176237583, + 0.008452566340565681, + 0.008812569081783295, + 0.012682710774242878, + 0.004924837499856949, + 0.00679917074739933, + 0.005927106365561485, + 0.005771312862634659, + 0.008120224811136723, + 0.004811688791960478, + 0.005951179191470146, + 0.003901127027347684, + 0.004040522500872612, + 0.006258814595639706, + 0.02356574870646, + 0.017066817730665207, + 0.034215863794088364, + 0.009329942055046558, + 0.009773767553269863, + 0.016315607354044914, + 0.004667019937187433, + 0.007508394308388233, + 0.019090769812464714, + 0.01610911637544632, + 0.022773733362555504, + 0.013950006105005741, + 0.008959065191447735, + 0.022716021165251732, + 0.006209279410541058, + 0.010367246344685555, + 0.006022683810442686, + 0.0067885806784033775, + 0.00664972560480237, + 0.005000143777579069, + 0.007915518246591091, + 0.0049290405586361885, + 0.004897313192486763, + 0.005915245972573757, + 0.007242649327963591, + 0.00538540817797184, + 0.007706404197961092, + 0.004401457495987415, + 0.00898344162851572, + 0.004921082407236099, + 0.0035709107760339975, + 0.00597996124997735, + 0.009826571680605412, + 0.004582050256431103, + 0.009046480059623718, + 0.004344090353697538, + 0.004561969079077244, + 0.004152866546064615, + 0.002152434317395091, + 0.004344522953033447, + 0.05177547410130501, + 0.024294424802064896, + 0.06159688159823418, + 0.019189663231372833, + 0.02442629635334015, + 0.024859141558408737, + 0.01261057611554861, + 0.02057947777211666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.023294596001505852 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a static, slightly elevated perspective, the extreme close-up renders the scene in a delicate watercolor and pen-and-ink style. Fine ink lines define the subject's elegant contours, while soft, translucent colors create a dreamlike quality. The photograph focuses on a pair of slender, elegant hands, their skin smooth and pale, their nails immaculately manicured and painted with a pure, neutral sheen. One hand gently cups the chin, fingers resting lightly on it, while the other carefully applies lipstick from a classic, polished black tube with a shimmering gold ribbon. The lipstick itself is a vibrant, creamy crimson, and as the bullet glides smoothly across the lower lip, the camera captures deliberate, steady movement, depositing layers of rich, saturated color. The background disappears into a beautiful, soft-focus blur of a sunlit, luxurious room; warm golden light streams from an unseen source, creating a brilliant bokeh effect and suggesting the grandeur of the mansion, the blurred gleam of gilded furniture, and the vast, airy space of a soaring ceiling. The atmosphere is one of tranquility, focused intimacy and quiet luxury, a personal ritual transformed into a living work of art.", + "seed": 33164, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_5/index_1683_seed_33164_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1624789234250784, + "vlm_logprob": -0.1624789234250784, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.1624789234250784, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5803229808807373 + ], + "opposite_logprob_tokens": [ + -0.8303229808807373 + ], + "logprob_reward": -0.5803229808807373, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5803229808807373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3005430996417999, + 0.1295185536146164, + 0.07841546088457108, + 0.10332050174474716, + 0.04347696900367737, + 0.025213193148374557, + 0.07334617525339127, + 0.06350673735141754, + 0.08879813551902771, + 0.0938611701130867 + ], + "credit_map": [ + 0.02574271522462368, + 0.056730251759290695, + 0.05233273282647133, + 0.043294262140989304, + 0.02428646944463253, + 0.04063905403017998, + 0.03181026503443718, + 0.025707324966788292, + 0.01164791826158762, + 0.025524882599711418, + 0.02191801741719246, + 0.024103807285428047, + 0.01160532794892788, + 0.015334797091782093, + 0.009712745435535908, + 0.009671046398580074, + 0.008134940639138222, + 0.008222914300858974, + 0.013397546485066414, + 0.011934088543057442, + 0.010599065572023392, + 0.005336215253919363, + 0.013292720541357994, + 0.007497972343116999, + 0.00793139636516571, + 0.013348132371902466, + 0.01366407610476017, + 0.019154485315084457, + 0.02282780036330223, + 0.0072046746499836445, + 0.011754979379475117, + 0.007434961386024952, + 0.003815373405814171, + 0.006568048149347305, + 0.007415159605443478, + 0.0085962675511837, + 0.00450314674526453, + 0.003753658849745989, + 0.004805047530680895, + 0.0040202671661973, + 0.0025583452079445124, + 0.0036007787566632032, + 0.0032446756958961487, + 0.0038680073339492083, + 0.0027384234126657248, + 0.0022740582935512066, + 0.004204771015793085, + 0.002724133199080825, + 0.007368726190179586, + 0.009845534339547157, + 0.013330969959497452, + 0.01571909338235855, + 0.006738664582371712, + 0.006095387041568756, + 0.007676542270928621, + 0.006571256089955568, + 0.0065903193317353725, + 0.00993926078081131, + 0.011523670516908169, + 0.008446709252893925, + 0.009331255219876766, + 0.0053230165503919125, + 0.007409685291349888, + 0.004942819010466337, + 0.007329236716032028, + 0.00964474305510521, + 0.013265672139823437, + 0.015291395597159863, + 0.007871349342167377, + 0.007367018144577742, + 0.015396039932966232, + 0.012632683850824833, + 0.00887301005423069, + 0.011713959276676178, + 0.011168106459081173, + 0.010956075973808765, + 0.011058619245886803, + 0.009267748333513737, + 0.0166257806122303, + 0.014197862707078457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.5803229808807373 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03374570235610008 + ], + "opposite_logprob_tokens": [ + -3.408745765686035 + ], + "logprob_reward": -0.03374570235610008, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03374570235610008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2887285053730011, + 0.13096648454666138, + 0.09836474806070328, + 0.09887595474720001, + 0.060224831104278564, + 0.03128468245267868, + 0.06961636990308762, + 0.05071596801280975, + 0.05884367227554321, + 0.11237876117229462 + ], + "credit_map": [ + 0.024105923250317574, + 0.058656513690948486, + 0.04430633783340454, + 0.05969778448343277, + 0.025707511231303215, + 0.026450587436556816, + 0.03442392870783806, + 0.015379950404167175, + 0.010376336984336376, + 0.027795610949397087, + 0.022155513986945152, + 0.018974926322698593, + 0.010212032124400139, + 0.015507405623793602, + 0.012586814351379871, + 0.013357855379581451, + 0.008008883334696293, + 0.017544275149703026, + 0.01862606592476368, + 0.01580742932856083, + 0.008345617912709713, + 0.009310645051300526, + 0.01015855185687542, + 0.010563291609287262, + 0.007446311414241791, + 0.013825828209519386, + 0.01431257463991642, + 0.02156950905919075, + 0.016107071191072464, + 0.007156932260841131, + 0.010047465562820435, + 0.008410273119807243, + 0.006637807469815016, + 0.007519274018704891, + 0.011821744032204151, + 0.011048334650695324, + 0.007798593491315842, + 0.005390174686908722, + 0.005247161723673344, + 0.004761749878525734, + 0.004031862132251263, + 0.0036865631118416786, + 0.0063631534576416016, + 0.005309424363076687, + 0.0032392528373748064, + 0.0021666688844561577, + 0.0035714085679501295, + 0.0029163516592234373, + 0.006787542253732681, + 0.011394376866519451, + 0.011538033373653889, + 0.008515115827322006, + 0.007520895451307297, + 0.005845046602189541, + 0.010784466750919819, + 0.007230905815958977, + 0.006498186849057674, + 0.007966297678649426, + 0.007024967111647129, + 0.005973102059215307, + 0.008664392866194248, + 0.005142797715961933, + 0.004183379467576742, + 0.005262849852442741, + 0.004441063851118088, + 0.007451323326677084, + 0.008086908608675003, + 0.011539923958480358, + 0.006013983860611916, + 0.005671036895364523, + 0.01108191255480051, + 0.004557523410767317, + 0.009783212095499039, + 0.01437913067638874, + 0.012562665157020092, + 0.01387528982013464, + 0.011904064565896988, + 0.011418606154620647, + 0.02020869217813015, + 0.018247107043862343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.03374570235610008 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009793094359338284 + ], + "opposite_logprob_tokens": [ + -4.634793281555176 + ], + "logprob_reward": -0.009793094359338284, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009793094359338284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23957356810569763, + 0.15894772112369537, + 0.2583986818790436, + 0.10919477045536041, + 0.03766219690442085, + 0.04164744168519974, + 0.0337379053235054, + 0.03450068458914757, + 0.04469970986247063, + 0.04163730889558792 + ], + "credit_map": [ + 0.019637329503893852, + 0.05395430698990822, + 0.028604550287127495, + 0.04056877642869949, + 0.025639422237873077, + 0.03323615714907646, + 0.020390402525663376, + 0.017542632296681404, + 0.013557075522840023, + 0.03470143303275108, + 0.026565123349428177, + 0.023698311299085617, + 0.013496783562004566, + 0.01752537116408348, + 0.01400868408381939, + 0.015394944697618484, + 0.02794942446053028, + 0.04093797132372856, + 0.0333283394575119, + 0.03567571938037872, + 0.04206579551100731, + 0.01915613003075123, + 0.04034971073269844, + 0.018935587257146835, + 0.008252814412117004, + 0.015457040630280972, + 0.014890183694660664, + 0.02356775850057602, + 0.021035879850387573, + 0.007912054657936096, + 0.010692315176129341, + 0.007386722601950169, + 0.00333096063695848, + 0.005245288368314505, + 0.007526816334575415, + 0.008733291178941727, + 0.0035244000609964132, + 0.0028924462385475636, + 0.003934716805815697, + 0.002474277513101697, + 0.003938791807740927, + 0.006024579983204603, + 0.004268179647624493, + 0.005598105955868959, + 0.004547527059912682, + 0.0035497057251632214, + 0.009490130469202995, + 0.004230421967804432, + 0.0030389868188649416, + 0.0038730220403522253, + 0.005916725378483534, + 0.004893298726528883, + 0.003157305298373103, + 0.0029285019263625145, + 0.006491214036941528, + 0.003438851097598672, + 0.003906888421624899, + 0.004429862834513187, + 0.006369506474584341, + 0.0049910275265574455, + 0.0038722690660506487, + 0.0028076444286853075, + 0.004605344962328672, + 0.0035181387793272734, + 0.0034930678084492683, + 0.007028271444141865, + 0.005039849318563938, + 0.008018889464437962, + 0.004885922651737928, + 0.005397961474955082, + 0.00698090111836791, + 0.0038548456504940987, + 0.0038216428365558386, + 0.005504042375832796, + 0.0055937012657523155, + 0.004888614639639854, + 0.004574344493448734, + 0.0036141015589237213, + 0.006332325283437967, + 0.007308539003133774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.009793094359338284 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.179892435669899 + ], + "opposite_logprob_tokens": [ + -1.8048924207687378 + ], + "logprob_reward": -0.179892435669899, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.179892435669899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16626088321208954, + 0.18365104496479034, + 0.05618331581354141, + 0.09105539321899414, + 0.1019148901104927, + 0.064399853348732, + 0.073920838534832, + 0.028149886056780815, + 0.0949430987238884, + 0.13952088356018066 + ], + "credit_map": [ + 0.012127709574997425, + 0.025184843689203262, + 0.02025659754872322, + 0.027320843189954758, + 0.018539302051067352, + 0.025179419666528702, + 0.023705702275037766, + 0.013946476392447948, + 0.013928811997175217, + 0.035761792212724686, + 0.0318729467689991, + 0.028373660519719124, + 0.0193390641361475, + 0.016909098252654076, + 0.01969040557742119, + 0.017775261774659157, + 0.006347602233290672, + 0.006789838895201683, + 0.009468740783631802, + 0.007276426535099745, + 0.005692838691174984, + 0.007822639308869839, + 0.0073993150144815445, + 0.005385915748775005, + 0.00701198261231184, + 0.011919836513698101, + 0.011374784633517265, + 0.01297890767455101, + 0.02240118570625782, + 0.0060457573272287846, + 0.010840378701686859, + 0.00848255306482315, + 0.008360463194549084, + 0.013153580948710442, + 0.014895834028720856, + 0.02150692790746689, + 0.008483770303428173, + 0.0075319125317037106, + 0.016828496009111404, + 0.011153904721140862, + 0.005249041132628918, + 0.007385885808616877, + 0.010458607226610184, + 0.010321546345949173, + 0.0068758390843868256, + 0.0043615857139229774, + 0.012511912733316422, + 0.007235435768961906, + 0.007682551629841328, + 0.009549765847623348, + 0.013697092421352863, + 0.011172072030603886, + 0.007046337239444256, + 0.005704313516616821, + 0.011313987895846367, + 0.0077547128312289715, + 0.0026091253384947777, + 0.00609612138941884, + 0.003406290430575609, + 0.002575912745669484, + 0.0029128689784556627, + 0.002998062176629901, + 0.0034760029520839453, + 0.004075500648468733, + 0.007574970833957195, + 0.010889529250562191, + 0.011387785896658897, + 0.011440646834671497, + 0.009222365915775299, + 0.007744798436760902, + 0.025252865627408028, + 0.011430133134126663, + 0.010654597543179989, + 0.017402611672878265, + 0.015610786154866219, + 0.013178086839616299, + 0.01981823518872261, + 0.011710903607308865, + 0.029002755880355835, + 0.022142907604575157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.179892435669899 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640403859317303 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640403859317303, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640403859317303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33436334133148193, + 0.30873236060142517, + 0.03974044695496559, + 0.037431977689266205, + 0.029652053490281105, + 0.019096866250038147, + 0.05434897914528847, + 0.054086100310087204, + 0.0772513821721077, + 0.045296505093574524 + ], + "credit_map": [ + 0.02592470310628414, + 0.0549849234521389, + 0.04371000453829765, + 0.07321428507566452, + 0.03571312129497528, + 0.041652973741292953, + 0.03374253213405609, + 0.025420790538191795, + 0.028302205726504326, + 0.04566028341650963, + 0.05657636746764183, + 0.051443345844745636, + 0.037872314453125, + 0.02702031284570694, + 0.032129090279340744, + 0.029728421941399574, + 0.003634057939052582, + 0.005289067979902029, + 0.007234991993755102, + 0.008278687484562397, + 0.00348874949850142, + 0.00380001706071198, + 0.004338069353252649, + 0.0036768047139048576, + 0.0026378801558166742, + 0.0056628272868692875, + 0.004961916711181402, + 0.006566887255758047, + 0.007556729018688202, + 0.002818932756781578, + 0.004308586940169334, + 0.0029182182624936104, + 0.0021985312923789024, + 0.004677094519138336, + 0.005679845344275236, + 0.006288798060268164, + 0.0023464097175747156, + 0.002751170424744487, + 0.003300222335383296, + 0.0024099794682115316, + 0.0019118429627269506, + 0.003011245047673583, + 0.003008059225976467, + 0.003468515584245324, + 0.0020270105451345444, + 0.0012970290845260024, + 0.00295228767208755, + 0.0014208771754056215, + 0.0049282279796898365, + 0.007443924900144339, + 0.011160151101648808, + 0.011322884820401669, + 0.005847479682415724, + 0.0030929595232009888, + 0.005808616988360882, + 0.004744730889797211, + 0.0060000941157341, + 0.007295500952750444, + 0.010122653096914291, + 0.006690713576972485, + 0.009163467213511467, + 0.005077108275145292, + 0.006178524345159531, + 0.0035580419935286045, + 0.007410428952425718, + 0.010207999497652054, + 0.013251464813947678, + 0.017511852085590363, + 0.00804153922945261, + 0.007075141184031963, + 0.008167382329702377, + 0.005585573147982359, + 0.00421550078317523, + 0.005863932892680168, + 0.00563876423984766, + 0.005551780574023724, + 0.0049294098280370235, + 0.004011021926999092, + 0.007238489110022783, + 0.00784760806709528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.008640403859317303 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a static, slightly elevated perspective, the extreme close-up renders the scene in a delicate watercolor and pen-and-ink style. Fine ink lines define the subject's elegant contours, while soft, translucent colors create a dreamlike quality. The photograph focuses on a pair of slender, elegant hands, their skin smooth and pale, their nails immaculately manicured and painted with a pure, neutral sheen. One hand gently cups the chin, fingers resting lightly on it, while the other carefully applies lipstick from a classic, polished black tube with a shimmering gold ribbon. The lipstick itself is a vibrant, creamy crimson, and as the bullet glides smoothly across the lower lip, the camera captures deliberate, steady movement, depositing layers of rich, saturated color. The background disappears into a beautiful, soft-focus blur of a sunlit, luxurious room; warm golden light streams from an unseen source, creating a brilliant bokeh effect and suggesting the grandeur of the mansion, the blurred gleam of gilded furniture, and the vast, airy space of a soaring ceiling. The atmosphere is one of tranquility, focused intimacy and quiet luxury, a personal ritual transformed into a living work of art.", + "seed": 33165, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_5/index_1683_seed_33165_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.058083906676620244, + "vlm_logprob": -0.058083906676620244, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.058083906676620244, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11749042570590973 + ], + "opposite_logprob_tokens": [ + -2.242490530014038 + ], + "logprob_reward": -0.11749042570590973, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11749042570590973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24404583871364594, + 0.13021200895309448, + 0.0367027223110199, + 0.05361052230000496, + 0.0938088521361351, + 0.05891011282801628, + 0.0549713633954525, + 0.1070646345615387, + 0.10414408892393112, + 0.11652984470129013 + ], + "credit_map": [ + 0.024127056822180748, + 0.03633032739162445, + 0.03093249537050724, + 0.03653344139456749, + 0.014458544552326202, + 0.02813725359737873, + 0.037826646119356155, + 0.03570007160305977, + 0.016572732478380203, + 0.032801926136016846, + 0.018245017156004906, + 0.015556241385638714, + 0.0060998182743787766, + 0.010610444471240044, + 0.014102576300501823, + 0.0162232406437397, + 0.004485254175961018, + 0.005154693499207497, + 0.005357297137379646, + 0.005879351403564215, + 0.0018361109541729093, + 0.00339896441437304, + 0.004470574203878641, + 0.006120476871728897, + 0.006666135508567095, + 0.008691196329891682, + 0.007864611223340034, + 0.010069827549159527, + 0.0029200385324656963, + 0.004647187888622284, + 0.006913993041962385, + 0.005837531294673681, + 0.011779186315834522, + 0.01114019937813282, + 0.0135386623442173, + 0.012768412940204144, + 0.007494818419218063, + 0.011956286616623402, + 0.012461892329156399, + 0.012669394724071026, + 0.007200091611593962, + 0.00888753030449152, + 0.009451502002775669, + 0.007131817750632763, + 0.004654229618608952, + 0.007365778554230928, + 0.006369550712406635, + 0.007849611341953278, + 0.005353617016226053, + 0.009099399670958519, + 0.008795625530183315, + 0.008983984589576721, + 0.0038075833581387997, + 0.006038804072886705, + 0.006099892780184746, + 0.006792459171265364, + 0.010688677430152893, + 0.014571522362530231, + 0.014202157966792583, + 0.014275952242314816, + 0.006834737956523895, + 0.01134446356445551, + 0.02018263190984726, + 0.01496448926627636, + 0.0112361004576087, + 0.01975267007946968, + 0.015432048588991165, + 0.01779204234480858, + 0.008641788735985756, + 0.007207198068499565, + 0.012992299161851406, + 0.011089942418038845, + 0.01337744016200304, + 0.020809747278690338, + 0.017740868031978607, + 0.02171599306166172, + 0.007609587628394365, + 0.007852382026612759, + 0.01611832156777382, + 0.01130551379173994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.11749042570590973 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906367421150208 + ], + "opposite_logprob_tokens": [ + -2.4640636444091797 + ], + "logprob_reward": -0.08906367421150208, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906367421150208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06771776080131531, + 0.06743098795413971, + 0.14535103738307953, + 0.08477946370840073, + 0.05344280228018761, + 0.07409629970788956, + 0.05687284842133522, + 0.05661144107580185, + 0.299873948097229, + 0.0938233956694603 + ], + "credit_map": [ + 0.00774463452398777, + 0.009300212375819683, + 0.008657424710690975, + 0.01259828545153141, + 0.004602853674441576, + 0.006800319999456406, + 0.00994018092751503, + 0.008073847740888596, + 0.007190209813416004, + 0.014962898567318916, + 0.008352071046829224, + 0.00978291779756546, + 0.003845471888780594, + 0.006116217467933893, + 0.007165031507611275, + 0.010016171261668205, + 0.017726978287100792, + 0.02397180162370205, + 0.021152643486857414, + 0.031172532588243484, + 0.0076645188964903355, + 0.01046372763812542, + 0.019128238782286644, + 0.014070579782128334, + 0.011609485372900963, + 0.012568740174174309, + 0.011201320216059685, + 0.013325585052371025, + 0.005389723926782608, + 0.007241108454763889, + 0.013744133524596691, + 0.009699362330138683, + 0.0056269122287631035, + 0.011121862567961216, + 0.007259595673531294, + 0.00822098646312952, + 0.003521140431985259, + 0.005077694077044725, + 0.00784093514084816, + 0.00477367639541626, + 0.009108842350542545, + 0.008510707877576351, + 0.010884452611207962, + 0.01371126901358366, + 0.004830119200050831, + 0.006558793596923351, + 0.01025883387774229, + 0.01023328211158514, + 0.006784074939787388, + 0.008939272724092007, + 0.007798023987561464, + 0.009263808839023113, + 0.004534543491899967, + 0.007154756225645542, + 0.005032182205468416, + 0.007366185076534748, + 0.005482162348926067, + 0.010273519903421402, + 0.008621023036539555, + 0.008594450540840626, + 0.004720002878457308, + 0.005591982509940863, + 0.0076225195080041885, + 0.00570578221231699, + 0.02968434989452362, + 0.06117723882198334, + 0.03964361548423767, + 0.04470490291714668, + 0.028258154168725014, + 0.025058045983314514, + 0.04112501069903374, + 0.030222654342651367, + 0.011615928262472153, + 0.01566249504685402, + 0.015191142447292805, + 0.018685320392251015, + 0.0050985561683773994, + 0.008810979314148426, + 0.012689022347331047, + 0.006069952622056007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.08906367421150208 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200941659510136 + ], + "opposite_logprob_tokens": [ + -4.2642011642456055 + ], + "logprob_reward": -0.014200941659510136, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200941659510136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11742537468671799, + 0.12842155992984772, + 0.11071828752756119, + 0.08439900726079941, + 0.06861420720815659, + 0.04134921357035637, + 0.040172211825847626, + 0.1437150537967682, + 0.14129069447517395, + 0.12389436364173889 + ], + "credit_map": [ + 0.011680510826408863, + 0.020088186487555504, + 0.015354033559560776, + 0.02692919597029686, + 0.006463637575507164, + 0.010632471181452274, + 0.013794745318591595, + 0.012482589110732079, + 0.019070401787757874, + 0.01967390440404415, + 0.017475416883826256, + 0.01958853378891945, + 0.0074104988016188145, + 0.010060750879347324, + 0.018855268135666847, + 0.016286786645650864, + 0.013533041812479496, + 0.02298402413725853, + 0.015781737864017487, + 0.02321632020175457, + 0.005652675870805979, + 0.0073142629116773605, + 0.012922132387757301, + 0.009314080700278282, + 0.010265686549246311, + 0.023731831461191177, + 0.011770810931921005, + 0.01584039069712162, + 0.0035246440675109625, + 0.006088865455240011, + 0.006514658685773611, + 0.006662118248641491, + 0.00785713829100132, + 0.008951133117079735, + 0.010176673531532288, + 0.010229953564703465, + 0.004862964618951082, + 0.007288592401891947, + 0.009071714244782925, + 0.01017603650689125, + 0.004499993287026882, + 0.004695866256952286, + 0.006783061195164919, + 0.005194558296352625, + 0.003277281066402793, + 0.006278418470174074, + 0.005275071132928133, + 0.005344966892153025, + 0.0038192023057490587, + 0.007324814796447754, + 0.00524860480800271, + 0.007388468831777573, + 0.003099881811067462, + 0.003580891527235508, + 0.00484937010332942, + 0.004860975779592991, + 0.012469791807234287, + 0.01825372874736786, + 0.02044503390789032, + 0.022122425958514214, + 0.01163653377443552, + 0.012809057720005512, + 0.028645453974604607, + 0.01733304001390934, + 0.013669964857399464, + 0.017248444259166718, + 0.02042192406952381, + 0.028309207409620285, + 0.01385736558586359, + 0.012487541884183884, + 0.020783696323633194, + 0.014512551948428154, + 0.013495401479303837, + 0.021110491827130318, + 0.018647629767656326, + 0.01947171799838543, + 0.007581604178994894, + 0.009523655287921429, + 0.022844182327389717, + 0.01121967937797308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.014200941659510136 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203708052635193 + ], + "opposite_logprob_tokens": [ + -2.8120369911193848 + ], + "logprob_reward": -0.06203708052635193, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203708052635193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04646652564406395, + 0.05634456127882004, + 0.04878493770956993, + 0.06791108101606369, + 0.0879514291882515, + 0.05556948110461235, + 0.04000811651349068, + 0.05565984919667244, + 0.3227553963661194, + 0.21854865550994873 + ], + "credit_map": [ + 0.004252173006534576, + 0.010432451032102108, + 0.006242494098842144, + 0.006203820928931236, + 0.0036863316781818867, + 0.00436647841706872, + 0.0052686347626149654, + 0.0060141440480947495, + 0.009657368063926697, + 0.006709423381835222, + 0.008561122231185436, + 0.010891402140259743, + 0.002703865757212043, + 0.004969629924744368, + 0.00661254720762372, + 0.006239206995815039, + 0.009685501456260681, + 0.005894343368709087, + 0.007051007356494665, + 0.006389016285538673, + 0.0026202453300356865, + 0.004225119482725859, + 0.006413660012185574, + 0.006506041623651981, + 0.007842916995286942, + 0.010014155879616737, + 0.009251649491488934, + 0.012429744005203247, + 0.004317900165915489, + 0.0056630875915288925, + 0.010614988394081593, + 0.007776646874845028, + 0.008852380327880383, + 0.012940745800733566, + 0.01271835807710886, + 0.010106799192726612, + 0.007803913205862045, + 0.010534001514315605, + 0.0122450552880764, + 0.012750182300806046, + 0.007070473395287991, + 0.009210151620209217, + 0.008445979095995426, + 0.009494098834693432, + 0.0031948022078722715, + 0.004974352195858955, + 0.00669475132599473, + 0.006484868470579386, + 0.004759669303894043, + 0.007319982163608074, + 0.006198402494192123, + 0.005801857449114323, + 0.002271259669214487, + 0.005187256261706352, + 0.004298611078411341, + 0.0041710808873176575, + 0.007555922958999872, + 0.008000979200005531, + 0.008106711320579052, + 0.01020518597215414, + 0.0034664759878069162, + 0.006919187027961016, + 0.005848830100148916, + 0.005556554067879915, + 0.047549232840538025, + 0.056238628923892975, + 0.048760466277599335, + 0.06911394000053406, + 0.026670260354876518, + 0.02240927703678608, + 0.025162523612380028, + 0.026851067319512367, + 0.04431602731347084, + 0.0437568761408329, + 0.03421749174594879, + 0.0365569032728672, + 0.01497145090252161, + 0.01676996611058712, + 0.01629352755844593, + 0.011666396632790565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.06203708052635193 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a watercolor and pen-and-ink style with fine ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a pair of slender, pale hands with immaculately manicured, neutral nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one hand cupping the chin while the other applies lipstick from a black tube with a gold ribbon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lipstick a vibrant, creamy crimson color being applied to the lower lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of a sunlit room with warm golden light and bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627411279827356 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.007627411279827356, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627411279827356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39091575145721436, + 0.06142793968319893, + 0.09713657200336456, + 0.24379277229309082, + 0.04990384727716446, + 0.033842507749795914, + 0.03243868798017502, + 0.014770176261663437, + 0.03412281721830368, + 0.04164892062544823 + ], + "credit_map": [ + 0.0443006195127964, + 0.048256803303956985, + 0.05389676243066788, + 0.05207911506295204, + 0.025945361703634262, + 0.05885486304759979, + 0.055102717131376266, + 0.052479542791843414, + 0.00989393051713705, + 0.008399327285587788, + 0.00889668334275484, + 0.010053887031972408, + 0.0033178720623254776, + 0.005182025488466024, + 0.008042746223509312, + 0.007641471456736326, + 0.01359812542796135, + 0.020237216725945473, + 0.014315003529191017, + 0.02067401260137558, + 0.00457350816577673, + 0.006699146702885628, + 0.009311802685260773, + 0.007727764546871185, + 0.030224522575736046, + 0.07811705023050308, + 0.033631473779678345, + 0.040604569017887115, + 0.011108211241662502, + 0.01794597879052162, + 0.01810266077518463, + 0.014058306813240051, + 0.00579931540414691, + 0.009609749540686607, + 0.00675129983574152, + 0.008369551971554756, + 0.0031529993284493685, + 0.0052426704205572605, + 0.005675246473401785, + 0.005303014535456896, + 0.0036655855365097523, + 0.009737852960824966, + 0.005169623997062445, + 0.004718309734016657, + 0.0018641105853021145, + 0.0029500580858439207, + 0.0033592728432267904, + 0.0023776921443641186, + 0.002892416901886463, + 0.006461850367486477, + 0.004666721913963556, + 0.0049886200577020645, + 0.0022021359764039516, + 0.00403169309720397, + 0.003961045295000076, + 0.003234204137697816, + 0.0013060871278867126, + 0.0020987526513636112, + 0.001989977201446891, + 0.0022125777322798967, + 0.0012131500989198685, + 0.0012867511250078678, + 0.002472595777362585, + 0.0021902837324887514, + 0.004479459021240473, + 0.006075754761695862, + 0.005164211615920067, + 0.006157797295600176, + 0.0027961060404777527, + 0.0026039965450763702, + 0.003536574076861143, + 0.0033089190255850554, + 0.004544903989881277, + 0.006110782735049725, + 0.007146908901631832, + 0.009541240520775318, + 0.0024972492828965187, + 0.0030917401891201735, + 0.004684230778366327, + 0.004031865857541561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.007627411279827356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_6/index_tensor([1762, 1762])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_6/index_tensor([1762, 1762])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6ecc29c5eb53b0080a4e1cf2fc3c62571055b7ba --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_6/index_tensor([1762, 1762])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera begins with a medium shot at eye level, slowly and steadily panning to the right, following the action in a brightly lit modern kitchen. A young man with neatly combed short brown hair and wearing a soft, heather-gray cotton T-shirt leans slightly over a gleaming stainless steel sink. Steady in both hands, he holds a small, rectangular glass fish tank under a modern, gooseneck chrome faucet, a stream of clear water filling it. With serene concentration, he watches the water level rise. When it's nearly full, he deftly turns the faucet handle with his right hand, and the flow abruptly stops. Then, he carefully lifts the now-heavy tank, the water sloshing gently inside, and turns his body toward the adjacent, sleek, light-gray crystal countertop. The camera pan ends as he deliberately and quietly sets the tank down. After a brief pause, his gaze shifts, and he carefully picks up a small fishing net with a delicate green mesh and a thin metal handle, which has been placed next to the tank. The entire scene is bathed in bright, indirect morning light, streaming in through an invisible window, creating soft shadows that accentuate the clean lines of the white subway tile backsplash, creating an atmosphere of quiet, focused communality and gentle preparation.", + "seed": 1221, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_6/index_1762_seed_1221_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01943294806405902, + "vlm_logprob": -0.01943294806405902, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.01943294806405902, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04972541704773903 + ], + "opposite_logprob_tokens": [ + -3.4247255325317383 + ], + "logprob_reward": -0.04972541704773903, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04972541704773903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24982503056526184, + 0.04742446169257164, + 0.029578763991594315, + 0.14383462071418762, + 0.1303752213716507, + 0.14124250411987305, + 0.05290945991873741, + 0.03112371824681759, + 0.0858534500002861, + 0.08783280849456787 + ], + "credit_map": [ + 0.10032808035612106, + 0.07576794922351837, + 0.02086706832051277, + 0.01092702616006136, + 0.012864725664258003, + 0.007475607562810183, + 0.005945602431893349, + 0.01564897783100605, + 0.01815071888267994, + 0.013392062857747078, + 0.0047227488830685616, + 0.0023794777225703, + 0.0026926903519779444, + 0.0015263534151017666, + 0.0015791587065905333, + 0.002981250174343586, + 0.008218126371502876, + 0.0058296541683375835, + 0.004258815664798021, + 0.0027294836472719908, + 0.0019979230128228664, + 0.0024352790787816048, + 0.0021193900611251593, + 0.0019900917541235685, + 0.05732567235827446, + 0.04496278986334801, + 0.014232879504561424, + 0.007341870106756687, + 0.006199345458298922, + 0.003813013667240739, + 0.004372126888483763, + 0.005586935207247734, + 0.05014953762292862, + 0.042275570333004, + 0.01094411127269268, + 0.007011421024799347, + 0.007583833299577236, + 0.0032269807998090982, + 0.004061897750943899, + 0.0051218620501458645, + 0.02105618827044964, + 0.02632940374314785, + 0.018541984260082245, + 0.031477563083171844, + 0.013829725794494152, + 0.012180994264781475, + 0.008468402549624443, + 0.009358248673379421, + 0.006573916878551245, + 0.0054547772742807865, + 0.005361074116080999, + 0.007158791180700064, + 0.00646559102460742, + 0.006750680971890688, + 0.008821484632790089, + 0.006323137786239386, + 0.0034538712352514267, + 0.004115876741707325, + 0.002833592239767313, + 0.0043418388813734055, + 0.005276262294501066, + 0.003222361672669649, + 0.005025884136557579, + 0.0028540308121591806, + 0.013369601219892502, + 0.015746360644698143, + 0.019125066697597504, + 0.011549290269613266, + 0.00847422331571579, + 0.00733311427757144, + 0.005744572728872299, + 0.004511219449341297, + 0.007378576323390007, + 0.013460645452141762, + 0.009670594707131386, + 0.011320825666189194, + 0.01040202658623457, + 0.019032083451747894, + 0.009582034312188625, + 0.006986018270254135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.04972541704773903 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028348281048238277 + ], + "opposite_logprob_tokens": [ + -5.877834796905518 + ], + "logprob_reward": -0.0028348281048238277, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028348281048238277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2404075711965561, + 0.04257524386048317, + 0.05481969565153122, + 0.020114826038479805, + 0.24181737005710602, + 0.04005863517522812, + 0.0293518528342247, + 0.08918201178312302, + 0.13239941000938416, + 0.10927338898181915 + ], + "credit_map": [ + 0.09459855407476425, + 0.07125385850667953, + 0.019208211451768875, + 0.009856440126895905, + 0.013617975637316704, + 0.00990642886608839, + 0.0066487169824540615, + 0.015317391604185104, + 0.01398969255387783, + 0.011282610706984997, + 0.004059591796249151, + 0.002189720282331109, + 0.003943748772144318, + 0.0021675904281437397, + 0.0025939843617379665, + 0.002348304493352771, + 0.01548382081091404, + 0.012184714898467064, + 0.005699578206986189, + 0.0063234115950763226, + 0.004521086812019348, + 0.004478386603295803, + 0.0030615930445492268, + 0.0030671076383441687, + 0.0035292033571749926, + 0.002398615935817361, + 0.0025127355474978685, + 0.0024641435593366623, + 0.0026764925569295883, + 0.0029399730265140533, + 0.001707569812424481, + 0.001886092941276729, + 0.09087306261062622, + 0.07494078576564789, + 0.019047414883971214, + 0.0063736047595739365, + 0.02344493567943573, + 0.010649184696376324, + 0.006696720607578754, + 0.009791669435799122, + 0.007725181989371777, + 0.007095638196915388, + 0.004277882631868124, + 0.004139482509344816, + 0.004789817146956921, + 0.004780974704772234, + 0.00442590331658721, + 0.0028237539809197187, + 0.002739948919042945, + 0.0026609660126268864, + 0.002967429580166936, + 0.003246518550440669, + 0.0033568968065083027, + 0.004752750508487225, + 0.005986117757856846, + 0.003641225164756179, + 0.008673416450619698, + 0.010319228284060955, + 0.008203141391277313, + 0.012941894121468067, + 0.013056144118309021, + 0.009712465107440948, + 0.01853955164551735, + 0.007736168336123228, + 0.0168418288230896, + 0.018387500196695328, + 0.023190515115857124, + 0.018349794670939445, + 0.016549138352274895, + 0.012570833787322044, + 0.012509952299296856, + 0.01399984210729599, + 0.01000350434333086, + 0.018471794202923775, + 0.008863097056746483, + 0.010819981805980206, + 0.015244310721755028, + 0.02137192338705063, + 0.009004348888993263, + 0.015494432300329208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.0028348281048238277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260316655039787 + ], + "opposite_logprob_tokens": [ + -5.255260467529297 + ], + "logprob_reward": -0.005260316655039787, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260316655039787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08835254609584808, + 0.06515102088451385, + 0.04007996991276741, + 0.05141845718026161, + 0.06603558361530304, + 0.13349539041519165, + 0.047990430146455765, + 0.053212348371744156, + 0.2137717604637146, + 0.24049246311187744 + ], + "credit_map": [ + 0.02758006751537323, + 0.02066393382847309, + 0.008285369724035263, + 0.00594774354249239, + 0.0076927547343075275, + 0.007010062225162983, + 0.004952398128807545, + 0.0062202210538089275, + 0.022713271901011467, + 0.016420049592852592, + 0.005777634214609861, + 0.005492707714438438, + 0.003943706396967173, + 0.004834288265556097, + 0.0025497586466372013, + 0.0034196036867797375, + 0.012943826615810394, + 0.010611829347908497, + 0.0031718784011900425, + 0.003404229413717985, + 0.003046210389584303, + 0.0023237979039549828, + 0.0021689680870622396, + 0.0024092290550470352, + 0.011077485047280788, + 0.0082576097920537, + 0.00895702000707388, + 0.008096934296190739, + 0.003700014902278781, + 0.003987318370491266, + 0.0038865534588694572, + 0.0034555250313133, + 0.018064023926854134, + 0.014486747793853283, + 0.007937922142446041, + 0.006460581440478563, + 0.0047601498663425446, + 0.006634270306676626, + 0.0033483493607491255, + 0.004343544133007526, + 0.02349410019814968, + 0.020301219075918198, + 0.014817440882325172, + 0.020232154056429863, + 0.0140942158177495, + 0.014930288307368755, + 0.017004642635583878, + 0.008621329441666603, + 0.0062548210844397545, + 0.004849614109843969, + 0.004730895161628723, + 0.004388467874377966, + 0.0061254194006323814, + 0.00906119029968977, + 0.007313904352486134, + 0.005266118329018354, + 0.007132141385227442, + 0.006180028896778822, + 0.007164297625422478, + 0.00769413635134697, + 0.006127523258328438, + 0.004458568524569273, + 0.009630560874938965, + 0.004825091455131769, + 0.025275370106101036, + 0.0319826602935791, + 0.025263845920562744, + 0.0200202539563179, + 0.04275836423039436, + 0.030447838827967644, + 0.018269846215844154, + 0.019753586500883102, + 0.018162626773118973, + 0.0411960668861866, + 0.022796738892793655, + 0.029245827347040176, + 0.025930114090442657, + 0.035429179668426514, + 0.029107803478837013, + 0.0386241152882576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.005260316655039787 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606367528438568 + ], + "opposite_logprob_tokens": [ + -4.141063690185547 + ], + "logprob_reward": -0.01606367528438568, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606367528438568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08973858505487442, + 0.08908882737159729, + 0.026797441765666008, + 0.069410540163517, + 0.09110484272241592, + 0.06491802632808685, + 0.2668248414993286, + 0.1314903050661087, + 0.08733834326267242, + 0.0832882896065712 + ], + "credit_map": [ + 0.03188300505280495, + 0.024409517645835876, + 0.008488651365041733, + 0.004937503486871719, + 0.00634592492133379, + 0.00407446688041091, + 0.004655938595533371, + 0.00494357431307435, + 0.035312555730342865, + 0.027370354160666466, + 0.007562810555100441, + 0.004304655361920595, + 0.004781992640346289, + 0.0030221096239984035, + 0.00223403237760067, + 0.004500315524637699, + 0.006373007781803608, + 0.004434581845998764, + 0.002852744422852993, + 0.002771587809547782, + 0.003057274268940091, + 0.0019900386687368155, + 0.0024437669198960066, + 0.0028744381852447987, + 0.026087164878845215, + 0.021870335564017296, + 0.0060847424902021885, + 0.003411681391298771, + 0.0038724008481949568, + 0.002256396459415555, + 0.00244874507188797, + 0.003379063680768013, + 0.02907542884349823, + 0.0238665658980608, + 0.007898306474089622, + 0.005050381645560265, + 0.0066642058081924915, + 0.009762314148247242, + 0.0031509471591562033, + 0.005636678542941809, + 0.00757306395098567, + 0.008252535946667194, + 0.005147319752722979, + 0.005917437374591827, + 0.009911594912409782, + 0.009188737720251083, + 0.011640806682407856, + 0.007286528591066599, + 0.033035244792699814, + 0.026871245354413986, + 0.015598359517753124, + 0.026562022045254707, + 0.04483002796769142, + 0.041189685463905334, + 0.05518229678273201, + 0.023555917665362358, + 0.01367619726806879, + 0.01551295630633831, + 0.012051140889525414, + 0.021109160035848618, + 0.020424915477633476, + 0.015742449089884758, + 0.019222762435674667, + 0.013750714249908924, + 0.011669804342091084, + 0.02106800489127636, + 0.011551138013601303, + 0.007573941256850958, + 0.013848005793988705, + 0.009556270204484463, + 0.006210025865584612, + 0.0058611477725207806, + 0.008422276005148888, + 0.01053121779114008, + 0.009763158857822418, + 0.006801213137805462, + 0.00924130529165268, + 0.019970780238509178, + 0.008544729091227055, + 0.010013602674007416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.01606367528438568 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328050322830677 + ], + "opposite_logprob_tokens": [ + -3.773280620574951 + ], + "logprob_reward": -0.02328050322830677, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328050322830677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0291946679353714, + 0.017369406297802925, + 0.019640227779746056, + 0.08963467925786972, + 0.05788921192288399, + 0.05040678754448891, + 0.03701372817158699, + 0.05074717849493027, + 0.36799320578575134, + 0.2801108658313751 + ], + "credit_map": [ + 0.005588090512901545, + 0.004892991855740547, + 0.002622141968458891, + 0.0026834027376025915, + 0.004320875275880098, + 0.002752814907580614, + 0.0024592180270701647, + 0.0038751321844756603, + 0.004750549793243408, + 0.003579291980713606, + 0.0014965280424803495, + 0.001352465245872736, + 0.0019005339127033949, + 0.0012177586322650313, + 0.0011865696869790554, + 0.0018857090035453439, + 0.004081071820110083, + 0.003073537489399314, + 0.0035363691858947277, + 0.0028499369509518147, + 0.0015365961007773876, + 0.0014939800603315234, + 0.0019635537173599005, + 0.0011051804758608341, + 0.037999581545591354, + 0.03059377335011959, + 0.007395240478217602, + 0.0031951209530234337, + 0.0035164912696927786, + 0.0017217437271028757, + 0.001647047116421163, + 0.003565683029592037, + 0.0211980938911438, + 0.01758020929992199, + 0.005499750841408968, + 0.0032170708291232586, + 0.00360300624743104, + 0.002235146937891841, + 0.0021167194936424494, + 0.002439213450998068, + 0.00689365528523922, + 0.005730376113206148, + 0.007727475371211767, + 0.0038898643106222153, + 0.004585477523505688, + 0.009836267679929733, + 0.00657198391854763, + 0.005171687807887793, + 0.00421947194263339, + 0.0034786774776875973, + 0.003392433747649193, + 0.0035100101958960295, + 0.005258186254650354, + 0.00553585821762681, + 0.007646495942026377, + 0.00397259509190917, + 0.0070741306990385056, + 0.005834054201841354, + 0.0073296488262712955, + 0.007848846726119518, + 0.006529211066663265, + 0.004423337057232857, + 0.007393197156488895, + 0.004314752295613289, + 0.043141160160303116, + 0.04895097017288208, + 0.0505458302795887, + 0.051215119659900665, + 0.06064105033874512, + 0.04371489956974983, + 0.034908004105091095, + 0.034876178950071335, + 0.023117773234844208, + 0.0457521490752697, + 0.025734316557645798, + 0.02666734904050827, + 0.02722291275858879, + 0.04185108095407486, + 0.030485620722174644, + 0.059279654175043106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.02328050322830677 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins with a medium shot at eye level, slowly and steadily panning to the right, following the action in a brightly lit modern kitchen. A young man with neatly combed short brown hair and wearing a soft, heather-gray cotton T-shirt leans slightly over a gleaming stainless steel sink. Steady in both hands, he holds a small, rectangular glass fish tank under a modern, gooseneck chrome faucet, a stream of clear water filling it. With serene concentration, he watches the water level rise. When it's nearly full, he deftly turns the faucet handle with his right hand, and the flow abruptly stops. Then, he carefully lifts the now-heavy tank, the water sloshing gently inside, and turns his body toward the adjacent, sleek, light-gray crystal countertop. The camera pan ends as he deliberately and quietly sets the tank down. After a brief pause, his gaze shifts, and he carefully picks up a small fishing net with a delicate green mesh and a thin metal handle, which has been placed next to the tank. The entire scene is bathed in bright, indirect morning light, streaming in through an invisible window, creating soft shadows that accentuate the clean lines of the white subway tile backsplash, creating an atmosphere of quiet, focused communality and gentle preparation.", + "seed": 1222, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_6/index_1762_seed_1222_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06626735460013151, + "vlm_logprob": -0.06626735460013151, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.06626735460013151, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24234366416931152 + ], + "opposite_logprob_tokens": [ + -1.6173436641693115 + ], + "logprob_reward": -0.24234366416931152, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24234366416931152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08614979684352875, + 0.07894749939441681, + 0.268865704536438, + 0.21818184852600098, + 0.034215204417705536, + 0.04716663435101509, + 0.035468511283397675, + 0.07789021730422974, + 0.08425967395305634, + 0.06885489821434021 + ], + "credit_map": [ + 0.009949411265552044, + 0.01860162802040577, + 0.010306566953659058, + 0.00894792191684246, + 0.00850643403828144, + 0.01137606892734766, + 0.009739202447235584, + 0.00872256513684988, + 0.006331268697977066, + 0.009681628085672855, + 0.007270353380590677, + 0.009061982855200768, + 0.012103748507797718, + 0.013829538598656654, + 0.008960387669503689, + 0.011708583682775497, + 0.029112886637449265, + 0.0661153793334961, + 0.029931744560599327, + 0.01856793276965618, + 0.0549941211938858, + 0.04096266254782677, + 0.01346990279853344, + 0.015711063519120216, + 0.007591484114527702, + 0.005474078003317118, + 0.0035399412736296654, + 0.002878748346120119, + 0.1010025218129158, + 0.09061267226934433, + 0.0023158248513936996, + 0.004766573663800955, + 0.003361828625202179, + 0.004123985301703215, + 0.0045456308871507645, + 0.0038521497044712305, + 0.0038336024153977633, + 0.00500499876216054, + 0.004123056773096323, + 0.005369955208152533, + 0.004755075555294752, + 0.004795264918357134, + 0.004451782908290625, + 0.0040712798945605755, + 0.007263804320245981, + 0.009647421538829803, + 0.004533670376986265, + 0.007648334372788668, + 0.006544310133904219, + 0.003260684199631214, + 0.006310716271400452, + 0.003133772173896432, + 0.002642696723341942, + 0.0035536671057343483, + 0.0034801422152668238, + 0.006542520131915808, + 0.011477858759462833, + 0.0062218341045081615, + 0.01533435471355915, + 0.01297129224985838, + 0.008941647596657276, + 0.006455651950091124, + 0.010744760744273663, + 0.005742815323174, + 0.008940048515796661, + 0.006739384029060602, + 0.012734873220324516, + 0.009396456182003021, + 0.008603856898844242, + 0.010905803181231022, + 0.010102374479174614, + 0.01683688536286354, + 0.011146724224090576, + 0.005812503397464752, + 0.014982243999838829, + 0.005395542364567518, + 0.009159470908343792, + 0.009389063343405724, + 0.005668353289365768, + 0.007300994824618101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.24234366416931152 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009802537970244884 + ], + "opposite_logprob_tokens": [ + -4.634802341461182 + ], + "logprob_reward": -0.009802537970244884, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009802537970244884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09784641861915588, + 0.032694511115550995, + 0.06294212490320206, + 0.4172247052192688, + 0.04968319460749626, + 0.13562683761119843, + 0.041898615658283234, + 0.07689175754785538, + 0.04605243355035782, + 0.039139412343502045 + ], + "credit_map": [ + 0.011576338671147823, + 0.01730802096426487, + 0.014977114275097847, + 0.011777474544942379, + 0.00951585453003645, + 0.012212933972477913, + 0.011013791896402836, + 0.009464891627430916, + 0.0034316901583224535, + 0.003991818521171808, + 0.006968267727643251, + 0.0026404561940580606, + 0.003831860376521945, + 0.004327643662691116, + 0.0029802063945680857, + 0.004522569943219423, + 0.008534330874681473, + 0.010342846624553204, + 0.011530602350831032, + 0.0047951675951480865, + 0.010367512702941895, + 0.008417053148150444, + 0.004822185728698969, + 0.004132427740842104, + 0.014568839222192764, + 0.011056202463805676, + 0.006446979008615017, + 0.006214452441781759, + 0.192824125289917, + 0.17258191108703613, + 0.00483403354883194, + 0.00869815144687891, + 0.004905519541352987, + 0.007803772576153278, + 0.010564170777797699, + 0.008709922432899475, + 0.004314784426242113, + 0.004620111081749201, + 0.005299081560224295, + 0.0034658322110772133, + 0.02166489139199257, + 0.016765937209129333, + 0.021540366113185883, + 0.015670889988541603, + 0.02171957679092884, + 0.014455747790634632, + 0.01328859943896532, + 0.010520818643271923, + 0.00989206600934267, + 0.004213627893477678, + 0.008196459151804447, + 0.0038042727392166853, + 0.003933072090148926, + 0.004286346025764942, + 0.0029992181807756424, + 0.004573550075292587, + 0.010569089092314243, + 0.005778612568974495, + 0.011791856028139591, + 0.010264274664223194, + 0.009352139197289944, + 0.007162311114370823, + 0.012036869302392006, + 0.009936606511473656, + 0.004265460651367903, + 0.003750378033146262, + 0.007534744683653116, + 0.005243740044534206, + 0.004558421205729246, + 0.00570631492882967, + 0.005963230971246958, + 0.009030140936374664, + 0.006817577872425318, + 0.0032392393331974745, + 0.008770840242505074, + 0.003641540417447686, + 0.004589436575770378, + 0.005134076811373234, + 0.00307658314704895, + 0.0038701200392097235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.009802537970244884 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792031720280647 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.009792031720280647, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792031720280647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058023784309625626, + 0.03990346938371658, + 0.0892845019698143, + 0.4713909327983856, + 0.04481222480535507, + 0.07202993333339691, + 0.0406484380364418, + 0.0737646073102951, + 0.04575574770569801, + 0.06438640505075455 + ], + "credit_map": [ + 0.007871399633586407, + 0.010691855102777481, + 0.005998011212795973, + 0.007532818242907524, + 0.006968045141547918, + 0.008314711973071098, + 0.0057512009516358376, + 0.004895737860351801, + 0.003148322692140937, + 0.0037887783255428076, + 0.0042898450046777725, + 0.0030450436752289534, + 0.004444612190127373, + 0.006059783976525068, + 0.006003149785101414, + 0.009123927913606167, + 0.005677912849932909, + 0.012441977858543396, + 0.008239619433879852, + 0.00753713957965374, + 0.013462159782648087, + 0.011675436981022358, + 0.01817050576210022, + 0.012079746462404728, + 0.01682177372276783, + 0.012257114052772522, + 0.005565212573856115, + 0.005166091024875641, + 0.21765951812267303, + 0.19627690315246582, + 0.006694464012980461, + 0.01094979140907526, + 0.003998170606791973, + 0.0056317090056836605, + 0.004586559720337391, + 0.0037483861669898033, + 0.005445275921374559, + 0.007288151420652866, + 0.0059137847274541855, + 0.008200183510780334, + 0.0077718812972307205, + 0.007194765377789736, + 0.005987768992781639, + 0.006331431679427624, + 0.011960353702306747, + 0.010050772689282894, + 0.010898703709244728, + 0.011834249831736088, + 0.006187028717249632, + 0.005693626590073109, + 0.004778951406478882, + 0.004401796497404575, + 0.004384704865515232, + 0.004750421736389399, + 0.005578141659498215, + 0.0048737600445747375, + 0.008495282381772995, + 0.006285792216658592, + 0.011010078713297844, + 0.010405179113149643, + 0.009744319133460522, + 0.0075410581193864346, + 0.012143091298639774, + 0.008139791898429394, + 0.004739646799862385, + 0.004363877698779106, + 0.006922492757439613, + 0.004862688016146421, + 0.00474498001858592, + 0.006279181223362684, + 0.0051759034395217896, + 0.00866696983575821, + 0.01051334012299776, + 0.0058289202861487865, + 0.01168717723339796, + 0.007440355606377125, + 0.006700604222714901, + 0.008292932994663715, + 0.007487589027732611, + 0.006435481831431389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 7 + ], + "selected_reward": -0.009792031720280647 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340505108237267 + ], + "opposite_logprob_tokens": [ + -3.6513404846191406 + ], + "logprob_reward": -0.026340505108237267, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340505108237267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03590758889913559, + 0.031098781153559685, + 0.113413967192173, + 0.4278077781200409, + 0.03709271177649498, + 0.04028763622045517, + 0.15760450065135956, + 0.09148279577493668, + 0.02718581072986126, + 0.038118403404951096 + ], + "credit_map": [ + 0.0037298440001904964, + 0.006943152751773596, + 0.006308706942945719, + 0.004154896829277277, + 0.003383212722837925, + 0.003842801321297884, + 0.004141025710850954, + 0.0034039458259940147, + 0.00217125890776515, + 0.004044702742248774, + 0.005459547974169254, + 0.003142625093460083, + 0.002891790820285678, + 0.003927911166101694, + 0.005457058548927307, + 0.004003886599093676, + 0.011671677231788635, + 0.02716415748000145, + 0.011072500608861446, + 0.0072084516286849976, + 0.023802490904927254, + 0.01892498880624771, + 0.005267987959086895, + 0.008301720954477787, + 0.015301867388188839, + 0.013978047296404839, + 0.008903036825358868, + 0.006926493253558874, + 0.19549302756786346, + 0.17379826307296753, + 0.004808756057173014, + 0.008598266169428825, + 0.004176200833171606, + 0.004759547300636768, + 0.007322543300688267, + 0.003159115556627512, + 0.003458385355770588, + 0.004273915663361549, + 0.005317465402185917, + 0.004625539295375347, + 0.007089323829859495, + 0.00474003329873085, + 0.009248105809092522, + 0.003944378346204758, + 0.003963189665228128, + 0.0051369136199355125, + 0.0029613689985126257, + 0.0032043245155364275, + 0.02737177349627018, + 0.025402352213859558, + 0.02124260552227497, + 0.01480440329760313, + 0.015836263075470924, + 0.0175830889493227, + 0.01592171937227249, + 0.019442295655608177, + 0.012167977169156075, + 0.006937265861779451, + 0.009923516772687435, + 0.012961892411112785, + 0.011594366282224655, + 0.008442227728664875, + 0.017537591978907585, + 0.011917962692677975, + 0.0028324099257588387, + 0.0033649862743914127, + 0.0032248888164758682, + 0.003215617034584284, + 0.003600204596295953, + 0.0033513151574879885, + 0.0031432476826012135, + 0.004453143570572138, + 0.006037046667188406, + 0.0038758155424147844, + 0.0077782585285604, + 0.0032964458223432302, + 0.0038978843949735165, + 0.00440643634647131, + 0.003825600491836667, + 0.005000915378332138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.026340505108237267 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305803403258324 + ], + "opposite_logprob_tokens": [ + -3.168057918548584 + ], + "logprob_reward": -0.04305803403258324, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305803403258324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05472712218761444, + 0.043156109750270844, + 0.09600351750850677, + 0.341439425945282, + 0.052181318402290344, + 0.05674894154071808, + 0.07541050016880035, + 0.12872056663036346, + 0.06459782272577286, + 0.08701466768980026 + ], + "credit_map": [ + 0.00584125192835927, + 0.012181933969259262, + 0.0053750514052808285, + 0.005368857178837061, + 0.005857637617737055, + 0.008097689598798752, + 0.005969021935015917, + 0.006035677157342434, + 0.004266969859600067, + 0.004242588300257921, + 0.005211513489484787, + 0.003362704999744892, + 0.005377487745136023, + 0.006046571768820286, + 0.0064851026982069016, + 0.008163169026374817, + 0.008509707637131214, + 0.018779722973704338, + 0.009790180251002312, + 0.007164534647017717, + 0.015829971060156822, + 0.012560809031128883, + 0.015810001641511917, + 0.00755858700722456, + 0.012979255989193916, + 0.01367239374667406, + 0.011325166560709476, + 0.00901578925549984, + 0.14074942469596863, + 0.12455940991640091, + 0.00963079184293747, + 0.019507193937897682, + 0.005337245296686888, + 0.007290963549166918, + 0.006343253422528505, + 0.0048070745542645454, + 0.00658660801127553, + 0.009059696458280087, + 0.004936696961522102, + 0.007819775491952896, + 0.006030952092260122, + 0.00434970622882247, + 0.0077772075310349464, + 0.005019829608500004, + 0.007977576926350594, + 0.011412597261369228, + 0.004410990513861179, + 0.009770077653229237, + 0.012646574527025223, + 0.008151779882609844, + 0.009894216433167458, + 0.0074342661537230015, + 0.008831426501274109, + 0.007872862741351128, + 0.008241823874413967, + 0.012337548658251762, + 0.013881480321288109, + 0.010727583430707455, + 0.01624203659594059, + 0.024437712505459785, + 0.021520856767892838, + 0.013890576548874378, + 0.019168222323060036, + 0.008852102793753147, + 0.0067319003865122795, + 0.0035789182875305414, + 0.007011496927589178, + 0.006144259124994278, + 0.008129716850817204, + 0.008597529493272305, + 0.008330518379807472, + 0.016073480248451233, + 0.007710359059274197, + 0.010194840840995312, + 0.012345797382295132, + 0.009681439027190208, + 0.010535981506109238, + 0.011746351607143879, + 0.011274515651166439, + 0.01352537702769041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 2 + ], + "selected_reward": -0.04305803403258324 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins with a medium shot at eye level, slowly and steadily panning to the right, following the action in a brightly lit modern kitchen. A young man with neatly combed short brown hair and wearing a soft, heather-gray cotton T-shirt leans slightly over a gleaming stainless steel sink. Steady in both hands, he holds a small, rectangular glass fish tank under a modern, gooseneck chrome faucet, a stream of clear water filling it. With serene concentration, he watches the water level rise. When it's nearly full, he deftly turns the faucet handle with his right hand, and the flow abruptly stops. Then, he carefully lifts the now-heavy tank, the water sloshing gently inside, and turns his body toward the adjacent, sleek, light-gray crystal countertop. The camera pan ends as he deliberately and quietly sets the tank down. After a brief pause, his gaze shifts, and he carefully picks up a small fishing net with a delicate green mesh and a thin metal handle, which has been placed next to the tank. The entire scene is bathed in bright, indirect morning light, streaming in through an invisible window, creating soft shadows that accentuate the clean lines of the white subway tile backsplash, creating an atmosphere of quiet, focused communality and gentle preparation.", + "seed": 1223, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_6/index_1762_seed_1223_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.033700197702273726, + "vlm_logprob": -0.033700197702273726, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.033700197702273726, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08228305727243423 + ], + "opposite_logprob_tokens": [ + -2.8322830200195312 + ], + "logprob_reward": -0.08228305727243423, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08228305727243423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12379591166973114, + 0.044500600546598434, + 0.07867853343486786, + 0.15214401483535767, + 0.07538861036300659, + 0.23732732236385345, + 0.040362775325775146, + 0.04714663326740265, + 0.08398738503456116, + 0.11666826158761978 + ], + "credit_map": [ + 0.012607521377503872, + 0.009438328444957733, + 0.016573993489146233, + 0.010987754911184311, + 0.01339260395616293, + 0.015109427273273468, + 0.02773417718708515, + 0.017952091991901398, + 0.004811324644833803, + 0.0049287318252027035, + 0.00694991322234273, + 0.00477111991494894, + 0.004481871146708727, + 0.005175846163183451, + 0.00871167704463005, + 0.004670110996812582, + 0.004056570120155811, + 0.005362358409911394, + 0.003790422808378935, + 0.003292477922514081, + 0.02488035149872303, + 0.020603982731699944, + 0.007289568427950144, + 0.009402794763445854, + 0.0104571133852005, + 0.017374984920024872, + 0.013438842259347439, + 0.011275006458163261, + 0.0243388582020998, + 0.0348847322165966, + 0.023186229169368744, + 0.017188237980008125, + 0.007881229743361473, + 0.006824797950685024, + 0.00844099372625351, + 0.008630952797830105, + 0.01304276566952467, + 0.00923731829971075, + 0.009537486359477043, + 0.011793055571615696, + 0.029736075550317764, + 0.01724066399037838, + 0.02599300444126129, + 0.023163387551903725, + 0.026843009516596794, + 0.03733308985829353, + 0.03444873169064522, + 0.0425693616271019, + 0.005491004791110754, + 0.005012380890548229, + 0.005227756220847368, + 0.006417676340788603, + 0.004695164039731026, + 0.004022829234600067, + 0.004427432548254728, + 0.005068528465926647, + 0.009469163604080677, + 0.007820512168109417, + 0.006548105739057064, + 0.004774036817252636, + 0.0036577635910362005, + 0.003924622666090727, + 0.006170674227178097, + 0.004781750030815601, + 0.01801818422973156, + 0.012846038676798344, + 0.009531278163194656, + 0.011442928574979305, + 0.007574785966426134, + 0.008699842728674412, + 0.008415278978645802, + 0.007459042593836784, + 0.010528143495321274, + 0.009508706629276276, + 0.014931365847587585, + 0.018459485843777657, + 0.013643796555697918, + 0.01744127832353115, + 0.01875861920416355, + 0.013396854512393475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.08228305727243423 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003211580915376544 + ], + "opposite_logprob_tokens": [ + -5.753211498260498 + ], + "logprob_reward": -0.003211580915376544, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003211580915376544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15134654939174652, + 0.08968497067689896, + 0.0907113328576088, + 0.05107651650905609, + 0.1773429960012436, + 0.15356971323490143, + 0.07257518917322159, + 0.0659397691488266, + 0.08313676714897156, + 0.06461621820926666 + ], + "credit_map": [ + 0.01620394177734852, + 0.013365169987082481, + 0.033914048224687576, + 0.01742340438067913, + 0.01826755329966545, + 0.016493532806634903, + 0.02377774938941002, + 0.01190113928169012, + 0.011436941102147102, + 0.024690954014658928, + 0.007646440993994474, + 0.00825458113104105, + 0.013194835744798183, + 0.013255894184112549, + 0.005106925964355469, + 0.006098388694226742, + 0.007184636779129505, + 0.008411484770476818, + 0.009753488004207611, + 0.004399649798870087, + 0.027289528399705887, + 0.024098364636301994, + 0.003920519258826971, + 0.00565365282818675, + 0.004853107500821352, + 0.009134531952440739, + 0.007757553830742836, + 0.005206800997257233, + 0.006789437495172024, + 0.005518924444913864, + 0.005961319897323847, + 0.005854834336787462, + 0.022869877517223358, + 0.02760399505496025, + 0.047385603189468384, + 0.019249960780143738, + 0.020098550245165825, + 0.019763130694627762, + 0.008189301937818527, + 0.012182566337287426, + 0.020299945026636124, + 0.02757505141198635, + 0.033732831478118896, + 0.023180151358246803, + 0.008590578101575375, + 0.02063804678618908, + 0.01003239955753088, + 0.009520702995359898, + 0.010936914943158627, + 0.009656121022999287, + 0.01609094813466072, + 0.008312603458762169, + 0.005189626477658749, + 0.006992271635681391, + 0.008203228935599327, + 0.007193469908088446, + 0.014414729550480843, + 0.013022460043430328, + 0.01000248547643423, + 0.008415672928094864, + 0.004541473463177681, + 0.00539877125993371, + 0.005172351840883493, + 0.004971813876181841, + 0.01288729626685381, + 0.010784905403852463, + 0.00888893473893404, + 0.012340083718299866, + 0.010384627617895603, + 0.008772473782300949, + 0.00952520128339529, + 0.009553235024213791, + 0.004510374274104834, + 0.005864561069756746, + 0.007612228859215975, + 0.01157345063984394, + 0.01020638458430767, + 0.009849403984844685, + 0.008083479478955269, + 0.00691632367670536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.003211580915376544 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083230376243591 + ], + "opposite_logprob_tokens": [ + -4.511083126068115 + ], + "logprob_reward": -0.011083230376243591, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083230376243591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.134996697306633, + 0.04350478947162628, + 0.04150853306055069, + 0.07111435383558273, + 0.047623131424188614, + 0.1500033140182495, + 0.1541249305009842, + 0.10701839625835419, + 0.18225063383579254, + 0.06785512715578079 + ], + "credit_map": [ + 0.010976116172969341, + 0.010371336713433266, + 0.011930745095014572, + 0.011972649954259396, + 0.015044274739921093, + 0.01628798618912697, + 0.030977090820670128, + 0.02743651531636715, + 0.0035506277345120907, + 0.0037258362863212824, + 0.004586957860738039, + 0.0048030074685812, + 0.005151588935405016, + 0.006525502540171146, + 0.010082165710628033, + 0.005079106893390417, + 0.0021242243237793446, + 0.003623237134888768, + 0.002638459438458085, + 0.0020452733151614666, + 0.013398253358900547, + 0.011911364272236824, + 0.002762997755780816, + 0.0030047278851270676, + 0.007738483138382435, + 0.010851324535906315, + 0.009499583393335342, + 0.006390830036252737, + 0.009003255516290665, + 0.009754570201039314, + 0.010092695243656635, + 0.007783617824316025, + 0.00402489397674799, + 0.0039035880472511053, + 0.009619657881557941, + 0.003558830823749304, + 0.008490062318742275, + 0.005434100981801748, + 0.005097509361803532, + 0.007494496181607246, + 0.016715463250875473, + 0.018125241622328758, + 0.014738188125193119, + 0.010913529433310032, + 0.03466285765171051, + 0.026438744738698006, + 0.015153514221310616, + 0.013255782425403595, + 0.022060059010982513, + 0.019570278003811836, + 0.03269411623477936, + 0.023786265403032303, + 0.013158970512449741, + 0.015529677271842957, + 0.016537176445126534, + 0.01078840997070074, + 0.018552538007497787, + 0.013354080729186535, + 0.012012379243969917, + 0.013080203905701637, + 0.009393936954438686, + 0.009667419828474522, + 0.018172521144151688, + 0.012785324826836586, + 0.03346119448542595, + 0.02702447772026062, + 0.01787835732102394, + 0.01933687925338745, + 0.017575539648532867, + 0.019215665757656097, + 0.02616129070520401, + 0.021597251296043396, + 0.008025892078876495, + 0.005361432209610939, + 0.012548346072435379, + 0.01208808459341526, + 0.0067098126746714115, + 0.00907707680016756, + 0.0075548384338617325, + 0.006489654537290335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.011083230376243591 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280037567019463 + ], + "opposite_logprob_tokens": [ + -3.773280143737793 + ], + "logprob_reward": -0.023280037567019463, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280037567019463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06490530073642731, + 0.048059768974781036, + 0.060210902243852615, + 0.04961865767836571, + 0.04614430293440819, + 0.10297694057226181, + 0.11778741329908371, + 0.15157994627952576, + 0.3029470145702362, + 0.055769696831703186 + ], + "credit_map": [ + 0.007238165009766817, + 0.006604794878512621, + 0.0072715491987764835, + 0.005988835357129574, + 0.013163995929062366, + 0.012494674883782864, + 0.006140029989182949, + 0.0060032568871974945, + 0.005369533319026232, + 0.012488964013755322, + 0.004466013051569462, + 0.0036235786974430084, + 0.008061337284743786, + 0.007430901750922203, + 0.0033938142005354166, + 0.0032256210688501596, + 0.005780539009720087, + 0.005031455773860216, + 0.008253581821918488, + 0.004801812581717968, + 0.013615227304399014, + 0.009999358095228672, + 0.006401341874152422, + 0.0063275862485170364, + 0.003101521637290716, + 0.006196009460836649, + 0.005422375164926052, + 0.004200085997581482, + 0.008075758814811707, + 0.010997209697961807, + 0.005038212053477764, + 0.006587485782802105, + 0.0049975779838860035, + 0.008023610338568687, + 0.006969395559281111, + 0.004273667931556702, + 0.006710384506732225, + 0.005536383017897606, + 0.004503540229052305, + 0.005129744298756123, + 0.01592167280614376, + 0.01249208115041256, + 0.010953404009342194, + 0.01129467785358429, + 0.01133230235427618, + 0.016287198290228844, + 0.014698903076350689, + 0.009996703825891018, + 0.018364785239100456, + 0.014052384532988071, + 0.017911115661263466, + 0.015277388505637646, + 0.015539797022938728, + 0.012594504281878471, + 0.012342025525867939, + 0.011705415323376656, + 0.018166353926062584, + 0.020876610651612282, + 0.014042463153600693, + 0.021467378363013268, + 0.015237947925925255, + 0.013940819539129734, + 0.026973119005560875, + 0.020875249058008194, + 0.035747189074754715, + 0.047031015157699585, + 0.03367970883846283, + 0.03062027506530285, + 0.03742904216051102, + 0.034498002380132675, + 0.045407310128211975, + 0.0385345034301281, + 0.0049384525045752525, + 0.005327774211764336, + 0.009233381599187851, + 0.011120310984551907, + 0.00499710813164711, + 0.008640971034765244, + 0.005676354747265577, + 0.005835345014929771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.023280037567019463 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486430823802948 + ], + "opposite_logprob_tokens": [ + -3.048643112182617 + ], + "logprob_reward": -0.0486430823802948, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486430823802948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06696832925081253, + 0.021425874903798103, + 0.032967109233140945, + 0.04541492834687233, + 0.029933610931038857, + 0.06356668472290039, + 0.0950973629951477, + 0.14619778096675873, + 0.26602762937545776, + 0.23240059614181519 + ], + "credit_map": [ + 0.0074532818980515, + 0.0064538829028606415, + 0.0070554097183048725, + 0.0075576575472950935, + 0.008502709679305553, + 0.008067368529736996, + 0.013063917867839336, + 0.008814109489321709, + 0.0027396725490689278, + 0.002278632950037718, + 0.003540248144418001, + 0.0021544122137129307, + 0.002694733440876007, + 0.0025128729175776243, + 0.0025036402512341738, + 0.0030016633681952953, + 0.0026682026218622923, + 0.0028819746803492308, + 0.0035623335279524326, + 0.0023153815418481827, + 0.008714498952031136, + 0.0071228500455617905, + 0.0023414073511958122, + 0.0033604661002755165, + 0.005274517927318811, + 0.0052453260868787766, + 0.008860445581376553, + 0.0052787228487432, + 0.006055750884115696, + 0.006919788662344217, + 0.0034055649302899837, + 0.0043748123571276665, + 0.0027666643727570772, + 0.003458180697634816, + 0.006364687345921993, + 0.003003046615049243, + 0.004275846295058727, + 0.003667358309030533, + 0.0023221371229737997, + 0.004075693432241678, + 0.009505048394203186, + 0.006140722893178463, + 0.010179354809224606, + 0.006284245755523443, + 0.00702715665102005, + 0.009548516012728214, + 0.007346045691519976, + 0.0075356015004217625, + 0.01070922240614891, + 0.013190134428441525, + 0.013513895682990551, + 0.010597631335258484, + 0.014361965470016003, + 0.009992636740207672, + 0.010897566564381123, + 0.011834324337542057, + 0.026672067120671272, + 0.019597651436924934, + 0.015409550629556179, + 0.019354594871401787, + 0.01337308157235384, + 0.012532574124634266, + 0.020623700693249702, + 0.018634574487805367, + 0.022728895768523216, + 0.027865273877978325, + 0.01900332234799862, + 0.02883698232471943, + 0.045346006751060486, + 0.036595989018678665, + 0.04711785167455673, + 0.038533322513103485, + 0.018282542005181313, + 0.016634954139590263, + 0.038197897374629974, + 0.043320611119270325, + 0.029829589650034904, + 0.03484836965799332, + 0.02319071628153324, + 0.02809595689177513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.0486430823802948 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins with a medium shot at eye level, slowly and steadily panning to the right, following the action in a brightly lit modern kitchen. A young man with neatly combed short brown hair and wearing a soft, heather-gray cotton T-shirt leans slightly over a gleaming stainless steel sink. Steady in both hands, he holds a small, rectangular glass fish tank under a modern, gooseneck chrome faucet, a stream of clear water filling it. With serene concentration, he watches the water level rise. When it's nearly full, he deftly turns the faucet handle with his right hand, and the flow abruptly stops. Then, he carefully lifts the now-heavy tank, the water sloshing gently inside, and turns his body toward the adjacent, sleek, light-gray crystal countertop. The camera pan ends as he deliberately and quietly sets the tank down. After a brief pause, his gaze shifts, and he carefully picks up a small fishing net with a delicate green mesh and a thin metal handle, which has been placed next to the tank. The entire scene is bathed in bright, indirect morning light, streaming in through an invisible window, creating soft shadows that accentuate the clean lines of the white subway tile backsplash, creating an atmosphere of quiet, focused communality and gentle preparation.", + "seed": 1224, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_6/index_1762_seed_1224_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1121253508143127, + "vlm_logprob": -0.1121253508143127, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.1121253508143127, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.406501442193985 + ], + "opposite_logprob_tokens": [ + -1.1565014123916626 + ], + "logprob_reward": -0.406501442193985, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.406501442193985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07150780409574509, + 0.03005889616906643, + 0.061945389956235886, + 0.014381764456629753, + 0.06022448092699051, + 0.04205240309238434, + 0.0158296637237072, + 0.023540541529655457, + 0.044770535081624985, + 0.6356886029243469 + ], + "credit_map": [ + 0.01138270553201437, + 0.006512769032269716, + 0.007811977062374353, + 0.008593589067459106, + 0.008688204921782017, + 0.013215393759310246, + 0.008929156698286533, + 0.006374005693942308, + 0.00431921286508441, + 0.00300932046957314, + 0.0030029176268726587, + 0.0029474145267158747, + 0.0023525580763816833, + 0.004891271702945232, + 0.0033651955891400576, + 0.006171002984046936, + 0.012064865790307522, + 0.004891838412731886, + 0.008024950511753559, + 0.0059407236985862255, + 0.006898907013237476, + 0.010268201120197773, + 0.0063547235913574696, + 0.007501172833144665, + 0.0023677977733314037, + 0.0023070690222084522, + 0.0016439033206552267, + 0.0016561042284592986, + 0.0016744222957640886, + 0.0020533218048512936, + 0.0008180676377378404, + 0.0018610775005072355, + 0.009686966426670551, + 0.0052102734334766865, + 0.01086323894560337, + 0.00533628324046731, + 0.006484637036919594, + 0.011212020181119442, + 0.005683234892785549, + 0.005747826769948006, + 0.008232271298766136, + 0.004210957325994968, + 0.006073811557143927, + 0.004191338550299406, + 0.004290770273655653, + 0.007028361316770315, + 0.0035389261320233345, + 0.0044859605841338634, + 0.001889705308713019, + 0.0009624131489545107, + 0.0018200502963736653, + 0.001944619114510715, + 0.0011183961760252714, + 0.0023982911370694637, + 0.002382175996899605, + 0.0033140103332698345, + 0.0030091290827840567, + 0.0013686660677194595, + 0.0019456160953268409, + 0.0018494157120585442, + 0.002180879469960928, + 0.004674739204347134, + 0.0037950247060507536, + 0.004717067815363407, + 0.0099515151232481, + 0.006027156487107277, + 0.006995298899710178, + 0.004134108312427998, + 0.004076577723026276, + 0.00437032850459218, + 0.004501629155129194, + 0.0047139134258031845, + 0.09805043786764145, + 0.05260314792394638, + 0.10200471431016922, + 0.019921941682696342, + 0.10854021459817886, + 0.21326705813407898, + 0.014590454287827015, + 0.026710549369454384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.406501442193985 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004659390542656183 + ], + "opposite_logprob_tokens": [ + -5.379659175872803 + ], + "logprob_reward": -0.004659390542656183, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004659390542656183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17682099342346191, + 0.08675410598516464, + 0.06224839761853218, + 0.12256119400262833, + 0.12139425426721573, + 0.03289499878883362, + 0.027066687121987343, + 0.014242975041270256, + 0.01917204074561596, + 0.3368443548679352 + ], + "credit_map": [ + 0.028268510475754738, + 0.01412003580480814, + 0.024107491597533226, + 0.020565537735819817, + 0.019541583955287933, + 0.029970746487379074, + 0.02061300352215767, + 0.019634105265140533, + 0.008839081972837448, + 0.013864441774785519, + 0.01735445111989975, + 0.011359591968357563, + 0.011326543986797333, + 0.011942621320486069, + 0.005218430422246456, + 0.0068489438854157925, + 0.008168382570147514, + 0.0038312424439936876, + 0.0038346287328749895, + 0.006728391628712416, + 0.004709587898105383, + 0.01082515250891447, + 0.011426838114857674, + 0.012724179774522781, + 0.028072107583284378, + 0.013081354089081287, + 0.013072743080556393, + 0.010962080210447311, + 0.01130159106105566, + 0.019942034035921097, + 0.009165010415017605, + 0.016964277252554893, + 0.017096519470214844, + 0.014716582372784615, + 0.022126689553260803, + 0.010851121507585049, + 0.01401886623352766, + 0.021778691560029984, + 0.010736819356679916, + 0.010068969801068306, + 0.00483310129493475, + 0.003848753171041608, + 0.003578315954655409, + 0.002675152150914073, + 0.004218854941427708, + 0.005648158024996519, + 0.003262908663600683, + 0.0048297555185854435, + 0.003907882608473301, + 0.0018396106315776706, + 0.003923375159502029, + 0.0027635765727609396, + 0.002197747118771076, + 0.004577066749334335, + 0.0025152757298201323, + 0.005342155694961548, + 0.001864083344116807, + 0.001375966938212514, + 0.0018920012516900897, + 0.001054528751410544, + 0.0019732729997485876, + 0.002195321023464203, + 0.0018148727249354124, + 0.002072929171845317, + 0.0035221711732447147, + 0.0018985890783369541, + 0.0018105992348864675, + 0.0019717032555490732, + 0.0024256319738924503, + 0.002653842093423009, + 0.002413666807115078, + 0.002475838642567396, + 0.04988453537225723, + 0.029881605878472328, + 0.05125807225704193, + 0.009652967564761639, + 0.05731749162077904, + 0.11491723358631134, + 0.0093266312032938, + 0.014605807140469551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.004659390542656183 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006747082807123661 + ], + "opposite_logprob_tokens": [ + -5.006747245788574 + ], + "logprob_reward": -0.006747082807123661, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747082807123661, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23936933279037476, + 0.07246999442577362, + 0.08852336555719376, + 0.14389751851558685, + 0.02946014143526554, + 0.04571399837732315, + 0.032499562948942184, + 0.03546757623553276, + 0.08011224865913391, + 0.2324862778186798 + ], + "credit_map": [ + 0.03837091848254204, + 0.01781586743891239, + 0.033608436584472656, + 0.026296226307749748, + 0.028017360717058182, + 0.041000399738550186, + 0.02533009648323059, + 0.028930019587278366, + 0.010031231679022312, + 0.009674472734332085, + 0.012336596846580505, + 0.0072454968467354774, + 0.008473139256238937, + 0.010083156637847424, + 0.007553321309387684, + 0.007072577252984047, + 0.010641361586749554, + 0.005543522536754608, + 0.0065732658840715885, + 0.00993605051189661, + 0.004978463519364595, + 0.016427045688033104, + 0.012181855738162994, + 0.022241801023483276, + 0.032592371106147766, + 0.015650806948542595, + 0.015158399939537048, + 0.010940728709101677, + 0.013257525861263275, + 0.024047277867794037, + 0.009440953843295574, + 0.022809455171227455, + 0.004419176373630762, + 0.002340837148949504, + 0.006262787152081728, + 0.0028676355723291636, + 0.002645031316205859, + 0.005556339863687754, + 0.0025466124061495066, + 0.002821722300723195, + 0.003604007186368108, + 0.005813022144138813, + 0.004293655976653099, + 0.003484065644443035, + 0.005856513045728207, + 0.007480630651116371, + 0.006536244414746761, + 0.008645858615636826, + 0.005060082301497459, + 0.0025414435658603907, + 0.0033967967610806227, + 0.003543970175087452, + 0.002215343527495861, + 0.005486289970576763, + 0.005027194507420063, + 0.0052284421399235725, + 0.003823622362688184, + 0.0038100380916148424, + 0.0036572369281202555, + 0.0026632396038621664, + 0.00790776964277029, + 0.006540623493492603, + 0.0029351424891501665, + 0.004129904787987471, + 0.013690427877008915, + 0.011353849433362484, + 0.011097827926278114, + 0.00807538814842701, + 0.01055213250219822, + 0.008532269857823849, + 0.008844174444675446, + 0.007966181263327599, + 0.0342826284468174, + 0.02007339708507061, + 0.036369193345308304, + 0.007689246442168951, + 0.03929835557937622, + 0.07773144543170929, + 0.0065898895263671875, + 0.010452129878103733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.006747082807123661 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292202025651932 + ], + "opposite_logprob_tokens": [ + -2.237921953201294 + ], + "logprob_reward": -0.11292202025651932, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292202025651932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029653843492269516, + 0.021954642608761787, + 0.01821732148528099, + 0.04729656130075455, + 0.014545457437634468, + 0.026094622910022736, + 0.01528795063495636, + 0.026612037792801857, + 0.04949504882097244, + 0.7508425116539001 + ], + "credit_map": [ + 0.004446981009095907, + 0.0028367568738758564, + 0.0062076072208583355, + 0.00400143675506115, + 0.0030761156231164932, + 0.0037621574010699987, + 0.0028372970409691334, + 0.002485492965206504, + 0.0032169059850275517, + 0.0026727193035185337, + 0.003075759159401059, + 0.0025257114320993423, + 0.00184867181815207, + 0.0031738178804516792, + 0.0015728126745671034, + 0.0038682438898831606, + 0.002277268562465906, + 0.0014721974730491638, + 0.0017257841536775231, + 0.002109651919454336, + 0.0017023567343130708, + 0.002755415625870228, + 0.0033759172074496746, + 0.0027987302746623755, + 0.010178904049098492, + 0.005084756761789322, + 0.005825468804687262, + 0.004176578484475613, + 0.0045383647084236145, + 0.007521113846451044, + 0.003361428389325738, + 0.006609947420656681, + 0.0020137398969382048, + 0.0015501758316531777, + 0.002740737283602357, + 0.001408140640705824, + 0.0011634929105639458, + 0.0021930618677288294, + 0.0011260518804192543, + 0.0023500570096075535, + 0.004219776019454002, + 0.002415737137198448, + 0.003488635877147317, + 0.0030298815108835697, + 0.002166296588256955, + 0.0038787845987826586, + 0.002809644676744938, + 0.004085865803062916, + 0.0020292301196604967, + 0.000971405126620084, + 0.0018177905585616827, + 0.0011361034121364355, + 0.0013383928453549743, + 0.0028898674063384533, + 0.0011449713492766023, + 0.003960189875215292, + 0.003719948697835207, + 0.0025607554707676172, + 0.002769856248050928, + 0.0016872057458385825, + 0.004794060252606869, + 0.004806785378605127, + 0.0029161812271922827, + 0.0033572446554899216, + 0.007692352868616581, + 0.004348920192569494, + 0.006703130900859833, + 0.006675181910395622, + 0.005806990899145603, + 0.005274343304336071, + 0.006410891190171242, + 0.006583238951861858, + 0.11644210666418076, + 0.0630875751376152, + 0.12181912362575531, + 0.021801527589559555, + 0.12830334901809692, + 0.2526622712612152, + 0.016658611595630646, + 0.03006792813539505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.11292202025651932 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen with white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with short brown hair wearing a heather-gray cotton T-shirt visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man hold a small rectangular glass fish tank under a gooseneck chrome faucet and fill it with water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot at eye level and pan steadily to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man pick up a fishing net with a delicate green mesh and a thin metal handle after setting down the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796818271279335 + ], + "opposite_logprob_tokens": [ + -3.529796838760376 + ], + "logprob_reward": -0.029796818271279335, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796818271279335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07332363724708557, + 0.04053426906466484, + 0.03450321778655052, + 0.032195523381233215, + 0.024168075993657112, + 0.029312118887901306, + 0.027142295613884926, + 0.052771762013435364, + 0.08131203800439835, + 0.6047370433807373 + ], + "credit_map": [ + 0.009381149895489216, + 0.006639828905463219, + 0.010706375353038311, + 0.007350806146860123, + 0.009250283241271973, + 0.013473241589963436, + 0.008542864583432674, + 0.007979094982147217, + 0.004883817862719297, + 0.006569539196789265, + 0.005429872311651707, + 0.005075851455330849, + 0.004752214532345533, + 0.005645937751978636, + 0.004514271859079599, + 0.0036627687513828278, + 0.0059631261974573135, + 0.002866205759346485, + 0.003051315201446414, + 0.0032625426538288593, + 0.003122802125290036, + 0.006117855664342642, + 0.0033437449019402266, + 0.006775629241019487, + 0.005920099560171366, + 0.0032228343188762665, + 0.0037001061718910933, + 0.0046028573997318745, + 0.0025269712787121534, + 0.0053160712122917175, + 0.002594729419797659, + 0.004311854485422373, + 0.0037783058360219, + 0.0018517113057896495, + 0.003821991616860032, + 0.0018991312244907022, + 0.0016624554991722107, + 0.004153863061219454, + 0.0018458517733961344, + 0.005154768005013466, + 0.0031979484483599663, + 0.003706506220623851, + 0.005509029608219862, + 0.002417695475742221, + 0.003919126465916634, + 0.0042424313724040985, + 0.003421500790864229, + 0.002897881902754307, + 0.0028644991107285023, + 0.002459946321323514, + 0.002864900277927518, + 0.0032261719461530447, + 0.0023342568892985582, + 0.003738831263035536, + 0.005209201946854591, + 0.004444487392902374, + 0.009334228001534939, + 0.007288266438990831, + 0.006681425496935844, + 0.004478034097701311, + 0.005577126517891884, + 0.0076441047713160515, + 0.004866827744990587, + 0.006901751272380352, + 0.007724850904196501, + 0.00878890324383974, + 0.005264990497380495, + 0.0067197526805102825, + 0.012349812313914299, + 0.01949329860508442, + 0.008677076548337936, + 0.012293361127376556, + 0.09214752912521362, + 0.04902859777212143, + 0.09991570562124252, + 0.018087081611156464, + 0.10024978220462799, + 0.19749715924263, + 0.018025605008006096, + 0.029785603284835815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.029796818271279335 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_7/index_tensor([3392, 3392])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_7/index_tensor([3392, 3392])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..03720b191014ec057ebaa6442848e9c9cce35873 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_7/index_tensor([3392, 3392])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In stunning Makoto Shinkai style, a static low-angle medium shot captures the scene with crisp high frame rates and dramatic slow motion as the iconic Audi Sport Quattro S1 executes a violent launch across a gravel rally stage. The car's classic white chassis, adorned with signature red, gray, and black racing stripes, sways from its suspension under immense torque. The camera's focus is fixed on the rear wheel as a deeply treaded rally tire spins violently, its tread clawing and biting into the loose gravel surface, sending explosive showers of rock and thick plumes of fine dust into the air. The entire scene is backlit by a bright, low-hanging sun, creating a magnificent, expansive anamorphic lens flare that streaks horizontally across the frame with vibrant blues and oranges. The setting sun paints the sky in Shinkai's signature hues of deep purple, fiery orange, and soft pink, casting long, dramatic shadows and making every pebble in the air shimmer like a tiny suspended gem. The atmosphere is a powerful fusion of raw mechanical force and ethereal, ephemeral beauty, capturing a single, epic moment frozen in time.", + "seed": 29718, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_7/index_3392_seed_29718_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.015904851024970412, + "vlm_logprob": -0.015904851024970412, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.015904851024970412, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03173711895942688 + ], + "opposite_logprob_tokens": [ + -3.5317370891571045 + ], + "logprob_reward": -0.03173711895942688, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03173711895942688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15445587038993835, + 0.07850747555494308, + 0.09676128625869751, + 0.08920308202505112, + 0.12997804582118988, + 0.06120270863175392, + 0.09640021622180939, + 0.16576802730560303, + 0.06912285089492798, + 0.058600421994924545 + ], + "credit_map": [ + 0.020732751116156578, + 0.03955766558647156, + 0.011561904102563858, + 0.018707573413848877, + 0.02441348321735859, + 0.015472247265279293, + 0.015777166932821274, + 0.008233078755438328, + 0.015505003742873669, + 0.01812954992055893, + 0.006078772246837616, + 0.008188928477466106, + 0.007582144346088171, + 0.007811207324266434, + 0.009371915832161903, + 0.005839953664690256, + 0.02725125290453434, + 0.026267902925610542, + 0.006596439052373171, + 0.011149942874908447, + 0.004708119202405214, + 0.0076447357423603535, + 0.006549313198775053, + 0.006593577563762665, + 0.009017105214297771, + 0.009782138280570507, + 0.010671211406588554, + 0.012418292462825775, + 0.008539414033293724, + 0.014203154481947422, + 0.01625661365687847, + 0.00831515807658434, + 0.013873822055757046, + 0.010527788661420345, + 0.015215630643069744, + 0.025630595162510872, + 0.011007307097315788, + 0.010593024082481861, + 0.0325675830245018, + 0.010562295094132423, + 0.00697518652305007, + 0.007686142344027758, + 0.010761714540421963, + 0.014941050671041012, + 0.003659231821075082, + 0.004613602068275213, + 0.009230273775756359, + 0.0033355059567838907, + 0.008458076044917107, + 0.012281033210456371, + 0.014227801002562046, + 0.0169135183095932, + 0.013893538154661655, + 0.010610325261950493, + 0.012174378149211407, + 0.007841541431844234, + 0.01761053130030632, + 0.024073388427495956, + 0.022655170410871506, + 0.02178177610039711, + 0.024159535765647888, + 0.01793057844042778, + 0.020931750535964966, + 0.016625281423330307, + 0.0078899459913373, + 0.0080763204023242, + 0.008528098464012146, + 0.010026955045759678, + 0.0053841788321733475, + 0.011516178958117962, + 0.01013944298028946, + 0.007561732083559036, + 0.008027222938835621, + 0.014485684223473072, + 0.008616154082119465, + 0.006359506398439407, + 0.004276219289749861, + 0.004980185534805059, + 0.007437488064169884, + 0.0044179572723805904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.03173711895942688 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026364656165242195 + ], + "opposite_logprob_tokens": [ + -3.65136456489563 + ], + "logprob_reward": -0.026364656165242195, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026364656165242195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044652607291936874, + 0.19677913188934326, + 0.04094772785902023, + 0.05581897124648094, + 0.03457076847553253, + 0.22306819260120392, + 0.05170123279094696, + 0.2872762382030487, + 0.04168786108493805, + 0.02349729649722576 + ], + "credit_map": [ + 0.006087112240493298, + 0.00984863005578518, + 0.004627476446330547, + 0.004103789571672678, + 0.008799788542091846, + 0.003897527465596795, + 0.004356461577117443, + 0.0029318188317120075, + 0.056448351591825485, + 0.07199639081954956, + 0.008265289478003979, + 0.011767150834202766, + 0.014782889746129513, + 0.013804875314235687, + 0.012397335842251778, + 0.007316847797483206, + 0.007183430716395378, + 0.007454289589077234, + 0.0035523336846381426, + 0.004738996271044016, + 0.0045249550603330135, + 0.0046816980466246605, + 0.00551003310829401, + 0.00330199021846056, + 0.006515623535960913, + 0.006160689052194357, + 0.009571150876581669, + 0.006603062618523836, + 0.0045963916927576065, + 0.007009705528616905, + 0.009106158278882504, + 0.006256190128624439, + 0.00503637595102191, + 0.005749150644987822, + 0.005954574327915907, + 0.007850896567106247, + 0.0018930608639493585, + 0.002155269728973508, + 0.00420053955167532, + 0.0017308980459347367, + 0.025346605107188225, + 0.0294521301984787, + 0.03859975188970566, + 0.06044791638851166, + 0.012865958735346794, + 0.01998242363333702, + 0.024267764762043953, + 0.01210563164204359, + 0.007165470160543919, + 0.01478370651602745, + 0.006483389064669609, + 0.0063796560280025005, + 0.004603033885359764, + 0.0038716879207640886, + 0.004330347292125225, + 0.004083942621946335, + 0.033207423985004425, + 0.03607213869690895, + 0.046947598457336426, + 0.05561496317386627, + 0.023030441254377365, + 0.02413337491452694, + 0.034081168472766876, + 0.034189119935035706, + 0.0049245525151491165, + 0.005210794974118471, + 0.008042610250413418, + 0.007766794413328171, + 0.0019471836276352406, + 0.0035857006441801786, + 0.006063473876565695, + 0.004146748688071966, + 0.0038633879739791155, + 0.004316872451454401, + 0.002591049764305353, + 0.0035207599867135286, + 0.0012898542918264866, + 0.0022172462195158005, + 0.00321862637065351, + 0.0024794999044388533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.026364656165242195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036178866866976023 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.0036178866866976023, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036178866866976023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07302158325910568, + 0.15329709649085999, + 0.08936665207147598, + 0.18245764076709747, + 0.17324726283550262, + 0.11476092785596848, + 0.02040192484855652, + 0.08197278529405594, + 0.08648305386304855, + 0.024991022422909737 + ], + "credit_map": [ + 0.015310700982809067, + 0.020141780376434326, + 0.005902528762817383, + 0.005842151120305061, + 0.00880354642868042, + 0.004615040495991707, + 0.008001996204257011, + 0.0044038365595042706, + 0.04519786313176155, + 0.05697092041373253, + 0.007596401497721672, + 0.007319349329918623, + 0.011530403047800064, + 0.010402352549135685, + 0.007681105751544237, + 0.006598711479455233, + 0.011430060490965843, + 0.010696970857679844, + 0.008538472466170788, + 0.012493441812694073, + 0.010315134190022945, + 0.009196160361170769, + 0.018021659925580025, + 0.008674751967191696, + 0.01947023719549179, + 0.015039964579045773, + 0.023011688143014908, + 0.024009916931390762, + 0.014393623918294907, + 0.02527562901377678, + 0.04402380809187889, + 0.01723276637494564, + 0.01885903812944889, + 0.020111434161663055, + 0.027197476476430893, + 0.04042993113398552, + 0.013464273884892464, + 0.007935029454529285, + 0.03519301116466522, + 0.010057071223855019, + 0.013606430031359196, + 0.013608381152153015, + 0.02168549783527851, + 0.0326397530734539, + 0.005961371585726738, + 0.008675893768668175, + 0.012308149598538876, + 0.0062754470854997635, + 0.002511080587282777, + 0.0037158625200390816, + 0.0033191607799381018, + 0.0032079797238111496, + 0.0017085031140595675, + 0.0014510744949802756, + 0.0026557459495961666, + 0.00183251709677279, + 0.009090431034564972, + 0.010526753030717373, + 0.0131485341116786, + 0.017832042649388313, + 0.0072401403449475765, + 0.005223806947469711, + 0.01069722417742014, + 0.008213857188820839, + 0.00741660175845027, + 0.0075814388692379, + 0.012923161499202251, + 0.012892497703433037, + 0.0061498177237808704, + 0.01067086961120367, + 0.01774999313056469, + 0.011098666116595268, + 0.0037403625901788473, + 0.003418739652261138, + 0.0026972682680934668, + 0.003121608169749379, + 0.0024086132179945707, + 0.00298363552428782, + 0.004118741024285555, + 0.0025020530447363853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.0036178866866976023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188013039529324 + ], + "opposite_logprob_tokens": [ + -4.264187812805176 + ], + "logprob_reward": -0.014188013039529324, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188013039529324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34626227617263794, + 0.09072910249233246, + 0.11478046327829361, + 0.07718152552843094, + 0.05314725637435913, + 0.035027217119932175, + 0.03877415508031845, + 0.06473838537931442, + 0.0438675694167614, + 0.13549213111400604 + ], + "credit_map": [ + 0.06413815915584564, + 0.08060341328382492, + 0.028220172971487045, + 0.020094482228159904, + 0.07282640039920807, + 0.03278990834951401, + 0.03029734455049038, + 0.017292337492108345, + 0.015539911575615406, + 0.021496402099728584, + 0.011324156075716019, + 0.010541633702814579, + 0.010373148135840893, + 0.00719183636829257, + 0.007966441102325916, + 0.0062955585308372974, + 0.033428072929382324, + 0.03337246924638748, + 0.007852924056351185, + 0.010561203584074974, + 0.00828615203499794, + 0.00735129090026021, + 0.008396035060286522, + 0.005532301962375641, + 0.011791141703724861, + 0.011152266524732113, + 0.015438925474882126, + 0.012895146384835243, + 0.004069955553859472, + 0.007732380647212267, + 0.007547739893198013, + 0.006553959101438522, + 0.0057677216827869415, + 0.005976696964353323, + 0.007833420298993587, + 0.013298953883349895, + 0.004302953369915485, + 0.0032267712522298098, + 0.00898738019168377, + 0.003753348719328642, + 0.00435281777754426, + 0.0053472635336220264, + 0.005059623159468174, + 0.0072335731238126755, + 0.002185904886573553, + 0.0028546627145260572, + 0.005028138868510723, + 0.002965229097753763, + 0.005372676998376846, + 0.005780601873993874, + 0.007678670808672905, + 0.008765892125666142, + 0.0023638326674699783, + 0.002590805757790804, + 0.0034450185485184193, + 0.0027766514103859663, + 0.0065521481446921825, + 0.009430787526071072, + 0.007726170588284731, + 0.011255527846515179, + 0.00896365661174059, + 0.005915566347539425, + 0.009683935903012753, + 0.005210590548813343, + 0.004327478352934122, + 0.004427019506692886, + 0.0075735049322247505, + 0.008964856155216694, + 0.002895220648497343, + 0.004507251549512148, + 0.007063319440931082, + 0.004108913242816925, + 0.032399337738752365, + 0.022526973858475685, + 0.030876941978931427, + 0.012137087993323803, + 0.007808752823621035, + 0.009119969792664051, + 0.010624435730278492, + 0.009998629800975323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.014188013039529324 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036165802739560604 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.0036165802739560604, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036165802739560604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18445341289043427, + 0.37262195348739624, + 0.05345253273844719, + 0.0467366985976696, + 0.057866282761096954, + 0.05483688414096832, + 0.08466771245002747, + 0.05656680092215538, + 0.05436154082417488, + 0.03443611040711403 + ], + "credit_map": [ + 0.03202703595161438, + 0.036029521375894547, + 0.01565423607826233, + 0.022921225056052208, + 0.027262160554528236, + 0.017258042469620705, + 0.02195955626666546, + 0.011341650038957596, + 0.0978979840874672, + 0.13441000878810883, + 0.02009396441280842, + 0.026042575016617775, + 0.028965182602405548, + 0.02487349510192871, + 0.02483796328306198, + 0.015500822104513645, + 0.0074118697084486485, + 0.008400822058320045, + 0.008699907921254635, + 0.008749464526772499, + 0.0044045643880963326, + 0.005450394470244646, + 0.006582498550415039, + 0.003753014374524355, + 0.00664091669023037, + 0.0045677488669753075, + 0.00913584791123867, + 0.00811219122260809, + 0.0031845448538661003, + 0.005505796056240797, + 0.005012062843888998, + 0.004577596206218004, + 0.007611056789755821, + 0.007367002312093973, + 0.011800643987953663, + 0.014561529271304607, + 0.002963192295283079, + 0.002763122320175171, + 0.007797929458320141, + 0.0030018137767910957, + 0.00814878847450018, + 0.009463578462600708, + 0.009024268947541714, + 0.011972165666520596, + 0.003306084545329213, + 0.004195377696305513, + 0.005546781234443188, + 0.003179845865815878, + 0.00965848471969366, + 0.016080746427178383, + 0.015349218621850014, + 0.012313289567828178, + 0.008368561044335365, + 0.007820695638656616, + 0.009074005298316479, + 0.00600272323936224, + 0.006498673465102911, + 0.006850036792457104, + 0.006739577278494835, + 0.00695519195869565, + 0.0065480382181704044, + 0.007028977386653423, + 0.006761547178030014, + 0.009184763766825199, + 0.004611651878803968, + 0.0053502656519412994, + 0.008277345448732376, + 0.009291945025324821, + 0.003962366841733456, + 0.006479824427515268, + 0.01020988542586565, + 0.006178262177854776, + 0.006509797647595406, + 0.004284567199647427, + 0.009747075848281384, + 0.0029982151463627815, + 0.0020995414815843105, + 0.0031990897841751575, + 0.0026072878390550613, + 0.002990540349856019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0036165802739560604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In stunning Makoto Shinkai style, a static low-angle medium shot captures the scene with crisp high frame rates and dramatic slow motion as the iconic Audi Sport Quattro S1 executes a violent launch across a gravel rally stage. The car's classic white chassis, adorned with signature red, gray, and black racing stripes, sways from its suspension under immense torque. The camera's focus is fixed on the rear wheel as a deeply treaded rally tire spins violently, its tread clawing and biting into the loose gravel surface, sending explosive showers of rock and thick plumes of fine dust into the air. The entire scene is backlit by a bright, low-hanging sun, creating a magnificent, expansive anamorphic lens flare that streaks horizontally across the frame with vibrant blues and oranges. The setting sun paints the sky in Shinkai's signature hues of deep purple, fiery orange, and soft pink, casting long, dramatic shadows and making every pebble in the air shimmer like a tiny suspended gem. The atmosphere is a powerful fusion of raw mechanical force and ethereal, ephemeral beauty, capturing a single, epic moment frozen in time.", + "seed": 29719, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_7/index_3392_seed_29719_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.025702290702611208, + "vlm_logprob": -0.025702290702611208, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.025702290702611208, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07355351001024246 + ], + "opposite_logprob_tokens": [ + -2.6985535621643066 + ], + "logprob_reward": -0.07355351001024246, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07355351001024246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.932345449924469, + 0.006297251209616661, + 0.004860808607190847, + 0.012564130127429962, + 0.008161087520420551, + 0.0038959491066634655, + 0.0027620382606983185, + 0.0039865924045443535, + 0.010271583683788776, + 0.0148551594465971 + ], + "credit_map": [ + 0.009680389426648617, + 0.01857118494808674, + 0.005166645161807537, + 0.004231093917042017, + 0.44219282269477844, + 0.43324536085128784, + 0.010361882857978344, + 0.008896026760339737, + 0.00042720662895590067, + 0.0010007042437791824, + 0.0006072572432458401, + 0.0007903681835159659, + 0.0008103346917778254, + 0.0006567625096067786, + 0.0012840781128033996, + 0.0007205398287624121, + 0.00030684861121699214, + 0.0004154339840169996, + 0.0006203115917742252, + 0.0009103396441787481, + 0.000506918178871274, + 0.0005742918001487851, + 0.0009775636717677116, + 0.0005491010961122811, + 0.0009741061949171126, + 0.002779644913971424, + 0.0010142580140382051, + 0.0011059079552069306, + 0.0022528881672769785, + 0.00206605507992208, + 0.0014957152307033539, + 0.0008755544549785554, + 0.0006340579711832106, + 0.0009877968113869429, + 0.0011451521422713995, + 0.0011909845052286983, + 0.0014278891030699015, + 0.001041464158333838, + 0.0010104477405548096, + 0.0007232945063151419, + 0.00033052542130462825, + 0.000384727114578709, + 0.0007893906440585852, + 0.0004812204570043832, + 0.0004071626462973654, + 0.0004079598293174058, + 0.0006128909299150109, + 0.0004820718022529036, + 0.00018238177290186286, + 0.0002570135402493179, + 0.00046238876529969275, + 0.00039692368591204286, + 0.0002845856361091137, + 0.00037047057412564754, + 0.00040048634400591254, + 0.0004077881167177111, + 0.0002863560803234577, + 0.00031069963006302714, + 0.0005057610687799752, + 0.00043817900586873293, + 0.0006175810121931136, + 0.0006787822348996997, + 0.0005585850449278951, + 0.0005906482692807913, + 0.0007913789013400674, + 0.0028167327400296926, + 0.0009642012300901115, + 0.001020291936583817, + 0.0017055027419701219, + 0.0010031083365902305, + 0.000981449382379651, + 0.00098891812376678, + 0.0010164000559598207, + 0.0021740703377872705, + 0.0004959352663718164, + 0.00048765967949293554, + 0.005104073788970709, + 0.0032458011992275715, + 0.0013051528949290514, + 0.0010260655544698238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.07355351001024246 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020594101399183273 + ], + "opposite_logprob_tokens": [ + -3.8955941200256348 + ], + "logprob_reward": -0.020594101399183273, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594101399183273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6979968547821045, + 0.01943693310022354, + 0.04605861380696297, + 0.03288416191935539, + 0.0723506361246109, + 0.06062401086091995, + 0.014263132587075233, + 0.013622811064124107, + 0.02111121267080307, + 0.021651552990078926 + ], + "credit_map": [ + 0.007257526740431786, + 0.013911749236285686, + 0.0038836305029690266, + 0.0034486716613173485, + 0.33078229427337646, + 0.32485124468803406, + 0.007427797187119722, + 0.006434018723666668, + 0.0016072882572188973, + 0.002578823594376445, + 0.0023303870111703873, + 0.002822879236191511, + 0.002150840125977993, + 0.0020436446648091078, + 0.0044419732876122, + 0.0014610977377742529, + 0.0023009704891592264, + 0.0031480116304010153, + 0.007692810148000717, + 0.00937565602362156, + 0.0036709546111524105, + 0.005150127690285444, + 0.010211802087724209, + 0.0045082843862473965, + 0.0022511465940624475, + 0.004973640199750662, + 0.0066579850390553474, + 0.003965051379054785, + 0.0033028770703822374, + 0.0034118241164833307, + 0.004582199268043041, + 0.0037394415121525526, + 0.005343327298760414, + 0.008695604279637337, + 0.012567183934152126, + 0.01149127446115017, + 0.011255971156060696, + 0.010780482552945614, + 0.005668801721185446, + 0.006548001896589994, + 0.004242700524628162, + 0.01016695611178875, + 0.007522221654653549, + 0.009039477445185184, + 0.011480206623673439, + 0.007454922888427973, + 0.005744746420532465, + 0.004972779657691717, + 0.0008941981941461563, + 0.0014119717525318265, + 0.0021456964313983917, + 0.001991427270695567, + 0.0018464440945535898, + 0.0024035186506807804, + 0.0018240250647068024, + 0.0017458528745919466, + 0.0008549592457711697, + 0.00202499283477664, + 0.0012117401929572225, + 0.0011937579838559031, + 0.002138944575563073, + 0.003122017253190279, + 0.0015305456472560763, + 0.0015458540292456746, + 0.0015416856622323394, + 0.006655056960880756, + 0.0017584109446033835, + 0.0017563402652740479, + 0.004178557079285383, + 0.0025768703781068325, + 0.0012693664757534862, + 0.0013749260688200593, + 0.0010544111719354987, + 0.0026987239252775908, + 0.0011594517854973674, + 0.001270195934921503, + 0.006954846438020468, + 0.004872331395745277, + 0.0018825782462954521, + 0.0017590171191841364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.020594101399183273 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199766330420971 + ], + "opposite_logprob_tokens": [ + -4.264199733734131 + ], + "logprob_reward": -0.014199766330420971, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199766330420971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8332191109657288, + 0.014703967608511448, + 0.007888665422797203, + 0.01612156443297863, + 0.04376263543963432, + 0.0148085355758667, + 0.008412383496761322, + 0.01731177233159542, + 0.009584733285009861, + 0.034186623990535736 + ], + "credit_map": [ + 0.008509132079780102, + 0.01639387384057045, + 0.004767941776663065, + 0.0034355944953858852, + 0.39600592851638794, + 0.38867878913879395, + 0.007461797446012497, + 0.007966135628521442, + 0.0010328025091439486, + 0.001766491332091391, + 0.0012613178696483374, + 0.001744610141031444, + 0.0018454582896083593, + 0.0016311625950038433, + 0.0033650074619799852, + 0.002057117410004139, + 0.0005440372042357922, + 0.0006865224568173289, + 0.001489850226789713, + 0.0012915268307551742, + 0.0005426398711279035, + 0.0008374489843845367, + 0.0014133157674223185, + 0.0010833240812644362, + 0.001082971808500588, + 0.0013717330293729901, + 0.0038084553088992834, + 0.0026290116365998983, + 0.0020071337930858135, + 0.0021369801834225655, + 0.0015000493731349707, + 0.0015852302312850952, + 0.0029957315418869257, + 0.005290503613650799, + 0.009346379898488522, + 0.007584794424474239, + 0.0060124765150249004, + 0.006315321195870638, + 0.0030075707472860813, + 0.0032098593655973673, + 0.0012626202078536153, + 0.001170059316791594, + 0.0030094790272414684, + 0.0021423003636300564, + 0.00125764065887779, + 0.0015892800875008106, + 0.0023741868790239096, + 0.0020029714796692133, + 0.0005163892637938261, + 0.0006971223629079759, + 0.0014322235947474837, + 0.0011786059476435184, + 0.0013842915650457144, + 0.0012785621220245957, + 0.0007464790833182633, + 0.0011787095572799444, + 0.0011323930229991674, + 0.00204149866476655, + 0.0013258676044642925, + 0.001640382339246571, + 0.0046690646559000015, + 0.002909420756623149, + 0.0019035846926271915, + 0.0016895622247830033, + 0.0008271599654108286, + 0.0014869144652038813, + 0.0011306360829621553, + 0.0015499591827392578, + 0.0011259469902142882, + 0.0015739287482574582, + 0.0008089252514764667, + 0.0010812629479914904, + 0.00245900871232152, + 0.005602634511888027, + 0.0013063802616670728, + 0.0012658752966672182, + 0.011349190957844257, + 0.0068572042509913445, + 0.0030423032585531473, + 0.0023040305823087692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.014199766330420971 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253335177898407 + ], + "opposite_logprob_tokens": [ + -4.387533187866211 + ], + "logprob_reward": -0.01253335177898407, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253335177898407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9498777985572815, + 0.00935811921954155, + 0.00486052455380559, + 0.003286745399236679, + 0.002100627636536956, + 0.0029436887707561255, + 0.009613380767405033, + 0.010089344345033169, + 0.002247343072667718, + 0.005622402764856815 + ], + "credit_map": [ + 0.009528247639536858, + 0.01924050971865654, + 0.005264695268124342, + 0.003713367274031043, + 0.45242491364479065, + 0.44264042377471924, + 0.007921465672552586, + 0.009144158102571964, + 0.0009143450297415257, + 0.002233011182397604, + 0.0006268254946917295, + 0.0009884742321446538, + 0.002059635240584612, + 0.0011158828856423497, + 0.0008846118580549955, + 0.000535333703737706, + 0.00034936153679154813, + 0.0005004387930966914, + 0.0007352303364314139, + 0.000748642545659095, + 0.0005507266614586115, + 0.0005204015760682523, + 0.0009299047524109483, + 0.0005258187302388251, + 0.00022301063290797174, + 0.0006867651827633381, + 0.0005176312988623977, + 0.00033495610114187, + 0.0005540083511732519, + 0.00047854037256911397, + 0.0002687269588932395, + 0.00022310657368507236, + 0.00016512579168193042, + 0.00025686080334708095, + 0.0002919354010373354, + 0.00035335641587153077, + 0.0003338102251291275, + 0.0002603916509542614, + 0.00022430434182751924, + 0.0002148430357920006, + 0.00023522492847405374, + 0.0003813294169958681, + 0.0004813497653231025, + 0.00034448649967089295, + 0.0002888840972445905, + 0.00040934840217232704, + 0.00038361168117262423, + 0.0004194538341835141, + 0.0006105753709562123, + 0.0010575928026810288, + 0.0016884919023141265, + 0.001376237254589796, + 0.0011309500550851226, + 0.001440631807781756, + 0.0012644887901842594, + 0.0010444129584357142, + 0.0006472130189649761, + 0.0025445991195738316, + 0.0008049372700043023, + 0.0009957560105249286, + 0.001467881491407752, + 0.0019908552058041096, + 0.0008806088590063155, + 0.0007574921473860741, + 0.00015805203292984515, + 0.0005930169718340039, + 0.00033483700826764107, + 0.0002349954011151567, + 0.0003261232632212341, + 0.00024506778572686017, + 0.00018127843213733286, + 0.00017397216288372874, + 0.0003636481415014714, + 0.0009534384007565677, + 0.00025117030600085855, + 0.0002800970687530935, + 0.001755360746756196, + 0.0011766240932047367, + 0.00047109831939451396, + 0.00037096557207405567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.01253335177898407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630723994225264 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630723994225264, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630723994225264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8537470102310181, + 0.01969400979578495, + 0.016469407826662064, + 0.024754976853728294, + 0.008125166408717632, + 0.006191705819219351, + 0.011757083237171173, + 0.006429173983633518, + 0.018432823941111565, + 0.0343986451625824 + ], + "credit_map": [ + 0.00843885075300932, + 0.016758624464273453, + 0.004783689975738525, + 0.0036096491385251284, + 0.4040992558002472, + 0.39858126640319824, + 0.009371098130941391, + 0.00810463447123766, + 0.0018949147779494524, + 0.0049269250594079494, + 0.0011538236867636442, + 0.0018040059367194772, + 0.004556450992822647, + 0.0028297307435423136, + 0.0015943651087582111, + 0.0009337954688817263, + 0.0007850260008126497, + 0.001069301157258451, + 0.002782112220302224, + 0.0036000388208776712, + 0.0015131880063563585, + 0.0018795238574966788, + 0.0032394526060670614, + 0.0016007665544748306, + 0.0018446739995852113, + 0.0062726940959692, + 0.002732788911089301, + 0.002374769886955619, + 0.004933702293783426, + 0.004329559858888388, + 0.0012439667480066419, + 0.0010228216415271163, + 0.0005663166521117091, + 0.0008505741134285927, + 0.0016067603137344122, + 0.0012392515782266855, + 0.0011048160959035158, + 0.0013225538423284888, + 0.0007140616071410477, + 0.0007208332535810769, + 0.0004187397425994277, + 0.0006021364824846387, + 0.0009792279452085495, + 0.0009481985471211374, + 0.0007374502602033317, + 0.0008343547815456986, + 0.0007381151663139462, + 0.0009334838250651956, + 0.0007442376809194684, + 0.0011859422083944082, + 0.0019847177900373936, + 0.0018697050400078297, + 0.0013889801921322942, + 0.0016806955682113767, + 0.0015818659448996186, + 0.0013209402095526457, + 0.00033959970460273325, + 0.0005647802026942372, + 0.0008321463828906417, + 0.0009589196997694671, + 0.0011488973395898938, + 0.0012531352695077658, + 0.0006678602658212185, + 0.0006638356717303395, + 0.0012599442852661014, + 0.005788532085716724, + 0.002002998720854521, + 0.0015941167948767543, + 0.002573618432506919, + 0.0030086436308920383, + 0.0009116153814829886, + 0.0012933567631989717, + 0.002271557692438364, + 0.005375770851969719, + 0.0009271724848076701, + 0.0012004560558125377, + 0.011934798210859299, + 0.007346551399677992, + 0.003094174200668931, + 0.0022481693886220455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.007630723994225264 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In stunning Makoto Shinkai style, a static low-angle medium shot captures the scene with crisp high frame rates and dramatic slow motion as the iconic Audi Sport Quattro S1 executes a violent launch across a gravel rally stage. The car's classic white chassis, adorned with signature red, gray, and black racing stripes, sways from its suspension under immense torque. The camera's focus is fixed on the rear wheel as a deeply treaded rally tire spins violently, its tread clawing and biting into the loose gravel surface, sending explosive showers of rock and thick plumes of fine dust into the air. The entire scene is backlit by a bright, low-hanging sun, creating a magnificent, expansive anamorphic lens flare that streaks horizontally across the frame with vibrant blues and oranges. The setting sun paints the sky in Shinkai's signature hues of deep purple, fiery orange, and soft pink, casting long, dramatic shadows and making every pebble in the air shimmer like a tiny suspended gem. The atmosphere is a powerful fusion of raw mechanical force and ethereal, ephemeral beauty, capturing a single, epic moment frozen in time.", + "seed": 29720, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_7/index_3392_seed_29720_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.027507949993014334, + "vlm_logprob": -0.027507949993014334, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.027507949993014334, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09049361944198608 + ], + "opposite_logprob_tokens": [ + -2.465493679046631 + ], + "logprob_reward": -0.09049361944198608, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09049361944198608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11710862070322037, + 0.04117941856384277, + 0.036278557032346725, + 0.10650241374969482, + 0.045280419290065765, + 0.5300341248512268, + 0.03077526018023491, + 0.031201941892504692, + 0.03086147829890251, + 0.03077773191034794 + ], + "credit_map": [ + 0.012704562395811081, + 0.014245127327740192, + 0.009317007847130299, + 0.0106486352160573, + 0.013544677756726742, + 0.01703113503754139, + 0.01753774844110012, + 0.02207973785698414, + 0.0037090873811393976, + 0.0042799911461770535, + 0.0036048886831849813, + 0.004778643604367971, + 0.004380019847303629, + 0.006042012479156256, + 0.008313900791108608, + 0.006070880219340324, + 0.007372555788606405, + 0.008623648434877396, + 0.0027204372454434633, + 0.0033208923414349556, + 0.0027736311312764883, + 0.0037947644013911486, + 0.004607576411217451, + 0.003065054537728429, + 0.02449827641248703, + 0.030115323141217232, + 0.0075569432228803635, + 0.00911886990070343, + 0.007175470236688852, + 0.007145363837480545, + 0.009962115436792374, + 0.010930064134299755, + 0.006110525690019131, + 0.008244038559496403, + 0.004544532857835293, + 0.0062329573556780815, + 0.0031414527911692858, + 0.003306887112557888, + 0.004428971093147993, + 0.009271055459976196, + 0.11684482544660568, + 0.22416019439697266, + 0.008673946373164654, + 0.014522198587656021, + 0.12420640885829926, + 0.01977131888270378, + 0.012400271371006966, + 0.009454972110688686, + 0.0026649930514395237, + 0.00442122109234333, + 0.0032030802685767412, + 0.0056654177606105804, + 0.005401304457336664, + 0.0033683052752166986, + 0.0029083259869366884, + 0.003142615081742406, + 0.0015380942495539784, + 0.008014696650207043, + 0.0015606918605044484, + 0.0017242531757801771, + 0.011471073143184185, + 0.0036390821915119886, + 0.0015732445754110813, + 0.0016808079089969397, + 0.002430922584608197, + 0.0033464529551565647, + 0.0031603951938450336, + 0.004895503632724285, + 0.0037763752043247223, + 0.005041299853473902, + 0.0036379883531481028, + 0.004572541452944279, + 0.0021286671981215477, + 0.00978358555585146, + 0.0020133042708039284, + 0.0021015936508774757, + 0.003651147708296776, + 0.006958676502108574, + 0.0023714134003967047, + 0.0017693450208753347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.09049361944198608 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792622178792953 + ], + "opposite_logprob_tokens": [ + -4.634792804718018 + ], + "logprob_reward": -0.009792622178792953, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792622178792953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03206455335021019, + 0.007414294872432947, + 0.049352556467056274, + 0.02873101644217968, + 0.15259434282779694, + 0.5202179551124573, + 0.09826403111219406, + 0.02571164071559906, + 0.025843461975455284, + 0.059806156903505325 + ], + "credit_map": [ + 0.004983552265912294, + 0.004975443705916405, + 0.008326131850481033, + 0.0035528093576431274, + 0.0027329856529831886, + 0.0020262603648006916, + 0.002792954910546541, + 0.0026744173374027014, + 0.0006930556264705956, + 0.0008386753033846617, + 0.0006310879834927619, + 0.0008588590426370502, + 0.0009993489366024733, + 0.0008671575924381614, + 0.0012157162418588996, + 0.0013103937963023782, + 0.014129356481134892, + 0.0167106781154871, + 0.001890520448796451, + 0.003460843116044998, + 0.0033465970773249865, + 0.0027663903310894966, + 0.003454337827861309, + 0.0035938320215791464, + 0.005579438526183367, + 0.006774233188480139, + 0.005453127436339855, + 0.003146213013678789, + 0.0016381762688979506, + 0.0015488232020288706, + 0.002338060410693288, + 0.0022529440466314554, + 0.012538718059659004, + 0.01932966709136963, + 0.012670096009969711, + 0.023791907355189323, + 0.015662316232919693, + 0.013032895512878895, + 0.02312614768743515, + 0.03244258463382721, + 0.07507085055112839, + 0.19937486946582794, + 0.007052164524793625, + 0.012262401171028614, + 0.1669994294643402, + 0.036578934639692307, + 0.013754851184785366, + 0.009124396368861198, + 0.007478935178369284, + 0.009104249998927116, + 0.011054412461817265, + 0.01181221567094326, + 0.011659862473607063, + 0.011020875535905361, + 0.020170461386442184, + 0.01596301794052124, + 0.0016983391251415014, + 0.005756629630923271, + 0.0014358182670548558, + 0.0026539196260273457, + 0.006955512799322605, + 0.0024934217799454927, + 0.003001389792189002, + 0.0017166093457490206, + 0.0023280649911612272, + 0.002719977404922247, + 0.003278574673458934, + 0.003537679323926568, + 0.0034652450121939182, + 0.005091249477118254, + 0.0026135763619095087, + 0.0028090947307646275, + 0.003087247023358941, + 0.02149812877178192, + 0.002481755567714572, + 0.0033532564993947744, + 0.004171013832092285, + 0.0199835654348135, + 0.002376433229073882, + 0.002854753052815795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.009792622178792953 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016074467450380325 + ], + "opposite_logprob_tokens": [ + -4.1410746574401855 + ], + "logprob_reward": -0.016074467450380325, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016074467450380325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03299741446971893, + 0.04121214151382446, + 0.02174489200115204, + 0.03811700642108917, + 0.129794642329216, + 0.6187541484832764, + 0.05244852229952812, + 0.02466309443116188, + 0.023692164570093155, + 0.016575993970036507 + ], + "credit_map": [ + 0.0049817427061498165, + 0.006413963623344898, + 0.0038760537281632423, + 0.0028904119972139597, + 0.003467113245278597, + 0.002496125642210245, + 0.00440274178981781, + 0.0044692582450807095, + 0.0035839604679495096, + 0.004394357558339834, + 0.004494459833949804, + 0.008396808058023453, + 0.0039341747760772705, + 0.004366832319647074, + 0.006513143889605999, + 0.005528400186449289, + 0.0022225193679332733, + 0.0021160454489290714, + 0.0024056662805378437, + 0.004218921065330505, + 0.0017213732935488224, + 0.0030927241314202547, + 0.0030819065868854523, + 0.0028857344295829535, + 0.005919643212109804, + 0.006864595226943493, + 0.004800242371857166, + 0.007228122092783451, + 0.002375067211687565, + 0.0025928837712854147, + 0.0046288566663861275, + 0.0037075928412377834, + 0.009626352228224277, + 0.01258794404566288, + 0.011520643718540668, + 0.02339581400156021, + 0.013466513715684414, + 0.01198772992938757, + 0.02001919597387314, + 0.02719043381512165, + 0.11438059061765671, + 0.24887065589427948, + 0.009981095790863037, + 0.01775258779525757, + 0.16800540685653687, + 0.03174557164311409, + 0.01643940806388855, + 0.011578816920518875, + 0.004333307966589928, + 0.005303337704390287, + 0.00592480693012476, + 0.0068403584882617, + 0.005902744363993406, + 0.0063782427459955215, + 0.01010003313422203, + 0.007665683981031179, + 0.001295002643018961, + 0.005740315653383732, + 0.0020576240494847298, + 0.0020756616722792387, + 0.007788826711475849, + 0.002215171931311488, + 0.0014132164651528, + 0.002077273325994611, + 0.00224698637612164, + 0.0024983647745102644, + 0.003335214452818036, + 0.004239937756210566, + 0.002555529586970806, + 0.0026889184955507517, + 0.0024085810873657465, + 0.0037186299450695515, + 0.0020285097416490316, + 0.002683128695935011, + 0.0018963796319440007, + 0.0019868446979671717, + 0.0026312763802707195, + 0.001472192583605647, + 0.002613511634990573, + 0.0012641499051824212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.016074467450380325 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863721314817667 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.00863721314817667, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863721314817667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022015465423464775, + 0.011041236110031605, + 0.022833457216620445, + 0.009645072743296623, + 0.025399383157491684, + 0.8083102107048035, + 0.013595524244010448, + 0.02343105711042881, + 0.028902795165777206, + 0.03482585400342941 + ], + "credit_map": [ + 0.004843594040721655, + 0.005113608203828335, + 0.0029331515543162823, + 0.002157461130991578, + 0.0020178519189357758, + 0.0011908685555681586, + 0.0019145667320117354, + 0.0018443597946316004, + 0.001087109325453639, + 0.0014157751575112343, + 0.0016011596890166402, + 0.0011994914384558797, + 0.0017605200409889221, + 0.0011920025572180748, + 0.0014535613590851426, + 0.001331615960225463, + 0.00689879572018981, + 0.008314395323395729, + 0.0008430271409451962, + 0.0013737677363678813, + 0.0018123961053788662, + 0.000933452567551285, + 0.0014181060250848532, + 0.0012395123485475779, + 0.001157493912614882, + 0.0012578062014654279, + 0.001214880496263504, + 0.0015089217340573668, + 0.0007624384015798569, + 0.0007514184690080583, + 0.0012962811160832644, + 0.0016958300257101655, + 0.0034301055129617453, + 0.0051645757630467415, + 0.0026545010041445494, + 0.0026746063958853483, + 0.0023246798664331436, + 0.0020448898430913687, + 0.002664166036993265, + 0.004441856872290373, + 0.1518113762140274, + 0.33169832825660706, + 0.011358196847140789, + 0.01859348826110363, + 0.22278878092765808, + 0.0402991808950901, + 0.01855260506272316, + 0.013208222575485706, + 0.0013596514472737908, + 0.0015679027419537306, + 0.001653757062740624, + 0.00230883969925344, + 0.001627308432944119, + 0.0013184059644117951, + 0.0022263512946665287, + 0.0015333057381212711, + 0.0010402369080111384, + 0.006347020156681538, + 0.0012905477778986096, + 0.0020400863140821457, + 0.007483185268938541, + 0.002154518151655793, + 0.001841330318711698, + 0.0012341310502961278, + 0.0022007913794368505, + 0.0037552681751549244, + 0.0025654020719230175, + 0.004413737915456295, + 0.00385864544659853, + 0.005547626875340939, + 0.003036054316908121, + 0.003525265958160162, + 0.002249825745820999, + 0.011584499850869179, + 0.0014650971861556172, + 0.0017681397730484605, + 0.0024543977342545986, + 0.011548693291842937, + 0.0019752918742597103, + 0.0017799065681174397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.00863721314817667 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541827745735645 + ], + "opposite_logprob_tokens": [ + -4.387541770935059 + ], + "logprob_reward": -0.012541827745735645, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541827745735645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06093841418623924, + 0.06496349722146988, + 0.060160115361213684, + 0.033886466175317764, + 0.06759036332368851, + 0.2245008945465088, + 0.18026189506053925, + 0.026455411687493324, + 0.03803957253694534, + 0.24320346117019653 + ], + "credit_map": [ + 0.008627446368336678, + 0.00775620061904192, + 0.013187572360038757, + 0.0069734444841742516, + 0.005194585770368576, + 0.004496634937822819, + 0.007816716097295284, + 0.006885805167257786, + 0.009359026327729225, + 0.010908202268183231, + 0.008034913800656796, + 0.008855580352246761, + 0.009956389665603638, + 0.005192141514271498, + 0.006990530528128147, + 0.005666704382747412, + 0.019029220566153526, + 0.022538434714078903, + 0.002095194300636649, + 0.0046767583116889, + 0.0042604985646903515, + 0.0024898615665733814, + 0.0027744113467633724, + 0.002295734826475382, + 0.004275476559996605, + 0.0035427536349743605, + 0.009660374373197556, + 0.004230794031172991, + 0.0023720324970781803, + 0.002323285210877657, + 0.0039216321893036366, + 0.003560112090781331, + 0.013158849440515041, + 0.01910344697535038, + 0.005408756900578737, + 0.005752724129706621, + 0.004876480903476477, + 0.004619532730430365, + 0.006173992529511452, + 0.008496574126183987, + 0.043435852974653244, + 0.0903536006808281, + 0.004359976388514042, + 0.00770556041970849, + 0.054627809673547745, + 0.009470357559621334, + 0.008320196531713009, + 0.006227509118616581, + 0.009003101848065853, + 0.04319155961275101, + 0.009921028278768063, + 0.016388561576604843, + 0.06165777146816254, + 0.019006984308362007, + 0.011984686367213726, + 0.009108180180191994, + 0.002653147792443633, + 0.003799258265644312, + 0.0036585612688213587, + 0.0026683127507567406, + 0.005618617404252291, + 0.0035828198306262493, + 0.0020159033592790365, + 0.0024587884545326233, + 0.003222182858735323, + 0.005261212587356567, + 0.004058726131916046, + 0.00457965862005949, + 0.005946852266788483, + 0.006994225550442934, + 0.003327741287648678, + 0.004648969043046236, + 0.012313458137214184, + 0.09546230733394623, + 0.004729183856397867, + 0.009629509411752224, + 0.019092248752713203, + 0.08562000095844269, + 0.008985411375761032, + 0.007371299900114536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.012541827745735645 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In stunning Makoto Shinkai style, a static low-angle medium shot captures the scene with crisp high frame rates and dramatic slow motion as the iconic Audi Sport Quattro S1 executes a violent launch across a gravel rally stage. The car's classic white chassis, adorned with signature red, gray, and black racing stripes, sways from its suspension under immense torque. The camera's focus is fixed on the rear wheel as a deeply treaded rally tire spins violently, its tread clawing and biting into the loose gravel surface, sending explosive showers of rock and thick plumes of fine dust into the air. The entire scene is backlit by a bright, low-hanging sun, creating a magnificent, expansive anamorphic lens flare that streaks horizontally across the frame with vibrant blues and oranges. The setting sun paints the sky in Shinkai's signature hues of deep purple, fiery orange, and soft pink, casting long, dramatic shadows and making every pebble in the air shimmer like a tiny suspended gem. The atmosphere is a powerful fusion of raw mechanical force and ethereal, ephemeral beauty, capturing a single, epic moment frozen in time.", + "seed": 29721, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_7/index_3392_seed_29721_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01508423862978816, + "vlm_logprob": -0.01508423862978816, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.01508423862978816, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04058503732085228 + ], + "opposite_logprob_tokens": [ + -3.2905850410461426 + ], + "logprob_reward": -0.04058503732085228, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04058503732085228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19104664027690887, + 0.0694931298494339, + 0.12994523346424103, + 0.12055598199367523, + 0.10130928456783295, + 0.07411521673202515, + 0.13642513751983643, + 0.06367993354797363, + 0.07418768107891083, + 0.039241742342710495 + ], + "credit_map": [ + 0.018929939717054367, + 0.021130414679646492, + 0.036886442452669144, + 0.02469445765018463, + 0.020774435251951218, + 0.014671988785266876, + 0.028964677825570107, + 0.02499430812895298, + 0.0065056681632995605, + 0.007070986554026604, + 0.008194731548428535, + 0.011739528737962246, + 0.009145844727754593, + 0.006469602696597576, + 0.011064126156270504, + 0.009302645921707153, + 0.013121271505951881, + 0.01652870886027813, + 0.014961115084588528, + 0.018313776701688766, + 0.017649514600634575, + 0.012354078702628613, + 0.02291928231716156, + 0.014097504317760468, + 0.009397419169545174, + 0.013033391907811165, + 0.016933703795075417, + 0.020154548808932304, + 0.014480053447186947, + 0.01138017512857914, + 0.020780857652425766, + 0.014395837672054768, + 0.010903957299888134, + 0.010699956677854061, + 0.01183281559497118, + 0.017561335116624832, + 0.010915352962911129, + 0.006221343297511339, + 0.022878658026456833, + 0.0102958669885993, + 0.008048676885664463, + 0.009096682071685791, + 0.009702315554022789, + 0.011865123175084591, + 0.008707461878657341, + 0.007059208583086729, + 0.01263614185154438, + 0.0069996146485209465, + 0.00993665587157011, + 0.016211779788136482, + 0.015186823904514313, + 0.017589695751667023, + 0.021820295602083206, + 0.00897310022264719, + 0.033577196300029755, + 0.013129599392414093, + 0.005725921597331762, + 0.007268569897860289, + 0.008622284047305584, + 0.005717242136597633, + 0.005962618626654148, + 0.007078934460878372, + 0.0071159363724291325, + 0.01618843711912632, + 0.006036222912371159, + 0.00758374622091651, + 0.011056628078222275, + 0.011123418807983398, + 0.007219571620225906, + 0.00808674842119217, + 0.01371429581195116, + 0.009367059916257858, + 0.004995591007173061, + 0.004254127852618694, + 0.0037506886292248964, + 0.004561396315693855, + 0.005449788644909859, + 0.006939988117665052, + 0.0037714429199695587, + 0.005518720485270023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.04058503732085228 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005271582398563623 + ], + "opposite_logprob_tokens": [ + -5.2552714347839355 + ], + "logprob_reward": -0.005271582398563623, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005271582398563623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0902915745973587, + 0.042912807315588, + 0.14880219101905823, + 0.08613520860671997, + 0.06761331111192703, + 0.03729737922549248, + 0.09587398916482925, + 0.10607892274856567, + 0.19145464897155762, + 0.13353992998600006 + ], + "credit_map": [ + 0.013236761093139648, + 0.011972214095294476, + 0.01816444844007492, + 0.012045107781887054, + 0.0075009833090007305, + 0.004861972760409117, + 0.011446607299149036, + 0.011063487268984318, + 0.004241376183927059, + 0.004983187653124332, + 0.006469733081758022, + 0.007535587530583143, + 0.004857845604419708, + 0.003063216106966138, + 0.007134878542274237, + 0.0046269865706563, + 0.016560668125748634, + 0.016412485390901566, + 0.02252005785703659, + 0.047876209020614624, + 0.00989176519215107, + 0.008296425454318523, + 0.01445096917450428, + 0.012793617323040962, + 0.009184869937598705, + 0.00936916284263134, + 0.014619114808738232, + 0.02426609769463539, + 0.00745960371568799, + 0.004370307084172964, + 0.010051221586763859, + 0.006814831867814064, + 0.006658127531409264, + 0.00956377387046814, + 0.0070504434406757355, + 0.012331727892160416, + 0.008099309168756008, + 0.005109007004648447, + 0.011102059856057167, + 0.007698864210397005, + 0.003869313281029463, + 0.004178810864686966, + 0.0051994118839502335, + 0.006028333213180304, + 0.0044854930602014065, + 0.0030406550504267216, + 0.006284863222390413, + 0.004210504237562418, + 0.00996130146086216, + 0.01238070335239172, + 0.015820972621440887, + 0.013738748617470264, + 0.012069596908986568, + 0.007676694076508284, + 0.01324194110929966, + 0.010984033346176147, + 0.0072881970554590225, + 0.011084317229688168, + 0.00939391553401947, + 0.016688497737050056, + 0.010968448594212532, + 0.013568539172410965, + 0.01772530935704708, + 0.0193617045879364, + 0.014397639781236649, + 0.02500583603978157, + 0.024343425408005714, + 0.031050896272063255, + 0.020944051444530487, + 0.020521782338619232, + 0.030599351972341537, + 0.024591675028204918, + 0.009617545641958714, + 0.015958769246935844, + 0.015396255068480968, + 0.01480321679264307, + 0.017419034615159035, + 0.027249688282608986, + 0.018065623939037323, + 0.015029793605208397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.005271582398563623 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079810559749603 + ], + "opposite_logprob_tokens": [ + -4.511079788208008 + ], + "logprob_reward": -0.011079810559749603, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079810559749603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19469252228736877, + 0.07445012032985687, + 0.0637604296207428, + 0.04969567805528641, + 0.12856104969978333, + 0.06385262310504913, + 0.08184245973825455, + 0.1414383053779602, + 0.06723540276288986, + 0.13447146117687225 + ], + "credit_map": [ + 0.020341627299785614, + 0.025543859228491783, + 0.04202374443411827, + 0.02760341204702854, + 0.021154476329684258, + 0.015222345478832722, + 0.02246728725731373, + 0.020335765555500984, + 0.00873894989490509, + 0.007870904169976711, + 0.01219819113612175, + 0.013023545034229755, + 0.007898710668087006, + 0.006822583265602589, + 0.009065872058272362, + 0.008831359446048737, + 0.0072368960827589035, + 0.00762367295101285, + 0.008090018294751644, + 0.01433225255459547, + 0.006432955153286457, + 0.005894726607948542, + 0.007657281123101711, + 0.006492622662335634, + 0.004720694851130247, + 0.005177087616175413, + 0.0069253453984856606, + 0.011495298705995083, + 0.004507727921009064, + 0.00380856660194695, + 0.007022562436759472, + 0.006038387771695852, + 0.0153181292116642, + 0.015545198693871498, + 0.018400436267256737, + 0.01953541301190853, + 0.015013237483799458, + 0.01110493391752243, + 0.01867622882127762, + 0.014967457391321659, + 0.00590383680537343, + 0.008341833017766476, + 0.009804755449295044, + 0.011077533476054668, + 0.007452415302395821, + 0.004124854691326618, + 0.00928090326488018, + 0.00786648690700531, + 0.006310835015028715, + 0.009155421517789364, + 0.010013723745942116, + 0.01036336924880743, + 0.009981381706893444, + 0.008815730921924114, + 0.01613781787455082, + 0.011064168997108936, + 0.010623626410961151, + 0.012584232725203037, + 0.015651216730475426, + 0.021878641098737717, + 0.011965553276240826, + 0.016882721334695816, + 0.023359384387731552, + 0.028492921963334084, + 0.005203343462198973, + 0.006463374011218548, + 0.009869490750133991, + 0.011610032059252262, + 0.007840736769139767, + 0.008603104390203953, + 0.009706859476864338, + 0.007938455790281296, + 0.010910636745393276, + 0.018558712676167488, + 0.01365703996270895, + 0.012060489505529404, + 0.01798737421631813, + 0.028872979804873466, + 0.0168014969676733, + 0.01562272198498249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.011079810559749603 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253500021994114 + ], + "opposite_logprob_tokens": [ + -4.387535095214844 + ], + "logprob_reward": -0.01253500021994114, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253500021994114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10821002721786499, + 0.06913716346025467, + 0.08659303933382034, + 0.06165897473692894, + 0.2880329191684723, + 0.21260052919387817, + 0.06457333266735077, + 0.02154509536921978, + 0.05748257786035538, + 0.03016633167862892 + ], + "credit_map": [ + 0.011509009636938572, + 0.016294369474053383, + 0.02378232218325138, + 0.017641153186559677, + 0.010768543928861618, + 0.007209864445030689, + 0.012166015803813934, + 0.008838748559355736, + 0.010406212881207466, + 0.009557340294122696, + 0.015232168138027191, + 0.010065916925668716, + 0.007400196045637131, + 0.004143568687140942, + 0.006402477156370878, + 0.005929285194724798, + 0.009598634205758572, + 0.009850019589066505, + 0.014322330243885517, + 0.027002085000276566, + 0.007080917712301016, + 0.004162285942584276, + 0.00818485114723444, + 0.006391922011971474, + 0.007738626562058926, + 0.008153616450726986, + 0.011279013939201832, + 0.014885952696204185, + 0.0055674235336482525, + 0.003020089352503419, + 0.006229223683476448, + 0.004785030148923397, + 0.03414274379611015, + 0.034973230212926865, + 0.03821209445595741, + 0.06794171035289764, + 0.03057013638317585, + 0.014009919948875904, + 0.041713155806064606, + 0.026469942182302475, + 0.024573925882577896, + 0.028161175549030304, + 0.025379501283168793, + 0.032213542610406876, + 0.02920057438313961, + 0.019504018127918243, + 0.03283832222223282, + 0.020729467272758484, + 0.009585240855813026, + 0.009750843979418278, + 0.007766725495457649, + 0.006643290631473064, + 0.01222349051386118, + 0.00491774408146739, + 0.009215626865625381, + 0.004470375832170248, + 0.0027424886357039213, + 0.0021130014210939407, + 0.0026003532111644745, + 0.002757860580459237, + 0.002492643892765045, + 0.0021062209270894527, + 0.0033716135658323765, + 0.00336091429926455, + 0.004416048526763916, + 0.0057920911349356174, + 0.007901669479906559, + 0.010434223338961601, + 0.005303252954035997, + 0.0063934181816875935, + 0.010794737376272678, + 0.006447137799113989, + 0.002942269667983055, + 0.0038730704691261053, + 0.004118023905903101, + 0.0034484700299799442, + 0.003957421984523535, + 0.004703570622950792, + 0.0036156491842120886, + 0.003507854649797082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.01253500021994114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Audi Sport Quattro S1 rally car with a white chassis and red, gray, and black racing stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a violent launch across a gravel rally stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera focused on the rear wheel showing a deeply treaded tire spinning violently and kicking up rock and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene backlit by a low-hanging sun creating anamorphic lens flare with horizontal streaks of blue and orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky painted with deep purple, fiery orange, and soft pink hues characteristic of the Makoto Shinkai style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949762649834156 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.005949762649834156, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949762649834156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22193236649036407, + 0.04064496234059334, + 0.0535212941467762, + 0.07200779765844345, + 0.046768274158239365, + 0.11418621242046356, + 0.03984731063246727, + 0.12150397896766663, + 0.0993010401725769, + 0.190286785364151 + ], + "credit_map": [ + 0.03271126002073288, + 0.026951642706990242, + 0.05196993425488472, + 0.03476644307374954, + 0.020036274567246437, + 0.01507868617773056, + 0.021440738812088966, + 0.018977370113134384, + 0.005204063840210438, + 0.005263934377580881, + 0.007383184507489204, + 0.006123880855739117, + 0.004268564283847809, + 0.0031965509988367558, + 0.004619548097252846, + 0.004585239104926586, + 0.006535438820719719, + 0.006985770072788, + 0.006788832601159811, + 0.011777983047068119, + 0.0041943821124732494, + 0.0040897587314248085, + 0.0066723995842039585, + 0.006476726848632097, + 0.007018543314188719, + 0.00701187364757061, + 0.011018740013241768, + 0.0203875619918108, + 0.0060227541252970695, + 0.004331076517701149, + 0.00945567712187767, + 0.00676156859844923, + 0.0069560580886900425, + 0.006311336066573858, + 0.006220022216439247, + 0.006682771723717451, + 0.005839422810822725, + 0.0033103798050433397, + 0.006546201184391975, + 0.004902081564068794, + 0.016186481341719627, + 0.013123779557645321, + 0.013878229074180126, + 0.017377711832523346, + 0.015614938922226429, + 0.007441235240548849, + 0.01816841959953308, + 0.012395421974360943, + 0.006240027025341988, + 0.004325270187109709, + 0.0037648933939635754, + 0.004672482144087553, + 0.004146456718444824, + 0.004356104880571365, + 0.00592364277690649, + 0.006418433506041765, + 0.011761974543333054, + 0.01283078733831644, + 0.012949347496032715, + 0.016560925170779228, + 0.011469883844256401, + 0.013561503030359745, + 0.019467299804091454, + 0.02290225587785244, + 0.007479354739189148, + 0.009869146160781384, + 0.012643205001950264, + 0.01948590949177742, + 0.00915470253676176, + 0.013228228315711021, + 0.01740841194987297, + 0.010032076388597488, + 0.019257957115769386, + 0.030997805297374725, + 0.015657814219594002, + 0.017859481275081635, + 0.029450228437781334, + 0.03646273538470268, + 0.01813778653740883, + 0.022462980821728706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.005949762649834156 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_8/index_tensor([6551, 6551])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_8/index_tensor([6551, 6551])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0c635d672b81f3e4171e857031d07def8bcbd1dc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_8/index_tensor([6551, 6551])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a static, sweeping long shot, captured in strict adherence to Sebastián Salgado's signature high-contrast black-and-white style, a luminous ghostly express train glides silently across a vast, cracked, desolate landscape. The landscape is arid, its horizon broken by the silhouettes of jagged rocks and the twisted skeletons of a few dead trees, while above rises a stormy sky filled with thick, textured clouds. The train itself is a translucent, luminous object, its carriages composed not of solid steel but rather of shimmering outlines of pure white light. It exudes a soft, otherworldly inner glow, radiating against the inky blackness of the background. In the foreground, a lone, ethereal figure stands in silhouette, its form blurred, as if draped in a flowing robe that blends seamlessly with the shadows of the earth. Its arms are raised in a powerful, sweeping gesture, as if conducting a silent symphony, guiding the train majestically through the void. The entire scene is meticulously lit, with deep blacks swallowing up every detail in the shadows, while whites—particularly the train's glow and the highlights of the clouds—are dazzling, verging on overexposed. A fine grain of film adds a tactile quality to the image, creating an atmosphere of mythical grandeur, profound loneliness, and timeless epicness.", + "seed": 26442, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_8/index_6551_seed_26442_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0734198709949851, + "vlm_logprob": -0.0734198709949851, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.0734198709949851, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04773779958486557 + ], + "opposite_logprob_tokens": [ + -4.2977375984191895 + ], + "logprob_reward": -0.04773779958486557, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.04773779958486557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04070327430963516, + 0.021892569959163666, + 0.009637197479605675, + 0.0420752577483654, + 0.0076664206571877, + 0.7123888731002808, + 0.026267634704709053, + 0.006152268964797258, + 0.12306321412324905, + 0.01015330757945776 + ], + "credit_map": [ + 0.0030739428475499153, + 0.004854306578636169, + 0.0021070202346891165, + 0.0022497521713376045, + 0.011106159538030624, + 0.01261045504361391, + 0.00245430669747293, + 0.002247333060950041, + 0.0020411780569702387, + 0.005001397337764502, + 0.0025905196089297533, + 0.002399652963504195, + 0.0028519814368337393, + 0.0036325145047158003, + 0.001687766402028501, + 0.00168755988124758, + 0.0009819167898967862, + 0.0024516433477401733, + 0.0006904859328642488, + 0.0015514742117375135, + 0.0010229316540062428, + 0.0015205375384539366, + 0.0007834240677766502, + 0.0006347842863760889, + 0.0031420616433024406, + 0.0032638669945299625, + 0.001684758928604424, + 0.001465577632188797, + 0.011785882525146008, + 0.017218980938196182, + 0.002048063324764371, + 0.0014660664601251483, + 0.0005827431450597942, + 0.0011547449976205826, + 0.0010862619383260608, + 0.0016428224043920636, + 0.0006323534180410206, + 0.0007476728060282767, + 0.0010193652706220746, + 0.0008004576084204018, + 0.05563463270664215, + 0.055070824921131134, + 0.008765030652284622, + 0.009146289899945259, + 0.23577074706554413, + 0.31790968775749207, + 0.015718456357717514, + 0.01437325682491064, + 0.0018523248145356774, + 0.0026427004486322403, + 0.0012078207219019532, + 0.0013358156429603696, + 0.006568977143615484, + 0.010197500698268414, + 0.0015387550229206681, + 0.0009237410849891603, + 0.0004578209191095084, + 0.0009365308796986938, + 0.0007524481043219566, + 0.0011433932231739163, + 0.0004931304138153791, + 0.0006285264971666038, + 0.0010795569978654385, + 0.0006608624244108796, + 0.00944435503333807, + 0.008773180656135082, + 0.0025688507594168186, + 0.0023346119560301304, + 0.04060004651546478, + 0.0542307086288929, + 0.0029537901282310486, + 0.0021576788276433945, + 0.0007699266425333917, + 0.0009302470134571195, + 0.0010419662576168776, + 0.0012996048899367452, + 0.0019670105539262295, + 0.0019716264214366674, + 0.0013251093914732337, + 0.0008478168747387826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.04773779958486557 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2814631462097168 + ], + "opposite_logprob_tokens": [ + -1.4064631462097168 + ], + "logprob_reward": -0.2814631462097168, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2814631462097168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031022407114505768, + 0.00981680303812027, + 0.004692609887570143, + 0.014074033126235008, + 0.006665104068815708, + 0.9103589653968811, + 0.006000994239002466, + 0.0021379562094807625, + 0.014040907844901085, + 0.0011902370024472475 + ], + "credit_map": [ + 0.002413633046671748, + 0.0028161779046058655, + 0.0010422896593809128, + 0.0010824715718626976, + 0.009546160697937012, + 0.011853886768221855, + 0.0011129797203466296, + 0.0011548075126484036, + 0.0007698479457758367, + 0.0006825182936154306, + 0.00063730200054124, + 0.0002685045765247196, + 0.002992733847349882, + 0.0037530541885644197, + 0.00034714251523837447, + 0.0003656999033410102, + 0.00038633905933238566, + 0.00036854110658168793, + 0.00015552241529803723, + 0.00021785669378004968, + 0.0014618643326684833, + 0.0017260415479540825, + 0.00020637831767089665, + 0.0001700667926343158, + 0.0011337181786075234, + 0.001063115312717855, + 0.00021121626195963472, + 0.00023716081341262907, + 0.004530234262347221, + 0.006230426952242851, + 0.0003561433986760676, + 0.0003120180917903781, + 0.0005128144985064864, + 0.0007494783494621515, + 0.00023799181508366019, + 0.00042497491813264787, + 0.001752369455061853, + 0.0023750856053084135, + 0.0003063692129217088, + 0.000306019966956228, + 0.07102144509553909, + 0.07065902650356293, + 0.011295794509351254, + 0.011720993556082249, + 0.3014979660511017, + 0.40600839257240295, + 0.020030617713928223, + 0.018124712631106377, + 0.0004133456386625767, + 0.0005553495138883591, + 0.000271078257355839, + 0.00030445371521636844, + 0.0015686762053519487, + 0.0022138666827231646, + 0.0004418221942614764, + 0.00023240181326400489, + 0.00016937582404352725, + 0.00023371078714262694, + 0.0001374407729599625, + 0.0001426410599378869, + 0.0005411842139437795, + 0.0006388855981640518, + 0.00017344843945465982, + 0.00010126980487257242, + 0.0010677839163690805, + 0.000999901327304542, + 0.00031999297789298, + 0.0004280638531781733, + 0.004501409828662872, + 0.006102594546973705, + 0.00035542904515750706, + 0.0002657310978975147, + 9.909724030876532e-05, + 9.754334314493462e-05, + 9.207945549860597e-05, + 7.625158468727022e-05, + 0.00028864361229352653, + 0.000338189274771139, + 0.00011260387691436335, + 8.582861482864246e-05 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.2814631462097168 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190231174230576 + ], + "opposite_logprob_tokens": [ + -4.018190383911133 + ], + "logprob_reward": -0.018190231174230576, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018190231174230576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05039762333035469, + 0.028995642438530922, + 0.00652329484000802, + 0.12381066381931305, + 0.020305266603827477, + 0.7202479839324951, + 0.005271773785352707, + 0.008099645376205444, + 0.016713354736566544, + 0.019634777680039406 + ], + "credit_map": [ + 0.003430720651522279, + 0.005772861186414957, + 0.005011603701859713, + 0.00307544507086277, + 0.012037266045808792, + 0.014926429837942123, + 0.002772331004962325, + 0.0033709658309817314, + 0.002252383856102824, + 0.002876449842005968, + 0.005295455455780029, + 0.003088774159550667, + 0.00472657848149538, + 0.00516773946583271, + 0.002878191415220499, + 0.0027100678998976946, + 0.0006373987416736782, + 0.0010161575628444552, + 0.0006904535694047809, + 0.0008393964963033795, + 0.0010759523138403893, + 0.0009368017199449241, + 0.0006901115411892533, + 0.0006370224291458726, + 0.010032148100435734, + 0.009727942757308483, + 0.002160443691536784, + 0.0019080901984125376, + 0.03923770412802696, + 0.05518987774848938, + 0.003063160227611661, + 0.0024912902154028416, + 0.001489080023020506, + 0.0015120416646823287, + 0.0011217722203582525, + 0.0022089718841016293, + 0.005052927881479263, + 0.0065202172845602036, + 0.0014090078184381127, + 0.000991246197372675, + 0.05615273490548134, + 0.05601339414715767, + 0.009171162731945515, + 0.009862477891147137, + 0.23802387714385986, + 0.320186585187912, + 0.016481095924973488, + 0.014356679283082485, + 0.0004252672952134162, + 0.0007220346597023308, + 0.0005025529535487294, + 0.0007855769945308566, + 0.0005594315007328987, + 0.0007538651116192341, + 0.00103155174292624, + 0.0004914931487292051, + 0.0005222994950599968, + 0.0008943419088609517, + 0.00047248718328773975, + 0.0007409208337776363, + 0.0014882271643728018, + 0.0019219268579035997, + 0.0014322828501462936, + 0.0006271582678891718, + 0.001175744691863656, + 0.0011309192050248384, + 0.0008344516973011196, + 0.0009098325972445309, + 0.004707218613475561, + 0.006514458451420069, + 0.0009817120153456926, + 0.00045901519479230046, + 0.0016142544336616993, + 0.0014148845802992582, + 0.0016043406212702394, + 0.0037731737829744816, + 0.003725603222846985, + 0.004426250699907541, + 0.0015553273260593414, + 0.0015209406847134233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.018190231174230576 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072855442762375 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072855442762375, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072855442762375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03566915914416313, + 0.010414808988571167, + 0.00420187134295702, + 0.14727845788002014, + 0.05198412388563156, + 0.5774216651916504, + 0.004486400168389082, + 0.011835820972919464, + 0.08239880204200745, + 0.07430891692638397 + ], + "credit_map": [ + 0.002537247259169817, + 0.0027293118182569742, + 0.0028672190383076668, + 0.0020562673453241587, + 0.009659440256655216, + 0.011972715146839619, + 0.0019149151630699635, + 0.0019320410210639238, + 0.0007723550661467016, + 0.0015407191822305322, + 0.0016288264887407422, + 0.0014043040573596954, + 0.0010446333326399326, + 0.0016667123418301344, + 0.0013331343652680516, + 0.0010241245618090034, + 0.0003658052009996027, + 0.0007384216296486557, + 0.0004132577159907669, + 0.0003942462499253452, + 0.00046206871047616005, + 0.0007897607865743339, + 0.000579667859710753, + 0.0004586431314237416, + 0.011659289710223675, + 0.011085893958806992, + 0.003139397595077753, + 0.0033094235695898533, + 0.0465969480574131, + 0.06356199085712433, + 0.004744280129671097, + 0.0031812377274036407, + 0.0037515065632760525, + 0.0033909196499735117, + 0.004522474016994238, + 0.008471399545669556, + 0.010941024869680405, + 0.012794643640518188, + 0.005159270949661732, + 0.0029528853483498096, + 0.045033980160951614, + 0.0447990745306015, + 0.0071397386491298676, + 0.007489718962460756, + 0.19137568771839142, + 0.2573707699775696, + 0.012748773209750652, + 0.011463885195553303, + 0.0003791726194322109, + 0.0006360693369060755, + 0.0005352626321837306, + 0.0007359182927757502, + 0.00034168490674346685, + 0.00055298040388152, + 0.0008226967765949667, + 0.00048261514166370034, + 0.0007992219761945307, + 0.0010184671264141798, + 0.0005055373185314238, + 0.0009062215103767812, + 0.0027487720362842083, + 0.0035235092509537935, + 0.0015112751862034202, + 0.0008228161022998393, + 0.006141115445643663, + 0.005644586402922869, + 0.0019246594747528434, + 0.0029193286318331957, + 0.025642039254307747, + 0.03497422859072685, + 0.003379057627171278, + 0.0017737910384312272, + 0.005981201305985451, + 0.005497857928276062, + 0.0062589566223323345, + 0.014589136466383934, + 0.013726596720516682, + 0.01614023558795452, + 0.0067732129245996475, + 0.005341724026948214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.011072855442762375 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0086353225633502 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.0086353225633502, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0086353225633502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04609539732336998, + 0.051708076149225235, + 0.01757699064910412, + 0.11241462826728821, + 0.011049145832657814, + 0.7024828195571899, + 0.013292143121361732, + 0.0055657848715782166, + 0.028716973960399628, + 0.011098034679889679 + ], + "credit_map": [ + 0.004103062674403191, + 0.006090780254453421, + 0.00598517432808876, + 0.006761420983821154, + 0.005145464092493057, + 0.005522980820387602, + 0.005910808220505714, + 0.00657570781186223, + 0.003650238737463951, + 0.005357672926038504, + 0.0034617381170392036, + 0.003138190833851695, + 0.012228413484990597, + 0.018218956887722015, + 0.0029305219650268555, + 0.002722341800108552, + 0.0017186874756589532, + 0.0033206066582351923, + 0.0020621393341571093, + 0.0015954617410898209, + 0.002479510847479105, + 0.0025900350883603096, + 0.0020065729040652514, + 0.0018039774149656296, + 0.008905245922505856, + 0.009629656560719013, + 0.00320127559825778, + 0.0033922656439244747, + 0.030815526843070984, + 0.04699144884943962, + 0.00559590570628643, + 0.003883293131366372, + 0.0008706293301656842, + 0.0010806007776409388, + 0.0013477386673912406, + 0.0010872475104406476, + 0.0017602703301236033, + 0.002877747407183051, + 0.0013433123240247369, + 0.0006816000095568597, + 0.054842326790094376, + 0.05408470705151558, + 0.008673045784235, + 0.009604278020560741, + 0.23304584622383118, + 0.31202682852745056, + 0.01599481701850891, + 0.014210963621735573, + 0.0008927443996071815, + 0.0017059522215276957, + 0.0007715795654803514, + 0.002320643747225404, + 0.0017331716371700168, + 0.0030637243762612343, + 0.001825018785893917, + 0.000979307689704001, + 0.0004527881683316082, + 0.0005961793940514326, + 0.0005186475464142859, + 0.0006575473234988749, + 0.0009041450102813542, + 0.0011464685667306185, + 0.0009177789324894547, + 0.00037223007529973984, + 0.0022184927947819233, + 0.0026415877509862185, + 0.0013961934018880129, + 0.001457523787394166, + 0.006605655420571566, + 0.010308101773262024, + 0.0026899611111730337, + 0.0013994608307257295, + 0.0011315784649923444, + 0.0012735759373754263, + 0.0013948394916951656, + 0.002320939442142844, + 0.0009393783984705806, + 0.001437045051716268, + 0.0014417541678994894, + 0.001158923376351595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.0086353225633502 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a static, sweeping long shot, captured in strict adherence to Sebastián Salgado's signature high-contrast black-and-white style, a luminous ghostly express train glides silently across a vast, cracked, desolate landscape. The landscape is arid, its horizon broken by the silhouettes of jagged rocks and the twisted skeletons of a few dead trees, while above rises a stormy sky filled with thick, textured clouds. The train itself is a translucent, luminous object, its carriages composed not of solid steel but rather of shimmering outlines of pure white light. It exudes a soft, otherworldly inner glow, radiating against the inky blackness of the background. In the foreground, a lone, ethereal figure stands in silhouette, its form blurred, as if draped in a flowing robe that blends seamlessly with the shadows of the earth. Its arms are raised in a powerful, sweeping gesture, as if conducting a silent symphony, guiding the train majestically through the void. The entire scene is meticulously lit, with deep blacks swallowing up every detail in the shadows, while whites—particularly the train's glow and the highlights of the clouds—are dazzling, verging on overexposed. A fine grain of film adds a tactile quality to the image, creating an atmosphere of mythical grandeur, profound loneliness, and timeless epicness.", + "seed": 26443, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_8/index_6551_seed_26443_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.039389312639832495, + "vlm_logprob": -0.039389312639832495, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.039389312639832495, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06738659739494324 + ], + "opposite_logprob_tokens": [ + -3.5673866271972656 + ], + "logprob_reward": -0.06738659739494324, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.06738659739494324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18660952150821686, + 0.24559296667575836, + 0.04057960957288742, + 0.07338973134756088, + 0.05286366865038872, + 0.09474191814661026, + 0.1023184061050415, + 0.07285091280937195, + 0.07988396286964417, + 0.05116939917206764 + ], + "credit_map": [ + 0.020482592284679413, + 0.02451186627149582, + 0.028480594977736473, + 0.02585745044052601, + 0.015900863334536552, + 0.022071849554777145, + 0.026900969445705414, + 0.022403312847018242, + 0.02108713798224926, + 0.02655722387135029, + 0.04673667997121811, + 0.030880646780133247, + 0.03393984213471413, + 0.034900568425655365, + 0.024775687605142593, + 0.02671513520181179, + 0.003239644691348076, + 0.007318925578147173, + 0.0052193752489984035, + 0.00563333835452795, + 0.004296806175261736, + 0.005711287725716829, + 0.004266733303666115, + 0.004893492441624403, + 0.0057143853046, + 0.008801422081887722, + 0.014249319210648537, + 0.01319347694516182, + 0.005452949553728104, + 0.008824490010738373, + 0.007893301546573639, + 0.009260378777980804, + 0.006171432323753834, + 0.007702811621129513, + 0.00664650509133935, + 0.009140292182564735, + 0.004022692795842886, + 0.010490630753338337, + 0.0036675597075372934, + 0.00502173975110054, + 0.009266044944524765, + 0.012517634779214859, + 0.01875954121351242, + 0.013732386752963066, + 0.007869280874729156, + 0.015449956990778446, + 0.006825024262070656, + 0.010322032496333122, + 0.008285547606647015, + 0.011882508173584938, + 0.013411491177976131, + 0.02812379226088524, + 0.007613246329128742, + 0.012982592917978764, + 0.009040264412760735, + 0.010978952050209045, + 0.008253180421888828, + 0.012624548748135567, + 0.007795735727995634, + 0.013414821587502956, + 0.005642786622047424, + 0.012553776614367962, + 0.006113705690950155, + 0.006452350877225399, + 0.00803422648459673, + 0.0172243881970644, + 0.012922893278300762, + 0.0093923918902874, + 0.007279967423528433, + 0.01226329617202282, + 0.005251020658761263, + 0.007515770848840475, + 0.005566755775362253, + 0.005845567211508751, + 0.007354244589805603, + 0.013319365680217743, + 0.004654391668736935, + 0.005213597789406776, + 0.0045032077468931675, + 0.004712261725217104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.06738659739494324 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08925776183605194 + ], + "opposite_logprob_tokens": [ + -2.4642577171325684 + ], + "logprob_reward": -0.08925776183605194, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08925776183605194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4488394558429718, + 0.20564410090446472, + 0.068463034927845, + 0.0248425230383873, + 0.031618114560842514, + 0.06131211668252945, + 0.051867466419935226, + 0.05303659662604332, + 0.025982748717069626, + 0.028393816202878952 + ], + "credit_map": [ + 0.03739622235298157, + 0.06400594860315323, + 0.04590415209531784, + 0.05896032974123955, + 0.03712700307369232, + 0.0978349894285202, + 0.057729918509721756, + 0.049880899488925934, + 0.015886109322309494, + 0.022961774840950966, + 0.04495581239461899, + 0.02397163398563862, + 0.015993503853678703, + 0.03767332434654236, + 0.018194973468780518, + 0.02600698545575142, + 0.005884936545044184, + 0.01057466771453619, + 0.007716478314250708, + 0.013308570720255375, + 0.00522236991673708, + 0.010788424871861935, + 0.007056914269924164, + 0.007910673506557941, + 0.002185661578550935, + 0.004459952469915152, + 0.0032454091124236584, + 0.003490451956167817, + 0.0023682136088609695, + 0.004258346278220415, + 0.0024695012252777815, + 0.002364987740293145, + 0.0036164792254567146, + 0.004985242150723934, + 0.0038055076729506254, + 0.004783392418175936, + 0.00247238390147686, + 0.005994110368192196, + 0.003086867742240429, + 0.0028741315472871065, + 0.007100358605384827, + 0.009973498061299324, + 0.007659096736460924, + 0.009598568081855774, + 0.007270359434187412, + 0.009690460748970509, + 0.004404401406645775, + 0.005615380592644215, + 0.004878467880189419, + 0.008746699430048466, + 0.006680333986878395, + 0.006651440169662237, + 0.004563341848552227, + 0.008535386063158512, + 0.005453457124531269, + 0.006358342710882425, + 0.0060281348414719105, + 0.005908084101974964, + 0.006004215218126774, + 0.010653886944055557, + 0.004611250013113022, + 0.009494670666754246, + 0.0046374863013625145, + 0.005698871798813343, + 0.0028024795465171337, + 0.0037681101821362972, + 0.004241135902702808, + 0.004051751922816038, + 0.0022043678909540176, + 0.0029995993245393038, + 0.002869729185476899, + 0.003045575227588415, + 0.0026345292571932077, + 0.0076290094293653965, + 0.0028050148393958807, + 0.0032505972776561975, + 0.00245550274848938, + 0.005384373478591442, + 0.0019936824683099985, + 0.0022411092650145292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.08925776183605194 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085351929068565 + ], + "opposite_logprob_tokens": [ + -4.511085510253906 + ], + "logprob_reward": -0.011085351929068565, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085351929068565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2086135447025299, + 0.27591341733932495, + 0.1285765916109085, + 0.03466754034161568, + 0.08450494706630707, + 0.11097513139247894, + 0.01772867515683174, + 0.027954846620559692, + 0.022769667208194733, + 0.0882955864071846 + ], + "credit_map": [ + 0.025386637076735497, + 0.03212810680270195, + 0.03563178703188896, + 0.025168608874082565, + 0.014999152161180973, + 0.02158770151436329, + 0.02878778614103794, + 0.024923786520957947, + 0.03270064294338226, + 0.03335411846637726, + 0.043969810009002686, + 0.04571182653307915, + 0.018872149288654327, + 0.03447483852505684, + 0.032589398324489594, + 0.034240640699863434, + 0.011925988830626011, + 0.016978468745946884, + 0.014405671507120132, + 0.026114074513316154, + 0.008063430897891521, + 0.013628710061311722, + 0.02034664712846279, + 0.017113596200942993, + 0.003657203633338213, + 0.004699635319411755, + 0.004939123056828976, + 0.006391436792910099, + 0.002735161455348134, + 0.004067257046699524, + 0.003945758566260338, + 0.00423196516931057, + 0.006806033663451672, + 0.009949716739356518, + 0.007974031381309032, + 0.019474824890494347, + 0.008783577010035515, + 0.01083680335432291, + 0.009737638756632805, + 0.010942327789962292, + 0.008359979838132858, + 0.010816412046551704, + 0.018683068454265594, + 0.027324294671416283, + 0.007019443903118372, + 0.013663352467119694, + 0.010413550771772861, + 0.014695031568408012, + 0.0018864191370084882, + 0.0026117758825421333, + 0.002393352100625634, + 0.003487279871478677, + 0.001605640398338437, + 0.0021262671798467636, + 0.0016470742411911488, + 0.001970865996554494, + 0.002547591459006071, + 0.005103961564600468, + 0.0037638829089701176, + 0.005053828004747629, + 0.002302567707374692, + 0.0033579678274691105, + 0.0025698700919747353, + 0.0032551754266023636, + 0.0023562731221318245, + 0.003127890871837735, + 0.003755209967494011, + 0.0042984364554286, + 0.0019011362455785275, + 0.0026982983108609915, + 0.001727089169435203, + 0.0029053345788270235, + 0.008191130124032497, + 0.016355184838175774, + 0.014749645255506039, + 0.01481702085584402, + 0.007849411107599735, + 0.008550826460123062, + 0.00852159596979618, + 0.009260771796107292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.011085351929068565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057822048664093 + ], + "opposite_logprob_tokens": [ + -3.895578145980835 + ], + "logprob_reward": -0.02057822048664093, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057822048664093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1974848061800003, + 0.23658794164657593, + 0.06765355914831161, + 0.03571459650993347, + 0.04750659316778183, + 0.26438629627227783, + 0.056059952825307846, + 0.026311466470360756, + 0.01706888899207115, + 0.0512259416282177 + ], + "credit_map": [ + 0.020577607676386833, + 0.025150658562779427, + 0.03563889488577843, + 0.03208744525909424, + 0.015422442927956581, + 0.022016391158103943, + 0.02288683131337166, + 0.023704534396529198, + 0.021643027663230896, + 0.025314001366496086, + 0.03495258092880249, + 0.035584427416324615, + 0.03127375990152359, + 0.033759333193302155, + 0.028557393699884415, + 0.02550341747701168, + 0.006353722419589758, + 0.00830661691725254, + 0.011379916220903397, + 0.011461448855698109, + 0.007172290701419115, + 0.007970592938363552, + 0.007957606576383114, + 0.007051366381347179, + 0.0032138300593942404, + 0.0054693762212991714, + 0.005044507794082165, + 0.005537652876228094, + 0.0030412848573178053, + 0.003746767295524478, + 0.004550304729491472, + 0.0051108761690557, + 0.004637676756829023, + 0.007298648357391357, + 0.007206064183264971, + 0.007402304094284773, + 0.005235055461525917, + 0.006250760052353144, + 0.004139570519328117, + 0.00533651327714324, + 0.02087242342531681, + 0.021579304710030556, + 0.037137724459171295, + 0.06901615858078003, + 0.019121600314974785, + 0.03194671496748924, + 0.03454465791583061, + 0.03016771376132965, + 0.004914399702101946, + 0.005540455225855112, + 0.009192110039293766, + 0.014925520867109299, + 0.0039024464786052704, + 0.005943584255874157, + 0.005285315681248903, + 0.0063561201095581055, + 0.0029249489307403564, + 0.004503151867538691, + 0.0029927832074463367, + 0.004066640045493841, + 0.003418971551582217, + 0.003179489402100444, + 0.002680474426597357, + 0.0025450061075389385, + 0.0021656653843820095, + 0.0028896850999444723, + 0.002221702830865979, + 0.0026393416337668896, + 0.001806602580472827, + 0.002210696693509817, + 0.001204935135319829, + 0.001930259633809328, + 0.0054442123509943485, + 0.005461905151605606, + 0.008286072872579098, + 0.011560664512217045, + 0.003854956477880478, + 0.005418098531663418, + 0.0053818607702851295, + 0.0058181737549602985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.02057822048664093 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863863155245781 + ], + "opposite_logprob_tokens": [ + -4.758638858795166 + ], + "logprob_reward": -0.00863863155245781, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863863155245781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39589923620224, + 0.16380058228969574, + 0.048514336347579956, + 0.06323500722646713, + 0.03845878690481186, + 0.07189592719078064, + 0.04148208349943161, + 0.04725067690014839, + 0.08246390521526337, + 0.04699940234422684 + ], + "credit_map": [ + 0.035134486854076385, + 0.04638521745800972, + 0.0558764822781086, + 0.051227159798145294, + 0.048194341361522675, + 0.06889721006155014, + 0.048761963844299316, + 0.04142238199710846, + 0.016304276883602142, + 0.01945498213171959, + 0.029360797256231308, + 0.024990882724523544, + 0.011134003289043903, + 0.020719429478049278, + 0.022464139387011528, + 0.01937209628522396, + 0.004883039742708206, + 0.007576487492769957, + 0.005928296595811844, + 0.006576233077794313, + 0.004869788885116577, + 0.006611849647015333, + 0.006199848372489214, + 0.0058687967248260975, + 0.006151845678687096, + 0.006653104443103075, + 0.008759335614740849, + 0.009426082484424114, + 0.008689513429999352, + 0.008263609372079372, + 0.007733957841992378, + 0.007557563483715057, + 0.003603189019486308, + 0.004110243637114763, + 0.00472569465637207, + 0.007023015525192022, + 0.0034761419519782066, + 0.005093126557767391, + 0.0051332032307982445, + 0.005294174887239933, + 0.006096240133047104, + 0.007100049871951342, + 0.00810925755649805, + 0.020283859223127365, + 0.004481638316065073, + 0.009951572865247726, + 0.00660700211301446, + 0.009266316890716553, + 0.003874866059049964, + 0.004383218940347433, + 0.0059125786647200584, + 0.010082845576107502, + 0.0030572577379643917, + 0.004362477455288172, + 0.004331856034696102, + 0.005476984195411205, + 0.005429199431091547, + 0.006166402716189623, + 0.007039136253297329, + 0.008395159617066383, + 0.003559583332389593, + 0.005509004928171635, + 0.00503931101411581, + 0.006112882401794195, + 0.008663207292556763, + 0.009028850123286247, + 0.014744410291314125, + 0.011384121142327785, + 0.007338549476116896, + 0.008745143190026283, + 0.010243549942970276, + 0.012316080741584301, + 0.004652443341910839, + 0.007196743041276932, + 0.008225698955357075, + 0.008716699667274952, + 0.0032599465921521187, + 0.005165530368685722, + 0.004024301189929247, + 0.005758040118962526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.00863863155245781 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a static, sweeping long shot, captured in strict adherence to Sebastián Salgado's signature high-contrast black-and-white style, a luminous ghostly express train glides silently across a vast, cracked, desolate landscape. The landscape is arid, its horizon broken by the silhouettes of jagged rocks and the twisted skeletons of a few dead trees, while above rises a stormy sky filled with thick, textured clouds. The train itself is a translucent, luminous object, its carriages composed not of solid steel but rather of shimmering outlines of pure white light. It exudes a soft, otherworldly inner glow, radiating against the inky blackness of the background. In the foreground, a lone, ethereal figure stands in silhouette, its form blurred, as if draped in a flowing robe that blends seamlessly with the shadows of the earth. Its arms are raised in a powerful, sweeping gesture, as if conducting a silent symphony, guiding the train majestically through the void. The entire scene is meticulously lit, with deep blacks swallowing up every detail in the shadows, while whites—particularly the train's glow and the highlights of the clouds—are dazzling, verging on overexposed. A fine grain of film adds a tactile quality to the image, creating an atmosphere of mythical grandeur, profound loneliness, and timeless epicness.", + "seed": 26444, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_8/index_6551_seed_26444_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04064919739030302, + "vlm_logprob": -0.04064919739030302, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.04064919739030302, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05169392377138138 + ], + "opposite_logprob_tokens": [ + -4.051693916320801 + ], + "logprob_reward": -0.05169392377138138, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.05169392377138138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18239514529705048, + 0.04357147589325905, + 0.03731771931052208, + 0.49343883991241455, + 0.029482804238796234, + 0.039784811437129974, + 0.035664189606904984, + 0.07570859789848328, + 0.03877240791916847, + 0.023863954469561577 + ], + "credit_map": [ + 0.015109975822269917, + 0.01884462870657444, + 0.016964197158813477, + 0.02088724635541439, + 0.03498274087905884, + 0.02662181667983532, + 0.02346990630030632, + 0.02551463432610035, + 0.003638464491814375, + 0.00436701625585556, + 0.003354068147018552, + 0.005988505203276873, + 0.005928660277277231, + 0.006456007249653339, + 0.006031769793480635, + 0.00780698424205184, + 0.0035183816216886044, + 0.0050086635164916515, + 0.0029096947982907295, + 0.002923243911936879, + 0.00787593238055706, + 0.004341204185038805, + 0.0038052950985729694, + 0.00693530822172761, + 0.03889259323477745, + 0.07461060583591461, + 0.031828053295612335, + 0.03093605302274227, + 0.11032774299383163, + 0.06557583808898926, + 0.06363127380609512, + 0.07763671875, + 0.002667077351361513, + 0.003949146252125502, + 0.0016402977053076029, + 0.0026039674412459135, + 0.00734897842630744, + 0.0025616639759391546, + 0.005644139368087053, + 0.0030675383750349283, + 0.004462142009288073, + 0.004880579654127359, + 0.00568507332354784, + 0.003096094587817788, + 0.006761518307030201, + 0.0030742741655558348, + 0.005790685303509235, + 0.006034445483237505, + 0.0032926974818110466, + 0.004527375567704439, + 0.0030123195610940456, + 0.004027465358376503, + 0.007953894324600697, + 0.0026316617149859667, + 0.006107208784669638, + 0.004111568909138441, + 0.006975471507757902, + 0.012709559872746468, + 0.005379979964345694, + 0.0058482326567173, + 0.02303425781428814, + 0.006547174416482449, + 0.01095676515251398, + 0.004257163032889366, + 0.002997847506776452, + 0.005312893074005842, + 0.001976467901840806, + 0.0036416405346244574, + 0.009526030160486698, + 0.003647947683930397, + 0.0063510434702038765, + 0.005318540148437023, + 0.0027989093214273453, + 0.0029002903029322624, + 0.002064502565190196, + 0.0029586616437882185, + 0.005210323724895716, + 0.0025409129448235035, + 0.0033264607191085815, + 0.002063894644379616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.05169392377138138 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10041911900043488 + ], + "opposite_logprob_tokens": [ + -2.350419044494629 + ], + "logprob_reward": -0.10041911900043488, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10041911900043488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3502899706363678, + 0.10157159715890884, + 0.04893479868769646, + 0.16400554776191711, + 0.06357930600643158, + 0.030125625431537628, + 0.04606472700834274, + 0.13361746072769165, + 0.013049958273768425, + 0.04876098036766052 + ], + "credit_map": [ + 0.03260696679353714, + 0.04463844373822212, + 0.025496358051896095, + 0.028409061953425407, + 0.0757429227232933, + 0.05592105910181999, + 0.03469441458582878, + 0.05278073623776436, + 0.008585023693740368, + 0.014395231381058693, + 0.005099656991660595, + 0.005843351595103741, + 0.022437267005443573, + 0.016532674431800842, + 0.01155065093189478, + 0.01712774485349655, + 0.0037470629904419184, + 0.007183785084635019, + 0.0026402268558740616, + 0.003147537587210536, + 0.01050170324742794, + 0.0065084355883300304, + 0.005203248001635075, + 0.0100028021261096, + 0.013573468662798405, + 0.024812763556838036, + 0.009781664237380028, + 0.008976589888334274, + 0.03487421199679375, + 0.023804835975170135, + 0.02029510773718357, + 0.027886928990483284, + 0.00572067778557539, + 0.008435330353677273, + 0.003711964702233672, + 0.004933200776576996, + 0.017443619668483734, + 0.005298310425132513, + 0.010377214290201664, + 0.007658992428332567, + 0.0026831456925719976, + 0.004049689508974552, + 0.003133209189400077, + 0.0028233746998012066, + 0.0074050575494766235, + 0.0028381755109876394, + 0.003220377489924431, + 0.003972597420215607, + 0.004419267177581787, + 0.006523576099425554, + 0.0034146320540457964, + 0.004580266773700714, + 0.009233980439603329, + 0.0034123011864721775, + 0.009624217636883259, + 0.004856490064412355, + 0.01183549128472805, + 0.022899657487869263, + 0.009554371237754822, + 0.009576927870512009, + 0.038930363953113556, + 0.012608347460627556, + 0.02023487538099289, + 0.007977435365319252, + 0.0014218456344678998, + 0.001580069656483829, + 0.0017988737672567368, + 0.0009722636896185577, + 0.00298413448035717, + 0.001046668621711433, + 0.0017695296555757523, + 0.0014765738742426038, + 0.004819782450795174, + 0.00721377320587635, + 0.003130888333544135, + 0.0028060004115104675, + 0.013611075468361378, + 0.0038835301529616117, + 0.009372679516673088, + 0.003923255484551191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.10041911900043488 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306819662451744 + ], + "opposite_logprob_tokens": [ + -3.1680681705474854 + ], + "logprob_reward": -0.04306819662451744, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306819662451744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18246658146381378, + 0.04941381886601448, + 0.046615809202194214, + 0.3394136428833008, + 0.13497011363506317, + 0.0505937822163105, + 0.025886395946145058, + 0.025470368564128876, + 0.015209086239337921, + 0.12996043264865875 + ], + "credit_map": [ + 0.016201866790652275, + 0.021191932260990143, + 0.0151191595941782, + 0.021205544471740723, + 0.031911011785268784, + 0.023052340373396873, + 0.02546614594757557, + 0.028318574652075768, + 0.004749075975269079, + 0.005541206803172827, + 0.00406541908159852, + 0.005064093042165041, + 0.008356849662959576, + 0.008218995295464993, + 0.005741880740970373, + 0.007676297333091497, + 0.004444333259016275, + 0.005759947467595339, + 0.0036752810701727867, + 0.004042734857648611, + 0.009772509336471558, + 0.00672101741656661, + 0.006501224357634783, + 0.005698761437088251, + 0.029354505240917206, + 0.040669407695531845, + 0.022318851202726364, + 0.022761981934309006, + 0.056124478578567505, + 0.05338083952665329, + 0.045338500291109085, + 0.06946510821580887, + 0.013355152681469917, + 0.016904611140489578, + 0.011189024895429611, + 0.011731008067727089, + 0.02787289023399353, + 0.011796563863754272, + 0.024400416761636734, + 0.01772044040262699, + 0.004411396570503712, + 0.007049435283988714, + 0.004390548449009657, + 0.0037574165035039186, + 0.011214231140911579, + 0.003668408142402768, + 0.007770553696900606, + 0.00833179336041212, + 0.002363508800044656, + 0.0037779519334435463, + 0.0022542434744536877, + 0.0024669922422617674, + 0.007015281356871128, + 0.0019669029861688614, + 0.003077500034123659, + 0.002964014420285821, + 0.0024333964101970196, + 0.0030652224086225033, + 0.0024920962750911713, + 0.0030042030848562717, + 0.004471636842936277, + 0.0030914675444364548, + 0.004086386412382126, + 0.0028259591199457645, + 0.0015514458063989878, + 0.001986150164157152, + 0.0012596360174939036, + 0.001727019320242107, + 0.003543566446751356, + 0.0014789107954129577, + 0.002527868375182152, + 0.001134488731622696, + 0.013371485285460949, + 0.01713578775525093, + 0.008716694079339504, + 0.009737732820212841, + 0.03104267083108425, + 0.010723976418375969, + 0.024869907647371292, + 0.014362188056111336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.04306819662451744 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818423556163907 + ], + "opposite_logprob_tokens": [ + -5.877818584442139 + ], + "logprob_reward": -0.002818423556163907, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818423556163907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2560327351093292, + 0.11823359876871109, + 0.06055360287427902, + 0.2502911388874054, + 0.04532182589173317, + 0.03908461332321167, + 0.03444996476173401, + 0.0889139249920845, + 0.05103336274623871, + 0.0560852587223053 + ], + "credit_map": [ + 0.032450463622808456, + 0.02948084846138954, + 0.017741164192557335, + 0.037366777658462524, + 0.04730135202407837, + 0.037160348147153854, + 0.02885921485722065, + 0.025672562420368195, + 0.011540413834154606, + 0.012371336109936237, + 0.008076804690063, + 0.019736165180802345, + 0.016400760039687157, + 0.01219594944268465, + 0.017435995861887932, + 0.020476166158914566, + 0.005369624122977257, + 0.006386619992554188, + 0.00570644298568368, + 0.00984656810760498, + 0.01018398255109787, + 0.006291977129876614, + 0.008665651082992554, + 0.008102737367153168, + 0.01966496743261814, + 0.04000980034470558, + 0.016030577942728996, + 0.012441249564290047, + 0.05945500358939171, + 0.03546402230858803, + 0.029582208022475243, + 0.037643302232027054, + 0.004636770114302635, + 0.006420063320547342, + 0.0037314172368496656, + 0.008948925882577896, + 0.00697886710986495, + 0.003767647547647357, + 0.006349713541567326, + 0.004488424398005009, + 0.0036795237101614475, + 0.005605475977063179, + 0.0033565876074135303, + 0.006008885335177183, + 0.009402789175510406, + 0.002533822087571025, + 0.003597050905227661, + 0.004900478757917881, + 0.0034866873174905777, + 0.004399796482175589, + 0.0033945124596357346, + 0.00484535563737154, + 0.0075770290568470955, + 0.0025070128031075, + 0.003452194854617119, + 0.004787375684827566, + 0.007013313937932253, + 0.01599789597094059, + 0.0065576983615756035, + 0.007091853767633438, + 0.02619812823832035, + 0.008593249134719372, + 0.012031154707074165, + 0.005430631805211306, + 0.005246395710855722, + 0.006962575018405914, + 0.0034849464427679777, + 0.005552472546696663, + 0.012844936922192574, + 0.0038668455090373755, + 0.009041136130690575, + 0.00403405399993062, + 0.005235078278928995, + 0.007332865614444017, + 0.005064737051725388, + 0.006790552753955126, + 0.011349356733262539, + 0.0055654896423220634, + 0.0055226595140993595, + 0.009224521927535534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.002818423556163907 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005246323999017477 + ], + "opposite_logprob_tokens": [ + -5.255246162414551 + ], + "logprob_reward": -0.005246323999017477, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005246323999017477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2828960120677948, + 0.0963287279009819, + 0.042008064687252045, + 0.12609869241714478, + 0.1099558025598526, + 0.11213668435811996, + 0.026801666244864464, + 0.02825395204126835, + 0.04461418837308884, + 0.13090616464614868 + ], + "credit_map": [ + 0.02430141158401966, + 0.03264658525586128, + 0.02206321619451046, + 0.030987044796347618, + 0.057968951761722565, + 0.029477085918188095, + 0.035766761749982834, + 0.04968496412038803, + 0.007746505551040173, + 0.01125890202820301, + 0.00729616591706872, + 0.008879355154931545, + 0.020281996577978134, + 0.013629632070660591, + 0.012124158442020416, + 0.015112015418708324, + 0.004313922952860594, + 0.005253992974758148, + 0.003368397941812873, + 0.0034257525112479925, + 0.009646943770349026, + 0.005759076215326786, + 0.004730755928903818, + 0.0055092256516218185, + 0.009841936640441418, + 0.013687783852219582, + 0.008360980078577995, + 0.011221885681152344, + 0.02293843775987625, + 0.017272457480430603, + 0.01889488659799099, + 0.023880325257778168, + 0.010653741657733917, + 0.014888232573866844, + 0.006996186450123787, + 0.009745655581355095, + 0.021649548783898354, + 0.008940848521888256, + 0.021947195753455162, + 0.015134398825466633, + 0.010247256606817245, + 0.016299406066536903, + 0.009315651841461658, + 0.007929367013275623, + 0.025324327871203423, + 0.009751816280186176, + 0.015144839882850647, + 0.018124019727110863, + 0.0022155416663736105, + 0.004244403447955847, + 0.0020741410553455353, + 0.0027199641335755587, + 0.006924545392394066, + 0.0021518091671168804, + 0.0035587146412581205, + 0.0029125474393367767, + 0.003119734115898609, + 0.003548981621861458, + 0.0027443545404821634, + 0.0025817053392529488, + 0.0060007208958268166, + 0.0031644541304558516, + 0.004767333157360554, + 0.002326669404283166, + 0.004313173703849316, + 0.006454674992710352, + 0.0029005249962210655, + 0.003760636318475008, + 0.012214944697916508, + 0.0037180546205490828, + 0.007433409336954355, + 0.0038187685422599316, + 0.012621970847249031, + 0.01668039709329605, + 0.009454605169594288, + 0.01094649638980627, + 0.027163038030266762, + 0.012594147585332394, + 0.019121715798974037, + 0.022323790937662125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.005246323999017477 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a static, sweeping long shot, captured in strict adherence to Sebastián Salgado's signature high-contrast black-and-white style, a luminous ghostly express train glides silently across a vast, cracked, desolate landscape. The landscape is arid, its horizon broken by the silhouettes of jagged rocks and the twisted skeletons of a few dead trees, while above rises a stormy sky filled with thick, textured clouds. The train itself is a translucent, luminous object, its carriages composed not of solid steel but rather of shimmering outlines of pure white light. It exudes a soft, otherworldly inner glow, radiating against the inky blackness of the background. In the foreground, a lone, ethereal figure stands in silhouette, its form blurred, as if draped in a flowing robe that blends seamlessly with the shadows of the earth. Its arms are raised in a powerful, sweeping gesture, as if conducting a silent symphony, guiding the train majestically through the void. The entire scene is meticulously lit, with deep blacks swallowing up every detail in the shadows, while whites—particularly the train's glow and the highlights of the clouds—are dazzling, verging on overexposed. A fine grain of film adds a tactile quality to the image, creating an atmosphere of mythical grandeur, profound loneliness, and timeless epicness.", + "seed": 26445, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_8/index_6551_seed_26445_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.016904898453503847, + "vlm_logprob": -0.016904898453503847, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.016904898453503847, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.047747571021318436 + ], + "opposite_logprob_tokens": [ + -4.297747611999512 + ], + "logprob_reward": -0.047747571021318436, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.047747571021318436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17480024695396423, + 0.08227385580539703, + 0.04974282160401344, + 0.04232625290751457, + 0.04223247990012169, + 0.05099659413099289, + 0.25913041830062866, + 0.19336964190006256, + 0.04917987808585167, + 0.05594784766435623 + ], + "credit_map": [ + 0.023814653977751732, + 0.019681300967931747, + 0.01800781488418579, + 0.023496219888329506, + 0.024948235601186752, + 0.02146747149527073, + 0.02912868931889534, + 0.014255845919251442, + 0.01081190537661314, + 0.009763444773852825, + 0.008745859377086163, + 0.007796036545187235, + 0.014328823424875736, + 0.009818283841013908, + 0.015075486153364182, + 0.005934007000178099, + 0.006567496340721846, + 0.007157064974308014, + 0.003236384829506278, + 0.0053246463648974895, + 0.00991765595972538, + 0.006984963547438383, + 0.0069028944708406925, + 0.003651715349406004, + 0.007113196421414614, + 0.007939430885016918, + 0.004332478623837233, + 0.0034106208477169275, + 0.007140668109059334, + 0.005315923132002354, + 0.003947159741073847, + 0.0031267725862562656, + 0.004736610688269138, + 0.006553743965923786, + 0.006597182713449001, + 0.004533182829618454, + 0.007692635990679264, + 0.005078005604445934, + 0.004007979761809111, + 0.003033131593838334, + 0.0039855544455349445, + 0.011207077652215958, + 0.004086468368768692, + 0.004712335765361786, + 0.013395537622272968, + 0.006211619358509779, + 0.004486622288823128, + 0.0029113751370459795, + 0.07104769349098206, + 0.0384342223405838, + 0.008409601636230946, + 0.00567871006205678, + 0.0903690829873085, + 0.022286945953965187, + 0.007696814835071564, + 0.015207323245704174, + 0.038853395730257034, + 0.03542240336537361, + 0.007836089469492435, + 0.005973856896162033, + 0.06108998507261276, + 0.022172406315803528, + 0.011912494897842407, + 0.010108986869454384, + 0.008419099263846874, + 0.00610751286149025, + 0.0038674487732350826, + 0.0052075376734137535, + 0.007189278956502676, + 0.008687221445143223, + 0.005766026675701141, + 0.003935749642550945, + 0.0064621614292263985, + 0.008771302178502083, + 0.005405244883149862, + 0.006449417676776648, + 0.008543601259589195, + 0.0077190096490085125, + 0.006762676872313023, + 0.0058344267308712006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.047747571021318436 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014253710396587849 + ], + "opposite_logprob_tokens": [ + -4.264253616333008 + ], + "logprob_reward": -0.014253710396587849, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014253710396587849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2690891623497009, + 0.07238984107971191, + 0.046073999255895615, + 0.040383387356996536, + 0.05607941746711731, + 0.10745564103126526, + 0.11246298998594284, + 0.08508661389350891, + 0.12259911000728607, + 0.08837976306676865 + ], + "credit_map": [ + 0.0372663214802742, + 0.03162841871380806, + 0.03301459923386574, + 0.03736953064799309, + 0.03816511854529381, + 0.04114914685487747, + 0.02577066794037819, + 0.024725375697016716, + 0.011299249716103077, + 0.008855663239955902, + 0.008289456367492676, + 0.007501089479774237, + 0.012911876663565636, + 0.010873195715248585, + 0.006414337083697319, + 0.006244981661438942, + 0.006227083969861269, + 0.006466344930231571, + 0.0036756854970008135, + 0.005727440118789673, + 0.0084577901288867, + 0.007130880374461412, + 0.004648262169212103, + 0.0037405129987746477, + 0.003975823055952787, + 0.005841027479618788, + 0.004052833188325167, + 0.004800707101821899, + 0.007184797897934914, + 0.0046449038200080395, + 0.006145448889583349, + 0.00373784895054996, + 0.00931963138282299, + 0.01147104799747467, + 0.006186121609061956, + 0.004516927059739828, + 0.009544918313622475, + 0.006240355782210827, + 0.005043906159698963, + 0.0037565124221146107, + 0.00830504298210144, + 0.03446768596768379, + 0.009645495563745499, + 0.006716922856867313, + 0.0323740690946579, + 0.007872740738093853, + 0.004273544531315565, + 0.0038001490756869316, + 0.020473182201385498, + 0.023238275200128555, + 0.006200387142598629, + 0.0048428913578391075, + 0.03789887949824333, + 0.00958840362727642, + 0.004481005482375622, + 0.005739974789321423, + 0.014251451939344406, + 0.012546942569315434, + 0.007325564976781607, + 0.006644382607191801, + 0.01825593411922455, + 0.013064184226095676, + 0.005824499297887087, + 0.007173661142587662, + 0.027203887701034546, + 0.018916495144367218, + 0.005214904434978962, + 0.008259646594524384, + 0.03687993437051773, + 0.010444305837154388, + 0.008792475797235966, + 0.006887468509376049, + 0.011519914492964745, + 0.0128248929977417, + 0.007137587293982506, + 0.008912605233490467, + 0.014667278155684471, + 0.011298326775431633, + 0.015540764667093754, + 0.006478396710008383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.014253710396587849 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792149998247623 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.009792149998247623, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792149998247623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19762568175792694, + 0.0406087227165699, + 0.03088018298149109, + 0.0348750501871109, + 0.039203375577926636, + 0.04551934078335762, + 0.07992591708898544, + 0.1678970903158188, + 0.089633047580719, + 0.27383163571357727 + ], + "credit_map": [ + 0.01935465633869171, + 0.022465594112873077, + 0.01644379459321499, + 0.02770315855741501, + 0.03230919688940048, + 0.024038314819335938, + 0.03884335979819298, + 0.01646757312119007, + 0.006618306506425142, + 0.004080827813595533, + 0.003926250617951155, + 0.005844469182193279, + 0.0051640598103404045, + 0.005752605851739645, + 0.00595046067610383, + 0.0032717380672693253, + 0.0040187593549489975, + 0.0034535902086645365, + 0.003984329756349325, + 0.003539340104907751, + 0.004033852368593216, + 0.0053641595877707005, + 0.003946518991142511, + 0.0025396288838237524, + 0.0040640560910105705, + 0.005939042195677757, + 0.0045470898039639, + 0.0032037419732660055, + 0.005815516225993633, + 0.0048922766000032425, + 0.0033605273347347975, + 0.0030527953058481216, + 0.005447983741760254, + 0.00779312290251255, + 0.005445932969450951, + 0.0041685584001243114, + 0.005943313706666231, + 0.003971766214817762, + 0.003638253780081868, + 0.0027944419998675585, + 0.007476082071661949, + 0.008374515920877457, + 0.005609855521470308, + 0.004816792905330658, + 0.006302721332758665, + 0.005713562946766615, + 0.0038581781554967165, + 0.003367629135027528, + 0.006389293819665909, + 0.022848917171359062, + 0.004341553896665573, + 0.006223846692591906, + 0.02196238376200199, + 0.010275471955537796, + 0.003907125908881426, + 0.003977321553975344, + 0.035072918981313705, + 0.03139679878950119, + 0.006195965688675642, + 0.005319086369127035, + 0.05677131190896034, + 0.01747881807386875, + 0.006723464000970125, + 0.008938724175095558, + 0.014515053480863571, + 0.01815042831003666, + 0.00410503800958395, + 0.006220680195838213, + 0.027109121903777122, + 0.007456755265593529, + 0.006253949832171202, + 0.0058220187202095985, + 0.048009902238845825, + 0.07654941082000732, + 0.02022317796945572, + 0.0285129863768816, + 0.037376344203948975, + 0.03722824901342392, + 0.013173040002584457, + 0.012758477590978146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.009792149998247623 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095738288015127 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095738288015127, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095738288015127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19225002825260162, + 0.09641159325838089, + 0.08547034114599228, + 0.03659965470433235, + 0.0745609849691391, + 0.06825365126132965, + 0.12477101385593414, + 0.07463086396455765, + 0.07860513031482697, + 0.16844673454761505 + ], + "credit_map": [ + 0.023123396560549736, + 0.021644629538059235, + 0.016648216173052788, + 0.03166704624891281, + 0.027232864871621132, + 0.020100602880120277, + 0.035111162811517715, + 0.016722118481993675, + 0.01129041239619255, + 0.011095438152551651, + 0.007087366189807653, + 0.015074688009917736, + 0.014764501713216305, + 0.011977313086390495, + 0.016252456232905388, + 0.0088694142177701, + 0.010398425161838531, + 0.011186151765286922, + 0.005239950492978096, + 0.012351213954389095, + 0.014985520392656326, + 0.010156267322599888, + 0.013999895192682743, + 0.007152915000915527, + 0.004202991724014282, + 0.004286980256438255, + 0.00459735281765461, + 0.003953329753130674, + 0.005977860651910305, + 0.004087597131729126, + 0.006473131477832794, + 0.0030204097274690866, + 0.007798735052347183, + 0.008309618569910526, + 0.009149248711764812, + 0.007721946574747562, + 0.01078889798372984, + 0.00786839984357357, + 0.01681303232908249, + 0.00611109659075737, + 0.007595103234052658, + 0.014700105413794518, + 0.006555058062076569, + 0.0063257054425776005, + 0.014135555364191532, + 0.007297181989997625, + 0.007699702400714159, + 0.003945241216570139, + 0.012753298506140709, + 0.021675223484635353, + 0.0063302102498710155, + 0.0086273904889822, + 0.026490552350878716, + 0.03194834291934967, + 0.009882040321826935, + 0.007063958328217268, + 0.011419032700359821, + 0.015234054997563362, + 0.0033320023212581873, + 0.005643362179398537, + 0.021257849410176277, + 0.008221831172704697, + 0.005339932162314653, + 0.004182799253612757, + 0.016523219645023346, + 0.011319332756102085, + 0.0067450315691530704, + 0.004941940773278475, + 0.02022668905556202, + 0.007399434689432383, + 0.006723013706505299, + 0.004726464860141277, + 0.02883412130177021, + 0.0414818711578846, + 0.01612018793821335, + 0.016583913937211037, + 0.019054848700761795, + 0.018349330872297287, + 0.018060626462101936, + 0.00996183417737484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.004095738288015127 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a static, sweeping long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent, luminous train with carriages made of shimmering white light visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, blurred, ethereal figure standing in silhouette in the foreground with raised arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a vast, cracked, desolate landscape with jagged rocks, dead trees, and a stormy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0086353225633502 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.0086353225633502, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0086353225633502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17299515008926392, + 0.043002985417842865, + 0.04991389811038971, + 0.04443008452653885, + 0.0641939640045166, + 0.08008561283349991, + 0.08667762577533722, + 0.031517188996076584, + 0.2705013155937195, + 0.15668214857578278 + ], + "credit_map": [ + 0.02401452511548996, + 0.016728447750210762, + 0.022016873583197594, + 0.02108769491314888, + 0.02948065847158432, + 0.02599014714360237, + 0.02015708200633526, + 0.013519728556275368, + 0.006509553175419569, + 0.004721779376268387, + 0.004132930655032396, + 0.005573732312768698, + 0.0068910568952560425, + 0.006692312657833099, + 0.005260058678686619, + 0.0032215637620538473, + 0.006982437334954739, + 0.005489677656441927, + 0.005015755072236061, + 0.0060690296813845634, + 0.0067166611552238464, + 0.008220994845032692, + 0.007066092919558287, + 0.004353254567831755, + 0.005773773416876793, + 0.004868155810981989, + 0.004658379126340151, + 0.0048937746323645115, + 0.006181648932397366, + 0.007925932295620441, + 0.006364846136420965, + 0.0037635776679962873, + 0.008110139518976212, + 0.009753524325788021, + 0.003711990313604474, + 0.005161507520824671, + 0.011047594249248505, + 0.009758047759532928, + 0.012226630933582783, + 0.004424531478434801, + 0.009645583108067513, + 0.024729328230023384, + 0.005089825019240379, + 0.00477673439309001, + 0.020271163433790207, + 0.008004152216017246, + 0.00439205439761281, + 0.003176783211529255, + 0.010433112271130085, + 0.018771620467305183, + 0.004439231473952532, + 0.005256597883999348, + 0.0244058296084404, + 0.01360554713755846, + 0.004754612687975168, + 0.005011086817830801, + 0.004490375053137541, + 0.006620620843023062, + 0.0021595731377601624, + 0.002321371342986822, + 0.005379417911171913, + 0.004973936825990677, + 0.0031881341710686684, + 0.0023837594781070948, + 0.057821497321128845, + 0.05212390795350075, + 0.008927782997488976, + 0.013676416128873825, + 0.08972732722759247, + 0.024853143841028214, + 0.009984289295971394, + 0.013386956416070461, + 0.024404125288128853, + 0.0383477658033371, + 0.01158322487026453, + 0.01654553972184658, + 0.022526977583765984, + 0.020881276577711105, + 0.011864540167152882, + 0.01052870787680149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.0086353225633502 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_9/index_tensor([7905, 7905])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_9/index_tensor([7905, 7905])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e85c74d4db23ba55a533cd7bca99baa425f5dcec --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_9/index_tensor([7905, 7905])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, wide-angle shot captures a dramatic moment in a dense bamboo forest at night, rendered with the powerful, stylized aesthetic of a modern comic book panel. A sudden, dazzling, jagged prong of electric white lightning violently splits the inky sky, its tendrils branching out in a tangled web. This momentary flash bathes the entire scene in dramatic, high-contrast light, captured with the razor-sharp precision of 4K footage. Each element is defined by bold, decisive ink lines and deep, graphic shadows, completely eliminating mid-tones. The towering bamboo stems become a rhythmic pattern, stark white vertical lines set against impenetrable black voids, their leaves catching the flash like sharp silver daggers. Below, a winding stone path winds through the forest. The jagged stones glow damply, contrasting the scene's only curves with the rigid lines of the bamboo, while the shadows between them are pools of pure, featureless black. The entire image crystallizes in this single, powerful moment, freezing the raw power of nature in a moment of stunning, intense, and hyper-stylized visual impact.", + "seed": 22069, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_9/index_7905_seed_22069_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02481557340361178, + "vlm_logprob": -0.02481557340361178, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.02481557340361178, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01342424750328064 + ], + "opposite_logprob_tokens": [ + -4.638424396514893 + ], + "logprob_reward": -0.01342424750328064, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01342424750328064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16662722826004028, + 0.22757548093795776, + 0.14209817349910736, + 0.03858479857444763, + 0.03214464336633682, + 0.1634267419576645, + 0.04000174626708031, + 0.05000948905944824, + 0.049534521996974945, + 0.08999716490507126 + ], + "credit_map": [ + 0.04109950363636017, + 0.013679885305464268, + 0.018152017146348953, + 0.008842727169394493, + 0.01415653619915247, + 0.04066254943609238, + 0.01026048231869936, + 0.019773514941334724, + 0.024963049218058586, + 0.02138010784983635, + 0.03477226197719574, + 0.02327723428606987, + 0.029758363962173462, + 0.030661433935165405, + 0.02130177617073059, + 0.04146125167608261, + 0.015675274655222893, + 0.013887825421988964, + 0.019489284604787827, + 0.01158104557543993, + 0.029293039813637733, + 0.01790877990424633, + 0.019569195806980133, + 0.014693738892674446, + 0.004025950096547604, + 0.0052192118018865585, + 0.00682535395026207, + 0.0045040263794362545, + 0.0028972821310162544, + 0.00631412910297513, + 0.004409517161548138, + 0.004389329347759485, + 0.0037943855859339237, + 0.004858137108385563, + 0.0042280410416424274, + 0.002853803336620331, + 0.0025355047546327114, + 0.006428467575460672, + 0.004383640363812447, + 0.00306266313418746, + 0.013848290778696537, + 0.00913265161216259, + 0.014746420085430145, + 0.032158467918634415, + 0.010940073989331722, + 0.023758022114634514, + 0.020166819915175438, + 0.03867598995566368, + 0.004348664078861475, + 0.002270132303237915, + 0.004717492964118719, + 0.002065985929220915, + 0.003525177016854286, + 0.003972596954554319, + 0.008869645185768604, + 0.010232051834464073, + 0.006326362956315279, + 0.004495199304074049, + 0.007062885910272598, + 0.006239074282348156, + 0.00541686499491334, + 0.007302108686417341, + 0.005743175745010376, + 0.007423819974064827, + 0.007188604213297367, + 0.005593088921159506, + 0.0067564756609499454, + 0.006688164081424475, + 0.004886529874056578, + 0.004638993181288242, + 0.0068571497686207294, + 0.006925520487129688, + 0.012636257335543633, + 0.007227281108498573, + 0.012159581296145916, + 0.010395093820989132, + 0.008508170023560524, + 0.010693916119635105, + 0.012162446975708008, + 0.01621442288160324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.01342424750328064 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028368488419800997 + ], + "opposite_logprob_tokens": [ + -5.87783670425415 + ], + "logprob_reward": -0.0028368488419800997, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028368488419800997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45599693059921265, + 0.08159666508436203, + 0.07945694029331207, + 0.02721240185201168, + 0.05022324249148369, + 0.05972382798790932, + 0.02670486643910408, + 0.07504604011774063, + 0.0298139825463295, + 0.11422508955001831 + ], + "credit_map": [ + 0.07029201090335846, + 0.08818686008453369, + 0.027677590027451515, + 0.021679185330867767, + 0.03036140650510788, + 0.14616647362709045, + 0.024439165368676186, + 0.047194235026836395, + 0.007850686088204384, + 0.006145718041807413, + 0.013326887972652912, + 0.011677000671625137, + 0.011488227173686028, + 0.011004189029335976, + 0.009085830301046371, + 0.011018127202987671, + 0.008097881451249123, + 0.0076508536003530025, + 0.012142570689320564, + 0.006803103256970644, + 0.01701289974153042, + 0.009985295124351978, + 0.009046492166817188, + 0.008717845194041729, + 0.0028336872346699238, + 0.003038623370230198, + 0.0036753194872289896, + 0.003739133244380355, + 0.004259170964360237, + 0.0034256353974342346, + 0.0029299354646354914, + 0.003310896921902895, + 0.006332476623356342, + 0.008116628043353558, + 0.0076035927049815655, + 0.004902385175228119, + 0.004375487100332975, + 0.007676600478589535, + 0.006372489966452122, + 0.004843584261834621, + 0.005265890154987574, + 0.004444865044206381, + 0.004783085081726313, + 0.013028285466134548, + 0.004475138615816832, + 0.0078006708063185215, + 0.006082727573812008, + 0.013843162916600704, + 0.004679408855736256, + 0.0026769116520881653, + 0.005311016459017992, + 0.001862542238086462, + 0.0027964352630078793, + 0.0027849730104207993, + 0.0026401886716485023, + 0.003953392617404461, + 0.010868874378502369, + 0.005604656413197517, + 0.009055465459823608, + 0.0060189347714185715, + 0.00877797044813633, + 0.013649151660501957, + 0.010665816254913807, + 0.010405168868601322, + 0.0030925872270017862, + 0.003772674361243844, + 0.003389861201867461, + 0.005216667428612709, + 0.0032722887117415667, + 0.0038317611906677485, + 0.0030058773700147867, + 0.004232264123857021, + 0.019145773723721504, + 0.010074240155518055, + 0.015094724483788013, + 0.011200238019227982, + 0.01146736554801464, + 0.01317885797470808, + 0.013912282884120941, + 0.020151596516370773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0028368488419800997 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866885766386986 + ], + "opposite_logprob_tokens": [ + -3.04866886138916 + ], + "logprob_reward": -0.04866885766386986, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866885766386986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29428815841674805, + 0.047049976885318756, + 0.04248762130737305, + 0.024947162717580795, + 0.035954274237155914, + 0.03335946425795555, + 0.027243342250585556, + 0.12956231832504272, + 0.0753469169139862, + 0.2897607684135437 + ], + "credit_map": [ + 0.032771509140729904, + 0.08395969867706299, + 0.018522515892982483, + 0.01699557527899742, + 0.030724000185728073, + 0.061158061027526855, + 0.018426064401865005, + 0.03173072636127472, + 0.00463154586032033, + 0.003875216469168663, + 0.006925012916326523, + 0.005314833018928766, + 0.006267415825277567, + 0.008233929984271526, + 0.004602439701557159, + 0.007199585437774658, + 0.004739508964121342, + 0.004256810061633587, + 0.0069886380806565285, + 0.0034326021559536457, + 0.006477157585322857, + 0.006820561829954386, + 0.004488804377615452, + 0.005283536855131388, + 0.0029744766652584076, + 0.002314847894012928, + 0.004272631369531155, + 0.002102099359035492, + 0.0025917396415024996, + 0.003629940329119563, + 0.003080917289480567, + 0.0039805094711482525, + 0.004262582864612341, + 0.00505260843783617, + 0.005233623553067446, + 0.003314785659313202, + 0.0026647436898201704, + 0.0067540863528847694, + 0.004968574736267328, + 0.0037032701075077057, + 0.0023920077364891768, + 0.0032012087758630514, + 0.0031179965008050203, + 0.005772936157882214, + 0.0029410789720714092, + 0.005475632380694151, + 0.0036776307970285416, + 0.006780974101275206, + 0.0031109610572457314, + 0.002128638094291091, + 0.0035560817923396826, + 0.0018388241296634078, + 0.002339621540158987, + 0.002823291113600135, + 0.00523063400760293, + 0.006215289235115051, + 0.022469324991106987, + 0.008332574740052223, + 0.017476819455623627, + 0.012288223020732403, + 0.01307266391813755, + 0.02319907955825329, + 0.012810083106160164, + 0.019913559779524803, + 0.010554855689406395, + 0.010925143957138062, + 0.008855331689119339, + 0.011532364413142204, + 0.007881077006459236, + 0.006525603588670492, + 0.008779579773545265, + 0.010292956605553627, + 0.03681318089365959, + 0.017933394759893417, + 0.03812990337610245, + 0.024585558101534843, + 0.029385991394519806, + 0.035467080771923065, + 0.0523889996111393, + 0.05505668371915817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.04866885766386986 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307778552174568 + ], + "opposite_logprob_tokens": [ + -3.1680777072906494 + ], + "logprob_reward": -0.04307778552174568, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307778552174568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40604540705680847, + 0.15733836591243744, + 0.09614311903715134, + 0.03186381235718727, + 0.04230540618300438, + 0.07074997574090958, + 0.03437032550573349, + 0.04891598969697952, + 0.05576420947909355, + 0.056503403931856155 + ], + "credit_map": [ + 0.06148546189069748, + 0.0789133608341217, + 0.026507453992962837, + 0.019932856783270836, + 0.0275440476834774, + 0.12648867070674896, + 0.022493695840239525, + 0.04267987236380577, + 0.018553197383880615, + 0.015655435621738434, + 0.02531249262392521, + 0.01400313712656498, + 0.018505321815609932, + 0.021013135090470314, + 0.014561101794242859, + 0.029734548181295395, + 0.012980387546122074, + 0.00943959690630436, + 0.015014040283858776, + 0.008146868087351322, + 0.014651766046881676, + 0.013638431206345558, + 0.00907937902957201, + 0.013192648068070412, + 0.004407266154885292, + 0.0031319204717874527, + 0.0054963137954473495, + 0.0022941776551306248, + 0.0026532826013863087, + 0.005723237991333008, + 0.003488818882033229, + 0.004668795503675938, + 0.005504597909748554, + 0.004624796099960804, + 0.005985857918858528, + 0.004469362553209066, + 0.003353292355313897, + 0.008702099323272705, + 0.0052374303340911865, + 0.004427969455718994, + 0.004945741966366768, + 0.006772854831069708, + 0.004991827066987753, + 0.01945442706346512, + 0.006300343666225672, + 0.005641038995236158, + 0.0049493820406496525, + 0.01769435964524746, + 0.004980445373803377, + 0.0023088911548256874, + 0.0047040884383022785, + 0.0020517169032245874, + 0.003839482320472598, + 0.004195721819996834, + 0.005775362253189087, + 0.006514618173241615, + 0.006489327643066645, + 0.0034328983165323734, + 0.005870061926543713, + 0.005174492951482534, + 0.004034912213683128, + 0.00816181767731905, + 0.006546271964907646, + 0.009206204675137997, + 0.007928314618766308, + 0.005723957438021898, + 0.006771376822143793, + 0.0060442714020609856, + 0.006110521499067545, + 0.008776456117630005, + 0.006536202505230904, + 0.0078731095418334, + 0.00775018660351634, + 0.005813885014504194, + 0.008704469539225101, + 0.005321434233337641, + 0.00699709914624691, + 0.005655412096530199, + 0.007213401608169079, + 0.009047518484294415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04307778552174568 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070127487182617 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.016070127487182617, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070127487182617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08275294303894043, + 0.1039975956082344, + 0.10534383356571198, + 0.04198528826236725, + 0.1056625247001648, + 0.2646181583404541, + 0.05080879479646683, + 0.06300520896911621, + 0.09944724291563034, + 0.08237847685813904 + ], + "credit_map": [ + 0.011246073991060257, + 0.006269159726798534, + 0.013168209232389927, + 0.007671758066862822, + 0.00899441633373499, + 0.008592979982495308, + 0.007562771439552307, + 0.019247565418481827, + 0.012587727047502995, + 0.007932125590741634, + 0.015397630631923676, + 0.008649240247905254, + 0.012764276936650276, + 0.013977276161313057, + 0.009469371289014816, + 0.02321993187069893, + 0.014676668681204319, + 0.008551220409572124, + 0.019963495433330536, + 0.010140982456505299, + 0.007571570575237274, + 0.016873622313141823, + 0.012765241786837578, + 0.014801016077399254, + 0.004447536543011665, + 0.006808050442487001, + 0.007477148436009884, + 0.0030502653680741787, + 0.006670818664133549, + 0.005286789499223232, + 0.004794075153768063, + 0.0034505988005548716, + 0.01409844495356083, + 0.01382145844399929, + 0.014645093120634556, + 0.011386395432054996, + 0.010074751451611519, + 0.016821354627609253, + 0.013858850114047527, + 0.01095616351813078, + 0.0180705226957798, + 0.026850750669836998, + 0.019531717523932457, + 0.06989632546901703, + 0.025309646502137184, + 0.0206321831792593, + 0.01869794726371765, + 0.0656290277838707, + 0.006786181125789881, + 0.010064288973808289, + 0.005945033393800259, + 0.004645591601729393, + 0.008650493808090687, + 0.004214269574731588, + 0.005136728286743164, + 0.005366202909499407, + 0.008891780860722065, + 0.00679384358227253, + 0.009788810275495052, + 0.006379870697855949, + 0.0076768891885876656, + 0.007314226124435663, + 0.008703184314072132, + 0.00745659414678812, + 0.020207328721880913, + 0.008585895411670208, + 0.013739203102886677, + 0.008653218857944012, + 0.009981861338019371, + 0.013367057777941227, + 0.008137275464832783, + 0.016775382682681084, + 0.011060510762035847, + 0.004870484117418528, + 0.017342504113912582, + 0.0072889067232608795, + 0.010022933594882488, + 0.00833530817180872, + 0.013032768853008747, + 0.010425056330859661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.016070127487182617 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures a dramatic moment in a dense bamboo forest at night, rendered with the powerful, stylized aesthetic of a modern comic book panel. A sudden, dazzling, jagged prong of electric white lightning violently splits the inky sky, its tendrils branching out in a tangled web. This momentary flash bathes the entire scene in dramatic, high-contrast light, captured with the razor-sharp precision of 4K footage. Each element is defined by bold, decisive ink lines and deep, graphic shadows, completely eliminating mid-tones. The towering bamboo stems become a rhythmic pattern, stark white vertical lines set against impenetrable black voids, their leaves catching the flash like sharp silver daggers. Below, a winding stone path winds through the forest. The jagged stones glow damply, contrasting the scene's only curves with the rigid lines of the bamboo, while the shadows between them are pools of pure, featureless black. The entire image crystallizes in this single, powerful moment, freezing the raw power of nature in a moment of stunning, intense, and hyper-stylized visual impact.", + "seed": 22070, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_9/index_7905_seed_22070_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.013939903490245343, + "vlm_logprob": -0.013939903490245343, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.013939903490245343, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018904630094766617 + ], + "opposite_logprob_tokens": [ + -4.143904685974121 + ], + "logprob_reward": -0.018904630094766617, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.018904630094766617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09212525188922882, + 0.16521278023719788, + 0.1606587916612625, + 0.12249363213777542, + 0.05415746197104454, + 0.04845857992768288, + 0.07615926116704941, + 0.05052186921238899, + 0.08014897257089615, + 0.1500634253025055 + ], + "credit_map": [ + 0.013688359409570694, + 0.011895162984728813, + 0.012748743407428265, + 0.00945485569536686, + 0.012723042629659176, + 0.0088207321241498, + 0.0138243418186903, + 0.008970016613602638, + 0.02631198614835739, + 0.015376942232251167, + 0.029224885627627373, + 0.0153989652171731, + 0.013811182230710983, + 0.02167883701622486, + 0.0128847137093544, + 0.030525263398885727, + 0.022373169660568237, + 0.030544614419341087, + 0.01951725408434868, + 0.014002072624862194, + 0.02814875729382038, + 0.0187241043895483, + 0.009854777716100216, + 0.01749403402209282, + 0.015528250485658646, + 0.011613043025135994, + 0.015066771768033504, + 0.015606006607413292, + 0.01631762646138668, + 0.018961839377880096, + 0.012523936107754707, + 0.01687616854906082, + 0.010902474634349346, + 0.009450477547943592, + 0.004684446845203638, + 0.005146115552634001, + 0.005676974542438984, + 0.00804182793945074, + 0.004671881906688213, + 0.005583262536674738, + 0.008390888571739197, + 0.005156523082405329, + 0.005092081613838673, + 0.005634747911244631, + 0.010255889035761356, + 0.005026400554925203, + 0.005065510515123606, + 0.0038365377113223076, + 0.013014622032642365, + 0.011115551926195621, + 0.006126321852207184, + 0.0066771125420928, + 0.0107216015458107, + 0.01301868911832571, + 0.006363199092447758, + 0.009122163988649845, + 0.009578831493854523, + 0.006765428464859724, + 0.007497844751924276, + 0.004279094282537699, + 0.005161162465810776, + 0.007923503406345844, + 0.0034635623451322317, + 0.005852439906448126, + 0.01129562221467495, + 0.01067337952554226, + 0.008190476335585117, + 0.005943702068179846, + 0.012133977375924587, + 0.009729127399623394, + 0.012442066334187984, + 0.009740614332258701, + 0.0262178685516119, + 0.030239107087254524, + 0.012308135628700256, + 0.00865152757614851, + 0.018621595576405525, + 0.032086748629808426, + 0.009231015108525753, + 0.012707415036857128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.018904630094766617 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011095255613327026 + ], + "opposite_logprob_tokens": [ + -4.51109504699707 + ], + "logprob_reward": -0.011095255613327026, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011095255613327026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28016266226768494, + 0.12960395216941833, + 0.1293964385986328, + 0.07570651918649673, + 0.10373693704605103, + 0.04719998314976692, + 0.029167301952838898, + 0.025219688192009926, + 0.0550021268427372, + 0.12480445206165314 + ], + "credit_map": [ + 0.04604458808898926, + 0.06817997992038727, + 0.03062334842979908, + 0.024280624464154243, + 0.046998023986816406, + 0.016277628019452095, + 0.03088526614010334, + 0.016873173415660858, + 0.020465558394789696, + 0.019182220101356506, + 0.011505547910928726, + 0.01723271794617176, + 0.01330661866813898, + 0.011195672675967216, + 0.019838163629174232, + 0.016877448186278343, + 0.024848125874996185, + 0.01620665192604065, + 0.018698111176490784, + 0.013601151295006275, + 0.01530396193265915, + 0.01667824573814869, + 0.007919597439467907, + 0.01614057458937168, + 0.01099157240241766, + 0.007930577732622623, + 0.00905065331608057, + 0.00571429543197155, + 0.006788135040551424, + 0.01602020300924778, + 0.005688994191586971, + 0.013522077351808548, + 0.02157443016767502, + 0.011877561919391155, + 0.016008980572223663, + 0.010368273593485355, + 0.013505877926945686, + 0.0114412447437644, + 0.008117650635540485, + 0.010842906311154366, + 0.014087324030697346, + 0.002852063160389662, + 0.004147855564951897, + 0.002043828833848238, + 0.0047630625776946545, + 0.01309194602072239, + 0.0026368037797510624, + 0.0035770945250988007, + 0.004586654715240002, + 0.004608153831213713, + 0.004093460738658905, + 0.0032520329114049673, + 0.003076297929510474, + 0.004255558829754591, + 0.0021045180037617683, + 0.003190623363479972, + 0.004613839089870453, + 0.0036161458119750023, + 0.0026093865744769573, + 0.0020336471498012543, + 0.002592015080153942, + 0.004518207628279924, + 0.00202564406208694, + 0.00321080069988966, + 0.008591536432504654, + 0.006910091266036034, + 0.006568605545908213, + 0.0037640088703483343, + 0.007899388670921326, + 0.007616945076733828, + 0.007253626361489296, + 0.006397915538400412, + 0.02239960990846157, + 0.01776016689836979, + 0.009466664865612984, + 0.007185134571045637, + 0.017908575013279915, + 0.026641584932804108, + 0.010697092860937119, + 0.01274560485035181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011095255613327026 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012548303231596947 + ], + "opposite_logprob_tokens": [ + -4.387548446655273 + ], + "logprob_reward": -0.012548303231596947, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012548303231596947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06470843404531479, + 0.06953318417072296, + 0.43887045979499817, + 0.13947059214115143, + 0.1970754861831665, + 0.039643295109272, + 0.012830140069127083, + 0.009154392406344414, + 0.008962813764810562, + 0.019751248881220818 + ], + "credit_map": [ + 0.010838277637958527, + 0.018061533570289612, + 0.007738177198916674, + 0.0044179740361869335, + 0.010637572035193443, + 0.005325010512024164, + 0.004298110492527485, + 0.0033917853143066168, + 0.009098016656935215, + 0.014085491187870502, + 0.00739516643807292, + 0.0035148337483406067, + 0.0073606958612799644, + 0.007595219649374485, + 0.003405228490009904, + 0.017078528180718422, + 0.06807805597782135, + 0.08927006274461746, + 0.05576934292912483, + 0.04094981774687767, + 0.0631554052233696, + 0.0600181445479393, + 0.021569183096289635, + 0.04006044566631317, + 0.02019352838397026, + 0.021638836711645126, + 0.01116161048412323, + 0.009609144181013107, + 0.014563201926648617, + 0.02339196763932705, + 0.009902501478791237, + 0.029009807854890823, + 0.04767932370305061, + 0.0415748655796051, + 0.01619182899594307, + 0.016136547550559044, + 0.018381750211119652, + 0.016833407804369926, + 0.014265080913901329, + 0.02601267211139202, + 0.009447143413126469, + 0.0027406716253608465, + 0.004422138445079327, + 0.002819472225382924, + 0.0040258667431771755, + 0.008752536028623581, + 0.0030032950453460217, + 0.004432175308465958, + 0.0023175454698503017, + 0.0018697535851970315, + 0.0011481551919132471, + 0.0017410324653610587, + 0.00155703688506037, + 0.001785662374459207, + 0.0013096693437546492, + 0.001101285801269114, + 0.0016714450903236866, + 0.0012458432465791702, + 0.0011683705961331725, + 0.0008118093246594071, + 0.0012932650279253721, + 0.001079836511053145, + 0.0009147985838353634, + 0.0009690238512121141, + 0.00128274189773947, + 0.001024094526655972, + 0.001217559678480029, + 0.0008852853206917644, + 0.0014991022180765867, + 0.0012613902799785137, + 0.0006721931276842952, + 0.0011204469483345747, + 0.0031198281794786453, + 0.002856555161997676, + 0.0023292286787182093, + 0.001405220478773117, + 0.002798896050080657, + 0.0030171102844178677, + 0.002101993653923273, + 0.002122415928170085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.012548303231596947 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075054183602333 + ], + "opposite_logprob_tokens": [ + -4.141075134277344 + ], + "logprob_reward": -0.016075054183602333, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075054183602333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20823194086551666, + 0.05969255045056343, + 0.2167813926935196, + 0.12639395892620087, + 0.05670914426445961, + 0.046533193439245224, + 0.04599384963512421, + 0.022721147164702415, + 0.03303651139140129, + 0.18390627205371857 + ], + "credit_map": [ + 0.03391975164413452, + 0.062214478850364685, + 0.02270885743200779, + 0.014487965032458305, + 0.03678353130817413, + 0.013229911215603352, + 0.014903012663125992, + 0.009984452277421951, + 0.008881540969014168, + 0.006991399917751551, + 0.009929610416293144, + 0.006899289786815643, + 0.006152179092168808, + 0.007584744598716497, + 0.005247551016509533, + 0.008006236515939236, + 0.0454757884144783, + 0.054586008191108704, + 0.024793945252895355, + 0.01080058328807354, + 0.02803242765367031, + 0.031529683619737625, + 0.008335704915225506, + 0.013227242976427078, + 0.01574035920202732, + 0.014212263748049736, + 0.01327822357416153, + 0.009736021049320698, + 0.014419020153582096, + 0.021674806252121925, + 0.010453910566866398, + 0.02687935344874859, + 0.00901759136468172, + 0.007015562616288662, + 0.007126701530069113, + 0.0041180383414030075, + 0.010371118783950806, + 0.007215557619929314, + 0.004892581608146429, + 0.006951996590942144, + 0.01351440604776144, + 0.0035792929120361805, + 0.004618718288838863, + 0.0026296302676200867, + 0.004150933586061001, + 0.012005051597952843, + 0.0030936808325350285, + 0.0029414831660687923, + 0.0076569076627492905, + 0.008988404646515846, + 0.00372842769138515, + 0.0046237194910645485, + 0.007682147901505232, + 0.0056967707350850105, + 0.003968729171901941, + 0.003648745594546199, + 0.003717507002875209, + 0.0032421385403722525, + 0.0030543419998139143, + 0.00219812523573637, + 0.00297363568097353, + 0.002943424042314291, + 0.0021009857300668955, + 0.0024909903295338154, + 0.004779062233865261, + 0.004350408911705017, + 0.00392067851498723, + 0.0022728331387043, + 0.004362182691693306, + 0.004454219713807106, + 0.004578133579343557, + 0.004318994004279375, + 0.033868398517370224, + 0.03103603422641754, + 0.013358615338802338, + 0.010478014126420021, + 0.023285970091819763, + 0.042303357273340225, + 0.013394076377153397, + 0.016181832179427147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.016075054183602333 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076274327933788 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076274327933788, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076274327933788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09438679367303848, + 0.07555852830410004, + 0.41803428530693054, + 0.0605083629488945, + 0.07784218341112137, + 0.06375131756067276, + 0.048442911356687546, + 0.02153012529015541, + 0.07040983438491821, + 0.06953564286231995 + ], + "credit_map": [ + 0.009763265028595924, + 0.021333463490009308, + 0.01079435832798481, + 0.011538145132362843, + 0.016480309888720512, + 0.007407898548990488, + 0.01062682829797268, + 0.006442526821047068, + 0.012128470465540886, + 0.007817151956260204, + 0.01080123521387577, + 0.007978728972375393, + 0.008429781533777714, + 0.010308598168194294, + 0.009918764233589172, + 0.008175801485776901, + 0.07965902984142303, + 0.10340490937232971, + 0.051743414252996445, + 0.026810355484485626, + 0.052550286054611206, + 0.05810971558094025, + 0.017206965014338493, + 0.028549611568450928, + 0.007528818212449551, + 0.008779718540608883, + 0.0067803519777953625, + 0.005897500552237034, + 0.008604362607002258, + 0.008695793338119984, + 0.005069056525826454, + 0.009152764454483986, + 0.014994017779827118, + 0.012050919234752655, + 0.008882145397365093, + 0.006441089790314436, + 0.010157040320336819, + 0.008838389068841934, + 0.005260185804218054, + 0.011218400672078133, + 0.017206301912665367, + 0.005978160072118044, + 0.006508569698780775, + 0.00477913161739707, + 0.011087141931056976, + 0.011146488599479198, + 0.003432013327255845, + 0.003613510634750128, + 0.007869867607951164, + 0.008563459850847721, + 0.004999813158065081, + 0.006379648577421904, + 0.007000565063208342, + 0.005089411977678537, + 0.003959835506975651, + 0.004580315668135881, + 0.004090096801519394, + 0.002155569614842534, + 0.0032680057920515537, + 0.0017262345645576715, + 0.002493574284017086, + 0.0039466870948672295, + 0.0015804753638803959, + 0.002269482472911477, + 0.01050924975425005, + 0.008180114440619946, + 0.007829836569726467, + 0.0054561379365623, + 0.011301657184958458, + 0.008905964903533459, + 0.008357263170182705, + 0.009869616478681564, + 0.011645740829408169, + 0.00702327536419034, + 0.006161561701446772, + 0.004475576337426901, + 0.011896857060492039, + 0.016993943601846695, + 0.006685526575893164, + 0.004653164651244879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.011076274327933788 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures a dramatic moment in a dense bamboo forest at night, rendered with the powerful, stylized aesthetic of a modern comic book panel. A sudden, dazzling, jagged prong of electric white lightning violently splits the inky sky, its tendrils branching out in a tangled web. This momentary flash bathes the entire scene in dramatic, high-contrast light, captured with the razor-sharp precision of 4K footage. Each element is defined by bold, decisive ink lines and deep, graphic shadows, completely eliminating mid-tones. The towering bamboo stems become a rhythmic pattern, stark white vertical lines set against impenetrable black voids, their leaves catching the flash like sharp silver daggers. Below, a winding stone path winds through the forest. The jagged stones glow damply, contrasting the scene's only curves with the rigid lines of the bamboo, while the shadows between them are pools of pure, featureless black. The entire image crystallizes in this single, powerful moment, freezing the raw power of nature in a moment of stunning, intense, and hyper-stylized visual impact.", + "seed": 22071, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_9/index_7905_seed_22071_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028353248769417406, + "vlm_logprob": -0.028353248769417406, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.028353248769417406, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01670140214264393 + ], + "opposite_logprob_tokens": [ + -4.2667012214660645 + ], + "logprob_reward": -0.01670140214264393, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01670140214264393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3665502369403839, + 0.08933133631944656, + 0.023302679881453514, + 0.025956690311431885, + 0.06030247360467911, + 0.048344094306230545, + 0.08224955201148987, + 0.06468833237886429, + 0.146896094083786, + 0.09237851202487946 + ], + "credit_map": [ + 0.03396732360124588, + 0.10201139003038406, + 0.015244475565850735, + 0.14038051664829254, + 0.02148308791220188, + 0.01736687868833542, + 0.02405559830367565, + 0.012040977366268635, + 0.008718980476260185, + 0.011129588820040226, + 0.0074293361976742744, + 0.012414100579917431, + 0.010161565616726875, + 0.014810262247920036, + 0.008389897644519806, + 0.016277605667710304, + 0.0018030621577054262, + 0.0021687126718461514, + 0.0016667627496644855, + 0.0030911839567124844, + 0.0048782662488520145, + 0.0030562281608581543, + 0.0046433936804533005, + 0.0019950703717768192, + 0.003002310171723366, + 0.002068874193355441, + 0.003565805032849312, + 0.0030516982078552246, + 0.0041780672036111355, + 0.004098966717720032, + 0.003443541470915079, + 0.002547426149249077, + 0.005226476117968559, + 0.008103661239147186, + 0.008481957949697971, + 0.00774363661184907, + 0.012692848220467567, + 0.0043190293945372105, + 0.0069648502394557, + 0.006770011968910694, + 0.0037962624337524176, + 0.0038600959815084934, + 0.006440191064029932, + 0.007502958178520203, + 0.006526971235871315, + 0.005073335487395525, + 0.0070343418046832085, + 0.008109941147267818, + 0.006878301501274109, + 0.010623450390994549, + 0.007030937820672989, + 0.013839400373399258, + 0.012633814476430416, + 0.006822195369750261, + 0.012638275511562824, + 0.0117831751704216, + 0.003409509314224124, + 0.005299134645611048, + 0.0033779186196625233, + 0.017955679446458817, + 0.016086578369140625, + 0.005370923783630133, + 0.007384940050542355, + 0.005803649313747883, + 0.008567443117499352, + 0.016007428988814354, + 0.00761657627299428, + 0.042041681706905365, + 0.025971615687012672, + 0.01130847167223692, + 0.02181417867541313, + 0.013568696565926075, + 0.01030183956027031, + 0.014463845640420914, + 0.006553419400006533, + 0.019260989502072334, + 0.013816621154546738, + 0.01100650429725647, + 0.00984737928956747, + 0.007127915043383837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.01670140214264393 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036258448380976915 + ], + "opposite_logprob_tokens": [ + -5.628625869750977 + ], + "logprob_reward": -0.0036258448380976915, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036258448380976915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1280609369277954, + 0.19495920836925507, + 0.07453031837940216, + 0.06990332156419754, + 0.0706477165222168, + 0.08587229251861572, + 0.1012764573097229, + 0.09383200109004974, + 0.09740184247493744, + 0.08351591974496841 + ], + "credit_map": [ + 0.01737043261528015, + 0.024273734539747238, + 0.00875888392329216, + 0.026289744302630424, + 0.01321488618850708, + 0.011292064562439919, + 0.017651628702878952, + 0.00920956116169691, + 0.0156417079269886, + 0.03282048553228378, + 0.012402628548443317, + 0.04856780543923378, + 0.03205466270446777, + 0.01875624991953373, + 0.022500742226839066, + 0.012214930728077888, + 0.004632875323295593, + 0.009008829481899738, + 0.00533596146851778, + 0.01967257261276245, + 0.01089807040989399, + 0.008250406011939049, + 0.010920073837041855, + 0.00581152830272913, + 0.006862392649054527, + 0.005591605789959431, + 0.009246580302715302, + 0.010753205046057701, + 0.01009555533528328, + 0.009000289253890514, + 0.008578033186495304, + 0.009775660932064056, + 0.005587446503341198, + 0.007520013954490423, + 0.010270027443766594, + 0.011151579208672047, + 0.012065816670656204, + 0.005135589279234409, + 0.00825670175254345, + 0.010660539381206036, + 0.006746198982000351, + 0.008759069256484509, + 0.010780640877783298, + 0.012497511692345142, + 0.01395121868699789, + 0.008988901041448116, + 0.012203074060380459, + 0.011945683509111404, + 0.009042516350746155, + 0.012952592223882675, + 0.007865930907428265, + 0.013152009807527065, + 0.024931199848651886, + 0.007722985930740833, + 0.016571510583162308, + 0.009037718176841736, + 0.005785960238426924, + 0.015989601612091064, + 0.006373060867190361, + 0.0179473664611578, + 0.020746974274516106, + 0.006227494683116674, + 0.011729461140930653, + 0.009032084606587887, + 0.00733914552256465, + 0.011699422262609005, + 0.009113925509154797, + 0.019704023376107216, + 0.015356866642832756, + 0.008544736541807652, + 0.015509023331105709, + 0.010134690441191196, + 0.00707386527210474, + 0.011795838363468647, + 0.005407066084444523, + 0.01499231532216072, + 0.018757397308945656, + 0.005799483973532915, + 0.01126830279827118, + 0.008421641774475574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0036258448380976915 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901863008737564 + ], + "opposite_logprob_tokens": [ + -2.5790185928344727 + ], + "logprob_reward": -0.07901863008737564, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901863008737564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09396427869796753, + 0.03812164440751076, + 0.019998926669359207, + 0.048002053052186966, + 0.09173601120710373, + 0.11675740033388138, + 0.043565742671489716, + 0.15246891975402832, + 0.15322960913181305, + 0.24215544760227203 + ], + "credit_map": [ + 0.011990022845566273, + 0.02005460113286972, + 0.006724759936332703, + 0.025577642023563385, + 0.0068971565924584866, + 0.006716813892126083, + 0.009835057891905308, + 0.006168217398226261, + 0.005324966739863157, + 0.004419777076691389, + 0.003085113363340497, + 0.005303040612488985, + 0.003952107857912779, + 0.006605284288525581, + 0.004533974453806877, + 0.004897379316389561, + 0.0017936329822987318, + 0.0017393676098436117, + 0.002426867838948965, + 0.0029880611691623926, + 0.003091629594564438, + 0.0028166123665869236, + 0.0035947321448475122, + 0.0015480213332921267, + 0.004642941057682037, + 0.0065371254459023476, + 0.005662085488438606, + 0.006087800953537226, + 0.008629953488707542, + 0.005600235424935818, + 0.006514163687825203, + 0.004327747039496899, + 0.008287506178021431, + 0.011699664406478405, + 0.01446154061704874, + 0.01085076853632927, + 0.01843887008726597, + 0.007226436398923397, + 0.00898398831486702, + 0.011787224560976028, + 0.009000870399177074, + 0.011780190281569958, + 0.01637599989771843, + 0.01701923832297325, + 0.017518049106001854, + 0.012061170302331448, + 0.015557472594082355, + 0.017444398254156113, + 0.004386785440146923, + 0.004064441192895174, + 0.004536605440080166, + 0.007285234052687883, + 0.0060982354916632175, + 0.0050679538398981094, + 0.006875247228890657, + 0.00525123393163085, + 0.01021154411137104, + 0.01138880755752325, + 0.01651979237794876, + 0.04008690267801285, + 0.029170982539653778, + 0.013567366637289524, + 0.01656084880232811, + 0.014962663874030113, + 0.008374480530619621, + 0.02004939131438732, + 0.012546030804514885, + 0.03945719450712204, + 0.02635134384036064, + 0.010499042458832264, + 0.02395191602408886, + 0.012000201269984245, + 0.020990191027522087, + 0.03585242107510567, + 0.014368781819939613, + 0.05613844469189644, + 0.041727859526872635, + 0.018321050330996513, + 0.030740978196263313, + 0.02401568740606308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.07901863008737564 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026352116838097572 + ], + "opposite_logprob_tokens": [ + -3.6513521671295166 + ], + "logprob_reward": -0.026352116838097572, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026352116838097572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2603203058242798, + 0.1677703559398651, + 0.052885789424180984, + 0.04532632231712341, + 0.05819631367921829, + 0.043409980833530426, + 0.06151372194290161, + 0.07064991444349289, + 0.072588711977005, + 0.16733862459659576 + ], + "credit_map": [ + 0.02725198119878769, + 0.05501881241798401, + 0.018045630306005478, + 0.07774453610181808, + 0.01861576735973358, + 0.025735488161444664, + 0.0186535082757473, + 0.019254591315984726, + 0.016361622139811516, + 0.02248508483171463, + 0.01291254349052906, + 0.025395173579454422, + 0.021171709522604942, + 0.028149334713816643, + 0.014765754342079163, + 0.026529114693403244, + 0.004965659696608782, + 0.004419967997819185, + 0.005138986743986607, + 0.007063072174787521, + 0.008483231998980045, + 0.009431536309421062, + 0.006705111358314753, + 0.006678223609924316, + 0.0038355730939656496, + 0.0034879574086517096, + 0.004410273861140013, + 0.007415447849780321, + 0.007530251517891884, + 0.008017969317734241, + 0.005901871714740992, + 0.004726977553218603, + 0.00518472446128726, + 0.0044272784143686295, + 0.007146164309233427, + 0.011480615474283695, + 0.011976242065429688, + 0.004566715564578772, + 0.007079761009663343, + 0.006334812846034765, + 0.0037946985103189945, + 0.004979670979082584, + 0.004812777508050203, + 0.006421093363314867, + 0.006947072688490152, + 0.005131938029080629, + 0.006029249168932438, + 0.005293479189276695, + 0.004159246105700731, + 0.006099410355091095, + 0.0056888749822974205, + 0.009287167340517044, + 0.015553686767816544, + 0.0049313753843307495, + 0.0096063781529665, + 0.00618758425116539, + 0.0046547213569283485, + 0.006962130311876535, + 0.0072105503641068935, + 0.015912318602204323, + 0.010045480914413929, + 0.008229825645685196, + 0.010151665657758713, + 0.007483218796551228, + 0.004032742232084274, + 0.011601563543081284, + 0.004246962256729603, + 0.014909050427377224, + 0.013428838923573494, + 0.006337035447359085, + 0.010777206160128117, + 0.007255316246300936, + 0.013744309544563293, + 0.02492203190922737, + 0.009360539726912975, + 0.025811869651079178, + 0.03679313883185387, + 0.016395997256040573, + 0.01683657616376877, + 0.023474160581827164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.026352116838097572 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016068249940872192 + ], + "opposite_logprob_tokens": [ + -4.141068458557129 + ], + "logprob_reward": -0.016068249940872192, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016068249940872192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19352102279663086, + 0.09247191250324249, + 0.03724395111203194, + 0.05668694153428078, + 0.08437040448188782, + 0.040380384773015976, + 0.06297945231199265, + 0.13969428837299347, + 0.11741480231285095, + 0.17523682117462158 + ], + "credit_map": [ + 0.0190814808011055, + 0.04076122120022774, + 0.012423758395016193, + 0.049912936985492706, + 0.01847199723124504, + 0.021447094157338142, + 0.0169081911444664, + 0.01451436709612608, + 0.009932412765920162, + 0.010483751073479652, + 0.007683359086513519, + 0.015164071694016457, + 0.01378037128597498, + 0.00891805812716484, + 0.010310995392501354, + 0.0161988977342844, + 0.0034674089401960373, + 0.003030744381248951, + 0.004003632813692093, + 0.006084159482270479, + 0.004459656309336424, + 0.007549358531832695, + 0.0041265846230089664, + 0.004522411618381739, + 0.006436103023588657, + 0.004389626905322075, + 0.008378296159207821, + 0.007218143902719021, + 0.008800843730568886, + 0.008511949330568314, + 0.006300082430243492, + 0.006651900243014097, + 0.008456696756184101, + 0.008568217046558857, + 0.012364014983177185, + 0.012143677100539207, + 0.011640099808573723, + 0.007063410710543394, + 0.009478587657213211, + 0.014655710197985172, + 0.003963160794228315, + 0.005297836381942034, + 0.005071520805358887, + 0.004818563815206289, + 0.005344924982637167, + 0.003421684494242072, + 0.0055581266060471535, + 0.006904568988829851, + 0.005973093677312136, + 0.007318188901990652, + 0.0040192981250584126, + 0.0074983881786465645, + 0.016418296843767166, + 0.006158800795674324, + 0.010247946716845036, + 0.005345440935343504, + 0.008473158814013004, + 0.01092533115297556, + 0.013726719655096531, + 0.025301320478320122, + 0.023198219016194344, + 0.012846398167312145, + 0.01665058173239231, + 0.028572579845786095, + 0.006702657788991928, + 0.012512034736573696, + 0.006506751757115126, + 0.027272086590528488, + 0.018874578177928925, + 0.011627180501818657, + 0.02302026003599167, + 0.010899259708821774, + 0.012790191918611526, + 0.023679165169596672, + 0.00934688188135624, + 0.04342813417315483, + 0.0330943837761879, + 0.013056222349405289, + 0.027247676625847816, + 0.012594168074429035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.016068249940872192 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot captures a dramatic moment in a dense bamboo forest at night, rendered with the powerful, stylized aesthetic of a modern comic book panel. A sudden, dazzling, jagged prong of electric white lightning violently splits the inky sky, its tendrils branching out in a tangled web. This momentary flash bathes the entire scene in dramatic, high-contrast light, captured with the razor-sharp precision of 4K footage. Each element is defined by bold, decisive ink lines and deep, graphic shadows, completely eliminating mid-tones. The towering bamboo stems become a rhythmic pattern, stark white vertical lines set against impenetrable black voids, their leaves catching the flash like sharp silver daggers. Below, a winding stone path winds through the forest. The jagged stones glow damply, contrasting the scene's only curves with the rigid lines of the bamboo, while the shadows between them are pools of pure, featureless black. The entire image crystallizes in this single, powerful moment, freezing the raw power of nature in a moment of stunning, intense, and hyper-stylized visual impact.", + "seed": 22072, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000024/group_9/index_7905_seed_22072_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.045628327131271365, + "vlm_logprob": -0.045628327131271365, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.045628327131271365, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10225468128919601 + ], + "opposite_logprob_tokens": [ + -2.352254629135132 + ], + "logprob_reward": -0.10225468128919601, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10225468128919601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5977813005447388, + 0.06492979824542999, + 0.042239245027303696, + 0.023957645520567894, + 0.044712334871292114, + 0.1457863301038742, + 0.048193614929914474, + 0.008025314658880234, + 0.017753632739186287, + 0.006620799656957388 + ], + "credit_map": [ + 0.08204139024019241, + 0.027765294536948204, + 0.061744485050439835, + 0.11378484219312668, + 0.03730624541640282, + 0.12949183583259583, + 0.08742143958806992, + 0.05822579562664032, + 0.011107285507023335, + 0.007058500312268734, + 0.0070264991372823715, + 0.0038459182251244783, + 0.006971993017941713, + 0.009544541127979755, + 0.013354037888348103, + 0.0060210274532437325, + 0.0050353314727544785, + 0.006798190530389547, + 0.0049939267337322235, + 0.0059774271212518215, + 0.004560426808893681, + 0.006191491149365902, + 0.0038839965127408504, + 0.004798457026481628, + 0.004373124800622463, + 0.002219225512817502, + 0.0025439911987632513, + 0.001722922781482339, + 0.0036063743755221367, + 0.004153348505496979, + 0.002605404006317258, + 0.002733254572376609, + 0.008588901720941067, + 0.003432053606957197, + 0.005038260482251644, + 0.004564568866044283, + 0.0034880819730460644, + 0.004558925982564688, + 0.0064657400362193584, + 0.008575805462896824, + 0.03648402541875839, + 0.009095948189496994, + 0.00775288138538599, + 0.01206333376467228, + 0.015590849332511425, + 0.016692856326699257, + 0.038642022758722305, + 0.00946442037820816, + 0.007252028677612543, + 0.004532811231911182, + 0.004973334260284901, + 0.005595938302576542, + 0.0059140827506780624, + 0.007003560196608305, + 0.005610555876046419, + 0.007311305496841669, + 0.0017525717848911881, + 0.0007224638829939067, + 0.0005975817330181599, + 0.0013289789203554392, + 0.0010315726976841688, + 0.0011691654799506068, + 0.0009919197764247656, + 0.0004310606454964727, + 0.002597648883238435, + 0.0013183518312871456, + 0.0020734339486807585, + 0.0018484911415725946, + 0.0020484852138906717, + 0.0023167941253632307, + 0.0016440649051219225, + 0.003906364552676678, + 0.0009817536920309067, + 0.0005227739457041025, + 0.0007814030977897346, + 0.0007176975486800075, + 0.0006202568765729666, + 0.0009084155899472535, + 0.0005838889628648758, + 0.0015046109911054373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.10225468128919601 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001956336200237274 + ], + "opposite_logprob_tokens": [ + -6.251956462860107 + ], + "logprob_reward": -0.001956336200237274, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001956336200237274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43952512741088867, + 0.1852712631225586, + 0.060065869241952896, + 0.05381350219249725, + 0.07810455560684204, + 0.07358469069004059, + 0.030812393873929977, + 0.022277764976024628, + 0.04617556184530258, + 0.0103693138808012 + ], + "credit_map": [ + 0.06092604994773865, + 0.016755476593971252, + 0.03319660946726799, + 0.08045680820941925, + 0.046442911028862, + 0.03230190649628639, + 0.1378912776708603, + 0.03155410289764404, + 0.04131104424595833, + 0.019759617745876312, + 0.020801519975066185, + 0.01054924726486206, + 0.015305401757359505, + 0.043067533522844315, + 0.017414378002285957, + 0.01706252247095108, + 0.007484747096896172, + 0.007595039904117584, + 0.006419577170163393, + 0.0069251167587935925, + 0.00622930470854044, + 0.009423174895346165, + 0.006129828747361898, + 0.009859076701104641, + 0.009157153777778149, + 0.004025712143629789, + 0.004747007042169571, + 0.005732433870434761, + 0.006674484349787235, + 0.006912014447152615, + 0.006050184369087219, + 0.01051451824605465, + 0.017267998307943344, + 0.005357400514185429, + 0.007250296883285046, + 0.006775082554668188, + 0.008150899782776833, + 0.011010468937456608, + 0.00985923781991005, + 0.012433166615664959, + 0.01621105521917343, + 0.00507793202996254, + 0.004886193200945854, + 0.006948994472622871, + 0.00859398115426302, + 0.007358727045357227, + 0.019061965867877007, + 0.0054458430968225, + 0.004708725027740002, + 0.0033878711983561516, + 0.0024042222648859024, + 0.0034099887125194073, + 0.005395950749516487, + 0.004422839265316725, + 0.004387751221656799, + 0.0026950454339385033, + 0.0052991569973528385, + 0.0019199226517230272, + 0.0014899682719260454, + 0.003228001296520233, + 0.0026708904188126326, + 0.003279387252405286, + 0.0025140566285699606, + 0.0018763808766379952, + 0.006829686928540468, + 0.0036697862669825554, + 0.0055364882573485374, + 0.003473924705758691, + 0.00616647582501173, + 0.00547516904771328, + 0.004668716341257095, + 0.010355313308537006, + 0.0016206912696361542, + 0.0008955489029176533, + 0.0015247090486809611, + 0.0013631180627271533, + 0.0011598945129662752, + 0.001303267665207386, + 0.0012346586445346475, + 0.0012674254830926657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.001956336200237274 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031075030565262 + ], + "opposite_logprob_tokens": [ + -2.350310802459717 + ], + "logprob_reward": -0.10031075030565262, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031075030565262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6300890445709229, + 0.10506980866193771, + 0.04024296998977661, + 0.13528110086917877, + 0.02602875605225563, + 0.023210156708955765, + 0.010934537276625633, + 0.011645141988992691, + 0.0056083486415445805, + 0.011890154331922531 + ], + "credit_map": [ + 0.07860914617776871, + 0.023539239540696144, + 0.050341472029685974, + 0.13767486810684204, + 0.06393583118915558, + 0.03114951401948929, + 0.20332056283950806, + 0.04151838645339012, + 0.011506491340696812, + 0.006460606120526791, + 0.005553869996219873, + 0.00575851509347558, + 0.010730843991041183, + 0.02596478909254074, + 0.030772067606449127, + 0.008322608657181263, + 0.005502748303115368, + 0.00537439901381731, + 0.005496206693351269, + 0.005975420121103525, + 0.0033814639318734407, + 0.007881165482103825, + 0.0031137606129050255, + 0.003517803968861699, + 0.04166560247540474, + 0.01034477073699236, + 0.010118645615875721, + 0.0073897019028663635, + 0.01176480669528246, + 0.018784385174512863, + 0.02795840986073017, + 0.007254771888256073, + 0.00530144153162837, + 0.0018954535480588675, + 0.0030144350603222847, + 0.0028674029745161533, + 0.0021683673840016127, + 0.003606379497796297, + 0.003965584095567465, + 0.0032096903305500746, + 0.005414318293333054, + 0.001449059578590095, + 0.0022834893316030502, + 0.0022973432205617428, + 0.002121520694345236, + 0.002482965588569641, + 0.004844539798796177, + 0.002316920319572091, + 0.001613002736121416, + 0.0010352948447689414, + 0.0011614937102422118, + 0.001207655412144959, + 0.0012335540959611535, + 0.001389708835631609, + 0.0013747455086559057, + 0.0019190808525308967, + 0.0018046082695946097, + 0.0011948950123041868, + 0.0012938595609739423, + 0.0015673772431910038, + 0.0013476447202265263, + 0.0015441939467564225, + 0.0014408944407477975, + 0.0014516683295369148, + 0.0009256456978619099, + 0.0005125635652802885, + 0.000612307689152658, + 0.0008314002770930529, + 0.0005720063927583396, + 0.0007847950328141451, + 0.0006650311406701803, + 0.0007045981474220753, + 0.0022332684602588415, + 0.0010716329561546445, + 0.001476081321015954, + 0.00155267131049186, + 0.0011944775469601154, + 0.0017348608234897256, + 0.0012149029644206166, + 0.0014122574357315898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.10031075030565262 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544183060526848 + ], + "opposite_logprob_tokens": [ + -4.38754415512085 + ], + "logprob_reward": -0.012544183060526848, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544183060526848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6649159789085388, + 0.04015886411070824, + 0.048976652324199677, + 0.08008059859275818, + 0.07244426757097244, + 0.04150824621319771, + 0.010424208827316761, + 0.01407411228865385, + 0.014170724898576736, + 0.013246331363916397 + ], + "credit_map": [ + 0.08274335414171219, + 0.021747490391135216, + 0.045538950711488724, + 0.12591473758220673, + 0.06774313747882843, + 0.06636098027229309, + 0.2111138105392456, + 0.04375355318188667, + 0.0057989610359072685, + 0.004223446827381849, + 0.004144615028053522, + 0.003246465465053916, + 0.004733713809400797, + 0.006970197893679142, + 0.005620596930384636, + 0.0054208701476454735, + 0.005920381285250187, + 0.006319454405456781, + 0.005117305088788271, + 0.004302696790546179, + 0.00539248064160347, + 0.009520131163299084, + 0.005565820261836052, + 0.006838387344032526, + 0.013825404457747936, + 0.006797700189054012, + 0.01007584948092699, + 0.008370812050998211, + 0.008124538697302341, + 0.013488870114088058, + 0.00876961462199688, + 0.01062781922519207, + 0.011594546027481556, + 0.006151407491415739, + 0.007160141598433256, + 0.0077922106720507145, + 0.009319158270955086, + 0.01068188063800335, + 0.009479221887886524, + 0.010265706107020378, + 0.009648253209888935, + 0.0030854835640639067, + 0.0026844653766602278, + 0.004091633949428797, + 0.0044706594198942184, + 0.0055152298882603645, + 0.008761740289628506, + 0.0032507856376469135, + 0.001615164685063064, + 0.0011062658159062266, + 0.0012295610504224896, + 0.0017798119224607944, + 0.0010231704218313098, + 0.0012786311563104391, + 0.0013204675633460283, + 0.0010711378417909145, + 0.0029393911827355623, + 0.0009666546247899532, + 0.001635703258216381, + 0.0011227659415453672, + 0.0009231814765371382, + 0.0019112980226054788, + 0.0014082014095038176, + 0.0031669167801737785, + 0.002020136918872595, + 0.001301066717132926, + 0.0017774499719962478, + 0.0015676168259233236, + 0.002057849895209074, + 0.00185287871863693, + 0.0015023861778900027, + 0.0020913395565003157, + 0.0017825077520683408, + 0.0012312312610447407, + 0.0015677743358537555, + 0.001600006246007979, + 0.0016424184432253242, + 0.0017233772668987513, + 0.0012856752146035433, + 0.002413342474028468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.012544183060526848 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dense bamboo forest at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged prong of electric white lightning visible splitting the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a winding stone path visible below the bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by bold ink lines and deep graphic shadows with no mid-tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075684800744057 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075684800744057, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075684800744057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.653637707233429, + 0.06673002988100052, + 0.05192629247903824, + 0.03471418097615242, + 0.03176455572247505, + 0.09141308814287186, + 0.029480241239070892, + 0.015452730469405651, + 0.012763561680912971, + 0.012117701582610607 + ], + "credit_map": [ + 0.08561620116233826, + 0.021522415801882744, + 0.04527178034186363, + 0.11738170683383942, + 0.067957304418087, + 0.0634492039680481, + 0.20806413888931274, + 0.044374894350767136, + 0.013763178139925003, + 0.006884716916829348, + 0.005911682266741991, + 0.00418236805126071, + 0.007174620404839516, + 0.012826450169086456, + 0.009804456494748592, + 0.006182555109262466, + 0.006743008736521006, + 0.004860103130340576, + 0.006871350575238466, + 0.005287852603942156, + 0.00464328657835722, + 0.011097053997218609, + 0.004394958261400461, + 0.008028676733374596, + 0.0073184301145374775, + 0.0023939330130815506, + 0.002850170014426112, + 0.002788308309391141, + 0.004274578299373388, + 0.004501947667449713, + 0.0076880743727087975, + 0.002898738719522953, + 0.007042558863759041, + 0.0021614152938127518, + 0.0029121125116944313, + 0.003142307512462139, + 0.0028979587368667126, + 0.004697680473327637, + 0.005162031855434179, + 0.0037484874483197927, + 0.025768596678972244, + 0.005490344483405352, + 0.005489393137395382, + 0.007572948466986418, + 0.007849497720599174, + 0.011542108841240406, + 0.02203277312219143, + 0.005667417775839567, + 0.004935987759381533, + 0.0025207067374140024, + 0.0029927182476967573, + 0.0034609129652380943, + 0.0026187817566096783, + 0.003959750756621361, + 0.0028784272726625204, + 0.006112954113632441, + 0.0038951344322413206, + 0.0010706002358347178, + 0.0011179610155522823, + 0.0016861528856679797, + 0.0012984761269763112, + 0.002208271762356162, + 0.001654012710787356, + 0.0025221200194209814, + 0.0017519393004477024, + 0.001251647132448852, + 0.0011904272250831127, + 0.0012908423086628318, + 0.0023345407098531723, + 0.0016720539424568415, + 0.0018654260784387589, + 0.0014066839357838035, + 0.0018147628288716078, + 0.001014156499877572, + 0.0017110157059505582, + 0.0014007462887093425, + 0.0014643348986282945, + 0.0015334176132455468, + 0.0013392240507528186, + 0.0018400429980829358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.011075684800744057 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_0/index_tensor([6980, 6980])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_0/index_tensor([6980, 6980])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2908bf1a9568b439efb9626837adb6ea11ec6c14 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_0/index_tensor([6980, 6980])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, full-body shot, shot from a low, upward angle, captures the majestic presence of Li Jing, Marshal of the Celestial Palace. He clad in a sacred suit of armor meticulously crafted from gold and white jade, each plate engraved with intricate reliefs of auspicious clouds and mythical beasts. His shoulder armor is shaped like a majestic dragon head, and the breastplate on his chest gleams with a restrained yet powerful aura beneath the radiance of the heavens. His face is resolute, his long beard flowing, and he wears a purple-gold crown inlaid with glittering gems, exuding a commanding presence without even anger. In his left palm, he steadily holds a shimmering, seven-story pagoda, shrouded in a faint, ethereal aura. Though small, it seems to contain the weight of the entire universe. He stands tall above a rolling sea of pristine white clouds. The distant buildings of the Celestial Palace loom through the mist, the skyline illuminated by a brilliant golden glow. Li Jing's figure was as tall as a mountain, and with his steady breathing, his body swayed extremely slightly but powerfully. Every sway caused the crimson cloak behind him to rise and fall slowly like waves, while the pagoda in his hand was as steady as a rock, and the golden bell on the top of the pagoda did not move at all. This extreme contrast between movement and stillness, in the main colors of brilliant gold and holy white, created an epic atmosphere that was sacred, solemn and full of absolute power.", + "seed": 10946, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_0/index_6980_seed_10946_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.20004517007619144, + "vlm_logprob": -0.20004517007619144, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.20004517007619144, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05231770873069763 + ], + "opposite_logprob_tokens": [ + -3.0523176193237305 + ], + "logprob_reward": -0.05231770873069763, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05231770873069763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18447551131248474, + 0.04745631292462349, + 0.10070826858282089, + 0.05680019035935402, + 0.08295921981334686, + 0.06113841012120247, + 0.09849520772695541, + 0.07863045483827591, + 0.26062262058258057, + 0.02871381677687168 + ], + "credit_map": [ + 0.0333319827914238, + 0.018185297027230263, + 0.03484168276190758, + 0.017420625314116478, + 0.024860672652721405, + 0.016966581344604492, + 0.018730906769633293, + 0.02013774961233139, + 0.0077537670731544495, + 0.005593549460172653, + 0.008853236213326454, + 0.004847274627536535, + 0.0059315841645002365, + 0.004425407852977514, + 0.003993825986981392, + 0.006057663820683956, + 0.017935948446393013, + 0.015421492978930473, + 0.013284389860928059, + 0.01524826604872942, + 0.012888452969491482, + 0.010085751302540302, + 0.008148521184921265, + 0.007695441134274006, + 0.007578515447676182, + 0.006592610385268927, + 0.009690238162875175, + 0.010050608776509762, + 0.007359295152127743, + 0.005348467733711004, + 0.004766773898154497, + 0.005413676146417856, + 0.014996462501585484, + 0.005844031926244497, + 0.027168698608875275, + 0.007387904915958643, + 0.006773571949452162, + 0.00542503921315074, + 0.007485137786716223, + 0.00787836778908968, + 0.008554222993552685, + 0.00982733629643917, + 0.015217388048768044, + 0.004557046107947826, + 0.005401242058724165, + 0.004784959368407726, + 0.008342470973730087, + 0.004453742410987616, + 0.010677921585738659, + 0.010909006930887699, + 0.015889909118413925, + 0.006886697839945555, + 0.014949718490242958, + 0.011692496947944164, + 0.01973055861890316, + 0.0077588832937181, + 0.00798021163791418, + 0.009392558597028255, + 0.014442984014749527, + 0.012242994271218777, + 0.011337650939822197, + 0.0072825136594474316, + 0.010527982376515865, + 0.005423543509095907, + 0.023536065593361855, + 0.02714519575238228, + 0.028885627165436745, + 0.01926676742732525, + 0.05723341554403305, + 0.034043364226818085, + 0.0484505221247673, + 0.022061645984649658, + 0.0032583740539848804, + 0.0023449978325515985, + 0.006017088890075684, + 0.002189725637435913, + 0.0035438593477010727, + 0.003956458531320095, + 0.004680027719587088, + 0.0027232805732637644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.05231770873069763 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029852472245693207 + ], + "opposite_logprob_tokens": [ + -3.5298523902893066 + ], + "logprob_reward": -0.029852472245693207, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029852472245693207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18944714963436127, + 0.12813030183315277, + 0.045836903154850006, + 0.16962654888629913, + 0.08262517303228378, + 0.1396363228559494, + 0.08920915424823761, + 0.053169842809438705, + 0.05211605504155159, + 0.05020257458090782 + ], + "credit_map": [ + 0.029422316700220108, + 0.02118036150932312, + 0.03309835493564606, + 0.015239602886140347, + 0.020600024610757828, + 0.013252470642328262, + 0.025423675775527954, + 0.031230345368385315, + 0.022201301530003548, + 0.02276836521923542, + 0.016200624406337738, + 0.011142334900796413, + 0.01578673906624317, + 0.014668653719127178, + 0.01152857020497322, + 0.013833715580403805, + 0.006593747995793819, + 0.006602941546589136, + 0.00699929753318429, + 0.006860882975161076, + 0.005271519999951124, + 0.004234235268086195, + 0.004934840369969606, + 0.004339438397437334, + 0.023056941106915474, + 0.02040211483836174, + 0.034801699221134186, + 0.025133952498435974, + 0.012449797242879868, + 0.01766986772418022, + 0.01452632900327444, + 0.021585853770375252, + 0.011521965265274048, + 0.007036307360976934, + 0.01643865928053856, + 0.007046626880764961, + 0.00938405655324459, + 0.006845868192613125, + 0.014782987534999847, + 0.009568706154823303, + 0.0198694858700037, + 0.024940893054008484, + 0.029436493292450905, + 0.011877262033522129, + 0.01300781313329935, + 0.013082588091492653, + 0.013440368697047234, + 0.013981415890157223, + 0.012746896594762802, + 0.019794633612036705, + 0.011028872802853584, + 0.006788490805774927, + 0.011698966845870018, + 0.009323944337666035, + 0.011870655231177807, + 0.005956696812063456, + 0.0065954215824604034, + 0.005314034875482321, + 0.008239587768912315, + 0.003796124132350087, + 0.008329876698553562, + 0.008764458820223808, + 0.00682823034003377, + 0.005302107892930508, + 0.004589615389704704, + 0.004099923186004162, + 0.007552634458988905, + 0.004374384414404631, + 0.009916623122990131, + 0.007169907446950674, + 0.008293963968753815, + 0.006119001656770706, + 0.0066223000176250935, + 0.0037786609027534723, + 0.00908227264881134, + 0.003947453573346138, + 0.00785038061439991, + 0.007311705034226179, + 0.006607492454349995, + 0.005002310033887625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.029852472245693207 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.898335337638855 + ], + "opposite_logprob_tokens": [ + -0.523335337638855 + ], + "logprob_reward": -0.898335337638855, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.898335337638855, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13647234439849854, + 0.22574985027313232, + 0.1869402378797531, + 0.0800689086318016, + 0.09193848073482513, + 0.07194677740335464, + 0.10153445601463318, + 0.050854913890361786, + 0.025893593207001686, + 0.02860042080283165 + ], + "credit_map": [ + 0.01567731611430645, + 0.011433306150138378, + 0.01576152816414833, + 0.017717014998197556, + 0.02225581742823124, + 0.012549076229333878, + 0.022803673520684242, + 0.018274620175361633, + 0.046651922166347504, + 0.042279526591300964, + 0.03753307834267616, + 0.01988283358514309, + 0.01982935331761837, + 0.016011236235499382, + 0.014317917637526989, + 0.029243994504213333, + 0.03334059938788414, + 0.02685852162539959, + 0.027848385274410248, + 0.027064787223935127, + 0.018999863415956497, + 0.017504319548606873, + 0.014605115167796612, + 0.0207186508923769, + 0.011435425840318203, + 0.011853460222482681, + 0.015353120863437653, + 0.006026772316545248, + 0.011209921911358833, + 0.008114648051559925, + 0.006556082051247358, + 0.009519483894109726, + 0.01576368324458599, + 0.008636759594082832, + 0.027286125347018242, + 0.0072973622009158134, + 0.00765968719497323, + 0.009717784821987152, + 0.007186009548604488, + 0.008391075767576694, + 0.008630803786218166, + 0.005911009386181831, + 0.01679275371134281, + 0.0053671570494771, + 0.009751351550221443, + 0.011340892873704433, + 0.007578391581773758, + 0.006574422121047974, + 0.0167476087808609, + 0.007675653323531151, + 0.03243063762784004, + 0.006280429195612669, + 0.012101755477488041, + 0.009725380688905716, + 0.00867715384811163, + 0.00789584219455719, + 0.005233215633779764, + 0.005372089799493551, + 0.010158674791455269, + 0.003915111534297466, + 0.0061858706176280975, + 0.007420506794005632, + 0.008513553999364376, + 0.004055895376950502, + 0.0026978151872754097, + 0.0035194202791899443, + 0.0034886253997683525, + 0.002513344632461667, + 0.004468712955713272, + 0.0029155833180993795, + 0.004448839463293552, + 0.0018412545323371887, + 0.0038529469165951014, + 0.0020820714998990297, + 0.0074286689050495625, + 0.0019942820072174072, + 0.003613400273025036, + 0.0035510475281625986, + 0.0033054696395993233, + 0.002772536827251315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.898335337638855 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080400086939335 + ], + "opposite_logprob_tokens": [ + -4.511080265045166 + ], + "logprob_reward": -0.011080400086939335, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080400086939335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.168457493185997, + 0.08011475950479507, + 0.3208463490009308, + 0.038659609854221344, + 0.10382995754480362, + 0.04399002343416214, + 0.12903232872486115, + 0.037578266113996506, + 0.038096994161605835, + 0.03939416632056236 + ], + "credit_map": [ + 0.033486053347587585, + 0.01830429583787918, + 0.03855576738715172, + 0.013937324285507202, + 0.01939687691628933, + 0.012083588168025017, + 0.016111455857753754, + 0.016582133248448372, + 0.01768592558801174, + 0.018066542223095894, + 0.010334177874028683, + 0.006867196876555681, + 0.008509745821356773, + 0.00664128502830863, + 0.006316314451396465, + 0.005693572107702494, + 0.07013802975416183, + 0.056914277374744415, + 0.05818251520395279, + 0.05029268562793732, + 0.0222832802683115, + 0.017571095377206802, + 0.021734226495027542, + 0.023730263113975525, + 0.004918977152556181, + 0.004496623761951923, + 0.005052343010902405, + 0.0087586073204875, + 0.00434724148362875, + 0.0036596234422177076, + 0.005198092665523291, + 0.0022281005512923002, + 0.02008288912475109, + 0.007704656105488539, + 0.0357745960354805, + 0.008422421291470528, + 0.007455081213265657, + 0.00787174142897129, + 0.006028504576534033, + 0.01049006450921297, + 0.005505645647644997, + 0.005651930347084999, + 0.00639614462852478, + 0.003942075651139021, + 0.0057987128384411335, + 0.007156257051974535, + 0.0054077086970210075, + 0.004131550434976816, + 0.022195614874362946, + 0.020559044554829597, + 0.032846152782440186, + 0.008658062666654587, + 0.011269615031778812, + 0.013378823176026344, + 0.011122716590762138, + 0.009002294391393661, + 0.0038850451819598675, + 0.0028788892086595297, + 0.00620479229837656, + 0.0034363020677119493, + 0.005262789316475391, + 0.006239891052246094, + 0.005936177913099527, + 0.0037343772128224373, + 0.0034855909179896116, + 0.0034376184921711683, + 0.003997725900262594, + 0.003042069962248206, + 0.009676341898739338, + 0.00553155830129981, + 0.0054566748440265656, + 0.0034694119822233915, + 0.00451068626716733, + 0.002724951133131981, + 0.007990695536136627, + 0.0030909706838428974, + 0.0052735875360667706, + 0.005078458227217197, + 0.004012898076325655, + 0.006711915135383606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.011080400086939335 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639931678771973 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639931678771973, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639931678771973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.134476438164711, + 0.07823041081428528, + 0.12568238377571106, + 0.31941908597946167, + 0.055388130247592926, + 0.046545471996068954, + 0.05261623114347458, + 0.06056180223822594, + 0.06245390325784683, + 0.06462610512971878 + ], + "credit_map": [ + 0.020875651389360428, + 0.013423221185803413, + 0.027628019452095032, + 0.011729849502444267, + 0.015143479220569134, + 0.012524603866040707, + 0.016385529190301895, + 0.01676609180867672, + 0.017498210072517395, + 0.015422391705214977, + 0.010212966240942478, + 0.005685858894139528, + 0.006878535728901625, + 0.007060601841658354, + 0.006214745342731476, + 0.009257101453840733, + 0.021712245419621468, + 0.020230138674378395, + 0.014875895343720913, + 0.02156820520758629, + 0.016358554363250732, + 0.010997219942510128, + 0.011338737793266773, + 0.00860138051211834, + 0.04413002356886864, + 0.04075704514980316, + 0.06573432683944702, + 0.03891985863447189, + 0.026082530617713928, + 0.0353388674557209, + 0.02340264432132244, + 0.04505380243062973, + 0.010944903828203678, + 0.0048661804758012295, + 0.018452266231179237, + 0.004296139348298311, + 0.004728866741061211, + 0.0039677307941019535, + 0.0029013166204094887, + 0.005230722948908806, + 0.006171028129756451, + 0.005812432616949081, + 0.009680776856839657, + 0.0034283960703760386, + 0.007623717188835144, + 0.005464474204927683, + 0.0045504262670874596, + 0.0038142220582813025, + 0.008517409674823284, + 0.011980690062046051, + 0.007673303131014109, + 0.004077092744410038, + 0.00505007803440094, + 0.00464545376598835, + 0.007266008760780096, + 0.003406188217923045, + 0.007275196723639965, + 0.007414697203785181, + 0.015048612840473652, + 0.004814295098185539, + 0.005975323263555765, + 0.005824427120387554, + 0.008563179522752762, + 0.005646075122058392, + 0.007485927548259497, + 0.005767535418272018, + 0.01277048047631979, + 0.004589817486703396, + 0.008932771161198616, + 0.00980446394532919, + 0.007311645895242691, + 0.00579126225784421, + 0.008968590758740902, + 0.004836274776607752, + 0.018525047227740288, + 0.004495159722864628, + 0.005504776258021593, + 0.006606522016227245, + 0.005756191443651915, + 0.009933543391525745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.008639931678771973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, full-body shot, shot from a low, upward angle, captures the majestic presence of Li Jing, Marshal of the Celestial Palace. He clad in a sacred suit of armor meticulously crafted from gold and white jade, each plate engraved with intricate reliefs of auspicious clouds and mythical beasts. His shoulder armor is shaped like a majestic dragon head, and the breastplate on his chest gleams with a restrained yet powerful aura beneath the radiance of the heavens. His face is resolute, his long beard flowing, and he wears a purple-gold crown inlaid with glittering gems, exuding a commanding presence without even anger. In his left palm, he steadily holds a shimmering, seven-story pagoda, shrouded in a faint, ethereal aura. Though small, it seems to contain the weight of the entire universe. He stands tall above a rolling sea of pristine white clouds. The distant buildings of the Celestial Palace loom through the mist, the skyline illuminated by a brilliant golden glow. Li Jing's figure was as tall as a mountain, and with his steady breathing, his body swayed extremely slightly but powerfully. Every sway caused the crimson cloak behind him to rise and fall slowly like waves, while the pagoda in his hand was as steady as a rock, and the golden bell on the top of the pagoda did not move at all. This extreme contrast between movement and stillness, in the main colors of brilliant gold and holy white, created an epic atmosphere that was sacred, solemn and full of absolute power.", + "seed": 10947, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_0/index_6980_seed_10947_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.18239729963243007, + "vlm_logprob": -0.18239729963243007, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.18239729963243007, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0935891792178154 + ], + "opposite_logprob_tokens": [ + -2.4685890674591064 + ], + "logprob_reward": -0.0935891792178154, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0935891792178154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22919686138629913, + 0.12203125655651093, + 0.1773451566696167, + 0.05257619544863701, + 0.024923866614699364, + 0.034488171339035034, + 0.05151115730404854, + 0.1396973729133606, + 0.10001997649669647, + 0.06820998340845108 + ], + "credit_map": [ + 0.05417218431830406, + 0.023438364267349243, + 0.02933724783360958, + 0.023755569010972977, + 0.02610263228416443, + 0.03707604855298996, + 0.01953982375562191, + 0.015774989500641823, + 0.021776434034109116, + 0.01911350153386593, + 0.016734085977077484, + 0.009559542872011662, + 0.011435137130320072, + 0.020877880975604057, + 0.01383715309202671, + 0.008697513490915298, + 0.02656797505915165, + 0.024583084508776665, + 0.021999463438987732, + 0.015762558206915855, + 0.02301371842622757, + 0.023869598284363747, + 0.021759964525699615, + 0.019788803532719612, + 0.010587913915514946, + 0.005645061377435923, + 0.00587316881865263, + 0.005024912301450968, + 0.005265349056571722, + 0.00978231243789196, + 0.0066640921868383884, + 0.003733384422957897, + 0.0035538349766284227, + 0.002398229669779539, + 0.003200840437784791, + 0.002047165995463729, + 0.00409582769498229, + 0.002929627662524581, + 0.0038121859543025494, + 0.0028861539904028177, + 0.00561530003324151, + 0.004657011944800615, + 0.0035822789650410414, + 0.0036009219475090504, + 0.004333686549216509, + 0.0033537186682224274, + 0.00641254847869277, + 0.002932704286649823, + 0.007775753270834684, + 0.00959770753979683, + 0.00669049471616745, + 0.00484835309907794, + 0.006696173455566168, + 0.008109496906399727, + 0.0049004703760147095, + 0.002892704913392663, + 0.017732048407197, + 0.036208633333444595, + 0.021531805396080017, + 0.01842592842876911, + 0.01380944438278675, + 0.011813664808869362, + 0.011618699878454208, + 0.008557154797017574, + 0.016728468239307404, + 0.0187886580824852, + 0.011551186442375183, + 0.013200344517827034, + 0.008112472482025623, + 0.013039608485996723, + 0.011064275167882442, + 0.007534955628216267, + 0.01774284616112709, + 0.009920390322804451, + 0.008997971192002296, + 0.003910432569682598, + 0.004144696518778801, + 0.013784784823656082, + 0.0057442281395196915, + 0.003964630421251059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.0935891792178154 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983812429010868 + ], + "opposite_logprob_tokens": [ + -3.5298380851745605 + ], + "logprob_reward": -0.02983812429010868, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983812429010868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28783679008483887, + 0.10944709926843643, + 0.05371319130063057, + 0.05338524654507637, + 0.06504005938768387, + 0.06008230894804001, + 0.08509184420108795, + 0.10109500586986542, + 0.12676206231117249, + 0.05754636973142624 + ], + "credit_map": [ + 0.07058676332235336, + 0.027925776317715645, + 0.04571279510855675, + 0.023381058126688004, + 0.023803550750017166, + 0.05160214751958847, + 0.027531854808330536, + 0.017292868345975876, + 0.017072785645723343, + 0.01937328279018402, + 0.014481195248663425, + 0.012740067206323147, + 0.010288928635418415, + 0.014913257211446762, + 0.011104803532361984, + 0.009472774341702461, + 0.008290060795843601, + 0.005588022992014885, + 0.00688399700447917, + 0.005523944739252329, + 0.008228468708693981, + 0.007409844547510147, + 0.006680631544440985, + 0.00510821770876646, + 0.010420748963952065, + 0.0054958779364824295, + 0.0064222365617752075, + 0.0051568653434515, + 0.004773297812789679, + 0.008414814248681068, + 0.008577116765081882, + 0.004124291241168976, + 0.0075092012993991375, + 0.00881373230367899, + 0.009156533516943455, + 0.004775551613420248, + 0.007462569046765566, + 0.007948020473122597, + 0.011197496205568314, + 0.008176951669156551, + 0.011429432779550552, + 0.007059486582875252, + 0.009386211633682251, + 0.0037838148418813944, + 0.0067843105643987656, + 0.006887966766953468, + 0.010441175661981106, + 0.004309904295951128, + 0.017634889110922813, + 0.02164113149046898, + 0.015669535845518112, + 0.006609415169805288, + 0.005625368561595678, + 0.008740408346056938, + 0.005496212746948004, + 0.003674878505989909, + 0.022850489243865013, + 0.013386455364525318, + 0.014198675751686096, + 0.009826713241636753, + 0.009572455659508705, + 0.014369525946676731, + 0.01194900181144476, + 0.004941695835441351, + 0.02427181787788868, + 0.018208390101790428, + 0.014253159053623676, + 0.017846662551164627, + 0.01111373770982027, + 0.016115941107273102, + 0.01398475468158722, + 0.010967598296701908, + 0.010203419253230095, + 0.007969017140567303, + 0.006758494768291712, + 0.0056273010559380054, + 0.006089323200285435, + 0.007858731783926487, + 0.008596844039857388, + 0.004443235229700804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.02983812429010868 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578300833702087 + ], + "opposite_logprob_tokens": [ + -0.6328300833702087 + ], + "logprob_reward": -0.7578300833702087, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578300833702087, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1533314287662506, + 0.05703689903020859, + 0.0458635650575161, + 0.08527421951293945, + 0.06515202671289444, + 0.05811489373445511, + 0.0446319542825222, + 0.1689557135105133, + 0.28427278995513916, + 0.03736650571227074 + ], + "credit_map": [ + 0.040183089673519135, + 0.01611867919564247, + 0.019221141934394836, + 0.012190012261271477, + 0.013928225263953209, + 0.029587892815470695, + 0.009962145239114761, + 0.012140237726271152, + 0.007627537474036217, + 0.006744359154254198, + 0.006912330165505409, + 0.006603258196264505, + 0.0074808429926633835, + 0.006284020841121674, + 0.008637560531497002, + 0.006746992003172636, + 0.010649068281054497, + 0.005376611370593309, + 0.005720896180719137, + 0.004209455102682114, + 0.005158849526196718, + 0.004641473293304443, + 0.004590838216245174, + 0.005516374483704567, + 0.01869995705783367, + 0.009002677164971828, + 0.009712161496281624, + 0.007827573455870152, + 0.010468662716448307, + 0.01373174600303173, + 0.008763347752392292, + 0.007068093400448561, + 0.01555549819022417, + 0.010112974792718887, + 0.00967799685895443, + 0.003853654721751809, + 0.005406551528722048, + 0.010091041214764118, + 0.007045025005936623, + 0.003409280441701412, + 0.012886609882116318, + 0.011257363483309746, + 0.010289925150573254, + 0.0033177812583744526, + 0.005128820426762104, + 0.008335285820066929, + 0.00394417904317379, + 0.0029549277387559414, + 0.008387681096792221, + 0.007937361486256123, + 0.007067916914820671, + 0.006299342028796673, + 0.003995729144662619, + 0.004685717169195414, + 0.0041302903555333614, + 0.0021279146894812584, + 0.029192674905061722, + 0.03979463875293732, + 0.02392628602683544, + 0.02023797668516636, + 0.013178624212741852, + 0.01778462529182434, + 0.015526678413152695, + 0.009314225986599922, + 0.06055780500173569, + 0.03506406396627426, + 0.03611208125948906, + 0.03268178552389145, + 0.021697238087654114, + 0.051131680607795715, + 0.0291032362729311, + 0.017924921587109566, + 0.006270634476095438, + 0.005054462235420942, + 0.004400772042572498, + 0.004623076878488064, + 0.004791514482349157, + 0.004557815380394459, + 0.0039506773464381695, + 0.0037175535690039396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.7578300833702087 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819116808474064 + ], + "opposite_logprob_tokens": [ + -4.018191337585449 + ], + "logprob_reward": -0.01819116808474064, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819116808474064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25941070914268494, + 0.037303660064935684, + 0.03689614310860634, + 0.060346633195877075, + 0.0642734095454216, + 0.04620711877942085, + 0.09200403839349747, + 0.10633835196495056, + 0.22624659538269043, + 0.07097338885068893 + ], + "credit_map": [ + 0.07475893199443817, + 0.02779858559370041, + 0.03726262226700783, + 0.017398515716195107, + 0.02270406112074852, + 0.044919200241565704, + 0.017920546233654022, + 0.01664823107421398, + 0.0060165077447891235, + 0.005252791102975607, + 0.004622968379408121, + 0.0037734992802143097, + 0.004697077441960573, + 0.005350086838006973, + 0.004202319774776697, + 0.003388406476005912, + 0.006019029300659895, + 0.007105963304638863, + 0.004885445348918438, + 0.0034579180646687746, + 0.005427956115454435, + 0.002999301301315427, + 0.004102111328393221, + 0.0028984190430492163, + 0.012744287960231304, + 0.009270229376852512, + 0.007434680126607418, + 0.005460516549646854, + 0.005329018924385309, + 0.010543323121964931, + 0.005834931507706642, + 0.0037296428345143795, + 0.008669751696288586, + 0.00623354921117425, + 0.00709889829158783, + 0.004362770356237888, + 0.011761452071368694, + 0.0071904947981238365, + 0.01354029681533575, + 0.005416198633611202, + 0.009045278653502464, + 0.007941599003970623, + 0.007125109899789095, + 0.003260411322116852, + 0.005119761917740107, + 0.0058713327161967754, + 0.004999697208404541, + 0.002843927824869752, + 0.011755102314054966, + 0.018891440704464912, + 0.013354037888348103, + 0.012237547896802425, + 0.00818106159567833, + 0.011171374469995499, + 0.011179886758327484, + 0.005233582574874163, + 0.016148436814546585, + 0.016945648938417435, + 0.013429450802505016, + 0.012303369119763374, + 0.013136330060660839, + 0.012460581958293915, + 0.015069706365466118, + 0.006844830699265003, + 0.04734523966908455, + 0.028375709429383278, + 0.028683293610811234, + 0.02472427859902382, + 0.016704926267266273, + 0.04414590448141098, + 0.024565232917666435, + 0.011701994575560093, + 0.011506356298923492, + 0.01277173962444067, + 0.007853558287024498, + 0.006081012077629566, + 0.009726367890834808, + 0.007914301007986069, + 0.011161811649799347, + 0.003958241082727909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.01819116808474064 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537943199276924 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012537943199276924, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537943199276924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43585771322250366, + 0.1287829726934433, + 0.10014194250106812, + 0.02975037507712841, + 0.033282019197940826, + 0.04583454132080078, + 0.05192667618393898, + 0.05817720293998718, + 0.059007324278354645, + 0.05723923444747925 + ], + "credit_map": [ + 0.1277914196252823, + 0.05068214610219002, + 0.0642271414399147, + 0.031130757182836533, + 0.03161943331360817, + 0.07881715148687363, + 0.027941983193159103, + 0.023647725582122803, + 0.022966524586081505, + 0.026282213628292084, + 0.016577478498220444, + 0.01222391240298748, + 0.013333580456674099, + 0.015868796035647392, + 0.013547854498028755, + 0.007982601411640644, + 0.026565829291939735, + 0.011035961098968983, + 0.015611819922924042, + 0.00794785376638174, + 0.008805236779153347, + 0.014912032522261143, + 0.008473686873912811, + 0.0067895096726715565, + 0.004225582350045443, + 0.004066282417625189, + 0.0032846375834196806, + 0.004870910197496414, + 0.003262103069573641, + 0.0032915507908910513, + 0.003630720777437091, + 0.0031185864936560392, + 0.0063690077513456345, + 0.00592547282576561, + 0.004988393746316433, + 0.0027130937669426203, + 0.0037726685404777527, + 0.004307000432163477, + 0.003584587248042226, + 0.0016217946540564299, + 0.01357702724635601, + 0.0047243558801710606, + 0.007639351766556501, + 0.003010423621162772, + 0.0038685998879373074, + 0.007908872328698635, + 0.0022371027152985334, + 0.002868807874619961, + 0.005849980749189854, + 0.016141721978783607, + 0.009738554246723652, + 0.00613592891022563, + 0.003911452367901802, + 0.0031912208069115877, + 0.004213966894894838, + 0.002743851626291871, + 0.013548273593187332, + 0.0068444013595581055, + 0.007979176938533783, + 0.005374904256314039, + 0.004845859482884407, + 0.007711262907832861, + 0.00847605150192976, + 0.0033972756937146187, + 0.011341537348926067, + 0.006856309249997139, + 0.006817067973315716, + 0.006915092933923006, + 0.006420252379029989, + 0.010054200887680054, + 0.006885443348437548, + 0.003717422252520919, + 0.015663713216781616, + 0.0067370557226240635, + 0.008298021741211414, + 0.0032303091138601303, + 0.004004505928605795, + 0.012017828412353992, + 0.0036622860934585333, + 0.0036255090963095427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.012537943199276924 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, full-body shot, shot from a low, upward angle, captures the majestic presence of Li Jing, Marshal of the Celestial Palace. He clad in a sacred suit of armor meticulously crafted from gold and white jade, each plate engraved with intricate reliefs of auspicious clouds and mythical beasts. His shoulder armor is shaped like a majestic dragon head, and the breastplate on his chest gleams with a restrained yet powerful aura beneath the radiance of the heavens. His face is resolute, his long beard flowing, and he wears a purple-gold crown inlaid with glittering gems, exuding a commanding presence without even anger. In his left palm, he steadily holds a shimmering, seven-story pagoda, shrouded in a faint, ethereal aura. Though small, it seems to contain the weight of the entire universe. He stands tall above a rolling sea of pristine white clouds. The distant buildings of the Celestial Palace loom through the mist, the skyline illuminated by a brilliant golden glow. Li Jing's figure was as tall as a mountain, and with his steady breathing, his body swayed extremely slightly but powerfully. Every sway caused the crimson cloak behind him to rise and fall slowly like waves, while the pagoda in his hand was as steady as a rock, and the golden bell on the top of the pagoda did not move at all. This extreme contrast between movement and stillness, in the main colors of brilliant gold and holy white, created an epic atmosphere that was sacred, solemn and full of absolute power.", + "seed": 10948, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_0/index_6980_seed_10948_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21886960752308368, + "vlm_logprob": -0.21886960752308368, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.21886960752308368, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35225674510002136 + ], + "opposite_logprob_tokens": [ + -1.2272567749023438 + ], + "logprob_reward": -0.35225674510002136, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35225674510002136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37287580966949463, + 0.13739079236984253, + 0.04425127059221268, + 0.0978594720363617, + 0.06235993653535843, + 0.10665518790483475, + 0.06816145032644272, + 0.037161003798246384, + 0.034675851464271545, + 0.038609277456998825 + ], + "credit_map": [ + 0.039693452417850494, + 0.04100583866238594, + 0.05203857645392418, + 0.03978809341788292, + 0.05137757956981659, + 0.0587301179766655, + 0.04746566340327263, + 0.04277645796537399, + 0.013778764754533768, + 0.012657158076763153, + 0.021610982716083527, + 0.010721128433942795, + 0.015540779568254948, + 0.026735670864582062, + 0.01597578264772892, + 0.02037050575017929, + 0.006027416326105595, + 0.005400445312261581, + 0.0056931874714791775, + 0.005163609981536865, + 0.005537175107747316, + 0.003749139839783311, + 0.00852961465716362, + 0.004150678403675556, + 0.019377732649445534, + 0.011986262165009975, + 0.016244778409600258, + 0.007312621921300888, + 0.008761228062212467, + 0.01406374666839838, + 0.011749697849154472, + 0.008363385684788227, + 0.008717546239495277, + 0.006337547674775124, + 0.011320331133902073, + 0.004855723585933447, + 0.00828499998897314, + 0.009911363944411278, + 0.007270741276443005, + 0.005661673843860626, + 0.012850292026996613, + 0.02147725410759449, + 0.018514949828386307, + 0.00863273162394762, + 0.010152967646718025, + 0.016980251297354698, + 0.011107509024441242, + 0.006939223501831293, + 0.01016860269010067, + 0.006173914764076471, + 0.011728302575647831, + 0.005283254198729992, + 0.004394365940243006, + 0.014648017473518848, + 0.00580496434122324, + 0.009960019960999489, + 0.005088907200843096, + 0.0035715617705136538, + 0.005466490983963013, + 0.003916952759027481, + 0.00410047359764576, + 0.00554687762632966, + 0.004756572190672159, + 0.004713159054517746, + 0.0038164984434843063, + 0.003432607278227806, + 0.004573412239551544, + 0.003517520846799016, + 0.004723024554550648, + 0.004691688809543848, + 0.005361609160900116, + 0.004559484776109457, + 0.004522419534623623, + 0.004471698775887489, + 0.004832842852920294, + 0.004081365652382374, + 0.003203370375558734, + 0.004747460130602121, + 0.008876917883753777, + 0.003873197827488184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.35225674510002136 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026393335312604904 + ], + "opposite_logprob_tokens": [ + -3.651393413543701 + ], + "logprob_reward": -0.026393335312604904, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026393335312604904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3108922243118286, + 0.11046841740608215, + 0.10487950593233109, + 0.07543273270130157, + 0.0690762996673584, + 0.0555080883204937, + 0.04729604721069336, + 0.023873789235949516, + 0.035198234021663666, + 0.1673746407032013 + ], + "credit_map": [ + 0.04180726036429405, + 0.03898290544748306, + 0.044293586164712906, + 0.028573980554938316, + 0.028529874980449677, + 0.05135960504412651, + 0.033605411648750305, + 0.043739620596170425, + 0.012360243126749992, + 0.0140597615391016, + 0.013707132078707218, + 0.01316868606954813, + 0.012121730484068394, + 0.012453551404178143, + 0.018384970724582672, + 0.014212331734597683, + 0.010115577839314938, + 0.01146659068763256, + 0.014459278434515, + 0.01009540818631649, + 0.010825049132108688, + 0.016856616362929344, + 0.021826384589076042, + 0.009234590455889702, + 0.013796131126582623, + 0.008046034723520279, + 0.012052968144416809, + 0.0058569577522575855, + 0.005420623812824488, + 0.01300105545669794, + 0.008983890525996685, + 0.008275067433714867, + 0.0064968084916472435, + 0.005446864292025566, + 0.010441454127430916, + 0.007733624894171953, + 0.00855327770113945, + 0.011179065331816673, + 0.009273290634155273, + 0.009951910004019737, + 0.007721239700913429, + 0.00649003591388464, + 0.009405185468494892, + 0.0041589331813156605, + 0.005300143267959356, + 0.011398774571716785, + 0.00555271003395319, + 0.005481070838868618, + 0.004726351238787174, + 0.004057776648551226, + 0.005925403442233801, + 0.005616733804345131, + 0.006647391710430384, + 0.004614520817995071, + 0.008325744420289993, + 0.007382123731076717, + 0.0039544119499623775, + 0.0025026905350387096, + 0.003168099094182253, + 0.0023372999858111143, + 0.003474835306406021, + 0.0033927748445421457, + 0.0029412724543362856, + 0.002102404134348035, + 0.006623697001487017, + 0.0042513455264270306, + 0.005843336693942547, + 0.0027783047407865524, + 0.004368607886135578, + 0.0058406684547662735, + 0.0025936306919902563, + 0.0028986409306526184, + 0.027784550562500954, + 0.016008740290999413, + 0.027557319030165672, + 0.01019326876848936, + 0.011140737682580948, + 0.0368996225297451, + 0.01948462799191475, + 0.018305771052837372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.026393335312604904 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933720111846924 + ], + "opposite_logprob_tokens": [ + -0.6933720111846924 + ], + "logprob_reward": -0.6933720111846924, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933720111846924, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.23347367346286774, + 0.06201249733567238, + 0.08905346691608429, + 0.04914279282093048, + 0.09722267091274261, + 0.024376535788178444, + 0.041538067162036896, + 0.11479933559894562, + 0.09743444621562958, + 0.19094645977020264 + ], + "credit_map": [ + 0.037943851202726364, + 0.025424329563975334, + 0.03141218051314354, + 0.024332035332918167, + 0.01889210380613804, + 0.03036952205002308, + 0.02013002149760723, + 0.04496964439749718, + 0.008103969506919384, + 0.006738354451954365, + 0.009586836211383343, + 0.004136096220463514, + 0.006084896624088287, + 0.011511982418596745, + 0.007629110477864742, + 0.008221253752708435, + 0.01120662596076727, + 0.008926336653530598, + 0.01118157897144556, + 0.011051961220800877, + 0.012026905082166195, + 0.012373706325888634, + 0.0143821956589818, + 0.007904169149696827, + 0.004862120375037193, + 0.004521581344306469, + 0.007172660902142525, + 0.00603095768019557, + 0.006672265008091927, + 0.00798063538968563, + 0.006247760262340307, + 0.005654816050082445, + 0.01331796869635582, + 0.0072695594280958176, + 0.01617981307208538, + 0.007961686700582504, + 0.013030223548412323, + 0.0156828835606575, + 0.010006832890212536, + 0.013773704878985882, + 0.004264011513441801, + 0.0034674277994781733, + 0.003905008314177394, + 0.0026297152508050203, + 0.002189558232203126, + 0.002909824252128601, + 0.0028802005108445883, + 0.0021307896822690964, + 0.004481000825762749, + 0.0036767260171473026, + 0.005432995967566967, + 0.0039933230727910995, + 0.004110674373805523, + 0.005195773672312498, + 0.005781356245279312, + 0.008866220712661743, + 0.017776325345039368, + 0.013089065439999104, + 0.018857842311263084, + 0.008724591694772243, + 0.011455447413027287, + 0.019906459376215935, + 0.011344971135258675, + 0.013644644990563393, + 0.0069271656684577465, + 0.009070072323083878, + 0.012029621750116348, + 0.01133759692311287, + 0.017108451575040817, + 0.010790356434881687, + 0.019200008362531662, + 0.010971187613904476, + 0.030411606654524803, + 0.017534123733639717, + 0.03213634714484215, + 0.009508011862635612, + 0.015564622357487679, + 0.04483591020107269, + 0.018181918188929558, + 0.022773930802941322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.6933720111846924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543476186692715 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543476186692715, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543476186692715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26592034101486206, + 0.06152649223804474, + 0.054170046001672745, + 0.04395702853798866, + 0.16767890751361847, + 0.04399310052394867, + 0.05073396861553192, + 0.06715729832649231, + 0.08594772219657898, + 0.15891516208648682 + ], + "credit_map": [ + 0.041837941855192184, + 0.03194403648376465, + 0.04564272612333298, + 0.02247917279601097, + 0.018590552732348442, + 0.05337972939014435, + 0.020817328244447708, + 0.03122885152697563, + 0.008016695268452168, + 0.006904543377459049, + 0.009304271079599857, + 0.006851469166576862, + 0.008797607384622097, + 0.00642735417932272, + 0.008704330772161484, + 0.006520214956253767, + 0.007383477408438921, + 0.01148688793182373, + 0.008625168353319168, + 0.005760008003562689, + 0.00485644768923521, + 0.006058456841856241, + 0.005784918554127216, + 0.004214680287986994, + 0.007544003892689943, + 0.0053559294901788235, + 0.007426686119288206, + 0.0031899602618068457, + 0.003188934875652194, + 0.006464883219450712, + 0.006446009501814842, + 0.004340618848800659, + 0.015078998170793056, + 0.016660651192069054, + 0.023552393540740013, + 0.02467142418026924, + 0.019766725599765778, + 0.022405939176678658, + 0.021687813103199005, + 0.023854978382587433, + 0.007061815354973078, + 0.008310962468385696, + 0.00726902624592185, + 0.0037419393192976713, + 0.0047800615429878235, + 0.005748830735683441, + 0.003537080716341734, + 0.0035433806478977203, + 0.006524627096951008, + 0.004417126998305321, + 0.007214774843305349, + 0.004444909747689962, + 0.004971886985003948, + 0.0063851140439510345, + 0.006400873884558678, + 0.01037465687841177, + 0.009109331294894218, + 0.0069789276458323, + 0.01010486576706171, + 0.0073389820754528046, + 0.007903463207185268, + 0.008973226882517338, + 0.009641025215387344, + 0.007107481826096773, + 0.009350812062621117, + 0.008025693707168102, + 0.011277691461145878, + 0.009331080131232738, + 0.01292334869503975, + 0.00885056797415018, + 0.015475613996386528, + 0.010712913237512112, + 0.020190509036183357, + 0.015327252447605133, + 0.02163281850516796, + 0.012308984994888306, + 0.0191758144646883, + 0.026431424543261528, + 0.02785380370914936, + 0.01599455438554287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.012543476186692715 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978246983140707 + ], + "opposite_logprob_tokens": [ + -4.634782314300537 + ], + "logprob_reward": -0.00978246983140707, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978246983140707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33106186985969543, + 0.12506061792373657, + 0.06913594901561737, + 0.04229271039366722, + 0.033517513424158096, + 0.05273474380373955, + 0.13107596337795258, + 0.10497697442770004, + 0.06382307410240173, + 0.04632055014371872 + ], + "credit_map": [ + 0.037819135934114456, + 0.021007493138313293, + 0.0528964027762413, + 0.02939811907708645, + 0.032762832939624786, + 0.07413595169782639, + 0.03944294527173042, + 0.04359901323914528, + 0.014739789068698883, + 0.010207090526819229, + 0.02171696163713932, + 0.00871361419558525, + 0.010099345818161964, + 0.028902357444167137, + 0.012788888067007065, + 0.017892563715577126, + 0.008332137949764729, + 0.006879555061459541, + 0.012560447677969933, + 0.004906026180833578, + 0.005840286146849394, + 0.01708952710032463, + 0.0062906560488045216, + 0.007237310521304607, + 0.005108710844069719, + 0.0038763738702982664, + 0.0066960155963897705, + 0.0035622427240014076, + 0.005857832729816437, + 0.007555900141596794, + 0.005649145692586899, + 0.003986488562077284, + 0.003377238055691123, + 0.0026074834167957306, + 0.004907806869596243, + 0.003442140528932214, + 0.0059018549509346485, + 0.00379896885715425, + 0.004677474964410067, + 0.004804546013474464, + 0.009421160444617271, + 0.0073951007798314095, + 0.009138145484030247, + 0.0042809476144611835, + 0.005560683086514473, + 0.007690810598433018, + 0.005142829846590757, + 0.004105068743228912, + 0.022135483101010323, + 0.011431400664150715, + 0.023166684433817863, + 0.011273711919784546, + 0.009591708891093731, + 0.029003584757447243, + 0.00902235135436058, + 0.015451047569513321, + 0.017312034964561462, + 0.009142884984612465, + 0.01535199023783207, + 0.00907072238624096, + 0.010319236665964127, + 0.02186785638332367, + 0.010865160264074802, + 0.011047083884477615, + 0.01419998798519373, + 0.007524731568992138, + 0.010235118679702282, + 0.005909733008593321, + 0.006696342024952173, + 0.008681437000632286, + 0.005616927985101938, + 0.004958801902830601, + 0.006044549401849508, + 0.0038834773004055023, + 0.00553444167599082, + 0.004266715608537197, + 0.005621068179607391, + 0.005996262654662132, + 0.00998564250767231, + 0.004988392814993858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.00978246983140707 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, full-body shot, shot from a low, upward angle, captures the majestic presence of Li Jing, Marshal of the Celestial Palace. He clad in a sacred suit of armor meticulously crafted from gold and white jade, each plate engraved with intricate reliefs of auspicious clouds and mythical beasts. His shoulder armor is shaped like a majestic dragon head, and the breastplate on his chest gleams with a restrained yet powerful aura beneath the radiance of the heavens. His face is resolute, his long beard flowing, and he wears a purple-gold crown inlaid with glittering gems, exuding a commanding presence without even anger. In his left palm, he steadily holds a shimmering, seven-story pagoda, shrouded in a faint, ethereal aura. Though small, it seems to contain the weight of the entire universe. He stands tall above a rolling sea of pristine white clouds. The distant buildings of the Celestial Palace loom through the mist, the skyline illuminated by a brilliant golden glow. Li Jing's figure was as tall as a mountain, and with his steady breathing, his body swayed extremely slightly but powerfully. Every sway caused the crimson cloak behind him to rise and fall slowly like waves, while the pagoda in his hand was as steady as a rock, and the golden bell on the top of the pagoda did not move at all. This extreme contrast between movement and stillness, in the main colors of brilliant gold and holy white, created an epic atmosphere that was sacred, solemn and full of absolute power.", + "seed": 10949, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_0/index_6980_seed_10949_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.19630097281187772, + "vlm_logprob": -0.19630097281187772, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.19630097281187772, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35460659861564636 + ], + "opposite_logprob_tokens": [ + -1.2296066284179688 + ], + "logprob_reward": -0.35460659861564636, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35460659861564636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3716568946838379, + 0.07493992149829865, + 0.03148951381444931, + 0.05736134201288223, + 0.023663807660341263, + 0.07894233614206314, + 0.04060807451605797, + 0.14669419825077057, + 0.08799652755260468, + 0.08664733916521072 + ], + "credit_map": [ + 0.04782281816005707, + 0.04171194136142731, + 0.055069297552108765, + 0.05144815146923065, + 0.0457974448800087, + 0.05806180089712143, + 0.04523485153913498, + 0.02651059627532959, + 0.010292010381817818, + 0.00763446232303977, + 0.013108965009450912, + 0.0092941178008914, + 0.007420075125992298, + 0.012368733063340187, + 0.007818976417183876, + 0.007002584170550108, + 0.0033423062413930893, + 0.004659332800656557, + 0.004576859995722771, + 0.006251891143620014, + 0.0028117960318922997, + 0.00282286130823195, + 0.004424723796546459, + 0.0025997415650635958, + 0.007827757857739925, + 0.008104685693979263, + 0.006617621518671513, + 0.007306412328034639, + 0.007626718375831842, + 0.010357155464589596, + 0.004927346482872963, + 0.004593643359839916, + 0.003142745466902852, + 0.002873252145946026, + 0.0025420801248401403, + 0.0034393095411360264, + 0.0034451959654688835, + 0.0018338793888688087, + 0.004031590186059475, + 0.002355752745643258, + 0.00986429862678051, + 0.010659725405275822, + 0.01522777508944273, + 0.011168736964464188, + 0.009392812848091125, + 0.008082479238510132, + 0.00767140556126833, + 0.00687510846182704, + 0.00502817053347826, + 0.005985615309327841, + 0.0031671146862208843, + 0.0052300929091870785, + 0.0054022702388465405, + 0.004666063003242016, + 0.00729140592738986, + 0.0038373437710106373, + 0.0172075517475605, + 0.016712166368961334, + 0.015555922873318195, + 0.019386377185583115, + 0.023605696856975555, + 0.01720397174358368, + 0.024996936321258545, + 0.01202556025236845, + 0.022444216534495354, + 0.005388650111854076, + 0.004807304125279188, + 0.006508483085781336, + 0.008205144666135311, + 0.027047671377658844, + 0.008597600273787975, + 0.0049974615685641766, + 0.014661879278719425, + 0.006932605989277363, + 0.007296018302440643, + 0.007733880542218685, + 0.01155055034905672, + 0.02070595696568489, + 0.010474459268152714, + 0.007291985210031271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.35460659861564636 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029869595542550087 + ], + "opposite_logprob_tokens": [ + -3.529869556427002 + ], + "logprob_reward": -0.029869595542550087, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029869595542550087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2756301760673523, + 0.08701957762241364, + 0.04687289521098137, + 0.03450939059257507, + 0.05359850823879242, + 0.045315712690353394, + 0.0723189115524292, + 0.14043158292770386, + 0.043468207120895386, + 0.20083501935005188 + ], + "credit_map": [ + 0.03941016644239426, + 0.030691951513290405, + 0.0448506735265255, + 0.037591252475976944, + 0.026329414919018745, + 0.047281209379434586, + 0.029878195375204086, + 0.019597316160798073, + 0.010730782523751259, + 0.011866564862430096, + 0.01739959977567196, + 0.014356685802340508, + 0.008861812762916088, + 0.009715926833450794, + 0.007411925587803125, + 0.006676292512565851, + 0.00573668023571372, + 0.006962306797504425, + 0.005580569617450237, + 0.007464672904461622, + 0.007116385735571384, + 0.005477637518197298, + 0.005964267998933792, + 0.002570376032963395, + 0.003995525185018778, + 0.004162410274147987, + 0.0039096432738006115, + 0.0050787352956831455, + 0.004417513031512499, + 0.005046003498136997, + 0.005412712227553129, + 0.0024868508335202932, + 0.006258935201913118, + 0.005586280487477779, + 0.0054775686003267765, + 0.0073202247731387615, + 0.00861659087240696, + 0.007626884151250124, + 0.007189264986664057, + 0.005522764753550291, + 0.005826939363032579, + 0.005163463763892651, + 0.00635152030736208, + 0.004478478338569403, + 0.005562149919569492, + 0.005078915040940046, + 0.007407873868942261, + 0.0054463730193674564, + 0.01016937755048275, + 0.010672158561646938, + 0.010400430299341679, + 0.009533373638987541, + 0.008303279988467693, + 0.010927080176770687, + 0.006044273730367422, + 0.0062689343467354774, + 0.017877908423542976, + 0.01610509306192398, + 0.01428945828229189, + 0.019575104117393494, + 0.020010191947221756, + 0.0190215352922678, + 0.022103313356637955, + 0.011448992416262627, + 0.006308848038315773, + 0.0059553878381848335, + 0.003712993348017335, + 0.005914533976465464, + 0.004062265157699585, + 0.007777400314807892, + 0.003893507644534111, + 0.005843272432684898, + 0.060531165450811386, + 0.011315767653286457, + 0.00997224822640419, + 0.008354024030268192, + 0.016949675977230072, + 0.07256952673196793, + 0.014303496107459068, + 0.0068391370587050915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.029869595542550087 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761662721633911 + ], + "opposite_logprob_tokens": [ + -0.8261662721633911 + ], + "logprob_reward": -0.5761662721633911, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761662721633911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10643074661493301, + 0.2054928094148636, + 0.043546538800001144, + 0.06401649117469788, + 0.1219547837972641, + 0.13746017217636108, + 0.020195437595248222, + 0.07587026804685593, + 0.17683763802051544, + 0.04819505289196968 + ], + "credit_map": [ + 0.013883726671338081, + 0.01228613592684269, + 0.020062007009983063, + 0.015223801136016846, + 0.009458836168050766, + 0.014387348666787148, + 0.011373416520655155, + 0.009755480103194714, + 0.027481725439429283, + 0.027676627039909363, + 0.03762960806488991, + 0.04535645991563797, + 0.013100532814860344, + 0.02033289149403572, + 0.015184100717306137, + 0.018730880692601204, + 0.005311062093824148, + 0.005487152840942144, + 0.006430129520595074, + 0.007413152605295181, + 0.0044833216816186905, + 0.00457990076392889, + 0.005077296867966652, + 0.004764521028846502, + 0.007698076777160168, + 0.008105247281491756, + 0.009740381501615047, + 0.009995302185416222, + 0.00906350091099739, + 0.006555103696882725, + 0.008132549934089184, + 0.004726330284029245, + 0.01697762869298458, + 0.016062211245298386, + 0.015330127440392971, + 0.011393880471587181, + 0.01651052199304104, + 0.01800376921892166, + 0.016348998993635178, + 0.011327659711241722, + 0.015830056741833687, + 0.017439264804124832, + 0.020845916122198105, + 0.014830094762146473, + 0.022165440022945404, + 0.012400161474943161, + 0.022982515394687653, + 0.010966732166707516, + 0.0027238510083407164, + 0.0030655357986688614, + 0.002009039046242833, + 0.0029731914401054382, + 0.002299484796822071, + 0.0026887748390436172, + 0.002598412102088332, + 0.001837151707150042, + 0.01126087550073862, + 0.0064825392328202724, + 0.0067293643951416016, + 0.009869316592812538, + 0.014324034564197063, + 0.012386264279484749, + 0.008808230981230736, + 0.006009650882333517, + 0.04215020313858986, + 0.011698651127517223, + 0.013380617834627628, + 0.012181829661130905, + 0.022045617923140526, + 0.048959970474243164, + 0.01648900844156742, + 0.009931757114827633, + 0.010711612179875374, + 0.0033785810228437185, + 0.004658905323594809, + 0.003617836395278573, + 0.005019693169742823, + 0.013612763024866581, + 0.004155540373176336, + 0.003040123265236616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 5 + ], + "selected_reward": -0.5761662721633911 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080872267484665 + ], + "opposite_logprob_tokens": [ + -4.511080741882324 + ], + "logprob_reward": -0.011080872267484665, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080872267484665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1676914244890213, + 0.15412767231464386, + 0.06835287064313889, + 0.06068754196166992, + 0.08143085241317749, + 0.11674878001213074, + 0.07847262173891068, + 0.105453722178936, + 0.08852704614400864, + 0.07850746065378189 + ], + "credit_map": [ + 0.023617226630449295, + 0.020513135939836502, + 0.027028527110815048, + 0.019187429919838905, + 0.01805979572236538, + 0.023479461669921875, + 0.019939394667744637, + 0.015866447240114212, + 0.02078479528427124, + 0.014383411034941673, + 0.029642770066857338, + 0.02113032527267933, + 0.015352561138570309, + 0.017725447192788124, + 0.0137455640360713, + 0.021362803876399994, + 0.008696305565536022, + 0.009860754944384098, + 0.01241360418498516, + 0.011446530930697918, + 0.004872298799455166, + 0.007667898666113615, + 0.006694980897009373, + 0.00670049199834466, + 0.0083177350461483, + 0.009209747426211834, + 0.005783325992524624, + 0.007580309174954891, + 0.00901450403034687, + 0.009526696056127548, + 0.005378459580242634, + 0.005876766052097082, + 0.011286052875220776, + 0.011529562063515186, + 0.010509978048503399, + 0.013181662186980247, + 0.008926301263272762, + 0.009239780716598034, + 0.00958385318517685, + 0.007173662073910236, + 0.014053155668079853, + 0.017603151500225067, + 0.01743004098534584, + 0.018067967146635056, + 0.01322821993380785, + 0.010148291476070881, + 0.01599128171801567, + 0.010226670652627945, + 0.010713250376284122, + 0.011281043291091919, + 0.012092223390936852, + 0.010954072698950768, + 0.00844029150903225, + 0.008811788633465767, + 0.009311584755778313, + 0.00686836801469326, + 0.017731215804815292, + 0.01079279463738203, + 0.00982870813459158, + 0.015661893412470818, + 0.013989663682878017, + 0.018753550946712494, + 0.011441991664469242, + 0.007253916934132576, + 0.019091526046395302, + 0.006524949800223112, + 0.005497170612215996, + 0.0061418479308485985, + 0.014697232283651829, + 0.021716533228754997, + 0.009419231675565243, + 0.005438555963337421, + 0.012942205183207989, + 0.007020624820142984, + 0.006596032064408064, + 0.009069567546248436, + 0.010355764999985695, + 0.012669738382101059, + 0.012364477850496769, + 0.00748905073851347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.011080872267484665 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Li Jing, Marshal of the Celestial Palace, visible in a full-body shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a sacred suit of armor made of gold and white jade with engraved reliefs of clouds and beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a shimmering seven-story pagoda in his left palm that remains completely still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's body sway slightly while his crimson cloak rises and falls like waves behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low, upward angle with the character standing above a sea of white clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978152547031641 + ], + "opposite_logprob_tokens": [ + -4.634781360626221 + ], + "logprob_reward": -0.00978152547031641, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978152547031641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.271180123090744, + 0.19091612100601196, + 0.027687808498740196, + 0.09274931997060776, + 0.048164114356040955, + 0.08135944604873657, + 0.06606835126876831, + 0.10885972529649734, + 0.07037395238876343, + 0.0426410511136055 + ], + "credit_map": [ + 0.03798236697912216, + 0.039364296942949295, + 0.04225142300128937, + 0.044155389070510864, + 0.02349286712706089, + 0.041041258722543716, + 0.022293251007795334, + 0.02059926651418209, + 0.02635309286415577, + 0.026307936757802963, + 0.037042755633592606, + 0.04097723215818405, + 0.009730387479066849, + 0.02125345915555954, + 0.013696623034775257, + 0.0155546385794878, + 0.0031772602815181017, + 0.004787080455571413, + 0.004246041644364595, + 0.004618439823389053, + 0.002851510886102915, + 0.002804891439154744, + 0.0034445191267877817, + 0.0017580661224201322, + 0.013978943228721619, + 0.011353189125657082, + 0.007261287420988083, + 0.013269315473735332, + 0.01168628130108118, + 0.0189659520983696, + 0.009147047996520996, + 0.007087303791195154, + 0.006509808357805014, + 0.007502492051571608, + 0.0061819348484277725, + 0.005875853355973959, + 0.0057064020074903965, + 0.005704719573259354, + 0.005974109750241041, + 0.004708793945610523, + 0.010308616794645786, + 0.01354743167757988, + 0.013311308808624744, + 0.012129228562116623, + 0.008808176964521408, + 0.009022815153002739, + 0.007142100483179092, + 0.007089771330356598, + 0.01083859708160162, + 0.0086748031899333, + 0.012272662483155727, + 0.008448903448879719, + 0.004799766931682825, + 0.00947093777358532, + 0.005866279825568199, + 0.005696405190974474, + 0.012771178968250751, + 0.01277628168463707, + 0.011344794183969498, + 0.015197434462606907, + 0.017802657559514046, + 0.013394218869507313, + 0.01646982692182064, + 0.009103333577513695, + 0.01707112416625023, + 0.005273227114230394, + 0.0040552676655352116, + 0.006360774859786034, + 0.007094698958098888, + 0.01989523321390152, + 0.006503436714410782, + 0.004120188765227795, + 0.004324981477111578, + 0.005885301623493433, + 0.0037060389295220375, + 0.006184520665556192, + 0.007065783254802227, + 0.004908400122076273, + 0.006674331612884998, + 0.00389169342815876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.00978152547031641 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_1/index_tensor([7510, 7510])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_1/index_tensor([7510, 7510])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..acabe98984a672de1644bb221bfb1f8b47deb157 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_1/index_tensor([7510, 7510])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, overhead tracking shot follows a ferocious, one-horned beast, the Nian, as it thunders through a narrow, ancient street. The beast, as massive as a lion, is covered in dark red scales that gleam metallically under the flickering light. Atop its head rises a massive, spiraling obsidian horn, its tip gleaming with an ominous red glow. Its blood-red eyes burn with fury, scorching white steam emanates from its nostrils, and sharp fangs loom in its gaping maw. The camera, close to the ground, races with it, capturing every powerful hooves' thump on the slippery bluestone pavement, sending up splashes of water and sparks. The streets were lined with quaint wooden structures with soaring eaves and angled corners, while across them hung densely hundreds of round red lanterns, each emitting a warm, saturated glow, painting the entire night a deep crimson. The fierce winds whipped up by the beast's ferocious charge caused the lanterns and their golden tassels to sway violently, and light and shadow danced wildly across the ground and buildings, creating a visual spectacle that intertwined myth and reality, a mixture of festive joy and extreme tension and chaos.", + "seed": 31427, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_1/index_7510_seed_31427_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06814206568524242, + "vlm_logprob": -0.06814206568524242, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.06814206568524242, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0055651492439210415 + ], + "opposite_logprob_tokens": [ + -5.630565166473389 + ], + "logprob_reward": -0.0055651492439210415, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0055651492439210415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07416148483753204, + 0.07863693684339523, + 0.04491938650608063, + 0.07041246443986893, + 0.0850258320569992, + 0.08348093926906586, + 0.20111775398254395, + 0.15873436629772186, + 0.1321946233510971, + 0.07131621241569519 + ], + "credit_map": [ + 0.006582536268979311, + 0.006069707218557596, + 0.013434037566184998, + 0.007619296200573444, + 0.009244091808795929, + 0.010416047647595406, + 0.007807633373886347, + 0.012988129630684853, + 0.012505787424743176, + 0.009502659551799297, + 0.010631213895976543, + 0.011292004026472569, + 0.008878793567419052, + 0.009849278256297112, + 0.006902324967086315, + 0.00907487515360117, + 0.005876793991774321, + 0.005511625669896603, + 0.005530165042728186, + 0.006038675084710121, + 0.005513839889317751, + 0.006316775921732187, + 0.005013255402445793, + 0.005118255037814379, + 0.007511391770094633, + 0.004548539407551289, + 0.010830658487975597, + 0.006688801571726799, + 0.010914905928075314, + 0.010375386103987694, + 0.010461707599461079, + 0.009081066586077213, + 0.00628661597147584, + 0.009718276560306549, + 0.00986343901604414, + 0.016728928312659264, + 0.009886960498988628, + 0.01291227713227272, + 0.008905130438506603, + 0.010724203661084175, + 0.005360995419323444, + 0.006269351579248905, + 0.01114534493535757, + 0.016597965732216835, + 0.008471524342894554, + 0.012353412806987762, + 0.0114237479865551, + 0.011858592741191387, + 0.017321404069662094, + 0.012363391928374767, + 0.01657835580408573, + 0.019478891044855118, + 0.023798743262887, + 0.030170796439051628, + 0.038100820034742355, + 0.04330535978078842, + 0.009986244142055511, + 0.01427119504660368, + 0.008904142305254936, + 0.02191700041294098, + 0.027684979140758514, + 0.016956737264990807, + 0.031066566705703735, + 0.027947505936026573, + 0.008959246799349785, + 0.012224149890244007, + 0.013033085502684116, + 0.023897171020507812, + 0.022591523826122284, + 0.020466605201363564, + 0.017549242824316025, + 0.013473602943122387, + 0.006042845081537962, + 0.006960602942854166, + 0.00843142531812191, + 0.012792014516890049, + 0.011226579546928406, + 0.010277396999299526, + 0.008830399252474308, + 0.006754948757588863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.0055651492439210415 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004649304784834385 + ], + "opposite_logprob_tokens": [ + -5.3796491622924805 + ], + "logprob_reward": -0.004649304784834385, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004649304784834385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2075100541114807, + 0.039921484887599945, + 0.049249570816755295, + 0.09754624962806702, + 0.04975522309541702, + 0.04657770320773125, + 0.30499929189682007, + 0.0777597725391388, + 0.07006385177373886, + 0.056616801768541336 + ], + "credit_map": [ + 0.024652965366840363, + 0.01393485814332962, + 0.05242078751325607, + 0.0164759811013937, + 0.030600832775235176, + 0.031951259821653366, + 0.015291040763258934, + 0.022182339802384377, + 0.004779853858053684, + 0.0033015760127454996, + 0.008743779733777046, + 0.003293177578598261, + 0.005048201885074377, + 0.006214156746864319, + 0.00342587404884398, + 0.005114861764013767, + 0.0075103892013430595, + 0.004742322955280542, + 0.007116645108908415, + 0.005746925249695778, + 0.007387382443994284, + 0.007224884815514088, + 0.004549037665128708, + 0.004971981048583984, + 0.0093096774071455, + 0.006852160673588514, + 0.013651426881551743, + 0.009507855400443077, + 0.018342332914471626, + 0.014102146960794926, + 0.01393101830035448, + 0.01184963807463646, + 0.006064584944397211, + 0.007201161701232195, + 0.006878033746033907, + 0.008224517107009888, + 0.004938616883009672, + 0.006114912685006857, + 0.004394683055579662, + 0.0059387111105024815, + 0.0043663508258759975, + 0.0038288086652755737, + 0.004904819652438164, + 0.007866845466196537, + 0.005628599319607019, + 0.008506315760314465, + 0.005124181974679232, + 0.006351776886731386, + 0.025021078065037727, + 0.020774397999048233, + 0.014582930132746696, + 0.03414393588900566, + 0.02252758853137493, + 0.021788764744997025, + 0.0822635143995285, + 0.08389707654714584, + 0.0036316323094069958, + 0.006535280961543322, + 0.0073595596477389336, + 0.012544555589556694, + 0.007711893878877163, + 0.008406342007219791, + 0.013983577489852905, + 0.017586931586265564, + 0.005135728977620602, + 0.005849773995578289, + 0.007385465782135725, + 0.011354888789355755, + 0.012949180789291859, + 0.011669048108160496, + 0.01136392168700695, + 0.004355845507234335, + 0.005236862692981958, + 0.005276397801935673, + 0.005281240679323673, + 0.008306164294481277, + 0.009984281845390797, + 0.00694948760792613, + 0.006857143249362707, + 0.008725223131477833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.004649304784834385 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947629921138287 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947629921138287, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947629921138287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20531687140464783, + 0.08590370416641235, + 0.14922289550304413, + 0.08322711288928986, + 0.06654304265975952, + 0.11972364783287048, + 0.12618206441402435, + 0.073218934237957, + 0.04111258685588837, + 0.0495491698384285 + ], + "credit_map": [ + 0.023037776350975037, + 0.01985936239361763, + 0.028224026784300804, + 0.02065545693039894, + 0.040957383811473846, + 0.028956012800335884, + 0.016064617782831192, + 0.027562225237488747, + 0.017962515354156494, + 0.008264248259365559, + 0.016721881926059723, + 0.007179273758083582, + 0.009802593849599361, + 0.012633414939045906, + 0.005556182935833931, + 0.007783589418977499, + 0.02365131862461567, + 0.011943656019866467, + 0.024726340547204018, + 0.009072313085198402, + 0.02957700937986374, + 0.0236485805362463, + 0.010849314741790295, + 0.015754345804452896, + 0.010992988012731075, + 0.005729208700358868, + 0.009572362527251244, + 0.0087336590513587, + 0.017166968435049057, + 0.012808683328330517, + 0.009705707430839539, + 0.008517526090145111, + 0.004237344488501549, + 0.006317897699773312, + 0.007608231157064438, + 0.008542663417756557, + 0.012690157629549503, + 0.011189790442585945, + 0.007670728024095297, + 0.008286228403449059, + 0.009733808226883411, + 0.00957402866333723, + 0.013646899722516537, + 0.01129639521241188, + 0.02284712716937065, + 0.020413527265191078, + 0.016572969034314156, + 0.01563888229429722, + 0.006987749133259058, + 0.011395314708352089, + 0.009108292870223522, + 0.02756013348698616, + 0.011672558262944221, + 0.016057096421718597, + 0.018522804602980614, + 0.024878108873963356, + 0.005144759546965361, + 0.00918315164744854, + 0.007046012673527002, + 0.008543531410396099, + 0.012670843861997128, + 0.007647611666470766, + 0.013106804341077805, + 0.009876211173832417, + 0.0031012806575745344, + 0.004213655833154917, + 0.004635617136955261, + 0.006289689335972071, + 0.006412869319319725, + 0.006579749286174774, + 0.0059965248219668865, + 0.003883199766278267, + 0.004260054789483547, + 0.0051699914038181305, + 0.0041695344261825085, + 0.008612210862338543, + 0.010066382586956024, + 0.004925829358398914, + 0.006519473157823086, + 0.005825689062476158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.005947629921138287 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076509952545166 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076509952545166, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076509952545166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1841929852962494, + 0.147854745388031, + 0.079441599547863, + 0.16604673862457275, + 0.0443912148475647, + 0.06084861606359482, + 0.10239103436470032, + 0.10117976367473602, + 0.07352687418460846, + 0.040126457810401917 + ], + "credit_map": [ + 0.025296766310930252, + 0.0223244596272707, + 0.022776495665311813, + 0.019953494891524315, + 0.02798595279455185, + 0.024611979722976685, + 0.01603708788752556, + 0.02520674839615822, + 0.0270893145352602, + 0.017596494406461716, + 0.025427602231502533, + 0.015694215893745422, + 0.017013926059007645, + 0.019734838977456093, + 0.012123228050768375, + 0.013175125233829021, + 0.008369834162294865, + 0.008949805051088333, + 0.0076158479787409306, + 0.010973970405757427, + 0.014428389258682728, + 0.011259087361395359, + 0.009326830506324768, + 0.008517827838659286, + 0.01826813630759716, + 0.010326133109629154, + 0.027080457657575607, + 0.01696014776825905, + 0.0180808212608099, + 0.021940944716334343, + 0.032206449657678604, + 0.021183636039495468, + 0.004303227178752422, + 0.004946856759488583, + 0.004922917112708092, + 0.005405418574810028, + 0.0067774211056530476, + 0.006010500248521566, + 0.005532704759389162, + 0.0064921630546450615, + 0.004263332113623619, + 0.00654748547822237, + 0.006015595979988575, + 0.009720900095999241, + 0.010448191314935684, + 0.007766880560666323, + 0.008495296351611614, + 0.007590928114950657, + 0.013767576776444912, + 0.010387388989329338, + 0.007976407185196877, + 0.02019350416958332, + 0.01268704328685999, + 0.010381854139268398, + 0.012347167357802391, + 0.014650074765086174, + 0.005206959787756205, + 0.012195892632007599, + 0.01008462905883789, + 0.015001673251390457, + 0.012491012923419476, + 0.011670959182083607, + 0.019421467557549477, + 0.015107161365449429, + 0.008141763508319855, + 0.00782274641096592, + 0.00898546352982521, + 0.013239548541605473, + 0.01028822734951973, + 0.008689459413290024, + 0.011303490027785301, + 0.005056170746684074, + 0.004648122470825911, + 0.003886529942974448, + 0.003664798568934202, + 0.00729739386588335, + 0.006179307587444782, + 0.004271613899618387, + 0.004911197815090418, + 0.005267491098493338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.011076509952545166 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134717345237732 + ], + "opposite_logprob_tokens": [ + -1.313471794128418 + ], + "logprob_reward": -0.3134717345237732, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134717345237732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1538512110710144, + 0.061362773180007935, + 0.04499706253409386, + 0.07303640246391296, + 0.09439355134963989, + 0.08254485577344894, + 0.08139798790216446, + 0.13544289767742157, + 0.18770621716976166, + 0.0852670893073082 + ], + "credit_map": [ + 0.01637372560799122, + 0.010914157144725323, + 0.033980607986450195, + 0.015813717618584633, + 0.02620595134794712, + 0.02309846505522728, + 0.011312615126371384, + 0.016151973977684975, + 0.007461214438080788, + 0.0047154780477285385, + 0.01399950124323368, + 0.005140024237334728, + 0.007703917566686869, + 0.011307398788630962, + 0.004030521959066391, + 0.007004716899245977, + 0.005606591235846281, + 0.00490003265440464, + 0.0055038221180438995, + 0.004819852765649557, + 0.007144905626773834, + 0.006559332832694054, + 0.0054261949844658375, + 0.005036328453570604, + 0.00936955027282238, + 0.006571881007403135, + 0.009902863763272762, + 0.00780119001865387, + 0.011633604764938354, + 0.011105482466518879, + 0.008437132462859154, + 0.008214695379137993, + 0.006035215687006712, + 0.008424242027103901, + 0.010450633242726326, + 0.00984184816479683, + 0.017753249034285545, + 0.01855793595314026, + 0.008599380031228065, + 0.014731047675013542, + 0.0057078078389167786, + 0.01025296188890934, + 0.007380587514489889, + 0.010376020334661007, + 0.010832269676029682, + 0.015342306345701218, + 0.010391846299171448, + 0.012261050753295422, + 0.00715810339897871, + 0.006695292890071869, + 0.008483868092298508, + 0.019949370995163918, + 0.00734913581982255, + 0.008769885636866093, + 0.009376429952681065, + 0.013615905307233334, + 0.007079912815243006, + 0.012602148577570915, + 0.007455222308635712, + 0.02372308075428009, + 0.020204294472932816, + 0.012700936757028103, + 0.019259819760918617, + 0.032417476177215576, + 0.022565094754099846, + 0.01635945960879326, + 0.025937115773558617, + 0.03236435726284981, + 0.02849818952381611, + 0.01765337400138378, + 0.03173816204071045, + 0.012590466067194939, + 0.011229547671973705, + 0.007321609649807215, + 0.010378657840192318, + 0.013034243136644363, + 0.012891530059278011, + 0.01107360515743494, + 0.010834086686372757, + 0.008503809571266174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.3134717345237732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, overhead tracking shot follows a ferocious, one-horned beast, the Nian, as it thunders through a narrow, ancient street. The beast, as massive as a lion, is covered in dark red scales that gleam metallically under the flickering light. Atop its head rises a massive, spiraling obsidian horn, its tip gleaming with an ominous red glow. Its blood-red eyes burn with fury, scorching white steam emanates from its nostrils, and sharp fangs loom in its gaping maw. The camera, close to the ground, races with it, capturing every powerful hooves' thump on the slippery bluestone pavement, sending up splashes of water and sparks. The streets were lined with quaint wooden structures with soaring eaves and angled corners, while across them hung densely hundreds of round red lanterns, each emitting a warm, saturated glow, painting the entire night a deep crimson. The fierce winds whipped up by the beast's ferocious charge caused the lanterns and their golden tassels to sway violently, and light and shadow danced wildly across the ground and buildings, creating a visual spectacle that intertwined myth and reality, a mixture of festive joy and extreme tension and chaos.", + "seed": 31428, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_1/index_7510_seed_31428_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08494850965216756, + "vlm_logprob": -0.08494850965216756, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.08494850965216756, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007291491609066725 + ], + "opposite_logprob_tokens": [ + -5.132291316986084 + ], + "logprob_reward": -0.007291491609066725, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.007291491609066725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13546141982078552, + 0.08236025273799896, + 0.09390248358249664, + 0.07659228146076202, + 0.04224569350481033, + 0.039108652621507645, + 0.20807301998138428, + 0.06915067881345749, + 0.06186874583363533, + 0.19123682379722595 + ], + "credit_map": [ + 0.02365965209901333, + 0.011785629205405712, + 0.017948154360055923, + 0.015164073556661606, + 0.02446643076837063, + 0.02042112685739994, + 0.010502947494387627, + 0.01151339989155531, + 0.009854717180132866, + 0.007949136197566986, + 0.01556787546724081, + 0.008610744960606098, + 0.008322006091475487, + 0.013777684420347214, + 0.008211254142224789, + 0.010066828690469265, + 0.010963145643472672, + 0.0052095577120780945, + 0.019252849742770195, + 0.0062350002117455006, + 0.007101903203874826, + 0.010842092335224152, + 0.007124624215066433, + 0.027173301205039024, + 0.00418538274243474, + 0.006345515605062246, + 0.01203959807753563, + 0.00979330763220787, + 0.012791449204087257, + 0.010674121789634228, + 0.012236501090228558, + 0.00852639228105545, + 0.005902340169996023, + 0.0033521228469908237, + 0.004104596097022295, + 0.005732735153287649, + 0.007228692527860403, + 0.00491362065076828, + 0.006175160873681307, + 0.004836423322558403, + 0.004548638593405485, + 0.0034173554740846157, + 0.004714292008429766, + 0.0028185942210257053, + 0.004233123268932104, + 0.008421779610216618, + 0.0056603108532726765, + 0.005294552072882652, + 0.014800213277339935, + 0.02402462437748909, + 0.03242764621973038, + 0.04515696316957474, + 0.0215828325599432, + 0.016466254368424416, + 0.03414011746644974, + 0.019474349915981293, + 0.00855939369648695, + 0.006114677060395479, + 0.010715249925851822, + 0.00600597495213151, + 0.00839065108448267, + 0.012126786634325981, + 0.010306049138307571, + 0.0069318898022174835, + 0.006559225730597973, + 0.007571861147880554, + 0.013952532783150673, + 0.006466945633292198, + 0.006131679750978947, + 0.006851342972368002, + 0.005750180687755346, + 0.008584973402321339, + 0.015875207260251045, + 0.03078906610608101, + 0.02292637526988983, + 0.026105470955371857, + 0.026895571500062943, + 0.01958303339779377, + 0.022792818024754524, + 0.02626926638185978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.007291491609066725 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067496877163648605 + ], + "opposite_logprob_tokens": [ + -5.006749629974365 + ], + "logprob_reward": -0.0067496877163648605, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067496877163648605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11660421639680862, + 0.16832013428211212, + 0.06273314356803894, + 0.053356051445007324, + 0.054349374026060104, + 0.0263595562428236, + 0.07460272312164307, + 0.18082813918590546, + 0.14522816240787506, + 0.11761848628520966 + ], + "credit_map": [ + 0.02056385576725006, + 0.011006948538124561, + 0.016690105199813843, + 0.01560969091951847, + 0.015547635965049267, + 0.01583394967019558, + 0.011957457289099693, + 0.00939458142966032, + 0.025110391899943352, + 0.01864486001431942, + 0.02150677517056465, + 0.028306471183896065, + 0.020980048924684525, + 0.01652051880955696, + 0.019639404490590096, + 0.017611678689718246, + 0.008948264643549919, + 0.005148795899003744, + 0.007465038448572159, + 0.005067541729658842, + 0.007939170114696026, + 0.009817034937441349, + 0.008525164797902107, + 0.009822134859859943, + 0.003295659786090255, + 0.005090765189379454, + 0.007176144979894161, + 0.006626138463616371, + 0.006632174365222454, + 0.006950354669243097, + 0.011942201294004917, + 0.005642617587000132, + 0.005319908261299133, + 0.0053782095201313496, + 0.005101614631712437, + 0.010485735721886158, + 0.007325488608330488, + 0.0064499168656766415, + 0.00803495105355978, + 0.006253551691770554, + 0.0033980123698711395, + 0.0027490758802741766, + 0.003624001517891884, + 0.0020118497777730227, + 0.002900304039940238, + 0.003376913256943226, + 0.005308554507791996, + 0.00299084628932178, + 0.00395148666575551, + 0.007560246624052525, + 0.011191805824637413, + 0.012935571372509003, + 0.009428702294826508, + 0.0070073665119707584, + 0.015384075231850147, + 0.007143465336412191, + 0.019473450258374214, + 0.018207112327218056, + 0.025552716106176376, + 0.01094757579267025, + 0.01691420003771782, + 0.046700187027454376, + 0.02016725391149521, + 0.022865649312734604, + 0.014551163651049137, + 0.024741172790527344, + 0.02202570252120495, + 0.021938953548669815, + 0.01737920381128788, + 0.018144896253943443, + 0.012623386457562447, + 0.013823683373630047, + 0.010671110823750496, + 0.0219802837818861, + 0.010447822511196136, + 0.014027209021151066, + 0.018251677975058556, + 0.014249933883547783, + 0.013026685453951359, + 0.014963772147893906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.0067496877163648605 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072383262217045 + ], + "opposite_logprob_tokens": [ + -4.511072158813477 + ], + "logprob_reward": -0.011072383262217045, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072383262217045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20811548829078674, + 0.16555717587471008, + 0.1527346819639206, + 0.1409044861793518, + 0.03159138560295105, + 0.03322641924023628, + 0.09186826646327972, + 0.035734497010707855, + 0.09271753579378128, + 0.0475500263273716 + ], + "credit_map": [ + 0.02070845104753971, + 0.022425038740038872, + 0.03300865739583969, + 0.02540498971939087, + 0.0282205268740654, + 0.033900342881679535, + 0.022825077176094055, + 0.02162240631878376, + 0.017504338175058365, + 0.01668667420744896, + 0.03176885470747948, + 0.019466552883386612, + 0.01673884317278862, + 0.02639705315232277, + 0.017636993899941444, + 0.019357876852154732, + 0.020124953240156174, + 0.016301609575748444, + 0.02100244164466858, + 0.01146798487752676, + 0.02326607145369053, + 0.019827863201498985, + 0.01585361547768116, + 0.024890130385756493, + 0.028958206996321678, + 0.008106794208288193, + 0.027413303032517433, + 0.010306290350854397, + 0.010090572759509087, + 0.022329291328787804, + 0.012900406494736671, + 0.02079962193965912, + 0.002886186819523573, + 0.00394925381988287, + 0.003213596297428012, + 0.005018972791731358, + 0.003634183667600155, + 0.003832734189927578, + 0.005485626868903637, + 0.0035708302166312933, + 0.004014261532574892, + 0.0024114788975566626, + 0.0032374628353863955, + 0.0021676570177078247, + 0.00366885494440794, + 0.006644138600677252, + 0.005742416717112064, + 0.005340151488780975, + 0.005110672675073147, + 0.007721835747361183, + 0.01141375582665205, + 0.011312825605273247, + 0.01335164625197649, + 0.018970029428601265, + 0.010285311378538609, + 0.01370218675583601, + 0.004098157398402691, + 0.0038501578383147717, + 0.004829090088605881, + 0.0028391380328685045, + 0.003991532605141401, + 0.007981189526617527, + 0.0044738841243088245, + 0.0036713448353111744, + 0.009289713576436043, + 0.009411318227648735, + 0.02004118449985981, + 0.010591729544103146, + 0.011001943610608578, + 0.010959516279399395, + 0.010003163479268551, + 0.011418969370424747, + 0.005188872572034597, + 0.0069825355894863605, + 0.006074799690395594, + 0.005023035686463118, + 0.006920286454260349, + 0.0057077486999332905, + 0.005429598968476057, + 0.006223150994628668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011072383262217045 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254018023610115 + ], + "opposite_logprob_tokens": [ + -4.387540340423584 + ], + "logprob_reward": -0.01254018023610115, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254018023610115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29911860823631287, + 0.2737221121788025, + 0.05822843685746193, + 0.046944163739681244, + 0.09260626137256622, + 0.03559095785021782, + 0.058160893619060516, + 0.04477321729063988, + 0.05638184770941734, + 0.03447341546416283 + ], + "credit_map": [ + 0.046715717762708664, + 0.024962684139609337, + 0.039314404129981995, + 0.040871646255254745, + 0.04697579890489578, + 0.038013093173503876, + 0.03350896015763283, + 0.028756339102983475, + 0.030965249985456467, + 0.027361992746591568, + 0.04200352355837822, + 0.04063430055975914, + 0.029202723875641823, + 0.035708218812942505, + 0.03710857033729553, + 0.030737563967704773, + 0.0036248748656362295, + 0.003842088859528303, + 0.01237414963543415, + 0.007044204510748386, + 0.004271176643669605, + 0.006898503750562668, + 0.005151907913386822, + 0.015021538361907005, + 0.0037668964359909296, + 0.0038854191079735756, + 0.007169037126004696, + 0.006290572229772806, + 0.008571128360927105, + 0.0055064186453819275, + 0.006547427270561457, + 0.005207268055528402, + 0.009068239480257034, + 0.008530122227966785, + 0.009104126133024693, + 0.0119169931858778, + 0.008718511089682579, + 0.011229777708649635, + 0.018097782507538795, + 0.015940720215439796, + 0.004774688743054867, + 0.0028506340458989143, + 0.0032250850927084684, + 0.0027226938400417566, + 0.0041710683144629, + 0.0073744445107877254, + 0.005709722638130188, + 0.004762623459100723, + 0.0049657067283988, + 0.0061554922722280025, + 0.00796709768474102, + 0.011959532275795937, + 0.005879729986190796, + 0.00593990832567215, + 0.007211050949990749, + 0.008082381449639797, + 0.0035041682422161102, + 0.003328730119392276, + 0.005605337209999561, + 0.005360342562198639, + 0.005513652227818966, + 0.009809131734073162, + 0.0071219028905034065, + 0.004529960919171572, + 0.006652228534221649, + 0.00852831732481718, + 0.010396439582109451, + 0.008112365379929543, + 0.005203044041991234, + 0.005660884082317352, + 0.004553500097244978, + 0.007275072857737541, + 0.0032111015170812607, + 0.00654952134937048, + 0.003133899299427867, + 0.004110532812774181, + 0.004938460420817137, + 0.004885999485850334, + 0.0029442422091960907, + 0.004699661396443844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.01254018023610115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.387088805437088 + ], + "opposite_logprob_tokens": [ + -1.1370887756347656 + ], + "logprob_reward": -0.387088805437088, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.387088805437088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1664716601371765, + 0.10575088113546371, + 0.0629439726471901, + 0.07217144966125488, + 0.109270378947258, + 0.07945282012224197, + 0.13207419216632843, + 0.05712118372321129, + 0.1108374297618866, + 0.1039060577750206 + ], + "credit_map": [ + 0.02893691696226597, + 0.01762501336634159, + 0.022681409493088722, + 0.019560949876904488, + 0.02057272382080555, + 0.02262755110859871, + 0.017131971195340157, + 0.017335116863250732, + 0.015065271407365799, + 0.013250922784209251, + 0.013979784213006496, + 0.016013875603675842, + 0.00982260424643755, + 0.014073304831981659, + 0.011294996365904808, + 0.012250126339495182, + 0.012604359537363052, + 0.0063816849142313, + 0.006736492272466421, + 0.005988295190036297, + 0.0096064992249012, + 0.0073348297737538815, + 0.007347154431045055, + 0.006944655906409025, + 0.008650346659123898, + 0.008544128388166428, + 0.014165138825774193, + 0.0060842689126729965, + 0.00695593049749732, + 0.009601229801774025, + 0.008190060965716839, + 0.00998033955693245, + 0.008843143470585346, + 0.013554477132856846, + 0.010169698856770992, + 0.016185693442821503, + 0.018865205347537994, + 0.01206138264387846, + 0.015540552325546741, + 0.014050224795937538, + 0.007652555126696825, + 0.0057146381586790085, + 0.009936368092894554, + 0.004861762747168541, + 0.008744853548705578, + 0.020520903170108795, + 0.009104925207793713, + 0.012916810810565948, + 0.010888163931667805, + 0.014413333497941494, + 0.018259085714817047, + 0.02456052228808403, + 0.02005963772535324, + 0.016940873116254807, + 0.014154549688100815, + 0.012798038311302662, + 0.005320168565958738, + 0.006187473423779011, + 0.0074616181664168835, + 0.004312749020755291, + 0.006859479937702417, + 0.012178760021924973, + 0.007796322461217642, + 0.007004610262811184, + 0.012946605682373047, + 0.016328446567058563, + 0.022803984582424164, + 0.014724452048540115, + 0.011096436530351639, + 0.01162413414567709, + 0.01028471440076828, + 0.011028660461306572, + 0.012211999855935574, + 0.01922682672739029, + 0.010061411187052727, + 0.01142065692692995, + 0.013367734849452972, + 0.008793659508228302, + 0.014189600944519043, + 0.014634167775511742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.387088805437088 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, overhead tracking shot follows a ferocious, one-horned beast, the Nian, as it thunders through a narrow, ancient street. The beast, as massive as a lion, is covered in dark red scales that gleam metallically under the flickering light. Atop its head rises a massive, spiraling obsidian horn, its tip gleaming with an ominous red glow. Its blood-red eyes burn with fury, scorching white steam emanates from its nostrils, and sharp fangs loom in its gaping maw. The camera, close to the ground, races with it, capturing every powerful hooves' thump on the slippery bluestone pavement, sending up splashes of water and sparks. The streets were lined with quaint wooden structures with soaring eaves and angled corners, while across them hung densely hundreds of round red lanterns, each emitting a warm, saturated glow, painting the entire night a deep crimson. The fierce winds whipped up by the beast's ferocious charge caused the lanterns and their golden tassels to sway violently, and light and shadow danced wildly across the ground and buildings, creating a visual spectacle that intertwined myth and reality, a mixture of festive joy and extreme tension and chaos.", + "seed": 31429, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_1/index_7510_seed_31429_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09006012715399266, + "vlm_logprob": -0.09006012715399266, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.09006012715399266, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016136404126882553 + ], + "opposite_logprob_tokens": [ + -4.266136169433594 + ], + "logprob_reward": -0.016136404126882553, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016136404126882553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07111815363168716, + 0.050261612981557846, + 0.05557163059711456, + 0.04251163825392723, + 0.26017284393310547, + 0.04121735319495201, + 0.03488343581557274, + 0.2376047670841217, + 0.07186552882194519, + 0.13479305803775787 + ], + "credit_map": [ + 0.007227097172290087, + 0.006681176833808422, + 0.007055106572806835, + 0.009739816188812256, + 0.010899237357079983, + 0.008624994195997715, + 0.011718649417161942, + 0.009172079153358936, + 0.004747081082314253, + 0.004468241240829229, + 0.004792267922312021, + 0.0053218090906739235, + 0.007540830411016941, + 0.006857153028249741, + 0.010343819856643677, + 0.0061904084868729115, + 0.003118715714663267, + 0.0036223132628947496, + 0.0033764359541237354, + 0.0029749241657555103, + 0.003832609159871936, + 0.017023006454110146, + 0.00689176470041275, + 0.014731860719621181, + 0.005355332978069782, + 0.0034549550618976355, + 0.0027039346750825644, + 0.002846148330718279, + 0.0046188002452254295, + 0.011886168271303177, + 0.004091876093298197, + 0.007554421667009592, + 0.03748632222414017, + 0.027156081050634384, + 0.00977598037570715, + 0.016416577622294426, + 0.031189527362585068, + 0.04040179401636124, + 0.036332640796899796, + 0.06141392141580582, + 0.0041704257018864155, + 0.0029450482688844204, + 0.004012942314147949, + 0.0029807279352098703, + 0.006511153187602758, + 0.006732132285833359, + 0.005254084710031748, + 0.00861083809286356, + 0.0031704085413366556, + 0.0037357620894908905, + 0.004546957090497017, + 0.0027621136978268623, + 0.0074989269487559795, + 0.005024272948503494, + 0.004294993821531534, + 0.0038499997463077307, + 0.023671802133321762, + 0.05709774047136307, + 0.01607285998761654, + 0.02715630643069744, + 0.039622873067855835, + 0.02985183708369732, + 0.02796807326376438, + 0.016163276508450508, + 0.006193455774337053, + 0.008672468364238739, + 0.011093830689787865, + 0.007441939786076546, + 0.012299456633627415, + 0.009337212890386581, + 0.008430412970483303, + 0.00839675310999155, + 0.014604132622480392, + 0.017791112884879112, + 0.01566915027797222, + 0.023643186315894127, + 0.015105453319847584, + 0.013837665319442749, + 0.021608348935842514, + 0.012534014880657196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.016136404126882553 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651867508888245 + ], + "opposite_logprob_tokens": [ + -4.7586517333984375 + ], + "logprob_reward": -0.008651867508888245, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651867508888245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14894311130046844, + 0.15533660352230072, + 0.054150231182575226, + 0.050577763468027115, + 0.20363496243953705, + 0.045098792761564255, + 0.10645276308059692, + 0.04474443569779396, + 0.07100226730108261, + 0.1200590431690216 + ], + "credit_map": [ + 0.016397560015320778, + 0.01428685151040554, + 0.015763262286782265, + 0.017631685361266136, + 0.02096150442957878, + 0.02524835243821144, + 0.01980479434132576, + 0.01884910836815834, + 0.016690563410520554, + 0.015547409653663635, + 0.017422202974557877, + 0.01963849365711212, + 0.023168392479419708, + 0.02245977893471718, + 0.025476548820734024, + 0.014933210797607899, + 0.005690440535545349, + 0.005027949344366789, + 0.005426469724625349, + 0.00535894650965929, + 0.00418090308085084, + 0.01151999644935131, + 0.008006949909031391, + 0.008938571438193321, + 0.005964759737253189, + 0.003542593214660883, + 0.002841843757778406, + 0.0028922343626618385, + 0.005780078936368227, + 0.01426184456795454, + 0.005687395576387644, + 0.009607015177607536, + 0.056553103029727936, + 0.019662223756313324, + 0.00976877473294735, + 0.011913284659385681, + 0.022967351600527763, + 0.0226619690656662, + 0.023633049800992012, + 0.03647519275546074, + 0.004601455293595791, + 0.0033586393110454082, + 0.005515008233487606, + 0.00394480163231492, + 0.005347772501409054, + 0.0066562918946146965, + 0.005466138478368521, + 0.010208684019744396, + 0.008236652240157127, + 0.013821832835674286, + 0.012816187925636768, + 0.008308861404657364, + 0.024226976558566093, + 0.013933110982179642, + 0.014755915850400925, + 0.010353225283324718, + 0.004980647005140781, + 0.005850351415574551, + 0.005520453676581383, + 0.004151557572185993, + 0.006234372034668922, + 0.007206107024103403, + 0.005418510176241398, + 0.005382439587265253, + 0.006176348309963942, + 0.008375568315386772, + 0.01228345651179552, + 0.008545485325157642, + 0.009833196178078651, + 0.010010185651481152, + 0.008337547071278095, + 0.007440478540956974, + 0.012850545346736908, + 0.01724984683096409, + 0.013772229664027691, + 0.021291188895702362, + 0.014111403375864029, + 0.011224575340747833, + 0.018962277099490166, + 0.010596973821520805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.008651867508888245 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634967729449272 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008634967729449272, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634967729449272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10141298919916153, + 0.055254016071558, + 0.09902121126651764, + 0.034941207617521286, + 0.2173273265361786, + 0.14586758613586426, + 0.025097208097577095, + 0.03752230480313301, + 0.04942084103822708, + 0.23413531482219696 + ], + "credit_map": [ + 0.012921391054987907, + 0.014228541404008865, + 0.014826800674200058, + 0.008193187415599823, + 0.010001178830862045, + 0.01722239889204502, + 0.009831011295318604, + 0.014188481494784355, + 0.006936261896044016, + 0.006137257441878319, + 0.006202626973390579, + 0.007528160698711872, + 0.007061130832880735, + 0.006231996696442366, + 0.007948583923280239, + 0.007207993883639574, + 0.010152089409530163, + 0.012535035610198975, + 0.009461999870836735, + 0.007110452279448509, + 0.007033162750303745, + 0.024049118161201477, + 0.0064263432286679745, + 0.022253014147281647, + 0.0050157527439296246, + 0.0035372707061469555, + 0.003483776468783617, + 0.002794282976537943, + 0.003359842114150524, + 0.008002244867384434, + 0.0038387735839933157, + 0.004909262526780367, + 0.06938964873552322, + 0.012116709724068642, + 0.00890787411481142, + 0.007798558566719294, + 0.01338159292936325, + 0.08234667032957077, + 0.009365790523588657, + 0.014020482078194618, + 0.013417944312095642, + 0.013217374682426453, + 0.008481990545988083, + 0.010964835993945599, + 0.019312065094709396, + 0.021057724952697754, + 0.022278962656855583, + 0.037136685103178024, + 0.002111023059114814, + 0.0033466711174696684, + 0.0023931313771754503, + 0.0024082500021904707, + 0.0038994152564555407, + 0.004426232073456049, + 0.0024734276812523603, + 0.00403905613347888, + 0.004636902827769518, + 0.007581484969705343, + 0.003943923860788345, + 0.003131344448775053, + 0.005173892714083195, + 0.004649993497878313, + 0.004437271971255541, + 0.003967490047216415, + 0.0059698475524783134, + 0.00879997480660677, + 0.007235330995172262, + 0.004530026577413082, + 0.006015992723405361, + 0.005995142739266157, + 0.005792304873466492, + 0.005082221236079931, + 0.01888495869934559, + 0.03424656391143799, + 0.020689791068434715, + 0.029676148667931557, + 0.03009265847504139, + 0.03316887840628624, + 0.03594283014535904, + 0.031433477997779846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.008634967729449272 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795892536640167 + ], + "opposite_logprob_tokens": [ + -3.5297958850860596 + ], + "logprob_reward": -0.029795892536640167, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795892536640167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09528400748968124, + 0.03165135905146599, + 0.05851394683122635, + 0.0524110347032547, + 0.5485630631446838, + 0.04338107630610466, + 0.05242719128727913, + 0.035880196839571, + 0.02691749483346939, + 0.0549706369638443 + ], + "credit_map": [ + 0.009639744646847248, + 0.011935376562178135, + 0.01146288774907589, + 0.01099223829805851, + 0.012362941168248653, + 0.014039631932973862, + 0.012221832759678364, + 0.012629358097910881, + 0.0032196638640016317, + 0.0027421098202466965, + 0.0033470839262008667, + 0.0030575660057365894, + 0.004593841731548309, + 0.004045448731631041, + 0.0067902239970862865, + 0.003855422604829073, + 0.003305545775219798, + 0.0037726382724940777, + 0.0048254127614200115, + 0.004190118983387947, + 0.00506643857806921, + 0.015273921191692352, + 0.00762578658759594, + 0.014454081654548645, + 0.004888462834060192, + 0.004788772203028202, + 0.0035700176376849413, + 0.0038089866284281015, + 0.005960455629974604, + 0.015477363020181656, + 0.004466264043003321, + 0.009450712241232395, + 0.12801411747932434, + 0.030718056485056877, + 0.019723739475011826, + 0.02164362743496895, + 0.03771398589015007, + 0.20257030427455902, + 0.03742283210158348, + 0.07075639814138412, + 0.003707981202751398, + 0.0029263764154165983, + 0.005111252889037132, + 0.0034443936310708523, + 0.00512883672490716, + 0.006515162065625191, + 0.007296570111066103, + 0.009250505827367306, + 0.004426203202456236, + 0.002964657498523593, + 0.006091127172112465, + 0.005018020048737526, + 0.005149830132722855, + 0.011992939747869968, + 0.004986814223229885, + 0.011797599494457245, + 0.0036706419195979834, + 0.004948040936142206, + 0.003339249873533845, + 0.0036546331830322742, + 0.005048099905252457, + 0.004541794303804636, + 0.006758722476661205, + 0.003919015172868967, + 0.002412644913420081, + 0.004217210225760937, + 0.003733925288543105, + 0.0027760895900428295, + 0.005098148249089718, + 0.003231137292459607, + 0.0033282688818871975, + 0.002120071789249778, + 0.005375741049647331, + 0.008010265417397022, + 0.006394647061824799, + 0.00967325083911419, + 0.0059638298116624355, + 0.00457246508449316, + 0.009387413039803505, + 0.005593022797256708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.029795892536640167 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.387081503868103 + ], + "opposite_logprob_tokens": [ + -1.137081503868103 + ], + "logprob_reward": -0.387081503868103, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.387081503868103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14235329627990723, + 0.06170163303613663, + 0.0786687508225441, + 0.04107262194156647, + 0.20351016521453857, + 0.0371435210108757, + 0.062037162482738495, + 0.07372888922691345, + 0.0540354885160923, + 0.24574846029281616 + ], + "credit_map": [ + 0.017375288531184196, + 0.019573159515857697, + 0.013518394902348518, + 0.015580510720610619, + 0.019047776237130165, + 0.028048226609826088, + 0.01695387251675129, + 0.012256074696779251, + 0.0073773860931396484, + 0.007403257768601179, + 0.0046197520568966866, + 0.006314207799732685, + 0.009045017883181572, + 0.008021856658160686, + 0.012099423445761204, + 0.006820731796324253, + 0.012615101411938667, + 0.00822499580681324, + 0.00577740790322423, + 0.006792163942009211, + 0.0060666510835289955, + 0.014690897427499294, + 0.007051833439618349, + 0.017449701204895973, + 0.006331580225378275, + 0.004360562656074762, + 0.0046831159852445126, + 0.00462745875120163, + 0.004281874746084213, + 0.003491966752335429, + 0.005591840948909521, + 0.007704221643507481, + 0.0637052059173584, + 0.012593403458595276, + 0.009009293280541897, + 0.008373590186238289, + 0.013553101569414139, + 0.07393282651901245, + 0.008941476233303547, + 0.013401261530816555, + 0.004404596518725157, + 0.003543849103152752, + 0.004269423894584179, + 0.0044359429739415646, + 0.004060074687004089, + 0.005192759912461042, + 0.00591342942789197, + 0.005323444027453661, + 0.0057534161023795605, + 0.007608088664710522, + 0.006892366334795952, + 0.00484548881649971, + 0.012414348311722279, + 0.008287755772471428, + 0.008218624629080296, + 0.008017076179385185, + 0.004482867196202278, + 0.00794235896319151, + 0.00870650727301836, + 0.008049281314015388, + 0.009875524789094925, + 0.011103639379143715, + 0.012343325652182102, + 0.0112253837287426, + 0.005703268107026815, + 0.007932675071060658, + 0.004279902670532465, + 0.006793114356696606, + 0.008913942612707615, + 0.008926483802497387, + 0.00680615846067667, + 0.004679943900555372, + 0.022801632061600685, + 0.0340692512691021, + 0.023699510842561722, + 0.03920857608318329, + 0.026558909565210342, + 0.029066454619169235, + 0.041984111070632935, + 0.028360020369291306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.387081503868103 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, overhead tracking shot follows a ferocious, one-horned beast, the Nian, as it thunders through a narrow, ancient street. The beast, as massive as a lion, is covered in dark red scales that gleam metallically under the flickering light. Atop its head rises a massive, spiraling obsidian horn, its tip gleaming with an ominous red glow. Its blood-red eyes burn with fury, scorching white steam emanates from its nostrils, and sharp fangs loom in its gaping maw. The camera, close to the ground, races with it, capturing every powerful hooves' thump on the slippery bluestone pavement, sending up splashes of water and sparks. The streets were lined with quaint wooden structures with soaring eaves and angled corners, while across them hung densely hundreds of round red lanterns, each emitting a warm, saturated glow, painting the entire night a deep crimson. The fierce winds whipped up by the beast's ferocious charge caused the lanterns and their golden tassels to sway violently, and light and shadow danced wildly across the ground and buildings, creating a visual spectacle that intertwined myth and reality, a mixture of festive joy and extreme tension and chaos.", + "seed": 31430, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_1/index_7510_seed_31430_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10896979719400406, + "vlm_logprob": -0.10896979719400406, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.10896979719400406, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0121212312951684 + ], + "opposite_logprob_tokens": [ + -4.512121200561523 + ], + "logprob_reward": -0.0121212312951684, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0121212312951684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11889449506998062, + 0.028376758098602295, + 0.14354485273361206, + 0.2565801739692688, + 0.14224225282669067, + 0.05841932073235512, + 0.08871182799339294, + 0.07257889211177826, + 0.04902881756424904, + 0.041622646152973175 + ], + "credit_map": [ + 0.010976127348840237, + 0.009199920110404491, + 0.0116674629971385, + 0.014775911346077919, + 0.020961491391062737, + 0.01616191677749157, + 0.019975263625383377, + 0.015176400542259216, + 0.0038321318570524454, + 0.003476772690191865, + 0.004202017094939947, + 0.003126363502815366, + 0.003528476459905505, + 0.003970997408032417, + 0.0035570859909057617, + 0.002682914026081562, + 0.011241598054766655, + 0.016786936670541763, + 0.012604299932718277, + 0.01802031882107258, + 0.029204614460468292, + 0.014986896887421608, + 0.02862687036395073, + 0.012073312886059284, + 0.018629949539899826, + 0.037782005965709686, + 0.022218136116862297, + 0.03195905312895775, + 0.05432868003845215, + 0.031653955578804016, + 0.03701457753777504, + 0.02299380861222744, + 0.01795787923038006, + 0.021878570318222046, + 0.01698337309062481, + 0.012731742113828659, + 0.02616848424077034, + 0.015558148734271526, + 0.016690051183104515, + 0.014274002984166145, + 0.0042447797022759914, + 0.005998431239277124, + 0.007416839711368084, + 0.007699940353631973, + 0.009069313295185566, + 0.00686772633343935, + 0.011107759550213814, + 0.00601452961564064, + 0.010670382529497147, + 0.009786983951926231, + 0.011288674548268318, + 0.012485520914196968, + 0.01079715695232153, + 0.01061203796416521, + 0.014891665428876877, + 0.008179405704140663, + 0.01005559042096138, + 0.0073943487368524075, + 0.014271472580730915, + 0.006228641141206026, + 0.008868470788002014, + 0.008152789436280727, + 0.012486250139772892, + 0.005121328867971897, + 0.003585876664146781, + 0.009942559525370598, + 0.005187724716961384, + 0.007708098739385605, + 0.007814397104084492, + 0.003349696984514594, + 0.006890654563903809, + 0.004549810197204351, + 0.004729247186332941, + 0.005409123841673136, + 0.007241250015795231, + 0.0063697719015181065, + 0.0033849214669317007, + 0.004361839033663273, + 0.004688560497015715, + 0.005437934305518866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.0121212312951684 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746253930032253 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746253930032253, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746253930032253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1612294316291809, + 0.050676386803388596, + 0.11818963289260864, + 0.0922846570611, + 0.09808840602636337, + 0.07069355994462967, + 0.11234854906797409, + 0.12843942642211914, + 0.06417496502399445, + 0.1038750484585762 + ], + "credit_map": [ + 0.014118168503046036, + 0.014038970693945885, + 0.015983615070581436, + 0.02340228669345379, + 0.03308161720633507, + 0.017039388418197632, + 0.02618665061891079, + 0.01737871766090393, + 0.005241872277110815, + 0.006540022324770689, + 0.006418098229914904, + 0.00828205794095993, + 0.005313320551067591, + 0.00572123983874917, + 0.006679655984044075, + 0.006480115465819836, + 0.010098926723003387, + 0.013235966674983501, + 0.012201597914099693, + 0.013688030652701855, + 0.022837873548269272, + 0.012932635843753815, + 0.020151114091277122, + 0.013043488375842571, + 0.007773499935865402, + 0.011074467562139034, + 0.01019138004630804, + 0.010393856093287468, + 0.016072502359747887, + 0.011331923305988312, + 0.01694161631166935, + 0.008505410514771938, + 0.00975767057389021, + 0.011451300233602524, + 0.012985975481569767, + 0.011314531788229942, + 0.014134990982711315, + 0.01788962259888649, + 0.01041446253657341, + 0.010139844380319118, + 0.005456719547510147, + 0.007953470572829247, + 0.00823940522968769, + 0.01039494201540947, + 0.01123618334531784, + 0.009327836334705353, + 0.011481057852506638, + 0.0066039361990988255, + 0.01621556095778942, + 0.009689037688076496, + 0.02327873930335045, + 0.011953341774642467, + 0.009164021350443363, + 0.012607122771441936, + 0.013610685244202614, + 0.015830032527446747, + 0.011734504252672195, + 0.019828874617815018, + 0.014180545695126057, + 0.01563285104930401, + 0.02249237149953842, + 0.014713598415255547, + 0.019929666072130203, + 0.009927013888955116, + 0.004729935899376869, + 0.011548466980457306, + 0.007233649492263794, + 0.009297415614128113, + 0.01022951677441597, + 0.00409685866907239, + 0.0114308325573802, + 0.0056082867085933685, + 0.010973104275763035, + 0.012481648474931717, + 0.013562929816544056, + 0.014176107943058014, + 0.010157003998756409, + 0.010430825874209404, + 0.013792429119348526, + 0.018300987780094147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.006746253930032253 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863473117351532 + ], + "opposite_logprob_tokens": [ + -4.758634567260742 + ], + "logprob_reward": -0.00863473117351532, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863473117351532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14063841104507446, + 0.19493497908115387, + 0.1325133889913559, + 0.07675226777791977, + 0.09138564020395279, + 0.04312926158308983, + 0.13315966725349426, + 0.06550182402133942, + 0.028883671388030052, + 0.09310087561607361 + ], + "credit_map": [ + 0.014858128502964973, + 0.017037250101566315, + 0.018888695165514946, + 0.018236566334962845, + 0.016657929867506027, + 0.01865355856716633, + 0.018667617812752724, + 0.017638668417930603, + 0.02393784373998642, + 0.019465865567326546, + 0.02990233153104782, + 0.024025633931159973, + 0.02115158550441265, + 0.02670464664697647, + 0.02336454577744007, + 0.02638252265751362, + 0.01844104565680027, + 0.012746292166411877, + 0.02016187645494938, + 0.0162296611815691, + 0.019207168370485306, + 0.01745045930147171, + 0.01649188995361328, + 0.011785004287958145, + 0.008268202655017376, + 0.010008217766880989, + 0.008505648002028465, + 0.006794349290430546, + 0.012693083845078945, + 0.009571786038577557, + 0.013390956446528435, + 0.0075200279243290424, + 0.011488666757941246, + 0.013050227425992489, + 0.012820879928767681, + 0.008077163249254227, + 0.01401180773973465, + 0.011614878661930561, + 0.010262553580105305, + 0.010059472173452377, + 0.004778007045388222, + 0.0032466345001012087, + 0.005875198170542717, + 0.007208830676972866, + 0.004346869885921478, + 0.005193730816245079, + 0.006508463993668556, + 0.005971524864435196, + 0.017291119322180748, + 0.012858852744102478, + 0.018264859914779663, + 0.01997455768287182, + 0.014270416460931301, + 0.01908513717353344, + 0.018830616027116776, + 0.012584103271365166, + 0.013166896067559719, + 0.0075950478203594685, + 0.007907707244157791, + 0.005259068217128515, + 0.0078762611374259, + 0.007481545675545931, + 0.01206677220761776, + 0.004148522391915321, + 0.002862921915948391, + 0.0032009391579777002, + 0.0037183749955147505, + 0.00406614039093256, + 0.0033426894806325436, + 0.0032423019874840975, + 0.0037197584751993418, + 0.00473054638132453, + 0.009555132128298283, + 0.010465923696756363, + 0.012538589537143707, + 0.011498701758682728, + 0.009260478429496288, + 0.01214276347309351, + 0.009074836038053036, + 0.018564455211162567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.00863473117351532 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305654764175415 + ], + "opposite_logprob_tokens": [ + -3.1680564880371094 + ], + "logprob_reward": -0.04305654764175415, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305654764175415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1112130656838417, + 0.15436218678951263, + 0.046390704810619354, + 0.12420977652072906, + 0.07325480133295059, + 0.18945249915122986, + 0.1553449183702469, + 0.04976677522063255, + 0.031032012775540352, + 0.06497325748205185 + ], + "credit_map": [ + 0.00896380003541708, + 0.009621940553188324, + 0.009406257420778275, + 0.017268430441617966, + 0.013217059895396233, + 0.01377443503588438, + 0.021290656179189682, + 0.01767049729824066, + 0.011215315200388432, + 0.013671554625034332, + 0.01779634691774845, + 0.020776433870196342, + 0.017210016027092934, + 0.024425921961665154, + 0.01982201635837555, + 0.02944459393620491, + 0.005710558034479618, + 0.004989867564290762, + 0.005283890757709742, + 0.005080106668174267, + 0.009117490611970425, + 0.0046888720244169235, + 0.006005666218698025, + 0.005514258053153753, + 0.009480439126491547, + 0.009750179946422577, + 0.015839600935578346, + 0.014554367400705814, + 0.013248459435999393, + 0.01889939233660698, + 0.027438871562480927, + 0.014998478814959526, + 0.011564183048903942, + 0.009721627458930016, + 0.011079459451138973, + 0.007447741460055113, + 0.00814956333488226, + 0.009390086866915226, + 0.007454620208591223, + 0.008447525091469288, + 0.015448547899723053, + 0.01966612972319126, + 0.028242798522114754, + 0.02064589411020279, + 0.02488357201218605, + 0.021018197759985924, + 0.035728808492422104, + 0.02381855994462967, + 0.019719544798135757, + 0.014735892415046692, + 0.03245172277092934, + 0.029888631775975227, + 0.015169738791882992, + 0.012556137517094612, + 0.016062689945101738, + 0.014760570600628853, + 0.007820811122655869, + 0.007198883220553398, + 0.007752621080726385, + 0.004714875482022762, + 0.006795426364988089, + 0.004800397902727127, + 0.006616082973778248, + 0.004067680798470974, + 0.002161232056096196, + 0.004621663596481085, + 0.004167946055531502, + 0.0038380352780222893, + 0.004977153614163399, + 0.0020115782972425222, + 0.005543538834899664, + 0.003710868302732706, + 0.006827508565038443, + 0.008541918359696865, + 0.008049102500081062, + 0.010688140988349915, + 0.006261198781430721, + 0.006449038162827492, + 0.006884189322590828, + 0.011272160336375237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.04305654764175415 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a one-horned beast resembling the Nian with dark red metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beast possess a massive spiraling obsidian horn with a red glowing tip and blood-red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow ancient street lined with wooden structures and hundreds of round red lanterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track closely to the ground, racing with the beast and capturing splashes of water and sparks on the pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the red lanterns and their golden tassels shown swaying violently due to the fierce winds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47429022192955017 + ], + "opposite_logprob_tokens": [ + -0.9742902517318726 + ], + "logprob_reward": -0.47429022192955017, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47429022192955017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22768034040927887, + 0.11781951040029526, + 0.08136925846338272, + 0.101987324655056, + 0.05694127827882767, + 0.11692579835653305, + 0.08312898129224777, + 0.07301248610019684, + 0.04662800580263138, + 0.09450704604387283 + ], + "credit_map": [ + 0.018575910478830338, + 0.016726085916161537, + 0.021105553954839706, + 0.03139540180563927, + 0.029785579070448875, + 0.025512313470244408, + 0.05737581104040146, + 0.027203692123293877, + 0.01264981459826231, + 0.015299922786653042, + 0.012343840673565865, + 0.014542306773364544, + 0.013004690408706665, + 0.01724325865507126, + 0.016091488301753998, + 0.0166441909968853, + 0.009687582030892372, + 0.008277026005089283, + 0.010195639915764332, + 0.011467142961919308, + 0.011221750639379025, + 0.00929985661059618, + 0.013082546181976795, + 0.008137717843055725, + 0.010027803480625153, + 0.011793578043580055, + 0.011518517509102821, + 0.013246522285044193, + 0.01855318807065487, + 0.012756834737956524, + 0.01470564492046833, + 0.009385231882333755, + 0.004077592398971319, + 0.005434578750282526, + 0.005797992460429668, + 0.004879145883023739, + 0.006537122651934624, + 0.011803587898612022, + 0.007373045664280653, + 0.011038213036954403, + 0.012244274839758873, + 0.010156365111470222, + 0.013726281002163887, + 0.013696094043552876, + 0.015170732513070107, + 0.012661950662732124, + 0.025800935924053192, + 0.013469158671796322, + 0.011371076107025146, + 0.008922741748392582, + 0.013945703394711018, + 0.009227334521710873, + 0.006592089775949717, + 0.014552746899425983, + 0.00770083861425519, + 0.010816450230777264, + 0.010593638755381107, + 0.009157645516097546, + 0.006504634395241737, + 0.006424333434551954, + 0.009855867363512516, + 0.0063773817382752895, + 0.017859183251857758, + 0.006239807233214378, + 0.006101020611822605, + 0.005394537001848221, + 0.008546177297830582, + 0.0040404764004051685, + 0.00528846075758338, + 0.005956614390015602, + 0.0053900862112641335, + 0.005910634063184261, + 0.010171044617891312, + 0.015086133033037186, + 0.01322232000529766, + 0.011543777771294117, + 0.008775686845183372, + 0.011629387736320496, + 0.01024674717336893, + 0.013831952586770058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.47429022192955017 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_10/index_tensor([6839, 6839])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_10/index_tensor([6839, 6839])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1ea68f22021c28a73a8faf6ffc20b0d923ef91f9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_10/index_tensor([6839, 6839])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up, static shot shot from a level angle, captured with a calm, restrained documentary style, precisely captures the moment of an Asian woman's emotional breakdown. She has long, straight, black hair, strands of which are messy and plastered to her cheeks with tears. Her eyes are red and swollen from prolonged crying, her brow furrowed, her lips slightly parted, her lower lip trembling as she struggles to suppress sobs. Large, glistening tears flow uncontrollably from her reddened eyes, silently sliding down her plain face, leaving streaks of wet light. Her shoulders tremble slightly from her uncontrollable sobs. She wears a simple gray turtleneck sweater, the soft material contrasting with her fragile emotions. The background is a simple, blurred interior space, and soft natural light filtering in from one side of the frame delicately outlines her facial features and makes each tear drop stand out with exceptional clarity. The whole picture has a cold tone and no dramatic camera movements. It just quietly and non-judgmentally records this primitive and deep sadness, creating a heartbreaking sense of reality and an immersive atmosphere of empathy.", + "seed": 26288, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_10/index_6839_seed_26288_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021956999320536852, + "vlm_logprob": -0.021956999320536852, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.021956999320536852, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009711874648928642 + ], + "opposite_logprob_tokens": [ + -4.759711742401123 + ], + "logprob_reward": -0.009711874648928642, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.009711874648928642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15251708030700684, + 0.25014010071754456, + 0.08163584768772125, + 0.018628068268299103, + 0.03257843852043152, + 0.0283560398966074, + 0.11961684376001358, + 0.05420806258916855, + 0.21385154128074646, + 0.048467930406332016 + ], + "credit_map": [ + 0.01988879404962063, + 0.010730685666203499, + 0.024085544049739838, + 0.013866739347577095, + 0.024075739085674286, + 0.016566062346100807, + 0.018804343417286873, + 0.024499190971255302, + 0.04419899731874466, + 0.0084311468526721, + 0.012712325900793076, + 0.012706472538411617, + 0.016268424689769745, + 0.10249527543783188, + 0.01870560646057129, + 0.03462187945842743, + 0.006515786051750183, + 0.006747904699295759, + 0.006593463942408562, + 0.007036725524812937, + 0.014950700104236603, + 0.010759113356471062, + 0.012120614759624004, + 0.016911553218960762, + 0.0018977908184751868, + 0.00240412843413651, + 0.002376620890572667, + 0.0017281455220654607, + 0.0017581151332706213, + 0.002717739436775446, + 0.0021928884088993073, + 0.003552641486749053, + 0.002746986923739314, + 0.0031521241180598736, + 0.004800927359610796, + 0.003603570396080613, + 0.00427285535261035, + 0.0037634416949003935, + 0.005144596099853516, + 0.005093940068036318, + 0.003372787730768323, + 0.002392685739323497, + 0.003239575307816267, + 0.0018612490966916084, + 0.0032502401154488325, + 0.004341654013842344, + 0.0029051813762634993, + 0.006992669776082039, + 0.013506032526493073, + 0.005901642609387636, + 0.014373202808201313, + 0.006474189460277557, + 0.007325673010200262, + 0.027459658682346344, + 0.010041942819952965, + 0.0345345139503479, + 0.0029876818880438805, + 0.003960530739277601, + 0.0024757222272455692, + 0.0029436589684337378, + 0.003607705933973193, + 0.01591322384774685, + 0.005269851069897413, + 0.017049692571163177, + 0.021688055247068405, + 0.03245924413204193, + 0.024285253137350082, + 0.027617720887064934, + 0.028683671727776527, + 0.019667552784085274, + 0.025172345340251923, + 0.034277696162462234, + 0.004707622807472944, + 0.004067964386194944, + 0.0051291645504534245, + 0.0029807724058628082, + 0.006558707915246487, + 0.008481909520924091, + 0.004563448950648308, + 0.01197834312915802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.009711874648928642 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983858808875084 + ], + "opposite_logprob_tokens": [ + -3.5298385620117188 + ], + "logprob_reward": -0.02983858808875084, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983858808875084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20907798409461975, + 0.2570439577102661, + 0.09029309451580048, + 0.035094231367111206, + 0.03159957006573677, + 0.02529088221490383, + 0.1139555275440216, + 0.06749466061592102, + 0.11939925700426102, + 0.05075083300471306 + ], + "credit_map": [ + 0.017508884891867638, + 0.013497618027031422, + 0.02024620957672596, + 0.020097332075238228, + 0.022101568058133125, + 0.025219328701496124, + 0.05150613933801651, + 0.03890090808272362, + 0.020620757713913918, + 0.01908540353178978, + 0.017799535766243935, + 0.025188203901052475, + 0.0639980360865593, + 0.02528013288974762, + 0.06453084945678711, + 0.020541034638881683, + 0.009148347191512585, + 0.00672796368598938, + 0.009891646914184093, + 0.008130494505167007, + 0.01737322472035885, + 0.01130517665296793, + 0.011900787241756916, + 0.015815455466508865, + 0.0022678314708173275, + 0.0037710960023105145, + 0.0049591753631830215, + 0.005083285737782717, + 0.005122319329530001, + 0.0043901377357542515, + 0.004327882081270218, + 0.005172505043447018, + 0.002834263490512967, + 0.002938818419352174, + 0.0049727498553693295, + 0.0033150387462228537, + 0.004255297128111124, + 0.002971096895635128, + 0.006351750809699297, + 0.0039605554193258286, + 0.0034151896834373474, + 0.0031826805789023638, + 0.0033455060329288244, + 0.0018520227167755365, + 0.0020129450131207705, + 0.003252737456932664, + 0.002642469946295023, + 0.005587329156696796, + 0.010648920200765133, + 0.005989656317979097, + 0.009238338097929955, + 0.006034926977008581, + 0.007662852294743061, + 0.030462326481938362, + 0.008787586353719234, + 0.035130929201841354, + 0.0033981341402977705, + 0.0024703615345060825, + 0.0038067158311605453, + 0.00315691321156919, + 0.003593648551031947, + 0.022893289104104042, + 0.005710553843528032, + 0.022465044632554054, + 0.008535253815352917, + 0.006613820791244507, + 0.00770593760535121, + 0.005894004367291927, + 0.01302428264170885, + 0.03218209743499756, + 0.013528071343898773, + 0.03191579133272171, + 0.005394838284701109, + 0.0036951531656086445, + 0.003337809583172202, + 0.003716372884809971, + 0.010503190569579601, + 0.007126791402697563, + 0.005918351002037525, + 0.011058329604566097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.02983858808875084 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054971784353256226 + ], + "opposite_logprob_tokens": [ + -2.929971694946289 + ], + "logprob_reward": -0.054971784353256226, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054971784353256226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07042999565601349, + 0.26225584745407104, + 0.09692852944135666, + 0.05849633738398552, + 0.029357535764575005, + 0.091749407351017, + 0.20711109042167664, + 0.12021519988775253, + 0.041384316980838776, + 0.022071706131100655 + ], + "credit_map": [ + 0.008275767788290977, + 0.00457737036049366, + 0.0076703294180333614, + 0.005695562344044447, + 0.009579028934240341, + 0.00703155854716897, + 0.017718380317091942, + 0.00988200306892395, + 0.043924182653427124, + 0.01513817347586155, + 0.019782688468694687, + 0.011559770442545414, + 0.021724101155996323, + 0.09798997640609741, + 0.01799747720360756, + 0.03413951396942139, + 0.012349504977464676, + 0.007235064171254635, + 0.009343440644443035, + 0.009442318230867386, + 0.014877930283546448, + 0.013557053171098232, + 0.012067063711583614, + 0.018056156113743782, + 0.006991272792220116, + 0.00618728855624795, + 0.007489798124879599, + 0.005528876092284918, + 0.006651327013969421, + 0.007772487122565508, + 0.008697913028299809, + 0.00917738489806652, + 0.002832856960594654, + 0.0018226144602522254, + 0.002482149051502347, + 0.003240118734538555, + 0.004637009464204311, + 0.004678602330386639, + 0.005405912641435862, + 0.004258276894688606, + 0.00874595157802105, + 0.009720495902001858, + 0.014117094688117504, + 0.007135453168302774, + 0.011665665544569492, + 0.01097197737544775, + 0.011237481608986855, + 0.018155300989747047, + 0.02799782156944275, + 0.01511941198259592, + 0.04528740420937538, + 0.015528121963143349, + 0.020249614492058754, + 0.02536659687757492, + 0.02216164581477642, + 0.03540048375725746, + 0.005210590083152056, + 0.0049278647638857365, + 0.005098307970911264, + 0.003359556198120117, + 0.0046183946542441845, + 0.046104058623313904, + 0.006903421599417925, + 0.0439930222928524, + 0.004875230602920055, + 0.005283754784613848, + 0.003990558907389641, + 0.004123501013964415, + 0.005214260425418615, + 0.0047701275907456875, + 0.004927413538098335, + 0.008199476636946201, + 0.00198428169824183, + 0.002288914518430829, + 0.0024570743553340435, + 0.0021604406647384167, + 0.0038023998495191336, + 0.0030261133797466755, + 0.00314345583319664, + 0.0032090272288769484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.054971784353256226 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763498293235898 + ], + "opposite_logprob_tokens": [ + -4.882635116577148 + ], + "logprob_reward": -0.00763498293235898, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763498293235898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07818948477506638, + 0.16369536519050598, + 0.14638911187648773, + 0.17297835648059845, + 0.0670475885272026, + 0.08442793041467667, + 0.06453043967485428, + 0.09984125196933746, + 0.08082269877195358, + 0.04207778722047806 + ], + "credit_map": [ + 0.00952029787003994, + 0.005126055330038071, + 0.008597202599048615, + 0.005650594364851713, + 0.00838868785649538, + 0.011193033307790756, + 0.010899007320404053, + 0.018814604729413986, + 0.02864130772650242, + 0.006526956334710121, + 0.011402430944144726, + 0.007970933802425861, + 0.013479338027536869, + 0.05901001766324043, + 0.015466957353055477, + 0.02119741588830948, + 0.016106555238366127, + 0.013595843687653542, + 0.018662981688976288, + 0.011498888023197651, + 0.029309500008821487, + 0.018184510990977287, + 0.01597333699464798, + 0.0230574831366539, + 0.02760755456984043, + 0.014138500206172466, + 0.018069634214043617, + 0.01566517911851406, + 0.01454172469675541, + 0.02773303911089897, + 0.018608998507261276, + 0.03661371394991875, + 0.008988833986222744, + 0.005589181557297707, + 0.013630388304591179, + 0.0052263857796788216, + 0.007851963862776756, + 0.007834780029952526, + 0.008026276715099812, + 0.009899768978357315, + 0.009370687417685986, + 0.01174292154610157, + 0.012276038527488708, + 0.006679178215563297, + 0.006420869380235672, + 0.011290719732642174, + 0.009244986809790134, + 0.0174025297164917, + 0.0040442682802677155, + 0.003974623512476683, + 0.008165115490555763, + 0.004387896042317152, + 0.005925805773586035, + 0.015791025012731552, + 0.005950910970568657, + 0.01629078947007656, + 0.013827254064381123, + 0.011256443336606026, + 0.008196963928639889, + 0.00870506651699543, + 0.011259312741458416, + 0.01711222343146801, + 0.014072895050048828, + 0.01541108638048172, + 0.006593985948711634, + 0.0071798088029026985, + 0.005754042882472277, + 0.007800782565027475, + 0.010938423685729504, + 0.01637502759695053, + 0.011639358475804329, + 0.014541268348693848, + 0.005190820898860693, + 0.002727619605138898, + 0.0048707276582717896, + 0.0027708993293344975, + 0.004676468204706907, + 0.006241761613637209, + 0.005498600658029318, + 0.01010088250041008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.00763498293235898 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627766579389572 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007627766579389572, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627766579389572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06938669085502625, + 0.27837541699409485, + 0.09234431385993958, + 0.06472375243902206, + 0.04804561659693718, + 0.05593082681298256, + 0.2216353416442871, + 0.029590126127004623, + 0.07460761815309525, + 0.06536029279232025 + ], + "credit_map": [ + 0.007016549818217754, + 0.006047068629413843, + 0.0073977247811853886, + 0.006540329195559025, + 0.01060075219720602, + 0.006897849030792713, + 0.013065876439213753, + 0.011820544488728046, + 0.047538649290800095, + 0.014047560282051563, + 0.02274426445364952, + 0.012377812527120113, + 0.019543003290891647, + 0.10612857341766357, + 0.021114561706781387, + 0.034880973398685455, + 0.008455375209450722, + 0.008201603777706623, + 0.009553845971822739, + 0.007431888021528721, + 0.015416767448186874, + 0.012036288157105446, + 0.013193284161388874, + 0.01805526204407215, + 0.007278006989508867, + 0.005484464578330517, + 0.005716747138649225, + 0.004875813610851765, + 0.006545842159539461, + 0.01446566078811884, + 0.005763113498687744, + 0.014594107866287231, + 0.004429906141012907, + 0.0036629594396799803, + 0.0062502166256308556, + 0.0030599029269069433, + 0.004397408105432987, + 0.008985490538179874, + 0.005831200163811445, + 0.011428531259298325, + 0.006694346200674772, + 0.007229072041809559, + 0.008706136606633663, + 0.004289919510483742, + 0.0055113984271883965, + 0.007241231855005026, + 0.004859895911067724, + 0.011398828588426113, + 0.020906155928969383, + 0.009775757789611816, + 0.019526397809386253, + 0.010777909308671951, + 0.013075157068669796, + 0.05943106487393379, + 0.01797463372349739, + 0.07016827911138535, + 0.002508483361452818, + 0.0024025558959692717, + 0.002161596901714802, + 0.002579432213678956, + 0.003823558334261179, + 0.006757812108844519, + 0.003056945512071252, + 0.006299741566181183, + 0.01287272572517395, + 0.007025971543043852, + 0.00580116081982851, + 0.005079973954707384, + 0.007121854927390814, + 0.008251032792031765, + 0.009736317209899426, + 0.018718579784035683, + 0.007904751226305962, + 0.007264663930982351, + 0.00851291511207819, + 0.006054211407899857, + 0.00933496281504631, + 0.006707894150167704, + 0.008402262814342976, + 0.011178635060787201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.007627766579389572 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot shot from a level angle, captured with a calm, restrained documentary style, precisely captures the moment of an Asian woman's emotional breakdown. She has long, straight, black hair, strands of which are messy and plastered to her cheeks with tears. Her eyes are red and swollen from prolonged crying, her brow furrowed, her lips slightly parted, her lower lip trembling as she struggles to suppress sobs. Large, glistening tears flow uncontrollably from her reddened eyes, silently sliding down her plain face, leaving streaks of wet light. Her shoulders tremble slightly from her uncontrollable sobs. She wears a simple gray turtleneck sweater, the soft material contrasting with her fragile emotions. The background is a simple, blurred interior space, and soft natural light filtering in from one side of the frame delicately outlines her facial features and makes each tear drop stand out with exceptional clarity. The whole picture has a cold tone and no dramatic camera movements. It just quietly and non-judgmentally records this primitive and deep sadness, creating a heartbreaking sense of reality and an immersive atmosphere of empathy.", + "seed": 26289, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_10/index_6839_seed_26289_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.015943042794242503, + "vlm_logprob": -0.015943042794242503, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.015943042794242503, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009375827386975288 + ], + "opposite_logprob_tokens": [ + -4.759376049041748 + ], + "logprob_reward": -0.009375827386975288, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.009375827386975288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4712525010108948, + 0.046720195561647415, + 0.02775687724351883, + 0.1401013433933258, + 0.04471570625901222, + 0.031225044280290604, + 0.05097579210996628, + 0.016100142151117325, + 0.13771605491638184, + 0.033436283469200134 + ], + "credit_map": [ + 0.03613262251019478, + 0.02255227230489254, + 0.102695994079113, + 0.020275412127375603, + 0.032689280807971954, + 0.1878562718629837, + 0.014474697411060333, + 0.05457594618201256, + 0.005709873512387276, + 0.004155952483415604, + 0.007629286032170057, + 0.005299656186252832, + 0.006227400619536638, + 0.004871907643973827, + 0.006281939335167408, + 0.00654418021440506, + 0.002939813071861863, + 0.0028757157269865274, + 0.00391969783231616, + 0.0024596047587692738, + 0.0056540812365710735, + 0.002715533133596182, + 0.003201327985152602, + 0.00399110559374094, + 0.020539361983537674, + 0.004495060071349144, + 0.041192930191755295, + 0.007553938776254654, + 0.007106407545506954, + 0.007950377650558949, + 0.007930220104753971, + 0.043333057314157486, + 0.0073738363571465015, + 0.004285187926143408, + 0.00871459860354662, + 0.006438448093831539, + 0.004628670867532492, + 0.0032046129927039146, + 0.006139396224170923, + 0.003930955659598112, + 0.006891151424497366, + 0.003148831659927964, + 0.006950072944164276, + 0.0022538998164236546, + 0.00340504408814013, + 0.0030272568110376596, + 0.0022972305305302143, + 0.003251557471230626, + 0.0067597744055092335, + 0.005269203335046768, + 0.007268471643328667, + 0.004818948917090893, + 0.006948171183466911, + 0.007339866366237402, + 0.004440158139914274, + 0.008131196722388268, + 0.0014235026901587844, + 0.0013398625887930393, + 0.0019719465635716915, + 0.0015535601414740086, + 0.0031887791119515896, + 0.0020353072322905064, + 0.0023644822649657726, + 0.0022227021399885416, + 0.013886639848351479, + 0.013477493077516556, + 0.016300804913043976, + 0.017449740320444107, + 0.017725801095366478, + 0.02060822769999504, + 0.017345082014799118, + 0.020922265946865082, + 0.004000050015747547, + 0.0029305564239621162, + 0.004906635731458664, + 0.003467443399131298, + 0.0051398733630776405, + 0.003428901545703411, + 0.005956016480922699, + 0.0036068048793822527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.009375827386975288 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007654029410332441 + ], + "opposite_logprob_tokens": [ + -4.882654190063477 + ], + "logprob_reward": -0.007654029410332441, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007654029410332441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2048293948173523, + 0.11633294820785522, + 0.11486409604549408, + 0.16321749985218048, + 0.10305777192115784, + 0.0871753990650177, + 0.03259347006678581, + 0.06143519654870033, + 0.10202057659626007, + 0.014473631046712399 + ], + "credit_map": [ + 0.016342854127287865, + 0.01127245370298624, + 0.03935673087835312, + 0.011609162203967571, + 0.021791070699691772, + 0.06395956128835678, + 0.019790207967162132, + 0.02070736698806286, + 0.014083530753850937, + 0.009942133910953999, + 0.015567188151180744, + 0.010820094496011734, + 0.017115788534283638, + 0.027223821729421616, + 0.008564787916839123, + 0.013015597127377987, + 0.012050765566527843, + 0.013327883556485176, + 0.016623975709080696, + 0.010858234949409962, + 0.017515268176794052, + 0.009804056957364082, + 0.01450361218303442, + 0.020180298015475273, + 0.028343234211206436, + 0.005588872358202934, + 0.045644767582416534, + 0.008839125744998455, + 0.01187289971858263, + 0.01106950081884861, + 0.011948599480092525, + 0.03991049528121948, + 0.01426917128264904, + 0.009957018308341503, + 0.017122844234108925, + 0.01108488067984581, + 0.009434346109628677, + 0.013694997876882553, + 0.012749181129038334, + 0.014745336025953293, + 0.019032442942261696, + 0.00879690982401371, + 0.018107421696186066, + 0.007298536133021116, + 0.008572946302592754, + 0.007390056271106005, + 0.010331233032047749, + 0.007645850069820881, + 0.004307945258915424, + 0.003217309480533004, + 0.0048890383914113045, + 0.0038981474936008453, + 0.004970095120370388, + 0.0025347594637423754, + 0.003589199623093009, + 0.005186975933611393, + 0.005007559433579445, + 0.007537398487329483, + 0.007005707826465368, + 0.00997836235910654, + 0.007699199486523867, + 0.006273729261010885, + 0.01158161647617817, + 0.006351620890200138, + 0.014264417812228203, + 0.009739555418491364, + 0.016343804076313972, + 0.009253773838281631, + 0.013983374461531639, + 0.015152747742831707, + 0.009612488560378551, + 0.013670414686203003, + 0.002211695769801736, + 0.0015957392752170563, + 0.0023200928699225187, + 0.0013319978024810553, + 0.0019788143690675497, + 0.001388933276757598, + 0.0014877307694405317, + 0.0021586264483630657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.007654029410332441 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054972801357507706 + ], + "opposite_logprob_tokens": [ + -2.9299728870391846 + ], + "logprob_reward": -0.054972801357507706, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054972801357507706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5167055726051331, + 0.16407552361488342, + 0.02066713757812977, + 0.1234142854809761, + 0.06637563556432724, + 0.009063267149031162, + 0.02108238823711872, + 0.04195763170719147, + 0.02097848616540432, + 0.015680059790611267 + ], + "credit_map": [ + 0.03690207004547119, + 0.021105438470840454, + 0.11537856608629227, + 0.02161904238164425, + 0.03019910305738449, + 0.21229228377342224, + 0.018103228881955147, + 0.061105843633413315, + 0.01885801926255226, + 0.015283026732504368, + 0.025582054629921913, + 0.022535869851708412, + 0.017927046865224838, + 0.02352563664317131, + 0.019283654168248177, + 0.02108020894229412, + 0.0020111696794629097, + 0.0011485242284834385, + 0.002850053133442998, + 0.0014752408023923635, + 0.004605703987181187, + 0.0021038183476775885, + 0.0031117061153054237, + 0.0033609208185225725, + 0.019157538190484047, + 0.0038420904893428087, + 0.03589221462607384, + 0.006274957675486803, + 0.007299843709915876, + 0.007967246696352959, + 0.006903139408677816, + 0.03607725352048874, + 0.013078272342681885, + 0.005844702012836933, + 0.013634460046887398, + 0.005747206974774599, + 0.008873244747519493, + 0.00538315623998642, + 0.007526558358222246, + 0.006288035307079554, + 0.0013109607389196754, + 0.0007666977471671999, + 0.001474336488172412, + 0.0006755201611667871, + 0.00139669852796942, + 0.0011371590662747622, + 0.0011402415111660957, + 0.0011616526171565056, + 0.002862604334950447, + 0.0021887212060391903, + 0.002807110548019409, + 0.0018991244724020362, + 0.0030581210739910603, + 0.002566936891525984, + 0.002543567679822445, + 0.0031562037765979767, + 0.0035585174337029457, + 0.004802049603313208, + 0.00530635891482234, + 0.007553302217274904, + 0.004658094607293606, + 0.004100257996469736, + 0.006822264287620783, + 0.005156788043677807, + 0.0024632031563669443, + 0.0022356403060257435, + 0.002577232662588358, + 0.0026799372863024473, + 0.002373536815866828, + 0.002171096159145236, + 0.003087559947744012, + 0.00339027913287282, + 0.0018492087256163359, + 0.0018250243738293648, + 0.002332051284611225, + 0.00202861032448709, + 0.001779915182851255, + 0.0010383082553744316, + 0.002383582526817918, + 0.0024433594662696123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.054972801357507706 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096688237041235 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.004096688237041235, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096688237041235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.631527841091156, + 0.04503316804766655, + 0.027787987142801285, + 0.09279577434062958, + 0.06378569453954697, + 0.02759474888443947, + 0.022897133603692055, + 0.0426492840051651, + 0.02738569863140583, + 0.018542665988206863 + ], + "credit_map": [ + 0.04563980922102928, + 0.02487029694020748, + 0.1394995003938675, + 0.02520078979432583, + 0.03348365053534508, + 0.2690642774105072, + 0.01710747554898262, + 0.07666203379631042, + 0.005838869139552116, + 0.004065205343067646, + 0.006575231906026602, + 0.005182641092687845, + 0.007161914370954037, + 0.005491242744028568, + 0.00557812349870801, + 0.005139937158674002, + 0.0029850308783352375, + 0.002662614220753312, + 0.004524098243564367, + 0.0023346226662397385, + 0.003926396369934082, + 0.0031989102717489004, + 0.0025795248802751303, + 0.00557679170742631, + 0.01406164187937975, + 0.003361037466675043, + 0.025911087170243263, + 0.004916992969810963, + 0.00491087744012475, + 0.006670042406767607, + 0.006455305498093367, + 0.026508791372179985, + 0.009221477434039116, + 0.006223896984010935, + 0.00990122091025114, + 0.0065805488266050816, + 0.007279324345290661, + 0.009211359545588493, + 0.0063247885555028915, + 0.009043081663548946, + 0.0030879268888384104, + 0.0027679179329425097, + 0.003955861087888479, + 0.0035613614600151777, + 0.0035076162312179804, + 0.004876534920185804, + 0.003302778583019972, + 0.0025347527116537094, + 0.002655991818755865, + 0.002572860335931182, + 0.0036264003720134497, + 0.002194038825109601, + 0.0037185135297477245, + 0.0023218372371047735, + 0.003073262283578515, + 0.0027342282701283693, + 0.003908307291567326, + 0.004475501365959644, + 0.005218409467488527, + 0.00723353261128068, + 0.004348136018961668, + 0.004856582265347242, + 0.006520902272313833, + 0.00608791271224618, + 0.003465411951765418, + 0.0035173336509615183, + 0.003535263007506728, + 0.003163821529597044, + 0.003483050037175417, + 0.0036946884356439114, + 0.00344813778065145, + 0.0030779934022575617, + 0.0027479310519993305, + 0.0016451627016067505, + 0.0030541059095412493, + 0.002052393974736333, + 0.0022712647914886475, + 0.0011954852379858494, + 0.00326828146353364, + 0.002308040391653776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.004096688237041235 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615867579355836 + ], + "opposite_logprob_tokens": [ + -5.628615856170654 + ], + "logprob_reward": -0.003615867579355836, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615867579355836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.269176185131073, + 0.03423956036567688, + 0.037506379187107086, + 0.30001500248908997, + 0.0864255428314209, + 0.0914212241768837, + 0.03447870537638664, + 0.07885771989822388, + 0.05523575842380524, + 0.012643917463719845 + ], + "credit_map": [ + 0.018931182101368904, + 0.011406566016376019, + 0.049671825021505356, + 0.011523211374878883, + 0.0158444382250309, + 0.11419510841369629, + 0.009116364642977715, + 0.0384875163435936, + 0.004340250510722399, + 0.00397438695654273, + 0.004227025900036097, + 0.002347020199522376, + 0.008503515273332596, + 0.00387382204644382, + 0.004223393276333809, + 0.0027501527220010757, + 0.0038833743892610073, + 0.002972680376842618, + 0.007199566345661879, + 0.0031044892966747284, + 0.007127181626856327, + 0.002672371920198202, + 0.0047220527194440365, + 0.0058246636763215065, + 0.0476662777364254, + 0.008596527390182018, + 0.08964040130376816, + 0.01491653360426426, + 0.01482371985912323, + 0.017772702500224113, + 0.014943660236895084, + 0.09165520966053009, + 0.009800593368709087, + 0.010828706435859203, + 0.01280565932393074, + 0.009172158315777779, + 0.01456938311457634, + 0.01115424558520317, + 0.009033583104610443, + 0.00906122662127018, + 0.017942026257514954, + 0.008604948408901691, + 0.014972387813031673, + 0.008335094898939133, + 0.010053496807813644, + 0.01108913030475378, + 0.009585683234035969, + 0.01083846390247345, + 0.006081051658838987, + 0.0031998688355088234, + 0.005737839266657829, + 0.002667850349098444, + 0.0030860428232699633, + 0.0028939563781023026, + 0.005559450946748257, + 0.0052526467479765415, + 0.005832134746015072, + 0.007094644941389561, + 0.010477283038198948, + 0.012992151081562042, + 0.009141085669398308, + 0.017236512154340744, + 0.00930042564868927, + 0.006783491466194391, + 0.007652915082871914, + 0.0050816526636481285, + 0.008271654136478901, + 0.006071256007999182, + 0.008611021563410759, + 0.007324483245611191, + 0.006114406511187553, + 0.00610837247222662, + 0.0014729989925399423, + 0.0012350308243185282, + 0.0018219643970951438, + 0.0016164117259904742, + 0.0013327323831617832, + 0.0012131549883633852, + 0.0022041057236492634, + 0.0017475190106779337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.003615867579355836 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot shot from a level angle, captured with a calm, restrained documentary style, precisely captures the moment of an Asian woman's emotional breakdown. She has long, straight, black hair, strands of which are messy and plastered to her cheeks with tears. Her eyes are red and swollen from prolonged crying, her brow furrowed, her lips slightly parted, her lower lip trembling as she struggles to suppress sobs. Large, glistening tears flow uncontrollably from her reddened eyes, silently sliding down her plain face, leaving streaks of wet light. Her shoulders tremble slightly from her uncontrollable sobs. She wears a simple gray turtleneck sweater, the soft material contrasting with her fragile emotions. The background is a simple, blurred interior space, and soft natural light filtering in from one side of the frame delicately outlines her facial features and makes each tear drop stand out with exceptional clarity. The whole picture has a cold tone and no dramatic camera movements. It just quietly and non-judgmentally records this primitive and deep sadness, creating a heartbreaking sense of reality and an immersive atmosphere of empathy.", + "seed": 26290, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_10/index_6839_seed_26290_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.030524871544912457, + "vlm_logprob": -0.030524871544912457, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.030524871544912457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01893516071140766 + ], + "opposite_logprob_tokens": [ + -4.018935203552246 + ], + "logprob_reward": -0.01893516071140766, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01893516071140766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14137990772724152, + 0.04985799640417099, + 0.053500112146139145, + 0.03918039798736572, + 0.06274247169494629, + 0.08287946879863739, + 0.04054759442806244, + 0.46907228231430054, + 0.02827858179807663, + 0.03256120905280113 + ], + "credit_map": [ + 0.016047203913331032, + 0.018676234409213066, + 0.015710702165961266, + 0.01734004355967045, + 0.018578507006168365, + 0.016196416690945625, + 0.01796756312251091, + 0.02086324617266655, + 0.004492693580687046, + 0.005940629169344902, + 0.005580986849963665, + 0.012567556463181973, + 0.005175992846488953, + 0.003796080593019724, + 0.004572370555251837, + 0.007731680292636156, + 0.008188185282051563, + 0.0026160108391195536, + 0.016159379854798317, + 0.005005101207643747, + 0.003505158703774214, + 0.0038831504061818123, + 0.00449839374050498, + 0.009644727222621441, + 0.00628491397947073, + 0.002207085257396102, + 0.009084356017410755, + 0.00349738379009068, + 0.002973443828523159, + 0.0030877201352268457, + 0.004411512985825539, + 0.007633981294929981, + 0.007226494140923023, + 0.0038902442902326584, + 0.009609793312847614, + 0.007034993264824152, + 0.006905493792146444, + 0.00529872439801693, + 0.011395185254514217, + 0.011381539516150951, + 0.01221510861068964, + 0.0062103066593408585, + 0.01743743196129799, + 0.00868128426373005, + 0.0073903584852814674, + 0.007900509983301163, + 0.01036082487553358, + 0.012683641165494919, + 0.0037443265318870544, + 0.004001235123723745, + 0.005439835134893656, + 0.004938030149787664, + 0.0041102939285337925, + 0.003955083899199963, + 0.005555999465286732, + 0.008802791126072407, + 0.05322643369436264, + 0.03230089321732521, + 0.05082757771015167, + 0.06483196467161179, + 0.06668156385421753, + 0.04578568786382675, + 0.09236446022987366, + 0.0630536898970604, + 0.003815935691818595, + 0.0029667753260582685, + 0.003437489503994584, + 0.0040151383727788925, + 0.004185172263532877, + 0.0026661958545446396, + 0.003709969809278846, + 0.00348190451040864, + 0.003489926690235734, + 0.001938381465151906, + 0.010062727145850658, + 0.003041916759684682, + 0.0019378192955628037, + 0.0021523963660001755, + 0.0024961400777101517, + 0.007441897876560688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.01893516071140766 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067648435942828655 + ], + "opposite_logprob_tokens": [ + -5.006764888763428 + ], + "logprob_reward": -0.0067648435942828655, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067648435942828655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24771985411643982, + 0.09155081957578659, + 0.12076862901449203, + 0.13787414133548737, + 0.05901806801557541, + 0.031008979305624962, + 0.0563211552798748, + 0.1525547057390213, + 0.0443715900182724, + 0.0588119737803936 + ], + "credit_map": [ + 0.02365107648074627, + 0.03167089447379112, + 0.023631755262613297, + 0.03262585029006004, + 0.035189446061849594, + 0.02921878546476364, + 0.04400118067860603, + 0.027730876579880714, + 0.010617760941386223, + 0.008012776263058186, + 0.01639433391392231, + 0.018194518983364105, + 0.009692184627056122, + 0.006819956470280886, + 0.008322234265506268, + 0.013497060164809227, + 0.013424997217953205, + 0.008427746593952179, + 0.01873491331934929, + 0.01572834514081478, + 0.015689635649323463, + 0.009917167015373707, + 0.01980746164917946, + 0.019038377329707146, + 0.023327432572841644, + 0.006488956976681948, + 0.03158217668533325, + 0.006838272791355848, + 0.01626475155353546, + 0.01105289813131094, + 0.012986886315047741, + 0.029332786798477173, + 0.006071705371141434, + 0.004067681264132261, + 0.009282073006033897, + 0.005645332857966423, + 0.011314024217426777, + 0.0050299749709665775, + 0.008175523020327091, + 0.009431757964193821, + 0.0031136011239141226, + 0.0023861294612288475, + 0.005085258278995752, + 0.0032710933592170477, + 0.0038492027670145035, + 0.003747060662135482, + 0.004078778438270092, + 0.005477855913341045, + 0.0061117494478821754, + 0.005168374162167311, + 0.008065847679972649, + 0.010629257187247276, + 0.0067788842134177685, + 0.004378747660666704, + 0.0065953731536865234, + 0.008592926897108555, + 0.018534988164901733, + 0.011121468618512154, + 0.018403824418783188, + 0.018617764115333557, + 0.0198739692568779, + 0.021210523322224617, + 0.023208780214190483, + 0.02158341556787491, + 0.005332141183316708, + 0.003404950490221381, + 0.0065315221436321735, + 0.003945959731936455, + 0.007147938013076782, + 0.004778372589498758, + 0.006662789732217789, + 0.0065679242834448814, + 0.004974746610969305, + 0.004468111786991358, + 0.008869774639606476, + 0.006070279516279697, + 0.006575660314410925, + 0.0036457297392189503, + 0.011740753427147865, + 0.012466919608414173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.0067648435942828655 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11302316933870316 + ], + "opposite_logprob_tokens": [ + -2.238023281097412 + ], + "logprob_reward": -0.11302316933870316, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11302316933870316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16544167697429657, + 0.1737966388463974, + 0.08717242628335953, + 0.04171495884656906, + 0.11940664052963257, + 0.0665273517370224, + 0.10101193934679031, + 0.1686478853225708, + 0.04153814539313316, + 0.0347423255443573 + ], + "credit_map": [ + 0.02970043197274208, + 0.02236887440085411, + 0.0252873282879591, + 0.01482764445245266, + 0.01954602636396885, + 0.015952102839946747, + 0.02000155672430992, + 0.017757710069417953, + 0.013776014558970928, + 0.01981537975370884, + 0.021162405610084534, + 0.03864073008298874, + 0.01648559793829918, + 0.01756046712398529, + 0.014883109368383884, + 0.03147292882204056, + 0.015483994968235493, + 0.006414331961423159, + 0.01747838780283928, + 0.010170738212764263, + 0.007914829067885876, + 0.006910975556820631, + 0.010736092925071716, + 0.012063080444931984, + 0.005206376314163208, + 0.0022914160508662462, + 0.01004413515329361, + 0.0027385263238102198, + 0.0036784063559025526, + 0.0035481094382703304, + 0.004814098123461008, + 0.00939389318227768, + 0.020878560841083527, + 0.008453837595880032, + 0.023456119000911713, + 0.011435301974415779, + 0.011369251646101475, + 0.008955865167081356, + 0.015649979934096336, + 0.019207721576094627, + 0.01140284352004528, + 0.0033983110915869474, + 0.017950670793652534, + 0.005117590073496103, + 0.004102164413779974, + 0.0070534804835915565, + 0.006003390531986952, + 0.011498904787003994, + 0.009388485923409462, + 0.00877401977777481, + 0.014271141961216927, + 0.01825428009033203, + 0.010312390513718128, + 0.007731700781732798, + 0.010937835089862347, + 0.02134208381175995, + 0.01707964576780796, + 0.011295469477772713, + 0.019395820796489716, + 0.022186052054166794, + 0.022398700937628746, + 0.01816834695637226, + 0.03542974591255188, + 0.022694092243909836, + 0.00325381918810308, + 0.0029369816184043884, + 0.006014491431415081, + 0.003976371604949236, + 0.00791886355727911, + 0.003937453497201204, + 0.007258773315697908, + 0.00624138955026865, + 0.003974385093897581, + 0.00408317381516099, + 0.0058832475915551186, + 0.0027675279416143894, + 0.0049798027612268925, + 0.0029282900504767895, + 0.0038041439838707447, + 0.006321753840893507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.11302316933870316 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011082522571086884 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.011082522571086884, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082522571086884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18300750851631165, + 0.26592376828193665, + 0.05015556514263153, + 0.08970894664525986, + 0.07883325964212418, + 0.03240618854761124, + 0.057253606617450714, + 0.16693326830863953, + 0.033241529017686844, + 0.04253634810447693 + ], + "credit_map": [ + 0.03386564925312996, + 0.02404782362282276, + 0.0267291571944952, + 0.015976497903466225, + 0.02155887894332409, + 0.014148596674203873, + 0.020543968304991722, + 0.02613694779574871, + 0.02215924859046936, + 0.03206411749124527, + 0.029574453830718994, + 0.06887824833393097, + 0.02737722173333168, + 0.02486506476998329, + 0.021188827231526375, + 0.03981660306453705, + 0.007645963691174984, + 0.003931868821382523, + 0.007015187293291092, + 0.00634631235152483, + 0.007147151045501232, + 0.004499984905123711, + 0.006505298428237438, + 0.007063799537718296, + 0.013872596435248852, + 0.0052465591579675674, + 0.016666390001773834, + 0.006125671789050102, + 0.010894901119172573, + 0.009168380871415138, + 0.009669207036495209, + 0.018065251410007477, + 0.009875413030385971, + 0.005076397210359573, + 0.01814851351082325, + 0.005948353558778763, + 0.00926550105214119, + 0.007282198406755924, + 0.009408343583345413, + 0.013828547671437263, + 0.003551671514287591, + 0.0020275809802114964, + 0.0049701896496117115, + 0.004029361065477133, + 0.0031128358095884323, + 0.003735386999323964, + 0.004674988798797131, + 0.006304173730313778, + 0.005988056771457195, + 0.006178468931466341, + 0.00891533400863409, + 0.009150587022304535, + 0.006404057610780001, + 0.005285229533910751, + 0.007382943294942379, + 0.007948936894536018, + 0.01569405198097229, + 0.011166977696120739, + 0.016777995973825455, + 0.023569392040371895, + 0.024170925840735435, + 0.01753433793783188, + 0.036352675408124924, + 0.02166692726314068, + 0.002835218096151948, + 0.002218296518549323, + 0.004146238788962364, + 0.003807765431702137, + 0.006445178296416998, + 0.002917608479037881, + 0.006422628182917833, + 0.00444859778508544, + 0.0037508951500058174, + 0.003675664309412241, + 0.009887157008051872, + 0.003285472746938467, + 0.004799234680831432, + 0.0029974975623190403, + 0.003933028783649206, + 0.010207404382526875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.011082522571086884 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028186615090817213 + ], + "opposite_logprob_tokens": [ + -5.877818584442139 + ], + "logprob_reward": -0.0028186615090817213, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028186615090817213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11501966416835785, + 0.12824001908302307, + 0.10873761028051376, + 0.04809606820344925, + 0.08649227023124695, + 0.05954309552907944, + 0.1602548211812973, + 0.10932298004627228, + 0.06707894802093506, + 0.11721453070640564 + ], + "credit_map": [ + 0.01667884550988674, + 0.012056445702910423, + 0.012862959876656532, + 0.011655759066343307, + 0.017723940312862396, + 0.010931742377579212, + 0.01993384025990963, + 0.01317613571882248, + 0.010903219692409039, + 0.0132056325674057, + 0.017406528815627098, + 0.02764522098004818, + 0.011208838783204556, + 0.012679433450102806, + 0.013172561302781105, + 0.022018583491444588, + 0.018512332811951637, + 0.005087648052722216, + 0.023116908967494965, + 0.00828964076936245, + 0.008586683310568333, + 0.0073699490167200565, + 0.009959802031517029, + 0.027814650908112526, + 0.005801263265311718, + 0.002425745129585266, + 0.011933984234929085, + 0.0041501205414533615, + 0.004263007547706366, + 0.003770067822188139, + 0.00522603327408433, + 0.010525844991207123, + 0.007352304179221392, + 0.004223647061735392, + 0.017235836014151573, + 0.006266115233302116, + 0.008598879911005497, + 0.009115810506045818, + 0.0104175740852952, + 0.02328210137784481, + 0.003582275239750743, + 0.003309177467599511, + 0.012139344587922096, + 0.004210083745419979, + 0.005427263211458921, + 0.0064521776512265205, + 0.006110667251050472, + 0.018312105908989906, + 0.013054229319095612, + 0.016137897968292236, + 0.0225091353058815, + 0.023920325562357903, + 0.014820472337305546, + 0.01487867720425129, + 0.01973252184689045, + 0.03520156070590019, + 0.017258228734135628, + 0.007178519852459431, + 0.012909114360809326, + 0.011550909839570522, + 0.014095213264226913, + 0.010165253654122353, + 0.01936398260295391, + 0.016801757737994194, + 0.0063123442232608795, + 0.0048712752759456635, + 0.008126449771225452, + 0.006777925416827202, + 0.013235625810921192, + 0.007066588383167982, + 0.01177255529910326, + 0.008916180580854416, + 0.010125224478542805, + 0.006501300726085901, + 0.04209029674530029, + 0.006386968307197094, + 0.006093252915889025, + 0.005237251054495573, + 0.007004503160715103, + 0.033775731921195984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 9 + ], + "selected_reward": -0.0028186615090817213 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot shot from a level angle, captured with a calm, restrained documentary style, precisely captures the moment of an Asian woman's emotional breakdown. She has long, straight, black hair, strands of which are messy and plastered to her cheeks with tears. Her eyes are red and swollen from prolonged crying, her brow furrowed, her lips slightly parted, her lower lip trembling as she struggles to suppress sobs. Large, glistening tears flow uncontrollably from her reddened eyes, silently sliding down her plain face, leaving streaks of wet light. Her shoulders tremble slightly from her uncontrollable sobs. She wears a simple gray turtleneck sweater, the soft material contrasting with her fragile emotions. The background is a simple, blurred interior space, and soft natural light filtering in from one side of the frame delicately outlines her facial features and makes each tear drop stand out with exceptional clarity. The whole picture has a cold tone and no dramatic camera movements. It just quietly and non-judgmentally records this primitive and deep sadness, creating a heartbreaking sense of reality and an immersive atmosphere of empathy.", + "seed": 26291, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_10/index_6839_seed_26291_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.013899588957428932, + "vlm_logprob": -0.013899588957428932, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.013899588957428932, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005912196822464466 + ], + "opposite_logprob_tokens": [ + -5.2559123039245605 + ], + "logprob_reward": -0.005912196822464466, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005912196822464466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4195776581764221, + 0.10598968714475632, + 0.06302584707736969, + 0.05962409824132919, + 0.08250033110380173, + 0.03482159227132797, + 0.05221603438258171, + 0.02280917577445507, + 0.12163504958152771, + 0.0378006175160408 + ], + "credit_map": [ + 0.09129085391759872, + 0.04091303423047066, + 0.06604954600334167, + 0.028356779366731644, + 0.0841742604970932, + 0.024968991056084633, + 0.058492083102464676, + 0.02533208206295967, + 0.014785947278141975, + 0.009068304672837257, + 0.009309428744018078, + 0.010823838412761688, + 0.015178103931248188, + 0.00501007167622447, + 0.013265576213598251, + 0.028548406437039375, + 0.005192465614527464, + 0.005447042640298605, + 0.008088406175374985, + 0.005419038236141205, + 0.01384578738361597, + 0.004223535303026438, + 0.013854039832949638, + 0.006955524440854788, + 0.004421357065439224, + 0.004226632881909609, + 0.0053495909087359905, + 0.011634563095867634, + 0.010214803740382195, + 0.005710295867174864, + 0.010924363508820534, + 0.007142486050724983, + 0.008154168725013733, + 0.012856529094278812, + 0.015117643401026726, + 0.00965369213372469, + 0.01112392544746399, + 0.0074752201326191425, + 0.01050350908190012, + 0.007615640759468079, + 0.005017964635044336, + 0.0036303491797298193, + 0.004738043062388897, + 0.004057036247104406, + 0.005549732130020857, + 0.002314155688509345, + 0.005447233561426401, + 0.004067075438797474, + 0.006758222822099924, + 0.00617654575034976, + 0.005814236123114824, + 0.006127695087343454, + 0.009401476942002773, + 0.003137868596240878, + 0.007355614099651575, + 0.007444371003657579, + 0.002703743288293481, + 0.003106639487668872, + 0.003389886114746332, + 0.002454200526699424, + 0.00442925188690424, + 0.0012107022339478135, + 0.002989315427839756, + 0.002525435062125325, + 0.011121602728962898, + 0.010333821177482605, + 0.01495019719004631, + 0.016055401414632797, + 0.02594069577753544, + 0.007350726053118706, + 0.02181965298950672, + 0.01406294945627451, + 0.00833466462790966, + 0.0031047752127051353, + 0.004616481717675924, + 0.003658877220004797, + 0.0061281342059373856, + 0.00217238487675786, + 0.005909065715968609, + 0.003876233007758856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.005912196822464466 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014226209372282028 + ], + "opposite_logprob_tokens": [ + -4.26422643661499 + ], + "logprob_reward": -0.014226209372282028, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014226209372282028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27835607528686523, + 0.14624512195587158, + 0.1333695650100708, + 0.04574523866176605, + 0.051289986819028854, + 0.12705637514591217, + 0.02998950146138668, + 0.026049872860312462, + 0.06585724651813507, + 0.09604103118181229 + ], + "credit_map": [ + 0.05344761908054352, + 0.02382109872996807, + 0.021987730637192726, + 0.02055749110877514, + 0.08609778434038162, + 0.015946177765727043, + 0.0367940254509449, + 0.01970416121184826, + 0.018114546313881874, + 0.01137208566069603, + 0.00866031926125288, + 0.014848139137029648, + 0.024228446185588837, + 0.007569711189717054, + 0.01717994548380375, + 0.04427195340394974, + 0.00909421406686306, + 0.01581650599837303, + 0.010401765815913677, + 0.015450267121195793, + 0.035873327404260635, + 0.007641277275979519, + 0.023419294506311417, + 0.015672916546463966, + 0.0033550190273672342, + 0.005071599967777729, + 0.003448561066761613, + 0.008178621530532837, + 0.007258075755089521, + 0.003315290901809931, + 0.007833953015506268, + 0.0072841220535337925, + 0.004651943221688271, + 0.004373745061457157, + 0.004300600849092007, + 0.004868964198976755, + 0.010275463573634624, + 0.003568535204976797, + 0.013799826614558697, + 0.005450911354273558, + 0.009859581477940083, + 0.016344930976629257, + 0.014164143241941929, + 0.012772233225405216, + 0.025043295696377754, + 0.009266320616006851, + 0.02335180528461933, + 0.016254065558314323, + 0.003986706957221031, + 0.0037323804572224617, + 0.004508517682552338, + 0.0033125081099569798, + 0.00498536741361022, + 0.0024652096908539534, + 0.003954518586397171, + 0.0030442969873547554, + 0.0032855041790753603, + 0.002367829205468297, + 0.0030425868462771177, + 0.002427334664389491, + 0.004786650184541941, + 0.0011078346287831664, + 0.004933756310492754, + 0.004098378587514162, + 0.004575197584927082, + 0.006413136143237352, + 0.006176284048706293, + 0.007844517007470131, + 0.014167697168886662, + 0.005049432627856731, + 0.015072680078446865, + 0.006558308843523264, + 0.011869235895574093, + 0.009364509955048561, + 0.007575614377856255, + 0.009179264307022095, + 0.020735755562782288, + 0.006214895751327276, + 0.019395818933844566, + 0.011705941520631313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014226209372282028 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306454211473465 + ], + "opposite_logprob_tokens": [ + -3.168064594268799 + ], + "logprob_reward": -0.04306454211473465, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306454211473465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08509507775306702, + 0.11192979663610458, + 0.06997741758823395, + 0.48858821392059326, + 0.03598208725452423, + 0.028832105919718742, + 0.01911470852792263, + 0.011125750839710236, + 0.12696371972560883, + 0.022391140460968018 + ], + "credit_map": [ + 0.008015645667910576, + 0.007663427386432886, + 0.0077166310511529446, + 0.008353310637176037, + 0.025228146463632584, + 0.006126388907432556, + 0.014894002117216587, + 0.0070975287817418575, + 0.01667628437280655, + 0.008229300379753113, + 0.00910618994385004, + 0.01134397927671671, + 0.011913897469639778, + 0.005915830377489328, + 0.013537304475903511, + 0.03520701453089714, + 0.006748660001903772, + 0.006703595630824566, + 0.01034811045974493, + 0.010201673023402691, + 0.011323788203299046, + 0.007348100654780865, + 0.008178859017789364, + 0.009124630130827427, + 0.0428752526640892, + 0.043308354914188385, + 0.06004808098077774, + 0.12824209034442902, + 0.05747494846582413, + 0.03441403806209564, + 0.043676506727933884, + 0.07854893058538437, + 0.003043937962502241, + 0.004218592774122953, + 0.003356731263920665, + 0.0046734134666621685, + 0.005158339627087116, + 0.003611260559409857, + 0.007185371592640877, + 0.004734443966299295, + 0.0024354197084903717, + 0.003480255138128996, + 0.0032855391036719084, + 0.0036966244224458933, + 0.0048043979331851006, + 0.001718289451673627, + 0.006244848016649485, + 0.0031667337752878666, + 0.002819374669343233, + 0.0021785562857985497, + 0.0036952835507690907, + 0.0020415603648871183, + 0.002661145757883787, + 0.0013851332478225231, + 0.0017484218114987016, + 0.0025852348189800978, + 0.0011602258309721947, + 0.0009317400981672108, + 0.0013160313246771693, + 0.0014796836767345667, + 0.0016893494175747037, + 0.0006212859298102558, + 0.0014890070306137204, + 0.002438428346067667, + 0.009917883202433586, + 0.011312977410852909, + 0.011463748291134834, + 0.011877452954649925, + 0.033898238092660904, + 0.0076775625348091125, + 0.027599340304732323, + 0.013216523453593254, + 0.00266694906167686, + 0.001721648615784943, + 0.0026603785809129477, + 0.0019106519175693393, + 0.004477549344301224, + 0.0013640022370964289, + 0.005266136489808559, + 0.0023238251451402903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 1 + ], + "selected_reward": -0.04306454211473465 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200088929384947 + ], + "opposite_logprob_tokens": [ + -6.127200126647949 + ], + "logprob_reward": -0.002200088929384947, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200088929384947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.207859605550766, + 0.17293792963027954, + 0.03148750588297844, + 0.04880969971418381, + 0.0483766570687294, + 0.2831883430480957, + 0.06196688488125801, + 0.019904807209968567, + 0.08301108330488205, + 0.04245750978589058 + ], + "credit_map": [ + 0.02649989165365696, + 0.01933564431965351, + 0.01593082956969738, + 0.014218957163393497, + 0.07383474707603455, + 0.01421826146543026, + 0.028292875736951828, + 0.015528395771980286, + 0.02835928462445736, + 0.013567839749157429, + 0.011451401747763157, + 0.018740808591246605, + 0.016544818878173828, + 0.01014624536037445, + 0.010858960449695587, + 0.06326857209205627, + 0.003709223587065935, + 0.0033868192695081234, + 0.00300366198644042, + 0.00419086217880249, + 0.005559190176427364, + 0.0026929122395813465, + 0.004093927331268787, + 0.004850910045206547, + 0.003250007750466466, + 0.0031738635152578354, + 0.0033602393232285976, + 0.011854304000735283, + 0.008535397239029408, + 0.0035524724517017603, + 0.006901850458234549, + 0.008181565441191196, + 0.004947606008499861, + 0.005539932753890753, + 0.0055363173596560955, + 0.0052909222431480885, + 0.008260774426162243, + 0.004345996305346489, + 0.009510069154202938, + 0.004945035092532635, + 0.04032213240861893, + 0.03470626473426819, + 0.04362165182828903, + 0.027310797944664955, + 0.037230879068374634, + 0.019219139590859413, + 0.041324958205223083, + 0.03945249691605568, + 0.007171728182584047, + 0.008881312794983387, + 0.006628843490034342, + 0.006220691837370396, + 0.010951136238873005, + 0.003761447034776211, + 0.010199613869190216, + 0.008152113296091557, + 0.0021707036066800356, + 0.001827721600420773, + 0.0022017499431967735, + 0.0026526632718741894, + 0.003363644937053323, + 0.0010697680991142988, + 0.0025898299645632505, + 0.00402872683480382, + 0.006347333081066608, + 0.0077442508190870285, + 0.007699136156588793, + 0.007391310762614012, + 0.02316225692629814, + 0.004905395675450563, + 0.017506882548332214, + 0.008254526183009148, + 0.011991921812295914, + 0.0035405331291258335, + 0.0065323831513524055, + 0.0026787729002535343, + 0.005735088139772415, + 0.0030042529106140137, + 0.0053230999037623405, + 0.0036514606326818466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.002200088929384947 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an Asian woman with long, straight, black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the woman's eyes red and swollen with large, glistening tears flowing down her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple gray turtleneck sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a simple, blurred interior space with soft natural light filtering from one side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up with a static camera at a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409490754827857 + ], + "opposite_logprob_tokens": [ + -5.504095077514648 + ], + "logprob_reward": -0.00409490754827857, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409490754827857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16789588332176208, + 0.11508917063474655, + 0.03400174155831337, + 0.08372534811496735, + 0.09457900375127792, + 0.10162193328142166, + 0.09549294412136078, + 0.038839902728796005, + 0.12359320372343063, + 0.14516091346740723 + ], + "credit_map": [ + 0.01711803302168846, + 0.018104013055562973, + 0.01498347707092762, + 0.01377087365835905, + 0.052245400846004486, + 0.01115417294204235, + 0.02664453163743019, + 0.0138753866776824, + 0.017785338684916496, + 0.009814116172492504, + 0.008051220327615738, + 0.010817030444741249, + 0.013110099360346794, + 0.0056344373151659966, + 0.015237832441926003, + 0.03463909402489662, + 0.00272157764993608, + 0.00327015807852149, + 0.0034541343338787556, + 0.005387020763009787, + 0.00610517431050539, + 0.0035415198653936386, + 0.005417336244136095, + 0.0041048224084079266, + 0.006599151063710451, + 0.005912266205996275, + 0.010350199416279793, + 0.01890050619840622, + 0.012723170220851898, + 0.005790791939944029, + 0.011828103102743626, + 0.011621162295341492, + 0.009100588038563728, + 0.010626964271068573, + 0.006146955769509077, + 0.012121287174522877, + 0.0217872504144907, + 0.007959775626659393, + 0.01679093949496746, + 0.010045244358479977, + 0.009117111563682556, + 0.014824441634118557, + 0.009813448414206505, + 0.01130241435021162, + 0.020077621564269066, + 0.007674342021346092, + 0.01596357859671116, + 0.012848975136876106, + 0.009692217223346233, + 0.008493456058204174, + 0.006946728099137545, + 0.014975168742239475, + 0.011933124624192715, + 0.0064034308306872845, + 0.015854280441999435, + 0.021194541826844215, + 0.0035201183054596186, + 0.0033303017262369394, + 0.003253018716350198, + 0.005873158574104309, + 0.00726973544806242, + 0.00235063792206347, + 0.004542008973658085, + 0.008700922131538391, + 0.009851362556219101, + 0.01119779422879219, + 0.010895376093685627, + 0.011584235355257988, + 0.03402700275182724, + 0.007279381621629, + 0.025862129405140877, + 0.012895915657281876, + 0.049339279532432556, + 0.015113191679120064, + 0.022215453907847404, + 0.009297354146838188, + 0.015414897352457047, + 0.010979611426591873, + 0.011503239162266254, + 0.011297892779111862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.00409490754827857 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_11/index_tensor([3420, 3420])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_11/index_tensor([3420, 3420])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..95b9a41e4addb07094b4969704792d2410973c3a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_11/index_tensor([3420, 3420])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, captured in the signature Studio Ghibli animation style, takes place from a slightly overhead angle. The camera remains still, focused on a textured, aged wooden table. Bathed in the warm, golden light slanting through the window in the late afternoon, the wood's grain and subtle scratches are clearly visible. A pair of hands, with healthy complexion, distinct knuckles, and a hint of warmth, rest upon the tabletop, their skin taking on a soft sheen in the light. Carefully and deliberately, these hands gently sweep aside a thin layer of fine dust that has accumulated on the tabletop with the sides of their palms. Each subtle movement is rendered with incredible fluidity by the extremely high frame rate. The sunlight illuminates the lifted dust particles, causing them to dance briefly, circling in the air like shimmering golden light before slowly settling. The warm, muted tones of the image evoke a sense of nostalgia and tranquility. The soft glow of the hands contrasts subtly with the matte texture of the wooden table, evoking a sense of focused and healing daily ritual.", + "seed": 3323, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_11/index_3420_seed_3323_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19832735061645507, + "vlm_logprob": -0.19832735061645507, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.19832735061645507, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033870629966259 + ], + "opposite_logprob_tokens": [ + -5.533870697021484 + ], + "logprob_reward": -0.033870629966259, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.033870629966259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09938360750675201, + 0.06355702877044678, + 0.06889712810516357, + 0.1314031332731247, + 0.09642590582370758, + 0.08753875643014908, + 0.12730510532855988, + 0.08119383454322815, + 0.10853900760412216, + 0.1357564479112625 + ], + "credit_map": [ + 0.009314049035310745, + 0.012286095879971981, + 0.008057775907218456, + 0.013443347066640854, + 0.016878273338079453, + 0.014239564538002014, + 0.013453565537929535, + 0.011710934340953827, + 0.006137733347713947, + 0.006950731854885817, + 0.0045328219421207905, + 0.012203577905893326, + 0.008783305063843727, + 0.009061755612492561, + 0.010147188790142536, + 0.005739910993725061, + 0.007159166503697634, + 0.010564465075731277, + 0.00586910592392087, + 0.011378489434719086, + 0.009098416194319725, + 0.007361031603068113, + 0.011133228428661823, + 0.006333222147077322, + 0.008621735498309135, + 0.019229991361498833, + 0.016978437080979347, + 0.021403618156909943, + 0.011859689839184284, + 0.014103420078754425, + 0.018630387261509895, + 0.020575854927301407, + 0.0063669318333268166, + 0.018971286714076996, + 0.006926184985786676, + 0.02413153648376465, + 0.007921003736555576, + 0.007986652664840221, + 0.012829750776290894, + 0.011292560957372189, + 0.0055656423792243, + 0.01770012266933918, + 0.006920980755239725, + 0.026534197852015495, + 0.005821758881211281, + 0.010332425124943256, + 0.008258827030658722, + 0.006404805462807417, + 0.009966863319277763, + 0.028468040749430656, + 0.007769597228616476, + 0.03912855684757233, + 0.008655420504510403, + 0.010587329976260662, + 0.012994201853871346, + 0.009735089726746082, + 0.005644662771373987, + 0.009267030283808708, + 0.00871133804321289, + 0.010395225137472153, + 0.00969853438436985, + 0.011426703073084354, + 0.016526944935321808, + 0.009523388929665089, + 0.010464159771800041, + 0.012514564208686352, + 0.01226110477000475, + 0.013301125727593899, + 0.017408758401870728, + 0.014589473605155945, + 0.016576960682868958, + 0.011422860436141491, + 0.012623085640370846, + 0.018590835854411125, + 0.01217322051525116, + 0.020330850034952164, + 0.018967412412166595, + 0.01625538244843483, + 0.024106111377477646, + 0.012709551490843296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.033870629966259 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983247876167297 + ], + "opposite_logprob_tokens": [ + -0.5233247876167297 + ], + "logprob_reward": -0.8983247876167297, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983247876167297, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08521155267953873, + 0.06632581353187561, + 0.24251800775527954, + 0.14335419237613678, + 0.1589166671037674, + 0.05862357094883919, + 0.059522222727537155, + 0.0561605840921402, + 0.04940212890505791, + 0.07996533811092377 + ], + "credit_map": [ + 0.006547156255692244, + 0.009659955278038979, + 0.010161870159208775, + 0.011453231796622276, + 0.011870602145791054, + 0.012430630624294281, + 0.012582850642502308, + 0.010505246929824352, + 0.00690498249605298, + 0.007446461822837591, + 0.008046457543969154, + 0.01183324120938778, + 0.006981892976909876, + 0.008766990154981613, + 0.01063553523272276, + 0.005710242316126823, + 0.03177405893802643, + 0.048538755625486374, + 0.01806904561817646, + 0.03452819213271141, + 0.028786076232790947, + 0.03203408792614937, + 0.029202207922935486, + 0.01958555355668068, + 0.008701731450855732, + 0.01953676901757717, + 0.014144974760711193, + 0.030665908008813858, + 0.012279516085982323, + 0.013956472277641296, + 0.01909804157912731, + 0.02497076988220215, + 0.010107499547302723, + 0.02456136792898178, + 0.012292634695768356, + 0.035867881029844284, + 0.013768823817372322, + 0.014675608836114407, + 0.02654970809817314, + 0.021093130111694336, + 0.00390638830140233, + 0.01258267555385828, + 0.0043881977908313274, + 0.014597728848457336, + 0.004650008399039507, + 0.004696566145867109, + 0.0071410215459764, + 0.006660982500761747, + 0.004124418366700411, + 0.01087121106684208, + 0.0050834594294428825, + 0.01612348109483719, + 0.00471632881090045, + 0.006063762120902538, + 0.007494852878153324, + 0.005044701974838972, + 0.004626784007996321, + 0.007488932926207781, + 0.004704452119767666, + 0.008085024543106556, + 0.007159475237131119, + 0.008642216213047504, + 0.007602475583553314, + 0.00785121414810419, + 0.003303682431578636, + 0.0052324701100587845, + 0.003317092312499881, + 0.007057641167193651, + 0.008558942005038261, + 0.007831648923456669, + 0.007953394204378128, + 0.006147252861410379, + 0.00655410485342145, + 0.01070446614176035, + 0.007598144002258778, + 0.011695564724504948, + 0.008582529611885548, + 0.009129513055086136, + 0.016988500952720642, + 0.008712513372302055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.8983247876167297 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009802302345633507 + ], + "opposite_logprob_tokens": [ + -4.634802341461182 + ], + "logprob_reward": -0.009802302345633507, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009802302345633507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17652875185012817, + 0.0910896435379982, + 0.12285119295120239, + 0.03248114511370659, + 0.05546896159648895, + 0.1707039624452591, + 0.03826021030545235, + 0.09696733951568604, + 0.033673036843538284, + 0.18197575211524963 + ], + "credit_map": [ + 0.015860725194215775, + 0.01832379586994648, + 0.01783028617501259, + 0.02709486149251461, + 0.02061736211180687, + 0.02285603992640972, + 0.03203478083014488, + 0.021910902112722397, + 0.008991767652332783, + 0.014340207912027836, + 0.007282541133463383, + 0.015521432273089886, + 0.013020312413573265, + 0.013187997974455357, + 0.011900108307600021, + 0.006845272611826658, + 0.011104759760200977, + 0.021586138755083084, + 0.011081168428063393, + 0.0137108713388443, + 0.022613028064370155, + 0.012430437840521336, + 0.018892960622906685, + 0.011431829072535038, + 0.0027405351866036654, + 0.0026627052575349808, + 0.004391406197100878, + 0.0038979952223598957, + 0.0031003288459032774, + 0.0043705119751393795, + 0.006314893718808889, + 0.0050027696415781975, + 0.00468925666064024, + 0.00926898792386055, + 0.004039316903799772, + 0.014966818504035473, + 0.004041879437863827, + 0.0069208876229822636, + 0.006024506408721209, + 0.00551730627194047, + 0.011380665004253387, + 0.023193785920739174, + 0.017311956733465195, + 0.022274592891335487, + 0.020248208194971085, + 0.015181634575128555, + 0.02656148374080658, + 0.03455163538455963, + 0.004326550289988518, + 0.0033917620312422514, + 0.003392360173165798, + 0.006422356702387333, + 0.0035527937579900026, + 0.003886257065460086, + 0.006618340499699116, + 0.006669789552688599, + 0.005464745685458183, + 0.010989035479724407, + 0.00792249757796526, + 0.01282094419002533, + 0.012936857528984547, + 0.014665938913822174, + 0.019217273220419884, + 0.012950054369866848, + 0.00252030068077147, + 0.0035224712919443846, + 0.0031336224637925625, + 0.004040826112031937, + 0.005519510246813297, + 0.006212770007550716, + 0.005481865722686052, + 0.003241667291149497, + 0.01490026619285345, + 0.0233993511646986, + 0.016997339203953743, + 0.023512402549386024, + 0.02462286874651909, + 0.0248388834297657, + 0.0339350625872612, + 0.0197695754468441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.009802302345633507 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026349445804953575 + ], + "opposite_logprob_tokens": [ + -3.6513495445251465 + ], + "logprob_reward": -0.026349445804953575, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349445804953575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06814991682767868, + 0.14398077130317688, + 0.10167307406663895, + 0.0820159837603569, + 0.08703054487705231, + 0.21024316549301147, + 0.05560251325368881, + 0.06971921026706696, + 0.03852669149637222, + 0.14305810630321503 + ], + "credit_map": [ + 0.005538587458431721, + 0.008764239028096199, + 0.007070059888064861, + 0.009005851112306118, + 0.01030805241316557, + 0.009395506232976913, + 0.009725223295390606, + 0.008342391811311245, + 0.012392472475767136, + 0.016103066504001617, + 0.011438123881816864, + 0.023804403841495514, + 0.02148308791220188, + 0.02421528287231922, + 0.02138346992433071, + 0.013160857371985912, + 0.010025790892541409, + 0.02164044789969921, + 0.00667354092001915, + 0.015935217961668968, + 0.01741054095327854, + 0.009572437033057213, + 0.012882215902209282, + 0.007532875053584576, + 0.005014156922698021, + 0.011062156409025192, + 0.00950698833912611, + 0.01673693023622036, + 0.007265287917107344, + 0.009086931124329567, + 0.011180964298546314, + 0.01216257456690073, + 0.0058165546506643295, + 0.010417775250971317, + 0.008804244920611382, + 0.013325280509889126, + 0.010444327257573605, + 0.008537136018276215, + 0.01709461212158203, + 0.01259061973541975, + 0.012843437492847443, + 0.017549317330121994, + 0.02249418944120407, + 0.03336329758167267, + 0.024280110374093056, + 0.019117305055260658, + 0.03400861844420433, + 0.04658688232302666, + 0.0061917961575090885, + 0.006830996368080378, + 0.00489750225096941, + 0.00835308339446783, + 0.007606405299156904, + 0.006344106048345566, + 0.007806363049894571, + 0.007572262082248926, + 0.004174241330474615, + 0.0058172415010631084, + 0.00566522404551506, + 0.010711701586842537, + 0.009913798421621323, + 0.009245183318853378, + 0.014583631418645382, + 0.009608190506696701, + 0.0033190096728503704, + 0.004330785479396582, + 0.0035145236179232597, + 0.005892524495720863, + 0.004935111850500107, + 0.006155543029308319, + 0.005448953714221716, + 0.004930238705128431, + 0.013786789029836655, + 0.019459782168269157, + 0.014915267936885357, + 0.018719200044870377, + 0.019408714026212692, + 0.017178593203425407, + 0.025176746770739555, + 0.014413025230169296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 9 + ], + "selected_reward": -0.026349445804953575 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328958734869957 + ], + "opposite_logprob_tokens": [ + -3.773289680480957 + ], + "logprob_reward": -0.02328958734869957, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328958734869957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2771633565425873, + 0.08708678930997849, + 0.07092684507369995, + 0.07984635978937149, + 0.08814475685358047, + 0.11572122573852539, + 0.10734502226114273, + 0.06461973488330841, + 0.06102251634001732, + 0.04812334105372429 + ], + "credit_map": [ + 0.027236362919211388, + 0.028148923069238663, + 0.026808269321918488, + 0.04399648308753967, + 0.033791057765483856, + 0.03834141418337822, + 0.04847244545817375, + 0.030368437990546227, + 0.008581099100410938, + 0.01168584544211626, + 0.009084961377084255, + 0.009772518649697304, + 0.012231482192873955, + 0.013828393071889877, + 0.012763964012265205, + 0.009138532914221287, + 0.00735560804605484, + 0.01195117924362421, + 0.007293693721294403, + 0.01683904230594635, + 0.008187253959476948, + 0.006301369518041611, + 0.007867623120546341, + 0.00513108167797327, + 0.007722186390310526, + 0.00704457052052021, + 0.011283475905656815, + 0.00943154189735651, + 0.007721124682575464, + 0.011126640252768993, + 0.0145815284922719, + 0.010935302823781967, + 0.00608399510383606, + 0.010852627456188202, + 0.010018536821007729, + 0.011026088148355484, + 0.009222643449902534, + 0.01024834904819727, + 0.013449138961732388, + 0.017243385314941406, + 0.00885385274887085, + 0.02359727770090103, + 0.011755836196243763, + 0.023881951346993446, + 0.009273122996091843, + 0.009412257000803947, + 0.01404609065502882, + 0.014900843612849712, + 0.009416000917553902, + 0.014379551634192467, + 0.00942113995552063, + 0.022116893902420998, + 0.00953060295432806, + 0.012594472616910934, + 0.01482311263680458, + 0.015063251368701458, + 0.005384329240769148, + 0.010470700450241566, + 0.007388412021100521, + 0.010024823248386383, + 0.008240082301199436, + 0.006199768744409084, + 0.010010989382863045, + 0.0069006322883069515, + 0.00474638631567359, + 0.006440060678869486, + 0.004502162802964449, + 0.009272409602999687, + 0.010489859618246555, + 0.008919216692447662, + 0.010004420764744282, + 0.006648003123700619, + 0.004721499979496002, + 0.006567280739545822, + 0.0042585330083966255, + 0.0073320139199495316, + 0.006835716776549816, + 0.005557290744036436, + 0.008492988534271717, + 0.0043580187484622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.02328958734869957 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, captured in the signature Studio Ghibli animation style, takes place from a slightly overhead angle. The camera remains still, focused on a textured, aged wooden table. Bathed in the warm, golden light slanting through the window in the late afternoon, the wood's grain and subtle scratches are clearly visible. A pair of hands, with healthy complexion, distinct knuckles, and a hint of warmth, rest upon the tabletop, their skin taking on a soft sheen in the light. Carefully and deliberately, these hands gently sweep aside a thin layer of fine dust that has accumulated on the tabletop with the sides of their palms. Each subtle movement is rendered with incredible fluidity by the extremely high frame rate. The sunlight illuminates the lifted dust particles, causing them to dance briefly, circling in the air like shimmering golden light before slowly settling. The warm, muted tones of the image evoke a sense of nostalgia and tranquility. The soft glow of the hands contrasts subtly with the matte texture of the wooden table, evoking a sense of focused and healing daily ritual.", + "seed": 3324, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_11/index_3420_seed_3324_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10790359871461988, + "vlm_logprob": -0.10790359871461988, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.10790359871461988, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03828524798154831 + ], + "opposite_logprob_tokens": [ + -5.413285255432129 + ], + "logprob_reward": -0.03828524798154831, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.03828524798154831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3706735074520111, + 0.08814273774623871, + 0.06676829606294632, + 0.09108489006757736, + 0.03537294641137123, + 0.20581452548503876, + 0.053930435329675674, + 0.0372823104262352, + 0.0230336245149374, + 0.027896679937839508 + ], + "credit_map": [ + 0.020833425223827362, + 0.02874160371720791, + 0.08870402723550797, + 0.09618508815765381, + 0.03257935494184494, + 0.019325030967593193, + 0.06312055140733719, + 0.021184412762522697, + 0.007873977534472942, + 0.008162221871316433, + 0.01718030497431755, + 0.009965729899704456, + 0.01721131056547165, + 0.00894371047616005, + 0.007296641822904348, + 0.011508838273584843, + 0.007315807044506073, + 0.0054740519262850285, + 0.009024351835250854, + 0.018387941643595695, + 0.005812551826238632, + 0.005073245149105787, + 0.008996332064270973, + 0.00668401550501585, + 0.008256100118160248, + 0.019356150180101395, + 0.010971182957291603, + 0.009860129095613956, + 0.012991742230951786, + 0.012803707271814346, + 0.007599943783134222, + 0.009245941415429115, + 0.004786314442753792, + 0.005026098806411028, + 0.004934605211019516, + 0.006667974404990673, + 0.004265562165528536, + 0.0026612915098667145, + 0.0035838244948536158, + 0.0034472746774554253, + 0.008250942453742027, + 0.009292436763644218, + 0.06531015038490295, + 0.07919372618198395, + 0.007077469956129789, + 0.006591962184756994, + 0.021024735644459724, + 0.00907311961054802, + 0.006219613831490278, + 0.007758853957056999, + 0.010343245230615139, + 0.00706520164385438, + 0.004724326077848673, + 0.005614510737359524, + 0.00481428625062108, + 0.007390392012894154, + 0.005755540914833546, + 0.008759156800806522, + 0.005598416551947594, + 0.004109212197363377, + 0.002881397260352969, + 0.003146531293168664, + 0.003767278976738453, + 0.0032647750340402126, + 0.003719559172168374, + 0.005064378958195448, + 0.0035149098839610815, + 0.002830859273672104, + 0.0019084499217569828, + 0.0016063135117292404, + 0.0022534525487571955, + 0.0021357000805437565, + 0.004743700847029686, + 0.003232373623177409, + 0.0058433073572814465, + 0.003479686565697193, + 0.0020269728265702724, + 0.002410688903182745, + 0.0030424529686570168, + 0.0031174945179373026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.03828524798154831 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42892441153526306 + ], + "opposite_logprob_tokens": [ + -1.0539244413375854 + ], + "logprob_reward": -0.42892441153526306, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42892441153526306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20619629323482513, + 0.15154947340488434, + 0.19412671029567719, + 0.07374857366085052, + 0.1093216985464096, + 0.12211217731237411, + 0.03950244188308716, + 0.021242978051304817, + 0.037497375160455704, + 0.04470229893922806 + ], + "credit_map": [ + 0.015587943606078625, + 0.017446519806981087, + 0.03021535463631153, + 0.024477176368236542, + 0.018109818920493126, + 0.021990109235048294, + 0.045917194336652756, + 0.032452166080474854, + 0.021065866574645042, + 0.012803458608686924, + 0.02108187973499298, + 0.023588381707668304, + 0.02586095966398716, + 0.012865547090768814, + 0.016206994652748108, + 0.018076376989483833, + 0.020020080730319023, + 0.017257345840334892, + 0.023326437920331955, + 0.03818724304437637, + 0.026132460683584213, + 0.020139778032898903, + 0.024264158681035042, + 0.02479921095073223, + 0.008537587709724903, + 0.009068437851965427, + 0.01135012786835432, + 0.010881387628614902, + 0.010248254984617233, + 0.008137376047670841, + 0.006998852360993624, + 0.008526554331183434, + 0.009609947912395, + 0.019828764721751213, + 0.01256803423166275, + 0.01695762388408184, + 0.016338050365447998, + 0.012592770159244537, + 0.0090538514778018, + 0.012372652068734169, + 0.008371934294700623, + 0.011524075642228127, + 0.033566925674676895, + 0.03828319534659386, + 0.0061973100528120995, + 0.004506109748035669, + 0.010681919753551483, + 0.0089807054027915, + 0.003287478117272258, + 0.006488845683634281, + 0.006201328709721565, + 0.0034581380896270275, + 0.006205422338098288, + 0.005519986618310213, + 0.0038959258235991, + 0.00444531487300992, + 0.0025848643854260445, + 0.004098385572433472, + 0.0035764826461672783, + 0.002320772735401988, + 0.0018292708555236459, + 0.002120664343237877, + 0.0020599274430423975, + 0.0026526087895035744, + 0.007439242210239172, + 0.006001833360642195, + 0.005559243261814117, + 0.003684906056150794, + 0.003960228059440851, + 0.004040358122438192, + 0.0035815683659166098, + 0.0032300003804266453, + 0.007481880020350218, + 0.00897607859224081, + 0.006112696602940559, + 0.004639127291738987, + 0.004776447545737028, + 0.004643938038498163, + 0.0037350384518504143, + 0.004337092395871878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.42892441153526306 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005966352764517069 + ], + "opposite_logprob_tokens": [ + -5.1309661865234375 + ], + "logprob_reward": -0.005966352764517069, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005966352764517069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20604953169822693, + 0.13378749787807465, + 0.07736900448799133, + 0.25427189469337463, + 0.04702066257596016, + 0.16828003525733948, + 0.02613486722111702, + 0.026982741430401802, + 0.031710680574178696, + 0.02839307114481926 + ], + "credit_map": [ + 0.012316410429775715, + 0.020525624975562096, + 0.03654855489730835, + 0.030705798417329788, + 0.01729409582912922, + 0.024201329797506332, + 0.04087723419070244, + 0.023580489680171013, + 0.008501040749251842, + 0.01386336050927639, + 0.018064798787236214, + 0.021480100229382515, + 0.022509664297103882, + 0.011496705003082752, + 0.02711801789700985, + 0.010753816924989223, + 0.005653555039316416, + 0.006260695867240429, + 0.00946304202079773, + 0.013044137507677078, + 0.00785911362618208, + 0.008124684914946556, + 0.01380377821624279, + 0.013159996829926968, + 0.02414143644273281, + 0.04340563341975212, + 0.035505134612321854, + 0.029913021251559258, + 0.039420194923877716, + 0.03221255540847778, + 0.023805364966392517, + 0.025868553668260574, + 0.004427020903676748, + 0.009045809507369995, + 0.0069470093585550785, + 0.010365964844822884, + 0.004371452145278454, + 0.003503351006656885, + 0.003556012874469161, + 0.004804043564945459, + 0.007701068185269833, + 0.009822407737374306, + 0.05318232253193855, + 0.06084117665886879, + 0.006500374060124159, + 0.005232522264122963, + 0.016370702534914017, + 0.008629458956420422, + 0.0028780896682292223, + 0.004440837074071169, + 0.0033284667879343033, + 0.002973656402900815, + 0.0037903469055891037, + 0.002879355801269412, + 0.003049273509532213, + 0.0027948433998972178, + 0.004103826824575663, + 0.003745364723727107, + 0.004368945490568876, + 0.0032360267359763384, + 0.0021678998600691557, + 0.0020523385610431433, + 0.00430241459980607, + 0.0030059251002967358, + 0.004068339709192514, + 0.008332286030054092, + 0.004231429658830166, + 0.003691556863486767, + 0.003454462392255664, + 0.00279281591065228, + 0.0027844796422868967, + 0.0023553119972348213, + 0.0024435657542198896, + 0.0020772074349224567, + 0.006507005076855421, + 0.004877070430666208, + 0.0017163400771096349, + 0.0021706661209464073, + 0.005615927744656801, + 0.002985290251672268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.005966352764517069 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328725904226303 + ], + "opposite_logprob_tokens": [ + -3.773287296295166 + ], + "logprob_reward": -0.02328725904226303, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328725904226303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28330332040786743, + 0.07772106677293777, + 0.1367599219083786, + 0.0729629173874855, + 0.0978776216506958, + 0.27174681425094604, + 0.017647452652454376, + 0.010957619175314903, + 0.020952554419636726, + 0.010070671327412128 + ], + "credit_map": [ + 0.015683433040976524, + 0.021125836297869682, + 0.05786842480301857, + 0.06763114035129547, + 0.02919219806790352, + 0.02498851716518402, + 0.04890868440270424, + 0.01790510304272175, + 0.005225832108408213, + 0.006370759569108486, + 0.010713081806898117, + 0.012183597311377525, + 0.008957664482295513, + 0.008982767350971699, + 0.01659052073955536, + 0.008696841076016426, + 0.009317279793322086, + 0.009542367421090603, + 0.017537636682391167, + 0.04080187529325485, + 0.01693042926490307, + 0.01173480600118637, + 0.01778116077184677, + 0.01311437413096428, + 0.00802729930728674, + 0.00573781318962574, + 0.010668846778571606, + 0.019488178193569183, + 0.007629210129380226, + 0.006392739247530699, + 0.006836134009063244, + 0.008182690478861332, + 0.00861589889973402, + 0.011384203098714352, + 0.022287528961896896, + 0.0244473684579134, + 0.00882111769169569, + 0.006339505314826965, + 0.008440835401415825, + 0.0075411670841276646, + 0.013941872864961624, + 0.0172197837382555, + 0.08214877545833588, + 0.09192300587892532, + 0.015544154681265354, + 0.011129430495202541, + 0.025766810402274132, + 0.014072995632886887, + 0.002026521833613515, + 0.0034100974444299936, + 0.003790149698033929, + 0.0016208692686632276, + 0.0017802993534132838, + 0.0014543668366968632, + 0.0013516547624021769, + 0.0022134939208626747, + 0.001251466921530664, + 0.0023412525188177824, + 0.0012282152893021703, + 0.0012189006665721536, + 0.0012602410279214382, + 0.0012154466239735484, + 0.0013851181138306856, + 0.0010569777805358171, + 0.003187784692272544, + 0.006197498645633459, + 0.0026802490465343, + 0.0024513774551451206, + 0.0019151367014274001, + 0.0015448095509782434, + 0.001323469216004014, + 0.0016522285295650363, + 0.0011661042226478457, + 0.0011304538929834962, + 0.001556617091409862, + 0.0013480719644576311, + 0.0010985920671373606, + 0.0013713770313188434, + 0.0013809027150273323, + 0.0010185523424297571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.02328725904226303 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043054722249507904 + ], + "opposite_logprob_tokens": [ + -3.1680548191070557 + ], + "logprob_reward": -0.043054722249507904, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043054722249507904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.134475439786911, + 0.059266552329063416, + 0.04319922253489494, + 0.13140350580215454, + 0.12026669085025787, + 0.44956842064857483, + 0.030182883143424988, + 0.016520075500011444, + 0.006663296837359667, + 0.00845387950539589 + ], + "credit_map": [ + 0.009093603119254112, + 0.010613291524350643, + 0.025397323071956635, + 0.03387168049812317, + 0.013056250289082527, + 0.008277631364762783, + 0.02351417765021324, + 0.01065149437636137, + 0.0039687370881438255, + 0.00541717791929841, + 0.009578459896147251, + 0.007577484007924795, + 0.015031997114419937, + 0.005600816570222378, + 0.007626923266798258, + 0.004464955534785986, + 0.002746679587289691, + 0.004145610146224499, + 0.0059206318110227585, + 0.006380204111337662, + 0.007038396317511797, + 0.002912849420681596, + 0.004767282400280237, + 0.009287568740546703, + 0.011112633161246777, + 0.020059917122125626, + 0.01863902062177658, + 0.01791846752166748, + 0.021632589399814606, + 0.017107568681240082, + 0.012986566871404648, + 0.011946741491556168, + 0.01328037679195404, + 0.02270290069282055, + 0.012311743572354317, + 0.016656918451189995, + 0.018919605761766434, + 0.013982276432216167, + 0.009009558707475662, + 0.013403302989900112, + 0.01896439678966999, + 0.020329363644123077, + 0.14755317568778992, + 0.16884084045886993, + 0.01556357927620411, + 0.011814638040959835, + 0.04536480829119682, + 0.021137624979019165, + 0.0035323542542755604, + 0.0034506069496273994, + 0.004404399544000626, + 0.003201359184458852, + 0.0041232602670788765, + 0.0036906818859279156, + 0.0034918554592877626, + 0.004288366064429283, + 0.001627958845347166, + 0.002428686246275902, + 0.0027924194000661373, + 0.0017028979491442442, + 0.0016108104027807713, + 0.0018617446767166257, + 0.0020069240126758814, + 0.0024886340834200382, + 0.0011867154389619827, + 0.0010014427825808525, + 0.001047816826030612, + 0.0007624506833963096, + 0.0006378968828357756, + 0.0005409685545600951, + 0.0007880197372287512, + 0.0006979857571423054, + 0.0012788893654942513, + 0.001108613912947476, + 0.0012286689598113298, + 0.0010682926513254642, + 0.000652171962428838, + 0.0009046016493812203, + 0.0011469235178083181, + 0.001065717195160687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.043054722249507904 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, captured in the signature Studio Ghibli animation style, takes place from a slightly overhead angle. The camera remains still, focused on a textured, aged wooden table. Bathed in the warm, golden light slanting through the window in the late afternoon, the wood's grain and subtle scratches are clearly visible. A pair of hands, with healthy complexion, distinct knuckles, and a hint of warmth, rest upon the tabletop, their skin taking on a soft sheen in the light. Carefully and deliberately, these hands gently sweep aside a thin layer of fine dust that has accumulated on the tabletop with the sides of their palms. Each subtle movement is rendered with incredible fluidity by the extremely high frame rate. The sunlight illuminates the lifted dust particles, causing them to dance briefly, circling in the air like shimmering golden light before slowly settling. The warm, muted tones of the image evoke a sense of nostalgia and tranquility. The soft glow of the hands contrasts subtly with the matte texture of the wooden table, evoking a sense of focused and healing daily ritual.", + "seed": 3325, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_11/index_3420_seed_3325_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1102654479444027, + "vlm_logprob": -0.1102654479444027, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.1102654479444027, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02851926162838936 + ], + "opposite_logprob_tokens": [ + -5.278519153594971 + ], + "logprob_reward": -0.02851926162838936, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.02851926162838936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15881788730621338, + 0.09095653891563416, + 0.07245952636003494, + 0.1372131109237671, + 0.13569845259189606, + 0.08309659361839294, + 0.06995396316051483, + 0.0791085958480835, + 0.056519441306591034, + 0.11617589741945267 + ], + "credit_map": [ + 0.01989191398024559, + 0.017286011949181557, + 0.018998583778738976, + 0.03314900025725365, + 0.009222514927387238, + 0.010753108188509941, + 0.018254097551107407, + 0.03126265108585358, + 0.01375130470842123, + 0.013571475632488728, + 0.009503320790827274, + 0.010565142147243023, + 0.006901596672832966, + 0.014138872735202312, + 0.011444690637290478, + 0.011080143041908741, + 0.010428550653159618, + 0.010705960914492607, + 0.00612312788143754, + 0.006647552363574505, + 0.004608852323144674, + 0.0071333940140903, + 0.015333905816078186, + 0.011478187516331673, + 0.021792534738779068, + 0.016418974846601486, + 0.010405171662569046, + 0.008754227310419083, + 0.013071500696241856, + 0.016353389248251915, + 0.027522560209035873, + 0.02289474941790104, + 0.031101156026124954, + 0.014881295152008533, + 0.014506231062114239, + 0.011715163476765156, + 0.010988039895892143, + 0.012216313742101192, + 0.020692843943834305, + 0.019597411155700684, + 0.014989648014307022, + 0.00778290256857872, + 0.011568712070584297, + 0.010167742148041725, + 0.0064044189639389515, + 0.010307618416845798, + 0.011894183233380318, + 0.009981376118957996, + 0.010557166300714016, + 0.004925362765789032, + 0.008709298446774483, + 0.011809900403022766, + 0.004919077269732952, + 0.006416209042072296, + 0.012759631499648094, + 0.009857317432761192, + 0.013352787122130394, + 0.005948479287326336, + 0.00999781396239996, + 0.010952966287732124, + 0.0061366320587694645, + 0.011041811667382717, + 0.012389522045850754, + 0.009288582019507885, + 0.008380505256354809, + 0.004810612183064222, + 0.00860767625272274, + 0.00941203162074089, + 0.004174343775957823, + 0.004668984562158585, + 0.008967647328972816, + 0.007497641257941723, + 0.01456031296402216, + 0.015464290045201778, + 0.016410909593105316, + 0.021786276251077652, + 0.007936199195683002, + 0.007854471914470196, + 0.016585038974881172, + 0.015578396618366241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.02851926162838936 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34865573048591614 + ], + "opposite_logprob_tokens": [ + -1.2236557006835938 + ], + "logprob_reward": -0.34865573048591614, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34865573048591614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13279134035110474, + 0.23588290810585022, + 0.11990968883037567, + 0.05208614841103554, + 0.09142296016216278, + 0.06595925241708755, + 0.05803265422582626, + 0.07788342982530594, + 0.09475220739841461, + 0.07127948850393295 + ], + "credit_map": [ + 0.014841838739812374, + 0.02101488970220089, + 0.020615365356206894, + 0.023661229759454727, + 0.00901908427476883, + 0.011868191882967949, + 0.012567766010761261, + 0.019202981144189835, + 0.020688513293862343, + 0.0223370511084795, + 0.026064718142151833, + 0.027394775301218033, + 0.011715088970959187, + 0.013918913900852203, + 0.06135820224881172, + 0.05240562930703163, + 0.014241237193346024, + 0.007972439751029015, + 0.013152709230780602, + 0.018014736473560333, + 0.0060773673467338085, + 0.008136878721415997, + 0.03365180268883705, + 0.01866251975297928, + 0.006711058784276247, + 0.005600119475275278, + 0.006058377213776112, + 0.004583298694342375, + 0.003482223954051733, + 0.004378125537186861, + 0.012633304111659527, + 0.008639639243483543, + 0.015185750089585781, + 0.013089626096189022, + 0.008853104896843433, + 0.007892249152064323, + 0.008465723134577274, + 0.009416659362614155, + 0.014574850909411907, + 0.013944996520876884, + 0.017081324011087418, + 0.006912874523550272, + 0.00865815207362175, + 0.008172059431672096, + 0.0036357806529849768, + 0.006102010607719421, + 0.007629254832863808, + 0.007767797447741032, + 0.00633909972384572, + 0.004363507032394409, + 0.007964994758367538, + 0.010350435972213745, + 0.005091738421469927, + 0.006607663817703724, + 0.010864868760108948, + 0.006450343877077103, + 0.013896794989705086, + 0.008100457489490509, + 0.010807211510837078, + 0.007379133254289627, + 0.008750212378799915, + 0.012617850676178932, + 0.007326213177293539, + 0.009005558677017689, + 0.017670486122369766, + 0.00884370505809784, + 0.0142361493781209, + 0.015795500949025154, + 0.005352464504539967, + 0.007922233082354069, + 0.014836069196462631, + 0.010095600970089436, + 0.007981948554515839, + 0.00637139193713665, + 0.00971763115376234, + 0.011819628067314625, + 0.008161201141774654, + 0.010315819643437862, + 0.00823926366865635, + 0.008672604337334633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.34865573048591614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012559015303850174 + ], + "opposite_logprob_tokens": [ + -4.387558937072754 + ], + "logprob_reward": -0.012559015303850174, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012559015303850174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1695840209722519, + 0.12422491610050201, + 0.05190346762537956, + 0.058281779289245605, + 0.04897832125425339, + 0.03282535448670387, + 0.09138084948062897, + 0.06392897665500641, + 0.15037725865840912, + 0.20851513743400574 + ], + "credit_map": [ + 0.01710524782538414, + 0.016654295846819878, + 0.013904641382396221, + 0.023507963865995407, + 0.008389820344746113, + 0.013749842531979084, + 0.02272600308060646, + 0.05354621261358261, + 0.01034939382225275, + 0.00859595462679863, + 0.007929486222565174, + 0.007828610949218273, + 0.006993998307734728, + 0.008400722406804562, + 0.03823436424136162, + 0.035892385989427567, + 0.007808694615960121, + 0.006176899187266827, + 0.0060971942730247974, + 0.005509245675057173, + 0.0040311082266271114, + 0.004757467657327652, + 0.010226248763501644, + 0.007296609692275524, + 0.0065705906599760056, + 0.008243652060627937, + 0.006778012961149216, + 0.006656317505985498, + 0.004669658374041319, + 0.0049134218133986, + 0.011883402243256569, + 0.008566726930439472, + 0.007010025903582573, + 0.006113728508353233, + 0.004187365993857384, + 0.008254731073975563, + 0.0025951373390853405, + 0.004574720747768879, + 0.010179426521062851, + 0.006063186097890139, + 0.006139381323009729, + 0.0033727940171957016, + 0.003987280651926994, + 0.004782079253345728, + 0.0021600183099508286, + 0.0026024356484413147, + 0.005270557012408972, + 0.004510807804763317, + 0.009533868171274662, + 0.006141938269138336, + 0.011650288477540016, + 0.017696494236588478, + 0.005756567232310772, + 0.010281041264533997, + 0.017554640769958496, + 0.012766013853251934, + 0.011938139796257019, + 0.006582184229046106, + 0.009122761897742748, + 0.007984192110598087, + 0.005230435635894537, + 0.007465599570423365, + 0.008310971781611443, + 0.007294694427400827, + 0.01473999209702015, + 0.016309576109051704, + 0.01979687251150608, + 0.03580233454704285, + 0.008751007728278637, + 0.011046928353607655, + 0.027114752680063248, + 0.0168157871812582, + 0.03146594390273094, + 0.02578539028763771, + 0.031237443909049034, + 0.039455875754356384, + 0.013471286743879318, + 0.01356247253715992, + 0.026436733081936836, + 0.027100007981061935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.012559015303850174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866738244891167 + ], + "opposite_logprob_tokens": [ + -3.0486674308776855 + ], + "logprob_reward": -0.04866738244891167, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866738244891167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21282660961151123, + 0.09286240488290787, + 0.12831413745880127, + 0.0764121413230896, + 0.07435450702905655, + 0.0740683525800705, + 0.03501866012811661, + 0.057203710079193115, + 0.09595584124326706, + 0.15298354625701904 + ], + "credit_map": [ + 0.02227015793323517, + 0.0339333675801754, + 0.026517627760767937, + 0.03916778415441513, + 0.015356474556028843, + 0.020738979801535606, + 0.024762390181422234, + 0.030079849064350128, + 0.010241355746984482, + 0.013511279597878456, + 0.005619020666927099, + 0.007399780210107565, + 0.005519058555364609, + 0.008400330320000648, + 0.02270233817398548, + 0.019469251856207848, + 0.017685946077108383, + 0.014001446776092052, + 0.01457587257027626, + 0.011912771500647068, + 0.008135135285556316, + 0.009618803858757019, + 0.03180238604545593, + 0.02058178000152111, + 0.010670121759176254, + 0.01110883429646492, + 0.005999418441206217, + 0.006624422036111355, + 0.0067020561546087265, + 0.006397883407771587, + 0.01816520094871521, + 0.01074421126395464, + 0.017402680590748787, + 0.008454899303615093, + 0.00850619375705719, + 0.008057006634771824, + 0.007444859482347965, + 0.005356680601835251, + 0.00916688796132803, + 0.009965305216610432, + 0.008739129640161991, + 0.01200870145112276, + 0.010401204228401184, + 0.01403950247913599, + 0.0057030413299798965, + 0.005697497632354498, + 0.008591735735535622, + 0.008887547999620438, + 0.0034326619934290648, + 0.0032136505469679832, + 0.0041519468650221825, + 0.004225284792482853, + 0.0036500021815299988, + 0.006418521516025066, + 0.005202281288802624, + 0.004724313039332628, + 0.008008996024727821, + 0.0054845912382006645, + 0.008993624709546566, + 0.008627312257885933, + 0.005310975015163422, + 0.006464702542871237, + 0.00833901111036539, + 0.005974497180432081, + 0.01770544797182083, + 0.009609783999621868, + 0.01396469958126545, + 0.01840350590646267, + 0.005201578140258789, + 0.006012818776071072, + 0.014253629371523857, + 0.01080439891666174, + 0.02309115044772625, + 0.018449578434228897, + 0.02095693349838257, + 0.02760615013539791, + 0.009415202774107456, + 0.010416046716272831, + 0.024522459134459496, + 0.018526041880249977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.04866738244891167 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292584985494614 + ], + "opposite_logprob_tokens": [ + -2.2379257678985596 + ], + "logprob_reward": -0.11292584985494614, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292584985494614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23795609176158905, + 0.09206345677375793, + 0.14845184981822968, + 0.11209642887115479, + 0.0711287409067154, + 0.04415766894817352, + 0.06025698781013489, + 0.08805825561285019, + 0.07913697510957718, + 0.066693514585495 + ], + "credit_map": [ + 0.024445831775665283, + 0.04382502660155296, + 0.022192662581801414, + 0.03771454468369484, + 0.01832101121544838, + 0.026801545172929764, + 0.024098483845591545, + 0.04055697098374367, + 0.01591488905251026, + 0.008544915355741978, + 0.013420979492366314, + 0.012297498993575573, + 0.006902589462697506, + 0.011669928207993507, + 0.009911669418215752, + 0.013400985859334469, + 0.025594700127840042, + 0.016042547300457954, + 0.01799706183373928, + 0.016115998849272728, + 0.010105970315635204, + 0.010408682748675346, + 0.03262989968061447, + 0.019557002931833267, + 0.019555438309907913, + 0.01865195855498314, + 0.013801674358546734, + 0.011982925236225128, + 0.00880817323923111, + 0.007896395400166512, + 0.014919775538146496, + 0.016480084508657455, + 0.008263854309916496, + 0.011101108975708485, + 0.007476028986275196, + 0.007828407920897007, + 0.0076701343059539795, + 0.008711866103112698, + 0.012360448017716408, + 0.007716894615441561, + 0.00881254393607378, + 0.0045433565974235535, + 0.006227597594261169, + 0.00420601898804307, + 0.0036641357000917196, + 0.006341623608022928, + 0.005401224363595247, + 0.004961166065186262, + 0.011040318757295609, + 0.008021212182939053, + 0.006263034883886576, + 0.008546416647732258, + 0.004177158698439598, + 0.006304294802248478, + 0.007919162511825562, + 0.007985389791429043, + 0.012902461923658848, + 0.012558918446302414, + 0.012652349658310413, + 0.008173593319952488, + 0.008962533436715603, + 0.012947852723300457, + 0.01010138913989067, + 0.009759160690009594, + 0.013301671482622623, + 0.009487044997513294, + 0.011322162114083767, + 0.01411687396466732, + 0.004436474293470383, + 0.0045753964222967625, + 0.012307333759963512, + 0.009590012021362782, + 0.0077962325885891914, + 0.006305935326963663, + 0.007869378663599491, + 0.01007862389087677, + 0.007233703043311834, + 0.00914338044822216, + 0.01020632591098547, + 0.00805993378162384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.11292584985494614 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, captured in the signature Studio Ghibli animation style, takes place from a slightly overhead angle. The camera remains still, focused on a textured, aged wooden table. Bathed in the warm, golden light slanting through the window in the late afternoon, the wood's grain and subtle scratches are clearly visible. A pair of hands, with healthy complexion, distinct knuckles, and a hint of warmth, rest upon the tabletop, their skin taking on a soft sheen in the light. Carefully and deliberately, these hands gently sweep aside a thin layer of fine dust that has accumulated on the tabletop with the sides of their palms. Each subtle movement is rendered with incredible fluidity by the extremely high frame rate. The sunlight illuminates the lifted dust particles, causing them to dance briefly, circling in the air like shimmering golden light before slowly settling. The warm, muted tones of the image evoke a sense of nostalgia and tranquility. The soft glow of the hands contrasts subtly with the matte texture of the wooden table, evoking a sense of focused and healing daily ritual.", + "seed": 3326, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_11/index_3420_seed_3326_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2470295530743897, + "vlm_logprob": -0.2470295530743897, + "vlm_yes_no_margin": 3.449999988079071, + "vlm_token_credit": -0.2470295530743897, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033389389514923096 + ], + "opposite_logprob_tokens": [ + -5.658389568328857 + ], + "logprob_reward": -0.033389389514923096, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.033389389514923096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17369891703128815, + 0.026432476937770844, + 0.07587416470050812, + 0.05314971134066582, + 0.06463892012834549, + 0.04211542755365372, + 0.03413939103484154, + 0.04523308202624321, + 0.4268660247325897, + 0.05785185098648071 + ], + "credit_map": [ + 0.02430698461830616, + 0.027515949681401253, + 0.05674998089671135, + 0.01937090791761875, + 0.010799813084304333, + 0.013156527653336525, + 0.009839939884841442, + 0.011958806775510311, + 0.0026044759433716536, + 0.0035545560531318188, + 0.0070241945795714855, + 0.003483859123662114, + 0.002508093137294054, + 0.003381006885319948, + 0.0024195113219320774, + 0.0014567796606570482, + 0.016182364895939827, + 0.02003251016139984, + 0.012800158932805061, + 0.009580308571457863, + 0.0036829132586717606, + 0.0056382352486252785, + 0.003824191866442561, + 0.00413347827270627, + 0.005358929745852947, + 0.00896976888179779, + 0.005387104116380215, + 0.0052133770659565926, + 0.008836663328111172, + 0.007891535758972168, + 0.006087547168135643, + 0.005404779687523842, + 0.012907249853014946, + 0.002917793346568942, + 0.0055288951843976974, + 0.02425326406955719, + 0.0027571129612624645, + 0.0034592929296195507, + 0.009068767540156841, + 0.003746544010937214, + 0.006168331019580364, + 0.0035504756961017847, + 0.006889196578413248, + 0.008485253900289536, + 0.0035958262160420418, + 0.002278734929859638, + 0.00523670157417655, + 0.005910907406359911, + 0.005878651514649391, + 0.0050537437200546265, + 0.003127139760181308, + 0.0062215281650424, + 0.0028632567264139652, + 0.0024392681661993265, + 0.004166788887232542, + 0.004389013163745403, + 0.009209327399730682, + 0.005349154118448496, + 0.007326013408601284, + 0.00769612193107605, + 0.0027857490349560976, + 0.004227524623274803, + 0.004764136392623186, + 0.003875058377161622, + 0.05170042812824249, + 0.17253810167312622, + 0.02172737941145897, + 0.1329273134469986, + 0.015789607539772987, + 0.007137839682400227, + 0.01657232828438282, + 0.008473044261336327, + 0.005704360082745552, + 0.005943079479038715, + 0.017654702067375183, + 0.008385778404772282, + 0.005846838466823101, + 0.004792981315404177, + 0.004349343478679657, + 0.00517477048560977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.033389389514923096 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370494365692139 + ], + "opposite_logprob_tokens": [ + -0.3870493769645691 + ], + "logprob_reward": -1.1370494365692139, + "yes_no_margin_reward": -0.7500000596046448, + "token_credit_reward": -1.1370494365692139, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1276046484708786, + 0.0649585947394371, + 0.046870797872543335, + 0.04157821461558342, + 0.03862437978386879, + 0.02809874527156353, + 0.030866941437125206, + 0.07483617961406708, + 0.37637558579444885, + 0.1701859086751938 + ], + "credit_map": [ + 0.01699569821357727, + 0.018524818122386932, + 0.037188008427619934, + 0.015016183257102966, + 0.008882653899490833, + 0.012617901898920536, + 0.008850482292473316, + 0.009528887458145618, + 0.00786355696618557, + 0.009612292982637882, + 0.020998086780309677, + 0.009714513085782528, + 0.004668996669352055, + 0.00477995490655303, + 0.0037787777837365866, + 0.0035424039233475924, + 0.008079023100435734, + 0.012294525280594826, + 0.005917625967413187, + 0.004823098890483379, + 0.003364107571542263, + 0.005208588670939207, + 0.0036991939414292574, + 0.0034846316557377577, + 0.004256622400134802, + 0.0054492042399942875, + 0.005069104954600334, + 0.00728689506649971, + 0.0054956162348389626, + 0.005449424032121897, + 0.0042662788182497025, + 0.004305065143853426, + 0.003948700614273548, + 0.0032861451618373394, + 0.007091089617460966, + 0.006396715063601732, + 0.003369199577718973, + 0.0032753469422459602, + 0.006604241672903299, + 0.004652935545891523, + 0.0031086639501154423, + 0.0037540472112596035, + 0.0028708206955343485, + 0.006654866971075535, + 0.003779306774958968, + 0.003027688479050994, + 0.003134441329166293, + 0.0017689047381281853, + 0.0045920610427856445, + 0.003639213042333722, + 0.004025063011795282, + 0.00569138815626502, + 0.0032670022919774055, + 0.0038924873806536198, + 0.0032937577925622463, + 0.002465965924784541, + 0.004933980293571949, + 0.01944253221154213, + 0.007193299476057291, + 0.026775231584906578, + 0.003791909897699952, + 0.004082227125763893, + 0.005299462471157312, + 0.0033175358548760414, + 0.03599466010928154, + 0.14923018217086792, + 0.020327003672719002, + 0.12733758985996246, + 0.014744064770638943, + 0.00830676406621933, + 0.013406580314040184, + 0.007028722669929266, + 0.01702083833515644, + 0.016188418492674828, + 0.03846930339932442, + 0.02169155515730381, + 0.02181144990026951, + 0.017450813204050064, + 0.018767401576042175, + 0.018786124885082245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -1.1370494365692139 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004655831027776003 + ], + "opposite_logprob_tokens": [ + -5.379655838012695 + ], + "logprob_reward": -0.004655831027776003, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004655831027776003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07785630226135254, + 0.03343695402145386, + 0.13645954430103302, + 0.06116047501564026, + 0.0762660950422287, + 0.040113478899002075, + 0.04427212104201317, + 0.02251993864774704, + 0.4507773816585541, + 0.05713775381445885 + ], + "credit_map": [ + 0.007889913395047188, + 0.009476871229708195, + 0.020662713795900345, + 0.00988111924380064, + 0.007179178297519684, + 0.008848731406033039, + 0.005961081013083458, + 0.007956692948937416, + 0.0035252971574664116, + 0.004550047684460878, + 0.005394658539444208, + 0.007327146828174591, + 0.003986582159996033, + 0.003808866022154689, + 0.0029383793007582426, + 0.0019059769110754132, + 0.038484033197164536, + 0.04855678603053093, + 0.011248422786593437, + 0.011696378700435162, + 0.006262077949941158, + 0.006330287083983421, + 0.0056698378175497055, + 0.008211715146899223, + 0.005812279414385557, + 0.009794466197490692, + 0.006752820685505867, + 0.006747255567461252, + 0.009795377030968666, + 0.009767946787178516, + 0.006461785174906254, + 0.006028544623404741, + 0.008707914501428604, + 0.005982886999845505, + 0.015181457623839378, + 0.012455706484615803, + 0.0072214603424072266, + 0.005471945274621248, + 0.011337222531437874, + 0.009907503612339497, + 0.005304871592670679, + 0.00330749130807817, + 0.006597374100238085, + 0.009110975079238415, + 0.002082457300275564, + 0.0023457747884094715, + 0.006177755072712898, + 0.005186781287193298, + 0.004544827621430159, + 0.004382545128464699, + 0.003993060439825058, + 0.005168789532035589, + 0.0043501961044967175, + 0.004782829433679581, + 0.007385011296719313, + 0.009664855897426605, + 0.0025403571780771017, + 0.0026981753762811422, + 0.003936276771128178, + 0.00423063850030303, + 0.001718859188258648, + 0.002970220288261771, + 0.0025282097049057484, + 0.0018972009420394897, + 0.04972749203443527, + 0.1793327033519745, + 0.02310822531580925, + 0.14462143182754517, + 0.01737293042242527, + 0.009150208905339241, + 0.01730070821940899, + 0.010163652710616589, + 0.005514465272426605, + 0.004893552046269178, + 0.00779931154102087, + 0.006905845832079649, + 0.008839452639222145, + 0.012064092792570591, + 0.004595733713358641, + 0.006525297649204731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.004655831027776003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635362558066845 + ], + "opposite_logprob_tokens": [ + -3.651353597640991 + ], + "logprob_reward": -0.02635362558066845, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635362558066845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13159744441509247, + 0.04042220488190651, + 0.05233738571405411, + 0.042826682329177856, + 0.06627316772937775, + 0.03757530450820923, + 0.04437047988176346, + 0.0985734760761261, + 0.3320893347263336, + 0.1539345234632492 + ], + "credit_map": [ + 0.015593198128044605, + 0.016588270664215088, + 0.01991460844874382, + 0.017881913110613823, + 0.01504451408982277, + 0.018693791702389717, + 0.01394524797797203, + 0.013935892842710018, + 0.004588805604726076, + 0.005565315950661898, + 0.010359375737607479, + 0.006046383175998926, + 0.0029999667312949896, + 0.004888124763965607, + 0.0035611670464277267, + 0.0024130651727318764, + 0.005574801936745644, + 0.0057843513786792755, + 0.012156052514910698, + 0.005982003640383482, + 0.004781785421073437, + 0.007882312871515751, + 0.005599083844572306, + 0.004576994106173515, + 0.004193042404949665, + 0.006994410417973995, + 0.005785074084997177, + 0.0048871892504394054, + 0.0070540569722652435, + 0.0037639846559613943, + 0.005113997031003237, + 0.005034928675740957, + 0.01078709214925766, + 0.005075201392173767, + 0.01008877158164978, + 0.011373259127140045, + 0.0040033552795648575, + 0.005296169780194759, + 0.010780508629977703, + 0.008868810720741749, + 0.005212474148720503, + 0.0038333465345203876, + 0.007228181231766939, + 0.00648554926738143, + 0.0029925801791250706, + 0.0026738347951322794, + 0.005332916043698788, + 0.0038164244033396244, + 0.005828905384987593, + 0.005478362552821636, + 0.0049245539121329784, + 0.005797273945063353, + 0.004590135533362627, + 0.0054298704490065575, + 0.005974540486931801, + 0.0063468380831182, + 0.028007762506604195, + 0.014131895266473293, + 0.006181139498949051, + 0.027927005663514137, + 0.00398542545735836, + 0.006718195043504238, + 0.007511782925575972, + 0.004110268317162991, + 0.03701186180114746, + 0.12441903352737427, + 0.0206136554479599, + 0.10005965828895569, + 0.012593730352818966, + 0.009521484375, + 0.01707257144153118, + 0.010797332040965557, + 0.01821601204574108, + 0.014411063864827156, + 0.03228235989809036, + 0.02657868154346943, + 0.014512363821268082, + 0.01977689191699028, + 0.01204958837479353, + 0.016107551753520966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.02635362558066845 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands with healthy skin and distinct knuckles visible on a textured, aged wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the hands sweep aside a thin layer of fine dust using the sides of their palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are dust particles illuminated by sunlight, appearing to dance and circle in the air before settling?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033699482679367065 + ], + "opposite_logprob_tokens": [ + -3.4086995124816895 + ], + "logprob_reward": -0.033699482679367065, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033699482679367065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1378014087677002, + 0.05835016071796417, + 0.14851301908493042, + 0.03380816802382469, + 0.02880132384598255, + 0.03859665244817734, + 0.03477999195456505, + 0.05444823205471039, + 0.3534294068813324, + 0.11147168278694153 + ], + "credit_map": [ + 0.01471051387488842, + 0.02015450969338417, + 0.03996559605002403, + 0.024854497984051704, + 0.010857976041734219, + 0.009980638511478901, + 0.009783264249563217, + 0.007494395598769188, + 0.005684189032763243, + 0.0078384755179286, + 0.015023447573184967, + 0.008904724381864071, + 0.005323238205164671, + 0.006133364979177713, + 0.006328636314719915, + 0.0031140800565481186, + 0.04411148279905319, + 0.05451909452676773, + 0.01104728039354086, + 0.012927714735269547, + 0.006487324368208647, + 0.005928907543420792, + 0.0050945826806128025, + 0.008396620862185955, + 0.004447249695658684, + 0.002577992156147957, + 0.00741159450262785, + 0.005633409600704908, + 0.0027387109585106373, + 0.0023213601671159267, + 0.003865693463012576, + 0.004812154918909073, + 0.004106089938431978, + 0.002391102025285363, + 0.004291235469281673, + 0.00616299407556653, + 0.0027350056916475296, + 0.0017970229964703321, + 0.00446480605751276, + 0.002853062003850937, + 0.006415717303752899, + 0.00311159691773355, + 0.004160808399319649, + 0.011494976468384266, + 0.002870711963623762, + 0.0020071747712790966, + 0.004816118162125349, + 0.0037195433396846056, + 0.004630048759281635, + 0.004978710785508156, + 0.0059645227156579494, + 0.005892842076718807, + 0.0032991948537528515, + 0.002371219452470541, + 0.003356811124831438, + 0.004286636598408222, + 0.006376162171363831, + 0.01169408019632101, + 0.0055903480388224125, + 0.020900744944810867, + 0.0014029141748324037, + 0.001689925673417747, + 0.0037833841051906347, + 0.00301067135296762, + 0.018023669719696045, + 0.13671444356441498, + 0.019822629168629646, + 0.14059865474700928, + 0.012772985734045506, + 0.006791882682591677, + 0.012980584986507893, + 0.005724539514631033, + 0.014869683422148228, + 0.008839203976094723, + 0.02236652560532093, + 0.01803310215473175, + 0.012585585936903954, + 0.015337230637669563, + 0.009141839109361172, + 0.010298502631485462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.033699482679367065 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_12/index_tensor([5470, 5470])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_12/index_tensor([5470, 5470])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..509b74c2ab32b999ce40530a9819b2dc1759ebfe --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_12/index_tensor([5470, 5470])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The shot begins at an extremely low angle, then shakes violently as the earth trembles violently and heavily. A massive crack rapidly tears open in the dry earth, sending rocks and earth scattering. The camera then rapidly pulls back and rises, revealing a vast underground cavity, sealed for millennia—the underground palace of the Qin Shi Huang Mausoleum—opening up with the collapse of the massive rock at the top. Dazzling sunlight pours through the cracks, slicing through the dark air thick with millennia of dust, creating Tyndall-effect beams of light. Beneath these beams, the thousands of Terracotta Army formations are revealed for the first time in the profound darkness. Each terracotta warrior is rendered with the breathtaking detail of next-generation gaming graphics, the realistic texture of their clay, the minute cracks etched by age, and the intricate braiding and metal clasps of their armor plates clearly visible. Their facial expressions vary, from resolute to pensive. Wielding patinated bronze spears, halberds, and powerful crossbows, they stand in endless formations with unwavering discipline, silently towering over the shattered ground. The entire scene's lighting and shadows are rendered with a physically based, real-time quality. The stark contrast between intense light and absolute darkness creates a solemn, mysterious, and oppressive epic atmosphere. All movement—from vibrations to drifting dust—is rendered at a smooth, high frame rate, delivering an ultimate visual immersion.", + "seed": 8701, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_12/index_5470_seed_8701_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17317942464724184, + "vlm_logprob": -0.17317942464724184, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.17317942464724184, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03175455704331398 + ], + "opposite_logprob_tokens": [ + -3.531754493713379 + ], + "logprob_reward": -0.03175455704331398, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03175455704331398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7253095507621765, + 0.04392462968826294, + 0.024999892339110374, + 0.0759521946310997, + 0.009178402833640575, + 0.00804528221487999, + 0.017732709646224976, + 0.0262001920491457, + 0.009816326200962067, + 0.05884083732962608 + ], + "credit_map": [ + 0.032602231949567795, + 0.039516381919384, + 0.30489158630371094, + 0.2848111689090729, + 0.010468298569321632, + 0.01316077820956707, + 0.027401136234402657, + 0.012458035722374916, + 0.003636940149590373, + 0.0032731383107602596, + 0.006237125489860773, + 0.005150287877768278, + 0.0037876167334616184, + 0.006164302118122578, + 0.00762566551566124, + 0.008049558848142624, + 0.002881609136238694, + 0.0031321803107857704, + 0.002229668665677309, + 0.0030321311205625534, + 0.0038581767585128546, + 0.004191538784652948, + 0.002678245073184371, + 0.0029963452834635973, + 0.004772256128489971, + 0.003291057189926505, + 0.02898765169084072, + 0.010324784554541111, + 0.0029603983275592327, + 0.017406441271305084, + 0.003365855198353529, + 0.004843764007091522, + 0.0009171335841529071, + 0.0012326419819146395, + 0.0010158577933907509, + 0.0011099973926320672, + 0.0009389256010763347, + 0.0013204285642132163, + 0.0007743912283331156, + 0.001869027502834797, + 0.0011999547714367509, + 0.0015643751248717308, + 0.0008551544160582125, + 0.0009024173486977816, + 0.0012179198674857616, + 0.0010211929911747575, + 0.0006801014533266425, + 0.0006041671731509268, + 0.0015763869741931558, + 0.0020400153007358313, + 0.0009349395986646414, + 0.000950376212131232, + 0.0049814521335065365, + 0.0047453115694224834, + 0.0011384442914277315, + 0.0013657854869961739, + 0.0024069705978035927, + 0.002986257430166006, + 0.0025206622667610645, + 0.0022603042889386415, + 0.004405585117638111, + 0.0033112557139247656, + 0.00452022347599268, + 0.0037889357190579176, + 0.0006742173573002219, + 0.0011487191077321768, + 0.0012135490542277694, + 0.0012136074947193265, + 0.0014106694143265486, + 0.0014445394044741988, + 0.0015633211005479097, + 0.0011477043153718114, + 0.0035794624127447605, + 0.0020750730764120817, + 0.0034884801134467125, + 0.0012990442337468266, + 0.002619074657559395, + 0.02331237494945526, + 0.001987727591767907, + 0.020479606464505196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.03175455704331398 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023322198539972305 + ], + "opposite_logprob_tokens": [ + -3.773322105407715 + ], + "logprob_reward": -0.023322198539972305, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023322198539972305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5889708995819092, + 0.05516810715198517, + 0.043762460350990295, + 0.13577865064144135, + 0.02490893565118313, + 0.039785683155059814, + 0.015907134860754013, + 0.032464947551488876, + 0.02055724523961544, + 0.04269598051905632 + ], + "credit_map": [ + 0.026764454320073128, + 0.03392805531620979, + 0.24316534399986267, + 0.22798216342926025, + 0.00875598844140768, + 0.012009943835437298, + 0.02488664537668228, + 0.011478242464363575, + 0.0040228418074548244, + 0.0039261216297745705, + 0.0071109929122030735, + 0.007124603260308504, + 0.006464012432843447, + 0.007612918969243765, + 0.010382541455328465, + 0.00852406956255436, + 0.004666266962885857, + 0.004817478824406862, + 0.0055345194414258, + 0.007540644612163305, + 0.0035819432232528925, + 0.005031019914895296, + 0.007219361141324043, + 0.005371225066483021, + 0.009827756322920322, + 0.005235578864812851, + 0.051312994211912155, + 0.020415736362338066, + 0.005753616336733103, + 0.03110741637647152, + 0.004889255855232477, + 0.007236288394778967, + 0.002481138100847602, + 0.0037240278907120228, + 0.002542948117479682, + 0.002746680285781622, + 0.0044150929898023605, + 0.0045396266505122185, + 0.0026306130457669497, + 0.0018288088031113148, + 0.0067152162082493305, + 0.007809530012309551, + 0.0022517433390021324, + 0.0020149373449385166, + 0.007678160909563303, + 0.00846696738153696, + 0.0026608959306031466, + 0.0021882306318730116, + 0.0017925542779266834, + 0.003615617984905839, + 0.0010884888470172882, + 0.0010291156359016895, + 0.0026851571165025234, + 0.00340158399194479, + 0.0013273971853777766, + 0.0009672211017459631, + 0.002791487844660878, + 0.0057588256895542145, + 0.002269746270030737, + 0.0028147229459136724, + 0.005373004823923111, + 0.003949503879994154, + 0.00526390178129077, + 0.004243755713105202, + 0.001598154311068356, + 0.002259555971249938, + 0.002623292850330472, + 0.0029678703285753727, + 0.0030058829579502344, + 0.002687848638743162, + 0.0027117347344756126, + 0.0027029046323150396, + 0.0029642367735505104, + 0.002238530432805419, + 0.0043983967043459415, + 0.0023599390406161547, + 0.0021465786267071962, + 0.014375902712345123, + 0.0023188763298094273, + 0.011893520131707191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.023322198539972305 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954858381301165 + ], + "opposite_logprob_tokens": [ + -5.130954742431641 + ], + "logprob_reward": -0.005954858381301165, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954858381301165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7459466457366943, + 0.0264437235891819, + 0.018051058053970337, + 0.05285309627652168, + 0.01149433571845293, + 0.026090048253536224, + 0.016226014122366905, + 0.04336504638195038, + 0.017944037914276123, + 0.041585974395275116 + ], + "credit_map": [ + 0.03232552856206894, + 0.041125696152448654, + 0.3129144608974457, + 0.29230302572250366, + 0.010201185010373592, + 0.01543249748647213, + 0.02787228859961033, + 0.01377194095402956, + 0.0022168371360749006, + 0.0019630612805485725, + 0.0048551373183727264, + 0.004179689567536116, + 0.0025448203086853027, + 0.0022460934706032276, + 0.004812964703887701, + 0.003625120036303997, + 0.0015393830835819244, + 0.0012632700381800532, + 0.0031305570155382156, + 0.0024592105764895678, + 0.0015154891880229115, + 0.0018612700514495373, + 0.0032084526028484106, + 0.0030734266620129347, + 0.002948034554719925, + 0.0024786072317510843, + 0.02076268009841442, + 0.011927044950425625, + 0.0021514124237000942, + 0.00737024936825037, + 0.0026635793037712574, + 0.0025514911394566298, + 0.001067176810465753, + 0.001634465530514717, + 0.001287770690396428, + 0.0015879909042268991, + 0.0015590166440233588, + 0.002118095988407731, + 0.0013449920807033777, + 0.0008948267204686999, + 0.0027678709011524916, + 0.005610082298517227, + 0.001482057967223227, + 0.001872227294370532, + 0.006483559496700764, + 0.004672762472182512, + 0.0017929682508111, + 0.001408520620316267, + 0.00198880466632545, + 0.003079212736338377, + 0.0010158686200156808, + 0.0012685003457590938, + 0.0027914131060242653, + 0.003287364263087511, + 0.0014867802383378148, + 0.001308069913648069, + 0.004306871443986893, + 0.004841245245188475, + 0.003926405217498541, + 0.003474222728982568, + 0.006285316310822964, + 0.005425554700195789, + 0.007922772318124771, + 0.007182660512626171, + 0.0010814761044457555, + 0.002537198830395937, + 0.0020849714055657387, + 0.001981719397008419, + 0.003367798402905464, + 0.00229992950335145, + 0.002496224595233798, + 0.0020947193261235952, + 0.0025252588093280792, + 0.002254932653158903, + 0.0024639873299747705, + 0.0018664489034563303, + 0.0018336169887334108, + 0.014085426926612854, + 0.0024449536576867104, + 0.014111347496509552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.005954858381301165 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234801769256592 + ], + "opposite_logprob_tokens": [ + -0.8984801769256592 + ], + "logprob_reward": -0.5234801769256592, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234801769256592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.335767924785614, + 0.035514622926712036, + 0.05343678221106529, + 0.0804852768778801, + 0.025266243144869804, + 0.015939630568027496, + 0.029966475442051888, + 0.028308996930718422, + 0.04691843315958977, + 0.34839555621147156 + ], + "credit_map": [ + 0.01605197973549366, + 0.0197354294359684, + 0.13401290774345398, + 0.13356687128543854, + 0.0056485822424292564, + 0.0077881659381091595, + 0.012698862701654434, + 0.006265103816986084, + 0.003152440069243312, + 0.0023775307927280664, + 0.004739248659461737, + 0.003233408322557807, + 0.003242646111175418, + 0.006010995712131262, + 0.006155728828161955, + 0.006602623965591192, + 0.006509975530207157, + 0.009554912336170673, + 0.006459957454353571, + 0.007445436902344227, + 0.0047817956656217575, + 0.005121811758726835, + 0.007308171130716801, + 0.006254718638956547, + 0.0065044560469686985, + 0.004717820789664984, + 0.02917572110891342, + 0.016837449744343758, + 0.0038277001585811377, + 0.01108443271368742, + 0.004382740240544081, + 0.003954955376684666, + 0.002382434904575348, + 0.0024424237199127674, + 0.0028676206711679697, + 0.004179939161986113, + 0.0036973513197153807, + 0.002891113515943289, + 0.003795056836679578, + 0.0030103055760264397, + 0.002204547869041562, + 0.003189126495271921, + 0.0012620034394785762, + 0.0008984520682133734, + 0.0030726685654371977, + 0.003659348003566265, + 0.0008534123189747334, + 0.0008000717498362064, + 0.0022725481539964676, + 0.0032291098032146692, + 0.0010478958720341325, + 0.0013699696864932775, + 0.008974895812571049, + 0.009499105624854565, + 0.0018008069600909948, + 0.0017721418989822268, + 0.0023331199772655964, + 0.006194583605974913, + 0.0023181524593383074, + 0.00201398809440434, + 0.006487266160547733, + 0.003805067390203476, + 0.0026428354904055595, + 0.002513984916731715, + 0.002860395470634103, + 0.009255480952560902, + 0.004221993498504162, + 0.004686970263719559, + 0.012037192471325397, + 0.005097795743495226, + 0.004882491659373045, + 0.003876114496961236, + 0.02260458655655384, + 0.01842682994902134, + 0.015960274264216423, + 0.014479239471256733, + 0.0147405369207263, + 0.1311354637145996, + 0.013123497366905212, + 0.11792515963315964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.5234801769256592 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813853323459625 + ], + "opposite_logprob_tokens": [ + -1.4063853025436401 + ], + "logprob_reward": -0.2813853323459625, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813853323459625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3507268726825714, + 0.02430333010852337, + 0.029811380431056023, + 0.07133777439594269, + 0.051799528300762177, + 0.15734274685382843, + 0.02796185575425625, + 0.10559769719839096, + 0.0518062561750412, + 0.12931260466575623 + ], + "credit_map": [ + 0.015098389238119125, + 0.02347872406244278, + 0.13765771687030792, + 0.14291714131832123, + 0.004330315627157688, + 0.006869808305054903, + 0.013985720463097095, + 0.006389024667441845, + 0.0021257372573018074, + 0.0017546092858538032, + 0.005132680293172598, + 0.004088882356882095, + 0.002257211832329631, + 0.0024305274710059166, + 0.0032998714596033096, + 0.0032138070091605186, + 0.0032870215363800526, + 0.006285879760980606, + 0.0033717304468154907, + 0.00327136623673141, + 0.0042495629750192165, + 0.00360596994869411, + 0.0032148933969438076, + 0.0025249531026929617, + 0.005582137499004602, + 0.002944707637652755, + 0.026203392073512077, + 0.011064610444009304, + 0.0028697624802589417, + 0.016004692763090134, + 0.0027856873348355293, + 0.0038827748503535986, + 0.0058434512466192245, + 0.005854362156242132, + 0.005618839990347624, + 0.006513539236038923, + 0.00894786138087511, + 0.009826784953474998, + 0.005023802164942026, + 0.0041708797216415405, + 0.017359701916575432, + 0.0337417796254158, + 0.006673464551568031, + 0.007685324642807245, + 0.04021671041846275, + 0.036413319408893585, + 0.009503052569925785, + 0.005749375559389591, + 0.002440165728330612, + 0.004023940768092871, + 0.0018881825963035226, + 0.0023138707038015127, + 0.005399096291512251, + 0.005137857981026173, + 0.003705530194565654, + 0.0030532071832567453, + 0.010309839621186256, + 0.013379420153796673, + 0.008723271079361439, + 0.008822870440781116, + 0.01594145968556404, + 0.01371206995099783, + 0.018331972882151604, + 0.016376785933971405, + 0.0033154760021716356, + 0.008661479689180851, + 0.004252738319337368, + 0.006230308674275875, + 0.010274745523929596, + 0.006014352664351463, + 0.008311689831316471, + 0.0047454568557441235, + 0.0071554724127054214, + 0.0041997069492936134, + 0.006079786457121372, + 0.002879688050597906, + 0.00514715863391757, + 0.04956376925110817, + 0.00454470980912447, + 0.04974229633808136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.2813853323459625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins at an extremely low angle, then shakes violently as the earth trembles violently and heavily. A massive crack rapidly tears open in the dry earth, sending rocks and earth scattering. The camera then rapidly pulls back and rises, revealing a vast underground cavity, sealed for millennia—the underground palace of the Qin Shi Huang Mausoleum—opening up with the collapse of the massive rock at the top. Dazzling sunlight pours through the cracks, slicing through the dark air thick with millennia of dust, creating Tyndall-effect beams of light. Beneath these beams, the thousands of Terracotta Army formations are revealed for the first time in the profound darkness. Each terracotta warrior is rendered with the breathtaking detail of next-generation gaming graphics, the realistic texture of their clay, the minute cracks etched by age, and the intricate braiding and metal clasps of their armor plates clearly visible. Their facial expressions vary, from resolute to pensive. Wielding patinated bronze spears, halberds, and powerful crossbows, they stand in endless formations with unwavering discipline, silently towering over the shattered ground. The entire scene's lighting and shadows are rendered with a physically based, real-time quality. The stark contrast between intense light and absolute darkness creates a solemn, mysterious, and oppressive epic atmosphere. All movement—from vibrations to drifting dust—is rendered at a smooth, high frame rate, delivering an ultimate visual immersion.", + "seed": 8702, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_12/index_5470_seed_8702_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03710415465757251, + "vlm_logprob": -0.03710415465757251, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03710415465757251, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014750470407307148 + ], + "opposite_logprob_tokens": [ + -4.3897504806518555 + ], + "logprob_reward": -0.014750470407307148, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014750470407307148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09091903269290924, + 0.1116367056965828, + 0.029176587238907814, + 0.14065957069396973, + 0.04622281342744827, + 0.10218005627393723, + 0.41429316997528076, + 0.01815466210246086, + 0.02476339600980282, + 0.021993961185216904 + ], + "credit_map": [ + 0.0103346798568964, + 0.010541497729718685, + 0.005645677447319031, + 0.005501795560121536, + 0.02154495194554329, + 0.023241518065333366, + 0.005866451654583216, + 0.008242455311119556, + 0.01548672653734684, + 0.014044984243810177, + 0.011753988452255726, + 0.008447032421827316, + 0.019516117870807648, + 0.011061524040997028, + 0.009803379885852337, + 0.021522946655750275, + 0.0030347497668117285, + 0.0044177221134305, + 0.004281212110072374, + 0.0039031279738992453, + 0.004290638491511345, + 0.003503517946228385, + 0.0028818517457693815, + 0.002863767556846142, + 0.014804830774664879, + 0.009755595587193966, + 0.02583848498761654, + 0.058703429996967316, + 0.005497968755662441, + 0.0050511714071035385, + 0.009905381128191948, + 0.01110271830111742, + 0.006433356087654829, + 0.006949182599782944, + 0.004309038631618023, + 0.0025925789959728718, + 0.005913343280553818, + 0.007529156282544136, + 0.006475049536675215, + 0.006021111272275448, + 0.014475733041763306, + 0.017681671306490898, + 0.01616099290549755, + 0.008325767703354359, + 0.017777342349290848, + 0.013013779185712337, + 0.008677585050463676, + 0.006067182403057814, + 0.025235936045646667, + 0.026783930137753487, + 0.010038753040134907, + 0.009068563580513, + 0.16545335948467255, + 0.16296400129795074, + 0.007502421271055937, + 0.007246213499456644, + 0.0025205393321812153, + 0.002217535860836506, + 0.0019306452013552189, + 0.0019465875811874866, + 0.002917218953371048, + 0.0030887832399457693, + 0.001560779637657106, + 0.001972571946680546, + 0.003286406397819519, + 0.002190183848142624, + 0.002541768131777644, + 0.00600495794788003, + 0.0018543959595263004, + 0.002528000622987747, + 0.002183937933295965, + 0.004173743538558483, + 0.005422589834779501, + 0.0038662569131702185, + 0.0023428979329764843, + 0.0018538993317633867, + 0.0017818263731896877, + 0.0015167692909017205, + 0.0019136139890179038, + 0.003296108217909932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.014750470407307148 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011101503856480122 + ], + "opposite_logprob_tokens": [ + -4.511101722717285 + ], + "logprob_reward": -0.011101503856480122, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011101503856480122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061342015862464905, + 0.05917005240917206, + 0.04416881129145622, + 0.17273686826229095, + 0.020845375955104828, + 0.029408076778054237, + 0.5479406118392944, + 0.02671082690358162, + 0.02045173943042755, + 0.017225567251443863 + ], + "credit_map": [ + 0.00556880934163928, + 0.006622912362217903, + 0.0046133180148899555, + 0.004504131618887186, + 0.012744125910103321, + 0.01501542329788208, + 0.0051618413999676704, + 0.007111455779522657, + 0.008629736490547657, + 0.006289394106715918, + 0.007116233464330435, + 0.006566229276359081, + 0.007870357483625412, + 0.009947571903467178, + 0.005490818526595831, + 0.0072597116231918335, + 0.0036453979555517435, + 0.0065484121441841125, + 0.007340529467910528, + 0.005060727708041668, + 0.008625115267932415, + 0.0057276575826108456, + 0.004819394554942846, + 0.0024015787057578564, + 0.019194569438695908, + 0.010765809565782547, + 0.03155812248587608, + 0.07163961976766586, + 0.00739056384190917, + 0.006020155269652605, + 0.012396618723869324, + 0.013771405443549156, + 0.003181918291375041, + 0.0024286387488245964, + 0.0021419625263661146, + 0.0031714430078864098, + 0.002377823693677783, + 0.002383096842095256, + 0.003414774779230356, + 0.0017457195790484548, + 0.005212075542658567, + 0.0033757444471120834, + 0.003515730844810605, + 0.003191140480339527, + 0.004405729006975889, + 0.00426152627915144, + 0.0031301514245569706, + 0.0023159808479249477, + 0.007381046656519175, + 0.0144663006067276, + 0.004361520521342754, + 0.00879322737455368, + 0.24593375623226166, + 0.2508804202079773, + 0.007346183992922306, + 0.008778185583651066, + 0.004509415477514267, + 0.002565817441791296, + 0.0038346436340361834, + 0.0029999965336173773, + 0.0056349667720496655, + 0.003071774961426854, + 0.0016704732552170753, + 0.0024237409234046936, + 0.0038821687921881676, + 0.0025155090261250734, + 0.002768680453300476, + 0.003625662997364998, + 0.001978762447834015, + 0.0014684320194646716, + 0.001838715048506856, + 0.0023738087620586157, + 0.003933690953999758, + 0.0018511386588215828, + 0.0026911275926977396, + 0.0018789342138916254, + 0.0016109753632918, + 0.0010747398482635617, + 0.0021308199502527714, + 0.002054140903055668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.011101503856480122 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003621212672442198 + ], + "opposite_logprob_tokens": [ + -5.6286211013793945 + ], + "logprob_reward": -0.003621212672442198, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003621212672442198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042809657752513885, + 0.031648095697164536, + 0.02945881336927414, + 0.040191080421209335, + 0.030055109411478043, + 0.12959274649620056, + 0.5479616522789001, + 0.04671606048941612, + 0.07917554676532745, + 0.022391287609934807 + ], + "credit_map": [ + 0.0040618302300572395, + 0.004595578648149967, + 0.0030946778133511543, + 0.0029441628139466047, + 0.010369529947638512, + 0.009856580756604671, + 0.003817858872935176, + 0.004069434013217688, + 0.003546046791598201, + 0.003873666049912572, + 0.005085685756057501, + 0.003118047956377268, + 0.004529034718871117, + 0.0041344501078128815, + 0.003932445775717497, + 0.00342871667817235, + 0.0032339454628527164, + 0.006576029118150473, + 0.0053694238886237144, + 0.003020769916474819, + 0.004221932962536812, + 0.0026415179017931223, + 0.0025199460797011852, + 0.0018752445466816425, + 0.0032565181609243155, + 0.003946241457015276, + 0.005858045537024736, + 0.011012953706085682, + 0.0036229537799954414, + 0.004593068268150091, + 0.004172381013631821, + 0.003728915937244892, + 0.0033293080050498247, + 0.0031050622928887606, + 0.0020270736422389746, + 0.002599318977445364, + 0.00332953711040318, + 0.0027311560697853565, + 0.00806101318448782, + 0.004872636869549751, + 0.015402792021632195, + 0.02379709668457508, + 0.025423774495720863, + 0.010330411605536938, + 0.023453254252672195, + 0.014585869386792183, + 0.009097679518163204, + 0.007501864340156317, + 0.024565991014242172, + 0.02710546739399433, + 0.010327122174203396, + 0.010139117948710918, + 0.23051683604717255, + 0.22945109009742737, + 0.007401068229228258, + 0.008454912342131138, + 0.008267823606729507, + 0.006092278752475977, + 0.00277724233455956, + 0.002078247256577015, + 0.016015974804759026, + 0.0069841668009757996, + 0.002136621857061982, + 0.0023636999540030956, + 0.012684526853263378, + 0.005286847241222858, + 0.011777772568166256, + 0.014161685481667519, + 0.005975388456135988, + 0.016603171825408936, + 0.006571428384631872, + 0.006114714778959751, + 0.003762677311897278, + 0.0028536475729197264, + 0.0030571070965379477, + 0.0031858610454946756, + 0.002244888572022319, + 0.0017469945596531034, + 0.002692595822736621, + 0.002847513882443309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.003621212672442198 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129860132932663 + ], + "opposite_logprob_tokens": [ + -2.2379860877990723 + ], + "logprob_reward": -0.1129860132932663, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129860132932663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01932954601943493, + 0.02877010405063629, + 0.02031876891851425, + 0.09707953035831451, + 0.017826015129685402, + 0.0802585780620575, + 0.5255329608917236, + 0.06511562317609787, + 0.10195275396108627, + 0.04381614550948143 + ], + "credit_map": [ + 0.002360285259783268, + 0.003188093425706029, + 0.0017824709648266435, + 0.002333597047254443, + 0.0036506368778645992, + 0.0029007752891629934, + 0.0015558867016807199, + 0.0015578018501400948, + 0.0038308328948915005, + 0.003810369409620762, + 0.003570746397599578, + 0.002520561683923006, + 0.003389250487089157, + 0.0033456713426858187, + 0.003411534009501338, + 0.004891142714768648, + 0.0017412776360288262, + 0.0029545207507908344, + 0.0034764951560646296, + 0.002245865063741803, + 0.003986476454883814, + 0.0028683997225016356, + 0.0020460861269384623, + 0.0009996498702093959, + 0.007779713254421949, + 0.01011040061712265, + 0.0160718634724617, + 0.03507334738969803, + 0.006738228723406792, + 0.005782588850706816, + 0.0077333166263997555, + 0.007790072355419397, + 0.0017461379757151008, + 0.002167756436392665, + 0.0016538597410544753, + 0.0027997589204460382, + 0.002011355711147189, + 0.0018963062902912498, + 0.004154817201197147, + 0.0013960242504253983, + 0.010189065709710121, + 0.014511339366436005, + 0.014224149286746979, + 0.00600504782050848, + 0.01421409659087658, + 0.010204339399933815, + 0.0061568752862513065, + 0.004753667861223221, + 0.006841564085334539, + 0.015603393316268921, + 0.00595104368403554, + 0.009301820769906044, + 0.23434056341648102, + 0.23695284128189087, + 0.00810310710221529, + 0.008438650518655777, + 0.00971680786460638, + 0.006941239815205336, + 0.0062200031243264675, + 0.005391149781644344, + 0.02003330923616886, + 0.009520716033875942, + 0.00328750628978014, + 0.004004896618425846, + 0.015293266624212265, + 0.006619591731578112, + 0.015473369508981705, + 0.017997482791543007, + 0.00810770783573389, + 0.021737391129136086, + 0.0087205795571208, + 0.00800336804240942, + 0.01073232851922512, + 0.00481095165014267, + 0.005776860285550356, + 0.004747997969388962, + 0.00355719611980021, + 0.0022595508489757776, + 0.005176120903342962, + 0.006755140144377947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.1129860132932663 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043061573058366776 + ], + "opposite_logprob_tokens": [ + -3.1680614948272705 + ], + "logprob_reward": -0.043061573058366776, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043061573058366776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03015255741775036, + 0.013346409425139427, + 0.005227504763752222, + 0.04302554950118065, + 0.017551712691783905, + 0.05109376832842827, + 0.8108362555503845, + 0.010875502601265907, + 0.008219144307076931, + 0.009671547450125217 + ], + "credit_map": [ + 0.003031840082257986, + 0.003021947806701064, + 0.0014257762813940644, + 0.001892396598123014, + 0.00738562224432826, + 0.00851235631853342, + 0.0020030997693538666, + 0.002879519946873188, + 0.0014841208467260003, + 0.0018158955499529839, + 0.001349119353108108, + 0.0012095290003344417, + 0.002265259390696883, + 0.002555352635681629, + 0.0012586161028593779, + 0.001408517244271934, + 0.0006008085911162198, + 0.0007367933867499232, + 0.0010341813322156668, + 0.0006362109561450779, + 0.0007194491918198764, + 0.0006507346988655627, + 0.0005420808447524905, + 0.000307245587464422, + 0.004969523288309574, + 0.0024458051193505526, + 0.008100477047264576, + 0.01800497993826866, + 0.0017452511237934232, + 0.001386631978675723, + 0.0032474366016685963, + 0.003125449176877737, + 0.003026326186954975, + 0.0028356448747217655, + 0.0018059979192912579, + 0.0018155936850234866, + 0.0026524250861257315, + 0.0025010292883962393, + 0.0014741930644959211, + 0.0014405034016817808, + 0.0073659177869558334, + 0.009765912778675556, + 0.008664251305162907, + 0.0036298329941928387, + 0.00932929664850235, + 0.006066718138754368, + 0.0034278191160410643, + 0.0028440235182642937, + 0.03417230397462845, + 0.036924246698617935, + 0.012793281115591526, + 0.015027116984128952, + 0.3441731035709381, + 0.3422626554965973, + 0.012792102061212063, + 0.012691444717347622, + 0.0016273135552182794, + 0.0011853971518576145, + 0.0009788457537069917, + 0.0007363039767369628, + 0.0033916665706783533, + 0.0017463897820562124, + 0.0005117733962833881, + 0.0006978130550123751, + 0.0012631703866645694, + 0.0005752643919549882, + 0.0008150304784066975, + 0.001336799468845129, + 0.0006777982343919575, + 0.001911344937980175, + 0.000772441562730819, + 0.0008672961848787963, + 0.002568596974015236, + 0.0010123798856511712, + 0.0012730314629152417, + 0.0009715229971334338, + 0.0007745250477455556, + 0.0005193313118070364, + 0.0010250435443595052, + 0.0015271170996129513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.043061573058366776 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins at an extremely low angle, then shakes violently as the earth trembles violently and heavily. A massive crack rapidly tears open in the dry earth, sending rocks and earth scattering. The camera then rapidly pulls back and rises, revealing a vast underground cavity, sealed for millennia—the underground palace of the Qin Shi Huang Mausoleum—opening up with the collapse of the massive rock at the top. Dazzling sunlight pours through the cracks, slicing through the dark air thick with millennia of dust, creating Tyndall-effect beams of light. Beneath these beams, the thousands of Terracotta Army formations are revealed for the first time in the profound darkness. Each terracotta warrior is rendered with the breathtaking detail of next-generation gaming graphics, the realistic texture of their clay, the minute cracks etched by age, and the intricate braiding and metal clasps of their armor plates clearly visible. Their facial expressions vary, from resolute to pensive. Wielding patinated bronze spears, halberds, and powerful crossbows, they stand in endless formations with unwavering discipline, silently towering over the shattered ground. The entire scene's lighting and shadows are rendered with a physically based, real-time quality. The stark contrast between intense light and absolute darkness creates a solemn, mysterious, and oppressive epic atmosphere. All movement—from vibrations to drifting dust—is rendered at a smooth, high frame rate, delivering an ultimate visual immersion.", + "seed": 8703, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_12/index_5470_seed_8703_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.056439651921391484, + "vlm_logprob": -0.056439651921391484, + "vlm_yes_no_margin": 3.499999976158142, + "vlm_token_credit": -0.056439651921391484, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09117790311574936 + ], + "opposite_logprob_tokens": [ + -2.4661779403686523 + ], + "logprob_reward": -0.09117790311574936, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09117790311574936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06056911498308182, + 0.16487625241279602, + 0.11022686213254929, + 0.0752236470580101, + 0.2072715163230896, + 0.2527654469013214, + 0.02720699831843376, + 0.044637225568294525, + 0.03984058275818825, + 0.017382416874170303 + ], + "credit_map": [ + 0.00672452338039875, + 0.00794164277613163, + 0.009112667292356491, + 0.008498011156916618, + 0.0069901542738080025, + 0.005287346430122852, + 0.007764560170471668, + 0.008250207640230656, + 0.010362783446907997, + 0.010802161879837513, + 0.04727619141340256, + 0.017942490056157112, + 0.010393129661679268, + 0.012220609001815319, + 0.045899927616119385, + 0.009978950023651123, + 0.007790443487465382, + 0.013394664041697979, + 0.019238639622926712, + 0.017405983060598373, + 0.014759653247892857, + 0.01011892594397068, + 0.014205667190253735, + 0.013312888331711292, + 0.006213003769516945, + 0.01693296991288662, + 0.007298686075955629, + 0.006399801932275295, + 0.01536883506923914, + 0.010736798867583275, + 0.006336245220154524, + 0.005937310867011547, + 0.021051893010735512, + 0.03529508039355278, + 0.02040352299809456, + 0.02081446722149849, + 0.03313577547669411, + 0.028220754116773605, + 0.027163349092006683, + 0.02118667960166931, + 0.07529834657907486, + 0.10130475461483002, + 0.013723397627472878, + 0.009691557846963406, + 0.01621873304247856, + 0.02687261253595352, + 0.004743116442114115, + 0.004912911914288998, + 0.004817738197743893, + 0.00935368612408638, + 0.0024126600474119186, + 0.0020407098345458508, + 0.0030781591776758432, + 0.002399429911747575, + 0.0017018148209899664, + 0.001402798923663795, + 0.007789408788084984, + 0.003921068273484707, + 0.004771326668560505, + 0.002859415952116251, + 0.003540185047313571, + 0.00645485520362854, + 0.004519224166870117, + 0.010781743563711643, + 0.00355651811696589, + 0.0038086697459220886, + 0.009266546927392483, + 0.003260454861447215, + 0.001974245300516486, + 0.0029216790571808815, + 0.003570175264030695, + 0.011482288129627705, + 0.0018286050762981176, + 0.0023831124417483807, + 0.0022690570913255215, + 0.0020005651749670506, + 0.0019368584034964442, + 0.002238309243693948, + 0.001946805277839303, + 0.002779102884232998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.09117790311574936 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636047638952732 + ], + "opposite_logprob_tokens": [ + -3.651360511779785 + ], + "logprob_reward": -0.02636047638952732, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636047638952732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0855216309428215, + 0.42296838760375977, + 0.050939854234457016, + 0.04759444668889046, + 0.1113676056265831, + 0.20422817766666412, + 0.010682502761483192, + 0.03161057084798813, + 0.01588256284594536, + 0.01920418255031109 + ], + "credit_map": [ + 0.01179235428571701, + 0.00953746773302555, + 0.011774249374866486, + 0.01231444999575615, + 0.00973694957792759, + 0.0065357922576367855, + 0.012491929344832897, + 0.011338437907397747, + 0.02500765025615692, + 0.012457596138119698, + 0.17716358602046967, + 0.054020050913095474, + 0.012124941684305668, + 0.013538097962737083, + 0.10617104917764664, + 0.022485412657260895, + 0.0036696831230074167, + 0.006311262492090464, + 0.008079198189079762, + 0.007590153254568577, + 0.008226212114095688, + 0.005360184703022242, + 0.005847019609063864, + 0.005856138188391924, + 0.004213918931782246, + 0.010442952625453472, + 0.004285652190446854, + 0.00385126075707376, + 0.009873691946268082, + 0.006650608032941818, + 0.0041272398084402084, + 0.004149123560637236, + 0.011246325448155403, + 0.018935665488243103, + 0.011662048287689686, + 0.011382420547306538, + 0.01788286305963993, + 0.014650634489953518, + 0.013928180560469627, + 0.011679474264383316, + 0.06350676715373993, + 0.083478182554245, + 0.010636118240654469, + 0.006163627374917269, + 0.01244489848613739, + 0.020764479413628578, + 0.0035397293977439404, + 0.0036943925078958273, + 0.0015993475681170821, + 0.002848690142855048, + 0.0009899772703647614, + 0.0013318766141310334, + 0.0010084197856485844, + 0.0011025471612811089, + 0.0010922910878434777, + 0.0007093536551110446, + 0.0034379658754915, + 0.003338191192597151, + 0.004525410011410713, + 0.003172062337398529, + 0.002881699940189719, + 0.006257650442421436, + 0.0034603439271450043, + 0.004537245724350214, + 0.0015808490570634604, + 0.0023275003768503666, + 0.0026492977049201727, + 0.0018753764452412724, + 0.001251873793080449, + 0.0015243045054376125, + 0.0024471620563417673, + 0.002226197626441717, + 0.0021141450852155685, + 0.002595379948616028, + 0.0022342982701957226, + 0.0026661199517548084, + 0.0018550250679254532, + 0.00238820374943316, + 0.0026256409473717213, + 0.0027253697626292706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.02636047638952732 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003616461530327797 + ], + "opposite_logprob_tokens": [ + -5.6286163330078125 + ], + "logprob_reward": -0.003616461530327797, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616461530327797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06088041886687279, + 0.23730246722698212, + 0.04392724111676216, + 0.06685402244329453, + 0.07619193941354752, + 0.3328098952770233, + 0.05385007709264755, + 0.07050581276416779, + 0.021158525720238686, + 0.036519575864076614 + ], + "credit_map": [ + 0.00681555550545454, + 0.006939002312719822, + 0.00791317131370306, + 0.008591287769377232, + 0.01165635883808136, + 0.005053769797086716, + 0.0065899379551410675, + 0.007321335840970278, + 0.016337569802999496, + 0.00890197791159153, + 0.09467042982578278, + 0.02986878715455532, + 0.008619598113000393, + 0.00964499730616808, + 0.05778878927230835, + 0.011470326222479343, + 0.0037787933833897114, + 0.004910601768642664, + 0.005708329379558563, + 0.005893814843147993, + 0.006631532683968544, + 0.005053408909589052, + 0.006923625711351633, + 0.005027135834097862, + 0.006862442009150982, + 0.010035707615315914, + 0.0076625836081802845, + 0.008978263475000858, + 0.00927016418427229, + 0.008957907557487488, + 0.008645059540867805, + 0.006441895850002766, + 0.0077832480892539024, + 0.009346548467874527, + 0.009338420815765858, + 0.007303539663553238, + 0.008327662944793701, + 0.010939916595816612, + 0.011392427608370781, + 0.011760176159441471, + 0.1019282415509224, + 0.13566870987415314, + 0.01772044040262699, + 0.01039425190538168, + 0.020516125485301018, + 0.03502568602561951, + 0.0058772568590939045, + 0.005679178051650524, + 0.007954621687531471, + 0.013225546106696129, + 0.005535058677196503, + 0.005256537348031998, + 0.007155755069106817, + 0.00681254593655467, + 0.004420642741024494, + 0.003489372320473194, + 0.01920865848660469, + 0.004679730627685785, + 0.00550645450130105, + 0.003512683790177107, + 0.003551098285242915, + 0.008425856940448284, + 0.0037227675784379244, + 0.02189856395125389, + 0.0021021850407123566, + 0.0031049817334860563, + 0.003455959027633071, + 0.002451443113386631, + 0.0019813573453575373, + 0.0018425099551677704, + 0.003450656309723854, + 0.0027694334276020527, + 0.003792372066527605, + 0.004755800124257803, + 0.0049986280500888824, + 0.00629384396597743, + 0.00312754325568676, + 0.004225782118737698, + 0.004776110406965017, + 0.004549493081867695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.003616461530327797 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428574025630951 + ], + "opposite_logprob_tokens": [ + -2.017857313156128 + ], + "logprob_reward": -0.1428574025630951, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1428574025630951, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06674781441688538, + 0.3692491054534912, + 0.03259968385100365, + 0.018436137586832047, + 0.052941616624593735, + 0.22825336456298828, + 0.01423778384923935, + 0.16325409710407257, + 0.018567243590950966, + 0.03571312874555588 + ], + "credit_map": [ + 0.008452989161014557, + 0.007605970837175846, + 0.008753000758588314, + 0.008057952858507633, + 0.013749336823821068, + 0.0055805835872888565, + 0.00613361457362771, + 0.008414369076490402, + 0.024441003799438477, + 0.010931655764579773, + 0.15689456462860107, + 0.053709231317043304, + 0.009652085602283478, + 0.010152989998459816, + 0.07948524504899979, + 0.02398233860731125, + 0.0018906418699771166, + 0.0035722125321626663, + 0.005183148663491011, + 0.004365337081253529, + 0.004980000667273998, + 0.0026092175394296646, + 0.0062904320657253265, + 0.0037086952943354845, + 0.001847071573138237, + 0.00308586354367435, + 0.0015756289940327406, + 0.0020336320158094168, + 0.0031548747792840004, + 0.00303249922581017, + 0.0019119741627946496, + 0.0017945946892723441, + 0.005797842517495155, + 0.008344636298716068, + 0.005283199716359377, + 0.005122464615851641, + 0.008476569317281246, + 0.008838401176035404, + 0.006374920718371868, + 0.0047035859897732735, + 0.07046740502119064, + 0.08966006338596344, + 0.0123302536085248, + 0.0081853112205863, + 0.014517368748784065, + 0.021644912660121918, + 0.00531139736995101, + 0.0061366683803498745, + 0.0017487385775893927, + 0.002043507993221283, + 0.0018134362762793899, + 0.0014618240529671311, + 0.002048040274530649, + 0.002523241564631462, + 0.0015721454983577132, + 0.0010268505429849029, + 0.04806141555309296, + 0.012359119951725006, + 0.012297281995415688, + 0.0076864929869771, + 0.006014683749526739, + 0.010727064684033394, + 0.009325167164206505, + 0.056782886385917664, + 0.0018967033829540014, + 0.0019728820770978928, + 0.003273558570072055, + 0.0027793715707957745, + 0.0014604192692786455, + 0.0012533431872725487, + 0.002344406908378005, + 0.003586560022085905, + 0.004776194225996733, + 0.00674815708771348, + 0.0044525135308504105, + 0.003979593049734831, + 0.003591828513890505, + 0.0032916248310357332, + 0.0035781969781965017, + 0.005295025184750557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.1428574025630951 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186016008257866 + ], + "opposite_logprob_tokens": [ + -4.018186092376709 + ], + "logprob_reward": -0.018186016008257866, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186016008257866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08372119069099426, + 0.2478269636631012, + 0.09808620810508728, + 0.10074023902416229, + 0.04003508388996124, + 0.26751771569252014, + 0.022004134953022003, + 0.05616459250450134, + 0.050161976367235184, + 0.033741965889930725 + ], + "credit_map": [ + 0.012392815202474594, + 0.009462486952543259, + 0.012552805244922638, + 0.009747008793056011, + 0.011353361420333385, + 0.00665591424331069, + 0.01007972564548254, + 0.01147706713527441, + 0.010563814081251621, + 0.010628389194607735, + 0.08588598668575287, + 0.029588406905531883, + 0.01443567220121622, + 0.012171220034360886, + 0.07063919305801392, + 0.013914248906075954, + 0.00788585189729929, + 0.011366431601345539, + 0.012121541425585747, + 0.010904504917562008, + 0.01789216324687004, + 0.010107604786753654, + 0.015793098136782646, + 0.012014995329082012, + 0.010873486287891865, + 0.01639847457408905, + 0.010913589969277382, + 0.00938986986875534, + 0.023632649332284927, + 0.015614192932844162, + 0.007054254412651062, + 0.006863715592771769, + 0.00411384180188179, + 0.0058571528643369675, + 0.0039435033686459064, + 0.004233213607221842, + 0.005966881290078163, + 0.004409835208207369, + 0.006303367204964161, + 0.005207284819334745, + 0.08303526043891907, + 0.10876792669296265, + 0.013047883287072182, + 0.008200987242162228, + 0.016800643876194954, + 0.028305310755968094, + 0.004503301344811916, + 0.004856375511735678, + 0.0023574456572532654, + 0.0029730359092354774, + 0.0029257694259285927, + 0.004003065638244152, + 0.0030036235693842173, + 0.0026509896852076054, + 0.0022814141120761633, + 0.001808789442293346, + 0.01527540385723114, + 0.004513850901275873, + 0.004790355917066336, + 0.003599112154915929, + 0.00315298093482852, + 0.005063202232122421, + 0.0035920178052037954, + 0.016177665442228317, + 0.005494323559105396, + 0.01015873160213232, + 0.007488820236176252, + 0.005873900838196278, + 0.00485504325479269, + 0.003797388169914484, + 0.005526516120880842, + 0.0069672479294240475, + 0.004823931492865086, + 0.00488633755594492, + 0.004262066446244717, + 0.003814837196841836, + 0.005101157817989588, + 0.0035513783805072308, + 0.003792890813201666, + 0.003509367583319545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.018186016008257866 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins at an extremely low angle, then shakes violently as the earth trembles violently and heavily. A massive crack rapidly tears open in the dry earth, sending rocks and earth scattering. The camera then rapidly pulls back and rises, revealing a vast underground cavity, sealed for millennia—the underground palace of the Qin Shi Huang Mausoleum—opening up with the collapse of the massive rock at the top. Dazzling sunlight pours through the cracks, slicing through the dark air thick with millennia of dust, creating Tyndall-effect beams of light. Beneath these beams, the thousands of Terracotta Army formations are revealed for the first time in the profound darkness. Each terracotta warrior is rendered with the breathtaking detail of next-generation gaming graphics, the realistic texture of their clay, the minute cracks etched by age, and the intricate braiding and metal clasps of their armor plates clearly visible. Their facial expressions vary, from resolute to pensive. Wielding patinated bronze spears, halberds, and powerful crossbows, they stand in endless formations with unwavering discipline, silently towering over the shattered ground. The entire scene's lighting and shadows are rendered with a physically based, real-time quality. The stark contrast between intense light and absolute darkness creates a solemn, mysterious, and oppressive epic atmosphere. All movement—from vibrations to drifting dust—is rendered at a smooth, high frame rate, delivering an ultimate visual immersion.", + "seed": 8704, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_12/index_5470_seed_8704_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07891059499233961, + "vlm_logprob": -0.07891059499233961, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.07891059499233961, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03342720493674278 + ], + "opposite_logprob_tokens": [ + -3.5334272384643555 + ], + "logprob_reward": -0.03342720493674278, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03342720493674278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2162855565547943, + 0.0825553759932518, + 0.1504339575767517, + 0.033925022929906845, + 0.09522703289985657, + 0.23962245881557465, + 0.10683383792638779, + 0.030156435444951057, + 0.025063030421733856, + 0.019897298887372017 + ], + "credit_map": [ + 0.02302371710538864, + 0.0223441980779171, + 0.022469962015748024, + 0.019363949075341225, + 0.0188919547945261, + 0.025194227695465088, + 0.03201008588075638, + 0.05298744887113571, + 0.007438948377966881, + 0.018168235197663307, + 0.005230099428445101, + 0.005380136892199516, + 0.019304055720567703, + 0.008664371445775032, + 0.011898250319063663, + 0.006471279077231884, + 0.019786793738603592, + 0.019906790927052498, + 0.016803238540887833, + 0.022290008142590523, + 0.015939902514219284, + 0.016054242849349976, + 0.024888066574931145, + 0.014764913357794285, + 0.004575527738779783, + 0.005730612203478813, + 0.003195427590981126, + 0.004417770076543093, + 0.00390414334833622, + 0.0037469372618943453, + 0.004880477208644152, + 0.0034741272684186697, + 0.012701399624347687, + 0.016432739794254303, + 0.010510552674531937, + 0.014600595459342003, + 0.014595044776797295, + 0.010920113883912563, + 0.009535710327327251, + 0.005930880084633827, + 0.018365340307354927, + 0.032114818692207336, + 0.01588571071624756, + 0.010830286890268326, + 0.07325677573680878, + 0.06444646418094635, + 0.014398918487131596, + 0.010324142873287201, + 0.011089291423559189, + 0.012989520095288754, + 0.017165392637252808, + 0.009496180340647697, + 0.011589930392801762, + 0.017349157482385635, + 0.00841265358030796, + 0.01874171569943428, + 0.003918352536857128, + 0.002380678430199623, + 0.004105374217033386, + 0.006706588435918093, + 0.0024706488475203514, + 0.00431692274287343, + 0.002332499949261546, + 0.0039253681898117065, + 0.0031612380407750607, + 0.0035081009846180677, + 0.00282652722671628, + 0.002637264784425497, + 0.0026620340067893267, + 0.005496256519109011, + 0.002508251229301095, + 0.0022633576299995184, + 0.0029737253207713366, + 0.0018874226370826364, + 0.0020349298138171434, + 0.00314654060639441, + 0.0027135047130286694, + 0.003285879036411643, + 0.0021447509061545134, + 0.001710544223897159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.03342720493674278 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012562076561152935 + ], + "opposite_logprob_tokens": [ + -4.387562274932861 + ], + "logprob_reward": -0.012562076561152935, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012562076561152935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22418206930160522, + 0.06857908517122269, + 0.15842118859291077, + 0.08004585653543472, + 0.08473381400108337, + 0.09282747656106949, + 0.15769796073436737, + 0.02790057845413685, + 0.06868243217468262, + 0.036929551512002945 + ], + "credit_map": [ + 0.02279576100409031, + 0.025570986792445183, + 0.02403281256556511, + 0.030991198495030403, + 0.019941404461860657, + 0.030259065330028534, + 0.03300346061587334, + 0.03758738562464714, + 0.008205431513488293, + 0.013786857947707176, + 0.005504006985574961, + 0.003737975377589464, + 0.0161843653768301, + 0.009004027582705021, + 0.006974313873797655, + 0.005182104185223579, + 0.01389746367931366, + 0.032009419053792953, + 0.019768264144659042, + 0.01684447005391121, + 0.02251582033932209, + 0.015108892694115639, + 0.021240193396806717, + 0.017036670818924904, + 0.008303900249302387, + 0.014784349128603935, + 0.01132948324084282, + 0.010140855796635151, + 0.009654352441430092, + 0.008744657039642334, + 0.010096614249050617, + 0.0069916401989758015, + 0.012544583529233932, + 0.015056867152452469, + 0.008024507202208042, + 0.010160976089537144, + 0.014169084839522839, + 0.012190365232527256, + 0.008010970428586006, + 0.004576463717967272, + 0.007933464832603931, + 0.01481933519244194, + 0.003848522435873747, + 0.003843289567157626, + 0.029936695471405983, + 0.024253521114587784, + 0.0039297365583479404, + 0.004262911155819893, + 0.0171811543405056, + 0.02689887396991253, + 0.01615368388593197, + 0.019979627802968025, + 0.017745571210980415, + 0.018631117418408394, + 0.013742775656282902, + 0.02736516296863556, + 0.0033482317812740803, + 0.002577185397967696, + 0.0038649574853479862, + 0.004041146486997604, + 0.0019457137677818537, + 0.006606247276067734, + 0.0026217931881546974, + 0.0028953009750694036, + 0.0038204609882086515, + 0.003415684914216399, + 0.02208290807902813, + 0.02363060973584652, + 0.0027735279873013496, + 0.006760952062904835, + 0.0037544213701039553, + 0.002443864708766341, + 0.003263218095526099, + 0.002031258074566722, + 0.002227863296866417, + 0.008847405202686787, + 0.0037756662350147963, + 0.011192657053470612, + 0.0030635863076895475, + 0.0025278981775045395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.012562076561152935 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409882515668869 + ], + "opposite_logprob_tokens": [ + -5.504098892211914 + ], + "logprob_reward": -0.00409882515668869, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409882515668869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1678330898284912, + 0.1630372852087021, + 0.04687375947833061, + 0.02959369495511055, + 0.04168842360377312, + 0.2746122181415558, + 0.1643606573343277, + 0.03795161098241806, + 0.05402890220284462, + 0.02002035267651081 + ], + "credit_map": [ + 0.018405728042125702, + 0.013266703113913536, + 0.01903931424021721, + 0.016198033466935158, + 0.01512975338846445, + 0.020352104678750038, + 0.019440213218331337, + 0.04600124433636665, + 0.006767110899090767, + 0.049363866448402405, + 0.006080209277570248, + 0.007712485268712044, + 0.05029378458857536, + 0.01622866839170456, + 0.018481628969311714, + 0.008109536953270435, + 0.005947473458945751, + 0.005818958859890699, + 0.004709876142442226, + 0.006174104288220406, + 0.006265817675739527, + 0.00747479684650898, + 0.005387508776038885, + 0.005095227621495724, + 0.00334947113879025, + 0.006257523782551289, + 0.003483786014840007, + 0.004147478379309177, + 0.00292464392259717, + 0.00312088243663311, + 0.0036415671929717064, + 0.0026683416217565536, + 0.00671226205304265, + 0.009419852867722511, + 0.003022085642442107, + 0.0044819265604019165, + 0.005765071138739586, + 0.006041824817657471, + 0.003424404188990593, + 0.0028209968004375696, + 0.01993275247514248, + 0.03417752683162689, + 0.019710920751094818, + 0.012438603676855564, + 0.08414681255817413, + 0.07463584840297699, + 0.016634894534945488, + 0.01293485052883625, + 0.016808653250336647, + 0.023068087175488472, + 0.023501703515648842, + 0.019041314721107483, + 0.017919695004820824, + 0.02009980008006096, + 0.014443590305745602, + 0.029477816075086594, + 0.003991409670561552, + 0.0029902386013418436, + 0.005767445079982281, + 0.008981588296592236, + 0.002920349594205618, + 0.005578478332608938, + 0.00293770176358521, + 0.004784399177879095, + 0.0031996958423405886, + 0.0023200977593660355, + 0.01862410269677639, + 0.017248455435037613, + 0.0019556398037821054, + 0.0063087414018809795, + 0.002467332873493433, + 0.001904838252812624, + 0.0018079797737300396, + 0.00154405296780169, + 0.001324919518083334, + 0.0039110854268074036, + 0.0020949195604771376, + 0.006206396501511335, + 0.0016766211483627558, + 0.0014543784782290459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.00409882515668869 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016482949256897 + ], + "opposite_logprob_tokens": [ + -1.701648235321045 + ], + "logprob_reward": -0.2016482949256897, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016482949256897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1105857715010643, + 0.036927152425050735, + 0.0725126639008522, + 0.03318508714437485, + 0.0640367716550827, + 0.1747383177280426, + 0.42762064933776855, + 0.02049974910914898, + 0.021808147430419922, + 0.03808574751019478 + ], + "credit_map": [ + 0.012656964361667633, + 0.008046917617321014, + 0.013488654978573322, + 0.012223634868860245, + 0.009383900091052055, + 0.01067428756505251, + 0.0164641086012125, + 0.027647294104099274, + 0.0020723899360746145, + 0.009318414144217968, + 0.0023702075704932213, + 0.0027426646556705236, + 0.010556098073720932, + 0.002951068338006735, + 0.004614882171154022, + 0.0023014198523014784, + 0.007967607118189335, + 0.007675617467612028, + 0.008622080087661743, + 0.008984604850411415, + 0.00767522631213069, + 0.006773530505597591, + 0.014430946670472622, + 0.010383039712905884, + 0.005223456304520369, + 0.006530337035655975, + 0.0028590539004653692, + 0.0044186050072312355, + 0.0036807728465646505, + 0.0029734999407082796, + 0.0045121535658836365, + 0.002987205982208252, + 0.01219609659165144, + 0.013451272621750832, + 0.004937035962939262, + 0.0068594650365412235, + 0.008757469244301319, + 0.007029856089502573, + 0.005416155327111483, + 0.005389404948800802, + 0.010826597921550274, + 0.016849810257554054, + 0.008686491288244724, + 0.006663221400231123, + 0.06259818375110626, + 0.052694667130708694, + 0.008615122176706791, + 0.00780420983210206, + 0.04418559744954109, + 0.06006396561861038, + 0.05502663925290108, + 0.047770820558071136, + 0.04743674397468567, + 0.05257866904139519, + 0.03965422511100769, + 0.08090394735336304, + 0.002263495931401849, + 0.0018749432638287544, + 0.002874282654374838, + 0.0045958575792610645, + 0.0016193805495277047, + 0.0035902773961424828, + 0.0018069373909384012, + 0.0018745709676295519, + 0.002301365602761507, + 0.0022663914132863283, + 0.0046058110892772675, + 0.002571707358583808, + 0.0016421994660049677, + 0.004358519800007343, + 0.002163211116567254, + 0.0018989397212862968, + 0.005362515803426504, + 0.0027058066334575415, + 0.0026799498591572046, + 0.007157802116125822, + 0.004079852718859911, + 0.009541401639580727, + 0.0034066743683069944, + 0.003151739714667201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.2016482949256897 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extremely low camera angle that shakes violently as the earth trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive crack shown rapidly tearing open in dry earth, scattering rocks and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera rapidly pull back and rise to reveal a vast underground cavity containing the Terracotta Army?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Terracotta warriors depicted with high-detail textures including minute cracks, intricate armor braiding, and varied facial expressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dazzling sunlight pouring through cracks, creating visible Tyndall-effect beams of light in the dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281657338142395 + ], + "opposite_logprob_tokens": [ + -2.0178165435791016 + ], + "logprob_reward": -0.14281657338142395, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14281657338142395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05502447485923767, + 0.03982419893145561, + 0.1953432410955429, + 0.06192190572619438, + 0.037433214485645294, + 0.4562976658344269, + 0.058077458292245865, + 0.03861845284700394, + 0.04437984898686409, + 0.013079472817480564 + ], + "credit_map": [ + 0.00844523310661316, + 0.006377778947353363, + 0.006112836301326752, + 0.005907516926527023, + 0.0059224944561719894, + 0.008680665865540504, + 0.00549993384629488, + 0.008078019134700298, + 0.004347031470388174, + 0.006795960012823343, + 0.003575195325538516, + 0.0026696689892560244, + 0.007142812013626099, + 0.004279552027583122, + 0.006234286352992058, + 0.004779694136232138, + 0.01760558970272541, + 0.04783923923969269, + 0.026123911142349243, + 0.023336242884397507, + 0.021916933357715607, + 0.013908726163208485, + 0.02739322930574417, + 0.01721939817070961, + 0.006717762444168329, + 0.013941732235252857, + 0.0069992817007005215, + 0.008388393558561802, + 0.008945942856371403, + 0.007298081181943417, + 0.005626806057989597, + 0.004003909882158041, + 0.004631116520613432, + 0.0046553295105695724, + 0.00490231066942215, + 0.005831221118569374, + 0.005406245589256287, + 0.004761080723255873, + 0.004514707718044519, + 0.002731208922341466, + 0.03690703585743904, + 0.06185280904173851, + 0.033839937299489975, + 0.021109653636813164, + 0.13622495532035828, + 0.12031389772891998, + 0.027218464761972427, + 0.018830979242920876, + 0.0055761560797691345, + 0.00641073752194643, + 0.008974416181445122, + 0.005925894249230623, + 0.006078827660530806, + 0.00909713376313448, + 0.005695016123354435, + 0.010319282300770283, + 0.004376315511763096, + 0.003214143216609955, + 0.004999256227165461, + 0.009526398964226246, + 0.0028858596924692392, + 0.005412399303168058, + 0.003109786193817854, + 0.0050942921079695225, + 0.004093034192919731, + 0.004728925880044699, + 0.010265873745083809, + 0.009338593110442162, + 0.0030109852086752653, + 0.006903438363224268, + 0.0027136653661727905, + 0.003325337776914239, + 0.001241394318640232, + 0.0013731411891058087, + 0.0023180951829999685, + 0.00213398109190166, + 0.0014703564811497927, + 0.002448670333251357, + 0.0011641157325357199, + 0.0009297187207266688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.14281657338142395 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_15/index_tensor([2632, 2632])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_15/index_tensor([2632, 2632])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..466fc124ce28f496d9dad9bba19ece91ead46b6e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_15/index_tensor([2632, 2632])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful low-angle shot, the camera almost close to the ground, tilts upward to capture the ground's sudden explosion. A massive, heavy-duty tractor, its bright yellow metal hull almost completely covered in thick, wet, dark brown mud, surges upward from the ground with unstoppable force, its massive, spiraling drill bit still spinning at high speed. The drill's sharp metal edge gleams coldly as it tears apart the soil. The intense centrifugal force hurls countless large and small rocks, clumps of earth, and uprooted grass into the air, creating a chaotic and spectacular rain of debris that instantly fills the frame. Under the bright, diffuse light of an overcast sky, each splatter of mud gleams with a wet sheen. The tractor's worn metal surface creates a striking textural contrast with the thick mud, and its massive, deeply treaded tires are also smeared with mud. The entire scene is imbued with a raw sense of industrial power and a dynamic, chaotic aesthetic, creating an awe-inspiring visual spectacle brimming with raw, kinetic energy.", + "seed": 26755, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_15/index_2632_seed_26755_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.054362804139964284, + "vlm_logprob": -0.054362804139964284, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.054362804139964284, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06588536500930786 + ], + "opposite_logprob_tokens": [ + -4.565885543823242 + ], + "logprob_reward": -0.06588536500930786, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.06588536500930786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1509319245815277, + 0.038989465683698654, + 0.0482492670416832, + 0.24477900564670563, + 0.11427510529756546, + 0.05775953829288483, + 0.045189280062913895, + 0.08170412480831146, + 0.06816931068897247, + 0.1499529331922531 + ], + "credit_map": [ + 0.02027936466038227, + 0.028234288096427917, + 0.011723767966032028, + 0.019019147381186485, + 0.02091846987605095, + 0.016645139083266258, + 0.01816427893936634, + 0.015947462990880013, + 0.004186311271041632, + 0.006159835495054722, + 0.004795154090970755, + 0.003947376273572445, + 0.004025044851005077, + 0.004412270616739988, + 0.006222554948180914, + 0.0052409181371331215, + 0.004727150779217482, + 0.007277873810380697, + 0.006042617838829756, + 0.0069348071701824665, + 0.005777811631560326, + 0.004408916458487511, + 0.006217791233211756, + 0.0068622976541519165, + 0.02352507784962654, + 0.03216922655701637, + 0.03384614363312721, + 0.042278725653886795, + 0.0341145284473896, + 0.024286171421408653, + 0.019705817103385925, + 0.03485332056879997, + 0.009099597111344337, + 0.02142511121928692, + 0.018922802060842514, + 0.017362454906105995, + 0.01786474511027336, + 0.00844881497323513, + 0.010429410263895988, + 0.01072216872125864, + 0.007707397919148207, + 0.004614958539605141, + 0.009094571694731712, + 0.007297015283256769, + 0.006414400413632393, + 0.0063752541318535805, + 0.0076723722741007805, + 0.00858356710523367, + 0.005304382648319006, + 0.005134408362209797, + 0.00499021215364337, + 0.004811089951545, + 0.00539844436571002, + 0.007127727381885052, + 0.005266580730676651, + 0.0071564363315701485, + 0.009913433343172073, + 0.010888242162764072, + 0.01119242887943983, + 0.014763550832867622, + 0.006899236235767603, + 0.009295575320720673, + 0.007142290938645601, + 0.011609361506998539, + 0.004412454552948475, + 0.01260717585682869, + 0.020488319918513298, + 0.006044481880962849, + 0.005612940527498722, + 0.003957715816795826, + 0.004740417469292879, + 0.010305801406502724, + 0.008017195388674736, + 0.008808830752968788, + 0.014681241475045681, + 0.03604689985513687, + 0.012241322547197342, + 0.014970727264881134, + 0.012929055839776993, + 0.042257655411958694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.06588536500930786 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296939849853516 + ], + "opposite_logprob_tokens": [ + -2.237969398498535 + ], + "logprob_reward": -0.11296939849853516, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296939849853516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08783130347728729, + 0.06464482098817825, + 0.06725054234266281, + 0.17303068935871124, + 0.16328240931034088, + 0.06901398301124573, + 0.08294851332902908, + 0.10682680457830429, + 0.0645398423075676, + 0.1206311285495758 + ], + "credit_map": [ + 0.012258411385118961, + 0.010548239573836327, + 0.007669266778975725, + 0.009340515360236168, + 0.009814674034714699, + 0.0101020447909832, + 0.010133192874491215, + 0.017964962869882584, + 0.009059970267117023, + 0.008532712236046791, + 0.006293876096606255, + 0.007183440495282412, + 0.007795088924467564, + 0.011040059849619865, + 0.0073346649296581745, + 0.007405013311654329, + 0.0073237549513578415, + 0.009126255288720131, + 0.006779637653380632, + 0.007881002500653267, + 0.009212353266775608, + 0.006317061837762594, + 0.010505441576242447, + 0.010105041787028313, + 0.018112875521183014, + 0.034099917858839035, + 0.022544657811522484, + 0.021533386781811714, + 0.027755705639719963, + 0.01626678928732872, + 0.011704424396157265, + 0.021012919023633003, + 0.015351971611380577, + 0.037235409021377563, + 0.01989351399242878, + 0.028778934851288795, + 0.02545904368162155, + 0.011774629354476929, + 0.012016725726425648, + 0.012772168964147568, + 0.010190379805862904, + 0.007896121591329575, + 0.008245560340583324, + 0.0061150663532316685, + 0.010418251156806946, + 0.0094376839697361, + 0.00912818405777216, + 0.007582737598568201, + 0.012850843369960785, + 0.011026623658835888, + 0.01307559572160244, + 0.006556912790983915, + 0.007824071682989597, + 0.010782678611576557, + 0.009671321138739586, + 0.011160468682646751, + 0.011525513604283333, + 0.013540386222302914, + 0.0170403141528368, + 0.02001240663230419, + 0.008027339354157448, + 0.011560752056539059, + 0.008602243848145008, + 0.016517847776412964, + 0.005947742145508528, + 0.005491928663104773, + 0.011364927515387535, + 0.012678802944719791, + 0.005972416140139103, + 0.004999018739908934, + 0.006508613470941782, + 0.011576391756534576, + 0.009199322201311588, + 0.00928556639701128, + 0.01852831244468689, + 0.026436539366841316, + 0.009390873834490776, + 0.008207634091377258, + 0.012680250220000744, + 0.026902630925178528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.11296939849853516 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017243054462596774 + ], + "opposite_logprob_tokens": [ + -6.3767242431640625 + ], + "logprob_reward": -0.0017243054462596774, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017243054462596774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1885058730840683, + 0.05764235183596611, + 0.0721168965101242, + 0.07819267362356186, + 0.04949561506509781, + 0.055076442658901215, + 0.1341799646615982, + 0.07692430168390274, + 0.15979896485805511, + 0.12806694209575653 + ], + "credit_map": [ + 0.024837268516421318, + 0.03821444511413574, + 0.01096369232982397, + 0.019386982545256615, + 0.035966068506240845, + 0.02075410820543766, + 0.020793922245502472, + 0.017589382827281952, + 0.006084696855396032, + 0.00811100099235773, + 0.007499805185943842, + 0.00795348547399044, + 0.0061393482610583305, + 0.0059361993335187435, + 0.008598926477134228, + 0.007318888790905476, + 0.005991578102111816, + 0.010266986675560474, + 0.007271275855600834, + 0.006820606533437967, + 0.00962877832353115, + 0.005772577133029699, + 0.01283571682870388, + 0.013529371470212936, + 0.009227357804775238, + 0.018414530903100967, + 0.009352389723062515, + 0.006027210969477892, + 0.013085582293570042, + 0.007190133910626173, + 0.005568488966673613, + 0.009326974861323833, + 0.004672531504184008, + 0.011414945125579834, + 0.00542153837159276, + 0.00750912306830287, + 0.009188107214868069, + 0.0033291711006313562, + 0.003991485573351383, + 0.003968709614127874, + 0.004901803098618984, + 0.006090160459280014, + 0.006090130656957626, + 0.007950401864945889, + 0.0067507061176002026, + 0.009370311163365841, + 0.006827996578067541, + 0.007094930857419968, + 0.013688323087990284, + 0.013756084255874157, + 0.01716567762196064, + 0.024301152676343918, + 0.014044594950973988, + 0.018694432452321053, + 0.01495010033249855, + 0.017579590901732445, + 0.008942502550780773, + 0.009395052678883076, + 0.009282193146646023, + 0.01373204868286848, + 0.006995169445872307, + 0.00790487416088581, + 0.008953734301030636, + 0.011718729510903358, + 0.011096535250544548, + 0.04003137722611427, + 0.051781680434942245, + 0.009519100189208984, + 0.015389380045235157, + 0.007221026811748743, + 0.007286354899406433, + 0.01747351512312889, + 0.00913303904235363, + 0.015786876901984215, + 0.014441534876823425, + 0.029565241187810898, + 0.011280057020485401, + 0.008878608234226704, + 0.011440358124673367, + 0.027541233226656914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.0017243054462596774 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904164284467697 + ], + "opposite_logprob_tokens": [ + -2.4640417098999023 + ], + "logprob_reward": -0.08904164284467697, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904164284467697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1450778692960739, + 0.07426031678915024, + 0.05606453865766525, + 0.047802042216062546, + 0.05229252576828003, + 0.05969235301017761, + 0.11633849143981934, + 0.09185265749692917, + 0.17687685787677765, + 0.17974238097667694 + ], + "credit_map": [ + 0.02177097089588642, + 0.028644347563385963, + 0.010020386427640915, + 0.013318351469933987, + 0.024675214663147926, + 0.01603413186967373, + 0.018745098263025284, + 0.011869358830153942, + 0.009795378893613815, + 0.01081161294132471, + 0.00649756146594882, + 0.007191340439021587, + 0.006700418423861265, + 0.01528190914541483, + 0.010266540572047234, + 0.007715546526014805, + 0.0049680136144161224, + 0.007675320841372013, + 0.005401714704930782, + 0.007375525310635567, + 0.008451434783637524, + 0.004782703705132008, + 0.008693752810359001, + 0.008716071024537086, + 0.004311793949455023, + 0.010543200187385082, + 0.004121451172977686, + 0.004510275553911924, + 0.010393171571195126, + 0.003647396806627512, + 0.0025987927801907063, + 0.00767595786601305, + 0.0037794269155710936, + 0.007451613433659077, + 0.010284987278282642, + 0.008520837873220444, + 0.009143280796706676, + 0.0034925583750009537, + 0.004234008491039276, + 0.005385811440646648, + 0.005258029792457819, + 0.00723125459626317, + 0.005909144878387451, + 0.006023896858096123, + 0.01072964072227478, + 0.00900545809417963, + 0.006821459624916315, + 0.00871346890926361, + 0.01178557239472866, + 0.011480401270091534, + 0.014566533267498016, + 0.021302664652466774, + 0.011481549590826035, + 0.01754787378013134, + 0.014602536335587502, + 0.013571356423199177, + 0.011055194772779942, + 0.009534604847431183, + 0.012964022345840931, + 0.01204029843211174, + 0.00944376178085804, + 0.01265687681734562, + 0.010038434527814388, + 0.01411945279687643, + 0.012121935375034809, + 0.04799216240644455, + 0.05734121799468994, + 0.008797851391136646, + 0.017292937263846397, + 0.007709273602813482, + 0.007965803146362305, + 0.017655659466981888, + 0.013707081787288189, + 0.02779179811477661, + 0.023033475503325462, + 0.032571788877248764, + 0.01961679570376873, + 0.012581435032188892, + 0.014774025417864323, + 0.03566597402095795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.08904164284467697 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002193308901041746 + ], + "opposite_logprob_tokens": [ + -6.127193450927734 + ], + "logprob_reward": -0.002193308901041746, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002193308901041746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17780843377113342, + 0.0575622133910656, + 0.08280045539140701, + 0.1114962249994278, + 0.10570503771305084, + 0.08106072247028351, + 0.07894989848136902, + 0.051279883831739426, + 0.13313086330890656, + 0.1202063038945198 + ], + "credit_map": [ + 0.022538235411047935, + 0.020657960325479507, + 0.02035658061504364, + 0.017516713589429855, + 0.028079921379685402, + 0.021709147840738297, + 0.019989075139164925, + 0.026960795745253563, + 0.006398837082087994, + 0.009649666957557201, + 0.0066232830286026, + 0.007406918331980705, + 0.007920353673398495, + 0.006202263757586479, + 0.006084254942834377, + 0.007276636548340321, + 0.006206709425896406, + 0.011133295483887196, + 0.007436814252287149, + 0.011357895098626614, + 0.013146508485078812, + 0.008117222227156162, + 0.010316436178982258, + 0.015085576102137566, + 0.01149915810674429, + 0.016316769644618034, + 0.015110772103071213, + 0.019946448504924774, + 0.016737855970859528, + 0.009630495682358742, + 0.011129247955977917, + 0.011125476099550724, + 0.009000317193567753, + 0.016123607754707336, + 0.019817285239696503, + 0.013751420192420483, + 0.020399656146764755, + 0.006927054841071367, + 0.00951910950243473, + 0.010166595689952374, + 0.007164215203374624, + 0.00948318000882864, + 0.009096761234104633, + 0.007131943479180336, + 0.012599042616784573, + 0.01337845902889967, + 0.008648551069200039, + 0.013558567501604557, + 0.00861374568194151, + 0.007956957444548607, + 0.009359106421470642, + 0.013766116462647915, + 0.008115382865071297, + 0.011630860157310963, + 0.008454715833067894, + 0.011053012683987617, + 0.005057309288531542, + 0.0051575773395597935, + 0.006730666849762201, + 0.006832560990005732, + 0.0047897351905703545, + 0.007549003232270479, + 0.008036475628614426, + 0.007126556243747473, + 0.009288350120186806, + 0.026398980990052223, + 0.040261510759592056, + 0.008493330329656601, + 0.01177994441241026, + 0.008198380470275879, + 0.007844319567084312, + 0.02086603455245495, + 0.009190109558403492, + 0.01981651782989502, + 0.01696932502090931, + 0.01739010214805603, + 0.015472502447664738, + 0.010263067670166492, + 0.010715516284108162, + 0.020389167591929436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.002193308901041746 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful low-angle shot, the camera almost close to the ground, tilts upward to capture the ground's sudden explosion. A massive, heavy-duty tractor, its bright yellow metal hull almost completely covered in thick, wet, dark brown mud, surges upward from the ground with unstoppable force, its massive, spiraling drill bit still spinning at high speed. The drill's sharp metal edge gleams coldly as it tears apart the soil. The intense centrifugal force hurls countless large and small rocks, clumps of earth, and uprooted grass into the air, creating a chaotic and spectacular rain of debris that instantly fills the frame. Under the bright, diffuse light of an overcast sky, each splatter of mud gleams with a wet sheen. The tractor's worn metal surface creates a striking textural contrast with the thick mud, and its massive, deeply treaded tires are also smeared with mud. The entire scene is imbued with a raw sense of industrial power and a dynamic, chaotic aesthetic, creating an awe-inspiring visual spectacle brimming with raw, kinetic energy.", + "seed": 26756, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_15/index_2632_seed_26756_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05893453301396221, + "vlm_logprob": -0.05893453301396221, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.05893453301396221, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09659986197948456 + ], + "opposite_logprob_tokens": [ + -3.096599817276001 + ], + "logprob_reward": -0.09659986197948456, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.09659986197948456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18890666961669922, + 0.1454780101776123, + 0.07097958773374557, + 0.1742183417081833, + 0.03288714215159416, + 0.13166743516921997, + 0.061292149126529694, + 0.03530075028538704, + 0.0658457949757576, + 0.09342413395643234 + ], + "credit_map": [ + 0.009346491657197475, + 0.014149941504001617, + 0.014970543794333935, + 0.01548723690211773, + 0.05145231634378433, + 0.0494648814201355, + 0.019525177776813507, + 0.014510086737573147, + 0.006663351319730282, + 0.004508118610829115, + 0.005301846656948328, + 0.0033149542286992073, + 0.057963598519563675, + 0.05736370012164116, + 0.0053474875167012215, + 0.005014946684241295, + 0.007467141840606928, + 0.005336402915418148, + 0.009136714972555637, + 0.009966248646378517, + 0.013060925528407097, + 0.009054295718669891, + 0.011283771134912968, + 0.005674084182828665, + 0.02027805708348751, + 0.018601063638925552, + 0.02423776686191559, + 0.016024673357605934, + 0.026836104691028595, + 0.020982690155506134, + 0.02831256575882435, + 0.018945425748825073, + 0.007218608167022467, + 0.0028498792089521885, + 0.0071528577245771885, + 0.004172072745859623, + 0.0024765413254499435, + 0.003014160320162773, + 0.00276452605612576, + 0.0032384926453232765, + 0.015330931171774864, + 0.011437608860433102, + 0.01951242983341217, + 0.011940092779695988, + 0.019574569538235664, + 0.018621567636728287, + 0.020810820162296295, + 0.014439413323998451, + 0.009563381783664227, + 0.00822407566010952, + 0.00790218822658062, + 0.01078137569129467, + 0.005081390496343374, + 0.003568990621715784, + 0.008084118366241455, + 0.00808662734925747, + 0.004474093671888113, + 0.003489315276965499, + 0.007078363560140133, + 0.008810029365122318, + 0.002805003896355629, + 0.0027126024942845106, + 0.0034110231790691614, + 0.0025203183759003878, + 0.005694885738193989, + 0.005473779980093241, + 0.008579272776842117, + 0.007971680723130703, + 0.00624617375433445, + 0.012716740369796753, + 0.006170887965708971, + 0.012992369942367077, + 0.006189167499542236, + 0.004951858893036842, + 0.005441596265882254, + 0.021797586232423782, + 0.004874394740909338, + 0.008704651147127151, + 0.010857025161385536, + 0.03060784935951233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.09659986197948456 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17995338141918182 + ], + "opposite_logprob_tokens": [ + -1.8049533367156982 + ], + "logprob_reward": -0.17995338141918182, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17995338141918182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19614604115486145, + 0.3827578127384186, + 0.09459932893514633, + 0.02222779579460621, + 0.030124222859740257, + 0.01793237216770649, + 0.04016602784395218, + 0.04631916433572769, + 0.031374573707580566, + 0.138352632522583 + ], + "credit_map": [ + 0.021136920899152756, + 0.016069315373897552, + 0.0324753038585186, + 0.016045356169342995, + 0.040278688073158264, + 0.04492900148034096, + 0.011754120700061321, + 0.013457325287163258, + 0.012161893770098686, + 0.008852051571011543, + 0.012886973097920418, + 0.01164915505796671, + 0.15996284782886505, + 0.15906089544296265, + 0.010939216241240501, + 0.007244768086820841, + 0.011239590123295784, + 0.009725040756165981, + 0.013257726095616817, + 0.014826763421297073, + 0.012642775662243366, + 0.009748774580657482, + 0.015352082438766956, + 0.007806578651070595, + 0.0023013330064713955, + 0.0024734281469136477, + 0.002033228985965252, + 0.002645289758220315, + 0.004485998768359423, + 0.003573862137272954, + 0.003241325728595257, + 0.0014733283314853907, + 0.005470520816743374, + 0.002373947063460946, + 0.004557463806122541, + 0.004276494961231947, + 0.003380812471732497, + 0.003006560727953911, + 0.0037059313617646694, + 0.0033524928148835897, + 0.0018984347116202116, + 0.002322266809642315, + 0.002738972892984748, + 0.0017552593490108848, + 0.0026055185589939356, + 0.0025280509144067764, + 0.001985451439395547, + 0.002098416443914175, + 0.005953649524599314, + 0.0060825166292488575, + 0.005000463221222162, + 0.007030540611594915, + 0.0030293494928628206, + 0.003292626468464732, + 0.005212694406509399, + 0.004564186558127403, + 0.007849016226828098, + 0.005180706270039082, + 0.0083763487637043, + 0.010285885073244572, + 0.004376669880002737, + 0.0036687327083200216, + 0.0037276383955031633, + 0.002854167716577649, + 0.0032363797072321177, + 0.0025403453037142754, + 0.0043113031424582005, + 0.0037916419096291065, + 0.002566243289038539, + 0.003819006960839033, + 0.004153282847255468, + 0.006956374738365412, + 0.009650737047195435, + 0.006250889506191015, + 0.006104602012783289, + 0.03639417141675949, + 0.005441491492092609, + 0.012820368632674217, + 0.012384659610688686, + 0.04930572956800461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.17995338141918182 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022103183437138796 + ], + "opposite_logprob_tokens": [ + -6.1272101402282715 + ], + "logprob_reward": -0.0022103183437138796, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022103183437138796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07210516184568405, + 0.13651394844055176, + 0.10057660937309265, + 0.046909891068935394, + 0.043151021003723145, + 0.059704992920160294, + 0.03547490015625954, + 0.10874094814062119, + 0.031300343573093414, + 0.36552226543426514 + ], + "credit_map": [ + 0.006907160393893719, + 0.008773046545684338, + 0.011781170032918453, + 0.009579497389495373, + 0.012011608108878136, + 0.014052332378923893, + 0.005355039611458778, + 0.003645308082923293, + 0.004663138650357723, + 0.0031638413202017546, + 0.005782485473901033, + 0.003429166506975889, + 0.05634134262800217, + 0.05605986714363098, + 0.004004485905170441, + 0.003069615224376321, + 0.008869809098541737, + 0.006845659110695124, + 0.011050374247133732, + 0.013735313899815083, + 0.01974675804376602, + 0.013934297487139702, + 0.018669219687581062, + 0.007725181058049202, + 0.004850273951888084, + 0.005072132218629122, + 0.007944461889564991, + 0.0043291691690683365, + 0.006458770018070936, + 0.0051574441604316235, + 0.007212689146399498, + 0.005884954705834389, + 0.006473001092672348, + 0.003031510394066572, + 0.009678552858531475, + 0.007884964346885681, + 0.003714664140716195, + 0.003830469213426113, + 0.00349820195697248, + 0.005039658863097429, + 0.008193464949727058, + 0.004841344431042671, + 0.008189419284462929, + 0.005674588959664106, + 0.006488222163170576, + 0.006903476547449827, + 0.00904160737991333, + 0.010372869670391083, + 0.005110148806124926, + 0.005758520215749741, + 0.004855165258049965, + 0.006205230485647917, + 0.0022367099300026894, + 0.0027356971986591816, + 0.004548949655145407, + 0.004024477209895849, + 0.014240162447094917, + 0.010321076028048992, + 0.021948523819446564, + 0.032019536942243576, + 0.008600478060543537, + 0.006736368406563997, + 0.009304995648562908, + 0.00556980911642313, + 0.002795264357700944, + 0.003991738893091679, + 0.0029472035821527243, + 0.0056554172188043594, + 0.0030111772939562798, + 0.0066509670577943325, + 0.0033140091691166162, + 0.002934566233307123, + 0.027173791080713272, + 0.014870070852339268, + 0.013304558582603931, + 0.09326063841581345, + 0.010716349817812443, + 0.03163492679595947, + 0.03464265912771225, + 0.139919251203537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.0022103183437138796 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197650365531445 + ], + "opposite_logprob_tokens": [ + -4.264197826385498 + ], + "logprob_reward": -0.014197650365531445, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197650365531445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0568668358027935, + 0.05730956420302391, + 0.24522839486598969, + 0.03556442633271217, + 0.027376484125852585, + 0.04018097370862961, + 0.05021097883582115, + 0.09484435617923737, + 0.060834404081106186, + 0.33158358931541443 + ], + "credit_map": [ + 0.004242869559675455, + 0.003284998005256057, + 0.007870590314269066, + 0.005206545349210501, + 0.013794198632240295, + 0.013914289884269238, + 0.0033544590696692467, + 0.005198884289711714, + 0.004551608115434647, + 0.0025836448185145855, + 0.007239135447889566, + 0.002688593463972211, + 0.017695439979434013, + 0.017626803368330002, + 0.0026735567953437567, + 0.002250781748443842, + 0.0249813012778759, + 0.016917210072278976, + 0.028896847739815712, + 0.043480005115270615, + 0.03987656906247139, + 0.028415028005838394, + 0.044787514954805374, + 0.01787390746176243, + 0.0034296817611902952, + 0.005164427682757378, + 0.003864114871248603, + 0.003238065866753459, + 0.005475158337503672, + 0.004910285584628582, + 0.006439664866775274, + 0.0030430268961936235, + 0.0035048264544457197, + 0.001883612829260528, + 0.007507065776735544, + 0.004268796183168888, + 0.0020048324950039387, + 0.002532919170334935, + 0.002461874857544899, + 0.0032125567086040974, + 0.004689461085945368, + 0.004499930888414383, + 0.00475708395242691, + 0.003926904872059822, + 0.006862848997116089, + 0.0059255678206682205, + 0.0038087978027760983, + 0.00571037782356143, + 0.006816624663770199, + 0.008604457601904869, + 0.0070107742212712765, + 0.00862001720815897, + 0.0035174679942429066, + 0.003998660948127508, + 0.007145763374865055, + 0.004497214686125517, + 0.013413160108029842, + 0.008685991168022156, + 0.018357643857598305, + 0.026853403076529503, + 0.006606813054531813, + 0.006189168896526098, + 0.00888745579868555, + 0.005850724410265684, + 0.005622774828225374, + 0.007530657108873129, + 0.006291147321462631, + 0.0076213013380765915, + 0.006305764429271221, + 0.014190003275871277, + 0.006653072312474251, + 0.0066196792759001255, + 0.020808571949601173, + 0.014627058058977127, + 0.015123427845537663, + 0.09528449922800064, + 0.012651865370571613, + 0.02813427709043026, + 0.02805744856595993, + 0.11689642071723938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.014197650365531445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017114529618993402 + ], + "opposite_logprob_tokens": [ + -6.376711368560791 + ], + "logprob_reward": -0.0017114529618993402, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017114529618993402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29622554779052734, + 0.0916302278637886, + 0.3569285571575165, + 0.03442623093724251, + 0.02140701562166214, + 0.06586546450853348, + 0.04153283312916756, + 0.022725269198417664, + 0.026158494874835014, + 0.043100353330373764 + ], + "credit_map": [ + 0.0064023882150650024, + 0.0112453056499362, + 0.00845684576779604, + 0.01023097150027752, + 0.11931243538856506, + 0.12290344387292862, + 0.010856105014681816, + 0.006818023510277271, + 0.006653687916696072, + 0.0038939770311117172, + 0.007448606193065643, + 0.003976224455982447, + 0.03033412992954254, + 0.030110904946923256, + 0.006000221241265535, + 0.00321247149258852, + 0.03389976918697357, + 0.028900105506181717, + 0.05181345343589783, + 0.03730469569563866, + 0.07171270996332169, + 0.04859654977917671, + 0.0521945096552372, + 0.03250674530863762, + 0.0044300672598183155, + 0.004829397425055504, + 0.0052605923265218735, + 0.003346688114106655, + 0.004662190563976765, + 0.004357608500868082, + 0.004771926440298557, + 0.002767759608104825, + 0.0027883045841008425, + 0.0022142804227769375, + 0.0036097937263548374, + 0.0025613976176828146, + 0.002900927560403943, + 0.0022954042069613934, + 0.002674683229997754, + 0.0023622221779078245, + 0.009220658801496029, + 0.005155674181878567, + 0.010968041606247425, + 0.007746018469333649, + 0.00642217556014657, + 0.00694633461534977, + 0.01198230218142271, + 0.007424253039062023, + 0.0052697681821882725, + 0.006586837582290173, + 0.005316280759871006, + 0.005649830214679241, + 0.003508238587528467, + 0.003112837905064225, + 0.006121678743511438, + 0.005967355333268642, + 0.002986992010846734, + 0.002169041195884347, + 0.00461358530446887, + 0.006079388782382011, + 0.0017356978496536613, + 0.0013558620121330023, + 0.0022189407609403133, + 0.0015657602343708277, + 0.0020460106898099184, + 0.0031646988354623318, + 0.0032369771506637335, + 0.004116720054298639, + 0.0027462001889944077, + 0.005899185314774513, + 0.0019795719999819994, + 0.002969128545373678, + 0.002756429836153984, + 0.0031929127871990204, + 0.0024602292105555534, + 0.009761098772287369, + 0.0035857188049703836, + 0.006587440147995949, + 0.004500933922827244, + 0.010255583561956882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0017114529618993402 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful low-angle shot, the camera almost close to the ground, tilts upward to capture the ground's sudden explosion. A massive, heavy-duty tractor, its bright yellow metal hull almost completely covered in thick, wet, dark brown mud, surges upward from the ground with unstoppable force, its massive, spiraling drill bit still spinning at high speed. The drill's sharp metal edge gleams coldly as it tears apart the soil. The intense centrifugal force hurls countless large and small rocks, clumps of earth, and uprooted grass into the air, creating a chaotic and spectacular rain of debris that instantly fills the frame. Under the bright, diffuse light of an overcast sky, each splatter of mud gleams with a wet sheen. The tractor's worn metal surface creates a striking textural contrast with the thick mud, and its massive, deeply treaded tires are also smeared with mud. The entire scene is imbued with a raw sense of industrial power and a dynamic, chaotic aesthetic, creating an awe-inspiring visual spectacle brimming with raw, kinetic energy.", + "seed": 26757, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_15/index_2632_seed_26757_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05118473223410547, + "vlm_logprob": -0.05118473223410547, + "vlm_yes_no_margin": 4.249999976158142, + "vlm_token_credit": -0.05118473223410547, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07784559577703476 + ], + "opposite_logprob_tokens": [ + -3.827845573425293 + ], + "logprob_reward": -0.07784559577703476, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.07784559577703476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21905215084552765, + 0.168753519654274, + 0.06438401341438293, + 0.04807943105697632, + 0.07088208198547363, + 0.08712954819202423, + 0.08477289229631424, + 0.06496727466583252, + 0.12694872915744781, + 0.06503033638000488 + ], + "credit_map": [ + 0.03385258838534355, + 0.0192378181964159, + 0.035584669560194016, + 0.03663419559597969, + 0.017615970224142075, + 0.024486379697918892, + 0.03361900523304939, + 0.01802152767777443, + 0.029066286981105804, + 0.017751332372426987, + 0.02339954301714897, + 0.0178530290722847, + 0.020865093916654587, + 0.01255748514086008, + 0.02460293471813202, + 0.02265782468020916, + 0.010727555491030216, + 0.007488828618079424, + 0.007492779288440943, + 0.006702486425638199, + 0.009497989900410175, + 0.010485910810530186, + 0.005130020435899496, + 0.006858444772660732, + 0.004983843769878149, + 0.008799450471997261, + 0.005481159780174494, + 0.004466445650905371, + 0.008296147920191288, + 0.007353383582085371, + 0.0050587947480380535, + 0.0036402044352144003, + 0.010428262874484062, + 0.008550741709768772, + 0.00855433102697134, + 0.008486163802444935, + 0.011759079992771149, + 0.007063951808959246, + 0.008911136537790298, + 0.007128414232283831, + 0.009755159728229046, + 0.007574964314699173, + 0.014504889957606792, + 0.017116613686084747, + 0.00932384468615055, + 0.009690786711871624, + 0.013282209634780884, + 0.005881074815988541, + 0.007102656178176403, + 0.005193409044295549, + 0.011999925598502159, + 0.007568336091935635, + 0.0073367152363061905, + 0.012523669749498367, + 0.021158022806048393, + 0.011890166439116001, + 0.0055763013660907745, + 0.004848231095820665, + 0.011637561954557896, + 0.011023755185306072, + 0.006516750901937485, + 0.007030313368886709, + 0.010306796990334988, + 0.008027561008930206, + 0.024488277733325958, + 0.005576570052653551, + 0.006647307425737381, + 0.007968776859343052, + 0.006073332391679287, + 0.04877840727567673, + 0.01021472830325365, + 0.01720132678747177, + 0.005181111861020327, + 0.005135605111718178, + 0.0054466999135911465, + 0.004668815061450005, + 0.006961564067751169, + 0.017424065619707108, + 0.008822573348879814, + 0.011389896273612976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.07784559577703476 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428426206111908 + ], + "opposite_logprob_tokens": [ + -2.0178425312042236 + ], + "logprob_reward": -0.1428426206111908, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1428426206111908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17298966646194458, + 0.17563888430595398, + 0.09551075845956802, + 0.1848524957895279, + 0.07876762747764587, + 0.03710366040468216, + 0.08120937645435333, + 0.03795409947633743, + 0.09400907158851624, + 0.04196435585618019 + ], + "credit_map": [ + 0.01642870344221592, + 0.024882346391677856, + 0.021634407341480255, + 0.03093290515244007, + 0.017982710152864456, + 0.021481122821569443, + 0.024836082011461258, + 0.014811400324106216, + 0.02759549207985401, + 0.01983925700187683, + 0.029081203043460846, + 0.020100882276892662, + 0.023202823475003242, + 0.013308010064065456, + 0.016491765156388283, + 0.026019463315606117, + 0.015847276896238327, + 0.012704501859843731, + 0.008504553698003292, + 0.008568133227527142, + 0.013765184208750725, + 0.01640707068145275, + 0.007540531922131777, + 0.012173512950539589, + 0.021740878000855446, + 0.03927287459373474, + 0.026357701048254967, + 0.01580873504281044, + 0.021207403391599655, + 0.024513281881809235, + 0.021623095497488976, + 0.01432853378355503, + 0.015172570943832397, + 0.00895167887210846, + 0.015769001096487045, + 0.010117585770785809, + 0.007657935842871666, + 0.009500984102487564, + 0.006920584011822939, + 0.004677292425185442, + 0.005498425103724003, + 0.003545396262779832, + 0.005515937693417072, + 0.006477812305092812, + 0.003230120288208127, + 0.0047284867614507675, + 0.005183998495340347, + 0.002923480235040188, + 0.006632508244365454, + 0.005654647946357727, + 0.010324742645025253, + 0.00811825878918171, + 0.008609965443611145, + 0.013424579054117203, + 0.017064115032553673, + 0.011380568146705627, + 0.004167450591921806, + 0.003985370043665171, + 0.005130439065396786, + 0.007466659415513277, + 0.0027955842670053244, + 0.0048245093785226345, + 0.004765982739627361, + 0.004818106535822153, + 0.011712035164237022, + 0.0029705397319048643, + 0.0041734897531569, + 0.004959410056471825, + 0.0040092081762850285, + 0.03235417231917381, + 0.009296332485973835, + 0.024533892050385475, + 0.004819605033844709, + 0.0043221116065979, + 0.004358302801847458, + 0.003963092342019081, + 0.003920095972716808, + 0.010893955826759338, + 0.004350786097347736, + 0.005336410831660032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.1428426206111908 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001347110839560628 + ], + "opposite_logprob_tokens": [ + -6.626347064971924 + ], + "logprob_reward": -0.001347110839560628, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001347110839560628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.307570219039917, + 0.08548364043235779, + 0.07256975769996643, + 0.06042347103357315, + 0.07268524169921875, + 0.031092088669538498, + 0.13063275814056396, + 0.030140556395053864, + 0.1457248479127884, + 0.06367741525173187 + ], + "credit_map": [ + 0.03827444463968277, + 0.030901266261935234, + 0.04737169295549393, + 0.04698962718248367, + 0.023544277995824814, + 0.03246159479022026, + 0.05625597760081291, + 0.031771350651979446, + 0.013446597382426262, + 0.009567865170538425, + 0.011365817859768867, + 0.007841169834136963, + 0.01275704801082611, + 0.00860529113560915, + 0.013000445440411568, + 0.008899417705833912, + 0.011963547207415104, + 0.00968210119754076, + 0.006904922891408205, + 0.006972472183406353, + 0.010899352841079235, + 0.009312541224062443, + 0.006897393148392439, + 0.009937436319887638, + 0.006701378151774406, + 0.011782224290072918, + 0.008338788524270058, + 0.005594644695520401, + 0.0077772680670022964, + 0.008461660705506802, + 0.00664869649335742, + 0.005118817090988159, + 0.006417233031243086, + 0.009855902753770351, + 0.014147560112178326, + 0.009677036665380001, + 0.009800051338970661, + 0.007176908664405346, + 0.009219542145729065, + 0.006391013506799936, + 0.0033049408812075853, + 0.0024712267331779003, + 0.004685659427195787, + 0.003918008413165808, + 0.004058199469000101, + 0.0036934404633939266, + 0.006083385553210974, + 0.0028772312216460705, + 0.010409332811832428, + 0.0077771032229065895, + 0.011743327602744102, + 0.01161011029034853, + 0.011051688343286514, + 0.01744747720658779, + 0.03854364901781082, + 0.022050073370337486, + 0.0039011796470731497, + 0.002742068376392126, + 0.004319291561841965, + 0.004110119771212339, + 0.002780033741146326, + 0.005511508788913488, + 0.0033222485799342394, + 0.0034541075583547354, + 0.022673700004816055, + 0.005730722099542618, + 0.006781040690839291, + 0.008764401078224182, + 0.0066938139498233795, + 0.052443161606788635, + 0.012559259310364723, + 0.03007875196635723, + 0.005888277664780617, + 0.005818364676088095, + 0.0063206166960299015, + 0.007214662153273821, + 0.005889525171369314, + 0.017998596653342247, + 0.006862395443022251, + 0.007684972137212753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.001347110839560628 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799362644553185 + ], + "opposite_logprob_tokens": [ + -3.529799461364746 + ], + "logprob_reward": -0.029799362644553185, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799362644553185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22893530130386353, + 0.17478430271148682, + 0.0863647311925888, + 0.03860054537653923, + 0.13196055591106415, + 0.07203163951635361, + 0.09103857725858688, + 0.035961490124464035, + 0.0978676825761795, + 0.04245518520474434 + ], + "credit_map": [ + 0.022719118744134903, + 0.035378433763980865, + 0.03034563548862934, + 0.033821310847997665, + 0.01975341886281967, + 0.029941527172923088, + 0.03044755756855011, + 0.026528313755989075, + 0.026980379596352577, + 0.021148404106497765, + 0.023497987538576126, + 0.019647669047117233, + 0.02611365355551243, + 0.014386329799890518, + 0.02046877145767212, + 0.022541100159287453, + 0.013294337317347527, + 0.010460629127919674, + 0.008289463818073273, + 0.00944951456040144, + 0.01269823033362627, + 0.012762838043272495, + 0.008808077313005924, + 0.010601642541587353, + 0.004995081573724747, + 0.007013896480202675, + 0.003480032319203019, + 0.0035315549466758966, + 0.005784987006336451, + 0.0057543679140508175, + 0.003999984823167324, + 0.004040641710162163, + 0.018764778971672058, + 0.014217003248631954, + 0.024419022724032402, + 0.019622189924120903, + 0.017443934455513954, + 0.012343261390924454, + 0.016947109252214432, + 0.008203258737921715, + 0.0051605235785245895, + 0.006793359760195017, + 0.010285667143762112, + 0.013253257609903812, + 0.009872575290501118, + 0.008060206659138203, + 0.011041403748095036, + 0.007564647123217583, + 0.00827798992395401, + 0.00798121653497219, + 0.012334080412983894, + 0.012053223326802254, + 0.008260245434939861, + 0.009198848158121109, + 0.017932334914803505, + 0.015000646002590656, + 0.0036668686661869287, + 0.003802195191383362, + 0.006634872872382402, + 0.006400887854397297, + 0.0026934058405458927, + 0.004743010271340609, + 0.004452092107385397, + 0.0035681594163179398, + 0.010637464933097363, + 0.003953820094466209, + 0.004497881513088942, + 0.0058978768065571785, + 0.004323454573750496, + 0.03064737282693386, + 0.01067512296140194, + 0.027234697714447975, + 0.004132864996790886, + 0.003968523349612951, + 0.005995882209390402, + 0.007804456632584333, + 0.0034529431723058224, + 0.004261367488652468, + 0.007491971366107464, + 0.0053471764549613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.029799362644553185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004088971298187971 + ], + "opposite_logprob_tokens": [ + -5.504088878631592 + ], + "logprob_reward": -0.004088971298187971, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004088971298187971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2733183205127716, + 0.18644076585769653, + 0.1420173943042755, + 0.09430775791406631, + 0.061231039464473724, + 0.04913992062211037, + 0.0403166189789772, + 0.022192733362317085, + 0.056450456380844116, + 0.07458498328924179 + ], + "credit_map": [ + 0.03807453811168671, + 0.028109006583690643, + 0.04123230278491974, + 0.0342097170650959, + 0.01934974454343319, + 0.029091643169522285, + 0.05753815174102783, + 0.02571321837604046, + 0.026759883388876915, + 0.024081861600279808, + 0.02304423786699772, + 0.018846169114112854, + 0.03286035358905792, + 0.017317116260528564, + 0.017182689160108566, + 0.026348456740379333, + 0.017487429082393646, + 0.01859702169895172, + 0.015163499861955643, + 0.014684589579701424, + 0.02021079882979393, + 0.016765708103775978, + 0.016565008088946342, + 0.022543340921401978, + 0.010252518579363823, + 0.014488797634840012, + 0.010999016463756561, + 0.0073175374418497086, + 0.01719369925558567, + 0.014405905269086361, + 0.00882697757333517, + 0.010823301039636135, + 0.016080310568213463, + 0.005843058694154024, + 0.0066927396692335606, + 0.007299129851162434, + 0.005915660411119461, + 0.008263709023594856, + 0.005623903591185808, + 0.005512532312422991, + 0.004713535308837891, + 0.004205008503049612, + 0.006588560063391924, + 0.009201875887811184, + 0.006842736154794693, + 0.005572807509452105, + 0.007859935984015465, + 0.004155462607741356, + 0.003276441479101777, + 0.00387985585257411, + 0.00708033936098218, + 0.005226296838372946, + 0.00375392590649426, + 0.005138344597071409, + 0.006539283785969019, + 0.005422130227088928, + 0.0018236779142171144, + 0.001527120708487928, + 0.0038273201789706945, + 0.0037752599455416203, + 0.002518516732379794, + 0.0028082942590117455, + 0.003186452202498913, + 0.0027260908391326666, + 0.008762720972299576, + 0.0037314915098249912, + 0.004660641308873892, + 0.005321150179952383, + 0.0033497449476271868, + 0.017194440588355064, + 0.00519109470769763, + 0.008239173330366611, + 0.010244685225188732, + 0.00885109044611454, + 0.008673608303070068, + 0.008066064678132534, + 0.0074211363680660725, + 0.01448985654860735, + 0.008976447395980358, + 0.007862094789743423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.004088971298187971 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful low-angle shot, the camera almost close to the ground, tilts upward to capture the ground's sudden explosion. A massive, heavy-duty tractor, its bright yellow metal hull almost completely covered in thick, wet, dark brown mud, surges upward from the ground with unstoppable force, its massive, spiraling drill bit still spinning at high speed. The drill's sharp metal edge gleams coldly as it tears apart the soil. The intense centrifugal force hurls countless large and small rocks, clumps of earth, and uprooted grass into the air, creating a chaotic and spectacular rain of debris that instantly fills the frame. Under the bright, diffuse light of an overcast sky, each splatter of mud gleams with a wet sheen. The tractor's worn metal surface creates a striking textural contrast with the thick mud, and its massive, deeply treaded tires are also smeared with mud. The entire scene is imbued with a raw sense of industrial power and a dynamic, chaotic aesthetic, creating an awe-inspiring visual spectacle brimming with raw, kinetic energy.", + "seed": 26758, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_15/index_2632_seed_26758_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.045377960754558445, + "vlm_logprob": -0.045377960754558445, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.045377960754558445, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03686052933335304 + ], + "opposite_logprob_tokens": [ + -5.786860466003418 + ], + "logprob_reward": -0.03686052933335304, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.03686052933335304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1380508691072464, + 0.18281659483909607, + 0.04487800598144531, + 0.09689716249704361, + 0.03989076241850853, + 0.15349099040031433, + 0.053395453840494156, + 0.14296963810920715, + 0.07406535744667053, + 0.07354513555765152 + ], + "credit_map": [ + 0.014463888481259346, + 0.011941364035010338, + 0.01825171709060669, + 0.026802420616149902, + 0.011560015380382538, + 0.018236888572573662, + 0.023891789838671684, + 0.012902800925076008, + 0.028304336592555046, + 0.024715660139918327, + 0.04195764660835266, + 0.021927665919065475, + 0.01815059594810009, + 0.015532325021922588, + 0.011738496832549572, + 0.020489875227212906, + 0.004260560963302851, + 0.003981560934334993, + 0.005002127029001713, + 0.004615412559360266, + 0.004117946606129408, + 0.005563149694353342, + 0.008236580528318882, + 0.009100668132305145, + 0.01176031306385994, + 0.014490625821053982, + 0.007360472809523344, + 0.012910677120089531, + 0.014269856736063957, + 0.01560511626303196, + 0.011596106924116611, + 0.008903996087610722, + 0.005552996415644884, + 0.004466870799660683, + 0.003988166339695454, + 0.009374727495014668, + 0.004150380380451679, + 0.0043231346644461155, + 0.00455050403252244, + 0.0034839825239032507, + 0.024537526071071625, + 0.008771839551627636, + 0.06361839175224304, + 0.01392902061343193, + 0.008235260844230652, + 0.010248785838484764, + 0.015291317366063595, + 0.008858844637870789, + 0.0047609200701117516, + 0.006888716947287321, + 0.007034534588456154, + 0.005204896442592144, + 0.011980781331658363, + 0.004443813115358353, + 0.006628046743571758, + 0.006453744135797024, + 0.012587249279022217, + 0.018795838579535484, + 0.017556311562657356, + 0.017855951562523842, + 0.017001118510961533, + 0.025816043838858604, + 0.010131781920790672, + 0.023225337266921997, + 0.010042116045951843, + 0.00622263178229332, + 0.00637211324647069, + 0.009111406281590462, + 0.009919029660522938, + 0.020847829058766365, + 0.003658606205135584, + 0.00789162889122963, + 0.006521983537822962, + 0.015564320608973503, + 0.0050280471332371235, + 0.006661736872047186, + 0.017389457672834396, + 0.00742024602368474, + 0.00820311438292265, + 0.0067562321200966835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.03686052933335304 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281441271305084 + ], + "opposite_logprob_tokens": [ + -2.0178143978118896 + ], + "logprob_reward": -0.14281441271305084, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14281441271305084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05831976979970932, + 0.26300734281539917, + 0.06005758419632912, + 0.03820640966296196, + 0.04718755558133125, + 0.10497623682022095, + 0.03411763906478882, + 0.09885282814502716, + 0.2671804130077362, + 0.028094224631786346 + ], + "credit_map": [ + 0.006017283536493778, + 0.005841466598212719, + 0.007199707441031933, + 0.01195618137717247, + 0.005188732407987118, + 0.00862931925803423, + 0.00847737304866314, + 0.005009707063436508, + 0.033894505351781845, + 0.03349163755774498, + 0.06013840436935425, + 0.0356006994843483, + 0.02396819368004799, + 0.023269925266504288, + 0.020645659416913986, + 0.03199831768870354, + 0.006280442234128714, + 0.004902156535536051, + 0.009551474824547768, + 0.007436007726937532, + 0.005276485346257687, + 0.009498540312051773, + 0.01005271915346384, + 0.007059755735099316, + 0.004444247577339411, + 0.005384865682572126, + 0.0035617442335933447, + 0.004310848657041788, + 0.005289565306156874, + 0.006750698667019606, + 0.004776128567755222, + 0.0036883093416690826, + 0.005114263854920864, + 0.004417040850967169, + 0.007005779072642326, + 0.008335825055837631, + 0.006015880033373833, + 0.00605360884219408, + 0.005734398029744625, + 0.0045107584446668625, + 0.019802695140242577, + 0.016187625005841255, + 0.01925438456237316, + 0.010619770735502243, + 0.010576801374554634, + 0.01119478140026331, + 0.0082970280200243, + 0.00904315710067749, + 0.0028080244082957506, + 0.0042806873098015785, + 0.005451427772641182, + 0.0039981454610824585, + 0.006311153061687946, + 0.003001181408762932, + 0.003953480627387762, + 0.004313539247959852, + 0.013754844665527344, + 0.01368253119289875, + 0.014260463416576385, + 0.010448720306158066, + 0.013158874586224556, + 0.013329864479601383, + 0.009147463366389275, + 0.011070065200328827, + 0.026603620499372482, + 0.019366858527064323, + 0.017986081540584564, + 0.032455556094646454, + 0.036419980227947235, + 0.08595941960811615, + 0.012057849206030369, + 0.036331042647361755, + 0.0024342259857803583, + 0.006216726265847683, + 0.002213256899267435, + 0.001871043466962874, + 0.004925880581140518, + 0.0037402757443487644, + 0.004150944761931896, + 0.0025418726727366447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.14281441271305084 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002824367256835103 + ], + "opposite_logprob_tokens": [ + -5.877824306488037 + ], + "logprob_reward": -0.002824367256835103, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002824367256835103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10383673012256622, + 0.12923407554626465, + 0.0679270401597023, + 0.0825611799955368, + 0.0625104233622551, + 0.0972861796617508, + 0.04000702127814293, + 0.13570527732372284, + 0.17780132591724396, + 0.10313079506158829 + ], + "credit_map": [ + 0.01826031319797039, + 0.012988759204745293, + 0.015406974591314793, + 0.01516233291476965, + 0.011587114073336124, + 0.010641055181622505, + 0.011017026379704475, + 0.008773146197199821, + 0.01882907934486866, + 0.01731185056269169, + 0.015738781541585922, + 0.019416166469454765, + 0.012567764148116112, + 0.018366048112511635, + 0.012699376791715622, + 0.014304998330771923, + 0.008383207954466343, + 0.011693054810166359, + 0.006231930106878281, + 0.007195603102445602, + 0.008910230360925198, + 0.011173974722623825, + 0.00857777614146471, + 0.005761255510151386, + 0.010274652391672134, + 0.011404485441744328, + 0.008358636870980263, + 0.007377978879958391, + 0.014273208566009998, + 0.01270279847085476, + 0.010263068601489067, + 0.007906344719231129, + 0.005488514434546232, + 0.007225050590932369, + 0.009363503195345402, + 0.0077675445936620235, + 0.009923065081238747, + 0.010063027031719685, + 0.008527948521077633, + 0.004151767585426569, + 0.0066332146525382996, + 0.0069153993390500546, + 0.02952185459434986, + 0.014816556125879288, + 0.008184010162949562, + 0.006169419270008802, + 0.013371369801461697, + 0.01167435571551323, + 0.004313934128731489, + 0.004090436268597841, + 0.00444062752649188, + 0.004833861254155636, + 0.00706606637686491, + 0.0038869993295520544, + 0.0049789766781032085, + 0.006396117620170116, + 0.016143737360835075, + 0.0192800834774971, + 0.02279341034591198, + 0.015537073835730553, + 0.019954141229391098, + 0.016419673338532448, + 0.011083097197115421, + 0.014494042843580246, + 0.02418031170964241, + 0.015445572324097157, + 0.01361799705773592, + 0.0221425611525774, + 0.023762518540024757, + 0.0503135584294796, + 0.008962421678006649, + 0.01937636360526085, + 0.008217505179345608, + 0.023084254935383797, + 0.00640076445415616, + 0.008214417845010757, + 0.024079669266939163, + 0.013279005885124207, + 0.010429234243929386, + 0.009425938129425049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.002824367256835103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305814579129219 + ], + "opposite_logprob_tokens": [ + -3.168058156967163 + ], + "logprob_reward": -0.04305814579129219, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305814579129219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05688629299402237, + 0.1022288054227829, + 0.05959401652216911, + 0.06344546377658844, + 0.08591799437999725, + 0.06586766988039017, + 0.03545141965150833, + 0.22920776903629303, + 0.06399913877248764, + 0.23740142583847046 + ], + "credit_map": [ + 0.011733297258615494, + 0.00876056682318449, + 0.008123012259602547, + 0.005840409081429243, + 0.00666855601593852, + 0.005322541110217571, + 0.004282835870981216, + 0.006155076902359724, + 0.0135645205155015, + 0.012629061937332153, + 0.022541193291544914, + 0.010935239493846893, + 0.0111482422798872, + 0.015302550978958607, + 0.005551587790250778, + 0.010556412860751152, + 0.006143641658127308, + 0.0068899220786988735, + 0.0055662766098976135, + 0.008543320931494236, + 0.0070782676339149475, + 0.010678968392312527, + 0.007906035520136356, + 0.006787585094571114, + 0.008035298436880112, + 0.008481455966830254, + 0.005203908309340477, + 0.006982433144003153, + 0.010119340382516384, + 0.011324649676680565, + 0.007148068398237228, + 0.0061503080651164055, + 0.008864366449415684, + 0.008797813206911087, + 0.008207478560507298, + 0.019268548116087914, + 0.012863263487815857, + 0.011202819645404816, + 0.009660942479968071, + 0.007052766624838114, + 0.010974694974720478, + 0.00839169230312109, + 0.005947035737335682, + 0.007668355014175177, + 0.009193874895572662, + 0.007141741458326578, + 0.007070802617818117, + 0.009479466825723648, + 0.003207170870155096, + 0.004080124665051699, + 0.005559446290135384, + 0.0037871459499001503, + 0.00593150407075882, + 0.002991040935739875, + 0.004921891260892153, + 0.004973095376044512, + 0.033981334418058395, + 0.019480131566524506, + 0.05092692747712135, + 0.023374035954475403, + 0.020840659737586975, + 0.04740288481116295, + 0.009085230529308319, + 0.02411656826734543, + 0.009395183995366096, + 0.00687283044680953, + 0.00633005378767848, + 0.0078610023483634, + 0.007520187646150589, + 0.01261801179498434, + 0.005159339401870966, + 0.0082425307482481, + 0.017941797152161598, + 0.05723924562335014, + 0.015663912519812584, + 0.017304567620158195, + 0.0577768050134182, + 0.02755611389875412, + 0.023627685382962227, + 0.020291294902563095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.04305814579129219 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, almost close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward to capture the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive tractor with a bright yellow metal hull visible, covered in thick dark brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tractor shown surging upward from the ground with a spinning drill bit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with a chaotic rain of debris including rocks, earth, and uprooted grass?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013323486782610416 + ], + "opposite_logprob_tokens": [ + -6.6263322830200195 + ], + "logprob_reward": -0.0013323486782610416, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013323486782610416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08729847520589828, + 0.19756032526493073, + 0.1639597862958908, + 0.06650225818157196, + 0.07163149863481522, + 0.17149613797664642, + 0.021774861961603165, + 0.05514839291572571, + 0.03914891555905342, + 0.12547928094863892 + ], + "credit_map": [ + 0.018410010263323784, + 0.01434363890439272, + 0.007120040245354176, + 0.008900840766727924, + 0.0058790841139853, + 0.00881026592105627, + 0.01002498622983694, + 0.013809616677463055, + 0.029899341985583305, + 0.026824137195944786, + 0.03634093329310417, + 0.02634492702782154, + 0.018521981313824654, + 0.020863240584731102, + 0.016912437975406647, + 0.021853335201740265, + 0.0152904586866498, + 0.010439164936542511, + 0.029148660600185394, + 0.022936448454856873, + 0.012398852966725826, + 0.01597292721271515, + 0.03135668858885765, + 0.0264165960252285, + 0.009157474152743816, + 0.009426494129002094, + 0.0072270343080163, + 0.008220448158681393, + 0.009076228365302086, + 0.007670537102967501, + 0.00797899067401886, + 0.007745055947452784, + 0.009059990756213665, + 0.005825191270560026, + 0.009192807599902153, + 0.009085395373404026, + 0.011375858448445797, + 0.008397813886404037, + 0.012807088904082775, + 0.005887357983738184, + 0.017775297164916992, + 0.01305313128978014, + 0.06121267378330231, + 0.02597539685666561, + 0.009972866624593735, + 0.015045732259750366, + 0.013823562301695347, + 0.014637499116361141, + 0.0027319276705384254, + 0.0021570385433733463, + 0.0026727973017841578, + 0.0030320612713694572, + 0.0033854746725410223, + 0.0023959169629961252, + 0.0022249033208936453, + 0.003174742916598916, + 0.005504492670297623, + 0.007038527633994818, + 0.007395269814878702, + 0.006340258289128542, + 0.005826868582516909, + 0.007478255778551102, + 0.006038829684257507, + 0.009525896981358528, + 0.0032458107452839613, + 0.007001966238021851, + 0.004286719020456076, + 0.0040695625357329845, + 0.004497145302593708, + 0.007089072838425636, + 0.0026000412181019783, + 0.0063586002215743065, + 0.009306049905717373, + 0.032564468681812286, + 0.007872489280998707, + 0.010565556585788727, + 0.02984975278377533, + 0.013551760464906693, + 0.01200835034251213, + 0.009760863147675991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.0013323486782610416 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_3/index_tensor([1630, 1630])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_3/index_tensor([1630, 1630])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d7ca3f291912569741f7b5653acfc5f36b91377f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_3/index_tensor([1630, 1630])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a level angle, slowly inching inward, transports the viewer deep into a lush, light-dappled bamboo forest. A woman, clad in a simple white linen robe, her cascading emerald green hair as vibrant as the tender shoots of early spring, soft and lustrous. Her graceful figure forms a gentle curve, forming a graceful S-shaped composition alongside the shimmering, clear stream that meanders beside her. Her face, clear and beautiful, her expression serene and focused, grips a dark, textured wooden comb. Starting from the roots, she meticulously and deliberately combs through her hair, each stroke of the comb teeth creating a tranquil, ritualistic dialogue. The afternoon sun filters through the gaps between the towering bamboo leaves, casting countless shattered, golden specks of light across her white robe, her green hair, and the moist earth, creating a dreamlike effect. The entire scene is filled with an ethereal and poetic atmosphere, presenting the exquisite beauty unique to film animation, full of the harmonious beauty of oriental Zen and nature.", + "seed": 12737, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_3/index_1630_seed_12737_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.035293451882898806, + "vlm_logprob": -0.035293451882898806, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.035293451882898806, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0058249724097549915 + ], + "opposite_logprob_tokens": [ + -5.630825042724609 + ], + "logprob_reward": -0.0058249724097549915, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0058249724097549915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41760796308517456, + 0.1112392246723175, + 0.058222535997629166, + 0.09245821833610535, + 0.08112139254808426, + 0.06180966645479202, + 0.046329155564308167, + 0.031376201659440994, + 0.06394167244434357, + 0.035893961787223816 + ], + "credit_map": [ + 0.08470757305622101, + 0.04790801927447319, + 0.040497004985809326, + 0.018523888662457466, + 0.03330433741211891, + 0.12645982205867767, + 0.02774795889854431, + 0.038459356874227524, + 0.029627477750182152, + 0.025304405018687248, + 0.01832929067313671, + 0.007967638783156872, + 0.006710284389555454, + 0.00972716324031353, + 0.0069189430214464664, + 0.00665403762832284, + 0.019313380122184753, + 0.003911511041224003, + 0.018319664523005486, + 0.004167619626969099, + 0.0031413286924362183, + 0.0029034449253231287, + 0.0034809180069714785, + 0.002984673483297229, + 0.012107706628739834, + 0.02533596381545067, + 0.00806270632892847, + 0.007342349272221327, + 0.008225830271840096, + 0.004480150993913412, + 0.021581526845693588, + 0.005321994889527559, + 0.008190223015844822, + 0.011375016532838345, + 0.012350955046713352, + 0.012718357145786285, + 0.011197172105312347, + 0.008296065032482147, + 0.008989444002509117, + 0.00800416525453329, + 0.007002965081483126, + 0.011774147860705853, + 0.009990504942834377, + 0.0068830144591629505, + 0.006022839341312647, + 0.008177596144378185, + 0.006147108506411314, + 0.005811493843793869, + 0.004876566119492054, + 0.005771196447312832, + 0.007805756293237209, + 0.005231386981904507, + 0.006421093363314867, + 0.004649546463042498, + 0.0054198745638132095, + 0.006153739057481289, + 0.0038473617751151323, + 0.003319615265354514, + 0.004214223939925432, + 0.004261622671037912, + 0.004781488329172134, + 0.002753550186753273, + 0.003962870687246323, + 0.00423547113314271, + 0.009324544109404087, + 0.008058696053922176, + 0.012979628518223763, + 0.00909112486988306, + 0.005303769838064909, + 0.004680783022195101, + 0.0054490151815116405, + 0.009054112248122692, + 0.004453396890312433, + 0.0038821592461317778, + 0.005765571724623442, + 0.004979309625923634, + 0.004548322409391403, + 0.003650934435427189, + 0.004808368626981974, + 0.003805899526923895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0058249724097549915 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639812279492617 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.004639812279492617, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639812279492617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2080123871564865, + 0.08630643784999847, + 0.08644812554121017, + 0.09641717374324799, + 0.05322125554084778, + 0.08273336291313171, + 0.15503434836864471, + 0.06560240685939789, + 0.10291171073913574, + 0.06331275403499603 + ], + "credit_map": [ + 0.044489044696092606, + 0.022944573312997818, + 0.02236936055123806, + 0.012324744835495949, + 0.017122654244303703, + 0.048587389290332794, + 0.016140317544341087, + 0.024034317582845688, + 0.013080017641186714, + 0.02516140043735504, + 0.006277177948504686, + 0.009863262064754963, + 0.007720813155174255, + 0.010552243329584599, + 0.007329470477998257, + 0.006322052329778671, + 0.024762103334069252, + 0.011255982331931591, + 0.02993813343346119, + 0.005002290476113558, + 0.003830876899883151, + 0.0045701684430241585, + 0.004539819899946451, + 0.0025487616658210754, + 0.013285747729241848, + 0.029701141640543938, + 0.006939507555216551, + 0.007307345047593117, + 0.007435689214617014, + 0.005758953280746937, + 0.020091257989406586, + 0.00589754106476903, + 0.005277095828205347, + 0.006428991910070181, + 0.006243590265512466, + 0.006745527498424053, + 0.009337585419416428, + 0.005188071168959141, + 0.008894344791769981, + 0.005106051452457905, + 0.008065732195973396, + 0.010742941871285439, + 0.012796924449503422, + 0.008964311331510544, + 0.010191262699663639, + 0.011065344326198101, + 0.009931255131959915, + 0.010975598357617855, + 0.014424522407352924, + 0.01978822983801365, + 0.015403904020786285, + 0.013504911214113235, + 0.029482470825314522, + 0.02340359427034855, + 0.023600701242685318, + 0.015426025725901127, + 0.004844848997890949, + 0.008117477409541607, + 0.005049738567322493, + 0.006949838250875473, + 0.012152639217674732, + 0.006877942476421595, + 0.013631648384034634, + 0.007978278212249279, + 0.01039185095578432, + 0.011092457920312881, + 0.016070296987891197, + 0.012807881459593773, + 0.012498962692916393, + 0.008892755955457687, + 0.017866848036646843, + 0.013290660455822945, + 0.005432142410427332, + 0.0061239381320774555, + 0.0065946211107075214, + 0.009213151410222054, + 0.011372735723853111, + 0.005172420758754015, + 0.011333377100527287, + 0.008070370182394981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.004639812279492617 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796933755278587 + ], + "opposite_logprob_tokens": [ + -3.529796838760376 + ], + "logprob_reward": -0.029796933755278587, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796933755278587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3520217537879944, + 0.09624385088682175, + 0.053295932710170746, + 0.14350078999996185, + 0.058976706117391586, + 0.11808383464813232, + 0.07054391503334045, + 0.030910516157746315, + 0.03865909203886986, + 0.03776369243860245 + ], + "credit_map": [ + 0.08964569121599197, + 0.03207772970199585, + 0.041996922343969345, + 0.020348014310002327, + 0.025390654802322388, + 0.09570370614528656, + 0.019693154841661453, + 0.02716585248708725, + 0.014994216151535511, + 0.006291960831731558, + 0.025111891329288483, + 0.006842827424407005, + 0.005912597756832838, + 0.025450870394706726, + 0.0065013510175049305, + 0.005138130858540535, + 0.008856190368533134, + 0.009613784030079842, + 0.012333194725215435, + 0.0060912626795470715, + 0.004326587077230215, + 0.004547836259007454, + 0.004676819778978825, + 0.002850259654223919, + 0.01554401870816946, + 0.045518990606069565, + 0.00888222549110651, + 0.01059267669916153, + 0.012696284800767899, + 0.007886840030550957, + 0.03396167233586311, + 0.008418095298111439, + 0.006043992005288601, + 0.005903808865696192, + 0.006887662690132856, + 0.0054445876739919186, + 0.00631911214441061, + 0.005758543964475393, + 0.0073194862343370914, + 0.015299509279429913, + 0.009694433771073818, + 0.012501970864832401, + 0.01436727587133646, + 0.014391958713531494, + 0.01637376844882965, + 0.01585257425904274, + 0.016527006402611732, + 0.01837485283613205, + 0.0057823839597404, + 0.009874546900391579, + 0.008628652431070805, + 0.005437894258648157, + 0.013695885427296162, + 0.011228423565626144, + 0.008114558644592762, + 0.0077815731056034565, + 0.002673584036529064, + 0.005107779521495104, + 0.0033489796333014965, + 0.0029212499503046274, + 0.0060335127636790276, + 0.002793361432850361, + 0.003786080051213503, + 0.004245969001203775, + 0.006121853366494179, + 0.004631431773304939, + 0.0066749099642038345, + 0.005263035651296377, + 0.0042152078822255135, + 0.002706650411710143, + 0.004470078740268946, + 0.004575921688228846, + 0.003540720557793975, + 0.0033972638193517923, + 0.0036784650292247534, + 0.005298609379678965, + 0.0068880850449204445, + 0.0043501039035618305, + 0.00613561412319541, + 0.004474828019738197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.029796933755278587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232884231954813 + ], + "opposite_logprob_tokens": [ + -3.7732884883880615 + ], + "logprob_reward": -0.0232884231954813, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232884231954813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2075241208076477, + 0.27292898297309875, + 0.08694064617156982, + 0.08473525941371918, + 0.07676059752702713, + 0.12830233573913574, + 0.04988739639520645, + 0.03708747401833534, + 0.017849069088697433, + 0.03798411041498184 + ], + "credit_map": [ + 0.038148149847984314, + 0.02072390541434288, + 0.02134162001311779, + 0.0181905347853899, + 0.020460762083530426, + 0.04023156687617302, + 0.03133675456047058, + 0.017090823501348495, + 0.05134247988462448, + 0.06534919887781143, + 0.07439782470464706, + 0.019363708794116974, + 0.011689280159771442, + 0.02427908033132553, + 0.014189863577485085, + 0.01231754757463932, + 0.018951155245304108, + 0.004552728962153196, + 0.02280021645128727, + 0.005498757120221853, + 0.00883082952350378, + 0.009670089930295944, + 0.009586649015545845, + 0.007050217129290104, + 0.01925201714038849, + 0.011112536303699017, + 0.010609873570501804, + 0.006876625586301088, + 0.009101231582462788, + 0.005681629292666912, + 0.01463298685848713, + 0.007468358613550663, + 0.00805802084505558, + 0.008648717775940895, + 0.008019449189305305, + 0.009266545064747334, + 0.013373921625316143, + 0.0125718479976058, + 0.009735886938869953, + 0.00708621134981513, + 0.01182637270539999, + 0.015464292839169502, + 0.017078028991818428, + 0.01386963576078415, + 0.01625889539718628, + 0.014678862877190113, + 0.017289936542510986, + 0.021836308762431145, + 0.005009173881262541, + 0.006255025509744883, + 0.004658557008951902, + 0.004919447470456362, + 0.011614816263318062, + 0.005832793656736612, + 0.006409323774278164, + 0.005188256036490202, + 0.003427425166592002, + 0.004106155596673489, + 0.0036727129481732845, + 0.00459524430334568, + 0.007167174946516752, + 0.0030776450876146555, + 0.005780989304184914, + 0.005260126665234566, + 0.0018967309733852744, + 0.0016460057813674212, + 0.0024150244425982237, + 0.002433215035125613, + 0.0026921560056507587, + 0.0020829085260629654, + 0.002057885518297553, + 0.0026251429226249456, + 0.003124575363472104, + 0.0037101800553500652, + 0.0043673827312886715, + 0.0050534592010080814, + 0.00702710822224617, + 0.003970701713114977, + 0.005442249588668346, + 0.005288451444357634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.0232884231954813 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291711777448654 + ], + "opposite_logprob_tokens": [ + -2.237917184829712 + ], + "logprob_reward": -0.11291711777448654, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291711777448654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3451196849346161, + 0.2575119435787201, + 0.06491208076477051, + 0.06938701868057251, + 0.05767729505896568, + 0.06632162630558014, + 0.044179704040288925, + 0.018905911594629288, + 0.032809559255838394, + 0.04317520186305046 + ], + "credit_map": [ + 0.0403023324906826, + 0.04403722286224365, + 0.015315351076424122, + 0.017893735319375992, + 0.037779390811920166, + 0.12298494577407837, + 0.04236423596739769, + 0.024442486464977264, + 0.09052421897649765, + 0.014891806989908218, + 0.09412480890750885, + 0.018260205164551735, + 0.007273039780557156, + 0.009509121999144554, + 0.011946538463234901, + 0.010982204228639603, + 0.009342644363641739, + 0.007644069846719503, + 0.014056592248380184, + 0.007753514219075441, + 0.005808220244944096, + 0.007024834863841534, + 0.006597583647817373, + 0.006684623658657074, + 0.008997316472232342, + 0.0138997882604599, + 0.007345270365476608, + 0.005776173900812864, + 0.007785622030496597, + 0.003573794150725007, + 0.017384398728609085, + 0.004624664317816496, + 0.005487619899213314, + 0.009899388998746872, + 0.0061069056391716, + 0.008521007373929024, + 0.00536897499114275, + 0.006969332695007324, + 0.005924988072365522, + 0.009399082511663437, + 0.006923728156834841, + 0.008134562522172928, + 0.010878213681280613, + 0.008093161508440971, + 0.008085471577942371, + 0.006859172601252794, + 0.008923825807869434, + 0.008423496969044209, + 0.004593366757035255, + 0.005894757341593504, + 0.006017304491251707, + 0.00552604254335165, + 0.005027426406741142, + 0.005142532289028168, + 0.0051432461477816105, + 0.006835028994828463, + 0.0023420716170221567, + 0.002131304470822215, + 0.0025188561994582415, + 0.0021169891115278006, + 0.002282509347423911, + 0.0018746014684438705, + 0.0022732012439519167, + 0.0033663802314549685, + 0.004642653279006481, + 0.0042617106810212135, + 0.005728833377361298, + 0.003663774812594056, + 0.003448639065027237, + 0.0030105221085250378, + 0.0034593979362398386, + 0.004594033118337393, + 0.005726806819438934, + 0.003974590450525284, + 0.006321974564343691, + 0.007431051693856716, + 0.005735552869737148, + 0.0040605803951621056, + 0.005114664323627949, + 0.004809982143342495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.11291711777448654 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly inching inward, transports the viewer deep into a lush, light-dappled bamboo forest. A woman, clad in a simple white linen robe, her cascading emerald green hair as vibrant as the tender shoots of early spring, soft and lustrous. Her graceful figure forms a gentle curve, forming a graceful S-shaped composition alongside the shimmering, clear stream that meanders beside her. Her face, clear and beautiful, her expression serene and focused, grips a dark, textured wooden comb. Starting from the roots, she meticulously and deliberately combs through her hair, each stroke of the comb teeth creating a tranquil, ritualistic dialogue. The afternoon sun filters through the gaps between the towering bamboo leaves, casting countless shattered, golden specks of light across her white robe, her green hair, and the moist earth, creating a dreamlike effect. The entire scene is filled with an ethereal and poetic atmosphere, presenting the exquisite beauty unique to film animation, full of the harmonious beauty of oriental Zen and nature.", + "seed": 12738, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_3/index_1630_seed_12738_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.040103968838229774, + "vlm_logprob": -0.040103968838229774, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.040103968838229774, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013422718271613121 + ], + "opposite_logprob_tokens": [ + -4.63842248916626 + ], + "logprob_reward": -0.013422718271613121, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013422718271613121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7711838483810425, + 0.04899462312459946, + 0.09967588633298874, + 0.0105585977435112, + 0.004861634224653244, + 0.006704432424157858, + 0.007609244901686907, + 0.01589469239115715, + 0.015590062364935875, + 0.01892690360546112 + ], + "credit_map": [ + 0.04883969947695732, + 0.023017918691039085, + 0.04962831363081932, + 0.02797919698059559, + 0.2565421164035797, + 0.30795061588287354, + 0.025464707985520363, + 0.03176135569810867, + 0.0032214594539254904, + 0.0019354127580299973, + 0.002585392678156495, + 0.0017986522288993, + 0.016437018290162086, + 0.01833861880004406, + 0.00203490792773664, + 0.002643163548782468, + 0.0070980931632220745, + 0.0023763931822031736, + 0.004020712338387966, + 0.004147868603467941, + 0.035172611474990845, + 0.03853071108460426, + 0.0034607858397066593, + 0.004868719261139631, + 0.0013975038891658187, + 0.0006383116124197841, + 0.0008748254040256143, + 0.0006226194091141224, + 0.0023240293376147747, + 0.002514774212613702, + 0.0012727257562801242, + 0.0009138083551079035, + 0.0005204260814934969, + 0.00032640868448652327, + 0.0005522677092812955, + 0.00042112902156077325, + 0.0011488982709124684, + 0.0012459980789572, + 0.0002710291591938585, + 0.0003754778881557286, + 0.0011164696188643575, + 0.0007207612507045269, + 0.0007388823432847857, + 0.0007090274593792856, + 0.001154993544332683, + 0.001046543475240469, + 0.0005784120294265449, + 0.0006393429939635098, + 0.0009094189153984189, + 0.0008405622211284935, + 0.0005603766185231507, + 0.0006800986593589187, + 0.0010418565943837166, + 0.0011081494158133864, + 0.00137238553725183, + 0.0010963972890749574, + 0.0023361986968666315, + 0.0010992741445079446, + 0.0019669763278216124, + 0.0014738204190507531, + 0.0018638534238561988, + 0.0027690932620316744, + 0.0023766763042658567, + 0.0020088020246475935, + 0.0015703779645264149, + 0.0009432798833586276, + 0.001636249478906393, + 0.0011455070925876498, + 0.002600593026727438, + 0.003398030763491988, + 0.002859878120943904, + 0.0014361480716615915, + 0.0017175733810290694, + 0.001244941377080977, + 0.0010913320584222674, + 0.0010425951331853867, + 0.005789135582745075, + 0.005372757092118263, + 0.001301031094044447, + 0.0013675373047590256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.013422718271613121 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028269824106246233 + ], + "opposite_logprob_tokens": [ + -5.877827167510986 + ], + "logprob_reward": -0.0028269824106246233, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028269824106246233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9323444962501526, + 0.014473329298198223, + 0.008327202871441841, + 0.010536139830946922, + 0.002137106843292713, + 0.0038458690978586674, + 0.0021869109477847815, + 0.016976116225123405, + 0.004367888439446688, + 0.004804880823940039 + ], + "credit_map": [ + 0.05939943715929985, + 0.0270877443253994, + 0.059609416872262955, + 0.033598706126213074, + 0.3117104768753052, + 0.3714539706707001, + 0.030897384509444237, + 0.038587357848882675, + 0.0011815485777333379, + 0.0009456890402361751, + 0.0009070919477380812, + 0.0008597312262281775, + 0.003885050071403384, + 0.004070374183356762, + 0.0011494893115013838, + 0.001474353950470686, + 0.0006601458298973739, + 0.0004563844995573163, + 0.0003902011667378247, + 0.0004583695554174483, + 0.002598932245746255, + 0.0028018024750053883, + 0.00044322668691165745, + 0.0005181399756111205, + 0.0006881857989355922, + 0.00042251060949638486, + 0.0005198894068598747, + 0.0007205531001091003, + 0.0031079307664185762, + 0.003350462531670928, + 0.000982666271738708, + 0.0007439410546794534, + 0.00022981548681855202, + 0.00019599685037974268, + 0.00018479324353393167, + 0.00030532715027220547, + 0.00022887768864165992, + 0.00027348785079084337, + 0.00044417098979465663, + 0.00027463745209388435, + 0.0004211398772895336, + 0.0002961296704597771, + 0.00027605079230852425, + 0.00045696060988120735, + 0.0006251116283237934, + 0.0006887797499075532, + 0.0006516689900308847, + 0.00043002760503441095, + 0.0003060513990931213, + 0.0001733012031763792, + 0.0002557968255132437, + 0.00022113646264187992, + 0.00028705730801448226, + 0.00025430633104406297, + 0.00036933092633262277, + 0.0003199302882421762, + 0.0012738866498693824, + 0.0013242020504549146, + 0.0008530320483259857, + 0.0009640716598369181, + 0.0038103212136775255, + 0.004372290801256895, + 0.003036340931430459, + 0.0013419706374406815, + 0.0003909229126293212, + 0.0005044103600084782, + 0.0003172830038238317, + 0.0003809933550655842, + 0.0009692225721664727, + 0.0010490322019904852, + 0.00038976638461463153, + 0.00036625692155212164, + 0.0004310775257181376, + 0.00047058548079803586, + 0.0003447599592618644, + 0.00036580985761247575, + 0.0012356372317299247, + 0.0010965964756906033, + 0.00044392820564098656, + 0.0004164853598922491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0028269824106246233 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023286793380975723 + ], + "opposite_logprob_tokens": [ + -3.773286819458008 + ], + "logprob_reward": -0.023286793380975723, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023286793380975723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8755185604095459, + 0.019853776320815086, + 0.049869634211063385, + 0.008216221816837788, + 0.005102300550788641, + 0.006305768620222807, + 0.008068710565567017, + 0.014281731098890305, + 0.00813286006450653, + 0.004650420043617487 + ], + "credit_map": [ + 0.05591191351413727, + 0.02533992938697338, + 0.05611107498407364, + 0.03143785893917084, + 0.2918775975704193, + 0.3489953875541687, + 0.029196185991168022, + 0.03664856031537056, + 0.0013423473574221134, + 0.0009883205639198422, + 0.0012254125904291868, + 0.0008147818152792752, + 0.006364479660987854, + 0.006905433721840382, + 0.001066432218067348, + 0.0011465669376775622, + 0.0035160311963409185, + 0.001235553645528853, + 0.0019345112377777696, + 0.002095417818054557, + 0.01769743114709854, + 0.0192885622382164, + 0.001716479193419218, + 0.00238564470782876, + 0.0006549422396346927, + 0.0004165342834312469, + 0.000495264888741076, + 0.0005822668899782002, + 0.002354174619540572, + 0.002551524667069316, + 0.0006260574446059763, + 0.0005354566383175552, + 0.0005881937686353922, + 0.00048329346464015543, + 0.0004821906331926584, + 0.0006418628036044538, + 0.000590797804761678, + 0.0006529334932565689, + 0.0010003207717090845, + 0.0006627074908465147, + 0.00094655278371647, + 0.0008816223707981408, + 0.0007188706658780575, + 0.0006870828801766038, + 0.0008117060642689466, + 0.0008512792992405593, + 0.0007606244180351496, + 0.0006480297306552529, + 0.000669819419272244, + 0.0006971370312385261, + 0.0005895326612517238, + 0.0007984754047356546, + 0.0012343608541414142, + 0.001235010800883174, + 0.0018336407374590635, + 0.0010107324924319983, + 0.0011849291622638702, + 0.0025798201095312834, + 0.000914578209631145, + 0.0014647745992988348, + 0.002587392460554838, + 0.002960287733003497, + 0.0013141244417056441, + 0.0012758247321471572, + 0.0005957173416391015, + 0.0008898662636056542, + 0.00046164565719664097, + 0.0005264669889584184, + 0.002173434244468808, + 0.0023789156693965197, + 0.0005029634921811521, + 0.000603849533945322, + 0.00034864855115301907, + 0.0005430024466477334, + 0.00034378672717139125, + 0.0005191010423004627, + 0.0007049671839922667, + 0.0005995706305839121, + 0.0008525323355570436, + 0.0007388110389001667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.023286793380975723 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189411610364914 + ], + "opposite_logprob_tokens": [ + -4.018189430236816 + ], + "logprob_reward": -0.018189411610364914, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018189411610364914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5225591659545898, + 0.054350968450307846, + 0.2050601840019226, + 0.029318956658244133, + 0.03318845480680466, + 0.024590466171503067, + 0.02484254352748394, + 0.023398153483867645, + 0.033661019057035446, + 0.049030084162950516 + ], + "credit_map": [ + 0.033665046095848083, + 0.015988439321517944, + 0.03310669958591461, + 0.01934351585805416, + 0.17483186721801758, + 0.2062251716852188, + 0.01861758716404438, + 0.020780859515070915, + 0.004727871622890234, + 0.003154660575091839, + 0.004094320349395275, + 0.0028072046115994453, + 0.014032048173248768, + 0.0157634150236845, + 0.004735864698886871, + 0.00503558712080121, + 0.014109993353486061, + 0.0050345840863883495, + 0.008199887350201607, + 0.008711475878953934, + 0.07240743190050125, + 0.07848663628101349, + 0.008397286757826805, + 0.00971289724111557, + 0.002689287532120943, + 0.0027676045428961515, + 0.0029116347432136536, + 0.0023829403799027205, + 0.0034550365526229143, + 0.0035701196175068617, + 0.007216440048068762, + 0.004325893707573414, + 0.0032829276751726866, + 0.0031509895343333483, + 0.0028668642044067383, + 0.004977706354111433, + 0.0036158328875899315, + 0.004466549027711153, + 0.006678057834506035, + 0.004149528685957193, + 0.0017150529893115163, + 0.0016017151065170765, + 0.002368156099691987, + 0.002372780814766884, + 0.0048784432001411915, + 0.005755506455898285, + 0.0031324082519859076, + 0.0027664031367748976, + 0.0025247775483876467, + 0.001966263400390744, + 0.001962953247129917, + 0.0024651437997817993, + 0.0033805794082581997, + 0.003538276068866253, + 0.006222795695066452, + 0.002781753661110997, + 0.002115522976964712, + 0.002752790693193674, + 0.0016714532393962145, + 0.0024394653737545013, + 0.0024353626649826765, + 0.0029807493556290865, + 0.005635458510369062, + 0.003367348574101925, + 0.002728459658101201, + 0.002199449110776186, + 0.0017938960809260607, + 0.001796330907382071, + 0.00979131180793047, + 0.010487858206033707, + 0.0026361567433923483, + 0.0022275575902312994, + 0.003835177281871438, + 0.0030022412538528442, + 0.0027451275382190943, + 0.002378434408456087, + 0.015158126130700111, + 0.014919665642082691, + 0.0031224710401147604, + 0.0038688413333147764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.018189411610364914 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427939385175705 + ], + "opposite_logprob_tokens": [ + -2.017793893814087 + ], + "logprob_reward": -0.1427939385175705, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1427939385175705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6675244569778442, + 0.059291865676641464, + 0.02933203987777233, + 0.01041142176836729, + 0.014114183373749256, + 0.006465807091444731, + 0.005566846579313278, + 0.014702126383781433, + 0.0425444096326828, + 0.15004685521125793 + ], + "credit_map": [ + 0.04283895716071129, + 0.019444746896624565, + 0.04273656755685806, + 0.024047544226050377, + 0.22300562262535095, + 0.2654431462287903, + 0.02220010571181774, + 0.027807798236608505, + 0.004532152786850929, + 0.0027974166441708803, + 0.0031326906755566597, + 0.0024906026665121317, + 0.018435383215546608, + 0.021071389317512512, + 0.003138704691082239, + 0.00369353243149817, + 0.0019507521064952016, + 0.0007408726960420609, + 0.0014082288835197687, + 0.0012876575347036123, + 0.009825796820223331, + 0.010459505021572113, + 0.0016802464378997684, + 0.0019789852667599916, + 0.0011602562153711915, + 0.0007831725524738431, + 0.0008324700174853206, + 0.0006022940506227314, + 0.0024022820871323347, + 0.002710615051910281, + 0.0011728991521522403, + 0.000747433805372566, + 0.0011871112510561943, + 0.0009312398033216596, + 0.00110261095687747, + 0.0016901983181014657, + 0.00205916166305542, + 0.0020906724967062473, + 0.0030608244705945253, + 0.0019923648796975613, + 0.0007160339737311006, + 0.0007389125530607998, + 0.00047204020665958524, + 0.0007741677109152079, + 0.0010014892322942615, + 0.0009034255053848028, + 0.001054945751093328, + 0.0008047929732128978, + 0.0006666545523330569, + 0.0008166507468558848, + 0.00047407913370989263, + 0.0005656099529005587, + 0.0009664208046160638, + 0.0005774087039753795, + 0.0009978256421163678, + 0.0005021975375711918, + 0.0013683554716408253, + 0.0014120584819465876, + 0.0009842299623414874, + 0.001956168795004487, + 0.00272170128300786, + 0.0027063849847763777, + 0.0022903080098330975, + 0.001262919744476676, + 0.003209978574886918, + 0.002314287470653653, + 0.002218173583969474, + 0.001713474397547543, + 0.012584665790200233, + 0.013892298564314842, + 0.004342163447290659, + 0.002269366290420294, + 0.011924690566956997, + 0.008840810507535934, + 0.009428410790860653, + 0.006636731326580048, + 0.0494869239628315, + 0.046702563762664795, + 0.00878025684505701, + 0.008246480487287045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.1427939385175705 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly inching inward, transports the viewer deep into a lush, light-dappled bamboo forest. A woman, clad in a simple white linen robe, her cascading emerald green hair as vibrant as the tender shoots of early spring, soft and lustrous. Her graceful figure forms a gentle curve, forming a graceful S-shaped composition alongside the shimmering, clear stream that meanders beside her. Her face, clear and beautiful, her expression serene and focused, grips a dark, textured wooden comb. Starting from the roots, she meticulously and deliberately combs through her hair, each stroke of the comb teeth creating a tranquil, ritualistic dialogue. The afternoon sun filters through the gaps between the towering bamboo leaves, casting countless shattered, golden specks of light across her white robe, her green hair, and the moist earth, creating a dreamlike effect. The entire scene is filled with an ethereal and poetic atmosphere, presenting the exquisite beauty unique to film animation, full of the harmonious beauty of oriental Zen and nature.", + "seed": 12739, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_3/index_1630_seed_12739_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07240849174559116, + "vlm_logprob": -0.07240849174559116, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.07240849174559116, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06524363160133362 + ], + "opposite_logprob_tokens": [ + -2.815243721008301 + ], + "logprob_reward": -0.06524363160133362, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06524363160133362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14682462811470032, + 0.06297610700130463, + 0.030633289366960526, + 0.04605568200349808, + 0.025905432179570198, + 0.06719271093606949, + 0.3713887631893158, + 0.15241439640522003, + 0.0442812442779541, + 0.052327774465084076 + ], + "credit_map": [ + 0.038665469735860825, + 0.012517965398728848, + 0.010774325579404831, + 0.009057905524969101, + 0.013656516559422016, + 0.020694859325885773, + 0.013742444105446339, + 0.02771511860191822, + 0.012150107882916927, + 0.005813559051603079, + 0.005323709920048714, + 0.004984165076166391, + 0.0073823388665914536, + 0.008282672613859177, + 0.004532802850008011, + 0.014506745152175426, + 0.008481234312057495, + 0.002892969409003854, + 0.002185750287026167, + 0.0020447135902941227, + 0.002772680250927806, + 0.004688136279582977, + 0.0023592098150402308, + 0.005208590533584356, + 0.009225445799529552, + 0.004185276571661234, + 0.003997235558927059, + 0.005051218904554844, + 0.006171930115669966, + 0.005771773401647806, + 0.005943500902503729, + 0.0057093002833426, + 0.005808006040751934, + 0.0025256560184061527, + 0.0020996560342609882, + 0.002272726269438863, + 0.0034741044510155916, + 0.0036217253655195236, + 0.003280447795987129, + 0.002823108807206154, + 0.013982906937599182, + 0.0049850973300635815, + 0.006796291563659906, + 0.006000050343573093, + 0.006219102069735527, + 0.008137769997119904, + 0.008136143907904625, + 0.012935344129800797, + 0.07361258566379547, + 0.03355298936367035, + 0.03527093306183815, + 0.021084662526845932, + 0.04033925384283066, + 0.04893088713288307, + 0.02880163863301277, + 0.08979577571153641, + 0.02767960913479328, + 0.010326634161174297, + 0.016169428825378418, + 0.011755451560020447, + 0.014023962430655956, + 0.016277145594358444, + 0.01786455698311329, + 0.038317594677209854, + 0.014362664893269539, + 0.003042566357180476, + 0.004067968111485243, + 0.0029805826488882303, + 0.003977242391556501, + 0.00621740473434329, + 0.0032917093485593796, + 0.006341101136058569, + 0.015295672230422497, + 0.0029978526290506124, + 0.005800487007945776, + 0.0040114386938512325, + 0.004375731106847525, + 0.008025617338716984, + 0.00499270623549819, + 0.006828267127275467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.06524363160133362 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292383179068565 + ], + "opposite_logprob_tokens": [ + -3.773292303085327 + ], + "logprob_reward": -0.023292383179068565, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292383179068565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29023680090904236, + 0.07469254732131958, + 0.07860993593931198, + 0.04274629056453705, + 0.1027873232960701, + 0.06164681538939476, + 0.15934208035469055, + 0.06142641603946686, + 0.07652322947978973, + 0.051988646388053894 + ], + "credit_map": [ + 0.06609019637107849, + 0.038709692656993866, + 0.01894685998558998, + 0.02273901365697384, + 0.03981633111834526, + 0.04093244671821594, + 0.029986903071403503, + 0.03301529958844185, + 0.01199845876544714, + 0.012512867338955402, + 0.004165113437920809, + 0.006351636257022619, + 0.014982125721871853, + 0.009114402346313, + 0.008109434507787228, + 0.007458504289388657, + 0.009343840181827545, + 0.01560925878584385, + 0.0035060972440987825, + 0.008526226505637169, + 0.014718063175678253, + 0.008414402604103088, + 0.010210343636572361, + 0.00828169658780098, + 0.005928669590502977, + 0.005737568251788616, + 0.003234089585021138, + 0.005448522046208382, + 0.0074192979373037815, + 0.004787905141711235, + 0.005145158618688583, + 0.005045074503868818, + 0.016507316380739212, + 0.018287362530827522, + 0.005680567119270563, + 0.012767466716468334, + 0.01878024823963642, + 0.010294996201992035, + 0.010027976706624031, + 0.010441375896334648, + 0.008911238051950932, + 0.010395334102213383, + 0.0036611424293369055, + 0.007365405093878508, + 0.011508210562169552, + 0.006531002465635538, + 0.007305783685296774, + 0.00596869271248579, + 0.03228344768285751, + 0.01335927378386259, + 0.01517700869590044, + 0.008716128766536713, + 0.01680406741797924, + 0.0215381421148777, + 0.012515843845903873, + 0.03894815593957901, + 0.015125531703233719, + 0.0048333145678043365, + 0.004957418888807297, + 0.004916847683489323, + 0.008210727944970131, + 0.007463808637112379, + 0.007374866865575314, + 0.0085438909009099, + 0.014215925708413124, + 0.006164860911667347, + 0.0059428103268146515, + 0.008109809830784798, + 0.013873523101210594, + 0.008368234150111675, + 0.011639110743999481, + 0.008208943530917168, + 0.008099501952528954, + 0.005132054910063744, + 0.004717801231890917, + 0.006566270254552364, + 0.008205343969166279, + 0.006026966497302055, + 0.008557693101465702, + 0.004683006554841995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.023292383179068565 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486559122800827 + ], + "opposite_logprob_tokens": [ + -3.0486559867858887 + ], + "logprob_reward": -0.0486559122800827, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486559122800827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16002961993217468, + 0.08324114978313446, + 0.04555504769086838, + 0.04842064529657364, + 0.09482519328594208, + 0.10889440029859543, + 0.3217408061027527, + 0.03945346921682358, + 0.05335742235183716, + 0.04448221996426582 + ], + "credit_map": [ + 0.0428864061832428, + 0.015408480539917946, + 0.013969029299914837, + 0.009593438357114792, + 0.020563727244734764, + 0.024295954033732414, + 0.012707571499049664, + 0.02060500718653202, + 0.01765194721519947, + 0.011461934074759483, + 0.004288828000426292, + 0.006137289106845856, + 0.011669258587062359, + 0.011997497640550137, + 0.00885679665952921, + 0.011177599430084229, + 0.006938657257705927, + 0.005265714135020971, + 0.0024400437250733376, + 0.0038307548966258764, + 0.006187513004988432, + 0.005881723947823048, + 0.007571273483335972, + 0.0074393670074641705, + 0.006979478057473898, + 0.00573498010635376, + 0.004062911029905081, + 0.004866556264460087, + 0.009568993002176285, + 0.00503112655133009, + 0.005527552217245102, + 0.006649045739322901, + 0.016151290386915207, + 0.011895195581018925, + 0.006245909724384546, + 0.009368552826344967, + 0.01527357567101717, + 0.012153329327702522, + 0.012784640304744244, + 0.010952702723443508, + 0.020555943250656128, + 0.011165143921971321, + 0.00973396748304367, + 0.007942475378513336, + 0.015587077476084232, + 0.013172832317650318, + 0.008511596359312534, + 0.022225353866815567, + 0.0481739267706871, + 0.03146907314658165, + 0.030747924000024796, + 0.023280639201402664, + 0.0407196469604969, + 0.03599129244685173, + 0.029416197910904884, + 0.0819421112537384, + 0.009073397144675255, + 0.004810040816664696, + 0.00381876970641315, + 0.003712527686730027, + 0.0037643678952008486, + 0.004451812710613012, + 0.003411011304706335, + 0.00641154358163476, + 0.01626463793218136, + 0.003109914716333151, + 0.00512304762378335, + 0.003526005893945694, + 0.004070208873599768, + 0.0081712007522583, + 0.004879424814134836, + 0.008212978951632977, + 0.011262523010373116, + 0.002487520221620798, + 0.005174524150788784, + 0.003218647325411439, + 0.003478935919702053, + 0.006635191384702921, + 0.0034246742725372314, + 0.008800206705927849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.0486559122800827 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294713348150253 + ], + "opposite_logprob_tokens": [ + -3.773294687271118 + ], + "logprob_reward": -0.023294713348150253, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294713348150253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1780339628458023, + 0.09567076712846756, + 0.047739215195178986, + 0.06768926233053207, + 0.07581601291894913, + 0.049092039465904236, + 0.0808425024151802, + 0.23734256625175476, + 0.041139282286167145, + 0.1266343593597412 + ], + "credit_map": [ + 0.04224732890725136, + 0.02539093792438507, + 0.013035745359957218, + 0.012941149063408375, + 0.020554661750793457, + 0.023494550958275795, + 0.01815561018884182, + 0.02221398986876011, + 0.015970515087246895, + 0.015430239960551262, + 0.006827118340879679, + 0.007920919917523861, + 0.01608331874012947, + 0.010919509455561638, + 0.010390044189989567, + 0.012129097245633602, + 0.00953403115272522, + 0.006477664224803448, + 0.0028840594459325075, + 0.004032460507005453, + 0.00611843029037118, + 0.007679734379053116, + 0.0056671686470508575, + 0.00534566817805171, + 0.009576567448675632, + 0.008529004640877247, + 0.00532789109274745, + 0.007022498641163111, + 0.011720997281372547, + 0.007485619280487299, + 0.006081826984882355, + 0.011944862082600594, + 0.012923377566039562, + 0.009195270016789436, + 0.003986924421042204, + 0.007846865803003311, + 0.013818439096212387, + 0.010256219655275345, + 0.009190813638269901, + 0.008598103187978268, + 0.007979711517691612, + 0.006861583795398474, + 0.003956569824367762, + 0.0056243897415697575, + 0.007683673873543739, + 0.005999644752591848, + 0.005592616740614176, + 0.005393849685788155, + 0.016043340787291527, + 0.007272044662386179, + 0.007433666847646236, + 0.005689080338925123, + 0.009843210689723492, + 0.010536149144172668, + 0.006636274978518486, + 0.017388733103871346, + 0.038873568177223206, + 0.022025713697075844, + 0.02563878893852234, + 0.018458154052495956, + 0.021033771336078644, + 0.02436281368136406, + 0.026419540867209435, + 0.06053023040294647, + 0.008796724490821362, + 0.0040334113873541355, + 0.003642162773758173, + 0.0038960312958806753, + 0.006612847559154034, + 0.00528245372697711, + 0.0039025512523949146, + 0.0049731009639799595, + 0.02904582768678665, + 0.012554386630654335, + 0.012822885997593403, + 0.01034003309905529, + 0.013902167789638042, + 0.016482755541801453, + 0.009408964775502682, + 0.022077331319451332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.023294713348150253 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155581831932068 + ], + "opposite_logprob_tokens": [ + -1.701555848121643 + ], + "logprob_reward": -0.20155581831932068, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155581831932068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08754455298185349, + 0.05011714622378349, + 0.04441056028008461, + 0.04537767916917801, + 0.027022315189242363, + 0.02314780279994011, + 0.27590620517730713, + 0.31574228405952454, + 0.062232013791799545, + 0.06849943101406097 + ], + "credit_map": [ + 0.01763453334569931, + 0.009891446679830551, + 0.008582686074078083, + 0.007821341045200825, + 0.010896701365709305, + 0.0123412711545825, + 0.007968183606863022, + 0.012408395297825336, + 0.01767701469361782, + 0.00486369151622057, + 0.0038148476742208004, + 0.0024393706116825342, + 0.0037468138616532087, + 0.008455036208033562, + 0.0025834282860159874, + 0.0065369377844035625, + 0.009462540037930012, + 0.005157814361155033, + 0.0029646919574588537, + 0.003330470761284232, + 0.005565848667174578, + 0.008691702038049698, + 0.004339898005127907, + 0.004897594451904297, + 0.010250793769955635, + 0.004776382818818092, + 0.00388070335611701, + 0.0035446928814053535, + 0.00684721814468503, + 0.006123907398432493, + 0.0035086467396467924, + 0.006445334758609533, + 0.004856893792748451, + 0.0031265877187252045, + 0.001839127391576767, + 0.003073886502534151, + 0.004527517128735781, + 0.00331156887114048, + 0.0031437177676707506, + 0.0031430150847882032, + 0.00409413268789649, + 0.0025527835823595524, + 0.0015420172130689025, + 0.002419818192720413, + 0.0037394126411527395, + 0.003428039140999317, + 0.002177794463932514, + 0.003193807089701295, + 0.04388689622282982, + 0.026111042127013206, + 0.025226499885320663, + 0.019702007994055748, + 0.03369952738285065, + 0.0329127311706543, + 0.025029068812727928, + 0.06933844834566116, + 0.048330485820770264, + 0.02672845497727394, + 0.03361259400844574, + 0.024078993126749992, + 0.030212486162781715, + 0.03243894502520561, + 0.03598222881555557, + 0.0843580886721611, + 0.019275397062301636, + 0.004522210918366909, + 0.005628679413348436, + 0.0052709938026964664, + 0.006704781670123339, + 0.009264402091503143, + 0.004971634596586227, + 0.006593913305550814, + 0.016195107251405716, + 0.006285948678851128, + 0.007510855328291655, + 0.006050935480743647, + 0.0067995511926710606, + 0.009244280867278576, + 0.0049446444027125835, + 0.011468115262687206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.20155581831932068 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly inching inward, transports the viewer deep into a lush, light-dappled bamboo forest. A woman, clad in a simple white linen robe, her cascading emerald green hair as vibrant as the tender shoots of early spring, soft and lustrous. Her graceful figure forms a gentle curve, forming a graceful S-shaped composition alongside the shimmering, clear stream that meanders beside her. Her face, clear and beautiful, her expression serene and focused, grips a dark, textured wooden comb. Starting from the roots, she meticulously and deliberately combs through her hair, each stroke of the comb teeth creating a tranquil, ritualistic dialogue. The afternoon sun filters through the gaps between the towering bamboo leaves, casting countless shattered, golden specks of light across her white robe, her green hair, and the moist earth, creating a dreamlike effect. The entire scene is filled with an ethereal and poetic atmosphere, presenting the exquisite beauty unique to film animation, full of the harmonious beauty of oriental Zen and nature.", + "seed": 12740, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_3/index_1630_seed_12740_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03147066873498261, + "vlm_logprob": -0.03147066873498261, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.03147066873498261, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01006657350808382 + ], + "opposite_logprob_tokens": [ + -5.135066509246826 + ], + "logprob_reward": -0.01006657350808382, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01006657350808382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16730213165283203, + 0.07303670793771744, + 0.3389914929866791, + 0.038066472858190536, + 0.0610235370695591, + 0.04340917244553566, + 0.03177575767040253, + 0.07540182024240494, + 0.04933746159076691, + 0.1216554269194603 + ], + "credit_map": [ + 0.025312863290309906, + 0.017036309465765953, + 0.027991417795419693, + 0.012901045382022858, + 0.013435968197882175, + 0.027936648577451706, + 0.010830609127879143, + 0.03185727819800377, + 0.009425506927073002, + 0.008338911458849907, + 0.00850937981158495, + 0.00590304471552372, + 0.009693153202533722, + 0.014030341990292072, + 0.0052414340898394585, + 0.011894947849214077, + 0.0936349481344223, + 0.015269966796040535, + 0.020370125770568848, + 0.014660531654953957, + 0.013356903567910194, + 0.12432751804590225, + 0.010252979584038258, + 0.04711855202913284, + 0.005972659680992365, + 0.004828596021980047, + 0.004251478705555201, + 0.005027442239224911, + 0.00469565112143755, + 0.006252910476177931, + 0.002958127297461033, + 0.004079611971974373, + 0.006953481584787369, + 0.010432306677103043, + 0.007509710732847452, + 0.006077987141907215, + 0.008476834744215012, + 0.01002753246575594, + 0.004123459570109844, + 0.007422234863042831, + 0.005785835441201925, + 0.006411496549844742, + 0.004905497655272484, + 0.005840688478201628, + 0.00532189104706049, + 0.00638424139469862, + 0.0030413037165999413, + 0.005718223284929991, + 0.0044303350150585175, + 0.0037568684201687574, + 0.003908507991582155, + 0.003099186811596155, + 0.005250242538750172, + 0.0042562480084598064, + 0.0025869186501950026, + 0.00448745209723711, + 0.009352042339742184, + 0.00800646934658289, + 0.010462775826454163, + 0.008643455803394318, + 0.013009756803512573, + 0.011575202457606792, + 0.0064827632158994675, + 0.007869360037147999, + 0.006920618005096912, + 0.00436308654025197, + 0.006495035719126463, + 0.006241567898541689, + 0.007847721688449383, + 0.006874759681522846, + 0.004896975122392178, + 0.005697702523320913, + 0.015053292736411095, + 0.008408874273300171, + 0.02706894651055336, + 0.009700612165033817, + 0.021054940298199654, + 0.009306889958679676, + 0.015046819113194942, + 0.016015058383345604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.01006657350808382 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002203538315370679 + ], + "opposite_logprob_tokens": [ + -6.127203464508057 + ], + "logprob_reward": -0.002203538315370679, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002203538315370679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22353145480155945, + 0.12646178901195526, + 0.22329643368721008, + 0.03866667300462723, + 0.10052786767482758, + 0.023595303297042847, + 0.06166063994169235, + 0.0954681858420372, + 0.03312315791845322, + 0.07366843521595001 + ], + "credit_map": [ + 0.02665713056921959, + 0.030035007745027542, + 0.030415037646889687, + 0.030285466462373734, + 0.018704302608966827, + 0.035629626363515854, + 0.01701190136373043, + 0.03479300066828728, + 0.013082033954560757, + 0.01988324709236622, + 0.014364814385771751, + 0.018633771687746048, + 0.01157221570611, + 0.017778202891349792, + 0.01241725031286478, + 0.018730249255895615, + 0.05480750650167465, + 0.014382450841367245, + 0.014126989990472794, + 0.01459705363959074, + 0.010409393347799778, + 0.07639748603105545, + 0.007904397323727608, + 0.030671169981360435, + 0.005445373710244894, + 0.008390099741518497, + 0.004037823062390089, + 0.00622315239161253, + 0.0039549763314425945, + 0.003876775037497282, + 0.0027583115734159946, + 0.003980167210102081, + 0.010761749930679798, + 0.02654058113694191, + 0.009273255243897438, + 0.018308283761143684, + 0.010009730234742165, + 0.00959810335189104, + 0.008861955255270004, + 0.007174226455390453, + 0.002757444279268384, + 0.005338462069630623, + 0.0026226574555039406, + 0.0027795378118753433, + 0.0035527076106518507, + 0.0022270060144364834, + 0.0020720260217785835, + 0.002245464362204075, + 0.008158409968018532, + 0.008685911074280739, + 0.007554426323622465, + 0.012035194784402847, + 0.007028455380350351, + 0.0050246259197592735, + 0.006819067522883415, + 0.006354550831019878, + 0.01063959114253521, + 0.020262256264686584, + 0.007645801175385714, + 0.026259463280439377, + 0.009838968515396118, + 0.005599671043455601, + 0.009424727410078049, + 0.005797720514237881, + 0.004450574517250061, + 0.006513131782412529, + 0.003884401870891452, + 0.006096903700381517, + 0.003497075755149126, + 0.0025905503425747156, + 0.0030273846350610256, + 0.00306313787586987, + 0.00923109706491232, + 0.0062554641626775265, + 0.015332494862377644, + 0.0070465668104588985, + 0.010768350213766098, + 0.00619420874863863, + 0.009565821848809719, + 0.009274437092244625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.002203538315370679 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077335104346275 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.011077335104346275, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077335104346275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10176702588796616, + 0.07460157573223114, + 0.24813759326934814, + 0.057327818125486374, + 0.37035390734672546, + 0.010971145704388618, + 0.040066417306661606, + 0.029203444719314575, + 0.031861815601587296, + 0.035709284245967865 + ], + "credit_map": [ + 0.013049348257482052, + 0.013650244101881981, + 0.01483181118965149, + 0.009666061960160732, + 0.008617511950433254, + 0.0179933812469244, + 0.006745937746018171, + 0.017212731763720512, + 0.008533990941941738, + 0.009961996227502823, + 0.00856841541826725, + 0.008898590691387653, + 0.006617067847400904, + 0.013727860525250435, + 0.005342212505638599, + 0.012951438315212727, + 0.059649839997291565, + 0.016481708735227585, + 0.014863583259284496, + 0.014613984152674675, + 0.010812130756676197, + 0.08641088008880615, + 0.008756307885050774, + 0.036549169570207596, + 0.009555760771036148, + 0.013474255800247192, + 0.004631430841982365, + 0.00932303722947836, + 0.00524154445156455, + 0.006011331919580698, + 0.004254947882145643, + 0.004835507366806269, + 0.047548551112413406, + 0.09308566898107529, + 0.03527311608195305, + 0.06341990828514099, + 0.04228220507502556, + 0.031218230724334717, + 0.032958198338747025, + 0.024568021297454834, + 0.00137621839530766, + 0.0016082023503258824, + 0.001735833240672946, + 0.0020891516469419003, + 0.0012821056880056858, + 0.0009679425274953246, + 0.0008761801873333752, + 0.0010355116100981832, + 0.005667459685355425, + 0.006198922637850046, + 0.0045437938533723354, + 0.007308232132345438, + 0.004683454986661673, + 0.0033085113391280174, + 0.004150389228016138, + 0.004205652512609959, + 0.0038323518820106983, + 0.005056567955762148, + 0.002569987438619137, + 0.007953673601150513, + 0.0031554456800222397, + 0.0020259078592061996, + 0.002845475682988763, + 0.0017640349688008428, + 0.00473569193854928, + 0.006596374325454235, + 0.003657470690086484, + 0.006692645605653524, + 0.003063429379835725, + 0.002138820942491293, + 0.002831944264471531, + 0.0021454356610774994, + 0.004306196700781584, + 0.00703738396987319, + 0.005719038192182779, + 0.005734903737902641, + 0.004414830356836319, + 0.002590497024357319, + 0.0034776325337588787, + 0.002428801031783223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.011077335104346275 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033702366054058075 + ], + "opposite_logprob_tokens": [ + -3.4087023735046387 + ], + "logprob_reward": -0.033702366054058075, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033702366054058075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0957154631614685, + 0.053607285022735596, + 0.11395511776208878, + 0.0538574755191803, + 0.0674055814743042, + 0.044435445219278336, + 0.06102241575717926, + 0.28916803002357483, + 0.1868077665567398, + 0.03402537852525711 + ], + "credit_map": [ + 0.00964275747537613, + 0.012407983653247356, + 0.013256264850497246, + 0.010254339314997196, + 0.006878392770886421, + 0.014224426820874214, + 0.009430386126041412, + 0.01962091401219368, + 0.005798278842121363, + 0.008921557106077671, + 0.005044826306402683, + 0.008420851081609726, + 0.005258176475763321, + 0.006032262928783894, + 0.007089192513376474, + 0.007042150944471359, + 0.02577797882258892, + 0.014162604697048664, + 0.008675812743604183, + 0.012104211375117302, + 0.0068833548575639725, + 0.029209811240434647, + 0.0054841721430420876, + 0.011657190509140491, + 0.007118915673345327, + 0.009335051290690899, + 0.006275140214711428, + 0.00925306137651205, + 0.004511428065598011, + 0.0061470093205571175, + 0.004585928283631802, + 0.006630948279052973, + 0.011036948300898075, + 0.014604618772864342, + 0.005766242276877165, + 0.011239335872232914, + 0.0073552438989281654, + 0.006718339864164591, + 0.00482427841052413, + 0.005860584322363138, + 0.006902816239744425, + 0.006230940576642752, + 0.004202198702841997, + 0.009581068530678749, + 0.004927057307213545, + 0.0038641570135951042, + 0.004904345143586397, + 0.0038228658959269524, + 0.008590303361415863, + 0.012406430207192898, + 0.0072181690484285355, + 0.009525395929813385, + 0.006709436420351267, + 0.004906987305730581, + 0.006783221382647753, + 0.004882479086518288, + 0.03292340785264969, + 0.05474036931991577, + 0.02530825510621071, + 0.08136402815580368, + 0.028928864747285843, + 0.016903258860111237, + 0.029512453824281693, + 0.01948740892112255, + 0.02831897884607315, + 0.03976978361606598, + 0.021078821271657944, + 0.040610212832689285, + 0.017796235159039497, + 0.011693772859871387, + 0.016938425600528717, + 0.010601562447845936, + 0.004767092876136303, + 0.0052021644078195095, + 0.004332046024501324, + 0.005153309553861618, + 0.003735755570232868, + 0.002858328865841031, + 0.004324628040194511, + 0.0036520566791296005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.033702366054058075 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with light filtering through the leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a white linen robe visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have cascading emerald green hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding a dark, textured wooden comb and using it to comb her hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle and slowly inching inward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003035306930542 + ], + "opposite_logprob_tokens": [ + -2.3503036499023438 + ], + "logprob_reward": -0.1003035306930542, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003035306930542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1528945118188858, + 0.07099464535713196, + 0.331755131483078, + 0.05433773621916771, + 0.0452202670276165, + 0.05295417085289955, + 0.04342673346400261, + 0.14451277256011963, + 0.05043137073516846, + 0.05347271263599396 + ], + "credit_map": [ + 0.013615034520626068, + 0.022148067131638527, + 0.025804929435253143, + 0.021397437900304794, + 0.011981223709881306, + 0.023875800892710686, + 0.011950827203691006, + 0.02212117239832878, + 0.008740900084376335, + 0.007999821566045284, + 0.009362035430967808, + 0.010445334017276764, + 0.008601099252700806, + 0.010107695125043392, + 0.006029489450156689, + 0.009708268567919731, + 0.08328163623809814, + 0.026815172284841537, + 0.019991427659988403, + 0.025829071179032326, + 0.017664648592472076, + 0.10687952488660812, + 0.014259502291679382, + 0.03703412413597107, + 0.00719649950042367, + 0.011239263229072094, + 0.005218230653554201, + 0.010031105019152164, + 0.004573178011924028, + 0.0062140668742358685, + 0.0046798791736364365, + 0.0051855058409273624, + 0.0062067084945738316, + 0.009265908971428871, + 0.004294208250939846, + 0.009096680209040642, + 0.004391550086438656, + 0.004065749701112509, + 0.004096951335668564, + 0.0038025062531232834, + 0.006748217158019543, + 0.009926836006343365, + 0.005591589026153088, + 0.012375888414680958, + 0.0056547243148088455, + 0.004030479583889246, + 0.0046501061879098415, + 0.003976323176175356, + 0.007376690860837698, + 0.008595714345574379, + 0.005546709056943655, + 0.008438232354819775, + 0.0041945381090044975, + 0.0028370097279548645, + 0.0037587848491966724, + 0.002679049503058195, + 0.02434489130973816, + 0.029422283172607422, + 0.015974827110767365, + 0.03488488867878914, + 0.01115516759455204, + 0.007462394889444113, + 0.012626194395124912, + 0.008642121218144894, + 0.007065830752253532, + 0.009070887230336666, + 0.005127535201609135, + 0.006024025846272707, + 0.0065048933029174805, + 0.005261365789920092, + 0.005887097679078579, + 0.005489733070135117, + 0.006397583521902561, + 0.005295688286423683, + 0.008609662763774395, + 0.005871289409697056, + 0.010103058069944382, + 0.004487553145736456, + 0.006155013106763363, + 0.0065528629347682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.1003035306930542 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_4/index_tensor([136, 136])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_4/index_tensor([136, 136])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..686cdf429cc0e4be9488f870bdcc37741a8585bf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_4/index_tensor([136, 136])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium shot, held at a slightly high angle, captures a minimalist, polished dark mahogany desk, its surface a flawlessly reflective plane within a sleek, techno-futuristic room. A vast floor-to-ceiling window in the background reveals a vast cityscape at dusk, towering buildings aglow with faint streaks of neon and aircraft. The transformation subtly begins as subtle cracks of bright blue light trace geometric patterns across the desk's mirror-like surface. These lines rapidly and decisively widen as the wooden top deconstructs, its panels sliding and retracting inward to reveal a breathtakingly complex interior. Beneath, a dense network of cyan-glowing circuits pulses with energy, casting cool electric light upward, while a complex polished chrome machine—a symphony of tiny pistons, interlocking gears, and articulated arms—begins to steadily and deliberately reconfigure the desk's innards. Shot at a high frame rate for hyper-real, crystal clarity, the footage captures every minute shift in the mechanisms and the crisp, dancing reflections on the moving chrome and wood surfaces, creating an atmosphere of lifeless awe and technological wonder, focused entirely on the controlled, mesmerizing deformations.", + "seed": 10136, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_4/index_136_seed_10136_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02820061766542494, + "vlm_logprob": -0.02820061766542494, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.02820061766542494, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025837622582912445 + ], + "opposite_logprob_tokens": [ + -4.150837421417236 + ], + "logprob_reward": -0.025837622582912445, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.025837622582912445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34022098779678345, + 0.1407174915075302, + 0.0562606118619442, + 0.05911575257778168, + 0.04612104967236519, + 0.06015956774353981, + 0.04144555330276489, + 0.05365093797445297, + 0.12367330491542816, + 0.07863477617502213 + ], + "credit_map": [ + 0.03597185015678406, + 0.021160030737519264, + 0.026841510087251663, + 0.019768385216593742, + 0.05678492411971092, + 0.07107710838317871, + 0.054888490587472916, + 0.053728677332401276, + 0.011985961347818375, + 0.005933872424066067, + 0.01311233825981617, + 0.009918587282299995, + 0.04331490397453308, + 0.011584979481995106, + 0.0342998132109642, + 0.010567025281488895, + 0.006170616950839758, + 0.004671128932386637, + 0.007245604880154133, + 0.007539416663348675, + 0.005373379215598106, + 0.005777820944786072, + 0.01035017054527998, + 0.009132467210292816, + 0.008053942583501339, + 0.006817550398409367, + 0.01234457828104496, + 0.004979799967259169, + 0.005993078928440809, + 0.0060010249726474285, + 0.005983759183436632, + 0.00894201546907425, + 0.01029286440461874, + 0.006477825343608856, + 0.0034900528844445944, + 0.003254103474318981, + 0.007062978111207485, + 0.006888518109917641, + 0.003668308025225997, + 0.004986393731087446, + 0.009153278544545174, + 0.007208621129393578, + 0.0069344425573945045, + 0.009306661784648895, + 0.005616363137960434, + 0.006441914942115545, + 0.006778656970709562, + 0.008719627745449543, + 0.005222161766141653, + 0.004998557269573212, + 0.004111264366656542, + 0.005248323082923889, + 0.005423819646239281, + 0.003955962136387825, + 0.007990091107785702, + 0.004495367873460054, + 0.006971563678234816, + 0.011353878304362297, + 0.004561610985547304, + 0.005757693201303482, + 0.007984533905982971, + 0.0059373388066887856, + 0.006156483665108681, + 0.004927830304950476, + 0.009621996432542801, + 0.006455373018980026, + 0.008755655027925968, + 0.007501637563109398, + 0.014981728047132492, + 0.02540740557014942, + 0.025769731029868126, + 0.025179771706461906, + 0.01791319064795971, + 0.009471855126321316, + 0.0073249442502856255, + 0.007474485784769058, + 0.010539318434894085, + 0.009180293418467045, + 0.00680925277993083, + 0.009921429678797722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.025837622582912445 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868203029036522 + ], + "opposite_logprob_tokens": [ + -3.0486819744110107 + ], + "logprob_reward": -0.04868203029036522, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868203029036522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1966128647327423, + 0.45481225848197937, + 0.08595255762338638, + 0.05282619968056679, + 0.05297355726361275, + 0.022376103326678276, + 0.008794966153800488, + 0.02616751566529274, + 0.06727782636880875, + 0.03220615163445473 + ], + "credit_map": [ + 0.013786698691546917, + 0.007965818047523499, + 0.014484483748674393, + 0.010871324688196182, + 0.044647879898548126, + 0.02068633772432804, + 0.06506838649511337, + 0.019101938232779503, + 0.028149712830781937, + 0.011822273954749107, + 0.04316064715385437, + 0.025857191532850266, + 0.11173080652952194, + 0.03189944103360176, + 0.1627233326435089, + 0.03946882113814354, + 0.0068271467462182045, + 0.00636294623836875, + 0.010875648818910122, + 0.009858386591076851, + 0.007622074335813522, + 0.00820633489638567, + 0.018981752917170525, + 0.01721825823187828, + 0.005902755539864302, + 0.0050460342317819595, + 0.011854497715830803, + 0.004884948953986168, + 0.003930225502699614, + 0.005094326101243496, + 0.007235979661345482, + 0.00887742917984724, + 0.014386716298758984, + 0.006787779275327921, + 0.003214338095858693, + 0.004257007967680693, + 0.006684343796223402, + 0.005156001541763544, + 0.0048807053826749325, + 0.007606661878526211, + 0.0029741977341473103, + 0.002559968736022711, + 0.002523517468944192, + 0.004306042566895485, + 0.002533125225454569, + 0.0026486904826015234, + 0.0022366901393979788, + 0.0025938700418919325, + 0.0016965966206043959, + 0.0012163548963144422, + 0.0010028774850070477, + 0.0008915217476896942, + 0.0008457028307020664, + 0.0008615378173999488, + 0.0012265214463695884, + 0.0010538524948060513, + 0.002998405834659934, + 0.0027678627520799637, + 0.0035950029268860817, + 0.0033298672642558813, + 0.0050731441006064415, + 0.0024774358607828617, + 0.003195259952917695, + 0.0027305372059345245, + 0.011622506193816662, + 0.008113247342407703, + 0.004640755243599415, + 0.004578899592161179, + 0.00939265824854374, + 0.006075605750083923, + 0.007645913865417242, + 0.015208244323730469, + 0.004813037347048521, + 0.004254802130162716, + 0.004348231479525566, + 0.0034739451948553324, + 0.0034965132363140583, + 0.002856848994269967, + 0.0037987257819622755, + 0.005164043512195349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.04868203029036522 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200353994965553 + ], + "opposite_logprob_tokens": [ + -4.264200210571289 + ], + "logprob_reward": -0.014200353994965553, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200353994965553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15981636941432953, + 0.3452211320400238, + 0.07328102737665176, + 0.22036123275756836, + 0.01861199177801609, + 0.036277540028095245, + 0.020451916381716728, + 0.03555137291550636, + 0.05479840189218521, + 0.0356290228664875 + ], + "credit_map": [ + 0.010378542356193066, + 0.01218497660011053, + 0.012770009227097034, + 0.011024017818272114, + 0.03324295952916145, + 0.011473028920590878, + 0.05218840390443802, + 0.016554441303014755, + 0.023042909801006317, + 0.009288287721574306, + 0.020703140646219254, + 0.017337288707494736, + 0.10609299689531326, + 0.0270466860383749, + 0.11802974343299866, + 0.023680103942751884, + 0.009924748912453651, + 0.008168363943696022, + 0.008800096809864044, + 0.007078559137880802, + 0.01004661526530981, + 0.011865521781146526, + 0.00758188683539629, + 0.009815243072807789, + 0.030611416324973106, + 0.03813339024782181, + 0.023269323632121086, + 0.024199867621064186, + 0.029873495921492577, + 0.02329149842262268, + 0.0228976309299469, + 0.02808462828397751, + 0.0031357689294964075, + 0.002062407322227955, + 0.00241589336656034, + 0.001807456836104393, + 0.002571289660409093, + 0.0022464096546173096, + 0.0021670798305422068, + 0.0022056892048567533, + 0.004834811203181744, + 0.004793120082467794, + 0.004231435712426901, + 0.004570258315652609, + 0.004844615235924721, + 0.0037234649062156677, + 0.004873811732977629, + 0.004406026564538479, + 0.002140074037015438, + 0.0013748562196269631, + 0.0022370831575244665, + 0.002132162218913436, + 0.00224491348490119, + 0.0025445029605180025, + 0.0047401911579072475, + 0.003038134193047881, + 0.00413131806999445, + 0.005954206921160221, + 0.0032329310197383165, + 0.004274794831871986, + 0.005423168186098337, + 0.003195559373125434, + 0.005654770415276289, + 0.003684622934088111, + 0.0060416217893362045, + 0.0037194809410721064, + 0.0033282616641372442, + 0.0039019479881972075, + 0.0043358332477509975, + 0.004378809127956629, + 0.017842723056674004, + 0.011249728500843048, + 0.006393684074282646, + 0.004216092173010111, + 0.005207930691540241, + 0.003426256123930216, + 0.004394706338644028, + 0.003835764480754733, + 0.003063312266021967, + 0.005091279279440641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.014200353994965553 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036171742249280214 + ], + "opposite_logprob_tokens": [ + -5.628617286682129 + ], + "logprob_reward": -0.0036171742249280214, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036171742249280214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22578497231006622, + 0.23079949617385864, + 0.10058163851499557, + 0.08591574430465698, + 0.04715251177549362, + 0.03269767388701439, + 0.0942692756652832, + 0.028126178309321404, + 0.11565368622541428, + 0.03901878371834755 + ], + "credit_map": [ + 0.016034943982958794, + 0.010496187955141068, + 0.017515748739242554, + 0.013774499297142029, + 0.04708985984325409, + 0.01942257210612297, + 0.0789930447936058, + 0.02245810441672802, + 0.01478125061839819, + 0.006460622418671846, + 0.02179625630378723, + 0.013235989026725292, + 0.05626840144395828, + 0.015218821354210377, + 0.08334406465291977, + 0.019694099202752113, + 0.007117677014321089, + 0.005797437392175198, + 0.0177015233784914, + 0.014831015840172768, + 0.008245039731264114, + 0.008069935254752636, + 0.02205360122025013, + 0.016765406355261803, + 0.009419629350304604, + 0.00975620374083519, + 0.013654517941176891, + 0.008967178873717785, + 0.006773937959223986, + 0.009430645033717155, + 0.012275375425815582, + 0.015638254582881927, + 0.008026201277971268, + 0.007268568966537714, + 0.00367896375246346, + 0.003145434893667698, + 0.0056893653236329556, + 0.00559714250266552, + 0.008153416216373444, + 0.005593417678028345, + 0.004158498253673315, + 0.003962255083024502, + 0.00406264653429389, + 0.005045405123382807, + 0.003653445281088352, + 0.003952755592763424, + 0.004300752189010382, + 0.0035619158297777176, + 0.014469528570771217, + 0.008520347066223621, + 0.0089770732447505, + 0.010172090493142605, + 0.0120849022641778, + 0.007322240620851517, + 0.023358022794127464, + 0.009365066885948181, + 0.004153144545853138, + 0.0030869508627802134, + 0.002241461770609021, + 0.0031061910558491945, + 0.005341371521353722, + 0.0035439340863376856, + 0.0044502876698970795, + 0.002202836098149419, + 0.01223765965551138, + 0.008859006687998772, + 0.006077112630009651, + 0.007982230745255947, + 0.008461887016892433, + 0.00968761183321476, + 0.035624563694000244, + 0.02672361582517624, + 0.008653574623167515, + 0.004274685867130756, + 0.004278587643057108, + 0.0033943266607820988, + 0.0037689926102757454, + 0.006314977537840605, + 0.003558924188837409, + 0.0047747124917805195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0036171742249280214 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048665907233953476 + ], + "opposite_logprob_tokens": [ + -3.048666000366211 + ], + "logprob_reward": -0.048665907233953476, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048665907233953476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11147116869688034, + 0.0898730456829071, + 0.059645313769578934, + 0.08194809406995773, + 0.040892232209444046, + 0.030291691422462463, + 0.12299943715333939, + 0.08301457017660141, + 0.3081001937389374, + 0.0717642530798912 + ], + "credit_map": [ + 0.015246765688061714, + 0.0077239093370735645, + 0.010288660414516926, + 0.009157533757388592, + 0.01943291164934635, + 0.023538896813988686, + 0.014960086904466152, + 0.011122404597699642, + 0.008290230296552181, + 0.0035924979019910097, + 0.005720722489058971, + 0.00490688718855381, + 0.02847464755177498, + 0.009798257611691952, + 0.02158941887319088, + 0.0075003840029239655, + 0.005877210292965174, + 0.005286279134452343, + 0.00633667316287756, + 0.008082272484898567, + 0.006530392915010452, + 0.005613178014755249, + 0.0123736672103405, + 0.009545647539198399, + 0.00827463436871767, + 0.007569729816168547, + 0.016484247520565987, + 0.008328008465468884, + 0.0072807022370398045, + 0.006830214988440275, + 0.01203572005033493, + 0.01514484267681837, + 0.006513056345283985, + 0.004575306084007025, + 0.004774464759975672, + 0.002943722065538168, + 0.005469998344779015, + 0.005221514962613583, + 0.004551277495920658, + 0.006842894479632378, + 0.00390777038410306, + 0.002499083988368511, + 0.0038159280084073544, + 0.0040389010682702065, + 0.0031130879651755095, + 0.003945530392229557, + 0.004753553308546543, + 0.004217837937176228, + 0.020090918987989426, + 0.0105735519900918, + 0.012086722068488598, + 0.011509668081998825, + 0.013464122079312801, + 0.012188032269477844, + 0.032568469643592834, + 0.010517946444451809, + 0.009787053801119328, + 0.015796367079019547, + 0.006860715337097645, + 0.0071702878922224045, + 0.009042706340551376, + 0.01554208341985941, + 0.010921286419034004, + 0.0078940624371171, + 0.028155986219644547, + 0.02431802824139595, + 0.02279409021139145, + 0.025546181946992874, + 0.026066947728395462, + 0.025901440531015396, + 0.09310704469680786, + 0.06221047416329384, + 0.011334475129842758, + 0.008480138145387173, + 0.012657511048018932, + 0.008162769488990307, + 0.007440956309437752, + 0.007979621179401875, + 0.006237656343728304, + 0.009471125900745392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.048665907233953476 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, held at a slightly high angle, captures a minimalist, polished dark mahogany desk, its surface a flawlessly reflective plane within a sleek, techno-futuristic room. A vast floor-to-ceiling window in the background reveals a vast cityscape at dusk, towering buildings aglow with faint streaks of neon and aircraft. The transformation subtly begins as subtle cracks of bright blue light trace geometric patterns across the desk's mirror-like surface. These lines rapidly and decisively widen as the wooden top deconstructs, its panels sliding and retracting inward to reveal a breathtakingly complex interior. Beneath, a dense network of cyan-glowing circuits pulses with energy, casting cool electric light upward, while a complex polished chrome machine—a symphony of tiny pistons, interlocking gears, and articulated arms—begins to steadily and deliberately reconfigure the desk's innards. Shot at a high frame rate for hyper-real, crystal clarity, the footage captures every minute shift in the mechanisms and the crisp, dancing reflections on the moving chrome and wood surfaces, creating an atmosphere of lifeless awe and technological wonder, focused entirely on the controlled, mesmerizing deformations.", + "seed": 10137, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_4/index_136_seed_10137_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05083841439336538, + "vlm_logprob": -0.05083841439336538, + "vlm_yes_no_margin": 3.674999952316284, + "vlm_token_credit": -0.05083841439336538, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07284796983003616 + ], + "opposite_logprob_tokens": [ + -4.072847843170166 + ], + "logprob_reward": -0.07284796983003616, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.07284796983003616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39205697178840637, + 0.06289932876825333, + 0.04221836104989052, + 0.02941819094121456, + 0.15636411309242249, + 0.06491320580244064, + 0.11319371312856674, + 0.028923438861966133, + 0.08123718947172165, + 0.02877548150718212 + ], + "credit_map": [ + 0.03656701743602753, + 0.021758269518613815, + 0.12174707651138306, + 0.0659686028957367, + 0.016483360901474953, + 0.021270278841257095, + 0.03522735834121704, + 0.07303499430418015, + 0.0061865998432040215, + 0.005627600010484457, + 0.008428050205111504, + 0.010459673590958118, + 0.0055552395060658455, + 0.009378557093441486, + 0.008701113983988762, + 0.008562496863305569, + 0.0034710753243416548, + 0.0029451099690049887, + 0.008230119943618774, + 0.009075884707272053, + 0.0030589387752115726, + 0.0033121646847575903, + 0.0059930249117314816, + 0.006132048089057207, + 0.004738796502351761, + 0.0024489681236445904, + 0.0038818269968032837, + 0.004590664058923721, + 0.003990989178419113, + 0.0027544163167476654, + 0.004105267580598593, + 0.002907264046370983, + 0.03648556023836136, + 0.01288790162652731, + 0.046961460262537, + 0.016546031460165977, + 0.009632526896893978, + 0.011676041409373283, + 0.009877617470920086, + 0.012296972796320915, + 0.017025621607899666, + 0.0044096349738538265, + 0.007705168332904577, + 0.014655454084277153, + 0.006697050761431456, + 0.0030181310139596462, + 0.006482341792434454, + 0.004919801838696003, + 0.031549349427223206, + 0.005617780610918999, + 0.014227703213691711, + 0.04181346297264099, + 0.0038197110407054424, + 0.003488679649308324, + 0.008929773233830929, + 0.0037472527474164963, + 0.0038911832962185144, + 0.0028402735479176044, + 0.00475141080096364, + 0.002903801156207919, + 0.0045131114311516285, + 0.00330320349894464, + 0.0027606613002717495, + 0.003959793131798506, + 0.010815275833010674, + 0.006241966038942337, + 0.014137796126306057, + 0.025180833414196968, + 0.004283803515136242, + 0.007859411649405956, + 0.006129170302301645, + 0.0065889377146959305, + 0.0043387264013290405, + 0.004000296350568533, + 0.002981020836159587, + 0.0036917305551469326, + 0.004688042216002941, + 0.003822110826149583, + 0.0032240499276667833, + 0.002029502997174859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.07284796983003616 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296577751636505 + ], + "opposite_logprob_tokens": [ + -2.2379658222198486 + ], + "logprob_reward": -0.11296577751636505, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296577751636505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.508786141872406, + 0.061539892107248306, + 0.025576623156666756, + 0.03476618602871895, + 0.05300401896238327, + 0.024517161771655083, + 0.01856621913611889, + 0.08831074833869934, + 0.05992333963513374, + 0.12500964105129242 + ], + "credit_map": [ + 0.0502999871969223, + 0.014982846565544605, + 0.1986151486635208, + 0.046435821801424026, + 0.016793373972177505, + 0.01874021627008915, + 0.024361303076148033, + 0.13855747878551483, + 0.004637929145246744, + 0.004827375523746014, + 0.013312989845871925, + 0.01143623236566782, + 0.004415699280798435, + 0.003913826309144497, + 0.009415727108716965, + 0.009580124169588089, + 0.002264628652483225, + 0.001390594756230712, + 0.00555519200861454, + 0.0032756472937762737, + 0.002677023410797119, + 0.002319206250831485, + 0.0038583208806812763, + 0.004236012697219849, + 0.0043394058011472225, + 0.0038732192479074, + 0.00398104777559638, + 0.0052908556535840034, + 0.004251454956829548, + 0.002445800928398967, + 0.006875346414744854, + 0.003709058277308941, + 0.013671563938260078, + 0.0037237582728266716, + 0.016482343897223473, + 0.00551608158275485, + 0.00327767594717443, + 0.0031768186017870903, + 0.0032867209520190954, + 0.0038690632209181786, + 0.005074425134807825, + 0.0018290039151906967, + 0.00298035005107522, + 0.0043150135315954685, + 0.0023345600347965956, + 0.0019613762851804495, + 0.00270953681319952, + 0.003312900196760893, + 0.004268959164619446, + 0.0011112125357612967, + 0.0019728525076061487, + 0.004727569874376059, + 0.0012199000921100378, + 0.0012831984786316752, + 0.0020264156628400087, + 0.001956112449988723, + 0.02232326753437519, + 0.004466043319553137, + 0.011957182548940182, + 0.029791446402668953, + 0.0052551934495568275, + 0.0039728120900690556, + 0.006431210786104202, + 0.0041136066429317, + 0.016499927267432213, + 0.004398636054247618, + 0.006111762952059507, + 0.018631605431437492, + 0.0025545128155499697, + 0.003650703700259328, + 0.0048298705369234085, + 0.003246322274208069, + 0.027023207396268845, + 0.0062716444954276085, + 0.017079763114452362, + 0.04795442149043083, + 0.004332250915467739, + 0.005130033474415541, + 0.010432268492877483, + 0.006786057725548744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.11296577751636505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077453382313251 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.011077453382313251, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077453382313251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16410034894943237, + 0.33776354789733887, + 0.05942220613360405, + 0.046218693256378174, + 0.07709509879350662, + 0.02995045855641365, + 0.02226801961660385, + 0.10185236483812332, + 0.13835753500461578, + 0.022971687838435173 + ], + "credit_map": [ + 0.01673736609518528, + 0.013416115194559097, + 0.039123717695474625, + 0.022216681391000748, + 0.011506983079016209, + 0.011860176920890808, + 0.025960514321923256, + 0.023278793320059776, + 0.02287236973643303, + 0.020437482744455338, + 0.05105863884091377, + 0.07750242203474045, + 0.019908420741558075, + 0.034804388880729675, + 0.0681016594171524, + 0.04307815432548523, + 0.006405627354979515, + 0.004373888950794935, + 0.009550309740006924, + 0.006058576982468367, + 0.008430511690676212, + 0.010604928247630596, + 0.007844626903533936, + 0.006153736263513565, + 0.006704623810946941, + 0.0035791806876659393, + 0.004845747258514166, + 0.004130663815885782, + 0.010208946652710438, + 0.00594615051522851, + 0.0071937283501029015, + 0.003609653562307358, + 0.018448201939463615, + 0.004362557549029589, + 0.02442195452749729, + 0.007143440656363964, + 0.004530961159616709, + 0.007069322280585766, + 0.00552401365712285, + 0.005594651214778423, + 0.005267992150038481, + 0.0022540073841810226, + 0.00513110589236021, + 0.00341165903955698, + 0.0039747171103954315, + 0.0023158579133450985, + 0.0040397788397967815, + 0.0035553399939090014, + 0.00430445559322834, + 0.0012377881212159991, + 0.003971767146140337, + 0.00719144893810153, + 0.0012041974114254117, + 0.0014035642379894853, + 0.0019596994388848543, + 0.000995097914710641, + 0.029912296682596207, + 0.005925281904637814, + 0.011909711174666882, + 0.03473526984453201, + 0.00439869612455368, + 0.0031216517090797424, + 0.00806492194533348, + 0.003784544998779893, + 0.04525291174650192, + 0.010631326586008072, + 0.010391431860625744, + 0.042583759874105453, + 0.004745963029563427, + 0.005160018801689148, + 0.013073562644422054, + 0.006518566515296698, + 0.0038456071633845568, + 0.0019288640469312668, + 0.003068098798394203, + 0.006912802811712027, + 0.001714642159640789, + 0.001820689532905817, + 0.0023015965707600117, + 0.0013793890830129385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.011077453382313251 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864359550178051 + ], + "opposite_logprob_tokens": [ + -4.758643627166748 + ], + "logprob_reward": -0.00864359550178051, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864359550178051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1479913294315338, + 0.09737343341112137, + 0.06097561493515968, + 0.07007359713315964, + 0.1187087744474411, + 0.07854722440242767, + 0.11860163509845734, + 0.1317879855632782, + 0.11942878365516663, + 0.05651170760393143 + ], + "credit_map": [ + 0.008473227731883526, + 0.006056395824998617, + 0.03930535539984703, + 0.01306577492505312, + 0.008108416572213173, + 0.009174825623631477, + 0.012666765600442886, + 0.05114055797457695, + 0.008259700611233711, + 0.006875333841890097, + 0.014545146375894547, + 0.02053164690732956, + 0.005857166368514299, + 0.012217910960316658, + 0.015450061298906803, + 0.013636467047035694, + 0.00507555715739727, + 0.0032526887953281403, + 0.015286472626030445, + 0.012269912287592888, + 0.004578333348035812, + 0.005263929720968008, + 0.00824000220745802, + 0.007008717395365238, + 0.006668183021247387, + 0.004135740455240011, + 0.009204201400279999, + 0.015219224616885185, + 0.005208408925682306, + 0.0052982717752456665, + 0.01790589839220047, + 0.0064336638897657394, + 0.030159424990415573, + 0.005322773940861225, + 0.04075678437948227, + 0.010385162197053432, + 0.005792082753032446, + 0.008666169829666615, + 0.008215093053877354, + 0.00941129494458437, + 0.011714613996446133, + 0.005397625733166933, + 0.01609260030090809, + 0.008882895112037659, + 0.008609093725681305, + 0.0061604930087924, + 0.012035843916237354, + 0.009654062800109386, + 0.031492166221141815, + 0.006037380080670118, + 0.016494354233145714, + 0.043463755398988724, + 0.0040409318171441555, + 0.003618343733251095, + 0.00951272901147604, + 0.00394196342676878, + 0.0410519503057003, + 0.007732325233519077, + 0.014444857835769653, + 0.04545070603489876, + 0.004902087152004242, + 0.0038055377081036568, + 0.010065230540931225, + 0.004335284698754549, + 0.0399760939180851, + 0.010316265746951103, + 0.008376591838896275, + 0.03096736967563629, + 0.005155150778591633, + 0.005722388159483671, + 0.011547655798494816, + 0.007367262616753578, + 0.0119773019105196, + 0.0034264475107192993, + 0.00848367065191269, + 0.019768862053751945, + 0.0033022945281118155, + 0.002564153866842389, + 0.004652399569749832, + 0.002336577046662569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.00864359550178051 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865727573633194 + ], + "opposite_logprob_tokens": [ + -3.048657178878784 + ], + "logprob_reward": -0.04865727573633194, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865727573633194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17649440467357635, + 0.09323965758085251, + 0.05893745273351669, + 0.07262072712182999, + 0.03426544740796089, + 0.04443909227848053, + 0.031084295362234116, + 0.23300878703594208, + 0.11273495852947235, + 0.1431751251220703 + ], + "credit_map": [ + 0.018467457965016365, + 0.010871907696127892, + 0.028262652456760406, + 0.019712306559085846, + 0.013733976520597935, + 0.008572430349886417, + 0.020081890746951103, + 0.05679180845618248, + 0.005113258026540279, + 0.007607049774378538, + 0.015064148232340813, + 0.021281611174345016, + 0.007931604981422424, + 0.006448905449360609, + 0.01792711205780506, + 0.011865978129208088, + 0.006474449764937162, + 0.0061218952760100365, + 0.012308115139603615, + 0.007505751680582762, + 0.005602847319096327, + 0.005255446303635836, + 0.007793682627379894, + 0.007875273004174232, + 0.007896719500422478, + 0.00542829092592001, + 0.006120570003986359, + 0.009820225648581982, + 0.014273346401751041, + 0.008655260317027569, + 0.01409777719527483, + 0.006328548304736614, + 0.0078011322766542435, + 0.0029787197709083557, + 0.00925776083022356, + 0.0039052634965628386, + 0.002653063740581274, + 0.0016586438287049532, + 0.0028099059127271175, + 0.00320095825009048, + 0.007741104811429977, + 0.002878174651414156, + 0.00887127872556448, + 0.00670468807220459, + 0.003759338054805994, + 0.003095154883340001, + 0.007852954789996147, + 0.0035364043433219194, + 0.006278941873461008, + 0.001870346604846418, + 0.004353091586381197, + 0.009563198313117027, + 0.002457957249134779, + 0.0019368499051779509, + 0.0028008639346808195, + 0.0018230500863865018, + 0.05034720525145531, + 0.0136316092684865, + 0.030948027968406677, + 0.07057739794254303, + 0.0199880413711071, + 0.01570696383714676, + 0.016897600144147873, + 0.014911950565874577, + 0.01863204315304756, + 0.009585175663232803, + 0.017073996365070343, + 0.029662353917956352, + 0.00841443706303835, + 0.010147323831915855, + 0.0101827597245574, + 0.009036880917847157, + 0.024304745718836784, + 0.007623921148478985, + 0.02533351071178913, + 0.05127806216478348, + 0.006821774411946535, + 0.008711446076631546, + 0.011022900231182575, + 0.00807877816259861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.04865727573633194 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, held at a slightly high angle, captures a minimalist, polished dark mahogany desk, its surface a flawlessly reflective plane within a sleek, techno-futuristic room. A vast floor-to-ceiling window in the background reveals a vast cityscape at dusk, towering buildings aglow with faint streaks of neon and aircraft. The transformation subtly begins as subtle cracks of bright blue light trace geometric patterns across the desk's mirror-like surface. These lines rapidly and decisively widen as the wooden top deconstructs, its panels sliding and retracting inward to reveal a breathtakingly complex interior. Beneath, a dense network of cyan-glowing circuits pulses with energy, casting cool electric light upward, while a complex polished chrome machine—a symphony of tiny pistons, interlocking gears, and articulated arms—begins to steadily and deliberately reconfigure the desk's innards. Shot at a high frame rate for hyper-real, crystal clarity, the footage captures every minute shift in the mechanisms and the crisp, dancing reflections on the moving chrome and wood surfaces, creating an atmosphere of lifeless awe and technological wonder, focused entirely on the controlled, mesmerizing deformations.", + "seed": 10138, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_4/index_136_seed_10138_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.030798340030014516, + "vlm_logprob": -0.030798340030014516, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.030798340030014516, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038777898997068405 + ], + "opposite_logprob_tokens": [ + -4.413777828216553 + ], + "logprob_reward": -0.038777898997068405, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.038777898997068405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1551823765039444, + 0.09030438214540482, + 0.05282307416200638, + 0.020033059641718864, + 0.03654820844531059, + 0.15111735463142395, + 0.019945016130805016, + 0.1525907963514328, + 0.05660267919301987, + 0.26485300064086914 + ], + "credit_map": [ + 0.012083785608410835, + 0.04385947808623314, + 0.00888821855187416, + 0.013365544378757477, + 0.017313353717327118, + 0.035584744065999985, + 0.010120447725057602, + 0.013966805301606655, + 0.006500688381493092, + 0.03207563981413841, + 0.0027045656461268663, + 0.0027367218863219023, + 0.0072772265411913395, + 0.03212601691484451, + 0.0028278930112719536, + 0.004055632743984461, + 0.0037307546008378267, + 0.017778245732188225, + 0.0017459671944379807, + 0.0022332645021378994, + 0.0035853865556418896, + 0.018724707886576653, + 0.001758755068294704, + 0.003265991108492017, + 0.0019976794719696045, + 0.002170409308746457, + 0.0017938865348696709, + 0.0022378512658178806, + 0.0032492149621248245, + 0.0036572820972651243, + 0.0021119508892297745, + 0.0028147853445261717, + 0.0032610062044113874, + 0.00849880836904049, + 0.004404260311275721, + 0.0037947194650769234, + 0.004609528463333845, + 0.00690616387873888, + 0.002145699691027403, + 0.0029280215967446566, + 0.015316474251449108, + 0.02580137923359871, + 0.01729670725762844, + 0.013627447187900543, + 0.015024503692984581, + 0.025913575664162636, + 0.023400891572237015, + 0.014736363664269447, + 0.0018589398823678493, + 0.0035021500661969185, + 0.0030073923990130424, + 0.0031284266151487827, + 0.001983191352337599, + 0.002955000614747405, + 0.0019285378511995077, + 0.0015813785139471292, + 0.014773783273994923, + 0.02256421186029911, + 0.02491411566734314, + 0.011130680330097675, + 0.016063695773482323, + 0.03013279102742672, + 0.01688215509057045, + 0.016129378229379654, + 0.006339048966765404, + 0.008322234265506268, + 0.003956150729209185, + 0.006788493134081364, + 0.011384746059775352, + 0.010356166400015354, + 0.005355773493647575, + 0.0041000694036483765, + 0.02006339281797409, + 0.07338446378707886, + 0.008251655846834183, + 0.013410096056759357, + 0.01854413002729416, + 0.0886872261762619, + 0.022137196734547615, + 0.020374834537506104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.038777898997068405 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029819265007972717 + ], + "opposite_logprob_tokens": [ + -3.5298192501068115 + ], + "logprob_reward": -0.029819265007972717, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029819265007972717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22093190252780914, + 0.14092616736888885, + 0.22684825956821442, + 0.02700749598443508, + 0.01756560616195202, + 0.07314017415046692, + 0.032716453075408936, + 0.06036248430609703, + 0.04173969477415085, + 0.15876178443431854 + ], + "credit_map": [ + 0.015055463649332523, + 0.06945008784532547, + 0.010875954292714596, + 0.012653694488108158, + 0.016342146322131157, + 0.07075750827789307, + 0.010175654664635658, + 0.015621370635926723, + 0.00943743996322155, + 0.050336804240942, + 0.004260378889739513, + 0.004129738546907902, + 0.008074913173913956, + 0.05352657288312912, + 0.004309909418225288, + 0.00685042142868042, + 0.014211026020348072, + 0.07760585099458694, + 0.007549547124654055, + 0.010316488333046436, + 0.011043693870306015, + 0.08260096609592438, + 0.009528848342597485, + 0.013991845771670341, + 0.0020694981794804335, + 0.00849628634750843, + 0.001218052813783288, + 0.0015367736341431737, + 0.0024388213641941547, + 0.008691889233887196, + 0.0013190301833674312, + 0.0012371452758088708, + 0.0014240322634577751, + 0.0042292592115700245, + 0.002061098115518689, + 0.0012019799323752522, + 0.002118336735293269, + 0.003897555870935321, + 0.0008805120014585555, + 0.001752830226905644, + 0.007351510692387819, + 0.013633722439408302, + 0.007442685775458813, + 0.007531469222158194, + 0.008957712911069393, + 0.01161507610231638, + 0.010291695594787598, + 0.0063163042068481445, + 0.0036054670345038176, + 0.007864444516599178, + 0.0030535634141415358, + 0.003636440960690379, + 0.004574693273752928, + 0.004685538355261087, + 0.003042115829885006, + 0.0022541882935911417, + 0.007912304252386093, + 0.006787976715713739, + 0.010771067813038826, + 0.004350618459284306, + 0.0077106282114982605, + 0.006046636961400509, + 0.009508512914180756, + 0.00727474270388484, + 0.0038975169882178307, + 0.005138463340699673, + 0.004453173838555813, + 0.00599483260884881, + 0.006422167178243399, + 0.005439401604235172, + 0.0057693771086633205, + 0.004624762572348118, + 0.015142867341637611, + 0.038788847625255585, + 0.00848623737692833, + 0.010033206082880497, + 0.013465861789882183, + 0.04196527972817421, + 0.018270960077643394, + 0.012608521617949009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.029819265007972717 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638868108391762 + ], + "opposite_logprob_tokens": [ + -4.758638858795166 + ], + "logprob_reward": -0.008638868108391762, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638868108391762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09113883227109909, + 0.11203707754611969, + 0.22213339805603027, + 0.014897706918418407, + 0.030031459406018257, + 0.1677074283361435, + 0.10529538989067078, + 0.05297957360744476, + 0.03685140609741211, + 0.1669277399778366 + ], + "credit_map": [ + 0.009980041533708572, + 0.015363930724561214, + 0.009947806596755981, + 0.005789533723145723, + 0.014302024617791176, + 0.016487671062350273, + 0.009388280101120472, + 0.009879538789391518, + 0.008331255055963993, + 0.03497456759214401, + 0.0049186889082193375, + 0.0039441525004804134, + 0.00901723187416792, + 0.03919994458556175, + 0.00499596307054162, + 0.0066552720963954926, + 0.015396968461573124, + 0.0827844962477684, + 0.0054777576588094234, + 0.00473681278526783, + 0.010217352770268917, + 0.08927637338638306, + 0.004579776432365179, + 0.009663863107562065, + 0.0019013970158994198, + 0.0017196480184793472, + 0.0014122166903689504, + 0.0011455706553533673, + 0.0024850668851286173, + 0.001958040753379464, + 0.0020767326932400465, + 0.0021990351378917694, + 0.0030514795798808336, + 0.006205055862665176, + 0.0031108821276575327, + 0.003025414887815714, + 0.004957093391567469, + 0.004827588796615601, + 0.002501012524589896, + 0.0023529292084276676, + 0.017012719064950943, + 0.030224956572055817, + 0.01902577467262745, + 0.014478195458650589, + 0.015907537192106247, + 0.028119875118136406, + 0.026338379830121994, + 0.01660000905394554, + 0.011487838812172413, + 0.03000880777835846, + 0.010904299095273018, + 0.007557007484138012, + 0.013060071505606174, + 0.01411508396267891, + 0.013190904632210732, + 0.004971380345523357, + 0.004486944060772657, + 0.006960027385503054, + 0.004841862712055445, + 0.009132759645581245, + 0.0050339289009571075, + 0.005101410672068596, + 0.011651666834950447, + 0.0057709733955562115, + 0.005448388867080212, + 0.005903957411646843, + 0.003776049707084894, + 0.005141229368746281, + 0.005795367993414402, + 0.004028696101158857, + 0.0035007346887141466, + 0.003256981959566474, + 0.013273964636027813, + 0.044601187109947205, + 0.005466127302497625, + 0.011361286044120789, + 0.012702940963208675, + 0.05136208236217499, + 0.014866800978779793, + 0.01329333521425724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 9 + ], + "selected_reward": -0.008638868108391762 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067379651591181755 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.0067379651591181755, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067379651591181755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.085877925157547, + 0.04169556125998497, + 0.036224741488695145, + 0.03267575055360794, + 0.026347598060965538, + 0.13103362917900085, + 0.024241391569375992, + 0.06432987004518509, + 0.05109674111008644, + 0.5064768195152283 + ], + "credit_map": [ + 0.006705604027956724, + 0.022492380812764168, + 0.004521963186562061, + 0.004100788850337267, + 0.008442295715212822, + 0.02570849284529686, + 0.005261164158582687, + 0.008645230904221535, + 0.0030189985409379005, + 0.004535273183137178, + 0.004347736015915871, + 0.004769863095134497, + 0.004429956432431936, + 0.007109538652002811, + 0.004544773139059544, + 0.00893942266702652, + 0.002831765217706561, + 0.004004929680377245, + 0.0034590905997902155, + 0.004323122091591358, + 0.004248874727636576, + 0.005501515232026577, + 0.004702945705503225, + 0.0071524977684021, + 0.0027253045700490475, + 0.007992536760866642, + 0.0016522471560165286, + 0.0022209244780242443, + 0.003409238299354911, + 0.010089058429002762, + 0.001971995225176215, + 0.002614445984363556, + 0.0023257092107087374, + 0.005181367974728346, + 0.003025112207978964, + 0.0023309034295380116, + 0.002925607841461897, + 0.006811006460338831, + 0.0017143014119938016, + 0.002033590106293559, + 0.013018034398555756, + 0.0226533655077219, + 0.01522681675851345, + 0.013063650578260422, + 0.013502382673323154, + 0.022371118888258934, + 0.01847011037170887, + 0.012728148140013218, + 0.0024208908434957266, + 0.004234231077134609, + 0.003860234748572111, + 0.0020982264541089535, + 0.0029720792081207037, + 0.003263472579419613, + 0.0025527060497552156, + 0.00283955130726099, + 0.006235601846128702, + 0.0077918278984725475, + 0.010096054524183273, + 0.006075252778828144, + 0.00905352272093296, + 0.008837253786623478, + 0.00854503270238638, + 0.007695321459323168, + 0.008705971762537956, + 0.006547985598444939, + 0.005764861591160297, + 0.005402914714068174, + 0.007263180334120989, + 0.007440348155796528, + 0.00459761917591095, + 0.005373857449740171, + 0.03819407522678375, + 0.15279364585876465, + 0.01630033552646637, + 0.01929476112127304, + 0.028016414493322372, + 0.17184476554393768, + 0.04290509223937988, + 0.037127766758203506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0067379651591181755 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001770287752151 + ], + "opposite_logprob_tokens": [ + -2.6950178146362305 + ], + "logprob_reward": -0.07001770287752151, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001770287752151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15423683822155, + 0.2048823982477188, + 0.047089722007513046, + 0.04494672268629074, + 0.018155794590711594, + 0.1730378270149231, + 0.04753242805600166, + 0.15689250826835632, + 0.05316271260380745, + 0.10006310790777206 + ], + "credit_map": [ + 0.012896563857793808, + 0.03967299312353134, + 0.00930903758853674, + 0.011969603598117828, + 0.016989076510071754, + 0.032132912427186966, + 0.009574420750141144, + 0.021692218258976936, + 0.014502523466944695, + 0.07845605909824371, + 0.004862343892455101, + 0.0041984133422374725, + 0.011631377041339874, + 0.08008628338575363, + 0.004120938014239073, + 0.007024458609521389, + 0.0034658920485526323, + 0.015420232899487019, + 0.0022230211179703474, + 0.0023605681490153074, + 0.004495639819651842, + 0.014416702091693878, + 0.002717694966122508, + 0.0019899720791727304, + 0.0037447763606905937, + 0.01152737159281969, + 0.0034356056712567806, + 0.004412160255014896, + 0.004325364250689745, + 0.009990335442125797, + 0.004060717299580574, + 0.0034503925126045942, + 0.0015809269389137626, + 0.0034814071841537952, + 0.0018744387198239565, + 0.0018799402751028538, + 0.002148847794160247, + 0.0034060764592140913, + 0.0017532239435240626, + 0.0020309328101575375, + 0.015548763796687126, + 0.03106435015797615, + 0.018447760492563248, + 0.01604258269071579, + 0.01929948478937149, + 0.032533612102270126, + 0.025464266538619995, + 0.014637017622590065, + 0.0051062945276498795, + 0.007335151545703411, + 0.004436600022017956, + 0.007497199811041355, + 0.007869712077081203, + 0.0065354700200259686, + 0.004878442268818617, + 0.0038735615089535713, + 0.016694050282239914, + 0.015884287655353546, + 0.023441100493073463, + 0.016238737851381302, + 0.01740843430161476, + 0.014019266702234745, + 0.03507011756300926, + 0.018136508762836456, + 0.00561924185603857, + 0.00818360410630703, + 0.004923239815980196, + 0.007363389246165752, + 0.00784328393638134, + 0.007003267295658588, + 0.005792939569801092, + 0.006433746311813593, + 0.012276177294552326, + 0.014651099219918251, + 0.008155751042068005, + 0.01113829854875803, + 0.01289619505405426, + 0.014120043255388737, + 0.015931133180856705, + 0.010894415900111198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.07001770287752151 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, held at a slightly high angle, captures a minimalist, polished dark mahogany desk, its surface a flawlessly reflective plane within a sleek, techno-futuristic room. A vast floor-to-ceiling window in the background reveals a vast cityscape at dusk, towering buildings aglow with faint streaks of neon and aircraft. The transformation subtly begins as subtle cracks of bright blue light trace geometric patterns across the desk's mirror-like surface. These lines rapidly and decisively widen as the wooden top deconstructs, its panels sliding and retracting inward to reveal a breathtakingly complex interior. Beneath, a dense network of cyan-glowing circuits pulses with energy, casting cool electric light upward, while a complex polished chrome machine—a symphony of tiny pistons, interlocking gears, and articulated arms—begins to steadily and deliberately reconfigure the desk's innards. Shot at a high frame rate for hyper-real, crystal clarity, the footage captures every minute shift in the mechanisms and the crisp, dancing reflections on the moving chrome and wood surfaces, creating an atmosphere of lifeless awe and technological wonder, focused entirely on the controlled, mesmerizing deformations.", + "seed": 10139, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_4/index_136_seed_10139_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.022939255461096764, + "vlm_logprob": -0.022939255461096764, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.022939255461096764, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025226503610610962 + ], + "opposite_logprob_tokens": [ + -4.275226593017578 + ], + "logprob_reward": -0.025226503610610962, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.025226503610610962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20955662429332733, + 0.13743068277835846, + 0.06183019280433655, + 0.13420377671718597, + 0.12300220876932144, + 0.04703652858734131, + 0.03961748257279396, + 0.06595251709222794, + 0.10711740702390671, + 0.07425256818532944 + ], + "credit_map": [ + 0.01725129783153534, + 0.01616697758436203, + 0.054593443870544434, + 0.023902008309960365, + 0.01755380444228649, + 0.01361543033272028, + 0.03349633887410164, + 0.03297734633088112, + 0.009896027855575085, + 0.0119745759293437, + 0.01640629954636097, + 0.03553894907236099, + 0.007381494622677565, + 0.010732249356806278, + 0.027960192412137985, + 0.017540890723466873, + 0.005737943109124899, + 0.004547238349914551, + 0.007603229023516178, + 0.01218078937381506, + 0.00381690775975585, + 0.004353751894086599, + 0.013374339789152145, + 0.010215993970632553, + 0.012004056945443153, + 0.014444209635257721, + 0.02451465092599392, + 0.025060540065169334, + 0.011666522361338139, + 0.008783712983131409, + 0.020367596298456192, + 0.01736249215900898, + 0.011934595182538033, + 0.010733264498412609, + 0.01779244653880596, + 0.01817036233842373, + 0.013011732138693333, + 0.024059567600488663, + 0.012339998036623001, + 0.014960244297981262, + 0.004591938108205795, + 0.004380791448056698, + 0.004942215979099274, + 0.009093528613448143, + 0.007223022636026144, + 0.005090746562927961, + 0.006075230892747641, + 0.005639060400426388, + 0.0033854430075734854, + 0.0032453760504722595, + 0.0034582416992634535, + 0.004288967698812485, + 0.008401892147958279, + 0.0048734755255281925, + 0.005999015644192696, + 0.005965070333331823, + 0.005288504064083099, + 0.004564157221466303, + 0.004285625647753477, + 0.01906416192650795, + 0.00473359739407897, + 0.007380298804491758, + 0.007967590354382992, + 0.012668583542108536, + 0.012848295271396637, + 0.010720204561948776, + 0.01465743500739336, + 0.012084181420505047, + 0.01426657009869814, + 0.012483049184083939, + 0.014571066945791245, + 0.015486616641283035, + 0.00829283520579338, + 0.007241438142955303, + 0.006992656737565994, + 0.01460039708763361, + 0.009405803866684437, + 0.011283330619335175, + 0.007928458973765373, + 0.00850764662027359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.025226503610610962 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195733428001404 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.018195733428001404, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195733428001404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1435869038105011, + 0.0757768377661705, + 0.05397216975688934, + 0.5644083619117737, + 0.028637100011110306, + 0.012780260294675827, + 0.022120490670204163, + 0.019975047558546066, + 0.040903240442276, + 0.03783958777785301 + ], + "credit_map": [ + 0.013003687374293804, + 0.013146104291081429, + 0.020488841459155083, + 0.027625378221273422, + 0.014137576334178448, + 0.014143436215817928, + 0.019961291924118996, + 0.021080592647194862, + 0.006416394375264645, + 0.004424331244081259, + 0.010920044034719467, + 0.015219614841043949, + 0.0058509004302322865, + 0.006225366611033678, + 0.012586998753249645, + 0.014133187010884285, + 0.004944212734699249, + 0.002784821204841137, + 0.007273429539054632, + 0.01022290624678135, + 0.004362596198916435, + 0.005124617833644152, + 0.008716986514627934, + 0.010542597621679306, + 0.04497826099395752, + 0.05129000544548035, + 0.07477403432130814, + 0.09473147988319397, + 0.0782364159822464, + 0.07631251960992813, + 0.06240537017583847, + 0.0816802829504013, + 0.0038566626608371735, + 0.002933926647529006, + 0.003900476498529315, + 0.003800504608079791, + 0.002879349049180746, + 0.0032382262870669365, + 0.004630186129361391, + 0.0033977667335420847, + 0.0018990988610312343, + 0.0014745948137715459, + 0.0014969512121751904, + 0.0014446581481024623, + 0.0013518129708245397, + 0.0018351380713284016, + 0.001552079920656979, + 0.0017259269952774048, + 0.0023150150664150715, + 0.0015877829864621162, + 0.0023797382600605488, + 0.0026407220866531134, + 0.004051890689879656, + 0.0029275219421833754, + 0.003184917848557234, + 0.003032902954146266, + 0.0022998068016022444, + 0.0018900869181379676, + 0.001906209858134389, + 0.0027731440495699644, + 0.0021128705702722073, + 0.0035535988863557577, + 0.0028462286572903395, + 0.0025931044947355986, + 0.004140348173677921, + 0.0030426273588091135, + 0.0042283181101083755, + 0.006465168669819832, + 0.007248652633279562, + 0.0064380900003015995, + 0.005858668126165867, + 0.0034813641104847193, + 0.003029341110959649, + 0.003565171966329217, + 0.002823499497026205, + 0.004102626349776983, + 0.0071895429864525795, + 0.008246227167546749, + 0.004433887545019388, + 0.0044492920860648155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.018195733428001404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071793735027313 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.011071793735027313, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071793735027313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3165237605571747, + 0.1891002506017685, + 0.08156175911426544, + 0.1535845398902893, + 0.05358859524130821, + 0.05555565655231476, + 0.03377879783511162, + 0.042588554322719574, + 0.045560985803604126, + 0.028157105669379234 + ], + "credit_map": [ + 0.024169251322746277, + 0.026571068912744522, + 0.06400790065526962, + 0.04630456492304802, + 0.041574396193027496, + 0.031222892925143242, + 0.035971738398075104, + 0.046701956540346146, + 0.00959348026663065, + 0.012436575256288052, + 0.02607930637896061, + 0.033547453582286835, + 0.01222757063806057, + 0.01428237371146679, + 0.06659778952598572, + 0.014335695654153824, + 0.006739822216331959, + 0.00490436889231205, + 0.010108567774295807, + 0.012865038588643074, + 0.008683654479682446, + 0.011882628314197063, + 0.014103754423558712, + 0.012273925356566906, + 0.012958867475390434, + 0.012928823009133339, + 0.01992058753967285, + 0.023414025083184242, + 0.021348604932427406, + 0.022392600774765015, + 0.016803044825792313, + 0.023817984387278557, + 0.006044705398380756, + 0.00492488220334053, + 0.008960666134953499, + 0.00866444781422615, + 0.005778636783361435, + 0.00621121795848012, + 0.006697723176330328, + 0.006306316703557968, + 0.007228619884699583, + 0.005898415111005306, + 0.008434819057583809, + 0.009472139179706573, + 0.004902875050902367, + 0.006059244275093079, + 0.007427058182656765, + 0.006132486276328564, + 0.004616272170096636, + 0.0033814047928899527, + 0.0027466220781207085, + 0.003525432199239731, + 0.009115628898143768, + 0.0033582844771444798, + 0.0027663533110171556, + 0.004268798977136612, + 0.0034734043292701244, + 0.004824116826057434, + 0.0031921041663736105, + 0.010764827951788902, + 0.003941175527870655, + 0.00485741812735796, + 0.004540605470538139, + 0.006994900293648243, + 0.0042859409004449844, + 0.006598613224923611, + 0.0032562674023211002, + 0.0055184257216751575, + 0.008565539494156837, + 0.006639584898948669, + 0.005809876602143049, + 0.00488673709332943, + 0.002307025482878089, + 0.0032613002695143223, + 0.0025588185526430607, + 0.003975313156843185, + 0.003228911431506276, + 0.005130109377205372, + 0.003952501341700554, + 0.003743125591427088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.011071793735027313 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251897498965263 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251897498965263, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251897498965263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1888222098350525, + 0.11273670941591263, + 0.06512012332677841, + 0.2118394821882248, + 0.04771328717470169, + 0.04180534929037094, + 0.0851721465587616, + 0.07151509076356888, + 0.11194053292274475, + 0.06333501636981964 + ], + "credit_map": [ + 0.014913033694028854, + 0.01804761216044426, + 0.04202539473772049, + 0.02930397354066372, + 0.018854448571801186, + 0.016176430508494377, + 0.0240744911134243, + 0.025426829233765602, + 0.009846661239862442, + 0.005831367801874876, + 0.016555922105908394, + 0.02492082305252552, + 0.006622208748012781, + 0.008020389825105667, + 0.019302133470773697, + 0.021637199446558952, + 0.005921490490436554, + 0.003788737114518881, + 0.008707503788173199, + 0.011910909786820412, + 0.005100280977785587, + 0.00583374360576272, + 0.012217818759381771, + 0.011639640666544437, + 0.010052526369690895, + 0.01477237232029438, + 0.039702389389276505, + 0.0467747338116169, + 0.017895398661494255, + 0.01968330703675747, + 0.03641634061932564, + 0.0265424195677042, + 0.005246299784630537, + 0.004918910562992096, + 0.006968813017010689, + 0.009263399988412857, + 0.005076839122921228, + 0.005389086902141571, + 0.005724101793020964, + 0.005125833209604025, + 0.004387004300951958, + 0.004573279060423374, + 0.0041758110746741295, + 0.005386429373174906, + 0.004978075623512268, + 0.006073617376387119, + 0.004820624366402626, + 0.007410508580505848, + 0.008626560680568218, + 0.006074737291783094, + 0.012557662092149258, + 0.011930655688047409, + 0.009336589835584164, + 0.009223350323736668, + 0.01728256233036518, + 0.010140031576156616, + 0.006203236989676952, + 0.0062487502582371235, + 0.004898950923234224, + 0.018490443006157875, + 0.00797322578728199, + 0.007104657590389252, + 0.007246542256325483, + 0.01334928348660469, + 0.010007289238274097, + 0.011910409666597843, + 0.02035791240632534, + 0.011805557645857334, + 0.016834167763590813, + 0.013443938456475735, + 0.01248075719922781, + 0.015100490301847458, + 0.0050433543510735035, + 0.0064370171166956425, + 0.006454386282712221, + 0.00803441647440195, + 0.007968041114509106, + 0.008685384877026081, + 0.007831033319234848, + 0.01288138423115015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.005251897498965263 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot held at a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a polished dark mahogany desk with a flawlessly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background feature a floor-to-ceiling window revealing a cityscape at dusk with neon lights and aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk deconstruct by cracking with blue light, sliding panels inward, and revealing cyan-glowing circuits and a chrome machine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage shot at a high frame rate showing hyper-real clarity and crisp reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054950349032878876 + ], + "opposite_logprob_tokens": [ + -2.92995023727417 + ], + "logprob_reward": -0.054950349032878876, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054950349032878876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25423765182495117, + 0.08231330662965775, + 0.07694700360298157, + 0.1504301130771637, + 0.06072038412094116, + 0.044172149151563644, + 0.05292126163840294, + 0.058772917836904526, + 0.13334232568740845, + 0.08614284545183182 + ], + "credit_map": [ + 0.0148934880271554, + 0.021622199565172195, + 0.05033041536808014, + 0.048121269792318344, + 0.021101485937833786, + 0.021237919107079506, + 0.03618872910737991, + 0.04074215516448021, + 0.00719954539090395, + 0.007558708544820547, + 0.008130141533911228, + 0.01748650148510933, + 0.007296240888535976, + 0.012308400124311447, + 0.012173892930150032, + 0.010159877128899097, + 0.0052087861113250256, + 0.005225394852459431, + 0.007885290309786797, + 0.017188427969813347, + 0.003493248950690031, + 0.0036848406307399273, + 0.026597900316119194, + 0.007663119584321976, + 0.015627389773726463, + 0.018183238804340363, + 0.020086390897631645, + 0.024274669587612152, + 0.011923247016966343, + 0.012864106334745884, + 0.020113039761781693, + 0.0273580364882946, + 0.009664174169301987, + 0.005712707061320543, + 0.007458011619746685, + 0.007576276082545519, + 0.007423921953886747, + 0.0086268475279212, + 0.006777357775717974, + 0.0074810911901295185, + 0.007033119443804026, + 0.00442280201241374, + 0.004896420054137707, + 0.005029531195759773, + 0.003772198921069503, + 0.005902743898332119, + 0.006423087790608406, + 0.00669224513694644, + 0.004894052166491747, + 0.003875839989632368, + 0.006873325910419226, + 0.006105358246713877, + 0.010030671954154968, + 0.004903602879494429, + 0.007224811241030693, + 0.009013607166707516, + 0.0071033635176718235, + 0.006493256893008947, + 0.003819634672254324, + 0.011830492876470089, + 0.005273544695228338, + 0.010106571018695831, + 0.006620799191296101, + 0.007525256834924221, + 0.015357030555605888, + 0.013515446335077286, + 0.014959502033889294, + 0.019902314990758896, + 0.017665179446339607, + 0.014077884145081043, + 0.027121618390083313, + 0.010743368417024612, + 0.005702556110918522, + 0.005955605302006006, + 0.0049193683080375195, + 0.015138921327888966, + 0.012232786975800991, + 0.012687748298048973, + 0.010222487151622772, + 0.01928338222205639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.054950349032878876 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_6/index_tensor([9371, 9371])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_6/index_tensor([9371, 9371])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..80b8d9d9fd3b55109aad01320ba146a71b1e3140 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_6/index_tensor([9371, 9371])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic slow-motion shot, shot from a slight upward angle, the camera slowly pans to the right, capturing a breathtaking aerial duel amidst swirling, inky black storm clouds. A massive, sinuous, cyan dragon clashes fiercely with a muscular white tiger. The dragon, covered in shimmering metallic cyan scales, shimmers brilliantly against the occasional streaks of purple lightning deep within the clouds. Its majestic horns and eyes, ablaze with golden fire, exude a commanding presence. The tiger, with its massive, richly feathered white wings spread wide, its snow-white fur streaked with deep black stripes, its amber pupils gleaming with endless fighting spirit. In extreme slow motion, the dragon brandishes its claws as sharp as a knife, and the tiger extends its tiger paws, which are shining with cold light, to meet the attack with the same swift posture. At the moment when the dragon claws and tiger paws collide violently in the air, a cluster of dazzling golden sparks bursts and splashes. Each spark drags a tiny tail of light, which is clearly visible against the dark background. The lightning casts an ever-changing and dramatic outline light on the fighting beasts, highlighting every muscle tension and burst of strength. The whole picture is full of epic power and mythological tension, freezing this battle between the mythical beasts that changes the world into a moment full of visual impact.", + "seed": 25345, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_6/index_9371_seed_25345_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14279851452447473, + "vlm_logprob": -0.14279851452447473, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.14279851452447473, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5765494704246521 + ], + "opposite_logprob_tokens": [ + -0.8265494704246521 + ], + "logprob_reward": -0.5765494704246521, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5765494704246521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3444877862930298, + 0.20350776612758636, + 0.08807913959026337, + 0.16175970435142517, + 0.07534445822238922, + 0.03649890795350075, + 0.03768667206168175, + 0.0262016449123621, + 0.007643919438123703, + 0.01879004016518593 + ], + "credit_map": [ + 0.03071310743689537, + 0.03059566579759121, + 0.09114562720060349, + 0.028953813016414642, + 0.0347171351313591, + 0.07390675693750381, + 0.02774529717862606, + 0.026710381731390953, + 0.053413163870573044, + 0.008471234701573849, + 0.0698077529668808, + 0.011867515742778778, + 0.01034516654908657, + 0.013951627537608147, + 0.010844504460692406, + 0.024806801229715347, + 0.009915253147482872, + 0.007474399637430906, + 0.0075324843637645245, + 0.015906015411019325, + 0.010782018303871155, + 0.011598133482038975, + 0.019008774310350418, + 0.005862055346369743, + 0.022421833127737045, + 0.026983188465237617, + 0.02359703555703163, + 0.02037849649786949, + 0.015894001349806786, + 0.01264546811580658, + 0.019435198977589607, + 0.020404476672410965, + 0.009446930140256882, + 0.010025089606642723, + 0.008467836305499077, + 0.008761578239500523, + 0.008425395004451275, + 0.0067396885715425014, + 0.014402035623788834, + 0.00907590240240097, + 0.003359055845066905, + 0.0050806342624127865, + 0.0041809468530118465, + 0.004898865707218647, + 0.0044079553335905075, + 0.00478182639926672, + 0.005666767247021198, + 0.004122857004404068, + 0.00462390948086977, + 0.004050280898809433, + 0.007990939542651176, + 0.004852619953453541, + 0.004069565795361996, + 0.005676943343132734, + 0.003344739554449916, + 0.003077676985412836, + 0.003075580345466733, + 0.00437902333214879, + 0.004464416764676571, + 0.003946246579289436, + 0.0028395820409059525, + 0.001950556063093245, + 0.001760181039571762, + 0.003786060493439436, + 0.0009842392755672336, + 0.0011374911991879344, + 0.0008094855584204197, + 0.0006245196564123034, + 0.0009220537613146007, + 0.0009434689418412745, + 0.0012396576348692179, + 0.0009830035269260406, + 0.0023080212995409966, + 0.0027184775099158287, + 0.0029917878564447165, + 0.0030723819509148598, + 0.0024327770806849003, + 0.0017533758655190468, + 0.001563274534419179, + 0.0019499434856697917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.5765494704246521 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003199935657903552 + ], + "opposite_logprob_tokens": [ + -5.753200054168701 + ], + "logprob_reward": -0.003199935657903552, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003199935657903552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29674988985061646, + 0.1810385137796402, + 0.1061958521604538, + 0.14176245033740997, + 0.06841181218624115, + 0.05663900077342987, + 0.02766791544854641, + 0.09288768470287323, + 0.014771762304008007, + 0.013875109143555164 + ], + "credit_map": [ + 0.025601251050829887, + 0.027786243706941605, + 0.07805570960044861, + 0.024471577256917953, + 0.030791910365223885, + 0.06212024390697479, + 0.022313334047794342, + 0.025609608739614487, + 0.045901115983724594, + 0.007573442067950964, + 0.06695312261581421, + 0.01030236016958952, + 0.009317857213318348, + 0.009901327081024647, + 0.007506270427256823, + 0.02358303777873516, + 0.011595159769058228, + 0.008557225577533245, + 0.011035756208002567, + 0.01824280060827732, + 0.011627244763076305, + 0.015050122514367104, + 0.02274481952190399, + 0.0073427255265414715, + 0.020234784111380577, + 0.01895308494567871, + 0.020458057522773743, + 0.019033975899219513, + 0.013162363320589066, + 0.012408166192471981, + 0.02216731384396553, + 0.01534469984471798, + 0.007894689217209816, + 0.007988424971699715, + 0.00703376205638051, + 0.007213474251329899, + 0.0076592001132667065, + 0.005037171766161919, + 0.01801953837275505, + 0.007565560285001993, + 0.004469209350645542, + 0.005200134124606848, + 0.010653199627995491, + 0.008521579205989838, + 0.004542050417512655, + 0.01008815597742796, + 0.007401980459690094, + 0.005762691143900156, + 0.0035038506612181664, + 0.005109873134642839, + 0.0035800556652247906, + 0.004989865701645613, + 0.002965295920148492, + 0.003667904296889901, + 0.002019642386585474, + 0.0018314288463443518, + 0.0178321935236454, + 0.013650543056428432, + 0.014261946082115173, + 0.010647200047969818, + 0.008287288248538971, + 0.008715971373021603, + 0.0073816063813865185, + 0.012110930867493153, + 0.001571168890222907, + 0.0020521266851574183, + 0.0018084145849570632, + 0.0015700419899076223, + 0.0019414988346397877, + 0.0021651587449014187, + 0.001732083153910935, + 0.0019312690710648894, + 0.0016409186646342278, + 0.0014424477703869343, + 0.0021757916547358036, + 0.002138483105227351, + 0.0020233429968357086, + 0.0015251848381012678, + 0.0015675307950004935, + 0.001361409667879343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.003199935657903552 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626228500157595 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007626228500157595, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626228500157595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1572495996952057, + 0.1684514433145523, + 0.05291132628917694, + 0.14071187376976013, + 0.07356652617454529, + 0.041623882949352264, + 0.037133969366550446, + 0.2932104170322418, + 0.015313709154725075, + 0.01982726901769638 + ], + "credit_map": [ + 0.013963080011308193, + 0.015652596950531006, + 0.0368676483631134, + 0.014142688363790512, + 0.01761939935386181, + 0.029357237741351128, + 0.013264757581055164, + 0.016382187604904175, + 0.04508405923843384, + 0.006482819560915232, + 0.060518667101860046, + 0.008802283555269241, + 0.008347699418663979, + 0.008459437638521194, + 0.009321698918938637, + 0.021434789523482323, + 0.006352303549647331, + 0.0038956112693995237, + 0.004827544093132019, + 0.008875078521668911, + 0.007682591211050749, + 0.006166499108076096, + 0.011211497709155083, + 0.003900199895724654, + 0.018998442217707634, + 0.02631191536784172, + 0.019984442740678787, + 0.01694250851869583, + 0.01454257033765316, + 0.01110897958278656, + 0.016227586194872856, + 0.01659543253481388, + 0.007460619322955608, + 0.007721457630395889, + 0.007273833267390728, + 0.007038667798042297, + 0.007773705292493105, + 0.005518015474081039, + 0.022951366379857063, + 0.007828863337635994, + 0.003060242161154747, + 0.004387891851365566, + 0.005672536790370941, + 0.00563531881198287, + 0.004813858773559332, + 0.005126818083226681, + 0.007933915592730045, + 0.004993303678929806, + 0.003597748000174761, + 0.005211547948420048, + 0.005495570600032806, + 0.0066992612555623055, + 0.0037274642381817102, + 0.0028695224318653345, + 0.0038037244230508804, + 0.005729127675294876, + 0.03280699998140335, + 0.02799990400671959, + 0.04054085537791252, + 0.02454366907477379, + 0.029837707057595253, + 0.039016563445329666, + 0.039994969964027405, + 0.0584697462618351, + 0.001718061277642846, + 0.0022341874428093433, + 0.0016425251960754395, + 0.0016784683102741838, + 0.0020373123697936535, + 0.002081568818539381, + 0.0016314906533807516, + 0.0022900961339473724, + 0.0022877247538417578, + 0.0025748719926923513, + 0.0023217492271214724, + 0.0022314083762466908, + 0.003785645356401801, + 0.002633216790854931, + 0.0023212938103824854, + 0.001671358710154891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.007626228500157595 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026335513219237328 + ], + "opposite_logprob_tokens": [ + -3.6513354778289795 + ], + "logprob_reward": -0.026335513219237328, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026335513219237328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08011474460363388, + 0.07080883532762527, + 0.16174866259098053, + 0.1518520712852478, + 0.07413768768310547, + 0.06078241020441055, + 0.07110031694173813, + 0.06952457875013351, + 0.03812713548541069, + 0.22180354595184326 + ], + "credit_map": [ + 0.00870591402053833, + 0.011903917416930199, + 0.012709298171103, + 0.007187957409769297, + 0.01005516666918993, + 0.008938690647482872, + 0.009340665303170681, + 0.011273136362433434, + 0.009879296645522118, + 0.006137229967862368, + 0.008666441775858402, + 0.010089344345033169, + 0.00785596389323473, + 0.011343625374138355, + 0.010679500177502632, + 0.006157432682812214, + 0.018632613122463226, + 0.011999320238828659, + 0.01542480569332838, + 0.021492386236786842, + 0.021748585626482964, + 0.016334136947989464, + 0.043677058070898056, + 0.012439761310815811, + 0.01899978145956993, + 0.02870463952422142, + 0.019077958539128304, + 0.018721306696534157, + 0.018362201750278473, + 0.010472116060554981, + 0.019833654165267944, + 0.017680412158370018, + 0.008645453490316868, + 0.010282309725880623, + 0.00664602080360055, + 0.010099063627421856, + 0.008614678867161274, + 0.0060330526903271675, + 0.01604696735739708, + 0.007770140655338764, + 0.006117065902799368, + 0.006277991458773613, + 0.007295962423086166, + 0.009862180799245834, + 0.006353920791298151, + 0.008743549697101116, + 0.009125462733209133, + 0.007006275467574596, + 0.00579069321975112, + 0.009086064994335175, + 0.017758648842573166, + 0.010105100460350513, + 0.005226071923971176, + 0.00548173300921917, + 0.006989977788180113, + 0.010662024840712547, + 0.008256184868514538, + 0.007279778365045786, + 0.009409429505467415, + 0.006945039611309767, + 0.006150831468403339, + 0.00954727828502655, + 0.010294100269675255, + 0.011641934514045715, + 0.0037091728299856186, + 0.004767335020005703, + 0.0037950428668409586, + 0.0036967629566788673, + 0.005600666627287865, + 0.004642603453248739, + 0.006497709546238184, + 0.005417840555310249, + 0.025073084980249405, + 0.0261042769998312, + 0.02864767052233219, + 0.034222621470689774, + 0.03713703528046608, + 0.02523150108754635, + 0.020566867664456367, + 0.024820486083626747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.026335513219237328 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028142482042313 + ], + "opposite_logprob_tokens": [ + -2.3502814769744873 + ], + "logprob_reward": -0.10028142482042313, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028142482042313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17903432250022888, + 0.0986965075135231, + 0.03090321272611618, + 0.10030823945999146, + 0.041464272886514664, + 0.07687824964523315, + 0.19518528878688812, + 0.06262395530939102, + 0.08200734108686447, + 0.13289859890937805 + ], + "credit_map": [ + 0.02796919085085392, + 0.01901165023446083, + 0.03675553947687149, + 0.013968968763947487, + 0.019854532554745674, + 0.02774207293987274, + 0.01419916097074747, + 0.019533218815922737, + 0.025465350598096848, + 0.007759907748550177, + 0.01593676581978798, + 0.0106441555544734, + 0.009351801127195358, + 0.015356320887804031, + 0.007204853929579258, + 0.006977361161261797, + 0.004801342263817787, + 0.0030592961702495813, + 0.0038072753231972456, + 0.0028800114523619413, + 0.004171561449766159, + 0.005287819541990757, + 0.004639328457415104, + 0.002256578067317605, + 0.01175683829933405, + 0.021588971838355064, + 0.010148212313652039, + 0.011922867968678474, + 0.01330884825438261, + 0.00620169285684824, + 0.012793192639946938, + 0.012587612494826317, + 0.005194984842091799, + 0.005193525925278664, + 0.005055563524365425, + 0.005782336927950382, + 0.004022671841084957, + 0.003178259590640664, + 0.008685167878866196, + 0.004351763986051083, + 0.006735455244779587, + 0.0057814414612948895, + 0.011240138672292233, + 0.006621312815696001, + 0.007848335430026054, + 0.012673204764723778, + 0.016237108036875725, + 0.009741256013512611, + 0.01740710809826851, + 0.02935308963060379, + 0.06186381354928017, + 0.03887448459863663, + 0.013790839351713657, + 0.014562923461198807, + 0.009865508414804935, + 0.009467517957091331, + 0.00833298172801733, + 0.0073416163213551044, + 0.00900318380445242, + 0.0068341572768986225, + 0.006829168647527695, + 0.00741945207118988, + 0.006623867433518171, + 0.010239530354738235, + 0.010381471365690231, + 0.014940201304852962, + 0.009034307673573494, + 0.008167441934347153, + 0.011840024963021278, + 0.009791525080800056, + 0.008425617590546608, + 0.009426754899322987, + 0.010441923514008522, + 0.015143857337534428, + 0.020642274990677834, + 0.01818208396434784, + 0.033736180514097214, + 0.012776997871696949, + 0.011805598624050617, + 0.01016966626048088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.10028142482042313 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic slow-motion shot, shot from a slight upward angle, the camera slowly pans to the right, capturing a breathtaking aerial duel amidst swirling, inky black storm clouds. A massive, sinuous, cyan dragon clashes fiercely with a muscular white tiger. The dragon, covered in shimmering metallic cyan scales, shimmers brilliantly against the occasional streaks of purple lightning deep within the clouds. Its majestic horns and eyes, ablaze with golden fire, exude a commanding presence. The tiger, with its massive, richly feathered white wings spread wide, its snow-white fur streaked with deep black stripes, its amber pupils gleaming with endless fighting spirit. In extreme slow motion, the dragon brandishes its claws as sharp as a knife, and the tiger extends its tiger paws, which are shining with cold light, to meet the attack with the same swift posture. At the moment when the dragon claws and tiger paws collide violently in the air, a cluster of dazzling golden sparks bursts and splashes. Each spark drags a tiny tail of light, which is clearly visible against the dark background. The lightning casts an ever-changing and dramatic outline light on the fighting beasts, highlighting every muscle tension and burst of strength. The whole picture is full of epic power and mythological tension, freezing this battle between the mythical beasts that changes the world into a moment full of visual impact.", + "seed": 25346, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_6/index_9371_seed_25346_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.33793093767017124, + "vlm_logprob": -0.33793093767017124, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.33793093767017124, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5025215148925781 + ], + "opposite_logprob_tokens": [ + -0.2525215148925781 + ], + "logprob_reward": -1.5025215148925781, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5025215148925781, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18014390766620636, + 0.041951317340135574, + 0.053761135786771774, + 0.05230160802602768, + 0.42316433787345886, + 0.04189673438668251, + 0.022144531831145287, + 0.02999669499695301, + 0.02471207268536091, + 0.12992769479751587 + ], + "credit_map": [ + 0.020036663860082626, + 0.021321559324860573, + 0.017973097041249275, + 0.0288603063672781, + 0.015500221401453018, + 0.021000031381845474, + 0.025150123983621597, + 0.03030189499258995, + 0.0068827299401164055, + 0.0045330035500228405, + 0.003827108535915613, + 0.008795302361249924, + 0.004238244146108627, + 0.006008569151163101, + 0.005033804103732109, + 0.0026325564831495285, + 0.004803119692951441, + 0.007011390291154385, + 0.0037957115564495325, + 0.01180518139153719, + 0.00978870503604412, + 0.0034313155338168144, + 0.005712405312806368, + 0.00741330673918128, + 0.007463359739631414, + 0.00801522471010685, + 0.007566462736576796, + 0.007311752997338772, + 0.006025896407663822, + 0.005489835981279612, + 0.004483295604586601, + 0.0059457761235535145, + 0.05391974374651909, + 0.06802969425916672, + 0.03717841953039169, + 0.1418137103319168, + 0.022183362394571304, + 0.03957788646221161, + 0.02607792802155018, + 0.03438356891274452, + 0.004184558056294918, + 0.005276849027723074, + 0.00660276971757412, + 0.005921584088355303, + 0.006533995736390352, + 0.0041425712406635284, + 0.004807980265468359, + 0.004426422528922558, + 0.002808098681271076, + 0.0025432880502194166, + 0.002270827069878578, + 0.0025892904959619045, + 0.002475112909451127, + 0.002767715370282531, + 0.0038411146961152554, + 0.002849082462489605, + 0.003152076620608568, + 0.00417834147810936, + 0.0034394459798932076, + 0.004956257063895464, + 0.0028443802148103714, + 0.003977013286203146, + 0.0037815370596945286, + 0.003667643293738365, + 0.004160711076110601, + 0.0028263325802981853, + 0.0036785148549824953, + 0.0028490214608609676, + 0.0021905750036239624, + 0.002694260561838746, + 0.003120491746813059, + 0.0031921654008328915, + 0.03149085491895676, + 0.012546898797154427, + 0.021395254880189896, + 0.011173227801918983, + 0.010413453914225101, + 0.01325250044465065, + 0.016913706436753273, + 0.012741786427795887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -1.5025215148925781 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005975003354251385 + ], + "opposite_logprob_tokens": [ + -5.130974769592285 + ], + "logprob_reward": -0.005975003354251385, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005975003354251385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19409745931625366, + 0.04480293393135071, + 0.05315610393881798, + 0.052572641521692276, + 0.340312123298645, + 0.0532531775534153, + 0.03672465682029724, + 0.02861221879720688, + 0.040853749960660934, + 0.1556149572134018 + ], + "credit_map": [ + 0.023358313366770744, + 0.024026062339544296, + 0.01830386370420456, + 0.022375980392098427, + 0.01927805133163929, + 0.031942691653966904, + 0.02357158251106739, + 0.031240906566381454, + 0.0069860126823186874, + 0.005755491554737091, + 0.004248372744768858, + 0.006857459899038076, + 0.005693416576832533, + 0.006551853381097317, + 0.00513507891446352, + 0.0035752516705542803, + 0.005142716690897942, + 0.004975778516381979, + 0.004104975610971451, + 0.013127770274877548, + 0.006189232692122459, + 0.004517983645200729, + 0.008032910525798798, + 0.007064733188599348, + 0.007679576985538006, + 0.007804248481988907, + 0.0076481797732412815, + 0.006895795930176973, + 0.006796916481107473, + 0.005411683581769466, + 0.004423535894602537, + 0.00591270299628377, + 0.04482197389006615, + 0.05360081046819687, + 0.030282089486718178, + 0.11601050198078156, + 0.017534904181957245, + 0.029746972024440765, + 0.0214383602142334, + 0.02687651291489601, + 0.005897125694900751, + 0.006709576118737459, + 0.009267584420740604, + 0.007151281461119652, + 0.00708424486219883, + 0.0050417217426002026, + 0.006198332644999027, + 0.005903309676796198, + 0.004144031088799238, + 0.004905963782221079, + 0.0033931531943380833, + 0.004480831325054169, + 0.004353566560894251, + 0.00528006162494421, + 0.006268424913287163, + 0.00389862316660583, + 0.0026336186565458775, + 0.003730843774974346, + 0.003296004142612219, + 0.004089685622602701, + 0.0027902477886527777, + 0.004483347292989492, + 0.003734834725037217, + 0.003853636560961604, + 0.005292312707751989, + 0.004641297273337841, + 0.00711459293961525, + 0.004413648974150419, + 0.003849388100206852, + 0.0031588224228471518, + 0.0052286540158092976, + 0.007155036553740501, + 0.03404156491160393, + 0.017571337521076202, + 0.0244058296084404, + 0.014398769475519657, + 0.013475606217980385, + 0.015479976311326027, + 0.021697331219911575, + 0.014544546604156494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.005975003354251385 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075802147388458 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011075802147388458, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075802147388458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16393817961215973, + 0.06498119235038757, + 0.0634685531258583, + 0.06463801860809326, + 0.08393614739179611, + 0.07269096374511719, + 0.048302508890628815, + 0.03792126849293709, + 0.0610649548470974, + 0.3390582203865051 + ], + "credit_map": [ + 0.016358470544219017, + 0.016785435378551483, + 0.0194763895124197, + 0.01766381599009037, + 0.01589539274573326, + 0.020483173429965973, + 0.022181155160069466, + 0.035094358026981354, + 0.010312092490494251, + 0.006119594909250736, + 0.006531840655952692, + 0.009298224933445454, + 0.007177471648901701, + 0.008432658389210701, + 0.010075331665575504, + 0.007033984176814556, + 0.00465619470924139, + 0.006749980617314577, + 0.00466557452455163, + 0.015012797899544239, + 0.007703378330916166, + 0.005275243427604437, + 0.009812455624341965, + 0.009592934511601925, + 0.010047946125268936, + 0.008119232021272182, + 0.010467691347002983, + 0.00792058277875185, + 0.008188308216631413, + 0.006474331486970186, + 0.005596522241830826, + 0.007823409512639046, + 0.007831317372620106, + 0.015524307265877724, + 0.007504748646169901, + 0.015145502984523773, + 0.005895174574106932, + 0.014993427321314812, + 0.006627440918236971, + 0.010414231568574905, + 0.008336922153830528, + 0.009580304846167564, + 0.012310478836297989, + 0.010078990831971169, + 0.009314817376434803, + 0.006719669792801142, + 0.00842773262411356, + 0.007922056131064892, + 0.004257035441696644, + 0.004512316081672907, + 0.004226704593747854, + 0.00598071189597249, + 0.005607020575553179, + 0.007313798181712627, + 0.01039799116551876, + 0.006006935145705938, + 0.0042891912162303925, + 0.006497704423964024, + 0.004645951092243195, + 0.0043413955718278885, + 0.0035232650116086006, + 0.005885931197553873, + 0.0040681916289031506, + 0.004669640678912401, + 0.006174927577376366, + 0.0076173762790858746, + 0.00877468939870596, + 0.007121182046830654, + 0.005518999882042408, + 0.005776256322860718, + 0.008857881650328636, + 0.011223643086850643, + 0.0527019277215004, + 0.033752575516700745, + 0.03654811531305313, + 0.03712546452879906, + 0.04177981615066528, + 0.03788793832063675, + 0.062036119401454926, + 0.037226296961307526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.011075802147388458 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043060772120952606 + ], + "opposite_logprob_tokens": [ + -3.168060779571533 + ], + "logprob_reward": -0.043060772120952606, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043060772120952606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0767672210931778, + 0.1892000287771225, + 0.03826475515961647, + 0.02966758795082569, + 0.4554927945137024, + 0.03590700030326843, + 0.034075669944286346, + 0.02485397830605507, + 0.023662526160478592, + 0.09210847318172455 + ], + "credit_map": [ + 0.007672683335840702, + 0.009587408043444157, + 0.007218174170702696, + 0.007621618453413248, + 0.011025708168745041, + 0.008803032338619232, + 0.012368513271212578, + 0.012470081448554993, + 0.020586268976330757, + 0.014738054946064949, + 0.015125890262424946, + 0.028664400801062584, + 0.032013460993766785, + 0.02293475717306137, + 0.02549852803349495, + 0.02963867411017418, + 0.005047942046076059, + 0.0038711712695658207, + 0.0023840926587581635, + 0.005697447806596756, + 0.008943061344325542, + 0.0030961947049945593, + 0.004390159156173468, + 0.004834682215005159, + 0.0033594605047255754, + 0.00400190893560648, + 0.002544573275372386, + 0.003064127406105399, + 0.004947017878293991, + 0.004055315162986517, + 0.003175226040184498, + 0.004519954789429903, + 0.05166398733854294, + 0.0812426283955574, + 0.03354661166667938, + 0.14089451730251312, + 0.025470087304711342, + 0.045534588396549225, + 0.0312083400785923, + 0.045931991189718246, + 0.0034662208054214716, + 0.004136501811444759, + 0.006791660096496344, + 0.004014256410300732, + 0.0048830825835466385, + 0.0036388251464813948, + 0.004585057497024536, + 0.00439139548689127, + 0.002922800602391362, + 0.004099262412637472, + 0.0033555980771780014, + 0.004934346303343773, + 0.004391542170196772, + 0.003353467443957925, + 0.005388485733419657, + 0.005630161613225937, + 0.0027483885642141104, + 0.004485843703150749, + 0.00268356129527092, + 0.0035200435668230057, + 0.0022619664669036865, + 0.003003545571118593, + 0.0031491194386035204, + 0.003001505509018898, + 0.002466606441885233, + 0.003511076094582677, + 0.0035603030119091272, + 0.0026520690880715847, + 0.0031386956106871367, + 0.0021758554503321648, + 0.0025116957258433104, + 0.0036462247371673584, + 0.015809636563062668, + 0.009690769016742706, + 0.014063647948205471, + 0.010232148692011833, + 0.009424379095435143, + 0.008509102277457714, + 0.014895866625010967, + 0.009482919238507748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.043060772120952606 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702159583568573 + ], + "opposite_logprob_tokens": [ + -2.127021551132202 + ], + "logprob_reward": -0.12702159583568573, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702159583568573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07845799624919891, + 0.04688819870352745, + 0.06651877611875534, + 0.03662814572453499, + 0.5511116981506348, + 0.03596879914402962, + 0.05840194597840309, + 0.07156434655189514, + 0.01792273111641407, + 0.036537326872348785 + ], + "credit_map": [ + 0.007834216579794884, + 0.008241403847932816, + 0.008856303989887238, + 0.00782898347824812, + 0.006557598244398832, + 0.009859835729002953, + 0.010984636843204498, + 0.01829502545297146, + 0.00469999061897397, + 0.0050214254297316074, + 0.0032981180120259523, + 0.005049629136919975, + 0.008980910293757915, + 0.005448059644550085, + 0.0063965111039578915, + 0.007993558421730995, + 0.008277595974504948, + 0.0063354820013046265, + 0.004018545616418123, + 0.010019670240581036, + 0.014849849976599216, + 0.008137480355799198, + 0.006882227025926113, + 0.007997935637831688, + 0.0036433746572583914, + 0.008026131428778172, + 0.0027088916394859552, + 0.007026670966297388, + 0.004868279676884413, + 0.004161108285188675, + 0.003409520024433732, + 0.0027841718401759863, + 0.07431133836507797, + 0.0809982568025589, + 0.04535461217164993, + 0.17590557038784027, + 0.030354425311088562, + 0.05756096914410591, + 0.0359356589615345, + 0.050690907984972, + 0.0033205661457031965, + 0.006278382148593664, + 0.004698984324932098, + 0.0048973094671964645, + 0.004453316330909729, + 0.0038242873270064592, + 0.0039014737121760845, + 0.0045944820158183575, + 0.005949381738901138, + 0.008899095468223095, + 0.005703606642782688, + 0.00741280522197485, + 0.00740095367655158, + 0.007114776875823736, + 0.010288878343999386, + 0.005632451269775629, + 0.0071683828718960285, + 0.008748744614422321, + 0.0109195401892066, + 0.010302741080522537, + 0.0057700262404978275, + 0.01115520391613245, + 0.008605580776929855, + 0.00889412872493267, + 0.0021630655974149704, + 0.002117203315719962, + 0.0022262416314333677, + 0.0026377385947853327, + 0.0020498561207205057, + 0.0014206605264917016, + 0.002090094145387411, + 0.003217872232198715, + 0.005555684212595224, + 0.0029599221888929605, + 0.006400496233254671, + 0.00388272013515234, + 0.0029638076666742563, + 0.00354453781619668, + 0.006289745215326548, + 0.004940416198223829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.12702159583568573 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic slow-motion shot, shot from a slight upward angle, the camera slowly pans to the right, capturing a breathtaking aerial duel amidst swirling, inky black storm clouds. A massive, sinuous, cyan dragon clashes fiercely with a muscular white tiger. The dragon, covered in shimmering metallic cyan scales, shimmers brilliantly against the occasional streaks of purple lightning deep within the clouds. Its majestic horns and eyes, ablaze with golden fire, exude a commanding presence. The tiger, with its massive, richly feathered white wings spread wide, its snow-white fur streaked with deep black stripes, its amber pupils gleaming with endless fighting spirit. In extreme slow motion, the dragon brandishes its claws as sharp as a knife, and the tiger extends its tiger paws, which are shining with cold light, to meet the attack with the same swift posture. At the moment when the dragon claws and tiger paws collide violently in the air, a cluster of dazzling golden sparks bursts and splashes. Each spark drags a tiny tail of light, which is clearly visible against the dark background. The lightning casts an ever-changing and dramatic outline light on the fighting beasts, highlighting every muscle tension and burst of strength. The whole picture is full of epic power and mythological tension, freezing this battle between the mythical beasts that changes the world into a moment full of visual impact.", + "seed": 25347, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_6/index_9371_seed_25347_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.043265885580331086, + "vlm_logprob": -0.043265885580331086, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.043265885580331086, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09033973515033722 + ], + "opposite_logprob_tokens": [ + -2.4653396606445312 + ], + "logprob_reward": -0.09033973515033722, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09033973515033722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31376877427101135, + 0.05791071802377701, + 0.045965518802404404, + 0.03480343893170357, + 0.037301816046237946, + 0.019809193909168243, + 0.016903771087527275, + 0.022056620568037033, + 0.4233914613723755, + 0.028088729828596115 + ], + "credit_map": [ + 0.022481929510831833, + 0.04681168124079704, + 0.03448563814163208, + 0.0479445680975914, + 0.04168672487139702, + 0.03782306984066963, + 0.049064140766859055, + 0.0334710069000721, + 0.0072173429653048515, + 0.007070336956530809, + 0.007910856045782566, + 0.00617031566798687, + 0.0064708213321864605, + 0.010468081571161747, + 0.006669697351753712, + 0.005933261010795832, + 0.0048192646354436874, + 0.004316437989473343, + 0.007571709807962179, + 0.005552057176828384, + 0.006214822642505169, + 0.005431594792753458, + 0.005590978544205427, + 0.0064686499536037445, + 0.0032048814464360476, + 0.0049797226674854755, + 0.0047744400799274445, + 0.005568076390773058, + 0.004154282622039318, + 0.005093713290989399, + 0.003764823544770479, + 0.00326349725946784, + 0.005248805973678827, + 0.003664187388494611, + 0.003507893765345216, + 0.00654322374612093, + 0.005282952915877104, + 0.003810297464951873, + 0.006126796826720238, + 0.0031176554039120674, + 0.002012253040447831, + 0.002211077604442835, + 0.0036891947966068983, + 0.003817848162725568, + 0.0015696362825110555, + 0.0022335294634103775, + 0.0016356325941160321, + 0.0026400212664157152, + 0.0017267788061872125, + 0.0021178724709898233, + 0.0013977944618090987, + 0.002035195706412196, + 0.0023629635106772184, + 0.00255601666867733, + 0.002542446367442608, + 0.002164701232686639, + 0.002407630905508995, + 0.0028740845154970884, + 0.0021843742579221725, + 0.002754134824499488, + 0.002827106276527047, + 0.0021862685680389404, + 0.002881234511733055, + 0.003941785544157028, + 0.026983030140399933, + 0.03702966123819351, + 0.04378775879740715, + 0.0678548812866211, + 0.0935017541050911, + 0.048812638968229294, + 0.08441167324781418, + 0.02101004496216774, + 0.004497716668993235, + 0.006334714591503143, + 0.0037763698492199183, + 0.0030934547539800406, + 0.0026461200322955847, + 0.003035984467715025, + 0.002880651503801346, + 0.001823716564103961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.09033973515033722 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004108679015189409 + ], + "opposite_logprob_tokens": [ + -5.504108905792236 + ], + "logprob_reward": -0.004108679015189409, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004108679015189409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36518019437789917, + 0.10319247841835022, + 0.06678063422441483, + 0.04447248578071594, + 0.04525567963719368, + 0.03963285684585571, + 0.06293188780546188, + 0.06571722030639648, + 0.17271848022937775, + 0.034117959439754486 + ], + "credit_map": [ + 0.02819543331861496, + 0.048822563141584396, + 0.05559294670820236, + 0.047413378953933716, + 0.05044957622885704, + 0.045616090297698975, + 0.05010416731238365, + 0.03898607939481735, + 0.009232363663613796, + 0.019793620333075523, + 0.0134081756696105, + 0.01150799635797739, + 0.011402246542274952, + 0.01660527102649212, + 0.011963638477027416, + 0.009279179386794567, + 0.010302471928298473, + 0.008442667312920094, + 0.0063912030309438705, + 0.007272943388670683, + 0.008097002282738686, + 0.006649848539382219, + 0.007537115830928087, + 0.012087389826774597, + 0.0038699971046298742, + 0.005098688881844282, + 0.00508772861212492, + 0.005779968574643135, + 0.006104219239205122, + 0.005460534244775772, + 0.007479103282094002, + 0.005592250265181065, + 0.004065246321260929, + 0.006871912628412247, + 0.004583701025694609, + 0.006990745663642883, + 0.006171842105686665, + 0.004984014201909304, + 0.007539365906268358, + 0.004048858769237995, + 0.003962286747992039, + 0.006239189766347408, + 0.005736636929214001, + 0.004452851600944996, + 0.004383067134767771, + 0.004281036090105772, + 0.003660112852230668, + 0.0069176796823740005, + 0.00615732790902257, + 0.008223949931561947, + 0.006801632232964039, + 0.0077088214457035065, + 0.00877596065402031, + 0.007759391330182552, + 0.010077003389596939, + 0.007427805103361607, + 0.006744343787431717, + 0.010732641443610191, + 0.005954609718173742, + 0.011342327110469341, + 0.008248315192759037, + 0.006000687833875418, + 0.007981766946613789, + 0.008712534792721272, + 0.014331003651022911, + 0.02893051505088806, + 0.019974349066615105, + 0.05398900806903839, + 0.013125566765666008, + 0.01581301912665367, + 0.017782753333449364, + 0.0087722884491086, + 0.004966858774423599, + 0.005309335421770811, + 0.0046575628221035, + 0.0043088519014418125, + 0.0033316940534859896, + 0.003923946525901556, + 0.004869998432695866, + 0.0027497143018990755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.004108679015189409 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018191752955317497 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.018191752955317497, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018191752955317497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2290472537279129, + 0.06805263459682465, + 0.06049642339348793, + 0.05379444360733032, + 0.05364873632788658, + 0.06467312574386597, + 0.054564304649829865, + 0.07683878391981125, + 0.23493191599845886, + 0.10395238548517227 + ], + "credit_map": [ + 0.018988246098160744, + 0.030452776700258255, + 0.03978639841079712, + 0.0281731728464365, + 0.03311906009912491, + 0.026494678109884262, + 0.021166617050766945, + 0.030866289511322975, + 0.006449984852224588, + 0.00824875570833683, + 0.008363425731658936, + 0.006791966501623392, + 0.00811418704688549, + 0.012746002525091171, + 0.008080177940428257, + 0.009258134290575981, + 0.007750581484287977, + 0.009840759448707104, + 0.0070082624442875385, + 0.0076505993492901325, + 0.0074891457334160805, + 0.00641455315053463, + 0.007190775126218796, + 0.00715174525976181, + 0.004654040560126305, + 0.005312194116413593, + 0.006200527306646109, + 0.008147751912474632, + 0.007537050172686577, + 0.006605603266507387, + 0.010600906796753407, + 0.0047363704070448875, + 0.006212539505213499, + 0.00716647831723094, + 0.005275294650346041, + 0.0096819419413805, + 0.007303944788873196, + 0.005767310969531536, + 0.007738356012851, + 0.004502872470766306, + 0.004626709967851639, + 0.006423633079975843, + 0.010014476254582405, + 0.014587770216166973, + 0.0045457142405211926, + 0.005975859705358744, + 0.005517260637134314, + 0.012981705367565155, + 0.0042887525632977486, + 0.007638336159288883, + 0.00502486526966095, + 0.006065201014280319, + 0.011673773638904095, + 0.004710916429758072, + 0.00872237328439951, + 0.006440085358917713, + 0.00841422751545906, + 0.014368850737810135, + 0.005569828674197197, + 0.010797394439578056, + 0.008431047201156616, + 0.007955102249979973, + 0.011300110258162022, + 0.010002224706113338, + 0.019106747582554817, + 0.03592242673039436, + 0.019701344892382622, + 0.04311062768101692, + 0.04217849299311638, + 0.04534274339675903, + 0.018995802849531174, + 0.010573743842542171, + 0.010420873761177063, + 0.015575796365737915, + 0.009839684702455997, + 0.015359352342784405, + 0.012601207010447979, + 0.012163601815700531, + 0.019329087808728218, + 0.008662777952849865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.018191752955317497 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033689454197883606 + ], + "opposite_logprob_tokens": [ + -3.408689498901367 + ], + "logprob_reward": -0.033689454197883606, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033689454197883606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1564691960811615, + 0.09852776676416397, + 0.18479953706264496, + 0.07538634538650513, + 0.06260370463132858, + 0.1144983172416687, + 0.04168602451682091, + 0.06406611949205399, + 0.1395835131406784, + 0.062379419803619385 + ], + "credit_map": [ + 0.014360027387738228, + 0.018558379262685776, + 0.023993173614144325, + 0.022023096680641174, + 0.017435362562537193, + 0.02270425669848919, + 0.019472636282444, + 0.01792227476835251, + 0.010429004207253456, + 0.010064832866191864, + 0.01625991426408291, + 0.010662674903869629, + 0.01103219948709011, + 0.01983882114291191, + 0.010259860195219517, + 0.009980469942092896, + 0.016657039523124695, + 0.023195190355181694, + 0.027825424447655678, + 0.024174032732844353, + 0.020862707868218422, + 0.022041920572519302, + 0.025739073753356934, + 0.024304164573550224, + 0.006118943449109793, + 0.010688030160963535, + 0.010762812569737434, + 0.01395124476402998, + 0.008992573246359825, + 0.008597237057983875, + 0.010535506531596184, + 0.005740002728998661, + 0.007415694184601307, + 0.006364033091813326, + 0.007329028565436602, + 0.010598735883831978, + 0.009974083863198757, + 0.007348432205617428, + 0.008304126560688019, + 0.00526956794783473, + 0.00721754739060998, + 0.010645483620464802, + 0.029158469289541245, + 0.03869640827178955, + 0.0050750053487718105, + 0.006350555922836065, + 0.00820941012352705, + 0.009145449846982956, + 0.003667144337669015, + 0.005384344141930342, + 0.004396301694214344, + 0.005560882855206728, + 0.006586873438209295, + 0.0038324054330587387, + 0.0062009538523852825, + 0.00605712179094553, + 0.007624167948961258, + 0.006841489113867283, + 0.007499865721911192, + 0.01830659992992878, + 0.003440050408244133, + 0.007476279512047768, + 0.0061081694439053535, + 0.006769501604139805, + 0.011077937670052052, + 0.025868995115160942, + 0.009161611087620258, + 0.015161664225161076, + 0.0125419395044446, + 0.035370491445064545, + 0.02392863854765892, + 0.00647225184366107, + 0.00914977677166462, + 0.012532174587249756, + 0.007317803800106049, + 0.008046230301260948, + 0.006152068730443716, + 0.006331938784569502, + 0.007065423298627138, + 0.0057840123772621155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.033689454197883606 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0699998065829277 + ], + "opposite_logprob_tokens": [ + -2.6949996948242188 + ], + "logprob_reward": -0.0699998065829277, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0699998065829277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11473540961742401, + 0.060988012701272964, + 0.07335073500871658, + 0.031166547909379005, + 0.04511973261833191, + 0.20366834104061127, + 0.08172658830881119, + 0.06165299937129021, + 0.26491180062294006, + 0.06267981231212616 + ], + "credit_map": [ + 0.00969803985208273, + 0.015192244201898575, + 0.022419428452849388, + 0.013321583159267902, + 0.01228599064052105, + 0.015052389353513718, + 0.013034861534833908, + 0.013730873353779316, + 0.004236248787492514, + 0.01022761594504118, + 0.007309557870030403, + 0.008260304108262062, + 0.008265203796327114, + 0.008487094193696976, + 0.007913847453892231, + 0.006288138683885336, + 0.00530208507552743, + 0.010493932291865349, + 0.007629085332155228, + 0.008047981187701225, + 0.009868449531495571, + 0.007717072498053312, + 0.010102391242980957, + 0.014189730398356915, + 0.0044046915136277676, + 0.004118451848626137, + 0.003353809705004096, + 0.004485595505684614, + 0.0050715673714876175, + 0.003931629937142134, + 0.003247022395953536, + 0.0025537796318531036, + 0.004451678600162268, + 0.005886695813387632, + 0.004901154898107052, + 0.007396926637738943, + 0.005497857928276062, + 0.00467492314055562, + 0.007684851065278053, + 0.004625646863132715, + 0.0183509960770607, + 0.019245285540819168, + 0.03830558806657791, + 0.05685940012335777, + 0.020653001964092255, + 0.01392021682113409, + 0.016185346990823746, + 0.0201485026627779, + 0.0064280666410923, + 0.01333346962928772, + 0.009268105030059814, + 0.009225288406014442, + 0.014771866612136364, + 0.00549935270100832, + 0.016433630138635635, + 0.006766807287931442, + 0.006251777056604624, + 0.009545840322971344, + 0.006221569608896971, + 0.012047050520777702, + 0.00410632137209177, + 0.008241220377385616, + 0.0066132997162640095, + 0.008625921793282032, + 0.02333018369972706, + 0.07866590470075607, + 0.059073179960250854, + 0.023193247616291046, + 0.016548210754990578, + 0.042799729853868484, + 0.013555100187659264, + 0.007746228016912937, + 0.00661996565759182, + 0.00978903193026781, + 0.006006828509271145, + 0.008899643085896969, + 0.0065590753220021725, + 0.010341199114918709, + 0.008634976111352444, + 0.005829090252518654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.0699998065829277 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic slow-motion shot, shot from a slight upward angle, the camera slowly pans to the right, capturing a breathtaking aerial duel amidst swirling, inky black storm clouds. A massive, sinuous, cyan dragon clashes fiercely with a muscular white tiger. The dragon, covered in shimmering metallic cyan scales, shimmers brilliantly against the occasional streaks of purple lightning deep within the clouds. Its majestic horns and eyes, ablaze with golden fire, exude a commanding presence. The tiger, with its massive, richly feathered white wings spread wide, its snow-white fur streaked with deep black stripes, its amber pupils gleaming with endless fighting spirit. In extreme slow motion, the dragon brandishes its claws as sharp as a knife, and the tiger extends its tiger paws, which are shining with cold light, to meet the attack with the same swift posture. At the moment when the dragon claws and tiger paws collide violently in the air, a cluster of dazzling golden sparks bursts and splashes. Each spark drags a tiny tail of light, which is clearly visible against the dark background. The lightning casts an ever-changing and dramatic outline light on the fighting beasts, highlighting every muscle tension and burst of strength. The whole picture is full of epic power and mythological tension, freezing this battle between the mythical beasts that changes the world into a moment full of visual impact.", + "seed": 25348, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_6/index_9371_seed_25348_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.4167029220610857, + "vlm_logprob": -0.4167029220610857, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.4167029220610857, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9106123447418213 + ], + "opposite_logprob_tokens": [ + -0.16061238944530487 + ], + "logprob_reward": -1.9106123447418213, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9106123447418213, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24538013339042664, + 0.5596321821212769, + 0.012100142426788807, + 0.006528130266815424, + 0.005748946685343981, + 0.0673474669456482, + 0.035013001412153244, + 0.037529293447732925, + 0.019503312185406685, + 0.011217379942536354 + ], + "credit_map": [ + 0.023408589884638786, + 0.03729857876896858, + 0.01580905169248581, + 0.028364190831780434, + 0.02470804937183857, + 0.03597905859351158, + 0.040810320526361465, + 0.03900229558348656, + 0.15266463160514832, + 0.03540431708097458, + 0.21881026029586792, + 0.029688123613595963, + 0.03549877554178238, + 0.04719448462128639, + 0.02471233159303665, + 0.01565924845635891, + 0.0014868390280753374, + 0.001751747215166688, + 0.0017060480313375592, + 0.0015124126803129911, + 0.0010934140300378203, + 0.0016601932002231479, + 0.001191383576951921, + 0.001698105363175273, + 0.0010018831817433238, + 0.0010053091682493687, + 0.0007104347459971905, + 0.0007364943157881498, + 0.0011040451936423779, + 0.0006549658719450235, + 0.0006850644713267684, + 0.0006299333181232214, + 0.0008903030538931489, + 0.0005067351157777011, + 0.0005483662826009095, + 0.0008188463980332017, + 0.001054457388818264, + 0.00042287283577024937, + 0.0008156927651725709, + 0.000691673019900918, + 0.02463381364941597, + 0.002841387642547488, + 0.00292870937846601, + 0.002498716814443469, + 0.0031206400599330664, + 0.0032048551365733147, + 0.02543582022190094, + 0.0026835256721824408, + 0.0039927237667143345, + 0.0038844686932861805, + 0.0029194005765020847, + 0.004543533083051443, + 0.005528321955353022, + 0.003688088385388255, + 0.007060402072966099, + 0.0033960656728595495, + 0.0032144144643098116, + 0.012675032019615173, + 0.002189185470342636, + 0.0023110967595130205, + 0.003153775120154023, + 0.002576275262981653, + 0.002652843715623021, + 0.008756672032177448, + 0.0036891077179461718, + 0.0029941541142761707, + 0.003216448938474059, + 0.001652092789299786, + 0.0021852885838598013, + 0.0013884622603654861, + 0.002205734606832266, + 0.0021720225922763348, + 0.0015284297987818718, + 0.0009763887501321733, + 0.0015117712318897247, + 0.0015443048905581236, + 0.00162770866882056, + 0.001446432899683714, + 0.0011697477893903852, + 0.0014125964371487498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -1.9106123447418213 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007676861248910427 + ], + "opposite_logprob_tokens": [ + -4.88267707824707 + ], + "logprob_reward": -0.007676861248910427, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007676861248910427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25220829248428345, + 0.5543651580810547, + 0.014416910707950592, + 0.006807683035731316, + 0.007108620833605528, + 0.04239209368824959, + 0.037528734654188156, + 0.05002996325492859, + 0.015808477997779846, + 0.01933411881327629 + ], + "credit_map": [ + 0.024191129952669144, + 0.03860654681921005, + 0.016333989799022675, + 0.029590822756290436, + 0.02547018975019455, + 0.03629904240369797, + 0.042139433324337006, + 0.03957710415124893, + 0.15167459845542908, + 0.035206809639930725, + 0.21617156267166138, + 0.029435239732265472, + 0.03509613499045372, + 0.04668297618627548, + 0.024536732584238052, + 0.015561023727059364, + 0.0017001983942463994, + 0.0020286531653255224, + 0.0020801154896616936, + 0.0021060912404209375, + 0.001324268989264965, + 0.0017843531677499413, + 0.001505209831520915, + 0.0018880196148529649, + 0.0010274425148963928, + 0.0010480335913598537, + 0.0006848606863059103, + 0.0007543755928054452, + 0.0012001658324152231, + 0.00067029899219051, + 0.000739447888918221, + 0.0006830570637248456, + 0.0009617242030799389, + 0.0006519283051602542, + 0.0006879005813971162, + 0.0010122552048414946, + 0.0010895964223891497, + 0.00046937790466472507, + 0.0011733252322301269, + 0.0010625122813507915, + 0.015357245691120625, + 0.0018129869131371379, + 0.0015698513016104698, + 0.0016302926233038306, + 0.0022312551736831665, + 0.002138792769983411, + 0.0159105584025383, + 0.0017411063890904188, + 0.004143792670220137, + 0.004236713983118534, + 0.0030002095736563206, + 0.004652839619666338, + 0.005768465343862772, + 0.004033423960208893, + 0.007907952181994915, + 0.0037853331305086613, + 0.0036719033960253, + 0.01874818094074726, + 0.002008546143770218, + 0.0028303114231675863, + 0.0038353169802576303, + 0.002941570710390806, + 0.0026608859188854694, + 0.013333242386579514, + 0.0029880315996706486, + 0.002473992994055152, + 0.0024304636754095554, + 0.0014827055856585503, + 0.0018719007493928075, + 0.0012972300173714757, + 0.0016975841717794538, + 0.001566567923873663, + 0.003570819040760398, + 0.0021365107968449593, + 0.0034528684336692095, + 0.0022872532717883587, + 0.001862929086200893, + 0.0015048508066684008, + 0.0018720926018431783, + 0.0026467922143638134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.007676861248910427 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198708347976208 + ], + "opposite_logprob_tokens": [ + -4.2641987800598145 + ], + "logprob_reward": -0.014198708347976208, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198708347976208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1819300651550293, + 0.4298257827758789, + 0.01256074570119381, + 0.004406650550663471, + 0.01232187170535326, + 0.12334038317203522, + 0.03328779339790344, + 0.1124119982123375, + 0.013775808736681938, + 0.07613879442214966 + ], + "credit_map": [ + 0.01771356165409088, + 0.027404887601733208, + 0.011719495058059692, + 0.02129109762609005, + 0.01843010075390339, + 0.026431135833263397, + 0.030630428344011307, + 0.028309360146522522, + 0.11764227598905563, + 0.027169808745384216, + 0.16767223179340363, + 0.022779956459999084, + 0.027116572484374046, + 0.03627585247159004, + 0.01912928931415081, + 0.012039825320243835, + 0.0014508318854495883, + 0.0016823798650875688, + 0.0019891655538231134, + 0.0018906710902228951, + 0.0011752491118386388, + 0.0015471377409994602, + 0.0012900226283818483, + 0.0015352880582213402, + 0.0006231477600522339, + 0.0006769553874619305, + 0.000512470374815166, + 0.0005487738526426256, + 0.0006557347951456904, + 0.00039544302853755653, + 0.00047083949903026223, + 0.0005232858820818365, + 0.0016556939808651805, + 0.0015276585472747684, + 0.0015625167870894074, + 0.002226522658020258, + 0.0012881598668172956, + 0.000824553775601089, + 0.001759020029567182, + 0.0014777460601180792, + 0.04451692849397659, + 0.00533655658364296, + 0.005054065492004156, + 0.0048851859755814075, + 0.00556393526494503, + 0.005632809363305569, + 0.047245875000953674, + 0.00510502839460969, + 0.0033999059814959764, + 0.003122658235952258, + 0.0024964644107967615, + 0.0039224568754434586, + 0.004336681682616472, + 0.003761849133297801, + 0.008690092712640762, + 0.0035576827358454466, + 0.007471838500350714, + 0.04329095780849457, + 0.004738585092127323, + 0.0065057361498475075, + 0.008551718667149544, + 0.006635698955506086, + 0.00484887883067131, + 0.030368594452738762, + 0.0021681988146156073, + 0.0015722722746431828, + 0.0021988824009895325, + 0.001374880550429225, + 0.0018213525181636214, + 0.001365781994536519, + 0.0017300491454079747, + 0.0015443909214809537, + 0.015337666496634483, + 0.009355493821203709, + 0.01376657746732235, + 0.007932529784739017, + 0.006677227560430765, + 0.0051664263010025024, + 0.008088788017630577, + 0.009814084507524967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.014198708347976208 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038100630044937134 + ], + "opposite_logprob_tokens": [ + -3.2881007194519043 + ], + "logprob_reward": -0.038100630044937134, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038100630044937134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22936049103736877, + 0.6230643391609192, + 0.018411660566926003, + 0.006768002640455961, + 0.01748889498412609, + 0.01873444765806198, + 0.035803355276584625, + 0.01647435873746872, + 0.011282662861049175, + 0.0226118266582489 + ], + "credit_map": [ + 0.02195661887526512, + 0.03520552068948746, + 0.01457869540899992, + 0.026714636012911797, + 0.022782698273658752, + 0.03297938406467438, + 0.03779330104589462, + 0.03734961152076721, + 0.17004723846912384, + 0.039159152656793594, + 0.24363191425800323, + 0.03298792615532875, + 0.039261531084775925, + 0.05256205052137375, + 0.02783442661166191, + 0.017580049112439156, + 0.0020742544438689947, + 0.00238027423620224, + 0.0030539350118488073, + 0.002758560236543417, + 0.0015479364665225148, + 0.0021558618173003197, + 0.0022014908026903868, + 0.0022393467370420694, + 0.0010019888868555427, + 0.0009484969195909798, + 0.0008628420764580369, + 0.0008607179624959826, + 0.0010856162989512086, + 0.0006377209210768342, + 0.0006522667245008051, + 0.0007183520356193185, + 0.0023828651756048203, + 0.0012441694270819426, + 0.001921769930049777, + 0.002056807978078723, + 0.002232582541182637, + 0.0017168805934488773, + 0.0033780066296458244, + 0.0025558106135576963, + 0.0025035012513399124, + 0.0019558260682970285, + 0.002772115170955658, + 0.003172994125634432, + 0.002762681106105447, + 0.002117839874699712, + 0.0018709386931732297, + 0.0015785491559654474, + 0.0039802733808755875, + 0.003813355229794979, + 0.002809269120916724, + 0.004180918913334608, + 0.006375476252287626, + 0.0038271709345281124, + 0.007340009789913893, + 0.003476874902844429, + 0.0018908511847257614, + 0.004315164405852556, + 0.0011870772577822208, + 0.0013414602726697922, + 0.0013978298520669341, + 0.0013981573283672333, + 0.00162379106041044, + 0.003320025047287345, + 0.001922283205203712, + 0.001535894232802093, + 0.0017861726228147745, + 0.0011521636042743921, + 0.0015157052548602223, + 0.0009612999274395406, + 0.0012290353188291192, + 0.0011801072396337986, + 0.0038074422627687454, + 0.0025328914634883404, + 0.004028976429253817, + 0.002648403402417898, + 0.002305535599589348, + 0.0018637081375345588, + 0.0023311551194638014, + 0.00309371342882514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.038100630044937134 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set amidst swirling, inky black storm clouds with purple lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive cyan dragon with shimmering metallic scales and golden fiery eyes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a muscular white tiger with richly feathered white wings and black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragon's claws and the tiger's paws collide violently creating golden sparks with visible light tails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a slight upward angle with the camera slowly panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292606592178345 + ], + "opposite_logprob_tokens": [ + -2.2379260063171387 + ], + "logprob_reward": -0.11292606592178345, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292606592178345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1420190930366516, + 0.4965326488018036, + 0.023141926154494286, + 0.0235135518014431, + 0.02480960264801979, + 0.13335706293582916, + 0.04732184112071991, + 0.05112389102578163, + 0.03995794430375099, + 0.018222356215119362 + ], + "credit_map": [ + 0.01383634377270937, + 0.017793400213122368, + 0.009434713050723076, + 0.016349026933312416, + 0.01581740379333496, + 0.02087932638823986, + 0.02508922852575779, + 0.022819653153419495, + 0.13516779243946075, + 0.03143119812011719, + 0.19370298087596893, + 0.026338163763284683, + 0.031375009566545486, + 0.04176587983965874, + 0.0225375909358263, + 0.014214083552360535, + 0.0023802861105650663, + 0.00355063215829432, + 0.0027336599305272102, + 0.0030071334913372993, + 0.0026902537792921066, + 0.0028630502056330442, + 0.0030051395297050476, + 0.0029117725789546967, + 0.0030156506691128016, + 0.0036332940217107534, + 0.0033143681939691305, + 0.00201685749925673, + 0.0034427193459123373, + 0.002236106898635626, + 0.002922465791925788, + 0.002932088915258646, + 0.0030954766552895308, + 0.002039647428318858, + 0.0018631400307640433, + 0.0022849165834486485, + 0.00275592808611691, + 0.002547296928241849, + 0.004415954928845167, + 0.005807244218885899, + 0.045276861637830734, + 0.006624490488320589, + 0.008675153367221355, + 0.007844001054763794, + 0.006355402525514364, + 0.005814766976982355, + 0.046843837946653366, + 0.0059225535951554775, + 0.00551625806838274, + 0.005104992538690567, + 0.0036691881250590086, + 0.0051402682438492775, + 0.010995843447744846, + 0.0059236763045191765, + 0.005926595069468021, + 0.005045023281127214, + 0.0036120470613241196, + 0.01640675589442253, + 0.0028677501250058413, + 0.003314822679385543, + 0.004558356013149023, + 0.0037736406084150076, + 0.003326257923617959, + 0.013264261186122894, + 0.004421130288392305, + 0.004947689361870289, + 0.005505560431629419, + 0.005268369801342487, + 0.005960858426988125, + 0.004431870765984058, + 0.005600434727966785, + 0.0038220288697630167, + 0.005097299348562956, + 0.0020445124246180058, + 0.002875053556635976, + 0.0016413137782365084, + 0.0014242022298276424, + 0.0010843481868505478, + 0.0012142221676185727, + 0.002841407200321555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.11292606592178345 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_7/index_tensor([883, 883])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_7/index_tensor([883, 883])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..30d0c52d58c4a68aaeae3895eef9664ada9585e1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_7/index_tensor([883, 883])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, deliberate 360-degree orbiting shot begins, the camera moving steadily at eye level around a magnificent, sun-warmed peach tree. The scene is rendered with a deeply natural and earthy style, emphasizing organic textures. The tree's gnarled, weathered bark creates a tactile contrast with its canopy of lush, dark green leaves, still shimmering with the lingering crystal droplets of a recent rain shower. Its branches hang heavy with an abundance of ripe fruit; each peach is a gradient of color, from pale yellow to deep rose, its skin covered in a delicate, furry velvet that captures the flecked golden sunlight filtering through the leaves. As the camera continues its graceful and steady rotation, the pastoral sky in the background reveals itself, a canvas of soft blue and broken white clouds. Then, a stunning, vibrant, full-spectrum rainbow appears, shimmering against the clear sky, its perfect semicircle framing the tree at the peak of the camera's turn, creating an atmosphere of profound tranquility and wonder.", + "seed": 4565, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_7/index_883_seed_4565_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.060538627579808234, + "vlm_logprob": -0.060538627579808234, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.060538627579808234, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16769728064537048 + ], + "opposite_logprob_tokens": [ + -1.9176973104476929 + ], + "logprob_reward": -0.16769728064537048, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16769728064537048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0842192992568016, + 0.10061150044202805, + 0.11217830330133438, + 0.1661055088043213, + 0.18271678686141968, + 0.15341287851333618, + 0.019858207553625107, + 0.01137549988925457, + 0.1478738635778427, + 0.021648209542036057 + ], + "credit_map": [ + 0.012770541943609715, + 0.015729818493127823, + 0.011930647306144238, + 0.012009642086923122, + 0.012278101406991482, + 0.006330474745482206, + 0.00770873436704278, + 0.005461333319544792, + 0.010947433300316334, + 0.013389899395406246, + 0.007675380911678076, + 0.014204547740519047, + 0.010210458189249039, + 0.008498365059494972, + 0.02498374506831169, + 0.010701660066843033, + 0.007123383227735758, + 0.031108802184462547, + 0.008433878421783447, + 0.019128162413835526, + 0.00890961941331625, + 0.012577801011502743, + 0.01925174705684185, + 0.005644909106194973, + 0.01599431224167347, + 0.020608549937605858, + 0.01293796207755804, + 0.03625702112913132, + 0.014650796540081501, + 0.015167752280831337, + 0.03901512548327446, + 0.01147399004548788, + 0.01977609097957611, + 0.02241385169327259, + 0.014359094202518463, + 0.059791356325149536, + 0.01346167828887701, + 0.01774509623646736, + 0.023628203198313713, + 0.011541411280632019, + 0.01377859152853489, + 0.04403374716639519, + 0.022373443469405174, + 0.01975676417350769, + 0.014814579859375954, + 0.013242794200778008, + 0.013720305636525154, + 0.011692659929394722, + 0.001513139228336513, + 0.002482897602021694, + 0.0028351640794426203, + 0.0026656740810722113, + 0.0026614065282046795, + 0.0022428531665354967, + 0.0029195514507591724, + 0.0025375213008373976, + 0.0009535832214169204, + 0.0016158755170181394, + 0.0010740893194451928, + 0.0013700496638193727, + 0.0014285281067714095, + 0.0016800605226308107, + 0.0019399491138756275, + 0.0013133639004081488, + 0.0050336625427007675, + 0.003654817119240761, + 0.00382760027423501, + 0.007899895310401917, + 0.008733056485652924, + 0.05571361631155014, + 0.007390165701508522, + 0.05562104284763336, + 0.0015640128403902054, + 0.0016558917704969645, + 0.002518927212804556, + 0.0022122517693787813, + 0.0030457787215709686, + 0.004818062763661146, + 0.0024548093788325787, + 0.003378473687916994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.16769728064537048 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012552776373922825 + ], + "opposite_logprob_tokens": [ + -4.387552738189697 + ], + "logprob_reward": -0.012552776373922825, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012552776373922825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16811642050743103, + 0.1368570327758789, + 0.11497712880373001, + 0.062362633645534515, + 0.07789836078882217, + 0.06149819493293762, + 0.11256370693445206, + 0.045330680906772614, + 0.15953980386257172, + 0.060856014490127563 + ], + "credit_map": [ + 0.018520088866353035, + 0.02375810034573078, + 0.018115177750587463, + 0.030712436884641647, + 0.018329409882426262, + 0.012560905888676643, + 0.03186439722776413, + 0.014255901798605919, + 0.01638440042734146, + 0.025153614580631256, + 0.015624278225004673, + 0.01998417265713215, + 0.012900948524475098, + 0.013256673701107502, + 0.021253788843750954, + 0.01229915302246809, + 0.009795852936804295, + 0.03141720965504646, + 0.009149393066763878, + 0.01662261411547661, + 0.01192693505436182, + 0.017787514254450798, + 0.011022089049220085, + 0.00725552998483181, + 0.004311331082135439, + 0.007805518340319395, + 0.0036436039954423904, + 0.013563957996666431, + 0.004686189349740744, + 0.006036967970430851, + 0.013956599868834019, + 0.008358464576303959, + 0.00772740738466382, + 0.013329848647117615, + 0.0070802136324346066, + 0.01715698465704918, + 0.008822663687169552, + 0.006760170217603445, + 0.010362125933170319, + 0.006658948492258787, + 0.005127656273543835, + 0.01705210655927658, + 0.007326249498873949, + 0.008216438814997673, + 0.00566565478220582, + 0.005147089250385761, + 0.008577268570661545, + 0.004385731182992458, + 0.008092190138995647, + 0.006760783027857542, + 0.015171046368777752, + 0.023409992456436157, + 0.011710315011441708, + 0.013674427755177021, + 0.026211079210042953, + 0.007533874828368425, + 0.00313219940289855, + 0.005063832737505436, + 0.004548800643533468, + 0.005290609318763018, + 0.005973357241600752, + 0.006517552305012941, + 0.007668693084269762, + 0.007135636173188686, + 0.005939624737948179, + 0.004892816301435232, + 0.0036123290192335844, + 0.0089784050360322, + 0.010178450495004654, + 0.05764239281415939, + 0.00858532078564167, + 0.059710465371608734, + 0.005453422199934721, + 0.004178791306912899, + 0.006348065100610256, + 0.006854955106973648, + 0.011099748313426971, + 0.0123156001791358, + 0.00722180912271142, + 0.007383621297776699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.012552776373922825 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107415184378624 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.01107415184378624, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107415184378624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16818031668663025, + 0.11928264051675797, + 0.04392360523343086, + 0.07857957482337952, + 0.11585782468318939, + 0.13362275063991547, + 0.06584038585424423, + 0.03587016835808754, + 0.17912772297859192, + 0.05971501022577286 + ], + "credit_map": [ + 0.021836990490555763, + 0.02329530008137226, + 0.015691453590989113, + 0.028409933671355247, + 0.020326046273112297, + 0.012454385869204998, + 0.03085976652801037, + 0.015306444838643074, + 0.014840700663626194, + 0.014666211791336536, + 0.012203378602862358, + 0.0157115887850523, + 0.01634983718395233, + 0.009776157326996326, + 0.026158971711993217, + 0.009575795382261276, + 0.0037872169632464647, + 0.007197343744337559, + 0.003977911081165075, + 0.008213610388338566, + 0.0056244689039886, + 0.003059364389628172, + 0.008784250356256962, + 0.0032794380094856024, + 0.0069643594324588776, + 0.008479380048811436, + 0.006446501240134239, + 0.01648811064660549, + 0.006557169836014509, + 0.006736556068062782, + 0.018563859164714813, + 0.00834363792091608, + 0.012643503956496716, + 0.019309261813759804, + 0.009487389586865902, + 0.02983442321419716, + 0.011242452077567577, + 0.010787555016577244, + 0.014688695780932903, + 0.007864540442824364, + 0.011286362074315548, + 0.036947913467884064, + 0.018787512555718422, + 0.017870843410491943, + 0.01233498565852642, + 0.010350351221859455, + 0.015011676587164402, + 0.011033105663955212, + 0.004853257909417152, + 0.0071506192907691, + 0.01035525742918253, + 0.011158708482980728, + 0.007487436756491661, + 0.009066205471754074, + 0.010451359674334526, + 0.005317541770637035, + 0.0029560052789747715, + 0.0036845977883785963, + 0.0036449895706027746, + 0.004549641162157059, + 0.004473397973924875, + 0.006825888529419899, + 0.004653776530176401, + 0.005081873387098312, + 0.0066413492895662785, + 0.004757903050631285, + 0.0037132268771529198, + 0.01009189710021019, + 0.011065518483519554, + 0.06565016508102417, + 0.00938885472714901, + 0.06781880557537079, + 0.005654627922922373, + 0.005693311803042889, + 0.0060822744853794575, + 0.009596449322998524, + 0.008369944989681244, + 0.0076644690707325935, + 0.010007591918110847, + 0.00664634071290493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.01107415184378624 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1002974882721901 + ], + "opposite_logprob_tokens": [ + -2.350297451019287 + ], + "logprob_reward": -0.1002974882721901, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1002974882721901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2501995265483856, + 0.07299022376537323, + 0.05990772694349289, + 0.049134042114019394, + 0.0960702896118164, + 0.08471401035785675, + 0.05362289026379585, + 0.058849845081567764, + 0.24287882447242737, + 0.031632598489522934 + ], + "credit_map": [ + 0.03508657589554787, + 0.03377769887447357, + 0.02707832120358944, + 0.03824147582054138, + 0.04275796562433243, + 0.021538810804486275, + 0.03095674142241478, + 0.02076195739209652, + 0.011551057919859886, + 0.01108894869685173, + 0.006063036620616913, + 0.009359501302242279, + 0.009510920383036137, + 0.00734319444745779, + 0.01035265251994133, + 0.0077209193259477615, + 0.004061487969011068, + 0.015327219851315022, + 0.004244605544954538, + 0.011696337722241879, + 0.0048237754963338375, + 0.005645806901156902, + 0.011098899878561497, + 0.0030096019618213177, + 0.002942865714430809, + 0.008125748485326767, + 0.0035898841451853514, + 0.011300789192318916, + 0.005615707486867905, + 0.004066657740622759, + 0.008162827230989933, + 0.0053295595571398735, + 0.010027258656919003, + 0.012469080276787281, + 0.007897790521383286, + 0.03179143741726875, + 0.007508972659707069, + 0.008763505145907402, + 0.011369079351425171, + 0.006243172567337751, + 0.006934004370123148, + 0.017326518893241882, + 0.007354019209742546, + 0.013736365363001823, + 0.01204706635326147, + 0.006956557743251324, + 0.014833899214863777, + 0.005525580141693354, + 0.003908978309482336, + 0.0054964725859463215, + 0.004916525911539793, + 0.007910506799817085, + 0.006839272566139698, + 0.00879607442766428, + 0.008192219771444798, + 0.007562840823084116, + 0.003227138426154852, + 0.005531465169042349, + 0.00497393636032939, + 0.006248687859624624, + 0.0066684638150036335, + 0.01205255463719368, + 0.010846284218132496, + 0.009301322512328625, + 0.008642055094242096, + 0.004848834592849016, + 0.006366017274558544, + 0.012999325059354305, + 0.014517264440655708, + 0.09287862479686737, + 0.011417316272854805, + 0.09120938181877136, + 0.002438398776575923, + 0.0027021467685699463, + 0.0029825104866176844, + 0.004275702405720949, + 0.0051127271726727486, + 0.006477204151451588, + 0.004459655378013849, + 0.003184252418577671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.1002974882721901 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071440763771534 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.011071440763771534, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071440763771534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11399656534194946, + 0.10043736547231674, + 0.08138241618871689, + 0.05297571048140526, + 0.07144042104482651, + 0.07192805409431458, + 0.07008039206266403, + 0.030451852828264236, + 0.3331541121006012, + 0.07415309548377991 + ], + "credit_map": [ + 0.019086051732301712, + 0.01938568614423275, + 0.010965791530907154, + 0.017140695825219154, + 0.011780540458858013, + 0.01368048693984747, + 0.012072166427969933, + 0.009885167703032494, + 0.009814970195293427, + 0.01402114424854517, + 0.009093104861676693, + 0.016681067645549774, + 0.0077675688080489635, + 0.009136929176747799, + 0.023884281516075134, + 0.010038310661911964, + 0.007080327719449997, + 0.0161861814558506, + 0.008021682500839233, + 0.01802721433341503, + 0.005998204927891493, + 0.005995229817926884, + 0.013439859263598919, + 0.006633720826357603, + 0.00332982512190938, + 0.009699172340333462, + 0.0036818080116063356, + 0.009618639014661312, + 0.00498606264591217, + 0.00515024084597826, + 0.011996136978268623, + 0.004513832740485668, + 0.007019881159067154, + 0.009563164785504341, + 0.006230376660823822, + 0.01145000010728836, + 0.012103894725441933, + 0.00907923188060522, + 0.010247423313558102, + 0.005746449809521437, + 0.0053483485244214535, + 0.008735756389796734, + 0.010226947255432606, + 0.012015430256724358, + 0.007913948968052864, + 0.00739105511456728, + 0.014381363987922668, + 0.0059152147732675076, + 0.005370141007006168, + 0.009301837533712387, + 0.008910992182791233, + 0.008715063333511353, + 0.009117628447711468, + 0.012027615681290627, + 0.009828560054302216, + 0.006808556616306305, + 0.003171043936163187, + 0.003397181164473295, + 0.0029663785826414824, + 0.0036018320824950933, + 0.005176704842597246, + 0.003985822666436434, + 0.003917844500392675, + 0.004235049709677696, + 0.013657478615641594, + 0.015471435151994228, + 0.009853188879787922, + 0.018420226871967316, + 0.02156117372214794, + 0.11799829453229904, + 0.01739930547773838, + 0.11879302561283112, + 0.0079496493563056, + 0.0069281356409192085, + 0.007712113205343485, + 0.010404090397059917, + 0.00955147948116064, + 0.01356244646012783, + 0.006227492354810238, + 0.011817705817520618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.011071440763771534 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberate 360-degree orbiting shot begins, the camera moving steadily at eye level around a magnificent, sun-warmed peach tree. The scene is rendered with a deeply natural and earthy style, emphasizing organic textures. The tree's gnarled, weathered bark creates a tactile contrast with its canopy of lush, dark green leaves, still shimmering with the lingering crystal droplets of a recent rain shower. Its branches hang heavy with an abundance of ripe fruit; each peach is a gradient of color, from pale yellow to deep rose, its skin covered in a delicate, furry velvet that captures the flecked golden sunlight filtering through the leaves. As the camera continues its graceful and steady rotation, the pastoral sky in the background reveals itself, a canvas of soft blue and broken white clouds. Then, a stunning, vibrant, full-spectrum rainbow appears, shimmering against the clear sky, its perfect semicircle framing the tree at the peak of the camera's turn, creating an atmosphere of profound tranquility and wonder.", + "seed": 4566, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_7/index_883_seed_4566_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12099969126284123, + "vlm_logprob": -0.12099969126284123, + "vlm_yes_no_margin": 2.974999952316284, + "vlm_token_credit": -0.12099969126284123, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3566354811191559 + ], + "opposite_logprob_tokens": [ + -1.2316354513168335 + ], + "logprob_reward": -0.3566354811191559, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3566354811191559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09802035987377167, + 0.20253999531269073, + 0.10677818953990936, + 0.14206354320049286, + 0.0426294207572937, + 0.09132222086191177, + 0.16552188992500305, + 0.054358575493097305, + 0.05981205776333809, + 0.03695373237133026 + ], + "credit_map": [ + 0.008594372309744358, + 0.014460332691669464, + 0.013038917444646358, + 0.00978147517889738, + 0.014011341147124767, + 0.01118710171431303, + 0.009740994311869144, + 0.017205819487571716, + 0.022261908277869225, + 0.019200246781110764, + 0.022715343162417412, + 0.02853267453610897, + 0.02428686060011387, + 0.01997664012014866, + 0.026546329259872437, + 0.03901998698711395, + 0.016359206289052963, + 0.014584390446543694, + 0.013327277265489101, + 0.009372776374220848, + 0.010006703436374664, + 0.009288324043154716, + 0.015550423413515091, + 0.01828908734023571, + 0.01543249562382698, + 0.021169930696487427, + 0.02164008840918541, + 0.017839740961790085, + 0.014782844111323357, + 0.012067241594195366, + 0.021038658916950226, + 0.018092544749379158, + 0.003672004444524646, + 0.005666615907102823, + 0.005774049088358879, + 0.0055695450864732265, + 0.005342860706150532, + 0.0036814818158745766, + 0.007251919247210026, + 0.005670945625752211, + 0.022923186421394348, + 0.009311375208199024, + 0.014359189197421074, + 0.006225225981324911, + 0.0070855566300451756, + 0.00905202142894268, + 0.008028569631278515, + 0.014337096363306046, + 0.01624409854412079, + 0.02893826551735401, + 0.023926932364702225, + 0.022191867232322693, + 0.01891721412539482, + 0.015531717799603939, + 0.020015712827444077, + 0.019756078720092773, + 0.006029970478266478, + 0.005815492942929268, + 0.011894493363797665, + 0.007544633466750383, + 0.007730190176516771, + 0.0035622266586869955, + 0.007806399837136269, + 0.003975167404860258, + 0.0060401614755392075, + 0.008114657364785671, + 0.00832721870392561, + 0.016160910949110985, + 0.005308473948389292, + 0.006904134526848793, + 0.004654289223253727, + 0.004302212968468666, + 0.004323279019445181, + 0.006029512733221054, + 0.007215716410428286, + 0.005306175909936428, + 0.0030806874856352806, + 0.0038626352325081825, + 0.00273453863337636, + 0.004401189275085926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.3566354811191559 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018200531601905823 + ], + "opposite_logprob_tokens": [ + -4.018200397491455 + ], + "logprob_reward": -0.018200531601905823, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018200531601905823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1279500275850296, + 0.2120332568883896, + 0.03582374379038811, + 0.05476907268166542, + 0.039008159190416336, + 0.10164396464824677, + 0.08214318007230759, + 0.16801251471042633, + 0.04155668988823891, + 0.13705940544605255 + ], + "credit_map": [ + 0.010478348471224308, + 0.014770286157727242, + 0.021007755771279335, + 0.015867257490754128, + 0.024868007749319077, + 0.01041514053940773, + 0.01872986927628517, + 0.011813340708613396, + 0.02269955724477768, + 0.02074076235294342, + 0.026695942506194115, + 0.03420315682888031, + 0.024687740951776505, + 0.01958618499338627, + 0.027027677744627, + 0.0363922081887722, + 0.00306843314319849, + 0.0037148771807551384, + 0.004258202854543924, + 0.004002891015261412, + 0.006314446683973074, + 0.004099202807992697, + 0.006698855198919773, + 0.003666829550638795, + 0.00488444697111845, + 0.006839295849204063, + 0.00809524580836296, + 0.009066365659236908, + 0.008684818632900715, + 0.0036841463297605515, + 0.008581843227148056, + 0.004932902753353119, + 0.0033323976676911116, + 0.003937365487217903, + 0.006034189369529486, + 0.006652757991105318, + 0.0065535735338926315, + 0.002907885704189539, + 0.006073896307498217, + 0.0035160891711711884, + 0.024671144783496857, + 0.008560605347156525, + 0.01615748181939125, + 0.0067390198819339275, + 0.008040142245590687, + 0.010542710311710835, + 0.009726310148835182, + 0.01720653846859932, + 0.008700130507349968, + 0.011446371674537659, + 0.013239997439086437, + 0.011876952834427357, + 0.0065289754420518875, + 0.00702841579914093, + 0.01178413350135088, + 0.011538192629814148, + 0.01726691424846649, + 0.017219388857483864, + 0.03654775395989418, + 0.02153804339468479, + 0.01807526871562004, + 0.011635172180831432, + 0.03123578056693077, + 0.014494173228740692, + 0.004559386987239122, + 0.005394468083977699, + 0.006011458579450846, + 0.005755840800702572, + 0.00467190146446228, + 0.003830412868410349, + 0.004811602178961039, + 0.006521615199744701, + 0.017258355394005775, + 0.021636828780174255, + 0.030845584347844124, + 0.01736290566623211, + 0.015783878043293953, + 0.00953361764550209, + 0.013144226744771004, + 0.011494013480842113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 9 + ], + "selected_reward": -0.018200531601905823 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069775447249413 + ], + "opposite_logprob_tokens": [ + -4.1410698890686035 + ], + "logprob_reward": -0.016069775447249413, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069775447249413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12751731276512146, + 0.2037745863199234, + 0.0856543481349945, + 0.08651836961507797, + 0.03765716776251793, + 0.0382467582821846, + 0.06754256039857864, + 0.1973847895860672, + 0.0681600347161293, + 0.08754406124353409 + ], + "credit_map": [ + 0.015096891671419144, + 0.016092104837298393, + 0.015826905146241188, + 0.015412062406539917, + 0.019982652738690376, + 0.015387739054858685, + 0.014688536524772644, + 0.01503042597323656, + 0.02251964621245861, + 0.02088886685669422, + 0.027302926406264305, + 0.03363965451717377, + 0.023855432868003845, + 0.017521414905786514, + 0.025863375514745712, + 0.032183270901441574, + 0.012124834582209587, + 0.010631606914103031, + 0.014021449722349644, + 0.009433425031602383, + 0.012562884017825127, + 0.008114767260849476, + 0.011651212349534035, + 0.007114169653505087, + 0.009537111967802048, + 0.012201358564198017, + 0.012082776054739952, + 0.010330391116440296, + 0.011583087965846062, + 0.006458044983446598, + 0.014841406606137753, + 0.009484190493822098, + 0.0037086259108036757, + 0.004227533936500549, + 0.006136338692158461, + 0.00519945565611124, + 0.005997510626912117, + 0.00271165301091969, + 0.005954793654382229, + 0.003721254412084818, + 0.004103872925043106, + 0.005790228024125099, + 0.005087755154818296, + 0.005484225694090128, + 0.005458037834614515, + 0.0025995327159762383, + 0.004246100317686796, + 0.005477004684507847, + 0.006940431892871857, + 0.009452462196350098, + 0.00931959692388773, + 0.011129036545753479, + 0.006868736818432808, + 0.005327481776475906, + 0.010243731550872326, + 0.008261076174676418, + 0.014740468934178352, + 0.01800740882754326, + 0.04700617119669914, + 0.025936035439372063, + 0.02066499926149845, + 0.014067992568016052, + 0.03502259403467178, + 0.021939124912023544, + 0.006529992911964655, + 0.00879837665706873, + 0.009611601009964943, + 0.010681030340492725, + 0.005330467130988836, + 0.007934052497148514, + 0.006333871278911829, + 0.012940644286572933, + 0.00987402256578207, + 0.013777175918221474, + 0.022412588819861412, + 0.010310633108019829, + 0.009551100432872772, + 0.005695365369319916, + 0.007125916425138712, + 0.008797260001301765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.016069775447249413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155425369739532 + ], + "opposite_logprob_tokens": [ + -1.701554298400879 + ], + "logprob_reward": -0.20155425369739532, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155425369739532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03482372313737869, + 0.10312286764383316, + 0.016081053763628006, + 0.02455724962055683, + 0.013362388126552105, + 0.046050865203142166, + 0.07414054125547409, + 0.2840140163898468, + 0.17294873297214508, + 0.23089857399463654 + ], + "credit_map": [ + 0.002723688492551446, + 0.0038730367086827755, + 0.005571525078266859, + 0.006454328540712595, + 0.0046957735903561115, + 0.003383372910320759, + 0.005047199781984091, + 0.0030747964046895504, + 0.01144165638834238, + 0.00952743086963892, + 0.011626143008470535, + 0.015882359817624092, + 0.011018482968211174, + 0.010292568244040012, + 0.01313081569969654, + 0.02020341157913208, + 0.0017889582086354494, + 0.0017197657143697143, + 0.0017249839147552848, + 0.0018211769638583064, + 0.0017803144874051213, + 0.0014386550756171346, + 0.002480223774909973, + 0.0033269745763391256, + 0.0029874886386096478, + 0.003336819587275386, + 0.00338190165348351, + 0.0033061665017157793, + 0.0028940197080373764, + 0.0019792718812823296, + 0.004230772145092487, + 0.0024408092722296715, + 0.0012059316504746675, + 0.0017009548610076308, + 0.001842485275119543, + 0.0020547090098261833, + 0.002381053753197193, + 0.0009637812618166208, + 0.0019067219691351056, + 0.0013067498803138733, + 0.01180571224540472, + 0.0045066047459840775, + 0.006500925403088331, + 0.0024526065681129694, + 0.0033739639911800623, + 0.005066538695245981, + 0.003635861910879612, + 0.0087086521089077, + 0.0068785627372562885, + 0.009075027890503407, + 0.014537988230586052, + 0.008985957130789757, + 0.010940656997263432, + 0.007215065415948629, + 0.009929920546710491, + 0.006577363237738609, + 0.02748524770140648, + 0.02735176868736744, + 0.06331359595060349, + 0.037019357085227966, + 0.02959185652434826, + 0.019673995673656464, + 0.05368579179048538, + 0.02589237317442894, + 0.01615501195192337, + 0.02245340123772621, + 0.02167687565088272, + 0.04424816370010376, + 0.015918027609586716, + 0.02331903949379921, + 0.012698831036686897, + 0.016479386016726494, + 0.0241598729044199, + 0.020981160923838615, + 0.04576423764228821, + 0.031285062432289124, + 0.026933806017041206, + 0.01842782273888588, + 0.03440770134329796, + 0.028938913717865944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.20155425369739532 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253841444849968 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.01253841444849968, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253841444849968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11241012811660767, + 0.15465739369392395, + 0.046434976160526276, + 0.02223426103591919, + 0.031254660338163376, + 0.08990396559238434, + 0.04568690434098244, + 0.2796432673931122, + 0.13995882868766785, + 0.07781560719013214 + ], + "credit_map": [ + 0.011030351743102074, + 0.014128039591014385, + 0.01827264577150345, + 0.011435718275606632, + 0.01712118461728096, + 0.011968242935836315, + 0.014603002928197384, + 0.013850937597453594, + 0.01610271818935871, + 0.014384589157998562, + 0.018767142668366432, + 0.025832796469330788, + 0.017804598435759544, + 0.015693707391619682, + 0.019896138459444046, + 0.026175687089562416, + 0.004320230800658464, + 0.005889453925192356, + 0.006670330185443163, + 0.006423651706427336, + 0.005733109079301357, + 0.004376283381134272, + 0.006283680442720652, + 0.0067382315173745155, + 0.0019692687783390284, + 0.004299541004002094, + 0.0031387105118483305, + 0.0029513929039239883, + 0.003131840843707323, + 0.0014755105366930366, + 0.0030044433660805225, + 0.002263550413772464, + 0.003336497349664569, + 0.003088162513449788, + 0.0045240577310323715, + 0.005397337023168802, + 0.004890029784291983, + 0.0024331880267709494, + 0.005126777105033398, + 0.002458608942106366, + 0.02054773084819317, + 0.006563403643667698, + 0.014417236670851707, + 0.007429612334817648, + 0.007585184648633003, + 0.009655123576521873, + 0.008188813924789429, + 0.015516852959990501, + 0.004507924895733595, + 0.007191922515630722, + 0.007501581218093634, + 0.005500574596226215, + 0.005471412092447281, + 0.004936039447784424, + 0.0048546940088272095, + 0.0057227518409490585, + 0.024933792650699615, + 0.026741139590740204, + 0.06553474813699722, + 0.036780476570129395, + 0.027480317279696465, + 0.019798800349235535, + 0.050879672169685364, + 0.027494292706251144, + 0.014424222521483898, + 0.019526125863194466, + 0.02226915769279003, + 0.022761836647987366, + 0.011803312227129936, + 0.01539635006338358, + 0.011569290421903133, + 0.02220851369202137, + 0.007966513745486736, + 0.009019236080348492, + 0.017102401703596115, + 0.010122433304786682, + 0.008799021132290363, + 0.0073862154968082905, + 0.008530477061867714, + 0.008889302611351013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.01253841444849968 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberate 360-degree orbiting shot begins, the camera moving steadily at eye level around a magnificent, sun-warmed peach tree. The scene is rendered with a deeply natural and earthy style, emphasizing organic textures. The tree's gnarled, weathered bark creates a tactile contrast with its canopy of lush, dark green leaves, still shimmering with the lingering crystal droplets of a recent rain shower. Its branches hang heavy with an abundance of ripe fruit; each peach is a gradient of color, from pale yellow to deep rose, its skin covered in a delicate, furry velvet that captures the flecked golden sunlight filtering through the leaves. As the camera continues its graceful and steady rotation, the pastoral sky in the background reveals itself, a canvas of soft blue and broken white clouds. Then, a stunning, vibrant, full-spectrum rainbow appears, shimmering against the clear sky, its perfect semicircle framing the tree at the peak of the camera's turn, creating an atmosphere of profound tranquility and wonder.", + "seed": 4567, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_7/index_883_seed_4567_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06339373984374105, + "vlm_logprob": -0.06339373984374105, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.06339373984374105, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2842821776866913 + ], + "opposite_logprob_tokens": [ + -1.4092822074890137 + ], + "logprob_reward": -0.2842821776866913, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2842821776866913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15076413750648499, + 0.045017048716545105, + 0.17411133646965027, + 0.1604614555835724, + 0.07835907489061356, + 0.11375261843204498, + 0.09376183897256851, + 0.06839364022016525, + 0.06401323527097702, + 0.051365625113248825 + ], + "credit_map": [ + 0.015149051323533058, + 0.03282138332724571, + 0.01462221797555685, + 0.01908249780535698, + 0.022372353821992874, + 0.012711773626506329, + 0.017847605049610138, + 0.0161572452634573, + 0.00458849873393774, + 0.007986563257873058, + 0.006450697313994169, + 0.0061195469461381435, + 0.0035655649844557047, + 0.0054804133251309395, + 0.006699145305901766, + 0.004126622807234526, + 0.015061491169035435, + 0.009744485840201378, + 0.01303982362151146, + 0.013689393177628517, + 0.02499450370669365, + 0.03080182708799839, + 0.04767841100692749, + 0.019101399928331375, + 0.034432001411914825, + 0.012275127694010735, + 0.01542145200073719, + 0.019599445164203644, + 0.012508786283433437, + 0.03249761089682579, + 0.019607173278927803, + 0.01411985233426094, + 0.007063164841383696, + 0.009429443627595901, + 0.005486112553626299, + 0.011870583519339561, + 0.01703750528395176, + 0.01202131062746048, + 0.011099439114332199, + 0.0043515111319720745, + 0.008614066988229752, + 0.006990076974034309, + 0.02923044003546238, + 0.03948467969894409, + 0.006348316092044115, + 0.008615047670900822, + 0.007476455066353083, + 0.00699353963136673, + 0.009743872098624706, + 0.012067722156643867, + 0.009644261561334133, + 0.012294943444430828, + 0.012495686300098896, + 0.013673916459083557, + 0.008499429561197758, + 0.015342009253799915, + 0.007939882576465607, + 0.007407342083752155, + 0.009701433591544628, + 0.011231618002057076, + 0.006575806066393852, + 0.006275847554206848, + 0.009340887889266014, + 0.009920823387801647, + 0.007807378191500902, + 0.007664921227842569, + 0.005683402996510267, + 0.0069231255911290646, + 0.015306390821933746, + 0.007073236163705587, + 0.0075365155935287476, + 0.006018263287842274, + 0.0058868746273219585, + 0.0036761723458766937, + 0.016172239556908607, + 0.007244917098432779, + 0.003257082076743245, + 0.0029279529117047787, + 0.008493991568684578, + 0.003706395160406828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.2842821776866913 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978990737348795 + ], + "opposite_logprob_tokens": [ + -4.634789943695068 + ], + "logprob_reward": -0.00978990737348795, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978990737348795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1470864862203598, + 0.1367013305425644, + 0.33921822905540466, + 0.039494987577199936, + 0.03819134458899498, + 0.03561405837535858, + 0.05014363303780556, + 0.02971881814301014, + 0.08517429232597351, + 0.09865684807300568 + ], + "credit_map": [ + 0.01821989379823208, + 0.01730772666633129, + 0.014494158327579498, + 0.014819027855992317, + 0.020028933882713318, + 0.01351460162550211, + 0.016538290306925774, + 0.03216386213898659, + 0.01093426626175642, + 0.03859236463904381, + 0.0062071518041193485, + 0.010052642785012722, + 0.01026295218616724, + 0.007048205938190222, + 0.017295541241765022, + 0.036308202892541885, + 0.03828180208802223, + 0.012944059446454048, + 0.12477611750364304, + 0.012029685080051422, + 0.023327743634581566, + 0.08909384161233902, + 0.017305409535765648, + 0.02145954966545105, + 0.006966681219637394, + 0.0027088734786957502, + 0.0034003148321062326, + 0.0047899591736495495, + 0.0030455307569354773, + 0.008881325833499432, + 0.006188650615513325, + 0.0035136521328240633, + 0.003059308510273695, + 0.0040920451283454895, + 0.0035562380217015743, + 0.005742844194173813, + 0.00772226694971323, + 0.00490056537091732, + 0.00654187798500061, + 0.002576199360191822, + 0.002507539466023445, + 0.0037224343977868557, + 0.0054109529592096806, + 0.010011795908212662, + 0.004414686933159828, + 0.0037413162644952536, + 0.0026788609102368355, + 0.003126471769064665, + 0.004643602296710014, + 0.004889619071036577, + 0.005499267019331455, + 0.010874311439692974, + 0.0055183302611112595, + 0.005995369981974363, + 0.007735537830740213, + 0.004987592343240976, + 0.0033499838318675756, + 0.003603770863264799, + 0.0037456185091286898, + 0.004189054947346449, + 0.003762209555134177, + 0.002925839740782976, + 0.004606255330145359, + 0.0035360853653401136, + 0.012640533968806267, + 0.01296544075012207, + 0.008977512829005718, + 0.01435813307762146, + 0.009240040555596352, + 0.0078972727060318, + 0.009116480126976967, + 0.009978876449167728, + 0.007571642752736807, + 0.0032365487422794104, + 0.03667963668704033, + 0.019530806690454483, + 0.004730531480163336, + 0.004111116286367178, + 0.019042421132326126, + 0.0037541459314525127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.00978990737348795 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031922117341309786 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.0031922117341309786, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031922117341309786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4145161807537079, + 0.05268678441643715, + 0.1762656718492508, + 0.026557812467217445, + 0.03762947395443916, + 0.07212304323911667, + 0.046962130814790726, + 0.032547660171985626, + 0.02892114967107773, + 0.11179010570049286 + ], + "credit_map": [ + 0.05609089136123657, + 0.031443074345588684, + 0.025166144594550133, + 0.020822100341320038, + 0.06816962361335754, + 0.02177806757390499, + 0.036011386662721634, + 0.15503491461277008, + 0.004077615682035685, + 0.01220016647130251, + 0.0035370683763176203, + 0.005099631380289793, + 0.004503288771957159, + 0.004031894728541374, + 0.008346554823219776, + 0.010890565812587738, + 0.015646489337086678, + 0.007959588430821896, + 0.04588186368346214, + 0.009248350746929646, + 0.014909167774021626, + 0.0500965416431427, + 0.019951190799474716, + 0.012572470121085644, + 0.005346118472516537, + 0.0019144683610647917, + 0.002804409246891737, + 0.0026711814571172, + 0.0018286550184711814, + 0.005518788006156683, + 0.004174458794295788, + 0.0022997334599494934, + 0.00339180463925004, + 0.004956812132149935, + 0.003596568014472723, + 0.005418727174401283, + 0.0059255873784422874, + 0.0056556654162704945, + 0.006001554895192385, + 0.0026827517431229353, + 0.00517279701307416, + 0.005174404010176659, + 0.016800738871097565, + 0.020724933594465256, + 0.008501199074089527, + 0.0047062295489013195, + 0.006287858355790377, + 0.004754879977554083, + 0.0058523500338196754, + 0.004665070213377476, + 0.004806440323591232, + 0.007257711607962847, + 0.004605154041200876, + 0.005516645964235067, + 0.00823331531137228, + 0.006025440990924835, + 0.0031227373983711004, + 0.0034506656229496002, + 0.0046459343284368515, + 0.004568797070533037, + 0.0029706174973398447, + 0.003115310100838542, + 0.003954894375056028, + 0.0067187040112912655, + 0.0035404106602072716, + 0.002355532720685005, + 0.003733244724571705, + 0.004949785303324461, + 0.0030559126753360033, + 0.0024293321184813976, + 0.004354471806436777, + 0.004502459429204464, + 0.0085268784314394, + 0.004831859841942787, + 0.03663922846317291, + 0.019628625363111496, + 0.0063744098879396915, + 0.005450594704598188, + 0.022031424567103386, + 0.008307079784572124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.0031922117341309786 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011069907806813717 + ], + "opposite_logprob_tokens": [ + -4.5110697746276855 + ], + "logprob_reward": -0.011069907806813717, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011069907806813717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25464490056037903, + 0.01931932009756565, + 0.10736765712499619, + 0.028067227452993393, + 0.08496279269456863, + 0.1359986662864685, + 0.0472218357026577, + 0.11608529835939407, + 0.08204350620508194, + 0.12428883463144302 + ], + "credit_map": [ + 0.0326954685151577, + 0.018949786201119423, + 0.01399251352995634, + 0.014775463379919529, + 0.041929446160793304, + 0.012456783093512058, + 0.02413225919008255, + 0.09571319818496704, + 0.0017004720866680145, + 0.003854327369481325, + 0.001624363474547863, + 0.002897709608078003, + 0.0016398949082940817, + 0.0023676075506955385, + 0.003212367184460163, + 0.002022577915340662, + 0.008567377924919128, + 0.0059131174348294735, + 0.02589385211467743, + 0.006574213970452547, + 0.008907021023333073, + 0.025104479864239693, + 0.01800226978957653, + 0.008405329659581184, + 0.004532543011009693, + 0.002158984774723649, + 0.002872155513614416, + 0.0029380826745182276, + 0.002417541341856122, + 0.005704614799469709, + 0.005210682284086943, + 0.0022326249163597822, + 0.007074875757098198, + 0.009557854384183884, + 0.0070390538312494755, + 0.013972698710858822, + 0.017260128632187843, + 0.012086422182619572, + 0.012824940495193005, + 0.005146821960806847, + 0.010444540530443192, + 0.008380548097193241, + 0.036586444824934006, + 0.04412292316555977, + 0.013475318439304829, + 0.006658472120761871, + 0.008601532317698002, + 0.007728880271315575, + 0.005696124397218227, + 0.003921739757061005, + 0.00466747023165226, + 0.008324883878231049, + 0.005249889101833105, + 0.00517289200797677, + 0.006421950180083513, + 0.007766885217279196, + 0.014795416034758091, + 0.01649424061179161, + 0.01175351906567812, + 0.013037227094173431, + 0.020360838621854782, + 0.01013385783880949, + 0.014529884792864323, + 0.0149803152307868, + 0.008744040504097939, + 0.01034501101821661, + 0.010306152515113354, + 0.011082297191023827, + 0.006693234667181969, + 0.00980592705309391, + 0.012243837118148804, + 0.012823008000850677, + 0.015516013838350773, + 0.006468064617365599, + 0.033957548439502716, + 0.012210885062813759, + 0.0065568480640649796, + 0.00793225783854723, + 0.02862994372844696, + 0.013017269782721996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.011069907806813717 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634494617581367 + ], + "opposite_logprob_tokens": [ + -4.758634567260742 + ], + "logprob_reward": -0.008634494617581367, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634494617581367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15118516981601715, + 0.02293359860777855, + 0.5607509016990662, + 0.011148418299853802, + 0.029296400025486946, + 0.019296813756227493, + 0.027764756232500076, + 0.039030369371175766, + 0.048236746340990067, + 0.09035686403512955 + ], + "credit_map": [ + 0.020140253007411957, + 0.009366139769554138, + 0.007642456330358982, + 0.006976481061428785, + 0.025698965415358543, + 0.007642359007149935, + 0.013718430884182453, + 0.06000006943941116, + 0.0019475152948871255, + 0.0022488271351903677, + 0.0027469452470541, + 0.00335432100109756, + 0.002356969052925706, + 0.003410931210964918, + 0.004587214905768633, + 0.002280872082337737, + 0.06175120174884796, + 0.02135620079934597, + 0.20146526396274567, + 0.01985473372042179, + 0.03869354724884033, + 0.1546773612499237, + 0.026568200439214706, + 0.03638436645269394, + 0.002012308221310377, + 0.0007946859113872051, + 0.0012498124269768596, + 0.0009439191780984402, + 0.0009610280394554138, + 0.002532089129090309, + 0.0017219891306012869, + 0.0009325850405730307, + 0.002408406464383006, + 0.0030587883666157722, + 0.0022830066736787558, + 0.004665993619710207, + 0.0060996683314442635, + 0.004308932460844517, + 0.0044647157192230225, + 0.002006887225434184, + 0.0018429020419716835, + 0.002908130409196019, + 0.0016178617952391505, + 0.002233366947621107, + 0.004592372104525566, + 0.0017978283576667309, + 0.002436227397993207, + 0.001868123305030167, + 0.0027097580023109913, + 0.002437266055494547, + 0.0033872793428599834, + 0.003841501660645008, + 0.004374487325549126, + 0.005066363140940666, + 0.002749040024355054, + 0.0031990588176995516, + 0.004801705479621887, + 0.007091145496815443, + 0.005026661325246096, + 0.005643904209136963, + 0.0067101432941854, + 0.0034794602543115616, + 0.0029013645835220814, + 0.0033759870566427708, + 0.0054594287648797035, + 0.007754716090857983, + 0.004917053505778313, + 0.007778654806315899, + 0.006821249611675739, + 0.005049154162406921, + 0.005312473978847265, + 0.005144009832292795, + 0.007157683838158846, + 0.003955224994570017, + 0.02995983324944973, + 0.01623748615384102, + 0.00586292939260602, + 0.004634062759578228, + 0.016662131994962692, + 0.0058875055983662605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.008634494617581367 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberate 360-degree orbiting shot begins, the camera moving steadily at eye level around a magnificent, sun-warmed peach tree. The scene is rendered with a deeply natural and earthy style, emphasizing organic textures. The tree's gnarled, weathered bark creates a tactile contrast with its canopy of lush, dark green leaves, still shimmering with the lingering crystal droplets of a recent rain shower. Its branches hang heavy with an abundance of ripe fruit; each peach is a gradient of color, from pale yellow to deep rose, its skin covered in a delicate, furry velvet that captures the flecked golden sunlight filtering through the leaves. As the camera continues its graceful and steady rotation, the pastoral sky in the background reveals itself, a canvas of soft blue and broken white clouds. Then, a stunning, vibrant, full-spectrum rainbow appears, shimmering against the clear sky, its perfect semicircle framing the tree at the peak of the camera's turn, creating an atmosphere of profound tranquility and wonder.", + "seed": 4568, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_7/index_883_seed_4568_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2468880344182253, + "vlm_logprob": -0.2468880344182253, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.2468880344182253, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0637224912643433 + ], + "opposite_logprob_tokens": [ + -0.43872249126434326 + ], + "logprob_reward": -1.0637224912643433, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0637224912643433, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10035347193479538, + 0.11502908170223236, + 0.2518482804298401, + 0.0452110692858696, + 0.07061541080474854, + 0.10938897728919983, + 0.15441299974918365, + 0.06331395357847214, + 0.041148923337459564, + 0.04867775738239288 + ], + "credit_map": [ + 0.007843520492315292, + 0.009562528692185879, + 0.027389084920287132, + 0.007180799264460802, + 0.00787719339132309, + 0.008762399666011333, + 0.007479083724319935, + 0.02425886131823063, + 0.018990546464920044, + 0.015551868826150894, + 0.0199173204600811, + 0.011925973929464817, + 0.008316880092024803, + 0.013192937709391117, + 0.006823880132287741, + 0.020309675484895706, + 0.01636503078043461, + 0.06745288521051407, + 0.0970688909292221, + 0.016201741993427277, + 0.01150889415293932, + 0.012084056623280048, + 0.014670991338789463, + 0.016495773568749428, + 0.006988027133047581, + 0.006294548511505127, + 0.005984669551253319, + 0.005941805895417929, + 0.00407802639529109, + 0.006059102714061737, + 0.00425691669806838, + 0.005607972852885723, + 0.005669571924954653, + 0.009572740644216537, + 0.006552956998348236, + 0.008110826835036278, + 0.013860348612070084, + 0.010086484253406525, + 0.006432775408029556, + 0.010329704731702805, + 0.015763409435749054, + 0.015523788519203663, + 0.0062601822428405285, + 0.030247369781136513, + 0.01038350909948349, + 0.006988092325627804, + 0.010441174730658531, + 0.013781452551484108, + 0.032393116503953934, + 0.017215829342603683, + 0.01944039575755596, + 0.013745613396167755, + 0.006774371489882469, + 0.009774797596037388, + 0.01771486923098564, + 0.0373539999127388, + 0.008655937388539314, + 0.009457179345190525, + 0.0061940886080265045, + 0.0060603441670536995, + 0.008604805916547775, + 0.007484771776944399, + 0.008257207460701466, + 0.008599622175097466, + 0.006713749375194311, + 0.005287680309265852, + 0.004661389160901308, + 0.0053942217491567135, + 0.003331667510792613, + 0.004470410756766796, + 0.003911389037966728, + 0.007378417067229748, + 0.008380263112485409, + 0.003388196462765336, + 0.011949625797569752, + 0.005947161465883255, + 0.004172180313616991, + 0.0056937215849757195, + 0.0036914676893502474, + 0.00545514328405261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -1.0637224912643433 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018204277381300926 + ], + "opposite_logprob_tokens": [ + -4.018204212188721 + ], + "logprob_reward": -0.018204277381300926, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018204277381300926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11770164221525192, + 0.0928415060043335, + 0.28449735045433044, + 0.06635981053113937, + 0.1913059800863266, + 0.030403997749090195, + 0.11944510042667389, + 0.04540213197469711, + 0.028452707454562187, + 0.023589665070176125 + ], + "credit_map": [ + 0.012024381197988987, + 0.012762847356498241, + 0.02678024396300316, + 0.012457743287086487, + 0.014649356715381145, + 0.008661461062729359, + 0.011939767748117447, + 0.018425848335027695, + 0.01941077597439289, + 0.009661334566771984, + 0.015138671733438969, + 0.008309058845043182, + 0.006735583767294884, + 0.011442601680755615, + 0.005272315349429846, + 0.016871167346835136, + 0.024243060499429703, + 0.06712929159402847, + 0.10660621523857117, + 0.019652118906378746, + 0.014927530661225319, + 0.014688285067677498, + 0.016391843557357788, + 0.020859016105532646, + 0.005278856959193945, + 0.007661620154976845, + 0.007143702823668718, + 0.008332131430506706, + 0.016140369698405266, + 0.005643348675221205, + 0.008320110850036144, + 0.007839671336114407, + 0.016338225454092026, + 0.02515815570950508, + 0.014721027575433254, + 0.028629017993807793, + 0.04267973080277443, + 0.020313570275902748, + 0.021899549290537834, + 0.021566715091466904, + 0.0044992947950959206, + 0.003631874918937683, + 0.00456651346758008, + 0.00564859714359045, + 0.0030253285076469183, + 0.0028004285413771868, + 0.0025513642467558384, + 0.003680596360936761, + 0.032667867839336395, + 0.011879916302859783, + 0.006298976484686136, + 0.008813311345875263, + 0.0048472885973751545, + 0.008093861863017082, + 0.009065942838788033, + 0.037777941673994064, + 0.005605597980320454, + 0.003907428588718176, + 0.003935979679226875, + 0.0041953520849347115, + 0.004173304419964552, + 0.007099032402038574, + 0.008173161186277866, + 0.008312273770570755, + 0.00473144743591547, + 0.0035668236669152975, + 0.004732174798846245, + 0.0034951979760080576, + 0.002530488185584545, + 0.002319705905392766, + 0.004099590703845024, + 0.002977279946208, + 0.004372008144855499, + 0.0027857706882059574, + 0.003183200489729643, + 0.0029876476619392633, + 0.002109093125909567, + 0.002774871187284589, + 0.0019756669644266367, + 0.0034014061093330383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -0.018204277381300926 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781880304217339 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.009781880304217339, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781880304217339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18971827626228333, + 0.1370672583580017, + 0.12842828035354614, + 0.05492366477847099, + 0.07463151216506958, + 0.05315861850976944, + 0.21331995725631714, + 0.05687073990702629, + 0.07507847249507904, + 0.01680324599146843 + ], + "credit_map": [ + 0.01922416500747204, + 0.028062496334314346, + 0.02181665599346161, + 0.02545468509197235, + 0.035090476274490356, + 0.01627933606505394, + 0.02805178426206112, + 0.01573866792023182, + 0.022659925743937492, + 0.014623592607676983, + 0.02181725576519966, + 0.01381603628396988, + 0.013610966503620148, + 0.01537563931196928, + 0.01094905473291874, + 0.024214763194322586, + 0.02631155028939247, + 0.022406859323382378, + 0.0170013215392828, + 0.011138380505144596, + 0.012792528606951237, + 0.012969313189387321, + 0.013080889359116554, + 0.012727421708405018, + 0.006662400905042887, + 0.005974819418042898, + 0.0061415149830281734, + 0.007493346463888884, + 0.008290045894682407, + 0.008726502768695354, + 0.006130815949290991, + 0.0055042109452188015, + 0.009735329076647758, + 0.0102285360917449, + 0.005319687537848949, + 0.013863826170563698, + 0.011097247712314129, + 0.007149555720388889, + 0.009720784611999989, + 0.007516543846577406, + 0.007021040190011263, + 0.008616131730377674, + 0.004142866935580969, + 0.012967297807335854, + 0.004920992534607649, + 0.00515950471162796, + 0.004838529974222183, + 0.005492249503731728, + 0.05394198000431061, + 0.022521525621414185, + 0.01359203364700079, + 0.016923045739531517, + 0.010446296073496342, + 0.017052434384822845, + 0.017012275755405426, + 0.06183035299181938, + 0.009375574998557568, + 0.005841901991516352, + 0.00417207321152091, + 0.004781598225235939, + 0.006343793589621782, + 0.007185284048318863, + 0.00713814701884985, + 0.012032358907163143, + 0.010260539129376411, + 0.007184790447354317, + 0.012735077179968357, + 0.00918127503246069, + 0.006329750642180443, + 0.0073746745474636555, + 0.013901161961257458, + 0.008111195638775826, + 0.002919120481237769, + 0.0015561141772195697, + 0.002111639827489853, + 0.0020576156675815582, + 0.0015860323328524828, + 0.00246342271566391, + 0.0013400999596342444, + 0.0027691994328051805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.009781880304217339 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129358634352684 + ], + "opposite_logprob_tokens": [ + -2.237935781478882 + ], + "logprob_reward": -0.1129358634352684, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129358634352684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029905838891863823, + 0.06129448860883713, + 0.0692056342959404, + 0.04270473122596741, + 0.048463400453329086, + 0.08071627467870712, + 0.5552314519882202, + 0.04217161610722542, + 0.05041149631142616, + 0.019895033910870552 + ], + "credit_map": [ + 0.004203868564218283, + 0.003755440004169941, + 0.0050033810548484325, + 0.0038202335126698017, + 0.003287188708782196, + 0.0023347202222794294, + 0.003200501436367631, + 0.0043005081824958324, + 0.012677669525146484, + 0.008443402126431465, + 0.010832610540091991, + 0.005052810534834862, + 0.004275904502719641, + 0.0070463502779603004, + 0.003811862785369158, + 0.009153886698186398, + 0.004141159355640411, + 0.019011912867426872, + 0.026563407853245735, + 0.0049582780338823795, + 0.003222998231649399, + 0.0032018679194152355, + 0.00355770206078887, + 0.004548307973891497, + 0.005198991391807795, + 0.005921051371842623, + 0.005929387640208006, + 0.006203147582709789, + 0.0036241994239389896, + 0.006259535439312458, + 0.004475446883589029, + 0.005092973820865154, + 0.004005077760666609, + 0.006546176038682461, + 0.005356668494641781, + 0.00518065644428134, + 0.008659688755869865, + 0.007905608043074608, + 0.0039957440458238125, + 0.006813785061240196, + 0.01277138851583004, + 0.007962970994412899, + 0.015037385746836662, + 0.012548143975436687, + 0.0063317627646028996, + 0.007640119642019272, + 0.00885134469717741, + 0.009573161602020264, + 0.12972919642925262, + 0.04900756850838661, + 0.05058924853801727, + 0.040535833686590195, + 0.029658064246177673, + 0.04797687008976936, + 0.059747274965047836, + 0.14798744022846222, + 0.004711718764156103, + 0.0047581312246620655, + 0.0024079198483377695, + 0.0026121544651687145, + 0.006954928394407034, + 0.008815527893602848, + 0.004625349771231413, + 0.007285889238119125, + 0.005036487244069576, + 0.004610078409314156, + 0.0063828653655946255, + 0.006480022333562374, + 0.004877409432083368, + 0.008499096147716045, + 0.008348479866981506, + 0.00617706123739481, + 0.004240288399159908, + 0.0018287260318174958, + 0.002643587999045849, + 0.001887573511339724, + 0.0018086497439071536, + 0.0026410978753119707, + 0.0017978730611503124, + 0.003047238802537322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.1129358634352684 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, deliberate 360-degree orbiting shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a peach tree with gnarled, weathered bark and lush dark green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the peaches on the tree depicted as ripe with a gradient color from pale yellow to deep rose and a furry velvet texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant, full-spectrum rainbow visible in the background sky, forming a perfect semicircle around the tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a natural, earthy style featuring organic textures and crystal droplets on the leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795659705996513 + ], + "opposite_logprob_tokens": [ + -3.5297956466674805 + ], + "logprob_reward": -0.029795659705996513, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795659705996513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2411602884531021, + 0.06188347190618515, + 0.0881107896566391, + 0.04797062277793884, + 0.08077403157949448, + 0.08766546845436096, + 0.11880340427160263, + 0.06567855179309845, + 0.14984437823295593, + 0.05810904875397682 + ], + "credit_map": [ + 0.028576308861374855, + 0.03457780182361603, + 0.025124717503786087, + 0.032594259828329086, + 0.04461774230003357, + 0.02127900905907154, + 0.03410813957452774, + 0.020282268524169922, + 0.0100749172270298, + 0.008630462922155857, + 0.009131379425525665, + 0.00589704280719161, + 0.005638814531266689, + 0.007565207779407501, + 0.003942019306123257, + 0.011003624647855759, + 0.00666817557066679, + 0.0211990587413311, + 0.03178083896636963, + 0.006728576496243477, + 0.004864373244345188, + 0.005102865863591433, + 0.004955453798174858, + 0.0068114325404167175, + 0.006651848088949919, + 0.006762067787349224, + 0.006292396225035191, + 0.006338012870401144, + 0.0055037024430930614, + 0.006371337920427322, + 0.005188726354390383, + 0.004862526431679726, + 0.01117311604321003, + 0.011349896900355816, + 0.008758526295423508, + 0.009686961770057678, + 0.012584631331264973, + 0.011107943020761013, + 0.008693410083651543, + 0.007419537752866745, + 0.014343267306685448, + 0.007681021932512522, + 0.019755858927965164, + 0.01197896059602499, + 0.005690487567335367, + 0.008969692513346672, + 0.008827230893075466, + 0.010418939404189587, + 0.02444772981107235, + 0.015175120905041695, + 0.011682986281812191, + 0.00782321859151125, + 0.00624486617743969, + 0.010366986505687237, + 0.010924365371465683, + 0.03213812783360481, + 0.009182723239064217, + 0.005716133862733841, + 0.006714218761771917, + 0.007308902218937874, + 0.005959245841950178, + 0.008801722899079323, + 0.010577239096164703, + 0.011418357491493225, + 0.019219860434532166, + 0.01487718615680933, + 0.022077394649386406, + 0.01932816207408905, + 0.011721961200237274, + 0.018160706385970116, + 0.027648016810417175, + 0.016811078414320946, + 0.0063223871402442455, + 0.00572488596662879, + 0.0077246506698429585, + 0.00729196984320879, + 0.005624454002827406, + 0.008315307088196278, + 0.006975583732128143, + 0.010129803791642189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.029795659705996513 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_9/index_tensor([9127, 9127])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_9/index_tensor([9127, 9127])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e4df83d2b7c8e619deb0be10487b920b19dd6be0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_9/index_tensor([9127, 9127])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the style of sports photography, a powerful low-angle wide shot clings to the rough asphalt runway, capturing a B-17 bomber roaring in from the side of the frame with unstoppable force, instantly filling the frame with a decisive moment. Its iconic silver metal fuselage reflects a cool luster under the bright, glaring sunlight, its rivets and panel seams clearly visible at close range, exuding a pure industrial beauty. The propellers of its four engines spin at high speed, transforming into nearly transparent dynamic discs. The powerful airflow churns the air in front of the camera, stirring up fine dust on the runway surface. Simultaneously, the aircraft's massive silhouette casts a fast-moving, sharply defined dark shadow on the runway, almost in sync with the aircraft itself, flashing past the camera. The stark contrast between light and dark emphasizes its breathtaking speed. The entire scene is filled with mechanical power and breathtaking speed. The sky in the background and the distant horizon are blurred due to high-speed movement, focusing all visual focus on the stunning spectacle of this behemoth flying close to the ground, creating a tense and magnificent atmosphere.", + "seed": 22064, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_9/index_9127_seed_22064_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08762277690693736, + "vlm_logprob": -0.08762277690693736, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.08762277690693736, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12029936909675598 + ], + "opposite_logprob_tokens": [ + -2.2452993392944336 + ], + "logprob_reward": -0.12029936909675598, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12029936909675598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06253135204315186, + 0.08369182050228119, + 0.036231569945812225, + 0.33527520298957825, + 0.17927005887031555, + 0.10120658576488495, + 0.08745797723531723, + 0.03710847720503807, + 0.025409560650587082, + 0.051817387342453 + ], + "credit_map": [ + 0.007762768771499395, + 0.007644370198249817, + 0.004753174260258675, + 0.005782111082226038, + 0.009270308539271355, + 0.006239417940378189, + 0.012347024865448475, + 0.008732175454497337, + 0.009496883489191532, + 0.010727095417678356, + 0.007696168962866068, + 0.009117038920521736, + 0.0164637491106987, + 0.011508219875395298, + 0.008492816239595413, + 0.010189848951995373, + 0.006390022113919258, + 0.004212339408695698, + 0.004682982340455055, + 0.005379378329962492, + 0.005191165488213301, + 0.0037484485656023026, + 0.0037008982617408037, + 0.002926336135715246, + 0.019063783809542656, + 0.013086416758596897, + 0.032313454896211624, + 0.11825653165578842, + 0.016587037593126297, + 0.015511607751250267, + 0.026218725368380547, + 0.09423764050006866, + 0.02151947095990181, + 0.008800243027508259, + 0.03435828536748886, + 0.033198531717061996, + 0.012064791284501553, + 0.0121689448133111, + 0.02264571562409401, + 0.03451407328248024, + 0.008053973317146301, + 0.005466169212013483, + 0.01649608090519905, + 0.01377044152468443, + 0.005545598920434713, + 0.006989397574216127, + 0.019195597618818283, + 0.025689324364066124, + 0.008764482103288174, + 0.012924677692353725, + 0.007997858338057995, + 0.01687130145728588, + 0.0094936927780509, + 0.010447951965034008, + 0.008910324424505234, + 0.012047680094838142, + 0.0046629528515040874, + 0.0070496657863259315, + 0.005268801003694534, + 0.005323415156453848, + 0.005407911725342274, + 0.003919908776879311, + 0.0025848359800875187, + 0.0028909852262586355, + 0.003075982676818967, + 0.003461491549387574, + 0.003685835748910904, + 0.003925330936908722, + 0.0032073722686618567, + 0.003733100136741996, + 0.002673785900697112, + 0.0016466614324599504, + 0.008807790465652943, + 0.00701680826023221, + 0.010187612846493721, + 0.007256961893290281, + 0.004496148321777582, + 0.005694608204066753, + 0.004962802864611149, + 0.003394654719159007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.12029936909675598 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869633540511131 + ], + "opposite_logprob_tokens": [ + -3.048696279525757 + ], + "logprob_reward": -0.04869633540511131, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869633540511131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04840686917304993, + 0.07888268679380417, + 0.03228368982672691, + 0.4054821729660034, + 0.12946854531764984, + 0.06626608967781067, + 0.050455156713724136, + 0.11639802902936935, + 0.02841060981154442, + 0.04394618794322014 + ], + "credit_map": [ + 0.005692055448889732, + 0.007055445574223995, + 0.004401999060064554, + 0.004320594482123852, + 0.00847064983099699, + 0.005035362206399441, + 0.007946307770907879, + 0.005484451074153185, + 0.007261122576892376, + 0.007920686155557632, + 0.0045929234474897385, + 0.007271858397871256, + 0.018072834238409996, + 0.010941650718450546, + 0.012820319272577763, + 0.010001281276345253, + 0.004993942566215992, + 0.003147299401462078, + 0.004627215676009655, + 0.0039012294728308916, + 0.006037306040525436, + 0.0033855040092021227, + 0.003511573886498809, + 0.0026796136517077684, + 0.01856689155101776, + 0.01808825507760048, + 0.02551751583814621, + 0.1511615514755249, + 0.017426511272788048, + 0.013646730221807957, + 0.03735334426164627, + 0.12372133135795593, + 0.016056999564170837, + 0.007007712032645941, + 0.023531615734100342, + 0.02069893293082714, + 0.010367033071815968, + 0.009052786976099014, + 0.016033684834837914, + 0.02671978250145912, + 0.006311883218586445, + 0.004024506080895662, + 0.013658502139151096, + 0.010738097131252289, + 0.004790295846760273, + 0.0053022513166069984, + 0.009593143127858639, + 0.011847403831779957, + 0.006629317533224821, + 0.007746895309537649, + 0.004984593950212002, + 0.006345377769321203, + 0.005407342687249184, + 0.008877729997038841, + 0.005078764166682959, + 0.005385131109505892, + 0.010325032286345959, + 0.014631600119173527, + 0.013873935677111149, + 0.024436159059405327, + 0.017180392518639565, + 0.009768717922270298, + 0.013885818421840668, + 0.012296359986066818, + 0.002911330433562398, + 0.00408813776448369, + 0.003295315196737647, + 0.003928627353161573, + 0.0031431480310857296, + 0.004241760354489088, + 0.0038001909852027893, + 0.0030020952690392733, + 0.00592466676607728, + 0.005856943316757679, + 0.006669917609542608, + 0.01040361076593399, + 0.0038334745913743973, + 0.005130299367010593, + 0.003947633318603039, + 0.0021796366199851036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.04869633540511131 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005957228597253561 + ], + "opposite_logprob_tokens": [ + -5.130957126617432 + ], + "logprob_reward": -0.005957228597253561, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005957228597253561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03190504014492035, + 0.013013334944844246, + 0.018582696095108986, + 0.7578100562095642, + 0.03244738653302193, + 0.05243231728672981, + 0.02489496022462845, + 0.03666725382208824, + 0.007975786924362183, + 0.02427113801240921 + ], + "credit_map": [ + 0.0036153155378997326, + 0.003985000308603048, + 0.00239352323114872, + 0.003116790670901537, + 0.004865558817982674, + 0.0032507600262761116, + 0.006224099081009626, + 0.004453996662050486, + 0.0017477116780355573, + 0.001663088914938271, + 0.0012986380606889725, + 0.001204592757858336, + 0.0024269307032227516, + 0.0017247264040634036, + 0.0012347351294010878, + 0.0017129109473899007, + 0.0022976966574788094, + 0.0019744860474020243, + 0.0023054357152432203, + 0.002273520454764366, + 0.0036981136072427034, + 0.0022351318039000034, + 0.001883217366412282, + 0.001915094442665577, + 0.023383306339383125, + 0.025486353784799576, + 0.0382167287170887, + 0.31136566400527954, + 0.018175343051552773, + 0.024089859798550606, + 0.05822170153260231, + 0.25887101888656616, + 0.004393681418150663, + 0.001712384750135243, + 0.005378165747970343, + 0.0041210707277059555, + 0.0018242803635075688, + 0.002249398734420538, + 0.004163494799286127, + 0.008604911155998707, + 0.005664257798343897, + 0.00366949662566185, + 0.009052732959389687, + 0.006160184741020203, + 0.0045093572698533535, + 0.006111556198447943, + 0.0075270263478159904, + 0.009737706743180752, + 0.0024294601753354073, + 0.0038715198170393705, + 0.0022934935986995697, + 0.004726276267319918, + 0.0027313348837196827, + 0.0025816536508500576, + 0.0030690424609929323, + 0.003192179137840867, + 0.004134000279009342, + 0.0038185562007129192, + 0.004929773975163698, + 0.007899006828665733, + 0.0030016915407031775, + 0.002016482874751091, + 0.0056717428378760815, + 0.005195999052375555, + 0.001226227730512619, + 0.0009464483591727912, + 0.001026470330543816, + 0.0013096971670165658, + 0.0006416963296942413, + 0.00113067461643368, + 0.0010623139096423984, + 0.0006322595290839672, + 0.004218216985464096, + 0.002688734093680978, + 0.00451977364718914, + 0.004208241123706102, + 0.0025163409300148487, + 0.0023320207837969065, + 0.0021578955929726362, + 0.0016299152048304677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.005957228597253561 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520851492881775 + ], + "opposite_logprob_tokens": [ + -1.5020852088928223 + ], + "logprob_reward": -0.2520851492881775, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520851492881775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1688450425863266, + 0.11156024783849716, + 0.05392751097679138, + 0.35423484444618225, + 0.11800630390644073, + 0.08394583314657211, + 0.04459282383322716, + 0.021419476717710495, + 0.024117818102240562, + 0.019350122660398483 + ], + "credit_map": [ + 0.022628990933299065, + 0.021638061851263046, + 0.01068376749753952, + 0.014076774939894676, + 0.02423226088285446, + 0.013790820725262165, + 0.03173654526472092, + 0.030057815834879875, + 0.009227316826581955, + 0.010046635754406452, + 0.011024042963981628, + 0.011224462650716305, + 0.020240893587470055, + 0.014665133319795132, + 0.01855710707604885, + 0.016574658453464508, + 0.006334459874778986, + 0.005077828653156757, + 0.005335944704711437, + 0.008721702732145786, + 0.008393175899982452, + 0.005553147755563259, + 0.0075949858874082565, + 0.006916263606399298, + 0.024767866358160973, + 0.016226068139076233, + 0.016754642128944397, + 0.13453666865825653, + 0.009328149259090424, + 0.012667014263570309, + 0.021904507651925087, + 0.11804990470409393, + 0.010277229361236095, + 0.009068088605999947, + 0.024830617010593414, + 0.0233762264251709, + 0.007460369262844324, + 0.006937758531421423, + 0.011674248613417149, + 0.024381764233112335, + 0.005113870836794376, + 0.003672792576253414, + 0.019489077851176262, + 0.009195194579660892, + 0.006258328445255756, + 0.006612374912947416, + 0.015109133906662464, + 0.018495069816708565, + 0.0035563509445637465, + 0.005255206488072872, + 0.004707540851086378, + 0.008712206967175007, + 0.003769452916458249, + 0.00525290472432971, + 0.006931576877832413, + 0.0064075845293700695, + 0.0023277816362679005, + 0.002553430153056979, + 0.0027111296076327562, + 0.004218725021928549, + 0.002656959230080247, + 0.0020776374731212854, + 0.002368729095906019, + 0.0025050838012248278, + 0.0024809434544295073, + 0.002416160423308611, + 0.003067349549382925, + 0.0052161142230033875, + 0.002569249365478754, + 0.0018449707422405481, + 0.003329232335090637, + 0.0031937987077981234, + 0.0026277105789631605, + 0.0019803508184850216, + 0.003415545215830207, + 0.003974339459091425, + 0.001968392403796315, + 0.0018092818791046739, + 0.002224842319265008, + 0.0013496594037860632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.2520851492881775 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075802147388458 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011075802147388458, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075802147388458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0493677593767643, + 0.022340048104524612, + 0.05582186579704285, + 0.5355121493339539, + 0.046904657036066055, + 0.04165034741163254, + 0.037902284413576126, + 0.03814280033111572, + 0.02750246226787567, + 0.14485560357570648 + ], + "credit_map": [ + 0.007843529805541039, + 0.009845920838415623, + 0.004165605176240206, + 0.003944877535104752, + 0.00571824936196208, + 0.004709863569587469, + 0.006472935900092125, + 0.006666781846433878, + 0.0021288134157657623, + 0.00296819512732327, + 0.0025751213543117046, + 0.0027446062304079533, + 0.004113313276320696, + 0.002477878937497735, + 0.0026433360762894154, + 0.0026887867134064436, + 0.009936400689184666, + 0.005183805245906115, + 0.007090538740158081, + 0.006726590916514397, + 0.008470067754387856, + 0.006363429129123688, + 0.005407148040831089, + 0.006643889471888542, + 0.028783762827515602, + 0.02102343738079071, + 0.02859944850206375, + 0.19806121289730072, + 0.013839622028172016, + 0.02069765329360962, + 0.03589634597301483, + 0.1886107325553894, + 0.006606627721339464, + 0.002967279637232423, + 0.009697221219539642, + 0.006188961211591959, + 0.004842090420424938, + 0.003585305530577898, + 0.0056878807954490185, + 0.0073292977176606655, + 0.0032476731576025486, + 0.0026594242081046104, + 0.009187444113194942, + 0.005744852591305971, + 0.0033512136433273554, + 0.002943190513178706, + 0.006199958268553019, + 0.008316596038639545, + 0.003788431640714407, + 0.004880602937191725, + 0.0034946666564792395, + 0.008139274083077908, + 0.003946315031498671, + 0.00495741656050086, + 0.0037562749348580837, + 0.004939306993037462, + 0.0040046898648142815, + 0.003805532120168209, + 0.005338605493307114, + 0.008855453692376614, + 0.0027615681756287813, + 0.0028857081197202206, + 0.006007409654557705, + 0.004483836703002453, + 0.0066705914214253426, + 0.0033208969980478287, + 0.00408942811191082, + 0.00416176114231348, + 0.002300403080880642, + 0.0025896437000483274, + 0.0026395327877253294, + 0.0017302058404311538, + 0.021859612315893173, + 0.015547027811408043, + 0.03307846561074257, + 0.0237509086728096, + 0.009624581784009933, + 0.014617050997912884, + 0.0160230603069067, + 0.010354917496442795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 2 + ], + "selected_reward": -0.011075802147388458 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of sports photography, a powerful low-angle wide shot clings to the rough asphalt runway, capturing a B-17 bomber roaring in from the side of the frame with unstoppable force, instantly filling the frame with a decisive moment. Its iconic silver metal fuselage reflects a cool luster under the bright, glaring sunlight, its rivets and panel seams clearly visible at close range, exuding a pure industrial beauty. The propellers of its four engines spin at high speed, transforming into nearly transparent dynamic discs. The powerful airflow churns the air in front of the camera, stirring up fine dust on the runway surface. Simultaneously, the aircraft's massive silhouette casts a fast-moving, sharply defined dark shadow on the runway, almost in sync with the aircraft itself, flashing past the camera. The stark contrast between light and dark emphasizes its breathtaking speed. The entire scene is filled with mechanical power and breathtaking speed. The sky in the background and the distant horizon are blurred due to high-speed movement, focusing all visual focus on the stunning spectacle of this behemoth flying close to the ground, creating a tense and magnificent atmosphere.", + "seed": 22065, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_9/index_9127_seed_22065_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05331482663750649, + "vlm_logprob": -0.05331482663750649, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.05331482663750649, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1862705498933792 + ], + "opposite_logprob_tokens": [ + -1.8112705945968628 + ], + "logprob_reward": -0.1862705498933792, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1862705498933792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10896657407283783, + 0.07249721884727478, + 0.1588527411222458, + 0.12235432863235474, + 0.18408085405826569, + 0.11048191040754318, + 0.047920338809490204, + 0.07507849484682083, + 0.05222949758172035, + 0.06753804534673691 + ], + "credit_map": [ + 0.012840085662901402, + 0.019674649462103844, + 0.010747273452579975, + 0.01146882213652134, + 0.016026657074689865, + 0.01767808012664318, + 0.011124376207590103, + 0.009406629018485546, + 0.007319471333175898, + 0.014118428342044353, + 0.0069956062361598015, + 0.009523295797407627, + 0.00934534054249525, + 0.008620255626738071, + 0.007460486609488726, + 0.009114332497119904, + 0.04002680629491806, + 0.01793798618018627, + 0.037100739777088165, + 0.012381873093545437, + 0.013186600990593433, + 0.01614699326455593, + 0.013482128269970417, + 0.008589619770646095, + 0.01814819872379303, + 0.02634517475962639, + 0.020097101107239723, + 0.01751019060611725, + 0.015314830467104912, + 0.011533085256814957, + 0.00847270991653204, + 0.004933048505336046, + 0.032906051725149155, + 0.04054310545325279, + 0.022009044885635376, + 0.021731367334723473, + 0.01933707483112812, + 0.020132500678300858, + 0.013948548585176468, + 0.013473167084157467, + 0.013379418291151524, + 0.01676144078373909, + 0.02172781713306904, + 0.014804643578827381, + 0.011787943542003632, + 0.009889715351164341, + 0.010276637971401215, + 0.011854294687509537, + 0.0049108462408185005, + 0.008141934871673584, + 0.006277003791183233, + 0.006420493125915527, + 0.0069557600654661655, + 0.0070404489524662495, + 0.0037888260558247566, + 0.004385026171803474, + 0.007114571984857321, + 0.013338346034288406, + 0.013392959721386433, + 0.011124972254037857, + 0.009046122431755066, + 0.008729173801839352, + 0.006037995219230652, + 0.006294346414506435, + 0.009742245078086853, + 0.007516087032854557, + 0.0048498110845685005, + 0.007774468511343002, + 0.006241827271878719, + 0.005441267974674702, + 0.004710004664957523, + 0.005953787826001644, + 0.00573789793998003, + 0.01867995783686638, + 0.0046752155758440495, + 0.0047844708897173405, + 0.014004937373101711, + 0.00728860218077898, + 0.006060440558940172, + 0.006306523457169533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.1862705498933792 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337274894118309 + ], + "opposite_logprob_tokens": [ + -3.4087274074554443 + ], + "logprob_reward": -0.0337274894118309, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337274894118309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28171053528785706, + 0.06384214758872986, + 0.16928455233573914, + 0.06428922712802887, + 0.07863930612802505, + 0.07073941826820374, + 0.08305741101503372, + 0.08344554901123047, + 0.04410829395055771, + 0.060883574187755585 + ], + "credit_map": [ + 0.024591980502009392, + 0.04651854932308197, + 0.03874221071600914, + 0.04062814265489578, + 0.02956496551632881, + 0.03127151355147362, + 0.040987517684698105, + 0.029405640438199043, + 0.007729690056294203, + 0.01051811408251524, + 0.010189485736191273, + 0.006610733922570944, + 0.007087791338562965, + 0.009303993545472622, + 0.004959253594279289, + 0.0074430787935853004, + 0.04125380888581276, + 0.024525366723537445, + 0.036237653344869614, + 0.012328692711889744, + 0.014529114589095116, + 0.018234441056847572, + 0.012487954460084438, + 0.009687496349215508, + 0.0075577483512461185, + 0.006642463617026806, + 0.007895433343946934, + 0.009287490509450436, + 0.007281606085598469, + 0.013950848951935768, + 0.006633564364165068, + 0.005040069110691547, + 0.015650004148483276, + 0.020908871665596962, + 0.00888165645301342, + 0.0055053625255823135, + 0.008356744423508644, + 0.01076838281005621, + 0.004417218267917633, + 0.0041510495357215405, + 0.007630117703229189, + 0.01186325028538704, + 0.007855106145143509, + 0.008098321035504341, + 0.013007253408432007, + 0.007132321130484343, + 0.00601567467674613, + 0.009137366898357868, + 0.00796459149569273, + 0.013077917508780956, + 0.012040925212204456, + 0.012507947161793709, + 0.012270929291844368, + 0.011264339089393616, + 0.0071541862562298775, + 0.00677656102925539, + 0.007698986679315567, + 0.013570073060691357, + 0.01600971631705761, + 0.011814315803349018, + 0.011578493751585484, + 0.012617144733667374, + 0.004443633835762739, + 0.005713168065994978, + 0.004753337241709232, + 0.007572909351438284, + 0.005312506575137377, + 0.004329775460064411, + 0.005963325034826994, + 0.0056615169160068035, + 0.005727408453822136, + 0.004787511192262173, + 0.014081661589443684, + 0.008306375704705715, + 0.004263221751898527, + 0.003699458669871092, + 0.01007620058953762, + 0.011111914180219173, + 0.005014894064515829, + 0.004329842049628496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.0337274894118309 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052625699900090694 + ], + "opposite_logprob_tokens": [ + -5.25526237487793 + ], + "logprob_reward": -0.0052625699900090694, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052625699900090694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22038787603378296, + 0.10775364935398102, + 0.21592970192432404, + 0.14055535197257996, + 0.03755997493863106, + 0.047630563378334045, + 0.07410275191068649, + 0.03685405105352402, + 0.044491250067949295, + 0.07473482191562653 + ], + "credit_map": [ + 0.018231317400932312, + 0.047154489904642105, + 0.02632017247378826, + 0.024696743115782738, + 0.023791559040546417, + 0.032166700810194016, + 0.026030153036117554, + 0.021996719762682915, + 0.015151378698647022, + 0.029038233682513237, + 0.009382100775837898, + 0.010321117006242275, + 0.010622442699968815, + 0.017133284360170364, + 0.006721396930515766, + 0.00938369520008564, + 0.05709778890013695, + 0.028797505423426628, + 0.04899584874510765, + 0.015194418840110302, + 0.016780434176325798, + 0.022334132343530655, + 0.015714120119810104, + 0.011015460826456547, + 0.02158508449792862, + 0.026194939389824867, + 0.01573832333087921, + 0.015406149439513683, + 0.0195961594581604, + 0.021421145647764206, + 0.010466656647622585, + 0.010146900080144405, + 0.008180339820683002, + 0.007263696752488613, + 0.005536913406103849, + 0.002924147993326187, + 0.0035623328294605017, + 0.005063754040747881, + 0.0020729578100144863, + 0.0029558311216533184, + 0.006966453976929188, + 0.01013085339218378, + 0.006764029152691364, + 0.005218566861003637, + 0.0063846781849861145, + 0.0034184998366981745, + 0.004175498150289059, + 0.004571984056383371, + 0.007844340987503529, + 0.010900046676397324, + 0.011061020195484161, + 0.014252838678658009, + 0.007753034587949514, + 0.0094218198210001, + 0.006443673744797707, + 0.0064259786158800125, + 0.004802405834197998, + 0.006890366785228252, + 0.007894068956375122, + 0.0051106540486216545, + 0.0032810638658702374, + 0.0038366366643458605, + 0.0025411800015717745, + 0.002497673500329256, + 0.003356131725013256, + 0.005311578046530485, + 0.004685836378484964, + 0.005586662329733372, + 0.007410629652440548, + 0.007522664964199066, + 0.005646218080073595, + 0.00497153215110302, + 0.007312529720366001, + 0.021133054047822952, + 0.004882367793470621, + 0.004104753956198692, + 0.015335106290876865, + 0.00882688257843256, + 0.00662212772294879, + 0.006517994683235884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0052625699900090694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033686112612485886 + ], + "opposite_logprob_tokens": [ + -3.4086861610412598 + ], + "logprob_reward": -0.033686112612485886, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033686112612485886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2797973155975342, + 0.1026962473988533, + 0.10338108241558075, + 0.14062124490737915, + 0.12720204889774323, + 0.027385640889406204, + 0.07832566648721695, + 0.03729110583662987, + 0.029200324788689613, + 0.07409932464361191 + ], + "credit_map": [ + 0.028583068400621414, + 0.05122244730591774, + 0.02459907718002796, + 0.031283989548683167, + 0.051022037863731384, + 0.04285350441932678, + 0.031082160770893097, + 0.019151024520397186, + 0.010248146019876003, + 0.017438212409615517, + 0.014096022583544254, + 0.014120117761194706, + 0.01274092122912407, + 0.01408753078430891, + 0.007709174416959286, + 0.012256124056875706, + 0.02666814625263214, + 0.012638498097658157, + 0.0244595929980278, + 0.008083553984761238, + 0.009201643988490105, + 0.008943798020482063, + 0.007792580407112837, + 0.005593270529061556, + 0.022165827453136444, + 0.015503782778978348, + 0.019831329584121704, + 0.01286288071423769, + 0.019170058891177177, + 0.028726499527692795, + 0.011557210236787796, + 0.010803654789924622, + 0.01540568470954895, + 0.024349277839064598, + 0.015327060595154762, + 0.018436718732118607, + 0.015883706510066986, + 0.018912747502326965, + 0.009944768622517586, + 0.008942094631493092, + 0.003719751024618745, + 0.004537439905107021, + 0.003911003936082125, + 0.0019986170809715986, + 0.005157849285751581, + 0.003644710872322321, + 0.002083855215460062, + 0.0023324142675846815, + 0.011829668655991554, + 0.01048182137310505, + 0.01259399950504303, + 0.010673328302800655, + 0.010401003062725067, + 0.008940063416957855, + 0.006403426174074411, + 0.007002354133874178, + 0.0037281957920640707, + 0.005812079180032015, + 0.004537612199783325, + 0.005174172110855579, + 0.0072784158401191235, + 0.0061568752862513065, + 0.0021937584970146418, + 0.002409996697679162, + 0.0035434241872280836, + 0.0027491056825965643, + 0.005912581458687782, + 0.0051443735137581825, + 0.00305561488494277, + 0.0025295009836554527, + 0.0030958992429077625, + 0.0031698269303888083, + 0.006883298046886921, + 0.019657569006085396, + 0.00522554200142622, + 0.004327071364969015, + 0.015416605398058891, + 0.009699080139398575, + 0.006743916776031256, + 0.00614624097943306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.033686112612485886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627411279827356 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.007627411279827356, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627411279827356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16341891884803772, + 0.11332366615533829, + 0.08598782122135162, + 0.043679241091012955, + 0.03579748421907425, + 0.3102547228336334, + 0.12460721284151077, + 0.03217817470431328, + 0.06013287603855133, + 0.03061988204717636 + ], + "credit_map": [ + 0.0160536952316761, + 0.024503832682967186, + 0.02430630847811699, + 0.020541049540042877, + 0.021678797900676727, + 0.023284606635570526, + 0.016440974548459053, + 0.016609642654657364, + 0.015642428770661354, + 0.018706107512116432, + 0.015820728614926338, + 0.016577450558543205, + 0.00974759366363287, + 0.018508674576878548, + 0.0075728376396000385, + 0.010747842490673065, + 0.0200863778591156, + 0.010200342163443565, + 0.016589205712080002, + 0.00896105170249939, + 0.008168991655111313, + 0.009523360058665276, + 0.00633325194939971, + 0.006125247105956078, + 0.005841458681970835, + 0.004586894065141678, + 0.006142321974039078, + 0.004906259011477232, + 0.00570729048922658, + 0.008671676740050316, + 0.003949368372559547, + 0.003873971989378333, + 0.006864381954073906, + 0.007240349426865578, + 0.004380154423415661, + 0.002193459076806903, + 0.00578170595690608, + 0.004234032239764929, + 0.002578679472208023, + 0.0025247205048799515, + 0.042598381638526917, + 0.05424170568585396, + 0.05347047001123428, + 0.04368221387267113, + 0.031217271462082863, + 0.02903340943157673, + 0.023280195891857147, + 0.03273104876279831, + 0.01063069049268961, + 0.016429495066404343, + 0.019858891144394875, + 0.024557966738939285, + 0.015835903584957123, + 0.017543023452162743, + 0.010087861679494381, + 0.009663381613790989, + 0.003273499198257923, + 0.006152746733278036, + 0.0060349395498633385, + 0.004671428352594376, + 0.003207567846402526, + 0.0035890736617147923, + 0.0029039124492555857, + 0.0023450066801160574, + 0.006224809680134058, + 0.005428786389529705, + 0.008097470737993717, + 0.005816775839775801, + 0.007631182670593262, + 0.01167936809360981, + 0.007417792920023203, + 0.007836691103875637, + 0.005538512021303177, + 0.005225285422056913, + 0.0028949773404747248, + 0.0027031423524022102, + 0.0038797063753008842, + 0.0042032357305288315, + 0.0027782537508755922, + 0.0033967674244195223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.007627411279827356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of sports photography, a powerful low-angle wide shot clings to the rough asphalt runway, capturing a B-17 bomber roaring in from the side of the frame with unstoppable force, instantly filling the frame with a decisive moment. Its iconic silver metal fuselage reflects a cool luster under the bright, glaring sunlight, its rivets and panel seams clearly visible at close range, exuding a pure industrial beauty. The propellers of its four engines spin at high speed, transforming into nearly transparent dynamic discs. The powerful airflow churns the air in front of the camera, stirring up fine dust on the runway surface. Simultaneously, the aircraft's massive silhouette casts a fast-moving, sharply defined dark shadow on the runway, almost in sync with the aircraft itself, flashing past the camera. The stark contrast between light and dark emphasizes its breathtaking speed. The entire scene is filled with mechanical power and breathtaking speed. The sky in the background and the distant horizon are blurred due to high-speed movement, focusing all visual focus on the stunning spectacle of this behemoth flying close to the ground, creating a tense and magnificent atmosphere.", + "seed": 22066, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_9/index_9127_seed_22066_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05773577084764838, + "vlm_logprob": -0.05773577084764838, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.05773577084764838, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20650582015514374 + ], + "opposite_logprob_tokens": [ + -1.7065057754516602 + ], + "logprob_reward": -0.20650582015514374, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20650582015514374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12147194892168045, + 0.3666497468948364, + 0.12529194355010986, + 0.06553833186626434, + 0.056489959359169006, + 0.033598899841308594, + 0.020311012864112854, + 0.10208148509263992, + 0.09657622128725052, + 0.011990483850240707 + ], + "credit_map": [ + 0.012068177573382854, + 0.019964808598160744, + 0.019949616864323616, + 0.012935524806380272, + 0.014989037066698074, + 0.014229366555809975, + 0.015246247872710228, + 0.012089168652892113, + 0.033118873834609985, + 0.053292352706193924, + 0.07310270518064499, + 0.040604159235954285, + 0.057543668895959854, + 0.05581681802868843, + 0.028704272583127022, + 0.024466902017593384, + 0.013459431938827038, + 0.016293685883283615, + 0.008670066483318806, + 0.010410039685666561, + 0.010624703951179981, + 0.013328466564416885, + 0.018248504027724266, + 0.03425705060362816, + 0.005637208931148052, + 0.007163578644394875, + 0.006559770554304123, + 0.007487158291041851, + 0.00914439745247364, + 0.008316723629832268, + 0.010504791513085365, + 0.010724704712629318, + 0.010175072588026524, + 0.010299494490027428, + 0.004504292272031307, + 0.004738076590001583, + 0.00435930211097002, + 0.0046763187274336815, + 0.012807493098080158, + 0.00492990855127573, + 0.005297820549458265, + 0.004664095118641853, + 0.0052092974074184895, + 0.0041227564215660095, + 0.004983141552656889, + 0.0036312893498688936, + 0.0029736771248281, + 0.0027168223168700933, + 0.0034565201494842768, + 0.002151052001863718, + 0.0018033196683973074, + 0.0019826956558972597, + 0.002536015585064888, + 0.004129105247557163, + 0.0019204351119697094, + 0.0023318701423704624, + 0.01613556407392025, + 0.01411337498575449, + 0.009396864101290703, + 0.017230050638318062, + 0.010543144308030605, + 0.010011352598667145, + 0.014492571353912354, + 0.010158566758036613, + 0.018479615449905396, + 0.01584700495004654, + 0.008120482787489891, + 0.005536194890737534, + 0.014525151811540127, + 0.01749364472925663, + 0.006292238365858793, + 0.01028189156204462, + 0.001257551135495305, + 0.0014286546502262354, + 0.0015604126965627074, + 0.0008481739205308259, + 0.002004923764616251, + 0.0018674239981919527, + 0.001467877416871488, + 0.0015554663259536028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.20650582015514374 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038127824664115906 + ], + "opposite_logprob_tokens": [ + -3.288127899169922 + ], + "logprob_reward": -0.038127824664115906, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038127824664115906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17604032158851624, + 0.08617904037237167, + 0.10587630420923233, + 0.07141183316707611, + 0.19097451865673065, + 0.10891694575548172, + 0.07093233615159988, + 0.06741999834775925, + 0.09596540778875351, + 0.026283279061317444 + ], + "credit_map": [ + 0.01928354799747467, + 0.02757108211517334, + 0.024855589494109154, + 0.02004978246986866, + 0.019541416317224503, + 0.01981998048722744, + 0.02186974324285984, + 0.02304919622838497, + 0.008424574509263039, + 0.009797921404242516, + 0.011977367103099823, + 0.013298053294420242, + 0.014880957081913948, + 0.010646248236298561, + 0.006143462844192982, + 0.011010459624230862, + 0.011980351991951466, + 0.017855161800980568, + 0.00930729229003191, + 0.01204504631459713, + 0.014964966103434563, + 0.012964359484612942, + 0.016550416126847267, + 0.010208703577518463, + 0.01026215124875307, + 0.006846579723060131, + 0.014951935037970543, + 0.005869217682629824, + 0.005893238354474306, + 0.010282731615006924, + 0.009056472219526768, + 0.008249510079622269, + 0.041045546531677246, + 0.06442207843065262, + 0.007350896950811148, + 0.011110439896583557, + 0.015822110697627068, + 0.019021304324269295, + 0.020153425633907318, + 0.01204872690141201, + 0.01632138341665268, + 0.014638546854257584, + 0.010935376398265362, + 0.01022307574748993, + 0.013539081439375877, + 0.016282007098197937, + 0.014950194396078587, + 0.01202727947384119, + 0.01249387301504612, + 0.011161082424223423, + 0.007129595149308443, + 0.006906147114932537, + 0.007377192378044128, + 0.00970992911607027, + 0.006159990560263395, + 0.009994525462388992, + 0.011260190047323704, + 0.012169956229627132, + 0.006909431889653206, + 0.00732074910774827, + 0.007850905880331993, + 0.010291211307048798, + 0.005489709787070751, + 0.006127844098955393, + 0.01664356328547001, + 0.012823017314076424, + 0.008545247837901115, + 0.0077260881662368774, + 0.015874795615673065, + 0.017943428829312325, + 0.006409612018615007, + 0.009999657981097698, + 0.003143866080790758, + 0.0038105591665953398, + 0.0023011702578514814, + 0.002787681994959712, + 0.004107190296053886, + 0.004557355772703886, + 0.0025499057956039906, + 0.003025549463927746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.038127824664115906 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003624419681727886 + ], + "opposite_logprob_tokens": [ + -5.628624439239502 + ], + "logprob_reward": -0.003624419681727886, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003624419681727886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11765142530202866, + 0.5099805593490601, + 0.09593379497528076, + 0.04735994711518288, + 0.06031497195363045, + 0.024596799165010452, + 0.038267090916633606, + 0.048381444066762924, + 0.041478727012872696, + 0.01603527180850506 + ], + "credit_map": [ + 0.01205182820558548, + 0.02163742482662201, + 0.019171487540006638, + 0.007218556944280863, + 0.018129153177142143, + 0.013968080282211304, + 0.008288541808724403, + 0.017186351120471954, + 0.04642794281244278, + 0.07532931864261627, + 0.09221313148736954, + 0.05706297606229782, + 0.08660155534744263, + 0.0763891264796257, + 0.04027861729264259, + 0.03567790985107422, + 0.008049787022173405, + 0.009019432589411736, + 0.00666193850338459, + 0.009848757646977901, + 0.007796364836394787, + 0.010237406007945538, + 0.015305643901228905, + 0.0290144681930542, + 0.006444100756198168, + 0.005820670165121555, + 0.010002522729337215, + 0.0041284142062067986, + 0.005009004846215248, + 0.006729634013026953, + 0.005833487492054701, + 0.0033921163994818926, + 0.010242135263979435, + 0.014201068319380283, + 0.002975939540192485, + 0.004825306590646505, + 0.005262547172605991, + 0.004031658638268709, + 0.01428528968244791, + 0.004491036292165518, + 0.0031153629533946514, + 0.0022666577715426683, + 0.004316315986216068, + 0.005014160647988319, + 0.002805649768561125, + 0.0028446544893085957, + 0.002362606581300497, + 0.0018713920144364238, + 0.006458188407123089, + 0.004815075546503067, + 0.003266548039391637, + 0.004901503212749958, + 0.004732184577733278, + 0.005999173037707806, + 0.0027542475145310163, + 0.005340178031474352, + 0.007513608317822218, + 0.007087035104632378, + 0.004490986932069063, + 0.0076298341155052185, + 0.0049409870989620686, + 0.004959946498274803, + 0.007302897982299328, + 0.004456150811165571, + 0.006462047807872295, + 0.00867198035120964, + 0.0029688721988350153, + 0.0024608250241726637, + 0.009132090955972672, + 0.007313070353120565, + 0.0021762894466519356, + 0.0022935543674975634, + 0.0022412999533116817, + 0.001834573340602219, + 0.0009488951181992888, + 0.0005813658353872597, + 0.003812721697613597, + 0.004360915161669254, + 0.0010296027176082134, + 0.001225898158736527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.003624419681727886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033685192465782166 + ], + "opposite_logprob_tokens": [ + -3.4086852073669434 + ], + "logprob_reward": -0.033685192465782166, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033685192465782166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1944437175989151, + 0.3096359074115753, + 0.051921069622039795, + 0.18043342232704163, + 0.04742773249745369, + 0.041528888046741486, + 0.06630188226699829, + 0.05936931073665619, + 0.0242009237408638, + 0.024737145751714706 + ], + "credit_map": [ + 0.013793830759823322, + 0.019727293401956558, + 0.01723000593483448, + 0.030830875039100647, + 0.016454841941595078, + 0.018334178254008293, + 0.03983323648571968, + 0.03823944926261902, + 0.027562256902456284, + 0.04586183652281761, + 0.056241922080516815, + 0.03516465052962303, + 0.05306841805577278, + 0.04528411850333214, + 0.02466864511370659, + 0.021784039214253426, + 0.005855750758200884, + 0.006656625773757696, + 0.003928675781935453, + 0.006605239585042, + 0.006664934102445841, + 0.006005563773214817, + 0.008146676234900951, + 0.008057602681219578, + 0.02628416195511818, + 0.012376147322356701, + 0.04149675369262695, + 0.01909438520669937, + 0.012426172383129597, + 0.027412056922912598, + 0.02017905004322529, + 0.02116469480097294, + 0.006392970215529203, + 0.005380155518651009, + 0.0035855374298989773, + 0.004335850942879915, + 0.006086219567805529, + 0.008539988659322262, + 0.009457658976316452, + 0.0036493497900664806, + 0.006668953690677881, + 0.0055434005334973335, + 0.004936551209539175, + 0.005243503954261541, + 0.0074131847359240055, + 0.005475231911987066, + 0.0033803994301706553, + 0.002867663512006402, + 0.014345508068799973, + 0.008325173519551754, + 0.0065011209808290005, + 0.0040055629797279835, + 0.00909449066966772, + 0.010273545980453491, + 0.00667050713673234, + 0.0070859771221876144, + 0.01185655128210783, + 0.01100681722164154, + 0.005460970103740692, + 0.00694621168076992, + 0.006867299322038889, + 0.006678742356598377, + 0.004716737195849419, + 0.005835983902215958, + 0.0033962957095354795, + 0.003839396173134446, + 0.0016993723111227155, + 0.0014962913701310754, + 0.005462050903588533, + 0.004527737386524677, + 0.0016464076470583677, + 0.0021333720069378614, + 0.00233877613209188, + 0.003976528532803059, + 0.0027848726604133844, + 0.0024142407346516848, + 0.0029600036796182394, + 0.004908180795609951, + 0.0018433090299367905, + 0.003511235350742936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.033685192465782166 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735597271472216 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735597271472216, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735597271472216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22606052458286285, + 0.3994893729686737, + 0.05719945207238197, + 0.04057956486940384, + 0.09397826343774796, + 0.02334817498922348, + 0.06740961223840714, + 0.04537868872284889, + 0.030236240476369858, + 0.016320157796144485 + ], + "credit_map": [ + 0.029150689020752907, + 0.019666338339447975, + 0.061331648379564285, + 0.020489893853664398, + 0.020978085696697235, + 0.03411388769745827, + 0.019738202914595604, + 0.02059178426861763, + 0.03580700606107712, + 0.05935964360833168, + 0.0725998729467392, + 0.04305187612771988, + 0.06869454681873322, + 0.061484191566705704, + 0.03039311058819294, + 0.02809913642704487, + 0.008404306136071682, + 0.011357289738953114, + 0.005185719113796949, + 0.0061857071705162525, + 0.007042719051241875, + 0.005540498532354832, + 0.007246570196002722, + 0.0062366388738155365, + 0.004026345442980528, + 0.003763291984796524, + 0.005889160092920065, + 0.0035609453916549683, + 0.004343852866441011, + 0.005169123411178589, + 0.007410304620862007, + 0.006416539195924997, + 0.010048777796328068, + 0.013371276669204235, + 0.006497965659946203, + 0.007094811648130417, + 0.0070497626438736916, + 0.012011943385004997, + 0.02936789020895958, + 0.008535828441381454, + 0.003214915283024311, + 0.0017474059714004397, + 0.0026473284233361483, + 0.003071493934839964, + 0.0030121298041194677, + 0.003120802342891693, + 0.002505767624825239, + 0.004028331022709608, + 0.014691047370433807, + 0.008102708496153355, + 0.006394735537469387, + 0.005219405982643366, + 0.008436563424766064, + 0.0104225417599082, + 0.006654744502156973, + 0.007487867493182421, + 0.009777523577213287, + 0.00733796413987875, + 0.00385291432030499, + 0.004057092126458883, + 0.0049927146174013615, + 0.0059996554628014565, + 0.005162566434592009, + 0.004198254086077213, + 0.004834085702896118, + 0.005801748018711805, + 0.0027848873287439346, + 0.001829027314670384, + 0.005023020785301924, + 0.005793678108602762, + 0.002202198375016451, + 0.0019675958901643753, + 0.002169571816921234, + 0.0023792367428541183, + 0.0012950475793331861, + 0.0013529642019420862, + 0.0026273413095623255, + 0.0032428426202386618, + 0.001503943232819438, + 0.001749210525304079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.006735597271472216 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of sports photography, a powerful low-angle wide shot clings to the rough asphalt runway, capturing a B-17 bomber roaring in from the side of the frame with unstoppable force, instantly filling the frame with a decisive moment. Its iconic silver metal fuselage reflects a cool luster under the bright, glaring sunlight, its rivets and panel seams clearly visible at close range, exuding a pure industrial beauty. The propellers of its four engines spin at high speed, transforming into nearly transparent dynamic discs. The powerful airflow churns the air in front of the camera, stirring up fine dust on the runway surface. Simultaneously, the aircraft's massive silhouette casts a fast-moving, sharply defined dark shadow on the runway, almost in sync with the aircraft itself, flashing past the camera. The stark contrast between light and dark emphasizes its breathtaking speed. The entire scene is filled with mechanical power and breathtaking speed. The sky in the background and the distant horizon are blurred due to high-speed movement, focusing all visual focus on the stunning spectacle of this behemoth flying close to the ground, creating a tense and magnificent atmosphere.", + "seed": 22067, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000037/group_9/index_9127_seed_22067_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06804295973852277, + "vlm_logprob": -0.06804295973852277, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.06804295973852277, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23597797751426697 + ], + "opposite_logprob_tokens": [ + -1.6109780073165894 + ], + "logprob_reward": -0.23597797751426697, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23597797751426697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09448716789484024, + 0.18026785552501678, + 0.08886189758777618, + 0.0536307729780674, + 0.05100090429186821, + 0.28208523988723755, + 0.04151767119765282, + 0.11570321768522263, + 0.04502664878964424, + 0.04741867259144783 + ], + "credit_map": [ + 0.007744421251118183, + 0.016465788707137108, + 0.013899148441851139, + 0.012128702364861965, + 0.00805638451129198, + 0.007869960740208626, + 0.012450609356164932, + 0.015872148796916008, + 0.01765344850718975, + 0.011426237411797047, + 0.010943122208118439, + 0.01835109107196331, + 0.012627571821212769, + 0.011052657850086689, + 0.048125941306352615, + 0.05008777976036072, + 0.010584515519440174, + 0.015847766771912575, + 0.008642244152724743, + 0.022033896297216415, + 0.006396967452019453, + 0.008991504088044167, + 0.006549409590661526, + 0.009815596975386143, + 0.00646522082388401, + 0.00679070083424449, + 0.003947063349187374, + 0.0053231134079396725, + 0.005380138289183378, + 0.01063207071274519, + 0.00441383384168148, + 0.010678633116185665, + 0.008967412635684013, + 0.007736600935459137, + 0.003606452140957117, + 0.004961942322552204, + 0.009439026936888695, + 0.006520136725157499, + 0.004113109316676855, + 0.005656223278492689, + 0.02329285815358162, + 0.08175819367170334, + 0.07381164282560349, + 0.02008765935897827, + 0.02963826060295105, + 0.018979830667376518, + 0.01715475507080555, + 0.017362037673592567, + 0.00615689205005765, + 0.00659001711755991, + 0.004076752346009016, + 0.0034430816303938627, + 0.008956980891525745, + 0.003742215223610401, + 0.004697707947343588, + 0.003854026086628437, + 0.017394186928868294, + 0.01600910909473896, + 0.009574377909302711, + 0.010118846781551838, + 0.015212059952318668, + 0.019038096070289612, + 0.01311543583869934, + 0.015241104178130627, + 0.006495777051895857, + 0.005207786336541176, + 0.005662191659212112, + 0.003917424939572811, + 0.010267063975334167, + 0.004984571598470211, + 0.004482248332351446, + 0.004009584430605173, + 0.006474378984421492, + 0.005008475389331579, + 0.00907918531447649, + 0.009419906884431839, + 0.0042166681960225105, + 0.004057828336954117, + 0.004365135449916124, + 0.004797093104571104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.23597797751426697 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870212823152542 + ], + "opposite_logprob_tokens": [ + -3.0487022399902344 + ], + "logprob_reward": -0.04870212823152542, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870212823152542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10017098486423492, + 0.1689344197511673, + 0.12156321853399277, + 0.029876846820116043, + 0.02966184727847576, + 0.24260324239730835, + 0.10597944259643555, + 0.10503170639276505, + 0.035196106880903244, + 0.060982171446084976 + ], + "credit_map": [ + 0.013118050992488861, + 0.019066836684942245, + 0.01376189012080431, + 0.011462541297078133, + 0.012175614014267921, + 0.007863987237215042, + 0.010900702327489853, + 0.011821364983916283, + 0.022553814575076103, + 0.012009650468826294, + 0.01637701690196991, + 0.013638310134410858, + 0.011365282349288464, + 0.021214738488197327, + 0.045512985438108444, + 0.026262611150741577, + 0.027807852253317833, + 0.019492579624056816, + 0.007989452220499516, + 0.01754114404320717, + 0.01906578801572323, + 0.010962224565446377, + 0.008783360943198204, + 0.009920810349285603, + 0.003611208638176322, + 0.003921352792531252, + 0.0028711846098303795, + 0.0034142297226935625, + 0.003388785058632493, + 0.004540632013231516, + 0.002562918933108449, + 0.005566535983234644, + 0.0029509139712899923, + 0.0030770080629736185, + 0.0034714285284280777, + 0.00357823446393013, + 0.005592777393758297, + 0.0039336285553872585, + 0.004253861028701067, + 0.0028039959724992514, + 0.02148117683827877, + 0.07857004553079605, + 0.048437003046274185, + 0.011799832805991173, + 0.025761304423213005, + 0.019615117460489273, + 0.02161186747252941, + 0.015326903201639652, + 0.01898241601884365, + 0.017089879140257835, + 0.009503110311925411, + 0.007081759627908468, + 0.018283691257238388, + 0.010126974433660507, + 0.016361217945814133, + 0.00855038408190012, + 0.017633624374866486, + 0.01606697402894497, + 0.006945776287466288, + 0.007632268592715263, + 0.013850989751517773, + 0.02032916434109211, + 0.0120589230209589, + 0.010513987392187119, + 0.0033071618527173996, + 0.002687849337235093, + 0.007909215055406094, + 0.002686151536181569, + 0.008693743497133255, + 0.0035459473729133606, + 0.0034327926114201546, + 0.0029332430567592382, + 0.008575293235480785, + 0.010029966942965984, + 0.009408941492438316, + 0.011114713735878468, + 0.005151501391083002, + 0.007227452006191015, + 0.004171950276941061, + 0.005302353762090206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.04870212823152542 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007645393256098032 + ], + "opposite_logprob_tokens": [ + -4.882645606994629 + ], + "logprob_reward": -0.007645393256098032, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007645393256098032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12158194184303284, + 0.10541892796754837, + 0.040117714554071426, + 0.07818333804607391, + 0.028720824047923088, + 0.19729197025299072, + 0.10179765522480011, + 0.15069237351417542, + 0.10732068121433258, + 0.06887456774711609 + ], + "credit_map": [ + 0.012894697487354279, + 0.024604367092251778, + 0.014772969298064709, + 0.01304458174854517, + 0.0186292864382267, + 0.008440274745225906, + 0.012836150825023651, + 0.016359619796276093, + 0.012629269622266293, + 0.008520280010998249, + 0.013376710005104542, + 0.011275374330580235, + 0.008724592626094818, + 0.013302955776453018, + 0.024586273357272148, + 0.013003470376133919, + 0.005389242433011532, + 0.004306974355131388, + 0.005845965817570686, + 0.0051088775508105755, + 0.005623990204185247, + 0.004361770581454039, + 0.003104929579421878, + 0.006375961471349001, + 0.010172802023589611, + 0.007205998059362173, + 0.005873560439795256, + 0.006575705483555794, + 0.008925419300794601, + 0.011218782514333725, + 0.005655643995851278, + 0.022555427625775337, + 0.002576058730483055, + 0.003517945297062397, + 0.0020488263107836246, + 0.005580872762948275, + 0.004787961952388287, + 0.003834183793514967, + 0.0030921066645532846, + 0.0032828703988343477, + 0.01927228271961212, + 0.054644618183374405, + 0.05343532934784889, + 0.014788713306188583, + 0.021220579743385315, + 0.013510831631720066, + 0.010056928731501102, + 0.010362708009779453, + 0.015084322541952133, + 0.016363568603992462, + 0.008448829874396324, + 0.008462055586278439, + 0.01931217685341835, + 0.010249563492834568, + 0.014937685802578926, + 0.008939448744058609, + 0.025690097361803055, + 0.020819567143917084, + 0.010253561660647392, + 0.011738238856196404, + 0.01914823055267334, + 0.029783539474010468, + 0.01565677486360073, + 0.017602356150746346, + 0.007874129340052605, + 0.009515210054814816, + 0.026453595608472824, + 0.007442959118634462, + 0.03583936393260956, + 0.008393495343625546, + 0.005392059218138456, + 0.006409861613065004, + 0.009820697829127312, + 0.007011324632912874, + 0.01471260841935873, + 0.013148123398423195, + 0.007943934760987759, + 0.005070862360298634, + 0.005885446444153786, + 0.00528157502412796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.007645393256098032 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336914137005806 + ], + "opposite_logprob_tokens": [ + -3.40869140625 + ], + "logprob_reward": -0.0336914137005806, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336914137005806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10515300184488297, + 0.061967529356479645, + 0.05056153982877731, + 0.03923287242650986, + 0.026791920885443687, + 0.07351913303136826, + 0.06206420809030533, + 0.44944125413894653, + 0.04498201236128807, + 0.08628646284341812 + ], + "credit_map": [ + 0.0086196418851614, + 0.014854631386697292, + 0.01359503623098135, + 0.01361293625086546, + 0.02091805264353752, + 0.009025298058986664, + 0.013198724016547203, + 0.011328703723847866, + 0.006354909390211105, + 0.004015100654214621, + 0.004268816206604242, + 0.005723495967686176, + 0.004482355900108814, + 0.013089843094348907, + 0.01783125475049019, + 0.006201762706041336, + 0.00992379616945982, + 0.004223699681460857, + 0.0041987234726548195, + 0.005124812480062246, + 0.009558282792568207, + 0.006109502632170916, + 0.003621782409027219, + 0.007800948806107044, + 0.005671381484717131, + 0.004564744886010885, + 0.0033822094555944204, + 0.004191723186522722, + 0.004295067396014929, + 0.007683059200644493, + 0.003808708628639579, + 0.0056359837763011456, + 0.0029452741146087646, + 0.0033738326746970415, + 0.002503354102373123, + 0.00547837046906352, + 0.004946727305650711, + 0.002646551700308919, + 0.0022375129628926516, + 0.002660301048308611, + 0.006844901945441961, + 0.01669856533408165, + 0.006870557554066181, + 0.008192542940378189, + 0.010438192635774612, + 0.008056986145675182, + 0.009654825553297997, + 0.00676257349550724, + 0.012653451412916183, + 0.009608849883079529, + 0.0037106506060808897, + 0.004057801328599453, + 0.01038418896496296, + 0.006149611435830593, + 0.01016138680279255, + 0.005338274873793125, + 0.06928477436304092, + 0.05604613199830055, + 0.030672350898385048, + 0.03403310477733612, + 0.057586513459682465, + 0.08870918303728104, + 0.05611474812030792, + 0.05699452757835388, + 0.006195522379130125, + 0.004283152986317873, + 0.010263246484100819, + 0.0032887423876672983, + 0.011072192341089249, + 0.0037061276379972696, + 0.002607970265671611, + 0.0035650625359266996, + 0.012665617279708385, + 0.016427330672740936, + 0.015128607861697674, + 0.012810950167477131, + 0.007469596806913614, + 0.008055970072746277, + 0.006374097429215908, + 0.007354301866143942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.0336914137005806 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a B-17 bomber entering the frame from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft's fuselage silver with visible rivets and panel seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the four propellers spinning rapidly, appearing as nearly transparent dynamic discs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rough asphalt runway with dust being stirred up by the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot where the background and horizon are blurred due to high-speed movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197885990142822 + ], + "opposite_logprob_tokens": [ + -4.264197826385498 + ], + "logprob_reward": -0.014197885990142822, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197885990142822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04994682967662811, + 0.060174643993377686, + 0.049065910279750824, + 0.06013702228665352, + 0.023646250367164612, + 0.17122162878513336, + 0.11814414709806442, + 0.23273755609989166, + 0.17677907645702362, + 0.05814684182405472 + ], + "credit_map": [ + 0.004626233130693436, + 0.007325443904846907, + 0.006452090572565794, + 0.006135343108326197, + 0.003983224742114544, + 0.004192241933196783, + 0.006496630143374205, + 0.01073562353849411, + 0.005469616502523422, + 0.005365756340324879, + 0.00403321348130703, + 0.006518236827105284, + 0.0034241098910570145, + 0.008368545211851597, + 0.009205284528434277, + 0.01778988540172577, + 0.005653622560203075, + 0.007290640380233526, + 0.004335591569542885, + 0.011980119161307812, + 0.007796624209731817, + 0.004745964892208576, + 0.003562477184459567, + 0.003700871951878071, + 0.008172365836799145, + 0.00516137108206749, + 0.005499186925590038, + 0.006087328772991896, + 0.006664445158094168, + 0.011444240808486938, + 0.004222060553729534, + 0.01288602314889431, + 0.0037991940043866634, + 0.003599548013880849, + 0.001478122896514833, + 0.0032383715733885765, + 0.003993915859609842, + 0.0027630787808448076, + 0.0022559177596122026, + 0.0025181022938340902, + 0.013518039137125015, + 0.06383004039525986, + 0.03740323707461357, + 0.00977209210395813, + 0.012528929859399796, + 0.008649112656712532, + 0.015164685435593128, + 0.010355506092309952, + 0.014890474267303944, + 0.01706732250750065, + 0.014984372071921825, + 0.015583089552819729, + 0.015637218952178955, + 0.01041218638420105, + 0.015617425553500652, + 0.013952046632766724, + 0.034944985061883926, + 0.02884797565639019, + 0.016899876296520233, + 0.018650740385055542, + 0.029189910739660263, + 0.04468497633934021, + 0.029931262135505676, + 0.029587823897600174, + 0.020744210109114647, + 0.016562720760703087, + 0.04156045988202095, + 0.01073397696018219, + 0.05473360791802406, + 0.013043327257037163, + 0.008094795979559422, + 0.011305985040962696, + 0.008558114990592003, + 0.008906970731914043, + 0.010601876303553581, + 0.010061387903988361, + 0.006407819222658873, + 0.004662892781198025, + 0.003987407777458429, + 0.004960374906659126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.014197885990142822 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_0/index_tensor([4029, 4029])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_0/index_tensor([4029, 4029])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c0f6e2c8b3594abcacaf95374dde353455790fa7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_0/index_tensor([4029, 4029])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle, long-shot lens, rotating slowly, captures a young Middle Eastern girl, her thick, dark curls tied into a tight ponytail, practicing alone in a vast indoor gymnastics hall. She wears a shimmering, blue-blue, long-sleeved gymnastics leotard, a striking contrast to the wide, padded royal blue mat beneath. She is centered in the frame, moving with intense concentration along a low, tan balance beam, her bare feet pointing, each deliberate, measured step. Suddenly, her focus wanders, and her body teeters precariously to one side. For a brief, tense moment, her arms, previously held out for balance, flutter slightly. Then, through an apparent effort of will, she tightens her core, firmly stretches her arms straight out to her sides like wings, and successfully regains her stability. Her gaze refocuses on the end of the beam, her expression shifting from a fleeting panic to one of unwavering determination. Bright, even fluorescent lighting from the high ceilings illuminates the space, casting soft shadows that emphasize the quiet, focused atmosphere of individual training. Other gymnastic equipment, such as the uneven bars and vault, are visible but out of focus in the background.", + "seed": 10540, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_0/index_4029_seed_10540_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3097637298516929, + "vlm_logprob": -0.3097637298516929, + "vlm_yes_no_margin": 3.000000011920929, + "vlm_token_credit": -0.3097637298516929, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4696524441242218 + ], + "opposite_logprob_tokens": [ + -1.4696524143218994 + ], + "logprob_reward": -0.4696524441242218, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.4696524441242218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20880413055419922, + 0.09136223047971725, + 0.032856013625860214, + 0.13159123063087463, + 0.035672351717948914, + 0.18443642556667328, + 0.09433536976575851, + 0.04655816778540611, + 0.1258973777294159, + 0.048486676067113876 + ], + "credit_map": [ + 0.04198634251952171, + 0.034318409860134125, + 0.01208515651524067, + 0.017749758437275887, + 0.03766714781522751, + 0.030038192868232727, + 0.02000141143798828, + 0.014957706443965435, + 0.007764393463730812, + 0.014039275236427784, + 0.00964285433292389, + 0.01636025495827198, + 0.016065340489149094, + 0.007939739152789116, + 0.009274971671402454, + 0.010275394655764103, + 0.005544164218008518, + 0.004186607897281647, + 0.005000284872949123, + 0.004442430101335049, + 0.0046363575384020805, + 0.003650364466011524, + 0.0020502775441855192, + 0.0033455248922109604, + 0.008697140961885452, + 0.011513463221490383, + 0.04110671952366829, + 0.008792747743427753, + 0.009736679494380951, + 0.00806998461484909, + 0.007064624689519405, + 0.03660986199975014, + 0.0032172752544283867, + 0.0034842288587242365, + 0.005265793297439814, + 0.002691235626116395, + 0.004490664694458246, + 0.007904763333499432, + 0.002378007397055626, + 0.006240379065275192, + 0.007575021591037512, + 0.008999241515994072, + 0.0713246539235115, + 0.00847061350941658, + 0.006120611447840929, + 0.007819090038537979, + 0.005712556652724743, + 0.06841465085744858, + 0.009138526394963264, + 0.008148430846631527, + 0.012264384888112545, + 0.01279141753911972, + 0.010438528843224049, + 0.012645797803997993, + 0.01436249166727066, + 0.014545794576406479, + 0.006086389068514109, + 0.004741804674267769, + 0.008157354779541492, + 0.007281940896064043, + 0.004082075320184231, + 0.004312233533710241, + 0.0048347716219723225, + 0.007061596028506756, + 0.013064305298030376, + 0.020139791071414948, + 0.015249671414494514, + 0.010510720312595367, + 0.010060050524771214, + 0.03363803029060364, + 0.00787907000631094, + 0.015355747193098068, + 0.005660977214574814, + 0.005778079386800528, + 0.0054344795644283295, + 0.0043078139424324036, + 0.006002392619848251, + 0.005607753060758114, + 0.00910248700529337, + 0.0065926918759942055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.4696524441242218 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0538957118988037 + ], + "opposite_logprob_tokens": [ + -0.4288957715034485 + ], + "logprob_reward": -1.0538957118988037, + "yes_no_margin_reward": -0.6249999403953552, + "token_credit_reward": -1.0538957118988037, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06024463102221489, + 0.04272225871682167, + 0.040321819484233856, + 0.4224282503128052, + 0.026344265788793564, + 0.14202538132667542, + 0.05997079983353615, + 0.10692380368709564, + 0.07204917073249817, + 0.026969581842422485 + ], + "credit_map": [ + 0.009492834098637104, + 0.009374178946018219, + 0.005154019221663475, + 0.011124609038233757, + 0.007271211128681898, + 0.007387220393866301, + 0.005960893351584673, + 0.0044796643778681755, + 0.0037030356470495462, + 0.006076570134609938, + 0.003405107418075204, + 0.005590092856436968, + 0.010885900817811489, + 0.004217874258756638, + 0.0045212311670184135, + 0.004322446882724762, + 0.0051576499827206135, + 0.005930126179009676, + 0.004608054179698229, + 0.004352045711129904, + 0.008753492496907711, + 0.004477412905544043, + 0.0022831656970083714, + 0.004759869538247585, + 0.025340726599097252, + 0.019217856228351593, + 0.16918152570724487, + 0.0211319699883461, + 0.011221306398510933, + 0.014598560519516468, + 0.01551810558885336, + 0.1462181806564331, + 0.0023130988702178, + 0.002514597261324525, + 0.0044172084890306, + 0.0023711093235760927, + 0.002970612607896328, + 0.005074730608612299, + 0.0025051322299987078, + 0.0041777752339839935, + 0.005697168409824371, + 0.0057608396746218204, + 0.05624263361096382, + 0.0061638206243515015, + 0.0036626639775931835, + 0.006103480234742165, + 0.003729959251359105, + 0.05466480180621147, + 0.005607049446552992, + 0.004822337534278631, + 0.008173026144504547, + 0.009462233632802963, + 0.007047703023999929, + 0.006491628009825945, + 0.009612828493118286, + 0.008753989823162556, + 0.008611826226115227, + 0.012160902842879295, + 0.015883924439549446, + 0.018088068813085556, + 0.008175215683877468, + 0.015662170946598053, + 0.012334431521594524, + 0.01600726507604122, + 0.007088083308190107, + 0.008832842111587524, + 0.009986214339733124, + 0.005506303161382675, + 0.006137496326118708, + 0.018321305513381958, + 0.006935655605047941, + 0.009241272695362568, + 0.0038831045385450125, + 0.002585404785349965, + 0.0031673831399530172, + 0.002294710138812661, + 0.0034727705642580986, + 0.005228117108345032, + 0.003509907051920891, + 0.0028281821869313717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -1.0538957118988037 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004103455226868391 + ], + "opposite_logprob_tokens": [ + -5.504103660583496 + ], + "logprob_reward": -0.004103455226868391, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004103455226868391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13096347451210022, + 0.07805858552455902, + 0.029436416923999786, + 0.1963450014591217, + 0.07085064798593521, + 0.14311052858829498, + 0.07509589195251465, + 0.08081240952014923, + 0.05848492681980133, + 0.13684207201004028 + ], + "credit_map": [ + 0.015754584223031998, + 0.021311335265636444, + 0.011086225509643555, + 0.02279353141784668, + 0.01938013546168804, + 0.017821097746491432, + 0.01352482009679079, + 0.009291752241551876, + 0.0073093753308057785, + 0.011081166565418243, + 0.007360345683991909, + 0.00874306820333004, + 0.013787292875349522, + 0.006774761248379946, + 0.012931127101182938, + 0.010071447119116783, + 0.0035719904117286205, + 0.004281382542103529, + 0.0029473244212567806, + 0.0025616695638746023, + 0.00543858390301466, + 0.004097190219908953, + 0.0023402029182761908, + 0.004198072012513876, + 0.012415832839906216, + 0.012695224024355412, + 0.06627912819385529, + 0.008411840535700321, + 0.01092684455215931, + 0.01652272418141365, + 0.007657050620764494, + 0.061436355113983154, + 0.006172242108732462, + 0.007300244644284248, + 0.012933422811329365, + 0.00920044630765915, + 0.007863400503993034, + 0.007894712500274181, + 0.006985150743275881, + 0.012501025572419167, + 0.005905358586460352, + 0.003934402950108051, + 0.05545227229595184, + 0.006129876710474491, + 0.0045955488458275795, + 0.010127567686140537, + 0.004013499710708857, + 0.05295200273394585, + 0.006646720692515373, + 0.005281282588839531, + 0.013129997998476028, + 0.013683543540537357, + 0.006093414966017008, + 0.007730227895081043, + 0.0076630087569355965, + 0.014867693185806274, + 0.0072916350327432156, + 0.006354783196002245, + 0.010789144784212112, + 0.015605686232447624, + 0.006193238776177168, + 0.008137883618474007, + 0.014118470251560211, + 0.01232156716287136, + 0.004804131109267473, + 0.003590912092477083, + 0.01416770089417696, + 0.005335397087037563, + 0.0038555345963686705, + 0.004745101556181908, + 0.006273283157497644, + 0.015712862834334373, + 0.016359008848667145, + 0.019912108778953552, + 0.013264062814414501, + 0.011173543520271778, + 0.03318525850772858, + 0.018111424520611763, + 0.014853868633508682, + 0.009982808493077755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 9 + ], + "selected_reward": -0.004103455226868391 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638394996523857 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638394996523857, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638394996523857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14833690226078033, + 0.07298499345779419, + 0.04258012771606445, + 0.21940700709819794, + 0.034048713743686676, + 0.11263096332550049, + 0.17701661586761475, + 0.02352818101644516, + 0.11345531046390533, + 0.056011222302913666 + ], + "credit_map": [ + 0.020155975595116615, + 0.019247734919190407, + 0.014254716224968433, + 0.028300831094384193, + 0.02656252682209015, + 0.017786579206585884, + 0.011921906843781471, + 0.010106632485985756, + 0.007923460565507412, + 0.01017998717725277, + 0.01019185222685337, + 0.007552752271294594, + 0.013815589249134064, + 0.006874803453683853, + 0.009314185008406639, + 0.0071323630400002, + 0.006739750038832426, + 0.007503050845116377, + 0.005610039923340082, + 0.003777362871915102, + 0.00666952645406127, + 0.00347004528157413, + 0.003338388167321682, + 0.005471964832395315, + 0.012960094027221203, + 0.010165977291762829, + 0.0861746072769165, + 0.010270893573760986, + 0.0067734564654529095, + 0.008016053587198257, + 0.008590638637542725, + 0.07645528018474579, + 0.0032570536714047194, + 0.004181445576250553, + 0.005514976102858782, + 0.004670622758567333, + 0.005997910164296627, + 0.0038232298102229834, + 0.0026668738573789597, + 0.003936603665351868, + 0.005596188362687826, + 0.010330937802791595, + 0.030393360182642937, + 0.009526092559099197, + 0.011582049541175365, + 0.007632220163941383, + 0.008996665477752686, + 0.028573455289006233, + 0.021619020029902458, + 0.015540686435997486, + 0.02055087871849537, + 0.022260630503296852, + 0.015977099537849426, + 0.03492593765258789, + 0.021326584741473198, + 0.024815758690238, + 0.001976641360670328, + 0.0016704732552170753, + 0.003478499362245202, + 0.0034121214412152767, + 0.002886212430894375, + 0.0031731431372463703, + 0.00344968494027853, + 0.003481405321508646, + 0.013020980171859264, + 0.017664259299635887, + 0.012382046319544315, + 0.01154347974807024, + 0.011480377987027168, + 0.02435322105884552, + 0.011666065081954002, + 0.011344884522259235, + 0.006962734740227461, + 0.005408417899161577, + 0.006127266213297844, + 0.005486087407916784, + 0.007187701761722565, + 0.008580131456255913, + 0.009374025277793407, + 0.006884860340505838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.008638394996523857 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012528643012046814 + ], + "opposite_logprob_tokens": [ + -4.387528419494629 + ], + "logprob_reward": -0.012528643012046814, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012528643012046814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05991750210523605, + 0.07796133309602737, + 0.052079424262046814, + 0.06691694259643555, + 0.06624078005552292, + 0.20435182750225067, + 0.2054644674062729, + 0.18035496771335602, + 0.07228221744298935, + 0.01443052850663662 + ], + "credit_map": [ + 0.00736894691362977, + 0.00816368218511343, + 0.005984144750982523, + 0.009049743413925171, + 0.010667952708899975, + 0.008088555186986923, + 0.005403998773545027, + 0.005190474912524223, + 0.008228840306401253, + 0.011250843293964863, + 0.011237913742661476, + 0.016582144424319267, + 0.008804495446383953, + 0.006151763256639242, + 0.007139374502003193, + 0.008565956726670265, + 0.007051568478345871, + 0.006629002746194601, + 0.007706047967076302, + 0.004938091617077589, + 0.006134997587651014, + 0.00764986639842391, + 0.0039250534027814865, + 0.008044795133173466, + 0.004607012029737234, + 0.006917282938957214, + 0.020723490044474602, + 0.00385288055986166, + 0.004266646690666676, + 0.006374987307935953, + 0.004273667465895414, + 0.015900976955890656, + 0.00964714027941227, + 0.005061120726168156, + 0.007383983116596937, + 0.0055250003933906555, + 0.009530138224363327, + 0.01692080870270729, + 0.0048539526760578156, + 0.007318636402487755, + 0.00906150322407484, + 0.005373743828386068, + 0.07955486327409744, + 0.008334683254361153, + 0.007080413866788149, + 0.015387129969894886, + 0.004672235809266567, + 0.07488725334405899, + 0.012374856509268284, + 0.015669172629714012, + 0.034481920301914215, + 0.05334550514817238, + 0.014559282921254635, + 0.018486909568309784, + 0.02098618820309639, + 0.035560622811317444, + 0.015872810035943985, + 0.024275045841932297, + 0.030756385996937752, + 0.0204856526106596, + 0.013972737826406956, + 0.03514532372355461, + 0.012484930455684662, + 0.027362078428268433, + 0.009460496716201305, + 0.008717609569430351, + 0.007367648649960756, + 0.004653230309486389, + 0.005998083855956793, + 0.02285352349281311, + 0.005098567809909582, + 0.008133058436214924, + 0.0024916394613683224, + 0.0022240877151489258, + 0.0015065281186252832, + 0.0010601310059428215, + 0.002672336297109723, + 0.0014922564150765538, + 0.0013973322929814458, + 0.0015862170839682221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.012528643012046814 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, long-shot lens, rotating slowly, captures a young Middle Eastern girl, her thick, dark curls tied into a tight ponytail, practicing alone in a vast indoor gymnastics hall. She wears a shimmering, blue-blue, long-sleeved gymnastics leotard, a striking contrast to the wide, padded royal blue mat beneath. She is centered in the frame, moving with intense concentration along a low, tan balance beam, her bare feet pointing, each deliberate, measured step. Suddenly, her focus wanders, and her body teeters precariously to one side. For a brief, tense moment, her arms, previously held out for balance, flutter slightly. Then, through an apparent effort of will, she tightens her core, firmly stretches her arms straight out to her sides like wings, and successfully regains her stability. Her gaze refocuses on the end of the beam, her expression shifting from a fleeting panic to one of unwavering determination. Bright, even fluorescent lighting from the high ceilings illuminates the space, casting soft shadows that emphasize the quiet, focused atmosphere of individual training. Other gymnastic equipment, such as the uneven bars and vault, are visible but out of focus in the background.", + "seed": 10541, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_0/index_4029_seed_10541_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.28109973995015025, + "vlm_logprob": -0.28109973995015025, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.28109973995015025, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38775378465652466 + ], + "opposite_logprob_tokens": [ + -2.01275372505188 + ], + "logprob_reward": -0.38775378465652466, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.38775378465652466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22713202238082886, + 0.07507526874542236, + 0.1297481805086136, + 0.1964331418275833, + 0.05723349750041962, + 0.10811313986778259, + 0.030807755887508392, + 0.05472154542803764, + 0.0482727512717247, + 0.07246273756027222 + ], + "credit_map": [ + 0.047075241804122925, + 0.046320222318172455, + 0.024392137303948402, + 0.021051231771707535, + 0.018866267055273056, + 0.031304776668548584, + 0.021460117772221565, + 0.016662009060382843, + 0.012210226617753506, + 0.015190030448138714, + 0.008959650993347168, + 0.010297645814716816, + 0.008817557245492935, + 0.004974719136953354, + 0.01000360306352377, + 0.0046218326315283775, + 0.011912836693227291, + 0.009302333928644657, + 0.02306845597922802, + 0.03967078775167465, + 0.0073546841740608215, + 0.013174088671803474, + 0.0117160240188241, + 0.013548956252634525, + 0.00920000858604908, + 0.008732173591852188, + 0.06292959302663803, + 0.0685826912522316, + 0.008123118430376053, + 0.012755396775901318, + 0.011288252659142017, + 0.014821893535554409, + 0.007078183349221945, + 0.007324695121496916, + 0.010150604881346226, + 0.00904554221779108, + 0.004086457658559084, + 0.009139204397797585, + 0.004986507352441549, + 0.005422300659120083, + 0.014947149902582169, + 0.010564099065959454, + 0.020649386569857597, + 0.009758345782756805, + 0.006436607334762812, + 0.016821956261992455, + 0.008324720896780491, + 0.020610865205526352, + 0.003023162018507719, + 0.004090235102921724, + 0.0048814681358635426, + 0.002857627347111702, + 0.0042912401258945465, + 0.004146374762058258, + 0.003432267578318715, + 0.0040853810496628284, + 0.00656971987336874, + 0.006183954421430826, + 0.006734019611030817, + 0.005116905085742474, + 0.005425286944955587, + 0.005296947900205851, + 0.008056435734033585, + 0.011338273994624615, + 0.006470969878137112, + 0.005577229429036379, + 0.005244653206318617, + 0.0032870876602828503, + 0.005796799436211586, + 0.011174106039106846, + 0.003918217495083809, + 0.0068036820739507675, + 0.019524171948432922, + 0.007983473129570484, + 0.004857523366808891, + 0.005629335064440966, + 0.006544371135532856, + 0.011372710578143597, + 0.007661796640604734, + 0.008889350108802319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.38775378465652466 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742910265922546 + ], + "opposite_logprob_tokens": [ + -0.47429102659225464 + ], + "logprob_reward": -0.9742910265922546, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742910265922546, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08930964022874832, + 0.10884423553943634, + 0.09971221536397934, + 0.11585044115781784, + 0.03830545023083687, + 0.14654481410980225, + 0.03928603231906891, + 0.03859822452068329, + 0.07675540447235107, + 0.24679359793663025 + ], + "credit_map": [ + 0.01215535681694746, + 0.010133652947843075, + 0.011576443910598755, + 0.01050147321075201, + 0.012771373614668846, + 0.01024431549012661, + 0.014600521884858608, + 0.007326500490307808, + 0.012912208214402199, + 0.015401170589029789, + 0.016410617157816887, + 0.015902385115623474, + 0.012080051936209202, + 0.011851751245558262, + 0.014191903173923492, + 0.01009414903819561, + 0.00849583838135004, + 0.010096210055053234, + 0.014969526790082455, + 0.023318644613027573, + 0.011049987748265266, + 0.009862598963081837, + 0.01056670118123293, + 0.011352702975273132, + 0.014842458069324493, + 0.01598508283495903, + 0.01641991175711155, + 0.014929561875760555, + 0.007073500659316778, + 0.02102341316640377, + 0.011082417331635952, + 0.01449409406632185, + 0.005120372865349054, + 0.0031431959941983223, + 0.00597802409902215, + 0.005185815040022135, + 0.0033799249213188887, + 0.006063286680728197, + 0.004858615808188915, + 0.0045762150548398495, + 0.014920379966497421, + 0.023276912048459053, + 0.014139141887426376, + 0.021237559616565704, + 0.023058896884322166, + 0.01766469143331051, + 0.015354355797171593, + 0.016892872750759125, + 0.004325731657445431, + 0.0038715393748134375, + 0.004395870491862297, + 0.003364376025274396, + 0.005090913735330105, + 0.009719368070363998, + 0.003381638787686825, + 0.005136595573276281, + 0.005971911363303661, + 0.004198572598397732, + 0.0064742304384708405, + 0.0035140030086040497, + 0.004377270117402077, + 0.006321659777313471, + 0.003034752793610096, + 0.0047058239579200745, + 0.009362333454191685, + 0.009528526104986668, + 0.009717050939798355, + 0.0060982475988566875, + 0.01147561427205801, + 0.010875729843974113, + 0.009223470464348793, + 0.010474427603185177, + 0.06153343617916107, + 0.030326170846819878, + 0.01787562109529972, + 0.027485441416502, + 0.022528570145368576, + 0.03395078331232071, + 0.02372582256793976, + 0.02936776541173458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 3 + ], + "selected_reward": -0.9742910265922546 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007640306372195482 + ], + "opposite_logprob_tokens": [ + -4.882640361785889 + ], + "logprob_reward": -0.007640306372195482, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007640306372195482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15097607672214508, + 0.06580550223588943, + 0.05937867611646652, + 0.10871893167495728, + 0.18656101822853088, + 0.050220590084791183, + 0.17059141397476196, + 0.050779812037944794, + 0.045241110026836395, + 0.11172685027122498 + ], + "credit_map": [ + 0.012916447594761848, + 0.01564924046397209, + 0.020487401634454727, + 0.018664881587028503, + 0.02924063615500927, + 0.025923574343323708, + 0.019242893904447556, + 0.00885098148137331, + 0.007433472666889429, + 0.010493303649127483, + 0.009123188443481922, + 0.008842012844979763, + 0.008979539386928082, + 0.0051880753599107265, + 0.011724909767508507, + 0.004021008498966694, + 0.006155593786388636, + 0.006124373059719801, + 0.00803635735064745, + 0.00965187232941389, + 0.007754553109407425, + 0.005836728028953075, + 0.008242028765380383, + 0.007577171083539724, + 0.006316223181784153, + 0.004498308524489403, + 0.030844174325466156, + 0.03478124737739563, + 0.006336296442896128, + 0.006701855454593897, + 0.007823407649993896, + 0.011417414993047714, + 0.01873026229441166, + 0.02200886979699135, + 0.03819402679800987, + 0.032558441162109375, + 0.016393963247537613, + 0.018015403300523758, + 0.019758742302656174, + 0.020901305601000786, + 0.0050320811569690704, + 0.007783497683703899, + 0.005227380897849798, + 0.007012243382632732, + 0.005231878254562616, + 0.005698706489056349, + 0.007924066856503487, + 0.006310734432190657, + 0.017144424840807915, + 0.023374376818537712, + 0.015210796147584915, + 0.013200607150793076, + 0.04204271733760834, + 0.02458980306982994, + 0.013358636759221554, + 0.021670041605830193, + 0.008942225016653538, + 0.004761934280395508, + 0.007633569184690714, + 0.0032690553925931454, + 0.0049531604163348675, + 0.0060789454728364944, + 0.006083883345127106, + 0.009057041257619858, + 0.004593031480908394, + 0.00763181084766984, + 0.004637131933122873, + 0.004576656501740217, + 0.005197704304009676, + 0.005048365797847509, + 0.006127904634922743, + 0.007428502198308706, + 0.02720632590353489, + 0.0133061483502388, + 0.01236779149621725, + 0.011383208446204662, + 0.010616456158459187, + 0.01596049964427948, + 0.008510824292898178, + 0.012375598773360252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.007640306372195482 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254053320735693 + ], + "opposite_logprob_tokens": [ + -4.387540340423584 + ], + "logprob_reward": -0.01254053320735693, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254053320735693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15592795610427856, + 0.07187508046627045, + 0.11108596622943878, + 0.21547187864780426, + 0.07369216531515121, + 0.12505996227264404, + 0.09127917885780334, + 0.04393557459115982, + 0.05324913561344147, + 0.05842307209968567 + ], + "credit_map": [ + 0.01787208765745163, + 0.021122131496667862, + 0.01747499592602253, + 0.02031838148832321, + 0.0254831463098526, + 0.019319070503115654, + 0.02125406265258789, + 0.013084071688354015, + 0.011418810114264488, + 0.014445409178733826, + 0.009280616417527199, + 0.01094487588852644, + 0.0071081798523664474, + 0.0073113879188895226, + 0.006199383642524481, + 0.005166415590792894, + 0.014672691933810711, + 0.012716636061668396, + 0.012638140469789505, + 0.01446792483329773, + 0.008267722092568874, + 0.0185688529163599, + 0.008356970734894276, + 0.02139703370630741, + 0.017709989100694656, + 0.016147499904036522, + 0.056695740669965744, + 0.059327587485313416, + 0.01224996242672205, + 0.021097391843795776, + 0.011298212222754955, + 0.020945487543940544, + 0.007237133104354143, + 0.008028090000152588, + 0.00903896614909172, + 0.010371172800660133, + 0.010035080835223198, + 0.007782077416777611, + 0.013368114829063416, + 0.007831532508134842, + 0.013341441750526428, + 0.015043810941278934, + 0.018044408410787582, + 0.01841883361339569, + 0.009262490086257458, + 0.013717563822865486, + 0.015182821080088615, + 0.022048603743314743, + 0.008158469572663307, + 0.013040234334766865, + 0.008519946597516537, + 0.008035218343138695, + 0.023958425968885422, + 0.010951017960906029, + 0.008131707087159157, + 0.010484160855412483, + 0.005081703886389732, + 0.004481958691030741, + 0.004320652224123478, + 0.0029588520992547274, + 0.006445606239140034, + 0.011511082760989666, + 0.003968562465161085, + 0.00516715832054615, + 0.004849398974329233, + 0.003047672566026449, + 0.0050364527851343155, + 0.0045750802382826805, + 0.005388468969613314, + 0.007183337118476629, + 0.01235943753272295, + 0.010809287428855896, + 0.009786007925868034, + 0.006784284021705389, + 0.007013220340013504, + 0.009924067184329033, + 0.005352397449314594, + 0.00653088977560401, + 0.006332775577902794, + 0.006699427962303162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.01254053320735693 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023273048922419548 + ], + "opposite_logprob_tokens": [ + -3.77327299118042 + ], + "logprob_reward": -0.023273048922419548, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023273048922419548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10630104690790176, + 0.04659242182970047, + 0.1253194659948349, + 0.055624641478061676, + 0.04942701384425163, + 0.04891160875558853, + 0.08402982354164124, + 0.35179558396339417, + 0.06337817013263702, + 0.06862027943134308 + ], + "credit_map": [ + 0.01803213730454445, + 0.013315646909177303, + 0.014341430738568306, + 0.01696888543665409, + 0.008865230716764927, + 0.015905063599348068, + 0.008814774453639984, + 0.010057884268462658, + 0.006526084151118994, + 0.00797019898891449, + 0.0060012973845005035, + 0.007561933249235153, + 0.00601145951077342, + 0.003330925479531288, + 0.005763924214988947, + 0.0034265995491296053, + 0.012360489927232265, + 0.010223843157291412, + 0.02184101939201355, + 0.03212105110287666, + 0.00853397510945797, + 0.014003842137753963, + 0.010773105546832085, + 0.015462144277989864, + 0.00806211307644844, + 0.008539700880646706, + 0.007831095717847347, + 0.008513296023011208, + 0.0037975178565829992, + 0.008717107586562634, + 0.0028559183701872826, + 0.007307892665266991, + 0.006847443524748087, + 0.008606575429439545, + 0.006471754517406225, + 0.006254310254007578, + 0.003650296712294221, + 0.006892319303005934, + 0.0048707411624491215, + 0.005833571311086416, + 0.005246646702289581, + 0.00853397324681282, + 0.006386925932019949, + 0.006102266255766153, + 0.006075710523873568, + 0.005735169164836407, + 0.004735318012535572, + 0.006095601245760918, + 0.006988427601754665, + 0.00968900416046381, + 0.009365522302687168, + 0.007518324069678783, + 0.01661747507750988, + 0.017045998945832253, + 0.006029315292835236, + 0.010775752365589142, + 0.04289218783378601, + 0.02297803945839405, + 0.06467255204916, + 0.02176465094089508, + 0.029846234247088432, + 0.055695779621601105, + 0.04122903570532799, + 0.07271713018417358, + 0.006477687042206526, + 0.006281109992414713, + 0.00543719157576561, + 0.005028316751122475, + 0.006560247857123613, + 0.016123877838253975, + 0.007150208577513695, + 0.01031953003257513, + 0.015874208882451057, + 0.008940340019762516, + 0.007361024152487516, + 0.008239059709012508, + 0.006153267342597246, + 0.010295669548213482, + 0.00446622958406806, + 0.007290479727089405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.023273048922419548 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, long-shot lens, rotating slowly, captures a young Middle Eastern girl, her thick, dark curls tied into a tight ponytail, practicing alone in a vast indoor gymnastics hall. She wears a shimmering, blue-blue, long-sleeved gymnastics leotard, a striking contrast to the wide, padded royal blue mat beneath. She is centered in the frame, moving with intense concentration along a low, tan balance beam, her bare feet pointing, each deliberate, measured step. Suddenly, her focus wanders, and her body teeters precariously to one side. For a brief, tense moment, her arms, previously held out for balance, flutter slightly. Then, through an apparent effort of will, she tightens her core, firmly stretches her arms straight out to her sides like wings, and successfully regains her stability. Her gaze refocuses on the end of the beam, her expression shifting from a fleeting panic to one of unwavering determination. Bright, even fluorescent lighting from the high ceilings illuminates the space, casting soft shadows that emphasize the quiet, focused atmosphere of individual training. Other gymnastic equipment, such as the uneven bars and vault, are visible but out of focus in the background.", + "seed": 10542, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_0/index_4029_seed_10542_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.176565015129745, + "vlm_logprob": -0.176565015129745, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.176565015129745, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22284844517707825 + ], + "opposite_logprob_tokens": [ + -2.222848415374756 + ], + "logprob_reward": -0.22284844517707825, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.22284844517707825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2578272819519043, + 0.055032823234796524, + 0.08869616687297821, + 0.04252093285322189, + 0.05020090192556381, + 0.07366219907999039, + 0.0679221898317337, + 0.05711955949664116, + 0.22189843654632568, + 0.08511946350336075 + ], + "credit_map": [ + 0.021064799278974533, + 0.016149751842021942, + 0.06973502784967422, + 0.026526091620326042, + 0.013975782319903374, + 0.0687640830874443, + 0.014380124397575855, + 0.02723163552582264, + 0.007947288453578949, + 0.010235964320600033, + 0.00627563614398241, + 0.00885486789047718, + 0.0038900941144675016, + 0.008346562273800373, + 0.00433092238381505, + 0.00515149487182498, + 0.00984289962798357, + 0.007578630931675434, + 0.006173454225063324, + 0.009458239190280437, + 0.005421000998467207, + 0.02300213649868965, + 0.007783280685544014, + 0.019436528906226158, + 0.004403321538120508, + 0.006374166812747717, + 0.0053700353018939495, + 0.006169591564685106, + 0.00619079265743494, + 0.0033400985412299633, + 0.005257051903754473, + 0.00541587732732296, + 0.006372448988258839, + 0.006874525919556618, + 0.008078934624791145, + 0.005164183676242828, + 0.004991951864212751, + 0.0063982317224144936, + 0.005682629998773336, + 0.006637995131313801, + 0.007417893037199974, + 0.009856362827122211, + 0.009884358383715153, + 0.00709396880120039, + 0.006978033576160669, + 0.01314274501055479, + 0.006722559686750174, + 0.01256628893315792, + 0.004901286214590073, + 0.009515482001006603, + 0.011967530474066734, + 0.007590801455080509, + 0.006700174883008003, + 0.005089824087917805, + 0.008071786724030972, + 0.014085305854678154, + 0.0048281364142894745, + 0.005492434371262789, + 0.017444966360926628, + 0.006316968705505133, + 0.005179842468351126, + 0.009472041390836239, + 0.0038643295411020517, + 0.004520844668149948, + 0.023173760622739792, + 0.01563345454633236, + 0.014876139350235462, + 0.011007924564182758, + 0.06832122057676315, + 0.057619910687208176, + 0.011818945407867432, + 0.01944710500538349, + 0.003902627620846033, + 0.0043471031822264194, + 0.004995190538465977, + 0.005685134325176477, + 0.006517503410577774, + 0.024839524179697037, + 0.027784116566181183, + 0.007048268336802721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.22284844517707825 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327879428863525 + ], + "opposite_logprob_tokens": [ + -0.7577879428863525 + ], + "logprob_reward": -0.6327879428863525, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327879428863525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08546812832355499, + 0.046102169901132584, + 0.05970621854066849, + 0.07248123735189438, + 0.07728096097707748, + 0.19360747933387756, + 0.022005876526236534, + 0.03994598984718323, + 0.030667470768094063, + 0.3727344572544098 + ], + "credit_map": [ + 0.00926771480590105, + 0.008670026436448097, + 0.013181431218981743, + 0.014316207729279995, + 0.011511076241731644, + 0.008511905558407307, + 0.009099476970732212, + 0.01091028656810522, + 0.004230074118822813, + 0.005228055641055107, + 0.007248743437230587, + 0.00703436229377985, + 0.005783282220363617, + 0.0035783788189291954, + 0.006545453332364559, + 0.006453821901232004, + 0.005884226877242327, + 0.005069569684565067, + 0.007690028287470341, + 0.009473049081861973, + 0.002794128842651844, + 0.01600041426718235, + 0.004098006524145603, + 0.008696791715919971, + 0.007745513692498207, + 0.008508149534463882, + 0.01093386858701706, + 0.010219656862318516, + 0.007541336119174957, + 0.007764802780002356, + 0.010612892918288708, + 0.009155012667179108, + 0.015470168553292751, + 0.0065737818367779255, + 0.01287265494465828, + 0.005590927321463823, + 0.005381630267947912, + 0.012910185381770134, + 0.007140764035284519, + 0.011340848170220852, + 0.02693275734782219, + 0.03517260029911995, + 0.019382411614060402, + 0.022302092984318733, + 0.021041179075837135, + 0.029376612976193428, + 0.020072543993592262, + 0.019327281042933464, + 0.0018011488718912005, + 0.0028934325091540813, + 0.004505244083702564, + 0.0022840367164462805, + 0.0021757062058895826, + 0.0021208240650594234, + 0.0032320672180503607, + 0.002993418602272868, + 0.0035398497711867094, + 0.005802431143820286, + 0.006685457658022642, + 0.003914035391062498, + 0.0036152510438114405, + 0.008908058516681194, + 0.00361446850001812, + 0.0038664366584271193, + 0.003037805203348398, + 0.0036089252680540085, + 0.0028931472916156054, + 0.002831901889294386, + 0.0052359625697135925, + 0.0043198480270802975, + 0.0022355769760906696, + 0.0065043019130826, + 0.01649291254580021, + 0.006485044490545988, + 0.01932300627231598, + 0.014802364632487297, + 0.03502826392650604, + 0.12857545912265778, + 0.12463943660259247, + 0.027387987822294235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.6327879428863525 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790733456611633 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.009790733456611633, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790733456611633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14699141681194305, + 0.0845857709646225, + 0.07981333881616592, + 0.052627310156822205, + 0.09034840762615204, + 0.05221787095069885, + 0.03826254606246948, + 0.10601402074098587, + 0.11274387687444687, + 0.2363954782485962 + ], + "credit_map": [ + 0.0094830971211195, + 0.008821536786854267, + 0.040568605065345764, + 0.019627507776021957, + 0.009077753871679306, + 0.03127496689558029, + 0.013822032138705254, + 0.014315908774733543, + 0.010672815144062042, + 0.014217574149370193, + 0.01039259321987629, + 0.012716902419924736, + 0.008804690092802048, + 0.00883406400680542, + 0.01018756814301014, + 0.008759560994803905, + 0.008872633799910545, + 0.007100339978933334, + 0.005629443563520908, + 0.0077408477663993835, + 0.00439708586782217, + 0.021180300042033195, + 0.005414576269686222, + 0.01947811245918274, + 0.005061142612248659, + 0.007593255490064621, + 0.0057294247671961784, + 0.006059014238417149, + 0.009289891459047794, + 0.0045882719568908215, + 0.0064183929935097694, + 0.007887915708124638, + 0.011984108947217464, + 0.013487366959452629, + 0.011525665409862995, + 0.010446134954690933, + 0.012133708223700523, + 0.011176089756190777, + 0.009755698032677174, + 0.009839639067649841, + 0.008267014287412167, + 0.006154300179332495, + 0.00579450698569417, + 0.006176540162414312, + 0.0044032675214111805, + 0.007744475733488798, + 0.00755017064511776, + 0.006127599161118269, + 0.0034100941848009825, + 0.00678672082722187, + 0.006330537609755993, + 0.003985997289419174, + 0.004207192454487085, + 0.003624326316639781, + 0.004478429909795523, + 0.005439247936010361, + 0.012346372939646244, + 0.00523598724976182, + 0.033728744834661484, + 0.006899967789649963, + 0.004758782684803009, + 0.030279381200671196, + 0.006338572129607201, + 0.006426209583878517, + 0.00987876858562231, + 0.010684747248888016, + 0.009350580163300037, + 0.009882146492600441, + 0.030528653413057327, + 0.024250490590929985, + 0.0074718245305120945, + 0.010696656070649624, + 0.009844997897744179, + 0.009437691420316696, + 0.015249723568558693, + 0.011847005225718021, + 0.021422401070594788, + 0.075449638068676, + 0.07510766386985779, + 0.018036354333162308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.009790733456611633 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625281810760498 + ], + "opposite_logprob_tokens": [ + -4.882625102996826 + ], + "logprob_reward": -0.007625281810760498, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625281810760498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11959349364042282, + 0.04336259514093399, + 0.09883254021406174, + 0.043799933046102524, + 0.031044455245137215, + 0.05911115184426308, + 0.0470084473490715, + 0.13953997194766998, + 0.0844639390707016, + 0.33324354887008667 + ], + "credit_map": [ + 0.007813003845512867, + 0.005266629625111818, + 0.039515022188425064, + 0.018705876544117928, + 0.008189029060304165, + 0.02231484465301037, + 0.00918474979698658, + 0.008604333736002445, + 0.007665472570806742, + 0.011354967951774597, + 0.004237591754645109, + 0.00418814504519105, + 0.002739478135481477, + 0.0052563585340976715, + 0.003715237369760871, + 0.004205339588224888, + 0.009741067886352539, + 0.0063925194554030895, + 0.005354108288884163, + 0.008929247036576271, + 0.004707019776105881, + 0.030061300843954086, + 0.0059769670478999615, + 0.027670295909047127, + 0.0049510011449456215, + 0.007504408713430166, + 0.00685419887304306, + 0.005632760003209114, + 0.004617445636540651, + 0.004391959402710199, + 0.003969178535044193, + 0.005878975614905357, + 0.004001540597528219, + 0.003803844330832362, + 0.004701617639511824, + 0.0029215719550848007, + 0.004375857301056385, + 0.002351118251681328, + 0.00449713971465826, + 0.004391763359308243, + 0.007849384099245071, + 0.004326876252889633, + 0.009251246228814125, + 0.0052186655811965466, + 0.004497552756220102, + 0.015165407210588455, + 0.006323342677205801, + 0.006478673778474331, + 0.004440510645508766, + 0.005574231967329979, + 0.007825839333236217, + 0.005725186318159103, + 0.005053611006587744, + 0.005645786412060261, + 0.004453339148312807, + 0.008289936929941177, + 0.017874713987112045, + 0.017200175672769547, + 0.026794834062457085, + 0.008737828582525253, + 0.009928464889526367, + 0.03337688744068146, + 0.01108505018055439, + 0.014541996642947197, + 0.008750597946345806, + 0.00844452902674675, + 0.009795796126127243, + 0.0059262169525027275, + 0.022379891946911812, + 0.015155111439526081, + 0.006165755912661552, + 0.00784603413194418, + 0.014431565068662167, + 0.0051823873072862625, + 0.017382612451910973, + 0.012890240177512169, + 0.03080427274107933, + 0.11493495851755142, + 0.11229922622442245, + 0.025318270549178123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.007625281810760498 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009772672317922115 + ], + "opposite_logprob_tokens": [ + -4.634772777557373 + ], + "logprob_reward": -0.009772672317922115, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009772672317922115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2173633575439453, + 0.10039622336626053, + 0.32965925335884094, + 0.03000909648835659, + 0.021495671942830086, + 0.030724432319402695, + 0.03282063081860542, + 0.04697222262620926, + 0.11982221901416779, + 0.07073689997196198 + ], + "credit_map": [ + 0.02666749432682991, + 0.016698405146598816, + 0.049497924745082855, + 0.022840427234768867, + 0.014114879071712494, + 0.05260583385825157, + 0.012344690039753914, + 0.022593703120946884, + 0.011686130426824093, + 0.018098261207342148, + 0.018141023814678192, + 0.013912764377892017, + 0.005996224004775286, + 0.010872921906411648, + 0.0094599649310112, + 0.012228925712406635, + 0.031043605878949165, + 0.020414704456925392, + 0.03154660761356354, + 0.03067951649427414, + 0.011686034500598907, + 0.09447672218084335, + 0.014000355266034603, + 0.09581167995929718, + 0.003704232396557927, + 0.0023266500793397427, + 0.006008211988955736, + 0.0028799818828701973, + 0.0024188251700252295, + 0.0037930861581116915, + 0.003752301912754774, + 0.005125805269926786, + 0.003214414929971099, + 0.0021769762970507145, + 0.003159203100949526, + 0.0022046209778636694, + 0.002050537383183837, + 0.003612924600020051, + 0.002675906987860799, + 0.002401087898761034, + 0.003141232533380389, + 0.0038910494185984135, + 0.0036823430564254522, + 0.003221327206119895, + 0.0030043809674680233, + 0.005176955368369818, + 0.004063779022544622, + 0.004543366376310587, + 0.0028492731507867575, + 0.006945450324565172, + 0.004408720415085554, + 0.003180720144882798, + 0.003968308214098215, + 0.0027518433053046465, + 0.0038330298848450184, + 0.004883287008851767, + 0.005793875083327293, + 0.003259889082983136, + 0.012168305926024914, + 0.004410139750689268, + 0.0031514812726527452, + 0.011708439327776432, + 0.0033718931954354048, + 0.0031082001514732838, + 0.012705034576356411, + 0.009340334683656693, + 0.007361400872468948, + 0.007581048645079136, + 0.03617527708411217, + 0.029327772557735443, + 0.006696217693388462, + 0.010635127313435078, + 0.003894105553627014, + 0.0028866170905530453, + 0.0062348609790205956, + 0.0043424335308372974, + 0.005895659327507019, + 0.020399892702698708, + 0.021206818521022797, + 0.005876509472727776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.009772672317922115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, long-shot lens, rotating slowly, captures a young Middle Eastern girl, her thick, dark curls tied into a tight ponytail, practicing alone in a vast indoor gymnastics hall. She wears a shimmering, blue-blue, long-sleeved gymnastics leotard, a striking contrast to the wide, padded royal blue mat beneath. She is centered in the frame, moving with intense concentration along a low, tan balance beam, her bare feet pointing, each deliberate, measured step. Suddenly, her focus wanders, and her body teeters precariously to one side. For a brief, tense moment, her arms, previously held out for balance, flutter slightly. Then, through an apparent effort of will, she tightens her core, firmly stretches her arms straight out to her sides like wings, and successfully regains her stability. Her gaze refocuses on the end of the beam, her expression shifting from a fleeting panic to one of unwavering determination. Bright, even fluorescent lighting from the high ceilings illuminates the space, casting soft shadows that emphasize the quiet, focused atmosphere of individual training. Other gymnastic equipment, such as the uneven bars and vault, are visible but out of focus in the background.", + "seed": 10543, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_0/index_4029_seed_10543_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3084359237924218, + "vlm_logprob": -0.3084359237924218, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.3084359237924218, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34601420164108276 + ], + "opposite_logprob_tokens": [ + -1.8460142612457275 + ], + "logprob_reward": -0.34601420164108276, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.34601420164108276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23895148932933807, + 0.05592259019613266, + 0.11745389550924301, + 0.1571701616048813, + 0.1392316222190857, + 0.04356773942708969, + 0.07137082517147064, + 0.06749575585126877, + 0.04708000645041466, + 0.06175593286752701 + ], + "credit_map": [ + 0.021772729232907295, + 0.028071871027350426, + 0.029335983097553253, + 0.031776927411556244, + 0.047946199774742126, + 0.030760427936911583, + 0.025082610547542572, + 0.024204732850193977, + 0.007188704796135426, + 0.011518930085003376, + 0.007347301114350557, + 0.007088484242558479, + 0.0054167513735592365, + 0.0049697356298565865, + 0.006454606540501118, + 0.005938076879829168, + 0.011012078262865543, + 0.014195646159350872, + 0.015163143165409565, + 0.011530141346156597, + 0.014913912862539291, + 0.02458963543176651, + 0.01095723919570446, + 0.015092100948095322, + 0.023153331130743027, + 0.03965074196457863, + 0.013756264932453632, + 0.01896398700773716, + 0.01650151237845421, + 0.018139656633138657, + 0.010122108273208141, + 0.016882561147212982, + 0.013806584291160107, + 0.01288833748549223, + 0.01485811173915863, + 0.01802195981144905, + 0.02366027981042862, + 0.020170290023088455, + 0.014270581305027008, + 0.021555479615926743, + 0.007215507328510284, + 0.006182433571666479, + 0.0048227268271148205, + 0.007196871098130941, + 0.004662783350795507, + 0.005622073542326689, + 0.0036204918287694454, + 0.004244853276759386, + 0.008119049482047558, + 0.007716202177107334, + 0.007678791414946318, + 0.006594204809516668, + 0.012352582067251205, + 0.01596906967461109, + 0.006196071393787861, + 0.006744856480509043, + 0.009573245421051979, + 0.006989185232669115, + 0.009694213978946209, + 0.0075413440354168415, + 0.00594719173386693, + 0.010851806029677391, + 0.007025348022580147, + 0.00987341906875372, + 0.004636103753000498, + 0.006053578108549118, + 0.005263377446681261, + 0.0070982687175273895, + 0.0034428639337420464, + 0.010722997598350048, + 0.004199401009827852, + 0.00566341495141387, + 0.006590761244297028, + 0.007169670425355434, + 0.008968445472419262, + 0.006086181849241257, + 0.00893465057015419, + 0.008969626389443874, + 0.006457853596657515, + 0.008578741922974586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.34601420164108276 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.137035846710205 + ], + "opposite_logprob_tokens": [ + -0.3870358467102051 + ], + "logprob_reward": -1.137035846710205, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.137035846710205, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13730531930923462, + 0.07988424599170685, + 0.27460309863090515, + 0.11059608310461044, + 0.08361604809761047, + 0.04636310413479805, + 0.03876829147338867, + 0.038336172699928284, + 0.12210328131914139, + 0.06842440366744995 + ], + "credit_map": [ + 0.011884691193699837, + 0.01248196605592966, + 0.016152894124388695, + 0.021189013496041298, + 0.017529966309666634, + 0.02026245929300785, + 0.0216363538056612, + 0.01616797409951687, + 0.008706310763955116, + 0.01225193589925766, + 0.009372361935675144, + 0.011431961320340633, + 0.007806768640875816, + 0.009096560068428516, + 0.008379361592233181, + 0.012838984839618206, + 0.02501172199845314, + 0.03760237246751785, + 0.03321905806660652, + 0.025403272360563278, + 0.040068235248327255, + 0.05867655947804451, + 0.019627047702670097, + 0.03499479591846466, + 0.013371803797781467, + 0.011756442487239838, + 0.011472693644464016, + 0.018959257751703262, + 0.013686387799680233, + 0.015144621953368187, + 0.010305429808795452, + 0.015899432823061943, + 0.007285761646926403, + 0.007646982558071613, + 0.008614715188741684, + 0.009852643124759197, + 0.013615002855658531, + 0.014325848780572414, + 0.007214820012450218, + 0.015060262754559517, + 0.005399229470640421, + 0.006972871720790863, + 0.0057664127089083195, + 0.0072518447414040565, + 0.004619959741830826, + 0.004418798256665468, + 0.005132779944688082, + 0.00680120475590229, + 0.004739250522106886, + 0.007338136900216341, + 0.00487061683088541, + 0.0049127875827252865, + 0.004480276722460985, + 0.005537298042327166, + 0.0028828848153352737, + 0.004007034469395876, + 0.004792104009538889, + 0.004469636827707291, + 0.004457008559256792, + 0.004982444923371077, + 0.004611174575984478, + 0.005489508621394634, + 0.004526195582002401, + 0.0050080930814146996, + 0.01803438551723957, + 0.01502162404358387, + 0.015341285616159439, + 0.014239992946386337, + 0.013264480046927929, + 0.01776621863245964, + 0.015219327062368393, + 0.013215967454016209, + 0.007728684227913618, + 0.007772515527904034, + 0.01005484163761139, + 0.008689535781741142, + 0.007742032408714294, + 0.008503563702106476, + 0.008326245471835136, + 0.009606985375285149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -1.137035846710205 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023316606879234314 + ], + "opposite_logprob_tokens": [ + -3.7733166217803955 + ], + "logprob_reward": -0.023316606879234314, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023316606879234314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10517217218875885, + 0.06247348710894585, + 0.1426323503255844, + 0.17569631338119507, + 0.19933679699897766, + 0.05381619185209274, + 0.0799417495727539, + 0.07721216231584549, + 0.05737753584980965, + 0.046341270208358765 + ], + "credit_map": [ + 0.011018075980246067, + 0.010314121842384338, + 0.012373349629342556, + 0.014361613430082798, + 0.016548968851566315, + 0.01412919256836176, + 0.015889976173639297, + 0.010536880232393742, + 0.005772051401436329, + 0.007750487420707941, + 0.009321559220552444, + 0.010884675197303295, + 0.005155119113624096, + 0.007556084543466568, + 0.007834460586309433, + 0.008199050091207027, + 0.013526823371648788, + 0.026300737634301186, + 0.014324256218969822, + 0.015791254118084908, + 0.01735595613718033, + 0.022959383204579353, + 0.015101288445293903, + 0.017272647470235825, + 0.029101286083459854, + 0.044743139296770096, + 0.015501915477216244, + 0.016850724816322327, + 0.021090175956487656, + 0.02342350035905838, + 0.008620301261544228, + 0.01636526547372341, + 0.011893151327967644, + 0.044896915555000305, + 0.02133391797542572, + 0.024347441270947456, + 0.03224076330661774, + 0.018854603171348572, + 0.028940485790371895, + 0.01682952418923378, + 0.005894868168979883, + 0.011858096346259117, + 0.006770712789148092, + 0.009340488351881504, + 0.00574540626257658, + 0.004036338068544865, + 0.006164176389575005, + 0.004006108269095421, + 0.00825495645403862, + 0.0068954117596149445, + 0.008181278593838215, + 0.006019541993737221, + 0.01956324465572834, + 0.01911657489836216, + 0.005990214645862579, + 0.005920526571571827, + 0.011353046633303165, + 0.017392804846167564, + 0.008209158666431904, + 0.007920058444142342, + 0.007942046038806438, + 0.008833611384034157, + 0.007926823571324348, + 0.007634609472006559, + 0.0072677284479141235, + 0.008284743875265121, + 0.007883219048380852, + 0.008267458528280258, + 0.006485764868557453, + 0.006472111213952303, + 0.007293842267245054, + 0.005422667600214481, + 0.0034167435951530933, + 0.006771111860871315, + 0.007690960541367531, + 0.005923902150243521, + 0.007655564695596695, + 0.00500547606498003, + 0.003994463011622429, + 0.005883048754185438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.023316606879234314 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253711897879839 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.01253711897879839, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253711897879839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08936218172311783, + 0.09461555629968643, + 0.40128293633461, + 0.07941131293773651, + 0.08369172364473343, + 0.03780420869588852, + 0.0881894901394844, + 0.02949591539800167, + 0.03756964951753616, + 0.058577004820108414 + ], + "credit_map": [ + 0.007903694175183773, + 0.008501374162733555, + 0.014938491396605968, + 0.015209571458399296, + 0.012271222658455372, + 0.009694253094494343, + 0.011042936705052853, + 0.00980063620954752, + 0.0081018703058362, + 0.015159755945205688, + 0.012417376041412354, + 0.013124949298799038, + 0.010365381836891174, + 0.010612762533128262, + 0.013794438913464546, + 0.011039016768336296, + 0.043411385267972946, + 0.0283027496188879, + 0.05423238128423691, + 0.03324040398001671, + 0.0757104828953743, + 0.0939038097858429, + 0.03508426249027252, + 0.037397462874650955, + 0.0064512756653130054, + 0.008059690706431866, + 0.00815195869654417, + 0.015553265810012817, + 0.007638491224497557, + 0.010047249495983124, + 0.012101261876523495, + 0.01140812411904335, + 0.008261699229478836, + 0.009514293633401394, + 0.008332440629601479, + 0.01058795303106308, + 0.0105636827647686, + 0.013845488429069519, + 0.007912563160061836, + 0.014673604629933834, + 0.006769651081413031, + 0.003932819701731205, + 0.005088571924716234, + 0.004779419861733913, + 0.004091979470103979, + 0.003039867617189884, + 0.003779363352805376, + 0.006322537083178759, + 0.009893608279526234, + 0.014003658667206764, + 0.00928693450987339, + 0.01092195138335228, + 0.01756315864622593, + 0.013988562859594822, + 0.006651721429079771, + 0.005879892036318779, + 0.00413432065397501, + 0.005732669495046139, + 0.003975299187004566, + 0.005100509151816368, + 0.002604937180876732, + 0.0023408173583447933, + 0.00267585227265954, + 0.002931509166955948, + 0.0051176343113183975, + 0.0037381225265562534, + 0.005571267567574978, + 0.004306417889893055, + 0.0038525916170328856, + 0.004986051935702562, + 0.0031919507309794426, + 0.00680561363697052, + 0.005490609910339117, + 0.006656999234110117, + 0.008845698088407516, + 0.0071502141654491425, + 0.009389986284077168, + 0.006412209011614323, + 0.00644519180059433, + 0.008186091668903828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.01253711897879839 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle in a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera rotating slowly throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young Middle Eastern girl with dark curly hair in a ponytail wearing a blue long-sleeved leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the girl perform a sequence of balancing on a tan beam, teetering, fluttering her arms, and then regaining stability with arms stretched out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are out-of-focus gymnastics equipment like uneven bars and vault visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023275844752788544 + ], + "opposite_logprob_tokens": [ + -3.773275852203369 + ], + "logprob_reward": -0.023275844752788544, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023275844752788544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09725727140903473, + 0.24926337599754333, + 0.12462031096220016, + 0.09893781691789627, + 0.0959412232041359, + 0.04424482211470604, + 0.1340722143650055, + 0.05151551961898804, + 0.06635034084320068, + 0.03779710456728935 + ], + "credit_map": [ + 0.00839291512966156, + 0.00890109408646822, + 0.013759138062596321, + 0.01738007366657257, + 0.015868213027715683, + 0.009879537858068943, + 0.013814321719110012, + 0.009261976927518845, + 0.02840598113834858, + 0.04707450047135353, + 0.028175998479127884, + 0.033916134387254715, + 0.026424571871757507, + 0.02298804558813572, + 0.034070130437612534, + 0.028207985684275627, + 0.012719860300421715, + 0.01594225876033306, + 0.014722933061420918, + 0.015881359577178955, + 0.011449120007455349, + 0.02012133039534092, + 0.013136751018464565, + 0.020646685734391212, + 0.014288416132330894, + 0.022457333281636238, + 0.008773011155426502, + 0.009233465418219566, + 0.013831070624291897, + 0.015732567757368088, + 0.00551204290241003, + 0.009109901264309883, + 0.014709338545799255, + 0.010788796469569206, + 0.008719135075807571, + 0.011577736586332321, + 0.012886128388345242, + 0.013360979035496712, + 0.0101653803139925, + 0.013733725063502789, + 0.00797422043979168, + 0.004433915484696627, + 0.005754251964390278, + 0.004298028070479631, + 0.005347795318812132, + 0.0054539963603019714, + 0.003018740564584732, + 0.007963873445987701, + 0.014762424863874912, + 0.010375563986599445, + 0.014270136132836342, + 0.011362331919372082, + 0.028381269425153732, + 0.03408202901482582, + 0.009398017078638077, + 0.01144042331725359, + 0.00635677482932806, + 0.0061776465736329556, + 0.006867930758744478, + 0.009413938038051128, + 0.004646029323339462, + 0.005672496743500233, + 0.005737205967307091, + 0.0066434904001653194, + 0.009211317636072636, + 0.010870593599975109, + 0.009669901803135872, + 0.00826253555715084, + 0.006984359119087458, + 0.007882518693804741, + 0.007327123545110226, + 0.006141988560557365, + 0.004031778313219547, + 0.0037043255288153887, + 0.006858846638351679, + 0.005467941053211689, + 0.003960772417485714, + 0.004992398899048567, + 0.003329506143927574, + 0.005451533943414688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.023275844752788544 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_1/index_tensor([6989, 6989])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_1/index_tensor([6989, 6989])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4b1a1fcc3a603d502982ab0259c9f81ed7b1655c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_1/index_tensor([6989, 6989])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level close-up captures a tall, precariously balanced tower constructed from brightly colored wooden blocks, laid atop a polished, light-colored wood floor. The blocks, composed of primary-colored cubes, cylinders, and triangular prisms, reveal subtle wood grain beneath a smooth, semi-gloss painted surface. The entire scene is bathed in the soft, warm glow of real playroom lights from an unseen overhead source, casting soft, elongated shadows. A hand swiftly enters the frame from the right and, with decisive precision, slides a foundational red cube from the very bottom of the stack. For a moment, the tower hangs in suspended animation, teetering precariously with a slight, fatal sway before a chain reaction of collapse begins, a dynamic movement captured in delicate slow motion. The structure first implodes downward, then the blocks explode outward in a chaotic, energetic cascade, tumbling and rattling against each other as they scatter in all directions. As the movement subsides, the camera remains steady, eventually revealing the aftermath, with blocks widely scattered across the floor, motionless, a vibrant scene of playful chaos in a serene, unchanging light.", + "seed": 24238, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_1/index_6989_seed_24238_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.017170645762234926, + "vlm_logprob": -0.017170645762234926, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.017170645762234926, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012290691956877708 + ], + "opposite_logprob_tokens": [ + -5.3872904777526855 + ], + "logprob_reward": -0.012290691956877708, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.012290691956877708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0352969616651535, + 0.0675407350063324, + 0.08461348712444305, + 0.06241007149219513, + 0.02000003308057785, + 0.016369489952921867, + 0.025718802586197853, + 0.01909160055220127, + 0.08430285006761551, + 0.5846559405326843 + ], + "credit_map": [ + 0.0031249176245182753, + 0.003460174659267068, + 0.004452952183783054, + 0.0039399960078299046, + 0.003814740339294076, + 0.007615953218191862, + 0.003439173102378845, + 0.005449060350656509, + 0.0047638071700930595, + 0.0029144587460905313, + 0.01989835873246193, + 0.003481917316094041, + 0.005164258647710085, + 0.006469585467129946, + 0.0030758564826101065, + 0.02177249826490879, + 0.0055391425266861916, + 0.0030246670357882977, + 0.021598871797323227, + 0.004964612890034914, + 0.0051240576431155205, + 0.013769019395112991, + 0.005188880488276482, + 0.025404246523976326, + 0.00328322546556592, + 0.002821848727762699, + 0.019813761115074158, + 0.003549369750544429, + 0.0043195015750825405, + 0.005061359144747257, + 0.0024907593615353107, + 0.02107025682926178, + 0.00150114088319242, + 0.003362164366990328, + 0.002576563274487853, + 0.003324955701828003, + 0.0025582534726709127, + 0.002643166808411479, + 0.0019408534280955791, + 0.0020929360762238503, + 0.0017457256326451898, + 0.0017803607042878866, + 0.0018331160536035895, + 0.0015647377585992217, + 0.002654292853549123, + 0.002982962876558304, + 0.0017982523422688246, + 0.002010044874623418, + 0.0024984886404126883, + 0.0041223871521651745, + 0.002994697308167815, + 0.002474940149113536, + 0.004092876333743334, + 0.0045821163803339005, + 0.0027247383259236813, + 0.0022285606246441603, + 0.0013865859946236014, + 0.002091059461236, + 0.001711424090899527, + 0.0023269716184586287, + 0.0029647264163941145, + 0.003029152750968933, + 0.0036667603999376297, + 0.0019149220315739512, + 0.00271040596999228, + 0.007336342707276344, + 0.01212207693606615, + 0.0028046804945915937, + 0.03410516679286957, + 0.017454514279961586, + 0.0031721617560833693, + 0.004597515799105167, + 0.022475970908999443, + 0.2171480655670166, + 0.03713606297969818, + 0.009285501204431057, + 0.24550768733024597, + 0.039281923323869705, + 0.007428979966789484, + 0.006391812115907669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.012290691956877708 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308966174721718 + ], + "opposite_logprob_tokens": [ + -3.1680896282196045 + ], + "logprob_reward": -0.04308966174721718, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308966174721718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.011936748400330544, + 0.018394112586975098, + 0.13131451606750488, + 0.021543066948652267, + 0.009065463207662106, + 0.012300722301006317, + 0.015673505142331123, + 0.013145701959729195, + 0.14363828301429749, + 0.6229879260063171 + ], + "credit_map": [ + 0.00111825589556247, + 0.001219572382979095, + 0.0014637249987572432, + 0.001533924019895494, + 0.001800249912776053, + 0.0017684566555544734, + 0.0015207488322630525, + 0.0015118159353733063, + 0.001294266781769693, + 0.0011304512154310942, + 0.004028332885354757, + 0.001405727700330317, + 0.002363592619076371, + 0.0019574235193431377, + 0.00130607676692307, + 0.004908238537609577, + 0.006761753931641579, + 0.0032261498272418976, + 0.04303281381726265, + 0.004905535373836756, + 0.004726119805127382, + 0.01564669981598854, + 0.005009978078305721, + 0.048005443066358566, + 0.0016131011070683599, + 0.001184777356684208, + 0.005793989636003971, + 0.0012992449337616563, + 0.0014502519043162465, + 0.002799367066472769, + 0.0009898846037685871, + 0.0064124492928385735, + 0.000777012319304049, + 0.0015915782423689961, + 0.0008996957913041115, + 0.0013612533221021295, + 0.0013909831177443266, + 0.0010759260039776564, + 0.0010901918867602944, + 0.0008788222912698984, + 0.0014725172659382224, + 0.0015077879652380943, + 0.0015198776964098215, + 0.0008679159800522029, + 0.0021604658104479313, + 0.002332282019779086, + 0.001211434486322105, + 0.0012284392723813653, + 0.0011111878557130694, + 0.003341563744470477, + 0.0017200619913637638, + 0.0012655823957175016, + 0.001631063292734325, + 0.004030381329357624, + 0.0010683010332286358, + 0.0015053619863465428, + 0.0007494168821722269, + 0.0012231769505888224, + 0.0008457425865344703, + 0.001216145814396441, + 0.003548193257302046, + 0.003105323761701584, + 0.0013820870080962777, + 0.0010756151750683784, + 0.003866920480504632, + 0.014674498699605465, + 0.009161487221717834, + 0.002464389195665717, + 0.06523852795362473, + 0.041839685291051865, + 0.00317120598629117, + 0.0032215663231909275, + 0.025357825681567192, + 0.2341204285621643, + 0.04774659499526024, + 0.011880353093147278, + 0.22144624590873718, + 0.06437279284000397, + 0.008236115798354149, + 0.009827536530792713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.04308966174721718 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007641607895493507 + ], + "opposite_logprob_tokens": [ + -4.882641792297363 + ], + "logprob_reward": -0.007641607895493507, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007641607895493507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04279748350381851, + 0.05922025069594383, + 0.09896327555179596, + 0.04732907563447952, + 0.03977994620800018, + 0.053772520273923874, + 0.03523199260234833, + 0.012690260075032711, + 0.2374730110168457, + 0.3727421462535858 + ], + "credit_map": [ + 0.004757023882120848, + 0.005089597776532173, + 0.005141030065715313, + 0.003470619209110737, + 0.006855363491922617, + 0.007625858765095472, + 0.003948623314499855, + 0.005909366998821497, + 0.0032686120830476284, + 0.002077834913507104, + 0.01954716630280018, + 0.0033581878524273634, + 0.003425849135965109, + 0.0044299704022705555, + 0.0020039663650095463, + 0.021108662709593773, + 0.004889973439276218, + 0.003198822494596243, + 0.034149546176195145, + 0.003672100603580475, + 0.0034049800597131252, + 0.009657831862568855, + 0.0032417650800198317, + 0.0367482528090477, + 0.00543509004637599, + 0.003910321742296219, + 0.007181901019066572, + 0.0036839370150119066, + 0.006122124847024679, + 0.009325332939624786, + 0.003573982510715723, + 0.00809638760983944, + 0.003279252676293254, + 0.0027751820161938667, + 0.004475601948797703, + 0.0021990088280290365, + 0.005469071678817272, + 0.012503590434789658, + 0.002476438181474805, + 0.006601798813790083, + 0.006236110348254442, + 0.007747480180114508, + 0.005821872036904097, + 0.004995766561478376, + 0.008849210105836391, + 0.009406030178070068, + 0.004410541150718927, + 0.006305514369159937, + 0.003230930771678686, + 0.00884185265749693, + 0.0029310949612408876, + 0.0030089598149061203, + 0.005422511138021946, + 0.006310371682047844, + 0.002752516884356737, + 0.0027337507344782352, + 0.0008560643182136118, + 0.0012108376249670982, + 0.001082992646843195, + 0.0016253386856988072, + 0.0019969684071838856, + 0.002308641793206334, + 0.002466363599523902, + 0.0011430520098656416, + 0.006225676741451025, + 0.023640772327780724, + 0.006392110604792833, + 0.002283484674990177, + 0.10654021054506302, + 0.08068139851093292, + 0.004839136730879545, + 0.006870224606245756, + 0.010513149201869965, + 0.09706747531890869, + 0.03598649054765701, + 0.004808883182704449, + 0.16780710220336914, + 0.047218143939971924, + 0.0052877627313137054, + 0.004053155891597271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.007641607895493507 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016086315736174583 + ], + "opposite_logprob_tokens": [ + -4.141086101531982 + ], + "logprob_reward": -0.016086315736174583, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016086315736174583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030613716691732407, + 0.03301657363772392, + 0.036977097392082214, + 0.0793117955327034, + 0.014737793244421482, + 0.015386870130896568, + 0.040449358522892, + 0.07914862781763077, + 0.3531953692436218, + 0.31716275215148926 + ], + "credit_map": [ + 0.0026821529027074575, + 0.0024794612545520067, + 0.003934346605092287, + 0.004218685440719128, + 0.004923908039927483, + 0.004966401495039463, + 0.0038580046966671944, + 0.003550758119672537, + 0.002981151221320033, + 0.0029593389481306076, + 0.004469782579690218, + 0.003037347923964262, + 0.004766493570059538, + 0.004255274310708046, + 0.002747618593275547, + 0.0077995662577450275, + 0.002813569037243724, + 0.001092113321647048, + 0.009434289298951626, + 0.002023975132033229, + 0.0016192540060728788, + 0.006894807331264019, + 0.0019583976827561855, + 0.011140692047774792, + 0.004813852719962597, + 0.0026421945076435804, + 0.025606941431760788, + 0.0028409045189619064, + 0.004084589425474405, + 0.008772210218012333, + 0.0029074251651763916, + 0.027643676847219467, + 0.0010690796189010143, + 0.0023200667928904295, + 0.0015845269663259387, + 0.002393417526036501, + 0.0018342823022976518, + 0.0022725549060851336, + 0.001749195042066276, + 0.0015146690420806408, + 0.0015344833955168724, + 0.00343130761757493, + 0.0017703433986753225, + 0.0014630453661084175, + 0.0024446004536002874, + 0.0020422227680683136, + 0.00142631062772125, + 0.0012745565036311746, + 0.0033371958415955305, + 0.007513514254242182, + 0.004448587074875832, + 0.0024895973037928343, + 0.008293897844851017, + 0.0070071835070848465, + 0.0030975528061389923, + 0.004261830355972052, + 0.004719746299088001, + 0.02990526333451271, + 0.0027723079547286034, + 0.004038865678012371, + 0.014862185344099998, + 0.016128506511449814, + 0.003595534712076187, + 0.00312621146440506, + 0.009117280133068562, + 0.031644079834222794, + 0.009660698473453522, + 0.003008835017681122, + 0.15896938741207123, + 0.12554888427257538, + 0.006171037442982197, + 0.009075183421373367, + 0.009365151636302471, + 0.08925045281648636, + 0.02880270965397358, + 0.004144830629229546, + 0.14209279417991638, + 0.03607679530978203, + 0.004372612107545137, + 0.0030573992989957333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.016086315736174583 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067449514754116535 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.0067449514754116535, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067449514754116535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020964616909623146, + 0.05326047167181969, + 0.07050488889217377, + 0.018429923802614212, + 0.028033249080181122, + 0.026498887687921524, + 0.04338323324918747, + 0.15086203813552856, + 0.3658903241157532, + 0.22217242419719696 + ], + "credit_map": [ + 0.00205293670296669, + 0.0022310896310955286, + 0.0029383995570242405, + 0.0017096728552132845, + 0.00362350232899189, + 0.003710165387019515, + 0.0017559323459863663, + 0.002942917635664344, + 0.002736567286774516, + 0.0022593264002352953, + 0.017574910074472427, + 0.0024876510724425316, + 0.003079841146245599, + 0.004176229238510132, + 0.001632580766454339, + 0.01931336522102356, + 0.002906210022047162, + 0.001955130137503147, + 0.026429658755660057, + 0.002373272553086281, + 0.00242431927472353, + 0.004796131514012814, + 0.0015957762952893972, + 0.02802438661456108, + 0.0013687815517187119, + 0.0013492513680830598, + 0.0037542590871453285, + 0.002317255362868309, + 0.0021944832988083363, + 0.002662212587893009, + 0.0013592456234619021, + 0.0034244360867887735, + 0.002147928113117814, + 0.0051648663356900215, + 0.0029531309846788645, + 0.0041753752157092094, + 0.0038761107716709375, + 0.0033591717947274446, + 0.0033390200696885586, + 0.0030176453292369843, + 0.0023961905390024185, + 0.004183528013527393, + 0.002202133182436228, + 0.00303644803352654, + 0.004934418946504593, + 0.003512534312903881, + 0.003465051297098398, + 0.002768581500276923, + 0.004627035465091467, + 0.0073874532245099545, + 0.004030147101730108, + 0.004262389149516821, + 0.005034307483583689, + 0.010366931557655334, + 0.003279351396486163, + 0.004395617172122002, + 0.006337828002870083, + 0.05546769127249718, + 0.0038303867913782597, + 0.005272011738270521, + 0.05943325534462929, + 0.010612118989229202, + 0.004235303495079279, + 0.005673440173268318, + 0.00920008309185505, + 0.034935496747493744, + 0.016475042328238487, + 0.0035217972472310066, + 0.1664557307958603, + 0.1196008250117302, + 0.007617373019456863, + 0.008083988912403584, + 0.009452633559703827, + 0.009221434593200684, + 0.05174500122666359, + 0.0043695536442101, + 0.099317766726017, + 0.040781617164611816, + 0.004245996940881014, + 0.0030384138226509094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.0067449514754116535 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures a tall, precariously balanced tower constructed from brightly colored wooden blocks, laid atop a polished, light-colored wood floor. The blocks, composed of primary-colored cubes, cylinders, and triangular prisms, reveal subtle wood grain beneath a smooth, semi-gloss painted surface. The entire scene is bathed in the soft, warm glow of real playroom lights from an unseen overhead source, casting soft, elongated shadows. A hand swiftly enters the frame from the right and, with decisive precision, slides a foundational red cube from the very bottom of the stack. For a moment, the tower hangs in suspended animation, teetering precariously with a slight, fatal sway before a chain reaction of collapse begins, a dynamic movement captured in delicate slow motion. The structure first implodes downward, then the blocks explode outward in a chaotic, energetic cascade, tumbling and rattling against each other as they scatter in all directions. As the movement subsides, the camera remains steady, eventually revealing the aftermath, with blocks widely scattered across the floor, motionless, a vibrant scene of playful chaos in a serene, unchanging light.", + "seed": 24239, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_1/index_6989_seed_24239_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01807774193584919, + "vlm_logprob": -0.01807774193584919, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.01807774193584919, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02924310974776745 + ], + "opposite_logprob_tokens": [ + -4.029242992401123 + ], + "logprob_reward": -0.02924310974776745, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02924310974776745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06180088222026825, + 0.04900272935628891, + 0.023497041314840317, + 0.03936038911342621, + 0.1679287850856781, + 0.3552490472793579, + 0.08761584758758545, + 0.048380665481090546, + 0.04188178479671478, + 0.12528283894062042 + ], + "credit_map": [ + 0.006633472628891468, + 0.009668593294918537, + 0.010013574734330177, + 0.011686800979077816, + 0.005620240233838558, + 0.006045382469892502, + 0.0061152661219239235, + 0.006017552223056555, + 0.006345023401081562, + 0.005932052619755268, + 0.007550196256488562, + 0.010728895664215088, + 0.006103668361902237, + 0.0034978080075234175, + 0.005180321633815765, + 0.003664764342829585, + 0.0018141524633392692, + 0.002602991880849004, + 0.0029675608966499567, + 0.0038829289842396975, + 0.0036589137744158506, + 0.002720099175348878, + 0.003673699451610446, + 0.0021766931749880314, + 0.00441278051584959, + 0.004816047847270966, + 0.003992769401520491, + 0.00534132681787014, + 0.005141977220773697, + 0.004392195958644152, + 0.007319585885852575, + 0.003943705931305885, + 0.00895602349191904, + 0.008939968422055244, + 0.010373035445809364, + 0.010391294956207275, + 0.05101846531033516, + 0.04742114618420601, + 0.015087738633155823, + 0.015741121023893356, + 0.018691396340727806, + 0.038424015045166016, + 0.014137444086372852, + 0.013710849918425083, + 0.15028218924999237, + 0.09902244061231613, + 0.009691998362541199, + 0.011288701556622982, + 0.0070029315538704395, + 0.014595762826502323, + 0.012417708523571491, + 0.005205556284636259, + 0.02860453352332115, + 0.011373438872396946, + 0.004093669354915619, + 0.0043222494423389435, + 0.0030125484336167574, + 0.005611598026007414, + 0.004889843985438347, + 0.004873236641287804, + 0.01520819403231144, + 0.0084841875359416, + 0.0030537741258740425, + 0.003247281303629279, + 0.0036947207991033792, + 0.0036660167388617992, + 0.008058487437665462, + 0.002710401313379407, + 0.015315210446715355, + 0.003864181460812688, + 0.0026310980319976807, + 0.0019416671711951494, + 0.008270863443613052, + 0.018026022240519524, + 0.027456630021333694, + 0.0030993816908448935, + 0.053911156952381134, + 0.00794847309589386, + 0.003276078263297677, + 0.0032942290417850018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.02924310974776745 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821715384721756 + ], + "opposite_logprob_tokens": [ + -4.018217086791992 + ], + "logprob_reward": -0.01821715384721756, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01821715384721756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08765935897827148, + 0.06596583873033524, + 0.020630955696105957, + 0.029759546741843224, + 0.10130351781845093, + 0.2470378428697586, + 0.05594239383935928, + 0.032272305339574814, + 0.17013663053512573, + 0.18929161131381989 + ], + "credit_map": [ + 0.008230755105614662, + 0.01090298593044281, + 0.014204409904778004, + 0.011517913080751896, + 0.009533779695630074, + 0.012637262232601643, + 0.009429840371012688, + 0.011202412657439709, + 0.00584025401622057, + 0.007774836383759975, + 0.007503775414079428, + 0.009954879060387611, + 0.010458733886480331, + 0.007481136824935675, + 0.011526801623404026, + 0.0054254187270998955, + 0.0018829458858817816, + 0.002092879731208086, + 0.0036126133054494858, + 0.0034725703299045563, + 0.0026829277630895376, + 0.0018821454141288996, + 0.0021506864577531815, + 0.0028541868086904287, + 0.0034334459342062473, + 0.0038905006367713213, + 0.0031069288961589336, + 0.004640035796910524, + 0.003415422048419714, + 0.0038123014383018017, + 0.0044693793170154095, + 0.002991532441228628, + 0.007152488920837641, + 0.016065899282693863, + 0.006978054065257311, + 0.0065228864550590515, + 0.03025601990520954, + 0.016622232273221016, + 0.006471885833889246, + 0.01123405247926712, + 0.009211506694555283, + 0.018098382279276848, + 0.0050875176675617695, + 0.007416347041726112, + 0.10918011516332626, + 0.08606203645467758, + 0.0047822799533605576, + 0.007199652958661318, + 0.005208371672779322, + 0.005654350388795137, + 0.014093994162976742, + 0.0055087837390601635, + 0.00950122345238924, + 0.005235140677541494, + 0.004794929642230272, + 0.005945604760199785, + 0.0024656718596816063, + 0.003353278851136565, + 0.0033141481690108776, + 0.0032298548612743616, + 0.006683743093162775, + 0.0061219222843647, + 0.003383230185136199, + 0.003720455104485154, + 0.012023501098155975, + 0.0233770739287138, + 0.02175227180123329, + 0.007114825304597616, + 0.0715869888663292, + 0.0215801652520895, + 0.0065009393729269505, + 0.00620086956769228, + 0.011237489990890026, + 0.023982474580407143, + 0.040593381971120834, + 0.0041823480278253555, + 0.08405929803848267, + 0.016814813017845154, + 0.004466387443244457, + 0.003955427091568708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.01821715384721756 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018201351165771484 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.018201351165771484, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018201351165771484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05086982622742653, + 0.016366010531783104, + 0.01806468330323696, + 0.06418055295944214, + 0.13577894866466522, + 0.31014499068260193, + 0.08559315651655197, + 0.06203710287809372, + 0.07525638490915298, + 0.18170839548110962 + ], + "credit_map": [ + 0.005528048612177372, + 0.007078478112816811, + 0.007023362442851067, + 0.009803039021790028, + 0.004257278051227331, + 0.006332726217806339, + 0.004675623029470444, + 0.00617126515135169, + 0.001868973602540791, + 0.0023219811264425516, + 0.002501091221347451, + 0.002582067856565118, + 0.0015392390778288245, + 0.0016363020986318588, + 0.002541369292885065, + 0.0013749852078035474, + 0.001859232666902244, + 0.001789064728654921, + 0.0024315339978784323, + 0.002643098821863532, + 0.002544407034292817, + 0.0018219203921034932, + 0.002127357292920351, + 0.0028480670880526304, + 0.006261037662625313, + 0.00766559224575758, + 0.006231679115444422, + 0.007521861232817173, + 0.009942857548594475, + 0.008353738114237785, + 0.010575919412076473, + 0.007627859711647034, + 0.0061339461244642735, + 0.011628903448581696, + 0.0065912422724068165, + 0.006792204920202494, + 0.05320523679256439, + 0.039360299706459045, + 0.006289960350841284, + 0.005777140147984028, + 0.013444377109408379, + 0.024280885234475136, + 0.006426694802939892, + 0.00948717724531889, + 0.13577128946781158, + 0.10521405935287476, + 0.006456657778471708, + 0.009063818491995335, + 0.004500071983784437, + 0.011541768908500671, + 0.007603921927511692, + 0.002304277615621686, + 0.03414033353328705, + 0.018682459369301796, + 0.0031700143590569496, + 0.0036503029987215996, + 0.0043234494514763355, + 0.005889698397368193, + 0.0036613678093999624, + 0.002398391719907522, + 0.02635316364467144, + 0.01488005556166172, + 0.002532459795475006, + 0.0019985109101980925, + 0.0059052761644124985, + 0.010461006313562393, + 0.005759517662227154, + 0.0033728396520018578, + 0.03220219537615776, + 0.012323545292019844, + 0.00266565615311265, + 0.002566346200183034, + 0.01103433407843113, + 0.024774784222245216, + 0.03885425999760628, + 0.0034445703495293856, + 0.08028331398963928, + 0.014516834169626236, + 0.004487011581659317, + 0.004313281737267971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.018201351165771484 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160837359726429 + ], + "opposite_logprob_tokens": [ + -4.141083717346191 + ], + "logprob_reward": -0.0160837359726429, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160837359726429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056874945759773254, + 0.02545471489429474, + 0.013747402466833591, + 0.04388079419732094, + 0.07056088000535965, + 0.39873290061950684, + 0.0756847932934761, + 0.0202057845890522, + 0.05826159939169884, + 0.23659612238407135 + ], + "credit_map": [ + 0.00807527918368578, + 0.007871409878134727, + 0.008022990077733994, + 0.011523975059390068, + 0.004864860791712999, + 0.005856949836015701, + 0.004985952749848366, + 0.005673539359122515, + 0.002677277661859989, + 0.003612987231463194, + 0.004506124649196863, + 0.005845487583428621, + 0.0022362309973686934, + 0.0017934524221345782, + 0.002698219381272793, + 0.0020849371794611216, + 0.0014516195515170693, + 0.0013835072750225663, + 0.001385865150950849, + 0.0025986069813370705, + 0.001641550101339817, + 0.001482196617871523, + 0.0018797264201566577, + 0.001924330834299326, + 0.0036114929243922234, + 0.005619796458631754, + 0.00506112165749073, + 0.009717319160699844, + 0.00406744796782732, + 0.005253503564745188, + 0.005153736099600792, + 0.0053963810205459595, + 0.003274762537330389, + 0.00654953345656395, + 0.007119644898921251, + 0.006031522061675787, + 0.019037237390875816, + 0.016939256340265274, + 0.004945279099047184, + 0.0066636549308896065, + 0.017434781417250633, + 0.031135769560933113, + 0.00468585267663002, + 0.0072341980412602425, + 0.18070374429225922, + 0.13891884684562683, + 0.0073570613749325275, + 0.011262661777436733, + 0.007596169598400593, + 0.00842063594609499, + 0.007370969746261835, + 0.004775677807629108, + 0.020470814779400826, + 0.016078749671578407, + 0.0046234834007918835, + 0.0063483016565442085, + 0.0012081849854439497, + 0.001658054767176509, + 0.0017903316766023636, + 0.002043267246335745, + 0.004950232338160276, + 0.005532541777938604, + 0.0015866025350987911, + 0.0014365705428645015, + 0.003633603686466813, + 0.005286234896630049, + 0.01864122785627842, + 0.0025791365187615156, + 0.018277758732438087, + 0.005473040509968996, + 0.0022037490271031857, + 0.002166852355003357, + 0.015326479449868202, + 0.03269556164741516, + 0.04906928539276123, + 0.004671080037951469, + 0.1046149805188179, + 0.01879647932946682, + 0.006033043842762709, + 0.005389237776398659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.0160837359726429 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643358945846558 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643358945846558, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643358945846558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.017648393288254738, + 0.009765942580997944, + 0.008892939426004887, + 0.02050701342523098, + 0.14549575746059418, + 0.45278581976890564, + 0.11719957739114761, + 0.028728283941745758, + 0.036076128482818604, + 0.1629001498222351 + ], + "credit_map": [ + 0.001767024165019393, + 0.0027357195504009724, + 0.0026806266978383064, + 0.0035985864233225584, + 0.0017029172740876675, + 0.002018648898229003, + 0.0014291268307715654, + 0.0017157439142465591, + 0.0013333394890651107, + 0.0012925609480589628, + 0.0014097981620579958, + 0.002238447777926922, + 0.0007960632210597396, + 0.0007537487545050681, + 0.0009544545318931341, + 0.0009875291725620627, + 0.0008930139010772109, + 0.0010224636644124985, + 0.0013080238131806254, + 0.0015112755354493856, + 0.0007931065047159791, + 0.0012502792524173856, + 0.0006967378430999815, + 0.0014180389698594809, + 0.0021005759481340647, + 0.0024829860776662827, + 0.0022183568216860294, + 0.0025165711995214224, + 0.0038967847358435392, + 0.00246503041125834, + 0.0027853138744831085, + 0.0020413934253156185, + 0.00801005493849516, + 0.016843542456626892, + 0.012084797024726868, + 0.011389418505132198, + 0.032715387642383575, + 0.048144206404685974, + 0.008235659450292587, + 0.00807268638163805, + 0.021009590476751328, + 0.03489508107304573, + 0.011371898464858532, + 0.010979106649756432, + 0.19966907799243927, + 0.15306010842323303, + 0.00888926349580288, + 0.012911682948470116, + 0.006818488705903292, + 0.015391319058835506, + 0.010658946819603443, + 0.003828091314062476, + 0.04654526710510254, + 0.024522963911294937, + 0.004179623909294605, + 0.005254879593849182, + 0.002116801217198372, + 0.003578599775210023, + 0.003753469791263342, + 0.0026913387700915337, + 0.0062608541920781136, + 0.005907097365707159, + 0.001988109666854143, + 0.0024320126976817846, + 0.003240507561713457, + 0.005288212560117245, + 0.004487819503992796, + 0.003142222296446562, + 0.010619793087244034, + 0.0028699624817818403, + 0.0034212754108011723, + 0.003006335813552141, + 0.010509743355214596, + 0.02308080904185772, + 0.038174472749233246, + 0.004284783732146025, + 0.06934304535388947, + 0.008809325285255909, + 0.00428759353235364, + 0.004410382825881243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.008643358945846558 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures a tall, precariously balanced tower constructed from brightly colored wooden blocks, laid atop a polished, light-colored wood floor. The blocks, composed of primary-colored cubes, cylinders, and triangular prisms, reveal subtle wood grain beneath a smooth, semi-gloss painted surface. The entire scene is bathed in the soft, warm glow of real playroom lights from an unseen overhead source, casting soft, elongated shadows. A hand swiftly enters the frame from the right and, with decisive precision, slides a foundational red cube from the very bottom of the stack. For a moment, the tower hangs in suspended animation, teetering precariously with a slight, fatal sway before a chain reaction of collapse begins, a dynamic movement captured in delicate slow motion. The structure first implodes downward, then the blocks explode outward in a chaotic, energetic cascade, tumbling and rattling against each other as they scatter in all directions. As the movement subsides, the camera remains steady, eventually revealing the aftermath, with blocks widely scattered across the floor, motionless, a vibrant scene of playful chaos in a serene, unchanging light.", + "seed": 24240, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_1/index_6989_seed_24240_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01770043782889843, + "vlm_logprob": -0.01770043782889843, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.01770043782889843, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01929011195898056 + ], + "opposite_logprob_tokens": [ + -4.394289970397949 + ], + "logprob_reward": -0.01929011195898056, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01929011195898056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05098438635468483, + 0.03844722360372543, + 0.041835587471723557, + 0.02608729898929596, + 0.15920107066631317, + 0.07382284849882126, + 0.3394920527935028, + 0.02948608621954918, + 0.06722550094127655, + 0.173417866230011 + ], + "credit_map": [ + 0.006123325787484646, + 0.00733004929497838, + 0.004892573691904545, + 0.006040595937520266, + 0.007013645954430103, + 0.007604636717587709, + 0.005879506003111601, + 0.006100059486925602, + 0.004476745147258043, + 0.005369795020669699, + 0.0028389275539666414, + 0.004698474891483784, + 0.006440663244575262, + 0.004613140132278204, + 0.005372538696974516, + 0.004636941011995077, + 0.006582653149962425, + 0.003378592198714614, + 0.006800269242376089, + 0.003103502793237567, + 0.003456854959949851, + 0.009205041453242302, + 0.0027226933743804693, + 0.006585984490811825, + 0.002538597211241722, + 0.002912405878305435, + 0.001955337356775999, + 0.0021458573173731565, + 0.004165211226791143, + 0.007556781638413668, + 0.0023997731041163206, + 0.002413337118923664, + 0.012743827886879444, + 0.021706214174628258, + 0.010221723467111588, + 0.021456459537148476, + 0.030194422230124474, + 0.01859433203935623, + 0.02130001038312912, + 0.022984104230999947, + 0.005778132472187281, + 0.0067221419885754585, + 0.005502086598426104, + 0.00646201940253377, + 0.024312270805239677, + 0.009419984184205532, + 0.007993576116859913, + 0.007632644381374121, + 0.04293639585375786, + 0.05448690056800842, + 0.020691635087132454, + 0.029863843694329262, + 0.04540855064988136, + 0.06990130245685577, + 0.035907331854104996, + 0.040296122431755066, + 0.003411260200664401, + 0.002976939082145691, + 0.004358567763119936, + 0.002820210298523307, + 0.0038137659430503845, + 0.0035751189570873976, + 0.004168195649981499, + 0.004362030886113644, + 0.009682059288024902, + 0.0058876811526715755, + 0.008399582467973232, + 0.005398971028625965, + 0.01602037437260151, + 0.010747094638645649, + 0.0039000255055725574, + 0.007189717143774033, + 0.013249034993350506, + 0.004382177256047726, + 0.005528994370251894, + 0.004174388479441404, + 0.07300157845020294, + 0.0648680180311203, + 0.003262335667386651, + 0.004951327573508024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.01929011195898056 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014223858714103699 + ], + "opposite_logprob_tokens": [ + -4.264224052429199 + ], + "logprob_reward": -0.014223858714103699, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014223858714103699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06452246755361557, + 0.05757586285471916, + 0.13987170159816742, + 0.0319385752081871, + 0.1413334459066391, + 0.10396943986415863, + 0.11346190422773361, + 0.053451452404260635, + 0.0924965962767601, + 0.20137853920459747 + ], + "credit_map": [ + 0.010296011343598366, + 0.008814944885671139, + 0.00622867327183485, + 0.006991453934460878, + 0.006839465815573931, + 0.009510296396911144, + 0.0076845851726830006, + 0.008157040923833847, + 0.007516658864915371, + 0.007230022922158241, + 0.005418286193162203, + 0.007337735965847969, + 0.008507377468049526, + 0.008739667013287544, + 0.005935872904956341, + 0.006890248041599989, + 0.01820511370897293, + 0.017078036442399025, + 0.01968380995094776, + 0.012651010416448116, + 0.01788969151675701, + 0.021199533715844154, + 0.014164356514811516, + 0.019000153988599777, + 0.0028602275997400284, + 0.004346377681940794, + 0.0027873420622199774, + 0.003526120912283659, + 0.0057840896770358086, + 0.0048470692709088326, + 0.004550564102828503, + 0.003236785065382719, + 0.012541740201413631, + 0.01943155750632286, + 0.010046749375760555, + 0.020573412999510765, + 0.022531984373927116, + 0.016167620196938515, + 0.019178273156285286, + 0.02086210995912552, + 0.01113079208880663, + 0.009073936380445957, + 0.00759554747492075, + 0.008381913416087627, + 0.032154303044080734, + 0.01020113006234169, + 0.014490189030766487, + 0.010941632091999054, + 0.013048012740910053, + 0.01537511870265007, + 0.008285744115710258, + 0.011008033528923988, + 0.015032228082418442, + 0.0238194540143013, + 0.014123154804110527, + 0.012770168483257294, + 0.009245178662240505, + 0.004041630309075117, + 0.005597822833806276, + 0.004074552096426487, + 0.01337401568889618, + 0.007266195956617594, + 0.004420922603458166, + 0.0054311384446918964, + 0.010613573715090752, + 0.012007149867713451, + 0.011313355527818203, + 0.009536308236420155, + 0.0152334775775671, + 0.012939447537064552, + 0.0075099291279911995, + 0.013343364000320435, + 0.014731728471815586, + 0.007453062571585178, + 0.009170198813080788, + 0.006641733460128307, + 0.08078141510486603, + 0.0700947642326355, + 0.005358139052987099, + 0.007147497963160276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.014223858714103699 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330962009727955 + ], + "opposite_logprob_tokens": [ + -3.7733097076416016 + ], + "logprob_reward": -0.02330962009727955, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330962009727955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04831618070602417, + 0.027086293324828148, + 0.04413147270679474, + 0.16536656022071838, + 0.14881348609924316, + 0.019034937024116516, + 0.398680716753006, + 0.02887648157775402, + 0.0340217687189579, + 0.08567207306623459 + ], + "credit_map": [ + 0.0074881152249872684, + 0.005323654506355524, + 0.0051467204466462135, + 0.00551680289208889, + 0.006422909442335367, + 0.0064083971083164215, + 0.006009896285831928, + 0.00599968247115612, + 0.0029994116630405188, + 0.0032726572826504707, + 0.002297680824995041, + 0.0029031008016318083, + 0.0037863687612116337, + 0.004770946688950062, + 0.0037633178289979696, + 0.003292810171842575, + 0.00601087836548686, + 0.006556264590471983, + 0.005679667927324772, + 0.0048382398672401905, + 0.005808141082525253, + 0.0053440360352396965, + 0.004075890406966209, + 0.0058183567598462105, + 0.01828344725072384, + 0.017828194424510002, + 0.01312035322189331, + 0.017925268039107323, + 0.02935313992202282, + 0.03241999074816704, + 0.01495964452624321, + 0.021476512774825096, + 0.013326296582818031, + 0.01966078020632267, + 0.012062556110322475, + 0.021663937717676163, + 0.022362401708960533, + 0.016363948583602905, + 0.020419400185346603, + 0.022954171523451805, + 0.0025501244235783815, + 0.0025155970361083746, + 0.001879002433270216, + 0.001631061895750463, + 0.0037238236982375383, + 0.0019778343848884106, + 0.0017110593616962433, + 0.0030464334413409233, + 0.048284534364938736, + 0.0616142675280571, + 0.02616027183830738, + 0.03475869074463844, + 0.0511791817843914, + 0.08377277851104736, + 0.04513038322329521, + 0.047780606895685196, + 0.005636926274746656, + 0.0033940887078642845, + 0.002489968203008175, + 0.002449776977300644, + 0.004694389179348946, + 0.00462674954906106, + 0.0023421740625053644, + 0.0032424081582576036, + 0.004117527510970831, + 0.003375408938154578, + 0.004265446215867996, + 0.003580203279852867, + 0.007310665678232908, + 0.0049117691814899445, + 0.002894097939133644, + 0.0035666488111019135, + 0.005434622056782246, + 0.00455802446231246, + 0.0057715605944395065, + 0.004595029167830944, + 0.03095119632780552, + 0.02963230013847351, + 0.0019729090854525566, + 0.0027564300689846277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.02330962009727955 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059725485742092 + ], + "opposite_logprob_tokens": [ + -3.895597219467163 + ], + "logprob_reward": -0.02059725485742092, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059725485742092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022144120186567307, + 0.011726582422852516, + 0.01813529245555401, + 0.004060840699821711, + 0.01358295138925314, + 0.005079944618046284, + 0.8346226215362549, + 0.009982704184949398, + 0.006082732230424881, + 0.07458215206861496 + ], + "credit_map": [ + 0.0031621691305190325, + 0.0019604640547186136, + 0.0025456359144300222, + 0.002095837378874421, + 0.003304061247035861, + 0.0036964036989957094, + 0.0024510740768164396, + 0.002928478643298149, + 0.0015027334447950125, + 0.0013661934062838554, + 0.001286239828914404, + 0.0013960470678284764, + 0.0018884841119870543, + 0.0015225857496261597, + 0.001289689913392067, + 0.0014746102970093489, + 0.0027466705068945885, + 0.0019373181276023388, + 0.0028102281503379345, + 0.0019505732925608754, + 0.002137346426025033, + 0.0024399771355092525, + 0.0016381969908252358, + 0.00247498438693583, + 0.0003310237661935389, + 0.000509685545694083, + 0.00029162486316636205, + 0.00039576253038831055, + 0.0006915549165569246, + 0.0008621764136478305, + 0.0006201190408319235, + 0.00035889449645765126, + 0.0010707273613661528, + 0.0019364039180800319, + 0.0009689513826742768, + 0.0019147841958329082, + 0.0024987661745399237, + 0.0015645453240722418, + 0.0018667004769667983, + 0.001762074651196599, + 0.0005113811930641532, + 0.0004392729024402797, + 0.0007555779884569347, + 0.0004551468591671437, + 0.0008570206700824201, + 0.0005465320427902043, + 0.000858311599586159, + 0.000656702381093055, + 0.104584701359272, + 0.13179507851600647, + 0.05131043121218681, + 0.06845444440841675, + 0.11069522798061371, + 0.17827662825584412, + 0.08918710798025131, + 0.10031908005475998, + 0.0015904689207673073, + 0.0017178085399791598, + 0.0009537015575915575, + 0.0009560653707012534, + 0.001880364608950913, + 0.0009957874426618218, + 0.0008110895869322121, + 0.001077419612556696, + 0.0010238307295367122, + 0.0006838556146249175, + 0.0006462733144871891, + 0.00045817936188541353, + 0.0012119364691898227, + 0.0004895063466392457, + 0.0006695385091006756, + 0.0008996130200102925, + 0.003649238497018814, + 0.0016833411063998938, + 0.002275734907016158, + 0.0012966325739398599, + 0.031605083495378494, + 0.031122902408242226, + 0.0010615388164296746, + 0.0018876938847824931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.02059725485742092 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108134351670742 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.01108134351670742, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108134351670742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07983105629682541, + 0.04762592911720276, + 0.07271970808506012, + 0.04340716078877449, + 0.2718510031700134, + 0.052207332104444504, + 0.08150897175073624, + 0.04133997857570648, + 0.06915764510631561, + 0.24035120010375977 + ], + "credit_map": [ + 0.009991721250116825, + 0.008679846301674843, + 0.010049612261354923, + 0.010448613204061985, + 0.009413999505341053, + 0.010626520030200481, + 0.010756907053291798, + 0.009863837622106075, + 0.005712330341339111, + 0.006186462473124266, + 0.005210167728364468, + 0.0057979607954621315, + 0.007524639833718538, + 0.006448732689023018, + 0.005031854845583439, + 0.005713773425668478, + 0.011479618959128857, + 0.00939235370606184, + 0.009347154758870602, + 0.007045859005302191, + 0.009623480029404163, + 0.009151262231171131, + 0.006840507499873638, + 0.009839464910328388, + 0.004379726015031338, + 0.0070367916487157345, + 0.003826734609901905, + 0.006697030272334814, + 0.0068662832491099834, + 0.004129988607019186, + 0.006097251083701849, + 0.0043733506463468075, + 0.02074720337986946, + 0.037835508584976196, + 0.016011720523238182, + 0.03873191773891449, + 0.05293511226773262, + 0.03030325472354889, + 0.037525054067373276, + 0.03776122257113457, + 0.004666510969400406, + 0.004415119998157024, + 0.005551312584429979, + 0.0053672087378799915, + 0.013280370272696018, + 0.005244998726993799, + 0.007401346694678068, + 0.006280457600951195, + 0.009665449149906635, + 0.011609835550189018, + 0.007455695886164904, + 0.008227492682635784, + 0.009854722768068314, + 0.01563148945569992, + 0.008110098540782928, + 0.010954180732369423, + 0.006803794763982296, + 0.003431319957599044, + 0.004736230708658695, + 0.004377133212983608, + 0.005752186756581068, + 0.007331710308790207, + 0.004249608144164085, + 0.004657990299165249, + 0.009865093044936657, + 0.006440148688852787, + 0.0069733066484332085, + 0.004921456333249807, + 0.01713290996849537, + 0.011983649805188179, + 0.004502218682318926, + 0.0073388563469052315, + 0.018896104767918587, + 0.011377006769180298, + 0.017448455095291138, + 0.01059929933398962, + 0.09039518982172012, + 0.0789748802781105, + 0.005090096965432167, + 0.007570172194391489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.01108134351670742 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures a tall, precariously balanced tower constructed from brightly colored wooden blocks, laid atop a polished, light-colored wood floor. The blocks, composed of primary-colored cubes, cylinders, and triangular prisms, reveal subtle wood grain beneath a smooth, semi-gloss painted surface. The entire scene is bathed in the soft, warm glow of real playroom lights from an unseen overhead source, casting soft, elongated shadows. A hand swiftly enters the frame from the right and, with decisive precision, slides a foundational red cube from the very bottom of the stack. For a moment, the tower hangs in suspended animation, teetering precariously with a slight, fatal sway before a chain reaction of collapse begins, a dynamic movement captured in delicate slow motion. The structure first implodes downward, then the blocks explode outward in a chaotic, energetic cascade, tumbling and rattling against each other as they scatter in all directions. As the movement subsides, the camera remains steady, eventually revealing the aftermath, with blocks widely scattered across the floor, motionless, a vibrant scene of playful chaos in a serene, unchanging light.", + "seed": 24241, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_1/index_6989_seed_24241_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.034528106823563574, + "vlm_logprob": -0.034528106823563574, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.034528106823563574, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022033773362636566 + ], + "opposite_logprob_tokens": [ + -4.14703369140625 + ], + "logprob_reward": -0.022033773362636566, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.022033773362636566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0788247361779213, + 0.03903563320636749, + 0.08120989799499512, + 0.03975001722574234, + 0.022503219544887543, + 0.029344068840146065, + 0.08467726409435272, + 0.4776202440261841, + 0.1107562780380249, + 0.03627869859337807 + ], + "credit_map": [ + 0.01172262616455555, + 0.007495693396776915, + 0.013229144737124443, + 0.0049770367331802845, + 0.00453025009483099, + 0.020858364179730415, + 0.004302190151065588, + 0.01170942559838295, + 0.006558812689036131, + 0.004862293601036072, + 0.005447915755212307, + 0.00409634318202734, + 0.004397735930979252, + 0.005961410701274872, + 0.0032703285105526447, + 0.004440791439265013, + 0.006613351404666901, + 0.017687203362584114, + 0.008617560379207134, + 0.006770022679120302, + 0.01463682297617197, + 0.010632401332259178, + 0.006416186690330505, + 0.009836341254413128, + 0.004317748825997114, + 0.010409083217382431, + 0.004260748624801636, + 0.0029152054339647293, + 0.004996560513973236, + 0.005714293569326401, + 0.0028638597577810287, + 0.004272517282515764, + 0.002598372520878911, + 0.00358145241625607, + 0.0020655591506510973, + 0.002156841801479459, + 0.003316574264317751, + 0.0036419585812836885, + 0.0026553566567599773, + 0.0024871034547686577, + 0.005033404100686312, + 0.003294207388535142, + 0.003566665342077613, + 0.0030374687630683184, + 0.003931158222258091, + 0.004339710809290409, + 0.002761745359748602, + 0.0033797090873122215, + 0.005785282235592604, + 0.007148840930312872, + 0.01070094108581543, + 0.016980690881609917, + 0.0152423195540905, + 0.010361768305301666, + 0.012618057429790497, + 0.005839360412210226, + 0.0198175311088562, + 0.0495455227792263, + 0.031272418797016144, + 0.04095527157187462, + 0.0947689414024353, + 0.05925259366631508, + 0.09174036979675293, + 0.0902675986289978, + 0.013128217309713364, + 0.017904413864016533, + 0.01575036719441414, + 0.013798650354146957, + 0.011443664319813251, + 0.01429077424108982, + 0.011873866431415081, + 0.012566328048706055, + 0.003837894881144166, + 0.00546044809743762, + 0.005634626839309931, + 0.003857716452330351, + 0.003196725621819496, + 0.004245480056852102, + 0.0050069852732121944, + 0.005038822069764137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.022033773362636566 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636488899588585 + ], + "opposite_logprob_tokens": [ + -3.651364803314209 + ], + "logprob_reward": -0.02636488899588585, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636488899588585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09250376373529434, + 0.0440998375415802, + 0.0788613110780716, + 0.10055875033140182, + 0.03864830732345581, + 0.020725693553686142, + 0.057284582406282425, + 0.42444509267807007, + 0.10697844624519348, + 0.03589421138167381 + ], + "credit_map": [ + 0.01063248049467802, + 0.011825373396277428, + 0.012766573578119278, + 0.006124429404735565, + 0.007441596128046513, + 0.024154189974069595, + 0.006637223530560732, + 0.012921895831823349, + 0.010927355848252773, + 0.00430223299190402, + 0.004719748627394438, + 0.00350361131131649, + 0.0038570649921894073, + 0.008703292347490788, + 0.003040990559384227, + 0.005045538768172264, + 0.009675520472228527, + 0.013110688887536526, + 0.00678604282438755, + 0.007739929482340813, + 0.0142755638808012, + 0.010394887998700142, + 0.008163717575371265, + 0.008714963681995869, + 0.010985881090164185, + 0.03334793448448181, + 0.007151758763939142, + 0.007932300679385662, + 0.013113342225551605, + 0.013597024604678154, + 0.006117844954133034, + 0.008312659338116646, + 0.00640469416975975, + 0.004479510243982077, + 0.005399273708462715, + 0.0034203995019197464, + 0.004562231246381998, + 0.005710477475076914, + 0.003706809366121888, + 0.0049649132415652275, + 0.002815906424075365, + 0.00211510481312871, + 0.0022453225683420897, + 0.003019103081896901, + 0.0022715150844305754, + 0.002728376304730773, + 0.002796443644911051, + 0.0027339220978319645, + 0.002895853715017438, + 0.005915954243391752, + 0.005520923063158989, + 0.011997559107840061, + 0.009626595303416252, + 0.007636378984898329, + 0.009958242066204548, + 0.003733081044629216, + 0.018639596179127693, + 0.042862821370363235, + 0.02550831437110901, + 0.03790901228785515, + 0.08658424019813538, + 0.05319184437394142, + 0.07833629101514816, + 0.08141296356916428, + 0.01129292231053114, + 0.016313403844833374, + 0.014315197244286537, + 0.015292229130864143, + 0.011550229974091053, + 0.012452598661184311, + 0.01248178444802761, + 0.013280077837407589, + 0.00445513753220439, + 0.00695924973115325, + 0.00471495371311903, + 0.0027478996198624372, + 0.0033578339498490095, + 0.00562070868909359, + 0.0037063213530927896, + 0.004332106094807386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.02636488899588585 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059783972799778 + ], + "opposite_logprob_tokens": [ + -3.8955979347229004 + ], + "logprob_reward": -0.02059783972799778, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059783972799778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08117026835680008, + 0.060961514711380005, + 0.15375834703445435, + 0.034281644970178604, + 0.1199645921587944, + 0.19947056472301483, + 0.09988614916801453, + 0.1817656308412552, + 0.04786942899227142, + 0.020871806889772415 + ], + "credit_map": [ + 0.011479751206934452, + 0.009166963398456573, + 0.013279425911605358, + 0.004002029541879892, + 0.006228545680642128, + 0.022723672911524773, + 0.002951233647763729, + 0.011338655836880207, + 0.009385950863361359, + 0.008040090091526508, + 0.006825971882790327, + 0.0053503322415053844, + 0.008444515988230705, + 0.01012696884572506, + 0.0051520089618861675, + 0.00763568002730608, + 0.012128260917961597, + 0.033890269696712494, + 0.00873472634702921, + 0.011760466732084751, + 0.04169577360153198, + 0.017688704654574394, + 0.011997489258646965, + 0.015862667933106422, + 0.003497911849990487, + 0.006019128020852804, + 0.0025642705149948597, + 0.00513206934556365, + 0.005254665855318308, + 0.004206183832138777, + 0.004163661040365696, + 0.0034437577705830336, + 0.018949024379253387, + 0.014629244804382324, + 0.013597681187093258, + 0.009552926756441593, + 0.017601264640688896, + 0.019875792786478996, + 0.013933666981756687, + 0.01182499248534441, + 0.02541649527847767, + 0.030460970476269722, + 0.028331823647022247, + 0.02851259522140026, + 0.01990465633571148, + 0.031575024127960205, + 0.01730164885520935, + 0.017967374995350838, + 0.009602406993508339, + 0.012568693608045578, + 0.01566985435783863, + 0.013951394706964493, + 0.014739101752638817, + 0.010658576153218746, + 0.011409071274101734, + 0.011287065222859383, + 0.008248436264693737, + 0.018149113282561302, + 0.01468776073306799, + 0.020076138898730278, + 0.029727604240179062, + 0.01960444264113903, + 0.036245305091142654, + 0.03502683341503143, + 0.0031698220409452915, + 0.0049663325771689415, + 0.006441588047891855, + 0.009179958142340183, + 0.004810445476323366, + 0.0037392450030893087, + 0.007787943352013826, + 0.007774100638926029, + 0.0018445011228322983, + 0.003767954418435693, + 0.0029296218417584896, + 0.0026437134947627783, + 0.0020986893214285374, + 0.00244628987275064, + 0.0024876927491277456, + 0.002653347095474601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.02059783972799778 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868725314736366 + ], + "opposite_logprob_tokens": [ + -3.048687219619751 + ], + "logprob_reward": -0.04868725314736366, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868725314736366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07728604972362518, + 0.024760114029049873, + 0.020735453814268112, + 0.006734003312885761, + 0.008742253296077251, + 0.011158738285303116, + 0.04449518397450447, + 0.6856692433357239, + 0.09141840785741806, + 0.029000595211982727 + ], + "credit_map": [ + 0.012220668606460094, + 0.006402897648513317, + 0.013303445652127266, + 0.003739373991265893, + 0.004387098364531994, + 0.022319044917821884, + 0.0029586234595626593, + 0.011954890564084053, + 0.005198152270168066, + 0.002388570923358202, + 0.0037614612374454737, + 0.0020009404979646206, + 0.0024875402450561523, + 0.004730974789708853, + 0.0013756538974121213, + 0.002816818654537201, + 0.0018862624419853091, + 0.004211036022752523, + 0.0018026974285021424, + 0.0018113844562321901, + 0.004460039082914591, + 0.0026152837090194225, + 0.0016888920217752457, + 0.0022598577197641134, + 0.0007501834770664573, + 0.0011682495241984725, + 0.0009049632353708148, + 0.0005298412288539112, + 0.0009199927444569767, + 0.0010725022293627262, + 0.0006243041134439409, + 0.0007639664690941572, + 0.0009404432494193316, + 0.001509019872173667, + 0.0010372685501351953, + 0.000858140760101378, + 0.0011070168111473322, + 0.0014140787534415722, + 0.0009311262401752174, + 0.0009451585356146097, + 0.0017513689817860723, + 0.0015926267951726913, + 0.0013215154176577926, + 0.0008954571676440537, + 0.0017769132973626256, + 0.0017899357480928302, + 0.0010705870809033513, + 0.0009603330981917679, + 0.005257668439298868, + 0.006674609612673521, + 0.0074807810597121716, + 0.005802553612738848, + 0.00526605686172843, + 0.00521232420578599, + 0.005398325622081757, + 0.0034028601367026567, + 0.027457326650619507, + 0.06227806583046913, + 0.040806885808706284, + 0.06196165829896927, + 0.1394733041524887, + 0.08716025948524475, + 0.1369718313217163, + 0.12955984473228455, + 0.01037563756108284, + 0.015241452492773533, + 0.012632801197469234, + 0.011518807150423527, + 0.010685033164918423, + 0.011782838962972164, + 0.008694939315319061, + 0.01048689428716898, + 0.0035897952038794756, + 0.005785152316093445, + 0.003687124000862241, + 0.00254876259714365, + 0.002698242664337158, + 0.004339687526226044, + 0.0026718485169112682, + 0.00367998075671494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.04868725314736366 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of a tower made of brightly colored wooden blocks on a polished light-colored wood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the blocks consist of primary-colored cubes, cylinders, and triangular prisms with a smooth, semi-gloss painted surface showing wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand enter from the right to slide a foundational red cube from the bottom of the stack, causing the tower to teeter and then collapse in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft, warm overhead playroom lights casting soft, elongated shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain steady throughout the collapse and aftermath, revealing blocks widely scattered and motionless on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495677888393402 + ], + "opposite_logprob_tokens": [ + -2.9299566745758057 + ], + "logprob_reward": -0.05495677888393402, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495677888393402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0485602542757988, + 0.016029609367251396, + 0.038666535168886185, + 0.03144675865769386, + 0.0400552824139595, + 0.08954637497663498, + 0.2076345682144165, + 0.3823687732219696, + 0.12893854081630707, + 0.016753332689404488 + ], + "credit_map": [ + 0.005154412239789963, + 0.004518923349678516, + 0.006237603724002838, + 0.003516576951369643, + 0.004684610292315483, + 0.015030412003397942, + 0.002663942286744714, + 0.006753778085112572, + 0.002552291378378868, + 0.002365443855524063, + 0.0020216440316289663, + 0.0015208078548312187, + 0.001872145221568644, + 0.0021041803993284702, + 0.001534231472760439, + 0.002058865502476692, + 0.004537656903266907, + 0.00785134732723236, + 0.00277728377841413, + 0.003217712277546525, + 0.008956093341112137, + 0.004289182368665934, + 0.002889446448534727, + 0.004147810861468315, + 0.003560506971552968, + 0.007638499140739441, + 0.002812647493556142, + 0.002859446918591857, + 0.004180693533271551, + 0.0038917295169085264, + 0.002759740222245455, + 0.003743494162335992, + 0.00530680175870657, + 0.005145983304828405, + 0.004297836218029261, + 0.0046202391386032104, + 0.00502835912629962, + 0.006636166479438543, + 0.004809346981346607, + 0.004210548475384712, + 0.011615505442023277, + 0.012835459783673286, + 0.011306243017315865, + 0.01048359740525484, + 0.01162578072398901, + 0.01350470632314682, + 0.006883610039949417, + 0.011291471309959888, + 0.014435484074056149, + 0.01881069503724575, + 0.02884966880083084, + 0.0378059521317482, + 0.03021688386797905, + 0.026992132887244225, + 0.025162391364574432, + 0.02536134608089924, + 0.018476512283086777, + 0.04198702797293663, + 0.02965737134218216, + 0.03783511370420456, + 0.06608976423740387, + 0.041930265724658966, + 0.07291898876428604, + 0.07347369939088821, + 0.013310055248439312, + 0.019231077283620834, + 0.0202235896140337, + 0.017582876607775688, + 0.015581922605633736, + 0.014723413623869419, + 0.011737432330846786, + 0.01654818095266819, + 0.0022985204122960567, + 0.0027376115322113037, + 0.0025982095394283533, + 0.0018885801546275616, + 0.0016195367788895965, + 0.0022791267838329077, + 0.001277257571928203, + 0.0020544903818517923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.05495677888393402 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_10/index_tensor([5160, 5160])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_10/index_tensor([5160, 5160])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c9d61f9ab79d8107e35a979500769819a985a2c9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_10/index_tensor([5160, 5160])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, full-length photograph depicts a stoic female warrior amidst a tranquil, misty bamboo forest, the entire scene rendered in the elegant, fluid lines of traditional Chinese art. Her face is a mask of cold fury, her eyes narrowed to a deadly focus, her lips pressed into a thin, determined line. Her sleek black hair is curled into an intricate headdress, secured with a simple jade hairpin. She clads herself in a crimson silk robe that billows around her, its movement and intricate cloud-patterned embroidery captured with delicate calligraphic brushstrokes. As an arrow suddenly pierces the frame, she grasps a dark lacquer war fan in her right hand, and with a swift and economical flick, deftly deflecting the arrow. The movement is a blur of controlled grace; the fan's edge precisely connects to the arrowhead, producing a sharp flick of light that momentarily illuminates the scene before the arrow is sent tumbling away. The force of her movements causes her robes to swirl with painterly grace, and her furious expression does not shift as she maintains a powerful, grounded stance, creating an atmosphere of breathtakingly serene beauty juxtaposed with deadly, focused intent.", + "seed": 9954, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_10/index_5160_seed_9954_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05104774385690689, + "vlm_logprob": -0.05104774385690689, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.05104774385690689, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09213995933532715 + ], + "opposite_logprob_tokens": [ + -2.467139959335327 + ], + "logprob_reward": -0.09213995933532715, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09213995933532715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27743759751319885, + 0.12258622795343399, + 0.10379984974861145, + 0.119745172560215, + 0.15203845500946045, + 0.0784110426902771, + 0.052539560943841934, + 0.032567135989665985, + 0.03832273557782173, + 0.0225522443652153 + ], + "credit_map": [ + 0.05960964784026146, + 0.04035412147641182, + 0.035045213997364044, + 0.026329943910241127, + 0.03051527962088585, + 0.039721641689538956, + 0.02032354660332203, + 0.02553819864988327, + 0.019213514402508736, + 0.014672039076685905, + 0.013963134028017521, + 0.010363978333771229, + 0.018567025661468506, + 0.018128007650375366, + 0.008855165913701057, + 0.01882336102426052, + 0.016652114689350128, + 0.021482359617948532, + 0.010140598751604557, + 0.01007427740842104, + 0.012202830985188484, + 0.01551209855824709, + 0.00789092481136322, + 0.009844648651778698, + 0.019245250150561333, + 0.026493029668927193, + 0.008666138164699078, + 0.01770040951669216, + 0.01210301648825407, + 0.016444526612758636, + 0.00906996987760067, + 0.01002283114939928, + 0.04219300299882889, + 0.01380243618041277, + 0.025411788374185562, + 0.011604432947933674, + 0.013588778674602509, + 0.022365430369973183, + 0.01059370581060648, + 0.012478891760110855, + 0.011956797912716866, + 0.007155198138207197, + 0.011717522516846657, + 0.006867155432701111, + 0.015999553725123405, + 0.01126390416175127, + 0.006485932040959597, + 0.006964975968003273, + 0.006668351590633392, + 0.005234419833868742, + 0.005376932676881552, + 0.005436873994767666, + 0.012207724153995514, + 0.006025110837072134, + 0.006519213318824768, + 0.005070934537798166, + 0.009364516474306583, + 0.0032878685742616653, + 0.0029439562931656837, + 0.003149954602122307, + 0.004032536875456572, + 0.005128634162247181, + 0.0024915114045143127, + 0.002168157836422324, + 0.006409169640392065, + 0.006122743245214224, + 0.004541309550404549, + 0.00358039909042418, + 0.006044875830411911, + 0.0059416391886770725, + 0.0029354547150433064, + 0.0027471473440527916, + 0.005632294341921806, + 0.0025614628102630377, + 0.002574823796749115, + 0.001956840744242072, + 0.0033042666036635637, + 0.0031008718069642782, + 0.0016315908869728446, + 0.0017900940729305148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.09213995933532715 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592818036675453 + ], + "opposite_logprob_tokens": [ + -3.8955929279327393 + ], + "logprob_reward": -0.020592818036675453, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592818036675453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31771141290664673, + 0.17956329882144928, + 0.038502562791109085, + 0.09737187623977661, + 0.08668674528598785, + 0.10550272464752197, + 0.08202879130840302, + 0.03243516385555267, + 0.025780323892831802, + 0.03441717103123665 + ], + "credit_map": [ + 0.06593222171068192, + 0.0379372239112854, + 0.032430753111839294, + 0.02853795327246189, + 0.027246546000242233, + 0.04715270549058914, + 0.021222837269306183, + 0.05725112557411194, + 0.05634589120745659, + 0.019401008263230324, + 0.025999201461672783, + 0.009105735458433628, + 0.018242402002215385, + 0.03156226500868797, + 0.009187314659357071, + 0.009719465859234333, + 0.00754107441753149, + 0.0067298272624611855, + 0.005488554481416941, + 0.003386369440704584, + 0.004222834948450327, + 0.0057814414612948895, + 0.0027690683491528034, + 0.0025833870749920607, + 0.013342776335775852, + 0.023631904274225235, + 0.0071351672522723675, + 0.012157212011516094, + 0.011768658645451069, + 0.013892065733671188, + 0.007318350952118635, + 0.008125731721520424, + 0.026586858555674553, + 0.00926668755710125, + 0.011316591873764992, + 0.006859083194285631, + 0.00712723471224308, + 0.014236990362405777, + 0.005015096627175808, + 0.006278202868998051, + 0.028933046385645866, + 0.011575290933251381, + 0.012655320577323437, + 0.0074585601687431335, + 0.014647587202489376, + 0.016074493527412415, + 0.006534516345709562, + 0.007623903453350067, + 0.019962092861533165, + 0.009088663384318352, + 0.00942285731434822, + 0.006590097211301327, + 0.010835424065589905, + 0.0115257129073143, + 0.008224931545555592, + 0.006379002705216408, + 0.010092421434819698, + 0.0037541005294770002, + 0.0026571734342724085, + 0.0026179198175668716, + 0.003541153622791171, + 0.005577527452260256, + 0.002142339013516903, + 0.0020525301806628704, + 0.005385680589824915, + 0.003453694749623537, + 0.003905416466295719, + 0.0023110825568437576, + 0.002994903828948736, + 0.00364102004095912, + 0.0017709940439090133, + 0.0023175301030278206, + 0.009913679212331772, + 0.0026720145251601934, + 0.005273805931210518, + 0.0020293924026191235, + 0.0022693797945976257, + 0.007143907714635134, + 0.0016099060885608196, + 0.003505081171169877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.020592818036675453 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033708129078149796 + ], + "opposite_logprob_tokens": [ + -3.408708095550537 + ], + "logprob_reward": -0.033708129078149796, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033708129078149796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2396104484796524, + 0.09622699022293091, + 0.05796762555837631, + 0.1550435870885849, + 0.14006829261779785, + 0.0693599283695221, + 0.0827326700091362, + 0.0778355523943901, + 0.06215078383684158, + 0.019004086032509804 + ], + "credit_map": [ + 0.04323146864771843, + 0.030477451160550117, + 0.02930414490401745, + 0.021360814571380615, + 0.0245173629373312, + 0.03632739931344986, + 0.018740074709057808, + 0.03565175458788872, + 0.020054927095770836, + 0.00961991772055626, + 0.0154695063829422, + 0.007718991953879595, + 0.011946686543524265, + 0.013850340619683266, + 0.00687268516048789, + 0.01069394126534462, + 0.007844025269150734, + 0.010198323987424374, + 0.007218346931040287, + 0.00610728794708848, + 0.006934802979230881, + 0.00802510417997837, + 0.005574382375925779, + 0.006065353285521269, + 0.025851678103208542, + 0.033585675060749054, + 0.011718693189322948, + 0.018979137763381004, + 0.016702590510249138, + 0.02222893014550209, + 0.01224882248789072, + 0.013728066347539425, + 0.03221704438328743, + 0.014864184893667698, + 0.025993788614869118, + 0.01074887067079544, + 0.01347390841692686, + 0.01929987594485283, + 0.010427478700876236, + 0.01304315309971571, + 0.011387993581593037, + 0.0055206529796123505, + 0.012260978110134602, + 0.006722012534737587, + 0.010152190923690796, + 0.010172439739108086, + 0.006380781996995211, + 0.006762886885553598, + 0.011190993711352348, + 0.006918572820723057, + 0.00896404031664133, + 0.008640180341899395, + 0.01671481691300869, + 0.009607601910829544, + 0.012137974612414837, + 0.008558501489460468, + 0.01590244472026825, + 0.007538820616900921, + 0.007735589053481817, + 0.008310609497129917, + 0.012767873704433441, + 0.010337862186133862, + 0.006572147831320763, + 0.008670210838317871, + 0.010303855873644352, + 0.005970381200313568, + 0.009534112177789211, + 0.006617442239075899, + 0.005755297839641571, + 0.008686097338795662, + 0.005373293999582529, + 0.009910313412547112, + 0.0050852796994149685, + 0.00185811473056674, + 0.0024894364178180695, + 0.00164424953982234, + 0.0018341012764722109, + 0.0033254562877118587, + 0.0012826580787077546, + 0.0014847918646410108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.033708129078149796 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029810355976223946 + ], + "opposite_logprob_tokens": [ + -3.5298104286193848 + ], + "logprob_reward": -0.029810355976223946, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029810355976223946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21636681258678436, + 0.0585041344165802, + 0.06369756907224655, + 0.15584930777549744, + 0.09382522106170654, + 0.06542152911424637, + 0.17144452035427094, + 0.0829649269580841, + 0.05978502333164215, + 0.03214094415307045 + ], + "credit_map": [ + 0.06425312161445618, + 0.01943369023501873, + 0.03649858757853508, + 0.01660888083279133, + 0.013881001621484756, + 0.03429067134857178, + 0.011275426484644413, + 0.020125441253185272, + 0.010797840543091297, + 0.0071272351779043674, + 0.00833593774586916, + 0.006229391787201166, + 0.007920853793621063, + 0.008514598943293095, + 0.004131929483264685, + 0.005446346942335367, + 0.010328017175197601, + 0.00990588590502739, + 0.007021739613264799, + 0.006510468199849129, + 0.00713422242552042, + 0.008748263120651245, + 0.0059791370294988155, + 0.008069838397204876, + 0.02366454154253006, + 0.033320941030979156, + 0.012222911231219769, + 0.02328602410852909, + 0.018190085887908936, + 0.022147774696350098, + 0.011715331114828587, + 0.011301683261990547, + 0.02653571031987667, + 0.009522046893835068, + 0.012803370133042336, + 0.006361543666571379, + 0.009311485104262829, + 0.015292953699827194, + 0.006566897034645081, + 0.007431207690387964, + 0.015815328806638718, + 0.0054361941292881966, + 0.006912692449986935, + 0.007683723699301481, + 0.008993307128548622, + 0.008398950099945068, + 0.005460611078888178, + 0.006720718462020159, + 0.028470957651734352, + 0.016233010217547417, + 0.02153005264699459, + 0.019375305622816086, + 0.03448496758937836, + 0.020540742203593254, + 0.015007623471319675, + 0.015801871195435524, + 0.02644485794007778, + 0.008531739935278893, + 0.005817932076752186, + 0.005291593261063099, + 0.011219747364521027, + 0.013911630026996136, + 0.005129563622176647, + 0.006617862731218338, + 0.01691468618810177, + 0.004663168918341398, + 0.009953157044947147, + 0.006039805710315704, + 0.0045284004881978035, + 0.009741032496094704, + 0.003520616330206394, + 0.0044241552241146564, + 0.008761225268244743, + 0.002352519426494837, + 0.004295126534998417, + 0.00200777780264616, + 0.002554255072027445, + 0.00537033611908555, + 0.002326301531866193, + 0.004473403096199036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.029810355976223946 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898745685815811 + ], + "opposite_logprob_tokens": [ + -2.5789873600006104 + ], + "logprob_reward": -0.07898745685815811, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898745685815811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21980048716068268, + 0.11069068312644958, + 0.0643315240740776, + 0.15405720472335815, + 0.04619526490569115, + 0.07241324335336685, + 0.16988566517829895, + 0.04620083421468735, + 0.0443396121263504, + 0.07208546251058578 + ], + "credit_map": [ + 0.03519995138049126, + 0.03999350965023041, + 0.030261589214205742, + 0.01851518079638481, + 0.025276076048612595, + 0.03229902684688568, + 0.015576363541185856, + 0.022678803652524948, + 0.018214864656329155, + 0.01004431489855051, + 0.01640351302921772, + 0.007238968275487423, + 0.010825369507074356, + 0.01785576157271862, + 0.009062041528522968, + 0.021045850589871407, + 0.010027985088527203, + 0.010986215434968472, + 0.00722693745046854, + 0.005722574889659882, + 0.00733781885355711, + 0.009244732558727264, + 0.005506208166480064, + 0.008279051631689072, + 0.020503075793385506, + 0.030682886019349098, + 0.01118965819478035, + 0.02019534446299076, + 0.015461379662156105, + 0.020617414265871048, + 0.012869996950030327, + 0.02253745123744011, + 0.0106272641569376, + 0.003835164476186037, + 0.006129366345703602, + 0.004867304116487503, + 0.0052724964916706085, + 0.006916379556059837, + 0.003595768241211772, + 0.004951522685587406, + 0.015049638226628304, + 0.00771481916308403, + 0.009573816321790218, + 0.006615378428250551, + 0.010809714905917645, + 0.010643634013831615, + 0.00584881566464901, + 0.006157423369586468, + 0.034604787826538086, + 0.020135197788476944, + 0.01807045005261898, + 0.012317365035414696, + 0.029544875025749207, + 0.02452479489147663, + 0.01668580435216427, + 0.01400238648056984, + 0.011583385057747364, + 0.0042498065158724785, + 0.005403179209679365, + 0.005538961384445429, + 0.0047317068092525005, + 0.0068086134269833565, + 0.00416556978598237, + 0.0037196148186922073, + 0.01091249380260706, + 0.0047285049222409725, + 0.006454992573708296, + 0.005089163780212402, + 0.004128042608499527, + 0.006720422767102718, + 0.0029056130442768335, + 0.0034003802575170994, + 0.02322130650281906, + 0.005490981973707676, + 0.011429431848227978, + 0.003108937991783023, + 0.0037597608752548695, + 0.01759139820933342, + 0.0027195666916668415, + 0.004764071200042963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.07898745685815811 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, full-length photograph depicts a stoic female warrior amidst a tranquil, misty bamboo forest, the entire scene rendered in the elegant, fluid lines of traditional Chinese art. Her face is a mask of cold fury, her eyes narrowed to a deadly focus, her lips pressed into a thin, determined line. Her sleek black hair is curled into an intricate headdress, secured with a simple jade hairpin. She clads herself in a crimson silk robe that billows around her, its movement and intricate cloud-patterned embroidery captured with delicate calligraphic brushstrokes. As an arrow suddenly pierces the frame, she grasps a dark lacquer war fan in her right hand, and with a swift and economical flick, deftly deflecting the arrow. The movement is a blur of controlled grace; the fan's edge precisely connects to the arrowhead, producing a sharp flick of light that momentarily illuminates the scene before the arrow is sent tumbling away. The force of her movements causes her robes to swirl with painterly grace, and her furious expression does not shift as she maintains a powerful, grounded stance, creating an atmosphere of breathtakingly serene beauty juxtaposed with deadly, focused intent.", + "seed": 9955, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_10/index_5160_seed_9955_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0308455690741539, + "vlm_logprob": -0.0308455690741539, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.0308455690741539, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009933797642588615 + ], + "opposite_logprob_tokens": [ + -5.009933948516846 + ], + "logprob_reward": -0.009933797642588615, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009933797642588615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3043416440486908, + 0.059590552002191544, + 0.04368928447365761, + 0.0729515552520752, + 0.03855696693062782, + 0.09264810383319855, + 0.08983026444911957, + 0.09313973039388657, + 0.1618737280368805, + 0.043378107249736786 + ], + "credit_map": [ + 0.07026489078998566, + 0.039333876222372055, + 0.038434017449617386, + 0.024431027472019196, + 0.0229207593947649, + 0.07355934381484985, + 0.016564343124628067, + 0.018833408132195473, + 0.00741725554689765, + 0.005817364901304245, + 0.005880501586943865, + 0.007349066901952028, + 0.009415877051651478, + 0.0067464131861925125, + 0.008306892588734627, + 0.008657176047563553, + 0.00875235628336668, + 0.005512218456715345, + 0.0056753926910459995, + 0.0034831282682716846, + 0.0029699152801185846, + 0.008569513447582722, + 0.0035243211314082146, + 0.005202436354011297, + 0.010442303493618965, + 0.011726300232112408, + 0.011989875696599483, + 0.007093373686075211, + 0.006642879918217659, + 0.010512690991163254, + 0.005765937268733978, + 0.008778189308941364, + 0.005237600300461054, + 0.0038930349983274937, + 0.003528210800141096, + 0.005545428488403559, + 0.004843658301979303, + 0.005310303531587124, + 0.004171052947640419, + 0.006027677096426487, + 0.014334031380712986, + 0.010707039386034012, + 0.015485511161386967, + 0.009735926985740662, + 0.011783747933804989, + 0.013471203856170177, + 0.008229614235460758, + 0.008901027962565422, + 0.009667180478572845, + 0.007667019497603178, + 0.012895398773252964, + 0.01237246859818697, + 0.006907113362103701, + 0.0137330899015069, + 0.018469886854290962, + 0.008118108846247196, + 0.021852271631360054, + 0.009619608521461487, + 0.010257326997816563, + 0.00402575358748436, + 0.0120636485517025, + 0.022623024880886078, + 0.00442179199308157, + 0.008276302367448807, + 0.019254589453339577, + 0.022966057062149048, + 0.021486474201083183, + 0.03265474736690521, + 0.01607498899102211, + 0.01787562482059002, + 0.018087051808834076, + 0.013474204577505589, + 0.005847002845257521, + 0.00807215366512537, + 0.004151858389377594, + 0.005368748214095831, + 0.005136628169566393, + 0.0058973971754312515, + 0.004987895954400301, + 0.003916422836482525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.009933797642588615 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200353994965553 + ], + "opposite_logprob_tokens": [ + -4.264200210571289 + ], + "logprob_reward": -0.014200353994965553, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200353994965553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39476245641708374, + 0.02645130269229412, + 0.02419617958366871, + 0.028432007879018784, + 0.015119270421564579, + 0.021504634991288185, + 0.04342266544699669, + 0.4072090685367584, + 0.02929900772869587, + 0.009603355079889297 + ], + "credit_map": [ + 0.09953968226909637, + 0.04694470390677452, + 0.05935428664088249, + 0.02122429944574833, + 0.026907607913017273, + 0.10266879200935364, + 0.015336617827415466, + 0.022786516696214676, + 0.004781519528478384, + 0.0030561003368347883, + 0.002995101734995842, + 0.0041220043785870075, + 0.002054594224318862, + 0.003708089701831341, + 0.0024539686273783445, + 0.003279926488175988, + 0.00425859447568655, + 0.0028500347398221493, + 0.0047860885970294476, + 0.001761462539434433, + 0.0018865583697333932, + 0.005677275825291872, + 0.001299843192100525, + 0.0016763248713687062, + 0.006816709879785776, + 0.005215947981923819, + 0.003962724469602108, + 0.0022528336849063635, + 0.0025062093045562506, + 0.0041669877246022224, + 0.0016543937381356955, + 0.0018562065670266747, + 0.0028122367803007364, + 0.0016642450354993343, + 0.0034847771748900414, + 0.001528026768937707, + 0.0012950313976034522, + 0.0023846651893109083, + 0.0008676687139086425, + 0.001082621980458498, + 0.003730611177161336, + 0.0028668250888586044, + 0.0028666469734162092, + 0.0023514111526310444, + 0.003379428293555975, + 0.0027526402845978737, + 0.0018458929844200611, + 0.0017111824126914144, + 0.007512826472520828, + 0.004306240938603878, + 0.00552564999088645, + 0.0021239514462649822, + 0.0028071384876966476, + 0.011350051499903202, + 0.0033775577321648598, + 0.006419257260859013, + 0.09828082472085953, + 0.04194795340299606, + 0.047635775059461594, + 0.01712508127093315, + 0.04694165661931038, + 0.10099054127931595, + 0.019604062661528587, + 0.03468320146203041, + 0.007736715022474527, + 0.003242199309170246, + 0.004701863043010235, + 0.0030872204806655645, + 0.0024371652398258448, + 0.004289214499294758, + 0.0017980479169636965, + 0.0020065882708877325, + 0.0013548582792282104, + 0.0010998643701896071, + 0.0011464909184724092, + 0.0013570378068834543, + 0.0010106326080858707, + 0.0012190343113616109, + 0.001175123848952353, + 0.0012403145665302873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.014200353994965553 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370881825685501 + ], + "opposite_logprob_tokens": [ + -3.4087088108062744 + ], + "logprob_reward": -0.03370881825685501, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370881825685501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3025279939174652, + 0.1385185718536377, + 0.06046835705637932, + 0.05069255828857422, + 0.08055920153856277, + 0.07332960516214371, + 0.050215546041727066, + 0.0688178539276123, + 0.12711811065673828, + 0.04775221645832062 + ], + "credit_map": [ + 0.06496111303567886, + 0.034496650099754333, + 0.04608309641480446, + 0.027213776484131813, + 0.022316616028547287, + 0.07020479440689087, + 0.015781598165631294, + 0.021470334380865097, + 0.022334763780236244, + 0.015186944045126438, + 0.015996817499399185, + 0.022277990356087685, + 0.015003282576799393, + 0.017536349594593048, + 0.012113362550735474, + 0.018069058656692505, + 0.010455744341015816, + 0.006792168132960796, + 0.007570542395114899, + 0.006000102963298559, + 0.0059309182688593864, + 0.010131715796887875, + 0.004666335415095091, + 0.008920827880501747, + 0.00862579420208931, + 0.007402999792248011, + 0.0058912839740514755, + 0.0045176902785897255, + 0.005572747439146042, + 0.008793587796390057, + 0.004024370573461056, + 0.0058640833012759686, + 0.01806478388607502, + 0.009792913682758808, + 0.009123589843511581, + 0.007867179811000824, + 0.008007233962416649, + 0.012072340585291386, + 0.006727148313075304, + 0.00890401192009449, + 0.009867759421467781, + 0.009077920578420162, + 0.007680542767047882, + 0.00948979053646326, + 0.012373058125376701, + 0.011268571019172668, + 0.007626132108271122, + 0.00594582874327898, + 0.008905166760087013, + 0.005830869078636169, + 0.005914883688092232, + 0.00668077589944005, + 0.004047642461955547, + 0.006314815953373909, + 0.007745313458144665, + 0.004776079673320055, + 0.012394042685627937, + 0.007932478561997414, + 0.007816535420715809, + 0.004129353445023298, + 0.008619438856840134, + 0.014820928685367107, + 0.005609814077615738, + 0.0074952649883925915, + 0.016305534169077873, + 0.018035108223557472, + 0.015943720936775208, + 0.024198483675718307, + 0.012220893055200577, + 0.01456918939948082, + 0.014425636269152164, + 0.011419553309679031, + 0.012170564383268356, + 0.007549802307039499, + 0.004831704311072826, + 0.004304586909711361, + 0.005213882774114609, + 0.005959878675639629, + 0.0036616004072129726, + 0.0040601929649710655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.03370881825685501 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026353277266025543 + ], + "opposite_logprob_tokens": [ + -3.651353359222412 + ], + "logprob_reward": -0.026353277266025543, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026353277266025543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21799753606319427, + 0.035795655101537704, + 0.05417682230472565, + 0.054926905781030655, + 0.034864336252212524, + 0.20098021626472473, + 0.03257400542497635, + 0.07378674298524857, + 0.27687159180641174, + 0.018026186153292656 + ], + "credit_map": [ + 0.047777481377124786, + 0.023797454312443733, + 0.03229076787829399, + 0.013602769002318382, + 0.01637132652103901, + 0.05664388835430145, + 0.011307787150144577, + 0.016206037253141403, + 0.005253023933619261, + 0.0032749553211033344, + 0.003699881723150611, + 0.005418237764388323, + 0.005175233352929354, + 0.004371159244328737, + 0.004867067560553551, + 0.00373609340749681, + 0.01209364365786314, + 0.006535446736961603, + 0.00706437835469842, + 0.003710487624630332, + 0.004973633214831352, + 0.013247587718069553, + 0.0029337042942643166, + 0.0036179327871650457, + 0.00728157814592123, + 0.005824860651046038, + 0.007267525419592857, + 0.0066819218918681145, + 0.004788396880030632, + 0.008993247523903847, + 0.005501207429915667, + 0.00858816597610712, + 0.004927461966872215, + 0.004734495654702187, + 0.0038042261730879545, + 0.00558457663282752, + 0.004883299581706524, + 0.00461970642209053, + 0.0036804373376071453, + 0.002630130387842655, + 0.020322345197200775, + 0.028033334761857986, + 0.023479526862502098, + 0.022687923163175583, + 0.03730824962258339, + 0.02978833019733429, + 0.021621428430080414, + 0.017739061266183853, + 0.004118986427783966, + 0.0031441340688616037, + 0.0028048669919371605, + 0.004403835628181696, + 0.0025550099089741707, + 0.005311290733516216, + 0.005558110307902098, + 0.004677767865359783, + 0.017062153667211533, + 0.007789670489728451, + 0.00892581045627594, + 0.004187872167676687, + 0.008662546053528786, + 0.01637553796172142, + 0.0047097159549593925, + 0.00607343390583992, + 0.03166109323501587, + 0.0346900038421154, + 0.04027920588850975, + 0.056055452674627304, + 0.02849554643034935, + 0.027905739843845367, + 0.0335681326687336, + 0.024216394871473312, + 0.0018622319912537932, + 0.002571849850937724, + 0.002015444217249751, + 0.0025660444516688585, + 0.0021509574726223946, + 0.0020102602429687977, + 0.0024001284036785364, + 0.0024492673110216856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.026353277266025543 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003159821033478 + ], + "opposite_logprob_tokens": [ + -2.6950316429138184 + ], + "logprob_reward": -0.07003159821033478, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003159821033478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14410343766212463, + 0.07825914770364761, + 0.031074615195393562, + 0.09598474949598312, + 0.03731656074523926, + 0.06967813521623611, + 0.10850239545106888, + 0.12324643135070801, + 0.2751884460449219, + 0.03664609417319298 + ], + "credit_map": [ + 0.027225712314248085, + 0.012850865721702576, + 0.023133190348744392, + 0.015484019182622433, + 0.01638549566268921, + 0.027106873691082, + 0.009916095063090324, + 0.012001200579106808, + 0.010582508519291878, + 0.009570772759616375, + 0.006680442485958338, + 0.010366316884756088, + 0.0113192955031991, + 0.01123009342700243, + 0.008983809500932693, + 0.009525913745164871, + 0.004051883239299059, + 0.0034092338755726814, + 0.0061681875959038734, + 0.003052875166758895, + 0.002971131354570389, + 0.00621367571875453, + 0.0024544219486415386, + 0.002753206994384527, + 0.013224785216152668, + 0.01180131547152996, + 0.011099879629909992, + 0.010918648913502693, + 0.012456266209483147, + 0.015296608209609985, + 0.009895279072225094, + 0.011291969567537308, + 0.008372415788471699, + 0.0038300296291708946, + 0.004295650403946638, + 0.004633442033082247, + 0.0029628335032612085, + 0.005688580218702555, + 0.003281231503933668, + 0.00425237650051713, + 0.006551701109856367, + 0.0073778401128947735, + 0.00962519459426403, + 0.012441206723451614, + 0.009465294890105724, + 0.008259355090558529, + 0.008111025206744671, + 0.00784651841968298, + 0.011608869768679142, + 0.010489419102668762, + 0.011411153711378574, + 0.018124772235751152, + 0.012035843916237354, + 0.010796510614454746, + 0.022540543228387833, + 0.01149529218673706, + 0.024776170030236244, + 0.012414291501045227, + 0.014136657118797302, + 0.006795932538807392, + 0.015795374289155006, + 0.026586811989545822, + 0.008476918563246727, + 0.014264268800616264, + 0.030380791053175926, + 0.03533671423792839, + 0.03373230621218681, + 0.05817197263240814, + 0.028487181290984154, + 0.029340369626879692, + 0.034170400351285934, + 0.02556869015097618, + 0.005776118487119675, + 0.004137866664677858, + 0.0037614903412759304, + 0.004044420551508665, + 0.005038703791797161, + 0.004198758862912655, + 0.004763216711580753, + 0.004925519227981567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.07003159821033478 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, full-length photograph depicts a stoic female warrior amidst a tranquil, misty bamboo forest, the entire scene rendered in the elegant, fluid lines of traditional Chinese art. Her face is a mask of cold fury, her eyes narrowed to a deadly focus, her lips pressed into a thin, determined line. Her sleek black hair is curled into an intricate headdress, secured with a simple jade hairpin. She clads herself in a crimson silk robe that billows around her, its movement and intricate cloud-patterned embroidery captured with delicate calligraphic brushstrokes. As an arrow suddenly pierces the frame, she grasps a dark lacquer war fan in her right hand, and with a swift and economical flick, deftly deflecting the arrow. The movement is a blur of controlled grace; the fan's edge precisely connects to the arrowhead, producing a sharp flick of light that momentarily illuminates the scene before the arrow is sent tumbling away. The force of her movements causes her robes to swirl with painterly grace, and her furious expression does not shift as she maintains a powerful, grounded stance, creating an atmosphere of breathtakingly serene beauty juxtaposed with deadly, focused intent.", + "seed": 9956, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_10/index_5160_seed_9956_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04591590315103531, + "vlm_logprob": -0.04591590315103531, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.04591590315103531, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01357855275273323 + ], + "opposite_logprob_tokens": [ + -4.513578414916992 + ], + "logprob_reward": -0.01357855275273323, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01357855275273323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08981630951166153, + 0.03362935036420822, + 0.07898467034101486, + 0.006790607236325741, + 0.007111734244972467, + 0.5870199799537659, + 0.014553014189004898, + 0.024432379752397537, + 0.1452813595533371, + 0.012380639091134071 + ], + "credit_map": [ + 0.016818825155496597, + 0.009225222282111645, + 0.01150829903781414, + 0.010780774988234043, + 0.0114591671153903, + 0.012770522385835648, + 0.00872296653687954, + 0.008530535735189915, + 0.006644973997026682, + 0.0037320542614907026, + 0.004398444201797247, + 0.004105926491320133, + 0.0033892926294356585, + 0.005028793588280678, + 0.003053749445825815, + 0.003276115981861949, + 0.027485357597470284, + 0.004279589280486107, + 0.0027473028749227524, + 0.00240135146304965, + 0.004165458958595991, + 0.031065691262483597, + 0.002846317831426859, + 0.003993598278611898, + 0.0012212703004479408, + 0.0008238711743615568, + 0.0008333437726832926, + 0.0009165148949250579, + 0.0009068158688023686, + 0.0010088818380609155, + 0.00046658594510518014, + 0.0006133232964202762, + 0.0016742144944146276, + 0.0006893053650856018, + 0.0008031533798202872, + 0.000882025109604001, + 0.000791077793110162, + 0.0009842471918091178, + 0.000634947675280273, + 0.0006527630030177534, + 0.23934948444366455, + 0.022207798436284065, + 0.016068128868937492, + 0.007627959828823805, + 0.012593962252140045, + 0.2557762861251831, + 0.00920482911169529, + 0.024191543459892273, + 0.004544675815850496, + 0.0011983937583863735, + 0.0014993221266195178, + 0.0013676222879439592, + 0.0011421828530728817, + 0.002442795317620039, + 0.0009734313352964818, + 0.0013845899375155568, + 0.005393257364630699, + 0.002186278346925974, + 0.002947421045973897, + 0.0030733589082956314, + 0.0021314951591193676, + 0.0034214225597679615, + 0.0022479654289782047, + 0.003031182335689664, + 0.048821549862623215, + 0.014585149474442005, + 0.016081253066658974, + 0.010118101723492146, + 0.011303572915494442, + 0.02663169428706169, + 0.007337633520364761, + 0.010402403771877289, + 0.0020902405958622694, + 0.0013952034059911966, + 0.0016662291018292308, + 0.0014497735537588596, + 0.0015243426896631718, + 0.0014185447944328189, + 0.0012046357151120901, + 0.0016316692344844341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.01357855275273323 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023308921605348587 + ], + "opposite_logprob_tokens": [ + -3.7733089923858643 + ], + "logprob_reward": -0.023308921605348587, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023308921605348587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13002365827560425, + 0.13156457245349884, + 0.34566113352775574, + 0.018479134887456894, + 0.02159305289387703, + 0.12238440662622452, + 0.07469046115875244, + 0.010697437450289726, + 0.13296733796596527, + 0.011938771232962608 + ], + "credit_map": [ + 0.029517965391278267, + 0.011288423091173172, + 0.01640748605132103, + 0.014140774495899677, + 0.01004227064549923, + 0.020651642233133316, + 0.011026282794773579, + 0.016948798671364784, + 0.030881036072969437, + 0.010974044911563396, + 0.01832418330013752, + 0.016366269439458847, + 0.00839476753026247, + 0.019754920154809952, + 0.010149401612579823, + 0.016719946637749672, + 0.13409380614757538, + 0.01338509563356638, + 0.009747538715600967, + 0.005719234701246023, + 0.010136427357792854, + 0.14802983403205872, + 0.0060482448898255825, + 0.018500924110412598, + 0.0034745170269161463, + 0.001682616537436843, + 0.0024148845113813877, + 0.0018358764937147498, + 0.001246110419742763, + 0.0026182096917182207, + 0.0016765277832746506, + 0.003530388930812478, + 0.0045804837718605995, + 0.0023170080967247486, + 0.0028994001913815737, + 0.0028876811265945435, + 0.0015247289557009935, + 0.0032400484196841717, + 0.0016948445700109005, + 0.0024488535709679127, + 0.049512892961502075, + 0.0047770896926522255, + 0.0036079376004636288, + 0.0017761216731742024, + 0.0026107870507985353, + 0.052997589111328125, + 0.002053633099421859, + 0.005048347637057304, + 0.02447323687374592, + 0.006834296975284815, + 0.008615918457508087, + 0.006831803824752569, + 0.004949918016791344, + 0.012593052349984646, + 0.0045935967937111855, + 0.005798637866973877, + 0.0023505855351686478, + 0.000956868811044842, + 0.001419315580278635, + 0.0012468447675928473, + 0.0007676437962800264, + 0.0013082290533930063, + 0.0012007005279883742, + 0.0014472485054284334, + 0.04104766622185707, + 0.012931442819535732, + 0.01704481616616249, + 0.01094083208590746, + 0.010618113912642002, + 0.022750303149223328, + 0.007339098956435919, + 0.010295046493411064, + 0.0019283242290839553, + 0.0011609989451244473, + 0.0019410959212109447, + 0.0012210110435262322, + 0.0011421556118875742, + 0.0018713953904807568, + 0.001065487158484757, + 0.0016083007212728262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 1 + ], + "selected_reward": -0.023308921605348587 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496591702103615 + ], + "opposite_logprob_tokens": [ + -2.9299659729003906 + ], + "logprob_reward": -0.05496591702103615, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496591702103615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03001989610493183, + 0.01585899293422699, + 0.22678741812705994, + 0.006986444815993309, + 0.009715641848742962, + 0.6042776107788086, + 0.007296974770724773, + 0.02465924806892872, + 0.0662364587187767, + 0.008161320351064205 + ], + "credit_map": [ + 0.00622919388115406, + 0.0028230485040694475, + 0.0038762919139117002, + 0.0029607785400003195, + 0.0035170288756489754, + 0.004576109815388918, + 0.0031651475001126528, + 0.002872299402952194, + 0.003150427946820855, + 0.001585933961905539, + 0.0025498687755316496, + 0.0017988303443416953, + 0.0013224289286881685, + 0.00225054775364697, + 0.0014492693590000272, + 0.0017516877269372344, + 0.08816498517990112, + 0.008689558133482933, + 0.006486536003649235, + 0.0036273393779993057, + 0.006737804971635342, + 0.09756966680288315, + 0.003945212345570326, + 0.011566328816115856, + 0.0015773685881868005, + 0.0008041895925998688, + 0.000690967426635325, + 0.0004961960366927087, + 0.0009732254547998309, + 0.0009956220164895058, + 0.0005531434435397387, + 0.000895732780918479, + 0.002012562705203891, + 0.0008907282026484609, + 0.0010465745581313968, + 0.0016568152932450175, + 0.0009755032951943576, + 0.0012665834510698915, + 0.0008239113376475871, + 0.0010429637040942907, + 0.24615740776062012, + 0.02299901656806469, + 0.016660263761878014, + 0.007998938672244549, + 0.0129558639600873, + 0.26307734847068787, + 0.009431597776710987, + 0.024997208267450333, + 0.002107531065121293, + 0.0005917834350839257, + 0.0007220572442747653, + 0.0006600517081096768, + 0.0005364026874303818, + 0.0013419517781585455, + 0.000478769390610978, + 0.0008584277820773423, + 0.006895672529935837, + 0.0016685455339029431, + 0.004362417384982109, + 0.0032475327607244253, + 0.0011503181885927916, + 0.0038130355533212423, + 0.0016673306236043572, + 0.0018543977057561278, + 0.022047026082873344, + 0.006835040170699358, + 0.007351480424404144, + 0.004606191534548998, + 0.005411854479461908, + 0.011984935961663723, + 0.00334118097089231, + 0.004658750724047422, + 0.0011401831870898604, + 0.0008113420335575938, + 0.0016683490248396993, + 0.0010807295329868793, + 0.0005550965433940291, + 0.0011063721030950546, + 0.0007687910692766309, + 0.0010304577881470323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 8 + ], + "selected_reward": -0.05496591702103615 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890621468424797 + ], + "opposite_logprob_tokens": [ + -2.464062213897705 + ], + "logprob_reward": -0.0890621468424797, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890621468424797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03425354138016701, + 0.024688445031642914, + 0.028644446283578873, + 0.014065612107515335, + 0.013141626492142677, + 0.6987448334693909, + 0.01885259337723255, + 0.040274664759635925, + 0.043387267738580704, + 0.08394698798656464 + ], + "credit_map": [ + 0.0072205886244773865, + 0.0027163366321474314, + 0.00506429560482502, + 0.004039345774799585, + 0.0028659296222031116, + 0.004981364123523235, + 0.002677645068615675, + 0.004688034765422344, + 0.0061739180237054825, + 0.0021775553468614817, + 0.0034429202787578106, + 0.0030710746068507433, + 0.002036072313785553, + 0.0033688251860439777, + 0.0021220946218818426, + 0.002295984188094735, + 0.007817486301064491, + 0.001498765079304576, + 0.0032604187726974487, + 0.002680209232494235, + 0.0016933213919401169, + 0.008180754259228706, + 0.001809681998565793, + 0.0017038085497915745, + 0.002868789015337825, + 0.0018342273542657495, + 0.0015607746317982674, + 0.001339117530733347, + 0.0021454324014484882, + 0.001923603587783873, + 0.0013606275897473097, + 0.0010330406948924065, + 0.0023476548958569765, + 0.0012684400426223874, + 0.001596097368746996, + 0.0019615564960986376, + 0.00120653142221272, + 0.0017361383652314544, + 0.0012926616473123431, + 0.0017325457883998752, + 0.2847660481929779, + 0.02673766389489174, + 0.018952613696455956, + 0.009395948611199856, + 0.014889849349856377, + 0.304266095161438, + 0.0110761858522892, + 0.028660373762249947, + 0.0050890883430838585, + 0.001203710213303566, + 0.0022511605639010668, + 0.0020913949701935053, + 0.001435346552170813, + 0.0035134342033416033, + 0.0016156235942617059, + 0.0016528352862223983, + 0.009127171710133553, + 0.0031525571830570698, + 0.00712988618761301, + 0.004662888590246439, + 0.0032020462676882744, + 0.005747977178543806, + 0.003215101547539234, + 0.004037035629153252, + 0.011341170407831669, + 0.0039973268285393715, + 0.005960546433925629, + 0.004404452163726091, + 0.00434821005910635, + 0.006356241647154093, + 0.0031690294854342937, + 0.003810292575508356, + 0.012151715345680714, + 0.0073641384951770306, + 0.016619432717561722, + 0.011214821599423885, + 0.005682156886905432, + 0.01314743235707283, + 0.008250485174357891, + 0.009516805410385132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.0890621468424797 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866397753357887 + ], + "opposite_logprob_tokens": [ + -3.048664093017578 + ], + "logprob_reward": -0.04866397753357887, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866397753357887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033775653690099716, + 0.013039493933320045, + 0.17651337385177612, + 0.014588273130357265, + 0.007321057841181755, + 0.680987536907196, + 0.016339603811502457, + 0.015009154565632343, + 0.026564262807369232, + 0.01586158759891987 + ], + "credit_map": [ + 0.006452891509979963, + 0.0035041584633290768, + 0.003578346921131015, + 0.004043949767947197, + 0.004171631298959255, + 0.004617780447006226, + 0.003204480279237032, + 0.00420241616666317, + 0.002907036803662777, + 0.001150574185885489, + 0.0016973321326076984, + 0.0013369583757594228, + 0.0012214777525514364, + 0.002116925548762083, + 0.0010692591313272715, + 0.0015399303520098329, + 0.06676335632801056, + 0.007112364750355482, + 0.005415679421275854, + 0.003149184165522456, + 0.0070893787778913975, + 0.07474399358034134, + 0.0032349741086363792, + 0.009004445746541023, + 0.0030258879996836185, + 0.0016309595666825771, + 0.001497454009950161, + 0.0013328755740076303, + 0.0025174932088702917, + 0.0019673462957143784, + 0.0013728842604905367, + 0.0012433725642040372, + 0.0017808685079216957, + 0.0007826111977919936, + 0.0006982668419368565, + 0.0007275256211869419, + 0.0008680938626639545, + 0.001106170704588294, + 0.0006003938615322113, + 0.000757127592805773, + 0.2776634395122528, + 0.025909079238772392, + 0.01866644062101841, + 0.009549993090331554, + 0.014460163190960884, + 0.2962636947631836, + 0.010754176415503025, + 0.027720576152205467, + 0.0046213832683861256, + 0.0014324309304356575, + 0.0018216335447505116, + 0.0018162266351282597, + 0.0013849773677065969, + 0.0026244246400892735, + 0.0011664537014439702, + 0.0014720737235620618, + 0.0031821744050830603, + 0.0012877265689894557, + 0.002033477183431387, + 0.0015602760249748826, + 0.001403993577696383, + 0.0020915314089506865, + 0.001402423600666225, + 0.002047552727162838, + 0.007999726571142673, + 0.0024283609818667173, + 0.0030224774964153767, + 0.0024441543500870466, + 0.0021122267935425043, + 0.004454370122402906, + 0.0018449543276801705, + 0.002257994143292308, + 0.002212606370449066, + 0.0013210049364715815, + 0.003059113398194313, + 0.002079466823488474, + 0.001439471379853785, + 0.002220337511971593, + 0.001382793067023158, + 0.002146792830899358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.04866397753357887 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, full-length photograph depicts a stoic female warrior amidst a tranquil, misty bamboo forest, the entire scene rendered in the elegant, fluid lines of traditional Chinese art. Her face is a mask of cold fury, her eyes narrowed to a deadly focus, her lips pressed into a thin, determined line. Her sleek black hair is curled into an intricate headdress, secured with a simple jade hairpin. She clads herself in a crimson silk robe that billows around her, its movement and intricate cloud-patterned embroidery captured with delicate calligraphic brushstrokes. As an arrow suddenly pierces the frame, she grasps a dark lacquer war fan in her right hand, and with a swift and economical flick, deftly deflecting the arrow. The movement is a blur of controlled grace; the fan's edge precisely connects to the arrowhead, producing a sharp flick of light that momentarily illuminates the scene before the arrow is sent tumbling away. The force of her movements causes her robes to swirl with painterly grace, and her furious expression does not shift as she maintains a powerful, grounded stance, creating an atmosphere of breathtakingly serene beauty juxtaposed with deadly, focused intent.", + "seed": 9957, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_10/index_5160_seed_9957_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.040745290368795394, + "vlm_logprob": -0.040745290368795394, + "vlm_yes_no_margin": 3.3999999523162843, + "vlm_token_credit": -0.040745290368795394, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01800982467830181 + ], + "opposite_logprob_tokens": [ + -4.143009662628174 + ], + "logprob_reward": -0.01800982467830181, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01800982467830181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17910487949848175, + 0.25578969717025757, + 0.03580998256802559, + 0.18907593190670013, + 0.0683608204126358, + 0.11240756511688232, + 0.016880638897418976, + 0.03174858167767525, + 0.05734588950872421, + 0.05347604304552078 + ], + "credit_map": [ + 0.035373423248529434, + 0.02014094777405262, + 0.02606670744717121, + 0.0181011613458395, + 0.0161711648106575, + 0.024808872491121292, + 0.016571132466197014, + 0.021871469914913177, + 0.05407867580652237, + 0.044357504695653915, + 0.04107502102851868, + 0.018314074724912643, + 0.026156513020396233, + 0.03945380449295044, + 0.015689853578805923, + 0.016664238646626472, + 0.005251084920018911, + 0.003985194489359856, + 0.005974845960736275, + 0.003820901270955801, + 0.003382159164175391, + 0.005526251625269651, + 0.0028697692323476076, + 0.004999774973839521, + 0.019664853811264038, + 0.006591340526938438, + 0.06636688858270645, + 0.008934362791478634, + 0.009035966359078884, + 0.06319849193096161, + 0.007204664871096611, + 0.008079363964498043, + 0.009017275646328926, + 0.005914546083658934, + 0.008817028254270554, + 0.008205625228583813, + 0.008968925103545189, + 0.008117043413221836, + 0.007604110520333052, + 0.01171626802533865, + 0.01411612518131733, + 0.01302920002490282, + 0.011078535579144955, + 0.011439858004450798, + 0.022186249494552612, + 0.019321102648973465, + 0.011191236786544323, + 0.010045250877737999, + 0.0017348889959976077, + 0.0015592520358040929, + 0.001612774096429348, + 0.003369820537045598, + 0.002465385477989912, + 0.0018590990221127868, + 0.0024289065040647984, + 0.0018505124608054757, + 0.006610474083572626, + 0.002494042506441474, + 0.0051252045668661594, + 0.0019199613016098738, + 0.0022023366764187813, + 0.009225458838045597, + 0.0017358000623062253, + 0.0024353060871362686, + 0.017274996265769005, + 0.0026573240756988525, + 0.02044428512454033, + 0.00390754034742713, + 0.002764730714261532, + 0.004753429908305407, + 0.002999010030180216, + 0.0025445716455578804, + 0.007690714206546545, + 0.004701415076851845, + 0.0062270211055874825, + 0.005470122676342726, + 0.0060968706384301186, + 0.009112519212067127, + 0.006149303168058395, + 0.008028078824281693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.01800982467830181 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018191050738096237 + ], + "opposite_logprob_tokens": [ + -4.018190860748291 + ], + "logprob_reward": -0.018191050738096237, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018191050738096237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18862763047218323, + 0.20208851993083954, + 0.05100426450371742, + 0.05795248970389366, + 0.05063905939459801, + 0.032107461243867874, + 0.03846012055873871, + 0.23117204010486603, + 0.0826072171330452, + 0.06534121930599213 + ], + "credit_map": [ + 0.04022585228085518, + 0.023241885006427765, + 0.026779498904943466, + 0.01625056564807892, + 0.01610056310892105, + 0.02784472517669201, + 0.01613241620361805, + 0.022052118554711342, + 0.03872973471879959, + 0.04461026191711426, + 0.041901595890522, + 0.010956034064292908, + 0.014844127930700779, + 0.033259060233831406, + 0.008668024092912674, + 0.009119675494730473, + 0.010447739623486996, + 0.00809257011860609, + 0.008067913353443146, + 0.004389406647533178, + 0.004657802637666464, + 0.006001660600304604, + 0.0034995987080037594, + 0.005847573280334473, + 0.010118355043232441, + 0.005577694159001112, + 0.008201608434319496, + 0.004460638388991356, + 0.004279064014554024, + 0.015376164577901363, + 0.0041204351000487804, + 0.005818534642457962, + 0.00714636268094182, + 0.0047042593359947205, + 0.005721426103264093, + 0.006506371311843395, + 0.00634646276012063, + 0.006203223019838333, + 0.00694590900093317, + 0.007065046112984419, + 0.00731743685901165, + 0.003060410264879465, + 0.005158639512956142, + 0.00307743763551116, + 0.003325018798932433, + 0.0048582833260297775, + 0.0026785237714648247, + 0.0026317115407437086, + 0.00664873979985714, + 0.004275035113096237, + 0.004751407075673342, + 0.0048147994093596935, + 0.004705388564616442, + 0.004554944112896919, + 0.00421461695805192, + 0.004495192784816027, + 0.03586217388510704, + 0.014193005859851837, + 0.06817787885665894, + 0.011130951344966888, + 0.011867613531649113, + 0.0706394612789154, + 0.009152495302259922, + 0.010148459114134312, + 0.02168840914964676, + 0.004149902146309614, + 0.02555125392973423, + 0.006846771575510502, + 0.0044798036105930805, + 0.007605941034853458, + 0.006120299454778433, + 0.006164837628602982, + 0.00996541976928711, + 0.005479066167026758, + 0.008661678060889244, + 0.006857908796519041, + 0.006605471950024366, + 0.011624692939221859, + 0.006635456345975399, + 0.009511524811387062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.018191050738096237 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029813015833497047 + ], + "opposite_logprob_tokens": [ + -3.529813051223755 + ], + "logprob_reward": -0.029813015833497047, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029813015833497047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12331029027700424, + 0.1270294487476349, + 0.055110570043325424, + 0.23220142722129822, + 0.11806309968233109, + 0.046500928699970245, + 0.012283428572118282, + 0.024672407656908035, + 0.2220078706741333, + 0.038820531219244 + ], + "credit_map": [ + 0.027687696740031242, + 0.016315976157784462, + 0.017826450988650322, + 0.01110595278441906, + 0.011522411368787289, + 0.01747405342757702, + 0.009482408873736858, + 0.011895344592630863, + 0.02309434488415718, + 0.016242992132902145, + 0.018847564235329628, + 0.011030933819711208, + 0.01993733085691929, + 0.017862167209386826, + 0.01117414515465498, + 0.00883998442441225, + 0.00814055372029543, + 0.006157347001135349, + 0.005692626349627972, + 0.007405325770378113, + 0.008782872930169106, + 0.008327066898345947, + 0.004968020133674145, + 0.005636757239699364, + 0.02600890025496483, + 0.010847385041415691, + 0.06305590271949768, + 0.011750198900699615, + 0.009914344176650047, + 0.08605435490608215, + 0.010517537593841553, + 0.014052808284759521, + 0.016493430361151695, + 0.011880922131240368, + 0.01586286723613739, + 0.013834696263074875, + 0.013764886185526848, + 0.014929191209375858, + 0.01359728816896677, + 0.017699817195534706, + 0.0053568207658827305, + 0.004365916829556227, + 0.0049321637488901615, + 0.005711644422262907, + 0.0061957817524671555, + 0.007587312720716, + 0.005839136429131031, + 0.006512153428047895, + 0.0014815465547144413, + 0.0010470570996403694, + 0.001618000096641481, + 0.0023895499762147665, + 0.001416717073880136, + 0.0016214053612202406, + 0.001591266831383109, + 0.0011178854620084167, + 0.0034792227670550346, + 0.0021686868276447058, + 0.004308244213461876, + 0.002080546924844384, + 0.002664614235982299, + 0.004899952560663223, + 0.0025631184689700603, + 0.002508023055270314, + 0.0812729075551033, + 0.008784052915871143, + 0.08869041502475739, + 0.009175824001431465, + 0.004585111979395151, + 0.01851455122232437, + 0.004761721007525921, + 0.0062232790514826775, + 0.006015989929437637, + 0.0030774290207773447, + 0.004876103717833757, + 0.004349813796579838, + 0.005213781725615263, + 0.005294445436447859, + 0.004771859385073185, + 0.005221104249358177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 1 + ], + "selected_reward": -0.029813015833497047 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048665791749954224 + ], + "opposite_logprob_tokens": [ + -3.048665761947632 + ], + "logprob_reward": -0.048665791749954224, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048665791749954224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11358673125505447, + 0.048847708851099014, + 0.05651744082570076, + 0.18803580105304718, + 0.12014016509056091, + 0.03644924983382225, + 0.04163461551070213, + 0.08053335547447205, + 0.2718631625175476, + 0.042391784489154816 + ], + "credit_map": [ + 0.024827806279063225, + 0.016339777037501335, + 0.013005171902477741, + 0.01021094061434269, + 0.01471653115004301, + 0.01726263388991356, + 0.008708762004971504, + 0.00851510651409626, + 0.008225245401263237, + 0.0059471591375768185, + 0.006701704580336809, + 0.00397124607115984, + 0.007228789385408163, + 0.005953936371952295, + 0.005457771942019463, + 0.005361857824027538, + 0.009078328497707844, + 0.007748338393867016, + 0.007929056882858276, + 0.005526345688849688, + 0.0063202218152582645, + 0.010824061930179596, + 0.00396596547216177, + 0.00512512493878603, + 0.02572597935795784, + 0.011518272571265697, + 0.061731696128845215, + 0.011566662229597569, + 0.011445499025285244, + 0.042577601969242096, + 0.012540862895548344, + 0.010929230600595474, + 0.01638033427298069, + 0.009741727262735367, + 0.014865466393530369, + 0.016844235360622406, + 0.013121398165822029, + 0.015228825621306896, + 0.016636386513710022, + 0.017321797087788582, + 0.00447043776512146, + 0.003406885312870145, + 0.004937523510307074, + 0.004143924452364445, + 0.006954134441912174, + 0.006319814827293158, + 0.0034575103782117367, + 0.002759016817435622, + 0.005610487889498472, + 0.006345454603433609, + 0.003169437637552619, + 0.004684896674007177, + 0.008971348404884338, + 0.0048602609895169735, + 0.0037845904007554054, + 0.004208138212561607, + 0.006956696510314941, + 0.0028509385883808136, + 0.018107466399669647, + 0.0052248151041567326, + 0.00605854531750083, + 0.022378548979759216, + 0.007611365057528019, + 0.011344976723194122, + 0.09416387975215912, + 0.010212629102170467, + 0.10247471183538437, + 0.013683862052857876, + 0.008037671446800232, + 0.021241744980216026, + 0.010833804495632648, + 0.011214840225875378, + 0.006595725193619728, + 0.0041778297163546085, + 0.005869313608855009, + 0.0035947482101619244, + 0.005650077480822802, + 0.007557883858680725, + 0.004736347123980522, + 0.004209858365356922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 4 + ], + "selected_reward": -0.048665791749954224 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in the style of traditional Chinese art with elegant, fluid lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a tranquil, misty bamboo forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female warrior wearing a crimson silk robe with cloud-patterned embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior deflect an incoming arrow using a dark lacquer war fan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the warrior maintain a stoic, furious expression without shifting her face during the action?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904676884412766 + ], + "opposite_logprob_tokens": [ + -2.4640467166900635 + ], + "logprob_reward": -0.08904676884412766, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904676884412766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10196220129728317, + 0.051909249275922775, + 0.02670017071068287, + 0.34485507011413574, + 0.026820676401257515, + 0.030300727114081383, + 0.04071195423603058, + 0.09294527024030685, + 0.24958543479442596, + 0.0342092290520668 + ], + "credit_map": [ + 0.018537316471338272, + 0.016894491389393806, + 0.00847112387418747, + 0.01006812322884798, + 0.015196249820291996, + 0.013250615447759628, + 0.010396742261946201, + 0.009147544391453266, + 0.009535033255815506, + 0.005583588499575853, + 0.007588708307594061, + 0.004313359968364239, + 0.009689283557236195, + 0.006650941446423531, + 0.005402915179729462, + 0.0031454202253371477, + 0.003530841553583741, + 0.0026401448994874954, + 0.0023246416822075844, + 0.0026790343690663576, + 0.00641222158446908, + 0.003439022460952401, + 0.002643916057422757, + 0.003030348103493452, + 0.03584963455796242, + 0.010641349479556084, + 0.10959527641534805, + 0.01495172269642353, + 0.015467649325728416, + 0.13314609229564667, + 0.01215150486677885, + 0.013051842339336872, + 0.00335609819740057, + 0.0024387913290411234, + 0.0031780237331986427, + 0.0037388766650110483, + 0.003581112250685692, + 0.0030843662098050117, + 0.0034509410616010427, + 0.003992465790361166, + 0.0034700019750744104, + 0.0035239318385720253, + 0.002921339822933078, + 0.0031504910439252853, + 0.006657439284026623, + 0.005191562231630087, + 0.0030084499157965183, + 0.002377511700615287, + 0.003964259289205074, + 0.004152582958340645, + 0.005264721345156431, + 0.004252846352756023, + 0.006203957833349705, + 0.005110298749059439, + 0.0058464487083256245, + 0.005916835740208626, + 0.009766420349478722, + 0.0032633414957672358, + 0.022896675392985344, + 0.004112899303436279, + 0.006374055985361338, + 0.03341219574213028, + 0.005266775377094746, + 0.007852914743125439, + 0.09051661193370819, + 0.0105325011536479, + 0.09566865861415863, + 0.009937471710145473, + 0.006764581892639399, + 0.02122926525771618, + 0.008124620653688908, + 0.0068117305636405945, + 0.005977076478302479, + 0.0030608447268605232, + 0.004548605531454086, + 0.003252255031839013, + 0.004189697094261646, + 0.00534016964957118, + 0.0037390219513326883, + 0.004101558588445187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.08904676884412766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_11/index_tensor([5559, 5559])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_11/index_tensor([5559, 5559])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1977a0c852cddcf5e56cb7da24fdb4d0280be93f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_11/index_tensor([5559, 5559])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sleek, low-angle, single-shot tracking shot unfolds across a vast automobile manufacturing plant, rendered entirely in intense, high-contrast monochrome. A brand-new sedan, its aerodynamically flawless curves, its deep, obsidian-black paintwork reflecting the harsh yet majestic surroundings, has just emerged from the sparkling robotic assembly line. In the background, towering steel beams loom in the darkness, while welding stations in the distance intermittently erupt with showers of brilliant white sparks, each flash casting a fleeting, sharp shadow across the polished concrete floor. As its LED headlights silently illuminate, emitting two cool, piercing beams that pierce the gloom ahead, the wheels begin to turn calmly and steadily, and the vehicle glides forward with an almost solemn slowness. The camera, parallel to the vehicle, steadily advances, capturing how the glaring overhead lights cast flowing, dazzling highlights on the car's smooth surface, while the shadows beneath it stretch into a bottomless depth. The interplay of light and shadow across the surface of the car emphasizes the immense power of the heavy industrial environment, while also outlining the flawless contours of this new industrial product with a sculptural beauty, creating an atmosphere of power and awe.", + "seed": 23231, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_11/index_5559_seed_23231_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.962246409803629, + "vlm_logprob": -0.962246409803629, + "vlm_yes_no_margin": 1.175, + "vlm_token_credit": -0.962246409803629, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -4.3881144523620605 + ], + "opposite_logprob_tokens": [ + -0.013114506378769875 + ], + "logprob_reward": -4.3881144523620605, + "yes_no_margin_reward": -4.375, + "token_credit_reward": -4.3881144523620605, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20480568706989288, + 0.05612173676490784, + 0.032945360988378525, + 0.06136557087302208, + 0.17111071944236755, + 0.1669953167438507, + 0.06648451089859009, + 0.16587978601455688, + 0.046316977590322495, + 0.02797427400946617 + ], + "credit_map": [ + 0.016122635453939438, + 0.02041524462401867, + 0.03235768526792526, + 0.030386438593268394, + 0.021776722744107246, + 0.039702076464891434, + 0.021036440506577492, + 0.02300846204161644, + 0.008363688364624977, + 0.007424211595207453, + 0.006287307478487492, + 0.008211503736674786, + 0.0054807295091450214, + 0.005483424291014671, + 0.005824871826916933, + 0.00904600415378809, + 0.0020287323277443647, + 0.0045740338973701, + 0.003975621424615383, + 0.0045882719568908215, + 0.00720094284042716, + 0.0040730200707912445, + 0.003988274838775396, + 0.002516465727239847, + 0.0028406709898263216, + 0.002579075749963522, + 0.00819354597479105, + 0.009257964789867401, + 0.014898216351866722, + 0.005960093345493078, + 0.012603005394339561, + 0.0050330013036727905, + 0.007227279711514711, + 0.004251345060765743, + 0.013486393727362156, + 0.007979833520948887, + 0.04959013685584068, + 0.06890945136547089, + 0.011548632755875587, + 0.008117664605379105, + 0.012073795311152935, + 0.008269169367849827, + 0.028380239382386208, + 0.010507330298423767, + 0.01497520599514246, + 0.05738150700926781, + 0.013128446415066719, + 0.02227962762117386, + 0.0046706171706318855, + 0.003866329323500395, + 0.008017304353415966, + 0.004193924833089113, + 0.006770061794668436, + 0.022328920662403107, + 0.006792467087507248, + 0.009844888933002949, + 0.010857131332159042, + 0.010583669878542423, + 0.01559665147215128, + 0.010864505544304848, + 0.013055728748440742, + 0.06462699174880981, + 0.020381072536110878, + 0.0199140515178442, + 0.004772460553795099, + 0.005723617970943451, + 0.005593430250883102, + 0.004609516356140375, + 0.007945241406559944, + 0.005723637528717518, + 0.005534767638891935, + 0.006414314266294241, + 0.0036286702379584312, + 0.006245782133191824, + 0.002121663885191083, + 0.002963567152619362, + 0.004088534042239189, + 0.0026151540223509073, + 0.003892426146194339, + 0.002418478252366185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -4.3881144523620605 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28180450201034546 + ], + "opposite_logprob_tokens": [ + -1.4068045616149902 + ], + "logprob_reward": -0.28180450201034546, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28180450201034546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049808189272880554, + 0.048620037734508514, + 0.08810219913721085, + 0.2228373885154724, + 0.19256414473056793, + 0.27719470858573914, + 0.03442177549004555, + 0.05192078277468681, + 0.009536519646644592, + 0.024994246661663055 + ], + "credit_map": [ + 0.004521142225712538, + 0.005570235662162304, + 0.007930553518235683, + 0.006027759052813053, + 0.005525077227503061, + 0.00701892888173461, + 0.006854809354990721, + 0.006359686143696308, + 0.006050614640116692, + 0.006682585924863815, + 0.005502090789377689, + 0.0067052836529910564, + 0.007179279811680317, + 0.005592245142906904, + 0.006095640826970339, + 0.004812300205230713, + 0.0053487555123865604, + 0.016163373365998268, + 0.007505292072892189, + 0.012551370076835155, + 0.013440071605145931, + 0.007883463054895401, + 0.018135735765099525, + 0.007074148394167423, + 0.01090787909924984, + 0.010110579431056976, + 0.012197046540677547, + 0.04861581698060036, + 0.07056094706058502, + 0.023084422573447227, + 0.025205137208104134, + 0.022155577316880226, + 0.01035391353070736, + 0.00548345735296607, + 0.014963960275053978, + 0.007169060409069061, + 0.05331546068191528, + 0.07768158614635468, + 0.01441076397895813, + 0.009185964241623878, + 0.0316825807094574, + 0.014476670883595943, + 0.055576518177986145, + 0.01821945048868656, + 0.023521116003394127, + 0.08738137036561966, + 0.016546744853258133, + 0.029790271073579788, + 0.0027102031745016575, + 0.002293347381055355, + 0.004109415225684643, + 0.0025570131838321686, + 0.0034357192926108837, + 0.01079658791422844, + 0.003350031329318881, + 0.005169461481273174, + 0.003913986962288618, + 0.0028557535260915756, + 0.005110124126076698, + 0.0030149624217301607, + 0.0035169068723917007, + 0.021382251754403114, + 0.005955043714493513, + 0.006171758286654949, + 0.0009756213403306901, + 0.0013587441062554717, + 0.0013572524767369032, + 0.0009891473455354571, + 0.001349421450868249, + 0.001407305127941072, + 0.0010338900610804558, + 0.0010651383781805634, + 0.0027198006864637136, + 0.0032275666017085314, + 0.0025216401554644108, + 0.001960010966286063, + 0.0020296478178352118, + 0.0026381707284599543, + 0.005172334145754576, + 0.0047250771895051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.28180450201034546 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04321274161338806 + ], + "opposite_logprob_tokens": [ + -3.168212652206421 + ], + "logprob_reward": -0.04321274161338806, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04321274161338806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13902708888053894, + 0.06728286296129227, + 0.057078007608652115, + 0.04782375320792198, + 0.38294467329978943, + 0.07178002595901489, + 0.08435353636741638, + 0.04319704324007034, + 0.02569444850087166, + 0.08081856369972229 + ], + "credit_map": [ + 0.011889475397765636, + 0.014401025138795376, + 0.018731918185949326, + 0.025363700464367867, + 0.015856383368372917, + 0.025085845962166786, + 0.012339058332145214, + 0.015359661541879177, + 0.008074365556240082, + 0.009572134353220463, + 0.00764516182243824, + 0.009813344106078148, + 0.006782290525734425, + 0.007620618212968111, + 0.00822970736771822, + 0.00954523216933012, + 0.0032550415489822626, + 0.00929960422217846, + 0.0065797423012554646, + 0.008669044822454453, + 0.008246755227446556, + 0.00561535032466054, + 0.01141674816608429, + 0.003995716106146574, + 0.0021352695766836405, + 0.0024435948580503464, + 0.006644767243415117, + 0.006503656972199678, + 0.006656590383499861, + 0.0043781655840575695, + 0.013234020210802555, + 0.005827685352414846, + 0.01531202718615532, + 0.008283025585114956, + 0.030222773551940918, + 0.016549045220017433, + 0.1107860654592514, + 0.1567928045988083, + 0.028430035337805748, + 0.0165688619017601, + 0.007066132500767708, + 0.005096157547086477, + 0.010107699781656265, + 0.007741888053715229, + 0.008077169768512249, + 0.01937299780547619, + 0.005871528759598732, + 0.00844645220786333, + 0.0056605590507388115, + 0.004869986791163683, + 0.01041550561785698, + 0.004900375381112099, + 0.008010133169591427, + 0.029639404267072678, + 0.007890657521784306, + 0.012966911308467388, + 0.0034921527840197086, + 0.004018603824079037, + 0.004247511737048626, + 0.004100271034985781, + 0.00438813166692853, + 0.013086244463920593, + 0.005379031877964735, + 0.0044850935228168964, + 0.0021360618993639946, + 0.003112765261903405, + 0.00285992119461298, + 0.002420711098238826, + 0.0038690767250955105, + 0.0038873988669365644, + 0.003194303484633565, + 0.0042142062447965145, + 0.007037595380097628, + 0.007884310558438301, + 0.0077689639292657375, + 0.004977247677743435, + 0.006366054527461529, + 0.010683110915124416, + 0.016033411026000977, + 0.020067863166332245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.04321274161338806 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04311637952923775 + ], + "opposite_logprob_tokens": [ + -3.168116331100464 + ], + "logprob_reward": -0.04311637952923775, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04311637952923775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09179031848907471, + 0.05359819531440735, + 0.05025438591837883, + 0.057343658059835434, + 0.38567572832107544, + 0.1598142832517624, + 0.06522928923368454, + 0.10064606368541718, + 0.018409792333841324, + 0.017238279804587364 + ], + "credit_map": [ + 0.009877744130790234, + 0.01078425906598568, + 0.013994394801557064, + 0.013389887288212776, + 0.009866214357316494, + 0.014105957932770252, + 0.007181110791862011, + 0.012590750120580196, + 0.005505783017724752, + 0.008539401926100254, + 0.006474485155194998, + 0.010194039903581142, + 0.004621867556124926, + 0.005337053444236517, + 0.006261980626732111, + 0.006663581356406212, + 0.003483465639874339, + 0.008357996121048927, + 0.006183861754834652, + 0.009089462459087372, + 0.006453982554376125, + 0.005023600999265909, + 0.007726333569735289, + 0.003935679793357849, + 0.0031404735054820776, + 0.003053775755688548, + 0.004418160766363144, + 0.011437390930950642, + 0.019779521971940994, + 0.00656540272757411, + 0.005137808155268431, + 0.003811122151091695, + 0.01636422798037529, + 0.010096495971083641, + 0.03125933185219765, + 0.016584713011980057, + 0.11169122159481049, + 0.1553327441215515, + 0.026788292452692986, + 0.01755872182548046, + 0.011513067409396172, + 0.007396379951387644, + 0.03175277262926102, + 0.010256190784275532, + 0.012229444459080696, + 0.056504908949136734, + 0.010369687341153622, + 0.01979181170463562, + 0.004494093358516693, + 0.0038720774464309216, + 0.007826337590813637, + 0.003999487496912479, + 0.0059856753796339035, + 0.02289384789764881, + 0.005783532280474901, + 0.010374230332672596, + 0.007709593046456575, + 0.006152383517473936, + 0.010256991721689701, + 0.0064436341635882854, + 0.007040196564048529, + 0.03995095193386078, + 0.0118097560480237, + 0.011282551102340221, + 0.0015313772018998861, + 0.0023241424933075905, + 0.0016258088871836662, + 0.0015759755624458194, + 0.0018267069244757295, + 0.0021878068801015615, + 0.0024220384657382965, + 0.004915935453027487, + 0.002107979031279683, + 0.0029938514344394207, + 0.001514081726782024, + 0.0016254385700449347, + 0.0017448405269533396, + 0.0016192569164559245, + 0.002813629573211074, + 0.00281920051202178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.04311637952923775 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498397350311279 + ], + "opposite_logprob_tokens": [ + -2.9299840927124023 + ], + "logprob_reward": -0.05498397350311279, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498397350311279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11193551123142242, + 0.09720216691493988, + 0.13923080265522003, + 0.11349321156740189, + 0.13493016362190247, + 0.15477432310581207, + 0.13978154957294464, + 0.04725194722414017, + 0.03843224048614502, + 0.022968122735619545 + ], + "credit_map": [ + 0.011506508104503155, + 0.014159082435071468, + 0.01220766268670559, + 0.017673274502158165, + 0.015252511948347092, + 0.017188958823680878, + 0.012670721858739853, + 0.011276790872216225, + 0.012271972373127937, + 0.01258508674800396, + 0.01660129614174366, + 0.011310525238513947, + 0.01112195011228323, + 0.010238303802907467, + 0.013147632591426373, + 0.009925397112965584, + 0.009535428136587143, + 0.016545051708817482, + 0.024380141869187355, + 0.021788615733385086, + 0.021555786952376366, + 0.013355140574276447, + 0.02134060114622116, + 0.010730024427175522, + 0.004636919591575861, + 0.006752501241862774, + 0.008541974239051342, + 0.020177822560071945, + 0.027690758928656578, + 0.010871022939682007, + 0.02310255728662014, + 0.011719654314219952, + 0.007285204250365496, + 0.006369924172759056, + 0.010140557773411274, + 0.00805901549756527, + 0.03320583701133728, + 0.05072588846087456, + 0.012245970778167248, + 0.006897770334035158, + 0.01944630593061447, + 0.016063367947936058, + 0.012258527800440788, + 0.020172864198684692, + 0.026769377291202545, + 0.02232881262898445, + 0.018055999651551247, + 0.01967906579375267, + 0.009101076051592827, + 0.009166174568235874, + 0.01669885404407978, + 0.007319427561014891, + 0.012554764747619629, + 0.04920278862118721, + 0.013304456137120724, + 0.022433998063206673, + 0.003086920827627182, + 0.004277688916772604, + 0.003318140050396323, + 0.006440868601202965, + 0.005647709127515554, + 0.008352883160114288, + 0.010312343016266823, + 0.005815396085381508, + 0.003239949932321906, + 0.0033789295703172684, + 0.0041247145272791386, + 0.002785570453852415, + 0.003042299998924136, + 0.0037786210887134075, + 0.007825094275176525, + 0.010257061570882797, + 0.002849781885743141, + 0.0049436092376708984, + 0.003291349858045578, + 0.0026952580083161592, + 0.0021618534810841084, + 0.002234494313597679, + 0.002279071370139718, + 0.002512706443667412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.05498397350311279 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sleek, low-angle, single-shot tracking shot unfolds across a vast automobile manufacturing plant, rendered entirely in intense, high-contrast monochrome. A brand-new sedan, its aerodynamically flawless curves, its deep, obsidian-black paintwork reflecting the harsh yet majestic surroundings, has just emerged from the sparkling robotic assembly line. In the background, towering steel beams loom in the darkness, while welding stations in the distance intermittently erupt with showers of brilliant white sparks, each flash casting a fleeting, sharp shadow across the polished concrete floor. As its LED headlights silently illuminate, emitting two cool, piercing beams that pierce the gloom ahead, the wheels begin to turn calmly and steadily, and the vehicle glides forward with an almost solemn slowness. The camera, parallel to the vehicle, steadily advances, capturing how the glaring overhead lights cast flowing, dazzling highlights on the car's smooth surface, while the shadows beneath it stretch into a bottomless depth. The interplay of light and shadow across the surface of the car emphasizes the immense power of the heavy industrial environment, while also outlining the flawless contours of this new industrial product with a sculptural beauty, creating an atmosphere of power and awe.", + "seed": 23232, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_11/index_5559_seed_23232_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.5571759298443795, + "vlm_logprob": -0.5571759298443795, + "vlm_yes_no_margin": 2.1, + "vlm_token_credit": -0.5571759298443795, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.4650962352752686 + ], + "opposite_logprob_tokens": [ + -0.09009618312120438 + ], + "logprob_reward": -2.4650962352752686, + "yes_no_margin_reward": -2.375, + "token_credit_reward": -2.4650962352752686, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19645974040031433, + 0.051916707307100296, + 0.027575857937335968, + 0.026100942865014076, + 0.08895683288574219, + 0.3078334629535675, + 0.09549407660961151, + 0.09604357928037643, + 0.05919020622968674, + 0.050428539514541626 + ], + "credit_map": [ + 0.02734055556356907, + 0.016675230115652084, + 0.025295322760939598, + 0.016652407124638557, + 0.024281570687890053, + 0.016920920461416245, + 0.05155421420931816, + 0.017739545553922653, + 0.009038636460900307, + 0.004082732368260622, + 0.013475106097757816, + 0.004343872424215078, + 0.005201712250709534, + 0.0027777303475886583, + 0.00877173338085413, + 0.004225184675306082, + 0.0019761077128350735, + 0.0013886041706427932, + 0.0025408093351870775, + 0.0070228250697255135, + 0.0017317048041149974, + 0.0013512881705537438, + 0.00881823617964983, + 0.00274628191255033, + 0.002218252047896385, + 0.0037065385840833187, + 0.0023141184356063604, + 0.002564760623499751, + 0.004094458185136318, + 0.0028232880868017673, + 0.0038629351183772087, + 0.004516590852290392, + 0.005792917218059301, + 0.009627865627408028, + 0.006475864443928003, + 0.01883036270737648, + 0.014634158462285995, + 0.007732758764177561, + 0.013678316958248615, + 0.012184593826532364, + 0.015781845897436142, + 0.12332618236541748, + 0.010660951025784016, + 0.021546470001339912, + 0.034309983253479004, + 0.01264108344912529, + 0.017828624695539474, + 0.07173828035593033, + 0.00916950311511755, + 0.025852704420685768, + 0.006037377752363682, + 0.010461886413395405, + 0.007281572557985783, + 0.009614991955459118, + 0.00660359114408493, + 0.02047245018184185, + 0.01229952648282051, + 0.01026986725628376, + 0.006693604867905378, + 0.024128355085849762, + 0.008196964859962463, + 0.006728626787662506, + 0.015009310096502304, + 0.012717324309051037, + 0.005628758110105991, + 0.006281900219619274, + 0.005403051618486643, + 0.008165660314261913, + 0.005371051840484142, + 0.00495185237377882, + 0.01300442311912775, + 0.01038350723683834, + 0.0028626364655792713, + 0.0056302412413060665, + 0.004557437729090452, + 0.012321858666837215, + 0.0085335997864604, + 0.006296432577073574, + 0.005246984306722879, + 0.004979352466762066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -2.4650962352752686 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2523637115955353 + ], + "opposite_logprob_tokens": [ + -1.502363681793213 + ], + "logprob_reward": -0.2523637115955353, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2523637115955353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08667246997356415, + 0.058541517704725266, + 0.04426965117454529, + 0.07883109897375107, + 0.033920612186193466, + 0.16547247767448425, + 0.14717631042003632, + 0.03712647408246994, + 0.16318081319332123, + 0.18480858206748962 + ], + "credit_map": [ + 0.011148079298436642, + 0.012796210125088692, + 0.008647128008306026, + 0.008913060650229454, + 0.014507019892334938, + 0.007118141744285822, + 0.015405356884002686, + 0.008137471973896027, + 0.012697426602244377, + 0.002528867917135358, + 0.023066893219947815, + 0.0057985843159258366, + 0.00417824974283576, + 0.001909928978420794, + 0.004664659965783358, + 0.0036969128996133804, + 0.004343314561992884, + 0.0017020800150930882, + 0.00617947056889534, + 0.006385251879692078, + 0.002596186473965645, + 0.0019919706974178553, + 0.014607103541493416, + 0.0064642722718417645, + 0.00626308424398303, + 0.010034187696874142, + 0.005187620874494314, + 0.011412499472498894, + 0.01590993069112301, + 0.006712464150041342, + 0.013720045797526836, + 0.009591265581548214, + 0.0019524811068549752, + 0.002905245404690504, + 0.0021653955336660147, + 0.003885880811139941, + 0.009735139086842537, + 0.001858415431343019, + 0.005829045549035072, + 0.0055890073999762535, + 0.013415517285466194, + 0.033719077706336975, + 0.01737399585545063, + 0.019697442650794983, + 0.023341551423072815, + 0.02425750531256199, + 0.01383447926491499, + 0.019832905381917953, + 0.021297959610819817, + 0.03885740414261818, + 0.010532891377806664, + 0.012403572909533978, + 0.012338332831859589, + 0.015437123365700245, + 0.00921429879963398, + 0.027094727382063866, + 0.003644622163847089, + 0.003922920674085617, + 0.0027131817769259214, + 0.008972537703812122, + 0.003778198268264532, + 0.0029019806534051895, + 0.006804080680012703, + 0.004388954024761915, + 0.015896296128630638, + 0.009497364982962608, + 0.008445770479738712, + 0.016154581680893898, + 0.013235525228083134, + 0.011843757703900337, + 0.05160195380449295, + 0.036505553871393204, + 0.013549592345952988, + 0.013672919943928719, + 0.020381776615977287, + 0.022466273978352547, + 0.023174269124865532, + 0.02844114601612091, + 0.0327378548681736, + 0.030384747311472893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.2523637115955353 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014234671369194984 + ], + "opposite_logprob_tokens": [ + -4.26423454284668 + ], + "logprob_reward": -0.014234671369194984, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014234671369194984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19837135076522827, + 0.11422915756702423, + 0.01862471178174019, + 0.02451304718852043, + 0.08425851911306381, + 0.3347037434577942, + 0.07741932570934296, + 0.05974726378917694, + 0.03960902988910675, + 0.04852385073900223 + ], + "credit_map": [ + 0.026974521577358246, + 0.01608271524310112, + 0.024604767560958862, + 0.017529549077153206, + 0.02383820340037346, + 0.017182447016239166, + 0.05401495471596718, + 0.01814420148730278, + 0.02419348433613777, + 0.005829731933772564, + 0.04438553377985954, + 0.010738412849605083, + 0.007693295367062092, + 0.0038338913582265377, + 0.010855760425329208, + 0.006699042860418558, + 0.0018080344889312983, + 0.0009441902511753142, + 0.0021146656945347786, + 0.003322713542729616, + 0.0014625419862568378, + 0.001266570994630456, + 0.005183177534490824, + 0.002522818511351943, + 0.002152571687474847, + 0.0038247774355113506, + 0.0021480214782059193, + 0.0027022298891097307, + 0.003467106493189931, + 0.0032799008768051863, + 0.0028878350276499987, + 0.004050602670758963, + 0.005357743240892887, + 0.009201815351843834, + 0.006309088319540024, + 0.01793106086552143, + 0.013960950076580048, + 0.007043271325528622, + 0.013071922585368156, + 0.011382662691175938, + 0.01806231029331684, + 0.13223318755626678, + 0.011341150850057602, + 0.024643253535032272, + 0.03627095744013786, + 0.016116928309202194, + 0.01992560550570488, + 0.0761103555560112, + 0.010136536322534084, + 0.017043860629200935, + 0.006482632830739021, + 0.008780227042734623, + 0.007415009196847677, + 0.009706252254545689, + 0.006638871505856514, + 0.011215937323868275, + 0.007815541699528694, + 0.007494478020817041, + 0.0064107561483979225, + 0.009748652577400208, + 0.005733456928282976, + 0.0051046814769506454, + 0.010121784172952175, + 0.007317914627492428, + 0.005049973726272583, + 0.005648290738463402, + 0.003228411776944995, + 0.006917223799973726, + 0.00441773422062397, + 0.004857650958001614, + 0.005103112664073706, + 0.0043866317719221115, + 0.0024136665742844343, + 0.00555231561884284, + 0.004813237581402063, + 0.013175640255212784, + 0.00868883915245533, + 0.0056711891666054726, + 0.0043604630045592785, + 0.0038484991528093815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.014234671369194984 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016082679852843285 + ], + "opposite_logprob_tokens": [ + -4.141082763671875 + ], + "logprob_reward": -0.016082679852843285, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016082679852843285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13408216834068298, + 0.08085574209690094, + 0.22673574090003967, + 0.024516940116882324, + 0.06084699556231499, + 0.23940755426883698, + 0.06395581364631653, + 0.06707119196653366, + 0.051774557679891586, + 0.05075322836637497 + ], + "credit_map": [ + 0.01775323785841465, + 0.012996461242437363, + 0.01670176349580288, + 0.011754791252315044, + 0.015400649979710579, + 0.011661929078400135, + 0.03547488525509834, + 0.012338447384536266, + 0.017886510118842125, + 0.004800704307854176, + 0.030979467555880547, + 0.007489873096346855, + 0.005648779217153788, + 0.002940813312307, + 0.006734801921993494, + 0.004374793730676174, + 0.020345814526081085, + 0.019998693838715553, + 0.05967620760202408, + 0.020711947232484818, + 0.037657175213098526, + 0.012785742059350014, + 0.026918699964880943, + 0.02864144556224346, + 0.002162792021408677, + 0.0032285212073475122, + 0.002239489695057273, + 0.0024781757965683937, + 0.004097693599760532, + 0.0026793137658387423, + 0.003796679200604558, + 0.0038342743646353483, + 0.0038817273452878, + 0.006531554739922285, + 0.004939783364534378, + 0.012997006997466087, + 0.01011720672249794, + 0.004839109722524881, + 0.009510930627584457, + 0.008029675111174583, + 0.013878156431019306, + 0.09170957654714584, + 0.009349366649985313, + 0.02047305554151535, + 0.02650322951376438, + 0.011536571197211742, + 0.014447592198848724, + 0.05151001363992691, + 0.007512894459068775, + 0.015505925752222538, + 0.005046160891652107, + 0.007585976738482714, + 0.005143838003277779, + 0.007142446935176849, + 0.004929558839648962, + 0.011089014820754528, + 0.008394747041165829, + 0.008164318278431892, + 0.004693957045674324, + 0.016240565106272697, + 0.005663569085299969, + 0.004410140682011843, + 0.01055372878909111, + 0.00895016547292471, + 0.00494264205917716, + 0.007993221282958984, + 0.00493230065330863, + 0.008402128703892231, + 0.006869032979011536, + 0.006465915124863386, + 0.007285410538315773, + 0.00488390913233161, + 0.0026490511372685432, + 0.006478939671069384, + 0.004003455862402916, + 0.014398516155779362, + 0.007863590493798256, + 0.005492815747857094, + 0.005484295543283224, + 0.004382567945867777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.016082679852843285 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810235112905502 + ], + "opposite_logprob_tokens": [ + -3.288102388381958 + ], + "logprob_reward": -0.03810235112905502, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810235112905502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18020953238010406, + 0.06798730045557022, + 0.060900136828422546, + 0.02608540467917919, + 0.06442394852638245, + 0.26029589772224426, + 0.10163051635026932, + 0.08712908625602722, + 0.09720078855752945, + 0.05413735285401344 + ], + "credit_map": [ + 0.023126522079110146, + 0.017016181722283363, + 0.02013288624584675, + 0.01675623469054699, + 0.02179904282093048, + 0.01439808588474989, + 0.049378782510757446, + 0.017601804807782173, + 0.012883205898106098, + 0.004063168540596962, + 0.021845398470759392, + 0.0062978086061775684, + 0.005517445970326662, + 0.0030159803573042154, + 0.009601677767932415, + 0.004762617871165276, + 0.0058778151869773865, + 0.005044957622885704, + 0.015036812983453274, + 0.008555199950933456, + 0.009035935625433922, + 0.003349261824041605, + 0.006477512419223785, + 0.00752263655886054, + 0.0025243042036890984, + 0.003070343751460314, + 0.0020850731525570154, + 0.002975438255816698, + 0.0037372554652392864, + 0.003435680642724037, + 0.004012978170067072, + 0.004244332667440176, + 0.004116622731089592, + 0.008715899661183357, + 0.006333377677947283, + 0.012944836169481277, + 0.01000400260090828, + 0.004974090959876776, + 0.009886073879897594, + 0.007449048105627298, + 0.014945732429623604, + 0.10329480469226837, + 0.009614335373044014, + 0.01791542023420334, + 0.02943994291126728, + 0.013082843273878098, + 0.015378344804048538, + 0.056624483317136765, + 0.011273371987044811, + 0.02876489982008934, + 0.007015431299805641, + 0.008874528110027313, + 0.0078027634881436825, + 0.009833247400820255, + 0.0060375467874109745, + 0.022028734907507896, + 0.010303723625838757, + 0.00831263791769743, + 0.006901643704622984, + 0.016874823719263077, + 0.009716984815895557, + 0.008105688728392124, + 0.01591784879565239, + 0.010995735414326191, + 0.011833729222416878, + 0.012154556810855865, + 0.014442047104239464, + 0.016130467876791954, + 0.01308069285005331, + 0.01042359508574009, + 0.012398144230246544, + 0.006737553980201483, + 0.004130939021706581, + 0.004760236479341984, + 0.007718522567301989, + 0.00785615760833025, + 0.007784179877489805, + 0.007764163892716169, + 0.008354985155165195, + 0.005768171511590481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.03810235112905502 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sleek, low-angle, single-shot tracking shot unfolds across a vast automobile manufacturing plant, rendered entirely in intense, high-contrast monochrome. A brand-new sedan, its aerodynamically flawless curves, its deep, obsidian-black paintwork reflecting the harsh yet majestic surroundings, has just emerged from the sparkling robotic assembly line. In the background, towering steel beams loom in the darkness, while welding stations in the distance intermittently erupt with showers of brilliant white sparks, each flash casting a fleeting, sharp shadow across the polished concrete floor. As its LED headlights silently illuminate, emitting two cool, piercing beams that pierce the gloom ahead, the wheels begin to turn calmly and steadily, and the vehicle glides forward with an almost solemn slowness. The camera, parallel to the vehicle, steadily advances, capturing how the glaring overhead lights cast flowing, dazzling highlights on the car's smooth surface, while the shadows beneath it stretch into a bottomless depth. The interplay of light and shadow across the surface of the car emphasizes the immense power of the heavy industrial environment, while also outlining the flawless contours of this new industrial product with a sculptural beauty, creating an atmosphere of power and awe.", + "seed": 23233, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_11/index_5559_seed_23233_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.872364417463541, + "vlm_logprob": -0.872364417463541, + "vlm_yes_no_margin": 1.275, + "vlm_token_credit": -0.872364417463541, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -4.019273281097412 + ], + "opposite_logprob_tokens": [ + -0.019273389130830765 + ], + "logprob_reward": -4.019273281097412, + "yes_no_margin_reward": -4.0, + "token_credit_reward": -4.019273281097412, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2622578740119934, + 0.2271411418914795, + 0.10359249264001846, + 0.048360440880060196, + 0.1403205394744873, + 0.015199337154626846, + 0.09085530787706375, + 0.027447130531072617, + 0.036785759031772614, + 0.0480399988591671 + ], + "credit_map": [ + 0.05961011350154877, + 0.018595756962895393, + 0.03608201816678047, + 0.015541309490799904, + 0.02000638283789158, + 0.02813759259879589, + 0.019033556804060936, + 0.06525115668773651, + 0.0217355377972126, + 0.021427441388368607, + 0.02537977695465088, + 0.022035574540495872, + 0.04123126342892647, + 0.03440777584910393, + 0.024210328236222267, + 0.036713458597660065, + 0.015040231868624687, + 0.011198855005204678, + 0.015497536398470402, + 0.012529588304460049, + 0.008886418305337429, + 0.009708518162369728, + 0.022740937769412994, + 0.007990398444235325, + 0.005746270529925823, + 0.003616784466430545, + 0.004005370195955038, + 0.005173236597329378, + 0.005372755229473114, + 0.006050140596926212, + 0.01275874674320221, + 0.005637134425342083, + 0.012418225407600403, + 0.004898537881672382, + 0.012481242418289185, + 0.05823735520243645, + 0.004740601405501366, + 0.00910182949155569, + 0.033727772533893585, + 0.004714976996183395, + 0.001687519485130906, + 0.0014061961555853486, + 0.0016515664756298065, + 0.002015845850110054, + 0.0011553672375157475, + 0.002259282162413001, + 0.0037531547714024782, + 0.001270405249670148, + 0.010869871824979782, + 0.005223080050200224, + 0.008732223883271217, + 0.011058810167014599, + 0.005976460408419371, + 0.017206251621246338, + 0.01846596784889698, + 0.01332264393568039, + 0.002585807116702199, + 0.001942901057191193, + 0.002953809220343828, + 0.003752768272534013, + 0.0016501802019774914, + 0.0034574968740344048, + 0.004560947418212891, + 0.0065432204864919186, + 0.003942620474845171, + 0.0018525830237194896, + 0.004725337494164705, + 0.007903019897639751, + 0.0019189720042049885, + 0.003336140885949135, + 0.006543250288814306, + 0.006563833449035883, + 0.002987232292070985, + 0.0036030274350196123, + 0.007118314504623413, + 0.007928211241960526, + 0.00480074854567647, + 0.007670284248888493, + 0.00628974474966526, + 0.007642436772584915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -4.019273281097412 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18057525157928467 + ], + "opposite_logprob_tokens": [ + -1.8055752515792847 + ], + "logprob_reward": -0.18057525157928467, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18057525157928467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2305779755115509, + 0.08476501703262329, + 0.0724959596991539, + 0.10117022693157196, + 0.06035446375608444, + 0.02727041020989418, + 0.06082480400800705, + 0.13242535293102264, + 0.17663948237895966, + 0.05347628518939018 + ], + "credit_map": [ + 0.0618475116789341, + 0.01490766555070877, + 0.028767496347427368, + 0.009817947633564472, + 0.014089821837842464, + 0.02669757604598999, + 0.01751508004963398, + 0.05693485960364342, + 0.007998364977538586, + 0.007831530645489693, + 0.00902080349624157, + 0.0084187937900424, + 0.01844099909067154, + 0.011311705224215984, + 0.009652337059378624, + 0.01209048368036747, + 0.007773055694997311, + 0.008195863105356693, + 0.013866609893739223, + 0.007789289578795433, + 0.006107928231358528, + 0.005811092909425497, + 0.01824181340634823, + 0.004710302222520113, + 0.010408803820610046, + 0.0066133541986346245, + 0.01100900024175644, + 0.00908890925347805, + 0.01000850461423397, + 0.01563807763159275, + 0.02124197408556938, + 0.017161598429083824, + 0.007152271922677755, + 0.002443069824948907, + 0.012108029797673225, + 0.022675223648548126, + 0.0030060941353440285, + 0.005975327920168638, + 0.004690818954259157, + 0.0023036280181258917, + 0.0034754604566842318, + 0.0019923660438507795, + 0.0018877765396609902, + 0.003520312951877713, + 0.0034710087347775698, + 0.00556854996830225, + 0.005288909655064344, + 0.002066026907414198, + 0.007722458336502314, + 0.0036858932580798864, + 0.006648645736277103, + 0.00773878488689661, + 0.004264015704393387, + 0.012633273378014565, + 0.010243331082165241, + 0.007888400927186012, + 0.0026466124691069126, + 0.003958709537982941, + 0.0035902124363929033, + 0.054617367684841156, + 0.006256288383156061, + 0.002879211213439703, + 0.05147150903940201, + 0.007005438674241304, + 0.026466412469744682, + 0.01273629255592823, + 0.02962433360517025, + 0.028744759038090706, + 0.013015232980251312, + 0.017500899732112885, + 0.026087680831551552, + 0.022463886067271233, + 0.0033644691575318575, + 0.0028249716851860285, + 0.0077285016886889935, + 0.010493433102965355, + 0.005913374945521355, + 0.006240857765078545, + 0.007384160533547401, + 0.009526513516902924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.18057525157928467 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04318762570619583 + ], + "opposite_logprob_tokens": [ + -3.1681876182556152 + ], + "logprob_reward": -0.04318762570619583, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04318762570619583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2591591477394104, + 0.15446774661540985, + 0.04653814807534218, + 0.06828072667121887, + 0.30232301354408264, + 0.019823698326945305, + 0.06256816536188126, + 0.04175570234656334, + 0.024613983929157257, + 0.02046957053244114 + ], + "credit_map": [ + 0.06318244338035583, + 0.016591092571616173, + 0.03897859528660774, + 0.010965071618556976, + 0.015206371434032917, + 0.024736076593399048, + 0.018440719693899155, + 0.07105881720781326, + 0.01442013680934906, + 0.014324408024549484, + 0.017320241779088974, + 0.013916688039898872, + 0.027034422382712364, + 0.02617534063756466, + 0.015205916948616505, + 0.026070620864629745, + 0.006290358491241932, + 0.006210626568645239, + 0.005257037002593279, + 0.006532803177833557, + 0.005663247313350439, + 0.005579283460974693, + 0.006130150519311428, + 0.004874648526310921, + 0.0066958582028746605, + 0.00332748144865036, + 0.010598127730190754, + 0.006793771870434284, + 0.005961379501968622, + 0.005267920438200235, + 0.02553819864988327, + 0.00409799674525857, + 0.025952501222491264, + 0.010365208610892296, + 0.020894834771752357, + 0.12586060166358948, + 0.011136443354189396, + 0.0175111535936594, + 0.08020270615816116, + 0.010399607010185719, + 0.0023192870430648327, + 0.0019068366382271051, + 0.0024622022174298763, + 0.0022202215623110533, + 0.0017064433777704835, + 0.002419463125988841, + 0.005207989364862442, + 0.00158125685993582, + 0.0076181055046617985, + 0.003784861881285906, + 0.006133546587079763, + 0.007516013458371162, + 0.004110522568225861, + 0.011598422192037106, + 0.012281861156225204, + 0.009524841792881489, + 0.001978489337489009, + 0.0014918464003130794, + 0.0027870074845850468, + 0.014269432052969933, + 0.0019917581230401993, + 0.002301084576174617, + 0.012765595689415932, + 0.004170491360127926, + 0.00329839950427413, + 0.0016713379882276058, + 0.00387028930708766, + 0.004470909014344215, + 0.001874900539405644, + 0.0024711652658879757, + 0.0038437123876065016, + 0.0031132737640291452, + 0.0014180534053593874, + 0.002664861036464572, + 0.003526127664372325, + 0.002990479115396738, + 0.0019414258422330022, + 0.0028055375441908836, + 0.0020400332286953926, + 0.003083055140450597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.04318762570619583 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0487213172018528 + ], + "opposite_logprob_tokens": [ + -3.0487213134765625 + ], + "logprob_reward": -0.0487213172018528, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0487213172018528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13927459716796875, + 0.11181001365184784, + 0.04760945215821266, + 0.29775771498680115, + 0.19781044125556946, + 0.04349812865257263, + 0.056979212909936905, + 0.06008410453796387, + 0.020529529079794884, + 0.02464679442346096 + ], + "credit_map": [ + 0.030490344390273094, + 0.011187371797859669, + 0.019876638427376747, + 0.010133149102330208, + 0.010066858492791653, + 0.014968602918088436, + 0.011309281922876835, + 0.03124234266579151, + 0.010541488416492939, + 0.011518311686813831, + 0.012421301566064358, + 0.01056578103452921, + 0.017821816727519035, + 0.02037668041884899, + 0.010231933556497097, + 0.018332701176404953, + 0.007834583520889282, + 0.005706148222088814, + 0.0055266558192670345, + 0.004756317939609289, + 0.005177291575819254, + 0.006282466929405928, + 0.007967707701027393, + 0.004358280450105667, + 0.0259004607796669, + 0.009974883869290352, + 0.059389300644397736, + 0.026660824194550514, + 0.021755652502179146, + 0.018576882779598236, + 0.11907393485307693, + 0.016425777226686478, + 0.017213786020874977, + 0.006861511152237654, + 0.014158962294459343, + 0.08285150676965714, + 0.0069045270793139935, + 0.01144255232065916, + 0.0517086386680603, + 0.006668975111097097, + 0.005670937243849039, + 0.0025947142858058214, + 0.006562469527125359, + 0.006454261485487223, + 0.0029438540805131197, + 0.005242621060460806, + 0.011410156264901161, + 0.002619115635752678, + 0.007575178053230047, + 0.003963778726756573, + 0.006543668452650309, + 0.0072954148054122925, + 0.003689925419166684, + 0.010160095989704132, + 0.009719652123749256, + 0.008031499572098255, + 0.002340817591175437, + 0.0021208333782851696, + 0.003960629925131798, + 0.0218514297157526, + 0.0031774735543876886, + 0.003216775832697749, + 0.019742978736758232, + 0.003673166036605835, + 0.0017431703163310885, + 0.001260043471120298, + 0.002467549405992031, + 0.004364456981420517, + 0.0011687297374010086, + 0.0022612649481743574, + 0.0037796893157064915, + 0.003484623972326517, + 0.0015825628070160747, + 0.0019185898127034307, + 0.0043873111717402935, + 0.003686963813379407, + 0.002328726230189204, + 0.0043389094062149525, + 0.003166418056935072, + 0.0032373121939599514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.0487213172018528 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006461173295975 + ], + "opposite_logprob_tokens": [ + -2.6950645446777344 + ], + "logprob_reward": -0.07006461173295975, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006461173295975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14175905287265778, + 0.05489100515842438, + 0.04738754779100418, + 0.16187110543251038, + 0.1075996533036232, + 0.05001978203654289, + 0.041268959641456604, + 0.3155866265296936, + 0.0359603688120842, + 0.043655939400196075 + ], + "credit_map": [ + 0.0303699541836977, + 0.008586202748119831, + 0.02433900348842144, + 0.006081270053982735, + 0.008633924648165703, + 0.012174240313470364, + 0.00946782436221838, + 0.04210663214325905, + 0.005957101937383413, + 0.005027262959629297, + 0.006110871210694313, + 0.005156407132744789, + 0.007166776806116104, + 0.011530997231602669, + 0.004585697781294584, + 0.00935589149594307, + 0.00438241520896554, + 0.004534316714853048, + 0.009756187908351421, + 0.0061560203321278095, + 0.003952965606004, + 0.0034177880734205246, + 0.012242128141224384, + 0.0029457295313477516, + 0.013996711932122707, + 0.005740988999605179, + 0.034127723425626755, + 0.01454920880496502, + 0.011309172958135605, + 0.010011550039052963, + 0.06383539736270905, + 0.00830036774277687, + 0.010094523429870605, + 0.0033119504805654287, + 0.016040578484535217, + 0.016596348956227303, + 0.0068565309047698975, + 0.007823893800377846, + 0.0405048169195652, + 0.006371004041284323, + 0.006334317848086357, + 0.002331075957044959, + 0.006748565007001162, + 0.007715573534369469, + 0.003983314614742994, + 0.008337173610925674, + 0.011356999166309834, + 0.003212762763723731, + 0.004632741678506136, + 0.0023781582713127136, + 0.003499578684568405, + 0.0050274464301764965, + 0.0028016010764986277, + 0.006390729453414679, + 0.010475736111402512, + 0.006062967702746391, + 0.006976091768592596, + 0.009271752089262009, + 0.008233250118792057, + 0.13120175898075104, + 0.0153737748041749, + 0.006562275346368551, + 0.12496639788150787, + 0.013001332990825176, + 0.004116714932024479, + 0.0031143396627157927, + 0.00661812536418438, + 0.0069322786293923855, + 0.0026197307743132114, + 0.002837598789483309, + 0.005525879096239805, + 0.004195700399577618, + 0.00300150946713984, + 0.0023803221993148327, + 0.00438410509377718, + 0.00903936568647623, + 0.005318868905305862, + 0.00510747916996479, + 0.007203757297247648, + 0.007220529019832611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.07006461173295975 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sleek, low-angle, single-shot tracking shot unfolds across a vast automobile manufacturing plant, rendered entirely in intense, high-contrast monochrome. A brand-new sedan, its aerodynamically flawless curves, its deep, obsidian-black paintwork reflecting the harsh yet majestic surroundings, has just emerged from the sparkling robotic assembly line. In the background, towering steel beams loom in the darkness, while welding stations in the distance intermittently erupt with showers of brilliant white sparks, each flash casting a fleeting, sharp shadow across the polished concrete floor. As its LED headlights silently illuminate, emitting two cool, piercing beams that pierce the gloom ahead, the wheels begin to turn calmly and steadily, and the vehicle glides forward with an almost solemn slowness. The camera, parallel to the vehicle, steadily advances, capturing how the glaring overhead lights cast flowing, dazzling highlights on the car's smooth surface, while the shadows beneath it stretch into a bottomless depth. The interplay of light and shadow across the surface of the car emphasizes the immense power of the heavy industrial environment, while also outlining the flawless contours of this new industrial product with a sculptural beauty, creating an atmosphere of power and awe.", + "seed": 23234, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_11/index_5559_seed_23234_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.8703842349350452, + "vlm_logprob": -0.8703842349350452, + "vlm_yes_no_margin": 1.325, + "vlm_token_credit": -0.8703842349350452, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -4.018933296203613 + ], + "opposite_logprob_tokens": [ + -0.018933288753032684 + ], + "logprob_reward": -4.018933296203613, + "yes_no_margin_reward": -4.0, + "token_credit_reward": -4.018933296203613, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05922537297010422, + 0.030588604509830475, + 0.05493203178048134, + 0.01815143972635269, + 0.25345584750175476, + 0.045539408922195435, + 0.30346086621284485, + 0.02543792501091957, + 0.08330696821212769, + 0.12590152025222778 + ], + "credit_map": [ + 0.008743613958358765, + 0.009003506042063236, + 0.006000070367008448, + 0.006238070782274008, + 0.007206283044070005, + 0.008352834731340408, + 0.0066614095121622086, + 0.007019587326794863, + 0.0035687254276126623, + 0.0034387751948088408, + 0.002965952269732952, + 0.0035743890330195427, + 0.0027413207571953535, + 0.004051445052027702, + 0.004761585965752602, + 0.0054864115081727505, + 0.007602150551974773, + 0.0065855178982019424, + 0.005159369669854641, + 0.010193991474807262, + 0.005949472542852163, + 0.008507231250405312, + 0.006741748191416264, + 0.004192547872662544, + 0.0021080533042550087, + 0.0021844818256795406, + 0.0020274093840271235, + 0.002216203836724162, + 0.0024725785478949547, + 0.002754621207714081, + 0.002909403294324875, + 0.0014786871615797281, + 0.010705976746976376, + 0.012328309006989002, + 0.05378109589219093, + 0.033811260014772415, + 0.014029344543814659, + 0.03170941397547722, + 0.018973570317029953, + 0.07811687141656876, + 0.005294227972626686, + 0.006628307048231363, + 0.006669751834124327, + 0.007683978881686926, + 0.007930339314043522, + 0.00413105683401227, + 0.0037195412442088127, + 0.0034822053276002407, + 0.010972187854349613, + 0.013952281326055527, + 0.014403979294002056, + 0.11597681045532227, + 0.014949030242860317, + 0.012535970658063889, + 0.10940800607204437, + 0.011262619867920876, + 0.0033080943394452333, + 0.0023882135283201933, + 0.0038152795750647783, + 0.003575409296900034, + 0.0015441891737282276, + 0.004183328244835138, + 0.0031107692047953606, + 0.0035126430448144674, + 0.007643581368029118, + 0.006619441322982311, + 0.006169741041958332, + 0.015413457527756691, + 0.007903141900897026, + 0.008197961375117302, + 0.020288899540901184, + 0.011070744134485722, + 0.009337430819869041, + 0.008982894942164421, + 0.009498500265181065, + 0.02621634304523468, + 0.01670251041650772, + 0.006559162400662899, + 0.021777641028165817, + 0.02682703547179699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -4.018933296203613 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18039032816886902 + ], + "opposite_logprob_tokens": [ + -1.8053903579711914 + ], + "logprob_reward": -0.18039032816886902, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18039032816886902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03828302025794983, + 0.10597817599773407, + 0.41256582736968994, + 0.03905702009797096, + 0.08380689471960068, + 0.02617083303630352, + 0.050323810428380966, + 0.018053274601697922, + 0.05241706594824791, + 0.17334413528442383 + ], + "credit_map": [ + 0.005041478667408228, + 0.0041178325191140175, + 0.0038209366612136364, + 0.0034122501965612173, + 0.004886461887508631, + 0.005603755358606577, + 0.004943369887769222, + 0.006456933915615082, + 0.014895946718752384, + 0.022994374856352806, + 0.009291345253586769, + 0.009081648662686348, + 0.014264087192714214, + 0.012897099368274212, + 0.009941906668245792, + 0.012611780315637589, + 0.06184208020567894, + 0.049589674919843674, + 0.041559960693120956, + 0.07461446523666382, + 0.04568915814161301, + 0.06341052800416946, + 0.043457090854644775, + 0.0324028842151165, + 0.005697319749742746, + 0.004451934713870287, + 0.004152387380599976, + 0.004723529797047377, + 0.004861666355282068, + 0.004354261327534914, + 0.00630169827491045, + 0.004514222964644432, + 0.003744087414816022, + 0.004392614588141441, + 0.01927368715405464, + 0.009856919758021832, + 0.005397876724600792, + 0.01087175589054823, + 0.006181671284139156, + 0.024088287726044655, + 0.0032615847885608673, + 0.003865194506943226, + 0.0028069387190043926, + 0.0038176870439201593, + 0.006049461662769318, + 0.00226552109234035, + 0.002333638956770301, + 0.0017708046361804008, + 0.001896512578241527, + 0.0025617009960114956, + 0.003923958633095026, + 0.01878596469759941, + 0.0036163057666271925, + 0.0032063911203294992, + 0.014139319770038128, + 0.0021936590783298016, + 0.002552102552726865, + 0.0011959332041442394, + 0.0025154061149805784, + 0.0020780260674655437, + 0.0016772508388385177, + 0.003018198534846306, + 0.0021077445708215237, + 0.0029086109716445208, + 0.004773372784256935, + 0.00815537478774786, + 0.0068536195904016495, + 0.008306129835546017, + 0.00674572354182601, + 0.006222508382052183, + 0.005439744796603918, + 0.005920597817748785, + 0.013129614293575287, + 0.010329054668545723, + 0.012448592111468315, + 0.03795705363154411, + 0.01240709237754345, + 0.011396723799407482, + 0.0334242507815361, + 0.04225175082683563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.18039032816886902 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048782624304294586 + ], + "opposite_logprob_tokens": [ + -3.0487825870513916 + ], + "logprob_reward": -0.048782624304294586, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048782624304294586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059938594698905945, + 0.0784030333161354, + 0.04116231948137283, + 0.02677849866449833, + 0.09971828013658524, + 0.0421675443649292, + 0.44659730792045593, + 0.031002642586827278, + 0.06689497083425522, + 0.10733681172132492 + ], + "credit_map": [ + 0.0075078606605529785, + 0.006852787919342518, + 0.007055814377963543, + 0.005800278391689062, + 0.007157515734434128, + 0.008619757369160652, + 0.008107041008770466, + 0.008837534114718437, + 0.011798559688031673, + 0.013366205617785454, + 0.006361792329698801, + 0.0076297796331346035, + 0.00993560254573822, + 0.010014502331614494, + 0.009887310676276684, + 0.009409278631210327, + 0.005000776145607233, + 0.004484552424401045, + 0.00395579170435667, + 0.006628398317843676, + 0.004750072956085205, + 0.00530589884147048, + 0.007572848815470934, + 0.00346397259272635, + 0.0030471354257315397, + 0.0029030791483819485, + 0.0033100536093115807, + 0.00365834659896791, + 0.0031032655388116837, + 0.0032136670779436827, + 0.004461769945919514, + 0.0030811787582933903, + 0.00432360777631402, + 0.0053199282847344875, + 0.021377798169851303, + 0.012698860839009285, + 0.005879382602870464, + 0.012534543871879578, + 0.007382727228105068, + 0.030201420187950134, + 0.005235081538558006, + 0.006461440119892359, + 0.006446989718824625, + 0.005981596652418375, + 0.007774971425533295, + 0.00376888201572001, + 0.003253232454881072, + 0.0032453478779643774, + 0.013383377343416214, + 0.016291704028844833, + 0.017192192375659943, + 0.18341997265815735, + 0.01928279921412468, + 0.014972824603319168, + 0.16714660823345184, + 0.014907794073224068, + 0.0035175266675651073, + 0.0031282256823033094, + 0.005036348942667246, + 0.004095061682164669, + 0.0024124360643327236, + 0.005894767586141825, + 0.003299582051113248, + 0.0036186925135552883, + 0.008086510002613068, + 0.005874766036868095, + 0.006400754675269127, + 0.0117229288443923, + 0.006210417952388525, + 0.005005444400012493, + 0.014470351859927177, + 0.009123798459768295, + 0.007717588450759649, + 0.006800161674618721, + 0.006994115188717842, + 0.021536605432629585, + 0.011937140487134457, + 0.006566731724888086, + 0.01928595080971718, + 0.026498503983020782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.048782624304294586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03374743089079857 + ], + "opposite_logprob_tokens": [ + -3.408747434616089 + ], + "logprob_reward": -0.03374743089079857, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03374743089079857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07284283638000488, + 0.05128895491361618, + 0.09308769553899765, + 0.07504907250404358, + 0.0956050381064415, + 0.029743367806077003, + 0.3517758250236511, + 0.023905502632260323, + 0.08706283569335938, + 0.11963886022567749 + ], + "credit_map": [ + 0.009137571789324284, + 0.008821175433695316, + 0.008657188154757023, + 0.007086542434990406, + 0.010833196341991425, + 0.010911207646131516, + 0.00811249390244484, + 0.009283456951379776, + 0.004290958866477013, + 0.006511250510811806, + 0.004453776404261589, + 0.007238054648041725, + 0.00709203677251935, + 0.00711027579382062, + 0.0078041027300059795, + 0.006788500584661961, + 0.014552897773683071, + 0.010769275017082691, + 0.00864237267524004, + 0.01634564809501171, + 0.010407065972685814, + 0.01445598155260086, + 0.010992647148668766, + 0.006921813823282719, + 0.009807350113987923, + 0.0070725916884839535, + 0.00868942216038704, + 0.009577828459441662, + 0.009295282885432243, + 0.009335326962172985, + 0.009826263412833214, + 0.011445007286965847, + 0.004176292102783918, + 0.0055998764000833035, + 0.019974930211901665, + 0.012335442937910557, + 0.005303045269101858, + 0.012293525040149689, + 0.007123397197574377, + 0.028798529878258705, + 0.0037096459418535233, + 0.0049210842698812485, + 0.004538383334875107, + 0.0040955631993710995, + 0.004941042046993971, + 0.0027768483851104975, + 0.0024540729355067015, + 0.002306726761162281, + 0.00925320852547884, + 0.011060576885938644, + 0.012972227297723293, + 0.14830909669399261, + 0.015051482245326042, + 0.011510821990668774, + 0.13203924894332886, + 0.011579162441194057, + 0.0028465071227401495, + 0.004665462765842676, + 0.003951022867113352, + 0.0030645495280623436, + 0.0021270259749144316, + 0.0026311196852475405, + 0.0019475914305076003, + 0.0026722229085862637, + 0.008699622936546803, + 0.00732080452144146, + 0.0068799881264567375, + 0.01586584746837616, + 0.008887267671525478, + 0.007543738931417465, + 0.02034849300980568, + 0.011517075821757317, + 0.007983366958796978, + 0.00834784097969532, + 0.008576584048569202, + 0.02595888264477253, + 0.013312028720974922, + 0.00670810230076313, + 0.02162773907184601, + 0.027124326676130295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.03374743089079857 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast automobile manufacturing plant with towering steel beams and welding stations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered entirely in intense, high-contrast monochrome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a brand-new sedan with obsidian-black paintwork visible, emerging from an assembly line and gliding forward with wheels turning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's LED headlights emit two cool, piercing beams that illuminate the path ahead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, single-shot tracking shot that advances parallel to the vehicle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006749510765076 + ], + "opposite_logprob_tokens": [ + -2.6950674057006836 + ], + "logprob_reward": -0.07006749510765076, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006749510765076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03286995366215706, + 0.03143034130334854, + 0.031062031164765358, + 0.023633942008018494, + 0.21355809271335602, + 0.06659870594739914, + 0.4478747546672821, + 0.05044617876410484, + 0.06003718450665474, + 0.04248873516917229 + ], + "credit_map": [ + 0.00443280627951026, + 0.00410520238801837, + 0.00325412442907691, + 0.003333036322146654, + 0.0035401806235313416, + 0.0047725485637784, + 0.004115542396903038, + 0.005316511262208223, + 0.0029903165996074677, + 0.003812296548858285, + 0.002170480787754059, + 0.004860876128077507, + 0.004669457674026489, + 0.0046050455421209335, + 0.004618777893483639, + 0.003703091759234667, + 0.005186145659536123, + 0.004139547236263752, + 0.0032812615390866995, + 0.005016125272959471, + 0.0038381144404411316, + 0.004469234496355057, + 0.0025951433926820755, + 0.002536458894610405, + 0.0029734878335148096, + 0.0026004682295024395, + 0.0024354816414415836, + 0.0030328670982271433, + 0.003145252587273717, + 0.002548850141465664, + 0.004177029710263014, + 0.0027205052319914103, + 0.00938852783292532, + 0.009965977631509304, + 0.046072885394096375, + 0.029258860275149345, + 0.011331649497151375, + 0.02571565844118595, + 0.01591300591826439, + 0.06591153889894485, + 0.0073026372119784355, + 0.008578475564718246, + 0.00890351366251707, + 0.011072409339249134, + 0.012482699938118458, + 0.00784576591104269, + 0.005509661510586739, + 0.0049035390838980675, + 0.02334017865359783, + 0.02226858213543892, + 0.0251608956605196, + 0.1676817685365677, + 0.024970557540655136, + 0.018120834603905678, + 0.15078309178352356, + 0.01554886344820261, + 0.006107007618993521, + 0.0036887091118842363, + 0.00839570164680481, + 0.004653207957744598, + 0.0034624901600182056, + 0.010907227173447609, + 0.007160544861108065, + 0.006071293260902166, + 0.006039436440914869, + 0.004528725519776344, + 0.004264154005795717, + 0.011123192496597767, + 0.005241929553449154, + 0.005588584113866091, + 0.014492085203528404, + 0.00875907577574253, + 0.0029776333831250668, + 0.0026719621382653713, + 0.0029579265974462032, + 0.007727344986051321, + 0.004268276505172253, + 0.0029607610777020454, + 0.007933818735182285, + 0.010991012677550316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.07006749510765076 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_12/index_tensor([6942, 6942])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_12/index_tensor([6942, 6942])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8b761ae432c07d190c405f19c689c48bd63f618e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_12/index_tensor([6942, 6942])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium shot, shot upward from the ground, is intermittently punctuated by distorted scan lines and brief bursts of snowy noise due to analog signal transmission glitches. The scene is set in a narrow, dank cyberpunk city back alley, the air thick with icy steam. Magenta and electric blue neon billboards flicker endlessly overhead, casting a harsh, unstable light on brick walls covered in graffiti and exposed pipes. On the ground, a bottomless pool of oily, dark green sewage is suddenly stirred by a tremendous force, and then, with an explosive gesture, a massive Australian lobster rises from the water. Its wet crimson carapace reflects a fluid, metallic sheen in the flickering neon lights, while droplets of water and foul fluids continuously flow down its segmented body and sharp back. Its huge, asymmetrical claws opened threateningly in the air, its black compound eyes flickered an unnatural glow in the shadows, and its long tentacles twitched sensitively in the humid air. The whole picture was filled with a sudden sense of grotesque horror and strong visual impact.", + "seed": 17610, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_12/index_6942_seed_17610_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.015007893578149379, + "vlm_logprob": -0.015007893578149379, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.015007893578149379, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05327311158180237 + ], + "opposite_logprob_tokens": [ + -3.0532732009887695 + ], + "logprob_reward": -0.05327311158180237, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05327311158180237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26509883999824524, + 0.24964354932308197, + 0.2111879289150238, + 0.05281972885131836, + 0.07508257031440735, + 0.03221210464835167, + 0.016019942238926888, + 0.0165387149900198, + 0.044947750866413116, + 0.03644886240363121 + ], + "credit_map": [ + 0.028804229572415352, + 0.03132259473204613, + 0.024523897096514702, + 0.039552465081214905, + 0.04454943165183067, + 0.021598564460873604, + 0.041685551404953, + 0.03306212276220322, + 0.047880999743938446, + 0.046567972749471664, + 0.03191421180963516, + 0.017434341832995415, + 0.04522906243801117, + 0.02164880558848381, + 0.01907275803387165, + 0.019895415753126144, + 0.0348665788769722, + 0.015570933930575848, + 0.07775299996137619, + 0.01692620851099491, + 0.0074436548165977, + 0.012774178758263588, + 0.02082120254635811, + 0.025032201781868935, + 0.006125411484390497, + 0.006131527479737997, + 0.0060927108861505985, + 0.006544869393110275, + 0.00607762997969985, + 0.009777838364243507, + 0.00565651198849082, + 0.006413235794752836, + 0.008312668651342392, + 0.003112186910584569, + 0.004321351647377014, + 0.004547021817415953, + 0.003921655938029289, + 0.018415771424770355, + 0.006006179377436638, + 0.026445744559168816, + 0.004547602962702513, + 0.002369093243032694, + 0.0027535443659871817, + 0.002387691754847765, + 0.0022088855039328337, + 0.0057611106894910336, + 0.0021341065876185894, + 0.010050074197351933, + 0.003486175090074539, + 0.0011914269998669624, + 0.002014513360336423, + 0.0015222164802253246, + 0.0017946023726835847, + 0.0015966910868883133, + 0.0014400744112208486, + 0.00297424360178411, + 0.0027750120498239994, + 0.001564952777698636, + 0.0020891674794256687, + 0.0019279320258647203, + 0.0015495145926252007, + 0.0019048948306590319, + 0.0028276562225073576, + 0.00189958733972162, + 0.0033731088042259216, + 0.005036380607634783, + 0.0027510144282132387, + 0.004991194698959589, + 0.0025455900467932224, + 0.01382975373417139, + 0.0032728195656090975, + 0.009147889912128448, + 0.006211179308593273, + 0.003531830618157983, + 0.003925052471458912, + 0.002629237249493599, + 0.00838674046099186, + 0.005282132420688868, + 0.002792337676510215, + 0.0036903556901961565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.05327311158180237 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015353093622252345 + ], + "opposite_logprob_tokens": [ + -6.501535415649414 + ], + "logprob_reward": -0.0015353093622252345, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015353093622252345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.229689359664917, + 0.15688054263591766, + 0.18085213005542755, + 0.05030469968914986, + 0.09619441628456116, + 0.09300754219293594, + 0.018577896058559418, + 0.046678222715854645, + 0.04353877529501915, + 0.08427635580301285 + ], + "credit_map": [ + 0.01945139467716217, + 0.024138929322361946, + 0.02215738408267498, + 0.02809467352926731, + 0.026370566338300705, + 0.02275470830500126, + 0.03436771407723427, + 0.05235400050878525, + 0.028026528656482697, + 0.028323756530880928, + 0.02027861773967743, + 0.012289315462112427, + 0.029759544879198074, + 0.014286153949797153, + 0.011641927063465118, + 0.012274700216948986, + 0.027060266584157944, + 0.01464566309005022, + 0.057402513921260834, + 0.021608775481581688, + 0.01073120255023241, + 0.010829586535692215, + 0.01802413910627365, + 0.020549986511468887, + 0.006584523245692253, + 0.005948320031166077, + 0.003169342875480652, + 0.006954724434763193, + 0.009761350229382515, + 0.006907457951456308, + 0.007035841699689627, + 0.003943141549825668, + 0.013773715123534203, + 0.004323963075876236, + 0.005433334968984127, + 0.004529149271547794, + 0.004648217000067234, + 0.021022701635956764, + 0.006240630988031626, + 0.036222703754901886, + 0.010741208679974079, + 0.006261068861931562, + 0.007623768411576748, + 0.008974464610219002, + 0.005211660172790289, + 0.01464210543781519, + 0.008416598662734032, + 0.03113667108118534, + 0.0027253893204033375, + 0.001489912741817534, + 0.002011851640418172, + 0.0025126247201114893, + 0.0026225587353110313, + 0.0023898999206721783, + 0.0022499626502394676, + 0.002575696911662817, + 0.006999583449214697, + 0.00483541376888752, + 0.0044258963316679, + 0.004664761014282703, + 0.005980951711535454, + 0.006292365957051516, + 0.007599558215588331, + 0.005879689939320087, + 0.004220197908580303, + 0.007417798507958651, + 0.00407906761392951, + 0.006140872370451689, + 0.004465772304683924, + 0.005712308455258608, + 0.005758966784924269, + 0.005743793211877346, + 0.013424658216536045, + 0.007732626982033253, + 0.011591359041631222, + 0.005166193470358849, + 0.020188918337225914, + 0.010051415301859379, + 0.0056827133521437645, + 0.010438471101224422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0015353093622252345 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629185914993286 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.007629185914993286, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629185914993286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4211525619029999, + 0.08725585043430328, + 0.05958830192685127, + 0.0761050209403038, + 0.10010898858308792, + 0.022140413522720337, + 0.01488415151834488, + 0.03988167271018028, + 0.10378451645374298, + 0.07509854435920715 + ], + "credit_map": [ + 0.06631588935852051, + 0.04964899271726608, + 0.0345725379884243, + 0.045364752411842346, + 0.08420218527317047, + 0.05262414738535881, + 0.053323108702898026, + 0.03510088101029396, + 0.01562252826988697, + 0.013605641201138496, + 0.012358077801764011, + 0.007398291956633329, + 0.013608605600893497, + 0.009313221089541912, + 0.007112882565706968, + 0.008236590772867203, + 0.008741997182369232, + 0.006712521892040968, + 0.013439097441732883, + 0.009281366132199764, + 0.003943051677197218, + 0.005022310186177492, + 0.006397114600986242, + 0.006050831638276577, + 0.008535214699804783, + 0.007392827421426773, + 0.00788636039942503, + 0.011686500161886215, + 0.008014545775949955, + 0.00926531944423914, + 0.01519718486815691, + 0.008127058856189251, + 0.012669399380683899, + 0.004646157380193472, + 0.0069130645133554935, + 0.005641778465360403, + 0.005611084867268801, + 0.021870965138077736, + 0.008207426406443119, + 0.034549109637737274, + 0.0027860065456479788, + 0.0019645364955067635, + 0.0031374909449368715, + 0.002786031225696206, + 0.002027802402153611, + 0.0027340957894921303, + 0.0027349938172847033, + 0.003969453740864992, + 0.002552812220528722, + 0.00135087373200804, + 0.0014404448447749019, + 0.0019064333755522966, + 0.0017592976801097393, + 0.0020312180276960135, + 0.001334320055320859, + 0.0025087506510317326, + 0.0037251999601721764, + 0.004035114776343107, + 0.004119975492358208, + 0.005230963695794344, + 0.003991341684013605, + 0.006836747284978628, + 0.005644727498292923, + 0.006297596264630556, + 0.00858050212264061, + 0.009061183780431747, + 0.007167896721512079, + 0.008028389886021614, + 0.005255145486444235, + 0.036059603095054626, + 0.006006264127790928, + 0.023625509813427925, + 0.00935427937656641, + 0.011393631808459759, + 0.009769607335329056, + 0.005791738163679838, + 0.01637807860970497, + 0.009203647263348103, + 0.005335347726941109, + 0.007872207090258598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.007629185914993286 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028217521030455828 + ], + "opposite_logprob_tokens": [ + -5.877821922302246 + ], + "logprob_reward": -0.0028217521030455828, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028217521030455828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0887543186545372, + 0.26205578446388245, + 0.13477195799350739, + 0.05809715762734413, + 0.1242583841085434, + 0.13863804936408997, + 0.03547784686088562, + 0.03578248247504234, + 0.06332279741764069, + 0.05884118750691414 + ], + "credit_map": [ + 0.007527300156652927, + 0.009895502589643002, + 0.008823394775390625, + 0.01924990303814411, + 0.008896421641111374, + 0.007629903499037027, + 0.014170577749609947, + 0.012561316601932049, + 0.052430808544158936, + 0.04717136546969414, + 0.03418809548020363, + 0.019395070150494576, + 0.04555337876081467, + 0.02440500259399414, + 0.01758185215294361, + 0.021330201998353004, + 0.022798193618655205, + 0.009921662509441376, + 0.04530806094408035, + 0.014434779062867165, + 0.005151727702468634, + 0.00833001360297203, + 0.014817343093454838, + 0.014010181650519371, + 0.007870707660913467, + 0.01185959205031395, + 0.0070695155300199986, + 0.0069749499671161175, + 0.007387338671833277, + 0.005500657483935356, + 0.005962745286524296, + 0.005471646785736084, + 0.016207164153456688, + 0.0054285526275634766, + 0.007860176265239716, + 0.006590797565877438, + 0.006239163223654032, + 0.02807684987783432, + 0.008654684759676456, + 0.045200999826192856, + 0.0168959628790617, + 0.010846148245036602, + 0.011557416059076786, + 0.012552611529827118, + 0.00839464645832777, + 0.02143530175089836, + 0.01221396867185831, + 0.04474199190735817, + 0.0036718931514769793, + 0.004121321253478527, + 0.003518818411976099, + 0.003942809067666531, + 0.004240271169692278, + 0.004437154158949852, + 0.004400674719363451, + 0.007144903764128685, + 0.003967747092247009, + 0.004569993354380131, + 0.0035878599155694246, + 0.005136388354003429, + 0.0038722041063010693, + 0.00435134069994092, + 0.005432236008346081, + 0.004864712245762348, + 0.0051959361881017685, + 0.0082183126360178, + 0.00507894903421402, + 0.004970853216946125, + 0.003981127869337797, + 0.020442310720682144, + 0.003922497387975454, + 0.011512813158333302, + 0.012215719558298588, + 0.004645576700568199, + 0.005731468554586172, + 0.003675872692838311, + 0.016236688941717148, + 0.006555305328220129, + 0.004059466999024153, + 0.0057210917584598064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.0028217521030455828 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978010892868042 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.00978010892868042, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978010892868042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21293793618679047, + 0.19974546134471893, + 0.047537557780742645, + 0.05044231936335564, + 0.16488513350486755, + 0.09105732291936874, + 0.033050183206796646, + 0.049898937344551086, + 0.05237133055925369, + 0.09807378798723221 + ], + "credit_map": [ + 0.024108823388814926, + 0.020495420321822166, + 0.01876377873122692, + 0.02364843524992466, + 0.033457931131124496, + 0.02081313356757164, + 0.030058063566684723, + 0.041592370718717575, + 0.038470882922410965, + 0.03456636145710945, + 0.0294538214802742, + 0.016920039430260658, + 0.03249158337712288, + 0.016770215705037117, + 0.01444939523935318, + 0.01662316359579563, + 0.007958191446959972, + 0.003877336625009775, + 0.01037291157990694, + 0.006654786877334118, + 0.003922600299119949, + 0.0038135447539389133, + 0.005199378356337547, + 0.005738810636103153, + 0.005037735681980848, + 0.006620983127504587, + 0.0059892828576266766, + 0.006934174802154303, + 0.0077455732971429825, + 0.004513715393841267, + 0.009010030888020992, + 0.004590826109051704, + 0.02610025741159916, + 0.007267344743013382, + 0.007881072349846363, + 0.007659072056412697, + 0.007132381666451693, + 0.035212624818086624, + 0.010072323493659496, + 0.06356006860733032, + 0.009547446854412556, + 0.006094178184866905, + 0.008126589469611645, + 0.009774323552846909, + 0.0055749258026480675, + 0.014095631428062916, + 0.008573682978749275, + 0.029270552098751068, + 0.003887656144797802, + 0.003052870510146022, + 0.002815636107698083, + 0.005108507815748453, + 0.0045297471806406975, + 0.00405507767572999, + 0.004112386144697666, + 0.005488302558660507, + 0.006945156492292881, + 0.003940714988857508, + 0.006493748631328344, + 0.004731157794594765, + 0.003385064424946904, + 0.009810852818191051, + 0.006958368234336376, + 0.007633876521140337, + 0.005285302177071571, + 0.007173734251409769, + 0.0056381975300610065, + 0.006084482651203871, + 0.004783375188708305, + 0.010857867076992989, + 0.005135354585945606, + 0.007413020357489586, + 0.012947600334882736, + 0.011670717038214207, + 0.009229740127921104, + 0.010505752637982368, + 0.017709722742438316, + 0.011407693848013878, + 0.0124898636713624, + 0.012112694792449474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.00978010892868042 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot, shot upward from the ground, is intermittently punctuated by distorted scan lines and brief bursts of snowy noise due to analog signal transmission glitches. The scene is set in a narrow, dank cyberpunk city back alley, the air thick with icy steam. Magenta and electric blue neon billboards flicker endlessly overhead, casting a harsh, unstable light on brick walls covered in graffiti and exposed pipes. On the ground, a bottomless pool of oily, dark green sewage is suddenly stirred by a tremendous force, and then, with an explosive gesture, a massive Australian lobster rises from the water. Its wet crimson carapace reflects a fluid, metallic sheen in the flickering neon lights, while droplets of water and foul fluids continuously flow down its segmented body and sharp back. Its huge, asymmetrical claws opened threateningly in the air, its black compound eyes flickered an unnatural glow in the shadows, and its long tentacles twitched sensitively in the humid air. The whole picture was filled with a sudden sense of grotesque horror and strong visual impact.", + "seed": 17611, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_12/index_6942_seed_17611_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03237039850791916, + "vlm_logprob": -0.03237039850791916, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.03237039850791916, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14623329043388367 + ], + "opposite_logprob_tokens": [ + -2.021233320236206 + ], + "logprob_reward": -0.14623329043388367, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14623329043388367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2080005705356598, + 0.03780510276556015, + 0.045667700469493866, + 0.3112046718597412, + 0.13542746007442474, + 0.046699654310941696, + 0.13117636740207672, + 0.026335347443819046, + 0.012810091488063335, + 0.04487305507063866 + ], + "credit_map": [ + 0.03636390343308449, + 0.02607370726764202, + 0.01906738430261612, + 0.02142101526260376, + 0.029935108497738838, + 0.029659900814294815, + 0.025569211691617966, + 0.019910335540771484, + 0.005083476193249226, + 0.00578508572652936, + 0.003252789843827486, + 0.004564996808767319, + 0.0046913474798202515, + 0.006145506631582975, + 0.005075606983155012, + 0.0032062919344753027, + 0.0053785452619194984, + 0.005238354206085205, + 0.00457084272056818, + 0.010122636333107948, + 0.0065827276557683945, + 0.004904766101390123, + 0.005638377740979195, + 0.003231450915336609, + 0.03686391934752464, + 0.05347836762666702, + 0.020639369264245033, + 0.02376888319849968, + 0.09786497056484222, + 0.029853707179427147, + 0.030748754739761353, + 0.017986683174967766, + 0.029229748994112015, + 0.004416670650243759, + 0.05618039891123772, + 0.009757262654602528, + 0.003950288984924555, + 0.006229298654943705, + 0.00983867421746254, + 0.01582510769367218, + 0.00468446547165513, + 0.0060980068519711494, + 0.005350032355636358, + 0.005752000492066145, + 0.007154226303100586, + 0.006295803468674421, + 0.005355615168809891, + 0.006009501405060291, + 0.010125766508281231, + 0.006939558312296867, + 0.008417608216404915, + 0.013503745198249817, + 0.008460929617285728, + 0.0077367820776999, + 0.04108735918998718, + 0.034904610365629196, + 0.0036463052965700626, + 0.001606413978151977, + 0.005321817472577095, + 0.002028312999755144, + 0.0014513212954625487, + 0.0023271730169653893, + 0.0022608391009271145, + 0.007693161256611347, + 0.00206359033472836, + 0.0014791416469961405, + 0.0019588025752454996, + 0.0012723453110083938, + 0.00130682741291821, + 0.0019885587971657515, + 0.0011299614561721683, + 0.0016108629060909152, + 0.005422242451459169, + 0.0022131074219942093, + 0.00537694850936532, + 0.0041420902125537395, + 0.002157482085749507, + 0.012111014686524868, + 0.003132615238428116, + 0.010317553766071796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.14623329043388367 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000942858459893614 + ], + "opposite_logprob_tokens": [ + -7.000942707061768 + ], + "logprob_reward": -0.000942858459893614, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.000942858459893614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04626617208123207, + 0.01601821556687355, + 0.03132570534944534, + 0.24624404311180115, + 0.07526590675115585, + 0.020645812153816223, + 0.47430419921875, + 0.010881480760872364, + 0.017847519367933273, + 0.061200931668281555 + ], + "credit_map": [ + 0.009617293253540993, + 0.005465340334922075, + 0.0056475806050002575, + 0.0040964046493172646, + 0.006943643558770418, + 0.005592521280050278, + 0.004867075942456722, + 0.004036314319819212, + 0.0024464658927172422, + 0.002284096786752343, + 0.0014058338711038232, + 0.002031253185123205, + 0.0018824603175744414, + 0.002742940792813897, + 0.0018299398943781853, + 0.0013952252920717, + 0.003035820322111249, + 0.0024804340209811926, + 0.0038894477766007185, + 0.009136688895523548, + 0.0030262686777859926, + 0.0025539195630699396, + 0.004739993717521429, + 0.002463135402649641, + 0.029128514230251312, + 0.04098089784383774, + 0.017307203263044357, + 0.019640794023871422, + 0.07648999989032745, + 0.023386478424072266, + 0.02487039752304554, + 0.014439760707318783, + 0.016682418063282967, + 0.00315312878228724, + 0.03013627417385578, + 0.005772845353931189, + 0.0020429876167327166, + 0.0036258436739444733, + 0.00532024260610342, + 0.008532162755727768, + 0.003146799048408866, + 0.0023197655100375414, + 0.002521599642932415, + 0.003229528432711959, + 0.001854788395576179, + 0.0018669062992557883, + 0.0025526173412799835, + 0.0031538072507828474, + 0.029102494940161705, + 0.013516225852072239, + 0.024057218804955482, + 0.053018905222415924, + 0.015892939642071724, + 0.019003594294190407, + 0.17232757806777954, + 0.14738525450229645, + 0.0015272398013621569, + 0.0009174129227176309, + 0.0015883940504863858, + 0.0010858395835384727, + 0.000937282107770443, + 0.0013792059617117047, + 0.0011981435818597674, + 0.002247963100671768, + 0.005245369393378496, + 0.0013011805713176727, + 0.0030242973007261753, + 0.0015117812436074018, + 0.0010572230676189065, + 0.0017110321205109358, + 0.0011590628419071436, + 0.0028375734109431505, + 0.0046671293675899506, + 0.0018101511523127556, + 0.007018704432994127, + 0.004211902618408203, + 0.002393457107245922, + 0.022571032866835594, + 0.0039802901446819305, + 0.01454826258122921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.000942858459893614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096925724297762 + ], + "opposite_logprob_tokens": [ + -5.504096984863281 + ], + "logprob_reward": -0.004096925724297762, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096925724297762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10885749757289886, + 0.03612193837761879, + 0.03789928928017616, + 0.11578452587127686, + 0.0903010442852974, + 0.039666492491960526, + 0.47032982110977173, + 0.02457176148891449, + 0.022874588146805763, + 0.0535929836332798 + ], + "credit_map": [ + 0.015777399763464928, + 0.011209933087229729, + 0.014398790895938873, + 0.012690603733062744, + 0.008726390078663826, + 0.0165630504488945, + 0.01694343611598015, + 0.012547900900244713, + 0.004070139955729246, + 0.00365069299004972, + 0.0037388810887932777, + 0.007425711024552584, + 0.0031647528521716595, + 0.004984345752745867, + 0.00572328083217144, + 0.003364137141034007, + 0.00418644305318594, + 0.002834504935890436, + 0.004834156949073076, + 0.008513703010976315, + 0.0023076729848980904, + 0.0024739594664424658, + 0.008608442731201649, + 0.004140407778322697, + 0.012545503675937653, + 0.01787768490612507, + 0.008561034686863422, + 0.011838225647807121, + 0.03246790170669556, + 0.0110300462692976, + 0.013019893318414688, + 0.008444242179393768, + 0.018598997965455055, + 0.006771106738597155, + 0.03219453990459442, + 0.006700422614812851, + 0.004583457950502634, + 0.004935876931995153, + 0.006864391267299652, + 0.009652264416217804, + 0.005304816644638777, + 0.002622035099193454, + 0.005026812199503183, + 0.008157322183251381, + 0.003159541403874755, + 0.004968676250427961, + 0.007793500553816557, + 0.002633787924423814, + 0.03015062026679516, + 0.012062040157616138, + 0.02243492193520069, + 0.05339981988072395, + 0.015374956652522087, + 0.018504535779356956, + 0.17324939370155334, + 0.14515355229377747, + 0.0032977419905364513, + 0.0028863779734820127, + 0.0056704021990299225, + 0.003000882687047124, + 0.001988145289942622, + 0.0017676996067166328, + 0.002547657350078225, + 0.003412856487557292, + 0.0026492848992347717, + 0.0029544206336140633, + 0.004506481811404228, + 0.0017886512214317918, + 0.0021416766103357077, + 0.0043152510188519955, + 0.001477917656302452, + 0.0030409065075218678, + 0.005118584726005793, + 0.0017043224070221186, + 0.004087450914084911, + 0.004643630236387253, + 0.002388100139796734, + 0.018750440329313278, + 0.003913607448339462, + 0.012986850924789906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.004096925724297762 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019421778852120042 + ], + "opposite_logprob_tokens": [ + -6.251942157745361 + ], + "logprob_reward": -0.0019421778852120042, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019421778852120042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09173997491598129, + 0.03771860525012016, + 0.05406106635928154, + 0.31770431995391846, + 0.08758220076560974, + 0.04569103568792343, + 0.2714187502861023, + 0.028091181069612503, + 0.03747702017426491, + 0.02851586975157261 + ], + "credit_map": [ + 0.014491480775177479, + 0.010813074186444283, + 0.009778324514627457, + 0.013301504775881767, + 0.011114531196653843, + 0.011140292510390282, + 0.011643994599580765, + 0.009456771425902843, + 0.005075140856206417, + 0.005735424812883139, + 0.003383997594937682, + 0.0046561951749026775, + 0.004563488066196442, + 0.00618231063708663, + 0.005044732708483934, + 0.0030773135367780924, + 0.006023503374308348, + 0.005487836431711912, + 0.004264636430889368, + 0.011201882734894753, + 0.0056929984129965305, + 0.006065958645194769, + 0.010371691547334194, + 0.004952558316290379, + 0.03715154156088829, + 0.05334808677434921, + 0.022590141743421555, + 0.03022167831659317, + 0.09540005028247833, + 0.02958700992166996, + 0.0293695330619812, + 0.02003627084195614, + 0.015197888016700745, + 0.010002543218433857, + 0.014552591368556023, + 0.01202111504971981, + 0.005310146138072014, + 0.007515096105635166, + 0.014621487818658352, + 0.008361329324543476, + 0.00508333183825016, + 0.004315621685236692, + 0.008049411699175835, + 0.0062433937564492226, + 0.005205843131989241, + 0.004879198502749205, + 0.0077641974203288555, + 0.00415003951638937, + 0.017120009288191795, + 0.007150752004235983, + 0.012439032085239887, + 0.03102809563279152, + 0.008437203243374825, + 0.011210782453417778, + 0.09929903596639633, + 0.08473384380340576, + 0.003515149699524045, + 0.003144932445138693, + 0.002910971874371171, + 0.0031113694421947002, + 0.0023795622400939465, + 0.002284953836351633, + 0.0034958559554070234, + 0.007248386275023222, + 0.008462576195597649, + 0.0027196111623197794, + 0.007924994453787804, + 0.0029931545723229647, + 0.0022114200983196497, + 0.005017613992094994, + 0.0023889304138720036, + 0.005758719518780708, + 0.003936541732400656, + 0.0028689801692962646, + 0.0040603214874863625, + 0.0038154625799506903, + 0.002701860386878252, + 0.004196475259959698, + 0.003434285754337907, + 0.0035019421484321356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.0019421778852120042 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636740036308765 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636740036308765, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636740036308765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16879509389400482, + 0.09125399589538574, + 0.05489916726946831, + 0.23326358199119568, + 0.07007389515638351, + 0.05053185299038887, + 0.15907461941242218, + 0.042395517230033875, + 0.0585942342877388, + 0.07111808657646179 + ], + "credit_map": [ + 0.03371085599064827, + 0.024599922820925713, + 0.01710561104118824, + 0.015651892870664597, + 0.02409709244966507, + 0.019558757543563843, + 0.019315365701913834, + 0.014755564741790295, + 0.010769137181341648, + 0.010164345614612103, + 0.00985436886548996, + 0.013037440367043018, + 0.009843140840530396, + 0.013138853944838047, + 0.014906702563166618, + 0.009539993479847908, + 0.006211499683558941, + 0.005704580806195736, + 0.004219237249344587, + 0.013375493697822094, + 0.005377080291509628, + 0.006729974877089262, + 0.007915735244750977, + 0.005365560296922922, + 0.02901020646095276, + 0.03963043540716171, + 0.013431064784526825, + 0.018353650346398354, + 0.07449489831924438, + 0.022272996604442596, + 0.023924540728330612, + 0.012145766988396645, + 0.013710446655750275, + 0.00593453086912632, + 0.023088611662387848, + 0.006557018496096134, + 0.003622970078140497, + 0.0036416309885680676, + 0.00651912298053503, + 0.006999560631811619, + 0.0051637450233101845, + 0.0051064989529550076, + 0.008280813694000244, + 0.007211444433778524, + 0.007130496669560671, + 0.006525104865431786, + 0.007269388530403376, + 0.0038443550001829863, + 0.011264491826295853, + 0.006496040616184473, + 0.013791514560580254, + 0.016645003110170364, + 0.007303529419004917, + 0.00878275465220213, + 0.05269479751586914, + 0.04209647327661514, + 0.0039423564448952675, + 0.0027373027987778187, + 0.007826405577361584, + 0.0042985850013792515, + 0.0030929879285395145, + 0.002628588816151023, + 0.004821102134883404, + 0.013048183172941208, + 0.016118597239255905, + 0.003152095479890704, + 0.01307449210435152, + 0.004396153148263693, + 0.0025890320539474487, + 0.006281145848333836, + 0.003510944312438369, + 0.009471766650676727, + 0.006661205552518368, + 0.002765987766906619, + 0.00828531663864851, + 0.006059655919671059, + 0.0034446250647306442, + 0.02515828050673008, + 0.005585505627095699, + 0.013157500885426998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.008636740036308765 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot, shot upward from the ground, is intermittently punctuated by distorted scan lines and brief bursts of snowy noise due to analog signal transmission glitches. The scene is set in a narrow, dank cyberpunk city back alley, the air thick with icy steam. Magenta and electric blue neon billboards flicker endlessly overhead, casting a harsh, unstable light on brick walls covered in graffiti and exposed pipes. On the ground, a bottomless pool of oily, dark green sewage is suddenly stirred by a tremendous force, and then, with an explosive gesture, a massive Australian lobster rises from the water. Its wet crimson carapace reflects a fluid, metallic sheen in the flickering neon lights, while droplets of water and foul fluids continuously flow down its segmented body and sharp back. Its huge, asymmetrical claws opened threateningly in the air, its black compound eyes flickered an unnatural glow in the shadows, and its long tentacles twitched sensitively in the humid air. The whole picture was filled with a sudden sense of grotesque horror and strong visual impact.", + "seed": 17612, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_12/index_6942_seed_17612_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14531908265780658, + "vlm_logprob": -0.14531908265780658, + "vlm_yes_no_margin": 4.124999952316284, + "vlm_token_credit": -0.14531908265780658, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6946923732757568 + ], + "opposite_logprob_tokens": [ + -0.6946923732757568 + ], + "logprob_reward": -0.6946923732757568, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6946923732757568, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.23871831595897675, + 0.12993331253528595, + 0.050390809774398804, + 0.01995670236647129, + 0.05913469195365906, + 0.061439961194992065, + 0.04332289472222328, + 0.012496099807322025, + 0.01106957346200943, + 0.3735376298427582 + ], + "credit_map": [ + 0.029148947447538376, + 0.014615415595471859, + 0.05914989486336708, + 0.02852824702858925, + 0.019180743023753166, + 0.019512977451086044, + 0.04730404168367386, + 0.02127804234623909, + 0.006946122739464045, + 0.009409920312464237, + 0.009308310225605965, + 0.005715573206543922, + 0.04747788608074188, + 0.03492315486073494, + 0.007511167321354151, + 0.008641178719699383, + 0.005308896768838167, + 0.006431670859456062, + 0.004606076516211033, + 0.004005667753517628, + 0.006519115064293146, + 0.009447037242352962, + 0.006592863239347935, + 0.007479480933398008, + 0.00210587284527719, + 0.002537064952775836, + 0.0025338763371109962, + 0.0030136897694319487, + 0.002629549242556095, + 0.0026315434370189905, + 0.0022327948827296495, + 0.0022723102010786533, + 0.0033832460176199675, + 0.003577336436137557, + 0.004302474670112133, + 0.0024918955750763416, + 0.022616932168602943, + 0.017504140734672546, + 0.0028135215397924185, + 0.002445144345983863, + 0.005391737446188927, + 0.006415906827896833, + 0.004727251827716827, + 0.004967015702277422, + 0.010540985502302647, + 0.017717953771352768, + 0.004143019672483206, + 0.0075360978953540325, + 0.004243360832333565, + 0.0025163181126117706, + 0.013781949877738953, + 0.004294725600630045, + 0.002257256768643856, + 0.0032395555172115564, + 0.002585567766800523, + 0.010404162108898163, + 0.0019512370927259326, + 0.0010463084327057004, + 0.0018960926681756973, + 0.0015152900014072657, + 0.0011508309980854392, + 0.0016317011322826147, + 0.0014330917038023472, + 0.0018715482437983155, + 0.0011773410951718688, + 0.0008294313447549939, + 0.0035207190085202456, + 0.0009029826032929122, + 0.0008323643123731017, + 0.0015587114030495286, + 0.0010925499955192208, + 0.0011554736411198974, + 0.03473588451743126, + 0.008676699362695217, + 0.025070587173104286, + 0.017682915553450584, + 0.025729645043611526, + 0.11453507095575333, + 0.12159235030412674, + 0.025514500215649605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.6946923732757568 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00173049361910671 + ], + "opposite_logprob_tokens": [ + -6.376730442047119 + ], + "logprob_reward": -0.00173049361910671, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.00173049361910671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18938031792640686, + 0.3173545002937317, + 0.07629764825105667, + 0.024052675813436508, + 0.1333180069923401, + 0.06874873489141464, + 0.053064823150634766, + 0.01627449132502079, + 0.014127626083791256, + 0.10738123208284378 + ], + "credit_map": [ + 0.024363882839679718, + 0.012211134657263756, + 0.051677156239748, + 0.02101977728307247, + 0.012860841117799282, + 0.015330338850617409, + 0.032728374004364014, + 0.019188793376088142, + 0.017522288486361504, + 0.023508509621024132, + 0.020443499088287354, + 0.010595006868243217, + 0.11923428624868393, + 0.09267926216125488, + 0.0172457043081522, + 0.016125919297337532, + 0.009373614564538002, + 0.008199633099138737, + 0.007600591517984867, + 0.006764188874512911, + 0.011817555874586105, + 0.01479930616915226, + 0.008152774535119534, + 0.00958998128771782, + 0.0027979619335383177, + 0.0035971393808722496, + 0.003092851722612977, + 0.003383058123290539, + 0.002765307668596506, + 0.0033498737029731274, + 0.002884840127080679, + 0.002181639662012458, + 0.008526613004505634, + 0.008060737513005733, + 0.012091727927327156, + 0.002982061356306076, + 0.05461156368255615, + 0.038997478783130646, + 0.0042183417826890945, + 0.003829469671472907, + 0.005617539398372173, + 0.007905533537268639, + 0.006082499399781227, + 0.005382166709750891, + 0.012315916828811169, + 0.019217727705836296, + 0.0052766925655305386, + 0.006950653623789549, + 0.00466526485979557, + 0.003015500260517001, + 0.016639549285173416, + 0.00520938728004694, + 0.003220570972189307, + 0.004566893447190523, + 0.00262037618085742, + 0.013127276673913002, + 0.002364927204325795, + 0.0017554264049977064, + 0.0020998925901949406, + 0.0018502003513276577, + 0.0016142266103997827, + 0.0027016347739845514, + 0.001975974766537547, + 0.0019122080411762, + 0.0015878108097240329, + 0.0012968784430995584, + 0.0032785949297249317, + 0.0013364542974159122, + 0.0011107875034213066, + 0.0024652827996760607, + 0.0016037210589274764, + 0.0014480953104794025, + 0.009616903960704803, + 0.002965726191177964, + 0.007723034825176001, + 0.0052351984195411205, + 0.007446914911270142, + 0.031405020505189896, + 0.03518564626574516, + 0.007802778854966164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.00173049361910671 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818765513598919 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.01818765513598919, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818765513598919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1168023943901062, + 0.054556433111429214, + 0.044144801795482635, + 0.04382253810763359, + 0.38630756735801697, + 0.030167492106556892, + 0.030497116968035698, + 0.00955243594944477, + 0.01438913680613041, + 0.2697600722312927 + ], + "credit_map": [ + 0.01547874417155981, + 0.007147834170609713, + 0.03103184700012207, + 0.013294601812958717, + 0.008861186914145947, + 0.0093063460662961, + 0.017140697687864304, + 0.01454113144427538, + 0.006336652673780918, + 0.006978670135140419, + 0.007808515802025795, + 0.005250674672424793, + 0.009739804081618786, + 0.006297027692198753, + 0.006189033854752779, + 0.005956056993454695, + 0.005141522269695997, + 0.005231882911175489, + 0.003692902158945799, + 0.004171131178736687, + 0.006357923615723848, + 0.008037114515900612, + 0.005109501536935568, + 0.0064028226770460606, + 0.005011386703699827, + 0.004672912415117025, + 0.005847607273608446, + 0.007240213453769684, + 0.0048810215666890144, + 0.00427641998976469, + 0.005586965940892696, + 0.006306009367108345, + 0.023141466081142426, + 0.0200810544192791, + 0.032887253910303116, + 0.009038979187607765, + 0.15959447622299194, + 0.11665070801973343, + 0.013331211172044277, + 0.011582417413592339, + 0.0035143133718520403, + 0.003963392693549395, + 0.0037082878407090902, + 0.0027165955398231745, + 0.004237855784595013, + 0.005210462491959333, + 0.0031216780189424753, + 0.0036949063651263714, + 0.0037387690972536802, + 0.0017631907248869538, + 0.004696752410382032, + 0.003246551612392068, + 0.0024931509979069233, + 0.00480670016258955, + 0.0027226456440985203, + 0.007029356434941292, + 0.0011694971472024918, + 0.0005111076170578599, + 0.001577530987560749, + 0.0012069819495081902, + 0.0008668942609801888, + 0.001074169180355966, + 0.0011606308398768306, + 0.0019856239669024944, + 0.0014407283160835505, + 0.0006991395493969321, + 0.005066612735390663, + 0.0010860339971259236, + 0.0010925832903012633, + 0.0023623695597052574, + 0.0011432920582592487, + 0.0014983761357143521, + 0.02514602243900299, + 0.00640492420643568, + 0.01834215223789215, + 0.012969942763447762, + 0.018484629690647125, + 0.083010733127594, + 0.08696835488080978, + 0.018433310091495514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.01818765513598919 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252964794635773 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.005252964794635773, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252964794635773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09522788971662521, + 0.20587201416492462, + 0.026733027771115303, + 0.039614468812942505, + 0.02178458496928215, + 0.04736142233014107, + 0.04486642777919769, + 0.01122822891920805, + 0.012323207221925259, + 0.49498873949050903 + ], + "credit_map": [ + 0.011011824011802673, + 0.006178955547511578, + 0.02674233354628086, + 0.01323908381164074, + 0.007811601273715496, + 0.008310877718031406, + 0.013426538556814194, + 0.008506674319505692, + 0.010143321007490158, + 0.01563551276922226, + 0.012672752141952515, + 0.007830570451915264, + 0.07869695127010345, + 0.05729007348418236, + 0.012396702542901039, + 0.011206110008060932, + 0.002500249072909355, + 0.0032664018217474222, + 0.002480635419487953, + 0.0026171847712248564, + 0.003875565715134144, + 0.0039666080847382545, + 0.0036476741079241037, + 0.00437870854511857, + 0.0040704067796468735, + 0.004059372004121542, + 0.005441291723400354, + 0.005394297186285257, + 0.006721451412886381, + 0.0048274691216647625, + 0.00493268808349967, + 0.00416748970746994, + 0.003021669341251254, + 0.003409916302189231, + 0.002182759577408433, + 0.0024719666689634323, + 0.003400110639631748, + 0.0028992595616728067, + 0.0026204208843410015, + 0.0017784798983484507, + 0.0094721345230937, + 0.004171370062977076, + 0.005867332220077515, + 0.006355212070047855, + 0.004890795797109604, + 0.005548436660319567, + 0.004198634997010231, + 0.006857507396489382, + 0.0036906362511217594, + 0.0025438645388931036, + 0.01501953974366188, + 0.004604894667863846, + 0.0022388503421097994, + 0.0029239479918032885, + 0.0024995270650833845, + 0.011345161125063896, + 0.0012759488308802247, + 0.0009792789351195097, + 0.001768707763403654, + 0.0018754572374746203, + 0.0008827856509014964, + 0.0011791513534262776, + 0.0017341329948976636, + 0.0015327654546126723, + 0.0014994392404332757, + 0.0012021505972370505, + 0.0024566310457885265, + 0.0015424841549247503, + 0.0008777364855632186, + 0.0014476398937404156, + 0.0018728384748101234, + 0.0014242858160287142, + 0.04620842635631561, + 0.011374274268746376, + 0.033088281750679016, + 0.023303335532546043, + 0.03397100046277046, + 0.1510888785123825, + 0.16178421676158905, + 0.034170303493738174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.005252964794635773 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731926463544369 + ], + "opposite_logprob_tokens": [ + -5.006731986999512 + ], + "logprob_reward": -0.006731926463544369, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731926463544369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20868578553199768, + 0.4235175848007202, + 0.15516671538352966, + 0.017105959355831146, + 0.04550107195973396, + 0.04609265178442001, + 0.05802815780043602, + 0.0077409325167536736, + 0.01644759811460972, + 0.021713517606258392 + ], + "credit_map": [ + 0.024449942633509636, + 0.014000903815031052, + 0.04931703954935074, + 0.025702403858304024, + 0.01581335812807083, + 0.018930776044726372, + 0.03945772722363472, + 0.021013638004660606, + 0.025546567514538765, + 0.03231411054730415, + 0.030068758875131607, + 0.012987283058464527, + 0.15484823286533356, + 0.12486139684915543, + 0.022597691044211388, + 0.020293574780225754, + 0.020962640643119812, + 0.015229702927172184, + 0.015695353969931602, + 0.013033214956521988, + 0.031349536031484604, + 0.0305708609521389, + 0.01228991150856018, + 0.016035497188568115, + 0.0017772106220945716, + 0.002548601245507598, + 0.0016891013365238905, + 0.0018632011488080025, + 0.0034218698274344206, + 0.0026262421160936356, + 0.0013404425699263811, + 0.0018392896745353937, + 0.0026448939461261034, + 0.003720321459695697, + 0.0033243168145418167, + 0.002271513920277357, + 0.016419753432273865, + 0.012279951944947243, + 0.002447657287120819, + 0.0023926629219204187, + 0.004396448377519846, + 0.004904513712972403, + 0.004533990286290646, + 0.0038579795509576797, + 0.007811350282281637, + 0.012118963524699211, + 0.0033756662160158157, + 0.005093739368021488, + 0.0049280948005616665, + 0.0029701997991651297, + 0.01966727524995804, + 0.005864201113581657, + 0.002755086636170745, + 0.0044261920265853405, + 0.002941340673714876, + 0.014475771225988865, + 0.001082777394913137, + 0.0007226397283375263, + 0.0013093341840431094, + 0.0008236893336288631, + 0.0007099441136233509, + 0.0011274259304627776, + 0.0006791349151171744, + 0.0012859873240813613, + 0.0015977997099980712, + 0.0010081881191581488, + 0.0054628909565508366, + 0.0015915248077362776, + 0.0011385324178263545, + 0.0020225278567522764, + 0.0017503495328128338, + 0.0018757865764200687, + 0.001985291950404644, + 0.0011268986854702234, + 0.002004098379984498, + 0.0012321306858211756, + 0.0015147639205679297, + 0.005643450189381838, + 0.006554082967340946, + 0.0016528032720088959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.006731926463544369 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot, shot upward from the ground, is intermittently punctuated by distorted scan lines and brief bursts of snowy noise due to analog signal transmission glitches. The scene is set in a narrow, dank cyberpunk city back alley, the air thick with icy steam. Magenta and electric blue neon billboards flicker endlessly overhead, casting a harsh, unstable light on brick walls covered in graffiti and exposed pipes. On the ground, a bottomless pool of oily, dark green sewage is suddenly stirred by a tremendous force, and then, with an explosive gesture, a massive Australian lobster rises from the water. Its wet crimson carapace reflects a fluid, metallic sheen in the flickering neon lights, while droplets of water and foul fluids continuously flow down its segmented body and sharp back. Its huge, asymmetrical claws opened threateningly in the air, its black compound eyes flickered an unnatural glow in the shadows, and its long tentacles twitched sensitively in the humid air. The whole picture was filled with a sudden sense of grotesque horror and strong visual impact.", + "seed": 17613, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_12/index_6942_seed_17613_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.007559518679045141, + "vlm_logprob": -0.007559518679045141, + "vlm_yes_no_margin": 5.35, + "vlm_token_credit": -0.007559518679045141, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020391235128045082 + ], + "opposite_logprob_tokens": [ + -4.020391464233398 + ], + "logprob_reward": -0.020391235128045082, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.020391235128045082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2011534869670868, + 0.1699017882347107, + 0.049534063786268234, + 0.09660796076059341, + 0.11517361551523209, + 0.015217701904475689, + 0.06809011846780777, + 0.05042683705687523, + 0.21804913878440857, + 0.015845339745283127 + ], + "credit_map": [ + 0.021424511447548866, + 0.01986805349588394, + 0.025988362729549408, + 0.03050502948462963, + 0.024008100852370262, + 0.026221338659524918, + 0.03299473971128464, + 0.020143333822488785, + 0.013374453410506248, + 0.025462571531534195, + 0.0304978396743536, + 0.019152425229549408, + 0.021550122648477554, + 0.016787204891443253, + 0.0202333927154541, + 0.022843772545456886, + 0.005146439652889967, + 0.0038906424306333065, + 0.0070376647636294365, + 0.006345722358673811, + 0.005412368569523096, + 0.0055336845107376575, + 0.007636866066604853, + 0.008530672639608383, + 0.01290152594447136, + 0.009768993593752384, + 0.010843268595635891, + 0.0100716482847929, + 0.01323021948337555, + 0.017325041815638542, + 0.010418904945254326, + 0.012048366479575634, + 0.020370187237858772, + 0.014131874777376652, + 0.01439152006059885, + 0.009484546259045601, + 0.015550700016319752, + 0.014327815733850002, + 0.01034195814281702, + 0.016575010493397713, + 0.002121271099895239, + 0.0019463410135358572, + 0.001527492655441165, + 0.0022480967454612255, + 0.0019160305382683873, + 0.0016918429173529148, + 0.0022871627006679773, + 0.0014794631861150265, + 0.00880028773099184, + 0.005258381832391024, + 0.008394076488912106, + 0.007747223135083914, + 0.00787921529263258, + 0.007005613762885332, + 0.010079133324325085, + 0.012926186434924603, + 0.009603109210729599, + 0.004789543338119984, + 0.007345735561102629, + 0.004707077518105507, + 0.004989757668226957, + 0.005793463904410601, + 0.00484058540314436, + 0.008357563056051731, + 0.018600832670927048, + 0.01187460869550705, + 0.054034266620874405, + 0.011974315159022808, + 0.022036926820874214, + 0.020066823810338974, + 0.011711341328918934, + 0.06775001436471939, + 0.0019033994758501649, + 0.0015114900888875127, + 0.001854152767919004, + 0.0023969688918441534, + 0.0026289739180356264, + 0.0023056063801050186, + 0.0015207616379484534, + 0.001723986933939159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.020391235128045082 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010568277211859822 + ], + "opposite_logprob_tokens": [ + -6.876056671142578 + ], + "logprob_reward": -0.0010568277211859822, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010568277211859822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21426507830619812, + 0.20376969873905182, + 0.05535627529025078, + 0.06983942538499832, + 0.07248064875602722, + 0.04004322364926338, + 0.10422077029943466, + 0.06254895776510239, + 0.06752730906009674, + 0.10994857549667358 + ], + "credit_map": [ + 0.02337447740137577, + 0.021270612254738808, + 0.03408905491232872, + 0.031039400026202202, + 0.022985152900218964, + 0.02585005760192871, + 0.025615744292736053, + 0.03004058077931404, + 0.015313241630792618, + 0.028532864525914192, + 0.031085943803191185, + 0.022144854068756104, + 0.0332893431186676, + 0.018660886213183403, + 0.020332666113972664, + 0.03440991789102554, + 0.005953714717179537, + 0.00421435059979558, + 0.006998372729867697, + 0.007055632770061493, + 0.006402932107448578, + 0.006644182838499546, + 0.008811704814434052, + 0.009275389835238457, + 0.007544730789959431, + 0.005018223542720079, + 0.009084159508347511, + 0.011598554439842701, + 0.006799143739044666, + 0.007994726300239563, + 0.010643289424479008, + 0.011156597174704075, + 0.009505209513008595, + 0.002669375156983733, + 0.010793302208185196, + 0.008006550371646881, + 0.0033969422802329063, + 0.0060677011497318745, + 0.00876627303659916, + 0.023275300860404968, + 0.004279138054698706, + 0.003455884288996458, + 0.003933110274374485, + 0.005691480822861195, + 0.00453962292522192, + 0.006325458642095327, + 0.007185271941125393, + 0.004633256699889898, + 0.01308721024543047, + 0.008088963106274605, + 0.012973558157682419, + 0.010031825862824917, + 0.01185630727559328, + 0.012479608878493309, + 0.01513796765357256, + 0.02056533470749855, + 0.011599396355450153, + 0.005576817784458399, + 0.009360701777040958, + 0.00569786224514246, + 0.007125596050173044, + 0.006807778030633926, + 0.006878085900098085, + 0.009502721950411797, + 0.01013463456183672, + 0.005369469989091158, + 0.010657408274710178, + 0.006068140733987093, + 0.009245130233466625, + 0.010329218581318855, + 0.0065169683657586575, + 0.009206347167491913, + 0.015164953656494617, + 0.017214402556419373, + 0.013299431651830673, + 0.014416840858757496, + 0.014384389854967594, + 0.010184688493609428, + 0.010519959032535553, + 0.014763912186026573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0010568277211859822 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094195086508989 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004094195086508989, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094195086508989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1807328313589096, + 0.21040712296962738, + 0.09228515625, + 0.08787361532449722, + 0.09093067795038223, + 0.032979272305965424, + 0.028274739161133766, + 0.09174255281686783, + 0.14261268079280853, + 0.042161278426647186 + ], + "credit_map": [ + 0.014851303771138191, + 0.0147999944165349, + 0.02484213002026081, + 0.028001917526125908, + 0.02077043615281582, + 0.012039577588438988, + 0.038211606442928314, + 0.027215881273150444, + 0.01755542866885662, + 0.022151632234454155, + 0.03102828375995159, + 0.03210365027189255, + 0.023909464478492737, + 0.020270489156246185, + 0.027938341721892357, + 0.03544984385371208, + 0.00892979372292757, + 0.005708250682801008, + 0.013522886671125889, + 0.01885986514389515, + 0.007272721733897924, + 0.009796911850571632, + 0.014515583403408527, + 0.013679150491952896, + 0.010677612386643887, + 0.005044783465564251, + 0.01088270265609026, + 0.018646134063601494, + 0.005713146645575762, + 0.007538038771599531, + 0.014642992056906223, + 0.014728215523064137, + 0.009325976483523846, + 0.004357143305242062, + 0.013451489619910717, + 0.009494959376752377, + 0.005209431983530521, + 0.007249138783663511, + 0.01077307015657425, + 0.031069474294781685, + 0.0034957837779074907, + 0.00546173844486475, + 0.002739576157182455, + 0.007059621158987284, + 0.0032342844642698765, + 0.003911329433321953, + 0.003992255311459303, + 0.003084686817601323, + 0.004265670198947191, + 0.0019351547816768289, + 0.004683959763497114, + 0.0052592265419662, + 0.0021133713889867067, + 0.0032378218602389097, + 0.002910622861236334, + 0.0038689144421368837, + 0.009254531003534794, + 0.005471160169690847, + 0.012784735299646854, + 0.017427170649170876, + 0.008547985926270485, + 0.00733215594664216, + 0.01847933419048786, + 0.012445495463907719, + 0.013593272306025028, + 0.005644845776259899, + 0.03853254020214081, + 0.013761879876255989, + 0.005773942451924086, + 0.008681687526404858, + 0.011658570729196072, + 0.0449659638106823, + 0.0048390659503638744, + 0.0034182157833129168, + 0.005301013123244047, + 0.008193395100533962, + 0.0026985867880284786, + 0.004221075680106878, + 0.007022595964372158, + 0.006467332597821951 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.004094195086508989 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036176491994410753 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.0036176491994410753, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036176491994410753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14799228310585022, + 0.21706891059875488, + 0.12770098447799683, + 0.06854624301195145, + 0.12222866714000702, + 0.04870074987411499, + 0.09448443353176117, + 0.05113418772816658, + 0.08160222321748734, + 0.04054131358861923 + ], + "credit_map": [ + 0.008975398726761341, + 0.01041752565652132, + 0.03317616879940033, + 0.03321782499551773, + 0.009299020282924175, + 0.009545736014842987, + 0.02370697818696499, + 0.01965363696217537, + 0.015567203052341938, + 0.016430331394076347, + 0.0290700551122427, + 0.058931730687618256, + 0.02508838102221489, + 0.01676112413406372, + 0.0328390970826149, + 0.022380994632840157, + 0.009767775423824787, + 0.004753767047077417, + 0.01846308819949627, + 0.042605768889188766, + 0.004280793480575085, + 0.008069347590208054, + 0.024597739800810814, + 0.015162710100412369, + 0.007125079166144133, + 0.005239621736109257, + 0.008230172097682953, + 0.011811305768787861, + 0.005235128570348024, + 0.008249625563621521, + 0.010871926322579384, + 0.01178338099271059, + 0.010865379124879837, + 0.007438963744789362, + 0.015665721148252487, + 0.026362614706158638, + 0.007147836498916149, + 0.009051541797816753, + 0.024829968810081482, + 0.020866641774773598, + 0.005114424042403698, + 0.005641057621687651, + 0.0038870142307132483, + 0.013060919009149075, + 0.005566410254687071, + 0.004228856414556503, + 0.007376309484243393, + 0.0038257616106420755, + 0.016686679795384407, + 0.008241253904998302, + 0.01280535664409399, + 0.013826081529259682, + 0.009379959665238857, + 0.009404602460563183, + 0.012195679359138012, + 0.011944815516471863, + 0.0061059677973389626, + 0.00369448633864522, + 0.008094722405076027, + 0.004672285169363022, + 0.006418078206479549, + 0.006056764628738165, + 0.0056515843607485294, + 0.010440299287438393, + 0.010027850046753883, + 0.00434158556163311, + 0.01644674874842167, + 0.009084494784474373, + 0.005003550089895725, + 0.00631735660135746, + 0.009467086754739285, + 0.02091355435550213, + 0.004575194790959358, + 0.0032100651878863573, + 0.00481752585619688, + 0.009723333641886711, + 0.0022468387614935637, + 0.0035020788200199604, + 0.0063475375063717365, + 0.00611873809248209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0036176491994410753 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a narrow, dank cyberpunk city back alley with brick walls and exposed pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are magenta and electric blue neon billboards visible overhead, casting flickering light on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive Australian lobster rise from a bottomless pool of oily, dark green sewage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's carapace wet, crimson, and reflecting a fluid, metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low angle looking upward, punctuated by distorted scan lines and snowy noise?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637686260044575 + ], + "opposite_logprob_tokens": [ + -4.75863790512085 + ], + "logprob_reward": -0.008637686260044575, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637686260044575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3006575405597687, + 0.16936838626861572, + 0.09568654745817184, + 0.06191491708159447, + 0.03244580328464508, + 0.022717377170920372, + 0.11845294386148453, + 0.10284881293773651, + 0.05582384020090103, + 0.040083907544612885 + ], + "credit_map": [ + 0.02193022146821022, + 0.017669562250375748, + 0.042893726378679276, + 0.05387118458747864, + 0.02783489041030407, + 0.023820428177714348, + 0.060697562992572784, + 0.0519399493932724, + 0.012486157938838005, + 0.01828879676759243, + 0.02220679633319378, + 0.029051244258880615, + 0.02757330983877182, + 0.01847277395427227, + 0.023293757811188698, + 0.017995532602071762, + 0.014001647010445595, + 0.0046704355627298355, + 0.012955225072801113, + 0.02375658228993416, + 0.006031661294400692, + 0.009620357304811478, + 0.016116047278046608, + 0.008534576743841171, + 0.0055947983637452126, + 0.004708331543952227, + 0.008298087865114212, + 0.00881283637136221, + 0.005362704861909151, + 0.005443890579044819, + 0.009679174050688744, + 0.014015082269906998, + 0.00405843323096633, + 0.002189398044720292, + 0.006183839403092861, + 0.0037355974782258272, + 0.0030095435213297606, + 0.002750672399997711, + 0.004414066206663847, + 0.0061042532324790955, + 0.003233329625800252, + 0.0017153279623016715, + 0.0020566200837492943, + 0.00350695732049644, + 0.0022505782544612885, + 0.003332975087687373, + 0.00394994905218482, + 0.0026716350112110376, + 0.01548784039914608, + 0.009974763728678226, + 0.01456182450056076, + 0.011687417514622211, + 0.0148993618786335, + 0.017802925780415535, + 0.013801970519125462, + 0.020236823707818985, + 0.011811185628175735, + 0.007580758538097143, + 0.012358328327536583, + 0.013090895488858223, + 0.010715148411691189, + 0.012791190296411514, + 0.018633896484971046, + 0.01586739905178547, + 0.006823996547609568, + 0.003946022596210241, + 0.009946910664439201, + 0.006737705320119858, + 0.006387125700712204, + 0.005355461034923792, + 0.008064244873821735, + 0.008562366478145123, + 0.004461104050278664, + 0.002475195797160268, + 0.005556998308748007, + 0.008971292525529861, + 0.0034717535600066185, + 0.0035000438801944256, + 0.005861992482095957, + 0.005785524379462004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.008637686260044575 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_13/index_tensor([6865, 6865])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_13/index_tensor([6865, 6865])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6b73b54046a2e4c9d73046755dc83a24d8fda45a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_13/index_tensor([6865, 6865])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle close-up shot captured with a telephoto lens, the camera slowly tilts upward, tracking a massive rectangular stone monument as it rises majestically and inexorably from a forest floor blanketed in damp black soil and decaying leaves. The monument is constructed of weathered, dark gray granite, its surface covered in mottled emerald green and deep black moss clinging to its cracks and damp areas. As the monument rises, dry dust and fine stone chips are shaken from the ancient inscriptions, eroded by time and nearly unreadable. Like tiny waterfalls of sand, they flow down its rough surface, visible in the few faint, grayish-white rays of light filtering through the dense canopy. The telephoto lens compresses space, making the blurred, gnarled ancient trees in the background appear both distant and oppressive. The entire scene, shrouded in a near-stagnant silence, is imbued with an atmosphere of antiquity, mystery, and solemnity.", + "seed": 8410, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_13/index_6865_seed_8410_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.055383218452334404, + "vlm_logprob": -0.055383218452334404, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.055383218452334404, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07313498854637146 + ], + "opposite_logprob_tokens": [ + -2.8231348991394043 + ], + "logprob_reward": -0.07313498854637146, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07313498854637146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15418456494808197, + 0.041162874549627304, + 0.5970667600631714, + 0.07304714620113373, + 0.029868464916944504, + 0.040571458637714386, + 0.015790849924087524, + 0.02901460789144039, + 0.010308285243809223, + 0.00898496713489294 + ], + "credit_map": [ + 0.010852969251573086, + 0.015721000730991364, + 0.03957544267177582, + 0.049082476645708084, + 0.013725081458687782, + 0.007647755555808544, + 0.009332996793091297, + 0.008246834389865398, + 0.0034883797634392977, + 0.007665247656404972, + 0.010648073628544807, + 0.005120656453073025, + 0.0035290594678372145, + 0.002772561740130186, + 0.005508685018867254, + 0.002430212451145053, + 0.021640341728925705, + 0.17656391859054565, + 0.05781308189034462, + 0.2451188713312149, + 0.02827843837440014, + 0.013363629579544067, + 0.03778362646698952, + 0.016504867002367973, + 0.0055105555802583694, + 0.01777079328894615, + 0.0065102530643343925, + 0.010197912342846394, + 0.013071730732917786, + 0.005273620598018169, + 0.009244409389793873, + 0.00546787166967988, + 0.002172225620597601, + 0.003928057383745909, + 0.0032287132926285267, + 0.004531990736722946, + 0.005357138812541962, + 0.0020099456887692213, + 0.005792334210127592, + 0.0028480617329478264, + 0.003618521150201559, + 0.004469345323741436, + 0.004480346571654081, + 0.005385607481002808, + 0.008502453565597534, + 0.0030934163369238377, + 0.007821477949619293, + 0.003200291423127055, + 0.0022111511789262295, + 0.001901910058222711, + 0.002063438296318054, + 0.001972053898498416, + 0.0017936069052666426, + 0.0021815223153680563, + 0.0018908223137259483, + 0.0017763463547453284, + 0.001624074880965054, + 0.0024254347663372755, + 0.004341722931712866, + 0.004889351315796375, + 0.004856804385781288, + 0.0014113192446529865, + 0.0070196189917624, + 0.0024462819565087557, + 0.0007052120636217296, + 0.001046700868755579, + 0.001449208240956068, + 0.0013765007024630904, + 0.0014250099193304777, + 0.0008462888654321432, + 0.002121256198734045, + 0.0013381096068769693, + 0.000880443665664643, + 0.0010516962502151728, + 0.001219296595081687, + 0.001188946538604796, + 0.0014067217707633972, + 0.0009585337829776108, + 0.0016315921675413847, + 0.0006477368297055364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.07313498854637146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498972535133362 + ], + "opposite_logprob_tokens": [ + -2.929989814758301 + ], + "logprob_reward": -0.05498972535133362, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498972535133362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06238338723778725, + 0.04420540854334831, + 0.7349553108215332, + 0.024358194321393967, + 0.024037864059209824, + 0.03779647871851921, + 0.028072848916053772, + 0.0338103212416172, + 0.0039993766695261, + 0.006380817387253046 + ], + "credit_map": [ + 0.010065102949738503, + 0.006522746756672859, + 0.008140425197780132, + 0.010796072892844677, + 0.007214734796434641, + 0.005581354722380638, + 0.006345516070723534, + 0.007717431057244539, + 0.0029885342810302973, + 0.007830287329852581, + 0.013996089808642864, + 0.0055394358932971954, + 0.003782750805839896, + 0.0020514281932264566, + 0.0058593060821294785, + 0.0021575733553618193, + 0.026008211076259613, + 0.22003422677516937, + 0.07208558917045593, + 0.30424249172210693, + 0.03394114598631859, + 0.014268961735069752, + 0.044461119920015335, + 0.019913583993911743, + 0.001875035697594285, + 0.004362352658063173, + 0.0029814396984875202, + 0.003589313244447112, + 0.003061239141970873, + 0.0022896795999258757, + 0.003599543822929263, + 0.002599589293822646, + 0.0015613185241818428, + 0.004700315650552511, + 0.0026388010010123253, + 0.0033858397509902716, + 0.004486085381358862, + 0.0017849182477220893, + 0.003804525127634406, + 0.0016760571161285043, + 0.0029002083465456963, + 0.005451378878206015, + 0.004333668388426304, + 0.00455956207588315, + 0.007642312906682491, + 0.0033056559041142464, + 0.006772994063794613, + 0.0028306927997618914, + 0.004033379722386599, + 0.0029929978772997856, + 0.004098464734852314, + 0.003069024533033371, + 0.0028628103900700808, + 0.00381705304607749, + 0.0032827071845531464, + 0.003916412126272917, + 0.002417563460767269, + 0.002926537301391363, + 0.004837125074118376, + 0.005454119294881821, + 0.004635360557585955, + 0.002588891191408038, + 0.008353106677532196, + 0.0025976160541176796, + 0.0004279470886103809, + 0.0004209309699945152, + 0.0005185524933040142, + 0.00056023511569947, + 0.0005530192865990102, + 0.00039021859993226826, + 0.0007704790332354605, + 0.0003579938493203372, + 0.0005955987144261599, + 0.0008135466487146914, + 0.0011551030911505222, + 0.0008847733261063695, + 0.0008698227466084063, + 0.0005871783941984177, + 0.0009155673906207085, + 0.0005592263187281787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.05498972535133362 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330542728304863 + ], + "opposite_logprob_tokens": [ + -3.7733054161071777 + ], + "logprob_reward": -0.02330542728304863, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330542728304863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15617981553077698, + 0.03067190572619438, + 0.5817610621452332, + 0.0721035823225975, + 0.05723417550325394, + 0.04343035817146301, + 0.016536764800548553, + 0.030102260410785675, + 0.005871947854757309, + 0.006108148954808712 + ], + "credit_map": [ + 0.010294553823769093, + 0.014490910805761814, + 0.04455113410949707, + 0.05096512287855148, + 0.012073932215571404, + 0.006699930410832167, + 0.009028389118611813, + 0.008075844496488571, + 0.003427698975428939, + 0.005630291532725096, + 0.005109299439936876, + 0.0036281368229538202, + 0.003791382070630789, + 0.0025377138517796993, + 0.004421117715537548, + 0.0021262667141854763, + 0.020072149112820625, + 0.1714687943458557, + 0.05502054840326309, + 0.23547524213790894, + 0.028874311596155167, + 0.014476761221885681, + 0.03910929337143898, + 0.017263993620872498, + 0.005287048872560263, + 0.015304969623684883, + 0.006445892155170441, + 0.010654889047145844, + 0.011529302224516869, + 0.006691587157547474, + 0.010151578113436699, + 0.00603831373155117, + 0.003946341574192047, + 0.012009130790829659, + 0.006761155556887388, + 0.008493512868881226, + 0.008876964449882507, + 0.0037829901557415724, + 0.00846417061984539, + 0.004899910185486078, + 0.002773669082671404, + 0.004511155188083649, + 0.005233657080680132, + 0.005613779183477163, + 0.00907338596880436, + 0.003085590898990631, + 0.009412226267158985, + 0.003726891940459609, + 0.0024448337499052286, + 0.0019986124243587255, + 0.0020759215112775564, + 0.0019191455794498324, + 0.0023451121523976326, + 0.001704871072433889, + 0.0021189949475228786, + 0.0019292725482955575, + 0.0022821880411356688, + 0.0027736411429941654, + 0.004223995376378298, + 0.0046350122429430485, + 0.005466545466333628, + 0.0018304676050320268, + 0.006353258155286312, + 0.0025371520314365625, + 0.00032290059607475996, + 0.0005603161407634616, + 0.000824387650936842, + 0.000863756809849292, + 0.0007271562935784459, + 0.0003588969411794096, + 0.0013860482722520828, + 0.0008284851210191846, + 0.0005143446614965796, + 0.0007690540514886379, + 0.0007636596565134823, + 0.0008572462829761207, + 0.001096803112886846, + 0.00055363355204463, + 0.0010961167281493545, + 0.00045729088014923036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.02330542728304863 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294565349817276 + ], + "opposite_logprob_tokens": [ + -2.237945556640625 + ], + "logprob_reward": -0.11294565349817276, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294565349817276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.016100138425827026, + 0.03148047626018524, + 0.8131846189498901, + 0.025918571278452873, + 0.04640630632638931, + 0.028355766087770462, + 0.019464785233139992, + 0.00947648473083973, + 0.0052407835610210896, + 0.004372024908661842 + ], + "credit_map": [ + 0.0019646482542157173, + 0.002272643381729722, + 0.0019570982549339533, + 0.002033552387729287, + 0.0022359751164913177, + 0.001781468978151679, + 0.0018630876438692212, + 0.0019916645251214504, + 0.0019174288026988506, + 0.005234086886048317, + 0.009156164713203907, + 0.0037977148313075304, + 0.0036579773295670748, + 0.0016312742372974753, + 0.0043639265932142735, + 0.0017219026340171695, + 0.028869779780507088, + 0.24985623359680176, + 0.07698270678520203, + 0.3352743089199066, + 0.039085566997528076, + 0.015381377190351486, + 0.04744906350970268, + 0.02028564363718033, + 0.0016226330772042274, + 0.004020643886178732, + 0.003083149902522564, + 0.003973298240453005, + 0.004101227968931198, + 0.002957800170406699, + 0.0038229201454669237, + 0.0023368974216282368, + 0.0033705602400004864, + 0.009578326717019081, + 0.005063088610768318, + 0.005988614168018103, + 0.008194122463464737, + 0.0025782096199691296, + 0.007306921295821667, + 0.004326462280005217, + 0.0017159085255116224, + 0.0032399999909102917, + 0.0035770817194133997, + 0.003638381836935878, + 0.005726756528019905, + 0.0021592341363430023, + 0.005886291153728962, + 0.002412112895399332, + 0.00238577788695693, + 0.00313496682792902, + 0.002330514369532466, + 0.002643665997311473, + 0.003389309160411358, + 0.001857903553172946, + 0.0024796389043331146, + 0.0012430099304765463, + 0.0009528638329356909, + 0.0010550591396167874, + 0.0018472138326615095, + 0.0013644970022141933, + 0.0012018632842227817, + 0.0009101324831135571, + 0.0012285088887438178, + 0.0009163454524241388, + 0.0002600054140202701, + 0.0005398268694989383, + 0.000697957759257406, + 0.0007896458264440298, + 0.0007349333609454334, + 0.0002725774538703263, + 0.0013480656780302525, + 0.0005977711407467723, + 0.00046235101763159037, + 0.0004919726634398103, + 0.000555042177438736, + 0.0005467270966619253, + 0.0007251405040733516, + 0.00038870747084729373, + 0.000823161390144378, + 0.0003789227339439094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.11294565349817276 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540297582745552 + ], + "opposite_logprob_tokens": [ + -4.387540340423584 + ], + "logprob_reward": -0.012540297582745552, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540297582745552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08618083596229553, + 0.031672898679971695, + 0.4136286973953247, + 0.2149912267923355, + 0.0571841262280941, + 0.07752776145935059, + 0.0400821715593338, + 0.0427098423242569, + 0.012649690732359886, + 0.02337273396551609 + ], + "credit_map": [ + 0.008261436596512794, + 0.010442161932587624, + 0.018938705325126648, + 0.021885262802243233, + 0.010414103977382183, + 0.005624815821647644, + 0.005973256193101406, + 0.004641098901629448, + 0.0025135171599686146, + 0.0057019805535674095, + 0.006706980988383293, + 0.0038995954673737288, + 0.0030986384954303503, + 0.002573051257058978, + 0.004809832666069269, + 0.0023693046532571316, + 0.015048804692924023, + 0.12495438754558563, + 0.039745405316352844, + 0.1707458347082138, + 0.019577961415052414, + 0.00785438995808363, + 0.025196298956871033, + 0.01050560176372528, + 0.013565672561526299, + 0.0549306757748127, + 0.018941227346658707, + 0.028292033821344376, + 0.04550221562385559, + 0.015051502734422684, + 0.025005368515849113, + 0.013702528551220894, + 0.004676736891269684, + 0.008155585266649723, + 0.006804449483752251, + 0.009346261620521545, + 0.007758812978863716, + 0.006676838733255863, + 0.00850951112806797, + 0.005255930125713348, + 0.004732257686555386, + 0.009852016344666481, + 0.009352312423288822, + 0.00976733397692442, + 0.015929700806736946, + 0.005692111793905497, + 0.015657572075724602, + 0.006544460076838732, + 0.004938229452818632, + 0.005515855271369219, + 0.006090790033340454, + 0.00476880045607686, + 0.004673253744840622, + 0.00409208657220006, + 0.005281270947307348, + 0.004721884615719318, + 0.0032471800222992897, + 0.004478891845792532, + 0.0053785331547260284, + 0.006188890431076288, + 0.008119386620819569, + 0.0034625264815986156, + 0.008329399861395359, + 0.0035050322767347097, + 0.0010864523937925696, + 0.0012917593121528625, + 0.0016263504512608051, + 0.0018558442825451493, + 0.0019212729530408978, + 0.0011657768627628684, + 0.0024611379485577345, + 0.0012410966446623206, + 0.0030081619042903185, + 0.002428104169666767, + 0.003457955550402403, + 0.0030138150323182344, + 0.002629367168992758, + 0.002887123264372349, + 0.003531601745635271, + 0.0024166046641767025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.012540297582745552 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up shot captured with a telephoto lens, the camera slowly tilts upward, tracking a massive rectangular stone monument as it rises majestically and inexorably from a forest floor blanketed in damp black soil and decaying leaves. The monument is constructed of weathered, dark gray granite, its surface covered in mottled emerald green and deep black moss clinging to its cracks and damp areas. As the monument rises, dry dust and fine stone chips are shaken from the ancient inscriptions, eroded by time and nearly unreadable. Like tiny waterfalls of sand, they flow down its rough surface, visible in the few faint, grayish-white rays of light filtering through the dense canopy. The telephoto lens compresses space, making the blurred, gnarled ancient trees in the background appear both distant and oppressive. The entire scene, shrouded in a near-stagnant silence, is imbued with an atmosphere of antiquity, mystery, and solemnity.", + "seed": 8411, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_13/index_6865_seed_8411_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02965650148689747, + "vlm_logprob": -0.02965650148689747, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.02965650148689747, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0192894097417593 + ], + "opposite_logprob_tokens": [ + -4.394289493560791 + ], + "logprob_reward": -0.0192894097417593, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0192894097417593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18956655263900757, + 0.05196800082921982, + 0.3339979946613312, + 0.0975278913974762, + 0.06758315116167068, + 0.048203110694885254, + 0.03886568918824196, + 0.04308729246258736, + 0.07810888439416885, + 0.05109137296676636 + ], + "credit_map": [ + 0.009938950650393963, + 0.012866973876953125, + 0.009620903059840202, + 0.00962086021900177, + 0.05611966550350189, + 0.06633877754211426, + 0.010477123782038689, + 0.014583305455744267, + 0.005246998276561499, + 0.010295125655829906, + 0.007682752329856157, + 0.006743832491338253, + 0.005761298816651106, + 0.005450497847050428, + 0.005712002981454134, + 0.005075495224446058, + 0.01593766175210476, + 0.02160402201116085, + 0.037581268697977066, + 0.025966089218854904, + 0.04769960045814514, + 0.012854771688580513, + 0.12762445211410522, + 0.04473014548420906, + 0.0040234969928860664, + 0.013943427242338657, + 0.008337018080055714, + 0.029480336233973503, + 0.009272290393710136, + 0.008194204419851303, + 0.013461376540362835, + 0.010815739631652832, + 0.004715620540082455, + 0.008318946696817875, + 0.008316781371831894, + 0.013231584802269936, + 0.01053459569811821, + 0.004943661857396364, + 0.008992074057459831, + 0.008529890328645706, + 0.0038661241997033358, + 0.010733646340668201, + 0.006159971002489328, + 0.0069937012158334255, + 0.006522544659674168, + 0.003698621643707156, + 0.005838312208652496, + 0.0043901922181248665, + 0.0031836426351219416, + 0.004508530255407095, + 0.003669269150123, + 0.005677116569131613, + 0.007813634350895882, + 0.0028759778942912817, + 0.005222202744334936, + 0.005915315356105566, + 0.002142607234418392, + 0.0034326985478401184, + 0.0026136913802474737, + 0.009838605299592018, + 0.006865154020488262, + 0.003398351604118943, + 0.010240389965474606, + 0.004555798135697842, + 0.00744247529655695, + 0.00828706007450819, + 0.0057486980222165585, + 0.00898239016532898, + 0.013190396130084991, + 0.008253113366663456, + 0.013472111895680428, + 0.012732640840113163, + 0.0037557377945631742, + 0.005032026208937168, + 0.0055730040185153484, + 0.0087751355022192, + 0.009243347682058811, + 0.004153692163527012, + 0.010385244153439999, + 0.004173186141997576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.0192894097417593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982921525835991 + ], + "opposite_logprob_tokens": [ + -3.529829263687134 + ], + "logprob_reward": -0.02982921525835991, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982921525835991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08925464004278183, + 0.08038928359746933, + 0.15939505398273468, + 0.12832044064998627, + 0.05515463277697563, + 0.06082294508814812, + 0.05425240471959114, + 0.04615505039691925, + 0.1614890694618225, + 0.16476644575595856 + ], + "credit_map": [ + 0.004940683953464031, + 0.008533667773008347, + 0.004310460761189461, + 0.0034733160864561796, + 0.030459610745310783, + 0.027255408465862274, + 0.005249429028481245, + 0.005032071843743324, + 0.00812497641891241, + 0.015386145561933517, + 0.012261992320418358, + 0.008840945549309254, + 0.007067047990858555, + 0.009787534363567829, + 0.01090375054627657, + 0.008016897365450859, + 0.006845312658697367, + 0.025230899453163147, + 0.018453100696206093, + 0.012351879850029945, + 0.00964379496872425, + 0.007214068900793791, + 0.06411737948656082, + 0.01553863100707531, + 0.005089945625513792, + 0.021856095641851425, + 0.012651854194700718, + 0.036360081285238266, + 0.015792816877365112, + 0.008368336595594883, + 0.016783155500888824, + 0.011418180540204048, + 0.003157079918310046, + 0.007576895412057638, + 0.006307682022452354, + 0.01113766897469759, + 0.010338311083614826, + 0.0048814513720571995, + 0.00637351768091321, + 0.005382033530622721, + 0.00443057157099247, + 0.01071209367364645, + 0.007904446683824062, + 0.006904342211782932, + 0.013422504998743534, + 0.005193930584937334, + 0.007288228254765272, + 0.004966831766068935, + 0.008401177823543549, + 0.005372818559408188, + 0.008323179557919502, + 0.009590409696102142, + 0.007001521531492472, + 0.005591595079749823, + 0.005601690150797367, + 0.00437002070248127, + 0.0023582996800541878, + 0.004026261158287525, + 0.002447563223540783, + 0.01035352423787117, + 0.010033241473138332, + 0.0037341054994612932, + 0.009020818397402763, + 0.004181241616606712, + 0.008311448618769646, + 0.008258171379566193, + 0.010132435709238052, + 0.04447023198008537, + 0.01648748852312565, + 0.007765272632241249, + 0.05363602936267853, + 0.012427997775375843, + 0.012091139331459999, + 0.010482572950422764, + 0.012164326384663582, + 0.03537064790725708, + 0.019362930208444595, + 0.015467827208340168, + 0.04581526294350624, + 0.014011755585670471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.02982921525835991 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012552187778055668 + ], + "opposite_logprob_tokens": [ + -4.387552261352539 + ], + "logprob_reward": -0.012552187778055668, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012552187778055668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1555641144514084, + 0.07710167020559311, + 0.4090900719165802, + 0.05276469141244888, + 0.05679414048790932, + 0.05170898512005806, + 0.09115307778120041, + 0.033898334950208664, + 0.029792848974466324, + 0.04213210567831993 + ], + "credit_map": [ + 0.01094227284193039, + 0.01442980207502842, + 0.008462387137115002, + 0.01329062320291996, + 0.029915116727352142, + 0.04748351126909256, + 0.012811875902116299, + 0.018228529021143913, + 0.0070954784750938416, + 0.011044985614717007, + 0.009155736304819584, + 0.011992469429969788, + 0.006624375004321337, + 0.010470006614923477, + 0.013510305434465408, + 0.007208319380879402, + 0.017401158809661865, + 0.11557700484991074, + 0.03969219699501991, + 0.026884563267230988, + 0.0517941489815712, + 0.014916520565748215, + 0.12026922404766083, + 0.02255525253713131, + 0.0030037867836654186, + 0.008611079305410385, + 0.005584988743066788, + 0.014031296595931053, + 0.007181561551988125, + 0.003253163304179907, + 0.0063931443728506565, + 0.004705674480646849, + 0.004911148454993963, + 0.007386540528386831, + 0.006631866097450256, + 0.010647395625710487, + 0.008987556211650372, + 0.00543771218508482, + 0.007057005073875189, + 0.005734918173402548, + 0.0035937028005719185, + 0.014302913099527359, + 0.004297396168112755, + 0.006790278013795614, + 0.007844769395887852, + 0.004965119995176792, + 0.005658336915075779, + 0.0042564705945551395, + 0.012291966937482357, + 0.008904962800443172, + 0.014173387549817562, + 0.015598606318235397, + 0.010664504021406174, + 0.01064405869692564, + 0.008794190362095833, + 0.010081402026116848, + 0.0038051381707191467, + 0.003367939731106162, + 0.003468332113698125, + 0.0045349677093327045, + 0.00512648094445467, + 0.003637610236182809, + 0.006611446850001812, + 0.003346417797729373, + 0.0025348204653710127, + 0.001708106487058103, + 0.003005714388564229, + 0.003723932895809412, + 0.0049244435504078865, + 0.0019755421672016382, + 0.008027199655771255, + 0.0038930904120206833, + 0.0016288224142044783, + 0.0024453008081763983, + 0.002054848475381732, + 0.01042510848492384, + 0.006669369526207447, + 0.003337803529575467, + 0.012293487787246704, + 0.0032773632556200027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.012552187778055668 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789823904633522 + ], + "opposite_logprob_tokens": [ + -2.578982353210449 + ], + "logprob_reward": -0.0789823904633522, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789823904633522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10168257355690002, + 0.059455856680870056, + 0.1492445170879364, + 0.01484387181699276, + 0.04545474797487259, + 0.032677531242370605, + 0.04679353907704353, + 0.06240390986204147, + 0.15611602365970612, + 0.33132752776145935 + ], + "credit_map": [ + 0.005780432373285294, + 0.008860118687152863, + 0.006412379443645477, + 0.006737194024026394, + 0.03174835443496704, + 0.026968279853463173, + 0.008125361055135727, + 0.00705045135691762, + 0.005752252414822578, + 0.00842318031936884, + 0.008537924848496914, + 0.006907766219228506, + 0.006443264428526163, + 0.008802318014204502, + 0.008028346113860607, + 0.006560801062732935, + 0.006355365738272667, + 0.02390093356370926, + 0.01862693391740322, + 0.010710282251238823, + 0.008714761584997177, + 0.0063416617922484875, + 0.058821532875299454, + 0.015773028135299683, + 0.00114375667180866, + 0.0020442602690309286, + 0.0018654187442734838, + 0.002058088779449463, + 0.002441325457766652, + 0.0018462822772562504, + 0.0019041489576920867, + 0.0015405892627313733, + 0.0025573712773621082, + 0.0066015031188726425, + 0.0041590407490730286, + 0.007829328067600727, + 0.01024760864675045, + 0.004532465245574713, + 0.004949792753905058, + 0.004577636253088713, + 0.0026248223148286343, + 0.005305016878992319, + 0.002950544934719801, + 0.004076460376381874, + 0.006956906523555517, + 0.0029779679607599974, + 0.004725807812064886, + 0.0030600037425756454, + 0.0038543089758604765, + 0.004179947543889284, + 0.0035962536931037903, + 0.00958109088242054, + 0.010206509381532669, + 0.004204932600259781, + 0.0056747389025986195, + 0.005495755467563868, + 0.002168065868318081, + 0.004326635040342808, + 0.002108590677380562, + 0.02113865315914154, + 0.006455251947045326, + 0.0025439015589654446, + 0.01981658861041069, + 0.0038462199736386538, + 0.00903661921620369, + 0.008940929546952248, + 0.011428373865783215, + 0.04173440486192703, + 0.016782958060503006, + 0.008110583759844303, + 0.0505988746881485, + 0.00948326662182808, + 0.020806334912776947, + 0.023545492440462112, + 0.01717080920934677, + 0.06703972071409225, + 0.06436106562614441, + 0.02575927972793579, + 0.09157568216323853, + 0.0210691150277853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.0789823904633522 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629304192960262 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.007629304192960262, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629304192960262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18177355825901031, + 0.13383729755878448, + 0.07193731516599655, + 0.04270228371024132, + 0.10765988379716873, + 0.04487502574920654, + 0.04429428651928902, + 0.049629583954811096, + 0.18346580862998962, + 0.13982489705085754 + ], + "credit_map": [ + 0.008824368007481098, + 0.009957448579370975, + 0.009339847601950169, + 0.00884242169559002, + 0.06284363567829132, + 0.05835775285959244, + 0.011585701256990433, + 0.012022379785776138, + 0.01455786731094122, + 0.015314855612814426, + 0.020780988037586212, + 0.02260204777121544, + 0.008841464295983315, + 0.01947721652686596, + 0.021395333111286163, + 0.010867536999285221, + 0.003409448079764843, + 0.012595048174262047, + 0.008159885182976723, + 0.006732031237334013, + 0.005315918009728193, + 0.004341469146311283, + 0.02433202974498272, + 0.007051480934023857, + 0.0021203779615461826, + 0.008865345269441605, + 0.005207496229559183, + 0.009308182634413242, + 0.006150056142359972, + 0.0029128624591976404, + 0.004992080386728048, + 0.00314588425680995, + 0.007445953320711851, + 0.012152648530900478, + 0.01505633257329464, + 0.02216506004333496, + 0.015349636785686016, + 0.008550148457288742, + 0.01813662052154541, + 0.008803486824035645, + 0.004668207373470068, + 0.007551942951977253, + 0.006436219904571772, + 0.004971577785909176, + 0.008334952406585217, + 0.00408064853399992, + 0.0052297404035925865, + 0.003601731499657035, + 0.005336202681064606, + 0.004629075061529875, + 0.005768706556409597, + 0.007691442966461182, + 0.0048833005130290985, + 0.007450057193636894, + 0.004447117447853088, + 0.004088384564965963, + 0.0026935068890452385, + 0.003903697244822979, + 0.003001539269462228, + 0.012075033038854599, + 0.005713154561817646, + 0.0049182698130607605, + 0.01096119824796915, + 0.006363180465996265, + 0.012757638469338417, + 0.010587435215711594, + 0.016154246404767036, + 0.050123944878578186, + 0.015240137465298176, + 0.011447958648204803, + 0.053267497569322586, + 0.01388696301728487, + 0.008400053717195988, + 0.0071603767573833466, + 0.01180027425289154, + 0.03276349604129791, + 0.013992249965667725, + 0.013870049268007278, + 0.04091397300362587, + 0.010924427770078182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.007629304192960262 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up shot captured with a telephoto lens, the camera slowly tilts upward, tracking a massive rectangular stone monument as it rises majestically and inexorably from a forest floor blanketed in damp black soil and decaying leaves. The monument is constructed of weathered, dark gray granite, its surface covered in mottled emerald green and deep black moss clinging to its cracks and damp areas. As the monument rises, dry dust and fine stone chips are shaken from the ancient inscriptions, eroded by time and nearly unreadable. Like tiny waterfalls of sand, they flow down its rough surface, visible in the few faint, grayish-white rays of light filtering through the dense canopy. The telephoto lens compresses space, making the blurred, gnarled ancient trees in the background appear both distant and oppressive. The entire scene, shrouded in a near-stagnant silence, is imbued with an atmosphere of antiquity, mystery, and solemnity.", + "seed": 8412, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_13/index_6865_seed_8412_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04150550216436386, + "vlm_logprob": -0.04150550216436386, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.04150550216436386, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019735872745513916 + ], + "opposite_logprob_tokens": [ + -4.519735813140869 + ], + "logprob_reward": -0.019735872745513916, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.019735872745513916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13683608174324036, + 0.024913888424634933, + 0.02229558490216732, + 0.046601273119449615, + 0.08732113987207413, + 0.01938239112496376, + 0.029954638332128525, + 0.2979954779148102, + 0.03714267164468765, + 0.2975568175315857 + ], + "credit_map": [ + 0.02345813624560833, + 0.022873930633068085, + 0.017300626263022423, + 0.0168945100158453, + 0.016446849331259727, + 0.011903275735676289, + 0.014019080437719822, + 0.01393968891352415, + 0.0025721602141857147, + 0.002617760794237256, + 0.002596086822450161, + 0.004271689802408218, + 0.0035612257197499275, + 0.00242090062238276, + 0.003076875116676092, + 0.0037971902638673782, + 0.0013121486408635974, + 0.0029896514024585485, + 0.0027355605270713568, + 0.003460767213255167, + 0.0026745533104985952, + 0.0020271027460694313, + 0.005123324692249298, + 0.0019724767189472914, + 0.004552328027784824, + 0.0033912588842213154, + 0.006756596267223358, + 0.015290250070393085, + 0.002835265127941966, + 0.0024276694748550653, + 0.007706710137426853, + 0.003641190705820918, + 0.0030496534891426563, + 0.0029142391867935658, + 0.007302033714950085, + 0.022625522688031197, + 0.00709132757037878, + 0.003046209691092372, + 0.03589107096195221, + 0.005401088390499353, + 0.000899437814950943, + 0.0015407877508550882, + 0.0018848065519705415, + 0.0030679996125400066, + 0.0031523958314210176, + 0.0016488240798935294, + 0.005272529553622007, + 0.0019156098132953048, + 0.0014141176361590624, + 0.001739798579365015, + 0.0033141938038170338, + 0.006752392277121544, + 0.0031513262074440718, + 0.0016998467035591602, + 0.008806968107819557, + 0.0030759964138269424, + 0.004940069280564785, + 0.008120305836200714, + 0.027530884370207787, + 0.12826043367385864, + 0.007966701872646809, + 0.008784480392932892, + 0.10052981972694397, + 0.011862807907164097, + 0.0027111710514873266, + 0.003499483224004507, + 0.007173823192715645, + 0.0064992839470505714, + 0.004662077873945236, + 0.003545887302607298, + 0.004251537844538689, + 0.004799410235136747, + 0.01275703590363264, + 0.009277704171836376, + 0.07641634345054626, + 0.12478604167699814, + 0.00698614539578557, + 0.014231123961508274, + 0.038925331085920334, + 0.014177090488374233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.019735872745513916 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637220360338688 + ], + "opposite_logprob_tokens": [ + -3.651372194290161 + ], + "logprob_reward": -0.02637220360338688, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637220360338688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038655608892440796, + 0.016677886247634888, + 0.01623733900487423, + 0.0405309684574604, + 0.10739343613386154, + 0.02105129323899746, + 0.029450634494423866, + 0.3434743583202362, + 0.0644589513540268, + 0.3220694959163666 + ], + "credit_map": [ + 0.003013857174664736, + 0.006402383558452129, + 0.00584814790636301, + 0.005178573541343212, + 0.005565758794546127, + 0.004646610468626022, + 0.004162044730037451, + 0.0038382394704967737, + 0.0019319275161251426, + 0.0016112372977659106, + 0.0015911224763840437, + 0.0022087260149419308, + 0.001498509431257844, + 0.0016000973992049694, + 0.0030408469028770924, + 0.003195421304553747, + 0.0008062621927820146, + 0.0018101745517924428, + 0.0016559207579120994, + 0.0037551678251475096, + 0.0018454876262694597, + 0.0014303469797596335, + 0.0036538797430694103, + 0.001280102995224297, + 0.004339680541306734, + 0.003234029980376363, + 0.004897942300885916, + 0.01033819280564785, + 0.003882189281284809, + 0.0030834113713353872, + 0.006351165473461151, + 0.004404361825436354, + 0.004571876954287291, + 0.003927337005734444, + 0.008729171939194202, + 0.026240766048431396, + 0.00743145402520895, + 0.004148051142692566, + 0.041612256318330765, + 0.010732538066804409, + 0.0010495390743017197, + 0.0011507985182106495, + 0.00163427391089499, + 0.005242700222879648, + 0.002190870000049472, + 0.0012053351383656263, + 0.007301222998648882, + 0.0012765571009367704, + 0.0014345466624945402, + 0.0019306308822706342, + 0.004170150961726904, + 0.006156282499432564, + 0.0031914941500872374, + 0.00203684251755476, + 0.006917007267475128, + 0.0036136843264102936, + 0.00573124410584569, + 0.009547115303575993, + 0.024332499131560326, + 0.14657346904277802, + 0.009224805980920792, + 0.010269022546708584, + 0.12484986335039139, + 0.01294640637934208, + 0.004032932221889496, + 0.002678604330867529, + 0.01979728415608406, + 0.009240223094820976, + 0.0062479437328875065, + 0.006105361040681601, + 0.010499829426407814, + 0.0058567882515490055, + 0.012099293991923332, + 0.008116439916193485, + 0.1302156150341034, + 0.1136038675904274, + 0.004946899600327015, + 0.017875496298074722, + 0.026746610179543495, + 0.00846533477306366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.02637220360338688 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023300303146243095 + ], + "opposite_logprob_tokens": [ + -3.7733004093170166 + ], + "logprob_reward": -0.023300303146243095, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023300303146243095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046774972230196, + 0.02346978150308132, + 0.02903035655617714, + 0.09299684315919876, + 0.06440990418195724, + 0.04013645648956299, + 0.03308993577957153, + 0.3686370551586151, + 0.019178446382284164, + 0.2822762429714203 + ], + "credit_map": [ + 0.005110003054141998, + 0.00929101649671793, + 0.005737510044127703, + 0.0052682566456496716, + 0.005642144475132227, + 0.004813733510673046, + 0.004669643472880125, + 0.006242663133889437, + 0.0012485877377912402, + 0.0037658147048205137, + 0.0022813782561570406, + 0.004410787485539913, + 0.0023668925277888775, + 0.0019223103299736977, + 0.005882151424884796, + 0.0015918591525405645, + 0.0016273665241897106, + 0.0029992160852998495, + 0.003980682231485844, + 0.0064313337206840515, + 0.0024221634957939386, + 0.003356791567057371, + 0.006083569955080748, + 0.002129232045263052, + 0.007028750143945217, + 0.00586642324924469, + 0.01397608034312725, + 0.03606298565864563, + 0.005374712869524956, + 0.003317568451166153, + 0.017864210531115532, + 0.003506107721477747, + 0.0020831460133194923, + 0.0023446157574653625, + 0.005787814036011696, + 0.016886694356799126, + 0.004672587849199772, + 0.0021661093924194574, + 0.02628360316157341, + 0.004185330588370562, + 0.001965717412531376, + 0.00204735086299479, + 0.002793654566630721, + 0.00927708949893713, + 0.00521328067407012, + 0.0021058658603578806, + 0.014600767754018307, + 0.0021327307913452387, + 0.0015119337476789951, + 0.001984346192330122, + 0.005141475237905979, + 0.007090842351317406, + 0.0039820014499127865, + 0.002216096268966794, + 0.007805598899722099, + 0.003357639769092202, + 0.006893085781484842, + 0.010288262739777565, + 0.03393169119954109, + 0.1594112664461136, + 0.009351379238069057, + 0.010829521343111992, + 0.12411880493164062, + 0.013813064433634281, + 0.0009474060498178005, + 0.0010794654954224825, + 0.005889969412237406, + 0.0022242176346480846, + 0.0017773150466382504, + 0.0009123761556111276, + 0.004898369777947664, + 0.0014493275666609406, + 0.011874535121023655, + 0.007980006746947765, + 0.07964850217103958, + 0.11892516165971756, + 0.006123012397438288, + 0.012067830190062523, + 0.03469615429639816, + 0.010961046442389488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.023300303146243095 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704533338546753 + ], + "opposite_logprob_tokens": [ + -2.1270453929901123 + ], + "logprob_reward": -0.12704533338546753, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704533338546753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04245021939277649, + 0.04421759024262428, + 0.02502666786313057, + 0.02634946070611477, + 0.029055221006274223, + 0.022234007716178894, + 0.04995967447757721, + 0.2383597046136856, + 0.0725603997707367, + 0.44978711009025574 + ], + "credit_map": [ + 0.005949115380644798, + 0.007717025466263294, + 0.004318972583860159, + 0.004733251873403788, + 0.00447479821741581, + 0.005613960791379213, + 0.0044478606432676315, + 0.005195233505219221, + 0.003467354690656066, + 0.0036125415936112404, + 0.003206180175766349, + 0.007417125161737204, + 0.004597329068928957, + 0.0028820026200264692, + 0.012659566476941109, + 0.006375488825142384, + 0.001993053127080202, + 0.0034035909920930862, + 0.0036947226617485285, + 0.0036438466049730778, + 0.0026095130015164614, + 0.002903114538639784, + 0.0038506402634084225, + 0.0029281836468726397, + 0.004066308960318565, + 0.002784627489745617, + 0.0031998534686863422, + 0.004154228139668703, + 0.002260421169921756, + 0.002925561275333166, + 0.0025788319762796164, + 0.004379625897854567, + 0.002464814344421029, + 0.0025379436556249857, + 0.003677557921037078, + 0.005162358749657869, + 0.0037216844502836466, + 0.0016300574643537402, + 0.006692611146718264, + 0.003168190363794565, + 0.0024506021291017532, + 0.002417051000520587, + 0.0024361773394048214, + 0.0038047621492296457, + 0.0027951044030487537, + 0.0024595041759312153, + 0.003911773208528757, + 0.001959033077582717, + 0.002940719248726964, + 0.002625590655952692, + 0.008214634843170643, + 0.009306865744292736, + 0.008118901401758194, + 0.0028670188039541245, + 0.011064285412430763, + 0.004821657668799162, + 0.004024730995297432, + 0.006200161762535572, + 0.013307911343872547, + 0.09784337878227234, + 0.007931642234325409, + 0.0073075369000434875, + 0.09133297950029373, + 0.01041134912520647, + 0.004529570695012808, + 0.00296615669503808, + 0.021937362849712372, + 0.009809518232941628, + 0.0063866074196994305, + 0.00685507245361805, + 0.013989209197461605, + 0.006086896173655987, + 0.018315007910132408, + 0.011638070456683636, + 0.1540672779083252, + 0.18486998975276947, + 0.007425339426845312, + 0.012227202765643597, + 0.047574810683727264, + 0.01366937905550003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.12704533338546753 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073797941207886 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073797941207886, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073797941207886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035341113805770874, + 0.011383083648979664, + 0.017268575727939606, + 0.023232290521264076, + 0.08232303708791733, + 0.01631215773522854, + 0.03134533762931824, + 0.4868183732032776, + 0.05430067330598831, + 0.24167534708976746 + ], + "credit_map": [ + 0.004494707100093365, + 0.00700882263481617, + 0.0044873980805277824, + 0.003966848831623793, + 0.003926754929125309, + 0.0037454920820891857, + 0.003793898969888687, + 0.003917190711945295, + 0.0005716372397728264, + 0.0012565234210342169, + 0.001052616280503571, + 0.002115936018526554, + 0.0015033447416499257, + 0.0007156904321163893, + 0.0032887207344174385, + 0.000878614722751081, + 0.0012079504085704684, + 0.0018925908952951431, + 0.002313984790816903, + 0.0034430869854986668, + 0.0013051173882558942, + 0.001838688855059445, + 0.003738939296454191, + 0.0015282187378033996, + 0.0019603886175900698, + 0.001837020507082343, + 0.003320267191156745, + 0.00791227724403143, + 0.001782685867510736, + 0.0013295739190652966, + 0.003932788502424955, + 0.0011572865769267082, + 0.0022084254305809736, + 0.0028565493412315845, + 0.0070620388723909855, + 0.02112123742699623, + 0.006345858797430992, + 0.003007427090778947, + 0.03350907936692238, + 0.006212424952536821, + 0.0006880349828861654, + 0.00109010876622051, + 0.0016043306095525622, + 0.00323352194391191, + 0.0020297865848988295, + 0.000981898047029972, + 0.005632374435663223, + 0.0010521038202568889, + 0.0014060231624171138, + 0.001572780660353601, + 0.003078395500779152, + 0.008086098358035088, + 0.0026833373121917248, + 0.001665827352553606, + 0.010347593575716019, + 0.0025052842684090137, + 0.00790299940854311, + 0.012996507808566093, + 0.043156981468200684, + 0.21027196943759918, + 0.012070207856595516, + 0.014126323163509369, + 0.16798822581768036, + 0.018305111676454544, + 0.0038643016014248133, + 0.002892378717660904, + 0.017074795439839363, + 0.00800050888210535, + 0.004477673210203648, + 0.004338426981121302, + 0.00866985134780407, + 0.00498274015262723, + 0.00918929185718298, + 0.005606018006801605, + 0.09760277718305588, + 0.08726763725280762, + 0.0039537339471280575, + 0.011276839300990105, + 0.020344268530607224, + 0.006434781011193991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.011073797941207886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up shot captured with a telephoto lens, the camera slowly tilts upward, tracking a massive rectangular stone monument as it rises majestically and inexorably from a forest floor blanketed in damp black soil and decaying leaves. The monument is constructed of weathered, dark gray granite, its surface covered in mottled emerald green and deep black moss clinging to its cracks and damp areas. As the monument rises, dry dust and fine stone chips are shaken from the ancient inscriptions, eroded by time and nearly unreadable. Like tiny waterfalls of sand, they flow down its rough surface, visible in the few faint, grayish-white rays of light filtering through the dense canopy. The telephoto lens compresses space, making the blurred, gnarled ancient trees in the background appear both distant and oppressive. The entire scene, shrouded in a near-stagnant silence, is imbued with an atmosphere of antiquity, mystery, and solemnity.", + "seed": 8413, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_13/index_6865_seed_8413_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10556844640523196, + "vlm_logprob": -0.10556844640523196, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.10556844640523196, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028610320761799812 + ], + "opposite_logprob_tokens": [ + -4.4036102294921875 + ], + "logprob_reward": -0.028610320761799812, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.028610320761799812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1586018055677414, + 0.10505008697509766, + 0.1281818449497223, + 0.04366407170891762, + 0.04718392342329025, + 0.39586326479911804, + 0.05461392551660538, + 0.01719278283417225, + 0.033509258180856705, + 0.01613904908299446 + ], + "credit_map": [ + 0.010290548205375671, + 0.022667957469820976, + 0.04717051610350609, + 0.029889235273003578, + 0.010001791641116142, + 0.010095259174704552, + 0.012520920485258102, + 0.015965590253472328, + 0.008754530921578407, + 0.009288495406508446, + 0.031095977872610092, + 0.014043277129530907, + 0.006244300398975611, + 0.004115880932658911, + 0.01062064990401268, + 0.020886974409222603, + 0.010417533107101917, + 0.008575073443353176, + 0.01866939663887024, + 0.04347698390483856, + 0.005864476319402456, + 0.004873637109994888, + 0.01084103249013424, + 0.025463711470365524, + 0.0033305659890174866, + 0.00686390046030283, + 0.007435462903231382, + 0.005815477110445499, + 0.004206163343042135, + 0.0025930795818567276, + 0.0053309490904212, + 0.008088473230600357, + 0.0059086354449391365, + 0.007431956008076668, + 0.007069086655974388, + 0.005825158674269915, + 0.003208204172551632, + 0.004746789578348398, + 0.0070540765300393105, + 0.005940014962106943, + 0.024956801906228065, + 0.10472548753023148, + 0.0574314258992672, + 0.043961066752672195, + 0.03141326084733009, + 0.0156986303627491, + 0.032317545264959335, + 0.08535905927419662, + 0.005539456382393837, + 0.007060325238853693, + 0.011455545201897621, + 0.009523861110210419, + 0.0036688395775854588, + 0.0029817703180015087, + 0.009089618921279907, + 0.005294507835060358, + 0.0018843822181224823, + 0.00214303913526237, + 0.0031603900715708733, + 0.0027736646588891745, + 0.0014917708467692137, + 0.0017940994584932923, + 0.0028105173259973526, + 0.001134918536990881, + 0.004917220678180456, + 0.0059307343326509, + 0.006314093247056007, + 0.004398027900606394, + 0.0025537684559822083, + 0.0021707862615585327, + 0.003692599479109049, + 0.0035320264287292957, + 0.002261860528960824, + 0.003305913880467415, + 0.002104830229654908, + 0.0019555063918232918, + 0.0017498378874734044, + 0.000909567519556731, + 0.0024874762166291475, + 0.0013640556717291474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.028610320761799812 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373221680521965 + ], + "opposite_logprob_tokens": [ + -3.4087321758270264 + ], + "logprob_reward": -0.03373221680521965, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373221680521965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2367641031742096, + 0.06710341572761536, + 0.22761619091033936, + 0.03296813368797302, + 0.024327868595719337, + 0.058335497975349426, + 0.051347434520721436, + 0.14817975461483002, + 0.13062375783920288, + 0.022733788937330246 + ], + "credit_map": [ + 0.027541616931557655, + 0.038593221455812454, + 0.04389505088329315, + 0.03811178356409073, + 0.019972827285528183, + 0.012415369972586632, + 0.029049457982182503, + 0.027184799313545227, + 0.008482905104756355, + 0.009395838715136051, + 0.010634231381118298, + 0.00810233037918806, + 0.006137010641396046, + 0.002855838742107153, + 0.010106869973242283, + 0.011388400569558144, + 0.019619300961494446, + 0.01615060865879059, + 0.03412313014268875, + 0.07431947439908981, + 0.012376728467643261, + 0.009232007898390293, + 0.019175782799720764, + 0.04261917993426323, + 0.002940294798463583, + 0.006131946109235287, + 0.006939061917364597, + 0.004331750795245171, + 0.0037982177454978228, + 0.001403670758008957, + 0.004017008934170008, + 0.0034061840269714594, + 0.002112151123583317, + 0.0032083301339298487, + 0.0037916151341050863, + 0.0041449349373579025, + 0.0020457443315535784, + 0.002031318610534072, + 0.003683624556288123, + 0.0033101532608270645, + 0.002826378447934985, + 0.010934899561107159, + 0.0053710066713392735, + 0.019742397591471672, + 0.0030183950439095497, + 0.002335228491574526, + 0.00642943661659956, + 0.007677762769162655, + 0.0034788004122674465, + 0.008742501959204674, + 0.008188804611563683, + 0.008379664272069931, + 0.0046380069106817245, + 0.0031058366876095533, + 0.01011049933731556, + 0.004703323356807232, + 0.009567896835505962, + 0.009054089896380901, + 0.053017374128103256, + 0.04857664927840233, + 0.006265407893806696, + 0.004652948118746281, + 0.00928580667823553, + 0.007759594824165106, + 0.01602332293987274, + 0.01780714839696884, + 0.021354809403419495, + 0.016653599217534065, + 0.010164031758904457, + 0.019234007224440575, + 0.014040173031389713, + 0.015346679836511612, + 0.0033758701756596565, + 0.0036097047850489616, + 0.002804218791425228, + 0.0032218366395682096, + 0.0027037402614951134, + 0.0015344033017754555, + 0.003555289236828685, + 0.0019287287723273039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.03373221680521965 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029828637838363647 + ], + "opposite_logprob_tokens": [ + -3.5298285484313965 + ], + "logprob_reward": -0.029828637838363647, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029828637838363647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20766432583332062, + 0.2639138102531433, + 0.07063038647174835, + 0.08262573182582855, + 0.06547752767801285, + 0.1511339545249939, + 0.037416644394397736, + 0.05915548652410507, + 0.03336301073431969, + 0.028619034215807915 + ], + "credit_map": [ + 0.011231103911995888, + 0.01647825352847576, + 0.07141019403934479, + 0.047841209918260574, + 0.012868511490523815, + 0.008869674056768417, + 0.013591435737907887, + 0.025373956188559532, + 0.020518921315670013, + 0.027758682146668434, + 0.07037968933582306, + 0.03534325212240219, + 0.022311875596642494, + 0.012442103587090969, + 0.03211873397231102, + 0.04304054379463196, + 0.006401165388524532, + 0.012861738912761211, + 0.008909481577575207, + 0.009627795778214931, + 0.007533856201916933, + 0.004266661591827869, + 0.009846472181379795, + 0.011183212511241436, + 0.005073441658169031, + 0.03149384260177612, + 0.015064241364598274, + 0.009637384675443172, + 0.00633349409326911, + 0.0025355154648423195, + 0.004861774854362011, + 0.00762603897601366, + 0.005827354732900858, + 0.008472636342048645, + 0.009960011579096317, + 0.014848086051642895, + 0.004173395689576864, + 0.00411265017464757, + 0.010117375291883945, + 0.007966023869812489, + 0.009854990057647228, + 0.05190108343958855, + 0.015542208217084408, + 0.025931892916560173, + 0.010661222971975803, + 0.005733175203204155, + 0.01051383838057518, + 0.020995546132326126, + 0.0036007112357765436, + 0.007035895250737667, + 0.004897056147456169, + 0.004468217957764864, + 0.0037047655787318945, + 0.001950894482433796, + 0.008668101392686367, + 0.0030910030473023653, + 0.0035067314747720957, + 0.0034449943341314793, + 0.023464513942599297, + 0.017786767333745956, + 0.00307164853438735, + 0.0014730332186445594, + 0.0036352069582790136, + 0.0027725882828235626, + 0.0035250489600002766, + 0.0059999157674610615, + 0.005106102209538221, + 0.005059268791228533, + 0.0028974944725632668, + 0.0037315140943974257, + 0.004301473032683134, + 0.002742191543802619, + 0.0036069396883249283, + 0.0049336994998157024, + 0.0045506833121180534, + 0.003236257703974843, + 0.0035707007627934217, + 0.0015422850847244263, + 0.00460563600063324, + 0.002572831464931369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.029828637838363647 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42892977595329285 + ], + "opposite_logprob_tokens": [ + -1.0539298057556152 + ], + "logprob_reward": -0.42892977595329285, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42892977595329285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15716448426246643, + 0.1681661456823349, + 0.1777065396308899, + 0.06399964541196823, + 0.1431257724761963, + 0.09882369637489319, + 0.036045536398887634, + 0.05687480792403221, + 0.0731482058763504, + 0.024945102632045746 + ], + "credit_map": [ + 0.009288354776799679, + 0.013922207988798618, + 0.05344304442405701, + 0.03377700224518776, + 0.008652972057461739, + 0.006826365366578102, + 0.012472506612539291, + 0.01878204755485058, + 0.012460945174098015, + 0.022099968045949936, + 0.04630720242857933, + 0.027472075074911118, + 0.014876597560942173, + 0.006717325188219547, + 0.015763889998197556, + 0.02246815897524357, + 0.015364660881459713, + 0.015023174695670605, + 0.03269511088728905, + 0.05818624421954155, + 0.01047943439334631, + 0.005741694010794163, + 0.01583682931959629, + 0.02437940426170826, + 0.005362050142139196, + 0.019085874781012535, + 0.012919045053422451, + 0.005832264199852943, + 0.00570821575820446, + 0.0037058466114103794, + 0.005041357129812241, + 0.006344999186694622, + 0.019289588555693626, + 0.02083795890212059, + 0.023250821977853775, + 0.024911953136324883, + 0.009887698106467724, + 0.010345026850700378, + 0.01753973588347435, + 0.01706300862133503, + 0.006419166456907988, + 0.026928329840302467, + 0.010298490524291992, + 0.012764990329742432, + 0.007446719333529472, + 0.003949868492782116, + 0.008854533545672894, + 0.022161604836583138, + 0.0034686727449297905, + 0.00473301624879241, + 0.004483480006456375, + 0.005446235649287701, + 0.002626348752528429, + 0.0026345669757574797, + 0.008601106703281403, + 0.004052110947668552, + 0.003262448823079467, + 0.002963503124192357, + 0.023340987041592598, + 0.017207924276590347, + 0.002722004195675254, + 0.0012831693748012185, + 0.003468140959739685, + 0.0026266328059136868, + 0.0083108926191926, + 0.008358451537787914, + 0.012093115597963333, + 0.009888283908367157, + 0.005487009882926941, + 0.011279978789389133, + 0.009114709682762623, + 0.008615769445896149, + 0.003115246770903468, + 0.00416285265237093, + 0.004547097720205784, + 0.003547919914126396, + 0.0022592637687921524, + 0.0019121355144307017, + 0.003458535997197032, + 0.0019420520402491093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.42892977595329285 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive rectangular stone monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward while tracking the monument as it rises from the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monument made of weathered dark gray granite covered in mottled emerald green and deep black moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dry dust and fine stone chips visible flowing down the monument's surface as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background show blurred, gnarled ancient trees with a telephoto lens compression effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741280667483807 + ], + "opposite_logprob_tokens": [ + -5.006741046905518 + ], + "logprob_reward": -0.006741280667483807, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741280667483807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09408188611268997, + 0.08964379131793976, + 0.20423485338687897, + 0.03893892094492912, + 0.0451931431889534, + 0.28401926159858704, + 0.09396326541900635, + 0.1002357080578804, + 0.028263766318559647, + 0.02142537757754326 + ], + "credit_map": [ + 0.01068152580410242, + 0.01142097357660532, + 0.017898542806506157, + 0.015311291441321373, + 0.013018539175391197, + 0.006724878214299679, + 0.008937835693359375, + 0.010088303126394749, + 0.010472476482391357, + 0.016400368884205818, + 0.014061314053833485, + 0.01352804433554411, + 0.008782892487943172, + 0.0043282900005578995, + 0.012428312562406063, + 0.009642093442380428, + 0.02307801879942417, + 0.02810760587453842, + 0.022876212373375893, + 0.043704621493816376, + 0.015565440058708191, + 0.010851120576262474, + 0.01926877722144127, + 0.04078305512666702, + 0.0043474650010466576, + 0.007254311349242926, + 0.00606116559356451, + 0.004352950025349855, + 0.004194241017103195, + 0.0028241325635463, + 0.004949264694005251, + 0.004955391865223646, + 0.0055307564325630665, + 0.004430458880960941, + 0.007452695164829493, + 0.0056719426065683365, + 0.0029875533655285835, + 0.003479747334495187, + 0.009993361309170723, + 0.005646624602377415, + 0.01457208301872015, + 0.10016809403896332, + 0.023925527930259705, + 0.060498930513858795, + 0.01932031288743019, + 0.01064090896397829, + 0.020497512072324753, + 0.03439589962363243, + 0.009685385040938854, + 0.00819745846092701, + 0.013324371539056301, + 0.016791168600320816, + 0.005420028232038021, + 0.007502357941120863, + 0.020918400958180428, + 0.012124096974730492, + 0.005873755551874638, + 0.0042756120674312115, + 0.04059280827641487, + 0.032097846269607544, + 0.00470610661432147, + 0.0027968978974968195, + 0.005015652161091566, + 0.004877034574747086, + 0.0026873864699155092, + 0.0028575032483786345, + 0.00528451194986701, + 0.0035998912062495947, + 0.002292117103934288, + 0.003953695762902498, + 0.0038996695075184107, + 0.003688992466777563, + 0.003392570884898305, + 0.0030441514682024717, + 0.0034404180478304625, + 0.0027613977435976267, + 0.002332694362848997, + 0.0016505795065313578, + 0.002793780295178294, + 0.002009786432608962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.006741280667483807 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_14/index_tensor([1769, 1769])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_14/index_tensor([1769, 1769])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a6e83bd2eaf64a7369e3dcab4b7707f39552f163 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_14/index_tensor([1769, 1769])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful overhead shot begins, followed by a steady, slow pull-back to reveal a magnificent, epic scene. Atop a steep, black mountaintop shrouded in mist, the legendary hero Sun Wukong stands proudly. He clad in a suit of intricately carved golden chainmail armor, each plate reflecting divinely inspired golden beams that pierce the dark clouds. The dark red lining at the armor's joints is faintly visible. He wears a phoenix-winged purple-gold crown, two slender pheasant feathers dancing fiercely in the mountaintop wind. Behind him, a vast scarlet robe rustles in the wind, almost blending into the swirling leaden sea of clouds. With his legs slightly bent, he remains rooted to the jagged ground like a rock. One hand grips his Ruyi Jingu Bang, embedded in a crevice in the rock, the other clenched into a fist. He jerked his head back, his features tense with utter fury. His fiery eyes burned with unyielding flames, his jaws gaping as if to devour the heavens and earth, and he let out a long, earth-shaking roar. This roar was so powerful that the air around him distorted and rippled visibly, and the gravel and dust beneath his feet trembled with the vibrations of an invisible force. The entire scene, with its striking contrast of gold, crimson, and dark gray, exuded a tragic atmosphere of aloofness, unyielding determination, and a mythical power.", + "seed": 2978, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_14/index_1769_seed_2978_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17621070481836795, + "vlm_logprob": -0.17621070481836795, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.17621070481836795, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6418384909629822 + ], + "opposite_logprob_tokens": [ + -0.7668384909629822 + ], + "logprob_reward": -0.6418384909629822, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6418384909629822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1652233749628067, + 0.07389617711305618, + 0.03408486023545265, + 0.09137717634439468, + 0.08377449214458466, + 0.07772326469421387, + 0.07776615023612976, + 0.19627206027507782, + 0.10336592048406601, + 0.09651659429073334 + ], + "credit_map": [ + 0.02430826984345913, + 0.017261717468500137, + 0.017689842730760574, + 0.029477206990122795, + 0.0247912909835577, + 0.014888008125126362, + 0.021998560056090355, + 0.014808481559157372, + 0.012967144139111042, + 0.00654850248247385, + 0.008700531907379627, + 0.009821583516895771, + 0.010030007921159267, + 0.006888434756547213, + 0.011101758107542992, + 0.00783821102231741, + 0.006046933587640524, + 0.0023064734414219856, + 0.006640428677201271, + 0.0032695410773158073, + 0.0031477920711040497, + 0.003520720638334751, + 0.005269298329949379, + 0.0038836717139929533, + 0.010624530725181103, + 0.003788059577345848, + 0.006979167927056551, + 0.0050938804633915424, + 0.006542026065289974, + 0.02684909477829933, + 0.006272413302212954, + 0.025227999314665794, + 0.012302895076572895, + 0.008018702268600464, + 0.008974403142929077, + 0.010068791918456554, + 0.013149814680218697, + 0.009452160447835922, + 0.011054557748138905, + 0.010753165930509567, + 0.010570359416306019, + 0.004302382934838533, + 0.0098872696980834, + 0.007464967668056488, + 0.0061220284551382065, + 0.016544876620173454, + 0.0068995170295238495, + 0.01593187265098095, + 0.01053721271455288, + 0.0067319078370928764, + 0.012406065128743649, + 0.009302468039095402, + 0.00843613501638174, + 0.009649927727878094, + 0.010366660542786121, + 0.010335770435631275, + 0.010387378744781017, + 0.009223767556250095, + 0.00910249911248684, + 0.0056440867483615875, + 0.012666312977671623, + 0.07583896070718765, + 0.007221049629151821, + 0.06618800014257431, + 0.006637287326157093, + 0.006014539860188961, + 0.009208902716636658, + 0.005333396140486002, + 0.007675019092857838, + 0.03168950974941254, + 0.005328994709998369, + 0.03147825971245766, + 0.014205092564225197, + 0.0118298614397645, + 0.011993071995675564, + 0.008317394182085991, + 0.01315088290721178, + 0.0124656418338418, + 0.010353141464293003, + 0.014201508834958076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.6418384909629822 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550657615065575 + ], + "opposite_logprob_tokens": [ + -4.3875508308410645 + ], + "logprob_reward": -0.012550657615065575, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550657615065575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20061078667640686, + 0.04291509836912155, + 0.07504764199256897, + 0.06289003044366837, + 0.021467771381139755, + 0.053723208606243134, + 0.2125888466835022, + 0.21932348608970642, + 0.08829658478498459, + 0.023136429488658905 + ], + "credit_map": [ + 0.020678212866187096, + 0.02029522880911827, + 0.01837230660021305, + 0.03789093345403671, + 0.03479268029332161, + 0.023083476349711418, + 0.02530721202492714, + 0.020190751180052757, + 0.005027998238801956, + 0.004257312975823879, + 0.005121891386806965, + 0.003946838900446892, + 0.00780430156737566, + 0.00365935149602592, + 0.008463908918201923, + 0.004633498378098011, + 0.012009814381599426, + 0.005062715616077185, + 0.012962086126208305, + 0.008479638025164604, + 0.009151156060397625, + 0.009370294399559498, + 0.008368689566850662, + 0.0096432538703084, + 0.007775590289384127, + 0.0031727042514830828, + 0.0051933941431343555, + 0.004180342890322208, + 0.0036122289020568132, + 0.018348529934883118, + 0.004743169993162155, + 0.015864074230194092, + 0.003082506125792861, + 0.001988323638215661, + 0.0016008689999580383, + 0.002295171143487096, + 0.003445893060415983, + 0.003152488963678479, + 0.002201016992330551, + 0.0037015045527368784, + 0.005392021965235472, + 0.002061853650957346, + 0.002590100048109889, + 0.0025570052675902843, + 0.003683417337015271, + 0.018944328650832176, + 0.0021623338107019663, + 0.016332155093550682, + 0.03394412621855736, + 0.013132519088685513, + 0.019261948764324188, + 0.018944235518574715, + 0.013541174121201038, + 0.04765673354268074, + 0.01801595278084278, + 0.04809219762682915, + 0.01273954939097166, + 0.007902367040514946, + 0.007827289402484894, + 0.006699614226818085, + 0.011166230775415897, + 0.08587268739938736, + 0.00733975088223815, + 0.07977601885795593, + 0.003988754935562611, + 0.004475863184779882, + 0.003595094196498394, + 0.0025232303887605667, + 0.005651196930557489, + 0.034821949899196625, + 0.0038497939240187407, + 0.02939070574939251, + 0.00392124941572547, + 0.001970631768926978, + 0.002758314600214362, + 0.0016378556611016393, + 0.0027428718749433756, + 0.0042626033537089825, + 0.002541837515309453, + 0.0033010696060955524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.012550657615065575 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0125464191660285 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.0125464191660285, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0125464191660285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19702623784542084, + 0.09560434520244598, + 0.04031522944569588, + 0.17466220259666443, + 0.0915447250008583, + 0.10992076247930527, + 0.07545845955610275, + 0.10542198270559311, + 0.07870189845561981, + 0.03134411573410034 + ], + "credit_map": [ + 0.024118252098560333, + 0.023342736065387726, + 0.017512427642941475, + 0.033497415482997894, + 0.029733162373304367, + 0.01840774528682232, + 0.03229329362511635, + 0.01812121272087097, + 0.01011103019118309, + 0.009246119298040867, + 0.010202350094914436, + 0.01131333690136671, + 0.018215933814644814, + 0.008129743859171867, + 0.017714733257889748, + 0.010671105235815048, + 0.005019401665776968, + 0.005400012247264385, + 0.003618452465161681, + 0.005554735194891691, + 0.006938563194125891, + 0.003424534108489752, + 0.006104843225330114, + 0.004254691302776337, + 0.027923090383410454, + 0.0076620387844741344, + 0.013662394136190414, + 0.008557363413274288, + 0.011485866270959377, + 0.04813729226589203, + 0.014136281795799732, + 0.04309787228703499, + 0.011430210433900356, + 0.00910345558077097, + 0.013069060631096363, + 0.011404313147068024, + 0.013640249148011208, + 0.008656158111989498, + 0.012156217359006405, + 0.012085061520338058, + 0.014161021448671818, + 0.0065360828302800655, + 0.008371763862669468, + 0.006373043637722731, + 0.009291909635066986, + 0.03178265690803528, + 0.008185043931007385, + 0.02521924115717411, + 0.009297918528318405, + 0.006019330117851496, + 0.013315999880433083, + 0.006787996739149094, + 0.007538514211773872, + 0.012327596545219421, + 0.005527027416974306, + 0.014644073322415352, + 0.006231721490621567, + 0.0062006572261452675, + 0.005914679728448391, + 0.004408029839396477, + 0.007049370091408491, + 0.03754245117306709, + 0.004723490681499243, + 0.03335157781839371, + 0.0066332509741187096, + 0.00433602649718523, + 0.004598947241902351, + 0.0045112124644219875, + 0.004586352035403252, + 0.02454247698187828, + 0.0048950547352433205, + 0.02459857426583767, + 0.005123551934957504, + 0.003080076305195689, + 0.0030143978074193, + 0.0024035221431404352, + 0.004603056702762842, + 0.005207614973187447, + 0.003780520521104336, + 0.004131377674639225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0125464191660285 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015843689441681 + ], + "opposite_logprob_tokens": [ + -1.7015843391418457 + ], + "logprob_reward": -0.2015843689441681, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015843689441681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11347726732492447, + 0.02967929281294346, + 0.025542644783854485, + 0.03196718916296959, + 0.025336813181638718, + 0.01820817030966282, + 0.08522132784128189, + 0.45576462149620056, + 0.1643804907798767, + 0.05042219161987305 + ], + "credit_map": [ + 0.012121563777327538, + 0.012465065345168114, + 0.010213978588581085, + 0.021188106387853622, + 0.01828654669225216, + 0.011304233223199844, + 0.017671920359134674, + 0.010225851088762283, + 0.004806446842849255, + 0.0026488613802939653, + 0.0029416559264063835, + 0.003427012823522091, + 0.004569666925817728, + 0.0029471791349351406, + 0.004948292393237352, + 0.0033901759888976812, + 0.0040030996315181255, + 0.0021104554180055857, + 0.00364879728294909, + 0.0031903681810945272, + 0.003540025558322668, + 0.0030866090673953295, + 0.003014111425727606, + 0.002949177287518978, + 0.0058440351858735085, + 0.0015917842974886298, + 0.0022065979428589344, + 0.0016231194604188204, + 0.002410061191767454, + 0.008488603867590427, + 0.002530904021114111, + 0.007272082846611738, + 0.003245777217671275, + 0.0015861414140090346, + 0.00333676440641284, + 0.002979114418849349, + 0.003301149234175682, + 0.003551711793988943, + 0.003697965294122696, + 0.003638189285993576, + 0.0032681641168892384, + 0.0015556399011984468, + 0.0026864141691476107, + 0.0020084120333194733, + 0.0024146598298102617, + 0.001901636365801096, + 0.0018342044204473495, + 0.0025390414521098137, + 0.015599850565195084, + 0.006398738827556372, + 0.0065412698313593864, + 0.00479520857334137, + 0.005949806421995163, + 0.022966036573052406, + 0.006530487909913063, + 0.016439927741885185, + 0.028752844780683517, + 0.01647636480629444, + 0.017773868516087532, + 0.015492657199501991, + 0.02239437773823738, + 0.17259643971920013, + 0.016935188323259354, + 0.16534291207790375, + 0.00690415920689702, + 0.004547512158751488, + 0.005612658802419901, + 0.0040519749745726585, + 0.006437645759433508, + 0.06954072415828705, + 0.0036601268220692873, + 0.06362568587064743, + 0.010636456310749054, + 0.0035760693717747927, + 0.006223110947757959, + 0.005794811528176069, + 0.005446429830044508, + 0.006636931095272303, + 0.006819782312959433, + 0.0052885995246469975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.2015843689441681 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533587403595448 + ], + "opposite_logprob_tokens": [ + -4.387533664703369 + ], + "logprob_reward": -0.012533587403595448, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533587403595448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12286096811294556, + 0.03889520466327667, + 0.040816377848386765, + 0.028108498081564903, + 0.03199358657002449, + 0.03491862490773201, + 0.0921272337436676, + 0.44594573974609375, + 0.14946117997169495, + 0.014872610569000244 + ], + "credit_map": [ + 0.01854049041867256, + 0.011798309162259102, + 0.013685056939721107, + 0.0140015734359622, + 0.018957925960421562, + 0.014925343915820122, + 0.01992153376340866, + 0.011030730791389942, + 0.004674722906202078, + 0.003760622814297676, + 0.0051095993258059025, + 0.004359476268291473, + 0.005091383121907711, + 0.0036177444271743298, + 0.007375891786068678, + 0.004905765876173973, + 0.005903052631765604, + 0.0030838882084935904, + 0.006548709236085415, + 0.004465759731829166, + 0.005544337444007397, + 0.00470383558422327, + 0.0056994142942130566, + 0.00486737908795476, + 0.0033610376995056868, + 0.0017483248375356197, + 0.003260416677221656, + 0.0018431617645546794, + 0.0026625562459230423, + 0.006739847362041473, + 0.002626701956614852, + 0.005866450257599354, + 0.0049790143966674805, + 0.002427454339340329, + 0.003272299887612462, + 0.0035372700076550245, + 0.00443621352314949, + 0.0038686057087033987, + 0.005025617778301239, + 0.004447112791240215, + 0.0052190818823874, + 0.0020196898840367794, + 0.003021703567355871, + 0.0023586954921483994, + 0.003333239583298564, + 0.009281854145228863, + 0.0024484542664140463, + 0.007235903292894363, + 0.015481287613511086, + 0.006498612929135561, + 0.010647274553775787, + 0.005855608265846968, + 0.006592554040253162, + 0.020534655079245567, + 0.008320954628288746, + 0.018196282908320427, + 0.026016047224402428, + 0.016900788992643356, + 0.01657460257411003, + 0.01461099274456501, + 0.0225664135068655, + 0.17182601988315582, + 0.01569896563887596, + 0.1617518812417984, + 0.006401603575795889, + 0.004163953009992838, + 0.00488286092877388, + 0.0037740375846624374, + 0.005980569403618574, + 0.06334643810987473, + 0.003945017233490944, + 0.05696669965982437, + 0.002528377575799823, + 0.0012089293450117111, + 0.0016947381664067507, + 0.001412961632013321, + 0.0023525471333414316, + 0.002019258448854089, + 0.0014068286400288343, + 0.0022489691618829966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.012533587403595448 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful overhead shot begins, followed by a steady, slow pull-back to reveal a magnificent, epic scene. Atop a steep, black mountaintop shrouded in mist, the legendary hero Sun Wukong stands proudly. He clad in a suit of intricately carved golden chainmail armor, each plate reflecting divinely inspired golden beams that pierce the dark clouds. The dark red lining at the armor's joints is faintly visible. He wears a phoenix-winged purple-gold crown, two slender pheasant feathers dancing fiercely in the mountaintop wind. Behind him, a vast scarlet robe rustles in the wind, almost blending into the swirling leaden sea of clouds. With his legs slightly bent, he remains rooted to the jagged ground like a rock. One hand grips his Ruyi Jingu Bang, embedded in a crevice in the rock, the other clenched into a fist. He jerked his head back, his features tense with utter fury. His fiery eyes burned with unyielding flames, his jaws gaping as if to devour the heavens and earth, and he let out a long, earth-shaking roar. This roar was so powerful that the air around him distorted and rippled visibly, and the gravel and dust beneath his feet trembled with the vibrations of an invisible force. The entire scene, with its striking contrast of gold, crimson, and dark gray, exuded a tragic atmosphere of aloofness, unyielding determination, and a mythical power.", + "seed": 2979, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_14/index_1769_seed_2979_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1540666501969099, + "vlm_logprob": -0.1540666501969099, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.1540666501969099, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6422194242477417 + ], + "opposite_logprob_tokens": [ + -0.7672194242477417 + ], + "logprob_reward": -0.6422194242477417, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6422194242477417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0409100167453289, + 0.06014546379446983, + 0.02881801687180996, + 0.018614234402775764, + 0.025261133909225464, + 0.027730658650398254, + 0.07624851912260056, + 0.014942246489226818, + 0.6842975616455078, + 0.023032095283269882 + ], + "credit_map": [ + 0.0032127946615219116, + 0.002591072814539075, + 0.002902301959693432, + 0.0024217080790549517, + 0.005587478168308735, + 0.01342984102666378, + 0.005224096588790417, + 0.0055407267063856125, + 0.011137186549603939, + 0.0038529466837644577, + 0.011162776499986649, + 0.005749952979385853, + 0.009115827269852161, + 0.005987622309476137, + 0.005208885297179222, + 0.007930269464850426, + 0.003848264692351222, + 0.0037705644499510527, + 0.003633189247921109, + 0.0032628378830850124, + 0.003578983945772052, + 0.0035894704051315784, + 0.002942565595731139, + 0.004192139953374863, + 0.0024538955185562372, + 0.0013431970728561282, + 0.0023636838886886835, + 0.0012496367562562227, + 0.0025547868572175503, + 0.0025845044292509556, + 0.002220741007477045, + 0.003843786893412471, + 0.0020070187747478485, + 0.0015573236159980297, + 0.0029298393055796623, + 0.0017575517995283008, + 0.004833612125366926, + 0.005545519292354584, + 0.0033510716166347265, + 0.003279198193922639, + 0.0043596369214355946, + 0.002200022106990218, + 0.006842862349003553, + 0.0020840237848460674, + 0.0030421633273363113, + 0.0028702833224087954, + 0.003401938360184431, + 0.002929728478193283, + 0.005202147178351879, + 0.0030831170734018087, + 0.003685832256451249, + 0.003079640679061413, + 0.023920314386487007, + 0.02720668353140354, + 0.002843830967321992, + 0.007226950488984585, + 0.0017930485773831606, + 0.0019012304255738854, + 0.0030756841879338026, + 0.0017915982753038406, + 0.0016691845376044512, + 0.0017727540107443929, + 0.0014380201464518905, + 0.001500727259553969, + 0.012695971876382828, + 0.010520576499402523, + 0.009350837208330631, + 0.00686543807387352, + 0.25802862644195557, + 0.2967720925807953, + 0.021525854244828224, + 0.06853823363780975, + 0.0017811933066695929, + 0.0023973954375833273, + 0.0024240571074187756, + 0.002163891913369298, + 0.0044865235686302185, + 0.004105822183191776, + 0.0033082745503634214, + 0.0023649362847208977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.6422194242477417 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820170320570469 + ], + "opposite_logprob_tokens": [ + -4.01820182800293 + ], + "logprob_reward": -0.01820170320570469, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820170320570469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055902838706970215, + 0.03128078952431679, + 0.015697140246629715, + 0.01631670817732811, + 0.02976859360933304, + 0.021330740302801132, + 0.021173393353819847, + 0.006642080377787352, + 0.7942406535148621, + 0.0076470403000712395 + ], + "credit_map": [ + 0.004752202425152063, + 0.002552273916080594, + 0.0039365957491099834, + 0.003539430210366845, + 0.008301567286252975, + 0.01413259468972683, + 0.009932504035532475, + 0.008755671791732311, + 0.004321707878261805, + 0.0017567992908880115, + 0.0045083859004080296, + 0.0028596173506230116, + 0.0035426702816039324, + 0.005220327526330948, + 0.004079021979123354, + 0.004992261063307524, + 0.002339709550142288, + 0.0011696165893226862, + 0.0032504021655768156, + 0.001647974131628871, + 0.0017906915163621306, + 0.0020645386539399624, + 0.0016532816225662827, + 0.0017809263663366437, + 0.0016578725771978498, + 0.0009841283317655325, + 0.0018714547622948885, + 0.0008781465003266931, + 0.002339711645618081, + 0.00277334894053638, + 0.002615115372464061, + 0.0031969300471246243, + 0.0012391979107633233, + 0.0014283423079177737, + 0.0016187119763344526, + 0.001402654917910695, + 0.00888274610042572, + 0.009495974518358707, + 0.0022933895234018564, + 0.0034075782168656588, + 0.0028380630537867546, + 0.0010213851928710938, + 0.003976849373430014, + 0.0011983991134911776, + 0.004548330325633287, + 0.002987332409247756, + 0.0024748395662754774, + 0.0022855428978800774, + 0.0011538867838680744, + 0.0007256579701788723, + 0.0008855292107909918, + 0.0006876299157738686, + 0.006572456564754248, + 0.007969491183757782, + 0.0009979853639379144, + 0.002180757699534297, + 0.0004788823425769806, + 0.00035123969428241253, + 0.0007334189722314477, + 0.0003172034630551934, + 0.0010874840663746, + 0.0013301691506057978, + 0.00130040617659688, + 0.0010432767448946834, + 0.013236502185463905, + 0.011568937450647354, + 0.009656538255512714, + 0.007614833768457174, + 0.3017726540565491, + 0.3442809283733368, + 0.02627713792026043, + 0.07983314990997314, + 0.0006559434696100652, + 0.0012951527023687959, + 0.000953957496676594, + 0.0008089503389783204, + 0.0007679933914914727, + 0.0010730697540566325, + 0.0010993911419063807, + 0.0009925822960212827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.01820170320570469 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195733428001404 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.018195733428001404, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195733428001404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09284026175737381, + 0.13487298786640167, + 0.07590942829847336, + 0.06560881435871124, + 0.0546792596578598, + 0.10452647507190704, + 0.026783641427755356, + 0.006940995343029499, + 0.39524045586586, + 0.04259764030575752 + ], + "credit_map": [ + 0.011787686496973038, + 0.004862989764660597, + 0.00966810155659914, + 0.007658119313418865, + 0.011852550320327282, + 0.026697678491473198, + 0.008113963529467583, + 0.01219917181879282, + 0.016442248597741127, + 0.0056943767704069614, + 0.01653098315000534, + 0.006981351878494024, + 0.014084942638874054, + 0.02725258655846119, + 0.015379027463495731, + 0.03250746801495552, + 0.008593026548624039, + 0.009897944517433643, + 0.008120455779135227, + 0.010638011619448662, + 0.009422200731933117, + 0.01390695758163929, + 0.006785357370972633, + 0.008545471355319023, + 0.012694895267486572, + 0.003937494475394487, + 0.011491999961435795, + 0.004984097555279732, + 0.008703498169779778, + 0.008623428642749786, + 0.006923045963048935, + 0.008250357583165169, + 0.004927229136228561, + 0.0034591841977089643, + 0.006722020450979471, + 0.002907239133492112, + 0.01364116184413433, + 0.014540993608534336, + 0.0035961540415883064, + 0.004885274451225996, + 0.021234557032585144, + 0.006358144339174032, + 0.022835077717900276, + 0.006841678638011217, + 0.016331786289811134, + 0.009756837971508503, + 0.008667358197271824, + 0.012501033022999763, + 0.0013242611894384027, + 0.0008930934709496796, + 0.0016254690708592534, + 0.001423140405677259, + 0.007712813559919596, + 0.009055917151272297, + 0.001974918879568577, + 0.002774027409031987, + 0.0007687875768169761, + 0.0005067879101261497, + 0.000924350053537637, + 0.0004342150059528649, + 0.0010071248980239034, + 0.00106732954736799, + 0.0013118108035996556, + 0.0009205895476043224, + 0.007032656576484442, + 0.005923752207309008, + 0.005299062933772802, + 0.00402144156396389, + 0.1490568220615387, + 0.17100270092487335, + 0.013731412589550018, + 0.03917261213064194, + 0.0021508007775992155, + 0.0026647918857634068, + 0.002228040946647525, + 0.0019580028019845486, + 0.0075734504498541355, + 0.014110208489000797, + 0.004631776362657547, + 0.007280571851879358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.018195733428001404 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866794869303703 + ], + "opposite_logprob_tokens": [ + -3.0486679077148438 + ], + "logprob_reward": -0.04866794869303703, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866794869303703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09297332167625427, + 0.181264728307724, + 0.046063896268606186, + 0.07769975811243057, + 0.04868920147418976, + 0.09657695889472961, + 0.07045000791549683, + 0.043710581958293915, + 0.22394119203090668, + 0.11863036453723907 + ], + "credit_map": [ + 0.016180863603949547, + 0.00692400336265564, + 0.011890808120369911, + 0.00788608007133007, + 0.00853062141686678, + 0.02079916186630726, + 0.009691714309155941, + 0.0110700698569417, + 0.023338954895734787, + 0.007837164215743542, + 0.02198566496372223, + 0.011881744489073753, + 0.017836209386587143, + 0.03702261298894882, + 0.01952989026904106, + 0.04183249548077583, + 0.0071661388501524925, + 0.0038489012513309717, + 0.008169426582753658, + 0.0035922692622989416, + 0.0052109938114881516, + 0.007369757164269686, + 0.005116643384099007, + 0.005589762702584267, + 0.0159382876008749, + 0.007823740132153034, + 0.010490665212273598, + 0.005842303391546011, + 0.010171000845730305, + 0.009960755705833435, + 0.008020096458494663, + 0.00945290643721819, + 0.005512167699635029, + 0.005344661418348551, + 0.006425068713724613, + 0.0043528126552701, + 0.007516084238886833, + 0.00719868578016758, + 0.006816048175096512, + 0.005523669999092817, + 0.013541562482714653, + 0.008030959405004978, + 0.021975001320242882, + 0.00796416774392128, + 0.013210699893534184, + 0.010651562362909317, + 0.010195301845669746, + 0.011007697321474552, + 0.002161318901926279, + 0.0031497834715992212, + 0.0025877507869154215, + 0.0038169887848198414, + 0.02326248399913311, + 0.026203634217381477, + 0.002911108545958996, + 0.006356936413794756, + 0.0062745907343924046, + 0.00394340930506587, + 0.01011417806148529, + 0.0039862170815467834, + 0.0049814991652965546, + 0.005150848533958197, + 0.004207211080938578, + 0.005052629392594099, + 0.005405282601714134, + 0.005429826211184263, + 0.0065086400136351585, + 0.006504704710096121, + 0.07121579349040985, + 0.08296461403369904, + 0.021451396867632866, + 0.0244609247893095, + 0.01031647901982069, + 0.008753587491810322, + 0.009078683331608772, + 0.007214452605694532, + 0.01994159258902073, + 0.0332045741379261, + 0.01007653959095478, + 0.02004445716738701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.04866794869303703 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430484414100647 + ], + "opposite_logprob_tokens": [ + -3.16804838180542 + ], + "logprob_reward": -0.0430484414100647, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430484414100647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10627903789281845, + 0.07592454552650452, + 0.06099928542971611, + 0.06296558678150177, + 0.046016424894332886, + 0.08108093589544296, + 0.015446461737155914, + 0.02107406035065651, + 0.46904364228248596, + 0.06117001175880432 + ], + "credit_map": [ + 0.01578926108777523, + 0.008608220145106316, + 0.01427501067519188, + 0.009139003232121468, + 0.009703919291496277, + 0.02264142967760563, + 0.014108432456851006, + 0.012013762257993221, + 0.015293180011212826, + 0.004774586763232946, + 0.015431956388056278, + 0.007805051282048225, + 0.008202698081731796, + 0.009013282135128975, + 0.006412126123905182, + 0.008991665206849575, + 0.009645500220358372, + 0.004991014488041401, + 0.010344293899834156, + 0.006214657332748175, + 0.007781764958053827, + 0.006842759437859058, + 0.005985589232295752, + 0.00919371098279953, + 0.012310883030295372, + 0.003444765228778124, + 0.011207780800759792, + 0.003841002006083727, + 0.008474458009004593, + 0.0075595988892018795, + 0.007680707145482302, + 0.00844639167189598, + 0.004250357858836651, + 0.0033195731230080128, + 0.005468281451612711, + 0.003082283539697528, + 0.0065206801518797874, + 0.008142005652189255, + 0.0076333992183208466, + 0.007599840871989727, + 0.01460875291377306, + 0.005391573999077082, + 0.021882478147745132, + 0.00622852798551321, + 0.008745511062443256, + 0.00816423911601305, + 0.006986350752413273, + 0.009073510766029358, + 0.0014540287666022778, + 0.0017359510529786348, + 0.0016005149809643626, + 0.0018477211706340313, + 0.0021659694612026215, + 0.0018132022814825177, + 0.0026672424282878637, + 0.002161832293495536, + 0.003486370900645852, + 0.0014780362835153937, + 0.004798542242497206, + 0.0019421506440266967, + 0.0023183568846434355, + 0.0021170293912291527, + 0.0023203459568321705, + 0.0026132280472666025, + 0.008872795850038528, + 0.007197179831564426, + 0.006416243501007557, + 0.004887086804956198, + 0.17523640394210815, + 0.2034483402967453, + 0.01526715513318777, + 0.04771846532821655, + 0.006064896937459707, + 0.005941904149949551, + 0.007724383845925331, + 0.005034216679632664, + 0.007450832054018974, + 0.012749720364809036, + 0.0063840108923614025, + 0.00982004776597023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.0430484414100647 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful overhead shot begins, followed by a steady, slow pull-back to reveal a magnificent, epic scene. Atop a steep, black mountaintop shrouded in mist, the legendary hero Sun Wukong stands proudly. He clad in a suit of intricately carved golden chainmail armor, each plate reflecting divinely inspired golden beams that pierce the dark clouds. The dark red lining at the armor's joints is faintly visible. He wears a phoenix-winged purple-gold crown, two slender pheasant feathers dancing fiercely in the mountaintop wind. Behind him, a vast scarlet robe rustles in the wind, almost blending into the swirling leaden sea of clouds. With his legs slightly bent, he remains rooted to the jagged ground like a rock. One hand grips his Ruyi Jingu Bang, embedded in a crevice in the rock, the other clenched into a fist. He jerked his head back, his features tense with utter fury. His fiery eyes burned with unyielding flames, his jaws gaping as if to devour the heavens and earth, and he let out a long, earth-shaking roar. This roar was so powerful that the air around him distorted and rippled visibly, and the gravel and dust beneath his feet trembled with the vibrations of an invisible force. The entire scene, with its striking contrast of gold, crimson, and dark gray, exuded a tragic atmosphere of aloofness, unyielding determination, and a mythical power.", + "seed": 2980, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_14/index_1769_seed_2980_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13657539244741201, + "vlm_logprob": -0.13657539244741201, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.13657539244741201, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5315073728561401 + ], + "opposite_logprob_tokens": [ + -0.9065073728561401 + ], + "logprob_reward": -0.5315073728561401, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5315073728561401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1042337566614151, + 0.056854359805583954, + 0.0617198571562767, + 0.051403164863586426, + 0.025864580646157265, + 0.1353452205657959, + 0.18029062449932098, + 0.11025914549827576, + 0.2119179666042328, + 0.062111321836709976 + ], + "credit_map": [ + 0.01066327653825283, + 0.012140915729105473, + 0.013983769342303276, + 0.010711395181715488, + 0.009902939200401306, + 0.01416641753166914, + 0.013227768242359161, + 0.01943727768957615, + 0.011098567396402359, + 0.006253570783883333, + 0.004708226770162582, + 0.0050779664888978004, + 0.005886998027563095, + 0.008627018891274929, + 0.008319737389683723, + 0.006882273592054844, + 0.006395617034286261, + 0.0065051475539803505, + 0.008681604638695717, + 0.009014692157506943, + 0.005461875814944506, + 0.00528748519718647, + 0.01181458868086338, + 0.008558842353522778, + 0.006265082396566868, + 0.005878920201212168, + 0.005511289928108454, + 0.006922513246536255, + 0.004279029555618763, + 0.005263498052954674, + 0.009777556173503399, + 0.0075052776373922825, + 0.0025666854344308376, + 0.005055298563092947, + 0.003298191586509347, + 0.002991016488522291, + 0.002953744027763605, + 0.0025135278701782227, + 0.003434104612097144, + 0.0030520125292241573, + 0.02616194635629654, + 0.010156959295272827, + 0.014143317937850952, + 0.009139899164438248, + 0.010075809434056282, + 0.02967039868235588, + 0.01284545473754406, + 0.023151438683271408, + 0.03899849206209183, + 0.015854662284255028, + 0.023497693240642548, + 0.014350281096994877, + 0.01660480722784996, + 0.02991451509296894, + 0.019464805722236633, + 0.021605370566248894, + 0.016458500176668167, + 0.011144212447106838, + 0.015374564565718174, + 0.010261026211082935, + 0.019928669556975365, + 0.013660102151334286, + 0.009094765409827232, + 0.014337309636175632, + 0.028537679463624954, + 0.021955668926239014, + 0.02990015596151352, + 0.019489139318466187, + 0.047518759965896606, + 0.030277196317911148, + 0.015421745367348194, + 0.01881762221455574, + 0.010924375616014004, + 0.00883367657661438, + 0.004513793624937534, + 0.004712141118943691, + 0.012652505189180374, + 0.008470993489027023, + 0.003427172312512994, + 0.008576665073633194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.5315073728561401 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371031582355499 + ], + "opposite_logprob_tokens": [ + -3.408710241317749 + ], + "logprob_reward": -0.03371031582355499, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371031582355499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2553282380104065, + 0.1040891706943512, + 0.04830898717045784, + 0.031806379556655884, + 0.02915383316576481, + 0.08386065810918808, + 0.09419337660074234, + 0.17986515164375305, + 0.130814790725708, + 0.04257942736148834 + ], + "credit_map": [ + 0.07569052278995514, + 0.02395332045853138, + 0.02494615875184536, + 0.017200546339154243, + 0.01575966365635395, + 0.03909972310066223, + 0.02443389780819416, + 0.03424441069364548, + 0.028474316000938416, + 0.012286774814128876, + 0.01056007482111454, + 0.007634572684764862, + 0.0072349668480455875, + 0.014480817131698132, + 0.008585885167121887, + 0.014831766486167908, + 0.013421901501715183, + 0.0037708906456828117, + 0.004658697172999382, + 0.003847689600661397, + 0.0030732478480786085, + 0.007880114950239658, + 0.0049566286616027355, + 0.006699813064187765, + 0.0053442842327058315, + 0.002561978530138731, + 0.0039160773158073425, + 0.0025995103642344475, + 0.0019439206225797534, + 0.004121145233511925, + 0.006390662398189306, + 0.004928798880428076, + 0.004082344006747007, + 0.003124144859611988, + 0.003137357300147414, + 0.003048036713153124, + 0.0031136316247284412, + 0.004400426987558603, + 0.0038736972492188215, + 0.004374195821583271, + 0.018453432247042656, + 0.007864916697144508, + 0.010826709680259228, + 0.00525082740932703, + 0.005356518551707268, + 0.01182998064905405, + 0.010446556843817234, + 0.013831716030836105, + 0.022516243159770966, + 0.008755345828831196, + 0.014765210449695587, + 0.006940534338355064, + 0.007868642918765545, + 0.013355987146496773, + 0.006901239976286888, + 0.013090173713862896, + 0.029253441840410233, + 0.025654945522546768, + 0.020998893305659294, + 0.015315537340939045, + 0.018404044210910797, + 0.019909018650650978, + 0.01968265324831009, + 0.03064662218093872, + 0.01446809247136116, + 0.01425935234874487, + 0.010793672874569893, + 0.009243781678378582, + 0.039544738829135895, + 0.022015085443854332, + 0.007691989652812481, + 0.012798086740076542, + 0.005808464251458645, + 0.006297196727246046, + 0.0034273765049874783, + 0.0038542584516108036, + 0.007913543842732906, + 0.005432414822280407, + 0.0032475099433213472, + 0.006598661653697491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.03371031582355499 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254300493746996 + ], + "opposite_logprob_tokens": [ + -4.387543201446533 + ], + "logprob_reward": -0.01254300493746996, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254300493746996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1948976367712021, + 0.09989304840564728, + 0.0823768675327301, + 0.038812581449747086, + 0.0815034955739975, + 0.17091666162014008, + 0.07041854411363602, + 0.12601466476917267, + 0.04997316002845764, + 0.08519336581230164 + ], + "credit_map": [ + 0.031879130750894547, + 0.026518216356635094, + 0.019576234742999077, + 0.020444447174668312, + 0.015643801540136337, + 0.02922937460243702, + 0.027145275846123695, + 0.024461153894662857, + 0.01087745651602745, + 0.011963913217186928, + 0.01067516952753067, + 0.012946105562150478, + 0.0069890134036540985, + 0.013485164381563663, + 0.017133241519331932, + 0.01582297310233116, + 0.008116768673062325, + 0.009807353839278221, + 0.008358295075595379, + 0.013667011633515358, + 0.006753889843821526, + 0.010270556434988976, + 0.013678502291440964, + 0.011724486015737057, + 0.0037430671509355307, + 0.005211777985095978, + 0.004918006248772144, + 0.005384658928960562, + 0.0038203089497983456, + 0.003460975829511881, + 0.006877663545310497, + 0.005396117921918631, + 0.00744996452704072, + 0.014801974408328533, + 0.01020092610269785, + 0.0108862379565835, + 0.008171983063220978, + 0.008977905847132206, + 0.010428767651319504, + 0.010585727170109749, + 0.03884569928050041, + 0.013204237446188927, + 0.022524118423461914, + 0.009400918148458004, + 0.012406528927385807, + 0.030211953446269035, + 0.021557312458753586, + 0.0227658674120903, + 0.014386538416147232, + 0.007743713911622763, + 0.011363420635461807, + 0.0062102060765028, + 0.005927845370024443, + 0.007972132414579391, + 0.007568874396383762, + 0.009245810098946095, + 0.014980942942202091, + 0.018003739416599274, + 0.016921643167734146, + 0.01362161710858345, + 0.0211957935243845, + 0.014760554768145084, + 0.012393712997436523, + 0.014136656187474728, + 0.007224643137305975, + 0.005772638134658337, + 0.005934135057032108, + 0.004783730022609234, + 0.01041016448289156, + 0.006566444877535105, + 0.0037103400100022554, + 0.00557105615735054, + 0.008855395019054413, + 0.015918543562293053, + 0.009111687541007996, + 0.011285115033388138, + 0.015143589116632938, + 0.011092028580605984, + 0.006290046498179436, + 0.007496953941881657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.01254300493746996 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890565812587738 + ], + "opposite_logprob_tokens": [ + -2.4640564918518066 + ], + "logprob_reward": -0.0890565812587738, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890565812587738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15533392131328583, + 0.0577198788523674, + 0.0669630914926529, + 0.1012691855430603, + 0.059263937175273895, + 0.1533433198928833, + 0.08818036317825317, + 0.05246587097644806, + 0.15161316096782684, + 0.11384721845388412 + ], + "credit_map": [ + 0.025605613365769386, + 0.011181067675352097, + 0.0183599554002285, + 0.019634829834103584, + 0.012209204025566578, + 0.02261032909154892, + 0.0246901698410511, + 0.02104276977479458, + 0.010338295251131058, + 0.006463813129812479, + 0.0061400411650538445, + 0.006833953317254782, + 0.005416332744061947, + 0.007690059021115303, + 0.0077233510091900826, + 0.007114039268344641, + 0.01019243709743023, + 0.006500238552689552, + 0.008915452286601067, + 0.007616564631462097, + 0.005408712197095156, + 0.009924676269292831, + 0.009612278081476688, + 0.008792738430202007, + 0.012316018342971802, + 0.013437551446259022, + 0.009839375503361225, + 0.015543939545750618, + 0.008608334697782993, + 0.009229304268956184, + 0.01788644678890705, + 0.014408224262297153, + 0.004704187158495188, + 0.005301401484757662, + 0.010371160693466663, + 0.010018900036811829, + 0.0052671185694634914, + 0.006785669829696417, + 0.008240476250648499, + 0.00857502780854702, + 0.02830471470952034, + 0.010988438501954079, + 0.01356225460767746, + 0.006429747678339481, + 0.010411935858428478, + 0.04446267709136009, + 0.01188256312161684, + 0.027300996705889702, + 0.02354178950190544, + 0.007331775967031717, + 0.010451561771333218, + 0.0063723246566951275, + 0.0064978087320923805, + 0.01749308779835701, + 0.006630317773669958, + 0.009861699305474758, + 0.0077557940967381, + 0.007599980104714632, + 0.006758412811905146, + 0.004637313075363636, + 0.005807873792946339, + 0.005662377458065748, + 0.005263549275696278, + 0.008980577811598778, + 0.016536040231585503, + 0.01376925315707922, + 0.027086004614830017, + 0.015060171484947205, + 0.027272101491689682, + 0.019748438149690628, + 0.012255084700882435, + 0.019886095076799393, + 0.01914023794233799, + 0.02098899707198143, + 0.01240019965916872, + 0.012758099474012852, + 0.012705527245998383, + 0.014351973310112953, + 0.008442079648375511, + 0.013060111552476883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.0890565812587738 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016059687361121178 + ], + "opposite_logprob_tokens": [ + -4.141059875488281 + ], + "logprob_reward": -0.016059687361121178, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016059687361121178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19744455814361572, + 0.061233192682266235, + 0.043147068470716476, + 0.0253425482660532, + 0.06430182605981827, + 0.09835676103830338, + 0.10054855048656464, + 0.06753688305616379, + 0.15936046838760376, + 0.1827281266450882 + ], + "credit_map": [ + 0.043585628271102905, + 0.027812333777546883, + 0.016714613884687424, + 0.01799985021352768, + 0.0142236752435565, + 0.029668232426047325, + 0.021808570250868797, + 0.02563166245818138, + 0.011408321559429169, + 0.00921336654573679, + 0.005722392816096544, + 0.0059844027273356915, + 0.00477296207100153, + 0.008490465581417084, + 0.008399908430874348, + 0.007241374813020229, + 0.006398333702236414, + 0.004954158328473568, + 0.005056976806372404, + 0.004752275999635458, + 0.004494562279433012, + 0.004912232514470816, + 0.006944663822650909, + 0.005633860360831022, + 0.002960185520350933, + 0.004210418090224266, + 0.00243863626383245, + 0.0027336550410836935, + 0.002422092715278268, + 0.0023070124443620443, + 0.004880600143224001, + 0.0033899496775120497, + 0.0053762756288051605, + 0.007574958726763725, + 0.009264081716537476, + 0.009521554224193096, + 0.006038874853402376, + 0.00867078173905611, + 0.008602947928011417, + 0.009252350777387619, + 0.020572124049067497, + 0.009413943625986576, + 0.011313720606267452, + 0.005894610658288002, + 0.008172912523150444, + 0.020698364824056625, + 0.007935082539916039, + 0.014355995692312717, + 0.023247256875038147, + 0.009358825162053108, + 0.014109096489846706, + 0.006225063465535641, + 0.008826219476759434, + 0.014408490620553493, + 0.011406129226088524, + 0.012967468239367008, + 0.014058013446629047, + 0.006849169731140137, + 0.007680070120841265, + 0.004390288144350052, + 0.012660879641771317, + 0.010631580837070942, + 0.0038493836764246225, + 0.007417495362460613, + 0.016744602471590042, + 0.0256829671561718, + 0.01487414538860321, + 0.0320601649582386, + 0.022521542385220528, + 0.018455462530255318, + 0.014302024617791176, + 0.014719563536345959, + 0.014646921306848526, + 0.033174656331539154, + 0.022479714825749397, + 0.03927715867757797, + 0.022918378934264183, + 0.02155788242816925, + 0.015272353775799274, + 0.013401063159108162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.016059687361121178 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful overhead shot begins, followed by a steady, slow pull-back to reveal a magnificent, epic scene. Atop a steep, black mountaintop shrouded in mist, the legendary hero Sun Wukong stands proudly. He clad in a suit of intricately carved golden chainmail armor, each plate reflecting divinely inspired golden beams that pierce the dark clouds. The dark red lining at the armor's joints is faintly visible. He wears a phoenix-winged purple-gold crown, two slender pheasant feathers dancing fiercely in the mountaintop wind. Behind him, a vast scarlet robe rustles in the wind, almost blending into the swirling leaden sea of clouds. With his legs slightly bent, he remains rooted to the jagged ground like a rock. One hand grips his Ruyi Jingu Bang, embedded in a crevice in the rock, the other clenched into a fist. He jerked his head back, his features tense with utter fury. His fiery eyes burned with unyielding flames, his jaws gaping as if to devour the heavens and earth, and he let out a long, earth-shaking roar. This roar was so powerful that the air around him distorted and rippled visibly, and the gravel and dust beneath his feet trembled with the vibrations of an invisible force. The entire scene, with its striking contrast of gold, crimson, and dark gray, exuded a tragic atmosphere of aloofness, unyielding determination, and a mythical power.", + "seed": 2981, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_14/index_1769_seed_2981_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08410202451050282, + "vlm_logprob": -0.08410202451050282, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.08410202451050282, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35282838344573975 + ], + "opposite_logprob_tokens": [ + -1.2278283834457397 + ], + "logprob_reward": -0.35282838344573975, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35282838344573975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23915314674377441, + 0.016021044924855232, + 0.015027657151222229, + 0.02205831930041313, + 0.2037794291973114, + 0.1002206802368164, + 0.30971062183380127, + 0.04558005556464195, + 0.03551644831895828, + 0.01293257623910904 + ], + "credit_map": [ + 0.010911780409514904, + 0.005733927711844444, + 0.00987336691468954, + 0.01040640939027071, + 0.005216839257627726, + 0.10200504213571548, + 0.00670271273702383, + 0.08830307424068451, + 0.0016997414641082287, + 0.0018996770959347486, + 0.002259837696328759, + 0.0015506087802350521, + 0.002106706378981471, + 0.002117660827934742, + 0.001645393786020577, + 0.00274141994304955, + 0.0014752316055819392, + 0.0013892517890781164, + 0.0021205637603998184, + 0.0016986069967970252, + 0.0016308330232277513, + 0.002416429342702031, + 0.001683118287473917, + 0.0026136236265301704, + 0.0016472372226417065, + 0.0015409262850880623, + 0.0037779316771775484, + 0.0016079641645774245, + 0.0018800812540575862, + 0.006265157368034124, + 0.001677033375017345, + 0.0036619899328798056, + 0.007932865060865879, + 0.005435397382825613, + 0.021668124943971634, + 0.005345440004020929, + 0.006226617377251387, + 0.07259561866521835, + 0.006594625301659107, + 0.07798075675964355, + 0.011629411950707436, + 0.006988596171140671, + 0.021909590810537338, + 0.007553070317953825, + 0.006580174434930086, + 0.02263287454843521, + 0.010026396252214909, + 0.012900574132800102, + 0.03278433531522751, + 0.021048441529273987, + 0.06843080371618271, + 0.018052082508802414, + 0.019142309203743935, + 0.05824572220444679, + 0.02864554524421692, + 0.06336141377687454, + 0.004107809159904718, + 0.0026267373468726873, + 0.010146989487111568, + 0.00275163515470922, + 0.0020794107113033533, + 0.008227582089602947, + 0.004523651674389839, + 0.011116241104900837, + 0.0030848863534629345, + 0.004163108300417662, + 0.005603440571576357, + 0.0036765479017049074, + 0.0026932116597890854, + 0.005680200178176165, + 0.006421368103474379, + 0.00419369013980031, + 0.00097280292538926, + 0.0006907194619998336, + 0.0027768530417233706, + 0.0009438538108952343, + 0.0008082606946118176, + 0.0018777224468067288, + 0.0020167608745396137, + 0.002845603972673416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.35282838344573975 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651040494441986 + ], + "opposite_logprob_tokens": [ + -4.758651256561279 + ], + "logprob_reward": -0.008651040494441986, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651040494441986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4503866732120514, + 0.02020302228629589, + 0.0459294319152832, + 0.02553989365696907, + 0.16435906291007996, + 0.03835561126470566, + 0.11886485666036606, + 0.05758662521839142, + 0.04797062650322914, + 0.030804192647337914 + ], + "credit_map": [ + 0.01926019974052906, + 0.008789652027189732, + 0.011514314450323582, + 0.019316280260682106, + 0.007388334255665541, + 0.19920344650745392, + 0.012938493862748146, + 0.1719759702682495, + 0.0016492988215759397, + 0.001996455015614629, + 0.0020878720097243786, + 0.0025169767905026674, + 0.0018634072039276361, + 0.0036618856247514486, + 0.001904572476632893, + 0.0045225536450743675, + 0.00315863243304193, + 0.002435005037114024, + 0.006322183646261692, + 0.0021988938096910715, + 0.003694943618029356, + 0.016292212530970573, + 0.00278619397431612, + 0.00904136709868908, + 0.002542823553085327, + 0.0015736074419692159, + 0.005733331199735403, + 0.001979768043383956, + 0.0020842431113123894, + 0.006528004072606564, + 0.0013955754693597555, + 0.00370253948494792, + 0.004158776719123125, + 0.0022242385894060135, + 0.011228087358176708, + 0.0022811738308519125, + 0.002682880498468876, + 0.06856352090835571, + 0.0029180191922932863, + 0.07030236721038818, + 0.0035982229746878147, + 0.0022624912671744823, + 0.009355287998914719, + 0.002919127931818366, + 0.002209868049249053, + 0.008895490318536758, + 0.003934380132704973, + 0.005180742125958204, + 0.01071296539157629, + 0.0037412040401250124, + 0.030479885637760162, + 0.004983231425285339, + 0.006053837016224861, + 0.02536524087190628, + 0.01123951468616724, + 0.02628897689282894, + 0.005295244045555592, + 0.0032926551066339016, + 0.015835508704185486, + 0.002823009854182601, + 0.0037210991140455008, + 0.010811596177518368, + 0.006180251482874155, + 0.009627258405089378, + 0.00390451168641448, + 0.002771561499685049, + 0.013306749984622002, + 0.003456278471276164, + 0.00274273962713778, + 0.008074956014752388, + 0.0047799963504076, + 0.008933834731578827, + 0.003352789208292961, + 0.002633740659803152, + 0.006050944793969393, + 0.0032572150230407715, + 0.0020197078119963408, + 0.0046868291683495045, + 0.003941692877560854, + 0.004861272871494293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.008651040494441986 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952132865786552 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952132865786552, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952132865786552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5279409289360046, + 0.03698199242353439, + 0.03387884795665741, + 0.04640476033091545, + 0.10641420632600784, + 0.0234702005982399, + 0.1501942276954651, + 0.02287456765770912, + 0.031867168843746185, + 0.019973093643784523 + ], + "credit_map": [ + 0.02534501999616623, + 0.015724100172519684, + 0.017890654504299164, + 0.030448289588093758, + 0.013554475270211697, + 0.22192443907260895, + 0.016383321955800056, + 0.18667057156562805, + 0.0045986478216946125, + 0.003947337623685598, + 0.00433738948777318, + 0.005482777021825314, + 0.0029660300351679325, + 0.0058761038817465305, + 0.0027961665764451027, + 0.006977537181228399, + 0.00577213428914547, + 0.002791032660752535, + 0.003955271560698748, + 0.0032477667555212975, + 0.0033741772640496492, + 0.006487860810011625, + 0.0029372095596045256, + 0.005313396453857422, + 0.004776474554091692, + 0.00277736340649426, + 0.007806670852005482, + 0.0036448759492486715, + 0.0037013678811490536, + 0.013646237552165985, + 0.002840018831193447, + 0.007211754098534584, + 0.003217934165149927, + 0.0022983981762081385, + 0.006660127080976963, + 0.001845255377702415, + 0.0027304564137011766, + 0.043234601616859436, + 0.002526714699342847, + 0.04390072822570801, + 0.003737444756552577, + 0.0027265322860330343, + 0.0035464034881442785, + 0.0019958089105784893, + 0.002736444119364023, + 0.004053252749145031, + 0.002276182174682617, + 0.0023981311824172735, + 0.01409245003014803, + 0.00853265356272459, + 0.03507029637694359, + 0.007889975793659687, + 0.008317911066114902, + 0.02974812500178814, + 0.014505543746054173, + 0.03203727304935455, + 0.0025191791355609894, + 0.0023661653976887465, + 0.004241178743541241, + 0.0016996326157823205, + 0.001825551618821919, + 0.0037803074810653925, + 0.0031508225947618484, + 0.0032917300704866648, + 0.002004612935706973, + 0.002170110121369362, + 0.007571820169687271, + 0.002509847516193986, + 0.0014129278715699911, + 0.005418913438916206, + 0.004798920825123787, + 0.005980018060654402, + 0.0015062454622238874, + 0.0011485478607937694, + 0.005279350560158491, + 0.0015359591925516725, + 0.0013055377639830112, + 0.0028280497062951326, + 0.002645258093252778, + 0.0037241443060338497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.005952132865786552 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029802370816469193 + ], + "opposite_logprob_tokens": [ + -3.5298023223876953 + ], + "logprob_reward": -0.029802370816469193, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029802370816469193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.338683158159256, + 0.03858368843793869, + 0.048257194459438324, + 0.05541945993900299, + 0.09574851393699646, + 0.028445087373256683, + 0.11627721786499023, + 0.12456578761339188, + 0.10009966045618057, + 0.05392029508948326 + ], + "credit_map": [ + 0.016273053362965584, + 0.00899070966988802, + 0.011349565349519253, + 0.017258865758776665, + 0.008014066144824028, + 0.143327996134758, + 0.01063693780452013, + 0.12283195555210114, + 0.006245742551982403, + 0.004200146999210119, + 0.0040295361541211605, + 0.003961554728448391, + 0.004458706825971603, + 0.006044344510883093, + 0.0038154476787894964, + 0.005828211084008217, + 0.007828089408576488, + 0.004705868661403656, + 0.00604931591078639, + 0.003883309895172715, + 0.005589142441749573, + 0.009765587747097015, + 0.002979222685098648, + 0.007456657942384481, + 0.0068939803168177605, + 0.008635909296572208, + 0.007673157379031181, + 0.004296232480555773, + 0.006452325731515884, + 0.009942164644598961, + 0.004016943275928497, + 0.007508744485676289, + 0.0028525847010314465, + 0.002049343893304467, + 0.004500577691942453, + 0.0018329316517338157, + 0.0025609666481614113, + 0.04010947421193123, + 0.0017453866312280297, + 0.04009724780917168, + 0.0032488645520061255, + 0.0021429727785289288, + 0.005032699555158615, + 0.0021716002374887466, + 0.0023948191665112972, + 0.0059298379346728325, + 0.0027329300064593554, + 0.004791361279785633, + 0.01346391998231411, + 0.005657791160047054, + 0.029199950397014618, + 0.0062174201011657715, + 0.008309375494718552, + 0.023951206356287003, + 0.010107171721756458, + 0.01937037520110607, + 0.00870122667402029, + 0.007289598695933819, + 0.02799651212990284, + 0.007478892337530851, + 0.0075532919727265835, + 0.023574193939566612, + 0.0177365243434906, + 0.024235544726252556, + 0.006041661370545626, + 0.007924429140985012, + 0.02288474142551422, + 0.008710966445505619, + 0.005027612671256065, + 0.016850469633936882, + 0.01700960472226143, + 0.015650177374482155, + 0.004125862382352352, + 0.002635946962982416, + 0.013919135555624962, + 0.0038974015042185783, + 0.0036297861952334642, + 0.0075822630897164345, + 0.007478426210582256, + 0.010651472955942154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.029802370816469193 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an overhead shot that pulls back slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, black mountaintop shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Sun Wukong wearing intricately carved golden chainmail armor with dark red lining at the joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is Sun Wukong holding his Ruyi Jingu Bang embedded in a rock crevice with one hand while the other is clenched into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character exhibit a roaring expression with fiery eyes, a gaping jaw, and visible air distortion around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232761949300766 + ], + "opposite_logprob_tokens": [ + -3.7732760906219482 + ], + "logprob_reward": -0.0232761949300766, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232761949300766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6115971207618713, + 0.011282611638307571, + 0.026827551424503326, + 0.02481166645884514, + 0.10324500501155853, + 0.01453434955328703, + 0.1422879844903946, + 0.03636598959565163, + 0.021541854366660118, + 0.007505846209824085 + ], + "credit_map": [ + 0.02602837048470974, + 0.012856653891503811, + 0.015707263723015785, + 0.026292549446225166, + 0.010624376125633717, + 0.27022314071655273, + 0.017189135774970055, + 0.23267559707164764, + 0.0016483233775943518, + 0.0014180578291416168, + 0.0012679318897426128, + 0.0014575779205188155, + 0.0011870436137542129, + 0.0016461667837575078, + 0.0009962511248886585, + 0.0016612574690952897, + 0.004247347358614206, + 0.002756620291620493, + 0.0032741776667535305, + 0.002866052556782961, + 0.002757746959105134, + 0.0044403537176549435, + 0.002344694221392274, + 0.004140555392950773, + 0.002491028979420662, + 0.003414982696995139, + 0.003972701262682676, + 0.0026129870675504208, + 0.002530217170715332, + 0.004035043064504862, + 0.002110699424520135, + 0.0036440046969801188, + 0.0032204147428274155, + 0.0015953484689816833, + 0.005549890920519829, + 0.0013854880817234516, + 0.001986760413274169, + 0.043798696249723434, + 0.0017525405855849385, + 0.043955858796834946, + 0.001926477998495102, + 0.0011612176895141602, + 0.0022804192267358303, + 0.0011802123626694083, + 0.001149230170994997, + 0.002720147604122758, + 0.0014486483996734023, + 0.0026679947040975094, + 0.013361234217882156, + 0.007840707898139954, + 0.03247208148241043, + 0.007845872081816196, + 0.007963434793055058, + 0.027955634519457817, + 0.013898183591663837, + 0.030950821936130524, + 0.003918084315955639, + 0.0022262237034738064, + 0.006016766186803579, + 0.002319966210052371, + 0.0020903802942484617, + 0.00640786113217473, + 0.0043680802918970585, + 0.00901862420141697, + 0.0018580507021397352, + 0.0016344680916517973, + 0.0042652408592402935, + 0.002049572765827179, + 0.0015997568843886256, + 0.003424234688282013, + 0.0033281671348959208, + 0.0033823621924966574, + 0.0007660685223527253, + 0.0006993802380748093, + 0.0014113779179751873, + 0.0007714872481301427, + 0.000608051719609648, + 0.0012189738918095827, + 0.0010901938658207655, + 0.0009403122821822762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0232761949300766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_15/index_tensor([5189, 5189])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_15/index_tensor([5189, 5189])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d138997bb10ee0010be81071d0fd0892b5d10ec8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_15/index_tensor([5189, 5189])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, medium shot captures a standard one-gallon metal paint can, its exterior a vibrant primary red, precariously balanced on the cracked, dusty concrete floor of a garage. Bright, direct daylight streams in from a high, off-camera source, casting sharp shadows and gleaming across the can's semi-glossy, slightly worn surface. The unsealed metal lid tilts askew on its rim, a clear sign of its looseness. Pulled by gravity, the can begins a slow, deliberate, and irreversible tilt, its motion accelerating as it loses its balance and tips over. It strikes the gravel with a sharp, metallic thud, the impact instantly forcing the loose lid to spring free and fly aside. Simultaneously, a glob of viscous red paint violently erupts from the hole in a dramatic arcing wave, breaking into small globules and tiny droplets in mid-air and splattering a chaotic, irregular, sheen of crimson on the cold gray floor. This contrasts sharply with the dull, stained concrete, creating a scene of sudden, chaotic disaster.", + "seed": 24899, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_15/index_5189_seed_24899_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.014565777126699686, + "vlm_logprob": -0.014565777126699686, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.014565777126699686, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02018061839044094 + ], + "opposite_logprob_tokens": [ + -4.395180702209473 + ], + "logprob_reward": -0.02018061839044094, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02018061839044094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0877169817686081, + 0.05233026668429375, + 0.06123421713709831, + 0.08574914932250977, + 0.26985156536102295, + 0.10288142412900925, + 0.1482374668121338, + 0.04345721751451492, + 0.06580749899148941, + 0.08273418247699738 + ], + "credit_map": [ + 0.009681452065706253, + 0.007971962913870811, + 0.014195350930094719, + 0.022622235119342804, + 0.008805410005152225, + 0.0072827422991395, + 0.006661284249275923, + 0.010496554896235466, + 0.005528269335627556, + 0.0054813954047858715, + 0.005279717966914177, + 0.009966370649635792, + 0.007668210193514824, + 0.006030408199876547, + 0.006837909575551748, + 0.005537988152354956, + 0.0070026470348238945, + 0.011991755105555058, + 0.00559981819242239, + 0.01061223354190588, + 0.006949176080524921, + 0.005603746511042118, + 0.005543573759496212, + 0.007931273430585861, + 0.01638973318040371, + 0.005704169627279043, + 0.01632341369986534, + 0.012520931661128998, + 0.008009405806660652, + 0.007696271874010563, + 0.010483164340257645, + 0.008622064255177975, + 0.020902464166283607, + 0.01872081868350506, + 0.09445660561323166, + 0.015325005166232586, + 0.016953065991401672, + 0.013129686936736107, + 0.07294926792383194, + 0.017414679750800133, + 0.014803090132772923, + 0.02193727344274521, + 0.01488521322607994, + 0.009356708265841007, + 0.0092784333974123, + 0.008388756774365902, + 0.0124933160841465, + 0.011738637462258339, + 0.045297231525182724, + 0.012383869849145412, + 0.008252090774476528, + 0.005522145889699459, + 0.005218574311584234, + 0.013200962916016579, + 0.006457649637013674, + 0.05190495401620865, + 0.007581484504044056, + 0.0042335521429777145, + 0.003991689067333937, + 0.004758425056934357, + 0.0022856055293232203, + 0.006885232403874397, + 0.003210528986528516, + 0.0105107007548213, + 0.017326179891824722, + 0.0072471038438379765, + 0.0193485114723444, + 0.00396785419434309, + 0.006290608085691929, + 0.0035696353297680616, + 0.003811026457697153, + 0.004246584139764309, + 0.011033873073756695, + 0.01293200347572565, + 0.014134875498712063, + 0.00926329754292965, + 0.009226273745298386, + 0.009402030147612095, + 0.009140688925981522, + 0.007601152174174786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.02018061839044094 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372391685843468 + ], + "opposite_logprob_tokens": [ + -3.408723831176758 + ], + "logprob_reward": -0.03372391685843468, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372391685843468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.158258318901062, + 0.0977068543434143, + 0.07884903252124786, + 0.061625342816114426, + 0.2604508399963379, + 0.1083732396364212, + 0.03515777736902237, + 0.04960091412067413, + 0.10933414101600647, + 0.04064354673027992 + ], + "credit_map": [ + 0.020114298909902573, + 0.021636981517076492, + 0.022307295352220535, + 0.031738538295030594, + 0.025590652599930763, + 0.01539777871221304, + 0.010660319589078426, + 0.01081245020031929, + 0.012709305621683598, + 0.011370250955224037, + 0.01088270079344511, + 0.01574486866593361, + 0.011520308442413807, + 0.013662335462868214, + 0.010077554732561111, + 0.011739526875317097, + 0.020232610404491425, + 0.01083870604634285, + 0.007609352003782988, + 0.007039492484182119, + 0.006451405584812164, + 0.005346821621060371, + 0.00831754133105278, + 0.013013101182878017, + 0.011051901616156101, + 0.0035770623944699764, + 0.012289349921047688, + 0.004864929243922234, + 0.005433123558759689, + 0.005026993341743946, + 0.008850021287798882, + 0.010531962849199772, + 0.01920454390347004, + 0.012412873096764088, + 0.08663073182106018, + 0.014711995609104633, + 0.017441730946302414, + 0.01257589552551508, + 0.07912857085466385, + 0.01834450103342533, + 0.010575015097856522, + 0.014625932089984417, + 0.02482467330992222, + 0.014956789091229439, + 0.00919283740222454, + 0.01005118153989315, + 0.01110828761011362, + 0.013038520701229572, + 0.006523532792925835, + 0.006725086364895105, + 0.0034930414985865355, + 0.0029903429094702005, + 0.002802102593705058, + 0.0032660728320479393, + 0.0018502274760976434, + 0.0075073703192174435, + 0.006779639050364494, + 0.004676932469010353, + 0.005802610889077187, + 0.005295104812830687, + 0.003959932364523411, + 0.01121832150965929, + 0.004174159839749336, + 0.007694212254136801, + 0.026246512308716774, + 0.010896030813455582, + 0.029961571097373962, + 0.008034724742174149, + 0.00992223434150219, + 0.010213113389909267, + 0.006557208485901356, + 0.00750274583697319, + 0.006395315285772085, + 0.00739851500838995, + 0.0072404323145747185, + 0.0034725230652838945, + 0.004002544097602367, + 0.004957491997629404, + 0.00393039733171463, + 0.0032463273964822292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.03372391685843468 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089949868619442 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.011089949868619442, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089949868619442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19507722556591034, + 0.09082008898258209, + 0.06887756288051605, + 0.07454497367143631, + 0.13633759319782257, + 0.14078669250011444, + 0.09511128813028336, + 0.08898252248764038, + 0.05581693351268768, + 0.05364512652158737 + ], + "credit_map": [ + 0.02143481746315956, + 0.02261815406382084, + 0.02478846162557602, + 0.03527107089757919, + 0.03038412146270275, + 0.02580253966152668, + 0.019147198647260666, + 0.015630878508090973, + 0.00984838604927063, + 0.009991437196731567, + 0.008840260095894337, + 0.013212120160460472, + 0.012950305826961994, + 0.012308154255151749, + 0.010852736420929432, + 0.01281669270247221, + 0.02008223533630371, + 0.010893096216022968, + 0.007138760760426521, + 0.006105565465986729, + 0.006130322348326445, + 0.005796818528324366, + 0.006598479580134153, + 0.006132290232926607, + 0.013587916269898415, + 0.005000036675482988, + 0.01201116107404232, + 0.009746264666318893, + 0.006688022520393133, + 0.00663814228028059, + 0.012152622453868389, + 0.008720814250409603, + 0.012206572107970715, + 0.01241826731711626, + 0.04226476326584816, + 0.010401234030723572, + 0.011720716953277588, + 0.009709146805107594, + 0.028873126953840256, + 0.00874378252774477, + 0.01830778457224369, + 0.01882331632077694, + 0.035660553723573685, + 0.010853751562535763, + 0.013134052976965904, + 0.013232077471911907, + 0.013308433815836906, + 0.017466722056269646, + 0.030032530426979065, + 0.008153518661856651, + 0.004707189276814461, + 0.003543191822245717, + 0.003354114480316639, + 0.00822839792817831, + 0.004041185136884451, + 0.03305116295814514, + 0.011914609000086784, + 0.01198136992752552, + 0.012496751733124256, + 0.009902949444949627, + 0.007834955118596554, + 0.01784340851008892, + 0.0067807878367602825, + 0.01022769883275032, + 0.009629280306398869, + 0.010757913812994957, + 0.005776984617114067, + 0.006152661517262459, + 0.004889505449682474, + 0.0094867879524827, + 0.003969880286604166, + 0.005153925158083439, + 0.010316374711692333, + 0.009354744106531143, + 0.011265481822192669, + 0.003928604070097208, + 0.004690423142164946, + 0.005624592769891024, + 0.0038334536366164684, + 0.00463145412504673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.011089949868619442 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319446949288249 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.00319446949288249, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319446949288249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06195303425192833, + 0.044861771166324615, + 0.026744795963168144, + 0.05764298886060715, + 0.16976919770240784, + 0.21088062226772308, + 0.14795784652233124, + 0.1786319464445114, + 0.032059621065855026, + 0.06949818134307861 + ], + "credit_map": [ + 0.00512497266754508, + 0.006765617989003658, + 0.0073641822673380375, + 0.010654956102371216, + 0.01200832612812519, + 0.006708839908242226, + 0.006724224425852299, + 0.006601914297789335, + 0.003851584391668439, + 0.005831947550177574, + 0.004321600776165724, + 0.007003135979175568, + 0.007281185127794743, + 0.006151314824819565, + 0.0053455764427781105, + 0.005075423512607813, + 0.005655655171722174, + 0.004258052445948124, + 0.002536663319915533, + 0.003305460326373577, + 0.002665334613993764, + 0.002749897073954344, + 0.002419069642201066, + 0.0031546626705676317, + 0.008195186965167522, + 0.0027489501517266035, + 0.0075501808896660805, + 0.005136427469551563, + 0.004013995174318552, + 0.005292237736284733, + 0.01215894054621458, + 0.012547069229185581, + 0.012905930168926716, + 0.009689176455140114, + 0.05788134038448334, + 0.010224477387964725, + 0.011209828779101372, + 0.008827853947877884, + 0.047862790524959564, + 0.01116781122982502, + 0.02703426219522953, + 0.014324851334095001, + 0.06101922318339348, + 0.013490002602338791, + 0.019699713215231895, + 0.021483397111296654, + 0.025250928476452827, + 0.02857823483645916, + 0.044533949345350266, + 0.011781488545238972, + 0.012273512780666351, + 0.006018576677888632, + 0.004823149647563696, + 0.013597794808447361, + 0.006505373865365982, + 0.048423994332551956, + 0.030138134956359863, + 0.023051735013723373, + 0.02068346180021763, + 0.02215881459414959, + 0.015504769049584866, + 0.028241071850061417, + 0.012441658414900303, + 0.02641231007874012, + 0.007328476291149855, + 0.004196130204945803, + 0.007892132736742496, + 0.002305798465386033, + 0.0032684809993952513, + 0.0023458299692720175, + 0.0020270918030291796, + 0.00269567989744246, + 0.011204085312783718, + 0.012310962192714214, + 0.010156916454434395, + 0.005702046677470207, + 0.007295700255781412, + 0.009505012072622776, + 0.005398457404226065, + 0.007925000041723251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.00319446949288249 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463993102312088 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.00463993102312088, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463993102312088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22143231332302094, + 0.06225042790174484, + 0.05707776919007301, + 0.06864384561777115, + 0.21858538687229156, + 0.09081355482339859, + 0.09329817444086075, + 0.08370599150657654, + 0.02654346264898777, + 0.0776490867137909 + ], + "credit_map": [ + 0.025738392025232315, + 0.03325441852211952, + 0.03113796003162861, + 0.04485359415411949, + 0.030483625829219818, + 0.021608129143714905, + 0.014607388526201248, + 0.019748814404010773, + 0.006657975260168314, + 0.005975809413939714, + 0.006002042442560196, + 0.01260366104543209, + 0.008894074708223343, + 0.007715808227658272, + 0.006688993889838457, + 0.007712060119956732, + 0.014535491354763508, + 0.00806535966694355, + 0.006588749587535858, + 0.004858970642089844, + 0.005380196496844292, + 0.004217161796987057, + 0.0057864002883434296, + 0.007645441219210625, + 0.013740679249167442, + 0.0031732236966490746, + 0.01059905905276537, + 0.005774027202278376, + 0.0036421581171453, + 0.005186017602682114, + 0.013953088782727718, + 0.012575591914355755, + 0.014947915449738503, + 0.010101212188601494, + 0.07392855733633041, + 0.012019661255180836, + 0.014586846344172955, + 0.009997379034757614, + 0.06732924282550812, + 0.015674568712711334, + 0.01045526098459959, + 0.011828385293483734, + 0.017768941819667816, + 0.008209805935621262, + 0.00838506780564785, + 0.009800570085644722, + 0.011451426893472672, + 0.012914096936583519, + 0.0258356761187315, + 0.008219612762331963, + 0.010850638151168823, + 0.0044622719287872314, + 0.003730381140485406, + 0.008499679155647755, + 0.0037519459147006273, + 0.02794797159731388, + 0.014308186247944832, + 0.009209776297211647, + 0.011021455749869347, + 0.008044609799981117, + 0.008760133758187294, + 0.014834287576377392, + 0.005198589060455561, + 0.012328947894275188, + 0.004864037968218327, + 0.004194133449345827, + 0.004794248845428228, + 0.002571208169683814, + 0.0032808026298880577, + 0.002889449940994382, + 0.0015642507933080196, + 0.0023853329475969076, + 0.008405613712966442, + 0.016828417778015137, + 0.010395477525889874, + 0.00667203776538372, + 0.009701346047222614, + 0.010167859494686127, + 0.00751616433262825, + 0.007962171919643879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.00463993102312088 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium shot captures a standard one-gallon metal paint can, its exterior a vibrant primary red, precariously balanced on the cracked, dusty concrete floor of a garage. Bright, direct daylight streams in from a high, off-camera source, casting sharp shadows and gleaming across the can's semi-glossy, slightly worn surface. The unsealed metal lid tilts askew on its rim, a clear sign of its looseness. Pulled by gravity, the can begins a slow, deliberate, and irreversible tilt, its motion accelerating as it loses its balance and tips over. It strikes the gravel with a sharp, metallic thud, the impact instantly forcing the loose lid to spring free and fly aside. Simultaneously, a glob of viscous red paint violently erupts from the hole in a dramatic arcing wave, breaking into small globules and tiny droplets in mid-air and splattering a chaotic, irregular, sheen of crimson on the cold gray floor. This contrasts sharply with the dull, stained concrete, creating a scene of sudden, chaotic disaster.", + "seed": 24900, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_15/index_5189_seed_24900_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03379634679295122, + "vlm_logprob": -0.03379634679295122, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.03379634679295122, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02237526886165142 + ], + "opposite_logprob_tokens": [ + -4.647375106811523 + ], + "logprob_reward": -0.02237526886165142, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.02237526886165142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11166337132453918, + 0.09148271381855011, + 0.16028296947479248, + 0.034265466034412384, + 0.10862140357494354, + 0.02732798457145691, + 0.02466900646686554, + 0.3624308109283447, + 0.03936677798628807, + 0.03988952189683914 + ], + "credit_map": [ + 0.012041036039590836, + 0.022439366206526756, + 0.019690727815032005, + 0.01037371065467596, + 0.016262656077742577, + 0.013203078880906105, + 0.006267414893954992, + 0.011385388672351837, + 0.006997949443757534, + 0.011623076163232327, + 0.013050300069153309, + 0.007645017467439175, + 0.018164079636335373, + 0.013820672407746315, + 0.011080795899033546, + 0.009100833907723427, + 0.01400931365787983, + 0.010354980826377869, + 0.028211692348122597, + 0.010431663133203983, + 0.020512791350483894, + 0.04034459590911865, + 0.017579231411218643, + 0.01883871667087078, + 0.0031280319672077894, + 0.003984026610851288, + 0.004567608702927828, + 0.0026613345835357904, + 0.0037774285301566124, + 0.006262375507503748, + 0.00448335986584425, + 0.00540130352601409, + 0.006544750649482012, + 0.007382160983979702, + 0.02069040946662426, + 0.007623021025210619, + 0.007859106175601482, + 0.024401813745498657, + 0.012893964536488056, + 0.021226191893219948, + 0.0021998474840074778, + 0.003605361795052886, + 0.0045585329644382, + 0.0025674731004983187, + 0.002779843984171748, + 0.0064314319752156734, + 0.0029310628306120634, + 0.0022544325329363346, + 0.002018920611590147, + 0.00261447811499238, + 0.003991786390542984, + 0.001898634247481823, + 0.0031573560554534197, + 0.005285043269395828, + 0.0025246357545256615, + 0.003178154118359089, + 0.024402720853686333, + 0.02289188839495182, + 0.06198182702064514, + 0.022402074187994003, + 0.024573689326643944, + 0.04542873054742813, + 0.04914001747965813, + 0.11160989850759506, + 0.0030842681881040335, + 0.003859590506181121, + 0.004810221493244171, + 0.002868321491405368, + 0.0035159317776560783, + 0.006598828826099634, + 0.006508592050522566, + 0.008121025748550892, + 0.0035054783802479506, + 0.006281048525124788, + 0.004312662873417139, + 0.003945620264858007, + 0.006068115588277578, + 0.005153728183358908, + 0.00453522615134716, + 0.006087649147957563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.02237526886165142 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902006804943085 + ], + "opposite_logprob_tokens": [ + -2.5790200233459473 + ], + "logprob_reward": -0.07902006804943085, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902006804943085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27840185165405273, + 0.04959341883659363, + 0.052911777049303055, + 0.16574127972126007, + 0.09358227252960205, + 0.027613792568445206, + 0.051278483122587204, + 0.19064098596572876, + 0.07594478875398636, + 0.014291385188698769 + ], + "credit_map": [ + 0.01911398023366928, + 0.03891777992248535, + 0.0390433743596077, + 0.01681315153837204, + 0.0516154021024704, + 0.05954970419406891, + 0.0185705479234457, + 0.03477790579199791, + 0.004665665794163942, + 0.004303110763430595, + 0.008618594147264957, + 0.002792560961097479, + 0.005897803697735071, + 0.011909598484635353, + 0.005328466184437275, + 0.006077620200812817, + 0.0038491350132972, + 0.004084480926394463, + 0.008695892058312893, + 0.003128957934677601, + 0.006611154414713383, + 0.014997456222772598, + 0.006113906856626272, + 0.005430789664387703, + 0.01475534588098526, + 0.01745324768126011, + 0.028498440980911255, + 0.01198683213442564, + 0.017893271520733833, + 0.04294930025935173, + 0.013026359491050243, + 0.019178470596671104, + 0.006664159242063761, + 0.0068548619747161865, + 0.017045658081769943, + 0.004858810920268297, + 0.00715664541348815, + 0.02275756374001503, + 0.009468194097280502, + 0.01877637766301632, + 0.0018861680291593075, + 0.003281560493633151, + 0.0042744409292936325, + 0.0026886628475040197, + 0.0020694350823760033, + 0.005659447051584721, + 0.0033504748716950417, + 0.004403604660183191, + 0.00406329520046711, + 0.004535211715847254, + 0.007169587537646294, + 0.0040142168290913105, + 0.006816972512751818, + 0.007344200741499662, + 0.007826565764844418, + 0.00950843095779419, + 0.016094839200377464, + 0.01124628260731697, + 0.026833107694983482, + 0.014252844266593456, + 0.02050141431391239, + 0.03710836172103882, + 0.03127405792474747, + 0.033330075442790985, + 0.007387144491076469, + 0.006308727897703648, + 0.011185931041836739, + 0.0044793798588216305, + 0.011387373320758343, + 0.0159947257488966, + 0.006994583643972874, + 0.012206926941871643, + 0.001196284662000835, + 0.001786707784049213, + 0.0016580915544182062, + 0.0011564285960048437, + 0.001969012198969722, + 0.0022733386140316725, + 0.0016727494075894356, + 0.00257877167314291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.07902006804943085 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009800530970096588 + ], + "opposite_logprob_tokens": [ + -4.634800434112549 + ], + "logprob_reward": -0.009800530970096588, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009800530970096588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2135489284992218, + 0.09206897765398026, + 0.10240074247121811, + 0.06745713204145432, + 0.16196736693382263, + 0.045567307621240616, + 0.02855616994202137, + 0.213505357503891, + 0.02792649157345295, + 0.04700159281492233 + ], + "credit_map": [ + 0.014880701899528503, + 0.031419798731803894, + 0.034346580505371094, + 0.014068867079913616, + 0.03773939236998558, + 0.0470832996070385, + 0.015572636388242245, + 0.018437635153532028, + 0.008128602989017963, + 0.013410793617367744, + 0.013532598502933979, + 0.007273897062987089, + 0.012300094589591026, + 0.018741214647889137, + 0.008996797725558281, + 0.009684977121651173, + 0.007100685499608517, + 0.007609227206557989, + 0.01810433715581894, + 0.00552668422460556, + 0.009295657277107239, + 0.028425196185708046, + 0.01397903636097908, + 0.012359908781945705, + 0.005606073886156082, + 0.005933053325861692, + 0.011666715145111084, + 0.004282016307115555, + 0.008188052102923393, + 0.018350953236222267, + 0.004994162358343601, + 0.008436106145381927, + 0.009843677282333374, + 0.016922667622566223, + 0.028833957388997078, + 0.013181962072849274, + 0.01347000990062952, + 0.034991730004549026, + 0.018331823870539665, + 0.026391534134745598, + 0.0035160542465746403, + 0.005241899285465479, + 0.0070202243514359, + 0.0035401303321123123, + 0.0042401510290801525, + 0.009495348669588566, + 0.005784350913017988, + 0.006729144137352705, + 0.002429262036457658, + 0.0034805985633283854, + 0.0037688612937927246, + 0.002783816307783127, + 0.003483436070382595, + 0.004751873668283224, + 0.0035947684664279222, + 0.004263549577444792, + 0.013581317849457264, + 0.013157088309526443, + 0.034841302782297134, + 0.01360713317990303, + 0.016657013446092606, + 0.030733883380889893, + 0.032044991850852966, + 0.05888260155916214, + 0.0029170061461627483, + 0.0030420294497162104, + 0.004129453096538782, + 0.002080794656649232, + 0.003205833025276661, + 0.005163153167814016, + 0.002737468108534813, + 0.004650751128792763, + 0.0031884328927844763, + 0.003737657330930233, + 0.0056740776635706425, + 0.0029196925461292267, + 0.003943109884858131, + 0.0055825128220021725, + 0.00663883937522769, + 0.015317266806960106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.009800530970096588 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496411398053169 + ], + "opposite_logprob_tokens": [ + -2.929964065551758 + ], + "logprob_reward": -0.05496411398053169, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496411398053169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04119144007563591, + 0.018200458958745003, + 0.20225375890731812, + 0.05046906694769859, + 0.059901162981987, + 0.022716641426086426, + 0.014531230553984642, + 0.5575228333473206, + 0.01153442170470953, + 0.0216789823025465 + ], + "credit_map": [ + 0.004620274994522333, + 0.005790590308606625, + 0.006407047621905804, + 0.003490663832053542, + 0.005218910984694958, + 0.006301945075392723, + 0.00301771960221231, + 0.006344292312860489, + 0.0014695801073685288, + 0.0022541824728250504, + 0.002849381184205413, + 0.0019136789487674832, + 0.0021917938720434904, + 0.0038977954536676407, + 0.0014864664990454912, + 0.0021375827491283417, + 0.018718669191002846, + 0.01153622753918171, + 0.04284020885825157, + 0.011803827248513699, + 0.01757790893316269, + 0.057152409106492996, + 0.017649633809924126, + 0.02497490495443344, + 0.0039961449801921844, + 0.0070912777446210384, + 0.007524057291448116, + 0.004015016369521618, + 0.005255849100649357, + 0.0110050980001688, + 0.004550727549940348, + 0.007030901033431292, + 0.003696944797411561, + 0.004729850217700005, + 0.010887417010962963, + 0.004403975792229176, + 0.00448593869805336, + 0.013652846217155457, + 0.007822265848517418, + 0.010221932083368301, + 0.001823626458644867, + 0.002605264075100422, + 0.0036552574019879103, + 0.002156586851924658, + 0.0020790838170796633, + 0.004850723315030336, + 0.0025804585311561823, + 0.0029656440019607544, + 0.001262171776033938, + 0.0010181560646742582, + 0.0025071415584534407, + 0.0009273012401536107, + 0.0009125220240093768, + 0.003109809709712863, + 0.0015192755963653326, + 0.003274854738265276, + 0.03558749705553055, + 0.032428160309791565, + 0.09702896326780319, + 0.033370137214660645, + 0.03475319221615791, + 0.06798116117715836, + 0.07750874757766724, + 0.178864985704422, + 0.0009712440078146756, + 0.001238649245351553, + 0.0015485472977161407, + 0.0012327387230470777, + 0.0009915032424032688, + 0.0021882213186472654, + 0.0013827849179506302, + 0.0019807342905551195, + 0.001368174096569419, + 0.0016743441810831428, + 0.002768744947388768, + 0.0015028384514153004, + 0.0017193752573803067, + 0.0019063320942223072, + 0.0033725241664797068, + 0.0073666502721607685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.05496411398053169 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028217521030455828 + ], + "opposite_logprob_tokens": [ + -5.877821922302246 + ], + "logprob_reward": -0.0028217521030455828, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028217521030455828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2218077927827835, + 0.08139625191688538, + 0.13021735846996307, + 0.15642665326595306, + 0.1318303942680359, + 0.028985125944018364, + 0.03150216117501259, + 0.1502537876367569, + 0.03935796394944191, + 0.02822253853082657 + ], + "credit_map": [ + 0.015825526788830757, + 0.03593859449028969, + 0.030749524012207985, + 0.017377199605107307, + 0.03275366500020027, + 0.048000335693359375, + 0.014599012210965157, + 0.026563916355371475, + 0.005521327722817659, + 0.006023662630468607, + 0.013661116361618042, + 0.006093128118664026, + 0.011194705963134766, + 0.01791168935596943, + 0.008594261482357979, + 0.012396351434290409, + 0.009719413705170155, + 0.010627402924001217, + 0.025363853201270103, + 0.007768332492560148, + 0.011217410676181316, + 0.03517724573612213, + 0.01375353429466486, + 0.01659015566110611, + 0.015124921686947346, + 0.017472542822360992, + 0.026479387655854225, + 0.01058091688901186, + 0.016607273370027542, + 0.03993448242545128, + 0.012090129777789116, + 0.01813698001205921, + 0.010562666691839695, + 0.00948543194681406, + 0.022944729775190353, + 0.010024833492934704, + 0.01318004634231329, + 0.029799504205584526, + 0.015032816678285599, + 0.020800350233912468, + 0.002380107529461384, + 0.003883777419105172, + 0.004087307956069708, + 0.0022111409343779087, + 0.0032706288620829582, + 0.005093185231089592, + 0.004325035959482193, + 0.0037339376285672188, + 0.001923597534187138, + 0.0025546159595251083, + 0.005333892535418272, + 0.0023327944800257683, + 0.00214388151653111, + 0.0052232989110052586, + 0.00464477576315403, + 0.007345301099121571, + 0.011551064439117908, + 0.009146634489297867, + 0.024056939408183098, + 0.00945220235735178, + 0.01080822292715311, + 0.01837955042719841, + 0.020822348073124886, + 0.04603681340813637, + 0.0029357450548559427, + 0.003497200785204768, + 0.0053112138994038105, + 0.0031730926129966974, + 0.003952872008085251, + 0.007007199339568615, + 0.005571605637669563, + 0.007909027859568596, + 0.0025753609370440245, + 0.004329037852585316, + 0.0029066232964396477, + 0.0023563506547361612, + 0.004865049850195646, + 0.003876057453453541, + 0.0029633583035320044, + 0.004350696690380573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.0028217521030455828 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium shot captures a standard one-gallon metal paint can, its exterior a vibrant primary red, precariously balanced on the cracked, dusty concrete floor of a garage. Bright, direct daylight streams in from a high, off-camera source, casting sharp shadows and gleaming across the can's semi-glossy, slightly worn surface. The unsealed metal lid tilts askew on its rim, a clear sign of its looseness. Pulled by gravity, the can begins a slow, deliberate, and irreversible tilt, its motion accelerating as it loses its balance and tips over. It strikes the gravel with a sharp, metallic thud, the impact instantly forcing the loose lid to spring free and fly aside. Simultaneously, a glob of viscous red paint violently erupts from the hole in a dramatic arcing wave, breaking into small globules and tiny droplets in mid-air and splattering a chaotic, irregular, sheen of crimson on the cold gray floor. This contrasts sharply with the dull, stained concrete, creating a scene of sudden, chaotic disaster.", + "seed": 24901, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_15/index_5189_seed_24901_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.037983722332865, + "vlm_logprob": -0.037983722332865, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.037983722332865, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07670552283525467 + ], + "opposite_logprob_tokens": [ + -2.7017054557800293 + ], + "logprob_reward": -0.07670552283525467, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07670552283525467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0753696858882904, + 0.08731665462255478, + 0.31215423345565796, + 0.15447668731212616, + 0.0734555646777153, + 0.0693826824426651, + 0.050807513296604156, + 0.07760514318943024, + 0.052683450281620026, + 0.046748336404561996 + ], + "credit_map": [ + 0.012405418790876865, + 0.01017376221716404, + 0.008341566659510136, + 0.007588053587824106, + 0.0100698946043849, + 0.009172456339001656, + 0.010135799646377563, + 0.00748274615034461, + 0.012980504892766476, + 0.008746101520955563, + 0.008137902244925499, + 0.010509553365409374, + 0.012378685176372528, + 0.010855020955204964, + 0.012208683416247368, + 0.01150020956993103, + 0.03189203888177872, + 0.029964175075292587, + 0.03491036593914032, + 0.039771899580955505, + 0.06241091713309288, + 0.02278384566307068, + 0.06643787771463394, + 0.023983124643564224, + 0.019201140850782394, + 0.013147957623004913, + 0.018898161128163338, + 0.016754280775785446, + 0.03669920563697815, + 0.015801692381501198, + 0.01691535674035549, + 0.01705891266465187, + 0.009741337038576603, + 0.0056836255826056, + 0.011476137675344944, + 0.005821527447551489, + 0.01609334722161293, + 0.01049529854208231, + 0.00929693877696991, + 0.0048473551869392395, + 0.007566084619611502, + 0.005987382493913174, + 0.013063306920230389, + 0.008558208122849464, + 0.012664208188652992, + 0.007983383722603321, + 0.00918150320649147, + 0.004378614481538534, + 0.006891003809869289, + 0.004038996063172817, + 0.00563449552282691, + 0.005831625312566757, + 0.007099959999322891, + 0.009758544154465199, + 0.007385916542261839, + 0.00416697608307004, + 0.010916572995483875, + 0.006426256615668535, + 0.008356775157153606, + 0.0079942112788558, + 0.019400427117943764, + 0.012054886668920517, + 0.005914892069995403, + 0.006541134789586067, + 0.006864093244075775, + 0.0060858759097754955, + 0.009454251267015934, + 0.004876311868429184, + 0.005761613603681326, + 0.006917121820151806, + 0.006726615596562624, + 0.00599757069721818, + 0.004457532428205013, + 0.004092068877071142, + 0.005098010879009962, + 0.003480812069028616, + 0.005662457551807165, + 0.0090134022757411, + 0.0072433180175721645, + 0.00770074175670743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.07670552283525467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.079041987657547 + ], + "opposite_logprob_tokens": [ + -2.5790419578552246 + ], + "logprob_reward": -0.079041987657547, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.079041987657547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13603462278842926, + 0.07674926519393921, + 0.11406643688678741, + 0.11128479242324829, + 0.11066055297851562, + 0.05971997603774071, + 0.13414373993873596, + 0.050311461091041565, + 0.02294306270778179, + 0.1840861290693283 + ], + "credit_map": [ + 0.025838909670710564, + 0.016727624461054802, + 0.01947823539376259, + 0.012683157809078693, + 0.014493010006844997, + 0.018968479707837105, + 0.016882263123989105, + 0.010962922126054764, + 0.010594664141535759, + 0.009001828730106354, + 0.006877689156681299, + 0.009266712702810764, + 0.010092487558722496, + 0.012323757633566856, + 0.010667943395674229, + 0.007924176752567291, + 0.015482548624277115, + 0.014303184114396572, + 0.008850540034472942, + 0.013951743021607399, + 0.020310811698436737, + 0.008411304093897343, + 0.022797588258981705, + 0.009958704933524132, + 0.012750894762575626, + 0.01088274922221899, + 0.011484191752970219, + 0.013589839451014996, + 0.0259693693369627, + 0.008163108490407467, + 0.014377451501786709, + 0.014067175798118114, + 0.014192547649145126, + 0.006660095416009426, + 0.030421219766139984, + 0.014216356910765171, + 0.014439016580581665, + 0.014565633609890938, + 0.007814491167664528, + 0.008351175114512444, + 0.0069468338042497635, + 0.0040060547180473804, + 0.0183776393532753, + 0.011197625659406185, + 0.004840029403567314, + 0.005455126520246267, + 0.004161186050623655, + 0.004735474940389395, + 0.01916046254336834, + 0.008371814154088497, + 0.033218804746866226, + 0.012075374834239483, + 0.013895023614168167, + 0.016990579664707184, + 0.014523105695843697, + 0.015908565372228622, + 0.006854419596493244, + 0.004359607584774494, + 0.006475359667092562, + 0.00662481551989913, + 0.006018882151693106, + 0.010784235782921314, + 0.0038947921711951494, + 0.005299343727529049, + 0.003150129457935691, + 0.0030266030225902796, + 0.0027454050723463297, + 0.0026258749421685934, + 0.0026039716321974993, + 0.0027758197393268347, + 0.0035626988392323256, + 0.002452556975185871, + 0.014252908527851105, + 0.015217163600027561, + 0.02311805821955204, + 0.01616155356168747, + 0.02319735661149025, + 0.032671574503183365, + 0.034790582954883575, + 0.02467690408229828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.079041987657547 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011093251407146454 + ], + "opposite_logprob_tokens": [ + -4.5110931396484375 + ], + "logprob_reward": -0.011093251407146454, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011093251407146454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07812897115945816, + 0.05766464024782181, + 0.2445223331451416, + 0.055393971502780914, + 0.020400196313858032, + 0.28924596309661865, + 0.061737317591905594, + 0.11266661435365677, + 0.02903345227241516, + 0.051206521689891815 + ], + "credit_map": [ + 0.013495148159563541, + 0.012640289030969143, + 0.00891453679651022, + 0.008444187231361866, + 0.012277322821319103, + 0.008862216025590897, + 0.006565849296748638, + 0.006929413415491581, + 0.011358274146914482, + 0.004629680421203375, + 0.005938241723924875, + 0.005358703434467316, + 0.006594352889806032, + 0.010050917975604534, + 0.007082019466906786, + 0.00665245158597827, + 0.03498515859246254, + 0.028846634551882744, + 0.02514992468059063, + 0.02631474658846855, + 0.047372572124004364, + 0.02046847715973854, + 0.04070926085114479, + 0.02067556045949459, + 0.00791038479655981, + 0.005081808660179377, + 0.007027016021311283, + 0.006827567704021931, + 0.012789339758455753, + 0.004666546825319529, + 0.00521639920771122, + 0.005874906200915575, + 0.002913652453571558, + 0.0009945239871740341, + 0.003434200771152973, + 0.0015271565644070506, + 0.004154758993536234, + 0.003124220995232463, + 0.0018342325929552317, + 0.002417451236397028, + 0.030594602227211, + 0.00737884733825922, + 0.11661534011363983, + 0.06755593419075012, + 0.013830171898007393, + 0.03084726072847843, + 0.012137148529291153, + 0.010286668315529823, + 0.012385041452944279, + 0.006387574132531881, + 0.00948770809918642, + 0.005575709510594606, + 0.008081736043095589, + 0.00739716412499547, + 0.0077699352987110615, + 0.0046524484641849995, + 0.01741540990769863, + 0.008937346749007702, + 0.015090425498783588, + 0.012496382929384708, + 0.022784139961004257, + 0.01743161678314209, + 0.008689146488904953, + 0.009822146967053413, + 0.0026260679587721825, + 0.002535480074584484, + 0.003579898504540324, + 0.0027983877807855606, + 0.0038472230080515146, + 0.004418159369379282, + 0.004479278344660997, + 0.0047489553689956665, + 0.004360878374427557, + 0.004979744087904692, + 0.004112802445888519, + 0.0038685209583491087, + 0.006962564308196306, + 0.010782136581838131, + 0.006860601715743542, + 0.009279274381697178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.011093251407146454 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020587095990777016 + ], + "opposite_logprob_tokens": [ + -3.895587205886841 + ], + "logprob_reward": -0.020587095990777016, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020587095990777016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06622103601694107, + 0.059475935995578766, + 0.07478626817464828, + 0.097620889544487, + 0.023575464263558388, + 0.05350487306714058, + 0.16346265375614166, + 0.15730951726436615, + 0.06349793076515198, + 0.24054546654224396 + ], + "credit_map": [ + 0.009762260131537914, + 0.007379522547125816, + 0.007713282480835915, + 0.008992047049105167, + 0.009676915593445301, + 0.008533326908946037, + 0.008942864835262299, + 0.00522081321105361, + 0.008933000266551971, + 0.007386223413050175, + 0.0043460046872496605, + 0.007294060196727514, + 0.00964452140033245, + 0.0076292273588478565, + 0.008934611454606056, + 0.005308292806148529, + 0.010103614069521427, + 0.00934764463454485, + 0.0054467422887682915, + 0.007301538251340389, + 0.016866475343704224, + 0.006342582870274782, + 0.013272716663777828, + 0.006104954052716494, + 0.011782911606132984, + 0.008555357344448566, + 0.011708301492035389, + 0.011937296949326992, + 0.02532357908785343, + 0.006696284282952547, + 0.010849187150597572, + 0.010767980478703976, + 0.002952947048470378, + 0.001955784158781171, + 0.0027054864913225174, + 0.0016561818774789572, + 0.00654032826423645, + 0.002484272699803114, + 0.0035030462313443422, + 0.0017774158623069525, + 0.006848382297903299, + 0.003687870455905795, + 0.005273981019854546, + 0.0035831844434142113, + 0.013051483780145645, + 0.00955547858029604, + 0.007632364518940449, + 0.0038721291348338127, + 0.025469385087490082, + 0.014302128925919533, + 0.01069254893809557, + 0.015804436057806015, + 0.018538685515522957, + 0.027970310300588608, + 0.02646338753402233, + 0.024221772328019142, + 0.02346055395901203, + 0.017674103379249573, + 0.023002130910754204, + 0.021495401859283447, + 0.020878991112113, + 0.024648670107126236, + 0.011587530374526978, + 0.014562140218913555, + 0.0048307618126273155, + 0.0038049037102609873, + 0.011044543236494064, + 0.005783370696008205, + 0.00801843497902155, + 0.009000163525342941, + 0.008504259400069714, + 0.012511498294770718, + 0.01613660901784897, + 0.024554206058382988, + 0.01112387701869011, + 0.015984661877155304, + 0.03172362223267555, + 0.047051429748535156, + 0.04244665428996086, + 0.05152441933751106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.020587095990777016 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002490753773599863 + ], + "opposite_logprob_tokens": [ + -6.002490520477295 + ], + "logprob_reward": -0.002490753773599863, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002490753773599863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11945219337940216, + 0.05591016635298729, + 0.11712931841611862, + 0.05720675364136696, + 0.12342827767133713, + 0.09510401636362076, + 0.048784129321575165, + 0.04770943894982338, + 0.019415516406297684, + 0.31586024165153503 + ], + "credit_map": [ + 0.017076678574085236, + 0.011320183984935284, + 0.011521130800247192, + 0.01492619514465332, + 0.01817016489803791, + 0.021150439977645874, + 0.01498029287904501, + 0.010307095013558865, + 0.008346155285835266, + 0.005896198097616434, + 0.004572889767587185, + 0.006124501582235098, + 0.008791115134954453, + 0.008193379268050194, + 0.0074412706308066845, + 0.006544653791934252, + 0.016285603865981102, + 0.015564468689262867, + 0.008145376108586788, + 0.012792016379535198, + 0.025449369102716446, + 0.009772570803761482, + 0.020608754828572273, + 0.008511158637702465, + 0.008564003743231297, + 0.004525757860392332, + 0.007320116739720106, + 0.007657674141228199, + 0.010510982945561409, + 0.006921406835317612, + 0.0055925073102116585, + 0.006114299409091473, + 0.01942109875380993, + 0.008222872391343117, + 0.027840808033943176, + 0.012057526968419552, + 0.022075185552239418, + 0.013961915858089924, + 0.009610606357455254, + 0.01023826003074646, + 0.011573183350265026, + 0.004101360682398081, + 0.032603077590465546, + 0.018643692135810852, + 0.00924978218972683, + 0.008368803188204765, + 0.006267412565648556, + 0.004296703729778528, + 0.008428891189396381, + 0.003204154083505273, + 0.00945412926375866, + 0.0038351998664438725, + 0.006595228333026171, + 0.007028704974800348, + 0.006320522166788578, + 0.003917297348380089, + 0.007650997024029493, + 0.00426509790122509, + 0.006832635495811701, + 0.006011645775288343, + 0.007451863028109074, + 0.007772913668304682, + 0.0030436019878834486, + 0.00468067592009902, + 0.0023886943235993385, + 0.001565398066304624, + 0.0029303135816007853, + 0.002175592817366123, + 0.00223514880053699, + 0.002719955053180456, + 0.002797677181661129, + 0.0026027357671409845, + 0.02671803906559944, + 0.024143552407622337, + 0.0352468304336071, + 0.028665924444794655, + 0.045150112360715866, + 0.06165801361203194, + 0.04780784621834755, + 0.04646989703178406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.002490753773599863 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium shot captures a standard one-gallon metal paint can, its exterior a vibrant primary red, precariously balanced on the cracked, dusty concrete floor of a garage. Bright, direct daylight streams in from a high, off-camera source, casting sharp shadows and gleaming across the can's semi-glossy, slightly worn surface. The unsealed metal lid tilts askew on its rim, a clear sign of its looseness. Pulled by gravity, the can begins a slow, deliberate, and irreversible tilt, its motion accelerating as it loses its balance and tips over. It strikes the gravel with a sharp, metallic thud, the impact instantly forcing the loose lid to spring free and fly aside. Simultaneously, a glob of viscous red paint violently erupts from the hole in a dramatic arcing wave, breaking into small globules and tiny droplets in mid-air and splattering a chaotic, irregular, sheen of crimson on the cold gray floor. This contrasts sharply with the dull, stained concrete, creating a scene of sudden, chaotic disaster.", + "seed": 24902, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_15/index_5189_seed_24902_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.016485346574336292, + "vlm_logprob": -0.016485346574336292, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.016485346574336292, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03617598116397858 + ], + "opposite_logprob_tokens": [ + -3.6611759662628174 + ], + "logprob_reward": -0.03617598116397858, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03617598116397858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16195034980773926, + 0.09431798756122589, + 0.06473402678966522, + 0.025460554286837578, + 0.041530199348926544, + 0.046962521970272064, + 0.07426934689283371, + 0.3616015315055847, + 0.06474414467811584, + 0.06442936509847641 + ], + "credit_map": [ + 0.025445668026804924, + 0.02699880301952362, + 0.02789069339632988, + 0.01797090284526348, + 0.014438308775424957, + 0.012932667508721352, + 0.013188292272388935, + 0.023084988817572594, + 0.014408678747713566, + 0.013565176166594028, + 0.015081726014614105, + 0.010104814544320107, + 0.009533400647342205, + 0.009650172665715218, + 0.008875081315636635, + 0.013098926283419132, + 0.010636450722813606, + 0.009354581125080585, + 0.01107334066182375, + 0.006100465077906847, + 0.005208081565797329, + 0.009005057625472546, + 0.005411280319094658, + 0.007944762706756592, + 0.004208789207041264, + 0.0025755444075912237, + 0.004794783890247345, + 0.0026624512393027544, + 0.001753521617501974, + 0.002613137010484934, + 0.0036275607999414206, + 0.003224764484912157, + 0.005779686383903027, + 0.005108359269797802, + 0.004745992831885815, + 0.006015920080244541, + 0.0038779783062636852, + 0.00378790614195168, + 0.005966838914901018, + 0.006247516255825758, + 0.006301697809249163, + 0.004109945613890886, + 0.007980749011039734, + 0.004594224970787764, + 0.004570942837744951, + 0.007078773342072964, + 0.005947969388216734, + 0.0063782124780118465, + 0.012422313913702965, + 0.007932677865028381, + 0.008495086804032326, + 0.0141141377389431, + 0.0066625322215259075, + 0.007281854748725891, + 0.007356461137533188, + 0.010004275478422642, + 0.05091216042637825, + 0.0218960028141737, + 0.04262766242027283, + 0.06775586307048798, + 0.04983613267540932, + 0.03452789783477783, + 0.045949824154376984, + 0.048095960170030594, + 0.01067399326711893, + 0.002932862611487508, + 0.005640482529997826, + 0.008154830895364285, + 0.016394706442952156, + 0.01065274141728878, + 0.006443288642913103, + 0.0038512307219207287, + 0.007050164043903351, + 0.00925994012504816, + 0.0076963044703006744, + 0.005868860054761171, + 0.010478503070771694, + 0.013977165333926678, + 0.0053319125436246395, + 0.004766508005559444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.03617598116397858 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014214222319424152 + ], + "opposite_logprob_tokens": [ + -4.264214038848877 + ], + "logprob_reward": -0.014214222319424152, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014214222319424152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19764187932014465, + 0.06747347116470337, + 0.035288870334625244, + 0.044218726456165314, + 0.09853435307741165, + 0.05443467199802399, + 0.030587829649448395, + 0.05842622369527817, + 0.34488439559936523, + 0.06850960850715637 + ], + "credit_map": [ + 0.027719365432858467, + 0.03039846196770668, + 0.04138747602701187, + 0.02173047512769699, + 0.018934262916445732, + 0.01617366448044777, + 0.015194824896752834, + 0.02610335685312748, + 0.010615760460495949, + 0.00770061369985342, + 0.01324592437595129, + 0.008417192846536636, + 0.006000801920890808, + 0.005620419979095459, + 0.007417896296828985, + 0.008454862982034683, + 0.005867689847946167, + 0.004074209835380316, + 0.008198262192308903, + 0.004250702913850546, + 0.0022726685274392366, + 0.0031551446299999952, + 0.0032615310046821833, + 0.004208660684525967, + 0.006742551922798157, + 0.0059498995542526245, + 0.009546197019517422, + 0.004675315693020821, + 0.002894588513299823, + 0.0037073874846100807, + 0.0044213347136974335, + 0.006281448993831873, + 0.01493843737989664, + 0.013700959272682667, + 0.014468303881585598, + 0.011488599702715874, + 0.006480671931058168, + 0.014932856895029545, + 0.011997104622423649, + 0.010527417063713074, + 0.00872683897614479, + 0.008949271403253078, + 0.011736128479242325, + 0.006092030089348555, + 0.004340400453656912, + 0.004961874336004257, + 0.005383170209825039, + 0.004244954325258732, + 0.004780993796885014, + 0.0037239755038172007, + 0.005434483289718628, + 0.004020254593342543, + 0.0026582356076687574, + 0.003426044713705778, + 0.0021733837202191353, + 0.004370460286736488, + 0.007756717503070831, + 0.005921050440520048, + 0.00769555801525712, + 0.006947389338165522, + 0.005360460840165615, + 0.00675733108073473, + 0.007379909977316856, + 0.010607806965708733, + 0.071573905646801, + 0.025187605991959572, + 0.03326418250799179, + 0.04143213853240013, + 0.048328615725040436, + 0.047644197940826416, + 0.05206490308046341, + 0.025388838723301888, + 0.009668209590017796, + 0.0037315546069294214, + 0.005317972972989082, + 0.0033636644948273897, + 0.02427889034152031, + 0.016447171568870544, + 0.0031299819238483906, + 0.0025721623096615076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.014214222319424152 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011096198111772537 + ], + "opposite_logprob_tokens": [ + -4.511096000671387 + ], + "logprob_reward": -0.011096198111772537, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011096198111772537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24780628085136414, + 0.05734606087207794, + 0.03828537091612816, + 0.04700784385204315, + 0.043868180364370346, + 0.039728276431560516, + 0.09546245634555817, + 0.15929566323757172, + 0.12727542221546173, + 0.14392441511154175 + ], + "credit_map": [ + 0.03605978563427925, + 0.0328289195895195, + 0.04470515623688698, + 0.03485744819045067, + 0.016796311363577843, + 0.028881195932626724, + 0.022719381377100945, + 0.03095809370279312, + 0.00961865670979023, + 0.006294491235166788, + 0.0098903588950634, + 0.005725300405174494, + 0.004864639602601528, + 0.006158508360385895, + 0.006386787164956331, + 0.008407319895923138, + 0.0066803316585719585, + 0.004543829709291458, + 0.005310808774083853, + 0.004201212432235479, + 0.003211884992197156, + 0.005092937499284744, + 0.0034575338941067457, + 0.005786836612969637, + 0.007260663900524378, + 0.004535048268735409, + 0.008419817313551903, + 0.005379222333431244, + 0.0030507436022162437, + 0.00412746099755168, + 0.006923213135451078, + 0.007311674766242504, + 0.005783322267234325, + 0.004053853917866945, + 0.006385880522429943, + 0.006467237137258053, + 0.0033551338128745556, + 0.004658191930502653, + 0.006899199448525906, + 0.006265361327677965, + 0.005759352818131447, + 0.003065132535994053, + 0.005056566093116999, + 0.004354699049144983, + 0.0037938342429697514, + 0.006838089320808649, + 0.004545736592262983, + 0.006314865313470364, + 0.012720308266580105, + 0.008765085600316525, + 0.010415269061923027, + 0.012677700258791447, + 0.0068106637336313725, + 0.0158150065690279, + 0.008944512344896793, + 0.019313909113407135, + 0.021511347964406013, + 0.015563329681754112, + 0.015511682257056236, + 0.02352747693657875, + 0.017064087092876434, + 0.01402838435024023, + 0.021926432847976685, + 0.030162926763296127, + 0.027963561937212944, + 0.01041357684880495, + 0.012660366483032703, + 0.018753046169877052, + 0.019733553752303123, + 0.016821378841996193, + 0.013492459431290627, + 0.007437479682266712, + 0.03403737023472786, + 0.012224195525050163, + 0.01731177419424057, + 0.008227134123444557, + 0.040036123245954514, + 0.019625408574938774, + 0.00654991902410984, + 0.005912481341511011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.011096198111772537 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200471341609955 + ], + "opposite_logprob_tokens": [ + -4.264200687408447 + ], + "logprob_reward": -0.014200471341609955, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200471341609955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11218246072530746, + 0.04530740901827812, + 0.03584209084510803, + 0.08250674605369568, + 0.051123324781656265, + 0.06342608481645584, + 0.1524541974067688, + 0.12239692360162735, + 0.21248210966587067, + 0.12227856367826462 + ], + "credit_map": [ + 0.01759885624051094, + 0.012379174120724201, + 0.01825433038175106, + 0.0142739899456501, + 0.009095166809856892, + 0.015394581481814384, + 0.010755038820207119, + 0.014431326650083065, + 0.007154461462050676, + 0.0055411397479474545, + 0.006206085439771414, + 0.005051193293184042, + 0.003553955350071192, + 0.004701907746493816, + 0.005582235753536224, + 0.007516433019191027, + 0.005341814365237951, + 0.004095038399100304, + 0.006161272991448641, + 0.004001035820692778, + 0.0029043175745755434, + 0.003925035707652569, + 0.004115600138902664, + 0.00529797887429595, + 0.010164797306060791, + 0.00929681584239006, + 0.00914611667394638, + 0.01056645531207323, + 0.007002468686550856, + 0.008824503049254417, + 0.013027212582528591, + 0.014478379860520363, + 0.0065683345310389996, + 0.0038970613386482, + 0.005192291457206011, + 0.00901615247130394, + 0.005120418034493923, + 0.005242838989943266, + 0.008355884812772274, + 0.007730345707386732, + 0.008248502388596535, + 0.005603582598268986, + 0.008022579364478588, + 0.0097508504986763, + 0.005874014459550381, + 0.009227870032191277, + 0.00900595635175705, + 0.007692735642194748, + 0.022473953664302826, + 0.01573529653251171, + 0.017290612682700157, + 0.03625831380486488, + 0.012126789428293705, + 0.015402081422507763, + 0.01652160845696926, + 0.01664554327726364, + 0.019198931753635406, + 0.008667726069688797, + 0.014886414632201195, + 0.012933189980685711, + 0.015134238637983799, + 0.0141828004270792, + 0.01780281774699688, + 0.019590813666582108, + 0.04686027765274048, + 0.014709999784827232, + 0.020927224308252335, + 0.023738108575344086, + 0.028339242562651634, + 0.03284192457795143, + 0.029808521270751953, + 0.015256829559803009, + 0.02687540277838707, + 0.00540558248758316, + 0.008269825018942356, + 0.008403894491493702, + 0.038181666284799576, + 0.014594743959605694, + 0.012722894549369812, + 0.007824568077921867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.014200471341609955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a standard one-gallon metal paint can with a vibrant primary red exterior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a garage with a cracked, dusty concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a low angle capturing a medium shot of the can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and fall over, causing the loose lid to fly off and red paint to erupt in a chaotic splatter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and direct, casting sharp shadows and creating a gleam on the semi-glossy surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739859934896231 + ], + "opposite_logprob_tokens": [ + -5.006740093231201 + ], + "logprob_reward": -0.006739859934896231, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739859934896231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2579059302806854, + 0.05001060292124748, + 0.04084983840584755, + 0.040173593908548355, + 0.04249846190214157, + 0.047458164393901825, + 0.03613311052322388, + 0.12807145714759827, + 0.3121056854724884, + 0.04479314386844635 + ], + "credit_map": [ + 0.03801329806447029, + 0.029372500255703926, + 0.047374870628118515, + 0.032178860157728195, + 0.01976529322564602, + 0.029200376942753792, + 0.023594271391630173, + 0.038406457751989365, + 0.008009118027985096, + 0.0055574108846485615, + 0.009265491738915443, + 0.00710959080606699, + 0.0037262719124555588, + 0.004350502043962479, + 0.0050859698094427586, + 0.006906250026077032, + 0.0064728884026408195, + 0.00609575817361474, + 0.008703866973519325, + 0.0039569903165102005, + 0.0022228388115763664, + 0.003967187833040953, + 0.003646373748779297, + 0.005783935077488422, + 0.0059228697791695595, + 0.004662378691136837, + 0.0067259883508086205, + 0.004724307917058468, + 0.0034968967083841562, + 0.004499753937125206, + 0.0047450498677790165, + 0.005396346561610699, + 0.005891118664294481, + 0.008217933587729931, + 0.006176636554300785, + 0.0038385423831641674, + 0.0034553029108792543, + 0.006669818889349699, + 0.003979858476668596, + 0.0042692553251981735, + 0.008065136149525642, + 0.0065019805915653706, + 0.010631185956299305, + 0.004371593240648508, + 0.0033438962418586016, + 0.004268843214958906, + 0.00486530177295208, + 0.005410227458924055, + 0.005337865091860294, + 0.004152288660407066, + 0.006339801009744406, + 0.005837744101881981, + 0.0026901697274297476, + 0.0037116319872438908, + 0.002553077880293131, + 0.005510532297194004, + 0.018413828685879707, + 0.009125467389822006, + 0.01122951041907072, + 0.022177495062351227, + 0.011549107730388641, + 0.010663463734090328, + 0.017572075128555298, + 0.027340512722730637, + 0.07168827205896378, + 0.02034773863852024, + 0.028131892904639244, + 0.038523025810718536, + 0.05834891274571419, + 0.03244251385331154, + 0.043200988322496414, + 0.01942233182489872, + 0.005440725013613701, + 0.00352187710814178, + 0.0046950168907642365, + 0.004925879649817944, + 0.008226661942899227, + 0.010489213280379772, + 0.004609452094882727, + 0.002884316723793745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.006739859934896231 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_2/index_tensor([4793, 4793])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_2/index_tensor([4793, 4793])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..49085ebd669a068a576a9420130be22067543dbf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_2/index_tensor([4793, 4793])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic tracking shot, shot from a low angle, closely follows the young Nezha as he navigates the turbulent deep blue sea. He wears his signature twin buns, a few unruly strands of black hair dancing in the wind, his face brimming with a rebellious yet playful smile. He wears a bright red bellyband intricately embroidered with golden lotus patterns. His bare arms are wrapped with the Hun Tian Ling, the silk ribbons rustling in the fierce wind like two living red serpents. His feet rest firmly on a pair of fiery wind-fire wheels. The wheels spin at high speed, leaving long trails of golden flame, instantly vaporizing any seawater they touch into hissing white steam. With a near-dance, he deftly glides and leaps over the crests of the mountainous waves, leaving two vast white wakes in his wake, churning the water behind him. The water, stirred by the wind-fire wheels, quickly converges into a single, towering wave. The sky is thick with dark clouds, punctuated by thunder and lightning. Occasionally, wisps of divine golden light peek through the cracks in the clouds, interplaying beautifully with the fiery glow of the Wind Fire Wheel and the pale lightning streaking across the sky. Light and shadow playfully shift across the inky sea. The dominant tones of the painting are deep, cool blues and grays, ignited by the vibrant red of Nezha's figure and the fiery golden orange of the Wind Fire Wheel. This creates an atmosphere of epic mythology, tension, and power, showcasing the powerful aura of the young deity, defying heaven and earth and possessing unbridled freedom.", + "seed": 25720, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_2/index_4793_seed_25720_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0160495444200933, + "vlm_logprob": -0.0160495444200933, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.0160495444200933, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05691896751523018 + ], + "opposite_logprob_tokens": [ + -2.9319188594818115 + ], + "logprob_reward": -0.05691896751523018, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05691896751523018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16879430413246155, + 0.14418675005435944, + 0.09307800978422165, + 0.2297951579093933, + 0.08043540269136429, + 0.08665783703327179, + 0.029449084773659706, + 0.04116907715797424, + 0.056007176637649536, + 0.07042718678712845 + ], + "credit_map": [ + 0.0201435424387455, + 0.02305821143090725, + 0.01760479435324669, + 0.03517243266105652, + 0.014480456709861755, + 0.015243520960211754, + 0.019718807190656662, + 0.023372536525130272, + 0.013678203336894512, + 0.016940176486968994, + 0.02340560220181942, + 0.029114525765180588, + 0.02534017711877823, + 0.014975045807659626, + 0.01155502162873745, + 0.009178007021546364, + 0.008152149617671967, + 0.01210206001996994, + 0.011703026480972767, + 0.009761037304997444, + 0.019108662381768227, + 0.013429611921310425, + 0.00744377076625824, + 0.01137769129127264, + 0.020812049508094788, + 0.023645957931876183, + 0.017067890614271164, + 0.04106849059462547, + 0.027156613767147064, + 0.05373017117381096, + 0.028239741921424866, + 0.01807422935962677, + 0.00804007425904274, + 0.011669003404676914, + 0.008450775407254696, + 0.0110568106174469, + 0.01168965082615614, + 0.014063837006688118, + 0.008289369754493237, + 0.007175881415605545, + 0.008508698083460331, + 0.012617865577340126, + 0.011042903177440166, + 0.011519551277160645, + 0.01788192056119442, + 0.010101412422955036, + 0.007120402995496988, + 0.007865075953304768, + 0.004609753377735615, + 0.004391261842101812, + 0.004314477555453777, + 0.002749999286606908, + 0.002971402369439602, + 0.0031535408925265074, + 0.002593869809061289, + 0.00466478057205677, + 0.0041964068077504635, + 0.005243358667939901, + 0.005714739207178354, + 0.004385451786220074, + 0.004532297607511282, + 0.0035370655823498964, + 0.00471346452832222, + 0.008846296928822994, + 0.013017267920076847, + 0.009710944257676601, + 0.00574528519064188, + 0.005104315001517534, + 0.006069594528526068, + 0.00558937294408679, + 0.007225947920233011, + 0.003544447012245655, + 0.008029009215533733, + 0.00982199888676405, + 0.012284521944820881, + 0.006500594783574343, + 0.011855274438858032, + 0.005568711087107658, + 0.009972892701625824, + 0.0063941883854568005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.05691896751523018 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008656476624310017 + ], + "opposite_logprob_tokens": [ + -4.7586565017700195 + ], + "logprob_reward": -0.008656476624310017, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008656476624310017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18470606207847595, + 0.11008388549089432, + 0.07637733221054077, + 0.1414041370153427, + 0.1520869880914688, + 0.12296588718891144, + 0.03757947310805321, + 0.04392051324248314, + 0.04210372641682625, + 0.08877196907997131 + ], + "credit_map": [ + 0.024518845602869987, + 0.025964999571442604, + 0.018960634246468544, + 0.04144008830189705, + 0.015352128073573112, + 0.01828409917652607, + 0.020554140210151672, + 0.01963113807141781, + 0.014438308775424957, + 0.018566086888313293, + 0.00984889455139637, + 0.015457486733794212, + 0.017105795443058014, + 0.013403336517512798, + 0.010739820078015327, + 0.010524161159992218, + 0.008065021596848965, + 0.012194378301501274, + 0.009423266164958477, + 0.009234634228050709, + 0.01255898829549551, + 0.0112282270565629, + 0.006932688876986504, + 0.0067401365377008915, + 0.01165748480707407, + 0.01701958477497101, + 0.0167181845754385, + 0.022105878219008446, + 0.020276514813303947, + 0.013939138501882553, + 0.024203941226005554, + 0.015483424067497253, + 0.014897473156452179, + 0.02051580883562565, + 0.01736675389111042, + 0.02154773660004139, + 0.028246760368347168, + 0.02090425416827202, + 0.013329935260117054, + 0.015278268605470657, + 0.011217066086828709, + 0.019000152125954628, + 0.014491244219243526, + 0.019818540662527084, + 0.023401064798235893, + 0.015343617647886276, + 0.009184148162603378, + 0.010510062798857689, + 0.005832870956510305, + 0.005344945006072521, + 0.0052774264477193356, + 0.0038017993792891502, + 0.004450723994523287, + 0.003824838437139988, + 0.0038520514499396086, + 0.0051948209293186665, + 0.004492742009460926, + 0.006196914240717888, + 0.00544839259237051, + 0.005582195706665516, + 0.005568147171288729, + 0.004166342783719301, + 0.006134960800409317, + 0.006330822594463825, + 0.007644021883606911, + 0.005838347598910332, + 0.004744050092995167, + 0.005629681516438723, + 0.004886955954134464, + 0.003820030717179179, + 0.005879636388272047, + 0.0036610064562410116, + 0.010890310630202293, + 0.01119440421462059, + 0.019373120740056038, + 0.01008003018796444, + 0.01146960724145174, + 0.007266346365213394, + 0.011328605934977531, + 0.007169547490775585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.008656476624310017 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097756464034319 + ], + "opposite_logprob_tokens": [ + -5.504097938537598 + ], + "logprob_reward": -0.004097756464034319, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097756464034319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1046755313873291, + 0.06709983199834824, + 0.10853520035743713, + 0.11568941920995712, + 0.13578030467033386, + 0.0879243016242981, + 0.03922640532255173, + 0.10570872575044632, + 0.13839666545391083, + 0.09696365892887115 + ], + "credit_map": [ + 0.013331430032849312, + 0.015560810454189777, + 0.013591337017714977, + 0.018913105130195618, + 0.00785849243402481, + 0.012051953934133053, + 0.014514950104057789, + 0.008853459730744362, + 0.007968076504766941, + 0.009695188142359257, + 0.007355640642344952, + 0.013791942037642002, + 0.007578129414469004, + 0.008072884753346443, + 0.007395194843411446, + 0.0052427733317017555, + 0.012214780785143375, + 0.01491209864616394, + 0.015899674966931343, + 0.012057111598551273, + 0.017715327441692352, + 0.014575906097888947, + 0.010358977131545544, + 0.010801323689520359, + 0.009970160201191902, + 0.013800944201648235, + 0.012892043218016624, + 0.016582684591412544, + 0.015087553299963474, + 0.014527271501719952, + 0.021824171766638756, + 0.01100458949804306, + 0.013454487547278404, + 0.02037653513252735, + 0.018428495153784752, + 0.026555068790912628, + 0.017583569511771202, + 0.017396796494722366, + 0.010517453774809837, + 0.01146790198981762, + 0.00847509317100048, + 0.013774079270660877, + 0.013722582720220089, + 0.015738997608423233, + 0.013052064925432205, + 0.00928275752812624, + 0.006072329822927713, + 0.007806397974491119, + 0.006387471687048674, + 0.00563281262293458, + 0.003547796979546547, + 0.004770794417709112, + 0.00553495017811656, + 0.004021910019218922, + 0.005999104585498571, + 0.0033315648324787617, + 0.013861480168998241, + 0.02015872299671173, + 0.009092816151678562, + 0.014503208920359612, + 0.013087267987430096, + 0.01156385987997055, + 0.012740248814225197, + 0.01070111058652401, + 0.03423637896776199, + 0.02618563361465931, + 0.014170612208545208, + 0.010625816881656647, + 0.01599031686782837, + 0.013769334182143211, + 0.011827168054878712, + 0.011591409333050251, + 0.007740967441350222, + 0.009499650448560715, + 0.0191316120326519, + 0.011191985569894314, + 0.012644142843782902, + 0.008237195201218128, + 0.01309807039797306, + 0.015420041047036648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.004097756464034319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019386084750294685 + ], + "opposite_logprob_tokens": [ + -6.251938819885254 + ], + "logprob_reward": -0.0019386084750294685, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019386084750294685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11347389966249466, + 0.0581323504447937, + 0.0571674220263958, + 0.24115748703479767, + 0.07631499320268631, + 0.021198507398366928, + 0.019125232473015785, + 0.016687989234924316, + 0.23673298954963684, + 0.16000914573669434 + ], + "credit_map": [ + 0.01478737872093916, + 0.023907450959086418, + 0.008442619815468788, + 0.013969953171908855, + 0.013351097702980042, + 0.012955870479345322, + 0.01002354733645916, + 0.016035977751016617, + 0.007853589951992035, + 0.008092430420219898, + 0.008624181151390076, + 0.010185083374381065, + 0.0061441524885594845, + 0.007409060373902321, + 0.005506141576915979, + 0.004317714832723141, + 0.005772601813077927, + 0.008426270447671413, + 0.007256705313920975, + 0.0077386414632201195, + 0.008457764983177185, + 0.006950444541871548, + 0.005569737870246172, + 0.006995257921516895, + 0.024516118690371513, + 0.034167755395174026, + 0.025617197155952454, + 0.03677679970860481, + 0.035013746470212936, + 0.029864829033613205, + 0.03041386790573597, + 0.024787170812487602, + 0.007561127655208111, + 0.010829485021531582, + 0.009052949026226997, + 0.011421115137636662, + 0.012025229632854462, + 0.010526489466428757, + 0.0068413312546908855, + 0.008057264611124992, + 0.001879339455626905, + 0.0023640196304768324, + 0.003075579646974802, + 0.0034600310027599335, + 0.0035497983917593956, + 0.0024884785525500774, + 0.0015873690135776997, + 0.0027938915882259607, + 0.002794646890833974, + 0.0028281714767217636, + 0.0022622039541602135, + 0.0020053701009601355, + 0.002684270031750202, + 0.0019821650348603725, + 0.0017811293946579099, + 0.002787275006994605, + 0.002275919308885932, + 0.00228609680198133, + 0.0023134953808039427, + 0.001987324096262455, + 0.0019924738444387913, + 0.0014845140976831317, + 0.0019277186365798116, + 0.0024204470682889223, + 0.06478115916252136, + 0.045390937477350235, + 0.02573380619287491, + 0.017099760472774506, + 0.02333936281502247, + 0.02191874012351036, + 0.01720478944480419, + 0.021264424547553062, + 0.04190613701939583, + 0.027890654280781746, + 0.023400355130434036, + 0.017602087929844856, + 0.013955707661807537, + 0.012773457914590836, + 0.013521184213459492, + 0.008959565311670303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.0019386084750294685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635913021862507 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.008635913021862507, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635913021862507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19645176827907562, + 0.061780914664268494, + 0.10175246000289917, + 0.13018105924129486, + 0.13878093659877777, + 0.04194895178079605, + 0.05090845748782158, + 0.059731896966695786, + 0.13572664558887482, + 0.08273691684007645 + ], + "credit_map": [ + 0.022212469950318336, + 0.02147567830979824, + 0.02409624680876732, + 0.05401421710848808, + 0.01532960869371891, + 0.019358348101377487, + 0.021605461835861206, + 0.018359722569584846, + 0.006582821253687143, + 0.0073158652521669865, + 0.008877781219780445, + 0.010661857202649117, + 0.0068762581795454025, + 0.00659132469445467, + 0.006148264743387699, + 0.00872674211859703, + 0.01152825728058815, + 0.016354653984308243, + 0.012143757194280624, + 0.019477689638733864, + 0.010806554928421974, + 0.01281569991260767, + 0.010488323867321014, + 0.00813752319663763, + 0.011296824552118778, + 0.019681261852383614, + 0.015662703663110733, + 0.016684554517269135, + 0.018239270895719528, + 0.023010313510894775, + 0.011527751572430134, + 0.014078392647206783, + 0.012964820489287376, + 0.021060507744550705, + 0.015517101623117924, + 0.020302817225456238, + 0.02105732262134552, + 0.02362428791821003, + 0.01180994976311922, + 0.01244413387030363, + 0.0035998865496367216, + 0.005020662676542997, + 0.004745551384985447, + 0.005340054631233215, + 0.004766926169395447, + 0.005093814339488745, + 0.004477113019675016, + 0.008904940448701382, + 0.009004862979054451, + 0.007223024033010006, + 0.005221107974648476, + 0.005235577467828989, + 0.008025412447750568, + 0.006044312845915556, + 0.004966329783201218, + 0.005187826231122017, + 0.005839325487613678, + 0.007652753964066505, + 0.008220107294619083, + 0.007540946360677481, + 0.007716928608715534, + 0.005015370436012745, + 0.008314077742397785, + 0.009432386606931686, + 0.040672991424798965, + 0.023948755115270615, + 0.015610401518642902, + 0.009652998298406601, + 0.013324647210538387, + 0.01156278233975172, + 0.00918711256235838, + 0.011766958050429821, + 0.012554842978715897, + 0.01119089312851429, + 0.013585129752755165, + 0.007160475943237543, + 0.012888060882687569, + 0.006666665896773338, + 0.011150520294904709, + 0.007540327496826649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.008635913021862507 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, shot from a low angle, closely follows the young Nezha as he navigates the turbulent deep blue sea. He wears his signature twin buns, a few unruly strands of black hair dancing in the wind, his face brimming with a rebellious yet playful smile. He wears a bright red bellyband intricately embroidered with golden lotus patterns. His bare arms are wrapped with the Hun Tian Ling, the silk ribbons rustling in the fierce wind like two living red serpents. His feet rest firmly on a pair of fiery wind-fire wheels. The wheels spin at high speed, leaving long trails of golden flame, instantly vaporizing any seawater they touch into hissing white steam. With a near-dance, he deftly glides and leaps over the crests of the mountainous waves, leaving two vast white wakes in his wake, churning the water behind him. The water, stirred by the wind-fire wheels, quickly converges into a single, towering wave. The sky is thick with dark clouds, punctuated by thunder and lightning. Occasionally, wisps of divine golden light peek through the cracks in the clouds, interplaying beautifully with the fiery glow of the Wind Fire Wheel and the pale lightning streaking across the sky. Light and shadow playfully shift across the inky sea. The dominant tones of the painting are deep, cool blues and grays, ignited by the vibrant red of Nezha's figure and the fiery golden orange of the Wind Fire Wheel. This creates an atmosphere of epic mythology, tension, and power, showcasing the powerful aura of the young deity, defying heaven and earth and possessing unbridled freedom.", + "seed": 25721, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_2/index_4793_seed_25721_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01187181866262108, + "vlm_logprob": -0.01187181866262108, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.01187181866262108, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03524540737271309 + ], + "opposite_logprob_tokens": [ + -3.410245418548584 + ], + "logprob_reward": -0.03524540737271309, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03524540737271309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22604505717754364, + 0.046448614448308945, + 0.023715365678071976, + 0.06104681268334389, + 0.029182102531194687, + 0.0404304563999176, + 0.05812041461467743, + 0.10097183287143707, + 0.18217474222183228, + 0.23186466097831726 + ], + "credit_map": [ + 0.020084623247385025, + 0.04288389906287193, + 0.01378032099455595, + 0.016574343666434288, + 0.011760815046727657, + 0.013112140819430351, + 0.04119858518242836, + 0.06665030121803284, + 0.0022780357394367456, + 0.003332060994580388, + 0.0048189833760261536, + 0.006200338713824749, + 0.005501869134604931, + 0.008933009579777718, + 0.0073178596794605255, + 0.008066453039646149, + 0.0017148560145869851, + 0.0030242137145251036, + 0.002079558093100786, + 0.0057862973771989346, + 0.002762144897133112, + 0.0025136638432741165, + 0.003925912082195282, + 0.0019087155815213919, + 0.005922779906541109, + 0.007061322219669819, + 0.011297516524791718, + 0.008968967944383621, + 0.004903303924947977, + 0.0077126519754529, + 0.00861206091940403, + 0.006568206008523703, + 0.0020380623172968626, + 0.003445371752604842, + 0.0033584439661353827, + 0.005380081478506327, + 0.0044016288593411446, + 0.002297311555594206, + 0.004772758577018976, + 0.0034884409978985786, + 0.0026995970401912928, + 0.0049883099272847176, + 0.006114764604717493, + 0.005505337845534086, + 0.005315038841217756, + 0.0035195776727050543, + 0.007405054289847612, + 0.004882770124822855, + 0.007461049128323793, + 0.004650920629501343, + 0.008657137863337994, + 0.007888898253440857, + 0.006647849455475807, + 0.007795216049998999, + 0.007673468440771103, + 0.007345869205892086, + 0.007753643672913313, + 0.02448705956339836, + 0.018109465017914772, + 0.018348487094044685, + 0.008682136423885822, + 0.008235039189457893, + 0.007735821418464184, + 0.00762017210945487, + 0.012473003007471561, + 0.025199677795171738, + 0.011730175465345383, + 0.015006826259195805, + 0.014808948151767254, + 0.031031498685479164, + 0.025085728615522385, + 0.04683886468410492, + 0.026087405160069466, + 0.03542337566614151, + 0.030966201797127724, + 0.05860563740134239, + 0.015334057621657848, + 0.020398423075675964, + 0.02154870145022869, + 0.023500844836235046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.03524540737271309 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004111172165721655 + ], + "opposite_logprob_tokens": [ + -5.504111289978027 + ], + "logprob_reward": -0.004111172165721655, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004111172165721655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21493621170520782, + 0.09867647290229797, + 0.04030214250087738, + 0.06277361512184143, + 0.07109551131725311, + 0.08119044452905655, + 0.09057646244764328, + 0.12925253808498383, + 0.13028040528297424, + 0.08091617375612259 + ], + "credit_map": [ + 0.017696861177682877, + 0.028256047517061234, + 0.018343038856983185, + 0.026645692065358162, + 0.028644395992159843, + 0.01564711704850197, + 0.03346942737698555, + 0.046233631670475006, + 0.005539658945053816, + 0.009321474470198154, + 0.008482744917273521, + 0.012539382092654705, + 0.012513703666627407, + 0.01890813559293747, + 0.015815988183021545, + 0.015555388294160366, + 0.003118731314316392, + 0.0029538990929722786, + 0.0038511459715664387, + 0.00915435142815113, + 0.004320190288126469, + 0.0044271270744502544, + 0.008456569164991379, + 0.004020126070827246, + 0.0063045755960047245, + 0.009613344445824623, + 0.008666988462209702, + 0.00748681602999568, + 0.006806609686464071, + 0.00887537281960249, + 0.009535140357911587, + 0.005484769120812416, + 0.006566853728145361, + 0.009001843631267548, + 0.008059905841946602, + 0.009827153757214546, + 0.009920608252286911, + 0.004987786058336496, + 0.013620611280202866, + 0.009110753424465656, + 0.005864760838449001, + 0.013006247580051422, + 0.013458524830639362, + 0.012640661559998989, + 0.011464220471680164, + 0.0066560786217451096, + 0.011784957721829414, + 0.006314992439001799, + 0.00957733765244484, + 0.00889371894299984, + 0.011623547412455082, + 0.014508034102618694, + 0.010931925848126411, + 0.00892145000398159, + 0.012928362004458904, + 0.013192093931138515, + 0.009831439703702927, + 0.025509899482131004, + 0.022823341190814972, + 0.02037215791642666, + 0.016432713717222214, + 0.010244674049317837, + 0.012422299012541771, + 0.011616019532084465, + 0.012076050043106079, + 0.027160510420799255, + 0.013775644823908806, + 0.017908815294504166, + 0.013144413009285927, + 0.013491938821971416, + 0.014539106748998165, + 0.018183927983045578, + 0.01090266928076744, + 0.012605637311935425, + 0.009107089601457119, + 0.011844173073768616, + 0.007932032458484173, + 0.006951281800866127, + 0.008988912217319012, + 0.01258437242358923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.004111172165721655 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097638186067343 + ], + "opposite_logprob_tokens": [ + -5.5040974617004395 + ], + "logprob_reward": -0.004097638186067343, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097638186067343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14549919962882996, + 0.049028798937797546, + 0.04082289710640907, + 0.04973917827010155, + 0.051638100296258926, + 0.059912074357271194, + 0.13313205540180206, + 0.18870748579502106, + 0.2289115935564041, + 0.05260857194662094 + ], + "credit_map": [ + 0.012048452161252499, + 0.02549162693321705, + 0.0064804330468177795, + 0.012026671320199966, + 0.013652117922902107, + 0.00872440729290247, + 0.02576535940170288, + 0.04131012409925461, + 0.00285886088386178, + 0.003967825323343277, + 0.004017610102891922, + 0.005001118406653404, + 0.0048623597249388695, + 0.009023447521030903, + 0.010169935412704945, + 0.009127645753324032, + 0.0023958368692547083, + 0.002518359338864684, + 0.004397635348141193, + 0.010795882903039455, + 0.0037368221674114466, + 0.0049980999901890755, + 0.006506409961730242, + 0.005473849829286337, + 0.004773035645484924, + 0.00838483776897192, + 0.006529548671096563, + 0.007566491607576609, + 0.005912297405302525, + 0.005907774902880192, + 0.006054368335753679, + 0.004610828123986721, + 0.0038409456610679626, + 0.006943919230252504, + 0.005125146359205246, + 0.006211481988430023, + 0.007646594662219286, + 0.004189403727650642, + 0.011202503927052021, + 0.006478106137365103, + 0.004067887552082539, + 0.0077286772429943085, + 0.01370425708591938, + 0.00857724342495203, + 0.007538835518062115, + 0.004868685733526945, + 0.00737459072843194, + 0.006051894277334213, + 0.010987709276378155, + 0.015818331390619278, + 0.017365412786602974, + 0.015825124457478523, + 0.016528034582734108, + 0.010861877351999283, + 0.025807427242398262, + 0.019938146695494652, + 0.017374152317643166, + 0.04143442213535309, + 0.03275164216756821, + 0.03327030688524246, + 0.014092516154050827, + 0.014717875979840755, + 0.017844680696725845, + 0.01722189225256443, + 0.007044696714729071, + 0.005808854941278696, + 0.005732889752835035, + 0.004857323598116636, + 0.016823191195726395, + 0.0561053529381752, + 0.0452466681599617, + 0.0872926115989685, + 0.006905577145516872, + 0.0062539889477193356, + 0.005631915293633938, + 0.009720198810100555, + 0.005921560805290937, + 0.004925518296658993, + 0.0054486761800944805, + 0.0078011383302509785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.004097638186067343 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017116909148171544 + ], + "opposite_logprob_tokens": [ + -6.376711845397949 + ], + "logprob_reward": -0.0017116909148171544, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017116909148171544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2686347961425781, + 0.07701317220926285, + 0.07513970136642456, + 0.07848972082138062, + 0.0707610696554184, + 0.15461787581443787, + 0.056122709065675735, + 0.07102426141500473, + 0.10877615958452225, + 0.03942056745290756 + ], + "credit_map": [ + 0.02208566665649414, + 0.045797500759363174, + 0.018172012642025948, + 0.02582847885787487, + 0.021888410672545433, + 0.0159077700227499, + 0.0470457598567009, + 0.07190920412540436, + 0.004331313073635101, + 0.004778381437063217, + 0.00841288361698389, + 0.01071317121386528, + 0.005397484637796879, + 0.006114477291703224, + 0.019887926056981087, + 0.017377538606524467, + 0.004734421148896217, + 0.004470465239137411, + 0.008069959469139576, + 0.01790931262075901, + 0.007421424146741629, + 0.0071518816985189915, + 0.017044872045516968, + 0.008337358012795448, + 0.00816913042217493, + 0.010933617129921913, + 0.015036444179713726, + 0.010035069659352303, + 0.007153727114200592, + 0.008670833893120289, + 0.010109592229127884, + 0.008381301537156105, + 0.005645175464451313, + 0.007699904497712851, + 0.007310147862881422, + 0.010095533914864063, + 0.011568350717425346, + 0.005726100876927376, + 0.01294176746159792, + 0.009774084202945232, + 0.01575012318789959, + 0.026728400960564613, + 0.033257000148296356, + 0.031145408749580383, + 0.014099079184234142, + 0.010289600118994713, + 0.012558374553918839, + 0.010789887048304081, + 0.007573696319013834, + 0.010756810195744038, + 0.005748913157731295, + 0.005227426532655954, + 0.007463005371391773, + 0.00561723904684186, + 0.006718640681356192, + 0.007016979623585939, + 0.005536586046218872, + 0.01337815448641777, + 0.00837190356105566, + 0.010014149360358715, + 0.01298042293637991, + 0.0070009673945605755, + 0.006785559002310038, + 0.006956517230719328, + 0.006359488237649202, + 0.013572021387517452, + 0.006136323790997267, + 0.00803042110055685, + 0.01233394630253315, + 0.01761874556541443, + 0.016367634758353233, + 0.028357574716210365, + 0.004547903314232826, + 0.006497830618172884, + 0.004251481033861637, + 0.008050397969782352, + 0.003797353245317936, + 0.0036403320264071226, + 0.003504826221615076, + 0.005130441393703222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.0017116909148171544 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193184673786163 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014193184673786163, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193184673786163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18452709913253784, + 0.04980059713125229, + 0.05278167873620987, + 0.03814604878425598, + 0.022711830213665962, + 0.02382047474384308, + 0.04705742001533508, + 0.05442412197589874, + 0.489962100982666, + 0.03676864132285118 + ], + "credit_map": [ + 0.015417711809277534, + 0.02568923495709896, + 0.01661631464958191, + 0.023995256051421165, + 0.015282385051250458, + 0.011739242821931839, + 0.031939033418893814, + 0.04384789988398552, + 0.0034414203837513924, + 0.004893973469734192, + 0.005271316505968571, + 0.006348205730319023, + 0.005907750688493252, + 0.007804397959262133, + 0.008191746659576893, + 0.007941779680550098, + 0.0047590541653335094, + 0.004763586446642876, + 0.006995661184191704, + 0.010286180302500725, + 0.005079424008727074, + 0.005117061547935009, + 0.010067705065011978, + 0.005713003221899271, + 0.0035316594876348972, + 0.003785761073231697, + 0.004783186595886946, + 0.002644026419147849, + 0.0035412819124758244, + 0.0074814376421272755, + 0.00864854734390974, + 0.0037301410920917988, + 0.002084142528474331, + 0.0024346522986888885, + 0.002694403985515237, + 0.0029993585776537657, + 0.0034292000345885754, + 0.0020676918793469667, + 0.004665781743824482, + 0.0023365961387753487, + 0.001635587657801807, + 0.002700544660910964, + 0.005001181736588478, + 0.004375231917947531, + 0.0030963290482759476, + 0.001816488686017692, + 0.0029102128464728594, + 0.0022848958615213633, + 0.008000344038009644, + 0.005180825479328632, + 0.008193792775273323, + 0.004083198495209217, + 0.005546264816075563, + 0.005314873997122049, + 0.005075525958091021, + 0.005662590730935335, + 0.0038359863683581352, + 0.009513022378087044, + 0.006421108264476061, + 0.008338764309883118, + 0.010131793096661568, + 0.00576595077291131, + 0.00485960440710187, + 0.005557887721806765, + 0.014923014678061008, + 0.012433587573468685, + 0.012779825367033482, + 0.010467813350260258, + 0.03492669016122818, + 0.12101323902606964, + 0.09593432396650314, + 0.18748357892036438, + 0.005029304418712854, + 0.004290197975933552, + 0.003974960185587406, + 0.007640991825610399, + 0.00401335209608078, + 0.0035628790501505136, + 0.0035882031079381704, + 0.00466874847188592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.014193184673786163 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, shot from a low angle, closely follows the young Nezha as he navigates the turbulent deep blue sea. He wears his signature twin buns, a few unruly strands of black hair dancing in the wind, his face brimming with a rebellious yet playful smile. He wears a bright red bellyband intricately embroidered with golden lotus patterns. His bare arms are wrapped with the Hun Tian Ling, the silk ribbons rustling in the fierce wind like two living red serpents. His feet rest firmly on a pair of fiery wind-fire wheels. The wheels spin at high speed, leaving long trails of golden flame, instantly vaporizing any seawater they touch into hissing white steam. With a near-dance, he deftly glides and leaps over the crests of the mountainous waves, leaving two vast white wakes in his wake, churning the water behind him. The water, stirred by the wind-fire wheels, quickly converges into a single, towering wave. The sky is thick with dark clouds, punctuated by thunder and lightning. Occasionally, wisps of divine golden light peek through the cracks in the clouds, interplaying beautifully with the fiery glow of the Wind Fire Wheel and the pale lightning streaking across the sky. Light and shadow playfully shift across the inky sea. The dominant tones of the painting are deep, cool blues and grays, ignited by the vibrant red of Nezha's figure and the fiery golden orange of the Wind Fire Wheel. This creates an atmosphere of epic mythology, tension, and power, showcasing the powerful aura of the young deity, defying heaven and earth and possessing unbridled freedom.", + "seed": 25722, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_2/index_4793_seed_25722_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.011540059256367386, + "vlm_logprob": -0.011540059256367386, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.011540059256367386, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03984392434358597 + ], + "opposite_logprob_tokens": [ + -3.289844036102295 + ], + "logprob_reward": -0.03984392434358597, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03984392434358597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35050109028816223, + 0.05763160437345505, + 0.1366320103406906, + 0.1098453477025032, + 0.0447375550866127, + 0.09747464954853058, + 0.029607845470309258, + 0.036174848675727844, + 0.07510047405958176, + 0.06229463219642639 + ], + "credit_map": [ + 0.03378455340862274, + 0.041346363723278046, + 0.05623576045036316, + 0.0697440356016159, + 0.038261059671640396, + 0.04256599396467209, + 0.04013269767165184, + 0.02843061089515686, + 0.01052343100309372, + 0.008948187343776226, + 0.006654086522758007, + 0.00831827986985445, + 0.008090781047940254, + 0.006308646406978369, + 0.0038165398873388767, + 0.0049716513603925705, + 0.020707320421934128, + 0.022484345361590385, + 0.012712356634438038, + 0.013982648029923439, + 0.015784328803420067, + 0.01576913520693779, + 0.012113354168832302, + 0.02307851053774357, + 0.01155881304293871, + 0.02046988718211651, + 0.01077057234942913, + 0.013935505412518978, + 0.016311360523104668, + 0.01115628145635128, + 0.010969012044370174, + 0.014673913829028606, + 0.00668419199064374, + 0.00860101729631424, + 0.004004506394267082, + 0.006969837471842766, + 0.004561872687190771, + 0.004919738508760929, + 0.0051138680428266525, + 0.0038825238589197397, + 0.01956758461892605, + 0.024735763669013977, + 0.006516297813504934, + 0.005741140339523554, + 0.01576913148164749, + 0.013977093622088432, + 0.004590270109474659, + 0.006577373947948217, + 0.006179469637572765, + 0.0052875252440571785, + 0.00200012419372797, + 0.0024307682178914547, + 0.005598769523203373, + 0.004845903255045414, + 0.0017355357995256782, + 0.0015297482023015618, + 0.007146190386265516, + 0.003389418590813875, + 0.0025380519218742847, + 0.005417493171989918, + 0.004831062629818916, + 0.003869228530675173, + 0.004996159579604864, + 0.0039872415363788605, + 0.00797862745821476, + 0.01417409349232912, + 0.0072625684551894665, + 0.010089168325066566, + 0.00880573783069849, + 0.013403480872511864, + 0.008143608458340168, + 0.005243187304586172, + 0.011427236720919609, + 0.00920625776052475, + 0.0068963817320764065, + 0.00636812811717391, + 0.006449534557759762, + 0.007584288250654936, + 0.006975088734179735, + 0.00738771166652441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.03984392434358597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004651084542274475 + ], + "opposite_logprob_tokens": [ + -5.379651069641113 + ], + "logprob_reward": -0.004651084542274475, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004651084542274475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4062832295894623, + 0.06918437033891678, + 0.08767597377300262, + 0.0646633729338646, + 0.04636240750551224, + 0.09342249482870102, + 0.06882143765687943, + 0.04346459358930588, + 0.0473223440349102, + 0.07279985398054123 + ], + "credit_map": [ + 0.04370700195431709, + 0.05356844887137413, + 0.06205415353178978, + 0.08414686471223831, + 0.03942137584090233, + 0.043129514902830124, + 0.05392736941576004, + 0.02632843889296055, + 0.014946089126169682, + 0.01006056647747755, + 0.00831009354442358, + 0.007875064387917519, + 0.010037407279014587, + 0.008269191719591618, + 0.003996274434030056, + 0.005689676385372877, + 0.011232494376599789, + 0.012490847148001194, + 0.016286853700876236, + 0.009089769795536995, + 0.012985831126570702, + 0.010068537667393684, + 0.008082251064479351, + 0.007439377252012491, + 0.00739396084100008, + 0.011205979622900486, + 0.0063131144270300865, + 0.007738939486443996, + 0.009118935093283653, + 0.008078615181148052, + 0.006417953409254551, + 0.008395863696932793, + 0.008333957754075527, + 0.008139698766171932, + 0.00416975375264883, + 0.0065806591883301735, + 0.005155330523848534, + 0.006309953052550554, + 0.0037269059102982283, + 0.003946144133806229, + 0.021439017727971077, + 0.023915350437164307, + 0.0055375006049871445, + 0.005293055437505245, + 0.01116181816905737, + 0.015287423506379128, + 0.004674245603382587, + 0.0061140661127865314, + 0.009439106099307537, + 0.014451351948082447, + 0.0056130653247237206, + 0.006014215759932995, + 0.008818995207548141, + 0.01602032221853733, + 0.0033987367060035467, + 0.0050656297244131565, + 0.010657216422259808, + 0.005881181452423334, + 0.0023566815070807934, + 0.0026026980485767126, + 0.006270148325711489, + 0.007985484786331654, + 0.00365197891369462, + 0.004059198312461376, + 0.005939536262303591, + 0.007584740407764912, + 0.005308612249791622, + 0.006673326250165701, + 0.00476813456043601, + 0.0068374453112483025, + 0.006267087068408728, + 0.003943456802517176, + 0.008836575783789158, + 0.009909492917358875, + 0.009143244475126266, + 0.008937506005167961, + 0.008752329275012016, + 0.008888058364391327, + 0.00918314978480339, + 0.009149493649601936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.004651084542274475 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001715499092824757 + ], + "opposite_logprob_tokens": [ + -6.376715660095215 + ], + "logprob_reward": -0.001715499092824757, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001715499092824757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18397146463394165, + 0.10617268085479736, + 0.1366562694311142, + 0.07026978582143784, + 0.07865927368402481, + 0.11360529810190201, + 0.05413832515478134, + 0.059783756732940674, + 0.0837622582912445, + 0.11298089474439621 + ], + "credit_map": [ + 0.020330334082245827, + 0.01998264715075493, + 0.029112419113516808, + 0.03532567247748375, + 0.023138148710131645, + 0.018998639658093452, + 0.02018681913614273, + 0.016896778717637062, + 0.020163124427199364, + 0.014827900566160679, + 0.013902172446250916, + 0.013486451469361782, + 0.01292875874787569, + 0.011724052019417286, + 0.008976363576948643, + 0.010163858532905579, + 0.01573343016207218, + 0.017400603741407394, + 0.025922250002622604, + 0.013497606851160526, + 0.01919691450893879, + 0.014600266702473164, + 0.011742498725652695, + 0.018562693148851395, + 0.00819496437907219, + 0.011785202659666538, + 0.006122114136815071, + 0.007838737219572067, + 0.010386128909885883, + 0.007890603505074978, + 0.0075483438558876514, + 0.010503692552447319, + 0.01370308455079794, + 0.0161738321185112, + 0.006742421537637711, + 0.009151396341621876, + 0.011723138391971588, + 0.0068480041809380054, + 0.005131251644343138, + 0.00918614398688078, + 0.032183244824409485, + 0.04076549410820007, + 0.006515949033200741, + 0.004509329795837402, + 0.009113741107285023, + 0.009839791804552078, + 0.0036833519116044044, + 0.0069943987764418125, + 0.007740535773336887, + 0.010744773782789707, + 0.005493655800819397, + 0.004623966291546822, + 0.008178161457180977, + 0.010218645446002483, + 0.002874452853575349, + 0.004264131188392639, + 0.015910709276795387, + 0.007370400708168745, + 0.002801570575684309, + 0.004377881530672312, + 0.009301395155489445, + 0.008674927987158298, + 0.005949812475591898, + 0.005397055763751268, + 0.012656113132834435, + 0.022304385900497437, + 0.004263489041477442, + 0.006133944261819124, + 0.006467482540756464, + 0.01841900497674942, + 0.00601237965747714, + 0.007505455054342747, + 0.017070457339286804, + 0.015060607343912125, + 0.011092533357441425, + 0.010519620962440968, + 0.01717105694115162, + 0.012483473867177963, + 0.01922386884689331, + 0.010359283536672592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.001715499092824757 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001711095916107297 + ], + "opposite_logprob_tokens": [ + -6.376710891723633 + ], + "logprob_reward": -0.001711095916107297, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001711095916107297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16602960228919983, + 0.05119400471448898, + 0.05476763844490051, + 0.02646586485207081, + 0.055036891251802444, + 0.14819268882274628, + 0.15581156313419342, + 0.10253527760505676, + 0.11477252095937729, + 0.1251939833164215 + ], + "credit_map": [ + 0.013151943683624268, + 0.024469701573252678, + 0.014982934109866619, + 0.014461943879723549, + 0.019754135981202126, + 0.028374088928103447, + 0.032441359013319016, + 0.01839347369968891, + 0.004003699403256178, + 0.005076761357486248, + 0.004812181461602449, + 0.005697987042367458, + 0.010304608382284641, + 0.007137823849916458, + 0.0055079273879528046, + 0.008653009310364723, + 0.005773080047219992, + 0.008493638597428799, + 0.007972905412316322, + 0.005754973739385605, + 0.009446339681744576, + 0.006360290106385946, + 0.0055189551785588264, + 0.005447452422231436, + 0.003343035001307726, + 0.004888957366347313, + 0.002487110672518611, + 0.0030332040041685104, + 0.004215479362756014, + 0.003076119115576148, + 0.0019224407151341438, + 0.0034995153546333313, + 0.007428967393934727, + 0.007949408143758774, + 0.005849837325513363, + 0.006789020262658596, + 0.005540172103792429, + 0.006404550280421972, + 0.005004503298550844, + 0.01007042732089758, + 0.03954320400953293, + 0.050457462668418884, + 0.005954033695161343, + 0.006081223022192717, + 0.016320211812853813, + 0.017145875841379166, + 0.0052346717566251755, + 0.007455986924469471, + 0.027732789516448975, + 0.02159249037504196, + 0.012439647689461708, + 0.016181478276848793, + 0.02956344187259674, + 0.029522303491830826, + 0.00746978959068656, + 0.011309592053294182, + 0.02812526933848858, + 0.013069742359220982, + 0.006396871525794268, + 0.009155143983662128, + 0.015683485195040703, + 0.013937150128185749, + 0.007962520234286785, + 0.008205090649425983, + 0.025576239451766014, + 0.017404751852154732, + 0.010293981991708279, + 0.011346371844410896, + 0.014885627664625645, + 0.018024055287241936, + 0.008689417503774166, + 0.0085520651191473, + 0.025570927187800407, + 0.017945216968655586, + 0.011025244370102882, + 0.0109401261433959, + 0.015663785859942436, + 0.016810297966003418, + 0.014073890633881092, + 0.013164477422833443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.001711095916107297 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977869238704443 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.00977869238704443, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977869238704443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2529394030570984, + 0.06090093404054642, + 0.09494362771511078, + 0.05541030690073967, + 0.06064550206065178, + 0.06079639121890068, + 0.05329101160168648, + 0.1285458505153656, + 0.12677909433841705, + 0.10574789345264435 + ], + "credit_map": [ + 0.029896173626184464, + 0.03991929814219475, + 0.024069197475910187, + 0.021631112322211266, + 0.029857957735657692, + 0.04165934771299362, + 0.037549808621406555, + 0.028356501832604408, + 0.0067771789617836475, + 0.00989762507379055, + 0.00538831390440464, + 0.007871891371905804, + 0.009713482111692429, + 0.0064317318610847, + 0.00678610522300005, + 0.008034603670239449, + 0.011238247156143188, + 0.015738803893327713, + 0.01025273185223341, + 0.010687668807804585, + 0.014150670729577541, + 0.01138575840741396, + 0.008131706155836582, + 0.013358035124838352, + 0.007687930017709732, + 0.01065424457192421, + 0.006010608281940222, + 0.00543105648830533, + 0.007716700434684753, + 0.006172172259539366, + 0.00452460115775466, + 0.007212992757558823, + 0.010356355458498001, + 0.007529321126639843, + 0.0050727203488349915, + 0.008549025282263756, + 0.011629821732640266, + 0.005736570805311203, + 0.005390000995248556, + 0.006381686311215162, + 0.010430593974888325, + 0.011791713535785675, + 0.004301173612475395, + 0.004698087926954031, + 0.00853596068918705, + 0.012353074736893177, + 0.004062958061695099, + 0.0046228282153606415, + 0.007523390464484692, + 0.010104291141033173, + 0.003263284219428897, + 0.006400407757610083, + 0.010198800824582577, + 0.008028891868889332, + 0.0045145112089812756, + 0.003257433883845806, + 0.03445577621459961, + 0.016434766352176666, + 0.006777612026780844, + 0.009651285596191883, + 0.017055833712220192, + 0.015067854896187782, + 0.017723139375448227, + 0.011379587464034557, + 0.025204692035913467, + 0.02675740048289299, + 0.008861059322953224, + 0.010332509875297546, + 0.012597691267728806, + 0.024678675457835197, + 0.009350345470011234, + 0.008996719494462013, + 0.023229237645864487, + 0.015539764426648617, + 0.007730612996965647, + 0.009092827327549458, + 0.012506969273090363, + 0.01491261925548315, + 0.01234347652643919, + 0.010392391122877598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.00977869238704443 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, shot from a low angle, closely follows the young Nezha as he navigates the turbulent deep blue sea. He wears his signature twin buns, a few unruly strands of black hair dancing in the wind, his face brimming with a rebellious yet playful smile. He wears a bright red bellyband intricately embroidered with golden lotus patterns. His bare arms are wrapped with the Hun Tian Ling, the silk ribbons rustling in the fierce wind like two living red serpents. His feet rest firmly on a pair of fiery wind-fire wheels. The wheels spin at high speed, leaving long trails of golden flame, instantly vaporizing any seawater they touch into hissing white steam. With a near-dance, he deftly glides and leaps over the crests of the mountainous waves, leaving two vast white wakes in his wake, churning the water behind him. The water, stirred by the wind-fire wheels, quickly converges into a single, towering wave. The sky is thick with dark clouds, punctuated by thunder and lightning. Occasionally, wisps of divine golden light peek through the cracks in the clouds, interplaying beautifully with the fiery glow of the Wind Fire Wheel and the pale lightning streaking across the sky. Light and shadow playfully shift across the inky sea. The dominant tones of the painting are deep, cool blues and grays, ignited by the vibrant red of Nezha's figure and the fiery golden orange of the Wind Fire Wheel. This creates an atmosphere of epic mythology, tension, and power, showcasing the powerful aura of the young deity, defying heaven and earth and possessing unbridled freedom.", + "seed": 25723, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_2/index_4793_seed_25723_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.030562536604702473, + "vlm_logprob": -0.030562536604702473, + "vlm_yes_no_margin": 4.624999976158142, + "vlm_token_credit": -0.030562536604702473, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13015325367450714 + ], + "opposite_logprob_tokens": [ + -2.130153179168701 + ], + "logprob_reward": -0.13015325367450714, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.13015325367450714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1163274422287941, + 0.27624082565307617, + 0.08642975240945816, + 0.11538288742303848, + 0.026372328400611877, + 0.05559629201889038, + 0.10064181685447693, + 0.022716082632541656, + 0.13837189972400665, + 0.06192067265510559 + ], + "credit_map": [ + 0.017402924597263336, + 0.018482716754078865, + 0.0163401011377573, + 0.011647559702396393, + 0.015182459726929665, + 0.00946812704205513, + 0.014889813028275967, + 0.012913732789456844, + 0.037216462194919586, + 0.030865516513586044, + 0.04065391421318054, + 0.026210946962237358, + 0.03858545422554016, + 0.0228298120200634, + 0.028393443673849106, + 0.05148524418473244, + 0.011895350180566311, + 0.012621184810996056, + 0.01586073823273182, + 0.005763128399848938, + 0.010287224315106869, + 0.010314426384866238, + 0.006860369350761175, + 0.012827323749661446, + 0.016542300581932068, + 0.019686203449964523, + 0.01215040497481823, + 0.012139830738306046, + 0.016329655423760414, + 0.013032822869718075, + 0.014189102686941624, + 0.01131253968924284, + 0.003698582062497735, + 0.003328025573864579, + 0.003856768598780036, + 0.002004190580919385, + 0.004184028133749962, + 0.003948305267840624, + 0.0024470514617860317, + 0.0029053741600364447, + 0.006040253676474094, + 0.007978796027600765, + 0.005962434224784374, + 0.004354466218501329, + 0.011205637827515602, + 0.007440952118486166, + 0.006505447905510664, + 0.006108299363404512, + 0.011984877288341522, + 0.017450138926506042, + 0.008880694396793842, + 0.01146889477968216, + 0.01462993398308754, + 0.021391302347183228, + 0.009144282899796963, + 0.00569168059155345, + 0.0026163423899561167, + 0.005290130618959665, + 0.001366994227282703, + 0.0017800283385440707, + 0.0034380678553134203, + 0.0025143628008663654, + 0.0024526456836611032, + 0.0032575076911598444, + 0.019058644771575928, + 0.015881745144724846, + 0.019204402342438698, + 0.011630569584667683, + 0.01729394868016243, + 0.02783377282321453, + 0.011562447063624859, + 0.015906358137726784, + 0.0072880759835243225, + 0.011916948482394218, + 0.007908941246569157, + 0.006410098168998957, + 0.008391963317990303, + 0.007898983545601368, + 0.007619995623826981, + 0.004485657438635826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.13015325367450714 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008653522469103336 + ], + "opposite_logprob_tokens": [ + -4.75865364074707 + ], + "logprob_reward": -0.008653522469103336, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008653522469103336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10849040001630783, + 0.07396045327186584, + 0.08501468598842621, + 0.09380035847425461, + 0.025589650496840477, + 0.027771463617682457, + 0.08955048769712448, + 0.1160835549235344, + 0.2968422472476959, + 0.08289671689271927 + ], + "credit_map": [ + 0.014145527966320515, + 0.01845156028866768, + 0.010293382219970226, + 0.012763986364006996, + 0.015498187392950058, + 0.008681436069309711, + 0.015954922884702682, + 0.01270139031112194, + 0.011672664433717728, + 0.008344240486621857, + 0.008946258574724197, + 0.005581527948379517, + 0.011018139310181141, + 0.006065803579986095, + 0.008968507871031761, + 0.013363312929868698, + 0.012039043940603733, + 0.014053244143724442, + 0.013732418417930603, + 0.004803346935659647, + 0.012706640176475048, + 0.010672450996935368, + 0.0053542316891252995, + 0.01165330782532692, + 0.014085464179515839, + 0.01594644784927368, + 0.010498874820768833, + 0.008638173341751099, + 0.01225967425853014, + 0.011259088292717934, + 0.010823159478604794, + 0.010289478115737438, + 0.0031224831473082304, + 0.0031305127777159214, + 0.002841431647539139, + 0.002404791535809636, + 0.005647234618663788, + 0.0034584368113428354, + 0.002677370561286807, + 0.002307388000190258, + 0.0031378769781440496, + 0.003409377532079816, + 0.0039094965904951096, + 0.002129720989614725, + 0.005083938594907522, + 0.0032750265672802925, + 0.003438781015574932, + 0.003387247445061803, + 0.009650588035583496, + 0.011840207502245903, + 0.01003695372492075, + 0.009222579188644886, + 0.01202599611133337, + 0.019274728372693062, + 0.009314123541116714, + 0.008185312151908875, + 0.00914839468896389, + 0.020627012476325035, + 0.007281245198100805, + 0.0093154301866889, + 0.0218998733907938, + 0.022629421204328537, + 0.014998475089669228, + 0.010183694772422314, + 0.03610425814986229, + 0.019337767735123634, + 0.058602750301361084, + 0.027959195896983147, + 0.031553518027067184, + 0.043848976492881775, + 0.022353198379278183, + 0.05708259716629982, + 0.016560224816203117, + 0.011927508749067783, + 0.009624787606298923, + 0.006544565316289663, + 0.010908144526183605, + 0.014852411113679409, + 0.007646685000509024, + 0.004832392558455467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.008653522469103336 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193043638020754 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.003193043638020754, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193043638020754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09902497380971909, + 0.28189969062805176, + 0.07548879832029343, + 0.14886334538459778, + 0.018112674355506897, + 0.023449204862117767, + 0.0351773239672184, + 0.020909486338496208, + 0.24233345687389374, + 0.05474107339978218 + ], + "credit_map": [ + 0.013187715783715248, + 0.017141073942184448, + 0.010003551840782166, + 0.00950619112700224, + 0.015373485162854195, + 0.008154276758432388, + 0.016515888273715973, + 0.009142783470451832, + 0.03572438284754753, + 0.027497196570038795, + 0.04397899657487869, + 0.024456754326820374, + 0.03944522887468338, + 0.02358809858560562, + 0.033593155443668365, + 0.05361584946513176, + 0.009937808848917484, + 0.011291446164250374, + 0.013318742625415325, + 0.004527952056378126, + 0.009810865856707096, + 0.008731589652597904, + 0.005846122279763222, + 0.012024263851344585, + 0.01978091150522232, + 0.025847794488072395, + 0.008183103054761887, + 0.011129106394946575, + 0.04115506634116173, + 0.015992099419236183, + 0.0178329236805439, + 0.008942322805523872, + 0.002324212808161974, + 0.0025608434807509184, + 0.0015195627929642797, + 0.001354895532131195, + 0.0034758085384964943, + 0.0020275949500501156, + 0.0021334784105420113, + 0.0027162772603332996, + 0.0025007897056639194, + 0.002775084925815463, + 0.002562637673690915, + 0.001839829026721418, + 0.00528831547126174, + 0.0036636388394981623, + 0.002925901673734188, + 0.0018930054502561688, + 0.0037742352578788996, + 0.00550567451864481, + 0.004906762856990099, + 0.0034634594339877367, + 0.004112790338695049, + 0.005376600660383701, + 0.0029178329277783632, + 0.005119962152093649, + 0.002147697377949953, + 0.0034211985766887665, + 0.0020905504934489727, + 0.0017858099890872836, + 0.0030303841922432184, + 0.0027049430646002293, + 0.004063586704432964, + 0.0016653130296617746, + 0.02738308720290661, + 0.016705622896552086, + 0.04644254222512245, + 0.022949235513806343, + 0.0230247899889946, + 0.04129350185394287, + 0.01790863461792469, + 0.0466260127723217, + 0.012080620974302292, + 0.006254390347748995, + 0.005050954408943653, + 0.0050697335973382, + 0.0071277315728366375, + 0.010829577222466469, + 0.004761262331157923, + 0.0035668006166815758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.003193043638020754 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031884093768894672 + ], + "opposite_logprob_tokens": [ + -5.753188610076904 + ], + "logprob_reward": -0.0031884093768894672, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031884093768894672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23916050791740417, + 0.18034204840660095, + 0.08412417024374008, + 0.07861846685409546, + 0.09749636054039001, + 0.10509002208709717, + 0.035573072731494904, + 0.046316228806972504, + 0.07940112799406052, + 0.0538780651986599 + ], + "credit_map": [ + 0.03290025144815445, + 0.03624183312058449, + 0.029842559248209, + 0.02270227111876011, + 0.031012780964374542, + 0.021040398627519608, + 0.027581702917814255, + 0.03783869370818138, + 0.026525219902396202, + 0.01917976699769497, + 0.028861915692687035, + 0.017687421292066574, + 0.023187221959233284, + 0.014544758014380932, + 0.016054334118962288, + 0.03430140018463135, + 0.010282095521688461, + 0.010778616182506084, + 0.01403884869068861, + 0.007137078791856766, + 0.012045882642269135, + 0.009287038818001747, + 0.009742106311023235, + 0.010812492109835148, + 0.01120722759515047, + 0.013960528187453747, + 0.006714182905852795, + 0.007513371761888266, + 0.011976388283073902, + 0.007938489317893982, + 0.011224907822906971, + 0.008083360269665718, + 0.01210886612534523, + 0.011766224168241024, + 0.011993576772511005, + 0.00901362020522356, + 0.01586782932281494, + 0.01060402113944292, + 0.010272402316331863, + 0.015869809314608574, + 0.017061887308955193, + 0.012592158280313015, + 0.010522737167775631, + 0.005569149740040302, + 0.020859410986304283, + 0.014331457205116749, + 0.009549536742269993, + 0.014603672549128532, + 0.003945864271372557, + 0.0074196564964950085, + 0.0037174213211983442, + 0.004995645489543676, + 0.004819977562874556, + 0.0043273488990962505, + 0.0031657570507377386, + 0.0031813974492251873, + 0.0054334416054189205, + 0.006077860482037067, + 0.004403335507959127, + 0.0032767888624221087, + 0.006805452052503824, + 0.009816542267799377, + 0.004338060040026903, + 0.006164744962006807, + 0.0073821935802698135, + 0.0035068760626018047, + 0.008657791651785374, + 0.0028710567858070135, + 0.021349668502807617, + 0.025826307013630867, + 0.0028160708025097847, + 0.0069911545142531395, + 0.008037938736379147, + 0.006214913912117481, + 0.009021415375173092, + 0.00530696427449584, + 0.0066361455246806145, + 0.007796873804181814, + 0.005005148239433765, + 0.005858665332198143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0031884093768894672 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young character named Nezha with twin buns and unruly black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a bright red bellyband with golden lotus patterns and bare arms wrapped in red silk ribbons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character standing on fiery wind-fire wheels that spin at high speed leaving golden flame trails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the character gliding and leaping over mountainous waves in a turbulent deep blue sea with dark clouds and lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low angle with a dynamic tracking motion following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624453864991665 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624453864991665, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624453864991665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10850193351507187, + 0.13058510422706604, + 0.06091796234250069, + 0.10380364954471588, + 0.01753341592848301, + 0.015476108528673649, + 0.061719369143247604, + 0.05607995390892029, + 0.4092194736003876, + 0.036163054406642914 + ], + "credit_map": [ + 0.019816545769572258, + 0.017121169716119766, + 0.012756734155118465, + 0.010018336586654186, + 0.013987146317958832, + 0.009216099977493286, + 0.010164435021579266, + 0.015421462245285511, + 0.019169310107827187, + 0.011570642702281475, + 0.02111106552183628, + 0.01082297507673502, + 0.017901495099067688, + 0.010530692525207996, + 0.013271075673401356, + 0.02620784193277359, + 0.008429291658103466, + 0.00788511335849762, + 0.009494712576270103, + 0.003879606258124113, + 0.010253828950226307, + 0.008333178237080574, + 0.004090581554919481, + 0.008551652543246746, + 0.012719111517071724, + 0.017952896654605865, + 0.010660526342689991, + 0.009439944289624691, + 0.020500315353274345, + 0.010650795884430408, + 0.01033061370253563, + 0.011549452319741249, + 0.0022622928954660892, + 0.002579983789473772, + 0.002356997225433588, + 0.001961668021976948, + 0.002531324280425906, + 0.0017997275572270155, + 0.001769919297657907, + 0.0022715041413903236, + 0.0016892848070710897, + 0.0019131405279040337, + 0.0020980655681341887, + 0.001367568620480597, + 0.0023936505895107985, + 0.00245221727527678, + 0.0020109645556658506, + 0.0015512165846303105, + 0.007309604436159134, + 0.010401112958788872, + 0.008502859622240067, + 0.005618379916995764, + 0.008813025429844856, + 0.009912487119436264, + 0.004512932617217302, + 0.006648968905210495, + 0.005732952617108822, + 0.009013715200126171, + 0.005339327268302441, + 0.0042907497845590115, + 0.009130781516432762, + 0.010111786425113678, + 0.005205375142395496, + 0.007255263160914183, + 0.04873037710785866, + 0.029634635895490646, + 0.07626108825206757, + 0.038201864808797836, + 0.0397561714053154, + 0.0686035007238388, + 0.03128763288259506, + 0.0767442137002945, + 0.00736978929489851, + 0.004230143502354622, + 0.00367629062384367, + 0.0031085654627531767, + 0.004835594445466995, + 0.006613502744585276, + 0.0035231190267950296, + 0.002806049771606922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.007624453864991665 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_3/index_tensor([5614, 5614])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_3/index_tensor([5614, 5614])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5b32f46dbc16838efa050938a62e2d264771c142 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_3/index_tensor([5614, 5614])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up, static shot captured from a slight downward angle focuses squarely on a nautilus with a magnificent spiral shell, resting peacefully on a bed of fine white sand. Its shell, a perfect Fibonacci spiral, is smooth as jade, its pearly, opalescent base punctuated by delicate russet-brown stripes radiating outward from the center. A soft, diffuse light filters down from the crystal-clear azure water, casting a dreamy halo over the smooth shell. Pastel-colored anemones sway gently in the background, blurring slightly. As the camera gazes, the nautilus carefully and slowly unfurls dozens of soft, translucent, opalescent tentacles from its shell. Like newly opened petals, these tentacles tentatively reach outward, their movements gentle and curious. The light softly illuminates the texture of each tentacle, making them look exceptionally delicate. The entire picture, with its tranquil blue and warm white tones, creates an atmosphere of ultimate tranquility, innocence, and a fairy tale.", + "seed": 9023, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_3/index_5614_seed_9023_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.024012535717338322, + "vlm_logprob": -0.024012535717338322, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.024012535717338322, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002253494691103697 + ], + "opposite_logprob_tokens": [ + -7.002253532409668 + ], + "logprob_reward": -0.002253494691103697, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.002253494691103697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1042269617319107, + 0.13260675966739655, + 0.02940591610968113, + 0.11536132544279099, + 0.0731038972735405, + 0.07779999822378159, + 0.08034274727106094, + 0.1328163594007492, + 0.1723809689283371, + 0.08195509016513824 + ], + "credit_map": [ + 0.011517791077494621, + 0.005893053486943245, + 0.01894857920706272, + 0.01196438167244196, + 0.006125674117356539, + 0.02891872636973858, + 0.008505779318511486, + 0.012352973222732544, + 0.022402649745345116, + 0.0076299323700368404, + 0.024184109643101692, + 0.0104446392506361, + 0.008295310661196709, + 0.03608761727809906, + 0.007851923815906048, + 0.015710581094026566, + 0.004864683374762535, + 0.002099303063005209, + 0.0043070693500339985, + 0.0034435121342539787, + 0.0026889960281550884, + 0.003816393204033375, + 0.004635733086615801, + 0.003550225170329213, + 0.017139634117484093, + 0.004769298247992992, + 0.0290752612054348, + 0.007605916820466518, + 0.006104655098170042, + 0.02911187894642353, + 0.006770275067538023, + 0.014784403145313263, + 0.005818923003971577, + 0.006395564880222082, + 0.012557991780340672, + 0.012018677778542042, + 0.006525993812829256, + 0.00994874071329832, + 0.011283524334430695, + 0.00855447817593813, + 0.006014256738126278, + 0.009680979885160923, + 0.010546723380684853, + 0.009565046057105064, + 0.010630505159497261, + 0.010918437503278255, + 0.014323585666716099, + 0.006120460107922554, + 0.012965795584022999, + 0.008037334308028221, + 0.010398431681096554, + 0.011956410482525826, + 0.006707462482154369, + 0.01055054273456335, + 0.00949648767709732, + 0.010230284184217453, + 0.01610676757991314, + 0.014592607505619526, + 0.017143970355391502, + 0.011027797125279903, + 0.02772107720375061, + 0.018147138878703117, + 0.01298571564257145, + 0.015091289766132832, + 0.017672419548034668, + 0.012986508198082447, + 0.03704724833369255, + 0.014394731260836124, + 0.011246073059737682, + 0.044904161244630814, + 0.01944255270063877, + 0.014687270857393742, + 0.006406843196600676, + 0.006838728673756123, + 0.010739823803305626, + 0.0051335180178284645, + 0.007893611676990986, + 0.013625840656459332, + 0.009727906435728073, + 0.021588817238807678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.002253494691103697 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980329655110836 + ], + "opposite_logprob_tokens": [ + -3.5298032760620117 + ], + "logprob_reward": -0.02980329655110836, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980329655110836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057951342314481735, + 0.11972931027412415, + 0.12829771637916565, + 0.06956747174263, + 0.10180871188640594, + 0.08284131437540054, + 0.14319157600402832, + 0.12671656906604767, + 0.12671750783920288, + 0.043178509920835495 + ], + "credit_map": [ + 0.010172390379011631, + 0.0062370337545871735, + 0.0076158493757247925, + 0.005393543746322393, + 0.006950187496840954, + 0.006882780697196722, + 0.006236938759684563, + 0.008462616242468357, + 0.02031899057328701, + 0.006841251160949469, + 0.027545979246497154, + 0.009414919652044773, + 0.006225026678293943, + 0.030376704409718513, + 0.0056955162435770035, + 0.013310915790498257, + 0.023645956069231033, + 0.007161838002502918, + 0.029770366847515106, + 0.006976759992539883, + 0.006060646381229162, + 0.03202493116259575, + 0.008505878038704395, + 0.014151329174637794, + 0.008433669805526733, + 0.002859120722860098, + 0.01942685805261135, + 0.004461425356566906, + 0.0028930052649229765, + 0.021818315610289574, + 0.0037519049365073442, + 0.005923178978264332, + 0.00854006502777338, + 0.010123339481651783, + 0.016438359394669533, + 0.011909683234989643, + 0.01515913661569357, + 0.013736060820519924, + 0.015583891421556473, + 0.010318179614841938, + 0.009787109680473804, + 0.0027148197405040264, + 0.024793799966573715, + 0.004734379705041647, + 0.0021981706377118826, + 0.029817933216691017, + 0.005028930958360434, + 0.003766172332689166, + 0.023120112717151642, + 0.015425130724906921, + 0.018153350800275803, + 0.011344294995069504, + 0.024724949151277542, + 0.016773775219917297, + 0.011693326756358147, + 0.02195665054023266, + 0.027583755552768707, + 0.012256418354809284, + 0.018940012902021408, + 0.01008536759763956, + 0.018803825601935387, + 0.013817014172673225, + 0.010559914633631706, + 0.014670254662632942, + 0.013515998609364033, + 0.010273806750774384, + 0.02839147485792637, + 0.012911481782793999, + 0.007448997348546982, + 0.03735937550663948, + 0.009002355858683586, + 0.0078140152618289, + 0.003189081558957696, + 0.005783299915492535, + 0.00539119029417634, + 0.007562555838376284, + 0.00428431574255228, + 0.004083045292645693, + 0.004166824277490377, + 0.008718199096620083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.02980329655110836 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866454377770424 + ], + "opposite_logprob_tokens": [ + -3.0486645698547363 + ], + "logprob_reward": -0.04866454377770424, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866454377770424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1676267683506012, + 0.16417920589447021, + 0.0927530825138092, + 0.09526602178812027, + 0.14572449028491974, + 0.08299359679222107, + 0.13631264865398407, + 0.028342559933662415, + 0.058617737144231796, + 0.028183860704302788 + ], + "credit_map": [ + 0.025419993326067924, + 0.008422590792179108, + 0.03863692283630371, + 0.014823849312961102, + 0.010038028471171856, + 0.044594306498765945, + 0.008593500591814518, + 0.017097586765885353, + 0.028735410422086716, + 0.009850658476352692, + 0.03660471364855766, + 0.01249642577022314, + 0.010643737390637398, + 0.04067089036107063, + 0.008589295670390129, + 0.016588082537055016, + 0.018303966149687767, + 0.007765776012092829, + 0.015591686591506004, + 0.008399088867008686, + 0.009619900956749916, + 0.011569592170417309, + 0.008821499533951283, + 0.012681578285992146, + 0.008753374218940735, + 0.006167739164084196, + 0.02103988640010357, + 0.008531749248504639, + 0.007861996069550514, + 0.025964587926864624, + 0.009579559788107872, + 0.007367134094238281, + 0.01327924057841301, + 0.013342476449906826, + 0.02302551455795765, + 0.017015617340803146, + 0.021035684272646904, + 0.019547611474990845, + 0.02259252779185772, + 0.01588582806289196, + 0.007932412438094616, + 0.004979356657713652, + 0.02084549143910408, + 0.00626293895766139, + 0.005349101964384317, + 0.023669682443141937, + 0.00840920489281416, + 0.005545413587242365, + 0.021885907277464867, + 0.012279261834919453, + 0.01933407224714756, + 0.015136930160224438, + 0.012840705923736095, + 0.02274663746356964, + 0.01436289120465517, + 0.017726251855492592, + 0.005670123267918825, + 0.0025980323553085327, + 0.0033941364381462336, + 0.0023134725634008646, + 0.005310892593115568, + 0.003108423901721835, + 0.003084195079281926, + 0.002863285131752491, + 0.006029958371073008, + 0.005431142169982195, + 0.010243686847388744, + 0.005493525881320238, + 0.005537029355764389, + 0.011487068608403206, + 0.007639603223651648, + 0.006755727343261242, + 0.0023597509134560823, + 0.0035542515106499195, + 0.0043680849485099316, + 0.0042465184815227985, + 0.003185006557032466, + 0.004101435653865337, + 0.0021443902514874935, + 0.004224423784762621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.04866454377770424 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016058044508099556 + ], + "opposite_logprob_tokens": [ + -4.141057968139648 + ], + "logprob_reward": -0.016058044508099556, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016058044508099556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15892954170703888, + 0.11221794039011002, + 0.0899467021226883, + 0.057167816907167435, + 0.0774841457605362, + 0.2228955626487732, + 0.0900510922074318, + 0.10503895580768585, + 0.03360747918486595, + 0.05266068875789642 + ], + "credit_map": [ + 0.023550204932689667, + 0.009779267013072968, + 0.0324498675763607, + 0.015350177884101868, + 0.010956864804029465, + 0.03794405609369278, + 0.009105495177209377, + 0.01979362964630127, + 0.01554536446928978, + 0.006853599566966295, + 0.022091342136263847, + 0.009821489453315735, + 0.007096179760992527, + 0.027860503643751144, + 0.007865632884204388, + 0.015083844773471355, + 0.012204714119434357, + 0.006948887836188078, + 0.019661586731672287, + 0.005289864260703325, + 0.004897065926343203, + 0.023938393220305443, + 0.007296586874872446, + 0.009709613397717476, + 0.010101094841957092, + 0.005714112427085638, + 0.009318457916378975, + 0.004639271646738052, + 0.005449547432363033, + 0.009554834105074406, + 0.005784405395388603, + 0.006606101989746094, + 0.00714363157749176, + 0.007775275502353907, + 0.009982914663851261, + 0.009656747803092003, + 0.013007152825593948, + 0.007558673620223999, + 0.007824255153536797, + 0.014535498805344105, + 0.025828436017036438, + 0.012562267482280731, + 0.06025175005197525, + 0.013523014262318611, + 0.014307701960206032, + 0.07145939022302628, + 0.011424549855291843, + 0.013538460247218609, + 0.013105028308928013, + 0.010951654985547066, + 0.011367551051080227, + 0.007431470789015293, + 0.016985537484288216, + 0.00993229541927576, + 0.007437526248395443, + 0.012840038165450096, + 0.02603398635983467, + 0.007952806539833546, + 0.017245376482605934, + 0.009640560485422611, + 0.008330632001161575, + 0.01497956644743681, + 0.009806022979319096, + 0.011050019413232803, + 0.0033914323430508375, + 0.0030314510222524405, + 0.007946779020130634, + 0.0032325617503374815, + 0.002609092742204666, + 0.007354749366641045, + 0.0024655377492308617, + 0.0035758796148002148, + 0.004222188610583544, + 0.005007068160921335, + 0.010222656652331352, + 0.006199195981025696, + 0.005087817553430796, + 0.007175011094659567, + 0.004515565931797028, + 0.010231190361082554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.016058044508099556 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283299058675766 + ], + "opposite_logprob_tokens": [ + -3.7732832431793213 + ], + "logprob_reward": -0.023283299058675766, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283299058675766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0719623789191246, + 0.2020183950662613, + 0.11833950877189636, + 0.0441083163022995, + 0.045723073184490204, + 0.04583710432052612, + 0.05816566199064255, + 0.13640238344669342, + 0.22908291220664978, + 0.04836026579141617 + ], + "credit_map": [ + 0.012433861382305622, + 0.004736632574349642, + 0.013354765251278877, + 0.006356784608215094, + 0.004949901718646288, + 0.01693648472428322, + 0.004337035119533539, + 0.008856916800141335, + 0.03731268644332886, + 0.011602056212723255, + 0.043145157396793365, + 0.015806356444954872, + 0.012463713064789772, + 0.05047141760587692, + 0.01143337320536375, + 0.019783634692430496, + 0.02306286059319973, + 0.006607619114220142, + 0.026291072368621826, + 0.006138203199952841, + 0.00605880469083786, + 0.02588311955332756, + 0.010933788493275642, + 0.013364032842218876, + 0.010625561699271202, + 0.003786938264966011, + 0.007002115715295076, + 0.0035011591389775276, + 0.004237801302224398, + 0.005871416535228491, + 0.00458596320822835, + 0.0044973609037697315, + 0.006441066041588783, + 0.0043584322556853294, + 0.005176216829568148, + 0.007931278087198734, + 0.004277779255062342, + 0.0066754925064742565, + 0.007294701412320137, + 0.003568105399608612, + 0.004382425453513861, + 0.005659488029778004, + 0.006417131517082453, + 0.006500334478914738, + 0.005486877169460058, + 0.005857157055288553, + 0.006963834632188082, + 0.004569856449961662, + 0.007661818061023951, + 0.006164172198623419, + 0.008215859532356262, + 0.00854951236397028, + 0.005403341259807348, + 0.006365151610225439, + 0.006984866224229336, + 0.008820941671729088, + 0.03248615562915802, + 0.010633635334670544, + 0.01888427510857582, + 0.011700756847858429, + 0.013449551537632942, + 0.01479783933609724, + 0.01361974235624075, + 0.02083042450249195, + 0.022603407502174377, + 0.019046753644943237, + 0.049832139164209366, + 0.022704577073454857, + 0.013772504404187202, + 0.06006486713886261, + 0.021883348003029823, + 0.019175320863723755, + 0.0038330773822963238, + 0.0038455103058367968, + 0.008614152669906616, + 0.005736625753343105, + 0.005139129236340523, + 0.005857778247445822, + 0.004559156484901905, + 0.010774831287562847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.023283299058675766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot captured from a slight downward angle focuses squarely on a nautilus with a magnificent spiral shell, resting peacefully on a bed of fine white sand. Its shell, a perfect Fibonacci spiral, is smooth as jade, its pearly, opalescent base punctuated by delicate russet-brown stripes radiating outward from the center. A soft, diffuse light filters down from the crystal-clear azure water, casting a dreamy halo over the smooth shell. Pastel-colored anemones sway gently in the background, blurring slightly. As the camera gazes, the nautilus carefully and slowly unfurls dozens of soft, translucent, opalescent tentacles from its shell. Like newly opened petals, these tentacles tentatively reach outward, their movements gentle and curious. The light softly illuminates the texture of each tentacle, making them look exceptionally delicate. The entire picture, with its tranquil blue and warm white tones, creates an atmosphere of ultimate tranquility, innocence, and a fairy tale.", + "seed": 9024, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_3/index_5614_seed_9024_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026786500681191684, + "vlm_logprob": -0.026786500681191684, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.026786500681191684, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005437228363007307 + ], + "opposite_logprob_tokens": [ + -5.50543737411499 + ], + "logprob_reward": -0.005437228363007307, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005437228363007307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6060104370117188, + 0.059810999780893326, + 0.04253779724240303, + 0.04101920872926712, + 0.032367706298828125, + 0.1365603655576706, + 0.014079286716878414, + 0.042488545179367065, + 0.010503564029932022, + 0.014622143469750881 + ], + "credit_map": [ + 0.09940952062606812, + 0.01722782291471958, + 0.14330708980560303, + 0.0324876569211483, + 0.03339987248182297, + 0.23410604894161224, + 0.015017146244645119, + 0.03105522319674492, + 0.01136027928441763, + 0.006171447690576315, + 0.004184350837022066, + 0.00528049236163497, + 0.005313424859195948, + 0.017654327675700188, + 0.005602322984486818, + 0.004244345705956221, + 0.00417330814525485, + 0.0035864871460944414, + 0.006802867632359266, + 0.003616996807977557, + 0.00311420951038599, + 0.008599456399679184, + 0.005278642289340496, + 0.007365824189037085, + 0.003954810090363026, + 0.005457693245261908, + 0.004552785772830248, + 0.004636203404515982, + 0.005057831294834614, + 0.006990003399550915, + 0.004369998816400766, + 0.005999877583235502, + 0.0032618872355669737, + 0.004973634146153927, + 0.0036526755429804325, + 0.004300391301512718, + 0.0032724374905228615, + 0.005072608590126038, + 0.003688461845740676, + 0.004145604558289051, + 0.01735226809978485, + 0.013239777646958828, + 0.021528948098421097, + 0.013288464397192001, + 0.010240771807730198, + 0.025247367098927498, + 0.016024818643927574, + 0.0196379367262125, + 0.0009430567151866853, + 0.0022543941158801317, + 0.0018713639583438635, + 0.002041804138571024, + 0.0014136562822386622, + 0.0019514437299221754, + 0.0016032221028581262, + 0.0020003432873636484, + 0.0021728130523115396, + 0.005139872431755066, + 0.005988952703773975, + 0.007345952559262514, + 0.004448687192052603, + 0.005618974566459656, + 0.0065118493512272835, + 0.00526143703609705, + 0.0009851889917626977, + 0.001517514348961413, + 0.0014929460594430566, + 0.0012950635282322764, + 0.0013923736987635493, + 0.001564272679388523, + 0.0011363961966708302, + 0.0011198066640645266, + 0.0011362064396962523, + 0.0020561988931149244, + 0.0023625902831554413, + 0.00281178904697299, + 0.0019152660388499498, + 0.0013970256550237536, + 0.0016906328964978456, + 0.0012524323537945747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.005437228363007307 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430787019431591 + ], + "opposite_logprob_tokens": [ + -3.168078660964966 + ], + "logprob_reward": -0.0430787019431591, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430787019431591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29102084040641785, + 0.11469543725252151, + 0.155622661113739, + 0.07250968366861343, + 0.04334238916635513, + 0.1009339690208435, + 0.036866385489702225, + 0.059850893914699554, + 0.06005479395389557, + 0.06510291248559952 + ], + "credit_map": [ + 0.050234608352184296, + 0.01634957268834114, + 0.030117614194750786, + 0.013411360792815685, + 0.01742238737642765, + 0.11209719628095627, + 0.017852457240223885, + 0.033535659313201904, + 0.015415234491229057, + 0.007106904871761799, + 0.015083487145602703, + 0.010762701742351055, + 0.006632054690271616, + 0.024240804836153984, + 0.01704557240009308, + 0.018408676609396935, + 0.014645294286310673, + 0.02841070666909218, + 0.02538941614329815, + 0.023088933899998665, + 0.013734005391597748, + 0.01818634197115898, + 0.018054965883493423, + 0.014113003388047218, + 0.007472096011042595, + 0.011231238953769207, + 0.010853946208953857, + 0.010271222330629826, + 0.0051292250864207745, + 0.011691573075950146, + 0.00734670739620924, + 0.008513674139976501, + 0.0048158797435462475, + 0.006132364273071289, + 0.004852553363889456, + 0.006118680816143751, + 0.0037999253254383802, + 0.007794401608407497, + 0.004185498226433992, + 0.005643085576593876, + 0.011812681332230568, + 0.01145379152148962, + 0.015789981931447983, + 0.010082281194627285, + 0.0075866845436394215, + 0.01828768290579319, + 0.011480871587991714, + 0.014439987950026989, + 0.002804902382194996, + 0.005229929927736521, + 0.005020525772124529, + 0.005888756830245256, + 0.0030975458212196827, + 0.005295535549521446, + 0.003983364440500736, + 0.00554582430049777, + 0.003267707535997033, + 0.007838836871087551, + 0.008541255258023739, + 0.01132690254598856, + 0.006329060066491365, + 0.00711061991751194, + 0.008823070675134659, + 0.006613448262214661, + 0.004759168718010187, + 0.008816683664917946, + 0.008510329760611057, + 0.009536964818835258, + 0.009732416830956936, + 0.00547855393961072, + 0.007472575642168522, + 0.0057480973191559315, + 0.005255591589957476, + 0.010589065961539745, + 0.00915289856493473, + 0.008503216318786144, + 0.009366931393742561, + 0.00820881687104702, + 0.0071436939761042595, + 0.006882700137794018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0430787019431591 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308772087097168 + ], + "opposite_logprob_tokens": [ + -3.1680877208709717 + ], + "logprob_reward": -0.04308772087097168, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308772087097168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5280669927597046, + 0.03955293819308281, + 0.07280568778514862, + 0.04509785771369934, + 0.01913578249514103, + 0.03361310809850693, + 0.02553154155611992, + 0.07160472124814987, + 0.025411324575543404, + 0.1391800343990326 + ], + "credit_map": [ + 0.09035595506429672, + 0.016144340857863426, + 0.11039895564317703, + 0.02377036027610302, + 0.02893153205513954, + 0.21275411546230316, + 0.016257673501968384, + 0.029454080387949944, + 0.006378135643899441, + 0.002835057210177183, + 0.0043582553043961525, + 0.0038469010032713413, + 0.0025052838027477264, + 0.01079234853386879, + 0.003619672730565071, + 0.005217283498495817, + 0.012998904101550579, + 0.004550204146653414, + 0.026127547025680542, + 0.006854617968201637, + 0.0030611578840762377, + 0.00864917691797018, + 0.004953052848577499, + 0.005611022934317589, + 0.00445695361122489, + 0.006724730599671602, + 0.0068016983568668365, + 0.005847558379173279, + 0.0035263982135802507, + 0.006965804845094681, + 0.003726543392986059, + 0.007048170082271099, + 0.0017844238318502903, + 0.0017986971652135253, + 0.002994305221363902, + 0.002112523652613163, + 0.0018758804071694613, + 0.003127706702798605, + 0.0020680108573287725, + 0.0033742343075573444, + 0.0032995801884680986, + 0.00417568301782012, + 0.003762634005397558, + 0.0055376579985022545, + 0.002946448978036642, + 0.005614673253148794, + 0.00429472466930747, + 0.003981704358011484, + 0.0022148245479911566, + 0.002806869801133871, + 0.003653588704764843, + 0.003657541237771511, + 0.0014960896223783493, + 0.004220820032060146, + 0.00270054186694324, + 0.004781267140060663, + 0.0036751413717865944, + 0.008864239789545536, + 0.010249023325741291, + 0.014827524311840534, + 0.0073637524619698524, + 0.0069844676181674, + 0.011634664610028267, + 0.008005911484360695, + 0.0022057299502193928, + 0.00372899929061532, + 0.0033336321357637644, + 0.003196387318894267, + 0.00244887568987906, + 0.0041305553168058395, + 0.0023364019580185413, + 0.0040307422168552876, + 0.008536169305443764, + 0.02352868765592575, + 0.021223776042461395, + 0.021183254197239876, + 0.020372698083519936, + 0.013772796839475632, + 0.016311656683683395, + 0.01425099465996027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.04308772087097168 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636976592242718 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636976592242718, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636976592242718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3931387960910797, + 0.09679588675498962, + 0.12280860543251038, + 0.055697958916425705, + 0.07409592717885971, + 0.08154425770044327, + 0.029517415910959244, + 0.08778131008148193, + 0.029843676835298538, + 0.028776193037629128 + ], + "credit_map": [ + 0.09148027002811432, + 0.02091478556394577, + 0.043356962502002716, + 0.015658700838685036, + 0.020198822021484375, + 0.14888225495815277, + 0.020599937066435814, + 0.03204704821109772, + 0.012088198214769363, + 0.006288633216172457, + 0.009368711151182652, + 0.009173844940960407, + 0.007402882445603609, + 0.025019269436597824, + 0.014348282478749752, + 0.013106068596243858, + 0.016408665105700493, + 0.008433030918240547, + 0.0319170244038105, + 0.009443286806344986, + 0.008230576291680336, + 0.01937130279839039, + 0.01086348108947277, + 0.01814122498035431, + 0.00543783837929368, + 0.006384131498634815, + 0.008427023887634277, + 0.005969063378870487, + 0.005890496075153351, + 0.008928593248128891, + 0.0060600051656365395, + 0.008600810542702675, + 0.007527806330472231, + 0.008946415968239307, + 0.011137058958411217, + 0.008303745649755001, + 0.007987381890416145, + 0.011189338751137257, + 0.007871696725487709, + 0.01113248523324728, + 0.010062861256301403, + 0.0084906667470932, + 0.01340923085808754, + 0.009033622220158577, + 0.006195123307406902, + 0.012605971656739712, + 0.010313259437680244, + 0.011433521285653114, + 0.0027016836684197187, + 0.0024712432641535997, + 0.004935182631015778, + 0.0030473198276013136, + 0.0023297430016100407, + 0.005114100873470306, + 0.0033438571263104677, + 0.0055742873810231686, + 0.004217860288918018, + 0.010012330487370491, + 0.012460276484489441, + 0.016534361988306046, + 0.009345931932330132, + 0.009925038553774357, + 0.014351727440953255, + 0.010933786630630493, + 0.002545360941439867, + 0.00339957932010293, + 0.004254912957549095, + 0.004327869974076748, + 0.002775676781311631, + 0.0049596079625189304, + 0.0031433305703103542, + 0.004437338560819626, + 0.002521080896258354, + 0.004500799346715212, + 0.004205239936709404, + 0.003989306278526783, + 0.0033361611422151327, + 0.0035683412570506334, + 0.0032037163618952036, + 0.0034515492152422667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.008636976592242718 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033691875636577606 + ], + "opposite_logprob_tokens": [ + -3.408691883087158 + ], + "logprob_reward": -0.033691875636577606, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033691875636577606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5390060544013977, + 0.055747538805007935, + 0.0460464246571064, + 0.048934388905763626, + 0.026863235980272293, + 0.10091972351074219, + 0.06223127990961075, + 0.04285676032304764, + 0.03862936049699783, + 0.03876521438360214 + ], + "credit_map": [ + 0.08809888362884521, + 0.015572089701890945, + 0.1304006576538086, + 0.028719378635287285, + 0.02991906926035881, + 0.20717912912368774, + 0.01416612509638071, + 0.024950722232460976, + 0.009539077989757061, + 0.0043176040053367615, + 0.003884978126734495, + 0.004063567146658897, + 0.003808594774454832, + 0.018787594512104988, + 0.004621586762368679, + 0.006724531762301922, + 0.006720912177115679, + 0.0032718630973249674, + 0.012882506474852562, + 0.00552079314365983, + 0.0030565992929041386, + 0.00645051384344697, + 0.0038181187119334936, + 0.00432511605322361, + 0.004854964092373848, + 0.00758800283074379, + 0.006009974516928196, + 0.006825943477451801, + 0.004671887960284948, + 0.006967410445213318, + 0.005429893732070923, + 0.006586307194083929, + 0.0027849285397678614, + 0.004454963840544224, + 0.002966383472084999, + 0.004369266331195831, + 0.0023481971584260464, + 0.003346943063661456, + 0.003428069641813636, + 0.0031644836999475956, + 0.010109847411513329, + 0.010228803381323814, + 0.011149865575134754, + 0.009232421405613422, + 0.009387938305735588, + 0.02304934710264206, + 0.010415722616016865, + 0.017345786094665527, + 0.004236151464283466, + 0.007798260543495417, + 0.008940856903791428, + 0.00914765801280737, + 0.004633757751435041, + 0.008673633448779583, + 0.006593828089535236, + 0.012207133695483208, + 0.003257100936025381, + 0.004688860848546028, + 0.006285384763032198, + 0.007041509263217449, + 0.005229187663644552, + 0.005262988153845072, + 0.005147931165993214, + 0.005943797063082457, + 0.0033296493347734213, + 0.005104099866002798, + 0.0053795999847352505, + 0.00466873636469245, + 0.004262311849743128, + 0.005615576636046171, + 0.003855231450870633, + 0.006414152216166258, + 0.0035496677737683058, + 0.005446021445095539, + 0.0052373576909303665, + 0.004948937799781561, + 0.005912251304835081, + 0.0042012715712189674, + 0.004757706541568041, + 0.004711999092251062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.033691875636577606 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot captured from a slight downward angle focuses squarely on a nautilus with a magnificent spiral shell, resting peacefully on a bed of fine white sand. Its shell, a perfect Fibonacci spiral, is smooth as jade, its pearly, opalescent base punctuated by delicate russet-brown stripes radiating outward from the center. A soft, diffuse light filters down from the crystal-clear azure water, casting a dreamy halo over the smooth shell. Pastel-colored anemones sway gently in the background, blurring slightly. As the camera gazes, the nautilus carefully and slowly unfurls dozens of soft, translucent, opalescent tentacles from its shell. Like newly opened petals, these tentacles tentatively reach outward, their movements gentle and curious. The light softly illuminates the texture of each tentacle, making them look exceptionally delicate. The entire picture, with its tranquil blue and warm white tones, creates an atmosphere of ultimate tranquility, innocence, and a fairy tale.", + "seed": 9025, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_3/index_5614_seed_9025_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03385668224655092, + "vlm_logprob": -0.03385668224655092, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.03385668224655092, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036601710598915815 + ], + "opposite_logprob_tokens": [ + -6.253660202026367 + ], + "logprob_reward": -0.0036601710598915815, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0036601710598915815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13252569735050201, + 0.043780941516160965, + 0.10297714918851852, + 0.07874440401792526, + 0.05174841359257698, + 0.044966015964746475, + 0.04807078838348389, + 0.14446166157722473, + 0.28370288014411926, + 0.06902205944061279 + ], + "credit_map": [ + 0.011104521341621876, + 0.01092191506177187, + 0.02518426813185215, + 0.023229530081152916, + 0.013411620631814003, + 0.009877072647213936, + 0.027161601930856705, + 0.01163516566157341, + 0.004947034642100334, + 0.004295557737350464, + 0.0054938760586082935, + 0.006879062857478857, + 0.005310597829520702, + 0.0037838879507035017, + 0.006594677455723286, + 0.006476248148828745, + 0.016370302066206932, + 0.007744804956018925, + 0.009776447899639606, + 0.014003277756273746, + 0.009949478320777416, + 0.010179351083934307, + 0.007947186939418316, + 0.027006294578313828, + 0.010051283054053783, + 0.00786014273762703, + 0.007582741789519787, + 0.0091929966583848, + 0.006697975564748049, + 0.006950566079467535, + 0.00745426258072257, + 0.022954436019062996, + 0.0055595687590539455, + 0.005464032292366028, + 0.009333117865025997, + 0.007776715327054262, + 0.005507394205778837, + 0.004422602243721485, + 0.006724911276251078, + 0.0069600725546479225, + 0.005264626815915108, + 0.004522768780589104, + 0.007132644299417734, + 0.004613168071955442, + 0.004320163745433092, + 0.005276372656226158, + 0.00617379043251276, + 0.007662477903068066, + 0.006813959684222937, + 0.00528640765696764, + 0.009587575681507587, + 0.0052755726501345634, + 0.003722051391378045, + 0.004708583932369947, + 0.00491494731977582, + 0.007761688902974129, + 0.018985386937856674, + 0.014755556359887123, + 0.027453619986772537, + 0.015034507028758526, + 0.01506244670599699, + 0.01121916901320219, + 0.017759935930371284, + 0.024191034957766533, + 0.03275240957736969, + 0.02902192808687687, + 0.03589912876486778, + 0.027054438367486, + 0.032895173877477646, + 0.020082352682948112, + 0.05145112797617912, + 0.05454633757472038, + 0.00776288565248251, + 0.010575591586530209, + 0.007370578590780497, + 0.011020155623555183, + 0.007678356487303972, + 0.007171564269810915, + 0.00877847708761692, + 0.008664448745548725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0036601710598915815 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020600058138370514 + ], + "opposite_logprob_tokens": [ + -3.8956000804901123 + ], + "logprob_reward": -0.020600058138370514, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020600058138370514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09200888127088547, + 0.039797160774469376, + 0.07756061106920242, + 0.0696839988231659, + 0.04628867283463478, + 0.1060158833861351, + 0.0465620718896389, + 0.03726942092180252, + 0.39554014801979065, + 0.0892731249332428 + ], + "credit_map": [ + 0.009068585000932217, + 0.007305072154849768, + 0.013337460346519947, + 0.019057124853134155, + 0.009918449446558952, + 0.007529228460043669, + 0.014715597964823246, + 0.011077363044023514, + 0.004658158402889967, + 0.004189542960375547, + 0.005141969304531813, + 0.006901714950799942, + 0.004636503756046295, + 0.003685606177896261, + 0.005115908570587635, + 0.005467756185680628, + 0.010607032105326653, + 0.007830439135432243, + 0.010564378462731838, + 0.009391572326421738, + 0.007641421630978584, + 0.007355170790106058, + 0.006781913805752993, + 0.017388686537742615, + 0.0056609525345265865, + 0.008015504106879234, + 0.00619086530059576, + 0.014271307736635208, + 0.0063497028313577175, + 0.005787446163594723, + 0.009695210494101048, + 0.01371301244944334, + 0.004826703574508429, + 0.006345393601804972, + 0.005489033181220293, + 0.0051755765452980995, + 0.004345469176769257, + 0.00382237508893013, + 0.007332064211368561, + 0.008952056057751179, + 0.011248067952692509, + 0.013250754214823246, + 0.010635134764015675, + 0.01769644021987915, + 0.010340524837374687, + 0.01059239637106657, + 0.018081586807966232, + 0.014170984737575054, + 0.006831142585724592, + 0.006372287403792143, + 0.0068632024340331554, + 0.005368697922676802, + 0.0062878429889678955, + 0.005879658740013838, + 0.0045005809515714645, + 0.004458658862859011, + 0.00502375653013587, + 0.004682037979364395, + 0.006289161741733551, + 0.0031845388002693653, + 0.004237073939293623, + 0.0035389058757573366, + 0.004348594229668379, + 0.00596535112708807, + 0.04061169922351837, + 0.04160480201244354, + 0.05242118611931801, + 0.040626153349876404, + 0.0452711284160614, + 0.02595500461757183, + 0.06598550081253052, + 0.08306468278169632, + 0.009268084540963173, + 0.009470499120652676, + 0.011697688139975071, + 0.014992153272032738, + 0.009428274817764759, + 0.0060449810698628426, + 0.013389001600444317, + 0.01498244795948267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.020600058138370514 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026358850300312042 + ], + "opposite_logprob_tokens": [ + -3.6513588428497314 + ], + "logprob_reward": -0.026358850300312042, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358850300312042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08757854998111725, + 0.05077112093567848, + 0.1798323690891266, + 0.09496606141328812, + 0.03893397003412247, + 0.06653591245412827, + 0.08946283906698227, + 0.05567163974046707, + 0.18668101727962494, + 0.14956648647785187 + ], + "credit_map": [ + 0.009069886989891529, + 0.007851019501686096, + 0.013887091539800167, + 0.01953393965959549, + 0.007009741384536028, + 0.006753494963049889, + 0.01339738816022873, + 0.010075985454022884, + 0.004284445662051439, + 0.0047167446464300156, + 0.006172205321490765, + 0.014197401702404022, + 0.0065756309777498245, + 0.0037570386193692684, + 0.006269885692745447, + 0.004797767382115126, + 0.029698720201849937, + 0.013075454160571098, + 0.02143012546002865, + 0.016930531710386276, + 0.014562883414328098, + 0.018121523782610893, + 0.016706902533769608, + 0.0493062362074852, + 0.010877640917897224, + 0.008945590816438198, + 0.009729556739330292, + 0.013245796784758568, + 0.007203353568911552, + 0.008261974900960922, + 0.009630359709262848, + 0.02707178331911564, + 0.003611735999584198, + 0.004877800587564707, + 0.005134952254593372, + 0.005654210224747658, + 0.0032423699740320444, + 0.0033216315787285566, + 0.005709752906113863, + 0.007381515111774206, + 0.00667062820866704, + 0.007997202686965466, + 0.010292607359588146, + 0.009390621446073055, + 0.005816188175231218, + 0.006066839676350355, + 0.00931583158671856, + 0.010985998436808586, + 0.010564175434410572, + 0.013500571250915527, + 0.009084315039217472, + 0.014716054312884808, + 0.011519519612193108, + 0.009400052949786186, + 0.012899037450551987, + 0.007779110688716173, + 0.005621673073619604, + 0.007160541135817766, + 0.009379450231790543, + 0.00566994259133935, + 0.00692143477499485, + 0.00428227661177516, + 0.0066887750290334225, + 0.009947548620402813, + 0.014208200387656689, + 0.027893682941794395, + 0.01688281260430813, + 0.0421590693295002, + 0.02213365212082863, + 0.012880118563771248, + 0.03036867268383503, + 0.020154796540737152, + 0.013361153192818165, + 0.02801266498863697, + 0.0135806230828166, + 0.03529123216867447, + 0.017521552741527557, + 0.010229089297354221, + 0.017953725531697273, + 0.013616448268294334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.026358850300312042 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001136988401413 + ], + "opposite_logprob_tokens": [ + -2.6950113773345947 + ], + "logprob_reward": -0.07001136988401413, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001136988401413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0954333245754242, + 0.05153518542647362, + 0.023286009207367897, + 0.015295568853616714, + 0.012353453785181046, + 0.021953489631414413, + 0.024590400978922844, + 0.02560119330883026, + 0.6972976922988892, + 0.03265365585684776 + ], + "credit_map": [ + 0.012111686170101166, + 0.009342361241579056, + 0.020214522257447243, + 0.016969412565231323, + 0.007828956469893456, + 0.007509557995945215, + 0.010054230690002441, + 0.011402608826756477, + 0.009127320721745491, + 0.004328275565057993, + 0.006770854815840721, + 0.005354207009077072, + 0.003527385648339987, + 0.006465213838964701, + 0.004862100351601839, + 0.011099829338490963, + 0.0038569937460124493, + 0.0014617907581850886, + 0.0028113326989114285, + 0.0032456826884299517, + 0.0015394347719848156, + 0.002346271649003029, + 0.0024840885307639837, + 0.0055404165759682655, + 0.0019601702224463224, + 0.002101817401126027, + 0.0015436934772878885, + 0.0023270873352885246, + 0.002380563411861658, + 0.0015612633433192968, + 0.0014498330419883132, + 0.001971140969544649, + 0.0012390578631311655, + 0.0013439350295811892, + 0.0015400038100779057, + 0.0017733616987243295, + 0.0015078160213306546, + 0.0010459618642926216, + 0.0019422568148002028, + 0.001961062429472804, + 0.0024420188274234533, + 0.002257870975881815, + 0.0029177635442465544, + 0.002035793848335743, + 0.0017363170627504587, + 0.0020243492908775806, + 0.003251006593927741, + 0.005288372747600079, + 0.0036568078212440014, + 0.0029851829167455435, + 0.0037083090282976627, + 0.002576488768681884, + 0.003050194587558508, + 0.002612314186990261, + 0.002738968236371875, + 0.0032621396239846945, + 0.0031329754274338484, + 0.0030198432505130768, + 0.004574112128466368, + 0.0023718865122646093, + 0.0021106635686010122, + 0.002501304494217038, + 0.002538522006943822, + 0.005351889878511429, + 0.06890930235385895, + 0.07583796232938766, + 0.09340906888246536, + 0.0713762417435646, + 0.08007136732339859, + 0.045127417892217636, + 0.10970263928174973, + 0.1528637558221817, + 0.004027417860925198, + 0.004893959034234285, + 0.0039506214670836926, + 0.00432194396853447, + 0.004231312312185764, + 0.002566727576777339, + 0.003009025938808918, + 0.005652650725096464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.07001136988401413 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865296185016632 + ], + "opposite_logprob_tokens": [ + -3.0486528873443604 + ], + "logprob_reward": -0.04865296185016632, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865296185016632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10344229638576508, + 0.0534403920173645, + 0.09457533061504364, + 0.11962644010782242, + 0.07012785971164703, + 0.09597474336624146, + 0.09024755656719208, + 0.0676041916012764, + 0.17511484026908875, + 0.12984634935855865 + ], + "credit_map": [ + 0.011941308155655861, + 0.0075587560422718525, + 0.015614164061844349, + 0.016780463978648186, + 0.011684030294418335, + 0.009694840759038925, + 0.021030638366937637, + 0.009138092398643494, + 0.007284378632903099, + 0.004908120259642601, + 0.006979363039135933, + 0.009947240352630615, + 0.004566829651594162, + 0.00529133528470993, + 0.007101785391569138, + 0.007361340802162886, + 0.013262540102005005, + 0.0072797429747879505, + 0.010115515440702438, + 0.008986703120172024, + 0.01185593567788601, + 0.00927478726953268, + 0.009070239961147308, + 0.02472987212240696, + 0.0142146535217762, + 0.011226946488022804, + 0.012327983975410461, + 0.013510829769074917, + 0.00883050449192524, + 0.010868220590054989, + 0.012744029983878136, + 0.03590327128767967, + 0.006061253137886524, + 0.0063347043469548225, + 0.009194167330861092, + 0.007621405646204948, + 0.005792585201561451, + 0.005607464350759983, + 0.010725523345172405, + 0.01879076287150383, + 0.010436427779495716, + 0.012360623106360435, + 0.013043753802776337, + 0.012567012570798397, + 0.00912033673375845, + 0.00820363499224186, + 0.012134073302149773, + 0.018108880147337914, + 0.006899858359247446, + 0.011861322447657585, + 0.010256986133754253, + 0.013143910095095634, + 0.011134402826428413, + 0.007988271303474903, + 0.017130741849541664, + 0.011832066811621189, + 0.008017637766897678, + 0.010297535918653011, + 0.005544804036617279, + 0.01061051432043314, + 0.009558018296957016, + 0.006155914627015591, + 0.010044466704130173, + 0.007375300861895084, + 0.024614110589027405, + 0.0215244609862566, + 0.021147212013602257, + 0.01870611682534218, + 0.017352543771266937, + 0.0157573614269495, + 0.028709130361676216, + 0.027303891256451607, + 0.012963849119842052, + 0.020613474771380424, + 0.01623530313372612, + 0.021047130227088928, + 0.01829722709953785, + 0.0083383247256279, + 0.012660647742450237, + 0.019690386950969696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.04865296185016632 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot captured from a slight downward angle focuses squarely on a nautilus with a magnificent spiral shell, resting peacefully on a bed of fine white sand. Its shell, a perfect Fibonacci spiral, is smooth as jade, its pearly, opalescent base punctuated by delicate russet-brown stripes radiating outward from the center. A soft, diffuse light filters down from the crystal-clear azure water, casting a dreamy halo over the smooth shell. Pastel-colored anemones sway gently in the background, blurring slightly. As the camera gazes, the nautilus carefully and slowly unfurls dozens of soft, translucent, opalescent tentacles from its shell. Like newly opened petals, these tentacles tentatively reach outward, their movements gentle and curious. The light softly illuminates the texture of each tentacle, making them look exceptionally delicate. The entire picture, with its tranquil blue and warm white tones, creates an atmosphere of ultimate tranquility, innocence, and a fairy tale.", + "seed": 9026, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_3/index_5614_seed_9026_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.046880670730024575, + "vlm_logprob": -0.046880670730024575, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.046880670730024575, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0047646318562328815 + ], + "opposite_logprob_tokens": [ + -5.879764556884766 + ], + "logprob_reward": -0.0047646318562328815, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0047646318562328815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08722231537103653, + 0.38398095965385437, + 0.05308184400200844, + 0.055932819843292236, + 0.02578897215425968, + 0.05331794545054436, + 0.0608043447136879, + 0.14493012428283691, + 0.07951594144105911, + 0.055424708873033524 + ], + "credit_map": [ + 0.014833522960543633, + 0.00899883359670639, + 0.011944709345698357, + 0.011699012480676174, + 0.010476695373654366, + 0.00840379111468792, + 0.010946069844067097, + 0.009919682517647743, + 0.03192882984876633, + 0.022374378517270088, + 0.12451336532831192, + 0.029073545709252357, + 0.018174365162849426, + 0.12720920145511627, + 0.011671568267047405, + 0.01903570257127285, + 0.006442501675337553, + 0.0052297115325927734, + 0.007051271852105856, + 0.008066045120358467, + 0.005870752036571503, + 0.005121542606502771, + 0.00753143522888422, + 0.007768583949655294, + 0.00725941825658083, + 0.0030088231433182955, + 0.013367582112550735, + 0.0070237210020422935, + 0.004474039189517498, + 0.01402172539383173, + 0.0031047009397298098, + 0.0036728105042129755, + 0.004566347226500511, + 0.003927342593669891, + 0.003144810674712062, + 0.003302466357126832, + 0.003444239730015397, + 0.0021018958650529385, + 0.002495442284271121, + 0.002806425327435136, + 0.004747486673295498, + 0.007816900499165058, + 0.007303043268620968, + 0.0057108765468001366, + 0.008625131100416183, + 0.007463164161890745, + 0.005027495324611664, + 0.006623850204050541, + 0.008278224617242813, + 0.00776999955996871, + 0.006140587385743856, + 0.009054983966052532, + 0.008768695406615734, + 0.005427374038845301, + 0.007425901014357805, + 0.007938580587506294, + 0.02674541063606739, + 0.01332776341587305, + 0.015498570166528225, + 0.010107733309268951, + 0.011309375055134296, + 0.017944680526852608, + 0.015359321609139442, + 0.03463725000619888, + 0.012960466556251049, + 0.008079019375145435, + 0.01187506690621376, + 0.007739542052149773, + 0.008860606700181961, + 0.011308195069432259, + 0.007039831485599279, + 0.011653205379843712, + 0.00955674983561039, + 0.005723870359361172, + 0.0062512243166565895, + 0.005166405346244574, + 0.005734230391681194, + 0.008279547095298767, + 0.006467326544225216, + 0.008245358243584633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.0047646318562328815 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811829909682274 + ], + "opposite_logprob_tokens": [ + -3.288118362426758 + ], + "logprob_reward": -0.03811829909682274, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811829909682274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09627598524093628, + 0.2568609416484833, + 0.06934715062379837, + 0.05080332234501839, + 0.043757081031799316, + 0.2101835310459137, + 0.08379413932561874, + 0.08236418664455414, + 0.059500571340322495, + 0.04711310192942619 + ], + "credit_map": [ + 0.020504185929894447, + 0.008628358133137226, + 0.014919525943696499, + 0.010050194337964058, + 0.010311659425497055, + 0.011082571931183338, + 0.010339013300836086, + 0.01044047437608242, + 0.02144017070531845, + 0.01565900258719921, + 0.08281952142715454, + 0.01804240606725216, + 0.013715722598135471, + 0.08475751429796219, + 0.007924572564661503, + 0.012502034194767475, + 0.009024735540151596, + 0.008230656385421753, + 0.009241449646651745, + 0.009139304049313068, + 0.007237506564706564, + 0.011414785869419575, + 0.006808618549257517, + 0.008250094018876553, + 0.006138649769127369, + 0.0036756033077836037, + 0.010970563627779484, + 0.00599317392334342, + 0.004373606294393539, + 0.011858026497066021, + 0.0035310874227434397, + 0.004262611269950867, + 0.0074129002168774605, + 0.005274856463074684, + 0.004582877736538649, + 0.0058296737261116505, + 0.0059320214204490185, + 0.005771438591182232, + 0.004468862898647785, + 0.004484451375901699, + 0.02742619812488556, + 0.027579758316278458, + 0.02318558655679226, + 0.019530270248651505, + 0.04867716133594513, + 0.024288397282361984, + 0.01940367929637432, + 0.020092468708753586, + 0.0146997245028615, + 0.009610516019165516, + 0.008473354391753674, + 0.010751085355877876, + 0.012821422889828682, + 0.0083944546058774, + 0.010915453545749187, + 0.008128127083182335, + 0.016066238284111023, + 0.009928030893206596, + 0.009520347230136395, + 0.009576072916388512, + 0.007064162287861109, + 0.009392743930220604, + 0.009094282053411007, + 0.011722302064299583, + 0.007092693354934454, + 0.008432600647211075, + 0.006256056018173695, + 0.005616882815957069, + 0.008309843949973583, + 0.007316237781196833, + 0.007574049755930901, + 0.008902203291654587, + 0.006464105099439621, + 0.005336049944162369, + 0.0051230136305093765, + 0.0054257819429039955, + 0.005086573772132397, + 0.006850740406662226, + 0.005459281615912914, + 0.007367553189396858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.03811829909682274 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12708091735839844 + ], + "opposite_logprob_tokens": [ + -2.1270809173583984 + ], + "logprob_reward": -0.12708091735839844, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12708091735839844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0793304368853569, + 0.26107725501060486, + 0.07205567508935928, + 0.06784254312515259, + 0.09579344838857651, + 0.14589525759220123, + 0.03239660710096359, + 0.09052600711584091, + 0.09530887752771378, + 0.059773899614810944 + ], + "credit_map": [ + 0.011640054173767567, + 0.01068583782762289, + 0.008769332431256771, + 0.011409916914999485, + 0.010567554272711277, + 0.008136900141835213, + 0.00988343358039856, + 0.008237401954829693, + 0.022610891610383987, + 0.015939226374030113, + 0.07836536318063736, + 0.0196085125207901, + 0.01571660116314888, + 0.08130820840597153, + 0.011517712846398354, + 0.016010748222470284, + 0.008457870222628117, + 0.009666378609836102, + 0.009340316988527775, + 0.00839211791753769, + 0.008734668605029583, + 0.01102885976433754, + 0.007481477688997984, + 0.008953983895480633, + 0.0075675989501178265, + 0.009137935005128384, + 0.009195568040013313, + 0.011029645800590515, + 0.008249728940427303, + 0.010487932711839676, + 0.006791879888623953, + 0.005382247734814882, + 0.010806909762322903, + 0.015728870406746864, + 0.01131688617169857, + 0.01633838377892971, + 0.014261914417147636, + 0.007944495417177677, + 0.009882904589176178, + 0.009513082914054394, + 0.015914302319288254, + 0.02081778645515442, + 0.017621392384171486, + 0.012672634795308113, + 0.030832309275865555, + 0.019190022721886635, + 0.014115877449512482, + 0.014730929397046566, + 0.007551457732915878, + 0.002687335480004549, + 0.004210075829178095, + 0.00357245746999979, + 0.0030078794807195663, + 0.004537670407444239, + 0.0030557108111679554, + 0.0037740194238722324, + 0.012942072935402393, + 0.008438188582658768, + 0.012942695990204811, + 0.00857524573802948, + 0.007890013977885246, + 0.010897155851125717, + 0.009585458785295486, + 0.019255179911851883, + 0.010664481669664383, + 0.011690056882798672, + 0.012257153168320656, + 0.0109367985278368, + 0.01253911480307579, + 0.01000231970101595, + 0.013236667960882187, + 0.01398228295147419, + 0.007360327523201704, + 0.006832779850810766, + 0.007209434639662504, + 0.005968880839645863, + 0.00698630977421999, + 0.0066229673102498055, + 0.006819233298301697, + 0.011973964981734753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.12708091735839844 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809110447764397 + ], + "opposite_logprob_tokens": [ + -3.2880911827087402 + ], + "logprob_reward": -0.03809110447764397, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809110447764397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30284053087234497, + 0.11260498315095901, + 0.03361205384135246, + 0.037311654537916183, + 0.04698970913887024, + 0.1405114084482193, + 0.0282034520059824, + 0.20653985440731049, + 0.021720098331570625, + 0.06966623663902283 + ], + "credit_map": [ + 0.06384848058223724, + 0.02814676985144615, + 0.04343106225132942, + 0.031472571194171906, + 0.03430088609457016, + 0.041678499430418015, + 0.030733345076441765, + 0.0292289387434721, + 0.009382410906255245, + 0.007248894311487675, + 0.03460729867219925, + 0.00864897109568119, + 0.006417489610612392, + 0.035554438829422, + 0.00486968457698822, + 0.005875794216990471, + 0.005210540723055601, + 0.0035216298419982195, + 0.005493586417287588, + 0.003988304175436497, + 0.003157964674755931, + 0.0033333851024508476, + 0.004115066956728697, + 0.004791577346622944, + 0.007584804203361273, + 0.0025833582039922476, + 0.006827397271990776, + 0.003920938353985548, + 0.0027028368785977364, + 0.007687215693295002, + 0.00281204329803586, + 0.003193061565980315, + 0.010037858039140701, + 0.006726425141096115, + 0.004666969180107117, + 0.006960242986679077, + 0.005551755893975496, + 0.005466068163514137, + 0.0040547773241996765, + 0.003525617066770792, + 0.013478314504027367, + 0.021759148687124252, + 0.018857743591070175, + 0.013879534788429737, + 0.02430335432291031, + 0.01936284825205803, + 0.014014851301908493, + 0.01485561951994896, + 0.005253934301435947, + 0.003014947287738323, + 0.0026466732379049063, + 0.002987898886203766, + 0.0038236200343817472, + 0.004156076814979315, + 0.0027721161022782326, + 0.00354818650521338, + 0.04068395495414734, + 0.019555900245904922, + 0.023161200806498528, + 0.017297513782978058, + 0.017894482240080833, + 0.022978952154517174, + 0.02469043619930744, + 0.04027743265032768, + 0.003215866396203637, + 0.0023514560889452696, + 0.002837799722328782, + 0.002706710249185562, + 0.0021015219390392303, + 0.002520288573578, + 0.0026555079966783524, + 0.003330948995426297, + 0.009799947962164879, + 0.009166045114398003, + 0.007258270401507616, + 0.007202512118965387, + 0.007561969570815563, + 0.009098920039832592, + 0.00879165343940258, + 0.010786923579871655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.03809110447764397 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a nautilus with a spiral shell resting on white sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nautilus shell display a smooth jade-like texture with pearly opalescent tones and russet-brown stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do dozens of soft, translucent, opalescent tentacles slowly unfurl from the nautilus shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are pastel-colored anemones visible swaying gently in the blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a static shot captured from a slight downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348400861024857 + ], + "opposite_logprob_tokens": [ + -3.651348352432251 + ], + "logprob_reward": -0.026348400861024857, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348400861024857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07279475033283234, + 0.5282083749771118, + 0.039092376828193665, + 0.055655818432569504, + 0.04202713072299957, + 0.11967535316944122, + 0.022610967978835106, + 0.07102008908987045, + 0.031069226562976837, + 0.0178458783775568 + ], + "credit_map": [ + 0.009601264260709286, + 0.00644348282366991, + 0.011200055480003357, + 0.00985770858824253, + 0.009056112729012966, + 0.007706536911427975, + 0.009498241357505322, + 0.00943135004490614, + 0.04318910092115402, + 0.030246799811720848, + 0.17031633853912354, + 0.041455142199993134, + 0.024446820840239525, + 0.17638534307479858, + 0.016957784071564674, + 0.02521107718348503, + 0.006633153650909662, + 0.004615355283021927, + 0.005006202030926943, + 0.004140717443078756, + 0.004513353109359741, + 0.004477882292121649, + 0.004995493218302727, + 0.004710222128778696, + 0.007600190583616495, + 0.0055051944218575954, + 0.009574567899107933, + 0.008878059685230255, + 0.004065896384418011, + 0.011676885187625885, + 0.004456740338355303, + 0.0038982874248176813, + 0.005893681198358536, + 0.0054232096299529076, + 0.0056021991185843945, + 0.007494928315281868, + 0.006639712490141392, + 0.003585833590477705, + 0.003880466101691127, + 0.0035071030724793673, + 0.014278147369623184, + 0.015095138922333717, + 0.014939608052372932, + 0.014048331417143345, + 0.022662581875920296, + 0.015115982852876186, + 0.011864520609378815, + 0.011671042069792747, + 0.003215398406609893, + 0.002487410791218281, + 0.0024116388522088528, + 0.00324231362901628, + 0.0037321995478123426, + 0.0024743955582380295, + 0.0022344605531543493, + 0.0028131529688835144, + 0.012097512371838093, + 0.006936443038284779, + 0.008160104975104332, + 0.00889952015131712, + 0.005212358199059963, + 0.009363485500216484, + 0.007272464223206043, + 0.013078204356133938, + 0.004988808184862137, + 0.004170672036707401, + 0.0031863495241850615, + 0.0032587565947324038, + 0.0036245123483240604, + 0.0038318471051752567, + 0.004531269893050194, + 0.003477012738585472, + 0.002685724990442395, + 0.00181991804856807, + 0.002291633514687419, + 0.0015648624394088984, + 0.001772603951394558, + 0.002892599208280444, + 0.0016373249236494303, + 0.0031812128145247698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.026348400861024857 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_4/index_tensor([3843, 3843])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_4/index_tensor([3843, 3843])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..97d28cff20224cf0eea4fabb5522e91904d027f3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_4/index_tensor([3843, 3843])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, shot from a frontal angle, first focuses on a static e-book cover displayed on a digital screen, imbued with a cyberpunk aesthetic. The background depicts a futuristic nightscape drenched in icy rain. Towering skyscrapers gleam with giant holographic advertisements, and the wet streets reflect magenta and cyan neon lights. Suddenly, the previously static female figure at the center of the cover—with asymmetrical bright pink hair, a face etched with glowing geometric circuitry, and clad in a tight-fitting black metallic tactical suit with a carbon fiber feel—suddenly acquires a sense of three-dimensional volume. In an explosive, dynamic pose, she bursts forward from the two-dimensional plane, seemingly piercing the screen itself. The entire movement, captured with incredibly smooth, high-frame-rate cinematography, appears incredibly fluent. Leaning forward, one arm extended, fingers spread, the other grips the grip of a high-tech pistol. Several modules, each with a faint, blinking indicator light, hang from a belt at her waist. As she bursts out of the frame, her sharp eyes stare forward, her face a look of resolute concentration. The movement of her flying hair and the folds of her clothing are clearly visible. The previously static raindrops and neon lights behind her seem to be torn apart by her movements, creating a slow-motion effect and a slight spatial distortion. The entire scene, in its cool tones of light and shadow, creates a unique atmosphere that blends futuristic, tense, and surreal aesthetics, imbued with a breathtaking sense of reality and visual impact.", + "seed": 4179, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_4/index_3843_seed_4179_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04181237714365125, + "vlm_logprob": -0.04181237714365125, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.04181237714365125, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054608847945928574 + ], + "opposite_logprob_tokens": [ + -3.0546088218688965 + ], + "logprob_reward": -0.054608847945928574, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.054608847945928574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13148175179958344, + 0.11944650858640671, + 0.25213131308555603, + 0.10486333817243576, + 0.0673801377415657, + 0.07256776839494705, + 0.029695045202970505, + 0.09405813366174698, + 0.03231842815876007, + 0.09605757892131805 + ], + "credit_map": [ + 0.010042158886790276, + 0.022179335355758667, + 0.015616193413734436, + 0.0191347673535347, + 0.021737724542617798, + 0.013654678128659725, + 0.014219268225133419, + 0.014897623099386692, + 0.00785766914486885, + 0.019887659698724747, + 0.012228008359670639, + 0.020524300634860992, + 0.014716999605298042, + 0.01579330675303936, + 0.013757645152509212, + 0.014680912718176842, + 0.0243998970836401, + 0.030442124232649803, + 0.0513654425740242, + 0.02844865433871746, + 0.027454311028122902, + 0.03435017168521881, + 0.02166079729795456, + 0.03400993347167969, + 0.009934784844517708, + 0.014250671491026878, + 0.010388301685452461, + 0.011400589719414711, + 0.011823284439742565, + 0.023020291700959206, + 0.010234002955257893, + 0.013811413198709488, + 0.012662232853472233, + 0.007917365990579128, + 0.009157819673418999, + 0.006609006784856319, + 0.005797977093607187, + 0.010237151756882668, + 0.0059408750385046005, + 0.009057707153260708, + 0.01090301014482975, + 0.006747991777956486, + 0.006889109965413809, + 0.006658567115664482, + 0.0091932974755764, + 0.015951966866850853, + 0.006304075941443443, + 0.009919748641550541, + 0.00431815953925252, + 0.0017693240661174059, + 0.009387842379510403, + 0.0031091556884348392, + 0.002159570110961795, + 0.0030223021749407053, + 0.002046933164820075, + 0.003881757380440831, + 0.015249330550432205, + 0.006960525643080473, + 0.028781212866306305, + 0.00832111481577158, + 0.010311276651918888, + 0.007666149642318487, + 0.006206690799444914, + 0.01056183222681284, + 0.005122434813529253, + 0.002830778481438756, + 0.009802563115954399, + 0.002628939924761653, + 0.0037638619542121887, + 0.003740353509783745, + 0.0014880483504384756, + 0.0029414459131658077, + 0.015538670122623444, + 0.007612223736941814, + 0.032658230513334274, + 0.007567449938505888, + 0.006676031742244959, + 0.011750039644539356, + 0.006232698913663626, + 0.00802223477512598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.054608847945928574 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007643500808626413 + ], + "opposite_logprob_tokens": [ + -4.882643699645996 + ], + "logprob_reward": -0.007643500808626413, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007643500808626413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2630831003189087, + 0.11085553467273712, + 0.15138980746269226, + 0.18801866471767426, + 0.050210632383823395, + 0.034272726625204086, + 0.03120756521821022, + 0.06583694368600845, + 0.08130960911512375, + 0.02381538227200508 + ], + "credit_map": [ + 0.02060795947909355, + 0.035325974225997925, + 0.02488589845597744, + 0.025212761014699936, + 0.03324136510491371, + 0.04613441973924637, + 0.05118498578667641, + 0.026489727199077606, + 0.008122115395963192, + 0.019199661910533905, + 0.010896730236709118, + 0.012037049978971481, + 0.01428281981498003, + 0.021704863756895065, + 0.011440064758062363, + 0.01317222137004137, + 0.01595030538737774, + 0.01602044701576233, + 0.036314964294433594, + 0.01615646854043007, + 0.010355724021792412, + 0.023372232913970947, + 0.013600094243884087, + 0.01961955986917019, + 0.019916940480470657, + 0.009426882490515709, + 0.036212604492902756, + 0.013004115782678127, + 0.01131551992148161, + 0.05455522611737251, + 0.013227489776909351, + 0.030359875410795212, + 0.0059944866225123405, + 0.004147084895521402, + 0.006339369807392359, + 0.0036202252376824617, + 0.004111893009394407, + 0.01451075728982687, + 0.004437061958014965, + 0.007049757521599531, + 0.0048899464309215546, + 0.003643615869805217, + 0.005259965546429157, + 0.002844941569492221, + 0.00419076532125473, + 0.005274971015751362, + 0.0029585259035229683, + 0.005209996830672026, + 0.005709810182452202, + 0.001657615415751934, + 0.009371930733323097, + 0.002849288284778595, + 0.0024846019223332405, + 0.0027020112611353397, + 0.0022797423880547285, + 0.004152566194534302, + 0.009464710019528866, + 0.008685927838087082, + 0.009841548278927803, + 0.007157949265092611, + 0.010874846018850803, + 0.006350183859467506, + 0.005606912076473236, + 0.007854866795241833, + 0.014672957360744476, + 0.005473197903484106, + 0.0274900384247303, + 0.007380111142992973, + 0.005267293192446232, + 0.006676641758531332, + 0.005483274348080158, + 0.008866101503372192, + 0.003844105638563633, + 0.0015201091300696135, + 0.008634079247713089, + 0.0014861281961202621, + 0.002017112448811531, + 0.002445732243359089, + 0.0014042954426258802, + 0.002463818760588765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.007643500808626413 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096094518899918 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004096094518899918, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096094518899918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10764266550540924, + 0.16747845709323883, + 0.07363779097795486, + 0.03505329415202141, + 0.03466308116912842, + 0.0499597005546093, + 0.24655453860759735, + 0.09645213931798935, + 0.12972745299339294, + 0.058830901980400085 + ], + "credit_map": [ + 0.007452318910509348, + 0.018196284770965576, + 0.01636701263487339, + 0.010353808291256428, + 0.014277021400630474, + 0.012890397571027279, + 0.016758812591433525, + 0.011347007937729359, + 0.014140399172902107, + 0.03024611435830593, + 0.03182816505432129, + 0.022370032966136932, + 0.017207518219947815, + 0.020367028191685677, + 0.013584630563855171, + 0.017734576016664505, + 0.0082358717918396, + 0.008789148181676865, + 0.014576735906302929, + 0.005974520929157734, + 0.00798872672021389, + 0.009425396099686623, + 0.006647587288171053, + 0.011999805457890034, + 0.0043342760764062405, + 0.005161300767213106, + 0.004962508101016283, + 0.003983998671174049, + 0.0036616523284465075, + 0.005529321264475584, + 0.002751236781477928, + 0.004668997600674629, + 0.0032987017184495926, + 0.0047653173096477985, + 0.00620614318177104, + 0.003013249719515443, + 0.004278308246284723, + 0.006095061544328928, + 0.002649025758728385, + 0.004357274156063795, + 0.008172428235411644, + 0.003943008836358786, + 0.007730982266366482, + 0.0036476270761340857, + 0.003915308974683285, + 0.01063468772917986, + 0.003945005126297474, + 0.007970652543008327, + 0.04758445546030998, + 0.011414028704166412, + 0.08540486544370651, + 0.02243203856050968, + 0.01897548697888851, + 0.017582232132554054, + 0.01685265824198723, + 0.02630876563489437, + 0.01380592305213213, + 0.009618900716304779, + 0.023013608530163765, + 0.011441776528954506, + 0.011937587521970272, + 0.008275642059743404, + 0.007473550736904144, + 0.010885157622396946, + 0.021661533042788506, + 0.007898072712123394, + 0.037652306258678436, + 0.012511571869254112, + 0.008845668286085129, + 0.015289848670363426, + 0.009291144087910652, + 0.01657732203602791, + 0.009547523222863674, + 0.004951579496264458, + 0.017492758110165596, + 0.005096472334116697, + 0.0058677648194134235, + 0.0059462194330990314, + 0.0034927749074995518, + 0.006435812450945377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 8 + ], + "selected_reward": -0.004096094518899918 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794618487358093 + ], + "opposite_logprob_tokens": [ + -3.529794692993164 + ], + "logprob_reward": -0.029794618487358093, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794618487358093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3591156005859375, + 0.20461441576480865, + 0.07239445298910141, + 0.05046626180410385, + 0.03408903256058693, + 0.04768577963113785, + 0.1344209462404251, + 0.04020296409726143, + 0.03158676251769066, + 0.0254238061606884 + ], + "credit_map": [ + 0.031428080052137375, + 0.03393529728055, + 0.041364990174770355, + 0.03668795898556709, + 0.04856448993086815, + 0.059703726321458817, + 0.07175785303115845, + 0.03567320480942726, + 0.017699941992759705, + 0.04153766483068466, + 0.026027904823422432, + 0.027158673852682114, + 0.0281402375549078, + 0.02679239772260189, + 0.01719348318874836, + 0.020064108073711395, + 0.007213052827864885, + 0.01300649344921112, + 0.012159113772213459, + 0.008331891149282455, + 0.009908923879265785, + 0.007984873838722706, + 0.0054423389956355095, + 0.008347760885953903, + 0.005632351618260145, + 0.008607675321400166, + 0.004981635604053736, + 0.006078575272113085, + 0.007178258150815964, + 0.006434429902583361, + 0.005718881729990244, + 0.005834453273564577, + 0.004469837062060833, + 0.003919539041817188, + 0.006137935444712639, + 0.0026692000683397055, + 0.002839193446561694, + 0.006953893229365349, + 0.0025629859883338213, + 0.004536445718258619, + 0.006507254205644131, + 0.006280358415096998, + 0.00825594738125801, + 0.002883296925574541, + 0.004798937123268843, + 0.008613405749201775, + 0.003837986383587122, + 0.006508590653538704, + 0.025289317592978477, + 0.007098699454218149, + 0.045754462480545044, + 0.014043418690562248, + 0.00736951595172286, + 0.008439349941909313, + 0.011524796485900879, + 0.014901375398039818, + 0.005741614382714033, + 0.00397493364289403, + 0.01082372386008501, + 0.0038707703351974487, + 0.004686309490352869, + 0.003762805601581931, + 0.00275972718372941, + 0.0045830789022147655, + 0.004480221774429083, + 0.003937205299735069, + 0.005428363103419542, + 0.0033087541814893484, + 0.0039289831183850765, + 0.003791568800806999, + 0.0022632547188550234, + 0.004448411520570517, + 0.0039051754865795374, + 0.0021452298387885094, + 0.008078094571828842, + 0.0019807361532002687, + 0.0025089180562645197, + 0.0025632602628320456, + 0.001587595441378653, + 0.0026547955349087715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.029794618487358093 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291882395744324 + ], + "opposite_logprob_tokens": [ + -2.2379188537597656 + ], + "logprob_reward": -0.11291882395744324, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291882395744324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10318096727132797, + 0.11897964030504227, + 0.1312788873910904, + 0.08614737540483475, + 0.055153608322143555, + 0.08520257472991943, + 0.086618572473526, + 0.051965679973363876, + 0.22215254604816437, + 0.05932018533349037 + ], + "credit_map": [ + 0.007828513160347939, + 0.014751487411558628, + 0.009998970665037632, + 0.012290038168430328, + 0.01565432734787464, + 0.019648469984531403, + 0.009014626033604145, + 0.013994531705975533, + 0.008248289115726948, + 0.0176527202129364, + 0.02241629734635353, + 0.01631253957748413, + 0.013215138576924801, + 0.011637385003268719, + 0.013238595798611641, + 0.01625867187976837, + 0.01405003760010004, + 0.02091720513999462, + 0.022172991186380386, + 0.012252943590283394, + 0.01339494064450264, + 0.02097439393401146, + 0.012050308287143707, + 0.015466065146028996, + 0.008427545428276062, + 0.006068226415663958, + 0.02008896879851818, + 0.006507082376629114, + 0.005842973478138447, + 0.021544666960835457, + 0.0051421672105789185, + 0.01252574659883976, + 0.008393124677240849, + 0.005029854364693165, + 0.013168442994356155, + 0.0046814498491585255, + 0.0045472802594304085, + 0.007690310478210449, + 0.004290051758289337, + 0.00735309487208724, + 0.012535136193037033, + 0.004685094114392996, + 0.011176392436027527, + 0.006476621609181166, + 0.005837873090058565, + 0.027702633291482925, + 0.005347855389118195, + 0.011440960690379143, + 0.015452521853148937, + 0.006019412074238062, + 0.02832731604576111, + 0.0074675427749753, + 0.008137301541864872, + 0.0067603085190057755, + 0.005342715885490179, + 0.00911145843565464, + 0.006825461518019438, + 0.00445663882419467, + 0.013049842789769173, + 0.004511657170951366, + 0.007169675547629595, + 0.005822293926030397, + 0.004166215192526579, + 0.005963893141597509, + 0.037676457315683365, + 0.02662174589931965, + 0.06249440833926201, + 0.018666815012693405, + 0.02865711972117424, + 0.014848457649350166, + 0.01284924428910017, + 0.020338281989097595, + 0.00845924112945795, + 0.005330410785973072, + 0.019353661686182022, + 0.0060842023231089115, + 0.004836881533265114, + 0.005360007751733065, + 0.004172700922936201, + 0.005723083391785622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.11291882395744324 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a frontal angle, first focuses on a static e-book cover displayed on a digital screen, imbued with a cyberpunk aesthetic. The background depicts a futuristic nightscape drenched in icy rain. Towering skyscrapers gleam with giant holographic advertisements, and the wet streets reflect magenta and cyan neon lights. Suddenly, the previously static female figure at the center of the cover—with asymmetrical bright pink hair, a face etched with glowing geometric circuitry, and clad in a tight-fitting black metallic tactical suit with a carbon fiber feel—suddenly acquires a sense of three-dimensional volume. In an explosive, dynamic pose, she bursts forward from the two-dimensional plane, seemingly piercing the screen itself. The entire movement, captured with incredibly smooth, high-frame-rate cinematography, appears incredibly fluent. Leaning forward, one arm extended, fingers spread, the other grips the grip of a high-tech pistol. Several modules, each with a faint, blinking indicator light, hang from a belt at her waist. As she bursts out of the frame, her sharp eyes stare forward, her face a look of resolute concentration. The movement of her flying hair and the folds of her clothing are clearly visible. The previously static raindrops and neon lights behind her seem to be torn apart by her movements, creating a slow-motion effect and a slight spatial distortion. The entire scene, in its cool tones of light and shadow, creates a unique atmosphere that blends futuristic, tense, and surreal aesthetics, imbued with a breathtaking sense of reality and visual impact.", + "seed": 4180, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_4/index_3843_seed_4180_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020526950852945446, + "vlm_logprob": -0.020526950852945446, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.020526950852945446, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03840995207428932 + ], + "opposite_logprob_tokens": [ + -3.538409948348999 + ], + "logprob_reward": -0.03840995207428932, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03840995207428932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11854797601699829, + 0.26915857195854187, + 0.06682615727186203, + 0.09894663840532303, + 0.09275036305189133, + 0.07143479585647583, + 0.1342790424823761, + 0.029356591403484344, + 0.07194200158119202, + 0.046757880598306656 + ], + "credit_map": [ + 0.010975207202136517, + 0.021534305065870285, + 0.01610126905143261, + 0.014434878714382648, + 0.025923386216163635, + 0.010119842365384102, + 0.010225649923086166, + 0.009233439341187477, + 0.020703807473182678, + 0.008603961206972599, + 0.011344585567712784, + 0.006833183579146862, + 0.09740979224443436, + 0.10262193530797958, + 0.009438352659344673, + 0.012202965095639229, + 0.008658744394779205, + 0.006012534257024527, + 0.01678866520524025, + 0.00695809256285429, + 0.007251374423503876, + 0.008440088480710983, + 0.004127879161387682, + 0.008588779717683792, + 0.012141893617808819, + 0.007752446457743645, + 0.025947025045752525, + 0.008412371389567852, + 0.010321622714400291, + 0.016210393980145454, + 0.006863278336822987, + 0.011297601275146008, + 0.01348467543721199, + 0.021825265139341354, + 0.012006783857941628, + 0.010702635161578655, + 0.009346257895231247, + 0.010176128707826138, + 0.006378632970154285, + 0.008829976432025433, + 0.010792913846671581, + 0.011097055859863758, + 0.00872352346777916, + 0.007278908044099808, + 0.010453754104673862, + 0.010226253420114517, + 0.005268832203000784, + 0.007593561429530382, + 0.018214091658592224, + 0.024204939603805542, + 0.020489297807216644, + 0.012599489651620388, + 0.015559748746454716, + 0.024729248136281967, + 0.009874649345874786, + 0.008607584051787853, + 0.0029047413263469934, + 0.004105222877115011, + 0.0030659951735287905, + 0.0038166996091604233, + 0.004574657417833805, + 0.003453619545325637, + 0.003979206085205078, + 0.0034564496017992496, + 0.008246537297964096, + 0.004067054484039545, + 0.024238498881459236, + 0.005062086973339319, + 0.004146185703575611, + 0.004705487750470638, + 0.005559122189879417, + 0.015917029231786728, + 0.0037921657785773277, + 0.00656886724755168, + 0.005159273743629456, + 0.0065874503925442696, + 0.008143438957631588, + 0.004441581200808287, + 0.006433519069105387, + 0.005631581414490938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.03840995207428932 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002836730098351836 + ], + "opposite_logprob_tokens": [ + -5.87783670425415 + ], + "logprob_reward": -0.002836730098351836, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002836730098351836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22085461020469666, + 0.08406354486942291, + 0.13602937757968903, + 0.05731944739818573, + 0.14278313517570496, + 0.03528730571269989, + 0.04923881217837334, + 0.014903941191732883, + 0.1683592051267624, + 0.0911606177687645 + ], + "credit_map": [ + 0.021150153130292892, + 0.029025964438915253, + 0.018357427790760994, + 0.029731320217251778, + 0.04131793975830078, + 0.04183994606137276, + 0.01777692511677742, + 0.02165493182837963, + 0.012316846288740635, + 0.0053973873145878315, + 0.014737540856003761, + 0.0070821866393089294, + 0.006339649669826031, + 0.01574130915105343, + 0.00663619302213192, + 0.015812428668141365, + 0.021694639697670937, + 0.010054734535515308, + 0.028623681515455246, + 0.010454817675054073, + 0.022224662825465202, + 0.018575359135866165, + 0.010306673124432564, + 0.014094811864197254, + 0.008329842239618301, + 0.0036091175861656666, + 0.017388593405485153, + 0.005584377329796553, + 0.0038742548786103725, + 0.007383518386632204, + 0.0032556825317442417, + 0.007894066162407398, + 0.048957619816064835, + 0.00863464456051588, + 0.04656635597348213, + 0.010567726567387581, + 0.006027274765074253, + 0.007586146239191294, + 0.006858517415821552, + 0.007584849372506142, + 0.007155353669077158, + 0.005535762757062912, + 0.007589402608573437, + 0.0033876642119139433, + 0.0037346510216593742, + 0.003029177663847804, + 0.0021117976866662502, + 0.002743497956544161, + 0.0059441556222736835, + 0.007103892974555492, + 0.007934503257274628, + 0.006912595126777887, + 0.005783375818282366, + 0.006597917061299086, + 0.004394694697111845, + 0.004567676689475775, + 0.0015788364689797163, + 0.0019771282095462084, + 0.0015004152664914727, + 0.0016646089497953653, + 0.0026190527714788914, + 0.002010508207604289, + 0.0017774205189198256, + 0.0017759718466550112, + 0.021498369053006172, + 0.0055161952041089535, + 0.06099734827876091, + 0.009966920129954815, + 0.006661472376435995, + 0.009100357070565224, + 0.013480803929269314, + 0.04113774746656418, + 0.007384682074189186, + 0.023737194016575813, + 0.011156254447996616, + 0.010599722154438496, + 0.01319479662925005, + 0.007813014090061188, + 0.008548936806619167, + 0.008726012893021107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.002836730098351836 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252134520560503 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252134520560503, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252134520560503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41512393951416016, + 0.06376145780086517, + 0.08128141611814499, + 0.10954409837722778, + 0.09277549386024475, + 0.04080952703952789, + 0.0371011421084404, + 0.02036232501268387, + 0.10004385560750961, + 0.03919677808880806 + ], + "credit_map": [ + 0.034357670694589615, + 0.05517454817891121, + 0.0558755025267601, + 0.06954429298639297, + 0.07169019430875778, + 0.05784452334046364, + 0.032638344913721085, + 0.037998851388692856, + 0.004293408710509539, + 0.00425918260589242, + 0.007112250197678804, + 0.004618453793227673, + 0.018058840185403824, + 0.013345650397241116, + 0.0059646363370120525, + 0.00610903836786747, + 0.011275466531515121, + 0.009195911698043346, + 0.014078100211918354, + 0.009386717341840267, + 0.012166260741651058, + 0.010486938059329987, + 0.007807803340256214, + 0.006884217262268066, + 0.013338381424546242, + 0.015051770024001598, + 0.0220944881439209, + 0.010710618458688259, + 0.013120174407958984, + 0.010543077252805233, + 0.011081391014158726, + 0.013604198582470417, + 0.027310503646731377, + 0.008400509133934975, + 0.02496536448597908, + 0.009035794995725155, + 0.0047006686218082905, + 0.006507049314677715, + 0.005623822100460529, + 0.006231779232621193, + 0.008116945624351501, + 0.006486734375357628, + 0.008607971481978893, + 0.004781844560056925, + 0.0038669321220368147, + 0.003919668961316347, + 0.0026122028939425945, + 0.0024172242265194654, + 0.00536707229912281, + 0.00511658051982522, + 0.005531794857233763, + 0.00416514603421092, + 0.005064387805759907, + 0.005046946462243795, + 0.0031904559582471848, + 0.003618754679337144, + 0.002339862287044525, + 0.0024950499646365643, + 0.0020965286530554295, + 0.0024373221676796675, + 0.0022274863440543413, + 0.00242515723221004, + 0.0028123657684773207, + 0.0035285528283566236, + 0.012949427589774132, + 0.003523101331666112, + 0.03608221933245659, + 0.005900443531572819, + 0.004185416270047426, + 0.00550633342936635, + 0.007782790344208479, + 0.02411412075161934, + 0.003064107848331332, + 0.00522732175886631, + 0.003963252063840628, + 0.005272525362670422, + 0.008557253517210484, + 0.003485490567982197, + 0.0063433353789150715, + 0.003283493686467409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.005252134520560503 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026341434568166733 + ], + "opposite_logprob_tokens": [ + -3.651341438293457 + ], + "logprob_reward": -0.026341434568166733, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026341434568166733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14932529628276825, + 0.12134537100791931, + 0.07272657752037048, + 0.12403185665607452, + 0.14725546538829803, + 0.08304288238286972, + 0.07230135798454285, + 0.028868423774838448, + 0.15893685817718506, + 0.042165905237197876 + ], + "credit_map": [ + 0.01845966838300228, + 0.0200065765529871, + 0.02819558046758175, + 0.016585102304816246, + 0.022543692961335182, + 0.019342953339219093, + 0.011558137834072113, + 0.012633568607270718, + 0.012357243336737156, + 0.008136529475450516, + 0.016398876905441284, + 0.008143543265759945, + 0.009939526207745075, + 0.02766239084303379, + 0.01110968366265297, + 0.027597570791840553, + 0.012750775553286076, + 0.006334702949970961, + 0.008819800801575184, + 0.009377620182931423, + 0.0071096536703407764, + 0.009755294770002365, + 0.007345833349972963, + 0.011232891120016575, + 0.014193769544363022, + 0.01348895113915205, + 0.015463214367628098, + 0.014130672439932823, + 0.02616710402071476, + 0.014587943442165852, + 0.013818629086017609, + 0.012181561440229416, + 0.03368751332163811, + 0.008409254252910614, + 0.03168695420026779, + 0.010623572394251823, + 0.011872353032231331, + 0.024276914075016975, + 0.012679221108555794, + 0.014019681140780449, + 0.013959514908492565, + 0.013262789696455002, + 0.013541823253035545, + 0.007910985499620438, + 0.008762583136558533, + 0.012904582545161247, + 0.005351673346012831, + 0.007348925806581974, + 0.008123132400214672, + 0.009415318258106709, + 0.011712080799043179, + 0.008016585372388363, + 0.008327852934598923, + 0.011825027875602245, + 0.00728003541007638, + 0.007601314689964056, + 0.004410485737025738, + 0.0030186637304723263, + 0.0027254181914031506, + 0.0029324181377887726, + 0.00378014356829226, + 0.0035210007335990667, + 0.004126980900764465, + 0.004353312775492668, + 0.019434625282883644, + 0.009618230164051056, + 0.05336443707346916, + 0.009236127138137817, + 0.013153346255421638, + 0.009310835041105747, + 0.011600279249250889, + 0.03321896120905876, + 0.0030063914600759745, + 0.008110297843813896, + 0.00603210274130106, + 0.004768461920320988, + 0.006983505096286535, + 0.0041547976434230804, + 0.004084741696715355, + 0.005025605671107769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.026341434568166733 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979450300335884 + ], + "opposite_logprob_tokens": [ + -3.529794454574585 + ], + "logprob_reward": -0.02979450300335884, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979450300335884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17214633524417877, + 0.16171514987945557, + 0.08834655582904816, + 0.0779225304722786, + 0.0939306691288948, + 0.04430432245135307, + 0.06394373625516891, + 0.026020333170890808, + 0.1718670129776001, + 0.09980335086584091 + ], + "credit_map": [ + 0.013793016783893108, + 0.02944077178835869, + 0.016543032601475716, + 0.028503797948360443, + 0.03639334440231323, + 0.017258724197745323, + 0.0188460610806942, + 0.011367591097950935, + 0.006960007362067699, + 0.00555498106405139, + 0.010314056649804115, + 0.0056222244165837765, + 0.04138760268688202, + 0.06127196550369263, + 0.007496796082705259, + 0.02310752309858799, + 0.016611739993095398, + 0.009401736781001091, + 0.010822730138897896, + 0.006491087377071381, + 0.013689587824046612, + 0.012917127460241318, + 0.0091116176918149, + 0.00930093228816986, + 0.010298958979547024, + 0.007589615415781736, + 0.01840844377875328, + 0.007691817358136177, + 0.010049665346741676, + 0.00748366629704833, + 0.00709170289337635, + 0.009308663196861744, + 0.02775416150689125, + 0.007360001094639301, + 0.026003871113061905, + 0.005882167723029852, + 0.0051000951789319515, + 0.008246170356869698, + 0.004703431855887175, + 0.008880767971277237, + 0.008793982677161694, + 0.004583870060741901, + 0.008770013228058815, + 0.0037239573430269957, + 0.004522035364061594, + 0.006704023107886314, + 0.003206143155694008, + 0.004000295419245958, + 0.007216328289359808, + 0.008200390264391899, + 0.006634978111833334, + 0.01141471415758133, + 0.0068532926961779594, + 0.011978477239608765, + 0.006335563957691193, + 0.00530998595058918, + 0.0026913147885352373, + 0.0023888638243079185, + 0.0029811544809490442, + 0.0031587472185492516, + 0.002555867424234748, + 0.003819458419457078, + 0.004010691307485104, + 0.004414234310388565, + 0.021033454686403275, + 0.008339504711329937, + 0.05892254039645195, + 0.010348602198064327, + 0.009833214804530144, + 0.010092686861753464, + 0.012898659333586693, + 0.040398333221673965, + 0.007891472429037094, + 0.026731746271252632, + 0.011682411655783653, + 0.01071917824447155, + 0.0163293294608593, + 0.008363714441657066, + 0.010326962918043137, + 0.007758538704365492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.02979450300335884 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a frontal angle, first focuses on a static e-book cover displayed on a digital screen, imbued with a cyberpunk aesthetic. The background depicts a futuristic nightscape drenched in icy rain. Towering skyscrapers gleam with giant holographic advertisements, and the wet streets reflect magenta and cyan neon lights. Suddenly, the previously static female figure at the center of the cover—with asymmetrical bright pink hair, a face etched with glowing geometric circuitry, and clad in a tight-fitting black metallic tactical suit with a carbon fiber feel—suddenly acquires a sense of three-dimensional volume. In an explosive, dynamic pose, she bursts forward from the two-dimensional plane, seemingly piercing the screen itself. The entire movement, captured with incredibly smooth, high-frame-rate cinematography, appears incredibly fluent. Leaning forward, one arm extended, fingers spread, the other grips the grip of a high-tech pistol. Several modules, each with a faint, blinking indicator light, hang from a belt at her waist. As she bursts out of the frame, her sharp eyes stare forward, her face a look of resolute concentration. The movement of her flying hair and the folds of her clothing are clearly visible. The previously static raindrops and neon lights behind her seem to be torn apart by her movements, creating a slow-motion effect and a slight spatial distortion. The entire scene, in its cool tones of light and shadow, creates a unique atmosphere that blends futuristic, tense, and surreal aesthetics, imbued with a breathtaking sense of reality and visual impact.", + "seed": 4181, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_4/index_3843_seed_4181_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02830293159931898, + "vlm_logprob": -0.02830293159931898, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.02830293159931898, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04980209097266197 + ], + "opposite_logprob_tokens": [ + -3.17480206489563 + ], + "logprob_reward": -0.04980209097266197, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04980209097266197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04956554248929024, + 0.5583577752113342, + 0.03222295641899109, + 0.04504149779677391, + 0.06415153294801712, + 0.01871020533144474, + 0.016812404617667198, + 0.009833777323365211, + 0.19380266964435577, + 0.011501632630825043 + ], + "credit_map": [ + 0.004582779947668314, + 0.005832963157445192, + 0.005237450823187828, + 0.004984026309102774, + 0.004702593665570021, + 0.009759812615811825, + 0.00516354339197278, + 0.009302372112870216, + 0.08074600994586945, + 0.00915808230638504, + 0.24582600593566895, + 0.020410122349858284, + 0.013796099461615086, + 0.02410813421010971, + 0.012669515796005726, + 0.1516438126564026, + 0.006843844428658485, + 0.0026045574340969324, + 0.005572797264903784, + 0.003074629232287407, + 0.0021413657814264297, + 0.005066236015409231, + 0.0031839716248214245, + 0.0037355530075728893, + 0.010565930046141148, + 0.00397182023152709, + 0.01475584413856268, + 0.004293239209800959, + 0.0019474751316010952, + 0.003192819422110915, + 0.002806668868288398, + 0.0035076995845884085, + 0.017178134992718697, + 0.003289748914539814, + 0.006029024720191956, + 0.0030656319577246904, + 0.002892146585509181, + 0.01802603341639042, + 0.0035618399269878864, + 0.010108966380357742, + 0.0023669758811593056, + 0.0020611209329217672, + 0.0028826454654335976, + 0.002050472889095545, + 0.0014981606509536505, + 0.00243129744194448, + 0.0022548639681190252, + 0.0031646667048335075, + 0.0018324173288419843, + 0.001998281106352806, + 0.0023536719381809235, + 0.0022151011507958174, + 0.00244901399128139, + 0.0019172042375430465, + 0.0016950716963037848, + 0.0023516423534601927, + 0.0012435362441465259, + 0.0011134300148114562, + 0.0012124241329729557, + 0.0010059424676001072, + 0.0008714553550817072, + 0.0018265508115291595, + 0.0008306204108521342, + 0.0017298177117481828, + 0.02755492739379406, + 0.018926778808236122, + 0.03029492124915123, + 0.01710786670446396, + 0.020574618130922318, + 0.04041801765561104, + 0.017490597441792488, + 0.02143494226038456, + 0.0012167891254648566, + 0.0014160997234284878, + 0.0017387355910614133, + 0.002000476000830531, + 0.0014131000498309731, + 0.0009043500176630914, + 0.0014100708067417145, + 0.0014020110247656703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 4 + ], + "selected_reward": -0.04980209097266197 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004113190341740847 + ], + "opposite_logprob_tokens": [ + -5.50411319732666 + ], + "logprob_reward": -0.004113190341740847, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004113190341740847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03171747177839279, + 0.8122689127922058, + 0.01923031359910965, + 0.03800119832158089, + 0.04102133959531784, + 0.008073040284216404, + 0.005340163130313158, + 0.004092950373888016, + 0.033687349408864975, + 0.006567241158336401 + ], + "credit_map": [ + 0.0065140933729708195, + 0.0027119398582726717, + 0.004976891912519932, + 0.002616962417960167, + 0.002642907202243805, + 0.004187732934951782, + 0.002898914273828268, + 0.005168032366782427, + 0.12211020290851593, + 0.012079532258212566, + 0.3620578944683075, + 0.028499772772192955, + 0.019304528832435608, + 0.03342827409505844, + 0.016238082200288773, + 0.2185506522655487, + 0.004618651699274778, + 0.0014803409576416016, + 0.0031083780340850353, + 0.0018867762992158532, + 0.001336717396043241, + 0.0029921955429017544, + 0.0016516995383426547, + 0.0021555537823587656, + 0.009871997870504856, + 0.00152818881906569, + 0.013616350479424, + 0.00265403650701046, + 0.001601142925210297, + 0.0038975083734840155, + 0.0018584956414997578, + 0.00297348340973258, + 0.010840245522558689, + 0.0015421068528667092, + 0.005723652895539999, + 0.001853741006925702, + 0.0014301914488896728, + 0.013051520101726055, + 0.001642139395698905, + 0.004937737248837948, + 0.0009696025517769158, + 0.0006588423275388777, + 0.0014290378894656897, + 0.0007349528023041785, + 0.0005423725815489888, + 0.0011010811431333423, + 0.0010400733444839716, + 0.001597077352926135, + 0.0005562468431890011, + 0.0006495678098872304, + 0.000775233784224838, + 0.0007118000066839159, + 0.0005750744603574276, + 0.0006046216003596783, + 0.0006571075064130127, + 0.0008105110609903932, + 0.0005038424860686064, + 0.0004544846306089312, + 0.000587266928050667, + 0.0004373063857201487, + 0.0003902914759237319, + 0.0006375358789227903, + 0.0004627546004485339, + 0.0006194678135216236, + 0.004950449801981449, + 0.0034947143867611885, + 0.005563632119446993, + 0.0030127381905913353, + 0.0035785280633717775, + 0.006629338022321463, + 0.0027954441029578447, + 0.0036625079810619354, + 0.0011374345049262047, + 0.0004542834940366447, + 0.0012226476101204753, + 0.0009243091335520148, + 0.0006878073909319937, + 0.0006313034682534635, + 0.0007065265090204775, + 0.0008029285818338394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.004113190341740847 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254624877125025 + ], + "opposite_logprob_tokens": [ + -5.255254745483398 + ], + "logprob_reward": -0.005254624877125025, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254624877125025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0436672568321228, + 0.6373640298843384, + 0.07557021826505661, + 0.0681488886475563, + 0.030613387003540993, + 0.0334891639649868, + 0.037284404039382935, + 0.014891290105879307, + 0.04600583389401436, + 0.012965571135282516 + ], + "credit_map": [ + 0.00480352807790041, + 0.006043637171387672, + 0.004246339667588472, + 0.0038841699715703726, + 0.004373972304165363, + 0.008765102364122868, + 0.004425905644893646, + 0.007124600000679493, + 0.09768989682197571, + 0.009070559404790401, + 0.2860478460788727, + 0.02271796017885208, + 0.01427417527884245, + 0.024844186380505562, + 0.012616646476089954, + 0.17010270059108734, + 0.010622011497616768, + 0.007866710424423218, + 0.011130819097161293, + 0.008974511176347733, + 0.005789806600660086, + 0.01022025290876627, + 0.008373183198273182, + 0.0125929219648242, + 0.017662782222032547, + 0.0051245070062577724, + 0.02492283098399639, + 0.004484058823436499, + 0.0024786810390651226, + 0.005305103026330471, + 0.0033827798906713724, + 0.004788139369338751, + 0.006727645173668861, + 0.0027898913249373436, + 0.003278913674876094, + 0.002706252969801426, + 0.0019022986525669694, + 0.007850408554077148, + 0.0019867434166371822, + 0.003371231723576784, + 0.003935191314667463, + 0.003400189336389303, + 0.0056148492731153965, + 0.0038966056890785694, + 0.002217720728367567, + 0.004437719937413931, + 0.0036292285658419132, + 0.006357656326144934, + 0.005151352379471064, + 0.0038077232893556356, + 0.0037178276106715202, + 0.00510307727381587, + 0.0044561042450368404, + 0.003276652190834284, + 0.006684230174869299, + 0.005087435711175203, + 0.0017679122975096107, + 0.0021484282333403826, + 0.0016928332624956965, + 0.0018982663750648499, + 0.0017638691933825612, + 0.0023852151352912188, + 0.0012961513129994273, + 0.0019386126659810543, + 0.006595931947231293, + 0.004226775839924812, + 0.008919489569962025, + 0.004031775519251823, + 0.004728924948722124, + 0.00914621539413929, + 0.0032749189995229244, + 0.005081798415631056, + 0.0018520448356866837, + 0.0013229043688625097, + 0.0021201546769589186, + 0.0014888931764289737, + 0.0015865956665948033, + 0.0015965692000463605, + 0.0013845603680238128, + 0.001613847678527236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.005254624877125025 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048650916665792465 + ], + "opposite_logprob_tokens": [ + -3.0486509799957275 + ], + "logprob_reward": -0.048650916665792465, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048650916665792465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2727286219596863, + 0.42351508140563965, + 0.043527159839868546, + 0.029762228950858116, + 0.04094705730676651, + 0.024508949369192123, + 0.05065562203526497, + 0.008171264082193375, + 0.09290599822998047, + 0.013278043828904629 + ], + "credit_map": [ + 0.020915674045681953, + 0.04001062735915184, + 0.022557144984602928, + 0.032673485577106476, + 0.024768363684415817, + 0.0502956323325634, + 0.02973710186779499, + 0.05177060142159462, + 0.07590752840042114, + 0.012019982561469078, + 0.18761587142944336, + 0.016156762838363647, + 0.009469754062592983, + 0.017948191612958908, + 0.010212855413556099, + 0.09418415278196335, + 0.009265769273042679, + 0.004828119184821844, + 0.006713494658470154, + 0.005400364752858877, + 0.0026455845218151808, + 0.005817010998725891, + 0.0042778439819812775, + 0.004578970838338137, + 0.005239652935415506, + 0.002867682371288538, + 0.005749536678195, + 0.0031279423274099827, + 0.0018472104566171765, + 0.004562538117170334, + 0.0027606021612882614, + 0.003607061691582203, + 0.009988619945943356, + 0.0022152785677462816, + 0.005857064388692379, + 0.0019257455132901669, + 0.0017899571685120463, + 0.011875808238983154, + 0.0019192314939573407, + 0.005375352222472429, + 0.002793411258608103, + 0.002642913954332471, + 0.0034225573763251305, + 0.002569501521065831, + 0.0022094296291470528, + 0.0029571717604994774, + 0.002458084374666214, + 0.0054558804258704185, + 0.005744912661612034, + 0.006462789140641689, + 0.003607959719374776, + 0.006876051425933838, + 0.0067390869371593, + 0.005100118461996317, + 0.008162295445799828, + 0.007962404750287533, + 0.000978193012997508, + 0.0009012985974550247, + 0.0011024880222976208, + 0.0006466928753070533, + 0.000998804927803576, + 0.0014069817261770368, + 0.0008130661444738507, + 0.0013237388338893652, + 0.012986334040760994, + 0.008580081164836884, + 0.01579475589096546, + 0.008856231346726418, + 0.009749261662364006, + 0.018501605838537216, + 0.00796643178910017, + 0.010471295565366745, + 0.0022796145640313625, + 0.0013606554130092263, + 0.001569868065416813, + 0.0022057031746953726, + 0.0016506888205185533, + 0.0011592437513172626, + 0.001587345264852047, + 0.0014649259392172098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.048650916665792465 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336938351392746 + ], + "opposite_logprob_tokens": [ + -3.408693790435791 + ], + "logprob_reward": -0.0336938351392746, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336938351392746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08947763592004776, + 0.6085551977157593, + 0.08901659399271011, + 0.03133365139365196, + 0.028486263006925583, + 0.032799914479255676, + 0.0261654332280159, + 0.018527470529079437, + 0.0527193620800972, + 0.02291850931942463 + ], + "credit_map": [ + 0.012130131945014, + 0.007270607631653547, + 0.011638931930065155, + 0.006985157262533903, + 0.0073204850777983665, + 0.01715954951941967, + 0.0069592855870723724, + 0.02001347951591015, + 0.09153539687395096, + 0.009687143377959728, + 0.27124419808387756, + 0.02382548339664936, + 0.0132899796590209, + 0.022613581269979477, + 0.012953681871294975, + 0.16340570151805878, + 0.01483003981411457, + 0.007884927093982697, + 0.013759477995336056, + 0.007727950811386108, + 0.00655090156942606, + 0.013850212097167969, + 0.00816773809492588, + 0.016245339065790176, + 0.004859304055571556, + 0.0030399500392377377, + 0.007143991068005562, + 0.0031096937600523233, + 0.0017790429992601275, + 0.003655018052086234, + 0.0031332080252468586, + 0.00461344001814723, + 0.004479172173887491, + 0.0031357225961983204, + 0.004088247194886208, + 0.0028481169138103724, + 0.003199067199602723, + 0.0030733251478523016, + 0.003088661003857851, + 0.004573948681354523, + 0.0037506178487092257, + 0.00403186259791255, + 0.0045698257163167, + 0.003421282861381769, + 0.0026010761503130198, + 0.004416674841195345, + 0.0028807802591472864, + 0.007127789780497551, + 0.0030039306730031967, + 0.0029288074001669884, + 0.0034183086827397346, + 0.0035005100071430206, + 0.0037224635016173124, + 0.002867263974621892, + 0.0036006723530590534, + 0.003123474307358265, + 0.0025915703736245632, + 0.001827222527936101, + 0.002602602820843458, + 0.002378911944106221, + 0.0015669540734961629, + 0.0032880655489861965, + 0.0017041824758052826, + 0.002567959949374199, + 0.007718455046415329, + 0.004941783845424652, + 0.008314716629683971, + 0.003799143712967634, + 0.005428444594144821, + 0.012566080316901207, + 0.0037741910200566053, + 0.006176541559398174, + 0.0027064592577517033, + 0.002316621830686927, + 0.003166450187563896, + 0.004484488628804684, + 0.0027619805186986923, + 0.002393952338024974, + 0.002365524647757411, + 0.002723031910136342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0336938351392746 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a frontal angle, first focuses on a static e-book cover displayed on a digital screen, imbued with a cyberpunk aesthetic. The background depicts a futuristic nightscape drenched in icy rain. Towering skyscrapers gleam with giant holographic advertisements, and the wet streets reflect magenta and cyan neon lights. Suddenly, the previously static female figure at the center of the cover—with asymmetrical bright pink hair, a face etched with glowing geometric circuitry, and clad in a tight-fitting black metallic tactical suit with a carbon fiber feel—suddenly acquires a sense of three-dimensional volume. In an explosive, dynamic pose, she bursts forward from the two-dimensional plane, seemingly piercing the screen itself. The entire movement, captured with incredibly smooth, high-frame-rate cinematography, appears incredibly fluent. Leaning forward, one arm extended, fingers spread, the other grips the grip of a high-tech pistol. Several modules, each with a faint, blinking indicator light, hang from a belt at her waist. As she bursts out of the frame, her sharp eyes stare forward, her face a look of resolute concentration. The movement of her flying hair and the folds of her clothing are clearly visible. The previously static raindrops and neon lights behind her seem to be torn apart by her movements, creating a slow-motion effect and a slight spatial distortion. The entire scene, in its cool tones of light and shadow, creates a unique atmosphere that blends futuristic, tense, and surreal aesthetics, imbued with a breathtaking sense of reality and visual impact.", + "seed": 4182, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_4/index_3843_seed_4182_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.046783382259309295, + "vlm_logprob": -0.046783382259309295, + "vlm_yes_no_margin": 3.799999976158142, + "vlm_token_credit": -0.046783382259309295, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024150127544999123 + ], + "opposite_logprob_tokens": [ + -4.0241498947143555 + ], + "logprob_reward": -0.024150127544999123, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.024150127544999123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1323816180229187, + 0.5012901425361633, + 0.05637156218290329, + 0.20879118144512177, + 0.06711450964212418, + 0.00807502306997776, + 0.006470923312008381, + 0.005652334541082382, + 0.009337212890386581, + 0.004515509586781263 + ], + "credit_map": [ + 0.0045782350935041904, + 0.009619411081075668, + 0.004965354688465595, + 0.02856033854186535, + 0.008289359509944916, + 0.018224403262138367, + 0.02234836295247078, + 0.0357961468398571, + 0.01785043440759182, + 0.0073071434162557125, + 0.02314499393105507, + 0.008324924856424332, + 0.022745903581380844, + 0.16550596058368683, + 0.019167302176356316, + 0.23724347352981567, + 0.005263973027467728, + 0.010382145643234253, + 0.008689787238836288, + 0.005630905274301767, + 0.009599297307431698, + 0.007508887443691492, + 0.0050636883825063705, + 0.0042328801937401295, + 0.015521207824349403, + 0.009712371043860912, + 0.07318111509084702, + 0.008380723185837269, + 0.01592199131846428, + 0.06948333233594894, + 0.007341573014855385, + 0.009248866699635983, + 0.008797354064881802, + 0.008117123506963253, + 0.010906756855547428, + 0.007881908677518368, + 0.006019400432705879, + 0.010189398191869259, + 0.005977825727313757, + 0.009224745444953442, + 0.0018139942549169064, + 0.0007558667566627264, + 0.0011407280107960105, + 0.0005729845142923295, + 0.0008172187954187393, + 0.0014282860793173313, + 0.0006820712587796152, + 0.0008638737490400672, + 0.0008245794451795518, + 0.0005192087264731526, + 0.0010189013555645943, + 0.0006794098298996687, + 0.000723558128811419, + 0.0013043500948697329, + 0.0006046129274182022, + 0.0007963027455843985, + 0.0006311099859885871, + 0.0006282098474912345, + 0.0006896598497405648, + 0.0005412301397882402, + 0.0006369570037350059, + 0.0007754914695397019, + 0.0006322708213701844, + 0.0011174057144671679, + 0.0008229598170146346, + 0.0015039267018437386, + 0.001387367956340313, + 0.0015037668636068702, + 0.0010285101598128676, + 0.0008326507522724569, + 0.0012916503474116325, + 0.0009663805249147117, + 0.0007576537900604308, + 0.0006093470728956163, + 0.0005892951157875359, + 0.0004901491920463741, + 0.00034294382203370333, + 0.000685869890730828, + 0.0004430450208019465, + 0.0005972055369056761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.024150127544999123 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006757147144526243 + ], + "opposite_logprob_tokens": [ + -5.0067572593688965 + ], + "logprob_reward": -0.006757147144526243, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006757147144526243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07612007111310959, + 0.5515608787536621, + 0.1159110888838768, + 0.13038748502731323, + 0.021757131442427635, + 0.023038102313876152, + 0.020201679319143295, + 0.011317268013954163, + 0.02934213913977146, + 0.020364128053188324 + ], + "credit_map": [ + 0.004991563502699137, + 0.00843839067965746, + 0.0065234084613621235, + 0.01831478253006935, + 0.007143101189285517, + 0.006656207609921694, + 0.01109451986849308, + 0.012958096340298653, + 0.01981962099671364, + 0.0071876393631100655, + 0.025926481932401657, + 0.009994324296712875, + 0.026146020740270615, + 0.18136104941368103, + 0.020407235249876976, + 0.26071852445602417, + 0.014109104871749878, + 0.012101485393941402, + 0.021706001833081245, + 0.009286677464842796, + 0.017468303442001343, + 0.013402913697063923, + 0.00934907328337431, + 0.018487539142370224, + 0.010657301172614098, + 0.0038371963892132044, + 0.0492064468562603, + 0.004910355433821678, + 0.009445486590266228, + 0.043301407247781754, + 0.003484450513496995, + 0.005544832441955805, + 0.0022975739557296038, + 0.0020024210680276155, + 0.003761380212381482, + 0.0017464038683101535, + 0.0013846822548657656, + 0.0038385645020753145, + 0.0021995767019689083, + 0.004526529461145401, + 0.004139795899391174, + 0.0021934867836534977, + 0.0030905071180313826, + 0.0023555587977170944, + 0.00212989654392004, + 0.004789321683347225, + 0.002258614171296358, + 0.0020809222478419542, + 0.002555669052526355, + 0.002007599687203765, + 0.002662456128746271, + 0.0036063254810869694, + 0.002112404443323612, + 0.002412322210147977, + 0.002659711055457592, + 0.002185191260650754, + 0.0017571293283253908, + 0.0017829498974606395, + 0.0014412547461688519, + 0.0009019515127874911, + 0.0011715036816895008, + 0.0013442115159705281, + 0.0012243191013112664, + 0.0016939477063715458, + 0.004217894282191992, + 0.003232849296182394, + 0.004735930822789669, + 0.003846776206046343, + 0.00212677288800478, + 0.0036422181874513626, + 0.0031746996100991964, + 0.004364994820207357, + 0.002911351853981614, + 0.00170116254594177, + 0.004453805275261402, + 0.0018306543352082372, + 0.0013451739214360714, + 0.003082155017182231, + 0.003235537325963378, + 0.0018042868468910456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.006757147144526243 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255454685539007 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.005255454685539007, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255454685539007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06701439619064331, + 0.8180909752845764, + 0.015847932547330856, + 0.03130612149834633, + 0.026442037895321846, + 0.010051644407212734, + 0.00898651871830225, + 0.01242976076900959, + 0.004455006215721369, + 0.0053756157867610455 + ], + "credit_map": [ + 0.002928101224824786, + 0.007055626716464758, + 0.003924396820366383, + 0.01648356392979622, + 0.004050505347549915, + 0.007055230438709259, + 0.011529958806931973, + 0.013987019658088684, + 0.02911994792521, + 0.009995441883802414, + 0.037637919187545776, + 0.01296080183237791, + 0.03785543143749237, + 0.27118656039237976, + 0.03053753823041916, + 0.3887973725795746, + 0.0012364255962893367, + 0.0026785654481500387, + 0.002507889876142144, + 0.00150919402949512, + 0.00260414881631732, + 0.001989691751077771, + 0.0011180391302332282, + 0.002203976036980748, + 0.0023151312489062548, + 0.0012527162907645106, + 0.009526819922029972, + 0.001688274322077632, + 0.0025409541558474302, + 0.009682470001280308, + 0.001572262728586793, + 0.0027274908497929573, + 0.0030595038551837206, + 0.003509273286908865, + 0.004396738484501839, + 0.0030463298317044973, + 0.002138599054887891, + 0.0035832419525831938, + 0.0032915284391492605, + 0.003416824620217085, + 0.0013586919521912932, + 0.0010463722283020616, + 0.0016233919886872172, + 0.0010500545613467693, + 0.0010681543499231339, + 0.0015546667855232954, + 0.0015102486358955503, + 0.0008400633232668042, + 0.0009491132223047316, + 0.001272320980206132, + 0.0011747581884264946, + 0.0012404491426423192, + 0.0009300657548010349, + 0.0009511203970760107, + 0.0012202630750834942, + 0.0012484280159696937, + 0.0011304864892736077, + 0.0018429860938340425, + 0.0011104962322860956, + 0.001644805772230029, + 0.0015444131568074226, + 0.0011951469350606203, + 0.001887154532596469, + 0.002074270974844694, + 0.0005021432880312204, + 0.0004934922326356173, + 0.0006495134439319372, + 0.0007564388797618449, + 0.0004019323387183249, + 0.00044832765706814826, + 0.0005158652784302831, + 0.0006872927770018578, + 0.0005713872378692031, + 0.0007021391647867858, + 0.0011698042508214712, + 0.0006109264213591814, + 0.0005141512374393642, + 0.0006611947319470346, + 0.0006580299814231694, + 0.00048798276111483574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.005255454685539007 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549561008810997 + ], + "opposite_logprob_tokens": [ + -2.9299561977386475 + ], + "logprob_reward": -0.0549561008810997, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549561008810997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23739635944366455, + 0.6750075817108154, + 0.019710643216967583, + 0.010290161706507206, + 0.01158315408974886, + 0.009266315959393978, + 0.007470687385648489, + 0.010191687382757664, + 0.005883208476006985, + 0.013200169429183006 + ], + "credit_map": [ + 0.007213717792183161, + 0.016390200704336166, + 0.009238619357347488, + 0.053819652646780014, + 0.011910943314433098, + 0.0339648611843586, + 0.03864821419119835, + 0.06621015071868896, + 0.02402973547577858, + 0.008319799788296223, + 0.031631674617528915, + 0.011117634363472462, + 0.031670957803726196, + 0.22232098877429962, + 0.025132866576313972, + 0.32078394293785095, + 0.0015968125080689788, + 0.002141843317076564, + 0.0036524715833365917, + 0.0013203483540564775, + 0.004149607848376036, + 0.00366612640209496, + 0.0014784991508349776, + 0.0017049351008608937, + 0.00105806824285537, + 0.0007446740055456758, + 0.0025731027126312256, + 0.0009797740494832397, + 0.0008403994725085795, + 0.0023284878116101027, + 0.0006974606076255441, + 0.0010681949788704515, + 0.0015030534705147147, + 0.0024984683841466904, + 0.0015487605705857277, + 0.0010677657555788755, + 0.0009605002123862505, + 0.0012826622696593404, + 0.001293977489694953, + 0.0014279664028435946, + 0.0018302962416782975, + 0.0006073564291000366, + 0.0014249542728066444, + 0.0007672483916394413, + 0.0009611509740352631, + 0.001736628357321024, + 0.001179136917926371, + 0.0007595439092256129, + 0.0010725735919550061, + 0.0009731724858283997, + 0.0009959283052012324, + 0.0008763007936067879, + 0.000855226768180728, + 0.0012894560350105166, + 0.0007967618876136839, + 0.000611267052590847, + 0.0011125183664262295, + 0.001344015239737928, + 0.0008721919148229063, + 0.0012297395151108503, + 0.0010927662951871753, + 0.0013122704112902284, + 0.0012973123230040073, + 0.0019308732589706779, + 0.0008371652220375836, + 0.0008733305730856955, + 0.000821812660433352, + 0.0007640354451723397, + 0.000541974266525358, + 0.0007867381209507585, + 0.0005659444723278284, + 0.0006922078318893909, + 0.0017016257625073195, + 0.0007709902711212635, + 0.003039450151845813, + 0.001008355524390936, + 0.0007589930901303887, + 0.0017995773814618587, + 0.002950014779344201, + 0.0011711616534739733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0549561008810997 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from a frontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene feature a futuristic nightscape with icy rain, towering skyscrapers, and magenta/cyan neon reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central female figure wearing a tight-fitting black metallic tactical suit with asymmetrical bright pink hair and glowing geometric circuitry on her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character burst forward from a 2D e-book cover into 3D space in an explosive pose holding a high-tech pistol?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the movement captured with smooth, high-frame-rate cinematography showing visible hair motion and clothing folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279808104038239 + ], + "opposite_logprob_tokens": [ + -2.0177981853485107 + ], + "logprob_reward": -0.14279808104038239, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14279808104038239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1024332046508789, + 0.7913540005683899, + 0.0218177679926157, + 0.031834617257118225, + 0.013115464709699154, + 0.006262924522161484, + 0.0117074279114604, + 0.00710905808955431, + 0.005374139174818993, + 0.008991463109850883 + ], + "credit_map": [ + 0.003922301810234785, + 0.008610636927187443, + 0.004952516872435808, + 0.023623771965503693, + 0.006399838719516993, + 0.013572610914707184, + 0.014348579570651054, + 0.027002934366464615, + 0.02832036092877388, + 0.01017184928059578, + 0.03749025613069534, + 0.01272880844771862, + 0.036866940557956696, + 0.2603292465209961, + 0.029709700495004654, + 0.3757367730140686, + 0.0020948811434209347, + 0.003243739251047373, + 0.0055506257340312, + 0.0017003925750032067, + 0.0027648196555674076, + 0.002225951524451375, + 0.0014738328754901886, + 0.0027635213918983936, + 0.0024777562357485294, + 0.0027306547854095697, + 0.009139078669250011, + 0.002515471074730158, + 0.0021712402813136578, + 0.007705135736614466, + 0.001647902769036591, + 0.003447376424446702, + 0.0012891448568552732, + 0.0012603605864569545, + 0.002180343260988593, + 0.0013860424514859915, + 0.001031109830364585, + 0.0016253339126706123, + 0.0013108632992953062, + 0.0030322647653520107, + 0.0013360664015635848, + 0.0007042621145956218, + 0.0009284742409363389, + 0.0005996814579702914, + 0.0005055941292084754, + 0.0008613422396592796, + 0.0005201788153499365, + 0.0008073243661783636, + 0.0011213283287361264, + 0.0016924377996474504, + 0.0016039827605709434, + 0.0020068574231117964, + 0.0012010697973892093, + 0.0017214257968589664, + 0.0013280484126880765, + 0.001032276195473969, + 0.000889850954990834, + 0.0009877171833068132, + 0.0007864089566282928, + 0.0008754616719670594, + 0.000801656162366271, + 0.0007474067970179021, + 0.0008683653431944549, + 0.0011521897977218032, + 0.0006555244908668101, + 0.0006721006357111037, + 0.0008089327602647245, + 0.0007097532507032156, + 0.0005689713871106505, + 0.0006263236864469945, + 0.0006281321402639151, + 0.0007044002413749695, + 0.001098194858059287, + 0.001035504974424839, + 0.0018702466040849686, + 0.0007658206159248948, + 0.0005211519310250878, + 0.001072951010428369, + 0.0017520354595035315, + 0.0008755561430007219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.14279808104038239 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_5/index_tensor([9606, 9606])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_5/index_tensor([9606, 9606])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f5dfdf0970289cd054e3a8fab507b9f4fc30ac59 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_5/index_tensor([9606, 9606])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera slowly and steadily pans to the right, maintaining a slightly elevated angle above the water. A sweeping wide-angle shot captures the scene. A sleek, modern luxury yacht, its bright white fiberglass hull gleaming in the warm light, glides effortlessly across the frame from left to right. It moves with majestic stability across the calm, glassy water, trailing a soft, V-shaped trail of white foam in its wake. The tranquil surface of the water acts like a perfect mirror, reflecting the magnificent colors of the sunset. The sky above is a breathtaking gradient, transitioning from a fiery red-orange near the horizon to a soft, rose-pink, finally fading into a deep, tranquil violet at the peaks. In the distant background, a series of majestic mountains stand out, their sharp, deep indigo outlines against the vibrant sky. To the left of the frame, near the shore, a cluster of small houses can be seen, their windows beginning to emit a warm, inviting glow. Near the left edge, a smaller, dark-hulled boat lies anchored, motionless, swaying almost imperceptibly on the gentle waves, providing a resting point for the yacht's steady motion. The entire scene is bathed in the soft golden light of the setting sun, creating an atmosphere of profound tranquility, majestic beauty.", + "seed": 10583, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_5/index_9606_seed_10583_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06026862901635468, + "vlm_logprob": -0.06026862901635468, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.06026862901635468, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007462127599865198 + ], + "opposite_logprob_tokens": [ + -5.257462024688721 + ], + "logprob_reward": -0.007462127599865198, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.007462127599865198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2144782692193985, + 0.03321626037359238, + 0.03170137107372284, + 0.05614601820707321, + 0.13170117139816284, + 0.02375044859945774, + 0.1465786248445511, + 0.12369116395711899, + 0.06296844780445099, + 0.1757681667804718 + ], + "credit_map": [ + 0.053136296570301056, + 0.034873731434345245, + 0.015477033331990242, + 0.013849953189492226, + 0.028615092858672142, + 0.03278467059135437, + 0.014940510503947735, + 0.020800983533263206, + 0.0036891759373247623, + 0.003478155704215169, + 0.005319083109498024, + 0.003051616484299302, + 0.0033970943186432123, + 0.0042878794483840466, + 0.005724587943404913, + 0.0042686681263148785, + 0.0027129286900162697, + 0.0028187616262584925, + 0.003713575191795826, + 0.0029668572824448347, + 0.004518239293247461, + 0.0036035729572176933, + 0.004740443546324968, + 0.006626991555094719, + 0.01139462273567915, + 0.01015531551092863, + 0.005225348751991987, + 0.004022601060569286, + 0.007407414726912975, + 0.0059383888728916645, + 0.005593144800513983, + 0.006409184075891972, + 0.005987559445202351, + 0.05502443388104439, + 0.0036140766460448503, + 0.0041090394370257854, + 0.04946383833885193, + 0.004899340681731701, + 0.005163151770830154, + 0.003439724212512374, + 0.003584835212677717, + 0.003651128616183996, + 0.0025388130452483892, + 0.00409866776317358, + 0.0030157717410475016, + 0.0021730922162532806, + 0.0026548265013843775, + 0.0020333135034888983, + 0.011014264076948166, + 0.03968106210231781, + 0.03663952276110649, + 0.019105849787592888, + 0.01156834326684475, + 0.010922678746283054, + 0.010657292790710926, + 0.006989601533859968, + 0.007568018976598978, + 0.012927848845720291, + 0.01364120002835989, + 0.02044498175382614, + 0.016571592539548874, + 0.011999803595244884, + 0.023409364745020866, + 0.01712835021317005, + 0.005336792673915625, + 0.006299890577793121, + 0.010812184773385525, + 0.011780099011957645, + 0.004372410010546446, + 0.005312182940542698, + 0.010593720711767673, + 0.008461165241897106, + 0.017674798145890236, + 0.018728360533714294, + 0.01999787427484989, + 0.016858648508787155, + 0.029207946732640266, + 0.0187950748950243, + 0.03122151643037796, + 0.02328394167125225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.007462127599865198 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813495993614197 + ], + "opposite_logprob_tokens": [ + -1.4063496589660645 + ], + "logprob_reward": -0.2813495993614197, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813495993614197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23976482450962067, + 0.050379037857055664, + 0.07786615937948227, + 0.05980157479643822, + 0.07405825704336166, + 0.053640179336071014, + 0.03904921934008598, + 0.13678325712680817, + 0.21605490148067474, + 0.052602533251047134 + ], + "credit_map": [ + 0.058372557163238525, + 0.03162973001599312, + 0.01867901347577572, + 0.019558101892471313, + 0.018541883677244186, + 0.03156888112425804, + 0.01983189396560192, + 0.041582778096199036, + 0.006048383191227913, + 0.006830334663391113, + 0.004684827756136656, + 0.006896871607750654, + 0.004914321471005678, + 0.007319784723222256, + 0.008216499350965023, + 0.005468018352985382, + 0.006603889632970095, + 0.006268246099352837, + 0.005821146070957184, + 0.005812714342027903, + 0.02141375094652176, + 0.013121308758854866, + 0.008012529462575912, + 0.01081258524209261, + 0.007773331366479397, + 0.010435841046273708, + 0.011693880893290043, + 0.005842227954417467, + 0.008039175532758236, + 0.007399380672723055, + 0.004398804157972336, + 0.004218935500830412, + 0.005350586026906967, + 0.027826596051454544, + 0.0034985514357686043, + 0.0037213678006082773, + 0.021971672773361206, + 0.005592441186308861, + 0.003379600355401635, + 0.0027174423448741436, + 0.005079695023596287, + 0.008266338147222996, + 0.0056194220669567585, + 0.008383375592529774, + 0.005286065395921469, + 0.004647236317396164, + 0.006174422334879637, + 0.010183632373809814, + 0.004368428606539965, + 0.005415743216872215, + 0.0033667003735899925, + 0.004534807987511158, + 0.0050817145965993404, + 0.006153556052595377, + 0.004720907658338547, + 0.005407366435974836, + 0.010964491404592991, + 0.012433616444468498, + 0.01718844287097454, + 0.024111997336149216, + 0.020133325830101967, + 0.011951255612075329, + 0.026524290442466736, + 0.013475845567882061, + 0.017083333805203438, + 0.01894484832882881, + 0.04477909207344055, + 0.0383429117500782, + 0.0157791581004858, + 0.019236775115132332, + 0.030234655365347862, + 0.03165414184331894, + 0.004283167887479067, + 0.006334928795695305, + 0.0103034982457757, + 0.006207169499248266, + 0.008409283123910427, + 0.005739189684391022, + 0.0055336132645606995, + 0.005791685078293085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.2813495993614197 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948222242295742 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005948222242295742, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948222242295742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18240545690059662, + 0.055678993463516235, + 0.09460859000682831, + 0.1645573079586029, + 0.20332419872283936, + 0.08047881722450256, + 0.10815832018852234, + 0.04022038355469704, + 0.0474129319190979, + 0.02315508760511875 + ], + "credit_map": [ + 0.043278224766254425, + 0.027054358273744583, + 0.012549202889204025, + 0.0120301004499197, + 0.01962396688759327, + 0.030492156744003296, + 0.014325346797704697, + 0.02305210754275322, + 0.007333735004067421, + 0.00853314995765686, + 0.005360739771276712, + 0.006329669617116451, + 0.006934558507055044, + 0.008729897439479828, + 0.006488644052296877, + 0.005968599114567041, + 0.009401698596775532, + 0.008882748894393444, + 0.0076546319760382175, + 0.005303151439875364, + 0.03023923933506012, + 0.016866976395249367, + 0.008245639503002167, + 0.008014500141143799, + 0.009522303007543087, + 0.054486218839883804, + 0.009840169921517372, + 0.006090277340263128, + 0.061929378658533096, + 0.00885731540620327, + 0.00756254093721509, + 0.006269094068557024, + 0.009880010969936848, + 0.07551174610853195, + 0.005323834717273712, + 0.00881118793040514, + 0.0795799046754837, + 0.009074878878891468, + 0.009188181720674038, + 0.005954443011432886, + 0.006403755862265825, + 0.008780776523053646, + 0.008290613070130348, + 0.01658852957189083, + 0.008468139916658401, + 0.006887170020490885, + 0.010540587827563286, + 0.014519243501126766, + 0.007760555017739534, + 0.027113301679491997, + 0.018138745799660683, + 0.011879312805831432, + 0.012644680216908455, + 0.012139033526182175, + 0.010595833882689476, + 0.007886849343776703, + 0.004318146966397762, + 0.0061263092793524265, + 0.004317506216466427, + 0.004802236799150705, + 0.004131744150072336, + 0.005133861675858498, + 0.006006208714097738, + 0.005384371615946293, + 0.004062285181134939, + 0.0037055984139442444, + 0.006851271726191044, + 0.010224821977317333, + 0.0036893479991704226, + 0.003937948495149612, + 0.007680857088416815, + 0.007260797079652548, + 0.002899362938478589, + 0.0024867376778274775, + 0.0023278819862753153, + 0.0031158786732703447, + 0.003108201315626502, + 0.0025391450617462397, + 0.003090166486799717, + 0.003587712999433279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.005948222242295742 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013367536012083292 + ], + "opposite_logprob_tokens": [ + -6.626336574554443 + ], + "logprob_reward": -0.0013367536012083292, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013367536012083292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07286713272333145, + 0.08665399998426437, + 0.1149420514702797, + 0.06470736861228943, + 0.47935259342193604, + 0.02425314113497734, + 0.06310846656560898, + 0.03570302575826645, + 0.02989094704389572, + 0.02852134220302105 + ], + "credit_map": [ + 0.013824598863720894, + 0.009233532473444939, + 0.006678506266325712, + 0.0072890049777925014, + 0.00943534541875124, + 0.009861537255346775, + 0.007465969771146774, + 0.009078639559447765, + 0.015098697505891323, + 0.010885626077651978, + 0.01214456558227539, + 0.009011693298816681, + 0.01063771452754736, + 0.013116953894495964, + 0.008487697690725327, + 0.0072710528038442135, + 0.01659597083926201, + 0.010555023327469826, + 0.014221844263374805, + 0.00755321467295289, + 0.021341953426599503, + 0.01743987761437893, + 0.013339630328118801, + 0.01389454398304224, + 0.008093753829598427, + 0.0077013070695102215, + 0.014064434915781021, + 0.00789124146103859, + 0.015349826775491238, + 0.0038804011419415474, + 0.003872527275234461, + 0.0038538798689842224, + 0.0233612097799778, + 0.19907692074775696, + 0.013356441631913185, + 0.014340001158416271, + 0.18169523775577545, + 0.01730303466320038, + 0.01842484064400196, + 0.011794887483119965, + 0.0028889174573123455, + 0.00328924972563982, + 0.002875832375138998, + 0.005277798976749182, + 0.0025361201260238886, + 0.001627736957743764, + 0.0027871178463101387, + 0.0029703681357204914, + 0.004965398460626602, + 0.015541454777121544, + 0.021500147879123688, + 0.010192984715104103, + 0.003130511147901416, + 0.0031358900014311075, + 0.002468980150297284, + 0.0021730978041887283, + 0.0035901623778045177, + 0.0036677152384072542, + 0.005294952541589737, + 0.004678758326917887, + 0.0054380944930016994, + 0.0038612422067672014, + 0.005009748041629791, + 0.004162350669503212, + 0.00258057564496994, + 0.002813626080751419, + 0.005803762469440699, + 0.00536772096529603, + 0.0024117359425872564, + 0.00288880942389369, + 0.0039956336840987206, + 0.004029083997011185, + 0.004697374999523163, + 0.003166633890941739, + 0.005036215763539076, + 0.0029915578197687864, + 0.0028740393463522196, + 0.0031521152704954147, + 0.0024203252978622913, + 0.0041830819100141525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.0013367536012083292 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005246442276984453 + ], + "opposite_logprob_tokens": [ + -5.255246639251709 + ], + "logprob_reward": -0.005246442276984453, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005246442276984453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23117601871490479, + 0.08188627660274506, + 0.03489169850945473, + 0.1496575027704239, + 0.19948990643024445, + 0.06314157694578171, + 0.03673557564616203, + 0.08338502794504166, + 0.062258802354335785, + 0.05737763270735741 + ], + "credit_map": [ + 0.05619410425424576, + 0.03400714695453644, + 0.01812068186700344, + 0.019016679376363754, + 0.02076936885714531, + 0.0323016531765461, + 0.019193507730960846, + 0.03157287836074829, + 0.010355217382311821, + 0.012411771342158318, + 0.011250758543610573, + 0.007592950016260147, + 0.010974423959851265, + 0.015165681950747967, + 0.007601710502058268, + 0.006533762905746698, + 0.0036299957428127527, + 0.0034624517429620028, + 0.004339657723903656, + 0.002532030688598752, + 0.004909670911729336, + 0.006016477011144161, + 0.005051656626164913, + 0.0049497573636472225, + 0.012683386914432049, + 0.021734002977609634, + 0.029213368892669678, + 0.014114034362137318, + 0.04859358072280884, + 0.009176050312817097, + 0.007409260608255863, + 0.006733819376677275, + 0.012922635301947594, + 0.0670153871178627, + 0.007915107533335686, + 0.007170744705945253, + 0.07569225877523422, + 0.011792967095971107, + 0.009525456465780735, + 0.007455333136022091, + 0.008707349188625813, + 0.009808109141886234, + 0.007655459921807051, + 0.011154474690556526, + 0.008205804973840714, + 0.006471144966781139, + 0.004750542808324099, + 0.006388696376234293, + 0.004801340401172638, + 0.005818141624331474, + 0.007348164916038513, + 0.004592267796397209, + 0.003592589870095253, + 0.003931188024580479, + 0.002953047165647149, + 0.0036988360807299614, + 0.01260435488075018, + 0.014497856609523296, + 0.008704889565706253, + 0.010113556869328022, + 0.009766806848347187, + 0.01343529298901558, + 0.006947694346308708, + 0.007314577698707581, + 0.0077441176399588585, + 0.008590350858867168, + 0.007402521092444658, + 0.007858891040086746, + 0.005994430277496576, + 0.01160902064293623, + 0.006248170509934425, + 0.0068112993612885475, + 0.008374853990972042, + 0.006774838548153639, + 0.0059641157276928425, + 0.006718125194311142, + 0.00772224273532629, + 0.0068823592737317085, + 0.006392530165612698, + 0.008548567071557045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.005246442276984453 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera slowly and steadily pans to the right, maintaining a slightly elevated angle above the water. A sweeping wide-angle shot captures the scene. A sleek, modern luxury yacht, its bright white fiberglass hull gleaming in the warm light, glides effortlessly across the frame from left to right. It moves with majestic stability across the calm, glassy water, trailing a soft, V-shaped trail of white foam in its wake. The tranquil surface of the water acts like a perfect mirror, reflecting the magnificent colors of the sunset. The sky above is a breathtaking gradient, transitioning from a fiery red-orange near the horizon to a soft, rose-pink, finally fading into a deep, tranquil violet at the peaks. In the distant background, a series of majestic mountains stand out, their sharp, deep indigo outlines against the vibrant sky. To the left of the frame, near the shore, a cluster of small houses can be seen, their windows beginning to emit a warm, inviting glow. Near the left edge, a smaller, dark-hulled boat lies anchored, motionless, swaying almost imperceptibly on the gentle waves, providing a resting point for the yacht's steady motion. The entire scene is bathed in the soft golden light of the setting sun, creating an atmosphere of profound tranquility, majestic beauty.", + "seed": 10584, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_5/index_9606_seed_10584_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.22713578240945936, + "vlm_logprob": -0.22713578240945936, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.22713578240945936, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8270805478096008 + ], + "opposite_logprob_tokens": [ + -0.5770805478096008 + ], + "logprob_reward": -0.8270805478096008, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8270805478096008, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22322723269462585, + 0.03309525176882744, + 0.024645710363984108, + 0.03701268136501312, + 0.03961000591516495, + 0.02043561078608036, + 0.11383680999279022, + 0.07346907258033752, + 0.038375742733478546, + 0.39629197120666504 + ], + "credit_map": [ + 0.026988795027136803, + 0.03073452226817608, + 0.038657113909721375, + 0.018614284694194794, + 0.03058420494198799, + 0.021109016612172127, + 0.018676383420825005, + 0.03786288946866989, + 0.003682136768475175, + 0.004706362262368202, + 0.004302732180804014, + 0.004358412232249975, + 0.0032213861122727394, + 0.004043987952172756, + 0.004100873600691557, + 0.004679359029978514, + 0.0027485608588904142, + 0.003447812283411622, + 0.0025131020229309797, + 0.002897696802392602, + 0.003322017379105091, + 0.00249181198887527, + 0.004464171826839447, + 0.002760533941909671, + 0.0031903323251754045, + 0.0065109520219266415, + 0.002090435940772295, + 0.0031630578450858593, + 0.010383395478129387, + 0.004882927052676678, + 0.004030443727970123, + 0.002761135110631585, + 0.004094522912055254, + 0.005916296970099211, + 0.0033698156476020813, + 0.0034946321975439787, + 0.007314040791243315, + 0.005892901215702295, + 0.006256780121475458, + 0.0032710100058466196, + 0.0022230499889701605, + 0.003397590247914195, + 0.0021790212485939264, + 0.0018044676398858428, + 0.004044555593281984, + 0.0023514502681791782, + 0.0028873446863144636, + 0.001548128086142242, + 0.019458701834082603, + 0.01807008869946003, + 0.018213748931884766, + 0.013457278721034527, + 0.016007712110877037, + 0.008877786807715893, + 0.011393954046070576, + 0.00835752859711647, + 0.010534953325986862, + 0.010460141114890575, + 0.019320553168654442, + 0.006846758536994457, + 0.007544186897575855, + 0.005824381485581398, + 0.007285999599844217, + 0.005652090534567833, + 0.005226494278758764, + 0.0047975885681807995, + 0.005196522455662489, + 0.0050464472733438015, + 0.0030859822873026133, + 0.005134425591677427, + 0.004804646130651236, + 0.005083633586764336, + 0.06747917830944061, + 0.06735216081142426, + 0.09584793448448181, + 0.06218205764889717, + 0.029658526182174683, + 0.021232696250081062, + 0.03132084757089615, + 0.02121850661933422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.8270805478096008 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28140467405319214 + ], + "opposite_logprob_tokens": [ + -1.406404733657837 + ], + "logprob_reward": -0.28140467405319214, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28140467405319214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08568255603313446, + 0.019101371988654137, + 0.08970725536346436, + 0.07976260036230087, + 0.030558759346604347, + 0.02653459832072258, + 0.10699676722288132, + 0.12158311158418655, + 0.03191336616873741, + 0.4081595838069916 + ], + "credit_map": [ + 0.009287196211516857, + 0.012097262777388096, + 0.010722764767706394, + 0.009876172058284283, + 0.0104249631986022, + 0.015282046049833298, + 0.009076866321265697, + 0.008915290236473083, + 0.0026939951349049807, + 0.0027413431089371443, + 0.002503243274986744, + 0.003396947169676423, + 0.0024345892015844584, + 0.0014517551753669977, + 0.0023696066346019506, + 0.0015098931035026908, + 0.008911254815757275, + 0.00930106546729803, + 0.010371703654527664, + 0.013938716612756252, + 0.010675908997654915, + 0.009483538568019867, + 0.016103819012641907, + 0.010921253822743893, + 0.010392396710813046, + 0.013133201748132706, + 0.010504134930670261, + 0.007369844242930412, + 0.010991139337420464, + 0.005664828233420849, + 0.013485260307788849, + 0.008221798576414585, + 0.005829061381518841, + 0.006137970834970474, + 0.004397734068334103, + 0.002736559836193919, + 0.0026623655576258898, + 0.002593024866655469, + 0.002783597679808736, + 0.0034184479154646397, + 0.0027800695970654488, + 0.0038615737576037645, + 0.0034225040581077337, + 0.002258548280224204, + 0.005152988713234663, + 0.00358297280035913, + 0.0031702511478215456, + 0.002305692760273814, + 0.018237391486763954, + 0.018050475046038628, + 0.016539912670850754, + 0.013229679316282272, + 0.014859749004244804, + 0.007573584094643593, + 0.010101412422955036, + 0.008404567837715149, + 0.013648295775055885, + 0.012924709357321262, + 0.013690618798136711, + 0.01796027272939682, + 0.01671035774052143, + 0.011905412189662457, + 0.02029464766383171, + 0.014448809437453747, + 0.004515667445957661, + 0.004915995057672262, + 0.007684915792196989, + 0.004390977323055267, + 0.0029659634456038475, + 0.002013107994571328, + 0.0029677057173103094, + 0.0024590366519987583, + 0.0670771673321724, + 0.06631479412317276, + 0.10096076875925064, + 0.06038147956132889, + 0.03092481754720211, + 0.023203307762742043, + 0.034149594604969025, + 0.02514767460525036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.28140467405319214 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607341133058071 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.01607341133058071, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607341133058071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18188151717185974, + 0.031082339584827423, + 0.0907105952501297, + 0.07323535531759262, + 0.04557276517152786, + 0.03639541566371918, + 0.2707497477531433, + 0.1391759216785431, + 0.03873587027192116, + 0.092460498213768 + ], + "credit_map": [ + 0.024186305701732635, + 0.022853754460811615, + 0.03184039145708084, + 0.01913963258266449, + 0.016530878841876984, + 0.023514175787568092, + 0.01630464754998684, + 0.02751171961426735, + 0.003098070388659835, + 0.004420854151248932, + 0.005108409561216831, + 0.0034209340810775757, + 0.004525086842477322, + 0.0036919584963470697, + 0.003880091942846775, + 0.0029369329567998648, + 0.008528674952685833, + 0.01082622166723013, + 0.00857547577470541, + 0.01468745805323124, + 0.009929844178259373, + 0.008793177083134651, + 0.018349021673202515, + 0.011020722799003124, + 0.008868314325809479, + 0.011966852471232414, + 0.009039136581122875, + 0.006608537398278713, + 0.011344783008098602, + 0.005899581126868725, + 0.012444819323718548, + 0.007063327357172966, + 0.005169119220227003, + 0.004761404823511839, + 0.005853299517184496, + 0.006011273246258497, + 0.005072614643722773, + 0.0062956479378044605, + 0.006473714020103216, + 0.0059356908313930035, + 0.004468675237149, + 0.003904846031218767, + 0.00536442780867219, + 0.005212701857089996, + 0.005153396166861057, + 0.0044850935228168964, + 0.004701407626271248, + 0.003104865550994873, + 0.03944241255521774, + 0.0408577136695385, + 0.04043864086270332, + 0.03875897824764252, + 0.0331464447081089, + 0.020735278725624084, + 0.03237738832831383, + 0.024992885068058968, + 0.026440532878041267, + 0.029485011473298073, + 0.016371406614780426, + 0.014400765299797058, + 0.017924752086400986, + 0.007781520951539278, + 0.017551271244883537, + 0.009220669977366924, + 0.005519661121070385, + 0.006773768458515406, + 0.004060604143887758, + 0.004131359979510307, + 0.005553725175559521, + 0.004347956273704767, + 0.004951443988829851, + 0.003397352062165737, + 0.015247088856995106, + 0.01661626063287258, + 0.020882615819573402, + 0.014032748527824879, + 0.008294928818941116, + 0.005494407378137112, + 0.007079905364662409, + 0.0048125372268259525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.01607341133058071 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00248968368396163 + ], + "opposite_logprob_tokens": [ + -6.0024895668029785 + ], + "logprob_reward": -0.00248968368396163, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00248968368396163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24174390733242035, + 0.09440919756889343, + 0.04308260604739189, + 0.06330284476280212, + 0.06560876965522766, + 0.037301741540431976, + 0.08587198704481125, + 0.08993390202522278, + 0.03681531548500061, + 0.24192973971366882 + ], + "credit_map": [ + 0.027151992544531822, + 0.03707057982683182, + 0.03366934880614281, + 0.020371297374367714, + 0.04308664798736572, + 0.022751949727535248, + 0.02139899507164955, + 0.03624310716986656, + 0.01766100525856018, + 0.0169718936085701, + 0.01260635070502758, + 0.008694099262356758, + 0.014389187097549438, + 0.00882597267627716, + 0.00788132008165121, + 0.007379372604191303, + 0.006095521617680788, + 0.0060838572680950165, + 0.005792167037725449, + 0.005431185010820627, + 0.005655386485159397, + 0.004375968128442764, + 0.00443803146481514, + 0.0052104913629591465, + 0.012363860383629799, + 0.01351278368383646, + 0.006924147251993418, + 0.006017671432346106, + 0.006956381723284721, + 0.005155094433575869, + 0.005719615146517754, + 0.006653284654021263, + 0.012371120043098927, + 0.010608801618218422, + 0.005790595430880785, + 0.005084665957838297, + 0.01186488289386034, + 0.005714451428502798, + 0.00820806436240673, + 0.0059661865234375, + 0.006864376366138458, + 0.00727075943723321, + 0.003089293371886015, + 0.0022763805463910103, + 0.007177797611802816, + 0.003758254460990429, + 0.0037720026448369026, + 0.003092879429459572, + 0.013588902540504932, + 0.014668204821646214, + 0.011037931777536869, + 0.01152979489415884, + 0.013637781143188477, + 0.00558127649128437, + 0.00943478848785162, + 0.0063933078199625015, + 0.013777850195765495, + 0.014475548639893532, + 0.016843097284436226, + 0.009733179584145546, + 0.01154405903071165, + 0.006646645721048117, + 0.008776508271694183, + 0.00813701469451189, + 0.004958098754286766, + 0.005726121366024017, + 0.007288628723472357, + 0.00508125452324748, + 0.0037278980016708374, + 0.002829750068485737, + 0.004149323329329491, + 0.0030542416498064995, + 0.04205629229545593, + 0.04350413382053375, + 0.056941717863082886, + 0.03147990256547928, + 0.021532827988266945, + 0.014237201772630215, + 0.019302139058709145, + 0.012875516898930073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.00248968368396163 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008630595169961452 + ], + "opposite_logprob_tokens": [ + -4.758630752563477 + ], + "logprob_reward": -0.008630595169961452, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008630595169961452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21646758913993835, + 0.08834796398878098, + 0.053668782114982605, + 0.02661973610520363, + 0.10677695274353027, + 0.03258475661277771, + 0.13498568534851074, + 0.08135867118835449, + 0.04818550497293472, + 0.2110043615102768 + ], + "credit_map": [ + 0.0223378986120224, + 0.0309830904006958, + 0.028504744172096252, + 0.023235216736793518, + 0.042832791805267334, + 0.01905684731900692, + 0.025166548788547516, + 0.02435043826699257, + 0.009222712367773056, + 0.012431441806256771, + 0.01306233648210764, + 0.01071014441549778, + 0.012441108003258705, + 0.009158391505479813, + 0.011256848461925983, + 0.010064984671771526, + 0.006697422359138727, + 0.006390175316482782, + 0.006980021018534899, + 0.00785026140511036, + 0.008057599887251854, + 0.00469624251127243, + 0.006428033113479614, + 0.006569029297679663, + 0.0036551528610289097, + 0.004035403952002525, + 0.0030000617261976004, + 0.0026602307334542274, + 0.005152223631739616, + 0.003238895907998085, + 0.0030226754024624825, + 0.001855093170888722, + 0.010428345762193203, + 0.013069897890090942, + 0.012566243298351765, + 0.012578753754496574, + 0.014603175222873688, + 0.01214332040399313, + 0.014629035256803036, + 0.016758183017373085, + 0.004552794620394707, + 0.005849950946867466, + 0.0036778387147933245, + 0.00327644357457757, + 0.004803670570254326, + 0.0035124975256621838, + 0.003762786975130439, + 0.003148771822452545, + 0.02611456997692585, + 0.03288416191935539, + 0.010895618237555027, + 0.006949098315089941, + 0.029637545347213745, + 0.00963615719228983, + 0.011664144694805145, + 0.007204389199614525, + 0.014660203829407692, + 0.015948479995131493, + 0.009314277209341526, + 0.008330749347805977, + 0.01252139825373888, + 0.005666178651154041, + 0.008256345055997372, + 0.006661034654825926, + 0.0070729246363043785, + 0.006947868037968874, + 0.007025226019322872, + 0.006944036111235619, + 0.004806112963706255, + 0.005082644056528807, + 0.005872447043657303, + 0.00443424703553319, + 0.03609825298190117, + 0.03613001108169556, + 0.05147991701960564, + 0.03031514585018158, + 0.017107965424656868, + 0.01216564979404211, + 0.01606174372136593, + 0.011645694263279438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.008630595169961452 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera slowly and steadily pans to the right, maintaining a slightly elevated angle above the water. A sweeping wide-angle shot captures the scene. A sleek, modern luxury yacht, its bright white fiberglass hull gleaming in the warm light, glides effortlessly across the frame from left to right. It moves with majestic stability across the calm, glassy water, trailing a soft, V-shaped trail of white foam in its wake. The tranquil surface of the water acts like a perfect mirror, reflecting the magnificent colors of the sunset. The sky above is a breathtaking gradient, transitioning from a fiery red-orange near the horizon to a soft, rose-pink, finally fading into a deep, tranquil violet at the peaks. In the distant background, a series of majestic mountains stand out, their sharp, deep indigo outlines against the vibrant sky. To the left of the frame, near the shore, a cluster of small houses can be seen, their windows beginning to emit a warm, inviting glow. Near the left edge, a smaller, dark-hulled boat lies anchored, motionless, swaying almost imperceptibly on the gentle waves, providing a resting point for the yacht's steady motion. The entire scene is bathed in the soft golden light of the setting sun, creating an atmosphere of profound tranquility, majestic beauty.", + "seed": 10585, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_5/index_9606_seed_10585_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06049276061821729, + "vlm_logprob": -0.06049276061821729, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.06049276061821729, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0035388958640396595 + ], + "opposite_logprob_tokens": [ + -6.628539085388184 + ], + "logprob_reward": -0.0035388958640396595, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0035388958640396595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07819663733243942, + 0.055515918880701065, + 0.057782188057899475, + 0.12508170306682587, + 0.12891171872615814, + 0.143592968583107, + 0.10166226327419281, + 0.08204085379838943, + 0.039716221392154694, + 0.187499538064003 + ], + "credit_map": [ + 0.007702534552663565, + 0.01014372892677784, + 0.011467893607914448, + 0.009755845181643963, + 0.010664082132279873, + 0.00803462415933609, + 0.010312928818166256, + 0.010115006007254124, + 0.005683210678398609, + 0.010725111700594425, + 0.004022662993520498, + 0.0059651401825249195, + 0.009467181749641895, + 0.008520647883415222, + 0.006506867706775665, + 0.004625093657523394, + 0.006186161655932665, + 0.01275885570794344, + 0.004703952465206385, + 0.0048156320117414, + 0.010254433378577232, + 0.006845536641776562, + 0.006339473649859428, + 0.005878141615539789, + 0.0096127949655056, + 0.011461739428341389, + 0.03818307816982269, + 0.01653340831398964, + 0.007190002594143152, + 0.011645055375993252, + 0.016647087410092354, + 0.013808535411953926, + 0.010407732799649239, + 0.017507774755358696, + 0.03863603249192238, + 0.015370219014585018, + 0.009341602213680744, + 0.007509930990636349, + 0.01799812540411949, + 0.012140298262238503, + 0.020432963967323303, + 0.011556414887309074, + 0.031898487359285355, + 0.01349143497645855, + 0.01779298298060894, + 0.023405296728014946, + 0.013354301452636719, + 0.011661093682050705, + 0.01177690178155899, + 0.021808777004480362, + 0.011281995102763176, + 0.018149681389331818, + 0.01082669012248516, + 0.006996855139732361, + 0.012749643996357918, + 0.008071718737483025, + 0.007218529935926199, + 0.017270546406507492, + 0.01713145524263382, + 0.01304639782756567, + 0.007424602285027504, + 0.004725142382085323, + 0.008747144602239132, + 0.006477041635662317, + 0.004806788172572851, + 0.00500436732545495, + 0.009700390510261059, + 0.006009154953062534, + 0.003303953679278493, + 0.003047622973099351, + 0.004616217687726021, + 0.0032277239952236414, + 0.023893389850854874, + 0.04210377112030983, + 0.0580422542989254, + 0.017829149961471558, + 0.015589648857712746, + 0.008792589418590069, + 0.012656313367187977, + 0.008592422120273113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.0035388958640396595 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813335955142975 + ], + "opposite_logprob_tokens": [ + -1.406333565711975 + ], + "logprob_reward": -0.2813335955142975, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813335955142975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08221551775932312, + 0.03111613355576992, + 0.07104629278182983, + 0.11355137079954147, + 0.08151112496852875, + 0.39549729228019714, + 0.0625891163945198, + 0.04500091075897217, + 0.07107310742139816, + 0.04639912024140358 + ], + "credit_map": [ + 0.008791433647274971, + 0.01592041738331318, + 0.009509516879916191, + 0.0075239138677716255, + 0.011457270011305809, + 0.010073027573525906, + 0.008105110377073288, + 0.010834828950464725, + 0.0028742391150444746, + 0.005985379219055176, + 0.0029521621763706207, + 0.0023596312385052443, + 0.005004304926842451, + 0.004770136438310146, + 0.0037194567266851664, + 0.003450825810432434, + 0.007071743719279766, + 0.013845876790583134, + 0.008231301791965961, + 0.010212906636297703, + 0.007634596899151802, + 0.005930543877184391, + 0.009026186540722847, + 0.009093135595321655, + 0.009571780450642109, + 0.0070631010457873344, + 0.034960243850946426, + 0.014987535774707794, + 0.007688434328883886, + 0.010989361442625523, + 0.013855424709618092, + 0.014435484074056149, + 0.005524865351617336, + 0.01739032194018364, + 0.02274191565811634, + 0.009188869968056679, + 0.004857698455452919, + 0.004284885246306658, + 0.009367319755256176, + 0.008155247196555138, + 0.05736321210861206, + 0.0422767698764801, + 0.12324069440364838, + 0.03340695798397064, + 0.03262653946876526, + 0.05035163089632988, + 0.025787169113755226, + 0.030444344505667686, + 0.007983841001987457, + 0.015229595825076103, + 0.011849550530314445, + 0.007042651064693928, + 0.006223499774932861, + 0.0050166333094239235, + 0.005462738685309887, + 0.003780604572966695, + 0.00543062062934041, + 0.008034419268369675, + 0.010718886740505695, + 0.006843362469226122, + 0.0033716775942593813, + 0.002643426414579153, + 0.004294029902666807, + 0.0036644882056862116, + 0.007549866568297148, + 0.01047301385551691, + 0.008188396692276001, + 0.015290613286197186, + 0.005947920493781567, + 0.004746823105961084, + 0.010019621811807156, + 0.008856849744915962, + 0.0052659776993095875, + 0.008601767010986805, + 0.006078704725950956, + 0.005801370833069086, + 0.004612460266798735, + 0.004525423515588045, + 0.006694409530609846, + 0.004819002468138933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.2813335955142975 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637567982077599 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.008637567982077599, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637567982077599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10506893694400787, + 0.07615893334150314, + 0.0558713935315609, + 0.06749282777309418, + 0.0792819932103157, + 0.13823020458221436, + 0.06867789477109909, + 0.038579102605581284, + 0.05902045592665672, + 0.31161829829216003 + ], + "credit_map": [ + 0.009717429988086224, + 0.0175099465996027, + 0.010991370305418968, + 0.011255654506385326, + 0.014329438097774982, + 0.021824035793542862, + 0.008135015144944191, + 0.01130603812634945, + 0.0068148113787174225, + 0.019959881901741028, + 0.010780520737171173, + 0.006175024434924126, + 0.009381983429193497, + 0.007336454000324011, + 0.006854786537587643, + 0.008855471387505531, + 0.0071133882738649845, + 0.014518823474645615, + 0.004488248843699694, + 0.00472291698679328, + 0.008756542578339577, + 0.006836198270320892, + 0.0035227967891842127, + 0.005912470631301403, + 0.007266562897711992, + 0.008393933065235615, + 0.008941181004047394, + 0.008981864899396896, + 0.0068788547068834305, + 0.0100506991147995, + 0.00731712905690074, + 0.009662594646215439, + 0.010847368277609348, + 0.023768121376633644, + 0.014838662929832935, + 0.005582695361226797, + 0.005777824204415083, + 0.005258753430098295, + 0.0056048547849059105, + 0.007603711448609829, + 0.016138942912220955, + 0.017742011696100235, + 0.047011762857437134, + 0.013072692789137363, + 0.009994004853069782, + 0.01519829872995615, + 0.00900760293006897, + 0.010064894333481789, + 0.008738012984395027, + 0.01636982336640358, + 0.01121418084949255, + 0.0064786272123456, + 0.006557498127222061, + 0.008207950741052628, + 0.006258625071495771, + 0.0048531703650951385, + 0.003239960176870227, + 0.005221258383244276, + 0.006777468603104353, + 0.006053426768630743, + 0.0037138580810278654, + 0.0028572003357112408, + 0.005737972911447287, + 0.004977952688932419, + 0.0076367976143956184, + 0.010012817569077015, + 0.006598250940442085, + 0.012560289353132248, + 0.004974640905857086, + 0.004754496738314629, + 0.007109450176358223, + 0.00537370890378952, + 0.042839985340833664, + 0.07268740981817245, + 0.08981731534004211, + 0.028600899502635002, + 0.0254229623824358, + 0.016398346051573753, + 0.022176498547196388, + 0.013674856163561344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.008637567982077599 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001335086883045733 + ], + "opposite_logprob_tokens": [ + -6.626335144042969 + ], + "logprob_reward": -0.001335086883045733, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001335086883045733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17716296017169952, + 0.055035050958395004, + 0.058244045823812485, + 0.15846043825149536, + 0.10699206590652466, + 0.15620465576648712, + 0.08448955416679382, + 0.045775026082992554, + 0.10797704756259918, + 0.04965914040803909 + ], + "credit_map": [ + 0.017528222873806953, + 0.023899026215076447, + 0.01598561555147171, + 0.017075728625059128, + 0.03615761175751686, + 0.025060156360268593, + 0.022297417744994164, + 0.019159194082021713, + 0.0046830312348902225, + 0.006457701791077852, + 0.007475461810827255, + 0.006048315670341253, + 0.007767548784613609, + 0.007086944300681353, + 0.007442840840667486, + 0.008073205128312111, + 0.004998732358217239, + 0.009564234875142574, + 0.007063109893351793, + 0.007740881759673357, + 0.0077922833152115345, + 0.00596710154786706, + 0.007695140317082405, + 0.007422559894621372, + 0.019999228417873383, + 0.01336495392024517, + 0.055318865925073624, + 0.01380317471921444, + 0.009296434931457043, + 0.017682503908872604, + 0.012143531814217567, + 0.016851741820573807, + 0.008007209748029709, + 0.024800246581435204, + 0.035051267594099045, + 0.010394973680377007, + 0.006398133002221584, + 0.005766922142356634, + 0.00815646629780531, + 0.008416850119829178, + 0.022333817556500435, + 0.02475973218679428, + 0.051073312759399414, + 0.01439411286264658, + 0.009708008728921413, + 0.012484372593462467, + 0.01036361139267683, + 0.011087705381214619, + 0.009193086065351963, + 0.0116912005469203, + 0.015759076923131943, + 0.019262157380580902, + 0.006042277440428734, + 0.005871260538697243, + 0.010601187124848366, + 0.006069310940802097, + 0.00552916107699275, + 0.0066267880611121655, + 0.00911062490195036, + 0.007478825747966766, + 0.0038719780277460814, + 0.003420802066102624, + 0.005400372669100761, + 0.004336473532021046, + 0.012892649509012699, + 0.021109096705913544, + 0.014649329707026482, + 0.018604019656777382, + 0.008563017472624779, + 0.007190658710896969, + 0.013516983948647976, + 0.01145129557698965, + 0.005992614198476076, + 0.006815916392952204, + 0.010774146765470505, + 0.0060800532810389996, + 0.006235551089048386, + 0.0030030054040253162, + 0.007531318813562393, + 0.0032265360932797194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.001335086883045733 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007618656847625971 + ], + "opposite_logprob_tokens": [ + -4.882618427276611 + ], + "logprob_reward": -0.007618656847625971, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007618656847625971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10267110913991928, + 0.07955098897218704, + 0.08242076635360718, + 0.07240954786539078, + 0.11095910519361496, + 0.15797340869903564, + 0.1309618502855301, + 0.21248728036880493, + 0.024422986432909966, + 0.026142947375774384 + ], + "credit_map": [ + 0.009317588061094284, + 0.012829883024096489, + 0.012394167482852936, + 0.01581137254834175, + 0.010886873118579388, + 0.020674990490078926, + 0.008224076591432095, + 0.012532173655927181, + 0.00639014458283782, + 0.017097486183047295, + 0.010674596764147282, + 0.007676086854189634, + 0.009715565480291843, + 0.00632553081959486, + 0.011079490184783936, + 0.01059209369122982, + 0.0090089812874794, + 0.01374074537307024, + 0.00884495209902525, + 0.009277054108679295, + 0.011702295392751694, + 0.009282301180064678, + 0.009960195049643517, + 0.010604252107441425, + 0.006488078273832798, + 0.005744450725615025, + 0.02169964089989662, + 0.009376512840390205, + 0.0055087958462536335, + 0.008099070750176907, + 0.007367739453911781, + 0.00812526885420084, + 0.007673036307096481, + 0.023788899183273315, + 0.042619261890649796, + 0.010687715373933315, + 0.007156341336667538, + 0.004022654611617327, + 0.006954605225473642, + 0.008056610822677612, + 0.023007933050394058, + 0.013092135079205036, + 0.03782512620091438, + 0.014553993940353394, + 0.017577869817614555, + 0.025812461972236633, + 0.013078413903713226, + 0.013025498948991299, + 0.015630288049578667, + 0.022724609822034836, + 0.01889902539551258, + 0.028484228998422623, + 0.009970438666641712, + 0.009655403904616833, + 0.014136657118797302, + 0.011461209505796432, + 0.022997990250587463, + 0.040507540106773376, + 0.031200634315609932, + 0.035813093185424805, + 0.021127210929989815, + 0.014227952808141708, + 0.022794896736741066, + 0.023817989975214005, + 0.0027552242390811443, + 0.003791113616898656, + 0.0043554180301725864, + 0.003267725696787238, + 0.002395824296399951, + 0.0017913805786520243, + 0.003618111601099372, + 0.0024481911677867174, + 0.0036656036972999573, + 0.004523256793618202, + 0.005007975734770298, + 0.0031276026275008917, + 0.0031751259230077267, + 0.0016664909198880196, + 0.0032843775115907192, + 0.001692514750175178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.007618656847625971 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera slowly and steadily pans to the right, maintaining a slightly elevated angle above the water. A sweeping wide-angle shot captures the scene. A sleek, modern luxury yacht, its bright white fiberglass hull gleaming in the warm light, glides effortlessly across the frame from left to right. It moves with majestic stability across the calm, glassy water, trailing a soft, V-shaped trail of white foam in its wake. The tranquil surface of the water acts like a perfect mirror, reflecting the magnificent colors of the sunset. The sky above is a breathtaking gradient, transitioning from a fiery red-orange near the horizon to a soft, rose-pink, finally fading into a deep, tranquil violet at the peaks. In the distant background, a series of majestic mountains stand out, their sharp, deep indigo outlines against the vibrant sky. To the left of the frame, near the shore, a cluster of small houses can be seen, their windows beginning to emit a warm, inviting glow. Near the left edge, a smaller, dark-hulled boat lies anchored, motionless, swaying almost imperceptibly on the gentle waves, providing a resting point for the yacht's steady motion. The entire scene is bathed in the soft golden light of the setting sun, creating an atmosphere of profound tranquility, majestic beauty.", + "seed": 10586, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_5/index_9606_seed_10586_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09220415726304054, + "vlm_logprob": -0.09220415726304054, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.09220415726304054, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.049580786377191544 + ], + "opposite_logprob_tokens": [ + -3.0495808124542236 + ], + "logprob_reward": -0.049580786377191544, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.049580786377191544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.396761953830719, + 0.06370773911476135, + 0.010676335543394089, + 0.0758950486779213, + 0.028763750568032265, + 0.15073412656784058, + 0.14923511445522308, + 0.05424027144908905, + 0.04765491932630539, + 0.02233077958226204 + ], + "credit_map": [ + 0.02752482146024704, + 0.02811039797961712, + 0.04159993305802345, + 0.07293111830949783, + 0.04338628798723221, + 0.031052257865667343, + 0.09475553035736084, + 0.05740157514810562, + 0.004419535864144564, + 0.005206541158258915, + 0.007517210673540831, + 0.010283940471708775, + 0.006959406193345785, + 0.004832100123167038, + 0.01704203151166439, + 0.0074469707906246185, + 0.001153722871094942, + 0.0011530346237123013, + 0.0011023455299437046, + 0.0013703800505027175, + 0.001410762662999332, + 0.0014620593283325434, + 0.0016472918214276433, + 0.0013767374912276864, + 0.00731826713308692, + 0.005257136188447475, + 0.009254607371985912, + 0.011787394993007183, + 0.009543359279632568, + 0.005682453978806734, + 0.017266185954213142, + 0.009785627014935017, + 0.0020911486353725195, + 0.0025920106563717127, + 0.0026396322064101696, + 0.004251356702297926, + 0.004680287092924118, + 0.0038139454554766417, + 0.004998167511075735, + 0.003697197185829282, + 0.01096603088080883, + 0.012279978953301907, + 0.017607521265745163, + 0.02198483794927597, + 0.023486880585551262, + 0.014979398809373379, + 0.03053048625588417, + 0.018898973241448402, + 0.01147019024938345, + 0.01170023251324892, + 0.014168735593557358, + 0.028418274596333504, + 0.0179018285125494, + 0.012830287218093872, + 0.03755320608615875, + 0.015192345716059208, + 0.0031344559974968433, + 0.004014210309833288, + 0.005238169804215431, + 0.009908411651849747, + 0.007088558282703161, + 0.005413089878857136, + 0.012986414134502411, + 0.006456954404711723, + 0.002828531665727496, + 0.004654104355722666, + 0.004137814976274967, + 0.006974403280764818, + 0.008458680473268032, + 0.004758140537887812, + 0.009273450821638107, + 0.0065697873942554, + 0.0014209917280822992, + 0.0017058821395039558, + 0.0017354871379211545, + 0.003352227620780468, + 0.00300985318608582, + 0.0017412068555131555, + 0.004111865535378456, + 0.0052532609552145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.049580786377191544 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871018886566162 + ], + "opposite_logprob_tokens": [ + -1.1371018886566162 + ], + "logprob_reward": -0.3871018886566162, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871018886566162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3373849093914032, + 0.05500440672039986, + 0.026847904548048973, + 0.18535040318965912, + 0.024653486907482147, + 0.21308982372283936, + 0.10190407931804657, + 0.025914933532476425, + 0.021832283586263657, + 0.008017772808670998 + ], + "credit_map": [ + 0.024358531460165977, + 0.023677224293351173, + 0.032266244292259216, + 0.06288592517375946, + 0.037900954484939575, + 0.027791274711489677, + 0.08173755556344986, + 0.0467672124505043, + 0.006240291986614466, + 0.004871671088039875, + 0.005269082263112068, + 0.007463173009455204, + 0.007757625076919794, + 0.007940518669784069, + 0.009242490865290165, + 0.0062195537611842155, + 0.0024803411215543747, + 0.0026263482868671417, + 0.0029859058558940887, + 0.003785195294767618, + 0.004016268067061901, + 0.003725724993273616, + 0.00396230723708868, + 0.0032658115960657597, + 0.020213982090353966, + 0.012957969680428505, + 0.02473800629377365, + 0.027849186211824417, + 0.022391369566321373, + 0.014814967289566994, + 0.03948574885725975, + 0.022899175062775612, + 0.0017941605765372515, + 0.002139275660738349, + 0.0035226631443947554, + 0.004255479201674461, + 0.002691597444936633, + 0.002595227211713791, + 0.004250336438417435, + 0.0034047476947307587, + 0.016260946169495583, + 0.016964862123131752, + 0.026266956701874733, + 0.03149979189038277, + 0.03214830905199051, + 0.02262689732015133, + 0.041288018226623535, + 0.026034047827124596, + 0.008022880181670189, + 0.008137645199894905, + 0.009853322058916092, + 0.018872667104005814, + 0.01128050684928894, + 0.009584512561559677, + 0.025318963453173637, + 0.010833575390279293, + 0.0017113317735493183, + 0.002315895864740014, + 0.0025040856562554836, + 0.00453768577426672, + 0.002939317375421524, + 0.002797258086502552, + 0.005828055087476969, + 0.0032813050784170628, + 0.0014077725354582071, + 0.00205765338614583, + 0.0020889765582978725, + 0.0029824431985616684, + 0.0038778798189014196, + 0.00213692057877779, + 0.0044943951070308685, + 0.0027862428687512875, + 0.000654239731375128, + 0.0007680103299207985, + 0.0009622922516427934, + 0.0012029367499053478, + 0.001043409574776888, + 0.000700253585819155, + 0.0013623040867969394, + 0.0013243264984339476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.3871018886566162 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076745577156544 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.011076745577156544, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076745577156544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1874741017818451, + 0.047584131360054016, + 0.08438150584697723, + 0.08495616912841797, + 0.08706650137901306, + 0.17821936309337616, + 0.15581928193569183, + 0.034896641969680786, + 0.05611792951822281, + 0.08348432183265686 + ], + "credit_map": [ + 0.010620648972690105, + 0.014254044741392136, + 0.02112637460231781, + 0.03785282373428345, + 0.01857542246580124, + 0.011415619403123856, + 0.05097949877381325, + 0.02264968492090702, + 0.003834093688055873, + 0.005177699960768223, + 0.00638611800968647, + 0.007376913912594318, + 0.005456821992993355, + 0.0026779244653880596, + 0.011254887096583843, + 0.005419674329459667, + 0.010718166828155518, + 0.008917040191590786, + 0.009450985118746758, + 0.010143047198653221, + 0.012654104270040989, + 0.010842839255928993, + 0.011148061603307724, + 0.010507267899811268, + 0.007338556926697493, + 0.0068865567445755005, + 0.009566589258611202, + 0.016340060159564018, + 0.0101691959425807, + 0.008613537065684795, + 0.01637110859155655, + 0.009670570492744446, + 0.005213118623942137, + 0.003850935259833932, + 0.019697364419698715, + 0.010483363643288612, + 0.007157675921916962, + 0.006821128539741039, + 0.01511220633983612, + 0.018730703741312027, + 0.01203528605401516, + 0.015050041489303112, + 0.020440388470888138, + 0.029927458614110947, + 0.028389498591423035, + 0.018280139192938805, + 0.03240468353033066, + 0.02169189043343067, + 0.011577127501368523, + 0.01794898323714733, + 0.01963314414024353, + 0.02793927676975727, + 0.016184596344828606, + 0.020993715152144432, + 0.02515205182135105, + 0.01639038883149624, + 0.003081055823713541, + 0.0038325712084770203, + 0.0034066596999764442, + 0.005729049444198608, + 0.0051937703974545, + 0.004945766646414995, + 0.004593169316649437, + 0.0041146003641188145, + 0.004778747446835041, + 0.006042073015123606, + 0.005706624127924442, + 0.008609386160969734, + 0.00829567201435566, + 0.006817793007940054, + 0.007963472977280617, + 0.007904165424406528, + 0.00554008549079299, + 0.006643375847488642, + 0.006321365013718605, + 0.013549561612308025, + 0.010325866751372814, + 0.008580180816352367, + 0.013997373171150684, + 0.01852651871740818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.011076745577156544 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002197353169322014 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.002197353169322014, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002197353169322014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20082944631576538, + 0.06423690170049667, + 0.10001374036073685, + 0.13395921885967255, + 0.044527530670166016, + 0.10672760754823685, + 0.10124808549880981, + 0.0861547440290451, + 0.11760981380939484, + 0.04469294100999832 + ], + "credit_map": [ + 0.018587229773402214, + 0.01735760271549225, + 0.02565188705921173, + 0.027270246297121048, + 0.025750577449798584, + 0.019215894863009453, + 0.044452160596847534, + 0.022543838247656822, + 0.007476855535060167, + 0.005513489246368408, + 0.00893467292189598, + 0.011133837513625622, + 0.007022361736744642, + 0.005874553229659796, + 0.011097528040409088, + 0.0071836006827652454, + 0.011452470906078815, + 0.009826784953474998, + 0.016499536111950874, + 0.012650273740291595, + 0.011317326687276363, + 0.013096889480948448, + 0.013447130098938942, + 0.011723319999873638, + 0.01273251697421074, + 0.014187297783792019, + 0.014215433970093727, + 0.016290858387947083, + 0.021352363750338554, + 0.017739862203598022, + 0.022422611713409424, + 0.015018259175121784, + 0.00401691161096096, + 0.004818753805011511, + 0.006935028359293938, + 0.0055107856169342995, + 0.004147375468164682, + 0.003919021226465702, + 0.006351632531732321, + 0.008828016929328442, + 0.008836735971271992, + 0.009585008956491947, + 0.012988177128136158, + 0.01835029572248459, + 0.01424493920058012, + 0.012490957044064999, + 0.017974892631173134, + 0.012256594374775887, + 0.00795810017734766, + 0.009524636901915073, + 0.010547851212322712, + 0.018935952335596085, + 0.013090413995087147, + 0.008826646953821182, + 0.02145485207438469, + 0.010909630917012691, + 0.006054793018847704, + 0.005419920198619366, + 0.00863958615809679, + 0.014603533782064915, + 0.010606946423649788, + 0.007983014918863773, + 0.021417269483208656, + 0.011429672129452229, + 0.008370642550289631, + 0.012800868600606918, + 0.011467985808849335, + 0.015577682293951511, + 0.020033441483974457, + 0.01268423069268465, + 0.022590383887290955, + 0.014084573835134506, + 0.006238063797354698, + 0.006532753352075815, + 0.004006457515060902, + 0.004560531582683325, + 0.006891387049108744, + 0.005010910797864199, + 0.00639741774648428, + 0.005055417772382498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.002197353169322014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a sleek, modern luxury yacht with a bright white fiberglass hull visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the yacht glide from left to right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera pan slowly and steadily to the right while maintaining a slightly elevated angle above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky a gradient transitioning from red-orange near the horizon to rose-pink and then deep violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, deep indigo mountain outlines visible in the distant background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011064012534916401 + ], + "opposite_logprob_tokens": [ + -4.511064052581787 + ], + "logprob_reward": -0.011064012534916401, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011064012534916401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34518879652023315, + 0.029826482757925987, + 0.05229434743523598, + 0.05039894953370094, + 0.08128606528043747, + 0.20339268445968628, + 0.06737590581178665, + 0.04403049126267433, + 0.0935983881354332, + 0.03260791301727295 + ], + "credit_map": [ + 0.022489020600914955, + 0.025405278429389, + 0.03628662973642349, + 0.06326606869697571, + 0.0371316634118557, + 0.025838695466518402, + 0.08503345400094986, + 0.04973798617720604, + 0.0035589123144745827, + 0.0024361396208405495, + 0.004005759488791227, + 0.0043597775511443615, + 0.0027572757098823786, + 0.0033965674228966236, + 0.005381576716899872, + 0.003930474165827036, + 0.006499854847788811, + 0.004540872760117054, + 0.005693440325558186, + 0.007154492661356926, + 0.006120325997471809, + 0.00667557492852211, + 0.008134362287819386, + 0.0074754199013113976, + 0.005478520877659321, + 0.005968967918306589, + 0.006127609871327877, + 0.006475489120930433, + 0.006389190908521414, + 0.0062813591212034225, + 0.008061138913035393, + 0.005616671871393919, + 0.005447076633572578, + 0.00453113904222846, + 0.01723417267203331, + 0.009991263039410114, + 0.006127016618847847, + 0.006555613595992327, + 0.014671112410724163, + 0.016728678718209267, + 0.01805466040968895, + 0.022821933031082153, + 0.022729603573679924, + 0.03004295937716961, + 0.021279603242874146, + 0.02821939066052437, + 0.03775898367166519, + 0.02248554676771164, + 0.005625638645142317, + 0.007329421583563089, + 0.007419211324304342, + 0.007536786608397961, + 0.010761728510260582, + 0.008895372971892357, + 0.011528384871780872, + 0.00827935803681612, + 0.003577160183340311, + 0.0034159976057708263, + 0.004145575687289238, + 0.007666077930480242, + 0.005834407173097134, + 0.0036869191098958254, + 0.010436062701046467, + 0.005268290638923645, + 0.008979850448668003, + 0.008361432701349258, + 0.009132171981036663, + 0.012343606911599636, + 0.012111599557101727, + 0.011051129549741745, + 0.017425093799829483, + 0.01419350691139698, + 0.00310410070233047, + 0.0034745235461741686, + 0.002639644779264927, + 0.0048560891300439835, + 0.00326372217386961, + 0.0045389411970973015, + 0.004180158954113722, + 0.006550733931362629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.011064012534916401 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_6/index_tensor([9896, 9896])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_6/index_tensor([9896, 9896])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fcb15ef17874360de376f506b98491cade06c8ef --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_6/index_tensor([9896, 9896])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-frame-rate close-up of a competitive game, captured in powerful slow motion, focuses on a hand, marked by the marks of time, yet steady and powerful. Between its slender thumb and index finger, this hand elegantly and decisively grasps a smooth black Go piece, its surface reflecting a jade-like luster under the soft, diffused sunlight. With a slight twist of the wrist, the piece is steadily and deliberately moved toward an ancient wooden chessboard near the water's edge, landing precisely on a star position with a crisp, resonant thud. The minute vibration of this movement instantly stirs up extremely fine ripples on the calm surface of a small, crystal-clear rock pond near the edge of the chessboard. These ripples, clearly visible in slow motion, intertwine with the reflection of the smooth pebbles underwater. In the background, the other end of the chessboard and the blurred figure of the opponent are blurred by the shallow depth of field. The whole scene presents an extreme sense of concentration, tranquility and inner strength in the cold tones, perfectly capturing the tense atmosphere of the duel between masters where one move decides the outcome.", + "seed": 239, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_6/index_9896_seed_239_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07809835523366929, + "vlm_logprob": -0.07809835523366929, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.07809835523366929, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012277974747121334 + ], + "opposite_logprob_tokens": [ + -6.012278079986572 + ], + "logprob_reward": -0.012277974747121334, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.012277974747121334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14141945540905, + 0.15764689445495605, + 0.04815709590911865, + 0.020845923572778702, + 0.04580249264836311, + 0.058024968951940536, + 0.18467094004154205, + 0.034995194524526596, + 0.2834111750125885, + 0.025025827810168266 + ], + "credit_map": [ + 0.014917613007128239, + 0.01505222450941801, + 0.014886018820106983, + 0.014153183437883854, + 0.017196882516145706, + 0.020061705261468887, + 0.016550250351428986, + 0.02860158123075962, + 0.014073843136429787, + 0.02474108524620533, + 0.02009325660765171, + 0.012199543416500092, + 0.02950628101825714, + 0.026426438242197037, + 0.013426794670522213, + 0.017179660499095917, + 0.004390248097479343, + 0.0037678920198231936, + 0.010528435930609703, + 0.008399021811783314, + 0.003941152710467577, + 0.004076831042766571, + 0.006635705940425396, + 0.006417808122932911, + 0.0017430701991543174, + 0.0020370502024888992, + 0.004108747933059931, + 0.0025219612289220095, + 0.0035752274561673403, + 0.0016299334820359945, + 0.0020853905007243156, + 0.0031445410568267107, + 0.006885843817144632, + 0.005111475009471178, + 0.006524225231260061, + 0.004275631159543991, + 0.006590103264898062, + 0.004257570952177048, + 0.00554312951862812, + 0.006614512763917446, + 0.0057733068242669106, + 0.0050208414904773235, + 0.010809273459017277, + 0.004859476815909147, + 0.011054877191781998, + 0.006895379163324833, + 0.0054148598574101925, + 0.00819694809615612, + 0.016992365941405296, + 0.019754959270358086, + 0.027817150577902794, + 0.016201570630073547, + 0.027241231873631477, + 0.022760527208447456, + 0.021251777186989784, + 0.03265136107802391, + 0.0038901735097169876, + 0.003004102036356926, + 0.007141788490116596, + 0.0040852525271475315, + 0.004574764519929886, + 0.00515112467110157, + 0.0031239932868629694, + 0.004023993853479624, + 0.03220351040363312, + 0.017820313572883606, + 0.03322073444724083, + 0.02125069685280323, + 0.02289852499961853, + 0.10132358223199844, + 0.018293771892786026, + 0.036400068551301956, + 0.002460296032950282, + 0.003566436003893614, + 0.0033964463509619236, + 0.0023050198797136545, + 0.00453978031873703, + 0.0033736522309482098, + 0.0032113485503941774, + 0.002172849839553237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.012277974747121334 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906727284193039 + ], + "opposite_logprob_tokens": [ + -2.464067220687866 + ], + "logprob_reward": -0.08906727284193039, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906727284193039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1413215696811676, + 0.07638633996248245, + 0.2256307601928711, + 0.09969079494476318, + 0.06237709894776344, + 0.020075123757123947, + 0.25962358713150024, + 0.03595476225018501, + 0.047694213688373566, + 0.03124573826789856 + ], + "credit_map": [ + 0.015329936519265175, + 0.015058819204568863, + 0.021788625046610832, + 0.019602922722697258, + 0.01561881322413683, + 0.019704241305589676, + 0.018302755430340767, + 0.015915459021925926, + 0.006460169330239296, + 0.006163895130157471, + 0.01042107306420803, + 0.011411145329475403, + 0.011634481139481068, + 0.009774656035006046, + 0.00902073085308075, + 0.01150018721818924, + 0.022771555930376053, + 0.02079768292605877, + 0.04913133382797241, + 0.02951202355325222, + 0.025053858757019043, + 0.02098025567829609, + 0.02803015150129795, + 0.0293539147824049, + 0.007785744033753872, + 0.008743640966713428, + 0.02215588092803955, + 0.009262443520128727, + 0.016258979216217995, + 0.007710774429142475, + 0.00966698955744505, + 0.018106337636709213, + 0.009139849804341793, + 0.006661435589194298, + 0.008122045546770096, + 0.006206544116139412, + 0.008281257003545761, + 0.00658407574519515, + 0.009580148383975029, + 0.007801741361618042, + 0.0019775822293013334, + 0.0021848217584192753, + 0.0029878385830670595, + 0.0012252958258613944, + 0.0033832418266683817, + 0.002729675965383649, + 0.001521472237072885, + 0.004065195098519325, + 0.02369534969329834, + 0.022007033228874207, + 0.056921396404504776, + 0.020605331286787987, + 0.021640190854668617, + 0.024307364597916603, + 0.023907270282506943, + 0.06653964519500732, + 0.004571736790239811, + 0.0029725651256740093, + 0.003324284218251705, + 0.0022319667041301727, + 0.00367092527449131, + 0.010525280609726906, + 0.0026443549431860447, + 0.006013647653162479, + 0.005311600398272276, + 0.003822054946795106, + 0.00457619596272707, + 0.006148382090032101, + 0.005130151752382517, + 0.011594436131417751, + 0.006566975731402636, + 0.004544418770819902, + 0.002879892708733678, + 0.004196468275040388, + 0.003700598841533065, + 0.00281030242331326, + 0.0061700185760855675, + 0.004404562991112471, + 0.0034813517704606056, + 0.0036025429144501686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.08906727284193039 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20153720676898956 + ], + "opposite_logprob_tokens": [ + -1.7015372514724731 + ], + "logprob_reward": -0.20153720676898956, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20153720676898956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21642251312732697, + 0.034084029495716095, + 0.05562853813171387, + 0.16294069588184357, + 0.18745309114456177, + 0.029810933396220207, + 0.0909142941236496, + 0.15079234540462494, + 0.043424881994724274, + 0.028528718277812004 + ], + "credit_map": [ + 0.020335782319307327, + 0.02149036154150963, + 0.04977342113852501, + 0.02161591872572899, + 0.027508985251188278, + 0.030984006822109222, + 0.02031465619802475, + 0.02439938299357891, + 0.0032518573570996523, + 0.0033569501247256994, + 0.004454782232642174, + 0.005791360046714544, + 0.00502530112862587, + 0.0028975647874176502, + 0.0037202853709459305, + 0.005585929378867149, + 0.006327152252197266, + 0.005849368404597044, + 0.007318652234971523, + 0.008273874409496784, + 0.004709475673735142, + 0.004654355812817812, + 0.006771693006157875, + 0.011723966337740421, + 0.008737105876207352, + 0.008331743068993092, + 0.05762171000242233, + 0.012332137674093246, + 0.034103620797395706, + 0.008212094195187092, + 0.01195246446877718, + 0.021649828180670738, + 0.02685767225921154, + 0.02152092754840851, + 0.027107462286949158, + 0.01915554329752922, + 0.02772601507604122, + 0.016999850049614906, + 0.025201793760061264, + 0.02288382686674595, + 0.0033325799740850925, + 0.0029847044497728348, + 0.004099148325622082, + 0.0027780933305621147, + 0.003984929993748665, + 0.004286794923245907, + 0.003172538708895445, + 0.005172142293304205, + 0.009239754639565945, + 0.00852876529097557, + 0.017649315297603607, + 0.011347352527081966, + 0.012546466663479805, + 0.009862920269370079, + 0.011154713109135628, + 0.010584999807178974, + 0.016342202201485634, + 0.009283333085477352, + 0.024420101195573807, + 0.007110341917723417, + 0.010912214405834675, + 0.040273942053318024, + 0.009511022828519344, + 0.032939184457063675, + 0.004645136184990406, + 0.0032869980204850435, + 0.0034483959898352623, + 0.004343722946941853, + 0.004498444031924009, + 0.011680619791150093, + 0.005566905718296766, + 0.005954657215625048, + 0.0026298065204173326, + 0.0038424234371632338, + 0.003260925877839327, + 0.0029337089508771896, + 0.005382720846682787, + 0.004212663974612951, + 0.003386500058695674, + 0.0028799681458622217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.20153720676898956 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789678543806076 + ], + "opposite_logprob_tokens": [ + -2.578967809677124 + ], + "logprob_reward": -0.0789678543806076, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789678543806076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21366117894649506, + 0.04245688393712044, + 0.14944371581077576, + 0.08229588717222214, + 0.1502838134765625, + 0.04242350906133652, + 0.13020887970924377, + 0.12174145132303238, + 0.05219047889113426, + 0.015294189564883709 + ], + "credit_map": [ + 0.023226423189044, + 0.026881076395511627, + 0.03603995963931084, + 0.02561330981552601, + 0.025300193578004837, + 0.03276760131120682, + 0.022772450000047684, + 0.021060168743133545, + 0.003467731876298785, + 0.003619183087721467, + 0.005014623515307903, + 0.0065641398541629314, + 0.0076823849231004715, + 0.004927859641611576, + 0.005429533775895834, + 0.005751425866037607, + 0.01700253039598465, + 0.017985232174396515, + 0.02598530799150467, + 0.020470723509788513, + 0.01609261892735958, + 0.01378338411450386, + 0.01858375035226345, + 0.019540172070264816, + 0.005180725362151861, + 0.005046873353421688, + 0.02629643678665161, + 0.006722571328282356, + 0.01737470179796219, + 0.005402682814747095, + 0.007149974349886179, + 0.009121921844780445, + 0.01690096966922283, + 0.010784734040498734, + 0.029504425823688507, + 0.014298268593847752, + 0.018603721633553505, + 0.02316112443804741, + 0.022256582975387573, + 0.01477398443967104, + 0.003494550473988056, + 0.0026576046366244555, + 0.00894262082874775, + 0.004803219344466925, + 0.0049903178587555885, + 0.005265119019895792, + 0.0048378934152424335, + 0.007432182785123587, + 0.010558666661381721, + 0.012796079739928246, + 0.017560333013534546, + 0.014513691887259483, + 0.017510443925857544, + 0.0161602720618248, + 0.015130600892007351, + 0.025978798046708107, + 0.015254714526236057, + 0.006876136176288128, + 0.016361216083168983, + 0.0051395162008702755, + 0.008148446679115295, + 0.036324311047792435, + 0.007629691623151302, + 0.02600741572678089, + 0.003780831815674901, + 0.002144722966477275, + 0.01677069626748562, + 0.0054812440648674965, + 0.002836797386407852, + 0.012881276197731495, + 0.004479116294533014, + 0.0038157920353114605, + 0.0018473421223461628, + 0.002208521356806159, + 0.002025625901296735, + 0.0013200658140704036, + 0.0020182495936751366, + 0.0018959036096930504, + 0.001974817831069231, + 0.0020036627538502216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.0789678543806076 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641467429697514 + ], + "opposite_logprob_tokens": [ + -4.758641242980957 + ], + "logprob_reward": -0.008641467429697514, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641467429697514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13804097473621368, + 0.07329513132572174, + 0.0635584145784378, + 0.08233009278774261, + 0.12008548527956009, + 0.10115696489810944, + 0.31313401460647583, + 0.04709935188293457, + 0.03655709698796272, + 0.024742476642131805 + ], + "credit_map": [ + 0.014466620050370693, + 0.01631448045372963, + 0.02609701082110405, + 0.014616669155657291, + 0.017554238438606262, + 0.02165212295949459, + 0.012766952626407146, + 0.014572889544069767, + 0.00823336374014616, + 0.0116542037576437, + 0.013492685742676258, + 0.007589464541524649, + 0.009586269967257977, + 0.00805099867284298, + 0.006165346130728722, + 0.008522812277078629, + 0.008976747281849384, + 0.007715523708611727, + 0.010839986614882946, + 0.007199241314083338, + 0.007914756424725056, + 0.004929665010422468, + 0.006630830001085997, + 0.009351667948067188, + 0.006539607420563698, + 0.00803798995912075, + 0.019814131781458855, + 0.007192963734269142, + 0.012106020003557205, + 0.008493753150105476, + 0.00816334318369627, + 0.011982292868196964, + 0.013662493787705898, + 0.016262566670775414, + 0.024537470191717148, + 0.01096649281680584, + 0.012930495664477348, + 0.010666732676327229, + 0.01598675549030304, + 0.01507248729467392, + 0.008590802550315857, + 0.012589987367391586, + 0.016030723229050636, + 0.006711085792630911, + 0.01948191039264202, + 0.013690012507140636, + 0.007700446527451277, + 0.016362009570002556, + 0.03496246784925461, + 0.03644537553191185, + 0.059550799429416656, + 0.03708465397357941, + 0.03990549221634865, + 0.02645127661526203, + 0.04415079578757286, + 0.0345831997692585, + 0.007472246885299683, + 0.0034403991885483265, + 0.004386961925774813, + 0.0033457616809755564, + 0.004726730287075043, + 0.012539250776171684, + 0.0040650092996656895, + 0.007122995797544718, + 0.002872980199754238, + 0.002379682846367359, + 0.007967938669025898, + 0.003857614705339074, + 0.0041325767524540424, + 0.008092985488474369, + 0.0039459592662751675, + 0.0033073618542402983, + 0.004436816088855267, + 0.0025848066434264183, + 0.003936602268368006, + 0.001980541041120887, + 0.0033604330383241177, + 0.0028453527484089136, + 0.0028290599584579468, + 0.0027688692789524794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.008641467429697514 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-frame-rate close-up of a competitive game, captured in powerful slow motion, focuses on a hand, marked by the marks of time, yet steady and powerful. Between its slender thumb and index finger, this hand elegantly and decisively grasps a smooth black Go piece, its surface reflecting a jade-like luster under the soft, diffused sunlight. With a slight twist of the wrist, the piece is steadily and deliberately moved toward an ancient wooden chessboard near the water's edge, landing precisely on a star position with a crisp, resonant thud. The minute vibration of this movement instantly stirs up extremely fine ripples on the calm surface of a small, crystal-clear rock pond near the edge of the chessboard. These ripples, clearly visible in slow motion, intertwine with the reflection of the smooth pebbles underwater. In the background, the other end of the chessboard and the blurred figure of the opponent are blurred by the shallow depth of field. The whole scene presents an extreme sense of concentration, tranquility and inner strength in the cold tones, perfectly capturing the tense atmosphere of the duel between masters where one move decides the outcome.", + "seed": 240, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_6/index_9896_seed_240_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07217459818348289, + "vlm_logprob": -0.07217459818348289, + "vlm_yes_no_margin": 3.374999976158142, + "vlm_token_credit": -0.07217459818348289, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020166831091046333 + ], + "opposite_logprob_tokens": [ + -5.520166873931885 + ], + "logprob_reward": -0.020166831091046333, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.020166831091046333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09788964688777924, + 0.16151081025600433, + 0.1477271169424057, + 0.1729014366865158, + 0.07004786282777786, + 0.032121751457452774, + 0.08605615049600601, + 0.09867550432682037, + 0.06089397519826889, + 0.07217573374509811 + ], + "credit_map": [ + 0.007849859073758125, + 0.012975234538316727, + 0.014662633650004864, + 0.013426634483039379, + 0.015483023598790169, + 0.009034345857799053, + 0.008853737264871597, + 0.015604179352521896, + 0.035128649324178696, + 0.056090086698532104, + 0.00769819738343358, + 0.007545653730630875, + 0.014048287644982338, + 0.012328127399086952, + 0.007858791388571262, + 0.02081301249563694, + 0.012283245101571083, + 0.025311902165412903, + 0.008818482980132103, + 0.006419968791306019, + 0.031248128041625023, + 0.03142068162560463, + 0.008843052200973034, + 0.023381652310490608, + 0.027081293985247612, + 0.024155763909220695, + 0.015560206025838852, + 0.012477731332182884, + 0.028765173628926277, + 0.026350369676947594, + 0.023529421538114548, + 0.014981481246650219, + 0.020462673157453537, + 0.02306867204606533, + 0.0043085794895887375, + 0.0034678317606449127, + 0.004902432207018137, + 0.0045228684321045876, + 0.0036993110552430153, + 0.005615489557385445, + 0.004620697814971209, + 0.0038464537356048822, + 0.004076263401657343, + 0.005452833138406277, + 0.003286853665485978, + 0.0036176959984004498, + 0.0037678575608879328, + 0.0034530966076999903, + 0.023211892694234848, + 0.029897384345531464, + 0.0037636233028024435, + 0.0038731826934963465, + 0.004298361483961344, + 0.007867609150707722, + 0.0053663006983697414, + 0.007777801714837551, + 0.020733851939439774, + 0.0211134385317564, + 0.005585059989243746, + 0.006265637464821339, + 0.014830959029495716, + 0.013197135180234909, + 0.00802687183022499, + 0.008922554552555084, + 0.004364322870969772, + 0.006862654350697994, + 0.0072593679651618, + 0.012538899667561054, + 0.007330536376684904, + 0.007100138813257217, + 0.006088323891162872, + 0.009349728934466839, + 0.005410871468484402, + 0.006977809127420187, + 0.008593806996941566, + 0.014728994108736515, + 0.008869005367159843, + 0.008073536679148674, + 0.008330097422003746, + 0.011191604658961296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.020166831091046333 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900255173444748 + ], + "opposite_logprob_tokens": [ + -2.579002618789673 + ], + "logprob_reward": -0.07900255173444748, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900255173444748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1399073600769043, + 0.1244615763425827, + 0.11524057388305664, + 0.15352779626846313, + 0.06944859027862549, + 0.1276148557662964, + 0.0681803822517395, + 0.06341920793056488, + 0.07471128553152084, + 0.06348837912082672 + ], + "credit_map": [ + 0.015042251907289028, + 0.01628127694129944, + 0.016930105164647102, + 0.013165678828954697, + 0.016562815755605698, + 0.017870035022497177, + 0.02514142543077469, + 0.01891377754509449, + 0.015560484491288662, + 0.011090349406003952, + 0.016801826655864716, + 0.014240461401641369, + 0.011041367426514626, + 0.02188349887728691, + 0.010352798737585545, + 0.023490797728300095, + 0.011799956671893597, + 0.019057421013712883, + 0.008552866987884045, + 0.005683116614818573, + 0.02261388674378395, + 0.0251119676977396, + 0.007904013618826866, + 0.01451734360307455, + 0.04137958586215973, + 0.04429423063993454, + 0.010102668777108192, + 0.008123699575662613, + 0.013599824160337448, + 0.012070881202816963, + 0.013961647637188435, + 0.00999525934457779, + 0.015883760526776314, + 0.018717672675848007, + 0.0035605088341981173, + 0.00385680771432817, + 0.005307952873408794, + 0.008713546209037304, + 0.005884133744984865, + 0.00752421235665679, + 0.04750943183898926, + 0.05378922075033188, + 0.0036567815113812685, + 0.003349220845848322, + 0.0042349714785814285, + 0.003024416510015726, + 0.00394998025149107, + 0.00810082908719778, + 0.007215025834739208, + 0.006102443672716618, + 0.008508062921464443, + 0.007040036842226982, + 0.010209969244897366, + 0.011925866827368736, + 0.009018764831125736, + 0.008160209283232689, + 0.011792957782745361, + 0.011427969671785831, + 0.005213110242038965, + 0.005154675804078579, + 0.008524376899003983, + 0.008954833261668682, + 0.00545797822996974, + 0.006893303710967302, + 0.004982593469321728, + 0.009892381727695465, + 0.006535998545587063, + 0.01455298624932766, + 0.012915805913507938, + 0.007511083502322435, + 0.0077768913470208645, + 0.010543541051447392, + 0.006163270678371191, + 0.005157647654414177, + 0.01078889612108469, + 0.015997085720300674, + 0.004907823633402586, + 0.006355386693030596, + 0.005326839163899422, + 0.008791429921984673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.07900255173444748 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270359754562378 + ], + "opposite_logprob_tokens": [ + -2.1270360946655273 + ], + "logprob_reward": -0.1270359754562378, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270359754562378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10409372299909592, + 0.06182956323027611, + 0.07975826412439346, + 0.1339375227689743, + 0.4372192919254303, + 0.061392877250909805, + 0.040853824466466904, + 0.020179061219096184, + 0.026250744238495827, + 0.03448515385389328 + ], + "credit_map": [ + 0.008915976621210575, + 0.009587561711668968, + 0.015539505518972874, + 0.017189592123031616, + 0.012314590625464916, + 0.011574910022318363, + 0.014992015436291695, + 0.013979572802782059, + 0.006806704215705395, + 0.005837800446897745, + 0.005966581869870424, + 0.007236625533550978, + 0.010554888285696507, + 0.012100286781787872, + 0.005834558978676796, + 0.007492118049412966, + 0.008412903174757957, + 0.01327255368232727, + 0.008027586154639721, + 0.007647912483662367, + 0.008521636947989464, + 0.01243163738399744, + 0.007856392301619053, + 0.013587642461061478, + 0.022224796935915947, + 0.024157455191016197, + 0.015034361742436886, + 0.012839010916650295, + 0.010999045334756374, + 0.013535249046981335, + 0.011112295091152191, + 0.024035314098000526, + 0.16707336902618408, + 0.18679994344711304, + 0.012352412566542625, + 0.00881770346313715, + 0.014784809201955795, + 0.008835072629153728, + 0.009263665415346622, + 0.029292287304997444, + 0.019235951825976372, + 0.02135235071182251, + 0.0040299962274730206, + 0.004262229893356562, + 0.0033220630139112473, + 0.0031997202895581722, + 0.002218834822997451, + 0.0037717297673225403, + 0.010317938402295113, + 0.012418673373758793, + 0.002734162611886859, + 0.0030367288272827864, + 0.0021211188286542892, + 0.0035457410849630833, + 0.003083599964156747, + 0.0035958625376224518, + 0.0017546414164826274, + 0.0017511524492874742, + 0.0031553322914987803, + 0.003298697993159294, + 0.0026990671176463366, + 0.0025459877215325832, + 0.0021187083330005407, + 0.0028554743621498346, + 0.0016664060531184077, + 0.002896577352657914, + 0.0035890289582312107, + 0.0054192133247852325, + 0.004315690603107214, + 0.002148152096197009, + 0.002987262327224016, + 0.003228413173928857, + 0.002946860156953335, + 0.004480676259845495, + 0.004327993839979172, + 0.00496493699029088, + 0.004738825839012861, + 0.003480583196505904, + 0.0049459608271718025, + 0.004599316976964474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.1270359754562378 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270361989736557 + ], + "opposite_logprob_tokens": [ + -2.1270360946655273 + ], + "logprob_reward": -0.1270361989736557, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270361989736557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08021469414234161, + 0.20894990861415863, + 0.1847591996192932, + 0.15579205751419067, + 0.14472995698451996, + 0.06322009861469269, + 0.020100045949220657, + 0.03491327539086342, + 0.05990774184465408, + 0.04741304740309715 + ], + "credit_map": [ + 0.012802422046661377, + 0.01038662064820528, + 0.007641795091331005, + 0.007411017548292875, + 0.009711159393191338, + 0.010001776739954948, + 0.008974078111350536, + 0.01328581478446722, + 0.03373125568032265, + 0.08072561770677567, + 0.006358469370752573, + 0.006124301813542843, + 0.015125478617846966, + 0.016134269535541534, + 0.006659702397882938, + 0.044090818613767624, + 0.017711589112877846, + 0.024140560999512672, + 0.016869952902197838, + 0.012037528678774834, + 0.04056824743747711, + 0.041325442492961884, + 0.012123861350119114, + 0.01998201385140419, + 0.04559916630387306, + 0.05237284675240517, + 0.007899271324276924, + 0.009340254589915276, + 0.009855786338448524, + 0.008947068825364113, + 0.007896020077168941, + 0.013881642371416092, + 0.04136433079838753, + 0.049812160432338715, + 0.008980072103440762, + 0.008541258983314037, + 0.006820560898631811, + 0.010370220057666302, + 0.007760454900562763, + 0.011080889962613583, + 0.017179852351546288, + 0.019770769402384758, + 0.0034977197647094727, + 0.004198912996798754, + 0.0032952348701655865, + 0.0035269942600280046, + 0.004462011158466339, + 0.00728860218077898, + 0.0024725422263145447, + 0.0023199105635285378, + 0.0019512538565322757, + 0.001975782448425889, + 0.0017881442327052355, + 0.002538782311603427, + 0.0035799904726445675, + 0.0034736408852040768, + 0.003074032487347722, + 0.0025372845120728016, + 0.00477377837523818, + 0.004982108250260353, + 0.003994652070105076, + 0.007040897849947214, + 0.003085298463702202, + 0.005425224546343088, + 0.003995860926806927, + 0.007617370691150427, + 0.004327178932726383, + 0.009137419052422047, + 0.010492225177586079, + 0.0062383245676755905, + 0.007581240963190794, + 0.010518119670450687, + 0.00410707900300622, + 0.004091927781701088, + 0.007263181731104851, + 0.011315270327031612, + 0.004630133043974638, + 0.00451212702319026, + 0.004845861811190844, + 0.0066474685445427895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.1270361989736557 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631433662027121 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631433662027121, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631433662027121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14130988717079163, + 0.1189565360546112, + 0.045558203011751175, + 0.08271268755197525, + 0.18398979306221008, + 0.03127530962228775, + 0.02272186055779457, + 0.03887636587023735, + 0.18294069170951843, + 0.15165869891643524 + ], + "credit_map": [ + 0.01802236959338188, + 0.019698085263371468, + 0.010842765681445599, + 0.012082855217158794, + 0.026708504185080528, + 0.014900144189596176, + 0.016709914430975914, + 0.022345256060361862, + 0.013803763315081596, + 0.03844214975833893, + 0.00788390077650547, + 0.007603947538882494, + 0.010552608408033848, + 0.011537068523466587, + 0.006394848693162203, + 0.022738248109817505, + 0.00467636389657855, + 0.01040433719754219, + 0.005302339792251587, + 0.0029274539556354284, + 0.003734951140359044, + 0.004459862597286701, + 0.0037768592592328787, + 0.010276034474372864, + 0.011162430047988892, + 0.010315582156181335, + 0.010791318491101265, + 0.00752319535240531, + 0.0110709797590971, + 0.01005852036178112, + 0.012006738223135471, + 0.009783925488591194, + 0.056382860988378525, + 0.06429723650217056, + 0.007702856790274382, + 0.008226748555898666, + 0.007351010572165251, + 0.01221696101129055, + 0.010876702144742012, + 0.016935423016548157, + 0.008415092714130878, + 0.009126036427915096, + 0.0022894912399351597, + 0.0029284032061696053, + 0.0018990554381161928, + 0.0020940802060067654, + 0.001964423805475235, + 0.0025587230920791626, + 0.0034174765460193157, + 0.004392253700643778, + 0.001494634780101478, + 0.001767895300872624, + 0.003158130683004856, + 0.0034039111342281103, + 0.0028201621025800705, + 0.00226739514619112, + 0.008711503818631172, + 0.00918494164943695, + 0.0022238106466829777, + 0.0025207132566720247, + 0.004171199630945921, + 0.004680446349084377, + 0.0034808253403753042, + 0.0039029247127473354, + 0.013683084398508072, + 0.02762967348098755, + 0.010885381139814854, + 0.017881933599710464, + 0.03920666500926018, + 0.019676920026540756, + 0.02127966657280922, + 0.03269735723733902, + 0.01273549348115921, + 0.013534718193113804, + 0.02162775583565235, + 0.03285781294107437, + 0.015140699222683907, + 0.015712793916463852, + 0.016947677358984947, + 0.023101747035980225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.007631433662027121 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-frame-rate close-up of a competitive game, captured in powerful slow motion, focuses on a hand, marked by the marks of time, yet steady and powerful. Between its slender thumb and index finger, this hand elegantly and decisively grasps a smooth black Go piece, its surface reflecting a jade-like luster under the soft, diffused sunlight. With a slight twist of the wrist, the piece is steadily and deliberately moved toward an ancient wooden chessboard near the water's edge, landing precisely on a star position with a crisp, resonant thud. The minute vibration of this movement instantly stirs up extremely fine ripples on the calm surface of a small, crystal-clear rock pond near the edge of the chessboard. These ripples, clearly visible in slow motion, intertwine with the reflection of the smooth pebbles underwater. In the background, the other end of the chessboard and the blurred figure of the opponent are blurred by the shallow depth of field. The whole scene presents an extreme sense of concentration, tranquility and inner strength in the cold tones, perfectly capturing the tense atmosphere of the duel between masters where one move decides the outcome.", + "seed": 241, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_6/index_9896_seed_241_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05969896260648966, + "vlm_logprob": -0.05969896260648966, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.05969896260648966, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.047851670533418655 + ], + "opposite_logprob_tokens": [ + -3.4228515625 + ], + "logprob_reward": -0.047851670533418655, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.047851670533418655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3682253062725067, + 0.04270273074507713, + 0.04813152179121971, + 0.04483484849333763, + 0.10869063436985016, + 0.1276673823595047, + 0.012952231802046299, + 0.0957934558391571, + 0.027657821774482727, + 0.1233440414071083 + ], + "credit_map": [ + 0.03135259449481964, + 0.03770881146192551, + 0.034990329295396805, + 0.05081355571746826, + 0.0535161979496479, + 0.03844839334487915, + 0.054016225039958954, + 0.06737922877073288, + 0.005133927799761295, + 0.004159983713179827, + 0.005945770535618067, + 0.005410433746874332, + 0.004380785394459963, + 0.006589321419596672, + 0.005359416361898184, + 0.0057230982929468155, + 0.004807818681001663, + 0.0055692726746201515, + 0.004965425468981266, + 0.010372990742325783, + 0.0035176854580640793, + 0.003573754569515586, + 0.009303472936153412, + 0.006021106615662575, + 0.004259362351149321, + 0.003791841911152005, + 0.0038035933393985033, + 0.005717134103178978, + 0.004798229318112135, + 0.005095349159091711, + 0.0068731652572751045, + 0.010496174916625023, + 0.0057263607159256935, + 0.005982837174087763, + 0.009558289311826229, + 0.03294725343585014, + 0.006168977357447147, + 0.02508910745382309, + 0.010157442651689053, + 0.01306038349866867, + 0.008126689121127129, + 0.009472673758864403, + 0.026285802945494652, + 0.0363183468580246, + 0.005056286696344614, + 0.011628211475908756, + 0.018431317061185837, + 0.012348057702183723, + 0.0010779739823192358, + 0.0012211643625050783, + 0.0026168604381382465, + 0.002366038504987955, + 0.0009417863329872489, + 0.0012012353399768472, + 0.0022803398314863443, + 0.0012468344066292048, + 0.004836238455027342, + 0.004801292438060045, + 0.006448810920119286, + 0.04004942625761032, + 0.0020841022487729788, + 0.002663003746420145, + 0.012765019200742245, + 0.022145571187138557, + 0.0024367920123040676, + 0.0036903335712850094, + 0.005170659627765417, + 0.004806550219655037, + 0.001788275665603578, + 0.001917636371217668, + 0.002781482180580497, + 0.0050660958513617516, + 0.004706182982772589, + 0.005308857187628746, + 0.015123757533729076, + 0.05389999970793724, + 0.0030364301055669785, + 0.004256368149071932, + 0.01419144682586193, + 0.022821007296442986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.047851670533418655 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909289538860321 + ], + "opposite_logprob_tokens": [ + -2.464092969894409 + ], + "logprob_reward": -0.08909289538860321, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909289538860321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06474342197179794, + 0.03126750513911247, + 0.03011137805879116, + 0.01943213865160942, + 0.11412614583969116, + 0.08708054572343826, + 0.01584288477897644, + 0.057015400379896164, + 0.47490817308425903, + 0.10547235608100891 + ], + "credit_map": [ + 0.007775167468935251, + 0.0071107265539467335, + 0.00803880114108324, + 0.01178726740181446, + 0.00714491680264473, + 0.005282615311443806, + 0.008947947062551975, + 0.008655992336571217, + 0.0026726305950433016, + 0.002569178119301796, + 0.0044687953777611256, + 0.005785580258816481, + 0.0030956279952079058, + 0.004151485860347748, + 0.004650637041777372, + 0.0038735715206712484, + 0.0028050497639924288, + 0.0033590104430913925, + 0.0037099160254001617, + 0.004562725313007832, + 0.00314586260356009, + 0.002319074235856533, + 0.006581653840839863, + 0.0036280907224863768, + 0.0019022708293050528, + 0.0011450211750343442, + 0.002407802501693368, + 0.0031668096780776978, + 0.0015132739208638668, + 0.00209617568179965, + 0.003227740293368697, + 0.003973046317696571, + 0.005117395427078009, + 0.006360199768096209, + 0.008742772974073887, + 0.037564292550086975, + 0.004877854138612747, + 0.03116820752620697, + 0.008001748472452164, + 0.012293686158955097, + 0.004586732480674982, + 0.005640510469675064, + 0.018777281045913696, + 0.027503006160259247, + 0.00297942734323442, + 0.006531130522489548, + 0.012244126759469509, + 0.008818340487778187, + 0.0019165523117408156, + 0.0016040951013565063, + 0.0031195294577628374, + 0.0028227658476680517, + 0.0009123299387283623, + 0.0013206407893449068, + 0.002484988421201706, + 0.001661983784288168, + 0.003115361789241433, + 0.002517721150070429, + 0.003746317932382226, + 0.024366233497858047, + 0.0012478061253204942, + 0.001571688917465508, + 0.00827969890087843, + 0.012170573696494102, + 0.02188934199512005, + 0.011074953712522984, + 0.02136228233575821, + 0.02992376685142517, + 0.010831075720489025, + 0.01812194660305977, + 0.19845682382583618, + 0.16324801743030548, + 0.004537817090749741, + 0.005022457800805569, + 0.013382215052843094, + 0.044546935707330704, + 0.0029875750187784433, + 0.00375130632892251, + 0.011506447568535805, + 0.0197376050055027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.08909289538860321 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002915441989899 + ], + "opposite_logprob_tokens": [ + -2.6950292587280273 + ], + "logprob_reward": -0.07002915441989899, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002915441989899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040595635771751404, + 0.023497402667999268, + 0.022362161427736282, + 0.02055767923593521, + 0.06518109887838364, + 0.15109258890151978, + 0.028510496020317078, + 0.07107938081026077, + 0.4560958445072174, + 0.12102770805358887 + ], + "credit_map": [ + 0.0036098293494433165, + 0.0039450922049582005, + 0.0063879587687551975, + 0.005872457753866911, + 0.004025915637612343, + 0.0034417561255395412, + 0.005382193252444267, + 0.007930435240268707, + 0.00233985623344779, + 0.001828396925702691, + 0.0036539568100124598, + 0.0032018120400607586, + 0.002344927517697215, + 0.003232240444049239, + 0.0037553668953478336, + 0.003140845103189349, + 0.0022416303399950266, + 0.0018685392569750547, + 0.0046392143703997135, + 0.003071787301450968, + 0.0017355065792798996, + 0.001069496152922511, + 0.0046818372793495655, + 0.003054150380194187, + 0.001762495725415647, + 0.0017834301106631756, + 0.0024144919589161873, + 0.004066125024110079, + 0.00166268739849329, + 0.0023443889804184437, + 0.00263617979362607, + 0.0038878805935382843, + 0.003316302318125963, + 0.0032015489414334297, + 0.0052842083387076855, + 0.023616213351488113, + 0.0025658858940005302, + 0.015570237301290035, + 0.005171178374439478, + 0.006455525290220976, + 0.008796284906566143, + 0.009258883073925972, + 0.032025083899497986, + 0.0508631207048893, + 0.005413900129497051, + 0.012510250322520733, + 0.016492977738380432, + 0.015732087194919586, + 0.0022020002361387014, + 0.0018365841824561357, + 0.003260861150920391, + 0.004593228921294212, + 0.0015657132025808096, + 0.0022113847080618143, + 0.009622476063668728, + 0.0032182454597204924, + 0.005152569618076086, + 0.005237945355474949, + 0.006558844819664955, + 0.022944295778870583, + 0.003017283510416746, + 0.0032421371433883905, + 0.010181030258536339, + 0.014745273627340794, + 0.021567996591329575, + 0.012160452082753181, + 0.024328408762812614, + 0.030882082879543304, + 0.011065829545259476, + 0.01768823154270649, + 0.18654663860797882, + 0.1518561840057373, + 0.005442806053906679, + 0.005816010292619467, + 0.014966878108680248, + 0.04923199862241745, + 0.0034677661024034023, + 0.0049936482682824135, + 0.013112901709973812, + 0.02399570494890213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.07002915441989899 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897479087114334 + ], + "opposite_logprob_tokens": [ + -2.578974723815918 + ], + "logprob_reward": -0.07897479087114334, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897479087114334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17106887698173523, + 0.10810451209545135, + 0.030542666092514992, + 0.04370937868952751, + 0.22012265026569366, + 0.0596061609685421, + 0.029971156269311905, + 0.1354096233844757, + 0.15263749659061432, + 0.048827510327100754 + ], + "credit_map": [ + 0.014612028375267982, + 0.013880880549550056, + 0.022669510915875435, + 0.03121548891067505, + 0.01681770384311676, + 0.019563795998692513, + 0.0230946633964777, + 0.029214810580015182, + 0.010858278721570969, + 0.00938998069614172, + 0.017648207023739815, + 0.021767036989331245, + 0.00952951330691576, + 0.009799487888813019, + 0.01541917398571968, + 0.013692840002477169, + 0.0027352082543075085, + 0.0023708385415375233, + 0.009175118058919907, + 0.004541241098195314, + 0.00215203408151865, + 0.0022711693309247494, + 0.003891369327902794, + 0.0034056880977004766, + 0.0052676512859761715, + 0.004050186835229397, + 0.005339597351849079, + 0.006217005662620068, + 0.0036836175713688135, + 0.003935276065021753, + 0.00611081114038825, + 0.009105230681598186, + 0.010620971210300922, + 0.013009017333388329, + 0.015299884602427483, + 0.07074970006942749, + 0.010251357220113277, + 0.06075352057814598, + 0.015404478646814823, + 0.024033714085817337, + 0.005322271957993507, + 0.005436982028186321, + 0.0075363898649811745, + 0.010764896869659424, + 0.0033002011477947235, + 0.004498108755797148, + 0.015299414284527302, + 0.007447894662618637, + 0.002628929214552045, + 0.0030540148727595806, + 0.002695507137104869, + 0.003993507474660873, + 0.0023732625413686037, + 0.0029269263613969088, + 0.007844571024179459, + 0.004454436711966991, + 0.0059777721762657166, + 0.005420702509582043, + 0.01162361167371273, + 0.05806427448987961, + 0.003962870687246323, + 0.004265839699655771, + 0.01934407278895378, + 0.026750480756163597, + 0.0074061984196305275, + 0.004907994531095028, + 0.00802241824567318, + 0.011490631848573685, + 0.004404792096465826, + 0.0063486248254776, + 0.06213707476854324, + 0.047919757664203644, + 0.005922895390540361, + 0.004387823399156332, + 0.007732380647212267, + 0.009353983215987682, + 0.004831324797123671, + 0.004236129112541676, + 0.006129236426204443, + 0.006233735475689173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.07897479087114334 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546301819384098 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.012546301819384098, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546301819384098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02749057486653328, + 0.021657105535268784, + 0.05103223770856857, + 0.02802829071879387, + 0.2841684818267822, + 0.05795079469680786, + 0.021534934639930725, + 0.0326261930167675, + 0.3031550645828247, + 0.17235633730888367 + ], + "credit_map": [ + 0.0033607869409024715, + 0.0021958413999527693, + 0.004746103193610907, + 0.004690979141741991, + 0.0025986896362155676, + 0.0024915067479014397, + 0.003390433033928275, + 0.004016234073787928, + 0.002669030101969838, + 0.0021585829090327024, + 0.0030725980177521706, + 0.0030933329835534096, + 0.0021312078461050987, + 0.0036335717886686325, + 0.002741647884249687, + 0.0021571337711066008, + 0.00494812848046422, + 0.0043827262707054615, + 0.010317526757717133, + 0.004625219851732254, + 0.005929340608417988, + 0.0027361486572772264, + 0.012174617499113083, + 0.005918529815971851, + 0.0016907935496419668, + 0.002107938751578331, + 0.002481326460838318, + 0.004845296498388052, + 0.0037513445131480694, + 0.002473261673003435, + 0.004001666326075792, + 0.0066766636446118355, + 0.01216153148561716, + 0.012789440341293812, + 0.023528367280960083, + 0.10591208934783936, + 0.009295973926782608, + 0.07246201485395432, + 0.019253253936767578, + 0.02876581810414791, + 0.0047094374895095825, + 0.00433462206274271, + 0.010655759833753109, + 0.015921222046017647, + 0.002492116065695882, + 0.004780815448611975, + 0.009110026992857456, + 0.005946795921772718, + 0.0018934085965156555, + 0.0019780390430241823, + 0.0018556490540504456, + 0.0026649374049156904, + 0.001347889075987041, + 0.0017041449900716543, + 0.006256537511944771, + 0.003834328381344676, + 0.0023465207777917385, + 0.00219447398558259, + 0.0026330251712352037, + 0.011970669962465763, + 0.0013194706989452243, + 0.0012792666675522923, + 0.005106653552502394, + 0.005776110105216503, + 0.013268214650452137, + 0.009496210142970085, + 0.01439910288900137, + 0.01877533458173275, + 0.007044584956020117, + 0.011164492927491665, + 0.12003818154335022, + 0.10896895080804825, + 0.005872691515833139, + 0.006458641961216927, + 0.020897896960377693, + 0.07781680673360825, + 0.0034760667476803064, + 0.005079847294837236, + 0.02033778466284275, + 0.03241658955812454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.012546301819384098 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-frame-rate close-up of a competitive game, captured in powerful slow motion, focuses on a hand, marked by the marks of time, yet steady and powerful. Between its slender thumb and index finger, this hand elegantly and decisively grasps a smooth black Go piece, its surface reflecting a jade-like luster under the soft, diffused sunlight. With a slight twist of the wrist, the piece is steadily and deliberately moved toward an ancient wooden chessboard near the water's edge, landing precisely on a star position with a crisp, resonant thud. The minute vibration of this movement instantly stirs up extremely fine ripples on the calm surface of a small, crystal-clear rock pond near the edge of the chessboard. These ripples, clearly visible in slow motion, intertwine with the reflection of the smooth pebbles underwater. In the background, the other end of the chessboard and the blurred figure of the opponent are blurred by the shallow depth of field. The whole scene presents an extreme sense of concentration, tranquility and inner strength in the cold tones, perfectly capturing the tense atmosphere of the duel between masters where one move decides the outcome.", + "seed": 242, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_6/index_9896_seed_242_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03908300111070275, + "vlm_logprob": -0.03908300111070275, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03908300111070275, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01782235875725746 + ], + "opposite_logprob_tokens": [ + -5.017822265625 + ], + "logprob_reward": -0.01782235875725746, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.01782235875725746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1324131190776825, + 0.0759909600019455, + 0.08271229267120361, + 0.1263549029827118, + 0.06459758430719376, + 0.10662119090557098, + 0.06674951314926147, + 0.08764643222093582, + 0.11444422602653503, + 0.14246979355812073 + ], + "credit_map": [ + 0.01605084352195263, + 0.012912934646010399, + 0.01984078623354435, + 0.021074069663882256, + 0.016461588442325592, + 0.015392394736409187, + 0.012837015092372894, + 0.017843492329120636, + 0.006268269382417202, + 0.005342371296137571, + 0.010820228606462479, + 0.012621752917766571, + 0.009166670963168144, + 0.012628758326172829, + 0.008190208114683628, + 0.01095269713550806, + 0.006100325845181942, + 0.006521261762827635, + 0.017741495743393898, + 0.010397515259683132, + 0.00690116873010993, + 0.007104301359504461, + 0.02086915634572506, + 0.007077058777213097, + 0.009411427192389965, + 0.0093993553891778, + 0.011663622222840786, + 0.01208624430000782, + 0.025316044688224792, + 0.019363896921277046, + 0.018460482358932495, + 0.020653827115893364, + 0.004653650335967541, + 0.00471043074503541, + 0.008718331344425678, + 0.008511736989021301, + 0.006253658793866634, + 0.007361528929322958, + 0.00943636056035757, + 0.014951885677874088, + 0.008315207436680794, + 0.006726637948304415, + 0.016292152926325798, + 0.031147759407758713, + 0.009675241075456142, + 0.008409903384745121, + 0.016680864617228508, + 0.009373420849442482, + 0.006766799837350845, + 0.006206574849784374, + 0.007855795323848724, + 0.013213306665420532, + 0.010669974610209465, + 0.008169449865818024, + 0.008274548687040806, + 0.005593059118837118, + 0.006611454300582409, + 0.00526980496942997, + 0.014290163293480873, + 0.022089295089244843, + 0.0075241997838020325, + 0.005807724315673113, + 0.00886489450931549, + 0.017188899219036102, + 0.009056135080754757, + 0.005519968923181295, + 0.026693003252148628, + 0.035696517676115036, + 0.006337299011647701, + 0.005517370533198118, + 0.013981182128190994, + 0.011642753146588802, + 0.007997368462383747, + 0.010448635555803776, + 0.012550991028547287, + 0.048301856964826584, + 0.02324100211262703, + 0.010927933268249035, + 0.022526567801833153, + 0.006475438829511404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.01782235875725746 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207752600312233 + ], + "opposite_logprob_tokens": [ + -2.812077522277832 + ], + "logprob_reward": -0.06207752600312233, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207752600312233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14909879863262177, + 0.1996021419763565, + 0.06210990250110626, + 0.04173082485795021, + 0.03467389941215515, + 0.14081412553787231, + 0.060435544699430466, + 0.077504463493824, + 0.06011643633246422, + 0.173913836479187 + ], + "credit_map": [ + 0.012136977165937424, + 0.013692584820091724, + 0.02424469403922558, + 0.03275217488408089, + 0.016065089032053947, + 0.013088485226035118, + 0.019937746226787567, + 0.01718105562031269, + 0.012532282620668411, + 0.016331875696778297, + 0.021002333611249924, + 0.023962276056408882, + 0.05087321251630783, + 0.03402618318796158, + 0.024418078362941742, + 0.016455894336104393, + 0.0034840211737900972, + 0.004723889287561178, + 0.012250315397977829, + 0.007837284356355667, + 0.005880170036107302, + 0.005423010792583227, + 0.015994152054190636, + 0.006517054513096809, + 0.0033149071969091892, + 0.0034191065933555365, + 0.008877011947333813, + 0.006217699497938156, + 0.004864169750362635, + 0.0038393663708120584, + 0.0053937858901917934, + 0.005804775282740593, + 0.002879636362195015, + 0.002641314174979925, + 0.0041411686688661575, + 0.005428207106888294, + 0.003753199242055416, + 0.004256921354681253, + 0.004158894997090101, + 0.007414558436721563, + 0.010693739168345928, + 0.00886500533670187, + 0.02296176552772522, + 0.0388997383415699, + 0.011038066819310188, + 0.012850720435380936, + 0.02036166377365589, + 0.015143418684601784, + 0.008468891493976116, + 0.00492248497903347, + 0.0074114371091127396, + 0.009362849406898022, + 0.008419591933488846, + 0.005023509729653597, + 0.009160519577562809, + 0.007666260004043579, + 0.006434405688196421, + 0.004423878621309996, + 0.0102383429184556, + 0.010527022182941437, + 0.00912958662956953, + 0.006867988035082817, + 0.010454794391989708, + 0.019428441300988197, + 0.0029339196626096964, + 0.0026089323218911886, + 0.012824928387999535, + 0.02269938215613365, + 0.007158173713833094, + 0.002907439833506942, + 0.005694644991308451, + 0.0032890134025365114, + 0.00989364180713892, + 0.01191556453704834, + 0.016664577648043633, + 0.059331368654966354, + 0.028060823678970337, + 0.014064202085137367, + 0.026074886322021484, + 0.007908767089247704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.06207752600312233 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980225533246994 + ], + "opposite_logprob_tokens": [ + -3.5298023223876953 + ], + "logprob_reward": -0.02980225533246994, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980225533246994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13372039794921875, + 0.09176042675971985, + 0.0633310005068779, + 0.05616392195224762, + 0.0680806115269661, + 0.2821525037288666, + 0.05499798804521561, + 0.07863981276750565, + 0.06421105563640594, + 0.10694228112697601 + ], + "credit_map": [ + 0.01841437630355358, + 0.013305021449923515, + 0.015426811762154102, + 0.021806098520755768, + 0.01431352086365223, + 0.013256882317364216, + 0.020804498344659805, + 0.01639319397509098, + 0.005155616905540228, + 0.006820470094680786, + 0.016554461792111397, + 0.01310146413743496, + 0.01273407693952322, + 0.013543656095862389, + 0.011792908422648907, + 0.0120577747002244, + 0.004876519087702036, + 0.004091528709977865, + 0.012210299260914326, + 0.008520890027284622, + 0.00746062770485878, + 0.0046255216002464294, + 0.013223768211901188, + 0.008321849629282951, + 0.005503914784640074, + 0.005374429747462273, + 0.007313113659620285, + 0.008094881661236286, + 0.010677081532776356, + 0.007633790373802185, + 0.006665916182100773, + 0.00490079540759325, + 0.004118743352591991, + 0.004426593892276287, + 0.019101552665233612, + 0.016092611476778984, + 0.005339653696864843, + 0.00559880118817091, + 0.007557274308055639, + 0.005845380015671253, + 0.021093757823109627, + 0.02313302643597126, + 0.04575392231345177, + 0.07499492168426514, + 0.026440782472491264, + 0.025575920939445496, + 0.03956640139222145, + 0.025593768805265427, + 0.005925304256379604, + 0.0041426438838243484, + 0.007245898246765137, + 0.008771605789661407, + 0.008157520554959774, + 0.006635736674070358, + 0.007553763687610626, + 0.006565514486283064, + 0.009558476507663727, + 0.01463691983371973, + 0.009988654404878616, + 0.011212553828954697, + 0.005656092427670956, + 0.010827621445059776, + 0.007107704412192106, + 0.009651784785091877, + 0.00586789520457387, + 0.004199208691716194, + 0.015444071032106876, + 0.012050804682075977, + 0.0038327407091856003, + 0.004173064138740301, + 0.009693633764982224, + 0.008949637413024902, + 0.007781847845762968, + 0.008540828712284565, + 0.010394113138318062, + 0.03524928167462349, + 0.009546002373099327, + 0.007750893943011761, + 0.02060483954846859, + 0.007074479479342699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.02980225533246994 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897016406059265 + ], + "opposite_logprob_tokens": [ + -2.578970193862915 + ], + "logprob_reward": -0.07897016406059265, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897016406059265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.147483229637146, + 0.10505640506744385, + 0.056639932096004486, + 0.040656909346580505, + 0.08076275140047073, + 0.05970447510480881, + 0.18777737021446228, + 0.11307332664728165, + 0.045798029750585556, + 0.16304756700992584 + ], + "credit_map": [ + 0.018051695078611374, + 0.0073553831316530704, + 0.015422237105667591, + 0.02728632651269436, + 0.015750547870993614, + 0.012336883693933487, + 0.024317249655723572, + 0.026962900534272194, + 0.008146083913743496, + 0.008589533157646656, + 0.015569206327199936, + 0.014848231337964535, + 0.013756906613707542, + 0.012592138722538948, + 0.016055332496762276, + 0.01549895852804184, + 0.005919830873608589, + 0.0029286567587405443, + 0.012367088347673416, + 0.0065856194123625755, + 0.003802767489105463, + 0.004442556761205196, + 0.012960134074091911, + 0.007633275352418423, + 0.004387602210044861, + 0.005540842656046152, + 0.0049850791692733765, + 0.005090660415589809, + 0.005085681565105915, + 0.004242182709276676, + 0.007475762628018856, + 0.0038490945007652044, + 0.007946085184812546, + 0.007675115950405598, + 0.009725130163133144, + 0.009591363370418549, + 0.010105679742991924, + 0.010580870322883129, + 0.010980948805809021, + 0.0141575513407588, + 0.007228467613458633, + 0.008316852152347565, + 0.008320211432874203, + 0.006656273268163204, + 0.006097342818975449, + 0.007965912111103535, + 0.007172545418143272, + 0.007946865633130074, + 0.014620461501181126, + 0.012532751075923443, + 0.027094071730971336, + 0.038243915885686874, + 0.03330390527844429, + 0.01975541189312935, + 0.022547805681824684, + 0.01967903971672058, + 0.009147527627646923, + 0.0126862283796072, + 0.023555440828204155, + 0.02754805237054825, + 0.010424238629639149, + 0.00554004916921258, + 0.00961256306618452, + 0.014559214934706688, + 0.0034644564148038626, + 0.003353291656821966, + 0.010415853001177311, + 0.014334799721837044, + 0.002236009808257222, + 0.0017007228452712297, + 0.005438484251499176, + 0.0048544080927968025, + 0.016097476705908775, + 0.021641621366143227, + 0.02845165506005287, + 0.025369836017489433, + 0.008916113525629044, + 0.018052278086543083, + 0.024727357551455498, + 0.019791217520833015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.07897016406059265 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a close-up focusing on a hand grasping a Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Go piece black with a smooth surface reflecting a jade-like luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand moving the piece toward an ancient wooden chessboard located near the water's edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are fine ripples visible on a small crystal-clear rock pond caused by the piece landing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background featuring the rest of the chessboard and an opponent blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067427014000713825 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.0067427014000713825, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067427014000713825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2564083933830261, + 0.14170342683792114, + 0.049726493656635284, + 0.06626947224140167, + 0.05687723681330681, + 0.1291745901107788, + 0.14701278507709503, + 0.06769286096096039, + 0.06134021654725075, + 0.02379450388252735 + ], + "credit_map": [ + 0.03266027197241783, + 0.03097819723188877, + 0.02738267555832863, + 0.03536933660507202, + 0.030282163992524147, + 0.028662987053394318, + 0.035199448466300964, + 0.03587330877780914, + 0.008560869842767715, + 0.010127117857336998, + 0.02333119325339794, + 0.01547590084373951, + 0.01837325096130371, + 0.035984840244054794, + 0.01278328150510788, + 0.01706697605550289, + 0.003673865692690015, + 0.0038027281407266855, + 0.005602861288934946, + 0.0052103870548307896, + 0.009892987087368965, + 0.006990071851760149, + 0.009434225969016552, + 0.005119360517710447, + 0.006384735461324453, + 0.005854200571775436, + 0.008781780488789082, + 0.00857452116906643, + 0.009407556615769863, + 0.008700934238731861, + 0.009269608184695244, + 0.009296129457652569, + 0.0028331843204796314, + 0.0030666696839034557, + 0.014202817343175411, + 0.011100688017904758, + 0.005626109428703785, + 0.006714410148561001, + 0.005718223750591278, + 0.007615129463374615, + 0.009343970566987991, + 0.01107360702008009, + 0.019801616668701172, + 0.031366996467113495, + 0.013431291095912457, + 0.011147495359182358, + 0.020768634974956512, + 0.012240956537425518, + 0.008427213877439499, + 0.014424366876482964, + 0.017235666513442993, + 0.027478814125061035, + 0.03122812509536743, + 0.02361639216542244, + 0.012910142540931702, + 0.011692057363688946, + 0.004697054158896208, + 0.003707712749019265, + 0.01097217295318842, + 0.01925971917808056, + 0.00575617142021656, + 0.0058013200759887695, + 0.0055432626977562904, + 0.011955447494983673, + 0.003534612013027072, + 0.003521989332512021, + 0.015682965517044067, + 0.020367402583360672, + 0.003919108305126429, + 0.003005590057000518, + 0.006027848459780216, + 0.005280697252601385, + 0.0028487604577094316, + 0.00249840528704226, + 0.003122923895716667, + 0.003697697538882494, + 0.002704385668039322, + 0.002696841722354293, + 0.0034999658819288015, + 0.002725522965192795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0067427014000713825 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_7/index_tensor([6349, 6349])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_7/index_tensor([6349, 6349])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c5bf7e57eae2cac49a009a06eb3d97857f9f947f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_7/index_tensor([6349, 6349])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme top-down close-up, held in perfect stillness, focuses with macroscopic precision on a pristine white ceramic slab resting on a dark, textured slate surface. The slab, with its smooth, glossy glaze, is illuminated by a single, direct overhead light, creating brilliant reflections across its flawless surface. The scene is held in a state of tense stillness, spurred by invisible thermal pressure. Suddenly, a thin, hairline crack appears at the edge of the slab. With swift and decisive force, the crack extends relentlessly across the polished ceramic, branching out in delicate, lightning-like patterns and spreading across the entire surface. The harsh light captures the crack's newly formed, sharp edges, causing them to shimmer and cast intricate micro-shadows, transforming the once-perfect object into a complex web of destruction and creating a stark, focused atmosphere of irreversible change.", + "seed": 21465, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_7/index_6349_seed_21465_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.020206376141868532, + "vlm_logprob": -0.020206376141868532, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.020206376141868532, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009947016835212708 + ], + "opposite_logprob_tokens": [ + -5.759946823120117 + ], + "logprob_reward": -0.009947016835212708, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.009947016835212708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17562898993492126, + 0.07452400028705597, + 0.20686431229114532, + 0.1010662317276001, + 0.06222473084926605, + 0.06809510290622711, + 0.053129538893699646, + 0.09900573641061783, + 0.10766346007585526, + 0.05179782956838608 + ], + "credit_map": [ + 0.009865039959549904, + 0.01746065355837345, + 0.012200710363686085, + 0.025713862851262093, + 0.0183865986764431, + 0.017780564725399017, + 0.03211610019207001, + 0.04210546240210533, + 0.004845224320888519, + 0.006434247363358736, + 0.00720843905583024, + 0.010359802283346653, + 0.008519191294908524, + 0.007810452952980995, + 0.011257525533437729, + 0.018089117482304573, + 0.017322249710559845, + 0.022937197238206863, + 0.016385788097977638, + 0.037504494190216064, + 0.01857965812087059, + 0.03253527730703354, + 0.03318311646580696, + 0.028416544198989868, + 0.009384208358824253, + 0.012333141639828682, + 0.006816616747528315, + 0.01389926578849554, + 0.009149824269115925, + 0.013265038840472698, + 0.010661732405424118, + 0.025556407868862152, + 0.006352390628308058, + 0.010931180790066719, + 0.00635699974372983, + 0.00718311220407486, + 0.010223368182778358, + 0.008594936691224575, + 0.006021284032613039, + 0.006561460439115763, + 0.006185889709740877, + 0.0105802733451128, + 0.006376137491315603, + 0.006613216828554869, + 0.012410735711455345, + 0.0068540312349796295, + 0.009384886361658573, + 0.009689928963780403, + 0.00792878121137619, + 0.008638059720396996, + 0.005726000759750605, + 0.006615852937102318, + 0.006790356244891882, + 0.004766153637319803, + 0.0081567307934165, + 0.00450760405510664, + 0.012753074057400227, + 0.007236990612000227, + 0.031052347272634506, + 0.011941402219235897, + 0.007025224156677723, + 0.00674773845821619, + 0.0077871037647128105, + 0.01446185726672411, + 0.020806465297937393, + 0.0077329338528215885, + 0.00776125630363822, + 0.010337361134588718, + 0.01098860427737236, + 0.02425939403474331, + 0.009803874418139458, + 0.015973567962646484, + 0.006032065954059362, + 0.006704977247864008, + 0.00570184551179409, + 0.007553542032837868, + 0.003917524125427008, + 0.010532131418585777, + 0.00410778122022748, + 0.00724796112626791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.009947016835212708 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011099381372332573 + ], + "opposite_logprob_tokens": [ + -4.511099338531494 + ], + "logprob_reward": -0.011099381372332573, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011099381372332573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.119420126080513, + 0.07498188316822052, + 0.03369152545928955, + 0.08261391520500183, + 0.18921709060668945, + 0.07202828675508499, + 0.05869123712182045, + 0.14603380858898163, + 0.09410201758146286, + 0.1292201578617096 + ], + "credit_map": [ + 0.008342239074409008, + 0.00886512454599142, + 0.007685048971325159, + 0.015442983247339725, + 0.010314073413610458, + 0.014208264648914337, + 0.021655460819602013, + 0.03290692716836929, + 0.006810605991631746, + 0.0088081955909729, + 0.00585351325571537, + 0.01018160954117775, + 0.007808656431734562, + 0.010619241744279861, + 0.010979454033076763, + 0.01392060425132513, + 0.0020028797443956137, + 0.0040687997825443745, + 0.0022265382576733828, + 0.00639035739004612, + 0.0030338249634951353, + 0.003206603229045868, + 0.007080206647515297, + 0.005682314746081829, + 0.004366464447230101, + 0.012428943067789078, + 0.00629996694624424, + 0.01679656095802784, + 0.010502884164452553, + 0.006796914152801037, + 0.01564892567694187, + 0.009773247875273228, + 0.006484463810920715, + 0.0730040967464447, + 0.005907483398914337, + 0.007963260635733604, + 0.07384475320577621, + 0.009556657634675503, + 0.00660195155069232, + 0.005854413378983736, + 0.006354003679007292, + 0.011810609139502048, + 0.007006793282926083, + 0.007463159039616585, + 0.013355628587305546, + 0.00777081586420536, + 0.007774420082569122, + 0.01049285288900137, + 0.007820861414074898, + 0.004492567386478186, + 0.0058341664262115955, + 0.0072144209407269955, + 0.007250672206282616, + 0.005749409087002277, + 0.012552930042147636, + 0.0077762058936059475, + 0.017214637249708176, + 0.012382043525576591, + 0.05295814573764801, + 0.012327210046350956, + 0.008410640992224216, + 0.00826059840619564, + 0.008045651018619537, + 0.026434864848852158, + 0.008667068555951118, + 0.01604800671339035, + 0.012511780485510826, + 0.014673723839223385, + 0.00924716517329216, + 0.011874803341925144, + 0.006437807809561491, + 0.014641650021076202, + 0.03180303797125816, + 0.014645248651504517, + 0.012790462002158165, + 0.011358882300555706, + 0.011918987147510052, + 0.029295625165104866, + 0.006665459834039211, + 0.01074244361370802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.011099381372332573 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015169790713116527 + ], + "opposite_logprob_tokens": [ + -6.501516819000244 + ], + "logprob_reward": -0.0015169790713116527, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015169790713116527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08905397355556488, + 0.052356261759996414, + 0.11785326153039932, + 0.08297186344861984, + 0.2676781117916107, + 0.05785149335861206, + 0.04402666911482811, + 0.10174572467803955, + 0.08410079032182693, + 0.10236181318759918 + ], + "credit_map": [ + 0.0072553749196231365, + 0.009604154154658318, + 0.008913961239159107, + 0.021872546523809433, + 0.007663337979465723, + 0.008513621054589748, + 0.015888258814811707, + 0.009342729113996029, + 0.004319415893405676, + 0.0065027931705117226, + 0.005031924694776535, + 0.006616312079131603, + 0.004873606376349926, + 0.006804806180298328, + 0.007964123971760273, + 0.010243283584713936, + 0.009536891244351864, + 0.012646435759961605, + 0.009602846577763557, + 0.01899883709847927, + 0.0116995545104146, + 0.01852888986468315, + 0.019771745428442955, + 0.017068063840270042, + 0.005520328879356384, + 0.009396480396389961, + 0.005385131109505892, + 0.02006746269762516, + 0.00910145603120327, + 0.0074455649591982365, + 0.011902157217264175, + 0.014153294265270233, + 0.009183356538414955, + 0.11070603877305984, + 0.006301247980445623, + 0.007950834929943085, + 0.1073649525642395, + 0.012706080451607704, + 0.0072999391704797745, + 0.006165665108710527, + 0.0037388920318335295, + 0.009866167791187763, + 0.0034408927895128727, + 0.004953013267368078, + 0.01305507030338049, + 0.006537066772580147, + 0.006826876662671566, + 0.009433520026504993, + 0.0066485414281487465, + 0.006428122520446777, + 0.003911483101546764, + 0.004396901931613684, + 0.005206826142966747, + 0.0042208158411085606, + 0.007839364930987358, + 0.005374616011977196, + 0.011693987064063549, + 0.009141448885202408, + 0.034260645508766174, + 0.010178320109844208, + 0.0073357801884412766, + 0.006089877337217331, + 0.009259829297661781, + 0.013785848394036293, + 0.017335308715701103, + 0.01175240520387888, + 0.009832225739955902, + 0.005583632737398148, + 0.007460281252861023, + 0.013508806936442852, + 0.006227330770343542, + 0.012400808744132519, + 0.02899605967104435, + 0.009473721496760845, + 0.00835124496370554, + 0.012145073153078556, + 0.006699887569993734, + 0.022298507392406464, + 0.006789126433432102, + 0.007608198560774326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.0015169790713116527 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029808389022946358 + ], + "opposite_logprob_tokens": [ + -3.529808282852173 + ], + "logprob_reward": -0.029808389022946358, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029808389022946358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0720658227801323, + 0.03204770013689995, + 0.023305045440793037, + 0.10930110514163971, + 0.18196317553520203, + 0.07268594205379486, + 0.07026972621679306, + 0.07367508113384247, + 0.12306775152683258, + 0.24161866307258606 + ], + "credit_map": [ + 0.00420242827385664, + 0.00678088515996933, + 0.0056913504377007484, + 0.01003031525760889, + 0.006252539809793234, + 0.008392684161663055, + 0.013419141992926598, + 0.017296478152275085, + 0.0026064328849315643, + 0.003413714934140444, + 0.003256134921684861, + 0.00465820636600256, + 0.0031903283670544624, + 0.004131600726395845, + 0.005181651562452316, + 0.00560962688177824, + 0.0018006536411121488, + 0.0027506095357239246, + 0.001980024855583906, + 0.0035618413239717484, + 0.002114060102030635, + 0.0023872305173426867, + 0.004696939140558243, + 0.004013685509562492, + 0.009212986566126347, + 0.01523785013705492, + 0.005965201649814844, + 0.01817486248910427, + 0.010459850542247295, + 0.012159045785665512, + 0.012946071103215218, + 0.02514522895216942, + 0.006859400775283575, + 0.0684492364525795, + 0.004584293346852064, + 0.011154909618198872, + 0.06528940796852112, + 0.011010861024260521, + 0.007668370380997658, + 0.00694669596850872, + 0.005543205887079239, + 0.011086035519838333, + 0.004374110605567694, + 0.004715448245406151, + 0.01402414869517088, + 0.010181179270148277, + 0.01040057372301817, + 0.012361235916614532, + 0.010750551708042622, + 0.0075971330516040325, + 0.008810559287667274, + 0.008149408735334873, + 0.00820303987711668, + 0.0077301873825490475, + 0.011740533635020256, + 0.007288305088877678, + 0.01045320089906454, + 0.006281391717493534, + 0.02479633316397667, + 0.0070822895504534245, + 0.005283353850245476, + 0.005097223445773125, + 0.00564155075699091, + 0.009039726108312607, + 0.012547105550765991, + 0.012329526245594025, + 0.0104215731844306, + 0.019761869683861732, + 0.016383560374379158, + 0.025118527933955193, + 0.011966314166784286, + 0.014539269730448723, + 0.04844120144844055, + 0.03558347746729851, + 0.02770325355231762, + 0.040162283927202225, + 0.020459860563278198, + 0.031329214572906494, + 0.021532423794269562, + 0.016406938433647156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.029808389022946358 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866011440753937 + ], + "opposite_logprob_tokens": [ + -3.0486600399017334 + ], + "logprob_reward": -0.04866011440753937, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866011440753937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12476172298192978, + 0.08277124166488647, + 0.044444937258958817, + 0.13062255084514618, + 0.20146732032299042, + 0.04217120632529259, + 0.05500035732984543, + 0.0705353170633316, + 0.1922617256641388, + 0.05596365034580231 + ], + "credit_map": [ + 0.009025906212627888, + 0.010010847821831703, + 0.0084780752658844, + 0.021557047963142395, + 0.010545573197305202, + 0.011902560479938984, + 0.02702794410288334, + 0.02621375396847725, + 0.006311225704848766, + 0.008281157352030277, + 0.008776018396019936, + 0.01204159576445818, + 0.011417785659432411, + 0.009696630761027336, + 0.01659994199872017, + 0.009646885097026825, + 0.0035974658094346523, + 0.005235016345977783, + 0.003427953226491809, + 0.00931215938180685, + 0.004128307569772005, + 0.005872882902622223, + 0.007973065599799156, + 0.004898081533610821, + 0.010839401744306087, + 0.0187603160738945, + 0.009527310729026794, + 0.01910334825515747, + 0.014553227461874485, + 0.014443793334066868, + 0.014096780680119991, + 0.02929835580289364, + 0.010226063430309296, + 0.04866547882556915, + 0.012260748073458672, + 0.024396130815148354, + 0.04998357966542244, + 0.018483396619558334, + 0.019264396280050278, + 0.018187513574957848, + 0.0036883277352899313, + 0.005670124664902687, + 0.005578464362770319, + 0.008202264085412025, + 0.004406070336699486, + 0.003934494219720364, + 0.006863294634968042, + 0.003828162793070078, + 0.004956372547894716, + 0.007784553803503513, + 0.00559269031509757, + 0.0076632145792245865, + 0.007727441843599081, + 0.005918539594858885, + 0.008020931854844093, + 0.007336607202887535, + 0.009342127479612827, + 0.005699318367987871, + 0.012089051306247711, + 0.00987296737730503, + 0.01070560421794653, + 0.007969779893755913, + 0.008368780836462975, + 0.006487678736448288, + 0.01798972859978676, + 0.04975612089037895, + 0.03846320882439613, + 0.01614869385957718, + 0.013759980909526348, + 0.013342500664293766, + 0.008694385178387165, + 0.03410709276795387, + 0.006761029362678528, + 0.008109427988529205, + 0.006870965473353863, + 0.009702976793050766, + 0.00574084185063839, + 0.006490549072623253, + 0.005156301893293858, + 0.007131556514650583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.04866011440753937 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme top-down close-up, held in perfect stillness, focuses with macroscopic precision on a pristine white ceramic slab resting on a dark, textured slate surface. The slab, with its smooth, glossy glaze, is illuminated by a single, direct overhead light, creating brilliant reflections across its flawless surface. The scene is held in a state of tense stillness, spurred by invisible thermal pressure. Suddenly, a thin, hairline crack appears at the edge of the slab. With swift and decisive force, the crack extends relentlessly across the polished ceramic, branching out in delicate, lightning-like patterns and spreading across the entire surface. The harsh light captures the crack's newly formed, sharp edges, causing them to shimmer and cast intricate micro-shadows, transforming the once-perfect object into a complex web of destruction and creating a stark, focused atmosphere of irreversible change.", + "seed": 21466, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_7/index_6349_seed_21466_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03294927144888789, + "vlm_logprob": -0.03294927144888789, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.03294927144888789, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06465200334787369 + ], + "opposite_logprob_tokens": [ + -3.0646519660949707 + ], + "logprob_reward": -0.06465200334787369, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06465200334787369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08752478659152985, + 0.029919356107711792, + 0.062397319823503494, + 0.11014866828918457, + 0.050266802310943604, + 0.054388321936130524, + 0.14690740406513214, + 0.17940111458301544, + 0.22135724127292633, + 0.05768899992108345 + ], + "credit_map": [ + 0.017057295888662338, + 0.010094830766320229, + 0.009722302667796612, + 0.005139477550983429, + 0.011726221069693565, + 0.013951048254966736, + 0.004654642194509506, + 0.015178975649178028, + 0.00466183852404356, + 0.00383540871553123, + 0.003283560974523425, + 0.0017454063054174185, + 0.004885971080511808, + 0.004525656811892986, + 0.0020455040503293276, + 0.0049360101111233234, + 0.01056217122823, + 0.00616441061720252, + 0.005528397858142853, + 0.002624369692057371, + 0.01136760227382183, + 0.012377702631056309, + 0.0031782486476004124, + 0.010594416409730911, + 0.016740359365940094, + 0.010049974545836449, + 0.009692413732409477, + 0.003884111065417528, + 0.016592998057603836, + 0.027365384623408318, + 0.004341482650488615, + 0.021481946110725403, + 0.009149988181889057, + 0.004133652430027723, + 0.0038810435216873884, + 0.003015183610841632, + 0.007229491136968136, + 0.011758382432162762, + 0.002585748676210642, + 0.0085133146494627, + 0.00698880897834897, + 0.011352356523275375, + 0.005178192164748907, + 0.005459149833768606, + 0.006128737237304449, + 0.006638951599597931, + 0.004476383328437805, + 0.008165739476680756, + 0.026807861402630806, + 0.008017266169190407, + 0.011776096187531948, + 0.013490593992173672, + 0.018699726089835167, + 0.03133663907647133, + 0.011708210222423077, + 0.025071006268262863, + 0.02956153079867363, + 0.009273423813283443, + 0.01758503168821335, + 0.018679285421967506, + 0.016539668664336205, + 0.04300684854388237, + 0.016940349712967873, + 0.02781498432159424, + 0.030470434576272964, + 0.02188096009194851, + 0.023633114993572235, + 0.029204493388533592, + 0.020096810534596443, + 0.03011866845190525, + 0.04071745276451111, + 0.02523529902100563, + 0.012212391942739487, + 0.005250759422779083, + 0.006309360731393099, + 0.004648787900805473, + 0.004778530914336443, + 0.012311532162129879, + 0.0036326884292066097, + 0.00854494795203209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.06465200334787369 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020619910210371017 + ], + "opposite_logprob_tokens": [ + -3.8956198692321777 + ], + "logprob_reward": -0.020619910210371017, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020619910210371017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0798177570104599, + 0.03060336410999298, + 0.1109730675816536, + 0.07785686105489731, + 0.02039535902440548, + 0.08373776823282242, + 0.07505333423614502, + 0.45059826970100403, + 0.053304027765989304, + 0.017660120502114296 + ], + "credit_map": [ + 0.012079644948244095, + 0.0113582918420434, + 0.011833139695227146, + 0.004838425666093826, + 0.008219168521463871, + 0.011477050371468067, + 0.00739066069945693, + 0.012621382251381874, + 0.0030587275978177786, + 0.002795091364532709, + 0.00299184862524271, + 0.001831361441873014, + 0.0031092495191842318, + 0.00815348606556654, + 0.0027631581760942936, + 0.005900444462895393, + 0.01670045591890812, + 0.013509770855307579, + 0.012417774647474289, + 0.007404429838061333, + 0.0199995469301939, + 0.013191308826208115, + 0.010241571813821793, + 0.017508219927549362, + 0.011996395885944366, + 0.006704823579639196, + 0.005766316782683134, + 0.0025028313975781202, + 0.014057536609470844, + 0.01856285333633423, + 0.0034257338847965, + 0.014840374700725079, + 0.0025626930873841047, + 0.0025713916402310133, + 0.0023248805664479733, + 0.0023778791073709726, + 0.001889022416435182, + 0.0032447923440486193, + 0.0021115292329341173, + 0.0033131721429526806, + 0.00990388821810484, + 0.016372675076127052, + 0.008431297726929188, + 0.006877530366182327, + 0.012847619131207466, + 0.011055167764425278, + 0.0063935741782188416, + 0.01185602881014347, + 0.01422046683728695, + 0.00450756773352623, + 0.006194965913891792, + 0.005243208259344101, + 0.010895771905779839, + 0.016176002100110054, + 0.004876930266618729, + 0.012938427738845348, + 0.07591214776039124, + 0.02320844866335392, + 0.04275854676961899, + 0.043426476418972015, + 0.041412245482206345, + 0.11068443208932877, + 0.04136287048459053, + 0.07183313369750977, + 0.007506140042096376, + 0.004556306172162294, + 0.00709078274667263, + 0.005533711053431034, + 0.005318968091160059, + 0.008829493075609207, + 0.008384873159229755, + 0.0060837590135633945, + 0.0034929399844259024, + 0.0016380292363464832, + 0.0018792191985994577, + 0.0017427837010473013, + 0.0014283697819337249, + 0.0034890417009592056, + 0.0011730444384738803, + 0.002816692693158984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 5 + ], + "selected_reward": -0.020619910210371017 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013423488708212972 + ], + "opposite_logprob_tokens": [ + -6.626342296600342 + ], + "logprob_reward": -0.0013423488708212972, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013423488708212972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06824515014886856, + 0.03398188576102257, + 0.0496843084692955, + 0.08553778380155563, + 0.12362479418516159, + 0.07028953731060028, + 0.0791647806763649, + 0.41172391176223755, + 0.05361491069197655, + 0.02413303777575493 + ], + "credit_map": [ + 0.006735749542713165, + 0.010139544494450092, + 0.010754769667983055, + 0.005178258754312992, + 0.007843134924769402, + 0.011414912529289722, + 0.005517980083823204, + 0.010660798288881779, + 0.0046936096623539925, + 0.0033944181632250547, + 0.0023703041952103376, + 0.0015026943292468786, + 0.00783145148307085, + 0.006434571463614702, + 0.0025159530341625214, + 0.005238879006356001, + 0.0071036978624761105, + 0.007180679589509964, + 0.0078048910945653915, + 0.0038767799269407988, + 0.0054025291465222836, + 0.008012124337255955, + 0.0030657609459012747, + 0.0072378418408334255, + 0.016414640471339226, + 0.008475297130644321, + 0.0093642957508564, + 0.0052372547797858715, + 0.006642548367381096, + 0.018683232367038727, + 0.004703843034803867, + 0.016016658395528793, + 0.027158381417393684, + 0.011852970346808434, + 0.010739894583821297, + 0.006344056688249111, + 0.016057774424552917, + 0.022511962801218033, + 0.006812069099396467, + 0.022147677838802338, + 0.012401173822581768, + 0.008404870517551899, + 0.004889848176389933, + 0.005711723584681749, + 0.009300105273723602, + 0.012290269136428833, + 0.005150239448994398, + 0.012141305021941662, + 0.01368968840688467, + 0.004842373542487621, + 0.0057159243151545525, + 0.005412653088569641, + 0.011223609559237957, + 0.01891472190618515, + 0.005672452971339226, + 0.013693350367248058, + 0.06742275506258011, + 0.02133898064494133, + 0.03960081934928894, + 0.039295218884944916, + 0.03951289504766464, + 0.1046031266450882, + 0.03452025353908539, + 0.06542980670928955, + 0.006737492512911558, + 0.005415493622422218, + 0.005522694438695908, + 0.007918817922472954, + 0.004593146964907646, + 0.007665702607482672, + 0.009503102861344814, + 0.006258459761738777, + 0.003689421806484461, + 0.0023443892132490873, + 0.003458446590229869, + 0.0023670957889407873, + 0.0022812557872384787, + 0.004261935595422983, + 0.0023532446939498186, + 0.0033772480674088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.0013423488708212972 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608537696301937 + ], + "opposite_logprob_tokens": [ + -4.141085147857666 + ], + "logprob_reward": -0.01608537696301937, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608537696301937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06178971379995346, + 0.03024890460073948, + 0.043027035892009735, + 0.06545918434858322, + 0.07080195099115372, + 0.0470210537314415, + 0.06530218571424484, + 0.5147106647491455, + 0.04612407088279724, + 0.05551523715257645 + ], + "credit_map": [ + 0.007203084882348776, + 0.009656274691224098, + 0.010669982992112637, + 0.005278993397951126, + 0.006906105205416679, + 0.007147653494030237, + 0.005426612216979265, + 0.009501009248197079, + 0.004089323338121176, + 0.0035604876466095448, + 0.003524170722812414, + 0.0020907309371978045, + 0.004566150717437267, + 0.005521918646991253, + 0.002138775074854493, + 0.004757346585392952, + 0.005074472166597843, + 0.00575225381180644, + 0.0058054691180586815, + 0.0036393236368894577, + 0.005539727862924337, + 0.007257977034896612, + 0.003880735021084547, + 0.006077077705413103, + 0.011256727389991283, + 0.0055308048613369465, + 0.005088937468826771, + 0.003474991535767913, + 0.006513503380119801, + 0.017026366665959358, + 0.0038052217569202185, + 0.01276263128966093, + 0.00834851898252964, + 0.008608422242105007, + 0.00892203114926815, + 0.006560116540640593, + 0.0070463744923472404, + 0.013964510522782803, + 0.005794689524918795, + 0.011557289399206638, + 0.005926155485212803, + 0.009246235713362694, + 0.0052159917540848255, + 0.004553684964776039, + 0.006118797697126865, + 0.005011264700442553, + 0.004324379377067089, + 0.006624538451433182, + 0.015204742550849915, + 0.004433445166796446, + 0.006442094687372446, + 0.0060072727501392365, + 0.0054693035781383514, + 0.011663392186164856, + 0.005161879118531942, + 0.010920051485300064, + 0.08141405135393143, + 0.0298105888068676, + 0.04966771602630615, + 0.05321177840232849, + 0.05243212357163429, + 0.1217927560210228, + 0.04729690030217171, + 0.07908475399017334, + 0.006655631586909294, + 0.003964889328926802, + 0.005283464211970568, + 0.005132834892719984, + 0.004967277869582176, + 0.007519082631915808, + 0.006422389764338732, + 0.006178498733788729, + 0.010616584680974483, + 0.005900677293539047, + 0.005902797915041447, + 0.006752310320734978, + 0.004630527924746275, + 0.00837397575378418, + 0.004850700031965971, + 0.008487663231790066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.01608537696301937 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204671785235405 + ], + "opposite_logprob_tokens": [ + -2.812046766281128 + ], + "logprob_reward": -0.06204671785235405, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204671785235405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07471150159835815, + 0.04151841625571251, + 0.06166834756731987, + 0.12080621719360352, + 0.023444747552275658, + 0.08772698044776917, + 0.09374512732028961, + 0.44146960973739624, + 0.04158022627234459, + 0.01332879438996315 + ], + "credit_map": [ + 0.012280473485589027, + 0.0071837459690868855, + 0.007710426580160856, + 0.0055418461561203, + 0.009375529363751411, + 0.011909594759345055, + 0.008311493322253227, + 0.012398392893373966, + 0.005231799092143774, + 0.005009953863918781, + 0.004069848917424679, + 0.0030752974562346935, + 0.004624349530786276, + 0.008905633352696896, + 0.003661954775452614, + 0.006939580198377371, + 0.008866003714501858, + 0.005238229408860207, + 0.005250477232038975, + 0.002911739982664585, + 0.007968483492732048, + 0.016782917082309723, + 0.0024462328292429447, + 0.01220425870269537, + 0.02332840859889984, + 0.012752464972436428, + 0.015634844079613686, + 0.007915012538433075, + 0.010330159217119217, + 0.022741300985217094, + 0.006936069577932358, + 0.021167967468500137, + 0.005046434234827757, + 0.0016929968260228634, + 0.0022601524833589792, + 0.001815755502320826, + 0.0023921357933431864, + 0.004579775966703892, + 0.0017722656484693289, + 0.003885231912136078, + 0.01010045688599348, + 0.016124863177537918, + 0.008694409392774105, + 0.005784453824162483, + 0.015304102562367916, + 0.012501288205385208, + 0.006350004114210606, + 0.012867406010627747, + 0.017360379919409752, + 0.005465415772050619, + 0.008455582894384861, + 0.0070180646143853664, + 0.012802662327885628, + 0.020843273028731346, + 0.005772977601736784, + 0.01602676697075367, + 0.07167523354291916, + 0.023536736145615578, + 0.042108550667762756, + 0.044307779520750046, + 0.042282432317733765, + 0.11001721769571304, + 0.03752797096967697, + 0.07001367956399918, + 0.005367261357605457, + 0.00459845457226038, + 0.0038798998575657606, + 0.0061808363534510136, + 0.004032655153423548, + 0.005264756735414267, + 0.007496352307498455, + 0.004760011099278927, + 0.002571744378656149, + 0.001553666195832193, + 0.001507667824625969, + 0.001364520052447915, + 0.0012324074050411582, + 0.00198123836889863, + 0.0010984035907313228, + 0.002019146690145135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.06204671785235405 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme top-down close-up, held in perfect stillness, focuses with macroscopic precision on a pristine white ceramic slab resting on a dark, textured slate surface. The slab, with its smooth, glossy glaze, is illuminated by a single, direct overhead light, creating brilliant reflections across its flawless surface. The scene is held in a state of tense stillness, spurred by invisible thermal pressure. Suddenly, a thin, hairline crack appears at the edge of the slab. With swift and decisive force, the crack extends relentlessly across the polished ceramic, branching out in delicate, lightning-like patterns and spreading across the entire surface. The harsh light captures the crack's newly formed, sharp edges, causing them to shimmer and cast intricate micro-shadows, transforming the once-perfect object into a complex web of destruction and creating a stark, focused atmosphere of irreversible change.", + "seed": 21467, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_7/index_6349_seed_21467_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03708377375733107, + "vlm_logprob": -0.03708377375733107, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03708377375733107, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04953506961464882 + ], + "opposite_logprob_tokens": [ + -3.674535036087036 + ], + "logprob_reward": -0.04953506961464882, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.04953506961464882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5673400163650513, + 0.1782776117324829, + 0.024858390912413597, + 0.03248916566371918, + 0.07542984932661057, + 0.03067651204764843, + 0.029409103095531464, + 0.02496705763041973, + 0.023411085829138756, + 0.013141152448952198 + ], + "credit_map": [ + 0.11274538934230804, + 0.09784122556447983, + 0.0841887965798378, + 0.010174177587032318, + 0.1689801812171936, + 0.053428277373313904, + 0.016157999634742737, + 0.023823952302336693, + 0.03233743831515312, + 0.031854238361120224, + 0.023693855851888657, + 0.0038924855180084705, + 0.054609015583992004, + 0.017512589693069458, + 0.005678057670593262, + 0.008699921891093254, + 0.004554342478513718, + 0.0031416090205311775, + 0.005273399408906698, + 0.0017402279190719128, + 0.0045617809519171715, + 0.0022553240414708853, + 0.0013858223101124167, + 0.0019458843162283301, + 0.007657183799892664, + 0.003449586685746908, + 0.009625671431422234, + 0.0009943291079252958, + 0.005928649567067623, + 0.001528258086182177, + 0.0010676950914785266, + 0.002237791894003749, + 0.02275826781988144, + 0.005029189400374889, + 0.02739989198744297, + 0.0023148292675614357, + 0.00901174359023571, + 0.0030284046661108732, + 0.0026257927529513836, + 0.0032617247197777033, + 0.003718350315466523, + 0.003209956455975771, + 0.004726700950413942, + 0.004856933373957872, + 0.00366520369425416, + 0.003699891734868288, + 0.0034744925796985626, + 0.0033249834086745977, + 0.00443238066509366, + 0.0036566941998898983, + 0.005381849128752947, + 0.0026465796399861574, + 0.0052408017218112946, + 0.0028785918839275837, + 0.0028060043696314096, + 0.0023661996237933636, + 0.003611734602600336, + 0.002603141823783517, + 0.003863480407744646, + 0.0017813907470554113, + 0.0032244515605270863, + 0.003380808047950268, + 0.00283733825199306, + 0.003664711955934763, + 0.005346647929400206, + 0.0026499552186578512, + 0.00273977336473763, + 0.0014676909195259213, + 0.0021003743167966604, + 0.0033179116435348988, + 0.0015467142220586538, + 0.004242016933858395, + 0.0018231926951557398, + 0.001437349827028811, + 0.0020249648950994015, + 0.0008575212559662759, + 0.0015894711250439286, + 0.002108363201841712, + 0.0017083443235605955, + 0.0015919455327093601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.04953506961464882 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430966280400753 + ], + "opposite_logprob_tokens": [ + -3.1680965423583984 + ], + "logprob_reward": -0.0430966280400753, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430966280400753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28756746649742126, + 0.06526661664247513, + 0.03453485295176506, + 0.1579984724521637, + 0.051194675266742706, + 0.19156311452388763, + 0.05970842391252518, + 0.03366953879594803, + 0.0647578090429306, + 0.053739070892333984 + ], + "credit_map": [ + 0.05479809269309044, + 0.049233343452215195, + 0.041404545307159424, + 0.006351241376250982, + 0.08512807637453079, + 0.028375444933772087, + 0.008980841375887394, + 0.013295888900756836, + 0.01101500540971756, + 0.011004922911524773, + 0.009157364256680012, + 0.0023832309525460005, + 0.0186196006834507, + 0.0068250615149736404, + 0.0024139967281371355, + 0.00384743744507432, + 0.0079710204154253, + 0.00372866727411747, + 0.009196807630360126, + 0.0015201539499685168, + 0.005545584950596094, + 0.002800893969833851, + 0.0015587258385494351, + 0.0022130010183900595, + 0.050319135189056396, + 0.00911788735538721, + 0.06207472085952759, + 0.004750135820358992, + 0.013099877163767815, + 0.007031486369669437, + 0.005309556610882282, + 0.00629566190764308, + 0.012933004647493362, + 0.00398943480104208, + 0.015481063164770603, + 0.0020439692307263613, + 0.006869837641716003, + 0.00442540692165494, + 0.00222761114127934, + 0.003224348183721304, + 0.021784910932183266, + 0.02067623659968376, + 0.03015073761343956, + 0.03211106359958649, + 0.019646648317575455, + 0.02690199203789234, + 0.016685809940099716, + 0.023605715483427048, + 0.008593017235398293, + 0.009494427591562271, + 0.007547292392700911, + 0.003407669486477971, + 0.015549063682556152, + 0.006031384691596031, + 0.004453488625586033, + 0.004632080439478159, + 0.005218053236603737, + 0.0031995626632124186, + 0.0036427367012947798, + 0.0026644845493137836, + 0.005137734580785036, + 0.00623454712331295, + 0.003455297788605094, + 0.004117123316973448, + 0.010363242588937283, + 0.008330048993229866, + 0.009251336567103863, + 0.004276567604392767, + 0.0061939554288983345, + 0.00929764099419117, + 0.00530279241502285, + 0.011742225848138332, + 0.00899546965956688, + 0.005830706097185612, + 0.007487216964364052, + 0.004375605843961239, + 0.006513319443911314, + 0.007605746854096651, + 0.0070017059333622456, + 0.005929300561547279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0430966280400753 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010477773612365127 + ], + "opposite_logprob_tokens": [ + -6.876047611236572 + ], + "logprob_reward": -0.0010477773612365127, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010477773612365127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4152624309062958, + 0.09797701239585876, + 0.053232304751873016, + 0.0805327296257019, + 0.17108029127120972, + 0.03493087738752365, + 0.03595786914229393, + 0.022105975076556206, + 0.04910506308078766, + 0.03981544449925423 + ], + "credit_map": [ + 0.08461050689220428, + 0.06865308433771133, + 0.0673830658197403, + 0.007765330374240875, + 0.11786804348230362, + 0.0393998846411705, + 0.011070664040744305, + 0.01851186715066433, + 0.016239497810602188, + 0.015409108251333237, + 0.017007166519761086, + 0.0024997000582516193, + 0.02616235986351967, + 0.010453736409544945, + 0.002917138859629631, + 0.007288303691893816, + 0.009535511955618858, + 0.005455320701003075, + 0.01590675860643387, + 0.0017014476470649242, + 0.008088376373052597, + 0.006373849231749773, + 0.002118683885782957, + 0.004052360542118549, + 0.0248109120875597, + 0.004900897853076458, + 0.030805619433522224, + 0.002618104452267289, + 0.007336731534451246, + 0.003937807399779558, + 0.0019163815304636955, + 0.00420627323910594, + 0.05622630566358566, + 0.00998951867222786, + 0.06726604700088501, + 0.004267851822078228, + 0.018795432522892952, + 0.0060674045234918594, + 0.002800160087645054, + 0.005667572841048241, + 0.0037622249219566584, + 0.004348176531493664, + 0.004963070619851351, + 0.00663546659052372, + 0.0037780948914587498, + 0.0043494002893567085, + 0.0031157119665294886, + 0.003978729248046875, + 0.007481110282242298, + 0.003292257431894541, + 0.009038933552801609, + 0.0020185017492622137, + 0.004425233695656061, + 0.004514545202255249, + 0.0021303000394254923, + 0.0030569867230951786, + 0.003810992930084467, + 0.002556334948167205, + 0.002785906195640564, + 0.0015151232946664095, + 0.004047383088618517, + 0.0032188419718295336, + 0.0014642623718827963, + 0.002707130741328001, + 0.008231205865740776, + 0.005953902844339609, + 0.00641686050221324, + 0.005792447831481695, + 0.0053185089491307735, + 0.006318069528788328, + 0.0059113879688084126, + 0.005162681452929974, + 0.00523441331461072, + 0.0034377933479845524, + 0.005701753776520491, + 0.0022593780886381865, + 0.004160950426012278, + 0.0061077577993273735, + 0.005253427196294069, + 0.007659971248358488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0010477773612365127 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043080754578113556 + ], + "opposite_logprob_tokens": [ + -3.1680808067321777 + ], + "logprob_reward": -0.043080754578113556, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043080754578113556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22901612520217896, + 0.04814145341515541, + 0.02183576114475727, + 0.23053504526615143, + 0.18694975972175598, + 0.04386616125702858, + 0.051181498914957047, + 0.042964622378349304, + 0.05018818750977516, + 0.0953214168548584 + ], + "credit_map": [ + 0.04374440759420395, + 0.03843556344509125, + 0.03561677411198616, + 0.004720441531389952, + 0.06593352556228638, + 0.02145213820040226, + 0.007367546204477549, + 0.011745743453502655, + 0.007867022417485714, + 0.008322112262248993, + 0.0065193623304367065, + 0.001402556081302464, + 0.014527672901749611, + 0.005252310540527105, + 0.0017317108577117324, + 0.0025187113787978888, + 0.004241846967488527, + 0.0021976770367473364, + 0.005595135968178511, + 0.0014850615989416838, + 0.002190616214647889, + 0.002121336292475462, + 0.0017288951203227043, + 0.002275193575769663, + 0.07041360437870026, + 0.016349537298083305, + 0.08876650035381317, + 0.006098496727645397, + 0.031120741739869118, + 0.007161737885326147, + 0.0034740662667900324, + 0.007150369696319103, + 0.06080929934978485, + 0.00995666068047285, + 0.07267802953720093, + 0.005421425215899944, + 0.019523276016116142, + 0.007690027356147766, + 0.0033155721612274647, + 0.0075554680079221725, + 0.004091824404895306, + 0.005876229144632816, + 0.005768297240138054, + 0.00939128641039133, + 0.004983143415302038, + 0.004474438726902008, + 0.004577646963298321, + 0.00470329774543643, + 0.007272934541106224, + 0.006114642601460218, + 0.00812010932713747, + 0.003381682327017188, + 0.009892012923955917, + 0.007854579947888851, + 0.0036542557645589113, + 0.00489128427579999, + 0.006166495848447084, + 0.006824387237429619, + 0.004716182593256235, + 0.0029371841810643673, + 0.0097951116040349, + 0.004764739889651537, + 0.0036560215521603823, + 0.0041045015677809715, + 0.006173995323479176, + 0.005183060187846422, + 0.00688515231013298, + 0.005384120158851147, + 0.0050218235701322556, + 0.009575394913554192, + 0.0051106540486216545, + 0.006853990722447634, + 0.012559088878333569, + 0.007740410044789314, + 0.011405818164348602, + 0.005486928857862949, + 0.011934798210859299, + 0.014012912288308144, + 0.013087044470012188, + 0.019094416871666908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.043080754578113556 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865863919258118 + ], + "opposite_logprob_tokens": [ + -3.048658609390259 + ], + "logprob_reward": -0.04865863919258118, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865863919258118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08043773472309113, + 0.01430158969014883, + 0.008635704405605793, + 0.2671976089477539, + 0.3584343492984772, + 0.12428636103868484, + 0.02837122604250908, + 0.06312493979930878, + 0.02936525084078312, + 0.025845173746347427 + ], + "credit_map": [ + 0.01444860827177763, + 0.012132285162806511, + 0.014636731706559658, + 0.0024339461233466864, + 0.020123666152358055, + 0.008047768846154213, + 0.0031700849067419767, + 0.005444637965410948, + 0.0028695499058812857, + 0.0010486922692507505, + 0.004742143675684929, + 0.0005244013154879212, + 0.0009075108100660145, + 0.0021051252260804176, + 0.0006925418856553733, + 0.001411624369211495, + 0.0013859830796718597, + 0.000978586496785283, + 0.0021056151017546654, + 0.00050323735922575, + 0.001392945647239685, + 0.0009307279251515865, + 0.0005211401148699224, + 0.0008174683898687363, + 0.08443106710910797, + 0.016029415652155876, + 0.10590177774429321, + 0.006892894394695759, + 0.031570520251989365, + 0.008524097502231598, + 0.004095357842743397, + 0.009752500802278519, + 0.11714435368776321, + 0.019967705011367798, + 0.14093232154846191, + 0.0108150290325284, + 0.035481974482536316, + 0.016102738678455353, + 0.0066403960809111595, + 0.011349845677614212, + 0.013421764597296715, + 0.015625273808836937, + 0.017600318416953087, + 0.027337176725268364, + 0.013237830251455307, + 0.013743065297603607, + 0.009855127893388271, + 0.013465804047882557, + 0.005122707225382328, + 0.0032293135300278664, + 0.005593465641140938, + 0.002273101592436433, + 0.003958649933338165, + 0.0030225024092942476, + 0.0019298172555863857, + 0.00324166938662529, + 0.008661695756018162, + 0.005206851754337549, + 0.010862676426768303, + 0.0058755711652338505, + 0.00782109797000885, + 0.01186382956802845, + 0.005529993213713169, + 0.007303223479539156, + 0.003933069761842489, + 0.003161562606692314, + 0.00317648914642632, + 0.002361305756494403, + 0.003072738414630294, + 0.006115816067904234, + 0.003673433791846037, + 0.0038708369247615337, + 0.0040020025335252285, + 0.0027201115153729916, + 0.004220479633659124, + 0.0018352392362430692, + 0.0030697002075612545, + 0.0031158290803432465, + 0.00358206476084888, + 0.0032997471280395985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.04865863919258118 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme top-down close-up, held in perfect stillness, focuses with macroscopic precision on a pristine white ceramic slab resting on a dark, textured slate surface. The slab, with its smooth, glossy glaze, is illuminated by a single, direct overhead light, creating brilliant reflections across its flawless surface. The scene is held in a state of tense stillness, spurred by invisible thermal pressure. Suddenly, a thin, hairline crack appears at the edge of the slab. With swift and decisive force, the crack extends relentlessly across the polished ceramic, branching out in delicate, lightning-like patterns and spreading across the entire surface. The harsh light captures the crack's newly formed, sharp edges, causing them to shimmer and cast intricate micro-shadows, transforming the once-perfect object into a complex web of destruction and creating a stark, focused atmosphere of irreversible change.", + "seed": 21468, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_7/index_6349_seed_21468_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.022069193003699184, + "vlm_logprob": -0.022069193003699184, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.022069193003699184, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024136977270245552 + ], + "opposite_logprob_tokens": [ + -5.149137020111084 + ], + "logprob_reward": -0.024136977270245552, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.024136977270245552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07455141097307205, + 0.025249244645237923, + 0.8347834944725037, + 0.010915635153651237, + 0.008317133411765099, + 0.011106492951512337, + 0.009503684937953949, + 0.01405960414558649, + 0.00741946417838335, + 0.0040938230231404305 + ], + "credit_map": [ + 0.007583070080727339, + 0.005777174606919289, + 0.018875036388635635, + 0.008607810363173485, + 0.0052286856807768345, + 0.0083170710131526, + 0.011251998133957386, + 0.008910572156310081, + 0.001872641732916236, + 0.0024151557590812445, + 0.0030682517681270838, + 0.005434281192719936, + 0.0014576821122318506, + 0.0031785601750016212, + 0.004455878399312496, + 0.003366794204339385, + 0.01615741103887558, + 0.023170633241534233, + 0.02134641259908676, + 0.37134429812431335, + 0.024117877706885338, + 0.011323468759655952, + 0.3512430787086487, + 0.016080331057310104, + 0.0012251337757334113, + 0.0014456685166805983, + 0.000901712803170085, + 0.0015139211900532246, + 0.0017392029985785484, + 0.0015772725455462933, + 0.001330394996330142, + 0.0011823289096355438, + 0.0010082131484523416, + 0.0008962478605099022, + 0.0009071339154615998, + 0.0012538876617327332, + 0.0012924998300150037, + 0.0010042710928246379, + 0.0012663204688578844, + 0.0006885592010803521, + 0.001863095792941749, + 0.0012856253888458014, + 0.0011694130953401327, + 0.001077389228157699, + 0.001683560898527503, + 0.0016452828422188759, + 0.001230330322869122, + 0.0011517951497808099, + 0.0016778315184637904, + 0.0010184391867369413, + 0.0009316467330791056, + 0.001082526403479278, + 0.0008044959395192564, + 0.001941663445904851, + 0.0012296567438170314, + 0.0008174249087460339, + 0.001688483520410955, + 0.0012164332438260317, + 0.0021113494876772165, + 0.0018323436379432678, + 0.0013653923524543643, + 0.0017417296767234802, + 0.0020522200502455235, + 0.0020516521763056517, + 0.0007868329994380474, + 0.0010729690548032522, + 0.0004863496287725866, + 0.0011654294794425368, + 0.001014219829812646, + 0.0006109657697379589, + 0.0012302198447287083, + 0.0010524778626859188, + 0.0003257759381085634, + 0.0003797631652560085, + 0.00030092502129264176, + 0.0007040029740892351, + 0.0006690822192467749, + 0.00029719166923314333, + 0.0009432807564735413, + 0.0004738014831673354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.024136977270245552 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023321032524108887 + ], + "opposite_logprob_tokens": [ + -3.7733211517333984 + ], + "logprob_reward": -0.023321032524108887, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023321032524108887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12604346871376038, + 0.12971900403499603, + 0.3304539620876312, + 0.022878985852003098, + 0.04808292165398598, + 0.10170921683311462, + 0.05584986135363579, + 0.08515363931655884, + 0.07619402557611465, + 0.02391497977077961 + ], + "credit_map": [ + 0.010428565554320812, + 0.010434546507894993, + 0.02378457970917225, + 0.018235255032777786, + 0.00933845341205597, + 0.012835873290896416, + 0.022919993847608566, + 0.01806618459522724, + 0.012386126443743706, + 0.012353054247796535, + 0.0189663153141737, + 0.028992347419261932, + 0.008651531301438808, + 0.015346609987318516, + 0.023116523399949074, + 0.009906478226184845, + 0.008796559646725655, + 0.008178215473890305, + 0.007903874851763248, + 0.13783638179302216, + 0.010659824125468731, + 0.006783289834856987, + 0.1351078450679779, + 0.015187950804829597, + 0.0020486218854784966, + 0.0025850944221019745, + 0.0021813814528286457, + 0.003653257619589567, + 0.004181185737252235, + 0.0030321029480546713, + 0.003603688906878233, + 0.00159364880528301, + 0.005014557391405106, + 0.006125316955149174, + 0.005119870882481337, + 0.006855595391243696, + 0.008173377253115177, + 0.006658625323325396, + 0.006605465896427631, + 0.0035301109310239553, + 0.011506355367600918, + 0.013249165378510952, + 0.013845078647136688, + 0.01105500292032957, + 0.012064961716532707, + 0.019941961392760277, + 0.010295044630765915, + 0.009751633740961552, + 0.007284369319677353, + 0.008529574610292912, + 0.0037730212789028883, + 0.0076872739009559155, + 0.00785919465124607, + 0.008251803927123547, + 0.006675837095826864, + 0.005788779817521572, + 0.009212927892804146, + 0.008860420435667038, + 0.012058574706315994, + 0.012163335457444191, + 0.010106067173182964, + 0.010844104923307896, + 0.01329429354518652, + 0.008613910526037216, + 0.006252886261790991, + 0.012248795479536057, + 0.006782926619052887, + 0.012221922166645527, + 0.009786290116608143, + 0.0060926214791834354, + 0.014035288244485855, + 0.008773291483521461, + 0.0027235541492700577, + 0.0028371575754135847, + 0.0016961264191195369, + 0.0036142091266810894, + 0.0030149449594318867, + 0.0015719023067504168, + 0.005403577815741301, + 0.003053504740819335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.023321032524108887 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028248426970094442 + ], + "opposite_logprob_tokens": [ + -5.877824783325195 + ], + "logprob_reward": -0.0028248426970094442, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028248426970094442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1741047501564026, + 0.07572145015001297, + 0.6258121132850647, + 0.009483233094215393, + 0.01428304798901081, + 0.06360198557376862, + 0.013493726029992104, + 0.006524037104099989, + 0.009480848908424377, + 0.007494875695556402 + ], + "credit_map": [ + 0.01917499676346779, + 0.01652710512280464, + 0.04224465414881706, + 0.019794080406427383, + 0.012678330764174461, + 0.018499895930290222, + 0.023454997688531876, + 0.021730685606598854, + 0.007683214265853167, + 0.011746871285140514, + 0.0063903299160301685, + 0.010413797572255135, + 0.007340090349316597, + 0.012973390519618988, + 0.009883170947432518, + 0.009290586225688457, + 0.01237728912383318, + 0.017314665019512177, + 0.016080208122730255, + 0.2774370610713959, + 0.018338127061724663, + 0.00801393948495388, + 0.26368972659111023, + 0.01256108283996582, + 0.0012158903991803527, + 0.0008033922058530152, + 0.0008336976170539856, + 0.0010989322327077389, + 0.0010588380973786116, + 0.002523925621062517, + 0.0009865835309028625, + 0.0009619726915843785, + 0.0019214877393096685, + 0.0013833564007654786, + 0.0010005490621551871, + 0.002133262110874057, + 0.001791128539480269, + 0.0029547023586928844, + 0.0019831983372569084, + 0.0011153630912303925, + 0.007712784223258495, + 0.008332387544214725, + 0.008560034446418285, + 0.006536074448376894, + 0.007412254344671965, + 0.012695323675870895, + 0.0061737908981740475, + 0.006179334130138159, + 0.0020846298430114985, + 0.0015171375125646591, + 0.001892191357910633, + 0.001869352301582694, + 0.0017971331253647804, + 0.0015930215595290065, + 0.0017493992345407605, + 0.0009908619103953242, + 0.0008373060845769942, + 0.0006232791347429156, + 0.0008528733742423356, + 0.0008449526503682137, + 0.00060701445909217, + 0.0009549926617182791, + 0.0009794948855414987, + 0.0008241244358941913, + 0.0009803191060200334, + 0.0014469128800556064, + 0.0010414477437734604, + 0.0012549830134958029, + 0.0010116429766640067, + 0.0009070303640328348, + 0.0015063600149005651, + 0.0013321528676897287, + 0.0012016487307846546, + 0.0009097547736018896, + 0.0008411288727074862, + 0.0009960057213902473, + 0.0007975987391546369, + 0.0006659714854322374, + 0.001096152700483799, + 0.0009866143809631467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0028248426970094442 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350490748882294 + ], + "opposite_logprob_tokens": [ + -3.651350498199463 + ], + "logprob_reward": -0.026350490748882294, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350490748882294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0697050392627716, + 0.03173534572124481, + 0.5773779153823853, + 0.01769139990210533, + 0.03664704039692879, + 0.05401773005723953, + 0.035453636199235916, + 0.07757752388715744, + 0.07845290750265121, + 0.021341457962989807 + ], + "credit_map": [ + 0.00810098648071289, + 0.00877460278570652, + 0.010361879132688046, + 0.007722392678260803, + 0.007162647321820259, + 0.00861209537833929, + 0.010020601563155651, + 0.008949830196797848, + 0.003110358491539955, + 0.0030811934266239405, + 0.003788548056036234, + 0.005438832100480795, + 0.003201040206477046, + 0.003401032416149974, + 0.005830707028508186, + 0.003883634926751256, + 0.012623608112335205, + 0.01634443737566471, + 0.015673648566007614, + 0.2545468807220459, + 0.016619950532913208, + 0.009984959848225117, + 0.2393391877412796, + 0.012245229445397854, + 0.0020824093371629715, + 0.0015050310175865889, + 0.001996466191485524, + 0.0019984382670372725, + 0.0021004530135542154, + 0.0041498346254229546, + 0.0018609234830364585, + 0.0019978440832346678, + 0.004819301888346672, + 0.003449873998761177, + 0.004935009870678186, + 0.004335238132625818, + 0.004832723643630743, + 0.004946594592183828, + 0.004803930874913931, + 0.004524365533143282, + 0.006223744247108698, + 0.00730938883498311, + 0.00679922616109252, + 0.005708331707865, + 0.006998947821557522, + 0.009833179414272308, + 0.005587756168097258, + 0.005557152908295393, + 0.004450744949281216, + 0.004496348090469837, + 0.002779367147013545, + 0.004112365655601025, + 0.004868254065513611, + 0.005377728026360273, + 0.005191225092858076, + 0.004177605267614126, + 0.008685898035764694, + 0.00778356334194541, + 0.00953834317624569, + 0.011328798718750477, + 0.009529009461402893, + 0.010052025318145752, + 0.012983083724975586, + 0.007676801644265652, + 0.009447851218283176, + 0.010307018645107746, + 0.010126452893018723, + 0.010908233933150768, + 0.00864209234714508, + 0.008520185016095638, + 0.012279042042791843, + 0.008222024887800217, + 0.0023535406216979027, + 0.001924336887896061, + 0.00196477142162621, + 0.0031822475139051676, + 0.0025645671412348747, + 0.0017077357042580843, + 0.004166389349848032, + 0.0034778700210154057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.026350490748882294 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main object a pristine white ceramic slab with a smooth, glossy glaze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the slab resting on a dark, textured slate surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline crack appear at the edge and extend across the entire surface in branching patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single, direct overhead light creating brilliant reflections?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371262177824974 + ], + "opposite_logprob_tokens": [ + -3.40871262550354 + ], + "logprob_reward": -0.03371262177824974, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371262177824974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3529590368270874, + 0.04093439131975174, + 0.3875102698802948, + 0.03370736539363861, + 0.04366038367152214, + 0.015823686495423317, + 0.052998680621385574, + 0.03997122123837471, + 0.022131524980068207, + 0.010303436778485775 + ], + "credit_map": [ + 0.03025154583156109, + 0.04858581721782684, + 0.05949785187840462, + 0.046322572976350784, + 0.03954051807522774, + 0.04071253165602684, + 0.05195300281047821, + 0.03609520196914673, + 0.0035509588196873665, + 0.0036397811491042376, + 0.006035452708601952, + 0.008019591681659222, + 0.0038413875736296177, + 0.004223338793963194, + 0.007664757780730724, + 0.003959128633141518, + 0.008196388371288776, + 0.009937119670212269, + 0.009385254234075546, + 0.168514683842659, + 0.012142496183514595, + 0.005564024671912193, + 0.16239628195762634, + 0.011374032124876976, + 0.00420549837872386, + 0.003605379955843091, + 0.0031163522507995367, + 0.004875324200838804, + 0.005266169086098671, + 0.004847753793001175, + 0.004626905545592308, + 0.0031639887019991875, + 0.004444963298738003, + 0.003929549362510443, + 0.005305357277393341, + 0.005622146651148796, + 0.006512918043881655, + 0.006630408111959696, + 0.007041417062282562, + 0.004173628520220518, + 0.002073731040582061, + 0.001805605017580092, + 0.00178768381010741, + 0.0019525681855157018, + 0.0019299662671983242, + 0.002059518825262785, + 0.0020070867612957954, + 0.002207526471465826, + 0.00876778643578291, + 0.007238178048282862, + 0.00392084801569581, + 0.006719677709043026, + 0.006275271065533161, + 0.008472023531794548, + 0.0062937489710748196, + 0.005311151500791311, + 0.0043233479373157024, + 0.005507258232682943, + 0.0047562154941260815, + 0.005257731769233942, + 0.003966126590967178, + 0.0050386907532811165, + 0.0060455831699073315, + 0.00507626635953784, + 0.001977259060367942, + 0.002076601143926382, + 0.001797000877559185, + 0.003674170933663845, + 0.0028323938604444265, + 0.0017773431027308106, + 0.005358051974326372, + 0.0026387053076177835, + 0.0014894531341269612, + 0.0013478855835273862, + 0.0015529850497841835, + 0.0012098824372515082, + 0.001051014056429267, + 0.0010229303734377027, + 0.0014623458264395595, + 0.0011669397354125977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.03371262177824974 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_8/index_tensor([1943, 1943])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_8/index_tensor([1943, 1943])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c8f9c59dd35e592ec0c7352135df86e3ce428caf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_8/index_tensor([1943, 1943])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a high-angle perspective, the camera slowly and gracefully moves across a sun-drenched modern living room. Bright afternoon light filters through floor-to-ceiling windows, illuminating the dancing dust particles in the air. The lens descends toward the gleaming, polished hardwood floors, reflecting the minimalist interior. As the camera settles into a low angle, it reveals a long-haired, pure white Persian cat, its fur unusually fluffy, its large, bronze eyes wide with excitement and comic panic. The cat crouches, clumsily hunting, chasing a sleek, perfectly spherical smart toy that emits a vibrant, pulsating purple-red light. The glowing sphere zips across the floor at an unpredictable, electric speed. With more enthusiasm than grace, the cat scrambles to keep up, its paws slipping on the slippery surface. It attempts a sharp turn, but its momentum sends it tumbling into a soft, undignified tumble, a ball of white fur tumbling before it quickly recovers. Just as the cat regains its footing and prepares to pounce, the sphere suddenly makes an absurd 90-degree turn and scurries away, the light flickering momentarily. The cat skidded to a stop, tilting its head to stare blankly at the empty space where the sphere had just been, its furry tail frozen in twitching, creating an atmosphere of serene yet comedic, playful chaos.", + "seed": 27498, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_8/index_1943_seed_27498_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03557032290846109, + "vlm_logprob": -0.03557032290846109, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.03557032290846109, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014581895433366299 + ], + "opposite_logprob_tokens": [ + -5.139581680297852 + ], + "logprob_reward": -0.014581895433366299, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.014581895433366299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1774812489748001, + 0.04086076840758324, + 0.14277319610118866, + 0.059707727283239365, + 0.09190820902585983, + 0.10211076587438583, + 0.05397912487387657, + 0.12759794294834137, + 0.13350337743759155, + 0.07007770240306854 + ], + "credit_map": [ + 0.021506279706954956, + 0.028102576732635498, + 0.014048462733626366, + 0.02376989834010601, + 0.034824591130018234, + 0.011710450053215027, + 0.024826793000102043, + 0.018692193552851677, + 0.004189039580523968, + 0.004128667060285807, + 0.003357531735673547, + 0.005356807727366686, + 0.005097439978271723, + 0.0045960405841469765, + 0.009477141313254833, + 0.004658099263906479, + 0.013582712970674038, + 0.016514120623469353, + 0.013740518130362034, + 0.01969345659017563, + 0.020032675936818123, + 0.02510487474501133, + 0.016951417550444603, + 0.01715342327952385, + 0.005933082662522793, + 0.007695164065808058, + 0.005947080906480551, + 0.009801224805414677, + 0.007461393717676401, + 0.004649022128432989, + 0.008243991062045097, + 0.009976768866181374, + 0.007748798467218876, + 0.016772719100117683, + 0.011591303162276745, + 0.012223935686051846, + 0.012054406106472015, + 0.0071132611483335495, + 0.01317308098077774, + 0.011230703443288803, + 0.013336949050426483, + 0.019481047987937927, + 0.015209060162305832, + 0.00898856669664383, + 0.010067463852465153, + 0.007330472115427256, + 0.014274125918745995, + 0.013423074968159199, + 0.005934944376349449, + 0.013672823086380959, + 0.00581674138084054, + 0.008492711000144482, + 0.004380157683044672, + 0.003271278226748109, + 0.006421383004635572, + 0.005989084951579571, + 0.007673725951462984, + 0.026113374158740044, + 0.015323495492339134, + 0.03404778614640236, + 0.014623362571001053, + 0.00639680540189147, + 0.012508345767855644, + 0.010911053977906704, + 0.016509879380464554, + 0.013942398130893707, + 0.020996620878577232, + 0.013985573314130306, + 0.027626199647784233, + 0.010855020955204964, + 0.012735634110867977, + 0.016852058470249176, + 0.007529072929173708, + 0.007381094619631767, + 0.010056697763502598, + 0.0081854909658432, + 0.012579457834362984, + 0.007835710421204567, + 0.007738116662949324, + 0.008772064931690693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.014581895433366299 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011099381372332573 + ], + "opposite_logprob_tokens": [ + -4.511099338531494 + ], + "logprob_reward": -0.011099381372332573, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011099381372332573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07594726234674454, + 0.028187578544020653, + 0.034731604158878326, + 0.036458246409893036, + 0.08677013218402863, + 0.06595586985349655, + 0.07964902371168137, + 0.44852980971336365, + 0.055229753255844116, + 0.08854071795940399 + ], + "credit_map": [ + 0.009115852415561676, + 0.010813545435667038, + 0.006666066590696573, + 0.010364355519413948, + 0.015253170393407345, + 0.005888322368264198, + 0.011184382252395153, + 0.006661565974354744, + 0.003150296863168478, + 0.0032706409692764282, + 0.0024845886509865522, + 0.0042847334407269955, + 0.002841387176886201, + 0.0038257695268839598, + 0.005642526783049107, + 0.002687635598704219, + 0.0040701767429709435, + 0.004754121880978346, + 0.0047078379429876804, + 0.005448125768452883, + 0.00357979629188776, + 0.004104710184037685, + 0.004454877693206072, + 0.003611956024542451, + 0.0038907823618501425, + 0.005143997725099325, + 0.003492138348519802, + 0.008149410597980022, + 0.0031557916663587093, + 0.0027848961763083935, + 0.004766987636685371, + 0.005074241664260626, + 0.006724867504090071, + 0.011008943431079388, + 0.007260031532496214, + 0.01108548603951931, + 0.012382552027702332, + 0.00523307966068387, + 0.017502032220363617, + 0.015573137439787388, + 0.006877236533910036, + 0.012583896517753601, + 0.007862335070967674, + 0.011129950173199177, + 0.005114415194839239, + 0.003969473298639059, + 0.010807554237544537, + 0.007611007895320654, + 0.007308941800147295, + 0.015077314339578152, + 0.013641995377838612, + 0.01189173012971878, + 0.0074189151637256145, + 0.00479465676471591, + 0.011346744373440742, + 0.00816873274743557, + 0.02972964383661747, + 0.06072692945599556, + 0.05625903606414795, + 0.08887124061584473, + 0.0741051584482193, + 0.029617594555020332, + 0.06258073449134827, + 0.046639468520879745, + 0.004785611294209957, + 0.005550803616642952, + 0.0063372086733579636, + 0.005185329355299473, + 0.015777697786688805, + 0.0038036443293094635, + 0.008451386354863644, + 0.005338072311133146, + 0.00599681306630373, + 0.01240581925958395, + 0.010730382986366749, + 0.010946491733193398, + 0.018490931019186974, + 0.00892416387796402, + 0.011347359977662563, + 0.009698756039142609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.011099381372332573 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818496361374855 + ], + "opposite_logprob_tokens": [ + -4.018185138702393 + ], + "logprob_reward": -0.01818496361374855, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818496361374855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06845541298389435, + 0.053349219262599945, + 0.05647106468677521, + 0.04069319739937782, + 0.3085671663284302, + 0.1656816601753235, + 0.05855073407292366, + 0.11411284655332565, + 0.08041555434465408, + 0.05370314046740532 + ], + "credit_map": [ + 0.009185811504721642, + 0.009521164931356907, + 0.005462462082505226, + 0.008069056086242199, + 0.014015455730259418, + 0.004633164498955011, + 0.012346125207841396, + 0.005222175270318985, + 0.006014497485011816, + 0.0061257570050656796, + 0.003657953580841422, + 0.006946691311895847, + 0.009424065239727497, + 0.007326486054807901, + 0.00858498364686966, + 0.005268785636872053, + 0.006019098684191704, + 0.007755784317851067, + 0.006554148625582457, + 0.010261822491884232, + 0.006594229023903608, + 0.005742998328059912, + 0.008045971393585205, + 0.005497018340975046, + 0.004890152718871832, + 0.0044666738249361515, + 0.0043443855829536915, + 0.008779515512287617, + 0.003507778747007251, + 0.003161192871630192, + 0.006044931244105101, + 0.0054985713213682175, + 0.021997909992933273, + 0.032344751060009, + 0.026176748797297478, + 0.03545413538813591, + 0.023038923740386963, + 0.016148671507835388, + 0.07275699824094772, + 0.08064902573823929, + 0.02121524140238762, + 0.034263141453266144, + 0.020703570917248726, + 0.016272829845547676, + 0.013534883968532085, + 0.016107190400362015, + 0.019973034039139748, + 0.023611772805452347, + 0.004766056314110756, + 0.012658854946494102, + 0.008216536603868008, + 0.010038000531494617, + 0.004575388040393591, + 0.002719440031796694, + 0.00949693750590086, + 0.006079523358494043, + 0.008813677355647087, + 0.01861523650586605, + 0.011650622822344303, + 0.02470514364540577, + 0.016678325831890106, + 0.00887354463338852, + 0.012128672562539577, + 0.012647625058889389, + 0.00498654879629612, + 0.012591596692800522, + 0.010633294470608234, + 0.013156537897884846, + 0.017358511686325073, + 0.005191759672015905, + 0.009382865391671658, + 0.007114451844245195, + 0.00299218506552279, + 0.009874262847006321, + 0.005788019858300686, + 0.008344529196619987, + 0.009814415127038956, + 0.004188573453575373, + 0.007043926045298576, + 0.005657229106873274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.01818496361374855 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029684007167816 + ], + "opposite_logprob_tokens": [ + -2.35029673576355 + ], + "logprob_reward": -0.10029684007167816, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029684007167816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09395723789930344, + 0.060745108872652054, + 0.13631607592105865, + 0.21154096722602844, + 0.10980826616287231, + 0.11736379563808441, + 0.059338632971048355, + 0.08988172560930252, + 0.09816203266382217, + 0.022886181250214577 + ], + "credit_map": [ + 0.01017309632152319, + 0.01586262881755829, + 0.008602172136306763, + 0.012184182181954384, + 0.01547672413289547, + 0.006425699219107628, + 0.014998563565313816, + 0.01023417990654707, + 0.006323274690657854, + 0.008259550668299198, + 0.005546801723539829, + 0.007914222776889801, + 0.00845431163907051, + 0.008585233241319656, + 0.008819581009447575, + 0.006842132657766342, + 0.015065707266330719, + 0.015016701072454453, + 0.01378177385777235, + 0.01786048710346222, + 0.0181270781904459, + 0.02265443466603756, + 0.018583495169878006, + 0.015226399526000023, + 0.034675419330596924, + 0.022729620337486267, + 0.02664509415626526, + 0.026681149378418922, + 0.023350901901721954, + 0.01778005063533783, + 0.028981508687138557, + 0.030697228386998177, + 0.008820593357086182, + 0.02141190506517887, + 0.004651482217013836, + 0.012858155183494091, + 0.017370298504829407, + 0.005878556054085493, + 0.017949847504496574, + 0.020867424085736275, + 0.014039097353816032, + 0.026604479178786278, + 0.015587104484438896, + 0.01073581911623478, + 0.009223174303770065, + 0.009557677432894707, + 0.01491046603769064, + 0.016705980524420738, + 0.006150675471872091, + 0.010953659191727638, + 0.010959130711853504, + 0.007112592924386263, + 0.005639658309519291, + 0.004015129525214434, + 0.007887588813900948, + 0.006620198953896761, + 0.005812519229948521, + 0.010924769565463066, + 0.011559667997062206, + 0.01892700232565403, + 0.017689766362309456, + 0.006430674344301224, + 0.012128779664635658, + 0.006408546585589647, + 0.007302106358110905, + 0.012005496770143509, + 0.01040900219231844, + 0.011723821051418781, + 0.02845027856528759, + 0.007394551765173674, + 0.011972472071647644, + 0.008904306218028069, + 0.0014531201450154185, + 0.0027039663400501013, + 0.0030674065928906202, + 0.0023990939371287823, + 0.0043063657358288765, + 0.002098666038364172, + 0.0037734268698841333, + 0.003084135241806507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.10029684007167816 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033688534051179886 + ], + "opposite_logprob_tokens": [ + -3.408688545227051 + ], + "logprob_reward": -0.033688534051179886, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033688534051179886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12931041419506073, + 0.05440492555499077, + 0.034688714891672134, + 0.05157376453280449, + 0.11317369341850281, + 0.04113296419382095, + 0.08315233886241913, + 0.090981625020504, + 0.2907606363296509, + 0.11082087457180023 + ], + "credit_map": [ + 0.009053724817931652, + 0.013524482026696205, + 0.015185125172138214, + 0.01783628948032856, + 0.032465532422065735, + 0.008775957860052586, + 0.020374201238155365, + 0.01209511049091816, + 0.004556423053145409, + 0.0036139695439487696, + 0.011047333478927612, + 0.010091333650052547, + 0.0060453880578279495, + 0.004135907627642155, + 0.009867641143500805, + 0.005046929232776165, + 0.003162587061524391, + 0.003708556294441223, + 0.004306005779653788, + 0.007230194751173258, + 0.002865755930542946, + 0.002630485687404871, + 0.006278767250478268, + 0.004506363999098539, + 0.005752414930611849, + 0.0047874790616333485, + 0.0057927146553993225, + 0.01427031122148037, + 0.005016629584133625, + 0.004387015476822853, + 0.005610177759081125, + 0.0059570265002548695, + 0.007310358341783285, + 0.014651727862656116, + 0.006544931326061487, + 0.010755069553852081, + 0.010662705637514591, + 0.005984428338706493, + 0.02583482302725315, + 0.03142965957522392, + 0.004851473495364189, + 0.006793222855776548, + 0.003975426312536001, + 0.006935671903192997, + 0.003700931090861559, + 0.0026987018063664436, + 0.007630171254277229, + 0.004547367803752422, + 0.007569991517812014, + 0.020525313913822174, + 0.013359987176954746, + 0.01261295285075903, + 0.007505001034587622, + 0.004598536994308233, + 0.009153077378869057, + 0.00782748218625784, + 0.006687216926366091, + 0.01170229073613882, + 0.012446808628737926, + 0.021066391840577126, + 0.014922157861292362, + 0.006093458738178015, + 0.01059346366673708, + 0.007469845935702324, + 0.016243761405348778, + 0.04450717195868492, + 0.031124703586101532, + 0.048477914184331894, + 0.06672760099172592, + 0.020757943391799927, + 0.03752505034208298, + 0.02539650909602642, + 0.006492828018963337, + 0.017818812280893326, + 0.010527083650231361, + 0.01764790713787079, + 0.023661429062485695, + 0.010227596387267113, + 0.013426193967461586, + 0.011019035242497921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.033688534051179886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a high-angle perspective, the camera slowly and gracefully moves across a sun-drenched modern living room. Bright afternoon light filters through floor-to-ceiling windows, illuminating the dancing dust particles in the air. The lens descends toward the gleaming, polished hardwood floors, reflecting the minimalist interior. As the camera settles into a low angle, it reveals a long-haired, pure white Persian cat, its fur unusually fluffy, its large, bronze eyes wide with excitement and comic panic. The cat crouches, clumsily hunting, chasing a sleek, perfectly spherical smart toy that emits a vibrant, pulsating purple-red light. The glowing sphere zips across the floor at an unpredictable, electric speed. With more enthusiasm than grace, the cat scrambles to keep up, its paws slipping on the slippery surface. It attempts a sharp turn, but its momentum sends it tumbling into a soft, undignified tumble, a ball of white fur tumbling before it quickly recovers. Just as the cat regains its footing and prepares to pounce, the sphere suddenly makes an absurd 90-degree turn and scurries away, the light flickering momentarily. The cat skidded to a stop, tilting its head to stare blankly at the empty space where the sphere had just been, its furry tail frozen in twitching, creating an atmosphere of serene yet comedic, playful chaos.", + "seed": 27499, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_8/index_1943_seed_27499_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07392163779586554, + "vlm_logprob": -0.07392163779586554, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.07392163779586554, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0796210914850235 + ], + "opposite_logprob_tokens": [ + -2.7046210765838623 + ], + "logprob_reward": -0.0796210914850235, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0796210914850235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13394686579704285, + 0.2917819917201996, + 0.05650101602077484, + 0.09882204979658127, + 0.036077987402677536, + 0.09995261579751968, + 0.03878672420978546, + 0.03388022258877754, + 0.05672018975019455, + 0.15353034436702728 + ], + "credit_map": [ + 0.012811683118343353, + 0.011227929964661598, + 0.01677071861922741, + 0.009961182251572609, + 0.013770113699138165, + 0.020610904321074486, + 0.014656954444944859, + 0.03413736820220947, + 0.01627299189567566, + 0.01550550851970911, + 0.03933529183268547, + 0.023372357711195946, + 0.04391198605298996, + 0.021713441237807274, + 0.08612876385450363, + 0.04554164037108421, + 0.006355006713420153, + 0.0071080150082707405, + 0.006593378260731697, + 0.003151714103296399, + 0.010129859670996666, + 0.006190587766468525, + 0.01030632946640253, + 0.006666121073067188, + 0.010294490493834019, + 0.018005307763814926, + 0.009489837102591991, + 0.009729763492941856, + 0.016614247113466263, + 0.012829208746552467, + 0.008010522462427616, + 0.013848666101694107, + 0.0032934260088950396, + 0.010155177675187588, + 0.0038976604118943214, + 0.003607081947848201, + 0.0029950712341815233, + 0.004377906210720539, + 0.0018352994229644537, + 0.0059163616970181465, + 0.008108771406114101, + 0.011485149152576923, + 0.007685455493628979, + 0.015814775601029396, + 0.015034711919724941, + 0.01124335452914238, + 0.014957794919610023, + 0.015622594393789768, + 0.0028185704723000526, + 0.005755133461207151, + 0.005828293040394783, + 0.007284714840352535, + 0.004089237656444311, + 0.0027368420269340277, + 0.007397265173494816, + 0.002876665210351348, + 0.00382397323846817, + 0.0034744790755212307, + 0.003275088034570217, + 0.004970792680978775, + 0.004391605034470558, + 0.005798515398055315, + 0.004046629182994366, + 0.0040991357527673244, + 0.007316602393984795, + 0.007619291078299284, + 0.007966759614646435, + 0.008365561254322529, + 0.008355295285582542, + 0.005830851849168539, + 0.006412644870579243, + 0.004853173159062862, + 0.02681182697415352, + 0.017075520008802414, + 0.015460148453712463, + 0.020999116823077202, + 0.015927251428365707, + 0.023715900257229805, + 0.021506652235984802, + 0.012033914215862751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.0796210914850235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014223741367459297 + ], + "opposite_logprob_tokens": [ + -4.264223575592041 + ], + "logprob_reward": -0.014223741367459297, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014223741367459297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10641469061374664, + 0.18844003975391388, + 0.022076109424233437, + 0.047289662063121796, + 0.048445604741573334, + 0.08422394096851349, + 0.19711743295192719, + 0.07451637089252472, + 0.05120176076889038, + 0.1802743673324585 + ], + "credit_map": [ + 0.009226856753230095, + 0.010496784001588821, + 0.017377808690071106, + 0.010375776328146458, + 0.009775206446647644, + 0.012674420140683651, + 0.018389344215393066, + 0.018098492175340652, + 0.0095581766217947, + 0.008400889113545418, + 0.02726602740585804, + 0.007278493605554104, + 0.02983929216861725, + 0.011767727322876453, + 0.05937879905104637, + 0.03495064005255699, + 0.001814796938560903, + 0.001869981992058456, + 0.002650124253705144, + 0.0020082122646272182, + 0.003271806053817272, + 0.002015378326177597, + 0.004530107136815786, + 0.0039157019928097725, + 0.005098096560686827, + 0.009343479759991169, + 0.004117887932807207, + 0.003859502263367176, + 0.008193843998014927, + 0.005324065685272217, + 0.0045724245719611645, + 0.00678035756573081, + 0.003771076677367091, + 0.010602151043713093, + 0.006340001709759235, + 0.004657229408621788, + 0.004614109173417091, + 0.003986929077655077, + 0.006701297126710415, + 0.0077728088945150375, + 0.007370973937213421, + 0.013676367700099945, + 0.006904251873493195, + 0.01968558505177498, + 0.007539046928286552, + 0.0046426113694906235, + 0.012345952913165092, + 0.012059151194989681, + 0.01109204813838005, + 0.03058076836168766, + 0.017084702849388123, + 0.038504861295223236, + 0.023368485271930695, + 0.01342987921088934, + 0.044174209237098694, + 0.01888248138129711, + 0.006921906489878893, + 0.008321565575897694, + 0.00812411680817604, + 0.006302430760115385, + 0.01837339624762535, + 0.00856697466224432, + 0.009715245105326176, + 0.008190729655325413, + 0.008689023554325104, + 0.005168322939425707, + 0.006551841273903847, + 0.004979112651199102, + 0.005859022960066795, + 0.006612075958400965, + 0.008375522680580616, + 0.00496683781966567, + 0.020170412957668304, + 0.01694890297949314, + 0.01416043471544981, + 0.04305897653102875, + 0.017742959782481194, + 0.012834487482905388, + 0.03829960152506828, + 0.017058582976460457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 9 + ], + "selected_reward": -0.014223741367459297 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763415452092886 + ], + "opposite_logprob_tokens": [ + -4.882634162902832 + ], + "logprob_reward": -0.00763415452092886, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763415452092886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1584383100271225, + 0.045808095484972, + 0.1189466193318367, + 0.10195492953062057, + 0.20704182982444763, + 0.04908527806401253, + 0.0883893370628357, + 0.04733045771718025, + 0.10358142852783203, + 0.07942367345094681 + ], + "credit_map": [ + 0.01342667918652296, + 0.015937356278300285, + 0.027820980176329613, + 0.02103126049041748, + 0.013262154534459114, + 0.013153619132936, + 0.014096440747380257, + 0.03970980644226074, + 0.0035227928310632706, + 0.002568495459854603, + 0.005324891768395901, + 0.003971773199737072, + 0.007136955391615629, + 0.00284003047272563, + 0.012264512479305267, + 0.008178643882274628, + 0.006109924055635929, + 0.01256132684648037, + 0.015400834381580353, + 0.027997437864542007, + 0.01901406981050968, + 0.008128543384373188, + 0.018405715003609657, + 0.01132876705378294, + 0.010704290121793747, + 0.017680633813142776, + 0.009453490376472473, + 0.010739637538790703, + 0.01685653068125248, + 0.008732528425753117, + 0.011277454905211926, + 0.016510356217622757, + 0.01514885388314724, + 0.07846620678901672, + 0.01076093316078186, + 0.012780457735061646, + 0.01676049269735813, + 0.015121687203645706, + 0.012558534741401672, + 0.04544467106461525, + 0.0052427807822823524, + 0.0075215124525129795, + 0.004217721056193113, + 0.009503670036792755, + 0.005051142536103725, + 0.005094672553241253, + 0.006904170848429203, + 0.0055496059358119965, + 0.006394912488758564, + 0.012808491475880146, + 0.012818759307265282, + 0.015701238065958023, + 0.008771972730755806, + 0.004593601915985346, + 0.02032138593494892, + 0.0069789765402674675, + 0.004224955569952726, + 0.005927185993641615, + 0.005262837745249271, + 0.005671353079378605, + 0.007826432585716248, + 0.0050217024981975555, + 0.008186436258256435, + 0.0052095577120780945, + 0.011353126727044582, + 0.012294156476855278, + 0.010985727421939373, + 0.011227481998503208, + 0.019689101725816727, + 0.015927784144878387, + 0.013926668092608452, + 0.008177380077540874, + 0.0064676073379814625, + 0.007351625710725784, + 0.006612695287913084, + 0.012110678479075432, + 0.009849772788584232, + 0.014578227885067463, + 0.015744315460324287, + 0.0067087505012750626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.00763415452092886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520628273487091 + ], + "opposite_logprob_tokens": [ + -1.5020627975463867 + ], + "logprob_reward": -0.2520628273487091, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520628273487091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08493886888027191, + 0.3493790924549103, + 0.04661843180656433, + 0.06960514932870865, + 0.1083330288529396, + 0.06439082324504852, + 0.07002823054790497, + 0.03294593468308449, + 0.08525040000677109, + 0.08851002901792526 + ], + "credit_map": [ + 0.0062651983462274075, + 0.005577718373388052, + 0.011850341223180294, + 0.007223017513751984, + 0.007854130119085312, + 0.010601899586617947, + 0.010323935188353062, + 0.025242630392313004, + 0.01665148325264454, + 0.013239477761089802, + 0.05114182084798813, + 0.01333128847181797, + 0.05750930309295654, + 0.022988196462392807, + 0.11399812251329422, + 0.06051941588521004, + 0.004248038399964571, + 0.004803008865565062, + 0.005391570273786783, + 0.006025171373039484, + 0.005661261733621359, + 0.004117344506084919, + 0.008492009714245796, + 0.007880029268562794, + 0.006826682481914759, + 0.012587706558406353, + 0.0068632434122264385, + 0.007114499807357788, + 0.011375853791832924, + 0.0060212211683392525, + 0.007719246204942465, + 0.011096694506704807, + 0.016910642385482788, + 0.022895999252796173, + 0.01244320534169674, + 0.008688249625265598, + 0.01554236188530922, + 0.00878207478672266, + 0.006345896981656551, + 0.0167246013879776, + 0.0053483289666473866, + 0.011759362183511257, + 0.004481411073356867, + 0.013303776271641254, + 0.006867084186524153, + 0.0044683120213449, + 0.009959407150745392, + 0.008203142322599888, + 0.003925376106053591, + 0.010165298357605934, + 0.008255809545516968, + 0.015581389889121056, + 0.007955919951200485, + 0.005077105946838856, + 0.013217675499618053, + 0.0058496566489338875, + 0.0043165297247469425, + 0.004009367432445288, + 0.0044242339208722115, + 0.0036866769660264254, + 0.005165677517652512, + 0.0030784860718995333, + 0.0050787716172635555, + 0.0031861893367022276, + 0.00897855032235384, + 0.01168944500386715, + 0.014678592793643475, + 0.014375811442732811, + 0.008196271024644375, + 0.007365059107542038, + 0.01039187517017126, + 0.009574792347848415, + 0.007203175686299801, + 0.006190252956002951, + 0.005422050599008799, + 0.008814620785415173, + 0.009619096294045448, + 0.028138823807239532, + 0.015184447169303894, + 0.007937556132674217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.2520628273487091 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066374257206917 + ], + "opposite_logprob_tokens": [ + -4.141066551208496 + ], + "logprob_reward": -0.016066374257206917, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066374257206917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04665585979819298, + 0.1535564363002777, + 0.05685781314969063, + 0.07337459921836853, + 0.13565599918365479, + 0.0713374987244606, + 0.07506075501441956, + 0.14622478187084198, + 0.13966171443462372, + 0.1016145646572113 + ], + "credit_map": [ + 0.0053136516362428665, + 0.0050737736746668816, + 0.007004162762314081, + 0.009050371125340462, + 0.004209443461149931, + 0.005420604255050421, + 0.0038998133968561888, + 0.006684042979031801, + 0.008065319620072842, + 0.006679134909063578, + 0.021311961114406586, + 0.007186389062553644, + 0.02318057231605053, + 0.009877940639853477, + 0.04743972048163414, + 0.029815392568707466, + 0.003076452761888504, + 0.0064476169645786285, + 0.007207055110484362, + 0.01316686999052763, + 0.010200388729572296, + 0.003842342644929886, + 0.007767366711050272, + 0.005149723496288061, + 0.006558063440024853, + 0.014192008413374424, + 0.007500173058360815, + 0.009246230125427246, + 0.008761252276599407, + 0.008396539837121964, + 0.008580024354159832, + 0.010140309110283852, + 0.008863171562552452, + 0.052930593490600586, + 0.008109010756015778, + 0.006570360157638788, + 0.011773093603551388, + 0.008873439393937588, + 0.00456685945391655, + 0.033969465643167496, + 0.006968870759010315, + 0.011603704653680325, + 0.005553228780627251, + 0.01351913996040821, + 0.00674592237919569, + 0.004525449592620134, + 0.013046160340309143, + 0.00937502458691597, + 0.005071139894425869, + 0.013134284876286983, + 0.008245693519711494, + 0.015781724825501442, + 0.007472318597137928, + 0.005862269084900618, + 0.013086549937725067, + 0.006406774278730154, + 0.009556086733937263, + 0.016010751947760582, + 0.014399932697415352, + 0.01735786907374859, + 0.030866853892803192, + 0.025556150823831558, + 0.017021412029862404, + 0.015455708838999271, + 0.014821918681263924, + 0.021427219733595848, + 0.019046220928430557, + 0.02213393710553646, + 0.014571918174624443, + 0.015681149438023567, + 0.01759490929543972, + 0.014384442009031773, + 0.009548600763082504, + 0.008105323649942875, + 0.008204017765820026, + 0.012489398941397667, + 0.009724795818328857, + 0.030516240745782852, + 0.012711930088698864, + 0.010314262472093105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.016066374257206917 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a high-angle perspective, the camera slowly and gracefully moves across a sun-drenched modern living room. Bright afternoon light filters through floor-to-ceiling windows, illuminating the dancing dust particles in the air. The lens descends toward the gleaming, polished hardwood floors, reflecting the minimalist interior. As the camera settles into a low angle, it reveals a long-haired, pure white Persian cat, its fur unusually fluffy, its large, bronze eyes wide with excitement and comic panic. The cat crouches, clumsily hunting, chasing a sleek, perfectly spherical smart toy that emits a vibrant, pulsating purple-red light. The glowing sphere zips across the floor at an unpredictable, electric speed. With more enthusiasm than grace, the cat scrambles to keep up, its paws slipping on the slippery surface. It attempts a sharp turn, but its momentum sends it tumbling into a soft, undignified tumble, a ball of white fur tumbling before it quickly recovers. Just as the cat regains its footing and prepares to pounce, the sphere suddenly makes an absurd 90-degree turn and scurries away, the light flickering momentarily. The cat skidded to a stop, tilting its head to stare blankly at the empty space where the sphere had just been, its furry tail frozen in twitching, creating an atmosphere of serene yet comedic, playful chaos.", + "seed": 27500, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_8/index_1943_seed_27500_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.030930010601878166, + "vlm_logprob": -0.030930010601878166, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.030930010601878166, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013579493388533592 + ], + "opposite_logprob_tokens": [ + -4.888579368591309 + ], + "logprob_reward": -0.013579493388533592, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.013579493388533592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3454130291938782, + 0.08692928403615952, + 0.07881040126085281, + 0.0611640140414238, + 0.05100274458527565, + 0.11411821842193604, + 0.07760422676801682, + 0.08023246377706528, + 0.037123020738363266, + 0.06760262697935104 + ], + "credit_map": [ + 0.04312058538198471, + 0.09636956453323364, + 0.08009885996580124, + 0.030408287420868874, + 0.02410714700818062, + 0.035123150795698166, + 0.020479168742895126, + 0.01570628210902214, + 0.009593506343662739, + 0.009672427549958229, + 0.011532259173691273, + 0.012267086654901505, + 0.008134408853948116, + 0.009165745228528976, + 0.01206150371581316, + 0.014502352103590965, + 0.009031005203723907, + 0.010184217244386673, + 0.005524274427443743, + 0.008688944391906261, + 0.010059368796646595, + 0.014786843210458755, + 0.009185000322759151, + 0.011350749991834164, + 0.005625023972243071, + 0.005843875929713249, + 0.012358058243989944, + 0.007020126562565565, + 0.006854616105556488, + 0.010367040522396564, + 0.009081629104912281, + 0.004013644997030497, + 0.004265503026545048, + 0.0055290148593485355, + 0.0040854220278561115, + 0.006399034988135099, + 0.011817729100584984, + 0.006351725198328495, + 0.0080184955149889, + 0.004535813815891743, + 0.01717310957610607, + 0.01050343457609415, + 0.01187247596681118, + 0.01430624257773161, + 0.010090958327054977, + 0.02166377753019333, + 0.015126824378967285, + 0.013381398282945156, + 0.00702407443895936, + 0.007965261116623878, + 0.011232716031372547, + 0.006506600882858038, + 0.010142414830625057, + 0.015950052067637444, + 0.009060531854629517, + 0.009722575545310974, + 0.015325927175581455, + 0.008969067595899105, + 0.010340806096792221, + 0.008190310560166836, + 0.013646571896970272, + 0.007283364422619343, + 0.011026037856936455, + 0.005450378637760878, + 0.0036151681561022997, + 0.004288466181606054, + 0.005219507962465286, + 0.004957564175128937, + 0.004175979178398848, + 0.005027636419981718, + 0.0045584323816001415, + 0.0052802674472332, + 0.0076029193587601185, + 0.007135355845093727, + 0.006522855255752802, + 0.007716097868978977, + 0.014523371122777462, + 0.01034647785127163, + 0.007498806808143854, + 0.006256746593862772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.013579493388533592 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866999387741089 + ], + "opposite_logprob_tokens": [ + -3.0486700534820557 + ], + "logprob_reward": -0.04866999387741089, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866999387741089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2143661379814148, + 0.20979748666286469, + 0.1525287628173828, + 0.12594233453273773, + 0.04429933428764343, + 0.0291570033878088, + 0.03429006412625313, + 0.06913154572248459, + 0.022909732535481453, + 0.09757748246192932 + ], + "credit_map": [ + 0.024650713428854942, + 0.062258075922727585, + 0.0519728921353817, + 0.016186190769076347, + 0.01922737807035446, + 0.020253745838999748, + 0.012115003541111946, + 0.007702161557972431, + 0.021754629909992218, + 0.022733716294169426, + 0.026758691295981407, + 0.0284435898065567, + 0.019159890711307526, + 0.041721757501363754, + 0.019055165350437164, + 0.03017008677124977, + 0.01185842975974083, + 0.017040500417351723, + 0.009396946988999844, + 0.015350093133747578, + 0.011164371855556965, + 0.031224926933646202, + 0.024120090529322624, + 0.03237342834472656, + 0.01023187581449747, + 0.012397002428770065, + 0.018858574330806732, + 0.010769900865852833, + 0.017486155033111572, + 0.02089828997850418, + 0.02069510519504547, + 0.014605456031858921, + 0.004092053044587374, + 0.005067223682999611, + 0.004215598572045565, + 0.0067515322007238865, + 0.0070441896095871925, + 0.006136411800980568, + 0.006518768146634102, + 0.004473562818020582, + 0.002862559398636222, + 0.002846203511580825, + 0.0030801284592598677, + 0.0032015962060540915, + 0.003976447973400354, + 0.004386968445032835, + 0.005503689404577017, + 0.0032994120847433805, + 0.004434226080775261, + 0.0037113388534635305, + 0.008586370386183262, + 0.003457125276327133, + 0.0036350423470139503, + 0.0036450743209570646, + 0.0030286419205367565, + 0.003792247036471963, + 0.009138647466897964, + 0.006929214112460613, + 0.007062823511660099, + 0.011791521683335304, + 0.010928362607955933, + 0.005418357439339161, + 0.011891496367752552, + 0.005971134640276432, + 0.0024683952797204256, + 0.002452296670526266, + 0.0026179880369454622, + 0.0035181939601898193, + 0.0025229714810848236, + 0.0029234772082418203, + 0.0036321599036455154, + 0.002774251392111182, + 0.009560794569551945, + 0.010197284631431103, + 0.009771138429641724, + 0.01296200230717659, + 0.01945965178310871, + 0.012549636885523796, + 0.011487483978271484, + 0.011589503847062588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04866999387741089 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097829420119524 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.0097829420119524, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097829420119524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16714197397232056, + 0.07641203701496124, + 0.1301133781671524, + 0.21475756168365479, + 0.08576706051826477, + 0.07864013314247131, + 0.060860615223646164, + 0.05070914700627327, + 0.04364462196826935, + 0.09195342659950256 + ], + "credit_map": [ + 0.02123008482158184, + 0.04416248947381973, + 0.03742823749780655, + 0.018059680238366127, + 0.01100833062082529, + 0.016793489456176758, + 0.0099588418379426, + 0.008500833995640278, + 0.010159797966480255, + 0.006962483283132315, + 0.010700210928916931, + 0.008828028105199337, + 0.0055493819527328014, + 0.016552768647670746, + 0.006393149960786104, + 0.011266222223639488, + 0.010958089493215084, + 0.013637138530611992, + 0.009854818694293499, + 0.014521834440529346, + 0.012976574711501598, + 0.028267575427889824, + 0.013888665474951267, + 0.026008695363998413, + 0.018482638522982597, + 0.027807064354419708, + 0.028776714578270912, + 0.019893188029527664, + 0.02447672188282013, + 0.03227224573493004, + 0.04246186837553978, + 0.02058713324368, + 0.00630713626742363, + 0.007756747771054506, + 0.007346364203840494, + 0.008830493316054344, + 0.02157127670943737, + 0.008737554773688316, + 0.01755460351705551, + 0.007662886753678322, + 0.008296353742480278, + 0.013244450092315674, + 0.009734640829265118, + 0.010872755199670792, + 0.008493096567690372, + 0.012626120820641518, + 0.007888789288699627, + 0.007483930326998234, + 0.007004996296018362, + 0.00595893757417798, + 0.014659106731414795, + 0.0046130092814564705, + 0.007222870830446482, + 0.006400572136044502, + 0.005140114575624466, + 0.009861014783382416, + 0.007994531653821468, + 0.005481211934238672, + 0.005382410250604153, + 0.010979813523590565, + 0.004859920125454664, + 0.004883316811174154, + 0.007135547697544098, + 0.0039924015291035175, + 0.00458548916503787, + 0.006281384266912937, + 0.006697673816233873, + 0.004508147947490215, + 0.005885999649763107, + 0.006050453055649996, + 0.004360171966254711, + 0.005275302566587925, + 0.00934487022459507, + 0.010095332749187946, + 0.013806178234517574, + 0.014014249667525291, + 0.01115298829972744, + 0.010071803815662861, + 0.011612924747169018, + 0.011855085380375385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.0097829420119524 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204290688037872 + ], + "opposite_logprob_tokens": [ + -2.8120429515838623 + ], + "logprob_reward": -0.06204290688037872, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204290688037872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10529308766126633, + 0.08231043070554733, + 0.15495598316192627, + 0.06380948424339294, + 0.092391237616539, + 0.17636366188526154, + 0.1394839733839035, + 0.042659685015678406, + 0.1088234931230545, + 0.03390898182988167 + ], + "credit_map": [ + 0.013458909466862679, + 0.027771705761551857, + 0.01976902037858963, + 0.00939969066530466, + 0.008559106849133968, + 0.011807076632976532, + 0.008211546577513218, + 0.006316040176898241, + 0.012311545200645924, + 0.009094326756894588, + 0.01137564517557621, + 0.011637882329523563, + 0.006351133342832327, + 0.011188141070306301, + 0.00651202816516161, + 0.013839730061590672, + 0.011799781583249569, + 0.01975048892199993, + 0.009230893105268478, + 0.013052497059106827, + 0.014069296419620514, + 0.042214274406433105, + 0.012464912608265877, + 0.032373830676078796, + 0.005359380040317774, + 0.005859865806996822, + 0.008489036932587624, + 0.008000900968909264, + 0.007611030247062445, + 0.010075364261865616, + 0.011354646645486355, + 0.007059261202812195, + 0.01024908572435379, + 0.008984635584056377, + 0.008650225587189198, + 0.00876427162438631, + 0.014670169912278652, + 0.011868282221257687, + 0.0194101445376873, + 0.009794418700039387, + 0.01699560508131981, + 0.013943318277597427, + 0.019470270723104477, + 0.03015254996716976, + 0.015085924416780472, + 0.03264903649687767, + 0.025699177756905556, + 0.022367769852280617, + 0.015542080625891685, + 0.011272069998085499, + 0.030697284266352654, + 0.00876446720212698, + 0.020526112988591194, + 0.01986943557858467, + 0.013765998184680939, + 0.019046518951654434, + 0.004132048226892948, + 0.0050703296437859535, + 0.005162039306014776, + 0.008615076541900635, + 0.006104352883994579, + 0.003174412762746215, + 0.0067824688740074635, + 0.003618954448029399, + 0.009038345888257027, + 0.019103989005088806, + 0.008982909843325615, + 0.011722216382622719, + 0.012031286023557186, + 0.013632712885737419, + 0.01750755123794079, + 0.01680447719991207, + 0.0036433469504117966, + 0.0038829497061669827, + 0.0032999147661030293, + 0.0035648939665406942, + 0.006831934209913015, + 0.005950110498815775, + 0.0030745628755539656, + 0.003661267925053835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.06204290688037872 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020574716851115227 + ], + "opposite_logprob_tokens": [ + -3.8955748081207275 + ], + "logprob_reward": -0.020574716851115227, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020574716851115227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1071992814540863, + 0.04849207028746605, + 0.08333247154951096, + 0.11215823143720627, + 0.14842848479747772, + 0.07587892562150955, + 0.07908167690038681, + 0.048472024500370026, + 0.11189688742160797, + 0.18505997955799103 + ], + "credit_map": [ + 0.010963925160467625, + 0.026890190318226814, + 0.021328426897525787, + 0.007676796521991491, + 0.011376031674444675, + 0.01112267468124628, + 0.010211179032921791, + 0.0076300473883748055, + 0.005747129209339619, + 0.0052833156660199165, + 0.00589854596182704, + 0.007604325655847788, + 0.00355682335793972, + 0.007441401015967131, + 0.004580681212246418, + 0.008379846811294556, + 0.006189222913235426, + 0.010221277363598347, + 0.005750610493123531, + 0.008187295868992805, + 0.007388455327600241, + 0.020973946899175644, + 0.011410889215767384, + 0.013210766948759556, + 0.009148282930254936, + 0.01743258349597454, + 0.016829591244459152, + 0.012031177058815956, + 0.011756554245948792, + 0.015186704695224762, + 0.02041599713265896, + 0.009357335045933723, + 0.011844698339700699, + 0.013116220012307167, + 0.013543187640607357, + 0.01549553032964468, + 0.02906758151948452, + 0.019654663279652596, + 0.03092559613287449, + 0.014780997298657894, + 0.005755084101110697, + 0.0071265376172959805, + 0.007864493876695633, + 0.012363454326987267, + 0.006504384335130453, + 0.015226280316710472, + 0.010559399612247944, + 0.010479294694960117, + 0.0077929990366101265, + 0.0071420203894376755, + 0.012103581801056862, + 0.007420300971716642, + 0.013703752309083939, + 0.013079272583127022, + 0.008175374008715153, + 0.00966437254101038, + 0.008332181721925735, + 0.006095452234148979, + 0.006313239689916372, + 0.005914604756981134, + 0.006270819343626499, + 0.004858528263866901, + 0.006294908933341503, + 0.004392285831272602, + 0.009607053361833096, + 0.023151755332946777, + 0.005418926943093538, + 0.01059491466730833, + 0.012782137840986252, + 0.011161754839122295, + 0.019872434437274933, + 0.019307894632220268, + 0.015982648357748985, + 0.024157991632819176, + 0.015169816091656685, + 0.026437871158123016, + 0.030591972172260284, + 0.035965029150247574, + 0.021417580544948578, + 0.015337065793573856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.020574716851115227 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a high-angle perspective, the camera slowly and gracefully moves across a sun-drenched modern living room. Bright afternoon light filters through floor-to-ceiling windows, illuminating the dancing dust particles in the air. The lens descends toward the gleaming, polished hardwood floors, reflecting the minimalist interior. As the camera settles into a low angle, it reveals a long-haired, pure white Persian cat, its fur unusually fluffy, its large, bronze eyes wide with excitement and comic panic. The cat crouches, clumsily hunting, chasing a sleek, perfectly spherical smart toy that emits a vibrant, pulsating purple-red light. The glowing sphere zips across the floor at an unpredictable, electric speed. With more enthusiasm than grace, the cat scrambles to keep up, its paws slipping on the slippery surface. It attempts a sharp turn, but its momentum sends it tumbling into a soft, undignified tumble, a ball of white fur tumbling before it quickly recovers. Just as the cat regains its footing and prepares to pounce, the sphere suddenly makes an absurd 90-degree turn and scurries away, the light flickering momentarily. The cat skidded to a stop, tilting its head to stare blankly at the empty space where the sphere had just been, its furry tail frozen in twitching, creating an atmosphere of serene yet comedic, playful chaos.", + "seed": 27501, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_8/index_1943_seed_27501_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06711023394018412, + "vlm_logprob": -0.06711023394018412, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.06711023394018412, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2491735816001892 + ], + "opposite_logprob_tokens": [ + -1.624173641204834 + ], + "logprob_reward": -0.2491735816001892, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2491735816001892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11538270115852356, + 0.17773346602916718, + 0.13752548396587372, + 0.10170816630125046, + 0.20415551960468292, + 0.047994691878557205, + 0.04797125607728958, + 0.047002580016851425, + 0.044874660670757294, + 0.07565145939588547 + ], + "credit_map": [ + 0.014544676057994366, + 0.01540635246783495, + 0.015671828761696815, + 0.017080798745155334, + 0.010117240250110626, + 0.014066524803638458, + 0.018305839970707893, + 0.010189441032707691, + 0.020673057064414024, + 0.032038867473602295, + 0.022142328321933746, + 0.02144467458128929, + 0.027513718232512474, + 0.022769272327423096, + 0.012620539404451847, + 0.01853102073073387, + 0.014361117035150528, + 0.017749177291989326, + 0.023337343707680702, + 0.013042007572948933, + 0.01317190658301115, + 0.024403534829616547, + 0.01168918889015913, + 0.01977122202515602, + 0.012233739718794823, + 0.01172640174627304, + 0.014889812096953392, + 0.01585768721997738, + 0.012936276383697987, + 0.014153030700981617, + 0.007746819872409105, + 0.012164411135017872, + 0.012154093012213707, + 0.030611632391810417, + 0.05198642611503601, + 0.057909805327653885, + 0.013128763064742088, + 0.007080278359353542, + 0.025785494595766068, + 0.005499036516994238, + 0.005660263355821371, + 0.003250196808949113, + 0.0071158078499138355, + 0.0038674036040902138, + 0.007956648245453835, + 0.00934529583901167, + 0.005504099652171135, + 0.005294984672218561, + 0.006673537660390139, + 0.0043842908926308155, + 0.007304613012820482, + 0.007247191853821278, + 0.00808155070990324, + 0.0042154197581112385, + 0.004971296992152929, + 0.0050933570601046085, + 0.003760955296456814, + 0.011128987185657024, + 0.0033480653073638678, + 0.0037981707137078047, + 0.00510654179379344, + 0.00454912893474102, + 0.01163000613451004, + 0.0036807293072342873, + 0.0038799899630248547, + 0.004014540929347277, + 0.005462725181132555, + 0.011964382603764534, + 0.003804380539804697, + 0.00321631389670074, + 0.008557644672691822, + 0.0039746868424117565, + 0.006435425952076912, + 0.019012127071619034, + 0.005239120218902826, + 0.008880384266376495, + 0.017557546496391296, + 0.007681811228394508, + 0.005596982315182686, + 0.005248068831861019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.2491735816001892 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014230557717382908 + ], + "opposite_logprob_tokens": [ + -4.264230728149414 + ], + "logprob_reward": -0.014230557717382908, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014230557717382908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03277748450636864, + 0.13895736634731293, + 0.027463847771286964, + 0.019073061645030975, + 0.08771228790283203, + 0.03135949745774269, + 0.0705975890159607, + 0.09816361963748932, + 0.0849449560046196, + 0.4089502692222595 + ], + "credit_map": [ + 0.0028287519235163927, + 0.005000297445803881, + 0.002876950427889824, + 0.002185674151405692, + 0.006766986567527056, + 0.005011920351535082, + 0.0038374231662601233, + 0.004269476514309645, + 0.013235459104180336, + 0.023142505437135696, + 0.01390206627547741, + 0.015262785367667675, + 0.02858826518058777, + 0.02060030959546566, + 0.009847301989793777, + 0.014378665946424007, + 0.0033429095055907965, + 0.005792603828012943, + 0.002649651374667883, + 0.0022298027761280537, + 0.004570122342556715, + 0.004359273239970207, + 0.002013289835304022, + 0.002506192307919264, + 0.0022838362492620945, + 0.0029103714041411877, + 0.00234247837215662, + 0.0033785447012633085, + 0.0016652786871418357, + 0.0022876025177538395, + 0.0017772553255781531, + 0.0024276950862258673, + 0.004231845960021019, + 0.01790773868560791, + 0.018926424905657768, + 0.029033789411187172, + 0.003310528816655278, + 0.002534025115892291, + 0.009728915058076382, + 0.0020390187855809927, + 0.004146892577409744, + 0.002539300359785557, + 0.005718638189136982, + 0.0033232884015887976, + 0.004279569257050753, + 0.005096851848065853, + 0.0027983917389065027, + 0.003456561593338847, + 0.008671239949762821, + 0.009686020202934742, + 0.009789465926587582, + 0.006505655590444803, + 0.011938406154513359, + 0.010065881535410881, + 0.008000076748430729, + 0.005940836854279041, + 0.01347567979246378, + 0.01594565436244011, + 0.012845254503190517, + 0.010650108568370342, + 0.010498364455997944, + 0.006705703213810921, + 0.02063969150185585, + 0.007403152529150248, + 0.008936777710914612, + 0.006781450007110834, + 0.005060374736785889, + 0.016005149111151695, + 0.00838322751224041, + 0.010260965675115585, + 0.013367136009037495, + 0.016149869188666344, + 0.02777877077460289, + 0.08807507157325745, + 0.03175238519906998, + 0.05256123095750809, + 0.07696255296468735, + 0.029702067375183105, + 0.07416418194770813, + 0.027954017743468285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.014230557717382908 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012548303231596947 + ], + "opposite_logprob_tokens": [ + -4.387548446655273 + ], + "logprob_reward": -0.012548303231596947, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012548303231596947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06774737685918808, + 0.07616408169269562, + 0.04938014969229698, + 0.05135722458362579, + 0.08198316395282745, + 0.04248567298054695, + 0.05203695967793465, + 0.17763733863830566, + 0.35778528451919556, + 0.04342281073331833 + ], + "credit_map": [ + 0.0071400669403374195, + 0.010341244749724865, + 0.006661641877144575, + 0.005775488447397947, + 0.013966879807412624, + 0.007951008155941963, + 0.009692675434052944, + 0.006218373775482178, + 0.0067604766227304935, + 0.011195684783160686, + 0.0063093556091189384, + 0.007265460677444935, + 0.01800665445625782, + 0.0077064367942512035, + 0.009425212629139423, + 0.009494802914559841, + 0.005682769697159529, + 0.008177490904927254, + 0.005656943190842867, + 0.0032642290461808443, + 0.007757079321891069, + 0.007387511432170868, + 0.004129915963858366, + 0.0073242103680968285, + 0.0057193138636648655, + 0.010972566902637482, + 0.00494022062048316, + 0.005762520246207714, + 0.0070721800439059734, + 0.005504000000655651, + 0.005310643929988146, + 0.006075777113437653, + 0.004960249178111553, + 0.020917048677802086, + 0.0057672904804348946, + 0.020824452862143517, + 0.010596994310617447, + 0.006303951609879732, + 0.006879113614559174, + 0.005734060890972614, + 0.00797845795750618, + 0.002091483911499381, + 0.005629465915262699, + 0.0044737025164067745, + 0.006561731919646263, + 0.008861144073307514, + 0.0026306414511054754, + 0.004259044770151377, + 0.007840454578399658, + 0.004316987469792366, + 0.006416645832359791, + 0.0046612839214503765, + 0.008563768118619919, + 0.008116214536130428, + 0.005527282599359751, + 0.006594321224838495, + 0.015003311447799206, + 0.03322381153702736, + 0.01827099360525608, + 0.01692267693579197, + 0.02331330068409443, + 0.014478622004389763, + 0.04185333102941513, + 0.014571284875273705, + 0.03974528610706329, + 0.03217749297618866, + 0.03261660411953926, + 0.035372041165828705, + 0.03394477441906929, + 0.05677454173564911, + 0.042539142072200775, + 0.08461540192365646, + 0.006395571865141392, + 0.0047136759385466576, + 0.005433009006083012, + 0.004727525636553764, + 0.00613612774759531, + 0.005250779911875725, + 0.005080394446849823, + 0.005685730371624231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.012548303231596947 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029805496335029602 + ], + "opposite_logprob_tokens": [ + -3.5298054218292236 + ], + "logprob_reward": -0.029805496335029602, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029805496335029602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1104549691081047, + 0.07617221027612686, + 0.07938721030950546, + 0.06409372389316559, + 0.14321109652519226, + 0.1799202561378479, + 0.09177189320325851, + 0.12891462445259094, + 0.07083012908697128, + 0.0552438385784626 + ], + "credit_map": [ + 0.015383024699985981, + 0.014588041231036186, + 0.013756392523646355, + 0.0085061090067029, + 0.020507946610450745, + 0.011521700769662857, + 0.013174458406865597, + 0.013017295859754086, + 0.006047364324331284, + 0.00797336921095848, + 0.007567143067717552, + 0.007058321498334408, + 0.017384760081768036, + 0.010207816027104855, + 0.010497301816940308, + 0.009436135180294514, + 0.007486048154532909, + 0.010444257408380508, + 0.01349269412457943, + 0.014114296063780785, + 0.006908294279128313, + 0.011740847490727901, + 0.004712711088359356, + 0.010488064959645271, + 0.006413659546524286, + 0.007400913629680872, + 0.007064211647957563, + 0.01381648425012827, + 0.00577310798689723, + 0.005823823623359203, + 0.005955710541456938, + 0.011845814064145088, + 0.012426200322806835, + 0.019493503496050835, + 0.027421176433563232, + 0.031249569728970528, + 0.021190473809838295, + 0.007072407752275467, + 0.01691632904112339, + 0.007441450841724873, + 0.02491435594856739, + 0.015082522295415401, + 0.02413700893521309, + 0.012543722987174988, + 0.02850007638335228, + 0.038423240184783936, + 0.020137684419751167, + 0.016181662678718567, + 0.010713241994380951, + 0.005605990532785654, + 0.010604215785861015, + 0.010338936932384968, + 0.01526788342744112, + 0.019089961424469948, + 0.008939503692090511, + 0.011212166398763657, + 0.013149679638445377, + 0.025565166026353836, + 0.012019089423120022, + 0.010877426713705063, + 0.011806298978626728, + 0.01067275833338499, + 0.030802281573414803, + 0.01402192935347557, + 0.0073775858618319035, + 0.006857624743133783, + 0.009638468734920025, + 0.007736184634268284, + 0.00718427961692214, + 0.009322991594672203, + 0.008566987700760365, + 0.014146015048027039, + 0.0038319099694490433, + 0.01690492033958435, + 0.002702487865462899, + 0.004386230371892452, + 0.015232465229928493, + 0.004500846844166517, + 0.004403551574796438, + 0.0032814322039484978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.029805496335029602 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched modern living room with floor-to-ceiling windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long-haired, pure white Persian cat with unusually fluffy fur visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek, perfectly spherical smart toy emitting a vibrant, pulsating purple-red light present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a high-angle perspective and slowly descend to a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cat tumble onto the hardwood floor and then recover to stare blankly at the empty space after the toy leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029793230816721916 + ], + "opposite_logprob_tokens": [ + -3.5297932624816895 + ], + "logprob_reward": -0.029793230816721916, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029793230816721916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07794758677482605, + 0.053769536316394806, + 0.044610969722270966, + 0.029888205230236053, + 0.2303289771080017, + 0.14848975837230682, + 0.1232152134180069, + 0.10499034076929092, + 0.12363533675670624, + 0.06312403827905655 + ], + "credit_map": [ + 0.006728095933794975, + 0.011010268703103065, + 0.00799416471272707, + 0.007284201681613922, + 0.016169529408216476, + 0.009168333373963833, + 0.012910380028188229, + 0.00668261107057333, + 0.004789767321199179, + 0.006673313211649656, + 0.0051973857916891575, + 0.004659460857510567, + 0.011082164011895657, + 0.005627418402582407, + 0.009540960192680359, + 0.0061990637332201, + 0.003860571188852191, + 0.008001215755939484, + 0.005184616427868605, + 0.008782442659139633, + 0.004574095364660025, + 0.005968865007162094, + 0.00449422188103199, + 0.0037449405062943697, + 0.003333375556394458, + 0.004613048397004604, + 0.0032879405189305544, + 0.004695046227425337, + 0.003517098492011428, + 0.00352348736487329, + 0.0029429879505187273, + 0.003975221887230873, + 0.01220749132335186, + 0.04591045901179314, + 0.045147333294153214, + 0.07211504131555557, + 0.015206225216388702, + 0.0062392824329435825, + 0.02489813044667244, + 0.008605009876191616, + 0.01600276120007038, + 0.012795863673090935, + 0.020097799599170685, + 0.012951517477631569, + 0.024711430072784424, + 0.03025125525891781, + 0.017508873715996742, + 0.014170250855386257, + 0.013770278543233871, + 0.009808912873268127, + 0.017130494117736816, + 0.009882570244371891, + 0.02925911173224449, + 0.019352054223418236, + 0.013565737754106522, + 0.01044605951756239, + 0.01771584339439869, + 0.01594001054763794, + 0.013573607429862022, + 0.012143142521381378, + 0.01696852222084999, + 0.00814824365079403, + 0.014066552743315697, + 0.006434421055018902, + 0.014784667640924454, + 0.012447441928088665, + 0.016314763575792313, + 0.01587999239563942, + 0.013631945475935936, + 0.016650259494781494, + 0.011838502250611782, + 0.022087756544351578, + 0.007853608578443527, + 0.007311629131436348, + 0.005151091609150171, + 0.008457559160888195, + 0.0075403968803584576, + 0.008019860833883286, + 0.009946946054697037, + 0.008842947892844677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.029793230816721916 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_9/index_tensor([273, 273])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_9/index_tensor([273, 273])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7a1f0240cddc6e34444725b5c6b0a65863fba43d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_9/index_tensor([273, 273])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, horizontal shot captured with a telephoto lens draws the viewer's attention to an ancient, black-sailed boat, anchored in the center of the frame, gently swaying rhythmically with the waves. The boat's hull is constructed of dark, weathered planks, wet with water marks near the waterline. Its black sail, supported by bamboo skeins, its oiled surface gleaming faintly in the setting sun. Golden afternoon sunlight slants across the shimmering river, each rocking of the boat creating tiny, shimmering specks of golden light. The intense compression effect of the telephoto lens renders the ancient stone arch bridge in the distance both imposing and imminent. The arch and its reflection in the water form a nearly perfect circle, clearly revealing the mottled texture of the stone, etched with age, and dotted with moss. Weeping willows lining the riverbank sway gently in the breeze, their strands occasionally touching the still water, creating ripples. The entire picture is shrouded in a warm and soft golden halo. The main colors are a blend of sunset gold and river blue, dotted with the inky black of the black sailboats and the emerald green of the willow trees, creating a quiet, distant and poetic atmosphere of the Jiangnan water town.", + "seed": 30462, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_9/index_273_seed_30462_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02099919430911541, + "vlm_logprob": -0.02099919430911541, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.02099919430911541, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019498461857438087 + ], + "opposite_logprob_tokens": [ + -5.269498348236084 + ], + "logprob_reward": -0.019498461857438087, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.019498461857438087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21685312688350677, + 0.08637993782758713, + 0.056937042623758316, + 0.05440685898065567, + 0.0529799610376358, + 0.042633675038814545, + 0.17616449296474457, + 0.08646690100431442, + 0.12179672718048096, + 0.10538125783205032 + ], + "credit_map": [ + 0.03163132444024086, + 0.039539165794849396, + 0.024426935240626335, + 0.035637106746435165, + 0.027893267571926117, + 0.020557913929224014, + 0.020123686641454697, + 0.01704373210668564, + 0.011788329109549522, + 0.015265091322362423, + 0.00789874792098999, + 0.014771525748074055, + 0.012414708733558655, + 0.008949619717895985, + 0.010043969377875328, + 0.005247947294265032, + 0.008690211921930313, + 0.008800076320767403, + 0.004726092331111431, + 0.006856788881123066, + 0.010378779843449593, + 0.007694492116570473, + 0.005195224191993475, + 0.004595377016812563, + 0.0068572754971683025, + 0.008378642611205578, + 0.0048465365543961525, + 0.0065489355474710464, + 0.008565891534090042, + 0.007440196815878153, + 0.005129744298756123, + 0.006639634259045124, + 0.008152037858963013, + 0.00898387748748064, + 0.007040360476821661, + 0.005696214735507965, + 0.0059296078979969025, + 0.00647088373079896, + 0.004756769631057978, + 0.005950208753347397, + 0.00584775023162365, + 0.006810476072132587, + 0.004450610373169184, + 0.0039054653607308865, + 0.006845234893262386, + 0.006226453464478254, + 0.00376307126134634, + 0.004784612916409969, + 0.023570600897073746, + 0.020978635177016258, + 0.01797860488295555, + 0.013531955890357494, + 0.04166330769658089, + 0.023981284350156784, + 0.019286470487713814, + 0.015173637308180332, + 0.013111799955368042, + 0.013947736471891403, + 0.007828352972865105, + 0.008274910971522331, + 0.01285639125853777, + 0.012444944120943546, + 0.010434848256409168, + 0.007567914668470621, + 0.01818903535604477, + 0.030265698209404945, + 0.013656694442033768, + 0.010649852454662323, + 0.014258562587201595, + 0.009543187916278839, + 0.017979122698307037, + 0.0072545697912573814, + 0.014447457157075405, + 0.016091885045170784, + 0.008118879050016403, + 0.013370170257985592, + 0.018479403108358383, + 0.008380941115319729, + 0.019455812871456146, + 0.007036707829684019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.019498461857438087 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982354536652565 + ], + "opposite_logprob_tokens": [ + -3.5298235416412354 + ], + "logprob_reward": -0.02982354536652565, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982354536652565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30138513445854187, + 0.08087436854839325, + 0.08116348832845688, + 0.043788351118564606, + 0.03165455535054207, + 0.0220175851136446, + 0.23945418000221252, + 0.07297812402248383, + 0.038932908326387405, + 0.08775129169225693 + ], + "credit_map": [ + 0.0453372485935688, + 0.05812516808509827, + 0.033822283148765564, + 0.04978644475340843, + 0.03948121890425682, + 0.02131381444633007, + 0.028751805424690247, + 0.024767128750681877, + 0.010800768621265888, + 0.014307951554656029, + 0.008956017903983593, + 0.008160856552422047, + 0.017282474786043167, + 0.008758609183132648, + 0.007909598760306835, + 0.004698098637163639, + 0.010359385050833225, + 0.010774951428174973, + 0.010235365480184555, + 0.008319295011460781, + 0.015144694596529007, + 0.008847150951623917, + 0.009233364835381508, + 0.008249285630881786, + 0.005628020968288183, + 0.004756907466799021, + 0.0039965007454156876, + 0.004311383236199617, + 0.009994659572839737, + 0.005319874733686447, + 0.00502240564674139, + 0.004758600145578384, + 0.003937073051929474, + 0.004443613346666098, + 0.0038178686518222094, + 0.003937745001167059, + 0.004859003704041243, + 0.004761259537190199, + 0.0029495737981051207, + 0.0029484210535883904, + 0.0029131327755749226, + 0.0032823067158460617, + 0.0027872424107044935, + 0.002311090240254998, + 0.002575694816187024, + 0.002813813043758273, + 0.002665419364348054, + 0.002668885514140129, + 0.03373977914452553, + 0.02834531106054783, + 0.024337204173207283, + 0.01772221177816391, + 0.06422603875398636, + 0.028241358697414398, + 0.02524692751467228, + 0.017595352604985237, + 0.009844671003520489, + 0.012068502604961395, + 0.0069508482702076435, + 0.006562178023159504, + 0.0134709682315588, + 0.00940408930182457, + 0.008607211522758007, + 0.006069660186767578, + 0.006745416671037674, + 0.006787662394344807, + 0.003498603357002139, + 0.003213393036276102, + 0.006062576547265053, + 0.0044628931209445, + 0.004313244018703699, + 0.0038491173181682825, + 0.01149691827595234, + 0.013324208557605743, + 0.008566731587052345, + 0.013076220639050007, + 0.01077338308095932, + 0.006860577501356602, + 0.01700882986187935, + 0.0066444240510463715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.02982354536652565 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820182055234909 + ], + "opposite_logprob_tokens": [ + -4.01820182800293 + ], + "logprob_reward": -0.01820182055234909, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820182055234909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15684780478477478, + 0.05978969484567642, + 0.09277180582284927, + 0.08663022518157959, + 0.06080372631549835, + 0.0221706535667181, + 0.06134956330060959, + 0.12031954526901245, + 0.032573338598012924, + 0.30674365162849426 + ], + "credit_map": [ + 0.024603888392448425, + 0.02935917302966118, + 0.014480484649538994, + 0.022646546363830566, + 0.02071065455675125, + 0.016154542565345764, + 0.014918704517185688, + 0.013973813503980637, + 0.008877101354300976, + 0.009834842756390572, + 0.005915297195315361, + 0.005738128907978535, + 0.013103753328323364, + 0.006356625352054834, + 0.0052452669478952885, + 0.00471867760643363, + 0.012139520607888699, + 0.014121360145509243, + 0.010939986445009708, + 0.008241835050284863, + 0.017491595819592476, + 0.010772795416414738, + 0.009179269894957542, + 0.009885437786579132, + 0.011679829098284245, + 0.011923720128834248, + 0.008848068304359913, + 0.006631968542933464, + 0.015299118123948574, + 0.01304036658257246, + 0.00635510915890336, + 0.012852041982114315, + 0.00826425664126873, + 0.014099983498454094, + 0.007357078604400158, + 0.005346841644495726, + 0.006834612227976322, + 0.008170116692781448, + 0.0043705180287361145, + 0.006360317580401897, + 0.0035849842242896557, + 0.0046139867044985294, + 0.0018663995433598757, + 0.002863082801923156, + 0.002375619253143668, + 0.0026677227579057217, + 0.002377596218138933, + 0.00182126194704324, + 0.008347253315150738, + 0.013687712140381336, + 0.006383110769093037, + 0.007079641334712505, + 0.006127435714006424, + 0.006327688228338957, + 0.007206689566373825, + 0.0061900317668914795, + 0.01817445084452629, + 0.029478639364242554, + 0.014810435473918915, + 0.01181037724018097, + 0.009799261577427387, + 0.010789047926664352, + 0.012848705053329468, + 0.012608624063432217, + 0.005679180845618248, + 0.005675111431628466, + 0.0027747817803174257, + 0.002810893813148141, + 0.005238765850663185, + 0.003854552051052451, + 0.003837883472442627, + 0.0027021660935133696, + 0.04112724959850311, + 0.042297933250665665, + 0.030361603945493698, + 0.05184164270758629, + 0.034535396844148636, + 0.01950392872095108, + 0.06509106606245041, + 0.021984823048114777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.01820182055234909 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418977603316307 + ], + "opposite_logprob_tokens": [ + -4.264189720153809 + ], + "logprob_reward": -0.01418977603316307, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418977603316307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1732283979654312, + 0.145672008395195, + 0.10142631828784943, + 0.09962962567806244, + 0.044802188873291016, + 0.05025434121489525, + 0.04741813614964485, + 0.12529537081718445, + 0.1277480274438858, + 0.08452557027339935 + ], + "credit_map": [ + 0.026480799540877342, + 0.0306208748370409, + 0.018168855458498, + 0.020151205360889435, + 0.019996078684926033, + 0.024898646399378777, + 0.01392786018550396, + 0.018984070047736168, + 0.022252537310123444, + 0.025391776114702225, + 0.010003779083490372, + 0.011344620026648045, + 0.03039650432765484, + 0.020483894273638725, + 0.011989103630185127, + 0.013809789903461933, + 0.015373148024082184, + 0.015282507054507732, + 0.0100882388651371, + 0.01114663016051054, + 0.010763945057988167, + 0.012704666703939438, + 0.013105602003633976, + 0.012961586005985737, + 0.01342387031763792, + 0.014465337619185448, + 0.010835962370038033, + 0.013988474383950233, + 0.015338284894824028, + 0.010998556390404701, + 0.011404856108129025, + 0.009174280799925327, + 0.006255695596337318, + 0.005338518414646387, + 0.006787944119423628, + 0.004073456861078739, + 0.005188090726733208, + 0.00749333156272769, + 0.0038062932435423136, + 0.0058588627725839615, + 0.007196629419922829, + 0.00933719053864479, + 0.00526264077052474, + 0.0035254843533039093, + 0.0077318972907960415, + 0.007379305548965931, + 0.004454617388546467, + 0.005366574041545391, + 0.007450159173458815, + 0.009066029451787472, + 0.004501170478761196, + 0.004773658700287342, + 0.009219169616699219, + 0.004775105509907007, + 0.0045052203349769115, + 0.0031276189256459475, + 0.0202163215726614, + 0.03306644409894943, + 0.010140759870409966, + 0.008933336474001408, + 0.014316650107502937, + 0.0161973275244236, + 0.012267217971384525, + 0.010157318785786629, + 0.018843406811356544, + 0.019701197743415833, + 0.014683632180094719, + 0.01588226482272148, + 0.012428502552211285, + 0.01509989332407713, + 0.020389290526509285, + 0.010719829238951206, + 0.015381182543933392, + 0.014036939479410648, + 0.00900734681636095, + 0.008279516361653805, + 0.010467593558132648, + 0.008766324259340763, + 0.011059909127652645, + 0.0075267525389790535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.01418977603316307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328236773610115 + ], + "opposite_logprob_tokens": [ + -3.773282289505005 + ], + "logprob_reward": -0.02328236773610115, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328236773610115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11090946942567825, + 0.09444233030080795, + 0.2786099910736084, + 0.1290842443704605, + 0.030040331184864044, + 0.02571677789092064, + 0.13970649242401123, + 0.05943545326590538, + 0.069593645632267, + 0.062461331486701965 + ], + "credit_map": [ + 0.015279140323400497, + 0.015721028670668602, + 0.013857409358024597, + 0.01425100490450859, + 0.010794670321047306, + 0.016360551118850708, + 0.011691324412822723, + 0.012954328209161758, + 0.014587478712201118, + 0.015049860812723637, + 0.00846325233578682, + 0.008286233991384506, + 0.018089963123202324, + 0.015260308980941772, + 0.005568054970353842, + 0.009137165732681751, + 0.03836534917354584, + 0.040208831429481506, + 0.029646387323737144, + 0.02266920730471611, + 0.07161574065685272, + 0.03115389496088028, + 0.021243611350655556, + 0.023706942796707153, + 0.017756741493940353, + 0.016461901366710663, + 0.012194068171083927, + 0.011860352009534836, + 0.027937056496739388, + 0.017855610698461533, + 0.010395302437245846, + 0.014623197726905346, + 0.004565403331071138, + 0.0034236274659633636, + 0.0044871545396745205, + 0.0041082557290792465, + 0.0038033367600291967, + 0.004005576483905315, + 0.0037451358512043953, + 0.001901834737509489, + 0.004077215678989887, + 0.004922050517052412, + 0.0036716863978654146, + 0.002363887382671237, + 0.00279192253947258, + 0.003125432413071394, + 0.0022944409865885973, + 0.0024701403453946114, + 0.019747328013181686, + 0.01702790893614292, + 0.015065334737300873, + 0.011347263120114803, + 0.03588433191180229, + 0.015306543558835983, + 0.014896713197231293, + 0.010431045666337013, + 0.00867434497922659, + 0.011026732623577118, + 0.005895963404327631, + 0.005704768002033234, + 0.009399261325597763, + 0.0067807831801474094, + 0.006727184168994427, + 0.0052264039404690266, + 0.011481106281280518, + 0.011699089780449867, + 0.00699494918808341, + 0.006802062503993511, + 0.009654897265136242, + 0.006995555944740772, + 0.010949531570076942, + 0.00501645402982831, + 0.007804848253726959, + 0.012216639705002308, + 0.0071031078696250916, + 0.0062597282230854034, + 0.00875515304505825, + 0.005209237802773714, + 0.009443315677344799, + 0.0056692915968596935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.02328236773610115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, horizontal shot captured with a telephoto lens draws the viewer's attention to an ancient, black-sailed boat, anchored in the center of the frame, gently swaying rhythmically with the waves. The boat's hull is constructed of dark, weathered planks, wet with water marks near the waterline. Its black sail, supported by bamboo skeins, its oiled surface gleaming faintly in the setting sun. Golden afternoon sunlight slants across the shimmering river, each rocking of the boat creating tiny, shimmering specks of golden light. The intense compression effect of the telephoto lens renders the ancient stone arch bridge in the distance both imposing and imminent. The arch and its reflection in the water form a nearly perfect circle, clearly revealing the mottled texture of the stone, etched with age, and dotted with moss. Weeping willows lining the riverbank sway gently in the breeze, their strands occasionally touching the still water, creating ripples. The entire picture is shrouded in a warm and soft golden halo. The main colors are a blend of sunset gold and river blue, dotted with the inky black of the black sailboats and the emerald green of the willow trees, creating a quiet, distant and poetic atmosphere of the Jiangnan water town.", + "seed": 30463, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_9/index_273_seed_30463_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01715188715606928, + "vlm_logprob": -0.01715188715606928, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.01715188715606928, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017444511875510216 + ], + "opposite_logprob_tokens": [ + -4.642444610595703 + ], + "logprob_reward": -0.017444511875510216, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.017444511875510216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2103591412305832, + 0.07644698023796082, + 0.07825683802366257, + 0.10962021350860596, + 0.12117090076208115, + 0.05745452642440796, + 0.10300715267658234, + 0.0628645122051239, + 0.0844922885298729, + 0.09632743895053864 + ], + "credit_map": [ + 0.05814554914832115, + 0.013585184700787067, + 0.07295700907707214, + 0.018215656280517578, + 0.01503483485430479, + 0.011607824824750423, + 0.010356467217206955, + 0.010456602089107037, + 0.024454733356833458, + 0.005273750051856041, + 0.02385610155761242, + 0.006147740874439478, + 0.005827616434544325, + 0.0033405851572752, + 0.0046721589751541615, + 0.0028742963913828135, + 0.020988784730434418, + 0.010272955521941185, + 0.01562024001032114, + 0.006698674988001585, + 0.007698862347751856, + 0.006214994005858898, + 0.006874986924231052, + 0.003887339262291789, + 0.02100890874862671, + 0.016523564234375954, + 0.015380516648292542, + 0.012313543818891048, + 0.013061868026852608, + 0.010866474360227585, + 0.013025017455220222, + 0.007440315559506416, + 0.034030769020318985, + 0.009309104643762112, + 0.027908969670534134, + 0.012759779579937458, + 0.011077137663960457, + 0.010565448552370071, + 0.010226817801594734, + 0.005292877554893494, + 0.011782210320234299, + 0.0069163707084953785, + 0.007498957682400942, + 0.005716174840927124, + 0.008025581017136574, + 0.008067863993346691, + 0.006024269387125969, + 0.0034230954479426146, + 0.015759138390421867, + 0.017219016328454018, + 0.009173242375254631, + 0.014151524752378464, + 0.014885869808495045, + 0.013449960388243198, + 0.011413403786718845, + 0.006954996380954981, + 0.01280828844755888, + 0.005352369975298643, + 0.012920713983476162, + 0.00584125192835927, + 0.008006523363292217, + 0.006587478332221508, + 0.006804252043366432, + 0.004543629940599203, + 0.020073989406228065, + 0.008957671001553535, + 0.02033010497689247, + 0.0069480049423873425, + 0.00944957323372364, + 0.0087743466719985, + 0.005520075559616089, + 0.004438517149537802, + 0.01460331305861473, + 0.0136668486520648, + 0.008289345540106297, + 0.012906341813504696, + 0.013889760710299015, + 0.013548347167670727, + 0.011222627945244312, + 0.008200851269066334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.017444511875510216 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020595736801624298 + ], + "opposite_logprob_tokens": [ + -3.8955957889556885 + ], + "logprob_reward": -0.020595736801624298, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020595736801624298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2615402340888977, + 0.052453529089689255, + 0.05383610352873802, + 0.07002287358045578, + 0.15169593691825867, + 0.06534182280302048, + 0.09749003499746323, + 0.07255961000919342, + 0.060161519795656204, + 0.11489836126565933 + ], + "credit_map": [ + 0.07628924399614334, + 0.02137323096394539, + 0.0814797654747963, + 0.018453676253557205, + 0.021329214796423912, + 0.021387869492173195, + 0.011829809285700321, + 0.009397408924996853, + 0.0097566032782197, + 0.004667473025619984, + 0.013311400078237057, + 0.005949733778834343, + 0.005399987567216158, + 0.004523345269262791, + 0.005544684361666441, + 0.0033002961426973343, + 0.014301210641860962, + 0.004362167324870825, + 0.01426734495908022, + 0.004703357815742493, + 0.004650528077036142, + 0.004082018975168467, + 0.004295226186513901, + 0.0031742467544972897, + 0.012859358452260494, + 0.009034932591021061, + 0.015521463006734848, + 0.005284291226416826, + 0.01003397535532713, + 0.007347031030803919, + 0.005838699638843536, + 0.004103115294128656, + 0.03510236367583275, + 0.011800657026469707, + 0.03877640515565872, + 0.01588476076722145, + 0.014220038428902626, + 0.011580748483538628, + 0.016384759917855263, + 0.007946193218231201, + 0.012784275226294994, + 0.007305764127522707, + 0.008293929509818554, + 0.009230826050043106, + 0.007382933050394058, + 0.007726916577666998, + 0.008670247159898281, + 0.003946925047785044, + 0.01573042944073677, + 0.013374811969697475, + 0.0076230363920331, + 0.008798904716968536, + 0.017154525965452194, + 0.016780361533164978, + 0.009745062328875065, + 0.00828289333730936, + 0.015601127408444881, + 0.007391719147562981, + 0.014233316294848919, + 0.008171583525836468, + 0.006739651784300804, + 0.006112303119152784, + 0.008729543536901474, + 0.005580362863838673, + 0.011605695821344852, + 0.006342373322695494, + 0.0075028459541499615, + 0.008689343929290771, + 0.007610387168824673, + 0.007027449086308479, + 0.006829965393990278, + 0.004553451668471098, + 0.02467656508088112, + 0.010057437233626842, + 0.029920045286417007, + 0.013237304985523224, + 0.009269406087696552, + 0.007871594280004501, + 0.012418004684150219, + 0.007447993382811546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.020595736801624298 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086649261415005 + ], + "opposite_logprob_tokens": [ + -4.511086463928223 + ], + "logprob_reward": -0.011086649261415005, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086649261415005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21716730296611786, + 0.05207429826259613, + 0.03280693292617798, + 0.04656105116009712, + 0.24335704743862152, + 0.0620686411857605, + 0.10852610319852829, + 0.0783848986029625, + 0.12335294485092163, + 0.0357007272541523 + ], + "credit_map": [ + 0.06527997553348541, + 0.021921971812844276, + 0.07368212193250656, + 0.015544107183814049, + 0.013717671856284142, + 0.01075630821287632, + 0.008339263498783112, + 0.00792588572949171, + 0.009647361934185028, + 0.00564230652526021, + 0.013346994295716286, + 0.006022145040333271, + 0.0036255724262446165, + 0.003437188919633627, + 0.005893087945878506, + 0.00445963628590107, + 0.008324145339429379, + 0.0031999675557017326, + 0.006508240010589361, + 0.0027245432138442993, + 0.0030454108491539955, + 0.002771596424281597, + 0.0039449469186365604, + 0.0022880854085087776, + 0.008205950260162354, + 0.006896784529089928, + 0.007915947586297989, + 0.004430077970027924, + 0.006389569025486708, + 0.00419765617698431, + 0.004934233613312244, + 0.00359083223156631, + 0.07007003575563431, + 0.019089721143245697, + 0.054982807487249374, + 0.028402680531144142, + 0.0218817126005888, + 0.014727188274264336, + 0.0219755657017231, + 0.012227346189320087, + 0.014655956998467445, + 0.006361646577715874, + 0.011693667620420456, + 0.007741493638604879, + 0.005905205383896828, + 0.004840580280870199, + 0.0070227510295808315, + 0.0038473389577120543, + 0.021110419183969498, + 0.014353923499584198, + 0.01500793918967247, + 0.009795771911740303, + 0.019115250557661057, + 0.013171859085559845, + 0.009429694153368473, + 0.006541249342262745, + 0.012123788706958294, + 0.007448202930390835, + 0.009790265932679176, + 0.009789819829165936, + 0.012861812487244606, + 0.011117007583379745, + 0.009177797473967075, + 0.006076205056160688, + 0.016288723796606064, + 0.018518172204494476, + 0.01601235195994377, + 0.01760188490152359, + 0.01504652388393879, + 0.01303758006542921, + 0.016073763370513916, + 0.010773940943181515, + 0.00629153847694397, + 0.002635866403579712, + 0.0053545283153653145, + 0.007024578750133514, + 0.0031070068944245577, + 0.0028641773387789726, + 0.005414529703557491, + 0.003008499974384904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.011086649261415005 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057226374745369 + ], + "opposite_logprob_tokens": [ + -3.8955721855163574 + ], + "logprob_reward": -0.02057226374745369, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057226374745369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11712108552455902, + 0.23155760765075684, + 0.060748081654310226, + 0.14525508880615234, + 0.07794944941997528, + 0.05784192308783531, + 0.12243665009737015, + 0.04485334828495979, + 0.07659731060266495, + 0.06563951075077057 + ], + "credit_map": [ + 0.015096654184162617, + 0.017828309908509254, + 0.011411179788410664, + 0.011414553970098495, + 0.02013593353331089, + 0.01851457916200161, + 0.01367418933659792, + 0.009045680053532124, + 0.06174340844154358, + 0.028376571834087372, + 0.06022873520851135, + 0.02156675048172474, + 0.01722240447998047, + 0.013067696243524551, + 0.01742277294397354, + 0.011929250322282314, + 0.014707349240779877, + 0.007302076555788517, + 0.012322012335062027, + 0.005638144910335541, + 0.005962038412690163, + 0.005295214243233204, + 0.0058097755536437035, + 0.003711463650688529, + 0.0376477874815464, + 0.012907417491078377, + 0.040266525000333786, + 0.01286461390554905, + 0.013081501238048077, + 0.010636493563652039, + 0.01068940106779337, + 0.007161343470215797, + 0.010813726112246513, + 0.008984670042991638, + 0.008014758117496967, + 0.00814782828092575, + 0.013454165309667587, + 0.013510762713849545, + 0.009190103970468044, + 0.005833425093442202, + 0.009412026032805443, + 0.007105736061930656, + 0.007381978444755077, + 0.008318357169628143, + 0.008180691860616207, + 0.00721777742728591, + 0.006103466730564833, + 0.004121882375329733, + 0.018753012642264366, + 0.019658194854855537, + 0.013502730056643486, + 0.01457296870648861, + 0.017406225204467773, + 0.017452599480748177, + 0.01280052401125431, + 0.008290378376841545, + 0.012222694233059883, + 0.004122292157262564, + 0.012315787374973297, + 0.004522555973380804, + 0.0037895115092396736, + 0.0024448642507195473, + 0.0033187982626259327, + 0.0021168370731174946, + 0.016228294000029564, + 0.007030266337096691, + 0.017044205218553543, + 0.006417439319193363, + 0.009419629350304604, + 0.009960134513676167, + 0.0052979388274252415, + 0.005199394654482603, + 0.014970891177654266, + 0.00562677439302206, + 0.016105463728308678, + 0.006797150243073702, + 0.007843660190701485, + 0.005120023153722286, + 0.0058275507763028145, + 0.003347985213622451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.02057226374745369 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016060274094343185 + ], + "opposite_logprob_tokens": [ + -4.1410603523254395 + ], + "logprob_reward": -0.016060274094343185, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016060274094343185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23919080197811127, + 0.12326134741306305, + 0.048214055597782135, + 0.24866288900375366, + 0.050287049263715744, + 0.02655957080423832, + 0.09436162561178207, + 0.05860380455851555, + 0.048024777323007584, + 0.06283412873744965 + ], + "credit_map": [ + 0.06423641741275787, + 0.024351006373763084, + 0.07096638530492783, + 0.022003861144185066, + 0.02086555026471615, + 0.012773496098816395, + 0.01445011142641306, + 0.009543937630951405, + 0.038748301565647125, + 0.008876360952854156, + 0.042083509266376495, + 0.008701454848051071, + 0.007830506190657616, + 0.006447460036724806, + 0.005889724940061569, + 0.004684017971158028, + 0.013746091164648533, + 0.004268063697963953, + 0.014315107837319374, + 0.0032957852818071842, + 0.0033118445426225662, + 0.0032012236770242453, + 0.0038839068729430437, + 0.0021920306608080864, + 0.039674870669841766, + 0.042110323905944824, + 0.025352265685796738, + 0.027475424110889435, + 0.037131473422050476, + 0.028244923800230026, + 0.029180623590946198, + 0.019492948427796364, + 0.011347576975822449, + 0.0048659988678991795, + 0.008894812315702438, + 0.004542867187410593, + 0.0061059752479195595, + 0.005921036470681429, + 0.00542703690007329, + 0.003181738778948784, + 0.0046774111688137054, + 0.00306700705550611, + 0.003071397077292204, + 0.003938074689358473, + 0.0037291832268238068, + 0.003912047948688269, + 0.0023156844545155764, + 0.0018487614579498768, + 0.024730918928980827, + 0.006618010811507702, + 0.025709088891744614, + 0.008051713928580284, + 0.009988831356167793, + 0.007587427739053965, + 0.006424963939934969, + 0.005250663496553898, + 0.009288948029279709, + 0.00654608802869916, + 0.004093095660209656, + 0.00775173120200634, + 0.009998667053878307, + 0.008714061230421066, + 0.007511910516768694, + 0.004699298180639744, + 0.005033848341554403, + 0.006584034767001867, + 0.006014000624418259, + 0.0063688973896205425, + 0.006491277366876602, + 0.006206423509865999, + 0.006541622802615166, + 0.004784670192748308, + 0.01511996891349554, + 0.0052229538559913635, + 0.018081223592162132, + 0.007210828829556704, + 0.00463530607521534, + 0.004321808461099863, + 0.005464307032525539, + 0.0027777221985161304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.016060274094343185 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, horizontal shot captured with a telephoto lens draws the viewer's attention to an ancient, black-sailed boat, anchored in the center of the frame, gently swaying rhythmically with the waves. The boat's hull is constructed of dark, weathered planks, wet with water marks near the waterline. Its black sail, supported by bamboo skeins, its oiled surface gleaming faintly in the setting sun. Golden afternoon sunlight slants across the shimmering river, each rocking of the boat creating tiny, shimmering specks of golden light. The intense compression effect of the telephoto lens renders the ancient stone arch bridge in the distance both imposing and imminent. The arch and its reflection in the water form a nearly perfect circle, clearly revealing the mottled texture of the stone, etched with age, and dotted with moss. Weeping willows lining the riverbank sway gently in the breeze, their strands occasionally touching the still water, creating ripples. The entire picture is shrouded in a warm and soft golden halo. The main colors are a blend of sunset gold and river blue, dotted with the inky black of the black sailboats and the emerald green of the willow trees, creating a quiet, distant and poetic atmosphere of the Jiangnan water town.", + "seed": 30464, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_9/index_273_seed_30464_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.022743193432688713, + "vlm_logprob": -0.022743193432688713, + "vlm_yes_no_margin": 3.974999952316284, + "vlm_token_credit": -0.022743193432688713, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017833132296800613 + ], + "opposite_logprob_tokens": [ + -4.517833232879639 + ], + "logprob_reward": -0.017833132296800613, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.017833132296800613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22455750405788422, + 0.16045759618282318, + 0.03093002177774906, + 0.0431823693215847, + 0.02954018861055374, + 0.028418274596333504, + 0.04933818802237511, + 0.2501915395259857, + 0.12245883047580719, + 0.06092555448412895 + ], + "credit_map": [ + 0.02361369878053665, + 0.02126425690948963, + 0.013709082268178463, + 0.032591771334409714, + 0.019803155213594437, + 0.03358129784464836, + 0.017957892268896103, + 0.06203630939126015, + 0.011702542193233967, + 0.01134366262704134, + 0.00856321956962347, + 0.045450177043676376, + 0.007989934645593166, + 0.014043522998690605, + 0.013742887414991856, + 0.04762163385748863, + 0.002875883597880602, + 0.0034555881284177303, + 0.004252977203577757, + 0.005165834911167622, + 0.0033714452292770147, + 0.003886215388774872, + 0.004246084950864315, + 0.003675985848531127, + 0.007512409705668688, + 0.005307681392878294, + 0.005844832397997379, + 0.0053987870924174786, + 0.005370383616536856, + 0.004648334812372923, + 0.004882951267063618, + 0.004216982517391443, + 0.0027730788569897413, + 0.0036607878282666206, + 0.0029916828498244286, + 0.0048988936468958855, + 0.003550502471625805, + 0.003829211462289095, + 0.00463142478838563, + 0.003204602748155594, + 0.003344692289829254, + 0.003877266775816679, + 0.00248132785782218, + 0.0032808207906782627, + 0.0030476460233330727, + 0.0061842892318964005, + 0.002578127197921276, + 0.003624100238084793, + 0.006367987021803856, + 0.005015539471060038, + 0.003707894589751959, + 0.006228980142623186, + 0.007258511148393154, + 0.00961347483098507, + 0.0054078553803265095, + 0.005737938452512026, + 0.01720607280731201, + 0.03609014302492142, + 0.025711778551340103, + 0.04826167970895767, + 0.0284469835460186, + 0.01679094508290291, + 0.048829276114702225, + 0.02885463647544384, + 0.013229107484221458, + 0.013399575836956501, + 0.010239935480058193, + 0.01764431595802307, + 0.015493905171751976, + 0.012124168686568737, + 0.02547287382185459, + 0.014854931272566319, + 0.006812666542828083, + 0.010291473008692265, + 0.00758587708696723, + 0.008740180172026157, + 0.00658103683963418, + 0.006313010584563017, + 0.008834941312670708, + 0.005766361020505428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.017833132296800613 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038119103759527206 + ], + "opposite_logprob_tokens": [ + -3.288119077682495 + ], + "logprob_reward": -0.038119103759527206, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038119103759527206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33757930994033813, + 0.11119427531957626, + 0.06812112778425217, + 0.026248617097735405, + 0.05892852321267128, + 0.042788393795490265, + 0.06026312708854675, + 0.1457882672548294, + 0.02770140953361988, + 0.12138701230287552 + ], + "credit_map": [ + 0.03539976105093956, + 0.02946801669895649, + 0.02713361382484436, + 0.09105291217565536, + 0.030459163710474968, + 0.03842135891318321, + 0.038044799119234085, + 0.047599636018276215, + 0.009168684482574463, + 0.007970038801431656, + 0.006221272051334381, + 0.02775355614721775, + 0.007856380194425583, + 0.011087020859122276, + 0.011097104288637638, + 0.030040211975574493, + 0.007810937240719795, + 0.008787051774561405, + 0.006007466930896044, + 0.00965889822691679, + 0.007135879714041948, + 0.008084633387625217, + 0.00892015639692545, + 0.011716095730662346, + 0.003213335294276476, + 0.0031963263172656298, + 0.002570463577285409, + 0.004279927350580692, + 0.0034270642790943384, + 0.002707236912101507, + 0.0041859629563987255, + 0.0026682980824261904, + 0.009204988367855549, + 0.005846698302775621, + 0.0068821110762655735, + 0.0067051732912659645, + 0.008861622773110867, + 0.009033091366291046, + 0.006848917808383703, + 0.0055459183640778065, + 0.007641523610800505, + 0.005501734092831612, + 0.003928483463823795, + 0.004449361469596624, + 0.005509476643055677, + 0.007576156873255968, + 0.0038134423084557056, + 0.00436820974573493, + 0.008883890695869923, + 0.0058458768762648106, + 0.006102373357862234, + 0.007683189120143652, + 0.008831336162984371, + 0.00807748083025217, + 0.007284082472324371, + 0.007554890587925911, + 0.010187315754592419, + 0.022536447271704674, + 0.014729080721735954, + 0.02768886648118496, + 0.016447298228740692, + 0.009840521030128002, + 0.02660905383527279, + 0.017749661579728127, + 0.0036480918060988188, + 0.0033338326029479504, + 0.0031660853419452906, + 0.003192236414179206, + 0.0037189119029790163, + 0.003887064754962921, + 0.0036511507350951433, + 0.003104033414274454, + 0.014723975211381912, + 0.02523024193942547, + 0.01235923357307911, + 0.013949821703135967, + 0.013561995700001717, + 0.014430197887122631, + 0.013405724428594112, + 0.013725802302360535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.038119103759527206 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819959655404091 + ], + "opposite_logprob_tokens": [ + -4.018199443817139 + ], + "logprob_reward": -0.01819959655404091, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819959655404091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2764933109283447, + 0.08698684722185135, + 0.08593561500310898, + 0.03519602492451668, + 0.04854537174105644, + 0.041078634560108185, + 0.08711748570203781, + 0.1206057146191597, + 0.03620219975709915, + 0.1818387657403946 + ], + "credit_map": [ + 0.023615216836333275, + 0.026133937761187553, + 0.02166583202779293, + 0.06684096902608871, + 0.02671506069600582, + 0.025460127741098404, + 0.030029136687517166, + 0.05603302642703056, + 0.007752778474241495, + 0.007831611670553684, + 0.007199740502983332, + 0.01806025393307209, + 0.007349331397563219, + 0.0088943000882864, + 0.008717037737369537, + 0.021181795746088028, + 0.008770258165895939, + 0.009341073222458363, + 0.011509038507938385, + 0.013235128484666348, + 0.011715578846633434, + 0.009724299423396587, + 0.010268093086779118, + 0.011372143402695656, + 0.004125221632421017, + 0.004917174112051725, + 0.0031352033838629723, + 0.005052467342466116, + 0.005591507535427809, + 0.005394685547798872, + 0.003608147846534848, + 0.0033716149628162384, + 0.005372216925024986, + 0.005356233101338148, + 0.004967795219272375, + 0.008072989992797375, + 0.006652279756963253, + 0.006136925425380468, + 0.006089056842029095, + 0.005897874478250742, + 0.006164608057588339, + 0.006345346104353666, + 0.003907088655978441, + 0.00524229696020484, + 0.0034495180007070303, + 0.005926555022597313, + 0.005146094597876072, + 0.004897126462310553, + 0.012592334300279617, + 0.010309134609997272, + 0.008549347519874573, + 0.009154561907052994, + 0.010431608185172081, + 0.01463242620229721, + 0.010561242699623108, + 0.010886830277740955, + 0.007820533588528633, + 0.016352999955415726, + 0.01390906609594822, + 0.023738713935017586, + 0.012936901301145554, + 0.007571284659206867, + 0.023760877549648285, + 0.014515338465571404, + 0.004291783086955547, + 0.004204616416245699, + 0.0038319509476423264, + 0.0042147450149059296, + 0.0049342019483447075, + 0.00484356377273798, + 0.003996217157691717, + 0.005885122809559107, + 0.0206795334815979, + 0.035630058497190475, + 0.022043777629733086, + 0.023710990324616432, + 0.021081555634737015, + 0.017524968832731247, + 0.021480169147253036, + 0.01968771405518055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.01819959655404091 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776921942830086 + ], + "opposite_logprob_tokens": [ + -4.634777069091797 + ], + "logprob_reward": -0.009776921942830086, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776921942830086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15662962198257446, + 0.07131238281726837, + 0.060052648186683655, + 0.040256526321172714, + 0.041218072175979614, + 0.03355467692017555, + 0.03237491846084595, + 0.30658525228500366, + 0.03541556000709534, + 0.2226003110408783 + ], + "credit_map": [ + 0.01994927041232586, + 0.015915647149086, + 0.011998987756669521, + 0.030399372801184654, + 0.016705892980098724, + 0.0307061318308115, + 0.01744706928730011, + 0.01350725069642067, + 0.004929274786263704, + 0.005511708091944456, + 0.005063273478299379, + 0.01686575636267662, + 0.0035048394929617643, + 0.007999755442142487, + 0.004978307522833347, + 0.022459477186203003, + 0.006544465199112892, + 0.0065399641171097755, + 0.01043542381376028, + 0.007154598832130432, + 0.0059366472996771336, + 0.00963568314909935, + 0.005849570967257023, + 0.007956298068165779, + 0.0057297502644360065, + 0.005498537793755531, + 0.003980169538408518, + 0.004375355318188667, + 0.004774052184075117, + 0.006849752273410559, + 0.0049196877516806126, + 0.00412922166287899, + 0.005883840378373861, + 0.004361539613455534, + 0.003353717038407922, + 0.003581642173230648, + 0.005733340047299862, + 0.008026751689612865, + 0.004681440535932779, + 0.00559580372646451, + 0.00576282711699605, + 0.0046337018720805645, + 0.0036119683645665646, + 0.0033640414476394653, + 0.0031951041892170906, + 0.005772816017270088, + 0.0037767435424029827, + 0.0034374799579381943, + 0.005517417564988136, + 0.003313192166388035, + 0.002913424978032708, + 0.0031306997407227755, + 0.004286601673811674, + 0.005062885116785765, + 0.003988130949437618, + 0.004162568598985672, + 0.020521873608231544, + 0.0465397983789444, + 0.031042257323861122, + 0.05901876464486122, + 0.0373329222202301, + 0.020494868978857994, + 0.057305000722408295, + 0.03432980552315712, + 0.0028580098878592253, + 0.0043684071861207485, + 0.006101568695157766, + 0.006306719034910202, + 0.003229580819606781, + 0.003394279396161437, + 0.004572956822812557, + 0.004584039095789194, + 0.023762866854667664, + 0.04140550643205643, + 0.03280859440565109, + 0.028897056356072426, + 0.024064645171165466, + 0.017315248027443886, + 0.03050863742828369, + 0.023837760090827942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.009776921942830086 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978721261024475 + ], + "opposite_logprob_tokens": [ + -3.529787302017212 + ], + "logprob_reward": -0.02978721261024475, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978721261024475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2955990433692932, + 0.08365828543901443, + 0.10961423069238663, + 0.03755262866616249, + 0.03420436382293701, + 0.051173239946365356, + 0.062412817031145096, + 0.0896928459405899, + 0.08819715678691864, + 0.14789538085460663 + ], + "credit_map": [ + 0.029951084405183792, + 0.02677999995648861, + 0.022476471960544586, + 0.0678352415561676, + 0.030301596969366074, + 0.047558121383190155, + 0.035787712782621384, + 0.03490882366895676, + 0.005864843726158142, + 0.006298489402979612, + 0.006509746424853802, + 0.01903527043759823, + 0.006148906424641609, + 0.009112363681197166, + 0.007542875595390797, + 0.02314579300582409, + 0.011716151610016823, + 0.011572214774787426, + 0.021086513996124268, + 0.018035147339105606, + 0.010879157111048698, + 0.011811617761850357, + 0.012854407541453838, + 0.011659018695354462, + 0.0040396424010396, + 0.005241121165454388, + 0.002945926273241639, + 0.004257322754710913, + 0.004995245952159166, + 0.006907802540808916, + 0.005178100895136595, + 0.003987463656812906, + 0.003979659639298916, + 0.003839087439700961, + 0.0030246260575950146, + 0.005191293079406023, + 0.00464292848482728, + 0.004665716551244259, + 0.004646699875593185, + 0.0042143515311181545, + 0.0068382155150175095, + 0.007068045902997255, + 0.0047543724067509174, + 0.007592353038489819, + 0.005438567139208317, + 0.006779382936656475, + 0.0073655517771840096, + 0.005336749367415905, + 0.010474592447280884, + 0.006365014240145683, + 0.006872452329844236, + 0.005720342043787241, + 0.008483597077429295, + 0.010731113143265247, + 0.006060181651264429, + 0.007705522235482931, + 0.006919681094586849, + 0.012861684896051884, + 0.008686217479407787, + 0.015553176403045654, + 0.011817674152553082, + 0.007105094846338034, + 0.01605410687625408, + 0.010695206932723522, + 0.007734265644103289, + 0.01049184799194336, + 0.008892004378139973, + 0.013211443088948727, + 0.010509808547794819, + 0.008698209188878536, + 0.017355138435959816, + 0.011304441839456558, + 0.015154394321143627, + 0.029054518789052963, + 0.019605379551649094, + 0.018902411684393883, + 0.016462750732898712, + 0.011912300251424313, + 0.022160222753882408, + 0.014643410220742226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.02978721261024475 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, horizontal shot captured with a telephoto lens draws the viewer's attention to an ancient, black-sailed boat, anchored in the center of the frame, gently swaying rhythmically with the waves. The boat's hull is constructed of dark, weathered planks, wet with water marks near the waterline. Its black sail, supported by bamboo skeins, its oiled surface gleaming faintly in the setting sun. Golden afternoon sunlight slants across the shimmering river, each rocking of the boat creating tiny, shimmering specks of golden light. The intense compression effect of the telephoto lens renders the ancient stone arch bridge in the distance both imposing and imminent. The arch and its reflection in the water form a nearly perfect circle, clearly revealing the mottled texture of the stone, etched with age, and dotted with moss. Weeping willows lining the riverbank sway gently in the breeze, their strands occasionally touching the still water, creating ripples. The entire picture is shrouded in a warm and soft golden halo. The main colors are a blend of sunset gold and river blue, dotted with the inky black of the black sailboats and the emerald green of the willow trees, creating a quiet, distant and poetic atmosphere of the Jiangnan water town.", + "seed": 30465, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000040/group_9/index_273_seed_30465_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.020663913898169996, + "vlm_logprob": -0.020663913898169996, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.020663913898169996, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017839808017015457 + ], + "opposite_logprob_tokens": [ + -5.6428399085998535 + ], + "logprob_reward": -0.017839808017015457, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.017839808017015457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14647896587848663, + 0.11721276491880417, + 0.03772711753845215, + 0.03365011513233185, + 0.03334764763712883, + 0.0667359009385109, + 0.030603647232055664, + 0.029175983741879463, + 0.4835824966430664, + 0.02148537151515484 + ], + "credit_map": [ + 0.02292056940495968, + 0.019432103261351585, + 0.016973771154880524, + 0.028804048895835876, + 0.01056662667542696, + 0.007868064567446709, + 0.027454936876893044, + 0.012458843179047108, + 0.03175013139843941, + 0.01266903430223465, + 0.039221763610839844, + 0.009212619625031948, + 0.006846141070127487, + 0.0038543499540537596, + 0.009327137842774391, + 0.004331599920988083, + 0.006172884721308947, + 0.006189592182636261, + 0.005057399161159992, + 0.004206307232379913, + 0.0038360634353011847, + 0.002864984329789877, + 0.007235982920974493, + 0.002163904719054699, + 0.005190787836909294, + 0.0029388752300292253, + 0.003522323677316308, + 0.004334932193160057, + 0.0037855729460716248, + 0.0028623698744922876, + 0.0075863199308514595, + 0.0034289327450096607, + 0.00567731074988842, + 0.005156213417649269, + 0.004092424642294645, + 0.004828902892768383, + 0.002675232943147421, + 0.0031531669665127993, + 0.00504361791536212, + 0.0027207776438444853, + 0.009909972548484802, + 0.00910457968711853, + 0.005654993001371622, + 0.007551050279289484, + 0.008782004937529564, + 0.0055092559196054935, + 0.015025632455945015, + 0.005198415368795395, + 0.006734623108059168, + 0.0028139932546764612, + 0.0073197828605771065, + 0.0043617794290184975, + 0.0015860655112192035, + 0.0015628981636837125, + 0.003880811156705022, + 0.002343693282455206, + 0.007324855774641037, + 0.00301427417434752, + 0.009265395812690258, + 0.002437545685097575, + 0.0018433755030855536, + 0.0011856992496177554, + 0.0025242315605282784, + 0.0015806065639480948, + 0.15556076169013977, + 0.02210584655404091, + 0.19777090847492218, + 0.026544656604528427, + 0.020381314679980278, + 0.010817552916705608, + 0.03524259105324745, + 0.01515883021056652, + 0.0031522016506642103, + 0.0025172620080411434, + 0.002257732441648841, + 0.003453879151493311, + 0.0021898637060076, + 0.0016542136436328292, + 0.0040518855676054955, + 0.002208334393799305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.017839808017015457 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023308223113417625 + ], + "opposite_logprob_tokens": [ + -3.773308277130127 + ], + "logprob_reward": -0.023308223113417625, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023308223113417625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33405235409736633, + 0.14671744406223297, + 0.050581395626068115, + 0.04557011276483536, + 0.04446879029273987, + 0.0666482001543045, + 0.02914281189441681, + 0.027012042701244354, + 0.18062150478363037, + 0.0751853957772255 + ], + "credit_map": [ + 0.09269478917121887, + 0.03231450170278549, + 0.12371624261140823, + 0.019655032083392143, + 0.015098197385668755, + 0.009570155292749405, + 0.027594227343797684, + 0.013409141451120377, + 0.04056772589683533, + 0.015574141405522823, + 0.05198746174573898, + 0.00976648274809122, + 0.008386814035475254, + 0.004597986117005348, + 0.010327570140361786, + 0.005509242415428162, + 0.00801881868392229, + 0.005938312038779259, + 0.0079916100949049, + 0.00944723840802908, + 0.004217989277094603, + 0.0028628765139728785, + 0.00765779335051775, + 0.004446750041097403, + 0.006446394138038158, + 0.006846877280622721, + 0.004369495436549187, + 0.007639086805284023, + 0.004815659951418638, + 0.003189720446243882, + 0.007962162606418133, + 0.004300713539123535, + 0.007234212011098862, + 0.008192581124603748, + 0.005559045821428299, + 0.007427498232573271, + 0.0034695244394242764, + 0.004046601243317127, + 0.005404896102845669, + 0.0031344282906502485, + 0.01061010267585516, + 0.009872817434370518, + 0.006205863784998655, + 0.008096717298030853, + 0.009034517221152782, + 0.005855720955878496, + 0.01240823045372963, + 0.004564221948385239, + 0.004378298297524452, + 0.003299969481304288, + 0.0033354703336954117, + 0.0042318920604884624, + 0.0024508715141564608, + 0.00232866033911705, + 0.005320087540894747, + 0.003797559766098857, + 0.005783600267022848, + 0.004209740087389946, + 0.007039795629680157, + 0.0020548913162201643, + 0.0027352147735655308, + 0.0017048969166353345, + 0.0021390984766185284, + 0.001344800926744938, + 0.05716735124588013, + 0.009807505644857883, + 0.07192539423704147, + 0.010484496131539345, + 0.00821742881089449, + 0.004292013123631477, + 0.012934020720422268, + 0.005793269723653793, + 0.015574990771710873, + 0.011149470694363117, + 0.017813552170991898, + 0.007475585211068392, + 0.008046969771385193, + 0.0050076572224497795, + 0.0063520800322294235, + 0.0037650777958333492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.023308223113417625 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819409430027008 + ], + "opposite_logprob_tokens": [ + -4.018194198608398 + ], + "logprob_reward": -0.01819409430027008, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819409430027008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.118942990899086, + 0.22392049431800842, + 0.06496105343103409, + 0.07383287698030472, + 0.02808559313416481, + 0.02183961682021618, + 0.03671099618077278, + 0.05482642352581024, + 0.34171101450920105, + 0.035168927162885666 + ], + "credit_map": [ + 0.022986415773630142, + 0.020480243489146233, + 0.02393355779349804, + 0.016051482409238815, + 0.010070291347801685, + 0.006788710597902536, + 0.011840637773275375, + 0.006791653111577034, + 0.06017250567674637, + 0.026162313297390938, + 0.07490864396095276, + 0.020040377974510193, + 0.012119822204113007, + 0.007614328525960445, + 0.015248707495629787, + 0.007653777953237295, + 0.009585080668330193, + 0.01003468781709671, + 0.014016925357282162, + 0.008267946541309357, + 0.005754203535616398, + 0.004090267699211836, + 0.008465542457997799, + 0.004746399819850922, + 0.010857285931706429, + 0.009806470945477486, + 0.010622587986290455, + 0.014281951822340488, + 0.007840447127819061, + 0.005709364544600248, + 0.010115478187799454, + 0.004599289037287235, + 0.00499641103670001, + 0.0032128498423844576, + 0.004489243496209383, + 0.005885357968509197, + 0.0015887325862422585, + 0.0017768505495041609, + 0.0040479060262441635, + 0.0020882419776171446, + 0.003385744523257017, + 0.002444029785692692, + 0.004309582524001598, + 0.003483093110844493, + 0.001477167010307312, + 0.001580649521201849, + 0.0029548227321356535, + 0.002204527612775564, + 0.006987945642322302, + 0.002660175319761038, + 0.006201224401593208, + 0.005610250402241945, + 0.0035811234265565872, + 0.002347825327888131, + 0.0064595602452754974, + 0.0028628925792872906, + 0.011485329829156399, + 0.004054035525768995, + 0.015772059559822083, + 0.00546646723523736, + 0.003449964104220271, + 0.0032065853010863066, + 0.00727071100845933, + 0.004121273290365934, + 0.1104394719004631, + 0.013670251704752445, + 0.1428610235452652, + 0.018055496737360954, + 0.013980639167129993, + 0.007009898778051138, + 0.02533070743083954, + 0.010363533161580563, + 0.005650701001286507, + 0.004463580437004566, + 0.004974636249244213, + 0.005245653446763754, + 0.0035813041031360626, + 0.002612412441521883, + 0.005292925518006086, + 0.003347712801769376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.01819409430027008 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189188368618488 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.014189188368618488, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189188368618488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34715795516967773, + 0.12558065354824066, + 0.17551881074905396, + 0.03490207716822624, + 0.022171109914779663, + 0.02509460784494877, + 0.049938399344682693, + 0.04852179437875748, + 0.1270296722650528, + 0.04408496990799904 + ], + "credit_map": [ + 0.05415276810526848, + 0.05901208519935608, + 0.03839082643389702, + 0.07027598470449448, + 0.02554105967283249, + 0.019223937764763832, + 0.053590331226587296, + 0.02697095274925232, + 0.033356454223394394, + 0.011144198477268219, + 0.0395074337720871, + 0.013927625492215157, + 0.005571063607931137, + 0.005499578546732664, + 0.010752853006124496, + 0.005821448750793934, + 0.026843566447496414, + 0.018473966047167778, + 0.02189185470342636, + 0.027069106698036194, + 0.01717664860188961, + 0.010109232738614082, + 0.04045257717370987, + 0.013501860201358795, + 0.005890040658414364, + 0.0027121102903038263, + 0.004046082962304354, + 0.004535192158073187, + 0.0031302969437092543, + 0.002535527106374502, + 0.009055120870471, + 0.0029977052472531796, + 0.0037271741311997175, + 0.0026357281021773815, + 0.0033314856700599194, + 0.004859532695263624, + 0.0015108621446415782, + 0.0019055104348808527, + 0.0024428588803857565, + 0.0017579577397555113, + 0.004003642126917839, + 0.005173849407583475, + 0.00303430063650012, + 0.0030618610326200724, + 0.0032430998980998993, + 0.0019491373095661402, + 0.0030278973281383514, + 0.0016008191742002964, + 0.011113721877336502, + 0.003922239877283573, + 0.011225784197449684, + 0.006479149218648672, + 0.004473934881389141, + 0.003191559575498104, + 0.006734491791576147, + 0.0027975139673799276, + 0.008943360298871994, + 0.008355481550097466, + 0.010894854553043842, + 0.0049705784767866135, + 0.004302285611629486, + 0.002888598246499896, + 0.005086363293230534, + 0.003080271650105715, + 0.0390758290886879, + 0.007057223469018936, + 0.04892421141266823, + 0.007859780453145504, + 0.0061676762998104095, + 0.00391295226290822, + 0.009652882814407349, + 0.004379123914986849, + 0.00847587175667286, + 0.0037437868304550648, + 0.009532371535897255, + 0.005657666828483343, + 0.0038231676444411278, + 0.0032459511421620846, + 0.006301871966570616, + 0.0033042817376554012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.014189188368618488 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot static and horizontal, capturing an ancient black-sailed boat anchored in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull constructed of dark, weathered planks with wet water marks near the waterline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boat gently swaying rhythmically with the waves while golden sunlight creates shimmering specks on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ancient stone arch bridge visible in the distance with a mottled texture, moss, and a reflection forming a nearly perfect circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are weeping willows lining the riverbank swaying gently with strands touching the water, under a warm golden halo with sunset gold and river blue colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978825569152832 + ], + "opposite_logprob_tokens": [ + -3.5297882556915283 + ], + "logprob_reward": -0.02978825569152832, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978825569152832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14433720707893372, + 0.1466570645570755, + 0.10438170284032822, + 0.06513293832540512, + 0.030771875753998756, + 0.043679460883140564, + 0.05154075846076012, + 0.06802933663129807, + 0.23182900249958038, + 0.11364059150218964 + ], + "credit_map": [ + 0.028332414105534554, + 0.03109596110880375, + 0.024107087403535843, + 0.01696767657995224, + 0.011562476865947247, + 0.008096400648355484, + 0.01617398113012314, + 0.008001217618584633, + 0.04419231787323952, + 0.007051980122923851, + 0.05589129030704498, + 0.013508138246834278, + 0.004751925356686115, + 0.004673202522099018, + 0.010812959633767605, + 0.00577526492998004, + 0.016044512391090393, + 0.009515712037682533, + 0.01068484503775835, + 0.015396804548799992, + 0.010827522724866867, + 0.006255213171243668, + 0.027255898341536522, + 0.008401199243962765, + 0.009934114292263985, + 0.007386582903563976, + 0.01015994418412447, + 0.014618024230003357, + 0.005211923271417618, + 0.004576048348098993, + 0.008917929604649544, + 0.004328378010541201, + 0.004774346947669983, + 0.003981026355177164, + 0.004092223942279816, + 0.006129004061222076, + 0.0022542995866388083, + 0.0020769229158759117, + 0.005081098061054945, + 0.0023829557467252016, + 0.007769336458295584, + 0.005032385233789682, + 0.006283036898821592, + 0.006093366537243128, + 0.005228763446211815, + 0.0036159115843474865, + 0.007086702156811953, + 0.002569963224232197, + 0.012925393879413605, + 0.005037847440689802, + 0.014364873059093952, + 0.005475268699228764, + 0.003277245908975601, + 0.002761246869340539, + 0.00463259918615222, + 0.003066291566938162, + 0.014648349024355412, + 0.008931479416787624, + 0.019220897927880287, + 0.007892309688031673, + 0.003364638891071081, + 0.003464294131845236, + 0.00682807108387351, + 0.0036793057806789875, + 0.07314083725214005, + 0.010609441436827183, + 0.09594166278839111, + 0.013786587864160538, + 0.008951633237302303, + 0.004670239519327879, + 0.017463497817516327, + 0.007265119347721338, + 0.02021128498017788, + 0.015213335864245892, + 0.018007755279541016, + 0.014760030433535576, + 0.011057829484343529, + 0.008728159591555595, + 0.017267916351556778, + 0.008394299075007439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.02978825569152832 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_0/index_tensor([608, 608])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_0/index_tensor([608, 608])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0744a77016b17b553015d68f9bf09f23185b77d6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_0/index_tensor([608, 608])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium-length shot triggers a slow, sweeping pan around a majestic African bush elephant, instantly establishing its immense size. The warm, golden light of the setting sun meticulously sculpts the creature's form, capturing the deep, intricate network of wrinkles in its weathered, gray-brown hide and illuminating the fine lines in its gently flapping, fan-shaped ears. Its long, curved tusks gleam softly in the dim light. The elephant advances with a steady, deliberate gait, its powerful footsteps firmly pounding the dry earth and stirring up tiny, shimmering clouds of dust suspended in the still air. As the camera continues its circular motion, the scope of its epic trek is revealed: endless expanses of sun-scorched golden grass stretch to the horizon, dotted with the stark silhouettes of scattered acacia trees. The entire scene is imbued with an atmosphere of profound tranquility and majesty, capturing a timeless moment of a solitary pilgrimage through the vast, untamed wilderness.", + "seed": 28453, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_0/index_608_seed_28453_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10440622288733721, + "vlm_logprob": -0.10440622288733721, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.10440622288733721, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0776551216840744 + ], + "opposite_logprob_tokens": [ + -2.952655076980591 + ], + "logprob_reward": -0.0776551216840744, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0776551216840744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1418880820274353, + 0.20625583827495575, + 0.07445238530635834, + 0.05962325632572174, + 0.06182020902633667, + 0.0919426679611206, + 0.07708984613418579, + 0.11034666001796722, + 0.057522378861904144, + 0.11905872076749802 + ], + "credit_map": [ + 0.01880236342549324, + 0.024672189727425575, + 0.01560977753251791, + 0.01712184213101864, + 0.018112726509571075, + 0.025178834795951843, + 0.01462620124220848, + 0.007764140143990517, + 0.029606791213154793, + 0.03444230556488037, + 0.025999078527092934, + 0.021769098937511444, + 0.02239682525396347, + 0.030917594209313393, + 0.022079162299633026, + 0.019044969230890274, + 0.01208691205829382, + 0.009938505478203297, + 0.007963228039443493, + 0.009946703910827637, + 0.009481903165578842, + 0.009408844634890556, + 0.005873943213373423, + 0.009752340614795685, + 0.007438475266098976, + 0.00800375360995531, + 0.007598691154271364, + 0.007042359560728073, + 0.004918260034173727, + 0.007093736901879311, + 0.005974763073027134, + 0.011553213931620121, + 0.009266860783100128, + 0.00893340166658163, + 0.006857285276055336, + 0.009940866380929947, + 0.0071517471224069595, + 0.009967932477593422, + 0.004672800190746784, + 0.005029311869293451, + 0.012409971095621586, + 0.009172861464321613, + 0.01987391524016857, + 0.011254988610744476, + 0.01022722665220499, + 0.010446726344525814, + 0.007094565313309431, + 0.01146240346133709, + 0.01287134736776352, + 0.00871430803090334, + 0.013566369190812111, + 0.007643610239028931, + 0.00783760566264391, + 0.012708612717688084, + 0.00593167170882225, + 0.007816319353878498, + 0.01210635807365179, + 0.016580801457166672, + 0.016542738303542137, + 0.01699201762676239, + 0.01506973896175623, + 0.0115634361281991, + 0.010626260191202164, + 0.010865302756428719, + 0.008090555667877197, + 0.006929784081876278, + 0.00613194378092885, + 0.004320915788412094, + 0.0065866815857589245, + 0.015157775953412056, + 0.004072565585374832, + 0.00623214989900589, + 0.015006416477262974, + 0.015711359679698944, + 0.012494448572397232, + 0.014133544638752937, + 0.01620342954993248, + 0.021589478477835655, + 0.007450122851878405, + 0.016469908878207207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0776551216840744 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016257792711258 + ], + "opposite_logprob_tokens": [ + -1.7016258239746094 + ], + "logprob_reward": -0.2016257792711258, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016257792711258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35086825489997864, + 0.0664595365524292, + 0.04568899795413017, + 0.02041299268603325, + 0.14039885997772217, + 0.03557974472641945, + 0.03655899688601494, + 0.12759247422218323, + 0.08442670851945877, + 0.0920134037733078 + ], + "credit_map": [ + 0.0349758043885231, + 0.044208381325006485, + 0.04710561782121658, + 0.07894179970026016, + 0.04550449177622795, + 0.057290218770504, + 0.023724514991044998, + 0.01911742426455021, + 0.011661180295050144, + 0.006875839550048113, + 0.0076575204730033875, + 0.008784051984548569, + 0.006842012982815504, + 0.009923003613948822, + 0.0073900166898965836, + 0.007325913291424513, + 0.008221816271543503, + 0.0063481805846095085, + 0.005930015817284584, + 0.0057211751118302345, + 0.004924692213535309, + 0.006156862713396549, + 0.003230118425562978, + 0.0051561384461820126, + 0.003029661485925317, + 0.002609407529234886, + 0.00251796399243176, + 0.0018485202454030514, + 0.002061770763248205, + 0.0036519255954772234, + 0.0022029222454875708, + 0.0024908205959945917, + 0.02365857921540737, + 0.019486993551254272, + 0.01806466281414032, + 0.01964866928756237, + 0.013862313702702522, + 0.022128740325570107, + 0.010796788148581982, + 0.012752103619277477, + 0.003867724444717169, + 0.005627261009067297, + 0.0034682212863117456, + 0.004007945768535137, + 0.0033252984285354614, + 0.007312352769076824, + 0.002204706659540534, + 0.005766235291957855, + 0.003730956930667162, + 0.0063878921791911125, + 0.004814865533262491, + 0.004560163710266352, + 0.005150923505425453, + 0.004379528574645519, + 0.003487321315333247, + 0.004047346767038107, + 0.015825405716896057, + 0.01657005399465561, + 0.020962540060281754, + 0.01538466289639473, + 0.015715613961219788, + 0.01761050522327423, + 0.011656422168016434, + 0.013867267407476902, + 0.012800606898963451, + 0.012300354428589344, + 0.007828171364963055, + 0.010219278745353222, + 0.01037874910980463, + 0.016461357474327087, + 0.005258248653262854, + 0.009179938584566116, + 0.014177219942212105, + 0.013138736598193645, + 0.010722351260483265, + 0.012209567241370678, + 0.010455786250531673, + 0.013885905034840107, + 0.005906855687499046, + 0.011516977101564407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.2016257792711258 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988058924674988 + ], + "opposite_logprob_tokens": [ + -1.8048806190490723 + ], + "logprob_reward": -0.17988058924674988, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988058924674988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17734842002391815, + 0.03859097883105278, + 0.05084250494837761, + 0.066067174077034, + 0.06044768914580345, + 0.12187239527702332, + 0.12143047153949738, + 0.15759843587875366, + 0.09169276803731918, + 0.11410919576883316 + ], + "credit_map": [ + 0.0178988054394722, + 0.019868286326527596, + 0.023687683045864105, + 0.04078303650021553, + 0.02167131006717682, + 0.03029109537601471, + 0.013354292139410973, + 0.009793905541300774, + 0.007031616289168596, + 0.00396030442789197, + 0.005710233002901077, + 0.0033986414782702923, + 0.0031436558347195387, + 0.008311543613672256, + 0.0020245304331183434, + 0.005010452587157488, + 0.007012816611677408, + 0.005773360375314951, + 0.006503531243652105, + 0.00753534073010087, + 0.005578584969043732, + 0.005381443537771702, + 0.005513595882803202, + 0.0075438255444169044, + 0.007846036925911903, + 0.008245262317359447, + 0.007797425612807274, + 0.010706793516874313, + 0.007417005952447653, + 0.007661794312298298, + 0.007817707024514675, + 0.008575141429901123, + 0.007953260093927383, + 0.009576457552611828, + 0.006872615776956081, + 0.010713772848248482, + 0.0055139814503490925, + 0.00841523241251707, + 0.004985236097127199, + 0.006417125463485718, + 0.021331438794732094, + 0.01124257780611515, + 0.032569095492362976, + 0.014698098413646221, + 0.009193552657961845, + 0.013865658082067966, + 0.006115165539085865, + 0.012856795452535152, + 0.015988105908036232, + 0.018397770822048187, + 0.021122286096215248, + 0.01598234660923481, + 0.012931299395859241, + 0.016214342787861824, + 0.010551970452070236, + 0.010242336429655552, + 0.011299899779260159, + 0.021050473675131798, + 0.02125733532011509, + 0.019867757335305214, + 0.024168310686945915, + 0.02064092457294464, + 0.0178282018750906, + 0.02148551680147648, + 0.01845497079193592, + 0.013019188307225704, + 0.010328758507966995, + 0.008526990190148354, + 0.01126101054251194, + 0.014865336008369923, + 0.005872790701687336, + 0.009363723918795586, + 0.016028719022870064, + 0.013920673169195652, + 0.012682621367275715, + 0.010637345723807812, + 0.013509358279407024, + 0.025198446586728096, + 0.006205972749739885, + 0.0159260556101799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.17988058924674988 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866715520620346 + ], + "opposite_logprob_tokens": [ + -3.0486671924591064 + ], + "logprob_reward": -0.04866715520620346, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866715520620346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2444058358669281, + 0.04401879757642746, + 0.058935970067977905, + 0.01980896294116974, + 0.06018860638141632, + 0.10624232888221741, + 0.04517173394560814, + 0.1588587760925293, + 0.15821267664432526, + 0.10415638238191605 + ], + "credit_map": [ + 0.028083346784114838, + 0.028015505522489548, + 0.031742993742227554, + 0.04933843016624451, + 0.03356992453336716, + 0.04145375266671181, + 0.01750790700316429, + 0.014693963341414928, + 0.006772753316909075, + 0.006661382503807545, + 0.004948955029249191, + 0.005238588433712721, + 0.005477220751345158, + 0.0066572390496730804, + 0.003716312116011977, + 0.004546342883259058, + 0.007420144509524107, + 0.01028938964009285, + 0.006672627758234739, + 0.010578632354736328, + 0.007486133370548487, + 0.005830190144479275, + 0.006284549366682768, + 0.004374297801405191, + 0.0031086630187928677, + 0.003043234348297119, + 0.002547140698879957, + 0.002656706143170595, + 0.002255433239042759, + 0.002641444094479084, + 0.001781838247552514, + 0.0017745021032169461, + 0.008496998809278011, + 0.009053045883774757, + 0.007334454916417599, + 0.009617150761187077, + 0.005329721607267857, + 0.007909711450338364, + 0.00560140423476696, + 0.006846113130450249, + 0.014985133893787861, + 0.010215546935796738, + 0.027787577360868454, + 0.01383963506668806, + 0.009231195785105228, + 0.013292660936713219, + 0.005075153429061174, + 0.011815417557954788, + 0.006449502892792225, + 0.005681847687810659, + 0.008324404247105122, + 0.003917096182703972, + 0.006208414677530527, + 0.007082193624228239, + 0.0027986601926386356, + 0.004709612112492323, + 0.012274887412786484, + 0.020978346467018127, + 0.02094762958586216, + 0.021195244044065475, + 0.024867510423064232, + 0.019190268591046333, + 0.01878684014081955, + 0.020618055015802383, + 0.02740195393562317, + 0.024226883426308632, + 0.01575980894267559, + 0.012928014621138573, + 0.016977280378341675, + 0.03632477670907974, + 0.00809336919337511, + 0.01650058850646019, + 0.014746597968041897, + 0.015424799174070358, + 0.01174866408109665, + 0.010998503305017948, + 0.011487318202853203, + 0.020439933985471725, + 0.005677992478013039, + 0.01363256573677063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.04866715520620346 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202469028532505 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202469028532505, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202469028532505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1431509256362915, + 0.08584335446357727, + 0.05376758798956871, + 0.058791741728782654, + 0.08333092927932739, + 0.09823819994926453, + 0.1990550458431244, + 0.1433013528585434, + 0.07836011052131653, + 0.05616078898310661 + ], + "credit_map": [ + 0.020217137411236763, + 0.02022293210029602, + 0.017835717648267746, + 0.022364523261785507, + 0.014020362868905067, + 0.03060307726264, + 0.009665240533649921, + 0.00822193268686533, + 0.014176913537085056, + 0.00957044493407011, + 0.011427192017436028, + 0.010906307026743889, + 0.007283845450729132, + 0.015094938687980175, + 0.0061418176628649235, + 0.01124188955873251, + 0.011954125016927719, + 0.008320435881614685, + 0.006881871726363897, + 0.007078970782458782, + 0.005495287477970123, + 0.0061335270293056965, + 0.0033104142639786005, + 0.004592950455844402, + 0.008520638570189476, + 0.007485765498131514, + 0.00946712400764227, + 0.005576690658926964, + 0.005025249440222979, + 0.00925474800169468, + 0.004347075708210468, + 0.00911444891244173, + 0.012880763970315456, + 0.011532747186720371, + 0.009404974989593029, + 0.01444259937852621, + 0.009594093076884747, + 0.01250496320426464, + 0.006524732336401939, + 0.006446051876991987, + 0.014298553578555584, + 0.009385914541780949, + 0.02413018047809601, + 0.011427628807723522, + 0.008374598808586597, + 0.010282126255333424, + 0.006176450289785862, + 0.014162739738821983, + 0.023478349670767784, + 0.025646384805440903, + 0.03360169380903244, + 0.018900208175182343, + 0.022891899570822716, + 0.03830467909574509, + 0.01204818394035101, + 0.024183647707104683, + 0.012336536310613155, + 0.019015535712242126, + 0.0179347712546587, + 0.019766736775636673, + 0.023143678903579712, + 0.01658986322581768, + 0.016838738694787025, + 0.017675485461950302, + 0.01306950580328703, + 0.012587936595082283, + 0.009660783223807812, + 0.0066473581828176975, + 0.008836117573082447, + 0.016088902950286865, + 0.004327922593802214, + 0.007141576148569584, + 0.00917538721114397, + 0.007574114017188549, + 0.006430264096707106, + 0.0054109483025968075, + 0.00576860411092639, + 0.010676284320652485, + 0.004164618905633688, + 0.006960560102015734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.014202469028532505 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-length shot triggers a slow, sweeping pan around a majestic African bush elephant, instantly establishing its immense size. The warm, golden light of the setting sun meticulously sculpts the creature's form, capturing the deep, intricate network of wrinkles in its weathered, gray-brown hide and illuminating the fine lines in its gently flapping, fan-shaped ears. Its long, curved tusks gleam softly in the dim light. The elephant advances with a steady, deliberate gait, its powerful footsteps firmly pounding the dry earth and stirring up tiny, shimmering clouds of dust suspended in the still air. As the camera continues its circular motion, the scope of its epic trek is revealed: endless expanses of sun-scorched golden grass stretch to the horizon, dotted with the stark silhouettes of scattered acacia trees. The entire scene is imbued with an atmosphere of profound tranquility and majesty, capturing a timeless moment of a solitary pilgrimage through the vast, untamed wilderness.", + "seed": 28454, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_0/index_608_seed_28454_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1177666824311018, + "vlm_logprob": -0.1177666824311018, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.1177666824311018, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09269740432500839 + ], + "opposite_logprob_tokens": [ + -2.7176973819732666 + ], + "logprob_reward": -0.09269740432500839, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09269740432500839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10644151270389557, + 0.04523571953177452, + 0.03698896989226341, + 0.03242425620555878, + 0.02693856880068779, + 0.011126714758574963, + 0.0973634421825409, + 0.5921894907951355, + 0.012752244248986244, + 0.038539063185453415 + ], + "credit_map": [ + 0.026550177484750748, + 0.012618159875273705, + 0.0063925436697900295, + 0.004664580337703228, + 0.025920717045664787, + 0.017271725460886955, + 0.00510628055781126, + 0.007917333394289017, + 0.005239467136561871, + 0.008699292317032814, + 0.00357623677700758, + 0.004706819076091051, + 0.005758348386734724, + 0.01155727356672287, + 0.0024575423449277878, + 0.0032407392282038927, + 0.005865611135959625, + 0.012245140969753265, + 0.0024495129473507404, + 0.003869503503665328, + 0.004826114512979984, + 0.002584411296993494, + 0.0024886264000087976, + 0.0026600516866892576, + 0.003607427002862096, + 0.00487600639462471, + 0.003710563760250807, + 0.004902251996099949, + 0.00412624841555953, + 0.0043444521725177765, + 0.0030193468555808067, + 0.003837959375232458, + 0.0030612340196967125, + 0.003900340758264065, + 0.010812654159963131, + 0.0018725375412032008, + 0.0017965050647035241, + 0.0029056111816316843, + 0.0012956511927768588, + 0.0012940349988639355, + 0.0010615757200866938, + 0.0016281567513942719, + 0.0013725939206779003, + 0.0013045641826465726, + 0.0018660937203094363, + 0.001411376753821969, + 0.0009113233536481857, + 0.0015710300067439675, + 0.005827707704156637, + 0.015090336091816425, + 0.010238900780677795, + 0.03405597060918808, + 0.013077112846076488, + 0.006890947464853525, + 0.006113908719271421, + 0.006068551912903786, + 0.050570953637361526, + 0.037657614797353745, + 0.03358212858438492, + 0.056395530700683594, + 0.10342475026845932, + 0.2386060655117035, + 0.016829919070005417, + 0.05512254685163498, + 0.0007717196131125093, + 0.003932431805878878, + 0.0012814176734536886, + 0.0031500766053795815, + 0.0011747698299586773, + 0.0009429727215319872, + 0.0007523638196289539, + 0.0007464921800419688, + 0.004288746975362301, + 0.004133278504014015, + 0.012417874298989773, + 0.00409702118486166, + 0.0030255867168307304, + 0.0038026554975658655, + 0.0015759309753775597, + 0.005197967868298292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.09269740432500839 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25217103958129883 + ], + "opposite_logprob_tokens": [ + -1.5021710395812988 + ], + "logprob_reward": -0.25217103958129883, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25217103958129883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2194889485836029, + 0.04492753744125366, + 0.04924530163407326, + 0.05931901931762695, + 0.07035087049007416, + 0.027967363595962524, + 0.05727555230259895, + 0.1958521157503128, + 0.06993136554956436, + 0.2056419402360916 + ], + "credit_map": [ + 0.05813410505652428, + 0.02225903794169426, + 0.01452881470322609, + 0.008367134258151054, + 0.03194000944495201, + 0.04270373284816742, + 0.010509472340345383, + 0.031046638265252113, + 0.004836258012801409, + 0.011765769682824612, + 0.004690851550549269, + 0.003672226332128048, + 0.006474383641034365, + 0.008348511531949043, + 0.002645710948854685, + 0.0024938276037573814, + 0.010879269801080227, + 0.016202576458454132, + 0.0038050550501793623, + 0.003026028163731098, + 0.005682999733835459, + 0.005018526688218117, + 0.0021583738271147013, + 0.0024724688846617937, + 0.012430379167199135, + 0.010823550634086132, + 0.00771422078832984, + 0.0070772236213088036, + 0.006723673082888126, + 0.005537224933505058, + 0.004183613695204258, + 0.0048291333951056, + 0.008902343921363354, + 0.01573236472904682, + 0.029398590326309204, + 0.004364019259810448, + 0.004295869264751673, + 0.0031347277108579874, + 0.0023976529482752085, + 0.0021253018639981747, + 0.0036630616523325443, + 0.003548699663951993, + 0.0028424072079360485, + 0.0035700188018381596, + 0.004137896001338959, + 0.004298756830394268, + 0.0020007826387882233, + 0.003905742196366191, + 0.018231181427836418, + 0.016435891389846802, + 0.004168559331446886, + 0.00502065708860755, + 0.005949744954705238, + 0.003763127140700817, + 0.0021051315125077963, + 0.0016012591077014804, + 0.012631099671125412, + 0.01138564758002758, + 0.008832963183522224, + 0.011207197792828083, + 0.03265704959630966, + 0.07154221832752228, + 0.007870377041399479, + 0.03972557187080383, + 0.0031508298125118017, + 0.025453558191657066, + 0.005360413808375597, + 0.02376200258731842, + 0.004088487941771746, + 0.0033895177766680717, + 0.002774380147457123, + 0.0019521762151271105, + 0.020276881754398346, + 0.04767253249883652, + 0.06417699158191681, + 0.01968572661280632, + 0.010440809652209282, + 0.020304111763834953, + 0.00552300876006484, + 0.017561882734298706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.25217103958129883 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372138366103172 + ], + "opposite_logprob_tokens": [ + -3.408721446990967 + ], + "logprob_reward": -0.03372138366103172, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372138366103172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09024975448846817, + 0.08068997412919998, + 0.14874720573425293, + 0.1056465357542038, + 0.1037340983748436, + 0.06997112184762955, + 0.136876180768013, + 0.19134241342544556, + 0.03848313167691231, + 0.03425963595509529 + ], + "credit_map": [ + 0.02067134901881218, + 0.009899782948195934, + 0.005893779452890158, + 0.0046092551201581955, + 0.02106085605919361, + 0.013120015151798725, + 0.004473474808037281, + 0.010521245189011097, + 0.017658617347478867, + 0.008304106071591377, + 0.011531435884535313, + 0.008280064910650253, + 0.0066273086704313755, + 0.00788861233741045, + 0.003302444703876972, + 0.017097381874918938, + 0.018617184832692146, + 0.05499314144253731, + 0.00874156691133976, + 0.008843169547617435, + 0.01783944107592106, + 0.010089272633194923, + 0.005155620165169239, + 0.024467790499329567, + 0.015629559755325317, + 0.027026083320379257, + 0.01127912662923336, + 0.0112448213621974, + 0.016123991459608078, + 0.01096944697201252, + 0.0074557410553097725, + 0.005917766597121954, + 0.006964144762605429, + 0.036946799606084824, + 0.025720015168190002, + 0.0069451965391635895, + 0.010152159258723259, + 0.006996867246925831, + 0.004742166958749294, + 0.005266751162707806, + 0.007669416721910238, + 0.00744160870090127, + 0.01026392262428999, + 0.012185259722173214, + 0.010277402587234974, + 0.008340016007423401, + 0.0057073840871453285, + 0.008086113259196281, + 0.007952108047902584, + 0.02102230302989483, + 0.014877340756356716, + 0.04957437142729759, + 0.0193418450653553, + 0.007372918538749218, + 0.008517938666045666, + 0.008217363618314266, + 0.02462095208466053, + 0.014039430767297745, + 0.011378247290849686, + 0.018090471625328064, + 0.02790665440261364, + 0.07701011747121811, + 0.0062627592124044895, + 0.012033781968057156, + 0.0020628359634429216, + 0.012853818945586681, + 0.0028273966163396835, + 0.011652117595076561, + 0.0026107539888471365, + 0.002888311166316271, + 0.0018093634862452745, + 0.0017785305390134454, + 0.003948865924030542, + 0.0030725549440830946, + 0.008347836323082447, + 0.005242019891738892, + 0.0036867938470095396, + 0.004958681762218475, + 0.002161248354241252, + 0.0028416351415216923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.03372138366103172 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159868896007538 + ], + "opposite_logprob_tokens": [ + -1.7015986442565918 + ], + "logprob_reward": -0.20159868896007538, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159868896007538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14955422282218933, + 0.10082730650901794, + 0.08987126499414444, + 0.04359298199415207, + 0.04066247120499611, + 0.1769551783800125, + 0.14290067553520203, + 0.10273661464452744, + 0.04798572137951851, + 0.10491356998682022 + ], + "credit_map": [ + 0.036925774067640305, + 0.015873931348323822, + 0.008876344189047813, + 0.005794961005449295, + 0.017051730304956436, + 0.03267887979745865, + 0.00541250966489315, + 0.02694009058177471, + 0.017655907198786736, + 0.013637091964483261, + 0.00921482965350151, + 0.00580830592662096, + 0.009783638641238213, + 0.030375992879271507, + 0.003253456437960267, + 0.011098089627921581, + 0.01107395812869072, + 0.0274858046323061, + 0.005674772895872593, + 0.006754464469850063, + 0.009677719324827194, + 0.011155445128679276, + 0.004319597966969013, + 0.013729498721659184, + 0.00741297472268343, + 0.00558061758056283, + 0.005984335672110319, + 0.004915382247418165, + 0.0054562026634812355, + 0.005171018652617931, + 0.0037051811814308167, + 0.005367267411202192, + 0.005843929015100002, + 0.006531768012791872, + 0.010433787479996681, + 0.0027275942265987396, + 0.005125746596604586, + 0.0058656781911849976, + 0.0018711478915065527, + 0.00226282118819654, + 0.016240336000919342, + 0.0166027694940567, + 0.023714227601885796, + 0.02049887552857399, + 0.028868025168776512, + 0.02900530956685543, + 0.014520774595439434, + 0.027504848316311836, + 0.016202494502067566, + 0.01703251153230667, + 0.014550377614796162, + 0.04624782130122185, + 0.018073681741952896, + 0.01376320794224739, + 0.009185614995658398, + 0.007844961248338223, + 0.015644501894712448, + 0.009974328801035881, + 0.006329420953989029, + 0.010879422537982464, + 0.009572048671543598, + 0.03935627266764641, + 0.004314772319048643, + 0.006665853783488274, + 0.004969420842826366, + 0.005519424565136433, + 0.006504049524664879, + 0.0069613405503332615, + 0.006203147117048502, + 0.0056289974600076675, + 0.006600033026188612, + 0.005599308293312788, + 0.00948259886354208, + 0.028361760079860687, + 0.02322227507829666, + 0.014968782663345337, + 0.008135979063808918, + 0.011306283995509148, + 0.004379275720566511, + 0.005056618247181177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.20159868896007538 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644895628094673 + ], + "opposite_logprob_tokens": [ + -4.758645057678223 + ], + "logprob_reward": -0.008644895628094673, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644895628094673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2225535362958908, + 0.04412419721484184, + 0.08022356033325195, + 0.03760484978556633, + 0.026367180049419403, + 0.012480384670197964, + 0.1905439794063568, + 0.09553248435258865, + 0.038234639912843704, + 0.25233519077301025 + ], + "credit_map": [ + 0.057506103068590164, + 0.02616928331553936, + 0.015146536752581596, + 0.009044064208865166, + 0.029652874916791916, + 0.051356177777051926, + 0.00952859129756689, + 0.02414989098906517, + 0.004956542048603296, + 0.00891637708991766, + 0.0040762899443507195, + 0.004153559450060129, + 0.004333693068474531, + 0.01071318332105875, + 0.002895491197705269, + 0.004079061560332775, + 0.010659088380634785, + 0.023449545726180077, + 0.006094528827816248, + 0.006903172004967928, + 0.00842174980789423, + 0.009110265411436558, + 0.004804533440619707, + 0.010780679062008858, + 0.004483343567699194, + 0.0051141157746315, + 0.004126262851059437, + 0.005706212017685175, + 0.005416261497884989, + 0.005599572788923979, + 0.0034648084547370672, + 0.003694274928420782, + 0.007647279184311628, + 0.004703545942902565, + 0.002249552169814706, + 0.0017738984897732735, + 0.0026691569946706295, + 0.003807625500485301, + 0.001483740285038948, + 0.0020323789212852716, + 0.0014418680220842361, + 0.0027099556755274534, + 0.0016549108549952507, + 0.0016340796137228608, + 0.001804789644666016, + 0.001274466747418046, + 0.0010881474008783698, + 0.0008721665944904089, + 0.01118611078709364, + 0.03144131600856781, + 0.01947184093296528, + 0.06433344632387161, + 0.02490427903831005, + 0.014832384884357452, + 0.012643031775951385, + 0.011731571517884731, + 0.010227451100945473, + 0.009372849017381668, + 0.007582611870020628, + 0.01915762573480606, + 0.011871623806655407, + 0.02521475963294506, + 0.004621465224772692, + 0.007484093774110079, + 0.002576869446784258, + 0.007540663704276085, + 0.0048706186935305595, + 0.01010866742581129, + 0.004404779989272356, + 0.003568812506273389, + 0.0025215360801666975, + 0.0026426922995597124, + 0.02946051023900509, + 0.010717220604419708, + 0.07975631207227707, + 0.027979493141174316, + 0.023175235837697983, + 0.025155529379844666, + 0.011383390985429287, + 0.0447075217962265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.008644895628094673 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-length shot triggers a slow, sweeping pan around a majestic African bush elephant, instantly establishing its immense size. The warm, golden light of the setting sun meticulously sculpts the creature's form, capturing the deep, intricate network of wrinkles in its weathered, gray-brown hide and illuminating the fine lines in its gently flapping, fan-shaped ears. Its long, curved tusks gleam softly in the dim light. The elephant advances with a steady, deliberate gait, its powerful footsteps firmly pounding the dry earth and stirring up tiny, shimmering clouds of dust suspended in the still air. As the camera continues its circular motion, the scope of its epic trek is revealed: endless expanses of sun-scorched golden grass stretch to the horizon, dotted with the stark silhouettes of scattered acacia trees. The entire scene is imbued with an atmosphere of profound tranquility and majesty, capturing a timeless moment of a solitary pilgrimage through the vast, untamed wilderness.", + "seed": 28455, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_0/index_608_seed_28455_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2235686868429184, + "vlm_logprob": -0.2235686868429184, + "vlm_yes_no_margin": 1.925, + "vlm_token_credit": -0.2235686868429184, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17367121577262878 + ], + "opposite_logprob_tokens": [ + -2.048671245574951 + ], + "logprob_reward": -0.17367121577262878, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.17367121577262878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05694340541958809, + 0.7141480445861816, + 0.016963157802820206, + 0.027804097160696983, + 0.017788071185350418, + 0.009821792133152485, + 0.07114134728908539, + 0.012462728656828403, + 0.039058178663253784, + 0.03386915847659111 + ], + "credit_map": [ + 0.009243561886250973, + 0.007751813158392906, + 0.006175687070935965, + 0.009427186101675034, + 0.005298416130244732, + 0.004641328006982803, + 0.004829205572605133, + 0.009576206095516682, + 0.04789469763636589, + 0.0870993509888649, + 0.03481109067797661, + 0.022801393643021584, + 0.21483974158763885, + 0.036312002688646317, + 0.04544226452708244, + 0.22494752705097198, + 0.003385738003998995, + 0.0033693891018629074, + 0.0009780380642041564, + 0.0011051162146031857, + 0.0024177508894354105, + 0.002492139581590891, + 0.0013042391510680318, + 0.0019107464468106627, + 0.003108217613771558, + 0.001712451339699328, + 0.0016764226602390409, + 0.001844858517870307, + 0.0019387707579880953, + 0.001590532949194312, + 0.00965604092925787, + 0.00627680029720068, + 0.0024452225770801306, + 0.0027338704094290733, + 0.0027578985318541527, + 0.0020159692503511906, + 0.002515452215448022, + 0.0023159505799412727, + 0.0014779771445319057, + 0.001525730942375958, + 0.0015939793083816767, + 0.0015408117324113846, + 0.0018185466760769486, + 0.0008503713761456311, + 0.0011062711710110307, + 0.0009818850085139275, + 0.0010119557846337557, + 0.0009179702028632164, + 0.005929325241595507, + 0.020924268290400505, + 0.017693595960736275, + 0.005123354960232973, + 0.007559203542768955, + 0.003966080490499735, + 0.0043769702315330505, + 0.0055685448460280895, + 0.0017596391262486577, + 0.0020429512951523066, + 0.0013420821633189917, + 0.0012472603702917695, + 0.0015983765479177237, + 0.0014982247957959771, + 0.0012421195860952139, + 0.0017320739571005106, + 0.005418776068836451, + 0.006369641050696373, + 0.003745968919247389, + 0.004651119001209736, + 0.004996199626475573, + 0.00476125767454505, + 0.00414357939735055, + 0.004971639718860388, + 0.0077699632383883, + 0.0053887153044342995, + 0.0045352475717663765, + 0.003429740434512496, + 0.004351947922259569, + 0.0030326847918331623, + 0.0025158978532999754, + 0.0028449594974517822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.17367121577262878 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234436988830566 + ], + "opposite_logprob_tokens": [ + -0.8984436988830566 + ], + "logprob_reward": -0.5234436988830566, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234436988830566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.102028027176857, + 0.08861961960792542, + 0.1047019511461258, + 0.16410700976848602, + 0.1163216233253479, + 0.1352887600660324, + 0.13900631666183472, + 0.03498675301671028, + 0.04659143462777138, + 0.06834843754768372 + ], + "credit_map": [ + 0.016293706372380257, + 0.01649465039372444, + 0.01126044150441885, + 0.018995851278305054, + 0.011904058046638966, + 0.008853706531226635, + 0.007213474716991186, + 0.01101214624941349, + 0.013282204046845436, + 0.012140452861785889, + 0.009871650487184525, + 0.00539153628051281, + 0.012844034470617771, + 0.019174065440893173, + 0.00580570986494422, + 0.010109968483448029, + 0.02134711481630802, + 0.01473966147750616, + 0.006369933485984802, + 0.0072623989544808865, + 0.011680681258440018, + 0.019882045686244965, + 0.010161487385630608, + 0.013258635066449642, + 0.022402724251151085, + 0.014377736486494541, + 0.008511073887348175, + 0.016115127131342888, + 0.008865008130669594, + 0.011160530149936676, + 0.05280337110161781, + 0.029871441423892975, + 0.013522130437195301, + 0.01480186078697443, + 0.0179478507488966, + 0.01465083658695221, + 0.011026999913156033, + 0.022061171010136604, + 0.01079489104449749, + 0.011515885591506958, + 0.015712842345237732, + 0.023462381213903427, + 0.022245269268751144, + 0.012698420323431492, + 0.01484355516731739, + 0.017455972731113434, + 0.0108915064483881, + 0.017978809773921967, + 0.011566958390176296, + 0.03320576623082161, + 0.03543412312865257, + 0.012372915633022785, + 0.011682520620524883, + 0.008843456394970417, + 0.010506377555429935, + 0.015394200570881367, + 0.004919915460050106, + 0.007222112268209457, + 0.005266196560114622, + 0.0027778020594269037, + 0.004880101885646582, + 0.0026073188055306673, + 0.002992903580889106, + 0.004320403560996056, + 0.006841118913143873, + 0.008936560712754726, + 0.004538950510323048, + 0.0042612748220562935, + 0.007256112992763519, + 0.006740213371813297, + 0.0033571149688214064, + 0.004660089034587145, + 0.008896516636013985, + 0.009280364960432053, + 0.008743420243263245, + 0.006981609854847193, + 0.010256495326757431, + 0.0089218495413661, + 0.006878652144223452, + 0.008389531634747982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.5234436988830566 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905135095119476 + ], + "opposite_logprob_tokens": [ + -2.4640512466430664 + ], + "logprob_reward": -0.08905135095119476, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905135095119476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06742823868989944, + 0.37359338998794556, + 0.040720511227846146, + 0.21642516553401947, + 0.02530379220843315, + 0.04232994467020035, + 0.0962105393409729, + 0.034542374312877655, + 0.03648983687162399, + 0.06695614010095596 + ], + "credit_map": [ + 0.008550150319933891, + 0.008215297013521194, + 0.006712743081152439, + 0.010892114602029324, + 0.006496865767985582, + 0.007582843769341707, + 0.005096799228340387, + 0.013881433755159378, + 0.024908510968089104, + 0.0459042452275753, + 0.018451368436217308, + 0.011778758838772774, + 0.11326855421066284, + 0.019365886226296425, + 0.023360926657915115, + 0.11655519157648087, + 0.007000305689871311, + 0.005326723214238882, + 0.0023421591613441706, + 0.0022331220097839832, + 0.010453765280544758, + 0.00377084338106215, + 0.003762981155887246, + 0.005830618552863598, + 0.017407117411494255, + 0.02160855196416378, + 0.012057899497449398, + 0.017018355429172516, + 0.01792026124894619, + 0.009744509123265743, + 0.07989618927240372, + 0.04077230766415596, + 0.0032854946330189705, + 0.003360212314873934, + 0.004674050956964493, + 0.0027978564612567425, + 0.00292520341463387, + 0.0032319144811481237, + 0.0022070342674851418, + 0.002822030568495393, + 0.004924768581986427, + 0.00879964791238308, + 0.005701216869056225, + 0.0038502411916851997, + 0.005064935889095068, + 0.006386140827089548, + 0.004543450195342302, + 0.0030595497228205204, + 0.007206114474684, + 0.027811141684651375, + 0.023091565817594528, + 0.008512635715305805, + 0.010173875838518143, + 0.004499440081417561, + 0.0072280168533325195, + 0.007687754463404417, + 0.0046143499203026295, + 0.00570728350430727, + 0.004554565064609051, + 0.0027670825365930796, + 0.003812802257016301, + 0.003455802798271179, + 0.004154813475906849, + 0.005475680343806744, + 0.005171287339180708, + 0.00632445327937603, + 0.0039120386354625225, + 0.002977875992655754, + 0.008282971568405628, + 0.00451640272513032, + 0.001985943177714944, + 0.003318872069939971, + 0.012292379513382912, + 0.00920811016112566, + 0.00900886207818985, + 0.0092775858938694, + 0.009773851372301579, + 0.006224330514669418, + 0.005807460751384497, + 0.005363570526242256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.08905135095119476 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134806156158447 + ], + "opposite_logprob_tokens": [ + -1.3134806156158447 + ], + "logprob_reward": -0.3134806156158447, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134806156158447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06751684099435806, + 0.10938122868537903, + 0.0264352448284626, + 0.19877907633781433, + 0.1510205715894699, + 0.049278564751148224, + 0.2731474041938782, + 0.047265637665987015, + 0.02778473310172558, + 0.049390681087970734 + ], + "credit_map": [ + 0.010967492125928402, + 0.026381071656942368, + 0.003866812912747264, + 0.007262353785336018, + 0.005786499008536339, + 0.004210575018078089, + 0.002667664084583521, + 0.006374372635036707, + 0.00786383543163538, + 0.012808108702301979, + 0.006033948622643948, + 0.0035913651809096336, + 0.03271843120455742, + 0.0064534833654761314, + 0.006463784258812666, + 0.03344827890396118, + 0.002837913343682885, + 0.0035169266629964113, + 0.0018993925768882036, + 0.0018939698347821832, + 0.005460470449179411, + 0.002809759695082903, + 0.0025196862407028675, + 0.005497125443071127, + 0.01985662244260311, + 0.03261899575591087, + 0.009092677384614944, + 0.020717887207865715, + 0.017921866849064827, + 0.0078076692298054695, + 0.06346328556537628, + 0.02730005979537964, + 0.01697486825287342, + 0.017511719837784767, + 0.023197107017040253, + 0.01858823373913765, + 0.015338126569986343, + 0.02898688055574894, + 0.014598512090742588, + 0.01582513004541397, + 0.0064110835082829, + 0.0069161406718194485, + 0.00877673365175724, + 0.004873243160545826, + 0.0059984903782606125, + 0.005971830803900957, + 0.006648322567343712, + 0.003682720474898815, + 0.018676159903407097, + 0.0809304416179657, + 0.09042854607105255, + 0.01797887124121189, + 0.018082795664668083, + 0.010254777036607265, + 0.01129954494535923, + 0.02549627423286438, + 0.004806749057024717, + 0.007988964207470417, + 0.007684761192649603, + 0.0038117903750389814, + 0.007638331037014723, + 0.004767043516039848, + 0.0037418121937662363, + 0.006826186086982489, + 0.006064141169190407, + 0.0036960053257644176, + 0.0031985074747353792, + 0.0023608270566910505, + 0.003935868386179209, + 0.0031023076735436916, + 0.0019817701540887356, + 0.003445305163040757, + 0.010379603132605553, + 0.0072349486872553825, + 0.006430908106267452, + 0.006410570815205574, + 0.0057251956313848495, + 0.004180798307061195, + 0.005374193657189608, + 0.0036544594913721085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.3134806156158447 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018196552991867065 + ], + "opposite_logprob_tokens": [ + -4.0181965827941895 + ], + "logprob_reward": -0.018196552991867065, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018196552991867065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10167399048805237, + 0.37995120882987976, + 0.013196306303143501, + 0.04675807058811188, + 0.02978178858757019, + 0.019058285281062126, + 0.07166732102632523, + 0.029033806174993515, + 0.09001300483942032, + 0.21886616945266724 + ], + "credit_map": [ + 0.019259808585047722, + 0.015080807730555534, + 0.008077678270637989, + 0.011193548329174519, + 0.005762461107224226, + 0.005775063764303923, + 0.004586949944496155, + 0.031937677413225174, + 0.025160379707813263, + 0.04498685896396637, + 0.018718533217906952, + 0.012141351588070393, + 0.11332786083221436, + 0.01973632723093033, + 0.024755273014307022, + 0.12112461775541306, + 0.002229887992143631, + 0.001870789099484682, + 0.0008571104845032096, + 0.001193682779558003, + 0.0018849070183932781, + 0.0020515958312898874, + 0.0016473089344799519, + 0.0014610250946134329, + 0.007775945123285055, + 0.01780678518116474, + 0.00229094410315156, + 0.006519245449453592, + 0.004520230460911989, + 0.0020271220710128546, + 0.0027994222473353148, + 0.0030183778144419193, + 0.003947642631828785, + 0.004194553941488266, + 0.003912817221134901, + 0.003656512824818492, + 0.003659273963421583, + 0.005073506850749254, + 0.0027604957576841116, + 0.002576985629275441, + 0.0021735334303230047, + 0.004199545830488205, + 0.0023284652270376682, + 0.0024655931629240513, + 0.002593283075839281, + 0.0016205242136493325, + 0.0021739790681749582, + 0.0015033610397949815, + 0.004545753821730614, + 0.02095293626189232, + 0.021710680797696114, + 0.004800795111805201, + 0.005963883828371763, + 0.0030005862936377525, + 0.00362185831181705, + 0.007070824038237333, + 0.004425174091011286, + 0.005620824638754129, + 0.00511133624240756, + 0.0021242541261017323, + 0.003188265021890402, + 0.0021309684962034225, + 0.002350859809666872, + 0.004082123748958111, + 0.013683774508535862, + 0.011919539421796799, + 0.011718221940100193, + 0.009211587719619274, + 0.00957524124532938, + 0.009403547272086143, + 0.009136154316365719, + 0.015364938415586948, + 0.0576401948928833, + 0.0401698462665081, + 0.027140481397509575, + 0.023725327104330063, + 0.02288729138672352, + 0.017202967777848244, + 0.012208807282149792, + 0.017891259863972664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.018196552991867065 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-length shot triggers a slow, sweeping pan around a majestic African bush elephant, instantly establishing its immense size. The warm, golden light of the setting sun meticulously sculpts the creature's form, capturing the deep, intricate network of wrinkles in its weathered, gray-brown hide and illuminating the fine lines in its gently flapping, fan-shaped ears. Its long, curved tusks gleam softly in the dim light. The elephant advances with a steady, deliberate gait, its powerful footsteps firmly pounding the dry earth and stirring up tiny, shimmering clouds of dust suspended in the still air. As the camera continues its circular motion, the scope of its epic trek is revealed: endless expanses of sun-scorched golden grass stretch to the horizon, dotted with the stark silhouettes of scattered acacia trees. The entire scene is imbued with an atmosphere of profound tranquility and majesty, capturing a timeless moment of a solitary pilgrimage through the vast, untamed wilderness.", + "seed": 28456, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_0/index_608_seed_28456_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06798460390418767, + "vlm_logprob": -0.06798460390418767, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.06798460390418767, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06726578623056412 + ], + "opposite_logprob_tokens": [ + -2.942265748977661 + ], + "logprob_reward": -0.06726578623056412, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06726578623056412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0955246314406395, + 0.15795792639255524, + 0.1958441585302353, + 0.10501208156347275, + 0.056341223418712616, + 0.08069239556789398, + 0.08724454790353775, + 0.13464829325675964, + 0.05006880685687065, + 0.036665935069322586 + ], + "credit_map": [ + 0.01386070903390646, + 0.011802584864199162, + 0.012638019397854805, + 0.010374326258897781, + 0.008373417891561985, + 0.017514530569314957, + 0.007278180215507746, + 0.013682856224477291, + 0.02132471278309822, + 0.01754688285291195, + 0.01300322636961937, + 0.019250335171818733, + 0.015099016018211842, + 0.040435731410980225, + 0.01214427687227726, + 0.01915375143289566, + 0.04575046896934509, + 0.020099330693483353, + 0.019644370302557945, + 0.015436810441315174, + 0.01930011622607708, + 0.04050374776124954, + 0.01083092950284481, + 0.024278393015265465, + 0.014919704757630825, + 0.01728024333715439, + 0.015907298773527145, + 0.00901045836508274, + 0.010833472944796085, + 0.018908336758613586, + 0.005628114566206932, + 0.012524452060461044, + 0.006599354092031717, + 0.007264281157404184, + 0.010136724449694157, + 0.003965985029935837, + 0.005684733856469393, + 0.012774099595844746, + 0.002573359990492463, + 0.0073426831513643265, + 0.013179011642932892, + 0.009878366254270077, + 0.01163067389279604, + 0.011611498892307281, + 0.008045737631618977, + 0.01074489951133728, + 0.00552151957526803, + 0.01008069422096014, + 0.014381466433405876, + 0.010543969459831715, + 0.01345611922442913, + 0.007156405597925186, + 0.012281225994229317, + 0.012669505551457405, + 0.005194562021642923, + 0.011561297811567783, + 0.021808069199323654, + 0.016314009204506874, + 0.021610606461763382, + 0.011282915249466896, + 0.010107878595590591, + 0.02669217437505722, + 0.007166788447648287, + 0.019665848463773727, + 0.011846471577882767, + 0.0046954345889389515, + 0.007054957095533609, + 0.0042455666698515415, + 0.0039930278435349464, + 0.0078429551795125, + 0.002721411408856511, + 0.007668984122574329, + 0.012480581179261208, + 0.004619261249899864, + 0.0068669975735247135, + 0.0028085908852517605, + 0.002940269187092781, + 0.00245620496571064, + 0.0012168856337666512, + 0.0032771462574601173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.06726578623056412 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17993496358394623 + ], + "opposite_logprob_tokens": [ + -1.8049349784851074 + ], + "logprob_reward": -0.17993496358394623, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17993496358394623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0945281833410263, + 0.18816514313220978, + 0.05286240950226784, + 0.20787391066551208, + 0.028427086770534515, + 0.04877132922410965, + 0.10038066655397415, + 0.18171969056129456, + 0.05670732632279396, + 0.04056432843208313 + ], + "credit_map": [ + 0.013347640633583069, + 0.009600781835615635, + 0.015960300341248512, + 0.010120073333382607, + 0.006762031931430101, + 0.016644736751914024, + 0.006861458998173475, + 0.015231146477162838, + 0.029876042157411575, + 0.01904365047812462, + 0.012490255758166313, + 0.020159492269158363, + 0.019958900287747383, + 0.05102809518575668, + 0.012536057271063328, + 0.023072633892297745, + 0.010873569175601006, + 0.006196819245815277, + 0.004026393406093121, + 0.0030664585065096617, + 0.0028410705272108316, + 0.015549116767942905, + 0.0024699606001377106, + 0.007839017547667027, + 0.035974256694316864, + 0.03478071838617325, + 0.02786070667207241, + 0.01647171936929226, + 0.018932443112134933, + 0.03877321258187294, + 0.011150623671710491, + 0.02393021248281002, + 0.0031523907091468573, + 0.0045417118817567825, + 0.00401085801422596, + 0.0027932031080126762, + 0.003374056424945593, + 0.005821311380714178, + 0.0015542524633929133, + 0.0031793012749403715, + 0.007238638121634722, + 0.00504195224493742, + 0.007626898121088743, + 0.005479229148477316, + 0.008147086016833782, + 0.0058347005397081375, + 0.0035265623591840267, + 0.005876258946955204, + 0.02810600772500038, + 0.010846569202840328, + 0.014848796650767326, + 0.007509272545576096, + 0.007991952821612358, + 0.01343501266092062, + 0.0042360397055745125, + 0.013407002203166485, + 0.03005138412117958, + 0.021857479587197304, + 0.028725411742925644, + 0.012686122208833694, + 0.01356825977563858, + 0.03845100477337837, + 0.009381948038935661, + 0.026998063549399376, + 0.010281339287757874, + 0.004558599554002285, + 0.00843418575823307, + 0.004519527778029442, + 0.004915178753435612, + 0.009269678965210915, + 0.0034384489990770817, + 0.01129036396741867, + 0.010914899408817291, + 0.005915675777941942, + 0.007085318677127361, + 0.0029057199135422707, + 0.003921965602785349, + 0.003057196270674467, + 0.0017971611814573407, + 0.004966387525200844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.17993496358394623 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810567781329155 + ], + "opposite_logprob_tokens": [ + -3.2881057262420654 + ], + "logprob_reward": -0.03810567781329155, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810567781329155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04591372609138489, + 0.03801034763455391, + 0.05772421881556511, + 0.2420150488615036, + 0.049156732857227325, + 0.0718238428235054, + 0.03707269951701164, + 0.37547218799591064, + 0.03655114769935608, + 0.046260010451078415 + ], + "credit_map": [ + 0.008119544945657253, + 0.007010661531239748, + 0.0055473241955041885, + 0.005977250635623932, + 0.003515163203701377, + 0.007081002928316593, + 0.0031599905341863632, + 0.005502789281308651, + 0.006035728845745325, + 0.004396659322082996, + 0.0036543693859130144, + 0.003332416992634535, + 0.005146413575857878, + 0.00764293409883976, + 0.003177741775289178, + 0.004624084569513798, + 0.017721649259328842, + 0.004949130583554506, + 0.00615728972479701, + 0.003957672975957394, + 0.003492474090307951, + 0.010358673520386219, + 0.0029148724861443043, + 0.008172459900379181, + 0.03749622404575348, + 0.04157004877924919, + 0.03179742023348808, + 0.02051910012960434, + 0.023383572697639465, + 0.046557359397411346, + 0.013334548100829124, + 0.02735677920281887, + 0.005173998419195414, + 0.01003377977758646, + 0.007322207093238831, + 0.004939632955938578, + 0.006498103961348534, + 0.007512928452342749, + 0.0022276416420936584, + 0.005448439624160528, + 0.015107864513993263, + 0.006839819252490997, + 0.01011685747653246, + 0.010929824784398079, + 0.008589544333517551, + 0.007775875274091959, + 0.004417716059833765, + 0.008046334609389305, + 0.004928575363010168, + 0.003956831060349941, + 0.006944756023585796, + 0.003077487228438258, + 0.005018180701881647, + 0.004887389484792948, + 0.0026824467349797487, + 0.00557703198865056, + 0.05616398900747299, + 0.045974891632795334, + 0.06592080742120743, + 0.02523786388337612, + 0.028001585975289345, + 0.06908536702394485, + 0.017949765548110008, + 0.06713791191577911, + 0.01015583798289299, + 0.0034000580199062824, + 0.005154708866029978, + 0.0022090754937380552, + 0.0029277910944074392, + 0.005299408920109272, + 0.0021868713665753603, + 0.005217393394559622, + 0.013937515206634998, + 0.005166208371520042, + 0.008378352969884872, + 0.0034690331667661667, + 0.002950347028672695, + 0.0028289512265473604, + 0.00210403255186975, + 0.007425573654472828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 5 + ], + "selected_reward": -0.03810567781329155 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866647347807884 + ], + "opposite_logprob_tokens": [ + -3.048666477203369 + ], + "logprob_reward": -0.04866647347807884, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866647347807884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08387291431427002, + 0.07247934490442276, + 0.07111969590187073, + 0.15566100180149078, + 0.05668896809220314, + 0.0934676006436348, + 0.064027801156044, + 0.15512171387672424, + 0.22041019797325134, + 0.027150733396410942 + ], + "credit_map": [ + 0.011697127483785152, + 0.011237812228500843, + 0.01147165335714817, + 0.010384304448962212, + 0.009258815087378025, + 0.01240105926990509, + 0.007585784420371056, + 0.009836366400122643, + 0.011016255244612694, + 0.007864813320338726, + 0.009382938966155052, + 0.006650060415267944, + 0.005009821150451899, + 0.01806533709168434, + 0.005406241398304701, + 0.009083887562155724, + 0.012800058349967003, + 0.008757311850786209, + 0.005264136474579573, + 0.005675964523106813, + 0.008158868178725243, + 0.016451440751552582, + 0.004820141009986401, + 0.009191777557134628, + 0.024787141010165215, + 0.02137772925198078, + 0.02795853465795517, + 0.010497422888875008, + 0.01315103005617857, + 0.031572941690683365, + 0.008117971941828728, + 0.018198244273662567, + 0.006476521026343107, + 0.011486418545246124, + 0.00728574488312006, + 0.004231705795973539, + 0.006588628049939871, + 0.011599976569414139, + 0.0022242481354624033, + 0.006795727647840977, + 0.011705170385539532, + 0.011897160671651363, + 0.010774844326078892, + 0.008879490196704865, + 0.014663612470030785, + 0.017416713759303093, + 0.005025597754865885, + 0.013105015270411968, + 0.011851021088659763, + 0.006236101966351271, + 0.008763925172388554, + 0.005267156288027763, + 0.00831068679690361, + 0.010993474163115025, + 0.004764285404235125, + 0.007841154932975769, + 0.01945587806403637, + 0.01983603648841381, + 0.03182872757315636, + 0.010574457235634327, + 0.011952687986195087, + 0.022255102172493935, + 0.007044376339763403, + 0.03217445686459541, + 0.04457927122712135, + 0.017606033012270927, + 0.0328776054084301, + 0.017245037481188774, + 0.016625069081783295, + 0.036684032529592514, + 0.012354043312370777, + 0.04243911802768707, + 0.006568001117557287, + 0.004213406704366207, + 0.004229144658893347, + 0.00375757017172873, + 0.0026953169144690037, + 0.0018088998040184379, + 0.0010426911758258939, + 0.0028357054106891155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.04866647347807884 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an African bush elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, sweeping circular pan around the elephant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the elephant walking with a steady, deliberate gait that stirs up dust clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of sun-scorched golden grass and scattered acacia trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950118415057659 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.005950118415057659, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950118415057659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045231159776449203, + 0.05562460049986839, + 0.1357690393924713, + 0.10637643188238144, + 0.07262694090604782, + 0.04254184290766716, + 0.025064265355467796, + 0.3634323477745056, + 0.12409785389900208, + 0.029235510155558586 + ], + "credit_map": [ + 0.00884067639708519, + 0.005212442949414253, + 0.005416302941739559, + 0.005184619687497616, + 0.0039779129438102245, + 0.00693883141502738, + 0.0033838236704468727, + 0.00627655116841197, + 0.006078224163502455, + 0.0039475178346037865, + 0.003572160145267844, + 0.00514698401093483, + 0.005255376920104027, + 0.020061612129211426, + 0.003365893382579088, + 0.008196836337447166, + 0.02272840403020382, + 0.015118005685508251, + 0.010909976437687874, + 0.008329656906425953, + 0.013099638745188713, + 0.03760996833443642, + 0.008412474766373634, + 0.019560925662517548, + 0.026614781469106674, + 0.012348979711532593, + 0.009961631149053574, + 0.006088626105338335, + 0.009603282436728477, + 0.02347143553197384, + 0.0053690713830292225, + 0.012918629683554173, + 0.008937478065490723, + 0.007974968291819096, + 0.0132621293887496, + 0.0037834050599485636, + 0.006875996943563223, + 0.018989507108926773, + 0.0027778211515396833, + 0.01002564188092947, + 0.009590710513293743, + 0.005297303199768066, + 0.006888709496706724, + 0.0030998182483017445, + 0.0043658423237502575, + 0.005271102301776409, + 0.001970018958672881, + 0.006058339029550552, + 0.005145060829818249, + 0.003286112565547228, + 0.00418838718906045, + 0.0017258018488064408, + 0.002329605631530285, + 0.003522513434290886, + 0.0013088814448565245, + 0.003557903226464987, + 0.049684036523103714, + 0.046232763677835464, + 0.06824541091918945, + 0.024540353566408157, + 0.02643943391740322, + 0.060920171439647675, + 0.01771850697696209, + 0.06965168565511703, + 0.022269990295171738, + 0.009797394275665283, + 0.018937066197395325, + 0.009218576364219189, + 0.00964577030390501, + 0.022007131949067116, + 0.007285095285624266, + 0.024936839938163757, + 0.006277458276599646, + 0.003467450151219964, + 0.004668021108955145, + 0.003040457610040903, + 0.0026821389328688383, + 0.002332918345928192, + 0.0014315858716145158, + 0.005335482303053141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.005950118415057659 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_1/index_tensor([4424, 4424])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_1/index_tensor([4424, 4424])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..79f98c297bdc17dc78a74af0a80a24225a99a460 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_1/index_tensor([4424, 4424])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "majestic, sweeping, high-angle drone shot slowly glides forward and slightly downward, revealing a grand Gothic ballroom rendered in stunning, surreal 3D. A soaring vaulted ceiling with dark stone ribs reaches into the shadows above, flanked by towering stained-glass arched windows that depict a somber and romantic scene. The entire opulent space is illuminated by the warm, shimmering glow of a massive crystal chandelier, whose thousands of facets cast dancing golden rays across the highly polished black marble floor, creating near-perfect, shimmering reflections. Below, hundreds of guests gathered in intimate groups, dressed in exquisite gowns of deep velvet, shimmering silks, and intricate lace in rich jewel tones like crimson, emerald, and sapphire, alongside men in sharp, dark tuxedos. Each guest's identity is concealed behind an intricate phantom mask, ranging from simple, elegant dominoes to ornate creations composed of wire and dark feathers. As the camera continues its steady, almost imperceptible descent, the quiet nature of their conversation becomes apparent, not through sound but through their subtle, deliberate gestures: a fan tapping gently against a gloved palm, a masked head tilting slightly, conspiratorially, a hand gesturing elegantly to emphasize a silent point. The atmosphere is rich with luxurious mystery and exquisite elegance, a breathtaking picture of secret grandeur, where warm light catches the glint of jewels or the metallic sheen of masks, casting long, soft shadows that twist and dance with the flickering candlelight.", + "seed": 17765, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_1/index_4424_seed_17765_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03751967614516616, + "vlm_logprob": -0.03751967614516616, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.03751967614516616, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07245276868343353 + ], + "opposite_logprob_tokens": [ + -2.6974527835845947 + ], + "logprob_reward": -0.07245276868343353, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07245276868343353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1800716370344162, + 0.18369042873382568, + 0.16979756951332092, + 0.08702586591243744, + 0.10008033365011215, + 0.03377334028482437, + 0.04326901584863663, + 0.027103275060653687, + 0.08635415136814117, + 0.08883440494537354 + ], + "credit_map": [ + 0.02352069690823555, + 0.02721797302365303, + 0.013050640001893044, + 0.01728353463113308, + 0.040632907301187515, + 0.018786532804369926, + 0.023162133991718292, + 0.016417216509580612, + 0.0179557204246521, + 0.03623832389712334, + 0.01800405979156494, + 0.03519584611058235, + 0.021797845140099525, + 0.014708098955452442, + 0.018221966922283173, + 0.02156856842339039, + 0.01370774395763874, + 0.022655511274933815, + 0.04298527166247368, + 0.027729181572794914, + 0.015388413332402706, + 0.010365627706050873, + 0.016909699887037277, + 0.020056137815117836, + 0.008959705010056496, + 0.006710174027830362, + 0.016465960070490837, + 0.021044811233878136, + 0.007636055815964937, + 0.005784851498901844, + 0.011271842755377293, + 0.009152463637292385, + 0.006469789892435074, + 0.012136348523199558, + 0.01588340289890766, + 0.02020602487027645, + 0.009930942207574844, + 0.007167973089963198, + 0.014541033655405045, + 0.013744819909334183, + 0.00249427929520607, + 0.002754233544692397, + 0.005810449365526438, + 0.008027165196835995, + 0.0030500732827931643, + 0.0028274941723793745, + 0.0033641017507761717, + 0.005445541813969612, + 0.003951153717935085, + 0.005562979262322187, + 0.006569039076566696, + 0.009857499971985817, + 0.0029043173417448997, + 0.003315214067697525, + 0.005042983219027519, + 0.006065825931727886, + 0.0016970904543995857, + 0.002243245719000697, + 0.005288015119731426, + 0.006287301890552044, + 0.00264038797467947, + 0.002261262619867921, + 0.002920454367995262, + 0.003765515750274062, + 0.005558494012802839, + 0.007813776843249798, + 0.021816851571202278, + 0.015651201829314232, + 0.005721465218812227, + 0.006636485457420349, + 0.0068822274915874004, + 0.016273649409413338, + 0.006272902712225914, + 0.006967579945921898, + 0.022260066121816635, + 0.019177895039319992, + 0.00708748446777463, + 0.007197909988462925, + 0.007972253486514091, + 0.011898316442966461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.07245276868343353 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637173980474472 + ], + "opposite_logprob_tokens": [ + -3.651371717453003 + ], + "logprob_reward": -0.02637173980474472, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637173980474472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22808769345283508, + 0.14852502942085266, + 0.10469422489404678, + 0.23011957108974457, + 0.06033939868211746, + 0.029103528708219528, + 0.038867462426424026, + 0.05079278722405434, + 0.06568454951047897, + 0.04378582537174225 + ], + "credit_map": [ + 0.034508299082517624, + 0.051985003054142, + 0.013349096290767193, + 0.019529307261109352, + 0.05044309049844742, + 0.020017605274915695, + 0.018943743780255318, + 0.019311530515551567, + 0.013250606134533882, + 0.024622080847620964, + 0.013152657076716423, + 0.02118312194943428, + 0.01663162373006344, + 0.018676046282052994, + 0.01712738163769245, + 0.02388150431215763, + 0.01061959471553564, + 0.018262961879372597, + 0.01830037124454975, + 0.015089926309883595, + 0.012713105417788029, + 0.009494422934949398, + 0.011662706732749939, + 0.008551127277314663, + 0.023937111720442772, + 0.016283946111798286, + 0.060030512511730194, + 0.06297982484102249, + 0.02024216763675213, + 0.012784830294549465, + 0.014867792837321758, + 0.01899336278438568, + 0.003903607139363885, + 0.006263719405978918, + 0.009669953025877476, + 0.011453148908913136, + 0.006803956814110279, + 0.005234257783740759, + 0.008169801905751228, + 0.008840948343276978, + 0.001929900492541492, + 0.002022262429818511, + 0.005028433166444302, + 0.00773429311811924, + 0.0022298600524663925, + 0.002036856021732092, + 0.003640703158453107, + 0.004481217823922634, + 0.003409121185541153, + 0.004564299248158932, + 0.0063239699229598045, + 0.008870616555213928, + 0.0033949108328670263, + 0.0026377488393336535, + 0.004168089013546705, + 0.0054987031035125256, + 0.003312528133392334, + 0.004095821175724268, + 0.008142148144543171, + 0.012843048200011253, + 0.00422348128631711, + 0.0034088403917849064, + 0.005451848264783621, + 0.00931506883352995, + 0.0051460107788443565, + 0.00623576994985342, + 0.01477380096912384, + 0.012742731720209122, + 0.004009114112704992, + 0.00489380769431591, + 0.004861854948103428, + 0.013021453283727169, + 0.003615387249737978, + 0.0030088459607213736, + 0.008520476520061493, + 0.008532837964594364, + 0.004517355468124151, + 0.004108751658350229, + 0.005189937073737383, + 0.006292229052633047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.02637173980474472 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631670217961073 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631670217961073, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631670217961073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3812789022922516, + 0.11086251586675644, + 0.1884305775165558, + 0.04752162843942642, + 0.06590168178081512, + 0.020546210929751396, + 0.051420532166957855, + 0.05102981626987457, + 0.04071815684437752, + 0.04229000210762024 + ], + "credit_map": [ + 0.049184754490852356, + 0.10669108480215073, + 0.018715638667345047, + 0.02720014937222004, + 0.0918237641453743, + 0.02888549491763115, + 0.029887506738305092, + 0.028890497982501984, + 0.01170344091951847, + 0.015816450119018555, + 0.011021571233868599, + 0.015403538011014462, + 0.014900913462042809, + 0.01336603332310915, + 0.012770022265613079, + 0.015880541875958443, + 0.015582782216370106, + 0.027679981663823128, + 0.025620074942708015, + 0.042851436883211136, + 0.020240627229213715, + 0.010466125793755054, + 0.027447378262877464, + 0.01854216493666172, + 0.003266952931880951, + 0.005014149006456137, + 0.010092522017657757, + 0.010667073540389538, + 0.005480313673615456, + 0.0037518737372010946, + 0.004617876373231411, + 0.004630867391824722, + 0.004137655254453421, + 0.008762864395976067, + 0.009509183466434479, + 0.012452615424990654, + 0.0072794146835803986, + 0.005625945050269365, + 0.00914587453007698, + 0.008988129906356335, + 0.001980161527171731, + 0.0022495724260807037, + 0.0033087239135056734, + 0.0045516518875956535, + 0.0015279853250831366, + 0.001564400503411889, + 0.0025491199921816587, + 0.0028145951218903065, + 0.0038970964960753918, + 0.0057096402160823345, + 0.008133696392178535, + 0.0112454229965806, + 0.004056386649608612, + 0.004231294617056847, + 0.004965850152075291, + 0.009181147441267967, + 0.003329166676849127, + 0.00435932632535696, + 0.008625595830380917, + 0.013260315172374249, + 0.0032954923808574677, + 0.0036485197488218546, + 0.005437240470200777, + 0.00907416082918644, + 0.003065271070227027, + 0.00390529353171587, + 0.007448425050824881, + 0.010391137562692165, + 0.003095618449151516, + 0.0031809816136956215, + 0.004369865637272596, + 0.005261564627289772, + 0.003747162874788046, + 0.002872103825211525, + 0.006306496448814869, + 0.010008067823946476, + 0.004431314766407013, + 0.0036503428127616644, + 0.0056063453666865826, + 0.005668165627866983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.007631670217961073 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809695690870285 + ], + "opposite_logprob_tokens": [ + -3.2880969047546387 + ], + "logprob_reward": -0.03809695690870285, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809695690870285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10738649219274521, + 0.07453567534685135, + 0.0794723778963089, + 0.24935667216777802, + 0.12020552903413773, + 0.0289730504155159, + 0.05287010595202446, + 0.06790955364704132, + 0.09855111688375473, + 0.1207394003868103 + ], + "credit_map": [ + 0.013449549674987793, + 0.01918499357998371, + 0.007596137002110481, + 0.010894112288951874, + 0.021930065006017685, + 0.01039055921137333, + 0.01487303338944912, + 0.009068062528967857, + 0.00765584921464324, + 0.011081154458224773, + 0.006387318018823862, + 0.00793539173901081, + 0.011330856941640377, + 0.007941395044326782, + 0.01110796257853508, + 0.011095745489001274, + 0.006405627820640802, + 0.007130759768188, + 0.018000205978751183, + 0.014822889119386673, + 0.006485803984105587, + 0.004504660610109568, + 0.008689881302416325, + 0.013432556763291359, + 0.013624239712953568, + 0.009698967449367046, + 0.07974034547805786, + 0.09349080920219421, + 0.008561939932405949, + 0.007136859465390444, + 0.021618852391839027, + 0.015484665520489216, + 0.009939661249518394, + 0.011824870482087135, + 0.019489115104079247, + 0.028178326785564423, + 0.009908368811011314, + 0.008415141142904758, + 0.013751795515418053, + 0.018698250874876976, + 0.0018772928742691875, + 0.0025766915641725063, + 0.004635647404938936, + 0.006919732317328453, + 0.002195539651438594, + 0.002148489700630307, + 0.0036900416016578674, + 0.004929616115987301, + 0.003036161419004202, + 0.004617782309651375, + 0.0088827945291996, + 0.012767909094691277, + 0.006417502649128437, + 0.0038637493271380663, + 0.005809440277516842, + 0.0074747661128640175, + 0.005199469160288572, + 0.004714776761829853, + 0.010826264508068562, + 0.013383445329964161, + 0.006132065784186125, + 0.006387153174728155, + 0.006788871716707945, + 0.014477507211267948, + 0.007399882189929485, + 0.008807733654975891, + 0.0217234306037426, + 0.024469444528222084, + 0.006744828075170517, + 0.00734241958707571, + 0.009556855075061321, + 0.012506530620157719, + 0.008177584037184715, + 0.009614699520170689, + 0.024011746048927307, + 0.029834730550646782, + 0.009853136725723743, + 0.009885136969387531, + 0.012952856719493866, + 0.01640952192246914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.03809695690870285 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304524511098862 + ], + "opposite_logprob_tokens": [ + -3.1680452823638916 + ], + "logprob_reward": -0.04304524511098862, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304524511098862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11239693313837051, + 0.08524096012115479, + 0.08894997090101242, + 0.21309755742549896, + 0.0705527812242508, + 0.0625392273068428, + 0.17885100841522217, + 0.06364479660987854, + 0.07795518636703491, + 0.04677160829305649 + ], + "credit_map": [ + 0.0196163821965456, + 0.015360415913164616, + 0.007797762285917997, + 0.010487853549420834, + 0.01792062632739544, + 0.009123995900154114, + 0.022642800584435463, + 0.009447095915675163, + 0.00833098217844963, + 0.010392836295068264, + 0.006958717945963144, + 0.008944813162088394, + 0.0144457733258605, + 0.011428382247686386, + 0.009430631063878536, + 0.015308825299143791, + 0.006368129048496485, + 0.007967359386384487, + 0.021936997771263123, + 0.014048258773982525, + 0.00871287565678358, + 0.006467177998274565, + 0.008085164241492748, + 0.015364001505076885, + 0.012969187460839748, + 0.011237366124987602, + 0.06403195112943649, + 0.07871868461370468, + 0.010478390380740166, + 0.006779532413929701, + 0.014857667498290539, + 0.014024769887328148, + 0.007952953688800335, + 0.00745221646502614, + 0.01051413919776678, + 0.010477433912456036, + 0.010706973262131214, + 0.00792055856436491, + 0.006337943021208048, + 0.009190556593239307, + 0.004597894847393036, + 0.0055554453283548355, + 0.011660938151180744, + 0.013849550858139992, + 0.004122910555452108, + 0.005394936539232731, + 0.007732522673904896, + 0.009625029750168324, + 0.013499189168214798, + 0.014121033251285553, + 0.030935589224100113, + 0.039079535752534866, + 0.011029547080397606, + 0.01321349386125803, + 0.02285616658627987, + 0.034116458147764206, + 0.004245541524142027, + 0.005612543318420649, + 0.01154282595962286, + 0.01752065308392048, + 0.004294451326131821, + 0.004443876910954714, + 0.006374433171004057, + 0.00961047038435936, + 0.005430329125374556, + 0.00541008822619915, + 0.022043351083993912, + 0.012814908288419247, + 0.004840548615902662, + 0.004808908328413963, + 0.005823726765811443, + 0.016783328726887703, + 0.0043282401748001575, + 0.0038706292398273945, + 0.007951244711875916, + 0.010244232602417469, + 0.00424689706414938, + 0.00404982129111886, + 0.005849942564964294, + 0.006230603903532028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.04304524511098862 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic, sweeping, high-angle drone shot slowly glides forward and slightly downward, revealing a grand Gothic ballroom rendered in stunning, surreal 3D. A soaring vaulted ceiling with dark stone ribs reaches into the shadows above, flanked by towering stained-glass arched windows that depict a somber and romantic scene. The entire opulent space is illuminated by the warm, shimmering glow of a massive crystal chandelier, whose thousands of facets cast dancing golden rays across the highly polished black marble floor, creating near-perfect, shimmering reflections. Below, hundreds of guests gathered in intimate groups, dressed in exquisite gowns of deep velvet, shimmering silks, and intricate lace in rich jewel tones like crimson, emerald, and sapphire, alongside men in sharp, dark tuxedos. Each guest's identity is concealed behind an intricate phantom mask, ranging from simple, elegant dominoes to ornate creations composed of wire and dark feathers. As the camera continues its steady, almost imperceptible descent, the quiet nature of their conversation becomes apparent, not through sound but through their subtle, deliberate gestures: a fan tapping gently against a gloved palm, a masked head tilting slightly, conspiratorially, a hand gesturing elegantly to emphasize a silent point. The atmosphere is rich with luxurious mystery and exquisite elegance, a breathtaking picture of secret grandeur, where warm light catches the glint of jewels or the metallic sheen of masks, casting long, soft shadows that twist and dance with the flickering candlelight.", + "seed": 17766, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_1/index_4424_seed_17766_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024170941021293402, + "vlm_logprob": -0.024170941021293402, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.024170941021293402, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02229168452322483 + ], + "opposite_logprob_tokens": [ + -3.897291660308838 + ], + "logprob_reward": -0.02229168452322483, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02229168452322483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2363746613264084, + 0.021694477647542953, + 0.20165956020355225, + 0.043217889964580536, + 0.11170200258493423, + 0.041467491537332535, + 0.07624805718660355, + 0.19469450414180756, + 0.02595776878297329, + 0.046983502805233 + ], + "credit_map": [ + 0.02906774915754795, + 0.05681372061371803, + 0.01106836088001728, + 0.012044047936797142, + 0.03398554399609566, + 0.07100708037614822, + 0.012741855345666409, + 0.00964631699025631, + 0.0025644018314778805, + 0.002498264890164137, + 0.002848171629011631, + 0.002606054302304983, + 0.004137688782066107, + 0.002618086989969015, + 0.0024098388385027647, + 0.0020119703840464354, + 0.02034698985517025, + 0.02000170387327671, + 0.06339679658412933, + 0.008502444252371788, + 0.045570652931928635, + 0.009736424311995506, + 0.011552087031304836, + 0.022552471607923508, + 0.004289687145501375, + 0.004326586611568928, + 0.01222211867570877, + 0.004015354905277491, + 0.005409781821072102, + 0.0033483216539025307, + 0.002829627599567175, + 0.006776413414627314, + 0.016150906682014465, + 0.014239897020161152, + 0.014221038669347763, + 0.009378276765346527, + 0.020686371251940727, + 0.013692477717995644, + 0.013478033244609833, + 0.009854999370872974, + 0.0038647507317364216, + 0.005891211796551943, + 0.009031374007463455, + 0.00404741708189249, + 0.006440778728574514, + 0.004400474950671196, + 0.004085102118551731, + 0.003706380957737565, + 0.00559663400053978, + 0.008412688039243221, + 0.02139771357178688, + 0.01948811300098896, + 0.004030891694128513, + 0.005333330016583204, + 0.007204112596809864, + 0.004784571938216686, + 0.016059326007962227, + 0.018221251666545868, + 0.019816039130091667, + 0.04000344127416611, + 0.01922937110066414, + 0.032184649258852005, + 0.025333503261208534, + 0.023846931755542755, + 0.0030564391054213047, + 0.0035165625158697367, + 0.003103518858551979, + 0.003163715358823538, + 0.003391129896044731, + 0.0029068475123494864, + 0.003507642075419426, + 0.0033119129948318005, + 0.0035716129932552576, + 0.005593039095401764, + 0.008202573284506798, + 0.007181389257311821, + 0.004666555672883987, + 0.004841105546802282, + 0.0053864577785134315, + 0.007540770340710878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.02229168452322483 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372564539313316 + ], + "opposite_logprob_tokens": [ + -3.4087257385253906 + ], + "logprob_reward": -0.03372564539313316, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372564539313316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19245962798595428, + 0.018057435750961304, + 0.08426413685083389, + 0.060025375336408615, + 0.19715027511119843, + 0.04057862237095833, + 0.17568190395832062, + 0.10270915180444717, + 0.020289739593863487, + 0.10878373682498932 + ], + "credit_map": [ + 0.024821538478136063, + 0.04494207724928856, + 0.009780586697161198, + 0.011794056743383408, + 0.025643879547715187, + 0.05708732455968857, + 0.00903776753693819, + 0.009352393448352814, + 0.0015485096955671906, + 0.0029215810354799032, + 0.002406272804364562, + 0.002312849508598447, + 0.0035635028034448624, + 0.002001568442210555, + 0.0016816642601042986, + 0.0016214875504374504, + 0.005260723177343607, + 0.014153311029076576, + 0.00608640955761075, + 0.004190459381788969, + 0.028874816372990608, + 0.005344853736460209, + 0.004703809041529894, + 0.01564975269138813, + 0.006229848135262728, + 0.005090395454317331, + 0.01357593946158886, + 0.007271184120327234, + 0.007228049915283918, + 0.004906344693154097, + 0.005121670197695494, + 0.010601945221424103, + 0.029172997921705246, + 0.023655572906136513, + 0.026484111323952675, + 0.01358635164797306, + 0.04726941138505936, + 0.019068127498030663, + 0.024794619530439377, + 0.013119067996740341, + 0.004689734894782305, + 0.006713849492371082, + 0.006996162701398134, + 0.0038120117969810963, + 0.005581859033554792, + 0.0047597154043614864, + 0.0041307974606752396, + 0.003894492518156767, + 0.010376591235399246, + 0.01801898516714573, + 0.0627196654677391, + 0.058540716767311096, + 0.005044565536081791, + 0.005029995925724506, + 0.010151103138923645, + 0.005800274666398764, + 0.008193185552954674, + 0.009970582090318203, + 0.010420628823339939, + 0.0226396843791008, + 0.009749453514814377, + 0.014743677340447903, + 0.014303240925073624, + 0.012688696384429932, + 0.0025927249807864428, + 0.003572981571778655, + 0.0021750398445874453, + 0.002340859267860651, + 0.0023320158943533897, + 0.001948973280377686, + 0.0023293038830161095, + 0.0029978430829942226, + 0.010111105628311634, + 0.01928633823990822, + 0.017255345359444618, + 0.014647483825683594, + 0.012799527496099472, + 0.010930575430393219, + 0.009548735804855824, + 0.014204626902937889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.03372564539313316 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646931774914265 + ], + "opposite_logprob_tokens": [ + -5.3796467781066895 + ], + "logprob_reward": -0.004646931774914265, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646931774914265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1653556227684021, + 0.021963169798254967, + 0.2231544852256775, + 0.07433020323514938, + 0.02564050257205963, + 0.04133837670087814, + 0.06972326338291168, + 0.1598002016544342, + 0.03195227310061455, + 0.1867419183254242 + ], + "credit_map": [ + 0.021008998155593872, + 0.037557389587163925, + 0.009373508393764496, + 0.009420222602784634, + 0.021069979295134544, + 0.05181217938661575, + 0.008108865469694138, + 0.007004470098763704, + 0.002509977435693145, + 0.003341750940307975, + 0.004830649588257074, + 0.0020135140512138605, + 0.0032868743874132633, + 0.0026363262441009283, + 0.0018644578522071242, + 0.001479618833400309, + 0.013583263382315636, + 0.04026641696691513, + 0.016757912933826447, + 0.006457228679209948, + 0.0812591016292572, + 0.012874006293714046, + 0.01007518544793129, + 0.04188138246536255, + 0.007631036918610334, + 0.0036946681793779135, + 0.023955630138516426, + 0.006224866025149822, + 0.003838096046820283, + 0.005399077665060759, + 0.0038785336073487997, + 0.019708288833498955, + 0.0028967014513909817, + 0.0036736030597239733, + 0.0030796846840530634, + 0.002025273395702243, + 0.003995523322373629, + 0.0033278765622526407, + 0.0024151611141860485, + 0.004226678982377052, + 0.004537602886557579, + 0.006190067622810602, + 0.009466471150517464, + 0.003342659445479512, + 0.005661785136908293, + 0.004276306834071875, + 0.004763647448271513, + 0.003099834779277444, + 0.0046402704901993275, + 0.007318733260035515, + 0.020800933241844177, + 0.020699219778180122, + 0.0036224606446921825, + 0.003937805537134409, + 0.0052488502115011215, + 0.003454993013292551, + 0.01649431884288788, + 0.01858210191130638, + 0.01646360009908676, + 0.029987353831529617, + 0.015577279031276703, + 0.023444917052984238, + 0.020222505554556847, + 0.019028130918741226, + 0.0038835934828966856, + 0.005106369964778423, + 0.0038148469757288694, + 0.004304270725697279, + 0.003756077727302909, + 0.003467835718765855, + 0.004215465858578682, + 0.0034038128796964884, + 0.019769495353102684, + 0.04182162135839462, + 0.027694297954440117, + 0.02640325203537941, + 0.02377394586801529, + 0.013618276454508305, + 0.015541164204478264, + 0.018119845539331436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.004646931774914265 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252846051007509 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.005252846051007509, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252846051007509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27773866057395935, + 0.05085283890366554, + 0.15305477380752563, + 0.11742381751537323, + 0.040785036981105804, + 0.051222190260887146, + 0.1471610963344574, + 0.03085874579846859, + 0.027410611510276794, + 0.10349225997924805 + ], + "credit_map": [ + 0.032708510756492615, + 0.06828824430704117, + 0.01618015766143799, + 0.01663961261510849, + 0.038582753390073776, + 0.07899578660726547, + 0.013771108351647854, + 0.012572471052408218, + 0.00594269810244441, + 0.007893640547990799, + 0.012487836182117462, + 0.004533157683908939, + 0.005841692443937063, + 0.004689894150942564, + 0.005041782278567553, + 0.004422139376401901, + 0.01278135646134615, + 0.01792665757238865, + 0.03520077094435692, + 0.005740089807659388, + 0.032643161714076996, + 0.007302858401089907, + 0.006716627161949873, + 0.03474324941635132, + 0.011406142264604568, + 0.007464907132089138, + 0.036879900842905045, + 0.010446717031300068, + 0.00949858408421278, + 0.008131700567901134, + 0.006028619594871998, + 0.027567245066165924, + 0.004623087588697672, + 0.00490296445786953, + 0.004972346592694521, + 0.004487550817430019, + 0.00649844016879797, + 0.005076235625892878, + 0.005398687440901995, + 0.004825724754482508, + 0.004606638569384813, + 0.007925588637590408, + 0.012598847039043903, + 0.004139108117669821, + 0.006588086951524019, + 0.004731787834316492, + 0.004819524474442005, + 0.005812604911625385, + 0.00956403836607933, + 0.016593191772699356, + 0.051139120012521744, + 0.04452112689614296, + 0.0061721280217170715, + 0.00534941628575325, + 0.00842678640037775, + 0.005395287647843361, + 0.0036481071729213, + 0.0042194221168756485, + 0.0036040926352143288, + 0.005599913652986288, + 0.0032911261077970266, + 0.0032867416739463806, + 0.0032791942358016968, + 0.0039301482029259205, + 0.003017681185156107, + 0.003396007465198636, + 0.0035267441999167204, + 0.004490705672651529, + 0.002736842492595315, + 0.0039566135965287685, + 0.003285527229309082, + 0.003000489203259349, + 0.009525139816105366, + 0.018389740958809853, + 0.016218330711126328, + 0.011814813129603863, + 0.011400450952351093, + 0.006986312568187714, + 0.01665295660495758, + 0.012504510581493378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.005252846051007509 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493759736418724 + ], + "opposite_logprob_tokens": [ + -2.9299376010894775 + ], + "logprob_reward": -0.05493759736418724, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493759736418724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06824392825365067, + 0.01602787710726261, + 0.24186670780181885, + 0.20893898606300354, + 0.01777462288737297, + 0.015225040726363659, + 0.11145524680614471, + 0.18935105204582214, + 0.047368988394737244, + 0.08374753594398499 + ], + "credit_map": [ + 0.01062209252268076, + 0.011059226468205452, + 0.004578817635774612, + 0.0068610478192567825, + 0.006619041319936514, + 0.020541362464427948, + 0.004912111908197403, + 0.003050229512155056, + 0.002214695094153285, + 0.0015449428465217352, + 0.0026015671901404858, + 0.0029440412763506174, + 0.001263712882064283, + 0.0012228814885020256, + 0.0020978376269340515, + 0.002138200681656599, + 0.01596953719854355, + 0.04206118732690811, + 0.03273846209049225, + 0.008093653246760368, + 0.0885963961482048, + 0.014149757102131844, + 0.011781820096075535, + 0.02847590669989586, + 0.027464795857667923, + 0.011612453497946262, + 0.055817484855651855, + 0.01709464192390442, + 0.011661379598081112, + 0.01885031908750534, + 0.01686609350144863, + 0.04957182705402374, + 0.0027673770673573017, + 0.002751808613538742, + 0.002353296848013997, + 0.002066225279122591, + 0.0026873203460127115, + 0.0021371536422520876, + 0.0014582062140107155, + 0.0015532344114035368, + 0.0014452867908403277, + 0.0020344513468444347, + 0.003043557284399867, + 0.0017059496603906155, + 0.0014248369261622429, + 0.0017126566963270307, + 0.002121505094692111, + 0.001736797858029604, + 0.007071542087942362, + 0.012018034234642982, + 0.03869542106986046, + 0.034740377217531204, + 0.003615250578150153, + 0.004415637347847223, + 0.006479617208242416, + 0.004419379867613316, + 0.017863448709249496, + 0.023133765906095505, + 0.018144872039556503, + 0.03945503011345863, + 0.019989365711808205, + 0.023826416581869125, + 0.023795826360583305, + 0.023142343387007713, + 0.005607736762613058, + 0.005606857128441334, + 0.006006441079080105, + 0.006585846655070782, + 0.0061371480114758015, + 0.006026179529726505, + 0.006190236657857895, + 0.005208548158407211, + 0.008312818594276905, + 0.017450982704758644, + 0.012189295142889023, + 0.010939630679786205, + 0.010584341362118721, + 0.0060209110379219055, + 0.008783296681940556, + 0.009466265328228474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.05493759736418724 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic, sweeping, high-angle drone shot slowly glides forward and slightly downward, revealing a grand Gothic ballroom rendered in stunning, surreal 3D. A soaring vaulted ceiling with dark stone ribs reaches into the shadows above, flanked by towering stained-glass arched windows that depict a somber and romantic scene. The entire opulent space is illuminated by the warm, shimmering glow of a massive crystal chandelier, whose thousands of facets cast dancing golden rays across the highly polished black marble floor, creating near-perfect, shimmering reflections. Below, hundreds of guests gathered in intimate groups, dressed in exquisite gowns of deep velvet, shimmering silks, and intricate lace in rich jewel tones like crimson, emerald, and sapphire, alongside men in sharp, dark tuxedos. Each guest's identity is concealed behind an intricate phantom mask, ranging from simple, elegant dominoes to ornate creations composed of wire and dark feathers. As the camera continues its steady, almost imperceptible descent, the quiet nature of their conversation becomes apparent, not through sound but through their subtle, deliberate gestures: a fan tapping gently against a gloved palm, a masked head tilting slightly, conspiratorially, a hand gesturing elegantly to emphasize a silent point. The atmosphere is rich with luxurious mystery and exquisite elegance, a breathtaking picture of secret grandeur, where warm light catches the glint of jewels or the metallic sheen of masks, casting long, soft shadows that twist and dance with the flickering candlelight.", + "seed": 17767, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_1/index_4424_seed_17767_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.035072869155555964, + "vlm_logprob": -0.035072869155555964, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.035072869155555964, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.051461149007081985 + ], + "opposite_logprob_tokens": [ + -3.0514612197875977 + ], + "logprob_reward": -0.051461149007081985, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.051461149007081985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1522180140018463, + 0.23315811157226562, + 0.13315503299236298, + 0.08357620239257812, + 0.14824961125850677, + 0.09921205043792725, + 0.036797381937503815, + 0.030637336894869804, + 0.04244668036699295, + 0.040549516677856445 + ], + "credit_map": [ + 0.01608063280582428, + 0.014918219298124313, + 0.04626685380935669, + 0.02136550471186638, + 0.011886291205883026, + 0.012496662326157093, + 0.015096936374902725, + 0.01410691812634468, + 0.0309443362057209, + 0.015245137736201286, + 0.049364760518074036, + 0.033924687653779984, + 0.021087070927023888, + 0.03157748281955719, + 0.026860928162932396, + 0.024153726175427437, + 0.02057243138551712, + 0.009807213209569454, + 0.018960226327180862, + 0.016463810577988625, + 0.017763394862413406, + 0.01463074516505003, + 0.0170395877212286, + 0.017917636781930923, + 0.008477259427309036, + 0.009125685319304466, + 0.012368779629468918, + 0.008734053932130337, + 0.010598353110253811, + 0.012638970278203487, + 0.009400355629622936, + 0.012232746928930283, + 0.016127871349453926, + 0.01893359236419201, + 0.024506933987140656, + 0.019061246886849403, + 0.011782638728618622, + 0.009343196637928486, + 0.024518445134162903, + 0.023975692689418793, + 0.008308429270982742, + 0.010813276283442974, + 0.012680872343480587, + 0.012631956487894058, + 0.008955327793955803, + 0.007005107123404741, + 0.024327538907527924, + 0.01448954176157713, + 0.0033640842884778976, + 0.003325482364743948, + 0.00486871087923646, + 0.0038270270451903343, + 0.0034618559293448925, + 0.0038339742459356785, + 0.007887730374932289, + 0.00622851587831974, + 0.004032009281218052, + 0.0028460160829126835, + 0.005022892728447914, + 0.0028400272130966187, + 0.0024194251745939255, + 0.0018162780907005072, + 0.006559777073562145, + 0.005100911017507315, + 0.0051028490997850895, + 0.005311756860464811, + 0.005815338343381882, + 0.0034223939292132854, + 0.004526339005678892, + 0.0032867470290511847, + 0.008143821731209755, + 0.0068374318070709705, + 0.0044362773187458515, + 0.007997776381671429, + 0.0044557759538292885, + 0.0036129767540842295, + 0.0029234641697257757, + 0.0027743580285459757, + 0.004613079130649567, + 0.009735806845128536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.051461149007081985 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372460976243019 + ], + "opposite_logprob_tokens": [ + -3.408724546432495 + ], + "logprob_reward": -0.03372460976243019, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372460976243019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.080381378531456, + 0.1319940835237503, + 0.18651798367500305, + 0.19570893049240112, + 0.049343157559633255, + 0.12211848795413971, + 0.031337544322013855, + 0.028710028156638145, + 0.1328543871641159, + 0.041034065186977386 + ], + "credit_map": [ + 0.009717966429889202, + 0.0053226519376039505, + 0.013686233200132847, + 0.014541850425302982, + 0.009152601473033428, + 0.008126071654260159, + 0.008803807199001312, + 0.011030188761651516, + 0.01563185453414917, + 0.014101733453571796, + 0.01784968376159668, + 0.015840068459510803, + 0.01443177554756403, + 0.019820386543869972, + 0.01457810029387474, + 0.019740469753742218, + 0.0162894856184721, + 0.020767655223608017, + 0.02406364306807518, + 0.024420250207185745, + 0.02319183386862278, + 0.02364135906100273, + 0.03172311186790466, + 0.02242063730955124, + 0.01877758279442787, + 0.035236019641160965, + 0.02744067646563053, + 0.02132905088365078, + 0.023669080808758736, + 0.018354812636971474, + 0.02410823665559292, + 0.026793453842401505, + 0.00625237263739109, + 0.006332047283649445, + 0.005198570899665356, + 0.00698045315220952, + 0.0047386083751916885, + 0.0063417875207960606, + 0.007027302403002977, + 0.006472012959420681, + 0.01009803730994463, + 0.009389912709593773, + 0.027653314173221588, + 0.017092619091272354, + 0.008570518344640732, + 0.006419254932552576, + 0.023963624611496925, + 0.01893119513988495, + 0.003781333100050688, + 0.002817395143210888, + 0.004850365221500397, + 0.003361936192959547, + 0.0024391652550548315, + 0.003429885720834136, + 0.005414014682173729, + 0.005243447609245777, + 0.004520303104072809, + 0.003434777492657304, + 0.00405988609418273, + 0.00237530586309731, + 0.0024170728866010904, + 0.0018784754211083055, + 0.005433286540210247, + 0.004590919241309166, + 0.01693429797887802, + 0.01835152506828308, + 0.017182625830173492, + 0.010325336828827858, + 0.015533673577010632, + 0.015138179995119572, + 0.018318548798561096, + 0.02107018604874611, + 0.005054810084402561, + 0.006557789631187916, + 0.004385306965559721, + 0.004231276921927929, + 0.002623221604153514, + 0.0035921307280659676, + 0.005231454502791166, + 0.009358071722090244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.03372460976243019 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005261383950710297 + ], + "opposite_logprob_tokens": [ + -5.255261421203613 + ], + "logprob_reward": -0.005261383950710297, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005261383950710297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05041603371500969, + 0.06170393154025078, + 0.12490573525428772, + 0.5537186861038208, + 0.046833284199237823, + 0.05000073462724686, + 0.01827637292444706, + 0.03497493267059326, + 0.04141191393136978, + 0.017758408561348915 + ], + "credit_map": [ + 0.005726556293666363, + 0.0055094133131206036, + 0.008648334071040154, + 0.010642175562679768, + 0.004568914417177439, + 0.003995132632553577, + 0.005796173121780157, + 0.005529333837330341, + 0.006576813757419586, + 0.006782755255699158, + 0.008699927479028702, + 0.009215251542627811, + 0.00599670922383666, + 0.007512433920055628, + 0.008833849802613258, + 0.0080861896276474, + 0.013923073187470436, + 0.007735670544207096, + 0.0341210812330246, + 0.023317793384194374, + 0.009250117465853691, + 0.008061189204454422, + 0.015590636059641838, + 0.012906182557344437, + 0.049242053180933, + 0.13083302974700928, + 0.06140722706913948, + 0.05233394354581833, + 0.094681516289711, + 0.04092651605606079, + 0.048532597720623016, + 0.07576180249452591, + 0.0035911870654672384, + 0.009750432334840298, + 0.006312140263617039, + 0.0059326086193323135, + 0.004986113402992487, + 0.002491020830348134, + 0.00736600486561656, + 0.006403777282685041, + 0.003474356373772025, + 0.003906435798853636, + 0.012140032835304737, + 0.006253042723983526, + 0.003255352843552828, + 0.0020514107309281826, + 0.01092567015439272, + 0.00799443107098341, + 0.002262309193611145, + 0.0018595138099044561, + 0.0025297091342508793, + 0.0034468453377485275, + 0.0013101872755214572, + 0.0010093202581629157, + 0.0030595797579735518, + 0.002798908157274127, + 0.004800376947969198, + 0.003698777873069048, + 0.005808165762573481, + 0.0027027891483157873, + 0.0026237210258841515, + 0.0022304269950836897, + 0.0069348993711173534, + 0.006175772286951542, + 0.00460665300488472, + 0.005210184957832098, + 0.005525541491806507, + 0.004854740109294653, + 0.0038050319999456406, + 0.0032737329602241516, + 0.007164996117353439, + 0.006971035152673721, + 0.002202715491876006, + 0.0024316776543855667, + 0.002586713992059231, + 0.002144777914509177, + 0.0014249187661334872, + 0.002071759197860956, + 0.0024642623029649258, + 0.0024315831251442432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.005261383950710297 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951777566224337 + ], + "opposite_logprob_tokens": [ + -5.130951881408691 + ], + "logprob_reward": -0.005951777566224337, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951777566224337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.277793824672699, + 0.14315856993198395, + 0.08884738385677338, + 0.056617118418216705, + 0.053506478667259216, + 0.12454862147569656, + 0.05502866953611374, + 0.05151490494608879, + 0.06887814402580261, + 0.08010624349117279 + ], + "credit_map": [ + 0.028687646612524986, + 0.04434100165963173, + 0.09719933569431305, + 0.027932438999414444, + 0.020875651389360428, + 0.0219995379447937, + 0.017792215570807457, + 0.018966015428304672, + 0.019671356305480003, + 0.012781664729118347, + 0.021495027467608452, + 0.01729726605117321, + 0.01757395640015602, + 0.02137078531086445, + 0.018775949254631996, + 0.014192578382790089, + 0.008733080700039864, + 0.00973033532500267, + 0.017192717641592026, + 0.011722216382622719, + 0.007014918606728315, + 0.00879101175814867, + 0.009975606575608253, + 0.015687493607401848, + 0.004548376426100731, + 0.011044272221624851, + 0.007856477983295918, + 0.008277269080281258, + 0.007803834974765778, + 0.00413819495588541, + 0.006078446749597788, + 0.006870246957987547, + 0.004757308401167393, + 0.007031926419585943, + 0.010712506249547005, + 0.00648080138489604, + 0.0038912752643227577, + 0.0044487821869552135, + 0.0074670878238976, + 0.008716791868209839, + 0.006550283636897802, + 0.01140976045280695, + 0.019436804577708244, + 0.023633873090147972, + 0.005467467475682497, + 0.00401195278391242, + 0.03408230468630791, + 0.019956190139055252, + 0.005282757803797722, + 0.010236047208309174, + 0.007859781384468079, + 0.007350518833845854, + 0.0044960519298911095, + 0.0018629974219948053, + 0.010143662802875042, + 0.007796857040375471, + 0.007128030061721802, + 0.005354730412364006, + 0.008933676406741142, + 0.007241710089147091, + 0.0029501114040613174, + 0.0029488662257790565, + 0.009275328367948532, + 0.007682456634938717, + 0.005678440909832716, + 0.007602999452501535, + 0.014717916026711464, + 0.0068806535564363, + 0.004223493859171867, + 0.004424885381013155, + 0.010981754399836063, + 0.014368012547492981, + 0.008981520310044289, + 0.008691534399986267, + 0.015107347629964352, + 0.011045923456549644, + 0.0035501145757734776, + 0.004803144372999668, + 0.014181246049702168, + 0.01374542061239481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.005951777566224337 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896542549133301 + ], + "opposite_logprob_tokens": [ + -2.578965425491333 + ], + "logprob_reward": -0.07896542549133301, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896542549133301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1695307195186615, + 0.16504321992397308, + 0.11468754708766937, + 0.1692509651184082, + 0.14596587419509888, + 0.08510259538888931, + 0.04657452180981636, + 0.050437916070222855, + 0.031290166079998016, + 0.02211645059287548 + ], + "credit_map": [ + 0.017590845003724098, + 0.012142462655901909, + 0.047213103622198105, + 0.02274801954627037, + 0.013201367110013962, + 0.02011946775019169, + 0.019878223538398743, + 0.01663723960518837, + 0.023147065192461014, + 0.011920368298888206, + 0.04011276736855507, + 0.02947014942765236, + 0.013199220411479473, + 0.014943944290280342, + 0.015475465916097164, + 0.016774246469140053, + 0.012132367119193077, + 0.00871456041932106, + 0.022988779470324516, + 0.015959400683641434, + 0.010491103865206242, + 0.009919409640133381, + 0.020336078479886055, + 0.014145848341286182, + 0.016240285709500313, + 0.031352657824754715, + 0.020538033917546272, + 0.015945592895150185, + 0.02400089241564274, + 0.01576484739780426, + 0.020928770303726196, + 0.024479899555444717, + 0.015980534255504608, + 0.018992677330970764, + 0.022643638774752617, + 0.023796314373612404, + 0.007487177848815918, + 0.008527991361916065, + 0.025334550067782402, + 0.023203004151582718, + 0.005848918575793505, + 0.006963507272303104, + 0.0185929574072361, + 0.015223011374473572, + 0.004542630631476641, + 0.003177100559696555, + 0.016431597992777824, + 0.014322878792881966, + 0.006499671842902899, + 0.004731652792543173, + 0.006716717965900898, + 0.0066566793248057365, + 0.003792861942201853, + 0.003278236137703061, + 0.008208520710468292, + 0.006690183654427528, + 0.008578013628721237, + 0.004486483056098223, + 0.007758012041449547, + 0.006007606629282236, + 0.0036664947401732206, + 0.0032248469069600105, + 0.008507664315402508, + 0.00820879451930523, + 0.0037898817099630833, + 0.0034979921765625477, + 0.004533003084361553, + 0.004869252443313599, + 0.0017211706144735217, + 0.002201635157689452, + 0.005531854461878538, + 0.005145380273461342, + 0.0026314265560358763, + 0.0028627023566514254, + 0.00386793608777225, + 0.002582772634923458, + 0.0013121658703312278, + 0.002145119709894061, + 0.0035129261668771505, + 0.0032014029566198587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.07896542549133301 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic, sweeping, high-angle drone shot slowly glides forward and slightly downward, revealing a grand Gothic ballroom rendered in stunning, surreal 3D. A soaring vaulted ceiling with dark stone ribs reaches into the shadows above, flanked by towering stained-glass arched windows that depict a somber and romantic scene. The entire opulent space is illuminated by the warm, shimmering glow of a massive crystal chandelier, whose thousands of facets cast dancing golden rays across the highly polished black marble floor, creating near-perfect, shimmering reflections. Below, hundreds of guests gathered in intimate groups, dressed in exquisite gowns of deep velvet, shimmering silks, and intricate lace in rich jewel tones like crimson, emerald, and sapphire, alongside men in sharp, dark tuxedos. Each guest's identity is concealed behind an intricate phantom mask, ranging from simple, elegant dominoes to ornate creations composed of wire and dark feathers. As the camera continues its steady, almost imperceptible descent, the quiet nature of their conversation becomes apparent, not through sound but through their subtle, deliberate gestures: a fan tapping gently against a gloved palm, a masked head tilting slightly, conspiratorially, a hand gesturing elegantly to emphasize a silent point. The atmosphere is rich with luxurious mystery and exquisite elegance, a breathtaking picture of secret grandeur, where warm light catches the glint of jewels or the metallic sheen of masks, casting long, soft shadows that twist and dance with the flickering candlelight.", + "seed": 17768, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_1/index_4424_seed_17768_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12801257586106657, + "vlm_logprob": -0.12801257586106657, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.12801257586106657, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4755517840385437 + ], + "opposite_logprob_tokens": [ + -0.9755517840385437 + ], + "logprob_reward": -0.4755517840385437, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4755517840385437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11208532750606537, + 0.19177144765853882, + 0.11155828833580017, + 0.13959921896457672, + 0.14702928066253662, + 0.023705458268523216, + 0.047205548733472824, + 0.06186698377132416, + 0.07726014405488968, + 0.08791834115982056 + ], + "credit_map": [ + 0.019207296893000603, + 0.01037965901196003, + 0.013536477461457253, + 0.019204026088118553, + 0.008478973992168903, + 0.018022341653704643, + 0.01389030460268259, + 0.009366244077682495, + 0.019728606566786766, + 0.01685270294547081, + 0.024015216156840324, + 0.04531453549861908, + 0.01716281846165657, + 0.013631847687065601, + 0.01626993529498577, + 0.03879578784108162, + 0.013355962932109833, + 0.008407366462051868, + 0.01720944605767727, + 0.01767227239906788, + 0.007523598149418831, + 0.01193886436522007, + 0.010327938944101334, + 0.025122839957475662, + 0.008195382542908192, + 0.016931096091866493, + 0.013249669224023819, + 0.01876082271337509, + 0.020866135135293007, + 0.013630461879074574, + 0.023680657148361206, + 0.02428499236702919, + 0.006001163274049759, + 0.008550186641514301, + 0.011316714808344841, + 0.04645325988531113, + 0.006305901799350977, + 0.008616329170763493, + 0.013259661383926868, + 0.046526066958904266, + 0.0015032347291707993, + 0.0030781410168856382, + 0.0030551180243492126, + 0.004938643425703049, + 0.0022651185281574726, + 0.003106002463027835, + 0.0031469285022467375, + 0.002612270647659898, + 0.004304312169551849, + 0.005096224136650562, + 0.006386168766766787, + 0.008844707161188126, + 0.0037732445634901524, + 0.005214279051870108, + 0.007338305469602346, + 0.006248308811336756, + 0.003949684090912342, + 0.004994641989469528, + 0.007043788675218821, + 0.01584070920944214, + 0.004307603929191828, + 0.004577263258397579, + 0.014593206346035004, + 0.00656008068472147, + 0.003710668534040451, + 0.005118518136441708, + 0.009352890774607658, + 0.027350978925824165, + 0.003392645390704274, + 0.004544551949948072, + 0.008605550974607468, + 0.015184341929852962, + 0.0053056092001497746, + 0.009286325424909592, + 0.01569703035056591, + 0.013885152526199818, + 0.009482449851930141, + 0.008657642640173435, + 0.013809099793434143, + 0.011795034632086754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.4755517840385437 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890740305185318 + ], + "opposite_logprob_tokens": [ + -2.46407413482666 + ], + "logprob_reward": -0.0890740305185318, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890740305185318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1111249029636383, + 0.2046469897031784, + 0.10350879281759262, + 0.050166353583335876, + 0.09535712748765945, + 0.019342826679348946, + 0.03874494880437851, + 0.26566508412361145, + 0.054156895726919174, + 0.057286087423563004 + ], + "credit_map": [ + 0.02247568778693676, + 0.009768420830368996, + 0.01211864035576582, + 0.021078091114759445, + 0.010065124370157719, + 0.018605370074510574, + 0.007847103290259838, + 0.009166470728814602, + 0.021955065429210663, + 0.021613016724586487, + 0.026072144508361816, + 0.03848598152399063, + 0.02088727056980133, + 0.02187437377870083, + 0.01842421665787697, + 0.035334937274456024, + 0.009817486628890038, + 0.007441755384206772, + 0.01421093475073576, + 0.018236804753541946, + 0.009508120827376842, + 0.009620198979973793, + 0.008486239239573479, + 0.026187263429164886, + 0.003222391940653324, + 0.005618254188448191, + 0.005284624639898539, + 0.006184462457895279, + 0.006792540196329355, + 0.006206916179507971, + 0.00966513529419899, + 0.0071920352056622505, + 0.0037431942764669657, + 0.004967674147337675, + 0.006891816854476929, + 0.03192543610930443, + 0.0027295590844005346, + 0.0057573760859668255, + 0.008054147474467754, + 0.03128792718052864, + 0.0010538325877860188, + 0.0031982259824872017, + 0.0022363485768437386, + 0.0036312195006757975, + 0.0019132469315081835, + 0.001634520129300654, + 0.0035230759531259537, + 0.00215235841460526, + 0.0038284100592136383, + 0.0037919662427157164, + 0.004686923697590828, + 0.006440242752432823, + 0.004499782808125019, + 0.004867828916758299, + 0.00534092215821147, + 0.005288873799145222, + 0.008601530455052853, + 0.013655567541718483, + 0.035532839596271515, + 0.1068677231669426, + 0.009955295361578465, + 0.00819181464612484, + 0.06745581328868866, + 0.015404506586492062, + 0.003366343677043915, + 0.0039239078760147095, + 0.007489768788218498, + 0.01470824796706438, + 0.004969604313373566, + 0.0034466024953871965, + 0.00786488875746727, + 0.00838753767311573, + 0.003575557144358754, + 0.007891790941357613, + 0.007277444936335087, + 0.009304282255470753, + 0.006166040897369385, + 0.006098969839513302, + 0.0072580487467348576, + 0.009713954292237759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.0890740305185318 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036216876469552517 + ], + "opposite_logprob_tokens": [ + -5.628621578216553 + ], + "logprob_reward": -0.0036216876469552517, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036216876469552517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11241636425256729, + 0.08827103674411774, + 0.07332336157560349, + 0.05292017385363579, + 0.06425394862890244, + 0.018817992880940437, + 0.03671741485595703, + 0.27108505368232727, + 0.2097088098526001, + 0.07248588651418686 + ], + "credit_map": [ + 0.03597529977560043, + 0.006604761816561222, + 0.006938903592526913, + 0.034915946424007416, + 0.0053354064002633095, + 0.007082396186888218, + 0.007814179174602032, + 0.007749459240585566, + 0.006220147479325533, + 0.004818721208721399, + 0.006786112207919359, + 0.031465210020542145, + 0.004257264547049999, + 0.00417367136105895, + 0.005065459292382002, + 0.025484446436166763, + 0.006724721752107143, + 0.005771935917437077, + 0.007953289896249771, + 0.012756568379700184, + 0.008016142062842846, + 0.0060195778496563435, + 0.006910555996000767, + 0.01917055808007717, + 0.0036426447331905365, + 0.005157426930963993, + 0.00632950384169817, + 0.007307779043912888, + 0.005640274845063686, + 0.005161105189472437, + 0.009943017736077309, + 0.009738416410982609, + 0.003254444571211934, + 0.0045392755419015884, + 0.006270962301641703, + 0.017490366473793983, + 0.004202515352517366, + 0.004306072369217873, + 0.006861939560621977, + 0.017328381538391113, + 0.0012703296961262822, + 0.0026084580458700657, + 0.0024723042733967304, + 0.0033303042873740196, + 0.0017471309984102845, + 0.0016635933425277472, + 0.0032750226091593504, + 0.0024508493952453136, + 0.0028226107824593782, + 0.0033517819829285145, + 0.0046210335567593575, + 0.006602582987397909, + 0.0033233410213142633, + 0.004245224874466658, + 0.005571113433688879, + 0.006179723422974348, + 0.009496892802417278, + 0.01443992368876934, + 0.03600170835852623, + 0.10730596631765366, + 0.009599925018846989, + 0.008670834824442863, + 0.06786565482616425, + 0.017704138532280922, + 0.013561073690652847, + 0.013524433597922325, + 0.0347362719476223, + 0.0839868038892746, + 0.014027547091245651, + 0.008437955752015114, + 0.022579364478588104, + 0.018855353817343712, + 0.005537218414247036, + 0.009062801487743855, + 0.013732059858739376, + 0.011644667945802212, + 0.005491953808814287, + 0.005104852840304375, + 0.011121029034256935, + 0.010791296139359474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.0036216876469552517 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784122928977013 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784122928977013, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784122928977013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08587296307086945, + 0.04669671133160591, + 0.057698823511600494, + 0.06429900974035263, + 0.06255783140659332, + 0.03762408345937729, + 0.03844594582915306, + 0.39689773321151733, + 0.14639350771903992, + 0.06351330876350403 + ], + "credit_map": [ + 0.014416910707950592, + 0.008639550767838955, + 0.008938740007579327, + 0.01819506660103798, + 0.008492776192724705, + 0.010776209644973278, + 0.007917921058833599, + 0.008495790883898735, + 0.0039215972647070885, + 0.004191488027572632, + 0.006635447032749653, + 0.012411242350935936, + 0.004538082983344793, + 0.0030699842609465122, + 0.0036429595202207565, + 0.008285916410386562, + 0.003524859668686986, + 0.005117154214531183, + 0.00403607077896595, + 0.01178182102739811, + 0.0081919701769948, + 0.0038975353818386793, + 0.003720942186191678, + 0.01742846891283989, + 0.005791212432086468, + 0.006274647545069456, + 0.007762387860566378, + 0.008773267269134521, + 0.0057643684558570385, + 0.00624705757945776, + 0.011626874096691608, + 0.012059192173182964, + 0.0036000264808535576, + 0.005260145757347345, + 0.006287398282438517, + 0.017917385324835777, + 0.0036305224057286978, + 0.005047845654189587, + 0.005582889541983604, + 0.01523161493241787, + 0.0026464557740837336, + 0.005948579404503107, + 0.005175097845494747, + 0.006432448513805866, + 0.0033227785024791956, + 0.0034060219768434763, + 0.006076240912079811, + 0.004616462625563145, + 0.0024273889139294624, + 0.005385581869632006, + 0.004712305031716824, + 0.006880845874547958, + 0.0030278225895017385, + 0.00301003223285079, + 0.006165821570903063, + 0.0068361503072083, + 0.013879576697945595, + 0.02399442344903946, + 0.05306030064821243, + 0.1557365208864212, + 0.018831724300980568, + 0.012774665839970112, + 0.09662599861621857, + 0.021994536742568016, + 0.00968275684863329, + 0.012403863482177258, + 0.022563308477401733, + 0.05653243139386177, + 0.00786131527274847, + 0.006179886870086193, + 0.015053728595376015, + 0.016116203740239143, + 0.004422810394316912, + 0.009056994691491127, + 0.011561259627342224, + 0.00959167443215847, + 0.006493676453828812, + 0.004904593341052532, + 0.008074638433754444, + 0.00940766092389822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.009784122928977013 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a grand Gothic ballroom with a soaring vaulted ceiling and dark stone ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there towering stained-glass arched windows depicting a somber and romantic scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crystal chandelier visible, casting dancing golden rays and shimmering reflections on a black marble floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hundreds of guests visible wearing masks (such as dominoes or wire/feather creations) and dressed in jewel-toned gowns and dark tuxedos?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a high-angle shot that slowly glides forward and slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062031254172325134 + ], + "opposite_logprob_tokens": [ + -2.8120312690734863 + ], + "logprob_reward": -0.062031254172325134, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062031254172325134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03481835126876831, + 0.03438247740268707, + 0.009525099769234657, + 0.016790049150586128, + 0.026031361892819405, + 0.01321823988109827, + 0.0260575320571661, + 0.637259840965271, + 0.14818744361400604, + 0.053729597479104996 + ], + "credit_map": [ + 0.009320860728621483, + 0.0023954063653945923, + 0.0032279076986014843, + 0.006872645113617182, + 0.0025217095389962196, + 0.0028708872850984335, + 0.0034338561818003654, + 0.004175079986453056, + 0.0021372537594288588, + 0.0026520267128944397, + 0.0021365215070545673, + 0.010794885456562042, + 0.003599327290430665, + 0.0019368637586012483, + 0.0023692038375884295, + 0.008756395429372787, + 0.0009813688229769468, + 0.0007417305023409426, + 0.0013325539184734225, + 0.0016503770602867007, + 0.0009194082813337445, + 0.0008648125221952796, + 0.000943028018809855, + 0.002091821515932679, + 0.0012849183985963464, + 0.00207942770794034, + 0.0020779322367161512, + 0.002315155928954482, + 0.002045045606791973, + 0.0018576879519969225, + 0.002351113362237811, + 0.0027787680737674236, + 0.0012494581751525402, + 0.001991222146898508, + 0.0022959760390222073, + 0.006818183697760105, + 0.0016458878526464105, + 0.0019681581761687994, + 0.002776846056804061, + 0.007285630330443382, + 0.0008195539703592658, + 0.0019484944641590118, + 0.0010839772876352072, + 0.00273098680190742, + 0.0011880185920745134, + 0.0013344257604330778, + 0.002675518626347184, + 0.0014372647274285555, + 0.0026392366271466017, + 0.0025227924343198538, + 0.003115906845778227, + 0.004352364223450422, + 0.0022694068029522896, + 0.003346165642142296, + 0.003954113461077213, + 0.00385754625312984, + 0.02193312719464302, + 0.032945554703474045, + 0.08453067392110825, + 0.2550528645515442, + 0.021709049120545387, + 0.019564641639590263, + 0.1635444164276123, + 0.03797950968146324, + 0.009750308468937874, + 0.010036153718829155, + 0.025058040395379066, + 0.05444851890206337, + 0.014272362925112247, + 0.006124007049947977, + 0.01478620432317257, + 0.013711836189031601, + 0.004676249343901873, + 0.006977073848247528, + 0.0070069339126348495, + 0.00869003776460886, + 0.004985183011740446, + 0.004385741427540779, + 0.006876659579575062, + 0.010131718590855598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.062031254172325134 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_10/index_tensor([8135, 8135])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_10/index_tensor([8135, 8135])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fdd2f42ef9da8c669857a4c2772d10c7a99d539e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_10/index_tensor([8135, 8135])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, taken from a horizontal angle, focuses on a hospital bed covered in white sheets. A male patient, with a sallow complexion and chapped lips, lies on it. He wears a loose, wrinkled hospital gown with light blue stripes, and his sparse, short black hair is matted to his forehead by cold sweat. Suddenly, a violent, rapid coughing fit seizes him. His chest heaves violently, his eyes squeeze shut in agony, and his brow knits together. With each heart-wrenching cough, his body involuntarily curls inward, his back arching in agony like a shrimp. One hand weakly covers his mouth, while the other clutches the sheets beneath him, his knuckles turning white from the strain. The cold, white light from the fluorescent ceiling lights pours down relentlessly, making his pale skin appear even more pale and casting faint yet sharp shadows on his twitching body and the dishevelled sheets. The entire scene, set against the monotonous pale green wall, is permeated with a suffocating atmosphere of pain, vulnerability, and helplessness.", + "seed": 7518, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_10/index_8135_seed_7518_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1450716469436884, + "vlm_logprob": -0.1450716469436884, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.1450716469436884, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019711211323738098 + ], + "opposite_logprob_tokens": [ + -4.144711017608643 + ], + "logprob_reward": -0.019711211323738098, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019711211323738098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09542237967252731, + 0.12334348261356354, + 0.06929633021354675, + 0.060003992170095444, + 0.05685450881719589, + 0.03338449075818062, + 0.04223499819636345, + 0.0903002917766571, + 0.38011622428894043, + 0.04904331639409065 + ], + "credit_map": [ + 0.00904186349362135, + 0.011242592707276344, + 0.010809944011271, + 0.011019153520464897, + 0.011945158243179321, + 0.012656762264668941, + 0.013984445482492447, + 0.014722458086907864, + 0.006722841411828995, + 0.013512246310710907, + 0.010803982615470886, + 0.008562244474887848, + 0.0364379920065403, + 0.02854834496974945, + 0.007038284093141556, + 0.011717547662556171, + 0.0048557971604168415, + 0.009387188591063023, + 0.00602624611929059, + 0.006848326418548822, + 0.011564772576093674, + 0.00848489161580801, + 0.012848664075136185, + 0.009280441328883171, + 0.004574896767735481, + 0.007461267989128828, + 0.0049703470431268215, + 0.006110549438744783, + 0.009648018516600132, + 0.009665546007454395, + 0.007924520410597324, + 0.009648844599723816, + 0.0055010938085615635, + 0.006945303175598383, + 0.005227726884186268, + 0.006048947107046843, + 0.008551658131182194, + 0.011327714659273624, + 0.0059990654699504375, + 0.007253000978380442, + 0.0021078260615468025, + 0.003598638344556093, + 0.003043551230803132, + 0.0026224011089652777, + 0.008771238848567009, + 0.007443159352988005, + 0.0022627434227615595, + 0.0035349309910088778, + 0.002937308279797435, + 0.0034070478286594152, + 0.003994878847151995, + 0.0021143758203834295, + 0.012096037156879902, + 0.013260615058243275, + 0.002091050148010254, + 0.002333681331947446, + 0.0073800282552838326, + 0.00576955359429121, + 0.008795354515314102, + 0.006066822446882725, + 0.01925729773938656, + 0.030213134363293648, + 0.005674111656844616, + 0.007143983617424965, + 0.04074283689260483, + 0.04692515358328819, + 0.05464967340230942, + 0.03297648951411247, + 0.046668391674757004, + 0.050914205610752106, + 0.03106686659157276, + 0.07617257535457611, + 0.0038160374388098717, + 0.006553920917212963, + 0.006596686318516731, + 0.006688855588436127, + 0.006978844292461872, + 0.005899250973016024, + 0.005720535293221474, + 0.006789186038076878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.019711211323738098 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790409967303276 + ], + "opposite_logprob_tokens": [ + -2.579041004180908 + ], + "logprob_reward": -0.0790409967303276, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790409967303276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07315845787525177, + 0.10260148346424103, + 0.058817993849515915, + 0.06067731976509094, + 0.03833327069878578, + 0.02694372460246086, + 0.032438974827528, + 0.04882526770234108, + 0.5006416440010071, + 0.05756179243326187 + ], + "credit_map": [ + 0.007197329308837652, + 0.008533257059752941, + 0.009072481654584408, + 0.010220288299024105, + 0.008799889124929905, + 0.009684624150395393, + 0.012057548388838768, + 0.007593052927404642, + 0.007273033261299133, + 0.009766662493348122, + 0.00833776406943798, + 0.0056334808468818665, + 0.024892065674066544, + 0.021802866831421852, + 0.008102825842797756, + 0.016792792826890945, + 0.003496752120554447, + 0.010429839603602886, + 0.0051393150351941586, + 0.006312832236289978, + 0.008708818815648556, + 0.006872527301311493, + 0.008873647078871727, + 0.00898426678031683, + 0.005457314196974039, + 0.006972287315875292, + 0.0039014280773699284, + 0.0058995806612074375, + 0.011253155767917633, + 0.012682925909757614, + 0.006961312610656023, + 0.007549320347607136, + 0.003805173560976982, + 0.004734883550554514, + 0.0042625851929187775, + 0.004720635246485472, + 0.005387346725910902, + 0.0054717944003641605, + 0.003851864719763398, + 0.0060989889316260815, + 0.0015538545558229089, + 0.0037405528128147125, + 0.002864236943423748, + 0.0038751941174268723, + 0.004717488773167133, + 0.004502493888139725, + 0.0021542333997786045, + 0.0035356711596250534, + 0.0028515153098851442, + 0.003704761853441596, + 0.004591995384544134, + 0.0024420134723186493, + 0.005747444462031126, + 0.006509372964501381, + 0.003544992534443736, + 0.0030468828044831753, + 0.0033002139534801245, + 0.002965570194646716, + 0.006618213374167681, + 0.004010058473795652, + 0.00947059690952301, + 0.013569698669016361, + 0.0036226992961019278, + 0.005268222186714411, + 0.05172001197934151, + 0.06507854163646698, + 0.07230553030967712, + 0.04578306898474693, + 0.06041083484888077, + 0.060573428869247437, + 0.04514220729470253, + 0.0996280387043953, + 0.004974001552909613, + 0.006682633422315121, + 0.00735319871455431, + 0.007182675879448652, + 0.008182322606444359, + 0.007337267976254225, + 0.006916805170476437, + 0.008932890370488167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.0790409967303276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058989740908146 + ], + "opposite_logprob_tokens": [ + -3.895589828491211 + ], + "logprob_reward": -0.02058989740908146, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058989740908146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06277766823768616, + 0.08163738250732422, + 0.1629209965467453, + 0.16493922472000122, + 0.05983887240290642, + 0.06794004887342453, + 0.12964002788066864, + 0.031692277640104294, + 0.2068464756011963, + 0.03176701068878174 + ], + "credit_map": [ + 0.00577103067189455, + 0.009141521528363228, + 0.006515567656606436, + 0.005600068252533674, + 0.013923801481723785, + 0.00907361414283514, + 0.006072554271668196, + 0.006679511163383722, + 0.006860910914838314, + 0.006884525530040264, + 0.013707487843930721, + 0.006224523298442364, + 0.011587539687752724, + 0.00989361759275198, + 0.008423218503594398, + 0.018055561929941177, + 0.011547368951141834, + 0.01573359966278076, + 0.010016885586082935, + 0.018554812297225, + 0.024775587022304535, + 0.015321286395192146, + 0.032663289457559586, + 0.034308165311813354, + 0.011666493490338326, + 0.018785929307341576, + 0.009170684032142162, + 0.014514835551381111, + 0.0266701839864254, + 0.03020581603050232, + 0.024303806945681572, + 0.02962147258222103, + 0.008396472781896591, + 0.005321783013641834, + 0.007722849026322365, + 0.006469771731644869, + 0.007309736218303442, + 0.007786646950989962, + 0.0057457974180579185, + 0.011085815727710724, + 0.004470774903893471, + 0.005985157564282417, + 0.006052485667169094, + 0.006829403340816498, + 0.011879866011440754, + 0.013668450526893139, + 0.0061318534426391125, + 0.012922055087983608, + 0.0044510336592793465, + 0.004865940660238266, + 0.003176196711137891, + 0.00412680534645915, + 0.050155382603406906, + 0.04881080985069275, + 0.006689358968287706, + 0.007364499848335981, + 0.002612351207062602, + 0.002911290153861046, + 0.0027295523323118687, + 0.002486054552718997, + 0.006294219288975, + 0.01024594809859991, + 0.0020093226339668036, + 0.0024035403039306402, + 0.023717552423477173, + 0.026640253141522408, + 0.03195672109723091, + 0.02026675082743168, + 0.024766897782683372, + 0.022895390167832375, + 0.01926461048424244, + 0.03733830153942108, + 0.0026890661101788282, + 0.004056820645928383, + 0.0036991352681070566, + 0.003685277421027422, + 0.004369676578789949, + 0.004065636079758406, + 0.0045104664750397205, + 0.00469093257561326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.02058989740908146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761953592300415 + ], + "opposite_logprob_tokens": [ + -0.8261953592300415 + ], + "logprob_reward": -0.5761953592300415, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761953592300415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05234505608677864, + 0.29219356179237366, + 0.03060895949602127, + 0.08454574644565582, + 0.04090065881609917, + 0.03167951852083206, + 0.07313387840986252, + 0.022436020895838737, + 0.33305320143699646, + 0.03910338506102562 + ], + "credit_map": [ + 0.005057063419371843, + 0.0057355621829628944, + 0.007236349396407604, + 0.006926980800926685, + 0.0062402281910181046, + 0.006257736124098301, + 0.008388437330722809, + 0.006502698175609112, + 0.02090114913880825, + 0.01843324303627014, + 0.036418844014406204, + 0.01518174260854721, + 0.08741933852434158, + 0.06429515779018402, + 0.012722543440759182, + 0.03682154789566994, + 0.0021539328154176474, + 0.005256053060293198, + 0.002793576568365097, + 0.0032223102170974016, + 0.005757411476224661, + 0.004241561982780695, + 0.004327886272221804, + 0.0028562273364514112, + 0.006610475946217775, + 0.009922148659825325, + 0.004805669188499451, + 0.007369583938270807, + 0.015819555148482323, + 0.019482972100377083, + 0.010747954249382019, + 0.009787389077246189, + 0.0046740020625293255, + 0.006093931384384632, + 0.004402156453579664, + 0.0049170698039233685, + 0.00633519422262907, + 0.004367383662611246, + 0.0037498073652386665, + 0.006361112929880619, + 0.0024451403878629208, + 0.0035658718552440405, + 0.0021347629372030497, + 0.0027878431137651205, + 0.006980346515774727, + 0.008498555980622768, + 0.002032586606219411, + 0.0032344122882932425, + 0.003926489967852831, + 0.0034360303543508053, + 0.002925131469964981, + 0.002244379138574004, + 0.02810201793909073, + 0.0250397976487875, + 0.004507088102400303, + 0.002952949609607458, + 0.0015617914032191038, + 0.0019239437533542514, + 0.0032039745710790157, + 0.002066457411274314, + 0.003965272102504969, + 0.005558613687753677, + 0.0018972420366480947, + 0.0022587240673601627, + 0.03408941999077797, + 0.04259787127375603, + 0.050312407314777374, + 0.03131651133298874, + 0.04023318737745285, + 0.037655748426914215, + 0.030949752777814865, + 0.06589829176664352, + 0.0033697180915623903, + 0.0044363620691001415, + 0.0048485868610441685, + 0.004761877469718456, + 0.005597712937742472, + 0.004940420389175415, + 0.005236332770437002, + 0.005912376567721367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.5761953592300415 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029820770025253296 + ], + "opposite_logprob_tokens": [ + -3.529820680618286 + ], + "logprob_reward": -0.029820770025253296, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029820770025253296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10927410423755646, + 0.19962486624717712, + 0.048984456807374954, + 0.08237297832965851, + 0.035925690084695816, + 0.035209864377975464, + 0.15104041993618011, + 0.03860606625676155, + 0.15680353343486786, + 0.14215797185897827 + ], + "credit_map": [ + 0.010409076698124409, + 0.013272508978843689, + 0.013573131524026394, + 0.01118085253983736, + 0.018887925893068314, + 0.015519095584750175, + 0.01541480515152216, + 0.01101671066135168, + 0.019308745861053467, + 0.01437836792320013, + 0.03586721047759056, + 0.011213768273591995, + 0.05356232449412346, + 0.03658643364906311, + 0.008850710466504097, + 0.019857287406921387, + 0.0035276913549751043, + 0.006901795044541359, + 0.004386704880744219, + 0.006002713926136494, + 0.006356450263410807, + 0.005254206247627735, + 0.009547367691993713, + 0.007007525768131018, + 0.009809538722038269, + 0.010698274709284306, + 0.006439125165343285, + 0.0078175263479352, + 0.010780610144138336, + 0.011330717243254185, + 0.01542600430548191, + 0.010071180760860443, + 0.004415702540427446, + 0.004088021349161863, + 0.004324879962950945, + 0.003568545915186405, + 0.005546363070607185, + 0.006337693426758051, + 0.003329468658193946, + 0.0043150135315954685, + 0.00301577290520072, + 0.00345050566829741, + 0.0030635602306574583, + 0.003409719793125987, + 0.007677610032260418, + 0.008884815499186516, + 0.0029906602576375008, + 0.0027172237168997526, + 0.007416137959808111, + 0.00899420864880085, + 0.007984369061887264, + 0.005175426136702299, + 0.053797513246536255, + 0.05551653355360031, + 0.005401575937867165, + 0.006754660978913307, + 0.0038513285107910633, + 0.003796046134084463, + 0.005886053666472435, + 0.0033580115996301174, + 0.006673939526081085, + 0.009074465371668339, + 0.002188653452321887, + 0.0037775696255266666, + 0.015687255188822746, + 0.021219410002231598, + 0.023062987253069878, + 0.014659770764410496, + 0.01851310208439827, + 0.017238058149814606, + 0.014910494908690453, + 0.031512461602687836, + 0.017483895644545555, + 0.012199072167277336, + 0.02313624881207943, + 0.022079577669501305, + 0.017850514501333237, + 0.018345464020967484, + 0.010063309222459793, + 0.020999902859330177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -0.029820770025253296 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, taken from a horizontal angle, focuses on a hospital bed covered in white sheets. A male patient, with a sallow complexion and chapped lips, lies on it. He wears a loose, wrinkled hospital gown with light blue stripes, and his sparse, short black hair is matted to his forehead by cold sweat. Suddenly, a violent, rapid coughing fit seizes him. His chest heaves violently, his eyes squeeze shut in agony, and his brow knits together. With each heart-wrenching cough, his body involuntarily curls inward, his back arching in agony like a shrimp. One hand weakly covers his mouth, while the other clutches the sheets beneath him, his knuckles turning white from the strain. The cold, white light from the fluorescent ceiling lights pours down relentlessly, making his pale skin appear even more pale and casting faint yet sharp shadows on his twitching body and the dishevelled sheets. The entire scene, set against the monotonous pale green wall, is permeated with a suffocating atmosphere of pain, vulnerability, and helplessness.", + "seed": 7519, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_10/index_8135_seed_7519_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14748194850981236, + "vlm_logprob": -0.14748194850981236, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.14748194850981236, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02141414023935795 + ], + "opposite_logprob_tokens": [ + -4.021414279937744 + ], + "logprob_reward": -0.02141414023935795, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02141414023935795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12920692563056946, + 0.11346778273582458, + 0.08616887032985687, + 0.14261576533317566, + 0.036965303122997284, + 0.07508205622434616, + 0.06558594852685928, + 0.20171724259853363, + 0.06513303518295288, + 0.08405705541372299 + ], + "credit_map": [ + 0.010508567094802856, + 0.01811853237450123, + 0.015569792129099369, + 0.017505336552858353, + 0.0163153987377882, + 0.020089516416192055, + 0.017048802226781845, + 0.014050975441932678, + 0.006023750174790621, + 0.009993905201554298, + 0.014864069409668446, + 0.011568126268684864, + 0.028553279116749763, + 0.016947725787758827, + 0.013776285573840141, + 0.01174064353108406, + 0.009187959134578705, + 0.015180673450231552, + 0.012515890412032604, + 0.0071794819086790085, + 0.013373383320868015, + 0.012797894887626171, + 0.008765090256929398, + 0.007168490905314684, + 0.0199973713606596, + 0.023771537467837334, + 0.01837010681629181, + 0.019204147160053253, + 0.013475370593369007, + 0.018978390842676163, + 0.015590310096740723, + 0.013228527270257473, + 0.0031416069250553846, + 0.004534765146672726, + 0.005117514636367559, + 0.004523742478340864, + 0.004825262352824211, + 0.005364974960684776, + 0.005245818290859461, + 0.004211616236716509, + 0.007971901446580887, + 0.009584498591721058, + 0.0101719219237566, + 0.012433723546564579, + 0.008253311738371849, + 0.009775691665709019, + 0.008441140875220299, + 0.00844987016171217, + 0.007333195768296719, + 0.004920370411127806, + 0.007591546513140202, + 0.00981112103909254, + 0.007366794161498547, + 0.008027984760701656, + 0.011712788604199886, + 0.00882214680314064, + 0.01687639392912388, + 0.02146914415061474, + 0.026597503572702408, + 0.030714044347405434, + 0.042441029101610184, + 0.025715811178088188, + 0.019611850380897522, + 0.018291454762220383, + 0.00447159493342042, + 0.007580715697258711, + 0.008015922270715237, + 0.00774016510695219, + 0.009037487208843231, + 0.009103690274059772, + 0.010654565878212452, + 0.008528892882168293, + 0.008626138791441917, + 0.009380963630974293, + 0.01248884666711092, + 0.008285504765808582, + 0.01377278845757246, + 0.01112541276961565, + 0.012439976446330547, + 0.007937425747513771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.02141414023935795 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500473454594612 + ], + "opposite_logprob_tokens": [ + -2.930004835128784 + ], + "logprob_reward": -0.05500473454594612, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500473454594612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08927209675312042, + 0.07527636736631393, + 0.05669400840997696, + 0.20289801061153412, + 0.025988340377807617, + 0.03261226788163185, + 0.08945545554161072, + 0.13287483155727386, + 0.048879362642765045, + 0.24604924023151398 + ], + "credit_map": [ + 0.015272147953510284, + 0.012977338396012783, + 0.009710315614938736, + 0.010069573298096657, + 0.014978041872382164, + 0.012551643885672092, + 0.008187406696379185, + 0.0055256313644349575, + 0.005325217265635729, + 0.007604437414556742, + 0.009647633880376816, + 0.009261641651391983, + 0.011478066444396973, + 0.010279633104801178, + 0.011708179488778114, + 0.009971552528440952, + 0.006072099320590496, + 0.005052641499787569, + 0.006019580643624067, + 0.004044892732053995, + 0.014368883334100246, + 0.009019278921186924, + 0.006141405086964369, + 0.005975227802991867, + 0.021189497783780098, + 0.01997552625834942, + 0.025959232822060585, + 0.044062379747629166, + 0.017352089285850525, + 0.028424503281712532, + 0.019952639937400818, + 0.025982147082686424, + 0.0026558830868452787, + 0.0024181464686989784, + 0.0038358878809958696, + 0.0025395562406629324, + 0.0029178147669881582, + 0.0033270095009356737, + 0.0037897403817623854, + 0.004504302050918341, + 0.0029728165827691555, + 0.0036069948691874743, + 0.006043366156518459, + 0.005771227180957794, + 0.0037894845008850098, + 0.0045090243220329285, + 0.0029754287097603083, + 0.002943924628198147, + 0.008884848095476627, + 0.007753155659884214, + 0.010654240846633911, + 0.017937328666448593, + 0.00780072808265686, + 0.01010267436504364, + 0.016947532072663307, + 0.00937495194375515, + 0.011511784978210926, + 0.016485055908560753, + 0.01850564405322075, + 0.020174909383058548, + 0.025976842269301414, + 0.017827019095420837, + 0.01175898127257824, + 0.0106345871463418, + 0.00366791314445436, + 0.005825574975460768, + 0.00930802058428526, + 0.004270290955901146, + 0.008229030296206474, + 0.008211756125092506, + 0.0048463051207363605, + 0.004520470276474953, + 0.02297152765095234, + 0.027655744925141335, + 0.03870664909482002, + 0.02623101696372032, + 0.044850628823041916, + 0.03334823623299599, + 0.03011992573738098, + 0.022165507078170776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 7 + ], + "selected_reward": -0.05500473454594612 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029807578772306442 + ], + "opposite_logprob_tokens": [ + -3.5298075675964355 + ], + "logprob_reward": -0.029807578772306442, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807578772306442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13733793795108795, + 0.04896268993616104, + 0.07647909969091415, + 0.10896013677120209, + 0.17407478392124176, + 0.11818131804466248, + 0.09318767488002777, + 0.05854322388768196, + 0.01812174916267395, + 0.16615135967731476 + ], + "credit_map": [ + 0.011264867149293423, + 0.011576378718018532, + 0.016348570585250854, + 0.016324514523148537, + 0.01651548594236374, + 0.02011682465672493, + 0.022160042077302933, + 0.02303125709295273, + 0.003791381139308214, + 0.005170329939574003, + 0.006697304081171751, + 0.006063312292098999, + 0.006861495785415173, + 0.007446242496371269, + 0.007125681731849909, + 0.005806946195662022, + 0.006839829497039318, + 0.0077865757048130035, + 0.009847868233919144, + 0.009574202820658684, + 0.011084861122071743, + 0.010592048987746239, + 0.010507973842322826, + 0.010245748795568943, + 0.009683873504400253, + 0.011890798807144165, + 0.014378219842910767, + 0.021975921466946602, + 0.009020423516631126, + 0.014476046897470951, + 0.011224708519876003, + 0.016310155391693115, + 0.014378146268427372, + 0.019243711605668068, + 0.02588251419365406, + 0.021316275000572205, + 0.01885654591023922, + 0.02338077314198017, + 0.02550904080271721, + 0.025507794693112373, + 0.0077872443944215775, + 0.018675262108445168, + 0.017478132620453835, + 0.020013194531202316, + 0.0105172423645854, + 0.017725905403494835, + 0.01328365784138441, + 0.012700687162578106, + 0.009187263436615467, + 0.006354481913149357, + 0.009707588702440262, + 0.017128881067037582, + 0.006848013494163752, + 0.011281176470220089, + 0.02114163152873516, + 0.01153864711523056, + 0.005688839126378298, + 0.00647966843098402, + 0.009140538051724434, + 0.007084620650857687, + 0.010008239187300205, + 0.007758938707411289, + 0.006626672577112913, + 0.005755711812525988, + 0.0014850993175059557, + 0.0017585884779691696, + 0.003187687834724784, + 0.0020911474712193012, + 0.002924875356256962, + 0.003004340222105384, + 0.002043970162048936, + 0.0016260416014119983, + 0.016191335394978523, + 0.018791763111948967, + 0.02548370324075222, + 0.017897434532642365, + 0.03097594901919365, + 0.021427292376756668, + 0.020381037145853043, + 0.015002859756350517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.029807578772306442 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762039422988892 + ], + "opposite_logprob_tokens": [ + -0.8262039422988892 + ], + "logprob_reward": -0.5762039422988892, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762039422988892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055067822337150574, + 0.05152284726500511, + 0.051208410412073135, + 0.1449001580476761, + 0.025787770748138428, + 0.04077911004424095, + 0.06256972998380661, + 0.04169044643640518, + 0.05203710496425629, + 0.47443652153015137 + ], + "credit_map": [ + 0.009698044508695602, + 0.006121895741671324, + 0.009120424278080463, + 0.004390168935060501, + 0.00746737327426672, + 0.009028010070323944, + 0.005513713229447603, + 0.003728193463757634, + 0.0033126622438430786, + 0.007035816553980112, + 0.005181091371923685, + 0.005761953070759773, + 0.008186903782188892, + 0.0074433451518416405, + 0.008145893923938274, + 0.006455180700868368, + 0.004991048946976662, + 0.007099404465407133, + 0.0045962086878716946, + 0.003926311619579792, + 0.012686708942055702, + 0.00778365321457386, + 0.005775933153927326, + 0.00434914231300354, + 0.014354747720062733, + 0.017700070515275, + 0.017770234495401382, + 0.02966902405023575, + 0.01181566808372736, + 0.019297292456030846, + 0.016174763441085815, + 0.01811836287379265, + 0.001989571610465646, + 0.0037033590488135815, + 0.002462689997628331, + 0.003099133027717471, + 0.0026465265545994043, + 0.0037283734418451786, + 0.004531800281256437, + 0.0036263170186430216, + 0.003963161259889603, + 0.003955803345888853, + 0.007633446250110865, + 0.007202469278126955, + 0.004207412712275982, + 0.005606698337942362, + 0.004196516703814268, + 0.004013602156192064, + 0.007307818159461021, + 0.004755414556711912, + 0.00717982416972518, + 0.012297030538320541, + 0.004305546637624502, + 0.007470929529517889, + 0.012597393244504929, + 0.006655779667198658, + 0.004042783286422491, + 0.005286644212901592, + 0.005006857682019472, + 0.007729186676442623, + 0.0062079355120658875, + 0.0056310733780264854, + 0.004095599986612797, + 0.0036903666332364082, + 0.004411161411553621, + 0.0057495697401463985, + 0.00972339790314436, + 0.004673906601965427, + 0.008668126538395882, + 0.008908124640583992, + 0.005267472472041845, + 0.004635350313037634, + 0.04498117417097092, + 0.054957687854766846, + 0.07332821935415268, + 0.05072447657585144, + 0.08847860991954803, + 0.061959631741046906, + 0.05743515118956566, + 0.042571596801280975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.5762039422988892 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549793466925621 + ], + "opposite_logprob_tokens": [ + -2.9299793243408203 + ], + "logprob_reward": -0.0549793466925621, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549793466925621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12404324114322662, + 0.09696251153945923, + 0.1197432428598404, + 0.15496620535850525, + 0.09689605236053467, + 0.034710340201854706, + 0.09122361987829208, + 0.07266265898942947, + 0.031057734042406082, + 0.1777344048023224 + ], + "credit_map": [ + 0.010002908296883106, + 0.02243836596608162, + 0.013428791426122189, + 0.013679815456271172, + 0.02356470376253128, + 0.02062964253127575, + 0.012203707359731197, + 0.008095312863588333, + 0.0067924680188298225, + 0.01383317168802023, + 0.011920013464987278, + 0.010277431458234787, + 0.0190932285040617, + 0.014753827825188637, + 0.012563546188175678, + 0.007728823460638523, + 0.010493665002286434, + 0.018592480570077896, + 0.02365245297551155, + 0.012511429376900196, + 0.01701953262090683, + 0.018995564430952072, + 0.009930949658155441, + 0.008547165431082249, + 0.013061339035630226, + 0.013745561242103577, + 0.019112322479486465, + 0.029563317075371742, + 0.017006881535053253, + 0.022223742678761482, + 0.015862245112657547, + 0.024390792474150658, + 0.009390315040946007, + 0.014693472534418106, + 0.01306071039289236, + 0.010086003690958023, + 0.013883626088500023, + 0.013798095285892487, + 0.011221560649573803, + 0.010762265883386135, + 0.002767823403701186, + 0.0036092340014874935, + 0.00626319320872426, + 0.006746333092451096, + 0.003452523145824671, + 0.005309298634529114, + 0.0029685846529901028, + 0.0035933495964854956, + 0.0141429603099823, + 0.009772375226020813, + 0.010083740577101707, + 0.012158065102994442, + 0.0148574598133564, + 0.010373753495514393, + 0.012080795131623745, + 0.007754472550004721, + 0.006841947790235281, + 0.00948413833975792, + 0.008520444855093956, + 0.010413921438157558, + 0.013335750438272953, + 0.010609263554215431, + 0.007880778051912785, + 0.005576411262154579, + 0.002691246336326003, + 0.0028582459781318903, + 0.005870107561349869, + 0.0027787508442997932, + 0.0058656297624111176, + 0.004947586916387081, + 0.003427947638556361, + 0.0026182192377746105, + 0.016570372506976128, + 0.019912635907530785, + 0.027372542768716812, + 0.018498918041586876, + 0.033631205558776855, + 0.023105982691049576, + 0.022390572354197502, + 0.0162521842867136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.0549793466925621 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, taken from a horizontal angle, focuses on a hospital bed covered in white sheets. A male patient, with a sallow complexion and chapped lips, lies on it. He wears a loose, wrinkled hospital gown with light blue stripes, and his sparse, short black hair is matted to his forehead by cold sweat. Suddenly, a violent, rapid coughing fit seizes him. His chest heaves violently, his eyes squeeze shut in agony, and his brow knits together. With each heart-wrenching cough, his body involuntarily curls inward, his back arching in agony like a shrimp. One hand weakly covers his mouth, while the other clutches the sheets beneath him, his knuckles turning white from the strain. The cold, white light from the fluorescent ceiling lights pours down relentlessly, making his pale skin appear even more pale and casting faint yet sharp shadows on his twitching body and the dishevelled sheets. The entire scene, set against the monotonous pale green wall, is permeated with a suffocating atmosphere of pain, vulnerability, and helplessness.", + "seed": 7520, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_10/index_8135_seed_7520_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.16570603009313345, + "vlm_logprob": -0.16570603009313345, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.16570603009313345, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011488589458167553 + ], + "opposite_logprob_tokens": [ + -4.761488437652588 + ], + "logprob_reward": -0.011488589458167553, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011488589458167553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13348783552646637, + 0.07937491685152054, + 0.06256796419620514, + 0.126180499792099, + 0.07468556612730026, + 0.06373744457960129, + 0.08488333225250244, + 0.10403908044099808, + 0.0772717148065567, + 0.1937715709209442 + ], + "credit_map": [ + 0.01664615422487259, + 0.016346238553524017, + 0.019000370055437088, + 0.01643550768494606, + 0.019433587789535522, + 0.018533337861299515, + 0.016638468950986862, + 0.010454187169671059, + 0.007615517359226942, + 0.009289575740695, + 0.008928848430514336, + 0.007705840282142162, + 0.010350883938372135, + 0.011662323027849197, + 0.012637162581086159, + 0.01118476688861847, + 0.003834398463368416, + 0.0044428338296711445, + 0.008014343678951263, + 0.006205166690051556, + 0.009271593764424324, + 0.010929499752819538, + 0.006145123392343521, + 0.013725005090236664, + 0.012376162223517895, + 0.019545050337910652, + 0.013171225786209106, + 0.012217363342642784, + 0.021339692175388336, + 0.02152852714061737, + 0.0130462646484375, + 0.012956228107213974, + 0.009207907132804394, + 0.005883615463972092, + 0.011210208758711815, + 0.00878724455833435, + 0.007847221568226814, + 0.011559839360415936, + 0.009046848863363266, + 0.011142686940729618, + 0.005208507180213928, + 0.008565803989768028, + 0.0046679615043103695, + 0.00590982660651207, + 0.0107570830732584, + 0.011005165986716747, + 0.00725461496040225, + 0.010368488729000092, + 0.005667704623192549, + 0.00933160912245512, + 0.009682351723313332, + 0.016977138817310333, + 0.009513878263533115, + 0.01119848620146513, + 0.009207743220031261, + 0.01330442912876606, + 0.006410125643014908, + 0.010465801693499088, + 0.011728211306035519, + 0.007808188907802105, + 0.028355777263641357, + 0.023150907829403877, + 0.006555340252816677, + 0.009564738720655441, + 0.0044878763146698475, + 0.0088807987049222, + 0.009392350912094116, + 0.011930068954825401, + 0.017639227211475372, + 0.013752233237028122, + 0.006672537885606289, + 0.004516628570854664, + 0.020149704068899155, + 0.012110345996916294, + 0.021121349185705185, + 0.01598251238465309, + 0.04149099066853523, + 0.034647732973098755, + 0.022816959768533707, + 0.02545199543237686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.011488589458167553 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903031259775162 + ], + "opposite_logprob_tokens": [ + -2.5790302753448486 + ], + "logprob_reward": -0.07903031259775162, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903031259775162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1390874981880188, + 0.06552986800670624, + 0.09280566871166229, + 0.07252871245145798, + 0.0692955031991005, + 0.04490951821208, + 0.06913292407989502, + 0.05958182364702225, + 0.07730352878570557, + 0.3098249137401581 + ], + "credit_map": [ + 0.016553139314055443, + 0.015433278866112232, + 0.020119160413742065, + 0.017071528360247612, + 0.02455846033990383, + 0.02243032492697239, + 0.013586890883743763, + 0.009334725327789783, + 0.0035061873495578766, + 0.006785741075873375, + 0.008091622032225132, + 0.006091626361012459, + 0.010235837660729885, + 0.010698475874960423, + 0.012686257250607014, + 0.007434116210788488, + 0.006699714809656143, + 0.009681868366897106, + 0.007713291794061661, + 0.02098737098276615, + 0.010538140311837196, + 0.013483259826898575, + 0.008660845458507538, + 0.015041166916489601, + 0.006647229194641113, + 0.012717973440885544, + 0.009029099717736244, + 0.009955714456737041, + 0.009146582335233688, + 0.009728849865496159, + 0.008358697406947613, + 0.006944566499441862, + 0.0075418995693326, + 0.006369737908244133, + 0.010875712148845196, + 0.006963376421481371, + 0.010356280021369457, + 0.010105074383318424, + 0.008724156767129898, + 0.008359264582395554, + 0.010051376186311245, + 0.0044306498020887375, + 0.004292161203920841, + 0.003111192723736167, + 0.0088855791836977, + 0.007242345716804266, + 0.002544427290558815, + 0.004351786337792873, + 0.005324468016624451, + 0.007310675922781229, + 0.010081672109663486, + 0.009299550205469131, + 0.009894564747810364, + 0.009739932604134083, + 0.007502497173845768, + 0.009979561902582645, + 0.0030690000858157873, + 0.0048199244774878025, + 0.005901029799133539, + 0.006204755045473576, + 0.01589599996805191, + 0.013524501584470272, + 0.00422386871650815, + 0.00594274140894413, + 0.004739169962704182, + 0.009206031449139118, + 0.008004708215594292, + 0.005873688030987978, + 0.021378066390752792, + 0.013223586603999138, + 0.006756106857210398, + 0.008122175000607967, + 0.03163803741335869, + 0.02063189446926117, + 0.029780462384223938, + 0.024737946689128876, + 0.07035957276821136, + 0.053210582584142685, + 0.03832100331783295, + 0.041145406663417816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.07903031259775162 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607470214366913 + ], + "opposite_logprob_tokens": [ + -4.1410746574401855 + ], + "logprob_reward": -0.01607470214366913, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607470214366913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09045154601335526, + 0.1147606372833252, + 0.03843814134597778, + 0.06631620228290558, + 0.13201214373111725, + 0.08024930953979492, + 0.06466300785541534, + 0.25910234451293945, + 0.048404037952423096, + 0.10560256242752075 + ], + "credit_map": [ + 0.005982325877994299, + 0.011037534102797508, + 0.010535364039242268, + 0.012349501252174377, + 0.01756996661424637, + 0.014287728816270828, + 0.010958521626889706, + 0.007730613928288221, + 0.0066963559947907925, + 0.014616564847528934, + 0.008855818770825863, + 0.011545046232640743, + 0.016735458746552467, + 0.01744239404797554, + 0.02005571871995926, + 0.018813299015164375, + 0.002897859551012516, + 0.003037587506696582, + 0.0033321527298539877, + 0.0037643599789589643, + 0.005350055638700724, + 0.006746283732354641, + 0.0044042374938726425, + 0.008905608206987381, + 0.006000852212309837, + 0.00978037528693676, + 0.007275999523699284, + 0.0077046616934239864, + 0.01169278472661972, + 0.011547340080142021, + 0.006486392579972744, + 0.005827808286994696, + 0.00829821266233921, + 0.010304595343768597, + 0.017916997894644737, + 0.0178330447524786, + 0.02384914644062519, + 0.021395351737737656, + 0.01624583825469017, + 0.016168972477316856, + 0.007856274954974651, + 0.006075366400182247, + 0.004847656469792128, + 0.004500084090977907, + 0.021232279017567635, + 0.021769313141703606, + 0.006563837639987469, + 0.007404506206512451, + 0.003893696004524827, + 0.0065214550122618675, + 0.007294197101145983, + 0.010136145167052746, + 0.009483139961957932, + 0.00845741294324398, + 0.008393650874495506, + 0.010483319871127605, + 0.01697545498609543, + 0.0178588405251503, + 0.042090997099876404, + 0.025435389950871468, + 0.05566907301545143, + 0.04739599674940109, + 0.02464515157043934, + 0.02903147041797638, + 0.0031219038646668196, + 0.005569084547460079, + 0.004632904659956694, + 0.005418669432401657, + 0.01305313315242529, + 0.00836746022105217, + 0.004022843670099974, + 0.0042180404998362064, + 0.011315342970192432, + 0.007938671857118607, + 0.013351446948945522, + 0.01029941439628601, + 0.018409546464681625, + 0.017043398693203926, + 0.01461695320904255, + 0.012627809308469296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.01607470214366913 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328490972518921 + ], + "opposite_logprob_tokens": [ + -0.7578490972518921 + ], + "logprob_reward": -0.6328490972518921, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328490972518921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0635296180844307, + 0.0400337316095829, + 0.1209673285484314, + 0.04807397350668907, + 0.043167371302843094, + 0.013939312659204006, + 0.11026659607887268, + 0.12878654897212982, + 0.0610053688287735, + 0.37023019790649414 + ], + "credit_map": [ + 0.006908038165420294, + 0.006595910526812077, + 0.008027262054383755, + 0.00561411352828145, + 0.010736743919551373, + 0.00937245786190033, + 0.009016099385917187, + 0.007258993107825518, + 0.0033558134455233812, + 0.004077954683452845, + 0.006903097033500671, + 0.0034040638711303473, + 0.006944307126104832, + 0.006681634113192558, + 0.005288139451295137, + 0.003378723282366991, + 0.009883017279207706, + 0.011774694547057152, + 0.010833482258021832, + 0.024648530408740044, + 0.014094810001552105, + 0.01749204844236374, + 0.01228166464716196, + 0.019959082826972008, + 0.004840356297791004, + 0.00572920311242342, + 0.004616848658770323, + 0.004178009461611509, + 0.00986082199960947, + 0.010829606093466282, + 0.003909057937562466, + 0.004110070411115885, + 0.004894264508038759, + 0.0038635809905827045, + 0.007786671165376902, + 0.004890158772468567, + 0.0048612505197525024, + 0.006823704577982426, + 0.005408387165516615, + 0.004639357794076204, + 0.002763610566034913, + 0.0018960536690428853, + 0.0011601154692471027, + 0.0012324098497629166, + 0.0018171247793361545, + 0.002381101716309786, + 0.0012825790327042341, + 0.0014063179260119796, + 0.008213698863983154, + 0.010857539251446724, + 0.016466373577713966, + 0.010922655463218689, + 0.015878254547715187, + 0.016464678570628166, + 0.011978480964899063, + 0.01948491856455803, + 0.007545985281467438, + 0.00782967358827591, + 0.019709838554263115, + 0.011048146523535252, + 0.03259960561990738, + 0.026377016678452492, + 0.00939435325562954, + 0.014281927607953548, + 0.0036082123406231403, + 0.0070728883147239685, + 0.006866627838462591, + 0.005792421288788319, + 0.016226952895522118, + 0.010450518690049648, + 0.00534657621756196, + 0.005641168914735317, + 0.03999485820531845, + 0.027263518422842026, + 0.04160851612687111, + 0.031854696571826935, + 0.06929808109998703, + 0.06179961934685707, + 0.04999213293194771, + 0.04841877147555351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.6328490972518921 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908744901418686 + ], + "opposite_logprob_tokens": [ + -2.46408748626709 + ], + "logprob_reward": -0.08908744901418686, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908744901418686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07655277103185654, + 0.07226436585187912, + 0.12056326866149902, + 0.03510067239403725, + 0.04202616214752197, + 0.06609170883893967, + 0.13142094016075134, + 0.09563183039426804, + 0.15857800841331482, + 0.20177024602890015 + ], + "credit_map": [ + 0.005766251124441624, + 0.009300551377236843, + 0.007450247649103403, + 0.009675436653196812, + 0.014626446180045605, + 0.011311243288218975, + 0.010693959891796112, + 0.007728635799139738, + 0.0066090854816138744, + 0.007645058445632458, + 0.0128922238945961, + 0.007252039387822151, + 0.00822637602686882, + 0.011384258046746254, + 0.01058870367705822, + 0.007666620891541243, + 0.0073107099160552025, + 0.012677193619310856, + 0.011058436706662178, + 0.027975449338555336, + 0.017101852223277092, + 0.01932379975914955, + 0.012527170591056347, + 0.012588664889335632, + 0.003110597375780344, + 0.00484763877466321, + 0.0037568293046206236, + 0.0036454140208661556, + 0.005239066202193499, + 0.005938501097261906, + 0.0033543389290571213, + 0.005208284594118595, + 0.003442569635808468, + 0.003986331168562174, + 0.00844574999064207, + 0.004018794745206833, + 0.0056032887659966946, + 0.007722032256424427, + 0.0040075648576021194, + 0.00479982839897275, + 0.010367844253778458, + 0.006469798739999533, + 0.007764589507132769, + 0.005205462221056223, + 0.011793584562838078, + 0.011019495315849781, + 0.0058721681125462055, + 0.007598763797432184, + 0.008614570833742619, + 0.011108871549367905, + 0.019430121406912804, + 0.01424258854240179, + 0.0194858331233263, + 0.0199388787150383, + 0.01586131751537323, + 0.022738764062523842, + 0.00889281090348959, + 0.012374200858175755, + 0.015714311972260475, + 0.009320667944848537, + 0.013298995792865753, + 0.014131157658994198, + 0.010686809197068214, + 0.011212870478630066, + 0.011015204712748528, + 0.01975112035870552, + 0.021227804943919182, + 0.020532026886940002, + 0.033713970333337784, + 0.02691132389008999, + 0.012688910588622093, + 0.012737646698951721, + 0.020435579121112823, + 0.015110793523490429, + 0.02633889764547348, + 0.01717069000005722, + 0.03457917273044586, + 0.03556862100958824, + 0.027182139456272125, + 0.02538435347378254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.08908744901418686 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, taken from a horizontal angle, focuses on a hospital bed covered in white sheets. A male patient, with a sallow complexion and chapped lips, lies on it. He wears a loose, wrinkled hospital gown with light blue stripes, and his sparse, short black hair is matted to his forehead by cold sweat. Suddenly, a violent, rapid coughing fit seizes him. His chest heaves violently, his eyes squeeze shut in agony, and his brow knits together. With each heart-wrenching cough, his body involuntarily curls inward, his back arching in agony like a shrimp. One hand weakly covers his mouth, while the other clutches the sheets beneath him, his knuckles turning white from the strain. The cold, white light from the fluorescent ceiling lights pours down relentlessly, making his pale skin appear even more pale and casting faint yet sharp shadows on his twitching body and the dishevelled sheets. The entire scene, set against the monotonous pale green wall, is permeated with a suffocating atmosphere of pain, vulnerability, and helplessness.", + "seed": 7521, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_10/index_8135_seed_7521_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08075274005532265, + "vlm_logprob": -0.08075274005532265, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.08075274005532265, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014288261532783508 + ], + "opposite_logprob_tokens": [ + -4.514288425445557 + ], + "logprob_reward": -0.014288261532783508, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.014288261532783508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2719244360923767, + 0.07352692633867264, + 0.05622095242142677, + 0.07604043185710907, + 0.05395720526576042, + 0.13639695942401886, + 0.04059666395187378, + 0.031304217875003815, + 0.1785896271467209, + 0.08144254237413406 + ], + "credit_map": [ + 0.02775098942220211, + 0.030421340838074684, + 0.03590003401041031, + 0.03281044587492943, + 0.037093352526426315, + 0.03097842074930668, + 0.0476078987121582, + 0.02936197817325592, + 0.00755862332880497, + 0.02182745561003685, + 0.006239228881895542, + 0.0071100578643381596, + 0.009294220246374607, + 0.005342544987797737, + 0.007884697988629341, + 0.008270097896456718, + 0.005508225876837969, + 0.006477419286966324, + 0.005104450508952141, + 0.006681380793452263, + 0.007003686390817165, + 0.009351213462650776, + 0.007655081804841757, + 0.008439489640295506, + 0.007169212214648724, + 0.021984227001667023, + 0.01507583912461996, + 0.005619936157017946, + 0.005741540342569351, + 0.007123481947928667, + 0.005385709460824728, + 0.007940485142171383, + 0.007866408675909042, + 0.005374391097575426, + 0.0058852327056229115, + 0.005694057792425156, + 0.006620101165026426, + 0.007938580587506294, + 0.005427967756986618, + 0.009150465950369835, + 0.01283276453614235, + 0.015015054494142532, + 0.010119909420609474, + 0.01648484729230404, + 0.013824268244206905, + 0.013405400328338146, + 0.029106566682457924, + 0.025608155876398087, + 0.004922252148389816, + 0.00826263427734375, + 0.00419992208480835, + 0.005600983276963234, + 0.003858964890241623, + 0.005705630872398615, + 0.004636684432625771, + 0.003409590106457472, + 0.0040430654771625996, + 0.004749627783894539, + 0.0028086365200579166, + 0.003017275594174862, + 0.004783454816788435, + 0.0041550579480826855, + 0.0033982815220952034, + 0.004348821006715298, + 0.018627405166625977, + 0.03543853387236595, + 0.01652250625193119, + 0.02088616043329239, + 0.019541598856449127, + 0.022138243541121483, + 0.029141461476683617, + 0.016293732449412346, + 0.008181224577128887, + 0.008060810156166553, + 0.008582419715821743, + 0.009601759724318981, + 0.015415201894938946, + 0.010610741563141346, + 0.009964213706552982, + 0.011026166379451752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.014288261532783508 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869656264781952 + ], + "opposite_logprob_tokens": [ + -3.048696517944336 + ], + "logprob_reward": -0.04869656264781952, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869656264781952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14193223416805267, + 0.07400275766849518, + 0.09898041933774948, + 0.12350494414567947, + 0.07518181204795837, + 0.12762416899204254, + 0.025207234546542168, + 0.047565214335918427, + 0.12959349155426025, + 0.15640772879123688 + ], + "credit_map": [ + 0.016843779012560844, + 0.02177293971180916, + 0.018660854548215866, + 0.017512362450361252, + 0.01676059141755104, + 0.016922282055020332, + 0.015961358323693275, + 0.017498070374131203, + 0.008805687539279461, + 0.020194601267576218, + 0.008426347747445107, + 0.004887424875050783, + 0.00839422270655632, + 0.006186724174767733, + 0.007625342346727848, + 0.009482408873736858, + 0.008601592853665352, + 0.012061863206326962, + 0.009226428344845772, + 0.017171863466501236, + 0.010971887037158012, + 0.013480160385370255, + 0.012854762375354767, + 0.014611858874559402, + 0.009030099958181381, + 0.04985823854804039, + 0.0346267893910408, + 0.006452053319662809, + 0.0061881812289357185, + 0.004240576643496752, + 0.006638007238507271, + 0.00647099269554019, + 0.009411756880581379, + 0.006904348265379667, + 0.01035989634692669, + 0.008035702630877495, + 0.008845191448926926, + 0.008680682629346848, + 0.009261219762265682, + 0.013683013617992401, + 0.008837629109621048, + 0.014997248537838459, + 0.007561888545751572, + 0.013512565754354, + 0.017783185467123985, + 0.01263689436018467, + 0.02694721147418022, + 0.025347543880343437, + 0.003213034011423588, + 0.005113836843520403, + 0.0025923261418938637, + 0.003669827012345195, + 0.002646726556122303, + 0.0029216445982456207, + 0.003260187339037657, + 0.0017896528588607907, + 0.005509770009666681, + 0.006340846884995699, + 0.006025377660989761, + 0.006405227351933718, + 0.006896522827446461, + 0.005145085975527763, + 0.005836603697389364, + 0.005405784118920565, + 0.013478536158800125, + 0.026683250442147255, + 0.014188054017722607, + 0.01922932080924511, + 0.011629074811935425, + 0.01158765610307455, + 0.015948588028550148, + 0.016849013045430183, + 0.013986635021865368, + 0.022379135712981224, + 0.01507929340004921, + 0.01936350204050541, + 0.03192621469497681, + 0.02008288912475109, + 0.015802918002009392, + 0.017787141725420952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.04869656264781952 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033703748136758804 + ], + "opposite_logprob_tokens": [ + -3.4087038040161133 + ], + "logprob_reward": -0.033703748136758804, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033703748136758804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.294048011302948, + 0.08380137383937836, + 0.0687870904803276, + 0.07318411767482758, + 0.02762545645236969, + 0.15475361049175262, + 0.02011869288980961, + 0.01596698723733425, + 0.18118734657764435, + 0.08052732050418854 + ], + "credit_map": [ + 0.026449114084243774, + 0.03426997363567352, + 0.039571184664964676, + 0.03299197182059288, + 0.03525972366333008, + 0.033020440489053726, + 0.058961350470781326, + 0.03352425619959831, + 0.00929163210093975, + 0.008971071802079678, + 0.013537807390093803, + 0.009049362502992153, + 0.009746760129928589, + 0.007159429602324963, + 0.014603819698095322, + 0.011441490612924099, + 0.008838891983032227, + 0.008366362191736698, + 0.006361944600939751, + 0.007635570131242275, + 0.008053518831729889, + 0.010382280685007572, + 0.008329568430781364, + 0.010818954557180405, + 0.006641079671680927, + 0.024715812876820564, + 0.015064330771565437, + 0.004728979431092739, + 0.0052617620676755905, + 0.005002245306968689, + 0.0047907051630318165, + 0.00697920611128211, + 0.0038982732221484184, + 0.002768667181953788, + 0.004779389128088951, + 0.003293032292276621, + 0.002952523296698928, + 0.0028868222143501043, + 0.0028335819952189922, + 0.00421316921710968, + 0.015059500001370907, + 0.018215546384453773, + 0.011474477127194405, + 0.020911531522870064, + 0.01984420232474804, + 0.017599444836378098, + 0.028471803292632103, + 0.02317710965871811, + 0.0016349554061889648, + 0.002097749849781394, + 0.0033351955935359, + 0.002613834338262677, + 0.003683817572891712, + 0.0022397872526198626, + 0.002848188392817974, + 0.001665164134465158, + 0.001737432088702917, + 0.001875631627626717, + 0.0019521351205185056, + 0.0023455682676285505, + 0.0018419661791995168, + 0.001645898912101984, + 0.0019759079441428185, + 0.0025924467481672764, + 0.014645589515566826, + 0.024862393736839294, + 0.018027162179350853, + 0.023085827007889748, + 0.02036254107952118, + 0.022199880331754684, + 0.032503265887498856, + 0.025500699877738953, + 0.006658070255070925, + 0.009782718494534492, + 0.009299029596149921, + 0.012798165902495384, + 0.010884849354624748, + 0.010407453402876854, + 0.009989936836063862, + 0.010707100853323936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.033703748136758804 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521146237850189 + ], + "opposite_logprob_tokens": [ + -1.5021146535873413 + ], + "logprob_reward": -0.2521146237850189, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521146237850189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09981057792901993, + 0.057083286345005035, + 0.12398038059473038, + 0.17028120160102844, + 0.11768733710050583, + 0.18337975442409515, + 0.03317950293421745, + 0.018055161461234093, + 0.09190059453248978, + 0.10464221984148026 + ], + "credit_map": [ + 0.011103441938757896, + 0.014715873636305332, + 0.01377030648291111, + 0.014417055994272232, + 0.011016225442290306, + 0.008789923042058945, + 0.014850983396172523, + 0.011146767064929008, + 0.006097191479057074, + 0.01650088094174862, + 0.010236025787889957, + 0.004071710165590048, + 0.003210639813914895, + 0.004121058154851198, + 0.007510978262871504, + 0.0053347996436059475, + 0.01232847012579441, + 0.016757480800151825, + 0.014597904868423939, + 0.0217304490506649, + 0.010943997651338577, + 0.0151598509401083, + 0.014315604232251644, + 0.018146634101867676, + 0.015309094451367855, + 0.06232181191444397, + 0.03984590247273445, + 0.008744349703192711, + 0.012433809228241444, + 0.011555150151252747, + 0.009805476292967796, + 0.010265606455504894, + 0.018419671803712845, + 0.011959630064666271, + 0.01411670632660389, + 0.013382229022681713, + 0.015207179822027683, + 0.014162130653858185, + 0.016788993030786514, + 0.013650793582201004, + 0.013219425454735756, + 0.02141699381172657, + 0.011033592745661736, + 0.017420392483472824, + 0.027478918433189392, + 0.01924261823296547, + 0.039754822850227356, + 0.0338129960000515, + 0.0023571343626827, + 0.0029446128755807877, + 0.0060832262970507145, + 0.0023452872410416603, + 0.009416051208972931, + 0.004294779617339373, + 0.003168652066960931, + 0.0025697583332657814, + 0.0024455932434648275, + 0.0038846293464303017, + 0.0012500864686444402, + 0.0014114642981439829, + 0.0030984743498265743, + 0.0017796993488445878, + 0.0017056948272511363, + 0.0024795192293822765, + 0.008735530078411102, + 0.011664919555187225, + 0.007973004132509232, + 0.014602637849748135, + 0.010476388037204742, + 0.01342086773365736, + 0.011924213729798794, + 0.013103034347295761, + 0.010101052932441235, + 0.015015313401818275, + 0.010272032581269741, + 0.01166490837931633, + 0.022242607548832893, + 0.014833061955869198, + 0.00983712449669838, + 0.010676116682589054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.2521146237850189 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male patient lying in a hospital bed covered with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the patient exhibit signs of illness such as a sallow complexion, chapped lips, and sparse black hair matted by sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, wrinkled hospital gown with light blue stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the patient perform a violent coughing fit where his body curls inward, his back arches, and he clutches the sheets with white knuckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by cold white fluorescent ceiling lights against a monotonous pale green wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496050417423248 + ], + "opposite_logprob_tokens": [ + -2.9299604892730713 + ], + "logprob_reward": -0.05496050417423248, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496050417423248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19767041504383087, + 0.03445274382829666, + 0.18007978796958923, + 0.19684448838233948, + 0.02272423356771469, + 0.05289411172270775, + 0.028348984196782112, + 0.1324252486228943, + 0.04687577486038208, + 0.10768423229455948 + ], + "credit_map": [ + 0.016958333551883698, + 0.021636616438627243, + 0.029555359855294228, + 0.02321571484208107, + 0.022274356335401535, + 0.0229911170899868, + 0.032603975385427475, + 0.028434939682483673, + 0.0049110762774944305, + 0.004175164736807346, + 0.005383849143981934, + 0.003983601462095976, + 0.0033190525136888027, + 0.0036547407507896423, + 0.004366028122603893, + 0.004659229423850775, + 0.019333720207214355, + 0.025456806644797325, + 0.01939961686730385, + 0.033822689205408096, + 0.013877682387828827, + 0.019816312938928604, + 0.022633947432041168, + 0.02573901228606701, + 0.02049676701426506, + 0.07631173729896545, + 0.04085320979356766, + 0.011651420034468174, + 0.012714942917227745, + 0.012771825306117535, + 0.011738164350390434, + 0.010306425392627716, + 0.002401800360530615, + 0.003047415753826499, + 0.003182027256116271, + 0.0020844407845288515, + 0.0022676903754472733, + 0.002274740021675825, + 0.0033816045615822077, + 0.004084513522684574, + 0.0056492360308766365, + 0.0056901415809988976, + 0.004573356360197067, + 0.007808504160493612, + 0.006742472294718027, + 0.008675359189510345, + 0.00683726416900754, + 0.006917780265212059, + 0.002186791505664587, + 0.0020331458654254675, + 0.006176819559186697, + 0.0025746666360646486, + 0.007480148691684008, + 0.003166937967762351, + 0.0027325295377522707, + 0.0019979444332420826, + 0.017674662172794342, + 0.029049327597022057, + 0.008539458736777306, + 0.013597537763416767, + 0.01706850901246071, + 0.00857294537127018, + 0.013336608186364174, + 0.02458619885146618, + 0.0050781844183802605, + 0.004703945480287075, + 0.00752631202340126, + 0.004395709838718176, + 0.00547091756016016, + 0.007363615557551384, + 0.005770920775830746, + 0.0065661706030368805, + 0.012651702389121056, + 0.012609601952135563, + 0.00765632139518857, + 0.010104846209287643, + 0.012956146150827408, + 0.01544388197362423, + 0.01471939217299223, + 0.021542338654398918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.05496050417423248 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_11/index_tensor([224, 224])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_11/index_tensor([224, 224])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c5b03cff85232bc9246f75732e66726b65a26375 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_11/index_tensor([224, 224])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This dramatic low-angle close-up, the camera slowly and almost imperceptibly zooming in, focuses on the face of a stoic, classical, bearded warrior bronze statue. The scene is plunged into deep, monumental darkness, the statue's features sculpted by a single, harsh keylight from above and sideways, creating a stark contrast of shimmering highlights and deep shadows. Initially, the statue's surface is a dark, rusty bronze with a subtle metallic sheen. Then, decay sets in; a dense network of fine cracks deliberately spreads across the cold metal of the cheekbones and temples. The edges of these cracks inexorably begin to lift and curl, flaking the solid bronze into thin, brittle flakes with the texture and fragility of dried, dead leaves. These fragments of metal detach and drift downward in a slow, continuous, silent waterfall, catching the dramatic light and glittering like copper shavings before disappearing into the blackness below, revealing a rougher, darker, and corroded texture beneath. The atmosphere is one of solemn and inevitable destruction, as the monument's eternal stoicism is captured in an elegant yet melancholic state of disintegration.", + "seed": 4033, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_11/index_224_seed_4033_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04197924304753542, + "vlm_logprob": -0.04197924304753542, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.04197924304753542, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07524918019771576 + ], + "opposite_logprob_tokens": [ + -2.700249195098877 + ], + "logprob_reward": -0.07524918019771576, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07524918019771576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12698675692081451, + 0.04585225135087967, + 0.028742024675011635, + 0.05254373326897621, + 0.041454996913671494, + 0.020430324599146843, + 0.0466073676943779, + 0.1845378875732422, + 0.09411858767271042, + 0.35872602462768555 + ], + "credit_map": [ + 0.010887589305639267, + 0.01656898483633995, + 0.012233274057507515, + 0.013444693759083748, + 0.01286243461072445, + 0.02190467342734337, + 0.013674018904566765, + 0.02541111409664154, + 0.006206294056028128, + 0.005209936760365963, + 0.004229381680488586, + 0.0038043016102164984, + 0.003327804384753108, + 0.008940598927438259, + 0.004127515479922295, + 0.010006423108279705, + 0.002745238831266761, + 0.0037820476572960615, + 0.002085990272462368, + 0.003257215255871415, + 0.003503029001876712, + 0.005335367284715176, + 0.0020031784661114216, + 0.006029960233718157, + 0.005700398702174425, + 0.00710037536919117, + 0.005014091730117798, + 0.006509967613965273, + 0.0056066978722810745, + 0.008507287129759789, + 0.0053517259657382965, + 0.00875319354236126, + 0.002645632717758417, + 0.003172168042510748, + 0.004203821066766977, + 0.0032864429522305727, + 0.004118763841688633, + 0.009304001927375793, + 0.004519697744399309, + 0.01020447164773941, + 0.0017577200196683407, + 0.0020637447014451027, + 0.0011931451736018062, + 0.0018869790947064757, + 0.003969268407672644, + 0.0035976425278931856, + 0.0018747147405520082, + 0.004087111447006464, + 0.004048953764140606, + 0.005142838228493929, + 0.0037460203748196363, + 0.006407593842595816, + 0.00523958308622241, + 0.006759011186659336, + 0.007418552413582802, + 0.007844817824661732, + 0.02130320854485035, + 0.03650470823049545, + 0.012121180072426796, + 0.019202137365937233, + 0.02677292935550213, + 0.024142896756529808, + 0.024512389674782753, + 0.019978463649749756, + 0.00816256646066904, + 0.010875875130295753, + 0.0063177128322422504, + 0.008638220839202404, + 0.006564549170434475, + 0.020262284204363823, + 0.007964422926306725, + 0.025332961231470108, + 0.029358111321926117, + 0.032398760318756104, + 0.061962470412254333, + 0.05151155963540077, + 0.03226904943585396, + 0.0491703599691391, + 0.05388646945357323, + 0.04816925525665283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.07524918019771576 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02638520672917366 + ], + "opposite_logprob_tokens": [ + -3.6513853073120117 + ], + "logprob_reward": -0.02638520672917366, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638520672917366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13829857110977173, + 0.044031478464603424, + 0.027859043329954147, + 0.04282272234559059, + 0.03415454179048538, + 0.012433978728950024, + 0.029706211760640144, + 0.46601051092147827, + 0.07695186138153076, + 0.1277310997247696 + ], + "credit_map": [ + 0.013606474734842777, + 0.018413716927170753, + 0.014376449398696423, + 0.013788159936666489, + 0.015596842393279076, + 0.024195315316319466, + 0.013743211515247822, + 0.024578386917710304, + 0.003660131013020873, + 0.004280077759176493, + 0.004516856279224157, + 0.004517430439591408, + 0.003967428579926491, + 0.008778560906648636, + 0.005319541785866022, + 0.008991449140012264, + 0.003081040922552347, + 0.0041215079836547375, + 0.0018471925286576152, + 0.00246546883136034, + 0.0029701131861656904, + 0.005838364828377962, + 0.0019860269967466593, + 0.0055493260733783245, + 0.004035906866192818, + 0.006294528488069773, + 0.004371582064777613, + 0.005612247623503208, + 0.004528529942035675, + 0.006450606510043144, + 0.005345515441149473, + 0.006183796562254429, + 0.0038109717424958944, + 0.0038847457617521286, + 0.0031939209438860416, + 0.0036713313311338425, + 0.0037410135846585035, + 0.006827518343925476, + 0.002849749056622386, + 0.006175285670906305, + 0.0013314519310370088, + 0.001465212437324226, + 0.0013424172066152096, + 0.00184885342605412, + 0.0017450075829401612, + 0.0017466943245381117, + 0.001510567031800747, + 0.0014437745558097959, + 0.0026204108726233244, + 0.003958502318710089, + 0.00372896040789783, + 0.0038343307096511126, + 0.0028814298566430807, + 0.004004211630672216, + 0.0044032796286046505, + 0.0042750840075314045, + 0.05860836058855057, + 0.0988934338092804, + 0.029314925894141197, + 0.045024026185274124, + 0.07153508067131042, + 0.06002763658761978, + 0.05551910400390625, + 0.04708795249462128, + 0.006307054776698351, + 0.006862086709588766, + 0.004954980686306953, + 0.00783755723387003, + 0.0061671850271523, + 0.017420146614313126, + 0.007561656646430492, + 0.019841192290186882, + 0.009129107929766178, + 0.015444018878042698, + 0.015239829197525978, + 0.01149902492761612, + 0.011008958332240582, + 0.02315327525138855, + 0.013238410465419292, + 0.029018469154834747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.02638520672917366 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899285852909088 + ], + "opposite_logprob_tokens": [ + -2.5789928436279297 + ], + "logprob_reward": -0.07899285852909088, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899285852909088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08851153403520584, + 0.03407618775963783, + 0.039215296506881714, + 0.026741795241832733, + 0.037285055965185165, + 0.04901844635605812, + 0.15793141722679138, + 0.1781676709651947, + 0.3051299452781677, + 0.08392263948917389 + ], + "credit_map": [ + 0.008845980279147625, + 0.011338007636368275, + 0.008702416904270649, + 0.00780831091105938, + 0.009838460944592953, + 0.01628587394952774, + 0.0084223048761487, + 0.017270175740122795, + 0.0048309252597391605, + 0.0038724509067833424, + 0.003703452879562974, + 0.004329059273004532, + 0.004244862589985132, + 0.004951520822942257, + 0.003445760579779744, + 0.004698154516518116, + 0.0029066004790365696, + 0.0035348127130419016, + 0.002985328435897827, + 0.003732808632776141, + 0.0031464400235563517, + 0.009044921956956387, + 0.004070158116519451, + 0.009794224984943867, + 0.002609137911349535, + 0.004460378084331751, + 0.00228638737462461, + 0.0032347203232347965, + 0.003417125204578042, + 0.004004376940429211, + 0.002603324595838785, + 0.004126346670091152, + 0.0031913351267576218, + 0.004053860437124968, + 0.0034358438570052385, + 0.004187627229839563, + 0.004583839792758226, + 0.007466928102076054, + 0.0031982327345758677, + 0.007167387753725052, + 0.005910387262701988, + 0.007136881817132235, + 0.004142556805163622, + 0.00847206637263298, + 0.006544369272887707, + 0.007430363912135363, + 0.0037118918262422085, + 0.005669931415468454, + 0.013844109140336514, + 0.021099630743265152, + 0.00837419182062149, + 0.015367673709988594, + 0.016317317262291908, + 0.028495697304606438, + 0.020840417593717575, + 0.03359238803386688, + 0.021052220836281776, + 0.03370209038257599, + 0.011478407308459282, + 0.02060066908597946, + 0.025851121172308922, + 0.02243904583156109, + 0.02440417744219303, + 0.018639925867319107, + 0.032323841005563736, + 0.03886663168668747, + 0.026039492338895798, + 0.03292885795235634, + 0.025322867557406425, + 0.050769444555044174, + 0.04023922607302666, + 0.058639585971832275, + 0.008412408642470837, + 0.0060988194309175014, + 0.01353411003947258, + 0.007412319537252188, + 0.007962878793478012, + 0.0123405447229743, + 0.015173466876149178, + 0.012988083995878696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.07899285852909088 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060776576399803 + ], + "opposite_logprob_tokens": [ + -3.8956077098846436 + ], + "logprob_reward": -0.02060776576399803, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060776576399803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11745714396238327, + 0.039104919880628586, + 0.08026324212551117, + 0.0576634518802166, + 0.052583590149879456, + 0.06472443789243698, + 0.11040237545967102, + 0.2524261772632599, + 0.06707488000392914, + 0.1582997590303421 + ], + "credit_map": [ + 0.01342565193772316, + 0.012132538482546806, + 0.008421199396252632, + 0.008168281987309456, + 0.010332243517041206, + 0.026597706601023674, + 0.009217895567417145, + 0.029161641374230385, + 0.005176615435630083, + 0.004595285281538963, + 0.0040148464031517506, + 0.00369973573833704, + 0.0029519405215978622, + 0.007277330383658409, + 0.0035809134133160114, + 0.007808254566043615, + 0.006245660595595837, + 0.008116990327835083, + 0.008739751763641834, + 0.010140499100089073, + 0.008113227784633636, + 0.014490395784378052, + 0.008759642019867897, + 0.015657080337405205, + 0.006655162665992975, + 0.007319384254515171, + 0.004766233265399933, + 0.006363227032124996, + 0.00829789973795414, + 0.008941863663494587, + 0.006560249719768763, + 0.008759432472288609, + 0.004583774600178003, + 0.004986390005797148, + 0.005477814935147762, + 0.004854145925492048, + 0.005838972516357899, + 0.009066376835107803, + 0.007098739966750145, + 0.010677381418645382, + 0.006197868846356869, + 0.006209103856235743, + 0.005574812181293964, + 0.005971909966319799, + 0.008325744420289993, + 0.012114088982343674, + 0.005886637605726719, + 0.014444282278418541, + 0.010442187078297138, + 0.016714124009013176, + 0.012913537211716175, + 0.017347631976008415, + 0.011839020065963268, + 0.01626187562942505, + 0.010680494830012321, + 0.014203513041138649, + 0.024191034957766533, + 0.04384951293468475, + 0.02613641507923603, + 0.022954806685447693, + 0.04534745588898659, + 0.03338601812720299, + 0.025785062462091446, + 0.030775904655456543, + 0.007330277469009161, + 0.007303397171199322, + 0.006374749355018139, + 0.007252037059515715, + 0.006421471014618874, + 0.010417807847261429, + 0.010495034046471119, + 0.011480116285383701, + 0.008153514005243778, + 0.014237836003303528, + 0.026070116087794304, + 0.015819644555449486, + 0.014346056617796421, + 0.0254941675812006, + 0.02155950479209423, + 0.032618921250104904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.02060776576399803 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008661204017698765 + ], + "opposite_logprob_tokens": [ + -4.758661270141602 + ], + "logprob_reward": -0.008661204017698765, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008661204017698765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1198829934000969, + 0.0330970361828804, + 0.07598785310983658, + 0.053988017141819, + 0.049996837973594666, + 0.02873375080525875, + 0.09127683937549591, + 0.2015097290277481, + 0.21095168590545654, + 0.13457533717155457 + ], + "credit_map": [ + 0.01568235456943512, + 0.011914103291928768, + 0.0097578763961792, + 0.009207102470099926, + 0.012872067280113697, + 0.024531954899430275, + 0.00916333869099617, + 0.02675417810678482, + 0.0032931703608483076, + 0.004320436157286167, + 0.0036498247645795345, + 0.0033113991376012564, + 0.003227124223485589, + 0.00558744790032506, + 0.003682496026158333, + 0.006025134585797787, + 0.0062344856560230255, + 0.008277395740151405, + 0.008746009320020676, + 0.010170456022024155, + 0.007428053766489029, + 0.012950158677995205, + 0.008593667298555374, + 0.013587618246674538, + 0.005921187344938517, + 0.0065079545602202415, + 0.0037958815228194, + 0.006013598758727312, + 0.007413624785840511, + 0.009326179511845112, + 0.005532666575163603, + 0.009476919658482075, + 0.0046943374909460545, + 0.00564266461879015, + 0.005237376783043146, + 0.005142335779964924, + 0.005512649659067392, + 0.008206591941416264, + 0.005941653158515692, + 0.009619222022593021, + 0.00263094250112772, + 0.002232697792351246, + 0.0022739432752132416, + 0.002413785783573985, + 0.0035617006942629814, + 0.00584725197404623, + 0.0024888701736927032, + 0.007284556049853563, + 0.00923176296055317, + 0.013091549277305603, + 0.010274054482579231, + 0.012856654822826385, + 0.009897678159177303, + 0.013529807329177856, + 0.009681998752057552, + 0.012713323347270489, + 0.0258385818451643, + 0.03765716031193733, + 0.013081186451017857, + 0.021094724535942078, + 0.027325397357344627, + 0.026781829074025154, + 0.027450542896986008, + 0.02228030376136303, + 0.021487277001142502, + 0.02387380599975586, + 0.01640470325946808, + 0.023687392473220825, + 0.018765384331345558, + 0.037758298218250275, + 0.02537512220442295, + 0.04359967261552811, + 0.007807609625160694, + 0.013532302342355251, + 0.017278460785746574, + 0.014336245134472847, + 0.00942875538021326, + 0.023387925699353218, + 0.012967183254659176, + 0.0358368419110775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.008661204017698765 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This dramatic low-angle close-up, the camera slowly and almost imperceptibly zooming in, focuses on the face of a stoic, classical, bearded warrior bronze statue. The scene is plunged into deep, monumental darkness, the statue's features sculpted by a single, harsh keylight from above and sideways, creating a stark contrast of shimmering highlights and deep shadows. Initially, the statue's surface is a dark, rusty bronze with a subtle metallic sheen. Then, decay sets in; a dense network of fine cracks deliberately spreads across the cold metal of the cheekbones and temples. The edges of these cracks inexorably begin to lift and curl, flaking the solid bronze into thin, brittle flakes with the texture and fragility of dried, dead leaves. These fragments of metal detach and drift downward in a slow, continuous, silent waterfall, catching the dramatic light and glittering like copper shavings before disappearing into the blackness below, revealing a rougher, darker, and corroded texture beneath. The atmosphere is one of solemn and inevitable destruction, as the monument's eternal stoicism is captured in an elegant yet melancholic state of disintegration.", + "seed": 4034, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_11/index_224_seed_4034_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.012895070854574442, + "vlm_logprob": -0.012895070854574442, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.012895070854574442, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009950321167707443 + ], + "opposite_logprob_tokens": [ + -5.759950160980225 + ], + "logprob_reward": -0.009950321167707443, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.009950321167707443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10237056761980057, + 0.286086767911911, + 0.032283294945955276, + 0.07044270634651184, + 0.05828747898340225, + 0.052012570202350616, + 0.1345812976360321, + 0.09374193847179413, + 0.09932445734739304, + 0.07086893916130066 + ], + "credit_map": [ + 0.00974587444216013, + 0.015401310287415981, + 0.013924497179687023, + 0.010544640012085438, + 0.009385907091200352, + 0.011211820878088474, + 0.012942655012011528, + 0.019213860854506493, + 0.0316828191280365, + 0.03434130921959877, + 0.040887102484703064, + 0.03180765360593796, + 0.0302259624004364, + 0.033207427710294724, + 0.03742384910583496, + 0.04651062190532684, + 0.004371593240648508, + 0.0038945714477449656, + 0.004906415939331055, + 0.004166112747043371, + 0.003156308550387621, + 0.0032808438409119844, + 0.003427912015467882, + 0.0050795357674360275, + 0.009051933884620667, + 0.007084398064762354, + 0.00985320657491684, + 0.008830561302602291, + 0.0064841569401323795, + 0.00963456928730011, + 0.007961350493133068, + 0.011542527005076408, + 0.006971390452235937, + 0.00549464114010334, + 0.007666134275496006, + 0.008836624212563038, + 0.004809880163520575, + 0.006765589118003845, + 0.00840029213577509, + 0.009342924691736698, + 0.00845950748771429, + 0.005528782028704882, + 0.008253824897110462, + 0.00549060245975852, + 0.003236535005271435, + 0.007238155696541071, + 0.006896739359945059, + 0.006908420938998461, + 0.015228073112666607, + 0.012272171676158905, + 0.015671025961637497, + 0.022260887548327446, + 0.01188626978546381, + 0.02029438316822052, + 0.017570778727531433, + 0.019397705793380737, + 0.014609822072088718, + 0.009740229696035385, + 0.013921143487095833, + 0.009640216827392578, + 0.012692652642726898, + 0.011188034899532795, + 0.009549224749207497, + 0.012400618754327297, + 0.012820007279515266, + 0.013124112971127033, + 0.01193985529243946, + 0.012985515408217907, + 0.017512377351522446, + 0.012446736916899681, + 0.01004505529999733, + 0.008450798690319061, + 0.009922183118760586, + 0.011549250222742558, + 0.009321967139840126, + 0.006505732424557209, + 0.009381020441651344, + 0.011086561717092991, + 0.0069256918504834175, + 0.006176529917865992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.009950321167707443 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0182130578905344 + ], + "opposite_logprob_tokens": [ + -4.018213272094727 + ], + "logprob_reward": -0.0182130578905344, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0182130578905344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10792054980993271, + 0.09310565143823624, + 0.07549737393856049, + 0.11234163492918015, + 0.10370922088623047, + 0.13177697360515594, + 0.0699426606297493, + 0.09109127521514893, + 0.05280398949980736, + 0.16181066632270813 + ], + "credit_map": [ + 0.011898805387318134, + 0.015547594055533409, + 0.015331776812672615, + 0.013874233700335026, + 0.008170802146196365, + 0.0128633389249444, + 0.011774418875575066, + 0.018459578976035118, + 0.010001600719988346, + 0.010337689891457558, + 0.013397356495261192, + 0.01193562988191843, + 0.01114997174590826, + 0.00995680969208479, + 0.011946857906877995, + 0.014379727654159069, + 0.008868125267326832, + 0.006988245062530041, + 0.010225831530988216, + 0.011018122546374798, + 0.009015926159918308, + 0.008610944263637066, + 0.009530123323202133, + 0.011240057647228241, + 0.013711292296648026, + 0.012807673774659634, + 0.015412670560181141, + 0.015194208361208439, + 0.009955010376870632, + 0.012451131828129292, + 0.015170900151133537, + 0.01763874478638172, + 0.012595660984516144, + 0.010008442215621471, + 0.01396346464753151, + 0.014279291033744812, + 0.006905770860612392, + 0.016358990222215652, + 0.013318625278770924, + 0.01627896912395954, + 0.023076027631759644, + 0.012181566096842289, + 0.022488394752144814, + 0.01349640917032957, + 0.011676974594593048, + 0.015162765979766846, + 0.015122511424124241, + 0.018572326749563217, + 0.006414753384888172, + 0.007601540070027113, + 0.006814875174313784, + 0.012761235237121582, + 0.006783606484532356, + 0.008719578385353088, + 0.0095065264031291, + 0.011340548284351826, + 0.009054366499185562, + 0.015912044793367386, + 0.00919827539473772, + 0.009930627420544624, + 0.013876145705580711, + 0.011784840375185013, + 0.008631808683276176, + 0.012703167274594307, + 0.004717307165265083, + 0.006630427669733763, + 0.005285064689815044, + 0.007562953047454357, + 0.009519410319626331, + 0.008452569134533405, + 0.0047541591338813305, + 0.005882099736481905, + 0.02524489350616932, + 0.021015722304582596, + 0.02491132728755474, + 0.01700734719634056, + 0.013714239932596684, + 0.026767730712890625, + 0.018503546714782715, + 0.014645867049694061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 3 + ], + "selected_reward": -0.0182130578905344 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026356063783168793 + ], + "opposite_logprob_tokens": [ + -3.6513559818267822 + ], + "logprob_reward": -0.026356063783168793, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026356063783168793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05677606537938118, + 0.12361961603164673, + 0.04734647646546364, + 0.09897764027118683, + 0.08654243499040604, + 0.1204182580113411, + 0.17843788862228394, + 0.12269379943609238, + 0.13915248215198517, + 0.026035288348793983 + ], + "credit_map": [ + 0.006183450575917959, + 0.007931015454232693, + 0.0071712033823132515, + 0.007285094819962978, + 0.004792980384081602, + 0.0056701842695474625, + 0.007629685569554567, + 0.010112454183399677, + 0.013752159662544727, + 0.014383366331458092, + 0.016844645142555237, + 0.01327597163617611, + 0.01405569538474083, + 0.015156026929616928, + 0.01627068594098091, + 0.019881075248122215, + 0.005092218518257141, + 0.004939411301165819, + 0.005750764161348343, + 0.005944640375673771, + 0.0057302070781588554, + 0.0051859705708920956, + 0.00622588163241744, + 0.008477380499243736, + 0.012747490778565407, + 0.008680248633027077, + 0.013879151083528996, + 0.014856621623039246, + 0.010518024675548077, + 0.009079482406377792, + 0.018670903518795967, + 0.010545728728175163, + 0.008666474372148514, + 0.00927957147359848, + 0.011261658743023872, + 0.012567739933729172, + 0.006040265318006277, + 0.012604392133653164, + 0.012995938770473003, + 0.013126404024660587, + 0.019225727766752243, + 0.011578826233744621, + 0.019283058121800423, + 0.010380804538726807, + 0.010328887030482292, + 0.013848105445504189, + 0.015507261268794537, + 0.020265594124794006, + 0.025711489841341972, + 0.01855667121708393, + 0.016083884984254837, + 0.01955173723399639, + 0.012954854406416416, + 0.020386379212141037, + 0.035782478749752045, + 0.02941039390861988, + 0.016818637028336525, + 0.012859226204454899, + 0.01480015553534031, + 0.012925098650157452, + 0.01638086698949337, + 0.015501530840992928, + 0.015434822998940945, + 0.01797347143292427, + 0.019414817914366722, + 0.01593927852809429, + 0.017043529078364372, + 0.01686534285545349, + 0.02611674927175045, + 0.015594985336065292, + 0.015388956293463707, + 0.012788845226168633, + 0.003318520961329341, + 0.003102642949670553, + 0.003975001163780689, + 0.0029692782554775476, + 0.0029962460976094007, + 0.0033876076340675354, + 0.002310434589162469, + 0.003975556697696447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.026356063783168793 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032035005278885365 + ], + "opposite_logprob_tokens": [ + -5.753203392028809 + ], + "logprob_reward": -0.0032035005278885365, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032035005278885365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06398709863424301, + 0.09802208840847015, + 0.042030010372400284, + 0.08692780882120132, + 0.10442396253347397, + 0.1350531131029129, + 0.05474886670708656, + 0.07916834205389023, + 0.2801242470741272, + 0.05551450327038765 + ], + "credit_map": [ + 0.007021381054073572, + 0.008317485451698303, + 0.008975379168987274, + 0.006655035074800253, + 0.005422965623438358, + 0.006465645972639322, + 0.007495212368667126, + 0.013633987866342068, + 0.01061966735869646, + 0.011754756793379784, + 0.013123683631420135, + 0.011449014768004417, + 0.010527906008064747, + 0.011589833535254002, + 0.012705654837191105, + 0.016251560300588608, + 0.004285028669983149, + 0.003957971464842558, + 0.006043759174644947, + 0.005576225463300943, + 0.004584490321576595, + 0.00614393875002861, + 0.004434091970324516, + 0.007004500366747379, + 0.010590995661914349, + 0.00652034766972065, + 0.011730032041668892, + 0.01244322769343853, + 0.009678087197244167, + 0.009689852595329285, + 0.01616613380610943, + 0.010109128430485725, + 0.01177339069545269, + 0.009540228173136711, + 0.013807835057377815, + 0.01657404936850071, + 0.0071820286102592945, + 0.017115704715251923, + 0.015398479998111725, + 0.013032231479883194, + 0.022827088832855225, + 0.01239332277327776, + 0.023337164893746376, + 0.01051913108676672, + 0.011299630627036095, + 0.01648244448006153, + 0.017140090465545654, + 0.021054232493042946, + 0.005377038381993771, + 0.0055635785683989525, + 0.00582583574578166, + 0.008656027726829052, + 0.0046279169619083405, + 0.006456850096583366, + 0.009118984453380108, + 0.009122632443904877, + 0.011181240901350975, + 0.007191907148808241, + 0.007411947939544916, + 0.008004759438335896, + 0.00872874353080988, + 0.015626629814505577, + 0.009342914447188377, + 0.011680190451443195, + 0.02299443632364273, + 0.030102301388978958, + 0.02694426290690899, + 0.034011851996183395, + 0.052081260830163956, + 0.04431317374110222, + 0.03568524867296219, + 0.03399168699979782, + 0.0046120015904307365, + 0.007965178228914738, + 0.006573813036084175, + 0.00855064857751131, + 0.007026958279311657, + 0.008579005487263203, + 0.007749946787953377, + 0.004456949885934591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.0032035005278885365 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006752410903573036 + ], + "opposite_logprob_tokens": [ + -5.0067524909973145 + ], + "logprob_reward": -0.006752410903573036, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006752410903573036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11162753403186798, + 0.058281488716602325, + 0.04116429388523102, + 0.1251528114080429, + 0.16124364733695984, + 0.076176218688488, + 0.06466110050678253, + 0.1107737347483635, + 0.12567387521266937, + 0.12524525821208954 + ], + "credit_map": [ + 0.011874441988766193, + 0.01618555374443531, + 0.016138607636094093, + 0.012161048129200935, + 0.009909006766974926, + 0.010500255972146988, + 0.012332472018897533, + 0.022526158019900322, + 0.007648014929145575, + 0.0073063126765191555, + 0.008306624367833138, + 0.00701036024838686, + 0.005465902853757143, + 0.007327865809202194, + 0.005752941593527794, + 0.009463470429182053, + 0.0037600472569465637, + 0.0048104021698236465, + 0.005168939009308815, + 0.007002050988376141, + 0.004648873116821051, + 0.004019344225525856, + 0.00574844516813755, + 0.00600619800388813, + 0.014119867235422134, + 0.011458396911621094, + 0.015144206583499908, + 0.017725596204400063, + 0.01502364594489336, + 0.013536565005779266, + 0.023233715444803238, + 0.014910820871591568, + 0.015995243564248085, + 0.015011942014098167, + 0.0196291022002697, + 0.03120001219213009, + 0.01145277637988329, + 0.02016778290271759, + 0.02685198187828064, + 0.020934823900461197, + 0.013458061963319778, + 0.006674970965832472, + 0.013497420586645603, + 0.006166233215481043, + 0.006697777658700943, + 0.007261546794325113, + 0.009313678368926048, + 0.013106531463563442, + 0.00984373688697815, + 0.0068662832491099834, + 0.00615260424092412, + 0.009198219515383244, + 0.00559349125251174, + 0.008096180856227875, + 0.0074941907078027725, + 0.011416396126151085, + 0.011146923527121544, + 0.013276174664497375, + 0.012852729298174381, + 0.011584124527871609, + 0.013294244185090065, + 0.019000744447112083, + 0.013401685282588005, + 0.016217118129134178, + 0.01473270170390606, + 0.013378085568547249, + 0.017895817756652832, + 0.014451675117015839, + 0.022602101787924767, + 0.01286349631845951, + 0.017632458359003067, + 0.012117532081902027, + 0.015035813674330711, + 0.019352169707417488, + 0.017677759751677513, + 0.014465958811342716, + 0.013502702116966248, + 0.02108212374150753, + 0.014745268039405346, + 0.009383471682667732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.006752410903573036 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This dramatic low-angle close-up, the camera slowly and almost imperceptibly zooming in, focuses on the face of a stoic, classical, bearded warrior bronze statue. The scene is plunged into deep, monumental darkness, the statue's features sculpted by a single, harsh keylight from above and sideways, creating a stark contrast of shimmering highlights and deep shadows. Initially, the statue's surface is a dark, rusty bronze with a subtle metallic sheen. Then, decay sets in; a dense network of fine cracks deliberately spreads across the cold metal of the cheekbones and temples. The edges of these cracks inexorably begin to lift and curl, flaking the solid bronze into thin, brittle flakes with the texture and fragility of dried, dead leaves. These fragments of metal detach and drift downward in a slow, continuous, silent waterfall, catching the dramatic light and glittering like copper shavings before disappearing into the blackness below, revealing a rougher, darker, and corroded texture beneath. The atmosphere is one of solemn and inevitable destruction, as the monument's eternal stoicism is captured in an elegant yet melancholic state of disintegration.", + "seed": 4035, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_11/index_224_seed_4035_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02588890539482236, + "vlm_logprob": -0.02588890539482236, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.02588890539482236, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015399482101202011 + ], + "opposite_logprob_tokens": [ + -5.01539945602417 + ], + "logprob_reward": -0.015399482101202011, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.015399482101202011, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07347158342599869, + 0.051897794008255005, + 0.014009349048137665, + 0.02115720696747303, + 0.05530634894967079, + 0.03303930535912514, + 0.11840026080608368, + 0.5086718201637268, + 0.04949231818318367, + 0.07455404847860336 + ], + "credit_map": [ + 0.0078066266141831875, + 0.005025453865528107, + 0.010482453741133213, + 0.009622441604733467, + 0.007924683392047882, + 0.011241314001381397, + 0.011808409355580807, + 0.009560206905007362, + 0.004586416762322187, + 0.006264869123697281, + 0.008444475941359997, + 0.003940664231777191, + 0.01021950040012598, + 0.00806334801018238, + 0.005087149795144796, + 0.005291374865919352, + 0.001817252952605486, + 0.0014328673714771867, + 0.0023875569459050894, + 0.0015421627322211862, + 0.001740522333420813, + 0.0016686052549630404, + 0.0018458827398717403, + 0.0015745005803182721, + 0.0021042402368038893, + 0.001977519830688834, + 0.0025999106001108885, + 0.002302213804796338, + 0.004251564387232065, + 0.0026775782462209463, + 0.0029140529222786427, + 0.002330128103494644, + 0.0052125644870102406, + 0.005453074816614389, + 0.008694043383002281, + 0.006368919741362333, + 0.006324755493551493, + 0.008375375531613827, + 0.010411269962787628, + 0.004466348793357611, + 0.0027907711919397116, + 0.0037644985131919384, + 0.002866340335458517, + 0.004618783015757799, + 0.0051718708127737045, + 0.005036760587245226, + 0.005217910278588533, + 0.003572372952476144, + 0.0048522124998271465, + 0.006039422005414963, + 0.03517765551805496, + 0.007016642484813929, + 0.011053621768951416, + 0.03960167244076729, + 0.00793793797492981, + 0.006721111945807934, + 0.018979383632540703, + 0.015768466517329216, + 0.028274966403841972, + 0.007387438789010048, + 0.21862465143203735, + 0.17657586932182312, + 0.01846698485314846, + 0.024594055488705635, + 0.004158591851592064, + 0.004301704466342926, + 0.005197679623961449, + 0.003801734885200858, + 0.009612939320504665, + 0.010260148905217648, + 0.005652740132063627, + 0.006506781093776226, + 0.005746654700487852, + 0.008005940355360508, + 0.0093154301866889, + 0.005031266715377569, + 0.01938117854297161, + 0.01769995503127575, + 0.004565955605357885, + 0.004807670135051012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.015399482101202011 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011103743687272072 + ], + "opposite_logprob_tokens": [ + -4.511103630065918 + ], + "logprob_reward": -0.011103743687272072, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011103743687272072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06991827487945557, + 0.07582297921180725, + 0.08563117682933807, + 0.033488158136606216, + 0.05142689868807793, + 0.053157102316617966, + 0.3151216506958008, + 0.2049238681793213, + 0.0322704054415226, + 0.07823947072029114 + ], + "credit_map": [ + 0.007372765336185694, + 0.005292817018926144, + 0.01123732328414917, + 0.006722292862832546, + 0.006823463831096888, + 0.014304456301033497, + 0.009724020957946777, + 0.008441142737865448, + 0.0066920011304318905, + 0.006531509570777416, + 0.012864174321293831, + 0.0072631132788956165, + 0.012205353938043118, + 0.013478060252964497, + 0.007402242161333561, + 0.009386532008647919, + 0.007646029349416494, + 0.009488056413829327, + 0.013059142045676708, + 0.00895862840116024, + 0.0128818079829216, + 0.011822449043393135, + 0.014248619787395, + 0.0075264484621584415, + 0.003826897358521819, + 0.0032562725245952606, + 0.004340383689850569, + 0.00413756538182497, + 0.004052841104567051, + 0.004336982034146786, + 0.003976107574999332, + 0.005561112891882658, + 0.004710206761956215, + 0.004377786070108414, + 0.009086699225008488, + 0.004368540830910206, + 0.005741967353969812, + 0.012760465033352375, + 0.005864591337740421, + 0.004516644403338432, + 0.0035763741470873356, + 0.00465422123670578, + 0.006632680539041758, + 0.007376766297966242, + 0.006977441720664501, + 0.008263139985501766, + 0.008920331485569477, + 0.006756151560693979, + 0.011331431567668915, + 0.011649446561932564, + 0.1045922115445137, + 0.01635138690471649, + 0.02139892429113388, + 0.11967810988426208, + 0.015066493302583694, + 0.015053658746182919, + 0.007492487318813801, + 0.0064336690120399, + 0.010915682651102543, + 0.0038294021505862474, + 0.08831945806741714, + 0.07572632282972336, + 0.006351906340569258, + 0.0058549633249640465, + 0.0023244544863700867, + 0.0029602148570120335, + 0.004574609454721212, + 0.0031807373743504286, + 0.005651320330798626, + 0.004740470089018345, + 0.0043367440812289715, + 0.0045018563978374004, + 0.005402717273682356, + 0.009175031445920467, + 0.01481657288968563, + 0.007061563897877932, + 0.013943948782980442, + 0.011943004094064236, + 0.009267612360417843, + 0.006629026960581541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.011103743687272072 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903422951698303 + ], + "opposite_logprob_tokens": [ + -2.46403431892395 + ], + "logprob_reward": -0.08903422951698303, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903422951698303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07777192443609238, + 0.04683763533830643, + 0.061183128505945206, + 0.04142432287335396, + 0.050355855375528336, + 0.0698448121547699, + 0.1211220771074295, + 0.387148916721344, + 0.0454932264983654, + 0.09881805628538132 + ], + "credit_map": [ + 0.00916012842208147, + 0.004941075574606657, + 0.012106476351618767, + 0.007517009042203426, + 0.007148225326091051, + 0.01391383446753025, + 0.011826303787529469, + 0.011158882640302181, + 0.0054237633012235165, + 0.004480411298573017, + 0.007637161295861006, + 0.003545986255630851, + 0.005169482436031103, + 0.009886149317026138, + 0.004450252279639244, + 0.006244429387152195, + 0.0058021340519189835, + 0.0059874593280255795, + 0.009711678139865398, + 0.0038569506723433733, + 0.013868510723114014, + 0.008371246978640556, + 0.0077269114553928375, + 0.005858237389475107, + 0.0036206538788974285, + 0.00450056791305542, + 0.005968168843537569, + 0.005929793696850538, + 0.006379099562764168, + 0.004996458999812603, + 0.00556945102289319, + 0.004460130352526903, + 0.005480567924678326, + 0.0057250699028372765, + 0.006788273341953754, + 0.005747271701693535, + 0.0069179837591946125, + 0.0074163395911455154, + 0.007359652314335108, + 0.004920701030641794, + 0.005167095456272364, + 0.009089305996894836, + 0.008103768341243267, + 0.00883537344634533, + 0.010789102874696255, + 0.01034378819167614, + 0.010103139095008373, + 0.007413252256810665, + 0.006632912904024124, + 0.012046512216329575, + 0.018555589020252228, + 0.0101840915158391, + 0.013655826449394226, + 0.025298921391367912, + 0.018537508323788643, + 0.016210731118917465, + 0.01062727253884077, + 0.009998666122555733, + 0.019213402643799782, + 0.004905801732093096, + 0.17075903713703156, + 0.14430998265743256, + 0.012035186402499676, + 0.015299562364816666, + 0.003453534794971347, + 0.004904989153146744, + 0.006724698003381491, + 0.0036863177083432674, + 0.008371942676603794, + 0.006255056243389845, + 0.006103701423853636, + 0.005992990918457508, + 0.0068263038992881775, + 0.014346395619213581, + 0.02004271000623703, + 0.010084214620292187, + 0.015933368355035782, + 0.013775177299976349, + 0.009042100049555302, + 0.008767795749008656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.08903422951698303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004106185864657164 + ], + "opposite_logprob_tokens": [ + -5.504106044769287 + ], + "logprob_reward": -0.004106185864657164, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004106185864657164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07623609900474548, + 0.05031350627541542, + 0.06312602013349533, + 0.027888348326086998, + 0.045879293233156204, + 0.054510727524757385, + 0.33625999093055725, + 0.061017956584692, + 0.07825999706983566, + 0.20650812983512878 + ], + "credit_map": [ + 0.010723844170570374, + 0.00627199886366725, + 0.011410283856093884, + 0.009798025712370872, + 0.006097136996686459, + 0.012079349718987942, + 0.009973758831620216, + 0.00988169014453888, + 0.004905572161078453, + 0.006757700815796852, + 0.007479474879801273, + 0.004842550493776798, + 0.00726340152323246, + 0.007904415018856525, + 0.004982046317309141, + 0.0061783380806446075, + 0.006331846117973328, + 0.0054632616229355335, + 0.01141126174479723, + 0.005105717107653618, + 0.008999316953122616, + 0.012445501983165741, + 0.0068119121715426445, + 0.006557192653417587, + 0.003022814402356744, + 0.002655087038874626, + 0.004430586472153664, + 0.004009982571005821, + 0.004390411078929901, + 0.002976942341774702, + 0.0040860832668840885, + 0.0023164385929703712, + 0.00527145154774189, + 0.00479835644364357, + 0.007292754482477903, + 0.0046564131043851376, + 0.007074050605297089, + 0.006212948355823755, + 0.0060117607936263084, + 0.004561553243547678, + 0.004267305135726929, + 0.0076585994102060795, + 0.008134428411722183, + 0.005963737145066261, + 0.00836031325161457, + 0.006522859912365675, + 0.007745955139398575, + 0.005857523530721664, + 0.012246329337358475, + 0.013680635020136833, + 0.11011858284473419, + 0.01780155673623085, + 0.022405728697776794, + 0.1281825304031372, + 0.014642528258264065, + 0.017182065173983574, + 0.007704412564635277, + 0.00696967076510191, + 0.007049001753330231, + 0.004877931904047728, + 0.010704481042921543, + 0.013528402894735336, + 0.004390793386846781, + 0.005793253891170025, + 0.00572822755202651, + 0.006692002061754465, + 0.01139851938933134, + 0.0065963659435510635, + 0.013157001696527004, + 0.013164855539798737, + 0.011694596149027348, + 0.009828426875174046, + 0.013224142603576183, + 0.02395937405526638, + 0.020322401076555252, + 0.019497912377119064, + 0.04068013280630112, + 0.03814473748207092, + 0.02328781969845295, + 0.027391575276851654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.004106185864657164 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009800885803997517 + ], + "opposite_logprob_tokens": [ + -4.634800910949707 + ], + "logprob_reward": -0.009800885803997517, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009800885803997517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08912227302789688, + 0.08992821723222733, + 0.035036806017160416, + 0.03803963586688042, + 0.04491160437464714, + 0.07771282643079758, + 0.11136597394943237, + 0.1388208419084549, + 0.061599407345056534, + 0.31346243619918823 + ], + "credit_map": [ + 0.010746695101261139, + 0.006298763211816549, + 0.013558336533606052, + 0.009531110525131226, + 0.007031514309346676, + 0.019152244552969933, + 0.010032289661467075, + 0.0127713056281209, + 0.007388393860310316, + 0.009966857731342316, + 0.016227899119257927, + 0.00746030081063509, + 0.014839858748018742, + 0.014635657891631126, + 0.008240154013037682, + 0.011169089935719967, + 0.003597447881475091, + 0.0037237447686493397, + 0.005705195479094982, + 0.0036705201491713524, + 0.004142029210925102, + 0.006407913751900196, + 0.004198732785880566, + 0.003591218963265419, + 0.0038127887528389692, + 0.003927641548216343, + 0.006173767149448395, + 0.005330333951860666, + 0.0069898804649710655, + 0.00411043968051672, + 0.004617535974830389, + 0.0030772455502301455, + 0.003692054655402899, + 0.004343995358794928, + 0.007993984036147594, + 0.004962152801454067, + 0.005049071274697781, + 0.009291809983551502, + 0.004937869030982256, + 0.004640659317374229, + 0.006225469056516886, + 0.00977476965636015, + 0.007775857113301754, + 0.011830485425889492, + 0.012123812921345234, + 0.010020026005804539, + 0.010193070396780968, + 0.009769321419298649, + 0.005040201358497143, + 0.005677070003002882, + 0.034864459186792374, + 0.0065315780229866505, + 0.008780795149505138, + 0.03781558573246002, + 0.006618318613618612, + 0.006037955638021231, + 0.007565178442746401, + 0.006697969511151314, + 0.0095389224588871, + 0.004563564900308847, + 0.05521514266729355, + 0.04149197041988373, + 0.00755369383841753, + 0.006194373592734337, + 0.0036474629305303097, + 0.004661181941628456, + 0.006935336627066135, + 0.005454689264297485, + 0.01103855948895216, + 0.011858266778290272, + 0.010470163077116013, + 0.007533745840191841, + 0.016746165230870247, + 0.03357890993356705, + 0.04710108041763306, + 0.031861890107393265, + 0.06543450057506561, + 0.0548660084605217, + 0.03655991330742836, + 0.027313949540257454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.009800885803997517 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This dramatic low-angle close-up, the camera slowly and almost imperceptibly zooming in, focuses on the face of a stoic, classical, bearded warrior bronze statue. The scene is plunged into deep, monumental darkness, the statue's features sculpted by a single, harsh keylight from above and sideways, creating a stark contrast of shimmering highlights and deep shadows. Initially, the statue's surface is a dark, rusty bronze with a subtle metallic sheen. Then, decay sets in; a dense network of fine cracks deliberately spreads across the cold metal of the cheekbones and temples. The edges of these cracks inexorably begin to lift and curl, flaking the solid bronze into thin, brittle flakes with the texture and fragility of dried, dead leaves. These fragments of metal detach and drift downward in a slow, continuous, silent waterfall, catching the dramatic light and glittering like copper shavings before disappearing into the blackness below, revealing a rougher, darker, and corroded texture beneath. The atmosphere is one of solemn and inevitable destruction, as the monument's eternal stoicism is captured in an elegant yet melancholic state of disintegration.", + "seed": 4036, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_11/index_224_seed_4036_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.028513123095035554, + "vlm_logprob": -0.028513123095035554, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.028513123095035554, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013301562517881393 + ], + "opposite_logprob_tokens": [ + -4.763301372528076 + ], + "logprob_reward": -0.013301562517881393, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.013301562517881393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14780566096305847, + 0.06330955773591995, + 0.025184551253914833, + 0.07351499050855637, + 0.06572341918945312, + 0.08886074274778366, + 0.15515463054180145, + 0.12345443665981293, + 0.1333686262369156, + 0.1236233338713646 + ], + "credit_map": [ + 0.025778360664844513, + 0.01116254087537527, + 0.020891107618808746, + 0.013787575997412205, + 0.017110364511609077, + 0.023767957463860512, + 0.015049558132886887, + 0.020258193835616112, + 0.012047156691551208, + 0.0044175381772220135, + 0.00924709253013134, + 0.0058736782521009445, + 0.006067691836506128, + 0.008569568395614624, + 0.006330994889140129, + 0.010755836963653564, + 0.0033426512964069843, + 0.0024640026967972517, + 0.003759095910936594, + 0.002441831398755312, + 0.0032841674983501434, + 0.003435201942920685, + 0.002892176853492856, + 0.003565424121916294, + 0.010743317194283009, + 0.00577689241617918, + 0.012810377404093742, + 0.005900380667299032, + 0.011995898559689522, + 0.00977451540529728, + 0.007442819885909557, + 0.00907079316675663, + 0.006476842798292637, + 0.004874700680375099, + 0.010123834013938904, + 0.0058070640079677105, + 0.008570524863898754, + 0.008503267541527748, + 0.013585892505943775, + 0.007781293708831072, + 0.015390734188258648, + 0.007087784819304943, + 0.009561595506966114, + 0.00820201076567173, + 0.010293680243194103, + 0.012609435245394707, + 0.013212745077908039, + 0.012502755038440228, + 0.020032797008752823, + 0.012337671592831612, + 0.023037508130073547, + 0.019124915823340416, + 0.0199456587433815, + 0.022510815411806107, + 0.023698309436440468, + 0.014466960914433002, + 0.014588063582777977, + 0.01203931588679552, + 0.014433695003390312, + 0.01245974376797676, + 0.015932541340589523, + 0.017580317333340645, + 0.02228473871946335, + 0.014136017300188541, + 0.012619505636394024, + 0.013967476785182953, + 0.013960909098386765, + 0.014803560450673103, + 0.018099460750818253, + 0.015109277330338955, + 0.031809743493795395, + 0.01299869455397129, + 0.014785921201109886, + 0.016756704077124596, + 0.01661727763712406, + 0.016551755368709564, + 0.015759579837322235, + 0.01575339026749134, + 0.01598524861037731, + 0.011413462460041046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.013301562517881393 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009805607609450817 + ], + "opposite_logprob_tokens": [ + -4.634805679321289 + ], + "logprob_reward": -0.009805607609450817, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009805607609450817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16577425599098206, + 0.0476405955851078, + 0.031113162636756897, + 0.07735652476549149, + 0.038025207817554474, + 0.06242887303233147, + 0.08936280757188797, + 0.16679121553897858, + 0.15633553266525269, + 0.16517186164855957 + ], + "credit_map": [ + 0.03659924864768982, + 0.013186934404075146, + 0.021938301622867584, + 0.016862405464053154, + 0.01818190887570381, + 0.019740300253033638, + 0.017057452350854874, + 0.022207684814929962, + 0.008485455997288227, + 0.0035379366017878056, + 0.007884552702307701, + 0.004541795700788498, + 0.004645223263651133, + 0.006864666473120451, + 0.005216767080128193, + 0.006464187987148762, + 0.00545712374150753, + 0.002268092939630151, + 0.004912660922855139, + 0.0029149064794182777, + 0.003299952717497945, + 0.004170210566371679, + 0.00358085916377604, + 0.00450935447588563, + 0.009967220947146416, + 0.00605469336733222, + 0.009635013528168201, + 0.006140574812889099, + 0.009251846931874752, + 0.012639726512134075, + 0.007633877918124199, + 0.016033563762903214, + 0.004859534557908773, + 0.003439305815845728, + 0.005612385459244251, + 0.0033027464523911476, + 0.005083422176539898, + 0.006113206036388874, + 0.0060746814124286175, + 0.00353992055170238, + 0.011251995339989662, + 0.004981446545571089, + 0.007226400077342987, + 0.006049816496670246, + 0.0066322386264801025, + 0.00838547945022583, + 0.009060329757630825, + 0.00884115882217884, + 0.01214808039367199, + 0.007156189065426588, + 0.013675094582140446, + 0.0099910544231534, + 0.011224383488297462, + 0.01330569013953209, + 0.012438913807272911, + 0.00942339189350605, + 0.0183821152895689, + 0.01840859465301037, + 0.019856536760926247, + 0.01904022879898548, + 0.025350021198391914, + 0.02069997973740101, + 0.028807705268263817, + 0.0162460058927536, + 0.018073944374918938, + 0.01420669723302126, + 0.021275239065289497, + 0.015292945317924023, + 0.020292645320296288, + 0.022921951487660408, + 0.025706153362989426, + 0.0185659471899271, + 0.023083297535777092, + 0.02021912857890129, + 0.018409086391329765, + 0.024675235152244568, + 0.01812424510717392, + 0.020187649875879288, + 0.023945394903421402, + 0.0165278110653162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.009805607609450817 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904425799846649 + ], + "opposite_logprob_tokens": [ + -2.4640443325042725 + ], + "logprob_reward": -0.08904425799846649, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904425799846649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08770792931318283, + 0.07312851399183273, + 0.0606151707470417, + 0.08219445496797562, + 0.055213067680597305, + 0.1779145449399948, + 0.08657336980104446, + 0.07598887383937836, + 0.08188632130622864, + 0.21877777576446533 + ], + "credit_map": [ + 0.013501649722456932, + 0.006898183375597, + 0.009312129579484463, + 0.010622718371450901, + 0.009573549032211304, + 0.012457824312150478, + 0.011430234648287296, + 0.013911638408899307, + 0.013286354020237923, + 0.006876599509268999, + 0.011754117906093597, + 0.0075092320330441, + 0.00831824354827404, + 0.00858679972589016, + 0.008172690868377686, + 0.008624479174613953, + 0.008772640489041805, + 0.005683181807398796, + 0.008806432597339153, + 0.0056927939876914024, + 0.007157165091484785, + 0.007468816824257374, + 0.007914066314697266, + 0.009120069444179535, + 0.010232290253043175, + 0.006149523891508579, + 0.015396819449961185, + 0.006125182379037142, + 0.0116353714838624, + 0.012173149734735489, + 0.010215462185442448, + 0.010266655124723911, + 0.0063257100991904736, + 0.004542027600109577, + 0.01073518767952919, + 0.004213694483041763, + 0.0085781030356884, + 0.007838035002350807, + 0.008041413500905037, + 0.004938898142427206, + 0.030646100640296936, + 0.013062821701169014, + 0.03563384339213371, + 0.012495999224483967, + 0.021751729771494865, + 0.023604024201631546, + 0.02155626192688942, + 0.019163774326443672, + 0.011940994299948215, + 0.00841681007295847, + 0.010739816352725029, + 0.010345695540308952, + 0.012187317945063114, + 0.011594939976930618, + 0.013303364627063274, + 0.008044430054724216, + 0.009456383995711803, + 0.00608681607991457, + 0.010459283366799355, + 0.006687591318041086, + 0.010378443636000156, + 0.010297630913555622, + 0.011512519791722298, + 0.011110205203294754, + 0.009097214788198471, + 0.008371579460799694, + 0.015312857925891876, + 0.010833385400474072, + 0.01193521823734045, + 0.00708275381475687, + 0.01247288566082716, + 0.006780428346246481, + 0.028749223798513412, + 0.024439994245767593, + 0.030839625746011734, + 0.028428127989172935, + 0.02726716361939907, + 0.027530422434210777, + 0.030997421592473984, + 0.02052581124007702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.08904425799846649 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0206089336425066 + ], + "opposite_logprob_tokens": [ + -3.895608901977539 + ], + "logprob_reward": -0.0206089336425066, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0206089336425066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14383192360401154, + 0.10644954442977905, + 0.052227843552827835, + 0.03634518012404442, + 0.040383122861385345, + 0.11466831713914871, + 0.24987681210041046, + 0.1345374882221222, + 0.06687936931848526, + 0.05480041727423668 + ], + "credit_map": [ + 0.025908149778842926, + 0.010596499778330326, + 0.015780048444867134, + 0.012791438028216362, + 0.018115799874067307, + 0.022743742913007736, + 0.017411235719919205, + 0.020485006272792816, + 0.020736735314130783, + 0.009599552489817142, + 0.012527153827250004, + 0.010483996011316776, + 0.009362196549773216, + 0.014957723207771778, + 0.009858861565589905, + 0.018923336640000343, + 0.008359668776392937, + 0.004435528535395861, + 0.005363943986594677, + 0.005172149743884802, + 0.00524795800447464, + 0.007588076870888472, + 0.006865378003567457, + 0.00919513963162899, + 0.005323122721165419, + 0.0030943735036998987, + 0.006196446251124144, + 0.0035350050311535597, + 0.005681883078068495, + 0.004267462063580751, + 0.004064827226102352, + 0.004182062577456236, + 0.006340817082673311, + 0.004839837551116943, + 0.00535541633144021, + 0.0033549554646015167, + 0.005790090654045343, + 0.005895284470170736, + 0.004214063752442598, + 0.004592656623572111, + 0.019307227805256844, + 0.00920824147760868, + 0.020532401278614998, + 0.010903909802436829, + 0.013922205194830894, + 0.015434924513101578, + 0.013203072361648083, + 0.01215633936226368, + 0.03203388303518295, + 0.030351724475622177, + 0.024297336116433144, + 0.032272737473249435, + 0.03100368194282055, + 0.03679798170924187, + 0.0361466221511364, + 0.02697284147143364, + 0.014706894755363464, + 0.015559018589556217, + 0.016500001773238182, + 0.015360136516392231, + 0.01881367713212967, + 0.020402958616614342, + 0.021286221221089363, + 0.01190858893096447, + 0.008906335569918156, + 0.0057505457662045956, + 0.010187913663685322, + 0.007469538599252701, + 0.0068303365260362625, + 0.009236912243068218, + 0.013337958604097366, + 0.005159826949238777, + 0.005094489082694054, + 0.005803130567073822, + 0.004939483478665352, + 0.009708484634757042, + 0.007421509828418493, + 0.006748019251972437, + 0.010393780656158924, + 0.004691522102802992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.0206089336425066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up focusing on the face of a bearded warrior statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dark with a single harsh keylight source coming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fine cracks spread across the statue's cheekbones and temples, causing bronze flakes to detach and drift downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling metal flakes glitter like copper shavings before disappearing into the blackness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009805253706872463 + ], + "opposite_logprob_tokens": [ + -4.634805202484131 + ], + "logprob_reward": -0.009805253706872463, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009805253706872463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30279961228370667, + 0.08459926396608353, + 0.028225691989064217, + 0.08325587958097458, + 0.04271228611469269, + 0.07212838530540466, + 0.08906050771474838, + 0.14674979448318481, + 0.10967519134283066, + 0.040793389081954956 + ], + "credit_map": [ + 0.06166883558034897, + 0.027560079470276833, + 0.035944145172834396, + 0.03053276427090168, + 0.03643167391419411, + 0.04008500277996063, + 0.035686343908309937, + 0.034890785813331604, + 0.017873218283057213, + 0.007296564988791943, + 0.011340010911226273, + 0.00828314945101738, + 0.008768818341195583, + 0.010349740274250507, + 0.009613895788788795, + 0.011073872447013855, + 0.005105694755911827, + 0.0022198627702891827, + 0.003546482417732477, + 0.002721971832215786, + 0.00261080265045166, + 0.004454381298273802, + 0.0032305687200278044, + 0.004335926845669746, + 0.009987559169530869, + 0.006484988145530224, + 0.01215602271258831, + 0.007067063357681036, + 0.009979874826967716, + 0.013017663732171059, + 0.00887258630245924, + 0.01569012552499771, + 0.006370609160512686, + 0.0047524431720376015, + 0.006304568611085415, + 0.003600403666496277, + 0.006018989719450474, + 0.0046585253439843655, + 0.00641911244019866, + 0.004587636329233646, + 0.009280440397560596, + 0.0043900031596422195, + 0.014193680137395859, + 0.00814406480640173, + 0.007855327799916267, + 0.00956272054463625, + 0.010902742855250835, + 0.00779940839856863, + 0.010979953221976757, + 0.013260304927825928, + 0.010247020050883293, + 0.012897828593850136, + 0.010549557395279408, + 0.012054837308824062, + 0.011109842918813229, + 0.007961167953908443, + 0.016733532771468163, + 0.01739666797220707, + 0.018591880798339844, + 0.016444267705082893, + 0.01930692233145237, + 0.021887993440032005, + 0.020409798249602318, + 0.015978733077645302, + 0.012294785119593143, + 0.010700258426368237, + 0.011614528484642506, + 0.014302678406238556, + 0.012584575451910496, + 0.013668971136212349, + 0.02143103815615177, + 0.013078357093036175, + 0.005582117475569248, + 0.003397299675270915, + 0.005425769370049238, + 0.005338718648999929, + 0.005650212988257408, + 0.004358899313956499, + 0.0060439747758209705, + 0.004996399395167828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.009805253706872463 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_12/index_tensor([1427, 1427])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_12/index_tensor([1427, 1427])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f2f5c59c585d0cd1c607cdc15e9b125e2e34907d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_12/index_tensor([1427, 1427])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful first-person point-of-view (POV) shot, with the camera tilted slightly downward and a slight handheld shake, simulates the human reaction of observation. Focused on a glossy, white kitchen tile floor, a monstrous centipede stands motionless. The creature's carapace is distinctly segmented, exhibiting an unsettling metallic sheen of electric blue and flame orange, interwoven with deep black stripes. Under the cold white light of the kitchen overhead, its smooth shell reflects sharp highlights, a texture so realistic that it appears as if each segment were coated with a layer of wet varnish. In an instant, as if sensing a threat, its entire body suddenly curls inward in a swift, spasmodic motion, its countless slender, sharp legs instantly folding together into a nearly impenetrable, defensive spiral, burying its head deep within. This sudden, decisive movement creates a stark contrast between the bright, clean background and the creature's grotesque, vibrant colors, creating a thrilling atmosphere of tension and terror.", + "seed": 17708, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_12/index_1427_seed_17708_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05325551088899374, + "vlm_logprob": -0.05325551088899374, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.05325551088899374, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09654758870601654 + ], + "opposite_logprob_tokens": [ + -2.4715476036071777 + ], + "logprob_reward": -0.09654758870601654, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09654758870601654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27417248487472534, + 0.16561046242713928, + 0.047105077654123306, + 0.07774419337511063, + 0.0736590102314949, + 0.07091693580150604, + 0.056570444256067276, + 0.08860006183385849, + 0.09817290306091309, + 0.04744842275977135 + ], + "credit_map": [ + 0.037786468863487244, + 0.026991739869117737, + 0.046033307909965515, + 0.034656181931495667, + 0.02234242670238018, + 0.06456193327903748, + 0.02145257778465748, + 0.02034788392484188, + 0.02348940260708332, + 0.016210194677114487, + 0.03078041411936283, + 0.02331860549747944, + 0.015313818119466305, + 0.0230756476521492, + 0.018123691901564598, + 0.015298674814403057, + 0.007497461047023535, + 0.007554471027106047, + 0.006125001236796379, + 0.004592191427946091, + 0.004136604256927967, + 0.00583885982632637, + 0.0057585411705076694, + 0.005601945333182812, + 0.013057488948106766, + 0.021240998059511185, + 0.007228359114378691, + 0.007580646313726902, + 0.006587402429431677, + 0.008777210488915443, + 0.006190691143274307, + 0.007081401534378529, + 0.014963481575250626, + 0.007757139857858419, + 0.02142062596976757, + 0.006252048537135124, + 0.005614628549665213, + 0.006060973275452852, + 0.006270000711083412, + 0.005320110358297825, + 0.00848980713635683, + 0.01331526692956686, + 0.010908636264503002, + 0.013510307297110558, + 0.005045470781624317, + 0.008055991493165493, + 0.0052276658825576305, + 0.006363792810589075, + 0.0073058949783444405, + 0.006987425964325666, + 0.007287542335689068, + 0.0088803730905056, + 0.0056689404882490635, + 0.005904918536543846, + 0.006401825230568647, + 0.008133523166179657, + 0.017946606501936913, + 0.008171383291482925, + 0.006750061642378569, + 0.005614768713712692, + 0.007115772459656, + 0.0072503527626395226, + 0.00912292767316103, + 0.02662818878889084, + 0.009374461136758327, + 0.00744961341843009, + 0.00742133567109704, + 0.010194703005254269, + 0.006217826157808304, + 0.014841608703136444, + 0.030700307339429855, + 0.011973049491643906, + 0.004864370450377464, + 0.005600274540483952, + 0.005555588286370039, + 0.005735872313380241, + 0.0038046902045607567, + 0.007929808460175991, + 0.008323620073497295, + 0.005634194239974022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.09654758870601654 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029812553897500038 + ], + "opposite_logprob_tokens": [ + -3.5298125743865967 + ], + "logprob_reward": -0.029812553897500038, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029812553897500038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23486639559268951, + 0.3046158254146576, + 0.06734801828861237, + 0.04061390087008476, + 0.02544543147087097, + 0.02300454117357731, + 0.022475294768810272, + 0.09263141453266144, + 0.17444148659706116, + 0.014557676389813423 + ], + "credit_map": [ + 0.028171854093670845, + 0.01938788592815399, + 0.049552470445632935, + 0.025022374466061592, + 0.018562080338597298, + 0.06344685703516006, + 0.015772562474012375, + 0.0149503443390131, + 0.04639596864581108, + 0.030684825032949448, + 0.05775916203856468, + 0.042541202157735825, + 0.023268256336450577, + 0.046902429312467575, + 0.033025290817022324, + 0.024038728326559067, + 0.012597774155437946, + 0.009408388286828995, + 0.010292533785104752, + 0.008936983533203602, + 0.0054851919412612915, + 0.009758060798048973, + 0.005324540194123983, + 0.005544546991586685, + 0.008569519966840744, + 0.007277262397110462, + 0.003922794479876757, + 0.00334018561989069, + 0.0053191957995295525, + 0.004576386883854866, + 0.003656829474493861, + 0.003951729740947485, + 0.003452652134001255, + 0.003023681929334998, + 0.004475446883589029, + 0.002072741510346532, + 0.0028321456629782915, + 0.003076021559536457, + 0.003198833903297782, + 0.0033139120787382126, + 0.0024413547944277525, + 0.0041867815889418125, + 0.002533052349463105, + 0.0035164826549589634, + 0.0024641992058604956, + 0.002459850860759616, + 0.002491238759830594, + 0.002911583287641406, + 0.002289131749421358, + 0.002135318471118808, + 0.002746326383203268, + 0.0042081717401742935, + 0.0027898584958165884, + 0.0019986594561487436, + 0.002922395244240761, + 0.003385435091331601, + 0.007868248037993908, + 0.005598700139671564, + 0.004751904867589474, + 0.003454808611422777, + 0.00724830012768507, + 0.025924798101186752, + 0.008587918244302273, + 0.02919674664735794, + 0.015471452847123146, + 0.013615202158689499, + 0.014779295772314072, + 0.021016394719481468, + 0.017474841326475143, + 0.029625624418258667, + 0.04468674957752228, + 0.017771948128938675, + 0.001352463266812265, + 0.002123515121638775, + 0.0013761635636910796, + 0.001757535501383245, + 0.0010955306934192777, + 0.0021236788015812635, + 0.0023939956445246935, + 0.0023347956594079733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.029812553897500038 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015187645331025124 + ], + "opposite_logprob_tokens": [ + -6.501518726348877 + ], + "logprob_reward": -0.0015187645331025124, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015187645331025124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13151097297668457, + 0.13938874006271362, + 0.051880914717912674, + 0.150040403008461, + 0.06189984455704689, + 0.03268791735172272, + 0.056842319667339325, + 0.1766061633825302, + 0.13920845091342926, + 0.05993425473570824 + ], + "credit_map": [ + 0.009718673303723335, + 0.015231351368129253, + 0.021125255152583122, + 0.02079918049275875, + 0.011451837606728077, + 0.024677544832229614, + 0.013371939770877361, + 0.01513519324362278, + 0.021002978086471558, + 0.012401321902871132, + 0.02970198728144169, + 0.01918615587055683, + 0.007874280214309692, + 0.02531304769217968, + 0.013286362402141094, + 0.010622624307870865, + 0.012972680851817131, + 0.004441810306161642, + 0.007766177877783775, + 0.004189406521618366, + 0.0034778262488543987, + 0.011677311733365059, + 0.003350120736286044, + 0.004005586262792349, + 0.03082224167883396, + 0.05528901144862175, + 0.013666280545294285, + 0.008229006081819534, + 0.011952233500778675, + 0.017695287242531776, + 0.00579886045306921, + 0.006587503477931023, + 0.010271693579852581, + 0.006258264183998108, + 0.01260319259017706, + 0.005166754126548767, + 0.005432922393083572, + 0.0075121656991541386, + 0.008651357144117355, + 0.006003500893712044, + 0.0037269038148224354, + 0.005585762206465006, + 0.003445429028943181, + 0.0053774709813296795, + 0.0034874468110501766, + 0.004800291266292334, + 0.0030378694646060467, + 0.003226746339350939, + 0.008179360069334507, + 0.005977647844702005, + 0.00707474397495389, + 0.009938270784914494, + 0.0047518606297671795, + 0.007156290579587221, + 0.006827496457844973, + 0.00693665212020278, + 0.04488682374358177, + 0.011377238668501377, + 0.009986168704926968, + 0.0077325329184532166, + 0.012659430503845215, + 0.03153205290436745, + 0.0172993466258049, + 0.041132569313049316, + 0.014299710281193256, + 0.010448640212416649, + 0.015336201526224613, + 0.014746956527233124, + 0.007963570766150951, + 0.023181220516562462, + 0.037265777587890625, + 0.015966370701789856, + 0.006262642797082663, + 0.007349863648414612, + 0.007389831822365522, + 0.0077127921395003796, + 0.004549519624561071, + 0.01054617390036583, + 0.009314301423728466, + 0.006809134501963854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.0015187645331025124 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030536353588104 + ], + "opposite_logprob_tokens": [ + -2.3503053188323975 + ], + "logprob_reward": -0.10030536353588104, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030536353588104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03927310183644295, + 0.06458072364330292, + 0.0417412631213665, + 0.23315775394439697, + 0.06218988448381424, + 0.0475173145532608, + 0.038832888007164, + 0.17841538786888123, + 0.2544127106666565, + 0.03987898677587509 + ], + "credit_map": [ + 0.003079252550378442, + 0.005925443023443222, + 0.0061652204021811485, + 0.00456445524469018, + 0.0035896652843803167, + 0.007753078825771809, + 0.004329265095293522, + 0.0038667183835059404, + 0.006365754641592503, + 0.007390892598778009, + 0.013200588524341583, + 0.005914133507758379, + 0.007456411607563496, + 0.011269931681454182, + 0.006409818306565285, + 0.006573184859007597, + 0.009277105331420898, + 0.004956355784088373, + 0.0059383767656981945, + 0.0044311415404081345, + 0.0025733422953635454, + 0.0068296873942017555, + 0.003828417044132948, + 0.003906833939254284, + 0.049438633024692535, + 0.086232490837574, + 0.019292844459414482, + 0.011598795652389526, + 0.018119364976882935, + 0.028607292100787163, + 0.00976465456187725, + 0.010103665292263031, + 0.011636578477919102, + 0.005837989039719105, + 0.016554493457078934, + 0.004857970401644707, + 0.0058424449525773525, + 0.007230741437524557, + 0.004364551976323128, + 0.00586510868743062, + 0.005752981174737215, + 0.008210867643356323, + 0.006529964040964842, + 0.006777945440262556, + 0.003649443155154586, + 0.007160106673836708, + 0.004696378950029612, + 0.004739624913781881, + 0.005153122823685408, + 0.0037809235509485006, + 0.005195735488086939, + 0.00667991628870368, + 0.0029963732231408358, + 0.005220211576670408, + 0.005124429240822792, + 0.004682173021137714, + 0.046661920845508575, + 0.012495061382651329, + 0.01112471055239439, + 0.00746757909655571, + 0.013157949782907963, + 0.017135068774223328, + 0.017368288710713387, + 0.05300479382276535, + 0.0251584704965353, + 0.01491495780646801, + 0.021340034902095795, + 0.025692909955978394, + 0.013763098046183586, + 0.04938173294067383, + 0.07536356896162033, + 0.02879791148006916, + 0.004436112008988857, + 0.004600432235747576, + 0.0049666473641991615, + 0.00520732905715704, + 0.0031838365830481052, + 0.007360130548477173, + 0.006074617616832256, + 0.004049879033118486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.10030536353588104 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809328377246857 + ], + "opposite_logprob_tokens": [ + -3.288093328475952 + ], + "logprob_reward": -0.03809328377246857, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809328377246857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21449285745620728, + 0.15012207627296448, + 0.05783277377486229, + 0.059123773127794266, + 0.024996032938361168, + 0.07624272257089615, + 0.018812600523233414, + 0.13264818489551544, + 0.24497239291667938, + 0.02075660042464733 + ], + "credit_map": [ + 0.025166712701320648, + 0.020004598423838615, + 0.04206332936882973, + 0.02130742371082306, + 0.019880734384059906, + 0.0576598085463047, + 0.01495726965367794, + 0.013452977873384953, + 0.02293502166867256, + 0.015790970996022224, + 0.026874464005231857, + 0.021000823006033897, + 0.011649228632450104, + 0.022318240255117416, + 0.01672004535794258, + 0.012833280488848686, + 0.008937246166169643, + 0.00755311967805028, + 0.010092710144817829, + 0.008778789080679417, + 0.0039588576182723045, + 0.007306771818548441, + 0.0055127982050180435, + 0.0056924764066934586, + 0.010476735420525074, + 0.009320675395429134, + 0.006531985476613045, + 0.0061747231520712376, + 0.00686456635594368, + 0.007314448710530996, + 0.005736015271395445, + 0.006704618223011494, + 0.0032463776879012585, + 0.00395170459523797, + 0.004204216878861189, + 0.00201961281709373, + 0.003725094487890601, + 0.002639582846313715, + 0.0020260114688426256, + 0.0031834302935749292, + 0.008181195706129074, + 0.011592838913202286, + 0.006482318509370089, + 0.014406213536858559, + 0.010727672837674618, + 0.0074650743044912815, + 0.0075468095019459724, + 0.00984059739857912, + 0.0021860830020159483, + 0.0017792127327993512, + 0.002351585542783141, + 0.00262865680269897, + 0.002280903048813343, + 0.0021628595422953367, + 0.002829882549121976, + 0.002593415090814233, + 0.022726379334926605, + 0.0061813113279640675, + 0.00616078544408083, + 0.004438323900103569, + 0.007007563021034002, + 0.053021807223558426, + 0.008888067677617073, + 0.024223940446972847, + 0.022013701498508453, + 0.016266999766230583, + 0.01835288107395172, + 0.026989499107003212, + 0.019801435992121696, + 0.04529197886586189, + 0.07004591822624207, + 0.02620997093617916, + 0.001985337818041444, + 0.0032377243041992188, + 0.0015449592610821128, + 0.0022799980361014605, + 0.0019744294695556164, + 0.0034676245413720608, + 0.00321183237247169, + 0.003054695436730981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.03809328377246857 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful first-person point-of-view (POV) shot, with the camera tilted slightly downward and a slight handheld shake, simulates the human reaction of observation. Focused on a glossy, white kitchen tile floor, a monstrous centipede stands motionless. The creature's carapace is distinctly segmented, exhibiting an unsettling metallic sheen of electric blue and flame orange, interwoven with deep black stripes. Under the cold white light of the kitchen overhead, its smooth shell reflects sharp highlights, a texture so realistic that it appears as if each segment were coated with a layer of wet varnish. In an instant, as if sensing a threat, its entire body suddenly curls inward in a swift, spasmodic motion, its countless slender, sharp legs instantly folding together into a nearly impenetrable, defensive spiral, burying its head deep within. This sudden, decisive movement creates a stark contrast between the bright, clean background and the creature's grotesque, vibrant colors, creating a thrilling atmosphere of tension and terror.", + "seed": 17709, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_12/index_1427_seed_17709_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07188235986977816, + "vlm_logprob": -0.07188235986977816, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.07188235986977816, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21305136382579803 + ], + "opposite_logprob_tokens": [ + -1.7130513191223145 + ], + "logprob_reward": -0.21305136382579803, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21305136382579803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14602626860141754, + 0.07344682514667511, + 0.049757346510887146, + 0.06946689635515213, + 0.0640888661146164, + 0.08383236080408096, + 0.21708258986473083, + 0.12018121033906937, + 0.06596967577934265, + 0.1101478710770607 + ], + "credit_map": [ + 0.017825312912464142, + 0.010688519105315208, + 0.02804410643875599, + 0.015868227928876877, + 0.015497523359954357, + 0.022044016048312187, + 0.019616395235061646, + 0.016442185267806053, + 0.008329176343977451, + 0.006197280250489712, + 0.015079177916049957, + 0.009581152349710464, + 0.0061053442768752575, + 0.008512718603014946, + 0.008787975646555424, + 0.010854008607566357, + 0.009505004622042179, + 0.005861158948391676, + 0.007026079576462507, + 0.0070049697533249855, + 0.003449424169957638, + 0.004081962630152702, + 0.006167647894471884, + 0.006661102641373873, + 0.01573273539543152, + 0.006920622196048498, + 0.011721094138920307, + 0.007620096672326326, + 0.004986309912055731, + 0.004251826088875532, + 0.007456475403159857, + 0.010777738876640797, + 0.012061347253620625, + 0.005018332973122597, + 0.00834998395293951, + 0.008786654099822044, + 0.010746779851615429, + 0.004983327351510525, + 0.006094512529671192, + 0.00804793182760477, + 0.021215148270130157, + 0.004756287205964327, + 0.02266945317387581, + 0.006489515770226717, + 0.0050962213426828384, + 0.008924751542508602, + 0.00791005790233612, + 0.006770930718630552, + 0.05734369158744812, + 0.019293025135993958, + 0.05549176037311554, + 0.01556849293410778, + 0.014024656265974045, + 0.017801282927393913, + 0.018809491768479347, + 0.018750205636024475, + 0.01857086457312107, + 0.015876125544309616, + 0.01380101963877678, + 0.014063270762562752, + 0.015275336802005768, + 0.012463925406336784, + 0.019899947568774223, + 0.010230736806988716, + 0.009690210223197937, + 0.008288144133985043, + 0.007066504564136267, + 0.007270018104463816, + 0.006885480135679245, + 0.01014690101146698, + 0.01082837674766779, + 0.005794054828584194, + 0.01169904600828886, + 0.01386548113077879, + 0.013107080943882465, + 0.022580908611416817, + 0.009388632141053677, + 0.013571590185165405, + 0.011960811913013458, + 0.013974332250654697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.21305136382579803 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026360243558883667 + ], + "opposite_logprob_tokens": [ + -3.651360273361206 + ], + "logprob_reward": -0.026360243558883667, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026360243558883667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2087288498878479, + 0.06811536848545074, + 0.11363376677036285, + 0.04189648851752281, + 0.08557600528001785, + 0.07557670027017593, + 0.2720479965209961, + 0.07996392995119095, + 0.024497082456946373, + 0.02996380627155304 + ], + "credit_map": [ + 0.020143577829003334, + 0.014584967866539955, + 0.04310121759772301, + 0.02165887877345085, + 0.021081525832414627, + 0.02984572760760784, + 0.028520120307803154, + 0.029792850837111473, + 0.010003805160522461, + 0.004761336371302605, + 0.01063241995871067, + 0.005918306298553944, + 0.005255565512925386, + 0.012388424947857857, + 0.008378948085010052, + 0.010776566341519356, + 0.02334449626505375, + 0.009333711117506027, + 0.014284282922744751, + 0.01573203317821026, + 0.010399891063570976, + 0.007624090649187565, + 0.013997041620314121, + 0.018918229267001152, + 0.00660770945250988, + 0.004364709369838238, + 0.005998122040182352, + 0.005200363229960203, + 0.003680311143398285, + 0.0027636531740427017, + 0.0064774733036756516, + 0.00680414866656065, + 0.014924420043826103, + 0.012928120791912079, + 0.013329134322702885, + 0.010800964199006557, + 0.008474036119878292, + 0.009885570965707302, + 0.006758876610547304, + 0.008474890142679214, + 0.01067345216870308, + 0.010699281468987465, + 0.010746459476649761, + 0.009361699223518372, + 0.007488864008337259, + 0.006918476428836584, + 0.009844967164099216, + 0.009843510575592518, + 0.0732690691947937, + 0.024832524359226227, + 0.06698063015937805, + 0.020542984828352928, + 0.01751961000263691, + 0.022335102781653404, + 0.021236201748251915, + 0.025331877171993256, + 0.012347524054348469, + 0.011406522244215012, + 0.008526203222572803, + 0.01010650023818016, + 0.008953490294516087, + 0.009568732231855392, + 0.011873360723257065, + 0.00718160392716527, + 0.0044122422114014626, + 0.0022808655630797148, + 0.004216030705720186, + 0.003135435516014695, + 0.002208251738920808, + 0.002912848489359021, + 0.0027637621387839317, + 0.0025676479563117027, + 0.003904065815731883, + 0.0030743309762328863, + 0.0037876348942518234, + 0.005626525729894638, + 0.0026906784623861313, + 0.0038529953453689814, + 0.0036156687419861555, + 0.0034119104966521263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.026360243558883667 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015175743028521538 + ], + "opposite_logprob_tokens": [ + -6.5015177726745605 + ], + "logprob_reward": -0.0015175743028521538, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015175743028521538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16083663702011108, + 0.056050971150398254, + 0.07255127280950546, + 0.05972440913319588, + 0.17601743340492249, + 0.08003129065036774, + 0.1404573619365692, + 0.17123587429523468, + 0.03534785658121109, + 0.047746915370225906 + ], + "credit_map": [ + 0.01505257748067379, + 0.009832639247179031, + 0.030340060591697693, + 0.01850060746073723, + 0.014505167491734028, + 0.02280261181294918, + 0.02391214855015278, + 0.025890827178955078, + 0.005989512894302607, + 0.007964090444147587, + 0.01183993462473154, + 0.007325653452426195, + 0.005046747624874115, + 0.007673520129173994, + 0.005124329589307308, + 0.005087187513709068, + 0.014123544096946716, + 0.0074227130971848965, + 0.011954886838793755, + 0.010276488028466702, + 0.006045673042535782, + 0.0041094329208135605, + 0.008287511765956879, + 0.010331029072403908, + 0.010219013318419456, + 0.0062899962067604065, + 0.00805721990764141, + 0.0073705692775547504, + 0.005625616759061813, + 0.005271282512694597, + 0.007890389300882816, + 0.009000319056212902, + 0.022525345906615257, + 0.01659865863621235, + 0.020569533109664917, + 0.024049649015069008, + 0.0383162647485733, + 0.015343707986176014, + 0.016018884256482124, + 0.02259538508951664, + 0.016096636652946472, + 0.0046478682197630405, + 0.019326530396938324, + 0.005775130353868008, + 0.006059837527573109, + 0.007721580099314451, + 0.010920178145170212, + 0.009483525529503822, + 0.025343671441078186, + 0.015262764878571033, + 0.025477834045886993, + 0.01693364977836609, + 0.011354331858456135, + 0.014441944658756256, + 0.01847250945866108, + 0.013170660473406315, + 0.02426327019929886, + 0.017612282186746597, + 0.012703493237495422, + 0.01845058612525463, + 0.020039107650518417, + 0.010546569712460041, + 0.05466005578637123, + 0.0129605233669281, + 0.005112025421112776, + 0.004741165321320295, + 0.006283401977270842, + 0.0043313587084412575, + 0.0036594087723642588, + 0.0036741821095347404, + 0.00390897411853075, + 0.0036373394541442394, + 0.004197141155600548, + 0.006276901811361313, + 0.006383739877492189, + 0.008535428903996944, + 0.004109042696654797, + 0.00479882163926959, + 0.006290877237915993, + 0.007154962047934532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.0015175743028521538 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030104964971542 + ], + "opposite_logprob_tokens": [ + -2.3503010272979736 + ], + "logprob_reward": -0.10030104964971542, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030104964971542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08770861476659775, + 0.06488139927387238, + 0.04603317379951477, + 0.052750807255506516, + 0.07799135893583298, + 0.03553703799843788, + 0.06244346499443054, + 0.46193981170654297, + 0.08504775166511536, + 0.025666555389761925 + ], + "credit_map": [ + 0.007221890613436699, + 0.007947014644742012, + 0.015496172942221165, + 0.011772657744586468, + 0.008177966810762882, + 0.010682226158678532, + 0.014996817335486412, + 0.011413871310651302, + 0.009533382020890713, + 0.004239952191710472, + 0.021493393927812576, + 0.005673228297382593, + 0.0035113648045808077, + 0.010394744575023651, + 0.004983955528587103, + 0.005051380954682827, + 0.007170971017330885, + 0.005200223997235298, + 0.005816844291985035, + 0.009074750356376171, + 0.004169951658695936, + 0.0036610967945307493, + 0.005598108284175396, + 0.005341229494661093, + 0.008942213840782642, + 0.005534068215638399, + 0.007373734377324581, + 0.005883393343538046, + 0.005892777815461159, + 0.005567213520407677, + 0.006055701058357954, + 0.007501700893044472, + 0.010776020586490631, + 0.008225981146097183, + 0.009539335034787655, + 0.009394961409270763, + 0.018164455890655518, + 0.007077435962855816, + 0.006682943552732468, + 0.008130230009555817, + 0.007404152303934097, + 0.00232866033911705, + 0.0074480511248111725, + 0.002995933871716261, + 0.004385818727314472, + 0.0032632446382194757, + 0.004333027638494968, + 0.0033781512174755335, + 0.011083127930760384, + 0.006604496389627457, + 0.010302837938070297, + 0.005236259661614895, + 0.005028466694056988, + 0.00588314700871706, + 0.010745425708591938, + 0.007559701334685087, + 0.07287459820508957, + 0.0543535053730011, + 0.04111667722463608, + 0.04336536303162575, + 0.05492524430155754, + 0.040177926421165466, + 0.11638893932104111, + 0.03873755782842636, + 0.013219003565609455, + 0.013240191154181957, + 0.010231376625597477, + 0.011018864810466766, + 0.008661605417728424, + 0.010638570412993431, + 0.00909294094890356, + 0.00894519779831171, + 0.0028375175315886736, + 0.0029118501115590334, + 0.0029809291008859873, + 0.004269220866262913, + 0.0019919087644666433, + 0.0032366951927542686, + 0.003247277345508337, + 0.004191156942397356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.10030104964971542 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018181568011641502 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.018181568011641502, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018181568011641502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18849213421344757, + 0.05163457617163658, + 0.14941659569740295, + 0.1007099524140358, + 0.07429595291614532, + 0.049153245985507965, + 0.16760249435901642, + 0.14389070868492126, + 0.0408414751291275, + 0.03396286442875862 + ], + "credit_map": [ + 0.01775389537215233, + 0.01433208305388689, + 0.040667105466127396, + 0.01925678178668022, + 0.0171847902238369, + 0.029271332547068596, + 0.029471619054675102, + 0.020554522052407265, + 0.0051803975366055965, + 0.008295321837067604, + 0.006165353115648031, + 0.007550330832600594, + 0.005573230795562267, + 0.00559273362159729, + 0.007092142011970282, + 0.006185067351907492, + 0.04124313220381737, + 0.013291006907820702, + 0.020297789946198463, + 0.019160084426403046, + 0.00997493602335453, + 0.01157320011407137, + 0.01590687595307827, + 0.01796957664191723, + 0.021763350814580917, + 0.010606610216200352, + 0.0171049851924181, + 0.011571915820240974, + 0.007262289524078369, + 0.008101637475192547, + 0.00845776591449976, + 0.015841392800211906, + 0.011840679682791233, + 0.00691050523892045, + 0.010428009554743767, + 0.010676386766135693, + 0.00810608547180891, + 0.007522595580667257, + 0.006733905989676714, + 0.01207778975367546, + 0.010923819616436958, + 0.0038389372639358044, + 0.010484184138476849, + 0.004127070773392916, + 0.005415251944214106, + 0.0038410904817283154, + 0.007016147021204233, + 0.0035067452117800713, + 0.04108324274420738, + 0.01503292191773653, + 0.03673131391406059, + 0.013064856640994549, + 0.010998657904565334, + 0.01803818717598915, + 0.017016572877764702, + 0.015636738389730453, + 0.024357210844755173, + 0.01950870081782341, + 0.014025166630744934, + 0.015836698934435844, + 0.015278968028724194, + 0.017840782180428505, + 0.023708777502179146, + 0.013334397226572037, + 0.006455888971686363, + 0.006526848301291466, + 0.005043391138315201, + 0.005194463301450014, + 0.0035101226530969143, + 0.005390858743339777, + 0.004288970027118921, + 0.004430933389812708, + 0.0034653167240321636, + 0.003430836135521531, + 0.003786752698943019, + 0.003919960465282202, + 0.003114883555099368, + 0.0035880361683666706, + 0.00537834269925952, + 0.007278735749423504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.018181568011641502 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful first-person point-of-view (POV) shot, with the camera tilted slightly downward and a slight handheld shake, simulates the human reaction of observation. Focused on a glossy, white kitchen tile floor, a monstrous centipede stands motionless. The creature's carapace is distinctly segmented, exhibiting an unsettling metallic sheen of electric blue and flame orange, interwoven with deep black stripes. Under the cold white light of the kitchen overhead, its smooth shell reflects sharp highlights, a texture so realistic that it appears as if each segment were coated with a layer of wet varnish. In an instant, as if sensing a threat, its entire body suddenly curls inward in a swift, spasmodic motion, its countless slender, sharp legs instantly folding together into a nearly impenetrable, defensive spiral, burying its head deep within. This sudden, decisive movement creates a stark contrast between the bright, clean background and the creature's grotesque, vibrant colors, creating a thrilling atmosphere of tension and terror.", + "seed": 17710, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_12/index_1427_seed_17710_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05088891226332635, + "vlm_logprob": -0.05088891226332635, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.05088891226332635, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07669569551944733 + ], + "opposite_logprob_tokens": [ + -2.701695680618286 + ], + "logprob_reward": -0.07669569551944733, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07669569551944733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048246387392282486, + 0.1109263077378273, + 0.04009666293859482, + 0.0606829971075058, + 0.1467241793870926, + 0.09643933922052383, + 0.3066641688346863, + 0.03577878326177597, + 0.062129076570272446, + 0.09231210500001907 + ], + "credit_map": [ + 0.0073585049249231815, + 0.004525038879364729, + 0.007040632423013449, + 0.0056354277767241, + 0.004760373383760452, + 0.007262279279530048, + 0.005101243499666452, + 0.006562887690961361, + 0.02540639415383339, + 0.007373903878033161, + 0.02041560970246792, + 0.02139054611325264, + 0.0071421898901462555, + 0.009015239775180817, + 0.011183046735823154, + 0.008999384008347988, + 0.0071154371835291386, + 0.004190597217530012, + 0.004550247918814421, + 0.004685685038566589, + 0.005975435487926006, + 0.007128849625587463, + 0.002597660291939974, + 0.003852750873193145, + 0.00720115564763546, + 0.00901295617222786, + 0.011281297542154789, + 0.00682316068559885, + 0.006569353397935629, + 0.006777402013540268, + 0.005698489025235176, + 0.007319175638258457, + 0.008365780115127563, + 0.03746933117508888, + 0.01153575163334608, + 0.01783604919910431, + 0.017126979306340218, + 0.011113745160400867, + 0.032087359577417374, + 0.011189178563654423, + 0.011441502720117569, + 0.01234526839107275, + 0.012481655925512314, + 0.009906635619699955, + 0.010111149400472641, + 0.011194168590009212, + 0.015116347931325436, + 0.013842605985701084, + 0.022069398313760757, + 0.018684884533286095, + 0.021842436864972115, + 0.06686317920684814, + 0.013231162913143635, + 0.010112390853464603, + 0.11968351900577545, + 0.034177184104919434, + 0.004279812797904015, + 0.005597951356321573, + 0.006464438512921333, + 0.004566955845803022, + 0.002919339807704091, + 0.0027878484688699245, + 0.004648077301681042, + 0.004514362197369337, + 0.007037031929939985, + 0.006500339135527611, + 0.008674431592226028, + 0.006396121345460415, + 0.005953532177954912, + 0.006964909844100475, + 0.01103894878178835, + 0.00956376176327467, + 0.009869770146906376, + 0.013566398993134499, + 0.008116906508803368, + 0.011540016159415245, + 0.013178730383515358, + 0.011578704230487347, + 0.011597937904298306, + 0.012863644398748875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.07669569551944733 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048672039061784744 + ], + "opposite_logprob_tokens": [ + -3.0486719608306885 + ], + "logprob_reward": -0.048672039061784744, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048672039061784744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06636618077754974, + 0.15916091203689575, + 0.15051372349262238, + 0.12184889614582062, + 0.19363948702812195, + 0.05725106969475746, + 0.06420359760522842, + 0.05095870792865753, + 0.08124561607837677, + 0.05481182783842087 + ], + "credit_map": [ + 0.008994112722575665, + 0.005670339800417423, + 0.010390167124569416, + 0.008411006070673466, + 0.0058723511174321175, + 0.010958395898342133, + 0.007903393357992172, + 0.008166412822902203, + 0.030590258538722992, + 0.010680965147912502, + 0.030028127133846283, + 0.03595130518078804, + 0.011600504629313946, + 0.013422133401036263, + 0.013438473455607891, + 0.013449144549667835, + 0.0365341454744339, + 0.015934810042381287, + 0.02220352739095688, + 0.017949437722563744, + 0.01647329144179821, + 0.01872381754219532, + 0.00857787299901247, + 0.014116818085312843, + 0.01267787255346775, + 0.01771640218794346, + 0.01573696918785572, + 0.018869169056415558, + 0.015361023135483265, + 0.014283577911555767, + 0.013395766727626324, + 0.01380811259150505, + 0.01512244064360857, + 0.043336063623428345, + 0.01883457787334919, + 0.0270330049097538, + 0.021195709705352783, + 0.011638644151389599, + 0.03848639875650406, + 0.01799265295267105, + 0.006691867485642433, + 0.009200955741107464, + 0.007327525410801172, + 0.006204643286764622, + 0.006790487095713615, + 0.00621472392231226, + 0.007303990889340639, + 0.00751687353476882, + 0.00492726219817996, + 0.004880355671048164, + 0.007893215864896774, + 0.013802601024508476, + 0.003494712058454752, + 0.004537938628345728, + 0.019462572410702705, + 0.005204939749091864, + 0.005704937037080526, + 0.006398351397365332, + 0.009393876418471336, + 0.008600490167737007, + 0.004577793646603823, + 0.003394176485016942, + 0.007609914522618055, + 0.00527917267754674, + 0.010383139364421368, + 0.007645810022950172, + 0.010129928588867188, + 0.00835452787578106, + 0.008083813823759556, + 0.01047351211309433, + 0.013728720135986805, + 0.012446166947484016, + 0.006454430054873228, + 0.010356057435274124, + 0.004229638259857893, + 0.0065514082089066505, + 0.0063957879319787025, + 0.008181129582226276, + 0.005742608569562435, + 0.0069007668644189835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.048672039061784744 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019470559200271964 + ], + "opposite_logprob_tokens": [ + -6.251946926116943 + ], + "logprob_reward": -0.0019470559200271964, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019470559200271964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1501503735780716, + 0.07922026515007019, + 0.06583375483751297, + 0.15578453242778778, + 0.1013108491897583, + 0.05201295390725136, + 0.16748963296413422, + 0.05005606263875961, + 0.10347560793161392, + 0.07466592639684677 + ], + "credit_map": [ + 0.02847117930650711, + 0.01616564765572548, + 0.02387205697596073, + 0.017327992245554924, + 0.014612787403166294, + 0.02545056864619255, + 0.009929961524903774, + 0.0143201919272542, + 0.016019830480217934, + 0.005818940699100494, + 0.012194307520985603, + 0.009621546603739262, + 0.00814164336770773, + 0.007954961620271206, + 0.007395114284008741, + 0.012073923833668232, + 0.011524082161486149, + 0.008486174046993256, + 0.008147325366735458, + 0.008994204923510551, + 0.007805306930094957, + 0.010061465203762054, + 0.005396420136094093, + 0.005418786313384771, + 0.0158968735486269, + 0.022635668516159058, + 0.025633322075009346, + 0.021826734766364098, + 0.021039292216300964, + 0.01869945600628853, + 0.01331211905926466, + 0.0167410708963871, + 0.01008341833949089, + 0.013061968609690666, + 0.013188065029680729, + 0.01520926970988512, + 0.013161780312657356, + 0.010203770361840725, + 0.014257941395044327, + 0.012144642882049084, + 0.006466444116085768, + 0.006541575770825148, + 0.005562025588005781, + 0.005858769174665213, + 0.006361847743391991, + 0.004745903890579939, + 0.009500749409198761, + 0.0069756414741277695, + 0.01110313180834055, + 0.007714549079537392, + 0.01398509182035923, + 0.03369143605232239, + 0.008709189482033253, + 0.006476300302892923, + 0.06193658336997032, + 0.02387336827814579, + 0.005597063805907965, + 0.006585797294974327, + 0.010418846271932125, + 0.007118909154087305, + 0.004866874776780605, + 0.00470689544454217, + 0.006882205605506897, + 0.0038794782012701035, + 0.010974714532494545, + 0.010279963724315166, + 0.015265114605426788, + 0.015239464119076729, + 0.008609763346612453, + 0.00979463942348957, + 0.019517598673701286, + 0.013794363476336002, + 0.010629349388182163, + 0.009348265826702118, + 0.007006377913057804, + 0.009832865558564663, + 0.008973880670964718, + 0.010057380422949791, + 0.008877307176589966, + 0.009940503165125847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.0019470559200271964 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903281390666962 + ], + "opposite_logprob_tokens": [ + -2.4640328884124756 + ], + "logprob_reward": -0.08903281390666962, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903281390666962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041473936289548874, + 0.06115637347102165, + 0.11107856035232544, + 0.10824356228113174, + 0.13277101516723633, + 0.05161956697702408, + 0.22883392870426178, + 0.06756418943405151, + 0.12458078563213348, + 0.0726781114935875 + ], + "credit_map": [ + 0.006536435801535845, + 0.004041297361254692, + 0.005676198285073042, + 0.004437192343175411, + 0.004952830262482166, + 0.005036836490035057, + 0.004593187943100929, + 0.006199956871569157, + 0.007364985998719931, + 0.0040603578090667725, + 0.008920572698116302, + 0.01195274107158184, + 0.0063237459398806095, + 0.006686357781291008, + 0.007357683032751083, + 0.008489925414323807, + 0.022030623629689217, + 0.012872806750237942, + 0.016825200989842415, + 0.018240630626678467, + 0.010287067852914333, + 0.01142217218875885, + 0.009252282790839672, + 0.010147768072783947, + 0.012302840128540993, + 0.016844283789396286, + 0.019397160038352013, + 0.016029013320803642, + 0.011089208535850048, + 0.01181876566261053, + 0.010412081144750118, + 0.010350210592150688, + 0.01070502121001482, + 0.02764623425900936, + 0.012586286291480064, + 0.016517959535121918, + 0.017459025606513023, + 0.012880228459835052, + 0.025271344929933548, + 0.009704908356070518, + 0.005989191122353077, + 0.008077645674347878, + 0.006472774315625429, + 0.005592090077698231, + 0.005731646902859211, + 0.005162609275430441, + 0.007103024050593376, + 0.007490586955100298, + 0.016885245218873024, + 0.007910270243883133, + 0.015398931689560413, + 0.05118309333920479, + 0.009620703756809235, + 0.008315772749483585, + 0.08950235694646835, + 0.030017539858818054, + 0.008745474740862846, + 0.009562701918184757, + 0.010058690793812275, + 0.008411247283220291, + 0.0058763716369867325, + 0.007336371578276157, + 0.008745955303311348, + 0.008827370591461658, + 0.011022008955478668, + 0.012379362247884274, + 0.025322945788502693, + 0.021177731454372406, + 0.009905245155096054, + 0.011799784377217293, + 0.01249176636338234, + 0.020481932908296585, + 0.008536000736057758, + 0.010694294236600399, + 0.006762892939150333, + 0.007364123594015837, + 0.014080384746193886, + 0.008325339294970036, + 0.007480214815586805, + 0.009434866718947887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 8 + ], + "selected_reward": -0.08903281390666962 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809695690870285 + ], + "opposite_logprob_tokens": [ + -3.2880969047546387 + ], + "logprob_reward": -0.03809695690870285, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809695690870285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05234573408961296, + 0.19321398437023163, + 0.11272755265235901, + 0.16377051174640656, + 0.04054554924368858, + 0.022255193442106247, + 0.18153640627861023, + 0.13504721224308014, + 0.027395013719797134, + 0.07116290181875229 + ], + "credit_map": [ + 0.008035245351493359, + 0.006193253211677074, + 0.00649490300565958, + 0.005704839248210192, + 0.006860008463263512, + 0.007235093042254448, + 0.006465043872594833, + 0.0053573413752019405, + 0.05572592467069626, + 0.010893391445279121, + 0.03693098947405815, + 0.027886781841516495, + 0.014445220120251179, + 0.01808997057378292, + 0.013928338885307312, + 0.015313345938920975, + 0.02549782209098339, + 0.011066906154155731, + 0.01605546474456787, + 0.01493094488978386, + 0.010866275057196617, + 0.010966031812131405, + 0.007811196148395538, + 0.015532894060015678, + 0.016251491382718086, + 0.023771608248353004, + 0.019554700702428818, + 0.026508433744311333, + 0.021644264459609985, + 0.02004878968000412, + 0.018371259793639183, + 0.017619935795664787, + 0.004192216321825981, + 0.006133124697953463, + 0.005295473616570234, + 0.005491879303008318, + 0.003335891757160425, + 0.004689238034188747, + 0.0051687657833099365, + 0.006238957867026329, + 0.0024569150991737843, + 0.00348449288867414, + 0.0027736977208405733, + 0.0026866018306463957, + 0.0031125652603805065, + 0.002709709806367755, + 0.0025033089332282543, + 0.0025279000401496887, + 0.020902691408991814, + 0.009011481888592243, + 0.011616365984082222, + 0.03475780412554741, + 0.007444543298333883, + 0.006789777893573046, + 0.0632515698671341, + 0.027762167155742645, + 0.015349434688687325, + 0.016839925199747086, + 0.02806936204433441, + 0.021391524001955986, + 0.009614871814846992, + 0.011507651768624783, + 0.019271453842520714, + 0.013002977706491947, + 0.0029189654160290956, + 0.003328926395624876, + 0.003954081796109676, + 0.0029980377294123173, + 0.0022393057588487864, + 0.0033897091634571552, + 0.004078418482095003, + 0.004487565718591213, + 0.006536118686199188, + 0.013473635539412498, + 0.005737856030464172, + 0.008294820785522461, + 0.01357953529804945, + 0.006942339241504669, + 0.00885036401450634, + 0.007748227566480637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.03809695690870285 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful first-person point-of-view (POV) shot, with the camera tilted slightly downward and a slight handheld shake, simulates the human reaction of observation. Focused on a glossy, white kitchen tile floor, a monstrous centipede stands motionless. The creature's carapace is distinctly segmented, exhibiting an unsettling metallic sheen of electric blue and flame orange, interwoven with deep black stripes. Under the cold white light of the kitchen overhead, its smooth shell reflects sharp highlights, a texture so realistic that it appears as if each segment were coated with a layer of wet varnish. In an instant, as if sensing a threat, its entire body suddenly curls inward in a swift, spasmodic motion, its countless slender, sharp legs instantly folding together into a nearly impenetrable, defensive spiral, burying its head deep within. This sudden, decisive movement creates a stark contrast between the bright, clean background and the creature's grotesque, vibrant colors, creating a thrilling atmosphere of tension and terror.", + "seed": 17711, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_12/index_1427_seed_17711_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11811034558340908, + "vlm_logprob": -0.11811034558340908, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.11811034558340908, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1364666372537613 + ], + "opposite_logprob_tokens": [ + -2.1364667415618896 + ], + "logprob_reward": -0.1364666372537613, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1364666372537613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1404271125793457, + 0.09867754578590393, + 0.12194634974002838, + 0.2644819915294647, + 0.06176663562655449, + 0.05175835266709328, + 0.09404554218053818, + 0.0510488897562027, + 0.053544558584690094, + 0.06230296194553375 + ], + "credit_map": [ + 0.013895293697714806, + 0.009897083044052124, + 0.018522489815950394, + 0.013096659444272518, + 0.011446025222539902, + 0.018819576129317284, + 0.01888556219637394, + 0.03586443141102791, + 0.014507492072880268, + 0.009573087096214294, + 0.017185181379318237, + 0.01212537381798029, + 0.010751240886747837, + 0.013607154600322247, + 0.008597247302532196, + 0.012330785393714905, + 0.010909905657172203, + 0.011591902002692223, + 0.011285393498837948, + 0.01784350909292698, + 0.024023691192269325, + 0.017897196114063263, + 0.014272626489400864, + 0.014122141525149345, + 0.04193780571222305, + 0.0324527807533741, + 0.024178309366106987, + 0.03191789984703064, + 0.04377460107207298, + 0.03681822121143341, + 0.034653954207897186, + 0.018748434260487556, + 0.008337888866662979, + 0.009263049811124802, + 0.007124502211809158, + 0.006855278741568327, + 0.004365305881947279, + 0.009279381483793259, + 0.006110472604632378, + 0.01043076254427433, + 0.007401711773127317, + 0.005109712481498718, + 0.005312195047736168, + 0.005666788201779127, + 0.005710082594305277, + 0.007983771152794361, + 0.007808768656104803, + 0.006765326950699091, + 0.021440207958221436, + 0.007349816150963306, + 0.007055223919451237, + 0.007884791120886803, + 0.008230110630393028, + 0.024750178679823875, + 0.008702930063009262, + 0.008632294833660126, + 0.007639680989086628, + 0.006196730770170689, + 0.006077583879232407, + 0.005555085372179747, + 0.005476618651300669, + 0.0058539132587611675, + 0.005222403910011053, + 0.009026879444718361, + 0.009431292302906513, + 0.008022386580705643, + 0.0059431991539895535, + 0.003341937903314829, + 0.0049323816783726215, + 0.009682263247668743, + 0.004985986277461052, + 0.007205116096884012, + 0.007577631156891584, + 0.004371934104710817, + 0.011378582566976547, + 0.007119155954569578, + 0.004696074407547712, + 0.010241672396659851, + 0.005572252906858921, + 0.011345666833221912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.1364666372537613 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899825274944305 + ], + "opposite_logprob_tokens": [ + -2.578998327255249 + ], + "logprob_reward": -0.07899825274944305, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899825274944305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2501218020915985, + 0.10078100115060806, + 0.052431438118219376, + 0.22109568119049072, + 0.041822634637355804, + 0.02674657106399536, + 0.07823260873556137, + 0.045617278665304184, + 0.10255400091409683, + 0.08059697598218918 + ], + "credit_map": [ + 0.02036142721772194, + 0.013046489097177982, + 0.034179359674453735, + 0.02316911518573761, + 0.013387143611907959, + 0.035318516194820404, + 0.02629411406815052, + 0.08436563611030579, + 0.012978659942746162, + 0.009633957408368587, + 0.012867527082562447, + 0.016578948125243187, + 0.01040263194590807, + 0.014877733774483204, + 0.011619187891483307, + 0.01182235311716795, + 0.004182129167020321, + 0.004934311378747225, + 0.0045535992830991745, + 0.007382930722087622, + 0.010848434641957283, + 0.008103456348180771, + 0.005733465775847435, + 0.006693113129585981, + 0.037397753447294235, + 0.02760949917137623, + 0.021676186472177505, + 0.022483982145786285, + 0.04315213859081268, + 0.03305148705840111, + 0.02134585566818714, + 0.014378777705132961, + 0.00535980798304081, + 0.004540087655186653, + 0.00471438467502594, + 0.004036280792206526, + 0.004329223185777664, + 0.009121326729655266, + 0.003926278557628393, + 0.005795246455818415, + 0.0031077759340405464, + 0.002341534709557891, + 0.0028288408648222685, + 0.0037212655879557133, + 0.003947301767766476, + 0.0034301639534533024, + 0.004306575283408165, + 0.0030631127301603556, + 0.01537048164755106, + 0.005191095173358917, + 0.007361018564552069, + 0.006328597664833069, + 0.006841588299721479, + 0.01799606718122959, + 0.00904790684580803, + 0.010095849633216858, + 0.005001692101359367, + 0.004830575548112392, + 0.005659082904458046, + 0.005031476262956858, + 0.0045737917535007, + 0.006721037905663252, + 0.004157617688179016, + 0.009642005898058414, + 0.010281363502144814, + 0.011435536667704582, + 0.008392427116632462, + 0.006091184914112091, + 0.007748444098979235, + 0.02384071983397007, + 0.008112919516861439, + 0.026651401072740555, + 0.010377542115747929, + 0.007314234506338835, + 0.018537236377596855, + 0.009644058533012867, + 0.005612924229353666, + 0.007582307793200016, + 0.008007600903511047, + 0.01352106872946024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.07899825274944305 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031974404118955135 + ], + "opposite_logprob_tokens": [ + -5.75319766998291 + ], + "logprob_reward": -0.0031974404118955135, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031974404118955135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15835051238536835, + 0.09975311160087585, + 0.05082094669342041, + 0.15552736818790436, + 0.15121659636497498, + 0.04507438838481903, + 0.1215318888425827, + 0.05812432989478111, + 0.11626148223876953, + 0.04333941265940666 + ], + "credit_map": [ + 0.020619599148631096, + 0.01885303296148777, + 0.020465239882469177, + 0.015868069604039192, + 0.014945989474654198, + 0.018398597836494446, + 0.013158144429326057, + 0.03604184836149216, + 0.011217529885470867, + 0.008170213550329208, + 0.012700042687356472, + 0.011064532212913036, + 0.01605977676808834, + 0.014664221554994583, + 0.01099072489887476, + 0.014886068180203438, + 0.004345860797911882, + 0.007500028237700462, + 0.004359328653663397, + 0.006582011003047228, + 0.010060683824121952, + 0.007148186210542917, + 0.004687918815761805, + 0.006136931013315916, + 0.02624591626226902, + 0.021364198997616768, + 0.014348762109875679, + 0.017580917105078697, + 0.031163876876235008, + 0.02148437686264515, + 0.010904993861913681, + 0.012434320524334908, + 0.018849367275834084, + 0.01944291591644287, + 0.018634624779224396, + 0.016056541353464127, + 0.015033571980893612, + 0.020276596769690514, + 0.014262286946177483, + 0.028660695999860764, + 0.006773766595870256, + 0.005220122169703245, + 0.0047045862302184105, + 0.0053193699568510056, + 0.0036975310649722815, + 0.007337152026593685, + 0.0066387830302119255, + 0.005383077077567577, + 0.03849401697516441, + 0.006970858201384544, + 0.006375457160174847, + 0.006406617816537619, + 0.005082255229353905, + 0.042752690613269806, + 0.006391557864844799, + 0.009058430790901184, + 0.006054820958524942, + 0.006068745627999306, + 0.005709751974791288, + 0.00697825662791729, + 0.005686984863132238, + 0.007403010968118906, + 0.011256986297667027, + 0.00896577537059784, + 0.03738038241863251, + 0.006098210345953703, + 0.005123327951878309, + 0.0040369764901697636, + 0.004436623305082321, + 0.03848518431186676, + 0.005666438955813646, + 0.015034347772598267, + 0.00474083237349987, + 0.00341047253459692, + 0.00459522008895874, + 0.006327415816485882, + 0.003719993168488145, + 0.0065126982517540455, + 0.006031893193721771, + 0.008000885136425495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.0031974404118955135 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329389750957489 + ], + "opposite_logprob_tokens": [ + -3.773293972015381 + ], + "logprob_reward": -0.02329389750957489, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329389750957489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15460072457790375, + 0.08994977921247482, + 0.05047730728983879, + 0.22890910506248474, + 0.056821491569280624, + 0.03689359501004219, + 0.1429281383752823, + 0.16989411413669586, + 0.03732362017035484, + 0.03220216557383537 + ], + "credit_map": [ + 0.020197495818138123, + 0.014891228638589382, + 0.01618250645697117, + 0.01714383438229561, + 0.009552912786602974, + 0.02213922329246998, + 0.022991398349404335, + 0.0315021350979805, + 0.010793643072247505, + 0.01400034874677658, + 0.011030102148652077, + 0.01298731379210949, + 0.007390044163912535, + 0.010605628602206707, + 0.011539843864738941, + 0.011602857150137424, + 0.0044808173552155495, + 0.004997661802917719, + 0.005428650416433811, + 0.00791831873357296, + 0.009774511680006981, + 0.007476121187210083, + 0.005211105104535818, + 0.005190122872591019, + 0.02543899603188038, + 0.026161523535847664, + 0.024965830147266388, + 0.033254221081733704, + 0.028096362948417664, + 0.03138953819870949, + 0.03947269544005394, + 0.020129935815930367, + 0.008811183273792267, + 0.008615086786448956, + 0.0056917606852948666, + 0.007043647579848766, + 0.00494006322696805, + 0.008283210918307304, + 0.004619851242750883, + 0.00881668645888567, + 0.005868289619684219, + 0.003018817864358425, + 0.004249422810971737, + 0.004848291631788015, + 0.003183609340339899, + 0.00529678538441658, + 0.0054393610917031765, + 0.004989017266780138, + 0.02730242721736431, + 0.011751212179660797, + 0.015995386987924576, + 0.009972904808819294, + 0.015308846719563007, + 0.03089478239417076, + 0.013307745568454266, + 0.018394835293293, + 0.028455514460802078, + 0.024370022118091583, + 0.0177676510065794, + 0.021444594487547874, + 0.02392338216304779, + 0.020153870806097984, + 0.014116956852376461, + 0.019662130624055862, + 0.007719818968325853, + 0.0037877794820815325, + 0.002381286583840847, + 0.0029227116610854864, + 0.0019009464886039495, + 0.011483016423881054, + 0.003509787144139409, + 0.0036182741168886423, + 0.0036221069749444723, + 0.0036329447757452726, + 0.00379331037402153, + 0.006348031107336283, + 0.003101217793300748, + 0.004370140377432108, + 0.0035730882082134485, + 0.0037613280583173037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.02329389750957489 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person point-of-view (POV) with a slight handheld shake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilted slightly downward to focus on a glossy white kitchen tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monstrous centipede with a segmented carapace featuring electric blue, flame orange, and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the centipede perform a sudden motion where its body curls inward into a defensive spiral with legs folding together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cold white overhead kitchen light causing sharp highlights on the creature's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3485954999923706 + ], + "opposite_logprob_tokens": [ + -1.2235954999923706 + ], + "logprob_reward": -0.3485954999923706, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3485954999923706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34660133719444275, + 0.07756787538528442, + 0.05242088809609413, + 0.15666474401950836, + 0.09026607125997543, + 0.053070634603500366, + 0.06210556998848915, + 0.040463026612997055, + 0.05779733881354332, + 0.0630425214767456 + ], + "credit_map": [ + 0.0338946096599102, + 0.02237161621451378, + 0.04192128777503967, + 0.0367911197245121, + 0.020941127091646194, + 0.04591602459549904, + 0.051607489585876465, + 0.0931580513715744, + 0.010748676024377346, + 0.007587650790810585, + 0.010426869615912437, + 0.01307649351656437, + 0.010838414542376995, + 0.011442543938755989, + 0.006257486063987017, + 0.007189743220806122, + 0.0045522102154791355, + 0.004485383164137602, + 0.005072387401014566, + 0.009281272068619728, + 0.008934940211474895, + 0.007932215929031372, + 0.006116872187703848, + 0.006045603659003973, + 0.023915721103549004, + 0.02174665406346321, + 0.011539428494870663, + 0.013987723737955093, + 0.03205077722668648, + 0.02195839025080204, + 0.020976059138774872, + 0.010489992797374725, + 0.013310244306921959, + 0.014110847376286983, + 0.010866058059036732, + 0.009173684753477573, + 0.008795146830379963, + 0.01239778846502304, + 0.007216656114906073, + 0.014395644888281822, + 0.008013966493308544, + 0.004268554039299488, + 0.00500473752617836, + 0.00606239540502429, + 0.005419059190899134, + 0.008869100362062454, + 0.008915184997022152, + 0.006517641246318817, + 0.00739056384190917, + 0.006207524333149195, + 0.006808373145759106, + 0.006028113421052694, + 0.011372283101081848, + 0.006220931187272072, + 0.010113600641489029, + 0.007964182645082474, + 0.004095804411917925, + 0.004796119872480631, + 0.004831782076507807, + 0.004593317396938801, + 0.0036580911837518215, + 0.004826684948056936, + 0.004548290744423866, + 0.009112940169870853, + 0.006047728005796671, + 0.0066258469596505165, + 0.004518234636634588, + 0.0033095329999923706, + 0.003958956338465214, + 0.014958951622247696, + 0.003949317615479231, + 0.014428770169615746, + 0.007358703296631575, + 0.006300021428614855, + 0.013672946020960808, + 0.007294897921383381, + 0.004854968283325434, + 0.00538844196125865, + 0.0071624997071921825, + 0.011010045185685158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.3485954999923706 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_13/index_tensor([8143, 8143])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_13/index_tensor([8143, 8143])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2a572d7b7fd4e94293f420ad2f82bcb4fe7f0901 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_13/index_tensor([8143, 8143])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static shot, captured from a slight overhead angle, focuses squarely on a heavy black cast iron pan. A thin layer of clear cooking oil blankets the bottom, bubbling intensely under the intense heat, creating a continuous, melodious sizzling sound. A pair of wooden chopsticks carefully and steadily grasps one side of a square block of tender white tofu, delicately flipping it over with a single, fluid, and precise motion. As the tofu flips, the side facing downwards is revealed, a perfectly golden and crispy color. The tiny, uneven, crispy crust, formed by the intense heat, gleams enticingly under the warm kitchen overhead light. Then, the unfried side of the tofu, still retaining its creamy, delicate texture, meets the scalding oil, instantly erupting with an even more intense sound and sending forth tiny, shimmering splashes of oil. The entire scene is imbued with the dynamic beauty of cooking and the focused ritual of dedication, creating an atmosphere of warmth, healing, and a sense of life.", + "seed": 20293, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_13/index_8143_seed_20293_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.025076406169682743, + "vlm_logprob": -0.025076406169682743, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.025076406169682743, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03952953964471817 + ], + "opposite_logprob_tokens": [ + -3.53952956199646 + ], + "logprob_reward": -0.03952953964471817, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03952953964471817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1226700097322464, + 0.0854976624250412, + 0.07636941224336624, + 0.11500620096921921, + 0.04017516225576401, + 0.0841924250125885, + 0.26002973318099976, + 0.15830695629119873, + 0.03446570411324501, + 0.02328670583665371 + ], + "credit_map": [ + 0.02759597823023796, + 0.030996762216091156, + 0.02090480551123619, + 0.008246849291026592, + 0.010807928629219532, + 0.00694824755191803, + 0.006657900754362345, + 0.010511538945138454, + 0.016339300200343132, + 0.012987788766622543, + 0.019549399614334106, + 0.009348887950181961, + 0.005657721310853958, + 0.0051040248945355415, + 0.007928143255412579, + 0.008582399226725101, + 0.007208333816379309, + 0.004564379807561636, + 0.014015586115419865, + 0.00934691447764635, + 0.0045162225142121315, + 0.0022160487715154886, + 0.013278660364449024, + 0.021223271265625954, + 0.019971176981925964, + 0.013728775084018707, + 0.02043737843632698, + 0.01634315960109234, + 0.009225084446370602, + 0.006670825183391571, + 0.01104711089283228, + 0.017582684755325317, + 0.006398157216608524, + 0.004827238153666258, + 0.0056789037771523, + 0.0030552998650819063, + 0.007618858944624662, + 0.00298930867575109, + 0.004426502622663975, + 0.005180892068892717, + 0.015278181061148643, + 0.010581659153103828, + 0.02143893763422966, + 0.007920264266431332, + 0.008772071450948715, + 0.004274475388228893, + 0.0062237875536084175, + 0.009703048504889011, + 0.04920196160674095, + 0.02007754147052765, + 0.06490669399499893, + 0.020625799894332886, + 0.031800802797079086, + 0.011739437468349934, + 0.02572277933359146, + 0.03595472127199173, + 0.02453051321208477, + 0.013024885207414627, + 0.021569058299064636, + 0.015119612216949463, + 0.01650596410036087, + 0.014659776352345943, + 0.017112083733081818, + 0.03578506410121918, + 0.004981193691492081, + 0.003692364552989602, + 0.007093736436218023, + 0.0032814061269164085, + 0.005046023987233639, + 0.0023078275844454765, + 0.0032400700729340315, + 0.004823080264031887, + 0.0035435990430414677, + 0.0031916426960378885, + 0.0021813984494656324, + 0.0032775045838207006, + 0.0026357672177255154, + 0.001331172650679946, + 0.004504204262048006, + 0.00262141739949584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.03952953964471817 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02638857439160347 + ], + "opposite_logprob_tokens": [ + -3.651388645172119 + ], + "logprob_reward": -0.02638857439160347, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638857439160347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10108177363872528, + 0.11886382848024368, + 0.13261571526527405, + 0.13340511918067932, + 0.0697857066988945, + 0.10289980471134186, + 0.11453146487474442, + 0.14860543608665466, + 0.030746102333068848, + 0.047465045005083084 + ], + "credit_map": [ + 0.02233581617474556, + 0.024312319234013557, + 0.022435825318098068, + 0.006415912415832281, + 0.006775927264243364, + 0.006239458452910185, + 0.008027559146285057, + 0.004538960754871368, + 0.02385847084224224, + 0.027802014723420143, + 0.023729028180241585, + 0.009376579895615578, + 0.008038748055696487, + 0.00822976790368557, + 0.006841900758445263, + 0.010987316258251667, + 0.015446296893060207, + 0.009092608466744423, + 0.02388104237616062, + 0.017497509717941284, + 0.009333737194538116, + 0.0063140192069113255, + 0.020058169960975647, + 0.030992332845926285, + 0.028692809864878654, + 0.009312228299677372, + 0.025117188692092896, + 0.007885267026722431, + 0.011117103509604931, + 0.005059625022113323, + 0.008489028550684452, + 0.037731874734163284, + 0.00974995270371437, + 0.008150779642164707, + 0.010933835990726948, + 0.010558288544416428, + 0.007272182032465935, + 0.003029121784493327, + 0.010356905870139599, + 0.009734650142490864, + 0.02005983702838421, + 0.008029799908399582, + 0.022792572155594826, + 0.007405107840895653, + 0.015875758603215218, + 0.00707798870280385, + 0.006663523614406586, + 0.01499522477388382, + 0.018689312040805817, + 0.012265613302588463, + 0.018834281712770462, + 0.00897021684795618, + 0.017898881807923317, + 0.013566930778324604, + 0.010039841756224632, + 0.014266395010054111, + 0.0233883298933506, + 0.015299074351787567, + 0.017074016854166985, + 0.011951012536883354, + 0.025522395968437195, + 0.013339567929506302, + 0.018068674951791763, + 0.02396237477660179, + 0.005143787711858749, + 0.0036053124349564314, + 0.004445952828973532, + 0.003187265945598483, + 0.004884566646069288, + 0.0018821810372173786, + 0.0032431427389383316, + 0.004353893455117941, + 0.0076622734777629375, + 0.00932414922863245, + 0.006514045875519514, + 0.004067698959261179, + 0.0053576030768454075, + 0.003304289421066642, + 0.004873987752944231, + 0.006361004430800676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.02638857439160347 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006756199989467859 + ], + "opposite_logprob_tokens": [ + -5.00675630569458 + ], + "logprob_reward": -0.006756199989467859, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006756199989467859, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3840767741203308, + 0.0368276946246624, + 0.06301059573888779, + 0.10198146104812622, + 0.13836364448070526, + 0.08246293663978577, + 0.0635870099067688, + 0.08218201249837875, + 0.031215276569128036, + 0.016292627900838852 + ], + "credit_map": [ + 0.07565126568078995, + 0.10033576935529709, + 0.054401494562625885, + 0.025325430557131767, + 0.04213595390319824, + 0.02021385170519352, + 0.03666692599654198, + 0.029346082359552383, + 0.007047952618449926, + 0.004327500704675913, + 0.00945971254259348, + 0.004401415586471558, + 0.002915942342951894, + 0.003898245980963111, + 0.0023889155127108097, + 0.0023880109656602144, + 0.006124409846961498, + 0.004552362021058798, + 0.010745738632977009, + 0.006191300693899393, + 0.005058246664702892, + 0.005801747553050518, + 0.009119903668761253, + 0.015416889451444149, + 0.014830104075372219, + 0.009657655842602253, + 0.012571629136800766, + 0.01934320665895939, + 0.00949771422892809, + 0.006779755465686321, + 0.014530646614730358, + 0.0147707499563694, + 0.019481679424643517, + 0.018305445089936256, + 0.02319202572107315, + 0.020488671958446503, + 0.01169414073228836, + 0.007299463730305433, + 0.01924917846918106, + 0.018653038889169693, + 0.01766842044889927, + 0.0054555488750338554, + 0.02652047760784626, + 0.006226222030818462, + 0.006972497329115868, + 0.004470005165785551, + 0.005179912317544222, + 0.009969847276806831, + 0.011016016826033592, + 0.006233407650142908, + 0.011456104926764965, + 0.006338374689221382, + 0.004557950422167778, + 0.003963993396610022, + 0.010252850130200386, + 0.009768307209014893, + 0.015264632180333138, + 0.0110918078571558, + 0.009333343245089054, + 0.007300886325538158, + 0.015415321104228497, + 0.004730344749987125, + 0.006917413789778948, + 0.012128259055316448, + 0.00467611663043499, + 0.0038614957593381405, + 0.003767328104004264, + 0.003409456927329302, + 0.006284274626523256, + 0.0020246391650289297, + 0.003268617205321789, + 0.00392334908246994, + 0.002652966184541583, + 0.0023181820288300514, + 0.001954730600118637, + 0.0023793361615389585, + 0.0018755601486191154, + 0.0011120717972517014, + 0.002554596634581685, + 0.0014451846946030855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.006756199989467859 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026360591873526573 + ], + "opposite_logprob_tokens": [ + -3.651360511779785 + ], + "logprob_reward": -0.026360591873526573, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026360591873526573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15661302208900452, + 0.033638305962085724, + 0.06167178228497505, + 0.16327255964279175, + 0.07544723898172379, + 0.18994711339473724, + 0.1680404394865036, + 0.08837271481752396, + 0.03166868910193443, + 0.03132816404104233 + ], + "credit_map": [ + 0.026854179799556732, + 0.027579842135310173, + 0.013178718276321888, + 0.006346109323203564, + 0.04337138682603836, + 0.01644853502511978, + 0.011118164286017418, + 0.011716083623468876, + 0.005166580434888601, + 0.004151052329689264, + 0.0048889280296862125, + 0.0032737827859818935, + 0.004126715939491987, + 0.005331810563802719, + 0.00312166940420866, + 0.003577765543013811, + 0.005785828921943903, + 0.004559091292321682, + 0.00993337668478489, + 0.006924693007022142, + 0.00433966051787138, + 0.0051210555247962475, + 0.009831276722252369, + 0.015176800079643726, + 0.029889391735196114, + 0.019300593063235283, + 0.027928275987505913, + 0.014267776161432266, + 0.014370860531926155, + 0.010110369883477688, + 0.013670033775269985, + 0.033735256642103195, + 0.014486119151115417, + 0.006812448613345623, + 0.018009981140494347, + 0.007257494609802961, + 0.00591809768229723, + 0.0037492027040570974, + 0.005364833865314722, + 0.01384905818849802, + 0.04225209355354309, + 0.012117041274905205, + 0.06779737025499344, + 0.014864590018987656, + 0.012353207916021347, + 0.008196826092898846, + 0.008579907938838005, + 0.02378607913851738, + 0.02834807150065899, + 0.016191978007555008, + 0.0341789610683918, + 0.014653309248387814, + 0.02647898904979229, + 0.007803009822964668, + 0.021114539355039597, + 0.01927158422768116, + 0.014831682667136192, + 0.011464794166386127, + 0.013229059055447578, + 0.007822426967322826, + 0.012726681306958199, + 0.005643211305141449, + 0.008477413095533848, + 0.014177449978888035, + 0.004422302357852459, + 0.00394430523738265, + 0.004480003844946623, + 0.003150895470753312, + 0.0061479066498577595, + 0.0018838507821783423, + 0.003823372535407543, + 0.00381605071015656, + 0.004915004130452871, + 0.00545028131455183, + 0.0023286473006010056, + 0.003984019625931978, + 0.004456629045307636, + 0.0020287949591875076, + 0.004394985735416412, + 0.0037698012311011553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.026360591873526573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347124949097633 + ], + "opposite_logprob_tokens": [ + -3.6513471603393555 + ], + "logprob_reward": -0.026347124949097633, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347124949097633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07795366644859314, + 0.12283461540937424, + 0.06982750445604324, + 0.1595822423696518, + 0.08723319321870804, + 0.07781695574522018, + 0.24156306684017181, + 0.08899649977684021, + 0.04530738666653633, + 0.028884869068861008 + ], + "credit_map": [ + 0.012062490917742252, + 0.01052522286772728, + 0.01123469416052103, + 0.006786307319998741, + 0.01153262797743082, + 0.0073561095632612705, + 0.010825056582689285, + 0.007631154265254736, + 0.024425847455859184, + 0.023860028013586998, + 0.027620406821370125, + 0.010516015812754631, + 0.010006783530116081, + 0.007282284554094076, + 0.008678159676492214, + 0.010445090010762215, + 0.008466370403766632, + 0.005212122108787298, + 0.01149748358875513, + 0.009976573288440704, + 0.006296411622315645, + 0.004116055555641651, + 0.010081116110086441, + 0.014181370846927166, + 0.022331615909934044, + 0.02358928509056568, + 0.02067628689110279, + 0.03416191041469574, + 0.011348179541528225, + 0.009084442630410194, + 0.01944444514811039, + 0.018946083262562752, + 0.015991030260920525, + 0.00725418608635664, + 0.013195984996855259, + 0.007898866198956966, + 0.0115746995434165, + 0.008256564848124981, + 0.005920866038650274, + 0.01714099571108818, + 0.015129848383367062, + 0.006899033207446337, + 0.017102982848882675, + 0.006652192212641239, + 0.006503747310489416, + 0.004007353447377682, + 0.006161051336675882, + 0.015360747464001179, + 0.0411430299282074, + 0.024856489151716232, + 0.03786659240722656, + 0.020214665681123734, + 0.049025312066078186, + 0.014646516181528568, + 0.025612113997340202, + 0.02819833904504776, + 0.014509012922644615, + 0.01363152451813221, + 0.012250111438333988, + 0.00813157856464386, + 0.014934957027435303, + 0.006799451075494289, + 0.007571155205368996, + 0.011168710887432098, + 0.008003673516213894, + 0.004674708470702171, + 0.006213332526385784, + 0.0032090051099658012, + 0.008797178976237774, + 0.003137860679998994, + 0.004670573864132166, + 0.006601054687052965, + 0.004140014760196209, + 0.003448792500421405, + 0.004646191373467445, + 0.003232192248106003, + 0.004186692647635937, + 0.003610026789829135, + 0.0022713197395205498, + 0.0033496387768536806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.026347124949097633 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot, captured from a slight overhead angle, focuses squarely on a heavy black cast iron pan. A thin layer of clear cooking oil blankets the bottom, bubbling intensely under the intense heat, creating a continuous, melodious sizzling sound. A pair of wooden chopsticks carefully and steadily grasps one side of a square block of tender white tofu, delicately flipping it over with a single, fluid, and precise motion. As the tofu flips, the side facing downwards is revealed, a perfectly golden and crispy color. The tiny, uneven, crispy crust, formed by the intense heat, gleams enticingly under the warm kitchen overhead light. Then, the unfried side of the tofu, still retaining its creamy, delicate texture, meets the scalding oil, instantly erupting with an even more intense sound and sending forth tiny, shimmering splashes of oil. The entire scene is imbued with the dynamic beauty of cooking and the focused ritual of dedication, creating an atmosphere of warmth, healing, and a sense of life.", + "seed": 20294, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_13/index_8143_seed_20294_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.031978411972522734, + "vlm_logprob": -0.031978411972522734, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.031978411972522734, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08231445401906967 + ], + "opposite_logprob_tokens": [ + -2.8323144912719727 + ], + "logprob_reward": -0.08231445401906967, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08231445401906967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17848815023899078, + 0.10678189247846603, + 0.04173930361866951, + 0.053916770964860916, + 0.05614854767918587, + 0.08242311328649521, + 0.07612951844930649, + 0.12063340842723846, + 0.1308625042438507, + 0.15287688374519348 + ], + "credit_map": [ + 0.028008367866277695, + 0.010754161514341831, + 0.05739202722907066, + 0.012484495528042316, + 0.008872350677847862, + 0.011218315921723843, + 0.011429162696003914, + 0.03832925856113434, + 0.012712672352790833, + 0.006490266881883144, + 0.016225850209593773, + 0.010062905959784985, + 0.02422560378909111, + 0.008202871307730675, + 0.011536745354533195, + 0.01732497662305832, + 0.005444901529699564, + 0.0027123040053993464, + 0.007575637195259333, + 0.0046257865615189075, + 0.006839554291218519, + 0.0034676494542509317, + 0.0040894122794270515, + 0.006984056439250708, + 0.007662865798920393, + 0.002245964016765356, + 0.00563048105686903, + 0.003808327717706561, + 0.003946204669773579, + 0.003752480261027813, + 0.008594107814133167, + 0.018276339396834373, + 0.0055255102925002575, + 0.0017339609330520034, + 0.005943950265645981, + 0.004521599970757961, + 0.0027613246347755194, + 0.0021920199505984783, + 0.017645960673689842, + 0.015824222937226295, + 0.011708664707839489, + 0.005491016432642937, + 0.01938796415925026, + 0.007200733758509159, + 0.005990482866764069, + 0.004941828083246946, + 0.017675895243883133, + 0.010026522912085056, + 0.00767195550724864, + 0.00611779373139143, + 0.01573115959763527, + 0.009464584290981293, + 0.013953515328466892, + 0.004977586213499308, + 0.006570794153958559, + 0.011642126366496086, + 0.009173359721899033, + 0.015388509258627892, + 0.013856983743607998, + 0.025477532297372818, + 0.022995566949248314, + 0.00644893990829587, + 0.01387377642095089, + 0.013418740592896938, + 0.012378009036183357, + 0.010914849117398262, + 0.01823505572974682, + 0.024702368304133415, + 0.0137283680960536, + 0.010160218924283981, + 0.017354032024741173, + 0.023389603942632675, + 0.011241204105317593, + 0.015246342867612839, + 0.015257144346833229, + 0.04146396368741989, + 0.019141247496008873, + 0.010812886990606785, + 0.020586300641298294, + 0.01912778615951538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.08231445401906967 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01611294411122799 + ], + "opposite_logprob_tokens": [ + -4.141112804412842 + ], + "logprob_reward": -0.01611294411122799, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01611294411122799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2313142716884613, + 0.052038323134183884, + 0.026896124705672264, + 0.05179877579212189, + 0.3688241243362427, + 0.027212446555495262, + 0.10188459604978561, + 0.053585849702358246, + 0.030842147767543793, + 0.05560331419110298 + ], + "credit_map": [ + 0.03984557092189789, + 0.015663033351302147, + 0.06453048437833786, + 0.027445239946246147, + 0.012518104165792465, + 0.017071861773729324, + 0.017313173040747643, + 0.03692680224776268, + 0.005011213012039661, + 0.0039534918032586575, + 0.005786076188087463, + 0.003581067081540823, + 0.014220261946320534, + 0.0035620350390672684, + 0.004956633783876896, + 0.01096754614263773, + 0.0023555904626846313, + 0.0020276294089853764, + 0.0031229802407324314, + 0.002494209911674261, + 0.007572989445179701, + 0.0017937475349754095, + 0.003759954357519746, + 0.003769021714106202, + 0.005573047325015068, + 0.001338689005933702, + 0.004531112965196371, + 0.004033433273434639, + 0.004126417450606823, + 0.0023701421450823545, + 0.01556624285876751, + 0.014259693212807178, + 0.07194662094116211, + 0.010178202763199806, + 0.0410824716091156, + 0.011499239131808281, + 0.00919895526021719, + 0.0172777958214283, + 0.05489369109272957, + 0.15274712443351746, + 0.005465715657919645, + 0.0014620668953284621, + 0.007892420515418053, + 0.00400154385715723, + 0.0012116105062887073, + 0.0009096633875742555, + 0.002443841192871332, + 0.0038255846593528986, + 0.009946988895535469, + 0.005147565621882677, + 0.03051929362118244, + 0.008316094987094402, + 0.023431947454810143, + 0.0036601873580366373, + 0.0059280660934746265, + 0.014934447593986988, + 0.004182802978903055, + 0.006971197668462992, + 0.005600378382951021, + 0.010442961007356644, + 0.009736218489706516, + 0.00328329810872674, + 0.007185212802141905, + 0.0061837793327867985, + 0.003039844101294875, + 0.002402894664555788, + 0.00334266759455204, + 0.007650571875274181, + 0.0030200835317373276, + 0.002633704338222742, + 0.004117763135582209, + 0.00463462108746171, + 0.0034420322626829147, + 0.005841386504471302, + 0.005241043400019407, + 0.014050581492483616, + 0.009229425340890884, + 0.005506663583219051, + 0.0068716853857040405, + 0.005420494358986616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.01611294411122799 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005278697237372398 + ], + "opposite_logprob_tokens": [ + -5.255278587341309 + ], + "logprob_reward": -0.005278697237372398, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005278697237372398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10353844612836838, + 0.10126946121454239, + 0.1746201366186142, + 0.018192294985055923, + 0.29779112339019775, + 0.023131515830755234, + 0.038718465715646744, + 0.07774284482002258, + 0.08856859058141708, + 0.0764271467924118 + ], + "credit_map": [ + 0.02659142203629017, + 0.00552326999604702, + 0.017162170261144638, + 0.009884314611554146, + 0.005128865595906973, + 0.006425245199352503, + 0.007790242787450552, + 0.02503291703760624, + 0.010218213312327862, + 0.008025860413908958, + 0.00989343598484993, + 0.007690833881497383, + 0.028137294575572014, + 0.008324961178004742, + 0.010057764127850533, + 0.018921097740530968, + 0.014280348084867, + 0.015262309461832047, + 0.018751274794340134, + 0.012861926108598709, + 0.050962429493665695, + 0.012565210461616516, + 0.02038269117474556, + 0.02955394797027111, + 0.0016112856101244688, + 0.0010753904934972525, + 0.002228556666523218, + 0.002330557443201542, + 0.0019282613648101687, + 0.0009727327851578593, + 0.005211121402680874, + 0.0028343903832137585, + 0.05293760448694229, + 0.007603830192238092, + 0.032574404031038284, + 0.010767919942736626, + 0.007489004172384739, + 0.013192974030971527, + 0.05361306294798851, + 0.11961235105991364, + 0.004095502197742462, + 0.00117787707131356, + 0.0034606358967721462, + 0.002901565982028842, + 0.0019018420716747642, + 0.001444130321033299, + 0.0054204706102609634, + 0.0027294917963445187, + 0.004153518006205559, + 0.004033053293824196, + 0.0055321757681667805, + 0.005020963493734598, + 0.004468761384487152, + 0.004114231094717979, + 0.0055627464316785336, + 0.005833015311509371, + 0.006716075353324413, + 0.007205369416624308, + 0.009118636138737202, + 0.01421114057302475, + 0.011463332921266556, + 0.005647622514516115, + 0.011950526386499405, + 0.01143014244735241, + 0.007731358986347914, + 0.00653735501691699, + 0.009363739751279354, + 0.020340530201792717, + 0.009832261130213737, + 0.006600298453122377, + 0.012154790572822094, + 0.016008250415325165, + 0.0060164849273860455, + 0.007805990986526012, + 0.006777135655283928, + 0.013972231186926365, + 0.012739401310682297, + 0.009183602407574654, + 0.009009272791445255, + 0.010923027992248535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.005278697237372398 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029832802712917328 + ], + "opposite_logprob_tokens": [ + -3.5298328399658203 + ], + "logprob_reward": -0.029832802712917328, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029832802712917328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2680976092815399, + 0.03757021948695183, + 0.01780249923467636, + 0.02596057392656803, + 0.13026294112205505, + 0.03578201308846474, + 0.11822494119405746, + 0.09823907911777496, + 0.1227545216679573, + 0.1453056037425995 + ], + "credit_map": [ + 0.04095078632235527, + 0.019169876351952553, + 0.07019232958555222, + 0.03399154543876648, + 0.017677901312708855, + 0.024081986397504807, + 0.0215020589530468, + 0.04053112864494324, + 0.00432747695595026, + 0.0028015358839184046, + 0.006204415578395128, + 0.003557222429662943, + 0.008446716703474522, + 0.002557269297540188, + 0.0044605010189116, + 0.0052150823175907135, + 0.001993522047996521, + 0.0009572806884534657, + 0.0021448968909680843, + 0.0014642541063949466, + 0.003667353419587016, + 0.001245655119419098, + 0.0018796281656250358, + 0.004449910018593073, + 0.003964236006140709, + 0.0012404798762872815, + 0.0036605321802198887, + 0.0024200109764933586, + 0.0019442620687186718, + 0.001792613766156137, + 0.00376566918566823, + 0.007172769401222467, + 0.03209058940410614, + 0.004983110819011927, + 0.014634199440479279, + 0.004172770772129297, + 0.004505441058427095, + 0.007131614722311497, + 0.006378496997058392, + 0.05636671930551529, + 0.0077383569441735744, + 0.0010446679079905152, + 0.012617907486855984, + 0.002687414875254035, + 0.0014037168584764004, + 0.0014399576466530561, + 0.0035265968181192875, + 0.005323393736034632, + 0.010789002291858196, + 0.007811830844730139, + 0.03168323636054993, + 0.012500422075390816, + 0.023529479280114174, + 0.004930843599140644, + 0.011311371810734272, + 0.015668760985136032, + 0.008550665341317654, + 0.00917110312730074, + 0.012797970324754715, + 0.018166158348321915, + 0.019317690283060074, + 0.005820377264171839, + 0.011094380170106888, + 0.013320731930434704, + 0.011258535087108612, + 0.010634336620569229, + 0.013706468045711517, + 0.026378987357020378, + 0.012869643978774548, + 0.00940956175327301, + 0.01681262068450451, + 0.021684370934963226, + 0.011068334802985191, + 0.0132828988134861, + 0.013991539366543293, + 0.03530612960457802, + 0.0205391775816679, + 0.014067078940570354, + 0.018581047654151917, + 0.01846940442919731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.029832802712917328 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635316178202629 + ], + "opposite_logprob_tokens": [ + -3.651353120803833 + ], + "logprob_reward": -0.02635316178202629, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635316178202629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07860302180051804, + 0.14149193465709686, + 0.06512023508548737, + 0.061509858816862106, + 0.12239427119493484, + 0.11185633391141891, + 0.08832836151123047, + 0.1741379201412201, + 0.08944718539714813, + 0.06711092591285706 + ], + "credit_map": [ + 0.01179509237408638, + 0.004759073723107576, + 0.022971557453274727, + 0.00850258581340313, + 0.004166615195572376, + 0.005710752680897713, + 0.00602634297683835, + 0.01467099878937006, + 0.01391428429633379, + 0.010942823253571987, + 0.016839120537042618, + 0.009174508973956108, + 0.040919892489910126, + 0.010827786289155483, + 0.013624188490211964, + 0.02524932101368904, + 0.005295726004987955, + 0.006137458141893148, + 0.007475303951650858, + 0.0049200416542589664, + 0.020586490631103516, + 0.005066475365310907, + 0.008077318780124187, + 0.007561422418802977, + 0.006729475222527981, + 0.001496967044658959, + 0.006200807634741068, + 0.0033175514545291662, + 0.0033138717990368605, + 0.002400362631306052, + 0.018005764111876488, + 0.02004506066441536, + 0.026241334155201912, + 0.0059800660237669945, + 0.009379776194691658, + 0.01066654548048973, + 0.0076578957960009575, + 0.004697904456406832, + 0.0362585075199604, + 0.021512234583497047, + 0.023414330556988716, + 0.004825260024517775, + 0.035022467374801636, + 0.010960252024233341, + 0.006586139090359211, + 0.004694486036896706, + 0.0088967839255929, + 0.01745660975575447, + 0.009359510615468025, + 0.0037691688630729914, + 0.026043174788355827, + 0.008425368927419186, + 0.018925605341792107, + 0.003159938845783472, + 0.005068112164735794, + 0.013577479869127274, + 0.01304022129625082, + 0.02345980703830719, + 0.021069081500172615, + 0.03454162925481796, + 0.029291577637195587, + 0.011300205253064632, + 0.016741249710321426, + 0.024694150313735008, + 0.007572900503873825, + 0.006963626481592655, + 0.010136399418115616, + 0.019765598699450493, + 0.009269733913242817, + 0.005952148698270321, + 0.013590680435299873, + 0.016196096315979958, + 0.0040047746151685715, + 0.005733837373554707, + 0.005730192642658949, + 0.022734282538294792, + 0.010099099017679691, + 0.004324542824178934, + 0.010375951416790485, + 0.004108243156224489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.02635316178202629 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot, captured from a slight overhead angle, focuses squarely on a heavy black cast iron pan. A thin layer of clear cooking oil blankets the bottom, bubbling intensely under the intense heat, creating a continuous, melodious sizzling sound. A pair of wooden chopsticks carefully and steadily grasps one side of a square block of tender white tofu, delicately flipping it over with a single, fluid, and precise motion. As the tofu flips, the side facing downwards is revealed, a perfectly golden and crispy color. The tiny, uneven, crispy crust, formed by the intense heat, gleams enticingly under the warm kitchen overhead light. Then, the unfried side of the tofu, still retaining its creamy, delicate texture, meets the scalding oil, instantly erupting with an even more intense sound and sending forth tiny, shimmering splashes of oil. The entire scene is imbued with the dynamic beauty of cooking and the focused ritual of dedication, creating an atmosphere of warmth, healing, and a sense of life.", + "seed": 20295, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_13/index_8143_seed_20295_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06260394975543022, + "vlm_logprob": -0.06260394975543022, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.06260394975543022, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23597250878810883 + ], + "opposite_logprob_tokens": [ + -1.61097252368927 + ], + "logprob_reward": -0.23597250878810883, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23597250878810883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09601453691720963, + 0.0671766996383667, + 0.28541216254234314, + 0.062143474817276, + 0.11381743103265762, + 0.029512764886021614, + 0.24119971692562103, + 0.03914881497621536, + 0.013271871022880077, + 0.0523025207221508 + ], + "credit_map": [ + 0.005534001160413027, + 0.019465837627649307, + 0.009416064247488976, + 0.011787006631493568, + 0.013971143402159214, + 0.0050952075980603695, + 0.018296726047992706, + 0.01244855485856533, + 0.004098160192370415, + 0.006756166461855173, + 0.005529576446861029, + 0.013581315986812115, + 0.015842344611883163, + 0.004914131481200457, + 0.01175836194306612, + 0.004696642514318228, + 0.017601357772946358, + 0.030811872333288193, + 0.014966779388487339, + 0.06878958642482758, + 0.09702041000127792, + 0.012890543788671494, + 0.028674596920609474, + 0.01465702150017023, + 0.007131676189601421, + 0.009172195568680763, + 0.005947257857769728, + 0.010520247742533684, + 0.005234392825514078, + 0.005069577135145664, + 0.006918394006788731, + 0.012149736285209656, + 0.006040120963007212, + 0.018040893599390984, + 0.007581178564578295, + 0.01739552430808544, + 0.028441615402698517, + 0.008154962211847305, + 0.022841431200504303, + 0.005321716424077749, + 0.0030783922411501408, + 0.0034207147546112537, + 0.0021326972637325525, + 0.008533664979040623, + 0.0031545041128993034, + 0.0020406856201589108, + 0.003405170515179634, + 0.0037469379603862762, + 0.01491963118314743, + 0.05327385663986206, + 0.007499776314944029, + 0.08872775733470917, + 0.024688590317964554, + 0.011759425513446331, + 0.031098254024982452, + 0.009232442826032639, + 0.00353793753311038, + 0.00369269959628582, + 0.0025310434866696596, + 0.009106534533202648, + 0.005496283993124962, + 0.006069105584174395, + 0.005772255826741457, + 0.0029429588466882706, + 0.0012653806479647756, + 0.001671340549364686, + 0.00108660280238837, + 0.0018196332966908813, + 0.0021814743522554636, + 0.0012403699802234769, + 0.0019379734294489026, + 0.0020690972451120615, + 0.00434849550947547, + 0.005250261630862951, + 0.005140526220202446, + 0.0087570920586586, + 0.013174137100577354, + 0.005717157386243343, + 0.00525146909058094, + 0.004663386382162571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 4 + ], + "selected_reward": -0.23597250878810883 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011116004548966885 + ], + "opposite_logprob_tokens": [ + -4.511116027832031 + ], + "logprob_reward": -0.011116004548966885, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011116004548966885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03399984911084175, + 0.08015192300081253, + 0.30333229899406433, + 0.07162874191999435, + 0.2353142499923706, + 0.01494379062205553, + 0.09691775590181351, + 0.11576103419065475, + 0.02510221116244793, + 0.022848114371299744 + ], + "credit_map": [ + 0.002431401750072837, + 0.007877475582063198, + 0.0028059440664947033, + 0.005022350698709488, + 0.004673731047660112, + 0.002492304425686598, + 0.0054279593750834465, + 0.003268678905442357, + 0.004565682262182236, + 0.009800306521356106, + 0.006221944000571966, + 0.014928920194506645, + 0.023627357557415962, + 0.0047880373895168304, + 0.009866056963801384, + 0.006353616248816252, + 0.01878337748348713, + 0.033247027546167374, + 0.01705906167626381, + 0.06786518543958664, + 0.09835626929998398, + 0.014118277467787266, + 0.035198286175727844, + 0.018704818561673164, + 0.00844611693173647, + 0.010135903023183346, + 0.0068266019225120544, + 0.011570006608963013, + 0.00745194498449564, + 0.005992872640490532, + 0.007247753906995058, + 0.01395754236727953, + 0.011906396597623825, + 0.03922280669212341, + 0.009808043949306011, + 0.036205776035785675, + 0.06130945682525635, + 0.01604105345904827, + 0.050316937267780304, + 0.010503789409995079, + 0.0014753062278032303, + 0.001864398131147027, + 0.0014660694869235158, + 0.002068108879029751, + 0.0030530320946127176, + 0.001279162708669901, + 0.002132039749994874, + 0.0016056732274591923, + 0.006074110046029091, + 0.02154381386935711, + 0.003233799245208502, + 0.039392806589603424, + 0.006125915329903364, + 0.004221655428409576, + 0.011501111090183258, + 0.004824544303119183, + 0.013659089803695679, + 0.012145129963755608, + 0.0067391106858849525, + 0.024936778470873833, + 0.017455995082855225, + 0.018245941027998924, + 0.01570153422653675, + 0.00687746237963438, + 0.00135098397731781, + 0.008615397848188877, + 0.0008257395238615572, + 0.005595515947788954, + 0.0028290573973208666, + 0.0017477436922490597, + 0.002770734252408147, + 0.0013670389307662845, + 0.002156865317374468, + 0.0050370898097753525, + 0.0010769589571282268, + 0.003457252634689212, + 0.004510100465267897, + 0.0023771561682224274, + 0.0030459153931587934, + 0.0011867774883285165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.011116004548966885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008660022169351578 + ], + "opposite_logprob_tokens": [ + -4.758659839630127 + ], + "logprob_reward": -0.008660022169351578, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008660022169351578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10771378874778748, + 0.10667585581541061, + 0.1234569326043129, + 0.0580926388502121, + 0.07703248411417007, + 0.036758460104465485, + 0.3038807511329651, + 0.05557470768690109, + 0.09284593909978867, + 0.0379684641957283 + ], + "credit_map": [ + 0.008355691097676754, + 0.021764252334833145, + 0.008439063094556332, + 0.01616535894572735, + 0.01843670941889286, + 0.006431215442717075, + 0.019289525225758553, + 0.008831976912915707, + 0.006665468215942383, + 0.012544006109237671, + 0.0069185406900942326, + 0.02324717305600643, + 0.025672590360045433, + 0.007301495876163244, + 0.017953386530280113, + 0.006373193114995956, + 0.006032156758010387, + 0.010386544279754162, + 0.005329106468707323, + 0.03961827605962753, + 0.04124850034713745, + 0.005228804424405098, + 0.011177429929375648, + 0.0044361199252307415, + 0.003945548087358475, + 0.00797425489872694, + 0.004026197828352451, + 0.015019294805824757, + 0.009947219863533974, + 0.0036326865665614605, + 0.007291823625564575, + 0.0062556141056120396, + 0.0057867723517119884, + 0.011058412492275238, + 0.005111084319651127, + 0.01221300009638071, + 0.01818809099495411, + 0.0049803671427071095, + 0.014353260397911072, + 0.005341490264981985, + 0.003529596608132124, + 0.003996274434030056, + 0.004214281216263771, + 0.006773156113922596, + 0.004354450851678848, + 0.0023976119700819254, + 0.006352260708808899, + 0.005140829831361771, + 0.01863865926861763, + 0.06565408408641815, + 0.009138351306319237, + 0.10912546515464783, + 0.03561960905790329, + 0.01587606593966484, + 0.03864714130759239, + 0.011181356385350227, + 0.006289076525717974, + 0.006891017779707909, + 0.00652351975440979, + 0.009330119006335735, + 0.007938575930893421, + 0.005347295198589563, + 0.007860193029046059, + 0.0053949072025716305, + 0.00702805258333683, + 0.021296516060829163, + 0.006700852420181036, + 0.012103412300348282, + 0.012243310920894146, + 0.01048076618462801, + 0.011999138630926609, + 0.010993891395628452, + 0.0031309854239225388, + 0.005377594381570816, + 0.0026355194859206676, + 0.006580382585525513, + 0.00800777692347765, + 0.0023037283681333065, + 0.006266288459300995, + 0.0036661862395703793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.008660022169351578 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209755696356297 + ], + "opposite_logprob_tokens": [ + -4.264209747314453 + ], + "logprob_reward": -0.014209755696356297, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209755696356297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18825571238994598, + 0.09710942208766937, + 0.13463379442691803, + 0.04808700084686279, + 0.06730902194976807, + 0.036513812839984894, + 0.1421038657426834, + 0.21711468696594238, + 0.028468534350395203, + 0.04040409252047539 + ], + "credit_map": [ + 0.010014694184064865, + 0.065213143825531, + 0.009513403289020061, + 0.030419686809182167, + 0.030589589849114418, + 0.008163802325725555, + 0.022486165165901184, + 0.011855232529342175, + 0.005933495704084635, + 0.017342370003461838, + 0.006266972981393337, + 0.01923026144504547, + 0.011809666641056538, + 0.00888867024332285, + 0.01926586963236332, + 0.008372110314667225, + 0.008413701318204403, + 0.012013958767056465, + 0.006390679627656937, + 0.05492943525314331, + 0.024822482839226723, + 0.004924488253891468, + 0.015604891814291477, + 0.007534154690802097, + 0.00373650761321187, + 0.008582402020692825, + 0.004600603133440018, + 0.005101273767650127, + 0.006914117373526096, + 0.005011918488889933, + 0.00869189202785492, + 0.0054482826963067055, + 0.004352905787527561, + 0.011915775015950203, + 0.004458709619939327, + 0.010873042047023773, + 0.011699485592544079, + 0.005003719124943018, + 0.01465540286153555, + 0.004349985625594854, + 0.003338788403198123, + 0.004771981853991747, + 0.0039890846237540245, + 0.004337722435593605, + 0.005033624824136496, + 0.0037319830153137445, + 0.007435433100908995, + 0.003875193651765585, + 0.009268400259315968, + 0.027003509923815727, + 0.004802503623068333, + 0.05379585921764374, + 0.014342083595693111, + 0.007083249744027853, + 0.018079357221722603, + 0.007728902157396078, + 0.025773629546165466, + 0.021904706954956055, + 0.013190719299018383, + 0.04056466743350029, + 0.03728680685162544, + 0.0345923975110054, + 0.03095994144678116, + 0.012841816060245037, + 0.0018287546699866652, + 0.0031740013509988785, + 0.0018593925051391125, + 0.004625605419278145, + 0.005611561704427004, + 0.002084679203107953, + 0.00534789776429534, + 0.003936640452593565, + 0.004047829657793045, + 0.0065875304862856865, + 0.003040408482775092, + 0.0063339355401694775, + 0.005227222573012114, + 0.0036449977196753025, + 0.008943396620452404, + 0.002578770276159048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.014209755696356297 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306145757436752 + ], + "opposite_logprob_tokens": [ + -3.1680614948272705 + ], + "logprob_reward": -0.04306145757436752, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306145757436752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08857107162475586, + 0.051294341683387756, + 0.08068300783634186, + 0.11732882261276245, + 0.06489153951406479, + 0.02661222778260708, + 0.308748722076416, + 0.13597190380096436, + 0.10254562646150589, + 0.023352688178420067 + ], + "credit_map": [ + 0.004325575195252895, + 0.029891103506088257, + 0.005549149587750435, + 0.014788093976676464, + 0.014741136692464352, + 0.0032687222119420767, + 0.009948763996362686, + 0.006058535072952509, + 0.0032418735790997744, + 0.008531330153346062, + 0.00404798099771142, + 0.010335156694054604, + 0.007012397050857544, + 0.0042141517624258995, + 0.00920682493597269, + 0.004704630468040705, + 0.0067824567668139935, + 0.012729011476039886, + 0.004515215288847685, + 0.009592748247087002, + 0.027299251407384872, + 0.004547163378447294, + 0.009415541775524616, + 0.005801624618470669, + 0.01552401389926672, + 0.013744809664785862, + 0.01063507329672575, + 0.02224023826420307, + 0.009570485912263393, + 0.010127849876880646, + 0.0112824821844697, + 0.024203883484005928, + 0.0038753680419176817, + 0.00921189971268177, + 0.004318494815379381, + 0.009288031607866287, + 0.014685249887406826, + 0.005065691191703081, + 0.012842921540141106, + 0.00560388620942831, + 0.0024213180877268314, + 0.0034445333294570446, + 0.0021847407333552837, + 0.0028282993007451296, + 0.006215738132596016, + 0.002311518881469965, + 0.003875410184264183, + 0.003330671228468418, + 0.019978046417236328, + 0.07293805480003357, + 0.010273627936840057, + 0.11550924181938171, + 0.02742239460349083, + 0.013860530219972134, + 0.03606114536523819, + 0.01270569022744894, + 0.014108393341302872, + 0.011118078604340553, + 0.01886884495615959, + 0.026104487478733063, + 0.020478321239352226, + 0.014074763283133507, + 0.01895270310342312, + 0.01226632110774517, + 0.007647627033293247, + 0.029309267178177834, + 0.005839316640049219, + 0.013293257914483547, + 0.01665346324443817, + 0.010578311048448086, + 0.011747465468943119, + 0.007476926315575838, + 0.0015771089820191264, + 0.0054529751650989056, + 0.001098540611565113, + 0.003255167743191123, + 0.004953102208673954, + 0.0019573436584323645, + 0.0036113678943365812, + 0.0014470848254859447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.04306145757436752 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot, captured from a slight overhead angle, focuses squarely on a heavy black cast iron pan. A thin layer of clear cooking oil blankets the bottom, bubbling intensely under the intense heat, creating a continuous, melodious sizzling sound. A pair of wooden chopsticks carefully and steadily grasps one side of a square block of tender white tofu, delicately flipping it over with a single, fluid, and precise motion. As the tofu flips, the side facing downwards is revealed, a perfectly golden and crispy color. The tiny, uneven, crispy crust, formed by the intense heat, gleams enticingly under the warm kitchen overhead light. Then, the unfried side of the tofu, still retaining its creamy, delicate texture, meets the scalding oil, instantly erupting with an even more intense sound and sending forth tiny, shimmering splashes of oil. The entire scene is imbued with the dynamic beauty of cooking and the focused ritual of dedication, creating an atmosphere of warmth, healing, and a sense of life.", + "seed": 20296, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_13/index_8143_seed_20296_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.015451657585799694, + "vlm_logprob": -0.015451657585799694, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.015451657585799694, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030038390308618546 + ], + "opposite_logprob_tokens": [ + -3.780038356781006 + ], + "logprob_reward": -0.030038390308618546, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.030038390308618546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1391269713640213, + 0.08229172229766846, + 0.07810446619987488, + 0.04973924532532692, + 0.16029047966003418, + 0.15996049344539642, + 0.06169002875685692, + 0.09450656920671463, + 0.07076054066419601, + 0.1035294458270073 + ], + "credit_map": [ + 0.01639203168451786, + 0.011540418490767479, + 0.020174749195575714, + 0.01806854084134102, + 0.012479275465011597, + 0.012247703969478607, + 0.014660567045211792, + 0.03356368467211723, + 0.00913437083363533, + 0.013917587697505951, + 0.012579343281686306, + 0.008665615692734718, + 0.009705265052616596, + 0.010290919803082943, + 0.008381454274058342, + 0.009617166593670845, + 0.009313171729445457, + 0.008681907318532467, + 0.004823477938771248, + 0.011491559445858002, + 0.0061874231323599815, + 0.007763613481074572, + 0.022482501342892647, + 0.007360812276601791, + 0.004355727694928646, + 0.0046088965609669685, + 0.0046444423496723175, + 0.009643405675888062, + 0.005367395933717489, + 0.006054754368960857, + 0.010288192890584469, + 0.00477642985060811, + 0.0318169929087162, + 0.012441882863640785, + 0.00971304252743721, + 0.02001296728849411, + 0.011836639605462551, + 0.014373046346008778, + 0.04796693101525307, + 0.012128971517086029, + 0.022432642057538033, + 0.01008581928908825, + 0.007709864992648363, + 0.018606793135404587, + 0.01113307848572731, + 0.01728004962205887, + 0.056265704333782196, + 0.0164465494453907, + 0.005880479700863361, + 0.004422615747898817, + 0.004456906113773584, + 0.01144526805728674, + 0.005862252786755562, + 0.011242149397730827, + 0.01355062797665596, + 0.004829731769859791, + 0.007749677170068026, + 0.0067182136699557304, + 0.009192668832838535, + 0.02162047103047371, + 0.0072582922875881195, + 0.017712416127324104, + 0.01651579700410366, + 0.007739044260233641, + 0.010753259062767029, + 0.004151099361479282, + 0.004521017894148827, + 0.007889795117080212, + 0.0052766562439501286, + 0.00922086276113987, + 0.021746162325143814, + 0.007201685570180416, + 0.010056711733341217, + 0.015176587738096714, + 0.015993228182196617, + 0.016833532601594925, + 0.009949452243745327, + 0.011095613241195679, + 0.012439661659300327, + 0.011984662152826786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.030038390308618546 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020623063668608665 + ], + "opposite_logprob_tokens": [ + -3.895622968673706 + ], + "logprob_reward": -0.020623063668608665, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020623063668608665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16372138261795044, + 0.05164269730448723, + 0.05594164878129959, + 0.04817327484488487, + 0.17056871950626373, + 0.22879323363304138, + 0.03240557014942169, + 0.11875549703836441, + 0.06294098496437073, + 0.0670570358633995 + ], + "credit_map": [ + 0.020255403593182564, + 0.015138721093535423, + 0.02705208584666252, + 0.014673082157969475, + 0.01677132584154606, + 0.022233599796891212, + 0.011864189989864826, + 0.035732951015233994, + 0.005026243627071381, + 0.007116842083632946, + 0.005474753677845001, + 0.006090336479246616, + 0.008112902753055096, + 0.00782308354973793, + 0.005509000737220049, + 0.0064895316027104855, + 0.006455376744270325, + 0.009667904116213322, + 0.004881778731942177, + 0.006746317725628614, + 0.00871960073709488, + 0.005708556156605482, + 0.00773313082754612, + 0.00602897722274065, + 0.005014361813664436, + 0.004220475442707539, + 0.004364706575870514, + 0.005542011000216007, + 0.005709903780370951, + 0.0057849762961268425, + 0.011928458698093891, + 0.005608375649899244, + 0.026770442724227905, + 0.009508238174021244, + 0.010215101763606071, + 0.022173557430505753, + 0.012177087366580963, + 0.015976138412952423, + 0.06066908687353134, + 0.01307904627174139, + 0.02835877612233162, + 0.010309306904673576, + 0.009609936736524105, + 0.02318016067147255, + 0.009442581795156002, + 0.03991340845823288, + 0.08820492774248123, + 0.01977410912513733, + 0.004668575711548328, + 0.0024519118014723063, + 0.0028315375093370676, + 0.005243197083473206, + 0.0023786292877048254, + 0.0041780318133533, + 0.006693271920084953, + 0.003960411064326763, + 0.01035741064697504, + 0.0062430077232420444, + 0.006275929976254702, + 0.03596905991435051, + 0.011475752107799053, + 0.01177580188959837, + 0.027447868138551712, + 0.009210658259689808, + 0.009674841538071632, + 0.00401296466588974, + 0.0036891186609864235, + 0.007070683408528566, + 0.0038295150734484196, + 0.005964422132819891, + 0.021517930552363396, + 0.007181499619036913, + 0.00731186056509614, + 0.006684836465865374, + 0.013923694379627705, + 0.008255889639258385, + 0.007042299956083298, + 0.0070495326071977615, + 0.00831657089293003, + 0.008472342975437641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.020623063668608665 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005270159337669611 + ], + "opposite_logprob_tokens": [ + -5.255270004272461 + ], + "logprob_reward": -0.005270159337669611, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005270159337669611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10384571552276611, + 0.05634313076734543, + 0.07375600934028625, + 0.021557187661528587, + 0.258832186460495, + 0.18695199489593506, + 0.02479931153357029, + 0.12216158211231232, + 0.11999323964118958, + 0.03175969421863556 + ], + "credit_map": [ + 0.013090507127344608, + 0.006028938107192516, + 0.011006499640643597, + 0.010203855112195015, + 0.009409165009856224, + 0.01754789799451828, + 0.008277099579572678, + 0.028281738981604576, + 0.005342292599380016, + 0.00792360957711935, + 0.006798395421355963, + 0.0060750748962163925, + 0.007687221746891737, + 0.006872403901070356, + 0.008984946645796299, + 0.006659182719886303, + 0.008379687555134296, + 0.0074146585538983345, + 0.006227247882634401, + 0.010650060139596462, + 0.012041866779327393, + 0.010368868708610535, + 0.010774972848594189, + 0.007898642681539059, + 0.0023275932762771845, + 0.002356543205678463, + 0.0018587999511510134, + 0.003693370381370187, + 0.0024586666841059923, + 0.0032642045989632607, + 0.0035290776286274195, + 0.002068930771201849, + 0.054220907390117645, + 0.017053576186299324, + 0.012861806899309158, + 0.02703755348920822, + 0.01522132195532322, + 0.021401895210146904, + 0.09460128098726273, + 0.01643383875489235, + 0.03162999451160431, + 0.011982609517872334, + 0.007279357872903347, + 0.02098057232797146, + 0.007671710569411516, + 0.025002576410770416, + 0.06913086771965027, + 0.013274278491735458, + 0.0021268511191010475, + 0.0014703435590490699, + 0.001521811354905367, + 0.0033276337198913097, + 0.0023102466948330402, + 0.005864529870450497, + 0.005664448719471693, + 0.002513442886993289, + 0.01153256930410862, + 0.00874507799744606, + 0.01091302651911974, + 0.02709464728832245, + 0.012567484751343727, + 0.016600171104073524, + 0.022793618962168694, + 0.01191496942192316, + 0.020344631746411324, + 0.006743757985532284, + 0.005512220785021782, + 0.012771778739988804, + 0.008703852072358131, + 0.014567854814231396, + 0.039418309926986694, + 0.01193082146346569, + 0.0031011716928333044, + 0.005662111099809408, + 0.002978728385642171, + 0.004618003964424133, + 0.004370932001620531, + 0.004900206811726093, + 0.0032305908389389515, + 0.0028979438357055187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.005270159337669611 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258894059807062 + ], + "opposite_logprob_tokens": [ + -5.255259037017822 + ], + "logprob_reward": -0.005258894059807062, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258894059807062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05682310461997986, + 0.0505501963198185, + 0.13187412917613983, + 0.053749047219753265, + 0.30530649423599243, + 0.08603142946958542, + 0.10994606465101242, + 0.0408855602145195, + 0.10150406509637833, + 0.06332986056804657 + ], + "credit_map": [ + 0.00789682287722826, + 0.006639702245593071, + 0.007887435145676136, + 0.0059650251641869545, + 0.006815811153501272, + 0.006760156247764826, + 0.005252031609416008, + 0.009606122970581055, + 0.005361804738640785, + 0.0077620311640203, + 0.007892832159996033, + 0.006410460919141769, + 0.004553177859634161, + 0.005903291516005993, + 0.005698244087398052, + 0.006968359462916851, + 0.015433148480951786, + 0.016496188938617706, + 0.007459662854671478, + 0.022014988586306572, + 0.015755927190184593, + 0.011282265186309814, + 0.03235546499490738, + 0.011076501570641994, + 0.006057966500520706, + 0.006438497453927994, + 0.004104250110685825, + 0.011190544813871384, + 0.005231085233390331, + 0.005652441177517176, + 0.010101482272148132, + 0.004972782917320728, + 0.06261007487773895, + 0.020106973126530647, + 0.015031810849905014, + 0.03431650251150131, + 0.017568083480000496, + 0.02434263564646244, + 0.11292938143014908, + 0.018401049077510834, + 0.00761417718604207, + 0.005114350467920303, + 0.006381466053426266, + 0.013172294944524765, + 0.008752115070819855, + 0.008993146941065788, + 0.024830274283885956, + 0.011173606850206852, + 0.014220846816897392, + 0.008982645347714424, + 0.017216583713889122, + 0.01645607501268387, + 0.007570226676762104, + 0.012199070304632187, + 0.021195881068706512, + 0.012104745022952557, + 0.004725976847112179, + 0.002121571684256196, + 0.003049644408747554, + 0.012265998870134354, + 0.002697161165997386, + 0.004452344961464405, + 0.008562028408050537, + 0.0030108392238616943, + 0.017887476831674576, + 0.005670815706253052, + 0.005126754753291607, + 0.010580427944660187, + 0.005423172377049923, + 0.009465961717069149, + 0.03611140325665474, + 0.011238057166337967, + 0.005907305981963873, + 0.007315977476537228, + 0.006424120161682367, + 0.009232766926288605, + 0.008887647651135921, + 0.010941280983388424, + 0.007331349421292543, + 0.007289416156709194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.005258894059807062 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a heavy black cast iron pan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a slight overhead view with a static shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do wooden chopsticks flip a square block of white tofu in a single fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the underside of the tofu revealed as perfectly golden and crispy after the flip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tofu's unfried side meet the oil and cause tiny shimmering splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067780554294586 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.016067780554294586, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067780554294586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10661615431308746, + 0.04923698306083679, + 0.13112737238407135, + 0.0957287847995758, + 0.1350252330303192, + 0.09255111962556839, + 0.05331290140748024, + 0.1761547178030014, + 0.09939780831336975, + 0.060848936438560486 + ], + "credit_map": [ + 0.015330967493355274, + 0.009864049032330513, + 0.011934414505958557, + 0.011309079825878143, + 0.007994605228304863, + 0.011690988205373287, + 0.008094929158687592, + 0.030397122725844383, + 0.004911559168249369, + 0.006186891812831163, + 0.005336516536772251, + 0.008408908732235432, + 0.006864228285849094, + 0.007146866992115974, + 0.004415424074977636, + 0.005966586526483297, + 0.0124774519354105, + 0.02286645583808422, + 0.009640026837587357, + 0.01726231351494789, + 0.009956616908311844, + 0.012016749940812588, + 0.03509965538978577, + 0.011808106675744057, + 0.009891169145703316, + 0.012869521975517273, + 0.015240322798490524, + 0.013196228072047234, + 0.011111542582511902, + 0.010137422941625118, + 0.012795689515769482, + 0.010486887767910957, + 0.02311684936285019, + 0.007547808811068535, + 0.008294685743749142, + 0.016314368695020676, + 0.008695957250893116, + 0.011806794442236423, + 0.04998554289340973, + 0.009263219311833382, + 0.013624632731080055, + 0.007857322692871094, + 0.005026719532907009, + 0.009882982820272446, + 0.004337910562753677, + 0.013943467289209366, + 0.024754606187343597, + 0.01312347687780857, + 0.006527580786496401, + 0.004424238111823797, + 0.00722433440387249, + 0.0066169449128210545, + 0.0037471679970622063, + 0.012252471409738064, + 0.008314171805977821, + 0.004205991048365831, + 0.015402169898152351, + 0.012338737957179546, + 0.014195454306900501, + 0.04798988252878189, + 0.014349384233355522, + 0.019954144954681396, + 0.037524159997701645, + 0.014400779269635677, + 0.015573225915431976, + 0.006069624330848455, + 0.00582797871902585, + 0.010226598009467125, + 0.006014678161591291, + 0.00904153659939766, + 0.03524963930249214, + 0.011394526809453964, + 0.00651975441724062, + 0.011646740138530731, + 0.010845055803656578, + 0.0070812818594276905, + 0.006230453494936228, + 0.005417545326054096, + 0.005934364628046751, + 0.007173740770667791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.016067780554294586 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_14/index_tensor([9419, 9419])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_14/index_tensor([9419, 9419])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c43c5cd8827ee0bee6dfb730b7def62f1ec24dbb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_14/index_tensor([9419, 9419])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot from a static, top-down perspective, the scene unfolds like a living Chinese landscape painting. A long, unbroken column of red and black ants traverses the surface of a large, mossy boulder, rendered not with photorealistic detail but rather through soft, wet-on-wet ink washes on the canvas, with charcoal and gray tones blending seamlessly to suggest an ancient, weathered texture. Patches of moss are depicted as blooming, spreading areas of pale celadon and deep green, their colors softly seeping into the damp rock surface. Against this atmospheric backdrop, the ants themselves are sharp calligraphic figures, each a vibrant stroke of deep black or vibrant crimson ink. Their slender legs and antennae are delicate, vibrant lines, giving them a sense of precise, rhythmic movement as they steadily and deliberately carry tiny, pale ochre seeds. The entire work is bathed in the diffuse, ethereal light characteristic of scroll painting, creating a profound atmosphere of tranquil concentration and minimalist elegance. The sharp, fluid energy of the ants forms a striking contrast with the soft, contemplative tranquility of the ink world.", + "seed": 12386, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_14/index_9419_seed_12386_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06945866728201508, + "vlm_logprob": -0.06945866728201508, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.06945866728201508, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12494243681430817 + ], + "opposite_logprob_tokens": [ + -2.2499425411224365 + ], + "logprob_reward": -0.12494243681430817, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12494243681430817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23596149682998657, + 0.12523990869522095, + 0.07566887140274048, + 0.08837806433439255, + 0.09010940045118332, + 0.07246403396129608, + 0.11340980231761932, + 0.06989538669586182, + 0.08399025350809097, + 0.04488284885883331 + ], + "credit_map": [ + 0.0195784792304039, + 0.022594498470425606, + 0.019290262833237648, + 0.017853835597634315, + 0.07034240663051605, + 0.06301817297935486, + 0.009727704338729382, + 0.013556103222072124, + 0.0094365943223238, + 0.017725517973303795, + 0.012012011371552944, + 0.01409352570772171, + 0.030107878148555756, + 0.019182737916707993, + 0.011893791146576405, + 0.01078784093260765, + 0.007323798723518848, + 0.009502960368990898, + 0.008780419826507568, + 0.010023802518844604, + 0.014379407279193401, + 0.009056122973561287, + 0.0100112808868289, + 0.006591066252440214, + 0.00834784097969532, + 0.010402018204331398, + 0.009582283906638622, + 0.007813573814928532, + 0.012645229697227478, + 0.021543040871620178, + 0.00901538971811533, + 0.009028684347867966, + 0.022593697533011436, + 0.009149571880698204, + 0.008720720186829567, + 0.00780477374792099, + 0.010136574506759644, + 0.01274340134114027, + 0.006882773712277412, + 0.012077880091965199, + 0.01395588368177414, + 0.009198164567351341, + 0.009131872095167637, + 0.0075825899839401245, + 0.008156905882060528, + 0.010448207147419453, + 0.005405003670603037, + 0.00858539342880249, + 0.011116190813481808, + 0.014507459476590157, + 0.009217941202223301, + 0.012819139286875725, + 0.025247277691960335, + 0.01793770305812359, + 0.012405604124069214, + 0.0101584792137146, + 0.00840681605041027, + 0.010436082258820534, + 0.0054715899750590324, + 0.006453768815845251, + 0.01374280359596014, + 0.015891702845692635, + 0.004414636641740799, + 0.005077978130429983, + 0.011172553524374962, + 0.010473004542291164, + 0.008542414754629135, + 0.008795669302344322, + 0.015294043347239494, + 0.010413323529064655, + 0.008093444630503654, + 0.011205796152353287, + 0.006243460346013308, + 0.004428591579198837, + 0.005054989363998175, + 0.005162909161299467, + 0.006154571194201708, + 0.008925148285925388, + 0.0033013001084327698, + 0.005611874628812075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.12494243681430817 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004109272733330727 + ], + "opposite_logprob_tokens": [ + -5.5041093826293945 + ], + "logprob_reward": -0.004109272733330727, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004109272733330727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27110469341278076, + 0.08318088203668594, + 0.0678662434220314, + 0.07446730881929398, + 0.11674036085605621, + 0.13530175387859344, + 0.04811699688434601, + 0.11807233840227127, + 0.03801443427801132, + 0.04713500663638115 + ], + "credit_map": [ + 0.019217856228351593, + 0.028743715956807137, + 0.02311566099524498, + 0.01917283609509468, + 0.0828254371881485, + 0.06703970581293106, + 0.013156825676560402, + 0.017832670360803604, + 0.01186162419617176, + 0.014012870378792286, + 0.009397240355610847, + 0.007718577515333891, + 0.01685996539890766, + 0.008928648196160793, + 0.006930419243872166, + 0.0074715339578688145, + 0.007518282625824213, + 0.010179068893194199, + 0.008453949354588985, + 0.00926604587584734, + 0.010596889071166515, + 0.009331295266747475, + 0.006433947943150997, + 0.0060867611318826675, + 0.007953274063766003, + 0.006376185454428196, + 0.0075100259855389595, + 0.007525978609919548, + 0.008165823295712471, + 0.011733965948224068, + 0.012299388647079468, + 0.012902669608592987, + 0.01734805665910244, + 0.023317262530326843, + 0.009539591148495674, + 0.010456390678882599, + 0.014511101879179478, + 0.014235317707061768, + 0.011796415783464909, + 0.015536223538219929, + 0.01471335906535387, + 0.014080461114645004, + 0.012624874711036682, + 0.00893157348036766, + 0.03115212917327881, + 0.03142048791050911, + 0.00887709204107523, + 0.01350177638232708, + 0.005764753092080355, + 0.003908500075340271, + 0.006637080106884241, + 0.0043062553741037846, + 0.010460580699145794, + 0.005797047633677721, + 0.006044668145477772, + 0.005198111291974783, + 0.01256754994392395, + 0.014721301384270191, + 0.011194342747330666, + 0.020071962848305702, + 0.02141505852341652, + 0.01999462954699993, + 0.008605264127254486, + 0.009502223692834377, + 0.004831263329833746, + 0.0037416997365653515, + 0.0038013362791389227, + 0.0033779777586460114, + 0.0065197027288377285, + 0.004344934597611427, + 0.004874683916568756, + 0.006522835697978735, + 0.005937569309026003, + 0.008766512386500835, + 0.0029005438555032015, + 0.0037180783692747355, + 0.01001072209328413, + 0.007030979264527559, + 0.003927712328732014, + 0.004842886235564947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.004109272733330727 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990289628505707 + ], + "opposite_logprob_tokens": [ + -1.8049029111862183 + ], + "logprob_reward": -0.17990289628505707, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990289628505707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42295148968696594, + 0.040294479578733444, + 0.17064988613128662, + 0.02882484346628189, + 0.11018142104148865, + 0.13524721562862396, + 0.0301941130310297, + 0.017285889014601707, + 0.024281056597828865, + 0.020089594647288322 + ], + "credit_map": [ + 0.054605744779109955, + 0.05835316330194473, + 0.049622416496276855, + 0.0561629980802536, + 0.07071961462497711, + 0.06389153748750687, + 0.03704993799328804, + 0.03254610672593117, + 0.004314221907407045, + 0.008200816810131073, + 0.0034282151609659195, + 0.0036406703293323517, + 0.008351541124284267, + 0.004937433637678623, + 0.0031862978357821703, + 0.00423528254032135, + 0.013126109726727009, + 0.018344642594456673, + 0.015339716337621212, + 0.019827663898468018, + 0.02756330370903015, + 0.04014229029417038, + 0.02148740366101265, + 0.014818752184510231, + 0.0024605882354080677, + 0.004964551422744989, + 0.005126442294567823, + 0.004806592594832182, + 0.0030940219294279814, + 0.003129142802208662, + 0.0026242141611874104, + 0.0026192902587354183, + 0.024714797735214233, + 0.009724005125463009, + 0.013905932195484638, + 0.013174003921449184, + 0.012297110632061958, + 0.017291251569986343, + 0.008141186088323593, + 0.01093313004821539, + 0.01342013105750084, + 0.009990161284804344, + 0.008499899879097939, + 0.008356411941349506, + 0.033415939658880234, + 0.030783193185925484, + 0.010005914606153965, + 0.0207755696028471, + 0.003583099227398634, + 0.004323332104831934, + 0.0027614908758550882, + 0.0032826841343194246, + 0.006286062765866518, + 0.0034151545260101557, + 0.003133981255814433, + 0.003408310003578663, + 0.0016927669057622552, + 0.0020876170601695776, + 0.0019197161309421062, + 0.0024084390606731176, + 0.0032134416978806257, + 0.0025171497836709023, + 0.0016529981512576342, + 0.0017937591765075922, + 0.0025427844375371933, + 0.0024243607185781, + 0.0031842158641666174, + 0.0024615786969661713, + 0.004107179585844278, + 0.0031726211309432983, + 0.002518330467864871, + 0.003869985695928335, + 0.002065763808786869, + 0.0015276421327143908, + 0.002379397861659527, + 0.0021238825283944607, + 0.0037365329917520285, + 0.004241651855409145, + 0.0019440504256635904, + 0.0020706732757389545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.17990289628505707 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033698562532663345 + ], + "opposite_logprob_tokens": [ + -3.408698558807373 + ], + "logprob_reward": -0.033698562532663345, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033698562532663345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20761890709400177, + 0.09820836782455444, + 0.0740123763680458, + 0.11957312375307083, + 0.22077423334121704, + 0.06253919005393982, + 0.08333110809326172, + 0.04750054329633713, + 0.02718392014503479, + 0.059258244931697845 + ], + "credit_map": [ + 0.01839803345501423, + 0.02234317921102047, + 0.015211179852485657, + 0.015200961381196976, + 0.06103671342134476, + 0.05269889160990715, + 0.011058014817535877, + 0.011671933345496655, + 0.007935019209980965, + 0.013995641842484474, + 0.013092393055558205, + 0.01138624269515276, + 0.011323421262204647, + 0.020692715421319008, + 0.009237989783287048, + 0.010544944554567337, + 0.008476180024445057, + 0.013461881317198277, + 0.007290580775588751, + 0.008218665607273579, + 0.012007172219455242, + 0.007638786919414997, + 0.0070454878732562065, + 0.00987362302839756, + 0.010211952961981297, + 0.010564154013991356, + 0.015578439459204674, + 0.013504908420145512, + 0.013073277659714222, + 0.02300994284451008, + 0.01604086346924305, + 0.017589585855603218, + 0.019599026069045067, + 0.029644886031746864, + 0.017419807612895966, + 0.026402492076158524, + 0.030718469992280006, + 0.02772882580757141, + 0.04086454212665558, + 0.02839619107544422, + 0.0090279970318079, + 0.007486499380320311, + 0.005612564273178577, + 0.006977423559874296, + 0.00679375883191824, + 0.006100360304117203, + 0.007982226088643074, + 0.012558365240693092, + 0.00947375874966383, + 0.011463968083262444, + 0.008513694629073143, + 0.00861682090908289, + 0.017663121223449707, + 0.011267182417213917, + 0.008609593845903873, + 0.007722968235611916, + 0.004502379801124334, + 0.00593161815777421, + 0.0036102451849728823, + 0.005278684664517641, + 0.009196390397846699, + 0.009312220849096775, + 0.0046541872434318066, + 0.005014817696064711, + 0.0034471359103918076, + 0.004215490538626909, + 0.0027835527434945107, + 0.002494915621355176, + 0.004003781359642744, + 0.0027278487104922533, + 0.002878724131733179, + 0.004632472526282072, + 0.007947719655930996, + 0.004853999707847834, + 0.0037676533684134483, + 0.004564839880913496, + 0.017268437892198563, + 0.011871864087879658, + 0.005097897723317146, + 0.0038858328480273485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.033698562532663345 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464016804471612 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.00464016804471612, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464016804471612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30887430906295776, + 0.09315386414527893, + 0.05170817673206329, + 0.11119130253791809, + 0.07755106687545776, + 0.11688900738954544, + 0.09568118304014206, + 0.027108965441584587, + 0.04248587042093277, + 0.07535631209611893 + ], + "credit_map": [ + 0.025997746735811234, + 0.03984196484088898, + 0.030232688412070274, + 0.030412083491683006, + 0.08120636641979218, + 0.06649567931890488, + 0.013887588866055012, + 0.020800182595849037, + 0.01133769378066063, + 0.0128082986921072, + 0.009914720430970192, + 0.010212738066911697, + 0.020990142598748207, + 0.012681849300861359, + 0.007615776266902685, + 0.007592638023197651, + 0.004872420337051153, + 0.00787017960101366, + 0.006813479587435722, + 0.006444389931857586, + 0.008114011026918888, + 0.005729747470468283, + 0.005950488615781069, + 0.005913462955504656, + 0.01402444951236248, + 0.012862784788012505, + 0.01208453718572855, + 0.015434050932526588, + 0.011764037422835827, + 0.020535724237561226, + 0.011590554378926754, + 0.012895171530544758, + 0.013222183100879192, + 0.007564437109977007, + 0.008564433082938194, + 0.009123947471380234, + 0.007982886396348476, + 0.009843147359788418, + 0.011748051270842552, + 0.009501982480287552, + 0.02208421193063259, + 0.01176315825432539, + 0.01310654915869236, + 0.013758597895503044, + 0.011790856719017029, + 0.012570016086101532, + 0.010856869630515575, + 0.02095874398946762, + 0.00923625286668539, + 0.009331632405519485, + 0.011854096315801144, + 0.010138235054910183, + 0.02450862154364586, + 0.016517918556928635, + 0.007770543918013573, + 0.006323882378637791, + 0.0037448923103511333, + 0.003000555792823434, + 0.0034497056622058153, + 0.0034301048144698143, + 0.004287922289222479, + 0.0028270233888179064, + 0.003402747679501772, + 0.0029660153668373823, + 0.005970072001218796, + 0.004019124433398247, + 0.00592816760763526, + 0.004863143898546696, + 0.009195362217724323, + 0.005948598962277174, + 0.003519796533510089, + 0.0030416061636060476, + 0.00845698919147253, + 0.007684112060815096, + 0.006633255630731583, + 0.009209636598825455, + 0.019382119178771973, + 0.012900370173156261, + 0.0058889552019536495, + 0.005200868472456932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.00464016804471612 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a static, top-down perspective, the scene unfolds like a living Chinese landscape painting. A long, unbroken column of red and black ants traverses the surface of a large, mossy boulder, rendered not with photorealistic detail but rather through soft, wet-on-wet ink washes on the canvas, with charcoal and gray tones blending seamlessly to suggest an ancient, weathered texture. Patches of moss are depicted as blooming, spreading areas of pale celadon and deep green, their colors softly seeping into the damp rock surface. Against this atmospheric backdrop, the ants themselves are sharp calligraphic figures, each a vibrant stroke of deep black or vibrant crimson ink. Their slender legs and antennae are delicate, vibrant lines, giving them a sense of precise, rhythmic movement as they steadily and deliberately carry tiny, pale ochre seeds. The entire work is bathed in the diffuse, ethereal light characteristic of scroll painting, creating a profound atmosphere of tranquil concentration and minimalist elegance. The sharp, fluid energy of the ants forms a striking contrast with the soft, contemplative tranquility of the ink world.", + "seed": 12387, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_14/index_9419_seed_12387_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07127811515238136, + "vlm_logprob": -0.07127811515238136, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.07127811515238136, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18418270349502563 + ], + "opposite_logprob_tokens": [ + -1.8091826438903809 + ], + "logprob_reward": -0.18418270349502563, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18418270349502563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17855271697044373, + 0.4004186987876892, + 0.05090196803212166, + 0.10634421557188034, + 0.05716356635093689, + 0.04422277584671974, + 0.04579373449087143, + 0.032493650913238525, + 0.029845885932445526, + 0.05426279455423355 + ], + "credit_map": [ + 0.03441841900348663, + 0.02253989689052105, + 0.01782403513789177, + 0.023379717022180557, + 0.022798215970396996, + 0.020417355000972748, + 0.021778231486678123, + 0.015396839007735252, + 0.06317883729934692, + 0.038730770349502563, + 0.03623373061418533, + 0.08389363437891006, + 0.037854380905628204, + 0.033542659133672714, + 0.07388312369585037, + 0.03310154750943184, + 0.00955369882285595, + 0.005664198659360409, + 0.006446110084652901, + 0.007057508919388056, + 0.00537074776366353, + 0.007726235780864954, + 0.005470941308885813, + 0.0036125313490629196, + 0.028881631791591644, + 0.015309267677366734, + 0.008371508680284023, + 0.009598526172339916, + 0.010527373291552067, + 0.013489850796759129, + 0.00904837716370821, + 0.01111767627298832, + 0.013762633316218853, + 0.007706552278250456, + 0.004556841216981411, + 0.007583268918097019, + 0.006155958399176598, + 0.007462434470653534, + 0.005631000269204378, + 0.004304877948015928, + 0.011494437232613564, + 0.006456736009567976, + 0.003929070197045803, + 0.004626346752047539, + 0.004328662063926458, + 0.0041739726439118385, + 0.0032858080230653286, + 0.005927741527557373, + 0.012112803757190704, + 0.005818189587444067, + 0.005281946621835232, + 0.005356503650546074, + 0.00432357843965292, + 0.004541429225355387, + 0.004344564862549305, + 0.004014718811959028, + 0.006677879486232996, + 0.0036905633751302958, + 0.002573567209765315, + 0.004618776962161064, + 0.0037610847502946854, + 0.0051094586960971355, + 0.003535158233717084, + 0.002527163829654455, + 0.004931239876896143, + 0.0031111612915992737, + 0.0024913300294429064, + 0.004035785794258118, + 0.0032665536273270845, + 0.005837699864059687, + 0.003463620552793145, + 0.0027084939647465944, + 0.010087044909596443, + 0.008135375566780567, + 0.004820167552679777, + 0.006814009044319391, + 0.006947478279471397, + 0.009607882238924503, + 0.004078249912708998, + 0.0037725879810750484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.18418270349502563 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007637703791260719 + ], + "opposite_logprob_tokens": [ + -4.8826375007629395 + ], + "logprob_reward": -0.007637703791260719, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007637703791260719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05906163528561592, + 0.2656405568122864, + 0.05118544027209282, + 0.03930472210049629, + 0.13316023349761963, + 0.06966084241867065, + 0.11676663160324097, + 0.08898081630468369, + 0.12493359297513962, + 0.05130554363131523 + ], + "credit_map": [ + 0.00808731745928526, + 0.0063455612398684025, + 0.005260035395622253, + 0.011274641379714012, + 0.006880444008857012, + 0.0074129654094576836, + 0.009316143579781055, + 0.00448452727869153, + 0.04331975057721138, + 0.026869384571909904, + 0.020427728071808815, + 0.052170027047395706, + 0.027556944638490677, + 0.01875893585383892, + 0.052736517041921616, + 0.023801255971193314, + 0.00813972670584917, + 0.005043528042733669, + 0.006406743545085192, + 0.007778480648994446, + 0.005374084692448378, + 0.010047716088593006, + 0.004304328467696905, + 0.0040908316150307655, + 0.008757688105106354, + 0.004878988489508629, + 0.0031495008151978254, + 0.004070672672241926, + 0.005222632084041834, + 0.006679000798612833, + 0.003103708615526557, + 0.0034425300545990467, + 0.034464988857507706, + 0.019570855423808098, + 0.011277368292212486, + 0.018952924758195877, + 0.012674606405198574, + 0.013776225969195366, + 0.01269218698143959, + 0.009751081466674805, + 0.015250762924551964, + 0.008615687489509583, + 0.007589524146169424, + 0.005865050479769707, + 0.007264638785272837, + 0.010268716141581535, + 0.004505249671638012, + 0.010301212780177593, + 0.025854669511318207, + 0.014820030890405178, + 0.012763457372784615, + 0.00795743241906166, + 0.013809874653816223, + 0.02124444581568241, + 0.009813751094043255, + 0.010502973571419716, + 0.014879928901791573, + 0.009078259579837322, + 0.008990499190986156, + 0.011802298948168755, + 0.009691228158771992, + 0.017041118815541267, + 0.009135645814239979, + 0.00836184248328209, + 0.028684651479125023, + 0.016003549098968506, + 0.011564315296709538, + 0.01724216900765896, + 0.013712706975638866, + 0.013357543386518955, + 0.013531670905649662, + 0.010836989618837833, + 0.007852723821997643, + 0.005928774829953909, + 0.00714068952947855, + 0.007128797471523285, + 0.004059769678860903, + 0.007204437628388405, + 0.00794738158583641, + 0.004042965359985828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 8 + ], + "selected_reward": -0.007637703791260719 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428080052137375 + ], + "opposite_logprob_tokens": [ + -2.017807960510254 + ], + "logprob_reward": -0.1428080052137375, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428080052137375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1676064133644104, + 0.14995598793029785, + 0.06976735591888428, + 0.09439413994550705, + 0.09021326154470444, + 0.13012978434562683, + 0.09411381185054779, + 0.04827241227030754, + 0.10263276845216751, + 0.0529140941798687 + ], + "credit_map": [ + 0.029506755992770195, + 0.017587214708328247, + 0.020839586853981018, + 0.02160835452377796, + 0.021737119182944298, + 0.02329178899526596, + 0.017162973061203957, + 0.01587262935936451, + 0.032033778727054596, + 0.018191630020737648, + 0.015135476365685463, + 0.01748606003820896, + 0.016190774738788605, + 0.021334413439035416, + 0.01672598347067833, + 0.01285786647349596, + 0.009809983894228935, + 0.00761397834867239, + 0.00532496115192771, + 0.0073826550506055355, + 0.010309714823961258, + 0.016477618366479874, + 0.007769353222101927, + 0.005079089663922787, + 0.010926485992968082, + 0.008171590976417065, + 0.009126991033554077, + 0.014502324163913727, + 0.012793741188943386, + 0.017511511221528053, + 0.013058017008006573, + 0.008303469978272915, + 0.019194841384887695, + 0.011267045512795448, + 0.00922362133860588, + 0.014056989923119545, + 0.009784801863133907, + 0.010197490453720093, + 0.007657171227037907, + 0.008831296116113663, + 0.03060574270784855, + 0.0173630528151989, + 0.01304465252906084, + 0.02046538144350052, + 0.014695942401885986, + 0.016736900433897972, + 0.008647781796753407, + 0.008570323698222637, + 0.019021451473236084, + 0.01043321006000042, + 0.012783452868461609, + 0.0098012275993824, + 0.008450113236904144, + 0.013472066260874271, + 0.009864812716841698, + 0.010287472046911716, + 0.013577600941061974, + 0.007095743902027607, + 0.0046490272507071495, + 0.003990534693002701, + 0.005179957952350378, + 0.0064353193156421185, + 0.0034822640009224415, + 0.003861958160996437, + 0.028497682884335518, + 0.014989586547017097, + 0.009146464988589287, + 0.011942158453166485, + 0.01155287679284811, + 0.01167047955095768, + 0.007359866518527269, + 0.007473636884242296, + 0.007800905499607325, + 0.005806318949908018, + 0.005382174625992775, + 0.008508368395268917, + 0.0070056794211268425, + 0.008304799906909466, + 0.006748615298420191, + 0.0033572285901755095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.1428080052137375 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581139251589775 + ], + "opposite_logprob_tokens": [ + -3.8955812454223633 + ], + "logprob_reward": -0.020581139251589775, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581139251589775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29581862688064575, + 0.23955675959587097, + 0.056613337248563766, + 0.06459061801433563, + 0.060578662902116776, + 0.051842667162418365, + 0.052172012627124786, + 0.07476373016834259, + 0.03961874917149544, + 0.06444485485553741 + ], + "credit_map": [ + 0.04065841808915138, + 0.029440512880682945, + 0.03852175921201706, + 0.05554664134979248, + 0.039102841168642044, + 0.035148750990629196, + 0.03643598034977913, + 0.020963722839951515, + 0.04996710270643234, + 0.02770509012043476, + 0.016929935663938522, + 0.03937641158699989, + 0.022028077393770218, + 0.020334912464022636, + 0.04182584211230278, + 0.021389393135905266, + 0.01022595725953579, + 0.005641213618218899, + 0.007032011169940233, + 0.00967829953879118, + 0.0062356796115636826, + 0.007940823212265968, + 0.0057878149673342705, + 0.004071536473929882, + 0.014330941252410412, + 0.006900508422404528, + 0.00548184709623456, + 0.010194703005254269, + 0.007263687439262867, + 0.008124537765979767, + 0.005344535689800978, + 0.00694985780864954, + 0.012791546061635017, + 0.007089748978614807, + 0.004805257078260183, + 0.008048022165894508, + 0.006764576770365238, + 0.009874792769551277, + 0.005367410834878683, + 0.005837307311594486, + 0.007134428713470697, + 0.005507908295840025, + 0.004808381199836731, + 0.0070715113542973995, + 0.005850682500749826, + 0.011014606803655624, + 0.005152252968400717, + 0.0053028990514576435, + 0.015282007865607738, + 0.0077013070695102215, + 0.004796758759766817, + 0.00457459082826972, + 0.0057628280483186245, + 0.007790778297930956, + 0.002973598428070545, + 0.003290142398327589, + 0.02317635901272297, + 0.012031608261168003, + 0.005729238968342543, + 0.005829646252095699, + 0.007621000986546278, + 0.01020266953855753, + 0.005138031207025051, + 0.005035178270190954, + 0.006440395023673773, + 0.00406342139467597, + 0.00428137369453907, + 0.004764569457620382, + 0.005751101300120354, + 0.006051399279385805, + 0.004570061806589365, + 0.0036964265163987875, + 0.011779140681028366, + 0.010070931166410446, + 0.0061441645957529545, + 0.008404496125876904, + 0.006539574358612299, + 0.008740758523344994, + 0.008652444928884506, + 0.004113348666578531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.020581139251589775 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011810240102931857 + ], + "opposite_logprob_tokens": [ + -6.751181125640869 + ], + "logprob_reward": -0.0011810240102931857, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011810240102931857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12351230531930923, + 0.5262421369552612, + 0.04909071326255798, + 0.05399780347943306, + 0.08082310110330582, + 0.05623120442032814, + 0.024243725463747978, + 0.03145579621195793, + 0.020456461235880852, + 0.033946696668863297 + ], + "credit_map": [ + 0.02684391289949417, + 0.014296897687017918, + 0.013288258574903011, + 0.01443928200751543, + 0.01725907437503338, + 0.013668236322700977, + 0.012098925188183784, + 0.011617723852396011, + 0.0601331889629364, + 0.047417208552360535, + 0.04466569796204567, + 0.11556544899940491, + 0.05192547291517258, + 0.055657681077718735, + 0.1081237867474556, + 0.04275364801287651, + 0.014432725496590137, + 0.007823770865797997, + 0.0038751380052417517, + 0.004835137631744146, + 0.0043798089027404785, + 0.006227807607501745, + 0.0034739114344120026, + 0.0040424116887152195, + 0.015747258439660072, + 0.007825837470591068, + 0.004191781859844923, + 0.004884060472249985, + 0.005334050394594669, + 0.00544583797454834, + 0.004349723923951387, + 0.006219253409653902, + 0.020957959815859795, + 0.011712310835719109, + 0.006463642232120037, + 0.010771317407488823, + 0.007880155928432941, + 0.009335261769592762, + 0.008261529728770256, + 0.005440921988338232, + 0.013949988409876823, + 0.00751937460154295, + 0.004632008261978626, + 0.005736464634537697, + 0.005503525026142597, + 0.006510619074106216, + 0.0045205228962004185, + 0.007858704775571823, + 0.006725654937326908, + 0.0038564454298466444, + 0.0021039738785475492, + 0.003476578975096345, + 0.0025720656849443913, + 0.0024033295921981335, + 0.0017393643502146006, + 0.0013663118006661534, + 0.009386518970131874, + 0.004855603910982609, + 0.0032342588528990746, + 0.003261181293055415, + 0.0031986450776457787, + 0.0035564296413213015, + 0.0020807702094316483, + 0.0018823874415829778, + 0.004116704221814871, + 0.0029102338012307882, + 0.0025031212717294693, + 0.002281824592500925, + 0.0019936184398829937, + 0.0019589609000831842, + 0.002534599043428898, + 0.0021573996637016535, + 0.008898559026420116, + 0.009135395288467407, + 0.0026964375283569098, + 0.0023227017372846603, + 0.003837786614894867, + 0.00334166525863111, + 0.002142129000276327, + 0.0015720220981165767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0011810240102931857 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a static, top-down perspective, the scene unfolds like a living Chinese landscape painting. A long, unbroken column of red and black ants traverses the surface of a large, mossy boulder, rendered not with photorealistic detail but rather through soft, wet-on-wet ink washes on the canvas, with charcoal and gray tones blending seamlessly to suggest an ancient, weathered texture. Patches of moss are depicted as blooming, spreading areas of pale celadon and deep green, their colors softly seeping into the damp rock surface. Against this atmospheric backdrop, the ants themselves are sharp calligraphic figures, each a vibrant stroke of deep black or vibrant crimson ink. Their slender legs and antennae are delicate, vibrant lines, giving them a sense of precise, rhythmic movement as they steadily and deliberately carry tiny, pale ochre seeds. The entire work is bathed in the diffuse, ethereal light characteristic of scroll painting, creating a profound atmosphere of tranquil concentration and minimalist elegance. The sharp, fluid energy of the ants forms a striking contrast with the soft, contemplative tranquility of the ink world.", + "seed": 12388, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_14/index_9419_seed_12388_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04884760770946741, + "vlm_logprob": -0.04884760770946741, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.04884760770946741, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16654814779758453 + ], + "opposite_logprob_tokens": [ + -1.9165481328964233 + ], + "logprob_reward": -0.16654814779758453, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16654814779758453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10461504012346268, + 0.5088769197463989, + 0.11305765062570572, + 0.10832415521144867, + 0.04130411893129349, + 0.03200513869524002, + 0.016806865110993385, + 0.041180629283189774, + 0.0181654691696167, + 0.015663985162973404 + ], + "credit_map": [ + 0.015894655138254166, + 0.018513519316911697, + 0.013283115811645985, + 0.015414830297231674, + 0.010420842096209526, + 0.010902957990765572, + 0.00909311044961214, + 0.011092009954154491, + 0.21902209520339966, + 0.2200952172279358, + 0.01208861917257309, + 0.00716201588511467, + 0.020194215700030327, + 0.00983535498380661, + 0.011233393102884293, + 0.009246006608009338, + 0.019273528829216957, + 0.01679401472210884, + 0.01928717829287052, + 0.008486003614962101, + 0.012133301235735416, + 0.009932155720889568, + 0.00923330057412386, + 0.017918165773153305, + 0.01986698806285858, + 0.014716909267008305, + 0.009018299169838428, + 0.008563593961298466, + 0.014836124144494534, + 0.009818265214562416, + 0.015466355718672276, + 0.016037622466683388, + 0.006683106068521738, + 0.005810190923511982, + 0.0071435547433793545, + 0.004905114881694317, + 0.0031767776235938072, + 0.0037215203046798706, + 0.003922541160136461, + 0.005941310431808233, + 0.005870405118912458, + 0.004435115028172731, + 0.0032909810543060303, + 0.0035820906050503254, + 0.004669887945055962, + 0.00267610652372241, + 0.003249065950512886, + 0.0042314850725233555, + 0.0024272880982607603, + 0.002486339770257473, + 0.00166886649094522, + 0.001962385606020689, + 0.0016158194048330188, + 0.002363050589337945, + 0.0012928233481943607, + 0.0029902919195592403, + 0.00508485222235322, + 0.004888154566287994, + 0.0032459208741784096, + 0.006402876693755388, + 0.0034001434687525034, + 0.004754940047860146, + 0.00802485179156065, + 0.005378890782594681, + 0.0031039202585816383, + 0.00241970457136631, + 0.001653685001656413, + 0.0014953183708712459, + 0.0014615710824728012, + 0.0020803611259907484, + 0.003158008214086294, + 0.002792901126667857, + 0.002113625407218933, + 0.001761634717695415, + 0.0017973965732380748, + 0.0018324152333661914, + 0.0013569553848356009, + 0.002469039522111416, + 0.0021009198389947414, + 0.0022319990675896406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.16654814779758453 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059059962630272 + ], + "opposite_logprob_tokens": [ + -3.8955905437469482 + ], + "logprob_reward": -0.02059059962630272, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059059962630272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11946175247430801, + 0.38464558124542236, + 0.1037554070353508, + 0.025699572637677193, + 0.0659165009856224, + 0.04673480615019798, + 0.0598127506673336, + 0.1323087513446808, + 0.030885128304362297, + 0.0307797659188509 + ], + "credit_map": [ + 0.029154229909181595, + 0.02929578721523285, + 0.008365892805159092, + 0.00952134095132351, + 0.011877581477165222, + 0.012164318934082985, + 0.009686404839158058, + 0.009396201930940151, + 0.1613570898771286, + 0.16413618624210358, + 0.008562257513403893, + 0.006375257857143879, + 0.017927642911672592, + 0.008130183443427086, + 0.009531935676932335, + 0.008625037036836147, + 0.02356633171439171, + 0.018459973856806755, + 0.015295718796551228, + 0.006638375110924244, + 0.009025538340210915, + 0.007137539330869913, + 0.009432440623641014, + 0.014199481345713139, + 0.0040872362442314625, + 0.003744284389540553, + 0.0028211569879204035, + 0.0020074702333658934, + 0.003468456445261836, + 0.0025015866849571466, + 0.003511138493195176, + 0.003558244090527296, + 0.010020497255027294, + 0.011373180896043777, + 0.007347616832703352, + 0.005783400032669306, + 0.012862824834883213, + 0.006063805427402258, + 0.005480601452291012, + 0.00698457658290863, + 0.00561124412342906, + 0.005823850631713867, + 0.007119150832295418, + 0.0050265612080693245, + 0.005914970301091671, + 0.006239281967282295, + 0.005248017143458128, + 0.005751732271164656, + 0.007511659525334835, + 0.007313184905797243, + 0.005378044210374355, + 0.00798369012773037, + 0.00857604295015335, + 0.008581160567700863, + 0.00609629275277257, + 0.00837267842143774, + 0.019809016957879066, + 0.01829690858721733, + 0.01227322593331337, + 0.013027267530560493, + 0.012073580175638199, + 0.020490679889917374, + 0.01346190832555294, + 0.022876160219311714, + 0.0054505495354533195, + 0.003815153380855918, + 0.0026157686952501535, + 0.0025901184417307377, + 0.004164493642747402, + 0.003407529555261135, + 0.004040900617837906, + 0.004800612572580576, + 0.00417112372815609, + 0.004418994300067425, + 0.002737773582339287, + 0.0026407663244754076, + 0.004592562094330788, + 0.003908430226147175, + 0.0046546440571546555, + 0.0036554697435349226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.02059059962630272 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370063751935959 + ], + "opposite_logprob_tokens": [ + -3.408700704574585 + ], + "logprob_reward": -0.03370063751935959, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370063751935959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17168517410755157, + 0.44917032122612, + 0.06805230677127838, + 0.06593703478574753, + 0.047960612922906876, + 0.04281512275338173, + 0.04497319459915161, + 0.047478463500738144, + 0.019263211637735367, + 0.042664557695388794 + ], + "credit_map": [ + 0.034488074481487274, + 0.03595644235610962, + 0.011399243026971817, + 0.015929516404867172, + 0.02511356957256794, + 0.015854885801672935, + 0.014175839722156525, + 0.01876761019229889, + 0.19415594637393951, + 0.19502083957195282, + 0.008626896888017654, + 0.006867038086056709, + 0.017752258107066154, + 0.00923994556069374, + 0.009850737638771534, + 0.0076566580682992935, + 0.008835372515022755, + 0.009383635595440865, + 0.011355794966220856, + 0.006457771174609661, + 0.006012099329382181, + 0.007201062515377998, + 0.007296003866940737, + 0.011510570533573627, + 0.009973718784749508, + 0.009057422168552876, + 0.005773636512458324, + 0.005244066007435322, + 0.017383942380547523, + 0.007659338414669037, + 0.005812194664031267, + 0.00503271259367466, + 0.0061759622767567635, + 0.005840830039232969, + 0.006720966659486294, + 0.006229650229215622, + 0.005056382156908512, + 0.005028911400586367, + 0.005364162381738424, + 0.007543750107288361, + 0.006328469142317772, + 0.005529850255697966, + 0.005610393825918436, + 0.006428065709769726, + 0.006561800371855497, + 0.004070710856467485, + 0.003538004821166396, + 0.0047478280030190945, + 0.007937656715512276, + 0.006272254511713982, + 0.0045764935202896595, + 0.006796903908252716, + 0.006735318806022406, + 0.003835753072053194, + 0.00356947910040617, + 0.005249337293207645, + 0.006398944649845362, + 0.005781634710729122, + 0.003241083351895213, + 0.006531449034810066, + 0.006500471383333206, + 0.004697266034781933, + 0.008046607486903667, + 0.006281006615608931, + 0.0034401502925902605, + 0.0025128317065536976, + 0.0019381993915885687, + 0.0016461840132251382, + 0.002635055920109153, + 0.0016980066429823637, + 0.0034736257512122393, + 0.0019191594328731298, + 0.006537056062370539, + 0.004930880386382341, + 0.006010133773088455, + 0.003342109266668558, + 0.005910147912800312, + 0.004445201717317104, + 0.006327436771243811, + 0.005161590874195099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.03370063751935959 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057880349457264 + ], + "opposite_logprob_tokens": [ + -3.8955788612365723 + ], + "logprob_reward": -0.02057880349457264, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057880349457264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15651459991931915, + 0.45845848321914673, + 0.05958282947540283, + 0.07941682636737823, + 0.03833191096782684, + 0.05111609026789665, + 0.07500527054071426, + 0.04006131738424301, + 0.013926060870289803, + 0.02758665569126606 + ], + "credit_map": [ + 0.02799389325082302, + 0.031784117221832275, + 0.013177844695746899, + 0.017021460458636284, + 0.016353093087673187, + 0.014662444591522217, + 0.013923601247370243, + 0.021598145365715027, + 0.19465689361095428, + 0.19661840796470642, + 0.01052763406187296, + 0.008874040096998215, + 0.018895944580435753, + 0.009885426610708237, + 0.010585498996078968, + 0.008414625190198421, + 0.010418027639389038, + 0.008614432998001575, + 0.007357717491686344, + 0.004499726463109255, + 0.00563777843490243, + 0.005851683206856251, + 0.006295967381447554, + 0.010907493531703949, + 0.009973512962460518, + 0.010933401063084602, + 0.009403962641954422, + 0.005308091174811125, + 0.011578383855521679, + 0.006806184072047472, + 0.012291595339775085, + 0.013121693395078182, + 0.006651659030467272, + 0.006595499347895384, + 0.003946660086512566, + 0.0050202952697873116, + 0.004491293337196112, + 0.003093551844358444, + 0.003895752364769578, + 0.004637200850993395, + 0.005032435059547424, + 0.005812580231577158, + 0.006886438932269812, + 0.008963166736066341, + 0.007908107712864876, + 0.005575265269726515, + 0.005716713611036539, + 0.00522138737142086, + 0.013396937400102615, + 0.011429641395807266, + 0.007334811147302389, + 0.01035300176590681, + 0.007679720409214497, + 0.007879951037466526, + 0.007696159649640322, + 0.009235047735273838, + 0.006200933363288641, + 0.005479560233652592, + 0.003989129792898893, + 0.005542328581213951, + 0.0041403984650969505, + 0.004151984583586454, + 0.004655986092984676, + 0.00590099673718214, + 0.002128127496689558, + 0.0017065871506929398, + 0.0014991344651207328, + 0.0018603882053866982, + 0.0016476649325340986, + 0.0016094250604510307, + 0.0018061030423268676, + 0.0016686314484104514, + 0.004453832749277353, + 0.0034506802912801504, + 0.003131053876131773, + 0.0026747246738523245, + 0.002430250635370612, + 0.003565790131688118, + 0.004169133026152849, + 0.003711190540343523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.02057880349457264 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028198501095175743 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.0028198501095175743, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028198501095175743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17121468484401703, + 0.29429176449775696, + 0.04319611191749573, + 0.23463697731494904, + 0.06485647708177567, + 0.04109985753893852, + 0.033855658024549484, + 0.03769979625940323, + 0.01886546052992344, + 0.06028328090906143 + ], + "credit_map": [ + 0.023717530071735382, + 0.02842259407043457, + 0.018696073442697525, + 0.02478352002799511, + 0.01731165312230587, + 0.01638728380203247, + 0.016649890691041946, + 0.02524612843990326, + 0.12472204864025116, + 0.12594842910766602, + 0.007923249155282974, + 0.004820648580789566, + 0.01206202618777752, + 0.0062875449657440186, + 0.006971721537411213, + 0.00555608281865716, + 0.009415742009878159, + 0.006403901148587465, + 0.004055995028465986, + 0.002660318510606885, + 0.0038282310124486685, + 0.0037027003709226847, + 0.006029233802109957, + 0.007099985610693693, + 0.04942534863948822, + 0.03834224119782448, + 0.021750206127762794, + 0.016968900337815285, + 0.023313477635383606, + 0.01748700812458992, + 0.03218235448002815, + 0.035167396068573, + 0.009587306529283524, + 0.009237601421773434, + 0.011414861306548119, + 0.007317123003304005, + 0.007289459463208914, + 0.005485300440341234, + 0.006093865260481834, + 0.00843095127493143, + 0.005486732814460993, + 0.005445678718388081, + 0.004667078610509634, + 0.0038016948383301497, + 0.0057694558054208755, + 0.0048726885579526424, + 0.0048347595147788525, + 0.0062217628583312035, + 0.006951179355382919, + 0.005135823972523212, + 0.0029004178941249847, + 0.0028528496623039246, + 0.003955053631216288, + 0.0034345637541264296, + 0.00434418860822916, + 0.004281579051166773, + 0.004963629879057407, + 0.004920464474707842, + 0.002759962109848857, + 0.005309052299708128, + 0.004770089406520128, + 0.005165692884474993, + 0.004048728384077549, + 0.00576217332854867, + 0.004691970534622669, + 0.003063375595957041, + 0.0021851505152881145, + 0.0012377864914014935, + 0.0017049949383363128, + 0.001659294357523322, + 0.0023566712625324726, + 0.0019662135746330023, + 0.009004354476928711, + 0.007763405796140432, + 0.007023976184427738, + 0.0053071752190589905, + 0.00703782495111227, + 0.00646110950037837, + 0.00822740513831377, + 0.009458029642701149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.0028198501095175743 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a static, top-down perspective, the scene unfolds like a living Chinese landscape painting. A long, unbroken column of red and black ants traverses the surface of a large, mossy boulder, rendered not with photorealistic detail but rather through soft, wet-on-wet ink washes on the canvas, with charcoal and gray tones blending seamlessly to suggest an ancient, weathered texture. Patches of moss are depicted as blooming, spreading areas of pale celadon and deep green, their colors softly seeping into the damp rock surface. Against this atmospheric backdrop, the ants themselves are sharp calligraphic figures, each a vibrant stroke of deep black or vibrant crimson ink. Their slender legs and antennae are delicate, vibrant lines, giving them a sense of precise, rhythmic movement as they steadily and deliberately carry tiny, pale ochre seeds. The entire work is bathed in the diffuse, ethereal light characteristic of scroll painting, creating a profound atmosphere of tranquil concentration and minimalist elegance. The sharp, fluid energy of the ants forms a striking contrast with the soft, contemplative tranquility of the ink world.", + "seed": 12389, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_14/index_9419_seed_12389_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10438224691897631, + "vlm_logprob": -0.10438224691897631, + "vlm_yes_no_margin": 3.325000011920929, + "vlm_token_credit": -0.10438224691897631, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3998696208000183 + ], + "opposite_logprob_tokens": [ + -1.149869680404663 + ], + "logprob_reward": -0.3998696208000183, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3998696208000183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14811761677265167, + 0.03763248026371002, + 0.03499497473239899, + 0.03574345260858536, + 0.025421874597668648, + 0.08214924484491348, + 0.07305790483951569, + 0.11109589040279388, + 0.3033350706100464, + 0.14845146238803864 + ], + "credit_map": [ + 0.01775040663778782, + 0.017490295693278313, + 0.015728069469332695, + 0.012849561870098114, + 0.014894695952534676, + 0.03191690519452095, + 0.014761930331587791, + 0.02272576279938221, + 0.005681105889379978, + 0.004181656986474991, + 0.005463373381644487, + 0.004621933680027723, + 0.004327680915594101, + 0.004425736144185066, + 0.003913172520697117, + 0.005017823074012995, + 0.005790079012513161, + 0.0033883010037243366, + 0.0054570152424275875, + 0.003565264865756035, + 0.002962724305689335, + 0.0037502071354538202, + 0.004949753172695637, + 0.0051316251046955585, + 0.005785245914012194, + 0.0035656478721648455, + 0.006254345178604126, + 0.004209455102682114, + 0.003250336507335305, + 0.0037401735316962004, + 0.003936640452593565, + 0.00500160688534379, + 0.004354557953774929, + 0.002913388190791011, + 0.003350757295265794, + 0.002877473132684827, + 0.002422033343464136, + 0.0035436893813312054, + 0.002577216597273946, + 0.0033827582374215126, + 0.013787630014121532, + 0.008231914602220058, + 0.009822167456150055, + 0.012440442107617855, + 0.006903818808495998, + 0.008508306927978992, + 0.012135719880461693, + 0.010319242253899574, + 0.012893225997686386, + 0.007693578954786062, + 0.008437358774244785, + 0.00898750126361847, + 0.005682806018739939, + 0.005148476455360651, + 0.009038825519382954, + 0.015176131390035152, + 0.01852514035999775, + 0.01196378841996193, + 0.011237013153731823, + 0.011490286327898502, + 0.0076734102331101894, + 0.015728598460555077, + 0.01753181405365467, + 0.016945840790867805, + 0.04745456948876381, + 0.039002176374197006, + 0.03353280574083328, + 0.036673739552497864, + 0.024759655818343163, + 0.03909991681575775, + 0.04224756732583046, + 0.040564633905887604, + 0.021243680268526077, + 0.022370224818587303, + 0.02517004683613777, + 0.00921835470944643, + 0.013221743516623974, + 0.01087911706417799, + 0.007897194474935532, + 0.038451097905635834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.3998696208000183 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008654467761516571 + ], + "opposite_logprob_tokens": [ + -4.758654594421387 + ], + "logprob_reward": -0.008654467761516571, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008654467761516571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.259212851524353, + 0.0844220444560051, + 0.0465519055724144, + 0.16763368248939514, + 0.0669112503528595, + 0.07972182333469391, + 0.11073347926139832, + 0.11139693111181259, + 0.03845369070768356, + 0.03496239334344864 + ], + "credit_map": [ + 0.04248587042093277, + 0.03086410090327263, + 0.03654123842716217, + 0.02497806027531624, + 0.023765308782458305, + 0.03422503545880318, + 0.027722487226128578, + 0.038630735129117966, + 0.01682283543050289, + 0.009934622794389725, + 0.011700738221406937, + 0.007809041067957878, + 0.008630530908703804, + 0.010732738301157951, + 0.007174126338213682, + 0.011617408134043217, + 0.007257891818881035, + 0.005237075034528971, + 0.006302123423665762, + 0.00545179657638073, + 0.005195755045861006, + 0.004175903741270304, + 0.006514668930321932, + 0.006416686344891787, + 0.027426041662693024, + 0.019993113353848457, + 0.030094916000962257, + 0.020210087299346924, + 0.014063353650271893, + 0.018131472170352936, + 0.016958802938461304, + 0.020755885168910027, + 0.008688542060554028, + 0.006213972344994545, + 0.012868721969425678, + 0.008151128888130188, + 0.007041949313133955, + 0.006214991211891174, + 0.009357018396258354, + 0.008374921977519989, + 0.012291456572711468, + 0.008485266007483006, + 0.011249958537518978, + 0.01225473452359438, + 0.006778234615921974, + 0.006708353292196989, + 0.013374256901443005, + 0.008579554967582226, + 0.01866520196199417, + 0.013780521228909492, + 0.011655440554022789, + 0.013809855096042156, + 0.01113450899720192, + 0.017257606610655785, + 0.011208313517272472, + 0.013222028501331806, + 0.017775926738977432, + 0.01361944992095232, + 0.011167160235345364, + 0.012289006263017654, + 0.009475046768784523, + 0.015609250403940678, + 0.011158840730786324, + 0.020302241668105125, + 0.006128646899014711, + 0.004493031185120344, + 0.004404065664857626, + 0.005635659210383892, + 0.002577307401224971, + 0.0050368779338896275, + 0.00497702369466424, + 0.005201077088713646, + 0.005368584766983986, + 0.0077088032849133015, + 0.0033005140721797943, + 0.0033361848909407854, + 0.0026029585860669613, + 0.004251579754054546, + 0.0037662992253899574, + 0.0046274662017822266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.008654467761516571 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700320452451706 + ], + "opposite_logprob_tokens": [ + -2.6950321197509766 + ], + "logprob_reward": -0.0700320452451706, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700320452451706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.228204607963562, + 0.1208273321390152, + 0.1353888362646103, + 0.11302445828914642, + 0.05387188866734505, + 0.06831828504800797, + 0.07275751978158951, + 0.03835880756378174, + 0.10815848410129547, + 0.06108981743454933 + ], + "credit_map": [ + 0.03423990681767464, + 0.02507951855659485, + 0.023351192474365234, + 0.01853250525891781, + 0.018010510131716728, + 0.057705946266651154, + 0.014800293371081352, + 0.0364847294986248, + 0.01804882287979126, + 0.011773398146033287, + 0.015471719205379486, + 0.015582061372697353, + 0.023229273036122322, + 0.013521116226911545, + 0.013327706605196, + 0.009873242117464542, + 0.0201435424387455, + 0.014064095914363861, + 0.00940030813217163, + 0.01853242889046669, + 0.02808396704494953, + 0.015907416120171547, + 0.015746919438242912, + 0.013510164804756641, + 0.01681942120194435, + 0.014186594635248184, + 0.010734138078987598, + 0.019028307870030403, + 0.01660812273621559, + 0.010844584554433823, + 0.013078097254037857, + 0.011725196614861488, + 0.009007427841424942, + 0.006047232076525688, + 0.006238413508981466, + 0.006747564300894737, + 0.008530332706868649, + 0.004891930613666773, + 0.00549683952704072, + 0.0069121443666517735, + 0.01210649311542511, + 0.006247005425393581, + 0.00781281292438507, + 0.007956892251968384, + 0.008663324639201164, + 0.005471234209835529, + 0.011749387718737125, + 0.008311134763062, + 0.012778238393366337, + 0.007148630451411009, + 0.0059614782221615314, + 0.011101697571575642, + 0.006757015362381935, + 0.007510542403906584, + 0.010424240492284298, + 0.01107567735016346, + 0.005167437717318535, + 0.003868839703500271, + 0.004089553840458393, + 0.009347070939838886, + 0.003162698820233345, + 0.0035594257060438395, + 0.0046938457526266575, + 0.00446993438526988, + 0.01883338764309883, + 0.02030898444354534, + 0.00784812681376934, + 0.007766854949295521, + 0.012811781838536263, + 0.008641314692795277, + 0.014027649536728859, + 0.017920387908816338, + 0.009581699036061764, + 0.006551702506840229, + 0.008829501457512379, + 0.005503229331225157, + 0.008101330138742924, + 0.006745265331119299, + 0.00699533149600029, + 0.008781759068369865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0700320452451706 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809810429811478 + ], + "opposite_logprob_tokens": [ + -3.288098096847534 + ], + "logprob_reward": -0.03809810429811478, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809810429811478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3662153482437134, + 0.050793882459402084, + 0.06507227569818497, + 0.03887750953435898, + 0.0891052782535553, + 0.23393762111663818, + 0.034878674894571304, + 0.04222503677010536, + 0.038391660898923874, + 0.04050269350409508 + ], + "credit_map": [ + 0.04613718017935753, + 0.04069823771715164, + 0.05022922903299332, + 0.037073004990816116, + 0.03617693483829498, + 0.07114063203334808, + 0.026968376711010933, + 0.05779176577925682, + 0.006980596575886011, + 0.0056004757061600685, + 0.007066556718200445, + 0.004375472664833069, + 0.008094334974884987, + 0.005549489054828882, + 0.003871840424835682, + 0.009255114942789078, + 0.011183507740497589, + 0.007047037594020367, + 0.008662385866045952, + 0.00716372998431325, + 0.007240616716444492, + 0.008003082126379013, + 0.006606169976294041, + 0.009165752679109573, + 0.00489847082644701, + 0.003624459495767951, + 0.005555069074034691, + 0.006106056738644838, + 0.006232149433344603, + 0.003999682608991861, + 0.004478371236473322, + 0.0039832498878240585, + 0.012906399555504322, + 0.011152239516377449, + 0.007624723017215729, + 0.01363337878137827, + 0.015923719853162766, + 0.005900673568248749, + 0.010782407596707344, + 0.011181733570992947, + 0.034307897090911865, + 0.022353917360305786, + 0.03536894544959068, + 0.028600746765732765, + 0.03778217360377312, + 0.02186409756541252, + 0.025199271738529205, + 0.028460558503866196, + 0.0061425757594406605, + 0.0038136104121804237, + 0.0033688650000840425, + 0.0034829506184905767, + 0.0026257121935486794, + 0.0033461949788033962, + 0.0037282295525074005, + 0.008370534516870975, + 0.006249354686588049, + 0.0036207966040819883, + 0.0057784863747656345, + 0.005294270813465118, + 0.004205222241580486, + 0.003240118036046624, + 0.0046579595655202866, + 0.00917882565408945, + 0.007801906205713749, + 0.003744487650692463, + 0.004614821635186672, + 0.0036219411995261908, + 0.005504239350557327, + 0.003934149630367756, + 0.005779883824288845, + 0.0033902288414537907, + 0.007036462426185608, + 0.007250955794006586, + 0.005545131396502256, + 0.0038820665795356035, + 0.00292813777923584, + 0.0037194497417658567, + 0.003245771862566471, + 0.0068947202526032925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.03809810429811478 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a static, top-down perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large mossy boulder visible as the surface where the ants are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the ants depicted as a long, unbroken column of red and black figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the ants shown carrying tiny, pale ochre seeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style rendered with soft, wet-on-wet ink washes and charcoal tones rather than photorealistic detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256996490061283 + ], + "opposite_logprob_tokens": [ + -5.2552571296691895 + ], + "logprob_reward": -0.005256996490061283, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256996490061283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35524100065231323, + 0.08811262995004654, + 0.041058264672756195, + 0.0675957202911377, + 0.04869300127029419, + 0.0718475729227066, + 0.0897621363401413, + 0.050864752382040024, + 0.08047216385602951, + 0.10635273903608322 + ], + "credit_map": [ + 0.056328170001506805, + 0.043347399681806564, + 0.046582240611314774, + 0.028507567942142487, + 0.025789201259613037, + 0.06461843103170395, + 0.032708290964365005, + 0.05735969915986061, + 0.012947308830916882, + 0.011088794097304344, + 0.013643525540828705, + 0.009170536883175373, + 0.008077580481767654, + 0.009248178452253342, + 0.011278709396719933, + 0.012657997198402882, + 0.006942736450582743, + 0.004136492032557726, + 0.005587814375758171, + 0.006177288014441729, + 0.005752855446189642, + 0.003842341946437955, + 0.003837081603705883, + 0.004781653173267841, + 0.01002873107790947, + 0.006905687507241964, + 0.010347605682909489, + 0.009633795358240604, + 0.005601975601166487, + 0.006907044444233179, + 0.006533422973006964, + 0.011637454852461815, + 0.005897600669413805, + 0.0050555625930428505, + 0.00670318678021431, + 0.006875315681099892, + 0.007088139187544584, + 0.003934979904443026, + 0.006321718450635672, + 0.006816500797867775, + 0.011187485419213772, + 0.007414879743009806, + 0.008549260906875134, + 0.01306245755404234, + 0.006974048912525177, + 0.006678932812064886, + 0.009506868198513985, + 0.008473643101751804, + 0.014063501730561256, + 0.01043398305773735, + 0.011599831283092499, + 0.00998523086309433, + 0.006595979910343885, + 0.012636863626539707, + 0.009590420871973038, + 0.014856326393783092, + 0.00889042392373085, + 0.006359105929732323, + 0.004683693405240774, + 0.005045624915510416, + 0.0037936768494546413, + 0.005913080647587776, + 0.006716311909258366, + 0.009462833404541016, + 0.013425692915916443, + 0.010859734378755093, + 0.00902737956494093, + 0.010489927604794502, + 0.010437965393066406, + 0.006543738301843405, + 0.01187813002616167, + 0.007809597533196211, + 0.013929920271039009, + 0.01582074910402298, + 0.016796140000224113, + 0.007765206973999739, + 0.012809879146516323, + 0.009649714455008507, + 0.007378213107585907, + 0.02220291644334793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.005256996490061283 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_15/index_tensor([9111, 9111])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_15/index_tensor([9111, 9111])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..51beecf497763b321cb3f72528e4f7d6b965b5ef --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_15/index_tensor([9111, 9111])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static close-up captures a slender, unfinished wooden stick, its pale brown fibrous texture clearly visible, standing perfectly vertically on the undulating surface of a weathered gray stone. The scene is bathed in a soft, diffuse light, suggestive of an overcast sky or a shady forest floor, which creates gentle, subtle shadows around the base of the stick. The background is a gentle, out-of-focus blur of dark greens and earthy browns, creating a quiet, contemplative atmosphere. After a moment of perfect stillness, the stick begins to sway almost imperceptibly, then slowly and deliberately tilts, its shape gracefully arcing through the air. Upon impact with the stone surface, the fragile wood instantly breaks with sharp cracks, disintegrating into a small cloud of fibrous fragments and dust that scatters from the point of impact, leaving the shattered remains in a final state of stillness.", + "seed": 21777, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_15/index_9111_seed_21777_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01235173000022769, + "vlm_logprob": -0.01235173000022769, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.01235173000022769, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009355396032333374 + ], + "opposite_logprob_tokens": [ + -5.259355545043945 + ], + "logprob_reward": -0.009355396032333374, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.009355396032333374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20527242124080658, + 0.04725545644760132, + 0.04413929581642151, + 0.04309634864330292, + 0.046353936195373535, + 0.1267446130514145, + 0.10249066352844238, + 0.14952020347118378, + 0.13608397543430328, + 0.0990431159734726 + ], + "credit_map": [ + 0.029602818191051483, + 0.035026513040065765, + 0.022024331614375114, + 0.019867440685629845, + 0.017533499747514725, + 0.02752131223678589, + 0.02265426330268383, + 0.031042227521538734, + 0.005295948591083288, + 0.006122838705778122, + 0.0056589944288134575, + 0.004793894477188587, + 0.006291936617344618, + 0.0049189492128789425, + 0.006788105703890324, + 0.00738478684797883, + 0.004339870531111956, + 0.005163511261343956, + 0.006248584017157555, + 0.003857911564409733, + 0.005290057975798845, + 0.004995863419026136, + 0.006603735964745283, + 0.007639757823199034, + 0.005238888319581747, + 0.004919226747006178, + 0.007577169220894575, + 0.004167341627180576, + 0.003790018381550908, + 0.004656133241951466, + 0.004433941096067429, + 0.008313625119626522, + 0.004455030895769596, + 0.009468724019825459, + 0.0061265709809958935, + 0.004361475817859173, + 0.00506758876144886, + 0.003186611458659172, + 0.0038563464768230915, + 0.009831584058701992, + 0.01274282205849886, + 0.01504527498036623, + 0.012665720656514168, + 0.012282982468605042, + 0.03154555335640907, + 0.011243627406656742, + 0.0179250780493021, + 0.013293535448610783, + 0.007425625342875719, + 0.013994209468364716, + 0.014950145967304707, + 0.013150004670023918, + 0.009449372068047523, + 0.01037706807255745, + 0.017959432676434517, + 0.015184802003204823, + 0.0133796576410532, + 0.015861544758081436, + 0.021354949101805687, + 0.026687176898121834, + 0.012133357115089893, + 0.015505586750805378, + 0.028411371633410454, + 0.016186537221074104, + 0.012501181103289127, + 0.01646282523870468, + 0.016291417181491852, + 0.0239275973290205, + 0.011469805613160133, + 0.018699169158935547, + 0.019047332927584648, + 0.017684638500213623, + 0.008455561473965645, + 0.011666218750178814, + 0.015705715864896774, + 0.018067093566060066, + 0.008113649673759937, + 0.009823166765272617, + 0.016370350494980812, + 0.01084135565906763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.009355396032333374 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001720973290503025 + ], + "opposite_logprob_tokens": [ + -6.376720905303955 + ], + "logprob_reward": -0.001720973290503025, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001720973290503025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13562826812267303, + 0.08291293680667877, + 0.054735880345106125, + 0.04178779572248459, + 0.03909890353679657, + 0.21847544610500336, + 0.22392693161964417, + 0.10238639265298843, + 0.04645499214529991, + 0.05459245666861534 + ], + "credit_map": [ + 0.013748729601502419, + 0.014583200216293335, + 0.019521314650774002, + 0.012672601267695427, + 0.01797877810895443, + 0.016011754050850868, + 0.017476845532655716, + 0.02363504283130169, + 0.009458094835281372, + 0.02178649604320526, + 0.006923656910657883, + 0.004762359894812107, + 0.0075125498697161674, + 0.007846793159842491, + 0.007283900864422321, + 0.01733909174799919, + 0.005166279152035713, + 0.008262529037892818, + 0.006098443642258644, + 0.0049580056220293045, + 0.007151759695261717, + 0.0056874193251132965, + 0.006781000178307295, + 0.010630441829562187, + 0.0038770369719713926, + 0.007929330691695213, + 0.004095139913260937, + 0.003364142496138811, + 0.0054391250014305115, + 0.004891477059572935, + 0.0051593235693871975, + 0.007032220251858234, + 0.003892499255016446, + 0.005982179194688797, + 0.005583730526268482, + 0.0036372074391692877, + 0.006073854863643646, + 0.004441804252564907, + 0.004274801351130009, + 0.005212825257331133, + 0.018557319417595863, + 0.026139525696635246, + 0.021101420745253563, + 0.02310493029654026, + 0.061871349811553955, + 0.016646990552544594, + 0.03007856011390686, + 0.020975353196263313, + 0.012926687486469746, + 0.025354886427521706, + 0.02477523870766163, + 0.047586966305971146, + 0.017658725380897522, + 0.0204352717846632, + 0.044917069375514984, + 0.030272074043750763, + 0.009184516035020351, + 0.014095887541770935, + 0.012869436293840408, + 0.017929065972566605, + 0.007375149056315422, + 0.007745933253318071, + 0.021874036639928818, + 0.011312367394566536, + 0.00419554254040122, + 0.006222890689969063, + 0.007347036153078079, + 0.006566387135535479, + 0.005986192263662815, + 0.006439758464694023, + 0.0054182205349206924, + 0.004278964828699827, + 0.004485811572521925, + 0.005962415132671595, + 0.0065037403255701065, + 0.00959277618676424, + 0.004959628451615572, + 0.005627236794680357, + 0.010864177718758583, + 0.006596672348678112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.001720973290503025 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010476582683622837 + ], + "opposite_logprob_tokens": [ + -6.876047611236572 + ], + "logprob_reward": -0.0010476582683622837, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010476582683622837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10290495306253433, + 0.04624245688319206, + 0.028043298050761223, + 0.02384999580681324, + 0.03585921600461006, + 0.1147453710436821, + 0.2736678719520569, + 0.1492399424314499, + 0.16351500153541565, + 0.061931803822517395 + ], + "credit_map": [ + 0.011851577088236809, + 0.014888159930706024, + 0.011447809636592865, + 0.010668017901480198, + 0.010048828087747097, + 0.010902371257543564, + 0.014741158112883568, + 0.0183570459485054, + 0.0041230591014027596, + 0.005654366221278906, + 0.006209207233041525, + 0.005534071009606123, + 0.0055092088878154755, + 0.004868624731898308, + 0.0054592411033809185, + 0.00888468511402607, + 0.0024994222912937403, + 0.002803107723593712, + 0.0047757732681930065, + 0.003399485955014825, + 0.0033009005710482597, + 0.002488444559276104, + 0.003277579555287957, + 0.005498585291206837, + 0.002310621552169323, + 0.0033924987073987722, + 0.002978687407448888, + 0.002260809298604727, + 0.002747122198343277, + 0.0028073498979210854, + 0.002665117383003235, + 0.004687792155891657, + 0.004311738535761833, + 0.009102868847548962, + 0.0027745007537305355, + 0.0025773392990231514, + 0.003979438915848732, + 0.0038315984420478344, + 0.0032709462102502584, + 0.006010787095874548, + 0.009957642294466496, + 0.012937862426042557, + 0.010852886363863945, + 0.013607276603579521, + 0.03013070859014988, + 0.009396987035870552, + 0.01731260120868683, + 0.010549422353506088, + 0.01632365956902504, + 0.030295133590698242, + 0.02951330691576004, + 0.057537224143743515, + 0.021319467574357986, + 0.026559703052043915, + 0.054427534341812134, + 0.037691880017519, + 0.012828870676457882, + 0.02046431414783001, + 0.01503180991858244, + 0.024335408583283424, + 0.01409363653510809, + 0.014054752886295319, + 0.029296569526195526, + 0.019134586676955223, + 0.013063762336969376, + 0.015943307429552078, + 0.02582029066979885, + 0.03536846116185188, + 0.012544244527816772, + 0.02089594677090645, + 0.024009261280298233, + 0.0158697459846735, + 0.005844724364578724, + 0.006938999984413385, + 0.009295573458075523, + 0.008608262054622173, + 0.0055472711101174355, + 0.006469035986810923, + 0.01216830499470234, + 0.007059638854116201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.0010476582683622837 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026344917714595795 + ], + "opposite_logprob_tokens": [ + -3.6513450145721436 + ], + "logprob_reward": -0.026344917714595795, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026344917714595795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08732949197292328, + 0.0819164589047432, + 0.03322547301650047, + 0.04199070855975151, + 0.03520237281918526, + 0.24153174459934235, + 0.3040848672389984, + 0.07219269126653671, + 0.09086451679468155, + 0.01166167389601469 + ], + "credit_map": [ + 0.008603798225522041, + 0.009085498750209808, + 0.01125923078507185, + 0.007552385330200195, + 0.01310751587152481, + 0.011099080555140972, + 0.012138928286731243, + 0.01448305044323206, + 0.008031880483031273, + 0.019389266148209572, + 0.009012780152261257, + 0.005939800292253494, + 0.008417549543082714, + 0.007980107329785824, + 0.008556580170989037, + 0.014588495716452599, + 0.002835786435753107, + 0.005552203394472599, + 0.003189962822943926, + 0.0031913984566926956, + 0.005076596513390541, + 0.0036052842624485493, + 0.004357574973255396, + 0.005416666157543659, + 0.004118543583899736, + 0.008485747501254082, + 0.004125118255615234, + 0.003119085915386677, + 0.00618612440302968, + 0.0041665080934762955, + 0.00589618319645524, + 0.005893395747989416, + 0.003067510202527046, + 0.004921719431877136, + 0.005536200478672981, + 0.0033391506876796484, + 0.005096131004393101, + 0.004012698773294687, + 0.004051521886140108, + 0.005177439656108618, + 0.021098803728818893, + 0.03067949414253235, + 0.028977777808904648, + 0.028618108481168747, + 0.05536355823278427, + 0.019096313044428825, + 0.03212399780750275, + 0.025573695078492165, + 0.018265336751937866, + 0.034172236919403076, + 0.03533833846449852, + 0.06204095110297203, + 0.02476515620946884, + 0.028523167595267296, + 0.05908960849046707, + 0.04189007356762886, + 0.0063785528764128685, + 0.010500947944819927, + 0.008745178580284119, + 0.012750478461384773, + 0.006080129649490118, + 0.005729472264647484, + 0.014202669262886047, + 0.007805268280208111, + 0.006404041312634945, + 0.011516595259308815, + 0.010473749600350857, + 0.018698615953326225, + 0.007486780174076557, + 0.010290290229022503, + 0.014661891385912895, + 0.01133254636079073, + 0.0008896949584595859, + 0.0019273384241387248, + 0.0012934505939483643, + 0.0015374758513644338, + 0.0014274389250203967, + 0.0012048016069456935, + 0.002010871423408389, + 0.0013706013560295105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.026344917714595795 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328970469534397 + ], + "opposite_logprob_tokens": [ + -3.773289680480957 + ], + "logprob_reward": -0.02328970469534397, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328970469534397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11213037371635437, + 0.07048960775136948, + 0.030310843139886856, + 0.041052769869565964, + 0.03277970105409622, + 0.1908779740333557, + 0.276218980550766, + 0.14233802258968353, + 0.09006377309560776, + 0.013737943954765797 + ], + "credit_map": [ + 0.014524023048579693, + 0.01874437741935253, + 0.014766250737011433, + 0.010782822035253048, + 0.01026109792292118, + 0.013988492079079151, + 0.011915373615920544, + 0.017147941514849663, + 0.007349725347012281, + 0.020127059891819954, + 0.005939986556768417, + 0.005791036877781153, + 0.005922896787524223, + 0.009156561456620693, + 0.00638878159224987, + 0.00981355831027031, + 0.0027059316635131836, + 0.005446579307317734, + 0.003472923766821623, + 0.0033256649039685726, + 0.00363993551582098, + 0.0032916213385760784, + 0.004117687698453665, + 0.004310497082769871, + 0.004580184351652861, + 0.011598002165555954, + 0.0034397279378026724, + 0.00294099235907197, + 0.003394700586795807, + 0.005650741048157215, + 0.0037337264511734247, + 0.005714694503694773, + 0.0027177727315574884, + 0.004126021638512611, + 0.005019247066229582, + 0.003543084254488349, + 0.003735319711267948, + 0.005295911803841591, + 0.003140550572425127, + 0.005201792344450951, + 0.016732731834053993, + 0.021816259250044823, + 0.021059708669781685, + 0.022400593385100365, + 0.0488777719438076, + 0.015364413149654865, + 0.025685692206025124, + 0.01894080452620983, + 0.015863558277487755, + 0.029681093990802765, + 0.033445511013269424, + 0.05536166578531265, + 0.023667292669415474, + 0.026448804885149002, + 0.053754162043333054, + 0.03799687325954437, + 0.01037571020424366, + 0.0149254584684968, + 0.01879061944782734, + 0.022501880303025246, + 0.014511617831885815, + 0.014972720295190811, + 0.02671031281352043, + 0.01954970508813858, + 0.006588929332792759, + 0.011116022244095802, + 0.010816310532391071, + 0.019182605668902397, + 0.008265056647360325, + 0.010190670378506184, + 0.01363072358071804, + 0.01027345284819603, + 0.0011168507626280189, + 0.001669467892497778, + 0.0018554425332695246, + 0.0024226189125329256, + 0.0015900565776973963, + 0.001361930277198553, + 0.0021748121362179518, + 0.0015467648627236485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.02328970469534397 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static close-up captures a slender, unfinished wooden stick, its pale brown fibrous texture clearly visible, standing perfectly vertically on the undulating surface of a weathered gray stone. The scene is bathed in a soft, diffuse light, suggestive of an overcast sky or a shady forest floor, which creates gentle, subtle shadows around the base of the stick. The background is a gentle, out-of-focus blur of dark greens and earthy browns, creating a quiet, contemplative atmosphere. After a moment of perfect stillness, the stick begins to sway almost imperceptibly, then slowly and deliberately tilts, its shape gracefully arcing through the air. Upon impact with the stone surface, the fragile wood instantly breaks with sharp cracks, disintegrating into a small cloud of fibrous fragments and dust that scatters from the point of impact, leaving the shattered remains in a final state of stillness.", + "seed": 21778, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_15/index_9111_seed_21778_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.017877213354222477, + "vlm_logprob": -0.017877213354222477, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.017877213354222477, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010047690942883492 + ], + "opposite_logprob_tokens": [ + -5.510047912597656 + ], + "logprob_reward": -0.010047690942883492, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.010047690942883492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12845869362354279, + 0.04856358468532562, + 0.03441695123910904, + 0.030036652460694313, + 0.15835002064704895, + 0.11432632803916931, + 0.18213096261024475, + 0.21387900412082672, + 0.044040560722351074, + 0.045797199010849 + ], + "credit_map": [ + 0.0164742823690176, + 0.016069898381829262, + 0.010618053376674652, + 0.01669994369149208, + 0.0259418748319149, + 0.02435794100165367, + 0.008368145674467087, + 0.009928569197654724, + 0.00606993492692709, + 0.005321464966982603, + 0.004784958902746439, + 0.008620058186352253, + 0.008254984393715858, + 0.007511341944336891, + 0.003911817446351051, + 0.004089026246219873, + 0.004749246407300234, + 0.0034855485428124666, + 0.003506258362904191, + 0.0030534605029970407, + 0.008496471680700779, + 0.004973957780748606, + 0.0028070295229554176, + 0.003344980301335454, + 0.0033201093319803476, + 0.0037555433809757233, + 0.003056952729821205, + 0.0027213527355343103, + 0.00644490635022521, + 0.004717680625617504, + 0.0024138917215168476, + 0.0036062190774828196, + 0.013256837613880634, + 0.017038436606526375, + 0.032145578414201736, + 0.01632911153137684, + 0.01963689737021923, + 0.018862096592783928, + 0.01475734356790781, + 0.026323728263378143, + 0.01110449805855751, + 0.012580971233546734, + 0.018844710662961006, + 0.013874104246497154, + 0.013685592450201511, + 0.018300695344805717, + 0.009908686392009258, + 0.016027076169848442, + 0.01544151920825243, + 0.027172885835170746, + 0.031744107604026794, + 0.017932269722223282, + 0.022038117051124573, + 0.016978196799755096, + 0.03060569427907467, + 0.020218171179294586, + 0.02836296707391739, + 0.03021502122282982, + 0.01821565069258213, + 0.014077898114919662, + 0.0391179583966732, + 0.02495642937719822, + 0.037063319236040115, + 0.02186977118253708, + 0.005517791025340557, + 0.006653865333646536, + 0.006789649836719036, + 0.005392676219344139, + 0.007993306033313274, + 0.004752728156745434, + 0.0036951021756976843, + 0.00324544170871377, + 0.004950085189193487, + 0.005339393857866526, + 0.00989495124667883, + 0.004883900750428438, + 0.007654814049601555, + 0.005330271553248167, + 0.00429177051410079, + 0.003452013013884425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.010047690942883492 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028254371136426926 + ], + "opposite_logprob_tokens": [ + -5.8778252601623535 + ], + "logprob_reward": -0.0028254371136426926, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028254371136426926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1815575510263443, + 0.1524857133626938, + 0.07744254171848297, + 0.043814267963171005, + 0.1223999559879303, + 0.17624922096729279, + 0.11024932563304901, + 0.05990855395793915, + 0.037867769598960876, + 0.038025110960006714 + ], + "credit_map": [ + 0.020485468208789825, + 0.022948436439037323, + 0.019004540517926216, + 0.015878723934292793, + 0.039303142577409744, + 0.03292134031653404, + 0.014138218946754932, + 0.01687767170369625, + 0.013129825703799725, + 0.020513463765382767, + 0.014722954481840134, + 0.024299344047904015, + 0.03045528195798397, + 0.02304564230144024, + 0.015250695869326591, + 0.011068505235016346, + 0.007030974142253399, + 0.008602880872786045, + 0.0067275529727339745, + 0.008163313381373882, + 0.018515095114707947, + 0.014184840023517609, + 0.007923507131636143, + 0.006294371094554663, + 0.004368763882666826, + 0.0048589264042675495, + 0.004087589215487242, + 0.004700584802776575, + 0.010616529732942581, + 0.006629283539950848, + 0.004300888627767563, + 0.004251702222973108, + 0.010798835195600986, + 0.013245568610727787, + 0.02274443954229355, + 0.010819492861628532, + 0.020345594733953476, + 0.015024659223854542, + 0.010691315867006779, + 0.01873004622757435, + 0.017534954473376274, + 0.028479857370257378, + 0.012678246945142746, + 0.029368439689278603, + 0.027262302115559578, + 0.024397050961852074, + 0.023694626986980438, + 0.012833750806748867, + 0.008411172777414322, + 0.012880174443125725, + 0.01913521997630596, + 0.013167127966880798, + 0.013127626851201057, + 0.010386447422206402, + 0.020256325602531433, + 0.012885232456028461, + 0.008452965877950191, + 0.010665164329111576, + 0.006274486891925335, + 0.00498089799657464, + 0.009498050436377525, + 0.0057075549848377705, + 0.008091142401099205, + 0.006238293368369341, + 0.0053847176022827625, + 0.007918797433376312, + 0.0039732628501951694, + 0.002457668539136648, + 0.007527349051088095, + 0.004333773627877235, + 0.0036147336941212416, + 0.0026574679650366306, + 0.004864612128585577, + 0.00564305204898119, + 0.005312829744070768, + 0.0030943374149501324, + 0.0069578466936945915, + 0.0038999621756374836, + 0.005786463152617216, + 0.0024660092312842607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0028254371136426926 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001517931348644197 + ], + "opposite_logprob_tokens": [ + -6.5015177726745605 + ], + "logprob_reward": -0.001517931348644197, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001517931348644197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17702896893024445, + 0.14252880215644836, + 0.06881633400917053, + 0.0226946622133255, + 0.10586412250995636, + 0.13449494540691376, + 0.10560262948274612, + 0.14769482612609863, + 0.05409042537212372, + 0.04118425399065018 + ], + "credit_map": [ + 0.01572435349225998, + 0.018813058733940125, + 0.01730203814804554, + 0.01746000535786152, + 0.04669102281332016, + 0.03454975783824921, + 0.01383878756314516, + 0.01264994964003563, + 0.01163251232355833, + 0.02002996765077114, + 0.013869247399270535, + 0.02379574440419674, + 0.02780856005847454, + 0.02161143161356449, + 0.01446452271193266, + 0.009316820651292801, + 0.006315345410257578, + 0.009345931932330132, + 0.007237615529447794, + 0.011148116551339626, + 0.01227731816470623, + 0.01027139276266098, + 0.006956478115171194, + 0.00526414206251502, + 0.00271063600666821, + 0.002765249228104949, + 0.0024817127268761396, + 0.0023620878346264362, + 0.004347600042819977, + 0.0028165439143776894, + 0.0019399404991418123, + 0.003270893357694149, + 0.008586919866502285, + 0.00994904525578022, + 0.022751085460186005, + 0.011088642291724682, + 0.013765833340585232, + 0.012006210163235664, + 0.009449363686144352, + 0.01826702617108822, + 0.012235118076205254, + 0.01764632947742939, + 0.011700664646923542, + 0.020196709781885147, + 0.025200283154845238, + 0.019668901339173317, + 0.016408655792474747, + 0.011438285000622272, + 0.010911648161709309, + 0.01905152201652527, + 0.01838555559515953, + 0.010321583598852158, + 0.012312064878642559, + 0.01295418106019497, + 0.012463640421628952, + 0.009202436544001102, + 0.01650460623204708, + 0.017811477184295654, + 0.029452167451381683, + 0.01975773088634014, + 0.017153890803456306, + 0.014967565424740314, + 0.016438474878668785, + 0.015608918853104115, + 0.006685486529022455, + 0.010767724364995956, + 0.007157489191740751, + 0.005010389722883701, + 0.00720078032463789, + 0.005308615043759346, + 0.005812338087707758, + 0.006147607229650021, + 0.006744098849594593, + 0.007055018562823534, + 0.004524759482592344, + 0.0029036677442491055, + 0.006890275981277227, + 0.0046739736571908, + 0.005652700085192919, + 0.0027397635858505964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.001517931348644197 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343408972024918 + ], + "opposite_logprob_tokens": [ + -3.65134334564209 + ], + "logprob_reward": -0.026343408972024918, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343408972024918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1413401961326599, + 0.1796526163816452, + 0.05296531319618225, + 0.04384699463844299, + 0.07778023928403854, + 0.20505213737487793, + 0.1590127944946289, + 0.09808506816625595, + 0.02760244719684124, + 0.014662198722362518 + ], + "credit_map": [ + 0.015411829575896263, + 0.01806425116956234, + 0.012531046755611897, + 0.0181638952344656, + 0.03235621750354767, + 0.021841557696461678, + 0.013985984027385712, + 0.008985423482954502, + 0.016412721946835518, + 0.026511741802096367, + 0.017995040863752365, + 0.0350610725581646, + 0.029420817270874977, + 0.02333645522594452, + 0.018696926534175873, + 0.012217828072607517, + 0.006365697365254164, + 0.008167484775185585, + 0.0038785221986472607, + 0.008018316701054573, + 0.010129844769835472, + 0.006969823036342859, + 0.004301538225263357, + 0.005134090315550566, + 0.004962582141160965, + 0.005189164541661739, + 0.0034451414830982685, + 0.004788684658706188, + 0.011579232290387154, + 0.006427621003240347, + 0.003956032916903496, + 0.0034985344391316175, + 0.008020157925784588, + 0.009216466918587685, + 0.015444094315171242, + 0.007173089310526848, + 0.011997999623417854, + 0.00820204894989729, + 0.006712132599204779, + 0.011014249175786972, + 0.022008119150996208, + 0.026828132569789886, + 0.016722122207283974, + 0.03493821993470192, + 0.0337408147752285, + 0.02776019647717476, + 0.0292025376111269, + 0.013851983472704887, + 0.01947549730539322, + 0.028959117829799652, + 0.01908905617892742, + 0.014938347972929478, + 0.017059432342648506, + 0.017103035002946854, + 0.02477184869349003, + 0.01761646382510662, + 0.011549086309969425, + 0.01348462700843811, + 0.008782443590462208, + 0.008226740173995495, + 0.017894644290208817, + 0.011601671576499939, + 0.01735425926744938, + 0.009191597811877728, + 0.0039046346209943295, + 0.005153820849955082, + 0.003142053261399269, + 0.0021720770746469498, + 0.005270467139780521, + 0.0030393321067094803, + 0.0027525757905095816, + 0.002167487284168601, + 0.0017024942208081484, + 0.0014919344102963805, + 0.003227951005101204, + 0.0016391250537708402, + 0.002564854919910431, + 0.001550497836433351, + 0.0015354064526036382, + 0.000949934939853847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.026343408972024918 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048651598393917084 + ], + "opposite_logprob_tokens": [ + -3.048651695251465 + ], + "logprob_reward": -0.048651598393917084, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048651598393917084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13296157121658325, + 0.07264057546854019, + 0.04947620630264282, + 0.025626404210925102, + 0.1448913812637329, + 0.12276098877191544, + 0.3179389238357544, + 0.06872578710317612, + 0.0439448319375515, + 0.021033283323049545 + ], + "credit_map": [ + 0.012872789055109024, + 0.013977278023958206, + 0.016721414402127266, + 0.018434932455420494, + 0.022448426112532616, + 0.02364560030400753, + 0.0103069469332695, + 0.014554185792803764, + 0.008011482656002045, + 0.009242287836968899, + 0.0074733286164700985, + 0.0071705905720591545, + 0.01787467487156391, + 0.013234086334705353, + 0.004443773068487644, + 0.005190358031541109, + 0.005571096669882536, + 0.005450705997645855, + 0.005691626574844122, + 0.004012661520391703, + 0.01100889127701521, + 0.00907918531447649, + 0.0030282465741038322, + 0.005633790511637926, + 0.0031177899800240993, + 0.002946657594293356, + 0.0032124724239110947, + 0.002893989672884345, + 0.003985374700278044, + 0.0028745762538164854, + 0.002070617862045765, + 0.0045249247923493385, + 0.015844790264964104, + 0.018817974254488945, + 0.030933000147342682, + 0.012813009321689606, + 0.015288105234503746, + 0.017991971224546432, + 0.009661031886935234, + 0.02354149892926216, + 0.01022161915898323, + 0.017050670459866524, + 0.012611729092895985, + 0.019527370110154152, + 0.023320864886045456, + 0.017586421221494675, + 0.01296447403728962, + 0.009477831423282623, + 0.0210711769759655, + 0.02705407328903675, + 0.08865432441234589, + 0.04263978824019432, + 0.027780577540397644, + 0.03475812450051308, + 0.04232390224933624, + 0.033656973391771317, + 0.008955939672887325, + 0.01093814242631197, + 0.005755754187703133, + 0.005344400182366371, + 0.012166180647909641, + 0.007876526564359665, + 0.01125430129468441, + 0.006434541195631027, + 0.0052871061488986015, + 0.00643134955316782, + 0.006829588208347559, + 0.0043653505854308605, + 0.00552434241399169, + 0.004080652724951506, + 0.0050380597822368145, + 0.006388382520526648, + 0.0023204011376947165, + 0.002666348358616233, + 0.002855688566341996, + 0.0025407508946955204, + 0.0037445530761033297, + 0.0028585961554199457, + 0.0021877887193113565, + 0.0018591557163745165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.048651598393917084 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static close-up captures a slender, unfinished wooden stick, its pale brown fibrous texture clearly visible, standing perfectly vertically on the undulating surface of a weathered gray stone. The scene is bathed in a soft, diffuse light, suggestive of an overcast sky or a shady forest floor, which creates gentle, subtle shadows around the base of the stick. The background is a gentle, out-of-focus blur of dark greens and earthy browns, creating a quiet, contemplative atmosphere. After a moment of perfect stillness, the stick begins to sway almost imperceptibly, then slowly and deliberately tilts, its shape gracefully arcing through the air. Upon impact with the stone surface, the fragile wood instantly breaks with sharp cracks, disintegrating into a small cloud of fibrous fragments and dust that scatters from the point of impact, leaving the shattered remains in a final state of stillness.", + "seed": 21779, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_15/index_9111_seed_21779_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.00899627641774714, + "vlm_logprob": -0.00899627641774714, + "vlm_yes_no_margin": 5.475, + "vlm_token_credit": -0.00899627641774714, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012270909734070301 + ], + "opposite_logprob_tokens": [ + -5.387270927429199 + ], + "logprob_reward": -0.012270909734070301, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.012270909734070301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14192403852939606, + 0.021244090050458908, + 0.018788954243063927, + 0.01903371326625347, + 0.5809967517852783, + 0.060669805854558945, + 0.028667455539107323, + 0.057463593780994415, + 0.042922452092170715, + 0.028289202600717545 + ], + "credit_map": [ + 0.014339916408061981, + 0.013785024173557758, + 0.017207462340593338, + 0.03611946851015091, + 0.0065874382853507996, + 0.023442458361387253, + 0.009565425105392933, + 0.02087683044373989, + 0.0018139607273042202, + 0.002925788052380085, + 0.003234189236536622, + 0.0025159793440252542, + 0.0032588010653853416, + 0.0025188871659338474, + 0.002656062599271536, + 0.0023204174358397722, + 0.0019996955525130033, + 0.001781168975867331, + 0.0025477977469563484, + 0.002921484876424074, + 0.002423900179564953, + 0.0027240535710006952, + 0.0021534247789531946, + 0.002237426582723856, + 0.001606541103683412, + 0.0020716474391520023, + 0.0033992489334195852, + 0.0022851130925118923, + 0.003076602704823017, + 0.002527925418689847, + 0.002129702363163233, + 0.001936931163072586, + 0.17081838846206665, + 0.044043924659490585, + 0.03347230702638626, + 0.03731568530201912, + 0.020861884579062462, + 0.23493635654449463, + 0.020088443532586098, + 0.019459718838334084, + 0.007640293799340725, + 0.0055569251999258995, + 0.005520422477275133, + 0.006802775897085667, + 0.008209899999201298, + 0.013521026819944382, + 0.00884916540235281, + 0.004569285083562136, + 0.003130632219836116, + 0.00290210940875113, + 0.0033341138623654842, + 0.0040578776970505714, + 0.003333501983433962, + 0.005139712709933519, + 0.004098760895431042, + 0.0026707411743700504, + 0.005180256441235542, + 0.005911798216402531, + 0.005102602299302816, + 0.007601489778608084, + 0.006283774506300688, + 0.007911953143775463, + 0.012362434528768063, + 0.007109277416020632, + 0.004636377561837435, + 0.003734553698450327, + 0.0041505396366119385, + 0.005626593250781298, + 0.00457436079159379, + 0.008011071011424065, + 0.00725052785128355, + 0.0049384222365915775, + 0.0023183671291917562, + 0.004142523277550936, + 0.004319520201534033, + 0.002677350537851453, + 0.0043794079683721066, + 0.0034601453226059675, + 0.004198150709271431, + 0.0027937341947108507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.012270909734070301 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00172216328792274 + ], + "opposite_logprob_tokens": [ + -6.37672233581543 + ], + "logprob_reward": -0.00172216328792274, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.00172216328792274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18531395494937897, + 0.06959987431764603, + 0.043544139713048935, + 0.036144427955150604, + 0.31854403018951416, + 0.06263209134340286, + 0.07075292617082596, + 0.14195840060710907, + 0.032649409025907516, + 0.038860756903886795 + ], + "credit_map": [ + 0.018357232213020325, + 0.021938631311058998, + 0.028541531413793564, + 0.03703922778367996, + 0.0197858065366745, + 0.02102108858525753, + 0.01696416363120079, + 0.0216662734746933, + 0.0064296601340174675, + 0.00875599030405283, + 0.009163293987512589, + 0.015402130782604218, + 0.005634977016597986, + 0.007349977269768715, + 0.006897988263517618, + 0.009965862147510052, + 0.004797981120646, + 0.004775729961693287, + 0.007867521606385708, + 0.006492364685982466, + 0.00421714736148715, + 0.006679411046206951, + 0.0047591920010745525, + 0.003954790998250246, + 0.004460936412215233, + 0.003699531080201268, + 0.00663229962810874, + 0.004668374080210924, + 0.002723632613196969, + 0.004423567093908787, + 0.0035914951004087925, + 0.005944591481238604, + 0.05727923661470413, + 0.028150366619229317, + 0.051828138530254364, + 0.045464176684617996, + 0.01656149886548519, + 0.07614922523498535, + 0.026532601565122604, + 0.01657881774008274, + 0.007262668572366238, + 0.007624703925102949, + 0.0077344984747469425, + 0.007122069131582975, + 0.009285676293075085, + 0.008444111794233322, + 0.009104043245315552, + 0.006054318509995937, + 0.00686805322766304, + 0.006854495033621788, + 0.010208610445261002, + 0.010521703399717808, + 0.008674203418195248, + 0.00892653688788414, + 0.008594542741775513, + 0.01010478287935257, + 0.012656211853027344, + 0.014270938001573086, + 0.028504684567451477, + 0.01816619746387005, + 0.017354344949126244, + 0.016331154853105545, + 0.022044628858566284, + 0.012630232609808445, + 0.0031073098070919514, + 0.003592623397707939, + 0.00750789325684309, + 0.005315990652889013, + 0.003089577192440629, + 0.004088728688657284, + 0.003177447011694312, + 0.0027698390185832977, + 0.004461360163986683, + 0.003593046683818102, + 0.007643441203981638, + 0.005995626095682383, + 0.0035600776318460703, + 0.005976931191980839, + 0.0038747231010347605, + 0.0037555524613708258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.00172216328792274 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007240534760057926 + ], + "opposite_logprob_tokens": [ + -7.250723838806152 + ], + "logprob_reward": -0.0007240534760057926, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007240534760057926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07362185418605804, + 0.028820309787988663, + 0.012896398082375526, + 0.025547320023179054, + 0.668681263923645, + 0.044809672981500626, + 0.04425645247101784, + 0.040897030383348465, + 0.021223876625299454, + 0.0392458438873291 + ], + "credit_map": [ + 0.008051889017224312, + 0.008267201483249664, + 0.012925323098897934, + 0.013277296908199787, + 0.006063567008823156, + 0.00978747196495533, + 0.006935824174433947, + 0.008313284255564213, + 0.00270185642875731, + 0.002554186386987567, + 0.0044637322425842285, + 0.007434219121932983, + 0.002583135152235627, + 0.003170452546328306, + 0.0023619525600224733, + 0.00355077488347888, + 0.0017737114103510976, + 0.0014253252884373069, + 0.0015969335800036788, + 0.0018712914315983653, + 0.0017096510855481029, + 0.0017526590963825583, + 0.0016614950727671385, + 0.0011053307680413127, + 0.003250029170885682, + 0.0028596746269613504, + 0.003446606220677495, + 0.002423207275569439, + 0.0031259534880518913, + 0.0040934719145298, + 0.0029184038285166025, + 0.003429973730817437, + 0.18259552121162415, + 0.06085098534822464, + 0.046624328941106796, + 0.02283306047320366, + 0.031721990555524826, + 0.2639068365097046, + 0.03314993903040886, + 0.02699856087565422, + 0.005259869620203972, + 0.004547005984932184, + 0.004903458524495363, + 0.005177019163966179, + 0.005814551375806332, + 0.009672114625573158, + 0.0052789379842579365, + 0.004156716633588076, + 0.004153519868850708, + 0.004336253274232149, + 0.007664120756089687, + 0.008160616271197796, + 0.004593179561197758, + 0.006698973476886749, + 0.005342858377844095, + 0.0033069339115172625, + 0.004067773465067148, + 0.00490775890648365, + 0.008434589020907879, + 0.005524047650396824, + 0.0044159418903291225, + 0.0046426537446677685, + 0.005211759824305773, + 0.0036925054155290127, + 0.002200349234044552, + 0.0016660828841850162, + 0.0023928750306367874, + 0.0022361427545547485, + 0.0027129666414111853, + 0.0030902090948075056, + 0.004110861569643021, + 0.0028143899980932474, + 0.00406660046428442, + 0.0035542547702789307, + 0.0038775731809437275, + 0.0049490551464259624, + 0.004471900872886181, + 0.005695604253560305, + 0.00772899342700839, + 0.0049018641002476215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.0007240534760057926 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196122996509075 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014196122996509075, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196122996509075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02908347174525261, + 0.01524419616907835, + 0.01718180440366268, + 0.020560838282108307, + 0.7440465688705444, + 0.07148269563913345, + 0.06193717569112778, + 0.014490767382085323, + 0.020726650953292847, + 0.0052458313293755054 + ], + "credit_map": [ + 0.0033528388012200594, + 0.0038351246621459723, + 0.0051643215119838715, + 0.003419123124331236, + 0.0036219959147274494, + 0.0036619710735976696, + 0.0029458864592015743, + 0.0030822118278592825, + 0.0014947450254112482, + 0.002262062393128872, + 0.00217171898111701, + 0.0023087237495929003, + 0.0015554362908005714, + 0.0019159576622769237, + 0.0015089531661942601, + 0.002026599831879139, + 0.001957167172804475, + 0.002433714922517538, + 0.0025028346572071314, + 0.0021278129424899817, + 0.001369183766655624, + 0.002681278856471181, + 0.001629549078643322, + 0.0024802626576274633, + 0.0020036727655678988, + 0.002102441620081663, + 0.0021515749394893646, + 0.0038631830830127, + 0.0014511365443468094, + 0.002883301116526127, + 0.002138324547559023, + 0.003967205993831158, + 0.20113743841648102, + 0.06623100489377975, + 0.05475788190960884, + 0.038076002150774, + 0.03383544459939003, + 0.284060537815094, + 0.03672679886221886, + 0.02922147326171398, + 0.008684595115482807, + 0.006515635177493095, + 0.010013097897171974, + 0.007694613188505173, + 0.00821262039244175, + 0.013400481082499027, + 0.010595878586173058, + 0.006365770008414984, + 0.008489190600812435, + 0.0049799662083387375, + 0.004861568566411734, + 0.00601794570684433, + 0.006583394482731819, + 0.01272922195494175, + 0.00960320234298706, + 0.008672688156366348, + 0.0013693706132471561, + 0.0020486328285187483, + 0.0013268133625388145, + 0.0010979310609400272, + 0.002989517990499735, + 0.002521565882489085, + 0.0015378613024950027, + 0.0015990756219252944, + 0.0017753777792677283, + 0.0019489324185997248, + 0.0023372650612145662, + 0.0035982809495180845, + 0.001936515560373664, + 0.0021571049001067877, + 0.00417507765814662, + 0.0027980993036180735, + 0.0004982320824638009, + 0.0006490621599368751, + 0.0006020034779794514, + 0.0007030550623312593, + 0.0007272472139447927, + 0.0005791868316009641, + 0.0008376789628528059, + 0.0006493653636425734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.014196122996509075 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606813259422779 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.01606813259422779, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606813259422779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03735944256186485, + 0.01652904413640499, + 0.010859265923500061, + 0.026737039908766747, + 0.6904041171073914, + 0.08763698488473892, + 0.0547361895442009, + 0.05411680415272713, + 0.016540464013814926, + 0.005080639850348234 + ], + "credit_map": [ + 0.0038047742564231157, + 0.004519113339483738, + 0.005847095977514982, + 0.007025981787592173, + 0.0025397203862667084, + 0.005261377897113562, + 0.0032971876207739115, + 0.005064188037067652, + 0.001408558920957148, + 0.0015212241560220718, + 0.0021957841236144304, + 0.004646812099963427, + 0.001024984405376017, + 0.001726709888316691, + 0.0017045310232788324, + 0.002300440799444914, + 0.0009980701142922044, + 0.0013267119647935033, + 0.001601038733497262, + 0.001701798872090876, + 0.0010709569323807955, + 0.0015730062732473016, + 0.0011589186033234, + 0.0014287640806287527, + 0.0023449137806892395, + 0.002279239008203149, + 0.0031263057608157396, + 0.0073209586553275585, + 0.001187062356621027, + 0.0033450154587626457, + 0.002180357463657856, + 0.004953186959028244, + 0.2043984830379486, + 0.06045671924948692, + 0.030695578083395958, + 0.017195818945765495, + 0.03035198152065277, + 0.29067784547805786, + 0.030137905851006508, + 0.02648983709514141, + 0.011825205758213997, + 0.007669792044907808, + 0.010801303200423717, + 0.008524528704583645, + 0.011066773906350136, + 0.022102870047092438, + 0.009186925366520882, + 0.006459587253630161, + 0.007465471047908068, + 0.00449421675875783, + 0.0045140632428228855, + 0.004944419022649527, + 0.0063073416240513325, + 0.011271794326603413, + 0.008246944285929203, + 0.0074919406324625015, + 0.005398119799792767, + 0.0038685707841068506, + 0.0057454826310276985, + 0.007446382660418749, + 0.004965133499354124, + 0.008892656303942204, + 0.010496431961655617, + 0.007304027676582336, + 0.0012866572942584753, + 0.0014069449389353395, + 0.0018385332077741623, + 0.0026511165779083967, + 0.0013542603701353073, + 0.002110233996063471, + 0.0032889002468436956, + 0.0026038161013275385, + 0.0004767808422911912, + 0.0006223092204891145, + 0.0007818429730832577, + 0.0007447234238497913, + 0.0007028532563708723, + 0.0006438890704885125, + 0.0006821380229666829, + 0.00042610306991264224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.01606813259422779 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static close-up captures a slender, unfinished wooden stick, its pale brown fibrous texture clearly visible, standing perfectly vertically on the undulating surface of a weathered gray stone. The scene is bathed in a soft, diffuse light, suggestive of an overcast sky or a shady forest floor, which creates gentle, subtle shadows around the base of the stick. The background is a gentle, out-of-focus blur of dark greens and earthy browns, creating a quiet, contemplative atmosphere. After a moment of perfect stillness, the stick begins to sway almost imperceptibly, then slowly and deliberately tilts, its shape gracefully arcing through the air. Upon impact with the stone surface, the fragile wood instantly breaks with sharp cracks, disintegrating into a small cloud of fibrous fragments and dust that scatters from the point of impact, leaving the shattered remains in a final state of stillness.", + "seed": 21780, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_15/index_9111_seed_21780_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03890844299457967, + "vlm_logprob": -0.03890844299457967, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.03890844299457967, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023979870602488518 + ], + "opposite_logprob_tokens": [ + -4.648979663848877 + ], + "logprob_reward": -0.023979870602488518, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.023979870602488518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.073697030544281, + 0.02498912438750267, + 0.014494814909994602, + 0.03057098761200905, + 0.33272379636764526, + 0.049568597227334976, + 0.08730565011501312, + 0.26002687215805054, + 0.06730864197015762, + 0.059314485639333725 + ], + "credit_map": [ + 0.007856717333197594, + 0.010524528101086617, + 0.013698790222406387, + 0.011632227338850498, + 0.004850768018513918, + 0.010322578251361847, + 0.00790688581764698, + 0.006904534064233303, + 0.0022105248644948006, + 0.0037201244849711657, + 0.004520758520811796, + 0.005003837402909994, + 0.001532147522084415, + 0.0026201328728348017, + 0.0028766500763595104, + 0.0025049475952982903, + 0.0015147741651162505, + 0.002006898168474436, + 0.0022339504212141037, + 0.003012764500454068, + 0.0010003970237448812, + 0.0016753837699070573, + 0.0014357009204104543, + 0.001614946173503995, + 0.0028797334525734186, + 0.004548437427729368, + 0.0040393248200416565, + 0.005993851460516453, + 0.0019250271143391728, + 0.004007916431874037, + 0.003180272178724408, + 0.003996424376964569, + 0.021350525319576263, + 0.07179641723632812, + 0.013157455250620842, + 0.01252200361341238, + 0.1246308907866478, + 0.042481668293476105, + 0.015615876764059067, + 0.031168941408395767, + 0.005958058871328831, + 0.006716530304402113, + 0.007289669010788202, + 0.006538107991218567, + 0.006434171926230192, + 0.006298786494880915, + 0.005809090565890074, + 0.004524181596934795, + 0.005986454896628857, + 0.01881757192313671, + 0.00845846626907587, + 0.00891518872231245, + 0.022716553881764412, + 0.008301407098770142, + 0.007630127482116222, + 0.006479877978563309, + 0.016610968858003616, + 0.04702896252274513, + 0.01763809472322464, + 0.018898559734225273, + 0.07979895919561386, + 0.02619452215731144, + 0.01753142662346363, + 0.036325376480817795, + 0.005583364982157946, + 0.007944656535983086, + 0.008310548029839993, + 0.009123682975769043, + 0.011257391422986984, + 0.006239764858037233, + 0.00594742689281702, + 0.012901808135211468, + 0.0037838665302842855, + 0.013765075244009495, + 0.0031642839312553406, + 0.003641288261860609, + 0.020548086613416672, + 0.006460874807089567, + 0.003848387161269784, + 0.004102619830518961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.023979870602488518 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005966945551335812 + ], + "opposite_logprob_tokens": [ + -5.130967140197754 + ], + "logprob_reward": -0.005966945551335812, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005966945551335812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13889417052268982, + 0.02881786786019802, + 0.04684049263596535, + 0.04698822647333145, + 0.23586666584014893, + 0.0539979562163353, + 0.041754670441150665, + 0.18284717202186584, + 0.08348138630390167, + 0.1405114233493805 + ], + "credit_map": [ + 0.012938923202455044, + 0.02589118666946888, + 0.02217484824359417, + 0.025216620415449142, + 0.005932229105383158, + 0.016835685819387436, + 0.013347918167710304, + 0.016556749120354652, + 0.002576058264821768, + 0.004706728272140026, + 0.004333045333623886, + 0.004963686689734459, + 0.0020458067301660776, + 0.0035923519171774387, + 0.0038441247306764126, + 0.0027560642920434475, + 0.00404933700338006, + 0.008418559096753597, + 0.006818723399192095, + 0.007506579626351595, + 0.0030251836869865656, + 0.006678983569145203, + 0.004732201807200909, + 0.005610920488834381, + 0.0046073696576058865, + 0.008078648708760738, + 0.005017969757318497, + 0.008556777611374855, + 0.0029919634107500315, + 0.007374636363238096, + 0.004154199734330177, + 0.006206661928445101, + 0.029255393892526627, + 0.030383875593543053, + 0.03016606532037258, + 0.0384281724691391, + 0.02699713222682476, + 0.020867260172963142, + 0.03686479479074478, + 0.022903943434357643, + 0.004516263026744127, + 0.006975771393626928, + 0.008467257022857666, + 0.006590082310140133, + 0.008218934759497643, + 0.0064100828021764755, + 0.005976060871034861, + 0.006843500770628452, + 0.0033889366313815117, + 0.006324919406324625, + 0.006933889351785183, + 0.005222856532782316, + 0.006301834713667631, + 0.004388133995234966, + 0.004281720146536827, + 0.004912376403808594, + 0.01171798910945654, + 0.032524820417165756, + 0.014829687774181366, + 0.011975527741014957, + 0.05597556382417679, + 0.02073846571147442, + 0.010432454757392406, + 0.02465265616774559, + 0.005992632359266281, + 0.014018350280821323, + 0.009485816583037376, + 0.009008198976516724, + 0.017216095700860023, + 0.01111755333840847, + 0.006627198308706284, + 0.010015533305704594, + 0.007750723045319319, + 0.03185717761516571, + 0.007027870509773493, + 0.006001680623739958, + 0.049668438732624054, + 0.015541785396635532, + 0.00657957186922431, + 0.016084160655736923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.005966945551335812 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002208415186032653 + ], + "opposite_logprob_tokens": [ + -6.127208232879639 + ], + "logprob_reward": -0.002208415186032653, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002208415186032653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1658906489610672, + 0.023289542645215988, + 0.024632146582007408, + 0.026923656463623047, + 0.25313857197761536, + 0.04999059811234474, + 0.03848911076784134, + 0.2637319564819336, + 0.048597075045108795, + 0.10531678050756454 + ], + "credit_map": [ + 0.01458246260881424, + 0.029646694660186768, + 0.025599047541618347, + 0.030625177547335625, + 0.007638484705239534, + 0.02026638574898243, + 0.018506882712244987, + 0.019025495275855064, + 0.0017524580471217632, + 0.003022726159542799, + 0.0043545616790652275, + 0.00397027563303709, + 0.001491363043896854, + 0.003575253998860717, + 0.0022737544495612383, + 0.002849146258085966, + 0.002537253312766552, + 0.004094560630619526, + 0.004102842882275581, + 0.0037025376223027706, + 0.0017541759880259633, + 0.0030415805522352457, + 0.002696127165108919, + 0.0027030641213059425, + 0.002250684192404151, + 0.0030429782345891, + 0.004117058124393225, + 0.003829683642834425, + 0.002545606344938278, + 0.002831901190802455, + 0.005556581076234579, + 0.002749157138168812, + 0.03266185149550438, + 0.030737943947315216, + 0.034462496638298035, + 0.039065320044755936, + 0.02643253654241562, + 0.02156735025346279, + 0.04180934652686119, + 0.02640167996287346, + 0.004168450832366943, + 0.005967542063444853, + 0.005667745601385832, + 0.004106453619897366, + 0.010404878295958042, + 0.006864689756184816, + 0.0036357007920742035, + 0.009175132028758526, + 0.0030523433815687895, + 0.007311962079256773, + 0.0046730018220841885, + 0.004351180512458086, + 0.00839664600789547, + 0.0034430590458214283, + 0.0034130446147173643, + 0.0038478693459182978, + 0.01738722249865532, + 0.04920471832156181, + 0.015242722816765308, + 0.01581614278256893, + 0.0853457897901535, + 0.029090063646435738, + 0.015812264755368233, + 0.03583298251032829, + 0.004003603011369705, + 0.006981650833040476, + 0.006262180395424366, + 0.006062465254217386, + 0.008272364735603333, + 0.005524205043911934, + 0.004635998047888279, + 0.006854597479104996, + 0.00576674984768033, + 0.023374000564217567, + 0.00602445425465703, + 0.005232404451817274, + 0.03623659163713455, + 0.011749185621738434, + 0.005293266382068396, + 0.011640106327831745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.002208415186032653 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032682120800018 + ], + "opposite_logprob_tokens": [ + -2.3503267765045166 + ], + "logprob_reward": -0.10032682120800018, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032682120800018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1868092268705368, + 0.03261752054095268, + 0.031977444887161255, + 0.0908525139093399, + 0.3187972903251648, + 0.04169073700904846, + 0.025525953620672226, + 0.13458427786827087, + 0.03191500902175903, + 0.10523001104593277 + ], + "credit_map": [ + 0.01513967476785183, + 0.034407760947942734, + 0.02630884200334549, + 0.031480930745601654, + 0.009581640362739563, + 0.021863484755158424, + 0.02358127571642399, + 0.024445632472634315, + 0.0021067762281745672, + 0.0033585766796022654, + 0.0054355342872440815, + 0.006406444124877453, + 0.002541682915762067, + 0.0038097992073744535, + 0.005084263626486063, + 0.0038744418416172266, + 0.0029419464990496635, + 0.004440477583557367, + 0.004739157389849424, + 0.005210387520492077, + 0.0021292180754244328, + 0.0032051638700067997, + 0.0046804482117295265, + 0.00463064294308424, + 0.006665622349828482, + 0.014737003482878208, + 0.0119878388941288, + 0.01540442369878292, + 0.006425953935831785, + 0.01098589226603508, + 0.01308301743119955, + 0.011562760919332504, + 0.03902624174952507, + 0.042344287037849426, + 0.03757735714316368, + 0.04904324188828468, + 0.049411412328481674, + 0.027309082448482513, + 0.04764588549733162, + 0.026439780369400978, + 0.004115129355341196, + 0.004928686656057835, + 0.00555391563102603, + 0.005784247070550919, + 0.006793096195906401, + 0.004890286363661289, + 0.004296906292438507, + 0.0053284685127437115, + 0.001916259410791099, + 0.0043129394762218, + 0.0033627753145992756, + 0.003476417390629649, + 0.003939724527299404, + 0.002798450877889991, + 0.0028880941681563854, + 0.002831291174516082, + 0.008639003150165081, + 0.024916978552937508, + 0.008703724481165409, + 0.0077671632170677185, + 0.043414466083049774, + 0.01563461683690548, + 0.007377673406153917, + 0.018130647018551826, + 0.00243098521605134, + 0.005336932837963104, + 0.003366296412423253, + 0.0036425550933927298, + 0.007219675928354263, + 0.0035478705540299416, + 0.0030072922818362713, + 0.003363401396200061, + 0.005822196137160063, + 0.023739701136946678, + 0.005048159044235945, + 0.004804147407412529, + 0.036747176200151443, + 0.011369043961167336, + 0.005030519794672728, + 0.012669059447944164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.10032682120800018 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle in a static close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a slender, unfinished wooden stick with a pale brown fibrous texture standing vertically on a weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by soft, diffuse light creating gentle shadows, with a background of out-of-focus dark greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick sway imperceptibly, then slowly tilt and arc through the air before breaking upon impact with the stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the broken stick disintegrate into a cloud of fibrous fragments and dust that scatters, leaving shattered remains in a final state of stillness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620601624250412 + ], + "opposite_logprob_tokens": [ + -2.8120601177215576 + ], + "logprob_reward": -0.0620601624250412, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620601624250412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24794739484786987, + 0.024374358355998993, + 0.02806391380727291, + 0.04840545356273651, + 0.3917067050933838, + 0.03476615622639656, + 0.043521225452423096, + 0.040437884628772736, + 0.06361924111843109, + 0.07715767621994019 + ], + "credit_map": [ + 0.018657874315977097, + 0.03935956954956055, + 0.040366291999816895, + 0.04940304532647133, + 0.012989949434995651, + 0.029091771692037582, + 0.030276091769337654, + 0.027802830561995506, + 0.0018779558595269918, + 0.002486508572474122, + 0.004693477880209684, + 0.004919145721942186, + 0.0018606237135827541, + 0.0029858234338462353, + 0.0027651546988636255, + 0.0027856682427227497, + 0.002373760798946023, + 0.0036249333061277866, + 0.0052352831698954105, + 0.005300531163811684, + 0.002243902301415801, + 0.003229176625609398, + 0.0027568174991756678, + 0.003299508709460497, + 0.003684518625959754, + 0.006643155589699745, + 0.005883269943296909, + 0.009901951067149639, + 0.0034895394928753376, + 0.0072224815376102924, + 0.004010260570794344, + 0.007570275571197271, + 0.037429552525281906, + 0.07497220486402512, + 0.028359783813357353, + 0.031257811933755875, + 0.12144720554351807, + 0.04330878332257271, + 0.028878595679998398, + 0.026052797213196754, + 0.003320832271128893, + 0.004800556693226099, + 0.004646590910851955, + 0.003790869377553463, + 0.005494740791618824, + 0.0042706760577857494, + 0.003377946326509118, + 0.005063943564891815, + 0.003478746162727475, + 0.0059592644684016705, + 0.0059344093315303326, + 0.00563194090500474, + 0.005449557676911354, + 0.005311293993145227, + 0.004535561893135309, + 0.007220450788736343, + 0.002700811717659235, + 0.006768639199435711, + 0.002778459805995226, + 0.003046268131583929, + 0.012241490185260773, + 0.004235729109495878, + 0.0024819113314151764, + 0.006184575613588095, + 0.005435914266854525, + 0.010038561187684536, + 0.00826583057641983, + 0.0074602896347641945, + 0.010869811289012432, + 0.005857103504240513, + 0.0072985063306987286, + 0.008393223397433758, + 0.004502877593040466, + 0.017519978806376457, + 0.004174104426056147, + 0.00419236347079277, + 0.0265616737306118, + 0.007781482767313719, + 0.003919497597962618, + 0.00850569549947977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.0620601624250412 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_2/index_tensor([3556, 3556])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_2/index_tensor([3556, 3556])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..55e365e7d79c93e455e8a2d069991bbb93fb143c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_2/index_tensor([3556, 3556])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera maintains a perfectly static, extreme close-up, immersing the viewer in the complex world of a single kumquat's surface. The fruit's rind is a vibrant, fiery orange canvas, its natural dimples and waxy texture rendered in stunning detail. A fine, shimmering mist covers the entire visible surface, with countless tiny droplets catching the light. The scene is illuminated by a single, soft-toned light emanating from the upper left corner, which sculpts the fruit's oval shape with gentle gradients of light and shadow and creates bright, star-shaped, mirrored highlights on each bead. Against a deep, out-of-focus charcoal background, the action unfolds in a satisfyingly physical animation style; smaller droplets carefully merge and coalesce into larger, heavier, crystal-clear spheres. A prominent droplet, having gathered sufficient mass, finally breaks surface tension and begins to drip with surreal movement and palpable weight. Slowly and deliberately, it glides along the fruit's curved surface, its descent carving a wet, shimmering trail that distorts the reflection of the orange skin beneath it, creating an enchanting and tranquil atmosphere focused on the beauty of natural physics.", + "seed": 24404, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_2/index_3556_seed_24404_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2762761004269123, + "vlm_logprob": -0.2762761004269123, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.2762761004269123, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016351042315363884 + ], + "opposite_logprob_tokens": [ + -4.516351222991943 + ], + "logprob_reward": -0.016351042315363884, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.016351042315363884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24191899597644806, + 0.022990280762314796, + 0.02214762009680271, + 0.05750083178281784, + 0.03961530327796936, + 0.02208220772445202, + 0.023070592433214188, + 0.028961334377527237, + 0.38819149136543274, + 0.1535213738679886 + ], + "credit_map": [ + 0.01583072729408741, + 0.07661431282758713, + 0.009368916973471642, + 0.02326219528913498, + 0.06921201944351196, + 0.011260068975389004, + 0.014682638458907604, + 0.02168811857700348, + 0.0029632793739438057, + 0.003884667530655861, + 0.002359377685934305, + 0.0032476535998284817, + 0.0037438557483255863, + 0.001556816860102117, + 0.0031940576154738665, + 0.0020405726972967386, + 0.0019842693582177162, + 0.003436038736253977, + 0.0018236765172332525, + 0.003644173266366124, + 0.00429689884185791, + 0.001753691816702485, + 0.003465625923126936, + 0.001743246684782207, + 0.008080795407295227, + 0.006375562399625778, + 0.006198802962899208, + 0.008321862667798996, + 0.009366507641971111, + 0.005475908052176237, + 0.0066753169521689415, + 0.007006076164543629, + 0.00359301152639091, + 0.009302671998739243, + 0.002297263126820326, + 0.007090629078447819, + 0.007971161045134068, + 0.002535034203901887, + 0.003361594630405307, + 0.0034639411605894566, + 0.002039834391325712, + 0.004294767510145903, + 0.0011729261605069041, + 0.005037971306592226, + 0.0035133534111082554, + 0.0013961698859930038, + 0.0024058865383267403, + 0.0022212984040379524, + 0.003074890235438943, + 0.0038063828833401203, + 0.001767476904205978, + 0.0035887742415070534, + 0.0036695280577987432, + 0.001682952861301601, + 0.003702378598973155, + 0.001778209232725203, + 0.0029461830854415894, + 0.006772736087441444, + 0.002218237379565835, + 0.003953663166612387, + 0.0035269041545689106, + 0.0025105683598667383, + 0.0029894867911934853, + 0.004043555818498135, + 0.031005218625068665, + 0.06990334391593933, + 0.016953101381659508, + 0.03207109123468399, + 0.0677589625120163, + 0.08214139938354492, + 0.024799644947052002, + 0.06355870515108109, + 0.010059203021228313, + 0.006820524577051401, + 0.00468548433855176, + 0.008697839453816414, + 0.01003944594413042, + 0.06098555400967598, + 0.005131472367793322, + 0.04710186645388603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.016351042315363884 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539659261703491 + ], + "opposite_logprob_tokens": [ + -0.4289659559726715 + ], + "logprob_reward": -1.0539659261703491, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539659261703491, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17171430587768555, + 0.042961910367012024, + 0.10237734764814377, + 0.03306780382990837, + 0.05734989047050476, + 0.017470918595790863, + 0.017171425744891167, + 0.02810131385922432, + 0.313655287027359, + 0.21612973511219025 + ], + "credit_map": [ + 0.015348316170275211, + 0.04356769844889641, + 0.009796001017093658, + 0.026654640212655067, + 0.030546927824616432, + 0.011045649647712708, + 0.015667635947465897, + 0.019087443128228188, + 0.002974253846332431, + 0.009972996078431606, + 0.0029606972821056843, + 0.007084475364536047, + 0.007797046564519405, + 0.0032389541156589985, + 0.006254815962165594, + 0.0026786786038428545, + 0.009473271667957306, + 0.018915867432951927, + 0.008383268490433693, + 0.016753965988755226, + 0.017998574301600456, + 0.0074162897653877735, + 0.013328501023352146, + 0.010107623413205147, + 0.0032168792095035315, + 0.00595559598878026, + 0.001741599990054965, + 0.005931139923632145, + 0.006651408039033413, + 0.002515933243557811, + 0.004385300911962986, + 0.0026699472218751907, + 0.0040183402597904205, + 0.01804383285343647, + 0.0023090862669050694, + 0.015496977604925632, + 0.007589999586343765, + 0.0027252042200416327, + 0.0037379013374447823, + 0.003428550437092781, + 0.0019401913741603494, + 0.002563916379585862, + 0.001495549688115716, + 0.003013432491570711, + 0.0026628889609128237, + 0.001635062275454402, + 0.0020991400815546513, + 0.0020607386250048876, + 0.0019173488253727555, + 0.0025608206633478403, + 0.001149738673120737, + 0.0031708297319710255, + 0.003793225856497884, + 0.0009700220543891191, + 0.002577003091573715, + 0.0010324390605092049, + 0.003043379867449403, + 0.0038716355338692665, + 0.0026638254057615995, + 0.003851387882605195, + 0.003528102533891797, + 0.0030213852878659964, + 0.0030827659647911787, + 0.005038832779973745, + 0.022743120789527893, + 0.030454624444246292, + 0.013174740597605705, + 0.02218644693493843, + 0.03249146416783333, + 0.10468842089176178, + 0.02059793472290039, + 0.06731854379177094, + 0.020996788516640663, + 0.07385434955358505, + 0.010444959625601768, + 0.029261743649840355, + 0.04172542691230774, + 0.012252537533640862, + 0.011824428103864193, + 0.015769504010677338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -1.0539659261703491 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02333151362836361 + ], + "opposite_logprob_tokens": [ + -3.7733314037323 + ], + "logprob_reward": -0.02333151362836361, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02333151362836361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10047683864831924, + 0.02899039536714554, + 0.022061724215745926, + 0.034728579223155975, + 0.4789682626724243, + 0.016398197039961815, + 0.008564033545553684, + 0.0214815866202116, + 0.10177676379680634, + 0.18655356764793396 + ], + "credit_map": [ + 0.007256859447807074, + 0.028748266398906708, + 0.005237349774688482, + 0.008657229132950306, + 0.031914692372083664, + 0.005343180149793625, + 0.006413290277123451, + 0.006905966904014349, + 0.00280580366961658, + 0.005454296246170998, + 0.0017061351099982858, + 0.003983229398727417, + 0.006167527753859758, + 0.0017913990886881948, + 0.0040061878971755505, + 0.003075817134231329, + 0.001722361659631133, + 0.00340021844021976, + 0.0022576225455850363, + 0.0030004270374774933, + 0.0039016001392155886, + 0.0019137933850288391, + 0.0032427285332232714, + 0.002622972009703517, + 0.003316891146823764, + 0.005522935185581446, + 0.0029791221022605896, + 0.00565597228705883, + 0.006323625799268484, + 0.003123525995761156, + 0.0030881199054419994, + 0.004718384705483913, + 0.031581852585077286, + 0.11550531536340714, + 0.022747913375496864, + 0.15688921511173248, + 0.05894916504621506, + 0.022413332015275955, + 0.02907530777156353, + 0.041806165128946304, + 0.00144022551830858, + 0.0034663588739931583, + 0.001224603271111846, + 0.0027965251356363297, + 0.0024500847794115543, + 0.0012473518727347255, + 0.0020396383479237556, + 0.0017334092408418655, + 0.0010319337015971541, + 0.001299758441746235, + 0.000659084995277226, + 0.0014841429656371474, + 0.0013342424063012004, + 0.0006224215030670166, + 0.0015120639000087976, + 0.0006203859811648726, + 0.001869507716037333, + 0.004120540339499712, + 0.001138519262894988, + 0.0030296803452074528, + 0.002769755432382226, + 0.0028141916263848543, + 0.0019067260436713696, + 0.003832666203379631, + 0.00800311379134655, + 0.009374741464853287, + 0.004427385050803423, + 0.0068409135565161705, + 0.011825147084891796, + 0.0339253693819046, + 0.006246712524443865, + 0.02113337628543377, + 0.009944331832230091, + 0.06109358370304108, + 0.005645166151225567, + 0.020682241767644882, + 0.03372318297624588, + 0.021958785131573677, + 0.0066225784830749035, + 0.0268836859613657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.02333151362836361 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06210598349571228 + ], + "opposite_logprob_tokens": [ + -2.812105894088745 + ], + "logprob_reward": -0.06210598349571228, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06210598349571228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1561529040336609, + 0.07691781967878342, + 0.1386091709136963, + 0.07211190462112427, + 0.07723075896501541, + 0.019229639321565628, + 0.015596242621541023, + 0.03900184854865074, + 0.10678663849830627, + 0.29836305975914 + ], + "credit_map": [ + 0.013169938698410988, + 0.0474683903157711, + 0.005969073623418808, + 0.01291683316230774, + 0.03961571678519249, + 0.00711134634912014, + 0.012607269920408726, + 0.01729433424770832, + 0.007549377623945475, + 0.014162499457597733, + 0.005999464076012373, + 0.010961310938000679, + 0.01337986160069704, + 0.004568716511130333, + 0.01527499407529831, + 0.005021593999117613, + 0.013094588182866573, + 0.026056936010718346, + 0.014566875994205475, + 0.018239330500364304, + 0.022419298067688942, + 0.010946995578706264, + 0.016868799924850464, + 0.01641635224223137, + 0.006320118438452482, + 0.010063068009912968, + 0.005674859508872032, + 0.01252808328717947, + 0.01181864459067583, + 0.006978283170610666, + 0.009437383152544498, + 0.009291468188166618, + 0.008485988713800907, + 0.014605889096856117, + 0.006154222879558802, + 0.012698928825557232, + 0.013842865824699402, + 0.004828182049095631, + 0.008793092332780361, + 0.007821589708328247, + 0.0016404826892539859, + 0.0028283356223255396, + 0.0014226733474060893, + 0.004748655948787928, + 0.0029045098926872015, + 0.001365043455734849, + 0.0026946866419166327, + 0.0016252524219453335, + 0.001388865290209651, + 0.002954458585008979, + 0.0013744396856054664, + 0.0035029263235628605, + 0.0023697284050285816, + 0.000970659835729748, + 0.0018366981530562043, + 0.0011984661687165499, + 0.0033909615594893694, + 0.007231523282825947, + 0.003600906813517213, + 0.006357564590871334, + 0.0049431161023676395, + 0.004341110587120056, + 0.0036647673696279526, + 0.005471899174153805, + 0.0050948248244822025, + 0.03316763415932655, + 0.002788842422887683, + 0.015283224172890186, + 0.01593777723610401, + 0.02070305496454239, + 0.005357751622796059, + 0.008453522808849812, + 0.01455051638185978, + 0.09787140041589737, + 0.007652013096958399, + 0.03578070551156998, + 0.04871664196252823, + 0.053213730454444885, + 0.0088855205103755, + 0.03169254586100578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.06210598349571228 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22562603652477264 + ], + "opposite_logprob_tokens": [ + -1.600625991821289 + ], + "logprob_reward": -0.22562603652477264, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22562603652477264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19584248960018158, + 0.03623552247881889, + 0.09780524671077728, + 0.05869286134839058, + 0.02702474407851696, + 0.023190397769212723, + 0.0274872537702322, + 0.06273728609085083, + 0.17698970437049866, + 0.2939944565296173 + ], + "credit_map": [ + 0.013671845197677612, + 0.0591416098177433, + 0.00785303208976984, + 0.020234404131770134, + 0.055762771517038345, + 0.009312605485320091, + 0.01532195694744587, + 0.014544273726642132, + 0.0027676112949848175, + 0.008658899925649166, + 0.002175760455429554, + 0.005997882224619389, + 0.006954181008040905, + 0.0023814677260816097, + 0.004966641776263714, + 0.0023330766707658768, + 0.008497846312820911, + 0.019069215282797813, + 0.009620392695069313, + 0.01279342919588089, + 0.014960539527237415, + 0.007217323873192072, + 0.013297430239617825, + 0.012349066324532032, + 0.00585536751896143, + 0.008204332552850246, + 0.0040204450488090515, + 0.011370368301868439, + 0.010724246501922607, + 0.004730203188955784, + 0.007316633127629757, + 0.006471266038715839, + 0.0017415902111679316, + 0.0073785982094705105, + 0.0013448518002405763, + 0.006128882057964802, + 0.004926621913909912, + 0.0015603398205712438, + 0.0020512097980827093, + 0.001892652129754424, + 0.0020506237633526325, + 0.0035634301602840424, + 0.002121635712683201, + 0.004172707442194223, + 0.003635210683569312, + 0.0022921073250472546, + 0.002994276350364089, + 0.002360407030209899, + 0.0026076696813106537, + 0.004229223355650902, + 0.0022434997372329235, + 0.004655260127037764, + 0.005132635124027729, + 0.0017250762321054935, + 0.005126995034515858, + 0.001766894944012165, + 0.005910676904022694, + 0.007900348864495754, + 0.00399969844147563, + 0.009137213230133057, + 0.008269666694104671, + 0.008488268591463566, + 0.006934361066669226, + 0.012097052298486233, + 0.013090865686535835, + 0.01628188043832779, + 0.008490851148962975, + 0.020916562527418137, + 0.02180463820695877, + 0.052185043692588806, + 0.014108410105109215, + 0.03011145070195198, + 0.010968251153826714, + 0.08330021053552628, + 0.006719684228301048, + 0.02781037986278534, + 0.04217274487018585, + 0.07007498294115067, + 0.007821780629456043, + 0.045126453042030334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.22562603652477264 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera maintains a perfectly static, extreme close-up, immersing the viewer in the complex world of a single kumquat's surface. The fruit's rind is a vibrant, fiery orange canvas, its natural dimples and waxy texture rendered in stunning detail. A fine, shimmering mist covers the entire visible surface, with countless tiny droplets catching the light. The scene is illuminated by a single, soft-toned light emanating from the upper left corner, which sculpts the fruit's oval shape with gentle gradients of light and shadow and creates bright, star-shaped, mirrored highlights on each bead. Against a deep, out-of-focus charcoal background, the action unfolds in a satisfyingly physical animation style; smaller droplets carefully merge and coalesce into larger, heavier, crystal-clear spheres. A prominent droplet, having gathered sufficient mass, finally breaks surface tension and begins to drip with surreal movement and palpable weight. Slowly and deliberately, it glides along the fruit's curved surface, its descent carving a wet, shimmering trail that distorts the reflection of the orange skin beneath it, creating an enchanting and tranquil atmosphere focused on the beauty of natural physics.", + "seed": 24405, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_2/index_3556_seed_24405_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.29660825356841086, + "vlm_logprob": -0.29660825356841086, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.29660825356841086, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020210525020956993 + ], + "opposite_logprob_tokens": [ + -4.3952107429504395 + ], + "logprob_reward": -0.020210525020956993, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.020210525020956993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23198802769184113, + 0.16822467744350433, + 0.2338518351316452, + 0.047433074563741684, + 0.07708245515823364, + 0.04358161240816116, + 0.058329302817583084, + 0.06096203625202179, + 0.054782409220933914, + 0.023764563724398613 + ], + "credit_map": [ + 0.02567051723599434, + 0.012925004586577415, + 0.024241527542471886, + 0.022335689514875412, + 0.017527487128973007, + 0.03782693296670914, + 0.0697750523686409, + 0.021685827523469925, + 0.03023819997906685, + 0.010186396539211273, + 0.009821407496929169, + 0.017191000282764435, + 0.013603338040411472, + 0.01681465655565262, + 0.05397263541817665, + 0.01639707013964653, + 0.037359271198511124, + 0.016861578449606895, + 0.012024928815662861, + 0.020778687670826912, + 0.02032800391316414, + 0.037118155509233475, + 0.05854886397719383, + 0.030832374468445778, + 0.01336982473731041, + 0.002980804769322276, + 0.0025916199665516615, + 0.005602645222097635, + 0.003199645783752203, + 0.0023345141671597958, + 0.012854658998548985, + 0.004499363247305155, + 0.016880860552191734, + 0.004004573915153742, + 0.0034548253752291203, + 0.00792644638568163, + 0.00487739872187376, + 0.006220159586519003, + 0.026870161294937134, + 0.006848039571195841, + 0.005126051139086485, + 0.0036135814152657986, + 0.003349313046783209, + 0.0058651152066886425, + 0.005561117548495531, + 0.0068956296890974045, + 0.009318288415670395, + 0.003852519439533353, + 0.009604510851204395, + 0.003456065198406577, + 0.0038407668471336365, + 0.0077424002811312675, + 0.005534888245165348, + 0.005703031085431576, + 0.015587267465889454, + 0.006860376801341772, + 0.0075781322084367275, + 0.004495123401284218, + 0.0060441140085458755, + 0.00836543831974268, + 0.0064137098379433155, + 0.009948132559657097, + 0.01110889483243227, + 0.007008498068898916, + 0.00888005830347538, + 0.004383299965411425, + 0.004375612363219261, + 0.00869500357657671, + 0.006239145062863827, + 0.00518426951020956, + 0.011176676489412785, + 0.005848345812410116, + 0.00358060491271317, + 0.0016196563374251127, + 0.002098876517266035, + 0.004090888891369104, + 0.0018671423895284534, + 0.0019060601480305195, + 0.005678421352058649, + 0.0029229167848825455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.020210525020956993 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.223686695098877 + ], + "opposite_logprob_tokens": [ + -0.34868669509887695 + ], + "logprob_reward": -1.223686695098877, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.223686695098877, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11731229722499847, + 0.14185373485088348, + 0.19050614535808563, + 0.058067552745342255, + 0.1604461669921875, + 0.06386565417051315, + 0.1554056704044342, + 0.09114813059568405, + 0.012865900062024593, + 0.008528727106750011 + ], + "credit_map": [ + 0.014872642233967781, + 0.010324480943381786, + 0.01621342822909355, + 0.01721358858048916, + 0.013571974821388721, + 0.00917931366711855, + 0.02210308611392975, + 0.013833790086209774, + 0.031878504902124405, + 0.008005788549780846, + 0.007254838012158871, + 0.020168056711554527, + 0.010332430712878704, + 0.00937004666775465, + 0.04152583330869675, + 0.013318237848579884, + 0.024448087438941002, + 0.023255301639437675, + 0.013595037162303925, + 0.015391331166028976, + 0.027258271351456642, + 0.03225524351000786, + 0.024701161310076714, + 0.02960170991718769, + 0.011924115009605885, + 0.004967867862433195, + 0.004754848312586546, + 0.008726021274924278, + 0.004235250875353813, + 0.004482379648834467, + 0.013101442717015743, + 0.005875626113265753, + 0.036160968244075775, + 0.010778638534247875, + 0.009875661693513393, + 0.016667179763317108, + 0.012532224878668785, + 0.01055795419961214, + 0.0493716225028038, + 0.014501932077109814, + 0.008055450394749641, + 0.0033818171359598637, + 0.0045752329751849174, + 0.008102938532829285, + 0.008553995750844479, + 0.00793695542961359, + 0.014757199212908745, + 0.008502064272761345, + 0.026935843750834465, + 0.013302971608936787, + 0.012040308676660061, + 0.01564440131187439, + 0.01958554796874523, + 0.012304860167205334, + 0.03800486400723457, + 0.017586877569556236, + 0.02741767093539238, + 0.004490502178668976, + 0.004475951194763184, + 0.009019101969897747, + 0.004740612581372261, + 0.004783547017723322, + 0.028128735721111298, + 0.008092010393738747, + 0.0022639392409473658, + 0.00161882140673697, + 0.0010377581929787993, + 0.0020137701649218798, + 0.001462517655454576, + 0.000869900337420404, + 0.002197326859459281, + 0.0014018655056133866, + 0.0018729185685515404, + 0.0006354805082082748, + 0.0005534037481993437, + 0.001054032938554883, + 0.000687230087351054, + 0.000398099742596969, + 0.002471842570230365, + 0.0008557187975384295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -1.223686695098877 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016098516061902046 + ], + "opposite_logprob_tokens": [ + -4.141098499298096 + ], + "logprob_reward": -0.016098516061902046, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016098516061902046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15925835072994232, + 0.2789894938468933, + 0.12689942121505737, + 0.09819971024990082, + 0.13319231569766998, + 0.03581387177109718, + 0.027106337249279022, + 0.06740099936723709, + 0.03146437928080559, + 0.04167511314153671 + ], + "credit_map": [ + 0.024443453177809715, + 0.012319779954850674, + 0.02851334773004055, + 0.016826296225190163, + 0.01280830055475235, + 0.015576625242829323, + 0.02678975649178028, + 0.02198079414665699, + 0.07007307559251785, + 0.016810908913612366, + 0.011089548468589783, + 0.03023873083293438, + 0.02650952711701393, + 0.014570316299796104, + 0.08296174556016922, + 0.02673562988638878, + 0.03248896077275276, + 0.006135913543403149, + 0.005102510564029217, + 0.010139785706996918, + 0.008200733922421932, + 0.01300365012139082, + 0.04209406301379204, + 0.009733802638947964, + 0.023781852796673775, + 0.006226599216461182, + 0.005160684231668711, + 0.006997270043939352, + 0.006076637655496597, + 0.015230872668325901, + 0.021575571969151497, + 0.01315021887421608, + 0.036650411784648895, + 0.006702416110783815, + 0.007145155221223831, + 0.008952817879617214, + 0.009034437127411366, + 0.017740881070494652, + 0.02654028870165348, + 0.020425904542207718, + 0.004704625811427832, + 0.0025128417182713747, + 0.002342597348615527, + 0.004858403466641903, + 0.005491858813911676, + 0.005521396175026894, + 0.005992976948618889, + 0.0043891724199056625, + 0.0030421216506510973, + 0.0029847228433936834, + 0.002055682707577944, + 0.0037778674159199, + 0.005474779289215803, + 0.0026021511293947697, + 0.003597379894927144, + 0.0035716311540454626, + 0.01967134326696396, + 0.0024559081066399813, + 0.0017955757211893797, + 0.006471135187894106, + 0.0034345940221101046, + 0.003979139029979706, + 0.02347109094262123, + 0.00612221471965313, + 0.006729703396558762, + 0.002114590723067522, + 0.0019283865112811327, + 0.004380704835057259, + 0.003269595792517066, + 0.0017810488352552056, + 0.008192832581698895, + 0.0030675148591399193, + 0.003696116851642728, + 0.0031129620037972927, + 0.0027516870759427547, + 0.0039032611530274153, + 0.003504127962514758, + 0.008330532349646091, + 0.011153076775372028, + 0.005223351996392012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.016098516061902046 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043108273297548294 + ], + "opposite_logprob_tokens": [ + -3.1681082248687744 + ], + "logprob_reward": -0.043108273297548294, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043108273297548294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0861118733882904, + 0.0815364196896553, + 0.16763265430927277, + 0.07061237096786499, + 0.04811815172433853, + 0.07915166765451431, + 0.1131478101015091, + 0.20404458045959473, + 0.13995416462421417, + 0.009690338745713234 + ], + "credit_map": [ + 0.010855046100914478, + 0.004773592110723257, + 0.008413659408688545, + 0.008584413677453995, + 0.00677389744669199, + 0.012935689650475979, + 0.02397594042122364, + 0.009799633175134659, + 0.016631610691547394, + 0.004811340011656284, + 0.0040362230502069, + 0.009891497902572155, + 0.008499139919877052, + 0.006301420275121927, + 0.024799535050988197, + 0.0065656546503305435, + 0.03067888133227825, + 0.008504359982907772, + 0.0078484071418643, + 0.012917804531753063, + 0.014764305204153061, + 0.02437022514641285, + 0.049398232251405716, + 0.019150421023368835, + 0.01574048399925232, + 0.004633471369743347, + 0.004417563322931528, + 0.006060084793716669, + 0.00433101924136281, + 0.006882913410663605, + 0.02123420499265194, + 0.00731263030320406, + 0.009571193717420101, + 0.004058241844177246, + 0.0038006678223609924, + 0.006341447122395039, + 0.00461867731064558, + 0.00553378090262413, + 0.008200893178582191, + 0.005993248894810677, + 0.02051168866455555, + 0.004994935356080532, + 0.0051894523203372955, + 0.008631441742181778, + 0.0055076065473258495, + 0.005537447053939104, + 0.020038090646266937, + 0.00874100811779499, + 0.019146757200360298, + 0.01090816967189312, + 0.008588957600295544, + 0.015100331045687199, + 0.015101436525583267, + 0.009159562177956104, + 0.022985083982348442, + 0.01215751189738512, + 0.04790191724896431, + 0.009148328565061092, + 0.011591494083404541, + 0.016178922727704048, + 0.014394750818610191, + 0.018419597297906876, + 0.06733594834804535, + 0.01907361112535, + 0.026470083743333817, + 0.011544464156031609, + 0.009807480499148369, + 0.01694156974554062, + 0.01370201725512743, + 0.010457216762006283, + 0.03622206673026085, + 0.014809259213507175, + 0.001784354797564447, + 0.0010080011561512947, + 0.0007932721055112779, + 0.000994784408248961, + 0.0008061985718086362, + 0.0011259570019319654, + 0.0020865059923380613, + 0.0010912641882896423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.043108273297548294 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17993725836277008 + ], + "opposite_logprob_tokens": [ + -1.8049372434616089 + ], + "logprob_reward": -0.17993725836277008, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17993725836277008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1669553965330124, + 0.10297661274671555, + 0.3781932592391968, + 0.043308623135089874, + 0.01811944507062435, + 0.04647684842348099, + 0.12616387009620667, + 0.06531836092472076, + 0.043181195855140686, + 0.009306449443101883 + ], + "credit_map": [ + 0.02262265980243683, + 0.0127053651958704, + 0.018625283613801003, + 0.0226871557533741, + 0.015476335771381855, + 0.01551129762083292, + 0.03507493808865547, + 0.024252349510788918, + 0.030369263142347336, + 0.005278907250612974, + 0.0040636505000293255, + 0.009403951466083527, + 0.007281288504600525, + 0.006115205120295286, + 0.029255690053105354, + 0.01120867021381855, + 0.040713559836149216, + 0.03853651136159897, + 0.02309384010732174, + 0.03778255730867386, + 0.07024583965539932, + 0.05745173245668411, + 0.049871888011693954, + 0.060497306287288666, + 0.009274356998503208, + 0.003197324927896261, + 0.0020846789702773094, + 0.005621784366667271, + 0.0028620713856071234, + 0.0037434969563037157, + 0.012151138857007027, + 0.004373769275844097, + 0.0030478094704449177, + 0.0013938101474195719, + 0.0016915567684918642, + 0.0017338781617581844, + 0.0017571980133652687, + 0.003725471207872033, + 0.0018698998028412461, + 0.0028998220805078745, + 0.015685947611927986, + 0.002656033495441079, + 0.0026029141154140234, + 0.0038843743968755007, + 0.002541070803999901, + 0.00353291560895741, + 0.010838833637535572, + 0.004734760615974665, + 0.020048893988132477, + 0.012485915794968605, + 0.00989434216171503, + 0.016698678955435753, + 0.019730228930711746, + 0.009421437047421932, + 0.024144375696778297, + 0.013740005902945995, + 0.017948945984244347, + 0.0022530644200742245, + 0.002588867209851742, + 0.0049535008147358894, + 0.0042236871086061, + 0.005295948591083288, + 0.022014888003468513, + 0.00603945879265666, + 0.008283707313239574, + 0.0041041867807507515, + 0.0031044769566506147, + 0.005644364282488823, + 0.004647166468203068, + 0.0029084356501698494, + 0.010084174573421478, + 0.004404683131724596, + 0.002075474476441741, + 0.0007500157225877047, + 0.0006863578455522656, + 0.0011070901528000832, + 0.0008530667400918901, + 0.0017037533689290285, + 0.001007072045467794, + 0.0011236188001930714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.17993725836277008 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera maintains a perfectly static, extreme close-up, immersing the viewer in the complex world of a single kumquat's surface. The fruit's rind is a vibrant, fiery orange canvas, its natural dimples and waxy texture rendered in stunning detail. A fine, shimmering mist covers the entire visible surface, with countless tiny droplets catching the light. The scene is illuminated by a single, soft-toned light emanating from the upper left corner, which sculpts the fruit's oval shape with gentle gradients of light and shadow and creates bright, star-shaped, mirrored highlights on each bead. Against a deep, out-of-focus charcoal background, the action unfolds in a satisfyingly physical animation style; smaller droplets carefully merge and coalesce into larger, heavier, crystal-clear spheres. A prominent droplet, having gathered sufficient mass, finally breaks surface tension and begins to drip with surreal movement and palpable weight. Slowly and deliberately, it glides along the fruit's curved surface, its descent carving a wet, shimmering trail that distorts the reflection of the orange skin beneath it, creating an enchanting and tranquil atmosphere focused on the beauty of natural physics.", + "seed": 24406, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_2/index_3556_seed_24406_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.30295051336288453, + "vlm_logprob": -0.30295051336288453, + "vlm_yes_no_margin": 1.775, + "vlm_token_credit": -0.30295051336288453, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10486047714948654 + ], + "opposite_logprob_tokens": [ + -2.354860544204712 + ], + "logprob_reward": -0.10486047714948654, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10486047714948654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5706810355186462, + 0.11026600003242493, + 0.057524945586919785, + 0.08096543699502945, + 0.025522056967020035, + 0.013394888490438461, + 0.03630037605762482, + 0.014404472894966602, + 0.03776214271783829, + 0.053178571164608 + ], + "credit_map": [ + 0.025279203429818153, + 0.03255131095647812, + 0.027009762823581696, + 0.07451453059911728, + 0.09179429709911346, + 0.053453341126441956, + 0.07246875017881393, + 0.19360986351966858, + 0.004681961145251989, + 0.010834184475243092, + 0.008322558365762234, + 0.017911169677972794, + 0.013104096055030823, + 0.011591554619371891, + 0.017322778701782227, + 0.026497697457671165, + 0.002585240639746189, + 0.0050934311002492905, + 0.004537736531347036, + 0.007967259734869003, + 0.005995649378746748, + 0.009675069712102413, + 0.008553854189813137, + 0.01311669871211052, + 0.00463761854916811, + 0.008982780389487743, + 0.0064561436884105206, + 0.011838400736451149, + 0.009249295108020306, + 0.009080404415726662, + 0.011559562757611275, + 0.019161226227879524, + 0.0015914436662569642, + 0.004007280804216862, + 0.003291674889624119, + 0.003029313636943698, + 0.0035135371144860983, + 0.0034914491698145866, + 0.0031381044536828995, + 0.003459256375208497, + 0.0009681478841230273, + 0.0020008194260299206, + 0.001185544184409082, + 0.0015807700110599399, + 0.0016232472844421864, + 0.001429966650903225, + 0.0018537392606958747, + 0.0027526523917913437, + 0.0013347441563382745, + 0.005259771831333637, + 0.0017526868032291532, + 0.006955578923225403, + 0.010882668197154999, + 0.0017888749716803432, + 0.0029885454569011927, + 0.0053375037387013435, + 0.0009620044729672372, + 0.003556442679837346, + 0.0007916264585219324, + 0.0014475566567853093, + 0.002127930289134383, + 0.0012625535018742085, + 0.001586540718562901, + 0.002669818000867963, + 0.0023508076556026936, + 0.006465878803282976, + 0.003068361198529601, + 0.0038571516051888466, + 0.006144828163087368, + 0.005265179555863142, + 0.004201157949864864, + 0.006408775690943003, + 0.002850933466106653, + 0.007088845130056143, + 0.006823946721851826, + 0.006289677694439888, + 0.006786482874304056, + 0.007817236706614494, + 0.0064162095077335835, + 0.009105237200856209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.10486047714948654 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539385080337524 + ], + "opposite_logprob_tokens": [ + -0.42893853783607483 + ], + "logprob_reward": -1.0539385080337524, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539385080337524, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21739740669727325, + 0.32576411962509155, + 0.05466382950544357, + 0.04061977565288544, + 0.03103630058467388, + 0.0316314771771431, + 0.1326511949300766, + 0.03388411924242973, + 0.07036198675632477, + 0.06198979914188385 + ], + "credit_map": [ + 0.00784843135625124, + 0.016273675486445427, + 0.008093171752989292, + 0.0261337012052536, + 0.037589870393276215, + 0.013811235316097736, + 0.02753239870071411, + 0.0801149308681488, + 0.013254005461931229, + 0.04185353219509125, + 0.024198800325393677, + 0.030541887506842613, + 0.049855947494506836, + 0.02618030272424221, + 0.04733791947364807, + 0.09254170954227448, + 0.00469421548768878, + 0.006128245033323765, + 0.006676258984953165, + 0.005390068516135216, + 0.007267010863870382, + 0.006099266931414604, + 0.005600668024271727, + 0.012808091938495636, + 0.0019342813175171614, + 0.005032011773437262, + 0.003241445869207382, + 0.0066949836909770966, + 0.004748315084725618, + 0.0040785204619169235, + 0.00605992553755641, + 0.008830289356410503, + 0.002291695214807987, + 0.004756011534482241, + 0.0033746117260307074, + 0.003466793103143573, + 0.004297327250242233, + 0.0037913713604211807, + 0.00354675087146461, + 0.005511739291250706, + 0.0017867068527266383, + 0.004347320180386305, + 0.002110925270244479, + 0.0031113633885979652, + 0.005589955020695925, + 0.002806782955303788, + 0.004254434257745743, + 0.0076239872723817825, + 0.005090062972158194, + 0.019380344077944756, + 0.006744245998561382, + 0.022318782284855843, + 0.03940608724951744, + 0.01003139466047287, + 0.010717899538576603, + 0.018962373957037926, + 0.0025205276906490326, + 0.005130468867719173, + 0.002943112514913082, + 0.0042534759268164635, + 0.004410149529576302, + 0.004687383770942688, + 0.00466609513387084, + 0.00527290441095829, + 0.005604967009276152, + 0.010962114669382572, + 0.007899129763245583, + 0.007500699255615473, + 0.010441114194691181, + 0.009346318431198597, + 0.007888900116086006, + 0.010718749836087227, + 0.006144074257463217, + 0.010364284738898277, + 0.007573693059384823, + 0.0072612641379237175, + 0.0076586599461734295, + 0.004119073506444693, + 0.006211973261088133, + 0.012656779028475285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -1.0539385080337524 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.070074163377285 + ], + "opposite_logprob_tokens": [ + -2.6950740814208984 + ], + "logprob_reward": -0.070074163377285, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.070074163377285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21452446281909943, + 0.23957745730876923, + 0.10202697664499283, + 0.09774527698755264, + 0.09070766717195511, + 0.047178808599710464, + 0.06612665206193924, + 0.022948630154132843, + 0.03989000990986824, + 0.07927405089139938 + ], + "credit_map": [ + 0.011955264024436474, + 0.02833939529955387, + 0.016269812360405922, + 0.03427351266145706, + 0.0230574794113636, + 0.02301642671227455, + 0.04104243963956833, + 0.03657013922929764, + 0.010135410353541374, + 0.027961980551481247, + 0.015975380316376686, + 0.023674998432397842, + 0.034090008586645126, + 0.017101382836699486, + 0.0360136479139328, + 0.07462462782859802, + 0.008972580544650555, + 0.015967881307005882, + 0.010983457788825035, + 0.011678623966872692, + 0.0135151082649827, + 0.008227133192121983, + 0.011598359793424606, + 0.021083831787109375, + 0.004805779550224543, + 0.00966021791100502, + 0.007713251747190952, + 0.013553024269640446, + 0.011504133231937885, + 0.012994352728128433, + 0.01307518221437931, + 0.024439333006739616, + 0.007554025389254093, + 0.016388241201639175, + 0.00969710573554039, + 0.010578878223896027, + 0.011221127584576607, + 0.008524258621037006, + 0.011114945635199547, + 0.01562909409403801, + 0.0023765750229358673, + 0.0072210608050227165, + 0.0021432028152048588, + 0.005084910895675421, + 0.007169557735323906, + 0.003256740514189005, + 0.007360436487942934, + 0.012566323392093182, + 0.0022889962419867516, + 0.006208636332303286, + 0.0024233469739556313, + 0.01328890398144722, + 0.02124762162566185, + 0.003123319009318948, + 0.006331913638859987, + 0.011213915422558784, + 0.0015994684072211385, + 0.0054965633898973465, + 0.0016446382505819201, + 0.002006006892770529, + 0.0034736511297523975, + 0.0015828616451472044, + 0.0023110767360776663, + 0.0048343646340072155, + 0.002595392521470785, + 0.00630480982363224, + 0.005258830264210701, + 0.004709217231720686, + 0.004736066795885563, + 0.005548113491386175, + 0.0043444521725177765, + 0.006393128540366888, + 0.009051288478076458, + 0.010866551660001278, + 0.013171483762562275, + 0.009636597707867622, + 0.009377562440931797, + 0.00809421855956316, + 0.008858026005327702, + 0.010218320414423943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.070074163377285 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033723458647727966 + ], + "opposite_logprob_tokens": [ + -3.4087233543395996 + ], + "logprob_reward": -0.033723458647727966, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033723458647727966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20679208636283875, + 0.15247325599193573, + 0.060477908700704575, + 0.07983294129371643, + 0.2604519724845886, + 0.054060790687799454, + 0.04100708290934563, + 0.03237156197428703, + 0.04937232658267021, + 0.06316012144088745 + ], + "credit_map": [ + 0.00890068244189024, + 0.009978976100683212, + 0.008043539710342884, + 0.023817753419280052, + 0.03889453411102295, + 0.015740925446152687, + 0.023662999272346497, + 0.07775267213582993, + 0.007133181672543287, + 0.018536560237407684, + 0.011715345084667206, + 0.014420794323086739, + 0.023025663569569588, + 0.013798314146697521, + 0.019912773743271828, + 0.043930619955062866, + 0.0040772478096187115, + 0.0054438794031739235, + 0.005158490035682917, + 0.0077958558686077595, + 0.0069229197688400745, + 0.007057827897369862, + 0.0075722429901361465, + 0.016449449583888054, + 0.004832376725971699, + 0.014719181694090366, + 0.006147891283035278, + 0.009227369911968708, + 0.012270980514585972, + 0.007545233704149723, + 0.00869964063167572, + 0.01639026589691639, + 0.014145641587674618, + 0.060088563710451126, + 0.016982141882181168, + 0.021565206348896027, + 0.049974214285612106, + 0.02222912386059761, + 0.02652398869395256, + 0.04894308000802994, + 0.004184219986200333, + 0.013309668749570847, + 0.003611844265833497, + 0.005319732706993818, + 0.00807386077940464, + 0.004942071624100208, + 0.005432494916021824, + 0.009186900220811367, + 0.001610057195648551, + 0.004301920998841524, + 0.0018406083108857274, + 0.009898418560624123, + 0.010652469471096992, + 0.0032801220659166574, + 0.0049703288823366165, + 0.004453158937394619, + 0.0022244893480092287, + 0.004994362127035856, + 0.002765586832538247, + 0.004133710637688637, + 0.003997736144810915, + 0.004639110062271357, + 0.0039718253538012505, + 0.005644736811518669, + 0.0036006206646561623, + 0.0072885374538600445, + 0.004800277762115002, + 0.005610928870737553, + 0.006500282324850559, + 0.005798587575554848, + 0.005089535377919674, + 0.010683558881282806, + 0.0046835411339998245, + 0.010052107274532318, + 0.006092809606343508, + 0.008703294210135937, + 0.007347147911787033, + 0.004884057678282261, + 0.00910705141723156, + 0.01229011919349432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.033723458647727966 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25215595960617065 + ], + "opposite_logprob_tokens": [ + -1.5021560192108154 + ], + "logprob_reward": -0.25215595960617065, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25215595960617065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2045401781797409, + 0.12371429055929184, + 0.04003751277923584, + 0.12097033858299255, + 0.17150679230690002, + 0.11962155252695084, + 0.05981504172086716, + 0.07850166410207748, + 0.044357359409332275, + 0.036935243755578995 + ], + "credit_map": [ + 0.011887772008776665, + 0.022570064291357994, + 0.015150099992752075, + 0.025882568210363388, + 0.029918484389781952, + 0.017219895496964455, + 0.030248796567320824, + 0.05166248604655266, + 0.007321051321923733, + 0.0186205692589283, + 0.015087603591382504, + 0.01333597395569086, + 0.017393821850419044, + 0.01176904421299696, + 0.01541428454220295, + 0.02477194182574749, + 0.0035188833717256784, + 0.00442051887512207, + 0.003744151210412383, + 0.005253582261502743, + 0.004235506057739258, + 0.006119648925960064, + 0.005015007685869932, + 0.0077302162535488605, + 0.004325282759964466, + 0.017639320343732834, + 0.008796201087534428, + 0.01916458085179329, + 0.01406039111316204, + 0.012179655022919178, + 0.016522206366062164, + 0.028282711282372475, + 0.007441441994160414, + 0.03519027680158615, + 0.013296810910105705, + 0.02045915275812149, + 0.02068493328988552, + 0.019163012504577637, + 0.01980692520737648, + 0.035464249551296234, + 0.006419243291020393, + 0.024393638595938683, + 0.005368402693420649, + 0.009619037620723248, + 0.017815830186009407, + 0.00982504803687334, + 0.014187349006533623, + 0.031992994248867035, + 0.002825530944392085, + 0.005236020311713219, + 0.0032341431360691786, + 0.01427783165127039, + 0.01817896217107773, + 0.004265215247869492, + 0.005706076975911856, + 0.006091266870498657, + 0.005641448777168989, + 0.017127683386206627, + 0.005881279706954956, + 0.008853775449097157, + 0.010659473948180676, + 0.008823617361485958, + 0.009622529149055481, + 0.011891855858266354, + 0.0038085812702775, + 0.007490587420761585, + 0.0069556282833218575, + 0.004782691597938538, + 0.005481358617544174, + 0.005034745670855045, + 0.005167112685739994, + 0.005636658053845167, + 0.003528999863192439, + 0.00573075283318758, + 0.0038885141257196665, + 0.004217406269162893, + 0.0045947907492518425, + 0.0027155831921845675, + 0.00422296766191721, + 0.008036229759454727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.25215595960617065 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera maintains a perfectly static, extreme close-up, immersing the viewer in the complex world of a single kumquat's surface. The fruit's rind is a vibrant, fiery orange canvas, its natural dimples and waxy texture rendered in stunning detail. A fine, shimmering mist covers the entire visible surface, with countless tiny droplets catching the light. The scene is illuminated by a single, soft-toned light emanating from the upper left corner, which sculpts the fruit's oval shape with gentle gradients of light and shadow and creates bright, star-shaped, mirrored highlights on each bead. Against a deep, out-of-focus charcoal background, the action unfolds in a satisfyingly physical animation style; smaller droplets carefully merge and coalesce into larger, heavier, crystal-clear spheres. A prominent droplet, having gathered sufficient mass, finally breaks surface tension and begins to drip with surreal movement and palpable weight. Slowly and deliberately, it glides along the fruit's curved surface, its descent carving a wet, shimmering trail that distorts the reflection of the orange skin beneath it, creating an enchanting and tranquil atmosphere focused on the beauty of natural physics.", + "seed": 24407, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_2/index_3556_seed_24407_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2611043196171522, + "vlm_logprob": -0.2611043196171522, + "vlm_yes_no_margin": 2.699999988079071, + "vlm_token_credit": -0.2611043196171522, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011916285380721092 + ], + "opposite_logprob_tokens": [ + -5.136916160583496 + ], + "logprob_reward": -0.011916285380721092, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.011916285380721092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24685868620872498, + 0.06555110216140747, + 0.031351298093795776, + 0.047148577868938446, + 0.17495183646678925, + 0.06080875173211098, + 0.07240872830152512, + 0.13742485642433167, + 0.09427288174629211, + 0.06922324746847153 + ], + "credit_map": [ + 0.013386519625782967, + 0.008017726242542267, + 0.044011592864990234, + 0.011339827440679073, + 0.013799925334751606, + 0.00820526760071516, + 0.08970393240451813, + 0.05839388817548752, + 0.005737312138080597, + 0.002715039299800992, + 0.012599133886396885, + 0.0031056085135787725, + 0.003058159025385976, + 0.002824794501066208, + 0.020415710285305977, + 0.015095345675945282, + 0.0015218835324048996, + 0.0018660813802853227, + 0.0012556869769468904, + 0.0015895258402451873, + 0.0017978480318561196, + 0.0017054247437044978, + 0.011621331796050072, + 0.009993516840040684, + 0.005241479258984327, + 0.0021573363337665796, + 0.01202285848557949, + 0.0020882338285446167, + 0.0019025555811822414, + 0.0021984437480568886, + 0.013543866574764252, + 0.007993806153535843, + 0.009949427098035812, + 0.022140715271234512, + 0.0109196025878191, + 0.01615537330508232, + 0.016265857964754105, + 0.020870158448815346, + 0.042769014835357666, + 0.035881686955690384, + 0.006015897262841463, + 0.003907026723027229, + 0.005582308396697044, + 0.008095341734588146, + 0.006093743722885847, + 0.00399519270285964, + 0.01238489430397749, + 0.01473434641957283, + 0.0049292827025055885, + 0.0059645515866577625, + 0.003722794819623232, + 0.0058134328573942184, + 0.011854086071252823, + 0.005161438137292862, + 0.020155971869826317, + 0.014807171188294888, + 0.009061542339622974, + 0.005751299671828747, + 0.006527949124574661, + 0.01108164805918932, + 0.0155798913910985, + 0.008755304850637913, + 0.04641102999448776, + 0.03425620496273041, + 0.008665301837027073, + 0.0035698949359357357, + 0.004647078923881054, + 0.005447289440780878, + 0.005162193905562162, + 0.0034977903123944998, + 0.035626087337732315, + 0.027657242491841316, + 0.004336191341280937, + 0.003872079774737358, + 0.0034546649549156427, + 0.007326265797019005, + 0.010186359286308289, + 0.005641602911055088, + 0.0174803975969553, + 0.016925686970353127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.011916285380721092 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.137089490890503 + ], + "opposite_logprob_tokens": [ + -0.38708943128585815 + ], + "logprob_reward": -1.137089490890503, + "yes_no_margin_reward": -0.7500000596046448, + "token_credit_reward": -1.137089490890503, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2761227488517761, + 0.07324168086051941, + 0.07108799368143082, + 0.21596761047840118, + 0.05103439465165138, + 0.17967581748962402, + 0.09388696402311325, + 0.020997609943151474, + 0.010059063322842121, + 0.00792608316987753 + ], + "credit_map": [ + 0.009984290227293968, + 0.005468105431646109, + 0.043084122240543365, + 0.008803687058389187, + 0.00984374899417162, + 0.006078466307371855, + 0.1153206080198288, + 0.07753975689411163, + 0.01005642767995596, + 0.008959249593317509, + 0.01946340687572956, + 0.005701296031475067, + 0.00606036651879549, + 0.006278141867369413, + 0.0074507188983261585, + 0.009272073395550251, + 0.002732324181124568, + 0.0020684609189629555, + 0.008796579204499722, + 0.0024838484823703766, + 0.0027230100240558386, + 0.0018936751876026392, + 0.02884523943066597, + 0.021544862538576126, + 0.011578889563679695, + 0.004711751360446215, + 0.04011763632297516, + 0.007315835449844599, + 0.006666778586804867, + 0.005425933748483658, + 0.08361854404211044, + 0.05653226003050804, + 0.005700203124433756, + 0.005419126711785793, + 0.004685789346694946, + 0.0057241166941821575, + 0.0068699815310537815, + 0.005146988667547703, + 0.00943749025464058, + 0.008050701580941677, + 0.01206905860453844, + 0.02331235632300377, + 0.007960923947393894, + 0.013363232836127281, + 0.022090155631303787, + 0.01297042891383171, + 0.04773908853530884, + 0.0401705838739872, + 0.005532031878829002, + 0.008448240347206593, + 0.002932376228272915, + 0.0061346073634922504, + 0.005390842445194721, + 0.004538095090538263, + 0.03325098380446434, + 0.027659788727760315, + 0.00106519041582942, + 0.0005838572978973389, + 0.0006844148156233132, + 0.0012683597160503268, + 0.001195946941152215, + 0.0009473480749875307, + 0.008398681879043579, + 0.006853812839835882, + 0.0013208524323999882, + 0.0005249576643109322, + 0.0008057412342168391, + 0.0006704656407237053, + 0.0005482042906805873, + 0.0004595161881297827, + 0.0032223581802099943, + 0.002506968332454562, + 0.0005134661914780736, + 0.0004670896159950644, + 0.0004951934097334743, + 0.0007817425648681819, + 0.0010189746972173452, + 0.0012106845388188958, + 0.0017849855357781053, + 0.0016539463540539145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -1.137089490890503 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337321013212204 + ], + "opposite_logprob_tokens": [ + -3.4087321758270264 + ], + "logprob_reward": -0.0337321013212204, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337321013212204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08155378699302673, + 0.14702998101711273, + 0.020257504656910896, + 0.12141504883766174, + 0.06837699562311172, + 0.05771613493561745, + 0.19686473906040192, + 0.13852961361408234, + 0.060291845351457596, + 0.1079644188284874 + ], + "credit_map": [ + 0.006068599410355091, + 0.003972872160375118, + 0.01593378186225891, + 0.00522019574418664, + 0.0053682695142924786, + 0.0036678689066320658, + 0.024988366290926933, + 0.016333824023604393, + 0.018360290676355362, + 0.011648163199424744, + 0.0414106510579586, + 0.008178476244211197, + 0.007462981157004833, + 0.007796024903655052, + 0.034265678375959396, + 0.017907705157995224, + 0.002167182508856058, + 0.0014240613672882318, + 0.003896721638739109, + 0.0015457759145647287, + 0.0022879382595419884, + 0.0014104789588600397, + 0.00395707692950964, + 0.0035682686138898134, + 0.00865601934492588, + 0.0034491342958062887, + 0.025977911427617073, + 0.0040982915088534355, + 0.003568673273548484, + 0.0035274315159767866, + 0.04390878602862358, + 0.02822880446910858, + 0.006096031051129103, + 0.004070260562002659, + 0.004000035580247641, + 0.00409088097512722, + 0.004486200399696827, + 0.003101970301941037, + 0.024429187178611755, + 0.01810242421925068, + 0.0026178378611803055, + 0.0037777526304125786, + 0.004929710645228624, + 0.0032092835754156113, + 0.00568033754825592, + 0.0028305042069405317, + 0.02021664008498192, + 0.014454062096774578, + 0.011723854579031467, + 0.01047977153211832, + 0.00606287969276309, + 0.011486200615763664, + 0.011598457582294941, + 0.00858535896986723, + 0.07576809823513031, + 0.06116010248661041, + 0.00807899422943592, + 0.0030027020256966352, + 0.0047470396384596825, + 0.006452259607613087, + 0.007410626858472824, + 0.005253545939922333, + 0.057761307805776596, + 0.045823127031326294, + 0.007648192811757326, + 0.010574768297374249, + 0.005311538465321064, + 0.005259721074253321, + 0.006668108515441418, + 0.004558676388114691, + 0.010624460875988007, + 0.009646378457546234, + 0.013890947215259075, + 0.0058297282084822655, + 0.015897665172815323, + 0.008423671126365662, + 0.01398224662989378, + 0.01584096997976303, + 0.017664896324276924, + 0.01643429510295391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.0337321013212204 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372691571712494 + ], + "opposite_logprob_tokens": [ + -3.408726930618286 + ], + "logprob_reward": -0.03372691571712494, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372691571712494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06970628350973129, + 0.6339812278747559, + 0.025684857740998268, + 0.06925172358751297, + 0.0582352951169014, + 0.03893638774752617, + 0.012286455370485783, + 0.02333795465528965, + 0.03459548205137253, + 0.03398424759507179 + ], + "credit_map": [ + 0.0049812630750238895, + 0.003073025494813919, + 0.0038572645280510187, + 0.006049133837223053, + 0.005877566058188677, + 0.0037228106521070004, + 0.022819017991423607, + 0.019326208159327507, + 0.07956778258085251, + 0.044874925166368484, + 0.1738792508840561, + 0.03307538107037544, + 0.02835320495069027, + 0.03069908171892166, + 0.1569923311471939, + 0.08653933554887772, + 0.0018019997514784336, + 0.0011511306511238217, + 0.0013535413891077042, + 0.00158989941701293, + 0.0021112007088959217, + 0.0016667734598740935, + 0.008177372626960278, + 0.00783294066786766, + 0.0023042764514684677, + 0.0012655681930482388, + 0.008744989521801472, + 0.002166279125958681, + 0.0029024011455476284, + 0.0015815530205145478, + 0.029065167531371117, + 0.02122148871421814, + 0.0040562874637544155, + 0.004257061518728733, + 0.004222941119223833, + 0.0032552573829889297, + 0.004662650171667337, + 0.0043822238221764565, + 0.01908312737941742, + 0.014315749518573284, + 0.0035390176344662905, + 0.005599750205874443, + 0.0035480763763189316, + 0.0037644249387085438, + 0.006491596344858408, + 0.0033990154042840004, + 0.007415872998535633, + 0.0051786405965685844, + 0.0007370849489234388, + 0.0008880570530891418, + 0.00046766671584919095, + 0.0008632260141894221, + 0.0007862644270062447, + 0.0006719902739860117, + 0.004281853791326284, + 0.0035903132520616055, + 0.001250502304174006, + 0.0006715058698318899, + 0.0006881266017444432, + 0.0012543012853711843, + 0.0010297740809619427, + 0.0008938700775615871, + 0.00959690473973751, + 0.007952972315251827, + 0.0035177660174667835, + 0.004136215429753065, + 0.0025130112189799547, + 0.002036243909969926, + 0.002799034584313631, + 0.0017895398195832968, + 0.009822319261729717, + 0.007981356233358383, + 0.002733300905674696, + 0.0015885144239291549, + 0.001791608752682805, + 0.00320530217140913, + 0.0038834838196635246, + 0.0026129784528166056, + 0.00944527331739664, + 0.0087237898260355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.03372691571712494 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up of a single kumquat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, out-of-focus charcoal color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single soft-toned light source illuminate the scene from the upper left corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a prominent droplet break surface tension and glide along the fruit's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905680477619171 + ], + "opposite_logprob_tokens": [ + -2.4640567302703857 + ], + "logprob_reward": -0.08905680477619171, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905680477619171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1893286108970642, + 0.18921396136283875, + 0.03671994060277939, + 0.12539801001548767, + 0.051564473658800125, + 0.29594162106513977, + 0.09498555958271027, + 0.004227961879223585, + 0.006528571248054504, + 0.006091328337788582 + ], + "credit_map": [ + 0.006536356173455715, + 0.0037139635533094406, + 0.02411620505154133, + 0.006990014109760523, + 0.006747048813849688, + 0.004453161731362343, + 0.0808616653084755, + 0.05591018125414848, + 0.024948619306087494, + 0.01574757695198059, + 0.05499729886651039, + 0.011027530767023563, + 0.010184499435126781, + 0.010448369197547436, + 0.04075147584080696, + 0.02110856957733631, + 0.0018343478441238403, + 0.0013912502909079194, + 0.005212578922510147, + 0.0015561893815174699, + 0.0019108932465314865, + 0.0014229402877390385, + 0.013472774997353554, + 0.009918960742652416, + 0.007341661024838686, + 0.0029420785140246153, + 0.024574603885412216, + 0.004316484089940786, + 0.003801625920459628, + 0.0034234407357871532, + 0.04733363911509514, + 0.031664468348026276, + 0.004734441637992859, + 0.0021529081277549267, + 0.0020592103246599436, + 0.0033048850018531084, + 0.0033235265873372555, + 0.0026728659868240356, + 0.019133219495415688, + 0.014183409512043, + 0.0219934843480587, + 0.04430050030350685, + 0.014581943862140179, + 0.024255283176898956, + 0.042013391852378845, + 0.024087844416499138, + 0.06729359924793243, + 0.057415563613176346, + 0.0056019010953605175, + 0.006624946370720863, + 0.0028340553399175406, + 0.00566406873986125, + 0.004999037366360426, + 0.004197754431515932, + 0.03587700054049492, + 0.02918679267168045, + 0.00032440933864563704, + 0.00044510391307994723, + 0.00027911795768886805, + 0.0003569532127585262, + 0.000340050901286304, + 0.0003064652264583856, + 0.0011509554460644722, + 0.001024905708618462, + 0.0006851794314570725, + 0.00038551882607862353, + 0.00035935433697886765, + 0.0003654058091342449, + 0.00040562200592830777, + 0.0002564180176705122, + 0.0022984629031270742, + 0.0017726095393300056, + 0.0005973278894089162, + 0.0002900005201809108, + 0.0005648296209983528, + 0.0006074864650145173, + 0.0007741256849840283, + 0.0004712266963906586, + 0.0014544884907081723, + 0.0013318427372723818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.08905680477619171 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_3/index_tensor([5672, 5672])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_3/index_tensor([5672, 5672])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..baee258ef5689e5ab94faabacbea4aadfb268fb4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_3/index_tensor([5672, 5672])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a medium shot shot, shot slightly from above, and executed in extreme slow motion, a chivalrous hero, clad in a resplendent royal blue silk robe, his long, jet-black hair tied high, a few strands flowing around his cheeks, stands in a dimly lit scripture library. His expression is resolute, his eyes focused. A stack of ancient scrolls, their edges curled with age and their pages yellowed with age, rests on a heavy mahogany desk before him. He thrusts out his leather-guarded right hand and strikes the bottom of the pile with a swift, precise force. As if ignited by an invisible force, the entire pile explodes upward and in all directions. Countless yellowed pages unfold, twirl, and spin in a frozen passage of time, ancient calligraphy glimmering in the dance. A dramatic beam of light, filtering through a high, carved wooden window, captures this moment with pinpoint precision, making each stirred particle of ancient dust shimmer like golden stardust within the beam, the fragile edges of the paper appearing brightly outlined against the backlight. The entire painting is a play of light and shadow, rich in detail. The deep wooden background contrasts sharply with the hero's brightly colored attire, capturing a classical martial arts aesthetic imbued with power and destiny in this frozen moment, creating a solemn and visually striking atmosphere.", + "seed": 5222, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_3/index_5672_seed_5222_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01629364341497421, + "vlm_logprob": -0.01629364341497421, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.01629364341497421, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021380886435508728 + ], + "opposite_logprob_tokens": [ + -4.02138090133667 + ], + "logprob_reward": -0.021380886435508728, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.021380886435508728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.210987389087677, + 0.16777664422988892, + 0.09728497266769409, + 0.16095849871635437, + 0.06453295052051544, + 0.035545989871025085, + 0.052168358117341995, + 0.028277123346924782, + 0.14302664995193481, + 0.039441391825675964 + ], + "credit_map": [ + 0.015837321057915688, + 0.013274919241666794, + 0.04281110689043999, + 0.020124346017837524, + 0.01726483926177025, + 0.02104688063263893, + 0.02455892600119114, + 0.05606904998421669, + 0.009568275883793831, + 0.01495958212763071, + 0.01537142600864172, + 0.011970769613981247, + 0.012780706398189068, + 0.05880250036716461, + 0.014334255829453468, + 0.02998914010822773, + 0.006492702756077051, + 0.008253470063209534, + 0.014062169939279556, + 0.00932620745152235, + 0.006911620032042265, + 0.03455911949276924, + 0.008770679123699665, + 0.008909003809094429, + 0.01155979372560978, + 0.013678904622793198, + 0.013122666627168655, + 0.013427555561065674, + 0.01851409673690796, + 0.05286795645952225, + 0.01564325951039791, + 0.022144276648759842, + 0.004336641635745764, + 0.00909968838095665, + 0.010323136113584042, + 0.0055112880654633045, + 0.007954027503728867, + 0.008704693987965584, + 0.0080260606482625, + 0.010577411390841007, + 0.002693410497158766, + 0.0065861358307302, + 0.006908806972205639, + 0.0030289478600025177, + 0.004410580266267061, + 0.003478922415524721, + 0.0030484581366181374, + 0.005390728823840618, + 0.003508192952722311, + 0.010792492888867855, + 0.005912082735449076, + 0.006635344121605158, + 0.008417549543082714, + 0.005468826275318861, + 0.005635301116853952, + 0.005798570346087217, + 0.002050854964181781, + 0.0044365073554217815, + 0.0021908925846219063, + 0.0038656694814562798, + 0.007766822353005409, + 0.002175259869545698, + 0.003416745224967599, + 0.0023743705824017525, + 0.008878020569682121, + 0.045990169048309326, + 0.03393232822418213, + 0.013236936181783676, + 0.01285621803253889, + 0.010241631418466568, + 0.008226409554481506, + 0.009664942510426044, + 0.002989864209666848, + 0.008366156369447708, + 0.008021366782486439, + 0.004422427155077457, + 0.003911594860255718, + 0.0026586721651256084, + 0.005103368777781725, + 0.003967939410358667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.021380886435508728 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978849083185196 + ], + "opposite_logprob_tokens": [ + -4.634788513183594 + ], + "logprob_reward": -0.00978849083185196, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978849083185196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09133478254079819, + 0.20377707481384277, + 0.12400585412979126, + 0.26664215326309204, + 0.06958287209272385, + 0.04136129468679428, + 0.0704532116651535, + 0.040885407477617264, + 0.046077027916908264, + 0.045880407094955444 + ], + "credit_map": [ + 0.008390093222260475, + 0.00827107485383749, + 0.011064345017075539, + 0.011622591875493526, + 0.012626590207219124, + 0.010135002434253693, + 0.01608191803097725, + 0.013143165037035942, + 0.011147139593958855, + 0.018353890627622604, + 0.02094036340713501, + 0.012857753783464432, + 0.0120805399492383, + 0.07108696550130844, + 0.014445852488279343, + 0.04286456108093262, + 0.009985450655221939, + 0.010431963950395584, + 0.023124387487769127, + 0.008752863854169846, + 0.008557071909308434, + 0.04348664730787277, + 0.00955531932413578, + 0.010112136602401733, + 0.01940951496362686, + 0.022971047088503838, + 0.03465345501899719, + 0.020905541256070137, + 0.030195532366633415, + 0.08893406391143799, + 0.021598322317004204, + 0.027974706143140793, + 0.004850976634770632, + 0.0041529275476932526, + 0.01601947471499443, + 0.005998979322612286, + 0.005984609946608543, + 0.004864959046244621, + 0.005107671953737736, + 0.022603262215852737, + 0.003595386864617467, + 0.011098215356469154, + 0.004420949146151543, + 0.004671425558626652, + 0.006948318798094988, + 0.0036529861390590668, + 0.003809361020103097, + 0.0031646504066884518, + 0.0054068234749138355, + 0.017646754160523415, + 0.011278255842626095, + 0.007383505813777447, + 0.011236000806093216, + 0.005625099875032902, + 0.005033691413700581, + 0.006843077950179577, + 0.002908524824306369, + 0.007471882272511721, + 0.0038458157796412706, + 0.005496402736753225, + 0.009920869953930378, + 0.002947860164567828, + 0.004435091745108366, + 0.0038589586038142443, + 0.004104495979845524, + 0.006508201826363802, + 0.006936390418559313, + 0.006588882766664028, + 0.006869744509458542, + 0.006019131280481815, + 0.00473183486610651, + 0.004318345338106155, + 0.0036714563611894846, + 0.01231017429381609, + 0.008901044726371765, + 0.004454914480447769, + 0.0038782923948019743, + 0.0038804123178124428, + 0.00455428333953023, + 0.0042298296466469765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.00978849083185196 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191891998052597 + ], + "opposite_logprob_tokens": [ + -4.2641921043396 + ], + "logprob_reward": -0.014191891998052597, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191891998052597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1312297284603119, + 0.06910717487335205, + 0.11390171200037003, + 0.16802966594696045, + 0.07378694415092468, + 0.08616341650485992, + 0.06907324492931366, + 0.049747344106435776, + 0.20655053853988647, + 0.03241022676229477 + ], + "credit_map": [ + 0.010198063217103481, + 0.007445155177265406, + 0.03270825743675232, + 0.009388216771185398, + 0.008301526308059692, + 0.009512891061604023, + 0.010908345691859722, + 0.04276726022362709, + 0.003940517548471689, + 0.006476277485489845, + 0.0069849854335188866, + 0.007888317108154297, + 0.005239718593657017, + 0.023260347545146942, + 0.005587116815149784, + 0.009729892946779728, + 0.0070290337316691875, + 0.010288321413099766, + 0.011718997731804848, + 0.006839717272669077, + 0.007204035762697458, + 0.0423579216003418, + 0.007246599067002535, + 0.02121707983314991, + 0.012005978263914585, + 0.013230494223535061, + 0.0317787304520607, + 0.009710895828902721, + 0.016510264948010445, + 0.05575093626976013, + 0.011481404304504395, + 0.017560966312885284, + 0.004040639381855726, + 0.010297597385942936, + 0.012119175866246223, + 0.007582130376249552, + 0.008474993519484997, + 0.004892381373792887, + 0.00725697074085474, + 0.019123053178191185, + 0.006061890162527561, + 0.017975669354200363, + 0.015303904190659523, + 0.007314774207770824, + 0.010118947364389896, + 0.007975883781909943, + 0.008801656775176525, + 0.01261068694293499, + 0.005016859620809555, + 0.01737690530717373, + 0.008217754773795605, + 0.007832039147615433, + 0.009526222944259644, + 0.007428843528032303, + 0.006397909484803677, + 0.007276710122823715, + 0.005751283373683691, + 0.007069775369018316, + 0.004183993674814701, + 0.004769155755639076, + 0.007857686839997768, + 0.00415772246196866, + 0.005105884280055761, + 0.01085184421390295, + 0.015220333822071552, + 0.05921798199415207, + 0.04593612998723984, + 0.022567039355635643, + 0.021189847961068153, + 0.014300449751317501, + 0.01488307025283575, + 0.013235684484243393, + 0.0036223523784428835, + 0.005055075976997614, + 0.0059408582746982574, + 0.004155176691710949, + 0.00337145384401083, + 0.003217614721506834, + 0.0035859716590493917, + 0.003461720421910286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.014191891998052597 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775033220648766 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.009775033220648766, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775033220648766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1568450629711151, + 0.07031721621751785, + 0.029382716864347458, + 0.11789782345294952, + 0.10488387197256088, + 0.15581156313419342, + 0.11207078397274017, + 0.021354245021939278, + 0.19700194895267487, + 0.03443477675318718 + ], + "credit_map": [ + 0.011283377185463905, + 0.008290569297969341, + 0.04076182097196579, + 0.010692098177969456, + 0.011843929998576641, + 0.012996261939406395, + 0.013970427215099335, + 0.047006577253341675, + 0.004242385737597942, + 0.005639573559165001, + 0.007928655482828617, + 0.00845979806035757, + 0.004936256445944309, + 0.017659427598118782, + 0.005979777779430151, + 0.015471341088414192, + 0.0028847292996942997, + 0.0024176852311939, + 0.0031222524121403694, + 0.0028258380480110645, + 0.0034641253296285868, + 0.007403254508972168, + 0.003575993934646249, + 0.003688835771754384, + 0.009047988802194595, + 0.009839472360908985, + 0.020513176918029785, + 0.008649281226098537, + 0.012635629624128342, + 0.031029129400849342, + 0.00999145582318306, + 0.01619168557226658, + 0.006254294421523809, + 0.009409565478563309, + 0.021651964634656906, + 0.00714130187407136, + 0.009866457432508469, + 0.008616798557341099, + 0.007386715151369572, + 0.03455676883459091, + 0.011227371171116829, + 0.03132869675755501, + 0.029405148699879646, + 0.012536090798676014, + 0.01803506165742874, + 0.014400933869183064, + 0.015402469784021378, + 0.02347578853368759, + 0.008805207908153534, + 0.031546451151371, + 0.012860417366027832, + 0.0110336197540164, + 0.01753617636859417, + 0.010252401232719421, + 0.008382324129343033, + 0.01165419165045023, + 0.0023325018119066954, + 0.002878340194001794, + 0.0021399925462901592, + 0.002744872821494937, + 0.003962788265198469, + 0.001962019130587578, + 0.0022296642418950796, + 0.0031040662433952093, + 0.018136808648705482, + 0.04255853220820427, + 0.03348162770271301, + 0.0262504443526268, + 0.025324884802103043, + 0.017689481377601624, + 0.018855538219213486, + 0.014704634435474873, + 0.002417605370283127, + 0.008416671305894852, + 0.00808509811758995, + 0.003135903039947152, + 0.0026444089598953724, + 0.0030337951611727476, + 0.0029032519087195396, + 0.0037980417255312204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.009775033220648766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026331914588809013 + ], + "opposite_logprob_tokens": [ + -3.651331901550293 + ], + "logprob_reward": -0.026331914588809013, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026331914588809013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08242306113243103, + 0.0433771014213562, + 0.1170278787612915, + 0.3628070056438446, + 0.02212880738079548, + 0.1367313414812088, + 0.05849070101976395, + 0.0900280624628067, + 0.06962360441684723, + 0.01736241579055786 + ], + "credit_map": [ + 0.0063970619812607765, + 0.00588091230019927, + 0.019663652405142784, + 0.006189534440636635, + 0.005884438753128052, + 0.008812949992716312, + 0.0077137453481554985, + 0.021880771964788437, + 0.003040231065824628, + 0.004167019855231047, + 0.006919161882251501, + 0.00423060730099678, + 0.0024682108778506517, + 0.00966561958193779, + 0.0031189103610813618, + 0.009767336770892143, + 0.01041097566485405, + 0.013114748522639275, + 0.010940227657556534, + 0.008604573085904121, + 0.00774451345205307, + 0.03991377726197243, + 0.0060090539045631886, + 0.020290011540055275, + 0.034255098551511765, + 0.029973968863487244, + 0.0640997663140297, + 0.01680932752788067, + 0.02637743018567562, + 0.13383689522743225, + 0.018369024619460106, + 0.039085503667593, + 0.0015064772451296449, + 0.0027792176697403193, + 0.0033507791813462973, + 0.0019273418001830578, + 0.003442413406446576, + 0.0014656942803412676, + 0.0026825335808098316, + 0.0049743494018912315, + 0.01078955177217722, + 0.020373737439513206, + 0.029642004519701004, + 0.010661140084266663, + 0.01798871159553528, + 0.014111875556409359, + 0.015088307671248913, + 0.018076026812195778, + 0.004756077192723751, + 0.014846251346170902, + 0.006297907792031765, + 0.007684113457798958, + 0.009362776763737202, + 0.005035560578107834, + 0.005367188714444637, + 0.005140827968716621, + 0.011179623194038868, + 0.01600242406129837, + 0.016393261030316353, + 0.010958355851471424, + 0.012799111194908619, + 0.008464362472295761, + 0.006070255301892757, + 0.00816066563129425, + 0.007621792145073414, + 0.015851696953177452, + 0.00878469180315733, + 0.00792186614125967, + 0.009193281643092632, + 0.00874208752065897, + 0.006083767861127853, + 0.005424420814961195, + 0.0016330870566889644, + 0.0034532954450696707, + 0.0038696706760674715, + 0.0015717644710093737, + 0.0015467365738004446, + 0.0014112457865849137, + 0.0015737893991172314, + 0.0023028263822197914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.026331914588809013 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a medium shot shot, shot slightly from above, and executed in extreme slow motion, a chivalrous hero, clad in a resplendent royal blue silk robe, his long, jet-black hair tied high, a few strands flowing around his cheeks, stands in a dimly lit scripture library. His expression is resolute, his eyes focused. A stack of ancient scrolls, their edges curled with age and their pages yellowed with age, rests on a heavy mahogany desk before him. He thrusts out his leather-guarded right hand and strikes the bottom of the pile with a swift, precise force. As if ignited by an invisible force, the entire pile explodes upward and in all directions. Countless yellowed pages unfold, twirl, and spin in a frozen passage of time, ancient calligraphy glimmering in the dance. A dramatic beam of light, filtering through a high, carved wooden window, captures this moment with pinpoint precision, making each stirred particle of ancient dust shimmer like golden stardust within the beam, the fragile edges of the paper appearing brightly outlined against the backlight. The entire painting is a play of light and shadow, rich in detail. The deep wooden background contrasts sharply with the hero's brightly colored attire, capturing a classical martial arts aesthetic imbued with power and destiny in this frozen moment, creating a solemn and visually striking atmosphere.", + "seed": 5223, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_3/index_5672_seed_5223_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05552466660737991, + "vlm_logprob": -0.05552466660737991, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.05552466660737991, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05807754024863243 + ], + "opposite_logprob_tokens": [ + -2.933077573776245 + ], + "logprob_reward": -0.05807754024863243, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05807754024863243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17538143694400787, + 0.0707746297121048, + 0.05570416897535324, + 0.13079951703548431, + 0.040304023772478104, + 0.03973302245140076, + 0.2054470330476761, + 0.06885264068841934, + 0.16720527410507202, + 0.04579827934503555 + ], + "credit_map": [ + 0.026612209156155586, + 0.02134912833571434, + 0.02198825217783451, + 0.024140173569321632, + 0.021518368273973465, + 0.016972601413726807, + 0.02487334981560707, + 0.017927372828125954, + 0.0075631397776305676, + 0.009047328494489193, + 0.010152638889849186, + 0.010203410871326923, + 0.008422660641372204, + 0.007901489734649658, + 0.011231744661927223, + 0.00625221710652113, + 0.006515662185847759, + 0.007489332463592291, + 0.008592364378273487, + 0.011041847988963127, + 0.006194517016410828, + 0.005863394122570753, + 0.005378519184887409, + 0.004628533497452736, + 0.011163397692143917, + 0.025856129825115204, + 0.016502732411026955, + 0.017047900706529617, + 0.01444984506815672, + 0.008384744636714458, + 0.02549627050757408, + 0.011898485012352467, + 0.003931035287678242, + 0.0052267760038375854, + 0.006313113961368799, + 0.006505999248474836, + 0.005363275296986103, + 0.005358545109629631, + 0.0041296775452792645, + 0.0034756010863929987, + 0.004312966484576464, + 0.0033703697845339775, + 0.005514739081263542, + 0.004324649926275015, + 0.006193215027451515, + 0.008573795668780804, + 0.0034340808633714914, + 0.004009207710623741, + 0.025409093126654625, + 0.020656464621424675, + 0.026774978265166283, + 0.030463991686701775, + 0.023717064410448074, + 0.038953039795160294, + 0.01568826101720333, + 0.023784123361110687, + 0.010033308528363705, + 0.005648160353302956, + 0.013529674150049686, + 0.008994038216769695, + 0.011989179998636246, + 0.006411177106201649, + 0.005889532621949911, + 0.006357566919177771, + 0.02147585153579712, + 0.012802166864275932, + 0.03178917244076729, + 0.02124323695898056, + 0.02152346260845661, + 0.020507333800196648, + 0.0235544852912426, + 0.014309568330645561, + 0.003920516464859247, + 0.0036045375745743513, + 0.007938600145280361, + 0.005929566919803619, + 0.009585360065102577, + 0.00434046471491456, + 0.005099301226437092, + 0.005379931069910526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.05807754024863243 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059129998087883 + ], + "opposite_logprob_tokens": [ + -3.8955912590026855 + ], + "logprob_reward": -0.02059129998087883, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059129998087883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20581749081611633, + 0.07854226976633072, + 0.07908715307712555, + 0.09451831877231598, + 0.03578690439462662, + 0.032909948378801346, + 0.09297569841146469, + 0.1076669991016388, + 0.21974797546863556, + 0.052947260439395905 + ], + "credit_map": [ + 0.02642514742910862, + 0.026168011128902435, + 0.02324017323553562, + 0.02377261221408844, + 0.024199970066547394, + 0.02156870812177658, + 0.03811012580990791, + 0.022332748398184776, + 0.010388115420937538, + 0.010190150700509548, + 0.011032314039766788, + 0.009083548560738564, + 0.011008926667273045, + 0.00881857518106699, + 0.011083271354436874, + 0.00693737156689167, + 0.009060181677341461, + 0.007245258893817663, + 0.01301666721701622, + 0.020479267463088036, + 0.008088101632893085, + 0.00772989634424448, + 0.00660586915910244, + 0.006861917674541473, + 0.00849639531224966, + 0.011331973597407341, + 0.0125697311013937, + 0.012121599167585373, + 0.01440955139696598, + 0.007262334693223238, + 0.01892833039164543, + 0.009398398920893669, + 0.00332793896086514, + 0.0030143228359520435, + 0.004394688177853823, + 0.0038639588747173548, + 0.005415652878582478, + 0.004595371428877115, + 0.007105689961463213, + 0.0040692780166864395, + 0.0036892949137836695, + 0.004940398037433624, + 0.0048199002631008625, + 0.004540849477052689, + 0.00341026927344501, + 0.0034290377516299486, + 0.004682132974267006, + 0.00339806592091918, + 0.009200138971209526, + 0.007699097506701946, + 0.017111310735344887, + 0.010390754789113998, + 0.011647891253232956, + 0.00894386600703001, + 0.008747797459363937, + 0.01923483982682228, + 0.011606262065470219, + 0.006698226556181908, + 0.020931880921125412, + 0.009562230668962002, + 0.02015378512442112, + 0.014127999544143677, + 0.011690933257341385, + 0.012895680963993073, + 0.02711089514195919, + 0.025165878236293793, + 0.03523274511098862, + 0.03557910770177841, + 0.023812972009181976, + 0.024487169459462166, + 0.03144358843564987, + 0.016915619373321533, + 0.005796588491648436, + 0.004775742068886757, + 0.009755796752870083, + 0.005497005768120289, + 0.008967832662165165, + 0.006642424035817385, + 0.007431940641254187, + 0.004079928621649742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.02059129998087883 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864785075187683 + ], + "opposite_logprob_tokens": [ + -3.048647880554199 + ], + "logprob_reward": -0.04864785075187683, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864785075187683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09604968875646591, + 0.06305105239152908, + 0.057575471699237823, + 0.06615007668733597, + 0.05012647807598114, + 0.05738630145788193, + 0.1392187774181366, + 0.13915877044200897, + 0.20967508852481842, + 0.12160836905241013 + ], + "credit_map": [ + 0.01626894436776638, + 0.011124571785330772, + 0.011187558993697166, + 0.01039561815559864, + 0.011099609546363354, + 0.011673841625452042, + 0.01516913902014494, + 0.009130394086241722, + 0.007424050010740757, + 0.007134910207241774, + 0.008450203575193882, + 0.007368527352809906, + 0.006713880226016045, + 0.009783574379980564, + 0.010499296709895134, + 0.005676599219441414, + 0.003763732500374317, + 0.00706062326207757, + 0.008810019120573997, + 0.0137550700455904, + 0.00662990054115653, + 0.007855938747525215, + 0.0042678057216107845, + 0.0054323747754096985, + 0.008003375492990017, + 0.009566900320351124, + 0.007576860953122377, + 0.0072958022356033325, + 0.006929312366992235, + 0.006527193821966648, + 0.011858819983899593, + 0.008391805924475193, + 0.007600418291985989, + 0.004211536608636379, + 0.006178373005241156, + 0.0051584006287157536, + 0.005811096169054508, + 0.00881941244006157, + 0.006788161583244801, + 0.005559075158089399, + 0.007675874978303909, + 0.00897037424147129, + 0.009829729795455933, + 0.007275199517607689, + 0.008867310360074043, + 0.004380675032734871, + 0.005052017513662577, + 0.005335111171007156, + 0.014566379599273205, + 0.006486567202955484, + 0.03215408697724342, + 0.010422955267131329, + 0.009960716590285301, + 0.015931902453303337, + 0.009437873028218746, + 0.040258269757032394, + 0.016413116827607155, + 0.006561424117535353, + 0.02923758514225483, + 0.012560892850160599, + 0.02677414007484913, + 0.01460783276706934, + 0.018850522115826607, + 0.014153236523270607, + 0.02519557997584343, + 0.017910104244947433, + 0.037199895828962326, + 0.02483164146542549, + 0.026469092816114426, + 0.025779714807868004, + 0.03197932243347168, + 0.020309723913669586, + 0.013584713451564312, + 0.0170707069337368, + 0.021493293344974518, + 0.0141445342451334, + 0.021350273862481117, + 0.00961795449256897, + 0.01514987088739872, + 0.009197008796036243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.04864785075187683 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270291656255722 + ], + "opposite_logprob_tokens": [ + -2.1270291805267334 + ], + "logprob_reward": -0.1270291656255722, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270291656255722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08654533326625824, + 0.06330019235610962, + 0.09187724441289902, + 0.0988672524690628, + 0.07061326503753662, + 0.07717366516590118, + 0.11437717080116272, + 0.0887051448225975, + 0.18795661628246307, + 0.12058410048484802 + ], + "credit_map": [ + 0.010156271979212761, + 0.012774032540619373, + 0.010463693179190159, + 0.0077521600760519505, + 0.01570167951285839, + 0.009233446791768074, + 0.012107184156775475, + 0.008356863632798195, + 0.011975247412919998, + 0.006850020028650761, + 0.008306137286126614, + 0.006422699894756079, + 0.006979580502957106, + 0.011241902597248554, + 0.006464987061917782, + 0.005059618968516588, + 0.006916973739862442, + 0.011073576286435127, + 0.014916989021003246, + 0.024466389790177345, + 0.009983670897781849, + 0.006781923118978739, + 0.007222530897706747, + 0.01051519438624382, + 0.009422553703188896, + 0.014501206576824188, + 0.012081930413842201, + 0.012131030671298504, + 0.01229813601821661, + 0.008139343000948429, + 0.019995877519249916, + 0.01029717456549406, + 0.00933421216905117, + 0.005684655625373125, + 0.00846518948674202, + 0.005273376125842333, + 0.009597373194992542, + 0.011920436285436153, + 0.01170125138014555, + 0.008636771701276302, + 0.008187110535800457, + 0.013565801084041595, + 0.011447505094110966, + 0.010219667106866837, + 0.013972335495054722, + 0.005942987743765116, + 0.005975568201392889, + 0.00786268524825573, + 0.009696664288640022, + 0.005938393529504538, + 0.028070049360394478, + 0.010011487640440464, + 0.00904788263142109, + 0.00913945585489273, + 0.0085108932107687, + 0.033962346613407135, + 0.00681106885895133, + 0.004563918802887201, + 0.009973471984267235, + 0.006417976226657629, + 0.01221365574747324, + 0.013402030803263187, + 0.01841062679886818, + 0.016912398859858513, + 0.021345151588320732, + 0.017895393073558807, + 0.0326068140566349, + 0.024371441453695297, + 0.025278639048337936, + 0.01993635483086109, + 0.030215779319405556, + 0.0163070447742939, + 0.011852147988975048, + 0.01380074117332697, + 0.02565363049507141, + 0.01356191374361515, + 0.02702036313712597, + 0.008043196983635426, + 0.012903568334877491, + 0.007748528849333525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.1270291656255722 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327747642993927 + ], + "opposite_logprob_tokens": [ + -3.773277521133423 + ], + "logprob_reward": -0.02327747642993927, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327747642993927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03261376544833183, + 0.01292517501860857, + 0.03487662971019745, + 0.019492952153086662, + 0.031374186277389526, + 0.09842143207788467, + 0.7319177389144897, + 0.013499406166374683, + 0.013488409109413624, + 0.01139027252793312 + ], + "credit_map": [ + 0.003728674491867423, + 0.0039170584641397, + 0.004153714515268803, + 0.003181817475706339, + 0.0042831129394471645, + 0.003653649240732193, + 0.0061656879261136055, + 0.0035300508607178926, + 0.0011763368966057897, + 0.001232586451806128, + 0.0017385570099577308, + 0.0016443438362330198, + 0.0019017530139535666, + 0.002932713832706213, + 0.0011855213670060039, + 0.0011133634252473712, + 0.004635761957615614, + 0.004417021293193102, + 0.005072474014014006, + 0.0036663583014160395, + 0.005616877693682909, + 0.003920760937035084, + 0.0037528465036302805, + 0.00379453063942492, + 0.0015460352879017591, + 0.002263673348352313, + 0.0026837503537535667, + 0.002320068422704935, + 0.0034954166039824486, + 0.0015422820579260588, + 0.0036234217695891857, + 0.002018303144723177, + 0.00443545076996088, + 0.0030575129203498363, + 0.0047994088381528854, + 0.002396167954429984, + 0.004422713536769152, + 0.00474929204210639, + 0.004016987979412079, + 0.0034966529347002506, + 0.00899631530046463, + 0.011408486403524876, + 0.011500845663249493, + 0.00872083567082882, + 0.020578235387802124, + 0.01114015094935894, + 0.014374549500644207, + 0.01170201413333416, + 0.08349210768938065, + 0.036102090030908585, + 0.14895939826965332, + 0.044962137937545776, + 0.057722922414541245, + 0.12528230249881744, + 0.04058556258678436, + 0.19481123983860016, + 0.002397736068814993, + 0.000888968410436064, + 0.002126909326761961, + 0.0011210881639271975, + 0.001927278353832662, + 0.0012899191351607442, + 0.0020265160128474236, + 0.0017209912184625864, + 0.0016750231152400374, + 0.0010436356533318758, + 0.0024912708904594183, + 0.0016270000487565994, + 0.001930302125401795, + 0.0012639043852686882, + 0.0019366753986105323, + 0.0015205963281914592, + 0.001375940046273172, + 0.000853072851896286, + 0.0020622850861400366, + 0.001140307867899537, + 0.0018728761933743954, + 0.0013360818848013878, + 0.0013640684774145484, + 0.0013856403529644012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.02327747642993927 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a medium shot shot, shot slightly from above, and executed in extreme slow motion, a chivalrous hero, clad in a resplendent royal blue silk robe, his long, jet-black hair tied high, a few strands flowing around his cheeks, stands in a dimly lit scripture library. His expression is resolute, his eyes focused. A stack of ancient scrolls, their edges curled with age and their pages yellowed with age, rests on a heavy mahogany desk before him. He thrusts out his leather-guarded right hand and strikes the bottom of the pile with a swift, precise force. As if ignited by an invisible force, the entire pile explodes upward and in all directions. Countless yellowed pages unfold, twirl, and spin in a frozen passage of time, ancient calligraphy glimmering in the dance. A dramatic beam of light, filtering through a high, carved wooden window, captures this moment with pinpoint precision, making each stirred particle of ancient dust shimmer like golden stardust within the beam, the fragile edges of the paper appearing brightly outlined against the backlight. The entire painting is a play of light and shadow, rich in detail. The deep wooden background contrasts sharply with the hero's brightly colored attire, capturing a classical martial arts aesthetic imbued with power and destiny in this frozen moment, creating a solemn and visually striking atmosphere.", + "seed": 5224, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_3/index_5672_seed_5224_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05480197407305241, + "vlm_logprob": -0.05480197407305241, + "vlm_yes_no_margin": 3.174999952316284, + "vlm_token_credit": -0.05480197407305241, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.056885406374931335 + ], + "opposite_logprob_tokens": [ + -2.9318854808807373 + ], + "logprob_reward": -0.056885406374931335, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.056885406374931335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16865231096744537, + 0.044837526977062225, + 0.15109439194202423, + 0.036260709166526794, + 0.08636247366666794, + 0.14420631527900696, + 0.06693622469902039, + 0.1416923701763153, + 0.1328592449426651, + 0.02709846943616867 + ], + "credit_map": [ + 0.024746378883719444, + 0.016324540600180626, + 0.016580162569880486, + 0.019147520884871483, + 0.02765715681016445, + 0.029936600476503372, + 0.017749393358826637, + 0.01651054620742798, + 0.004735150374472141, + 0.004826246295124292, + 0.0032573945354670286, + 0.005841847974807024, + 0.007461801171302795, + 0.004984055180102587, + 0.007524577900767326, + 0.006206451915204525, + 0.027007469907402992, + 0.010199177078902721, + 0.010562681593000889, + 0.01823965273797512, + 0.018090523779392242, + 0.030799902975559235, + 0.017923953011631966, + 0.018271030858159065, + 0.005402592476457357, + 0.0038788323290646076, + 0.004033020231872797, + 0.0037136045284569263, + 0.004826525691896677, + 0.005283263977617025, + 0.0038937970530241728, + 0.0052290745079517365, + 0.022745000198483467, + 0.008466263301670551, + 0.008452017791569233, + 0.004303184803575277, + 0.009262297302484512, + 0.019192680716514587, + 0.004697296302765608, + 0.009243734180927277, + 0.015988949686288834, + 0.016246479004621506, + 0.009504997171461582, + 0.012731588445603848, + 0.020473364740610123, + 0.02055172622203827, + 0.029887834563851357, + 0.01882137730717659, + 0.004995743744075298, + 0.007406867109239101, + 0.004959695972502232, + 0.009824172593653202, + 0.014034834690392017, + 0.015118812210857868, + 0.005410673096776009, + 0.005185428075492382, + 0.03022388555109501, + 0.011047936044633389, + 0.007207048125565052, + 0.017158644273877144, + 0.0186647716909647, + 0.015875661745667458, + 0.01112061645835638, + 0.03039381094276905, + 0.016875863075256348, + 0.01358266081660986, + 0.010456816293299198, + 0.02642136625945568, + 0.02482001483440399, + 0.01920883171260357, + 0.011708340607583523, + 0.009785356000065804, + 0.0028484787326306105, + 0.004286001902073622, + 0.00203039706684649, + 0.003999366424977779, + 0.006696030497550964, + 0.0026498872321099043, + 0.00287139555439353, + 0.0017169115599244833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.056885406374931335 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819819025695324 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.01819819025695324, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819819025695324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11811710149049759, + 0.07026638835668564, + 0.1451139748096466, + 0.051579851657152176, + 0.10641194880008698, + 0.15282826125621796, + 0.0696922019124031, + 0.1336890012025833, + 0.10182575136423111, + 0.05047551542520523 + ], + "credit_map": [ + 0.012710805982351303, + 0.011029351502656937, + 0.010803897865116596, + 0.01385777909308672, + 0.020760061219334602, + 0.015885768458247185, + 0.014739765785634518, + 0.018329693004488945, + 0.011255934834480286, + 0.006340527907013893, + 0.005160320084542036, + 0.007497751619666815, + 0.011524075642228127, + 0.008989162743091583, + 0.010190143249928951, + 0.009308479726314545, + 0.03385593742132187, + 0.009948166087269783, + 0.00772172911092639, + 0.011347531341016293, + 0.017227573320269585, + 0.04115532338619232, + 0.010173329152166843, + 0.013684393838047981, + 0.006473059765994549, + 0.005925509613007307, + 0.005550803150981665, + 0.006193187087774277, + 0.008872490376234055, + 0.006154049187898636, + 0.006312685552984476, + 0.006098071578890085, + 0.02761075273156166, + 0.008038816042244434, + 0.00889157596975565, + 0.005570652894675732, + 0.011160172522068024, + 0.024928735569119453, + 0.0070360261015594006, + 0.013175219297409058, + 0.015240512788295746, + 0.01629386842250824, + 0.0093240300193429, + 0.014452128671109676, + 0.021236181259155273, + 0.0245102196931839, + 0.02882743440568447, + 0.022943895310163498, + 0.005912587977945805, + 0.00587060209363699, + 0.004095219075679779, + 0.0064489468932151794, + 0.016567396000027657, + 0.01678425818681717, + 0.007311432622373104, + 0.006701766978949308, + 0.03147983178496361, + 0.011408118531107903, + 0.006653403863310814, + 0.015052045695483685, + 0.016078809276223183, + 0.014701002277433872, + 0.009438528679311275, + 0.02887727878987789, + 0.019578997045755386, + 0.007144683040678501, + 0.0076248664408922195, + 0.00783990602940321, + 0.029119649901986122, + 0.009238316677510738, + 0.009097261354327202, + 0.012182071805000305, + 0.004714446607977152, + 0.004956979304552078, + 0.0021479602437466383, + 0.00433300482109189, + 0.015820534899830818, + 0.012036029249429703, + 0.003624647855758667, + 0.002841916400939226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.01819819025695324 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795313253998756 + ], + "opposite_logprob_tokens": [ + -3.5297954082489014 + ], + "logprob_reward": -0.029795313253998756, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795313253998756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11245204508304596, + 0.07622397691011429, + 0.10697507858276367, + 0.07094017416238785, + 0.19631458818912506, + 0.09612283855676651, + 0.04661881923675537, + 0.06670333445072174, + 0.1683645248413086, + 0.05928460881114006 + ], + "credit_map": [ + 0.01660129427909851, + 0.01235966570675373, + 0.008052713237702847, + 0.011823983862996101, + 0.018217196688055992, + 0.022598639130592346, + 0.01402914710342884, + 0.008769407868385315, + 0.012359188869595528, + 0.0072326031513512135, + 0.005922945681959391, + 0.010225817561149597, + 0.015470757149159908, + 0.01124089490622282, + 0.006773096974939108, + 0.006998675875365734, + 0.022484950721263885, + 0.008299602195620537, + 0.006964624393731356, + 0.010591300204396248, + 0.013345491141080856, + 0.023628683760762215, + 0.008919891901314259, + 0.012740532867610455, + 0.009755038656294346, + 0.006388168781995773, + 0.0054234606213867664, + 0.007710320875048637, + 0.01146087795495987, + 0.012841779738664627, + 0.008760346099734306, + 0.008600185625255108, + 0.03557007014751434, + 0.02084299921989441, + 0.01573118008673191, + 0.015398929826915264, + 0.022389423102140427, + 0.04554218798875809, + 0.016406027600169182, + 0.024433759972453117, + 0.009776406921446323, + 0.011455070227384567, + 0.005616372916847467, + 0.008786183781921864, + 0.014654731377959251, + 0.01519586332142353, + 0.01713985949754715, + 0.01349834818392992, + 0.004699049051851034, + 0.00343270949088037, + 0.0026734210550785065, + 0.0042310431599617004, + 0.010558603331446648, + 0.011488345451653004, + 0.005355722736567259, + 0.0041799237951636314, + 0.009344463236629963, + 0.006632098462432623, + 0.004945673048496246, + 0.014914968982338905, + 0.009241627529263496, + 0.006038553547114134, + 0.006096152123063803, + 0.009489801712334156, + 0.029293686151504517, + 0.013172920793294907, + 0.013550921343266964, + 0.016184872016310692, + 0.045292165130376816, + 0.016882648691534996, + 0.0162793081253767, + 0.017707983031868935, + 0.006310565862804651, + 0.004875853192061186, + 0.003404512768611312, + 0.006741138640791178, + 0.016957268118858337, + 0.012422959320247173, + 0.005101339425891638, + 0.003470965661108494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.029795313253998756 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.142788365483284 + ], + "opposite_logprob_tokens": [ + -2.0177884101867676 + ], + "logprob_reward": -0.142788365483284, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.142788365483284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09934897720813751, + 0.05088185518980026, + 0.0865180715918541, + 0.05976073071360588, + 0.24866822361946106, + 0.056245725601911545, + 0.05087145417928696, + 0.08765293657779694, + 0.22248475253582, + 0.03756731376051903 + ], + "credit_map": [ + 0.013520658947527409, + 0.012313222512602806, + 0.00907900184392929, + 0.009635329246520996, + 0.014509587548673153, + 0.01586821675300598, + 0.01315928902477026, + 0.011263662949204445, + 0.006670290604233742, + 0.004742363002151251, + 0.00315101002342999, + 0.006342827342450619, + 0.013624843209981918, + 0.006536788772791624, + 0.004850307479500771, + 0.004963415674865246, + 0.01798211969435215, + 0.005757895763963461, + 0.004663215484470129, + 0.006402004510164261, + 0.014719696715474129, + 0.02128107286989689, + 0.008273933082818985, + 0.007438127417117357, + 0.010466523468494415, + 0.005600058473646641, + 0.005853390786796808, + 0.005407555960118771, + 0.007211863528937101, + 0.015632526949048042, + 0.0050373743288218975, + 0.004551427438855171, + 0.05444445461034775, + 0.01807653158903122, + 0.018893098458647728, + 0.015736812725663185, + 0.027676016092300415, + 0.06460143625736237, + 0.017319802194833755, + 0.03192005679011345, + 0.005458597559481859, + 0.007225570734590292, + 0.003518609795719385, + 0.006342889275401831, + 0.008560368791222572, + 0.009282518178224564, + 0.0073363264091312885, + 0.008520838804543018, + 0.005759326741099358, + 0.004247370176017284, + 0.0033667089883238077, + 0.006188824772834778, + 0.009997090324759483, + 0.010546650737524033, + 0.005837324075400829, + 0.004928152076900005, + 0.013180959038436413, + 0.008479186333715916, + 0.006598406936973333, + 0.019406454637646675, + 0.011459234170615673, + 0.008088826201856136, + 0.007150738965719938, + 0.01328912191092968, + 0.034568071365356445, + 0.01633358933031559, + 0.018781837075948715, + 0.019674114882946014, + 0.054502420127391815, + 0.022857293486595154, + 0.023823926225304604, + 0.03194347023963928, + 0.004206562880426645, + 0.003648444777354598, + 0.0025185258127748966, + 0.006290358956903219, + 0.007826173678040504, + 0.005235292948782444, + 0.005097803194075823, + 0.002744149649515748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.142788365483284 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026342594996094704 + ], + "opposite_logprob_tokens": [ + -3.6513426303863525 + ], + "logprob_reward": -0.026342594996094704, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026342594996094704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18533417582511902, + 0.32283854484558105, + 0.06925451755523682, + 0.06817629933357239, + 0.09125667810440063, + 0.05956891551613808, + 0.039182767271995544, + 0.03594077751040459, + 0.04896005988121033, + 0.07948731631040573 + ], + "credit_map": [ + 0.024542387574911118, + 0.019838443025946617, + 0.01867922581732273, + 0.02016371302306652, + 0.03427927568554878, + 0.024160517379641533, + 0.024324316531419754, + 0.01934630237519741, + 0.04397959262132645, + 0.023991618305444717, + 0.020389221608638763, + 0.03219267725944519, + 0.06320884078741074, + 0.06657161563634872, + 0.04546505585312843, + 0.02703990414738655, + 0.01023640763014555, + 0.0075332182459533215, + 0.005186337511986494, + 0.007083261385560036, + 0.011317994445562363, + 0.008987825363874435, + 0.008910071104764938, + 0.009999395348131657, + 0.009622647427022457, + 0.007683051750063896, + 0.006563269998878241, + 0.00629054382443428, + 0.009701351635158062, + 0.015229769051074982, + 0.006720807868987322, + 0.006364858243614435, + 0.010472622700035572, + 0.014530993066728115, + 0.009083284065127373, + 0.008038799278438091, + 0.010480482131242752, + 0.01668323017656803, + 0.00943734496831894, + 0.012529924511909485, + 0.006203053053468466, + 0.007928282022476196, + 0.006445672828704119, + 0.007070884574204683, + 0.009341374039649963, + 0.006144554819911718, + 0.007788544520735741, + 0.008646546863019466, + 0.0051586320623755455, + 0.004008445423096418, + 0.0031741568818688393, + 0.0049741920083761215, + 0.008147266693413258, + 0.006328370422124863, + 0.003477000631392002, + 0.003914698958396912, + 0.004358620382845402, + 0.0033244481310248375, + 0.0024149157106876373, + 0.004140147008001804, + 0.007718792650848627, + 0.004476866219192743, + 0.003934492357075214, + 0.005572494585067034, + 0.010526854544878006, + 0.004568894859403372, + 0.003230110974982381, + 0.00654372526332736, + 0.008001982234418392, + 0.007713154423981905, + 0.003882718738168478, + 0.004492621403187513, + 0.007635836023837328, + 0.006130363792181015, + 0.00327268592081964, + 0.004979015793651342, + 0.026290427893400192, + 0.021598195657134056, + 0.0050349426455795765, + 0.004545849282294512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.026342594996094704 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a medium shot shot, shot slightly from above, and executed in extreme slow motion, a chivalrous hero, clad in a resplendent royal blue silk robe, his long, jet-black hair tied high, a few strands flowing around his cheeks, stands in a dimly lit scripture library. His expression is resolute, his eyes focused. A stack of ancient scrolls, their edges curled with age and their pages yellowed with age, rests on a heavy mahogany desk before him. He thrusts out his leather-guarded right hand and strikes the bottom of the pile with a swift, precise force. As if ignited by an invisible force, the entire pile explodes upward and in all directions. Countless yellowed pages unfold, twirl, and spin in a frozen passage of time, ancient calligraphy glimmering in the dance. A dramatic beam of light, filtering through a high, carved wooden window, captures this moment with pinpoint precision, making each stirred particle of ancient dust shimmer like golden stardust within the beam, the fragile edges of the paper appearing brightly outlined against the backlight. The entire painting is a play of light and shadow, rich in detail. The deep wooden background contrasts sharply with the hero's brightly colored attire, capturing a classical martial arts aesthetic imbued with power and destiny in this frozen moment, creating a solemn and visually striking atmosphere.", + "seed": 5225, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_3/index_5672_seed_5225_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09754881076514721, + "vlm_logprob": -0.09754881076514721, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.09754881076514721, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033014263957738876 + ], + "opposite_logprob_tokens": [ + -3.5330142974853516 + ], + "logprob_reward": -0.033014263957738876, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.033014263957738876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16738779842853546, + 0.11711636930704117, + 0.2109072208404541, + 0.05563248321413994, + 0.04485626518726349, + 0.2210438847541809, + 0.028507892042398453, + 0.06590107083320618, + 0.06518715620040894, + 0.02345985360443592 + ], + "credit_map": [ + 0.027742676436901093, + 0.020697861909866333, + 0.025373999029397964, + 0.02699722722172737, + 0.01804533042013645, + 0.014518260024487972, + 0.015247700735926628, + 0.018764769658446312, + 0.010419017635285854, + 0.007515162229537964, + 0.021753504872322083, + 0.01915128342807293, + 0.009100416675209999, + 0.010014001280069351, + 0.023753467947244644, + 0.015409521758556366, + 0.025467460975050926, + 0.01574038714170456, + 0.03106350265443325, + 0.023651616647839546, + 0.02151601016521454, + 0.020206861197948456, + 0.021414868533611298, + 0.051846522837877274, + 0.0068243821151554585, + 0.0072424523532390594, + 0.009952477179467678, + 0.008591352961957455, + 0.007335251662880182, + 0.003871148219332099, + 0.006490104831755161, + 0.005325316451489925, + 0.010744906030595303, + 0.004923918750137091, + 0.004172549117356539, + 0.002971569774672389, + 0.0033642391208559275, + 0.002723804209381342, + 0.0039929295890033245, + 0.011962354183197021, + 0.0370781347155571, + 0.024585969746112823, + 0.024745065718889236, + 0.018506985157728195, + 0.022139469161629677, + 0.0291445255279541, + 0.030356546863913536, + 0.034487199038267136, + 0.0021272979211062193, + 0.002895921701565385, + 0.003799992147833109, + 0.0036047210451215506, + 0.005865247920155525, + 0.0024618995375931263, + 0.0040462929755449295, + 0.003706521587446332, + 0.007061975542455912, + 0.007030174136161804, + 0.009426908567547798, + 0.010243742726743221, + 0.009716687723994255, + 0.00659453310072422, + 0.00974921602755785, + 0.006077835336327553, + 0.006676685996353626, + 0.004634495824575424, + 0.01322750560939312, + 0.006859804969280958, + 0.005071522202342749, + 0.007654749322682619, + 0.006341948639601469, + 0.014720447361469269, + 0.0023245441261678934, + 0.003443293273448944, + 0.0035907516721636057, + 0.004208300728350878, + 0.0025527230463922024, + 0.002237166976556182, + 0.00329170492477715, + 0.0018113709520548582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.033014263957738876 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589547231793404 + ], + "opposite_logprob_tokens": [ + -3.895589590072632 + ], + "logprob_reward": -0.020589547231793404, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589547231793404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20724794268608093, + 0.13037289679050446, + 0.17347195744514465, + 0.028311874717473984, + 0.05458139255642891, + 0.16629885137081146, + 0.0252988263964653, + 0.09377158433198929, + 0.085732601583004, + 0.034912075847387314 + ], + "credit_map": [ + 0.03915766254067421, + 0.026436230167746544, + 0.031504444777965546, + 0.02888786979019642, + 0.024455051869153976, + 0.017574287950992584, + 0.016479242593050003, + 0.022753151133656502, + 0.01279302779585123, + 0.00849274080246687, + 0.02083594724535942, + 0.020202895626425743, + 0.011356083676218987, + 0.010567632503807545, + 0.028779665008187294, + 0.017344892024993896, + 0.018272139132022858, + 0.014432365074753761, + 0.02854875475168228, + 0.02282867580652237, + 0.017945531755685806, + 0.015235337428748608, + 0.01694224588572979, + 0.039266906678676605, + 0.0036753257736563683, + 0.0033125707414001226, + 0.004121209494769573, + 0.0035020518116652966, + 0.0036358018405735493, + 0.003379784058779478, + 0.0026328167878091335, + 0.004052316769957542, + 0.005479433108121157, + 0.006251489743590355, + 0.007120778784155846, + 0.009964780882000923, + 0.004335954785346985, + 0.002936507808044553, + 0.009847162291407585, + 0.008645284920930862, + 0.028451411053538322, + 0.018371369689702988, + 0.018506715074181557, + 0.014585170894861221, + 0.015823958441615105, + 0.021662035956978798, + 0.02241547964513302, + 0.026482706889510155, + 0.001829645480029285, + 0.002575075253844261, + 0.0031958017498254776, + 0.003322238801047206, + 0.005080928560346365, + 0.0020731682889163494, + 0.004384260158985853, + 0.0028377065900713205, + 0.009694202803075314, + 0.00959707796573639, + 0.01378353126347065, + 0.01344415731728077, + 0.014415677636861801, + 0.009744050912559032, + 0.014308351092040539, + 0.008784529753029346, + 0.009325575083494186, + 0.006708734203130007, + 0.01598062366247177, + 0.009398016147315502, + 0.0075981575064361095, + 0.010165601037442684, + 0.008605719543993473, + 0.017950179055333138, + 0.0028508875984698534, + 0.004676799289882183, + 0.004290275741368532, + 0.004752970766276121, + 0.0043195923790335655, + 0.006847739219665527, + 0.003970901947468519, + 0.003202911000698805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.020589547231793404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495903640985489 + ], + "opposite_logprob_tokens": [ + -2.9299590587615967 + ], + "logprob_reward": -0.05495903640985489, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495903640985489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09030809998512268, + 0.06953521817922592, + 0.10960696637630463, + 0.06916745752096176, + 0.032089963555336, + 0.17397527396678925, + 0.08285027742385864, + 0.2656933069229126, + 0.08453913033008575, + 0.02223437838256359 + ], + "credit_map": [ + 0.02084592916071415, + 0.009760967455804348, + 0.015149451792240143, + 0.00970318354666233, + 0.008426071144640446, + 0.008777744136750698, + 0.005650434177368879, + 0.011994311586022377, + 0.00964223314076662, + 0.004746342543512583, + 0.010937596671283245, + 0.010543414391577244, + 0.006575118284672499, + 0.007467711344361305, + 0.010342271067202091, + 0.009280525147914886, + 0.011317012831568718, + 0.010890807025134563, + 0.01681327633559704, + 0.01664736680686474, + 0.010256313718855381, + 0.010366274043917656, + 0.011205700226128101, + 0.02211020514369011, + 0.01212590653449297, + 0.00642781239002943, + 0.007558340206742287, + 0.008810442872345448, + 0.007730836980044842, + 0.00705150980502367, + 0.006564520299434662, + 0.012898070737719536, + 0.0034056943841278553, + 0.0030480483546853065, + 0.00441244849935174, + 0.00514873256906867, + 0.0031681342516094446, + 0.002405047183856368, + 0.006325053051114082, + 0.0041767992079257965, + 0.030389538034796715, + 0.018937911838293076, + 0.018362844362854958, + 0.015232150442898273, + 0.017097823321819305, + 0.0241240207105875, + 0.022809172049164772, + 0.027021780610084534, + 0.005407601594924927, + 0.010470629669725895, + 0.012438412755727768, + 0.012165982276201248, + 0.01940044015645981, + 0.005432543810456991, + 0.008945532143115997, + 0.0085891243070364, + 0.024200720712542534, + 0.030122552067041397, + 0.03569982945919037, + 0.043428052216768265, + 0.04125745967030525, + 0.023743441328406334, + 0.04278271272778511, + 0.024458501487970352, + 0.005545372609049082, + 0.007706714794039726, + 0.015340842306613922, + 0.015603145584464073, + 0.012070546858012676, + 0.004983365070074797, + 0.00742173520848155, + 0.01586739718914032, + 0.0031784602906554937, + 0.0018347454024478793, + 0.0027155973948538303, + 0.001865101046860218, + 0.002317737555131316, + 0.004922852851450443, + 0.0026452832389622927, + 0.00275459885597229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.05495903640985489 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252124547958374 + ], + "opposite_logprob_tokens": [ + -1.502124547958374 + ], + "logprob_reward": -0.252124547958374, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252124547958374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15312764048576355, + 0.1164071261882782, + 0.1384889781475067, + 0.047425076365470886, + 0.04074655845761299, + 0.1235620379447937, + 0.06230781599879265, + 0.18542253971099854, + 0.09677489101886749, + 0.03573731705546379 + ], + "credit_map": [ + 0.0346401073038578, + 0.015736905857920647, + 0.026107681915163994, + 0.016618480905890465, + 0.014367516152560711, + 0.01608952321112156, + 0.010795822367072105, + 0.0187715794891119, + 0.015467747114598751, + 0.007307648193091154, + 0.017844846472144127, + 0.018294984474778175, + 0.010141151025891304, + 0.009220816195011139, + 0.022851604968309402, + 0.015278328210115433, + 0.01388877909630537, + 0.01197210606187582, + 0.022578252479434013, + 0.01831260323524475, + 0.013899925164878368, + 0.012149281799793243, + 0.013394002802670002, + 0.032294027507305145, + 0.005635185167193413, + 0.00535901915282011, + 0.006356443744152784, + 0.005058737937361002, + 0.007372142281383276, + 0.0064674788154661655, + 0.0048612384125590324, + 0.006314827129244804, + 0.004296450410038233, + 0.003254874376580119, + 0.004795392043888569, + 0.0047663738951087, + 0.0035470034927129745, + 0.0038688075728714466, + 0.007518022786825895, + 0.008699632249772549, + 0.021382542327046394, + 0.01269733626395464, + 0.013334625400602818, + 0.012823828496038914, + 0.011318993754684925, + 0.01635115221142769, + 0.015993736684322357, + 0.01965980976819992, + 0.003811561269685626, + 0.007220979779958725, + 0.008336484432220459, + 0.009215980768203735, + 0.015437405556440353, + 0.003820640267804265, + 0.007493290584534407, + 0.006971466820687056, + 0.017943140119314194, + 0.02026800438761711, + 0.02492193877696991, + 0.03129315376281738, + 0.02838664874434471, + 0.017405793070793152, + 0.028742440044879913, + 0.01646142080426216, + 0.008001932874321938, + 0.008675744757056236, + 0.016107656061649323, + 0.016154984012246132, + 0.014852801337838173, + 0.006836278829723597, + 0.009020721539855003, + 0.017124760895967484, + 0.005894767586141825, + 0.0033218502067029476, + 0.004285365343093872, + 0.0021762950345873833, + 0.004671095870435238, + 0.007359127048403025, + 0.003993842750787735, + 0.00403496902436018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.252124547958374 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the hero wearing a royal blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hero's hair long, jet-black, and tied high with strands flowing around his cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hero thrust his leather-guarded right hand to strike a stack of ancient yellowed scrolls on a mahogany desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scrolls explode upward and outward with pages unfolding, twirling, and spinning in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dramatic beam of light filtering through a high carved wooden window that illuminates dust particles and outlines the paper edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705665826797485 + ], + "opposite_logprob_tokens": [ + -2.12705659866333 + ], + "logprob_reward": -0.12705665826797485, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705665826797485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1273433417081833, + 0.1733873039484024, + 0.09848302602767944, + 0.054759614169597626, + 0.033598802983760834, + 0.09718917310237885, + 0.03177884593605995, + 0.06383740156888962, + 0.28094714879989624, + 0.03867535665631294 + ], + "credit_map": [ + 0.025014517828822136, + 0.011168421246111393, + 0.020197559148073196, + 0.013673416338860989, + 0.010736706666648388, + 0.013548973016440868, + 0.01987406238913536, + 0.01312969159334898, + 0.016359098255634308, + 0.011820330284535885, + 0.03723134472966194, + 0.025094185024499893, + 0.013347900472581387, + 0.015068724751472473, + 0.03261449933052063, + 0.02185121923685074, + 0.008396890014410019, + 0.008882809430360794, + 0.015234710648655891, + 0.014289218932390213, + 0.009687178768217564, + 0.008876994252204895, + 0.010753577575087547, + 0.022361649200320244, + 0.006533775478601456, + 0.005261976737529039, + 0.006513563450425863, + 0.00459251319989562, + 0.0055686915293335915, + 0.005401027854532003, + 0.004514660220593214, + 0.01637340523302555, + 0.00652006221935153, + 0.003346092300489545, + 0.0033555743284523487, + 0.003035574685782194, + 0.002259397879242897, + 0.002307572402060032, + 0.00295059266500175, + 0.009823936969041824, + 0.01617513969540596, + 0.01075781136751175, + 0.01144378911703825, + 0.008742534555494785, + 0.008972582407295704, + 0.012532173655927181, + 0.013083509169518948, + 0.01548162940889597, + 0.0024090036749839783, + 0.003573755268007517, + 0.004675392061471939, + 0.0036869309842586517, + 0.005028121173381805, + 0.0036809854209423065, + 0.004643342457711697, + 0.004081317223608494, + 0.006532672327011824, + 0.007822382263839245, + 0.0098110968247056, + 0.007934999652206898, + 0.008931806311011314, + 0.007223879918456078, + 0.00989856943488121, + 0.0056820036843419075, + 0.04957256093621254, + 0.02415165863931179, + 0.036701373755931854, + 0.018678603693842888, + 0.02471213787794113, + 0.05449364334344864, + 0.034911349415779114, + 0.037725843489170074, + 0.004660114645957947, + 0.00395867507904768, + 0.004556469619274139, + 0.0028438849840313196, + 0.006350986659526825, + 0.00813752319663763, + 0.0038590002804994583, + 0.0043087005615234375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.12705665826797485 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_4/index_tensor([297, 297])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_4/index_tensor([297, 297])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5db2a983cd78c047e28574f719c891589b50a0d7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_4/index_tensor([297, 297])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, long tracking shot follows a legion of heavily armored cavalry as they charge forward with unstoppable force. Each knight rests securely on his horse, clad in battle-worn steel plate that reflects the cold sunlight, the crimson plumes of his helmets whipping violently in the wind. Beneath their steeds, muscular and dark-haired, their nostrils blazing with white air, their hooves pounding the cracked earth with a devastating rhythm. Their environment is a vast, desolate plain. Beneath a leaden, gloomy sky, sparse, withered grass dotted the parched earth. The jagged outlines of mountains loom on the distant horizon. As the cavalry surged past like a tide, their horses' hooves stirred up a thick, rolling wave of ochre dust and gravel, almost engulfing the cavalry at the rear, leaving a chaotic and long trail behind them. The entire picture, under the flat and diffuse light, was filled with an urgent, harsh and unstoppable epic atmosphere.", + "seed": 4590, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_4/index_297_seed_4590_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.44482677448540925, + "vlm_logprob": -0.44482677448540925, + "vlm_yes_no_margin": 1.675, + "vlm_token_credit": -0.44482677448540925, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5667046308517456 + ], + "opposite_logprob_tokens": [ + -0.9417046308517456 + ], + "logprob_reward": -0.5667046308517456, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5667046308517456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25448840856552124, + 0.06895559281110764, + 0.1137467697262764, + 0.042590368539094925, + 0.03700414299964905, + 0.09936250001192093, + 0.058697812259197235, + 0.023036567494273186, + 0.12361336499452591, + 0.17850445210933685 + ], + "credit_map": [ + 0.028217976912856102, + 0.025394754484295845, + 0.02779255248606205, + 0.05261282995343208, + 0.027183271944522858, + 0.0382976196706295, + 0.030535992234945297, + 0.024453453719615936, + 0.008236675523221493, + 0.007668786216527224, + 0.008428463712334633, + 0.009896201081573963, + 0.009102694690227509, + 0.00817906018346548, + 0.010902008973062038, + 0.006541706155985594, + 0.013987862505018711, + 0.012859328649938107, + 0.015374303795397282, + 0.012823693454265594, + 0.011981057934463024, + 0.018941868096590042, + 0.012048824690282345, + 0.015729840844869614, + 0.0035990311298519373, + 0.004150385037064552, + 0.00639405008405447, + 0.005749034695327282, + 0.006411291193217039, + 0.005862236488610506, + 0.0051704696379601955, + 0.005253873765468597, + 0.0035336969885975122, + 0.003444869304075837, + 0.005213684402406216, + 0.0036957471165806055, + 0.0033184716012328863, + 0.004593283403664827, + 0.005543801002204418, + 0.007660592906177044, + 0.010312904603779316, + 0.009725868701934814, + 0.016840258613228798, + 0.01360249798744917, + 0.010981577448546886, + 0.00936190877109766, + 0.014780374243855476, + 0.013757111504673958, + 0.006512075196951628, + 0.005664112512022257, + 0.0057641263119876385, + 0.011966835707426071, + 0.00536560732871294, + 0.008199986070394516, + 0.0063386345282197, + 0.00888644065707922, + 0.0022484133951365948, + 0.002736295573413372, + 0.0034052233677357435, + 0.003952096216380596, + 0.0017226733034476638, + 0.0025377043057233095, + 0.00330316717736423, + 0.0031309970654547215, + 0.010519483126699924, + 0.014747759327292442, + 0.01584656536579132, + 0.02503584325313568, + 0.009198798798024654, + 0.011952458880841732, + 0.02210526540875435, + 0.014207198284566402, + 0.010398177430033684, + 0.008801235817372799, + 0.055802855640649796, + 0.05737024545669556, + 0.008806333877146244, + 0.006302901078015566, + 0.01915826089680195, + 0.01186446100473404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.5667046308517456 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039562225341797 + ], + "opposite_logprob_tokens": [ + -1.910395622253418 + ], + "logprob_reward": -0.16039562225341797, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039562225341797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11339451372623444, + 0.0602804496884346, + 0.05536536127328873, + 0.07601402699947357, + 0.05256298556923866, + 0.09585147351026535, + 0.03704717755317688, + 0.02446177415549755, + 0.04265158623456955, + 0.4423706531524658 + ], + "credit_map": [ + 0.016681170091032982, + 0.012541159987449646, + 0.01470376830548048, + 0.022051652893424034, + 0.00810313317924738, + 0.01275814138352871, + 0.014879082329571247, + 0.011676403693854809, + 0.00668503250926733, + 0.005802267696708441, + 0.005312005989253521, + 0.006749641615897417, + 0.009079952724277973, + 0.00937645323574543, + 0.010474045760929585, + 0.006801052019000053, + 0.006864375900477171, + 0.005320895928889513, + 0.007350170519202948, + 0.006794733926653862, + 0.0077225095592439175, + 0.010441581718623638, + 0.006036771461367607, + 0.004834321793168783, + 0.008205350488424301, + 0.009195412509143353, + 0.013718955218791962, + 0.010545594617724419, + 0.008935786783695221, + 0.009710739366710186, + 0.006205969024449587, + 0.009496216662228107, + 0.005006107036024332, + 0.003841879777610302, + 0.009101474657654762, + 0.004222537390887737, + 0.003981361631304026, + 0.0076419999822974205, + 0.010188632644712925, + 0.008578991517424583, + 0.010490534827113152, + 0.010391883552074432, + 0.016625115647912025, + 0.014616432599723339, + 0.008255045861005783, + 0.00797086488455534, + 0.016356442123651505, + 0.011145154014229774, + 0.004261368419975042, + 0.003421423491090536, + 0.006010347045958042, + 0.004957280587404966, + 0.0045235357247292995, + 0.004564137198030949, + 0.005248570814728737, + 0.004060511011630297, + 0.005189291667193174, + 0.0025852189864963293, + 0.0024389284662902355, + 0.002959959674626589, + 0.0021581645123660564, + 0.0027041900902986526, + 0.003181044477969408, + 0.003244976978749037, + 0.005011893808841705, + 0.005282211117446423, + 0.006771152839064598, + 0.006202671676874161, + 0.0044112177565693855, + 0.00435988325625658, + 0.0063951327465474606, + 0.004217423498630524, + 0.02077319845557213, + 0.01684148609638214, + 0.16946521401405334, + 0.17923463881015778, + 0.008591677993535995, + 0.005680170841515064, + 0.024363171309232712, + 0.01742112822830677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.16039562225341797 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3134548664093018 + ], + "opposite_logprob_tokens": [ + -0.31345483660697937 + ], + "logprob_reward": -1.3134548664093018, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3134548664093018, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21909759938716888, + 0.07302771508693695, + 0.08803798258304596, + 0.12024055421352386, + 0.06090078875422478, + 0.1543193906545639, + 0.03385613486170769, + 0.03183220699429512, + 0.13512641191482544, + 0.08356127142906189 + ], + "credit_map": [ + 0.027681130915880203, + 0.02371179312467575, + 0.024367384612560272, + 0.04758303984999657, + 0.015250741504132748, + 0.029615886509418488, + 0.030443459749221802, + 0.020444152876734734, + 0.007281114347279072, + 0.007201637141406536, + 0.005244969855993986, + 0.008264168165624142, + 0.010374829173088074, + 0.009855316951870918, + 0.016412513330578804, + 0.008393161930143833, + 0.010855771601200104, + 0.01056059543043375, + 0.01152236107736826, + 0.010587400756776333, + 0.010332681238651276, + 0.01698809117078781, + 0.009967315010726452, + 0.007223757449537516, + 0.012539902701973915, + 0.013490899465978146, + 0.02032354287803173, + 0.012062721885740757, + 0.016199782490730286, + 0.01539002638310194, + 0.011786742135882378, + 0.01844693347811699, + 0.006331483367830515, + 0.005284237675368786, + 0.009728411212563515, + 0.005066066514700651, + 0.005903139710426331, + 0.007493975572288036, + 0.010583274997770786, + 0.010510201565921307, + 0.014232994988560677, + 0.017205897718667984, + 0.019975638017058372, + 0.016611913219094276, + 0.024208948016166687, + 0.015910018235445023, + 0.026719313114881516, + 0.01945466920733452, + 0.002798769623041153, + 0.003967017866671085, + 0.0038135142531245947, + 0.004419720731675625, + 0.004154982045292854, + 0.005841253325343132, + 0.005090816877782345, + 0.003770061768591404, + 0.007057950831949711, + 0.0029506892897188663, + 0.0027989589143544436, + 0.00464314641430974, + 0.002331141382455826, + 0.003517041215673089, + 0.005100688431411982, + 0.0034325947053730488, + 0.013269532471895218, + 0.017017655074596405, + 0.01787969283759594, + 0.02868759259581566, + 0.013241172768175602, + 0.012791087850928307, + 0.020191214978694916, + 0.012048475444316864, + 0.005045679397881031, + 0.0055939313024282455, + 0.025020090863108635, + 0.024170726537704468, + 0.005472865421324968, + 0.0033585168421268463, + 0.009741518646478653, + 0.0051579466089606285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -1.3134548664093018 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17996205389499664 + ], + "opposite_logprob_tokens": [ + -1.8049620389938354 + ], + "logprob_reward": -0.17996205389499664, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17996205389499664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11297381669282913, + 0.03897172585129738, + 0.05562150478363037, + 0.05969591438770294, + 0.0286130178719759, + 0.06484615057706833, + 0.024190939962863922, + 0.019534461200237274, + 0.04782429337501526, + 0.5477281212806702 + ], + "credit_map": [ + 0.009432666935026646, + 0.009997852146625519, + 0.013568539172410965, + 0.023899463936686516, + 0.007148996461182833, + 0.013165757060050964, + 0.018756920471787453, + 0.017003625631332397, + 0.0033363529946655035, + 0.003733874298632145, + 0.004020752850919962, + 0.005563441198319197, + 0.006280114408582449, + 0.0048555186949670315, + 0.006238075904548168, + 0.00494359852746129, + 0.006653065327554941, + 0.007274468895047903, + 0.00834178552031517, + 0.005811283830553293, + 0.005717350170016289, + 0.00833190232515335, + 0.005382446106523275, + 0.008109207265079021, + 0.006259183399379253, + 0.006823196075856686, + 0.00944907683879137, + 0.00618142681196332, + 0.008619415573775768, + 0.0074904304929077625, + 0.005898964125663042, + 0.008974222466349602, + 0.0027920384891331196, + 0.002080026315525174, + 0.004574057646095753, + 0.002603520406410098, + 0.0023115200456231833, + 0.0035400683991611004, + 0.0056982520036399364, + 0.005013535264879465, + 0.00765198701992631, + 0.007123015355318785, + 0.007177051156759262, + 0.007581617217510939, + 0.00800315197557211, + 0.006925487425178289, + 0.010089416056871414, + 0.010294424369931221, + 0.0031826994381844997, + 0.0021607608068734407, + 0.0026604528538882732, + 0.004056056495755911, + 0.003034105058759451, + 0.002581611042842269, + 0.0035517767537385225, + 0.002963477745652199, + 0.003237053519114852, + 0.0020678965374827385, + 0.001954789040610194, + 0.0029253128450363874, + 0.0020402518566697836, + 0.002253239043056965, + 0.0031219797674566507, + 0.00193393777590245, + 0.004092641174793243, + 0.0036008544266223907, + 0.005370013415813446, + 0.009450988844037056, + 0.00500975176692009, + 0.004007843788713217, + 0.00811411626636982, + 0.008178082294762135, + 0.02538745477795601, + 0.020817097276449203, + 0.21451134979724884, + 0.21993938088417053, + 0.009751616977155209, + 0.0074755423702299595, + 0.029770152643322945, + 0.020075535401701927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.17996205389499664 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003616699017584324 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.003616699017584324, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616699017584324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15683338046073914, + 0.05908922851085663, + 0.09637442231178284, + 0.0631658136844635, + 0.04405838996171951, + 0.24056720733642578, + 0.052424706518650055, + 0.04458094388246536, + 0.09246789664030075, + 0.1504380702972412 + ], + "credit_map": [ + 0.021722745150327682, + 0.018049033358693123, + 0.019425880163908005, + 0.027479300275444984, + 0.014323944225907326, + 0.017451999709010124, + 0.024776499718427658, + 0.013603967614471912, + 0.004678291734308004, + 0.00581502728164196, + 0.005643615499138832, + 0.007903719320893288, + 0.00967491790652275, + 0.006550157442688942, + 0.012376101687550545, + 0.006447394844144583, + 0.011560232378542423, + 0.012223553843796253, + 0.014910017140209675, + 0.009331654757261276, + 0.007381221745163202, + 0.016922011971473694, + 0.009249763563275337, + 0.014795968309044838, + 0.006410608999431133, + 0.007931572385132313, + 0.013284106738865376, + 0.006567805539816618, + 0.0072532049380242825, + 0.008255652152001858, + 0.006124642211943865, + 0.007338220719248056, + 0.00428766431286931, + 0.0028073375578969717, + 0.007182874716818333, + 0.0038463331293314695, + 0.003832706017419696, + 0.005344455596059561, + 0.00799515936523676, + 0.008761858567595482, + 0.03292088955640793, + 0.022422218695282936, + 0.022838467732071877, + 0.015945876017212868, + 0.043944861739873886, + 0.023851579055190086, + 0.03834617882966995, + 0.040297139436006546, + 0.005161528941243887, + 0.0064150067046284676, + 0.004320492967963219, + 0.01068142894655466, + 0.00501969875767827, + 0.007082702592015266, + 0.0075343335047364235, + 0.006209515035152435, + 0.009465192444622517, + 0.005390973296016455, + 0.004546734504401684, + 0.005821832921355963, + 0.003477713093161583, + 0.0051972209475934505, + 0.005853327922523022, + 0.004827952012419701, + 0.00799373909831047, + 0.012318002060055733, + 0.010253853164613247, + 0.020561620593070984, + 0.0074932752177119255, + 0.010379662737250328, + 0.015960069373250008, + 0.007507668808102608, + 0.007907582446932793, + 0.006831125821918249, + 0.05686989799141884, + 0.0560266375541687, + 0.004164882469922304, + 0.003385088639333844, + 0.008831906132400036, + 0.006420953664928675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.003616699017584324 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, long tracking shot follows a legion of heavily armored cavalry as they charge forward with unstoppable force. Each knight rests securely on his horse, clad in battle-worn steel plate that reflects the cold sunlight, the crimson plumes of his helmets whipping violently in the wind. Beneath their steeds, muscular and dark-haired, their nostrils blazing with white air, their hooves pounding the cracked earth with a devastating rhythm. Their environment is a vast, desolate plain. Beneath a leaden, gloomy sky, sparse, withered grass dotted the parched earth. The jagged outlines of mountains loom on the distant horizon. As the cavalry surged past like a tide, their horses' hooves stirred up a thick, rolling wave of ochre dust and gravel, almost engulfing the cavalry at the rear, leaving a chaotic and long trail behind them. The entire picture, under the flat and diffuse light, was filled with an urgent, harsh and unstoppable epic atmosphere.", + "seed": 4591, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_4/index_297_seed_4591_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.4116749646142125, + "vlm_logprob": -0.4116749646142125, + "vlm_yes_no_margin": 1.65, + "vlm_token_credit": -0.4116749646142125, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4937775433063507 + ], + "opposite_logprob_tokens": [ + -1.1187775135040283 + ], + "logprob_reward": -0.4937775433063507, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4937775433063507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18213558197021484, + 0.11151861399412155, + 0.09188581258058548, + 0.044575292617082596, + 0.035726841539144516, + 0.11983300745487213, + 0.12128977477550507, + 0.0678647980093956, + 0.09585943818092346, + 0.12931086122989655 + ], + "credit_map": [ + 0.023452069610357285, + 0.01951737143099308, + 0.019523747265338898, + 0.024883126839995384, + 0.031079469248652458, + 0.022816520184278488, + 0.02023947425186634, + 0.02062379941344261, + 0.017419183626770973, + 0.02890094183385372, + 0.008520065806806087, + 0.006577649153769016, + 0.012213868089020252, + 0.020607098937034607, + 0.00816384982317686, + 0.009115954861044884, + 0.012380771338939667, + 0.013516665436327457, + 0.008489414118230343, + 0.009683954529464245, + 0.009852733463048935, + 0.01604383811354637, + 0.010449089109897614, + 0.011469350196421146, + 0.004764683078974485, + 0.005627955310046673, + 0.0053563108667731285, + 0.0037207165732979774, + 0.007207808084785938, + 0.006714144255965948, + 0.006100891157984734, + 0.005082786548882723, + 0.003343699499964714, + 0.0049887485802173615, + 0.0032344793435186148, + 0.0031850438099354506, + 0.005620747804641724, + 0.007827985100448132, + 0.0031279046088457108, + 0.004398229997605085, + 0.018325453624129295, + 0.017633942887187004, + 0.01489343773573637, + 0.020419633015990257, + 0.010613273829221725, + 0.013571099378168583, + 0.013467775657773018, + 0.0109083978459239, + 0.01644114777445793, + 0.010881801135838032, + 0.012735257856547832, + 0.022122744470834732, + 0.014014534652233124, + 0.014737737365067005, + 0.01331192348152399, + 0.017044629901647568, + 0.008720016106963158, + 0.009185167960822582, + 0.010817646980285645, + 0.007826535031199455, + 0.007072044536471367, + 0.00807881634682417, + 0.008383016102015972, + 0.00778155168518424, + 0.01102646067738533, + 0.014684641733765602, + 0.008006133139133453, + 0.012996871024370193, + 0.014332904480397701, + 0.008924768306314945, + 0.013428275473415852, + 0.012459381483495235, + 0.011690800078213215, + 0.008622964844107628, + 0.009448465891182423, + 0.009906529448926449, + 0.01827886700630188, + 0.02864934131503105, + 0.025799894705414772, + 0.016913997009396553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.4937775433063507 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295939236879349 + ], + "opposite_logprob_tokens": [ + -2.237959384918213 + ], + "logprob_reward": -0.11295939236879349, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295939236879349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0947532057762146, + 0.10477914661169052, + 0.04700960963964462, + 0.06557927280664444, + 0.02138318121433258, + 0.17824675142765045, + 0.10585091263055801, + 0.10075642168521881, + 0.22334468364715576, + 0.05829688534140587 + ], + "credit_map": [ + 0.013085472397506237, + 0.009150221012532711, + 0.008461030200123787, + 0.01279973704367876, + 0.010789277963340282, + 0.011685867793858051, + 0.018321316689252853, + 0.010460281744599342, + 0.011244394816458225, + 0.016858451068401337, + 0.008245694451034069, + 0.012007299810647964, + 0.022910544648766518, + 0.013452451676130295, + 0.013220204971730709, + 0.006840114016085863, + 0.00899691041558981, + 0.004351769108325243, + 0.0033879103139042854, + 0.004787537269294262, + 0.0063386475667357445, + 0.009584762156009674, + 0.004596660379320383, + 0.004965414758771658, + 0.004303647670894861, + 0.006344717461615801, + 0.004331286530941725, + 0.0038498917128890753, + 0.01667475327849388, + 0.019861014559864998, + 0.00440639303997159, + 0.005807570181787014, + 0.002263338305056095, + 0.0033730866853147745, + 0.0021446484606713057, + 0.0018901051953434944, + 0.003053839085623622, + 0.0026976822409778833, + 0.003270244225859642, + 0.0026902358513325453, + 0.03041490726172924, + 0.020918311551213264, + 0.018403364345431328, + 0.016020821407437325, + 0.027282681316137314, + 0.038262493908405304, + 0.014118907041847706, + 0.012825274839997292, + 0.018275944516062737, + 0.007017889525741339, + 0.00880851037800312, + 0.019571315497159958, + 0.00872436910867691, + 0.010229712352156639, + 0.014699572697281837, + 0.018523601815104485, + 0.0185278058052063, + 0.015257065184414387, + 0.010982087813317776, + 0.010732442140579224, + 0.010085445828735828, + 0.014142475090920925, + 0.009838927537202835, + 0.011190174147486687, + 0.03109404630959034, + 0.030503755435347557, + 0.017439639195799828, + 0.024103596806526184, + 0.027161724865436554, + 0.02114206552505493, + 0.03550530970096588, + 0.03639454022049904, + 0.010216059163212776, + 0.007247067056596279, + 0.006183825433254242, + 0.0038136185612529516, + 0.0070955208502709866, + 0.006671363487839699, + 0.012101695872843266, + 0.0049677323549985886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.11295939236879349 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3134486675262451 + ], + "opposite_logprob_tokens": [ + -0.3134486675262451 + ], + "logprob_reward": -1.3134486675262451, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3134486675262451, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1435994952917099, + 0.05340229719877243, + 0.025590701028704643, + 0.09379252046346664, + 0.025147780776023865, + 0.0816565528512001, + 0.09125331789255142, + 0.04024064168334007, + 0.3606308102607727, + 0.08468583226203918 + ], + "credit_map": [ + 0.013451670296490192, + 0.013340362347662449, + 0.012041132897138596, + 0.02630866877734661, + 0.021067125722765923, + 0.021960705518722534, + 0.019497763365507126, + 0.015932079404592514, + 0.006089941598474979, + 0.0063658286817371845, + 0.0047475360333919525, + 0.00508140679448843, + 0.009198253974318504, + 0.009702164679765701, + 0.006726297549903393, + 0.0054908739402890205, + 0.00405990332365036, + 0.0031272401101887226, + 0.0022007219959050417, + 0.0023190693464130163, + 0.003796785371378064, + 0.004792810417711735, + 0.00231378385797143, + 0.0029803893994539976, + 0.01007440872490406, + 0.014431703835725784, + 0.008435102179646492, + 0.00959919672459364, + 0.015672923997044563, + 0.013407508842647076, + 0.009708411060273647, + 0.012463272549211979, + 0.002568785334005952, + 0.005056614521890879, + 0.0019232755294069648, + 0.0025810557417571545, + 0.003512406488880515, + 0.00400423863902688, + 0.0025678034871816635, + 0.0029336034785956144, + 0.01139201782643795, + 0.008386284112930298, + 0.00930608157068491, + 0.008471939712762833, + 0.012727852910757065, + 0.018147440627217293, + 0.007008977234363556, + 0.006215970031917095, + 0.013766017742455006, + 0.00816438440233469, + 0.010182238183915615, + 0.012334054335951805, + 0.007689737249165773, + 0.010614002123475075, + 0.0156698115170002, + 0.012833081185817719, + 0.00685776025056839, + 0.003857259638607502, + 0.004125638399273157, + 0.005526539869606495, + 0.003751983167603612, + 0.007237311452627182, + 0.005188797600567341, + 0.0036953540984541178, + 0.04451877251267433, + 0.04611516743898392, + 0.029757382348179817, + 0.04921833798289299, + 0.046228356659412384, + 0.032169315963983536, + 0.060343749821186066, + 0.052279774099588394, + 0.007452487014234066, + 0.006964323576539755, + 0.00608377531170845, + 0.006600419525057077, + 0.014376800507307053, + 0.009311280213296413, + 0.023708781227469444, + 0.01018797792494297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -1.3134486675262451 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12711483240127563 + ], + "opposite_logprob_tokens": [ + -2.127114772796631 + ], + "logprob_reward": -0.12711483240127563, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12711483240127563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1641741693019867, + 0.17155902087688446, + 0.07530679553747177, + 0.08541622757911682, + 0.05836187303066254, + 0.0882057398557663, + 0.07398828864097595, + 0.06957074999809265, + 0.12104266881942749, + 0.09237439930438995 + ], + "credit_map": [ + 0.013111172243952751, + 0.014514070935547352, + 0.014802721329033375, + 0.022543206810951233, + 0.029233895242214203, + 0.026609526947140694, + 0.026660259813070297, + 0.016699334606528282, + 0.025284456089138985, + 0.020655667409300804, + 0.01348823867738247, + 0.01701144501566887, + 0.03332414850592613, + 0.028403257951140404, + 0.017274299636483192, + 0.016117526218295097, + 0.007347870152443647, + 0.008970104157924652, + 0.005825545638799667, + 0.006277657113969326, + 0.017160603776574135, + 0.013265127316117287, + 0.009153461083769798, + 0.0073064337484538555, + 0.011648942716419697, + 0.011891732923686504, + 0.007867141626775265, + 0.008249379694461823, + 0.011405600234866142, + 0.009044266305863857, + 0.013705017045140266, + 0.011604158207774162, + 0.0052664815448224545, + 0.016224922612309456, + 0.0031514272559434175, + 0.005811595357954502, + 0.008566176518797874, + 0.008914856240153313, + 0.004032706841826439, + 0.0063937073573470116, + 0.01564493589103222, + 0.01430730801075697, + 0.008864560164511204, + 0.010185048915445805, + 0.009361198171973228, + 0.012671036645770073, + 0.009057107381522655, + 0.00811455026268959, + 0.012739820405840874, + 0.006897490005940199, + 0.007060100324451923, + 0.008370989933609962, + 0.007558105047792196, + 0.008298765867948532, + 0.011567600071430206, + 0.011495423503220081, + 0.01274248119443655, + 0.009718024171888828, + 0.008348515257239342, + 0.006742074154317379, + 0.0072297896258533, + 0.007647973019629717, + 0.006533049046993256, + 0.01060884352773428, + 0.010386412963271141, + 0.018904123455286026, + 0.012772754766047001, + 0.027357598766684532, + 0.012512070126831532, + 0.010720093734562397, + 0.01731582172214985, + 0.011073803529143333, + 0.011754604987800121, + 0.009260062128305435, + 0.00858637597411871, + 0.009053634479641914, + 0.012391837313771248, + 0.01395768765360117, + 0.012720183469355106, + 0.014650021679699421 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.12711483240127563 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074387468397617 + ], + "opposite_logprob_tokens": [ + -4.511074542999268 + ], + "logprob_reward": -0.011074387468397617, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074387468397617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21814991533756256, + 0.2143723964691162, + 0.10334786027669907, + 0.1765916794538498, + 0.0384882427752018, + 0.04597112908959389, + 0.05583258345723152, + 0.024916792288422585, + 0.038006797432899475, + 0.08432261645793915 + ], + "credit_map": [ + 0.018293052911758423, + 0.019863096997141838, + 0.019146375358104706, + 0.0405263788998127, + 0.03764669597148895, + 0.028296317905187607, + 0.025769490748643875, + 0.028608504682779312, + 0.03253047913312912, + 0.05058756470680237, + 0.01917588897049427, + 0.015137054957449436, + 0.02398734539747238, + 0.04050351679325104, + 0.015077868476510048, + 0.01737269014120102, + 0.017597416415810585, + 0.012356218881905079, + 0.006774114444851875, + 0.009467449970543385, + 0.010915767401456833, + 0.028690097853541374, + 0.00812805350869894, + 0.009418734349310398, + 0.01433591078966856, + 0.01499016024172306, + 0.014317439869046211, + 0.01554791908711195, + 0.036924850195646286, + 0.03172217309474945, + 0.01920928806066513, + 0.02954394556581974, + 0.003784329164773226, + 0.002802441595122218, + 0.0029968905728310347, + 0.002871492877602577, + 0.005285405553877354, + 0.009343058802187443, + 0.004287708550691605, + 0.00711691752076149, + 0.005348362028598785, + 0.005894671194255352, + 0.0033888481557369232, + 0.0034964082296937704, + 0.010172686539590359, + 0.010410278104245663, + 0.003327897982671857, + 0.0039319791831076145, + 0.012594534084200859, + 0.004385886248201132, + 0.00524141313508153, + 0.006741483695805073, + 0.004516793880611658, + 0.00596191082149744, + 0.00764012336730957, + 0.008750437758862972, + 0.0033998731523752213, + 0.002326608169823885, + 0.002661241916939616, + 0.0032835681922733784, + 0.0036583549808710814, + 0.0030787864234298468, + 0.002235341351479292, + 0.004273017402738333, + 0.0022752091754227877, + 0.00484316423535347, + 0.004971854854375124, + 0.010353841818869114, + 0.0029383921064436436, + 0.0035103545524179935, + 0.005626873578876257, + 0.0034871103707700968, + 0.010213324800133705, + 0.006674385163933039, + 0.006199017632752657, + 0.007352638524025679, + 0.009763511829078197, + 0.0174467321485281, + 0.013551710173487663, + 0.013121294789016247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.011074387468397617 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, long tracking shot follows a legion of heavily armored cavalry as they charge forward with unstoppable force. Each knight rests securely on his horse, clad in battle-worn steel plate that reflects the cold sunlight, the crimson plumes of his helmets whipping violently in the wind. Beneath their steeds, muscular and dark-haired, their nostrils blazing with white air, their hooves pounding the cracked earth with a devastating rhythm. Their environment is a vast, desolate plain. Beneath a leaden, gloomy sky, sparse, withered grass dotted the parched earth. The jagged outlines of mountains loom on the distant horizon. As the cavalry surged past like a tide, their horses' hooves stirred up a thick, rolling wave of ochre dust and gravel, almost engulfing the cavalry at the rear, leaving a chaotic and long trail behind them. The entire picture, under the flat and diffuse light, was filled with an urgent, harsh and unstoppable epic atmosphere.", + "seed": 4592, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_4/index_297_seed_4592_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.39374729739502073, + "vlm_logprob": -0.39374729739502073, + "vlm_yes_no_margin": 1.7999999880790711, + "vlm_token_credit": -0.39374729739502073, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.37379589676856995 + ], + "opposite_logprob_tokens": [ + -1.3737958669662476 + ], + "logprob_reward": -0.37379589676856995, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.37379589676856995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11517585068941116, + 0.06345045566558838, + 0.1600964367389679, + 0.024683354422450066, + 0.05723404884338379, + 0.3345279395580292, + 0.042186152189970016, + 0.06379800289869308, + 0.04910854622721672, + 0.08973927050828934 + ], + "credit_map": [ + 0.013057181611657143, + 0.00969049520790577, + 0.01186743937432766, + 0.013524295762181282, + 0.013727175071835518, + 0.018895667046308517, + 0.013822550885379314, + 0.02059103548526764, + 0.006616804748773575, + 0.004825092386454344, + 0.012268497608602047, + 0.007701096590608358, + 0.006698663346469402, + 0.007870959118008614, + 0.008183738216757774, + 0.009285599924623966, + 0.01125340536236763, + 0.03204141929745674, + 0.007995542138814926, + 0.006614034995436668, + 0.009706007316708565, + 0.017773864790797234, + 0.016471078619360924, + 0.05824108049273491, + 0.0022031599655747414, + 0.0032392218708992004, + 0.0021289214491844177, + 0.0031063947826623917, + 0.005597791168838739, + 0.0034772008657455444, + 0.0026274817064404488, + 0.0023031795863062143, + 0.005232582334429026, + 0.0076094260439276695, + 0.004371698014438152, + 0.010725545696914196, + 0.00853994581848383, + 0.00938531756401062, + 0.006176624447107315, + 0.005192901473492384, + 0.034427281469106674, + 0.0656323954463005, + 0.01926075853407383, + 0.02071395516395569, + 0.046795621514320374, + 0.020886346697807312, + 0.10140436887741089, + 0.025407185778021812, + 0.0043593705631792545, + 0.008337052538990974, + 0.003462305059656501, + 0.0056292093358933926, + 0.005407892167568207, + 0.005529273301362991, + 0.004519284702837467, + 0.004941761493682861, + 0.007977364584803581, + 0.010643508285284042, + 0.00560257863253355, + 0.006610306445509195, + 0.009415450505912304, + 0.008824334479868412, + 0.00693141994997859, + 0.007793036289513111, + 0.005858997814357281, + 0.007913336157798767, + 0.004795897752046585, + 0.0054727946408092976, + 0.011531248688697815, + 0.004806419834494591, + 0.004024846013635397, + 0.004705002065747976, + 0.02518530935049057, + 0.030795453116297722, + 0.0036909780465066433, + 0.006605461239814758, + 0.009066374972462654, + 0.0061637586914002895, + 0.005263760220259428, + 0.002968163462355733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.37379589676856995 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743073582649231 + ], + "opposite_logprob_tokens": [ + -0.9743073582649231 + ], + "logprob_reward": -0.4743073582649231, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743073582649231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042914267629384995, + 0.03404935449361801, + 0.08625462651252747, + 0.024481775239109993, + 0.020403200760483742, + 0.6358622908592224, + 0.030501538887619972, + 0.08298837393522263, + 0.019362570717930794, + 0.02318192645907402 + ], + "credit_map": [ + 0.004489503800868988, + 0.0030187813099473715, + 0.005409119185060263, + 0.005355645902454853, + 0.006172939669340849, + 0.007588092237710953, + 0.005390461999922991, + 0.005489726085215807, + 0.0031819804571568966, + 0.0027870757039636374, + 0.0056334976106882095, + 0.00341184064745903, + 0.0044425372034311295, + 0.006307482253760099, + 0.003569765482097864, + 0.004715179558843374, + 0.006327916402369738, + 0.019565364345908165, + 0.004171687178313732, + 0.003954347688704729, + 0.005083863623440266, + 0.005729249212890863, + 0.008884037844836712, + 0.032538171857595444, + 0.002010195516049862, + 0.0030601760372519493, + 0.0022585992701351643, + 0.003832676913589239, + 0.003107436466962099, + 0.0030617876909673214, + 0.004146208520978689, + 0.0030046964529901743, + 0.00203043594956398, + 0.00277814082801342, + 0.0022226644214242697, + 0.0035616508685052395, + 0.001928976969793439, + 0.0036832585465162992, + 0.0017337823519483209, + 0.0024642942007631063, + 0.05417627841234207, + 0.024522840976715088, + 0.0404503270983696, + 0.031994618475437164, + 0.13479672372341156, + 0.03265472874045372, + 0.2764686942100525, + 0.04079815372824669, + 0.00325592327862978, + 0.00782197155058384, + 0.0021201111376285553, + 0.002034939592704177, + 0.004885938484221697, + 0.0044655464589595795, + 0.002810729667544365, + 0.0031063829082995653, + 0.00893495511263609, + 0.005830471869558096, + 0.010487678460776806, + 0.010369163006544113, + 0.006294784136116505, + 0.009444613941013813, + 0.021595178171992302, + 0.010031539015471935, + 0.002691548550501466, + 0.0032048530410975218, + 0.002856186358258128, + 0.001994200749322772, + 0.0026676978450268507, + 0.001525614527054131, + 0.0022180296946316957, + 0.0022044426295906305, + 0.0048050410114228725, + 0.008166253566741943, + 0.0009720196248963475, + 0.00100472173653543, + 0.0025012448895722628, + 0.0016090369317680597, + 0.0015062259044498205, + 0.0026173843070864677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.4743073582649231 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539112091064453 + ], + "opposite_logprob_tokens": [ + -0.42891114950180054 + ], + "logprob_reward": -1.0539112091064453, + "yes_no_margin_reward": -0.6250000596046448, + "token_credit_reward": -1.0539112091064453, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09628625214099884, + 0.02914399281144142, + 0.10656417161226273, + 0.02539738267660141, + 0.019024301320314407, + 0.459403395652771, + 0.04100225865840912, + 0.11874985694885254, + 0.03564183786511421, + 0.06878659874200821 + ], + "credit_map": [ + 0.013534272089600563, + 0.00915292464196682, + 0.010137791745364666, + 0.008627002127468586, + 0.013407228514552116, + 0.015833158046007156, + 0.009775957092642784, + 0.015817897394299507, + 0.0027897271793335676, + 0.0024101180024445057, + 0.0035554671194404364, + 0.004766333382576704, + 0.004070280585438013, + 0.0042046126909554005, + 0.0031683549750596285, + 0.004179095383733511, + 0.007606723811477423, + 0.02363792061805725, + 0.005239370744675398, + 0.004811306484043598, + 0.006305625196546316, + 0.007404016796499491, + 0.010954609140753746, + 0.04060458019375801, + 0.0023262668401002884, + 0.0032377794850617647, + 0.0021671701688319445, + 0.0034976666793227196, + 0.003654682543128729, + 0.0038863900117576122, + 0.003720018779858947, + 0.002907406073063612, + 0.0016957870684564114, + 0.0027156416326761246, + 0.0021639345213770866, + 0.0027381086256355047, + 0.0029680829029530287, + 0.0030055271927267313, + 0.001770028960891068, + 0.001967186573892832, + 0.039993271231651306, + 0.03597388043999672, + 0.027791764587163925, + 0.024720434099435806, + 0.09031329303979874, + 0.023948529735207558, + 0.1881185621023178, + 0.028543621301651, + 0.0034086876548826694, + 0.0050207143649458885, + 0.003660584334284067, + 0.008711939677596092, + 0.0043876152485609055, + 0.006905947811901569, + 0.005634294357150793, + 0.0032724684569984674, + 0.008974169380962849, + 0.007312430534511805, + 0.011569718830287457, + 0.014363953843712807, + 0.008330295793712139, + 0.01632564328610897, + 0.03645291179418564, + 0.015420719981193542, + 0.003861081786453724, + 0.007285518571734428, + 0.003418802050873637, + 0.0038645013701170683, + 0.007503252476453781, + 0.00372852711006999, + 0.002990875393152237, + 0.0029892742168158293, + 0.006526090670377016, + 0.017989175394177437, + 0.006647614296525717, + 0.004059108905494213, + 0.008139763958752155, + 0.006797873880714178, + 0.006112260278314352, + 0.012514702044427395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -1.0539112091064453 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208660081028938 + ], + "opposite_logprob_tokens": [ + -2.812086582183838 + ], + "logprob_reward": -0.06208660081028938, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208660081028938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10486441850662231, + 0.038400717079639435, + 0.050201840698719025, + 0.01929638348519802, + 0.07056494802236557, + 0.4271605312824249, + 0.049129076302051544, + 0.05339398235082626, + 0.13014119863510132, + 0.05684695765376091 + ], + "credit_map": [ + 0.012931201606988907, + 0.008593975566327572, + 0.014256242662668228, + 0.009133311919867992, + 0.017213113605976105, + 0.01509841252118349, + 0.012667858973145485, + 0.014970287680625916, + 0.004542413633316755, + 0.002673664828762412, + 0.008002120070159435, + 0.0048548756167292595, + 0.0031679298263043165, + 0.005588332191109657, + 0.004159308038651943, + 0.005412065889686346, + 0.003768106224015355, + 0.012509103864431381, + 0.0025666740257292986, + 0.002886538626626134, + 0.0037216870114207268, + 0.003907650243490934, + 0.005290968343615532, + 0.01555110514163971, + 0.001618951209820807, + 0.002603117609396577, + 0.0020185615867376328, + 0.0018205313244834542, + 0.004054220858961344, + 0.0026443293318152428, + 0.0022061136551201344, + 0.0023305551148951054, + 0.0060004000551998615, + 0.010782463476061821, + 0.003668383229523897, + 0.005281203426420689, + 0.023645957931876183, + 0.01240613218396902, + 0.0035274007823318243, + 0.00525299645960331, + 0.04543059691786766, + 0.1017402857542038, + 0.02068810537457466, + 0.02547421306371689, + 0.06114020198583603, + 0.03508216515183449, + 0.11449603736400604, + 0.02310887724161148, + 0.005119071342051029, + 0.01153636910021305, + 0.003919097129255533, + 0.00620932225137949, + 0.005971408449113369, + 0.007614025380462408, + 0.005071148741990328, + 0.0036886304151266813, + 0.004799343645572662, + 0.006230005528777838, + 0.004427460953593254, + 0.004928681068122387, + 0.014009646140038967, + 0.008367094211280346, + 0.005363805685192347, + 0.005267939995974302, + 0.019485874101519585, + 0.01740287058055401, + 0.017380012199282646, + 0.012796152383089066, + 0.024820029735565186, + 0.01531711220741272, + 0.0112534798681736, + 0.011685661040246487, + 0.006786414422094822, + 0.01282014511525631, + 0.0061263274401426315, + 0.006280038971453905, + 0.005209938157349825, + 0.004807565826922655, + 0.007599002681672573, + 0.007217517122626305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.06208660081028938 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635422024875879 + ], + "opposite_logprob_tokens": [ + -5.379635334014893 + ], + "logprob_reward": -0.004635422024875879, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635422024875879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2586643695831299, + 0.026638207957148552, + 0.3512061536312103, + 0.10556572675704956, + 0.0499974861741066, + 0.021661654114723206, + 0.028580257669091225, + 0.06807032227516174, + 0.023118652403354645, + 0.06649716943502426 + ], + "credit_map": [ + 0.02718898467719555, + 0.023113539442420006, + 0.019101815298199654, + 0.020276060327887535, + 0.052390214055776596, + 0.04993331804871559, + 0.022130155935883522, + 0.044530294835567474, + 0.003255733521655202, + 0.0021161509212106466, + 0.004254793748259544, + 0.0025419665034860373, + 0.002235010266304016, + 0.003994348458945751, + 0.0028578832279890776, + 0.0053823222406208515, + 0.02478739060461521, + 0.07623638212680817, + 0.015920624136924744, + 0.013018828816711903, + 0.021129129454493523, + 0.03696255013346672, + 0.0343201644718647, + 0.12883110344409943, + 0.013955969363451004, + 0.013760654255747795, + 0.006138404365628958, + 0.004939763341099024, + 0.013765587471425533, + 0.030135203152894974, + 0.009917883202433586, + 0.012952263467013836, + 0.00406160531565547, + 0.004061147104948759, + 0.003101021284237504, + 0.0142855579033494, + 0.007964774034917355, + 0.008396816439926624, + 0.003063696436583996, + 0.005062870215624571, + 0.002744016470387578, + 0.002734976354986429, + 0.002457470167428255, + 0.0023693786934018135, + 0.003488440066576004, + 0.0026004775427281857, + 0.002930954098701477, + 0.0023359423503279686, + 0.002928712870925665, + 0.006718343589454889, + 0.0020110253244638443, + 0.0023482036776840687, + 0.004011868499219418, + 0.004050291609019041, + 0.0031393971294164658, + 0.0033724147360771894, + 0.0051314751617610455, + 0.00427400553599, + 0.006483954843133688, + 0.0079716881737113, + 0.004990102723240852, + 0.008791645057499409, + 0.02134258672595024, + 0.009084862656891346, + 0.003452321281656623, + 0.0029298763256520033, + 0.002207268727943301, + 0.0035931409802287817, + 0.0043642050586640835, + 0.0023779633920639753, + 0.0022495028097182512, + 0.0019443763885647058, + 0.013957654125988483, + 0.006899167783558369, + 0.0037252488546073437, + 0.006445871200412512, + 0.005157850217074156, + 0.007012438494712114, + 0.0064605893567204475, + 0.016838354989886284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.004635422024875879 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, long tracking shot follows a legion of heavily armored cavalry as they charge forward with unstoppable force. Each knight rests securely on his horse, clad in battle-worn steel plate that reflects the cold sunlight, the crimson plumes of his helmets whipping violently in the wind. Beneath their steeds, muscular and dark-haired, their nostrils blazing with white air, their hooves pounding the cracked earth with a devastating rhythm. Their environment is a vast, desolate plain. Beneath a leaden, gloomy sky, sparse, withered grass dotted the parched earth. The jagged outlines of mountains loom on the distant horizon. As the cavalry surged past like a tide, their horses' hooves stirred up a thick, rolling wave of ochre dust and gravel, almost engulfing the cavalry at the rear, leaving a chaotic and long trail behind them. The entire picture, under the flat and diffuse light, was filled with an urgent, harsh and unstoppable epic atmosphere.", + "seed": 4593, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_4/index_297_seed_4593_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.5052747623994946, + "vlm_logprob": -0.5052747623994946, + "vlm_yes_no_margin": 1.25, + "vlm_token_credit": -0.5052747623994946, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6879647970199585 + ], + "opposite_logprob_tokens": [ + -0.8129647970199585 + ], + "logprob_reward": -0.6879647970199585, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6879647970199585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12313870340585709, + 0.09171388298273087, + 0.04395876079797745, + 0.10062260180711746, + 0.1434335857629776, + 0.08890713006258011, + 0.03277207538485527, + 0.09958502650260925, + 0.17486542463302612, + 0.1010027676820755 + ], + "credit_map": [ + 0.009262450970709324, + 0.010450459085404873, + 0.018633520230650902, + 0.022080009803175926, + 0.013987896032631397, + 0.016200048848986626, + 0.018735283985733986, + 0.013789035379886627, + 0.009131638333201408, + 0.008165492676198483, + 0.007941094227135181, + 0.007037365809082985, + 0.028456497937440872, + 0.012464810162782669, + 0.009831001982092857, + 0.008685988374054432, + 0.004762980155646801, + 0.0066106608137488365, + 0.0036606923677027225, + 0.0038357723969966173, + 0.007779198698699474, + 0.006554706487804651, + 0.005970937665551901, + 0.0047838119789958, + 0.018774880096316338, + 0.00913214311003685, + 0.010891580022871494, + 0.01065668836236, + 0.015896258875727654, + 0.01638229563832283, + 0.011620873585343361, + 0.00726788816973567, + 0.012835881672799587, + 0.012693537399172783, + 0.008511641062796116, + 0.01297973282635212, + 0.04212619736790657, + 0.01566571556031704, + 0.017951926216483116, + 0.02066895179450512, + 0.010021368972957134, + 0.017925363034009933, + 0.006648324895650148, + 0.006862153299152851, + 0.02061249315738678, + 0.012184709310531616, + 0.008391357026994228, + 0.0062613580375909805, + 0.00404729600995779, + 0.003488416550680995, + 0.0028224140405654907, + 0.00380504853092134, + 0.004463442135602236, + 0.006248855497688055, + 0.00478814821690321, + 0.0031084518413990736, + 0.014502155594527721, + 0.01215796172618866, + 0.008875791914761066, + 0.008424918167293072, + 0.015476575121283531, + 0.00952798593789339, + 0.01852422021329403, + 0.012095424346625805, + 0.025698214769363403, + 0.02688535675406456, + 0.012418619357049465, + 0.013212107121944427, + 0.017430638894438744, + 0.05514301732182503, + 0.013532044366002083, + 0.010545413941144943, + 0.008702423423528671, + 0.005771988537162542, + 0.007931578904390335, + 0.005989835597574711, + 0.006600424647331238, + 0.028865721076726913, + 0.007332662586122751, + 0.029808131977915764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.6879647970199585 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135104179382324 + ], + "opposite_logprob_tokens": [ + -1.3135104179382324 + ], + "logprob_reward": -0.3135104179382324, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135104179382324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2701985239982605, + 0.0640547126531601, + 0.05283594876527786, + 0.06174054369330406, + 0.0759577676653862, + 0.05016050115227699, + 0.09407120943069458, + 0.09372279047966003, + 0.12262854725122452, + 0.11462944000959396 + ], + "credit_map": [ + 0.023700566962361336, + 0.041945140808820724, + 0.031714245676994324, + 0.04058685526251793, + 0.038064517080783844, + 0.032958608120679855, + 0.0321517176926136, + 0.029076870530843735, + 0.005602192133665085, + 0.005786160472780466, + 0.009561351500451565, + 0.008774221874773502, + 0.00834848452359438, + 0.008872583508491516, + 0.008675118908286095, + 0.008434596471488476, + 0.005386326927691698, + 0.008326890878379345, + 0.005478603765368462, + 0.00626815389841795, + 0.008122340776026249, + 0.006762666627764702, + 0.007941484451293945, + 0.004549482837319374, + 0.011941208504140377, + 0.005378614645451307, + 0.007296408526599407, + 0.005246741697192192, + 0.011385444551706314, + 0.009107910096645355, + 0.006991883274167776, + 0.004392334260046482, + 0.012835327535867691, + 0.007467726711183786, + 0.006887497380375862, + 0.007412957958877087, + 0.009914151392877102, + 0.0070832883939146996, + 0.010855110362172127, + 0.013501706533133984, + 0.006553450133651495, + 0.0062818643637001514, + 0.0054730321280658245, + 0.006534848362207413, + 0.007220245432108641, + 0.006037265993654728, + 0.006503427866846323, + 0.0055563682690262794, + 0.012172529473900795, + 0.014488635584712029, + 0.007404649164527655, + 0.008791379630565643, + 0.012073902413249016, + 0.024103177711367607, + 0.0074906377121806145, + 0.007546295411884785, + 0.013625211082398891, + 0.00841255858540535, + 0.009631301276385784, + 0.009385880082845688, + 0.01734969951212406, + 0.01525032427161932, + 0.01223390456289053, + 0.007833912968635559, + 0.015810726210474968, + 0.022167993709445, + 0.00950766820460558, + 0.016627714037895203, + 0.014444724656641483, + 0.02170713245868683, + 0.010990818962454796, + 0.011371773667633533, + 0.008826635777950287, + 0.007723154034465551, + 0.010255727916955948, + 0.007357491180300713, + 0.0070862919092178345, + 0.03248554468154907, + 0.009388381615281105, + 0.031506214290857315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.3135104179382324 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.313462495803833 + ], + "opposite_logprob_tokens": [ + -0.313462495803833 + ], + "logprob_reward": -1.313462495803833, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.313462495803833, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1468774676322937, + 0.08517711609601974, + 0.046631064265966415, + 0.09109105169773102, + 0.08691322058439255, + 0.0919157862663269, + 0.0931604653596878, + 0.08540351688861847, + 0.12288521975278854, + 0.14994509518146515 + ], + "credit_map": [ + 0.012350967153906822, + 0.011544688604772091, + 0.02071104757487774, + 0.02284792624413967, + 0.017937110736966133, + 0.02108011767268181, + 0.021633263677358627, + 0.01877233386039734, + 0.010243712924420834, + 0.007508315611630678, + 0.007005501538515091, + 0.00498089799657464, + 0.026277869939804077, + 0.012599606066942215, + 0.008314055390655994, + 0.008247164078056812, + 0.003770155366510153, + 0.005918496288359165, + 0.005737459287047386, + 0.004740288946777582, + 0.007784850895404816, + 0.00718372967094183, + 0.0075704604387283325, + 0.003925621509552002, + 0.011994521133601665, + 0.009944338351488113, + 0.012302820570766926, + 0.008492770604789257, + 0.01434989646077156, + 0.017316577956080437, + 0.00884267408400774, + 0.007847444154322147, + 0.010855446569621563, + 0.010675405152142048, + 0.007438165135681629, + 0.008832967840135098, + 0.008726261556148529, + 0.008467921055853367, + 0.012383506633341312, + 0.019533542916178703, + 0.011326185427606106, + 0.020194752141833305, + 0.006154287606477737, + 0.006794251035898924, + 0.021170910447835922, + 0.010824869386851788, + 0.009484087117016315, + 0.005966444034129381, + 0.010939780622720718, + 0.014462193474173546, + 0.011370033957064152, + 0.015551230870187283, + 0.007320215925574303, + 0.013331257738173008, + 0.011709457263350487, + 0.00847629550844431, + 0.00938889104872942, + 0.010227170772850513, + 0.010514657944440842, + 0.00969691202044487, + 0.007799256592988968, + 0.009335232898592949, + 0.016636213287711143, + 0.011805187910795212, + 0.014488983899354935, + 0.016226476058363914, + 0.009732401929795742, + 0.012070629745721817, + 0.018396075814962387, + 0.02956855483353138, + 0.013216682709753513, + 0.009185418486595154, + 0.01902299001812935, + 0.016481269150972366, + 0.013758852146565914, + 0.014016713947057724, + 0.015673678368330002, + 0.03985048085451126, + 0.0156166460365057, + 0.015524455346167088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -1.313462495803833 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016492635011673 + ], + "opposite_logprob_tokens": [ + -1.7016493082046509 + ], + "logprob_reward": -0.2016492635011673, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016492635011673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.262876033782959, + 0.08753146976232529, + 0.10855519026517868, + 0.0827750414609909, + 0.11350183933973312, + 0.0454372763633728, + 0.032865945249795914, + 0.07734484225511551, + 0.05668235197663307, + 0.13242998719215393 + ], + "credit_map": [ + 0.021712061017751694, + 0.03738034889101982, + 0.02567267045378685, + 0.027756638824939728, + 0.036587897688150406, + 0.03639940917491913, + 0.05505078658461571, + 0.022316215559840202, + 0.008075661025941372, + 0.008632874116301537, + 0.0057475632056593895, + 0.006424016784876585, + 0.029936466366052628, + 0.012570440769195557, + 0.008972539566457272, + 0.007171907462179661, + 0.00972954835742712, + 0.013637459836900234, + 0.017870670184493065, + 0.01199154369533062, + 0.009151851758360863, + 0.015918409451842308, + 0.016985062509775162, + 0.013270649127662182, + 0.009818424470722675, + 0.010128520429134369, + 0.008130486123263836, + 0.005946517456322908, + 0.016031663864850998, + 0.019389720633625984, + 0.006346384063363075, + 0.00698332954198122, + 0.0093789491802454, + 0.013094601221382618, + 0.008450929075479507, + 0.008889539167284966, + 0.026852501556277275, + 0.013323125429451466, + 0.009522111155092716, + 0.023990090936422348, + 0.004827714990824461, + 0.006115838885307312, + 0.004406050778925419, + 0.004849754273891449, + 0.007651203311979771, + 0.008025833405554295, + 0.006012790836393833, + 0.0035480877850204706, + 0.0033603694755584, + 0.003623020136728883, + 0.0032552708871662617, + 0.007159365341067314, + 0.0027172439731657505, + 0.005421333480626345, + 0.004065189976245165, + 0.003264151979237795, + 0.008907212875783443, + 0.007000611629337072, + 0.011824124492704868, + 0.007831700146198273, + 0.0047033559530973434, + 0.007340686395764351, + 0.016869308426976204, + 0.012867840006947517, + 0.006507265381515026, + 0.007638436742126942, + 0.004928880836814642, + 0.004570083227008581, + 0.007389952894300222, + 0.014058385044336319, + 0.006730753928422928, + 0.0048585934564471245, + 0.011706246063113213, + 0.00806246418505907, + 0.009507141076028347, + 0.009625439532101154, + 0.00923524983227253, + 0.039781004190444946, + 0.008982710540294647, + 0.0355297289788723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.2016492635011673 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low, looking up at the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a long tracking shot following the cavalry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knights wearing heavily armored steel plate that reflects sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the knights' helmets feature crimson plumes whipping in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a vast, desolate plain with a leaden, gloomy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786837734282017 + ], + "opposite_logprob_tokens": [ + -4.634786605834961 + ], + "logprob_reward": -0.009786837734282017, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786837734282017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18744458258152008, + 0.09973481297492981, + 0.04035063460469246, + 0.06634065508842468, + 0.08697424829006195, + 0.06053384020924568, + 0.03910993039608002, + 0.0670805498957634, + 0.04762246087193489, + 0.30480828881263733 + ], + "credit_map": [ + 0.020527172833681107, + 0.03170676901936531, + 0.020947426557540894, + 0.018709968775510788, + 0.03485281765460968, + 0.027356091886758804, + 0.016199003905057907, + 0.017145324498414993, + 0.009386027231812477, + 0.009040715172886848, + 0.006404140964150429, + 0.009618311189115047, + 0.029622342437505722, + 0.014348620548844337, + 0.010532050393521786, + 0.010782606899738312, + 0.0034207836724817753, + 0.006894804537296295, + 0.003957923036068678, + 0.0034716681111603975, + 0.0074918270111083984, + 0.005992230027914047, + 0.00466351630166173, + 0.004457884468138218, + 0.010188272222876549, + 0.00548143545165658, + 0.0058973608538508415, + 0.004736436065286398, + 0.014795984141528606, + 0.014673091471195221, + 0.0059410883113741875, + 0.004626990761607885, + 0.006949120666831732, + 0.009502938017249107, + 0.004900486208498478, + 0.005644462537020445, + 0.02465684339404106, + 0.010251231491565704, + 0.0072133103385567665, + 0.01785586029291153, + 0.012241184711456299, + 0.00847458653151989, + 0.005673589650541544, + 0.0037951949052512646, + 0.009548007510602474, + 0.008873370476067066, + 0.006771768443286419, + 0.005156140308827162, + 0.0048654270358383656, + 0.0070498306304216385, + 0.002508200006559491, + 0.002999264979735017, + 0.004097962286323309, + 0.009840821847319603, + 0.004247497301548719, + 0.003500925609841943, + 0.01263864804059267, + 0.007029800675809383, + 0.0067359572276473045, + 0.006103122606873512, + 0.013971160165965557, + 0.010175036266446114, + 0.0062404414638876915, + 0.004186383914202452, + 0.004747909959405661, + 0.005744879133999348, + 0.004601027350872755, + 0.00432839198037982, + 0.007712952792644501, + 0.010452567599713802, + 0.005674539599567652, + 0.004360191058367491, + 0.022361012175679207, + 0.017467189580202103, + 0.022723183035850525, + 0.01866573840379715, + 0.018290704116225243, + 0.07988102734088898, + 0.02664785273373127, + 0.0987715795636177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.009786837734282017 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_5/index_tensor([5561, 5561])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_5/index_tensor([5561, 5561])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5bd443545659bbb6666e13994934deb72c81d3c9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_5/index_tensor([5561, 5561])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot follows closely behind, capturing a heroic general clad in brilliant silver armor as he rides a muscular white steed at full speed across a dusty ancient battlefield. The general's armor, meticulously polished to a mirror-smooth surface that reflects the setting sun, is intricately engraved with heraldry. He wears a helmet crowned with crimson plumes, his face stern and his gaze piercing, revealing the composure and determination of a lifetime of battle. His steed is powerful and agile, its pure white coat matted with battlefield dust and sweat, its long mane dancing wildly in the wind. Their setting is a vast, desolate plain, littered with broken flags and abandoned weapons. On the distant horizon, the setting sun paints the sky a fiery orange and deep purple, casting long, dramatic shadows. The general grips the reins tightly, his body rising and falling powerfully to the rhythm of his steed. His scarlet cape billows violently behind him like flames at the high speed. Each thunderous thump of the steed's hooves on the cracked earth stirs up vast clouds of dust, dyed golden by the setting sun, enveloping them. When the camera is angled towards the sun, a dazzling, cinematic horizontal lens flare periodically sweeps across the frame, adding a powerful visual impact and solemn atmosphere to this epic scene of power and determination.", + "seed": 32478, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_5/index_5561_seed_32478_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03210883168503642, + "vlm_logprob": -0.03210883168503642, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03210883168503642, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10377798974514008 + ], + "opposite_logprob_tokens": [ + -2.3537778854370117 + ], + "logprob_reward": -0.10377798974514008, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10377798974514008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16944167017936707, + 0.10677462816238403, + 0.06259099394083023, + 0.25329142808914185, + 0.1383880376815796, + 0.08930276334285736, + 0.041553281247615814, + 0.06355787068605423, + 0.036940690129995346, + 0.038158662617206573 + ], + "credit_map": [ + 0.02227182500064373, + 0.01977892778813839, + 0.018851079046726227, + 0.027591442689299583, + 0.016139648854732513, + 0.021205900236964226, + 0.01953546516597271, + 0.024067381396889687, + 0.015122320502996445, + 0.010955560021102428, + 0.012130175717175007, + 0.010162466205656528, + 0.010194581001996994, + 0.014999447390437126, + 0.015927527099847794, + 0.017282545566558838, + 0.010098287835717201, + 0.010532292537391186, + 0.006030675023794174, + 0.006936578545719385, + 0.007797013036906719, + 0.005110882222652435, + 0.008910526521503925, + 0.007174738682806492, + 0.04153772071003914, + 0.022568153217434883, + 0.013039453886449337, + 0.009714142419397831, + 0.04870373755693436, + 0.07441996783018112, + 0.02230929397046566, + 0.020998962223529816, + 0.015169495716691017, + 0.013040316291153431, + 0.016010798513889313, + 0.007453195285052061, + 0.022741936147212982, + 0.029757382348179817, + 0.011815263889729977, + 0.022399647161364555, + 0.013345212675631046, + 0.01127912662923336, + 0.010662493295967579, + 0.008119851350784302, + 0.012227221392095089, + 0.015588940121233463, + 0.008746949955821037, + 0.009332968853414059, + 0.0059685613960027695, + 0.004319627769291401, + 0.007063155062496662, + 0.0060219247825443745, + 0.006194681860506535, + 0.004235501401126385, + 0.004304677713662386, + 0.003445148468017578, + 0.007787493057549, + 0.008415731601417065, + 0.009276768192648888, + 0.0067427209578454494, + 0.008322110399603844, + 0.008254358544945717, + 0.008644705638289452, + 0.006113984156399965, + 0.00593972485512495, + 0.009249386377632618, + 0.00403164979070425, + 0.003732047276571393, + 0.003384771989658475, + 0.0035172475036233664, + 0.004299644380807877, + 0.0027862172573804855, + 0.0028411750681698322, + 0.005245622713118792, + 0.0075630974024534225, + 0.004780405666679144, + 0.004639435093849897, + 0.004218645393848419, + 0.004145990125834942, + 0.004724288824945688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.10377798974514008 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070948913693428 + ], + "opposite_logprob_tokens": [ + -4.14107084274292 + ], + "logprob_reward": -0.016070948913693428, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070948913693428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22359077632427216, + 0.15482909977436066, + 0.11126995086669922, + 0.08133867383003235, + 0.13522468507289886, + 0.054981082677841187, + 0.08598130196332932, + 0.03725888952612877, + 0.04316733032464981, + 0.07235819101333618 + ], + "credit_map": [ + 0.022435329854488373, + 0.02650342509150505, + 0.018955841660499573, + 0.03315991535782814, + 0.026956791058182716, + 0.041395340114831924, + 0.02324330247938633, + 0.0309408288449049, + 0.020382748916745186, + 0.0188608355820179, + 0.019801655784249306, + 0.018620533868670464, + 0.01960655115544796, + 0.01282323244959116, + 0.02401905506849289, + 0.020714474841952324, + 0.011279266327619553, + 0.013126157224178314, + 0.012104520574212074, + 0.013468306511640549, + 0.02096834033727646, + 0.012771531008183956, + 0.014008070342242718, + 0.013543759472668171, + 0.011841613799333572, + 0.010349828749895096, + 0.00941690243780613, + 0.00610707700252533, + 0.012347519397735596, + 0.012539274990558624, + 0.008778619579970837, + 0.009957837872207165, + 0.011140606366097927, + 0.01157366018742323, + 0.018282903358340263, + 0.010801369324326515, + 0.01845954917371273, + 0.02640419825911522, + 0.01290243212133646, + 0.025659961625933647, + 0.007595314644277096, + 0.0057334755547344685, + 0.008441220037639141, + 0.007226207759231329, + 0.0062022823840379715, + 0.009062235243618488, + 0.005394467152655125, + 0.005325877107679844, + 0.009615986607968807, + 0.01201856043189764, + 0.012568531557917595, + 0.015159368515014648, + 0.010493026115000248, + 0.009454991668462753, + 0.008119341917335987, + 0.008551490493118763, + 0.007034171838313341, + 0.005250147543847561, + 0.006984012201428413, + 0.004197291098535061, + 0.004216053523123264, + 0.0037258153315633535, + 0.0032137364614754915, + 0.002637662924826145, + 0.003221972845494747, + 0.0062123853713274, + 0.006109523121267557, + 0.005566534586250782, + 0.0055098943412303925, + 0.0071079423651099205, + 0.0043656728230416775, + 0.005073405336588621, + 0.004784672521054745, + 0.010563423857092857, + 0.012266476638615131, + 0.010493358597159386, + 0.009807582013309002, + 0.008800848387181759, + 0.008292155340313911, + 0.007349679246544838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.016070948913693428 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732873618602753 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.006732873618602753, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732873618602753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23618051409721375, + 0.1470845341682434, + 0.06600186973810196, + 0.21784314513206482, + 0.06878907978534698, + 0.12518155574798584, + 0.03946981951594353, + 0.029573101550340652, + 0.033601149916648865, + 0.036275263875722885 + ], + "credit_map": [ + 0.03164907544851303, + 0.03728896379470825, + 0.0332188755273819, + 0.03794100508093834, + 0.019344404339790344, + 0.029896151274442673, + 0.027927525341510773, + 0.01891452632844448, + 0.02391584776341915, + 0.019670361652970314, + 0.01763177290558815, + 0.02624492160975933, + 0.016458673402667046, + 0.01383892446756363, + 0.017094774171710014, + 0.012229264713823795, + 0.013948657549917698, + 0.008302411064505577, + 0.006981167010962963, + 0.008915524929761887, + 0.008947107940912247, + 0.005804531741887331, + 0.006826347671449184, + 0.006276118569076061, + 0.028280969709157944, + 0.027722066268324852, + 0.027123238891363144, + 0.023344460874795914, + 0.03010004572570324, + 0.033809490501880646, + 0.02350468561053276, + 0.023958196863532066, + 0.00586163904517889, + 0.006947411224246025, + 0.011379196308553219, + 0.005646192468702793, + 0.008900247514247894, + 0.01010608859360218, + 0.006304023787379265, + 0.013644278980791569, + 0.021497130393981934, + 0.01579132117331028, + 0.031735505908727646, + 0.012859652750194073, + 0.01113115530461073, + 0.01117133628576994, + 0.01062006875872612, + 0.010375382378697395, + 0.005078780464828014, + 0.005424061790108681, + 0.007840605452656746, + 0.003760617459192872, + 0.0056989239528775215, + 0.004801060538738966, + 0.0032375638838857412, + 0.0036282073706388474, + 0.005186022724956274, + 0.0033086249604821205, + 0.006107015069574118, + 0.004372354596853256, + 0.0024911051150411367, + 0.002511776750907302, + 0.003086938988417387, + 0.002509261481463909, + 0.004514018539339304, + 0.006888256408274174, + 0.006212031468749046, + 0.0036708267871290445, + 0.0037130711134523153, + 0.003551093628630042, + 0.0026503214612603188, + 0.0024015295784920454, + 0.004122935701161623, + 0.00722115533426404, + 0.00630358699709177, + 0.004348693881183863, + 0.00390321621671319, + 0.0038265518378466368, + 0.003159567480906844, + 0.0033895564265549183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.006732873618602753 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076275295577943325 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.0076275295577943325, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076275295577943325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14690186083316803, + 0.07324588298797607, + 0.05042277276515961, + 0.0677555575966835, + 0.42689448595046997, + 0.0688847079873085, + 0.019070284441113472, + 0.03010796196758747, + 0.0445876270532608, + 0.07212880253791809 + ], + "credit_map": [ + 0.013227807357907295, + 0.020992429926991463, + 0.019122352823615074, + 0.03439967334270477, + 0.010780581273138523, + 0.014978120103478432, + 0.0195234976708889, + 0.013877409510314465, + 0.0081752585247159, + 0.009739909321069717, + 0.007376234512776136, + 0.012353302910923958, + 0.007420034147799015, + 0.008503761142492294, + 0.00966671109199524, + 0.0100106718018651, + 0.00928930751979351, + 0.009189504198729992, + 0.00434808200225234, + 0.007883172482252121, + 0.005913306027650833, + 0.005490828771144152, + 0.004320523235946894, + 0.003988049924373627, + 0.010631818324327469, + 0.0064545851200819016, + 0.008373871445655823, + 0.006171376444399357, + 0.0075957621447741985, + 0.013473129831254482, + 0.0066135600209236145, + 0.008441461250185966, + 0.052969254553318024, + 0.02908993698656559, + 0.10108397901058197, + 0.0215190090239048, + 0.022882675752043724, + 0.035334665328264236, + 0.021097170189023018, + 0.14291782677173615, + 0.010541980154812336, + 0.009541582316160202, + 0.014889572747051716, + 0.008872248232364655, + 0.0055536204017698765, + 0.007033012341707945, + 0.005832898896187544, + 0.006619802676141262, + 0.002751734107732773, + 0.0022847778163850307, + 0.0029515770729631186, + 0.0022852164693176746, + 0.0026878416538238525, + 0.0024864773731678724, + 0.001843006699346006, + 0.0017796535976231098, + 0.0058999452739953995, + 0.00403023324906826, + 0.00670410692691803, + 0.0034707498271018267, + 0.003447621362283826, + 0.0018634260632097721, + 0.002172894775867462, + 0.002518987050279975, + 0.0068935975432395935, + 0.011588593013584614, + 0.007131311576813459, + 0.0036547258496284485, + 0.004477819427847862, + 0.004211242310702801, + 0.003701830981299281, + 0.00292851310223341, + 0.00630231574177742, + 0.00953768938779831, + 0.015430036000907421, + 0.006205315236002207, + 0.009718543849885464, + 0.009195763617753983, + 0.007171382196247578, + 0.00856775976717472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0076275295577943325 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026334816589951515 + ], + "opposite_logprob_tokens": [ + -3.651334762573242 + ], + "logprob_reward": -0.026334816589951515, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026334816589951515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11255435645580292, + 0.10501591116189957, + 0.14252085983753204, + 0.11194619536399841, + 0.0714372918009758, + 0.05258313566446304, + 0.045115116983652115, + 0.08222059160470963, + 0.04338160529732704, + 0.23322495818138123 + ], + "credit_map": [ + 0.015282254666090012, + 0.013642343692481518, + 0.012156229466199875, + 0.019973305985331535, + 0.010115068405866623, + 0.01323187630623579, + 0.014872277155518532, + 0.013281011022627354, + 0.012590743601322174, + 0.013246774673461914, + 0.010769673623144627, + 0.019906580448150635, + 0.012506254017353058, + 0.00901718344539404, + 0.012644737027585506, + 0.014333958737552166, + 0.01854608580470085, + 0.02069535106420517, + 0.014324089512228966, + 0.019494572654366493, + 0.020606696605682373, + 0.016391931101679802, + 0.014581551775336266, + 0.017880581319332123, + 0.021388689056038857, + 0.015545936301350594, + 0.008528389036655426, + 0.008296706713736057, + 0.016855118796229362, + 0.024078531190752983, + 0.008247940801084042, + 0.00900487694889307, + 0.01174595020711422, + 0.006482733879238367, + 0.008182739838957787, + 0.005082028917968273, + 0.009481843560934067, + 0.011458395048975945, + 0.005678643938153982, + 0.013324951753020287, + 0.007456430699676275, + 0.006221921648830175, + 0.006380857899785042, + 0.006152073387056589, + 0.006841996684670448, + 0.007952451705932617, + 0.005557899363338947, + 0.006019507069140673, + 0.00706952391192317, + 0.006214558612555265, + 0.005169386509805918, + 0.006175462622195482, + 0.006397534161806107, + 0.004779816139489412, + 0.005037236958742142, + 0.00427160132676363, + 0.021327003836631775, + 0.008182594552636147, + 0.02620088867843151, + 0.007431489415466785, + 0.005532295908778906, + 0.005447430070489645, + 0.003457415848970413, + 0.004641473758965731, + 0.005531789734959602, + 0.006200132891535759, + 0.008605102077126503, + 0.0068292273208498955, + 0.004511888604611158, + 0.0035790724214166403, + 0.004070100840181112, + 0.004054293967783451, + 0.02128240279853344, + 0.031973857432603836, + 0.049496326595544815, + 0.021009385585784912, + 0.029346322640776634, + 0.03181375935673714, + 0.021806208416819572, + 0.026496699079871178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.026334816589951515 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows closely behind, capturing a heroic general clad in brilliant silver armor as he rides a muscular white steed at full speed across a dusty ancient battlefield. The general's armor, meticulously polished to a mirror-smooth surface that reflects the setting sun, is intricately engraved with heraldry. He wears a helmet crowned with crimson plumes, his face stern and his gaze piercing, revealing the composure and determination of a lifetime of battle. His steed is powerful and agile, its pure white coat matted with battlefield dust and sweat, its long mane dancing wildly in the wind. Their setting is a vast, desolate plain, littered with broken flags and abandoned weapons. On the distant horizon, the setting sun paints the sky a fiery orange and deep purple, casting long, dramatic shadows. The general grips the reins tightly, his body rising and falling powerfully to the rhythm of his steed. His scarlet cape billows violently behind him like flames at the high speed. Each thunderous thump of the steed's hooves on the cracked earth stirs up vast clouds of dust, dyed golden by the setting sun, enveloping them. When the camera is angled towards the sun, a dazzling, cinematic horizontal lens flare periodically sweeps across the frame, adding a powerful visual impact and solemn atmosphere to this epic scene of power and determination.", + "seed": 32479, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_5/index_5561_seed_32479_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.022247753292322158, + "vlm_logprob": -0.022247753292322158, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.022247753292322158, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05711478367447853 + ], + "opposite_logprob_tokens": [ + -2.932114839553833 + ], + "logprob_reward": -0.05711478367447853, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05711478367447853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2365093231201172, + 0.14917248487472534, + 0.09121791273355484, + 0.17550915479660034, + 0.07429762184619904, + 0.04851164668798447, + 0.05216657742857933, + 0.051867593079805374, + 0.045928530395030975, + 0.07481922209262848 + ], + "credit_map": [ + 0.022705553099513054, + 0.02538352832198143, + 0.03970993310213089, + 0.03412686660885811, + 0.025466542690992355, + 0.03327055647969246, + 0.030468475073575974, + 0.025377877056598663, + 0.014207866042852402, + 0.016104254871606827, + 0.013773814775049686, + 0.021865684539079666, + 0.01905260980129242, + 0.030532529577612877, + 0.017575513571500778, + 0.016060203313827515, + 0.008605490438640118, + 0.010783808305859566, + 0.009489470161497593, + 0.011610493995249271, + 0.010764040984213352, + 0.01746145263314247, + 0.011778479442000389, + 0.010724682360887527, + 0.01197095401585102, + 0.01753096655011177, + 0.016188064590096474, + 0.0143484678119421, + 0.025708256289362907, + 0.04683937132358551, + 0.021509971469640732, + 0.02141309343278408, + 0.005891995970159769, + 0.007609606254845858, + 0.007165257818996906, + 0.01005373802036047, + 0.010685648769140244, + 0.012842390686273575, + 0.011551083996891975, + 0.008497904986143112, + 0.004674003459513187, + 0.004427922423928976, + 0.005942252930253744, + 0.008091731928288937, + 0.0071331229992210865, + 0.004955552984029055, + 0.006707620341330767, + 0.006579437758773565, + 0.005197749473154545, + 0.004175152629613876, + 0.008041818626224995, + 0.008107437752187252, + 0.006664678454399109, + 0.006781439296901226, + 0.007101752329617739, + 0.0060965511947870255, + 0.008868701756000519, + 0.0038071589078754187, + 0.008812768384814262, + 0.009618421085178852, + 0.0070360018871724606, + 0.00419303635135293, + 0.005061222240328789, + 0.00447028037160635, + 0.007267299573868513, + 0.002964332466945052, + 0.01198643445968628, + 0.008977094665169716, + 0.003686553332954645, + 0.004257874563336372, + 0.003720359643921256, + 0.0030685828533023596, + 0.019083533436059952, + 0.008768709376454353, + 0.020785335451364517, + 0.007330281659960747, + 0.006066410802304745, + 0.004052522126585245, + 0.005051518324762583, + 0.0036809039302170277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.05711478367447853 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539238668978214 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539238668978214, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539238668978214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1926155388355255, + 0.08602774143218994, + 0.10172293335199356, + 0.04231017082929611, + 0.046106964349746704, + 0.03829994797706604, + 0.04271756485104561, + 0.06650716066360474, + 0.17209354043006897, + 0.21159833669662476 + ], + "credit_map": [ + 0.019744189456105232, + 0.021981598809361458, + 0.03374577686190605, + 0.028397826477885246, + 0.016041167080402374, + 0.029243456199765205, + 0.02540072426199913, + 0.018060823902487755, + 0.009284701198339462, + 0.010582664050161839, + 0.012555358931422234, + 0.012556664645671844, + 0.010161390528082848, + 0.013429693877696991, + 0.006900020409375429, + 0.010557251051068306, + 0.008683815598487854, + 0.010678029619157314, + 0.012035995721817017, + 0.023569593206048012, + 0.013127215206623077, + 0.009893050417304039, + 0.014690838754177094, + 0.0090444041416049, + 0.004576815292239189, + 0.004086190834641457, + 0.00604519946500659, + 0.004941562656313181, + 0.004926473833620548, + 0.006291268393397331, + 0.006491342559456825, + 0.004951318725943565, + 0.005188272334635258, + 0.005167888477444649, + 0.004647109657526016, + 0.005769597832113504, + 0.0066498382948338985, + 0.006761595141142607, + 0.006671152543276548, + 0.005251518450677395, + 0.0033331161830574274, + 0.0032974081113934517, + 0.005584810860455036, + 0.006953331176191568, + 0.00490791629999876, + 0.0043383375741541386, + 0.004810564685612917, + 0.005074464716017246, + 0.004536814521998167, + 0.002950477646663785, + 0.0050197383388876915, + 0.006022299639880657, + 0.005242147948592901, + 0.005879293195903301, + 0.007220573257654905, + 0.005846222396939993, + 0.011190257035195827, + 0.0028293346986174583, + 0.019243281334638596, + 0.012637214735150337, + 0.0049171773716807365, + 0.004942682106047869, + 0.004395778756588697, + 0.006351438816636801, + 0.02292224019765854, + 0.013898526318371296, + 0.03076346032321453, + 0.03003295697271824, + 0.026542766019701958, + 0.011371192522346973, + 0.023099953308701515, + 0.013462460599839687, + 0.04460152983665466, + 0.02041577734053135, + 0.04651658609509468, + 0.029527870938181877, + 0.01495975349098444, + 0.019798824563622475, + 0.01932799071073532, + 0.016450034454464912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.012539238668978214 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626346778124571 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626346778124571, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626346778124571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29245540499687195, + 0.11366570740938187, + 0.07807334512472153, + 0.05397196114063263, + 0.12240679562091827, + 0.05612782761454582, + 0.029371902346611023, + 0.06699712574481964, + 0.0548366904258728, + 0.132093146443367 + ], + "credit_map": [ + 0.03413878381252289, + 0.03037947230041027, + 0.06599391996860504, + 0.047434426844120026, + 0.02056952379643917, + 0.0349227711558342, + 0.03140190243721008, + 0.027614634484052658, + 0.015332214534282684, + 0.013422729447484016, + 0.01860613375902176, + 0.021875133737921715, + 0.010129113681614399, + 0.009413561783730984, + 0.014398615807294846, + 0.010488204658031464, + 0.006315318401902914, + 0.007317358627915382, + 0.011582738719880581, + 0.013765445910394192, + 0.011053080670535564, + 0.010264629498124123, + 0.00830770842730999, + 0.009467077441513538, + 0.006397433578968048, + 0.005404472351074219, + 0.009411468170583248, + 0.00692167179659009, + 0.005857003852725029, + 0.00546528585255146, + 0.007794357370585203, + 0.006720270495861769, + 0.020556993782520294, + 0.014789044857025146, + 0.024639908224344254, + 0.021474605426192284, + 0.010728229768574238, + 0.009356544353067875, + 0.009829682298004627, + 0.011031798087060452, + 0.007056655362248421, + 0.005149967037141323, + 0.00897910911589861, + 0.010101551190018654, + 0.006741848774254322, + 0.00530164735391736, + 0.00666575925424695, + 0.00613129697740078, + 0.0028614436741918325, + 0.0031492209527641535, + 0.004998963791877031, + 0.0043181125074625015, + 0.003536326577886939, + 0.0027945477049797773, + 0.0036394880153238773, + 0.004073802847415209, + 0.010071750730276108, + 0.004751726984977722, + 0.010487566702067852, + 0.018118171021342278, + 0.0068415929563343525, + 0.004613584838807583, + 0.006275051273405552, + 0.005837689153850079, + 0.006955851800739765, + 0.005741087719798088, + 0.01275684405118227, + 0.008240374736487865, + 0.006314917001873255, + 0.003724209498614073, + 0.006721349433064461, + 0.004382065497338772, + 0.0399739146232605, + 0.008746905252337456, + 0.046999599784612656, + 0.015429235994815826, + 0.0040558576583862305, + 0.0045190309174358845, + 0.006557247135788202, + 0.005811371374875307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.007626346778124571 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076275295577943325 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.0076275295577943325, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076275295577943325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07196293026208878, + 0.04439178481698036, + 0.06826645135879517, + 0.061128079891204834, + 0.04302201420068741, + 0.030339088290929794, + 0.026134410873055458, + 0.05646856874227524, + 0.09271464496850967, + 0.5055720210075378 + ], + "credit_map": [ + 0.008042515255510807, + 0.007867303676903248, + 0.01424808707088232, + 0.013446207158267498, + 0.005018183961510658, + 0.009194052778184414, + 0.007529564667493105, + 0.006617013365030289, + 0.004994634073227644, + 0.005309867672622204, + 0.003952191211283207, + 0.0050324383191764355, + 0.005067333113402128, + 0.0111858444288373, + 0.003861968871206045, + 0.004987507592886686, + 0.008941244333982468, + 0.008627625182271004, + 0.012842228636145592, + 0.018094006925821304, + 0.005777956452220678, + 0.0058308737352490425, + 0.004777892492711544, + 0.003374628257006407, + 0.006028014700859785, + 0.005594819784164429, + 0.009219435043632984, + 0.006846388336271048, + 0.007563479244709015, + 0.013495015911757946, + 0.006191241554915905, + 0.006189681123942137, + 0.0050533185712993145, + 0.005834616720676422, + 0.006158730015158653, + 0.007572506554424763, + 0.0037153458688408136, + 0.00622039707377553, + 0.0035028059501200914, + 0.004964294843375683, + 0.005329038482159376, + 0.003256079275161028, + 0.004800432361662388, + 0.0052960105240345, + 0.0025555877946317196, + 0.00308975949883461, + 0.0031673379708081484, + 0.002844843314960599, + 0.002846538322046399, + 0.001973219681531191, + 0.003763491054996848, + 0.0046808454208076, + 0.002649173606187105, + 0.002167720580473542, + 0.0044371141120791435, + 0.003616306697949767, + 0.005216430407017469, + 0.0038185915909707546, + 0.008344089612364769, + 0.014158623293042183, + 0.007894386537373066, + 0.005436609033495188, + 0.005605779588222504, + 0.005994059145450592, + 0.014402099885046482, + 0.008163114078342915, + 0.02737317979335785, + 0.0199135672301054, + 0.006980258040130138, + 0.004202661104500294, + 0.0065110959112644196, + 0.0051686749793589115, + 0.1309647560119629, + 0.0742684155702591, + 0.1488896608352661, + 0.04928312823176384, + 0.027423664927482605, + 0.02642185613512993, + 0.027537398040294647, + 0.02078310213983059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0076275295577943325 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026330867782235146 + ], + "opposite_logprob_tokens": [ + -3.6513309478759766 + ], + "logprob_reward": -0.026330867782235146, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026330867782235146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1388085037469864, + 0.07174526154994965, + 0.06532315164804459, + 0.03629717230796814, + 0.025327082723379135, + 0.04200108349323273, + 0.07103808224201202, + 0.22723190486431122, + 0.2315971851348877, + 0.09063056856393814 + ], + "credit_map": [ + 0.0205061137676239, + 0.025019556283950806, + 0.020988095551729202, + 0.028345050290226936, + 0.0077326493337750435, + 0.012404452078044415, + 0.013508313335478306, + 0.01030427310615778, + 0.008318454958498478, + 0.007865484803915024, + 0.00840248167514801, + 0.010759498924016953, + 0.006658205296844244, + 0.01525883749127388, + 0.007372637279331684, + 0.007109656929969788, + 0.006072466727346182, + 0.007217788137495518, + 0.008549056947231293, + 0.006998090539127588, + 0.007553099188953638, + 0.016038641333580017, + 0.0057909442111849785, + 0.007103067357093096, + 0.0028433790430426598, + 0.003884225618094206, + 0.0055312346667051315, + 0.003594952402636409, + 0.004604021552950144, + 0.006828837562352419, + 0.0054576946422457695, + 0.0035528282169252634, + 0.0029436235781759024, + 0.0027857229579240084, + 0.003732952056452632, + 0.004213915206491947, + 0.002900023479014635, + 0.0027174020651727915, + 0.0032610914204269648, + 0.002772353822365403, + 0.004866664297878742, + 0.004008364863693714, + 0.008066464215517044, + 0.007165026850998402, + 0.00402029138058424, + 0.005666409619152546, + 0.003751647425815463, + 0.004456214606761932, + 0.0067406101152300835, + 0.005371171981096268, + 0.009132815524935722, + 0.02092025801539421, + 0.007274255156517029, + 0.00542557192966342, + 0.007986224256455898, + 0.008187171071767807, + 0.037484511733055115, + 0.008366743102669716, + 0.08018018305301666, + 0.044755175709724426, + 0.013513968326151371, + 0.013217651285231113, + 0.012197908014059067, + 0.017515774816274643, + 0.036481764167547226, + 0.023335609585046768, + 0.06556497514247894, + 0.035180509090423584, + 0.028697270900011063, + 0.010477673262357712, + 0.0208706296980381, + 0.010988765396177769, + 0.026851287111639977, + 0.005783137399703264, + 0.030237272381782532, + 0.010577994398772717, + 0.0033539680298417807, + 0.004521118942648172, + 0.005547942128032446, + 0.003757848171517253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.026330867782235146 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows closely behind, capturing a heroic general clad in brilliant silver armor as he rides a muscular white steed at full speed across a dusty ancient battlefield. The general's armor, meticulously polished to a mirror-smooth surface that reflects the setting sun, is intricately engraved with heraldry. He wears a helmet crowned with crimson plumes, his face stern and his gaze piercing, revealing the composure and determination of a lifetime of battle. His steed is powerful and agile, its pure white coat matted with battlefield dust and sweat, its long mane dancing wildly in the wind. Their setting is a vast, desolate plain, littered with broken flags and abandoned weapons. On the distant horizon, the setting sun paints the sky a fiery orange and deep purple, casting long, dramatic shadows. The general grips the reins tightly, his body rising and falling powerfully to the rhythm of his steed. His scarlet cape billows violently behind him like flames at the high speed. Each thunderous thump of the steed's hooves on the cracked earth stirs up vast clouds of dust, dyed golden by the setting sun, enveloping them. When the camera is angled towards the sun, a dazzling, cinematic horizontal lens flare periodically sweeps across the frame, adding a powerful visual impact and solemn atmosphere to this epic scene of power and determination.", + "seed": 32480, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_5/index_5561_seed_32480_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05968152992427349, + "vlm_logprob": -0.05968152992427349, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.05968152992427349, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20415207743644714 + ], + "opposite_logprob_tokens": [ + -1.7041521072387695 + ], + "logprob_reward": -0.20415207743644714, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20415207743644714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05465452000498772, + 0.027328750118613243, + 0.03317256271839142, + 0.03921196237206459, + 0.025410562753677368, + 0.01827375218272209, + 0.7285770177841187, + 0.01534618716686964, + 0.02771059051156044, + 0.030314069241285324 + ], + "credit_map": [ + 0.006372993811964989, + 0.005069108214229345, + 0.01012642402201891, + 0.006739253643900156, + 0.006269342731684446, + 0.0072761643677949905, + 0.0058681536465883255, + 0.006933083292096853, + 0.0026815992314368486, + 0.002409901935607195, + 0.004815436899662018, + 0.003563618753105402, + 0.002918049693107605, + 0.0029993378557264805, + 0.004255560226738453, + 0.003685247851535678, + 0.004749532788991928, + 0.002538501052185893, + 0.004262308124452829, + 0.0036900993436574936, + 0.0026849773712456226, + 0.006099863443523645, + 0.0035378923639655113, + 0.0056093912571668625, + 0.00836207065731287, + 0.0031611823942512274, + 0.005294493865221739, + 0.005137438420206308, + 0.002522469963878393, + 0.0035847558174282312, + 0.0063265603967010975, + 0.0048229931853711605, + 0.005116769578307867, + 0.0027211098931729794, + 0.0036041175480931997, + 0.0027260466013103724, + 0.0021837358362972736, + 0.002615258563309908, + 0.0032421418000012636, + 0.0032013857271522284, + 0.0038673225790262222, + 0.0015759284142404795, + 0.0026663686148822308, + 0.0026178820990025997, + 0.0019873566925525665, + 0.001424515270628035, + 0.0023412350565195084, + 0.0017931448528543115, + 0.09658988565206528, + 0.06865499913692474, + 0.12287818640470505, + 0.09555386006832123, + 0.06783483177423477, + 0.04030073434114456, + 0.07313573360443115, + 0.16362881660461426, + 0.0019129033898934722, + 0.001856275019235909, + 0.001715832739137113, + 0.0017561664571985602, + 0.003357900306582451, + 0.0008129389607347548, + 0.002102258848026395, + 0.0018319125520065427, + 0.0034331728238612413, + 0.0028463227208703756, + 0.004112279042601585, + 0.0033446261659264565, + 0.0032965021673589945, + 0.0038192456122487783, + 0.0034382224548608065, + 0.0034202218521386385, + 0.004171083681285381, + 0.0036066167522221804, + 0.003712944919243455, + 0.003132500685751438, + 0.007960936985909939, + 0.002057651523500681, + 0.003618301823735237, + 0.0020540340337902308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.20415207743644714 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023297041654586792 + ], + "opposite_logprob_tokens": [ + -3.773297071456909 + ], + "logprob_reward": -0.023297041654586792, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297041654586792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08856253325939178, + 0.09149954468011856, + 0.12712286412715912, + 0.044108811765909195, + 0.05772644281387329, + 0.041999466717243195, + 0.4317094385623932, + 0.0384090431034565, + 0.02149277552962303, + 0.05736902728676796 + ], + "credit_map": [ + 0.013607325963675976, + 0.0089309923350811, + 0.014733592048287392, + 0.010856173001229763, + 0.009577571414411068, + 0.011046714149415493, + 0.007558938581496477, + 0.012251228094100952, + 0.014171340502798557, + 0.0073103695176541805, + 0.014646140858530998, + 0.01309172809123993, + 0.008159888908267021, + 0.012257536873221397, + 0.009002371691167355, + 0.012860171496868134, + 0.017109805718064308, + 0.008217922411859035, + 0.01530026737600565, + 0.008587993681430817, + 0.006466692313551903, + 0.031005483120679855, + 0.01315248105674982, + 0.0272822268307209, + 0.00469393702223897, + 0.004183617886155844, + 0.005235071759670973, + 0.006489428225904703, + 0.003050314262509346, + 0.00579150440171361, + 0.00735542131587863, + 0.007309513632208109, + 0.010307790711522102, + 0.006761098280549049, + 0.006711853668093681, + 0.006374628748744726, + 0.007371978834271431, + 0.00519732479006052, + 0.007815524004399776, + 0.007186243776232004, + 0.006024164613336325, + 0.004576212260872126, + 0.0068983156234025955, + 0.005637129303067923, + 0.004091199487447739, + 0.0030747046694159508, + 0.0070922900922596455, + 0.00460545439273119, + 0.056971266865730286, + 0.03983106091618538, + 0.07104998081922531, + 0.05809732899069786, + 0.0401480495929718, + 0.023418160155415535, + 0.04439341649413109, + 0.09780019521713257, + 0.004892196971923113, + 0.004574592690914869, + 0.006509364582598209, + 0.005582060664892197, + 0.0048514800146222115, + 0.0021084377076476812, + 0.00530395470559597, + 0.004586956463754177, + 0.0034438692964613438, + 0.0017976363888010383, + 0.0033129355870187283, + 0.002075767610222101, + 0.0028853958938270807, + 0.002687909174710512, + 0.002331233350560069, + 0.0029580278787761927, + 0.008166089653968811, + 0.010582475923001766, + 0.00860220193862915, + 0.01105133444070816, + 0.004892428871244192, + 0.0034814123064279556, + 0.005870374385267496, + 0.004722710698843002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.023297041654586792 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779755026102066 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779755026102066, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779755026102066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10545264929533005, + 0.09298740327358246, + 0.1810849905014038, + 0.1068752259016037, + 0.075858935713768, + 0.07816861569881439, + 0.21064677834510803, + 0.03297565504908562, + 0.04095905274152756, + 0.0749906674027443 + ], + "credit_map": [ + 0.013029498979449272, + 0.011116941459476948, + 0.02011188119649887, + 0.012852118350565434, + 0.01229185238480568, + 0.009835106320679188, + 0.013318479992449284, + 0.012896774336695671, + 0.010269079357385635, + 0.008644700050354004, + 0.014323199167847633, + 0.013852966949343681, + 0.008238890208303928, + 0.013641538098454475, + 0.011063594371080399, + 0.012953422963619232, + 0.03126306086778641, + 0.013202241621911526, + 0.021950935944914818, + 0.015472250990569592, + 0.008280126377940178, + 0.041979219764471054, + 0.013978364877402782, + 0.03495877608656883, + 0.01514926552772522, + 0.009219739586114883, + 0.01422111876308918, + 0.016852540895342827, + 0.006116722710430622, + 0.012856301851570606, + 0.017820235341787338, + 0.014639300294220448, + 0.020234903320670128, + 0.0075781075283885, + 0.010486725717782974, + 0.009554063901305199, + 0.004381880164146423, + 0.007182920351624489, + 0.006903876084834337, + 0.009536460973322392, + 0.021988747641444206, + 0.005965370684862137, + 0.014551874250173569, + 0.007456496823579073, + 0.0047037480399012566, + 0.005956872366368771, + 0.009479301050305367, + 0.008066200651228428, + 0.032252948731184006, + 0.018278302624821663, + 0.038739100098609924, + 0.025228427723050117, + 0.017681892961263657, + 0.012226369231939316, + 0.0207632128149271, + 0.045476529747247696, + 0.005482043139636517, + 0.003529567737132311, + 0.006257840897887945, + 0.004577999468892813, + 0.0040422966703772545, + 0.0022784231696277857, + 0.003272632835432887, + 0.0035348518285900354, + 0.008420521393418312, + 0.0041593946516513824, + 0.0050809248350560665, + 0.004423738922923803, + 0.0033613473642617464, + 0.009308906272053719, + 0.0034156963229179382, + 0.0027885197196155787, + 0.012901020236313343, + 0.013423353433609009, + 0.009609629400074482, + 0.013509120792150497, + 0.008660105057060719, + 0.003627764992415905, + 0.007141379173845053, + 0.006118291057646275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.009779755026102066 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538531795144081 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538531795144081, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538531795144081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03735441341996193, + 0.09192066639661789, + 0.6306727528572083, + 0.02656266838312149, + 0.01153395976871252, + 0.012554013170301914, + 0.06887587904930115, + 0.019626347348093987, + 0.02007131278514862, + 0.0808279812335968 + ], + "credit_map": [ + 0.00589792849496007, + 0.005844587925821543, + 0.006477463524788618, + 0.0042594196274876595, + 0.0027473510708659887, + 0.004163783509284258, + 0.002940167672932148, + 0.005023712757974863, + 0.016647977754473686, + 0.007289221044629812, + 0.012436204589903355, + 0.009394098073244095, + 0.007349320221692324, + 0.015813522040843964, + 0.010968347080051899, + 0.012021980248391628, + 0.08916451036930084, + 0.04942799359560013, + 0.07729402184486389, + 0.05200666934251785, + 0.031660210341215134, + 0.15071885287761688, + 0.05116812512278557, + 0.12923234701156616, + 0.005498002283275127, + 0.0024187681265175343, + 0.003997421357780695, + 0.003531910013407469, + 0.0018843275029212236, + 0.0024043393786996603, + 0.0036515805404633284, + 0.0031763191800564528, + 0.002780429320409894, + 0.0011018336517736316, + 0.0019054447766393423, + 0.0012452879454940557, + 0.0007326810737140477, + 0.0011416531633585691, + 0.0010120635852217674, + 0.0016145664267241955, + 0.0017750832485035062, + 0.0008732908754609525, + 0.00340462033636868, + 0.0013798720901831985, + 0.0009285162086598575, + 0.0009088569204322994, + 0.0013790760422125459, + 0.0019046971574425697, + 0.009034622460603714, + 0.0067794970236718655, + 0.011991101317107677, + 0.00820858869701624, + 0.006989991292357445, + 0.0039418647065758705, + 0.00640568882226944, + 0.015524519607424736, + 0.0029087646398693323, + 0.0020472395699471235, + 0.0031766854226589203, + 0.0030763200484216213, + 0.002960457233712077, + 0.0010562408715486526, + 0.0023313937708735466, + 0.0020692460238933563, + 0.004152604844421148, + 0.0018774846103042364, + 0.004160943441092968, + 0.0025895608123391867, + 0.0020208884961903095, + 0.0014151075156405568, + 0.0021440356504172087, + 0.0017106856685131788, + 0.01390584371984005, + 0.015595180913805962, + 0.010826010257005692, + 0.013811829499900341, + 0.006365637760609388, + 0.0042676497250795364, + 0.00865013338625431, + 0.007405702956020832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.012538531795144081 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864024370908737 + ], + "opposite_logprob_tokens": [ + -3.048640251159668 + ], + "logprob_reward": -0.04864024370908737, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864024370908737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17084988951683044, + 0.04345317929983139, + 0.11203068494796753, + 0.034029170870780945, + 0.030966859310865402, + 0.04223330318927765, + 0.17481432855129242, + 0.08314249664545059, + 0.15219882130622864, + 0.1562812626361847 + ], + "credit_map": [ + 0.02439083717763424, + 0.016456974670290947, + 0.028124568983912468, + 0.022925889119505882, + 0.020476024597883224, + 0.01623801700770855, + 0.022946951910853386, + 0.0192906204611063, + 0.005181873682886362, + 0.004257179331034422, + 0.00796828418970108, + 0.005110412370413542, + 0.003858331823721528, + 0.005045878700911999, + 0.005772027652710676, + 0.006259192246943712, + 0.016827179118990898, + 0.008396442048251629, + 0.013514825142920017, + 0.011151759885251522, + 0.009755217470228672, + 0.021563926711678505, + 0.012887467630207539, + 0.017933867871761322, + 0.004215828143060207, + 0.0024912937078624964, + 0.005294329021126032, + 0.005391464568674564, + 0.0032893167808651924, + 0.0033910167403519154, + 0.004980192054063082, + 0.0049757324159145355, + 0.00661492021754384, + 0.002949634101241827, + 0.0036624984350055456, + 0.0038449750281870365, + 0.0034331970382481813, + 0.0028434123378247023, + 0.0034224214032292366, + 0.004195801448076963, + 0.008464178070425987, + 0.002637450350448489, + 0.01014015544205904, + 0.006209604442119598, + 0.0028995636384934187, + 0.002691616304218769, + 0.004565962124615908, + 0.004624772816896439, + 0.027413861826062202, + 0.013869315385818481, + 0.04142739623785019, + 0.023304598405957222, + 0.012161399237811565, + 0.010681046172976494, + 0.017994916066527367, + 0.027961786836385727, + 0.010560738854110241, + 0.0064442455768585205, + 0.013455905020236969, + 0.010809088125824928, + 0.015978381037712097, + 0.005009497050195932, + 0.009261703118681908, + 0.011622939258813858, + 0.04577251896262169, + 0.011146781966090202, + 0.022870251908898354, + 0.008330355398356915, + 0.014004205353558064, + 0.026990516111254692, + 0.011490369215607643, + 0.011593829840421677, + 0.02293410710990429, + 0.0210686307400465, + 0.020246222615242004, + 0.02411426045000553, + 0.017176061868667603, + 0.015711527317762375, + 0.01827029138803482, + 0.016760166734457016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.04864024370908737 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows closely behind, capturing a heroic general clad in brilliant silver armor as he rides a muscular white steed at full speed across a dusty ancient battlefield. The general's armor, meticulously polished to a mirror-smooth surface that reflects the setting sun, is intricately engraved with heraldry. He wears a helmet crowned with crimson plumes, his face stern and his gaze piercing, revealing the composure and determination of a lifetime of battle. His steed is powerful and agile, its pure white coat matted with battlefield dust and sweat, its long mane dancing wildly in the wind. Their setting is a vast, desolate plain, littered with broken flags and abandoned weapons. On the distant horizon, the setting sun paints the sky a fiery orange and deep purple, casting long, dramatic shadows. The general grips the reins tightly, his body rising and falling powerfully to the rhythm of his steed. His scarlet cape billows violently behind him like flames at the high speed. Each thunderous thump of the steed's hooves on the cracked earth stirs up vast clouds of dust, dyed golden by the setting sun, enveloping them. When the camera is angled towards the sun, a dazzling, cinematic horizontal lens flare periodically sweeps across the frame, adding a powerful visual impact and solemn atmosphere to this epic scene of power and determination.", + "seed": 32481, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_5/index_5561_seed_32481_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04413480767980218, + "vlm_logprob": -0.04413480767980218, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.04413480767980218, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10270035266876221 + ], + "opposite_logprob_tokens": [ + -2.3527002334594727 + ], + "logprob_reward": -0.10270035266876221, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10270035266876221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16148163378238678, + 0.03158501908183098, + 0.06527648866176605, + 0.08453617990016937, + 0.15950383245944977, + 0.09506451338529587, + 0.13702130317687988, + 0.14134584367275238, + 0.046353116631507874, + 0.07783207297325134 + ], + "credit_map": [ + 0.012949545867741108, + 0.020127711817622185, + 0.02234562113881111, + 0.020055430009961128, + 0.0325714610517025, + 0.023726023733615875, + 0.016775114461779594, + 0.012930727563798428, + 0.003365489188581705, + 0.0035011437721550465, + 0.00316445785574615, + 0.0071996646001935005, + 0.0038509902078658342, + 0.00289698364213109, + 0.004118646029382944, + 0.0034876433201134205, + 0.006806512363255024, + 0.007802219595760107, + 0.007179582491517067, + 0.008155952207744122, + 0.006325921975076199, + 0.016383029520511627, + 0.006992904469370842, + 0.005630370695143938, + 0.008731071837246418, + 0.010519349947571754, + 0.006463209167122841, + 0.0064489273354411125, + 0.019256459549069405, + 0.0119132986292243, + 0.013838337734341621, + 0.007365523371845484, + 0.026036646217107773, + 0.03625291585922241, + 0.011888762935996056, + 0.01901574246585369, + 0.018543506041169167, + 0.021290361881256104, + 0.013472035527229309, + 0.013003847561776638, + 0.014255065470933914, + 0.010101471096277237, + 0.010504834353923798, + 0.010693994350731373, + 0.016208453103899956, + 0.00835308525711298, + 0.01232825592160225, + 0.012619350105524063, + 0.01689104735851288, + 0.012609046883881092, + 0.011216521263122559, + 0.018760766834020615, + 0.014422067441046238, + 0.01686694845557213, + 0.02121683582663536, + 0.025038056075572968, + 0.012892581522464752, + 0.016483470797538757, + 0.014650958590209484, + 0.017202695831656456, + 0.024594919756054878, + 0.022430645301938057, + 0.016109228134155273, + 0.016981346532702446, + 0.0047656474635005, + 0.004810766316950321, + 0.0051696402952075005, + 0.006380137987434864, + 0.005729797296226025, + 0.0074654221534729, + 0.0063650943338871, + 0.005666612647473812, + 0.013035967014729977, + 0.007754239719361067, + 0.005366142839193344, + 0.0062239402905106544, + 0.009783645160496235, + 0.00853915885090828, + 0.009064888581633568, + 0.018064092844724655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.10270035266876221 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023290636017918587 + ], + "opposite_logprob_tokens": [ + -3.7732906341552734 + ], + "logprob_reward": -0.023290636017918587, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290636017918587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15943872928619385, + 0.07344747334718704, + 0.03211443871259689, + 0.06092827767133713, + 0.06375772505998611, + 0.08889400959014893, + 0.12271062284708023, + 0.0947994589805603, + 0.05020572617650032, + 0.2537035644054413 + ], + "credit_map": [ + 0.014347054064273834, + 0.019559361040592194, + 0.026734303683042526, + 0.023492304608225822, + 0.022699696943163872, + 0.017422573640942574, + 0.019722316414117813, + 0.015461118891835213, + 0.007058995310217142, + 0.007787957321852446, + 0.009947791695594788, + 0.009839676320552826, + 0.012408106587827206, + 0.00906579103320837, + 0.010319559834897518, + 0.007019592449069023, + 0.0038959356024861336, + 0.0044334158301353455, + 0.0038495382759720087, + 0.0035474232863634825, + 0.003914162516593933, + 0.0047830697149038315, + 0.003874995745718479, + 0.0038158935494720936, + 0.0069771092385053635, + 0.008390195667743683, + 0.00903051532804966, + 0.00608564680442214, + 0.007819725200533867, + 0.008616359904408455, + 0.008431698195636272, + 0.005577032454311848, + 0.009169034659862518, + 0.010204709134995937, + 0.007014519069343805, + 0.008219163864850998, + 0.007414040621370077, + 0.010115026496350765, + 0.005276453215628862, + 0.006344778463244438, + 0.013301738537847996, + 0.008962662890553474, + 0.007450105622410774, + 0.0090279346331954, + 0.018341558054089546, + 0.009381459094583988, + 0.008854812942445278, + 0.013573738746345043, + 0.015619765967130661, + 0.010807652957737446, + 0.010896336287260056, + 0.01831139251589775, + 0.012430264614522457, + 0.015430864877998829, + 0.018147608265280724, + 0.021066734567284584, + 0.010224108584225178, + 0.013755976222455502, + 0.009671367704868317, + 0.012047432363033295, + 0.017712606117129326, + 0.011452740989625454, + 0.009843969717621803, + 0.010091261938214302, + 0.007129062432795763, + 0.005251321010291576, + 0.005170353222638369, + 0.005536200478672981, + 0.005690307356417179, + 0.006558532360941172, + 0.005165220703929663, + 0.009704729542136192, + 0.035809118300676346, + 0.02602088823914528, + 0.0340866893529892, + 0.02481691539287567, + 0.02381044067442417, + 0.02264532260596752, + 0.020747877657413483, + 0.06576632708311081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.023290636017918587 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945378448814154 + ], + "opposite_logprob_tokens": [ + -5.130945205688477 + ], + "logprob_reward": -0.005945378448814154, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945378448814154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14121802151203156, + 0.09569432586431503, + 0.06671611219644547, + 0.21488118171691895, + 0.26994919776916504, + 0.034701116383075714, + 0.04423907399177551, + 0.06821693480014801, + 0.026463927701115608, + 0.03792014345526695 + ], + "credit_map": [ + 0.019953271374106407, + 0.021296197548508644, + 0.018574953079223633, + 0.021716320887207985, + 0.01882893778383732, + 0.012519453652203083, + 0.014069478027522564, + 0.014259402640163898, + 0.011077246628701687, + 0.019182303920388222, + 0.01575491949915886, + 0.017052538692951202, + 0.008623589761555195, + 0.007289470173418522, + 0.008212079294025898, + 0.008502175100147724, + 0.012025001458823681, + 0.006067042704671621, + 0.019758589565753937, + 0.009929078631103039, + 0.006147132255136967, + 0.0037973416037857533, + 0.003791524562984705, + 0.00520040187984705, + 0.028004644438624382, + 0.04243645444512367, + 0.049424633383750916, + 0.030226899310946465, + 0.018212996423244476, + 0.015266974456608295, + 0.015114502981305122, + 0.016194073483347893, + 0.061029788106679916, + 0.07994271069765091, + 0.020852867513895035, + 0.023718159645795822, + 0.01929544098675251, + 0.02568165212869644, + 0.014688627794384956, + 0.024739962071180344, + 0.0050423843786120415, + 0.003879575291648507, + 0.003606592770665884, + 0.004474249668419361, + 0.004531115293502808, + 0.0030640908516943455, + 0.0038981852121651173, + 0.006204924080520868, + 0.005970243364572525, + 0.004240117967128754, + 0.0046269241720438, + 0.006787868682295084, + 0.004349694587290287, + 0.005579805932939053, + 0.004561389796435833, + 0.008123030886054039, + 0.008800430223345757, + 0.012069565244019032, + 0.006707813590764999, + 0.00879648420959711, + 0.0116915637627244, + 0.005950077436864376, + 0.0074923173524439335, + 0.006708681117743254, + 0.004111959133297205, + 0.002876392798498273, + 0.003144495654851198, + 0.003554798662662506, + 0.002831311197951436, + 0.0035119110252708197, + 0.0022699444089084864, + 0.0041631171479821205, + 0.004640572238713503, + 0.002434610156342387, + 0.00902401190251112, + 0.004995075520128012, + 0.004790851846337318, + 0.002965400693938136, + 0.003430754179134965, + 0.005638865754008293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.005945378448814154 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782706387341022 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.009782706387341022, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782706387341022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0971846953034401, + 0.27709752321243286, + 0.11989840120077133, + 0.07439420372247696, + 0.07054070383310318, + 0.04014209657907486, + 0.0512474849820137, + 0.14490512013435364, + 0.05743618682026863, + 0.06715355813503265 + ], + "credit_map": [ + 0.010108159855008125, + 0.013177023269236088, + 0.013577657751739025, + 0.014120671898126602, + 0.01110103540122509, + 0.013437562622129917, + 0.010069175623357296, + 0.011593404226005077, + 0.02977687120437622, + 0.05283819139003754, + 0.04162346571683884, + 0.04043489322066307, + 0.029915181919932365, + 0.024795986711978912, + 0.02907918021082878, + 0.028633764013648033, + 0.02253059297800064, + 0.018711084499955177, + 0.025041481480002403, + 0.01948271133005619, + 0.008671173825860023, + 0.007082080468535423, + 0.006582946982234716, + 0.011796334758400917, + 0.012724504806101322, + 0.01394963450729847, + 0.005432452540844679, + 0.0061548734083771706, + 0.00667598657310009, + 0.01231299713253975, + 0.009718388319015503, + 0.0074253687635064125, + 0.0107658002525568, + 0.008788819424808025, + 0.011898554861545563, + 0.007643934804946184, + 0.0050587342120707035, + 0.00938729103654623, + 0.006137940101325512, + 0.010859633795917034, + 0.007335379254072905, + 0.004536173772066832, + 0.004188759718090296, + 0.004431021399796009, + 0.0072655766271054745, + 0.003338028211146593, + 0.0040479255840182304, + 0.004999230150133371, + 0.007252030074596405, + 0.0045364671386778355, + 0.004261698108166456, + 0.005524912383407354, + 0.006131753325462341, + 0.008054845035076141, + 0.007493204902857542, + 0.007992574945092201, + 0.016880357638001442, + 0.025126051157712936, + 0.016247786581516266, + 0.019247109070420265, + 0.02482214756309986, + 0.013160067610442638, + 0.013805774040520191, + 0.015615829266607761, + 0.008596288040280342, + 0.006083257496356964, + 0.00635506771504879, + 0.005763585679233074, + 0.005768767558038235, + 0.008818699046969414, + 0.005385064519941807, + 0.010665456764400005, + 0.013614446856081486, + 0.003812872339040041, + 0.010249301791191101, + 0.006113314535468817, + 0.0042795720510184765, + 0.005089864134788513, + 0.0038414790760725737, + 0.020152708515524864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 2 + ], + "selected_reward": -0.009782706387341022 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general wearing brilliant silver armor with crimson plumes on his helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general riding a muscular white steed that is matted with dust and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate ancient battlefield littered with broken flags and abandoned weapons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with fiery orange and deep purple colors due to a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a dynamic low-angle tracking shot with a horizontal lens flare sweeping across the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895496487617493 + ], + "opposite_logprob_tokens": [ + -2.5789549350738525 + ], + "logprob_reward": -0.07895496487617493, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895496487617493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2438320517539978, + 0.17076756060123444, + 0.09066895395517349, + 0.174587681889534, + 0.07436548918485641, + 0.07987306267023087, + 0.01799812540411949, + 0.07342702895402908, + 0.04107975959777832, + 0.033400289714336395 + ], + "credit_map": [ + 0.018402311950922012, + 0.02773083932697773, + 0.02594134397804737, + 0.02354835532605648, + 0.03433988243341446, + 0.04269596189260483, + 0.04197804629802704, + 0.029195299372076988, + 0.018660763278603554, + 0.018681485205888748, + 0.018932433798909187, + 0.0172549057751894, + 0.03409896045923233, + 0.025988811627030373, + 0.02128235064446926, + 0.015867844223976135, + 0.020032282918691635, + 0.01760486513376236, + 0.008279815316200256, + 0.007690082769840956, + 0.010751289315521717, + 0.008850425481796265, + 0.010648832656443119, + 0.0068113552406430244, + 0.035747379064559937, + 0.040278371423482895, + 0.008875385858118534, + 0.009837768040597439, + 0.02588915266096592, + 0.02524528279900551, + 0.01674433797597885, + 0.011969982646405697, + 0.015766238793730736, + 0.020408235490322113, + 0.008071711286902428, + 0.00551173510029912, + 0.00627492181956768, + 0.006987215951085091, + 0.005959618836641312, + 0.005385808181017637, + 0.01000714860856533, + 0.007407812867313623, + 0.007452876307070255, + 0.007141343783587217, + 0.018342893570661545, + 0.008671890944242477, + 0.010524298064410686, + 0.010324789211153984, + 0.00194073049351573, + 0.0021547311916947365, + 0.001400901353918016, + 0.0015794679056853056, + 0.003132164478302002, + 0.002342090243473649, + 0.002972337417304516, + 0.0024757010396569967, + 0.012102052569389343, + 0.007977415807545185, + 0.013329414650797844, + 0.008993159048259258, + 0.010172970592975616, + 0.007051760330796242, + 0.005989380180835724, + 0.007810874376446009, + 0.0061737485229969025, + 0.003966410178691149, + 0.00482641626149416, + 0.004771654028445482, + 0.0039888364262878895, + 0.005878305062651634, + 0.003668746678158641, + 0.007805635686963797, + 0.003337481524795294, + 0.0022919378243386745, + 0.007161612156778574, + 0.0033605536445975304, + 0.0031310850754380226, + 0.0033964626491069794, + 0.003109774086624384, + 0.007611381355673075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.07895496487617493 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_6/index_tensor([4856, 4856])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_6/index_tensor([4856, 4856])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e8b2307e46856ea8b14d14049835320b64ab91e3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_6/index_tensor([4856, 4856])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level close-up captures two rock pigeons standing side by side at the edge of a shallow, clear puddle on a damp, gray cobblestone street. The bird on the left, with its classic blue-gray plumage and iridescent green and purple shimmering across its neck, captures the soft, diffuse light of an overcast sky. The bird on the right is slightly darker, with more pronounced black markings on its wings. In a swift, almost mechanical, and perfectly synchronized motion, both birds simultaneously dip their dark gray beaks into the water. Then, swiftly raising their heads, they jerk back and forth erratically in a staccato rhythm. This sharp, spastic movement ejects a cloud of shimmering droplets into the air, which, like scattered diamonds, hover against a silent background for a microsecond. The repetition of the action creates a vibrant and slightly comical spectacle of nature, the splashing water disturbing the puddle's glassy surface and creating concentric ripples that radiate outward, all contributing to the tranquil yet lively atmosphere of a quiet urban moment.", + "seed": 15263, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_6/index_4856_seed_15263_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03840513490140438, + "vlm_logprob": -0.03840513490140438, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.03840513490140438, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02656944841146469 + ], + "opposite_logprob_tokens": [ + -3.901569366455078 + ], + "logprob_reward": -0.02656944841146469, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02656944841146469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10653135925531387, + 0.07352404296398163, + 0.053928207606077194, + 0.09764841198921204, + 0.09467200189828873, + 0.08449292927980423, + 0.033496566116809845, + 0.044151876121759415, + 0.2384452521800995, + 0.17310935258865356 + ], + "credit_map": [ + 0.011879711411893368, + 0.014000781811773777, + 0.012508183717727661, + 0.013446073047816753, + 0.017299793660640717, + 0.01713007502257824, + 0.009069453924894333, + 0.011197289451956749, + 0.014977332204580307, + 0.009298969060182571, + 0.010414824821054935, + 0.006900502368807793, + 0.010518544353544712, + 0.01113829668611288, + 0.0046837409026920795, + 0.005591835360974073, + 0.006411283276975155, + 0.007090285886079073, + 0.00644787447527051, + 0.0066698165610432625, + 0.008490273728966713, + 0.007422618567943573, + 0.0054637352004647255, + 0.005932319443672895, + 0.012286215089261532, + 0.01105496846139431, + 0.011755132116377354, + 0.011797678656876087, + 0.01185102853924036, + 0.01287741307169199, + 0.011196289211511612, + 0.014829679392278194, + 0.01412439439445734, + 0.011396030895411968, + 0.012840082868933678, + 0.01622282713651657, + 0.010607765056192875, + 0.011318433098495007, + 0.007330749649554491, + 0.01083171647042036, + 0.010879884473979473, + 0.012553269043564796, + 0.011538404040038586, + 0.016144566237926483, + 0.009497934952378273, + 0.010306213051080704, + 0.005956320092082024, + 0.007616342511028051, + 0.0041317385621368885, + 0.0044650668278336525, + 0.0033685441594570875, + 0.0036412416957318783, + 0.004181989934295416, + 0.004064455628395081, + 0.004862669389694929, + 0.004780860617756844, + 0.005971031729131937, + 0.007866019383072853, + 0.005746576469391584, + 0.007393908221274614, + 0.005829669069498777, + 0.0043837884441018105, + 0.0038924128748476505, + 0.003068470163270831, + 0.02707946114242077, + 0.042165059596300125, + 0.03283504396677017, + 0.03355611488223076, + 0.02633158676326275, + 0.034664276987314224, + 0.022031467407941818, + 0.019782237708568573, + 0.013827521353960037, + 0.025013422593474388, + 0.019441217184066772, + 0.020753061398863792, + 0.02775133028626442, + 0.018986016511917114, + 0.024470418691635132, + 0.02286636270582676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.02656944841146469 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026371043175458908 + ], + "opposite_logprob_tokens": [ + -3.6513710021972656 + ], + "logprob_reward": -0.026371043175458908, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026371043175458908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21656465530395508, + 0.07503969222307205, + 0.19992408156394958, + 0.12655825912952423, + 0.07672717422246933, + 0.07280221581459045, + 0.04362912476062775, + 0.04363721236586571, + 0.08941501379013062, + 0.0557025708258152 + ], + "credit_map": [ + 0.024430163204669952, + 0.02683839201927185, + 0.026136768981814384, + 0.03477364405989647, + 0.03187423199415207, + 0.02912626415491104, + 0.01851116493344307, + 0.024874018505215645, + 0.01041597779840231, + 0.009147089906036854, + 0.009821181185543537, + 0.005833334289491177, + 0.01272224448621273, + 0.013895324431359768, + 0.00596638023853302, + 0.007238159421831369, + 0.048894744366407394, + 0.021328633651137352, + 0.0313059501349926, + 0.026185331866145134, + 0.020702512934803963, + 0.01565215177834034, + 0.011954765766859055, + 0.023899998515844345, + 0.017967678606510162, + 0.013658869080245495, + 0.020479237660765648, + 0.028435274958610535, + 0.013630149886012077, + 0.01038238313049078, + 0.010776454582810402, + 0.011228207498788834, + 0.010192273184657097, + 0.007485455833375454, + 0.012228601612150669, + 0.020562365651130676, + 0.006481385324150324, + 0.007432140409946442, + 0.006232664920389652, + 0.006112285424023867, + 0.015016208402812481, + 0.010802938602864742, + 0.0109037384390831, + 0.012556913308799267, + 0.006946236360818148, + 0.007228143513202667, + 0.0033636458683758974, + 0.005984391085803509, + 0.007623812649399042, + 0.006120109464973211, + 0.005542587023228407, + 0.004236963577568531, + 0.004212019965052605, + 0.005528731271624565, + 0.003138985252007842, + 0.007225915789604187, + 0.0054033976048231125, + 0.0064511350356042385, + 0.005552125629037619, + 0.0071123759262263775, + 0.0046233125030994415, + 0.005051618907600641, + 0.0038147212471812963, + 0.005628525745123625, + 0.010658174753189087, + 0.014750327914953232, + 0.012495598755776882, + 0.012681764550507069, + 0.010895341634750366, + 0.011489765718579292, + 0.008534364402294159, + 0.007909676060080528, + 0.005883168429136276, + 0.008084475994110107, + 0.007301656063646078, + 0.011894158087670803, + 0.005806044675409794, + 0.006798189133405685, + 0.0043940469622612, + 0.005540832877159119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.026371043175458908 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058090642094612 + ], + "opposite_logprob_tokens": [ + -3.895581007003784 + ], + "logprob_reward": -0.02058090642094612, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058090642094612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11516416817903519, + 0.09650233387947083, + 0.051454007625579834, + 0.08166589587926865, + 0.06672462821006775, + 0.05493296682834625, + 0.17702776193618774, + 0.0878763273358345, + 0.0892895758152008, + 0.17936235666275024 + ], + "credit_map": [ + 0.015766862779855728, + 0.015759358182549477, + 0.01606924831867218, + 0.017449546605348587, + 0.015245224349200726, + 0.013436837121844292, + 0.009483618661761284, + 0.011953461915254593, + 0.015367349609732628, + 0.011023510247468948, + 0.012959834188222885, + 0.007512967102229595, + 0.01518243458122015, + 0.018994661048054695, + 0.0065421778708696365, + 0.00891939178109169, + 0.006364360451698303, + 0.00544585008174181, + 0.006526100914925337, + 0.005615161266177893, + 0.00719081424176693, + 0.00684269517660141, + 0.005477361846715212, + 0.007991663180291653, + 0.007293027359992266, + 0.009798813611268997, + 0.01063563209027052, + 0.016637789085507393, + 0.013571006245911121, + 0.007771211676299572, + 0.008488248102366924, + 0.007470152340829372, + 0.006645725108683109, + 0.00913973804563284, + 0.007868229411542416, + 0.010262844152748585, + 0.008729550056159496, + 0.01037613395601511, + 0.007498821243643761, + 0.006203579716384411, + 0.00646193977445364, + 0.007428303826600313, + 0.006595651153475046, + 0.00933812279254198, + 0.0065726544708013535, + 0.008248457685112953, + 0.004102594684809446, + 0.006185237783938646, + 0.023064173758029938, + 0.029410306364297867, + 0.02202175371348858, + 0.014547799713909626, + 0.022552886977791786, + 0.025374481454491615, + 0.017986176535487175, + 0.022070180624723434, + 0.010364638641476631, + 0.012265222147107124, + 0.010934999212622643, + 0.011778806336224079, + 0.010592743754386902, + 0.011159887537360191, + 0.009763981215655804, + 0.011016043834388256, + 0.01157798245549202, + 0.01722520776093006, + 0.012272256426513195, + 0.01208230759948492, + 0.01003752090036869, + 0.01043388620018959, + 0.008577964268624783, + 0.007082452066242695, + 0.018998771905899048, + 0.030111538246273994, + 0.021917326375842094, + 0.01987868919968605, + 0.03144882246851921, + 0.022422194480895996, + 0.02259228378534317, + 0.01199272833764553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.02058090642094612 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195031210780144 + ], + "opposite_logprob_tokens": [ + -4.018195152282715 + ], + "logprob_reward": -0.018195031210780144, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195031210780144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1185012087225914, + 0.06324513256549835, + 0.059404972940683365, + 0.06065766513347626, + 0.11648828536272049, + 0.07608599215745926, + 0.10821786522865295, + 0.06713566929101944, + 0.13904787600040436, + 0.19121529161930084 + ], + "credit_map": [ + 0.013740255497395992, + 0.016474027186632156, + 0.015600220300257206, + 0.016220802441239357, + 0.019261559471488, + 0.01850937493145466, + 0.008433526381850243, + 0.01026144064962864, + 0.012776199728250504, + 0.007800472434610128, + 0.008898002095520496, + 0.007019047159701586, + 0.009807329624891281, + 0.007155937608331442, + 0.004061636049300432, + 0.005726511124521494, + 0.010269282385706902, + 0.007233596406877041, + 0.007849248126149178, + 0.008159926161170006, + 0.0073034497909247875, + 0.006279720459133387, + 0.005566962994635105, + 0.006742788013070822, + 0.008430279791355133, + 0.008901665918529034, + 0.007462495006620884, + 0.009504000656306744, + 0.009331690147519112, + 0.006179911084473133, + 0.005907381419092417, + 0.004940241575241089, + 0.017254265025258064, + 0.016356416046619415, + 0.015366122126579285, + 0.014032436534762383, + 0.01824268326163292, + 0.01378010306507349, + 0.010116824880242348, + 0.01133942510932684, + 0.008301682770252228, + 0.01057454850524664, + 0.009013650007545948, + 0.011126948520541191, + 0.010706184431910515, + 0.012313210405409336, + 0.006354173645377159, + 0.007695596199482679, + 0.013684929348528385, + 0.01747298613190651, + 0.012906297110021114, + 0.013941925950348377, + 0.012356679886579514, + 0.01404997892677784, + 0.013154116459190845, + 0.010650958865880966, + 0.008557048626244068, + 0.007690254133194685, + 0.00908253900706768, + 0.011022493243217468, + 0.007898321375250816, + 0.006989395711570978, + 0.006437437143176794, + 0.009458182379603386, + 0.016231082379817963, + 0.023621516302227974, + 0.018425961956381798, + 0.019954942166805267, + 0.015580187551677227, + 0.020685918629169464, + 0.013008391484618187, + 0.01153988391160965, + 0.023559289053082466, + 0.031128520146012306, + 0.02339855208992958, + 0.020967844873666763, + 0.03323885798454285, + 0.022647591307759285, + 0.020068276673555374, + 0.01620635949075222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.018195031210780144 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030924528837204 + ], + "opposite_logprob_tokens": [ + -2.350309133529663 + ], + "logprob_reward": -0.10030924528837204, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030924528837204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15573427081108093, + 0.06774698197841644, + 0.06005285680294037, + 0.14266623556613922, + 0.06137744337320328, + 0.09181869029998779, + 0.050736986100673676, + 0.16846096515655518, + 0.07030259072780609, + 0.13110294938087463 + ], + "credit_map": [ + 0.02125994674861431, + 0.020047197118401527, + 0.01894020102918148, + 0.024761242792010307, + 0.01576731726527214, + 0.019323449581861496, + 0.016738660633563995, + 0.018896261230111122, + 0.014650626108050346, + 0.007923085242509842, + 0.009605616331100464, + 0.005713451188057661, + 0.010261870920658112, + 0.00826533604413271, + 0.005024695303291082, + 0.006302302703261375, + 0.008334685117006302, + 0.0057281446643173695, + 0.008708403445780277, + 0.007041725795716047, + 0.007261188235133886, + 0.007723276503384113, + 0.005080074537545443, + 0.010175359435379505, + 0.017913099378347397, + 0.01210461650043726, + 0.025323841720819473, + 0.03949413821101189, + 0.008988380432128906, + 0.00924909021705389, + 0.012194118462502956, + 0.017398962751030922, + 0.00885904673486948, + 0.008979855105280876, + 0.008410213515162468, + 0.00902768224477768, + 0.007766044698655605, + 0.006559999193996191, + 0.004966767504811287, + 0.00680783623829484, + 0.011066571809351444, + 0.012039882130920887, + 0.012693343684077263, + 0.020521022379398346, + 0.009817576967179775, + 0.01056758500635624, + 0.00539309112355113, + 0.009719622321426868, + 0.010396573692560196, + 0.006758018396794796, + 0.0072926972061395645, + 0.004938601516187191, + 0.004590170457959175, + 0.004666931461542845, + 0.005093972198665142, + 0.007000023499131203, + 0.023424142971634865, + 0.019864322617650032, + 0.020689265802502632, + 0.030029604211449623, + 0.015098465606570244, + 0.01667964644730091, + 0.014541289769113064, + 0.02813422866165638, + 0.00804592203348875, + 0.01005700696259737, + 0.01049997191876173, + 0.013467608019709587, + 0.006651550997048616, + 0.007199979852885008, + 0.006354425102472305, + 0.00802612490952015, + 0.01369135919958353, + 0.021536028012633324, + 0.016013750806450844, + 0.02217947505414486, + 0.0185124222189188, + 0.01188821904361248, + 0.012954948469996452, + 0.014326754957437515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.10030924528837204 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures two rock pigeons standing side by side at the edge of a shallow, clear puddle on a damp, gray cobblestone street. The bird on the left, with its classic blue-gray plumage and iridescent green and purple shimmering across its neck, captures the soft, diffuse light of an overcast sky. The bird on the right is slightly darker, with more pronounced black markings on its wings. In a swift, almost mechanical, and perfectly synchronized motion, both birds simultaneously dip their dark gray beaks into the water. Then, swiftly raising their heads, they jerk back and forth erratically in a staccato rhythm. This sharp, spastic movement ejects a cloud of shimmering droplets into the air, which, like scattered diamonds, hover against a silent background for a microsecond. The repetition of the action creates a vibrant and slightly comical spectacle of nature, the splashing water disturbing the puddle's glassy surface and creating concentric ripples that radiate outward, all contributing to the tranquil yet lively atmosphere of a quiet urban moment.", + "seed": 15264, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_6/index_4856_seed_15264_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.051143769174814224, + "vlm_logprob": -0.051143769174814224, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.051143769174814224, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03743966296315193 + ], + "opposite_logprob_tokens": [ + -3.5374395847320557 + ], + "logprob_reward": -0.03743966296315193, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03743966296315193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12712334096431732, + 0.12264414876699448, + 0.05200241878628731, + 0.07617951184511185, + 0.27735450863838196, + 0.08175135403871536, + 0.10271301120519638, + 0.055213768035173416, + 0.06295431405305862, + 0.04206360504031181 + ], + "credit_map": [ + 0.027992529794573784, + 0.0217252466827631, + 0.01776760071516037, + 0.013776872307062149, + 0.013163182884454727, + 0.014504942111670971, + 0.008111114613711834, + 0.010081849992275238, + 0.03203727304935455, + 0.012064618989825249, + 0.01590149477124214, + 0.00676773814484477, + 0.012341856025159359, + 0.026690782979130745, + 0.010189173743128777, + 0.006651211995631456, + 0.009513218887150288, + 0.007488278206437826, + 0.009138345718383789, + 0.0057533434592187405, + 0.0051134913228452206, + 0.004188107326626778, + 0.006130088120698929, + 0.004677546210587025, + 0.014427055604755878, + 0.008509409613907337, + 0.009978671558201313, + 0.006959115620702505, + 0.0055665746331214905, + 0.01230591256171465, + 0.009315171279013157, + 0.00911760050803423, + 0.04180242866277695, + 0.045822639018297195, + 0.02862934209406376, + 0.02036314830183983, + 0.054072920233011246, + 0.03707287833094597, + 0.024679034948349, + 0.02491212822496891, + 0.011925862170755863, + 0.010655817575752735, + 0.012836502864956856, + 0.007249511778354645, + 0.017634283751249313, + 0.012064321897923946, + 0.004210257902741432, + 0.005174793303012848, + 0.013975017704069614, + 0.015923844650387764, + 0.009938156232237816, + 0.013292912393808365, + 0.014183208346366882, + 0.017128391191363335, + 0.008937510661780834, + 0.009333975613117218, + 0.011355815455317497, + 0.0061180479824543, + 0.006715595256537199, + 0.0062975650653243065, + 0.009290606714785099, + 0.006790101528167725, + 0.0036823174450546503, + 0.004963716492056847, + 0.007209407165646553, + 0.013635881245136261, + 0.005566707346588373, + 0.005233071278780699, + 0.010028349235653877, + 0.010091237723827362, + 0.004222423769533634, + 0.006967239081859589, + 0.00632456224411726, + 0.004563296213746071, + 0.004949968308210373, + 0.0032137956004589796, + 0.007206371985375881, + 0.007919271476566792, + 0.004117797128856182, + 0.003768542082980275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.03743966296315193 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07009172439575195 + ], + "opposite_logprob_tokens": [ + -2.695091724395752 + ], + "logprob_reward": -0.07009172439575195, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07009172439575195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2269357591867447, + 0.3279793858528137, + 0.049537062644958496, + 0.05478715896606445, + 0.06583768129348755, + 0.040203750133514404, + 0.05809909850358963, + 0.08188284933567047, + 0.06386083364486694, + 0.03087642416357994 + ], + "credit_map": [ + 0.060321152210235596, + 0.04270143061876297, + 0.036696501076221466, + 0.02060241438448429, + 0.014355815947055817, + 0.028602879494428635, + 0.009506131522357464, + 0.014149431139230728, + 0.09416425228118896, + 0.023197116330266, + 0.05155978351831436, + 0.011274788528680801, + 0.02728591114282608, + 0.07717970758676529, + 0.026557398959994316, + 0.016760429367423058, + 0.015671931207180023, + 0.005706251598894596, + 0.013249375857412815, + 0.003817748511210084, + 0.002815903862938285, + 0.002715809503570199, + 0.003075306536629796, + 0.0024847318418323994, + 0.00822548195719719, + 0.005287814885377884, + 0.007464616559445858, + 0.005879946053028107, + 0.004760851617902517, + 0.008421639911830425, + 0.008686131797730923, + 0.006060679443180561, + 0.008427953347563744, + 0.012031066231429577, + 0.006603253073990345, + 0.005470376927405596, + 0.013839471153914928, + 0.008418700657784939, + 0.005985175725072622, + 0.005061686504632235, + 0.010173710063099861, + 0.004485230892896652, + 0.0064740898087620735, + 0.0022530448623001575, + 0.005256696604192257, + 0.006862583104521036, + 0.002002477878704667, + 0.0026959162205457687, + 0.007806582376360893, + 0.009691717103123665, + 0.005891500040888786, + 0.008129237219691277, + 0.007197197992354631, + 0.008122043684124947, + 0.0057380604557693005, + 0.005522755905985832, + 0.02134559489786625, + 0.005689694546163082, + 0.01130943838506937, + 0.009854298084974289, + 0.013460979796946049, + 0.008922118693590164, + 0.0056513212621212006, + 0.005649395752698183, + 0.007585907820612192, + 0.015398062765598297, + 0.006345126777887344, + 0.005529507528990507, + 0.006778670474886894, + 0.011283992789685726, + 0.004128138534724712, + 0.0068114278838038445, + 0.0038565522991120815, + 0.0031072120182216167, + 0.003806204302236438, + 0.0025758356787264347, + 0.005988375283777714, + 0.0058491649106144905, + 0.0026295878924429417, + 0.003063490614295006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.07009172439575195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204615533351898 + ], + "opposite_logprob_tokens": [ + -2.8120460510253906 + ], + "logprob_reward": -0.06204615533351898, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204615533351898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0655016377568245, + 0.38748955726623535, + 0.05310114100575447, + 0.12099390476942062, + 0.08338339626789093, + 0.11108270287513733, + 0.07792723178863525, + 0.03804711252450943, + 0.035594917833805084, + 0.0268784761428833 + ], + "credit_map": [ + 0.015944788232445717, + 0.012013091705739498, + 0.008905861526727676, + 0.0049520218744874, + 0.006005556788295507, + 0.008772523142397404, + 0.0033521761652082205, + 0.005555618554353714, + 0.11456333845853806, + 0.027174077928066254, + 0.059006575495004654, + 0.012589126825332642, + 0.0318559966981411, + 0.09286393970251083, + 0.030448034405708313, + 0.018988467752933502, + 0.015076285228133202, + 0.0067867194302380085, + 0.012692319229245186, + 0.00371252978220582, + 0.003884043777361512, + 0.0036133206449449062, + 0.004373383242636919, + 0.0029625410679727793, + 0.01646399311721325, + 0.01568671315908432, + 0.012948095798492432, + 0.016380611807107925, + 0.012059276923537254, + 0.013309787027537823, + 0.018162474036216736, + 0.015982942655682564, + 0.014421725645661354, + 0.014484460465610027, + 0.009396555833518505, + 0.005640454590320587, + 0.015519152395427227, + 0.010104070417582989, + 0.006966304033994675, + 0.006850669626146555, + 0.031990986317396164, + 0.01168047171086073, + 0.02155889943242073, + 0.005426052957773209, + 0.012704375199973583, + 0.016970157623291016, + 0.004635101184248924, + 0.006116649601608515, + 0.010482772253453732, + 0.011818760074675083, + 0.007312051486223936, + 0.009438063018023968, + 0.013376598246395588, + 0.00999569334089756, + 0.0076224165968596935, + 0.007880877703428268, + 0.007041471544653177, + 0.004197878763079643, + 0.004440969321876764, + 0.0031745124142616987, + 0.008171298541128635, + 0.0038398532196879387, + 0.0036120396107435226, + 0.003569090273231268, + 0.004469241481274366, + 0.007507681380957365, + 0.003545810002833605, + 0.003659163834527135, + 0.005042839329689741, + 0.006426411215215921, + 0.0018489969661459327, + 0.0030947737395763397, + 0.0022466653026640415, + 0.002922085579484701, + 0.0029992293566465378, + 0.004163472913205624, + 0.003739862237125635, + 0.0035692264791578054, + 0.0032553854398429394, + 0.003982550464570522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.06204615533351898 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307710379362106 + ], + "opposite_logprob_tokens": [ + -3.168076992034912 + ], + "logprob_reward": -0.04307710379362106, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307710379362106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16531918942928314, + 0.23759646713733673, + 0.04739298298954964, + 0.08351774513721466, + 0.25158271193504333, + 0.07061309367418289, + 0.03469821438193321, + 0.030495265498757362, + 0.04420933499932289, + 0.0345749594271183 + ], + "credit_map": [ + 0.03915274888277054, + 0.032055262476205826, + 0.022981027141213417, + 0.015652509406208992, + 0.012958286330103874, + 0.022101495414972305, + 0.008628062903881073, + 0.01178978756070137, + 0.06758293509483337, + 0.018582919612526894, + 0.035733092576265335, + 0.008187779225409031, + 0.022209851071238518, + 0.056500013917684555, + 0.017221838235855103, + 0.011578027158975601, + 0.011843469925224781, + 0.005591524764895439, + 0.01139871682971716, + 0.003584192832931876, + 0.0038614352233707905, + 0.0027413105126470327, + 0.005057521630078554, + 0.003314811736345291, + 0.01321420818567276, + 0.007879802957177162, + 0.010088533163070679, + 0.008385281078517437, + 0.00577398668974638, + 0.011078871786594391, + 0.011754469946026802, + 0.01534259133040905, + 0.04077931493520737, + 0.04029525816440582, + 0.026967421174049377, + 0.016462212428450584, + 0.049210429191589355, + 0.035548239946365356, + 0.022540291771292686, + 0.019779562950134277, + 0.015640446916222572, + 0.009379119612276554, + 0.008608205243945122, + 0.005825243424624205, + 0.012930464930832386, + 0.009289821609854698, + 0.004143041558563709, + 0.00479674618691206, + 0.004118358716368675, + 0.005270032212138176, + 0.003660056972876191, + 0.00463912496343255, + 0.006287140306085348, + 0.00394123001024127, + 0.003375907428562641, + 0.0034063654020428658, + 0.0045861490070819855, + 0.005713538732379675, + 0.003334041917696595, + 0.003097603330388665, + 0.0036223307251930237, + 0.0030813170596957207, + 0.003295817179605365, + 0.003764469176530838, + 0.005531576462090015, + 0.008556663990020752, + 0.004546046257019043, + 0.003176616970449686, + 0.0067602526396512985, + 0.008000205270946026, + 0.002880917862057686, + 0.004757055081427097, + 0.004332565702497959, + 0.003598709823563695, + 0.0037006495986133814, + 0.0036722635850310326, + 0.004273693542927504, + 0.005958514288067818, + 0.002912064315751195, + 0.006126500200480223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.04307710379362106 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306419938802719 + ], + "opposite_logprob_tokens": [ + -3.1680641174316406 + ], + "logprob_reward": -0.04306419938802719, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306419938802719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041889291256666183, + 0.5892764329910278, + 0.019826864823698997, + 0.06330131739377975, + 0.0446491539478302, + 0.05473984777927399, + 0.03873536363244057, + 0.015181463211774826, + 0.104515939950943, + 0.027884459123015404 + ], + "credit_map": [ + 0.011182926595211029, + 0.007415418978780508, + 0.0053992122411727905, + 0.0034102594945579767, + 0.003998534753918648, + 0.00599849596619606, + 0.0017839507199823856, + 0.0027004852890968323, + 0.17042550444602966, + 0.04158658906817436, + 0.09134392440319061, + 0.01886262744665146, + 0.04901978373527527, + 0.14078612625598907, + 0.04804408550262451, + 0.029207663610577583, + 0.003559124656021595, + 0.0031917146407067776, + 0.0021438300609588623, + 0.0028741727583110332, + 0.0017691892571747303, + 0.0020414693281054497, + 0.0023861557710915804, + 0.0018612046260386705, + 0.01310630701482296, + 0.008472243323922157, + 0.008772073313593864, + 0.0059052384458482265, + 0.006183255463838577, + 0.008447244763374329, + 0.005387237295508385, + 0.0070277052000164986, + 0.006283693481236696, + 0.008339563384652138, + 0.004808447323739529, + 0.0038018713239580393, + 0.008021446876227856, + 0.005414151586592197, + 0.0037983711808919907, + 0.004181601107120514, + 0.016804683953523636, + 0.0050398679450154305, + 0.012248212471604347, + 0.002319351304322481, + 0.005385814234614372, + 0.008228125981986523, + 0.0020754134748131037, + 0.0026383674703538418, + 0.004163466859608889, + 0.007002105005085468, + 0.0035174076911062002, + 0.0054248301312327385, + 0.006893185433000326, + 0.004748196806758642, + 0.0030780283268541098, + 0.003908138256520033, + 0.0023992732167243958, + 0.002080318983644247, + 0.0019292726647108793, + 0.0014084738213568926, + 0.0024661009665578604, + 0.0016309880884364247, + 0.0016440770123153925, + 0.0016229551983997226, + 0.016615035012364388, + 0.010542974807322025, + 0.013391702435910702, + 0.008861382491886616, + 0.00918574444949627, + 0.030082089826464653, + 0.006316264159977436, + 0.009520730935037136, + 0.004240255802869797, + 0.004725173581391573, + 0.0028545462992042303, + 0.0024272373411804438, + 0.0035050269216299057, + 0.004607599228620529, + 0.0018108334625139832, + 0.003713783575221896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.04306419938802719 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures two rock pigeons standing side by side at the edge of a shallow, clear puddle on a damp, gray cobblestone street. The bird on the left, with its classic blue-gray plumage and iridescent green and purple shimmering across its neck, captures the soft, diffuse light of an overcast sky. The bird on the right is slightly darker, with more pronounced black markings on its wings. In a swift, almost mechanical, and perfectly synchronized motion, both birds simultaneously dip their dark gray beaks into the water. Then, swiftly raising their heads, they jerk back and forth erratically in a staccato rhythm. This sharp, spastic movement ejects a cloud of shimmering droplets into the air, which, like scattered diamonds, hover against a silent background for a microsecond. The repetition of the action creates a vibrant and slightly comical spectacle of nature, the splashing water disturbing the puddle's glassy surface and creating concentric ripples that radiate outward, all contributing to the tranquil yet lively atmosphere of a quiet urban moment.", + "seed": 15265, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_6/index_4856_seed_15265_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04197399746626616, + "vlm_logprob": -0.04197399746626616, + "vlm_yes_no_margin": 3.624999976158142, + "vlm_token_credit": -0.04197399746626616, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012914001010358334 + ], + "opposite_logprob_tokens": [ + -4.887914180755615 + ], + "logprob_reward": -0.012914001010358334, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012914001010358334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0830690786242485, + 0.05043444037437439, + 0.025290250778198242, + 0.2983265519142151, + 0.08236004412174225, + 0.04429015517234802, + 0.04666575416922569, + 0.033376745879650116, + 0.03929538279771805, + 0.29689157009124756 + ], + "credit_map": [ + 0.01702312007546425, + 0.0090488251298666, + 0.013052036985754967, + 0.006812763400375843, + 0.0100895706564188, + 0.00870293378829956, + 0.00838195439428091, + 0.009957884438335896, + 0.010017311200499535, + 0.007072520442306995, + 0.005710991565138102, + 0.004441538825631142, + 0.005394706968218088, + 0.008508732542395592, + 0.003618581686168909, + 0.0056700618006289005, + 0.00392451835796237, + 0.003086521290242672, + 0.002354919444769621, + 0.004183778539299965, + 0.0026956009678542614, + 0.003182740416377783, + 0.0027875672094523907, + 0.003074606880545616, + 0.011298805475234985, + 0.011100616306066513, + 0.017217500135302544, + 0.007436835672706366, + 0.037993330508470535, + 0.12699367105960846, + 0.007987785153090954, + 0.07829800993204117, + 0.007641119416803122, + 0.011824232526123524, + 0.01287919469177723, + 0.004842500202357769, + 0.014181451871991158, + 0.018572581931948662, + 0.006822090595960617, + 0.005596871953457594, + 0.008883491158485413, + 0.005925293546169996, + 0.004924194421619177, + 0.0037279841490089893, + 0.009235897101461887, + 0.005721824709326029, + 0.0031547972466796637, + 0.002716674702242017, + 0.007146927993744612, + 0.005692844744771719, + 0.0034667481668293476, + 0.0036757804919034243, + 0.01457933895289898, + 0.007599457632750273, + 0.002465426456183195, + 0.002039234386757016, + 0.00473408168181777, + 0.0048752641305327415, + 0.0043942308984696865, + 0.004275026731193066, + 0.004350688774138689, + 0.004884712398052216, + 0.003180411411449313, + 0.002682330086827278, + 0.008231713436543941, + 0.004262286238372326, + 0.00483242841437459, + 0.002037514466792345, + 0.007812008261680603, + 0.007142679300159216, + 0.002578738145530224, + 0.002398014534264803, + 0.0719432681798935, + 0.011707728728652, + 0.025268975645303726, + 0.009878814220428467, + 0.033827293664216995, + 0.02016136236488819, + 0.11116737872362137, + 0.01293676532804966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.012914001010358334 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023317305371165276 + ], + "opposite_logprob_tokens": [ + -3.773317337036133 + ], + "logprob_reward": -0.023317305371165276, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023317305371165276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1387043595314026, + 0.08251828700304031, + 0.060933686792850494, + 0.4033045768737793, + 0.0266580730676651, + 0.13300301134586334, + 0.034792497754096985, + 0.03628866374492645, + 0.02121911011636257, + 0.0625777617096901 + ], + "credit_map": [ + 0.027792008593678474, + 0.01658855564892292, + 0.02124789170920849, + 0.012556186877191067, + 0.01785995438694954, + 0.017074821516871452, + 0.0114084267988801, + 0.014176495373249054, + 0.01723625510931015, + 0.008632566779851913, + 0.012761957012116909, + 0.007947658188641071, + 0.009631230495870113, + 0.008069985546171665, + 0.006094371899962425, + 0.012144259177148342, + 0.009515059180557728, + 0.006934489589184523, + 0.009291442111134529, + 0.007043951656669378, + 0.0050058513879776, + 0.010209009051322937, + 0.004298597574234009, + 0.008635277859866619, + 0.01747766137123108, + 0.016164645552635193, + 0.050451748073101044, + 0.009137069806456566, + 0.07212088257074356, + 0.1644735485315323, + 0.010630138218402863, + 0.06284885108470917, + 0.0030828702729195356, + 0.005514154210686684, + 0.0024021249264478683, + 0.002023746958002448, + 0.004312904551625252, + 0.003427598625421524, + 0.004070493392646313, + 0.0018241771031171083, + 0.02793697826564312, + 0.01540316827595234, + 0.014543690718710423, + 0.011552558280527592, + 0.027656588703393936, + 0.016028711572289467, + 0.008765965700149536, + 0.011115345172584057, + 0.005348788574337959, + 0.0034805648028850555, + 0.0028158770874142647, + 0.0025146950501948595, + 0.01204853504896164, + 0.005809306167066097, + 0.0014651564415544271, + 0.00130957062356174, + 0.006977180950343609, + 0.004172415006905794, + 0.0038414306472986937, + 0.004152303095906973, + 0.006828548852354288, + 0.005158823449164629, + 0.0023989572655409575, + 0.0027590009849518538, + 0.003615116933360696, + 0.002991813002154231, + 0.0027904068119823933, + 0.001400477602146566, + 0.004465087316930294, + 0.0033124706242233515, + 0.0013362501049414277, + 0.0013074851594865322, + 0.012933975085616112, + 0.00291256932541728, + 0.005219048820436001, + 0.0023437724448740482, + 0.007014920003712177, + 0.005575949791818857, + 0.023638620972633362, + 0.0029389006085693836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.023317305371165276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289471864700317 + ], + "opposite_logprob_tokens": [ + -3.773289442062378 + ], + "logprob_reward": -0.023289471864700317, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289471864700317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020216887816786766, + 0.02327251434326172, + 0.015963386744260788, + 0.6377021074295044, + 0.020867938175797462, + 0.026884548366069794, + 0.04269886761903763, + 0.014993508346378803, + 0.023288337513804436, + 0.17411193251609802 + ], + "credit_map": [ + 0.00422662915661931, + 0.0021966176573187113, + 0.0031735890079289675, + 0.002036287449300289, + 0.002647781977429986, + 0.0020483278203755617, + 0.0020889013539999723, + 0.0017987508326768875, + 0.004292443860322237, + 0.0036393676418811083, + 0.0028007891960442066, + 0.0016189257148653269, + 0.002521735616028309, + 0.004629079718142748, + 0.0016671257326379418, + 0.002103044418618083, + 0.002720830263569951, + 0.002012623008340597, + 0.0014825079124420881, + 0.0018166954396292567, + 0.0013747166376560926, + 0.0031954278238117695, + 0.0015663743251934648, + 0.001794209354557097, + 0.022155726328492165, + 0.014967088587582111, + 0.055755969136953354, + 0.010844332166016102, + 0.08413016051054001, + 0.28643694519996643, + 0.013314997777342796, + 0.1500968188047409, + 0.0022041492629796267, + 0.004026871640235186, + 0.0029757055453956127, + 0.0015637363540008664, + 0.002263482427224517, + 0.0038278168067336082, + 0.001999505329877138, + 0.0020066683646291494, + 0.0037876872811466455, + 0.002766968682408333, + 0.0026340559124946594, + 0.0015407263999804854, + 0.006845691706985235, + 0.003189149545505643, + 0.0024296981282532215, + 0.003690567798912525, + 0.006158201955258846, + 0.002992400899529457, + 0.003955182619392872, + 0.0034002186730504036, + 0.014698321931064129, + 0.007631219457834959, + 0.0019702978897839785, + 0.0018930205842480063, + 0.002124723047018051, + 0.001593838562257588, + 0.001391434809193015, + 0.002431685570627451, + 0.002743413671851158, + 0.0021224492229521275, + 0.0009619618649594486, + 0.0016240011900663376, + 0.00446883263066411, + 0.0026459451764822006, + 0.002156465779989958, + 0.0024018411058932543, + 0.003665769239887595, + 0.0038756420835852623, + 0.00229568500071764, + 0.0017781542846933007, + 0.041631828993558884, + 0.006406181957572699, + 0.014638761058449745, + 0.00542402733117342, + 0.020901750773191452, + 0.007853055372834206, + 0.06985943764448166, + 0.007396856788545847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 6 + ], + "selected_reward": -0.023289471864700317 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023297274485230446 + ], + "opposite_logprob_tokens": [ + -3.7732973098754883 + ], + "logprob_reward": -0.023297274485230446, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297274485230446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022543786093592644, + 0.022099027410149574, + 0.02683481015264988, + 0.7054516077041626, + 0.01563410833477974, + 0.022954124957323074, + 0.021770954132080078, + 0.01783514767885208, + 0.01909216307103634, + 0.12578430771827698 + ], + "credit_map": [ + 0.0034845834597945213, + 0.003077669767662883, + 0.0031282866839319468, + 0.0018648034892976284, + 0.002647387096658349, + 0.0033789759036153555, + 0.001792249851860106, + 0.003169829258695245, + 0.004064980894327164, + 0.0030594938434660435, + 0.002821321366354823, + 0.0017790970159694552, + 0.0022172066383063793, + 0.00428830785676837, + 0.0015965088969096541, + 0.0022721104323863983, + 0.004103511571884155, + 0.0032595207449048758, + 0.003821591380983591, + 0.0024921095464378595, + 0.0024398190435022116, + 0.00580129399895668, + 0.0016336627304553986, + 0.0032833016011863947, + 0.024769453331828117, + 0.017060821875929832, + 0.05972529202699661, + 0.012116114608943462, + 0.09185637533664703, + 0.3162488639354706, + 0.014878044836223125, + 0.16879665851593018, + 0.0021300336811691523, + 0.0035307423677295446, + 0.0017800289206206799, + 0.0012497205752879381, + 0.0018290909938514233, + 0.0020662664901465178, + 0.0020359649788588285, + 0.0010122604435309768, + 0.002684334758669138, + 0.00321066752076149, + 0.0022466243244707584, + 0.0018760627135634422, + 0.005092771258205175, + 0.0032937447540462017, + 0.0018213476287201047, + 0.0027285718824714422, + 0.003071641316637397, + 0.0016845533391460776, + 0.002104210201650858, + 0.0018836800009012222, + 0.007215294986963272, + 0.003657375229522586, + 0.0011029659071937203, + 0.0010512341978028417, + 0.0021422712597995996, + 0.002015786711126566, + 0.0019397088326513767, + 0.0030648852698504925, + 0.0032980453688651323, + 0.0023144937586039305, + 0.0010765448678284883, + 0.0019834127742797136, + 0.004657888785004616, + 0.0023588237818330526, + 0.0019413481932133436, + 0.0013838770100846887, + 0.0026743735652416945, + 0.003915533889085054, + 0.0009705346892587841, + 0.0011897827498614788, + 0.03076678141951561, + 0.004709073808044195, + 0.010562947019934654, + 0.004065127111971378, + 0.014661897905170918, + 0.006367031019181013, + 0.04914103075861931, + 0.00551042752340436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 2 + ], + "selected_reward": -0.023297274485230446 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270519345998764 + ], + "opposite_logprob_tokens": [ + -2.127051830291748 + ], + "logprob_reward": -0.1270519345998764, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270519345998764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06872090697288513, + 0.05238080024719238, + 0.17664839327335358, + 0.2613922655582428, + 0.032838497310876846, + 0.04087290167808533, + 0.030535178259015083, + 0.13120074570178986, + 0.016539646312594414, + 0.18887075781822205 + ], + "credit_map": [ + 0.015673652291297913, + 0.007460654713213444, + 0.011823441833257675, + 0.005148904863744974, + 0.009946661069989204, + 0.006610837299376726, + 0.007597877644002438, + 0.004458873998373747, + 0.008764597587287426, + 0.006312007550150156, + 0.007239200174808502, + 0.005623308476060629, + 0.004918267484754324, + 0.007567350286990404, + 0.003939509391784668, + 0.0080165546387434, + 0.025925874710083008, + 0.02182747982442379, + 0.02942795865237713, + 0.020253930240869522, + 0.014901604503393173, + 0.025169143453240395, + 0.012451684102416039, + 0.026690710335969925, + 0.010428466834127903, + 0.012285025790333748, + 0.011186671443283558, + 0.007567197550088167, + 0.03600294515490532, + 0.10010480135679245, + 0.008563356474041939, + 0.07525377720594406, + 0.0028838380239903927, + 0.005471810698509216, + 0.0029662130400538445, + 0.0032919126097112894, + 0.004227209836244583, + 0.003504982450976968, + 0.007663719356060028, + 0.0028288098983466625, + 0.0056717125698924065, + 0.006239126902073622, + 0.003938720561563969, + 0.004293257836252451, + 0.007276425138115883, + 0.006187499035149813, + 0.003618373768404126, + 0.0036477847024798393, + 0.003960642032325268, + 0.005005514714866877, + 0.004542913753539324, + 0.0031287309248000383, + 0.0068781026639044285, + 0.0039106858894228935, + 0.0014246890787035227, + 0.001683895941823721, + 0.02384527400135994, + 0.013020426966249943, + 0.014254878275096416, + 0.02836170420050621, + 0.017027905210852623, + 0.013627459295094013, + 0.007801464758813381, + 0.013261616230010986, + 0.0028860170859843493, + 0.0018599305767565966, + 0.0013566457200795412, + 0.0014028536388650537, + 0.003373128129169345, + 0.0033978186547756195, + 0.0011582692386582494, + 0.001104980823583901, + 0.03727346658706665, + 0.009197925217449665, + 0.017980845645070076, + 0.00639666523784399, + 0.020858919247984886, + 0.021476946771144867, + 0.06551435589790344, + 0.010171632282435894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 2 + ], + "selected_reward": -0.1270519345998764 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures two rock pigeons standing side by side at the edge of a shallow, clear puddle on a damp, gray cobblestone street. The bird on the left, with its classic blue-gray plumage and iridescent green and purple shimmering across its neck, captures the soft, diffuse light of an overcast sky. The bird on the right is slightly darker, with more pronounced black markings on its wings. In a swift, almost mechanical, and perfectly synchronized motion, both birds simultaneously dip their dark gray beaks into the water. Then, swiftly raising their heads, they jerk back and forth erratically in a staccato rhythm. This sharp, spastic movement ejects a cloud of shimmering droplets into the air, which, like scattered diamonds, hover against a silent background for a microsecond. The repetition of the action creates a vibrant and slightly comical spectacle of nature, the splashing water disturbing the puddle's glassy surface and creating concentric ripples that radiate outward, all contributing to the tranquil yet lively atmosphere of a quiet urban moment.", + "seed": 15266, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_6/index_4856_seed_15266_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03087567500770092, + "vlm_logprob": -0.03087567500770092, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.03087567500770092, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017208080738782883 + ], + "opposite_logprob_tokens": [ + -4.392208099365234 + ], + "logprob_reward": -0.017208080738782883, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.017208080738782883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07738980650901794, + 0.07526933401823044, + 0.060985736548900604, + 0.0470241941511631, + 0.15384866297245026, + 0.192543625831604, + 0.08782531321048737, + 0.12160629034042358, + 0.05438908562064171, + 0.12911787629127502 + ], + "credit_map": [ + 0.012025794014334679, + 0.010035831481218338, + 0.01051498856395483, + 0.008736604824662209, + 0.01024418231099844, + 0.009816489182412624, + 0.007616888266056776, + 0.008399025537073612, + 0.010756024159491062, + 0.009503020904958248, + 0.010868518613278866, + 0.005540756508708, + 0.009753170423209667, + 0.010183694772422314, + 0.007077316753566265, + 0.011586831882596016, + 0.007224950008094311, + 0.006917973980307579, + 0.009999696165323257, + 0.007804321125149727, + 0.006274719722568989, + 0.005717764142900705, + 0.006491024512797594, + 0.010555285029113293, + 0.0073418947868049145, + 0.008357762359082699, + 0.005333944223821163, + 0.005482199601829052, + 0.004849988501518965, + 0.004661163315176964, + 0.004804467316716909, + 0.006192774511873722, + 0.01739271730184555, + 0.013214967213571072, + 0.012979977764189243, + 0.009982115589082241, + 0.01880512200295925, + 0.035227660089731216, + 0.02895074523985386, + 0.017295358702540398, + 0.04898224025964737, + 0.009676617570221424, + 0.006267174147069454, + 0.022114671766757965, + 0.06931470334529877, + 0.010782357305288315, + 0.007704205811023712, + 0.017701664939522743, + 0.02666180022060871, + 0.008946022018790245, + 0.007508707232773304, + 0.018179060891270638, + 0.0072088465094566345, + 0.005995156243443489, + 0.005910313222557306, + 0.0074154045432806015, + 0.041423387825489044, + 0.03767138347029686, + 0.006923368200659752, + 0.005348805338144302, + 0.01140093058347702, + 0.004640727769583464, + 0.007506998721510172, + 0.0066906847059726715, + 0.009016836062073708, + 0.006090780720114708, + 0.005872849840670824, + 0.005975222680717707, + 0.008551460690796375, + 0.0037614593748003244, + 0.005472387187182903, + 0.009648089297115803, + 0.03208969533443451, + 0.010705511085689068, + 0.008108599111437798, + 0.01881064847111702, + 0.01330650970339775, + 0.011670568026602268, + 0.010271097533404827, + 0.02415524795651436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.017208080738782883 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983708307147026 + ], + "opposite_logprob_tokens": [ + -3.529837131500244 + ], + "logprob_reward": -0.02983708307147026, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983708307147026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07290046662092209, + 0.09935685992240906, + 0.05642973259091377, + 0.029125669971108437, + 0.12834569811820984, + 0.08247366547584534, + 0.04249778389930725, + 0.19113226234912872, + 0.08328092098236084, + 0.21445700526237488 + ], + "credit_map": [ + 0.014190380461513996, + 0.010834382846951485, + 0.00864235870540142, + 0.007074299734085798, + 0.011428333818912506, + 0.007780507206916809, + 0.005579287651926279, + 0.007370913401246071, + 0.01922803744673729, + 0.015179425477981567, + 0.008755129761993885, + 0.007565259467810392, + 0.014512835070490837, + 0.011753380298614502, + 0.009831883013248444, + 0.01253090612590313, + 0.008191745728254318, + 0.006786392070353031, + 0.007053605746477842, + 0.008693147450685501, + 0.005071275867521763, + 0.004714768845587969, + 0.004692737013101578, + 0.011226058006286621, + 0.004559327382594347, + 0.004162297584116459, + 0.0031245227437466383, + 0.0026658212300390005, + 0.004313200246542692, + 0.003392511745914817, + 0.002768339589238167, + 0.004139648750424385, + 0.01729693077504635, + 0.012131256982684135, + 0.015281938947737217, + 0.02745448425412178, + 0.014912663958966732, + 0.013460292480885983, + 0.009081925265491009, + 0.018726205453276634, + 0.02703702077269554, + 0.004949796479195356, + 0.003474057884886861, + 0.0047951736487448215, + 0.025837430730462074, + 0.004566604737192392, + 0.003870853688567877, + 0.007942729629576206, + 0.012255697511136532, + 0.0038158458191901445, + 0.003932218533009291, + 0.008975649252533913, + 0.0032018322963267565, + 0.003030517604202032, + 0.0042650639079511166, + 0.0030209566466510296, + 0.06505068391561508, + 0.07586846500635147, + 0.005802581086754799, + 0.0072126672603189945, + 0.017419269308447838, + 0.0067040990106761456, + 0.006400353740900755, + 0.0066741458140313625, + 0.025557061657309532, + 0.005231427028775215, + 0.004762475844472647, + 0.003775457851588726, + 0.007108662277460098, + 0.0048258546739816666, + 0.005462362430989742, + 0.026557615026831627, + 0.02569931000471115, + 0.012945744208991528, + 0.014802164398133755, + 0.04218599572777748, + 0.01940043829381466, + 0.02433120086789131, + 0.021494632586836815, + 0.05359750986099243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.02983708307147026 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809695690870285 + ], + "opposite_logprob_tokens": [ + -3.2880969047546387 + ], + "logprob_reward": -0.03809695690870285, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809695690870285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11258552223443985, + 0.07019071280956268, + 0.08634080737829208, + 0.08886314183473587, + 0.1891430765390396, + 0.041153814643621445, + 0.064128078520298, + 0.23369474709033966, + 0.056337155401706696, + 0.05756296217441559 + ], + "credit_map": [ + 0.01216012891381979, + 0.01357500534504652, + 0.013619094155728817, + 0.010240786708891392, + 0.023479018360376358, + 0.012536263093352318, + 0.017073296010494232, + 0.009901930578052998, + 0.00870010070502758, + 0.00889627169817686, + 0.0067882533185184, + 0.005780884064733982, + 0.011905471794307232, + 0.009950758889317513, + 0.010605594143271446, + 0.00756338145583868, + 0.009736044332385063, + 0.008438163436949253, + 0.010787911713123322, + 0.00914002861827612, + 0.01250278390944004, + 0.010765383951365948, + 0.013295971788465977, + 0.011674521490931511, + 0.010203506797552109, + 0.01128251664340496, + 0.011510578915476799, + 0.010115866549313068, + 0.011504881083965302, + 0.008973287418484688, + 0.012657559476792812, + 0.012614943087100983, + 0.02028511092066765, + 0.015690462663769722, + 0.015365160070359707, + 0.012192496098577976, + 0.025668399408459663, + 0.04192926734685898, + 0.03609320521354675, + 0.021918967366218567, + 0.005274066235870123, + 0.003399292007088661, + 0.0019889166578650475, + 0.008305391296744347, + 0.01271061785519123, + 0.00418432243168354, + 0.0021747390273958445, + 0.0031164675019681454, + 0.010750293731689453, + 0.009474498219788074, + 0.008198531344532967, + 0.004196156747639179, + 0.008289683610200882, + 0.00657846312969923, + 0.010196618735790253, + 0.0064438325352966785, + 0.0648312121629715, + 0.06861704587936401, + 0.018737252801656723, + 0.012999134138226509, + 0.025148946791887283, + 0.014201960526406765, + 0.01522843074053526, + 0.013930757530033588, + 0.007390569429844618, + 0.006731396075338125, + 0.005956125445663929, + 0.005961115472018719, + 0.011602584272623062, + 0.006916707381606102, + 0.007156409788876772, + 0.004622251261025667, + 0.00586819089949131, + 0.003286226186901331, + 0.007894243113696575, + 0.01540687121450901, + 0.004394736606627703, + 0.004184036981314421, + 0.002931930823251605, + 0.013596721924841404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.03809695690870285 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589197054505348 + ], + "opposite_logprob_tokens": [ + -3.8955891132354736 + ], + "logprob_reward": -0.020589197054505348, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589197054505348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09327393770217896, + 0.08259250223636627, + 0.06867994368076324, + 0.04698871448636055, + 0.04389222338795662, + 0.11972687393426895, + 0.26972073316574097, + 0.17699752748012543, + 0.020311176776885986, + 0.07781640440225601 + ], + "credit_map": [ + 0.012077182531356812, + 0.010968644171953201, + 0.00982646457850933, + 0.009452428668737411, + 0.020947735756635666, + 0.009342923760414124, + 0.011549086309969425, + 0.009109469130635262, + 0.012558074668049812, + 0.011447776108980179, + 0.00749832671135664, + 0.004472202155739069, + 0.01641722582280636, + 0.012211790308356285, + 0.00829132180660963, + 0.009695778600871563, + 0.009731045924127102, + 0.0069334846921265125, + 0.008011950179934502, + 0.0070760734379291534, + 0.011126033961772919, + 0.009127367287874222, + 0.008296329528093338, + 0.008377648890018463, + 0.006917028222233057, + 0.007081435993313789, + 0.0047034225426614285, + 0.004194523207843304, + 0.00628149788826704, + 0.0033428354654461145, + 0.005267661064863205, + 0.009200307540595531, + 0.005324303638190031, + 0.0037106098607182503, + 0.00436041597276926, + 0.004682530649006367, + 0.0071390108205378056, + 0.0073603419587016106, + 0.0052861543372273445, + 0.006028853822499514, + 0.03821940720081329, + 0.007102501578629017, + 0.006026299670338631, + 0.006934076081961393, + 0.01949702389538288, + 0.007331897504627705, + 0.005575206130743027, + 0.029040461406111717, + 0.07689911872148514, + 0.03318821266293526, + 0.02597290650010109, + 0.059848539531230927, + 0.02327846549451351, + 0.015769479796290398, + 0.016069911420345306, + 0.01869407296180725, + 0.049518048763275146, + 0.05614122003316879, + 0.011005537584424019, + 0.009019511751830578, + 0.019654888659715652, + 0.011132486164569855, + 0.010720930062234402, + 0.009804900735616684, + 0.0027866947930306196, + 0.002378616016358137, + 0.0026405768003314734, + 0.00251575605943799, + 0.0026574963703751564, + 0.0019084705272689462, + 0.001852604909799993, + 0.0035709610674530268, + 0.018753010779619217, + 0.004262469243258238, + 0.008851704187691212, + 0.010385822504758835, + 0.0062337773852050304, + 0.006328979041427374, + 0.005685437470674515, + 0.017315199598670006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.020589197054505348 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two rock pigeons visible standing side by side at the edge of a shallow, clear puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a damp, gray cobblestone street under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both birds simultaneously dip their dark gray beaks into the water in a swift, synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the birds jerk their heads back and forth erratically in a staccato rhythm, ejecting a cloud of shimmering droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show concentric ripples radiating outward from the disturbed puddle surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864705726504326 + ], + "opposite_logprob_tokens": [ + -3.048647165298462 + ], + "logprob_reward": -0.04864705726504326, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864705726504326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0311751589179039, + 0.05654529482126236, + 0.0636318102478981, + 0.03371578827500343, + 0.1616896241903305, + 0.33552074432373047, + 0.1150646060705185, + 0.15012620389461517, + 0.03529367968440056, + 0.01723710633814335 + ], + "credit_map": [ + 0.005301628261804581, + 0.003883677301928401, + 0.003911043982952833, + 0.002689951565116644, + 0.005021009128540754, + 0.0035157622769474983, + 0.002656539436429739, + 0.004195545334368944, + 0.008996414951980114, + 0.008037365041673183, + 0.005792274139821529, + 0.005138208623975515, + 0.008460789918899536, + 0.0058477469719946384, + 0.004386322107166052, + 0.009886171668767929, + 0.008431672118604183, + 0.007593431510031223, + 0.008839019574224949, + 0.00836936291307211, + 0.006536259315907955, + 0.006242658011615276, + 0.00791182555258274, + 0.009707581251859665, + 0.005303890444338322, + 0.005314450711011887, + 0.0031114371959120035, + 0.0025601822417229414, + 0.003632471663877368, + 0.0031363314483314753, + 0.003193406853824854, + 0.00746361818164587, + 0.019419129937887192, + 0.013417107984423637, + 0.014821421355009079, + 0.025160929188132286, + 0.019017765298485756, + 0.02196899987757206, + 0.027508718892931938, + 0.020375549793243408, + 0.05562884360551834, + 0.023365046828985214, + 0.019458433613181114, + 0.021495895460247993, + 0.1142084002494812, + 0.026751955971121788, + 0.017492882907390594, + 0.057119302451610565, + 0.02922002039849758, + 0.013674064539372921, + 0.009187854826450348, + 0.02045218087732792, + 0.011547122150659561, + 0.008694312535226345, + 0.010807482525706291, + 0.011481567285954952, + 0.04389238730072975, + 0.052979618310928345, + 0.008970599621534348, + 0.0068901716731488705, + 0.014598654583096504, + 0.007141653914004564, + 0.0077024768106639385, + 0.007950653322041035, + 0.006937320344150066, + 0.002710913773626089, + 0.0028929547406733036, + 0.003928988706320524, + 0.004970748908817768, + 0.0033818120136857033, + 0.002552467631176114, + 0.00791847426444292, + 0.0034140946809202433, + 0.0016376855783164501, + 0.001992191653698683, + 0.003988325130194426, + 0.0021499856375157833, + 0.0015979140298441052, + 0.0011201555607840419, + 0.0013367547653615475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.04864705726504326 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_7/index_tensor([4701, 4701])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_7/index_tensor([4701, 4701])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..df487da51604116801597bb3bbeed0fa158c85b5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_7/index_tensor([4701, 4701])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a vibrant pop art style, a rapid 360-degree orbiting shot rotates around the featured bottle of fruit-flavored beer, capturing its exaggerated exhaustion from every angle. Made of glass, but depicted flatly in pale yellow with a bold black outline, the bottle features a simple cartoon face, with two downward-curving black lines representing closed, sleepy eyes. As the camera rotates, the bottle's mouth, the metal cap area, stretches impossibly in a vast, silent yawn, its glass body slightly distorting to emphasize the effort. A stream of perfectly circular, graphic-style bubbles, each with a single white highlight, escapes the opening and floats upward against the dynamic background. The setting is a non-representational explosion of color, with a flat, electric blue background overlaid with a precise grid of large, red Ben-Day dots, creating a stark graphic contrast. The entire scene is illuminated by a bright, even, shadowless light, reinforcing the bland comic book aesthetic and creating a playful, high-energy atmosphere tinged with cartoon fatigue.", + "seed": 3969, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_7/index_4701_seed_3969_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2397978536784649, + "vlm_logprob": -0.2397978536784649, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.2397978536784649, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009820717386901379 + ], + "opposite_logprob_tokens": [ + -4.759820938110352 + ], + "logprob_reward": -0.009820717386901379, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.009820717386901379, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11570107936859131, + 0.0757986530661583, + 0.09597060084342957, + 0.1318666636943817, + 0.16265109181404114, + 0.07475217431783676, + 0.09769296646118164, + 0.1021890789270401, + 0.059922657907009125, + 0.08345501869916916 + ], + "credit_map": [ + 0.016364838927984238, + 0.012697344645857811, + 0.02120070718228817, + 0.022108493372797966, + 0.009010017849504948, + 0.010994065552949905, + 0.01304739061743021, + 0.010278225876390934, + 0.010505037382245064, + 0.007877880707383156, + 0.012955820187926292, + 0.01122099906206131, + 0.006234683096408844, + 0.009600912220776081, + 0.006864480208605528, + 0.010538841597735882, + 0.015213973820209503, + 0.013180715963244438, + 0.013375366106629372, + 0.01968950591981411, + 0.009416106157004833, + 0.006504413206130266, + 0.007338388357311487, + 0.01125213410705328, + 0.016505327075719833, + 0.02210988663136959, + 0.01889646239578724, + 0.019354473799467087, + 0.011172211728990078, + 0.013348565436899662, + 0.0131459916010499, + 0.0173337422311306, + 0.02502397447824478, + 0.014598906971514225, + 0.032955072820186615, + 0.03736204653978348, + 0.011400154791772366, + 0.01006009429693222, + 0.014013909734785557, + 0.01723693124949932, + 0.011044255457818508, + 0.007764077745378017, + 0.008080671541392803, + 0.011183468624949455, + 0.007406748831272125, + 0.008398220874369144, + 0.006368081551045179, + 0.01450665108859539, + 0.009256723336875439, + 0.010935915634036064, + 0.00745501508936286, + 0.018037976697087288, + 0.01382079441100359, + 0.01383237075060606, + 0.012581468559801579, + 0.011772708036005497, + 0.018039368093013763, + 0.013779453933238983, + 0.02214277721941471, + 0.011547013185918331, + 0.009737752377986908, + 0.0071446578949689865, + 0.010812722146511078, + 0.008985335938632488, + 0.00700742844492197, + 0.004511047154664993, + 0.005441577639430761, + 0.010217213071882725, + 0.005686111748218536, + 0.009457371197640896, + 0.010008616372942924, + 0.007593293208628893, + 0.01575591415166855, + 0.00812726840376854, + 0.009407037869095802, + 0.015758415684103966, + 0.008015530183911324, + 0.006645487155765295, + 0.010224428959190845, + 0.009520933032035828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.009820717386901379 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786483831703663 + ], + "opposite_logprob_tokens": [ + -4.634786605834961 + ], + "logprob_reward": -0.009786483831703663, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786483831703663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09784026443958282, + 0.09678970277309418, + 0.15074379742145538, + 0.08691674470901489, + 0.11177156120538712, + 0.0857287347316742, + 0.162883922457695, + 0.05574503540992737, + 0.07665742933750153, + 0.0749228224158287 + ], + "credit_map": [ + 0.010957596823573112, + 0.009631898254156113, + 0.015668960288167, + 0.015153567306697369, + 0.009668736718595028, + 0.01439386885613203, + 0.011715454049408436, + 0.010650171898305416, + 0.017021499574184418, + 0.008643054403364658, + 0.014155389741063118, + 0.014981104992330074, + 0.007335487753152847, + 0.01393932942301035, + 0.008312209509313107, + 0.01240161806344986, + 0.027143659070134163, + 0.018577277660369873, + 0.022000683471560478, + 0.031900107860565186, + 0.017704134806990623, + 0.010672648437321186, + 0.01179770939052105, + 0.010947568342089653, + 0.01063181646168232, + 0.010200660675764084, + 0.013465373776853085, + 0.01350519061088562, + 0.008749842643737793, + 0.007813180796802044, + 0.011090969666838646, + 0.011459697037935257, + 0.01709163561463356, + 0.014201038517057896, + 0.018916422501206398, + 0.022952118888497353, + 0.008457004092633724, + 0.010189441032707691, + 0.009301736019551754, + 0.010662153363227844, + 0.010759979486465454, + 0.008570954203605652, + 0.011769040487706661, + 0.012482627294957638, + 0.00925949215888977, + 0.010450998321175575, + 0.0091199716553092, + 0.013315660879015923, + 0.014910024590790272, + 0.012899001128971577, + 0.008364392444491386, + 0.028640735894441605, + 0.036468587815761566, + 0.019515840336680412, + 0.02517852745950222, + 0.016906803473830223, + 0.0065646301954984665, + 0.004869669675827026, + 0.005777635611593723, + 0.006069813389331102, + 0.011619177646934986, + 0.005502001382410526, + 0.009279778227210045, + 0.006062323227524757, + 0.008770884945988655, + 0.011509480886161327, + 0.005634243134409189, + 0.009213301353156567, + 0.012442666105926037, + 0.007731690537184477, + 0.013198384083807468, + 0.008156772702932358, + 0.012259988114237785, + 0.008639702573418617, + 0.011166236363351345, + 0.00990930013358593, + 0.010415753349661827, + 0.00725923478603363, + 0.006973407231271267, + 0.008299185894429684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.009786483831703663 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605898328125477 + ], + "opposite_logprob_tokens": [ + -4.141058921813965 + ], + "logprob_reward": -0.01605898328125477, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605898328125477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0802370011806488, + 0.0782194435596466, + 0.08748031407594681, + 0.06770867854356766, + 0.14404107630252838, + 0.13111084699630737, + 0.13937225937843323, + 0.08156955987215042, + 0.10099893063306808, + 0.08926183730363846 + ], + "credit_map": [ + 0.010992111638188362, + 0.00931471586227417, + 0.012112337164580822, + 0.009450269863009453, + 0.008796414360404015, + 0.010560398921370506, + 0.008453630842268467, + 0.010557123459875584, + 0.010826105251908302, + 0.008131847716867924, + 0.013595297932624817, + 0.013519643805921078, + 0.006296148058027029, + 0.010884745977818966, + 0.005359248723834753, + 0.009606405161321163, + 0.014043805189430714, + 0.01262739859521389, + 0.010738649405539036, + 0.015417053364217281, + 0.007944713346660137, + 0.0063750552944839, + 0.00791342556476593, + 0.012420211918652058, + 0.010838653892278671, + 0.007857061922550201, + 0.010658184997737408, + 0.010887856595218182, + 0.0064694639295339584, + 0.004983264487236738, + 0.008000415749847889, + 0.008013773709535599, + 0.016617367044091225, + 0.021357698366045952, + 0.02782573737204075, + 0.03655557706952095, + 0.009278847835958004, + 0.010697211138904095, + 0.010724276304244995, + 0.010984367690980434, + 0.01580432802438736, + 0.011403451673686504, + 0.020022369921207428, + 0.030151870101690292, + 0.012097028084099293, + 0.014469322748482227, + 0.009410370141267776, + 0.017752112820744514, + 0.012024153023958206, + 0.01186007633805275, + 0.008095014840364456, + 0.02600077912211418, + 0.025378581136465073, + 0.020230883732438087, + 0.020758746191859245, + 0.015024018473923206, + 0.01492406241595745, + 0.010439918376505375, + 0.01187689695507288, + 0.009874210692942142, + 0.010424317792057991, + 0.0066057899966835976, + 0.011181418783962727, + 0.006242945324629545, + 0.0093727707862854, + 0.013752056285738945, + 0.008474594913423061, + 0.009434626437723637, + 0.017144573852419853, + 0.013426137156784534, + 0.01616591401398182, + 0.013228255324065685, + 0.01839073933660984, + 0.008431505411863327, + 0.01419791765511036, + 0.016143308952450752, + 0.0098292026668787, + 0.006142789032310247, + 0.007801407016813755, + 0.008324961178004742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.01605898328125477 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1369832754135132 + ], + "opposite_logprob_tokens": [ + -0.38698330521583557 + ], + "logprob_reward": -1.1369832754135132, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1369832754135132, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04887405410408974, + 0.06277987360954285, + 0.061223797500133514, + 0.09870712459087372, + 0.11407007277011871, + 0.11656740307807922, + 0.3854110836982727, + 0.04127536341547966, + 0.047193191945552826, + 0.023898012936115265 + ], + "credit_map": [ + 0.008624103851616383, + 0.004666912369430065, + 0.007403276395052671, + 0.007731771096587181, + 0.004755082074552774, + 0.006576938554644585, + 0.00404308270663023, + 0.005072887986898422, + 0.012799790129065514, + 0.006091534160077572, + 0.011431729421019554, + 0.012186511419713497, + 0.004901959095150232, + 0.0052239010110497475, + 0.004556981846690178, + 0.005587466061115265, + 0.012123286724090576, + 0.006003331393003464, + 0.007717906963080168, + 0.01022196002304554, + 0.00656649237498641, + 0.004809965379536152, + 0.0065397946164011955, + 0.007241064216941595, + 0.013646294362843037, + 0.015212390571832657, + 0.01413404569029808, + 0.014700294472277164, + 0.011232205666601658, + 0.008895556442439556, + 0.010439984500408173, + 0.010446349158883095, + 0.017109861597418785, + 0.014445463195443153, + 0.01972704753279686, + 0.01956883631646633, + 0.009003628976643085, + 0.009077918715775013, + 0.008608357049524784, + 0.016528965905308723, + 0.01907537877559662, + 0.012540988624095917, + 0.016512250527739525, + 0.013699225150048733, + 0.010219790041446686, + 0.01632702723145485, + 0.010953952558338642, + 0.017238792032003403, + 0.03593902289867401, + 0.032812219113111496, + 0.01871681958436966, + 0.06456232070922852, + 0.08993112295866013, + 0.04627755284309387, + 0.058036498725414276, + 0.03913556784391403, + 0.006674269679933786, + 0.006286449730396271, + 0.004873866681009531, + 0.003941272385418415, + 0.006186886690557003, + 0.004469755105674267, + 0.003795619821175933, + 0.005047242622822523, + 0.0049073947593569756, + 0.007307243999093771, + 0.0032799062319099903, + 0.007156290113925934, + 0.008066239766776562, + 0.005832493770867586, + 0.006800384726375341, + 0.0038432397413998842, + 0.003469038987532258, + 0.002595123602077365, + 0.003948998171836138, + 0.003337536007165909, + 0.0026126238517463207, + 0.002324941335245967, + 0.002552793361246586, + 0.003056956920772791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -1.1369832754135132 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026339808478951454 + ], + "opposite_logprob_tokens": [ + -3.6513397693634033 + ], + "logprob_reward": -0.026339808478951454, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026339808478951454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06879934668540955, + 0.05438404157757759, + 0.07111179083585739, + 0.06512989848852158, + 0.17906616628170013, + 0.07617455720901489, + 0.21027958393096924, + 0.05889936909079552, + 0.11165261268615723, + 0.1045026108622551 + ], + "credit_map": [ + 0.012227198109030724, + 0.007225581910461187, + 0.011444485746324062, + 0.012752941809594631, + 0.005189590156078339, + 0.006827151868492365, + 0.007261516526341438, + 0.005870883818715811, + 0.005886886268854141, + 0.006669752765446901, + 0.007514443714171648, + 0.008312856778502464, + 0.0056739868596196175, + 0.00613095099106431, + 0.005185442510992289, + 0.009009722620248795, + 0.01360770221799612, + 0.006911388598382473, + 0.010188157670199871, + 0.010542381554841995, + 0.00854652188718319, + 0.007831178605556488, + 0.00572572648525238, + 0.007758735213428736, + 0.009023318998515606, + 0.009490815922617912, + 0.010235391557216644, + 0.011760275810956955, + 0.005781649611890316, + 0.006052052136510611, + 0.005980313289910555, + 0.00680608069524169, + 0.031284719705581665, + 0.026057304814457893, + 0.032009758055210114, + 0.0310530848801136, + 0.01219593733549118, + 0.014155417680740356, + 0.012054985389113426, + 0.02025495283305645, + 0.008789683692157269, + 0.00772433215752244, + 0.009372393600642681, + 0.011826586909592152, + 0.010188028216362, + 0.009262176230549812, + 0.0077827088534832, + 0.01122865080833435, + 0.02016698755323887, + 0.01871410757303238, + 0.010090157389640808, + 0.03572756052017212, + 0.04778515174984932, + 0.02459581196308136, + 0.0320146307349205, + 0.021185167133808136, + 0.00864448118954897, + 0.008899720385670662, + 0.005693519487977028, + 0.006820733658969402, + 0.0058375513181090355, + 0.006413740571588278, + 0.007068455219268799, + 0.009521166794002056, + 0.012695720419287682, + 0.01631726510822773, + 0.00924421101808548, + 0.01578681729733944, + 0.018476704135537148, + 0.011728677898645401, + 0.016922589391469955, + 0.010480626486241817, + 0.019944677129387856, + 0.011543630622327328, + 0.018465667963027954, + 0.016704503446817398, + 0.010469583794474602, + 0.007222217041999102, + 0.010046682320535183, + 0.010105657391250134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 8 + ], + "selected_reward": -0.026339808478951454 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "In a vibrant pop art style, a rapid 360-degree orbiting shot rotates around the featured bottle of fruit-flavored beer, capturing its exaggerated exhaustion from every angle. Made of glass, but depicted flatly in pale yellow with a bold black outline, the bottle features a simple cartoon face, with two downward-curving black lines representing closed, sleepy eyes. As the camera rotates, the bottle's mouth, the metal cap area, stretches impossibly in a vast, silent yawn, its glass body slightly distorting to emphasize the effort. A stream of perfectly circular, graphic-style bubbles, each with a single white highlight, escapes the opening and floats upward against the dynamic background. The setting is a non-representational explosion of color, with a flat, electric blue background overlaid with a precise grid of large, red Ben-Day dots, creating a stark graphic contrast. The entire scene is illuminated by a bright, even, shadowless light, reinforcing the bland comic book aesthetic and creating a playful, high-energy atmosphere tinged with cartoon fatigue.", + "seed": 3970, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_7/index_4701_seed_3970_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.413414490967989, + "vlm_logprob": -0.413414490967989, + "vlm_yes_no_margin": 3.225000023841858, + "vlm_token_credit": -0.413414490967989, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009457547217607498 + ], + "opposite_logprob_tokens": [ + -4.759457588195801 + ], + "logprob_reward": -0.009457547217607498, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.009457547217607498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07317840307950974, + 0.059614554047584534, + 0.04839389771223068, + 0.2583860754966736, + 0.066319040954113, + 0.08273657411336899, + 0.02660321071743965, + 0.06737343966960907, + 0.040415313094854355, + 0.2769794762134552 + ], + "credit_map": [ + 0.009515972808003426, + 0.005865469109266996, + 0.015567407943308353, + 0.006922549568116665, + 0.006226075813174248, + 0.009044656530022621, + 0.007613457273691893, + 0.012422814965248108, + 0.009685366414487362, + 0.003330122446641326, + 0.010784056037664413, + 0.0032763665076345205, + 0.002936793491244316, + 0.005613639950752258, + 0.005945081822574139, + 0.01804313063621521, + 0.01232055202126503, + 0.003337038680911064, + 0.012336832471191883, + 0.0033123400062322617, + 0.0026033497415483, + 0.004015077371150255, + 0.003369645681232214, + 0.007099061273038387, + 0.03514425456523895, + 0.02834564633667469, + 0.08871816098690033, + 0.01400996558368206, + 0.012529147788882256, + 0.018906565383076668, + 0.04139005392789841, + 0.019342277199029922, + 0.0064706080593168736, + 0.004734822083264589, + 0.006092466413974762, + 0.004987331107258797, + 0.006106488406658173, + 0.01488638948649168, + 0.01758873648941517, + 0.005452194716781378, + 0.007663215976208448, + 0.00671274121850729, + 0.007307552266865969, + 0.00932574737817049, + 0.006949113216251135, + 0.014222610741853714, + 0.01759476400911808, + 0.012960831634700298, + 0.0029308705125004053, + 0.00254896585829556, + 0.005864446517080069, + 0.003397279651835561, + 0.0030539100989699364, + 0.0028625745326280594, + 0.002814863109961152, + 0.0031303022988140583, + 0.013761324808001518, + 0.005401177331805229, + 0.009881162084639072, + 0.006692349445074797, + 0.004838382359594107, + 0.012074652127921581, + 0.0073639508336782455, + 0.0073604402132332325, + 0.004243840929120779, + 0.0054562753066420555, + 0.007393313106149435, + 0.005569361615926027, + 0.004634090233594179, + 0.004716879688203335, + 0.003418264677748084, + 0.004983288701623678, + 0.03963089734315872, + 0.02664703130722046, + 0.10154443979263306, + 0.016920939087867737, + 0.0101366201415658, + 0.018338238820433617, + 0.007323163561522961, + 0.05643815919756889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 5 + ], + "selected_reward": -0.009457547217607498 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674376729875803 + ], + "opposite_logprob_tokens": [ + -5.006743907928467 + ], + "logprob_reward": -0.00674376729875803, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674376729875803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13952945172786713, + 0.18763624131679535, + 0.0454532653093338, + 0.10384945571422577, + 0.090902179479599, + 0.1284654587507248, + 0.029908278957009315, + 0.14327846467494965, + 0.02842259407043457, + 0.10255464911460876 + ], + "credit_map": [ + 0.02426554262638092, + 0.008595828898251057, + 0.02324184589087963, + 0.009572653099894524, + 0.00813250057399273, + 0.018160009756684303, + 0.013976440764963627, + 0.033584631979465485, + 0.03495079651474953, + 0.009469976648688316, + 0.0216533150523901, + 0.008146844804286957, + 0.007217878010123968, + 0.021879049018025398, + 0.020519813522696495, + 0.06379856169223785, + 0.011063070967793465, + 0.003120922949165106, + 0.009323657490313053, + 0.0031310352496802807, + 0.0027956687845289707, + 0.00562310079112649, + 0.003868510713800788, + 0.006527296733111143, + 0.017439553514122963, + 0.012222657911479473, + 0.03652561083436012, + 0.005931474734097719, + 0.006724191829562187, + 0.008069116622209549, + 0.00892026536166668, + 0.008016584441065788, + 0.01173635758459568, + 0.00732197891920805, + 0.01782672293484211, + 0.009184511378407478, + 0.010105512104928493, + 0.013877691701054573, + 0.012881307862699032, + 0.00796809233725071, + 0.011248480528593063, + 0.011237427592277527, + 0.012295235879719257, + 0.010558356530964375, + 0.011543149128556252, + 0.033850573003292084, + 0.024571387097239494, + 0.013160841539502144, + 0.003476367797702551, + 0.0025673869531601667, + 0.005092561710625887, + 0.003755530109629035, + 0.0032144596334546804, + 0.0038661938160657883, + 0.004168330226093531, + 0.003767448477447033, + 0.03626253455877304, + 0.009613160975277424, + 0.01995963416993618, + 0.012087258510291576, + 0.008193770423531532, + 0.024101756513118744, + 0.010702904313802719, + 0.02235744148492813, + 0.0028825532644987106, + 0.0036316129844635725, + 0.004063946194946766, + 0.00453727412968874, + 0.003030214225873351, + 0.0036129276268184185, + 0.0028751662466675043, + 0.003788899164646864, + 0.01638534665107727, + 0.013020394369959831, + 0.030092278495430946, + 0.006026572082191706, + 0.005460618529468775, + 0.010263701900839806, + 0.004351179581135511, + 0.016954561695456505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.00674376729875803 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536530382931232 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536530382931232, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536530382931232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13831119239330292, + 0.07268555462360382, + 0.040431223809719086, + 0.023721447214484215, + 0.049971744418144226, + 0.11203841865062714, + 0.2793693542480469, + 0.06573674082756042, + 0.05624512955546379, + 0.16148915886878967 + ], + "credit_map": [ + 0.025008490309119225, + 0.01049080491065979, + 0.013533844612538815, + 0.010658366605639458, + 0.011154856532812119, + 0.019641052931547165, + 0.01514187641441822, + 0.0326819121837616, + 0.007759829983115196, + 0.005099267698824406, + 0.015143959783017635, + 0.007383182179182768, + 0.006457026582211256, + 0.013340956531465054, + 0.007265124469995499, + 0.010236208327114582, + 0.004418650176376104, + 0.0029593128710985184, + 0.008383499458432198, + 0.004570935852825642, + 0.004213991109281778, + 0.005036916583776474, + 0.003889717161655426, + 0.006958201061934233, + 0.0028564094100147486, + 0.0018718746723607183, + 0.0035796593874692917, + 0.0015098043950274587, + 0.00176938995718956, + 0.0032638821285218, + 0.004344265908002853, + 0.004526161588728428, + 0.0056059276685118675, + 0.006043123546987772, + 0.0048586525954306126, + 0.006497271358966827, + 0.007063492201268673, + 0.008295340463519096, + 0.0080976951867342, + 0.003510244656354189, + 0.006950002629309893, + 0.004000928718596697, + 0.008261830545961857, + 0.007852206937968731, + 0.008006268180906773, + 0.01467222347855568, + 0.02882736548781395, + 0.033467601984739304, + 0.03080553188920021, + 0.02831730991601944, + 0.04487397521734238, + 0.042718302458524704, + 0.04202442616224289, + 0.026402238756418228, + 0.03143542259931564, + 0.03279216215014458, + 0.014603178016841412, + 0.004647740162909031, + 0.008738941513001919, + 0.006178883835673332, + 0.004037948325276375, + 0.013102578930556774, + 0.005116078071296215, + 0.009311392903327942, + 0.004747149534523487, + 0.008308184333145618, + 0.010255495086312294, + 0.00698953727260232, + 0.009342853911221027, + 0.005188033916056156, + 0.006171687506139278, + 0.005242188926786184, + 0.023189837113022804, + 0.015866247937083244, + 0.05879884585738182, + 0.00983602087944746, + 0.0056945811957120895, + 0.010937993414700031, + 0.004050300922244787, + 0.03311533480882645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.012536530382931232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.0177414417266846 + ], + "opposite_logprob_tokens": [ + -0.14274154603481293 + ], + "logprob_reward": -2.0177414417266846, + "yes_no_margin_reward": -1.8749998807907104, + "token_credit_reward": -2.0177414417266846, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.042634908109903336, + 0.17151671648025513, + 0.03029220923781395, + 0.03925054892897606, + 0.061604976654052734, + 0.019329123198986053, + 0.06741945445537567, + 0.17327292263507843, + 0.05407373979687691, + 0.34060537815093994 + ], + "credit_map": [ + 0.007618095725774765, + 0.0030742825474590063, + 0.006791302002966404, + 0.003441091161221266, + 0.002695733681321144, + 0.005900043062865734, + 0.004578971769660711, + 0.008535390719771385, + 0.02984686940908432, + 0.008397133089601994, + 0.01926678791642189, + 0.00784391537308693, + 0.0056604607962071896, + 0.016583776101469994, + 0.019873443990945816, + 0.06404431909322739, + 0.004610349889844656, + 0.00242715235799551, + 0.009313247166574001, + 0.00303258397616446, + 0.0020658511202782393, + 0.003003281308338046, + 0.0022720193956047297, + 0.003567724721506238, + 0.007708530407398939, + 0.003671320853754878, + 0.008258489891886711, + 0.0024855975061655045, + 0.003324618097394705, + 0.0056649609468877316, + 0.003216461045667529, + 0.004920575767755508, + 0.005801643244922161, + 0.0027965547051280737, + 0.004658718127757311, + 0.003697542706504464, + 0.003915851004421711, + 0.01880042813718319, + 0.017149318009614944, + 0.0047849202528595924, + 0.0021549302618950605, + 0.0013408922823145986, + 0.0021136763971298933, + 0.0015557310543954372, + 0.0012801532866433263, + 0.002989955013617873, + 0.004728137981146574, + 0.0031656476203352213, + 0.007103061769157648, + 0.00620696134865284, + 0.013165597803890705, + 0.00960760097950697, + 0.008729231543838978, + 0.008029451593756676, + 0.006895347964018583, + 0.007682204712182283, + 0.02454489655792713, + 0.014227649196982384, + 0.027536556124687195, + 0.015556859783828259, + 0.015519564040005207, + 0.03798219561576843, + 0.017614129930734634, + 0.020291069522500038, + 0.00531506072729826, + 0.00900451373308897, + 0.008555452339351177, + 0.0067891767248511314, + 0.008751322515308857, + 0.00507316691800952, + 0.0064939092844724655, + 0.004091136623173952, + 0.04702972248196602, + 0.03201087564229965, + 0.12803895771503448, + 0.020834676921367645, + 0.010904079303145409, + 0.02070777118206024, + 0.007874691858887672, + 0.07320459187030792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -2.0177414417266846 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059316821396351 + ], + "opposite_logprob_tokens": [ + -3.8955931663513184 + ], + "logprob_reward": -0.02059316821396351, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059316821396351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030689505860209465, + 0.23717555403709412, + 0.029081881046295166, + 0.07210420817136765, + 0.0354815348982811, + 0.025322284549474716, + 0.05698880925774574, + 0.12382452189922333, + 0.029920563101768494, + 0.3594111502170563 + ], + "credit_map": [ + 0.0050931465812027454, + 0.0023935819044709206, + 0.008146893233060837, + 0.0024408544413745403, + 0.001710506621748209, + 0.0045189643278717995, + 0.0026039667427539825, + 0.003781592007726431, + 0.04442606493830681, + 0.011878274381160736, + 0.021564798429608345, + 0.008322075009346008, + 0.006725183222442865, + 0.02326291799545288, + 0.0285485927015543, + 0.09244765341281891, + 0.0043710204772651196, + 0.0025166752748191357, + 0.010035865008831024, + 0.0024035691749304533, + 0.0019127605482935905, + 0.003768159309402108, + 0.0016481165075674653, + 0.0024257146287709475, + 0.0105033153668046, + 0.008272077888250351, + 0.023637764155864716, + 0.0036619664169847965, + 0.003928052727133036, + 0.005659342743456364, + 0.009807753376662731, + 0.006633935496211052, + 0.005471696145832539, + 0.0024404949508607388, + 0.008937487378716469, + 0.002892095595598221, + 0.002743674209341407, + 0.006492234766483307, + 0.0029749535024166107, + 0.003528896952047944, + 0.0024700439535081387, + 0.0019392670365050435, + 0.0023395572789013386, + 0.0019504387164488435, + 0.0016657138476148248, + 0.007690113503485918, + 0.004724045284092426, + 0.0025431043468415737, + 0.0059827775694429874, + 0.00563993351534009, + 0.009367132559418678, + 0.008630623109638691, + 0.008409319445490837, + 0.006116984412074089, + 0.006221188697963953, + 0.006620846223086119, + 0.02227838709950447, + 0.009598865173757076, + 0.01877506822347641, + 0.011003924533724785, + 0.010126570239663124, + 0.025081660598516464, + 0.011271604336798191, + 0.015688447281718254, + 0.003227368462830782, + 0.004829929210245609, + 0.004178279545158148, + 0.0036765558179467916, + 0.0036107602063566446, + 0.0036950174253433943, + 0.002738483250141144, + 0.003964169416576624, + 0.05057469382882118, + 0.033901505172252655, + 0.13439148664474487, + 0.021957911550998688, + 0.012014156207442284, + 0.022007161751389503, + 0.008486643433570862, + 0.07607755810022354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.02059316821396351 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "In a vibrant pop art style, a rapid 360-degree orbiting shot rotates around the featured bottle of fruit-flavored beer, capturing its exaggerated exhaustion from every angle. Made of glass, but depicted flatly in pale yellow with a bold black outline, the bottle features a simple cartoon face, with two downward-curving black lines representing closed, sleepy eyes. As the camera rotates, the bottle's mouth, the metal cap area, stretches impossibly in a vast, silent yawn, its glass body slightly distorting to emphasize the effort. A stream of perfectly circular, graphic-style bubbles, each with a single white highlight, escapes the opening and floats upward against the dynamic background. The setting is a non-representational explosion of color, with a flat, electric blue background overlaid with a precise grid of large, red Ben-Day dots, creating a stark graphic contrast. The entire scene is illuminated by a bright, even, shadowless light, reinforcing the bland comic book aesthetic and creating a playful, high-energy atmosphere tinged with cartoon fatigue.", + "seed": 3971, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_7/index_4701_seed_3971_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3152088919654489, + "vlm_logprob": -0.3152088919654489, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.3152088919654489, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00819061417132616 + ], + "opposite_logprob_tokens": [ + -5.508190631866455 + ], + "logprob_reward": -0.00819061417132616, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00819061417132616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19964854419231415, + 0.11638934165239334, + 0.03434159606695175, + 0.042634136974811554, + 0.1261192262172699, + 0.07465456426143646, + 0.214150071144104, + 0.05887860432267189, + 0.05151382461190224, + 0.08167014271020889 + ], + "credit_map": [ + 0.026926657184958458, + 0.02147139050066471, + 0.027632946148514748, + 0.020479289814829826, + 0.024600403383374214, + 0.02481672167778015, + 0.02802119217813015, + 0.025699952617287636, + 0.020936820656061172, + 0.015509928576648235, + 0.01698850281536579, + 0.00816951971501112, + 0.014342151582241058, + 0.01844712905585766, + 0.010390184819698334, + 0.011605104431509972, + 0.0019168435828760266, + 0.0034294133074581623, + 0.0029038169886916876, + 0.0024269495625048876, + 0.007312806323170662, + 0.006286425516009331, + 0.003969176206737757, + 0.006096165161579847, + 0.005334086716175079, + 0.004552503582090139, + 0.00815228745341301, + 0.004243760835379362, + 0.004602684639394283, + 0.007061007898300886, + 0.003352002939209342, + 0.005335804540663958, + 0.023153195157647133, + 0.022386584430933, + 0.018676690757274628, + 0.009416391141712666, + 0.012871545739471912, + 0.00983237661421299, + 0.011952085420489311, + 0.017830343917012215, + 0.006426219828426838, + 0.010171874426305294, + 0.009751962497830391, + 0.007722676265984774, + 0.013496320694684982, + 0.009223729372024536, + 0.008192867040634155, + 0.00966891460120678, + 0.05712633952498436, + 0.01892467774450779, + 0.06755285710096359, + 0.010570978745818138, + 0.014439661987125874, + 0.022386500611901283, + 0.010508394800126553, + 0.012640658766031265, + 0.005759926047176123, + 0.006896335631608963, + 0.007249700836837292, + 0.004287182353436947, + 0.008567544631659985, + 0.013369602151215076, + 0.005807339679449797, + 0.006940972525626421, + 0.0067081269808113575, + 0.006251945160329342, + 0.0061033256351947784, + 0.005046428646892309, + 0.005553229711949825, + 0.010694675147533417, + 0.005554059986025095, + 0.005602032411843538, + 0.011769132688641548, + 0.011534271761775017, + 0.016213903203606606, + 0.007019601296633482, + 0.007160905282944441, + 0.008170715533196926, + 0.007575905881822109, + 0.01222570426762104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.00819061417132616 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033708471804857254 + ], + "opposite_logprob_tokens": [ + -3.4087085723876953 + ], + "logprob_reward": -0.033708471804857254, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033708471804857254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4423889219760895, + 0.047825660556554794, + 0.03906795382499695, + 0.02291465736925602, + 0.23019926249980927, + 0.028251133859157562, + 0.11485784500837326, + 0.013726972974836826, + 0.03250455856323242, + 0.02826303243637085 + ], + "credit_map": [ + 0.0584724098443985, + 0.04780452698469162, + 0.05952833592891693, + 0.048515331000089645, + 0.05021766200661659, + 0.05409828945994377, + 0.06526331603527069, + 0.05848906189203262, + 0.004143236670643091, + 0.005931796971708536, + 0.005349497776478529, + 0.004914409015327692, + 0.00730481930077076, + 0.003922559320926666, + 0.00878523476421833, + 0.007474108599126339, + 0.0025322516448795795, + 0.0027690392453223467, + 0.002725475700572133, + 0.0030778038781136274, + 0.005520316772162914, + 0.009047210216522217, + 0.004025316797196865, + 0.009370538406074047, + 0.0019526174291968346, + 0.0014274996938183904, + 0.003760445397347212, + 0.0030173202976584435, + 0.004140777047723532, + 0.003460872219875455, + 0.003108514938503504, + 0.0020466104615479708, + 0.036485590040683746, + 0.03547956421971321, + 0.03322989121079445, + 0.021972401067614555, + 0.03212978318333626, + 0.01781262271106243, + 0.03157300502061844, + 0.021516399458050728, + 0.0029711017850786448, + 0.003024621168151498, + 0.003480293555185199, + 0.0030233643483370543, + 0.005625441204756498, + 0.0038982920814305544, + 0.0033973597455769777, + 0.002830658806487918, + 0.01793225109577179, + 0.027596086263656616, + 0.014090834185481071, + 0.0070424918085336685, + 0.010486919432878494, + 0.014760133810341358, + 0.00622535590082407, + 0.01672377623617649, + 0.0012048371136188507, + 0.00158028700388968, + 0.0016416520811617374, + 0.0011801542714238167, + 0.0020925356075167656, + 0.0027574317064136267, + 0.0017991241766139865, + 0.0014709519455209374, + 0.0032105250284075737, + 0.004073504824191332, + 0.004652190953493118, + 0.003238784847781062, + 0.004194853827357292, + 0.00553535670042038, + 0.0041691879741847515, + 0.0034301551058888435, + 0.007656043395400047, + 0.0023573345970362425, + 0.009363623335957527, + 0.0018301919335499406, + 0.002118340227752924, + 0.0013531219447031617, + 0.001613145461305976, + 0.001971231075003743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.033708471804857254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064496710896492 + ], + "opposite_logprob_tokens": [ + -4.141064643859863 + ], + "logprob_reward": -0.016064496710896492, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064496710896492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13214173913002014, + 0.04914064705371857, + 0.1272934228181839, + 0.04050326719880104, + 0.22110633552074432, + 0.06788039207458496, + 0.20593857765197754, + 0.044093016535043716, + 0.056739818304777145, + 0.055162861943244934 + ], + "credit_map": [ + 0.015620723366737366, + 0.01942429319024086, + 0.01696164906024933, + 0.013047960586845875, + 0.013959584757685661, + 0.013174470514059067, + 0.028737405315041542, + 0.011215655133128166, + 0.005860160104930401, + 0.007337560411542654, + 0.00789195392280817, + 0.0048883524723351, + 0.006401080172508955, + 0.006379624363034964, + 0.005519377999007702, + 0.004862537141889334, + 0.007219656836241484, + 0.010010280646383762, + 0.008026263676583767, + 0.010932104662060738, + 0.018651988357305527, + 0.029801853001117706, + 0.012820553965866566, + 0.029830725863575935, + 0.0036738247144967318, + 0.003869604552164674, + 0.006063671316951513, + 0.004620397463440895, + 0.008185811340808868, + 0.005113566294312477, + 0.004336390178650618, + 0.004639999475330114, + 0.03521803766489029, + 0.03211680427193642, + 0.029621267691254616, + 0.020744135603308678, + 0.0331069640815258, + 0.016093304380774498, + 0.034870445728302, + 0.019335390999913216, + 0.007526614237576723, + 0.007033331785351038, + 0.00989046785980463, + 0.007118194829672575, + 0.011880556121468544, + 0.009440675377845764, + 0.007533425930887461, + 0.007457128260284662, + 0.02720930613577366, + 0.049797579646110535, + 0.028689542785286903, + 0.010143335908651352, + 0.017561348155140877, + 0.03127850219607353, + 0.010229354724287987, + 0.03102959878742695, + 0.0054805888794362545, + 0.004971789661794901, + 0.005490910727530718, + 0.004983959719538689, + 0.0044852918945252895, + 0.009227300994098186, + 0.004575607366859913, + 0.004877569619566202, + 0.0067323180846869946, + 0.006038608029484749, + 0.005083661526441574, + 0.011315868236124516, + 0.005599338561296463, + 0.006931958720088005, + 0.009536266326904297, + 0.005501800682395697, + 0.008703085593879223, + 0.008025161921977997, + 0.010432137176394463, + 0.004800807684659958, + 0.006935551762580872, + 0.00614622887223959, + 0.004763728938996792, + 0.005356159992516041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.016064496710896492 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5020136833190918 + ], + "opposite_logprob_tokens": [ + -0.2520137429237366 + ], + "logprob_reward": -1.5020136833190918, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5020136833190918, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.02292679063975811, + 0.03213120996952057, + 0.010224058292806149, + 0.015647824853658676, + 0.018441475927829742, + 0.01967119798064232, + 0.7849440574645996, + 0.022581715136766434, + 0.06230813264846802, + 0.011123568750917912 + ], + "credit_map": [ + 0.003214158583432436, + 0.0025137329939752817, + 0.0030948403291404247, + 0.0020174472592771053, + 0.0024450267665088177, + 0.0024213725700974464, + 0.0030416005756706, + 0.00417861295863986, + 0.004942476749420166, + 0.002866474213078618, + 0.004558229353278875, + 0.002086872700601816, + 0.0034970322158187628, + 0.005989499390125275, + 0.002349795773625374, + 0.005840827710926533, + 0.0007998687797226012, + 0.0008067460148595273, + 0.0008643621113151312, + 0.0006877785781398416, + 0.0018026628531515598, + 0.002132238121703267, + 0.0009144007926806808, + 0.0022160012740641832, + 0.0015898827696219087, + 0.0017515033250674605, + 0.0026068121660500765, + 0.001930398866534233, + 0.002301927190274, + 0.0019616903737187386, + 0.0015394052024930716, + 0.0019662044942378998, + 0.002718335250392556, + 0.0027766490820795298, + 0.0025304032023996115, + 0.0011594797251746058, + 0.0020582620054483414, + 0.00214777491055429, + 0.0015203390503302217, + 0.003530233632773161, + 0.0020231681410223246, + 0.002286591799929738, + 0.002640602644532919, + 0.0020304543431848288, + 0.0031885476782917976, + 0.0028494035359472036, + 0.001954043284058571, + 0.0026983851566910744, + 0.12791365385055542, + 0.19639995694160461, + 0.09841253608465195, + 0.0424964539706707, + 0.06299158930778503, + 0.08152446150779724, + 0.03860732540488243, + 0.13659806549549103, + 0.002777209272608161, + 0.002920430153608322, + 0.0037559804040938616, + 0.001991251250728965, + 0.002196402521803975, + 0.004452744498848915, + 0.0018353848718106747, + 0.0026523133274167776, + 0.007396477274596691, + 0.005837219301611185, + 0.005975235253572464, + 0.010273274034261703, + 0.006767771672457457, + 0.006665571592748165, + 0.012221122160553932, + 0.0071714590303599834, + 0.002484266646206379, + 0.0010949388379231095, + 0.0030712515581399202, + 0.0009117817389778793, + 0.000990999978967011, + 0.0007751455414108932, + 0.0009299610392190516, + 0.0008652237593196332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -1.5020136833190918 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606719382107258 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.01606719382107258, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606719382107258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029774120077490807, + 0.038127537816762924, + 0.021058406680822372, + 0.01423654705286026, + 0.041773103177547455, + 0.019622212275862694, + 0.7334052324295044, + 0.03608480095863342, + 0.055344998836517334, + 0.010573039762675762 + ], + "credit_map": [ + 0.004207052290439606, + 0.0033819989766925573, + 0.0043989126570522785, + 0.003118951804935932, + 0.0026622824370861053, + 0.0033440100960433483, + 0.004226047545671463, + 0.004434871021658182, + 0.006515355780720711, + 0.003794396761804819, + 0.004815617110580206, + 0.0019896305166184902, + 0.0036742002703249454, + 0.007457556668668985, + 0.002784301759675145, + 0.007096482906490564, + 0.0022442785557359457, + 0.0023227378260344267, + 0.0031979489140212536, + 0.002088270615786314, + 0.0032762859482318163, + 0.0031005165074020624, + 0.0018302482785657048, + 0.0029981217812746763, + 0.001408080686815083, + 0.0011633278336375952, + 0.002881786087527871, + 0.0014435198390856385, + 0.0015049910871312022, + 0.002648330293595791, + 0.0011105539742857218, + 0.0020759599283337593, + 0.005215436220169067, + 0.005510897841304541, + 0.0063634952530264854, + 0.00260367663577199, + 0.0038749955128878355, + 0.005595840513706207, + 0.0032669391948729753, + 0.009341824799776077, + 0.0018482772866263986, + 0.0026482450775802135, + 0.002818136941641569, + 0.001818123972043395, + 0.0033126631751656532, + 0.002853006822988391, + 0.0018865644233301282, + 0.002437195274978876, + 0.12344443798065186, + 0.18254555761814117, + 0.08982079476118088, + 0.03929378092288971, + 0.058683935552835464, + 0.07517094165086746, + 0.03584790602326393, + 0.1285979151725769, + 0.006198083981871605, + 0.0047765509225428104, + 0.007507755421102047, + 0.002755269641056657, + 0.0030520488508045673, + 0.0056406715884804726, + 0.002111536916345358, + 0.004042886663228273, + 0.0057134805247187614, + 0.0043733324855566025, + 0.005434657447040081, + 0.009799621999263763, + 0.006075382232666016, + 0.005647888872772455, + 0.011356671340763569, + 0.006943971384316683, + 0.001512050279416144, + 0.0014395726611837745, + 0.0021807444281876087, + 0.0012012593215331435, + 0.0011812892043963075, + 0.0009969206294044852, + 0.0010356109123677015, + 0.0010255936067551374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.01606719382107258 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "In a vibrant pop art style, a rapid 360-degree orbiting shot rotates around the featured bottle of fruit-flavored beer, capturing its exaggerated exhaustion from every angle. Made of glass, but depicted flatly in pale yellow with a bold black outline, the bottle features a simple cartoon face, with two downward-curving black lines representing closed, sleepy eyes. As the camera rotates, the bottle's mouth, the metal cap area, stretches impossibly in a vast, silent yawn, its glass body slightly distorting to emphasize the effort. A stream of perfectly circular, graphic-style bubbles, each with a single white highlight, escapes the opening and floats upward against the dynamic background. The setting is a non-representational explosion of color, with a flat, electric blue background overlaid with a precise grid of large, red Ben-Day dots, creating a stark graphic contrast. The entire scene is illuminated by a bright, even, shadowless light, reinforcing the bland comic book aesthetic and creating a playful, high-energy atmosphere tinged with cartoon fatigue.", + "seed": 3972, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_7/index_4701_seed_3972_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.39007097156718373, + "vlm_logprob": -0.39007097156718373, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.39007097156718373, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007737545762211084 + ], + "opposite_logprob_tokens": [ + -5.257737636566162 + ], + "logprob_reward": -0.007737545762211084, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.007737545762211084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12251918762922287, + 0.05432892218232155, + 0.07807173579931259, + 0.23041781783103943, + 0.05275668203830719, + 0.08097437024116516, + 0.0486518032848835, + 0.15275587141513824, + 0.04709262400865555, + 0.1324310004711151 + ], + "credit_map": [ + 0.014823454432189465, + 0.01765318773686886, + 0.013122196309268475, + 0.010873162187635899, + 0.020384782925248146, + 0.023144792765378952, + 0.008919769898056984, + 0.013597848825156689, + 0.008024060167372227, + 0.005327937658876181, + 0.007728028576821089, + 0.00800042599439621, + 0.004760411102324724, + 0.008748477324843407, + 0.006181387696415186, + 0.0055581917986273766, + 0.0075208768248558044, + 0.010191764682531357, + 0.0067182015627622604, + 0.00959005020558834, + 0.009253213182091713, + 0.0073271458968520164, + 0.018459070473909378, + 0.00901141855865717, + 0.02080957032740116, + 0.020344668999314308, + 0.021655820310115814, + 0.02237670123577118, + 0.03535139933228493, + 0.04058234393596649, + 0.03796522319316864, + 0.031332094222307205, + 0.007004988379776478, + 0.0031194796320050955, + 0.007781148422509432, + 0.003680666908621788, + 0.00539943715557456, + 0.010286353528499603, + 0.005304632242769003, + 0.010179977864027023, + 0.00643231812864542, + 0.015238537453114986, + 0.0048581259325146675, + 0.006179250776767731, + 0.025202682241797447, + 0.006352207623422146, + 0.0059673115611076355, + 0.010743940249085426, + 0.00869087316095829, + 0.006624921225011349, + 0.004875711631029844, + 0.005487137008458376, + 0.007168188225477934, + 0.005247584544122219, + 0.005759493913501501, + 0.004797894041985273, + 0.019126219674944878, + 0.0225642379373312, + 0.014385866932570934, + 0.010922777466475964, + 0.05771239474415779, + 0.009594039991497993, + 0.01039318647235632, + 0.008057152852416039, + 0.004662840627133846, + 0.005096680484712124, + 0.0037171593867242336, + 0.008367674425244331, + 0.005105365067720413, + 0.007409061770886183, + 0.006186920218169689, + 0.006546921096742153, + 0.014383944682776928, + 0.052106570452451706, + 0.006529460661113262, + 0.010869813151657581, + 0.013258113525807858, + 0.0063935620710253716, + 0.017944594845175743, + 0.010944933630526066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 9 + ], + "selected_reward": -0.007737545762211084 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632262073457241 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632262073457241, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632262073457241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14015819132328033, + 0.13063006103038788, + 0.03484102338552475, + 0.08865986764431, + 0.18271316587924957, + 0.13505564630031586, + 0.08582381159067154, + 0.09830385446548462, + 0.059717174619436264, + 0.0440971702337265 + ], + "credit_map": [ + 0.014631210826337337, + 0.0220049936324358, + 0.018183404579758644, + 0.01503551471978426, + 0.017613328993320465, + 0.022130431607365608, + 0.012258567847311497, + 0.018300747498869896, + 0.013010963797569275, + 0.036433812230825424, + 0.017736680805683136, + 0.011648274026811123, + 0.01681886427104473, + 0.012701650150120258, + 0.009473927319049835, + 0.01280590146780014, + 0.003824612358585, + 0.003108509350568056, + 0.00295903324149549, + 0.00473403325304389, + 0.0029009815771132708, + 0.0038133650086820126, + 0.00879471655935049, + 0.0047057741321623325, + 0.011130042374134064, + 0.012508764863014221, + 0.006578364875167608, + 0.0072768270038068295, + 0.011612343601882458, + 0.020284553989768028, + 0.007171133533120155, + 0.012097842991352081, + 0.021065281704068184, + 0.027454707771539688, + 0.017139730975031853, + 0.008177808485925198, + 0.0615105926990509, + 0.016514290124177933, + 0.01082445215433836, + 0.020026328042149544, + 0.012520630843937397, + 0.022970646619796753, + 0.006835772190243006, + 0.009550963528454304, + 0.04375699907541275, + 0.01232755184173584, + 0.00862619373947382, + 0.018466902896761894, + 0.01357853040099144, + 0.009493005461990833, + 0.009724984876811504, + 0.008483294397592545, + 0.00843796692788601, + 0.00960591621696949, + 0.0153876394033432, + 0.011112483218312263, + 0.011819319799542427, + 0.01420014351606369, + 0.009916609153151512, + 0.0071623860858380795, + 0.03358042240142822, + 0.008223475888371468, + 0.007280777208507061, + 0.006120723206549883, + 0.005962098482996225, + 0.00704009598121047, + 0.006525113247334957, + 0.009000053629279137, + 0.004883675370365381, + 0.011906635947525501, + 0.008015294559299946, + 0.006384212989360094, + 0.005280024372041225, + 0.004875394981354475, + 0.004744763020426035, + 0.005696232430636883, + 0.0023919660598039627, + 0.0038641884457319975, + 0.009617157280445099, + 0.007627448998391628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.007632262073457241 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016061095520853996 + ], + "opposite_logprob_tokens": [ + -4.141061305999756 + ], + "logprob_reward": -0.016061095520853996, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016061095520853996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05724053829908371, + 0.06752101331949234, + 0.07966815680265427, + 0.057934656739234924, + 0.21964502334594727, + 0.1694972962141037, + 0.07133426517248154, + 0.048934392631053925, + 0.03360426425933838, + 0.19462040066719055 + ], + "credit_map": [ + 0.007328318897634745, + 0.008642695844173431, + 0.0059328568167984486, + 0.006830455735325813, + 0.00972756091505289, + 0.008216873742640018, + 0.004734255839139223, + 0.005827521439641714, + 0.009199852123856544, + 0.0187041237950325, + 0.009789896197617054, + 0.00449352664873004, + 0.008556892164051533, + 0.006869462784379721, + 0.004697903525084257, + 0.0052093579433858395, + 0.007843835279345512, + 0.009005033411085606, + 0.008285743184387684, + 0.009530842304229736, + 0.00870880950242281, + 0.0076049696654081345, + 0.015289907343685627, + 0.013399017974734306, + 0.006533319596201181, + 0.006306163500994444, + 0.003977441694587469, + 0.004907165188342333, + 0.008431564085185528, + 0.012783532962203026, + 0.0067277466878294945, + 0.008267722092568874, + 0.026153309270739555, + 0.031211931258440018, + 0.01753072626888752, + 0.010693511925637722, + 0.06803902238607407, + 0.02565472200512886, + 0.013372453860938549, + 0.026989342644810677, + 0.011218653060495853, + 0.03149701654911041, + 0.007235804107040167, + 0.008859378285706043, + 0.061752013862133026, + 0.011826297268271446, + 0.009937867522239685, + 0.027170266956090927, + 0.014883492141962051, + 0.008861295878887177, + 0.009093474596738815, + 0.00572447432205081, + 0.010356990620493889, + 0.007631762884557247, + 0.005718323867768049, + 0.009064452722668648, + 0.006329741328954697, + 0.007399664726108313, + 0.004875452723354101, + 0.0028548447880893946, + 0.015541884116828442, + 0.0051821316592395306, + 0.003230839967727661, + 0.003519836813211441, + 0.005048966500908136, + 0.0037424310576170683, + 0.003149349009618163, + 0.005881627090275288, + 0.002722111763432622, + 0.004742978606373072, + 0.0043901680037379265, + 0.003926629200577736, + 0.01731375977396965, + 0.07233375310897827, + 0.008882032707333565, + 0.016817985102534294, + 0.016632815822958946, + 0.008572756312787533, + 0.04062777757644653, + 0.013439507223665714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.016061095520853996 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.910281777381897 + ], + "opposite_logprob_tokens": [ + -0.16028179228305817 + ], + "logprob_reward": -1.910281777381897, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.910281777381897, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14227910339832306, + 0.09677625447511673, + 0.11707693338394165, + 0.0657891258597374, + 0.13960765302181244, + 0.1290029138326645, + 0.14695176482200623, + 0.06947499513626099, + 0.06561898440122604, + 0.027422303333878517 + ], + "credit_map": [ + 0.01548060867935419, + 0.022011110559105873, + 0.012085759080946445, + 0.010827168822288513, + 0.028281226754188538, + 0.022129546850919724, + 0.012604904361069202, + 0.018858781084418297, + 0.013835039921104908, + 0.019461683928966522, + 0.015046682208776474, + 0.008342884480953217, + 0.010362942703068256, + 0.010844605974853039, + 0.00932571105659008, + 0.009556705132126808, + 0.011959888972342014, + 0.014938910491764545, + 0.010565700009465218, + 0.014616364613175392, + 0.013183784671127796, + 0.011424310505390167, + 0.027174517512321472, + 0.01321345940232277, + 0.006047837436199188, + 0.005261499434709549, + 0.007040593773126602, + 0.007948262616991997, + 0.0069817714393138885, + 0.006891331169754267, + 0.01584276556968689, + 0.009775065816938877, + 0.01521999016404152, + 0.018557673320174217, + 0.010187311097979546, + 0.007093555293977261, + 0.047257933765649796, + 0.012871851213276386, + 0.010669968090951443, + 0.017749357968568802, + 0.019151121377944946, + 0.022328460589051247, + 0.008133448660373688, + 0.007504077162593603, + 0.037293314933776855, + 0.010168558917939663, + 0.009196215309202671, + 0.01522771641612053, + 0.0378900021314621, + 0.01597854495048523, + 0.027880465611815453, + 0.012861860916018486, + 0.013402841053903103, + 0.012107840739190578, + 0.014415786601603031, + 0.012414420954883099, + 0.00882265530526638, + 0.009790812619030476, + 0.005679752677679062, + 0.004163722507655621, + 0.0226273275911808, + 0.006042314227670431, + 0.006146313156932592, + 0.0062020947225391865, + 0.008575599640607834, + 0.008038477040827274, + 0.005689206067472696, + 0.008806047961115837, + 0.0064965891651809216, + 0.011102860793471336, + 0.009682008996605873, + 0.007228191010653973, + 0.002660411177203059, + 0.006589286029338837, + 0.002716431859880686, + 0.003119691740721464, + 0.0029219624120742083, + 0.0035641449503600597, + 0.0024621605407446623, + 0.0033882148563861847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -1.910281777381897 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bottle of fruit-flavored beer depicted in a vibrant pop art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle made of glass, depicted flatly in pale yellow with a bold black outline, and featuring a cartoon face with closed, sleepy eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bottle perform an exaggerated yawn where its mouth and metal cap area stretch impossibly while the glass body distorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a rapid 360-degree orbiting shot around the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flat electric blue with a precise grid of large red Ben-Day dots, illuminated by bright, even, shadowless light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864217709749937 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.00864217709749937, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864217709749937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1219530999660492, + 0.053692884743213654, + 0.04790755361318588, + 0.088943250477314, + 0.041176095604896545, + 0.07763824611902237, + 0.1362149715423584, + 0.08558173477649689, + 0.08716317266225815, + 0.25972893834114075 + ], + "credit_map": [ + 0.011127011850476265, + 0.018754035234451294, + 0.014229752123355865, + 0.012250838801264763, + 0.019818298518657684, + 0.0171528123319149, + 0.012952323071658611, + 0.015668032690882683, + 0.008651715703308582, + 0.007291844580322504, + 0.008987382054328918, + 0.004317425657063723, + 0.0052338093519210815, + 0.005717828404158354, + 0.006161681842058897, + 0.0073311952874064445, + 0.00466544134542346, + 0.005530378315597773, + 0.005572305992245674, + 0.00488334521651268, + 0.004709416534751654, + 0.004730281885713339, + 0.009859448298811913, + 0.00795693974941969, + 0.007191777229309082, + 0.007747636176645756, + 0.008746621198952198, + 0.009105239063501358, + 0.011959141120314598, + 0.012209141626954079, + 0.019014067947864532, + 0.012969637289643288, + 0.005449631251394749, + 0.0037339963018894196, + 0.004842061083763838, + 0.003232513787224889, + 0.007124049589037895, + 0.006197002250701189, + 0.0038875965401530266, + 0.006709245964884758, + 0.008120562881231308, + 0.014196265488862991, + 0.004356409423053265, + 0.005072714760899544, + 0.024429213255643845, + 0.006174208130687475, + 0.005904782563447952, + 0.009384092874825, + 0.03469046577811241, + 0.01376399863511324, + 0.02523844875395298, + 0.011051740497350693, + 0.01327687781304121, + 0.011338230222463608, + 0.01506265252828598, + 0.011792563833296299, + 0.010480431839823723, + 0.012107609771192074, + 0.008114535361528397, + 0.005416492931544781, + 0.0299041997641325, + 0.007214013021439314, + 0.006858221720904112, + 0.005486230365931988, + 0.011194683611392975, + 0.008677848614752293, + 0.008304405957460403, + 0.014308554120361805, + 0.007033901289105415, + 0.011816892772912979, + 0.015096157789230347, + 0.010730724781751633, + 0.022256655618548393, + 0.09122329205274582, + 0.013386341743171215, + 0.027247389778494835, + 0.020345449447631836, + 0.0107407933101058, + 0.05780597776174545, + 0.016723021864891052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.00864217709749937 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_8/index_tensor([9546, 9546])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_8/index_tensor([9546, 9546])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..78aee33bf7d152f245a9ef8f88734440d8b639ea --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_8/index_tensor([9546, 9546])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly low-angle wide shot, filmed with the crisp, high-fidelity quality of a historical documentary, surveys the scene from the foot of a long, dark gravel driveway. The camera is positioned to capture the unbroken, imposing facade of a Gothic Revival mansion constructed of weathered gray stone beneath a cool, cloudy sky. Flanked by neatly trimmed, dark-green yew hedges, the driveway draws the eye toward the central entrance. The central focal point is a pair of massive, ornate oak doors, their dark wood rich with an ancient patina, adorned with intricate pointed arches and square carvings, fitted with heavy black cast iron hinges and a massive, powerful ring handle. Without a sound, the doors begin to swing inward slowly and deliberately, their majestic movement revealing a small glimpse of the grand, shadowed entrance hall within. As the gap widens, the deep, cool darkness of the interior contrasts with the soft daylight, revealing a glimpse of the massive stone flagstone floor and the base of the winding staircase disappearing into the darkness of the upper floors, creating an atmosphere of solemn reverence and profound historical anticipation.", + "seed": 31696, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_8/index_9546_seed_31696_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.21726338975131512, + "vlm_logprob": -0.21726338975131512, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.21726338975131512, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.897800087928772 + ], + "opposite_logprob_tokens": [ + -0.647800087928772 + ], + "logprob_reward": -0.897800087928772, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.897800087928772, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2065661996603012, + 0.04555079713463783, + 0.08644048869609833, + 0.04473433643579483, + 0.05614562705159187, + 0.03056197613477707, + 0.4201015532016754, + 0.03130233660340309, + 0.052706293761730194, + 0.025890326127409935 + ], + "credit_map": [ + 0.020104942843317986, + 0.013487755320966244, + 0.04327697679400444, + 0.026881584897637367, + 0.02435406856238842, + 0.02140520140528679, + 0.019978877156972885, + 0.03707680106163025, + 0.003512629307806492, + 0.005832729861140251, + 0.0058791302144527435, + 0.0067717838101089, + 0.006724974140524864, + 0.004282576497644186, + 0.006984650623053312, + 0.005562321748584509, + 0.012247828766703606, + 0.007076449692249298, + 0.026770204305648804, + 0.00917920470237732, + 0.0060431333258748055, + 0.0057757459580898285, + 0.006014302838593721, + 0.013333621434867382, + 0.0033902153372764587, + 0.0036071916110813618, + 0.0066912309266626835, + 0.010871565900743008, + 0.003783764783293009, + 0.005169426091015339, + 0.0048087239265441895, + 0.006412217393517494, + 0.008542823605239391, + 0.0058351741172373295, + 0.008315467275679111, + 0.00815561506897211, + 0.004229757934808731, + 0.007547538261860609, + 0.005402534268796444, + 0.00811671745032072, + 0.0028174053877592087, + 0.003320202464237809, + 0.005180599167943001, + 0.003771263174712658, + 0.005249157547950745, + 0.0026449181605130434, + 0.0035503441467881203, + 0.00402808515354991, + 0.0315844863653183, + 0.013522627763450146, + 0.16329985857009888, + 0.022650908678770065, + 0.008900060318410397, + 0.01638852432370186, + 0.011142361909151077, + 0.15261273086071014, + 0.002321380190551281, + 0.0022476473823189735, + 0.006260741502046585, + 0.003965670708566904, + 0.00290879188105464, + 0.0034530116245150566, + 0.003522425889968872, + 0.006622666493058205, + 0.004625505302101374, + 0.004833924118429422, + 0.007755305152386427, + 0.011811385862529278, + 0.005571222398430109, + 0.0034147407859563828, + 0.00880381092429161, + 0.005890401545912027, + 0.00264212884940207, + 0.0026570286136120558, + 0.003549338551238179, + 0.0031388134229928255, + 0.0036651797126978636, + 0.0030804285779595375, + 0.0029831037390977144, + 0.0041743069887161255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.897800087928772 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10038288682699203 + ], + "opposite_logprob_tokens": [ + -2.3503828048706055 + ], + "logprob_reward": -0.10038288682699203, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10038288682699203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2082294374704361, + 0.12399406731128693, + 0.20164239406585693, + 0.09189268946647644, + 0.01788424700498581, + 0.03835944086313248, + 0.21549595892429352, + 0.04105907306075096, + 0.03140503540635109, + 0.03003762476146221 + ], + "credit_map": [ + 0.021267643198370934, + 0.01628587581217289, + 0.055612098425626755, + 0.0239866953343153, + 0.016365690156817436, + 0.013936790637671947, + 0.019064748659729958, + 0.041709933429956436, + 0.01675482466816902, + 0.011324835009872913, + 0.02292105183005333, + 0.01562551222741604, + 0.012478657998144627, + 0.012311364524066448, + 0.01654817722737789, + 0.016029657796025276, + 0.02278924360871315, + 0.02221217006444931, + 0.051043905317783356, + 0.024395743384957314, + 0.024456650018692017, + 0.014899891801178455, + 0.014770371839404106, + 0.0270744226872921, + 0.009300700388848782, + 0.011185724288225174, + 0.017739025875926018, + 0.01869596540927887, + 0.008644168265163898, + 0.005342496559023857, + 0.009625419043004513, + 0.011359203606843948, + 0.002431967994198203, + 0.002190357306972146, + 0.0027726192492991686, + 0.001913820393383503, + 0.0021424393635243177, + 0.002248656703159213, + 0.0015770223690196872, + 0.002607363974675536, + 0.0036803565453737974, + 0.0033786895219236612, + 0.007746838498860598, + 0.004389165434986353, + 0.0029210406355559826, + 0.00521821016445756, + 0.003975411877036095, + 0.0070497337728738785, + 0.016549117863178253, + 0.008621742948889732, + 0.08023277670145035, + 0.013852671720087528, + 0.007017751690000296, + 0.007519345264881849, + 0.008160261437296867, + 0.07354229688644409, + 0.0030194346327334642, + 0.004753931891173124, + 0.006143204402178526, + 0.005890377331525087, + 0.004398979712277651, + 0.004896258469671011, + 0.004102890379726887, + 0.007853999733924866, + 0.002886233851313591, + 0.002773275598883629, + 0.005681138020008802, + 0.005075458902865648, + 0.002514007268473506, + 0.0030196814332157373, + 0.004413554444909096, + 0.005041688214987516, + 0.00241642352193594, + 0.004781510215252638, + 0.0036115841940045357, + 0.003912199754267931, + 0.004096732009202242, + 0.0033852928318083286, + 0.00321534788236022, + 0.004618535749614239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.10038288682699203 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012563606724143028 + ], + "opposite_logprob_tokens": [ + -4.387563705444336 + ], + "logprob_reward": -0.012563606724143028, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012563606724143028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19037197530269623, + 0.07653253525495529, + 0.0921308621764183, + 0.027286581695079803, + 0.02140169031918049, + 0.04042503610253334, + 0.3661889433860779, + 0.07403366267681122, + 0.038374342024326324, + 0.07325436174869537 + ], + "credit_map": [ + 0.021431157365441322, + 0.008062762208282948, + 0.059400081634521484, + 0.020026762038469315, + 0.012891831807792187, + 0.012793791480362415, + 0.014090951532125473, + 0.04167463257908821, + 0.008002623915672302, + 0.004338670521974564, + 0.021767672151327133, + 0.007237576879560947, + 0.006177644710987806, + 0.006003174465149641, + 0.007452900055795908, + 0.015552271157503128, + 0.012676974758505821, + 0.005667638033628464, + 0.0320596806704998, + 0.00820170994848013, + 0.005204619839787483, + 0.006157963071018457, + 0.005050584673881531, + 0.017111698165535927, + 0.0035489280708134174, + 0.0023633723612874746, + 0.0039972830563783646, + 0.005077788606286049, + 0.0028751096688210964, + 0.0032474545296281576, + 0.002516903216019273, + 0.0036597424186766148, + 0.0015543760964646935, + 0.0025049590039998293, + 0.0032876566983759403, + 0.00406144792214036, + 0.00232395576313138, + 0.0020847611594945192, + 0.0026503342669457197, + 0.0029341988265514374, + 0.0032653717789798975, + 0.003744046436622739, + 0.008357079699635506, + 0.0061645410023629665, + 0.0051302253268659115, + 0.003969607874751091, + 0.00383352255448699, + 0.005960641894489527, + 0.027789613232016563, + 0.013610944151878357, + 0.13824757933616638, + 0.022112267091870308, + 0.010865007527172565, + 0.013185451738536358, + 0.012421408668160439, + 0.12795667350292206, + 0.007629780098795891, + 0.007675663102418184, + 0.013789910823106766, + 0.01064220629632473, + 0.00675022741779685, + 0.007184093352407217, + 0.007781575433909893, + 0.012580208480358124, + 0.0035623519215732813, + 0.005846323445439339, + 0.005041902884840965, + 0.005580550525337458, + 0.004127293825149536, + 0.003170158015564084, + 0.006261810194700956, + 0.0047839549370110035, + 0.005517028272151947, + 0.011043244041502476, + 0.009838566184043884, + 0.00941928569227457, + 0.00806917529553175, + 0.006180441938340664, + 0.013015097938477993, + 0.010171518661081791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.012563606724143028 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059643715620041 + ], + "opposite_logprob_tokens": [ + -3.895596504211426 + ], + "logprob_reward": -0.02059643715620041, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059643715620041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09285633265972137, + 0.04891436547040939, + 0.128239706158638, + 0.032694343477487564, + 0.08171562850475311, + 0.05546814948320389, + 0.18786592781543732, + 0.1411692351102829, + 0.16883166134357452, + 0.062244631350040436 + ], + "credit_map": [ + 0.012611255049705505, + 0.005864392966032028, + 0.025635335594415665, + 0.008811811916530132, + 0.010218576528131962, + 0.006543691270053387, + 0.007155219558626413, + 0.016016041859984398, + 0.0037330174818634987, + 0.006540299393236637, + 0.006061290856450796, + 0.005913487169891596, + 0.009735082276165485, + 0.00430566119030118, + 0.0066015333868563175, + 0.006023990456014872, + 0.01972654089331627, + 0.010845735669136047, + 0.03541098162531853, + 0.013687312602996826, + 0.012238331139087677, + 0.009900935925543308, + 0.008379056118428707, + 0.018050802871584892, + 0.005839708726853132, + 0.0030453552026301622, + 0.0060161929577589035, + 0.004662180319428444, + 0.0025433781556785107, + 0.0034758050460368395, + 0.002450519474223256, + 0.00466120382770896, + 0.012276985682547092, + 0.008001953363418579, + 0.012675873003900051, + 0.015285476110875607, + 0.0064058611169457436, + 0.00954840611666441, + 0.007507955655455589, + 0.010013123974204063, + 0.005330572370439768, + 0.00650213286280632, + 0.009226575493812561, + 0.00689734285697341, + 0.00871293991804123, + 0.006312057375907898, + 0.004480401519685984, + 0.008006127551198006, + 0.012792644090950489, + 0.0067882658913731575, + 0.06971511244773865, + 0.009594352915883064, + 0.005196995101869106, + 0.009737159125506878, + 0.005265555810183287, + 0.06877584010362625, + 0.013207999989390373, + 0.0055029126815497875, + 0.04670914635062218, + 0.010996358469128609, + 0.004641713108867407, + 0.009494860656559467, + 0.007249316666275263, + 0.0433669276535511, + 0.01737363077700138, + 0.017964035272598267, + 0.02388836070895195, + 0.029019299894571304, + 0.01826416701078415, + 0.012918231077492237, + 0.027163086459040642, + 0.022240841761231422, + 0.004899968393146992, + 0.0061574093997478485, + 0.008030176162719727, + 0.007021509110927582, + 0.008036051876842976, + 0.005616785492748022, + 0.010966351255774498, + 0.011516377329826355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.02059643715620041 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497393012046814 + ], + "opposite_logprob_tokens": [ + -2.929973840713501 + ], + "logprob_reward": -0.05497393012046814, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497393012046814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25331911444664, + 0.051623474806547165, + 0.05342787876725197, + 0.032622963190078735, + 0.02546234242618084, + 0.037029802799224854, + 0.27305588126182556, + 0.22077254951000214, + 0.02999822236597538, + 0.022687774151563644 + ], + "credit_map": [ + 0.025439219549298286, + 0.013128502294421196, + 0.07544243335723877, + 0.022958217188715935, + 0.018829302862286568, + 0.02060188725590706, + 0.015056009404361248, + 0.06186358258128166, + 0.004496907349675894, + 0.0038197431713342667, + 0.012773995287716389, + 0.005561511497944593, + 0.0038054119795560837, + 0.003964222501963377, + 0.005207103211432695, + 0.01199458260089159, + 0.00811118446290493, + 0.0035767180379480124, + 0.01706628128886223, + 0.005097789224237204, + 0.0036576841957867146, + 0.004346739500761032, + 0.003218925790861249, + 0.008352558128535748, + 0.004104013554751873, + 0.003054804867133498, + 0.006519102491438389, + 0.004174904432147741, + 0.0027447238098829985, + 0.004273813683539629, + 0.002630266360938549, + 0.005121334921568632, + 0.0023053595796227455, + 0.002254295162856579, + 0.003942619077861309, + 0.003128982847556472, + 0.002103652572259307, + 0.003432653844356537, + 0.002161977579817176, + 0.006132801994681358, + 0.003279985161498189, + 0.003065818687900901, + 0.007722064852714539, + 0.004615359008312225, + 0.0037066370714455843, + 0.005846572574228048, + 0.0034199883230030537, + 0.005373378749936819, + 0.019886983558535576, + 0.008025968447327614, + 0.10581223666667938, + 0.014578751288354397, + 0.005346772260963917, + 0.011719132773578167, + 0.006768461316823959, + 0.10091758519411087, + 0.021266113966703415, + 0.015163932926952839, + 0.06327846646308899, + 0.01709700934588909, + 0.008034314028918743, + 0.02651360258460045, + 0.008716019801795483, + 0.0607030987739563, + 0.003112280275672674, + 0.0023521415423601866, + 0.005212027579545975, + 0.004739623516798019, + 0.003360411152243614, + 0.002928681904450059, + 0.0031193075701594353, + 0.005173750687390566, + 0.0019598237704485655, + 0.0031254098284989595, + 0.0031834780238568783, + 0.0032135818619281054, + 0.0024194668512791395, + 0.002674282528460026, + 0.002382769249379635, + 0.0037289655301719904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.05497393012046814 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle wide shot, filmed with the crisp, high-fidelity quality of a historical documentary, surveys the scene from the foot of a long, dark gravel driveway. The camera is positioned to capture the unbroken, imposing facade of a Gothic Revival mansion constructed of weathered gray stone beneath a cool, cloudy sky. Flanked by neatly trimmed, dark-green yew hedges, the driveway draws the eye toward the central entrance. The central focal point is a pair of massive, ornate oak doors, their dark wood rich with an ancient patina, adorned with intricate pointed arches and square carvings, fitted with heavy black cast iron hinges and a massive, powerful ring handle. Without a sound, the doors begin to swing inward slowly and deliberately, their majestic movement revealing a small glimpse of the grand, shadowed entrance hall within. As the gap widens, the deep, cool darkness of the interior contrasts with the soft daylight, revealing a glimpse of the massive stone flagstone floor and the base of the winding staircase disappearing into the darkness of the upper floors, creating an atmosphere of solemn reverence and profound historical anticipation.", + "seed": 31697, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_8/index_9546_seed_31697_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07786451913416385, + "vlm_logprob": -0.07786451913416385, + "vlm_yes_no_margin": 2.9999999523162844, + "vlm_token_credit": -0.07786451913416385, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24564477801322937 + ], + "opposite_logprob_tokens": [ + -1.7456448078155518 + ], + "logprob_reward": -0.24564477801322937, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.24564477801322937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08918347209692001, + 0.10818490386009216, + 0.029706155881285667, + 0.08366398513317108, + 0.2371511608362198, + 0.10394089668989182, + 0.1044517233967781, + 0.046086639165878296, + 0.1467856764793396, + 0.050845373421907425 + ], + "credit_map": [ + 0.009009371511638165, + 0.010185526683926582, + 0.00973912701010704, + 0.009932396002113819, + 0.013655250892043114, + 0.020106012001633644, + 0.00778555404394865, + 0.00877023208886385, + 0.010241788811981678, + 0.00677043991163373, + 0.03348570317029953, + 0.009946412406861782, + 0.008772377856075764, + 0.007689062040299177, + 0.006619448773562908, + 0.024659672752022743, + 0.0027029872871935368, + 0.004000257700681686, + 0.0040459404699504375, + 0.0035941696260124445, + 0.0038336424622684717, + 0.0030983486212790012, + 0.004705470521003008, + 0.003725337563082576, + 0.0033005455043166876, + 0.0024414481595158577, + 0.012859509326517582, + 0.005584895610809326, + 0.0060514118522405624, + 0.0028001023456454277, + 0.031809523701667786, + 0.018816553056240082, + 0.012023908086121082, + 0.03946705535054207, + 0.03240901976823807, + 0.024155834689736366, + 0.036547619849443436, + 0.03436242789030075, + 0.029972869902849197, + 0.02821243181824684, + 0.005836840718984604, + 0.015058252029120922, + 0.012935481034219265, + 0.011920609511435032, + 0.011593767441809177, + 0.014763826504349709, + 0.015149268321692944, + 0.016682852059602737, + 0.010797436349093914, + 0.008587794378399849, + 0.014472459442913532, + 0.011343391612172127, + 0.019117144867777824, + 0.008815981447696686, + 0.017131801694631577, + 0.014185719192028046, + 0.004214606247842312, + 0.004444051068276167, + 0.005591136869043112, + 0.005305099301040173, + 0.0056808083318173885, + 0.007259434089064598, + 0.007691079750657082, + 0.005900423973798752, + 0.012832366861402988, + 0.01110734324902296, + 0.0119633125141263, + 0.01755729876458645, + 0.024611787870526314, + 0.019552486017346382, + 0.024840261787176132, + 0.0243208184838295, + 0.003619085531681776, + 0.003994941711425781, + 0.004337389953434467, + 0.006227125879377127, + 0.00860351137816906, + 0.008574554696679115, + 0.00935003999620676, + 0.006138727068901062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.24564477801322937 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03374074399471283 + ], + "opposite_logprob_tokens": [ + -3.408740758895874 + ], + "logprob_reward": -0.03374074399471283, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03374074399471283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1791338473558426, + 0.0965106263756752, + 0.04534865543246269, + 0.07251501083374023, + 0.09521093964576721, + 0.06389845162630081, + 0.060133304446935654, + 0.12012014538049698, + 0.1590699851512909, + 0.10805902630090714 + ], + "credit_map": [ + 0.0179221760481596, + 0.021492620930075645, + 0.01986195705831051, + 0.02114473283290863, + 0.02626756951212883, + 0.03856181725859642, + 0.01739395037293434, + 0.016489015892148018, + 0.008880396373569965, + 0.004386909306049347, + 0.029901079833507538, + 0.011253345757722855, + 0.00694697443395853, + 0.00703329686075449, + 0.0077218711376190186, + 0.020386751741170883, + 0.00317121553234756, + 0.004496603272855282, + 0.006977829150855541, + 0.004171133507043123, + 0.006219205446541309, + 0.003774807555601001, + 0.006341245956718922, + 0.01019661221653223, + 0.003098235698416829, + 0.003681847592815757, + 0.015494306571781635, + 0.007227517198771238, + 0.006789823994040489, + 0.005010849330574274, + 0.024062300100922585, + 0.007150130346417427, + 0.005323760211467743, + 0.016344958916306496, + 0.010002847760915756, + 0.010895553976297379, + 0.02087593451142311, + 0.010962729342281818, + 0.011415593326091766, + 0.009389552287757397, + 0.0025631540920585394, + 0.008306729607284069, + 0.01478230394423008, + 0.00697566382586956, + 0.006737970281392336, + 0.0037781945429742336, + 0.015414545312523842, + 0.005339892115443945, + 0.0036827882286161184, + 0.005483332555741072, + 0.009368833154439926, + 0.0077558280900120735, + 0.009946823120117188, + 0.005928142461925745, + 0.008540197275578976, + 0.009427361190319061, + 0.009268345311284065, + 0.011522721499204636, + 0.017250433564186096, + 0.011287765577435493, + 0.016232842579483986, + 0.01357308216392994, + 0.018727172166109085, + 0.022257788106799126, + 0.013607983477413654, + 0.0124317966401577, + 0.01373367104679346, + 0.017457913607358932, + 0.02658087946474552, + 0.022671964019536972, + 0.026854807510972023, + 0.02573096752166748, + 0.012902114540338516, + 0.009476463310420513, + 0.013637076131999493, + 0.012069654650986195, + 0.018817178905010223, + 0.015943661332130432, + 0.012562244199216366, + 0.012650636024773121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.03374074399471283 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820392720401287 + ], + "opposite_logprob_tokens": [ + -4.0182037353515625 + ], + "logprob_reward": -0.01820392720401287, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01820392720401287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06746336072683334, + 0.04325681924819946, + 0.06093582138419151, + 0.06829504668712616, + 0.12113435566425323, + 0.08277842402458191, + 0.1405067890882492, + 0.13512928783893585, + 0.1934005320072174, + 0.08709963411092758 + ], + "credit_map": [ + 0.006290866993367672, + 0.006854832172393799, + 0.011435926891863346, + 0.008607115596532822, + 0.00952423270791769, + 0.0069740768522024155, + 0.009053833782672882, + 0.008722463622689247, + 0.0046095531433820724, + 0.005285215098410845, + 0.006104919593781233, + 0.005434446968138218, + 0.006461819633841515, + 0.004831555299460888, + 0.005602742545306683, + 0.004926562309265137, + 0.002977960277348757, + 0.006074490491300821, + 0.008462882600724697, + 0.0070836348459124565, + 0.010463221929967403, + 0.0035157648380845785, + 0.013707397505640984, + 0.008650460280478, + 0.003770409384742379, + 0.0028081447817385197, + 0.006557101849466562, + 0.005108862649649382, + 0.004853455349802971, + 0.0034164534881711006, + 0.019848251715302467, + 0.02193235419690609, + 0.006388879381120205, + 0.02208462543785572, + 0.01641148515045643, + 0.011494945734739304, + 0.020888620987534523, + 0.014699331484735012, + 0.01162413228303194, + 0.017542323097586632, + 0.004956486169248819, + 0.014423979446291924, + 0.012895478866994381, + 0.010333743877708912, + 0.012546639889478683, + 0.008140122517943382, + 0.010246393270790577, + 0.009235567413270473, + 0.00719074159860611, + 0.0111231729388237, + 0.015699848532676697, + 0.019731828942894936, + 0.02296360582113266, + 0.011096678674221039, + 0.025844890624284744, + 0.02685600519180298, + 0.010599019005894661, + 0.00940744485706091, + 0.016650058329105377, + 0.012642455287277699, + 0.017637770622968674, + 0.025083027780056, + 0.023660363629460335, + 0.0194491446018219, + 0.015875497832894325, + 0.01719546690583229, + 0.021313896402716637, + 0.020543746650218964, + 0.03287140652537346, + 0.026250148192048073, + 0.03434968739748001, + 0.02500065602362156, + 0.007595821283757687, + 0.007634206674993038, + 0.009751896373927593, + 0.01168850902467966, + 0.016023928299546242, + 0.013612629845738411, + 0.010696765035390854, + 0.010095874778926373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.01820392720401287 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867294430732727 + ], + "opposite_logprob_tokens": [ + -3.048672914505005 + ], + "logprob_reward": -0.04867294430732727, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867294430732727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05667216330766678, + 0.06380341947078705, + 0.03235326707363129, + 0.10491232573986053, + 0.13222232460975647, + 0.12544789910316467, + 0.08394795656204224, + 0.16923411190509796, + 0.13982637226581573, + 0.09158019721508026 + ], + "credit_map": [ + 0.00575382960960269, + 0.004865744616836309, + 0.010344832204282284, + 0.007264386862516403, + 0.007934032939374447, + 0.0070850239135324955, + 0.00666164793074131, + 0.0067626661621034145, + 0.004464372526854277, + 0.003996615298092365, + 0.01811179332435131, + 0.0063210357911884785, + 0.0051399338990449905, + 0.005500219762325287, + 0.006386606488376856, + 0.013882840052247047, + 0.002131155924871564, + 0.003959231544286013, + 0.004931076895445585, + 0.005317545961588621, + 0.0065336935222148895, + 0.0025870820973068476, + 0.0037435307167470455, + 0.0031499515753239393, + 0.0035845290403813124, + 0.005332494154572487, + 0.010527023114264011, + 0.008650197647511959, + 0.008421259932219982, + 0.005747418384999037, + 0.029611818492412567, + 0.03303758054971695, + 0.0074449023231863976, + 0.023825855925679207, + 0.019036293029785156, + 0.011393125168979168, + 0.01976693980395794, + 0.017222244292497635, + 0.013697615824639797, + 0.019835352897644043, + 0.005888940300792456, + 0.013312602415680885, + 0.028984561562538147, + 0.011014115996658802, + 0.01804966665804386, + 0.011218280531466007, + 0.026676665991544724, + 0.010303067043423653, + 0.004564025439321995, + 0.00475210789591074, + 0.013165684416890144, + 0.008172624744474888, + 0.01732046902179718, + 0.006191735155880451, + 0.017269007861614227, + 0.012512296438217163, + 0.012788382358849049, + 0.015927573665976524, + 0.02261815220117569, + 0.018787799403071404, + 0.022288503125309944, + 0.01927506923675537, + 0.031237035989761353, + 0.026311591267585754, + 0.01136871986091137, + 0.013232635334134102, + 0.015572370029985905, + 0.015061096288263798, + 0.024569561704993248, + 0.018729107454419136, + 0.026191825047135353, + 0.015101048164069653, + 0.012774169445037842, + 0.00805875938385725, + 0.012941080145537853, + 0.010271724313497543, + 0.014484801329672337, + 0.010568656958639622, + 0.011008965782821178, + 0.011472039856016636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.04867294430732727 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306020215153694 + ], + "opposite_logprob_tokens": [ + -3.168060302734375 + ], + "logprob_reward": -0.04306020215153694, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306020215153694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0972013920545578, + 0.11801724880933762, + 0.020901842042803764, + 0.09534363448619843, + 0.05173918604850769, + 0.038489632308483124, + 0.045573700219392776, + 0.20256714522838593, + 0.07221411913633347, + 0.25795212388038635 + ], + "credit_map": [ + 0.006210707128047943, + 0.010533363558351994, + 0.015239272266626358, + 0.01805349439382553, + 0.020263278856873512, + 0.008155032061040401, + 0.008635113947093487, + 0.010111121460795403, + 0.014897598884999752, + 0.009271543473005295, + 0.03142363205552101, + 0.01346670463681221, + 0.01210148073732853, + 0.010398058220744133, + 0.007603608537465334, + 0.018854616209864616, + 0.0013129774015396833, + 0.0022579773794859648, + 0.0033356475178152323, + 0.0026064873673021793, + 0.004771743435412645, + 0.001923360163345933, + 0.0026780625339597464, + 0.0020155864767730236, + 0.007495370227843523, + 0.002866103546693921, + 0.009455319494009018, + 0.003647587727755308, + 0.005767025053501129, + 0.005170443095266819, + 0.027641048654913902, + 0.0333007387816906, + 0.003009067615494132, + 0.008441708981990814, + 0.0068787094205617905, + 0.005854216869920492, + 0.0058979131281375885, + 0.008149593137204647, + 0.005720443557947874, + 0.007787535432726145, + 0.0016511487774550915, + 0.006351060699671507, + 0.005362183786928654, + 0.005169786978513002, + 0.006074201315641403, + 0.0037385118193924427, + 0.005671490915119648, + 0.0044712466187775135, + 0.0030596491415053606, + 0.004986978136003017, + 0.00899291131645441, + 0.005993415601551533, + 0.0064009265042841434, + 0.0038404325023293495, + 0.004852104000747204, + 0.007447286508977413, + 0.012643514201045036, + 0.019361020997166634, + 0.01562463864684105, + 0.021506866440176964, + 0.02755178138613701, + 0.037002235651016235, + 0.030422646552324295, + 0.03845443204045296, + 0.007364709861576557, + 0.005951151717454195, + 0.007681330665946007, + 0.006784401834011078, + 0.012000217102468014, + 0.010061184875667095, + 0.010895281098783016, + 0.01147584430873394, + 0.037375308573246, + 0.021233126521110535, + 0.03827677294611931, + 0.026791447773575783, + 0.04137454554438591, + 0.029047587886452675, + 0.03479628637433052, + 0.02905706875026226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.04306020215153694 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle wide shot, filmed with the crisp, high-fidelity quality of a historical documentary, surveys the scene from the foot of a long, dark gravel driveway. The camera is positioned to capture the unbroken, imposing facade of a Gothic Revival mansion constructed of weathered gray stone beneath a cool, cloudy sky. Flanked by neatly trimmed, dark-green yew hedges, the driveway draws the eye toward the central entrance. The central focal point is a pair of massive, ornate oak doors, their dark wood rich with an ancient patina, adorned with intricate pointed arches and square carvings, fitted with heavy black cast iron hinges and a massive, powerful ring handle. Without a sound, the doors begin to swing inward slowly and deliberately, their majestic movement revealing a small glimpse of the grand, shadowed entrance hall within. As the gap widens, the deep, cool darkness of the interior contrasts with the soft daylight, revealing a glimpse of the massive stone flagstone floor and the base of the winding staircase disappearing into the darkness of the upper floors, creating an atmosphere of solemn reverence and profound historical anticipation.", + "seed": 31698, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_8/index_9546_seed_31698_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11212468519806862, + "vlm_logprob": -0.11212468519806862, + "vlm_yes_no_margin": 3.175000011920929, + "vlm_token_credit": -0.11212468519806862, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.434015691280365 + ], + "opposite_logprob_tokens": [ + -1.1840157508850098 + ], + "logprob_reward": -0.434015691280365, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.434015691280365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14411617815494537, + 0.049054984003305435, + 0.022600650787353516, + 0.058138247579336166, + 0.06858838349580765, + 0.04354969412088394, + 0.13772796094417572, + 0.3677895665168762, + 0.07545348256826401, + 0.03298092260956764 + ], + "credit_map": [ + 0.011463215574622154, + 0.015141219832003117, + 0.018061073496937752, + 0.019283801317214966, + 0.015840735286474228, + 0.013064300641417503, + 0.026221072301268578, + 0.025040753185749054, + 0.003812112146988511, + 0.004511128645390272, + 0.008698937483131886, + 0.006917391438037157, + 0.004866847302764654, + 0.00424010306596756, + 0.00767989456653595, + 0.008328567259013653, + 0.0026046745479106903, + 0.0023458569776266813, + 0.004436674062162638, + 0.0024578773882240057, + 0.002034294418990612, + 0.002036825055256486, + 0.002710691187530756, + 0.003973756916821003, + 0.003304677549749613, + 0.004042390733957291, + 0.018348075449466705, + 0.0036076928954571486, + 0.003520610509440303, + 0.003686001058667898, + 0.004372386261820793, + 0.017256412655115128, + 0.00346552231349051, + 0.002929827431216836, + 0.023515956476330757, + 0.006737253163009882, + 0.00277065415866673, + 0.0033513919916003942, + 0.004487592726945877, + 0.021330172196030617, + 0.0038893595337867737, + 0.006558706983923912, + 0.006810725200921297, + 0.005560382269322872, + 0.0050146859139204025, + 0.004632072523236275, + 0.004937934223562479, + 0.006145827937871218, + 0.007895310409367085, + 0.01212658453732729, + 0.04071813449263573, + 0.015796927735209465, + 0.008485372178256512, + 0.007815876975655556, + 0.011373241432011127, + 0.033516500145196915, + 0.01902969740331173, + 0.02140534296631813, + 0.1315460056066513, + 0.020774785429239273, + 0.02031281217932701, + 0.013667112216353416, + 0.019969947636127472, + 0.1210838034749031, + 0.006825366523116827, + 0.010663356631994247, + 0.01885903626680374, + 0.010532640852034092, + 0.0052059488371014595, + 0.003296059323474765, + 0.00591710489243269, + 0.014153963886201382, + 0.0038300768937915564, + 0.0040366146713495255, + 0.005814129952341318, + 0.004288768861442804, + 0.002898598788306117, + 0.001872799126431346, + 0.004728700499981642, + 0.0055112335830926895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.434015691280365 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006727904081345 + ], + "opposite_logprob_tokens": [ + -2.6950671672821045 + ], + "logprob_reward": -0.07006727904081345, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006727904081345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10533171147108078, + 0.061130013316869736, + 0.023537199944257736, + 0.12252108752727509, + 0.2898068130016327, + 0.0418497659265995, + 0.07851888984441757, + 0.22752273082733154, + 0.030400048941373825, + 0.01938171125948429 + ], + "credit_map": [ + 0.010878348723053932, + 0.017747944220900536, + 0.011536567471921444, + 0.013958062045276165, + 0.014197420328855515, + 0.011018939316272736, + 0.013524108566343784, + 0.012470314279198647, + 0.0036924818996340036, + 0.0096040740609169, + 0.007385315839201212, + 0.013269622810184956, + 0.006902105174958706, + 0.004791704472154379, + 0.006715229246765375, + 0.00876947958022356, + 0.0014652310637757182, + 0.0020502880215644836, + 0.004654688760638237, + 0.00510368263348937, + 0.0014144876040518284, + 0.0011644638143479824, + 0.0038858314510434866, + 0.0037985267117619514, + 0.010836847126483917, + 0.010093802586197853, + 0.029878951609134674, + 0.010624675080180168, + 0.013230285607278347, + 0.007532194256782532, + 0.008480045944452286, + 0.031844284385442734, + 0.016753973439335823, + 0.004961302038282156, + 0.11881110817193985, + 0.01482586469501257, + 0.006911832373589277, + 0.01126333698630333, + 0.008564219810068607, + 0.1077151969075203, + 0.003073030151426792, + 0.0024810887407511473, + 0.013779413886368275, + 0.002969806082546711, + 0.002781116869300604, + 0.0024364921264350414, + 0.0029092393815517426, + 0.01141957938671112, + 0.005166116636246443, + 0.008682695217430592, + 0.02081868425011635, + 0.00984913855791092, + 0.005748035851866007, + 0.005670894403010607, + 0.007489632349461317, + 0.015093701891601086, + 0.011732688173651695, + 0.012863392941653728, + 0.07700017094612122, + 0.015521662309765816, + 0.011989406310021877, + 0.00915268249809742, + 0.016514988616108894, + 0.07274773716926575, + 0.0022002544719725847, + 0.0022507919929921627, + 0.006393795367330313, + 0.0035407105460762978, + 0.0027177073061466217, + 0.0022838276345282793, + 0.005206014961004257, + 0.005806945264339447, + 0.0022880821488797665, + 0.003686281153932214, + 0.0031706488225609064, + 0.002492083702236414, + 0.0018772822804749012, + 0.0012264529941603541, + 0.0022947178222239017, + 0.002346161985769868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.07006727904081345 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067548975348472595 + ], + "opposite_logprob_tokens": [ + -5.0067548751831055 + ], + "logprob_reward": -0.0067548975348472595, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067548975348472595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0891939103603363, + 0.05021345987915993, + 0.058969613164663315, + 0.04654482379555702, + 0.045080579817295074, + 0.1170596107840538, + 0.04583783447742462, + 0.4404713809490204, + 0.05592505633831024, + 0.050703756511211395 + ], + "credit_map": [ + 0.010148468427360058, + 0.011144871823489666, + 0.012923944741487503, + 0.009163380600512028, + 0.010502870194613934, + 0.007064505945891142, + 0.008826123550534248, + 0.01941973716020584, + 0.004337100312113762, + 0.005436771083623171, + 0.007784358225762844, + 0.00652488274499774, + 0.0046902927570044994, + 0.0035710535012185574, + 0.010607779957354069, + 0.007261217571794987, + 0.009129928424954414, + 0.006693474017083645, + 0.012804253026843071, + 0.004892780911177397, + 0.005559537094086409, + 0.006058103404939175, + 0.004879504442214966, + 0.008952027186751366, + 0.0027080702129751444, + 0.008023465052247047, + 0.00687082763761282, + 0.006170815322548151, + 0.008006509393453598, + 0.0035173825453966856, + 0.005000069737434387, + 0.0062476834282279015, + 0.0033617476001381874, + 0.007113829720765352, + 0.006612520664930344, + 0.006574677769094706, + 0.005325341131538153, + 0.003184442175552249, + 0.007513043936342001, + 0.005394975654780865, + 0.008668635040521622, + 0.010426145046949387, + 0.03318231180310249, + 0.014140507206320763, + 0.009253446012735367, + 0.006494311150163412, + 0.010835890658199787, + 0.024058347567915916, + 0.0034557697363197803, + 0.00884613674134016, + 0.007305820472538471, + 0.005551752634346485, + 0.004283008631318808, + 0.004216203466057777, + 0.0059163523837924, + 0.00626278854906559, + 0.02052350342273712, + 0.022836651653051376, + 0.16322189569473267, + 0.023278912529349327, + 0.01824754849076271, + 0.014625178650021553, + 0.022362442687153816, + 0.15537524223327637, + 0.004096705932170153, + 0.008911599405109882, + 0.010483563877642155, + 0.010063017718493938, + 0.00551136489957571, + 0.004318948835134506, + 0.005568050313740969, + 0.006971798837184906, + 0.006274843588471413, + 0.004999171942472458, + 0.009374338202178478, + 0.00708759855479002, + 0.0035890720319002867, + 0.0048389011062681675, + 0.005055073648691177, + 0.009484751150012016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.0067548975348472595 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371308371424675 + ], + "opposite_logprob_tokens": [ + -3.4087131023406982 + ], + "logprob_reward": -0.03371308371424675, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371308371424675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028203926980495453, + 0.02699333429336548, + 0.034879859536886215, + 0.024175778031349182, + 0.3755472004413605, + 0.04295222461223602, + 0.06446637958288193, + 0.2052147537469864, + 0.08750057220458984, + 0.1100659966468811 + ], + "credit_map": [ + 0.0044700754806399345, + 0.002832431811839342, + 0.004510068334639072, + 0.005001319106668234, + 0.002595247235149145, + 0.002464846009388566, + 0.0032187269534915686, + 0.003111214144155383, + 0.0031940401531755924, + 0.0022733525838702917, + 0.004252096172422171, + 0.004265898372977972, + 0.002260820008814335, + 0.003086070530116558, + 0.0031188002321869135, + 0.004542255774140358, + 0.0054598115384578705, + 0.0037379679270088673, + 0.006669341586530209, + 0.004386360757052898, + 0.002534633968025446, + 0.0031212852336466312, + 0.003641977207735181, + 0.005328482948243618, + 0.0033912360668182373, + 0.0022555901668965816, + 0.004373912699520588, + 0.003015076043084264, + 0.0019716238602995872, + 0.0019708024337887764, + 0.0028640187811106443, + 0.0043335179798305035, + 0.021541889756917953, + 0.006735941860824823, + 0.1542259305715561, + 0.01753941923379898, + 0.009340356104075909, + 0.014588854275643826, + 0.011075086891651154, + 0.14049969613552094, + 0.0026521889958530664, + 0.003421430243179202, + 0.01243529561907053, + 0.006589229218661785, + 0.0028205218259245157, + 0.0019066105596721172, + 0.00311104953289032, + 0.010015899315476418, + 0.0032670204527676105, + 0.005178388208150864, + 0.021095551550388336, + 0.005366483237594366, + 0.002830170327797532, + 0.0030895243398845196, + 0.003123560221865773, + 0.02051568403840065, + 0.009488226845860481, + 0.009919429197907448, + 0.07573746144771576, + 0.011229461058974266, + 0.00904261413961649, + 0.006444784812629223, + 0.011256970465183258, + 0.07209579646587372, + 0.006573521997779608, + 0.00979574490338564, + 0.01956014521420002, + 0.010949578136205673, + 0.010772586800158024, + 0.006004997529089451, + 0.010207779705524445, + 0.013636217452585697, + 0.009236677549779415, + 0.015229940414428711, + 0.017002398148179054, + 0.019853726029396057, + 0.011476153507828712, + 0.006385516840964556, + 0.015055191703140736, + 0.015826400369405746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.03371308371424675 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072474420070648 + ], + "opposite_logprob_tokens": [ + -4.1410722732543945 + ], + "logprob_reward": -0.016072474420070648, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072474420070648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13751855492591858, + 0.08209692686796188, + 0.03137916699051857, + 0.07596352696418762, + 0.1291239708662033, + 0.06401846557855606, + 0.1173415407538414, + 0.24689428508281708, + 0.04964883252978325, + 0.06601472198963165 + ], + "credit_map": [ + 0.012276739813387394, + 0.008670590817928314, + 0.02731388807296753, + 0.01596459373831749, + 0.010631709359586239, + 0.010567009449005127, + 0.01974705047905445, + 0.032346971333026886, + 0.014993211254477501, + 0.006405906286090612, + 0.01404274720698595, + 0.008126534521579742, + 0.006641904357820749, + 0.011800154112279415, + 0.008579381741583347, + 0.011507092975080013, + 0.003172785509377718, + 0.0026708359364420176, + 0.0066628544591367245, + 0.0037841114681214094, + 0.0024437797255814075, + 0.0026632798835635185, + 0.0036497891414910555, + 0.006331729236990213, + 0.003526762593537569, + 0.005269760265946388, + 0.016558844596147537, + 0.013568555936217308, + 0.005879137199372053, + 0.005451401695609093, + 0.01066471729427576, + 0.015044343657791615, + 0.0071635134518146515, + 0.0038746921345591545, + 0.04881320148706436, + 0.006634259596467018, + 0.00545680383220315, + 0.006429231259971857, + 0.005270101595669985, + 0.04548216983675957, + 0.005192401818931103, + 0.005792729090899229, + 0.013163079507648945, + 0.007091819774359465, + 0.005953105632215738, + 0.005373390391469002, + 0.009544432163238525, + 0.011907502077519894, + 0.005928147118538618, + 0.007128109224140644, + 0.041962068527936935, + 0.0060782781802117825, + 0.004077968653291464, + 0.006456884555518627, + 0.006220090202987194, + 0.039489999413490295, + 0.011850607581436634, + 0.012150856666266918, + 0.09288802742958069, + 0.013634675182402134, + 0.009611628018319607, + 0.007949208840727806, + 0.01224985346198082, + 0.08655944466590881, + 0.003573224414139986, + 0.003874234389513731, + 0.007352686021476984, + 0.0088400449603796, + 0.005831559654325247, + 0.0036586690694093704, + 0.009895039722323418, + 0.006623375229537487, + 0.005912196822464466, + 0.006177494302392006, + 0.010091525502502918, + 0.010409899987280369, + 0.0054751005955040455, + 0.004743251483887434, + 0.009070700965821743, + 0.014134550467133522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.016072474420070648 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle wide shot, filmed with the crisp, high-fidelity quality of a historical documentary, surveys the scene from the foot of a long, dark gravel driveway. The camera is positioned to capture the unbroken, imposing facade of a Gothic Revival mansion constructed of weathered gray stone beneath a cool, cloudy sky. Flanked by neatly trimmed, dark-green yew hedges, the driveway draws the eye toward the central entrance. The central focal point is a pair of massive, ornate oak doors, their dark wood rich with an ancient patina, adorned with intricate pointed arches and square carvings, fitted with heavy black cast iron hinges and a massive, powerful ring handle. Without a sound, the doors begin to swing inward slowly and deliberately, their majestic movement revealing a small glimpse of the grand, shadowed entrance hall within. As the gap widens, the deep, cool darkness of the interior contrasts with the soft daylight, revealing a glimpse of the massive stone flagstone floor and the base of the winding staircase disappearing into the darkness of the upper floors, creating an atmosphere of solemn reverence and profound historical anticipation.", + "seed": 31699, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_8/index_9546_seed_31699_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12090456429868937, + "vlm_logprob": -0.12090456429868937, + "vlm_yes_no_margin": 2.949999988079071, + "vlm_token_credit": -0.12090456429868937, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4455798268318176 + ], + "opposite_logprob_tokens": [ + -1.1955797672271729 + ], + "logprob_reward": -0.4455798268318176, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.4455798268318176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06280646473169327, + 0.034378837794065475, + 0.0188518725335598, + 0.025659294798970222, + 0.11064333468675613, + 0.4738157093524933, + 0.05548587441444397, + 0.045504674315452576, + 0.042583972215652466, + 0.13026998937129974 + ], + "credit_map": [ + 0.009470322169363499, + 0.007228961680084467, + 0.007946686819195747, + 0.006970290560275316, + 0.00641945144161582, + 0.007536096032708883, + 0.007705136202275753, + 0.009529510512948036, + 0.0037209067959338427, + 0.002163264900445938, + 0.007949830032885075, + 0.0022742256987839937, + 0.002258261200040579, + 0.0064797354862093925, + 0.002776536624878645, + 0.00675607193261385, + 0.00209276145324111, + 0.0015764590352773666, + 0.0036197914741933346, + 0.00154957827180624, + 0.0017937846714630723, + 0.003214219817891717, + 0.002259735018014908, + 0.002745541278272867, + 0.0020820985082536936, + 0.001982598565518856, + 0.004427231848239899, + 0.003393179504200816, + 0.0018441511783748865, + 0.005285047926008701, + 0.0035581612028181553, + 0.0030868237372487783, + 0.0079202214255929, + 0.006059596780687571, + 0.03904302045702934, + 0.008580676279962063, + 0.004905541427433491, + 0.006598203908652067, + 0.006163202226161957, + 0.03137287124991417, + 0.021624481305480003, + 0.02194678597152233, + 0.18208074569702148, + 0.030952634289860725, + 0.014643869362771511, + 0.0181973148137331, + 0.01340340543538332, + 0.17096640169620514, + 0.0044690524227917194, + 0.007306764833629131, + 0.01030246913433075, + 0.007296639960259199, + 0.008643761277198792, + 0.003885313170030713, + 0.004762019030749798, + 0.008819852024316788, + 0.002957904478535056, + 0.0046183220110833645, + 0.006262572482228279, + 0.005374168511480093, + 0.004340282175689936, + 0.003767822403460741, + 0.006310720928013325, + 0.01187287736684084, + 0.003214616095647216, + 0.007290680427104235, + 0.005338655319064856, + 0.007146961987018585, + 0.006609764415770769, + 0.003271910361945629, + 0.004319081548601389, + 0.005392298102378845, + 0.005209316033869982, + 0.006764732766896486, + 0.0509844645857811, + 0.03853324055671692, + 0.0044730897061526775, + 0.005185539834201336, + 0.011329500004649162, + 0.007790085393935442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.4455798268318176 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908046782016754 + ], + "opposite_logprob_tokens": [ + -2.464080572128296 + ], + "logprob_reward": -0.08908046782016754, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908046782016754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09524673223495483, + 0.1100800633430481, + 0.015524212270975113, + 0.02644200623035431, + 0.02542179636657238, + 0.0632464662194252, + 0.023009587079286575, + 0.023740176111459732, + 0.06126033887267113, + 0.5560285449028015 + ], + "credit_map": [ + 0.014155837707221508, + 0.007225817535072565, + 0.014460640028119087, + 0.008927577175199986, + 0.009465791285037994, + 0.011300131678581238, + 0.011490113101899624, + 0.01822083070874214, + 0.010355459526181221, + 0.006985392887145281, + 0.01492330152541399, + 0.006640248466283083, + 0.009191013872623444, + 0.03202303126454353, + 0.009582229889929295, + 0.02037939243018627, + 0.0017273050034418702, + 0.0016380317974835634, + 0.001799199846573174, + 0.0018029386410489678, + 0.001769330003298819, + 0.0025519561022520065, + 0.0028781862929463387, + 0.0013572652824223042, + 0.0019867089577019215, + 0.0016832253895699978, + 0.006448109168559313, + 0.0025741078425198793, + 0.0017325931694358587, + 0.002776876324787736, + 0.002558008534833789, + 0.006682378239929676, + 0.0023861669469624758, + 0.0027534996625036, + 0.005125077441334724, + 0.003793965792283416, + 0.00291448668576777, + 0.0015138668240979314, + 0.002755177905783057, + 0.004179558716714382, + 0.002778132911771536, + 0.003556839656084776, + 0.0225431676954031, + 0.005250907503068447, + 0.0022718387190252542, + 0.0026655932888388634, + 0.0029197705443948507, + 0.021260220557451248, + 0.0015969849191606045, + 0.0032413301523774862, + 0.0032383191864937544, + 0.0034957739990204573, + 0.003172952216118574, + 0.0025619911029934883, + 0.0026637862902134657, + 0.0030384506098926067, + 0.0014453604817390442, + 0.0020335137378424406, + 0.004701115190982819, + 0.003820751328021288, + 0.001997266663238406, + 0.0017401011427864432, + 0.004229488782584667, + 0.0037725823931396008, + 0.00480834161862731, + 0.0055321757681667805, + 0.018452363088726997, + 0.010719986632466316, + 0.006590609904378653, + 0.0037910565733909607, + 0.005891880020499229, + 0.005473929923027754, + 0.021562356501817703, + 0.02192409709095955, + 0.2314116656780243, + 0.18393531441688538, + 0.011724530719220638, + 0.01394641026854515, + 0.04605688899755478, + 0.025467338040471077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.08908046782016754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009799941442906857 + ], + "opposite_logprob_tokens": [ + -4.634799957275391 + ], + "logprob_reward": -0.009799941442906857, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009799941442906857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05064355209469795, + 0.020391929894685745, + 0.008292263373732567, + 0.02039450779557228, + 0.09137741476297379, + 0.19744160771369934, + 0.027344031259417534, + 0.05280689150094986, + 0.036526069045066833, + 0.4947817325592041 + ], + "credit_map": [ + 0.008384555578231812, + 0.002841192064806819, + 0.009816420264542103, + 0.003592222463339567, + 0.003823985578492284, + 0.005830509588122368, + 0.003272016067057848, + 0.013082653284072876, + 0.0013817589497193694, + 0.002080676145851612, + 0.003558655036613345, + 0.0016983301611617208, + 0.001657489687204361, + 0.0026160506531596184, + 0.0033758163917809725, + 0.004023151472210884, + 0.0008632413228042424, + 0.0007469429401680827, + 0.0013054385781288147, + 0.0010652625933289528, + 0.0009120404720306396, + 0.0011161467991769314, + 0.0011646716156974435, + 0.0011185191106051207, + 0.001777359750121832, + 0.001544794999063015, + 0.004787728190422058, + 0.002107626525685191, + 0.0015381085686385632, + 0.0015846129972487688, + 0.0023089463356882334, + 0.004745330195873976, + 0.006475264206528664, + 0.005818151403218508, + 0.03236227110028267, + 0.006678622681647539, + 0.0045751845464110374, + 0.005226914305239916, + 0.0057454826310276985, + 0.024495529010891914, + 0.008126595057547092, + 0.011713864281773567, + 0.07004452496767044, + 0.013298653066158295, + 0.00864377897232771, + 0.009330631233751774, + 0.006735068745911121, + 0.06954850256443024, + 0.001801078673452139, + 0.004374869633466005, + 0.003323518903926015, + 0.004365956410765648, + 0.0046264478005468845, + 0.00255968258716166, + 0.003524489002302289, + 0.002767988946288824, + 0.0030871101189404726, + 0.005636819638311863, + 0.008052492514252663, + 0.009631559252738953, + 0.005471022333949804, + 0.003112349659204483, + 0.011262417770922184, + 0.006553123239427805, + 0.0027373924385756254, + 0.0035671016667038202, + 0.011111977510154247, + 0.006769052240997553, + 0.003648659447208047, + 0.0018357844091951847, + 0.003575862618163228, + 0.0032802403438836336, + 0.018696509301662445, + 0.019786955788731575, + 0.20617827773094177, + 0.1626916229724884, + 0.009918097406625748, + 0.013201464898884296, + 0.041591230779886246, + 0.02271762117743492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.009799941442906857 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635710872709751 + ], + "opposite_logprob_tokens": [ + -3.6513571739196777 + ], + "logprob_reward": -0.02635710872709751, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635710872709751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03083564154803753, + 0.030665114521980286, + 0.016557147726416588, + 0.043611977249383926, + 0.058615416288375854, + 0.4301435649394989, + 0.029855934903025627, + 0.03920203447341919, + 0.1123080849647522, + 0.20820507407188416 + ], + "credit_map": [ + 0.003189746756106615, + 0.005430364515632391, + 0.0031054657883942127, + 0.003887010272592306, + 0.0035208200570195913, + 0.004151847213506699, + 0.0035991952754557133, + 0.00395119097083807, + 0.0035756705328822136, + 0.0019035530276596546, + 0.003365748096257448, + 0.002164216013625264, + 0.0027645835652947426, + 0.010604101233184338, + 0.0017247380455955863, + 0.00456250412389636, + 0.0017290712567046285, + 0.002655541989952326, + 0.00229858560487628, + 0.001960552064701915, + 0.0016376387793570757, + 0.0016815993003547192, + 0.002314108656719327, + 0.0022800499573349953, + 0.0033124187029898167, + 0.0023754104040563107, + 0.010043411515653133, + 0.0038766413927078247, + 0.0028970593120902777, + 0.006774391978979111, + 0.004793869331479073, + 0.009538776241242886, + 0.0041014282032847404, + 0.003745450871065259, + 0.019545841962099075, + 0.0038597804959863424, + 0.0026928121224045753, + 0.004152361303567886, + 0.004090787842869759, + 0.01642695814371109, + 0.01839035376906395, + 0.020408539101481438, + 0.1606103479862213, + 0.026892470195889473, + 0.016638385131955147, + 0.017856694757938385, + 0.01582140289247036, + 0.15352539718151093, + 0.0022757090628147125, + 0.0043904902413487434, + 0.004775663372129202, + 0.004268564283847809, + 0.0033756475895643234, + 0.0023901239037513733, + 0.0028331337962299585, + 0.005546601489186287, + 0.0022224392741918564, + 0.0035671244841068983, + 0.006532568950206041, + 0.006226986646652222, + 0.0039114016108214855, + 0.002712195971980691, + 0.007227448280900717, + 0.0068018692545592785, + 0.008821449242532253, + 0.016157792881131172, + 0.02060389518737793, + 0.017113542184233665, + 0.015761535614728928, + 0.009508522227406502, + 0.011346324346959591, + 0.01299501582980156, + 0.012256193906068802, + 0.023778017610311508, + 0.04391556605696678, + 0.04221419617533684, + 0.01938321441411972, + 0.015547028742730618, + 0.031686924397945404, + 0.019423946738243103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.02635710872709751 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Gothic Revival mansion made of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive ornate oak doors with dark wood and ancient patina visible as the central focal point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the doors swing inward slowly and deliberately to reveal the interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene framed by a long dark gravel driveway flanked by neatly trimmed dark-green yew hedges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370547667145729 + ], + "opposite_logprob_tokens": [ + -3.408705472946167 + ], + "logprob_reward": -0.03370547667145729, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370547667145729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08785681426525116, + 0.06873487681150436, + 0.03056236542761326, + 0.03845862299203873, + 0.5155662894248962, + 0.04417800530791283, + 0.035055480897426605, + 0.08229910582304001, + 0.025222664698958397, + 0.0720658078789711 + ], + "credit_map": [ + 0.014206956140697002, + 0.005276393610984087, + 0.01884608156979084, + 0.006795995868742466, + 0.006158832460641861, + 0.0065071796998381615, + 0.005396521650254726, + 0.024668855592608452, + 0.00823277048766613, + 0.003587613580748439, + 0.008413262665271759, + 0.004944557324051857, + 0.005830190610140562, + 0.022650575265288353, + 0.00503596430644393, + 0.010039947926998138, + 0.004073321353644133, + 0.0021565244533121586, + 0.005176614969968796, + 0.0042235893197357655, + 0.0021980805322527885, + 0.004942032974213362, + 0.003637699643149972, + 0.004154503345489502, + 0.0031222484540194273, + 0.0020931779872626066, + 0.00945038627833128, + 0.0036408996675163507, + 0.002390434266999364, + 0.006465632934123278, + 0.003215261735022068, + 0.008080584928393364, + 0.03610176220536232, + 0.027402635663747787, + 0.18462292850017548, + 0.03377470001578331, + 0.021276867017149925, + 0.033214110881090164, + 0.02772999368607998, + 0.1514432579278946, + 0.0019688778556883335, + 0.0032185360323637724, + 0.014651133678853512, + 0.0034531820565462112, + 0.0025780550204217434, + 0.002039337297901511, + 0.0030757093336433172, + 0.013193173334002495, + 0.002350099617615342, + 0.005817038007080555, + 0.004490172956138849, + 0.004963644314557314, + 0.004391427151858807, + 0.005273999180644751, + 0.0037682554684579372, + 0.004000842571258545, + 0.004351441282778978, + 0.009077804163098335, + 0.012150388211011887, + 0.01577606424689293, + 0.007356536574661732, + 0.004752257838845253, + 0.01704740896821022, + 0.011787200346589088, + 0.0017410428263247013, + 0.0024523779284209013, + 0.007906200364232063, + 0.005044776480644941, + 0.0020076206419616938, + 0.001271474058739841, + 0.0027180102188140154, + 0.0020811620634049177, + 0.0034618834033608437, + 0.0044599915854632854, + 0.026318974792957306, + 0.02239910513162613, + 0.002383402083069086, + 0.0027873662766069174, + 0.00617948267608881, + 0.004075597506016493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.03370547667145729 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_9/index_tensor([6922, 6922])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_9/index_tensor([6922, 6922])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..39295a9fe4c2e7a811fae2a1c9225f43daaad5a5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_9/index_tensor([6922, 6922])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "strikingly dynamic, low-angle, side-on tracking shot follows a modern police sedan at high speed. The car, sporting a classic black and white paint scheme, its rooftop lights flashing red and blue at a rapid tempo. It travels along a straight, empty asphalt road, framed by a pure, uniform light-gray sky that stretches seamlessly to the horizon, devoid of any clouds or landmarks. The car's body tilts slightly at breakneck speed, its glossy paint reflecting a soft, shadowless diffuse light, creating a cool, metallic quality. The tires' rubbing against the road kick up a faint cloud of dust, further accentuating its rapid motion. The composition is remarkably minimalist, primarily black, white, and gray. The flashing red and blue police lights serve as the sole splash of color within this minimalist composition, shattering the monochromatic silence and creating a powerful sense of speed, urgency, and a surreal atmosphere of isolation.", + "seed": 28113, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_9/index_6922_seed_28113_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15481210984289645, + "vlm_logprob": -0.15481210984289645, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.15481210984289645, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030974164605140686 + ], + "opposite_logprob_tokens": [ + -3.6559741497039795 + ], + "logprob_reward": -0.030974164605140686, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.030974164605140686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17540043592453003, + 0.12173811346292496, + 0.026712823659181595, + 0.0665922462940216, + 0.11873967200517654, + 0.047656912356615067, + 0.11962991952896118, + 0.09049355983734131, + 0.041562993079423904, + 0.19147327542304993 + ], + "credit_map": [ + 0.014809689484536648, + 0.01727120950818062, + 0.02611633762717247, + 0.018756341189146042, + 0.020224371924996376, + 0.02181786671280861, + 0.03407811373472214, + 0.022326521575450897, + 0.005297538358718157, + 0.01265441719442606, + 0.008096625097095966, + 0.00834465678781271, + 0.04136142134666443, + 0.031589772552251816, + 0.007239104714244604, + 0.0071545750834047794, + 0.0017615124816074967, + 0.0025929452385753393, + 0.0022782699670642614, + 0.002011540113016963, + 0.005922322161495686, + 0.005353638902306557, + 0.003774772398173809, + 0.0030178234446793795, + 0.004473351873457432, + 0.009024154394865036, + 0.0063971481285989285, + 0.016425957903265953, + 0.009428326040506363, + 0.006189308129251003, + 0.008761944249272346, + 0.0058920541778206825, + 0.011029582470655441, + 0.013418049551546574, + 0.019223880022764206, + 0.014415619894862175, + 0.014694440178573132, + 0.014504635706543922, + 0.013424351811408997, + 0.0180291049182415, + 0.007892238907516003, + 0.004977087024599314, + 0.006287507247179747, + 0.004601296503096819, + 0.006317898165434599, + 0.006546326447278261, + 0.004563754424452782, + 0.006470805034041405, + 0.015107340179383755, + 0.015964288264513016, + 0.012988762930035591, + 0.011371868662536144, + 0.0295402854681015, + 0.009868352673947811, + 0.013353298418223858, + 0.011435720138251781, + 0.009015885181725025, + 0.01046337652951479, + 0.011296792887151241, + 0.013388699851930141, + 0.011854177340865135, + 0.011098292656242847, + 0.013272261247038841, + 0.01010406669229269, + 0.004913808312267065, + 0.006318636238574982, + 0.0053441789932549, + 0.004787768237292767, + 0.007902326993644238, + 0.005194289144128561, + 0.003116488689556718, + 0.003985497634857893, + 0.005659220740199089, + 0.08177781105041504, + 0.006326963659375906, + 0.002801249735057354, + 0.08259416371583939, + 0.005443675443530083, + 0.0034527690149843693, + 0.003417420666664839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.030974164605140686 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762509703636169 + ], + "opposite_logprob_tokens": [ + -0.8262509703636169 + ], + "logprob_reward": -0.5762509703636169, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762509703636169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13125817477703094, + 0.18327870965003967, + 0.10194860398769379, + 0.1685926914215088, + 0.027317607775330544, + 0.024809228256344795, + 0.0360882543027401, + 0.07404255121946335, + 0.044514115899801254, + 0.20815007388591766 + ], + "credit_map": [ + 0.008860476315021515, + 0.024313010275363922, + 0.01119465846568346, + 0.009609864093363285, + 0.02460549585521221, + 0.01633233204483986, + 0.023619351908564568, + 0.012722985818982124, + 0.007080679759383202, + 0.020340420305728912, + 0.008292382583022118, + 0.009097890928387642, + 0.07037974894046783, + 0.05181768536567688, + 0.00853784754872322, + 0.007732052356004715, + 0.00872756913304329, + 0.010954346507787704, + 0.00865265168249607, + 0.007679277565330267, + 0.016375532373785973, + 0.016325373202562332, + 0.013246618211269379, + 0.019987227395176888, + 0.011966054327785969, + 0.022733936086297035, + 0.018286826089024544, + 0.04421002045273781, + 0.021681571379303932, + 0.011833738535642624, + 0.021934043616056442, + 0.015946492552757263, + 0.0027114071417599916, + 0.002757738344371319, + 0.0066724736243486404, + 0.0031079130712896585, + 0.0027976329438388348, + 0.0027455082163214684, + 0.0024645801167935133, + 0.004060355015099049, + 0.004162027034908533, + 0.0023835566826164722, + 0.002982201986014843, + 0.0019260472618043423, + 0.0036568406503647566, + 0.0029733057599514723, + 0.0032860056962817907, + 0.003439241088926792, + 0.004883369896560907, + 0.005006332416087389, + 0.0040053208358585835, + 0.003158462466672063, + 0.006646069698035717, + 0.00467897392809391, + 0.003567772451788187, + 0.004141956567764282, + 0.009336956776678562, + 0.009485531598329544, + 0.00787913054227829, + 0.009601191617548466, + 0.009547250345349312, + 0.009625651873648167, + 0.010729403235018253, + 0.007837440818548203, + 0.006025275215506554, + 0.0057106781750917435, + 0.01075795665383339, + 0.005506808403879404, + 0.006727300118654966, + 0.003250132082030177, + 0.0028797376435250044, + 0.0036562259774655104, + 0.004707893822342157, + 0.09389500319957733, + 0.004348062444478273, + 0.0025604390539228916, + 0.09294872730970383, + 0.005038538482040167, + 0.0024272550363093615, + 0.0022241517435759306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.5762509703636169 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009793448261916637 + ], + "opposite_logprob_tokens": [ + -4.634793281555176 + ], + "logprob_reward": -0.009793448261916637, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009793448261916637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14804956316947937, + 0.30391770601272583, + 0.20778755843639374, + 0.0810231864452362, + 0.04569530859589577, + 0.04061228781938553, + 0.02735612727701664, + 0.023073595017194748, + 0.022000616416335106, + 0.10048401355743408 + ], + "credit_map": [ + 0.0079409284517169, + 0.019027573987841606, + 0.013547156937420368, + 0.010803187265992165, + 0.02527209185063839, + 0.018542295321822166, + 0.0274654608219862, + 0.025450890883803368, + 0.010580851696431637, + 0.03942614793777466, + 0.009980015456676483, + 0.01019987277686596, + 0.1181531623005867, + 0.08176183700561523, + 0.018502049148082733, + 0.015313784591853619, + 0.01953817531466484, + 0.01989104226231575, + 0.013009578920900822, + 0.013804462738335133, + 0.029172739014029503, + 0.036877769976854324, + 0.028884943574666977, + 0.046608857810497284, + 0.00737955654039979, + 0.009079327806830406, + 0.0038980627432465553, + 0.0052319779060781, + 0.014332257211208344, + 0.01673128828406334, + 0.007854608818888664, + 0.016516108065843582, + 0.0064804633148014545, + 0.005049095489084721, + 0.008073375560343266, + 0.003713845042511821, + 0.0040522669441998005, + 0.005640375893563032, + 0.004032403230667114, + 0.008653489872813225, + 0.008365720510482788, + 0.0043844422325491905, + 0.00552612217143178, + 0.003788206959143281, + 0.005147287622094154, + 0.0055079227313399315, + 0.004125768784433603, + 0.003766818903386593, + 0.004386244807392359, + 0.004466111771762371, + 0.0028926562517881393, + 0.0023490653838962317, + 0.005000075791031122, + 0.002710845088586211, + 0.0028226852882653475, + 0.0027284445241093636, + 0.0045854030176997185, + 0.0032261821907013655, + 0.002689096610993147, + 0.002843766938894987, + 0.0018083284376189113, + 0.002029676688835025, + 0.003456512466073036, + 0.0024346315767616034, + 0.0023341902997344732, + 0.003075624816119671, + 0.004902863875031471, + 0.0024407540913671255, + 0.0030882966239005327, + 0.002189626917243004, + 0.002126343548297882, + 0.0018429192714393139, + 0.0026918670628219843, + 0.043479952961206436, + 0.0038523622788488865, + 0.0015676126349717379, + 0.04263265058398247, + 0.002837744075804949, + 0.0016994838370010257, + 0.0017223487375304103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.009793448261916637 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428481936454773 + ], + "opposite_logprob_tokens": [ + -2.017848253250122 + ], + "logprob_reward": -0.1428481936454773, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428481936454773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27109676599502563, + 0.08621799200773239, + 0.06018957495689392, + 0.055463433265686035, + 0.0754241794347763, + 0.019919265061616898, + 0.018656160682439804, + 0.021022383123636246, + 0.017619263380765915, + 0.37439098954200745 + ], + "credit_map": [ + 0.021787887439131737, + 0.04115302115678787, + 0.03297147899866104, + 0.023171069100499153, + 0.02703368477523327, + 0.04900292679667473, + 0.041076961904764175, + 0.03489973023533821, + 0.005958739668130875, + 0.009782691486179829, + 0.01043714303523302, + 0.00847113411873579, + 0.019855601713061333, + 0.015389544889330864, + 0.009237703867256641, + 0.007085438817739487, + 0.002469732193276286, + 0.004514715168625116, + 0.0030924328602850437, + 0.002748006023466587, + 0.021409709006547928, + 0.018605448305606842, + 0.0034472534898668528, + 0.0039022788405418396, + 0.004933105781674385, + 0.009907870553433895, + 0.004581010434776545, + 0.007157263346016407, + 0.006630445830523968, + 0.00812835618853569, + 0.005746065638959408, + 0.008379315957427025, + 0.009441160596907139, + 0.007620177697390318, + 0.012845645658671856, + 0.006633681710809469, + 0.007751818746328354, + 0.009117676876485348, + 0.007761699613183737, + 0.014252321794629097, + 0.0026754073332995176, + 0.00183737906627357, + 0.003399631241336465, + 0.0023854700848460197, + 0.0023845427203923464, + 0.002171840751543641, + 0.0023007241543382406, + 0.0027642694767564535, + 0.0028040551114827394, + 0.0026459270156919956, + 0.0021035505924373865, + 0.0014893363695591688, + 0.003058507340028882, + 0.002929557580500841, + 0.0018323491094633937, + 0.0017928791930899024, + 0.0024112751707434654, + 0.0025966456159949303, + 0.0029083515983074903, + 0.0025862392503768206, + 0.0021251440048217773, + 0.002894678618758917, + 0.0029543903656303883, + 0.0025456598959863186, + 0.0021730356384068727, + 0.00210975157096982, + 0.004380366764962673, + 0.0019009846728295088, + 0.0017941243713721633, + 0.002051582792773843, + 0.0014051462057977915, + 0.0018042708979919553, + 0.008441467769443989, + 0.17165425419807434, + 0.004862048663198948, + 0.00279019633308053, + 0.17256687581539154, + 0.008172484114766121, + 0.0030374687630683184, + 0.002866203896701336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.1428481936454773 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193772338330746 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193772338330746, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193772338330746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027619877830147743, + 0.010859350673854351, + 0.009556950069963932, + 0.006427908316254616, + 0.004930576775223017, + 0.006070033181458712, + 0.010269874706864357, + 0.00849798135459423, + 0.017011871561408043, + 0.8987555503845215 + ], + "credit_map": [ + 0.0019246768206357956, + 0.002778352703899145, + 0.002493899082764983, + 0.0023379737976938486, + 0.004064818378537893, + 0.004595472477376461, + 0.00475281523540616, + 0.004671873524785042, + 0.000690075452439487, + 0.0027278559282422066, + 0.0009553359705023468, + 0.0007887347601354122, + 0.002776414854452014, + 0.000986081431619823, + 0.0008713533170521259, + 0.0010635004146024585, + 0.0006835244712419808, + 0.0009236767655238509, + 0.0010786234633997083, + 0.0007578147342428565, + 0.001994617749005556, + 0.001687365584075451, + 0.0011393075110390782, + 0.0012920205481350422, + 0.0004530636942945421, + 0.0008838996291160583, + 0.00041395961306989193, + 0.0008713534916751087, + 0.0012940973974764347, + 0.0011081863194704056, + 0.0005544916493818164, + 0.0008488575113005936, + 0.0005817227647639811, + 0.0006297272047959268, + 0.0005379452486522496, + 0.0005535798845812678, + 0.0005881682154722512, + 0.0005689804092980921, + 0.0004974562907591462, + 0.000972997397184372, + 0.0007304349564947188, + 0.0005392017774283886, + 0.0007187238661572337, + 0.0005464446148835123, + 0.0010043482761830091, + 0.0006739500677213073, + 0.0009563592029735446, + 0.0009005714673548937, + 0.001286303042434156, + 0.0015697844792157412, + 0.0010683505097404122, + 0.0009962780168280005, + 0.00216429284773767, + 0.00101514405105263, + 0.0011289279209449887, + 0.0010407953523099422, + 0.0014808546984568238, + 0.0010643050773069263, + 0.0010964983375743032, + 0.001091483049094677, + 0.0006145931547507644, + 0.0006611424032598734, + 0.0014186215121299028, + 0.0010704842861741781, + 0.0018814241047948599, + 0.002234348561614752, + 0.0038052101153880358, + 0.0019168922444805503, + 0.0026670522056519985, + 0.0012952272081747651, + 0.0016134819015860558, + 0.001598237082362175, + 0.019731270149350166, + 0.4174574613571167, + 0.00658406363800168, + 0.006680905818939209, + 0.416884183883667, + 0.019194582477211952, + 0.006597472820430994, + 0.005625701043754816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.014193772338330746 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly dynamic, low-angle, side-on tracking shot follows a modern police sedan at high speed. The car, sporting a classic black and white paint scheme, its rooftop lights flashing red and blue at a rapid tempo. It travels along a straight, empty asphalt road, framed by a pure, uniform light-gray sky that stretches seamlessly to the horizon, devoid of any clouds or landmarks. The car's body tilts slightly at breakneck speed, its glossy paint reflecting a soft, shadowless diffuse light, creating a cool, metallic quality. The tires' rubbing against the road kick up a faint cloud of dust, further accentuating its rapid motion. The composition is remarkably minimalist, primarily black, white, and gray. The flashing red and blue police lights serve as the sole splash of color within this minimalist composition, shattering the monochromatic silence and creating a powerful sense of speed, urgency, and a surreal atmosphere of isolation.", + "seed": 28114, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_9/index_6922_seed_28114_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16756333373486995, + "vlm_logprob": -0.16756333373486995, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.16756333373486995, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02736082673072815 + ], + "opposite_logprob_tokens": [ + -3.7773609161376953 + ], + "logprob_reward": -0.02736082673072815, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02736082673072815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5233120322227478, + 0.2641929090023041, + 0.019657626748085022, + 0.018350651487708092, + 0.03580009192228317, + 0.024067744612693787, + 0.03286008909344673, + 0.058355800807476044, + 0.008019438944756985, + 0.015383578836917877 + ], + "credit_map": [ + 0.013679767958819866, + 0.2135278284549713, + 0.01059707347303629, + 0.005655640736222267, + 0.23763659596443176, + 0.026903042569756508, + 0.008772220462560654, + 0.006539906375110149, + 0.005251158960163593, + 0.10548540204763412, + 0.003129406599327922, + 0.0021342416293919086, + 0.12380866706371307, + 0.019602064043283463, + 0.0026836504694074392, + 0.00209830142557621, + 0.000730067549739033, + 0.005477386061102152, + 0.0014972623903304338, + 0.0008892940240912139, + 0.00697666360065341, + 0.002160878386348486, + 0.0009302321705035865, + 0.0009958426235243678, + 0.0012216902105137706, + 0.0029483414255082607, + 0.002968066604807973, + 0.0029772305861115456, + 0.0029620351269841194, + 0.0017023388063535094, + 0.001295899273827672, + 0.002275049453601241, + 0.003072403371334076, + 0.009725377894937992, + 0.0045300680212676525, + 0.002834800397977233, + 0.0063865468837320805, + 0.0046303169801831245, + 0.0019634831696748734, + 0.0026570972986519337, + 0.0013337393756955862, + 0.005627606529742479, + 0.002923221793025732, + 0.0016985408728942275, + 0.006533811334520578, + 0.0021625536028295755, + 0.0019684117287397385, + 0.0018198589095845819, + 0.0009927765931934118, + 0.013699941337108612, + 0.0013478160835802555, + 0.0007957530324347317, + 0.011816659942269325, + 0.0023551471531391144, + 0.0006198182236403227, + 0.0012321746908128262, + 0.0013823880581185222, + 0.02212367206811905, + 0.0022785679902881384, + 0.0017384565435349941, + 0.025387512519955635, + 0.0031372811645269394, + 0.0010245863813906908, + 0.0012833394575864077, + 0.0006437661359086633, + 0.001665120362304151, + 0.0013371746754273772, + 0.001001778757199645, + 0.0009734794730320573, + 0.0014743105275556445, + 0.00041343754855915904, + 0.0005103717558085918, + 0.0015170545084401965, + 0.0018829801119863987, + 0.0035299581941217184, + 0.0025579212233424187, + 0.002060554688796401, + 0.0019661246333271265, + 0.001188334310427308, + 0.0006806503515690565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.02736082673072815 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234334468841553 + ], + "opposite_logprob_tokens": [ + -0.8984334468841553 + ], + "logprob_reward": -0.5234334468841553, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234334468841553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5047364830970764, + 0.2398701161146164, + 0.048560962080955505, + 0.03439813107252121, + 0.015230078250169754, + 0.025532003492116928, + 0.04144834354519844, + 0.01650780253112316, + 0.050410185009241104, + 0.02330598421394825 + ], + "credit_map": [ + 0.010856707580387592, + 0.16808070242404938, + 0.009810300543904305, + 0.007199910935014486, + 0.2321336269378662, + 0.05888636037707329, + 0.011032338254153728, + 0.006736456416547298, + 0.00474458048120141, + 0.09887026250362396, + 0.004834611900150776, + 0.004524432122707367, + 0.10832981765270233, + 0.011440714821219444, + 0.0036843265406787395, + 0.0034413475077599287, + 0.003016304224729538, + 0.01349796261638403, + 0.0036257838364690542, + 0.002835117047652602, + 0.012891392223536968, + 0.005798309575766325, + 0.0026430797297507524, + 0.004253006540238857, + 0.002736140741035342, + 0.004776958841830492, + 0.005112807732075453, + 0.004929103888571262, + 0.0035687878262251616, + 0.005470335017889738, + 0.0021100421436131, + 0.005693950690329075, + 0.0012825911398977041, + 0.0030521913431584835, + 0.002402976620942354, + 0.0014223998878151178, + 0.0030490977223962545, + 0.0019227428128942847, + 0.0010454516159370542, + 0.0010526267578825355, + 0.00284850737079978, + 0.005725253373384476, + 0.002826729556545615, + 0.0014045697171241045, + 0.0030716783367097378, + 0.007288879249244928, + 0.0011152076767757535, + 0.0012511772802099586, + 0.001669646124355495, + 0.013446581549942493, + 0.0027503122109919786, + 0.0015959820011630654, + 0.015679772943258286, + 0.003405417548492551, + 0.0009662607917562127, + 0.0019343658350408077, + 0.001456990372389555, + 0.0020664455369114876, + 0.002136679133400321, + 0.0015480861766263843, + 0.004086825996637344, + 0.002837895881384611, + 0.0009525794303044677, + 0.0014222983736544847, + 0.00451771030202508, + 0.007612185087054968, + 0.006192237604409456, + 0.006137609947472811, + 0.00783852580934763, + 0.010495945811271667, + 0.0037967527750879526, + 0.0038192132487893105, + 0.0015541411703452468, + 0.00489431619644165, + 0.002880518790334463, + 0.003384568030014634, + 0.0029890546575188637, + 0.002966819331049919, + 0.00263407570309937, + 0.002002488821744919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.5234334468841553 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023303797468543053 + ], + "opposite_logprob_tokens": [ + -3.773303747177124 + ], + "logprob_reward": -0.023303797468543053, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023303797468543053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4933791756629944, + 0.3809188902378082, + 0.009385887533426285, + 0.015040130354464054, + 0.00994805246591568, + 0.008461588993668556, + 0.004456930328160524, + 0.022903602570295334, + 0.012720595113933086, + 0.04278520122170448 + ], + "credit_map": [ + 0.014204727485775948, + 0.20121659338474274, + 0.014820490032434464, + 0.00823673140257597, + 0.21538470685482025, + 0.021730458363890648, + 0.009812921285629272, + 0.00797255802899599, + 0.007895668037235737, + 0.1542791724205017, + 0.003972020465880632, + 0.003069068770855665, + 0.1789027601480484, + 0.026662131771445274, + 0.0035448134876787663, + 0.0025932546705007553, + 0.0009239120408892632, + 0.0014800374628975987, + 0.0019864265341311693, + 0.0006694674375467002, + 0.0009052365785464644, + 0.0019485807279124856, + 0.0006338309613056481, + 0.0008383949752897024, + 0.0008347489056177437, + 0.003660726360976696, + 0.0018790052272379398, + 0.0012786936713382602, + 0.002104312414303422, + 0.002938664983958006, + 0.0006429307977668941, + 0.001701047644019127, + 0.0009211681317538023, + 0.0009882258018478751, + 0.0016513997688889503, + 0.0006166069651953876, + 0.0022956891916692257, + 0.002307382645085454, + 0.00039516756078228354, + 0.0007724113529548049, + 0.0009018086711876094, + 0.0016706667374819517, + 0.0014463308034464717, + 0.0006010137731209397, + 0.0012775001814588904, + 0.0016997059574350715, + 0.0003749942989088595, + 0.0004895685124211013, + 0.0005563970771618187, + 0.0006913103861734271, + 0.000928445952013135, + 0.0003440075961407274, + 0.0005488498136401176, + 0.0006649275892414153, + 0.00019066296226810664, + 0.0005323289660736918, + 0.0011552309151738882, + 0.00816471129655838, + 0.001987566240131855, + 0.0009759285603649914, + 0.007879144512116909, + 0.001481783576309681, + 0.0004950373549945652, + 0.0007641995325684547, + 0.0012781328987330198, + 0.002129633678123355, + 0.002132748020812869, + 0.0011003100080415606, + 0.0030977095011621714, + 0.0016791025409474969, + 0.0006056423881091177, + 0.0006973169511184096, + 0.0024411696940660477, + 0.01211154367774725, + 0.004576374311000109, + 0.004763379227370024, + 0.010589362122118473, + 0.004462994169443846, + 0.002047798130661249, + 0.0017925777938216925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.023303797468543053 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22562669217586517 + ], + "opposite_logprob_tokens": [ + -1.6006267070770264 + ], + "logprob_reward": -0.22562669217586517, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22562669217586517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04629475995898247, + 0.11582094430923462, + 0.07750318199396133, + 0.03759169951081276, + 0.02472102828323841, + 0.06360207498073578, + 0.09266284108161926, + 0.23599928617477417, + 0.057844165712594986, + 0.24796001613140106 + ], + "credit_map": [ + 0.0027449235785752535, + 0.00972472969442606, + 0.007364708464592695, + 0.0029097332153469324, + 0.007134162355214357, + 0.005439888220280409, + 0.006915324367582798, + 0.00406129565089941, + 0.002403843682259321, + 0.043837275356054306, + 0.002409864217042923, + 0.001905673649162054, + 0.052568964660167694, + 0.00955837219953537, + 0.0015938233118504286, + 0.0015431413194164634, + 0.004987977910786867, + 0.02304999530315399, + 0.0062603214755654335, + 0.003626131219789386, + 0.021811025217175484, + 0.008489503525197506, + 0.0037352675572037697, + 0.0055429707281291485, + 0.0025892341509461403, + 0.004919718485325575, + 0.005895922891795635, + 0.006652359385043383, + 0.0052242823876440525, + 0.003884483128786087, + 0.003236619522795081, + 0.0051890830509364605, + 0.001628460711799562, + 0.007071000523865223, + 0.0033800597302615643, + 0.0014305630465969443, + 0.006185468751937151, + 0.0026725702919065952, + 0.0009578693425282836, + 0.0013950368156656623, + 0.003675422165542841, + 0.019942771643400192, + 0.004899959079921246, + 0.003822626546025276, + 0.009450498037040234, + 0.015042109414935112, + 0.003303009318187833, + 0.0034656869247555733, + 0.0033777011558413506, + 0.03647340461611748, + 0.004208688158541918, + 0.003096505766734481, + 0.03389696404337883, + 0.006026728078722954, + 0.0018344662385061383, + 0.003748390357941389, + 0.0064740558154881, + 0.08511354774236679, + 0.009616526775062084, + 0.0079851895570755, + 0.10158421844244003, + 0.014279869385063648, + 0.004331222735345364, + 0.006614676211029291, + 0.006225882563740015, + 0.0067952158860862255, + 0.013940911740064621, + 0.0064121331088244915, + 0.007649260573089123, + 0.009446228854358196, + 0.003727773204445839, + 0.003646771190688014, + 0.01686149090528488, + 0.05514238029718399, + 0.04281982406973839, + 0.030753465369343758, + 0.04312223941087723, + 0.03484821692109108, + 0.012954684905707836, + 0.011457723565399647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.22562669217586517 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038091905415058136 + ], + "opposite_logprob_tokens": [ + -3.2880918979644775 + ], + "logprob_reward": -0.038091905415058136, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038091905415058136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19584165513515472, + 0.32531100511550903, + 0.021310390904545784, + 0.05254135653376579, + 0.05449090898036957, + 0.10420837253332138, + 0.05191272124648094, + 0.08807742595672607, + 0.025009874254465103, + 0.08129633218050003 + ], + "credit_map": [ + 0.010556299239397049, + 0.017990464344620705, + 0.016222763806581497, + 0.009861554950475693, + 0.054403431713581085, + 0.06133449450135231, + 0.014770432375371456, + 0.010702192783355713, + 0.007454602513462305, + 0.1285613477230072, + 0.0072184219025075436, + 0.004785115830600262, + 0.1466366946697235, + 0.02170446515083313, + 0.004550302866846323, + 0.004400030244141817, + 0.002313438104465604, + 0.0028329973574727774, + 0.004729290027171373, + 0.0014190793735906482, + 0.0027208314277231693, + 0.004227121360599995, + 0.0010798037983477116, + 0.0019878270104527473, + 0.0042190710082650185, + 0.006668763235211372, + 0.011368712410330772, + 0.006793626118451357, + 0.009652881883084774, + 0.006411412265151739, + 0.0015709707513451576, + 0.00585591234266758, + 0.002407171530649066, + 0.02179265394806862, + 0.0035337775480002165, + 0.002206195145845413, + 0.010473689064383507, + 0.01142757199704647, + 0.0009855973767116666, + 0.001664252602495253, + 0.006163505837321281, + 0.037033140659332275, + 0.010333108715713024, + 0.005094082094728947, + 0.026616262272000313, + 0.012771153822541237, + 0.0023260489106178284, + 0.0038710590451955795, + 0.002634260803461075, + 0.019658997654914856, + 0.0034915897995233536, + 0.001534795854240656, + 0.015682164579629898, + 0.0053182137198746204, + 0.0010375166311860085, + 0.0025551780126988888, + 0.003269593697041273, + 0.027237189933657646, + 0.005184793844819069, + 0.004189236555248499, + 0.03563343361020088, + 0.007166448514908552, + 0.0023927907459437847, + 0.003003934631124139, + 0.002042957814410329, + 0.0035057684872299433, + 0.002986787585541606, + 0.003473857184872031, + 0.004136297851800919, + 0.005276720505207777, + 0.0018139817984774709, + 0.0017735001165419817, + 0.008272544480860233, + 0.00940202921628952, + 0.0175456665456295, + 0.014157451689243317, + 0.011540206149220467, + 0.00961918756365776, + 0.006058746017515659, + 0.004700495861470699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.038091905415058136 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly dynamic, low-angle, side-on tracking shot follows a modern police sedan at high speed. The car, sporting a classic black and white paint scheme, its rooftop lights flashing red and blue at a rapid tempo. It travels along a straight, empty asphalt road, framed by a pure, uniform light-gray sky that stretches seamlessly to the horizon, devoid of any clouds or landmarks. The car's body tilts slightly at breakneck speed, its glossy paint reflecting a soft, shadowless diffuse light, creating a cool, metallic quality. The tires' rubbing against the road kick up a faint cloud of dust, further accentuating its rapid motion. The composition is remarkably minimalist, primarily black, white, and gray. The flashing red and blue police lights serve as the sole splash of color within this minimalist composition, shattering the monochromatic silence and creating a powerful sense of speed, urgency, and a surreal atmosphere of isolation.", + "seed": 28115, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_9/index_6922_seed_28115_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2234877362847328, + "vlm_logprob": -0.2234877362847328, + "vlm_yes_no_margin": 2.499999988079071, + "vlm_token_credit": -0.2234877362847328, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027381589636206627 + ], + "opposite_logprob_tokens": [ + -3.777381658554077 + ], + "logprob_reward": -0.027381589636206627, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027381589636206627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7865208983421326, + 0.01634947396814823, + 0.008600630797445774, + 0.06326217204332352, + 0.02274991385638714, + 0.018119530752301216, + 0.01591353677213192, + 0.014863144606351852, + 0.04271560162305832, + 0.010905112139880657 + ], + "credit_map": [ + 0.013885311782360077, + 0.30115658044815063, + 0.008238831534981728, + 0.008358163759112358, + 0.36890271306037903, + 0.0713900700211525, + 0.008914126083254814, + 0.005675096996128559, + 0.0013361726887524128, + 0.001544626196846366, + 0.002769887214526534, + 0.0015884601743891835, + 0.0034520705230534077, + 0.0026496059726923704, + 0.0015104033518582582, + 0.001498248428106308, + 0.0005537818069569767, + 0.0007522592786699533, + 0.0008621235028840601, + 0.0007164478884078562, + 0.0023173384834080935, + 0.0019423398189246655, + 0.0005579538992606103, + 0.0008983867592178285, + 0.0015800660476088524, + 0.028794381767511368, + 0.0010202634148299694, + 0.0011390839936211705, + 0.020467344671487808, + 0.008472170680761337, + 0.0008367221453227103, + 0.0009521375177428126, + 0.001476710312999785, + 0.004486082121729851, + 0.0019474094733595848, + 0.0018606401281431317, + 0.006105702370405197, + 0.002730252221226692, + 0.0019089790293946862, + 0.002234137849882245, + 0.0012193367583677173, + 0.00333987595513463, + 0.0019230416510254145, + 0.0024540803860872984, + 0.0018326316494494677, + 0.003941028844565153, + 0.0017971285851672292, + 0.0016124069225043058, + 0.0006937094149179757, + 0.005260095000267029, + 0.0011278429301455617, + 0.0007145479321479797, + 0.005730264820158482, + 0.0012370254844427109, + 0.0005022073164582253, + 0.0006478448049165308, + 0.0009141830378212035, + 0.0014971834607422352, + 0.0019368833163753152, + 0.00048367565614171326, + 0.004647215828299522, + 0.004003912676125765, + 0.0007597752846777439, + 0.0006203161319717765, + 0.0032576255034655333, + 0.00354182836599648, + 0.0037181596271693707, + 0.0032499281223863363, + 0.010127997025847435, + 0.01040357630699873, + 0.004105635453015566, + 0.004310850519686937, + 0.00045458500972017646, + 0.00420368043705821, + 0.000734019442461431, + 0.0003578884934540838, + 0.003395399311557412, + 0.0008771627326495945, + 0.00040843486203812063, + 0.00047394196735695004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.027381589636206627 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934697031974792 + ], + "opposite_logprob_tokens": [ + -0.6934697031974792 + ], + "logprob_reward": -0.6934697031974792, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934697031974792, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.46169513463974, + 0.03864511102437973, + 0.01738661527633667, + 0.20467041432857513, + 0.04504634439945221, + 0.043125513941049576, + 0.0340031273663044, + 0.022001566365361214, + 0.07358390092849731, + 0.059842295944690704 + ], + "credit_map": [ + 0.008182627148926258, + 0.129558727145195, + 0.010109197348356247, + 0.009128902107477188, + 0.21178941428661346, + 0.07903841137886047, + 0.0067332130856812, + 0.007154632825404406, + 0.0022481305059045553, + 0.009104722179472446, + 0.005235306452959776, + 0.0027795506175607443, + 0.010947777889668941, + 0.0035541271790862083, + 0.0020744025241583586, + 0.0027010932099074125, + 0.0009314729832112789, + 0.004106799140572548, + 0.0019715256057679653, + 0.0012296189088374376, + 0.0048341695219278336, + 0.0014444526750594378, + 0.0010047522373497486, + 0.0018638246692717075, + 0.004513611551374197, + 0.06534029543399811, + 0.0016874616267159581, + 0.002831770572811365, + 0.09578349441289902, + 0.03014618158340454, + 0.002225299132987857, + 0.002142309909686446, + 0.0017704122001305223, + 0.0030180688481777906, + 0.0027350103482604027, + 0.00385780306532979, + 0.015235922299325466, + 0.013356601819396019, + 0.0026063418481498957, + 0.0024661843199282885, + 0.001956938998773694, + 0.005709454882889986, + 0.003285976592451334, + 0.003080864204093814, + 0.014140048995614052, + 0.008981958031654358, + 0.003319381270557642, + 0.002650890499353409, + 0.0021030688658356667, + 0.009117565117776394, + 0.0034825250040739775, + 0.0025287363678216934, + 0.00760574359446764, + 0.0037208956200629473, + 0.0031696027144789696, + 0.0022749893832951784, + 0.00174370885360986, + 0.0028286653105169535, + 0.0028243069536983967, + 0.0012883603340014815, + 0.004162486642599106, + 0.006187784019857645, + 0.0014738584868609905, + 0.001492396811954677, + 0.0037838164716959, + 0.01081003900617361, + 0.004810039885342121, + 0.004133862443268299, + 0.015931544825434685, + 0.023980770260095596, + 0.005377490073442459, + 0.004756333772093058, + 0.0038050764705985785, + 0.007705347612500191, + 0.007309287786483765, + 0.0018153581768274307, + 0.020906273275613785, + 0.013738562352955341, + 0.0026041690725833178, + 0.00195821444503963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.6934697031974792 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009791677817702293 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.009791677817702293, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791677817702293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7980451583862305, + 0.02291167341172695, + 0.007961854338645935, + 0.10987014323472977, + 0.014517799951136112, + 0.012592973187565804, + 0.010620832443237305, + 0.0026945772115141153, + 0.008653105236589909, + 0.012131867930293083 + ], + "credit_map": [ + 0.013888959772884846, + 0.3068881034851074, + 0.003573501482605934, + 0.00576150743290782, + 0.38105401396751404, + 0.0767621174454689, + 0.006958703510463238, + 0.0031582978554069996, + 0.0020702832844108343, + 0.001972789177671075, + 0.005418888758867979, + 0.002152801491320133, + 0.003579381387680769, + 0.003617311595007777, + 0.0017254226841032505, + 0.0023747961968183517, + 0.00042704836232587695, + 0.001995702041313052, + 0.0010053947335109115, + 0.000482204370200634, + 0.00149833713658154, + 0.0011323466897010803, + 0.00035967316944152117, + 0.001061149057932198, + 0.0021303677931427956, + 0.04536035284399986, + 0.0006964409840293229, + 0.0010972080053761601, + 0.051292624324560165, + 0.007400640286505222, + 0.0010137245990335941, + 0.0008787987753748894, + 0.001577541115693748, + 0.0018387940945103765, + 0.002117648720741272, + 0.0016578617505729198, + 0.0019592090975493193, + 0.0021788463927805424, + 0.0010884659131988883, + 0.002099434146657586, + 0.0008232726831920445, + 0.0015726094134151936, + 0.0012736013159155846, + 0.001624205382540822, + 0.0018683495000004768, + 0.0030055653769522905, + 0.0012698008213192225, + 0.0011555703822523355, + 0.0005774881574325264, + 0.0014811287401244044, + 0.000981622375547886, + 0.000435618800111115, + 0.003768881317228079, + 0.0024506261106580496, + 0.0004550212761387229, + 0.00047044630628079176, + 0.0002009730669669807, + 0.0002951150818262249, + 0.00033414681092835963, + 0.0001509066205471754, + 0.000782558461651206, + 0.0006132275448180735, + 0.0001478872582083568, + 0.00016976261395029724, + 0.00068138592178002, + 0.0010140383383259177, + 0.0009490679949522018, + 0.0006467623170465231, + 0.0016218862729147077, + 0.0017177884001284838, + 0.0010749356588348746, + 0.0009472419624216855, + 0.0008255683351308107, + 0.003427766263484955, + 0.0016047367826104164, + 0.00046357422252185643, + 0.0036233654245734215, + 0.0010976827470585704, + 0.0005377903580665588, + 0.0005513852811418474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.009791677817702293 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3487008213996887 + ], + "opposite_logprob_tokens": [ + -1.223700761795044 + ], + "logprob_reward": -0.3487008213996887, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3487008213996887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5076281428337097, + 0.03887972608208656, + 0.027561144903302193, + 0.15000927448272705, + 0.06316475570201874, + 0.10915853083133698, + 0.02828548103570938, + 0.015434607863426208, + 0.025625349953770638, + 0.03425298631191254 + ], + "credit_map": [ + 0.010401714593172073, + 0.22491484880447388, + 0.0053164283744990826, + 0.005570669192820787, + 0.22929304838180542, + 0.02299247682094574, + 0.00452387286350131, + 0.004615106154233217, + 0.0027543262112885714, + 0.003891686210408807, + 0.007037143688648939, + 0.00351971248164773, + 0.00874087493866682, + 0.005917806178331375, + 0.003065455472096801, + 0.003952718805521727, + 0.0016898552421480417, + 0.007521138060837984, + 0.0030602463521063328, + 0.002154589630663395, + 0.002404715633019805, + 0.007125464268028736, + 0.0018186363158747554, + 0.00178649858571589, + 0.008293460123240948, + 0.05006908252835274, + 0.008557960391044617, + 0.013825024478137493, + 0.03403756022453308, + 0.01742311753332615, + 0.009484333917498589, + 0.008318734355270863, + 0.0027170958928763866, + 0.00786403939127922, + 0.0034757512621581554, + 0.0035836519673466682, + 0.02150709740817547, + 0.015394443646073341, + 0.003101982641965151, + 0.005520687438547611, + 0.005991167388856411, + 0.008829827420413494, + 0.008761242032051086, + 0.013249671086668968, + 0.028983671218156815, + 0.023866653442382812, + 0.010995696298778057, + 0.008480594493448734, + 0.0020184398163110018, + 0.006790419574826956, + 0.004189662169665098, + 0.0020994781516492367, + 0.0061703333631157875, + 0.002979639219120145, + 0.0021005591843277216, + 0.001936950022354722, + 0.0011503332061693072, + 0.004197811242192984, + 0.0018031069776043296, + 0.0010230132611468434, + 0.002456487389281392, + 0.0026187170296907425, + 0.0011607073247432709, + 0.0010244310833513737, + 0.0014491602778434753, + 0.003122687805444002, + 0.002187073929235339, + 0.0014762200880795717, + 0.005628054030239582, + 0.008310096338391304, + 0.0020331114064902067, + 0.0014189465437084436, + 0.0032810841221362352, + 0.006260734982788563, + 0.007249715272337198, + 0.0018262394005432725, + 0.006441435776650906, + 0.004654272459447384, + 0.002409469336271286, + 0.002130033215507865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.3487008213996887 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038094889372587204 + ], + "opposite_logprob_tokens": [ + -3.288094997406006 + ], + "logprob_reward": -0.038094889372587204, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038094889372587204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5441753268241882, + 0.0420108288526535, + 0.049655981361866, + 0.10760606825351715, + 0.02637118473649025, + 0.011205471120774746, + 0.027432916685938835, + 0.10584712028503418, + 0.007848208770155907, + 0.07784689962863922 + ], + "credit_map": [ + 0.010446920990943909, + 0.20928223431110382, + 0.005166669376194477, + 0.00469555240124464, + 0.2558863162994385, + 0.04967387020587921, + 0.005114251282066107, + 0.003909530583769083, + 0.0034510449040681124, + 0.005065206438302994, + 0.00892558041960001, + 0.003315496491268277, + 0.008762218058109283, + 0.005677854176610708, + 0.0026701781898736954, + 0.0041432492434978485, + 0.0028734151273965836, + 0.010491706430912018, + 0.004175223410129547, + 0.0022172224707901478, + 0.007735515013337135, + 0.01602538675069809, + 0.002096108626574278, + 0.004041407722979784, + 0.002042752457782626, + 0.034598395228385925, + 0.0010810412932187319, + 0.002013249322772026, + 0.04974067583680153, + 0.015060913749039173, + 0.0016799427103251219, + 0.001389104756526649, + 0.0015981554752215743, + 0.0018487406196072698, + 0.0025467216037213802, + 0.0021241498179733753, + 0.007577687967568636, + 0.006887945346534252, + 0.0021448477637022734, + 0.001642937189899385, + 0.0008662681211717427, + 0.002609420334920287, + 0.0012392426142469049, + 0.0011572913499549031, + 0.0024125799536705017, + 0.0009249934810213745, + 0.000925779459066689, + 0.001069897087290883, + 0.0014411858282983303, + 0.0040954300202429295, + 0.0024028574116528034, + 0.0013617206132039428, + 0.009589196182787418, + 0.005820257589221001, + 0.001569528947584331, + 0.0011527432361617684, + 0.005331552121788263, + 0.00846824236214161, + 0.009617025963962078, + 0.004332988057285547, + 0.030649984255433083, + 0.03537749871611595, + 0.00560959056019783, + 0.006460257340222597, + 0.00043846043990924954, + 0.0006983939092606306, + 0.0010077878832817078, + 0.0004039925697725266, + 0.002367396606132388, + 0.0019924682565033436, + 0.0004979473305866122, + 0.00044176215305924416, + 0.004428015556186438, + 0.014745675958693027, + 0.008345860987901688, + 0.002158511197194457, + 0.028762932866811752, + 0.014468803070485592, + 0.002670996356755495, + 0.0022661127150058746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.038094889372587204 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly dynamic, low-angle, side-on tracking shot follows a modern police sedan at high speed. The car, sporting a classic black and white paint scheme, its rooftop lights flashing red and blue at a rapid tempo. It travels along a straight, empty asphalt road, framed by a pure, uniform light-gray sky that stretches seamlessly to the horizon, devoid of any clouds or landmarks. The car's body tilts slightly at breakneck speed, its glossy paint reflecting a soft, shadowless diffuse light, creating a cool, metallic quality. The tires' rubbing against the road kick up a faint cloud of dust, further accentuating its rapid motion. The composition is remarkably minimalist, primarily black, white, and gray. The flashing red and blue police lights serve as the sole splash of color within this minimalist composition, shattering the monochromatic silence and creating a powerful sense of speed, urgency, and a surreal atmosphere of isolation.", + "seed": 28116, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000053/group_9/index_6922_seed_28116_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07145326286554336, + "vlm_logprob": -0.07145326286554336, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.07145326286554336, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012006985023617744 + ], + "opposite_logprob_tokens": [ + -5.012006759643555 + ], + "logprob_reward": -0.012006985023617744, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.012006985023617744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16545067727565765, + 0.042934488505125046, + 0.0899757668375969, + 0.05273640900850296, + 0.10300414264202118, + 0.07879883050918579, + 0.11327041685581207, + 0.12691470980644226, + 0.16437220573425293, + 0.0625423938035965 + ], + "credit_map": [ + 0.021381454542279243, + 0.01823532208800316, + 0.01756059005856514, + 0.01641770824790001, + 0.036881547421216965, + 0.023074572905898094, + 0.017334753647446632, + 0.014564727433025837, + 0.00538301095366478, + 0.004748089704662561, + 0.004241577815264463, + 0.003952780272811651, + 0.003565005026757717, + 0.005994731094688177, + 0.0074067628011107445, + 0.007642530836164951, + 0.01129264198243618, + 0.011172644793987274, + 0.010061918757855892, + 0.01028166338801384, + 0.012496136128902435, + 0.013664398342370987, + 0.00947527401149273, + 0.01153109036386013, + 0.004967130720615387, + 0.006536439526826143, + 0.004459158051759005, + 0.005895753391087055, + 0.0058247908018529415, + 0.007496533915400505, + 0.007262353785336018, + 0.010294253006577492, + 0.004679562523961067, + 0.03972689062356949, + 0.0015922111924737692, + 0.0033990389201790094, + 0.008985747583210468, + 0.03809839487075806, + 0.004213688429445028, + 0.002308606868609786, + 0.009701003320515156, + 0.009484644047915936, + 0.009850251488387585, + 0.006667308043688536, + 0.01105167530477047, + 0.01199313160032034, + 0.008823432959616184, + 0.011227386072278023, + 0.013511487282812595, + 0.013903921470046043, + 0.0129171097651124, + 0.0066286721266806126, + 0.013323080725967884, + 0.033225592225790024, + 0.006588289979845285, + 0.013172263279557228, + 0.013673296198248863, + 0.016419220715761185, + 0.01801304891705513, + 0.012891117483377457, + 0.01908905990421772, + 0.018466418609023094, + 0.011397122405469418, + 0.016965428367257118, + 0.01741102896630764, + 0.021631091833114624, + 0.012123430147767067, + 0.010521143674850464, + 0.042543575167655945, + 0.017529815435409546, + 0.013742746785283089, + 0.028869373723864555, + 0.008345206268131733, + 0.008453535847365856, + 0.007223019376397133, + 0.005514428950846195, + 0.008352991193532944, + 0.007019371725618839, + 0.007398913148790598, + 0.01023492868989706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.012006985023617744 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11304467916488647 + ], + "opposite_logprob_tokens": [ + -2.2380447387695312 + ], + "logprob_reward": -0.11304467916488647, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11304467916488647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06870803982019424, + 0.06554526835680008, + 0.1924324780702591, + 0.09205181151628494, + 0.14662490785121918, + 0.06251996010541916, + 0.07728523015975952, + 0.08186152577400208, + 0.19638997316360474, + 0.016580840572714806 + ], + "credit_map": [ + 0.012252593412995338, + 0.006780101452022791, + 0.007735079620033503, + 0.007061672396957874, + 0.01286865584552288, + 0.008102286607027054, + 0.004131005145609379, + 0.009776637889444828, + 0.009141817688941956, + 0.008705678395926952, + 0.005240917671471834, + 0.005745737347751856, + 0.009565322659909725, + 0.008817290887236595, + 0.007788882590830326, + 0.010539611801505089, + 0.018756218254566193, + 0.02592083252966404, + 0.012833760119974613, + 0.019042355939745903, + 0.04309088736772537, + 0.031620558351278305, + 0.021970516070723534, + 0.019197333604097366, + 0.008705626241862774, + 0.011713964864611626, + 0.008812724612653255, + 0.004485622048377991, + 0.020223522558808327, + 0.013776388950645924, + 0.010127232410013676, + 0.01420671958476305, + 0.012558290734887123, + 0.034882836043834686, + 0.006224638782441616, + 0.007573902606964111, + 0.03138144686818123, + 0.03146792948246002, + 0.010067506693303585, + 0.012468354776501656, + 0.005754097830504179, + 0.007826549001038074, + 0.004288522992283106, + 0.005157028324902058, + 0.016247497871518135, + 0.009406810626387596, + 0.007410502061247826, + 0.0064289444126188755, + 0.009621258825063705, + 0.010586266405880451, + 0.006656299345195293, + 0.005025562830269337, + 0.011017827317118645, + 0.02101997099816799, + 0.007339641451835632, + 0.006018390879034996, + 0.009213721379637718, + 0.01131352037191391, + 0.016085222363471985, + 0.007059234194457531, + 0.01033159252256155, + 0.009574678726494312, + 0.006438072305172682, + 0.011845473200082779, + 0.016631323844194412, + 0.028531044721603394, + 0.02282765507698059, + 0.019181115552783012, + 0.031360428780317307, + 0.03219852223992348, + 0.026206599548459053, + 0.019453268498182297, + 0.002110792091116309, + 0.002634084317833185, + 0.0014686300419270992, + 0.0012752630282193422, + 0.0031909006647765636, + 0.0024508077185600996, + 0.0015668317209929228, + 0.001883528078906238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.11304467916488647 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206347987055779 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.014206347987055779, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206347987055779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07402848452329636, + 0.0626797080039978, + 0.04907788336277008, + 0.19002851843833923, + 0.24786226451396942, + 0.10594535619020462, + 0.11733528971672058, + 0.038602083921432495, + 0.0714833065867424, + 0.04295710474252701 + ], + "credit_map": [ + 0.011258703656494617, + 0.008143461309373379, + 0.006694515701383352, + 0.007536543067544699, + 0.014070037752389908, + 0.00973194744437933, + 0.006430937442928553, + 0.01016232930123806, + 0.007857215590775013, + 0.0063134473748505116, + 0.007183085195720196, + 0.005623755976557732, + 0.014304953627288342, + 0.007752916309982538, + 0.00516459671780467, + 0.008479735814034939, + 0.005536890588700771, + 0.005058306735008955, + 0.005003668367862701, + 0.005618009716272354, + 0.00915012788027525, + 0.006964281667023897, + 0.0055443961173295975, + 0.006202203221619129, + 0.01891965977847576, + 0.022558914497494698, + 0.017650078982114792, + 0.012827711179852486, + 0.05356023833155632, + 0.026633117347955704, + 0.013536151498556137, + 0.024342646822333336, + 0.0118559580296278, + 0.09361615777015686, + 0.00509287416934967, + 0.006646833848208189, + 0.02220088802278042, + 0.09311357140541077, + 0.006224090699106455, + 0.009111886844038963, + 0.01434233970940113, + 0.013460114598274231, + 0.022360116243362427, + 0.008172173984348774, + 0.008843603543937206, + 0.010384678840637207, + 0.010466661304235458, + 0.017915664240717888, + 0.017320767045021057, + 0.015670832246541977, + 0.0223922748118639, + 0.009809095412492752, + 0.011518198996782303, + 0.02257499285042286, + 0.008095555007457733, + 0.009953576140105724, + 0.005060388706624508, + 0.004726216197013855, + 0.007997704669833183, + 0.0033862018026411533, + 0.0036300714127719402, + 0.004545322619378567, + 0.0039390320889651775, + 0.005317150615155697, + 0.007044254336506128, + 0.008805038407444954, + 0.006796083878725767, + 0.005173749756067991, + 0.015327539294958115, + 0.007790744304656982, + 0.0054256245493888855, + 0.01512027345597744, + 0.004989930894225836, + 0.004790653008967638, + 0.00563232833519578, + 0.0026376505848020315, + 0.007281898055225611, + 0.004975813906639814, + 0.0049916841089725494, + 0.007657146081328392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.014206347987055779 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991903424263 + ], + "opposite_logprob_tokens": [ + -1.8049190044403076 + ], + "logprob_reward": -0.17991903424263, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991903424263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1262740045785904, + 0.04837552458047867, + 0.061148736625909805, + 0.28336188197135925, + 0.3008395731449127, + 0.03279133513569832, + 0.03146112337708473, + 0.03340188413858414, + 0.06388770043849945, + 0.018458247184753418 + ], + "credit_map": [ + 0.016038665547966957, + 0.01456422172486782, + 0.013082426972687244, + 0.013996748253703117, + 0.025158124044537544, + 0.01732192188501358, + 0.011100860312581062, + 0.015011020936071873, + 0.007759512402117252, + 0.004444153048098087, + 0.0055402847938239574, + 0.005263075698167086, + 0.006428311578929424, + 0.006404948886483908, + 0.006758800242096186, + 0.00577643932774663, + 0.005958196707069874, + 0.006913069169968367, + 0.005357711110264063, + 0.005072516389191151, + 0.013690019026398659, + 0.008644786663353443, + 0.007925118319690228, + 0.007587311789393425, + 0.02826036512851715, + 0.03827863931655884, + 0.02725781314074993, + 0.015437480062246323, + 0.05437380075454712, + 0.04030103236436844, + 0.026822945103049278, + 0.05262978747487068, + 0.010303208604454994, + 0.13014085590839386, + 0.003750012256205082, + 0.0036852939520031214, + 0.012195988558232784, + 0.12736113369464874, + 0.006502791773527861, + 0.006900261156260967, + 0.005797631572932005, + 0.004867031238973141, + 0.00481194956228137, + 0.002623993204906583, + 0.0043144566006958485, + 0.004040601197630167, + 0.0029171255882829428, + 0.003418540582060814, + 0.003574274480342865, + 0.003988681361079216, + 0.004415517207235098, + 0.002000075066462159, + 0.0033953702077269554, + 0.007109425961971283, + 0.0026325336657464504, + 0.004345243331044912, + 0.00385348591953516, + 0.004330303054302931, + 0.007152904756367207, + 0.0022232194896787405, + 0.005372270010411739, + 0.0036996041890233755, + 0.0024989787489175797, + 0.004271115642040968, + 0.006197814829647541, + 0.009985098615288734, + 0.006378867197781801, + 0.005760340951383114, + 0.012725015170872211, + 0.008339899592101574, + 0.00538595300167799, + 0.009114709682762623, + 0.0018704000394791365, + 0.0023125214502215385, + 0.0017413259483873844, + 0.0015481036389246583, + 0.003750674892216921, + 0.002814520848914981, + 0.0019218107918277383, + 0.0024988879449665546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.17991903424263 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police sedan with a black and white paint scheme?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are red and blue lights flashing on the car's rooftop at a rapid tempo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car traveling along a straight, empty asphalt road with a light-gray cloudless sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's body tilt slightly while moving at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle side-on tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038089267909526825 + ], + "opposite_logprob_tokens": [ + -3.2880892753601074 + ], + "logprob_reward": -0.038089267909526825, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038089267909526825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12663869559764862, + 0.0660761296749115, + 0.13926920294761658, + 0.19917084276676178, + 0.07960157841444016, + 0.06201566383242607, + 0.03827190771698952, + 0.09834570437669754, + 0.08664290606975555, + 0.10396740585565567 + ], + "credit_map": [ + 0.01707502454519272, + 0.011373155750334263, + 0.014929805882275105, + 0.01586553454399109, + 0.02372889220714569, + 0.015348156914114952, + 0.011360932141542435, + 0.016957197338342667, + 0.007923784665763378, + 0.007481636945158243, + 0.009355196729302406, + 0.0058468240313231945, + 0.007706410717219114, + 0.008953699842095375, + 0.005235392600297928, + 0.013573186472058296, + 0.014388441108167171, + 0.018106568604707718, + 0.011319565586745739, + 0.013418100774288177, + 0.02990277297794819, + 0.02082659676671028, + 0.01672668382525444, + 0.01458046492189169, + 0.0220062043517828, + 0.026683064177632332, + 0.01695723831653595, + 0.01512712612748146, + 0.0422586053609848, + 0.024522384628653526, + 0.018937168642878532, + 0.03267902880907059, + 0.006205549463629723, + 0.02574930526316166, + 0.0039823828265070915, + 0.0022427444346249104, + 0.006356211379170418, + 0.02426452375948429, + 0.005694810301065445, + 0.005106043536216021, + 0.007633734028786421, + 0.0076129077933728695, + 0.009383536875247955, + 0.004803037270903587, + 0.00780530134215951, + 0.007167678326368332, + 0.006496370770037174, + 0.01111309602856636, + 0.0061475131660699844, + 0.005089900456368923, + 0.006498876493424177, + 0.002542270813137293, + 0.004730018321424723, + 0.00659826910123229, + 0.0037360989954322577, + 0.0029289578087627888, + 0.012375530786812305, + 0.013318966142833233, + 0.01914200186729431, + 0.007437653373926878, + 0.014276639558374882, + 0.01131937000900507, + 0.006991436704993248, + 0.013484107330441475, + 0.009368599392473698, + 0.011990547180175781, + 0.007100187707692385, + 0.007034365087747574, + 0.02315983548760414, + 0.009635662660002708, + 0.007709616329520941, + 0.010644094087183475, + 0.013331197202205658, + 0.013707916252315044, + 0.009857728146016598, + 0.007706086616963148, + 0.023631887510418892, + 0.012713940814137459, + 0.010254021733999252, + 0.01276461873203516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.038089267909526825 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_0/index_tensor([4108, 4108])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_0/index_tensor([4108, 4108])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b8ede131f1d5edcf3ca23a489b87434bcdf853ed --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_0/index_tensor([4108, 4108])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic shot, shot from an extremely low, upward angle, captures a vibrant young female character at the apex of a graceful leap. Her body unfurls in the center of the frame, limbs extended outward in a gesture so effortless that it seems to defy gravity. She wears a form-fitting, iridescent, silver-white, futuristic bodysuit, its smooth material reflecting the surrounding light. The backdrop is a vast, dark void, dotted with faint, distant stardust, accentuating the character's radiance. Her most striking feature is her long, completely translucent hair, which gently undulates and drifts in the weightless environment like luminous jellyfish tentacles. These strands emanate an intense, iridescent bioluminescence, shifting vibrantly between cyan, magenta, brilliant yellow, and electric blue. This light not only illuminates her face, brimming with pure joy and excitement, but also casts ever-changing colorful reflections on her silver-white garment. The entire image is imbued with a sense of movement and fantasy, creating an atmosphere of intense liveliness and vibrant energy.", + "seed": 23778, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_0/index_4108_seed_23778_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10749326336663216, + "vlm_logprob": -0.10749326336663216, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.10749326336663216, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4411674439907074 + ], + "opposite_logprob_tokens": [ + -1.0661674737930298 + ], + "logprob_reward": -0.4411674439907074, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4411674439907074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11583035439252853, + 0.05775104835629463, + 0.04767971113324165, + 0.13811276853084564, + 0.12396465241909027, + 0.3461310565471649, + 0.041624680161476135, + 0.041878145188093185, + 0.05318428948521614, + 0.033843282610177994 + ], + "credit_map": [ + 0.015976566821336746, + 0.018528161570429802, + 0.017241593450307846, + 0.014178408309817314, + 0.01078718714416027, + 0.008315039798617363, + 0.016693169251084328, + 0.014110223390161991, + 0.009074085392057896, + 0.011915531009435654, + 0.0074841477908194065, + 0.008453152142465115, + 0.0059454720467329025, + 0.0048405989073216915, + 0.004210818093270063, + 0.005827247630804777, + 0.005676915869116783, + 0.009592153131961823, + 0.00636371411383152, + 0.006520621012896299, + 0.004404445644468069, + 0.00452011451125145, + 0.004174153786152601, + 0.006427590269595385, + 0.01670372486114502, + 0.016417084261775017, + 0.019456705078482628, + 0.020283378660678864, + 0.013434047810733318, + 0.01762128248810768, + 0.0181251373142004, + 0.016071414574980736, + 0.013719758950173855, + 0.01508062519133091, + 0.017296424135565758, + 0.020913975313305855, + 0.017289740964770317, + 0.013420198112726212, + 0.011718811467289925, + 0.014525122009217739, + 0.02083355188369751, + 0.007391062099486589, + 0.12633387744426727, + 0.012674150988459587, + 0.007457511965185404, + 0.018460974097251892, + 0.010202491655945778, + 0.14277741312980652, + 0.003368576057255268, + 0.0035505471751093864, + 0.005911021959036589, + 0.005483068060129881, + 0.00623870175331831, + 0.004478445276618004, + 0.006801516283303499, + 0.005792803131043911, + 0.005101970862597227, + 0.004099768120795488, + 0.008868738077580929, + 0.002845424460247159, + 0.004473178181797266, + 0.005767189897596836, + 0.0038600419647991657, + 0.006861836649477482, + 0.0035262180026620626, + 0.009504199028015137, + 0.006156798452138901, + 0.00857100635766983, + 0.005715733859688044, + 0.005580877885222435, + 0.006663242820650339, + 0.007466214708983898, + 0.0036385152488946915, + 0.005487182177603245, + 0.005453660152852535, + 0.0036954774986952543, + 0.0035529681481420994, + 0.004877137020230293, + 0.002639141632243991, + 0.0044992039911448956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.4411674439907074 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331754006445408 + ], + "opposite_logprob_tokens": [ + -3.773317575454712 + ], + "logprob_reward": -0.02331754006445408, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331754006445408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0637141540646553, + 0.04269101843237877, + 0.04774654284119606, + 0.08128342032432556, + 0.04947257041931152, + 0.4257632791996002, + 0.03312474489212036, + 0.07343313843011856, + 0.06037598475813866, + 0.1223951205611229 + ], + "credit_map": [ + 0.009472537785768509, + 0.00919607374817133, + 0.010194788686931133, + 0.009053797461092472, + 0.005221404600888491, + 0.004426809027791023, + 0.008378016762435436, + 0.007770728785544634, + 0.008917725645005703, + 0.003746665082871914, + 0.006618732586503029, + 0.006238820496946573, + 0.00391432736068964, + 0.0030359558295458555, + 0.0036863205023109913, + 0.006532473023980856, + 0.005560562014579773, + 0.006178722251206636, + 0.007576825562864542, + 0.008317682892084122, + 0.005232080351561308, + 0.004046527203172445, + 0.004975890275090933, + 0.005858259275555611, + 0.00950460135936737, + 0.008229185827076435, + 0.011971128173172474, + 0.011342685669660568, + 0.009557656943798065, + 0.011532701551914215, + 0.01072594616562128, + 0.008419527672231197, + 0.00649495143443346, + 0.005240366328507662, + 0.007308498490601778, + 0.005682815797626972, + 0.004999998491257429, + 0.007611573673784733, + 0.005471865646541119, + 0.006662505213171244, + 0.02542203851044178, + 0.006624337751418352, + 0.15663611888885498, + 0.015421099029481411, + 0.009771234355866909, + 0.023453500121831894, + 0.013525390066206455, + 0.17490960657596588, + 0.0025563500821590424, + 0.0029970284085720778, + 0.004414419177919626, + 0.0036683257203549147, + 0.004230877384543419, + 0.003098951419815421, + 0.007001347839832306, + 0.005157448351383209, + 0.0070032477378845215, + 0.006334330420941114, + 0.01145397312939167, + 0.0064241597428917885, + 0.006359986495226622, + 0.018624447286128998, + 0.005686915013939142, + 0.011546093970537186, + 0.0033318018540740013, + 0.007448410615324974, + 0.004977970849722624, + 0.008827733807265759, + 0.009581646881997585, + 0.008228149265050888, + 0.010005727410316467, + 0.007974552921950817, + 0.01401443686336279, + 0.013032804243266582, + 0.014241654425859451, + 0.01282732468098402, + 0.02002381905913353, + 0.015743985772132874, + 0.014842952601611614, + 0.017668159678578377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 3 + ], + "selected_reward": -0.02331754006445408 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608455553650856 + ], + "opposite_logprob_tokens": [ + -4.141084671020508 + ], + "logprob_reward": -0.01608455553650856, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608455553650856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13774628937244415, + 0.10272341966629028, + 0.07325847446918488, + 0.07408615946769714, + 0.1796555370092392, + 0.25216957926750183, + 0.038835447281599045, + 0.06006985530257225, + 0.046962346881628036, + 0.03449287638068199 + ], + "credit_map": [ + 0.02046171762049198, + 0.02200193889439106, + 0.023118436336517334, + 0.02284238673746586, + 0.012240913696587086, + 0.008577441796660423, + 0.014189047738909721, + 0.01431440468877554, + 0.014075538143515587, + 0.013420114293694496, + 0.015558090060949326, + 0.021301070228219032, + 0.01273074559867382, + 0.0055508119985461235, + 0.011329511180520058, + 0.008757533505558968, + 0.007675521541386843, + 0.011487734504044056, + 0.011357687413692474, + 0.011073228903114796, + 0.008877132087945938, + 0.007248719222843647, + 0.006493165623396635, + 0.009045284241437912, + 0.008526462130248547, + 0.008061513304710388, + 0.010395699180662632, + 0.014874988235533237, + 0.006955714430660009, + 0.010742145590484142, + 0.008313819766044617, + 0.006215814035385847, + 0.02066769264638424, + 0.024974891915917397, + 0.025581397116184235, + 0.02379339002072811, + 0.021894600242376328, + 0.027287669479846954, + 0.016743239015340805, + 0.018712658435106277, + 0.01729266718029976, + 0.02164790779352188, + 0.06994400918483734, + 0.022899813950061798, + 0.019834132865071297, + 0.014101087115705013, + 0.014048242010176182, + 0.07240171730518341, + 0.004918485414236784, + 0.004298320040106773, + 0.005703614093363285, + 0.0037146550603210926, + 0.006283842492848635, + 0.0033364281989634037, + 0.004903288092464209, + 0.005676815751940012, + 0.005630880128592253, + 0.0052569289691746235, + 0.00869771558791399, + 0.005375293083488941, + 0.006493404507637024, + 0.012213573791086674, + 0.006711611058562994, + 0.009690449573099613, + 0.004137001931667328, + 0.007766078691929579, + 0.0046307481825351715, + 0.006164838559925556, + 0.005001936573535204, + 0.006241233553737402, + 0.006102059502154589, + 0.00691844942048192, + 0.0032333771232515574, + 0.003866248531267047, + 0.00344663392752409, + 0.003070433158427477, + 0.0045461952686309814, + 0.005355150438845158, + 0.004249291960150003, + 0.006725546903908253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.01608455553650856 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019480077316984534 + ], + "opposite_logprob_tokens": [ + -6.25194787979126 + ], + "logprob_reward": -0.0019480077316984534, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019480077316984534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1204967200756073, + 0.1517070233821869, + 0.051773812621831894, + 0.10049202293157578, + 0.14237931370735168, + 0.18712350726127625, + 0.07972754538059235, + 0.05600437894463539, + 0.05620725452899933, + 0.05408836901187897 + ], + "credit_map": [ + 0.016350936144590378, + 0.0180131234228611, + 0.020085353404283524, + 0.01649557799100876, + 0.011105956509709358, + 0.008816626854240894, + 0.013483553193509579, + 0.01614559441804886, + 0.025604693219065666, + 0.016051428392529488, + 0.026800421997904778, + 0.03522925078868866, + 0.010080724954605103, + 0.01070344727486372, + 0.015121287666261196, + 0.012115772813558578, + 0.0063600619323551655, + 0.007201859261840582, + 0.007512264885008335, + 0.006547292694449425, + 0.008057712577283382, + 0.004737016744911671, + 0.004947139881551266, + 0.006410466972738504, + 0.012760506942868233, + 0.010917626321315765, + 0.016513993963599205, + 0.01588387042284012, + 0.011007610708475113, + 0.01466321386396885, + 0.009561920538544655, + 0.009183277375996113, + 0.011728701181709766, + 0.011889995075762272, + 0.020798061043024063, + 0.012800686992704868, + 0.018462205305695534, + 0.01715034246444702, + 0.022348226979374886, + 0.027201103046536446, + 0.012624159455299377, + 0.008706982247531414, + 0.057294026017189026, + 0.010602949187159538, + 0.011813578195869923, + 0.01104616466909647, + 0.00984313152730465, + 0.06519252061843872, + 0.011385616846382618, + 0.005844892933964729, + 0.012451082468032837, + 0.00920090638101101, + 0.008729873225092888, + 0.00940658524632454, + 0.008534970693290234, + 0.014173619449138641, + 0.005084923934191465, + 0.0059453765861690044, + 0.00847846083343029, + 0.00575288524851203, + 0.0067369225434958935, + 0.007792379707098007, + 0.007385989185422659, + 0.008827444165945053, + 0.003427001880481839, + 0.00709600979462266, + 0.007025131955742836, + 0.009449461475014687, + 0.004952423740178347, + 0.005856023635715246, + 0.01067600678652525, + 0.007725196424871683, + 0.008390579372644424, + 0.006637477781623602, + 0.009712760336697102, + 0.0035258838906884193, + 0.004000951070338488, + 0.008116490207612514, + 0.005115382838994265, + 0.008588841184973717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.0019480077316984534 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494876950979233 + ], + "opposite_logprob_tokens": [ + -2.9299488067626953 + ], + "logprob_reward": -0.05494876950979233, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494876950979233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09718047082424164, + 0.44526803493499756, + 0.03181130439043045, + 0.0565636120736599, + 0.04919642582535744, + 0.13758978247642517, + 0.01934393309056759, + 0.03756612166762352, + 0.0793001726269722, + 0.04618019238114357 + ], + "credit_map": [ + 0.016488604247570038, + 0.01331926230341196, + 0.016304735094308853, + 0.014518971554934978, + 0.008887522853910923, + 0.006953581236302853, + 0.007160416804254055, + 0.013547373935580254, + 0.07511018961668015, + 0.06826657056808472, + 0.06650149077177048, + 0.08165723830461502, + 0.03762509301304817, + 0.03506370633840561, + 0.03777956962585449, + 0.04326419159770012, + 0.005757035221904516, + 0.003894147928804159, + 0.005231816787272692, + 0.00596219627186656, + 0.0033119595609605312, + 0.0026802902575582266, + 0.0022877498995512724, + 0.0026861082296818495, + 0.007760345935821533, + 0.006631968542933464, + 0.00759625481441617, + 0.006668558344244957, + 0.009124843403697014, + 0.008199439384043217, + 0.006057292222976685, + 0.004524907562881708, + 0.0043734111823141575, + 0.006741757970303297, + 0.007160452660173178, + 0.009281758219003677, + 0.006432981695979834, + 0.0058805798180401325, + 0.004629787523299456, + 0.004695693962275982, + 0.007764769252389669, + 0.006842893082648516, + 0.04093989357352257, + 0.00878441147506237, + 0.00802088063210249, + 0.010134360752999783, + 0.007437772583216429, + 0.04766479879617691, + 0.0020784505177289248, + 0.0018364118877798319, + 0.0030917529948055744, + 0.002692244481295347, + 0.0022308214101940393, + 0.0017656735144555569, + 0.002777357818558812, + 0.002871220698580146, + 0.003766447538509965, + 0.0037438166327774525, + 0.007184124551713467, + 0.003983688540756702, + 0.004291099961847067, + 0.004287710878998041, + 0.004014899022877216, + 0.006294331979006529, + 0.005005502142012119, + 0.013894593343138695, + 0.007466576993465424, + 0.014266105368733406, + 0.009273297153413296, + 0.006960636470466852, + 0.012406610883772373, + 0.01002684235572815, + 0.004839225206524134, + 0.006648125126957893, + 0.008367572911083698, + 0.0045256358571350574, + 0.005047073122113943, + 0.006467150524258614, + 0.004109417088329792, + 0.0061759911477565765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.05494876950979233 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic shot, shot from an extremely low, upward angle, captures a vibrant young female character at the apex of a graceful leap. Her body unfurls in the center of the frame, limbs extended outward in a gesture so effortless that it seems to defy gravity. She wears a form-fitting, iridescent, silver-white, futuristic bodysuit, its smooth material reflecting the surrounding light. The backdrop is a vast, dark void, dotted with faint, distant stardust, accentuating the character's radiance. Her most striking feature is her long, completely translucent hair, which gently undulates and drifts in the weightless environment like luminous jellyfish tentacles. These strands emanate an intense, iridescent bioluminescence, shifting vibrantly between cyan, magenta, brilliant yellow, and electric blue. This light not only illuminates her face, brimming with pure joy and excitement, but also casts ever-changing colorful reflections on her silver-white garment. The entire image is imbued with a sense of movement and fantasy, creating an atmosphere of intense liveliness and vibrant energy.", + "seed": 23779, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_0/index_4108_seed_23779_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12853642785921693, + "vlm_logprob": -0.12853642785921693, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.12853642785921693, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.36193907260894775 + ], + "opposite_logprob_tokens": [ + -1.2369390726089478 + ], + "logprob_reward": -0.36193907260894775, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.36193907260894775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33071181178092957, + 0.2148192673921585, + 0.0444517619907856, + 0.09006771445274353, + 0.04762829467654228, + 0.0696425661444664, + 0.035361845046281815, + 0.03841330111026764, + 0.07954667508602142, + 0.049356747418642044 + ], + "credit_map": [ + 0.03332148492336273, + 0.06320922076702118, + 0.03887145221233368, + 0.044793564826250076, + 0.030630754306912422, + 0.05417296290397644, + 0.03460860997438431, + 0.03110378608107567, + 0.018491530790925026, + 0.008073996752500534, + 0.013500561937689781, + 0.030040210112929344, + 0.015681594610214233, + 0.011580439284443855, + 0.07220510393381119, + 0.045245833694934845, + 0.0038988918531686068, + 0.005469475407153368, + 0.00540135009214282, + 0.006090388633310795, + 0.005137110594660044, + 0.006040948908776045, + 0.005862933583557606, + 0.00655066454783082, + 0.009105928242206573, + 0.010663359425961971, + 0.012432767078280449, + 0.014103303663432598, + 0.009263811632990837, + 0.010202554054558277, + 0.014871636405587196, + 0.00942435022443533, + 0.004409461747854948, + 0.006807529833167791, + 0.0056551918387413025, + 0.004369379486888647, + 0.007214786950498819, + 0.009239907376468182, + 0.004876251798123121, + 0.005055786110460758, + 0.008310705423355103, + 0.012674943543970585, + 0.010942571796476841, + 0.008610461838543415, + 0.008373974822461605, + 0.0062099602073431015, + 0.006551120895892382, + 0.007968829944729805, + 0.005590961314737797, + 0.003990833181887865, + 0.003759793471544981, + 0.0050892517901957035, + 0.0036678931210190058, + 0.00502587528899312, + 0.0040870895609259605, + 0.004150145687162876, + 0.0052208732813596725, + 0.004024786874651909, + 0.00705736456438899, + 0.0037572330329567194, + 0.004840684123337269, + 0.0038726076018065214, + 0.005005333106964827, + 0.004634418524801731, + 0.00975881703197956, + 0.008386819623410702, + 0.00701173534616828, + 0.011821294203400612, + 0.006340248975902796, + 0.013939080759882927, + 0.011911987327039242, + 0.01037669274955988, + 0.006092457100749016, + 0.009143882431089878, + 0.003202383406460285, + 0.004762861877679825, + 0.004555270541459322, + 0.006035665050148964, + 0.005934837274253368, + 0.009629388339817524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.36193907260894775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026384975761175156 + ], + "opposite_logprob_tokens": [ + -3.6513850688934326 + ], + "logprob_reward": -0.026384975761175156, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026384975761175156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2792816162109375, + 0.3406393826007843, + 0.07159416377544403, + 0.09038299322128296, + 0.07090982794761658, + 0.02881074883043766, + 0.033384259790182114, + 0.032495807856321335, + 0.02225494757294655, + 0.030246185138821602 + ], + "credit_map": [ + 0.027678580954670906, + 0.05380239337682724, + 0.03154200315475464, + 0.03369633108377457, + 0.02220243401825428, + 0.04014704376459122, + 0.04803526774048805, + 0.022177575156092644, + 0.02717941626906395, + 0.015079141594469547, + 0.022077836096286774, + 0.05831664055585861, + 0.02467404492199421, + 0.015796981751918793, + 0.1128745973110199, + 0.06464071571826935, + 0.005444742273539305, + 0.013832994736731052, + 0.006774929817765951, + 0.00942632369697094, + 0.009274137206375599, + 0.008158446289598942, + 0.009873620234429836, + 0.008808974176645279, + 0.009308446198701859, + 0.012378565967082977, + 0.01238896232098341, + 0.01420200988650322, + 0.009395704604685307, + 0.006821780931204557, + 0.017311496660113335, + 0.008576026186347008, + 0.005558013450354338, + 0.015630653128027916, + 0.007020098622888327, + 0.011488526128232479, + 0.009137260727584362, + 0.008547721430659294, + 0.007561623118817806, + 0.005965930409729481, + 0.0029995576478540897, + 0.0047489614225924015, + 0.0032187332399189472, + 0.003969175741076469, + 0.0032385620288550854, + 0.003999583888798952, + 0.002994653768837452, + 0.0036415208596736193, + 0.004246893804520369, + 0.0038143876008689404, + 0.0038757980801165104, + 0.003494878765195608, + 0.004895112477242947, + 0.0051098112016916275, + 0.002753856126219034, + 0.005193519871681929, + 0.0051467218436300755, + 0.0028798480052500963, + 0.006818743888288736, + 0.0031496118754148483, + 0.002581852488219738, + 0.003007100895047188, + 0.004558986518532038, + 0.004352943040430546, + 0.0034500425681471825, + 0.0032819805201143026, + 0.001950963051058352, + 0.002139653777703643, + 0.0021228347904980183, + 0.0027363612316548824, + 0.0034891183022409678, + 0.003083992749452591, + 0.0035561700351536274, + 0.004691314417868853, + 0.002930490067228675, + 0.0030331031884998083, + 0.0030709875281900167, + 0.003489487338811159, + 0.006356411147862673, + 0.0031182209495455027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.026384975761175156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048682939261198044 + ], + "opposite_logprob_tokens": [ + -3.048682928085327 + ], + "logprob_reward": -0.048682939261198044, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048682939261198044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1316305249929428, + 0.3223119080066681, + 0.05090554803609848, + 0.21737869083881378, + 0.05488428846001625, + 0.0474848747253418, + 0.08305488526821136, + 0.03673671931028366, + 0.02986891008913517, + 0.02574361301958561 + ], + "credit_map": [ + 0.013531270436942577, + 0.02349085919559002, + 0.012628212571144104, + 0.023812096565961838, + 0.011302003636956215, + 0.019605383276939392, + 0.016494980081915855, + 0.010765710845589638, + 0.028786657378077507, + 0.008429132401943207, + 0.02021690085530281, + 0.04175403714179993, + 0.02565581351518631, + 0.01628435216844082, + 0.10787657648324966, + 0.07330847531557083, + 0.004479760769754648, + 0.005733035039156675, + 0.005247349850833416, + 0.007106955628842115, + 0.006381694693118334, + 0.008008288219571114, + 0.007396848872303963, + 0.006551617290824652, + 0.020971084013581276, + 0.030475471168756485, + 0.028544580563902855, + 0.027611998841166496, + 0.03454554080963135, + 0.022834537550807, + 0.03061845153570175, + 0.021777015179395676, + 0.006268768571317196, + 0.007369756232947111, + 0.008321925066411495, + 0.0069787222892045975, + 0.007848097942769527, + 0.0075243739411234856, + 0.005121362395584583, + 0.005451281554996967, + 0.0050912522710859776, + 0.00789431668817997, + 0.004610396921634674, + 0.006173249334096909, + 0.005435274448245764, + 0.006631461903452873, + 0.005648041144013405, + 0.006000879220664501, + 0.009003322571516037, + 0.00872777123004198, + 0.009900463744997978, + 0.010882483795285225, + 0.011366254650056362, + 0.012585327960550785, + 0.008996019139885902, + 0.011593240313231945, + 0.004471698775887489, + 0.004245514050126076, + 0.005478712730109692, + 0.005229255184531212, + 0.00361946364864707, + 0.003321888158097863, + 0.005792755633592606, + 0.004577431827783585, + 0.0040786596946418285, + 0.004171054810285568, + 0.0026976943481713533, + 0.0032372851856052876, + 0.0028447865042835474, + 0.004919265862554312, + 0.003603606251999736, + 0.0043165553361177444, + 0.0031707617454230785, + 0.006083587650209665, + 0.0031240792013704777, + 0.0024461152497678995, + 0.0026317851152271032, + 0.0025786850601434708, + 0.0031067810487002134, + 0.0026018191128969193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.048682939261198044 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410511763766408 + ], + "opposite_logprob_tokens": [ + -5.504105091094971 + ], + "logprob_reward": -0.00410511763766408, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410511763766408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07561588287353516, + 0.15106140077114105, + 0.03348443657159805, + 0.44853103160858154, + 0.029581943526864052, + 0.09048615396022797, + 0.06793266534805298, + 0.0504215843975544, + 0.024001125246286392, + 0.028883717954158783 + ], + "credit_map": [ + 0.0068496414460241795, + 0.008592162281274796, + 0.009828399866819382, + 0.012693280354142189, + 0.007085294928401709, + 0.011207965202629566, + 0.010365162044763565, + 0.008993970230221748, + 0.011860236525535583, + 0.005646673496812582, + 0.009713174775242805, + 0.0261389072984457, + 0.010151127353310585, + 0.009299129247665405, + 0.05194304138422012, + 0.026309119537472725, + 0.003146169241517782, + 0.003247553249821067, + 0.004060675390064716, + 0.004213678650557995, + 0.003606863785535097, + 0.003304922953248024, + 0.004705764818936586, + 0.007198809180408716, + 0.04312220960855484, + 0.05171355605125427, + 0.06793447583913803, + 0.07197634130716324, + 0.04664471372961998, + 0.04497281089425087, + 0.08397147059440613, + 0.03819545730948448, + 0.002989819273352623, + 0.004583476576954126, + 0.0038805988151580095, + 0.0035528424195945263, + 0.0031482644844800234, + 0.003938383422791958, + 0.004519638139754534, + 0.0029689196962863207, + 0.008723612874746323, + 0.020174169912934303, + 0.008691915310919285, + 0.009190254844725132, + 0.012979450635612011, + 0.010221371427178383, + 0.008876007981598377, + 0.011629367247223854, + 0.005284245125949383, + 0.010367557406425476, + 0.010254640132188797, + 0.012997288256883621, + 0.007483799941837788, + 0.007444295100867748, + 0.006541101727634668, + 0.007559735327959061, + 0.00555404694750905, + 0.0058752126060426235, + 0.006249237339943647, + 0.008957944810390472, + 0.005328249651938677, + 0.004780918825417757, + 0.006462037097662687, + 0.007213936652988195, + 0.0035176374949514866, + 0.0026000412181019783, + 0.002429298823699355, + 0.002226970624178648, + 0.0020753636490553617, + 0.004125817678868771, + 0.003423161804676056, + 0.003602832555770874, + 0.003616379341110587, + 0.005584254860877991, + 0.0026070373132824898, + 0.002602851018309593, + 0.004120760131627321, + 0.0029767341911792755, + 0.0037987753748893738, + 0.003576925490051508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.00410511763766408 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015700340270996 + ], + "opposite_logprob_tokens": [ + -1.7015700340270996 + ], + "logprob_reward": -0.2015700340270996, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015700340270996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.398962140083313, + 0.2377278208732605, + 0.021130237728357315, + 0.13542160391807556, + 0.014645328745245934, + 0.05766148865222931, + 0.047330230474472046, + 0.01843762956559658, + 0.04834059253334999, + 0.020342955365777016 + ], + "credit_map": [ + 0.04669560119509697, + 0.04689633101224899, + 0.05604520067572594, + 0.05019465461373329, + 0.0371411107480526, + 0.05109274387359619, + 0.05856625735759735, + 0.05233023315668106, + 0.020528588443994522, + 0.006208007223904133, + 0.015078000724315643, + 0.03567339852452278, + 0.016493534669280052, + 0.010298623703420162, + 0.08369839936494827, + 0.04974927008152008, + 0.0024494556710124016, + 0.002854224992915988, + 0.0023109486792236567, + 0.003246848238632083, + 0.003040990326553583, + 0.001978997141122818, + 0.0028899505268782377, + 0.0023588216863572598, + 0.014261458069086075, + 0.01910915970802307, + 0.020728925243020058, + 0.022863907739520073, + 0.012216866947710514, + 0.010514133609831333, + 0.021825743839144707, + 0.013901417143642902, + 0.0016175600467249751, + 0.0029149188194423914, + 0.00178891874384135, + 0.0019163752440363169, + 0.0017959907418116927, + 0.0019017565064132214, + 0.001155260019004345, + 0.0015545475762337446, + 0.005445587914437056, + 0.008763831108808517, + 0.006487001199275255, + 0.006917377933859825, + 0.00816676951944828, + 0.008762099780142307, + 0.006520992610603571, + 0.006597830913960934, + 0.005269048269838095, + 0.007311652880162001, + 0.00589506421238184, + 0.005710446275770664, + 0.005769873503595591, + 0.005761657375842333, + 0.005652773659676313, + 0.005959715228527784, + 0.0027679570484906435, + 0.0017506532603874803, + 0.0026374903973191977, + 0.002304858062416315, + 0.0021792559418827295, + 0.001800845842808485, + 0.0025034756399691105, + 0.0024930951185524464, + 0.0075170001946389675, + 0.005435859318822622, + 0.004629913717508316, + 0.004560868721455336, + 0.0038509350270032883, + 0.0067116436548531055, + 0.008320054039359093, + 0.007314317859709263, + 0.0031893178820610046, + 0.002526360098272562, + 0.0022145549301058054, + 0.0016424967907369137, + 0.002084293868392706, + 0.003601094940677285, + 0.002526538912206888, + 0.002558297012001276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.2015700340270996 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic shot, shot from an extremely low, upward angle, captures a vibrant young female character at the apex of a graceful leap. Her body unfurls in the center of the frame, limbs extended outward in a gesture so effortless that it seems to defy gravity. She wears a form-fitting, iridescent, silver-white, futuristic bodysuit, its smooth material reflecting the surrounding light. The backdrop is a vast, dark void, dotted with faint, distant stardust, accentuating the character's radiance. Her most striking feature is her long, completely translucent hair, which gently undulates and drifts in the weightless environment like luminous jellyfish tentacles. These strands emanate an intense, iridescent bioluminescence, shifting vibrantly between cyan, magenta, brilliant yellow, and electric blue. This light not only illuminates her face, brimming with pure joy and excitement, but also casts ever-changing colorful reflections on her silver-white garment. The entire image is imbued with a sense of movement and fantasy, creating an atmosphere of intense liveliness and vibrant energy.", + "seed": 23780, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_0/index_4108_seed_23780_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1431198138743639, + "vlm_logprob": -0.1431198138743639, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.1431198138743639, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.589599072933197 + ], + "opposite_logprob_tokens": [ + -0.839599072933197 + ], + "logprob_reward": -0.589599072933197, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.589599072933197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1880534291267395, + 0.08519550412893295, + 0.06099178269505501, + 0.056084439158439636, + 0.2823188900947571, + 0.07112529128789902, + 0.051165562123060226, + 0.05596592277288437, + 0.07733017206192017, + 0.07176903635263443 + ], + "credit_map": [ + 0.01330435648560524, + 0.016636215150356293, + 0.01876099593937397, + 0.013027979992330074, + 0.010656063444912434, + 0.05475206300616264, + 0.013975625857710838, + 0.04694012925028801, + 0.006049297749996185, + 0.007536777760833502, + 0.014065430499613285, + 0.016243383288383484, + 0.005605159793049097, + 0.015113861300051212, + 0.007418745197355747, + 0.013162849470973015, + 0.00516847986727953, + 0.008213535882532597, + 0.011048090644180775, + 0.007993275299668312, + 0.005568983498960733, + 0.009960995987057686, + 0.003739729057997465, + 0.009298691526055336, + 0.004399619065225124, + 0.0042492784559726715, + 0.009663390927016735, + 0.006493619177490473, + 0.004200590308755636, + 0.010127127170562744, + 0.006342235952615738, + 0.01060857716947794, + 0.032798390835523605, + 0.023987963795661926, + 0.0673004761338234, + 0.037109579890966415, + 0.015982162207365036, + 0.026275504380464554, + 0.016715385019779205, + 0.06214941665530205, + 0.011658232659101486, + 0.008520405739545822, + 0.007345592137426138, + 0.005652216728776693, + 0.008228546939790249, + 0.0063719251193106174, + 0.007282622158527374, + 0.016065750271081924, + 0.007126423995941877, + 0.00806827750056982, + 0.010447274893522263, + 0.006551241967827082, + 0.004507327917963266, + 0.003744088113307953, + 0.0049694981426000595, + 0.005751431453973055, + 0.005811810959130526, + 0.010247325524687767, + 0.005694753956049681, + 0.00630323588848114, + 0.006847556680440903, + 0.006353459320962429, + 0.007332858629524708, + 0.007374922279268503, + 0.0076335109770298, + 0.009535354562103748, + 0.012140442617237568, + 0.009766650386154652, + 0.013003095053136349, + 0.0070438748225569725, + 0.009431453421711922, + 0.008775800466537476, + 0.00733867147937417, + 0.015168487094342709, + 0.011942107230424881, + 0.012835005298256874, + 0.007024622056633234, + 0.0045666820369660854, + 0.005620065610855818, + 0.007273397874087095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.589599072933197 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02639867551624775 + ], + "opposite_logprob_tokens": [ + -3.6513986587524414 + ], + "logprob_reward": -0.02639867551624775, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02639867551624775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3705718517303467, + 0.13493935763835907, + 0.036478593945503235, + 0.02417367324233055, + 0.04582240805029869, + 0.045138515532016754, + 0.04357276111841202, + 0.04540622979402542, + 0.1592300832271576, + 0.0946664959192276 + ], + "credit_map": [ + 0.01352948509156704, + 0.013805807568132877, + 0.05190441384911537, + 0.009153086692094803, + 0.007827994413673878, + 0.15484324097633362, + 0.012698227539658546, + 0.10680961608886719, + 0.006648635026067495, + 0.01074036955833435, + 0.03584953397512436, + 0.01655479520559311, + 0.007561971433460712, + 0.014801060780882835, + 0.007842181250452995, + 0.03494080901145935, + 0.0039796470664441586, + 0.005306856706738472, + 0.005270783323794603, + 0.0033565200865268707, + 0.0038072268944233656, + 0.004538919311016798, + 0.0033206406515091658, + 0.006898002233356237, + 0.0017053298652172089, + 0.0019475166918709874, + 0.004149258602410555, + 0.0030302798841148615, + 0.002263795817270875, + 0.004204807337373495, + 0.0018379660323262215, + 0.005034719593822956, + 0.005832709837704897, + 0.00267034862190485, + 0.007806393783539534, + 0.006842797622084618, + 0.002358451019972563, + 0.007148995529860258, + 0.002789407968521118, + 0.010373305529356003, + 0.00807871948927641, + 0.006651660893112421, + 0.005168190225958824, + 0.0035916403867304325, + 0.005334699060767889, + 0.004992613568902016, + 0.004451218526810408, + 0.006869771517813206, + 0.004108514171093702, + 0.010063993744552135, + 0.006698735058307648, + 0.005157290492206812, + 0.004523637238889933, + 0.0037033988628536463, + 0.005135850980877876, + 0.004181341268122196, + 0.005660341586917639, + 0.01122280303388834, + 0.004640568047761917, + 0.004917406011372805, + 0.0039048902690410614, + 0.004677601624280214, + 0.004943108186125755, + 0.005439509637653828, + 0.014243138954043388, + 0.030878672376275063, + 0.02315990999341011, + 0.029788615182042122, + 0.01911807991564274, + 0.013638646341860294, + 0.013104867190122604, + 0.015298150479793549, + 0.011784352362155914, + 0.01868908666074276, + 0.018223106861114502, + 0.010035242885351181, + 0.010420804843306541, + 0.007773575372993946, + 0.008077815175056458, + 0.009662517346441746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.02639867551624775 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026375919580459595 + ], + "opposite_logprob_tokens": [ + -3.6513760089874268 + ], + "logprob_reward": -0.026375919580459595, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026375919580459595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37990301847457886, + 0.16171002388000488, + 0.03384462371468544, + 0.02938280813395977, + 0.07918116450309753, + 0.030818162485957146, + 0.04732467234134674, + 0.0936642587184906, + 0.09294135868549347, + 0.051229946315288544 + ], + "credit_map": [ + 0.019445940852165222, + 0.016883302479982376, + 0.06239648163318634, + 0.01582551933825016, + 0.00997872930020094, + 0.14555904269218445, + 0.015440602786839008, + 0.09437335282564163, + 0.010206523351371288, + 0.011819485574960709, + 0.0519523099064827, + 0.008819593116641045, + 0.00801414716988802, + 0.03897988796234131, + 0.007284360006451607, + 0.024633701890707016, + 0.003122499445453286, + 0.0037712908815592527, + 0.00559835322201252, + 0.006041018292307854, + 0.003118615597486496, + 0.0038662245497107506, + 0.003134939819574356, + 0.005191676784306765, + 0.003424248658120632, + 0.003290105378255248, + 0.004988911096006632, + 0.003994031343609095, + 0.003165273927152157, + 0.0032831081189215183, + 0.0030853303615003824, + 0.004151798784732819, + 0.009868348948657513, + 0.0057447138242423534, + 0.015549774281680584, + 0.006687105633318424, + 0.004165293648838997, + 0.009449483826756477, + 0.005340449046343565, + 0.022375984117388725, + 0.0034928470849990845, + 0.003045678837224841, + 0.0038087789434939623, + 0.0028943743091076612, + 0.0028868266381323338, + 0.00423416355624795, + 0.0033169768285006285, + 0.007138512562960386, + 0.004856420215219259, + 0.011599819175899029, + 0.00748970964923501, + 0.005969186779111624, + 0.004966232925653458, + 0.003984702751040459, + 0.004630866926163435, + 0.0038277299609035254, + 0.013010572642087936, + 0.016796616837382317, + 0.010954657569527626, + 0.010210116393864155, + 0.010384519584476948, + 0.012209939770400524, + 0.009167618118226528, + 0.010930213145911694, + 0.012397080659866333, + 0.01264795009046793, + 0.013570686802268028, + 0.012729890644550323, + 0.012793930247426033, + 0.006295402999967337, + 0.011044814251363277, + 0.011461596935987473, + 0.005058394744992256, + 0.013341883197426796, + 0.009358922019600868, + 0.005149442702531815, + 0.004436562769114971, + 0.004428253509104252, + 0.002876936923712492, + 0.0065795485861599445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.026375919580459595 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032048076391220093 + ], + "opposite_logprob_tokens": [ + -5.753204822540283 + ], + "logprob_reward": -0.0032048076391220093, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032048076391220093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20821350812911987, + 0.12226938456296921, + 0.04747043922543526, + 0.07307767122983932, + 0.06302709877490997, + 0.09049935638904572, + 0.0965251624584198, + 0.09784496575593948, + 0.0892070010304451, + 0.11186537146568298 + ], + "credit_map": [ + 0.009841571561992168, + 0.011078674346208572, + 0.024465110152959824, + 0.007910510525107384, + 0.008604842238128185, + 0.076463483273983, + 0.010994138196110725, + 0.05885517597198486, + 0.006802902091294527, + 0.007421804126352072, + 0.03572472184896469, + 0.0056451307609677315, + 0.004666644148528576, + 0.014104552567005157, + 0.0037096692249178886, + 0.044193968176841736, + 0.005093524698168039, + 0.008100937120616436, + 0.007248414680361748, + 0.00499024149030447, + 0.005484664347022772, + 0.005902286618947983, + 0.004351744428277016, + 0.006298630032688379, + 0.004579054657369852, + 0.004457872360944748, + 0.01327571552246809, + 0.008815376088023186, + 0.00548454187810421, + 0.012152970768511295, + 0.006323395296931267, + 0.017988745123147964, + 0.007062443997710943, + 0.010047714225947857, + 0.01182286161929369, + 0.006783253513276577, + 0.004768076818436384, + 0.005734824109822512, + 0.0036872844211757183, + 0.01312064379453659, + 0.007200405467301607, + 0.02388477884232998, + 0.008576767519116402, + 0.006937993690371513, + 0.01511030551046133, + 0.006713585928082466, + 0.0069141751155257225, + 0.015161349438130856, + 0.007481666747480631, + 0.027879033237695694, + 0.010454107075929642, + 0.012265767902135849, + 0.010012540966272354, + 0.0076203420758247375, + 0.011191803961992264, + 0.00961990188807249, + 0.010748033411800861, + 0.02537076734006405, + 0.008138609118759632, + 0.010619455948472023, + 0.008996574208140373, + 0.010450811125338078, + 0.010801203548908234, + 0.012719512917101383, + 0.013667022809386253, + 0.009837743826210499, + 0.018497655168175697, + 0.012161514721810818, + 0.008643677458167076, + 0.007292395923286676, + 0.0067985933274030685, + 0.012308402918279171, + 0.012753616087138653, + 0.023525934666395187, + 0.016155077144503593, + 0.015561387874186039, + 0.015380592085421085, + 0.008641717955470085, + 0.009465651586651802, + 0.01038140058517456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0032048076391220093 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002059370279312 + ], + "opposite_logprob_tokens": [ + -2.6950206756591797 + ], + "logprob_reward": -0.07002059370279312, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002059370279312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7358121871948242, + 0.05508730560541153, + 0.03866381198167801, + 0.008782130666077137, + 0.04542114958167076, + 0.019558241590857506, + 0.011740687303245068, + 0.03551708534359932, + 0.02874700538814068, + 0.020670466125011444 + ], + "credit_map": [ + 0.028974508866667747, + 0.019588926807045937, + 0.09718229621648788, + 0.013490206561982632, + 0.017256079241633415, + 0.3103910982608795, + 0.021682657301425934, + 0.22724640369415283, + 0.0024964988697320223, + 0.003845500759780407, + 0.016348032280802727, + 0.005800968501716852, + 0.0028946821112185717, + 0.005641160067170858, + 0.0026767603121697903, + 0.015383699908852577, + 0.0033848618622869253, + 0.004956027492880821, + 0.00735731516033411, + 0.004638461861759424, + 0.0034877138677984476, + 0.0058883377350866795, + 0.002516302512958646, + 0.006434789393097162, + 0.0007794866687618196, + 0.0008385773980990052, + 0.001418480183929205, + 0.0009979570750147104, + 0.0008082616259343922, + 0.0017793017905205488, + 0.0007488197879865766, + 0.0014112445060163736, + 0.005703245755285025, + 0.004040144849568605, + 0.009002242237329483, + 0.003964956849813461, + 0.0026969353202730417, + 0.004943891428411007, + 0.0030902535654604435, + 0.011979475617408752, + 0.002026644069701433, + 0.003921299707144499, + 0.0024512961972504854, + 0.001522372360341251, + 0.0029174445662647486, + 0.0030438273679465055, + 0.001976468600332737, + 0.0016988877905532718, + 0.0011847177520394325, + 0.002366023138165474, + 0.001825807266868651, + 0.0011470947647467256, + 0.001247153733856976, + 0.001142717432230711, + 0.001654237974435091, + 0.0011729345424100757, + 0.003171126125380397, + 0.007672522682696581, + 0.0031322462018579245, + 0.003854162525385618, + 0.005660578142851591, + 0.003532008733600378, + 0.004821269307285547, + 0.0036731683649122715, + 0.0027131664101034403, + 0.0040603457018733025, + 0.0034369882196187973, + 0.0038524516858160496, + 0.005009809508919716, + 0.002839732449501753, + 0.003488908987492323, + 0.0033456000965088606, + 0.0029894413892179728, + 0.0026135328225791454, + 0.002997414208948612, + 0.0022262230049818754, + 0.0028227251023054123, + 0.0022286945022642612, + 0.002001398941501975, + 0.0027910342905670404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.07002059370279312 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic shot, shot from an extremely low, upward angle, captures a vibrant young female character at the apex of a graceful leap. Her body unfurls in the center of the frame, limbs extended outward in a gesture so effortless that it seems to defy gravity. She wears a form-fitting, iridescent, silver-white, futuristic bodysuit, its smooth material reflecting the surrounding light. The backdrop is a vast, dark void, dotted with faint, distant stardust, accentuating the character's radiance. Her most striking feature is her long, completely translucent hair, which gently undulates and drifts in the weightless environment like luminous jellyfish tentacles. These strands emanate an intense, iridescent bioluminescence, shifting vibrantly between cyan, magenta, brilliant yellow, and electric blue. This light not only illuminates her face, brimming with pure joy and excitement, but also casts ever-changing colorful reflections on her silver-white garment. The entire image is imbued with a sense of movement and fantasy, creating an atmosphere of intense liveliness and vibrant energy.", + "seed": 23781, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_0/index_4108_seed_23781_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06392550603486598, + "vlm_logprob": -0.06392550603486598, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.06392550603486598, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21497035026550293 + ], + "opposite_logprob_tokens": [ + -1.714970350265503 + ], + "logprob_reward": -0.21497035026550293, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21497035026550293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24666950106620789, + 0.12551765143871307, + 0.0809299498796463, + 0.11092010885477066, + 0.04721156880259514, + 0.037881623953580856, + 0.1395973116159439, + 0.055249299854040146, + 0.08789452910423279, + 0.06812848895788193 + ], + "credit_map": [ + 0.04240867495536804, + 0.03117309883236885, + 0.036212172359228134, + 0.0324455127120018, + 0.024269379675388336, + 0.02207663282752037, + 0.015740472823381424, + 0.04234353080391884, + 0.025558914989233017, + 0.016107788309454918, + 0.0198979452252388, + 0.011496404185891151, + 0.01399086881428957, + 0.014185798354446888, + 0.011752041056752205, + 0.0125278914347291, + 0.00954961683601141, + 0.0069778370670974255, + 0.011641746386885643, + 0.00853912066668272, + 0.007705229334533215, + 0.012055651284754276, + 0.010233576409518719, + 0.014227164909243584, + 0.02242586761713028, + 0.007914785295724869, + 0.024852700531482697, + 0.007830587215721607, + 0.006315572187304497, + 0.012081551365554333, + 0.008889439515769482, + 0.020609606057405472, + 0.006288330070674419, + 0.0063659450970590115, + 0.005764232948422432, + 0.0035105454735457897, + 0.00593726709485054, + 0.007589335553348064, + 0.005742785520851612, + 0.006013123318552971, + 0.005307271610945463, + 0.003913180436939001, + 0.005725739058107138, + 0.00422629714012146, + 0.0038094613701105118, + 0.007062605582177639, + 0.004005325026810169, + 0.0038317362777888775, + 0.016454918310046196, + 0.009653267450630665, + 0.01599419303238392, + 0.01614813506603241, + 0.015943292528390884, + 0.01626277156174183, + 0.02241883985698223, + 0.0267218966037035, + 0.005751070100814104, + 0.005216702353209257, + 0.005240230821073055, + 0.0036349822767078876, + 0.008710959926247597, + 0.011764606460928917, + 0.008180290460586548, + 0.006750452797859907, + 0.008564540185034275, + 0.009014991112053394, + 0.011601138859987259, + 0.00767169427126646, + 0.012931342236697674, + 0.01400450337678194, + 0.011767503805458546, + 0.012338807806372643, + 0.006298328284174204, + 0.004484740551561117, + 0.005529666785150766, + 0.0037760206032544374, + 0.007745387963950634, + 0.021694382652640343, + 0.006508978083729744, + 0.012090973556041718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.21497035026550293 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023321032524108887 + ], + "opposite_logprob_tokens": [ + -3.7733211517333984 + ], + "logprob_reward": -0.023321032524108887, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023321032524108887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13167060911655426, + 0.28492608666419983, + 0.056333865970373154, + 0.12869904935359955, + 0.1168336272239685, + 0.015666132792830467, + 0.03725211322307587, + 0.07305850088596344, + 0.026255840435624123, + 0.1293041706085205 + ], + "credit_map": [ + 0.021299589425325394, + 0.020188122987747192, + 0.01778961718082428, + 0.013734509237110615, + 0.012391015887260437, + 0.015906529501080513, + 0.010733299888670444, + 0.01962791197001934, + 0.04284944385290146, + 0.05025307089090347, + 0.0418313667178154, + 0.02629280649125576, + 0.03390912339091301, + 0.035638678818941116, + 0.029038196429610252, + 0.025113392621278763, + 0.005672052968293428, + 0.006445345468819141, + 0.008726345375180244, + 0.0071066259406507015, + 0.005006373394280672, + 0.008574511855840683, + 0.007121648173779249, + 0.007680960930883884, + 0.01860257238149643, + 0.006542000453919172, + 0.03440584987401962, + 0.01070458348840475, + 0.006280750501900911, + 0.011486929841339588, + 0.010090624913573265, + 0.030585728585720062, + 0.013137344270944595, + 0.022745266556739807, + 0.011754476465284824, + 0.01277724839746952, + 0.017020346596837044, + 0.009327869862318039, + 0.021528584882616997, + 0.008542491123080254, + 0.0024403438437730074, + 0.0014154111267998815, + 0.0018806784646585584, + 0.0011950468178838491, + 0.0019365566549822688, + 0.0036897780373692513, + 0.0011712189298123121, + 0.0019370978698134422, + 0.003334321081638336, + 0.005137432366609573, + 0.004442600067704916, + 0.005257796961814165, + 0.005265232175588608, + 0.004614496603608131, + 0.004822263494133949, + 0.004377968143671751, + 0.0064452397637069225, + 0.00866835005581379, + 0.007944279350340366, + 0.0053910138085484505, + 0.01146947592496872, + 0.014102901332080364, + 0.01012380514293909, + 0.0089134331792593, + 0.003082403214648366, + 0.002393280388787389, + 0.0030006386805325747, + 0.0016148132272064686, + 0.00348955811932683, + 0.005182836204767227, + 0.0034655125346034765, + 0.004026798531413078, + 0.015500277280807495, + 0.018218060955405235, + 0.015666307881474495, + 0.009823936969041824, + 0.018229398876428604, + 0.021213296800851822, + 0.012184946797788143, + 0.01846793293952942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.023321032524108887 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016094880178570747 + ], + "opposite_logprob_tokens": [ + -4.14109468460083 + ], + "logprob_reward": -0.016094880178570747, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016094880178570747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1681656837463379, + 0.1827530413866043, + 0.08817366510629654, + 0.08455182611942291, + 0.17308521270751953, + 0.03604642301797867, + 0.04828621447086334, + 0.0615314245223999, + 0.026471221819519997, + 0.13093526661396027 + ], + "credit_map": [ + 0.026409272104501724, + 0.014434294775128365, + 0.025291400030255318, + 0.025228621438145638, + 0.018572574481368065, + 0.01694687269628048, + 0.012769676744937897, + 0.0285129863768816, + 0.0359538197517395, + 0.027564870193600655, + 0.03462299332022667, + 0.015101228840649128, + 0.019893541932106018, + 0.01831519976258278, + 0.017627619206905365, + 0.013673753477633, + 0.010485482402145863, + 0.009787925519049168, + 0.013880306854844093, + 0.010725874453783035, + 0.008280518464744091, + 0.01171000674366951, + 0.01199271623045206, + 0.011310836300253868, + 0.012531111016869545, + 0.005183353554457426, + 0.019577423110604286, + 0.007690807804465294, + 0.00574590265750885, + 0.009600036777555943, + 0.007691723760217428, + 0.01653147116303444, + 0.02069694548845291, + 0.030686747282743454, + 0.015930406749248505, + 0.01862257719039917, + 0.027046972885727882, + 0.013449432328343391, + 0.03330607712268829, + 0.013346041552722454, + 0.005340904928743839, + 0.003922733012586832, + 0.005332960281521082, + 0.0030909543856978416, + 0.0031880810856819153, + 0.006746501196175814, + 0.0031920974142849445, + 0.005232188850641251, + 0.006205813959240913, + 0.0052998908795416355, + 0.005974124651402235, + 0.005400749389082193, + 0.006344171240925789, + 0.006265724077820778, + 0.00632315780967474, + 0.006472583394497633, + 0.005417318549007177, + 0.006094695068895817, + 0.005078683607280254, + 0.00472094165161252, + 0.006608579307794571, + 0.014201418496668339, + 0.011784330010414124, + 0.007625462021678686, + 0.003338464302942157, + 0.0030602910555899143, + 0.0037117914762347937, + 0.0020193541422486305, + 0.003959003835916519, + 0.00409488333389163, + 0.0029469290748238564, + 0.0033405032008886337, + 0.013430033810436726, + 0.015095801092684269, + 0.015887504443526268, + 0.009414286352694035, + 0.015355184674263, + 0.03018532134592533, + 0.011916473507881165, + 0.0196506567299366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.016094880178570747 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031998169142752886 + ], + "opposite_logprob_tokens": [ + -5.753200054168701 + ], + "logprob_reward": -0.0031998169142752886, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031998169142752886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06978242099285126, + 0.10188470035791397, + 0.03756619617342949, + 0.05248570442199707, + 0.07637263089418411, + 0.027322664856910706, + 0.07431473582983017, + 0.11701811850070953, + 0.0545610636472702, + 0.3886917531490326 + ], + "credit_map": [ + 0.011824372224509716, + 0.007875287905335426, + 0.01500675268471241, + 0.007795161567628384, + 0.006980746053159237, + 0.00555170513689518, + 0.006565390154719353, + 0.008183009922504425, + 0.018887264654040337, + 0.012142160907387733, + 0.018181821331381798, + 0.008557973429560661, + 0.014655006118118763, + 0.010427542962133884, + 0.00894649513065815, + 0.010086441412568092, + 0.005400039721280336, + 0.004186945501714945, + 0.006489226128906012, + 0.004288061521947384, + 0.0034593527670949697, + 0.004815420135855675, + 0.004787743091583252, + 0.004139409400522709, + 0.009329846128821373, + 0.0034906507935374975, + 0.010714066214859486, + 0.006133622955530882, + 0.004221766721457243, + 0.006171360146254301, + 0.005158788524568081, + 0.0072656054981052876, + 0.008311503566801548, + 0.014969374053180218, + 0.007082067430019379, + 0.007827205583453178, + 0.010288240388035774, + 0.006500208750367165, + 0.014832945540547371, + 0.006561083719134331, + 0.0040901401080191135, + 0.0039830077439546585, + 0.0034695647191256285, + 0.0030058776028454304, + 0.003606816753745079, + 0.0040293605998158455, + 0.003202013671398163, + 0.0019358827266842127, + 0.00743399141356349, + 0.0056626382283866405, + 0.010795855894684792, + 0.007620240096002817, + 0.010776863433420658, + 0.007606897968798876, + 0.012293411418795586, + 0.012124829925596714, + 0.009633613750338554, + 0.010703152976930141, + 0.010343987494707108, + 0.009797175414860249, + 0.01902569830417633, + 0.027091633528470993, + 0.016326958313584328, + 0.014095896854996681, + 0.009813541546463966, + 0.005800435319542885, + 0.004945060238242149, + 0.0036232806742191315, + 0.008903160691261292, + 0.009523112326860428, + 0.004591542296111584, + 0.0073609305545687675, + 0.04383213073015213, + 0.05714200437068939, + 0.04137931764125824, + 0.026449594646692276, + 0.053122274577617645, + 0.07533092796802521, + 0.03382378816604614, + 0.05761170759797096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.0031998169142752886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a form-fitting, iridescent, silver-white, futuristic bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have long, completely translucent hair that undulates like jellyfish tentacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hair emitting intense bioluminescence with colors shifting between cyan, magenta, yellow, and electric blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a vast, dark void dotted with faint, distant stardust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062041450291872025 + ], + "opposite_logprob_tokens": [ + -2.8120415210723877 + ], + "logprob_reward": -0.062041450291872025, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062041450291872025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16411525011062622, + 0.09264394640922546, + 0.03948821872472763, + 0.06037408858537674, + 0.07641369104385376, + 0.05616964399814606, + 0.06751963496208191, + 0.06655150651931763, + 0.025315290316939354, + 0.3514087200164795 + ], + "credit_map": [ + 0.02868572250008583, + 0.013466366566717625, + 0.031411975622177124, + 0.0150421392172575, + 0.01411919854581356, + 0.023120861500501633, + 0.013623441569507122, + 0.024645550176501274, + 0.017142893746495247, + 0.011205067858099937, + 0.015631459653377533, + 0.009194677695631981, + 0.011509598232805729, + 0.00749385729432106, + 0.010908287949860096, + 0.009558101184666157, + 0.004377337172627449, + 0.004419672302901745, + 0.0056685577146708965, + 0.004828256089240313, + 0.003947785589843988, + 0.00489741051569581, + 0.0037414678372442722, + 0.007607735227793455, + 0.010248556733131409, + 0.00335501367226243, + 0.015042097307741642, + 0.003920883871614933, + 0.003017929382622242, + 0.007403527852147818, + 0.004848165437579155, + 0.012537911534309387, + 0.009737723506987095, + 0.011985057964920998, + 0.008458740077912807, + 0.007336982060223818, + 0.0090818265452981, + 0.00849837064743042, + 0.012879962101578712, + 0.008435025811195374, + 0.007374190725386143, + 0.004938770085573196, + 0.007262464612722397, + 0.005613043438643217, + 0.0062626986764371395, + 0.011145509779453278, + 0.006379709579050541, + 0.007193256169557571, + 0.009392396546900272, + 0.006604366470128298, + 0.006407217588275671, + 0.006264974363148212, + 0.009187078103423119, + 0.007961693219840527, + 0.008396332152187824, + 0.013305576518177986, + 0.009735985659062862, + 0.005544944200664759, + 0.007606265600770712, + 0.006750508211553097, + 0.006570274941623211, + 0.009107576683163643, + 0.011154383420944214, + 0.010081573389470577, + 0.0027042278088629246, + 0.00207333080470562, + 0.0034260281827300787, + 0.0016129157738760114, + 0.004373482428491116, + 0.004902602639049292, + 0.0027987798675894737, + 0.0034239240922033787, + 0.03295079991221428, + 0.02152930200099945, + 0.018997304141521454, + 0.01328364759683609, + 0.035611849278211594, + 0.13205598294734955, + 0.031878672540187836, + 0.06510116904973984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.062041450291872025 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_1/index_tensor([778, 778])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_1/index_tensor([778, 778])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..faa40b1ab25aecb1586a0e99ea82f146ea00595b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_1/index_tensor([778, 778])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, orbiting medium close-up begins, the camera smoothly panning around a cylindrical, transparent glass container unit standing on a sterile, metal pedestal in the center of a desolate, dystopian bio-lab. The environment exhibits a stark sci-fi realism, featuring a focal wall of cold, brushed steel and a diagnostic screen displaying a static, flat diagram in the dark background. The only lighting comes from a harsh white overhead LED panel, which casts a sharp, clinical reflection on the glass and creates a subdued hum. Inside the container, a fragile specimen of a plant, its flower composed of paper-thin, withered white petals, begins to fail. Its once strong, slender green stem, now a slightly sickly yellow, trembles before slowly and inexorably drooping downward. The leaves, deprived of moisture, curl tightly inward in a final, protective gesture. As the camera continues its dispassionate pan, it captures the plant's final, graceful collapse in a continuous, detailed sequence: the main stem completely bends, and the flower head, bearing its dying petals, gently drops, landing on the cold, sterile enclosure—a poignant moment. Nature decays in a world of artificial perfection.", + "seed": 32997, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_1/index_778_seed_32997_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.029634845466352998, + "vlm_logprob": -0.029634845466352998, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.029634845466352998, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03934331610798836 + ], + "opposite_logprob_tokens": [ + -4.164343357086182 + ], + "logprob_reward": -0.03934331610798836, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.03934331610798836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.073102205991745, + 0.03574013710021973, + 0.11983843892812729, + 0.033683665096759796, + 0.03103281371295452, + 0.07568477094173431, + 0.04356444254517555, + 0.12783542275428772, + 0.42254552245140076, + 0.036972593516111374 + ], + "credit_map": [ + 0.009247798472642899, + 0.008854927495121956, + 0.01387918833643198, + 0.010091538541018963, + 0.006352716125547886, + 0.0067605809308588505, + 0.006334983743727207, + 0.011580470949411392, + 0.005345641635358334, + 0.003645807271823287, + 0.006258804351091385, + 0.005554880481213331, + 0.002994994167238474, + 0.0030698084738105536, + 0.0055215382017195225, + 0.003348660422489047, + 0.013256405480206013, + 0.009548122063279152, + 0.01951834373176098, + 0.01165318675339222, + 0.0138056306168437, + 0.016025220975279808, + 0.01856796070933342, + 0.017463570460677147, + 0.0040308465249836445, + 0.0024127704091370106, + 0.003771828021854162, + 0.00304548186250031, + 0.0028981149662286043, + 0.00575578399002552, + 0.004532738123089075, + 0.007236099801957607, + 0.0032528392039239407, + 0.002290000207722187, + 0.0035759469028562307, + 0.006045775022357702, + 0.0020708423107862473, + 0.004967908374965191, + 0.0051156519912183285, + 0.003713848302140832, + 0.007100870832800865, + 0.004745275713503361, + 0.012546233832836151, + 0.007772760931402445, + 0.006529320031404495, + 0.01623409427702427, + 0.008875193074345589, + 0.0118810273706913, + 0.005127658136188984, + 0.004975170828402042, + 0.005704838782548904, + 0.004808510188013315, + 0.004731728695333004, + 0.007136065047234297, + 0.006975431926548481, + 0.0041050356812775135, + 0.014300500974059105, + 0.0060651060193777084, + 0.013467155396938324, + 0.008028188720345497, + 0.004065122455358505, + 0.04853137955069542, + 0.006207791157066822, + 0.02717018313705921, + 0.055188387632369995, + 0.05439597740769386, + 0.04646841436624527, + 0.0812147706747055, + 0.037654951214790344, + 0.05691317468881607, + 0.055363304913043976, + 0.03534655272960663, + 0.0038757494185119867, + 0.0034309574402868748, + 0.004312859382480383, + 0.005183468107134104, + 0.0060868943110108376, + 0.00415156502276659, + 0.005959146656095982, + 0.003971950616687536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.03934331610798836 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007647286169230938 + ], + "opposite_logprob_tokens": [ + -4.882647514343262 + ], + "logprob_reward": -0.007647286169230938, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007647286169230938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030922571197152138, + 0.014553900808095932, + 0.04977649822831154, + 0.017596134915947914, + 0.01088440977036953, + 0.017574021592736244, + 0.04241658374667168, + 0.7665016651153564, + 0.03112816996872425, + 0.01864605024456978 + ], + "credit_map": [ + 0.004712133202701807, + 0.00259685842320323, + 0.00666759954765439, + 0.0032959305681288242, + 0.002747863531112671, + 0.003413204103708267, + 0.002923141699284315, + 0.004565835930407047, + 0.0021563314367085695, + 0.0014135838719084859, + 0.0016612008912488818, + 0.0025428058579564095, + 0.001484969980083406, + 0.0012366250157356262, + 0.0028368623461574316, + 0.0012215202441439033, + 0.009647665545344353, + 0.004988440778106451, + 0.007269192487001419, + 0.006486514583230019, + 0.003495770040899515, + 0.005444584414362907, + 0.007064817938953638, + 0.0053795077838003635, + 0.0020854196045547724, + 0.0015428283950313926, + 0.0022395646665245295, + 0.002438558265566826, + 0.0010616431245580316, + 0.002684704726561904, + 0.002431327011436224, + 0.003112087957561016, + 0.0015015164390206337, + 0.0012936255661770701, + 0.001601340132765472, + 0.0013719246489927173, + 0.0008763782097958028, + 0.0016967995325103402, + 0.0012740178499370813, + 0.0012688065180554986, + 0.0016530507709831, + 0.0012866691686213017, + 0.002544794697314501, + 0.0020162584260106087, + 0.002410563174635172, + 0.0028284036088734865, + 0.0021824936848133802, + 0.0026517845690250397, + 0.005552628077566624, + 0.004223233554512262, + 0.007304233033210039, + 0.00584587873890996, + 0.0038957386277616024, + 0.0068503934890031815, + 0.0034499350003898144, + 0.005294540897011757, + 0.09835144132375717, + 0.028366146609187126, + 0.07827946543693542, + 0.016924040392041206, + 0.016570406034588814, + 0.3247709572315216, + 0.018848393112421036, + 0.18439078330993652, + 0.00397245679050684, + 0.00360985123552382, + 0.003153252648189664, + 0.006064416375011206, + 0.003156010527163744, + 0.004181185737252235, + 0.004097828175872564, + 0.0028931647539138794, + 0.0020152777433395386, + 0.002016534795984626, + 0.0017426814883947372, + 0.0028356953989714384, + 0.002412464702501893, + 0.0026339355390518904, + 0.0027516998816281557, + 0.0022377583663910627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.007647286169230938 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982412464916706 + ], + "opposite_logprob_tokens": [ + -3.5298240184783936 + ], + "logprob_reward": -0.02982412464916706, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982412464916706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13486763834953308, + 0.08617904782295227, + 0.13694265484809875, + 0.075312539935112, + 0.033714260905981064, + 0.09691520780324936, + 0.09960786253213882, + 0.128774031996727, + 0.16668878495693207, + 0.0409979522228241 + ], + "credit_map": [ + 0.021007074043154716, + 0.007628831081092358, + 0.03235875815153122, + 0.011891920119524002, + 0.010926923714578152, + 0.021363863721489906, + 0.008053144440054893, + 0.021637117490172386, + 0.00902221817523241, + 0.008908369578421116, + 0.010676968842744827, + 0.013464163988828659, + 0.006354720331728458, + 0.011246154084801674, + 0.016900140792131424, + 0.009606312029063702, + 0.01724512130022049, + 0.010795490816235542, + 0.02097965218126774, + 0.012833843007683754, + 0.012058869935572147, + 0.021547699347138405, + 0.0198290403932333, + 0.021652942523360252, + 0.006890534423291683, + 0.005668302066624165, + 0.007936723530292511, + 0.007903465069830418, + 0.007852730341255665, + 0.013272091746330261, + 0.01195244025439024, + 0.013836248777806759, + 0.00368799502030015, + 0.0024321437813341618, + 0.0033256723545491695, + 0.004958245903253555, + 0.003050052560865879, + 0.00716148316860199, + 0.003366202814504504, + 0.005732466466724873, + 0.010877059772610664, + 0.007529410533607006, + 0.020655235275626183, + 0.008191606029868126, + 0.006624196190387011, + 0.023444555699825287, + 0.008972193114459515, + 0.010620956309139729, + 0.014482110738754272, + 0.008822140283882618, + 0.01674344763159752, + 0.008357606828212738, + 0.00545177236199379, + 0.02013505809009075, + 0.011829128488898277, + 0.013786598108708858, + 0.014162113890051842, + 0.012283761985599995, + 0.01709493063390255, + 0.0164748914539814, + 0.012688389047980309, + 0.022195981815457344, + 0.017404677346348763, + 0.016469284892082214, + 0.022052349522709846, + 0.020004570484161377, + 0.017287008464336395, + 0.033204011619091034, + 0.015624115243554115, + 0.022354530170559883, + 0.02249644137918949, + 0.013665761798620224, + 0.006510545965284109, + 0.005018956493586302, + 0.005447912961244583, + 0.00396697036921978, + 0.005022755824029446, + 0.008279732428491116, + 0.003193206386640668, + 0.0035578699316829443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.02982412464916706 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013423488708212972 + ], + "opposite_logprob_tokens": [ + -6.626342296600342 + ], + "logprob_reward": -0.0013423488708212972, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013423488708212972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029289448633790016, + 0.02178233489394188, + 0.08397191017866135, + 0.049789268523454666, + 0.020090054720640182, + 0.015543725341558456, + 0.029648851603269577, + 0.4245648980140686, + 0.2822651267051697, + 0.04305436834692955 + ], + "credit_map": [ + 0.004680721089243889, + 0.002998995827510953, + 0.006562771741300821, + 0.0033208823297172785, + 0.002471307758241892, + 0.0025195961352437735, + 0.0022882563062012196, + 0.004446918610483408, + 0.003707324620336294, + 0.0023665050975978374, + 0.002835661405697465, + 0.003619606141000986, + 0.002341970568522811, + 0.0019114272436127067, + 0.0026947197038680315, + 0.00230511836707592, + 0.010928642004728317, + 0.006189280189573765, + 0.014729678630828857, + 0.006997555028647184, + 0.008971990086138248, + 0.012165535241365433, + 0.011376282200217247, + 0.012612939812242985, + 0.0040689012967050076, + 0.005215131212025881, + 0.005353850312530994, + 0.005313056521117687, + 0.007452818565070629, + 0.007130385842174292, + 0.00839969888329506, + 0.006855425424873829, + 0.002281355205923319, + 0.002626152476295829, + 0.0022373241372406483, + 0.0023417919874191284, + 0.0026151402853429317, + 0.0026151868514716625, + 0.003149892669171095, + 0.002223211107775569, + 0.0010851682163774967, + 0.0015624305233359337, + 0.0038462092634290457, + 0.002191678388044238, + 0.0010155325289815664, + 0.0027650948613882065, + 0.001646082499064505, + 0.0014315296430140734, + 0.0034847811330109835, + 0.0040605803951621056, + 0.0040096137672662735, + 0.0029373732395470142, + 0.002630560426041484, + 0.005132890306413174, + 0.003946925979107618, + 0.0034461268223822117, + 0.05337296798825264, + 0.014394234865903854, + 0.041833773255348206, + 0.013493528589606285, + 0.009329349733889103, + 0.1815284937620163, + 0.010691232979297638, + 0.099921315908432, + 0.036477409303188324, + 0.036138683557510376, + 0.029722433537244797, + 0.055571869015693665, + 0.025238586589694023, + 0.03824345022439957, + 0.03784073144197464, + 0.023031944409012794, + 0.003533099079504609, + 0.003665205556899309, + 0.0028816822450608015, + 0.005603039171546698, + 0.006570240017026663, + 0.006930320058017969, + 0.008559913374483585, + 0.005310870707035065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.0013423488708212972 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001715153455734 + ], + "opposite_logprob_tokens": [ + -2.695017099380493 + ], + "logprob_reward": -0.07001715153455734, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001715153455734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.015306495130062103, + 0.013194265775382519, + 0.041961897164583206, + 0.0349736250936985, + 0.028280610218644142, + 0.01417120173573494, + 0.03066805563867092, + 0.40839359164237976, + 0.348623126745224, + 0.06442717462778091 + ], + "credit_map": [ + 0.0017542921705171466, + 0.001765976077876985, + 0.0026167510077357292, + 0.0023395293392241, + 0.001346304314211011, + 0.0017493288032710552, + 0.0016305105527862906, + 0.0021038020495325327, + 0.0024227388203144073, + 0.0013456976739689708, + 0.0020227476488798857, + 0.0022933713626116514, + 0.0012046032352373004, + 0.0011815953766927123, + 0.001559572876431048, + 0.0011639395961537957, + 0.005837582983076572, + 0.003043857403099537, + 0.005763995461165905, + 0.0038693745154887438, + 0.004942058585584164, + 0.005567646585404873, + 0.006685817148536444, + 0.006251562386751175, + 0.00273309787735343, + 0.004137091338634491, + 0.0027384604327380657, + 0.0031070071272552013, + 0.00601009838283062, + 0.005860737059265375, + 0.004909834358841181, + 0.0054772961884737015, + 0.002927251858636737, + 0.0024330951273441315, + 0.00673924433067441, + 0.0024545397609472275, + 0.0020485101267695427, + 0.003979584202170372, + 0.0023703237529844046, + 0.005328062921762466, + 0.0016416398575529456, + 0.0013465004740282893, + 0.0025285473093390465, + 0.001129606505855918, + 0.0012335011269897223, + 0.003478099824860692, + 0.0011186664924025536, + 0.0016946394462138414, + 0.003260523546487093, + 0.003251743968576193, + 0.0044137886725366116, + 0.0037772569339722395, + 0.0035598333925008774, + 0.003627980127930641, + 0.006177729927003384, + 0.002599200466647744, + 0.053390517830848694, + 0.013943669386208057, + 0.04004170745611191, + 0.010402938351035118, + 0.008393555879592896, + 0.17552797496318817, + 0.009495040401816368, + 0.09719817340373993, + 0.04539386183023453, + 0.04492310434579849, + 0.036793749779462814, + 0.0663769543170929, + 0.031886909157037735, + 0.04832897707819939, + 0.04570351168513298, + 0.029216056689620018, + 0.004509575664997101, + 0.006118425168097019, + 0.0047645834274590015, + 0.008853022940456867, + 0.01073144469410181, + 0.009826240129768848, + 0.009696154855191708, + 0.009927728213369846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.07001715153455734 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, orbiting medium close-up begins, the camera smoothly panning around a cylindrical, transparent glass container unit standing on a sterile, metal pedestal in the center of a desolate, dystopian bio-lab. The environment exhibits a stark sci-fi realism, featuring a focal wall of cold, brushed steel and a diagnostic screen displaying a static, flat diagram in the dark background. The only lighting comes from a harsh white overhead LED panel, which casts a sharp, clinical reflection on the glass and creates a subdued hum. Inside the container, a fragile specimen of a plant, its flower composed of paper-thin, withered white petals, begins to fail. Its once strong, slender green stem, now a slightly sickly yellow, trembles before slowly and inexorably drooping downward. The leaves, deprived of moisture, curl tightly inward in a final, protective gesture. As the camera continues its dispassionate pan, it captures the plant's final, graceful collapse in a continuous, detailed sequence: the main stem completely bends, and the flower head, bearing its dying petals, gently drops, landing on the cold, sterile enclosure—a poignant moment. Nature decays in a world of artificial perfection.", + "seed": 32998, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_1/index_778_seed_32998_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.019995130551978945, + "vlm_logprob": -0.019995130551978945, + "vlm_yes_no_margin": 4.349999952316284, + "vlm_token_credit": -0.019995130551978945, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03739798069000244 + ], + "opposite_logprob_tokens": [ + -4.287397861480713 + ], + "logprob_reward": -0.03739798069000244, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.03739798069000244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19499728083610535, + 0.12118077278137207, + 0.0627366229891777, + 0.06561311334371567, + 0.07046791911125183, + 0.060084957629442215, + 0.05518850311636925, + 0.0788419172167778, + 0.1317751407623291, + 0.1591138392686844 + ], + "credit_map": [ + 0.03600041940808296, + 0.011423255316913128, + 0.026318803429603577, + 0.02217748947441578, + 0.013686411082744598, + 0.025631245225667953, + 0.03491827845573425, + 0.024841343984007835, + 0.011488077230751514, + 0.016217004507780075, + 0.015998581424355507, + 0.015613707713782787, + 0.01090690866112709, + 0.016427703201770782, + 0.015606135129928589, + 0.01892264187335968, + 0.005852194968611002, + 0.0032921491656452417, + 0.01252070628106594, + 0.006360117346048355, + 0.0052897874265909195, + 0.010656931437551975, + 0.00985513161867857, + 0.008909596130251884, + 0.007066264282912016, + 0.004831796046346426, + 0.009240091778337955, + 0.006223659496754408, + 0.010625061579048634, + 0.007595913019031286, + 0.010781929828226566, + 0.009248393587768078, + 0.007428113836795092, + 0.0051758480258286, + 0.010077512823045254, + 0.013227136805653572, + 0.005281052086502314, + 0.007680766750127077, + 0.009092104621231556, + 0.012505379505455494, + 0.010274042375385761, + 0.004433397203683853, + 0.007758673280477524, + 0.01105592679232359, + 0.007932056672871113, + 0.007686452940106392, + 0.004877926781773567, + 0.006066473200917244, + 0.01218506507575512, + 0.004951913841068745, + 0.009318193420767784, + 0.008107662200927734, + 0.004948018118739128, + 0.005845186300575733, + 0.004687770269811153, + 0.005144691094756126, + 0.01032948773354292, + 0.005953139625489712, + 0.010045114904642105, + 0.012359937652945518, + 0.009078022092580795, + 0.008170736953616142, + 0.0068039679899811745, + 0.01610150747001171, + 0.01943540759384632, + 0.014171873219311237, + 0.018875205889344215, + 0.013828481547534466, + 0.01281480398029089, + 0.021592656150460243, + 0.012597016990184784, + 0.01845969632267952, + 0.017480993643403053, + 0.03295937553048134, + 0.020501306280493736, + 0.018246615305542946, + 0.01685658097267151, + 0.01955161988735199, + 0.024607859551906586, + 0.00890948623418808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.03739798069000244 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060052566230297 + ], + "opposite_logprob_tokens": [ + -3.8956005573272705 + ], + "logprob_reward": -0.02060052566230297, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060052566230297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06356238573789597, + 0.12829230725765228, + 0.06675561517477036, + 0.044185686856508255, + 0.1026020273566246, + 0.03912070766091347, + 0.059680793434381485, + 0.08607584983110428, + 0.239049032330513, + 0.17067556083202362 + ], + "credit_map": [ + 0.009741698391735554, + 0.004329984076321125, + 0.008654596284031868, + 0.009981755167245865, + 0.006674591451883316, + 0.007183637470006943, + 0.009170952253043652, + 0.00782516784965992, + 0.014360100030899048, + 0.01869519054889679, + 0.019895097240805626, + 0.0194721482694149, + 0.009738552384078503, + 0.018161917105317116, + 0.013254615478217602, + 0.014714685268700123, + 0.006127604283392429, + 0.004210163839161396, + 0.01456779707223177, + 0.008041366003453732, + 0.003526841290295124, + 0.010991016402840614, + 0.009660684503614902, + 0.009630140848457813, + 0.0066789076663553715, + 0.003593223402276635, + 0.006751293782144785, + 0.005607524886727333, + 0.005109661724418402, + 0.004324851557612419, + 0.0061010876670479774, + 0.006019135005772114, + 0.010054408572614193, + 0.00880196038633585, + 0.021787244826555252, + 0.017251888290047646, + 0.007536743767559528, + 0.011231573298573494, + 0.00836144294589758, + 0.01757676526904106, + 0.0054278988391160965, + 0.0033086517360061407, + 0.005722442641854286, + 0.006961384788155556, + 0.003636143635958433, + 0.004646133631467819, + 0.003913435619324446, + 0.00550461933016777, + 0.008631456643342972, + 0.006051302887499332, + 0.007853667251765728, + 0.007410596590489149, + 0.009154416620731354, + 0.00639156810939312, + 0.006062903441488743, + 0.0081248814240098, + 0.01388559304177761, + 0.0075841438956558704, + 0.016575271263718605, + 0.01683306135237217, + 0.007685779128223658, + 0.006018337327986956, + 0.00847371481359005, + 0.00901995599269867, + 0.03930962458252907, + 0.025589944794774055, + 0.03422043099999428, + 0.03519866243004799, + 0.01930045150220394, + 0.037682656198740005, + 0.024740885943174362, + 0.0230063796043396, + 0.016493219882249832, + 0.027617452666163445, + 0.02294716238975525, + 0.014007753692567348, + 0.01673680730164051, + 0.028005922213196754, + 0.025671258568763733, + 0.019195983186364174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.02060052566230297 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819678582251072 + ], + "opposite_logprob_tokens": [ + -4.0181965827941895 + ], + "logprob_reward": -0.01819678582251072, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819678582251072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19748282432556152, + 0.13208094239234924, + 0.12404578924179077, + 0.051865845918655396, + 0.07594068348407745, + 0.15084916353225708, + 0.04335029050707817, + 0.11462671309709549, + 0.04453413188457489, + 0.06522366404533386 + ], + "credit_map": [ + 0.039908308535814285, + 0.012532036751508713, + 0.025758031755685806, + 0.020390426740050316, + 0.011593405157327652, + 0.03898315876722336, + 0.018363559618592262, + 0.029953883960843086, + 0.014354800805449486, + 0.015614194795489311, + 0.021569645032286644, + 0.016261791810393333, + 0.010968809947371483, + 0.01618153229355812, + 0.015964161604642868, + 0.021165994927287102, + 0.009013086557388306, + 0.014369452372193336, + 0.018857596442103386, + 0.021339187398552895, + 0.008807078003883362, + 0.017824584618210793, + 0.01658528484404087, + 0.017249519005417824, + 0.005470059812068939, + 0.004423840902745724, + 0.010982665233314037, + 0.0070350696332752705, + 0.0040608616545796394, + 0.005918850190937519, + 0.006366003770381212, + 0.007608495652675629, + 0.006859822664409876, + 0.00577487563714385, + 0.01651323027908802, + 0.012828432023525238, + 0.004895029589533806, + 0.009139860980212688, + 0.006131357979029417, + 0.013798079453408718, + 0.02309887669980526, + 0.013412859290838242, + 0.01997668854892254, + 0.02705758437514305, + 0.01784958876669407, + 0.020806346088647842, + 0.01413680985569954, + 0.014510397799313068, + 0.00896124355494976, + 0.0048749991692602634, + 0.004770873114466667, + 0.004622750449925661, + 0.005905489902943373, + 0.0026740743778645992, + 0.0038931649178266525, + 0.007647692691534758, + 0.019645489752292633, + 0.010389034636318684, + 0.0233662948012352, + 0.01883521117269993, + 0.007233749143779278, + 0.01625189371407032, + 0.006718999706208706, + 0.012186042964458466, + 0.005532061215490103, + 0.004846411291509867, + 0.0068994867615401745, + 0.0046205357648432255, + 0.00639417814090848, + 0.006728694308549166, + 0.005164528265595436, + 0.004348239861428738, + 0.006731043104082346, + 0.008776000700891018, + 0.007849649526178837, + 0.008085742592811584, + 0.006738963536918163, + 0.006968229543417692, + 0.013912810012698174, + 0.006161228287965059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.01819678582251072 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031991039868444204 + ], + "opposite_logprob_tokens": [ + -5.753199100494385 + ], + "logprob_reward": -0.0031991039868444204, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031991039868444204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27048903703689575, + 0.0910186767578125, + 0.10003210604190826, + 0.1183573380112648, + 0.0314212292432785, + 0.03932711482048035, + 0.032130349427461624, + 0.10109962522983551, + 0.05742822214961052, + 0.15869630873203278 + ], + "credit_map": [ + 0.03528142347931862, + 0.017758723348379135, + 0.05277498438954353, + 0.03936931863427162, + 0.03326903283596039, + 0.02569640427827835, + 0.03203343227505684, + 0.03430570289492607, + 0.006408837623894215, + 0.006821624469012022, + 0.012580947950482368, + 0.01238774973899126, + 0.010555923916399479, + 0.011179548688232899, + 0.011951872147619724, + 0.01913216896355152, + 0.005118987988680601, + 0.00956402625888586, + 0.014543313533067703, + 0.008780141361057758, + 0.0158150065690279, + 0.013419774360954762, + 0.019125493243336678, + 0.013665362261235714, + 0.010324085131287575, + 0.011174424551427364, + 0.013387126848101616, + 0.014581659808754921, + 0.016782917082309723, + 0.019181272014975548, + 0.014753559604287148, + 0.018172286450862885, + 0.004297066014260054, + 0.002548507647588849, + 0.005070417653769255, + 0.004155843984335661, + 0.002872513374313712, + 0.004289908800274134, + 0.0038074844051152468, + 0.0043794852681458, + 0.006645652931183577, + 0.0033519051503390074, + 0.005759421270340681, + 0.005572827532887459, + 0.004578576423227787, + 0.0060278382152318954, + 0.003813941264525056, + 0.003576951567083597, + 0.005167621653527021, + 0.003381300950422883, + 0.005533225834369659, + 0.003469645045697689, + 0.0033587152138352394, + 0.0033192660193890333, + 0.0037013371475040913, + 0.004199235234409571, + 0.017429735511541367, + 0.007689422462135553, + 0.025141984224319458, + 0.015305927954614162, + 0.008193396031856537, + 0.01226020511239767, + 0.006797501817345619, + 0.008281450718641281, + 0.006316781044006348, + 0.0056588314473629, + 0.006425146479159594, + 0.0072685182094573975, + 0.008636357262730598, + 0.00805063359439373, + 0.009919474832713604, + 0.005152474157512188, + 0.016332345083355904, + 0.016978641971945763, + 0.019153958186507225, + 0.025461966171860695, + 0.015623502433300018, + 0.014199832454323769, + 0.03119989112019539, + 0.01974615454673767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.0031991039868444204 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581256598234177 + ], + "opposite_logprob_tokens": [ + -3.8955812454223633 + ], + "logprob_reward": -0.020581256598234177, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581256598234177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056153930723667145, + 0.03421822935342789, + 0.11162218451499939, + 0.050438109785318375, + 0.07343749701976776, + 0.0455484576523304, + 0.03836226835846901, + 0.1568264216184616, + 0.062332771718502045, + 0.37106016278266907 + ], + "credit_map": [ + 0.015065276063978672, + 0.004240983631461859, + 0.005901988595724106, + 0.0058405459858477116, + 0.003820792306214571, + 0.006929739844053984, + 0.007299979217350483, + 0.007054623682051897, + 0.0033840849064290524, + 0.0036954297684133053, + 0.005001432728022337, + 0.004075116477906704, + 0.004193155560642481, + 0.004290069919079542, + 0.004807273857295513, + 0.004771665204316378, + 0.005458284635096788, + 0.008595874533057213, + 0.007848125882446766, + 0.010132314637303352, + 0.02108142338693142, + 0.017113149166107178, + 0.02780369482934475, + 0.013589312322437763, + 0.004758643452078104, + 0.00395040912553668, + 0.006322473753243685, + 0.0058769239112734795, + 0.006083791609853506, + 0.005108399782329798, + 0.010775377042591572, + 0.007562088780105114, + 0.007385500706732273, + 0.005632209125906229, + 0.01598166488111019, + 0.011759571731090546, + 0.00500256335362792, + 0.008887049742043018, + 0.004570907447487116, + 0.014218018390238285, + 0.010191274806857109, + 0.0040253144688904285, + 0.005873965565115213, + 0.0060113174840807915, + 0.005395950749516487, + 0.005566667299717665, + 0.003557878779247403, + 0.00492608780041337, + 0.005686810240149498, + 0.004009967669844627, + 0.005545303225517273, + 0.0060949744656682014, + 0.004539508372545242, + 0.0032199618872255087, + 0.00546311866492033, + 0.003802620805799961, + 0.02430533617734909, + 0.012334281578660011, + 0.030912339687347412, + 0.02782050147652626, + 0.014404179528355598, + 0.010618556290864944, + 0.019002290442585945, + 0.017428923398256302, + 0.0083024762570858, + 0.005473065655678511, + 0.006395997479557991, + 0.009150701574981213, + 0.006639872211962938, + 0.010812024585902691, + 0.009140984155237675, + 0.0064176456071436405, + 0.027019143104553223, + 0.03249983862042427, + 0.05536501109600067, + 0.06531288474798203, + 0.03835822269320488, + 0.022039620205760002, + 0.08677385747432709, + 0.043691568076610565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.020581256598234177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, orbiting medium close-up begins, the camera smoothly panning around a cylindrical, transparent glass container unit standing on a sterile, metal pedestal in the center of a desolate, dystopian bio-lab. The environment exhibits a stark sci-fi realism, featuring a focal wall of cold, brushed steel and a diagnostic screen displaying a static, flat diagram in the dark background. The only lighting comes from a harsh white overhead LED panel, which casts a sharp, clinical reflection on the glass and creates a subdued hum. Inside the container, a fragile specimen of a plant, its flower composed of paper-thin, withered white petals, begins to fail. Its once strong, slender green stem, now a slightly sickly yellow, trembles before slowly and inexorably drooping downward. The leaves, deprived of moisture, curl tightly inward in a final, protective gesture. As the camera continues its dispassionate pan, it captures the plant's final, graceful collapse in a continuous, detailed sequence: the main stem completely bends, and the flower head, bearing its dying petals, gently drops, landing on the cold, sterile enclosure—a poignant moment. Nature decays in a world of artificial perfection.", + "seed": 32999, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_1/index_778_seed_32999_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03927346961572766, + "vlm_logprob": -0.03927346961572766, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.03927346961572766, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07085651904344559 + ], + "opposite_logprob_tokens": [ + -4.19585657119751 + ], + "logprob_reward": -0.07085651904344559, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.07085651904344559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10561869293451309, + 0.09203191101551056, + 0.07495550811290741, + 0.07012646645307541, + 0.1572066843509674, + 0.09342411160469055, + 0.10639884322881699, + 0.05778573825955391, + 0.12078415602445602, + 0.12166786938905716 + ], + "credit_map": [ + 0.01392168365418911, + 0.007123869378119707, + 0.01456624735146761, + 0.02374165877699852, + 0.008122331462800503, + 0.010100101120769978, + 0.013306836597621441, + 0.014735965989530087, + 0.02185044251382351, + 0.008046384900808334, + 0.01460733637213707, + 0.01233639381825924, + 0.008458429016172886, + 0.007208111695945263, + 0.010448792949318886, + 0.009076010435819626, + 0.008871456608176231, + 0.0059951599687337875, + 0.010724840685725212, + 0.011976353824138641, + 0.0081346919760108, + 0.006625056266784668, + 0.016609713435173035, + 0.00601823627948761, + 0.007506917230784893, + 0.006047853734344244, + 0.01101776771247387, + 0.011858366429805756, + 0.0043655987828969955, + 0.004651431925594807, + 0.012588737532496452, + 0.012089794501662254, + 0.014355638064444065, + 0.008487765677273273, + 0.04786578565835953, + 0.016042478382587433, + 0.010206114500761032, + 0.008445662446320057, + 0.017449159175157547, + 0.034354083240032196, + 0.00754176639020443, + 0.008597041480243206, + 0.026347743347287178, + 0.014936123974621296, + 0.005287489388138056, + 0.006594070699065924, + 0.013606280088424683, + 0.010513591580092907, + 0.00849334429949522, + 0.013515236787497997, + 0.012275218963623047, + 0.018265141174197197, + 0.011472649872303009, + 0.014507901854813099, + 0.016879452392458916, + 0.010989896953105927, + 0.008976839482784271, + 0.005197800230234861, + 0.009413842111825943, + 0.00655066454783082, + 0.006964303553104401, + 0.004959030542522669, + 0.007197526749223471, + 0.008525730110704899, + 0.017650414258241653, + 0.02043551206588745, + 0.01293204165995121, + 0.008959870785474777, + 0.02606508694589138, + 0.013109328225255013, + 0.015278120525181293, + 0.006353772711008787, + 0.02707548253238201, + 0.027195997536182404, + 0.015396618284285069, + 0.009735982865095139, + 0.012578289024531841, + 0.010715746320784092, + 0.009951555170118809, + 0.009018195793032646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.07085651904344559 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012574437074363232 + ], + "opposite_logprob_tokens": [ + -4.387574672698975 + ], + "logprob_reward": -0.012574437074363232, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012574437074363232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09800560772418976, + 0.0908370390534401, + 0.2517579197883606, + 0.07610412687063217, + 0.10603903979063034, + 0.05753384903073311, + 0.04105932638049126, + 0.07139275223016739, + 0.11514507234096527, + 0.092125304043293 + ], + "credit_map": [ + 0.014035855419933796, + 0.006830528378486633, + 0.01853816956281662, + 0.015413072891533375, + 0.007501979824155569, + 0.007908729836344719, + 0.013779419474303722, + 0.013997852802276611, + 0.02626897022128105, + 0.004911745898425579, + 0.015847640112042427, + 0.00927871372550726, + 0.003869100473821163, + 0.010862858965992928, + 0.007917298004031181, + 0.01188071258366108, + 0.022684108465909958, + 0.017929092049598694, + 0.04017060250043869, + 0.051034197211265564, + 0.02842472307384014, + 0.02649075724184513, + 0.04205621778964996, + 0.022968214005231857, + 0.005870386958122253, + 0.007165797986090183, + 0.016309117898344994, + 0.012103043496608734, + 0.004650535061955452, + 0.008319071494042873, + 0.010416846722364426, + 0.01126932818442583, + 0.009560427628457546, + 0.012526693753898144, + 0.011607080698013306, + 0.01713300682604313, + 0.01253236923366785, + 0.009055973961949348, + 0.020020978525280952, + 0.013602503575384617, + 0.0060800183564424515, + 0.006346951704472303, + 0.007853297516703606, + 0.007081194780766964, + 0.005190713331103325, + 0.008183670230209827, + 0.009217674843966961, + 0.007580333389341831, + 0.003421608591452241, + 0.005018648691475391, + 0.004095734562724829, + 0.005461685825139284, + 0.0047590904869139194, + 0.0056806826032698154, + 0.007466867100447416, + 0.005155009217560291, + 0.006822838447988033, + 0.006675546523183584, + 0.009605677798390388, + 0.017692118883132935, + 0.006621720269322395, + 0.00769415870308876, + 0.009102425538003445, + 0.007178264670073986, + 0.02040526270866394, + 0.017101025208830833, + 0.01262387353926897, + 0.008616780862212181, + 0.023251468315720558, + 0.010690171271562576, + 0.015388136729598045, + 0.007068352308124304, + 0.012507932260632515, + 0.010988523252308369, + 0.009832713752985, + 0.008391175419092178, + 0.01932605542242527, + 0.006021815352141857, + 0.016201728954911232, + 0.008855368942022324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.012574437074363232 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868929833173752 + ], + "opposite_logprob_tokens": [ + -3.048689365386963 + ], + "logprob_reward": -0.04868929833173752, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868929833173752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16017790138721466, + 0.04005725309252739, + 0.11542032659053802, + 0.1181410476565361, + 0.12561969459056854, + 0.06264309585094452, + 0.09037529677152634, + 0.08476503193378448, + 0.13144081830978394, + 0.07135950773954391 + ], + "credit_map": [ + 0.020041797310113907, + 0.010559221729636192, + 0.022595304995775223, + 0.026472385972738266, + 0.01254019234329462, + 0.016311312094330788, + 0.03019484505057335, + 0.02146284468472004, + 0.009187477640807629, + 0.002125097205862403, + 0.006675310432910919, + 0.005339952185750008, + 0.00194387580268085, + 0.003919921815395355, + 0.005100814625620842, + 0.005764804780483246, + 0.010774186812341213, + 0.007783459033817053, + 0.017969932407140732, + 0.02221275493502617, + 0.010930477641522884, + 0.013520440086722374, + 0.019653869792819023, + 0.01257520541548729, + 0.009578828699886799, + 0.008306194096803665, + 0.03266579285264015, + 0.018700454384088516, + 0.009118415415287018, + 0.007300222758203745, + 0.012782882899045944, + 0.019688259810209274, + 0.009940126910805702, + 0.010749602690339088, + 0.02997044287621975, + 0.013425000011920929, + 0.008395354263484478, + 0.007035507820546627, + 0.02155284956097603, + 0.02455080859363079, + 0.004634753335267305, + 0.005695965141057968, + 0.007285985630005598, + 0.007433814462274313, + 0.004630255047231913, + 0.00884427409619093, + 0.015368746593594551, + 0.008749299682676792, + 0.007764576468616724, + 0.012351591140031815, + 0.01103082112967968, + 0.014471831731498241, + 0.011538177728652954, + 0.011210581287741661, + 0.010730593465268612, + 0.011277125217020512, + 0.010863308794796467, + 0.011564710177481174, + 0.01065848208963871, + 0.011843726970255375, + 0.013119337148964405, + 0.012640003114938736, + 0.007914462126791477, + 0.006161000579595566, + 0.026086969301104546, + 0.02148827724158764, + 0.015323895961046219, + 0.011657793074846268, + 0.016523070633411407, + 0.011563539505004883, + 0.018259158357977867, + 0.010538105852901936, + 0.009656623005867004, + 0.009087846614420414, + 0.010940161533653736, + 0.008184300735592842, + 0.011452386155724525, + 0.00621148431673646, + 0.008243163116276264, + 0.007583535276353359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.04868929833173752 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022050845436751842 + ], + "opposite_logprob_tokens": [ + -6.127204895019531 + ], + "logprob_reward": -0.0022050845436751842, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022050845436751842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10159232467412949, + 0.06123512610793114, + 0.07180141657590866, + 0.1886843591928482, + 0.11061329394578934, + 0.10799045115709305, + 0.11182558536529541, + 0.07057038694620132, + 0.12109358608722687, + 0.05459348484873772 + ], + "credit_map": [ + 0.008342541754245758, + 0.009595809504389763, + 0.010672875680029392, + 0.016073592007160187, + 0.010209892876446247, + 0.01340683363378048, + 0.018778938800096512, + 0.014511840417981148, + 0.009673630818724632, + 0.007466494105756283, + 0.007739880587905645, + 0.007055000867694616, + 0.007728972006589174, + 0.005790147930383682, + 0.01084968913346529, + 0.004931311123073101, + 0.006920414976775646, + 0.0052023883908987045, + 0.011422998271882534, + 0.013097592629492283, + 0.005375530570745468, + 0.009138759225606918, + 0.013300741091370583, + 0.007342984434217215, + 0.012897090055048466, + 0.017862284556031227, + 0.04097670689225197, + 0.035171326249837875, + 0.016795149073004723, + 0.010885088704526424, + 0.027752766385674477, + 0.02634393982589245, + 0.008184345439076424, + 0.007803112734109163, + 0.03164764493703842, + 0.008325343020260334, + 0.010172619484364986, + 0.006442293990403414, + 0.013166103512048721, + 0.02487182803452015, + 0.011035540141165257, + 0.012975270859897137, + 0.016554487869143486, + 0.013631928712129593, + 0.010166359134018421, + 0.017264332622289658, + 0.015075568109750748, + 0.011286958120763302, + 0.008906498551368713, + 0.010537165217101574, + 0.013187875039875507, + 0.017430435866117477, + 0.009471972472965717, + 0.012809337116777897, + 0.027465270832180977, + 0.012017027474939823, + 0.006251065526157618, + 0.005982517264783382, + 0.008869314566254616, + 0.012258052825927734, + 0.0076295300386846066, + 0.00995076633989811, + 0.011322975158691406, + 0.008306164294481277, + 0.014900133945047855, + 0.016955697908997536, + 0.014568811282515526, + 0.010492715984582901, + 0.01835763268172741, + 0.00980207696557045, + 0.02607480250298977, + 0.0099417082965374, + 0.00818383228033781, + 0.008102146908640862, + 0.009475688450038433, + 0.004207693971693516, + 0.009160862304270267, + 0.004841219168156385, + 0.0063652354292571545, + 0.00425680773332715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.0022050845436751842 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062042009085416794 + ], + "opposite_logprob_tokens": [ + -2.812041997909546 + ], + "logprob_reward": -0.062042009085416794, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062042009085416794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07746988534927368, + 0.0715487077832222, + 0.06650839745998383, + 0.06710116565227509, + 0.19581040740013123, + 0.10625254362821579, + 0.057590827345848083, + 0.05356103926897049, + 0.12627670168876648, + 0.17788024246692657 + ], + "credit_map": [ + 0.008498079143464565, + 0.006501744966953993, + 0.009035836905241013, + 0.013157098554074764, + 0.005395489279180765, + 0.00783787202090025, + 0.010964889079332352, + 0.016078874468803406, + 0.01855558529496193, + 0.005993954837322235, + 0.011490550823509693, + 0.007842178456485271, + 0.005886789411306381, + 0.007367840968072414, + 0.006402838975191116, + 0.008008971810340881, + 0.005017252638936043, + 0.008065580390393734, + 0.0084620276466012, + 0.009780754335224628, + 0.009754552505910397, + 0.00650487095117569, + 0.013371330685913563, + 0.005552023649215698, + 0.005277679767459631, + 0.006135245785117149, + 0.0174045879393816, + 0.00991874560713768, + 0.007103261537849903, + 0.004469497129321098, + 0.0069127874448895454, + 0.00987935345619917, + 0.01725379191339016, + 0.023825732991099358, + 0.024934452027082443, + 0.022094350308179855, + 0.02861182764172554, + 0.017382090911269188, + 0.0366930216550827, + 0.02501515857875347, + 0.008047266863286495, + 0.013857223093509674, + 0.014457964338362217, + 0.013902313075959682, + 0.009439184330403805, + 0.014263143762946129, + 0.019832858815789223, + 0.012452585622668266, + 0.004765305668115616, + 0.0054568941704928875, + 0.005674679297953844, + 0.011120287701487541, + 0.004227341152727604, + 0.005668367259204388, + 0.014847302809357643, + 0.0058306441642344, + 0.004716931842267513, + 0.004932400304824114, + 0.00831590499728918, + 0.008770247921347618, + 0.005950454156845808, + 0.004548756871372461, + 0.010668911039829254, + 0.00565743213519454, + 0.015830498188734055, + 0.018208177760243416, + 0.012822912074625492, + 0.010355985723435879, + 0.0200357623398304, + 0.010352768935263157, + 0.023833954706788063, + 0.014836649410426617, + 0.05048929154872894, + 0.05276499688625336, + 0.019368598237633705, + 0.009154238738119602, + 0.017607783898711205, + 0.00809094961732626, + 0.011407631449401379, + 0.008996744640171528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.062042009085416794 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, orbiting medium close-up begins, the camera smoothly panning around a cylindrical, transparent glass container unit standing on a sterile, metal pedestal in the center of a desolate, dystopian bio-lab. The environment exhibits a stark sci-fi realism, featuring a focal wall of cold, brushed steel and a diagnostic screen displaying a static, flat diagram in the dark background. The only lighting comes from a harsh white overhead LED panel, which casts a sharp, clinical reflection on the glass and creates a subdued hum. Inside the container, a fragile specimen of a plant, its flower composed of paper-thin, withered white petals, begins to fail. Its once strong, slender green stem, now a slightly sickly yellow, trembles before slowly and inexorably drooping downward. The leaves, deprived of moisture, curl tightly inward in a final, protective gesture. As the camera continues its dispassionate pan, it captures the plant's final, graceful collapse in a continuous, detailed sequence: the main stem completely bends, and the flower head, bearing its dying petals, gently drops, landing on the cold, sterile enclosure—a poignant moment. Nature decays in a world of artificial perfection.", + "seed": 33000, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_1/index_778_seed_33000_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.020492793363519012, + "vlm_logprob": -0.020492793363519012, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.020492793363519012, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03237428143620491 + ], + "opposite_logprob_tokens": [ + -4.282374382019043 + ], + "logprob_reward": -0.03237428143620491, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.03237428143620491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4586387276649475, + 0.031168252229690552, + 0.04831961542367935, + 0.03709326311945915, + 0.027791691944003105, + 0.01953401230275631, + 0.023960508406162262, + 0.030488377436995506, + 0.16509012877941132, + 0.15791544318199158 + ], + "credit_map": [ + 0.02613496407866478, + 0.006071063224226236, + 0.010523035190999508, + 0.008136994205415249, + 0.008459731005132198, + 0.20478606224060059, + 0.010080656036734581, + 0.18444621562957764, + 0.006058908998966217, + 0.0024851213674992323, + 0.004400782752782106, + 0.0045073265209794044, + 0.002757528331130743, + 0.0042543173767626286, + 0.003755867714062333, + 0.0029483989346772432, + 0.01084047555923462, + 0.003124810755252838, + 0.007061365991830826, + 0.003860536264255643, + 0.003058506641536951, + 0.011285088025033474, + 0.005222254432737827, + 0.003866578219458461, + 0.008243585005402565, + 0.002218222711235285, + 0.006709014531224966, + 0.003879518248140812, + 0.00257120537571609, + 0.005690323654562235, + 0.003966757096350193, + 0.0038146376609802246, + 0.0031070876866579056, + 0.002613543067127466, + 0.0032459739595651627, + 0.0039253332652151585, + 0.0038074939511716366, + 0.004705691710114479, + 0.0038298554718494415, + 0.002556713530793786, + 0.0031985025852918625, + 0.0021473243832588196, + 0.0020990953780710697, + 0.002837754087522626, + 0.0022076598834246397, + 0.002782968571409583, + 0.0019998939242213964, + 0.002260811859741807, + 0.0023397738113999367, + 0.002138375537469983, + 0.0028488836251199245, + 0.002415357856079936, + 0.0011485619470477104, + 0.004804020281881094, + 0.001932881772518158, + 0.006332653108984232, + 0.0025453930720686913, + 0.002847383264452219, + 0.003007870865985751, + 0.0024403496645390987, + 0.002652107272297144, + 0.007917026989161968, + 0.004243627656251192, + 0.004834618419408798, + 0.016120171174407005, + 0.01651749759912491, + 0.01726982742547989, + 0.025607213377952576, + 0.016470246016979218, + 0.028787847608327866, + 0.028014283627271652, + 0.01630304753780365, + 0.018243687227368355, + 0.018797865137457848, + 0.017252959311008453, + 0.01926172524690628, + 0.014389649964869022, + 0.025769764557480812, + 0.019117699936032295, + 0.025082098320126534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.03237428143620491 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004110934678465128 + ], + "opposite_logprob_tokens": [ + -5.504110813140869 + ], + "logprob_reward": -0.004110934678465128, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004110934678465128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6471368074417114, + 0.018858997151255608, + 0.01106853224337101, + 0.012650017626583576, + 0.0186191126704216, + 0.019929880276322365, + 0.019285742193460464, + 0.019416680559515953, + 0.05875834450125694, + 0.1742759346961975 + ], + "credit_map": [ + 0.03630950674414635, + 0.010710390284657478, + 0.01505197212100029, + 0.014591187238693237, + 0.01338778343051672, + 0.28325116634368896, + 0.022278238087892532, + 0.25155648589134216, + 0.002366304863244295, + 0.0016381629975512624, + 0.002854498801752925, + 0.0028412931133061647, + 0.0015955413691699505, + 0.002343445084989071, + 0.003620911156758666, + 0.0015988403465598822, + 0.0030519594438374043, + 0.0008122059516608715, + 0.0013257863465696573, + 0.0009100645547732711, + 0.0006744182319380343, + 0.0018947955686599016, + 0.0012857230613008142, + 0.0011135793756693602, + 0.0018209624104201794, + 0.0012023409362882376, + 0.001671415171585977, + 0.0016580132069066167, + 0.0013058788608759642, + 0.0019892300479114056, + 0.0018060306319966912, + 0.0011961461277678609, + 0.0024056669790297747, + 0.0016783318715170026, + 0.002056747442111373, + 0.002903213258832693, + 0.0018461570143699646, + 0.0030854183714836836, + 0.0025573300663381815, + 0.0020862494129687548, + 0.0029344542417675257, + 0.0021970251109451056, + 0.0021924751345068216, + 0.002834704937413335, + 0.001981875393539667, + 0.002941506216302514, + 0.0023167370818555355, + 0.002531100297346711, + 0.0015567945083603263, + 0.0017132902285084128, + 0.002127936575561762, + 0.003562037367373705, + 0.0011156467953696847, + 0.0032691629603505135, + 0.0022480550687760115, + 0.0036928183399140835, + 0.001555110327899456, + 0.0019090172136202455, + 0.0016872929409146309, + 0.0019359794678166509, + 0.0017371929716318846, + 0.005312557332217693, + 0.0020028322469443083, + 0.0032766980584710836, + 0.005538338329643011, + 0.005407246295362711, + 0.006204329896718264, + 0.00884464755654335, + 0.006161234807223082, + 0.010746686719357967, + 0.009907176718115807, + 0.005948685575276613, + 0.021487032994627953, + 0.021320639178156853, + 0.019334152340888977, + 0.0226691123098135, + 0.01562481839209795, + 0.02811741642653942, + 0.01774313673377037, + 0.027979619801044464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.004110934678465128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026365121826529503 + ], + "opposite_logprob_tokens": [ + -3.651365041732788 + ], + "logprob_reward": -0.026365121826529503, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026365121826529503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.272724449634552, + 0.016712242737412453, + 0.015599372796714306, + 0.06885319948196411, + 0.06905658543109894, + 0.04812058433890343, + 0.0594744049012661, + 0.12537285685539246, + 0.2763899862766266, + 0.0476963110268116 + ], + "credit_map": [ + 0.0151132270693779, + 0.008359887637197971, + 0.009462452493607998, + 0.013231481425464153, + 0.008069541305303574, + 0.10775595158338547, + 0.010712485760450363, + 0.1000194102525711, + 0.0020878661889582872, + 0.0012001178693026304, + 0.002468004822731018, + 0.002070584800094366, + 0.002150035696104169, + 0.0021702360827475786, + 0.0033468687906861305, + 0.0012185294181108475, + 0.003092935774475336, + 0.0011985430028289557, + 0.002419138792902231, + 0.0016585587291046977, + 0.0015749590238556266, + 0.0021330274175852537, + 0.001692204736173153, + 0.0018300054362043738, + 0.016507575288414955, + 0.005827852059155703, + 0.011857871897518635, + 0.006156648974865675, + 0.00481568556278944, + 0.011202521622180939, + 0.005324796307832003, + 0.007160250097513199, + 0.010162720456719398, + 0.00551808625459671, + 0.006330912932753563, + 0.01196391973644495, + 0.006915780249983072, + 0.009485618211328983, + 0.009311486966907978, + 0.009368070401251316, + 0.010667631402611732, + 0.0046153892762959, + 0.004843158181756735, + 0.007718560751527548, + 0.0043401410803198814, + 0.007832123897969723, + 0.003785340581089258, + 0.004318240098655224, + 0.0041598048992455006, + 0.003101264825090766, + 0.007305805571377277, + 0.0062549845315515995, + 0.0021928204223513603, + 0.013575767166912556, + 0.00546139245852828, + 0.017422573640942574, + 0.008859138004481792, + 0.010170389898121357, + 0.008815601468086243, + 0.010881003923714161, + 0.008851975202560425, + 0.0404655821621418, + 0.01523091271519661, + 0.02209826558828354, + 0.02654234878718853, + 0.02751932665705681, + 0.02906837873160839, + 0.042965419590473175, + 0.02699373848736286, + 0.04878011718392372, + 0.04726685956120491, + 0.027253825217485428, + 0.007024201564490795, + 0.004600260406732559, + 0.010091487318277359, + 0.005105263087898493, + 0.004571467638015747, + 0.005368210840970278, + 0.004981043748557568, + 0.0059543815441429615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.026365121826529503 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015180503251031041 + ], + "opposite_logprob_tokens": [ + -6.501518249511719 + ], + "logprob_reward": -0.0015180503251031041, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015180503251031041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.258833646774292, + 0.019446665421128273, + 0.010850818827748299, + 0.006617823150008917, + 0.011603633873164654, + 0.01403918955475092, + 0.016062846407294273, + 0.05624229088425636, + 0.43977057933807373, + 0.1665325164794922 + ], + "credit_map": [ + 0.015435055829584599, + 0.00603487528860569, + 0.006513173691928387, + 0.007207490503787994, + 0.007805144879966974, + 0.11119173467159271, + 0.007684700191020966, + 0.09696146845817566, + 0.0030347886495292187, + 0.001708518946543336, + 0.0021692183800041676, + 0.0022574521135538816, + 0.0019974117167294025, + 0.002319849096238613, + 0.004207123536616564, + 0.0017523025162518024, + 0.002728451043367386, + 0.0007555348565801978, + 0.001660862355493009, + 0.0007938557537272573, + 0.0009674401953816414, + 0.0021703478414565325, + 0.0009800551924854517, + 0.0007942712982185185, + 0.0008468646556138992, + 0.0005403615068644285, + 0.001089310273528099, + 0.0007490465650334954, + 0.0006541696493513882, + 0.0009980847826227546, + 0.0007996734930202365, + 0.0009403117001056671, + 0.0019292992074042559, + 0.0009270964656025171, + 0.002125900937244296, + 0.001173267257399857, + 0.000759966962505132, + 0.002488442463800311, + 0.0010026843519881368, + 0.0011969764018431306, + 0.002022008877247572, + 0.0016110835131257772, + 0.0014453916810452938, + 0.00235953857190907, + 0.0017852470045909286, + 0.0019933655858039856, + 0.0013323829043656588, + 0.001490171067416668, + 0.0014540624106302857, + 0.001490656053647399, + 0.002209722762927413, + 0.0024376248475164175, + 0.0010719660203903913, + 0.0024270121939480305, + 0.0019255344523116946, + 0.0030462692957371473, + 0.005171747878193855, + 0.006219365634024143, + 0.0055132051929831505, + 0.008645672351121902, + 0.004405589308589697, + 0.010029442608356476, + 0.006889742333441973, + 0.009367521852254868, + 0.04248008131980896, + 0.04397593438625336, + 0.04604576528072357, + 0.06809315085411072, + 0.04331192374229431, + 0.07784470170736313, + 0.07439877092838287, + 0.04362024366855621, + 0.0203216802328825, + 0.018989015370607376, + 0.017674140632152557, + 0.020311910659074783, + 0.014760205522179604, + 0.02816501073539257, + 0.017747949808835983, + 0.028562597930431366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.0015180503251031041 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium close-up that orbits or pans smoothly around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a cylindrical, transparent glass container standing on a sterile metal pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a desolate, dystopian bio-lab with a focal wall of cold brushed steel and a diagnostic screen showing a static flat diagram?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the plant specimen inside the container have a stem that changes from green to sickly yellow and droops downward while its white petals wither and curl inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the plant's flower head eventually detach and land on the cold sterile enclosure of the container?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809557855129242 + ], + "opposite_logprob_tokens": [ + -3.288095474243164 + ], + "logprob_reward": -0.03809557855129242, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809557855129242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033501192927360535, + 0.006265370175242424, + 0.006309462711215019, + 0.004039266146719456, + 0.008036378771066666, + 0.011457657441496849, + 0.010102957487106323, + 0.013497211039066315, + 0.877912163734436, + 0.028878340497612953 + ], + "credit_map": [ + 0.0018961301539093256, + 0.0008441200479865074, + 0.0011275153374299407, + 0.0013639423996210098, + 0.0007418236345984042, + 0.01372500415891409, + 0.001282963901758194, + 0.012519698590040207, + 0.0006211176514625549, + 0.00047110405284911394, + 0.0008367787231691182, + 0.0010273901280015707, + 0.0007881632191129029, + 0.0007537486962974072, + 0.001275260467082262, + 0.0004918075865134597, + 0.0007416409789584577, + 0.0004253624938428402, + 0.0009910293156281114, + 0.0007544565596617758, + 0.0006287865689955652, + 0.0011673177359625697, + 0.0010420450707897544, + 0.0005588250933215022, + 0.0005106489988975227, + 0.0002543476002756506, + 0.0006981898914091289, + 0.0004455886082723737, + 0.00029098245431669056, + 0.0009219344356097281, + 0.0005053661298006773, + 0.00041220802813768387, + 0.001139288768172264, + 0.0009540400933474302, + 0.0009592178394086659, + 0.0010920921340584755, + 0.0009908818174153566, + 0.0011949571780860424, + 0.0009388513281010091, + 0.0007670497288927436, + 0.001390673452988267, + 0.0012812174391001463, + 0.0008188956417143345, + 0.0025532443542033434, + 0.001247247215360403, + 0.0016369536751881242, + 0.0013002530904486775, + 0.0012291731545701623, + 0.0008052767952904105, + 0.001353911473415792, + 0.0014977441169321537, + 0.0016187896253541112, + 0.000751345360185951, + 0.001591599895618856, + 0.0008805386023595929, + 0.001603752956725657, + 0.0011050158645957708, + 0.001496814307756722, + 0.0014119433471933007, + 0.0016085186507552862, + 0.0015926148043945432, + 0.002958484459668398, + 0.0012911733938381076, + 0.00203264644369483, + 0.0852120965719223, + 0.08694455027580261, + 0.09117110073566437, + 0.13678032159805298, + 0.08684060722589493, + 0.1543293297290802, + 0.14978814125061035, + 0.0868460014462471, + 0.003910131752490997, + 0.0027590696699917316, + 0.0031356094405055046, + 0.0035284587647765875, + 0.0031244901474565268, + 0.003996359184384346, + 0.003491754177957773, + 0.004932471551001072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.03809557855129242 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_10/index_tensor([5582, 5582])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_10/index_tensor([5582, 5582])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..368a204152a7a77fc02abe88b111c80a9221abb6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_10/index_tensor([5582, 5582])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, high-angle panoramic shot, as if executed by a drone, smoothly parallels a sleek intercity express train as it smoothly traverses a vast metropolis at dusk. The train's slender, articulated body, constructed of a dark, polished metal alloy that reflects the fading ambient light, appears to undulate gracefully in a powerful, serpentine pattern along a vast, elevated concrete track between towering skyscrapers. The cityscape is a dense forest of modern buildings, their concrete and glass facades a soft, non-reflective gray. In the distance, the streets are dark canyons, streaked only by the faint, creeping headlights and taillights of cars. The entire scene is captured in a deliberately desaturated palette: the cool, dominant gray of the city buildings and the deep blue of the dusk sky create a stark, dramatic contrast with the train's continuous, glowing rectangular windows, which emit a warm golden glow, and the scattered, sharp, electronic city lights. This visual dichotomy creates a powerful atmosphere of cool, cinematic grandeur, emphasizing the immense scale of the city and the relentless, elegant movement of technology within it.", + "seed": 17071, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_10/index_5582_seed_17071_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05946612781845033, + "vlm_logprob": -0.05946612781845033, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.05946612781845033, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025837738066911697 + ], + "opposite_logprob_tokens": [ + -4.0258378982543945 + ], + "logprob_reward": -0.025837738066911697, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.025837738066911697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3737643361091614, + 0.07236449420452118, + 0.04313753917813301, + 0.10536511987447739, + 0.08431297540664673, + 0.07665190100669861, + 0.05858132615685463, + 0.07773792743682861, + 0.06295238435268402, + 0.045131999999284744 + ], + "credit_map": [ + 0.09344802796840668, + 0.07492347806692123, + 0.06390166282653809, + 0.022341979667544365, + 0.042113978415727615, + 0.035297948867082596, + 0.02419598028063774, + 0.017541268840432167, + 0.008407321758568287, + 0.007003425154834986, + 0.0063246628269553185, + 0.0076317014172673225, + 0.011636406183242798, + 0.01228906400501728, + 0.009360161609947681, + 0.009711739607155323, + 0.005609404295682907, + 0.005672540981322527, + 0.003791438415646553, + 0.005427585914731026, + 0.004360554274171591, + 0.008829518221318722, + 0.004516797140240669, + 0.004929693881422281, + 0.014608431607484818, + 0.012801826000213623, + 0.012941263616085052, + 0.01663530245423317, + 0.010874913074076176, + 0.01101890578866005, + 0.012351904064416885, + 0.01413255650550127, + 0.010983995161950588, + 0.021297400817275047, + 0.007862967438995838, + 0.009260850958526134, + 0.017593104392290115, + 0.007354709319770336, + 0.004199228249490261, + 0.005760711617767811, + 0.010830794461071491, + 0.009283850900828838, + 0.005570912268012762, + 0.009961055591702461, + 0.009834877215325832, + 0.008667540736496449, + 0.010221205651760101, + 0.01228165440261364, + 0.008645192719995975, + 0.007906819693744183, + 0.0053461515344679356, + 0.0065383403562009335, + 0.012703506276011467, + 0.007554885931313038, + 0.005814237520098686, + 0.004072187934070826, + 0.01595449075102806, + 0.009772592224180698, + 0.006182176060974598, + 0.007203015498816967, + 0.013707342557609081, + 0.006895947270095348, + 0.009927023202180862, + 0.008095333352684975, + 0.009280133992433548, + 0.007604438345879316, + 0.004867890849709511, + 0.009667322970926762, + 0.00703743239864707, + 0.011292744427919388, + 0.005780446343123913, + 0.007421966642141342, + 0.009832482784986496, + 0.007552332244813442, + 0.0038257776759564877, + 0.0037047078367322683, + 0.005461740307509899, + 0.0068397605791687965, + 0.003563342150300741, + 0.004351852927356958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.025837738066911697 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004657254554331303 + ], + "opposite_logprob_tokens": [ + -5.37965726852417 + ], + "logprob_reward": -0.004657254554331303, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004657254554331303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29444554448127747, + 0.08116479963064194, + 0.03169144690036774, + 0.15391358733177185, + 0.036536309868097305, + 0.06693429499864578, + 0.06212391331791878, + 0.12914396822452545, + 0.0754888579249382, + 0.0685572475194931 + ], + "credit_map": [ + 0.061483096331357956, + 0.06856434047222137, + 0.02225416526198387, + 0.03511667996644974, + 0.030159683898091316, + 0.0240537840873003, + 0.03561454266309738, + 0.017199287191033363, + 0.014361354522407055, + 0.008352387696504593, + 0.0075634620152413845, + 0.01361566036939621, + 0.010252823121845722, + 0.010354075580835342, + 0.007667981553822756, + 0.008997047320008278, + 0.003840053454041481, + 0.00435858778655529, + 0.0028755643870681524, + 0.003931388258934021, + 0.00355731975287199, + 0.00568722328171134, + 0.0030080347787588835, + 0.0044332765974104404, + 0.01850246451795101, + 0.036982569843530655, + 0.017372991889715195, + 0.018511760979890823, + 0.017498891800642014, + 0.016303695738315582, + 0.013844078406691551, + 0.014897143468260765, + 0.0034741403069347143, + 0.006540424656122923, + 0.0029204702004790306, + 0.003739879233762622, + 0.0059052868746221066, + 0.0062642344273626804, + 0.004119734279811382, + 0.003572136629372835, + 0.00801981333643198, + 0.010842601768672466, + 0.006855875253677368, + 0.0056139701046049595, + 0.00701647624373436, + 0.01161985658109188, + 0.008941631764173508, + 0.00802407506853342, + 0.006682025734335184, + 0.007735460065305233, + 0.005363671109080315, + 0.007043065503239632, + 0.01968712918460369, + 0.005853781942278147, + 0.004681452643126249, + 0.005077327135950327, + 0.02497020736336708, + 0.01595895364880562, + 0.011012177914381027, + 0.014506261795759201, + 0.01889023184776306, + 0.020027151331305504, + 0.012051421217620373, + 0.011727561242878437, + 0.013595464639365673, + 0.008770597167313099, + 0.00815604254603386, + 0.01187399122864008, + 0.009366722777485847, + 0.00987319741398096, + 0.005696723237633705, + 0.008156110532581806, + 0.012905596755445004, + 0.014838830567896366, + 0.006970185786485672, + 0.006722254678606987, + 0.007231926545500755, + 0.008079512044787407, + 0.005410314071923494, + 0.006398625671863556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.004657254554331303 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549126520752907 + ], + "opposite_logprob_tokens": [ + -4.387548923492432 + ], + "logprob_reward": -0.012549126520752907, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549126520752907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30242660641670227, + 0.11302149295806885, + 0.03877529874444008, + 0.09032361209392548, + 0.060527097433805466, + 0.07861149311065674, + 0.07116033881902695, + 0.08766471594572067, + 0.044877149164676666, + 0.11261224746704102 + ], + "credit_map": [ + 0.030870314687490463, + 0.047430068254470825, + 0.06438053399324417, + 0.048807162791490555, + 0.023933300748467445, + 0.03005051426589489, + 0.03256653994321823, + 0.02438816986978054, + 0.0185423344373703, + 0.012906903401017189, + 0.010462166741490364, + 0.016620337963104248, + 0.015897542238235474, + 0.016450537368655205, + 0.011087368242442608, + 0.011054295115172863, + 0.0055211507715284824, + 0.004014895297586918, + 0.004108130000531673, + 0.0036752570886164904, + 0.004870488774031401, + 0.008436023257672787, + 0.003145563416182995, + 0.005003787111490965, + 0.00837266817688942, + 0.009802361018955708, + 0.012849589809775352, + 0.014001218602061272, + 0.00962908286601305, + 0.008519158698618412, + 0.012710284441709518, + 0.014439242891967297, + 0.007019973825663328, + 0.011199216358363628, + 0.0042688618414103985, + 0.006174716632813215, + 0.013549372553825378, + 0.00674347672611475, + 0.0061724260449409485, + 0.005399049259722233, + 0.011795328930020332, + 0.011269129812717438, + 0.007989012636244297, + 0.005801008082926273, + 0.011369435116648674, + 0.012172715738415718, + 0.011110753752291203, + 0.007104103919118643, + 0.009309608489274979, + 0.009040466509759426, + 0.006452909670770168, + 0.006570923607796431, + 0.010980161838233471, + 0.011080893687903881, + 0.007953867316246033, + 0.009771501645445824, + 0.018512096256017685, + 0.012571626342833042, + 0.00849535595625639, + 0.007836068980395794, + 0.01356023084372282, + 0.010476769879460335, + 0.007117628585547209, + 0.009094935841858387, + 0.006136790849268436, + 0.006262193899601698, + 0.003942570183426142, + 0.009174026548862457, + 0.005096818320453167, + 0.005412723403424025, + 0.003812864189967513, + 0.005039158742874861, + 0.0363578237593174, + 0.01594500243663788, + 0.00411852914839983, + 0.004143975209444761, + 0.03371216729283333, + 0.007656655739992857, + 0.00319458544254303, + 0.007483499590307474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.012549126520752907 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520865499973297 + ], + "opposite_logprob_tokens": [ + -1.5020865201950073 + ], + "logprob_reward": -0.2520865499973297, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520865499973297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2177763432264328, + 0.12503521144390106, + 0.03349343314766884, + 0.05427613481879234, + 0.05851919203996658, + 0.11012114584445953, + 0.12897290289402008, + 0.12484366446733475, + 0.04499819874763489, + 0.10196378827095032 + ], + "credit_map": [ + 0.04012254625558853, + 0.046109940856695175, + 0.02151663601398468, + 0.016809405758976936, + 0.0304114893078804, + 0.03180227801203728, + 0.01783457025885582, + 0.013169487938284874, + 0.02581016533076763, + 0.013446880504488945, + 0.008880086243152618, + 0.011496773920953274, + 0.02153112180531025, + 0.011195285245776176, + 0.012125636450946331, + 0.020549260079860687, + 0.004968902561813593, + 0.0035503823310136795, + 0.004013908561319113, + 0.0027458833064883947, + 0.005145653150975704, + 0.006791710387915373, + 0.002990484004840255, + 0.003286508144810796, + 0.004604358691722155, + 0.0066852509044110775, + 0.007464217953383923, + 0.009864557534456253, + 0.005961038637906313, + 0.005196218378841877, + 0.006231056060642004, + 0.008269435726106167, + 0.005400843918323517, + 0.01508367620408535, + 0.005195511970669031, + 0.006281627342104912, + 0.006497523281723261, + 0.007988451980054379, + 0.007326027844101191, + 0.004745530430227518, + 0.01613725535571575, + 0.014958946034312248, + 0.011360518634319305, + 0.01055420096963644, + 0.014235994778573513, + 0.01655522920191288, + 0.012102148495614529, + 0.014216847717761993, + 0.017911000177264214, + 0.018117062747478485, + 0.013539223931729794, + 0.013420343399047852, + 0.024504078552126884, + 0.01680779457092285, + 0.013182966969907284, + 0.011490441858768463, + 0.018034478649497032, + 0.016144318506121635, + 0.010654238983988762, + 0.01804506964981556, + 0.016172967851161957, + 0.01812674105167389, + 0.011063865385949612, + 0.016601985320448875, + 0.005277654621750116, + 0.0069855800829827785, + 0.0033334908075630665, + 0.010319421999156475, + 0.005403339862823486, + 0.005372412037104368, + 0.004019617568701506, + 0.004286685027182102, + 0.03669002279639244, + 0.017866360023617744, + 0.007405232172459364, + 0.004623305052518845, + 0.01878291927278042, + 0.007786930538713932, + 0.003715323284268379, + 0.005093701649457216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.2520865499973297 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021999699529260397 + ], + "opposite_logprob_tokens": [ + -6.127200126647949 + ], + "logprob_reward": -0.0021999699529260397, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021999699529260397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15982143580913544, + 0.04608484357595444, + 0.04090297967195511, + 0.08633318543434143, + 0.062253136187791824, + 0.19243453443050385, + 0.15648630261421204, + 0.15953665971755981, + 0.020992884412407875, + 0.07515405118465424 + ], + "credit_map": [ + 0.03989909216761589, + 0.026712758466601372, + 0.02836887538433075, + 0.013152719475328922, + 0.014122157357633114, + 0.014363541267812252, + 0.013674386776983738, + 0.009527914226055145, + 0.007043268531560898, + 0.006356552708894014, + 0.0041837613098323345, + 0.007014668546617031, + 0.005897826980799437, + 0.0057596974074840546, + 0.004328853450715542, + 0.005500216037034988, + 0.006516322493553162, + 0.004751355387270451, + 0.004182010423392057, + 0.004418289754539728, + 0.005079240072518587, + 0.007296167314052582, + 0.0038448679260909557, + 0.004814728628844023, + 0.008770952001214027, + 0.016674829646945, + 0.009901614859700203, + 0.014270508661866188, + 0.009360473603010178, + 0.00992310605943203, + 0.00997455045580864, + 0.007457154802978039, + 0.006077190861105919, + 0.011597249656915665, + 0.006457440089434385, + 0.006286339834332466, + 0.00948993768543005, + 0.01407574862241745, + 0.002991157118231058, + 0.005278069991618395, + 0.02856028825044632, + 0.028440356254577637, + 0.02013959176838398, + 0.019660884514451027, + 0.024519192054867744, + 0.020942935720086098, + 0.031064065173268318, + 0.019107215106487274, + 0.028414415195584297, + 0.024795878678560257, + 0.019691573455929756, + 0.016737574711441994, + 0.018905607983469963, + 0.021149542182683945, + 0.013762515038251877, + 0.01302918791770935, + 0.03409050777554512, + 0.0181420985609293, + 0.019326360896229744, + 0.018962213769555092, + 0.020176514983177185, + 0.01996104046702385, + 0.011829410679638386, + 0.017048511654138565, + 0.002767048077657819, + 0.0031489855609834194, + 0.0019632764160633087, + 0.003331660758703947, + 0.002535383915528655, + 0.0030767687130719423, + 0.0018013822846114635, + 0.0023683791514486074, + 0.013364478945732117, + 0.015061818063259125, + 0.004686263855546713, + 0.003781606676056981, + 0.026284055784344673, + 0.0052351257763803005, + 0.002991111483424902, + 0.0037495866417884827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.0021999699529260397 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, high-angle panoramic shot, as if executed by a drone, smoothly parallels a sleek intercity express train as it smoothly traverses a vast metropolis at dusk. The train's slender, articulated body, constructed of a dark, polished metal alloy that reflects the fading ambient light, appears to undulate gracefully in a powerful, serpentine pattern along a vast, elevated concrete track between towering skyscrapers. The cityscape is a dense forest of modern buildings, their concrete and glass facades a soft, non-reflective gray. In the distance, the streets are dark canyons, streaked only by the faint, creeping headlights and taillights of cars. The entire scene is captured in a deliberately desaturated palette: the cool, dominant gray of the city buildings and the deep blue of the dusk sky create a stark, dramatic contrast with the train's continuous, glowing rectangular windows, which emit a warm golden glow, and the scattered, sharp, electronic city lights. This visual dichotomy creates a powerful atmosphere of cool, cinematic grandeur, emphasizing the immense scale of the city and the relentless, elegant movement of technology within it.", + "seed": 17072, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_10/index_5582_seed_17072_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04857667274773121, + "vlm_logprob": -0.04857667274773121, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.04857667274773121, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.050533369183540344 + ], + "opposite_logprob_tokens": [ + -3.3005332946777344 + ], + "logprob_reward": -0.050533369183540344, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.050533369183540344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1446770578622818, + 0.07225455343723297, + 0.08113612234592438, + 0.09733513742685318, + 0.44615888595581055, + 0.0436684712767601, + 0.017119556665420532, + 0.03482808172702789, + 0.03424721583724022, + 0.028574904426932335 + ], + "credit_map": [ + 0.0201194416731596, + 0.016932152211666107, + 0.021357886493206024, + 0.02282518707215786, + 0.016203822568058968, + 0.014016346074640751, + 0.01975753903388977, + 0.013464671559631824, + 0.007462258916348219, + 0.0059316339902579784, + 0.012586350552737713, + 0.011592603288590908, + 0.011457006447017193, + 0.011121674440801144, + 0.00670374371111393, + 0.0053992802277207375, + 0.014342598617076874, + 0.007026802748441696, + 0.01174353901296854, + 0.011452553793787956, + 0.006969434674829245, + 0.008449248038232327, + 0.004879662301391363, + 0.01627228409051895, + 0.013770008459687233, + 0.01123952865600586, + 0.02193554863333702, + 0.007991689257323742, + 0.008449766784906387, + 0.01236586831510067, + 0.006984406150877476, + 0.014598322100937366, + 0.10905759781599045, + 0.02954043075442314, + 0.02367984503507614, + 0.015575183555483818, + 0.030196072533726692, + 0.06058325618505478, + 0.021958794444799423, + 0.1555677205324173, + 0.004840197041630745, + 0.005076729226857424, + 0.005332318600267172, + 0.004330685827881098, + 0.007294207811355591, + 0.004705527797341347, + 0.0066698286682367325, + 0.005418976303189993, + 0.002540176035836339, + 0.0021155516151338816, + 0.0014517871895805001, + 0.0010416613658890128, + 0.002055481309071183, + 0.0038644683081656694, + 0.002022639149799943, + 0.002027791691944003, + 0.0035878638736903667, + 0.005683841649442911, + 0.0027995433192700148, + 0.0032495418563485146, + 0.004671876318752766, + 0.003979090601205826, + 0.006788369733840227, + 0.0040679569356143475, + 0.005486651323735714, + 0.005108875222504139, + 0.003246741136536002, + 0.0029622914735227823, + 0.00394503865391016, + 0.005288620013743639, + 0.0037426610942929983, + 0.004466336220502853, + 0.004654410295188427, + 0.002578323008492589, + 0.0017568101175129414, + 0.0019666466396301985, + 0.0035336927976459265, + 0.006575326435267925, + 0.003130007768049836, + 0.004379688296467066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.050533369183540344 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020612552762031555 + ], + "opposite_logprob_tokens": [ + -3.8956124782562256 + ], + "logprob_reward": -0.020612552762031555, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020612552762031555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09461207687854767, + 0.05531580373644829, + 0.08605355769395828, + 0.054732546210289, + 0.5096283555030823, + 0.0334915965795517, + 0.019406285136938095, + 0.04495583847165108, + 0.06507916748523712, + 0.03672485798597336 + ], + "credit_map": [ + 0.014413153752684593, + 0.015300766564905643, + 0.012623355723917484, + 0.013543603010475636, + 0.009124460630118847, + 0.009200923144817352, + 0.011139770038425922, + 0.009266039356589317, + 0.005958298221230507, + 0.006093725096434355, + 0.007210525218397379, + 0.00679204985499382, + 0.012842529453337193, + 0.006463134195655584, + 0.006021020468324423, + 0.003934516105800867, + 0.013315527699887753, + 0.008448939770460129, + 0.01311239693313837, + 0.011508607305586338, + 0.007949153892695904, + 0.008866342715919018, + 0.005588614381849766, + 0.01726396381855011, + 0.009061294607818127, + 0.005190536379814148, + 0.009610230103135109, + 0.004668319132179022, + 0.004904760513454676, + 0.006552209611982107, + 0.005381647497415543, + 0.00936354510486126, + 0.12357574701309204, + 0.0323498360812664, + 0.0245427954941988, + 0.01768035814166069, + 0.03345499187707901, + 0.07222796976566315, + 0.021425390616059303, + 0.184371218085289, + 0.004315901082009077, + 0.004095951095223427, + 0.002949520479887724, + 0.002568164374679327, + 0.005094168707728386, + 0.0038042194209992886, + 0.006605967879295349, + 0.004057700280100107, + 0.0026735966093838215, + 0.0024008408654481173, + 0.0018829844193533063, + 0.0011228184448555112, + 0.0030133503023535013, + 0.003592174034565687, + 0.0023685013875365257, + 0.0023520183749496937, + 0.004648677539080381, + 0.006526859477162361, + 0.0042142788879573345, + 0.005511957220733166, + 0.005282440222799778, + 0.004464562050998211, + 0.008788174949586391, + 0.005518882069736719, + 0.011576613411307335, + 0.007707573473453522, + 0.0071173361502587795, + 0.0040103900246322155, + 0.0073720295913517475, + 0.013050209730863571, + 0.0057627665810287, + 0.008482244797050953, + 0.004338673781603575, + 0.004764507990330458, + 0.003210684284567833, + 0.0033189677633345127, + 0.004240705631673336, + 0.006803913041949272, + 0.00491422601044178, + 0.005133176688104868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.020612552762031555 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890849381685257 + ], + "opposite_logprob_tokens": [ + -2.4640848636627197 + ], + "logprob_reward": -0.0890849381685257, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890849381685257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07410456985235214, + 0.07943717390298843, + 0.07613907754421234, + 0.04187074303627014, + 0.5435229539871216, + 0.029133418574929237, + 0.028132757171988487, + 0.08200806379318237, + 0.026940736919641495, + 0.018710484728217125 + ], + "credit_map": [ + 0.011542894877493382, + 0.008921191096305847, + 0.012555527500808239, + 0.00959675945341587, + 0.0075263939797878265, + 0.006547571159899235, + 0.009281836450099945, + 0.008132400922477245, + 0.006188858766108751, + 0.006002327427268028, + 0.016247298568487167, + 0.013278104364871979, + 0.008762737736105919, + 0.01701590232551098, + 0.0069348737597465515, + 0.005007070489227772, + 0.010785609483718872, + 0.006265508010983467, + 0.013110654428601265, + 0.014064110815525055, + 0.0053955670446157455, + 0.006199606694281101, + 0.006458778399974108, + 0.01385924406349659, + 0.0065441676415503025, + 0.0036910639610141516, + 0.010034811682999134, + 0.004443406127393246, + 0.00393077963963151, + 0.0038564717397093773, + 0.00311899627558887, + 0.006251049228012562, + 0.13091319799423218, + 0.033649180084466934, + 0.026024678722023964, + 0.018504749983549118, + 0.03563394770026207, + 0.0773979052901268, + 0.02338840626180172, + 0.1980108767747879, + 0.003670488717034459, + 0.0039063370786607265, + 0.0027009828481823206, + 0.0025509546976536512, + 0.00530248461291194, + 0.0033319529611617327, + 0.004767832346260548, + 0.002902386011555791, + 0.004851026460528374, + 0.00415235897526145, + 0.0034058946184813976, + 0.001646157237701118, + 0.0036801109090447426, + 0.0049376958049833775, + 0.002516689710319042, + 0.002942822640761733, + 0.008076529018580914, + 0.012731783092021942, + 0.005411254707723856, + 0.006812776904553175, + 0.011059165932238102, + 0.007774642668664455, + 0.019190356135368347, + 0.010951554402709007, + 0.005212970543652773, + 0.0037143337540328503, + 0.0024048704653978348, + 0.001765987602993846, + 0.0031566335819661617, + 0.003787847701460123, + 0.003236018354073167, + 0.0036620742175728083, + 0.00270852935500443, + 0.002526892814785242, + 0.0014930980978533626, + 0.0017551915952935815, + 0.002043135929852724, + 0.0029581685084849596, + 0.0025765409227460623, + 0.002648927504196763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.0890849381685257 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062060385942459106 + ], + "opposite_logprob_tokens": [ + -2.8120603561401367 + ], + "logprob_reward": -0.062060385942459106, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062060385942459106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11540766060352325, + 0.19283784925937653, + 0.20086447894573212, + 0.0777529776096344, + 0.15975028276443481, + 0.03358383849263191, + 0.06875506043434143, + 0.08210349082946777, + 0.04960507154464722, + 0.01933930441737175 + ], + "credit_map": [ + 0.008654767647385597, + 0.007941978052258492, + 0.01803618296980858, + 0.019835833460092545, + 0.012952206656336784, + 0.011331693269312382, + 0.02623399719595909, + 0.01042100414633751, + 0.0099330460652709, + 0.012736777774989605, + 0.043437618762254715, + 0.023122461512684822, + 0.02057735063135624, + 0.04358997195959091, + 0.017115259543061256, + 0.02232534997165203, + 0.023407703265547752, + 0.012023613788187504, + 0.030332446098327637, + 0.03381560742855072, + 0.018931996077299118, + 0.02099873498082161, + 0.025401022285223007, + 0.035953350365161896, + 0.012152910232543945, + 0.0071758124977350235, + 0.018321756273508072, + 0.008754701353609562, + 0.00752409640699625, + 0.008382963016629219, + 0.005150543991476297, + 0.010290199890732765, + 0.03745976462960243, + 0.011000550352036953, + 0.008150778710842133, + 0.00494422996416688, + 0.01504430826753378, + 0.021694757044315338, + 0.008508059196174145, + 0.05294783040881157, + 0.005645067896693945, + 0.004212595988065004, + 0.004414629191160202, + 0.0025658903177827597, + 0.006109964568167925, + 0.003374493680894375, + 0.003713814076036215, + 0.003547383239492774, + 0.013387464918196201, + 0.010354213416576385, + 0.005649954080581665, + 0.004005156923085451, + 0.007356489077210426, + 0.011494722217321396, + 0.008282133378088474, + 0.008224926888942719, + 0.008033791556954384, + 0.01920439675450325, + 0.006177844479680061, + 0.008448745124042034, + 0.00992585439234972, + 0.008834704756736755, + 0.012830407358705997, + 0.008647744543850422, + 0.00872381404042244, + 0.005364630836993456, + 0.0060228267684578896, + 0.004279353655874729, + 0.005902990698814392, + 0.008824203163385391, + 0.004677030723541975, + 0.005810219328850508, + 0.0030288577545434237, + 0.001785485539585352, + 0.0016631975304335356, + 0.0018548232037574053, + 0.0032323456835001707, + 0.0022955285385251045, + 0.0023671910166740417, + 0.003111874684691429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.062060385942459106 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592117682099342 + ], + "opposite_logprob_tokens": [ + -3.895592212677002 + ], + "logprob_reward": -0.020592117682099342, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592117682099342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19087781012058258, + 0.12262754887342453, + 0.09088768064975739, + 0.08065401017665863, + 0.2962983250617981, + 0.029815414920449257, + 0.04416598752140999, + 0.08368291705846786, + 0.03469313308596611, + 0.026297232136130333 + ], + "credit_map": [ + 0.0268535278737545, + 0.02008073776960373, + 0.030810529366135597, + 0.02923196367919445, + 0.01708068884909153, + 0.01695903390645981, + 0.0274294875562191, + 0.02243182808160782, + 0.012996437028050423, + 0.010161039419472218, + 0.024132203310728073, + 0.014403578825294971, + 0.01754620485007763, + 0.029402753338217735, + 0.008393721655011177, + 0.00559161975979805, + 0.011416741646826267, + 0.008273907005786896, + 0.014935550279915333, + 0.015044488944113255, + 0.008930770680308342, + 0.008757263422012329, + 0.006161143071949482, + 0.017367815598845482, + 0.012913198210299015, + 0.007928199134767056, + 0.013523779809474945, + 0.006784065626561642, + 0.007851382717490196, + 0.010900452733039856, + 0.008138442412018776, + 0.012614483945071697, + 0.07087439298629761, + 0.01891091652214527, + 0.01419270969927311, + 0.010558542795479298, + 0.019643230363726616, + 0.042127370834350586, + 0.012544602155685425, + 0.10744655877351761, + 0.003913955297321081, + 0.0031409645453095436, + 0.003808438079431653, + 0.001983512891456485, + 0.004407341592013836, + 0.0036859307438135147, + 0.004693619906902313, + 0.0041816518642008305, + 0.007589469198137522, + 0.005945876706391573, + 0.004448249936103821, + 0.00346020027063787, + 0.004834464751183987, + 0.00782825518399477, + 0.0050316182896494865, + 0.005027851089835167, + 0.009521410800516605, + 0.010069770738482475, + 0.00989010464400053, + 0.008021263405680656, + 0.011654756031930447, + 0.008157121948897839, + 0.01609904132783413, + 0.01026945374906063, + 0.005608811508864164, + 0.005469521041959524, + 0.0031774311792105436, + 0.0018393386853858829, + 0.004530407954007387, + 0.005466101225465536, + 0.004109855741262436, + 0.004491661675274372, + 0.004443814978003502, + 0.003110946621745825, + 0.002129490254446864, + 0.002555346116423607, + 0.0029346083756536245, + 0.004543506074696779, + 0.003164483467116952, + 0.0034150369465351105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.020592117682099342 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, high-angle panoramic shot, as if executed by a drone, smoothly parallels a sleek intercity express train as it smoothly traverses a vast metropolis at dusk. The train's slender, articulated body, constructed of a dark, polished metal alloy that reflects the fading ambient light, appears to undulate gracefully in a powerful, serpentine pattern along a vast, elevated concrete track between towering skyscrapers. The cityscape is a dense forest of modern buildings, their concrete and glass facades a soft, non-reflective gray. In the distance, the streets are dark canyons, streaked only by the faint, creeping headlights and taillights of cars. The entire scene is captured in a deliberately desaturated palette: the cool, dominant gray of the city buildings and the deep blue of the dusk sky create a stark, dramatic contrast with the train's continuous, glowing rectangular windows, which emit a warm golden glow, and the scattered, sharp, electronic city lights. This visual dichotomy creates a powerful atmosphere of cool, cinematic grandeur, emphasizing the immense scale of the city and the relentless, elegant movement of technology within it.", + "seed": 17073, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_10/index_5582_seed_17073_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04154482427984476, + "vlm_logprob": -0.04154482427984476, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.04154482427984476, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0388469323515892 + ], + "opposite_logprob_tokens": [ + -3.663846969604492 + ], + "logprob_reward": -0.0388469323515892, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0388469323515892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3458593487739563, + 0.06828723847866058, + 0.10063718259334564, + 0.05422401800751686, + 0.07730340212583542, + 0.03218470513820648, + 0.08505405485630035, + 0.09593817591667175, + 0.060599032789468765, + 0.07991284877061844 + ], + "credit_map": [ + 0.0341116301715374, + 0.03504030406475067, + 0.07615853101015091, + 0.05839886888861656, + 0.029811052605509758, + 0.039526574313640594, + 0.03528531268239021, + 0.03752704709768295, + 0.007114586886018515, + 0.007173403166234493, + 0.012178126722574234, + 0.00914028286933899, + 0.006687413901090622, + 0.011598977260291576, + 0.00616122130304575, + 0.008233225904405117, + 0.00872480496764183, + 0.009884078986942768, + 0.013679059222340584, + 0.008738613687455654, + 0.029410695657134056, + 0.011384519748389721, + 0.01147971861064434, + 0.007335681468248367, + 0.0040839118883013725, + 0.009797940030694008, + 0.00618856493383646, + 0.007747170981019735, + 0.0069482591934502125, + 0.006189489271491766, + 0.00923185609281063, + 0.0040368251502513885, + 0.005608766805380583, + 0.009556998498737812, + 0.012176941148936749, + 0.013884863816201687, + 0.011586123146116734, + 0.008810729719698429, + 0.0068768528290092945, + 0.008802127093076706, + 0.004579419270157814, + 0.004778564441949129, + 0.004748048260807991, + 0.004354571457952261, + 0.004366341046988964, + 0.0035930476151406765, + 0.002800465328618884, + 0.002964247250929475, + 0.010897275991737843, + 0.011514934711158276, + 0.009216311387717724, + 0.01187659427523613, + 0.01298802811652422, + 0.008689513429999352, + 0.013615936040878296, + 0.006255459506064653, + 0.01142710167914629, + 0.014053127728402615, + 0.007526251021772623, + 0.007190224714577198, + 0.0210323054343462, + 0.014465064741671085, + 0.012301065027713776, + 0.007943039759993553, + 0.007574369199573994, + 0.013228083960711956, + 0.005709921941161156, + 0.004649346694350243, + 0.007297946140170097, + 0.010399979539215565, + 0.0057142749428749084, + 0.0060251131653785706, + 0.011981585063040257, + 0.010873774066567421, + 0.0178818479180336, + 0.007091827690601349, + 0.014740750193595886, + 0.006457814015448093, + 0.004064240958541632, + 0.006821007933467627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.0388469323515892 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014225504361093044 + ], + "opposite_logprob_tokens": [ + -4.264225482940674 + ], + "logprob_reward": -0.014225504361093044, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014225504361093044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1776469647884369, + 0.18018124997615814, + 0.12050645053386688, + 0.06654241681098938, + 0.1110328733921051, + 0.04319621995091438, + 0.032558612525463104, + 0.06596960127353668, + 0.043321046978235245, + 0.15904457867145538 + ], + "credit_map": [ + 0.017991797998547554, + 0.01947777345776558, + 0.036608047783374786, + 0.020009897649288177, + 0.022700432687997818, + 0.023977462202310562, + 0.01350411307066679, + 0.02337745763361454, + 0.03836001455783844, + 0.010838668793439865, + 0.024735497310757637, + 0.014560839161276817, + 0.01318055298179388, + 0.04911570996046066, + 0.015081193298101425, + 0.014308782294392586, + 0.00910220853984356, + 0.011566408909857273, + 0.017721999436616898, + 0.014843901619315147, + 0.03609480708837509, + 0.012613475322723389, + 0.008031883276998997, + 0.010531767271459103, + 0.004660936072468758, + 0.006106634624302387, + 0.009846377186477184, + 0.012962393462657928, + 0.008595211431384087, + 0.004921757150441408, + 0.01299781259149313, + 0.0064512938261032104, + 0.007463195826858282, + 0.01411423273384571, + 0.01544665452092886, + 0.01971994712948799, + 0.01522183045744896, + 0.010122241452336311, + 0.012400275096297264, + 0.01654449664056301, + 0.0035131420008838177, + 0.0052278777584433556, + 0.008743122220039368, + 0.00898667424917221, + 0.004095855634659529, + 0.00424708379432559, + 0.004653284326195717, + 0.0037291799671947956, + 0.0037026223726570606, + 0.004928549751639366, + 0.003225798485800624, + 0.003143600421026349, + 0.006078167352825403, + 0.004281753674149513, + 0.003947040997445583, + 0.0032510787714272738, + 0.007935645058751106, + 0.008564474061131477, + 0.005904715042561293, + 0.005486220587044954, + 0.013464662246406078, + 0.009343279525637627, + 0.010108625516295433, + 0.005161980167031288, + 0.006080315914005041, + 0.008479953743517399, + 0.004027732647955418, + 0.006416938733309507, + 0.005030364729464054, + 0.005916900932788849, + 0.0031730227638036013, + 0.004195819143205881, + 0.01675209030508995, + 0.021194754168391228, + 0.030019136145710945, + 0.01862136833369732, + 0.028625695034861565, + 0.011732780374586582, + 0.012507026083767414, + 0.01959172636270523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.014225504361093044 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373267874121666 + ], + "opposite_logprob_tokens": [ + -3.4087326526641846 + ], + "logprob_reward": -0.03373267874121666, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373267874121666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12054110318422318, + 0.04076235368847847, + 0.1813318282365799, + 0.054820720106363297, + 0.48018303513526917, + 0.026546912267804146, + 0.01405785046517849, + 0.013248374685645103, + 0.018920322880148888, + 0.049587495625019073 + ], + "credit_map": [ + 0.010130438022315502, + 0.010880684480071068, + 0.028446100652217865, + 0.018798425793647766, + 0.013981790281832218, + 0.014966522343456745, + 0.01129511184990406, + 0.01204201951622963, + 0.004803868941962719, + 0.0035726972855627537, + 0.007023361045867205, + 0.004417419899255037, + 0.004354880191385746, + 0.007137162610888481, + 0.003942506387829781, + 0.005510454997420311, + 0.012345894239842892, + 0.019816122949123383, + 0.027810880914330482, + 0.019147958606481552, + 0.045725274831056595, + 0.01791880466043949, + 0.019224662333726883, + 0.019342223182320595, + 0.003548797918483615, + 0.007400832138955593, + 0.007771093863993883, + 0.008545377291738987, + 0.005095180124044418, + 0.004852800630033016, + 0.008797739632427692, + 0.008808892220258713, + 0.03174562379717827, + 0.0629219114780426, + 0.06975803524255753, + 0.09045137465000153, + 0.06914331763982773, + 0.05476944148540497, + 0.04984220489859581, + 0.05155109241604805, + 0.002813079860061407, + 0.0026188166812062263, + 0.004719917196780443, + 0.004810469225049019, + 0.00377823319286108, + 0.002845580456778407, + 0.002760607050731778, + 0.00220020511187613, + 0.001698994543403387, + 0.0015629892004653811, + 0.0025622285902500153, + 0.0018836393719539046, + 0.0011554127559065819, + 0.0010666177840903401, + 0.0022284029982984066, + 0.0018995642894878983, + 0.0013988305581733584, + 0.0014531408669427037, + 0.0016631106846034527, + 0.0015881723957136273, + 0.001162820146419108, + 0.0013941318029537797, + 0.0025803144089877605, + 0.0020078527741134167, + 0.0021065466571599245, + 0.002716686809435487, + 0.001565324142575264, + 0.0015643442748114467, + 0.0034653134644031525, + 0.004357615485787392, + 0.0013272662181407213, + 0.0018172244308516383, + 0.005841483362019062, + 0.006582815665751696, + 0.01204225979745388, + 0.006172562018036842, + 0.00979174766689539, + 0.0035420998465269804, + 0.0027177410665899515, + 0.0028967850375920534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.03373267874121666 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032875835895538 + ], + "opposite_logprob_tokens": [ + -2.3503286838531494 + ], + "logprob_reward": -0.10032875835895538, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032875835895538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2175411581993103, + 0.0471041277050972, + 0.09490497410297394, + 0.03895183280110359, + 0.11787416785955429, + 0.09327621757984161, + 0.04403591901063919, + 0.03611212596297264, + 0.20722416043281555, + 0.10297541320323944 + ], + "credit_map": [ + 0.02669856697320938, + 0.02147478051483631, + 0.03445159271359444, + 0.02840375527739525, + 0.02576325833797455, + 0.03197422996163368, + 0.023702172562479973, + 0.02507278509438038, + 0.005944793578237295, + 0.005037068389356136, + 0.007941716350615025, + 0.005004000384360552, + 0.006901057902723551, + 0.00664547411724925, + 0.004520158749073744, + 0.0051098535768687725, + 0.008592321537435055, + 0.010103088803589344, + 0.012798996642231941, + 0.007447495125234127, + 0.03025682643055916, + 0.009696047753095627, + 0.008039742708206177, + 0.007970442995429039, + 0.0044850558042526245, + 0.007236477918922901, + 0.004458162467926741, + 0.004378709476441145, + 0.004763816483318806, + 0.006992025766521692, + 0.0034483554773032665, + 0.0031892224214971066, + 0.007964516058564186, + 0.016048742458224297, + 0.0176712479442358, + 0.02054532989859581, + 0.01570315845310688, + 0.017858192324638367, + 0.010889433324337006, + 0.011193539015948772, + 0.014837377704679966, + 0.013253017328679562, + 0.009342830628156662, + 0.00946956966072321, + 0.01571851223707199, + 0.01389237493276596, + 0.009336071088910103, + 0.007426445838063955, + 0.0058852555230259895, + 0.008746500127017498, + 0.002568036550655961, + 0.003929863218218088, + 0.007445004768669605, + 0.006319723557680845, + 0.0034352063667029142, + 0.005706324707716703, + 0.005085762590169907, + 0.00561516871675849, + 0.00310098216868937, + 0.0026036968920379877, + 0.008505415171384811, + 0.0046951728872954845, + 0.0038968147709965706, + 0.0026091078761965036, + 0.02929231896996498, + 0.03887331485748291, + 0.017799504101276398, + 0.01889054849743843, + 0.02954799309372902, + 0.04374741017818451, + 0.014271019026637077, + 0.014802027493715286, + 0.01501633133739233, + 0.016643553972244263, + 0.0071727680042386055, + 0.00932143535465002, + 0.025681864470243454, + 0.008058594539761543, + 0.00921562872827053, + 0.011865229345858097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.10032875835895538 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590247586369514 + ], + "opposite_logprob_tokens": [ + -3.895590305328369 + ], + "logprob_reward": -0.020590247586369514, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590247586369514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3404403626918793, + 0.05154995247721672, + 0.05902574583888054, + 0.023666543886065483, + 0.10155073553323746, + 0.05715395137667656, + 0.06497827917337418, + 0.02877609245479107, + 0.17365851998329163, + 0.0991998016834259 + ], + "credit_map": [ + 0.033939581364393234, + 0.03219006583094597, + 0.05587497353553772, + 0.053138598799705505, + 0.037291768938302994, + 0.05659741535782814, + 0.037400610744953156, + 0.034007370471954346, + 0.009783961810171604, + 0.005249266512691975, + 0.005953454412519932, + 0.0036666716914623976, + 0.006647407077252865, + 0.014250158332288265, + 0.0024617405142635107, + 0.0035372916609048843, + 0.004665101878345013, + 0.006322638597339392, + 0.007382918614894152, + 0.004157650284469128, + 0.01857188157737255, + 0.0055291736498475075, + 0.0048307511024177074, + 0.007565631065517664, + 0.0021458836272358894, + 0.005070401821285486, + 0.0025987166445702314, + 0.003186269896104932, + 0.0025670973118394613, + 0.003816155018284917, + 0.002574196783825755, + 0.001707822666503489, + 0.007366007659584284, + 0.01301665510982275, + 0.014546262100338936, + 0.015406343154609203, + 0.014369337819516659, + 0.014714332297444344, + 0.012004622258245945, + 0.010127182118594646, + 0.0061821346171200275, + 0.00870468094944954, + 0.00822244118899107, + 0.007119821850210428, + 0.007470287382602692, + 0.008274826221168041, + 0.006387310102581978, + 0.004792449530214071, + 0.008528302423655987, + 0.012217486277222633, + 0.009650086984038353, + 0.00780628714710474, + 0.006395155098289251, + 0.006106170359998941, + 0.009138599969446659, + 0.005136189516633749, + 0.003653316991403699, + 0.004191586282104254, + 0.003724839771166444, + 0.0022508620750159025, + 0.005118354689329863, + 0.004570810589939356, + 0.002929910086095333, + 0.0023364124353975058, + 0.024355223402380943, + 0.036091141402721405, + 0.013528093695640564, + 0.014200392179191113, + 0.029094651341438293, + 0.0331336185336113, + 0.010277044028043747, + 0.012978354468941689, + 0.009122522547841072, + 0.013430197723209858, + 0.014802574180066586, + 0.015452754683792591, + 0.01609622687101364, + 0.008001364767551422, + 0.009324809536337852, + 0.01296935323625803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.020590247586369514 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, high-angle panoramic shot, as if executed by a drone, smoothly parallels a sleek intercity express train as it smoothly traverses a vast metropolis at dusk. The train's slender, articulated body, constructed of a dark, polished metal alloy that reflects the fading ambient light, appears to undulate gracefully in a powerful, serpentine pattern along a vast, elevated concrete track between towering skyscrapers. The cityscape is a dense forest of modern buildings, their concrete and glass facades a soft, non-reflective gray. In the distance, the streets are dark canyons, streaked only by the faint, creeping headlights and taillights of cars. The entire scene is captured in a deliberately desaturated palette: the cool, dominant gray of the city buildings and the deep blue of the dusk sky create a stark, dramatic contrast with the train's continuous, glowing rectangular windows, which emit a warm golden glow, and the scattered, sharp, electronic city lights. This visual dichotomy creates a powerful atmosphere of cool, cinematic grandeur, emphasizing the immense scale of the city and the relentless, elegant movement of technology within it.", + "seed": 17074, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_10/index_5582_seed_17074_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08107380764558911, + "vlm_logprob": -0.08107380764558911, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.08107380764558911, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034714169800281525 + ], + "opposite_logprob_tokens": [ + -3.9097142219543457 + ], + "logprob_reward": -0.034714169800281525, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.034714169800281525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29200616478919983, + 0.03721345216035843, + 0.02196744829416275, + 0.046417396515607834, + 0.07517354935407639, + 0.07283101230859756, + 0.27343055605888367, + 0.024463685229420662, + 0.11065443605184555, + 0.04584241658449173 + ], + "credit_map": [ + 0.021839985623955727, + 0.01395221147686243, + 0.013098000548779964, + 0.013575587421655655, + 0.01812896691262722, + 0.05242554098367691, + 0.10472185909748077, + 0.05426399037241936, + 0.006170061882585287, + 0.004676596261560917, + 0.0030749416910111904, + 0.004156387876719236, + 0.005114739760756493, + 0.0047925920225679874, + 0.004798878915607929, + 0.004429252818226814, + 0.0020639896392822266, + 0.002226694254204631, + 0.003118323627859354, + 0.002582443645223975, + 0.0025297056417912245, + 0.003037884831428528, + 0.004121200647205114, + 0.0022872029803693295, + 0.003596341935917735, + 0.003478195983916521, + 0.008670915849506855, + 0.003498677397146821, + 0.0036006190348416567, + 0.006962039042264223, + 0.0047414712607860565, + 0.011869128793478012, + 0.007732597179710865, + 0.00604708818718791, + 0.012521034106612206, + 0.0161487627774477, + 0.007760023698210716, + 0.00818555522710085, + 0.009734590537846088, + 0.007043893449008465, + 0.008345779031515121, + 0.013926913030445576, + 0.010726350359618664, + 0.009240154176950455, + 0.007533225696533918, + 0.01061977632343769, + 0.006430842913687229, + 0.006007961928844452, + 0.022713929414749146, + 0.0214957557618618, + 0.013369778171181679, + 0.07313663512468338, + 0.016265226528048515, + 0.0898529514670372, + 0.01106793899089098, + 0.025528298690915108, + 0.0033973476383835077, + 0.0029830678831785917, + 0.00430970499292016, + 0.003027204191312194, + 0.003143744543194771, + 0.00305604818277061, + 0.001525825122371316, + 0.0030207394156605005, + 0.028321251273155212, + 0.01636507175862789, + 0.03325171768665314, + 0.006897462531924248, + 0.006632981356233358, + 0.0056139882653951645, + 0.002939088735729456, + 0.010632863268256187, + 0.00597825413569808, + 0.0052605848759412766, + 0.00912167876958847, + 0.004391767084598541, + 0.006444756407290697, + 0.004424505867063999, + 0.004994496237486601, + 0.005226368550211191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.034714169800281525 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018209077417850494 + ], + "opposite_logprob_tokens": [ + -4.018208980560303 + ], + "logprob_reward": -0.018209077417850494, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018209077417850494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14508675038814545, + 0.01610296033322811, + 0.014843257144093513, + 0.026924557983875275, + 0.03127537667751312, + 0.1956935077905655, + 0.35076016187667847, + 0.05567055195569992, + 0.1323842853307724, + 0.03125859051942825 + ], + "credit_map": [ + 0.008773911744356155, + 0.00651257298886776, + 0.004699813202023506, + 0.007231554016470909, + 0.008222538977861404, + 0.023853110149502754, + 0.05456409975886345, + 0.031229138374328613, + 0.002965000458061695, + 0.001961305970326066, + 0.001845993218012154, + 0.0013349262299016118, + 0.0019229419995099306, + 0.0020153801888227463, + 0.002334230812266469, + 0.001723182969726622, + 0.0014098865212872624, + 0.0013784359907731414, + 0.0018842702265828848, + 0.0014570074854418635, + 0.0026538195088505745, + 0.0019415105925872922, + 0.002476707799360156, + 0.001641618786379695, + 0.0022811954841017723, + 0.0021865724120289087, + 0.006720694247633219, + 0.003813396207988262, + 0.0030299597419798374, + 0.0028651864267885685, + 0.0030142413452267647, + 0.0030133125837892294, + 0.004065139219164848, + 0.002823371673002839, + 0.005950174760073423, + 0.004518051631748676, + 0.0036513665691018105, + 0.004839460365474224, + 0.0030092301312834024, + 0.002418581396341324, + 0.019592875614762306, + 0.05649543181061745, + 0.015651188790798187, + 0.01831248216331005, + 0.02080436982214451, + 0.04312589019536972, + 0.010852042585611343, + 0.010859213769435883, + 0.029716335237026215, + 0.02000572346150875, + 0.011813219636678696, + 0.09684453159570694, + 0.03672735020518303, + 0.12977901101112366, + 0.011948400177061558, + 0.013925605453550816, + 0.005037369672209024, + 0.006947431247681379, + 0.006911200471222401, + 0.0068945023231208324, + 0.011102134361863136, + 0.008048294112086296, + 0.0038906994741410017, + 0.006838918663561344, + 0.03208700194954872, + 0.012169589288532734, + 0.04695909097790718, + 0.00985419750213623, + 0.007300455588847399, + 0.007097688503563404, + 0.004362680949270725, + 0.012553580105304718, + 0.004569167271256447, + 0.0033268840052187443, + 0.00572314253076911, + 0.0018359832465648651, + 0.0032158500980585814, + 0.0029944577254354954, + 0.002306768437847495, + 0.007286339066922665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.018209077417850494 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372311219573021 + ], + "opposite_logprob_tokens": [ + -3.4087231159210205 + ], + "logprob_reward": -0.03372311219573021, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372311219573021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37547412514686584, + 0.031160075217485428, + 0.025343654677271843, + 0.11793025583028793, + 0.030688250437378883, + 0.042880892753601074, + 0.18570378422737122, + 0.047574516385793686, + 0.0648551881313324, + 0.07838932424783707 + ], + "credit_map": [ + 0.022916365414857864, + 0.01717442460358143, + 0.014462442137300968, + 0.01793428510427475, + 0.01952403411269188, + 0.06679855287075043, + 0.13662920892238617, + 0.0800347700715065, + 0.003355317749083042, + 0.006219178903847933, + 0.002808487741276622, + 0.002546776318922639, + 0.004170190542936325, + 0.00461990013718605, + 0.004642434883862734, + 0.00279778684489429, + 0.002415840746834874, + 0.0028726388700306416, + 0.0028005412314087152, + 0.0020730416290462017, + 0.004320468287914991, + 0.0036689855623990297, + 0.004721377044916153, + 0.0024707599077373743, + 0.009316666051745415, + 0.007975351996719837, + 0.03378821164369583, + 0.01851482503116131, + 0.008593657985329628, + 0.014660815708339214, + 0.012672561220824718, + 0.012408155016601086, + 0.0041049267165362835, + 0.0027509068604558706, + 0.005101889371871948, + 0.005172693636268377, + 0.003223758889362216, + 0.004385600797832012, + 0.00354786217212677, + 0.002400609664618969, + 0.008011126890778542, + 0.006241902709007263, + 0.0054880124516785145, + 0.00908080954104662, + 0.0036943675950169563, + 0.005090168211609125, + 0.0025526259560137987, + 0.0027218740433454514, + 0.01690368913114071, + 0.014282205142080784, + 0.009174957871437073, + 0.049974966794252396, + 0.012246991507709026, + 0.059115033596754074, + 0.007632175460457802, + 0.016373738646507263, + 0.006020667031407356, + 0.008612802252173424, + 0.0066568912006914616, + 0.0045167626813054085, + 0.007981176488101482, + 0.007381943985819817, + 0.0031729384791105986, + 0.003231327747926116, + 0.01470222044736147, + 0.015172571875154972, + 0.006767960265278816, + 0.003928415942937136, + 0.009862472303211689, + 0.00711470004171133, + 0.002710737520828843, + 0.004596109502017498, + 0.009739696979522705, + 0.009665804915130138, + 0.016680890694260597, + 0.004580947104841471, + 0.012149004265666008, + 0.008845984004437923, + 0.006945858243852854, + 0.009781131520867348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.03372311219573021 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134652078151703 + ], + "opposite_logprob_tokens": [ + -1.3134652376174927 + ], + "logprob_reward": -0.3134652078151703, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134652078151703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23524557054042816, + 0.02853836677968502, + 0.0336444154381752, + 0.059419553726911545, + 0.011822182685136795, + 0.03252970427274704, + 0.0989030972123146, + 0.04557555168867111, + 0.20464304089546204, + 0.2496785819530487 + ], + "credit_map": [ + 0.01481669582426548, + 0.008794724009931087, + 0.008048130199313164, + 0.010523405857384205, + 0.010485825128853321, + 0.03804255276918411, + 0.09241873770952225, + 0.05211549997329712, + 0.002545404015108943, + 0.003457516897469759, + 0.002878773957490921, + 0.0032829376868903637, + 0.0037543284706771374, + 0.0034208993893116713, + 0.006787397898733616, + 0.002411104505881667, + 0.0027275108732283115, + 0.0024911146610975266, + 0.0033993867691606283, + 0.0030841499101370573, + 0.005197593476623297, + 0.00500936433672905, + 0.008547828532755375, + 0.003187462454661727, + 0.0047269114293158054, + 0.005185888614505529, + 0.014347796328365803, + 0.007227551657706499, + 0.00531761022284627, + 0.0067469594068825245, + 0.0068596126511693, + 0.009007219225168228, + 0.0010719475103542209, + 0.0011694437125697732, + 0.0019371277885511518, + 0.001709508360363543, + 0.001865863217972219, + 0.0012696156045421958, + 0.0018046076875180006, + 0.0009940682211890817, + 0.0041117253713309765, + 0.006154348142445087, + 0.004735555499792099, + 0.0046516042202711105, + 0.003974623512476683, + 0.00467259157449007, + 0.002321347128599882, + 0.001907906262204051, + 0.011645685881376266, + 0.008230378851294518, + 0.005829852540045977, + 0.02361420728266239, + 0.008553668856620789, + 0.03391360118985176, + 0.003571645589545369, + 0.003544055623933673, + 0.0051229954697191715, + 0.007622557692229748, + 0.005170217715203762, + 0.004770967178046703, + 0.0077159772627055645, + 0.008021350018680096, + 0.0022375970147550106, + 0.00491388700902462, + 0.06522876769304276, + 0.02995995059609413, + 0.05520998686552048, + 0.011372752487659454, + 0.019617941230535507, + 0.010345926508307457, + 0.004937788005918264, + 0.007969909347593784, + 0.02687743306159973, + 0.019404439255595207, + 0.06245878338813782, + 0.01148369163274765, + 0.03986223042011261, + 0.02665027417242527, + 0.016419148072600365, + 0.046522561460733414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.3134652078151703 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle panoramic view taken from a drone-like perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly parallel a sleek intercity express train moving along an elevated concrete track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the train's body dark, polished metal alloy with glowing rectangular windows emitting a warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding cityscape composed of dense modern skyscrapers with non-reflective gray concrete and glass facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette deliberately desaturated with a cool dominant gray and deep blue sky contrasting the warm train lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525747099891305 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.00525747099891305, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525747099891305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10781443864107132, + 0.0478973425924778, + 0.05502290278673172, + 0.031460925936698914, + 0.02104438655078411, + 0.052248407155275345, + 0.19067087769508362, + 0.08945226669311523, + 0.23004239797592163, + 0.17434610426425934 + ], + "credit_map": [ + 0.012078797444701195, + 0.010127491317689419, + 0.009428475052118301, + 0.00921609252691269, + 0.013963554054498672, + 0.016044678166508675, + 0.016973799094557762, + 0.019981548190116882, + 0.006358780432492495, + 0.005968155339360237, + 0.004417012445628643, + 0.004285873845219612, + 0.0043507106602191925, + 0.004736654460430145, + 0.013868273235857487, + 0.003911882638931274, + 0.005323742050677538, + 0.005789151880890131, + 0.0047879149205982685, + 0.004392004106193781, + 0.008941511623561382, + 0.008009427227079868, + 0.012350828386843204, + 0.005428322125226259, + 0.0029234352987259626, + 0.0025075378362089396, + 0.006204079836606979, + 0.004108720924705267, + 0.0026745100039988756, + 0.0034553848672658205, + 0.006025225855410099, + 0.0035620289854705334, + 0.0020704108756035566, + 0.0020090185571461916, + 0.003353290492668748, + 0.0027779301162809134, + 0.0028932674322277308, + 0.0029062528628855944, + 0.0029369124677032232, + 0.0020973028149455786, + 0.006391310133039951, + 0.011481606401503086, + 0.00729766208678484, + 0.007564525585621595, + 0.006554631050676107, + 0.0071860565803945065, + 0.0030508481431752443, + 0.0027217671740800142, + 0.016587896272540092, + 0.0207492858171463, + 0.011292786337435246, + 0.04139962047338486, + 0.016679827123880386, + 0.05050934851169586, + 0.01036827266216278, + 0.023083848878741264, + 0.0125027010217309, + 0.012868516147136688, + 0.01218714751303196, + 0.008897708728909492, + 0.015713242813944817, + 0.014437069185078144, + 0.004858627449721098, + 0.00798726361244917, + 0.06687527149915695, + 0.0329742431640625, + 0.04169414937496185, + 0.012064991518855095, + 0.026684310287237167, + 0.017574507743120193, + 0.00992211140692234, + 0.022252818569540977, + 0.020062465220689774, + 0.014248130843043327, + 0.04401194304227829, + 0.008963948115706444, + 0.02679326757788658, + 0.01865527033805847, + 0.010391668416559696, + 0.031219404190778732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.00525747099891305 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_11/index_tensor([1433, 1433])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_11/index_tensor([1433, 1433])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4c0af5456ad582c434cf628b3022f7fed1c5f44c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_11/index_tensor([1433, 1433])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a static, aerial view established directly overhead, the camera executes a slow, almost imperceptible pan to follow the fluid movement of a lightweight silk scarf as it drifts across a polished, dark mahogany tabletop. The table's surface is impeccably smooth, its dark, rich veins reflecting soft, diffuse light from an unseen, off-camera source, creating a moody and contemplative canvas. The scarf itself is a masterpiece of exquisite fabric, dyed in ethereal, shifting shades of lavender and dusty rose, its edges woven with fine, shimmering silver threads. The scene comes alive when a gentle, unseen draft cautiously lifts a corner of the scarf; the fabric responds with slow, weightless grace, creating a series of silent, small waves that languidly spread across its surface. Each ripple catches the soft light, causing the silver threads to shimmer and the pastel colors to deepen and lighten in a mesmerizing dance, before the material deliberately lies flat again, its folds resting in a new, organic arrangement. The atmosphere is one of profound stillness and focused elegance, capturing fleeting, poetic moments of movement and stillness.", + "seed": 20240, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_11/index_1433_seed_20240_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.42087400164455174, + "vlm_logprob": -0.42087400164455174, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.42087400164455174, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7075846195220947 + ], + "opposite_logprob_tokens": [ + -0.2075846791267395 + ], + "logprob_reward": -1.7075846195220947, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7075846195220947, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.145870640873909, + 0.1307384967803955, + 0.1760834902524948, + 0.16062775254249573, + 0.0739968791604042, + 0.07105977088212967, + 0.10468626022338867, + 0.04709598422050476, + 0.04673948884010315, + 0.04310120269656181 + ], + "credit_map": [ + 0.01502661220729351, + 0.029100706800818443, + 0.010980956256389618, + 0.014574670232832432, + 0.01822502352297306, + 0.01972520351409912, + 0.01797080971300602, + 0.020266667008399963, + 0.012871870771050453, + 0.011668728664517403, + 0.009026872925460339, + 0.0127799641340971, + 0.03112746961414814, + 0.02898414060473442, + 0.009282726794481277, + 0.014996723271906376, + 0.019108425825834274, + 0.011315014213323593, + 0.025394536554813385, + 0.019452383741736412, + 0.01897231489419937, + 0.018882298842072487, + 0.028637945652008057, + 0.034320563077926636, + 0.028430312871932983, + 0.021677693352103233, + 0.01283292192965746, + 0.014131905511021614, + 0.02150828205049038, + 0.02944665029644966, + 0.015164625830948353, + 0.017435360699892044, + 0.006993211340159178, + 0.009305459447205067, + 0.007343336474150419, + 0.008282034657895565, + 0.008063988760113716, + 0.00970793329179287, + 0.012097084894776344, + 0.012203832156956196, + 0.008158208802342415, + 0.008642070926725864, + 0.004922435153275728, + 0.006306334864348173, + 0.01372744794934988, + 0.012639294378459454, + 0.006777416914701462, + 0.009886560030281544, + 0.006983208004385233, + 0.018884003162384033, + 0.00941769964993, + 0.00415989151224494, + 0.013245782814919949, + 0.03979044780135155, + 0.006297304295003414, + 0.005907924380153418, + 0.0048920647241175175, + 0.008614026010036469, + 0.0047278571873903275, + 0.0037768615875393152, + 0.0066144117154181, + 0.010654800571501255, + 0.004149230197072029, + 0.0036667324602603912, + 0.0031178013887256384, + 0.004208110272884369, + 0.0017881500534713268, + 0.0036965482868254185, + 0.006292133126407862, + 0.013806399889290333, + 0.0048719546757638454, + 0.008958391845226288, + 0.005019023083150387, + 0.008382953703403473, + 0.003374453168362379, + 0.003980197478085756, + 0.004355461336672306, + 0.009394516237080097, + 0.005462944973260164, + 0.0031316557433456182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -1.7075846195220947 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135158121585846 + ], + "opposite_logprob_tokens": [ + -1.3135157823562622 + ], + "logprob_reward": -0.3135158121585846, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135158121585846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.177871435880661, + 0.3019103705883026, + 0.14114226400852203, + 0.10870352387428284, + 0.03979221358895302, + 0.03499137610197067, + 0.03984017297625542, + 0.022096242755651474, + 0.0755084827542305, + 0.058143891394138336 + ], + "credit_map": [ + 0.019469648599624634, + 0.045864976942539215, + 0.010699471458792686, + 0.009116451255977154, + 0.04073481634259224, + 0.0240030400454998, + 0.012682854197919369, + 0.015300201252102852, + 0.01098270807415247, + 0.011037768796086311, + 0.005831008777022362, + 0.007963759824633598, + 0.12574729323387146, + 0.12446847558021545, + 0.008829228579998016, + 0.0070501468144357204, + 0.015903282910585403, + 0.008369467221200466, + 0.02286825329065323, + 0.016047678887844086, + 0.012628615833818913, + 0.015784865245223045, + 0.02183026634156704, + 0.027709836140275, + 0.02063317969441414, + 0.0141824372112751, + 0.00981102604418993, + 0.010061489418148994, + 0.014797436073422432, + 0.016030458733439445, + 0.012369295582175255, + 0.010818208567798138, + 0.0052010128274559975, + 0.0057350872084498405, + 0.004011078737676144, + 0.004801171366125345, + 0.004595656879246235, + 0.00526180025190115, + 0.004978048149496317, + 0.0052083623595535755, + 0.004639552906155586, + 0.009474139660596848, + 0.0019021689658984542, + 0.002354065887629986, + 0.0040541901253163815, + 0.005505993030965328, + 0.0029249205254018307, + 0.004136348143219948, + 0.005578990094363689, + 0.004525268450379372, + 0.004119865130633116, + 0.004322946537286043, + 0.004717405419796705, + 0.007179249543696642, + 0.004356822930276394, + 0.005039628129452467, + 0.004204654600471258, + 0.0026037145871669054, + 0.002347855595871806, + 0.0018271678127348423, + 0.0026594065129756927, + 0.0028148533310741186, + 0.0027217906899750233, + 0.002916803117841482, + 0.005011043511331081, + 0.006652438547462225, + 0.0031497764866799116, + 0.002164885401725769, + 0.00989014282822609, + 0.02711356431245804, + 0.002846760442480445, + 0.018679877743124962, + 0.004938876256346703, + 0.014584529213607311, + 0.0037074564024806023, + 0.005147819872945547, + 0.006613729987293482, + 0.011375113390386105, + 0.007395924534648657, + 0.004380445461720228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.3135158121585846 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207192316651344 + ], + "opposite_logprob_tokens": [ + -2.8120720386505127 + ], + "logprob_reward": -0.06207192316651344, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207192316651344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22287707030773163, + 0.11472262442111969, + 0.30393531918525696, + 0.047786589711904526, + 0.025221025571227074, + 0.034596037119627, + 0.03482963144779205, + 0.0840526819229126, + 0.11790002137422562, + 0.01407903153449297 + ], + "credit_map": [ + 0.020971326157450676, + 0.04660798981785774, + 0.018871624022722244, + 0.02035418711602688, + 0.028597278520464897, + 0.030262038111686707, + 0.027866419404745102, + 0.029346171766519547, + 0.0063246507197618484, + 0.004684820771217346, + 0.003741276916116476, + 0.004097640048712492, + 0.04280270263552666, + 0.04255504533648491, + 0.004719176329672337, + 0.0057972935028374195, + 0.03615039587020874, + 0.018523333594202995, + 0.04714217782020569, + 0.032755304127931595, + 0.030525214970111847, + 0.03417959064245224, + 0.04940801113843918, + 0.05525128170847893, + 0.00803923886269331, + 0.005306437611579895, + 0.003955783322453499, + 0.003586040809750557, + 0.006730711553245783, + 0.008726023137569427, + 0.005280483514070511, + 0.006161869969218969, + 0.004912792704999447, + 0.002504768781363964, + 0.0020963947754353285, + 0.001889088423922658, + 0.0030110233929008245, + 0.002641199855133891, + 0.0028911724220961332, + 0.005274581257253885, + 0.003058933187276125, + 0.004828480537980795, + 0.003257275791838765, + 0.003711970057338476, + 0.005492114461958408, + 0.004064409993588924, + 0.004259614739567041, + 0.005923235788941383, + 0.002982586156576872, + 0.00425743218511343, + 0.0029427811969071627, + 0.0031147405970841646, + 0.004828937817364931, + 0.004267789423465729, + 0.0059996843338012695, + 0.006435676012188196, + 0.010870247147977352, + 0.015222768299281597, + 0.0067719994112849236, + 0.006272038444876671, + 0.011727690696716309, + 0.019790414720773697, + 0.0069184391759335995, + 0.0064790756441652775, + 0.009142155759036541, + 0.008978221565485, + 0.005007331259548664, + 0.004778525326400995, + 0.015593028627336025, + 0.04223856329917908, + 0.005792694631963968, + 0.026369497179985046, + 0.0020234528928995132, + 0.0022176543716341257, + 0.001318525057286024, + 0.0017974411603063345, + 0.0017703184857964516, + 0.0019899241160601377, + 0.0019168227445334196, + 0.0010448906105011702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.06207192316651344 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012557249516248703 + ], + "opposite_logprob_tokens": [ + -4.387557029724121 + ], + "logprob_reward": -0.012557249516248703, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012557249516248703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19534575939178467, + 0.1959506720304489, + 0.17954972386360168, + 0.027087928727269173, + 0.05602264404296875, + 0.04802371934056282, + 0.10051649808883667, + 0.04949937015771866, + 0.03711069002747536, + 0.11089302599430084 + ], + "credit_map": [ + 0.01611955091357231, + 0.06332287192344666, + 0.010525715537369251, + 0.00864163227379322, + 0.05075906217098236, + 0.015360437333583832, + 0.011051726527512074, + 0.01956476829946041, + 0.008791058324277401, + 0.010704412125051022, + 0.00711841369047761, + 0.012247921898961067, + 0.06686550378799438, + 0.062320783734321594, + 0.015380199067294598, + 0.012522364966571331, + 0.02127804420888424, + 0.013170860707759857, + 0.02415703795850277, + 0.019717613235116005, + 0.018816892057657242, + 0.01955338940024376, + 0.032483380287885666, + 0.030372506007552147, + 0.003511885879561305, + 0.003940438851714134, + 0.00200462993234396, + 0.002239513210952282, + 0.005234939511865377, + 0.0038435161113739014, + 0.0019952133297920227, + 0.00431778933852911, + 0.0054747420363128185, + 0.005874625872820616, + 0.0058860755525529385, + 0.00547699024900794, + 0.006876187399029732, + 0.007222658954560757, + 0.010018207132816315, + 0.00919315405189991, + 0.004071156494319439, + 0.0056771705858409405, + 0.0030270013958215714, + 0.0040410892106592655, + 0.01326791848987341, + 0.0070158205926418304, + 0.005168306175619364, + 0.005755254533141851, + 0.007976059801876545, + 0.012368768453598022, + 0.005500125233083963, + 0.007293866015970707, + 0.016617756336927414, + 0.012181210331618786, + 0.017292441800236702, + 0.021286269649863243, + 0.005426079500466585, + 0.007892545312643051, + 0.0035029775463044643, + 0.004245765507221222, + 0.009485884569585323, + 0.007057391572743654, + 0.0046272785402834415, + 0.0072614471428096294, + 0.0034697712399065495, + 0.0045202006585896015, + 0.0025141516234725714, + 0.004026762209832668, + 0.00468041654676199, + 0.006762291770428419, + 0.005985090043395758, + 0.00515200337395072, + 0.007015772629529238, + 0.02953118085861206, + 0.004052300006151199, + 0.0050994157791137695, + 0.010418510995805264, + 0.04310721531510353, + 0.006106361746788025, + 0.00556225934997201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.012557249516248703 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640403859317303 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640403859317303, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640403859317303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2135220617055893, + 0.11207611113786697, + 0.1579846292734146, + 0.19933941960334778, + 0.04295383393764496, + 0.03834852948784828, + 0.09186244755983353, + 0.022706152871251106, + 0.06070118397474289, + 0.06050568073987961 + ], + "credit_map": [ + 0.02550520934164524, + 0.03013284131884575, + 0.025096764788031578, + 0.020047858357429504, + 0.03399137407541275, + 0.037115029990673065, + 0.022478003054857254, + 0.019154969602823257, + 0.008104193955659866, + 0.005416769068688154, + 0.0033432645723223686, + 0.0037224185653030872, + 0.04033615067601204, + 0.038915835320949554, + 0.005218364764004946, + 0.007019101642072201, + 0.01783389039337635, + 0.017189789563417435, + 0.01082146167755127, + 0.014113315381109715, + 0.02931452915072441, + 0.025489717721939087, + 0.02005036175251007, + 0.023171547800302505, + 0.03423945978283882, + 0.025772159919142723, + 0.015479139983654022, + 0.01876995898783207, + 0.02535327896475792, + 0.036178864538669586, + 0.021770527586340904, + 0.021776005625724792, + 0.004698703996837139, + 0.0058367010205984116, + 0.004299613181501627, + 0.005843651480972767, + 0.006085668690502644, + 0.00424388749524951, + 0.0068450006656348705, + 0.005100604146718979, + 0.004762865602970123, + 0.006448161322623491, + 0.003944214899092913, + 0.0039057566318660975, + 0.0037902023177593946, + 0.005520219448953867, + 0.0043298713862895966, + 0.005647234618663788, + 0.007115045562386513, + 0.014649592339992523, + 0.007612633518874645, + 0.003566402941942215, + 0.01247361395508051, + 0.03543715551495552, + 0.00584594439715147, + 0.005162053741514683, + 0.0031240032985806465, + 0.003446036484092474, + 0.0019511443097144365, + 0.0017845230177044868, + 0.003245218889787793, + 0.004501913208514452, + 0.0016312936786562204, + 0.003022018587216735, + 0.004996221512556076, + 0.005389399826526642, + 0.0034066142980009317, + 0.005880399141460657, + 0.0075025400146842, + 0.017427802085876465, + 0.007159102242439985, + 0.008939102292060852, + 0.007819592952728271, + 0.01377939060330391, + 0.004641989711672068, + 0.005689175799489021, + 0.004585005808621645, + 0.01312690507620573, + 0.007028309162706137, + 0.003835305105894804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.008640403859317303 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a static, aerial view established directly overhead, the camera executes a slow, almost imperceptible pan to follow the fluid movement of a lightweight silk scarf as it drifts across a polished, dark mahogany tabletop. The table's surface is impeccably smooth, its dark, rich veins reflecting soft, diffuse light from an unseen, off-camera source, creating a moody and contemplative canvas. The scarf itself is a masterpiece of exquisite fabric, dyed in ethereal, shifting shades of lavender and dusty rose, its edges woven with fine, shimmering silver threads. The scene comes alive when a gentle, unseen draft cautiously lifts a corner of the scarf; the fabric responds with slow, weightless grace, creating a series of silent, small waves that languidly spread across its surface. Each ripple catches the soft light, causing the silver threads to shimmer and the pastel colors to deepen and lighten in a mesmerizing dance, before the material deliberately lies flat again, its folds resting in a new, organic arrangement. The atmosphere is one of profound stillness and focused elegance, capturing fleeting, poetic moments of movement and stillness.", + "seed": 20241, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_11/index_1433_seed_20241_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.27428505108691753, + "vlm_logprob": -0.27428505108691753, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.27428505108691753, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.46183064579963684 + ], + "opposite_logprob_tokens": [ + -1.0868306159973145 + ], + "logprob_reward": -0.46183064579963684, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.46183064579963684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29711681604385376, + 0.0691407173871994, + 0.02549101784825325, + 0.4535678029060364, + 0.032174088060855865, + 0.01070698443800211, + 0.04071696475148201, + 0.0141248544678092, + 0.01955375075340271, + 0.03740698844194412 + ], + "credit_map": [ + 0.015162899158895016, + 0.01852777600288391, + 0.05531800910830498, + 0.019518887624144554, + 0.010499229654669762, + 0.08009792119264603, + 0.012327543459832668, + 0.08566455543041229, + 0.0063543496653437614, + 0.006294611841440201, + 0.008431798778474331, + 0.008315670304000378, + 0.012506382539868355, + 0.01096128299832344, + 0.006570407189428806, + 0.00970621220767498, + 0.0035647633485496044, + 0.002004406414926052, + 0.004614566918462515, + 0.0024724681861698627, + 0.0028343910817056894, + 0.004340284038335085, + 0.001783860381692648, + 0.003876276547089219, + 0.11448908597230911, + 0.04497712478041649, + 0.025167666375637054, + 0.019040152430534363, + 0.03861282020807266, + 0.0724669098854065, + 0.030174676328897476, + 0.10863936692476273, + 0.004747030325233936, + 0.00406496599316597, + 0.005421743728220463, + 0.005003792699426413, + 0.0036479909904301167, + 0.0027912750374525785, + 0.0022912933491170406, + 0.0042059957049787045, + 0.0017704350175336003, + 0.0013658746611326933, + 0.0010728941997513175, + 0.001488233101554215, + 0.0012780858669430017, + 0.0013215193757787347, + 0.0007227692985907197, + 0.0016871732659637928, + 0.004661806859076023, + 0.006928599905222654, + 0.005288832820951939, + 0.00438319705426693, + 0.006006503943353891, + 0.00387685582973063, + 0.002905074041336775, + 0.0066660926677286625, + 0.0015908530913293362, + 0.00160076969768852, + 0.002329885493963957, + 0.0018203147919848561, + 0.0026304814964532852, + 0.0013385225320234895, + 0.0011114540975540876, + 0.0017025721026584506, + 0.0028602818492799997, + 0.0021298243664205074, + 0.0028675920329988003, + 0.002450495958328247, + 0.0033136375714093447, + 0.002255789702758193, + 0.0014287879457697272, + 0.0022473426070064306, + 0.00402343412861228, + 0.004767676815390587, + 0.0027841771952807903, + 0.004038660787045956, + 0.008423215709626675, + 0.004048431292176247, + 0.0030088801868259907, + 0.0063125090673565865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.46183064579963684 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983514308929443 + ], + "opposite_logprob_tokens": [ + -0.5233514308929443 + ], + "logprob_reward": -0.8983514308929443, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983514308929443, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.29063427448272705, + 0.04420807585120201, + 0.02205047197639942, + 0.48343023657798767, + 0.029052751138806343, + 0.021631229668855667, + 0.029340874403715134, + 0.026872746646404266, + 0.034418985247612, + 0.018360333517193794 + ], + "credit_map": [ + 0.010629513300955296, + 0.010751406662166119, + 0.0304414015263319, + 0.01004073303192854, + 0.006535341031849384, + 0.1064152866601944, + 0.007709647994488478, + 0.1081109493970871, + 0.004293054807931185, + 0.004630283452570438, + 0.006518305279314518, + 0.0054226228967309, + 0.007814214564859867, + 0.005922913085669279, + 0.004297064617276192, + 0.005309623666107655, + 0.002158078830689192, + 0.001894757617264986, + 0.0024031903594732285, + 0.003006798680871725, + 0.0021695939358323812, + 0.004582379944622517, + 0.0020705380011349916, + 0.00376513646915555, + 0.13172967731952667, + 0.049991458654403687, + 0.02552560344338417, + 0.02202264964580536, + 0.04317784309387207, + 0.06418860703706741, + 0.029228487983345985, + 0.11756592988967896, + 0.0068106879480183125, + 0.00405790563672781, + 0.0029325555078685284, + 0.0024331791792064905, + 0.0038211573846638203, + 0.0029132096096873283, + 0.001824179315008223, + 0.004259878769516945, + 0.003222367027774453, + 0.00399828003719449, + 0.0018056571716442704, + 0.0022769165225327015, + 0.003351544262841344, + 0.0027910482604056597, + 0.0017745312070474029, + 0.002410885412245989, + 0.0034511706326156855, + 0.004682654980570078, + 0.003618102055042982, + 0.0036606460344046354, + 0.005540220066905022, + 0.0025093238800764084, + 0.001994191901758313, + 0.0038845704402774572, + 0.0034657965879887342, + 0.0029159497935324907, + 0.002877215389162302, + 0.0036636353470385075, + 0.005989233031868935, + 0.0027553518302738667, + 0.0022422007750719786, + 0.0029633650556206703, + 0.005822873208671808, + 0.0030849785543978214, + 0.00612275768071413, + 0.0029211523942649364, + 0.006522377021610737, + 0.004163943696767092, + 0.0024013970978558064, + 0.003379508852958679, + 0.002233698731288314, + 0.0024919810239225626, + 0.0014747095992788672, + 0.0019389571389183402, + 0.0032663519959896803, + 0.0016867599915713072, + 0.0015681263757869601, + 0.003699749242514372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.8983514308929443 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00250038574449718 + ], + "opposite_logprob_tokens": [ + -6.002500534057617 + ], + "logprob_reward": -0.00250038574449718, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00250038574449718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5821134448051453, + 0.08896749466657639, + 0.0330749973654747, + 0.19713352620601654, + 0.025817476212978363, + 0.012580727227032185, + 0.013199292123317719, + 0.013709348626434803, + 0.011465776711702347, + 0.02193792723119259 + ], + "credit_map": [ + 0.0200242567807436, + 0.01824018359184265, + 0.05649608001112938, + 0.017911970615386963, + 0.01290151197463274, + 0.22073005139827728, + 0.013820759952068329, + 0.22198860347270966, + 0.007427725940942764, + 0.009128623642027378, + 0.011347847990691662, + 0.013627516105771065, + 0.008872383274137974, + 0.018995624035596848, + 0.005358971655368805, + 0.014208813197910786, + 0.002678131451830268, + 0.0033128708600997925, + 0.0032070723827928305, + 0.00360670848749578, + 0.003865647129714489, + 0.008279450237751007, + 0.002014638390392065, + 0.006110482383519411, + 0.053569644689559937, + 0.018029291182756424, + 0.010730448178946972, + 0.009129338897764683, + 0.020360860973596573, + 0.03200620785355568, + 0.011348088271915913, + 0.041959669440984726, + 0.0038543306291103363, + 0.0040336246602237225, + 0.0046651423908770084, + 0.0034938850440084934, + 0.0023504928685724735, + 0.002611269475892186, + 0.0015033591771498322, + 0.003305373014882207, + 0.0016640156973153353, + 0.0013059743214398623, + 0.001781004830263555, + 0.0016827881336212158, + 0.0016275777015835047, + 0.0016409357776865363, + 0.0010172742186114192, + 0.0018611575942486525, + 0.0013989852741360664, + 0.001705633127130568, + 0.002271438715979457, + 0.0016550839645788074, + 0.0023443421814590693, + 0.0012231012806296349, + 0.0009311081375926733, + 0.0016695997910574079, + 0.0016509322449564934, + 0.0014279407914727926, + 0.0022807836066931486, + 0.001687863958068192, + 0.002045105677098036, + 0.00138060562312603, + 0.0011425937991589308, + 0.0020935237407684326, + 0.0016471013659611344, + 0.0012862792937085032, + 0.0011605496983975172, + 0.0017927984008565545, + 0.0017842119559645653, + 0.001409894903190434, + 0.0010656469967216253, + 0.0013192951446399093, + 0.0023531275801360607, + 0.002500323811545968, + 0.001445566420443356, + 0.00215704133734107, + 0.005619396455585957, + 0.0021941803861409426, + 0.001611283514648676, + 0.004057008773088455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.00250038574449718 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004103930201381445 + ], + "opposite_logprob_tokens": [ + -5.504104137420654 + ], + "logprob_reward": -0.004103930201381445, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004103930201381445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25464797019958496, + 0.08406182378530502, + 0.03152520954608917, + 0.41451382637023926, + 0.11292879283428192, + 0.019417382776737213, + 0.02277938462793827, + 0.01718704216182232, + 0.02404417097568512, + 0.018894417211413383 + ], + "credit_map": [ + 0.013157022185623646, + 0.020276155322790146, + 0.02930152788758278, + 0.0117512671276927, + 0.013454548083245754, + 0.08028394728899002, + 0.008799491450190544, + 0.07762397825717926, + 0.009168452583253384, + 0.007976402528584003, + 0.009953649714589119, + 0.008146331645548344, + 0.014337928965687752, + 0.014495889656245708, + 0.008044891059398651, + 0.011938264593482018, + 0.003734881989657879, + 0.0034931686241179705, + 0.004277724772691727, + 0.004347566980868578, + 0.004076154436916113, + 0.004620063118636608, + 0.0026552833151072264, + 0.004320364445447922, + 0.12281044572591782, + 0.042914390563964844, + 0.021101290360093117, + 0.01818927749991417, + 0.04218830540776253, + 0.04572828859090805, + 0.024578625336289406, + 0.09700319916009903, + 0.02377178706228733, + 0.014012996107339859, + 0.008059388026595116, + 0.014333878643810749, + 0.014517189934849739, + 0.012807813473045826, + 0.007022328674793243, + 0.018403401598334312, + 0.0036306732799857855, + 0.0025158256758004427, + 0.002452240791171789, + 0.0019676059018820524, + 0.002142185578122735, + 0.0026507964357733727, + 0.0014500792603939772, + 0.002607974922284484, + 0.0034908493980765343, + 0.004751918371766806, + 0.0021343843545764685, + 0.0031347491312772036, + 0.0032026495318859816, + 0.001875347807072103, + 0.0015851963544264436, + 0.0026042878162115812, + 0.0022841873578727245, + 0.0020836438052356243, + 0.00221825810149312, + 0.0020023216493427753, + 0.0029829740524291992, + 0.0016126214759424329, + 0.0012363961432129145, + 0.0027666385285556316, + 0.0037565426900982857, + 0.0025352018419653177, + 0.0023190435022115707, + 0.0025360980071127415, + 0.006125308573246002, + 0.0027508926577866077, + 0.0017907634610310197, + 0.002230316400527954, + 0.002257555490359664, + 0.0019891278352588415, + 0.0018913341918960214, + 0.002043156186118722, + 0.003537536133080721, + 0.0017984983278438449, + 0.001553458976559341, + 0.003823748091235757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.004103930201381445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638862796127796 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.004638862796127796, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638862796127796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29994937777519226, + 0.06706949323415756, + 0.04282797500491142, + 0.32293736934661865, + 0.05152217298746109, + 0.026334956288337708, + 0.055076051503419876, + 0.05503154173493385, + 0.061984267085790634, + 0.01726681925356388 + ], + "credit_map": [ + 0.02182210423052311, + 0.03695753589272499, + 0.06455779075622559, + 0.021739140152931213, + 0.021797209978103638, + 0.06078556925058365, + 0.017522213980555534, + 0.054767780005931854, + 0.006329679861664772, + 0.008526260033249855, + 0.009941433556377888, + 0.004736464470624924, + 0.008108883164823055, + 0.013346916064620018, + 0.006391049828380346, + 0.00968879647552967, + 0.0070388116873800755, + 0.004719255492091179, + 0.007285327184945345, + 0.0053720176219940186, + 0.0036907847970724106, + 0.005099842790514231, + 0.003455609316006303, + 0.006166323088109493, + 0.06513720750808716, + 0.02281646989285946, + 0.019033607095479965, + 0.01644008234143257, + 0.03669647499918938, + 0.08418361842632294, + 0.018524184823036194, + 0.06010570004582405, + 0.009950022213160992, + 0.006041806656867266, + 0.008470520377159119, + 0.0052123889327049255, + 0.006118752062320709, + 0.00474742753431201, + 0.00284819514490664, + 0.00813305750489235, + 0.004399842582643032, + 0.00327067612670362, + 0.003098459215834737, + 0.0037669294979423285, + 0.002903580665588379, + 0.0032364854123443365, + 0.0022361420560628176, + 0.0034228379372507334, + 0.00830897968262434, + 0.01304781623184681, + 0.0048462944105267525, + 0.006118094548583031, + 0.007638467475771904, + 0.004704449325799942, + 0.0035511914175003767, + 0.006860753986984491, + 0.007767918519675732, + 0.006769979372620583, + 0.007238604594022036, + 0.006861594505608082, + 0.00917037669569254, + 0.0050654360093176365, + 0.00445539178326726, + 0.007702231407165527, + 0.01070763822644949, + 0.005800827872008085, + 0.010366193018853664, + 0.005619044415652752, + 0.012511970475316048, + 0.006969484034925699, + 0.0040689073503017426, + 0.005940197966992855, + 0.0022606272250413895, + 0.001910352730192244, + 0.0018738567596301436, + 0.002305480884388089, + 0.0032727455254644156, + 0.001961569767445326, + 0.0010877299355342984, + 0.0025944567751139402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.004638862796127796 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a static, aerial view established directly overhead, the camera executes a slow, almost imperceptible pan to follow the fluid movement of a lightweight silk scarf as it drifts across a polished, dark mahogany tabletop. The table's surface is impeccably smooth, its dark, rich veins reflecting soft, diffuse light from an unseen, off-camera source, creating a moody and contemplative canvas. The scarf itself is a masterpiece of exquisite fabric, dyed in ethereal, shifting shades of lavender and dusty rose, its edges woven with fine, shimmering silver threads. The scene comes alive when a gentle, unseen draft cautiously lifts a corner of the scarf; the fabric responds with slow, weightless grace, creating a series of silent, small waves that languidly spread across its surface. Each ripple catches the soft light, causing the silver threads to shimmer and the pastel colors to deepen and lighten in a mesmerizing dance, before the material deliberately lies flat again, its folds resting in a new, organic arrangement. The atmosphere is one of profound stillness and focused elegance, capturing fleeting, poetic moments of movement and stillness.", + "seed": 20242, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_11/index_1433_seed_20242_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.29386505130678414, + "vlm_logprob": -0.29386505130678414, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.29386505130678414, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9215649962425232 + ], + "opposite_logprob_tokens": [ + -0.5465649962425232 + ], + "logprob_reward": -0.9215649962425232, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9215649962425232, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11744319647550583, + 0.10199160128831863, + 0.09617002308368683, + 0.10821042954921722, + 0.11242462694644928, + 0.052419181913137436, + 0.055303607136011124, + 0.1103895828127861, + 0.1201544851064682, + 0.12549324333667755 + ], + "credit_map": [ + 0.014741111546754837, + 0.010865829885005951, + 0.015935588628053665, + 0.009485676884651184, + 0.010605829767882824, + 0.017443856224417686, + 0.014372752048075199, + 0.023992547765374184, + 0.016401778906583786, + 0.010327757336199284, + 0.017260588705539703, + 0.011951806023716927, + 0.009137659333646297, + 0.008088587783277035, + 0.00978107564151287, + 0.019042346626520157, + 0.012464631348848343, + 0.009305366314947605, + 0.017086412757635117, + 0.007544691674411297, + 0.00820870604366064, + 0.017805065959692, + 0.009685703553259373, + 0.01406943891197443, + 0.023663697764277458, + 0.008294584229588509, + 0.015983542427420616, + 0.01206482294946909, + 0.011205978691577911, + 0.01147244032472372, + 0.009570449590682983, + 0.015954913571476936, + 0.013302812352776527, + 0.010681284591555595, + 0.018627788871526718, + 0.017701826989650726, + 0.011189783923327923, + 0.012551333755254745, + 0.013215936720371246, + 0.01515386812388897, + 0.0081771370023489, + 0.0030471133068203926, + 0.010159626603126526, + 0.013528089039027691, + 0.0036698661278933287, + 0.00474531389772892, + 0.003032241016626358, + 0.0060597918927669525, + 0.005748338531702757, + 0.004529237747192383, + 0.009230362251400948, + 0.008812349289655685, + 0.004620502702891827, + 0.005568044725805521, + 0.0059974901378154755, + 0.010797282680869102, + 0.010075744241476059, + 0.0074864947237074375, + 0.017053147777915, + 0.012309777550399303, + 0.007625292520970106, + 0.017524786293506622, + 0.009432222694158554, + 0.028882114216685295, + 0.011079343967139721, + 0.01543261855840683, + 0.027719100937247276, + 0.013806177303195, + 0.01015134621411562, + 0.01375950314104557, + 0.007060885429382324, + 0.021145515143871307, + 0.009835136123001575, + 0.011122288182377815, + 0.02077673375606537, + 0.03303506225347519, + 0.008402622304856777, + 0.007276076823472977, + 0.024309726431965828, + 0.010735606774687767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.9215649962425232 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233842730522156 + ], + "opposite_logprob_tokens": [ + -0.8983842730522156 + ], + "logprob_reward": -0.5233842730522156, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233842730522156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2602445185184479, + 0.03020579367876053, + 0.11124260723590851, + 0.14000776410102844, + 0.022655703127384186, + 0.05875649303197861, + 0.09908092767000198, + 0.059715863317251205, + 0.08869396150112152, + 0.12939630448818207 + ], + "credit_map": [ + 0.024184681475162506, + 0.039563581347465515, + 0.027674607932567596, + 0.03287564963102341, + 0.026799028739333153, + 0.033816348761320114, + 0.033765047788619995, + 0.04156559333205223, + 0.0061973268166184425, + 0.003334527602419257, + 0.005171455908566713, + 0.003119050059467554, + 0.002964091720059514, + 0.002231342950835824, + 0.0031352967489510775, + 0.004052701406180859, + 0.014996740967035294, + 0.008222081698477268, + 0.021125428378582, + 0.008623521775007248, + 0.009451751597225666, + 0.02504413202404976, + 0.00934689398854971, + 0.01443206425756216, + 0.018150312826037407, + 0.012379247695207596, + 0.021382255479693413, + 0.014160961844027042, + 0.02489754743874073, + 0.024671657010912895, + 0.014447256922721863, + 0.009918526746332645, + 0.003221341175958514, + 0.002281285123899579, + 0.00344073330052197, + 0.0034510144032537937, + 0.0030377451330423355, + 0.0019663390703499317, + 0.002563755726441741, + 0.0026934898924082518, + 0.009437344036996365, + 0.004566207993775606, + 0.009312594309449196, + 0.009707541204988956, + 0.005861909594386816, + 0.005930487532168627, + 0.006102823186665773, + 0.007837587036192417, + 0.014842461794614792, + 0.008107803761959076, + 0.014298372901976109, + 0.01371893472969532, + 0.009205486625432968, + 0.012274377048015594, + 0.008537329733371735, + 0.018096167594194412, + 0.006837153807282448, + 0.004923959728330374, + 0.008357659913599491, + 0.005379038862884045, + 0.005468699149787426, + 0.013378317467868328, + 0.005687409546226263, + 0.009683622978627682, + 0.009876561351120472, + 0.014890727587044239, + 0.01222823653370142, + 0.007345597259700298, + 0.012092496268451214, + 0.012075883336365223, + 0.005209787283092737, + 0.014974674209952354, + 0.011735641397535801, + 0.014013423584401608, + 0.01630297861993313, + 0.030985025689005852, + 0.009183964692056179, + 0.006499773822724819, + 0.030179640278220177, + 0.01049585547298193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.5233842730522156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647287540137768 + ], + "opposite_logprob_tokens": [ + -5.379647254943848 + ], + "logprob_reward": -0.004647287540137768, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647287540137768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25464996695518494, + 0.04894145950675011, + 0.033416759222745895, + 0.08859992772340775, + 0.17370405793190002, + 0.09476879984140396, + 0.04900931194424629, + 0.14982476830482483, + 0.06961054354906082, + 0.0374743826687336 + ], + "credit_map": [ + 0.03458060324192047, + 0.029012227430939674, + 0.037359751760959625, + 0.01863102987408638, + 0.020101290196180344, + 0.03109312430024147, + 0.026050807908177376, + 0.057821132242679596, + 0.006425632629543543, + 0.004956758581101894, + 0.00764845497906208, + 0.005392592400312424, + 0.005874699912965298, + 0.0035648690536618233, + 0.005609126761555672, + 0.00946932751685381, + 0.004968112800270319, + 0.002726583043113351, + 0.007109391503036022, + 0.003104821313172579, + 0.003209776012226939, + 0.00397130660712719, + 0.003416317980736494, + 0.004910450428724289, + 0.010230653919279575, + 0.00832796934992075, + 0.01375549379736185, + 0.010046699084341526, + 0.015270397998392582, + 0.014270906336605549, + 0.009718550369143486, + 0.006979256868362427, + 0.02601122111082077, + 0.013822264969348907, + 0.02722812630236149, + 0.033703308552503586, + 0.018541615456342697, + 0.01789010874927044, + 0.01993159018456936, + 0.016575828194618225, + 0.012610270641744137, + 0.006531492806971073, + 0.01763736642897129, + 0.020898358896374702, + 0.011079908348619938, + 0.008541442453861237, + 0.008455894887447357, + 0.009014072827994823, + 0.004809702280908823, + 0.0030195345170795918, + 0.010450494475662708, + 0.013393580913543701, + 0.0034836006816476583, + 0.004923220723867416, + 0.003350054146721959, + 0.0055791232734918594, + 0.02140544168651104, + 0.00828755833208561, + 0.024191809818148613, + 0.02831815741956234, + 0.013248542323708534, + 0.01788952946662903, + 0.010599837638437748, + 0.025883900001645088, + 0.007718110457062721, + 0.00841268990188837, + 0.01577026955783367, + 0.00918834563344717, + 0.007479365915060043, + 0.007977309636771679, + 0.005047500133514404, + 0.008016953244805336, + 0.0072478544898331165, + 0.005027412436902523, + 0.005722855683416128, + 0.004088201094418764, + 0.0029348726384341717, + 0.003151815850287676, + 0.003580614924430847, + 0.005720756482332945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.004647287540137768 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089595966041088 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.011089595966041088, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089595966041088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3171803951263428, + 0.058697257190942764, + 0.14838606119155884, + 0.11037905514240265, + 0.056552112102508545, + 0.03708936274051666, + 0.09931904822587967, + 0.07072177529335022, + 0.05389007180929184, + 0.047784898430109024 + ], + "credit_map": [ + 0.044052448123693466, + 0.037681978195905685, + 0.04184526950120926, + 0.020155636593699455, + 0.02765693888068199, + 0.051133837550878525, + 0.02986307442188263, + 0.06479119509458542, + 0.008353884331882, + 0.006701315753161907, + 0.00931001827120781, + 0.006874007638543844, + 0.005679372698068619, + 0.004669043701142073, + 0.00848991796374321, + 0.008619694039225578, + 0.0198504775762558, + 0.010042700916528702, + 0.030102141201496124, + 0.016256429255008698, + 0.02011767216026783, + 0.018301021307706833, + 0.016843389719724655, + 0.016872230917215347, + 0.014006733894348145, + 0.009797707200050354, + 0.01692366600036621, + 0.011159897781908512, + 0.01776876673102379, + 0.01759604923427105, + 0.014056691899895668, + 0.009069539606571198, + 0.00744953379034996, + 0.004667643923312426, + 0.009557454846799374, + 0.010809839703142643, + 0.0065023647621273994, + 0.004725308157503605, + 0.006622758693993092, + 0.006217204034328461, + 0.004964756779372692, + 0.003305265214294195, + 0.006184321362525225, + 0.005306152161210775, + 0.0035487827844917774, + 0.0031257998198270798, + 0.004088613670319319, + 0.006565670482814312, + 0.011390842497348785, + 0.009217988699674606, + 0.015414911322295666, + 0.015346086584031582, + 0.008360246196389198, + 0.009772548452019691, + 0.008716288022696972, + 0.021100129932165146, + 0.009314902126789093, + 0.00454651890322566, + 0.011233707889914513, + 0.007489518262445927, + 0.004880729131400585, + 0.01174919493496418, + 0.004998342599719763, + 0.016508854925632477, + 0.005972369108349085, + 0.00614550756290555, + 0.007641958072781563, + 0.005166292190551758, + 0.0072618029080331326, + 0.00806095078587532, + 0.005195192992687225, + 0.008445994928479195, + 0.010068400762975216, + 0.0067264363169670105, + 0.00821719691157341, + 0.003577633760869503, + 0.004589250311255455, + 0.0031066311057657003, + 0.007690259721130133, + 0.0038090862799435854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.011089595966041088 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863910373300314 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.00863910373300314, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863910373300314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31508439779281616, + 0.06997093558311462, + 0.0929822027683258, + 0.09801483899354935, + 0.088805191218853, + 0.058006592094898224, + 0.058769501745700836, + 0.09439557790756226, + 0.06545059382915497, + 0.05852017179131508 + ], + "credit_map": [ + 0.045556649565696716, + 0.04206501692533493, + 0.041355594992637634, + 0.025279097259044647, + 0.02644781768321991, + 0.039333466440439224, + 0.033446572721004486, + 0.06160016730427742, + 0.011036843061447144, + 0.006498510483652353, + 0.011100579053163528, + 0.008866354823112488, + 0.007257141638547182, + 0.00512167252600193, + 0.007735606282949448, + 0.012354233302175999, + 0.013218567706644535, + 0.00779455853626132, + 0.01593906618654728, + 0.008462860248982906, + 0.008883129805326462, + 0.018323825672268867, + 0.009010673500597477, + 0.011349520646035671, + 0.019658012315630913, + 0.007840896025300026, + 0.015051577240228653, + 0.010709847323596478, + 0.009268076159060001, + 0.010440289042890072, + 0.00872611440718174, + 0.016320019960403442, + 0.015307334251701832, + 0.007027534302324057, + 0.012844209559261799, + 0.014198342338204384, + 0.010479954071342945, + 0.010346522554755211, + 0.009364491328597069, + 0.009236805140972137, + 0.008710753172636032, + 0.0033543782774358988, + 0.011930041015148163, + 0.011980336159467697, + 0.0070538897998631, + 0.004211587831377983, + 0.004348980728536844, + 0.0064166211523115635, + 0.006091459188610315, + 0.005257400218397379, + 0.011100470088422298, + 0.014450198039412498, + 0.004744402132928371, + 0.005126143340021372, + 0.004634448792785406, + 0.007364974357187748, + 0.011336381547152996, + 0.007564219646155834, + 0.01393922045826912, + 0.011345230042934418, + 0.007745979819446802, + 0.021047644317150116, + 0.006778098177164793, + 0.014638807624578476, + 0.0057732765562832355, + 0.010142565704882145, + 0.01108099240809679, + 0.006773855071514845, + 0.008444077335298061, + 0.0055826022289693356, + 0.005849601700901985, + 0.011803620494902134, + 0.005199441686272621, + 0.007220753468573093, + 0.008948681876063347, + 0.008372252807021141, + 0.0056864600628614426, + 0.0036686407402157784, + 0.011968008242547512, + 0.007455933839082718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.00863910373300314 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a static, aerial view established directly overhead, the camera executes a slow, almost imperceptible pan to follow the fluid movement of a lightweight silk scarf as it drifts across a polished, dark mahogany tabletop. The table's surface is impeccably smooth, its dark, rich veins reflecting soft, diffuse light from an unseen, off-camera source, creating a moody and contemplative canvas. The scarf itself is a masterpiece of exquisite fabric, dyed in ethereal, shifting shades of lavender and dusty rose, its edges woven with fine, shimmering silver threads. The scene comes alive when a gentle, unseen draft cautiously lifts a corner of the scarf; the fabric responds with slow, weightless grace, creating a series of silent, small waves that languidly spread across its surface. Each ripple catches the soft light, causing the silver threads to shimmer and the pastel colors to deepen and lighten in a mesmerizing dance, before the material deliberately lies flat again, its folds resting in a new, organic arrangement. The atmosphere is one of profound stillness and focused elegance, capturing fleeting, poetic moments of movement and stillness.", + "seed": 20243, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_11/index_1433_seed_20243_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.22355876285582782, + "vlm_logprob": -0.22355876285582782, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.22355876285582782, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6666131019592285 + ], + "opposite_logprob_tokens": [ + -0.7916131019592285 + ], + "logprob_reward": -0.6666131019592285, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6666131019592285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14861035346984863, + 0.09111374616622925, + 0.0691487044095993, + 0.1261831670999527, + 0.201677143573761, + 0.09968914836645126, + 0.029291542246937752, + 0.03823374956846237, + 0.14865589141845703, + 0.04739655926823616 + ], + "credit_map": [ + 0.012630884535610676, + 0.02853928506374359, + 0.014386994764208794, + 0.011460630223155022, + 0.023782044649124146, + 0.01792200841009617, + 0.013497157953679562, + 0.026391364634037018, + 0.006068329326808453, + 0.01650591567158699, + 0.007891673594713211, + 0.008785529993474483, + 0.009568803943693638, + 0.007805051282048225, + 0.018034664914011955, + 0.016453774645924568, + 0.009874245151877403, + 0.011991129256784916, + 0.006294223014265299, + 0.008509083651006222, + 0.007851028814911842, + 0.008148860186338425, + 0.008573551662266254, + 0.007906586863100529, + 0.011453785933554173, + 0.023077236488461494, + 0.009534185752272606, + 0.021017486229538918, + 0.014254684560000896, + 0.012128520756959915, + 0.017826667055487633, + 0.016890600323677063, + 0.008841472677886486, + 0.04650436341762543, + 0.00958404503762722, + 0.013363867998123169, + 0.013855086639523506, + 0.00973100308328867, + 0.08157011866569519, + 0.01822717860341072, + 0.02214496023952961, + 0.006633478216826916, + 0.00984929595142603, + 0.011438839137554169, + 0.007254408206790686, + 0.01032845675945282, + 0.02506660670042038, + 0.006973110605031252, + 0.0053299241699278355, + 0.0026655676774680614, + 0.005414366256445646, + 0.0036923247389495373, + 0.002736644586548209, + 0.0033018304966390133, + 0.0029313096310943365, + 0.0032195725943893194, + 0.00636453740298748, + 0.003981128800660372, + 0.006642712280154228, + 0.004523936193436384, + 0.004631890915334225, + 0.005148295313119888, + 0.004105815198272467, + 0.0028354355599731207, + 0.01789849065244198, + 0.021959198638796806, + 0.017910920083522797, + 0.02667953446507454, + 0.012696067802608013, + 0.02022152952849865, + 0.015275003388524055, + 0.016015147790312767, + 0.007375246379524469, + 0.005980018060654402, + 0.005654598586261272, + 0.008781655691564083, + 0.003125554881989956, + 0.0039834617637097836, + 0.004276702646166086, + 0.008219318464398384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.6666131019592285 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42894241213798523 + ], + "opposite_logprob_tokens": [ + -1.0539424419403076 + ], + "logprob_reward": -0.42894241213798523, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42894241213798523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12707221508026123, + 0.07467080652713776, + 0.06447619199752808, + 0.07745812088251114, + 0.23693467676639557, + 0.0782242864370346, + 0.13221125304698944, + 0.05939355120062828, + 0.10162727534770966, + 0.04793157801032066 + ], + "credit_map": [ + 0.012856421992182732, + 0.014893695712089539, + 0.017690399661660194, + 0.010521164163947105, + 0.015990931540727615, + 0.014164600521326065, + 0.0115822060033679, + 0.029372816905379295, + 0.005125534720718861, + 0.010511183179914951, + 0.004485196899622679, + 0.007816828787326813, + 0.007987972348928452, + 0.005833673290908337, + 0.02313331514596939, + 0.009777109138667583, + 0.007514846511185169, + 0.0083718691021204, + 0.004896266385912895, + 0.013660166412591934, + 0.007671691011637449, + 0.007529154419898987, + 0.007365733850747347, + 0.007466467563062906, + 0.007882512174546719, + 0.011826308444142342, + 0.007330683991312981, + 0.014239051379263401, + 0.008839364163577557, + 0.008484234102070332, + 0.010736445896327496, + 0.008119520731270313, + 0.007939993403851986, + 0.05978706479072571, + 0.010464138351380825, + 0.013377503491938114, + 0.014672989025712013, + 0.009384391829371452, + 0.09736136347055435, + 0.02394726872444153, + 0.01471751183271408, + 0.007253473158925772, + 0.007857562974095345, + 0.008361547254025936, + 0.007246183697134256, + 0.008976487442851067, + 0.016782516613602638, + 0.007029014639556408, + 0.03187767416238785, + 0.00713937496766448, + 0.03161578252911568, + 0.027798999100923538, + 0.008890116587281227, + 0.01017242856323719, + 0.006812336854636669, + 0.007904552854597569, + 0.010801301337778568, + 0.0064031388610601425, + 0.0106821209192276, + 0.007627580780535936, + 0.006281526759266853, + 0.005858330521732569, + 0.007752392441034317, + 0.0039871614426374435, + 0.017614252865314484, + 0.007588020525872707, + 0.014385814778506756, + 0.01668432168662548, + 0.00854556541889906, + 0.013475046493113041, + 0.013217638246715069, + 0.01011662743985653, + 0.006459038704633713, + 0.004540480673313141, + 0.006249592173844576, + 0.009998492896556854, + 0.004178975708782673, + 0.0056298524141311646, + 0.005339782685041428, + 0.005535367410629988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.42894241213798523 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005959717091172934 + ], + "opposite_logprob_tokens": [ + -5.130959510803223 + ], + "logprob_reward": -0.005959717091172934, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005959717091172934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07031188905239105, + 0.2008141279220581, + 0.0817503035068512, + 0.03238607197999954, + 0.07978473603725433, + 0.1945573389530182, + 0.1749766319990158, + 0.05094897747039795, + 0.10173989087343216, + 0.012730044312775135 + ], + "credit_map": [ + 0.008887778967618942, + 0.00814750138670206, + 0.010469886474311352, + 0.007203457877039909, + 0.011119110509753227, + 0.008191702887415886, + 0.009124205447733402, + 0.007168245036154985, + 0.014960899949073792, + 0.02575661428272724, + 0.014345627278089523, + 0.016700515523552895, + 0.01814216934144497, + 0.015630299225449562, + 0.05688757449388504, + 0.03839041292667389, + 0.00717596709728241, + 0.012166142463684082, + 0.0074327378533780575, + 0.017779046669602394, + 0.009900589473545551, + 0.00930982455611229, + 0.00886991061270237, + 0.009116076864302158, + 0.0032670237123966217, + 0.003952867817133665, + 0.002298379084095359, + 0.006351515185087919, + 0.003949773497879505, + 0.0035226643085479736, + 0.005225272849202156, + 0.0038185748271644115, + 0.005299985874444246, + 0.016862163320183754, + 0.006601308472454548, + 0.0049818288534879684, + 0.004718205891549587, + 0.005025009158998728, + 0.02716118097305298, + 0.009135046042501926, + 0.044131483882665634, + 0.016511186957359314, + 0.024371987208724022, + 0.035363055765628815, + 0.017914073541760445, + 0.02451930195093155, + 0.020154468715190887, + 0.01159177627414465, + 0.03650665655732155, + 0.01052643358707428, + 0.04302283003926277, + 0.03757493942975998, + 0.013128874823451042, + 0.013380075804889202, + 0.010789524763822556, + 0.010047285817563534, + 0.008863312192261219, + 0.005172883626073599, + 0.009918160736560822, + 0.006816455628722906, + 0.005295857787132263, + 0.006124951411038637, + 0.005154297687113285, + 0.003603055141866207, + 0.01175154559314251, + 0.0152689004316926, + 0.012246387079358101, + 0.017811069265007973, + 0.00925920344889164, + 0.013760581612586975, + 0.01072480995208025, + 0.010917387902736664, + 0.002269417978823185, + 0.001299305004067719, + 0.0016298722475767136, + 0.001638453220948577, + 0.0010549151338636875, + 0.00165046367328614, + 0.0015590806724503636, + 0.0016285356832668185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.005959717091172934 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764006981626153 + ], + "opposite_logprob_tokens": [ + -4.8826398849487305 + ], + "logprob_reward": -0.00764006981626153, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764006981626153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04565267264842987, + 0.02540942281484604, + 0.07096406817436218, + 0.03834833577275276, + 0.1887756735086441, + 0.5007292032241821, + 0.01745089516043663, + 0.03427157923579216, + 0.05958510562777519, + 0.018813008442521095 + ], + "credit_map": [ + 0.004855927545577288, + 0.007991532795131207, + 0.0033627557568252087, + 0.003379703499376774, + 0.010386373847723007, + 0.004638452548533678, + 0.006079385057091713, + 0.004958536010235548, + 0.0025783153250813484, + 0.0036417748779058456, + 0.002193448133766651, + 0.0038543380796909332, + 0.0033935715910047293, + 0.0032310010865330696, + 0.0030759365763515234, + 0.003441033186390996, + 0.007836068980395794, + 0.01065156888216734, + 0.005230095703154802, + 0.01005034800618887, + 0.008173579350113869, + 0.012464601546525955, + 0.009571989066898823, + 0.00698581337928772, + 0.0026519566308707, + 0.005312248133122921, + 0.0024817369412630796, + 0.00754672521725297, + 0.004787397105246782, + 0.004442310426384211, + 0.006990264635533094, + 0.004135691095143557, + 0.006519573274999857, + 0.0462241992354393, + 0.0077590541914105415, + 0.01161287259310484, + 0.012493086978793144, + 0.00626421719789505, + 0.08031077682971954, + 0.01759188435971737, + 0.10836686193943024, + 0.0421757698059082, + 0.06110015884041786, + 0.09621313959360123, + 0.045379456132650375, + 0.05935772508382797, + 0.05753116309642792, + 0.03060489147901535, + 0.002121619414538145, + 0.0014709725510329008, + 0.0012997856829315424, + 0.006026398856192827, + 0.0021789742168039083, + 0.0019522774964571, + 0.0014854819746688008, + 0.0009153828141279519, + 0.00557637820020318, + 0.004160704091191292, + 0.006038605701178312, + 0.005332939326763153, + 0.0033577748108655214, + 0.0037600696086883545, + 0.0034881827887147665, + 0.0025569195859134197, + 0.006925183814018965, + 0.008930538780987263, + 0.0071035888977348804, + 0.010203732177615166, + 0.0053924149833619595, + 0.008291755802929401, + 0.006356785073876381, + 0.006381100043654442, + 0.0027211331762373447, + 0.0015441215364262462, + 0.0025202645920217037, + 0.00278390571475029, + 0.001470701303333044, + 0.002192388754338026, + 0.0023521718103438616, + 0.0032283205073326826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.00764006981626153 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a static, aerial view directly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a slow pan to follow the scarf's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lightweight silk scarf visible on a polished, dark mahogany tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scarf dyed in shades of lavender and dusty rose with fine, shimmering silver threads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a gentle draft lift a corner of the scarf, creating slow, weightless waves that spread across its surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638513274490833 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638513274490833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638513274490833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10697662830352783, + 0.03155509755015373, + 0.05424511060118675, + 0.060112934559583664, + 0.32823044061660767, + 0.1273195743560791, + 0.030182620510458946, + 0.09308446198701859, + 0.14319643378257751, + 0.025096720084547997 + ], + "credit_map": [ + 0.010789172723889351, + 0.014872057363390923, + 0.009460212662816048, + 0.010512564331293106, + 0.021149834617972374, + 0.012322048656642437, + 0.014446384273469448, + 0.013424346223473549, + 0.002461639465764165, + 0.0050277370028197765, + 0.0033813584595918655, + 0.003693670267239213, + 0.0042332191951572895, + 0.003980476874858141, + 0.0037015057168900967, + 0.005075489170849323, + 0.010161047801375389, + 0.014927715063095093, + 0.0035842605866491795, + 0.00540124299004674, + 0.00731092132627964, + 0.0045538004487752914, + 0.004138024523854256, + 0.004168092273175716, + 0.006947934627532959, + 0.005505164619535208, + 0.0055329278111457825, + 0.013366633094847202, + 0.007729401811957359, + 0.008106828667223454, + 0.0065284352749586105, + 0.006395597942173481, + 0.012174448929727077, + 0.07857529073953629, + 0.013947958126664162, + 0.020413657650351524, + 0.02217744290828705, + 0.012402782216668129, + 0.13827478885650635, + 0.03026406280696392, + 0.011330684646964073, + 0.00941267516463995, + 0.009970102459192276, + 0.017162002623081207, + 0.011095821857452393, + 0.024991940706968307, + 0.03462105989456177, + 0.008735285140573978, + 0.007300883997231722, + 0.0027575085405260324, + 0.007021293975412846, + 0.003743085078895092, + 0.002060109982267022, + 0.0023604733869433403, + 0.0027910710778087378, + 0.0021481902804225683, + 0.014667066745460033, + 0.01005504373461008, + 0.01653754897415638, + 0.013839318417012691, + 0.009946711361408234, + 0.011507445946335793, + 0.010032963007688522, + 0.006498364266008139, + 0.015433433465659618, + 0.024441247805953026, + 0.017687035724520683, + 0.02364329621195793, + 0.01332825981080532, + 0.01828816905617714, + 0.01583731546998024, + 0.014537664130330086, + 0.0046044522896409035, + 0.0017195377731695771, + 0.003249478992074728, + 0.004127036780118942, + 0.0016690085176378489, + 0.00269336742348969, + 0.0027463161386549473, + 0.00428751902654767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.008638513274490833 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_12/index_tensor([7671, 7671])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_12/index_tensor([7671, 7671])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9ac9807d9bd69f7b2e65f0ebe80a6f7cf487c234 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_12/index_tensor([7671, 7671])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "quick, slightly shaky first-person perspective shot, as if the camera holder has suddenly glanced back in alarm, locks onto a distant, black-and-white forest framed by towering birch trees. There, a snow-white fox lurks motionless, its pristine, fluffy fur striking against the dark background. Its body crouches low, its limbs tense, a poised and poised posture. Sharp light penetrates the sparse treetops, casting dappled, fragmented shadows on the damp, leaf-covered ground, heightening the scene's intensely graphic quality. The fox, a perfect ice sculpture, is pierced through the forest gloom by its amber eyes, gleaming with a watchful glint. Its motionless gaze pierces the camera, its pointed ears tilted forward to pick up any sound. The entire picture is stripped of all colors, leaving only pure black, white and gray, creating an extreme silence, high tension and a primitive, heart-pounding atmosphere of confrontation.", + "seed": 9322, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_12/index_7671_seed_9322_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05844517583027482, + "vlm_logprob": -0.05844517583027482, + "vlm_yes_no_margin": 3.499999976158142, + "vlm_token_credit": -0.05844517583027482, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1196214035153389 + ], + "opposite_logprob_tokens": [ + -2.244621515274048 + ], + "logprob_reward": -0.1196214035153389, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1196214035153389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18438097834587097, + 0.1482679545879364, + 0.07176190614700317, + 0.170765221118927, + 0.06554022431373596, + 0.1349201202392578, + 0.10978392511606216, + 0.041646551340818405, + 0.038902055472135544, + 0.034031011164188385 + ], + "credit_map": [ + 0.01840701512992382, + 0.0193728469312191, + 0.03508840501308441, + 0.028286417946219444, + 0.025303497910499573, + 0.015547814778983593, + 0.022686153650283813, + 0.019688833504915237, + 0.017103923484683037, + 0.016215389594435692, + 0.03615430369973183, + 0.012392259202897549, + 0.010073021054267883, + 0.010203609243035316, + 0.010760156437754631, + 0.03536528721451759, + 0.007531801238656044, + 0.007686732802540064, + 0.013784586451947689, + 0.00812323484569788, + 0.007410912308841944, + 0.006259560585021973, + 0.009596176445484161, + 0.011368902400135994, + 0.01571238785982132, + 0.013031563721597195, + 0.024120934307575226, + 0.01459500938653946, + 0.018235549330711365, + 0.01919008232653141, + 0.0369587317109108, + 0.028920967131853104, + 0.008877567946910858, + 0.00602510292083025, + 0.008998194709420204, + 0.0065167867578566074, + 0.008048555813729763, + 0.008634093217551708, + 0.009589729830622673, + 0.008850188925862312, + 0.01528911478817463, + 0.012182442471385002, + 0.019887184724211693, + 0.011756811290979385, + 0.018817482516169548, + 0.017739273607730865, + 0.022124215960502625, + 0.017123591154813766, + 0.012652341276407242, + 0.012480371631681919, + 0.01860538311302662, + 0.009822128340601921, + 0.010012845508754253, + 0.014304911717772484, + 0.016492558643221855, + 0.015413382090628147, + 0.004785140510648489, + 0.0033705714158713818, + 0.005877132061868906, + 0.003447930561378598, + 0.004846968222409487, + 0.004939395003020763, + 0.009001603350043297, + 0.005377807654440403, + 0.004825597163289785, + 0.0035697794519364834, + 0.007374988868832588, + 0.003904821118339896, + 0.003862689482048154, + 0.0039286185055971146, + 0.006823389325290918, + 0.00461217574775219, + 0.004136078525334597, + 0.002973517868667841, + 0.005674140062183142, + 0.0026646859478205442, + 0.0037601436488330364, + 0.006152092479169369, + 0.0038708732463419437, + 0.00479948241263628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.1196214035153389 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108936034142971 + ], + "opposite_logprob_tokens": [ + -4.511089324951172 + ], + "logprob_reward": -0.01108936034142971, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108936034142971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.245815709233284, + 0.16349968314170837, + 0.05399150028824806, + 0.04081789404153824, + 0.14110147953033447, + 0.1714312881231308, + 0.059162188321352005, + 0.0480613112449646, + 0.03543622046709061, + 0.040682703256607056 + ], + "credit_map": [ + 0.02673982083797455, + 0.031129151582717896, + 0.05114355683326721, + 0.016702765598893166, + 0.027795804664492607, + 0.02648979239165783, + 0.03221943974494934, + 0.03359537199139595, + 0.016273969784379005, + 0.015028406865894794, + 0.031673166900873184, + 0.012908507138490677, + 0.01866975799202919, + 0.016410984098911285, + 0.016726987436413765, + 0.03580789640545845, + 0.0055893538519740105, + 0.004417289048433304, + 0.01014980673789978, + 0.0035590175539255142, + 0.008072169497609138, + 0.00696791335940361, + 0.0056855762377381325, + 0.009550375863909721, + 0.004427762236446142, + 0.002683750120922923, + 0.00750358123332262, + 0.0019391825189813972, + 0.004890789277851582, + 0.007258241530507803, + 0.004430156666785479, + 0.0076844277791678905, + 0.031623292714357376, + 0.012085163034498692, + 0.026748886331915855, + 0.01203007623553276, + 0.010744265280663967, + 0.015177315101027489, + 0.009911749511957169, + 0.022780723869800568, + 0.03222455456852913, + 0.02287260629236698, + 0.026168720796704292, + 0.012602788396179676, + 0.013646559789776802, + 0.025052309036254883, + 0.015060421079397202, + 0.023803316056728363, + 0.005488836206495762, + 0.006274987477809191, + 0.008746406994760036, + 0.005237594712525606, + 0.007087194360792637, + 0.006791931111365557, + 0.012720488011837006, + 0.0068147508427500725, + 0.009377016685903072, + 0.0050294105894863605, + 0.009421885944902897, + 0.003126889932900667, + 0.003739179577678442, + 0.00565693061798811, + 0.005148355383425951, + 0.006561641581356525, + 0.0033491658978164196, + 0.0037326794117689133, + 0.006119875703006983, + 0.003752391319721937, + 0.0036934874951839447, + 0.003505379194393754, + 0.006745315622538328, + 0.0045379274524748325, + 0.005299049895256758, + 0.004545284900814295, + 0.007935515604913235, + 0.0031475303694605827, + 0.0032297195866703987, + 0.006703250575810671, + 0.004258173983544111, + 0.005564179737120867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.01108936034142971 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638150800019503 + ], + "opposite_logprob_tokens": [ + -5.379638195037842 + ], + "logprob_reward": -0.004638150800019503, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638150800019503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10533872246742249, + 0.22051209211349487, + 0.08176148682832718, + 0.09148235619068146, + 0.15232326090335846, + 0.08504947274923325, + 0.08817596733570099, + 0.09372851997613907, + 0.056928716599941254, + 0.024699319154024124 + ], + "credit_map": [ + 0.012602298520505428, + 0.007997489534318447, + 0.013977308757603168, + 0.019462766125798225, + 0.010513992980122566, + 0.010894599370658398, + 0.013320168480277061, + 0.016570094972848892, + 0.022011129185557365, + 0.013707609847187996, + 0.0632549375295639, + 0.021115917712450027, + 0.018331117928028107, + 0.015124780125916004, + 0.01782830059528351, + 0.04913830757141113, + 0.010350191034376621, + 0.011431630700826645, + 0.010726844891905785, + 0.012466104701161385, + 0.010301712900400162, + 0.00650188559666276, + 0.008504199795424938, + 0.011478911153972149, + 0.012549725361168385, + 0.0059861126355826855, + 0.015120076015591621, + 0.006922753993421793, + 0.009066849015653133, + 0.01273284014314413, + 0.014913047663867474, + 0.014190958812832832, + 0.01582791656255722, + 0.012811909429728985, + 0.01957375928759575, + 0.013964293524622917, + 0.01655271276831627, + 0.029255619272589684, + 0.023088620975613594, + 0.021248428151011467, + 0.011340719647705555, + 0.0062945191748440266, + 0.01352753210812807, + 0.005964373704046011, + 0.011033483780920506, + 0.013726550154387951, + 0.01180416252464056, + 0.011358129791915417, + 0.008575957268476486, + 0.009761791676282883, + 0.011386742815375328, + 0.008507379330694675, + 0.009730357676744461, + 0.011447329074144363, + 0.019091716036200523, + 0.00967468973249197, + 0.014936836436390877, + 0.007423547096550465, + 0.012902256101369858, + 0.005876996088773012, + 0.009426002390682697, + 0.012191345915198326, + 0.013123982585966587, + 0.017847556620836258, + 0.006336668506264687, + 0.00434041628614068, + 0.009963189251720905, + 0.004219074733555317, + 0.005642333533614874, + 0.006107596680521965, + 0.008149499073624611, + 0.012169940397143364, + 0.0033056028187274933, + 0.0020588121842592955, + 0.003989318385720253, + 0.0015672725858166814, + 0.002662027021870017, + 0.004137047566473484, + 0.0023873650934547186, + 0.0045918733812868595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.004638150800019503 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270793378353119 + ], + "opposite_logprob_tokens": [ + -2.1270792484283447 + ], + "logprob_reward": -0.1270793378353119, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270793378353119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1389743834733963, + 0.10673730075359344, + 0.09095633774995804, + 0.07104985415935516, + 0.31198370456695557, + 0.14926952123641968, + 0.0471867136657238, + 0.03713125362992287, + 0.0261977631598711, + 0.02051321603357792 + ], + "credit_map": [ + 0.023716455325484276, + 0.01486420352011919, + 0.022273536771535873, + 0.014294778928160667, + 0.012696614488959312, + 0.015266905538737774, + 0.01579141616821289, + 0.02007046900689602, + 0.014237216673791409, + 0.01162286102771759, + 0.01779060997068882, + 0.010635609738528728, + 0.010685938410460949, + 0.013160533271729946, + 0.01311561744660139, + 0.015488911420106888, + 0.009731017053127289, + 0.008885320276021957, + 0.017957815900444984, + 0.005297182127833366, + 0.011743893846869469, + 0.010136461816728115, + 0.01139349490404129, + 0.015811152756214142, + 0.006954766344279051, + 0.004783095791935921, + 0.011114222928881645, + 0.004454498644918203, + 0.009720390662550926, + 0.011516420170664787, + 0.009211217053234577, + 0.013295243494212627, + 0.0774010419845581, + 0.028122207149863243, + 0.05431198328733444, + 0.02466709166765213, + 0.022799748927354813, + 0.032620057463645935, + 0.02174440771341324, + 0.05031716823577881, + 0.024448275566101074, + 0.018194694072008133, + 0.026557467877864838, + 0.012530186213552952, + 0.011717465706169605, + 0.022077715024352074, + 0.012594519183039665, + 0.021149199455976486, + 0.008290125988423824, + 0.005061221309006214, + 0.009150122292339802, + 0.0028384337201714516, + 0.0030961360316723585, + 0.007275380659848452, + 0.004274464212357998, + 0.0072008296847343445, + 0.004120670724660158, + 0.004493411630392075, + 0.005605988670140505, + 0.0032226990442723036, + 0.002940094331279397, + 0.006087475456297398, + 0.004460279364138842, + 0.006200636737048626, + 0.0024026022292673588, + 0.0022263601422309875, + 0.0044744741171598434, + 0.002808019518852234, + 0.003117920132353902, + 0.0024696707259863615, + 0.004926609806716442, + 0.003772105323150754, + 0.003218948608264327, + 0.0017460836097598076, + 0.0034798281267285347, + 0.002311237156391144, + 0.0018234688322991133, + 0.0026428571436554193, + 0.002984452061355114, + 0.0023063416592776775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.1270793378353119 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0297976266592741 + ], + "opposite_logprob_tokens": [ + -3.5297975540161133 + ], + "logprob_reward": -0.0297976266592741, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0297976266592741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12578916549682617, + 0.09692041575908661, + 0.05739749222993851, + 0.045477818697690964, + 0.12413784116506577, + 0.21464785933494568, + 0.07675683498382568, + 0.13429118692874908, + 0.04617339372634888, + 0.07840792089700699 + ], + "credit_map": [ + 0.013027189299464226, + 0.0138859236612916, + 0.0174472127109766, + 0.022746937349438667, + 0.009006267413496971, + 0.009982023388147354, + 0.02097037434577942, + 0.018723251298069954, + 0.009362534619867802, + 0.015911025926470757, + 0.015439060516655445, + 0.010800999589264393, + 0.009732073172926903, + 0.009867041371762753, + 0.01152966357767582, + 0.014278030022978783, + 0.0061075733974576, + 0.009256651625037193, + 0.00769365020096302, + 0.007804178167134523, + 0.005882935598492622, + 0.004854213446378708, + 0.007582351565361023, + 0.008215945214033127, + 0.008518947288393974, + 0.003311451990157366, + 0.008282280527055264, + 0.0027815040666610003, + 0.004721785895526409, + 0.005769988987594843, + 0.004989352542907, + 0.007102510891854763, + 0.02283249981701374, + 0.009074537083506584, + 0.023359498009085655, + 0.01070323120802641, + 0.008873384445905685, + 0.01638069935142994, + 0.011274448595941067, + 0.02163955196738243, + 0.036224640905857086, + 0.019343167543411255, + 0.03898164629936218, + 0.018477141857147217, + 0.017308494076132774, + 0.03481645509600639, + 0.01859884150326252, + 0.030897485092282295, + 0.007183950860053301, + 0.00684055220335722, + 0.01392251905053854, + 0.0051687792874872684, + 0.006227212026715279, + 0.01278340257704258, + 0.010495568625628948, + 0.01413485873490572, + 0.014229873195290565, + 0.009229257702827454, + 0.023136887699365616, + 0.011564791202545166, + 0.009906299412250519, + 0.014696430414915085, + 0.021475899964571, + 0.030051756650209427, + 0.0074392203241586685, + 0.0040834988467395306, + 0.006417746655642986, + 0.004644836764782667, + 0.004003744572401047, + 0.006884006317704916, + 0.006808389909565449, + 0.005891955457627773, + 0.010370098985731602, + 0.005738166160881519, + 0.013983532786369324, + 0.007412170525640249, + 0.007367522455751896, + 0.007825121283531189, + 0.01143189799040556, + 0.014279414899647236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.0297976266592741 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "quick, slightly shaky first-person perspective shot, as if the camera holder has suddenly glanced back in alarm, locks onto a distant, black-and-white forest framed by towering birch trees. There, a snow-white fox lurks motionless, its pristine, fluffy fur striking against the dark background. Its body crouches low, its limbs tense, a poised and poised posture. Sharp light penetrates the sparse treetops, casting dappled, fragmented shadows on the damp, leaf-covered ground, heightening the scene's intensely graphic quality. The fox, a perfect ice sculpture, is pierced through the forest gloom by its amber eyes, gleaming with a watchful glint. Its motionless gaze pierces the camera, its pointed ears tilted forward to pick up any sound. The entire picture is stripped of all colors, leaving only pure black, white and gray, creating an extreme silence, high tension and a primitive, heart-pounding atmosphere of confrontation.", + "seed": 9323, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_12/index_7671_seed_9323_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04213382750749588, + "vlm_logprob": -0.04213382750749588, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.04213382750749588, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03163779899477959 + ], + "opposite_logprob_tokens": [ + -3.6566379070281982 + ], + "logprob_reward": -0.03163779899477959, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03163779899477959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17281371355056763, + 0.03280489891767502, + 0.027926912531256676, + 0.032891299575567245, + 0.04888349398970604, + 0.3112338185310364, + 0.08851837366819382, + 0.1975613683462143, + 0.07025668025016785, + 0.01710946299135685 + ], + "credit_map": [ + 0.019984638318419456, + 0.011365730315446854, + 0.01809769682586193, + 0.01253565400838852, + 0.019647836685180664, + 0.04324834793806076, + 0.01905212365090847, + 0.02888166904449463, + 0.0036414782516658306, + 0.003759876824915409, + 0.004573619458824396, + 0.004645826760679483, + 0.0042412797920405865, + 0.004037948790937662, + 0.004046875052154064, + 0.003857993520796299, + 0.0026116727385669947, + 0.0036728340201079845, + 0.0035818801261484623, + 0.003625453682616353, + 0.004020116291940212, + 0.003876797156408429, + 0.003319891868159175, + 0.003218265948817134, + 0.002907375805079937, + 0.0036308388225734234, + 0.0031646722927689552, + 0.003454015124589205, + 0.0036659601610153913, + 0.004508528858423233, + 0.00662775244563818, + 0.004932156298309565, + 0.006530482321977615, + 0.004796491004526615, + 0.008568857796490192, + 0.005383238662034273, + 0.00579957477748394, + 0.004991348832845688, + 0.008374731987714767, + 0.0044387695379555225, + 0.03531384840607643, + 0.033655088394880295, + 0.05031860992312431, + 0.06200205162167549, + 0.03616538643836975, + 0.03512191399931908, + 0.02514263428747654, + 0.03351426497101784, + 0.011829899623990059, + 0.009640447795391083, + 0.00728060957044363, + 0.01436601672321558, + 0.010172922164201736, + 0.012185469269752502, + 0.013520737178623676, + 0.00952227134257555, + 0.038189955055713654, + 0.018010804429650307, + 0.018986258655786514, + 0.0233125239610672, + 0.016992179676890373, + 0.034036505967378616, + 0.022585321217775345, + 0.02544783428311348, + 0.014461684040725231, + 0.006324131041765213, + 0.007463355548679829, + 0.007479276042431593, + 0.007403410505503416, + 0.009776346385478973, + 0.007656749803572893, + 0.009691732935607433, + 0.0030932663939893246, + 0.001250068424269557, + 0.002402359386906028, + 0.0019274483202025294, + 0.001609757891856134, + 0.0028582352679222822, + 0.002163779688999057, + 0.0018045470351353288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.03163779899477959 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329809032380581 + ], + "opposite_logprob_tokens": [ + -3.7732980251312256 + ], + "logprob_reward": -0.02329809032380581, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329809032380581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2185734212398529, + 0.08835798501968384, + 0.05038322135806084, + 0.07502099871635437, + 0.04569288715720177, + 0.22760753333568573, + 0.06598740071058273, + 0.1250728964805603, + 0.08407437056303024, + 0.019229255616664886 + ], + "credit_map": [ + 0.01659216172993183, + 0.014696072787046432, + 0.03858981281518936, + 0.02452296018600464, + 0.023555772379040718, + 0.02898843213915825, + 0.03766951337456703, + 0.033958714455366135, + 0.01561111118644476, + 0.010724913328886032, + 0.009130612947046757, + 0.01528279297053814, + 0.009760153479874134, + 0.012417498044669628, + 0.007911561988294125, + 0.007519342005252838, + 0.007658981718122959, + 0.005760414060205221, + 0.006653022486716509, + 0.005627082660794258, + 0.00535487150773406, + 0.008669642731547356, + 0.005486737471073866, + 0.0051724729128181934, + 0.007704147137701511, + 0.009587440639734268, + 0.008616741746664047, + 0.008284194394946098, + 0.01058867946267128, + 0.00808868557214737, + 0.011862906627357006, + 0.010288204997777939, + 0.006512912455946207, + 0.004038726910948753, + 0.007112276740372181, + 0.00503833731636405, + 0.004797859583050013, + 0.005252072587609291, + 0.008423896506428719, + 0.004516805987805128, + 0.027348432689905167, + 0.022182798013091087, + 0.03623853623867035, + 0.045653313398361206, + 0.02619374543428421, + 0.026959285140037537, + 0.01983347348868847, + 0.023197956383228302, + 0.008965659886598587, + 0.007475560065358877, + 0.010122977197170258, + 0.009461979381740093, + 0.0060027833096683025, + 0.010223998688161373, + 0.008206472732126713, + 0.005527976900339127, + 0.025264037773013115, + 0.010267364792525768, + 0.013303207233548164, + 0.01426506694406271, + 0.009617803618311882, + 0.020793285220861435, + 0.011762351728975773, + 0.01979978382587433, + 0.015824947506189346, + 0.006552505306899548, + 0.01200089417397976, + 0.007871467620134354, + 0.007954912260174751, + 0.014280527830123901, + 0.009326289407908916, + 0.010262827388942242, + 0.0036867731250822544, + 0.001121677691116929, + 0.003664580639451742, + 0.00207139877602458, + 0.001327943173237145, + 0.0030798364896327257, + 0.0012953480472788215, + 0.0029816990718245506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.02329809032380581 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067370180040597916 + ], + "opposite_logprob_tokens": [ + -5.006737232208252 + ], + "logprob_reward": -0.0067370180040597916, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067370180040597916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13345788419246674, + 0.05584443360567093, + 0.07531074434518814, + 0.06856536120176315, + 0.04683247208595276, + 0.33939141035079956, + 0.10115797817707062, + 0.06930004060268402, + 0.09517192095518112, + 0.014967733062803745 + ], + "credit_map": [ + 0.012385564856231213, + 0.021735375747084618, + 0.015324217267334461, + 0.020054500550031662, + 0.018492983654141426, + 0.017263291403651237, + 0.011791015043854713, + 0.016410933807492256, + 0.0063340868800878525, + 0.007562196347862482, + 0.004994637798517942, + 0.012983801774680614, + 0.005348746199160814, + 0.008919932879507542, + 0.005417110864073038, + 0.004283920396119356, + 0.005917403381317854, + 0.008946739137172699, + 0.007382598705589771, + 0.008350667543709278, + 0.010007907636463642, + 0.011563226580619812, + 0.012590519152581692, + 0.010551686398684978, + 0.005431895609945059, + 0.00908349547535181, + 0.008159076794981956, + 0.008247128687798977, + 0.00985807180404663, + 0.00829944759607315, + 0.012516205199062824, + 0.0069700428284704685, + 0.006172124296426773, + 0.005511248484253883, + 0.006549564655870199, + 0.007784000597894192, + 0.004306944552809, + 0.004799415823072195, + 0.006925380788743496, + 0.00478379474952817, + 0.038407102227211, + 0.04265768453478813, + 0.043349094688892365, + 0.060175612568855286, + 0.026146987453103065, + 0.0497036911547184, + 0.043560922145843506, + 0.03539032116532326, + 0.013780499808490276, + 0.0076803239062428474, + 0.0071327053010463715, + 0.012692847289144993, + 0.011275939643383026, + 0.013985015451908112, + 0.019648373126983643, + 0.014962280169129372, + 0.011013353243470192, + 0.006304142065346241, + 0.008465699851512909, + 0.009897453710436821, + 0.006928571965545416, + 0.01043917890638113, + 0.008117740042507648, + 0.008133902214467525, + 0.018839670345187187, + 0.0071419584564864635, + 0.014206988736987114, + 0.01051273848861456, + 0.007922726683318615, + 0.013679678551852703, + 0.009294888004660606, + 0.013573271222412586, + 0.003202851163223386, + 0.0008851918391883373, + 0.0021466773469001055, + 0.0011725869262591004, + 0.0010708235204219818, + 0.0035078534856438637, + 0.0010230076732113957, + 0.0019587406422942877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.0067370180040597916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034213215112686 + ], + "opposite_logprob_tokens": [ + -2.350342035293579 + ], + "logprob_reward": -0.10034213215112686, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034213215112686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1327318549156189, + 0.0782901868224144, + 0.13445308804512024, + 0.11490139365196228, + 0.053960688412189484, + 0.23650908470153809, + 0.08536333590745926, + 0.10422627627849579, + 0.03104224242269993, + 0.028521908447146416 + ], + "credit_map": [ + 0.013670987449586391, + 0.00846046768128872, + 0.025137383490800858, + 0.015022987499833107, + 0.016777915880084038, + 0.02119632065296173, + 0.014684976078569889, + 0.01778082363307476, + 0.011706026270985603, + 0.011930201202630997, + 0.0070177349261939526, + 0.006886352319270372, + 0.00952964648604393, + 0.017522910609841347, + 0.00636496813967824, + 0.007332347333431244, + 0.022909730672836304, + 0.014522206969559193, + 0.02108105458319187, + 0.01270491536706686, + 0.013409732840955257, + 0.025889238342642784, + 0.009164167568087578, + 0.014772040769457817, + 0.015780877321958542, + 0.009089087136089802, + 0.01372662652283907, + 0.008111565373837948, + 0.005777192767709494, + 0.017410309985280037, + 0.008515342138707638, + 0.03649039566516876, + 0.011992597952485085, + 0.005380956921726465, + 0.007220441475510597, + 0.0052627515979111195, + 0.004710156936198473, + 0.007296054158359766, + 0.006267675664275885, + 0.0058300551027059555, + 0.02946712262928486, + 0.023649249225854874, + 0.03874886408448219, + 0.04676847159862518, + 0.02597867324948311, + 0.030083712190389633, + 0.018138058483600616, + 0.02367492951452732, + 0.015859302133321762, + 0.008314988575875759, + 0.015220188535749912, + 0.00887399073690176, + 0.005208432674407959, + 0.014477074146270752, + 0.005260415840893984, + 0.012148943729698658, + 0.019422177225351334, + 0.008444756269454956, + 0.013960931450128555, + 0.01187356561422348, + 0.008195945993065834, + 0.014581163413822651, + 0.01035828422755003, + 0.0173894502222538, + 0.006227629259228706, + 0.0027816719375550747, + 0.0029191854409873486, + 0.003829789115116, + 0.0033855491783469915, + 0.004807691555470228, + 0.003548090113326907, + 0.003542634891346097, + 0.006666857283562422, + 0.001917071989737451, + 0.0029236660338938236, + 0.0025660982355475426, + 0.0018755781929939985, + 0.006083481479436159, + 0.002060315804556012, + 0.00442883837968111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.10034213215112686 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865409806370735 + ], + "opposite_logprob_tokens": [ + -3.048654079437256 + ], + "logprob_reward": -0.04865409806370735, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865409806370735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17306984961032867, + 0.0671805739402771, + 0.03197705000638962, + 0.10368599742650986, + 0.10841574519872665, + 0.19347813725471497, + 0.17149999737739563, + 0.06339649856090546, + 0.06070602685213089, + 0.026590246707201004 + ], + "credit_map": [ + 0.01658092625439167, + 0.011626733466982841, + 0.018998093903064728, + 0.017665402963757515, + 0.0162747073918581, + 0.04231493920087814, + 0.02010921761393547, + 0.029499785974621773, + 0.008673016913235188, + 0.007851338014006615, + 0.004805733449757099, + 0.011146989651024342, + 0.006415958981961012, + 0.01731330156326294, + 0.004647610709071159, + 0.006326611153781414, + 0.003401880618184805, + 0.003932863473892212, + 0.004346709232777357, + 0.002839381340891123, + 0.004155746661126614, + 0.005179258529096842, + 0.004915651865303516, + 0.003205551067367196, + 0.009076305665075779, + 0.010747001506388187, + 0.011450408957898617, + 0.009537462145090103, + 0.013222452253103256, + 0.012850451283156872, + 0.01989925466477871, + 0.01690264232456684, + 0.02318795770406723, + 0.010165831074118614, + 0.016296936199069023, + 0.0076520731672644615, + 0.010704749263823032, + 0.015886742621660233, + 0.013294854201376438, + 0.011226583272218704, + 0.021887382492423058, + 0.020241836085915565, + 0.031222134828567505, + 0.037070952355861664, + 0.021269556134939194, + 0.023386752232909203, + 0.017675815150141716, + 0.02072366140782833, + 0.025165049359202385, + 0.019707947969436646, + 0.01570284366607666, + 0.022048035636544228, + 0.02047593519091606, + 0.023249799385666847, + 0.025126593187451363, + 0.020023752003908157, + 0.011210869066417217, + 0.006182544864714146, + 0.009515894576907158, + 0.0070679811760783195, + 0.006085073109716177, + 0.011468937620520592, + 0.00818586628884077, + 0.0036793206818401814, + 0.012424888089299202, + 0.004913954995572567, + 0.009514044038951397, + 0.006893298588693142, + 0.00473281042650342, + 0.009683660231530666, + 0.005468854680657387, + 0.007074505556374788, + 0.004299665801227093, + 0.0024013465736061335, + 0.003105801297351718, + 0.003633739659562707, + 0.0021975559648126364, + 0.0043222480453550816, + 0.0028811923693865538, + 0.00374869117513299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.04865409806370735 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "quick, slightly shaky first-person perspective shot, as if the camera holder has suddenly glanced back in alarm, locks onto a distant, black-and-white forest framed by towering birch trees. There, a snow-white fox lurks motionless, its pristine, fluffy fur striking against the dark background. Its body crouches low, its limbs tense, a poised and poised posture. Sharp light penetrates the sparse treetops, casting dappled, fragmented shadows on the damp, leaf-covered ground, heightening the scene's intensely graphic quality. The fox, a perfect ice sculpture, is pierced through the forest gloom by its amber eyes, gleaming with a watchful glint. Its motionless gaze pierces the camera, its pointed ears tilted forward to pick up any sound. The entire picture is stripped of all colors, leaving only pure black, white and gray, creating an extreme silence, high tension and a primitive, heart-pounding atmosphere of confrontation.", + "seed": 9324, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_12/index_7671_seed_9324_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05919556031003594, + "vlm_logprob": -0.05919556031003594, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.05919556031003594, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07864636182785034 + ], + "opposite_logprob_tokens": [ + -2.703646421432495 + ], + "logprob_reward": -0.07864636182785034, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07864636182785034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05605780705809593, + 0.06243574991822243, + 0.032556477934122086, + 0.12030737102031708, + 0.01952989585697651, + 0.39643529057502747, + 0.037776704877614975, + 0.037175923585891724, + 0.06951053440570831, + 0.16821424663066864 + ], + "credit_map": [ + 0.016402050852775574, + 0.003302568569779396, + 0.00999602023512125, + 0.0037458380684256554, + 0.006028537172824144, + 0.00483487406745553, + 0.004850480239838362, + 0.006897439248859882, + 0.00940616149455309, + 0.005146017763763666, + 0.010799413546919823, + 0.005712065380066633, + 0.007673750165849924, + 0.00708540016785264, + 0.008729123510420322, + 0.00788381602615118, + 0.006684795953333378, + 0.003562917234376073, + 0.004348029848188162, + 0.003483339212834835, + 0.0031499438919126987, + 0.003537553595378995, + 0.003567825071513653, + 0.004222072195261717, + 0.013368630781769753, + 0.005893457680940628, + 0.03433400020003319, + 0.009118806570768356, + 0.0069313328713178635, + 0.014185132458806038, + 0.009304404258728027, + 0.027171602472662926, + 0.0027295604813843966, + 0.002894993405789137, + 0.002751984167844057, + 0.0023281772155314684, + 0.0019450264517217875, + 0.0028631980530917645, + 0.002399517921730876, + 0.0016174394404515624, + 0.034724969416856766, + 0.16217459738254547, + 0.015891147777438164, + 0.008867963217198849, + 0.10566385090351105, + 0.019407721236348152, + 0.013758460991084576, + 0.03594658523797989, + 0.004326092079281807, + 0.006751775275915861, + 0.0037772073410451412, + 0.005143597722053528, + 0.004583264235407114, + 0.004138822201639414, + 0.004358098842203617, + 0.004697848577052355, + 0.0035950050223618746, + 0.00895755272358656, + 0.002433660440146923, + 0.002149263396859169, + 0.004416090901941061, + 0.002999588381499052, + 0.0035234398674219847, + 0.009101320058107376, + 0.005652768071740866, + 0.007765662390738726, + 0.004884170368313789, + 0.0057728346437215805, + 0.009130004793405533, + 0.010513348504900932, + 0.010978986509144306, + 0.014812752604484558, + 0.01650104857981205, + 0.02165945991873741, + 0.009944044053554535, + 0.011412858963012695, + 0.03492114320397377, + 0.018010245636105537, + 0.017035620287060738, + 0.03872983902692795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 3 + ], + "selected_reward": -0.07864636182785034 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302283138036728 + ], + "opposite_logprob_tokens": [ + -3.7733023166656494 + ], + "logprob_reward": -0.023302283138036728, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302283138036728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07249492406845093, + 0.1606990545988083, + 0.051704466342926025, + 0.03805757686495781, + 0.0747726634144783, + 0.3963445723056793, + 0.03520204499363899, + 0.0444137342274189, + 0.03794560208916664, + 0.08836538344621658 + ], + "credit_map": [ + 0.02101920172572136, + 0.004265085328370333, + 0.008423777297139168, + 0.006212251260876656, + 0.004969248082488775, + 0.008217277936637402, + 0.006688264198601246, + 0.012699811719357967, + 0.03233276680111885, + 0.0170558150857687, + 0.0286839809268713, + 0.012877625413239002, + 0.009362422861158848, + 0.025567984208464622, + 0.008510121144354343, + 0.026308342814445496, + 0.006143487989902496, + 0.006293515209108591, + 0.006800143513828516, + 0.004019985441118479, + 0.007638116367161274, + 0.01128377765417099, + 0.005043809302151203, + 0.004481630399823189, + 0.004486049059778452, + 0.0036784601397812366, + 0.007569515611976385, + 0.003334491280838847, + 0.0048496234230697155, + 0.0058805630542337894, + 0.002612157491967082, + 0.005646713078022003, + 0.013364833779633045, + 0.01193208061158657, + 0.006898594554513693, + 0.003891188884153962, + 0.011439688503742218, + 0.015110661275684834, + 0.003761155763641, + 0.008374459110200405, + 0.03261224925518036, + 0.16077999770641327, + 0.017388056963682175, + 0.010769405402243137, + 0.10126468539237976, + 0.016471179202198982, + 0.015404541976749897, + 0.04165444150567055, + 0.00325393071398139, + 0.006257058121263981, + 0.0038745964411646128, + 0.0037133723963052034, + 0.004593897145241499, + 0.0025343282613903284, + 0.004613207187503576, + 0.006361654028296471, + 0.0029975352808833122, + 0.01270307507365942, + 0.0033209440298378468, + 0.0032420833595097065, + 0.008694643154740334, + 0.0035775629803538322, + 0.0046594650484621525, + 0.005218422971665859, + 0.003797173732891679, + 0.005721484310925007, + 0.0031384718604385853, + 0.003246483625844121, + 0.007485882844775915, + 0.004587539006024599, + 0.00368840666487813, + 0.0062801591120660305, + 0.005991664249449968, + 0.0184219591319561, + 0.007070318330079317, + 0.009168760851025581, + 0.015600358135998249, + 0.009319508448243141, + 0.010858791880309582, + 0.011934015899896622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 9 + ], + "selected_reward": -0.023302283138036728 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463898153975606 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.00463898153975606, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463898153975606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047219935804605484, + 0.06363101303577423, + 0.14487525820732117, + 0.09297122806310654, + 0.0433468334376812, + 0.3219698667526245, + 0.1353866010904312, + 0.10348950326442719, + 0.030482223257422447, + 0.016627579927444458 + ], + "credit_map": [ + 0.00623773317784071, + 0.0024655996821820736, + 0.011705164797604084, + 0.004600744228810072, + 0.005216539837419987, + 0.006177731789648533, + 0.004932134412229061, + 0.0058842869475483894, + 0.01238429918885231, + 0.004938825499266386, + 0.010798403061926365, + 0.006745357532054186, + 0.008588668890297413, + 0.005453349091112614, + 0.008586883544921875, + 0.006135222036391497, + 0.032911259680986404, + 0.012097703292965889, + 0.02144932746887207, + 0.013708001002669334, + 0.01135235745459795, + 0.02091672457754612, + 0.01607176661491394, + 0.016368111595511436, + 0.010111281648278236, + 0.004718967247754335, + 0.02779330313205719, + 0.006222837138921022, + 0.0048448024317622185, + 0.01009035762399435, + 0.005673836451023817, + 0.023515844717621803, + 0.01052782591432333, + 0.003548645181581378, + 0.008316810242831707, + 0.005152516532689333, + 0.0031924396753311157, + 0.003930277191102505, + 0.004160241689532995, + 0.0045180730521678925, + 0.027499660849571228, + 0.1302778571844101, + 0.013622161000967026, + 0.007268731947988272, + 0.08295434713363647, + 0.014913379214704037, + 0.012208430096507072, + 0.03322528302669525, + 0.012171046808362007, + 0.01732419617474079, + 0.012886252254247665, + 0.011530122719705105, + 0.023285221308469772, + 0.01199811790138483, + 0.022771261632442474, + 0.023420367389917374, + 0.006353678647428751, + 0.03547877445816994, + 0.004701030906289816, + 0.00658529344946146, + 0.01888158544898033, + 0.005857559852302074, + 0.010044335387647152, + 0.015587248839437962, + 0.0031744339503347874, + 0.00463398452848196, + 0.0031106232199817896, + 0.0027421871200203896, + 0.0038690569344908, + 0.00513111986219883, + 0.00455405842512846, + 0.003266755724325776, + 0.001685824361629784, + 0.0022440978791564703, + 0.0011413355823606253, + 0.0012557633453980088, + 0.0033185624051839113, + 0.00191345508210361, + 0.0017729219980537891, + 0.0032956188078969717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.00463898153975606 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907359838485718 + ], + "opposite_logprob_tokens": [ + -2.464073657989502 + ], + "logprob_reward": -0.08907359838485718, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907359838485718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17894767224788666, + 0.10681045800447464, + 0.21408338844776154, + 0.09633056074380875, + 0.03276101499795914, + 0.1105855330824852, + 0.0606563426554203, + 0.08767162263393402, + 0.04968079924583435, + 0.06247260421514511 + ], + "credit_map": [ + 0.03682759776711464, + 0.012335315346717834, + 0.022431140765547752, + 0.013815163634717464, + 0.02561808004975319, + 0.023266371339559555, + 0.01414917130023241, + 0.030504828318953514, + 0.021477025002241135, + 0.009058685973286629, + 0.016065707430243492, + 0.008149892091751099, + 0.012451771646738052, + 0.016417982056736946, + 0.008389928378164768, + 0.014799464493989944, + 0.04874509200453758, + 0.01761331781744957, + 0.031041355803608894, + 0.019304154440760612, + 0.019402354955673218, + 0.03151339292526245, + 0.02316899411380291, + 0.023294731974601746, + 0.010226309299468994, + 0.005648071877658367, + 0.02616073563694954, + 0.006985065061599016, + 0.004859015345573425, + 0.012486614286899567, + 0.006520516704767942, + 0.023444227874279022, + 0.004411056637763977, + 0.0048270742408931255, + 0.0025440927129238844, + 0.0016878402093425393, + 0.004307003226131201, + 0.006138644181191921, + 0.0015049419598653913, + 0.007340361829847097, + 0.01148957572877407, + 0.036789312958717346, + 0.005369420163333416, + 0.0041354261338710785, + 0.022128690034151077, + 0.010892841033637524, + 0.005047326907515526, + 0.014732945710420609, + 0.00523800402879715, + 0.015380692668259144, + 0.005320515483617783, + 0.004888188559561968, + 0.011635304428637028, + 0.005007999017834663, + 0.007129777688533068, + 0.00605586264282465, + 0.008237017318606377, + 0.010396397672593594, + 0.006768778432160616, + 0.00511559285223484, + 0.015297927893698215, + 0.010727995075285435, + 0.007781672291457653, + 0.02334623970091343, + 0.005220538005232811, + 0.009067731909453869, + 0.004323701374232769, + 0.0049265166744589806, + 0.009487731382250786, + 0.005830470472574234, + 0.005524157080799341, + 0.005299947690218687, + 0.004274318926036358, + 0.01454379502683878, + 0.005609358660876751, + 0.006623315159231424, + 0.009949966333806515, + 0.007533588912338018, + 0.00664688553661108, + 0.007291373331099749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.08907359838485718 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031657665967941 + ], + "opposite_logprob_tokens": [ + -2.3503165245056152 + ], + "logprob_reward": -0.10031657665967941, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031657665967941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06255552172660828, + 0.07248292863368988, + 0.15728570520877838, + 0.04941365122795105, + 0.019376538693904877, + 0.11989834159612656, + 0.1346718668937683, + 0.14237405359745026, + 0.19130536913871765, + 0.05063606798648834 + ], + "credit_map": [ + 0.01708689145743847, + 0.0038881960790604353, + 0.006589927710592747, + 0.0055955382995307446, + 0.005300017539411783, + 0.0056659081019461155, + 0.006542033515870571, + 0.011887004598975182, + 0.011187438853085041, + 0.006171061657369137, + 0.01789386197924614, + 0.005842254031449556, + 0.0041509876027703285, + 0.0075990199111402035, + 0.004702757578343153, + 0.01493554562330246, + 0.035525184124708176, + 0.012629043310880661, + 0.022548384964466095, + 0.015999337658286095, + 0.011743324808776379, + 0.022446582093834877, + 0.01752532087266445, + 0.01886853016912937, + 0.005316811613738537, + 0.0034088592510670424, + 0.01264860574156046, + 0.005029503256082535, + 0.0040654465556144714, + 0.0054841358214616776, + 0.0030209480319172144, + 0.010439339093863964, + 0.0033476653043180704, + 0.002280029933899641, + 0.004346462897956371, + 0.0018842329736799002, + 0.00169559835921973, + 0.0017289770767092705, + 0.0011203908361494541, + 0.002973180031403899, + 0.010419716127216816, + 0.04346437752246857, + 0.00417014816775918, + 0.004183849319815636, + 0.02550136297941208, + 0.007697113789618015, + 0.005865014158189297, + 0.018596768379211426, + 0.009278390556573868, + 0.05141401290893555, + 0.006225020624697208, + 0.006232616491615772, + 0.036482423543930054, + 0.00806897971779108, + 0.009890170767903328, + 0.007080244831740856, + 0.01010273676365614, + 0.040031563490629196, + 0.011106221936643124, + 0.010219557210803032, + 0.022015368565917015, + 0.011418967507779598, + 0.0219293013215065, + 0.01555032841861248, + 0.01771010272204876, + 0.03339935839176178, + 0.017660966143012047, + 0.01747245527803898, + 0.019172772765159607, + 0.041267260909080505, + 0.024662595242261887, + 0.01995987631380558, + 0.0038019632920622826, + 0.011277555488049984, + 0.004802045412361622, + 0.006299659144133329, + 0.0074165454134345055, + 0.005625233519822359, + 0.006034952588379383, + 0.005378111731261015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.10031657665967941 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "quick, slightly shaky first-person perspective shot, as if the camera holder has suddenly glanced back in alarm, locks onto a distant, black-and-white forest framed by towering birch trees. There, a snow-white fox lurks motionless, its pristine, fluffy fur striking against the dark background. Its body crouches low, its limbs tense, a poised and poised posture. Sharp light penetrates the sparse treetops, casting dappled, fragmented shadows on the damp, leaf-covered ground, heightening the scene's intensely graphic quality. The fox, a perfect ice sculpture, is pierced through the forest gloom by its amber eyes, gleaming with a watchful glint. Its motionless gaze pierces the camera, its pointed ears tilted forward to pick up any sound. The entire picture is stripped of all colors, leaving only pure black, white and gray, creating an extreme silence, high tension and a primitive, heart-pounding atmosphere of confrontation.", + "seed": 9325, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_12/index_7671_seed_9325_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04570482848212123, + "vlm_logprob": -0.04570482848212123, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.04570482848212123, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0635363757610321 + ], + "opposite_logprob_tokens": [ + -2.9385364055633545 + ], + "logprob_reward": -0.0635363757610321, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0635363757610321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22355937957763672, + 0.04347125440835953, + 0.054256170988082886, + 0.03860573098063469, + 0.13918481767177582, + 0.2681396007537842, + 0.05199812352657318, + 0.07543051242828369, + 0.06225517392158508, + 0.04309920594096184 + ], + "credit_map": [ + 0.06273459643125534, + 0.01077736634761095, + 0.01789969392120838, + 0.01624324731528759, + 0.014322854578495026, + 0.015776975080370903, + 0.06061778962612152, + 0.025186864659190178, + 0.010304088704288006, + 0.0035252340603619814, + 0.005337572656571865, + 0.0042792996391654015, + 0.004446765873581171, + 0.00798199325799942, + 0.003449346637353301, + 0.004146954510360956, + 0.012208688072860241, + 0.004275473766028881, + 0.005946559365838766, + 0.0063612512312829494, + 0.0069015915505588055, + 0.008526986464858055, + 0.005023628938943148, + 0.005011986009776592, + 0.004125345032662153, + 0.005225103348493576, + 0.004982030484825373, + 0.004130006767809391, + 0.0039212447591125965, + 0.006190483458340168, + 0.004605286289006472, + 0.005426231771707535, + 0.016667276620864868, + 0.012559368275105953, + 0.04959589242935181, + 0.011899719014763832, + 0.006070729345083237, + 0.007279505487531424, + 0.00659153750166297, + 0.02852078340947628, + 0.07819462567567825, + 0.01640995219349861, + 0.015839286148548126, + 0.09358584135770798, + 0.023028314113616943, + 0.01421936135739088, + 0.015261072665452957, + 0.011601153761148453, + 0.006433269940316677, + 0.005305267870426178, + 0.006170002743601799, + 0.007973493076860905, + 0.005373020190745592, + 0.005345270037651062, + 0.007199933286756277, + 0.008197863586246967, + 0.0038628089241683483, + 0.005141673609614372, + 0.0029915538616478443, + 0.02280249260365963, + 0.026548761874437332, + 0.004501394461840391, + 0.00512196309864521, + 0.004459865391254425, + 0.0067315478809177876, + 0.008469246327877045, + 0.005441061686724424, + 0.007898328825831413, + 0.0085742874071002, + 0.00678799906745553, + 0.00787641853094101, + 0.010476280003786087, + 0.006198267452418804, + 0.007832394912838936, + 0.004157276824116707, + 0.005151768215000629, + 0.0063193426467478275, + 0.004578524734824896, + 0.005712656769901514, + 0.0031489753164350986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0635363757610321 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033717118203639984 + ], + "opposite_logprob_tokens": [ + -3.408717155456543 + ], + "logprob_reward": -0.033717118203639984, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033717118203639984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26633062958717346, + 0.09062513709068298, + 0.041135940700769424, + 0.022372590377926826, + 0.048707231879234314, + 0.2819143831729889, + 0.050764232873916626, + 0.1462009698152542, + 0.019549774006009102, + 0.03239915519952774 + ], + "credit_map": [ + 0.07147214561700821, + 0.012899215333163738, + 0.021568525582551956, + 0.02086985856294632, + 0.01722012832760811, + 0.023466378450393677, + 0.07215868681669235, + 0.026675673201680183, + 0.016886359080672264, + 0.0068505494855344296, + 0.009728019125759602, + 0.008805260062217712, + 0.014410525560379028, + 0.018869100138545036, + 0.007152383681386709, + 0.007922941818833351, + 0.00849792268127203, + 0.0036113872192800045, + 0.006202700547873974, + 0.004795453045517206, + 0.003913847729563713, + 0.005978252738714218, + 0.0028120935894548893, + 0.005324283614754677, + 0.0021206880919635296, + 0.0035338953603059053, + 0.0028938883915543556, + 0.002663388615474105, + 0.0038961435202509165, + 0.0032266320195049047, + 0.002161255804821849, + 0.0018766988068819046, + 0.005828034598380327, + 0.004283054731786251, + 0.015709208324551582, + 0.004568550270050764, + 0.003979145083576441, + 0.003272728528827429, + 0.003196852281689644, + 0.00786965899169445, + 0.08640465885400772, + 0.018705107271671295, + 0.016416367143392563, + 0.09734746813774109, + 0.024348432198166847, + 0.014026880264282227, + 0.01583196222782135, + 0.008833477273583412, + 0.004510917235165834, + 0.004616162274032831, + 0.006261125206947327, + 0.010242131538689137, + 0.007112691178917885, + 0.004753328859806061, + 0.009608604945242405, + 0.003659276058897376, + 0.007723595015704632, + 0.008911862038075924, + 0.002511371625587344, + 0.05235026776790619, + 0.05784069374203682, + 0.009322063997387886, + 0.0050992099568247795, + 0.002441897289827466, + 0.00211564963683486, + 0.0029825344681739807, + 0.0011861162493005395, + 0.002436945680528879, + 0.003474486293271184, + 0.002803957089781761, + 0.002741327276453376, + 0.0018087561475113034, + 0.0053512295708060265, + 0.004264718387275934, + 0.00246458500623703, + 0.004321232438087463, + 0.004896881524473429, + 0.0034100383054465055, + 0.004044033586978912, + 0.003646436147391796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.033717118203639984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007635810878127813 + ], + "opposite_logprob_tokens": [ + -4.882635593414307 + ], + "logprob_reward": -0.007635810878127813, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007635810878127813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14814426004886627, + 0.012612416408956051, + 0.02800736017525196, + 0.01540373358875513, + 0.12781383097171783, + 0.3976523280143738, + 0.0402710847556591, + 0.1722232550382614, + 0.02513977512717247, + 0.03273198381066322 + ], + "credit_map": [ + 0.04441292956471443, + 0.006886462215334177, + 0.011249646544456482, + 0.00877279695123434, + 0.010306895710527897, + 0.010678672231733799, + 0.04083583876490593, + 0.015000984072685242, + 0.0021077084820717573, + 0.0008186528575606644, + 0.0016090287826955318, + 0.001730601885356009, + 0.0014368236297741532, + 0.0018097073771059513, + 0.001247749081812799, + 0.001852142857387662, + 0.004142330959439278, + 0.001676197163760662, + 0.005698462948203087, + 0.002571996534243226, + 0.0025641529355198145, + 0.0048790499567985535, + 0.0021665222011506557, + 0.004308645613491535, + 0.0015023350715637207, + 0.0013958996860310435, + 0.002643830142915249, + 0.002416009781882167, + 0.0015130932442843914, + 0.0022651858162134886, + 0.001690867356956005, + 0.001976512372493744, + 0.016118761152029037, + 0.01223317626863718, + 0.047897372394800186, + 0.008810932748019695, + 0.0034048063680529594, + 0.005037560593336821, + 0.005764954723417759, + 0.028546253219246864, + 0.11989027261734009, + 0.026445334777235985, + 0.022622661665081978, + 0.13843458890914917, + 0.03465212136507034, + 0.020218130201101303, + 0.022681672126054764, + 0.012707498855888844, + 0.0029577442910522223, + 0.0032774233259260654, + 0.006132649257779121, + 0.0074293604120612144, + 0.006237206514924765, + 0.006141168065369129, + 0.0052857305854558945, + 0.0028097990434616804, + 0.009059271775186062, + 0.011696730740368366, + 0.003294011577963829, + 0.05957181006669998, + 0.06605682522058487, + 0.010936231352388859, + 0.008466020226478577, + 0.0031423328910022974, + 0.0025026844814419746, + 0.0030851177871227264, + 0.0021193029824644327, + 0.0030524765606969595, + 0.002582716289907694, + 0.003011733992025256, + 0.00404511671513319, + 0.004740623757243156, + 0.004563445691019297, + 0.0068880789913237095, + 0.0024163012858480215, + 0.003405944677069783, + 0.0045151980593800545, + 0.004542214795947075, + 0.004619240295141935, + 0.0017815546598285437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.007635810878127813 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034536570310593 + ], + "opposite_logprob_tokens": [ + -2.3503453731536865 + ], + "logprob_reward": -0.10034536570310593, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034536570310593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08033370226621628, + 0.03815210983157158, + 0.0479891411960125, + 0.055448003113269806, + 0.1617753803730011, + 0.3784467577934265, + 0.06584152579307556, + 0.12032514810562134, + 0.026208138093352318, + 0.025480162352323532 + ], + "credit_map": [ + 0.014578944072127342, + 0.009333833120763302, + 0.009827911853790283, + 0.011907919310033321, + 0.007094122003763914, + 0.013170314021408558, + 0.00714816665276885, + 0.007272484712302685, + 0.007162551861256361, + 0.0032192938961088657, + 0.004161060322076082, + 0.004051631316542625, + 0.00626598484814167, + 0.006924450863152742, + 0.0029650472570210695, + 0.003402082482352853, + 0.011467247270047665, + 0.0033473623916506767, + 0.00680615333840251, + 0.004938654135912657, + 0.004121126141399145, + 0.0069120717234909534, + 0.003404696239158511, + 0.006991820875555277, + 0.00582326902076602, + 0.007113207597285509, + 0.007331951521337032, + 0.005804806016385555, + 0.00852582324296236, + 0.010702206753194332, + 0.006222126539796591, + 0.003924603573977947, + 0.02028634399175644, + 0.015518509782850742, + 0.06036931276321411, + 0.010483140125870705, + 0.006622492801398039, + 0.0065828850492835045, + 0.006882767658680677, + 0.03502989560365677, + 0.11605573445558548, + 0.025875357910990715, + 0.0225522518157959, + 0.12566013634204865, + 0.03304370492696762, + 0.019959351047873497, + 0.022135699167847633, + 0.013164437375962734, + 0.0052156406454741955, + 0.005652723368257284, + 0.01290412899106741, + 0.016204340383410454, + 0.008217808790504932, + 0.00672802422195673, + 0.0060534244403243065, + 0.004865429364144802, + 0.008232440799474716, + 0.01255282573401928, + 0.0036635722499340773, + 0.03422010689973831, + 0.03634476661682129, + 0.007231161464005709, + 0.013984785415232182, + 0.004095475189387798, + 0.0028179679065942764, + 0.0032646649051457644, + 0.0021600243635475636, + 0.003593415953218937, + 0.0030087034683674574, + 0.0034329714253544807, + 0.0029698736034333706, + 0.004960513673722744, + 0.003923915326595306, + 0.00412038154900074, + 0.0023804863449186087, + 0.003244384890422225, + 0.0031484656501561403, + 0.002874766243621707, + 0.002687092637643218, + 0.0031006664503365755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.10034536570310593 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot presented from a first-person perspective with a slightly shaky camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a black-and-white forest featuring towering birch trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white fox visible in a crouched, motionless posture with pointed ears tilted forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image contain only pure black, white, and gray tones with no other colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the fox's gaze directed straight at the camera with amber eyes gleaming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289471864700317 + ], + "opposite_logprob_tokens": [ + -3.773289442062378 + ], + "logprob_reward": -0.023289471864700317, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289471864700317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17487791180610657, + 0.05221071094274521, + 0.025928735733032227, + 0.043049927800893784, + 0.1332419365644455, + 0.24593505263328552, + 0.06749504059553146, + 0.10428077727556229, + 0.06263457238674164, + 0.09034530818462372 + ], + "credit_map": [ + 0.04385368525981903, + 0.009968448430299759, + 0.015078870579600334, + 0.0136046651750803, + 0.012146110646426678, + 0.015462791547179222, + 0.04514392092823982, + 0.019619418308138847, + 0.005149678327143192, + 0.004017900675535202, + 0.005668352358043194, + 0.0061894492246210575, + 0.009123510681092739, + 0.009985455311834812, + 0.004553608596324921, + 0.007522757165133953, + 0.00614737207069993, + 0.0026771232951432467, + 0.0026520811952650547, + 0.002775760367512703, + 0.0038086362183094025, + 0.004083802457898855, + 0.0014322216156870127, + 0.002351738279685378, + 0.004551888909190893, + 0.0063253785483539104, + 0.005218964070081711, + 0.005153303500264883, + 0.007122560869902372, + 0.0059876823797822, + 0.005787629168480635, + 0.0029025222174823284, + 0.016788918524980545, + 0.012454940006136894, + 0.05080534145236015, + 0.008310459554195404, + 0.0033624081406742334, + 0.005476287100464106, + 0.005723827984184027, + 0.030319755896925926, + 0.07427047938108444, + 0.016833655536174774, + 0.014739945530891418, + 0.08291801810264587, + 0.021240483969449997, + 0.013047481887042522, + 0.014338838867843151, + 0.008546173572540283, + 0.0065011875703930855, + 0.004856887273490429, + 0.010227146558463573, + 0.011930769309401512, + 0.007449828088283539, + 0.005432209465652704, + 0.013023307546973228, + 0.008073708973824978, + 0.0058978283777832985, + 0.006943596061319113, + 0.0027215920854359865, + 0.0351816825568676, + 0.03769513592123985, + 0.006967156659811735, + 0.004776074551045895, + 0.004097712691873312, + 0.010410293936729431, + 0.007620693184435368, + 0.007744739297777414, + 0.008293483406305313, + 0.00901771429926157, + 0.005830900277942419, + 0.006551073864102364, + 0.00716567924246192, + 0.012763077393174171, + 0.013459322974085808, + 0.006951488554477692, + 0.010949242860078812, + 0.014835290610790253, + 0.009608792141079903, + 0.01437842845916748, + 0.0073996614664793015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.023289471864700317 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_13/index_tensor([6316, 6316])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_13/index_tensor([6316, 6316])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..612f87a09a0796f86c85090d3274d14d8e029969 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_13/index_tensor([6316, 6316])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "majestic, eye-level wide shot begins, the camera moving forward with slow, deliberate grace through the breathtaking interior of a massive, rotating O'Neill cylindrical space habitat. Beneath the surface lies a vast, curving landscape of spotless green parkland, sprawling upward on either side to create an impossible, gravity-defying horizon, where sleek, white residential buildings constructed of a smooth, pearlescent material stretch along the arc of the world. A tranquil blue river flows along the curve, its surface perfectly flush with the ground it passes through. People are scattered throughout this idyllic scene, going about their peaceful, normal lives: a man in a tight, dark blue tracksuit jogs at a steady pace along a winding path; further ahead, a couple sits on a minimalist bench, absorbed in conversation, while nearby, two children laugh as they chase a small orb of soft, pulsating light. Looking directly upward, the \"sky\" is a vast transparent window, revealing the zero-gravity core of the habitat; a complex, elongated central hub is clearly visible, with an extended docking arm and blinking navigation lights, and tiny, silent spacecraft occasionally maneuver against the backdrop of a distant, star-filled void. The entire scene is bathed in bright, full-spectrum light, mimicking a perfect afternoon on Earth, casting soft, aimless shadows and creating an atmosphere of warmth, optimism, and profound peace.", + "seed": 32725, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_13/index_6316_seed_32725_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.00510862092487514, + "vlm_logprob": -0.00510862092487514, + "vlm_yes_no_margin": 5.875, + "vlm_token_credit": -0.00510862092487514, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010144460014998913 + ], + "opposite_logprob_tokens": [ + -6.510144233703613 + ], + "logprob_reward": -0.010144460014998913, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.010144460014998913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21283185482025146, + 0.07947608083486557, + 0.04633720591664314, + 0.07438313961029053, + 0.06742383539676666, + 0.07283965498209, + 0.06677136570215225, + 0.09212123602628708, + 0.18450409173965454, + 0.10331153124570847 + ], + "credit_map": [ + 0.016163956373929977, + 0.02718239091336727, + 0.022746851667761803, + 0.022977976128458977, + 0.027364298701286316, + 0.033940840512514114, + 0.04355230927467346, + 0.018903227522969246, + 0.008065666072070599, + 0.010288338176906109, + 0.011184679344296455, + 0.0088472506031394, + 0.010788192972540855, + 0.011070862412452698, + 0.010593538172543049, + 0.00863755214959383, + 0.004054854158312082, + 0.007352744694799185, + 0.005613174755126238, + 0.006554602179676294, + 0.007917497307062149, + 0.006492371205240488, + 0.003783364314585924, + 0.004568594042211771, + 0.005709818564355373, + 0.00985700823366642, + 0.008484773337841034, + 0.006906685419380665, + 0.01466487254947424, + 0.011659645475447178, + 0.009015667252242565, + 0.008084672503173351, + 0.005160126835107803, + 0.0070374407805502415, + 0.007246075663715601, + 0.006619067396968603, + 0.012588651850819588, + 0.011329871602356434, + 0.007415096275508404, + 0.010027503594756126, + 0.007312413305044174, + 0.01117625366896391, + 0.007890500128269196, + 0.007788326591253281, + 0.01117493212223053, + 0.015476876869797707, + 0.004527268931269646, + 0.007493088021874428, + 0.012029696255922318, + 0.009271902963519096, + 0.012696683406829834, + 0.006063614971935749, + 0.007452385500073433, + 0.008856326341629028, + 0.004547652788460255, + 0.005853105336427689, + 0.010549313388764858, + 0.008554448373615742, + 0.024096384644508362, + 0.01768900640308857, + 0.006357927341014147, + 0.00765285175293684, + 0.008866752497851849, + 0.008354554884135723, + 0.01139379758387804, + 0.010244052857160568, + 0.03353169932961464, + 0.055452995002269745, + 0.01255817897617817, + 0.013169960118830204, + 0.019100835546851158, + 0.02905256301164627, + 0.007105547469109297, + 0.005747449584305286, + 0.019144754856824875, + 0.01846659742295742, + 0.007707872893661261, + 0.009634971618652344, + 0.015345915220677853, + 0.020158421248197556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.010144460014998913 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019587157294154167 + ], + "opposite_logprob_tokens": [ + -6.251958847045898 + ], + "logprob_reward": -0.0019587157294154167, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019587157294154167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11360016465187073, + 0.08846655488014221, + 0.06822667270898819, + 0.09210848808288574, + 0.10308310389518738, + 0.0539376474916935, + 0.039235904812812805, + 0.1645578145980835, + 0.1423129439353943, + 0.13447065651416779 + ], + "credit_map": [ + 0.009426741860806942, + 0.013184529729187489, + 0.014755431562662125, + 0.013469009660184383, + 0.01760336197912693, + 0.016205551102757454, + 0.018110206350684166, + 0.01084533054381609, + 0.011839636601507664, + 0.011371981352567673, + 0.01141450461000204, + 0.009284253232181072, + 0.01633141003549099, + 0.012309925630688667, + 0.009565441869199276, + 0.006349402479827404, + 0.007626892067492008, + 0.011014541611075401, + 0.008702961727976799, + 0.007542394567281008, + 0.010142926126718521, + 0.009963048622012138, + 0.007501914165914059, + 0.005731992889195681, + 0.006570339202880859, + 0.011161125265061855, + 0.011850442737340927, + 0.011548475362360477, + 0.015001144260168076, + 0.014352640137076378, + 0.01138509251177311, + 0.010239223949611187, + 0.005837500095367432, + 0.00896748248487711, + 0.01246668491512537, + 0.013719282113015652, + 0.017564527690410614, + 0.016460521146655083, + 0.011896011419594288, + 0.01617109775543213, + 0.004814039450138807, + 0.00624769926071167, + 0.005749668460339308, + 0.005190656986087561, + 0.011153031140565872, + 0.009143941104412079, + 0.004561620764434338, + 0.007076992187649012, + 0.009393064305186272, + 0.0046560862101614475, + 0.008112408220767975, + 0.004752239678055048, + 0.0031862417235970497, + 0.0030244640074670315, + 0.0032350244000554085, + 0.002876375801861286, + 0.010336487554013729, + 0.006173229776322842, + 0.06467782706022263, + 0.038222137838602066, + 0.005933334585279226, + 0.007030946668237448, + 0.008322808891534805, + 0.02386104315519333, + 0.011460399255156517, + 0.007258287165313959, + 0.023865319788455963, + 0.03584793955087662, + 0.01021621935069561, + 0.012664026580750942, + 0.013710101135075092, + 0.027290653437376022, + 0.009206993505358696, + 0.006813158746808767, + 0.022625289857387543, + 0.028869634494185448, + 0.008381258696317673, + 0.012206093408167362, + 0.021597621962428093, + 0.024770615622401237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.0019587157294154167 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615511115640402 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.003615511115640402, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615511115640402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13453280925750732, + 0.07990936189889908, + 0.05458185449242592, + 0.049173641949892044, + 0.04754377901554108, + 0.04080181196331978, + 0.07734695076942444, + 0.24432212114334106, + 0.15507292747497559, + 0.11671476811170578 + ], + "credit_map": [ + 0.014786445535719395, + 0.014047905802726746, + 0.01628255285322666, + 0.015527394600212574, + 0.01986563950777054, + 0.02109343186020851, + 0.01956319808959961, + 0.013366235420107841, + 0.009175905957818031, + 0.010982697829604149, + 0.009890311397612095, + 0.011516343802213669, + 0.014042655937373638, + 0.01112239621579647, + 0.007872854359447956, + 0.0053061931394040585, + 0.005611990112811327, + 0.00876487884670496, + 0.006518191657960415, + 0.006683880928903818, + 0.008921623229980469, + 0.008205262012779713, + 0.004758859518915415, + 0.005117163527756929, + 0.0033408585004508495, + 0.0072707729414105415, + 0.005101757124066353, + 0.005097182467579842, + 0.00676362682133913, + 0.007704730611294508, + 0.006284891162067652, + 0.007609819062054157, + 0.003967109601944685, + 0.005147300194948912, + 0.006669448688626289, + 0.0055736456997692585, + 0.008498512208461761, + 0.007345946040004492, + 0.004472568165510893, + 0.005869247484952211, + 0.004375791177153587, + 0.004351272713392973, + 0.005224567372351885, + 0.005409108940511942, + 0.005152454599738121, + 0.007697666063904762, + 0.0027980091981589794, + 0.005792939104139805, + 0.01744271069765091, + 0.009144659154117107, + 0.01854492537677288, + 0.009117413312196732, + 0.0067712049931287766, + 0.0061203897930681705, + 0.004782866220921278, + 0.005422770511358976, + 0.017758602276444435, + 0.00836846325546503, + 0.09559068828821182, + 0.06071460247039795, + 0.006647341884672642, + 0.012767673470079899, + 0.015364672057330608, + 0.027110084891319275, + 0.013513037003576756, + 0.008104920387268066, + 0.026208920404314995, + 0.042395804077386856, + 0.011995861306786537, + 0.013685734011232853, + 0.014204295352101326, + 0.0249643437564373, + 0.008160876110196114, + 0.003893683198839426, + 0.023105522617697716, + 0.024750886484980583, + 0.004585872869938612, + 0.007842481136322021, + 0.018578190356492996, + 0.025797249749302864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.003615511115640402 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762551836669445 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.00762551836669445, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762551836669445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15143628418445587, + 0.12421850115060806, + 0.041140366345644, + 0.04775344207882881, + 0.05964722856879234, + 0.07203657180070877, + 0.04282812029123306, + 0.15928325057029724, + 0.2042650431394577, + 0.09739124774932861 + ], + "credit_map": [ + 0.014019758440554142, + 0.014475071802735329, + 0.017878344282507896, + 0.018568851053714752, + 0.017789894714951515, + 0.025815557688474655, + 0.029501359909772873, + 0.01338744256645441, + 0.008613409474492073, + 0.007002773229032755, + 0.01374544482678175, + 0.01749749481678009, + 0.010801440104842186, + 0.020384296774864197, + 0.03215130791068077, + 0.014022329822182655, + 0.003448989475145936, + 0.006596829742193222, + 0.004513709340244532, + 0.004035020712763071, + 0.007958268746733665, + 0.00615069642663002, + 0.0036506857722997665, + 0.004786164034157991, + 0.0028760740533471107, + 0.004558618646115065, + 0.006069705355912447, + 0.005238546524196863, + 0.006205452140420675, + 0.006919404957443476, + 0.008173051290214062, + 0.007712589111179113, + 0.005234964657574892, + 0.005378300789743662, + 0.008513054810464382, + 0.007807191926985979, + 0.008234841749072075, + 0.0100205447524786, + 0.006018875166773796, + 0.008439455181360245, + 0.007931348867714405, + 0.009795542806386948, + 0.006380785256624222, + 0.006308621261268854, + 0.013453067280352116, + 0.01070395763963461, + 0.005973383784294128, + 0.011489863507449627, + 0.008799168281257153, + 0.005464871879667044, + 0.005486569833010435, + 0.0043890951201319695, + 0.006400762591511011, + 0.004699708428233862, + 0.0038822696078568697, + 0.0037056771107017994, + 0.014381550252437592, + 0.009146468713879585, + 0.05453180521726608, + 0.03528422862291336, + 0.008765492588281631, + 0.01556267961859703, + 0.008524795062839985, + 0.013086223974823952, + 0.010413215495646, + 0.010162275284528732, + 0.03497887775301933, + 0.050969842821359634, + 0.013412232510745525, + 0.01427167747169733, + 0.020610220730304718, + 0.049446698278188705, + 0.0054724449291825294, + 0.00472552003338933, + 0.018337314948439598, + 0.024122027680277824, + 0.004890620242804289, + 0.006580940447747707, + 0.015720754861831665, + 0.01754162833094597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.00762551836669445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198899397626519 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198899397626519, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198899397626519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3328392505645752, + 0.10031910985708237, + 0.08121068775653839, + 0.11802110821008682, + 0.04400945082306862, + 0.05435881391167641, + 0.040125854313373566, + 0.08411947637796402, + 0.07335395365953445, + 0.07164230942726135 + ], + "credit_map": [ + 0.026310978457331657, + 0.02620312198996544, + 0.04018702358007431, + 0.03699054196476936, + 0.05419722944498062, + 0.055184777826070786, + 0.0595417246222496, + 0.03422386199235916, + 0.008168439380824566, + 0.013759702444076538, + 0.011482956819236279, + 0.012337606400251389, + 0.012562815099954605, + 0.01562812738120556, + 0.01708787865936756, + 0.00929158553481102, + 0.007582625839859247, + 0.014903798699378967, + 0.009320230223238468, + 0.009839225560426712, + 0.012972279451787472, + 0.012219281867146492, + 0.008100784383714199, + 0.006272456143051386, + 0.008709228597581387, + 0.009129693731665611, + 0.015451427549123764, + 0.012609564699232578, + 0.014991963282227516, + 0.01819494180381298, + 0.021379750221967697, + 0.01755453646183014, + 0.0034395207185298204, + 0.004773781169205904, + 0.006606632843613625, + 0.004715878050774336, + 0.007662651129066944, + 0.0060860333032906055, + 0.0043683722615242004, + 0.006356576923280954, + 0.007019587326794863, + 0.00662093423306942, + 0.006180231925100088, + 0.004072768148034811, + 0.008851682767271996, + 0.011989077553153038, + 0.0036555908154696226, + 0.005968938581645489, + 0.005395545158535242, + 0.006258840207010508, + 0.005501252133399248, + 0.0035981247201561928, + 0.005297013558447361, + 0.005368605721741915, + 0.004131346475332975, + 0.004575129132717848, + 0.00871602538973093, + 0.006765756290405989, + 0.022083576768636703, + 0.014078587293624878, + 0.005857436917722225, + 0.008521877229213715, + 0.006143905222415924, + 0.011952309869229794, + 0.006978181190788746, + 0.004818515852093697, + 0.00999557226896286, + 0.00953762885183096, + 0.006680100224912167, + 0.008675197139382362, + 0.00740272318944335, + 0.0192660354077816, + 0.005212982650846243, + 0.0033472059294581413, + 0.014699664898216724, + 0.017859768122434616, + 0.003562730038538575, + 0.004464174620807171, + 0.01014227606356144, + 0.012353511527180672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.002198899397626519 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic, eye-level wide shot begins, the camera moving forward with slow, deliberate grace through the breathtaking interior of a massive, rotating O'Neill cylindrical space habitat. Beneath the surface lies a vast, curving landscape of spotless green parkland, sprawling upward on either side to create an impossible, gravity-defying horizon, where sleek, white residential buildings constructed of a smooth, pearlescent material stretch along the arc of the world. A tranquil blue river flows along the curve, its surface perfectly flush with the ground it passes through. People are scattered throughout this idyllic scene, going about their peaceful, normal lives: a man in a tight, dark blue tracksuit jogs at a steady pace along a winding path; further ahead, a couple sits on a minimalist bench, absorbed in conversation, while nearby, two children laugh as they chase a small orb of soft, pulsating light. Looking directly upward, the \"sky\" is a vast transparent window, revealing the zero-gravity core of the habitat; a complex, elongated central hub is clearly visible, with an extended docking arm and blinking navigation lights, and tiny, silent spacecraft occasionally maneuver against the backdrop of a distant, star-filled void. The entire scene is bathed in bright, full-spectrum light, mimicking a perfect afternoon on Earth, casting soft, aimless shadows and creating an atmosphere of warmth, optimism, and profound peace.", + "seed": 32726, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_13/index_6316_seed_32726_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.007875265367329121, + "vlm_logprob": -0.007875265367329121, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.007875265367329121, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011719073168933392 + ], + "opposite_logprob_tokens": [ + -5.511719226837158 + ], + "logprob_reward": -0.011719073168933392, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.011719073168933392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35077497363090515, + 0.05286703631281853, + 0.11991656571626663, + 0.06595386564731598, + 0.025835271924734116, + 0.022040070965886116, + 0.04083223640918732, + 0.062305912375450134, + 0.1663510948419571, + 0.09312310814857483 + ], + "credit_map": [ + 0.033958613872528076, + 0.07613525539636612, + 0.04050666093826294, + 0.0272752083837986, + 0.09376201778650284, + 0.02556525729596615, + 0.028208747506141663, + 0.025363150984048843, + 0.004049958661198616, + 0.014885627664625645, + 0.005253185052424669, + 0.0043060434982180595, + 0.011302828788757324, + 0.006551777012646198, + 0.0037748226895928383, + 0.0027427810709923506, + 0.0046803210861980915, + 0.0504646860063076, + 0.007677219808101654, + 0.0038021046202629805, + 0.04289301484823227, + 0.004388092085719109, + 0.003886360675096512, + 0.002124750055372715, + 0.005062500946223736, + 0.015751635655760765, + 0.007131898310035467, + 0.0077813249081373215, + 0.010543480515480042, + 0.005684700794517994, + 0.009812084026634693, + 0.0041862232610583305, + 0.0018963406328111887, + 0.004551765508949757, + 0.004471926484256983, + 0.005578311160206795, + 0.0028439005836844444, + 0.0021634381264448166, + 0.0027789382729679346, + 0.0015506455674767494, + 0.0014202874153852463, + 0.00512246647849679, + 0.0026312540285289288, + 0.0019529397832229733, + 0.005848641507327557, + 0.0018842823337763548, + 0.001801328849978745, + 0.001378868124447763, + 0.002451573498547077, + 0.008874177001416683, + 0.004208036232739687, + 0.003698793239891529, + 0.010530933737754822, + 0.003942879382520914, + 0.0038503925316035748, + 0.0032754421699792147, + 0.004357425961643457, + 0.013857855461537838, + 0.0065879784524440765, + 0.006510899402201176, + 0.01283477246761322, + 0.006284039933234453, + 0.0077079045586287975, + 0.004165025427937508, + 0.005486397538334131, + 0.06419987976551056, + 0.015185397118330002, + 0.010083334520459175, + 0.0492258220911026, + 0.007706115022301674, + 0.00838964618742466, + 0.006074467208236456, + 0.00443531759083271, + 0.02623782306909561, + 0.006533255334943533, + 0.0052122632041573524, + 0.01823786087334156, + 0.010940882377326488, + 0.01183599978685379, + 0.009689684957265854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.011719073168933392 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003219304606318474 + ], + "opposite_logprob_tokens": [ + -5.753219127655029 + ], + "logprob_reward": -0.003219304606318474, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003219304606318474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34502995014190674, + 0.045842304825782776, + 0.03739007189869881, + 0.09991275519132614, + 0.07789599150419235, + 0.06621287763118744, + 0.08341041952371597, + 0.045573219656944275, + 0.10550644993782043, + 0.09322596341371536 + ], + "credit_map": [ + 0.038618117570877075, + 0.06669358909130096, + 0.041442401707172394, + 0.02683091163635254, + 0.07469531148672104, + 0.032113365828990936, + 0.03400323912501335, + 0.030632980167865753, + 0.004288624506443739, + 0.013964413665235043, + 0.004073393531143665, + 0.0026611119974404573, + 0.011817491613328457, + 0.003550552763044834, + 0.0029900260269641876, + 0.0024966823402792215, + 0.003119959030300379, + 0.009641758166253567, + 0.0038287541829049587, + 0.003523046849295497, + 0.007449145894497633, + 0.00371129484847188, + 0.0034298175014555454, + 0.0026862940285354853, + 0.008749101310968399, + 0.03223735839128494, + 0.011465104296803474, + 0.00867448840290308, + 0.016737215220928192, + 0.006559557281434536, + 0.010002716444432735, + 0.005487207323312759, + 0.0057219211012125015, + 0.010370020754635334, + 0.016336865723133087, + 0.012734158895909786, + 0.01018060464411974, + 0.0064557259902358055, + 0.009049637243151665, + 0.007047052960842848, + 0.006140491459518671, + 0.018110308796167374, + 0.012092338874936104, + 0.00640140101313591, + 0.008134538307785988, + 0.003942008130252361, + 0.007074699271470308, + 0.004317082464694977, + 0.0050865644589066505, + 0.030399572104215622, + 0.0075591071508824825, + 0.005604547448456287, + 0.01495615765452385, + 0.003912490792572498, + 0.011941220611333847, + 0.003950751386582851, + 0.002857385203242302, + 0.010971073992550373, + 0.005543950013816357, + 0.004146893974393606, + 0.009381203912198544, + 0.004297582898288965, + 0.004848174285143614, + 0.003526951652020216, + 0.00423005223274231, + 0.034440405666828156, + 0.011935390532016754, + 0.007215541321784258, + 0.02664526179432869, + 0.0055365487933158875, + 0.008568539284169674, + 0.006934705656021833, + 0.0046574026346206665, + 0.02074957825243473, + 0.009053937159478664, + 0.00656853336840868, + 0.012699848972260952, + 0.005122561939060688, + 0.013762646354734898, + 0.02061145007610321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.003219304606318474 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949288606643677 + ], + "opposite_logprob_tokens": [ + -5.1309494972229 + ], + "logprob_reward": -0.005949288606643677, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949288606643677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4219670295715332, + 0.02557438053190708, + 0.08974316716194153, + 0.04578201845288277, + 0.023143582046031952, + 0.014321007765829563, + 0.04450589790940285, + 0.06749352812767029, + 0.1153164952993393, + 0.1521529108285904 + ], + "credit_map": [ + 0.04685385152697563, + 0.08788230270147324, + 0.050077974796295166, + 0.02788722887635231, + 0.10116887092590332, + 0.03678112477064133, + 0.03689916804432869, + 0.03441649675369263, + 0.0024614944122731686, + 0.0050287628546357155, + 0.0028533886652439833, + 0.003480657236650586, + 0.00343781104311347, + 0.0034878014121204615, + 0.0028905945364385843, + 0.0019338716519996524, + 0.0035419599153101444, + 0.037013132125139236, + 0.006262044422328472, + 0.003203129395842552, + 0.028939710929989815, + 0.004088603891432285, + 0.003933923318982124, + 0.0027606673538684845, + 0.003327278420329094, + 0.007634890265762806, + 0.004975559655576944, + 0.00619347020983696, + 0.00912497565150261, + 0.006797695532441139, + 0.004410972353070974, + 0.003317174967378378, + 0.0017826550174504519, + 0.005760661792010069, + 0.0033610223326832056, + 0.003104129573330283, + 0.00311649520881474, + 0.001695612445473671, + 0.0029506420250982046, + 0.0013723636511713266, + 0.0011446618009358644, + 0.0021407781168818474, + 0.002868464682251215, + 0.0019772073719650507, + 0.0017369642155244946, + 0.0012555808061733842, + 0.0019015776924788952, + 0.0012957730796188116, + 0.0027498730923980474, + 0.016423767432570457, + 0.0040544080547988415, + 0.0033906775061041117, + 0.006522692274302244, + 0.002349311951547861, + 0.00644234474748373, + 0.0025728228501975536, + 0.0038928696885704994, + 0.015857884660363197, + 0.008008385077118874, + 0.007509985007345676, + 0.013322215527296066, + 0.006067374721169472, + 0.007286344189196825, + 0.005548464134335518, + 0.004470418207347393, + 0.03917526826262474, + 0.013761141337454319, + 0.008543749339878559, + 0.029318390414118767, + 0.005979103036224842, + 0.00811728835105896, + 0.0059511372819542885, + 0.006731314118951559, + 0.06180606782436371, + 0.00968512799590826, + 0.007698528468608856, + 0.029390376061201096, + 0.008838082663714886, + 0.013694548979401588, + 0.014308860525488853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.005949288606643677 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537001632153988 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.012537001632153988, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537001632153988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29956844449043274, + 0.0815340206027031, + 0.050034020096063614, + 0.031108440831303596, + 0.08307507634162903, + 0.030765362083911896, + 0.07039962708950043, + 0.07070133090019226, + 0.07800452411174774, + 0.20480920374393463 + ], + "credit_map": [ + 0.03314574807882309, + 0.06409233063459396, + 0.03587768226861954, + 0.025550004094839096, + 0.06213289126753807, + 0.025345830246806145, + 0.02405554987490177, + 0.029368404299020767, + 0.006517962086945772, + 0.028724761679768562, + 0.006273338105529547, + 0.003917655907571316, + 0.02326638624072075, + 0.004989448003470898, + 0.004853196442127228, + 0.002991266781464219, + 0.004128335043787956, + 0.015906130895018578, + 0.004834288265556097, + 0.0035934746265411377, + 0.013206934556365013, + 0.0035533099435269833, + 0.002846454968675971, + 0.0019650848116725683, + 0.003017653012648225, + 0.00526574207469821, + 0.005127603188157082, + 0.004329971503466368, + 0.004296845756471157, + 0.0030699081253260374, + 0.004439867567270994, + 0.0015608449466526508, + 0.0064863767474889755, + 0.0221308171749115, + 0.013503632508218288, + 0.013867742381989956, + 0.009345658123493195, + 0.0049376715905964375, + 0.006905984133481979, + 0.005897185765206814, + 0.0028976125176995993, + 0.007271394599229097, + 0.0037649848964065313, + 0.003230787580832839, + 0.006467674858868122, + 0.0022931562270969152, + 0.002963591832667589, + 0.0018761551473289728, + 0.005200795363634825, + 0.011633837595582008, + 0.012689256109297276, + 0.012679781764745712, + 0.007848729379475117, + 0.006599859334528446, + 0.008268048986792564, + 0.005479313898831606, + 0.004209927748888731, + 0.0163862444460392, + 0.007300985045731068, + 0.010607687756419182, + 0.010722082108259201, + 0.007220308296382427, + 0.008942609652876854, + 0.005311479791998863, + 0.0035355491563677788, + 0.026024114340543747, + 0.00689366739243269, + 0.006980520207434893, + 0.020366178825497627, + 0.004412938840687275, + 0.006043590139597654, + 0.003747952403500676, + 0.00832425244152546, + 0.08530069887638092, + 0.01461970154196024, + 0.011262895539402962, + 0.04609489068388939, + 0.012735183350741863, + 0.009516694582998753, + 0.016954872757196426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.012537001632153988 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951658822596073 + ], + "opposite_logprob_tokens": [ + -5.130951881408691 + ], + "logprob_reward": -0.005951658822596073, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951658822596073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38457369804382324, + 0.09261180460453033, + 0.209958016872406, + 0.028832118958234787, + 0.012006036937236786, + 0.012313458137214184, + 0.050408221781253815, + 0.04668033495545387, + 0.04853836074471474, + 0.11407794058322906 + ], + "credit_map": [ + 0.04405205696821213, + 0.07706045359373093, + 0.05487366020679474, + 0.03741515427827835, + 0.05651046335697174, + 0.02899855561554432, + 0.03365715220570564, + 0.05200619250535965, + 0.005675656255334616, + 0.03105976991355419, + 0.007312563247978687, + 0.006100937258452177, + 0.02651808224618435, + 0.007123281713575125, + 0.005421262234449387, + 0.0034002491738647223, + 0.010149327106773853, + 0.09513349831104279, + 0.01535767037421465, + 0.004548385739326477, + 0.06962820887565613, + 0.005176318343728781, + 0.006056526675820351, + 0.003908082377165556, + 0.002446981845423579, + 0.0050688632763922215, + 0.003400198882445693, + 0.004172918386757374, + 0.0048236302100121975, + 0.0037594125606119633, + 0.0032019864302128553, + 0.0019581287633627653, + 0.0008723686914891005, + 0.002448313869535923, + 0.0021282806992530823, + 0.0014597241533920169, + 0.0018474557437002659, + 0.001284372410736978, + 0.0012003402225673199, + 0.0007651818450540304, + 0.0010634479112923145, + 0.002854118123650551, + 0.0017852179007604718, + 0.0012151182163506746, + 0.002254039980471134, + 0.0008969951304607093, + 0.0012528527295216918, + 0.0009916684357449412, + 0.003324054181575775, + 0.01618458330631256, + 0.004948461428284645, + 0.004507813137024641, + 0.00917387567460537, + 0.0030976026318967342, + 0.006674888078123331, + 0.0024969412479549646, + 0.002600535284727812, + 0.013754318468272686, + 0.00431770458817482, + 0.004863637965172529, + 0.009807193651795387, + 0.0030042037833482027, + 0.0056325942277908325, + 0.002700147219002247, + 0.0019557622727006674, + 0.017351454123854637, + 0.0043132673017680645, + 0.004000879358500242, + 0.011269436217844486, + 0.0025497276801615953, + 0.004838445223867893, + 0.0022593874018639326, + 0.006148090586066246, + 0.03891042619943619, + 0.007631033193320036, + 0.007556559983640909, + 0.01706647500395775, + 0.004909854382276535, + 0.024651186540722847, + 0.007204311899840832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.005951658822596073 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic, eye-level wide shot begins, the camera moving forward with slow, deliberate grace through the breathtaking interior of a massive, rotating O'Neill cylindrical space habitat. Beneath the surface lies a vast, curving landscape of spotless green parkland, sprawling upward on either side to create an impossible, gravity-defying horizon, where sleek, white residential buildings constructed of a smooth, pearlescent material stretch along the arc of the world. A tranquil blue river flows along the curve, its surface perfectly flush with the ground it passes through. People are scattered throughout this idyllic scene, going about their peaceful, normal lives: a man in a tight, dark blue tracksuit jogs at a steady pace along a winding path; further ahead, a couple sits on a minimalist bench, absorbed in conversation, while nearby, two children laugh as they chase a small orb of soft, pulsating light. Looking directly upward, the \"sky\" is a vast transparent window, revealing the zero-gravity core of the habitat; a complex, elongated central hub is clearly visible, with an extended docking arm and blinking navigation lights, and tiny, silent spacecraft occasionally maneuver against the backdrop of a distant, star-filled void. The entire scene is bathed in bright, full-spectrum light, mimicking a perfect afternoon on Earth, casting soft, aimless shadows and creating an atmosphere of warmth, optimism, and profound peace.", + "seed": 32727, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_13/index_6316_seed_32727_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.004527351632714271, + "vlm_logprob": -0.004527351632714271, + "vlm_yes_no_margin": 5.925, + "vlm_token_credit": -0.004527351632714271, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006345955654978752 + ], + "opposite_logprob_tokens": [ + -6.381345748901367 + ], + "logprob_reward": -0.006345955654978752, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.006345955654978752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23745784163475037, + 0.10994902998209, + 0.10060637444257736, + 0.09905340522527695, + 0.08842869102954865, + 0.12705595791339874, + 0.08447609096765518, + 0.05313217639923096, + 0.04006115347146988, + 0.059779245406389236 + ], + "credit_map": [ + 0.027317361906170845, + 0.03078153170645237, + 0.026422131806612015, + 0.031155521050095558, + 0.03195487707853317, + 0.040114548057317734, + 0.021079817786812782, + 0.028632070869207382, + 0.014999214559793472, + 0.019279008731245995, + 0.012325616553425789, + 0.014066440053284168, + 0.010489766485989094, + 0.01686762273311615, + 0.010097703896462917, + 0.011823677457869053, + 0.012274457141757011, + 0.008813658729195595, + 0.011050011962652206, + 0.012805919162929058, + 0.013533134944736958, + 0.015988299623131752, + 0.01319890096783638, + 0.012941996566951275, + 0.008328440599143505, + 0.01262042298913002, + 0.019224882125854492, + 0.015212949365377426, + 0.007790284231305122, + 0.01531093567609787, + 0.012398052029311657, + 0.008167448453605175, + 0.0074175582267344, + 0.007580040954053402, + 0.013177327811717987, + 0.013599283993244171, + 0.011339431628584862, + 0.011181849054992199, + 0.013123310171067715, + 0.01100989244878292, + 0.006962570361793041, + 0.007723470684140921, + 0.021476395428180695, + 0.021054279059171677, + 0.00917868036776781, + 0.015584909357130527, + 0.016758430749177933, + 0.02831723913550377, + 0.007712602615356445, + 0.006674652919173241, + 0.011582409963011742, + 0.010558806359767914, + 0.010063761845231056, + 0.018220460042357445, + 0.009697390720248222, + 0.009966022334992886, + 0.005370844155550003, + 0.005372625309973955, + 0.006080498453229666, + 0.00863488670438528, + 0.0048586903139948845, + 0.006160684861242771, + 0.008495454676449299, + 0.008158499374985695, + 0.00340006360784173, + 0.003472918178886175, + 0.007905262522399426, + 0.004759596195071936, + 0.004608220886439085, + 0.005241904407739639, + 0.004958515055477619, + 0.005714681930840015, + 0.0071687824092805386, + 0.005527250003069639, + 0.008860823698341846, + 0.0059162708930671215, + 0.006988178938627243, + 0.010870167054235935, + 0.00603447062894702, + 0.008413305506110191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.006345955654978752 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0012029323261231184 + ], + "opposite_logprob_tokens": [ + -6.7512030601501465 + ], + "logprob_reward": -0.0012029323261231184, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0012029323261231184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20155808329582214, + 0.03214945271611214, + 0.02737109176814556, + 0.07643502205610275, + 0.07393769919872284, + 0.23065054416656494, + 0.06571836769580841, + 0.13322558999061584, + 0.04403472691774368, + 0.11491949111223221 + ], + "credit_map": [ + 0.017416782677173615, + 0.02593293972313404, + 0.03194433078169823, + 0.030527614057064056, + 0.01879430003464222, + 0.03361590951681137, + 0.019737418740987778, + 0.023588746786117554, + 0.004638053011149168, + 0.0030795899219810963, + 0.003661131951957941, + 0.004146596882492304, + 0.0043729739263653755, + 0.00470852293074131, + 0.002821950474753976, + 0.004720630124211311, + 0.0022180324885994196, + 0.0020198470447212458, + 0.005675545893609524, + 0.0043759881518781185, + 0.002845081267878413, + 0.0035210824571549892, + 0.0041648210026323795, + 0.002550691133365035, + 0.005001125391572714, + 0.004596738610416651, + 0.012621399946510792, + 0.011128878220915794, + 0.011610972695052624, + 0.012525767087936401, + 0.010085898451507092, + 0.00886424072086811, + 0.006820325274020433, + 0.004394305404275656, + 0.01028942596167326, + 0.0116043109446764, + 0.005679525434970856, + 0.01005962211638689, + 0.009515004232525826, + 0.015575175173580647, + 0.015028866939246655, + 0.011264761909842491, + 0.043804533779621124, + 0.030387727543711662, + 0.014566854573786259, + 0.02779594622552395, + 0.029833849519491196, + 0.05796799063682556, + 0.006412023212760687, + 0.004351688548922539, + 0.012523421086370945, + 0.008448285050690174, + 0.007863378152251244, + 0.009740759618580341, + 0.007696462795138359, + 0.008682348765432835, + 0.014183921739459038, + 0.015631338581442833, + 0.020313436165452003, + 0.016452336683869362, + 0.01362518034875393, + 0.014804743230342865, + 0.021628741174936295, + 0.016585879027843475, + 0.0031206400599330664, + 0.0022709539625793695, + 0.004585165996104479, + 0.00509774312376976, + 0.003540675388649106, + 0.008650070056319237, + 0.008381118066608906, + 0.008388357236981392, + 0.013441083021461964, + 0.00828564539551735, + 0.019576746970415115, + 0.014555878937244415, + 0.008970714174211025, + 0.01136425044387579, + 0.017314253374934196, + 0.02141091227531433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.0012029323261231184 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002819255692884326 + ], + "opposite_logprob_tokens": [ + -5.877819061279297 + ], + "logprob_reward": -0.002819255692884326, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002819255692884326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3130880892276764, + 0.0634765774011612, + 0.08258028328418732, + 0.14586366713047028, + 0.07482597231864929, + 0.08125591278076172, + 0.0518084317445755, + 0.07877550274133682, + 0.057842887938022614, + 0.05048266425728798 + ], + "credit_map": [ + 0.023023363202810287, + 0.028469553217291832, + 0.06022858992218971, + 0.053134314715862274, + 0.029053496196866035, + 0.05276263877749443, + 0.03177209571003914, + 0.034644048660993576, + 0.006530339363962412, + 0.005234868731349707, + 0.0073718661442399025, + 0.00904079806059599, + 0.006836059503257275, + 0.009683958254754543, + 0.008589038625359535, + 0.010189649648964405, + 0.00694383354857564, + 0.005648060701787472, + 0.007077284622937441, + 0.01327107660472393, + 0.00684459088370204, + 0.014939670450985432, + 0.015745382755994797, + 0.012110385112464428, + 0.00782312173396349, + 0.010302918963134289, + 0.011644024401903152, + 0.016345620155334473, + 0.027616318315267563, + 0.03457939997315407, + 0.019255556166172028, + 0.01829669624567032, + 0.008559870533645153, + 0.004506334196776152, + 0.012837951071560383, + 0.013055864721536636, + 0.00748036103323102, + 0.009377882815897465, + 0.010741259902715683, + 0.00826644990593195, + 0.00647567817941308, + 0.00786040723323822, + 0.012271338142454624, + 0.012564806267619133, + 0.00851408950984478, + 0.010072173550724983, + 0.01304197870194912, + 0.010455446317791939, + 0.006274685729295015, + 0.004827975295484066, + 0.007615283131599426, + 0.007040797732770443, + 0.005255233962088823, + 0.007048396393656731, + 0.0058602686040103436, + 0.007885793223977089, + 0.008268126286566257, + 0.010084131732583046, + 0.008714083582162857, + 0.008918732404708862, + 0.012095117941498756, + 0.01165627408772707, + 0.010728292167186737, + 0.008310748264193535, + 0.004159383475780487, + 0.0032931207679212093, + 0.0052032205276191235, + 0.004600295796990395, + 0.002931883092969656, + 0.010033924132585526, + 0.009238217957317829, + 0.018382839858531952, + 0.0063640703447163105, + 0.0037884509656578302, + 0.007083809468895197, + 0.00507013313472271, + 0.005924133118242025, + 0.006101664155721664, + 0.007878408767282963, + 0.008271993137896061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.002819255692884326 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778574109077454 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.009778574109077454, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778574109077454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4523673355579376, + 0.03658507019281387, + 0.027151385322213173, + 0.07235421985387802, + 0.030448416247963905, + 0.054920945316553116, + 0.048458825796842575, + 0.1334053874015808, + 0.07180221378803253, + 0.07250618189573288 + ], + "credit_map": [ + 0.03971266373991966, + 0.04701961204409599, + 0.07819931209087372, + 0.059423770755529404, + 0.059893108904361725, + 0.06622195988893509, + 0.05057142302393913, + 0.051325470209121704, + 0.003401203081011772, + 0.0032085690181702375, + 0.0046015577390789986, + 0.005742284934967756, + 0.004536694847047329, + 0.0056480769999325275, + 0.004395629744976759, + 0.00505105359479785, + 0.003716588718816638, + 0.002370733069255948, + 0.004150418099015951, + 0.002882534172385931, + 0.004570381715893745, + 0.0037525505758821964, + 0.0024448297917842865, + 0.0032633491791784763, + 0.004787086509168148, + 0.007575141731649637, + 0.009970136918127537, + 0.010534829460084438, + 0.006470112595707178, + 0.010245121084153652, + 0.010798129253089428, + 0.011973664164543152, + 0.0029725448694080114, + 0.0018895524553954601, + 0.004040932282805443, + 0.004275085404515266, + 0.0023502486292272806, + 0.003643002361059189, + 0.004440190736204386, + 0.0068368613719940186, + 0.004397973883897066, + 0.0037818343844264746, + 0.0076375785283744335, + 0.009153381921350956, + 0.004690644331276417, + 0.006653664633631706, + 0.009532366879284382, + 0.009073500521481037, + 0.004574210848659277, + 0.004119411576539278, + 0.008267810568213463, + 0.009645641781389713, + 0.0033180273603647947, + 0.004908926319330931, + 0.0068695093505084515, + 0.006755293346941471, + 0.011768710799515247, + 0.017046255990862846, + 0.01812458597123623, + 0.016649287194013596, + 0.014958465471863747, + 0.013450712896883488, + 0.026708899065852165, + 0.014698478393256664, + 0.005276345182210207, + 0.004051209427416325, + 0.011244240216910839, + 0.006358196958899498, + 0.004927636124193668, + 0.01769828051328659, + 0.008155127987265587, + 0.014091184362769127, + 0.007540899328887463, + 0.0065107401460409164, + 0.010648663155734539, + 0.007944336161017418, + 0.005465284921228886, + 0.011507083661854267, + 0.013130620121955872, + 0.009758553467690945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.009778574109077454 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024900403805077076 + ], + "opposite_logprob_tokens": [ + -6.002490043640137 + ], + "logprob_reward": -0.0024900403805077076, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024900403805077076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19560688734054565, + 0.13248728215694427, + 0.08082470297813416, + 0.08168213069438934, + 0.0899970605969429, + 0.06409957259893417, + 0.0735069215297699, + 0.16671325266361237, + 0.06416646391153336, + 0.05091580003499985 + ], + "credit_map": [ + 0.016014687716960907, + 0.034319207072257996, + 0.021729668602347374, + 0.03266835957765579, + 0.02266673743724823, + 0.022981395944952965, + 0.02239704504609108, + 0.02282976545393467, + 0.01001825649291277, + 0.010633321478962898, + 0.03495478257536888, + 0.01280254777520895, + 0.015759842470288277, + 0.02109258994460106, + 0.013085884973406792, + 0.014140053652226925, + 0.00929472129791975, + 0.008494723588228226, + 0.010103229433298111, + 0.011112512089312077, + 0.013223070651292801, + 0.012053301557898521, + 0.009089500643312931, + 0.007453633937984705, + 0.01050492562353611, + 0.009638553485274315, + 0.011007784865796566, + 0.01117789652198553, + 0.00854804739356041, + 0.009819692000746727, + 0.010109903290867805, + 0.010875330306589603, + 0.005811257753521204, + 0.009520884603261948, + 0.014091407880187035, + 0.017046090215444565, + 0.009279481135308743, + 0.012223486788570881, + 0.011204320937395096, + 0.010820126160979271, + 0.006985187996178865, + 0.006855313666164875, + 0.008201675489544868, + 0.009349759668111801, + 0.004646083805710077, + 0.0066298129968345165, + 0.00804476160556078, + 0.013386974111199379, + 0.007227320224046707, + 0.00837436132133007, + 0.00918420311063528, + 0.011544203385710716, + 0.008837716653943062, + 0.008074833080172539, + 0.01025630347430706, + 0.01000797376036644, + 0.015922911465168, + 0.023072374984622, + 0.01643052138388157, + 0.019451983273029327, + 0.024831801652908325, + 0.021457524970173836, + 0.02401135489344597, + 0.02153477631509304, + 0.005027014762163162, + 0.004151507280766964, + 0.00619459617882967, + 0.006130787078291178, + 0.0040735783986747265, + 0.008083291351795197, + 0.011756530031561852, + 0.018749156966805458, + 0.007491705473512411, + 0.0039633275009691715, + 0.004856228828430176, + 0.00470968009904027, + 0.004404784645885229, + 0.010528270155191422, + 0.00740460678935051, + 0.007557193282991648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0024900403805077076 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic, eye-level wide shot begins, the camera moving forward with slow, deliberate grace through the breathtaking interior of a massive, rotating O'Neill cylindrical space habitat. Beneath the surface lies a vast, curving landscape of spotless green parkland, sprawling upward on either side to create an impossible, gravity-defying horizon, where sleek, white residential buildings constructed of a smooth, pearlescent material stretch along the arc of the world. A tranquil blue river flows along the curve, its surface perfectly flush with the ground it passes through. People are scattered throughout this idyllic scene, going about their peaceful, normal lives: a man in a tight, dark blue tracksuit jogs at a steady pace along a winding path; further ahead, a couple sits on a minimalist bench, absorbed in conversation, while nearby, two children laugh as they chase a small orb of soft, pulsating light. Looking directly upward, the \"sky\" is a vast transparent window, revealing the zero-gravity core of the habitat; a complex, elongated central hub is clearly visible, with an extended docking arm and blinking navigation lights, and tiny, silent spacecraft occasionally maneuver against the backdrop of a distant, star-filled void. The entire scene is bathed in bright, full-spectrum light, mimicking a perfect afternoon on Earth, casting soft, aimless shadows and creating an atmosphere of warmth, optimism, and profound peace.", + "seed": 32728, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_13/index_6316_seed_32728_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0061164157232269645, + "vlm_logprob": -0.0061164157232269645, + "vlm_yes_no_margin": 5.6, + "vlm_token_credit": -0.0061164157232269645, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007277408614754677 + ], + "opposite_logprob_tokens": [ + -6.632277488708496 + ], + "logprob_reward": -0.007277408614754677, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.007277408614754677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2844898998737335, + 0.1249588280916214, + 0.05804594233632088, + 0.06424511224031448, + 0.14131410419940948, + 0.050566643476486206, + 0.055031802505254745, + 0.0697077214717865, + 0.08500271290540695, + 0.06663727015256882 + ], + "credit_map": [ + 0.016716640442609787, + 0.03000328317284584, + 0.029345352202653885, + 0.07056161761283875, + 0.04032899811863899, + 0.025138026103377342, + 0.03173694387078285, + 0.04065905138850212, + 0.008245300501585007, + 0.01342162024229765, + 0.01914951764047146, + 0.014310138300061226, + 0.011203347705304623, + 0.011020113714039326, + 0.03170371055603027, + 0.01590508036315441, + 0.005084982607513666, + 0.004842876922339201, + 0.008029366843402386, + 0.006790780462324619, + 0.007671765983104706, + 0.007838979363441467, + 0.00950908474624157, + 0.008278106339275837, + 0.008626601658761501, + 0.008741799741983414, + 0.0068624126724898815, + 0.007119759451597929, + 0.008821489289402962, + 0.0115842055529356, + 0.004829174838960171, + 0.007659665308892727, + 0.008208505809307098, + 0.00756733538582921, + 0.0067472937516868114, + 0.009099963121116161, + 0.02040170505642891, + 0.04480019211769104, + 0.007187641225755215, + 0.03730146959424019, + 0.004064916167408228, + 0.007343963719904423, + 0.004717777483165264, + 0.0041376929730176926, + 0.016403045505285263, + 0.006114272866398096, + 0.0034721335396170616, + 0.004312843084335327, + 0.005663914605975151, + 0.007640787400305271, + 0.006059970706701279, + 0.004323964938521385, + 0.009329003281891346, + 0.012747076340019703, + 0.003204131731763482, + 0.006062954664230347, + 0.006061519030481577, + 0.0065263002179563046, + 0.012398658320307732, + 0.00791560485959053, + 0.007776453159749508, + 0.008430344983935356, + 0.0077960435301065445, + 0.012802799232304096, + 0.011689036153256893, + 0.008152655325829983, + 0.009552986361086369, + 0.012035210616886616, + 0.009483138099312782, + 0.010042911395430565, + 0.005413657054305077, + 0.01863311231136322, + 0.005326454062014818, + 0.005815659649670124, + 0.013479240238666534, + 0.012290718965232372, + 0.007587769068777561, + 0.01026966329663992, + 0.005244390107691288, + 0.006623375229537487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.007277408614754677 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001731802592985332 + ], + "opposite_logprob_tokens": [ + -6.376731872558594 + ], + "logprob_reward": -0.001731802592985332, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001731802592985332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1581445187330246, + 0.21445439755916595, + 0.03822603076696396, + 0.030361995100975037, + 0.28090450167655945, + 0.041564520448446274, + 0.09255298227071762, + 0.0401962511241436, + 0.02710810676217079, + 0.07648672163486481 + ], + "credit_map": [ + 0.013100417330861092, + 0.01740555837750435, + 0.016111211851239204, + 0.030734974890947342, + 0.020802592858672142, + 0.013824555091559887, + 0.02408209815621376, + 0.022083118557929993, + 0.010881111025810242, + 0.011428440921008587, + 0.04450586810708046, + 0.014099853113293648, + 0.01618323288857937, + 0.012147385627031326, + 0.0744185671210289, + 0.030789928510785103, + 0.003124722745269537, + 0.003300574142485857, + 0.005207052454352379, + 0.005577566102147102, + 0.004290200769901276, + 0.003968860022723675, + 0.006107802037149668, + 0.006649250630289316, + 0.0023631909862160683, + 0.0027398618403822184, + 0.006494412664324045, + 0.004646004643291235, + 0.003097021020948887, + 0.004000782035291195, + 0.0036840313114225864, + 0.003336691064760089, + 0.02426241524517536, + 0.02068316377699375, + 0.012900632806122303, + 0.019839243963360786, + 0.03452189266681671, + 0.05530015006661415, + 0.02333037368953228, + 0.09006660431623459, + 0.003862487617880106, + 0.007925831712782383, + 0.005772971082478762, + 0.004685572814196348, + 0.004204020835459232, + 0.004843563307076693, + 0.004072790499776602, + 0.006197282113134861, + 0.013489490374922752, + 0.013021024875342846, + 0.010645080357789993, + 0.011876476928591728, + 0.01811450533568859, + 0.008489775471389294, + 0.005687169265002012, + 0.011229459196329117, + 0.003744706977158785, + 0.003914332948625088, + 0.007075567729771137, + 0.004678158555179834, + 0.00579549465328455, + 0.004235440865159035, + 0.004957905039191246, + 0.005794645752757788, + 0.005670459009706974, + 0.0038440159987658262, + 0.0030157079454511404, + 0.002993090311065316, + 0.002658192766830325, + 0.0027387470472604036, + 0.0027111556846648455, + 0.0034767361357808113, + 0.00926143117249012, + 0.009585064835846424, + 0.007106589153409004, + 0.013409679755568504, + 0.012378168292343616, + 0.010169903747737408, + 0.007430939003825188, + 0.00714494101703167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.001731802592985332 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251067224889994 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005251067224889994, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251067224889994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19457775354385376, + 0.2111596167087555, + 0.018172573298215866, + 0.015877651050686836, + 0.3957609236240387, + 0.03822082281112671, + 0.047807008028030396, + 0.02703111432492733, + 0.01703590340912342, + 0.034356631338596344 + ], + "credit_map": [ + 0.01221719104796648, + 0.025244703516364098, + 0.021735787391662598, + 0.05050814151763916, + 0.028275376185774803, + 0.014314482919871807, + 0.02352512627840042, + 0.018756942823529243, + 0.015079256147146225, + 0.020300060510635376, + 0.05936707928776741, + 0.019915049895644188, + 0.018053963780403137, + 0.011361205019056797, + 0.04263020306825638, + 0.02445279248058796, + 0.0013706558384001255, + 0.001540223485790193, + 0.0024502314627170563, + 0.0025267447344958782, + 0.001752391573973, + 0.002703362377360463, + 0.0026880870573222637, + 0.0031408772338181734, + 0.00138967321254313, + 0.0013059533666819334, + 0.002745770150795579, + 0.002559119602665305, + 0.001517044729553163, + 0.001920000882819295, + 0.002703147940337658, + 0.001736941165290773, + 0.017281226813793182, + 0.025343384593725204, + 0.013568593189120293, + 0.020958613604307175, + 0.1541462540626526, + 0.12011787295341492, + 0.019266368821263313, + 0.025078602135181427, + 0.006586242001503706, + 0.010657685808837414, + 0.003356833243742585, + 0.00259659206494689, + 0.005208691582083702, + 0.005075050983577967, + 0.001977454638108611, + 0.002762273885309696, + 0.008489315398037434, + 0.005131851881742477, + 0.0055807605385780334, + 0.005662624258548021, + 0.006699062883853912, + 0.005700836423784494, + 0.003304114332422614, + 0.0072384425438940525, + 0.0026519380044192076, + 0.0028295419178903103, + 0.004514642991125584, + 0.002801202004775405, + 0.003336317138746381, + 0.003049624152481556, + 0.0028244394343346357, + 0.005023408215492964, + 0.002123005222529173, + 0.0017910454189404845, + 0.002256587380543351, + 0.002448637969791889, + 0.002015433507040143, + 0.0019047545501962304, + 0.0022188564762473106, + 0.0022775838151574135, + 0.004421479534357786, + 0.004422812722623348, + 0.0041779931634664536, + 0.004267600364983082, + 0.005288912449032068, + 0.004061815328896046, + 0.004005007911473513, + 0.0037110119592398405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.005251067224889994 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011070732958614826 + ], + "opposite_logprob_tokens": [ + -4.511070728302002 + ], + "logprob_reward": -0.011070732958614826, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011070732958614826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06938248127698898, + 0.1237817257642746, + 0.022879499942064285, + 0.02573912963271141, + 0.5648124814033508, + 0.06267077475786209, + 0.03396530821919441, + 0.03886587172746658, + 0.031490862369537354, + 0.02641182951629162 + ], + "credit_map": [ + 0.006911490112543106, + 0.007612148299813271, + 0.008140123449265957, + 0.009370188228785992, + 0.009847261942923069, + 0.007504039444029331, + 0.010432567447423935, + 0.009564659558236599, + 0.005320571828633547, + 0.003763065207749605, + 0.03875896707177162, + 0.0068908752873539925, + 0.007797709200531244, + 0.006631976924836636, + 0.04139275103807449, + 0.013225816190242767, + 0.0014133491786196828, + 0.0016563732642680407, + 0.004069684073328972, + 0.003539303783327341, + 0.0020043321419507265, + 0.002287985058501363, + 0.0034481182228773832, + 0.004460354335606098, + 0.002518444089218974, + 0.0032709576189517975, + 0.0024095482658594847, + 0.003895408473908901, + 0.0032771506812423468, + 0.003333395579829812, + 0.00394104328006506, + 0.00309318327344954, + 0.022038569673895836, + 0.034331388771533966, + 0.01826758123934269, + 0.02904585190117359, + 0.21517477929592133, + 0.1833835244178772, + 0.025871554389595985, + 0.03669922426342964, + 0.008438356220722198, + 0.020636240020394325, + 0.006577429361641407, + 0.0042416718788445, + 0.005971456877887249, + 0.008830506354570389, + 0.003056515473872423, + 0.0049185985699296, + 0.0038437610492110252, + 0.0031210279557853937, + 0.005109156481921673, + 0.005524784326553345, + 0.0043564788065850735, + 0.004687928594648838, + 0.0027538577560335398, + 0.0045683132484555244, + 0.004383502062410116, + 0.0033895631786435843, + 0.004970936570316553, + 0.004509002901613712, + 0.005424574017524719, + 0.003930365666747093, + 0.004798527806997299, + 0.007459398824721575, + 0.004194977693259716, + 0.004141594748944044, + 0.003962321672588587, + 0.0035591854248195887, + 0.004153117537498474, + 0.0036059701815247536, + 0.0031554424203932285, + 0.004718253389000893, + 0.0037279052194207907, + 0.0037801906000822783, + 0.0026988943573087454, + 0.0038075277116149664, + 0.004492426291108131, + 0.003579640993848443, + 0.0019875129219144583, + 0.0023377318866550922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.011070732958614826 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a massive, rotating O'Neill cylindrical space habitat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ground covered in spotless green parkland with sleek, white residential buildings made of a smooth, pearlescent material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tranquil blue river flowing along the curve of the ground, perfectly flush with the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving forward with slow, deliberate grace through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the 'sky' a vast transparent window revealing a zero-gravity core with a central hub, docking arm, and blinking navigation lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251067224889994 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005251067224889994, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251067224889994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15523530542850494, + 0.19280265271663666, + 0.03782019019126892, + 0.034020401537418365, + 0.2821785807609558, + 0.1178843304514885, + 0.08255862444639206, + 0.04078857973217964, + 0.016324132680892944, + 0.040387239307165146 + ], + "credit_map": [ + 0.010987579822540283, + 0.02105599083006382, + 0.01922488585114479, + 0.027118735015392303, + 0.026694495230913162, + 0.014346010982990265, + 0.016966456547379494, + 0.018841132521629333, + 0.009338416159152985, + 0.011434030719101429, + 0.05171426013112068, + 0.012655800208449364, + 0.01633443869650364, + 0.010507860220968723, + 0.06078914925456047, + 0.02002870664000511, + 0.0033220003824681044, + 0.0026605308521538973, + 0.005963901057839394, + 0.006038511637598276, + 0.004899698309600353, + 0.0036099597346037626, + 0.0058855521492660046, + 0.005440036300569773, + 0.0038502905517816544, + 0.0039047119207680225, + 0.005656347144395113, + 0.0044430941343307495, + 0.003916418645530939, + 0.003481024643406272, + 0.0056332736276090145, + 0.003135238541290164, + 0.013334083370864391, + 0.01813160628080368, + 0.011257236823439598, + 0.015874721109867096, + 0.10597255080938339, + 0.08487023413181305, + 0.014824904501438141, + 0.0179132129997015, + 0.01862865500152111, + 0.02325340360403061, + 0.0071477084420621395, + 0.006777417380362749, + 0.024401212111115456, + 0.018432607874274254, + 0.008852616883814335, + 0.010390713810920715, + 0.0093923918902874, + 0.016857508569955826, + 0.007646812591701746, + 0.007758484687656164, + 0.01100548729300499, + 0.017254125326871872, + 0.0046616229228675365, + 0.007982193492352962, + 0.004196935798972845, + 0.0036562695167958736, + 0.005355288740247488, + 0.004803569056093693, + 0.006239750888198614, + 0.004366930574178696, + 0.005281117744743824, + 0.006888716947287321, + 0.0022397558204829693, + 0.001981350826099515, + 0.0019515568856149912, + 0.0018796646036207676, + 0.002041181083768606, + 0.0018119609449058771, + 0.0016509435372427106, + 0.0027677183970808983, + 0.004724995698779821, + 0.00394425867125392, + 0.006380995735526085, + 0.008937654085457325, + 0.003579153213649988, + 0.0038518260698765516, + 0.004865692462772131, + 0.0041026645340025425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.005251067224889994 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_14/index_tensor([5279, 5279])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_14/index_tensor([5279, 5279])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a11d6088795258720a630e90f41faab64d5eb6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_14/index_tensor([5279, 5279])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, deliberate, medium shot, at eye level, begins the film's depiction of a couple in a tender embrace on a sun-drenched wooden deck. The man, with soft brown hair and dressed in a loose-fitting cream linen shirt, rests his hands gently on the woman's back. The woman, her long blond hair gleaming in the sunlight, is dressed in a simple light blue sundress. Her arms gently drape around his neck, her head resting on his shoulder, her eyes peacefully closed. The bright afternoon sun casts a warm golden glow across the scene, highlighting the grain of the deck's weathered cedar planks and creating soft, long shadows. As the camera smoothly slides forward, the frame tightens into an intimate close-up, focusing on their serene expressions and the tender connection between them. The lush, out-of-focus greenery of the background garden blurs into a soft bokeh, enhancing the deeply romantic and tranquil atmosphere of the moment.", + "seed": 25403, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_14/index_5279_seed_25403_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.007358313049189746, + "vlm_logprob": -0.007358313049189746, + "vlm_yes_no_margin": 5.55, + "vlm_token_credit": -0.007358313049189746, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007207109127193689 + ], + "opposite_logprob_tokens": [ + -6.257206916809082 + ], + "logprob_reward": -0.007207109127193689, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.007207109127193689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06596700847148895, + 0.07115420699119568, + 0.03768470883369446, + 0.06823283433914185, + 0.07054523378610611, + 0.045513663440942764, + 0.046750783920288086, + 0.4112682342529297, + 0.09324034303426743, + 0.08964294195175171 + ], + "credit_map": [ + 0.006578715983778238, + 0.008993775583803654, + 0.008150815032422543, + 0.009608193300664425, + 0.005883209407329559, + 0.00791982188820839, + 0.009637990966439247, + 0.009194489568471909, + 0.006772185210138559, + 0.005022101569920778, + 0.005358894821256399, + 0.007458533626049757, + 0.015705006197094917, + 0.008210782893002033, + 0.006487047765403986, + 0.01613966003060341, + 0.004490734077990055, + 0.0033597189467400312, + 0.003002497600391507, + 0.005692131817340851, + 0.003938098903745413, + 0.006066541187465191, + 0.004002192988991737, + 0.007132797501981258, + 0.005648825783282518, + 0.01030199509114027, + 0.0074268970638513565, + 0.01463062223047018, + 0.007752093952149153, + 0.009746287018060684, + 0.007212045602500439, + 0.005514072719961405, + 0.005490290932357311, + 0.008921896107494831, + 0.008766122162342072, + 0.010565508157014847, + 0.009115414693951607, + 0.008336568251252174, + 0.011795004829764366, + 0.007554436102509499, + 0.0036387848667800426, + 0.007640021853148937, + 0.004581276327371597, + 0.007137860171496868, + 0.007706027943640947, + 0.0055664959363639355, + 0.00628213956952095, + 0.002961061429232359, + 0.004892904311418533, + 0.005509732291102409, + 0.006675016134977341, + 0.004860483109951019, + 0.007731355261057615, + 0.007721198722720146, + 0.005318602547049522, + 0.004041492007672787, + 0.015531576238572598, + 0.0188753679394722, + 0.026221638545393944, + 0.156570702791214, + 0.024869205430150032, + 0.09374690800905228, + 0.05249045044183731, + 0.022962428629398346, + 0.005875034257769585, + 0.009616798721253872, + 0.0093846395611763, + 0.01470122393220663, + 0.019877644255757332, + 0.005218842066824436, + 0.018185580149292946, + 0.010380590334534645, + 0.008512561209499836, + 0.005022173281759024, + 0.007051992230117321, + 0.010603216476738453, + 0.012584356591105461, + 0.009921688586473465, + 0.02251410111784935, + 0.013432854786515236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.007207109127193689 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004108560271561146 + ], + "opposite_logprob_tokens": [ + -5.504108428955078 + ], + "logprob_reward": -0.004108560271561146, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004108560271561146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06637658923864365, + 0.1129363402724266, + 0.043511759489774704, + 0.07330186665058136, + 0.07796727865934372, + 0.08147160708904266, + 0.0301519725471735, + 0.28792068362236023, + 0.15492327511310577, + 0.07143861055374146 + ], + "credit_map": [ + 0.006924571469426155, + 0.009082561358809471, + 0.007898969575762749, + 0.008568662218749523, + 0.0061485376209020615, + 0.007519530598074198, + 0.012365455739200115, + 0.00786830484867096, + 0.007846113294363022, + 0.006274467799812555, + 0.00961659662425518, + 0.011008611880242825, + 0.026231933385133743, + 0.009607844054698944, + 0.010721415281295776, + 0.03162935748696327, + 0.003579192329198122, + 0.00470163905993104, + 0.005242790561169386, + 0.010337907820940018, + 0.005708374083042145, + 0.0053566861897706985, + 0.0050699845887720585, + 0.0035151902120560408, + 0.005294055677950382, + 0.007687658071517944, + 0.010345576331019402, + 0.015252764336764812, + 0.007254663854837418, + 0.008829534985125065, + 0.011118731461465359, + 0.007518886588513851, + 0.005629322957247496, + 0.00708787003532052, + 0.008850700221955776, + 0.011257636360824108, + 0.010562780313193798, + 0.009156573563814163, + 0.009527571499347687, + 0.015894819051027298, + 0.005919665563851595, + 0.01930459961295128, + 0.008782495744526386, + 0.011700457893311977, + 0.010069839656352997, + 0.01101948693394661, + 0.008639417588710785, + 0.006035651545971632, + 0.0023309532552957535, + 0.004337524063885212, + 0.0032704651821404696, + 0.004575095139443874, + 0.005174017511308193, + 0.0031226337887346745, + 0.00415094755589962, + 0.003190335351973772, + 0.013772665522992611, + 0.02411746047437191, + 0.017680155113339424, + 0.11841069906949997, + 0.021358858793973923, + 0.013422246091067791, + 0.03687426820397377, + 0.042284317314624786, + 0.008332629688084126, + 0.008604872971773148, + 0.023843184113502502, + 0.0291318129748106, + 0.02564835362136364, + 0.009046072140336037, + 0.03542693331837654, + 0.014889426529407501, + 0.005105688702315092, + 0.0028986402321606874, + 0.006151419598609209, + 0.014079228043556213, + 0.011867296881973743, + 0.007719884160906076, + 0.012999456375837326, + 0.010617001913487911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.004108560271561146 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001716570113785565 + ], + "opposite_logprob_tokens": [ + -6.376716613769531 + ], + "logprob_reward": -0.001716570113785565, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001716570113785565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05244975537061691, + 0.07107937335968018, + 0.04962125048041344, + 0.04628928005695343, + 0.034006647765636444, + 0.08035264164209366, + 0.057811494916677475, + 0.4165017008781433, + 0.14295479655265808, + 0.048933032900094986 + ], + "credit_map": [ + 0.005218489561229944, + 0.005641959607601166, + 0.006571244914084673, + 0.006640969309955835, + 0.004749064799398184, + 0.00678989989683032, + 0.009518055245280266, + 0.007320078555494547, + 0.00626674760133028, + 0.0039641475304961205, + 0.004328073002398014, + 0.005017298739403486, + 0.018474753946065903, + 0.007123688701540232, + 0.006311662029474974, + 0.019593004137277603, + 0.003644695272669196, + 0.005309259984642267, + 0.006399719975888729, + 0.01138814352452755, + 0.004777196794748306, + 0.006022148299962282, + 0.007616931106895208, + 0.004463158082216978, + 0.003142525441944599, + 0.004637780133634806, + 0.006509178783744574, + 0.009450777433812618, + 0.004273728933185339, + 0.0054297782480716705, + 0.007708182092756033, + 0.005137337371706963, + 0.0023986126761883497, + 0.004156417213380337, + 0.00419260049238801, + 0.005136858206242323, + 0.0038594515062868595, + 0.0037932307459414005, + 0.005215606186538935, + 0.0052538723684847355, + 0.004947622772306204, + 0.01554760430008173, + 0.008073966950178146, + 0.011050895787775517, + 0.011747539043426514, + 0.00873239804059267, + 0.011581390164792538, + 0.008671226911246777, + 0.0043211220763623714, + 0.007483574561774731, + 0.005295854061841965, + 0.007184829097241163, + 0.010936141014099121, + 0.007851029746234417, + 0.009331118315458298, + 0.005407830700278282, + 0.01849507912993431, + 0.02078157290816307, + 0.02571839839220047, + 0.18430106341838837, + 0.02387933060526848, + 0.02860165573656559, + 0.05705658346414566, + 0.05766802281141281, + 0.009030833840370178, + 0.008160986937582493, + 0.031112147495150566, + 0.03154570981860161, + 0.016522180289030075, + 0.0075883143581449986, + 0.023717375472187996, + 0.015277261845767498, + 0.0033997350838035345, + 0.0021731483284384012, + 0.003956596367061138, + 0.012601668015122414, + 0.007102925330400467, + 0.004139679484069347, + 0.00890300888568163, + 0.006656275596469641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.001716570113785565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031914988067001104 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.0031914988067001104, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031914988067001104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16323710978031158, + 0.17052900791168213, + 0.03713567927479744, + 0.07089699059724808, + 0.13352574408054352, + 0.036477189511060715, + 0.031741268932819366, + 0.1584945172071457, + 0.15769624710083008, + 0.04026619344949722 + ], + "credit_map": [ + 0.016586346551775932, + 0.014524227008223534, + 0.014203361235558987, + 0.017945844680070877, + 0.015639640390872955, + 0.027527745813131332, + 0.022825775668025017, + 0.03398418053984642, + 0.015578100457787514, + 0.007705416996032, + 0.008288120850920677, + 0.012826015241444111, + 0.04727312922477722, + 0.01687808521091938, + 0.015144309028983116, + 0.04683583974838257, + 0.002885909518226981, + 0.0035732234828174114, + 0.003878597402945161, + 0.008052141405642033, + 0.0036815390922129154, + 0.00490522338077426, + 0.004580676089972258, + 0.005578370299190283, + 0.00625889515504241, + 0.007358137518167496, + 0.009375059977173805, + 0.011559883132576942, + 0.006802602205425501, + 0.008591984398663044, + 0.009905606508255005, + 0.011044824495911598, + 0.010828258469700813, + 0.01389247551560402, + 0.016484856605529785, + 0.01895633153617382, + 0.019381409510970116, + 0.01666942983865738, + 0.020080136135220528, + 0.017232844606041908, + 0.0024361000396311283, + 0.005512030329555273, + 0.0035029833670705557, + 0.005615490488708019, + 0.00756909279152751, + 0.004570831079035997, + 0.00445583276450634, + 0.0028148326091468334, + 0.003287845989689231, + 0.0037662081886082888, + 0.00339640025049448, + 0.004859223496168852, + 0.0051467036828398705, + 0.002917539793998003, + 0.004495659843087196, + 0.003871686989441514, + 0.009139223955571651, + 0.013994470238685608, + 0.008392158895730972, + 0.021689696237444878, + 0.01495768316090107, + 0.03666103631258011, + 0.014969184063374996, + 0.03869107738137245, + 0.010940373875200748, + 0.007136685773730278, + 0.032793693244457245, + 0.03268933668732643, + 0.022441793233156204, + 0.0086054103448987, + 0.031186237931251526, + 0.011902730911970139, + 0.006140830460935831, + 0.0023558204993605614, + 0.003971053287386894, + 0.007395077962428331, + 0.00516147306188941, + 0.00446234131231904, + 0.005274217110127211, + 0.005505383014678955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0031914988067001104 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02056782692670822 + ], + "opposite_logprob_tokens": [ + -3.8955678939819336 + ], + "logprob_reward": -0.02056782692670822, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02056782692670822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046558868139982224, + 0.02721550315618515, + 0.019894376397132874, + 0.06428384780883789, + 0.08466045558452606, + 0.025670398026704788, + 0.040324095636606216, + 0.20030345022678375, + 0.4342561960220337, + 0.056832797825336456 + ], + "credit_map": [ + 0.0047152964398264885, + 0.006460358388721943, + 0.005078150890767574, + 0.008159440010786057, + 0.0037594863679260015, + 0.005073725711554289, + 0.006260803434997797, + 0.007051603868603706, + 0.002802341477945447, + 0.002876436337828636, + 0.0031373423989862204, + 0.003912073094397783, + 0.003838359145447612, + 0.0035585081204771996, + 0.0035054569598287344, + 0.0035849865525960922, + 0.002125666243955493, + 0.002238149056211114, + 0.0022633192129433155, + 0.0046170256100595, + 0.0019593762699514627, + 0.002518707886338234, + 0.002115604467689991, + 0.0020565262529999018, + 0.007366042118519545, + 0.00890145730227232, + 0.006729262415319681, + 0.010221289470791817, + 0.0069236489944159985, + 0.008581601083278656, + 0.009984409436583519, + 0.005576135125011206, + 0.004068329930305481, + 0.014532842673361301, + 0.010477944277226925, + 0.014660930261015892, + 0.01143363956362009, + 0.01159281749278307, + 0.009173212572932243, + 0.008720742538571358, + 0.00196894071996212, + 0.004666893742978573, + 0.002651851624250412, + 0.003784718457609415, + 0.004324415232986212, + 0.0032795367296785116, + 0.00259591918438673, + 0.002398120239377022, + 0.004129115492105484, + 0.004790029022842646, + 0.004444146528840065, + 0.0064575932919979095, + 0.006371880881488323, + 0.00399670097976923, + 0.005661503411829472, + 0.004473128821700811, + 0.01493543479591608, + 0.01470552571117878, + 0.01344915758818388, + 0.061228472739458084, + 0.016213055700063705, + 0.01543381530791521, + 0.026513943448662758, + 0.037824053317308426, + 0.0338529609143734, + 0.022145569324493408, + 0.10781724005937576, + 0.09050872176885605, + 0.05093584954738617, + 0.025000035762786865, + 0.06713496148586273, + 0.03686084225773811, + 0.0042068553157150745, + 0.0028989091515541077, + 0.004539597779512405, + 0.011055015958845615, + 0.00762938242405653, + 0.0049397642724215984, + 0.014148031361401081, + 0.007415238302201033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.02056782692670822 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberate, medium shot, at eye level, begins the film's depiction of a couple in a tender embrace on a sun-drenched wooden deck. The man, with soft brown hair and dressed in a loose-fitting cream linen shirt, rests his hands gently on the woman's back. The woman, her long blond hair gleaming in the sunlight, is dressed in a simple light blue sundress. Her arms gently drape around his neck, her head resting on his shoulder, her eyes peacefully closed. The bright afternoon sun casts a warm golden glow across the scene, highlighting the grain of the deck's weathered cedar planks and creating soft, long shadows. As the camera smoothly slides forward, the frame tightens into an intimate close-up, focusing on their serene expressions and the tender connection between them. The lush, out-of-focus greenery of the background garden blurs into a soft bokeh, enhancing the deeply romantic and tranquil atmosphere of the moment.", + "seed": 25404, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_14/index_5279_seed_25404_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.010640733130276202, + "vlm_logprob": -0.010640733130276202, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.010640733130276202, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010617434978485107 + ], + "opposite_logprob_tokens": [ + -5.135617256164551 + ], + "logprob_reward": -0.010617434978485107, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.010617434978485107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17522384226322174, + 0.09766727685928345, + 0.0768890529870987, + 0.04030529409646988, + 0.06497935950756073, + 0.060764484107494354, + 0.09766053408384323, + 0.19243302941322327, + 0.05381452292203903, + 0.14026261866092682 + ], + "credit_map": [ + 0.01656913198530674, + 0.019458269700407982, + 0.022092511877417564, + 0.03142579272389412, + 0.025866106152534485, + 0.015533882193267345, + 0.021028980612754822, + 0.023249175399541855, + 0.007925610989332199, + 0.014462900348007679, + 0.009230004623532295, + 0.008919098414480686, + 0.021981533616781235, + 0.007940883748233318, + 0.013552000746130943, + 0.013655252754688263, + 0.006486102472990751, + 0.012163386680185795, + 0.010517796501517296, + 0.008646257221698761, + 0.012891184538602829, + 0.006710047367960215, + 0.009464472532272339, + 0.010009808465838432, + 0.004981662146747112, + 0.004013860132545233, + 0.005096176639199257, + 0.005652722902595997, + 0.005907451268285513, + 0.004046333022415638, + 0.005257295444607735, + 0.005349790211766958, + 0.004773751366883516, + 0.006344195455312729, + 0.005325680132955313, + 0.007079262286424637, + 0.010016805492341518, + 0.012915351428091526, + 0.007917752489447594, + 0.010606559924781322, + 0.004647120367735624, + 0.00914087239652872, + 0.007722823414951563, + 0.01745602674782276, + 0.005122519098222256, + 0.0037845002952963114, + 0.007685592398047447, + 0.0052050338126719, + 0.013700340874493122, + 0.011305874213576317, + 0.01295938715338707, + 0.014021861366927624, + 0.009152045473456383, + 0.007722151465713978, + 0.017531409859657288, + 0.011267466470599174, + 0.0314289890229702, + 0.01583932153880596, + 0.025959203019738197, + 0.05868706852197647, + 0.007576913107186556, + 0.01790246181190014, + 0.02164723351597786, + 0.01339184120297432, + 0.007318397052586079, + 0.00613701157271862, + 0.011542423628270626, + 0.006773913744837046, + 0.0035091459285467863, + 0.004083410836756229, + 0.008491864427924156, + 0.005958354566246271, + 0.02079935371875763, + 0.018011612817645073, + 0.02237246371805668, + 0.02349252812564373, + 0.009646402671933174, + 0.011735763400793076, + 0.01836879923939705, + 0.01583569124341011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.010617434978485107 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009807496331632137 + ], + "opposite_logprob_tokens": [ + -4.634807586669922 + ], + "logprob_reward": -0.009807496331632137, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009807496331632137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1702650636434555, + 0.09940160065889359, + 0.2858283817768097, + 0.05487325042486191, + 0.04277805984020233, + 0.058381617069244385, + 0.05759267508983612, + 0.09988399595022202, + 0.09942035377025604, + 0.03157501295208931 + ], + "credit_map": [ + 0.015598246827721596, + 0.01509300246834755, + 0.025202829390764236, + 0.022930346429347992, + 0.029298143461346626, + 0.020296702161431313, + 0.02033303491771221, + 0.021512750536203384, + 0.00543575081974268, + 0.01665622554719448, + 0.007583881728351116, + 0.0164403785020113, + 0.01479924563318491, + 0.007857532240450382, + 0.015792621299624443, + 0.01483597420156002, + 0.014298900961875916, + 0.030523449182510376, + 0.018086951225996017, + 0.021651893854141235, + 0.037310484796762466, + 0.016274144873023033, + 0.07632821798324585, + 0.07135434448719025, + 0.005563642363995314, + 0.005706761963665485, + 0.005502499174326658, + 0.007247439585626125, + 0.009895730763673782, + 0.0062242597341537476, + 0.008105703629553318, + 0.006627216469496489, + 0.0026740578468889, + 0.003975271712988615, + 0.0046661836095154285, + 0.005204970948398113, + 0.008717294782400131, + 0.0059706345200538635, + 0.005277248099446297, + 0.006292395293712616, + 0.004908654373139143, + 0.006806728430092335, + 0.006891098339110613, + 0.008466498926281929, + 0.01133253425359726, + 0.003077285597100854, + 0.009736891835927963, + 0.007161923684179783, + 0.006623422261327505, + 0.005724365822970867, + 0.007598654367029667, + 0.008638394996523857, + 0.005078272894024849, + 0.011507934890687466, + 0.007435580249875784, + 0.004986050073057413, + 0.01181129552423954, + 0.012306850403547287, + 0.014729623682796955, + 0.014810358174145222, + 0.008904057554900646, + 0.013755319640040398, + 0.009269029833376408, + 0.014297458343207836, + 0.013659441843628883, + 0.01239715050905943, + 0.016864243894815445, + 0.019228393211960793, + 0.0068323989398777485, + 0.00443838769569993, + 0.015486477874219418, + 0.010513859800994396, + 0.0038077530916780233, + 0.004600140266120434, + 0.004508324898779392, + 0.005132748745381832, + 0.0017979824915528297, + 0.002994921989738941, + 0.004065532702952623, + 0.004667607136070728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.009807496331632137 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028254371136426926 + ], + "opposite_logprob_tokens": [ + -5.8778252601623535 + ], + "logprob_reward": -0.0028254371136426926, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028254371136426926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14218804240226746, + 0.08552069962024689, + 0.08834715187549591, + 0.1270613968372345, + 0.04837469384074211, + 0.11660564690828323, + 0.09337230026721954, + 0.10213548690080643, + 0.13416153192520142, + 0.062233101576566696 + ], + "credit_map": [ + 0.013537627644836903, + 0.02393135242164135, + 0.01780053973197937, + 0.019336244091391563, + 0.025523265823721886, + 0.01139337383210659, + 0.017454251646995544, + 0.01321137510240078, + 0.006437551695853472, + 0.010488087311387062, + 0.008628243580460548, + 0.015390834771096706, + 0.012435242533683777, + 0.006705114152282476, + 0.0109636215493083, + 0.01447200682014227, + 0.007060491945594549, + 0.011822470463812351, + 0.012273100204765797, + 0.013862687163054943, + 0.015933316200971603, + 0.004743919707834721, + 0.01118293683975935, + 0.011468230746686459, + 0.007587638217955828, + 0.013822733424603939, + 0.01168391015380621, + 0.01785765215754509, + 0.016915882006287575, + 0.014517380855977535, + 0.025366749614477158, + 0.019309435039758682, + 0.004711460787802935, + 0.0052876463159918785, + 0.005202431697398424, + 0.00757874408736825, + 0.010242072865366936, + 0.003471790347248316, + 0.006295245606452227, + 0.005585302133113146, + 0.010394255630671978, + 0.016051754355430603, + 0.021181847900152206, + 0.027269601821899414, + 0.010300890542566776, + 0.0051071420311927795, + 0.013404374942183495, + 0.01289577130228281, + 0.01104125939309597, + 0.010320867411792278, + 0.010709683410823345, + 0.012207955121994019, + 0.008488716557621956, + 0.02035280503332615, + 0.01267680898308754, + 0.007574201561510563, + 0.014696740545332432, + 0.013609062880277634, + 0.012109853327274323, + 0.022318625822663307, + 0.007672186475247145, + 0.01287542562931776, + 0.010596293956041336, + 0.008257299661636353, + 0.019195396453142166, + 0.017127472907304764, + 0.020359158515930176, + 0.0291975699365139, + 0.00858480017632246, + 0.009794190526008606, + 0.021176515147089958, + 0.008726416155695915, + 0.008017982356250286, + 0.007763609755784273, + 0.008882444351911545, + 0.007906786166131496, + 0.004316427279263735, + 0.0050310431979596615, + 0.011123103089630604, + 0.009191700257360935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.0028254371136426926 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036184806376695633 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036184806376695633, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036184806376695633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09999377280473709, + 0.11678597331047058, + 0.07817722111940384, + 0.13681946694850922, + 0.0514654740691185, + 0.048273369669914246, + 0.22928301990032196, + 0.10004742443561554, + 0.07821252197027206, + 0.060941845178604126 + ], + "credit_map": [ + 0.010314562357962132, + 0.011632362380623817, + 0.012319817207753658, + 0.015860823914408684, + 0.017163021489977837, + 0.00890839658677578, + 0.011077605187892914, + 0.012717186473309994, + 0.005986335687339306, + 0.018813906237483025, + 0.010196622461080551, + 0.015997620299458504, + 0.02141484059393406, + 0.010182985104620457, + 0.01739433780312538, + 0.016799326986074448, + 0.007117424625903368, + 0.00891560222953558, + 0.00522198248654604, + 0.007892943918704987, + 0.014435581862926483, + 0.004466544836759567, + 0.016043979674577713, + 0.014083161018788815, + 0.011418050155043602, + 0.015621527098119259, + 0.011515581049025059, + 0.01636030338704586, + 0.02317512221634388, + 0.017556050792336464, + 0.02085445635020733, + 0.020318374037742615, + 0.003239616984501481, + 0.0052530220709741116, + 0.005159022286534309, + 0.008965753950178623, + 0.01269522588700056, + 0.0036727874539792538, + 0.007503383792936802, + 0.004976664669811726, + 0.0034737566020339727, + 0.007422198075801134, + 0.004313736222684383, + 0.013726696372032166, + 0.005224756896495819, + 0.002809120574966073, + 0.007096072193235159, + 0.004207031335681677, + 0.026190802454948425, + 0.026018282398581505, + 0.02467793971300125, + 0.033452801406383514, + 0.023305164650082588, + 0.048136863857507706, + 0.025719961151480675, + 0.021781204268336296, + 0.01751963049173355, + 0.008557508699595928, + 0.013225284405052662, + 0.02649087645113468, + 0.00544886477291584, + 0.01017099991440773, + 0.010970677249133587, + 0.007663580123335123, + 0.01232186984270811, + 0.0072943479754030704, + 0.019200120121240616, + 0.011658073402941227, + 0.005208487156778574, + 0.007439008913934231, + 0.009120154194533825, + 0.005970458500087261, + 0.00841071642935276, + 0.007752575445920229, + 0.010180078446865082, + 0.009981805458664894, + 0.0036418733652681112, + 0.0056878961622715, + 0.00773710897192359, + 0.007549789268523455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.0036184806376695633 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026334816589951515 + ], + "opposite_logprob_tokens": [ + -3.651334762573242 + ], + "logprob_reward": -0.026334816589951515, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026334816589951515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058642324060201645, + 0.10737745463848114, + 0.15787753462791443, + 0.05725902318954468, + 0.04212154448032379, + 0.12215308099985123, + 0.12923267483711243, + 0.18611818552017212, + 0.08244633674621582, + 0.056771740317344666 + ], + "credit_map": [ + 0.005143984220921993, + 0.006706235930323601, + 0.0059797377325594425, + 0.008740109391510487, + 0.00959835946559906, + 0.006458393298089504, + 0.007785068359225988, + 0.008230440318584442, + 0.005878460127860308, + 0.016997478902339935, + 0.009758826345205307, + 0.012686528265476227, + 0.01956942304968834, + 0.010568441823124886, + 0.017698142677545547, + 0.014220152050256729, + 0.01058737002313137, + 0.015931004658341408, + 0.009590133093297482, + 0.011426948010921478, + 0.01604597084224224, + 0.004973741713911295, + 0.04994283244013786, + 0.039379559457302094, + 0.007587313186377287, + 0.005131466314196587, + 0.005967508070170879, + 0.008179157972335815, + 0.010041606612503529, + 0.004408151376992464, + 0.009414168074727058, + 0.006529655773192644, + 0.00519814295694232, + 0.004332434851676226, + 0.005153011996299028, + 0.006608360912650824, + 0.00593903660774231, + 0.004516880959272385, + 0.005501362960785627, + 0.004872316960245371, + 0.008471142500638962, + 0.017545949667692184, + 0.011797400191426277, + 0.03812330216169357, + 0.010593931190669537, + 0.005701535847038031, + 0.017519406974315643, + 0.012400418519973755, + 0.009235024452209473, + 0.01575503870844841, + 0.023057347163558006, + 0.024378512054681778, + 0.011845128610730171, + 0.008255446329712868, + 0.027686184272170067, + 0.009020001627504826, + 0.031049644574522972, + 0.013691408559679985, + 0.03228238597512245, + 0.052974384278059006, + 0.008069263771176338, + 0.01031318586319685, + 0.025527168065309525, + 0.012210755608975887, + 0.01189951878041029, + 0.00998112466186285, + 0.014361673034727573, + 0.013500542379915714, + 0.006860662251710892, + 0.008813259191811085, + 0.009281345643103123, + 0.007748223841190338, + 0.007857520133256912, + 0.007512103300541639, + 0.010172130540013313, + 0.007818321697413921, + 0.004010842181742191, + 0.003975447732955217, + 0.007717905566096306, + 0.007707477081567049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.026334816589951515 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberate, medium shot, at eye level, begins the film's depiction of a couple in a tender embrace on a sun-drenched wooden deck. The man, with soft brown hair and dressed in a loose-fitting cream linen shirt, rests his hands gently on the woman's back. The woman, her long blond hair gleaming in the sunlight, is dressed in a simple light blue sundress. Her arms gently drape around his neck, her head resting on his shoulder, her eyes peacefully closed. The bright afternoon sun casts a warm golden glow across the scene, highlighting the grain of the deck's weathered cedar planks and creating soft, long shadows. As the camera smoothly slides forward, the frame tightens into an intimate close-up, focusing on their serene expressions and the tender connection between them. The lush, out-of-focus greenery of the background garden blurs into a soft bokeh, enhancing the deeply romantic and tranquil atmosphere of the moment.", + "seed": 25405, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_14/index_5279_seed_25405_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.013770864717662335, + "vlm_logprob": -0.013770864717662335, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.013770864717662335, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02232595719397068 + ], + "opposite_logprob_tokens": [ + -4.0223259925842285 + ], + "logprob_reward": -0.02232595719397068, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02232595719397068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09824405610561371, + 0.022939065471291542, + 0.048584870994091034, + 0.18002623319625854, + 0.05857638269662857, + 0.10083601623773575, + 0.0967814177274704, + 0.03935082256793976, + 0.1977100521326065, + 0.15695104002952576 + ], + "credit_map": [ + 0.01858040876686573, + 0.008417048491537571, + 0.008042803965508938, + 0.013357874006032944, + 0.011916573159396648, + 0.008489178493618965, + 0.018168054521083832, + 0.011272109113633633, + 0.0019514313898980618, + 0.0014127555768936872, + 0.002894400618970394, + 0.005125189665704966, + 0.002684905892238021, + 0.0013469774276018143, + 0.005004641599953175, + 0.002518764231353998, + 0.006749807856976986, + 0.003188385860994458, + 0.004591218661516905, + 0.008318332955241203, + 0.006151964422315359, + 0.005389409139752388, + 0.008548961952328682, + 0.005646795034408569, + 0.04281822219491005, + 0.009222240187227726, + 0.04625846818089485, + 0.027925783768296242, + 0.011230893433094025, + 0.02672671712934971, + 0.00899931788444519, + 0.006844609044492245, + 0.011628805659711361, + 0.004375738557428122, + 0.007602918893098831, + 0.00462345639243722, + 0.004451435524970293, + 0.015106317587196827, + 0.006778310518711805, + 0.004009395372122526, + 0.02503392845392227, + 0.004487937781959772, + 0.0068217492662370205, + 0.008429694920778275, + 0.006138462573289871, + 0.02848418802022934, + 0.01510221790522337, + 0.006337836384773254, + 0.016525711864233017, + 0.004937007091939449, + 0.009468749165534973, + 0.011956225149333477, + 0.004101808648556471, + 0.026482094079256058, + 0.010781960561871529, + 0.01252786349505186, + 0.00312402262352407, + 0.005176382139325142, + 0.004328163806349039, + 0.009053409099578857, + 0.003197851125150919, + 0.004182496108114719, + 0.007040123455226421, + 0.0032483760733157396, + 0.012881975620985031, + 0.03914108872413635, + 0.019671237096190453, + 0.025427306070923805, + 0.04081263765692711, + 0.01612250506877899, + 0.025522669777274132, + 0.01813063584268093, + 0.00938916765153408, + 0.025388041511178017, + 0.010364165529608727, + 0.03724806383252144, + 0.019310815259814262, + 0.014221684075891972, + 0.028651803731918335, + 0.012377315200865269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.02232595719397068 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008660613559186459 + ], + "opposite_logprob_tokens": [ + -4.758660793304443 + ], + "logprob_reward": -0.008660613559186459, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008660613559186459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.070672906935215, + 0.018667707219719887, + 0.06225626915693283, + 0.1056145504117012, + 0.02726389467716217, + 0.07114335894584656, + 0.05054929852485657, + 0.04830444976687431, + 0.48339077830314636, + 0.06213681027293205 + ], + "credit_map": [ + 0.010076937265694141, + 0.005687496159225702, + 0.006140417885035276, + 0.01139304880052805, + 0.008975263684988022, + 0.004804755095392466, + 0.014238801784813404, + 0.009356184862554073, + 0.002371591981500387, + 0.0013778829015791416, + 0.001803843304514885, + 0.003188838018104434, + 0.003044083248823881, + 0.0016480258200317621, + 0.0033666687086224556, + 0.0018667732365429401, + 0.006785563193261623, + 0.005202534142881632, + 0.00616227462887764, + 0.011740782298147678, + 0.00953417643904686, + 0.006740937475115061, + 0.01016089878976345, + 0.005929096601903439, + 0.027378765866160393, + 0.00634277006611228, + 0.02192908152937889, + 0.011667576618492603, + 0.007764420006424189, + 0.017506174743175507, + 0.008068581111729145, + 0.004957192577421665, + 0.004616496153175831, + 0.0021141173783689737, + 0.00354730524122715, + 0.004343717359006405, + 0.002267079195007682, + 0.004553959704935551, + 0.0034760264679789543, + 0.0023451964370906353, + 0.018811006098985672, + 0.003902141936123371, + 0.004055905621498823, + 0.009349368512630463, + 0.0033640654291957617, + 0.0216685701161623, + 0.0066709998063743114, + 0.003321297699585557, + 0.009180424734950066, + 0.0034431417006999254, + 0.0038073884788900614, + 0.004422344733029604, + 0.002746223472058773, + 0.014857131987810135, + 0.004019072744995356, + 0.00807357020676136, + 0.004042401444166899, + 0.006756582297384739, + 0.00349780498072505, + 0.009718202985823154, + 0.0034944589715451, + 0.006884723901748657, + 0.00793616846203804, + 0.005974109284579754, + 0.02690044231712818, + 0.09619570523500443, + 0.039237700402736664, + 0.06348580867052078, + 0.10630226135253906, + 0.051266517490148544, + 0.06827308237552643, + 0.03172927722334862, + 0.004212257917970419, + 0.009925322607159615, + 0.0039930082857608795, + 0.011458855122327805, + 0.009926932863891125, + 0.0035855870228260756, + 0.013269017450511456, + 0.005765831097960472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -0.008660613559186459 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028257938101887703 + ], + "opposite_logprob_tokens": [ + -5.877825736999512 + ], + "logprob_reward": -0.0028257938101887703, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028257938101887703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06629390269517899, + 0.021935580298304558, + 0.24733999371528625, + 0.2906075417995453, + 0.047700926661491394, + 0.058314014226198196, + 0.08585423976182938, + 0.07354485243558884, + 0.03867526724934578, + 0.06973369419574738 + ], + "credit_map": [ + 0.009638950228691101, + 0.005196231883019209, + 0.005311261862516403, + 0.008570093661546707, + 0.011500630527734756, + 0.005111867096275091, + 0.011840841732919216, + 0.009124021045863628, + 0.002813269617035985, + 0.0015561891486868262, + 0.002528227400034666, + 0.003811677685007453, + 0.0021206545643508434, + 0.001862053177319467, + 0.0050090281292796135, + 0.0022344801109284163, + 0.03333316370844841, + 0.020266007632017136, + 0.02500658668577671, + 0.03538762032985687, + 0.042108479887247086, + 0.03897465020418167, + 0.03168364614248276, + 0.020579833537340164, + 0.07493873685598373, + 0.009975473396480083, + 0.08033868670463562, + 0.045187197625637054, + 0.008656083606183529, + 0.04921264573931694, + 0.011270659044384956, + 0.011028031818568707, + 0.011468412354588509, + 0.0026198874693363905, + 0.005122288130223751, + 0.004580426961183548, + 0.002385031431913376, + 0.014502923004329205, + 0.004420969635248184, + 0.0026009820867329836, + 0.018159743398427963, + 0.0028997696936130524, + 0.002636892022565007, + 0.003108185948804021, + 0.002605604473501444, + 0.01985398307442665, + 0.005437968764454126, + 0.0036118673160672188, + 0.01694871485233307, + 0.004285968840122223, + 0.0053893704898655415, + 0.006031916476786137, + 0.0037945611402392387, + 0.029499048367142677, + 0.005382220726460218, + 0.01452243234962225, + 0.0059119039215147495, + 0.008155934512615204, + 0.005921078845858574, + 0.018836231902241707, + 0.0047849612310528755, + 0.01263198722153902, + 0.012726313434541225, + 0.0045764390379190445, + 0.004339202307164669, + 0.006657056976109743, + 0.0036360309459269047, + 0.00674406485632062, + 0.0035580918192863464, + 0.0051255314610898495, + 0.005433815531432629, + 0.0031814719550311565, + 0.003780962899327278, + 0.008244896307587624, + 0.008816767483949661, + 0.013614444062113762, + 0.009863123297691345, + 0.005730697885155678, + 0.013921831734478474, + 0.005760962143540382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 6 + ], + "selected_reward": -0.0028257938101887703 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052548618987202644 + ], + "opposite_logprob_tokens": [ + -5.255254745483398 + ], + "logprob_reward": -0.0052548618987202644, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052548618987202644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1187795028090477, + 0.034527167677879333, + 0.05173638463020325, + 0.14602595567703247, + 0.062379997223615646, + 0.10652907937765121, + 0.0951695665717125, + 0.039000727236270905, + 0.05110163986682892, + 0.2947499454021454 + ], + "credit_map": [ + 0.019588811323046684, + 0.009118851274251938, + 0.011189863085746765, + 0.01750594936311245, + 0.011432956904172897, + 0.00984308309853077, + 0.020608747377991676, + 0.019491251558065414, + 0.005338061135262251, + 0.002850001910701394, + 0.004094399977475405, + 0.004502670373767614, + 0.003727053524926305, + 0.0043219029903411865, + 0.005702253431081772, + 0.003990824334323406, + 0.010009736754000187, + 0.00445022527128458, + 0.006558607332408428, + 0.005957706831395626, + 0.00611139927059412, + 0.00736583024263382, + 0.006213878281414509, + 0.005069006234407425, + 0.037941474467515945, + 0.005076553672552109, + 0.04069971665740013, + 0.021966909989714622, + 0.004911799915134907, + 0.024270912632346153, + 0.005932911764830351, + 0.005225683096796274, + 0.013496299274265766, + 0.0040972549468278885, + 0.00967460684478283, + 0.008695151656866074, + 0.004454429727047682, + 0.011614971794188023, + 0.005962524097412825, + 0.004384760744869709, + 0.03509213402867317, + 0.004510539583861828, + 0.004403888247907162, + 0.009269542060792446, + 0.0038399372715502977, + 0.03793584927916527, + 0.006910674273967743, + 0.004566522315144539, + 0.017865650355815887, + 0.005889554973691702, + 0.006371951196342707, + 0.010393600910902023, + 0.0044556050561368465, + 0.028044650331139565, + 0.008438418619334698, + 0.013710145838558674, + 0.003571901237592101, + 0.004913718439638615, + 0.0027065800968557596, + 0.008177184499800205, + 0.0025143595412373543, + 0.008095861412584782, + 0.005889557301998138, + 0.0031315672677010298, + 0.004289087373763323, + 0.008367358706891537, + 0.007789197377860546, + 0.00764033617451787, + 0.007139434572309256, + 0.004974119830876589, + 0.0057912711054086685, + 0.005110836587846279, + 0.01531907171010971, + 0.04757789894938469, + 0.02711154893040657, + 0.06235117092728615, + 0.04512466862797737, + 0.019782833755016327, + 0.060037754476070404, + 0.017445016652345657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.0052548618987202644 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0297870971262455 + ], + "opposite_logprob_tokens": [ + -3.529787063598633 + ], + "logprob_reward": -0.0297870971262455, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0297870971262455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045528847724199295, + 0.024645118042826653, + 0.03825148195028305, + 0.09390061348676682, + 0.053552962839603424, + 0.14508698880672455, + 0.25612926483154297, + 0.23693516850471497, + 0.045702096074819565, + 0.06026744097471237 + ], + "credit_map": [ + 0.006459760945290327, + 0.0042551662772893906, + 0.0033347993157804012, + 0.006305094342678785, + 0.004822054412215948, + 0.0035050224978476763, + 0.010107259266078472, + 0.006739689968526363, + 0.0030357048381119967, + 0.00199267640709877, + 0.0025798161514103413, + 0.003953319974243641, + 0.004053115379065275, + 0.002143354620784521, + 0.005057192873209715, + 0.0018299365183338523, + 0.00505288178101182, + 0.002502351300790906, + 0.0025223696138709784, + 0.004719950258731842, + 0.006076715420931578, + 0.007422530557960272, + 0.005975206382572651, + 0.003979476168751717, + 0.012434659525752068, + 0.006726847030222416, + 0.008499880321323872, + 0.012597757391631603, + 0.012683957815170288, + 0.012595918029546738, + 0.015779711306095123, + 0.012581873685121536, + 0.010646100156009197, + 0.003832106711342931, + 0.008018176071345806, + 0.005679080728441477, + 0.004584446549415588, + 0.010745815932750702, + 0.00663352757692337, + 0.0034137063194066286, + 0.02684560976922512, + 0.012470545247197151, + 0.012750650756061077, + 0.028224430978298187, + 0.008948597125709057, + 0.027120515704154968, + 0.019401995465159416, + 0.009324636310338974, + 0.04980701208114624, + 0.012919262982904911, + 0.014712261967360973, + 0.01951790601015091, + 0.011470980942249298, + 0.08526899665594101, + 0.017806939780712128, + 0.0446259006857872, + 0.023647133260965347, + 0.025081999599933624, + 0.022922299802303314, + 0.04900289699435234, + 0.01584811322391033, + 0.05041911080479622, + 0.029264792799949646, + 0.020748818293213844, + 0.0040737297385931015, + 0.00842414703220129, + 0.005189047195017338, + 0.009407890029251575, + 0.00483451085165143, + 0.004502854309976101, + 0.006237296387553215, + 0.003032616339623928, + 0.0028605919796973467, + 0.009170141071081161, + 0.0056081656366586685, + 0.011467943899333477, + 0.008897610940039158, + 0.004420737735927105, + 0.01272410899400711, + 0.005118141416460276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.0297870971262455 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberate, medium shot, at eye level, begins the film's depiction of a couple in a tender embrace on a sun-drenched wooden deck. The man, with soft brown hair and dressed in a loose-fitting cream linen shirt, rests his hands gently on the woman's back. The woman, her long blond hair gleaming in the sunlight, is dressed in a simple light blue sundress. Her arms gently drape around his neck, her head resting on his shoulder, her eyes peacefully closed. The bright afternoon sun casts a warm golden glow across the scene, highlighting the grain of the deck's weathered cedar planks and creating soft, long shadows. As the camera smoothly slides forward, the frame tightens into an intimate close-up, focusing on their serene expressions and the tender connection between them. The lush, out-of-focus greenery of the background garden blurs into a soft bokeh, enhancing the deeply romantic and tranquil atmosphere of the moment.", + "seed": 25406, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_14/index_5279_seed_25406_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.015572916343808174, + "vlm_logprob": -0.015572916343808174, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.015572916343808174, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008165311068296432 + ], + "opposite_logprob_tokens": [ + -6.13316535949707 + ], + "logprob_reward": -0.008165311068296432, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.008165311068296432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2923662066459656, + 0.08600128442049026, + 0.21175038814544678, + 0.052519604563713074, + 0.049571432173252106, + 0.051672860980033875, + 0.03745061531662941, + 0.04683874174952507, + 0.03751380369067192, + 0.13431505858898163 + ], + "credit_map": [ + 0.05141717940568924, + 0.02993187867105007, + 0.012686308473348618, + 0.04297640547156334, + 0.012869942933321, + 0.013190262019634247, + 0.03645634278655052, + 0.09283789992332458, + 0.01136404275894165, + 0.006849692203104496, + 0.008788451552391052, + 0.008721254765987396, + 0.006644249428063631, + 0.007950486615300179, + 0.024397943168878555, + 0.01128516998142004, + 0.053510263562202454, + 0.086653932929039, + 0.010278832167387009, + 0.009664005599915981, + 0.009674802422523499, + 0.021014252677559853, + 0.014226020313799381, + 0.006728263106197119, + 0.008792524226009846, + 0.013312598690390587, + 0.005167036782950163, + 0.004157453775405884, + 0.005776533856987953, + 0.004359078127890825, + 0.0063079954124987125, + 0.004646380431950092, + 0.010112461633980274, + 0.017156125977635384, + 0.003108784556388855, + 0.007063254248350859, + 0.0030762923415750265, + 0.0022803996689617634, + 0.0045663584023714066, + 0.0022077555768191814, + 0.009653172455728054, + 0.013037366792559624, + 0.005118979141116142, + 0.008735841140151024, + 0.0038926557172089815, + 0.0023154262453317642, + 0.0061653973534703255, + 0.0027540214359760284, + 0.0065085687674582005, + 0.008085335604846478, + 0.005773867480456829, + 0.006301121320575476, + 0.002301840577274561, + 0.002346426947042346, + 0.003127551171928644, + 0.0030059039127081633, + 0.007799886167049408, + 0.006020166911184788, + 0.004184308461844921, + 0.007385161705315113, + 0.0037575853057205677, + 0.005113546270877123, + 0.006772241555154324, + 0.005805844906717539, + 0.006277449429035187, + 0.005007701925933361, + 0.0034013662952929735, + 0.004231403116136789, + 0.003080488182604313, + 0.00482052331790328, + 0.0051205093041062355, + 0.005574364215135574, + 0.044430267065763474, + 0.018857669085264206, + 0.008681496605277061, + 0.010991615243256092, + 0.007326487917453051, + 0.01309641357511282, + 0.010555418208241463, + 0.020375695079565048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.008165311068296432 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764338206499815 + ], + "opposite_logprob_tokens": [ + -4.882643222808838 + ], + "logprob_reward": -0.00764338206499815, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764338206499815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22347110509872437, + 0.05106374993920326, + 0.12430036067962646, + 0.0827159509062767, + 0.040377456694841385, + 0.10710739344358444, + 0.03371594101190567, + 0.03583666682243347, + 0.0513308010995388, + 0.25008055567741394 + ], + "credit_map": [ + 0.03490152955055237, + 0.021205062046647072, + 0.014384734444320202, + 0.030170517042279243, + 0.019374966621398926, + 0.015207161195576191, + 0.03136320412158966, + 0.0568639375269413, + 0.006834807805716991, + 0.004544919356703758, + 0.0037124718073755503, + 0.006627604365348816, + 0.0034608065616339445, + 0.004305779002606869, + 0.014719697646796703, + 0.006857665255665779, + 0.031133202835917473, + 0.050094880163669586, + 0.004410970956087112, + 0.0063084145076572895, + 0.0063131083734333515, + 0.012173265218734741, + 0.007858886383473873, + 0.0060076299123466015, + 0.019883451983332634, + 0.0298444963991642, + 0.006025049369782209, + 0.005794164724647999, + 0.006785059813410044, + 0.004121619742363691, + 0.006209122017025948, + 0.004052986856549978, + 0.005680765490978956, + 0.004431535955518484, + 0.002769788494333625, + 0.006819330621510744, + 0.0059711383655667305, + 0.0029845989774912596, + 0.006935090757906437, + 0.004785208962857723, + 0.0167890265583992, + 0.020965153351426125, + 0.006176862865686417, + 0.02473117783665657, + 0.005694583058357239, + 0.00439088512212038, + 0.0196506567299366, + 0.008709058165550232, + 0.0071173072792589664, + 0.011349611915647984, + 0.003809774061664939, + 0.002356380457058549, + 0.0018175668083131313, + 0.0024180873297154903, + 0.002656307304278016, + 0.002190906321629882, + 0.007424887269735336, + 0.004962428938597441, + 0.0026096117217093706, + 0.005153854377567768, + 0.0026925806887447834, + 0.003169245785102248, + 0.0057481746189296246, + 0.004075883887708187, + 0.008111858740448952, + 0.005989799741655588, + 0.0037428257055580616, + 0.006093773525208235, + 0.004680861718952656, + 0.006709431763738394, + 0.009695583954453468, + 0.006306667812168598, + 0.0993800088763237, + 0.039946503937244415, + 0.017452722415328026, + 0.017310835421085358, + 0.008434044197201729, + 0.01432949211448431, + 0.020092837512493134, + 0.033134110271930695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.00764338206499815 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464171078056097 + ], + "opposite_logprob_tokens": [ + -5.379641532897949 + ], + "logprob_reward": -0.00464171078056097, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464171078056097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18971958756446838, + 0.08264569938182831, + 0.19130180776119232, + 0.06217643991112709, + 0.11696617305278778, + 0.149592325091362, + 0.09364674240350723, + 0.042018625885248184, + 0.04044654220342636, + 0.03148599714040756 + ], + "credit_map": [ + 0.029441429302096367, + 0.017929015681147575, + 0.010316760279238224, + 0.02913815900683403, + 0.008787287399172783, + 0.008644850924611092, + 0.02446528524160385, + 0.060996804386377335, + 0.011329780332744122, + 0.005697048269212246, + 0.004271578509360552, + 0.009362831711769104, + 0.00621733907610178, + 0.007694969419389963, + 0.027667932212352753, + 0.010404235683381557, + 0.04806853458285332, + 0.0778631716966629, + 0.006841183640062809, + 0.008973360992968082, + 0.010592264123260975, + 0.01870417222380638, + 0.012729233130812645, + 0.007529911119490862, + 0.015147156082093716, + 0.022988999262452126, + 0.0050888340920209885, + 0.0038866971153765917, + 0.0036509670317173004, + 0.003100764472037554, + 0.004346924833953381, + 0.003966100979596376, + 0.021112410351634026, + 0.024581605568528175, + 0.007346959784626961, + 0.01162733044475317, + 0.01335398480296135, + 0.007342839147895575, + 0.016131484881043434, + 0.015469570644199848, + 0.01589113660156727, + 0.025326024740934372, + 0.008905823342502117, + 0.03574954345822334, + 0.00838281586766243, + 0.008484693244099617, + 0.02736530639231205, + 0.019486980512738228, + 0.020749010145664215, + 0.03428328037261963, + 0.006198040209710598, + 0.013473302125930786, + 0.004299303516745567, + 0.0037987553514540195, + 0.006507853977382183, + 0.004337202291935682, + 0.009180174209177494, + 0.00691566988825798, + 0.0035508691798895597, + 0.004647304303944111, + 0.0039678411558270454, + 0.003028903156518936, + 0.006131845526397228, + 0.0045960210263729095, + 0.0070608449168503284, + 0.005617510061711073, + 0.0027998739387840033, + 0.007428589276969433, + 0.0023214297834783792, + 0.0036903610453009605, + 0.004618134815245867, + 0.00690980302169919, + 0.005059557966887951, + 0.002654219977557659, + 0.002574037527665496, + 0.004583948757499456, + 0.0031050334218889475, + 0.0036457267124205828, + 0.004211817402392626, + 0.005651656072586775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.00464171078056097 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024871863424777985 + ], + "opposite_logprob_tokens": [ + -6.0024871826171875 + ], + "logprob_reward": -0.0024871863424777985, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024871863424777985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09418779611587524, + 0.07054611295461655, + 0.2502530515193939, + 0.1672530472278595, + 0.049826595932245255, + 0.10414429008960724, + 0.09404006600379944, + 0.06348302960395813, + 0.037733953446149826, + 0.0685320645570755 + ], + "credit_map": [ + 0.009816908277571201, + 0.010939202271401882, + 0.011911501176655293, + 0.015640048310160637, + 0.01173253171145916, + 0.010386358946561813, + 0.008298616856336594, + 0.01546263787895441, + 0.007949885912239552, + 0.008820513263344765, + 0.01120162196457386, + 0.007432611659169197, + 0.0064374725334346294, + 0.007211577612906694, + 0.013544362969696522, + 0.007948068901896477, + 0.06485655903816223, + 0.10420914739370346, + 0.00840704794973135, + 0.012017118744552135, + 0.010760083794593811, + 0.025647656992077827, + 0.015310985967516899, + 0.009044457226991653, + 0.04454416409134865, + 0.06715847551822662, + 0.011142246425151825, + 0.008344791829586029, + 0.00860179215669632, + 0.008825747296214104, + 0.012022214941680431, + 0.006613607052713633, + 0.009576375596225262, + 0.012145528569817543, + 0.005443538073450327, + 0.005292357411235571, + 0.00438686041161418, + 0.0026317844167351723, + 0.005447503179311752, + 0.004902651067823172, + 0.035688310861587524, + 0.022554952651262283, + 0.009569672867655754, + 0.013594520278275013, + 0.004415083676576614, + 0.002983572194352746, + 0.011514313519001007, + 0.0038238721899688244, + 0.02247653715312481, + 0.03592052683234215, + 0.008812721818685532, + 0.007478634361177683, + 0.005114517640322447, + 0.0038332133553922176, + 0.005701092537492514, + 0.004702826030552387, + 0.02111116610467434, + 0.013298721984028816, + 0.004625435918569565, + 0.004736984148621559, + 0.003940634895116091, + 0.0036330102011561394, + 0.004859265871345997, + 0.007277811411768198, + 0.006302583031356335, + 0.005019544158130884, + 0.002442365512251854, + 0.005763314198702574, + 0.0036429292522370815, + 0.0031655863858759403, + 0.006601550150662661, + 0.004796083550900221, + 0.0157232154160738, + 0.007160630077123642, + 0.003698944114148617, + 0.0059210010804235935, + 0.004567605908960104, + 0.009914484806358814, + 0.006959026213735342, + 0.014587164856493473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.0024871863424777985 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a man and a woman in a tender embrace on a wooden deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting cream linen shirt with soft brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a simple light blue sundress with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a sun-drenched environment with warm golden lighting on weathered cedar planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a medium shot and smoothly slide forward into an intimate close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05492699146270752 + ], + "opposite_logprob_tokens": [ + -2.929926872253418 + ], + "logprob_reward": -0.05492699146270752, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05492699146270752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07849918305873871, + 0.043304864317178726, + 0.10205191373825073, + 0.0504189133644104, + 0.1214938685297966, + 0.10210361331701279, + 0.06413448601961136, + 0.08133731782436371, + 0.13290081918239594, + 0.22375507652759552 + ], + "credit_map": [ + 0.00962697435170412, + 0.006495965179055929, + 0.007389855105429888, + 0.013230716809630394, + 0.005420300643891096, + 0.006168510764837265, + 0.010144313797354698, + 0.020022545009851456, + 0.004502208437770605, + 0.004773923195898533, + 0.005452306941151619, + 0.006556337233632803, + 0.004125936888158321, + 0.005601603537797928, + 0.00591404689475894, + 0.006378497462719679, + 0.019605642184615135, + 0.027157451957464218, + 0.00993111077696085, + 0.007525356486439705, + 0.01195838674902916, + 0.00989331491291523, + 0.009084990248084068, + 0.006895646918565035, + 0.006876239087432623, + 0.006941702216863632, + 0.004127683117985725, + 0.007452176883816719, + 0.0074249692261219025, + 0.0051542483270168304, + 0.007375463377684355, + 0.0050664301961660385, + 0.03240006044507027, + 0.05080661177635193, + 0.00458334106951952, + 0.008403895422816277, + 0.00620610173791647, + 0.00569823058322072, + 0.007322660647332668, + 0.006072955206036568, + 0.016489604488015175, + 0.021435769274830818, + 0.005862311460077763, + 0.024850478395819664, + 0.005074379500001669, + 0.0034744059666991234, + 0.01934535801410675, + 0.005571300163865089, + 0.005135624669492245, + 0.006612604018300772, + 0.005604943726211786, + 0.018842017278075218, + 0.0034206029959023, + 0.005764845758676529, + 0.014400576241314411, + 0.0043532694689929485, + 0.009680602699518204, + 0.008845672942698002, + 0.005927427671849728, + 0.014134196564555168, + 0.01002977043390274, + 0.009138597175478935, + 0.013076185248792171, + 0.01050486508756876, + 0.010877318680286407, + 0.011872660368680954, + 0.008860865607857704, + 0.02192106656730175, + 0.012222720310091972, + 0.029645638540387154, + 0.017421793192625046, + 0.020078744739294052, + 0.057933367788791656, + 0.025722170248627663, + 0.013104663230478764, + 0.017321031540632248, + 0.01377787720412016, + 0.02867450937628746, + 0.02151263691484928, + 0.045708805322647095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.05492699146270752 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_15/index_tensor([4396, 4396])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_15/index_tensor([4396, 4396])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..54aff3b6a35ad55db9236cf614d394e6b52d8b47 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_15/index_tensor([4396, 4396])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "perfectly static, cinematic long shot depicts the breathtaking expanse of a remote, rugged mountainside bathed in the warm, low-angle light of golden hour. A massive pine tree stands at the majestic center of the scene, its ancient, towering form and deep green needles silhouetted against the pale, clear sky. Suddenly, a sharp explosion shatters the profound silence, echoing throughout the valley. The pine's massive trunk shudders violently, and near the base, the dark, gnarled bark bursts open, revealing a chaotic explosion of pale, raw wood fibers splintering outward. The tree begins to tilt, its movement beginning with a slow, deliberate groan before accelerating with a terrifying, decisive momentum, carving a majestic, tragic arc through the crisp mountain air. As it falls, its massive shadow sweeps swiftly across the rocky terrain below, its branches streaking through the air with a final, violent hiss before impacting the ground with a deafening, earth-shattering crash, sending a huge whirlwind of dust, soil, and pine needles into golden light, creating an atmosphere of raw, natural power and a stunning finale.", + "seed": 27274, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_15/index_4396_seed_27274_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.008321270416490733, + "vlm_logprob": -0.008321270416490733, + "vlm_yes_no_margin": 5.5, + "vlm_token_credit": -0.008321270416490733, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006038990803062916 + ], + "opposite_logprob_tokens": [ + -6.256039142608643 + ], + "logprob_reward": -0.006038990803062916, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.006038990803062916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05736589804291725, + 0.03405170515179634, + 0.014509295113384724, + 0.022599224001169205, + 0.13097453117370605, + 0.10411318391561508, + 0.09047342091798782, + 0.3132624328136444, + 0.2195577621459961, + 0.013092519715428352 + ], + "credit_map": [ + 0.00496859522536397, + 0.003740760963410139, + 0.010777023620903492, + 0.00904299970716238, + 0.0065192561596632, + 0.008618376217782497, + 0.007710065692663193, + 0.00598882045596838, + 0.0029542602133005857, + 0.002080409089103341, + 0.008097412064671516, + 0.007298417389392853, + 0.003548641689121723, + 0.003794238204136491, + 0.003533555893227458, + 0.002744769211858511, + 0.0012825398007407784, + 0.0009414579253643751, + 0.0025964314118027687, + 0.002792476210743189, + 0.0014588942285627127, + 0.0019254974322393537, + 0.0019394662231206894, + 0.0015725318808108568, + 0.002390597714111209, + 0.0014573009684681892, + 0.005643831100314856, + 0.002927041845396161, + 0.0025554022286087275, + 0.002768927486613393, + 0.0026798017788678408, + 0.0021763222757726908, + 0.0102830296382308, + 0.0073635573498904705, + 0.037311069667339325, + 0.0308262649923563, + 0.01676187664270401, + 0.004864617250859737, + 0.015033007599413395, + 0.008531108498573303, + 0.008979519829154015, + 0.011799757368862629, + 0.015343740582466125, + 0.015089282765984535, + 0.012936810962855816, + 0.005571662448346615, + 0.026367787271738052, + 0.008024625480175018, + 0.015783796086907387, + 0.013614199124276638, + 0.009454098530113697, + 0.00967582780867815, + 0.009815300814807415, + 0.008145013824105263, + 0.016653764992952347, + 0.007331421133130789, + 0.03211858868598938, + 0.05579674243927002, + 0.03180306777358055, + 0.0378250889480114, + 0.04007120057940483, + 0.04297414794564247, + 0.04927830770611763, + 0.023395288735628128, + 0.03206370770931244, + 0.03900172561407089, + 0.0239776149392128, + 0.029372088611125946, + 0.03757190331816673, + 0.01350479293614626, + 0.02902769111096859, + 0.015038246288895607, + 0.0020866342820227146, + 0.0016899844631552696, + 0.0011024895356968045, + 0.0024436451494693756, + 0.0020496456418186426, + 0.0012132171541452408, + 0.0017317152814939618, + 0.0007751878583803773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.006038990803062916 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009344025747850537 + ], + "opposite_logprob_tokens": [ + -7.000934600830078 + ], + "logprob_reward": -0.0009344025747850537, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009344025747850537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06989038735628128, + 0.04429992660880089, + 0.04884343221783638, + 0.042760882526636124, + 0.08009101450443268, + 0.14533694088459015, + 0.07438159734010696, + 0.27099934220314026, + 0.21030010282993317, + 0.013096305541694164 + ], + "credit_map": [ + 0.00578988017514348, + 0.005846736487001181, + 0.013814674690365791, + 0.012705209665000439, + 0.007326554041355848, + 0.010439719073474407, + 0.007936937734484673, + 0.006030682008713484, + 0.004657502751797438, + 0.0026142215356230736, + 0.008496339432895184, + 0.00439654104411602, + 0.0056762779131531715, + 0.008030877448618412, + 0.005504535045474768, + 0.004923629574477673, + 0.004494484513998032, + 0.0022770033683627844, + 0.01372884213924408, + 0.005034065339714289, + 0.005790544208139181, + 0.007871626876294613, + 0.005290238652378321, + 0.004356624558568001, + 0.004213007632642984, + 0.004328092560172081, + 0.005506766028702259, + 0.006385982967913151, + 0.0062361350283026695, + 0.0060763913206756115, + 0.005988189484924078, + 0.004026317968964577, + 0.006299539003521204, + 0.004320100881159306, + 0.021474437788128853, + 0.026974312961101532, + 0.004109078552573919, + 0.004094513598829508, + 0.007308907341212034, + 0.005510119721293449, + 0.01114355307072401, + 0.016149966046214104, + 0.021014653146266937, + 0.02069717086851597, + 0.017467785626649857, + 0.007630430161952972, + 0.0394635833799839, + 0.011769798584282398, + 0.011745407246053219, + 0.010783934965729713, + 0.007233545184135437, + 0.008216451853513718, + 0.00860180240124464, + 0.007058178540319204, + 0.014669214375317097, + 0.0060730609111487865, + 0.027797551825642586, + 0.04702466353774071, + 0.02736910991370678, + 0.03330385684967041, + 0.033840570598840714, + 0.035796869546175, + 0.045746300369501114, + 0.02012043073773384, + 0.03163821995258331, + 0.036914777010679245, + 0.02365567535161972, + 0.027562785893678665, + 0.03672311455011368, + 0.012924755923449993, + 0.026784585788846016, + 0.014096199534833431, + 0.0015887191984802485, + 0.0021851772908121347, + 0.001097296248190105, + 0.00218402617610991, + 0.0018061937298625708, + 0.0014502633130177855, + 0.001924534561112523, + 0.0008600943838246167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.0009344025747850537 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00249610492028296 + ], + "opposite_logprob_tokens": [ + -6.002496242523193 + ], + "logprob_reward": -0.00249610492028296, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00249610492028296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08289099484682083, + 0.02514874003827572, + 0.014713955111801624, + 0.014313275925815105, + 0.1343379020690918, + 0.19091299176216125, + 0.07344097644090652, + 0.2618596851825714, + 0.1918511837720871, + 0.010530287399888039 + ], + "credit_map": [ + 0.006810646038502455, + 0.00561919342726469, + 0.017878254875540733, + 0.015120220370590687, + 0.008596059866249561, + 0.012140870094299316, + 0.009758155792951584, + 0.006967594381421804, + 0.0022233554627746344, + 0.0016259470721706748, + 0.0060900067910552025, + 0.005202965345233679, + 0.0024597139563411474, + 0.002820841735228896, + 0.002702288329601288, + 0.0020236228592693806, + 0.0013529673451557755, + 0.0006988208624534309, + 0.0041213748045265675, + 0.0018109966767951846, + 0.0013454003492370248, + 0.0020990886259824038, + 0.0016955765895545483, + 0.0015897295670583844, + 0.0015068957582116127, + 0.0013525862013921142, + 0.0019028664100915194, + 0.0015639690682291985, + 0.0017056893557310104, + 0.002695866860449314, + 0.002211083425208926, + 0.0013743184972554445, + 0.007167911622673273, + 0.006442480254918337, + 0.013519635424017906, + 0.03894314169883728, + 0.014477133750915527, + 0.005222770851105452, + 0.038564879447221756, + 0.009999953210353851, + 0.014970684424042702, + 0.020725172013044357, + 0.02708103321492672, + 0.02733035199344158, + 0.023929709568619728, + 0.010293974541127682, + 0.05097652226686478, + 0.015605539083480835, + 0.009309383109211922, + 0.009239678271114826, + 0.0070522925816476345, + 0.009425553493201733, + 0.008054579608142376, + 0.00706107122823596, + 0.017146917060017586, + 0.006151502951979637, + 0.026940930634737015, + 0.04526109993457794, + 0.027110600844025612, + 0.03153159096837044, + 0.032976169139146805, + 0.03665127605199814, + 0.04114754870533943, + 0.02024049125611782, + 0.0290616974234581, + 0.03452077880501747, + 0.021276671439409256, + 0.0244555976241827, + 0.03406162187457085, + 0.011758641339838505, + 0.023962784558534622, + 0.012753395363688469, + 0.0013754491228610277, + 0.0020449389703571796, + 0.0009738372755236924, + 0.0015108617953956127, + 0.0014436687342822552, + 0.0011015578638762236, + 0.0014040812384337187, + 0.0006758922827430069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.00249610492028296 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065670177340508 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.016065670177340508, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065670177340508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08157569915056229, + 0.02138230763375759, + 0.015174053609371185, + 0.026944857090711594, + 0.17076300084590912, + 0.11063764244318008, + 0.12539860606193542, + 0.20711971819400787, + 0.20721060037612915, + 0.033793505281209946 + ], + "credit_map": [ + 0.007271905895322561, + 0.005452023819088936, + 0.018052348867058754, + 0.012827243655920029, + 0.00872623361647129, + 0.013233593665063381, + 0.00889651384204626, + 0.0071158381178975105, + 0.0024795602075755596, + 0.0013366120401769876, + 0.004544470459222794, + 0.0034315744414925575, + 0.0018728652503341436, + 0.003087934572249651, + 0.0023680783342570066, + 0.0022612123284488916, + 0.0014300133334472775, + 0.0009451782098039985, + 0.004586696624755859, + 0.0017358249751850963, + 0.0016113134333863854, + 0.0017998985713347793, + 0.001622554613277316, + 0.0014425738481804729, + 0.0029443223029375076, + 0.002986564766615629, + 0.0036053939256817102, + 0.0030143801122903824, + 0.003172624623402953, + 0.005338899325579405, + 0.003332901978865266, + 0.0025497693568468094, + 0.009797554463148117, + 0.008031755685806274, + 0.036777399480342865, + 0.058023568242788315, + 0.00738061498850584, + 0.008435404859483242, + 0.03256465122103691, + 0.00975203700363636, + 0.010137923993170261, + 0.009457547217607498, + 0.01703559421002865, + 0.012215415015816689, + 0.01595219410955906, + 0.007748162839561701, + 0.027838056907057762, + 0.010252746753394604, + 0.015414601191878319, + 0.01713055558502674, + 0.00934402272105217, + 0.02107452042400837, + 0.017492184415459633, + 0.012015082873404026, + 0.024296533316373825, + 0.008631102740764618, + 0.021159490570425987, + 0.04290930554270744, + 0.016360148787498474, + 0.018625957891345024, + 0.030674993991851807, + 0.029071956872940063, + 0.030431319028139114, + 0.017886532470583916, + 0.032665006816387177, + 0.03673352673649788, + 0.02435469813644886, + 0.024461867287755013, + 0.038262706249952316, + 0.013199005275964737, + 0.024225473403930664, + 0.013308306224644184, + 0.0037958889733999968, + 0.007785984314978123, + 0.00297940568998456, + 0.004480171948671341, + 0.005342325195670128, + 0.003209230490028858, + 0.0039015929214656353, + 0.0022989052813500166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.016065670177340508 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607118360698223 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.01607118360698223, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607118360698223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10771143436431885, + 0.03804885223507881, + 0.03296534717082977, + 0.035881977528333664, + 0.32837897539138794, + 0.06826003640890121, + 0.03547520935535431, + 0.10868985205888748, + 0.08408541977405548, + 0.16050295531749725 + ], + "credit_map": [ + 0.013549626804888248, + 0.00769279059022665, + 0.017241785302758217, + 0.01741868257522583, + 0.009261995553970337, + 0.01599327102303505, + 0.011679669842123985, + 0.014873616397380829, + 0.003562159836292267, + 0.0018380810506641865, + 0.009217258542776108, + 0.006999081000685692, + 0.0033781591337174177, + 0.005284457001835108, + 0.004447212442755699, + 0.0033224420621991158, + 0.002907329937443137, + 0.0015573195414617658, + 0.008151227608323097, + 0.005816695746034384, + 0.0032184463925659657, + 0.004822578281164169, + 0.003775943536311388, + 0.0027158085722476244, + 0.004073684569448233, + 0.003301336895674467, + 0.005098496098071337, + 0.003376829670742154, + 0.0042138537392020226, + 0.008045792579650879, + 0.004589580465108156, + 0.003182405373081565, + 0.01888684369623661, + 0.013021925464272499, + 0.013695154339075089, + 0.0923662930727005, + 0.017960740253329277, + 0.012588747777044773, + 0.1318254917860031, + 0.0280337855219841, + 0.005816848482936621, + 0.007605698890984058, + 0.008868766948580742, + 0.007121937815099955, + 0.015640340745449066, + 0.00785723514854908, + 0.01055887434631586, + 0.004790338687598705, + 0.005430549383163452, + 0.005029418971389532, + 0.004908113740384579, + 0.0046643358655273914, + 0.004496978595852852, + 0.003916600253432989, + 0.0036092489026486874, + 0.0034199622459709644, + 0.012539010494947433, + 0.021644094958901405, + 0.008940263651311398, + 0.012174168601632118, + 0.015972821041941643, + 0.009547531604766846, + 0.02115955576300621, + 0.006712405942380428, + 0.010978865437209606, + 0.013308479450643063, + 0.009626239538192749, + 0.012970817275345325, + 0.013391848653554916, + 0.004729588516056538, + 0.013486681506037712, + 0.005592897534370422, + 0.013399731367826462, + 0.03590742126107216, + 0.013136127032339573, + 0.021949799731373787, + 0.026471218094229698, + 0.010810252279043198, + 0.027774596586823463, + 0.011053798720240593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.01607118360698223 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "perfectly static, cinematic long shot depicts the breathtaking expanse of a remote, rugged mountainside bathed in the warm, low-angle light of golden hour. A massive pine tree stands at the majestic center of the scene, its ancient, towering form and deep green needles silhouetted against the pale, clear sky. Suddenly, a sharp explosion shatters the profound silence, echoing throughout the valley. The pine's massive trunk shudders violently, and near the base, the dark, gnarled bark bursts open, revealing a chaotic explosion of pale, raw wood fibers splintering outward. The tree begins to tilt, its movement beginning with a slow, deliberate groan before accelerating with a terrifying, decisive momentum, carving a majestic, tragic arc through the crisp mountain air. As it falls, its massive shadow sweeps swiftly across the rocky terrain below, its branches streaking through the air with a final, violent hiss before impacting the ground with a deafening, earth-shattering crash, sending a huge whirlwind of dust, soil, and pine needles into golden light, creating an atmosphere of raw, natural power and a stunning finale.", + "seed": 27275, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_15/index_4396_seed_27275_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.009411299391649664, + "vlm_logprob": -0.009411299391649664, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.009411299391649664, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016325945034623146 + ], + "opposite_logprob_tokens": [ + -5.266325950622559 + ], + "logprob_reward": -0.016325945034623146, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.016325945034623146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10892914235591888, + 0.05470830947160721, + 0.04059567302465439, + 0.039268311113119125, + 0.04328251630067825, + 0.03845430538058281, + 0.13403217494487762, + 0.09954477846622467, + 0.14162728190422058, + 0.2995575964450836 + ], + "credit_map": [ + 0.015154310502111912, + 0.009007702581584454, + 0.02378116361796856, + 0.00719448970630765, + 0.01548532210290432, + 0.015799516811966896, + 0.010730397887527943, + 0.01177622564136982, + 0.007619068957865238, + 0.004495883826166391, + 0.014619356952607632, + 0.003679780289530754, + 0.0075789871625602245, + 0.007455663755536079, + 0.004640715196728706, + 0.004618851933628321, + 0.0050002154894173145, + 0.004835989326238632, + 0.009068439714610577, + 0.0030745412223041058, + 0.006413010414689779, + 0.005424470640718937, + 0.003700678702443838, + 0.003078324254602194, + 0.004812389612197876, + 0.004168229177594185, + 0.00677609583362937, + 0.0027100781444460154, + 0.0071641854010522366, + 0.005291654262691736, + 0.004313454497605562, + 0.0040322160348296165, + 0.00612172344699502, + 0.0031291511841118336, + 0.011641590856015682, + 0.0027199452742934227, + 0.005596798378974199, + 0.005395041778683662, + 0.0043727499432861805, + 0.004305513110011816, + 0.003732784651219845, + 0.004196117632091045, + 0.007710094563663006, + 0.002589714713394642, + 0.005623865406960249, + 0.005082097370177507, + 0.004011497367173433, + 0.005508129950612783, + 0.0117601677775383, + 0.0091137969866395, + 0.018985554575920105, + 0.008685105480253696, + 0.018558327108621597, + 0.020368406549096107, + 0.02618490532040596, + 0.02037590928375721, + 0.013041489757597446, + 0.013411353342235088, + 0.020188260823488235, + 0.010791441425681114, + 0.007108921650797129, + 0.008384967222809792, + 0.014040328562259674, + 0.012578005902469158, + 0.01946120522916317, + 0.03473829850554466, + 0.0209097508341074, + 0.014583416283130646, + 0.010975546203553677, + 0.0105031318962574, + 0.01686365343630314, + 0.013592279516160488, + 0.027700500562787056, + 0.03504858538508415, + 0.04128959774971008, + 0.024086838588118553, + 0.045637890696525574, + 0.04669591784477234, + 0.04342401772737503, + 0.03567421808838844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.016325945034623146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013533014571294188 + ], + "opposite_logprob_tokens": [ + -6.6263532638549805 + ], + "logprob_reward": -0.0013533014571294188, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013533014571294188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17696751654148102, + 0.08196937292814255, + 0.06242574006319046, + 0.06654182076454163, + 0.04694078117609024, + 0.02810417115688324, + 0.09123434871435165, + 0.20601746439933777, + 0.0918215811252594, + 0.14797717332839966 + ], + "credit_map": [ + 0.027117015793919563, + 0.014580203220248222, + 0.048186637461185455, + 0.011219270527362823, + 0.024543562904000282, + 0.022087031975388527, + 0.013963302597403526, + 0.015270496718585491, + 0.010717757977545261, + 0.007701955270022154, + 0.01579621061682701, + 0.004561424721032381, + 0.011724238283932209, + 0.015002785250544548, + 0.007464189548045397, + 0.009000807069242, + 0.007148015312850475, + 0.006679846439510584, + 0.010653520934283733, + 0.00452787522226572, + 0.009198683314025402, + 0.010806137695908546, + 0.0063150981441140175, + 0.007096567191183567, + 0.00794933084398508, + 0.005630269646644592, + 0.01034286618232727, + 0.0031273444183170795, + 0.013162972405552864, + 0.009873926639556885, + 0.00903311837464571, + 0.007421995513141155, + 0.006011382210999727, + 0.003310931846499443, + 0.01073352713137865, + 0.0029821344651281834, + 0.007161600049585104, + 0.00733641954138875, + 0.005243885796517134, + 0.004160899668931961, + 0.0038753338158130646, + 0.0025372833479195833, + 0.006520509719848633, + 0.002078836318105459, + 0.0031090902630239725, + 0.004494973924010992, + 0.002398665761575103, + 0.003089477075263858, + 0.008270167745649815, + 0.00636171642690897, + 0.014606637880206108, + 0.005392410792410374, + 0.013692328706383705, + 0.011010316200554371, + 0.018972322344779968, + 0.012928447686135769, + 0.029703961685299873, + 0.016813937574625015, + 0.04858625307679176, + 0.01864592917263508, + 0.01650124415755272, + 0.019199451431632042, + 0.028980853036046028, + 0.027585845440626144, + 0.009634566493332386, + 0.0116750942543149, + 0.015318556688725948, + 0.010663114488124847, + 0.008264533244073391, + 0.009595795534551144, + 0.012761966325342655, + 0.013907957822084427, + 0.012816266156733036, + 0.013697011396288872, + 0.020336050540208817, + 0.013937185518443584, + 0.017768949270248413, + 0.0272086001932621, + 0.020533116534352303, + 0.021679997444152832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.0013533014571294188 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101912025362253 + ], + "opposite_logprob_tokens": [ + -5.504101753234863 + ], + "logprob_reward": -0.004101912025362253, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101912025362253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08287172019481659, + 0.03766608610749245, + 0.02185092866420746, + 0.018635980784893036, + 0.04818987473845482, + 0.022402619943022728, + 0.10388349741697311, + 0.3586212992668152, + 0.13708022236824036, + 0.16879777610301971 + ], + "credit_map": [ + 0.012501631863415241, + 0.005801312625408173, + 0.018759947270154953, + 0.004581780172884464, + 0.010670065879821777, + 0.015723705291748047, + 0.006605730392038822, + 0.008227546699345112, + 0.005787632428109646, + 0.0035157823003828526, + 0.008212198503315449, + 0.0024088455829769373, + 0.004979867022484541, + 0.006426981184631586, + 0.002532470738515258, + 0.0038023118395358324, + 0.003697472857311368, + 0.0020092721097171307, + 0.005568987689912319, + 0.001531135756522417, + 0.002619851380586624, + 0.0026217056438326836, + 0.001576621551066637, + 0.002225881442427635, + 0.003169361501932144, + 0.0015522462781518698, + 0.005695899482816458, + 0.0010701745050027966, + 0.0017293004784733057, + 0.002564982743933797, + 0.0012237448245286942, + 0.0016302707372233272, + 0.007935237139463425, + 0.003035180736333132, + 0.014825101010501385, + 0.0030859666876494884, + 0.005937602836638689, + 0.004890204407274723, + 0.004319809377193451, + 0.004160772077739239, + 0.0034243471454828978, + 0.0019207363948225975, + 0.005683741997927427, + 0.0016054892912507057, + 0.002716250019147992, + 0.002903077518567443, + 0.001928450888954103, + 0.0022205254063010216, + 0.008313961327075958, + 0.005687923636287451, + 0.009864035993814468, + 0.0060719395987689495, + 0.016285724937915802, + 0.011658014729619026, + 0.029444647952914238, + 0.016557253897190094, + 0.04577257111668587, + 0.04682023078203201, + 0.050459668040275574, + 0.03560618683695793, + 0.027758648619055748, + 0.029796747490763664, + 0.06517381221055984, + 0.05723346769809723, + 0.015616613440215588, + 0.0259444210678339, + 0.013829006813466549, + 0.01691136136651039, + 0.014629381708800793, + 0.019324995577335358, + 0.017520416527986526, + 0.013304032385349274, + 0.0167403481900692, + 0.021969519555568695, + 0.020356900990009308, + 0.0255708210170269, + 0.01839139312505722, + 0.025446636602282524, + 0.02334219217300415, + 0.016979966312646866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.004101912025362253 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197767712175846 + ], + "opposite_logprob_tokens": [ + -4.264197826385498 + ], + "logprob_reward": -0.014197767712175846, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197767712175846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06409446895122528, + 0.019729655236005783, + 0.015388545580208302, + 0.027011888101696968, + 0.008926727809011936, + 0.020058806985616684, + 0.04424750432372093, + 0.15825746953487396, + 0.277168333530426, + 0.36511653661727905 + ], + "credit_map": [ + 0.01085271779447794, + 0.0036345256958156824, + 0.01797770895063877, + 0.002736026654019952, + 0.008637195453047752, + 0.008932718075811863, + 0.005659826565533876, + 0.005663754418492317, + 0.0031971342395991087, + 0.001314223394729197, + 0.005922976415604353, + 0.0010717931436374784, + 0.0025623266119509935, + 0.0020640764851123095, + 0.0018113633850589395, + 0.00178576132748276, + 0.001945195603184402, + 0.0014684819616377354, + 0.0033866274170577526, + 0.0009599667391739786, + 0.0022269573528319597, + 0.00224233022890985, + 0.0017833311576396227, + 0.001375654828734696, + 0.004227282013744116, + 0.0026248733047395945, + 0.006792542990297079, + 0.0020310520194470882, + 0.0030662042554467916, + 0.0028483981732279062, + 0.0024360872339457273, + 0.0029854478780180216, + 0.001186760258860886, + 0.0006212075240910053, + 0.0020585423335433006, + 0.0005514015210792422, + 0.0011739376932382584, + 0.0013241928536444902, + 0.0008997979457490146, + 0.0011108876205980778, + 0.0030266279354691505, + 0.0023308328818529844, + 0.0052153742872178555, + 0.0015115611022338271, + 0.0021662891376763582, + 0.0018231365829706192, + 0.0018346556462347507, + 0.0021503304596990347, + 0.003949997015297413, + 0.0049803778529167175, + 0.005729671102017164, + 0.0025895098224282265, + 0.005718988366425037, + 0.004502951167523861, + 0.01069714967161417, + 0.0060788607224822044, + 0.02385328710079193, + 0.021838614717125893, + 0.03894593194127083, + 0.017053402960300446, + 0.0104455491527915, + 0.01439081784337759, + 0.017185527831315994, + 0.014544342644512653, + 0.03016837127506733, + 0.025919022038578987, + 0.04662526026368141, + 0.04142218083143234, + 0.021559081971645355, + 0.034994106739759445, + 0.034413691610097885, + 0.042066626250743866, + 0.029217630624771118, + 0.02821621671319008, + 0.06365843117237091, + 0.03913361206650734, + 0.044370464980602264, + 0.04777059704065323, + 0.0625719502568245, + 0.05017763003706932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.014197767712175846 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077570728957653 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.011077570728957653, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077570728957653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1122056245803833, + 0.024619562551379204, + 0.01657463051378727, + 0.03262447193264961, + 0.024219242855906487, + 0.021951498463749886, + 0.04157768562436104, + 0.36539164185523987, + 0.1458875685930252, + 0.2149481177330017 + ], + "credit_map": [ + 0.020104287192225456, + 0.006529670208692551, + 0.039429064840078354, + 0.005914509296417236, + 0.013471969403326511, + 0.010585741139948368, + 0.006948530208319426, + 0.009221848100423813, + 0.003682629903778434, + 0.0017395906616002321, + 0.006892024073749781, + 0.001358690089546144, + 0.0035431855358183384, + 0.002737745177000761, + 0.0022921699564903975, + 0.0023735270369797945, + 0.00252083339728415, + 0.001437730505131185, + 0.004606931935995817, + 0.0009569533867761493, + 0.0022423185873776674, + 0.001973602920770645, + 0.001409027143381536, + 0.0014272327534854412, + 0.005141411442309618, + 0.004392476286739111, + 0.008768388070166111, + 0.0024554904084652662, + 0.0037660992238670588, + 0.003123221220448613, + 0.0025508610997349024, + 0.0024265230167657137, + 0.0038456348702311516, + 0.0013670393964275718, + 0.007963581010699272, + 0.0013848968083038926, + 0.002442411845549941, + 0.002801472321152687, + 0.0017932892078533769, + 0.0026209158822894096, + 0.003605636302381754, + 0.0015967466169968247, + 0.007290821056813002, + 0.00121814024168998, + 0.002360077342018485, + 0.0024574855342507362, + 0.0012890930520370603, + 0.0021334972698241472, + 0.005073810927569866, + 0.0020685915369540453, + 0.00864063948392868, + 0.0028885724022984505, + 0.004230407997965813, + 0.004208608530461788, + 0.008558395318686962, + 0.005908660590648651, + 0.0478927418589592, + 0.046658508479595184, + 0.0590251199901104, + 0.03703415393829346, + 0.028967345133423805, + 0.02868051268160343, + 0.06315381079912186, + 0.05397946387529373, + 0.01638137176632881, + 0.015754450112581253, + 0.02586200274527073, + 0.022741863504052162, + 0.01281483843922615, + 0.015547000803053379, + 0.01774575002491474, + 0.01904028281569481, + 0.023656733334064484, + 0.029604218900203705, + 0.02664138562977314, + 0.021910492330789566, + 0.019665447995066643, + 0.040441107004880905, + 0.02705700322985649, + 0.025971729308366776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.011077570728957653 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "perfectly static, cinematic long shot depicts the breathtaking expanse of a remote, rugged mountainside bathed in the warm, low-angle light of golden hour. A massive pine tree stands at the majestic center of the scene, its ancient, towering form and deep green needles silhouetted against the pale, clear sky. Suddenly, a sharp explosion shatters the profound silence, echoing throughout the valley. The pine's massive trunk shudders violently, and near the base, the dark, gnarled bark bursts open, revealing a chaotic explosion of pale, raw wood fibers splintering outward. The tree begins to tilt, its movement beginning with a slow, deliberate groan before accelerating with a terrifying, decisive momentum, carving a majestic, tragic arc through the crisp mountain air. As it falls, its massive shadow sweeps swiftly across the rocky terrain below, its branches streaking through the air with a final, violent hiss before impacting the ground with a deafening, earth-shattering crash, sending a huge whirlwind of dust, soil, and pine needles into golden light, creating an atmosphere of raw, natural power and a stunning finale.", + "seed": 27276, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_15/index_4396_seed_27276_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010098785907030106, + "vlm_logprob": -0.010098785907030106, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.010098785907030106, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022292615845799446 + ], + "opposite_logprob_tokens": [ + -4.647292613983154 + ], + "logprob_reward": -0.022292615845799446, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.022292615845799446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10886093974113464, + 0.05452657490968704, + 0.03475602716207504, + 0.034948188811540604, + 0.04157252237200737, + 0.2689017057418823, + 0.09495840221643448, + 0.1257658749818802, + 0.08924319595098495, + 0.14646662771701813 + ], + "credit_map": [ + 0.011127626523375511, + 0.012652774341404438, + 0.011865322478115559, + 0.012046342715620995, + 0.015017115511000156, + 0.008354081772267818, + 0.012397954240441322, + 0.0253997091203928, + 0.00502274464815855, + 0.00506007531657815, + 0.005247351247817278, + 0.010024582967162132, + 0.011005548760294914, + 0.005102192051708698, + 0.005435950588434935, + 0.007628125138580799, + 0.0039048667531460524, + 0.003263324499130249, + 0.003958742134273052, + 0.003791502211242914, + 0.0041104452684521675, + 0.0037480148021131754, + 0.0055544706992805, + 0.006424658000469208, + 0.0034932629205286503, + 0.002846709219738841, + 0.004253252875059843, + 0.005440075416117907, + 0.004400512203574181, + 0.002181875752285123, + 0.0071001495234668255, + 0.005232346709817648, + 0.0035433094017207623, + 0.0037970556877553463, + 0.003444112604483962, + 0.005091422703117132, + 0.007258636876940727, + 0.0032453478779643774, + 0.007793189957737923, + 0.0073994421400129795, + 0.02847743220627308, + 0.04342223331332207, + 0.023598486557602882, + 0.05070777237415314, + 0.03565915301442146, + 0.020905712619423866, + 0.03495784476399422, + 0.031173057854175568, + 0.008531731553375721, + 0.009683937765657902, + 0.004458016715943813, + 0.014027257449924946, + 0.03135184198617935, + 0.008260742761194706, + 0.008056181482970715, + 0.010588682256639004, + 0.010518050752580166, + 0.00914629828184843, + 0.011200341396033764, + 0.017463315278291702, + 0.03527022525668144, + 0.021276185289025307, + 0.00924699380993843, + 0.011644459329545498, + 0.010788707062602043, + 0.0182682853192091, + 0.006348840892314911, + 0.010926619172096252, + 0.014411616139113903, + 0.01115539949387312, + 0.01036796160042286, + 0.006975760217756033, + 0.01307437103241682, + 0.02168741263449192, + 0.014482187107205391, + 0.014523741789162159, + 0.025132423266768456, + 0.016674829646945, + 0.021795231848955154, + 0.019096417352557182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.022292615845799446 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005965286400169134 + ], + "opposite_logprob_tokens": [ + -5.130965232849121 + ], + "logprob_reward": -0.005965286400169134, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005965286400169134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1015106812119484, + 0.07281063497066498, + 0.05614873394370079, + 0.05658472701907158, + 0.13987137377262115, + 0.13731415569782257, + 0.06848863512277603, + 0.058062709867954254, + 0.2147107720375061, + 0.09449754655361176 + ], + "credit_map": [ + 0.008582832291722298, + 0.009482884779572487, + 0.008503347635269165, + 0.011364366859197617, + 0.025644712150096893, + 0.00815275963395834, + 0.015529846772551537, + 0.014249932020902634, + 0.006959123071283102, + 0.006202880758792162, + 0.007703412789851427, + 0.015862509608268738, + 0.009869357571005821, + 0.010019839741289616, + 0.008004715666174889, + 0.00818879995495081, + 0.006010867655277252, + 0.003735276870429516, + 0.005818950477987528, + 0.0057373070158064365, + 0.007670394144952297, + 0.005341026932001114, + 0.009898639284074306, + 0.011936272494494915, + 0.00563600892201066, + 0.004539058543741703, + 0.004617113154381514, + 0.005118154920637608, + 0.010004814714193344, + 0.004795561078935862, + 0.012218034826219082, + 0.00965598039329052, + 0.010756022296845913, + 0.012719999998807907, + 0.011983584612607956, + 0.018147168681025505, + 0.025237252935767174, + 0.009140511974692345, + 0.024457097053527832, + 0.027429742738604546, + 0.014358878135681152, + 0.024832069873809814, + 0.01169137004762888, + 0.026344401761889458, + 0.017513781785964966, + 0.011098027229309082, + 0.016106577590107918, + 0.01536904089152813, + 0.004318913444876671, + 0.007921166718006134, + 0.003146502422168851, + 0.0197893176227808, + 0.014379994943737984, + 0.003825838677585125, + 0.008823694661259651, + 0.006283204536885023, + 0.006934166885912418, + 0.007830275222659111, + 0.006265270058065653, + 0.011562135070562363, + 0.006418734323233366, + 0.005652829073369503, + 0.005898119416087866, + 0.007501175161451101, + 0.05503762140870094, + 0.045875802636146545, + 0.014431619085371494, + 0.01689772494137287, + 0.028607748448848724, + 0.013961069285869598, + 0.026698162779211998, + 0.013201027177274227, + 0.009291266091167927, + 0.013455291278660297, + 0.010148158296942711, + 0.015690729022026062, + 0.01310148648917675, + 0.008792602457106113, + 0.014740783721208572, + 0.009277227334678173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.005965286400169134 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008650331757962704 + ], + "opposite_logprob_tokens": [ + -4.758650302886963 + ], + "logprob_reward": -0.008650331757962704, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008650331757962704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08732456713914871, + 0.048042066395282745, + 0.05372537672519684, + 0.025369174778461456, + 0.125259667634964, + 0.1169290840625763, + 0.04050211235880852, + 0.07945956289768219, + 0.30706751346588135, + 0.11632085591554642 + ], + "credit_map": [ + 0.009111429564654827, + 0.008587840013206005, + 0.010242490097880363, + 0.012674713507294655, + 0.009672255255281925, + 0.007989298552274704, + 0.009976180270314217, + 0.01907035894691944, + 0.004799327347427607, + 0.003087013727054, + 0.005295783746987581, + 0.01094175036996603, + 0.005132915452122688, + 0.005763092078268528, + 0.006398489233106375, + 0.006623696535825729, + 0.005484535358846188, + 0.0037138250190764666, + 0.004713607020676136, + 0.010988903231918812, + 0.00464623561128974, + 0.005320793483406305, + 0.010818242095410824, + 0.008039235137403011, + 0.0027252777945250273, + 0.0023515974171459675, + 0.002746900077909231, + 0.0026279499288648367, + 0.003200170351192355, + 0.0024462498258799314, + 0.004905048757791519, + 0.004365979693830013, + 0.011516238562762737, + 0.012471198104321957, + 0.012301135808229446, + 0.01904536411166191, + 0.011692997999489307, + 0.008803119882941246, + 0.02251249924302101, + 0.0269171129912138, + 0.012437978759407997, + 0.018483014777302742, + 0.00975649245083332, + 0.025452857837080956, + 0.013637049123644829, + 0.010540302842855453, + 0.013210246339440346, + 0.013411141000688076, + 0.00472413282841444, + 0.0055244979448616505, + 0.0027521559968590736, + 0.010754236951470375, + 0.0030105947516858578, + 0.003324017394334078, + 0.006370123941451311, + 0.004042354878038168, + 0.009849066846072674, + 0.011281963437795639, + 0.00653607863932848, + 0.013615326024591923, + 0.014885161072015762, + 0.008086424320936203, + 0.006842336617410183, + 0.008363207802176476, + 0.07794729620218277, + 0.060528144240379333, + 0.023332219570875168, + 0.022619936615228653, + 0.039687253534793854, + 0.01947033777832985, + 0.04202845320105553, + 0.021453866735100746, + 0.012064334936439991, + 0.0210203155875206, + 0.011547790840268135, + 0.01761864498257637, + 0.015357425436377525, + 0.00865519791841507, + 0.020439665764570236, + 0.009617484174668789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.008650331757962704 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951777566224337 + ], + "opposite_logprob_tokens": [ + -5.130951881408691 + ], + "logprob_reward": -0.005951777566224337, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951777566224337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07752082496881485, + 0.022436216473579407, + 0.03854753077030182, + 0.021757617592811584, + 0.05888180807232857, + 0.026370897889137268, + 0.12368118017911911, + 0.13827772438526154, + 0.13451556861400604, + 0.35801056027412415 + ], + "credit_map": [ + 0.006516641937196255, + 0.006463642697781324, + 0.007981483824551105, + 0.007456909865140915, + 0.014367385767400265, + 0.005511848255991936, + 0.013486477546393871, + 0.01573643833398819, + 0.0018496086122468114, + 0.0015845289453864098, + 0.0022268351167440414, + 0.0026785265654325485, + 0.004706261213868856, + 0.0016416561556980014, + 0.003631893778219819, + 0.004116908181458712, + 0.003176447469741106, + 0.0030517736449837685, + 0.00343187409453094, + 0.008898084983229637, + 0.005232561845332384, + 0.004129412118345499, + 0.006286596413701773, + 0.0043407827615737915, + 0.0021769891027361155, + 0.0015554247656837106, + 0.0022761609870940447, + 0.0022422687616199255, + 0.0030462935101240873, + 0.0017207758501172066, + 0.005028230603784323, + 0.0037114759907126427, + 0.005685087759047747, + 0.006810123100876808, + 0.005833170842379332, + 0.009487238712608814, + 0.00504856463521719, + 0.003911652136594057, + 0.009260955266654491, + 0.012845021672546864, + 0.002257570158690214, + 0.002317344071343541, + 0.003634579246863723, + 0.004351446405053139, + 0.0030597345903515816, + 0.002425898564979434, + 0.004665450658649206, + 0.0036588753573596478, + 0.007858666591346264, + 0.014905850403010845, + 0.006627040449529886, + 0.035159625113010406, + 0.020464688539505005, + 0.011924090795218945, + 0.014809384942054749, + 0.011931844055652618, + 0.011397396214306355, + 0.010876321233808994, + 0.012212703935801983, + 0.016953986138105392, + 0.03614223375916481, + 0.02382007986307144, + 0.01062182430177927, + 0.016253190115094185, + 0.028319507837295532, + 0.025967678055167198, + 0.011298131197690964, + 0.012777109630405903, + 0.01689455844461918, + 0.010610283352434635, + 0.018102744594216347, + 0.010545571334660053, + 0.02864661067724228, + 0.04514872655272484, + 0.03488951921463013, + 0.06789039075374603, + 0.046646930277347565, + 0.031001748517155647, + 0.06925112009048462, + 0.034535542130470276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.005951777566224337 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633917964994907 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633917964994907, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633917964994907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19098301231861115, + 0.03874436020851135, + 0.033228542655706406, + 0.08256807923316956, + 0.25278255343437195, + 0.06848688423633575, + 0.09440038353204727, + 0.08199792355298996, + 0.05734119564294815, + 0.09946706891059875 + ], + "credit_map": [ + 0.019169483333826065, + 0.01633976586163044, + 0.02043217606842518, + 0.04783913493156433, + 0.01780429668724537, + 0.019553162157535553, + 0.017924772575497627, + 0.03192021697759628, + 0.0030055444221943617, + 0.0029035350307822227, + 0.0032609892077744007, + 0.003013073932379484, + 0.011067640967667103, + 0.0042138476856052876, + 0.00542967626824975, + 0.00585005059838295, + 0.003238004632294178, + 0.002106457483023405, + 0.004343770910054445, + 0.004560770466923714, + 0.006423230282962322, + 0.0035028245765715837, + 0.004781750030815601, + 0.004271732177585363, + 0.009112963452935219, + 0.00789656862616539, + 0.010045853443443775, + 0.011354592628777027, + 0.009662065654993057, + 0.007549742236733437, + 0.01540644746273756, + 0.011539831757545471, + 0.02169233188033104, + 0.022445207461714745, + 0.020891109481453896, + 0.040373850613832474, + 0.024303045123815536, + 0.017760509625077248, + 0.054151181131601334, + 0.05116530880331993, + 0.0076606180518865585, + 0.008868066594004631, + 0.0049154250882565975, + 0.01633407361805439, + 0.007780793122947216, + 0.006233200430870056, + 0.007459043059498072, + 0.00923565961420536, + 0.006307818926870823, + 0.009369435720145702, + 0.004195446614176035, + 0.01991674304008484, + 0.029359864071011543, + 0.006780178286135197, + 0.008991765789687634, + 0.00947912223637104, + 0.007406752556562424, + 0.005313344765454531, + 0.008133930154144764, + 0.009769383817911148, + 0.020658493041992188, + 0.015325340442359447, + 0.006580387707799673, + 0.008810288272798061, + 0.007911579683423042, + 0.009757660329341888, + 0.004602437373250723, + 0.008952456526458263, + 0.0068561239168047905, + 0.005175103899091482, + 0.00981888361275196, + 0.004266946576535702, + 0.009871349669992924, + 0.015829429030418396, + 0.011145668104290962, + 0.014640252105891705, + 0.016868945211172104, + 0.008403890766203403, + 0.013761857524514198, + 0.008945669047534466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.007633917964994907 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "perfectly static, cinematic long shot depicts the breathtaking expanse of a remote, rugged mountainside bathed in the warm, low-angle light of golden hour. A massive pine tree stands at the majestic center of the scene, its ancient, towering form and deep green needles silhouetted against the pale, clear sky. Suddenly, a sharp explosion shatters the profound silence, echoing throughout the valley. The pine's massive trunk shudders violently, and near the base, the dark, gnarled bark bursts open, revealing a chaotic explosion of pale, raw wood fibers splintering outward. The tree begins to tilt, its movement beginning with a slow, deliberate groan before accelerating with a terrifying, decisive momentum, carving a majestic, tragic arc through the crisp mountain air. As it falls, its massive shadow sweeps swiftly across the rocky terrain below, its branches streaking through the air with a final, violent hiss before impacting the ground with a deafening, earth-shattering crash, sending a huge whirlwind of dust, soil, and pine needles into golden light, creating an atmosphere of raw, natural power and a stunning finale.", + "seed": 27277, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_15/index_4396_seed_27277_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.034876062301918864, + "vlm_logprob": -0.034876062301918864, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.034876062301918864, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15173067152500153 + ], + "opposite_logprob_tokens": [ + -2.15173077583313 + ], + "logprob_reward": -0.15173067152500153, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.15173067152500153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22828738391399384, + 0.14063222706317902, + 0.07933329790830612, + 0.026453152298927307, + 0.14885418117046356, + 0.02890419214963913, + 0.04882093146443367, + 0.07094027101993561, + 0.06486672908067703, + 0.1629076600074768 + ], + "credit_map": [ + 0.027077695354819298, + 0.015856992453336716, + 0.05494103953242302, + 0.021250609308481216, + 0.026216689497232437, + 0.030384082347154617, + 0.023951521143317223, + 0.02860872447490692, + 0.013852627947926521, + 0.010902343317866325, + 0.024306733161211014, + 0.01337710116058588, + 0.02413202077150345, + 0.015866046771407127, + 0.023378731682896614, + 0.01481660921126604, + 0.00792589783668518, + 0.005535686854273081, + 0.018589342013001442, + 0.010679608210921288, + 0.010719675570726395, + 0.006872137077152729, + 0.007119436748325825, + 0.011891504749655724, + 0.0022382994648069143, + 0.0023115319199860096, + 0.0029712687246501446, + 0.004240944981575012, + 0.004287514835596085, + 0.002505781129002571, + 0.005003652069717646, + 0.002894155913963914, + 0.01600322127342224, + 0.016589820384979248, + 0.020164795219898224, + 0.031195158138871193, + 0.014285137876868248, + 0.015226498246192932, + 0.017757399007678032, + 0.017632141709327698, + 0.002888441551476717, + 0.0033241640776395798, + 0.004895276855677366, + 0.0033285825047641993, + 0.0020762155763804913, + 0.0024693747982382774, + 0.005562383681535721, + 0.004359752871096134, + 0.004298929590731859, + 0.003578053554520011, + 0.006058051250874996, + 0.01016019657254219, + 0.006017726846039295, + 0.0046713994815945625, + 0.006256863474845886, + 0.007779703475534916, + 0.008938736282289028, + 0.007512704934924841, + 0.005957040470093489, + 0.004697324708104134, + 0.010638394393026829, + 0.009744703769683838, + 0.01544921938329935, + 0.008002142421901226, + 0.005989340599626303, + 0.0054751913994550705, + 0.006665869615972042, + 0.012336785905063152, + 0.010235348716378212, + 0.006662210915237665, + 0.01062355749309063, + 0.006878415588289499, + 0.02963181585073471, + 0.016898449510335922, + 0.01824282482266426, + 0.015647049993276596, + 0.021787771955132484, + 0.020992834120988846, + 0.020914867520332336, + 0.018792031332850456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.15173067152500153 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00466105155646801 + ], + "opposite_logprob_tokens": [ + -5.3796610832214355 + ], + "logprob_reward": -0.00466105155646801, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00466105155646801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22554367780685425, + 0.1613408923149109, + 0.10985060036182404, + 0.16718696057796478, + 0.09487827867269516, + 0.025432799011468887, + 0.049029529094696045, + 0.058407317847013474, + 0.04130396991968155, + 0.06702597439289093 + ], + "credit_map": [ + 0.02446575090289116, + 0.025719422847032547, + 0.035080086439847946, + 0.019124578684568405, + 0.03854777291417122, + 0.024862907826900482, + 0.037218812853097916, + 0.020524369552731514, + 0.022335126996040344, + 0.014862338081002235, + 0.01713375374674797, + 0.016554104164242744, + 0.030564527958631516, + 0.01803632453083992, + 0.02519460767507553, + 0.01666010543704033, + 0.013262798078358173, + 0.00936905574053526, + 0.01758183166384697, + 0.014352350495755672, + 0.01987919770181179, + 0.011954478919506073, + 0.011909350752830505, + 0.011541544459760189, + 0.01360041182488203, + 0.015702761709690094, + 0.028302190825343132, + 0.020723188295960426, + 0.03379340097308159, + 0.016550682485103607, + 0.019199784845113754, + 0.01931455358862877, + 0.010603930801153183, + 0.010615821927785873, + 0.011217085644602776, + 0.01470026932656765, + 0.01358053833246231, + 0.008930950425565243, + 0.010522708296775818, + 0.01470697671175003, + 0.0032505833078175783, + 0.002639916492626071, + 0.004670649766921997, + 0.002783531788736582, + 0.0020955451764166355, + 0.0018295897170901299, + 0.0038310682866722345, + 0.0043319170363247395, + 0.006194978952407837, + 0.004361424595117569, + 0.004972004797309637, + 0.006787299178540707, + 0.006021803244948387, + 0.005481642205268145, + 0.00912631954997778, + 0.006084057502448559, + 0.00410036277025938, + 0.005522564519196749, + 0.0054972534999251366, + 0.0061495439149439335, + 0.008365130051970482, + 0.006571488454937935, + 0.012163229286670685, + 0.010037750005722046, + 0.0033416959922760725, + 0.0031727354507893324, + 0.003702367888763547, + 0.0031130986753851175, + 0.008627631701529026, + 0.005602745339274406, + 0.007859033532440662, + 0.005884666461497545, + 0.007522644940763712, + 0.004652973730117083, + 0.007300097029656172, + 0.0063749197870492935, + 0.01270200777798891, + 0.00906575471162796, + 0.010977661237120628, + 0.008429921232163906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.00466105155646801 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741517689079046 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741517689079046, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741517689079046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11657733470201492, + 0.07687684893608093, + 0.061368148773908615, + 0.0349944606423378, + 0.09153684228658676, + 0.0682077556848526, + 0.0771416425704956, + 0.1989220231771469, + 0.11708789318799973, + 0.1572871059179306 + ], + "credit_map": [ + 0.014734543859958649, + 0.007013333961367607, + 0.0268300361931324, + 0.013771139085292816, + 0.014470933005213737, + 0.00694600073620677, + 0.015466338023543358, + 0.01734500378370285, + 0.007784431800246239, + 0.005569095723330975, + 0.012747125700116158, + 0.011167877353727818, + 0.01234059501439333, + 0.005263142287731171, + 0.01241945568472147, + 0.009585125371813774, + 0.0064037214033305645, + 0.004465256817638874, + 0.010452259331941605, + 0.007856453768908978, + 0.008530586957931519, + 0.005049205385148525, + 0.006972745060920715, + 0.011637923307716846, + 0.005426755174994469, + 0.0025569263380020857, + 0.003721417160704732, + 0.00465503940358758, + 0.005473660305142403, + 0.003529322100803256, + 0.005696807987987995, + 0.003934534732252359, + 0.008907959796488285, + 0.011223609559237957, + 0.013038614764809608, + 0.018233878538012505, + 0.008710592985153198, + 0.008037864230573177, + 0.00799219124019146, + 0.015392124652862549, + 0.006406259257346392, + 0.006875273305922747, + 0.008852965198457241, + 0.008814645931124687, + 0.0061548808589577675, + 0.0052889445796608925, + 0.013227569870650768, + 0.012587212957441807, + 0.005843934137374163, + 0.005847943015396595, + 0.007725318428128958, + 0.007616107352077961, + 0.013487433083355427, + 0.01000235229730606, + 0.011301911436021328, + 0.015316643752157688, + 0.028896979987621307, + 0.021310068666934967, + 0.01783980242908001, + 0.020618710666894913, + 0.026525083929300308, + 0.022175101563334465, + 0.03426230326294899, + 0.027293967083096504, + 0.008356374688446522, + 0.008698998019099236, + 0.011172981932759285, + 0.01233640220016241, + 0.024304470047354698, + 0.015892023220658302, + 0.024518487975001335, + 0.011808155104517937, + 0.016149379312992096, + 0.011683442629873753, + 0.014644027687609196, + 0.017240187153220177, + 0.02761056460440159, + 0.021102849394083023, + 0.02594430185854435, + 0.022912345826625824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.006741517689079046 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629659026861191 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007629659026861191, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629659026861191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10314995795488358, + 0.07479526102542877, + 0.032969892024993896, + 0.03270690143108368, + 0.03525685891509056, + 0.0512731559574604, + 0.18308429419994354, + 0.17703582346439362, + 0.07141696661710739, + 0.23831090331077576 + ], + "credit_map": [ + 0.013462559320032597, + 0.006460968870669603, + 0.015899082645773888, + 0.009531904011964798, + 0.01476529985666275, + 0.012247002683579922, + 0.01939425617456436, + 0.01138888206332922, + 0.00814630463719368, + 0.006243492476642132, + 0.012094443663954735, + 0.008503494784235954, + 0.011821509338915348, + 0.0065001524053514, + 0.011929857544600964, + 0.009556008502840996, + 0.003055066103115678, + 0.0018684578826650977, + 0.006039503496140242, + 0.004839672241359949, + 0.004987555555999279, + 0.0025629675947129726, + 0.003982891328632832, + 0.0056337760761380196, + 0.0031912862323224545, + 0.0038672806695103645, + 0.00489464495331049, + 0.0043499162420630455, + 0.005977018270641565, + 0.0029366693925112486, + 0.004648374859243631, + 0.0028417084831744432, + 0.003609114559367299, + 0.004940475802868605, + 0.0046389964409172535, + 0.006500388495624065, + 0.004374855197966099, + 0.0031689966563135386, + 0.0030554626137018204, + 0.004968570079654455, + 0.00343413557857275, + 0.004720381461083889, + 0.006398289930075407, + 0.008907554671168327, + 0.003815216012299061, + 0.003918107133358717, + 0.009983289986848831, + 0.010096182115375996, + 0.012693989090621471, + 0.010543947108089924, + 0.016327237710356712, + 0.01783282309770584, + 0.030825229361653328, + 0.023236583918333054, + 0.04320512339472771, + 0.028419354930520058, + 0.012965367175638676, + 0.01620166003704071, + 0.0147812245413661, + 0.012639445252716541, + 0.03085211105644703, + 0.023917578160762787, + 0.046269964426755905, + 0.019408464431762695, + 0.005627810023725033, + 0.006868115626275539, + 0.008410255424678326, + 0.007828600704669952, + 0.016295388340950012, + 0.009119845926761627, + 0.010751822963356972, + 0.00651513459160924, + 0.024361461400985718, + 0.016973335295915604, + 0.025069186463952065, + 0.026898110285401344, + 0.04948733001947403, + 0.03381781652569771, + 0.03543068841099739, + 0.0262729711830616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.007629659026861191 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static cinematic long shot of a remote, rugged mountainside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive pine tree visible at the center of the scene with deep green needles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pine tree's trunk burst open to reveal pale, raw wood fibers splintering outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the falling tree create a huge whirlwind of dust, soil, and pine needles upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by the warm, low-angle light of golden hour?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036174117121845484 + ], + "opposite_logprob_tokens": [ + -5.628617286682129 + ], + "logprob_reward": -0.0036174117121845484, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036174117121845484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10620376467704773, + 0.09725040942430496, + 0.08890807628631592, + 0.07203903794288635, + 0.16888703405857086, + 0.04657868295907974, + 0.13920451700687408, + 0.10517824441194534, + 0.09099330008029938, + 0.08475691825151443 + ], + "credit_map": [ + 0.014028264209628105, + 0.006526846904307604, + 0.025314990431070328, + 0.010497516952455044, + 0.011351259425282478, + 0.013352364301681519, + 0.008111290633678436, + 0.017021244391798973, + 0.010128508321940899, + 0.007918632589280605, + 0.01230948232114315, + 0.008477075956761837, + 0.018855735659599304, + 0.012158919125795364, + 0.017502281814813614, + 0.009899784810841084, + 0.008285497315227985, + 0.006094719283282757, + 0.014947975054383278, + 0.014848205260932446, + 0.011104199104011059, + 0.008668028749525547, + 0.010304746218025684, + 0.014654716476798058, + 0.008036602288484573, + 0.004982160869985819, + 0.006168509367853403, + 0.008597754873335361, + 0.010685566812753677, + 0.011921966448426247, + 0.013655625283718109, + 0.007990852929651737, + 0.014729681424796581, + 0.01756127178668976, + 0.027783188968896866, + 0.03415806218981743, + 0.016240088269114494, + 0.015704743564128876, + 0.020573565736413002, + 0.022136438637971878, + 0.004204469732940197, + 0.005019236356019974, + 0.007613312918692827, + 0.006873166188597679, + 0.0037385555915534496, + 0.0034000403247773647, + 0.009047378785908222, + 0.006682526785880327, + 0.00992258358746767, + 0.00944794062525034, + 0.013215756043791771, + 0.015473599545657635, + 0.026162855327129364, + 0.01726566255092621, + 0.025494063273072243, + 0.022222071886062622, + 0.00755260419100523, + 0.005542673170566559, + 0.011332943104207516, + 0.00944909080862999, + 0.013826226815581322, + 0.010737896896898746, + 0.026260098442435265, + 0.02047671377658844, + 0.007152154576033354, + 0.008753573521971703, + 0.010113942436873913, + 0.010879198089241982, + 0.017606370151042938, + 0.011626830324530602, + 0.016961393877863884, + 0.007899845950305462, + 0.008236981928348541, + 0.006908500101417303, + 0.007997213862836361, + 0.009812532924115658, + 0.014893488958477974, + 0.010814275592565536, + 0.01331921387463808, + 0.012774713337421417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.0036174117121845484 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_2/index_tensor([6912, 6912])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_2/index_tensor([6912, 6912])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..75d0036fd1c6ba838f207a393523676091103f35 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_2/index_tensor([6912, 6912])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "subsequent low-angle, dynamic tracking shot captures a lone knight atop a majestic black warhorse, galloping across endless golden grasslands. The knight, clad in weathered steel plate armor, its surface covered in tiny scratches and battle scars, wears a full-face helmet that obscures his face, and a tattered crimson cape billows violently behind him at high speed. His steed, a powerfully muscled beast, its pure black fur glistening in the sunset, strikes the ground with a powerful, rhythmic cadence, each landing sending a swift and powerful kick of dry earth. In the background, a massive, orange-red sunset slowly sinks into the distant horizon, its final rays tinting the world with a warm orange-gold hue, lengthening the lone shadows of knight and horse across the grassland. This golden backlight sharply outlines the hard contours of the knight and the horse, and the dust stirred up by the horse's hooves is illuminated in the light, forming a golden mist flying around them. The whole picture is full of epic grandeur, speed and heroic loneliness.", + "seed": 11777, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_2/index_6912_seed_11777_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11320673625450581, + "vlm_logprob": -0.11320673625450581, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.11320673625450581, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1261448711156845 + ], + "opposite_logprob_tokens": [ + -2.5011448860168457 + ], + "logprob_reward": -0.1261448711156845, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.1261448711156845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05923338979482651, + 0.10408073663711548, + 0.24253243207931519, + 0.11658129096031189, + 0.255577951669693, + 0.057583391666412354, + 0.027172086760401726, + 0.051306888461112976, + 0.051675278693437576, + 0.034256625920534134 + ], + "credit_map": [ + 0.008647638373076916, + 0.006332295946776867, + 0.006504892371594906, + 0.005493353586643934, + 0.008841004222631454, + 0.0074468995444476604, + 0.008349872194230556, + 0.007617425173521042, + 0.010197851806879044, + 0.008244416676461697, + 0.011016285046935081, + 0.010854759253561497, + 0.018237272277474403, + 0.013966443948447704, + 0.018779665231704712, + 0.012784034013748169, + 0.026723036542534828, + 0.023640118539333344, + 0.03807194530963898, + 0.025758353993296623, + 0.0309403445571661, + 0.03396099805831909, + 0.03879513964056969, + 0.024642469361424446, + 0.013321293517947197, + 0.01337639894336462, + 0.012244717217981815, + 0.016180863603949547, + 0.015590768307447433, + 0.01699487492442131, + 0.01654798537492752, + 0.012324389070272446, + 0.031581420451402664, + 0.02588709443807602, + 0.02980095148086548, + 0.03504505380988121, + 0.03639071807265282, + 0.02608674392104149, + 0.036752935498952866, + 0.03403303399682045, + 0.006094536278396845, + 0.004523169249296188, + 0.011495827697217464, + 0.005077390465885401, + 0.004732216242700815, + 0.0069654835388064384, + 0.005908160470426083, + 0.012786606326699257, + 0.0038719032891094685, + 0.0032708868384361267, + 0.0032610774505883455, + 0.002654053969308734, + 0.0025886876974254847, + 0.0017811269499361515, + 0.004682247061282396, + 0.005062103737145662, + 0.007712248712778091, + 0.003669755533337593, + 0.009819570928812027, + 0.004867886658757925, + 0.004384705331176519, + 0.005449439398944378, + 0.004325870890170336, + 0.011077411472797394, + 0.0065610576421022415, + 0.011300020851194859, + 0.00797527190297842, + 0.006075466051697731, + 0.004589078016579151, + 0.004574696067720652, + 0.003797411685809493, + 0.006802273914217949, + 0.005196929443627596, + 0.006600871216505766, + 0.003895949572324753, + 0.003127424279227853, + 0.004687566310167313, + 0.0033919711131602526, + 0.0034569071140140295, + 0.003899009432643652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.1261448711156845 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42897114157676697 + ], + "opposite_logprob_tokens": [ + -1.0539711713790894 + ], + "logprob_reward": -0.42897114157676697, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42897114157676697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1657380610704422, + 0.0924367681145668, + 0.1371619552373886, + 0.074250727891922, + 0.1652221828699112, + 0.06426119059324265, + 0.09109374135732651, + 0.07846985012292862, + 0.08419779688119888, + 0.04716778174042702 + ], + "credit_map": [ + 0.023304255679249763, + 0.01747359335422516, + 0.016756173223257065, + 0.021722467616200447, + 0.025155609473586082, + 0.018221428617835045, + 0.02575298398733139, + 0.017351539805531502, + 0.008868915028870106, + 0.007300239522010088, + 0.010778370313346386, + 0.012176673859357834, + 0.014663336798548698, + 0.011814791709184647, + 0.014695346355438232, + 0.012139089405536652, + 0.014140088111162186, + 0.013416321016848087, + 0.021314572542905807, + 0.014785188250243664, + 0.01839696615934372, + 0.01962726004421711, + 0.022041523829102516, + 0.013440024107694626, + 0.008981695398688316, + 0.009687609039247036, + 0.008603636175394058, + 0.007485615089535713, + 0.010150245390832424, + 0.010535633191466331, + 0.010111036710441113, + 0.008695252239704132, + 0.02695431374013424, + 0.011070002801716328, + 0.027118312194943428, + 0.020504089072346687, + 0.015340672805905342, + 0.020392967388033867, + 0.020755304023623466, + 0.023086506873369217, + 0.012412511743605137, + 0.004518268164247274, + 0.01296046283096075, + 0.003610995365306735, + 0.003530676243826747, + 0.010550109669566154, + 0.004119011573493481, + 0.012559154070913792, + 0.014085110276937485, + 0.013417169451713562, + 0.011517157778143883, + 0.015518950298428535, + 0.00988183543086052, + 0.00833118800073862, + 0.010465278290212154, + 0.007877047173678875, + 0.012680578045547009, + 0.004637779202312231, + 0.010215566493570805, + 0.007732470985502005, + 0.005894099362194538, + 0.008833595551550388, + 0.007649535313248634, + 0.020826220512390137, + 0.012355023995041847, + 0.012811879627406597, + 0.015114795416593552, + 0.008213389664888382, + 0.009797211736440659, + 0.0069824643433094025, + 0.007705681957304478, + 0.011217345483601093, + 0.006250185891985893, + 0.008421954698860645, + 0.00533091276884079, + 0.004801091738045216, + 0.0038509792648255825, + 0.0033242334611713886, + 0.00681527704000473, + 0.008373142220079899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.42897114157676697 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644422516226768 + ], + "opposite_logprob_tokens": [ + -4.7586445808410645 + ], + "logprob_reward": -0.008644422516226768, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644422516226768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22655460238456726, + 0.07415422052145004, + 0.044650863856077194, + 0.04751303046941757, + 0.10747859627008438, + 0.05488564074039459, + 0.046950843185186386, + 0.0879790261387825, + 0.050605956465005875, + 0.2592272162437439 + ], + "credit_map": [ + 0.031067101284861565, + 0.02165966108441353, + 0.025617213919758797, + 0.023049360141158104, + 0.030932852998375893, + 0.026396069675683975, + 0.04638800770044327, + 0.021444331854581833, + 0.009158138185739517, + 0.0058961715549230576, + 0.010549628175795078, + 0.007629574742168188, + 0.010781652294099331, + 0.011079588904976845, + 0.011368999257683754, + 0.007690466474741697, + 0.0050610448233783245, + 0.0028386886697262526, + 0.006797106470912695, + 0.002952083945274353, + 0.007308999076485634, + 0.007725340314209461, + 0.00461860466748476, + 0.007348996587097645, + 0.006631418131291866, + 0.004378264304250479, + 0.005416556727141142, + 0.009578694589436054, + 0.005680926609784365, + 0.005424258764833212, + 0.005124971736222506, + 0.005277939140796661, + 0.01858050562441349, + 0.007952183485031128, + 0.019103363156318665, + 0.01642659492790699, + 0.01110262330621481, + 0.01387120597064495, + 0.009127859957516193, + 0.0113142691552639, + 0.007158155087381601, + 0.00541117787361145, + 0.008391893468797207, + 0.005863524507731199, + 0.005236867815256119, + 0.007652961649000645, + 0.005563609767705202, + 0.009607445448637009, + 0.006662783678621054, + 0.005874719005078077, + 0.006841340102255344, + 0.004247573204338551, + 0.005738097708672285, + 0.0033344493713229895, + 0.00891320500522852, + 0.005338673014193773, + 0.01296621561050415, + 0.006028598640114069, + 0.012868611142039299, + 0.008635241538286209, + 0.00751115009188652, + 0.006658533588051796, + 0.011564421467483044, + 0.021746253594756126, + 0.007520497776567936, + 0.0076492163352668285, + 0.007162101101130247, + 0.005095313303172588, + 0.006642501335591078, + 0.005570612847805023, + 0.004862368106842041, + 0.0061033437959849834, + 0.036675553768873215, + 0.058349624276161194, + 0.037514012306928635, + 0.020290348678827286, + 0.02929558791220188, + 0.021021217107772827, + 0.024779820814728737, + 0.03130105510354042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.008644422516226768 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005603650351986289 + ], + "opposite_logprob_tokens": [ + -7.500560283660889 + ], + "logprob_reward": -0.0005603650351986289, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005603650351986289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13153277337551117, + 0.17038361728191376, + 0.1623629629611969, + 0.04952662065625191, + 0.08201571553945541, + 0.0660773441195488, + 0.07081036269664764, + 0.07328301668167114, + 0.041907280683517456, + 0.1521003246307373 + ], + "credit_map": [ + 0.016244815662503242, + 0.014761854894459248, + 0.015416824258863926, + 0.01309171598404646, + 0.018075348809361458, + 0.016622737050056458, + 0.017436832189559937, + 0.01988263987004757, + 0.036293935030698776, + 0.014639115892350674, + 0.0347336046397686, + 0.011414865031838417, + 0.012993496842682362, + 0.029067473486065865, + 0.009209941141307354, + 0.022031189873814583, + 0.021916482597589493, + 0.010728908702731133, + 0.030431799590587616, + 0.013866322115063667, + 0.021837977692484856, + 0.021798670291900635, + 0.013283537700772285, + 0.028499262407422066, + 0.007224616594612598, + 0.006796499248594046, + 0.004933399613946676, + 0.0038074320182204247, + 0.008185898885130882, + 0.007872480899095535, + 0.005876189563423395, + 0.004830104764550924, + 0.011594253592193127, + 0.008868187665939331, + 0.01394860353320837, + 0.007943926379084587, + 0.01084040105342865, + 0.009698321111500263, + 0.007931359112262726, + 0.011190658435225487, + 0.010590345598757267, + 0.005336484406143427, + 0.010305949486792088, + 0.006832681130617857, + 0.004950863309204578, + 0.011341470293700695, + 0.005305267404764891, + 0.011414279229938984, + 0.00998449232429266, + 0.0064126295037567616, + 0.010666071437299252, + 0.0070856413803994656, + 0.007874288596212864, + 0.007192662451416254, + 0.012553582899272442, + 0.009040986187756062, + 0.012643997557461262, + 0.004542030394077301, + 0.011708654463291168, + 0.00770498625934124, + 0.006758436560630798, + 0.00620567100122571, + 0.009459585882723331, + 0.01425965316593647, + 0.008218508213758469, + 0.005028935149312019, + 0.00601666746661067, + 0.0027432062197476625, + 0.003987427800893784, + 0.002605704590678215, + 0.002652290742844343, + 0.010654544457793236, + 0.022018298506736755, + 0.030633335933089256, + 0.02587212063372135, + 0.010566364973783493, + 0.017656363546848297, + 0.011638360098004341, + 0.013089885003864765, + 0.020625604316592216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.0005603650351986289 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017128810286521912 + ], + "opposite_logprob_tokens": [ + -6.376712799072266 + ], + "logprob_reward": -0.0017128810286521912, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017128810286521912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1718658059835434, + 0.11347244679927826, + 0.11291259527206421, + 0.22538362443447113, + 0.0738881453871727, + 0.11775760352611542, + 0.03770190104842186, + 0.05804203078150749, + 0.026226414367556572, + 0.06274943798780441 + ], + "credit_map": [ + 0.024397622793912888, + 0.016464408487081528, + 0.017941484227776527, + 0.015544583089649677, + 0.02788849174976349, + 0.01927260123193264, + 0.034893013536930084, + 0.015463583171367645, + 0.020291198045015335, + 0.010952858254313469, + 0.012911571189761162, + 0.008282245136797428, + 0.016765862703323364, + 0.016343681141734123, + 0.008538702502846718, + 0.019386325031518936, + 0.011296363547444344, + 0.010940749198198318, + 0.011789161711931229, + 0.013772337697446346, + 0.0193692184984684, + 0.014624109491705894, + 0.020131316035985947, + 0.010989334434270859, + 0.0257114265114069, + 0.030896564945578575, + 0.02059173583984375, + 0.039336614310741425, + 0.030551576986908913, + 0.027549004182219505, + 0.024545812979340553, + 0.02620088867843151, + 0.014045912772417068, + 0.0059987218119204044, + 0.012849081307649612, + 0.005871865898370743, + 0.005623022094368935, + 0.01143695879727602, + 0.006200782489031553, + 0.011861804872751236, + 0.0241082813590765, + 0.009775212965905666, + 0.02228091098368168, + 0.008060024119913578, + 0.00743388757109642, + 0.020765211433172226, + 0.006363669876009226, + 0.01897040195763111, + 0.0056104399263858795, + 0.00415569543838501, + 0.007637999020516872, + 0.0034280975814908743, + 0.0035118882078677416, + 0.004336058162152767, + 0.003999154083430767, + 0.005022567696869373, + 0.00912315584719181, + 0.004364324267953634, + 0.01029134914278984, + 0.006467102561146021, + 0.006512459833174944, + 0.004104875959455967, + 0.005289963446557522, + 0.011888804845511913, + 0.004043754190206528, + 0.0043067061342298985, + 0.0033236979506909847, + 0.002182725118473172, + 0.0025936260353773832, + 0.003075656481087208, + 0.001605750061571598, + 0.005094499327242374, + 0.00870522391051054, + 0.011346434243023396, + 0.008348981849849224, + 0.00600782036781311, + 0.006785389501601458, + 0.004043630789965391, + 0.0072232927195727825, + 0.010288664139807224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.0017128810286521912 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, dynamic tracking shot captures a lone knight atop a majestic black warhorse, galloping across endless golden grasslands. The knight, clad in weathered steel plate armor, its surface covered in tiny scratches and battle scars, wears a full-face helmet that obscures his face, and a tattered crimson cape billows violently behind him at high speed. His steed, a powerfully muscled beast, its pure black fur glistening in the sunset, strikes the ground with a powerful, rhythmic cadence, each landing sending a swift and powerful kick of dry earth. In the background, a massive, orange-red sunset slowly sinks into the distant horizon, its final rays tinting the world with a warm orange-gold hue, lengthening the lone shadows of knight and horse across the grassland. This golden backlight sharply outlines the hard contours of the knight and the horse, and the dust stirred up by the horse's hooves is illuminated in the light, forming a golden mist flying around them. The whole picture is full of epic grandeur, speed and heroic loneliness.", + "seed": 11778, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_2/index_6912_seed_11778_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16646978817880154, + "vlm_logprob": -0.16646978817880154, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.16646978817880154, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.27628469467163086 + ], + "opposite_logprob_tokens": [ + -1.6512846946716309 + ], + "logprob_reward": -0.27628469467163086, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.27628469467163086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07542470097541809, + 0.04134742170572281, + 0.02287185750901699, + 0.032742951065301895, + 0.03212030604481697, + 0.10678355395793915, + 0.05080300197005272, + 0.22617381811141968, + 0.20761512219905853, + 0.2041172832250595 + ], + "credit_map": [ + 0.005108295939862728, + 0.00864360574632883, + 0.010251430794596672, + 0.012080264277756214, + 0.008288804441690445, + 0.00815372634679079, + 0.012204609811306, + 0.010693955235183239, + 0.0038737934082746506, + 0.004215711262077093, + 0.004262592643499374, + 0.006283972878009081, + 0.007492508739233017, + 0.004450392909348011, + 0.006330438889563084, + 0.004438009578734636, + 0.003096765372902155, + 0.0019940240308642387, + 0.004319246392697096, + 0.003109188750386238, + 0.002391804475337267, + 0.002580862957984209, + 0.002722837496548891, + 0.002657128032296896, + 0.0039757248014211655, + 0.0025538126938045025, + 0.0047404286451637745, + 0.004760671406984329, + 0.003481472609564662, + 0.0034243783447891474, + 0.005508573725819588, + 0.004297889769077301, + 0.00453274417668581, + 0.003127464558929205, + 0.00580451637506485, + 0.003552533918991685, + 0.003957103937864304, + 0.00362428929656744, + 0.0036358144134283066, + 0.003885839367285371, + 0.0105331065133214, + 0.013351424597203732, + 0.013882960192859173, + 0.01769423484802246, + 0.01548616960644722, + 0.010643458925187588, + 0.014696117490530014, + 0.010496080853044987, + 0.005411889404058456, + 0.003978736698627472, + 0.007394284009933472, + 0.005547821521759033, + 0.00431858841329813, + 0.004988300614058971, + 0.006849818862974644, + 0.012313567101955414, + 0.020771389827132225, + 0.007603107951581478, + 0.07277970761060715, + 0.016506845131516457, + 0.008729101158678532, + 0.06721937656402588, + 0.012707425281405449, + 0.019856860861182213, + 0.014888843521475792, + 0.012531016953289509, + 0.02403757907450199, + 0.04655064269900322, + 0.017146296799182892, + 0.035641249269247055, + 0.030864685773849487, + 0.02595481462776661, + 0.012944158166646957, + 0.014375241473317146, + 0.061044689267873764, + 0.0266598928719759, + 0.014168634079396725, + 0.03741779178380966, + 0.01651638373732567, + 0.020990492776036263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.27628469467163086 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234083533287048 + ], + "opposite_logprob_tokens": [ + -0.8984083533287048 + ], + "logprob_reward": -0.5234083533287048, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234083533287048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10482218116521835, + 0.05526699498295784, + 0.06917276233434677, + 0.0343681238591671, + 0.05202975869178772, + 0.1437193602323532, + 0.050358228385448456, + 0.05226833373308182, + 0.12658728659152985, + 0.3114069998264313 + ], + "credit_map": [ + 0.010679050348699093, + 0.009329852648079395, + 0.017322732135653496, + 0.016465740278363228, + 0.0119051244109869, + 0.011499096639454365, + 0.015414263121783733, + 0.012206326238811016, + 0.0048739914782345295, + 0.003781126579269767, + 0.009929150342941284, + 0.006730110850185156, + 0.007349268067628145, + 0.005134028848260641, + 0.009876841679215431, + 0.007592474110424519, + 0.006178682669997215, + 0.005334151908755302, + 0.011619693599641323, + 0.008839739486575127, + 0.011542698368430138, + 0.008105402812361717, + 0.008349592797458172, + 0.009202797897160053, + 0.003138998756185174, + 0.0031757636461406946, + 0.004167367238551378, + 0.0037206874694675207, + 0.004359686281532049, + 0.003065064549446106, + 0.007466729264706373, + 0.005273826885968447, + 0.0036579344887286425, + 0.004348041024059057, + 0.0072527313604950905, + 0.007863800041377544, + 0.008749788627028465, + 0.005513141863048077, + 0.00726674497127533, + 0.007377577945590019, + 0.010788066312670708, + 0.019122440367937088, + 0.01713937148451805, + 0.021172882989048958, + 0.028067972511053085, + 0.013570243492722511, + 0.022376948967576027, + 0.011481435038149357, + 0.005283876787871122, + 0.003824685700237751, + 0.0067489794455468655, + 0.006023561116307974, + 0.004079420119524002, + 0.0048895906656980515, + 0.007102746516466141, + 0.012405372224748135, + 0.0036434363573789597, + 0.004004682879894972, + 0.00788598321378231, + 0.010317528620362282, + 0.0061527760699391365, + 0.0041081877425313, + 0.008778850547969341, + 0.007376887369900942, + 0.007912029512226582, + 0.00835481658577919, + 0.013301370665431023, + 0.030454248189926147, + 0.010810467414557934, + 0.019276117905974388, + 0.020526543259620667, + 0.015951691195368767, + 0.019428877159953117, + 0.02070806547999382, + 0.09462843835353851, + 0.040356021374464035, + 0.01903211511671543, + 0.06454866379499435, + 0.02407849207520485, + 0.028626294806599617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.5234083533287048 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026359083130955696 + ], + "opposite_logprob_tokens": [ + -3.6513590812683105 + ], + "logprob_reward": -0.026359083130955696, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026359083130955696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10957178473472595, + 0.08963577449321747, + 0.042098116129636765, + 0.03759355470538139, + 0.058997441083192825, + 0.10548141598701477, + 0.028870176523923874, + 0.11829626560211182, + 0.05289661884307861, + 0.3565588891506195 + ], + "credit_map": [ + 0.010302884504199028, + 0.010455461218953133, + 0.01592596247792244, + 0.013945688493549824, + 0.014723824337124825, + 0.012147491797804832, + 0.01919909380376339, + 0.012871363200247288, + 0.006693297065794468, + 0.006916471756994724, + 0.010272604413330555, + 0.011082559823989868, + 0.020564530044794083, + 0.010134347714483738, + 0.015761684626340866, + 0.008210274390876293, + 0.0031091291457414627, + 0.0032355051953345537, + 0.005339985713362694, + 0.005558266770094633, + 0.007420273032039404, + 0.0041228532791137695, + 0.008107553236186504, + 0.005204546265304089, + 0.0030923334416002035, + 0.004078047350049019, + 0.004656240344047546, + 0.006352788303047419, + 0.004973616451025009, + 0.0039184787310659885, + 0.005969982594251633, + 0.004552067723125219, + 0.004004552960395813, + 0.004763994365930557, + 0.008008806966245174, + 0.010369171388447285, + 0.008670558221638203, + 0.005884310230612755, + 0.009518982842564583, + 0.007777058985084295, + 0.01032302901148796, + 0.012453385628759861, + 0.014257720671594143, + 0.01460838969796896, + 0.013920532539486885, + 0.010735750198364258, + 0.016277329996228218, + 0.012905267998576164, + 0.0032207085750997066, + 0.002183520933613181, + 0.0049967472441494465, + 0.0033774294424802065, + 0.0027839012909680605, + 0.0042105792090296745, + 0.003306108759716153, + 0.004791178274899721, + 0.009737718850374222, + 0.004736984148621559, + 0.03529396280646324, + 0.009402959607541561, + 0.006325367372483015, + 0.032257672399282455, + 0.008555179461836815, + 0.011986401863396168, + 0.0032557817175984383, + 0.0032552059274166822, + 0.005852698814123869, + 0.012854767963290215, + 0.004185862839221954, + 0.008224275894463062, + 0.009593931958079338, + 0.005674091167747974, + 0.023815952241420746, + 0.019585201516747475, + 0.10919284075498581, + 0.042053524404764175, + 0.015411111526191235, + 0.09140345454216003, + 0.025647340342402458, + 0.029449433088302612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.026359083130955696 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010432521812617779 + ], + "opposite_logprob_tokens": [ + -6.876043319702148 + ], + "logprob_reward": -0.0010432521812617779, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010432521812617779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16981464624404907, + 0.1041807159781456, + 0.05684966593980789, + 0.09375309199094772, + 0.040094293653964996, + 0.10652276128530502, + 0.034141454845666885, + 0.1008339524269104, + 0.0353601835668087, + 0.2584492564201355 + ], + "credit_map": [ + 0.02269802801311016, + 0.017230985686182976, + 0.029215890914201736, + 0.019378168508410454, + 0.022098062559962273, + 0.021391715854406357, + 0.018182003870606422, + 0.019619781523942947, + 0.012518053874373436, + 0.008150056004524231, + 0.01736762747168541, + 0.010929228737950325, + 0.017661208286881447, + 0.011620690114796162, + 0.015501279383897781, + 0.010432576760649681, + 0.007384283468127251, + 0.005128443241119385, + 0.01079318206757307, + 0.006019528955221176, + 0.0069681922905147076, + 0.006954692304134369, + 0.007859366945922375, + 0.005741974338889122, + 0.012262561358511448, + 0.009414232335984707, + 0.017658745869994164, + 0.014490327797830105, + 0.00941387563943863, + 0.01154274307191372, + 0.008829208090901375, + 0.010141396895051003, + 0.00643358938395977, + 0.003035227069631219, + 0.007381850387901068, + 0.005091409664601088, + 0.004297557752579451, + 0.005560981575399637, + 0.0040967888198792934, + 0.0041968864388763905, + 0.013965520076453686, + 0.012812582775950432, + 0.016728486865758896, + 0.011850719340145588, + 0.015426995232701302, + 0.014699666760861874, + 0.010901295579969883, + 0.010137504898011684, + 0.005145885515958071, + 0.0026569142937660217, + 0.006912812124937773, + 0.003955713473260403, + 0.0029649678617715836, + 0.00466937618330121, + 0.0036279272753745317, + 0.0042078569531440735, + 0.008851139806210995, + 0.0043766614980995655, + 0.028771644458174706, + 0.009133224375545979, + 0.0062797837890684605, + 0.024086011573672295, + 0.008611580356955528, + 0.010723913088440895, + 0.003107996890321374, + 0.002862766617909074, + 0.007543432991951704, + 0.004669636022299528, + 0.0020664935000240803, + 0.004132134839892387, + 0.003833336755633354, + 0.0071443854831159115, + 0.027838096022605896, + 0.022495480254292488, + 0.05033717304468155, + 0.024657459929585457, + 0.021739782765507698, + 0.07030000537633896, + 0.018825026229023933, + 0.022256216034293175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.0010432521812617779 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052535575814545155 + ], + "opposite_logprob_tokens": [ + -5.255253791809082 + ], + "logprob_reward": -0.0052535575814545155, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052535575814545155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1163334921002388, + 0.03096238151192665, + 0.03593188896775246, + 0.0580122247338295, + 0.06749153137207031, + 0.13515058159828186, + 0.021349741145968437, + 0.14604294300079346, + 0.026268158107995987, + 0.36245712637901306 + ], + "credit_map": [ + 0.015106914564967155, + 0.009747852571308613, + 0.024009395390748978, + 0.011083897203207016, + 0.007999544963240623, + 0.016287872567772865, + 0.014790588989853859, + 0.01730741746723652, + 0.003490982810035348, + 0.0021832729689776897, + 0.005360288079828024, + 0.0034446038771420717, + 0.004957709927111864, + 0.0035982662811875343, + 0.004101620987057686, + 0.0038256340194493532, + 0.005089887417852879, + 0.0028178419452160597, + 0.005466417875140905, + 0.0045618279837071896, + 0.004940409678965807, + 0.004131810739636421, + 0.0050209565088152885, + 0.0039027328602969646, + 0.005635952576994896, + 0.005382138770073652, + 0.008910207077860832, + 0.00830636266618967, + 0.0054969764314591885, + 0.006098994519561529, + 0.011120899580419064, + 0.00706068892031908, + 0.004880262538790703, + 0.0052318391390144825, + 0.008259348571300507, + 0.01069528330117464, + 0.010467087849974632, + 0.005732695106416941, + 0.012935010716319084, + 0.009289998561143875, + 0.012005730532109737, + 0.01712438464164734, + 0.016845399513840675, + 0.017469555139541626, + 0.025099661201238632, + 0.014511574991047382, + 0.021697482094168663, + 0.010396784171462059, + 0.001601161202415824, + 0.0016688919859007, + 0.002930000191554427, + 0.003260866040363908, + 0.0031716639641672373, + 0.0022418475709855556, + 0.003377024782821536, + 0.0030982838943600655, + 0.012717047706246376, + 0.005159405991435051, + 0.04575803130865097, + 0.01168065331876278, + 0.006667203735560179, + 0.04064112529158592, + 0.009963937103748322, + 0.013455526903271675, + 0.002318999031558633, + 0.0015341968974098563, + 0.00405587162822485, + 0.004092676565051079, + 0.0016456908779218793, + 0.004715402610599995, + 0.003231652081012726, + 0.004673664923757315, + 0.0374133475124836, + 0.020284833386540413, + 0.09491614997386932, + 0.03377040475606918, + 0.01436634548008442, + 0.11416169255971909, + 0.022299516946077347, + 0.02524477429687977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.0052535575814545155 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, dynamic tracking shot captures a lone knight atop a majestic black warhorse, galloping across endless golden grasslands. The knight, clad in weathered steel plate armor, its surface covered in tiny scratches and battle scars, wears a full-face helmet that obscures his face, and a tattered crimson cape billows violently behind him at high speed. His steed, a powerfully muscled beast, its pure black fur glistening in the sunset, strikes the ground with a powerful, rhythmic cadence, each landing sending a swift and powerful kick of dry earth. In the background, a massive, orange-red sunset slowly sinks into the distant horizon, its final rays tinting the world with a warm orange-gold hue, lengthening the lone shadows of knight and horse across the grassland. This golden backlight sharply outlines the hard contours of the knight and the horse, and the dust stirred up by the horse's hooves is illuminated in the light, forming a golden mist flying around them. The whole picture is full of epic grandeur, speed and heroic loneliness.", + "seed": 11779, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_2/index_6912_seed_11779_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07419375750469044, + "vlm_logprob": -0.07419375750469044, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.07419375750469044, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15391652286052704 + ], + "opposite_logprob_tokens": [ + -2.278916597366333 + ], + "logprob_reward": -0.15391652286052704, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.15391652286052704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15119190514087677, + 0.0509798526763916, + 0.08631417155265808, + 0.4540603756904602, + 0.02853425033390522, + 0.03514738008379936, + 0.06709399819374084, + 0.05986160784959793, + 0.013975497335195541, + 0.052840955555438995 + ], + "credit_map": [ + 0.02199482172727585, + 0.012016527354717255, + 0.01941995695233345, + 0.014999663457274437, + 0.016699325293302536, + 0.01817970722913742, + 0.014962497167289257, + 0.032919399440288544, + 0.003939033951610327, + 0.0068456693552434444, + 0.00615295534953475, + 0.009478666819632053, + 0.007658453192561865, + 0.005777498707175255, + 0.005830826703459024, + 0.005296746268868446, + 0.010228592902421951, + 0.008138143457472324, + 0.010941041633486748, + 0.01412280835211277, + 0.011009647510945797, + 0.009380904957652092, + 0.01101138535887003, + 0.01148164551705122, + 0.09152637422084808, + 0.03293520212173462, + 0.05686500668525696, + 0.04978156462311745, + 0.04519079253077507, + 0.06870885193347931, + 0.031631700694561005, + 0.07742089033126831, + 0.004127779975533485, + 0.0027186868246644735, + 0.00288075371645391, + 0.002544010290876031, + 0.00493816751986742, + 0.00353676313534379, + 0.0052035171538591385, + 0.00258457288146019, + 0.0049889786168932915, + 0.002483085496351123, + 0.007680599577724934, + 0.0027813592460006475, + 0.003852587891742587, + 0.00512418057769537, + 0.002591622993350029, + 0.005644968245178461, + 0.016698943451046944, + 0.0046164789237082005, + 0.0070736766792833805, + 0.004156382754445076, + 0.0053369770757853985, + 0.01802213117480278, + 0.00527418777346611, + 0.005915218032896519, + 0.01707294210791588, + 0.0056444997899234295, + 0.01554369181394577, + 0.004460304509848356, + 0.004605214111506939, + 0.003761110594496131, + 0.0035458351485431194, + 0.005228012800216675, + 0.0016491932328790426, + 0.0018642388749867678, + 0.001259914948605001, + 0.0009609495755285025, + 0.0031789294444024563, + 0.0016133412718772888, + 0.0017662663012742996, + 0.001682664966210723, + 0.010716891847550869, + 0.006132941227406263, + 0.007290407549589872, + 0.004762871656566858, + 0.007718448527157307, + 0.005334099754691124, + 0.005815605167299509, + 0.005069685634225607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.15391652286052704 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20163659751415253 + ], + "opposite_logprob_tokens": [ + -1.701636552810669 + ], + "logprob_reward": -0.20163659751415253, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20163659751415253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04834626242518425, + 0.032868918031454086, + 0.1661480963230133, + 0.08924482762813568, + 0.0314958430826664, + 0.030082926154136658, + 0.02068650722503662, + 0.09924328327178955, + 0.011723313480615616, + 0.4701600968837738 + ], + "credit_map": [ + 0.006825954653322697, + 0.005913018714636564, + 0.006355763413012028, + 0.00600137235596776, + 0.005334374960511923, + 0.005191688891500235, + 0.005221017636358738, + 0.007503075059503317, + 0.00305295642465353, + 0.0038428439293056726, + 0.0034530318807810545, + 0.006759903393685818, + 0.004452827852219343, + 0.002746493322774768, + 0.003866664133965969, + 0.0046941968612372875, + 0.018345342949032784, + 0.01645059697329998, + 0.02323751524090767, + 0.024630101397633553, + 0.022476786747574806, + 0.016510605812072754, + 0.01846378482878208, + 0.026033371686935425, + 0.019813470542430878, + 0.005392508581280708, + 0.012740371748805046, + 0.008723309263586998, + 0.008171097375452518, + 0.014356948435306549, + 0.005812546703964472, + 0.014234570786356926, + 0.004326891619712114, + 0.004373529460281134, + 0.004067687783390284, + 0.00317964400164783, + 0.0046737948432564735, + 0.0038854677695780993, + 0.004149573855102062, + 0.0028392544481903315, + 0.0037361865397542715, + 0.002915991935878992, + 0.003974054008722305, + 0.003403558861464262, + 0.0045161000452935696, + 0.0038605062291026115, + 0.004069264512509108, + 0.0036072663497179747, + 0.003070557489991188, + 0.0023341444320976734, + 0.00251062773168087, + 0.002062975661829114, + 0.0026809934061020613, + 0.0029773979913443327, + 0.0022316842805594206, + 0.002818128326907754, + 0.016156848520040512, + 0.012101947329938412, + 0.018017012625932693, + 0.008974147960543633, + 0.01653958112001419, + 0.008504330180585384, + 0.009177799336612225, + 0.009771606884896755, + 0.0016864417120814323, + 0.0015293655451387167, + 0.001167932292446494, + 0.0014652954414486885, + 0.002298886887729168, + 0.0013380880700424314, + 0.0012238150229677558, + 0.0010134894400835037, + 0.0906519666314125, + 0.1912146508693695, + 0.10492593050003052, + 0.01819448359310627, + 0.022354744374752045, + 0.016121869906783104, + 0.014319585636258125, + 0.012376813217997551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.20163659751415253 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081226170063019 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011081226170063019, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081226170063019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11632589995861053, + 0.05299314856529236, + 0.2605021297931671, + 0.28534644842147827, + 0.043188318610191345, + 0.03861235827207565, + 0.06136256828904152, + 0.07124301791191101, + 0.01163032092154026, + 0.058795809745788574 + ], + "credit_map": [ + 0.017268316820263863, + 0.007027992978692055, + 0.01697460003197193, + 0.012086920440196991, + 0.009654945693910122, + 0.014456541277468204, + 0.011031869798898697, + 0.027824709191918373, + 0.0053469459526240826, + 0.0062614018097519875, + 0.00587959960103035, + 0.008192158304154873, + 0.005971856415271759, + 0.005748362746089697, + 0.006136841140687466, + 0.009455980733036995, + 0.028745317831635475, + 0.026601819321513176, + 0.029010489583015442, + 0.028933128342032433, + 0.049430087208747864, + 0.027104802429676056, + 0.03133491799235344, + 0.03934154286980629, + 0.061000339686870575, + 0.019714942201972008, + 0.037091974169015884, + 0.029995443299412727, + 0.02803949825465679, + 0.04435097053647041, + 0.01917733997106552, + 0.0459759421646595, + 0.003581446595489979, + 0.005843718536198139, + 0.004477397538721561, + 0.004473850131034851, + 0.008737094700336456, + 0.004104288760572672, + 0.008453614078462124, + 0.003516905242577195, + 0.005051042418926954, + 0.003414785023778677, + 0.006319479085505009, + 0.00370760727673769, + 0.004896884318441153, + 0.00494350865483284, + 0.004248785320669413, + 0.006030261516571045, + 0.017568182200193405, + 0.004188118036836386, + 0.0043856254778802395, + 0.00471845967695117, + 0.003447039984166622, + 0.01815180294215679, + 0.004769224673509598, + 0.004134108312427998, + 0.02422800287604332, + 0.004780816379934549, + 0.02134268917143345, + 0.003402242437005043, + 0.0063764615915715694, + 0.0030689705163240433, + 0.0033477130346000195, + 0.004696112126111984, + 0.0010696436511352658, + 0.0010072094155475497, + 0.0012742426479235291, + 0.0018213974544778466, + 0.0019926843233406544, + 0.001529454835690558, + 0.001404465176165104, + 0.0015312227187678218, + 0.00982693675905466, + 0.01000509224832058, + 0.01484627928584814, + 0.004627627786248922, + 0.004544548690319061, + 0.004566938150674105, + 0.00432899734005332, + 0.006049385294318199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.011081226170063019 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007184546557255089 + ], + "opposite_logprob_tokens": [ + -7.250718593597412 + ], + "logprob_reward": -0.0007184546557255089, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007184546557255089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2724915146827698, + 0.05422785133123398, + 0.17833951115608215, + 0.19350960850715637, + 0.04470613971352577, + 0.07259373366832733, + 0.04540300369262695, + 0.08911186456680298, + 0.01795773208141327, + 0.03165910392999649 + ], + "credit_map": [ + 0.04453432187438011, + 0.016756154596805573, + 0.04154546558856964, + 0.02502870000898838, + 0.023558562621474266, + 0.03412731736898422, + 0.024017266929149628, + 0.06292372196912766, + 0.009048490785062313, + 0.005709616467356682, + 0.007529233116656542, + 0.008001050911843777, + 0.0070960004813969135, + 0.0064420257695019245, + 0.005602855235338211, + 0.00479858135804534, + 0.021392086520791054, + 0.016569750383496284, + 0.02082010544836521, + 0.02057615853846073, + 0.03298431262373924, + 0.018374981358647346, + 0.020471081137657166, + 0.027151035144925117, + 0.0428500771522522, + 0.013347450643777847, + 0.025294192135334015, + 0.01942555606365204, + 0.017586180940270424, + 0.02975662238895893, + 0.012934358790516853, + 0.032315175980329514, + 0.005891219712793827, + 0.004003835376352072, + 0.006320207845419645, + 0.004734379705041647, + 0.006637135520577431, + 0.005594221409410238, + 0.005807251203805208, + 0.005717887543141842, + 0.014778060838580132, + 0.004508537240326405, + 0.014840485528111458, + 0.004762060474604368, + 0.007234923075884581, + 0.013690534047782421, + 0.004188038408756256, + 0.008591102436184883, + 0.006117735989391804, + 0.004647563211619854, + 0.004973888862878084, + 0.004735732451081276, + 0.003976171836256981, + 0.008107275702059269, + 0.006251361221075058, + 0.006593275815248489, + 0.03172845020890236, + 0.006606757175177336, + 0.027254823595285416, + 0.004815320950001478, + 0.005825652275234461, + 0.003931719344109297, + 0.0035782589111477137, + 0.0053708781488239765, + 0.0021358069498091936, + 0.0019995870534330606, + 0.0018170796101912856, + 0.0017265629721805453, + 0.0037891895044595003, + 0.0021030413918197155, + 0.0023016496561467648, + 0.0020848161075264215, + 0.005592290312051773, + 0.0039031512569636106, + 0.004558214917778969, + 0.003527024295181036, + 0.003197185229510069, + 0.003724472364410758, + 0.003230083268135786, + 0.00392668042331934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0007184546557255089 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036159863229840994 + ], + "opposite_logprob_tokens": [ + -5.628615856170654 + ], + "logprob_reward": -0.0036159863229840994, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036159863229840994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15600086748600006, + 0.07208298146724701, + 0.14892268180847168, + 0.11147906631231308, + 0.04776943102478981, + 0.05441680923104286, + 0.11851254850625992, + 0.14022012054920197, + 0.022245701402425766, + 0.12834979593753815 + ], + "credit_map": [ + 0.029861820861697197, + 0.012550784274935722, + 0.017450924962759018, + 0.015359862707555294, + 0.022000648081302643, + 0.018269404768943787, + 0.018319686874747276, + 0.0221877321600914, + 0.007381809409707785, + 0.005050020292401314, + 0.009356735274195671, + 0.014203817583620548, + 0.008720261976122856, + 0.007363454904407263, + 0.006923255976289511, + 0.01308362279087305, + 0.020263273268938065, + 0.01577141135931015, + 0.02254638448357582, + 0.017196007072925568, + 0.0163132231682539, + 0.01497383788228035, + 0.01967664808034897, + 0.02218189463019371, + 0.022064335644245148, + 0.0074343811720609665, + 0.015543131157755852, + 0.01342831365764141, + 0.012080969288945198, + 0.016477292403578758, + 0.007876294665038586, + 0.01657434180378914, + 0.006725660525262356, + 0.004670554306358099, + 0.005064622964709997, + 0.006720710080116987, + 0.005408333148807287, + 0.005858094431459904, + 0.0075942641124129295, + 0.005727191921323538, + 0.010357216000556946, + 0.005355044733732939, + 0.009685256518423557, + 0.004230097867548466, + 0.006845737341791391, + 0.008025314658880234, + 0.0037784401793032885, + 0.006139704026281834, + 0.03240159526467323, + 0.005150666460394859, + 0.011160477064549923, + 0.006687745917588472, + 0.005993019789457321, + 0.03716748207807541, + 0.011057616211473942, + 0.008893944323062897, + 0.040394894778728485, + 0.01176986563950777, + 0.03770211711525917, + 0.010421638377010822, + 0.013730869628489017, + 0.008224871009588242, + 0.007675657514482737, + 0.010300200432538986, + 0.0020160384010523558, + 0.0023600608110427856, + 0.001966738374903798, + 0.0024529581423848867, + 0.004978000186383724, + 0.0031306110322475433, + 0.0028592757880687714, + 0.0024820207618176937, + 0.02210591547191143, + 0.04750285670161247, + 0.02708015777170658, + 0.005104364361613989, + 0.00904518086463213, + 0.006328205578029156, + 0.004368813708424568, + 0.006814306601881981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.0036159863229840994 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, dynamic tracking shot captures a lone knight atop a majestic black warhorse, galloping across endless golden grasslands. The knight, clad in weathered steel plate armor, its surface covered in tiny scratches and battle scars, wears a full-face helmet that obscures his face, and a tattered crimson cape billows violently behind him at high speed. His steed, a powerfully muscled beast, its pure black fur glistening in the sunset, strikes the ground with a powerful, rhythmic cadence, each landing sending a swift and powerful kick of dry earth. In the background, a massive, orange-red sunset slowly sinks into the distant horizon, its final rays tinting the world with a warm orange-gold hue, lengthening the lone shadows of knight and horse across the grassland. This golden backlight sharply outlines the hard contours of the knight and the horse, and the dust stirred up by the horse's hooves is illuminated in the light, forming a golden mist flying around them. The whole picture is full of epic grandeur, speed and heroic loneliness.", + "seed": 11780, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_2/index_6912_seed_11780_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05727622710401192, + "vlm_logprob": -0.05727622710401192, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.05727622710401192, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13303251564502716 + ], + "opposite_logprob_tokens": [ + -2.3830325603485107 + ], + "logprob_reward": -0.13303251564502716, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.13303251564502716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2105763703584671, + 0.03765961155295372, + 0.05219924449920654, + 0.03167571872472763, + 0.12400895357131958, + 0.10173507034778595, + 0.04524659365415573, + 0.25025027990341187, + 0.11144053190946579, + 0.03520766645669937 + ], + "credit_map": [ + 0.06503401696681976, + 0.007920538075268269, + 0.02644983120262623, + 0.009696080349385738, + 0.008047131821513176, + 0.06660264730453491, + 0.010428344830870628, + 0.016397789120674133, + 0.0051374901086091995, + 0.0025879787281155586, + 0.005674171727150679, + 0.0030700508505105972, + 0.004018843173980713, + 0.004537547938525677, + 0.003079899586737156, + 0.009553628042340279, + 0.006006207317113876, + 0.005105311516672373, + 0.005072271451354027, + 0.005780343897640705, + 0.0065731569193303585, + 0.0062850890681147575, + 0.010607869364321232, + 0.006768995895981789, + 0.004237603861838579, + 0.003402695059776306, + 0.003382078604772687, + 0.004930276423692703, + 0.004642115905880928, + 0.002278587780892849, + 0.0035294636618345976, + 0.005272896960377693, + 0.016833698377013206, + 0.007748525124043226, + 0.019926458597183228, + 0.03972071781754494, + 0.007859395816922188, + 0.008013330399990082, + 0.011854777112603188, + 0.012052040547132492, + 0.012496478855609894, + 0.011086707003414631, + 0.018741047009825706, + 0.01411531399935484, + 0.014891857281327248, + 0.006829927675426006, + 0.012938043102622032, + 0.010635698214173317, + 0.0034626382403075695, + 0.004645314998924732, + 0.008553102612495422, + 0.006413019262254238, + 0.0065760258585214615, + 0.0032737136352807283, + 0.006038508843630552, + 0.006284266710281372, + 0.010932582430541515, + 0.04807174578309059, + 0.05255161598324776, + 0.01303917821496725, + 0.08529859036207199, + 0.017339054495096207, + 0.009199656546115875, + 0.013817833736538887, + 0.010026216506958008, + 0.014157422818243504, + 0.02100353315472603, + 0.016726301982998848, + 0.009837063029408455, + 0.00838024914264679, + 0.015030348673462868, + 0.01627938821911812, + 0.00397935975342989, + 0.003454025136306882, + 0.0037451102398335934, + 0.003269827226176858, + 0.004393870942294598, + 0.006766072474420071, + 0.003400560934096575, + 0.006198838818818331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.13303251564502716 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285419881343842 + ], + "opposite_logprob_tokens": [ + -2.0178542137145996 + ], + "logprob_reward": -0.14285419881343842, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14285419881343842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0608598031103611, + 0.020497620105743408, + 0.05535305663943291, + 0.014737875200808048, + 0.15217071771621704, + 0.08240268379449844, + 0.03874915465712547, + 0.4564274251461029, + 0.0708317756652832, + 0.04796991124749184 + ], + "credit_map": [ + 0.01086611207574606, + 0.0033420720137655735, + 0.012263204902410507, + 0.004958969075232744, + 0.0047570341266691685, + 0.010493889451026917, + 0.005211705807596445, + 0.008966808207333088, + 0.002453424734994769, + 0.0018449937924742699, + 0.003175147110596299, + 0.002785211428999901, + 0.0033944561146199703, + 0.0023049486335366964, + 0.0021620464976876974, + 0.002377390395849943, + 0.006669621914625168, + 0.006395610980689526, + 0.005358204711228609, + 0.005190765950828791, + 0.008632618002593517, + 0.004952088464051485, + 0.011446247808635235, + 0.0067078955471515656, + 0.0015861478168517351, + 0.001618852955289185, + 0.0015139025636017323, + 0.0016666961601004004, + 0.003447331488132477, + 0.0013560731895267963, + 0.0019165297271683812, + 0.0016323408344760537, + 0.020893916487693787, + 0.009177567437291145, + 0.026069406419992447, + 0.0527888648211956, + 0.008717158809304237, + 0.0061385552398860455, + 0.01448819600045681, + 0.013897047378122807, + 0.009860842488706112, + 0.00800798088312149, + 0.010503290221095085, + 0.012820463627576828, + 0.01621095836162567, + 0.00462739635258913, + 0.012341393157839775, + 0.008030352182686329, + 0.0035944078117609024, + 0.004468022380024195, + 0.006392582319676876, + 0.006168773397803307, + 0.004011794459074736, + 0.003138096071779728, + 0.0055822706781327724, + 0.005393205676227808, + 0.024521032348275185, + 0.04722491279244423, + 0.14391836524009705, + 0.036574553698301315, + 0.12234558910131454, + 0.033950045704841614, + 0.017133668065071106, + 0.030759241431951523, + 0.016329171136021614, + 0.010908086784183979, + 0.0099196657538414, + 0.01024619210511446, + 0.0055753071792423725, + 0.004647813271731138, + 0.005243691615760326, + 0.007961844094097614, + 0.007370665203779936, + 0.005982010159641504, + 0.006613454781472683, + 0.004968330264091492, + 0.0041605764999985695, + 0.0066402386873960495, + 0.00305837020277977, + 0.009176261723041534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.14285419881343842 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076340362429618835 + ], + "opposite_logprob_tokens": [ + -4.882634162902832 + ], + "logprob_reward": -0.0076340362429618835, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076340362429618835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33296647667884827, + 0.049197155982255936, + 0.04766574129462242, + 0.021685387939214706, + 0.06189189478754997, + 0.13267526030540466, + 0.08315829187631607, + 0.16088566184043884, + 0.0538676418364048, + 0.0560065433382988 + ], + "credit_map": [ + 0.11096403747797012, + 0.011638280935585499, + 0.03348584473133087, + 0.014641770161688328, + 0.012641692534089088, + 0.11571957170963287, + 0.011657251976430416, + 0.022218002006411552, + 0.007782385218888521, + 0.00253510195761919, + 0.008234579116106033, + 0.004783299285918474, + 0.0066869910806417465, + 0.007443963084369898, + 0.004635035526007414, + 0.007095798850059509, + 0.005670290905982256, + 0.004773752298206091, + 0.004187772050499916, + 0.004520590882748365, + 0.0070710857398808, + 0.006758680567145348, + 0.009701388888061047, + 0.0049821785651147366, + 0.003073198487982154, + 0.002250820863991976, + 0.0024414551444351673, + 0.003220904851332307, + 0.0030528432689607143, + 0.002291626762598753, + 0.002352884504944086, + 0.003001653589308262, + 0.008221796713769436, + 0.0038796281442046165, + 0.010916062630712986, + 0.02059931866824627, + 0.003199642524123192, + 0.003884620498865843, + 0.005332502536475658, + 0.005858315620571375, + 0.014167393557727337, + 0.012524993158876896, + 0.027075067162513733, + 0.019999271258711815, + 0.022067781537771225, + 0.00912858173251152, + 0.015097435563802719, + 0.012614731676876545, + 0.006428858265280724, + 0.006344926543533802, + 0.011914622969925404, + 0.012082602828741074, + 0.007997525855898857, + 0.0066413492895662785, + 0.020259877666831017, + 0.011488514021039009, + 0.008649053983390331, + 0.021275948733091354, + 0.046934064477682114, + 0.013782898895442486, + 0.039098113775253296, + 0.009424899704754353, + 0.010984924621880054, + 0.010735747404396534, + 0.010502948425710201, + 0.007927801460027695, + 0.007083436008542776, + 0.009498710744082928, + 0.004592937882989645, + 0.0036008234601467848, + 0.0060318284668028355, + 0.004629149101674557, + 0.008106628432869911, + 0.007464949041604996, + 0.009410737082362175, + 0.005369008053094149, + 0.0065042669884860516, + 0.005576670169830322, + 0.006715426687151194, + 0.006858853157609701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.0076340362429618835 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009201106731779873 + ], + "opposite_logprob_tokens": [ + -7.000920295715332 + ], + "logprob_reward": -0.0009201106731779873, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009201106731779873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2536076009273529, + 0.10642290860414505, + 0.04143672436475754, + 0.03381185606122017, + 0.18895898759365082, + 0.061341896653175354, + 0.05248904228210449, + 0.14098024368286133, + 0.06934329122304916, + 0.0516074113547802 + ], + "credit_map": [ + 0.06603097170591354, + 0.012401171028614044, + 0.02793164923787117, + 0.016702041029930115, + 0.016340801492333412, + 0.07334934175014496, + 0.018063951283693314, + 0.022787677124142647, + 0.015763988718390465, + 0.007134695071727037, + 0.022005271166563034, + 0.008909903466701508, + 0.011146080680191517, + 0.017602674663066864, + 0.0066683595068752766, + 0.017191948369145393, + 0.004578799474984407, + 0.0025643515400588512, + 0.007205856963992119, + 0.006133595481514931, + 0.0042312974110245705, + 0.002695034258067608, + 0.00858068373054266, + 0.005447108764201403, + 0.004681348334997892, + 0.0027755359187722206, + 0.004942047409713268, + 0.008562416769564152, + 0.0035416793543845415, + 0.0027570524252951145, + 0.0032063734252005816, + 0.0033454070799052715, + 0.02518600784242153, + 0.011422100476920605, + 0.03644504025578499, + 0.05784057080745697, + 0.013114212080836296, + 0.009582608938217163, + 0.01892252452671528, + 0.0164459440857172, + 0.004921755753457546, + 0.003932253457605839, + 0.008669784292578697, + 0.010303915478289127, + 0.011220842599868774, + 0.003578381147235632, + 0.008508777245879173, + 0.010206199251115322, + 0.004751619882881641, + 0.005561769939959049, + 0.007666273508220911, + 0.006496946793049574, + 0.007768967188894749, + 0.004521474242210388, + 0.00801765825599432, + 0.0077043380588293076, + 0.006664814427495003, + 0.012524540536105633, + 0.046898555010557175, + 0.011264164932072163, + 0.03518032655119896, + 0.009632546454668045, + 0.007006515748798847, + 0.011808802373707294, + 0.006158837582916021, + 0.008747352287173271, + 0.011280791833996773, + 0.01292191632091999, + 0.0053298682905733585, + 0.004487177357077599, + 0.01006957795470953, + 0.01034777145832777, + 0.007497336249798536, + 0.005944532807916403, + 0.00903474073857069, + 0.00598644744604826, + 0.006146285217255354, + 0.005978703964501619, + 0.003760010004043579, + 0.007259360048919916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.0009201106731779873 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot a dynamic tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing weathered steel plate armor with a full-face helmet visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a black warhorse with a tattered crimson cape galloping across golden grasslands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a massive orange-red sunset creating a golden backlight and dust mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the knight and horse shown with long shadows and a golden hue due to the sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019402741454541683 + ], + "opposite_logprob_tokens": [ + -6.2519402503967285 + ], + "logprob_reward": -0.0019402741454541683, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019402741454541683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09462366998195648, + 0.025731420144438744, + 0.037001680582761765, + 0.011943500488996506, + 0.32084131240844727, + 0.024605412036180496, + 0.06439653784036636, + 0.21824927628040314, + 0.036490101367235184, + 0.16611704230308533 + ], + "credit_map": [ + 0.02997167408466339, + 0.0034880174789577723, + 0.00748934131115675, + 0.005091054830700159, + 0.006116923876106739, + 0.030913682654500008, + 0.005630990490317345, + 0.005921989679336548, + 0.0034041772596538067, + 0.0017914471682161093, + 0.003723345696926117, + 0.002876515733078122, + 0.0028357564006000757, + 0.0039576697163283825, + 0.0014101805863901973, + 0.005732327699661255, + 0.004388313740491867, + 0.0019574109464883804, + 0.007751095574349165, + 0.008403249084949493, + 0.0031027847435325384, + 0.0022564223036170006, + 0.005514421965926886, + 0.0036279847845435143, + 0.0015855822712182999, + 0.0012484211474657059, + 0.001434776233509183, + 0.002142817247658968, + 0.0011209124932065606, + 0.0010150134330615401, + 0.0011942082783207297, + 0.002201769733801484, + 0.04417787864804268, + 0.01937253773212433, + 0.05595061182975769, + 0.111811064183712, + 0.01884741522371769, + 0.013148892670869827, + 0.028440244495868683, + 0.029092643409967422, + 0.003371805651113391, + 0.002411331282928586, + 0.0029517619404941797, + 0.004540544468909502, + 0.0035710656084120274, + 0.0013051837449893355, + 0.0033465521410107613, + 0.003107166849076748, + 0.00623556412756443, + 0.007636187598109245, + 0.008593179285526276, + 0.010682971216738224, + 0.007203245535492897, + 0.005432163830846548, + 0.00829148106276989, + 0.010321750305593014, + 0.009402717463672161, + 0.030757397413253784, + 0.06531065702438354, + 0.01277098711580038, + 0.07093612849712372, + 0.009900621138513088, + 0.007910731248557568, + 0.011260030791163445, + 0.006441615987569094, + 0.004694235976785421, + 0.0058057582937181, + 0.006293442100286484, + 0.003417069325223565, + 0.001970747485756874, + 0.004691047128289938, + 0.003176187165081501, + 0.028801102191209793, + 0.02346537634730339, + 0.024212844669818878, + 0.013673285022377968, + 0.014822762459516525, + 0.02206249162554741, + 0.011882763355970383, + 0.027196431532502174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.0019402741454541683 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_3/index_tensor([4620, 4620])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_3/index_tensor([4620, 4620])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b4b250b530fc2d04928db96190cb21b7d38dded8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_3/index_tensor([4620, 4620])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a fixed, bird's-eye view, the world is constructed as a vivid 8-bit pixel art canvas, accurately recreating the visual experience described by users. The crowded Urumqi Grand Bazaar is depicted as a maze of vibrantly colored blocks. Beneath a red, yellow, and green pixelated canopy lies a dense network of stalls piled high with clusters of colorful pixels representing fruits, carpets, and handicrafts. Densely packed with pixel figures, each composed of a few simple blocks of color, they converge into a perpetual stream of colorful people, flowing steadily through narrow aisles. Against this dynamic backdrop, a central 8-bit character, wearing a small, pixelated flower hat and a striking bright blue coat, stands out from the crowd. With a unique, frame-by-frame animation rhythm, he deftly weaves through the gaps in the flow, moving purposefully in one direction. The entire scene has no complicated light and shadow, presenting a flat and uniform bright lighting. The strong color contrast and hard pixel edges together create a retro game atmosphere that is both chaotic and vibrant, like a vivid game map being calculated in real time.", + "seed": 26157, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_3/index_4620_seed_26157_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14010802321135998, + "vlm_logprob": -0.14010802321135998, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.14010802321135998, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5289360284805298 + ], + "opposite_logprob_tokens": [ + -0.9039360284805298 + ], + "logprob_reward": -0.5289360284805298, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5289360284805298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3493245542049408, + 0.11308413743972778, + 0.2352941930294037, + 0.051248129457235336, + 0.04936546832323074, + 0.029163014143705368, + 0.03805534914135933, + 0.06106705963611603, + 0.032876525074243546, + 0.040521543473005295 + ], + "credit_map": [ + 0.035585347563028336, + 0.05271461233496666, + 0.02860512211918831, + 0.041900064796209335, + 0.06361530721187592, + 0.03841843083500862, + 0.034671589732170105, + 0.05381409451365471, + 0.010688011534512043, + 0.01390449795871973, + 0.012226088903844357, + 0.022672781720757484, + 0.017291337251663208, + 0.016477461904287338, + 0.010854840278625488, + 0.008969116024672985, + 0.04011280834674835, + 0.03571828082203865, + 0.015051591210067272, + 0.016001202166080475, + 0.06267392635345459, + 0.03153450787067413, + 0.018346572294831276, + 0.015855304896831512, + 0.00881149061024189, + 0.006572287064045668, + 0.006135002244263887, + 0.007180463057011366, + 0.005111338570713997, + 0.006617318838834763, + 0.0052049243822693825, + 0.0056153032928705215, + 0.007078845519572496, + 0.009741757065057755, + 0.00735620129853487, + 0.003938195761293173, + 0.006774715147912502, + 0.0055086566135287285, + 0.004974285140633583, + 0.0039928145706653595, + 0.003684192430227995, + 0.0031304156873375177, + 0.0037519559264183044, + 0.005035028327256441, + 0.0036124647594988346, + 0.004641660954803228, + 0.0028144267853349447, + 0.0024928695056587458, + 0.006213262677192688, + 0.0035474568139761686, + 0.006793486885726452, + 0.005328034050762653, + 0.0037700978573411703, + 0.004140106495469809, + 0.0030305555555969477, + 0.005232348572462797, + 0.007165889721363783, + 0.013163545168936253, + 0.0038561534602195024, + 0.006051165517419577, + 0.018214935436844826, + 0.004941651597619057, + 0.0033593010157346725, + 0.004314417950809002, + 0.0030532500240951777, + 0.0026911634486168623, + 0.006390733178704977, + 0.0031800945289433002, + 0.00445484509691596, + 0.003819385776296258, + 0.003548156237229705, + 0.0057388972491025925, + 0.004583480302244425, + 0.006582515314221382, + 0.010374946519732475, + 0.0038148444145917892, + 0.004203731659799814, + 0.0036825225688517094, + 0.0027578503359109163, + 0.004521651659160852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.5289360284805298 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208698645234108 + ], + "opposite_logprob_tokens": [ + -4.264208793640137 + ], + "logprob_reward": -0.014208698645234108, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208698645234108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21013008058071136, + 0.07748012244701385, + 0.2674284875392914, + 0.058061353862285614, + 0.060537487268447876, + 0.05980512499809265, + 0.06268340349197388, + 0.06509816646575928, + 0.04629376903176308, + 0.09248200058937073 + ], + "credit_map": [ + 0.019341791048645973, + 0.02607915736734867, + 0.019144879654049873, + 0.031878240406513214, + 0.02879519760608673, + 0.0333850122988224, + 0.025829443708062172, + 0.025676356628537178, + 0.00936049409210682, + 0.006701166275888681, + 0.007427687291055918, + 0.020258286967873573, + 0.013766519725322723, + 0.005375552456825972, + 0.006487272214144468, + 0.008103138767182827, + 0.047354333102703094, + 0.03884045407176018, + 0.017342854291200638, + 0.016912655904889107, + 0.07433968037366867, + 0.034654371440410614, + 0.020937366411089897, + 0.01704677939414978, + 0.006243805401027203, + 0.015582543797791004, + 0.006383372005075216, + 0.004714805632829666, + 0.004934555385261774, + 0.012810717336833477, + 0.0033077343832701445, + 0.0040838164277374744, + 0.006523734889924526, + 0.01027838233858347, + 0.00977995339781046, + 0.006549919955432415, + 0.005100947804749012, + 0.010838084854185581, + 0.004828632343560457, + 0.006637830287218094, + 0.006245133932679892, + 0.008155455812811852, + 0.004788914695382118, + 0.006871710065752268, + 0.0066309436224401, + 0.012934776954352856, + 0.00831136666238308, + 0.005866821855306625, + 0.007717680186033249, + 0.0056031192652881145, + 0.006815648637712002, + 0.008447140455245972, + 0.009028826840221882, + 0.009203951805830002, + 0.006977629382163286, + 0.008889404125511646, + 0.007102481555193663, + 0.010364263318479061, + 0.0066419863142073154, + 0.007748892065137625, + 0.018654361367225647, + 0.004549880512058735, + 0.003929854370653629, + 0.0061064474284648895, + 0.003576140385121107, + 0.004746047779917717, + 0.008032634854316711, + 0.005187207367271185, + 0.007918234914541245, + 0.0057595656253397465, + 0.004820914473384619, + 0.006253024097532034, + 0.010536367073655128, + 0.019045380875468254, + 0.031451620161533356, + 0.006824252661317587, + 0.0071113016456365585, + 0.006788911763578653, + 0.004872442223131657, + 0.005851725582033396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.014208698645234108 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896608859300613 + ], + "opposite_logprob_tokens": [ + -2.5789661407470703 + ], + "logprob_reward": -0.07896608859300613, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896608859300613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2251056283712387, + 0.1507721245288849, + 0.11958274990320206, + 0.05487890541553497, + 0.060448914766311646, + 0.13585524260997772, + 0.03381386399269104, + 0.03118523210287094, + 0.11499980837106705, + 0.0733574628829956 + ], + "credit_map": [ + 0.019031764939427376, + 0.038631949573755264, + 0.01894558034837246, + 0.029473602771759033, + 0.03358495980501175, + 0.04058989882469177, + 0.019616318866610527, + 0.02523157000541687, + 0.015938252210617065, + 0.012069630436599255, + 0.014013984240591526, + 0.02630886808037758, + 0.02985558658838272, + 0.01750173419713974, + 0.019182881340384483, + 0.015901202335953712, + 0.020129628479480743, + 0.01674409583210945, + 0.015093992464244366, + 0.012440442107617855, + 0.02138531394302845, + 0.011818304657936096, + 0.01094899047166109, + 0.01102199126034975, + 0.007039817050099373, + 0.008272322826087475, + 0.00598347932100296, + 0.005128726363182068, + 0.00388990412466228, + 0.014254121109843254, + 0.00415249215438962, + 0.006158042699098587, + 0.008589272387325764, + 0.009192782454192638, + 0.007651487365365028, + 0.00524688046425581, + 0.012214369140565395, + 0.009183619171380997, + 0.004369471222162247, + 0.0040010325610637665, + 0.016073215752840042, + 0.011983076110482216, + 0.00866440124809742, + 0.014688175171613693, + 0.016599029302597046, + 0.03009025752544403, + 0.02600695937871933, + 0.011750146746635437, + 0.005845517851412296, + 0.003196451347321272, + 0.004461050499230623, + 0.004291690420359373, + 0.004481730051338673, + 0.0028767872136086226, + 0.004294078331440687, + 0.004366559442132711, + 0.00554853118956089, + 0.003878219984471798, + 0.003302142955362797, + 0.004021728876978159, + 0.004228393547236919, + 0.002401308622211218, + 0.0036952276714146137, + 0.004109681583940983, + 0.009663010016083717, + 0.012217681854963303, + 0.0175638385117054, + 0.014816708862781525, + 0.014725166372954845, + 0.012326404452323914, + 0.01207278948277235, + 0.021614212542772293, + 0.007544158026576042, + 0.00936930812895298, + 0.012397744692862034, + 0.007938500493764877, + 0.007720593828707933, + 0.007298266980797052, + 0.007997225970029831, + 0.013091669417917728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.07896608859300613 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863468185067177 + ], + "opposite_logprob_tokens": [ + -3.0486347675323486 + ], + "logprob_reward": -0.04863468185067177, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863468185067177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24401646852493286, + 0.13979597389698029, + 0.10017754137516022, + 0.10311861336231232, + 0.0769178718328476, + 0.10357631742954254, + 0.06433612108230591, + 0.05298663675785065, + 0.03375885263085365, + 0.08131557703018188 + ], + "credit_map": [ + 0.02454240433871746, + 0.04789628088474274, + 0.022215768694877625, + 0.032970916479825974, + 0.030432963743805885, + 0.025936832651495934, + 0.0315546989440918, + 0.02846660651266575, + 0.016187945380806923, + 0.009477649815380573, + 0.012365700677037239, + 0.040316399186849594, + 0.022445380687713623, + 0.011680887080729008, + 0.012332325801253319, + 0.01498967781662941, + 0.016076015308499336, + 0.013839812017977238, + 0.013783873990178108, + 0.010786721482872963, + 0.01698918454349041, + 0.011069965548813343, + 0.008654458448290825, + 0.008977512829005718, + 0.010291175916790962, + 0.03511042520403862, + 0.006754015106707811, + 0.006119718309491873, + 0.007762619759887457, + 0.025729626417160034, + 0.0055978684686124325, + 0.005753168370574713, + 0.008361870422959328, + 0.012217001989483833, + 0.0067937932908535, + 0.009369054809212685, + 0.008456408977508545, + 0.011505824513733387, + 0.010950167663395405, + 0.009263748303055763, + 0.015266562812030315, + 0.007564458064734936, + 0.008103784173727036, + 0.01259071845561266, + 0.009899736382067204, + 0.02434113807976246, + 0.01606082357466221, + 0.009749099612236023, + 0.011096397414803505, + 0.0056570530869066715, + 0.01127459853887558, + 0.00947915855795145, + 0.007228551432490349, + 0.008463118225336075, + 0.004947259556502104, + 0.006189984735101461, + 0.008107905276119709, + 0.005754798650741577, + 0.00702453451231122, + 0.004567570518702269, + 0.011300073936581612, + 0.005583828780800104, + 0.0050639002583920956, + 0.005584024824202061, + 0.0031167345587164164, + 0.004020576365292072, + 0.0035991619806736708, + 0.003437043633311987, + 0.0066836061887443066, + 0.003007640363648534, + 0.005808042827993631, + 0.004086045082658529, + 0.010873855091631413, + 0.008911523967981339, + 0.014557179994881153, + 0.009658326394855976, + 0.010671272873878479, + 0.012649619951844215, + 0.005017466377466917, + 0.008976330980658531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.04863468185067177 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794618487358093 + ], + "opposite_logprob_tokens": [ + -3.529794692993164 + ], + "logprob_reward": -0.029794618487358093, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794618487358093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46854308247566223, + 0.08855487406253815, + 0.11747895181179047, + 0.04730767756700516, + 0.06384090334177017, + 0.05479796603322029, + 0.03563360497355461, + 0.025736810639500618, + 0.03393327072262764, + 0.06417293101549149 + ], + "credit_map": [ + 0.04297671094536781, + 0.09080483764410019, + 0.030912956222891808, + 0.050130847841501236, + 0.09359849989414215, + 0.059849001467227936, + 0.036405932158231735, + 0.06386428326368332, + 0.010246583260595798, + 0.008140679448843002, + 0.010922673158347607, + 0.02149328961968422, + 0.01451465580612421, + 0.006231830921024084, + 0.008033799938857555, + 0.008971361443400383, + 0.015237180516123772, + 0.019505459815263748, + 0.008653628639876842, + 0.007514788769185543, + 0.035728879272937775, + 0.012732941657304764, + 0.009192951023578644, + 0.008913119323551655, + 0.005592653527855873, + 0.006439917720854282, + 0.006789470091462135, + 0.005002855788916349, + 0.005862534511834383, + 0.0053252531215548515, + 0.0049469610676169395, + 0.0073480322025716305, + 0.01193924155086279, + 0.008364349603652954, + 0.0073659950867295265, + 0.0066159749403595924, + 0.007712958846241236, + 0.010254405438899994, + 0.004555676598101854, + 0.007032302208244801, + 0.009057986550033092, + 0.004576918203383684, + 0.0033363138791173697, + 0.006628438364714384, + 0.005815989337861538, + 0.015686288475990295, + 0.004755004309117794, + 0.004941022489219904, + 0.00513994786888361, + 0.0033393457997590303, + 0.005751970689743757, + 0.005063519347459078, + 0.0035754134878516197, + 0.0036719199270009995, + 0.0029005007818341255, + 0.006190991960465908, + 0.0034883730113506317, + 0.002461491385474801, + 0.0031913260463625193, + 0.0033570672385394573, + 0.004691381938755512, + 0.0031140237115323544, + 0.0029900658410042524, + 0.0024430821649730206, + 0.002691097790375352, + 0.003073369385674596, + 0.004512223415076733, + 0.003719167085364461, + 0.004138017538934946, + 0.0050267972983419895, + 0.0033973550889641047, + 0.007375241722911596, + 0.006897846236824989, + 0.01618322730064392, + 0.022673098370432854, + 0.004239038564264774, + 0.003552411450073123, + 0.004094330593943596, + 0.0032595046795904636, + 0.003273477777838707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.029794618487358093 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a fixed, bird's-eye view, the world is constructed as a vivid 8-bit pixel art canvas, accurately recreating the visual experience described by users. The crowded Urumqi Grand Bazaar is depicted as a maze of vibrantly colored blocks. Beneath a red, yellow, and green pixelated canopy lies a dense network of stalls piled high with clusters of colorful pixels representing fruits, carpets, and handicrafts. Densely packed with pixel figures, each composed of a few simple blocks of color, they converge into a perpetual stream of colorful people, flowing steadily through narrow aisles. Against this dynamic backdrop, a central 8-bit character, wearing a small, pixelated flower hat and a striking bright blue coat, stands out from the crowd. With a unique, frame-by-frame animation rhythm, he deftly weaves through the gaps in the flow, moving purposefully in one direction. The entire scene has no complicated light and shadow, presenting a flat and uniform bright lighting. The strong color contrast and hard pixel edges together create a retro game atmosphere that is both chaotic and vibrant, like a vivid game map being calculated in real time.", + "seed": 26158, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_3/index_4620_seed_26158_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05288498802110553, + "vlm_logprob": -0.05288498802110553, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.05288498802110553, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1642739474773407 + ], + "opposite_logprob_tokens": [ + -1.914273977279663 + ], + "logprob_reward": -0.1642739474773407, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1642739474773407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1471892148256302, + 0.11217152327299118, + 0.030092161148786545, + 0.0819273367524147, + 0.06730011105537415, + 0.1752084344625473, + 0.2510364055633545, + 0.03995028883218765, + 0.07392603904008865, + 0.02119855210185051 + ], + "credit_map": [ + 0.013193261809647083, + 0.01352906133979559, + 0.030764281749725342, + 0.019058823585510254, + 0.010077115148305893, + 0.011917274445295334, + 0.022354573011398315, + 0.026294799521565437, + 0.008486414328217506, + 0.012378722429275513, + 0.016412025317549706, + 0.020413393154740334, + 0.007380115799605846, + 0.011435003019869328, + 0.018179014325141907, + 0.017486831173300743, + 0.0022662130650132895, + 0.0021321880631148815, + 0.0036969210486859083, + 0.007610063534229994, + 0.003035716712474823, + 0.0027117852587252855, + 0.00402143644168973, + 0.004617835395038128, + 0.007547692861407995, + 0.009582022204995155, + 0.010317754931747913, + 0.0070336246863007545, + 0.010939637199044228, + 0.020669234916567802, + 0.007770617492496967, + 0.008066747337579727, + 0.014764349907636642, + 0.0056547923013567924, + 0.009207877330482006, + 0.005702903959900141, + 0.008068926632404327, + 0.011145741678774357, + 0.005516104865819216, + 0.007239412982016802, + 0.010562562383711338, + 0.02100769802927971, + 0.04290485382080078, + 0.026748379692435265, + 0.010935373604297638, + 0.024227997288107872, + 0.017206262797117233, + 0.0216152872890234, + 0.04840253293514252, + 0.016079610213637352, + 0.025013700127601624, + 0.0166630819439888, + 0.024982359260320663, + 0.04323519766330719, + 0.0192373339086771, + 0.05742257088422775, + 0.003064708551391959, + 0.0030689372215420008, + 0.010214347392320633, + 0.005648558493703604, + 0.0034204237163066864, + 0.003261350793763995, + 0.004690205678343773, + 0.0065817562863230705, + 0.012603101320564747, + 0.00752281304448843, + 0.014917144551873207, + 0.008050583302974701, + 0.005312989000231028, + 0.006444515660405159, + 0.007559449411928654, + 0.011515439487993717, + 0.0017083775019273162, + 0.0022420547902584076, + 0.0027930280193686485, + 0.004010064993053675, + 0.001893526641651988, + 0.0018554911948740482, + 0.003316994523629546, + 0.0033790122251957655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.1642739474773407 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009797697886824608 + ], + "opposite_logprob_tokens": [ + -4.6347975730896 + ], + "logprob_reward": -0.009797697886824608, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009797697886824608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08527940511703491, + 0.06022805720567703, + 0.01972125843167305, + 0.05483870953321457, + 0.06697677075862885, + 0.15231116116046906, + 0.21303820610046387, + 0.03790046647191048, + 0.2949438989162445, + 0.014762029983103275 + ], + "credit_map": [ + 0.00903040636330843, + 0.006020590662956238, + 0.018382931128144264, + 0.011279418133199215, + 0.008066883310675621, + 0.006218886934220791, + 0.01240870077162981, + 0.013871584087610245, + 0.004500783048570156, + 0.0057083917781710625, + 0.010137851350009441, + 0.00940607488155365, + 0.004526851233094931, + 0.006178620271384716, + 0.010462361387908459, + 0.009307126514613628, + 0.001602507778443396, + 0.002500206930562854, + 0.0021974919363856316, + 0.0029230171348899603, + 0.0035609309561550617, + 0.0030332952737808228, + 0.0017278683371841908, + 0.0021759397350251675, + 0.005037659779191017, + 0.006347121205180883, + 0.010596765205264091, + 0.008365913294255733, + 0.004861847497522831, + 0.006012489087879658, + 0.005381867289543152, + 0.008235047571361065, + 0.01493090484291315, + 0.006337842438369989, + 0.0076773446053266525, + 0.005664593540132046, + 0.007797433063387871, + 0.011158742010593414, + 0.005533583927899599, + 0.007876322604715824, + 0.01420045830309391, + 0.019847631454467773, + 0.021065209060907364, + 0.02227654494345188, + 0.012255016714334488, + 0.018238963559269905, + 0.02200319431722164, + 0.022424137219786644, + 0.04227110743522644, + 0.013657390139997005, + 0.02060062624514103, + 0.013438757508993149, + 0.021185463294386864, + 0.036333754658699036, + 0.016388829797506332, + 0.04916228726506233, + 0.0035176717210561037, + 0.0036704521626234055, + 0.007502182386815548, + 0.007254195399582386, + 0.003349734703078866, + 0.0025234976783394814, + 0.004698502365499735, + 0.0053842272609472275, + 0.04778362438082695, + 0.029212957248091698, + 0.04462120309472084, + 0.03132143244147301, + 0.03480446711182594, + 0.029420902952551842, + 0.04113643243908882, + 0.036642905324697495, + 0.0014086348237469792, + 0.0017483100527897477, + 0.0020890901796519756, + 0.002591374097391963, + 0.0013722082367166877, + 0.0018646296812221408, + 0.0017475130734965205, + 0.001940269605256617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.009797697886824608 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036170552484691143 + ], + "opposite_logprob_tokens": [ + -5.628617286682129 + ], + "logprob_reward": -0.0036170552484691143, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036170552484691143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12070409208536148, + 0.1822364181280136, + 0.08175112307071686, + 0.08745355159044266, + 0.06943080574274063, + 0.09072286635637283, + 0.03525829315185547, + 0.08146542310714722, + 0.21073682606220245, + 0.040240541100502014 + ], + "credit_map": [ + 0.007780070882290602, + 0.00983385182917118, + 0.03389115259051323, + 0.013186302967369556, + 0.0129519859328866, + 0.0073889633640646935, + 0.02122354321181774, + 0.014448225498199463, + 0.01416840497404337, + 0.018701989203691483, + 0.024690622463822365, + 0.03329950571060181, + 0.014392116107046604, + 0.021961025893688202, + 0.031216952949762344, + 0.023805800825357437, + 0.005276096519082785, + 0.006178753916174173, + 0.009912324137985706, + 0.012641053646802902, + 0.007542897015810013, + 0.006047028116881847, + 0.0173739455640316, + 0.01677902787923813, + 0.006700889673084021, + 0.010221575386822224, + 0.014380991458892822, + 0.008200480602681637, + 0.011080975644290447, + 0.017130309715867043, + 0.0073259915225207806, + 0.012412335723638535, + 0.005396605469286442, + 0.006449808366596699, + 0.007950209081172943, + 0.011944201774895191, + 0.008951498195528984, + 0.006810257211327553, + 0.014282360672950745, + 0.007645861711353064, + 0.006142891012132168, + 0.010296340100467205, + 0.018001511693000793, + 0.016872745007276535, + 0.005600932985544205, + 0.011978618800640106, + 0.010343937203288078, + 0.011485897935926914, + 0.0034854337573051453, + 0.003004171885550022, + 0.007394713815301657, + 0.004518517758697271, + 0.0026729160454124212, + 0.0038010291755199432, + 0.0043474710546433926, + 0.006034038960933685, + 0.011369145475327969, + 0.007794237695634365, + 0.011868692003190517, + 0.01299247331917286, + 0.0066557214595377445, + 0.01027434878051281, + 0.0092797065153718, + 0.011231095530092716, + 0.03203224018216133, + 0.022031553089618683, + 0.031169729307293892, + 0.022249527275562286, + 0.025188835337758064, + 0.02033570408821106, + 0.030861154198646545, + 0.026868078857660294, + 0.0036830962635576725, + 0.004844639915972948, + 0.005266530904918909, + 0.007404836360365152, + 0.0036237051244825125, + 0.004016163758933544, + 0.005737436469644308, + 0.005664134863764048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.0036170552484691143 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808112069964409 + ], + "opposite_logprob_tokens": [ + -3.288081169128418 + ], + "logprob_reward": -0.03808112069964409, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808112069964409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10630045086145401, + 0.3089786171913147, + 0.0739312693476677, + 0.046148963272571564, + 0.09877438098192215, + 0.1187969371676445, + 0.059134334325790405, + 0.09296402335166931, + 0.06300753355026245, + 0.031963467597961426 + ], + "credit_map": [ + 0.007660329341888428, + 0.010032224468886852, + 0.0160787682980299, + 0.016033094376325607, + 0.010160673409700394, + 0.0050355009734630585, + 0.020721951499581337, + 0.020577911287546158, + 0.027345508337020874, + 0.038493167608976364, + 0.03489942103624344, + 0.04878107085824013, + 0.024215253069996834, + 0.05046383664011955, + 0.04834596440196037, + 0.036434389650821686, + 0.0059224627912044525, + 0.00694957235828042, + 0.009527510963380337, + 0.014900372363626957, + 0.008854047395288944, + 0.008569009602069855, + 0.010423289611935616, + 0.00878500659018755, + 0.0040190862491726875, + 0.005908993072807789, + 0.006221135146915913, + 0.006318188738077879, + 0.0050549632869660854, + 0.008865770883858204, + 0.004617328755557537, + 0.005143497139215469, + 0.017779937013983727, + 0.008620603010058403, + 0.018767496570944786, + 0.009483191184699535, + 0.007939167320728302, + 0.014299261383712292, + 0.010367654263973236, + 0.01151706650853157, + 0.011652812361717224, + 0.015373463742434978, + 0.01814066246151924, + 0.0182231105864048, + 0.01052437536418438, + 0.0104670450091362, + 0.018565306439995766, + 0.01585015468299389, + 0.010672218166291714, + 0.0049057588912546635, + 0.008451956324279308, + 0.00463577825576067, + 0.006000243593007326, + 0.008164169266819954, + 0.0042086332105100155, + 0.012095572426915169, + 0.012987473048269749, + 0.01140544656664133, + 0.010782666504383087, + 0.01449323259294033, + 0.008401739411056042, + 0.012883254326879978, + 0.011206156574189663, + 0.010804055258631706, + 0.007126449607312679, + 0.00838860310614109, + 0.009294280782341957, + 0.00667810020968318, + 0.008709217421710491, + 0.007170026656240225, + 0.009419236332178116, + 0.006221619434654713, + 0.002490360289812088, + 0.0032116009388118982, + 0.004015945363789797, + 0.005354816559702158, + 0.0036841777618974447, + 0.0027486616745591164, + 0.004101268481463194, + 0.006356638390570879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.03808112069964409 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865511879324913 + ], + "opposite_logprob_tokens": [ + -3.0486550331115723 + ], + "logprob_reward": -0.04865511879324913, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865511879324913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17624999582767487, + 0.0676998421549797, + 0.027169037610292435, + 0.06915920227766037, + 0.06773378700017929, + 0.09467790275812149, + 0.40025800466537476, + 0.022042492404580116, + 0.05679916590452194, + 0.018210534006357193 + ], + "credit_map": [ + 0.017063317820429802, + 0.0207014549523592, + 0.027040662243962288, + 0.028942212462425232, + 0.02600768394768238, + 0.016181934624910355, + 0.019993042573332787, + 0.02031971514225006, + 0.007182291243225336, + 0.006383825093507767, + 0.010786496102809906, + 0.007841486483812332, + 0.007223343942314386, + 0.008749929256737232, + 0.008380627259612083, + 0.01115185022354126, + 0.0023761182092130184, + 0.0032251987140625715, + 0.0025532895233482122, + 0.0027951421216130257, + 0.0051590073853731155, + 0.003758855164051056, + 0.0035490375012159348, + 0.003752391319721937, + 0.005753911565989256, + 0.008624718524515629, + 0.008372708223760128, + 0.006291165482252836, + 0.00897511187940836, + 0.017127400264143944, + 0.006192502565681934, + 0.007821687497198582, + 0.0139961251989007, + 0.005973349791020155, + 0.007442315574735403, + 0.006319004576653242, + 0.008079449646174908, + 0.010473634116351604, + 0.006375267636030912, + 0.009074650704860687, + 0.0075082480907440186, + 0.011063975282013416, + 0.0242670476436615, + 0.009805471636354923, + 0.007277006283402443, + 0.01525322999805212, + 0.009306712076067924, + 0.010196222923696041, + 0.07461672276258469, + 0.027317846193909645, + 0.040025319904088974, + 0.026607384905219078, + 0.039250846952199936, + 0.07034480571746826, + 0.03205123543739319, + 0.09004388749599457, + 0.0026466345880180597, + 0.0013967077247798443, + 0.004104929976165295, + 0.003486949484795332, + 0.0016869660466909409, + 0.002341044368222356, + 0.003265781793743372, + 0.00311348307877779, + 0.006587089970707893, + 0.005113303195685148, + 0.008019346743822098, + 0.007893831469118595, + 0.0066064572893083096, + 0.0062867687083780766, + 0.009193501435220242, + 0.007098871283233166, + 0.0014784674858674407, + 0.001722060376778245, + 0.0026505044661462307, + 0.003293032990768552, + 0.001547837513498962, + 0.001713923062197864, + 0.003285517916083336, + 0.002519192872568965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.04865511879324913 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a fixed, bird's-eye view, the world is constructed as a vivid 8-bit pixel art canvas, accurately recreating the visual experience described by users. The crowded Urumqi Grand Bazaar is depicted as a maze of vibrantly colored blocks. Beneath a red, yellow, and green pixelated canopy lies a dense network of stalls piled high with clusters of colorful pixels representing fruits, carpets, and handicrafts. Densely packed with pixel figures, each composed of a few simple blocks of color, they converge into a perpetual stream of colorful people, flowing steadily through narrow aisles. Against this dynamic backdrop, a central 8-bit character, wearing a small, pixelated flower hat and a striking bright blue coat, stands out from the crowd. With a unique, frame-by-frame animation rhythm, he deftly weaves through the gaps in the flow, moving purposefully in one direction. The entire scene has no complicated light and shadow, presenting a flat and uniform bright lighting. The strong color contrast and hard pixel edges together create a retro game atmosphere that is both chaotic and vibrant, like a vivid game map being calculated in real time.", + "seed": 26159, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_3/index_4620_seed_26159_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09017356662079692, + "vlm_logprob": -0.09017356662079692, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.09017356662079692, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35622501373291016 + ], + "opposite_logprob_tokens": [ + -1.2312250137329102 + ], + "logprob_reward": -0.35622501373291016, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35622501373291016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12705715000629425, + 0.04379688948392868, + 0.060267187654972076, + 0.11968804150819778, + 0.2695247232913971, + 0.15160413086414337, + 0.04993906617164612, + 0.06957890093326569, + 0.039225246757268906, + 0.06931861490011215 + ], + "credit_map": [ + 0.013489014469087124, + 0.021371565759181976, + 0.017780236899852753, + 0.014124758541584015, + 0.011745614930987358, + 0.01139884814620018, + 0.015342796221375465, + 0.02180430479347706, + 0.003045501187443733, + 0.0034422536846250296, + 0.005764537490904331, + 0.007183842360973358, + 0.006984103471040726, + 0.004897423554211855, + 0.006402789615094662, + 0.006076440215110779, + 0.007898318581283092, + 0.004062546417117119, + 0.011045818217098713, + 0.00867654662579298, + 0.009706353768706322, + 0.005689555779099464, + 0.006771535146981478, + 0.00641650939360261, + 0.02680467814207077, + 0.007959638722240925, + 0.018234407529234886, + 0.015835054218769073, + 0.006310984026640654, + 0.005916048772633076, + 0.009764987044036388, + 0.028862236067652702, + 0.020135613158345222, + 0.023500634357333183, + 0.040373362600803375, + 0.0876164510846138, + 0.013634239323437214, + 0.03436845913529396, + 0.026889683678746223, + 0.023006273433566093, + 0.007449930999428034, + 0.032721325755119324, + 0.017196327447891235, + 0.04667932167649269, + 0.009678035974502563, + 0.010542517527937889, + 0.01886921003460884, + 0.008467461913824081, + 0.004345608875155449, + 0.00392645038664341, + 0.007919244468212128, + 0.013995002023875713, + 0.004988175816833973, + 0.003554051509127021, + 0.006249330006539822, + 0.0049612028524279594, + 0.0061490233056247234, + 0.006092533934861422, + 0.011988930404186249, + 0.011500679887831211, + 0.005006211344152689, + 0.007744763512164354, + 0.00989094004034996, + 0.011205820366740227, + 0.006213454529643059, + 0.004689149558544159, + 0.0057105012238025665, + 0.004465446807444096, + 0.004133608657866716, + 0.004438048228621483, + 0.0049746399745345116, + 0.00460039870813489, + 0.010819937102496624, + 0.01012923289090395, + 0.007706529926508665, + 0.009110176004469395, + 0.008897091262042522, + 0.007620908319950104, + 0.00817456841468811, + 0.006860170979052782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.35622501373291016 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005266483407467604 + ], + "opposite_logprob_tokens": [ + -5.2552666664123535 + ], + "logprob_reward": -0.005266483407467604, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005266483407467604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09173040091991425, + 0.051887109875679016, + 0.03719298169016838, + 0.15281198918819427, + 0.3985896110534668, + 0.09505683928728104, + 0.09883967787027359, + 0.03947266936302185, + 0.014257973991334438, + 0.020160725340247154 + ], + "credit_map": [ + 0.011027144268155098, + 0.011291238479316235, + 0.014919840730726719, + 0.009781597182154655, + 0.005467935465276241, + 0.007246778346598148, + 0.020353062078356743, + 0.01164280530065298, + 0.0057141538709402084, + 0.004566442687064409, + 0.006796896457672119, + 0.00713465316221118, + 0.007682448253035545, + 0.0054883998818695545, + 0.008291292004287243, + 0.006212822627276182, + 0.005358366295695305, + 0.0034656154457479715, + 0.005438187159597874, + 0.0058538950979709625, + 0.0034372101072221994, + 0.0030335995834320784, + 0.0056476532481610775, + 0.004958452191203833, + 0.02752317115664482, + 0.014947992749512196, + 0.01775154285132885, + 0.01735193096101284, + 0.010069973766803741, + 0.00879637710750103, + 0.013225754722952843, + 0.04314525052905083, + 0.03115685097873211, + 0.037104517221450806, + 0.05816873908042908, + 0.12538012862205505, + 0.021443719044327736, + 0.05509769171476364, + 0.03815094381570816, + 0.03208703175187111, + 0.004790147300809622, + 0.0062934355810284615, + 0.036530982702970505, + 0.029839763417840004, + 0.00350528908893466, + 0.0038402914069592953, + 0.006892061792314053, + 0.0033648719545453787, + 0.010264219716191292, + 0.009090305306017399, + 0.01670529879629612, + 0.018640125170350075, + 0.013920342549681664, + 0.00851657148450613, + 0.008401045575737953, + 0.013301766477525234, + 0.003153415396809578, + 0.004043294116854668, + 0.005681559443473816, + 0.005010390654206276, + 0.006774668116122484, + 0.005257833749055862, + 0.005057925824075937, + 0.004493579268455505, + 0.0016894431319087744, + 0.0013493176084011793, + 0.0020599165000021458, + 0.0018780454993247986, + 0.001698582200333476, + 0.0015268573770299554, + 0.0022290246561169624, + 0.0018267881823703647, + 0.0022954612504690886, + 0.004242550116032362, + 0.0024451464414596558, + 0.002192231360822916, + 0.0020413403399288654, + 0.0028287763707339764, + 0.0022107840050011873, + 0.0019044374348595738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.005266483407467604 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794966802001 + ], + "opposite_logprob_tokens": [ + -3.529794931411743 + ], + "logprob_reward": -0.029794966802001, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794966802001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10137779265642166, + 0.1381937861442566, + 0.12952691316604614, + 0.10631278902292252, + 0.08720995485782623, + 0.14210107922554016, + 0.10324618220329285, + 0.1305556297302246, + 0.020957691594958305, + 0.040518131107091904 + ], + "credit_map": [ + 0.009456085041165352, + 0.014977853745222092, + 0.010468867607414722, + 0.010521686635911465, + 0.004978399258106947, + 0.007045609410852194, + 0.03128368780016899, + 0.012645604088902473, + 0.009155809879302979, + 0.010822800919413567, + 0.018481943756341934, + 0.022170783951878548, + 0.018016686663031578, + 0.008844872936606407, + 0.03312157467007637, + 0.01757930964231491, + 0.023692285642027855, + 0.0398910716176033, + 0.01937616989016533, + 0.014455677010118961, + 0.007224267348647118, + 0.006089414935559034, + 0.008338513784110546, + 0.010459513403475285, + 0.021747056394815445, + 0.010787525214254856, + 0.008559632115066051, + 0.008800976909697056, + 0.006584387272596359, + 0.005890482570976019, + 0.009724102914333344, + 0.03421862795948982, + 0.008299488574266434, + 0.010819178074598312, + 0.012471511028707027, + 0.023688428103923798, + 0.005651358515024185, + 0.011018593795597553, + 0.0076211062259972095, + 0.00764028774574399, + 0.008804179728031158, + 0.04027894511818886, + 0.03669596463441849, + 0.015226413495838642, + 0.011368251405656338, + 0.007332834415137768, + 0.014595874585211277, + 0.007798617705702782, + 0.010445886291563511, + 0.011072300374507904, + 0.014169215224683285, + 0.015229593962430954, + 0.021439099684357643, + 0.009670314379036427, + 0.007978350855410099, + 0.013241416774690151, + 0.00701659731566906, + 0.01322478149086237, + 0.017943348735570908, + 0.020716125145554543, + 0.014765634201467037, + 0.012408060021698475, + 0.03023718111217022, + 0.01424390822649002, + 0.0023120774421840906, + 0.0024438549298793077, + 0.0024611016269773245, + 0.0027782581746578217, + 0.0029083406552672386, + 0.002808315446600318, + 0.0027550337836146355, + 0.0024907090701162815, + 0.002821042202413082, + 0.004436948802322149, + 0.006665806286036968, + 0.005633365362882614, + 0.005154365673661232, + 0.003586243372410536, + 0.007454706821590662, + 0.004765649791806936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.029794966802001 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978570945560932 + ], + "opposite_logprob_tokens": [ + -3.529785633087158 + ], + "logprob_reward": -0.02978570945560932, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978570945560932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1728244125843048, + 0.08535302430391312, + 0.05866619199514389, + 0.11790817230939865, + 0.029620924964547157, + 0.12510675191879272, + 0.1722954362630844, + 0.06965386122465134, + 0.042460303753614426, + 0.1261109560728073 + ], + "credit_map": [ + 0.014413373544812202, + 0.042210448533296585, + 0.018279442563652992, + 0.015240984968841076, + 0.009812313131988049, + 0.010183319449424744, + 0.03665824234485626, + 0.0260262880474329, + 0.005294276401400566, + 0.005704604089260101, + 0.017110316082835197, + 0.013368220068514347, + 0.009251514449715614, + 0.006644413806498051, + 0.01733569987118244, + 0.0106439758092165, + 0.005644695367664099, + 0.00734739750623703, + 0.01259677205234766, + 0.005358261056244373, + 0.008430239744484425, + 0.005668992176651955, + 0.006231389939785004, + 0.0073884460143744946, + 0.011550047434866428, + 0.029261456802487373, + 0.010672209784388542, + 0.009816121309995651, + 0.006877881474792957, + 0.0074529582634568214, + 0.009545848704874516, + 0.03273164853453636, + 0.002317247446626425, + 0.003068463411182165, + 0.0036417494993656874, + 0.007192610763013363, + 0.002647507470101118, + 0.0037312202621251345, + 0.0045454008504748344, + 0.002476724097505212, + 0.005549051333218813, + 0.018562348559498787, + 0.02908167988061905, + 0.04436078295111656, + 0.00677987327799201, + 0.0054086497984826565, + 0.01071002148091793, + 0.004654340445995331, + 0.016084779053926468, + 0.015401262789964676, + 0.028559692203998566, + 0.039053965359926224, + 0.02678738720715046, + 0.012920246459543705, + 0.012789719738066196, + 0.02069837972521782, + 0.0053749908693134785, + 0.006051806267350912, + 0.009267511777579784, + 0.012198304757475853, + 0.007714553270488977, + 0.00477044889703393, + 0.015480738133192062, + 0.008795511908829212, + 0.004907853435724974, + 0.004604666959494352, + 0.005405995063483715, + 0.004350222181528807, + 0.006029224488884211, + 0.005062298383563757, + 0.0068027907982468605, + 0.0052972533740103245, + 0.016280751675367355, + 0.0182355847209692, + 0.02137926034629345, + 0.015617398545145988, + 0.013031388632953167, + 0.01572510227560997, + 0.014405252411961555, + 0.011436228640377522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.02978570945560932 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795659705996513 + ], + "opposite_logprob_tokens": [ + -3.5297956466674805 + ], + "logprob_reward": -0.029795659705996513, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795659705996513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1599082350730896, + 0.04812807962298393, + 0.08920654654502869, + 0.21589897572994232, + 0.1784406453371048, + 0.1542789191007614, + 0.05432600528001785, + 0.035613905638456345, + 0.028327103704214096, + 0.035871610045433044 + ], + "credit_map": [ + 0.02273937128484249, + 0.028369449079036713, + 0.021936679258942604, + 0.015521119348704815, + 0.013738403096795082, + 0.013468078337609768, + 0.020838936790823936, + 0.023296182975172997, + 0.004535221494734287, + 0.004196100868284702, + 0.006346529815346003, + 0.006853228900581598, + 0.007101231254637241, + 0.005793444346636534, + 0.006948110647499561, + 0.006354205310344696, + 0.01588713936507702, + 0.008561618626117706, + 0.017054669559001923, + 0.010655906051397324, + 0.010370523668825626, + 0.006497728172689676, + 0.010294085368514061, + 0.009884875267744064, + 0.04033881798386574, + 0.021927837282419205, + 0.022861972451210022, + 0.022235488519072533, + 0.012280185706913471, + 0.011074751615524292, + 0.01813817396759987, + 0.06704172492027283, + 0.014063776470720768, + 0.017761902883648872, + 0.027390342205762863, + 0.05439462140202522, + 0.010074612684547901, + 0.022863876074552536, + 0.017243895679712296, + 0.014647622592747211, + 0.009063968434929848, + 0.029629580676555634, + 0.05575979873538017, + 0.023438138887286186, + 0.00935274176299572, + 0.006106242071837187, + 0.014120296575129032, + 0.006808136589825153, + 0.004205486737191677, + 0.006020525004714727, + 0.007872224785387516, + 0.011668282561004162, + 0.007951641455292702, + 0.004744604229927063, + 0.0049848104827106, + 0.006878424435853958, + 0.0025516028981655836, + 0.005178190302103758, + 0.00502825528383255, + 0.0051903086714446545, + 0.004073087591677904, + 0.0038795555010437965, + 0.005049009807407856, + 0.004663893487304449, + 0.004085178952664137, + 0.0025856520514935255, + 0.004130301531404257, + 0.00348386587575078, + 0.0036561491433531046, + 0.0031969170086085796, + 0.00419622054323554, + 0.0029928151052445173, + 0.0033854760695248842, + 0.0040618726052343845, + 0.004748497158288956, + 0.004140839446336031, + 0.004759700503200293, + 0.005259523633867502, + 0.004902472719550133, + 0.004613226279616356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.029795659705996513 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a fixed, bird's-eye view, the world is constructed as a vivid 8-bit pixel art canvas, accurately recreating the visual experience described by users. The crowded Urumqi Grand Bazaar is depicted as a maze of vibrantly colored blocks. Beneath a red, yellow, and green pixelated canopy lies a dense network of stalls piled high with clusters of colorful pixels representing fruits, carpets, and handicrafts. Densely packed with pixel figures, each composed of a few simple blocks of color, they converge into a perpetual stream of colorful people, flowing steadily through narrow aisles. Against this dynamic backdrop, a central 8-bit character, wearing a small, pixelated flower hat and a striking bright blue coat, stands out from the crowd. With a unique, frame-by-frame animation rhythm, he deftly weaves through the gaps in the flow, moving purposefully in one direction. The entire scene has no complicated light and shadow, presenting a flat and uniform bright lighting. The strong color contrast and hard pixel edges together create a retro game atmosphere that is both chaotic and vibrant, like a vivid game map being calculated in real time.", + "seed": 26160, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_3/index_4620_seed_26160_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.014011135953478516, + "vlm_logprob": -0.014011135953478516, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.014011135953478516, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03442753851413727 + ], + "opposite_logprob_tokens": [ + -3.5344276428222656 + ], + "logprob_reward": -0.03442753851413727, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03442753851413727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2224932312965393, + 0.03269293159246445, + 0.0969243124127388, + 0.01761510781943798, + 0.11244864761829376, + 0.14448358118534088, + 0.0407419353723526, + 0.07300807535648346, + 0.20594455301761627, + 0.053647641092538834 + ], + "credit_map": [ + 0.021992115303874016, + 0.032286107540130615, + 0.047229792922735214, + 0.03198689967393875, + 0.01485493965446949, + 0.01865299604833126, + 0.020164282992482185, + 0.035326067358255386, + 0.004161390475928783, + 0.0029616665560752153, + 0.006536063738167286, + 0.004311354830861092, + 0.00171559932641685, + 0.005104603245854378, + 0.0024587628431618214, + 0.005443489644676447, + 0.007801428437232971, + 0.010203017853200436, + 0.021231437101960182, + 0.016345391049981117, + 0.004840678069740534, + 0.013940581120550632, + 0.009667563252151012, + 0.012894202955067158, + 0.002332323230803013, + 0.002204200718551874, + 0.0027774169575423002, + 0.0033948307391256094, + 0.0012136056320741773, + 0.0018819376127794385, + 0.0016554773319512606, + 0.002155314665287733, + 0.00836018193513155, + 0.016095442697405815, + 0.011363837867975235, + 0.011384772136807442, + 0.013322293758392334, + 0.01784682273864746, + 0.021291173994541168, + 0.012784110382199287, + 0.012289050966501236, + 0.015632206574082375, + 0.01207005511969328, + 0.028584040701389313, + 0.0234703179448843, + 0.022994114086031914, + 0.015024674125015736, + 0.014419102109968662, + 0.004692106507718563, + 0.003410671604797244, + 0.004147606436163187, + 0.004671890288591385, + 0.00661633163690567, + 0.006493789609521627, + 0.005595638882368803, + 0.005113894585520029, + 0.006725650280714035, + 0.0051248143427073956, + 0.007823100313544273, + 0.013416455127298832, + 0.006051244214177132, + 0.006903509609401226, + 0.018499750643968582, + 0.0084635429084301, + 0.013321728445589542, + 0.021184375509619713, + 0.01749788224697113, + 0.02796812541782856, + 0.023093268275260925, + 0.059995412826538086, + 0.017984755337238312, + 0.024898994714021683, + 0.004597134422510862, + 0.004147982690483332, + 0.010828251019120216, + 0.008224980905652046, + 0.004115174990147352, + 0.005159296095371246, + 0.0068957870826125145, + 0.009679026901721954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.03442753851413727 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011880489764735103 + ], + "opposite_logprob_tokens": [ + -6.751188278198242 + ], + "logprob_reward": -0.0011880489764735103, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011880489764735103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04535719007253647, + 0.04891429841518402, + 0.057058751583099365, + 0.020272141322493553, + 0.060055654495954514, + 0.5753971338272095, + 0.06589268147945404, + 0.02992253564298153, + 0.07909440994262695, + 0.018035169690847397 + ], + "credit_map": [ + 0.005119193810969591, + 0.0040733348578214645, + 0.00827621016651392, + 0.006488179787993431, + 0.003525904845446348, + 0.0068361796438694, + 0.004778997972607613, + 0.006259189452975988, + 0.003691765945404768, + 0.006262654438614845, + 0.008405071683228016, + 0.006973639130592346, + 0.007856070064008236, + 0.005319373682141304, + 0.004140292759984732, + 0.006265431176871061, + 0.009727598167955875, + 0.013940317556262016, + 0.00629878044128418, + 0.007219416089355946, + 0.0035746523644775152, + 0.007434125989675522, + 0.004151643719524145, + 0.004712216556072235, + 0.001722691347822547, + 0.0013542354572564363, + 0.003227504203096032, + 0.0029099772218614817, + 0.0029827679973095655, + 0.004137735348194838, + 0.0019327717600390315, + 0.002004457637667656, + 0.00931751448661089, + 0.013214915059506893, + 0.005650303792208433, + 0.007115575019270182, + 0.004720333963632584, + 0.006659976672381163, + 0.007564077153801918, + 0.0058129578828811646, + 0.05003637075424194, + 0.09913314133882523, + 0.02692054584622383, + 0.06317233294248581, + 0.15630337595939636, + 0.07824259251356125, + 0.0641675814986229, + 0.03742120414972305, + 0.00562812527641654, + 0.006361731793731451, + 0.006335092708468437, + 0.007738027721643448, + 0.005170856136828661, + 0.012969025410711765, + 0.0079877320677042, + 0.013702087104320526, + 0.0036476869136095047, + 0.004052876494824886, + 0.0023029630538076162, + 0.006175729911774397, + 0.0022132229059934616, + 0.0027712564915418625, + 0.006169897969812155, + 0.002588903298601508, + 0.006805521436035633, + 0.006888084579259157, + 0.0037678221706300974, + 0.011156865395605564, + 0.00879739411175251, + 0.02709355391561985, + 0.004000519402325153, + 0.010584651492536068, + 0.0024113012477755547, + 0.001454311772249639, + 0.0031653698533773422, + 0.0027138597797602415, + 0.0011818492785096169, + 0.0028083084616810083, + 0.0015584348002448678, + 0.002741734031587839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.0011880489764735103 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040937201119959354 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.0040937201119959354, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040937201119959354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07254461199045181, + 0.029475970193743706, + 0.15036888420581818, + 0.029737701639533043, + 0.029578713700175285, + 0.4699503183364868, + 0.014923837967216969, + 0.03170254826545715, + 0.15612849593162537, + 0.015588940121233463 + ], + "credit_map": [ + 0.009409002028405666, + 0.006890751887112856, + 0.014927994459867477, + 0.010282214730978012, + 0.005922403652220964, + 0.006642069201916456, + 0.006478443741798401, + 0.01199173741042614, + 0.0031533262226730585, + 0.0017180831637233496, + 0.0038209815975278616, + 0.005068502854555845, + 0.0056048305705189705, + 0.003745654597878456, + 0.0031845367047935724, + 0.00318005564622581, + 0.013925003819167614, + 0.01456855796277523, + 0.016259467229247093, + 0.039813511073589325, + 0.010383324697613716, + 0.019749730825424194, + 0.01986563205718994, + 0.01580364815890789, + 0.0033495842944830656, + 0.0024203371722251177, + 0.002709828782826662, + 0.005071448162198067, + 0.005601454060524702, + 0.0039585428312420845, + 0.003624536329880357, + 0.0030019707046449184, + 0.002616730984300375, + 0.002249254612252116, + 0.0036171888932585716, + 0.0033151546958833933, + 0.0032491805031895638, + 0.00527307391166687, + 0.005932221654802561, + 0.0033259086776524782, + 0.037495411932468414, + 0.09266511350870132, + 0.02326042577624321, + 0.0478825680911541, + 0.1269916594028473, + 0.053721167147159576, + 0.056034546345472336, + 0.03189942240715027, + 0.001401175162754953, + 0.0012977832229807973, + 0.0013464841758832335, + 0.001581166055984795, + 0.002196826273575425, + 0.0027511920779943466, + 0.002266193274408579, + 0.0020830181892961264, + 0.0022925245575606823, + 0.0021908320486545563, + 0.0031726243905723095, + 0.005500882863998413, + 0.003467296250164509, + 0.0035757797304540873, + 0.008265268057584763, + 0.0032373424619436264, + 0.015454789623618126, + 0.011286068707704544, + 0.009985266253352165, + 0.02130153775215149, + 0.018928835168480873, + 0.044079992920160294, + 0.015819966793060303, + 0.01927204243838787, + 0.0020385009702295065, + 0.002107304986566305, + 0.0022728329058736563, + 0.0022438846062868834, + 0.0010866213124245405, + 0.0017627180786803365, + 0.0017554019577801228, + 0.0023216751869767904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.0040937201119959354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020570512861013412 + ], + "opposite_logprob_tokens": [ + -3.8955705165863037 + ], + "logprob_reward": -0.020570512861013412, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020570512861013412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.135188028216362, + 0.056374963372945786, + 0.16010156273841858, + 0.05453145503997803, + 0.05385430529713631, + 0.1671837419271469, + 0.02605276368558407, + 0.027017436921596527, + 0.22063271701335907, + 0.0990629717707634 + ], + "credit_map": [ + 0.013180604204535484, + 0.01348918303847313, + 0.024630265310406685, + 0.019210247322916985, + 0.015995141118764877, + 0.012944596819579601, + 0.013697035610675812, + 0.02204095758497715, + 0.005583461374044418, + 0.0031734041403979063, + 0.011539287865161896, + 0.008958914317190647, + 0.007726815529167652, + 0.007058932911604643, + 0.005256118718534708, + 0.007078027818351984, + 0.011333002708852291, + 0.01585833914577961, + 0.017630290240049362, + 0.04670816659927368, + 0.01028536818921566, + 0.02110700123012066, + 0.017600102350115776, + 0.019579298794269562, + 0.0063794925808906555, + 0.004689962137490511, + 0.004856558050960302, + 0.004902306944131851, + 0.013017578981816769, + 0.007876311428844929, + 0.008184609934687614, + 0.004624638706445694, + 0.0054784961976110935, + 0.0051797255873680115, + 0.005342785269021988, + 0.00611977931112051, + 0.004882081411778927, + 0.008514207787811756, + 0.011808064766228199, + 0.006529164966195822, + 0.014992183074355125, + 0.025685368105769157, + 0.015309234149754047, + 0.01821381039917469, + 0.03885142505168915, + 0.01734827645123005, + 0.01948627643287182, + 0.01729716919362545, + 0.003875990863889456, + 0.0027288119308650494, + 0.0027585490606725216, + 0.0032197018153965473, + 0.0030513328965753317, + 0.004151517059653997, + 0.003460859414190054, + 0.002806000178679824, + 0.0021870320197194815, + 0.0023093828931450844, + 0.004043784458190203, + 0.007070071529597044, + 0.00234044436365366, + 0.002618246478959918, + 0.0030905401799827814, + 0.00335793592967093, + 0.018737832084298134, + 0.013323108665645123, + 0.010909805074334145, + 0.03181111440062523, + 0.02860497683286667, + 0.0705425962805748, + 0.01857956312596798, + 0.02812371216714382, + 0.008894042111933231, + 0.009313877671957016, + 0.02024621143937111, + 0.016277072951197624, + 0.006043076515197754, + 0.008799283765256405, + 0.011498942971229553, + 0.017990458756685257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.020570512861013412 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a fixed bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a vivid 8-bit pixel art style with hard pixel edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a central character wearing a small pixelated flower hat and a bright blue coat visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the central character weaving through a crowd of pixel figures in a narrow aisle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting flat and uniform without complicated light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977585930377245 + ], + "opposite_logprob_tokens": [ + -4.634775638580322 + ], + "logprob_reward": -0.00977585930377245, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977585930377245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08184250444173813, + 0.02724163606762886, + 0.22029709815979004, + 0.02130706049501896, + 0.059860195964574814, + 0.3069390654563904, + 0.03817952051758766, + 0.0254217479377985, + 0.1902385652065277, + 0.028672631829977036 + ], + "credit_map": [ + 0.006744582671672106, + 0.015098176896572113, + 0.01273413747549057, + 0.011447048746049404, + 0.008004114963114262, + 0.008421865291893482, + 0.009148734621703625, + 0.010243835859000683, + 0.002705660182982683, + 0.0013544425601139665, + 0.005298299249261618, + 0.004632845055311918, + 0.001731186406686902, + 0.002962816972285509, + 0.0027668860275298357, + 0.005789496004581451, + 0.035214614123106, + 0.052940960973501205, + 0.01543786283582449, + 0.02795267477631569, + 0.014830658212304115, + 0.02754727005958557, + 0.02553611621260643, + 0.020836923271417618, + 0.0031021847389638424, + 0.002586368005722761, + 0.004262086004018784, + 0.0027920168358832598, + 0.001565589802339673, + 0.003033712739124894, + 0.0020353214349597692, + 0.00192977802362293, + 0.0067332107573747635, + 0.01353931799530983, + 0.00663046445697546, + 0.0073434170335531235, + 0.007061026990413666, + 0.006393376737833023, + 0.007286830805242062, + 0.004872544668614864, + 0.026169493794441223, + 0.05014278367161751, + 0.017002489417791367, + 0.029446745291352272, + 0.08662950992584229, + 0.03797559440135956, + 0.037242330610752106, + 0.02233009599149227, + 0.0038689044304192066, + 0.003648775862529874, + 0.003642319468781352, + 0.004646035842597485, + 0.004479975439608097, + 0.00754400622099638, + 0.0046623810194432735, + 0.005687118507921696, + 0.002210778882727027, + 0.002696865238249302, + 0.0029485258273780346, + 0.004673567600548267, + 0.0029785705264657736, + 0.0024143930058926344, + 0.004585585091263056, + 0.0029134603682905436, + 0.01662495732307434, + 0.012938739731907845, + 0.00732372235506773, + 0.028173353523015976, + 0.021325694397091866, + 0.06696006655693054, + 0.010488823987543583, + 0.02640317939221859, + 0.0038064850959926844, + 0.0028977736365050077, + 0.004928018432110548, + 0.0046168300323188305, + 0.002004551235586405, + 0.0031750043854117393, + 0.0028028148226439953, + 0.004441150464117527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 8 + ], + "selected_reward": -0.00977585930377245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_4/index_tensor([2081, 2081])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_4/index_tensor([2081, 2081])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c65bbfb05fe20b492f5de03c5764c4962579ea10 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_4/index_tensor([2081, 2081])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fluid, uninterrupted long shot begins with a low-angle tracking shot, gliding smoothly and deliberately across the phosphorescent moss floor of a magical forest bathed in eternal twilight. Ancient, gnarled trees with bark like iridescent obsidian tower over an invisible canopy, their massive branches swaying rhythmically, not in sync with the wind but as if conducting a silent, grand orchestra. Their leaves, seemingly made of thin, translucent crystal, veined with liquid gold, vibrate not with a hiss but with a tangible vibration, emitting a soft, pulsating cyan and purplish-red glow that ripples through the air like visible sound waves. Below, clusters of low-growing shrubs join the ethereal chorus; their enormous, bell-shaped flowers, radiating an inner white light, pulsate in synchronized harmony, their petals gently opening and closing, releasing sparkling points of light into the gentle, luminous mist woven between the trunks. The entire scene is illuminated from within by this omnipresent bioluminescence, creating a cool, mysterious palette of deep indigos and emerald greens, casting long dancing shadows and creating a profound atmosphere of serene, otherworldly wonder as the camera continues its steady, dreamlike progress through the living symphony.", + "seed": 29447, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_4/index_2081_seed_29447_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1320307595655322, + "vlm_logprob": -0.1320307595655322, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.1320307595655322, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0625808835029602 + ], + "opposite_logprob_tokens": [ + -2.9375808238983154 + ], + "logprob_reward": -0.0625808835029602, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0625808835029602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1061074286699295, + 0.11528471112251282, + 0.05094830319285393, + 0.1473858505487442, + 0.16365963220596313, + 0.10669811069965363, + 0.05251917988061905, + 0.07573916018009186, + 0.07288450002670288, + 0.10877314954996109 + ], + "credit_map": [ + 0.015278168022632599, + 0.015378612093627453, + 0.013239051215350628, + 0.010943246074020863, + 0.012845080345869064, + 0.013060882687568665, + 0.011885593645274639, + 0.013476789928972721, + 0.014849952422082424, + 0.01818712428212166, + 0.006365858018398285, + 0.009901607409119606, + 0.011726027354598045, + 0.01709689572453499, + 0.01185185182839632, + 0.02530539222061634, + 0.009556383825838566, + 0.012162444181740284, + 0.003694914747029543, + 0.004175611771643162, + 0.006914034020155668, + 0.004680023528635502, + 0.004357718862593174, + 0.005407172255218029, + 0.04687987267971039, + 0.022571058943867683, + 0.02634389139711857, + 0.010968788526952267, + 0.014055753126740456, + 0.010855312459170818, + 0.007183757144957781, + 0.008527418598532677, + 0.032373856753110886, + 0.016222715377807617, + 0.022938672453165054, + 0.014799188822507858, + 0.012366198934614658, + 0.01745518110692501, + 0.012731662951409817, + 0.03477214649319649, + 0.015629544854164124, + 0.008680205792188644, + 0.019648004323244095, + 0.01040971465408802, + 0.010128523223102093, + 0.019528768956661224, + 0.006133398041129112, + 0.01653994806110859, + 0.007105338387191296, + 0.004687555134296417, + 0.005545143503695726, + 0.0061361053958535194, + 0.00635366840288043, + 0.005666972137987614, + 0.005977547727525234, + 0.011046850122511387, + 0.007208452560007572, + 0.008010122925043106, + 0.01013936661183834, + 0.007100630085915327, + 0.006964568048715591, + 0.01504258532077074, + 0.0066811940632760525, + 0.014592238701879978, + 0.006317439954727888, + 0.009614596143364906, + 0.008315537124872208, + 0.008985226042568684, + 0.010549896396696568, + 0.010525284335017204, + 0.009797326289117336, + 0.008779192343354225, + 0.008513597771525383, + 0.03225696086883545, + 0.006047158967703581, + 0.008447897620499134, + 0.018965784460306168, + 0.010698655620217323, + 0.009860396385192871, + 0.013982701115310192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.0625808835029602 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521824538707733 + ], + "opposite_logprob_tokens": [ + -1.5021824836730957 + ], + "logprob_reward": -0.2521824538707733, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521824538707733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12042125314474106, + 0.0799640566110611, + 0.033406224101781845, + 0.18726737797260284, + 0.2283179759979248, + 0.05427940562367439, + 0.032162122428417206, + 0.06277547776699066, + 0.0934164896607399, + 0.10798966884613037 + ], + "credit_map": [ + 0.014156338758766651, + 0.021745841950178146, + 0.012019466608762741, + 0.009214473888278008, + 0.01568218320608139, + 0.014293058775365353, + 0.015329958871006966, + 0.017979931086301804, + 0.008428200148046017, + 0.01918969303369522, + 0.004525986034423113, + 0.00906138401478529, + 0.007357813883572817, + 0.010245107114315033, + 0.007303736172616482, + 0.013852138072252274, + 0.004478325601667166, + 0.005134539678692818, + 0.003040504874661565, + 0.0028046430088579655, + 0.004413691349327564, + 0.004783578682690859, + 0.0027412252966314554, + 0.0060097151435911655, + 0.0642630085349083, + 0.028142394497990608, + 0.03265604376792908, + 0.011872571893036366, + 0.01821521483361721, + 0.015078095719218254, + 0.008430976420640945, + 0.00860907044261694, + 0.042169589549303055, + 0.026909450069069862, + 0.03366399556398392, + 0.020244866609573364, + 0.020205162465572357, + 0.029297247529029846, + 0.016629807651042938, + 0.03919784724712372, + 0.005954911932349205, + 0.005509926937520504, + 0.016380490735173225, + 0.005498009268194437, + 0.007699592504650354, + 0.0054228100925683975, + 0.003690319135785103, + 0.0041233450174331665, + 0.004436018411070108, + 0.003668995341286063, + 0.0031733375508338213, + 0.003869087900966406, + 0.0032998789101839066, + 0.004698328673839569, + 0.003664167830720544, + 0.005352308042347431, + 0.0068708909675478935, + 0.006351171527057886, + 0.010711563751101494, + 0.0072693657130002975, + 0.0051880283281207085, + 0.011357187293469906, + 0.005298978183418512, + 0.009728288277983665, + 0.010495306923985481, + 0.016381418332457542, + 0.0071159121580421925, + 0.008308140560984612, + 0.014340848661959171, + 0.01855744980275631, + 0.00769520690664649, + 0.010522209107875824, + 0.006936563178896904, + 0.03651050850749016, + 0.0050543565303087234, + 0.0057482351548969746, + 0.021750619634985924, + 0.01083961222320795, + 0.009027467109262943, + 0.012122297659516335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.2521824538707733 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22569233179092407 + ], + "opposite_logprob_tokens": [ + -1.6006922721862793 + ], + "logprob_reward": -0.22569233179092407, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22569233179092407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08204161375761032, + 0.1744173914194107, + 0.07069942355155945, + 0.05110941454768181, + 0.3445122241973877, + 0.11632521450519562, + 0.026347186416387558, + 0.04739472642540932, + 0.06511519104242325, + 0.022037554532289505 + ], + "credit_map": [ + 0.010764594189822674, + 0.013118594884872437, + 0.0073088714852929115, + 0.008228606544435024, + 0.008278295397758484, + 0.009413525462150574, + 0.010918395593762398, + 0.014010732062160969, + 0.018957484513521194, + 0.04042471945285797, + 0.007999776862561703, + 0.009965281002223492, + 0.013895567506551743, + 0.02913912571966648, + 0.01205409225076437, + 0.04198134317994118, + 0.008230146951973438, + 0.00936176162213087, + 0.004219524562358856, + 0.005597381852567196, + 0.010088473558425903, + 0.015111728571355343, + 0.005138469394296408, + 0.012951938435435295, + 0.009375397115945816, + 0.0073615568690001965, + 0.004533533472567797, + 0.004388115834444761, + 0.006884898990392685, + 0.007771738804876804, + 0.004115635063499212, + 0.006678539793938398, + 0.06994850188493729, + 0.03881433978676796, + 0.04484347626566887, + 0.03234691545367241, + 0.02962816320359707, + 0.04821758344769478, + 0.022168662399053574, + 0.05854461342096329, + 0.009342292323708534, + 0.004976207856088877, + 0.04753176122903824, + 0.006198452319949865, + 0.012808621861040592, + 0.018507568165659904, + 0.00564954848960042, + 0.011310763657093048, + 0.0047053261660039425, + 0.0029199346899986267, + 0.0031677470542490482, + 0.003557516261935234, + 0.0027250282000750303, + 0.003109973855316639, + 0.00323389726690948, + 0.002927762223407626, + 0.005083668511360884, + 0.00501667195931077, + 0.008530773222446442, + 0.008325710892677307, + 0.0035026627592742443, + 0.0067070662043988705, + 0.004222922958433628, + 0.006005249917507172, + 0.0069580296985805035, + 0.006957078818231821, + 0.008524920791387558, + 0.011524949222803116, + 0.007839024998247623, + 0.009139147587120533, + 0.006913091521710157, + 0.007258950732648373, + 0.0018349341116845608, + 0.0030932712834328413, + 0.002580874366685748, + 0.0031242258846759796, + 0.0017356047173961997, + 0.00433655921369791, + 0.0023062601685523987, + 0.0030258256010711193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.22569233179092407 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006758331321179867 + ], + "opposite_logprob_tokens": [ + -5.006758213043213 + ], + "logprob_reward": -0.006758331321179867, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006758331321179867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062434591352939606, + 0.3062496781349182, + 0.05496791750192642, + 0.15250647068023682, + 0.14897651970386505, + 0.04409008100628853, + 0.06331279128789902, + 0.04882843792438507, + 0.04880484566092491, + 0.06982868164777756 + ], + "credit_map": [ + 0.009992686100304127, + 0.007947263307869434, + 0.004819335415959358, + 0.004891733638942242, + 0.009255586192011833, + 0.006344282533973455, + 0.008066282607614994, + 0.011117414571344852, + 0.03470221161842346, + 0.07613710314035416, + 0.015449345111846924, + 0.019546253606677055, + 0.02255634218454361, + 0.044661279767751694, + 0.02060830220580101, + 0.07258881628513336, + 0.010678673163056374, + 0.01633581891655922, + 0.0032236911356449127, + 0.005208245478570461, + 0.006443158257752657, + 0.00483026122674346, + 0.003514170413836837, + 0.0047338916920125484, + 0.05307787284255028, + 0.02262108214199543, + 0.025251610204577446, + 0.009666704572737217, + 0.015231545083224773, + 0.01260143518447876, + 0.007040228229016066, + 0.007015974726527929, + 0.027464985847473145, + 0.014039507135748863, + 0.02134726010262966, + 0.017613815143704414, + 0.015453248284757137, + 0.020361047238111496, + 0.012396556325256824, + 0.020300090312957764, + 0.0042194402776658535, + 0.004224758595228195, + 0.011489558964967728, + 0.004154409747570753, + 0.004967909771949053, + 0.0064365919679403305, + 0.00382223236374557, + 0.0047751739621162415, + 0.00938655436038971, + 0.006806717719882727, + 0.00570399034768343, + 0.006481191609054804, + 0.008347335271537304, + 0.0075773089192807674, + 0.00604245625436306, + 0.012967225164175034, + 0.005812051240354776, + 0.004197793547064066, + 0.008859222754836082, + 0.006331325508654118, + 0.005213249009102583, + 0.004353496711701155, + 0.004619038663804531, + 0.009442253969609737, + 0.004780486226081848, + 0.008405216038227081, + 0.0049647134728729725, + 0.004328654147684574, + 0.012549185194075108, + 0.005249810870736837, + 0.004502097610384226, + 0.004024676512926817, + 0.0056040287017822266, + 0.02422848902642727, + 0.003648664802312851, + 0.0045616808347404, + 0.016102172434329987, + 0.0038556891959160566, + 0.005233720410615206, + 0.00659422529861331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.006758331321179867 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293979734182358 + ], + "opposite_logprob_tokens": [ + -2.2379398345947266 + ], + "logprob_reward": -0.11293979734182358, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293979734182358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07742086797952652, + 0.06690115481615067, + 0.0454450324177742, + 0.06318429112434387, + 0.1150689423084259, + 0.31948086619377136, + 0.04661261662840843, + 0.13855141401290894, + 0.055781979113817215, + 0.0715528056025505 + ], + "credit_map": [ + 0.009005268104374409, + 0.01230017188936472, + 0.0071831997483968735, + 0.005727764684706926, + 0.01042564120143652, + 0.007389257661998272, + 0.011006485670804977, + 0.014383084140717983, + 0.008673005737364292, + 0.013950719498097897, + 0.004599248059093952, + 0.005958870518952608, + 0.008024418726563454, + 0.0084634218364954, + 0.005841241218149662, + 0.011390235275030136, + 0.005638347007334232, + 0.008399873040616512, + 0.0028206012211740017, + 0.0026062228716909885, + 0.0050364527851343155, + 0.010287998244166374, + 0.003047845559194684, + 0.007607697509229183, + 0.019788451492786407, + 0.008518839254975319, + 0.007602442055940628, + 0.004645002074539661, + 0.008477491326630116, + 0.006454984191805124, + 0.003436856437474489, + 0.004260230343788862, + 0.024267585948109627, + 0.012325289659202099, + 0.014950628392398357, + 0.010128004476428032, + 0.010478843003511429, + 0.01506535429507494, + 0.007254039868712425, + 0.020599206909537315, + 0.02186116762459278, + 0.012900929898023605, + 0.12866736948490143, + 0.01772676780819893, + 0.029424363747239113, + 0.07232441753149033, + 0.014176526106894016, + 0.02239934541285038, + 0.005957723129540682, + 0.0044407970272004604, + 0.007420488633215427, + 0.006160220596939325, + 0.0060852160677313805, + 0.0043412805534899235, + 0.004389437381178141, + 0.007817459292709827, + 0.013111730106174946, + 0.013311391696333885, + 0.02799469605088234, + 0.020325200632214546, + 0.01251638401299715, + 0.013142748735845089, + 0.01490684226155281, + 0.023242438212037086, + 0.004504956305027008, + 0.008361836895346642, + 0.005645215976983309, + 0.008534858003258705, + 0.013699726201593876, + 0.0047467658296227455, + 0.004668273963034153, + 0.005620346870273352, + 0.005949132610112429, + 0.02139430306851864, + 0.005620402283966541, + 0.008594147861003876, + 0.011164466850459576, + 0.005226909648627043, + 0.007357032969594002, + 0.00624641589820385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.11293979734182358 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, uninterrupted long shot begins with a low-angle tracking shot, gliding smoothly and deliberately across the phosphorescent moss floor of a magical forest bathed in eternal twilight. Ancient, gnarled trees with bark like iridescent obsidian tower over an invisible canopy, their massive branches swaying rhythmically, not in sync with the wind but as if conducting a silent, grand orchestra. Their leaves, seemingly made of thin, translucent crystal, veined with liquid gold, vibrate not with a hiss but with a tangible vibration, emitting a soft, pulsating cyan and purplish-red glow that ripples through the air like visible sound waves. Below, clusters of low-growing shrubs join the ethereal chorus; their enormous, bell-shaped flowers, radiating an inner white light, pulsate in synchronized harmony, their petals gently opening and closing, releasing sparkling points of light into the gentle, luminous mist woven between the trunks. The entire scene is illuminated from within by this omnipresent bioluminescence, creating a cool, mysterious palette of deep indigos and emerald greens, casting long dancing shadows and creating a profound atmosphere of serene, otherworldly wonder as the camera continues its steady, dreamlike progress through the living symphony.", + "seed": 29448, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_4/index_2081_seed_29448_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11532216183841229, + "vlm_logprob": -0.11532216183841229, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.11532216183841229, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10686961561441422 + ], + "opposite_logprob_tokens": [ + -2.356869697570801 + ], + "logprob_reward": -0.10686961561441422, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10686961561441422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14848069846630096, + 0.08522607386112213, + 0.08333208411931992, + 0.15099716186523438, + 0.09425575286149979, + 0.18393829464912415, + 0.034753140062093735, + 0.10999463498592377, + 0.03694896772503853, + 0.07207323610782623 + ], + "credit_map": [ + 0.014193813316524029, + 0.017271697521209717, + 0.02554766647517681, + 0.023402918130159378, + 0.025370169430971146, + 0.012973337434232235, + 0.016076460480690002, + 0.013644643127918243, + 0.012981940060853958, + 0.010628491640090942, + 0.009033394046127796, + 0.008536005392670631, + 0.014812363311648369, + 0.007738939486443996, + 0.008411074057221413, + 0.01308386866003275, + 0.00926975253969431, + 0.01788434572517872, + 0.008246532641351223, + 0.007630529347807169, + 0.013634644448757172, + 0.01271163672208786, + 0.007971188984811306, + 0.005983454640954733, + 0.013034130446612835, + 0.018571272492408752, + 0.015917694196105003, + 0.01257662009447813, + 0.025977861136198044, + 0.03577011451125145, + 0.013463158160448074, + 0.01568630523979664, + 0.011631584726274014, + 0.014030739665031433, + 0.010026398114860058, + 0.010338637046515942, + 0.014086847193539143, + 0.018535174429416656, + 0.006779851391911507, + 0.008826523087918758, + 0.02101612277328968, + 0.021921278908848763, + 0.03285347297787666, + 0.020367184653878212, + 0.028896212577819824, + 0.0334366075694561, + 0.011080204509198666, + 0.014367198571562767, + 0.004349677357822657, + 0.007999070920050144, + 0.003928971476852894, + 0.002427724888548255, + 0.005516144912689924, + 0.004260342102497816, + 0.003201907966285944, + 0.0030692999716848135, + 0.0060000731609761715, + 0.03958966210484505, + 0.006972447969019413, + 0.005161552224308252, + 0.03538458049297333, + 0.007242412306368351, + 0.005589957814663649, + 0.004053948912769556, + 0.0025828946381807327, + 0.006136912852525711, + 0.004198106471449137, + 0.0019033084390684962, + 0.009777383878827095, + 0.006234826520085335, + 0.002474734792485833, + 0.003640797222033143, + 0.006921511609107256, + 0.010445905849337578, + 0.007602122146636248, + 0.00703324843198061, + 0.011982292868196964, + 0.007366809528321028, + 0.014420484192669392, + 0.0063008591532707214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.10686961561441422 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16042466461658478 + ], + "opposite_logprob_tokens": [ + -1.9104247093200684 + ], + "logprob_reward": -0.16042466461658478, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16042466461658478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08228861540555954, + 0.11657353490591049, + 0.06413757801055908, + 0.11967446655035019, + 0.05004766955971718, + 0.17246271669864655, + 0.0312756672501564, + 0.12896336615085602, + 0.05895989015698433, + 0.17561648786067963 + ], + "credit_map": [ + 0.008559311740100384, + 0.00669122813269496, + 0.02105538919568062, + 0.01565837860107422, + 0.00733585562556982, + 0.0068615153431892395, + 0.006499216891825199, + 0.009627723135054111, + 0.016569115221500397, + 0.016856402158737183, + 0.015108821913599968, + 0.010586155578494072, + 0.011981423944234848, + 0.024748781695961952, + 0.007698426488786936, + 0.013024413958191872, + 0.009259025566279888, + 0.009137054905295372, + 0.0057886382564902306, + 0.007049722131341696, + 0.01006557047367096, + 0.010153204202651978, + 0.006004588678479195, + 0.006679777055978775, + 0.008482002653181553, + 0.007851921953260899, + 0.010643824934959412, + 0.009753112681210041, + 0.02278798259794712, + 0.03955204784870148, + 0.004388865549117327, + 0.016214700415730476, + 0.004998710937798023, + 0.008595388382673264, + 0.006897754967212677, + 0.006170257460325956, + 0.005286104511469603, + 0.004871366545557976, + 0.005925075616687536, + 0.007303006947040558, + 0.01997534930706024, + 0.020034372806549072, + 0.028865737840533257, + 0.019079886376857758, + 0.026249852031469345, + 0.029904525727033615, + 0.010975201614201069, + 0.01737780123949051, + 0.00359075958840549, + 0.006646563299000263, + 0.0027402155101299286, + 0.002675894880667329, + 0.005191621836274862, + 0.004171065054833889, + 0.0035671377554535866, + 0.002692408859729767, + 0.0063331518322229385, + 0.04113086313009262, + 0.010033119469881058, + 0.006887223105877638, + 0.03904370218515396, + 0.008799121715128422, + 0.01019628718495369, + 0.006539896596223116, + 0.0040145497769117355, + 0.009866601787507534, + 0.005028635263442993, + 0.003715414786711335, + 0.01836308091878891, + 0.00976432953029871, + 0.003936198074370623, + 0.004271082580089569, + 0.016589032486081123, + 0.027428438887000084, + 0.01846722513437271, + 0.01793918013572693, + 0.02505001798272133, + 0.01642322726547718, + 0.039585139602422714, + 0.014134225435554981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.16042466461658478 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22569061815738678 + ], + "opposite_logprob_tokens": [ + -1.6006906032562256 + ], + "logprob_reward": -0.22569061815738678, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22569061815738678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09287533909082413, + 0.40059924125671387, + 0.03258214145898819, + 0.20409883558750153, + 0.044251468032598495, + 0.06917324662208557, + 0.032382622361183167, + 0.05043831467628479, + 0.013351726345717907, + 0.06024710834026337 + ], + "credit_map": [ + 0.011650584638118744, + 0.008577202446758747, + 0.019295046105980873, + 0.016184166073799133, + 0.01281486451625824, + 0.0076413643546402454, + 0.008558307774364948, + 0.008153804577887058, + 0.05837569013237953, + 0.03522135689854622, + 0.04332699999213219, + 0.04379885271191597, + 0.04100063443183899, + 0.09529859572649002, + 0.03980572894215584, + 0.04377137869596481, + 0.005270788446068764, + 0.006784243509173393, + 0.0027163794729858637, + 0.0025318677071481943, + 0.004556574393063784, + 0.0031925213988870382, + 0.0037768054753541946, + 0.00375296245329082, + 0.02370046079158783, + 0.022269129753112793, + 0.044282734394073486, + 0.020721035078167915, + 0.024639025330543518, + 0.023371754214167595, + 0.02631106600165367, + 0.018803631886839867, + 0.00470369216054678, + 0.0064340392127633095, + 0.005391887854784727, + 0.006479332689195871, + 0.005349146667867899, + 0.004513913299888372, + 0.00494724977761507, + 0.0064322068355977535, + 0.008790967985987663, + 0.009831180796027184, + 0.00880934577435255, + 0.0064292773604393005, + 0.011458986438810825, + 0.013058879412710667, + 0.004447652027010918, + 0.006346956826746464, + 0.003693884704262018, + 0.0065035694278776646, + 0.003638173220679164, + 0.0023910836316645145, + 0.004678727127611637, + 0.005136032123118639, + 0.0033199989702552557, + 0.003021156182512641, + 0.002851656638085842, + 0.012481442652642727, + 0.0054560499265789986, + 0.004119783639907837, + 0.012493577785789967, + 0.0032579887192696333, + 0.006370971444994211, + 0.0034068413078784943, + 0.0012664097594097257, + 0.001856502378359437, + 0.001545257167890668, + 0.0011097192764282227, + 0.0023575685918331146, + 0.0016563457902520895, + 0.001806418295018375, + 0.001753504853695631, + 0.0052298554219305515, + 0.007359441835433245, + 0.007768032606691122, + 0.007450983393937349, + 0.00872323289513588, + 0.00867635477334261, + 0.008600318804383278, + 0.006438888143748045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.22569061815738678 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004653101786971092 + ], + "opposite_logprob_tokens": [ + -5.379652976989746 + ], + "logprob_reward": -0.004653101786971092, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004653101786971092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07146836072206497, + 0.3044462502002716, + 0.03797342628240585, + 0.1778344064950943, + 0.09497419744729996, + 0.10756440460681915, + 0.023124342784285545, + 0.10208214819431305, + 0.03106043115258217, + 0.04947209730744362 + ], + "credit_map": [ + 0.007825296372175217, + 0.006514979060739279, + 0.014712193980813026, + 0.015072192065417767, + 0.0053816623985767365, + 0.006644301116466522, + 0.008811529725790024, + 0.006506204605102539, + 0.06161312386393547, + 0.02368086203932762, + 0.027872201055288315, + 0.01714828424155712, + 0.024949725717306137, + 0.06703603267669678, + 0.022914504632353783, + 0.059231504797935486, + 0.0045085870660841465, + 0.00557343615218997, + 0.005190705414861441, + 0.003638411872088909, + 0.0043540820479393005, + 0.004383449908345938, + 0.0057651810348033905, + 0.004559572786092758, + 0.020955415442585945, + 0.023462790995836258, + 0.03251047059893608, + 0.021770913153886795, + 0.024580352008342743, + 0.01853283867239952, + 0.021760383620858192, + 0.014261247590184212, + 0.01115147303789854, + 0.013884552754461765, + 0.013577710837125778, + 0.010886015370488167, + 0.012435532175004482, + 0.015260821208357811, + 0.007876871153712273, + 0.00990122091025114, + 0.013201546855270863, + 0.01443091407418251, + 0.019387606531381607, + 0.011615688912570477, + 0.015934351831674576, + 0.01586027443408966, + 0.007537540514022112, + 0.00959648098796606, + 0.002657405799254775, + 0.004232823383063078, + 0.0025582017842680216, + 0.0027974890545010567, + 0.003466776106506586, + 0.0030872386414557695, + 0.002290484495460987, + 0.002033921191468835, + 0.005983210634440184, + 0.030518580228090286, + 0.008760158903896809, + 0.006059650331735611, + 0.02789018675684929, + 0.006379407364875078, + 0.009079431183636189, + 0.007411526050418615, + 0.0027133170515298843, + 0.004701131489127874, + 0.003152799094095826, + 0.007062150165438652, + 0.0033479693811386824, + 0.0026055655907839537, + 0.004288516938686371, + 0.003188981208950281, + 0.004087319131940603, + 0.006867234129458666, + 0.005098122637718916, + 0.004437060095369816, + 0.011867386288940907, + 0.007477201987057924, + 0.006436453200876713, + 0.0032013223972171545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.004653101786971092 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897280901670456 + ], + "opposite_logprob_tokens": [ + -2.578972816467285 + ], + "logprob_reward": -0.07897280901670456, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897280901670456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05489465966820717, + 0.26508423686027527, + 0.02054743841290474, + 0.046466972678899765, + 0.11824102699756622, + 0.08566629886627197, + 0.024983111768960953, + 0.28777021169662476, + 0.023924151435494423, + 0.07242182642221451 + ], + "credit_map": [ + 0.006139217875897884, + 0.004279805347323418, + 0.012313053011894226, + 0.011947411112487316, + 0.00455657672137022, + 0.004874080885201693, + 0.006298589985817671, + 0.004485927522182465, + 0.060507889837026596, + 0.01905456930398941, + 0.022798899561166763, + 0.013058684766292572, + 0.018913811072707176, + 0.050140414386987686, + 0.02028718963265419, + 0.060322780162096024, + 0.0026647639460861683, + 0.003924691118299961, + 0.0015710267471149564, + 0.0022050724364817142, + 0.002495279535651207, + 0.0031846300698816776, + 0.001988401636481285, + 0.002513572573661804, + 0.003926877863705158, + 0.005488904658704996, + 0.0037519168108701706, + 0.0026015460025519133, + 0.008050980046391487, + 0.01392091903835535, + 0.002511248691007495, + 0.006214579101651907, + 0.01430859137326479, + 0.013607732020318508, + 0.01846267469227314, + 0.01267564669251442, + 0.014660513028502464, + 0.020753933116793633, + 0.009729757905006409, + 0.014042175374925137, + 0.010620180517435074, + 0.01207297295331955, + 0.013175591826438904, + 0.010275921784341335, + 0.013466139324009418, + 0.014558191411197186, + 0.0054063037969172, + 0.00609099306166172, + 0.0028452493716031313, + 0.005401554051786661, + 0.002951831091195345, + 0.001997482031583786, + 0.0036844240967184305, + 0.0031446372158825397, + 0.00234397454187274, + 0.0026139591354876757, + 0.013531328178942204, + 0.11087054014205933, + 0.012268452905118465, + 0.008713395334780216, + 0.10172229260206223, + 0.01608950085937977, + 0.013511343859136105, + 0.011063357815146446, + 0.0015402496792376041, + 0.0030357767827808857, + 0.0022906591184437275, + 0.004340241197496653, + 0.004387692082673311, + 0.0022693309001624584, + 0.003551286179572344, + 0.0025089166592806578, + 0.006238401867449284, + 0.009384923614561558, + 0.004896326456218958, + 0.004880030173808336, + 0.0196381825953722, + 0.011966757476329803, + 0.010908059775829315, + 0.004509141203016043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.07897280901670456 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, uninterrupted long shot begins with a low-angle tracking shot, gliding smoothly and deliberately across the phosphorescent moss floor of a magical forest bathed in eternal twilight. Ancient, gnarled trees with bark like iridescent obsidian tower over an invisible canopy, their massive branches swaying rhythmically, not in sync with the wind but as if conducting a silent, grand orchestra. Their leaves, seemingly made of thin, translucent crystal, veined with liquid gold, vibrate not with a hiss but with a tangible vibration, emitting a soft, pulsating cyan and purplish-red glow that ripples through the air like visible sound waves. Below, clusters of low-growing shrubs join the ethereal chorus; their enormous, bell-shaped flowers, radiating an inner white light, pulsate in synchronized harmony, their petals gently opening and closing, releasing sparkling points of light into the gentle, luminous mist woven between the trunks. The entire scene is illuminated from within by this omnipresent bioluminescence, creating a cool, mysterious palette of deep indigos and emerald greens, casting long dancing shadows and creating a profound atmosphere of serene, otherworldly wonder as the camera continues its steady, dreamlike progress through the living symphony.", + "seed": 29449, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_4/index_2081_seed_29449_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06656651496887207, + "vlm_logprob": -0.06656651496887207, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.06656651496887207, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027940502390265465 + ], + "opposite_logprob_tokens": [ + -3.777940511703491 + ], + "logprob_reward": -0.027940502390265465, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027940502390265465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10860489308834076, + 0.034536391496658325, + 0.04801227152347565, + 0.13882634043693542, + 0.36313265562057495, + 0.11026632785797119, + 0.04757300764322281, + 0.0843246579170227, + 0.02311067096889019, + 0.041612789034843445 + ], + "credit_map": [ + 0.025824448093771935, + 0.017788497731089592, + 0.009321462363004684, + 0.010267945006489754, + 0.015271763317286968, + 0.013298843055963516, + 0.00920177437365055, + 0.007630154490470886, + 0.004182330798357725, + 0.003928561229258776, + 0.0038425864186137915, + 0.0030844423454254866, + 0.005444743670523167, + 0.004989491309970617, + 0.004947299603372812, + 0.004116937983781099, + 0.005519011523574591, + 0.006090863142162561, + 0.005737295839935541, + 0.005989882629364729, + 0.005930415354669094, + 0.007986828684806824, + 0.004819799680262804, + 0.005938174203038216, + 0.012600667774677277, + 0.029650550335645676, + 0.013780048117041588, + 0.010392951779067516, + 0.01766880601644516, + 0.019635092467069626, + 0.01729988306760788, + 0.017798349261283875, + 0.03396452218294144, + 0.05490178242325783, + 0.02416960522532463, + 0.0685323178768158, + 0.05645381286740303, + 0.032538339495658875, + 0.05481510981917381, + 0.03775714337825775, + 0.012342792004346848, + 0.031947117298841476, + 0.010770724155008793, + 0.010244910605251789, + 0.01413017325103283, + 0.013919636607170105, + 0.007751897443085909, + 0.009159074164927006, + 0.006261538248509169, + 0.007727430667728186, + 0.005713805090636015, + 0.005628267303109169, + 0.008549933321774006, + 0.00505112437531352, + 0.003945637959986925, + 0.0046952697448432446, + 0.008669759146869183, + 0.012047716416418552, + 0.009750333614647388, + 0.006146712228655815, + 0.010044685564935207, + 0.019975727424025536, + 0.008250254206359386, + 0.009439473040401936, + 0.0016949318815022707, + 0.005461899098008871, + 0.0023694694973528385, + 0.002464322140440345, + 0.002388587687164545, + 0.005320446565747261, + 0.001889082370325923, + 0.0015219306806102395, + 0.005985849071294069, + 0.006515612360090017, + 0.004109118599444628, + 0.004709814675152302, + 0.006513271480798721, + 0.0044939154759049416, + 0.004072784446179867, + 0.005212422460317612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.027940502390265465 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062071699649095535 + ], + "opposite_logprob_tokens": [ + -2.8120718002319336 + ], + "logprob_reward": -0.062071699649095535, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062071699649095535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020580794662237167, + 0.023493226617574692, + 0.05748100206255913, + 0.15940384566783905, + 0.4792998135089874, + 0.05480991303920746, + 0.02096259780228138, + 0.1281273365020752, + 0.01979851722717285, + 0.03604302182793617 + ], + "credit_map": [ + 0.0026275457348674536, + 0.0028231181204319, + 0.002881215186789632, + 0.0025962945073843002, + 0.002870509633794427, + 0.002154303714632988, + 0.0021340642124414444, + 0.002493740990757942, + 0.003059353679418564, + 0.0026780064217746258, + 0.0018729186849668622, + 0.0018259588396176696, + 0.0036512913648039103, + 0.0034646806307137012, + 0.00440013175830245, + 0.002540882444009185, + 0.009311996400356293, + 0.007515690289437771, + 0.0029834711458534002, + 0.005099172703921795, + 0.009280773811042309, + 0.012573273852467537, + 0.003970911726355553, + 0.0067457095719873905, + 0.006383589003235102, + 0.007879821583628654, + 0.00826902687549591, + 0.005298333708196878, + 0.050239577889442444, + 0.010379587300121784, + 0.05901087075471878, + 0.01194301713258028, + 0.045803003013134, + 0.07463724166154861, + 0.03226445987820625, + 0.08991724997758865, + 0.07349371165037155, + 0.043802183121442795, + 0.0709960088133812, + 0.04838591441512108, + 0.006825926247984171, + 0.01679266430437565, + 0.00429827906191349, + 0.006773374974727631, + 0.0061077941209077835, + 0.006036592647433281, + 0.00340382126159966, + 0.004571456462144852, + 0.001992160687223077, + 0.0029731092508882284, + 0.0016849827952682972, + 0.001841363380663097, + 0.004261918365955353, + 0.004839796107262373, + 0.0016853382112458348, + 0.001683926093392074, + 0.013590062037110329, + 0.023114891722798347, + 0.01112109050154686, + 0.02196335233747959, + 0.015789546072483063, + 0.023061368614435196, + 0.010187075473368168, + 0.009299937635660172, + 0.0011098376708105206, + 0.005929844453930855, + 0.001318496186286211, + 0.001766642672009766, + 0.0014860037481412292, + 0.005347926169633865, + 0.0013698064722120762, + 0.0014699563616886735, + 0.006035583559423685, + 0.0040539284236729145, + 0.005504055414348841, + 0.003571734530851245, + 0.004502504598349333, + 0.004376687575131655, + 0.003445018781349063, + 0.004553505219519138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.062071699649095535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256363034248352 + ], + "opposite_logprob_tokens": [ + -1.6006362438201904 + ], + "logprob_reward": -0.2256363034248352, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256363034248352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.016204457730054855, + 0.0077726528979837894, + 0.03392966091632843, + 0.3299865126609802, + 0.4330776631832123, + 0.05881012976169586, + 0.009064753539860249, + 0.066025011241436, + 0.021609237417578697, + 0.02351989410817623 + ], + "credit_map": [ + 0.0019397250143811107, + 0.0024663046933710575, + 0.0020993815269321203, + 0.0019677418749779463, + 0.0023885725531727076, + 0.002041013678535819, + 0.0016680406406521797, + 0.0016336787957698107, + 0.0007384815835393965, + 0.0010178093798458576, + 0.0009495132835581899, + 0.0007309106294997036, + 0.0013703355798497796, + 0.0012383840512484312, + 0.0010118051432073116, + 0.0007154139457270503, + 0.0055106752552092075, + 0.00409714225679636, + 0.002787974663078785, + 0.0035913437604904175, + 0.004298223182559013, + 0.005620894022285938, + 0.002561656292527914, + 0.005461754743009806, + 0.013101189397275448, + 0.013720182701945305, + 0.015449006110429764, + 0.009591011330485344, + 0.108697809278965, + 0.021795494481921196, + 0.12318892776966095, + 0.024442892521619797, + 0.041102174669504166, + 0.06642435491085052, + 0.02872207760810852, + 0.08198049664497375, + 0.06639987975358963, + 0.03787926957011223, + 0.06484608352184296, + 0.045723337680101395, + 0.006140085868537426, + 0.014527454040944576, + 0.004586620721966028, + 0.008575349114835262, + 0.007093193009495735, + 0.007170060183852911, + 0.004004620481282473, + 0.006712752860039473, + 0.0010684238513931632, + 0.0017083737766370177, + 0.0007848841487430036, + 0.0009074857807718217, + 0.0019318810664117336, + 0.001404988463036716, + 0.0006065757479518652, + 0.0006521414034068584, + 0.007272432092577219, + 0.008778003975749016, + 0.006954025477170944, + 0.012518982402980328, + 0.008135206997394562, + 0.011326692998409271, + 0.005895559675991535, + 0.005144108552485704, + 0.0013221915578469634, + 0.006519094109535217, + 0.0014685963978990912, + 0.0018859419506043196, + 0.0016382543835788965, + 0.005823794286698103, + 0.0014368925476446748, + 0.0015144721837714314, + 0.0038032226730138063, + 0.002663885708898306, + 0.0031451706308871508, + 0.0029003384988754988, + 0.002589463721960783, + 0.002542155096307397, + 0.002814775798469782, + 0.0030608847737312317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.2256363034248352 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004649304784834385 + ], + "opposite_logprob_tokens": [ + -5.3796491622924805 + ], + "logprob_reward": -0.004649304784834385, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004649304784834385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06515704095363617, + 0.020917201414704323, + 0.024845484644174576, + 0.05932837724685669, + 0.5862857699394226, + 0.08070481568574905, + 0.0318915881216526, + 0.05660463124513626, + 0.02025509625673294, + 0.05400998145341873 + ], + "credit_map": [ + 0.014974893070757389, + 0.012812691740691662, + 0.007149744778871536, + 0.004444976802915335, + 0.009249852038919926, + 0.007855067029595375, + 0.004406311083585024, + 0.004263508133590221, + 0.002444059820845723, + 0.0023216214030981064, + 0.0019739484414458275, + 0.002888414077460766, + 0.003368483856320381, + 0.0028077319730073214, + 0.0028500996995717287, + 0.0022628423757851124, + 0.004042625892907381, + 0.0035590766929090023, + 0.0034281760454177856, + 0.0025572990998625755, + 0.00248823594301939, + 0.0033567287027835846, + 0.0014390572905540466, + 0.003974283579736948, + 0.00417206110432744, + 0.005642683245241642, + 0.0031798447016626596, + 0.003465629881247878, + 0.01775035448372364, + 0.0055849249474704266, + 0.015421411953866482, + 0.004111468326300383, + 0.055599041283130646, + 0.09088840335607529, + 0.03879310563206673, + 0.10884392261505127, + 0.0878530815243721, + 0.05209324136376381, + 0.08877089619636536, + 0.0634440928697586, + 0.009389023296535015, + 0.016336889937520027, + 0.010530536994338036, + 0.010612939484417439, + 0.010308369994163513, + 0.007873138412833214, + 0.007425208110362291, + 0.008228704333305359, + 0.002212608465924859, + 0.00442476337775588, + 0.00341770239174366, + 0.003528113244101405, + 0.006259385496377945, + 0.0040098377503454685, + 0.004629755392670631, + 0.003409423166885972, + 0.006121372804045677, + 0.009318429976701736, + 0.005294688045978546, + 0.011244907975196838, + 0.006659314967691898, + 0.007031020242720842, + 0.006341487634927034, + 0.004593405872583389, + 0.002609622897580266, + 0.0020593958906829357, + 0.0026027418207377195, + 0.0028389475774019957, + 0.0035270927473902702, + 0.002333408920094371, + 0.0021415797527879477, + 0.0021423071157187223, + 0.006343055982142687, + 0.008554348722100258, + 0.006069779861718416, + 0.0052797431126236916, + 0.008692456409335136, + 0.006954287178814411, + 0.005229244939982891, + 0.006887063384056091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.004649304784834385 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534764595329762 + ], + "opposite_logprob_tokens": [ + -4.3875346183776855 + ], + "logprob_reward": -0.012534764595329762, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534764595329762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060237035155296326, + 0.012089378200471401, + 0.03657696396112442, + 0.05995072424411774, + 0.15493237972259521, + 0.27484673261642456, + 0.02879244275391102, + 0.24345985054969788, + 0.03165558725595474, + 0.09745889157056808 + ], + "credit_map": [ + 0.01177813857793808, + 0.009049782529473305, + 0.006442466750741005, + 0.006144377402961254, + 0.008312935940921307, + 0.007412589155137539, + 0.0067187100648880005, + 0.004378036595880985, + 0.0011567927431315184, + 0.0016485872911289334, + 0.0013887904351577163, + 0.0014720915351063013, + 0.0015224948292598128, + 0.001543463789857924, + 0.0018559980671852827, + 0.0015011591603979468, + 0.005104961339384317, + 0.003732949960976839, + 0.004267194774001837, + 0.003907374106347561, + 0.005209419410675764, + 0.005631484091281891, + 0.003248658962547779, + 0.005474921781569719, + 0.00400906428694725, + 0.004519275389611721, + 0.004610105883330107, + 0.002998883603140712, + 0.01237662322819233, + 0.006090804934501648, + 0.019998790696263313, + 0.005347177851945162, + 0.014980509877204895, + 0.02380572445690632, + 0.009931524284183979, + 0.02667599357664585, + 0.022703584283590317, + 0.015658045187592506, + 0.024096721783280373, + 0.017080282792448997, + 0.033169664442539215, + 0.08955296874046326, + 0.023255933076143265, + 0.024961154907941818, + 0.03383675217628479, + 0.03561345115303993, + 0.01576400175690651, + 0.018692785874009132, + 0.0025020712055265903, + 0.003468614537268877, + 0.0026058738585561514, + 0.0025998419150710106, + 0.00469463923946023, + 0.0048935976810753345, + 0.0026705560740083456, + 0.005357249639928341, + 0.027776651084423065, + 0.03389015048742294, + 0.021535826846957207, + 0.05140414461493492, + 0.029290009289979935, + 0.04085615649819374, + 0.021781794726848602, + 0.01692511886358261, + 0.002834930084645748, + 0.004667514935135841, + 0.0030945155303925276, + 0.0030382126569747925, + 0.0034911788534373045, + 0.0054327757097780704, + 0.0049171485006809235, + 0.004179311916232109, + 0.010951252654194832, + 0.01228544395416975, + 0.012671773321926594, + 0.013729363679885864, + 0.011629790998995304, + 0.010336506180465221, + 0.013431495055556297, + 0.012423256412148476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.012534764595329762 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, uninterrupted long shot begins with a low-angle tracking shot, gliding smoothly and deliberately across the phosphorescent moss floor of a magical forest bathed in eternal twilight. Ancient, gnarled trees with bark like iridescent obsidian tower over an invisible canopy, their massive branches swaying rhythmically, not in sync with the wind but as if conducting a silent, grand orchestra. Their leaves, seemingly made of thin, translucent crystal, veined with liquid gold, vibrate not with a hiss but with a tangible vibration, emitting a soft, pulsating cyan and purplish-red glow that ripples through the air like visible sound waves. Below, clusters of low-growing shrubs join the ethereal chorus; their enormous, bell-shaped flowers, radiating an inner white light, pulsate in synchronized harmony, their petals gently opening and closing, releasing sparkling points of light into the gentle, luminous mist woven between the trunks. The entire scene is illuminated from within by this omnipresent bioluminescence, creating a cool, mysterious palette of deep indigos and emerald greens, casting long dancing shadows and creating a profound atmosphere of serene, otherworldly wonder as the camera continues its steady, dreamlike progress through the living symphony.", + "seed": 29450, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_4/index_2081_seed_29450_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08927734438329935, + "vlm_logprob": -0.08927734438329935, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.08927734438329935, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09422339498996735 + ], + "opposite_logprob_tokens": [ + -2.4692234992980957 + ], + "logprob_reward": -0.09422339498996735, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09422339498996735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0857967808842659, + 0.09385168552398682, + 0.1023755893111229, + 0.06492169201374054, + 0.15157461166381836, + 0.07532283663749695, + 0.08396472036838531, + 0.08166145533323288, + 0.20396918058395386, + 0.05656147003173828 + ], + "credit_map": [ + 0.007464142981916666, + 0.009016045369207859, + 0.013179788365960121, + 0.010406291112303734, + 0.011637192219495773, + 0.01595231145620346, + 0.009567802771925926, + 0.008573204278945923, + 0.007000735495239496, + 0.009497188031673431, + 0.006789092440158129, + 0.012213965877890587, + 0.008933410979807377, + 0.015286838635802269, + 0.027334151789546013, + 0.006796295754611492, + 0.010675432160496712, + 0.020212527364492416, + 0.008351728320121765, + 0.01041070744395256, + 0.023554636165499687, + 0.011995662935078144, + 0.00869680941104889, + 0.008478095754981041, + 0.006624847650527954, + 0.007675688248127699, + 0.007151199970394373, + 0.004345277324318886, + 0.0148495277389884, + 0.009516210295259953, + 0.009744176641106606, + 0.0050147660076618195, + 0.013860605657100677, + 0.02444332279264927, + 0.006938657257705927, + 0.009322866797447205, + 0.03973875194787979, + 0.03527044504880905, + 0.009328329935669899, + 0.012671638280153275, + 0.006566272117197514, + 0.00963288638740778, + 0.006417613476514816, + 0.007221750915050507, + 0.018013518303632736, + 0.014965107664465904, + 0.007124291267246008, + 0.00538139883428812, + 0.00768332090228796, + 0.01771005615592003, + 0.0072599174454808235, + 0.007770902942866087, + 0.016152162104845047, + 0.012907354161143303, + 0.005984515883028507, + 0.008496485650539398, + 0.007744370028376579, + 0.01474254671484232, + 0.007441692054271698, + 0.00968544464558363, + 0.01481568906456232, + 0.013674546033143997, + 0.007420487701892853, + 0.006136676762253046, + 0.009633861482143402, + 0.060691315680742264, + 0.010027511976659298, + 0.015326708555221558, + 0.07376228272914886, + 0.011733179911971092, + 0.010143320076167583, + 0.012651007622480392, + 0.005918360315263271, + 0.00692747114226222, + 0.007134059444069862, + 0.005666445009410381, + 0.009276296943426132, + 0.010702889412641525, + 0.005676364526152611, + 0.00525958277285099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -0.09422339498996735 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992520332336426 + ], + "opposite_logprob_tokens": [ + -1.8049252033233643 + ], + "logprob_reward": -0.17992520332336426, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992520332336426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08149578422307968, + 0.161970853805542, + 0.07008113712072372, + 0.03823322057723999, + 0.044981252402067184, + 0.016263844445347786, + 0.17241837084293365, + 0.1891794502735138, + 0.17036046087741852, + 0.05501563102006912 + ], + "credit_map": [ + 0.008923476561903954, + 0.008594119921326637, + 0.01098808366805315, + 0.008477170020341873, + 0.013144063763320446, + 0.013297961093485355, + 0.010617156513035297, + 0.007453745696693659, + 0.015217795968055725, + 0.015240127220749855, + 0.012081409804522991, + 0.01986785978078842, + 0.015063056722283363, + 0.02828027307987213, + 0.04689806327223778, + 0.009322257712483406, + 0.004943489097058773, + 0.011400061659514904, + 0.0065923333168029785, + 0.00873482134193182, + 0.014483182691037655, + 0.010063757188618183, + 0.007066912017762661, + 0.006796578876674175, + 0.004039045423269272, + 0.006502840667963028, + 0.003005111124366522, + 0.003542356425896287, + 0.005906498059630394, + 0.004757513292133808, + 0.005819486454129219, + 0.004660371225327253, + 0.0031018387526273727, + 0.005835114512592554, + 0.003540384117513895, + 0.005696929059922695, + 0.008112704381346703, + 0.010040812194347382, + 0.004035070072859526, + 0.0046184007078409195, + 0.0017296249279752374, + 0.0026299674063920975, + 0.0012218101182952523, + 0.0017521687550470233, + 0.0027440281119197607, + 0.00213093520142138, + 0.0017418772913515568, + 0.0023134334478527308, + 0.01310175284743309, + 0.04841243103146553, + 0.008465071208775043, + 0.013379242271184921, + 0.045234471559524536, + 0.022838681936264038, + 0.007372196763753891, + 0.013614527881145477, + 0.016065822914242744, + 0.04083770513534546, + 0.015928661450743675, + 0.017928263172507286, + 0.03660210967063904, + 0.031001748517155647, + 0.0160652045160532, + 0.014749935828149319, + 0.007812961004674435, + 0.06572254747152328, + 0.005475349724292755, + 0.008109580725431442, + 0.05901340767741203, + 0.011088777333498001, + 0.007104441523551941, + 0.006033409386873245, + 0.005031616892665625, + 0.01308529544621706, + 0.007317869924008846, + 0.003199840895831585, + 0.008455438539385796, + 0.010795659385621548, + 0.003274105256423354, + 0.003855800721794367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.17992520332336426 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06211617588996887 + ], + "opposite_logprob_tokens": [ + -2.8121161460876465 + ], + "logprob_reward": -0.06211617588996887, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06211617588996887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11085294187068939, + 0.19296614825725555, + 0.09921092540025711, + 0.10758326947689056, + 0.09688472747802734, + 0.025069886818528175, + 0.05725037679076195, + 0.06487757712602615, + 0.1913161724805832, + 0.05398791655898094 + ], + "credit_map": [ + 0.009468755684792995, + 0.014302917756140232, + 0.011706123128533363, + 0.009683809243142605, + 0.013346136547625065, + 0.01833305135369301, + 0.02174319326877594, + 0.01226896233856678, + 0.019633131101727486, + 0.024787530303001404, + 0.021105850115418434, + 0.023610973730683327, + 0.03556126356124878, + 0.03303709998726845, + 0.018679380416870117, + 0.01655092090368271, + 0.007947414182126522, + 0.015340018086135387, + 0.008662581443786621, + 0.012477923184633255, + 0.019808638840913773, + 0.012494499795138836, + 0.012718851678073406, + 0.009761005640029907, + 0.011239008978009224, + 0.02448403462767601, + 0.009553536772727966, + 0.009071297943592072, + 0.016846640035510063, + 0.014696243219077587, + 0.012555849738419056, + 0.009136646054685116, + 0.00783479493111372, + 0.014250924810767174, + 0.0070369853638112545, + 0.009204354137182236, + 0.01989186927676201, + 0.022078100591897964, + 0.0060859182849526405, + 0.010501782409846783, + 0.002456408692523837, + 0.004184595309197903, + 0.0033365655690431595, + 0.0026026726700365543, + 0.003068068064749241, + 0.0029724249616265297, + 0.0032518927473574877, + 0.003197257174178958, + 0.0052665951661765575, + 0.01590610109269619, + 0.0029231985099613667, + 0.0043199555948376656, + 0.013030115514993668, + 0.007632331922650337, + 0.0025758175179362297, + 0.005596263334155083, + 0.0050089433789253235, + 0.012197163887321949, + 0.007931931875646114, + 0.009312770329415798, + 0.010329796001315117, + 0.009565800428390503, + 0.00639269407838583, + 0.004138473886996508, + 0.012470917776226997, + 0.0676727443933487, + 0.009498128667473793, + 0.010708576999604702, + 0.05216874182224274, + 0.016238437965512276, + 0.009956739842891693, + 0.012601885944604874, + 0.006185573060065508, + 0.007906388491392136, + 0.005937603767961264, + 0.004767897538840771, + 0.006955175660550594, + 0.011092890985310078, + 0.005820424761623144, + 0.005321957170963287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.06211617588996887 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009805371053516865 + ], + "opposite_logprob_tokens": [ + -4.634805202484131 + ], + "logprob_reward": -0.009805371053516865, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009805371053516865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04223739728331566, + 0.09501542896032333, + 0.06709348410367966, + 0.13933753967285156, + 0.11274784803390503, + 0.04568909481167793, + 0.03466456010937691, + 0.03233485668897629, + 0.3987189829349518, + 0.03216082230210304 + ], + "credit_map": [ + 0.004182485863566399, + 0.004047506023198366, + 0.006622740998864174, + 0.004776179790496826, + 0.005960035137832165, + 0.005908779799938202, + 0.006184400524944067, + 0.004555270075798035, + 0.007390967570245266, + 0.00831126794219017, + 0.007490789517760277, + 0.009647419676184654, + 0.010468894615769386, + 0.02024402655661106, + 0.025084173306822777, + 0.006377892103046179, + 0.005389371886849403, + 0.013009208254516125, + 0.005679078865796328, + 0.005038105882704258, + 0.016331514343619347, + 0.010386161506175995, + 0.005415833089500666, + 0.005844212137162685, + 0.00975034013390541, + 0.01694217137992382, + 0.019343329593539238, + 0.009312180802226067, + 0.03178181126713753, + 0.019028108566999435, + 0.014714977703988552, + 0.01846461370587349, + 0.008623402565717697, + 0.009992493316531181, + 0.014461152255535126, + 0.012342764064669609, + 0.01424824446439743, + 0.01516751665621996, + 0.012125813402235508, + 0.025786466896533966, + 0.003242025151848793, + 0.0045317537151277065, + 0.0056972443126142025, + 0.004301773384213448, + 0.007001966703683138, + 0.005492872558534145, + 0.0045617325231432915, + 0.010859725996851921, + 0.0035592149943113327, + 0.00761900469660759, + 0.0026936251670122147, + 0.0031171839218586683, + 0.008352084085345268, + 0.004659702535718679, + 0.00182737922295928, + 0.0028363631572574377, + 0.0033420121762901545, + 0.004206646699458361, + 0.0034750117920339108, + 0.005413672421127558, + 0.004919766914099455, + 0.004488678183406591, + 0.00394374318420887, + 0.0025453264825046062, + 0.018245868384838104, + 0.1402645707130432, + 0.01987946219742298, + 0.02583964914083481, + 0.14524206519126892, + 0.02039559744298458, + 0.01884741708636284, + 0.010004343464970589, + 0.0030785121489316225, + 0.009213294833898544, + 0.0030391213949769735, + 0.002087114844471216, + 0.005883455742150545, + 0.003961149603128433, + 0.0021313184406608343, + 0.0027668573893606663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 4 + ], + "selected_reward": -0.009805371053516865 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle tracking shot that glides smoothly across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the trees depicted as ancient and gnarled with bark resembling iridescent obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the tree leaves appear as thin, translucent crystal veined with liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves and flowers emit a soft, pulsating cyan and purplish-red glow with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there low-growing shrubs with enormous, bell-shaped flowers radiating an inner white light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031657665967941 + ], + "opposite_logprob_tokens": [ + -2.3503165245056152 + ], + "logprob_reward": -0.10031657665967941, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031657665967941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06544578820466995, + 0.06960050761699677, + 0.05865970626473427, + 0.18358533084392548, + 0.28384193778038025, + 0.03517737612128258, + 0.06185854971408844, + 0.055682238191366196, + 0.14582559466362, + 0.040323011577129364 + ], + "credit_map": [ + 0.004817856941372156, + 0.011585989035665989, + 0.007429217454046011, + 0.008091704919934273, + 0.00675631919875741, + 0.008398905396461487, + 0.009260828606784344, + 0.009104968048632145, + 0.006136714946478605, + 0.005249127745628357, + 0.006541526410728693, + 0.008040751330554485, + 0.006687946151942015, + 0.01062279473990202, + 0.02194182761013508, + 0.00437981728464365, + 0.00551660405471921, + 0.011866604909300804, + 0.004747218918055296, + 0.006527556572109461, + 0.011637057177722454, + 0.008068157359957695, + 0.0054086619056761265, + 0.004887849558144808, + 0.014746141619980335, + 0.028056396171450615, + 0.02404855750501156, + 0.011939811520278454, + 0.03612066060304642, + 0.023618832230567932, + 0.018861224874854088, + 0.026193704456090927, + 0.02020109072327614, + 0.036644116044044495, + 0.03590456396341324, + 0.027163472026586533, + 0.04310286417603493, + 0.04447551071643829, + 0.030405614525079727, + 0.0459446907043457, + 0.003742391709238291, + 0.005673657171428204, + 0.0038686937186867, + 0.004229327663779259, + 0.00613003084436059, + 0.005074184387922287, + 0.0034360408317297697, + 0.0030230481643229723, + 0.005765685345977545, + 0.01274350006133318, + 0.004415699280798435, + 0.006813008338212967, + 0.012979075312614441, + 0.01036008633673191, + 0.0034648168366402388, + 0.005316682159900665, + 0.004931285977363586, + 0.008224465884268284, + 0.0054840827360749245, + 0.008292128331959248, + 0.009735479019582272, + 0.0076504298485815525, + 0.005620470270514488, + 0.005743892397731543, + 0.007669337559491396, + 0.0434684157371521, + 0.0069784861989319324, + 0.008861979469656944, + 0.04729229956865311, + 0.00881895050406456, + 0.014324795454740524, + 0.008411319926381111, + 0.004019981250166893, + 0.006140715908259153, + 0.005415167193859816, + 0.0044261617586016655, + 0.0057687158696353436, + 0.0051645319908857346, + 0.004788225516676903, + 0.004599509760737419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.10031657665967941 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_5/index_tensor([6531, 6531])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_5/index_tensor([6531, 6531])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..92e76eb3ebca754c7182abef3040b9f4c3bde95c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_5/index_tensor([6531, 6531])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-definition time-lapse unfolds from a static, slightly high-angle close-up to focus on an exquisite white porcelain water dropper perched on the dark, polished mahogany surface of a traditional scholar's desk. Shaped like a small, smooth gourd, the dropper features a tiny, elegant spout and boasts a flawless, glossy glaze, reflecting its surroundings with crystal clarity. As the timelapse progresses, a single, sharp beam of sunlight from an unseen window deliberately sweeps across the scene, its color temperature shifting noticeably from the cool, pale white of morning to the intense, bright glare of midday, finally fading into the warm, golden hues of late afternoon. The shadow cast by the dropper dramatically shortens, swirls, and then stretches across the rich texture, while the shadow of the window frame floats quietly over the background elements of ink washes and still calligraphic brushstrokes. With each subtle shift in the sun's arc, the gleaming porcelain surface catches the light, creating a mesmerizing pulsation as bright, mirrored highlights illuminate and gently fade, making the inanimate object feel alive. This silent, rhythmic play of light and shadow transforms a simple scene into a profound and meditative observation of the passage of time, imbuing the image with a deep tranquility and focused calm.", + "seed": 9180, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_5/index_6531_seed_9180_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.024161070678383113, + "vlm_logprob": -0.024161070678383113, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.024161070678383113, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007846249267458916 + ], + "opposite_logprob_tokens": [ + -5.757846355438232 + ], + "logprob_reward": -0.007846249267458916, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007846249267458916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1154031977057457, + 0.06503799557685852, + 0.04988308250904083, + 0.05353131517767906, + 0.1342519074678421, + 0.02736271545290947, + 0.0789196789264679, + 0.05650325492024422, + 0.18898622691631317, + 0.23012062907218933 + ], + "credit_map": [ + 0.009287858381867409, + 0.017707331106066704, + 0.019248725846409798, + 0.019771208986639977, + 0.014193981885910034, + 0.011435437947511673, + 0.013319025747478008, + 0.010439623147249222, + 0.005294125992804766, + 0.010987316258251667, + 0.008771857246756554, + 0.008693398907780647, + 0.00830900389701128, + 0.007574108894914389, + 0.007640086580067873, + 0.007768095470964909, + 0.0043773953802883625, + 0.00919927004724741, + 0.010116457007825375, + 0.004655080381780863, + 0.004093850497156382, + 0.006424747873097658, + 0.004224733915179968, + 0.006791544612497091, + 0.006816885434091091, + 0.006188841536641121, + 0.0159777719527483, + 0.00671956455335021, + 0.004532329272478819, + 0.0053625768050551414, + 0.0039137317799031734, + 0.004019609652459621, + 0.01881497912108898, + 0.0103410379961133, + 0.053433604538440704, + 0.014650947414338589, + 0.010600170120596886, + 0.00790342129766941, + 0.011856398545205593, + 0.006651351228356361, + 0.0019547860138118267, + 0.004813197534531355, + 0.003821484511718154, + 0.004338807892054319, + 0.0026539957616478205, + 0.003616000059992075, + 0.0036140619777143, + 0.0025503821671009064, + 0.006729459390044212, + 0.008529433980584145, + 0.010333760641515255, + 0.008320894092321396, + 0.009930525906383991, + 0.01446203701198101, + 0.008506439626216888, + 0.01210713479667902, + 0.0036779504735022783, + 0.009476781822741032, + 0.012227817438542843, + 0.015242849476635456, + 0.002706604776903987, + 0.0029102452099323273, + 0.006071749608963728, + 0.004189256113022566, + 0.018402332440018654, + 0.010764886625111103, + 0.06857743114233017, + 0.032697197049856186, + 0.005122464150190353, + 0.009580486454069614, + 0.009452738799154758, + 0.03438868373632431, + 0.014216403476893902, + 0.011945136822760105, + 0.01685957796871662, + 0.09723460674285889, + 0.005707316100597382, + 0.007978130131959915, + 0.012165963649749756, + 0.06401348114013672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.007846249267458916 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647880792617798 + ], + "opposite_logprob_tokens": [ + -5.379647731781006 + ], + "logprob_reward": -0.004647880792617798, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647880792617798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15640316903591156, + 0.11042367666959763, + 0.061252545565366745, + 0.05504460260272026, + 0.09628283232450485, + 0.05560584366321564, + 0.060100022703409195, + 0.04087376594543457, + 0.042675189673900604, + 0.32133838534355164 + ], + "credit_map": [ + 0.01508949976414442, + 0.017520753666758537, + 0.028437109664082527, + 0.02706858143210411, + 0.016361963003873825, + 0.0171473640948534, + 0.01789390854537487, + 0.0168839730322361, + 0.011659108102321625, + 0.010665930807590485, + 0.016861382871866226, + 0.01381184533238411, + 0.011861591599881649, + 0.015009475871920586, + 0.01578153669834137, + 0.01477278396487236, + 0.005656877998262644, + 0.006207714788615704, + 0.008682459592819214, + 0.007289920002222061, + 0.007487411145120859, + 0.010463140904903412, + 0.007295137271285057, + 0.008169878274202347, + 0.0057827443815767765, + 0.0044622342102229595, + 0.01820521615445614, + 0.006545960903167725, + 0.005193708930164576, + 0.006181646604090929, + 0.004304032307118177, + 0.004369055852293968, + 0.006668254267424345, + 0.00895555131137371, + 0.02616068162024021, + 0.016731208190321922, + 0.008449351415038109, + 0.008760906755924225, + 0.010955804027616978, + 0.009601067751646042, + 0.005358962807804346, + 0.008993329480290413, + 0.007620507851243019, + 0.007045283447951078, + 0.007377498783171177, + 0.0077558341436088085, + 0.006424569990485907, + 0.005029850173741579, + 0.005971002858132124, + 0.008171318098902702, + 0.007210942450910807, + 0.006025856826454401, + 0.007102173753082752, + 0.009911530651152134, + 0.007016953080892563, + 0.008690242655575275, + 0.0028490673284977674, + 0.005876698065549135, + 0.0058626532554626465, + 0.007805303204804659, + 0.0036890653427690268, + 0.0043799602426588535, + 0.005067215766757727, + 0.00534379668533802, + 0.004162889439612627, + 0.003291386878117919, + 0.012457791715860367, + 0.006890534423291683, + 0.0022005364298820496, + 0.004043543245643377, + 0.003595477668568492, + 0.006033026613295078, + 0.017490245401859283, + 0.016575807705521584, + 0.02246019057929516, + 0.13043323159217834, + 0.00944115687161684, + 0.014811147004365921, + 0.01656482368707657, + 0.09356175363063812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.004647880792617798 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904251456260681 + ], + "opposite_logprob_tokens": [ + -2.4640424251556396 + ], + "logprob_reward": -0.08904251456260681, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904251456260681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09149755537509918, + 0.06444210559129715, + 0.10300393402576447, + 0.12240269035100937, + 0.12125097215175629, + 0.12353099882602692, + 0.06116975471377373, + 0.041769057512283325, + 0.06177455186843872, + 0.20915842056274414 + ], + "credit_map": [ + 0.0069509148597717285, + 0.00953609962016344, + 0.016863465309143066, + 0.012100507505238056, + 0.012053235433995724, + 0.009736312553286552, + 0.012448878958821297, + 0.011808138340711594, + 0.006084552500396967, + 0.005896429531276226, + 0.012367865070700645, + 0.00756740802899003, + 0.009197973646223545, + 0.008896502666175365, + 0.00678798696026206, + 0.0076433829963207245, + 0.010309389792382717, + 0.011573501862585545, + 0.015151034109294415, + 0.008866655640304089, + 0.012467429041862488, + 0.017798233777284622, + 0.011868528090417385, + 0.014969155192375183, + 0.009719746187329292, + 0.011643920093774796, + 0.016253627836704254, + 0.011542171239852905, + 0.01741746813058853, + 0.026526641100645065, + 0.010577628389000893, + 0.018721487373113632, + 0.017533840611577034, + 0.00863974541425705, + 0.04677850753068924, + 0.012437889352440834, + 0.009451569058001041, + 0.008857226930558681, + 0.010327952913939953, + 0.007224226836115122, + 0.008932020515203476, + 0.014912081882357597, + 0.016973212361335754, + 0.017098942771553993, + 0.012911329977214336, + 0.022186778485774994, + 0.01625882461667061, + 0.014257815666496754, + 0.005933697801083326, + 0.008908660151064396, + 0.008097966201603413, + 0.010447544045746326, + 0.005063247866928577, + 0.007848929613828659, + 0.006826401222497225, + 0.008043302223086357, + 0.0031705754809081554, + 0.005231014918535948, + 0.007663756608963013, + 0.006887038238346577, + 0.0028353803791105747, + 0.0035062297247350216, + 0.0075696911662817, + 0.004905371926724911, + 0.004490477032959461, + 0.003924757242202759, + 0.018582189455628395, + 0.008339302614331245, + 0.0026975246146321297, + 0.0052094897255301476, + 0.004839778412133455, + 0.013691033236682415, + 0.010520624928176403, + 0.011718693189322948, + 0.012710857205092907, + 0.08099374175071716, + 0.007444528862833977, + 0.01127124484628439, + 0.010646731592714787, + 0.0638519898056984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 3 + ], + "selected_reward": -0.08904251456260681 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533940374851227 + ], + "opposite_logprob_tokens": [ + -4.387534141540527 + ], + "logprob_reward": -0.012533940374851227, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533940374851227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24195611476898193, + 0.12226749956607819, + 0.10519038885831833, + 0.09792196750640869, + 0.08667897433042526, + 0.029013998806476593, + 0.06796949356794357, + 0.033054251223802567, + 0.14280745387077332, + 0.07313983142375946 + ], + "credit_map": [ + 0.020884348079562187, + 0.02924000658094883, + 0.04172607511281967, + 0.04028655216097832, + 0.029050976037979126, + 0.024015793576836586, + 0.027353862300515175, + 0.02939850464463234, + 0.009783560410141945, + 0.016522424295544624, + 0.01570923626422882, + 0.016513895243406296, + 0.019787058234214783, + 0.015547296032309532, + 0.014151904731988907, + 0.014252126216888428, + 0.009235368110239506, + 0.008776748552918434, + 0.014179877005517483, + 0.010466834530234337, + 0.015064277686178684, + 0.018528863787651062, + 0.012001716531813145, + 0.01693669706583023, + 0.009307133965194225, + 0.005084603559225798, + 0.03207254782319069, + 0.007959449663758278, + 0.009914699010550976, + 0.012611915357410908, + 0.008528856560587883, + 0.01244275737553835, + 0.010858163237571716, + 0.008209381252527237, + 0.024903692305088043, + 0.012563556432723999, + 0.006935286335647106, + 0.008429296314716339, + 0.00732216564938426, + 0.007457435596734285, + 0.0026748809032142162, + 0.0033500962890684605, + 0.004579934757202864, + 0.004324737936258316, + 0.0030735344626009464, + 0.004008032381534576, + 0.003001287579536438, + 0.004001494497060776, + 0.005796370562165976, + 0.00818149745464325, + 0.008123805746436119, + 0.00827737059444189, + 0.00856674276292324, + 0.011773216538131237, + 0.007652927190065384, + 0.009597561322152615, + 0.00264279218390584, + 0.006246402394026518, + 0.004958895966410637, + 0.00803305208683014, + 0.0022065669763833284, + 0.0022290197666734457, + 0.004054826218634844, + 0.00268269469961524, + 0.014053210616111755, + 0.008233902975916862, + 0.050358064472675323, + 0.02350609190762043, + 0.005119698122143745, + 0.0092660216614604, + 0.006870149169117212, + 0.02540031634271145, + 0.004117053467780352, + 0.004597350023686886, + 0.00477457232773304, + 0.019576428458094597, + 0.004313830751925707, + 0.006783101707696915, + 0.004155084025114775, + 0.024822410196065903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.012533940374851227 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734768394380808 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.006734768394380808, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734768394380808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03668515756726265, + 0.024703852832317352, + 0.0706842914223671, + 0.04385710880160332, + 0.18059709668159485, + 0.047810379415750504, + 0.07075632363557816, + 0.07736562192440033, + 0.2395721971988678, + 0.20796801149845123 + ], + "credit_map": [ + 0.0028889530804008245, + 0.00506806094199419, + 0.006884210743010044, + 0.005141021218150854, + 0.00486733065918088, + 0.004788239952176809, + 0.0035470621660351753, + 0.0035002748481929302, + 0.0026234544347971678, + 0.0026057525537908077, + 0.0035483362153172493, + 0.004072509240359068, + 0.0034405263140797615, + 0.0027562291361391544, + 0.0028761206194758415, + 0.002780919661745429, + 0.006940745748579502, + 0.0070373909547924995, + 0.009472642093896866, + 0.006164494436234236, + 0.009347401559352875, + 0.013342714868485928, + 0.008360443636775017, + 0.010018453001976013, + 0.003933631349354982, + 0.004642190877348185, + 0.005671618040651083, + 0.0047608669847249985, + 0.00591341033577919, + 0.008675027638673782, + 0.004265323746949434, + 0.0059950328432023525, + 0.02838989719748497, + 0.012529340572655201, + 0.07337968796491623, + 0.018319347873330116, + 0.013241357170045376, + 0.009603417478501797, + 0.015722863376140594, + 0.009411166422069073, + 0.003994240891188383, + 0.007979447022080421, + 0.008062883280217648, + 0.008957418613135815, + 0.0035285106860101223, + 0.00523296557366848, + 0.005753911100327969, + 0.004300996195524931, + 0.007436690386384726, + 0.011459069326519966, + 0.008631366305053234, + 0.011099410243332386, + 0.006970234680920839, + 0.009935913607478142, + 0.006689523812383413, + 0.008534103631973267, + 0.005636931862682104, + 0.013455124571919441, + 0.012314068153500557, + 0.018031619489192963, + 0.005248379893600941, + 0.005962766241282225, + 0.010648153722286224, + 0.006068573333323002, + 0.023540951311588287, + 0.014255935326218605, + 0.0843123197555542, + 0.03886992856860161, + 0.008239240385591984, + 0.015593988820910454, + 0.011040537618100643, + 0.043719250708818436, + 0.016786664724349976, + 0.011923073790967464, + 0.021037589758634567, + 0.09008686244487762, + 0.005198119208216667, + 0.007036656141281128, + 0.011549744755029678, + 0.04434928297996521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.006734768394380808 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-definition time-lapse unfolds from a static, slightly high-angle close-up to focus on an exquisite white porcelain water dropper perched on the dark, polished mahogany surface of a traditional scholar's desk. Shaped like a small, smooth gourd, the dropper features a tiny, elegant spout and boasts a flawless, glossy glaze, reflecting its surroundings with crystal clarity. As the timelapse progresses, a single, sharp beam of sunlight from an unseen window deliberately sweeps across the scene, its color temperature shifting noticeably from the cool, pale white of morning to the intense, bright glare of midday, finally fading into the warm, golden hues of late afternoon. The shadow cast by the dropper dramatically shortens, swirls, and then stretches across the rich texture, while the shadow of the window frame floats quietly over the background elements of ink washes and still calligraphic brushstrokes. With each subtle shift in the sun's arc, the gleaming porcelain surface catches the light, creating a mesmerizing pulsation as bright, mirrored highlights illuminate and gently fade, making the inanimate object feel alive. This silent, rhythmic play of light and shadow transforms a simple scene into a profound and meditative observation of the passage of time, imbuing the image with a deep tranquility and focused calm.", + "seed": 9181, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_5/index_6531_seed_9181_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09956664443016053, + "vlm_logprob": -0.09956664443016053, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.09956664443016053, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021363966166973114 + ], + "opposite_logprob_tokens": [ + -4.146363735198975 + ], + "logprob_reward": -0.021363966166973114, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.021363966166973114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1129222959280014, + 0.1491328328847885, + 0.17468677461147308, + 0.1359584778547287, + 0.06814726442098618, + 0.03062501735985279, + 0.1584417223930359, + 0.03813989832997322, + 0.0392826572060585, + 0.0926630049943924 + ], + "credit_map": [ + 0.015203799121081829, + 0.0134471096098423, + 0.02004089578986168, + 0.0229691993445158, + 0.01037529855966568, + 0.008891411125659943, + 0.012417932972311974, + 0.009576661512255669, + 0.011669330298900604, + 0.02332904003560543, + 0.02905120514333248, + 0.029263632372021675, + 0.011421113274991512, + 0.02311910130083561, + 0.010613368824124336, + 0.010666051879525185, + 0.014380332082509995, + 0.020670058205723763, + 0.023683680221438408, + 0.028440438210964203, + 0.017678501084446907, + 0.032604582607746124, + 0.023330677300691605, + 0.013898511417210102, + 0.009255683049559593, + 0.012575831264257431, + 0.03788640350103378, + 0.0102436114102602, + 0.007651480380445719, + 0.04305774345993996, + 0.006158153992146254, + 0.00912958662956953, + 0.0064517296850681305, + 0.008933068253099918, + 0.005970606114715338, + 0.008290434256196022, + 0.012486354447901249, + 0.010250214487314224, + 0.007524401415139437, + 0.00824046228080988, + 0.002247687429189682, + 0.005807796493172646, + 0.0027343621477484703, + 0.006406072061508894, + 0.0037304337602108717, + 0.004125760402530432, + 0.0029272802639752626, + 0.0026456282939761877, + 0.012361745350062847, + 0.024127928540110588, + 0.013792098499834538, + 0.03555317223072052, + 0.01782090775668621, + 0.02101466991007328, + 0.022488543763756752, + 0.011282670311629772, + 0.003665400203317404, + 0.0035851295106112957, + 0.00433952733874321, + 0.004991204012185335, + 0.005727911368012428, + 0.0044909571297466755, + 0.004605742171406746, + 0.006734029855579138, + 0.007486910093575716, + 0.005097941495478153, + 0.0029480771627277136, + 0.004850034136325121, + 0.006366677116602659, + 0.005249242763966322, + 0.003748919116333127, + 0.0035348606761544943, + 0.010240231640636921, + 0.01983516663312912, + 0.005376307759433985, + 0.012967578135430813, + 0.015309739857912064, + 0.010532451793551445, + 0.009610739536583424, + 0.008790797553956509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.021363966166973114 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048663295805454254 + ], + "opposite_logprob_tokens": [ + -3.048663377761841 + ], + "logprob_reward": -0.048663295805454254, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048663295805454254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10537977516651154, + 0.06467477977275848, + 0.1544533520936966, + 0.16090187430381775, + 0.06692515313625336, + 0.020460234954953194, + 0.05899810418486595, + 0.27746641635894775, + 0.025167806074023247, + 0.06557248532772064 + ], + "credit_map": [ + 0.009359009563922882, + 0.008880835957825184, + 0.02058877795934677, + 0.016513053327798843, + 0.010005274787545204, + 0.009558185003697872, + 0.02022632583975792, + 0.01024832297116518, + 0.005158574320375919, + 0.005154326092451811, + 0.015449435450136662, + 0.008499848656356335, + 0.005860568955540657, + 0.017381099984049797, + 0.004278947599232197, + 0.0028919787146151066, + 0.011733843013644218, + 0.018606750294566154, + 0.02092760056257248, + 0.024698879569768906, + 0.017432978376746178, + 0.02804148755967617, + 0.021356482058763504, + 0.011655336245894432, + 0.019046716392040253, + 0.01770135574042797, + 0.04146828502416611, + 0.0069760652258992195, + 0.0075853485614061356, + 0.05192122235894203, + 0.005056003108620644, + 0.011146889068186283, + 0.008511701598763466, + 0.010030772536993027, + 0.006039353087544441, + 0.008439545519649982, + 0.010461810976266861, + 0.008503027260303497, + 0.007367169018834829, + 0.007571772206574678, + 0.0014732179697602987, + 0.0018294270848855376, + 0.002161515410989523, + 0.002063893945887685, + 0.003550863591954112, + 0.002916099037975073, + 0.0019311110954731703, + 0.004534109961241484, + 0.005507476162165403, + 0.006639560218900442, + 0.006106581073254347, + 0.004947444424033165, + 0.009194992482662201, + 0.00898888148367405, + 0.007937254384160042, + 0.009675919078290462, + 0.021990815177559853, + 0.03502162545919418, + 0.032848138362169266, + 0.016612384468317032, + 0.026078464463353157, + 0.05700613930821419, + 0.025492612272500992, + 0.06241624802350998, + 0.0030042119324207306, + 0.0027176453731954098, + 0.0016471234848722816, + 0.0017982119461521506, + 0.005235925782471895, + 0.00403132289648056, + 0.003247383050620556, + 0.003485983470454812, + 0.006956125143915415, + 0.014254621230065823, + 0.0034874919801950455, + 0.007310344371944666, + 0.01252531073987484, + 0.007885664701461792, + 0.0064361002296209335, + 0.006716833915561438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.048663295805454254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25211814045906067 + ], + "opposite_logprob_tokens": [ + -1.5021181106567383 + ], + "logprob_reward": -0.25211814045906067, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25211814045906067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1376214474439621, + 0.11940828710794449, + 0.14917121827602386, + 0.07925382256507874, + 0.056071966886520386, + 0.0775817260146141, + 0.04371169954538345, + 0.275950163602829, + 0.027573341503739357, + 0.03365631029009819 + ], + "credit_map": [ + 0.008704321458935738, + 0.012067302130162716, + 0.020065534859895706, + 0.01959504373371601, + 0.014215144328773022, + 0.014506558887660503, + 0.032090477645397186, + 0.016377069056034088, + 0.00957889761775732, + 0.006598948035389185, + 0.018999967724084854, + 0.009545460343360901, + 0.02335108257830143, + 0.037351880222558975, + 0.008302092552185059, + 0.005679958034306765, + 0.008265540935099125, + 0.014153416268527508, + 0.024598395451903343, + 0.019946806132793427, + 0.016897691413760185, + 0.03879556804895401, + 0.01285899430513382, + 0.013654800131917, + 0.0068714129738509655, + 0.012723186984658241, + 0.016027526929974556, + 0.006688578054308891, + 0.005084419157356024, + 0.01855618692934513, + 0.004652822855859995, + 0.008649682626128197, + 0.007967652752995491, + 0.006737865507602692, + 0.008046245202422142, + 0.0048204585909843445, + 0.003791266120970249, + 0.006176782771945, + 0.004839273169636726, + 0.013692423701286316, + 0.007440914399921894, + 0.01061118021607399, + 0.00823210645467043, + 0.01371012069284916, + 0.0068447343073785305, + 0.009911799803376198, + 0.010455259121954441, + 0.01037560310214758, + 0.004172917455434799, + 0.007027399726212025, + 0.004300029948353767, + 0.005991089157760143, + 0.007356603164225817, + 0.005646027158945799, + 0.006367730442434549, + 0.002849897835403681, + 0.022552676498889923, + 0.03139173239469528, + 0.04396025463938713, + 0.015669003129005432, + 0.02538757212460041, + 0.058023933321237564, + 0.023143701255321503, + 0.055821262300014496, + 0.0026396051980555058, + 0.0046610296703875065, + 0.001546862768009305, + 0.0028050346300005913, + 0.00379472179338336, + 0.004185288678854704, + 0.0043050074018538, + 0.003635788569226861, + 0.003133670426905155, + 0.0054442137479782104, + 0.002209512982517481, + 0.0037631248123943806, + 0.0065319431014359, + 0.004556352272629738, + 0.004628111608326435, + 0.003389377612620592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.25211814045906067 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270456463098526 + ], + "opposite_logprob_tokens": [ + -2.1270456314086914 + ], + "logprob_reward": -0.1270456463098526, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270456463098526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20120878517627716, + 0.08975591510534286, + 0.14015920460224152, + 0.2681906819343567, + 0.07763697952032089, + 0.0398881658911705, + 0.04023507237434387, + 0.01724456250667572, + 0.026966935023665428, + 0.09871366620063782 + ], + "credit_map": [ + 0.01932978630065918, + 0.020725583657622337, + 0.03567039221525192, + 0.035532113164663315, + 0.021875645965337753, + 0.019520573318004608, + 0.02986757457256317, + 0.018687114119529724, + 0.009800335392355919, + 0.015695536509156227, + 0.012745938263833523, + 0.017975924536585808, + 0.006576213985681534, + 0.00808133464306593, + 0.011156841181218624, + 0.007723791059106588, + 0.012799495831131935, + 0.015378203243017197, + 0.02045920491218567, + 0.018332378938794136, + 0.013368040323257446, + 0.0318751223385334, + 0.015295454300940037, + 0.01265130564570427, + 0.017064237967133522, + 0.023510480299592018, + 0.08180458843708038, + 0.01290124747902155, + 0.013423738069832325, + 0.09325405210256577, + 0.006787539459764957, + 0.019444793462753296, + 0.007489480078220367, + 0.011134803295135498, + 0.007397313602268696, + 0.007629695348441601, + 0.014762619510293007, + 0.011262352578341961, + 0.007931985892355442, + 0.010028728283941746, + 0.0036035492084920406, + 0.004361976403743029, + 0.003289321670308709, + 0.00606851140037179, + 0.006176925264298916, + 0.006502625532448292, + 0.004719546064734459, + 0.005165711510926485, + 0.003242070786654949, + 0.005016378127038479, + 0.004463819321244955, + 0.005541892722249031, + 0.007010741159319878, + 0.0051816245540976524, + 0.005104452837258577, + 0.004674093797802925, + 0.002114916918799281, + 0.002680423203855753, + 0.0022550546564161777, + 0.001901299343444407, + 0.0027573222760111094, + 0.0015765668358653784, + 0.002345829037949443, + 0.001613150117918849, + 0.004162115976214409, + 0.0034058475866913795, + 0.0021546182688325644, + 0.003117148531600833, + 0.0030226062517613173, + 0.0038007074035704136, + 0.0034126569516956806, + 0.0038912342861294746, + 0.010840769857168198, + 0.02115393802523613, + 0.0054151033982634544, + 0.012856720946729183, + 0.016899310052394867, + 0.011304248124361038, + 0.010901082307100296, + 0.00934249721467495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.1270456463098526 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048642173409461975 + ], + "opposite_logprob_tokens": [ + -3.048642158508301 + ], + "logprob_reward": -0.048642173409461975, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048642173409461975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058233730494976044, + 0.31286799907684326, + 0.1522097885608673, + 0.1472572684288025, + 0.03662749007344246, + 0.025937628000974655, + 0.033485181629657745, + 0.03191770985722542, + 0.06948733329772949, + 0.13197579979896545 + ], + "credit_map": [ + 0.005877303425222635, + 0.0078491335734725, + 0.008394544012844563, + 0.011416416615247726, + 0.0067212823778390884, + 0.0056863767094910145, + 0.007048184517771006, + 0.005240486469119787, + 0.03616739436984062, + 0.04779498279094696, + 0.04833589494228363, + 0.05389125645160675, + 0.04676998034119606, + 0.02806202694773674, + 0.02508612722158432, + 0.02676033042371273, + 0.010300823487341404, + 0.016081763431429863, + 0.021989626809954643, + 0.022961026057600975, + 0.01767658069729805, + 0.030318226665258408, + 0.019988836720585823, + 0.01289290376007557, + 0.01837330311536789, + 0.02170071378350258, + 0.031516436487436295, + 0.007915972732007504, + 0.009268376976251602, + 0.03986665606498718, + 0.005996611434966326, + 0.012619185261428356, + 0.007392827421426773, + 0.005538653582334518, + 0.004883092362433672, + 0.004369725473225117, + 0.0032727362122386694, + 0.00404850160703063, + 0.0023213999811559916, + 0.0048005543649196625, + 0.0017815984319895506, + 0.0032002374064177275, + 0.0025342244189232588, + 0.0031690725591033697, + 0.004154282622039318, + 0.0037785011809319258, + 0.002513090381398797, + 0.004806620068848133, + 0.004123196471482515, + 0.004169282969087362, + 0.003282112767919898, + 0.0037889908999204636, + 0.005290471948683262, + 0.0048654391430318356, + 0.0041974447667598724, + 0.0037682433612644672, + 0.0040698484517633915, + 0.0041555906645953655, + 0.005190412513911724, + 0.0028975405730307102, + 0.003865450620651245, + 0.0033422738779336214, + 0.003531327238306403, + 0.004865264054387808, + 0.013130906037986279, + 0.009172158315777779, + 0.005122094415128231, + 0.006378003861755133, + 0.008816703222692013, + 0.009368853643536568, + 0.007355724927037954, + 0.010142887011170387, + 0.012951594777405262, + 0.027373062446713448, + 0.0069186934269964695, + 0.01652703620493412, + 0.02507008984684944, + 0.01702469028532505, + 0.013977465219795704, + 0.012133179232478142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.048642173409461975 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-definition time-lapse unfolds from a static, slightly high-angle close-up to focus on an exquisite white porcelain water dropper perched on the dark, polished mahogany surface of a traditional scholar's desk. Shaped like a small, smooth gourd, the dropper features a tiny, elegant spout and boasts a flawless, glossy glaze, reflecting its surroundings with crystal clarity. As the timelapse progresses, a single, sharp beam of sunlight from an unseen window deliberately sweeps across the scene, its color temperature shifting noticeably from the cool, pale white of morning to the intense, bright glare of midday, finally fading into the warm, golden hues of late afternoon. The shadow cast by the dropper dramatically shortens, swirls, and then stretches across the rich texture, while the shadow of the window frame floats quietly over the background elements of ink washes and still calligraphic brushstrokes. With each subtle shift in the sun's arc, the gleaming porcelain surface catches the light, creating a mesmerizing pulsation as bright, mirrored highlights illuminate and gently fade, making the inanimate object feel alive. This silent, rhythmic play of light and shadow transforms a simple scene into a profound and meditative observation of the passage of time, imbuing the image with a deep tranquility and focused calm.", + "seed": 9182, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_5/index_6531_seed_9182_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.054358876682817935, + "vlm_logprob": -0.054358876682817935, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.054358876682817935, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009235164150595665 + ], + "opposite_logprob_tokens": [ + -6.009235382080078 + ], + "logprob_reward": -0.009235164150595665, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.009235164150595665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04841809347271919, + 0.03244059905409813, + 0.02744980715215206, + 0.10412343591451645, + 0.3603016436100006, + 0.050371892750263214, + 0.03218689188361168, + 0.04022383317351341, + 0.12409010529518127, + 0.180393785238266 + ], + "credit_map": [ + 0.007229646202176809, + 0.004047088790684938, + 0.007163171656429768, + 0.005458363797515631, + 0.004901613108813763, + 0.006609692703932524, + 0.005574128590524197, + 0.007434382103383541, + 0.002538381377235055, + 0.0022000903263688087, + 0.005922000389546156, + 0.0029269568622112274, + 0.0038220195565372705, + 0.004621041473001242, + 0.003991968929767609, + 0.006418135017156601, + 0.0025300036650151014, + 0.003115852829068899, + 0.005379998125135899, + 0.0029852965380996466, + 0.0032945817802101374, + 0.0031639044173061848, + 0.002494265791028738, + 0.004485900513827801, + 0.010000569745898247, + 0.010699880309402943, + 0.015916023403406143, + 0.00928308442234993, + 0.014630993828177452, + 0.015045557171106339, + 0.01274083461612463, + 0.01580647937953472, + 0.05128501355648041, + 0.05762877315282822, + 0.0281204991042614, + 0.03184932842850685, + 0.05919888615608215, + 0.053421832621097565, + 0.034973692148923874, + 0.04382358863949776, + 0.004085344262421131, + 0.004928684793412685, + 0.004247196950018406, + 0.006977658253163099, + 0.007375478278845549, + 0.004961439874023199, + 0.009419337846338749, + 0.00837674643844366, + 0.002626921748742461, + 0.004155317321419716, + 0.0035637104883790016, + 0.003716283245012164, + 0.005391254089772701, + 0.00385897234082222, + 0.005032071843743324, + 0.003842355217784643, + 0.003954789135605097, + 0.005152109079062939, + 0.005548433866351843, + 0.0040977331809699535, + 0.007358977105468512, + 0.003750202711671591, + 0.005628838203847408, + 0.004732747562229633, + 0.009504612535238266, + 0.01858445443212986, + 0.01703222468495369, + 0.013541070744395256, + 0.022013351321220398, + 0.012773414142429829, + 0.017864830791950226, + 0.01277613639831543, + 0.017005274072289467, + 0.017845019698143005, + 0.028434285894036293, + 0.013276834972202778, + 0.026558788493275642, + 0.04521334543824196, + 0.011067034676671028, + 0.0209931880235672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.009235164150595665 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674376729875803 + ], + "opposite_logprob_tokens": [ + -5.006743907928467 + ], + "logprob_reward": -0.00674376729875803, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674376729875803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08549489825963974, + 0.09046053886413574, + 0.051565542817115784, + 0.06837811321020126, + 0.5156394243240356, + 0.039193812757730484, + 0.033799946308135986, + 0.040194328874349594, + 0.028531445190310478, + 0.046741969883441925 + ], + "credit_map": [ + 0.012055044062435627, + 0.006116208154708147, + 0.020241230726242065, + 0.007246627938002348, + 0.006576185114681721, + 0.0076143029145896435, + 0.010175803676247597, + 0.015469495207071304, + 0.008176551200449467, + 0.005192725919187069, + 0.01981077715754509, + 0.009531835094094276, + 0.007786680478602648, + 0.010110958479344845, + 0.01056352723389864, + 0.019287481904029846, + 0.006770546082407236, + 0.0038194814696907997, + 0.00827314704656601, + 0.006862732581794262, + 0.004270859993994236, + 0.005672208499163389, + 0.007251022383570671, + 0.008645547553896904, + 0.007506937254220247, + 0.006588777061551809, + 0.009320023469626904, + 0.006703366059809923, + 0.00952825415879488, + 0.008737701922655106, + 0.009817793034017086, + 0.010175257921218872, + 0.07439258694648743, + 0.08265181630849838, + 0.04081097990274429, + 0.044656239449977875, + 0.0851185992360115, + 0.07626765966415405, + 0.04893465340137482, + 0.0628068670630455, + 0.004116343334317207, + 0.004379773512482643, + 0.0034910670947283506, + 0.004290974698960781, + 0.006235038861632347, + 0.004364794585853815, + 0.006770185194909573, + 0.005545636638998985, + 0.0041956715285778046, + 0.003576367162168026, + 0.003737544408068061, + 0.003808386391028762, + 0.005053384229540825, + 0.004002995789051056, + 0.005158873740583658, + 0.004266721196472645, + 0.0039063249714672565, + 0.004874805454164743, + 0.00504763750359416, + 0.004926780238747597, + 0.006901511922478676, + 0.003520278725773096, + 0.006338709965348244, + 0.0046782782301306725, + 0.00273995753377676, + 0.003514372743666172, + 0.002645829925313592, + 0.003117701970040798, + 0.005132591817528009, + 0.0045084040611982346, + 0.0030159850139170885, + 0.0038566032890230417, + 0.006369804497808218, + 0.004227288533002138, + 0.007407516706734896, + 0.0036048111505806446, + 0.005477230530232191, + 0.005259343888610601, + 0.0030821028631180525, + 0.01131387148052454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.00674376729875803 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155629515647888 + ], + "opposite_logprob_tokens": [ + -1.7015563249588013 + ], + "logprob_reward": -0.20155629515647888, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155629515647888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14913977682590485, + 0.045142583549022675, + 0.020207365974783897, + 0.04933707416057587, + 0.1545064002275467, + 0.04000962898135185, + 0.020050685852766037, + 0.07885821163654327, + 0.2559713125228882, + 0.18677696585655212 + ], + "credit_map": [ + 0.028186386451125145, + 0.008932825177907944, + 0.030885040760040283, + 0.011726968921720982, + 0.01050817221403122, + 0.011157434433698654, + 0.013571555726230145, + 0.03417138010263443, + 0.0052503012120723724, + 0.0033619629684835672, + 0.008690809831023216, + 0.003083698684349656, + 0.0050465757958590984, + 0.003693319857120514, + 0.005276153329759836, + 0.010739765129983425, + 0.003356192260980606, + 0.0013753528473898768, + 0.0036429238971322775, + 0.0013674269430339336, + 0.0021441723220050335, + 0.0023740604519844055, + 0.0025460433680564165, + 0.0034011914394795895, + 0.005020535085350275, + 0.004271123092621565, + 0.007692967541515827, + 0.004428033716976643, + 0.006278340704739094, + 0.007929765619337559, + 0.0063232057727873325, + 0.00739310123026371, + 0.020972300320863724, + 0.022402813658118248, + 0.0161052905023098, + 0.01291506364941597, + 0.023741543292999268, + 0.02442888729274273, + 0.014023397117853165, + 0.019917119294404984, + 0.0032974707428365946, + 0.008399063721299171, + 0.004871929995715618, + 0.0045332047156989574, + 0.005060260649770498, + 0.004586797673255205, + 0.004429668188095093, + 0.004831234458833933, + 0.0024943898897618055, + 0.0031884789932519197, + 0.002927527530118823, + 0.0021870238706469536, + 0.0024495788384228945, + 0.0022457363083958626, + 0.002512847539037466, + 0.0020451031159609556, + 0.008164436556398869, + 0.01115909218788147, + 0.010004314593970776, + 0.007627531420439482, + 0.013540251180529594, + 0.008711867965757847, + 0.010429968126118183, + 0.009220756590366364, + 0.017213664948940277, + 0.03798355907201767, + 0.03512395918369293, + 0.02793487347662449, + 0.04811866581439972, + 0.02798396535217762, + 0.03387601673603058, + 0.027736615389585495, + 0.01918512023985386, + 0.015402749180793762, + 0.03235286846756935, + 0.013912200927734375, + 0.023559628054499626, + 0.04807526245713234, + 0.010431822389364243, + 0.02385728992521763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.20155629515647888 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368911147117615 + ], + "opposite_logprob_tokens": [ + -3.408689022064209 + ], + "logprob_reward": -0.03368911147117615, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368911147117615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07977986335754395, + 0.03488992899656296, + 0.10492539405822754, + 0.501889169216156, + 0.14106978476047516, + 0.03714274242520332, + 0.019624805077910423, + 0.024604111909866333, + 0.03879339620471001, + 0.01728074438869953 + ], + "credit_map": [ + 0.0089961476624012, + 0.006415687035769224, + 0.016696114093065262, + 0.007836131379008293, + 0.008500364609062672, + 0.006158995907753706, + 0.013059325516223907, + 0.01211710274219513, + 0.004823207855224609, + 0.0018810877809301019, + 0.007447629701346159, + 0.0017236927524209023, + 0.0025896839797496796, + 0.0034852363169193268, + 0.002860956359654665, + 0.010078434832394123, + 0.011182554066181183, + 0.004940958693623543, + 0.02416187897324562, + 0.011437499895691872, + 0.007592205889523029, + 0.01197217870503664, + 0.012815194204449654, + 0.020822923630475998, + 0.05348479002714157, + 0.04638867452740669, + 0.0717100277543068, + 0.04536993056535721, + 0.06799188256263733, + 0.07685016840696335, + 0.0634210854768753, + 0.07667262852191925, + 0.01942555420100689, + 0.021991580724716187, + 0.011328221298754215, + 0.011905990540981293, + 0.022902440279722214, + 0.022279979661107063, + 0.013295041397213936, + 0.017940975725650787, + 0.0034870142117142677, + 0.0054379114881157875, + 0.003615858731791377, + 0.0048244185745716095, + 0.005163982044905424, + 0.0043986402451992035, + 0.00525969872251153, + 0.004955214448273182, + 0.002594108460471034, + 0.002838543150573969, + 0.003059558803215623, + 0.0019580970983952284, + 0.002776334062218666, + 0.002190688159316778, + 0.002297633094713092, + 0.0019098414340987802, + 0.0025773365050554276, + 0.0036889538168907166, + 0.0029888316057622433, + 0.0030207557138055563, + 0.0035846822429448366, + 0.0023324869107455015, + 0.0035007663536816835, + 0.002910299226641655, + 0.0029911347664892673, + 0.005543148145079613, + 0.0054422966204583645, + 0.004136288538575172, + 0.0073845405131578445, + 0.0038449259009212255, + 0.005262172780930996, + 0.004188889171928167, + 0.0019210013560950756, + 0.0016877446323633194, + 0.0024600320030003786, + 0.001492639072239399, + 0.0024895397946238518, + 0.0031054692808538675, + 0.001254052622243762, + 0.0028702670242637396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.03368911147117615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020570045337080956 + ], + "opposite_logprob_tokens": [ + -3.8955700397491455 + ], + "logprob_reward": -0.020570045337080956, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020570045337080956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1005585789680481, + 0.03943765163421631, + 0.058741096407175064, + 0.05286955088376999, + 0.13424114882946014, + 0.1700705736875534, + 0.04018282890319824, + 0.0807928517460823, + 0.22509697079658508, + 0.09800871461629868 + ], + "credit_map": [ + 0.01397673785686493, + 0.005494283512234688, + 0.022180777043104172, + 0.00884604174643755, + 0.006472827401012182, + 0.014031349681317806, + 0.01089321356266737, + 0.018663346767425537, + 0.003377442480996251, + 0.0020278559532016516, + 0.008171902038156986, + 0.002680410398170352, + 0.0032279326114803553, + 0.004179898649454117, + 0.005671140272170305, + 0.010101068764925003, + 0.005628006532788277, + 0.0030584183987230062, + 0.0114379758015275, + 0.007483717519789934, + 0.004819538444280624, + 0.005605206824839115, + 0.009806517511606216, + 0.01090171653777361, + 0.004352688789367676, + 0.003488237503916025, + 0.012778904289007187, + 0.0056121740490198135, + 0.004307410679757595, + 0.004316302016377449, + 0.005781053565442562, + 0.012232777662575245, + 0.020570626482367516, + 0.020505841821432114, + 0.010636171326041222, + 0.012964772991836071, + 0.021523380652070045, + 0.017027132213115692, + 0.015224114060401917, + 0.015789106488227844, + 0.013724583201110363, + 0.019587518647313118, + 0.014174613170325756, + 0.03209739923477173, + 0.0190754197537899, + 0.01950831152498722, + 0.025569183751940727, + 0.02633354254066944, + 0.0027143394108861685, + 0.0040219128131866455, + 0.004834811203181744, + 0.005224253516644239, + 0.006531061604619026, + 0.004785068798810244, + 0.006804768927395344, + 0.0052666133269667625, + 0.006201051641255617, + 0.00955218356102705, + 0.009609020315110683, + 0.008952075615525246, + 0.014906167984008789, + 0.00780489481985569, + 0.01345357857644558, + 0.01031387597322464, + 0.022736886516213417, + 0.028824273496866226, + 0.022321876138448715, + 0.03013267181813717, + 0.044544149190187454, + 0.02332783304154873, + 0.029529234394431114, + 0.023680049926042557, + 0.014134940691292286, + 0.011141354218125343, + 0.023161640390753746, + 0.008307471871376038, + 0.00729881227016449, + 0.011359251104295254, + 0.00536586157977581, + 0.017239386215806007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.020570045337080956 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-definition time-lapse unfolds from a static, slightly high-angle close-up to focus on an exquisite white porcelain water dropper perched on the dark, polished mahogany surface of a traditional scholar's desk. Shaped like a small, smooth gourd, the dropper features a tiny, elegant spout and boasts a flawless, glossy glaze, reflecting its surroundings with crystal clarity. As the timelapse progresses, a single, sharp beam of sunlight from an unseen window deliberately sweeps across the scene, its color temperature shifting noticeably from the cool, pale white of morning to the intense, bright glare of midday, finally fading into the warm, golden hues of late afternoon. The shadow cast by the dropper dramatically shortens, swirls, and then stretches across the rich texture, while the shadow of the window frame floats quietly over the background elements of ink washes and still calligraphic brushstrokes. With each subtle shift in the sun's arc, the gleaming porcelain surface catches the light, creating a mesmerizing pulsation as bright, mirrored highlights illuminate and gently fade, making the inanimate object feel alive. This silent, rhythmic play of light and shadow transforms a simple scene into a profound and meditative observation of the passage of time, imbuing the image with a deep tranquility and focused calm.", + "seed": 9183, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_5/index_6531_seed_9183_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2221583787817508, + "vlm_logprob": -0.2221583787817508, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.2221583787817508, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01438003871589899 + ], + "opposite_logprob_tokens": [ + -4.889379978179932 + ], + "logprob_reward": -0.01438003871589899, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01438003871589899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7146472930908203, + 0.02543715201318264, + 0.02077372372150421, + 0.03555523231625557, + 0.017467191442847252, + 0.05321485176682472, + 0.020959563553333282, + 0.0728575736284256, + 0.011517905630171299, + 0.027569418773055077 + ], + "credit_map": [ + 0.16023947298526764, + 0.3079777657985687, + 0.04096477851271629, + 0.02884344942867756, + 0.05789153277873993, + 0.07839535921812057, + 0.020786281675100327, + 0.01954873465001583, + 0.0046089268289506435, + 0.006558991968631744, + 0.002844415372237563, + 0.001751220435835421, + 0.0025267673190683126, + 0.0023075651843100786, + 0.0026743486523628235, + 0.0021649168338626623, + 0.002221924951300025, + 0.003225685330107808, + 0.0019001471810042858, + 0.0013314024545252323, + 0.0019002959597855806, + 0.0056827557273209095, + 0.0015187845565378666, + 0.0029927296563982964, + 0.003749406896531582, + 0.00470293452963233, + 0.003942003473639488, + 0.003597761271521449, + 0.0052646854892373085, + 0.00307885161601007, + 0.008920490741729736, + 0.0022991017904132605, + 0.0023775582667440176, + 0.00483822962269187, + 0.0020239900331944227, + 0.0012490152148529887, + 0.0017585456371307373, + 0.0018019848503172398, + 0.0022891699336469173, + 0.0011286998633295298, + 0.008979837410151958, + 0.019703147932887077, + 0.004602357279509306, + 0.002545863389968872, + 0.005461128894239664, + 0.004695802461355925, + 0.00411873497068882, + 0.003107981523498893, + 0.00392083590850234, + 0.0048798806965351105, + 0.002171093598008156, + 0.0015026162145659328, + 0.0021315678022801876, + 0.002044185297563672, + 0.0018524276092648506, + 0.002456958871334791, + 0.005014115944504738, + 0.014100013300776482, + 0.007968340069055557, + 0.005490491166710854, + 0.0045736790634691715, + 0.004062281455844641, + 0.0054702977649867535, + 0.02617836184799671, + 0.0009151333943009377, + 0.0019051702693104744, + 0.0011735431617125869, + 0.0012553957058116794, + 0.0014026318676769733, + 0.0012956118443980813, + 0.002122059464454651, + 0.0014483608538284898, + 0.0028163103852421045, + 0.005310231354087591, + 0.003098041983321309, + 0.0027424325235188007, + 0.003748283488675952, + 0.0032216054387390614, + 0.0034185070544481277, + 0.0032140109688043594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.01438003871589899 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550657615065575 + ], + "opposite_logprob_tokens": [ + -4.3875508308410645 + ], + "logprob_reward": -0.012550657615065575, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550657615065575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2521730363368988, + 0.12702667713165283, + 0.0546136200428009, + 0.05969777703285217, + 0.09236864000558853, + 0.07259654253721237, + 0.21863427758216858, + 0.040354106575250626, + 0.04910879582166672, + 0.033426519483327866 + ], + "credit_map": [ + 0.034865602850914, + 0.08575402200222015, + 0.019980374723672867, + 0.015095294453203678, + 0.02640771120786667, + 0.026837246492505074, + 0.021183816716074944, + 0.02204897440969944, + 0.017332881689071655, + 0.02907276526093483, + 0.014064105227589607, + 0.009141504764556885, + 0.012271604500710964, + 0.01780521124601364, + 0.012317678891122341, + 0.015020920895040035, + 0.007365906145423651, + 0.007693126332014799, + 0.005669570527970791, + 0.004404445644468069, + 0.006132008042186499, + 0.01190266665071249, + 0.006870744749903679, + 0.004575151018798351, + 0.0063782320357859135, + 0.009089041501283646, + 0.0063221389427781105, + 0.003902398282662034, + 0.008685961365699768, + 0.010994364507496357, + 0.009313579648733139, + 0.005012063775211573, + 0.013239248655736446, + 0.019481081515550613, + 0.011193506419658661, + 0.008623216301202774, + 0.010060730390250683, + 0.008659555576741695, + 0.012174909934401512, + 0.008936387486755848, + 0.006019825115799904, + 0.013148175552487373, + 0.012138724327087402, + 0.007902145385742188, + 0.009561638347804546, + 0.00759712141007185, + 0.010066528804600239, + 0.006162382662296295, + 0.04603153467178345, + 0.07382908463478088, + 0.019072290509939194, + 0.009453658945858479, + 0.02118280716240406, + 0.018470369279384613, + 0.01438702829182148, + 0.016207506880164146, + 0.006427241489291191, + 0.013816507533192635, + 0.003212001407518983, + 0.0028038471937179565, + 0.0036366074346005917, + 0.0025675331708043814, + 0.0033914244268089533, + 0.004498946014791727, + 0.006603774148970842, + 0.006184668745845556, + 0.0059694936498999596, + 0.003942055162042379, + 0.0039624907076358795, + 0.00455879233777523, + 0.003808110486716032, + 0.014079410582780838, + 0.0038115670904517174, + 0.007231604773551226, + 0.00377169088460505, + 0.0028414495754987, + 0.003215803299099207, + 0.00453949486836791, + 0.004116592463105917, + 0.003898319322615862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.012550657615065575 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539034605026245 + ], + "opposite_logprob_tokens": [ + -0.4289034605026245 + ], + "logprob_reward": -1.0539034605026245, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539034605026245, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.6632603406906128, + 0.02554033137857914, + 0.012841342948377132, + 0.014659109525382519, + 0.022202707827091217, + 0.09110432118177414, + 0.04101697728037834, + 0.10262002050876617, + 0.013187982141971588, + 0.013566907495260239 + ], + "credit_map": [ + 0.14675214886665344, + 0.2853166460990906, + 0.038298141211271286, + 0.02647864632308483, + 0.05388375744223595, + 0.07314056903123856, + 0.019999973475933075, + 0.01939043216407299, + 0.004398177843540907, + 0.00469968980178237, + 0.003758829552680254, + 0.0023807999677956104, + 0.0024978318251669407, + 0.0022711558267474174, + 0.0025918942410498857, + 0.0029419504571706057, + 0.0013505431124940515, + 0.0024532347451895475, + 0.001574073568917811, + 0.000950125337112695, + 0.0013118124334141612, + 0.0022080158814787865, + 0.0016089840792119503, + 0.001384551520459354, + 0.002802536590024829, + 0.003477469552308321, + 0.0014756947057321668, + 0.0012025132309645414, + 0.001420894404873252, + 0.0015783042181283236, + 0.0018339441157877445, + 0.0008677512523718178, + 0.003598604118451476, + 0.006763724144548178, + 0.0022377909626811743, + 0.0013433401472866535, + 0.002097922144457698, + 0.0025449232198297977, + 0.0022819917649030685, + 0.0013344089966267347, + 0.013252036646008492, + 0.031147871166467667, + 0.008010730147361755, + 0.004847055766731501, + 0.010140925645828247, + 0.009412001818418503, + 0.007756755221635103, + 0.006536941044032574, + 0.006325980182737112, + 0.01516913902014494, + 0.003226993838325143, + 0.0021045017056167126, + 0.004314427729696035, + 0.003214253345504403, + 0.004598331172019243, + 0.0020633472595363855, + 0.014768977649509907, + 0.0076115014962852, + 0.013607881031930447, + 0.008540141396224499, + 0.0038146288134157658, + 0.004602779168635607, + 0.0058102854527533054, + 0.04386381804943085, + 0.0017106137238442898, + 0.001418468775227666, + 0.0018106692004948854, + 0.0011345671955496073, + 0.001031136722303927, + 0.00103173463139683, + 0.0010527329286560416, + 0.003998058382421732, + 0.0009425312164239585, + 0.0018677347106859088, + 0.0018533830298110843, + 0.0012411022325977683, + 0.0017541402485221624, + 0.0011102566495537758, + 0.0021339785307645798, + 0.002663778606802225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -1.0539034605026245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343176141381264 + ], + "opposite_logprob_tokens": [ + -3.6513431072235107 + ], + "logprob_reward": -0.026343176141381264, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343176141381264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2731209099292755, + 0.03000338189303875, + 0.08481001108884811, + 0.0327996127307415, + 0.0820220410823822, + 0.10984078049659729, + 0.08174841850996017, + 0.10929857194423676, + 0.10191426426172256, + 0.09444205462932587 + ], + "credit_map": [ + 0.06774844229221344, + 0.10651598125696182, + 0.017579779028892517, + 0.01156801637262106, + 0.020122943446040154, + 0.029276741668581963, + 0.010514098219573498, + 0.00979489553719759, + 0.004125605337321758, + 0.006081029772758484, + 0.003956095781177282, + 0.0025989115238189697, + 0.003414392238482833, + 0.0034500700421631336, + 0.0035899376962333918, + 0.002787340199574828, + 0.010966606438159943, + 0.011768899857997894, + 0.006667112931609154, + 0.004440621007233858, + 0.008439595811069012, + 0.031730081886053085, + 0.003203250700607896, + 0.007593845948576927, + 0.002859418513253331, + 0.006319067440927029, + 0.0040131667628884315, + 0.0021758838556706905, + 0.004211501218378544, + 0.003454721299931407, + 0.006529290694743395, + 0.0032365662045776844, + 0.014596867375075817, + 0.025502335280179977, + 0.007040008902549744, + 0.005504406988620758, + 0.009618151932954788, + 0.005169035866856575, + 0.009368879720568657, + 0.005222357343882322, + 0.01551977638155222, + 0.039012715220451355, + 0.009480861946940422, + 0.005633679684251547, + 0.01233121007680893, + 0.010390905663371086, + 0.009464377537369728, + 0.008007257245481014, + 0.01747303269803524, + 0.03295633941888809, + 0.005442320369184017, + 0.0029553414788097143, + 0.007999586872756481, + 0.005745240021497011, + 0.005429787095636129, + 0.003746774746105075, + 0.005780171602964401, + 0.026255441829562187, + 0.011768486350774765, + 0.00826335046440363, + 0.007321774028241634, + 0.006329250521957874, + 0.009257543832063675, + 0.034322552382946014, + 0.011716381646692753, + 0.009583751671016216, + 0.013176844455301762, + 0.009096268564462662, + 0.006792976520955563, + 0.009144586510956287, + 0.008751466870307922, + 0.03365198150277138, + 0.008350745774805546, + 0.01616271212697029, + 0.013196148909628391, + 0.008807001635432243, + 0.011103243567049503, + 0.012159980833530426, + 0.010756544768810272, + 0.013905675150454044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.026343176141381264 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white porcelain water dropper shaped like a small, smooth gourd?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dropper placed on a dark, polished mahogany surface of a traditional scholar's desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a single beam of sunlight sweeping across the scene with color temperature shifting from cool morning white to warm late afternoon gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are background elements including ink washes and still calligraphic brushstrokes visible on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, slightly high-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036145609337836504 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.0036145609337836504, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036145609337836504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6842320561408997, + 0.01193265337496996, + 0.028256388381123543, + 0.012080444023013115, + 0.01554887369275093, + 0.02902122214436531, + 0.024518463760614395, + 0.10238301753997803, + 0.06559274345636368, + 0.026434190571308136 + ], + "credit_map": [ + 0.15099748969078064, + 0.29318130016326904, + 0.038499027490615845, + 0.027363182976841927, + 0.055390603840351105, + 0.07576637715101242, + 0.021503325551748276, + 0.02153071202337742, + 0.0014126390451565385, + 0.002194901229813695, + 0.0012342261616140604, + 0.0013284916058182716, + 0.0014883715193718672, + 0.0013023308711126447, + 0.0014183588791638613, + 0.0015533333644270897, + 0.0031868969090282917, + 0.007008739281445742, + 0.0024037412367761135, + 0.0016002776101231575, + 0.002311375457793474, + 0.006573551334440708, + 0.0023223068565130234, + 0.002849499462172389, + 0.0012419425183907151, + 0.0013040098128840327, + 0.0016841153847053647, + 0.0010505883255973458, + 0.0018063908210024238, + 0.0016245865263044834, + 0.0020950976759195328, + 0.001273712725378573, + 0.0028238375671207905, + 0.004728461615741253, + 0.0014245219063013792, + 0.0009274050826206803, + 0.0014911327743902802, + 0.001239531789906323, + 0.00169186491984874, + 0.0012221162905916572, + 0.0028228117153048515, + 0.00681894039735198, + 0.0037303436547517776, + 0.0025350104551762342, + 0.0040480950847268105, + 0.002181504387408495, + 0.004323272500187159, + 0.0025612444151192904, + 0.004784402437508106, + 0.008935718797147274, + 0.0018390932818874717, + 0.0011577538680285215, + 0.0023874694015830755, + 0.001899259863421321, + 0.001937482156790793, + 0.001577281393110752, + 0.012490122579038143, + 0.011125287041068077, + 0.013066108338534832, + 0.00838189572095871, + 0.004104312043637037, + 0.004353670869022608, + 0.005666544195264578, + 0.04319507256150246, + 0.008230719715356827, + 0.00662896828725934, + 0.00889258086681366, + 0.005524195730686188, + 0.0044532353058457375, + 0.005498070269823074, + 0.004847792908549309, + 0.021517179906368256, + 0.0026650228537619114, + 0.0050796037539839745, + 0.003116299631074071, + 0.0026721192989498377, + 0.0031836924608796835, + 0.0031286717858165503, + 0.0033543063327670097, + 0.003234471660107374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.0036145609337836504 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_6/index_tensor([3106, 3106])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_6/index_tensor([3106, 3106])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f3bf7ee7fa5e3965d388ec57288a57aa241ec872 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_6/index_tensor([3106, 3106])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up, static shot, shot from a slightly upward angle, focuses on a minimalist toilet brush, placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop, evoking the atmosphere of a high-end product photography studio. The brush features a slender, cylindrical handle made of matte white polymer. Its flawless, smooth surface reflects virtually no light under even, soft, diffuse lighting, while the brush head is composed of densely packed, uniformly arranged white synthetic fiber bristles. As the movement begins, a fiery, lava-like red glow quietly emanates from the very tips of the bristles, rapidly intensifying and steadily spreading upward along the length of the bristles. Captured in smooth, high-frame-rate footage, the once-sturdy white bristle structure instantly disintegrates wherever the red light passes, disintegrating inch by inch from bottom to top into countless, extremely fine black ash particles at a visible yet incredibly smooth speed. Instead of falling immediately, the embers gracefully drift and swirl on gentle currents before dissipating into the air. The entire process resembles a meticulously crafted CG special effects short film, imbued with a surreal visual beauty. The fiery red light becomes the sole dynamic source of light in the image, casting shifting, warm reflections on the matte white handle and dark background. This contrasts sharply with the initial, cool studio lighting, creating a unique atmosphere that is both tranquil and filled with devastating tension.", + "seed": 10061, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_6/index_3106_seed_10061_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05547601468861103, + "vlm_logprob": -0.05547601468861103, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.05547601468861103, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010608706623315811 + ], + "opposite_logprob_tokens": [ + -5.135608673095703 + ], + "logprob_reward": -0.010608706623315811, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.010608706623315811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19761396944522858, + 0.08095749467611313, + 0.07287529855966568, + 0.05700303614139557, + 0.08726071566343307, + 0.10226825624704361, + 0.0548546276986599, + 0.04207376390695572, + 0.1509118676185608, + 0.15418094396591187 + ], + "credit_map": [ + 0.027295706793665886, + 0.0356622114777565, + 0.03613971173763275, + 0.02679923176765442, + 0.010891959071159363, + 0.013857172802090645, + 0.022859778255224228, + 0.02410821057856083, + 0.012172619812190533, + 0.013436328619718552, + 0.015393358655273914, + 0.010088685899972916, + 0.005232645198702812, + 0.004233075771480799, + 0.011369072832167149, + 0.009031708352267742, + 0.011439280584454536, + 0.008761420845985413, + 0.010807746089994907, + 0.01036954764276743, + 0.004336318466812372, + 0.0035231022629886866, + 0.0100248446688056, + 0.01361303310841322, + 0.007331051863729954, + 0.006777619011700153, + 0.011700460687279701, + 0.007043634075671434, + 0.003765176748856902, + 0.002727038227021694, + 0.008898883126676083, + 0.00875917449593544, + 0.006082320585846901, + 0.026714636012911797, + 0.03008214756846428, + 0.004803613293915987, + 0.0029949075542390347, + 0.0033528178464621305, + 0.007672668434679508, + 0.005557607859373093, + 0.008867541328072548, + 0.021327413618564606, + 0.024020500481128693, + 0.00898479763418436, + 0.006443570833653212, + 0.007527174893766642, + 0.01597476191818714, + 0.009122498333454132, + 0.006116599775850773, + 0.008787555620074272, + 0.007139950525015593, + 0.0062250797636806965, + 0.0036953247617930174, + 0.004357881378382444, + 0.010030276142060757, + 0.008501959033310413, + 0.004604683257639408, + 0.007035349495708942, + 0.006378561723977327, + 0.005623364355415106, + 0.0030920167919248343, + 0.003104926785454154, + 0.006679898127913475, + 0.005554963834583759, + 0.020879404619336128, + 0.02378687635064125, + 0.024111250415444374, + 0.018761321902275085, + 0.012780416756868362, + 0.01583665795624256, + 0.014353567734360695, + 0.020402366295456886, + 0.026281047612428665, + 0.022122159600257874, + 0.028365088626742363, + 0.014028824865818024, + 0.0119704008102417, + 0.011149169877171516, + 0.017562588676810265, + 0.02270166575908661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.010608706623315811 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016092298552393913 + ], + "opposite_logprob_tokens": [ + -4.141092300415039 + ], + "logprob_reward": -0.016092298552393913, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016092298552393913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20449796319007874, + 0.09963494539260864, + 0.13008172810077667, + 0.06578358262777328, + 0.14083649218082428, + 0.047270625829696655, + 0.03366474062204361, + 0.03509170189499855, + 0.09761114418506622, + 0.14552709460258484 + ], + "credit_map": [ + 0.03652743995189667, + 0.032584842294454575, + 0.044135212898254395, + 0.024110548198223114, + 0.008068425580859184, + 0.011045664548873901, + 0.018846767023205757, + 0.029179051518440247, + 0.016939105466008186, + 0.012591774575412273, + 0.020845826715230942, + 0.010787996463477612, + 0.0051636407151818275, + 0.00578879751265049, + 0.014681553468108177, + 0.012836240231990814, + 0.02089017815887928, + 0.01378471590578556, + 0.019649507477879524, + 0.014136429876089096, + 0.008863848634064198, + 0.011274768970906734, + 0.022749586030840874, + 0.01873268187046051, + 0.008455055765807629, + 0.007877493277192116, + 0.010944459587335587, + 0.006947133224457502, + 0.004601746331900358, + 0.003952247090637684, + 0.013675161637365818, + 0.009330276399850845, + 0.010695423930883408, + 0.04161835089325905, + 0.04607878997921944, + 0.008258393034338951, + 0.005349469371140003, + 0.005028590559959412, + 0.011983857490122318, + 0.01182362250983715, + 0.0047866650857031345, + 0.008335724472999573, + 0.008830021135509014, + 0.005166736431419849, + 0.003080862108618021, + 0.0031260286923497915, + 0.008120539598166943, + 0.005824048537760973, + 0.00417118426412344, + 0.0065728528425097466, + 0.004903097171336412, + 0.0030167759396135807, + 0.0024555178824812174, + 0.002647734945639968, + 0.004678579978644848, + 0.00521900225430727, + 0.004862286150455475, + 0.006227674428373575, + 0.005389474797993898, + 0.003067637560889125, + 0.0024458859115839005, + 0.003747302107512951, + 0.004674382973462343, + 0.004677062388509512, + 0.012837539426982403, + 0.01725251041352749, + 0.014506726525723934, + 0.012459255754947662, + 0.007894454523921013, + 0.00835858378559351, + 0.01170546468347311, + 0.012596607208251953, + 0.024104535579681396, + 0.02626059763133526, + 0.028401153162121773, + 0.014402929693460464, + 0.011097421869635582, + 0.008680470287799835, + 0.015045594424009323, + 0.017534403130412102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.016092298552393913 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547243386507034 + ], + "opposite_logprob_tokens": [ + -4.387547016143799 + ], + "logprob_reward": -0.012547243386507034, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547243386507034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12692801654338837, + 0.0489252470433712, + 0.03179508447647095, + 0.020213494077324867, + 0.03264015540480614, + 0.036879222840070724, + 0.059944070875644684, + 0.04969233646988869, + 0.42874810099601746, + 0.16423426568508148 + ], + "credit_map": [ + 0.014729620888829231, + 0.016092313453555107, + 0.022222494706511497, + 0.019513564184308052, + 0.006890038959681988, + 0.01087581180036068, + 0.02125464752316475, + 0.015349536202847958, + 0.0053095449693500996, + 0.006304020527750254, + 0.008573696948587894, + 0.007296854164451361, + 0.003233677241951227, + 0.004214663524180651, + 0.007565184962004423, + 0.006427607499063015, + 0.003576745046302676, + 0.005041572730988264, + 0.006930784322321415, + 0.003570332657545805, + 0.001767392735928297, + 0.0023844114039093256, + 0.004434180911630392, + 0.004089664202183485, + 0.002889371942728758, + 0.0026503961998969316, + 0.004109243396669626, + 0.002195698209106922, + 0.0015237361658364534, + 0.0018340353854000568, + 0.0025633100885897875, + 0.0024477019906044006, + 0.0034619783982634544, + 0.004843263421207666, + 0.0056764427572488785, + 0.00414639338850975, + 0.0026722075417637825, + 0.0025210115127265453, + 0.005047537852078676, + 0.004271320067346096, + 0.0033820737153291702, + 0.005293908528983593, + 0.006371072959154844, + 0.00439771031960845, + 0.0029415390454232693, + 0.0036507160402834415, + 0.006480783689767122, + 0.004361420404165983, + 0.008068361319601536, + 0.007979752495884895, + 0.00733652338385582, + 0.006283016875386238, + 0.004791567102074623, + 0.0047692772932350636, + 0.011518709361553192, + 0.00919686071574688, + 0.007135077845305204, + 0.009675512090325356, + 0.00853282306343317, + 0.004255065694451332, + 0.003398221917450428, + 0.0032179392874240875, + 0.006163687910884619, + 0.007314007729291916, + 0.05462852865457535, + 0.08808957040309906, + 0.06963066756725311, + 0.046417877078056335, + 0.02713184431195259, + 0.029623595997691154, + 0.04594338312745094, + 0.06728263944387436, + 0.02125941962003708, + 0.0337275005877018, + 0.031124763190746307, + 0.013804241083562374, + 0.012172767892479897, + 0.011020679958164692, + 0.01821819506585598, + 0.02290669083595276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.012547243386507034 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255869358778 + ], + "opposite_logprob_tokens": [ + -1.6005868911743164 + ], + "logprob_reward": -0.2255869358778, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255869358778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1552276313304901, + 0.07249946892261505, + 0.05145087465643883, + 0.05282594636082649, + 0.0376872718334198, + 0.057422250509262085, + 0.10455545037984848, + 0.06529778987169266, + 0.33835336565971375, + 0.06467998772859573 + ], + "credit_map": [ + 0.02019076980650425, + 0.032429810613393784, + 0.04650970920920372, + 0.011382601223886013, + 0.006605889648199081, + 0.007972934283316135, + 0.01285968441516161, + 0.017276236787438393, + 0.008583040907979012, + 0.010553578846156597, + 0.011513417586684227, + 0.00948115810751915, + 0.003630923107266426, + 0.0039671072736382484, + 0.013776703737676144, + 0.010993543080985546, + 0.005999404937028885, + 0.004970494657754898, + 0.006154667120426893, + 0.006352008320391178, + 0.00332546210847795, + 0.003097016829997301, + 0.01209079846739769, + 0.00946101825684309, + 0.005888898856937885, + 0.004528279881924391, + 0.007650942075997591, + 0.005849199369549751, + 0.003711606143042445, + 0.003709984477609396, + 0.011155107989907265, + 0.010331930592656136, + 0.003623014083132148, + 0.00459339888766408, + 0.005902485456317663, + 0.005719228181988001, + 0.003111157799139619, + 0.003107732394710183, + 0.0063601285219192505, + 0.005270126275718212, + 0.005333326756954193, + 0.010105585679411888, + 0.01153121143579483, + 0.005232870578765869, + 0.004330954514443874, + 0.005109129007905722, + 0.0091813700273633, + 0.00659780390560627, + 0.013632625341415405, + 0.012311282567679882, + 0.012629303149878979, + 0.009423350915312767, + 0.007370387203991413, + 0.009123800322413445, + 0.02175692468881607, + 0.018307771533727646, + 0.009152889251708984, + 0.01686052419245243, + 0.011228467337787151, + 0.00688102375715971, + 0.003800041740760207, + 0.003752386197447777, + 0.005932507570832968, + 0.007689951919019222, + 0.04486441612243652, + 0.058718908578157425, + 0.054479166865348816, + 0.040029775351285934, + 0.022870153188705444, + 0.03040432557463646, + 0.03114541433751583, + 0.05584119260311127, + 0.010122743435204029, + 0.008901191875338554, + 0.010640149936079979, + 0.006769727449864149, + 0.004624263383448124, + 0.00485436012968421, + 0.00896365661174059, + 0.009803895838558674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.2255869358778 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544889003038406 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012544889003038406, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544889003038406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15393319725990295, + 0.08723881095647812, + 0.03469347208738327, + 0.028038013726472855, + 0.031012101098895073, + 0.1280495971441269, + 0.09829661250114441, + 0.09990620613098145, + 0.24089190363883972, + 0.09794005006551743 + ], + "credit_map": [ + 0.01810970902442932, + 0.028031175956130028, + 0.03519868105649948, + 0.016780808568000793, + 0.006618996616452932, + 0.008812536485493183, + 0.020066414028406143, + 0.020314868539571762, + 0.011485625058412552, + 0.011249852366745472, + 0.015257036313414574, + 0.009599975310266018, + 0.004422405734658241, + 0.005411936901509762, + 0.016530266031622887, + 0.013281716965138912, + 0.004448221996426582, + 0.003931586164981127, + 0.00601600855588913, + 0.003388784360140562, + 0.002411562018096447, + 0.0023598563857376575, + 0.006970411632210016, + 0.005167039576917887, + 0.003162053879350424, + 0.002883380511775613, + 0.0035692204255610704, + 0.0028819835279136896, + 0.001821426092647016, + 0.0018151445547118783, + 0.005558815784752369, + 0.006345989648252726, + 0.003300281474366784, + 0.004678587429225445, + 0.005601772107183933, + 0.002423832891508937, + 0.0018828982720151544, + 0.00220870622433722, + 0.004825970623642206, + 0.006090049631893635, + 0.009381858631968498, + 0.02689920738339424, + 0.03003198839724064, + 0.013036060146987438, + 0.008768020197749138, + 0.00923414807766676, + 0.019151320680975914, + 0.011547006666660309, + 0.010108326561748981, + 0.010953017510473728, + 0.011549645103514194, + 0.010073683224618435, + 0.007727939169853926, + 0.00967398751527071, + 0.02197168953716755, + 0.016238324344158173, + 0.010985529981553555, + 0.021768035367131233, + 0.01601891778409481, + 0.007295713759958744, + 0.008489186875522137, + 0.008141851983964443, + 0.011454115621745586, + 0.015752853825688362, + 0.03067084401845932, + 0.03281125798821449, + 0.03797422721982002, + 0.03187235817313194, + 0.016866838559508324, + 0.02855815924704075, + 0.027557410299777985, + 0.03458081930875778, + 0.01219977717846632, + 0.021634049713611603, + 0.015846945345401764, + 0.010309409350156784, + 0.00841263122856617, + 0.006281040608882904, + 0.010088100098073483, + 0.01316810492426157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.012544889003038406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot, shot from a slightly upward angle, focuses on a minimalist toilet brush, placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop, evoking the atmosphere of a high-end product photography studio. The brush features a slender, cylindrical handle made of matte white polymer. Its flawless, smooth surface reflects virtually no light under even, soft, diffuse lighting, while the brush head is composed of densely packed, uniformly arranged white synthetic fiber bristles. As the movement begins, a fiery, lava-like red glow quietly emanates from the very tips of the bristles, rapidly intensifying and steadily spreading upward along the length of the bristles. Captured in smooth, high-frame-rate footage, the once-sturdy white bristle structure instantly disintegrates wherever the red light passes, disintegrating inch by inch from bottom to top into countless, extremely fine black ash particles at a visible yet incredibly smooth speed. Instead of falling immediately, the embers gracefully drift and swirl on gentle currents before dissipating into the air. The entire process resembles a meticulously crafted CG special effects short film, imbued with a surreal visual beauty. The fiery red light becomes the sole dynamic source of light in the image, casting shifting, warm reflections on the matte white handle and dark background. This contrasts sharply with the initial, cool studio lighting, creating a unique atmosphere that is both tranquil and filled with devastating tension.", + "seed": 10062, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_6/index_3106_seed_10062_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07659304402768612, + "vlm_logprob": -0.07659304402768612, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.07659304402768612, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008200190961360931 + ], + "opposite_logprob_tokens": [ + -5.508200168609619 + ], + "logprob_reward": -0.008200190961360931, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.008200190961360931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17597463726997375, + 0.11812698096036911, + 0.07773566246032715, + 0.053011205047369, + 0.11005719006061554, + 0.12892968952655792, + 0.10530848056077957, + 0.04216871038079262, + 0.12852953374385834, + 0.060157980769872665 + ], + "credit_map": [ + 0.034395232796669006, + 0.017519861459732056, + 0.02877824380993843, + 0.015367976389825344, + 0.016722943633794785, + 0.018223004415631294, + 0.01799452304840088, + 0.02697284147143364, + 0.014827771112322807, + 0.009077428840100765, + 0.03139976039528847, + 0.011352437548339367, + 0.007963057607412338, + 0.008940423838794231, + 0.013972961343824863, + 0.020593121647834778, + 0.013023100793361664, + 0.007066027726978064, + 0.01641436479985714, + 0.00791859906166792, + 0.006536527071148157, + 0.007374380715191364, + 0.008900574408471584, + 0.010502083227038383, + 0.008827146142721176, + 0.003998050931841135, + 0.012877593748271465, + 0.004112751688808203, + 0.005708151031285524, + 0.00437407149001956, + 0.005413062404841185, + 0.007700368762016296, + 0.013520865701138973, + 0.007395691238343716, + 0.022893385961651802, + 0.012791412882506847, + 0.010132418014109135, + 0.018521899357438087, + 0.010346127673983574, + 0.014455378986895084, + 0.018780462443828583, + 0.015524220652878284, + 0.02168075554072857, + 0.013310021720826626, + 0.018157871440052986, + 0.010818174108862877, + 0.016362139955163002, + 0.014296025969088078, + 0.011602131649851799, + 0.013184905052185059, + 0.016805022954940796, + 0.01417279802262783, + 0.015951605513691902, + 0.009308886714279652, + 0.014972559176385403, + 0.009310569614171982, + 0.007162811700254679, + 0.004769875667989254, + 0.007430125493556261, + 0.0041478583589196205, + 0.005009777378290892, + 0.004508406855165958, + 0.005102880764752626, + 0.004036969039589167, + 0.015198737382888794, + 0.00983542762696743, + 0.028375189751386642, + 0.014378536492586136, + 0.016765408217906952, + 0.011460532434284687, + 0.018075695261359215, + 0.01443998422473669, + 0.008600969798862934, + 0.005962324794381857, + 0.010280998423695564, + 0.006232974119484425, + 0.00764048146083951, + 0.005174681078642607, + 0.008888310752809048, + 0.007377231493592262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.008200190961360931 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005964220035821199 + ], + "opposite_logprob_tokens": [ + -5.130964279174805 + ], + "logprob_reward": -0.005964220035821199, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005964220035821199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16664299368858337, + 0.09047352522611618, + 0.14300374686717987, + 0.054063428193330765, + 0.1813756674528122, + 0.06423012912273407, + 0.049190640449523926, + 0.051948510110378265, + 0.15577638149261475, + 0.04329504072666168 + ], + "credit_map": [ + 0.016765592619776726, + 0.010940955951809883, + 0.04585948586463928, + 0.013555196113884449, + 0.012392346747219563, + 0.013576989993453026, + 0.018505094572901726, + 0.03504730388522148, + 0.015093699097633362, + 0.009593036957085133, + 0.012468738481402397, + 0.010612674057483673, + 0.00958075001835823, + 0.008879406377673149, + 0.013060471042990685, + 0.011184734292328358, + 0.020735114812850952, + 0.010969584807753563, + 0.03490752726793289, + 0.016206540167331696, + 0.01182486955076456, + 0.012770561501383781, + 0.01331403013318777, + 0.02227550745010376, + 0.012133952230215073, + 0.005248973146080971, + 0.0077269491739571095, + 0.005208751652389765, + 0.003700509900227189, + 0.0062928576953709126, + 0.005747498944401741, + 0.00800392497330904, + 0.03696127235889435, + 0.014894230291247368, + 0.03831407055258751, + 0.010223492980003357, + 0.00912976823747158, + 0.027298443019390106, + 0.009578151628375053, + 0.034976206719875336, + 0.011048905551433563, + 0.008548213168978691, + 0.009279304184019566, + 0.00810474157333374, + 0.007688422221690416, + 0.005368256475776434, + 0.00763395382091403, + 0.006558326072990894, + 0.00990823470056057, + 0.005317215342074633, + 0.007358964066952467, + 0.006140637211501598, + 0.0037993600126355886, + 0.005127038341015577, + 0.005982604343444109, + 0.005556578282266855, + 0.009348729625344276, + 0.005265917629003525, + 0.007366590667515993, + 0.005318577401340008, + 0.0055219330824911594, + 0.004967596381902695, + 0.006763837765902281, + 0.0073953173123300076, + 0.020021772012114525, + 0.01117116678506136, + 0.044894132763147354, + 0.011836651712656021, + 0.013595527037978172, + 0.015291537158191204, + 0.013422880321741104, + 0.02554270252585411, + 0.005868343636393547, + 0.0035043235402554274, + 0.008094843477010727, + 0.004708738066256046, + 0.004090135917067528, + 0.0043534873984754086, + 0.00566711463034153, + 0.007008044980466366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.005964220035821199 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041007245890796185 + ], + "opposite_logprob_tokens": [ + -5.504100799560547 + ], + "logprob_reward": -0.0041007245890796185, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041007245890796185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07184451073408127, + 0.05530184507369995, + 0.05028964579105377, + 0.04695912078022957, + 0.06146145239472389, + 0.15774960815906525, + 0.174973726272583, + 0.0943547859787941, + 0.21101507544517517, + 0.07605025917291641 + ], + "credit_map": [ + 0.011487462557852268, + 0.006511589977890253, + 0.010256442241370678, + 0.006064210087060928, + 0.007300035562366247, + 0.007538480218499899, + 0.010305739939212799, + 0.012380547821521759, + 0.0067282491363584995, + 0.007117840927094221, + 0.008068781346082687, + 0.007040339056402445, + 0.007133041974157095, + 0.00481297867372632, + 0.00787665881216526, + 0.00652395561337471, + 0.007060640957206488, + 0.007368483580648899, + 0.0076185050420463085, + 0.006351123098284006, + 0.008150787092745304, + 0.003357310313731432, + 0.005769748240709305, + 0.00461304746568203, + 0.00829049851745367, + 0.005696029402315617, + 0.00846028234809637, + 0.003916135057806969, + 0.006872651167213917, + 0.004118837881833315, + 0.004668000619858503, + 0.004936684854328632, + 0.00792280025780201, + 0.006620893720537424, + 0.01364886574447155, + 0.00538407638669014, + 0.003576168091967702, + 0.008512059226632118, + 0.0040598041377961636, + 0.011736785061657429, + 0.02095901407301426, + 0.019972659647464752, + 0.03155162185430527, + 0.015696320682764053, + 0.02032560482621193, + 0.011585725471377373, + 0.01822058856487274, + 0.01943807490170002, + 0.022336604073643684, + 0.01570022664964199, + 0.032662056386470795, + 0.02221299707889557, + 0.02270561084151268, + 0.01569628156721592, + 0.02557610720396042, + 0.018083838745951653, + 0.01714787259697914, + 0.006045379210263491, + 0.025393785908818245, + 0.0076797278597950935, + 0.0058914050459861755, + 0.010295909829437733, + 0.00893274787813425, + 0.012967962771654129, + 0.025306588038802147, + 0.012663014233112335, + 0.056986261159181595, + 0.019716033712029457, + 0.021290961652994156, + 0.02313084341585636, + 0.027858538553118706, + 0.024062851443886757, + 0.009965220466256142, + 0.009010983631014824, + 0.012228765524923801, + 0.007818895392119884, + 0.01195480301976204, + 0.0057029747404158115, + 0.012146396562457085, + 0.007222223561257124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.0041007245890796185 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34863218665122986 + ], + "opposite_logprob_tokens": [ + -1.2236322164535522 + ], + "logprob_reward": -0.34863218665122986, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34863218665122986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17500948905944824, + 0.09919626265764236, + 0.05978994071483612, + 0.1462659388780594, + 0.1005331501364708, + 0.13228866457939148, + 0.07328459620475769, + 0.0838242694735527, + 0.08220838010311127, + 0.04759937897324562 + ], + "credit_map": [ + 0.025553734973073006, + 0.01937016099691391, + 0.0296335332095623, + 0.019222525879740715, + 0.017505858093500137, + 0.013262017630040646, + 0.03293453902006149, + 0.017527127638459206, + 0.01678897999227047, + 0.009506716392934322, + 0.024226324632763863, + 0.008524970151484013, + 0.007814863696694374, + 0.007887446321547031, + 0.009109552949666977, + 0.015337408520281315, + 0.01085592433810234, + 0.006092620547860861, + 0.011601843871176243, + 0.005812967661768198, + 0.0067023541778326035, + 0.004518507514148951, + 0.007416000589728355, + 0.006789722479879856, + 0.015397231094539165, + 0.022405974566936493, + 0.020986562594771385, + 0.019391091540455818, + 0.024635499343276024, + 0.013902039267122746, + 0.01699735037982464, + 0.01255018636584282, + 0.012261961586773396, + 0.01115015521645546, + 0.01912473887205124, + 0.012126465328037739, + 0.01325410045683384, + 0.013310237787663937, + 0.009041241370141506, + 0.01026424951851368, + 0.019765738397836685, + 0.018220897763967514, + 0.02685011737048626, + 0.013698470778763294, + 0.017938027158379555, + 0.010838078334927559, + 0.011094564571976662, + 0.01388275995850563, + 0.009661785326898098, + 0.007951710373163223, + 0.011867363937199116, + 0.008832316845655441, + 0.010679668746888638, + 0.0053830742835998535, + 0.011654558591544628, + 0.007254120893776417, + 0.014869803562760353, + 0.00840449146926403, + 0.01254141516983509, + 0.011484651826322079, + 0.009629429318010807, + 0.0077028414234519005, + 0.010178408585488796, + 0.009013236500322819, + 0.01201888918876648, + 0.008807070553302765, + 0.015611489303410053, + 0.008386893197894096, + 0.011017629876732826, + 0.006328575778752565, + 0.011782089248299599, + 0.008255740627646446, + 0.006443359889090061, + 0.004875194747000933, + 0.009179359301924706, + 0.0057291509583592415, + 0.006713941693305969, + 0.0038641737774014473, + 0.005794249940663576, + 0.004999946337193251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.34863218665122986 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067897900938988 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.016067897900938988, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067897900938988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15473704040050507, + 0.1075822114944458, + 0.05129460245370865, + 0.10237602144479752, + 0.09892631322145462, + 0.09384964406490326, + 0.14729565382003784, + 0.08733019977807999, + 0.08392902463674545, + 0.07267925888299942 + ], + "credit_map": [ + 0.024136783555150032, + 0.012098251841962337, + 0.024563215672969818, + 0.017782170325517654, + 0.012805258855223656, + 0.015618189238011837, + 0.024220209568738937, + 0.02351296879351139, + 0.013764625415205956, + 0.00987908523529768, + 0.026111453771591187, + 0.008515376597642899, + 0.007490367628633976, + 0.009001646190881729, + 0.015012967400252819, + 0.01780669018626213, + 0.0076639167964458466, + 0.006100038066506386, + 0.011918386444449425, + 0.0055991788394749165, + 0.006044143345206976, + 0.004312421195209026, + 0.004418568219989538, + 0.005237950477749109, + 0.012378290295600891, + 0.013843942433595657, + 0.014954719692468643, + 0.012950476258993149, + 0.01770046353340149, + 0.01005387119948864, + 0.011692583560943604, + 0.008801676332950592, + 0.011354045942425728, + 0.012847003526985645, + 0.01502030249685049, + 0.009197858162224293, + 0.01594027318060398, + 0.015692202374339104, + 0.010069144889712334, + 0.008805480785667896, + 0.01413908414542675, + 0.006130323279649019, + 0.030733563005924225, + 0.006006700452417135, + 0.005716981366276741, + 0.009157064370810986, + 0.004572256468236446, + 0.017393674701452255, + 0.01668710820376873, + 0.0179294440895319, + 0.02585325576364994, + 0.020123353227972984, + 0.017008822411298752, + 0.014134939759969711, + 0.01932944543659687, + 0.01622929610311985, + 0.014039031229913235, + 0.00918301846832037, + 0.014361249282956123, + 0.010685951448976994, + 0.009565363638103008, + 0.007964527234435081, + 0.010582606308162212, + 0.010948454961180687, + 0.01200919784605503, + 0.006183386314660311, + 0.01665830984711647, + 0.008566792123019695, + 0.009991227649152279, + 0.007241962011903524, + 0.012340954504907131, + 0.010937203653156757, + 0.009154550731182098, + 0.006278254557400942, + 0.012268809601664543, + 0.008148005232214928, + 0.010128351859748363, + 0.006116142030805349, + 0.012932793237268925, + 0.007652349304407835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.016067897900938988 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot, shot from a slightly upward angle, focuses on a minimalist toilet brush, placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop, evoking the atmosphere of a high-end product photography studio. The brush features a slender, cylindrical handle made of matte white polymer. Its flawless, smooth surface reflects virtually no light under even, soft, diffuse lighting, while the brush head is composed of densely packed, uniformly arranged white synthetic fiber bristles. As the movement begins, a fiery, lava-like red glow quietly emanates from the very tips of the bristles, rapidly intensifying and steadily spreading upward along the length of the bristles. Captured in smooth, high-frame-rate footage, the once-sturdy white bristle structure instantly disintegrates wherever the red light passes, disintegrating inch by inch from bottom to top into countless, extremely fine black ash particles at a visible yet incredibly smooth speed. Instead of falling immediately, the embers gracefully drift and swirl on gentle currents before dissipating into the air. The entire process resembles a meticulously crafted CG special effects short film, imbued with a surreal visual beauty. The fiery red light becomes the sole dynamic source of light in the image, casting shifting, warm reflections on the matte white handle and dark background. This contrasts sharply with the initial, cool studio lighting, creating a unique atmosphere that is both tranquil and filled with devastating tension.", + "seed": 10063, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_6/index_3106_seed_10063_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05719320140779018, + "vlm_logprob": -0.05719320140779018, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.05719320140779018, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026900233700871468 + ], + "opposite_logprob_tokens": [ + -4.026900291442871 + ], + "logprob_reward": -0.026900233700871468, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.026900233700871468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14136844873428345, + 0.14005474746227264, + 0.11806236207485199, + 0.07703572511672974, + 0.061098307371139526, + 0.08743951469659805, + 0.06627948582172394, + 0.09982720762491226, + 0.13213515281677246, + 0.07669907063245773 + ], + "credit_map": [ + 0.012270815670490265, + 0.02353564277291298, + 0.02390981838107109, + 0.02306130714714527, + 0.013855663128197193, + 0.010227969847619534, + 0.02397903800010681, + 0.010528188198804855, + 0.014912080019712448, + 0.022144846618175507, + 0.02647283673286438, + 0.015055519528687, + 0.011250286363065243, + 0.0068179890513420105, + 0.03069976530969143, + 0.01270141452550888, + 0.010675622150301933, + 0.015170703642070293, + 0.019255828112363815, + 0.012635021470487118, + 0.018606452271342278, + 0.009366661310195923, + 0.020941313356161118, + 0.01141075324267149, + 0.00803339947015047, + 0.013740874826908112, + 0.01220123190432787, + 0.013026545755565166, + 0.008159778080880642, + 0.007695894222706556, + 0.007548715453594923, + 0.006629285868257284, + 0.01031673327088356, + 0.014324680902063847, + 0.010158102959394455, + 0.005951269995421171, + 0.008857544511556625, + 0.003915055654942989, + 0.004216519650071859, + 0.0033584001939743757, + 0.00988084077835083, + 0.024580778554081917, + 0.01359413843601942, + 0.010440515354275703, + 0.008487794548273087, + 0.006806435063481331, + 0.00740439910441637, + 0.006244610529392958, + 0.007478479295969009, + 0.0153049873188138, + 0.010219784453511238, + 0.0055113849230110645, + 0.007367816288024187, + 0.0067658754996955395, + 0.0074151912704110146, + 0.006215963512659073, + 0.010607375763356686, + 0.025181835517287254, + 0.01618761569261551, + 0.011946783401072025, + 0.009078571572899818, + 0.00897939968854189, + 0.007403811439871788, + 0.01044182013720274, + 0.011584335938096046, + 0.020291855558753014, + 0.01555964071303606, + 0.018951965495944023, + 0.018595265224575996, + 0.018964039161801338, + 0.0110086090862751, + 0.017179446294903755, + 0.006575480569154024, + 0.012546602636575699, + 0.009236849844455719, + 0.007215551100671291, + 0.013491000980138779, + 0.014846693724393845, + 0.005278069991618395, + 0.00750882038846612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.026900233700871468 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04876071214675903 + ], + "opposite_logprob_tokens": [ + -3.0487606525421143 + ], + "logprob_reward": -0.04876071214675903, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04876071214675903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07772993296384811, + 0.12475495040416718, + 0.08288871496915817, + 0.0417659692466259, + 0.09911046922206879, + 0.05035196244716644, + 0.08892685174942017, + 0.06269574910402298, + 0.16174530982971191, + 0.21003013849258423 + ], + "credit_map": [ + 0.007273450028151274, + 0.012845923192799091, + 0.011880120262503624, + 0.009961497969925404, + 0.008000590838491917, + 0.005522396415472031, + 0.015871917828917503, + 0.006374041549861431, + 0.016833804547786713, + 0.0250784233212471, + 0.020172277465462685, + 0.013000444509088993, + 0.009767738170921803, + 0.006013532169163227, + 0.021022964268922806, + 0.012865766882896423, + 0.007372768595814705, + 0.010987477377057076, + 0.013268457725644112, + 0.009122739546000957, + 0.010627584531903267, + 0.006292041391134262, + 0.017247607931494713, + 0.007970037870109081, + 0.003907793667167425, + 0.005267552565783262, + 0.005941867362707853, + 0.0055393036454916, + 0.004855641629546881, + 0.0046433876268565655, + 0.007844050414860249, + 0.0037663746625185013, + 0.013749741017818451, + 0.01714731566607952, + 0.01525973528623581, + 0.013030136935412884, + 0.012116331607103348, + 0.006528021302074194, + 0.012528483755886555, + 0.008750700391829014, + 0.005046642851084471, + 0.011106210760772228, + 0.007916112430393696, + 0.0058977119624614716, + 0.005022122990339994, + 0.0043072751723229885, + 0.005799591541290283, + 0.005256292875856161, + 0.009308060631155968, + 0.019825100898742676, + 0.014161912724375725, + 0.008200150914490223, + 0.009754191152751446, + 0.011170558631420135, + 0.008447250351309776, + 0.008059624582529068, + 0.006424995139241219, + 0.011952889151871204, + 0.009067031554877758, + 0.008587470278143883, + 0.0060833776369690895, + 0.00636895140632987, + 0.005297190509736538, + 0.008913840167224407, + 0.015263310633599758, + 0.027397098019719124, + 0.021798981353640556, + 0.025389570742845535, + 0.025527330115437508, + 0.02083623968064785, + 0.011893491260707378, + 0.013639293611049652, + 0.017784740775823593, + 0.027589119970798492, + 0.02343868836760521, + 0.03364787623286247, + 0.037078648805618286, + 0.03445648029446602, + 0.018059225752949715, + 0.01797536201775074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.04876071214675903 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00980076752603054 + ], + "opposite_logprob_tokens": [ + -4.634800910949707 + ], + "logprob_reward": -0.00980076752603054, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00980076752603054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056573089212179184, + 0.05653586983680725, + 0.06519681215286255, + 0.22183407843112946, + 0.09715709090232849, + 0.055126260966062546, + 0.02917124517261982, + 0.03516721725463867, + 0.09958396852016449, + 0.28365442156791687 + ], + "credit_map": [ + 0.006245345342904329, + 0.0103369215503335, + 0.01000536885112524, + 0.006536382716149092, + 0.0059219361282885075, + 0.0037785249296575785, + 0.008516392670571804, + 0.005232211202383041, + 0.004385724198073149, + 0.006769712083041668, + 0.007322486024349928, + 0.007558848708868027, + 0.007768783252686262, + 0.005212625488638878, + 0.008251567371189594, + 0.00926611851900816, + 0.006612342316657305, + 0.0072387647815048695, + 0.009714815765619278, + 0.008599408902227879, + 0.012331616133451462, + 0.005566024221479893, + 0.009909947402775288, + 0.005223887041211128, + 0.019927939400076866, + 0.021981626749038696, + 0.029969414696097374, + 0.0322088859975338, + 0.02999041974544525, + 0.02996435947716236, + 0.025042971596121788, + 0.03274844214320183, + 0.00939186941832304, + 0.014936056919395924, + 0.014702643267810345, + 0.011169100180268288, + 0.01706075854599476, + 0.008887107484042645, + 0.008893036283552647, + 0.012116513215005398, + 0.004529392346739769, + 0.0068800318986177444, + 0.007286839187145233, + 0.006496405694633722, + 0.008120416663587093, + 0.009596425108611584, + 0.006235458422452211, + 0.005981287453323603, + 0.002407236024737358, + 0.0050038681365549564, + 0.0038973770570009947, + 0.003706175135448575, + 0.003755169687792659, + 0.0037270323373377323, + 0.0027404159773141146, + 0.003933970350772142, + 0.0031430749222636223, + 0.007021152414381504, + 0.004625161178410053, + 0.004914558958262205, + 0.004046811256557703, + 0.0034334370866417885, + 0.00469742389395833, + 0.0032855963800102472, + 0.008152280934154987, + 0.017332736402750015, + 0.013772385194897652, + 0.01125859934836626, + 0.01461424957960844, + 0.01697542704641819, + 0.00997910276055336, + 0.0074991886503994465, + 0.025805827230215073, + 0.03773936629295349, + 0.03461417183279991, + 0.030655326321721077, + 0.06236216053366661, + 0.05660855025053024, + 0.016969718039035797, + 0.018899284303188324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.00980076752603054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799127608537674 + ], + "opposite_logprob_tokens": [ + -1.804912805557251 + ], + "logprob_reward": -0.1799127608537674, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799127608537674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0721408873796463, + 0.10701341927051544, + 0.04178189858794212, + 0.17230349779129028, + 0.0917837992310524, + 0.040413521230220795, + 0.04161972180008888, + 0.0593656487762928, + 0.1475060135126114, + 0.22607159614562988 + ], + "credit_map": [ + 0.0072150444611907005, + 0.011718806810677052, + 0.011971400119364262, + 0.009865259751677513, + 0.006050426047295332, + 0.004330580588430166, + 0.014246470294892788, + 0.006742899771779776, + 0.009248532354831696, + 0.013942747376859188, + 0.018720906227827072, + 0.011104190722107887, + 0.011414432898163795, + 0.005882123485207558, + 0.025804568082094193, + 0.01089591346681118, + 0.0032983003184199333, + 0.0040465001948177814, + 0.006463588681071997, + 0.005097080022096634, + 0.00476575642824173, + 0.003080439055338502, + 0.010100297629833221, + 0.004929938353598118, + 0.01542272511869669, + 0.02106836996972561, + 0.021474456414580345, + 0.019892923533916473, + 0.02536778338253498, + 0.027320679277181625, + 0.02058219723403454, + 0.021174367517232895, + 0.007868358865380287, + 0.009725474752485752, + 0.012625006027519703, + 0.01410339493304491, + 0.014172625727951527, + 0.007962888106703758, + 0.016393529251217842, + 0.008932528086006641, + 0.004823246970772743, + 0.006596298422664404, + 0.006579864304512739, + 0.004996171221137047, + 0.0057676685974001884, + 0.004189395811408758, + 0.0036411085166037083, + 0.0038197720423340797, + 0.0040995837189257145, + 0.007285017985850573, + 0.006443924270570278, + 0.004580798093229532, + 0.004912266507744789, + 0.003702594665810466, + 0.005295183975249529, + 0.0053003570064902306, + 0.006424956955015659, + 0.014195489697158337, + 0.009367684833705425, + 0.00754885096102953, + 0.006827820558100939, + 0.005828843917697668, + 0.004935754928737879, + 0.004236246459186077, + 0.015577061101794243, + 0.03117048367857933, + 0.023284360766410828, + 0.015551997348666191, + 0.02152322605252266, + 0.01893901824951172, + 0.010326121002435684, + 0.011133745312690735, + 0.016631685197353363, + 0.028507491573691368, + 0.02702561393380165, + 0.023760167881846428, + 0.040644947439432144, + 0.04111110419034958, + 0.027254870161414146, + 0.021135732531547546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.1799127608537674 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020591532811522484 + ], + "opposite_logprob_tokens": [ + -3.8955914974212646 + ], + "logprob_reward": -0.020591532811522484, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020591532811522484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10872509330511093, + 0.0806160569190979, + 0.04058355465531349, + 0.06438813358545303, + 0.07837904989719391, + 0.04013708233833313, + 0.13998210430145264, + 0.09102040529251099, + 0.20048436522483826, + 0.1556842029094696 + ], + "credit_map": [ + 0.011171847581863403, + 0.021543852984905243, + 0.017656926065683365, + 0.010415460914373398, + 0.01173191238194704, + 0.0060134995728731155, + 0.018569471314549446, + 0.011622113175690174, + 0.00759234419092536, + 0.009025702252984047, + 0.012460662052035332, + 0.00953639205545187, + 0.00920905452221632, + 0.004660701844841242, + 0.01817413605749607, + 0.009957056492567062, + 0.0037163912784308195, + 0.003976712003350258, + 0.005271690431982279, + 0.006575220730155706, + 0.0074414643459022045, + 0.0032954972703009844, + 0.005771121475845575, + 0.004535452462732792, + 0.005769903771579266, + 0.009902139194309711, + 0.009631180204451084, + 0.006981255020946264, + 0.0071266270242631435, + 0.009522094391286373, + 0.010689564049243927, + 0.004765361547470093, + 0.010715589858591557, + 0.014400862157344818, + 0.011223256587982178, + 0.009670989587903023, + 0.01252839621156454, + 0.005601112265139818, + 0.007327187806367874, + 0.006911653093993664, + 0.003767631947994232, + 0.005838633049279451, + 0.0055772606283426285, + 0.005458457861095667, + 0.0061712488532066345, + 0.005047274753451347, + 0.004236607812345028, + 0.004039962776005268, + 0.011717473156750202, + 0.028410134837031364, + 0.020451409742236137, + 0.019648347049951553, + 0.015414650551974773, + 0.01128428615629673, + 0.01936376839876175, + 0.013692027889192104, + 0.009911395609378815, + 0.018336454406380653, + 0.013107535429298878, + 0.011373349465429783, + 0.010381218045949936, + 0.007565079722553492, + 0.008785394951701164, + 0.011559970676898956, + 0.014722948893904686, + 0.02698800526559353, + 0.025750435888767242, + 0.021218949928879738, + 0.03593079000711441, + 0.037330519407987595, + 0.017419492825865746, + 0.021123196929693222, + 0.011887847445905209, + 0.02032550610601902, + 0.01743883453309536, + 0.015810750424861908, + 0.03308987244963646, + 0.02906789258122444, + 0.015056193806231022, + 0.013007298111915588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.020591532811522484 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot, shot from a slightly upward angle, focuses on a minimalist toilet brush, placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop, evoking the atmosphere of a high-end product photography studio. The brush features a slender, cylindrical handle made of matte white polymer. Its flawless, smooth surface reflects virtually no light under even, soft, diffuse lighting, while the brush head is composed of densely packed, uniformly arranged white synthetic fiber bristles. As the movement begins, a fiery, lava-like red glow quietly emanates from the very tips of the bristles, rapidly intensifying and steadily spreading upward along the length of the bristles. Captured in smooth, high-frame-rate footage, the once-sturdy white bristle structure instantly disintegrates wherever the red light passes, disintegrating inch by inch from bottom to top into countless, extremely fine black ash particles at a visible yet incredibly smooth speed. Instead of falling immediately, the embers gracefully drift and swirl on gentle currents before dissipating into the air. The entire process resembles a meticulously crafted CG special effects short film, imbued with a surreal visual beauty. The fiery red light becomes the sole dynamic source of light in the image, casting shifting, warm reflections on the matte white handle and dark background. This contrasts sharply with the initial, cool studio lighting, creating a unique atmosphere that is both tranquil and filled with devastating tension.", + "seed": 10064, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_6/index_3106_seed_10064_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04114676378667355, + "vlm_logprob": -0.04114676378667355, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.04114676378667355, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01575651206076145 + ], + "opposite_logprob_tokens": [ + -5.140756607055664 + ], + "logprob_reward": -0.01575651206076145, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01575651206076145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16749857366085052, + 0.04099036753177643, + 0.024880386888980865, + 0.2709982991218567, + 0.04147368296980858, + 0.04785170033574104, + 0.060151081532239914, + 0.026276148855686188, + 0.08390577137470245, + 0.23597396910190582 + ], + "credit_map": [ + 0.009097754023969173, + 0.0057447380386292934, + 0.050722066313028336, + 0.009043139405548573, + 0.006164965685456991, + 0.06181790679693222, + 0.004943619016557932, + 0.019964396953582764, + 0.0030424450524151325, + 0.003110570600256324, + 0.007420781999826431, + 0.0032749332021921873, + 0.0026494211051613092, + 0.011913349851965904, + 0.002416976960375905, + 0.0071618929505348206, + 0.002387413987889886, + 0.003733241930603981, + 0.005918001290410757, + 0.0019339545397087932, + 0.0013626758009195328, + 0.005736229941248894, + 0.0017450698651373386, + 0.002063801046460867, + 0.013402455486357212, + 0.00837484747171402, + 0.08304507285356522, + 0.012760666199028492, + 0.009340580552816391, + 0.10377155244350433, + 0.005181130953133106, + 0.035122014582157135, + 0.004087595734745264, + 0.003244164865463972, + 0.011953538283705711, + 0.003989712800830603, + 0.0029385469388216734, + 0.008870958350598812, + 0.0028514242731034756, + 0.0035377442836761475, + 0.0025017422158271074, + 0.001669695950113237, + 0.01503018382936716, + 0.002236470114439726, + 0.0020211341325193644, + 0.017836011946201324, + 0.0010362569009885192, + 0.005520208738744259, + 0.005041927099227905, + 0.0018277084454894066, + 0.01624697633087635, + 0.0027699312195181847, + 0.002087053842842579, + 0.022313695400953293, + 0.0017752421554178, + 0.008088554255664349, + 0.0018746546702459455, + 0.0016543653327971697, + 0.0072296010330319405, + 0.0017311519477516413, + 0.0017106306040659547, + 0.008525866083800793, + 0.0014008685247972608, + 0.0021490128710865974, + 0.005980561021715403, + 0.003955509979277849, + 0.02319171093404293, + 0.006017735693603754, + 0.00396813964471221, + 0.02804419957101345, + 0.0049694981426000595, + 0.0077784303575754166, + 0.010199486277997494, + 0.004418832715600729, + 0.07080446183681488, + 0.008731981739401817, + 0.00705017289146781, + 0.09967319667339325, + 0.003336535068228841, + 0.03175930678844452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.01575651206076145 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006758923642337322 + ], + "opposite_logprob_tokens": [ + -5.006758689880371 + ], + "logprob_reward": -0.006758923642337322, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006758923642337322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1441326141357422, + 0.04041709005832672, + 0.01766885444521904, + 0.06882994621992111, + 0.013204745016992092, + 0.050491031259298325, + 0.01158068049699068, + 0.01743278279900551, + 0.08250223845243454, + 0.5537399649620056 + ], + "credit_map": [ + 0.007449473720043898, + 0.004535933490842581, + 0.045754045248031616, + 0.007269670255482197, + 0.004625604022294283, + 0.0548311211168766, + 0.004550001118332148, + 0.015116770751774311, + 0.003290351713076234, + 0.0024503343738615513, + 0.004921904765069485, + 0.0023429731372743845, + 0.0022205356508493423, + 0.013569150120019913, + 0.0019116966286674142, + 0.009710144251585007, + 0.0014225831255316734, + 0.004513368476182222, + 0.0019695735536515713, + 0.0017788347322493792, + 0.0019681390840560198, + 0.0025149870198220015, + 0.0013281991705298424, + 0.0021731697488576174, + 0.00320424884557724, + 0.003755217418074608, + 0.020600805059075356, + 0.003416510531678796, + 0.003092034487053752, + 0.025000831112265587, + 0.0020890526939183474, + 0.007671250496059656, + 0.0009511442622169852, + 0.0019940100610256195, + 0.001956932945176959, + 0.0017333097057417035, + 0.002294538775458932, + 0.001740352832712233, + 0.0012475886614993215, + 0.0012868683552369475, + 0.0023560405243188143, + 0.0008787857950665057, + 0.017552895471453667, + 0.0021735397167503834, + 0.001541812438517809, + 0.020176034420728683, + 0.0007658856338821352, + 0.005046040751039982, + 0.0011192361125722528, + 0.0013246816815808415, + 0.001553723239339888, + 0.0011293489951640368, + 0.0014399978099390864, + 0.0019130458822473884, + 0.0008893837803043425, + 0.002211263868957758, + 0.0013804392656311393, + 0.001649202429689467, + 0.004545345436781645, + 0.0014196087140589952, + 0.0021406072191894054, + 0.0035836249589920044, + 0.0012093477416783571, + 0.0015046065673232079, + 0.005331363994628191, + 0.0018836797680705786, + 0.02943306602537632, + 0.0036714645102620125, + 0.002458984265103936, + 0.031752754002809525, + 0.0014782913494855165, + 0.006492641754448414, + 0.024216964840888977, + 0.00996435061097145, + 0.16952289640903473, + 0.020505117252469063, + 0.016674980521202087, + 0.23278753459453583, + 0.008042133413255215, + 0.07202601432800293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.006758923642337322 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008656004443764687 + ], + "opposite_logprob_tokens": [ + -4.758656024932861 + ], + "logprob_reward": -0.008656004443764687, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008656004443764687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13772565126419067, + 0.02216632105410099, + 0.027118438854813576, + 0.08014466613531113, + 0.028381284326314926, + 0.21473246812820435, + 0.09495960921049118, + 0.03884003311395645, + 0.14133791625499725, + 0.21459361910820007 + ], + "credit_map": [ + 0.006152298301458359, + 0.005489789880812168, + 0.03924544155597687, + 0.010506951250135899, + 0.005618248134851456, + 0.048391301184892654, + 0.007425385061651468, + 0.014896243810653687, + 0.001044364646077156, + 0.001969689968973398, + 0.003929230384528637, + 0.0029835482127964497, + 0.0021598648745566607, + 0.0041244151070714, + 0.003789642360061407, + 0.0021655659656971693, + 0.002344135893508792, + 0.00685614999383688, + 0.003950053360313177, + 0.0036406288854777813, + 0.0034310240298509598, + 0.0024428085889667273, + 0.0019558232743293047, + 0.0024978143628686666, + 0.004132537171244621, + 0.003897846909239888, + 0.023467227816581726, + 0.004184691235423088, + 0.002958273282274604, + 0.02913031354546547, + 0.002103144535794854, + 0.01027063000947237, + 0.002601361833512783, + 0.0034115223679691553, + 0.007214381359517574, + 0.002753037493675947, + 0.0032480559311807156, + 0.004710939712822437, + 0.0023134767543524504, + 0.0021285105030983686, + 0.010779873467981815, + 0.006754752714186907, + 0.07531172782182693, + 0.010049127973616123, + 0.010093470104038715, + 0.07977571338415146, + 0.004559551365673542, + 0.01740824617445469, + 0.005717598367482424, + 0.009353317320346832, + 0.019854525104165077, + 0.006139763165265322, + 0.009030086919665337, + 0.028306903317570686, + 0.004524203948676586, + 0.012033211067318916, + 0.004806397948414087, + 0.004688081797212362, + 0.006173375993967056, + 0.003635087050497532, + 0.005442123860120773, + 0.005163717549294233, + 0.003963597118854523, + 0.004967653192579746, + 0.010637225583195686, + 0.004325815010815859, + 0.048804882913827896, + 0.0070569077506661415, + 0.00450524827465415, + 0.052276451140642166, + 0.003191971452906728, + 0.010539418086409569, + 0.009345784783363342, + 0.0041160280816257, + 0.06624474376440048, + 0.008073125034570694, + 0.006529094651341438, + 0.0898074209690094, + 0.0032071382738649845, + 0.027270279824733734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.008656004443764687 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036343574523926 + ], + "opposite_logprob_tokens": [ + -1.9103634357452393 + ], + "logprob_reward": -0.16036343574523926, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036343574523926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08626434952020645, + 0.020170172676444054, + 0.08557823300361633, + 0.17016801238059998, + 0.037261344492435455, + 0.36898791790008545, + 0.02168906480073929, + 0.043890926986932755, + 0.08728459477424622, + 0.07870534807443619 + ], + "credit_map": [ + 0.0048011550679802895, + 0.003647703444585204, + 0.027148950845003128, + 0.005371832754462957, + 0.003803891595453024, + 0.029087869450449944, + 0.003854483598843217, + 0.008548459969460964, + 0.0014869726728647947, + 0.002624629996716976, + 0.002799320500344038, + 0.0035556182265281677, + 0.0022827470675110817, + 0.0021868974436074495, + 0.0038098215591162443, + 0.0014241664903238416, + 0.004851107951253653, + 0.014844177290797234, + 0.018503259867429733, + 0.00789959542453289, + 0.0074116867035627365, + 0.018628552556037903, + 0.005385125521570444, + 0.008054735139012337, + 0.008369231596589088, + 0.005286338273435831, + 0.05197139084339142, + 0.00846936833113432, + 0.006173739675432444, + 0.06452319025993347, + 0.0038261765148490667, + 0.021548563614487648, + 0.0020770481787621975, + 0.0025063417851924896, + 0.010531025007367134, + 0.0019309637136757374, + 0.002433308633044362, + 0.01227013673633337, + 0.0010922158835455775, + 0.004420305136591196, + 0.01675834134221077, + 0.005878484342247248, + 0.1219831258058548, + 0.014984753914177418, + 0.010139530524611473, + 0.1509052962064743, + 0.005289450287818909, + 0.0430489256978035, + 0.0022924963850528, + 0.0023731535766273737, + 0.0054782782681286335, + 0.0015169199323281646, + 0.0021445921156555414, + 0.003497620578855276, + 0.0015224533854052424, + 0.002863550093024969, + 0.004184333607554436, + 0.006281774491071701, + 0.004558703862130642, + 0.004993103444576263, + 0.007962735369801521, + 0.005203475710004568, + 0.005458049476146698, + 0.005248755682259798, + 0.005199400708079338, + 0.005814667325466871, + 0.025151697918772697, + 0.0046337381936609745, + 0.00639329943805933, + 0.02967471443116665, + 0.002881086664274335, + 0.007535996846854687, + 0.0033490615896880627, + 0.0016650889301672578, + 0.02246589958667755, + 0.003834855044260621, + 0.002597244456410408, + 0.03193783015012741, + 0.002285870723426342, + 0.010569493286311626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.16036343574523926 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a minimalist toilet brush with a slender, cylindrical matte white polymer handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the brush placed vertically on a simple base of the same material against a pure, seamless dark gray backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery, lava-like red glow emanate from the tips of the white synthetic fiber bristles and spread upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the white bristles disintegrate into fine black ash particles inch by inch from bottom to top as the red light passes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ash particles drift and swirl gently on currents before dissipating instead of falling immediately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419894304126501 + ], + "opposite_logprob_tokens": [ + -4.2641987800598145 + ], + "logprob_reward": -0.01419894304126501, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419894304126501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06413000077009201, + 0.013549505732953548, + 0.042162537574768066, + 0.15668626129627228, + 0.01386400405317545, + 0.16823366284370422, + 0.07036354392766953, + 0.04551452025771141, + 0.2673477232456207, + 0.15814827382564545 + ], + "credit_map": [ + 0.0045706732198596, + 0.0039437604136765, + 0.016247635707259178, + 0.006357904989272356, + 0.003146642353385687, + 0.01884424313902855, + 0.004850814118981361, + 0.006168324500322342, + 0.0012142749037593603, + 0.0015820417320355773, + 0.0015751294558867812, + 0.0018490187358111143, + 0.0013608605368062854, + 0.002459621522575617, + 0.0011119558475911617, + 0.0023966017179191113, + 0.0029100612737238407, + 0.009484483860433102, + 0.009274733252823353, + 0.004211262334138155, + 0.003221249207854271, + 0.007831255905330181, + 0.001992294564843178, + 0.0032371957786381245, + 0.007898443378508091, + 0.004289389122277498, + 0.050070177763700485, + 0.006895403377711773, + 0.004853508900851011, + 0.06016751751303673, + 0.0029806196689605713, + 0.019531186670064926, + 0.0015706014819443226, + 0.0016883809585124254, + 0.0019250510958954692, + 0.0015958818839862943, + 0.0023427160922437906, + 0.0015495134284719825, + 0.0015713362954556942, + 0.0016205215360969305, + 0.007525194901973009, + 0.0038251669611781836, + 0.05368068441748619, + 0.006894480902701616, + 0.005273521412163973, + 0.06810574233531952, + 0.0026211238000541925, + 0.02030772715806961, + 0.0036873328499495983, + 0.0033896330278366804, + 0.017053663730621338, + 0.003152497811242938, + 0.0037566765677183867, + 0.026182202622294426, + 0.002077924320474267, + 0.011063608340919018, + 0.00630370294675231, + 0.0024977801367640495, + 0.012519722804427147, + 0.0030010761693120003, + 0.0028960732743144035, + 0.0082091698423028, + 0.0021915731485933065, + 0.007895420305430889, + 0.015797944739460945, + 0.0077240378595888615, + 0.09443879872560501, + 0.011336994357407093, + 0.009351051412522793, + 0.10213879495859146, + 0.006317609455436468, + 0.02024247497320175, + 0.0066921100951731205, + 0.0034537618048489094, + 0.048259977251291275, + 0.006041462067514658, + 0.0046946825459599495, + 0.06564168632030487, + 0.0033503854647278786, + 0.020014209672808647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.01419894304126501 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_7/index_tensor([958, 958])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_7/index_tensor([958, 958])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4abfeeaac431a9c01f6fbf7c85a9b46912840fc2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_7/index_tensor([958, 958])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "striking low-angle, profile tracking shot, filmed in ultra-high frame rate (slow motion), follows a dragon boat at full speed. Under the bright midday sun, a team of dragon boatmen, their muscles muscular and skin drenched with sweat and water spray, each wearing a uniform red quick-drying tank top, paddles with astonishing unison. The camera focuses on a few members mid-boat, their faces focused and determined, their teeth clenched, the muscles in their arms and backs clearly visible with every thrust. They grip the dark wooden oars tightly, and each powerful backward stroke unleashes a shower of shimmering water and white foam. In slow motion, each splashing droplet seems frozen in mid-air, glittering like a diamond in the sunlight before slowly falling to slap against the crimson boat, delicately carved with dragon scales. The background is the wide, sparkling river, and the distant shore is crowded with blurry but enthusiastic cheering crowds. The whole picture is full of intense competition, the beauty of teamwork and the moment of adrenaline rush. The main color is bright red, which forms a strong contrast with the green river water, creating a warm atmosphere full of vitality and passion.", + "seed": 29498, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_7/index_958_seed_29498_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.294437325745821, + "vlm_logprob": -0.294437325745821, + "vlm_yes_no_margin": 1.525, + "vlm_token_credit": -0.294437325745821, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5398253202438354 + ], + "opposite_logprob_tokens": [ + -1.0398253202438354 + ], + "logprob_reward": -0.5398253202438354, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.5398253202438354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23096135258674622, + 0.04796503111720085, + 0.19826725125312805, + 0.05298612266778946, + 0.04516570270061493, + 0.12125903367996216, + 0.057583414018154144, + 0.09006766229867935, + 0.08507288247346878, + 0.07067152112722397 + ], + "credit_map": [ + 0.025516588240861893, + 0.022976983338594437, + 0.03189367428421974, + 0.03155491128563881, + 0.021653268486261368, + 0.02785661444067955, + 0.036466341465711594, + 0.03304298222064972, + 0.00704759918153286, + 0.008588027209043503, + 0.006000131834298372, + 0.005325536709278822, + 0.005795971490442753, + 0.005341501906514168, + 0.005743079818785191, + 0.004123183432966471, + 0.013143952935934067, + 0.0062684365548193455, + 0.06190039590001106, + 0.012690620496869087, + 0.008917175233364105, + 0.011311916634440422, + 0.07324642688035965, + 0.010788348503410816, + 0.0029249803628772497, + 0.004148627631366253, + 0.00427151657640934, + 0.009122931398451328, + 0.003302126657217741, + 0.009116346947848797, + 0.014402488246560097, + 0.005697107873857021, + 0.003642355091869831, + 0.005988017655909061, + 0.006164561025798321, + 0.005338016431778669, + 0.00346061447635293, + 0.006306720431894064, + 0.007912300527095795, + 0.0063531179912388325, + 0.007436756510287523, + 0.006029041483998299, + 0.008016209118068218, + 0.009078273549675941, + 0.009211640805006027, + 0.04343195632100105, + 0.009438201785087585, + 0.028616968542337418, + 0.00839432142674923, + 0.006027822848409414, + 0.008238745853304863, + 0.008113110437989235, + 0.005246213171631098, + 0.005680883768945932, + 0.007077301386743784, + 0.008805018849670887, + 0.005298355594277382, + 0.006043741945177317, + 0.026732800528407097, + 0.03137974441051483, + 0.004481544718146324, + 0.002433023415505886, + 0.004846110008656979, + 0.008852348662912846, + 0.010306144133210182, + 0.01074607390910387, + 0.008168879896402359, + 0.010892648249864578, + 0.014585527591407299, + 0.007969391532242298, + 0.009842303581535816, + 0.012561914511024952, + 0.007849698886275291, + 0.010483928956091404, + 0.013629887253046036, + 0.008574374951422215, + 0.0069109247997403145, + 0.00890278909355402, + 0.007217183709144592, + 0.007102731615304947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.5398253202438354 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25208330154418945 + ], + "opposite_logprob_tokens": [ + -1.5020833015441895 + ], + "logprob_reward": -0.25208330154418945, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25208330154418945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16782209277153015, + 0.06197395175695419, + 0.19560761749744415, + 0.054708082228899, + 0.09828704595565796, + 0.08617957681417465, + 0.04243513196706772, + 0.06663724035024643, + 0.13122518360614777, + 0.09512408077716827 + ], + "credit_map": [ + 0.01923077367246151, + 0.015921201556921005, + 0.018629111349582672, + 0.027128547430038452, + 0.013945386745035648, + 0.02175414375960827, + 0.025924179702997208, + 0.025288769975304604, + 0.006375548429787159, + 0.008560924790799618, + 0.008400903083384037, + 0.007649207953363657, + 0.008871670812368393, + 0.007640080526471138, + 0.00850269477814436, + 0.005972930230200291, + 0.01606709323823452, + 0.009134810417890549, + 0.056813668459653854, + 0.016111714765429497, + 0.008344316855072975, + 0.013194692321121693, + 0.059681281447410583, + 0.016260040923953056, + 0.0036967541091144085, + 0.00477788969874382, + 0.005856840871274471, + 0.011349298059940338, + 0.004148677457123995, + 0.0076390160247683525, + 0.011490507051348686, + 0.005749105010181665, + 0.011381919495761395, + 0.009373688139021397, + 0.010239445604383945, + 0.012350935488939285, + 0.012931359000504017, + 0.01167540717869997, + 0.016784019768238068, + 0.013550280593335629, + 0.006966249085962772, + 0.007450091186910868, + 0.009017568081617355, + 0.012794509530067444, + 0.00958992913365364, + 0.01996849663555622, + 0.006742600351572037, + 0.0136501369997859, + 0.00533220823854208, + 0.006084633059799671, + 0.005299408920109272, + 0.005253243260085583, + 0.005703217815607786, + 0.005981292575597763, + 0.00426536425948143, + 0.0045157670974731445, + 0.00713819544762373, + 0.008010840974748135, + 0.007525418885052204, + 0.00707752862945199, + 0.006804595701396465, + 0.007695710752159357, + 0.00657442444935441, + 0.015810534358024597, + 0.012356875464320183, + 0.012668030336499214, + 0.017706455662846565, + 0.02146008238196373, + 0.017075588926672935, + 0.012570545077323914, + 0.019142096862196922, + 0.0182455163449049, + 0.011827977374196053, + 0.01210054848343134, + 0.015769347548484802, + 0.014012491330504417, + 0.0070779817178845406, + 0.010682478547096252, + 0.010996470227837563, + 0.012656796723604202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.25208330154418945 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821095123887062 + ], + "opposite_logprob_tokens": [ + -4.0182108879089355 + ], + "logprob_reward": -0.01821095123887062, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01821095123887062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09315293282270432, + 0.024682078510522842, + 0.0771055817604065, + 0.08834569156169891, + 0.04507159814238548, + 0.45029139518737793, + 0.024215945973992348, + 0.0366186685860157, + 0.04636808857321739, + 0.11414800584316254 + ], + "credit_map": [ + 0.009894455783069134, + 0.012118445709347725, + 0.011722338385879993, + 0.009962723590433598, + 0.010702826082706451, + 0.014226501807570457, + 0.01049103308469057, + 0.014034615829586983, + 0.003138175467029214, + 0.0028382353484630585, + 0.003175324061885476, + 0.0031561071518808603, + 0.002393827773630619, + 0.0028165101539343596, + 0.0036145669873803854, + 0.003549332730472088, + 0.005678752902895212, + 0.005501719657331705, + 0.020880060270428658, + 0.00695500522851944, + 0.004842998925596476, + 0.005356313660740852, + 0.021966494619846344, + 0.005924235563725233, + 0.005020380485802889, + 0.007312838453799486, + 0.0085042929276824, + 0.02291182242333889, + 0.005342570133507252, + 0.011168470606207848, + 0.01832614652812481, + 0.009759172797203064, + 0.00458621745929122, + 0.007208500523120165, + 0.005608507432043552, + 0.004580637905746698, + 0.007550292648375034, + 0.005592821165919304, + 0.0050213332287967205, + 0.004923290107399225, + 0.027997268363833427, + 0.021800890564918518, + 0.027727635577321053, + 0.02866842970252037, + 0.026028947904706, + 0.1806696355342865, + 0.020090406760573387, + 0.11730816960334778, + 0.0031815562397241592, + 0.002412778791040182, + 0.0033075681421905756, + 0.003295972477644682, + 0.0032877614721655846, + 0.002764178439974785, + 0.0030514735262840986, + 0.002914658049121499, + 0.003976178355515003, + 0.004734269808977842, + 0.006202372256666422, + 0.006999720819294453, + 0.003705863142386079, + 0.003084625117480755, + 0.0033685348462313414, + 0.004547106102108955, + 0.004387959372252226, + 0.0052504390478134155, + 0.004922542721033096, + 0.008112983778119087, + 0.006243031471967697, + 0.004459298215806484, + 0.007206673733890057, + 0.005785161163657904, + 0.012431123293936253, + 0.010705767199397087, + 0.013051957823336124, + 0.01222973596304655, + 0.013974035158753395, + 0.012745601125061512, + 0.0147511912509799, + 0.024258587509393692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.01821095123887062 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486471474170685 + ], + "opposite_logprob_tokens": [ + -1.223647117614746 + ], + "logprob_reward": -0.3486471474170685, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486471474170685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14791709184646606, + 0.03279699757695198, + 0.052028268575668335, + 0.05162511020898819, + 0.04902130737900734, + 0.3945663869380951, + 0.04514916613698006, + 0.14847689867019653, + 0.0543106272816658, + 0.024108126759529114 + ], + "credit_map": [ + 0.014384266920387745, + 0.01618972234427929, + 0.02125909924507141, + 0.01965704746544361, + 0.01742926426231861, + 0.019350236281752586, + 0.02109598182141781, + 0.01855146884918213, + 0.002823412185534835, + 0.004424715880304575, + 0.005683418828994036, + 0.004367772955447435, + 0.003624863689765334, + 0.004100755788385868, + 0.004219635855406523, + 0.0035524186678230762, + 0.0037865578196942806, + 0.0027502395678311586, + 0.01537784468382597, + 0.0037754857912659645, + 0.0037660510279238224, + 0.003421268193051219, + 0.015643730759620667, + 0.0035070886369794607, + 0.004025017376989126, + 0.005731577519327402, + 0.005161420442163944, + 0.011423386633396149, + 0.004778461530804634, + 0.008038720116019249, + 0.007860176265239716, + 0.004606352187693119, + 0.005057858768850565, + 0.0067858612164855, + 0.006312214303761721, + 0.00642875162884593, + 0.005301352124661207, + 0.005624530836939812, + 0.0062543293461203575, + 0.007256406359374523, + 0.024818509817123413, + 0.018869878724217415, + 0.024260802194476128, + 0.024866938591003418, + 0.021660972386598587, + 0.15984198451042175, + 0.01711234636604786, + 0.10313495248556137, + 0.006736014038324356, + 0.004902942571789026, + 0.00520417932420969, + 0.006888524163514376, + 0.0036891910713166, + 0.004834952764213085, + 0.005684950854629278, + 0.007208408787846565, + 0.00863700732588768, + 0.009955178014934063, + 0.04983291029930115, + 0.04417607933282852, + 0.006843091920018196, + 0.0049337237142026424, + 0.009080560877919197, + 0.015018356963992119, + 0.005369569640606642, + 0.0064358217641711235, + 0.007065243553370237, + 0.006096601486206055, + 0.008159712888300419, + 0.0049576470628380775, + 0.008461873047053814, + 0.0077641550451517105, + 0.0027172090485692024, + 0.0032291871029883623, + 0.00438062846660614, + 0.0029464655090123415, + 0.0023380678612738848, + 0.0020791504066437483, + 0.0030934640672057867, + 0.0033239529002457857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.3486471474170685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.313419908285141 + ], + "opposite_logprob_tokens": [ + -1.3134199380874634 + ], + "logprob_reward": -0.313419908285141, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.313419908285141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.066018246114254, + 0.022706566378474236, + 0.13490210473537445, + 0.14574375748634338, + 0.09270982444286346, + 0.32313981652259827, + 0.07669394463300705, + 0.054368771612644196, + 0.0657452866435051, + 0.017971696332097054 + ], + "credit_map": [ + 0.00592529121786356, + 0.006024041213095188, + 0.009747505187988281, + 0.009351424872875214, + 0.005287347827106714, + 0.0076602911576628685, + 0.014220268465578556, + 0.0078020780347287655, + 0.0028269013855606318, + 0.0024907535407692194, + 0.0031247877050191164, + 0.00263402983546257, + 0.003300974378362298, + 0.0032256534323096275, + 0.002597578801214695, + 0.002505888231098652, + 0.009730614721775055, + 0.0071934061124920845, + 0.03994109854102135, + 0.01096577849239111, + 0.006422823760658503, + 0.00735397357493639, + 0.046096522361040115, + 0.007197890430688858, + 0.008474350906908512, + 0.011819322593510151, + 0.014574089087545872, + 0.040017154067754745, + 0.009420870803296566, + 0.016149481758475304, + 0.028900472447276115, + 0.016388019546866417, + 0.009720641188323498, + 0.013265288434922695, + 0.01116017997264862, + 0.011501609347760677, + 0.013396714814007282, + 0.013838307000696659, + 0.010289200581610203, + 0.00953788310289383, + 0.020667580887675285, + 0.015482009388506413, + 0.02022491954267025, + 0.020757246762514114, + 0.018415842205286026, + 0.12922550737857819, + 0.014319512993097305, + 0.08404717594385147, + 0.010162636637687683, + 0.008175340481102467, + 0.009596100077033043, + 0.010867394506931305, + 0.00829312764108181, + 0.008352262899279594, + 0.010141014121472836, + 0.011106071062386036, + 0.0052603548392653465, + 0.006973973009735346, + 0.012328235432505608, + 0.010344592854380608, + 0.005273188930004835, + 0.003977188840508461, + 0.005027702543884516, + 0.005183537025004625, + 0.007161383517086506, + 0.0074500287882983685, + 0.007229488343000412, + 0.007090118248015642, + 0.008788847364485264, + 0.006754007190465927, + 0.010663273744285107, + 0.01060813944786787, + 0.0020029868464916945, + 0.0031247222796082497, + 0.002667403547093272, + 0.001695743529126048, + 0.0014436999335885048, + 0.0022482494823634624, + 0.0017656406853348017, + 0.003023249562829733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.313419908285141 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking low-angle, profile tracking shot, filmed in ultra-high frame rate (slow motion), follows a dragon boat at full speed. Under the bright midday sun, a team of dragon boatmen, their muscles muscular and skin drenched with sweat and water spray, each wearing a uniform red quick-drying tank top, paddles with astonishing unison. The camera focuses on a few members mid-boat, their faces focused and determined, their teeth clenched, the muscles in their arms and backs clearly visible with every thrust. They grip the dark wooden oars tightly, and each powerful backward stroke unleashes a shower of shimmering water and white foam. In slow motion, each splashing droplet seems frozen in mid-air, glittering like a diamond in the sunlight before slowly falling to slap against the crimson boat, delicately carved with dragon scales. The background is the wide, sparkling river, and the distant shore is crowded with blurry but enthusiastic cheering crowds. The whole picture is full of intense competition, the beauty of teamwork and the moment of adrenaline rush. The main color is bright red, which forms a strong contrast with the green river water, creating a warm atmosphere full of vitality and passion.", + "seed": 29499, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_7/index_958_seed_29499_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.314984068274498, + "vlm_logprob": -0.314984068274498, + "vlm_yes_no_margin": 1.525, + "vlm_token_credit": -0.314984068274498, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5771666765213013 + ], + "opposite_logprob_tokens": [ + -0.9521666765213013 + ], + "logprob_reward": -0.5771666765213013, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5771666765213013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.408308207988739, + 0.05971190333366394, + 0.06284357607364655, + 0.026046860963106155, + 0.153860405087471, + 0.043441399931907654, + 0.07182382047176361, + 0.050228316336870193, + 0.04523473232984543, + 0.07850082218647003 + ], + "credit_map": [ + 0.044330038130283356, + 0.015565336681902409, + 0.03861023485660553, + 0.016807174310088158, + 0.017548296600580215, + 0.026333728805184364, + 0.14611226320266724, + 0.10300111025571823, + 0.004333618562668562, + 0.0049543119966983795, + 0.012613307684659958, + 0.01251072995364666, + 0.004384680651128292, + 0.00577890919521451, + 0.004906900692731142, + 0.010229446925222874, + 0.008345149457454681, + 0.007941770367324352, + 0.008001659065485, + 0.00725797563791275, + 0.006782830227166414, + 0.009627689607441425, + 0.007856331765651703, + 0.00703017320483923, + 0.0028101447969675064, + 0.005098327063024044, + 0.0024281577207148075, + 0.001837821793742478, + 0.006009037606418133, + 0.0035273588728159666, + 0.0025087594985961914, + 0.001827253494411707, + 0.0201579537242651, + 0.011816033162176609, + 0.019276494160294533, + 0.011656484566628933, + 0.048102810978889465, + 0.011351138353347778, + 0.011485125869512558, + 0.02001437172293663, + 0.004653236363083124, + 0.003992170561105013, + 0.012168332003057003, + 0.01087899785488844, + 0.003077542409300804, + 0.0036584818735718727, + 0.0020712334662675858, + 0.0029414042364805937, + 0.006624152418226004, + 0.01420183852314949, + 0.008949565701186657, + 0.005268216133117676, + 0.008086996152997017, + 0.011440836824476719, + 0.00889564584940672, + 0.008356571197509766, + 0.00523701636120677, + 0.012198054231703281, + 0.004952060524374247, + 0.004803464747965336, + 0.006697860546410084, + 0.005217785015702248, + 0.005513733718544245, + 0.005608345381915569, + 0.0054478999227285385, + 0.010430391877889633, + 0.0038763920310884714, + 0.004303480964154005, + 0.004471465013921261, + 0.004306730814278126, + 0.006778612732887268, + 0.005619759205728769, + 0.00996729452162981, + 0.00918267946690321, + 0.013608096167445183, + 0.008396966382861137, + 0.005787198431789875, + 0.005516174715012312, + 0.009678849019110203, + 0.016363568603992462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.5771666765213013 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129569411277771 + ], + "opposite_logprob_tokens": [ + -2.237957000732422 + ], + "logprob_reward": -0.1129569411277771, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129569411277771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1563670039176941, + 0.1638457179069519, + 0.041841086000204086, + 0.05419310927391052, + 0.16448286175727844, + 0.05920649692416191, + 0.04524090886116028, + 0.03227894753217697, + 0.09334142506122589, + 0.1892024725675583 + ], + "credit_map": [ + 0.011662056669592857, + 0.00895420927554369, + 0.015409855172038078, + 0.00825753714889288, + 0.009597653523087502, + 0.009376270696520805, + 0.0525679886341095, + 0.04054141417145729, + 0.008265135809779167, + 0.011561490595340729, + 0.04302869364619255, + 0.05440349876880646, + 0.01005992665886879, + 0.00861852616071701, + 0.01210691500455141, + 0.015801507979631424, + 0.006737321149557829, + 0.005500582978129387, + 0.005337056703865528, + 0.005305538885295391, + 0.005889689549803734, + 0.00596074853092432, + 0.0034859497100114822, + 0.0036241943016648293, + 0.009398351423442364, + 0.006234541069716215, + 0.003932547755539417, + 0.0031265520956367254, + 0.017087075859308243, + 0.005237383767962456, + 0.005563525948673487, + 0.003613125765696168, + 0.01900300942361355, + 0.01319245994091034, + 0.02282997965812683, + 0.013614408671855927, + 0.047257568687200546, + 0.014057944528758526, + 0.01387192402034998, + 0.020655563101172447, + 0.006018475629389286, + 0.0053855013102293015, + 0.011465819552540779, + 0.013756432570517063, + 0.004060131497681141, + 0.009959070011973381, + 0.0041156248189508915, + 0.004445436876267195, + 0.0048295436426997185, + 0.004125588107854128, + 0.005061857867985964, + 0.004295213147997856, + 0.006486611440777779, + 0.008701389655470848, + 0.006630833260715008, + 0.00510987127199769, + 0.004116109572350979, + 0.005519150756299496, + 0.004373625852167606, + 0.0039297593757510185, + 0.004296414088457823, + 0.003580688964575529, + 0.0031229716259986162, + 0.0033402228727936745, + 0.008037925697863102, + 0.012039806693792343, + 0.011051036417484283, + 0.011106802150607109, + 0.013441285118460655, + 0.010952075943350792, + 0.014458133839070797, + 0.012254349887371063, + 0.01806412637233734, + 0.01936742477118969, + 0.01879400573670864, + 0.011645627208054066, + 0.03929229825735092, + 0.01820462755858898, + 0.0282259751111269, + 0.03560836240649223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.1129569411277771 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026384279131889343 + ], + "opposite_logprob_tokens": [ + -3.6513843536376953 + ], + "logprob_reward": -0.026384279131889343, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026384279131889343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4211253821849823, + 0.04678920656442642, + 0.026283835992217064, + 0.02282406948506832, + 0.08782251924276352, + 0.03671425208449364, + 0.12591339647769928, + 0.0686609148979187, + 0.028618991374969482, + 0.13524745404720306 + ], + "credit_map": [ + 0.1298503428697586, + 0.032166752964258194, + 0.019759954884648323, + 0.020038632676005363, + 0.027232060208916664, + 0.03835107013583183, + 0.0406293161213398, + 0.11309724301099777, + 0.0034804765600711107, + 0.0043976688757538795, + 0.008380611427128315, + 0.009710673242807388, + 0.005486803129315376, + 0.005641196388751268, + 0.004540406633168459, + 0.005151369608938694, + 0.004084492102265358, + 0.004894076846539974, + 0.003947635181248188, + 0.002816474763676524, + 0.0034096038434654474, + 0.002786132972687483, + 0.0023036247584968805, + 0.002041794592514634, + 0.0026903734542429447, + 0.003554312977939844, + 0.001533715520054102, + 0.0016591961029917002, + 0.007218718063086271, + 0.002714061178267002, + 0.0020723689813166857, + 0.0013813234400004148, + 0.0123569555580616, + 0.006634067744016647, + 0.011988509446382523, + 0.007078394293785095, + 0.02420666627585888, + 0.006832965649664402, + 0.0060461475513875484, + 0.012678810395300388, + 0.003468033391982317, + 0.004924147855490446, + 0.009582477621734142, + 0.004999232012778521, + 0.002808475401252508, + 0.00567484088242054, + 0.002975825686007738, + 0.002281219931319356, + 0.011372186243534088, + 0.01924390345811844, + 0.01615675911307335, + 0.009615771472454071, + 0.014599784277379513, + 0.025254681706428528, + 0.016069740056991577, + 0.013600561767816544, + 0.006401205435395241, + 0.012532904744148254, + 0.00811484083533287, + 0.006734172347933054, + 0.011892047710716724, + 0.0074169528670609, + 0.008251396007835865, + 0.00731739541515708, + 0.0034599846694618464, + 0.009287948720157146, + 0.002675039926543832, + 0.002699329052120447, + 0.002755231922492385, + 0.002774567576125264, + 0.002463237615302205, + 0.0025036518927663565, + 0.011406092904508114, + 0.015058625489473343, + 0.009870125912129879, + 0.01090062316507101, + 0.026740210130810738, + 0.00833939015865326, + 0.04214233160018921, + 0.010790059342980385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.026384279131889343 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328371167182922 + ], + "opposite_logprob_tokens": [ + -0.7578371167182922 + ], + "logprob_reward": -0.6328371167182922, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328371167182922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4472346901893616, + 0.0542423278093338, + 0.06186775863170624, + 0.1017046868801117, + 0.05153556168079376, + 0.059244006872177124, + 0.049649979919195175, + 0.04051452502608299, + 0.09353534877300262, + 0.040471192449331284 + ], + "credit_map": [ + 0.11751779913902283, + 0.031482476741075516, + 0.03393440693616867, + 0.021893197670578957, + 0.032859351485967636, + 0.046062011271715164, + 0.10373367369174957, + 0.05975176393985748, + 0.004796003457158804, + 0.006378485355526209, + 0.0074988980777561665, + 0.006762595381587744, + 0.006144380662590265, + 0.006669654976576567, + 0.006429036147892475, + 0.009563268162310123, + 0.0061912741512060165, + 0.008421109989285469, + 0.006615742109715939, + 0.008960351347923279, + 0.005638771690428257, + 0.010120498947799206, + 0.00805127713829279, + 0.007868730463087559, + 0.018367774784564972, + 0.01435632724314928, + 0.006994124501943588, + 0.0047348481602966785, + 0.03206022456288338, + 0.010474100708961487, + 0.009721565991640091, + 0.00499572092667222, + 0.00686431722715497, + 0.0038333965931087732, + 0.005133167374879122, + 0.004082908388227224, + 0.015838373452425003, + 0.0034478865563869476, + 0.004194546956568956, + 0.00814096350222826, + 0.006075398996472359, + 0.012960904277861118, + 0.013509795069694519, + 0.0050663105212152, + 0.004204821307212114, + 0.007433109916746616, + 0.003813505405560136, + 0.006180156487971544, + 0.0051328944973647594, + 0.00462399423122406, + 0.005422562826424837, + 0.004490312188863754, + 0.0071685523726046085, + 0.010168247856199741, + 0.007318025454878807, + 0.005325387232005596, + 0.003842160338535905, + 0.008449552580714226, + 0.00493622338399291, + 0.003886121790856123, + 0.007156034000217915, + 0.0041705030016601086, + 0.0037419674918055534, + 0.004331955220550299, + 0.009664958342909813, + 0.015721186995506287, + 0.009748722426593304, + 0.008930139243602753, + 0.011286369524896145, + 0.011359951458871365, + 0.013018363155424595, + 0.013805649243295193, + 0.004646761808544397, + 0.0033966165501624346, + 0.005575569812208414, + 0.0032288178335875273, + 0.0036482021678239107, + 0.0030229208059608936, + 0.006157395429909229, + 0.010794905945658684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.6328371167182922 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557532787322998 + ], + "opposite_logprob_tokens": [ + -1.60057532787323 + ], + "logprob_reward": -0.22557532787322998, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557532787322998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3943066895008087, + 0.06312217563390732, + 0.05963948369026184, + 0.15383806824684143, + 0.05911482870578766, + 0.11130556464195251, + 0.05165993422269821, + 0.023947248235344887, + 0.02388087473809719, + 0.05918509513139725 + ], + "credit_map": [ + 0.08238711208105087, + 0.02560487762093544, + 0.040514230728149414, + 0.020426053553819656, + 0.026501698419451714, + 0.027827559038996696, + 0.13850390911102295, + 0.032541241496801376, + 0.007553601171821356, + 0.02180960588157177, + 0.006958295591175556, + 0.006257540546357632, + 0.006179631222039461, + 0.005530608352273703, + 0.003935550805181265, + 0.004897336475551128, + 0.005568396300077438, + 0.008792162872850895, + 0.005772109143435955, + 0.008245657198131084, + 0.005252555478364229, + 0.009731539525091648, + 0.008380994200706482, + 0.00789607036858797, + 0.028736313804984093, + 0.028353219851851463, + 0.01179030817002058, + 0.006959185935556889, + 0.041278768330812454, + 0.01670895330607891, + 0.012181145139038563, + 0.007830158807337284, + 0.010544732213020325, + 0.0046366136521101, + 0.011229541152715683, + 0.006937980651855469, + 0.006561472080647945, + 0.006005197297781706, + 0.005627929233014584, + 0.007571361493319273, + 0.01135533768683672, + 0.017871445044875145, + 0.020886283367872238, + 0.015962887555360794, + 0.010644501075148582, + 0.01282335165888071, + 0.008044011890888214, + 0.01371774822473526, + 0.005173420067876577, + 0.013925842940807343, + 0.005999282468110323, + 0.0038923961110413074, + 0.004762192256748676, + 0.004474744666367769, + 0.007303299382328987, + 0.006128753535449505, + 0.002277676248922944, + 0.004717857111245394, + 0.0036534506361931562, + 0.002713675145059824, + 0.0031987670809030533, + 0.002389473607763648, + 0.0028726186137646437, + 0.0021237286273390055, + 0.0031556254252791405, + 0.0061588897369802, + 0.0026669418439269066, + 0.0025911261327564716, + 0.0021329852752387524, + 0.0023548381868749857, + 0.002671692054718733, + 0.002148775849491358, + 0.00519529078155756, + 0.006758348550647497, + 0.003733486868441105, + 0.005237241741269827, + 0.006105959415435791, + 0.0025261431001126766, + 0.022652821615338326, + 0.006975803058594465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.22557532787322998 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking low-angle, profile tracking shot, filmed in ultra-high frame rate (slow motion), follows a dragon boat at full speed. Under the bright midday sun, a team of dragon boatmen, their muscles muscular and skin drenched with sweat and water spray, each wearing a uniform red quick-drying tank top, paddles with astonishing unison. The camera focuses on a few members mid-boat, their faces focused and determined, their teeth clenched, the muscles in their arms and backs clearly visible with every thrust. They grip the dark wooden oars tightly, and each powerful backward stroke unleashes a shower of shimmering water and white foam. In slow motion, each splashing droplet seems frozen in mid-air, glittering like a diamond in the sunlight before slowly falling to slap against the crimson boat, delicately carved with dragon scales. The background is the wide, sparkling river, and the distant shore is crowded with blurry but enthusiastic cheering crowds. The whole picture is full of intense competition, the beauty of teamwork and the moment of adrenaline rush. The main color is bright red, which forms a strong contrast with the green river water, creating a warm atmosphere full of vitality and passion.", + "seed": 29500, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_7/index_958_seed_29500_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3315069295465946, + "vlm_logprob": -0.3315069295465946, + "vlm_yes_no_margin": 1.349999988079071, + "vlm_token_credit": -0.3315069295465946, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6859123706817627 + ], + "opposite_logprob_tokens": [ + -0.8109123706817627 + ], + "logprob_reward": -0.6859123706817627, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6859123706817627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21599683165550232, + 0.21550703048706055, + 0.1266453117132187, + 0.07977679371833801, + 0.048018842935562134, + 0.0606376975774765, + 0.02565605193376541, + 0.04804689437150955, + 0.042017482221126556, + 0.13769696652889252 + ], + "credit_map": [ + 0.0245975274592638, + 0.040581755340099335, + 0.026386097073554993, + 0.023441966623067856, + 0.026755543425679207, + 0.02849237620830536, + 0.024979762732982635, + 0.020761806517839432, + 0.023540526628494263, + 0.01410119142383337, + 0.024003304541110992, + 0.02757571078836918, + 0.026610465720295906, + 0.015650207176804543, + 0.06571207195520401, + 0.0183135736733675, + 0.010769104585051537, + 0.011949187144637108, + 0.03737250715494156, + 0.0070608328096568584, + 0.03225218504667282, + 0.009471778757870197, + 0.01194657851010561, + 0.005823139566928148, + 0.0138228265568614, + 0.010391277261078358, + 0.009453761391341686, + 0.011405783705413342, + 0.007189308293163776, + 0.009611258283257484, + 0.007315022870898247, + 0.010587556287646294, + 0.0066067869774997234, + 0.005850883666425943, + 0.003402438247576356, + 0.004208178725093603, + 0.011223806999623775, + 0.006568890996277332, + 0.0052553038112819195, + 0.0049025556072592735, + 0.008316789753735065, + 0.005572364665567875, + 0.005881284363567829, + 0.005834862589836121, + 0.006674746982753277, + 0.009905418381094933, + 0.008046523667871952, + 0.0104057090356946, + 0.003193796379491687, + 0.0027629495598375797, + 0.0027762700337916613, + 0.003724663984030485, + 0.003816785756498575, + 0.0037178602069616318, + 0.0028017074801027775, + 0.002862019930034876, + 0.003940754570066929, + 0.004396852105855942, + 0.0071445791982114315, + 0.006933624856173992, + 0.005891809705644846, + 0.005273076239973307, + 0.00924857147037983, + 0.00521762715652585, + 0.005451199132949114, + 0.004440873861312866, + 0.006509492639452219, + 0.004863400012254715, + 0.004730285611003637, + 0.005882171913981438, + 0.005096162669360638, + 0.005043894983828068, + 0.0162130706012249, + 0.018518676981329918, + 0.016826869919896126, + 0.01858997903764248, + 0.016720063984394073, + 0.022129084914922714, + 0.01233246922492981, + 0.01636675000190735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.6859123706817627 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037358343601227 + ], + "opposite_logprob_tokens": [ + -1.910373568534851 + ], + "logprob_reward": -0.16037358343601227, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037358343601227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15831045806407928, + 0.27874383330345154, + 0.10394051671028137, + 0.1473776251077652, + 0.05037172883749008, + 0.057059261947870255, + 0.027279507368803024, + 0.04725956544280052, + 0.033601101487874985, + 0.09605637937784195 + ], + "credit_map": [ + 0.017969369888305664, + 0.024090180173516273, + 0.01930905319750309, + 0.025190092623233795, + 0.0173292588442564, + 0.016907772049307823, + 0.020862214267253876, + 0.016652513295412064, + 0.04141799360513687, + 0.017250845208764076, + 0.03586006909608841, + 0.033409200608730316, + 0.02400391735136509, + 0.031532447785139084, + 0.06712517887353897, + 0.02814420685172081, + 0.0070865866728127, + 0.006948093883693218, + 0.03317464515566826, + 0.004548692610114813, + 0.030069760978221893, + 0.008317219093441963, + 0.01020803302526474, + 0.003587479703128338, + 0.023139551281929016, + 0.02201930247247219, + 0.020441250875592232, + 0.019912630319595337, + 0.013672967441380024, + 0.013874934986233711, + 0.019372237846255302, + 0.01494474895298481, + 0.006371951662003994, + 0.006252849940210581, + 0.004264494404196739, + 0.0042557488195598125, + 0.011905438266694546, + 0.005764735396951437, + 0.0072615318931639194, + 0.004294977989047766, + 0.005757455714046955, + 0.005137050058692694, + 0.005632185377180576, + 0.004655599594116211, + 0.007210585754364729, + 0.01631327159702778, + 0.004129655193537474, + 0.008223459124565125, + 0.003983480855822563, + 0.0033720543142408133, + 0.0022455560974776745, + 0.0035511585883796215, + 0.004220157861709595, + 0.0034514868166297674, + 0.0021921575535088778, + 0.004263454582542181, + 0.004482559859752655, + 0.00567555520683527, + 0.0070567261427640915, + 0.008692329749464989, + 0.0047023058868944645, + 0.004932266194373369, + 0.006430398672819138, + 0.005287427455186844, + 0.0033297440968453884, + 0.003147983457893133, + 0.003162162844091654, + 0.0058670504949986935, + 0.00446733646094799, + 0.004158876836299896, + 0.005393394734710455, + 0.004074551165103912, + 0.010671239346265793, + 0.015413176268339157, + 0.01057941373437643, + 0.01559050939977169, + 0.01110018603503704, + 0.013549127615988255, + 0.009589210152626038, + 0.009563518688082695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.16037358343601227 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033722881227731705 + ], + "opposite_logprob_tokens": [ + -3.4087228775024414 + ], + "logprob_reward": -0.033722881227731705, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033722881227731705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36649027466773987, + 0.10024823993444443, + 0.10813496261835098, + 0.13078083097934723, + 0.030820267274975777, + 0.0565236434340477, + 0.05134080722928047, + 0.037846751511096954, + 0.04701587185263634, + 0.0707983523607254 + ], + "credit_map": [ + 0.05228830501437187, + 0.032455138862133026, + 0.0563618466258049, + 0.0362672358751297, + 0.04255910590291023, + 0.06306836009025574, + 0.043545957654714584, + 0.03994431719183922, + 0.01123177818953991, + 0.005728968884795904, + 0.013773656450212002, + 0.01273910328745842, + 0.00895323883742094, + 0.01001095212996006, + 0.02733839862048626, + 0.010472138412296772, + 0.009865086525678635, + 0.00908316858112812, + 0.020841872319579124, + 0.008083608001470566, + 0.024733958765864372, + 0.010271057486534119, + 0.014786796644330025, + 0.010469418950378895, + 0.012730400077998638, + 0.020325113087892532, + 0.01967446133494377, + 0.01310378685593605, + 0.01743442378938198, + 0.021449171006679535, + 0.012304440140724182, + 0.013759028166532516, + 0.0038467722479254007, + 0.0029301282484084368, + 0.0028056285809725523, + 0.0029355548322200775, + 0.004490548744797707, + 0.005601856857538223, + 0.004074335563927889, + 0.004135443363338709, + 0.006810745224356651, + 0.004572554957121611, + 0.007932638749480247, + 0.009013454429805279, + 0.006065275054425001, + 0.006683127488940954, + 0.01011724304407835, + 0.005328606814146042, + 0.00728223379701376, + 0.005001546815037727, + 0.005664809141308069, + 0.006884527392685413, + 0.008927753195166588, + 0.005687213037163019, + 0.005920116323977709, + 0.00597260519862175, + 0.0032698463182896376, + 0.005010644905269146, + 0.00514609320089221, + 0.0052515314891934395, + 0.0037499950267374516, + 0.004223865922540426, + 0.007352778222411871, + 0.0038419924676418304, + 0.004104910418391228, + 0.005099354311823845, + 0.00303410436026752, + 0.006864231079816818, + 0.006313625257462263, + 0.00819613877683878, + 0.00429306086152792, + 0.009110447950661182, + 0.009911740198731422, + 0.009466146118938923, + 0.007929522544145584, + 0.006994805298745632, + 0.00969019066542387, + 0.011324109509587288, + 0.007914742454886436, + 0.0075670997612178326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.033722881227731705 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289124608039856 + ], + "opposite_logprob_tokens": [ + -1.0539124011993408 + ], + "logprob_reward": -0.4289124608039856, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.4289124608039856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1033814400434494, + 0.06593543291091919, + 0.5392967462539673, + 0.09783107787370682, + 0.049577511847019196, + 0.03411123901605606, + 0.014035340398550034, + 0.035665303468704224, + 0.01642179675400257, + 0.0437440425157547 + ], + "credit_map": [ + 0.013534769415855408, + 0.011585776694118977, + 0.01162646058946848, + 0.018815064802765846, + 0.011543034575879574, + 0.010518206283450127, + 0.014874473214149475, + 0.010883663780987263, + 0.012081857770681381, + 0.007766249589622021, + 0.008053446188569069, + 0.00821683183312416, + 0.008435150608420372, + 0.005870316177606583, + 0.00811917893588543, + 0.007392407860606909, + 0.03771905228495598, + 0.04419223219156265, + 0.1471300721168518, + 0.026349350810050964, + 0.14937767386436462, + 0.03994656726717949, + 0.06403561681509018, + 0.030546264722943306, + 0.008180230855941772, + 0.009750982746481895, + 0.011527598835527897, + 0.012719972059130669, + 0.011203230358660221, + 0.016985861584544182, + 0.012547052465379238, + 0.014916151762008667, + 0.008019298315048218, + 0.004658028483390808, + 0.004602209199219942, + 0.0036054893862456083, + 0.010674668475985527, + 0.005828387103974819, + 0.00723423482850194, + 0.004955200012773275, + 0.0029525086283683777, + 0.0034124827943742275, + 0.004093111492693424, + 0.0051154447719454765, + 0.003231508657336235, + 0.006544728763401508, + 0.005109742749482393, + 0.003651712089776993, + 0.001418462023139, + 0.0018140784231945872, + 0.0020907451398670673, + 0.001650795340538025, + 0.002406420186161995, + 0.00151624099817127, + 0.0012915085535496473, + 0.001847091014496982, + 0.0035305647179484367, + 0.006400646176189184, + 0.003844096325337887, + 0.004780810791999102, + 0.0031712809577584267, + 0.005186787806451321, + 0.004048595204949379, + 0.004702522419393063, + 0.0015718197682872415, + 0.0018850258784368634, + 0.0019402125617489219, + 0.0022438974119722843, + 0.0018080680165439844, + 0.002676398027688265, + 0.002394601935520768, + 0.0019017739687114954, + 0.005005135200917721, + 0.004224831238389015, + 0.005154580809175968, + 0.004639098420739174, + 0.004177868366241455, + 0.004017847590148449, + 0.00941895879805088, + 0.007105721160769463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.4289124608039856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34861335158348083 + ], + "opposite_logprob_tokens": [ + -1.2236133813858032 + ], + "logprob_reward": -0.34861335158348083, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34861335158348083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15129123628139496, + 0.1798778772354126, + 0.21078698337078094, + 0.06682687997817993, + 0.07722029089927673, + 0.14618800580501556, + 0.03041122294962406, + 0.07432539761066437, + 0.028295908123254776, + 0.03477620333433151 + ], + "credit_map": [ + 0.015883909538388252, + 0.01611947827041149, + 0.02123369462788105, + 0.022434284910559654, + 0.017198555171489716, + 0.019701382145285606, + 0.0219934843480587, + 0.01672643981873989, + 0.01730371080338955, + 0.010860719718039036, + 0.024670211598277092, + 0.023870686069130898, + 0.01771964505314827, + 0.018338249996304512, + 0.04961148649454117, + 0.017503170296549797, + 0.015116054564714432, + 0.015762941911816597, + 0.06278669834136963, + 0.00984756276011467, + 0.06033027917146683, + 0.014998627826571465, + 0.01914893090724945, + 0.012795883230865002, + 0.007283388637006283, + 0.010128839872777462, + 0.006832826416939497, + 0.008656412363052368, + 0.006222024094313383, + 0.010040570050477982, + 0.008782623335719109, + 0.008880198001861572, + 0.007578377611935139, + 0.00798705406486988, + 0.007692576386034489, + 0.005327671766281128, + 0.014777942560613155, + 0.010122756473720074, + 0.011625142768025398, + 0.012108772993087769, + 0.02123577520251274, + 0.014060582034289837, + 0.01833435520529747, + 0.019465459510684013, + 0.011983922682702541, + 0.02086271531879902, + 0.0203205868601799, + 0.01992461271584034, + 0.004440081771463156, + 0.0031052380800247192, + 0.0035124567802995443, + 0.004374693613499403, + 0.0038996960502117872, + 0.003627072786912322, + 0.0036858306266367435, + 0.003766152774915099, + 0.008508463390171528, + 0.008196589536964893, + 0.011182497255504131, + 0.009232722222805023, + 0.00987791083753109, + 0.006682228296995163, + 0.011397164314985275, + 0.009247826412320137, + 0.004203062504529953, + 0.002691888017579913, + 0.0033825072459876537, + 0.0030231447890400887, + 0.003145005088299513, + 0.006035359110683203, + 0.0027597350999712944, + 0.0030552048701792955, + 0.0066797747276723385, + 0.004363712854683399, + 0.004739854484796524, + 0.0033682698849588633, + 0.0029994978103786707, + 0.0033460562117397785, + 0.004944819025695324, + 0.004334218334406614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.34861335158348083 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking low-angle, profile tracking shot, filmed in ultra-high frame rate (slow motion), follows a dragon boat at full speed. Under the bright midday sun, a team of dragon boatmen, their muscles muscular and skin drenched with sweat and water spray, each wearing a uniform red quick-drying tank top, paddles with astonishing unison. The camera focuses on a few members mid-boat, their faces focused and determined, their teeth clenched, the muscles in their arms and backs clearly visible with every thrust. They grip the dark wooden oars tightly, and each powerful backward stroke unleashes a shower of shimmering water and white foam. In slow motion, each splashing droplet seems frozen in mid-air, glittering like a diamond in the sunlight before slowly falling to slap against the crimson boat, delicately carved with dragon scales. The background is the wide, sparkling river, and the distant shore is crowded with blurry but enthusiastic cheering crowds. The whole picture is full of intense competition, the beauty of teamwork and the moment of adrenaline rush. The main color is bright red, which forms a strong contrast with the green river water, creating a warm atmosphere full of vitality and passion.", + "seed": 29501, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_7/index_958_seed_29501_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.24400933384895324, + "vlm_logprob": -0.24400933384895324, + "vlm_yes_no_margin": 1.775, + "vlm_token_credit": -0.24400933384895324, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5461934804916382 + ], + "opposite_logprob_tokens": [ + -1.0461934804916382 + ], + "logprob_reward": -0.5461934804916382, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.5461934804916382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2364775687456131, + 0.03946895897388458, + 0.10054142028093338, + 0.07012277096509933, + 0.13239149749279022, + 0.05270994082093239, + 0.04727213829755783, + 0.11910601705312729, + 0.10764513164758682, + 0.09426457434892654 + ], + "credit_map": [ + 0.0330333337187767, + 0.026522086933255196, + 0.026300547644495964, + 0.02678130380809307, + 0.027387375012040138, + 0.0254583191126585, + 0.029150465503335, + 0.04184413701295853, + 0.005836655385792255, + 0.006255564745515585, + 0.003627012250944972, + 0.006090554874390364, + 0.004897136706858873, + 0.005524123087525368, + 0.0032445371616631746, + 0.003993374761193991, + 0.00875436794012785, + 0.015205743722617626, + 0.011410199105739594, + 0.011677329428493977, + 0.021768659353256226, + 0.01548369787633419, + 0.008978469297289848, + 0.007262950297445059, + 0.009762377478182316, + 0.008017306216061115, + 0.006295404862612486, + 0.008182663470506668, + 0.009884527884423733, + 0.010502535849809647, + 0.009155158884823322, + 0.00832279585301876, + 0.020023588091135025, + 0.0141823198646307, + 0.032627880573272705, + 0.0103226313367486, + 0.010082639753818512, + 0.018954964354634285, + 0.017005862668156624, + 0.009191612713038921, + 0.007339567877352238, + 0.006542342249304056, + 0.0058205812238156796, + 0.00763274310156703, + 0.006749266292899847, + 0.0062674591317772865, + 0.005592368543148041, + 0.0067656091414391994, + 0.00501752644777298, + 0.007644661236554384, + 0.004647072404623032, + 0.006863635033369064, + 0.005592795088887215, + 0.0075042168609797955, + 0.0036500459536910057, + 0.006352183874696493, + 0.013144118711352348, + 0.014748853631317616, + 0.011746433563530445, + 0.009701807983219624, + 0.02519182115793228, + 0.014543161727488041, + 0.01702946238219738, + 0.013000359758734703, + 0.009014475159347057, + 0.010160461068153381, + 0.03273438289761543, + 0.007675924804061651, + 0.007263739127665758, + 0.021549565717577934, + 0.013015306554734707, + 0.006231280975043774, + 0.010265853255987167, + 0.0048081842251122, + 0.009370541200041771, + 0.006621936336159706, + 0.03320197016000748, + 0.007699473761022091, + 0.00802265852689743, + 0.014273962005972862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.5461934804916382 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033480077981949 + ], + "opposite_logprob_tokens": [ + -2.350334882736206 + ], + "logprob_reward": -0.10033480077981949, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033480077981949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14945948123931885, + 0.04286765679717064, + 0.10694713145494461, + 0.06405894458293915, + 0.061802938580513, + 0.06162156164646149, + 0.07384052872657776, + 0.11112000048160553, + 0.24390308558940887, + 0.08437865227460861 + ], + "credit_map": [ + 0.02053224667906761, + 0.011085721664130688, + 0.018164312466979027, + 0.013984850607812405, + 0.025855856016278267, + 0.015204912051558495, + 0.027232836931943893, + 0.017398754134774208, + 0.005504295229911804, + 0.004556107334792614, + 0.0033804564736783504, + 0.0061188810504972935, + 0.005466627422720194, + 0.00861317291855812, + 0.004391795955598354, + 0.0048363227397203445, + 0.009036066010594368, + 0.013649197295308113, + 0.011086215265095234, + 0.010733889415860176, + 0.012104240246117115, + 0.018254967406392097, + 0.022683419287204742, + 0.00939913559705019, + 0.00924993772059679, + 0.009906819090247154, + 0.00579252140596509, + 0.0052214195020496845, + 0.007988485507667065, + 0.007405534386634827, + 0.010142220184206963, + 0.00835200771689415, + 0.01682639867067337, + 0.007366852834820747, + 0.004956535995006561, + 0.005133545957505703, + 0.006033675279468298, + 0.00666766008362174, + 0.01140849944204092, + 0.0034097719471901655, + 0.010262987576425076, + 0.011704372242093086, + 0.005738046485930681, + 0.005777556449174881, + 0.01133815385401249, + 0.00748946238309145, + 0.004403387196362019, + 0.004907594062387943, + 0.007497555110603571, + 0.01599513739347458, + 0.007002339698374271, + 0.012504403479397297, + 0.008444503881037235, + 0.00719305407255888, + 0.006669065449386835, + 0.008534476161003113, + 0.013245781883597374, + 0.011519428342580795, + 0.012188228778541088, + 0.010655689984560013, + 0.02650546096265316, + 0.008338356390595436, + 0.018390655517578125, + 0.01027639489620924, + 0.015193412080407143, + 0.013799106702208519, + 0.05929365009069443, + 0.011787955649197102, + 0.012823417782783508, + 0.03481573611497879, + 0.08256497234106064, + 0.013624818064272404, + 0.009849648922681808, + 0.009063961915671825, + 0.009808619506657124, + 0.010274817235767841, + 0.025843452662229538, + 0.005743800662457943, + 0.007809738162904978, + 0.0059846132062375546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.10033480077981949 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331509441137314 + ], + "opposite_logprob_tokens": [ + -3.773315191268921 + ], + "logprob_reward": -0.02331509441137314, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331509441137314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1629403978586197, + 0.019815955311059952, + 0.26240822672843933, + 0.026767341420054436, + 0.07709702104330063, + 0.03168930858373642, + 0.01626882515847683, + 0.035005584359169006, + 0.28949734568595886, + 0.07851001620292664 + ], + "credit_map": [ + 0.03361125290393829, + 0.014626986347138882, + 0.020141394808888435, + 0.01544420886784792, + 0.0242836382240057, + 0.015797289088368416, + 0.023084403946995735, + 0.01595122367143631, + 0.0023932140320539474, + 0.0022494688164442778, + 0.001641563605517149, + 0.0027146488428115845, + 0.002386916894465685, + 0.004332001321017742, + 0.0022937438916414976, + 0.0018043973250314593, + 0.01959703490138054, + 0.04049944877624512, + 0.024785244837403297, + 0.027877667918801308, + 0.0438474602997303, + 0.056266579777002335, + 0.03258213400840759, + 0.016952650621533394, + 0.004068403504788876, + 0.0054360018111765385, + 0.0016803044127300382, + 0.002518119290471077, + 0.003183696186169982, + 0.0029340358451008797, + 0.0027624384965747595, + 0.0041843438521027565, + 0.019069356843829155, + 0.008449682965874672, + 0.0063116769306361675, + 0.0042305937968194485, + 0.008562401868402958, + 0.007568519562482834, + 0.017501693218946457, + 0.005403090268373489, + 0.005186144262552261, + 0.0053068422712385654, + 0.002751758089289069, + 0.0025321978610008955, + 0.005967280827462673, + 0.0049653747119009495, + 0.0022265827283263206, + 0.002753127831965685, + 0.0017737626330927014, + 0.002436053706333041, + 0.001220064121298492, + 0.0020042697433382273, + 0.0021223174408078194, + 0.0024183194618672132, + 0.0024925542529672384, + 0.0018014837987720966, + 0.005008274223655462, + 0.00525972293689847, + 0.003690004348754883, + 0.0027637514285743237, + 0.006912570912390947, + 0.004436012357473373, + 0.003637615591287613, + 0.003297633258625865, + 0.018150465562939644, + 0.01678868755698204, + 0.06979772448539734, + 0.015828153118491173, + 0.014330409467220306, + 0.040385130792856216, + 0.09719708561897278, + 0.01701970212161541, + 0.006247872486710548, + 0.006103090010583401, + 0.021191611886024475, + 0.009348496794700623, + 0.021508676931262016, + 0.003804379142820835, + 0.005300784949213266, + 0.005005111452192068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.02331509441137314 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486550748348236 + ], + "opposite_logprob_tokens": [ + -1.223655104637146 + ], + "logprob_reward": -0.3486550748348236, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486550748348236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1468435525894165, + 0.03176495060324669, + 0.09097070246934891, + 0.05148432031273842, + 0.11135778576135635, + 0.0537370890378952, + 0.05296912416815758, + 0.0660930871963501, + 0.18639953434467316, + 0.2083798199892044 + ], + "credit_map": [ + 0.027631431818008423, + 0.013059698976576328, + 0.011795202270150185, + 0.014813419431447983, + 0.025438038632273674, + 0.013978064060211182, + 0.019864968955516815, + 0.020262736827135086, + 0.0036446833983063698, + 0.004483756609261036, + 0.0030188525561243296, + 0.00559769943356514, + 0.003620852716267109, + 0.004103187005966902, + 0.003731247503310442, + 0.003564670914784074, + 0.008076545782387257, + 0.016512667760252953, + 0.013404308818280697, + 0.009162147529423237, + 0.017643969506025314, + 0.011015334166586399, + 0.009238657541573048, + 0.00591707369312644, + 0.007521109655499458, + 0.011121116578578949, + 0.004572249483317137, + 0.00508188595995307, + 0.006174991838634014, + 0.005132399499416351, + 0.005833798088133335, + 0.006046771537512541, + 0.017566397786140442, + 0.012848827987909317, + 0.024465754628181458, + 0.012758546508848667, + 0.009382217191159725, + 0.01321425661444664, + 0.012763342820107937, + 0.008358445018529892, + 0.008696144446730614, + 0.009190735407173634, + 0.004244490992277861, + 0.004393809475004673, + 0.009711996652185917, + 0.008088668808341026, + 0.004091176204383373, + 0.005320069845765829, + 0.004213084001094103, + 0.008171775378286839, + 0.005180567502975464, + 0.004453347530215979, + 0.004890612792223692, + 0.004905509762465954, + 0.01406294759362936, + 0.007091281469911337, + 0.010025879368185997, + 0.01193930208683014, + 0.005941235926002264, + 0.004261198453605175, + 0.010042022913694382, + 0.01095713023096323, + 0.006387685425579548, + 0.006538630463182926, + 0.013331202790141106, + 0.0073022800497710705, + 0.06726524233818054, + 0.01036972925066948, + 0.007052515167742968, + 0.0142317283898592, + 0.05230147764086723, + 0.014545355923473835, + 0.04605996608734131, + 0.010054350830614567, + 0.03933146968483925, + 0.013008487410843372, + 0.04243813082575798, + 0.013285850174725056, + 0.012849660590291023, + 0.031351909041404724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.3486550748348236 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a profile tracking shot following the dragon boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dragon boatmen wearing red quick-drying tank tops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon boat dark in color with dragon scales carved on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in ultra-high frame rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20154821872711182 + ], + "opposite_logprob_tokens": [ + -1.7015482187271118 + ], + "logprob_reward": -0.20154821872711182, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20154821872711182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11935128271579742, + 0.07488718628883362, + 0.08312389999628067, + 0.12061542272567749, + 0.1269044131040573, + 0.17866572737693787, + 0.0637703388929367, + 0.07604172080755234, + 0.08690334111452103, + 0.06973659992218018 + ], + "credit_map": [ + 0.018132438883185387, + 0.011433672159910202, + 0.01531878113746643, + 0.01433663908392191, + 0.020433716475963593, + 0.01209072582423687, + 0.01280178315937519, + 0.01480354368686676, + 0.009740112349390984, + 0.011828883551061153, + 0.007203094661235809, + 0.009933809749782085, + 0.011092236265540123, + 0.010085850022733212, + 0.008790763095021248, + 0.006212444510310888, + 0.010159258730709553, + 0.009918375872075558, + 0.018965894356369972, + 0.006337794475257397, + 0.008221290074288845, + 0.014510356821119785, + 0.009427346289157867, + 0.005583590362221003, + 0.01839279942214489, + 0.015577161684632301, + 0.01373672392219305, + 0.015562053769826889, + 0.016239117830991745, + 0.01876288838684559, + 0.010458871722221375, + 0.01188582181930542, + 0.02160102315247059, + 0.012849238701164722, + 0.032039545476436615, + 0.013524512760341167, + 0.009897821582853794, + 0.014281030744314194, + 0.01361802313476801, + 0.009093225933611393, + 0.026385005563497543, + 0.02655905857682228, + 0.024425752460956573, + 0.012266493402421474, + 0.024858849123120308, + 0.035310808569192886, + 0.01315394602715969, + 0.015705831348896027, + 0.0053417072631418705, + 0.008113289251923561, + 0.008102637715637684, + 0.00610598037019372, + 0.00695756683126092, + 0.005316233262419701, + 0.016524475067853928, + 0.007308457978069782, + 0.009141558781266212, + 0.011195928789675236, + 0.008260127156972885, + 0.006713866721838713, + 0.012513143010437489, + 0.012145690619945526, + 0.00912465713918209, + 0.006946762092411518, + 0.006630921736359596, + 0.007471715100109577, + 0.0120368842035532, + 0.008085381239652634, + 0.006746417842805386, + 0.016485923901200294, + 0.024416614323854446, + 0.005029493477195501, + 0.013682118616998196, + 0.0048364391550421715, + 0.0045039779506623745, + 0.005742303095757961, + 0.023744042962789536, + 0.0051409099251031876, + 0.00536359753459692, + 0.006723214406520128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.20154821872711182 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_8/index_tensor([3340, 3340])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_8/index_tensor([3340, 3340])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a831a2f216384f878947c934481c19af690c7933 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_8/index_tensor([3340, 3340])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angle wide-angle shot captures a high-definition time-lapse of a dense swarm of Uca tullbergi crabs on a vast expanse of damp, dark sand, the entire scene rendered with the stark, authentic quality of photojournalism. As the timelapse progresses, the tide recedes away from the shore through a glassy, smooth sheet of sand, revealing a shimmering, rippled sand texture. In accelerated motion, the swarm transforms into a frenzied hive of activity; individual crabs, defined with incredible clarity, dash past one another. The males' large, vibrant orange-red claws contrast strikingly with their smaller feeding claws and dark, speckled carapaces, which they rhythmically wave in what appears to be territorial displays. Some crabs burrow purposefully into the damp sand, while others surface, their collective behavior captured as a dynamic living system. The scene is illuminated by bright, diffuse daylight from a cloudy sky, casting soft, minimal shadows and highlighting the moisture on every surface, creating a mesmerizing and scientifically accurate spectacle of natural life in motion.", + "seed": 13173, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_8/index_3340_seed_13173_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.037871900992468, + "vlm_logprob": -0.037871900992468, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.037871900992468, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.169661745429039 + ], + "opposite_logprob_tokens": [ + -1.9196617603302002 + ], + "logprob_reward": -0.169661745429039, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.169661745429039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06441247463226318, + 0.0643373653292656, + 0.04570712521672249, + 0.09314120560884476, + 0.048731882125139236, + 0.23348261415958405, + 0.2302359789609909, + 0.08634068071842194, + 0.07611795514822006, + 0.057492680847644806 + ], + "credit_map": [ + 0.01168777234852314, + 0.006264415103942156, + 0.008269088342785835, + 0.006754498463124037, + 0.01004375983029604, + 0.00804726593196392, + 0.006839083507657051, + 0.006506594829261303, + 0.008021358400583267, + 0.006187263410538435, + 0.00823812372982502, + 0.0068281227722764015, + 0.012665074318647385, + 0.005570663139224052, + 0.00978805124759674, + 0.0070387087762355804, + 0.005247525405138731, + 0.005088760517537594, + 0.00417672423645854, + 0.005015926901251078, + 0.007841476239264011, + 0.005927129648625851, + 0.005505367182195187, + 0.006904216017574072, + 0.013001732528209686, + 0.012796556577086449, + 0.013245015405118465, + 0.00881616398692131, + 0.010797849856317043, + 0.014448055997490883, + 0.009726598858833313, + 0.010309232398867607, + 0.005620680749416351, + 0.0066239614970982075, + 0.004668065346777439, + 0.00394578417763114, + 0.0076438868418335915, + 0.0060424585826694965, + 0.004113928414881229, + 0.01007311325520277, + 0.022205734625458717, + 0.022714480757713318, + 0.02366599626839161, + 0.019752059131860733, + 0.05452446639537811, + 0.030811013653874397, + 0.023134620860219002, + 0.036674246191978455, + 0.025700846686959267, + 0.01714772917330265, + 0.045217130333185196, + 0.018495801836252213, + 0.015648797154426575, + 0.06419377774000168, + 0.01683792844414711, + 0.02699395827949047, + 0.009388566948473454, + 0.009211001917719841, + 0.010766084305942059, + 0.008155117742717266, + 0.012080655433237553, + 0.013827258720993996, + 0.0114597762003541, + 0.011452224105596542, + 0.010446364991366863, + 0.006142013240605593, + 0.009843997657299042, + 0.0062969340942800045, + 0.015128659084439278, + 0.010007811710238457, + 0.007095432840287685, + 0.01115674339234829, + 0.005548423156142235, + 0.004960058256983757, + 0.009749185293912888, + 0.00704984413459897, + 0.00745401531457901, + 0.008852245286107063, + 0.00661626597866416, + 0.007262647617608309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.169661745429039 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076424358412623405 + ], + "opposite_logprob_tokens": [ + -4.8826422691345215 + ], + "logprob_reward": -0.0076424358412623405, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076424358412623405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05020149052143097, + 0.08897394686937332, + 0.08135158568620682, + 0.06846802681684494, + 0.03471746668219566, + 0.1282486468553543, + 0.23604770004749298, + 0.051684729754924774, + 0.0984705239534378, + 0.16183587908744812 + ], + "credit_map": [ + 0.009138811379671097, + 0.004902255721390247, + 0.005842057056725025, + 0.00449109124019742, + 0.007645449135452509, + 0.008089226670563221, + 0.005001055542379618, + 0.005091541912406683, + 0.01245205383747816, + 0.010911227203905582, + 0.008662172593176365, + 0.009165194816887379, + 0.01916605979204178, + 0.008278929628431797, + 0.01000167615711689, + 0.010336636565625668, + 0.010091147385537624, + 0.011361537501215935, + 0.00813109241425991, + 0.009432632476091385, + 0.008656062185764313, + 0.008512256667017937, + 0.012533823028206825, + 0.012633027508854866, + 0.007358070928603411, + 0.008698255755007267, + 0.006727862171828747, + 0.0060350545682013035, + 0.008950456976890564, + 0.014836630783975124, + 0.0067079924046993256, + 0.0091536995023489, + 0.004771261475980282, + 0.004672972019761801, + 0.003284063655883074, + 0.0027149657253175974, + 0.006491526495665312, + 0.0035062762908637524, + 0.003949383273720741, + 0.005327013786882162, + 0.012579815462231636, + 0.014820635318756104, + 0.007519007194787264, + 0.008419753052294254, + 0.039417561143636703, + 0.017340300604701042, + 0.013261319138109684, + 0.014890259131789207, + 0.039805177599191666, + 0.017560364678502083, + 0.04693475365638733, + 0.015564504079520702, + 0.01653757132589817, + 0.058053065091371536, + 0.015028296038508415, + 0.026563987135887146, + 0.006159808021038771, + 0.004892885684967041, + 0.01032256055623293, + 0.004969101864844561, + 0.007558566518127918, + 0.006537648383527994, + 0.005248155444860458, + 0.005996005143970251, + 0.008746185339987278, + 0.010440671816468239, + 0.013857893645763397, + 0.008707952685654163, + 0.01846846379339695, + 0.014554704539477825, + 0.011987252160906792, + 0.011707399040460587, + 0.01578919216990471, + 0.026829948648810387, + 0.02355758287012577, + 0.014232530258595943, + 0.018010858446359634, + 0.03389890119433403, + 0.009834756143391132, + 0.019682101905345917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.0076424358412623405 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003623231779783964 + ], + "opposite_logprob_tokens": [ + -5.628623008728027 + ], + "logprob_reward": -0.003623231779783964, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003623231779783964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12823767960071564, + 0.12390744686126709, + 0.07398726791143417, + 0.043030235916376114, + 0.07447783648967743, + 0.18797685205936432, + 0.08622799813747406, + 0.09790433198213577, + 0.054533977061510086, + 0.12971638143062592 + ], + "credit_map": [ + 0.01929888129234314, + 0.011575177311897278, + 0.017545588314533234, + 0.016614630818367004, + 0.019230641424655914, + 0.014811119064688683, + 0.01188570074737072, + 0.017275938764214516, + 0.017018061131238937, + 0.012393965385854244, + 0.015153168700635433, + 0.014466091990470886, + 0.023823808878660202, + 0.011596780270338058, + 0.016260499134659767, + 0.013195068575441837, + 0.00785908754914999, + 0.008991627022624016, + 0.0063004461117088795, + 0.008179588243365288, + 0.015812108293175697, + 0.008780339732766151, + 0.010790050961077213, + 0.007274019531905651, + 0.005057689268141985, + 0.0049084932543337345, + 0.004524245858192444, + 0.0032375387381762266, + 0.006002798210829496, + 0.00900641642510891, + 0.003953987266868353, + 0.006339069455862045, + 0.006846137344837189, + 0.009335493668913841, + 0.011399352923035622, + 0.009992902167141438, + 0.00541040301322937, + 0.005161621607840061, + 0.01055397279560566, + 0.015777956694364548, + 0.020296256989240646, + 0.02066003903746605, + 0.011048957705497742, + 0.012505143880844116, + 0.05881863087415695, + 0.025221483781933784, + 0.020038073882460594, + 0.01938825286924839, + 0.013140990398824215, + 0.006861328613013029, + 0.01661141961812973, + 0.00833315309137106, + 0.009110582061111927, + 0.014371595345437527, + 0.007621540687978268, + 0.010177389718592167, + 0.007679958362132311, + 0.008685239590704441, + 0.014046791940927505, + 0.015573336742818356, + 0.009273450821638107, + 0.01577882096171379, + 0.013194090686738491, + 0.013672647066414356, + 0.00846865028142929, + 0.005803218111395836, + 0.009879863820970058, + 0.0066691511310637, + 0.005307012237608433, + 0.008285438641905785, + 0.0042674061842262745, + 0.005853240843862295, + 0.014047391712665558, + 0.01964198611676693, + 0.016969595104455948, + 0.0122281638905406, + 0.01719106361269951, + 0.02611052617430687, + 0.008414712734520435, + 0.015112940222024918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.003623231779783964 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00248754327185452 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.00248754327185452, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00248754327185452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06026021018624306, + 0.04293282330036163, + 0.03457779437303543, + 0.0514477975666523, + 0.03637503460049629, + 0.3857983946800232, + 0.1773419827222824, + 0.06614410877227783, + 0.0637955591082573, + 0.08132627606391907 + ], + "credit_map": [ + 0.009430276229977608, + 0.005287518724799156, + 0.008088592439889908, + 0.006386156193912029, + 0.010722015984356403, + 0.00812312588095665, + 0.005841846112161875, + 0.006380678620189428, + 0.004549604374915361, + 0.005042181815952063, + 0.005058976821601391, + 0.00643779244273901, + 0.004377273377031088, + 0.00666497927159071, + 0.005139325279742479, + 0.005662690382450819, + 0.0037663725670427084, + 0.004001196473836899, + 0.002769626211374998, + 0.002917035948485136, + 0.008364452049136162, + 0.0046160402707755566, + 0.003957115113735199, + 0.004185952711850405, + 0.006706038489937782, + 0.007226199842989445, + 0.003937138710170984, + 0.002953625749796629, + 0.0076157161965966225, + 0.012454317882657051, + 0.00417181383818388, + 0.006382944528013468, + 0.004442587494850159, + 0.004700429271906614, + 0.004166176076978445, + 0.0034427782520651817, + 0.005995959974825382, + 0.004573823884129524, + 0.004212554544210434, + 0.004840725567191839, + 0.03950793296098709, + 0.045121464878320694, + 0.025902725756168365, + 0.026943204924464226, + 0.11026298999786377, + 0.04787837341427803, + 0.039484113454818726, + 0.050697553902864456, + 0.025538213551044464, + 0.012904457747936249, + 0.036241836845874786, + 0.012049256823956966, + 0.012617708183825016, + 0.04617542773485184, + 0.011445185169577599, + 0.02036990039050579, + 0.007668892852962017, + 0.005592683330178261, + 0.007688458543270826, + 0.008391603827476501, + 0.008285765536129475, + 0.009360672906041145, + 0.010800219140946865, + 0.008355813100934029, + 0.008444499224424362, + 0.006136449985206127, + 0.01170753501355648, + 0.005927470047026873, + 0.009897865355014801, + 0.00816440861672163, + 0.005806328263133764, + 0.007711002603173256, + 0.008214402012526989, + 0.009874683804810047, + 0.011700012721121311, + 0.008550306782126427, + 0.00955015979707241, + 0.01333712600171566, + 0.009353933855891228, + 0.01074565201997757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.00248754327185452 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944548640400171 + ], + "opposite_logprob_tokens": [ + -5.130944728851318 + ], + "logprob_reward": -0.005944548640400171, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944548640400171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09301834553480148, + 0.07861723750829697, + 0.05781375989317894, + 0.07435256987810135, + 0.07059668749570847, + 0.07576701045036316, + 0.21140064299106598, + 0.10029380768537521, + 0.11837506294250488, + 0.11976487189531326 + ], + "credit_map": [ + 0.014915961772203445, + 0.008358723483979702, + 0.008581227622926235, + 0.007594206370413303, + 0.020428430289030075, + 0.014585480093955994, + 0.008699060417711735, + 0.009855260141193867, + 0.009363136254251003, + 0.006198791787028313, + 0.007333414163440466, + 0.009363784454762936, + 0.01938638836145401, + 0.009387608617544174, + 0.010700046084821224, + 0.006884063594043255, + 0.007326357997953892, + 0.00616162084043026, + 0.004235077183693647, + 0.005850133020430803, + 0.014565933495759964, + 0.008237226866185665, + 0.006751669570803642, + 0.004685741849243641, + 0.008844837546348572, + 0.008548430167138577, + 0.005772354546934366, + 0.005148359574377537, + 0.011804415844380856, + 0.019777635112404823, + 0.005797985475510359, + 0.008658554404973984, + 0.005320829339325428, + 0.008786922320723534, + 0.0107565401121974, + 0.011007736437022686, + 0.008063644170761108, + 0.009259098209440708, + 0.007353613153100014, + 0.010048302821815014, + 0.007090259809046984, + 0.0067449696362018585, + 0.006310705561190844, + 0.0075352126732468605, + 0.010499120689928532, + 0.016955990344285965, + 0.008462055586278439, + 0.012168696150183678, + 0.041587237268686295, + 0.016424795612692833, + 0.03845927491784096, + 0.014384797774255276, + 0.018035341054201126, + 0.044800229370594025, + 0.013744181953370571, + 0.02396479994058609, + 0.009833195246756077, + 0.007934563793241978, + 0.01300411019474268, + 0.008996078744530678, + 0.011065847240388393, + 0.02407185174524784, + 0.011260442435741425, + 0.014127720147371292, + 0.016001632437109947, + 0.012120016850531101, + 0.017996154725551605, + 0.00938281137496233, + 0.02074529603123665, + 0.014167703688144684, + 0.011011198163032532, + 0.016950247809290886, + 0.01260322891175747, + 0.014762712642550468, + 0.021745765581727028, + 0.009811941534280777, + 0.01681915670633316, + 0.01636037416756153, + 0.008751762099564075, + 0.018909931182861328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.005944548640400171 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle wide-angle shot captures a high-definition time-lapse of a dense swarm of Uca tullbergi crabs on a vast expanse of damp, dark sand, the entire scene rendered with the stark, authentic quality of photojournalism. As the timelapse progresses, the tide recedes away from the shore through a glassy, smooth sheet of sand, revealing a shimmering, rippled sand texture. In accelerated motion, the swarm transforms into a frenzied hive of activity; individual crabs, defined with incredible clarity, dash past one another. The males' large, vibrant orange-red claws contrast strikingly with their smaller feeding claws and dark, speckled carapaces, which they rhythmically wave in what appears to be territorial displays. Some crabs burrow purposefully into the damp sand, while others surface, their collective behavior captured as a dynamic living system. The scene is illuminated by bright, diffuse daylight from a cloudy sky, casting soft, minimal shadows and highlighting the moisture on every surface, creating a mesmerizing and scientifically accurate spectacle of natural life in motion.", + "seed": 13174, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_8/index_3340_seed_13174_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06207101419568062, + "vlm_logprob": -0.06207101419568062, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.06207101419568062, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2668237090110779 + ], + "opposite_logprob_tokens": [ + -1.5168237686157227 + ], + "logprob_reward": -0.2668237090110779, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2668237090110779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2186080664396286, + 0.052690539509058, + 0.07245177030563354, + 0.041165027767419815, + 0.09887532889842987, + 0.21663783490657806, + 0.11311715096235275, + 0.09970399737358093, + 0.055201370269060135, + 0.031548913568258286 + ], + "credit_map": [ + 0.04514473304152489, + 0.0313839353621006, + 0.04824930429458618, + 0.01910722441971302, + 0.02080472558736801, + 0.020870037376880646, + 0.015542035922408104, + 0.017506061121821404, + 0.0047907703556120396, + 0.006082133390009403, + 0.00627366965636611, + 0.005765099544078112, + 0.008493204601109028, + 0.007042869459837675, + 0.006361049599945545, + 0.00788174569606781, + 0.010736261494457722, + 0.015664134174585342, + 0.01257933210581541, + 0.008822874166071415, + 0.007567256689071655, + 0.005025898106396198, + 0.0063037630170583725, + 0.005752256140112877, + 0.004647396970540285, + 0.004373625852167606, + 0.0034759060945361853, + 0.004512549843639135, + 0.005900785326957703, + 0.005192093551158905, + 0.008495423011481762, + 0.004567247349768877, + 0.013598187826573849, + 0.011274074204266071, + 0.00938414316624403, + 0.01577949710190296, + 0.014515041373670101, + 0.01154926884919405, + 0.011269022710621357, + 0.011506100185215473, + 0.031377360224723816, + 0.051891714334487915, + 0.017935460433363914, + 0.021211715415120125, + 0.04068709537386894, + 0.02308293804526329, + 0.016678795218467712, + 0.013772773556411266, + 0.010493598878383636, + 0.017140299081802368, + 0.005666373763233423, + 0.010872925631701946, + 0.01631210558116436, + 0.027749178931117058, + 0.007823650725185871, + 0.017059026286005974, + 0.013171411119401455, + 0.012274750508368015, + 0.008820275776088238, + 0.015737343579530716, + 0.0127873495221138, + 0.015295916236937046, + 0.009183687157928944, + 0.012433269061148167, + 0.00704534025862813, + 0.007452177815139294, + 0.004336277488619089, + 0.006863010581582785, + 0.0066908267326653, + 0.01063980907201767, + 0.005603166297078133, + 0.006570762954652309, + 0.004777822643518448, + 0.006741348188370466, + 0.0025197986979037523, + 0.004100413993000984, + 0.004043009597808123, + 0.0040022144094109535, + 0.0026041255332529545, + 0.00276018469594419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.2668237090110779 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014214691706001759 + ], + "opposite_logprob_tokens": [ + -4.264214515686035 + ], + "logprob_reward": -0.014214691706001759, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014214691706001759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17417263984680176, + 0.07538481056690216, + 0.058153796941041946, + 0.25711843371391296, + 0.06686770170927048, + 0.03586025536060333, + 0.08458076417446136, + 0.06434182077646255, + 0.15943552553653717, + 0.024084266275167465 + ], + "credit_map": [ + 0.023478949442505836, + 0.026549899950623512, + 0.015040133148431778, + 0.01836179941892624, + 0.02718687616288662, + 0.023777134716510773, + 0.016902783885598183, + 0.022875064983963966, + 0.006975763477385044, + 0.010121479630470276, + 0.009389356710016727, + 0.007457074709236622, + 0.011386641301214695, + 0.013186573050916195, + 0.00874021090567112, + 0.008127707988023758, + 0.0073647573590278625, + 0.011637032963335514, + 0.006135103292763233, + 0.005009799730032682, + 0.008595316670835018, + 0.007883623242378235, + 0.005554531700909138, + 0.005973626859486103, + 0.02471032738685608, + 0.028177380561828613, + 0.014552817679941654, + 0.014921378344297409, + 0.08149110525846481, + 0.016171889379620552, + 0.061620354652404785, + 0.015473154373466969, + 0.010024317540228367, + 0.014807862229645252, + 0.006024216767400503, + 0.006828809157013893, + 0.01099066250026226, + 0.006854813080281019, + 0.005796876270323992, + 0.005540135316550732, + 0.005016983952373266, + 0.0053657107055187225, + 0.0030656824819743633, + 0.004118955694139004, + 0.004318066406995058, + 0.0053561534732580185, + 0.004441233351826668, + 0.004177469294518232, + 0.014268036931753159, + 0.007512357085943222, + 0.00686879176646471, + 0.011743457056581974, + 0.010859226807951927, + 0.014837422408163548, + 0.008132471702992916, + 0.010358990170061588, + 0.007894710637629032, + 0.010721084661781788, + 0.003419271670281887, + 0.0076575991697609425, + 0.008085405454039574, + 0.013927814550697803, + 0.004890819080173969, + 0.007745112292468548, + 0.02934797666966915, + 0.02078493870794773, + 0.011416206136345863, + 0.014294320717453957, + 0.026962192729115486, + 0.02512899786233902, + 0.012168982066214085, + 0.01933191530406475, + 0.003056472400203347, + 0.0026518250815570354, + 0.0018500869628041983, + 0.0017925178399309516, + 0.00407964875921607, + 0.0037856833077967167, + 0.0023700259625911713, + 0.0044980039820075035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.014214691706001759 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589664578437805 + ], + "opposite_logprob_tokens": [ + -3.895589590072632 + ], + "logprob_reward": -0.020589664578437805, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589664578437805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2409375011920929, + 0.0698671042919159, + 0.10809997469186783, + 0.1650332659482956, + 0.1193074956536293, + 0.09577101469039917, + 0.0428619310259819, + 0.061678431928157806, + 0.06208014488220215, + 0.03436317667365074 + ], + "credit_map": [ + 0.03857431188225746, + 0.028719132766127586, + 0.05403890088200569, + 0.023447584360837936, + 0.025235440582036972, + 0.02626638300716877, + 0.0246921144425869, + 0.019963614642620087, + 0.007199202198535204, + 0.008754412643611431, + 0.007218561135232449, + 0.007174879778176546, + 0.013015590608119965, + 0.009745405986905098, + 0.006255219224840403, + 0.010503838770091534, + 0.01527299266308546, + 0.026530424132943153, + 0.00990370474755764, + 0.009285089559853077, + 0.016365336254239082, + 0.01610381156206131, + 0.006945604458451271, + 0.007693012244999409, + 0.01538827084004879, + 0.018240856006741524, + 0.00922801997512579, + 0.010190351866185665, + 0.052772507071495056, + 0.009081761352717876, + 0.04033863916993141, + 0.009792856872081757, + 0.021072331815958023, + 0.02138570323586464, + 0.0065127182751894, + 0.017415883019566536, + 0.019434407353401184, + 0.014810914173722267, + 0.009470547549426556, + 0.009204995818436146, + 0.013299601152539253, + 0.017735235393047333, + 0.009635763242840767, + 0.012159524485468864, + 0.013569675385951996, + 0.012303054332733154, + 0.009061460383236408, + 0.008006701245903969, + 0.006002991925925016, + 0.004849439486861229, + 0.004294399172067642, + 0.006395081989467144, + 0.006264389026910067, + 0.006084170658141375, + 0.005285791587084532, + 0.0036856657825410366, + 0.008597793988883495, + 0.006396956741809845, + 0.004630988463759422, + 0.013446219265460968, + 0.007111981976777315, + 0.007023697718977928, + 0.007175336591899395, + 0.007295461371541023, + 0.009549958631396294, + 0.010379387997090816, + 0.0037460816092789173, + 0.006393208168447018, + 0.00937727838754654, + 0.009118967689573765, + 0.004977100063115358, + 0.00853816233575344, + 0.004676029086112976, + 0.00513193104416132, + 0.0026618100237101316, + 0.006842074450105429, + 0.004017110448330641, + 0.00409512547776103, + 0.00399883883073926, + 0.0029402568470686674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.020589664578437805 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093363881111145 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.004093363881111145, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093363881111145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2096935361623764, + 0.10695421695709229, + 0.06712314486503601, + 0.053660910576581955, + 0.07213710993528366, + 0.14086513221263885, + 0.061285197734832764, + 0.15552476048469543, + 0.062460340559482574, + 0.07029560953378677 + ], + "credit_map": [ + 0.029465211555361748, + 0.041686780750751495, + 0.026214174926280975, + 0.023131098598241806, + 0.02393556572496891, + 0.02019377239048481, + 0.02250572107732296, + 0.02256120927631855, + 0.00724527845159173, + 0.012744051404297352, + 0.015926769003272057, + 0.008500836789608002, + 0.019039863720536232, + 0.017972826957702637, + 0.011510341428220272, + 0.014014258980751038, + 0.009638544172048569, + 0.01388419046998024, + 0.008054465986788273, + 0.006685573607683182, + 0.00846419483423233, + 0.0073962765745818615, + 0.006398920435458422, + 0.006600986234843731, + 0.0063793957233428955, + 0.006223044823855162, + 0.004912299569696188, + 0.004494404420256615, + 0.016280807554721832, + 0.004952318500727415, + 0.006373691838234663, + 0.004044949542731047, + 0.011424320749938488, + 0.00810379907488823, + 0.008339394815266132, + 0.009631091728806496, + 0.012544733472168446, + 0.008131386712193489, + 0.006937656085938215, + 0.007024733815342188, + 0.01730930618941784, + 0.02019670605659485, + 0.010878401808440685, + 0.01807462051510811, + 0.03217121213674545, + 0.015770161524415016, + 0.01560790091753006, + 0.01085683424025774, + 0.007818582467734814, + 0.00795542262494564, + 0.006140352226793766, + 0.009438987821340561, + 0.009684809483587742, + 0.008327680639922619, + 0.005795210599899292, + 0.006124160252511501, + 0.019489098340272903, + 0.017771435901522636, + 0.015424120239913464, + 0.02814614772796631, + 0.020103326067328453, + 0.01713874377310276, + 0.016623372212052345, + 0.020828528329730034, + 0.009360174648463726, + 0.007949195802211761, + 0.004750202409923077, + 0.0058887857012450695, + 0.009602334350347519, + 0.012610917910933495, + 0.005335797090083361, + 0.00696293730288744, + 0.00974676851183176, + 0.010909060016274452, + 0.007332431618124247, + 0.010282373987138271, + 0.010077942162752151, + 0.007240859791636467, + 0.007016871590167284, + 0.007689310237765312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.004093363881111145 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633641801774502 + ], + "opposite_logprob_tokens": [ + -5.37963342666626 + ], + "logprob_reward": -0.004633641801774502, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633641801774502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2059045284986496, + 0.0698358491063118, + 0.05895494669675827, + 0.32641807198524475, + 0.06637411564588547, + 0.04237421229481697, + 0.06106419861316681, + 0.04093175753951073, + 0.09742609411478043, + 0.03071625903248787 + ], + "credit_map": [ + 0.03610571101307869, + 0.03018730878829956, + 0.022375283762812614, + 0.017457248643040657, + 0.02887823060154915, + 0.030848076567053795, + 0.019152291119098663, + 0.020900385454297066, + 0.005662768147885799, + 0.01002881396561861, + 0.0103355273604393, + 0.008086513727903366, + 0.009475440718233585, + 0.011012461967766285, + 0.007697719614952803, + 0.007536603603512049, + 0.006408086512237787, + 0.011003553867340088, + 0.00450062146410346, + 0.00501397205516696, + 0.008333805948495865, + 0.011130057275295258, + 0.006155597046017647, + 0.00640924833714962, + 0.03238185867667198, + 0.0304242093116045, + 0.012434474192559719, + 0.016694994643330574, + 0.12111011892557144, + 0.016345929354429245, + 0.08048911392688751, + 0.016537362709641457, + 0.008499742485582829, + 0.007533361203968525, + 0.004671154543757439, + 0.006912359967827797, + 0.010299820452928543, + 0.015980003401637077, + 0.004519493784755468, + 0.007958182133734226, + 0.00738360732793808, + 0.008800710551440716, + 0.003442120738327503, + 0.00437263585627079, + 0.0068449294194579124, + 0.005217087455093861, + 0.0032607903704047203, + 0.0030523291788995266, + 0.005541135556995869, + 0.010751709342002869, + 0.0029319210443645716, + 0.006009870208799839, + 0.008816331624984741, + 0.014176269993185997, + 0.00448351027444005, + 0.008353451266884804, + 0.005035292357206345, + 0.005478646606206894, + 0.004501857329159975, + 0.0049840775318443775, + 0.005779695697128773, + 0.005668396595865488, + 0.004964884836226702, + 0.00451890891417861, + 0.010469588451087475, + 0.0146729601547122, + 0.004780485760420561, + 0.006022369954735041, + 0.014881855808198452, + 0.026482470333576202, + 0.006499791983515024, + 0.013616567477583885, + 0.0038685365580022335, + 0.0059605613350868225, + 0.002918688580393791, + 0.004428171087056398, + 0.004298292566090822, + 0.0038876489270478487, + 0.002717647235840559, + 0.0026367150712758303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.004633641801774502 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle wide-angle shot captures a high-definition time-lapse of a dense swarm of Uca tullbergi crabs on a vast expanse of damp, dark sand, the entire scene rendered with the stark, authentic quality of photojournalism. As the timelapse progresses, the tide recedes away from the shore through a glassy, smooth sheet of sand, revealing a shimmering, rippled sand texture. In accelerated motion, the swarm transforms into a frenzied hive of activity; individual crabs, defined with incredible clarity, dash past one another. The males' large, vibrant orange-red claws contrast strikingly with their smaller feeding claws and dark, speckled carapaces, which they rhythmically wave in what appears to be territorial displays. Some crabs burrow purposefully into the damp sand, while others surface, their collective behavior captured as a dynamic living system. The scene is illuminated by bright, diffuse daylight from a cloudy sky, casting soft, minimal shadows and highlighting the moisture on every surface, creating a mesmerizing and scientifically accurate spectacle of natural life in motion.", + "seed": 13175, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_8/index_3340_seed_13175_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04565947325900197, + "vlm_logprob": -0.04565947325900197, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.04565947325900197, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18509750068187714 + ], + "opposite_logprob_tokens": [ + -1.9350974559783936 + ], + "logprob_reward": -0.18509750068187714, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.18509750068187714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20966804027557373, + 0.08396916836500168, + 0.07009277492761612, + 0.14104066789150238, + 0.1581701636314392, + 0.051403746008872986, + 0.09727061539888382, + 0.031894512474536896, + 0.10287416726350784, + 0.05361616611480713 + ], + "credit_map": [ + 0.015158798545598984, + 0.02094038762152195, + 0.01712186634540558, + 0.011713597923517227, + 0.028865564614534378, + 0.04629790410399437, + 0.04838767647743225, + 0.02118222787976265, + 0.005961138289421797, + 0.014854047447443008, + 0.010614108294248581, + 0.007162554655224085, + 0.01620386727154255, + 0.012243897654116154, + 0.010673039592802525, + 0.006256510503590107, + 0.007821157574653625, + 0.00595251377671957, + 0.00812090840190649, + 0.010222764685750008, + 0.007619992829859257, + 0.013740240596234798, + 0.008668600581586361, + 0.007946589030325413, + 0.021984804421663284, + 0.020825736224651337, + 0.019421719014644623, + 0.015465120784938335, + 0.015255270525813103, + 0.01884281449019909, + 0.015543755143880844, + 0.013701451942324638, + 0.015310675837099552, + 0.013821432366967201, + 0.022981813177466393, + 0.018719898536801338, + 0.02975747175514698, + 0.01778928004205227, + 0.018797263503074646, + 0.02099231630563736, + 0.006611264310777187, + 0.01051327958703041, + 0.00605307100340724, + 0.01051898580044508, + 0.0029071830213069916, + 0.003922691103070974, + 0.004965724889189005, + 0.005911545362323523, + 0.013494256883859634, + 0.009956877678632736, + 0.012724028900265694, + 0.010213560424745083, + 0.015661586076021194, + 0.011363583616912365, + 0.01480772253125906, + 0.009049001149833202, + 0.006298082415014505, + 0.003492060350254178, + 0.00403042184188962, + 0.004095164127647877, + 0.0027481678407639265, + 0.004148863255977631, + 0.0037975169252604246, + 0.003284233855083585, + 0.00970389973372221, + 0.017639778554439545, + 0.017082126811146736, + 0.013954670168459415, + 0.0069263833574950695, + 0.018043799325823784, + 0.01034530345350504, + 0.009178216569125652, + 0.004845495335757732, + 0.005309293046593666, + 0.007194162346422672, + 0.006938679609447718, + 0.005107617471367121, + 0.01190704945474863, + 0.007300614845007658, + 0.005013254005461931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.18509750068187714 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014221390709280968 + ], + "opposite_logprob_tokens": [ + -4.26422119140625 + ], + "logprob_reward": -0.014221390709280968, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014221390709280968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35220855474472046, + 0.03404514864087105, + 0.07640164345502853, + 0.09466006606817245, + 0.07740402966737747, + 0.03033837489783764, + 0.04450707510113716, + 0.05864425003528595, + 0.14112350344657898, + 0.09066733717918396 + ], + "credit_map": [ + 0.021511351689696312, + 0.02193414606153965, + 0.041533131152391434, + 0.02591518685221672, + 0.030872248113155365, + 0.07537473738193512, + 0.0912119448184967, + 0.04385582357645035, + 0.0033036330714821815, + 0.004267239011824131, + 0.003987340722233057, + 0.006335923448204994, + 0.004873259924352169, + 0.0036078612320125103, + 0.004174589645117521, + 0.003495302749797702, + 0.010422877967357635, + 0.010208669118583202, + 0.011465819552540779, + 0.010015683248639107, + 0.011432714760303497, + 0.004958793055266142, + 0.009189741685986519, + 0.008707341738045216, + 0.008525113575160503, + 0.014424112625420094, + 0.011873358860611916, + 0.013886954635381699, + 0.01043196301907301, + 0.015526880510151386, + 0.010450965724885464, + 0.009540724568068981, + 0.0072847651317715645, + 0.006858945358544588, + 0.011407291516661644, + 0.011559443548321724, + 0.010468744672834873, + 0.008750657550990582, + 0.009080443531274796, + 0.011993741616606712, + 0.004167486913502216, + 0.004579463507980108, + 0.004035294055938721, + 0.005562543869018555, + 0.0021676684264093637, + 0.0033741118386387825, + 0.003188616130501032, + 0.0032631875947117805, + 0.0064557818695902824, + 0.004424773622304201, + 0.006562817841768265, + 0.005044382065534592, + 0.005565744359046221, + 0.0051761530339717865, + 0.0066954223439097404, + 0.0045820013619959354, + 0.006587529554963112, + 0.008714060299098492, + 0.010532999411225319, + 0.006398527417331934, + 0.0049466402269899845, + 0.007345106452703476, + 0.008675294928252697, + 0.005444093607366085, + 0.015588474459946156, + 0.017669780179858208, + 0.032509516924619675, + 0.027304260060191154, + 0.00719870813190937, + 0.019189925864338875, + 0.009149692952632904, + 0.01251314952969551, + 0.010294741950929165, + 0.010992263443768024, + 0.01304832473397255, + 0.01577775366604328, + 0.006885576527565718, + 0.01630132831633091, + 0.007555295247584581, + 0.009812054224312305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.014221390709280968 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016093002632260323 + ], + "opposite_logprob_tokens": [ + -4.141092777252197 + ], + "logprob_reward": -0.016093002632260323, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016093002632260323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.194065123796463, + 0.06995932757854462, + 0.12049154937267303, + 0.1494784653186798, + 0.09279162436723709, + 0.08816642314195633, + 0.04454172030091286, + 0.08860478550195694, + 0.04578477889299393, + 0.10611619800329208 + ], + "credit_map": [ + 0.021196894347667694, + 0.02995024248957634, + 0.017690490931272507, + 0.016899708658456802, + 0.029693350195884705, + 0.024707579985260963, + 0.023722905665636063, + 0.030203955247998238, + 0.005637053400278091, + 0.005279243923723698, + 0.009359034709632397, + 0.007103754673153162, + 0.014927101321518421, + 0.011197117157280445, + 0.01153406873345375, + 0.004921952728182077, + 0.011661567725241184, + 0.01011629868298769, + 0.011213215999305248, + 0.011672047898173332, + 0.033851344138383865, + 0.013551012612879276, + 0.015518865548074245, + 0.012907203286886215, + 0.013314003124833107, + 0.013495436869561672, + 0.025920668616890907, + 0.021852822974324226, + 0.015889011323451996, + 0.024456992745399475, + 0.019734464585781097, + 0.014815077185630798, + 0.007090211380273104, + 0.01020676176995039, + 0.0134353032335639, + 0.013403091579675674, + 0.009757946245372295, + 0.008744203485548496, + 0.012055260129272938, + 0.018098846077919006, + 0.008513291366398335, + 0.015095042064785957, + 0.010799983516335487, + 0.014689913019537926, + 0.006785489618778229, + 0.012499132193624973, + 0.00902754720300436, + 0.010756016708910465, + 0.007369563449174166, + 0.005163867492228746, + 0.006959735881537199, + 0.007234499324113131, + 0.004369470290839672, + 0.003953604958951473, + 0.005731245037168264, + 0.0037597320042550564, + 0.009224346838891506, + 0.010692544281482697, + 0.01668948493897915, + 0.01566370390355587, + 0.00690297270193696, + 0.007517077028751373, + 0.012528995051980019, + 0.009385662153363228, + 0.004817752167582512, + 0.005244560074061155, + 0.00818619690835476, + 0.008145446889102459, + 0.0028078770264983177, + 0.00817539356648922, + 0.004200880415737629, + 0.004206674173474312, + 0.00882166251540184, + 0.009679421782493591, + 0.015826430171728134, + 0.02096884325146675, + 0.009044991806149483, + 0.019887004047632217, + 0.013585217297077179, + 0.008302632719278336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.016093002632260323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525687774643302 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.00525687774643302, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525687774643302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16635757684707642, + 0.060649093240499496, + 0.16386909782886505, + 0.1397281438112259, + 0.2004719078540802, + 0.027113212272524834, + 0.04496088996529579, + 0.08262389898300171, + 0.04869401082396507, + 0.06553211063146591 + ], + "credit_map": [ + 0.012308596633374691, + 0.021137570962309837, + 0.014507808722555637, + 0.011506734415888786, + 0.02226065658032894, + 0.03608785942196846, + 0.0318826362490654, + 0.01666571944952011, + 0.004576283507049084, + 0.007177273277193308, + 0.0056620147079229355, + 0.009583882987499237, + 0.008263491094112396, + 0.013347459025681019, + 0.00647800462320447, + 0.005560689605772495, + 0.014405877329409122, + 0.030218511819839478, + 0.017639735713601112, + 0.01911180093884468, + 0.019829535856842995, + 0.029431533068418503, + 0.018197348341345787, + 0.01503476407378912, + 0.022442610934376717, + 0.016142500564455986, + 0.015745628625154495, + 0.01750035770237446, + 0.015245684422552586, + 0.021544013172388077, + 0.017183534801006317, + 0.013923822902143002, + 0.019713303074240685, + 0.021474363282322884, + 0.03258911520242691, + 0.02835744246840477, + 0.03233126923441887, + 0.023414410650730133, + 0.02187119796872139, + 0.0207208339124918, + 0.002111242851242423, + 0.00445773359388113, + 0.0032432787120342255, + 0.003431455697864294, + 0.0024621153715997934, + 0.004278494045138359, + 0.002725159050896764, + 0.0044037350453436375, + 0.006614937912672758, + 0.004619319923222065, + 0.006453193258494139, + 0.005143008194863796, + 0.0041090352460742, + 0.005647789221256971, + 0.008136162534356117, + 0.004237444140017033, + 0.009348745457828045, + 0.008948945440351963, + 0.010981119237840176, + 0.012297161854803562, + 0.007359130308032036, + 0.013685263693332672, + 0.012449031695723534, + 0.007554513867944479, + 0.004127497784793377, + 0.010228374041616917, + 0.007203024346381426, + 0.004621591884642839, + 0.003577570430934429, + 0.010634495876729488, + 0.004124130588024855, + 0.004177328199148178, + 0.006616024300456047, + 0.007918297313153744, + 0.010183746926486492, + 0.008763114921748638, + 0.007286215201020241, + 0.010147029533982277, + 0.008553994819521904, + 0.006063694600015879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.00525687774643302 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628594525158405 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628594525158405, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628594525158405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2602590024471283, + 0.13060691952705383, + 0.07416503876447678, + 0.09434725344181061, + 0.07062385231256485, + 0.061728090047836304, + 0.08993613719940186, + 0.08594534546136856, + 0.031711723655462265, + 0.10067659616470337 + ], + "credit_map": [ + 0.01584564335644245, + 0.012674213387072086, + 0.017320239916443825, + 0.018809683620929718, + 0.01910969614982605, + 0.05630926787853241, + 0.08501411974430084, + 0.03517615422606468, + 0.010000048205256462, + 0.0295916348695755, + 0.01726045273244381, + 0.014226295053958893, + 0.018654044717550278, + 0.012502391822636127, + 0.01857147365808487, + 0.009800584986805916, + 0.005057245492935181, + 0.011295843869447708, + 0.007435350213199854, + 0.010082200169563293, + 0.006690046284347773, + 0.019605353474617004, + 0.007528048474341631, + 0.00647096149623394, + 0.008760966360569, + 0.011792050674557686, + 0.01739802211523056, + 0.013213022612035275, + 0.009986776858568192, + 0.014747938141226768, + 0.009923077188432217, + 0.008525415323674679, + 0.005442501045763493, + 0.00633940240368247, + 0.010241926647722721, + 0.010323242284357548, + 0.007769386749714613, + 0.007745281793177128, + 0.009255294688045979, + 0.013506822288036346, + 0.007424513343721628, + 0.012332526966929436, + 0.00784109067171812, + 0.012104969471693039, + 0.003527818014845252, + 0.007330324966460466, + 0.005334584042429924, + 0.005832266993820667, + 0.011448580771684647, + 0.008955448865890503, + 0.010961713269352913, + 0.008423619903624058, + 0.016991229727864265, + 0.010438958182930946, + 0.014117563143372536, + 0.008599033579230309, + 0.013722943142056465, + 0.007138206157833338, + 0.01510771457105875, + 0.020077653229236603, + 0.004467765800654888, + 0.008645654655992985, + 0.006998891942203045, + 0.009786522015929222, + 0.002760030562058091, + 0.004275176208466291, + 0.00570157403126359, + 0.006173064932227135, + 0.001700728083960712, + 0.00408409908413887, + 0.003102092770859599, + 0.003914959728717804, + 0.009051435627043247, + 0.010509653948247433, + 0.01548775564879179, + 0.012876369059085846, + 0.007701904512941837, + 0.02722945623099804, + 0.010007566772401333, + 0.007812468335032463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.007628594525158405 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle wide-angle shot captures a high-definition time-lapse of a dense swarm of Uca tullbergi crabs on a vast expanse of damp, dark sand, the entire scene rendered with the stark, authentic quality of photojournalism. As the timelapse progresses, the tide recedes away from the shore through a glassy, smooth sheet of sand, revealing a shimmering, rippled sand texture. In accelerated motion, the swarm transforms into a frenzied hive of activity; individual crabs, defined with incredible clarity, dash past one another. The males' large, vibrant orange-red claws contrast strikingly with their smaller feeding claws and dark, speckled carapaces, which they rhythmically wave in what appears to be territorial displays. Some crabs burrow purposefully into the damp sand, while others surface, their collective behavior captured as a dynamic living system. The scene is illuminated by bright, diffuse daylight from a cloudy sky, casting soft, minimal shadows and highlighting the moisture on every surface, creating a mesmerizing and scientifically accurate spectacle of natural life in motion.", + "seed": 13176, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_8/index_3340_seed_13176_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.018535970896482467, + "vlm_logprob": -0.018535970896482467, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.018535970896482467, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07448309659957886 + ], + "opposite_logprob_tokens": [ + -2.8244831562042236 + ], + "logprob_reward": -0.07448309659957886, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07448309659957886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20538605749607086, + 0.11390026658773422, + 0.04774932563304901, + 0.058532506227493286, + 0.05238281190395355, + 0.13274343311786652, + 0.10130524635314941, + 0.07255339622497559, + 0.16421958804130554, + 0.05122731998562813 + ], + "credit_map": [ + 0.027601271867752075, + 0.03455927222967148, + 0.023936225101351738, + 0.022872524335980415, + 0.023808367550373077, + 0.03335501626133919, + 0.024227922782301903, + 0.015025459229946136, + 0.011821999214589596, + 0.013159670867025852, + 0.015256977640092373, + 0.014615330845117569, + 0.017380384728312492, + 0.01649361476302147, + 0.014377309009432793, + 0.01079497579485178, + 0.005998312961310148, + 0.006548840552568436, + 0.006081132218241692, + 0.0093816127628088, + 0.006478447001427412, + 0.004346604924649, + 0.004987343214452267, + 0.0039270296692848206, + 0.008672761730849743, + 0.007150223013013601, + 0.006482121534645557, + 0.006863247603178024, + 0.006835157051682472, + 0.007923987694084644, + 0.008076089434325695, + 0.006528919562697411, + 0.006095919758081436, + 0.006894908845424652, + 0.005535047966986895, + 0.006955471355468035, + 0.0056575993075966835, + 0.006105273962020874, + 0.008109597489237785, + 0.007028994616121054, + 0.01658519357442856, + 0.010530193336308002, + 0.020135341212153435, + 0.019821112975478172, + 0.013579840771853924, + 0.01750120520591736, + 0.01347887422889471, + 0.021111683920025826, + 0.015382492914795876, + 0.007799164392054081, + 0.012161538936197758, + 0.011017387732863426, + 0.014644726179540157, + 0.012691802345216274, + 0.012776868417859077, + 0.01483126450330019, + 0.006283754017204046, + 0.006061422638595104, + 0.013707232661545277, + 0.008402609266340733, + 0.010875900276005268, + 0.009925466030836105, + 0.006530288606882095, + 0.01076672226190567, + 0.014351851306855679, + 0.016361083835363388, + 0.018551042303442955, + 0.01443553902208805, + 0.02080591395497322, + 0.018073085695505142, + 0.03335751220583916, + 0.028283564373850822, + 0.004599152598530054, + 0.006053768564015627, + 0.0038963283877819777, + 0.003972562029957771, + 0.012028217315673828, + 0.011158212088048458, + 0.004598564468324184, + 0.004920513369143009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.07448309659957886 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004650965798646212 + ], + "opposite_logprob_tokens": [ + -5.379651069641113 + ], + "logprob_reward": -0.004650965798646212, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004650965798646212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20083779096603394, + 0.1056782603263855, + 0.13662667572498322, + 0.11236650496721268, + 0.03812310844659805, + 0.08663146942853928, + 0.05194883421063423, + 0.10258553177118301, + 0.11987747251987457, + 0.04532432183623314 + ], + "credit_map": [ + 0.022171389311552048, + 0.024792127311229706, + 0.027432866394519806, + 0.03436322882771492, + 0.022192899137735367, + 0.02740967832505703, + 0.02472454495728016, + 0.01775105483829975, + 0.013127213343977928, + 0.013244331814348698, + 0.015145879238843918, + 0.012656684033572674, + 0.0173653494566679, + 0.019012857228517532, + 0.007426368538290262, + 0.007699570618569851, + 0.021766817197203636, + 0.02483677864074707, + 0.0134831378236413, + 0.019763851538300514, + 0.01968809776008129, + 0.018357066437602043, + 0.009264330379664898, + 0.009466592222452164, + 0.0292349923402071, + 0.028705934062600136, + 0.00896469410508871, + 0.006696260068565607, + 0.007954498752951622, + 0.018068954348564148, + 0.007006907369941473, + 0.005734260194003582, + 0.004067591391503811, + 0.006147442851215601, + 0.003343599382787943, + 0.004463449120521545, + 0.0037573829758912325, + 0.006169221829622984, + 0.004604469984769821, + 0.00556994928047061, + 0.011584089137613773, + 0.009367885999381542, + 0.01607549376785755, + 0.011830659583210945, + 0.01015486940741539, + 0.00835440494120121, + 0.008432051166892052, + 0.010832017287611961, + 0.008347741328179836, + 0.0034457489382475615, + 0.008944934234023094, + 0.006645862944424152, + 0.004830923397094011, + 0.009209588170051575, + 0.003987227100878954, + 0.0065368046052753925, + 0.009260145016014576, + 0.0086821299046278, + 0.0177981685847044, + 0.012585769407451153, + 0.014341287314891815, + 0.013174930587410927, + 0.01005347166210413, + 0.01668962836265564, + 0.013521386310458183, + 0.014288282953202724, + 0.013753876090049744, + 0.010994713753461838, + 0.016420338302850723, + 0.014254642650485039, + 0.012743677012622356, + 0.023900561034679413, + 0.0038827802054584026, + 0.004728234838694334, + 0.004602271597832441, + 0.004193462431430817, + 0.01059598196297884, + 0.009444443508982658, + 0.00329449400305748, + 0.00458265095949173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.004650965798646212 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742583122104406 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.006742583122104406, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742583122104406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16392782330513, + 0.03758591040968895, + 0.11243109405040741, + 0.1346137821674347, + 0.027918551117181778, + 0.08826900273561478, + 0.24407508969306946, + 0.04929601028561592, + 0.10309959948062897, + 0.03878318890929222 + ], + "credit_map": [ + 0.01744786649942398, + 0.016376599669456482, + 0.01977638714015484, + 0.01714525744318962, + 0.020640602335333824, + 0.024045510217547417, + 0.01583322323858738, + 0.032662369310855865, + 0.004244119394570589, + 0.004972331691533327, + 0.005898064933717251, + 0.004859366454184055, + 0.0047973510809242725, + 0.00489777373149991, + 0.0038026412948966026, + 0.004114262759685516, + 0.01593446545302868, + 0.017040463164448738, + 0.010614144615828991, + 0.013679679483175278, + 0.015426469966769218, + 0.019414115697145462, + 0.009526348672807217, + 0.010795407928526402, + 0.03783604875206947, + 0.030503079295158386, + 0.009694138541817665, + 0.00882251188158989, + 0.013122289441525936, + 0.016267403960227966, + 0.011247574351727962, + 0.007120735011994839, + 0.002886589616537094, + 0.003676819382235408, + 0.0025069487746804953, + 0.00319301662966609, + 0.003379741683602333, + 0.004729770123958588, + 0.0030215915758162737, + 0.004524070303887129, + 0.010356739163398743, + 0.010196926072239876, + 0.008487989194691181, + 0.012543308548629284, + 0.014785682782530785, + 0.012181228958070278, + 0.008645780384540558, + 0.011071350425481796, + 0.032616518437862396, + 0.021620137616991997, + 0.030147574841976166, + 0.0316084586083889, + 0.03334169089794159, + 0.03416331484913826, + 0.026615386828780174, + 0.03396201133728027, + 0.004379965830594301, + 0.007002091500908136, + 0.008212766610085964, + 0.006412489805370569, + 0.007427050266414881, + 0.005624656565487385, + 0.00531940720975399, + 0.0049175843596458435, + 0.011095097288489342, + 0.014019283466041088, + 0.010504703037440777, + 0.010073085315525532, + 0.009449857287108898, + 0.010571847669780254, + 0.015846600756049156, + 0.021539123728871346, + 0.005005208775401115, + 0.005989275872707367, + 0.005929343868046999, + 0.005648639518767595, + 0.005347096361219883, + 0.0037698475643992424, + 0.0031395158730447292, + 0.003954258281737566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.006742583122104406 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319126108661294 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.00319126108661294, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319126108661294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1756903976202011, + 0.03965781256556511, + 0.04927593842148781, + 0.14175096154212952, + 0.040161967277526855, + 0.17241275310516357, + 0.2223622053861618, + 0.045759160071611404, + 0.06540346890687943, + 0.04752533510327339 + ], + "credit_map": [ + 0.01827993057668209, + 0.020218897610902786, + 0.031315386295318604, + 0.021143728867173195, + 0.025013387203216553, + 0.024013692513108253, + 0.020173601806163788, + 0.015531772747635841, + 0.005187343806028366, + 0.006152716930955648, + 0.004843037575483322, + 0.004747454542666674, + 0.005764476023614407, + 0.0051726000383496284, + 0.00449092173948884, + 0.003299260511994362, + 0.006402317900210619, + 0.008138979785144329, + 0.006517565343528986, + 0.005922568961977959, + 0.006286627147346735, + 0.005301953759044409, + 0.005733718164265156, + 0.004972212016582489, + 0.047870706766843796, + 0.042989056557416916, + 0.008869370445609093, + 0.00901669543236494, + 0.008691198192536831, + 0.009672832675278187, + 0.008723944425582886, + 0.0059171500615775585, + 0.004898327402770519, + 0.007519952021539211, + 0.0034171706065535545, + 0.004492994397878647, + 0.007507396396249533, + 0.003728647483512759, + 0.00390268093906343, + 0.004694798961281776, + 0.017165109515190125, + 0.018778767436742783, + 0.021101703867316246, + 0.023146143183112144, + 0.024852236732840538, + 0.02525928243994713, + 0.01896688900887966, + 0.023142626509070396, + 0.03812389820814133, + 0.015590516850352287, + 0.02493133209645748, + 0.02380094863474369, + 0.031037691980600357, + 0.028659725561738014, + 0.026657868176698685, + 0.033560216426849365, + 0.0035640329588204622, + 0.004933780990540981, + 0.006982681807130575, + 0.005849660839885473, + 0.009100625291466713, + 0.005797109100967646, + 0.004543252754956484, + 0.004988017026335001, + 0.008161296136677265, + 0.008612622506916523, + 0.0078035094775259495, + 0.005988136399537325, + 0.0076688299886882305, + 0.006870375946164131, + 0.011552453972399235, + 0.008746244944632053, + 0.004350168164819479, + 0.006506911478936672, + 0.006306875962764025, + 0.004890862386673689, + 0.008958058431744576, + 0.006898405496031046, + 0.003758166916668415, + 0.005855885334312916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.00319126108661294 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a high-definition time-lapse of a dense swarm of Uca tullbergi crabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of damp, dark sand with a glassy, smooth sheet receding to reveal rippled textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the male crabs shown with large, vibrant orange-red claws and dark, speckled carapaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crabs exhibit frenzied activity including dashing past one another, rhythmic claw waving, and purposeful burrowing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, slightly high-angle wide-angle view illuminated by bright, diffuse daylight from a cloudy sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003611947875469923 + ], + "opposite_logprob_tokens": [ + -5.628612041473389 + ], + "logprob_reward": -0.003611947875469923, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003611947875469923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12419600784778595, + 0.07663150131702423, + 0.09404435008764267, + 0.08748640865087509, + 0.1445336788892746, + 0.15234212577342987, + 0.20606712996959686, + 0.03450770303606987, + 0.05765623226761818, + 0.022534796968102455 + ], + "credit_map": [ + 0.013252146542072296, + 0.015624109655618668, + 0.02245270647108555, + 0.017141439020633698, + 0.013793077319860458, + 0.014233062975108624, + 0.01599753461778164, + 0.011701934970915318, + 0.00958964042365551, + 0.008445922285318375, + 0.010844415053725243, + 0.007554220035672188, + 0.009843181818723679, + 0.01682097278535366, + 0.007946435362100601, + 0.005586713552474976, + 0.013200744055211544, + 0.016890401020646095, + 0.011305385269224644, + 0.011749917641282082, + 0.012687692418694496, + 0.011884665116667747, + 0.008709125220775604, + 0.0076164137572050095, + 0.017118064686655998, + 0.017657939344644547, + 0.005674033425748348, + 0.008199268952012062, + 0.011177388951182365, + 0.01156015507876873, + 0.009711628779768944, + 0.006387932226061821, + 0.01669253595173359, + 0.025487324222922325, + 0.011126545257866383, + 0.01911001466214657, + 0.02001054771244526, + 0.018742632120847702, + 0.01339681725949049, + 0.01996726170182228, + 0.013740611262619495, + 0.014225896447896957, + 0.021537480875849724, + 0.021742260083556175, + 0.02245931327342987, + 0.018431318923830986, + 0.015090829692780972, + 0.02511441335082054, + 0.034169286489486694, + 0.014006508514285088, + 0.02515411004424095, + 0.02437291294336319, + 0.027801670134067535, + 0.028384128585457802, + 0.02190338261425495, + 0.0302751362323761, + 0.0028927975799888372, + 0.003329881001263857, + 0.006673631723970175, + 0.0034136567264795303, + 0.004716905299574137, + 0.004723913036286831, + 0.0030413635540753603, + 0.00571555458009243, + 0.006875663064420223, + 0.006213233340531588, + 0.008408554829657078, + 0.005644543096423149, + 0.007506839465349913, + 0.0066609978675842285, + 0.009629691019654274, + 0.006716712843626738, + 0.0023646277841180563, + 0.003389309858903289, + 0.0030024279840290546, + 0.00240364414639771, + 0.003576949005946517, + 0.002779500326141715, + 0.001925254357047379, + 0.0030930838547647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.003611947875469923 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_9/index_tensor([4548, 4548])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_9/index_tensor([4548, 4548])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..251f1366aa2447a3d1eff0725b054302f43e48d2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_9/index_tensor([4548, 4548])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up shot, shot from a slightly overhead angle, closely tracks the subtle movements of a boy's head. He is about seven or eight years old, with short, shaggy dark brown hair, a few strands playfully falling across his forehead. He wears a soft, light blue and white striped long-sleeved T-shirt and is absorbed in his work. He sits before a crystal-clear glass guardrail supported by matte silver aluminum columns. Beyond the guardrail, a gentle green hill and cotton-candy clouds appear. As he ponders, the camera steadily follows his head as it tilts slightly to the left. His bright, deep eyes glance upward, revealing the concentration characteristic of deep thought. His right hand, gripping a fine-tipped black fountain pen, briefly hovers a few millimeters above the drawing paper. Then, his gaze refocuses on the paper, and with steady, fluid movements of his wrist, he traces a precise line. The entire world is bathed in a bright and evenly diffused light, presenting a picture book style with fresh color matching and almost no shadows. The main colors of the picture are composed of soft blue, white and green, creating an atmosphere of tranquility, purity and creativity.", + "seed": 4006, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_9/index_4548_seed_4006_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1435389019548893, + "vlm_logprob": -0.1435389019548893, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.1435389019548893, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07260355353355408 + ], + "opposite_logprob_tokens": [ + -3.572603464126587 + ], + "logprob_reward": -0.07260355353355408, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.07260355353355408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1782834678888321, + 0.05261329561471939, + 0.03630652278661728, + 0.08103092014789581, + 0.1308453232049942, + 0.13765567541122437, + 0.1079646572470665, + 0.18732479214668274, + 0.05564497038722038, + 0.03233040124177933 + ], + "credit_map": [ + 0.01998000778257847, + 0.024856453761458397, + 0.03812146931886673, + 0.014053924940526485, + 0.018012069165706635, + 0.015978189185261726, + 0.030835676938295364, + 0.016445670276880264, + 0.007546528242528439, + 0.0110921710729599, + 0.009612816385924816, + 0.005484443157911301, + 0.0044484613463282585, + 0.0039061123970896006, + 0.006160282995551825, + 0.004362478852272034, + 0.005982973147183657, + 0.006224173586815596, + 0.00409604050219059, + 0.003978127148002386, + 0.0038815317675471306, + 0.003563199657946825, + 0.004613874014467001, + 0.0039666034281253815, + 0.014178253710269928, + 0.017517542466521263, + 0.011922953650355339, + 0.010878941975533962, + 0.008321018889546394, + 0.003938478417694569, + 0.009502374567091465, + 0.004771354142576456, + 0.02187149040400982, + 0.024793215095996857, + 0.02760191075503826, + 0.017216412350535393, + 0.010571768507361412, + 0.007896626368165016, + 0.010503342375159264, + 0.010390564799308777, + 0.028843186795711517, + 0.03724246099591255, + 0.01931995525956154, + 0.015007201582193375, + 0.01187894120812416, + 0.008637557737529278, + 0.008425941690802574, + 0.008300425484776497, + 0.02032105065882206, + 0.019076885655522346, + 0.014444537460803986, + 0.009342861361801624, + 0.012600895948708057, + 0.00963283609598875, + 0.013144281692802906, + 0.009401300922036171, + 0.07256035506725311, + 0.07411855459213257, + 0.009602404199540615, + 0.005212649703025818, + 0.0075590163469314575, + 0.00519837299361825, + 0.00712219625711441, + 0.005951232276856899, + 0.00869086105376482, + 0.008407906629145145, + 0.00852365605533123, + 0.006167698185890913, + 0.0058257137425243855, + 0.006434338632971048, + 0.007392053958028555, + 0.0042027439922094345, + 0.004210151266306639, + 0.004889136645942926, + 0.003409069264307618, + 0.006196607835590839, + 0.0028227129951119423, + 0.0038367826491594315, + 0.003326658857986331, + 0.0036392821930348873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.07260355353355408 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309525713324547 + ], + "opposite_logprob_tokens": [ + -3.168095350265503 + ], + "logprob_reward": -0.04309525713324547, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309525713324547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14848272502422333, + 0.04278479143977165, + 0.015679502859711647, + 0.03209454193711281, + 0.09934662282466888, + 0.04615755379199982, + 0.02940264716744423, + 0.5093543529510498, + 0.026131074875593185, + 0.050566088408231735 + ], + "credit_map": [ + 0.01827656477689743, + 0.02316521480679512, + 0.02371601015329361, + 0.013835184276103973, + 0.016676252707839012, + 0.01269674114882946, + 0.027204912155866623, + 0.012911844067275524, + 0.006154605187475681, + 0.007028679363429546, + 0.004647317808121443, + 0.0035896017216145992, + 0.004880768246948719, + 0.004177426919341087, + 0.00783319491893053, + 0.004473199136555195, + 0.003820517798885703, + 0.0037557005416601896, + 0.0018247278640046716, + 0.0018110860837623477, + 0.0011183477472513914, + 0.0010290318168699741, + 0.0012656411854550242, + 0.0010544491233304143, + 0.00517664710059762, + 0.004563576076179743, + 0.004722030833363533, + 0.0033591934479773045, + 0.003927797079086304, + 0.002290981588885188, + 0.005272021517157555, + 0.002782293129712343, + 0.01631510816514492, + 0.021577687934041023, + 0.021433498710393906, + 0.010765236802399158, + 0.009384799748659134, + 0.006109120789915323, + 0.00713026849552989, + 0.006630901712924242, + 0.010961577296257019, + 0.012367386370897293, + 0.004888385534286499, + 0.004526961594820023, + 0.004850806202739477, + 0.0023660745937377214, + 0.003052473533898592, + 0.0031438868027180433, + 0.0038412674330174923, + 0.00559193966910243, + 0.0050163064152002335, + 0.0032989168539643288, + 0.00339129869826138, + 0.002531976904720068, + 0.003846785519272089, + 0.0018841546261683106, + 0.21927644312381744, + 0.22480656206607819, + 0.008900989778339863, + 0.008167335763573647, + 0.015173283405601978, + 0.008406265638768673, + 0.015425488352775574, + 0.009198043495416641, + 0.003938859328627586, + 0.003931031562387943, + 0.003338089445605874, + 0.0034096194431185722, + 0.0024153278209269047, + 0.002506229327991605, + 0.004119311925023794, + 0.002472606021910906, + 0.007884879596531391, + 0.009047058410942554, + 0.003850123845040798, + 0.004269757773727179, + 0.004637723322957754, + 0.007350646425038576, + 0.00609868997707963, + 0.007427210919559002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.04309525713324547 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978010892868042 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.00978010892868042, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978010892868042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12335315346717834, + 0.1569766253232956, + 0.06469590216875076, + 0.07520243525505066, + 0.09543320536613464, + 0.025095123797655106, + 0.024148203432559967, + 0.3248926103115082, + 0.04665789753198624, + 0.06354483962059021 + ], + "credit_map": [ + 0.024338770657777786, + 0.027644820511341095, + 0.01531305629760027, + 0.013400360941886902, + 0.007930965162813663, + 0.008773976936936378, + 0.00985903199762106, + 0.016092166304588318, + 0.0338074192404747, + 0.04024729132652283, + 0.027895163744688034, + 0.016499150544404984, + 0.011280085891485214, + 0.00913259293884039, + 0.011226557195186615, + 0.00688836257904768, + 0.018236685544252396, + 0.02022690698504448, + 0.007390265818685293, + 0.004136100877076387, + 0.003203919157385826, + 0.0024744991678744555, + 0.0036056742537766695, + 0.005421850364655256, + 0.011815357953310013, + 0.012347864918410778, + 0.0090063801035285, + 0.00992269441485405, + 0.008446074090898037, + 0.008183774538338184, + 0.007154443766921759, + 0.008325845934450626, + 0.015679622069001198, + 0.019434969872236252, + 0.018736183643341064, + 0.011087680235505104, + 0.00999297946691513, + 0.007001491729170084, + 0.006920537445694208, + 0.006579735316336155, + 0.0038604356814175844, + 0.0038883890956640244, + 0.0033567515201866627, + 0.003284826409071684, + 0.002816754626110196, + 0.002481719246134162, + 0.0027149696834385395, + 0.002691277302801609, + 0.0035106067080050707, + 0.0048674289137125015, + 0.0033252565190196037, + 0.0034030654933303595, + 0.00256133358925581, + 0.0019893127027899027, + 0.0026977728120982647, + 0.0017934262286871672, + 0.13815192878246307, + 0.14287908375263214, + 0.006684370804578066, + 0.0058158026076853275, + 0.009325080551207066, + 0.005147647578269243, + 0.010308624245226383, + 0.006580063607543707, + 0.012291967868804932, + 0.012853408232331276, + 0.003536681178957224, + 0.00445659551769495, + 0.003057941095903516, + 0.003075603861361742, + 0.0040723602287471294, + 0.003313336754217744, + 0.009568778797984123, + 0.012060330249369144, + 0.01357920840382576, + 0.010774003341794014, + 0.004008299671113491, + 0.0043227155692875385, + 0.005824712570756674, + 0.003406787756830454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.00978010892868042 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761486291885376 + ], + "opposite_logprob_tokens": [ + -0.8261486291885376 + ], + "logprob_reward": -0.5761486291885376, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761486291885376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22898943722248077, + 0.09016311913728714, + 0.06770595163106918, + 0.07369960099458694, + 0.21322299540042877, + 0.09195329248905182, + 0.019080093130469322, + 0.12900985777378082, + 0.06761385500431061, + 0.018561815842986107 + ], + "credit_map": [ + 0.042620159685611725, + 0.04627254605293274, + 0.023029623553156853, + 0.02437850832939148, + 0.014780147932469845, + 0.016751499846577644, + 0.024905528873205185, + 0.03625141456723213, + 0.0216030552983284, + 0.023438934236764908, + 0.014499606564640999, + 0.008174993097782135, + 0.00594064686447382, + 0.005440754350274801, + 0.006888005882501602, + 0.00417712377384305, + 0.0191761814057827, + 0.021012810990214348, + 0.00769952405244112, + 0.004124425817281008, + 0.003916537389159203, + 0.0023856391198933125, + 0.004816804081201553, + 0.0045740338973701, + 0.016427582129836082, + 0.017975887283682823, + 0.007681555114686489, + 0.008733541704714298, + 0.006645864341408014, + 0.006097240839153528, + 0.005857541225850582, + 0.004280382767319679, + 0.03813755512237549, + 0.039160504937171936, + 0.041526611894369125, + 0.0290497075766325, + 0.013359148055315018, + 0.011552241630852222, + 0.01923399791121483, + 0.021203218027949333, + 0.016712505370378494, + 0.023316720500588417, + 0.012524512596428394, + 0.012234074994921684, + 0.008508166298270226, + 0.006399110425263643, + 0.0055984016507864, + 0.006659801583737135, + 0.0024513122625648975, + 0.003398482222110033, + 0.0030181417241692543, + 0.001991559285670519, + 0.0030396992806345224, + 0.0016722817672416568, + 0.002285107970237732, + 0.001223509549163282, + 0.05294445902109146, + 0.05454002693295479, + 0.0034435580018907785, + 0.0034850232768803835, + 0.003969882614910603, + 0.0029062433168292046, + 0.004703429061919451, + 0.0030172334518283606, + 0.019039975479245186, + 0.020975226536393166, + 0.006881945300847292, + 0.0043937130831182, + 0.003998564090579748, + 0.004183715209364891, + 0.004395928233861923, + 0.003744784975424409, + 0.003658984089270234, + 0.00456998823210597, + 0.002117598196491599, + 0.0023157731629908085, + 0.001163226435892284, + 0.001737330574542284, + 0.0013996139168739319, + 0.0015993000706657767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.5761486291885376 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066960990428925 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.016066960990428925, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066960990428925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19197210669517517, + 0.08122891932725906, + 0.0492890328168869, + 0.06055840849876404, + 0.147125244140625, + 0.12333433330059052, + 0.03369668498635292, + 0.16003331542015076, + 0.12285333126783371, + 0.029908599331974983 + ], + "credit_map": [ + 0.02942122332751751, + 0.043028876185417175, + 0.04583768546581268, + 0.018919840455055237, + 0.01330201793462038, + 0.011982572264969349, + 0.013337095268070698, + 0.016142800450325012, + 0.011028535664081573, + 0.016274109482765198, + 0.01985972188413143, + 0.009334745816886425, + 0.006077777128666639, + 0.004594443831592798, + 0.008150104433298111, + 0.005909487139433622, + 0.014487644657492638, + 0.015600107610225677, + 0.005478458013385534, + 0.003285725601017475, + 0.002512517385184765, + 0.0019061657367274165, + 0.002942452672868967, + 0.0030759654473513365, + 0.009881618432700634, + 0.01030692271888256, + 0.010090155526995659, + 0.007725854404270649, + 0.0049681332893669605, + 0.004924274515360594, + 0.0064554777927696705, + 0.00620597368106246, + 0.029039371758699417, + 0.030555695295333862, + 0.028745736926794052, + 0.0145722646266222, + 0.009289616718888283, + 0.009181034751236439, + 0.012797326780855656, + 0.012944212183356285, + 0.027798376977443695, + 0.03128021955490112, + 0.014426407404243946, + 0.010930473916232586, + 0.015555262565612793, + 0.006298617459833622, + 0.010532170534133911, + 0.006512804888188839, + 0.00561794126406312, + 0.006789220497012138, + 0.004814827349036932, + 0.00396106019616127, + 0.0032901098020374775, + 0.0029966586735099554, + 0.0037231240421533585, + 0.002503745723515749, + 0.06577455997467041, + 0.06730329245328903, + 0.005458993837237358, + 0.004271955695003271, + 0.005105352960526943, + 0.0033438962418586016, + 0.005568200722336769, + 0.003207078203558922, + 0.024082789197564125, + 0.025953739881515503, + 0.015905747190117836, + 0.018430408090353012, + 0.008254675194621086, + 0.009682221338152885, + 0.011842858046293259, + 0.008700900711119175, + 0.004501480143517256, + 0.005569279193878174, + 0.005215616896748543, + 0.0037924526259303093, + 0.0017463485710322857, + 0.002534442115575075, + 0.0038552233017981052, + 0.0026937604416161776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.016066960990428925 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a slightly overhead angle, closely tracks the subtle movements of a boy's head. He is about seven or eight years old, with short, shaggy dark brown hair, a few strands playfully falling across his forehead. He wears a soft, light blue and white striped long-sleeved T-shirt and is absorbed in his work. He sits before a crystal-clear glass guardrail supported by matte silver aluminum columns. Beyond the guardrail, a gentle green hill and cotton-candy clouds appear. As he ponders, the camera steadily follows his head as it tilts slightly to the left. His bright, deep eyes glance upward, revealing the concentration characteristic of deep thought. His right hand, gripping a fine-tipped black fountain pen, briefly hovers a few millimeters above the drawing paper. Then, his gaze refocuses on the paper, and with steady, fluid movements of his wrist, he traces a precise line. The entire world is bathed in a bright and evenly diffused light, presenting a picture book style with fresh color matching and almost no shadows. The main colors of the picture are composed of soft blue, white and green, creating an atmosphere of tranquility, purity and creativity.", + "seed": 4007, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_9/index_4548_seed_4007_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11124344933778048, + "vlm_logprob": -0.11124344933778048, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.11124344933778048, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03885885700583458 + ], + "opposite_logprob_tokens": [ + -4.413858890533447 + ], + "logprob_reward": -0.03885885700583458, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.03885885700583458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17322681844234467, + 0.0656747967004776, + 0.08492089807987213, + 0.027745133265852928, + 0.10562069714069366, + 0.06568621844053268, + 0.16477851569652557, + 0.0719311460852623, + 0.09006985276937485, + 0.15034589171409607 + ], + "credit_map": [ + 0.019272975623607635, + 0.026414206251502037, + 0.024618221446871758, + 0.03568074852228165, + 0.012908478267490864, + 0.014197411015629768, + 0.022243844345211983, + 0.017890939489006996, + 0.007536827586591244, + 0.007761687505990267, + 0.006985820364207029, + 0.006371883675456047, + 0.011426455341279507, + 0.006247963290661573, + 0.009750665165483952, + 0.009593488648533821, + 0.013240430504083633, + 0.016301479190587997, + 0.007762596011161804, + 0.008997843600809574, + 0.01421290822327137, + 0.008725074119865894, + 0.008074332028627396, + 0.00760623486712575, + 0.004144317004829645, + 0.003670753911137581, + 0.0038175215013325214, + 0.004094464238733053, + 0.0033216536976397038, + 0.003278556978330016, + 0.003345176111906767, + 0.0020726893562823534, + 0.03160965070128441, + 0.026659654453396797, + 0.008841277100145817, + 0.011013383977115154, + 0.007808689959347248, + 0.004857327323406935, + 0.008884390816092491, + 0.005946322809904814, + 0.009694233536720276, + 0.009898338466882706, + 0.009943502955138683, + 0.012669460847973824, + 0.005154666490852833, + 0.004213118460029364, + 0.008866515010595322, + 0.005246382672339678, + 0.02845117077231407, + 0.01958436332643032, + 0.01579391211271286, + 0.03256101533770561, + 0.019442453980445862, + 0.015531865879893303, + 0.015579527243971825, + 0.01783420331776142, + 0.011037289164960384, + 0.0100797638297081, + 0.005085252691060305, + 0.009781096130609512, + 0.008428542874753475, + 0.008511017076671124, + 0.011148378252983093, + 0.007859811186790466, + 0.011719491332769394, + 0.008740426041185856, + 0.01129963994026184, + 0.013526753522455692, + 0.00907982885837555, + 0.010409848764538765, + 0.012035703286528587, + 0.013258161023259163, + 0.025947820395231247, + 0.029342956840991974, + 0.010144930332899094, + 0.016328759491443634, + 0.012733452022075653, + 0.03110847994685173, + 0.007743474096059799, + 0.016996018588542938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.03885885700583458 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981579303741455 + ], + "opposite_logprob_tokens": [ + -3.529815673828125 + ], + "logprob_reward": -0.02981579303741455, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981579303741455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10785378515720367, + 0.053792908787727356, + 0.16909660398960114, + 0.03774842619895935, + 0.24379727244377136, + 0.05946188047528267, + 0.09105401486158371, + 0.13131794333457947, + 0.052312616258859634, + 0.05356459319591522 + ], + "credit_map": [ + 0.017185702919960022, + 0.014980293810367584, + 0.012209917418658733, + 0.013990766368806362, + 0.016135673969984055, + 0.007357114460319281, + 0.013760260306298733, + 0.012234053574502468, + 0.005719812586903572, + 0.006499030161648989, + 0.008691300638020039, + 0.004493823274970055, + 0.007060741074383259, + 0.003644169308245182, + 0.00848206877708435, + 0.009201962500810623, + 0.01740073226392269, + 0.023743657395243645, + 0.02257738821208477, + 0.025711920112371445, + 0.02460753172636032, + 0.017169583588838577, + 0.021030772477388382, + 0.016855018213391304, + 0.005807749927043915, + 0.005871995352208614, + 0.003702040296047926, + 0.003105918178334832, + 0.005963245406746864, + 0.003327382728457451, + 0.005599479656666517, + 0.004370615351945162, + 0.0826319083571434, + 0.07538346201181412, + 0.010317486710846424, + 0.019372910261154175, + 0.014419890008866787, + 0.012124487198889256, + 0.015876883640885353, + 0.013670247048139572, + 0.010625223629176617, + 0.010708021000027657, + 0.00907953828573227, + 0.005469200666993856, + 0.004759873263537884, + 0.005255688447505236, + 0.009942442178726196, + 0.0036218930035829544, + 0.012118817307054996, + 0.013311659917235374, + 0.00821816734969616, + 0.01936155930161476, + 0.010717530734837055, + 0.008876119740307331, + 0.008501456119120121, + 0.00994870439171791, + 0.019678998738527298, + 0.024391978979110718, + 0.008409631438553333, + 0.01696697063744068, + 0.009936892427504063, + 0.01586664654314518, + 0.016026532277464867, + 0.020040296018123627, + 0.010358370840549469, + 0.004109520465135574, + 0.004810992628335953, + 0.006061263848096132, + 0.0042125992476940155, + 0.009405553340911865, + 0.0060300095938146114, + 0.007324304897338152, + 0.007895580492913723, + 0.006195695139467716, + 0.0042032515630126, + 0.01167442835867405, + 0.005330255255103111, + 0.006584618706256151, + 0.006212052423506975, + 0.005468710791319609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.02981579303741455 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635065793991089 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004635065793991089, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635065793991089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17876756191253662, + 0.08341925591230392, + 0.041025616228580475, + 0.030989311635494232, + 0.11231150478124619, + 0.09798087924718857, + 0.07880131900310516, + 0.05485193058848381, + 0.08870811760425568, + 0.23314449191093445 + ], + "credit_map": [ + 0.02588796243071556, + 0.023153923451900482, + 0.02032802253961563, + 0.027198871597647667, + 0.01629532501101494, + 0.01727648451924324, + 0.019832467660307884, + 0.02879452146589756, + 0.008829884231090546, + 0.010161105543375015, + 0.007512534502893686, + 0.012451809830963612, + 0.011057648807764053, + 0.0064606876112520695, + 0.01907350681722164, + 0.007872078567743301, + 0.005053010769188404, + 0.006293291226029396, + 0.004471681080758572, + 0.005447999108582735, + 0.005686938762664795, + 0.003407308366149664, + 0.006230960600078106, + 0.004434423521161079, + 0.004343181382864714, + 0.0037008700892329216, + 0.002635024720802903, + 0.004931825678795576, + 0.004436772782355547, + 0.002350832102820277, + 0.004827967844903469, + 0.0037628368008881807, + 0.025530682876706123, + 0.02702442929148674, + 0.011688620783388615, + 0.016042208299040794, + 0.009162183851003647, + 0.006317337043583393, + 0.008563892915844917, + 0.00798215251415968, + 0.015550538897514343, + 0.01184843573719263, + 0.020890282467007637, + 0.011845664121210575, + 0.007176374550908804, + 0.008098420687019825, + 0.012972753494977951, + 0.009598405100405216, + 0.018833762034773827, + 0.011089323088526726, + 0.006319838110357523, + 0.010892137885093689, + 0.0067454492673277855, + 0.008570300415158272, + 0.007857263088226318, + 0.008493243716657162, + 0.008103839121758938, + 0.010711072012782097, + 0.003843860002234578, + 0.006958561949431896, + 0.005104057025164366, + 0.005268443375825882, + 0.0098807318136096, + 0.004981360863894224, + 0.019110271707177162, + 0.008454114198684692, + 0.008497081696987152, + 0.011235934682190418, + 0.007240945473313332, + 0.01576608046889305, + 0.007603752892464399, + 0.01079992949962616, + 0.041229091584682465, + 0.023966223001480103, + 0.011620519682765007, + 0.05910050868988037, + 0.020706316456198692, + 0.03534553945064545, + 0.022567057982087135, + 0.018609236925840378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.004635065793991089 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742707908153534 + ], + "opposite_logprob_tokens": [ + -0.9742708206176758 + ], + "logprob_reward": -0.4742707908153534, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742707908153534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13067272305488586, + 0.067332424223423, + 0.16417665779590607, + 0.022547127678990364, + 0.12015629559755325, + 0.22348934412002563, + 0.12634249031543732, + 0.0423133447766304, + 0.05956738442182541, + 0.04340227693319321 + ], + "credit_map": [ + 0.023897485807538033, + 0.026418866589665413, + 0.01753830350935459, + 0.010102586820721626, + 0.016549818217754364, + 0.01347940880805254, + 0.008784845471382141, + 0.013901399448513985, + 0.008638177998363972, + 0.008905316703021526, + 0.017354723066091537, + 0.006206723861396313, + 0.006630559451878071, + 0.004791023209691048, + 0.008636998943984509, + 0.006168902385979891, + 0.02824869565665722, + 0.04009414464235306, + 0.021737832576036453, + 0.021295148879289627, + 0.012789261527359486, + 0.009720534086227417, + 0.014708896167576313, + 0.015582145191729069, + 0.00366564211435616, + 0.004143152851611376, + 0.0026541631668806076, + 0.0039988732896745205, + 0.0017763589276000857, + 0.0015751369064673781, + 0.0026101679541170597, + 0.0021236331667751074, + 0.041194263845682144, + 0.03599664568901062, + 0.008161338977515697, + 0.01030588336288929, + 0.005626417696475983, + 0.006148080807179213, + 0.006583560723811388, + 0.00614010076969862, + 0.03844822943210602, + 0.030173011124134064, + 0.04813616722822189, + 0.028199119493365288, + 0.01532537117600441, + 0.015199067071080208, + 0.029179446399211884, + 0.018828926607966423, + 0.02470606565475464, + 0.017635557800531387, + 0.01032488513737917, + 0.02209470607340336, + 0.012982098385691643, + 0.012275183573365211, + 0.013304641470313072, + 0.013019349426031113, + 0.007059359923005104, + 0.008860532194375992, + 0.002436933107674122, + 0.005507540889084339, + 0.004083593375980854, + 0.00593708036467433, + 0.00547723239287734, + 0.0029510725289583206, + 0.011928881518542767, + 0.005479738581925631, + 0.007843230850994587, + 0.006922914180904627, + 0.003875178284943104, + 0.009589076973497868, + 0.005264701787382364, + 0.008663664571940899, + 0.006746395956724882, + 0.006855530198663473, + 0.0028971261344850063, + 0.00752067007124424, + 0.003771851770579815, + 0.007448785472661257, + 0.0039772456511855125, + 0.004184671677649021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.4742707908153534 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636740036308765 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636740036308765, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636740036308765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15674972534179688, + 0.06551461666822433, + 0.06837380677461624, + 0.06044640392065048, + 0.1530359983444214, + 0.11176764965057373, + 0.1568402647972107, + 0.05044962838292122, + 0.09888937324285507, + 0.07793255895376205 + ], + "credit_map": [ + 0.02469942346215248, + 0.028647402301430702, + 0.020272888243198395, + 0.015878494828939438, + 0.01745905913412571, + 0.01875336281955242, + 0.015649698674678802, + 0.01538939867168665, + 0.00800031702965498, + 0.010564960539340973, + 0.014879140071570873, + 0.007078608963638544, + 0.004824921488761902, + 0.0033520341385155916, + 0.009660981595516205, + 0.007153655402362347, + 0.00850619561970234, + 0.009904920123517513, + 0.01848752051591873, + 0.007579961325973272, + 0.006593042053282261, + 0.004894773941487074, + 0.006794797256588936, + 0.005612590815871954, + 0.012338060885667801, + 0.010648973286151886, + 0.008559313602745533, + 0.006191275082528591, + 0.004009661730378866, + 0.00497613986954093, + 0.005037334281951189, + 0.008685646578669548, + 0.040405914187431335, + 0.03217853978276253, + 0.011825932189822197, + 0.021139662712812424, + 0.010373233817517757, + 0.013796565122902393, + 0.013364383950829506, + 0.009951768442988396, + 0.02513149566948414, + 0.019221020862460136, + 0.019186828285455704, + 0.014003240503370762, + 0.00660924194380641, + 0.00801150593906641, + 0.011835538782179356, + 0.007768786977976561, + 0.02144516259431839, + 0.02111210860311985, + 0.013746989890933037, + 0.04003766179084778, + 0.015342196449637413, + 0.01260478887706995, + 0.0138734495267272, + 0.018677912652492523, + 0.008304434828460217, + 0.009314564988017082, + 0.005445435177534819, + 0.005899260751903057, + 0.006037087179720402, + 0.0063269915990531445, + 0.005493844859302044, + 0.00362801062874496, + 0.017700085416436195, + 0.011380204930901527, + 0.011343328282237053, + 0.01690707728266716, + 0.008367455564439297, + 0.0120090926066041, + 0.010889104567468166, + 0.010293022729456425, + 0.013540873304009438, + 0.014271371997892857, + 0.005387906916439533, + 0.009342263452708721, + 0.005785332527011633, + 0.01557246409356594, + 0.0055010197684168816, + 0.008531327359378338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.008636740036308765 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a slightly overhead angle, closely tracks the subtle movements of a boy's head. He is about seven or eight years old, with short, shaggy dark brown hair, a few strands playfully falling across his forehead. He wears a soft, light blue and white striped long-sleeved T-shirt and is absorbed in his work. He sits before a crystal-clear glass guardrail supported by matte silver aluminum columns. Beyond the guardrail, a gentle green hill and cotton-candy clouds appear. As he ponders, the camera steadily follows his head as it tilts slightly to the left. His bright, deep eyes glance upward, revealing the concentration characteristic of deep thought. His right hand, gripping a fine-tipped black fountain pen, briefly hovers a few millimeters above the drawing paper. Then, his gaze refocuses on the paper, and with steady, fluid movements of his wrist, he traces a precise line. The entire world is bathed in a bright and evenly diffused light, presenting a picture book style with fresh color matching and almost no shadows. The main colors of the picture are composed of soft blue, white and green, creating an atmosphere of tranquility, purity and creativity.", + "seed": 4008, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_9/index_4548_seed_4008_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.054231648053973916, + "vlm_logprob": -0.054231648053973916, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.054231648053973916, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.040518756955862045 + ], + "opposite_logprob_tokens": [ + -4.290518760681152 + ], + "logprob_reward": -0.040518756955862045, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.040518756955862045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18020038306713104, + 0.15576903522014618, + 0.08472204953432083, + 0.02984698861837387, + 0.05825158208608627, + 0.11950423568487167, + 0.11207831650972366, + 0.09637632220983505, + 0.06783677637577057, + 0.09541430324316025 + ], + "credit_map": [ + 0.01736854389309883, + 0.02247685007750988, + 0.020255545154213905, + 0.025412945076823235, + 0.018872717395424843, + 0.026576295495033264, + 0.027310363948345184, + 0.0219271257519722, + 0.011089152656495571, + 0.010907980613410473, + 0.01501180324703455, + 0.014975880272686481, + 0.035307835787534714, + 0.029065003618597984, + 0.021779997274279594, + 0.017631379887461662, + 0.007172394078224897, + 0.009629128500819206, + 0.012850492261350155, + 0.01219557411968708, + 0.009029250591993332, + 0.01330548245459795, + 0.010876060463488102, + 0.009663671255111694, + 0.0024807509034872055, + 0.0032547751907259226, + 0.004976756405085325, + 0.004247738514095545, + 0.002325665205717087, + 0.003736910643056035, + 0.004665303509682417, + 0.004159089177846909, + 0.008297361433506012, + 0.005735226906836033, + 0.008494817651808262, + 0.009389616549015045, + 0.004184462130069733, + 0.007456895895302296, + 0.008459287695586681, + 0.006233917083591223, + 0.023151665925979614, + 0.010404374450445175, + 0.013592185452580452, + 0.013347139582037926, + 0.00893444661051035, + 0.013882753439247608, + 0.022776424884796143, + 0.013415252789855003, + 0.01377206388860941, + 0.014776492491364479, + 0.015893107280135155, + 0.016068777069449425, + 0.014958241023123264, + 0.013183550909161568, + 0.013591999188065529, + 0.00983408559113741, + 0.010745245963335037, + 0.006750200409442186, + 0.012075291946530342, + 0.015583180822432041, + 0.006685944739729166, + 0.012756461277604103, + 0.020356522873044014, + 0.011423466727137566, + 0.007131853140890598, + 0.005528668407350779, + 0.01384789776057005, + 0.01037429366260767, + 0.0050533185712993145, + 0.007780563551932573, + 0.01123359426856041, + 0.006886578630656004, + 0.010288964956998825, + 0.0063033015467226505, + 0.011249971576035023, + 0.015521017834544182, + 0.0075694359838962555, + 0.012874424457550049, + 0.02156335860490799, + 0.010043825022876263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.040518756955862045 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372852876782417 + ], + "opposite_logprob_tokens": [ + -3.40872859954834 + ], + "logprob_reward": -0.03372852876782417, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372852876782417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22118347883224487, + 0.18173077702522278, + 0.08663325011730194, + 0.03349066153168678, + 0.09124448895454407, + 0.15348701179027557, + 0.047761835157871246, + 0.06571543961763382, + 0.06664635986089706, + 0.05210673063993454 + ], + "credit_map": [ + 0.02215677499771118, + 0.026331786066293716, + 0.02961825765669346, + 0.03618483990430832, + 0.019888950511813164, + 0.032237812876701355, + 0.031099075451493263, + 0.023665988817811012, + 0.016850991174578667, + 0.02971014194190502, + 0.020295241847634315, + 0.029082493856549263, + 0.020750077441334724, + 0.02394586242735386, + 0.02196655608713627, + 0.019129423424601555, + 0.007486269343644381, + 0.009618858806788921, + 0.009181040339171886, + 0.009864937514066696, + 0.014770428650081158, + 0.014017663896083832, + 0.010816231369972229, + 0.010877823457121849, + 0.0034155542962253094, + 0.00489312456920743, + 0.004323089495301247, + 0.004752286709845066, + 0.0029275433626025915, + 0.004045865964144468, + 0.004492416046559811, + 0.004640782717615366, + 0.014664063230156898, + 0.008573762141168118, + 0.016796346753835678, + 0.012773440219461918, + 0.008087534457445145, + 0.012795324437320232, + 0.010721586644649506, + 0.006832432001829147, + 0.012325667776167393, + 0.008895923383533955, + 0.01945764198899269, + 0.026982886716723442, + 0.010809103026986122, + 0.017086388543248177, + 0.03969721123576164, + 0.01823219656944275, + 0.0062234066426754, + 0.0055044712498784065, + 0.005179797764867544, + 0.008541240356862545, + 0.005467433016747236, + 0.0052347988821566105, + 0.005844108760356903, + 0.005766578018665314, + 0.00868633110076189, + 0.007575674913823605, + 0.008198224008083344, + 0.011094747111201286, + 0.005564978811889887, + 0.0063324798829853535, + 0.012206877581775188, + 0.0060561238788068295, + 0.008132710121572018, + 0.007380216848105192, + 0.008703557774424553, + 0.010794112458825111, + 0.004506334196776152, + 0.009833522140979767, + 0.011726072058081627, + 0.005569832865148783, + 0.005212875083088875, + 0.005468869581818581, + 0.006498920265585184, + 0.00921314675360918, + 0.005546108353883028, + 0.007365839555859566, + 0.009008665569126606, + 0.003792309667915106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.03372852876782417 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947866942733526 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005947866942733526, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947866942733526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1216970682144165, + 0.13133065402507782, + 0.0874876007437706, + 0.0475149005651474, + 0.08302091807126999, + 0.14943940937519073, + 0.0644373670220375, + 0.13220547139644623, + 0.07212725281715393, + 0.11073940992355347 + ], + "credit_map": [ + 0.010367446579039097, + 0.013871381059288979, + 0.013280296698212624, + 0.017053648829460144, + 0.0122727882117033, + 0.018742011860013008, + 0.019609002396464348, + 0.016500499099493027, + 0.013288761489093304, + 0.010608341544866562, + 0.01849263720214367, + 0.017164113000035286, + 0.017914073541760445, + 0.02170606516301632, + 0.01678292825818062, + 0.015373723581433296, + 0.007658103946596384, + 0.00681172963231802, + 0.009203343652188778, + 0.010870889760553837, + 0.012725879438221455, + 0.0173020102083683, + 0.011049680411815643, + 0.011865961365401745, + 0.004020045977085829, + 0.003872773377224803, + 0.009754437021911144, + 0.008226471021771431, + 0.0038742152974009514, + 0.006451168563216925, + 0.006296157371252775, + 0.005019632168114185, + 0.006323548965156078, + 0.007077950518578291, + 0.017557743936777115, + 0.015819018706679344, + 0.0060086557641625404, + 0.011961404234170914, + 0.011064615100622177, + 0.007207978051155806, + 0.01726747490465641, + 0.00959018524736166, + 0.01915998011827469, + 0.02331160381436348, + 0.011197197251021862, + 0.01758536882698536, + 0.03431744500994682, + 0.017010148614645004, + 0.008122158236801624, + 0.007932320237159729, + 0.009598413482308388, + 0.010300932452082634, + 0.00893114972859621, + 0.007495001424103975, + 0.007330100983381271, + 0.004727290011942387, + 0.019309835508465767, + 0.010220251977443695, + 0.019472600892186165, + 0.020928069949150085, + 0.008837291970849037, + 0.017026280984282494, + 0.025570198893547058, + 0.010840936563909054, + 0.007449861150234938, + 0.00710612628608942, + 0.013633357360959053, + 0.013507886789739132, + 0.004825977608561516, + 0.009372798725962639, + 0.008351271972060204, + 0.00787997804582119, + 0.009586870670318604, + 0.009132354520261288, + 0.015049497596919537, + 0.02689994126558304, + 0.009131151251494884, + 0.01674959436058998, + 0.01733734831213951, + 0.006852646358311176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.005947866942733526 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988905310630798 + ], + "opposite_logprob_tokens": [ + -1.8048890829086304 + ], + "logprob_reward": -0.17988905310630798, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988905310630798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3448769748210907, + 0.1799216866493225, + 0.12839530408382416, + 0.021616274490952492, + 0.047871481627225876, + 0.08997756987810135, + 0.051086489111185074, + 0.08435756713151932, + 0.028776880353689194, + 0.023119783028960228 + ], + "credit_map": [ + 0.03041902743279934, + 0.02301931008696556, + 0.0333016999065876, + 0.030137924477458, + 0.04615028575062752, + 0.05985426902770996, + 0.07787138968706131, + 0.044123079627752304, + 0.02256093919277191, + 0.014800267294049263, + 0.02349166013300419, + 0.017993751913309097, + 0.024752110242843628, + 0.029499055817723274, + 0.02420726604759693, + 0.02261662296950817, + 0.009867645800113678, + 0.00927208922803402, + 0.017498962581157684, + 0.010981651954352856, + 0.021925581619143486, + 0.02124847285449505, + 0.01641492173075676, + 0.021185969933867455, + 0.0017187725752592087, + 0.0019075707532465458, + 0.004540459718555212, + 0.0033923916053026915, + 0.001368826488032937, + 0.0028609505388885736, + 0.0027358762454241514, + 0.0030914254020899534, + 0.007899774238467216, + 0.004911063704639673, + 0.005567715968936682, + 0.005600722506642342, + 0.0036668935790657997, + 0.006126431282609701, + 0.006880853790789843, + 0.007218025624752045, + 0.009451348334550858, + 0.005072377622127533, + 0.010754293762147427, + 0.013451650738716125, + 0.006441732402890921, + 0.0101063447073102, + 0.023774154484272003, + 0.01092566829174757, + 0.01138892862945795, + 0.0062914094887673855, + 0.005458034109324217, + 0.004995040595531464, + 0.0050218310207128525, + 0.00583316944539547, + 0.004297955427318811, + 0.007800120394676924, + 0.009320437908172607, + 0.0051611922681331635, + 0.010857975110411644, + 0.014406150206923485, + 0.0055977338925004005, + 0.008367002941668034, + 0.020678183063864708, + 0.009968888945877552, + 0.0029112037736922503, + 0.0024239360354840755, + 0.0042510549537837505, + 0.004484109114855528, + 0.002019796520471573, + 0.003445386653766036, + 0.0051427194848656654, + 0.004098673351109028, + 0.002202013274654746, + 0.0028258096426725388, + 0.003051248611882329, + 0.003291565226390958, + 0.0021453830413520336, + 0.002907566726207733, + 0.0032366712111979723, + 0.0034595243632793427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.17988905310630798 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074034497141838 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.011074034497141838, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074034497141838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26300308108329773, + 0.09516753256320953, + 0.07918647676706314, + 0.030582884326577187, + 0.0533788837492466, + 0.0777239128947258, + 0.14820073544979095, + 0.09416952729225159, + 0.072382852435112, + 0.08620402216911316 + ], + "credit_map": [ + 0.024411892518401146, + 0.025000622496008873, + 0.03387366235256195, + 0.03486204892396927, + 0.030026553198695183, + 0.03978794068098068, + 0.04791203513741493, + 0.027128344401717186, + 0.010395430959761143, + 0.01015117485076189, + 0.009738372638821602, + 0.01151119451969862, + 0.012375548481941223, + 0.014740423299372196, + 0.012910322286188602, + 0.01334507204592228, + 0.00736318901181221, + 0.008745984174311161, + 0.009798114188015461, + 0.010864517651498318, + 0.010485438629984856, + 0.012148872949182987, + 0.011407189071178436, + 0.008373172022402287, + 0.0029640570282936096, + 0.0027736714109778404, + 0.005351252853870392, + 0.003756193444132805, + 0.0021613668650388718, + 0.0046295044012367725, + 0.004985976964235306, + 0.0039608594961464405, + 0.005107240751385689, + 0.00484828045591712, + 0.009946724399924278, + 0.008273101411759853, + 0.005434306804090738, + 0.0071845632046461105, + 0.007047244347631931, + 0.005537422839552164, + 0.00800277665257454, + 0.004794639069586992, + 0.01011411752551794, + 0.012120741419494152, + 0.005853425711393356, + 0.007844985462725163, + 0.019958827644586563, + 0.009034398943185806, + 0.03429684042930603, + 0.01841963641345501, + 0.016656463965773582, + 0.01408079918473959, + 0.015018610283732414, + 0.016649020835757256, + 0.013273629359900951, + 0.019805731251835823, + 0.018031291663646698, + 0.007308955769985914, + 0.014143413864076138, + 0.01137477532029152, + 0.005833761300891638, + 0.012576933950185776, + 0.01603233814239502, + 0.008868059143424034, + 0.010977358557283878, + 0.007752060890197754, + 0.007738184183835983, + 0.010016071610152721, + 0.004308097530156374, + 0.011695296503603458, + 0.013260731473565102, + 0.006635055877268314, + 0.012062632478773594, + 0.008591501973569393, + 0.006641385145485401, + 0.011146972887217999, + 0.007378365378826857, + 0.014531007036566734, + 0.017480576410889626, + 0.008371582254767418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.011074034497141838 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a slightly overhead angle, closely tracks the subtle movements of a boy's head. He is about seven or eight years old, with short, shaggy dark brown hair, a few strands playfully falling across his forehead. He wears a soft, light blue and white striped long-sleeved T-shirt and is absorbed in his work. He sits before a crystal-clear glass guardrail supported by matte silver aluminum columns. Beyond the guardrail, a gentle green hill and cotton-candy clouds appear. As he ponders, the camera steadily follows his head as it tilts slightly to the left. His bright, deep eyes glance upward, revealing the concentration characteristic of deep thought. His right hand, gripping a fine-tipped black fountain pen, briefly hovers a few millimeters above the drawing paper. Then, his gaze refocuses on the paper, and with steady, fluid movements of his wrist, he traces a precise line. The entire world is bathed in a bright and evenly diffused light, presenting a picture book style with fresh color matching and almost no shadows. The main colors of the picture are composed of soft blue, white and green, creating an atmosphere of tranquility, purity and creativity.", + "seed": 4009, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000059/group_9/index_4548_seed_4009_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.053656240459531546, + "vlm_logprob": -0.053656240459531546, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.053656240459531546, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04148949682712555 + ], + "opposite_logprob_tokens": [ + -4.041489601135254 + ], + "logprob_reward": -0.04148949682712555, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.04148949682712555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15810492634773254, + 0.07189683616161346, + 0.05397430807352066, + 0.05061885342001915, + 0.06782511621713638, + 0.05848134309053421, + 0.06708251684904099, + 0.03331966698169708, + 0.3601302206516266, + 0.07856619358062744 + ], + "credit_map": [ + 0.02510000392794609, + 0.022439967840909958, + 0.02404825948178768, + 0.0344809927046299, + 0.01587524637579918, + 0.010786901228129864, + 0.013296418823301792, + 0.012077150866389275, + 0.009339581243693829, + 0.010408771224319935, + 0.009387558326125145, + 0.011666380800306797, + 0.00872932281345129, + 0.006085929460823536, + 0.008450270630419254, + 0.007829018868505955, + 0.010862495750188828, + 0.007163235917687416, + 0.005525327753275633, + 0.005911252927035093, + 0.006211371626704931, + 0.0058848802000284195, + 0.00644852127879858, + 0.005967221688479185, + 0.007622234057635069, + 0.005332019180059433, + 0.01272617932409048, + 0.007139561697840691, + 0.0038554624188691378, + 0.004217823501676321, + 0.005538249853998423, + 0.004187317565083504, + 0.009785127826035023, + 0.007336669601500034, + 0.00991889275610447, + 0.007475925609469414, + 0.007817610166966915, + 0.007553616538643837, + 0.010160628706216812, + 0.0077766431495547295, + 0.010216303169727325, + 0.00608109962195158, + 0.00768664013594389, + 0.009262661449611187, + 0.008953169919550419, + 0.00448697991669178, + 0.006204183679074049, + 0.005590303335338831, + 0.020874351263046265, + 0.021724872291088104, + 0.003214483615010977, + 0.004075118340551853, + 0.005366562400013208, + 0.004302130080759525, + 0.003797606099396944, + 0.0037273962516337633, + 0.004511014558374882, + 0.007863247767090797, + 0.0026961981784552336, + 0.003123246831819415, + 0.0059249624609947205, + 0.003312495304271579, + 0.0034914177376776934, + 0.0023970846086740494, + 0.09393422305583954, + 0.1314207762479782, + 0.03847406059503555, + 0.021626092493534088, + 0.03388193994760513, + 0.014829420484602451, + 0.017787201330065727, + 0.00817649532109499, + 0.014816713519394398, + 0.020251916721463203, + 0.010384080931544304, + 0.0070802574045956135, + 0.005868169944733381, + 0.007314628455787897, + 0.0076208398677408695, + 0.00522959278896451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.04148949682712555 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867033287882805 + ], + "opposite_logprob_tokens": [ + -3.0486702919006348 + ], + "logprob_reward": -0.04867033287882805, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867033287882805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19902685284614563, + 0.0388825498521328, + 0.12170606106519699, + 0.056486982852220535, + 0.13184645771980286, + 0.09571336954832077, + 0.06219646334648132, + 0.053392160683870316, + 0.0814112052321434, + 0.15933792293071747 + ], + "credit_map": [ + 0.026423579081892967, + 0.024349216371774673, + 0.02819335088133812, + 0.040356460958719254, + 0.021021774038672447, + 0.016767196357250214, + 0.022163335233926773, + 0.01975194364786148, + 0.005927916616201401, + 0.0051100775599479675, + 0.005804999731481075, + 0.004710597917437553, + 0.004926591645926237, + 0.0038586410228163004, + 0.004968930501490831, + 0.0035747941583395004, + 0.015166624449193478, + 0.010954937897622585, + 0.01164957880973816, + 0.012048861011862755, + 0.011809144169092178, + 0.02217809483408928, + 0.019171208143234253, + 0.01872762106359005, + 0.006843550130724907, + 0.006536655128002167, + 0.007885421626269817, + 0.0071863033808767796, + 0.005771550349891186, + 0.006400901824235916, + 0.00957088265568018, + 0.006291715893894434, + 0.01362538617104292, + 0.014071633107960224, + 0.020993947982788086, + 0.020336298272013664, + 0.01720728538930416, + 0.011885866522789001, + 0.021168630570173264, + 0.01255741436034441, + 0.017495255917310715, + 0.008328959345817566, + 0.013103519566357136, + 0.013733823783695698, + 0.01555267907679081, + 0.008134759031236172, + 0.011365759186446667, + 0.007998614571988583, + 0.020987557247281075, + 0.02045089192688465, + 0.0030850847251713276, + 0.0024045277386903763, + 0.004314661957323551, + 0.0043012783862650394, + 0.003344763070344925, + 0.0033076938707381487, + 0.005764428526163101, + 0.007892437279224396, + 0.005724813789129257, + 0.007617691066116095, + 0.005640047136694193, + 0.006838143337517977, + 0.008024530485272408, + 0.005890069995075464, + 0.012277486734092236, + 0.02198782190680504, + 0.012831696309149265, + 0.009680666960775852, + 0.006455284543335438, + 0.0060858153738081455, + 0.0075736939907073975, + 0.004518736153841019, + 0.015910891816020012, + 0.024767599999904633, + 0.026864251121878624, + 0.030517268925905228, + 0.00901744607836008, + 0.014258045703172684, + 0.02513289265334606, + 0.012869530357420444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.04867033287882805 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531704269349575 + ], + "opposite_logprob_tokens": [ + -4.387531757354736 + ], + "logprob_reward": -0.012531704269349575, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531704269349575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10839205980300903, + 0.05826433375477791, + 0.04968021810054779, + 0.03745514526963234, + 0.10961589962244034, + 0.06301038712263107, + 0.11528941243886948, + 0.03704824298620224, + 0.07306679338216782, + 0.3481775224208832 + ], + "credit_map": [ + 0.014378366991877556, + 0.016109000891447067, + 0.01331400591880083, + 0.017428161576390266, + 0.017221670597791672, + 0.006821307819336653, + 0.013818959705531597, + 0.00930058117955923, + 0.006622364278882742, + 0.0071791610680520535, + 0.008183563128113747, + 0.006542961113154888, + 0.00810097623616457, + 0.005529001820832491, + 0.008800160139799118, + 0.007306144572794437, + 0.012006733566522598, + 0.005395293701440096, + 0.009842200204730034, + 0.0043229456059634686, + 0.003714952850714326, + 0.004143827594816685, + 0.005292441230267286, + 0.004961824044585228, + 0.0043732901103794575, + 0.004211893770843744, + 0.005504819098860025, + 0.004827846307307482, + 0.004344608634710312, + 0.004308902658522129, + 0.005847515072673559, + 0.00403626961633563, + 0.018749237060546875, + 0.025464238598942757, + 0.010907648131251335, + 0.01249009184539318, + 0.008461058139801025, + 0.008842882700264454, + 0.01509569026529789, + 0.009605055674910545, + 0.009509868919849396, + 0.008334329351782799, + 0.006401194259524345, + 0.007015608716756105, + 0.008896427229046822, + 0.007734884507954121, + 0.007370950188487768, + 0.007747122552245855, + 0.03889286518096924, + 0.0381450280547142, + 0.005835898220539093, + 0.008204651065170765, + 0.006911190692335367, + 0.005364732816815376, + 0.007103331387042999, + 0.004831717815250158, + 0.007891804911196232, + 0.005079885013401508, + 0.004262064583599567, + 0.003286480437964201, + 0.004084407351911068, + 0.0041218590922653675, + 0.00486571341753006, + 0.0034560298081487417, + 0.009220367297530174, + 0.006752062123268843, + 0.01313056144863367, + 0.008853055536746979, + 0.006621386855840683, + 0.00939903873950243, + 0.01152944564819336, + 0.007560869678854942, + 0.04941559582948685, + 0.06670248508453369, + 0.04333686828613281, + 0.04725749045610428, + 0.023044336587190628, + 0.028387004509568214, + 0.05909046158194542, + 0.030943287536501884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "selected_reward": -0.012531704269349575 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033966839313507 + ], + "opposite_logprob_tokens": [ + -1.9103397130966187 + ], + "logprob_reward": -0.16033966839313507, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033966839313507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1885266900062561, + 0.12684953212738037, + 0.03735482692718506, + 0.0531984381377697, + 0.1384008526802063, + 0.21987837553024292, + 0.019456826150417328, + 0.062464628368616104, + 0.1066845953464508, + 0.04718523845076561 + ], + "credit_map": [ + 0.02579544670879841, + 0.03444381803274155, + 0.028460999950766563, + 0.0390484593808651, + 0.023688094690442085, + 0.007699289359152317, + 0.0193155687302351, + 0.010075019672513008, + 0.018868863582611084, + 0.018475180491805077, + 0.023565897718071938, + 0.02698427252471447, + 0.01148462388664484, + 0.007116565480828285, + 0.013260869309306145, + 0.007093245629221201, + 0.007115567568689585, + 0.0077882674522697926, + 0.004383017309010029, + 0.006511966232210398, + 0.0030267296824604273, + 0.002565825590863824, + 0.003644560230895877, + 0.0023188942577689886, + 0.01341123878955841, + 0.0106473658233881, + 0.007854855619370937, + 0.006302675697952509, + 0.004573661834001541, + 0.0038074965123087168, + 0.0037418920546770096, + 0.002859249245375395, + 0.02213106118142605, + 0.031252846121788025, + 0.018936941400170326, + 0.02909368835389614, + 0.010195774957537651, + 0.007303336635231972, + 0.011169924400746822, + 0.008317283354699612, + 0.029580140486359596, + 0.029305920004844666, + 0.025761757045984268, + 0.05503678694367409, + 0.03284333273768425, + 0.013752523809671402, + 0.021645208820700645, + 0.01195270661264658, + 0.0028123275842517614, + 0.0049806637689471245, + 0.0016727154143154621, + 0.0033046279568225145, + 0.0022502050269395113, + 0.0014044094132259488, + 0.0015183547511696815, + 0.0015135194407775998, + 0.01253097876906395, + 0.015037097968161106, + 0.009770144708454609, + 0.008106963708996773, + 0.00628556776791811, + 0.004511860199272633, + 0.0036974847316741943, + 0.0025245321448892355, + 0.02981986291706562, + 0.03765396401286125, + 0.008289760909974575, + 0.01124329213052988, + 0.007439055945724249, + 0.003655999433249235, + 0.005853009410202503, + 0.0027296545449644327, + 0.00828262884169817, + 0.00992242619395256, + 0.007127266842871904, + 0.008831643499433994, + 0.0031032778788357973, + 0.002665918553248048, + 0.005006780847907066, + 0.0022452971898019314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.16033966839313507 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a boy aged seven or eight with short, shaggy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy wearing a light blue and white striped long-sleeved T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a crystal-clear glass guardrail supported by matte silver aluminum columns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera style a picture book style with bright, evenly diffused light and almost no shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249999929219484 + ], + "opposite_logprob_tokens": [ + -5.255249977111816 + ], + "logprob_reward": -0.005249999929219484, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249999929219484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18639586865901947, + 0.05251920223236084, + 0.08813667297363281, + 0.09752357751131058, + 0.19108761847019196, + 0.22208094596862793, + 0.0352495014667511, + 0.024405663833022118, + 0.06780929118394852, + 0.03479171544313431 + ], + "credit_map": [ + 0.028307054191827774, + 0.02757626213133335, + 0.02823086641728878, + 0.045709095895290375, + 0.016262885183095932, + 0.012269856408238411, + 0.01531314104795456, + 0.012726696208119392, + 0.008559935726225376, + 0.007849994115531445, + 0.006192399654537439, + 0.0067787631414830685, + 0.006633681710809469, + 0.004529128782451153, + 0.005860981531441212, + 0.006114312447607517, + 0.017965318635106087, + 0.012238794006407261, + 0.014947479590773582, + 0.009612374939024448, + 0.009306845255196095, + 0.008277587592601776, + 0.007916415110230446, + 0.007871845737099648, + 0.01171224657446146, + 0.010683764703571796, + 0.02043432742357254, + 0.013060875236988068, + 0.008136262185871601, + 0.008470609784126282, + 0.013519251719117165, + 0.01150624267756939, + 0.03887815773487091, + 0.03733805567026138, + 0.025578459724783897, + 0.02339457906782627, + 0.015623333863914013, + 0.014876841567456722, + 0.020478645339608192, + 0.014919537119567394, + 0.03697287663817406, + 0.0209292471408844, + 0.03248274326324463, + 0.03460806608200073, + 0.03571726754307747, + 0.016094496473670006, + 0.02771873027086258, + 0.017557501792907715, + 0.010605241172015667, + 0.011237005703151226, + 0.002541221445426345, + 0.002309032017365098, + 0.0029046114068478346, + 0.0020366408862173557, + 0.0018437517574056983, + 0.0017719923052936792, + 0.003668722463771701, + 0.003358191344887018, + 0.002472895197570324, + 0.004331029020249844, + 0.0025639210361987352, + 0.0021816708613187075, + 0.003032299689948559, + 0.0027969335205852985, + 0.011195882223546505, + 0.01750359870493412, + 0.007069301325827837, + 0.011935979127883911, + 0.007207526359707117, + 0.003274427028372884, + 0.006318836938589811, + 0.0033037334214895964, + 0.0033804040867835283, + 0.003351011546328664, + 0.0059227729216217995, + 0.008836375549435616, + 0.0025669196620583534, + 0.0027220635674893856, + 0.005622797645628452, + 0.00238936860114336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.005249999929219484 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_0/index_tensor([9363, 9363])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_0/index_tensor([9363, 9363])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1226b2a20da7b39338365e15f979349ae0e0c165 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_0/index_tensor([9363, 9363])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful first-person perspective shot places the viewer in the perspective of the mythical figure Nezha, whose body transforms into a stream of red and gold light, hurtling toward the ancient Great Wall at the speed of a cannonball. At the edge of the field of vision, clenched fists can be seen, wrists adorned with the Qiankun Circle, gleaming with a dark golden glow, and the faint afterimage of a red Hun Tian Ling silken thread dancing like flames. Beneath a gloomy, dark cloud, the Great Wall ahead rapidly zooms in, its weathered gray-brown masonry, its mottled texture, and the deep green moss growing in its crevices becoming crystal clear. With a roar like the shattering of heaven and earth, the camera jerks forward and trembles violently. A destructive force erupts, shattering the vast masonry of the wall before our eyes. Countless rocks, debris, and thick dust of varying sizes scatter violently in all directions and toward the camera, blurring the entire field of view and creating a breathtaking, primal, and violent impact, an atmosphere of a grand collision of divine power.", + "seed": 11532, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_0/index_9363_seed_11532_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.008626206312328577, + "vlm_logprob": -0.008626206312328577, + "vlm_yes_no_margin": 5.325, + "vlm_token_credit": -0.008626206312328577, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01646491140127182 + ], + "opposite_logprob_tokens": [ + -4.141464710235596 + ], + "logprob_reward": -0.01646491140127182, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01646491140127182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15552127361297607, + 0.06460701674222946, + 0.04571617767214775, + 0.015697672963142395, + 0.0871182456612587, + 0.4009239077568054, + 0.024309372529387474, + 0.15603509545326233, + 0.019555969163775444, + 0.0305152740329504 + ], + "credit_map": [ + 0.017279524356126785, + 0.04909384250640869, + 0.009015468880534172, + 0.009330324828624725, + 0.012819522991776466, + 0.011590011417865753, + 0.007646181620657444, + 0.03874639794230461, + 0.011037386022508144, + 0.0060506826266646385, + 0.0064212786965072155, + 0.0058226981200277805, + 0.006399054545909166, + 0.0077065955847501755, + 0.006568501237779856, + 0.01460082083940506, + 0.008695737458765507, + 0.004815166350454092, + 0.004037079866975546, + 0.004207448102533817, + 0.0057503157295286655, + 0.005741908214986324, + 0.004667029250413179, + 0.007801493164151907, + 0.002000561449676752, + 0.002238237764686346, + 0.0017372594447806478, + 0.001748231123201549, + 0.002173469867557287, + 0.0017231600359082222, + 0.0022156955674290657, + 0.0018610581755638123, + 0.009369034320116043, + 0.009030568413436413, + 0.01558687910437584, + 0.016321301460266113, + 0.008615322411060333, + 0.007843691855669022, + 0.008760407567024231, + 0.011591030284762383, + 0.15652543306350708, + 0.15368464589118958, + 0.015757549554109573, + 0.012780381366610527, + 0.016899531707167625, + 0.01519716065376997, + 0.01156943291425705, + 0.018509753048419952, + 0.0038217781111598015, + 0.002973207039758563, + 0.0026835778262466192, + 0.004567157942801714, + 0.0018234318122267723, + 0.0017110515618696809, + 0.003498868551105261, + 0.0032303002662956715, + 0.036631885915994644, + 0.05433952435851097, + 0.012377600185573101, + 0.007996148429811, + 0.007747558876872063, + 0.008476514369249344, + 0.010550685226917267, + 0.017915185540914536, + 0.002980679040774703, + 0.0033496569376438856, + 0.002948804059997201, + 0.002401904668658972, + 0.0020381060894578695, + 0.0015917159616947174, + 0.0026504448615014553, + 0.0015946581261232495, + 0.002566241892054677, + 0.0029579743277281523, + 0.0034285876899957657, + 0.0043951645493507385, + 0.004283657297492027, + 0.005014529451727867, + 0.0039661177434027195, + 0.0039029994513839483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.01646491140127182 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018204161897301674 + ], + "opposite_logprob_tokens": [ + -4.018204212188721 + ], + "logprob_reward": -0.018204161897301674, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018204161897301674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12673455476760864, + 0.12340054661035538, + 0.04263764247298241, + 0.026248814538121223, + 0.05127215385437012, + 0.2947804033756256, + 0.07649466395378113, + 0.1416250616312027, + 0.04868360236287117, + 0.06812258809804916 + ], + "credit_map": [ + 0.01964079588651657, + 0.014989177696406841, + 0.015561079606413841, + 0.022842558100819588, + 0.013914056122303009, + 0.013628664426505566, + 0.011772640980780125, + 0.014385579153895378, + 0.013854770921170712, + 0.013920959085226059, + 0.01098148338496685, + 0.012295149266719818, + 0.014811240136623383, + 0.017687439918518066, + 0.014206936582922935, + 0.02564256824553013, + 0.00477334251627326, + 0.004291502758860588, + 0.005030663218349218, + 0.005664343014359474, + 0.00543237617239356, + 0.006826914846897125, + 0.005271511618047953, + 0.00534698273986578, + 0.0029215626418590546, + 0.003461996093392372, + 0.002022929722443223, + 0.003115263534709811, + 0.003526934189721942, + 0.0037740375846624374, + 0.0038153030909597874, + 0.003610787680372596, + 0.004878993146121502, + 0.0065801115706563, + 0.008996102958917618, + 0.008844171650707722, + 0.00572511600330472, + 0.00515954801812768, + 0.005147480871528387, + 0.005940626375377178, + 0.09073418378829956, + 0.08980400860309601, + 0.02206435054540634, + 0.015115389600396156, + 0.01760825514793396, + 0.027870722115039825, + 0.013420294970273972, + 0.018163172528147697, + 0.024967556819319725, + 0.01720287650823593, + 0.005432979203760624, + 0.005169102922081947, + 0.005950567778199911, + 0.0036891698837280273, + 0.005487194284796715, + 0.008595213294029236, + 0.03401158004999161, + 0.053036320954561234, + 0.008737782016396523, + 0.005674844607710838, + 0.006788858212530613, + 0.007109843660145998, + 0.010196803137660027, + 0.016069035977125168, + 0.007753660902380943, + 0.008858460932970047, + 0.006417338270694017, + 0.004393141716718674, + 0.003876340575516224, + 0.004590144380927086, + 0.00594634935259819, + 0.006848164834082127, + 0.007595089264214039, + 0.006509431637823582, + 0.007416312582790852, + 0.00888737104833126, + 0.008631880395114422, + 0.010014655999839306, + 0.009958433918654919, + 0.009109409525990486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.018204161897301674 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001178999780677259 + ], + "opposite_logprob_tokens": [ + -6.751179218292236 + ], + "logprob_reward": -0.001178999780677259, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001178999780677259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07854492962360382, + 0.0313895083963871, + 0.02533547393977642, + 0.020543504506349564, + 0.03837420418858528, + 0.09605272859334946, + 0.15558888018131256, + 0.19613732397556305, + 0.08492229878902435, + 0.27311113476753235 + ], + "credit_map": [ + 0.00862423051148653, + 0.01664040796458721, + 0.00754681508988142, + 0.00851503387093544, + 0.008422496728599072, + 0.008284504525363445, + 0.006416186224669218, + 0.014095249585807323, + 0.004579261876642704, + 0.0029441961087286472, + 0.0034331749193370342, + 0.002692044712603092, + 0.00315143633633852, + 0.005500434432178736, + 0.003101871581748128, + 0.005987085402011871, + 0.004565270617604256, + 0.003104597330093384, + 0.002487343968823552, + 0.0022313855588436127, + 0.0028463818598538637, + 0.0034252009354531765, + 0.0026518625672906637, + 0.004023431800305843, + 0.0023153680376708508, + 0.002923712134361267, + 0.0020072234328836203, + 0.0015277116326615214, + 0.004494574386626482, + 0.0027407510206103325, + 0.0025845717173069715, + 0.0019495922606438398, + 0.003943026997148991, + 0.00420283293351531, + 0.005744019988924265, + 0.006567134987562895, + 0.004650295712053776, + 0.0035239080898463726, + 0.004919563885778189, + 0.004823420196771622, + 0.022166291251778603, + 0.021121583878993988, + 0.008643016219139099, + 0.005792667623609304, + 0.00913332775235176, + 0.00848395749926567, + 0.008558426983654499, + 0.012153462506830692, + 0.021103940904140472, + 0.02082783542573452, + 0.01672426611185074, + 0.023346779868006706, + 0.014998995698988438, + 0.0160718634724617, + 0.0222887322306633, + 0.02022646740078926, + 0.04691851884126663, + 0.07082321494817734, + 0.013735051266849041, + 0.009056690149009228, + 0.009643161669373512, + 0.01000652089715004, + 0.016571924090385437, + 0.019382242113351822, + 0.012782510370016098, + 0.011372992768883705, + 0.011907282285392284, + 0.010081452317535877, + 0.007930023595690727, + 0.011474991217255592, + 0.008913351222872734, + 0.010459691286087036, + 0.018912160769104958, + 0.025971395894885063, + 0.019521575421094894, + 0.019490960985422134, + 0.036353107541799545, + 0.08468834310770035, + 0.022745205089449883, + 0.04542839527130127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.001178999780677259 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946800112724304 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.005946800112724304, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946800112724304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28743094205856323, + 0.0200826246291399, + 0.01663864031434059, + 0.012082860805094242, + 0.037459809333086014, + 0.17302222549915314, + 0.12335089594125748, + 0.12479078769683838, + 0.10772181302309036, + 0.0974193662405014 + ], + "credit_map": [ + 0.025952791795134544, + 0.09670978039503098, + 0.012447006069123745, + 0.01547584030777216, + 0.023304222151637077, + 0.022118985652923584, + 0.012023361399769783, + 0.0793989822268486, + 0.003829154185950756, + 0.00212879478931427, + 0.002934711752459407, + 0.002190840430557728, + 0.0017888142028823495, + 0.002474440261721611, + 0.0019877194426953793, + 0.0027481515426188707, + 0.0020925647113472223, + 0.0016595887718722224, + 0.0017293548444285989, + 0.0016515335300937295, + 0.002689225832000375, + 0.0025875589344650507, + 0.0018137673614546657, + 0.0024150460958480835, + 0.0017108980100601912, + 0.00134233757853508, + 0.002144336234778166, + 0.0010156205389648676, + 0.001789592090062797, + 0.0014351202407851815, + 0.0014081147965043783, + 0.0012368429452180862, + 0.0038867725525051355, + 0.0037864577025175095, + 0.00496251042932272, + 0.006020139902830124, + 0.0055368514731526375, + 0.004470514133572578, + 0.004891408607363701, + 0.003905158257111907, + 0.04033071547746658, + 0.03587296977639198, + 0.015410834923386574, + 0.011433136649429798, + 0.011454052291810513, + 0.024830684065818787, + 0.013715790584683418, + 0.019974051043391228, + 0.039302609860897064, + 0.02911282889544964, + 0.009425170719623566, + 0.010310119017958641, + 0.0093461899086833, + 0.005416530650109053, + 0.00884049478918314, + 0.01159696839749813, + 0.030316157266497612, + 0.044994741678237915, + 0.010244347155094147, + 0.005221164785325527, + 0.0058906422927975655, + 0.006327040493488312, + 0.009862100705504417, + 0.011934607289731503, + 0.01481964997947216, + 0.020445631816983223, + 0.014760911464691162, + 0.009137286804616451, + 0.00871333573013544, + 0.009540969505906105, + 0.014109000563621521, + 0.016195043921470642, + 0.008213815279304981, + 0.010677454061806202, + 0.008104161359369755, + 0.008610955439507961, + 0.015483418479561806, + 0.021467629820108414, + 0.010101589374244213, + 0.014760347083210945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.005946800112724304 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001336158369667828 + ], + "opposite_logprob_tokens": [ + -6.626336097717285 + ], + "logprob_reward": -0.001336158369667828, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001336158369667828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0623498298227787, + 0.008251539431512356, + 0.014259417541325092, + 0.006695201620459557, + 0.01471682544797659, + 0.23299072682857513, + 0.0608268678188324, + 0.14023788273334503, + 0.0844876691699028, + 0.3751840591430664 + ], + "credit_map": [ + 0.007229594513773918, + 0.01798868738114834, + 0.004276171792298555, + 0.0034510167315602303, + 0.005925187841057777, + 0.005395673215389252, + 0.0037383507005870342, + 0.014345148578286171, + 0.0012910861987620592, + 0.0008483494748361409, + 0.0014579340349882841, + 0.0007160588284023106, + 0.0010094258468598127, + 0.0011303565697744489, + 0.0007115237531252205, + 0.0010868044337257743, + 0.0025633745826780796, + 0.0015063148457556963, + 0.0017450173618271947, + 0.0012466342886909842, + 0.0020660152658820152, + 0.001562109449878335, + 0.0012840965064242482, + 0.00228585465811193, + 0.0008983921725302935, + 0.000830041419249028, + 0.0006089197704568505, + 0.0006522335461340845, + 0.001320753013715148, + 0.0006742299883626401, + 0.0008345961105078459, + 0.0008760361233726144, + 0.0016249017789959908, + 0.001985214650630951, + 0.002010102616623044, + 0.0022999632637947798, + 0.0016352137317880988, + 0.0014610959915444255, + 0.0019613364711403847, + 0.0017389970598742366, + 0.08665017038583755, + 0.0857900008559227, + 0.010640188120305538, + 0.00759037584066391, + 0.011024822480976582, + 0.011497359722852707, + 0.007925406098365784, + 0.011872401461005211, + 0.018291549757122993, + 0.017182935029268265, + 0.003504727967083454, + 0.003957230597734451, + 0.004197058733552694, + 0.0028821120504289865, + 0.00486081512644887, + 0.005950441583991051, + 0.030987296253442764, + 0.05064930021762848, + 0.00801630038768053, + 0.006058893166482449, + 0.006610250566154718, + 0.007381447125226259, + 0.007568993605673313, + 0.022965405136346817, + 0.013415868394076824, + 0.014224305748939514, + 0.010738243348896503, + 0.008489608764648438, + 0.006806415505707264, + 0.009690705686807632, + 0.01170202437788248, + 0.009420487098395824, + 0.027727210894227028, + 0.05338606238365173, + 0.02708456665277481, + 0.042225852608680725, + 0.07234903424978256, + 0.060796529054641724, + 0.04971184581518173, + 0.04190296307206154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.001336158369667828 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful first-person perspective shot places the viewer in the perspective of the mythical figure Nezha, whose body transforms into a stream of red and gold light, hurtling toward the ancient Great Wall at the speed of a cannonball. At the edge of the field of vision, clenched fists can be seen, wrists adorned with the Qiankun Circle, gleaming with a dark golden glow, and the faint afterimage of a red Hun Tian Ling silken thread dancing like flames. Beneath a gloomy, dark cloud, the Great Wall ahead rapidly zooms in, its weathered gray-brown masonry, its mottled texture, and the deep green moss growing in its crevices becoming crystal clear. With a roar like the shattering of heaven and earth, the camera jerks forward and trembles violently. A destructive force erupts, shattering the vast masonry of the wall before our eyes. Countless rocks, debris, and thick dust of varying sizes scatter violently in all directions and toward the camera, blurring the entire field of view and creating a breathtaking, primal, and violent impact, an atmosphere of a grand collision of divine power.", + "seed": 11533, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_0/index_9363_seed_11533_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02058295561000705, + "vlm_logprob": -0.02058295561000705, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.02058295561000705, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04362274706363678 + ], + "opposite_logprob_tokens": [ + -3.1686227321624756 + ], + "logprob_reward": -0.04362274706363678, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04362274706363678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12341272830963135, + 0.39041897654533386, + 0.07092554867267609, + 0.056542154401540756, + 0.09959837794303894, + 0.17732417583465576, + 0.03145670145750046, + 0.015572633594274521, + 0.0124007323756814, + 0.022348029538989067 + ], + "credit_map": [ + 0.014378809370100498, + 0.01676557958126068, + 0.043741375207901, + 0.01778952032327652, + 0.007409248035401106, + 0.007451904006302357, + 0.010058850049972534, + 0.005817444995045662, + 0.025216970592737198, + 0.11528202891349792, + 0.024352625012397766, + 0.029925057664513588, + 0.12968268990516663, + 0.022978054359555244, + 0.027245979756116867, + 0.015735583379864693, + 0.008069029077887535, + 0.009725762531161308, + 0.004604157991707325, + 0.006341961678117514, + 0.010253352113068104, + 0.010329247452318668, + 0.010883263312280178, + 0.010718775913119316, + 0.003773566335439682, + 0.00476152915507555, + 0.0017539967084303498, + 0.0021971033420413733, + 0.003494084347039461, + 0.007114968728274107, + 0.018068429082632065, + 0.015378475189208984, + 0.009683211334049702, + 0.005081417970359325, + 0.006250948645174503, + 0.01511616911739111, + 0.006610940210521221, + 0.011803396977484226, + 0.02633943222463131, + 0.018712855875492096, + 0.011612619273364544, + 0.008830401115119457, + 0.05891168490052223, + 0.013601052574813366, + 0.005701854825019836, + 0.01021544635295868, + 0.013569937087595463, + 0.05488117039203644, + 0.0032776426523923874, + 0.0030194977298378944, + 0.0032171164639294147, + 0.003927037585526705, + 0.0032346888910979033, + 0.0030221433844417334, + 0.007298413664102554, + 0.004460162483155727, + 0.00287304213270545, + 0.001984539907425642, + 0.0013940177159383893, + 0.0018103389302268624, + 0.0018673943122848868, + 0.001681343768723309, + 0.002495724940672517, + 0.0014662320027127862, + 0.001245509134605527, + 0.0012662421213462949, + 0.0009912241948768497, + 0.0014152511721476912, + 0.0017719040624797344, + 0.0014921639813110232, + 0.0018757970537990332, + 0.0023426408879458904, + 0.0028432465624064207, + 0.0034196132328361273, + 0.0017134000081568956, + 0.0020596084650605917, + 0.0026778646279126406, + 0.004568215925246477, + 0.0024925554171204567, + 0.0025735246017575264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.04362274706363678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197957426309586 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.018197957426309586, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197957426309586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2055612951517105, + 0.10949181020259857, + 0.11072110384702682, + 0.3307545781135559, + 0.08841517567634583, + 0.047999441623687744, + 0.015468740835785866, + 0.03167091682553291, + 0.021817397326231003, + 0.03809952363371849 + ], + "credit_map": [ + 0.025333572179079056, + 0.04094092547893524, + 0.04335308447480202, + 0.03844226524233818, + 0.0161915123462677, + 0.009916257113218307, + 0.020235557109117508, + 0.011148129589855671, + 0.016190875321626663, + 0.01595604233443737, + 0.01197145227342844, + 0.015096115879714489, + 0.01682494767010212, + 0.010043595917522907, + 0.013897382654249668, + 0.009511398151516914, + 0.010550577193498611, + 0.010736209340393543, + 0.006520548369735479, + 0.006261556409299374, + 0.01887752301990986, + 0.02651391550898552, + 0.018396001309156418, + 0.012864773161709309, + 0.012229629792273045, + 0.009465166367590427, + 0.004667304921895266, + 0.007842387072741985, + 0.01967807114124298, + 0.10293203592300415, + 0.12255771458148956, + 0.05138227343559265, + 0.00834551453590393, + 0.007098384667187929, + 0.006949898321181536, + 0.009342118166387081, + 0.012047652155160904, + 0.014694451354444027, + 0.018902050331234932, + 0.01103510707616806, + 0.003679470159113407, + 0.0040693809278309345, + 0.00722083356231451, + 0.008769044652581215, + 0.003705711802467704, + 0.00468906294554472, + 0.009678332135081291, + 0.006187602411955595, + 0.0018932173261418939, + 0.0015534708509221673, + 0.0017492955084890127, + 0.002057349309325218, + 0.0017115523805841804, + 0.0013601297978311777, + 0.002897325437515974, + 0.002246399875730276, + 0.005332239903509617, + 0.00379120628349483, + 0.0030211268458515406, + 0.0038975412026047707, + 0.003362829564139247, + 0.0032857079058885574, + 0.005557790398597717, + 0.0034224737901240587, + 0.002317191567271948, + 0.0016564727993682027, + 0.0014997124671936035, + 0.0021379468962550163, + 0.002028533024713397, + 0.003577202558517456, + 0.00566102284938097, + 0.0029393157456070185, + 0.005964948330074549, + 0.0063601890578866005, + 0.0025058421306312084, + 0.00263104191981256, + 0.004205326084047556, + 0.00844989437609911, + 0.00396820018067956, + 0.004014082718640566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.018197957426309586 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072147637605667 + ], + "opposite_logprob_tokens": [ + -4.511072158813477 + ], + "logprob_reward": -0.011072147637605667, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072147637605667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03526299074292183, + 0.05349385365843773, + 0.04162324592471123, + 0.28926458954811096, + 0.10081398487091064, + 0.3115769326686859, + 0.03927707299590111, + 0.054074883460998535, + 0.029675083234906197, + 0.04493735358119011 + ], + "credit_map": [ + 0.0035606834571808577, + 0.005153168458491564, + 0.008006933145225048, + 0.006192820146679878, + 0.003137219464406371, + 0.0024610767140984535, + 0.004669343587011099, + 0.002081746468320489, + 0.005661210045218468, + 0.006707421969622374, + 0.006008671130985022, + 0.008328879252076149, + 0.0075448863208293915, + 0.008140714839100838, + 0.006718918681144714, + 0.004383149091154337, + 0.0051660495810210705, + 0.006862131878733635, + 0.0030073956586420536, + 0.00294083496555686, + 0.007474942598491907, + 0.005087632220238447, + 0.005877915304154158, + 0.005206340923905373, + 0.00878795888274908, + 0.007071377243846655, + 0.004010151140391827, + 0.006301939953118563, + 0.01079210452735424, + 0.06005357950925827, + 0.11653845012187958, + 0.0757090151309967, + 0.010939300060272217, + 0.006467020139098167, + 0.007552075199782848, + 0.01163691096007824, + 0.01069182064384222, + 0.012277116999030113, + 0.021837959066033363, + 0.019411779940128326, + 0.012225926853716373, + 0.009451473131775856, + 0.10146965086460114, + 0.029932547360658646, + 0.008174868300557137, + 0.01935425214469433, + 0.02896335907280445, + 0.10200485587120056, + 0.004239452071487904, + 0.0034197107888758183, + 0.004848286975175142, + 0.004869707394391298, + 0.004191727843135595, + 0.0045014638453722, + 0.00785084255039692, + 0.005355883855372667, + 0.010302776470780373, + 0.007469378877431154, + 0.005374671891331673, + 0.006414916832000017, + 0.004723994992673397, + 0.004712645895779133, + 0.009344993159174919, + 0.005731505807489157, + 0.0039874291978776455, + 0.0031638063956052065, + 0.0011910389875993133, + 0.003303906647488475, + 0.0019680880941450596, + 0.0037665721029043198, + 0.008115766569972038, + 0.004178475122898817, + 0.006701591890305281, + 0.005610102321952581, + 0.0030637916643172503, + 0.0025263740681111813, + 0.004132535308599472, + 0.012064306065440178, + 0.004766084719449282, + 0.006072565913200378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.011072147637605667 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023286210373044014 + ], + "opposite_logprob_tokens": [ + -3.7732861042022705 + ], + "logprob_reward": -0.023286210373044014, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023286210373044014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06860736757516861, + 0.1030142679810524, + 0.05916271731257439, + 0.15090149641036987, + 0.12156138569116592, + 0.21389968693256378, + 0.05365042760968208, + 0.051618073135614395, + 0.12505514919757843, + 0.05252945423126221 + ], + "credit_map": [ + 0.008294648490846157, + 0.011574296280741692, + 0.009475165978074074, + 0.012257488444447517, + 0.005712651647627354, + 0.0055482215248048306, + 0.010383125394582748, + 0.0053617702797055244, + 0.004226605407893658, + 0.03469298779964447, + 0.002878679661080241, + 0.004518367815762758, + 0.03753387928009033, + 0.010528802871704102, + 0.00327887199819088, + 0.005356075242161751, + 0.0070167845115065575, + 0.007346071768552065, + 0.003991677891463041, + 0.005714217200875282, + 0.00845426693558693, + 0.008501455187797546, + 0.00991150364279747, + 0.008226734586060047, + 0.010580072179436684, + 0.007427173666656017, + 0.0033295098692178726, + 0.0032029624562710524, + 0.012001914903521538, + 0.061344828456640244, + 0.02096961997449398, + 0.032045409083366394, + 0.01452399231493473, + 0.007763682398945093, + 0.008402364328503609, + 0.015296261757612228, + 0.010576480999588966, + 0.021390389651060104, + 0.01940409280359745, + 0.024204116314649582, + 0.009517082013189793, + 0.007789974100887775, + 0.070747010409832, + 0.021558791399002075, + 0.010125255212187767, + 0.014750039204955101, + 0.024167606607079506, + 0.055243924260139465, + 0.00690798694267869, + 0.0061540910974144936, + 0.006560204084962606, + 0.006869536824524403, + 0.0049641127698123455, + 0.005981005262583494, + 0.008954500779509544, + 0.007258987985551357, + 0.007252012845128775, + 0.00801828783005476, + 0.006211049854755402, + 0.00692241545766592, + 0.005174349062144756, + 0.004717414733022451, + 0.0075890012085437775, + 0.0057335407473146915, + 0.015235347673296928, + 0.008877517655491829, + 0.005312254652380943, + 0.011916094459593296, + 0.005854096729308367, + 0.013431373052299023, + 0.04759117215871811, + 0.016837285831570625, + 0.007890870794653893, + 0.007152621168643236, + 0.00476874178275466, + 0.004781594034284353, + 0.004819121677428484, + 0.010838103480637074, + 0.005979347508400679, + 0.006299052853137255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.023286210373044014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735715549439192 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735715549439192, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735715549439192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06001961976289749, + 0.054182540625333786, + 0.08474724739789963, + 0.08493147790431976, + 0.08359230309724808, + 0.15928871929645538, + 0.06333386152982712, + 0.10223017632961273, + 0.05598797649145126, + 0.25168606638908386 + ], + "credit_map": [ + 0.006254349835216999, + 0.00936086941510439, + 0.009010322391986847, + 0.014246256090700626, + 0.004712313413619995, + 0.004358756355941296, + 0.008928592316806316, + 0.00314815784804523, + 0.004705254919826984, + 0.010040300898253918, + 0.004400100093334913, + 0.005013170652091503, + 0.008100440725684166, + 0.011036171577870846, + 0.005211084149777889, + 0.005676013417541981, + 0.008584555238485336, + 0.0103905713185668, + 0.00705596012994647, + 0.007433817256242037, + 0.014241976663470268, + 0.01412918046116829, + 0.01405932568013668, + 0.008851859718561172, + 0.006929545197635889, + 0.0057618338614702225, + 0.0030409249011427164, + 0.0034959015902131796, + 0.004766032099723816, + 0.01879969798028469, + 0.01565883681178093, + 0.026478705927729607, + 0.009100563824176788, + 0.005968624725937843, + 0.005612250417470932, + 0.010041785426437855, + 0.009038555435836315, + 0.016439422965049744, + 0.01259251032024622, + 0.014798593707382679, + 0.012668338604271412, + 0.018055159598588943, + 0.045636244118213654, + 0.01197265274822712, + 0.00827370397746563, + 0.00904329214245081, + 0.011086096055805683, + 0.04255322366952896, + 0.008444341830909252, + 0.00782860815525055, + 0.0058235228061676025, + 0.007752069737762213, + 0.005804185755550861, + 0.004559638444334269, + 0.013769207522273064, + 0.009352289140224457, + 0.013712701387703419, + 0.014358851127326488, + 0.01238047331571579, + 0.01364792138338089, + 0.009821106679737568, + 0.010584789328277111, + 0.016836734488606453, + 0.010887589305639267, + 0.007264365442097187, + 0.009550857357680798, + 0.004538119304925203, + 0.006482532247900963, + 0.007491250056773424, + 0.0061677079647779465, + 0.006753645837306976, + 0.007739501539617777, + 0.03398820757865906, + 0.04384414106607437, + 0.021672163158655167, + 0.019147997722029686, + 0.02415993995964527, + 0.049084048718214035, + 0.02674468792974949, + 0.033044904470443726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.006735715549439192 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful first-person perspective shot places the viewer in the perspective of the mythical figure Nezha, whose body transforms into a stream of red and gold light, hurtling toward the ancient Great Wall at the speed of a cannonball. At the edge of the field of vision, clenched fists can be seen, wrists adorned with the Qiankun Circle, gleaming with a dark golden glow, and the faint afterimage of a red Hun Tian Ling silken thread dancing like flames. Beneath a gloomy, dark cloud, the Great Wall ahead rapidly zooms in, its weathered gray-brown masonry, its mottled texture, and the deep green moss growing in its crevices becoming crystal clear. With a roar like the shattering of heaven and earth, the camera jerks forward and trembles violently. A destructive force erupts, shattering the vast masonry of the wall before our eyes. Countless rocks, debris, and thick dust of varying sizes scatter violently in all directions and toward the camera, blurring the entire field of view and creating a breathtaking, primal, and violent impact, an atmosphere of a grand collision of divine power.", + "seed": 11534, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_0/index_9363_seed_11534_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.011873895523604006, + "vlm_logprob": -0.011873895523604006, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.011873895523604006, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026658017188310623 + ], + "opposite_logprob_tokens": [ + -3.651658058166504 + ], + "logprob_reward": -0.026658017188310623, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026658017188310623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19334900379180908, + 0.060581307858228683, + 0.05909108743071556, + 0.1775021255016327, + 0.06758811324834824, + 0.05923277884721756, + 0.09613964706659317, + 0.1696738302707672, + 0.0685364231467247, + 0.04830564185976982 + ], + "credit_map": [ + 0.026668250560760498, + 0.017132041975855827, + 0.02294955775141716, + 0.020489180460572243, + 0.025246357545256615, + 0.028318390250205994, + 0.02414032630622387, + 0.02840489149093628, + 0.00605729129165411, + 0.005145194474607706, + 0.007240416947752237, + 0.005924064666032791, + 0.0057999491691589355, + 0.009610679000616074, + 0.008631162345409393, + 0.012172550894320011, + 0.005285114515572786, + 0.005170408170670271, + 0.006285776849836111, + 0.005603695288300514, + 0.006274724379181862, + 0.009319919161498547, + 0.010041824541985989, + 0.011109625920653343, + 0.027438687160611153, + 0.016197534278035164, + 0.0173489972949028, + 0.021975649520754814, + 0.022199099883437157, + 0.024800628423690796, + 0.02199704758822918, + 0.02554447390139103, + 0.009625725448131561, + 0.005930765066295862, + 0.00870142225176096, + 0.006934774108231068, + 0.007828820496797562, + 0.009181661531329155, + 0.009883852675557137, + 0.009501096792519093, + 0.0061011905781924725, + 0.009299539029598236, + 0.004632013849914074, + 0.008356156758964062, + 0.005875873379409313, + 0.006505114492028952, + 0.00978380348533392, + 0.008679087273776531, + 0.009463413618505001, + 0.014460653997957706, + 0.012260956689715385, + 0.014965150505304337, + 0.009658473543822765, + 0.011287297122180462, + 0.013671201653778553, + 0.010372494347393513, + 0.017263958230614662, + 0.020803697407245636, + 0.026052288711071014, + 0.02016250602900982, + 0.013933317735791206, + 0.023878730833530426, + 0.019281772896647453, + 0.028297562152147293, + 0.007005731109529734, + 0.007304968312382698, + 0.005877469666302204, + 0.0051859598606824875, + 0.013061155565083027, + 0.010402402840554714, + 0.011729816906154156, + 0.007968920283019543, + 0.005469540599733591, + 0.00639516394585371, + 0.008187346160411835, + 0.005744066089391708, + 0.005689782090485096, + 0.004785044584423304, + 0.006518928799778223, + 0.005515770521014929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.026658017188310623 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078690066933632 + ], + "opposite_logprob_tokens": [ + -4.141078472137451 + ], + "logprob_reward": -0.016078690066933632, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078690066933632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3349872827529907, + 0.09617903083562851, + 0.10703082382678986, + 0.09013631194829941, + 0.04456518217921257, + 0.05653952807188034, + 0.06943992525339127, + 0.07873250544071198, + 0.0827602967619896, + 0.03962906077504158 + ], + "credit_map": [ + 0.03845404461026192, + 0.05359087511897087, + 0.04127107188105583, + 0.031809549778699875, + 0.03478676825761795, + 0.04777532070875168, + 0.04346034675836563, + 0.04383932799100876, + 0.016028229147195816, + 0.00722983293235302, + 0.011186370626091957, + 0.00819012988358736, + 0.012052319943904877, + 0.014565067365765572, + 0.009351984597742558, + 0.017575103789567947, + 0.00947914645075798, + 0.009108648635447025, + 0.010687604546546936, + 0.010275384411215782, + 0.01034320518374443, + 0.02545766904950142, + 0.012274322099983692, + 0.019404850900173187, + 0.011842185631394386, + 0.007567455526441336, + 0.009387446567416191, + 0.009299131110310555, + 0.011063503101468086, + 0.014538516290485859, + 0.011376991868019104, + 0.015061093494296074, + 0.007503382861614227, + 0.0036364628467708826, + 0.005574130918830633, + 0.005274372640997171, + 0.0030578651931136847, + 0.007136389147490263, + 0.006082437466830015, + 0.006300149019807577, + 0.007202178705483675, + 0.009036370553076267, + 0.0049070678651332855, + 0.0055464948527514935, + 0.007922862656414509, + 0.005866229068487883, + 0.007624355144798756, + 0.008433976210653782, + 0.006780998315662146, + 0.011447395198047161, + 0.008053223602473736, + 0.007671234663575888, + 0.007586369756609201, + 0.007257914170622826, + 0.011785987764596939, + 0.00885680876672268, + 0.00842970795929432, + 0.006362056825309992, + 0.01309856679290533, + 0.00834790337830782, + 0.007376380264759064, + 0.012609431520104408, + 0.009771713055670261, + 0.012736758217215538, + 0.010237215086817741, + 0.009185838513076305, + 0.008085928857326508, + 0.007439909502863884, + 0.017005298286676407, + 0.012566816061735153, + 0.008833976462483406, + 0.009405326098203659, + 0.005328989587724209, + 0.0064683035016059875, + 0.0037125477101653814, + 0.00692816311493516, + 0.004925798624753952, + 0.004509110935032368, + 0.004861845634877682, + 0.0028943081852048635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.016078690066933632 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074623093008995 + ], + "opposite_logprob_tokens": [ + -4.511074542999268 + ], + "logprob_reward": -0.011074623093008995, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074623093008995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16327327489852905, + 0.07618290930986404, + 0.07123742997646332, + 0.07915075123310089, + 0.2059597223997116, + 0.11653905361890793, + 0.15035772323608398, + 0.07875850796699524, + 0.03735736757516861, + 0.02118326723575592 + ], + "credit_map": [ + 0.021128099411725998, + 0.02049747295677662, + 0.020214742049574852, + 0.015779539942741394, + 0.01925733871757984, + 0.02179565839469433, + 0.019863542169332504, + 0.024736877530813217, + 0.005992473103106022, + 0.005094766616821289, + 0.00613546883687377, + 0.0052175866439938545, + 0.009478219784796238, + 0.015858987346291542, + 0.009379051625728607, + 0.01902635768055916, + 0.005544757936149836, + 0.0058146193623542786, + 0.006356450729072094, + 0.0061389063484966755, + 0.010295816697180271, + 0.015978222712874413, + 0.011343083344399929, + 0.009765577502548695, + 0.006896816659718752, + 0.006917763967067003, + 0.007352777291089296, + 0.012025569565594196, + 0.010048954747617245, + 0.011644457466900349, + 0.012828563340008259, + 0.01143585704267025, + 0.016907822340726852, + 0.01569274067878723, + 0.01980474777519703, + 0.03339114785194397, + 0.01939254067838192, + 0.029293516650795937, + 0.04134892299771309, + 0.03012828901410103, + 0.011018035002052784, + 0.01253251638263464, + 0.009849971160292625, + 0.018008925020694733, + 0.012413010001182556, + 0.017915796488523483, + 0.020036347210407257, + 0.014764457009732723, + 0.01721460372209549, + 0.02458471618592739, + 0.015290374867618084, + 0.016042230650782585, + 0.025786887854337692, + 0.016663093119859695, + 0.01779671013355255, + 0.01697910763323307, + 0.007708430290222168, + 0.007595644798129797, + 0.010836795903742313, + 0.006024837959557772, + 0.008376236073672771, + 0.016401279717683792, + 0.008832668885588646, + 0.012982618063688278, + 0.004414335824549198, + 0.0045335241593420506, + 0.003022352932021022, + 0.003563138423487544, + 0.005398095119744539, + 0.005766617134213448, + 0.0063028293661773205, + 0.004356473684310913, + 0.0024358530063182116, + 0.0030294503085315228, + 0.0026308088563382626, + 0.002457056660205126, + 0.0029359362088143826, + 0.002249704208225012, + 0.0030637369491159916, + 0.0023807210382074118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.011074623093008995 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636845551431179 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636845551431179, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636845551431179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27622681856155396, + 0.040716368705034256, + 0.07244028151035309, + 0.03936145082116127, + 0.06862718611955643, + 0.029193203896284103, + 0.12861092388629913, + 0.053918905556201935, + 0.1739102005958557, + 0.11699468642473221 + ], + "credit_map": [ + 0.03150532767176628, + 0.04034942388534546, + 0.03781028091907501, + 0.02622576616704464, + 0.027002060785889626, + 0.039948221296072006, + 0.03347517549991608, + 0.039910558611154556, + 0.006083660759031773, + 0.004632940981537104, + 0.006386843044310808, + 0.0042979116551578045, + 0.0047798859886825085, + 0.005527058150619268, + 0.004693361930549145, + 0.004314705729484558, + 0.007319745607674122, + 0.0054568336345255375, + 0.007810845039784908, + 0.0068679493851959705, + 0.008333851583302021, + 0.01168906595557928, + 0.01566448248922825, + 0.009297503158450127, + 0.003971282858401537, + 0.0025672672782093287, + 0.00314203929156065, + 0.006656639277935028, + 0.004825124982744455, + 0.0052277944050729275, + 0.008473502472043037, + 0.004497800022363663, + 0.005278084427118301, + 0.006039911415427923, + 0.007460380904376507, + 0.011721918359398842, + 0.00882304273545742, + 0.010425813496112823, + 0.009571661241352558, + 0.009306378662586212, + 0.002876102225854993, + 0.004683386534452438, + 0.002531558508053422, + 0.002549968659877777, + 0.00359143759123981, + 0.003088756464421749, + 0.004816780332475901, + 0.005055213812738657, + 0.014116361737251282, + 0.021775824949145317, + 0.011938506737351418, + 0.017191078513860703, + 0.013821687549352646, + 0.014460157603025436, + 0.019464882090687752, + 0.015842413529753685, + 0.005605760961771011, + 0.006092306692153215, + 0.006886227056384087, + 0.00492647523060441, + 0.0062050470151007175, + 0.006854894105345011, + 0.006914801429957151, + 0.010433390736579895, + 0.024778012186288834, + 0.020732825621962547, + 0.016159825026988983, + 0.016241462901234627, + 0.03290792182087898, + 0.020588181912899017, + 0.025463595986366272, + 0.017038367688655853, + 0.013720138929784298, + 0.015273099765181541, + 0.015783200040459633, + 0.013800148852169514, + 0.013498671352863312, + 0.01784205622971058, + 0.016611741855740547, + 0.01046562846750021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.004636845551431179 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009213017183355987 + ], + "opposite_logprob_tokens": [ + -7.000921249389648 + ], + "logprob_reward": -0.0009213017183355987, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009213017183355987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09998918324708939, + 0.030506635084748268, + 0.04833496734499931, + 0.048536110669374466, + 0.03919932246208191, + 0.038422051817178726, + 0.08100976049900055, + 0.15372735261917114, + 0.24931101500988007, + 0.210963636636734 + ], + "credit_map": [ + 0.011229434050619602, + 0.00969256553798914, + 0.013934999704360962, + 0.009467324241995811, + 0.012546072714030743, + 0.014240778982639313, + 0.010812441818416119, + 0.018065566197037697, + 0.003202600171789527, + 0.0023299329914152622, + 0.0035624043084681034, + 0.003958563320338726, + 0.0030053763184696436, + 0.0048231021501123905, + 0.003701424226164818, + 0.005923233926296234, + 0.005397689528763294, + 0.004075417295098305, + 0.0052652498707175255, + 0.003720739157870412, + 0.004816102795302868, + 0.007256046403199434, + 0.010718741454184055, + 0.007084980141371489, + 0.004264977294951677, + 0.00476313941180706, + 0.004558788612484932, + 0.007613827008754015, + 0.007034625392407179, + 0.006983790546655655, + 0.0073562380857765675, + 0.005960722919553518, + 0.005357079207897186, + 0.0028513912111520767, + 0.005173070821911097, + 0.004981494043022394, + 0.00265564676374197, + 0.006818204186856747, + 0.005540715530514717, + 0.005821723025292158, + 0.0034176339395344257, + 0.006019561551511288, + 0.004081944469362497, + 0.0037058829329907894, + 0.004131114110350609, + 0.005608882289379835, + 0.005365780089050531, + 0.006091252900660038, + 0.0075424364767968655, + 0.010171105153858662, + 0.008790399879217148, + 0.009685104712843895, + 0.008839031681418419, + 0.012958393432199955, + 0.013680762611329556, + 0.009342525154352188, + 0.015336548909544945, + 0.01667761243879795, + 0.018171526491642, + 0.012722986750304699, + 0.016785399988293648, + 0.03315262869000435, + 0.014735297299921513, + 0.026145344600081444, + 0.0384526289999485, + 0.032825373113155365, + 0.021152056753635406, + 0.019213968887925148, + 0.047004565596580505, + 0.029360182583332062, + 0.03892626613378525, + 0.022375965490937233, + 0.03010382130742073, + 0.0553070530295372, + 0.017587870359420776, + 0.020181750878691673, + 0.025825755670666695, + 0.02772454172372818, + 0.021426547318696976, + 0.012806296348571777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.0009213017183355987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful first-person perspective shot places the viewer in the perspective of the mythical figure Nezha, whose body transforms into a stream of red and gold light, hurtling toward the ancient Great Wall at the speed of a cannonball. At the edge of the field of vision, clenched fists can be seen, wrists adorned with the Qiankun Circle, gleaming with a dark golden glow, and the faint afterimage of a red Hun Tian Ling silken thread dancing like flames. Beneath a gloomy, dark cloud, the Great Wall ahead rapidly zooms in, its weathered gray-brown masonry, its mottled texture, and the deep green moss growing in its crevices becoming crystal clear. With a roar like the shattering of heaven and earth, the camera jerks forward and trembles violently. A destructive force erupts, shattering the vast masonry of the wall before our eyes. Countless rocks, debris, and thick dust of varying sizes scatter violently in all directions and toward the camera, blurring the entire field of view and creating a breathtaking, primal, and violent impact, an atmosphere of a grand collision of divine power.", + "seed": 11535, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_0/index_9363_seed_11535_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02741894875653088, + "vlm_logprob": -0.02741894875653088, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02741894875653088, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06285194307565689 + ], + "opposite_logprob_tokens": [ + -2.812851905822754 + ], + "logprob_reward": -0.06285194307565689, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06285194307565689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14630113542079926, + 0.0842328667640686, + 0.08505967259407043, + 0.07456788420677185, + 0.19238747656345367, + 0.08219602704048157, + 0.07199112325906754, + 0.08700094372034073, + 0.10107151418924332, + 0.07519131153821945 + ], + "credit_map": [ + 0.016444962471723557, + 0.016751211136579514, + 0.019552065059542656, + 0.015129347331821918, + 0.015302920714020729, + 0.01970616914331913, + 0.0241681020706892, + 0.019246358424425125, + 0.005901042837649584, + 0.004930646158754826, + 0.008848784491419792, + 0.006702801678329706, + 0.004919387400150299, + 0.021899789571762085, + 0.009095778688788414, + 0.02193463407456875, + 0.008876379579305649, + 0.009297912009060383, + 0.01101246289908886, + 0.007053971756249666, + 0.014858843758702278, + 0.013923030346632004, + 0.010838385671377182, + 0.009198679588735104, + 0.007998421788215637, + 0.01040350366383791, + 0.009335361421108246, + 0.0036339841317385435, + 0.008353324607014656, + 0.011939163319766521, + 0.012207647785544395, + 0.010696480982005596, + 0.030626535415649414, + 0.03750095143914223, + 0.023470768705010414, + 0.025426940992474556, + 0.018438320606946945, + 0.015118362382054329, + 0.01563112623989582, + 0.026174474507570267, + 0.01483171246945858, + 0.013217343017458916, + 0.00882982462644577, + 0.009110372513532639, + 0.008305677212774754, + 0.008609490469098091, + 0.010080622509121895, + 0.009210984222590923, + 0.007793104741722345, + 0.006135540548712015, + 0.011053015477955341, + 0.008126058615744114, + 0.009875633753836155, + 0.007645326666533947, + 0.012789050117135048, + 0.008573392406105995, + 0.008056508377194405, + 0.008863680064678192, + 0.009658869355916977, + 0.01799905113875866, + 0.010419752448797226, + 0.007090969476848841, + 0.011431199498474598, + 0.013480911031365395, + 0.015527657233178616, + 0.01593058742582798, + 0.010796991176903248, + 0.006987052503973246, + 0.015058659017086029, + 0.012317681685090065, + 0.009749279357492924, + 0.014703606255352497, + 0.006313913501799107, + 0.014442982152104378, + 0.012471525929868221, + 0.0095720449462533, + 0.00899264495819807, + 0.006753325928002596, + 0.00968415942043066, + 0.006960711907595396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.06285194307565689 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329622581601143 + ], + "opposite_logprob_tokens": [ + -3.7732961177825928 + ], + "logprob_reward": -0.02329622581601143, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329622581601143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11754562705755234, + 0.09948849678039551, + 0.03709428757429123, + 0.05609585717320442, + 0.050626207143068314, + 0.060404788702726364, + 0.09217175096273422, + 0.12671588361263275, + 0.23233571648597717, + 0.1275213658809662 + ], + "credit_map": [ + 0.015219636261463165, + 0.012087824754416943, + 0.01478913426399231, + 0.010661832056939602, + 0.009309330955147743, + 0.016132796183228493, + 0.024066079407930374, + 0.015278995037078857, + 0.005615785717964172, + 0.004617094527930021, + 0.0113824512809515, + 0.007333339191973209, + 0.008817659690976143, + 0.023892909288406372, + 0.0107446713373065, + 0.027084581553936005, + 0.004882900509983301, + 0.004697267431765795, + 0.005392922554165125, + 0.0027891602367162704, + 0.004632944241166115, + 0.006325883325189352, + 0.0037676719948649406, + 0.00460553914308548, + 0.005172492936253548, + 0.004555674735456705, + 0.006916764657944441, + 0.006160899996757507, + 0.0054571921937167645, + 0.012735196389257908, + 0.006832804065197706, + 0.008264833129942417, + 0.004626531153917313, + 0.005995142739266157, + 0.004500225652009249, + 0.009649449959397316, + 0.004674222320318222, + 0.005561238154768944, + 0.007173729129135609, + 0.008445666171610355, + 0.0062270984053611755, + 0.007358531001955271, + 0.005320328753441572, + 0.008921385742723942, + 0.009334939531981945, + 0.006450880318880081, + 0.009392458945512772, + 0.007399165071547031, + 0.008168795146048069, + 0.008562114089727402, + 0.010308881290256977, + 0.009924090467393398, + 0.019019998610019684, + 0.013556433841586113, + 0.011879459023475647, + 0.010751972906291485, + 0.010724689811468124, + 0.012730449438095093, + 0.014950466342270374, + 0.01639474369585514, + 0.014741169288754463, + 0.013938596472144127, + 0.01606016792356968, + 0.027175597846508026, + 0.026591308414936066, + 0.03597524017095566, + 0.02885650098323822, + 0.016510741785168648, + 0.03306488320231438, + 0.022398848086595535, + 0.017727533355355263, + 0.05121064558625221, + 0.01225188933312893, + 0.01974259875714779, + 0.021325625479221344, + 0.021198183298110962, + 0.014336391352117062, + 0.007595981005579233, + 0.02111918106675148, + 0.009951516054570675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.02329622581601143 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305985942482948 + ], + "opposite_logprob_tokens": [ + -3.168059825897217 + ], + "logprob_reward": -0.04305985942482948, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305985942482948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17902067303657532, + 0.11905141919851303, + 0.06466145813465118, + 0.02749796211719513, + 0.046352796256542206, + 0.033456817269325256, + 0.03857055678963661, + 0.039720915257930756, + 0.06319158524274826, + 0.38847586512565613 + ], + "credit_map": [ + 0.017109792679548264, + 0.013253956101834774, + 0.01669282466173172, + 0.01961664855480194, + 0.01737707108259201, + 0.030267253518104553, + 0.03600963205099106, + 0.02869347855448723, + 0.006321379914879799, + 0.00430184043943882, + 0.009783775545656681, + 0.005084949545562267, + 0.005888781510293484, + 0.03623827174305916, + 0.012708517722785473, + 0.038723912090063095, + 0.008246999233961105, + 0.006903988309204578, + 0.003991093952208757, + 0.0031187785789370537, + 0.006575321778655052, + 0.011963853612542152, + 0.007451413664966822, + 0.016410009935498238, + 0.0033467314206063747, + 0.002761691343039274, + 0.0034719533286988735, + 0.002021457301452756, + 0.0035034052561968565, + 0.0052172658033668995, + 0.0034837869461625814, + 0.0036916686221957207, + 0.0046446314081549644, + 0.004267876967787743, + 0.004589094314724207, + 0.006522510666400194, + 0.004115666262805462, + 0.005326139274984598, + 0.008837254717946053, + 0.0080496184527874, + 0.006156355608254671, + 0.003774945857003331, + 0.002792009385302663, + 0.003626142628490925, + 0.0035894003231078386, + 0.004192625172436237, + 0.005105307325720787, + 0.004220032133162022, + 0.003498231526464224, + 0.0035185660235583782, + 0.004181677009910345, + 0.007522068917751312, + 0.0055948407389223576, + 0.0034607842098921537, + 0.006914195604622364, + 0.0038801925256848335, + 0.003641526447609067, + 0.005775695666670799, + 0.004592422861605883, + 0.005054404493421316, + 0.005410889163613319, + 0.0031495944131165743, + 0.004066817462444305, + 0.00802956335246563, + 0.008765872567892075, + 0.007751660421490669, + 0.00796052347868681, + 0.004842717666178942, + 0.006875257007777691, + 0.005521268583834171, + 0.006009769160300493, + 0.015464515425264835, + 0.028759077191352844, + 0.04907497018575668, + 0.04987568035721779, + 0.0628625676035881, + 0.05894282087683678, + 0.03688143938779831, + 0.0549781508743763, + 0.04710114374756813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.04305985942482948 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059472741559147835 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.0059472741559147835, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059472741559147835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09236359596252441, + 0.04113157093524933, + 0.03214067965745926, + 0.031782906502485275, + 0.021700890734791756, + 0.0741036981344223, + 0.2305617779493332, + 0.06778764724731445, + 0.10094626247882843, + 0.30748099088668823 + ], + "credit_map": [ + 0.008978860452771187, + 0.007204579655081034, + 0.008722624741494656, + 0.010449817404150963, + 0.007897662930190563, + 0.012911158613860607, + 0.023034049198031425, + 0.013164839707314968, + 0.002554898150265217, + 0.001917508547194302, + 0.00360091426409781, + 0.0035331877879798412, + 0.001942928764037788, + 0.011811965145170689, + 0.003207595320418477, + 0.012562570162117481, + 0.005834722425788641, + 0.003976808395236731, + 0.0018805117579177022, + 0.0022301627323031425, + 0.002670645946636796, + 0.006523538380861282, + 0.0029652975499629974, + 0.006058995146304369, + 0.0029872371815145016, + 0.00408132141456008, + 0.0034127328544855118, + 0.0036583568435162306, + 0.003415802726522088, + 0.005350786726921797, + 0.004564756061881781, + 0.004311911761760712, + 0.002689564600586891, + 0.0031923861242830753, + 0.002176132518798113, + 0.0033848450984805822, + 0.0017575531965121627, + 0.001748222392052412, + 0.003448852337896824, + 0.0033033345825970173, + 0.007459464482963085, + 0.010746181942522526, + 0.005297230090945959, + 0.008704076521098614, + 0.010427719913423061, + 0.011860692873597145, + 0.01090669073164463, + 0.00870163831859827, + 0.020664962008595467, + 0.020337140187621117, + 0.020003676414489746, + 0.028364134952425957, + 0.04274880886077881, + 0.032166797667741776, + 0.03806368261575699, + 0.028212571516633034, + 0.005482180044054985, + 0.007188571151345968, + 0.008186870254576206, + 0.009676664136350155, + 0.009817319922149181, + 0.006804516073316336, + 0.00856051780283451, + 0.012071007862687111, + 0.01795494370162487, + 0.01641159877181053, + 0.011239869520068169, + 0.009024805389344692, + 0.014748051762580872, + 0.008678631857037544, + 0.009808309376239777, + 0.013080047443509102, + 0.029435448348522186, + 0.06703700870275497, + 0.04684045538306236, + 0.04069790244102478, + 0.04391540586948395, + 0.020190371200442314, + 0.028975769877433777, + 0.030388634651899338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.0059472741559147835 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing clenched fists with dark golden Qiankun Circle wristbands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character's body transform into a stream of red and gold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the target an ancient Great Wall featuring weathered gray-brown masonry with deep green moss in the crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera jerk forward and tremble violently while a destructive force shatters the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do countless rocks, debris, and thick dust scatter violently in all directions, blurring the field of view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019394413102418184 + ], + "opposite_logprob_tokens": [ + -6.251939296722412 + ], + "logprob_reward": -0.0019394413102418184, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019394413102418184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10887598246335983, + 0.032351747155189514, + 0.05258208140730858, + 0.03410577401518822, + 0.04569218307733536, + 0.08077501505613327, + 0.13723616302013397, + 0.03806155547499657, + 0.1832968145608902, + 0.2870227098464966 + ], + "credit_map": [ + 0.01450506504625082, + 0.012080369517207146, + 0.010509788990020752, + 0.0140610346570611, + 0.008932522498071194, + 0.012896936386823654, + 0.020206226035952568, + 0.015684030950069427, + 0.00285117095336318, + 0.0027375107165426016, + 0.004218034911900759, + 0.003975509665906429, + 0.00283321226015687, + 0.005007641389966011, + 0.004514175001531839, + 0.006214491557329893, + 0.00678697694092989, + 0.007368636317551136, + 0.005302598234266043, + 0.0037073453422635794, + 0.006494835484772921, + 0.007190953940153122, + 0.006488226354122162, + 0.009242505766451359, + 0.003847625106573105, + 0.003595227375626564, + 0.00376122840680182, + 0.0032846981193870306, + 0.004009476397186518, + 0.006522577255964279, + 0.0045251548290252686, + 0.004559783264994621, + 0.005347604863345623, + 0.0065274471417069435, + 0.004360591992735863, + 0.0076281665824353695, + 0.0032975724898278713, + 0.0032991673797369003, + 0.00860541220754385, + 0.006626221351325512, + 0.008340590633451939, + 0.012939434498548508, + 0.005353779066354036, + 0.010196330957114697, + 0.010974830016493797, + 0.0119914086535573, + 0.012221155688166618, + 0.00875747948884964, + 0.014188376255333424, + 0.011707989498972893, + 0.013409885577857494, + 0.01579876057803631, + 0.024610942229628563, + 0.019017741084098816, + 0.022006239742040634, + 0.01649622991681099, + 0.00383010134100914, + 0.004689157474786043, + 0.00536844739690423, + 0.00520422775298357, + 0.003868245519697666, + 0.00503897899761796, + 0.004111794289201498, + 0.005950602702796459, + 0.036970287561416626, + 0.020791005343198776, + 0.021985290572047234, + 0.014598687179386616, + 0.03669903054833412, + 0.015239452943205833, + 0.014159872196614742, + 0.02285318449139595, + 0.029157938435673714, + 0.06310239434242249, + 0.047391533851623535, + 0.036904484033584595, + 0.037294190376996994, + 0.01646900735795498, + 0.02957172878086567, + 0.02713143639266491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.0019394413102418184 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_1/index_tensor([8497, 8497])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_1/index_tensor([8497, 8497])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..32dad9c4955aa2baf796b7f98936c6cd556d36c5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_1/index_tensor([8497, 8497])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, at eye level, frames the scene, the camera panning slowly and steadily across the face of a grieving adult African-American woman. She's in a park on a blustery day, the wind whipping her long, dark, intricately braided hair, some of which hang down against her cheeks. She wears a simple, dark charcoal-colored hoodie, the soft fabric pulled up at her neck. The background is a beautiful, out-of-focus bokeh of autumn leaves, soft swirls of gold and brown under the flat, diffuse light of an overcast sky. Her eyes are deep and expressive, lowered and furrowed, as she reads something off-screen, revealing a profound sadness. As the camera pans across her features, a perfectly formed tear escapes from the corner of her eye, slowly and deliberately running down her smooth cheek, glistening. Her lips, slightly parted, tremble almost imperceptibly, creating a powerful, intimate, and melancholic atmosphere of quiet heartbreak.", + "seed": 25701, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_1/index_8497_seed_25701_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2251458117738366, + "vlm_logprob": -0.2251458117738366, + "vlm_yes_no_margin": 2.3500000178813933, + "vlm_token_credit": -0.2251458117738366, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013528100214898586 + ], + "opposite_logprob_tokens": [ + -5.013527870178223 + ], + "logprob_reward": -0.013528100214898586, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.013528100214898586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2780044674873352, + 0.12197792530059814, + 0.14622682332992554, + 0.07723589241504669, + 0.13454900681972504, + 0.050448425114154816, + 0.07824181020259857, + 0.044446155428886414, + 0.024346066638827324, + 0.04452342167496681 + ], + "credit_map": [ + 0.01793633960187435, + 0.03209720179438591, + 0.0235806405544281, + 0.043494321405887604, + 0.021229971200227737, + 0.02171015739440918, + 0.02538273110985756, + 0.09257308393716812, + 0.02032047137618065, + 0.011178486049175262, + 0.02712300419807434, + 0.016845358535647392, + 0.007210449315607548, + 0.017979657277464867, + 0.009979167021811008, + 0.011341331526637077, + 0.015389461070299149, + 0.017826572060585022, + 0.010932201519608498, + 0.031982630491256714, + 0.007352028973400593, + 0.029150111600756645, + 0.013299207203090191, + 0.02029460109770298, + 0.006928306072950363, + 0.008038995787501335, + 0.016009174287319183, + 0.014079063199460506, + 0.005333429668098688, + 0.009888374246656895, + 0.006637869402766228, + 0.010320684872567654, + 0.024449491873383522, + 0.018441155552864075, + 0.0138172572478652, + 0.0377865768969059, + 0.007660115137696266, + 0.008823770098388195, + 0.014095287770032883, + 0.009475353173911572, + 0.004245673771947622, + 0.006218872498720884, + 0.005617641378194094, + 0.00969509407877922, + 0.005453136749565601, + 0.0030350780580192804, + 0.007918432354927063, + 0.008264493197202682, + 0.01044623926281929, + 0.00727187329903245, + 0.012021159753203392, + 0.011611918918788433, + 0.010851611383259296, + 0.00830369908362627, + 0.010368132032454014, + 0.007367176469415426, + 0.004361454863101244, + 0.0033984477631747723, + 0.008129697293043137, + 0.005189781542867422, + 0.0043113213032484055, + 0.003980662208050489, + 0.005995009560137987, + 0.009079781360924244, + 0.002372383838519454, + 0.0019050899427384138, + 0.004996440839022398, + 0.004000977613031864, + 0.0023277767468243837, + 0.0012837863760069013, + 0.0032929065637290478, + 0.004166705533862114, + 0.005287094507366419, + 0.0034937667660415173, + 0.008406558074057102, + 0.005713757127523422, + 0.005069405771791935, + 0.005037884693592787, + 0.004850343801081181, + 0.006664608605206013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.013528100214898586 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020613370463252068 + ], + "opposite_logprob_tokens": [ + -3.895613431930542 + ], + "logprob_reward": -0.020613370463252068, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020613370463252068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21765685081481934, + 0.01779916323721409, + 0.10708748549222946, + 0.15430566668510437, + 0.4506005644798279, + 0.00702892430126667, + 0.01040429063141346, + 0.012402921915054321, + 0.008193779736757278, + 0.014520268887281418 + ], + "credit_map": [ + 0.013490455225110054, + 0.0331350676715374, + 0.018741946667432785, + 0.04023800417780876, + 0.01602752134203911, + 0.015484113246202469, + 0.017983919009566307, + 0.06255582720041275, + 0.001551157678477466, + 0.002228210214525461, + 0.00202564406208694, + 0.0027576875872910023, + 0.003370485035702586, + 0.001265780534595251, + 0.002918289275839925, + 0.0016819091979414225, + 0.010337138548493385, + 0.012839061208069324, + 0.007981870323419571, + 0.022730864584445953, + 0.005602849647402763, + 0.020350871607661247, + 0.008676531724631786, + 0.01856829784810543, + 0.013894673436880112, + 0.015676848590373993, + 0.03535086661577225, + 0.030246112495660782, + 0.00942213088274002, + 0.017560839653015137, + 0.013503876514732838, + 0.018650328740477562, + 0.0858556404709816, + 0.057978324592113495, + 0.04013122618198395, + 0.1389734447002411, + 0.021712323650717735, + 0.029372001066803932, + 0.04614214599132538, + 0.03043549507856369, + 0.0007850826368667185, + 0.000841831206344068, + 0.000761941191740334, + 0.0012306713033467531, + 0.0009224212844856083, + 0.0003231523442082107, + 0.001253588474355638, + 0.0009102363255806267, + 0.0010072689037770033, + 0.0010254140943288803, + 0.0014996316749602556, + 0.0015719843795523047, + 0.001328930607996881, + 0.0009098705486394465, + 0.0016379113076254725, + 0.0014232795219868422, + 0.0011122412979602814, + 0.0007011456764303148, + 0.00228285719640553, + 0.0015326979337260127, + 0.001224124920554459, + 0.0014211091911420226, + 0.0018481012666597962, + 0.002280645305290818, + 0.000984154874458909, + 0.0006377004901878536, + 0.0018302800599485636, + 0.0009202464716508985, + 0.0007208648603409529, + 0.0009526180219836533, + 0.0007700046407990158, + 0.0013779100263491273, + 0.0013575439807027578, + 0.0010251908097416162, + 0.002737481379881501, + 0.0019896114245057106, + 0.0019832318648695946, + 0.00123053218703717, + 0.0023045307025313377, + 0.0018921466544270515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.020613370463252068 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285863935947418 + ], + "opposite_logprob_tokens": [ + -2.0178587436676025 + ], + "logprob_reward": -0.14285863935947418, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14285863935947418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14378279447555542, + 0.11497395485639572, + 0.08501916378736496, + 0.24588657915592194, + 0.2537015974521637, + 0.0538518913090229, + 0.019866889342665672, + 0.035845424979925156, + 0.024017881602048874, + 0.023053834214806557 + ], + "credit_map": [ + 0.025855008512735367, + 0.01292258407920599, + 0.013042813166975975, + 0.025747494772076607, + 0.009996599517762661, + 0.02046719379723072, + 0.013406540267169476, + 0.02234455943107605, + 0.0198063962161541, + 0.011360125616192818, + 0.01903413236141205, + 0.0143778370693326, + 0.009431537240743637, + 0.016589488834142685, + 0.01225796528160572, + 0.012116461992263794, + 0.005793726071715355, + 0.008549170568585396, + 0.006746552884578705, + 0.017444586381316185, + 0.004834020044654608, + 0.005107136908918619, + 0.0075830030255019665, + 0.02896096557378769, + 0.02052895724773407, + 0.03553149476647377, + 0.04908223822712898, + 0.05127444863319397, + 0.020725037902593613, + 0.02099473588168621, + 0.02568640559911728, + 0.022063273936510086, + 0.04380599036812782, + 0.035877861082553864, + 0.02634493075311184, + 0.07300237566232681, + 0.016023991629481316, + 0.014732437208294868, + 0.02644781582057476, + 0.017466187477111816, + 0.006366624031215906, + 0.009511747397482395, + 0.007311770226806402, + 0.011803652159869671, + 0.005401107016950846, + 0.0025245139840990305, + 0.007314679212868214, + 0.003617800772190094, + 0.0020874233450740576, + 0.001781674800440669, + 0.0025950295384973288, + 0.002842091955244541, + 0.0017775528831407428, + 0.0013883172068744898, + 0.0017032958567142487, + 0.005691505502909422, + 0.0035590690094977617, + 0.0032573153730481863, + 0.008532249368727207, + 0.004595628939568996, + 0.00315635628066957, + 0.00432975310832262, + 0.003227190114557743, + 0.00518786022439599, + 0.003018787130713463, + 0.0018541030585765839, + 0.005121342837810516, + 0.002344735199585557, + 0.0021974851842969656, + 0.0041310167871415615, + 0.0016789891524240375, + 0.003671422367915511, + 0.002052083844318986, + 0.0016537867486476898, + 0.003704793518409133, + 0.0032554841600358486, + 0.002787493634968996, + 0.0021142144687473774, + 0.004657345358282328, + 0.002828630618751049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.14285863935947418 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47439777851104736 + ], + "opposite_logprob_tokens": [ + -0.9743977785110474 + ], + "logprob_reward": -0.47439777851104736, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47439777851104736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07799027115106583, + 0.09978519380092621, + 0.17929793894290924, + 0.32137537002563477, + 0.09235382080078125, + 0.020613370463252068, + 0.03052416443824768, + 0.09899064898490906, + 0.01796570047736168, + 0.0611034519970417 + ], + "credit_map": [ + 0.012992126867175102, + 0.007894659414887428, + 0.008155281655490398, + 0.015901975333690643, + 0.007852460257709026, + 0.008393465541303158, + 0.010387858375906944, + 0.0064124446362257, + 0.011094549670815468, + 0.011823690496385098, + 0.01362607255578041, + 0.015210258774459362, + 0.013887361623346806, + 0.007041106000542641, + 0.017307568341493607, + 0.009794596582651138, + 0.0155857028439641, + 0.020184621214866638, + 0.012383121065795422, + 0.03458940237760544, + 0.007374315522611141, + 0.03454173356294632, + 0.014576165936887264, + 0.040062859654426575, + 0.027718523517251015, + 0.03431349992752075, + 0.06961775571107864, + 0.06688212603330612, + 0.020858315750956535, + 0.03425907716155052, + 0.031107274815440178, + 0.036618802696466446, + 0.01707732304930687, + 0.011594814248383045, + 0.007809531409293413, + 0.02626660093665123, + 0.006629572715610266, + 0.005603766534477472, + 0.010992701165378094, + 0.006379513069987297, + 0.002506840042769909, + 0.0026268947403877974, + 0.002146258717402816, + 0.0029692654497921467, + 0.0030029164627194405, + 0.002271169563755393, + 0.0028957989998161793, + 0.002194227883592248, + 0.001925180433318019, + 0.00234400387853384, + 0.0035415024030953646, + 0.005342216696590185, + 0.004814004059880972, + 0.0017835728358477354, + 0.008014565333724022, + 0.0027591160032898188, + 0.008191701024770737, + 0.007239514496177435, + 0.01631321758031845, + 0.014192621223628521, + 0.011395979672670364, + 0.010849094949662685, + 0.02168251760303974, + 0.009125998243689537, + 0.0021945154294371605, + 0.001139810890890658, + 0.0038986990693956614, + 0.002853244077414274, + 0.0019917169120162725, + 0.0016127991257235408, + 0.002311736112460494, + 0.001963177928701043, + 0.00628712959587574, + 0.005373605526983738, + 0.00790379848331213, + 0.008604908362030983, + 0.008402818813920021, + 0.005499138496816158, + 0.011762798763811588, + 0.007269251625984907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.47439777851104736 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47433117032051086 + ], + "opposite_logprob_tokens": [ + -0.9743311405181885 + ], + "logprob_reward": -0.47433117032051086, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.47433117032051086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22958505153656006, + 0.1274447739124298, + 0.23416148126125336, + 0.06139520928263664, + 0.06106029078364372, + 0.03898204118013382, + 0.08052078634500504, + 0.06552109122276306, + 0.03270826116204262, + 0.06862103939056396 + ], + "credit_map": [ + 0.031029056757688522, + 0.03512515500187874, + 0.027833227068185806, + 0.030299928039312363, + 0.020700128749012947, + 0.016064371913671494, + 0.023690495640039444, + 0.044842686504125595, + 0.024708403274416924, + 0.010601588524878025, + 0.0231182724237442, + 0.014565112069249153, + 0.008731916546821594, + 0.014878460206091404, + 0.010689936578273773, + 0.020151076838374138, + 0.0251768771559, + 0.031082553789019585, + 0.017568228766322136, + 0.06173623725771904, + 0.013047287240624428, + 0.04454931989312172, + 0.021638741716742516, + 0.01936223730444908, + 0.004616840742528439, + 0.008241850882768631, + 0.011687553487718105, + 0.012146562337875366, + 0.006492244079709053, + 0.004900901112705469, + 0.007427316624671221, + 0.005881942342966795, + 0.009752897545695305, + 0.008026473224163055, + 0.007187183015048504, + 0.017288444563746452, + 0.004475639201700687, + 0.003759967628866434, + 0.0067232693545520306, + 0.0038464136887341738, + 0.004171752370893955, + 0.003982568625360727, + 0.004197889473289251, + 0.008627183735370636, + 0.004878323990851641, + 0.0019158467184752226, + 0.008078065700829029, + 0.003130408236756921, + 0.01050371490418911, + 0.007903234101831913, + 0.011512541212141514, + 0.010323544032871723, + 0.008697357028722763, + 0.008754348382353783, + 0.010635788552463055, + 0.012190258130431175, + 0.007276435382664204, + 0.008888062089681625, + 0.005942775867879391, + 0.011075802147388458, + 0.007987389340996742, + 0.0031234167981892824, + 0.010278059169650078, + 0.010949146933853626, + 0.00373952928930521, + 0.002869885880500078, + 0.003782636718824506, + 0.003949387464672327, + 0.004329612012952566, + 0.00653090188279748, + 0.003669534809887409, + 0.0038367738015949726, + 0.00711206765845418, + 0.008201575838029385, + 0.01152217946946621, + 0.009319798089563847, + 0.007573350332677364, + 0.006339491344988346, + 0.009659378789365292, + 0.008893200196325779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.47433117032051086 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, at eye level, frames the scene, the camera panning slowly and steadily across the face of a grieving adult African-American woman. She's in a park on a blustery day, the wind whipping her long, dark, intricately braided hair, some of which hang down against her cheeks. She wears a simple, dark charcoal-colored hoodie, the soft fabric pulled up at her neck. The background is a beautiful, out-of-focus bokeh of autumn leaves, soft swirls of gold and brown under the flat, diffuse light of an overcast sky. Her eyes are deep and expressive, lowered and furrowed, as she reads something off-screen, revealing a profound sadness. As the camera pans across her features, a perfectly formed tear escapes from the corner of her eye, slowly and deliberately running down her smooth cheek, glistening. Her lips, slightly parted, tremble almost imperceptibly, creating a powerful, intimate, and melancholic atmosphere of quiet heartbreak.", + "seed": 25702, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_1/index_8497_seed_25702_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13013069033622743, + "vlm_logprob": -0.13013069033622743, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.13013069033622743, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016589319333434105 + ], + "opposite_logprob_tokens": [ + -4.641589164733887 + ], + "logprob_reward": -0.016589319333434105, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.016589319333434105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2665964961051941, + 0.10933385044336319, + 0.0862225592136383, + 0.04188058525323868, + 0.1347016990184784, + 0.02543851174414158, + 0.07961210608482361, + 0.06816919893026352, + 0.15859299898147583, + 0.029451992362737656 + ], + "credit_map": [ + 0.04999176040291786, + 0.01201990619301796, + 0.09823726117610931, + 0.03349081054329872, + 0.01824243739247322, + 0.014806543476879597, + 0.019929081201553345, + 0.019878705963492393, + 0.010022365488111973, + 0.007668675389140844, + 0.02497873827815056, + 0.02015025168657303, + 0.007637634873390198, + 0.008002597838640213, + 0.01987328752875328, + 0.011000299826264381, + 0.016013406217098236, + 0.005211973562836647, + 0.014628271572291851, + 0.012419265694916248, + 0.00457268999889493, + 0.01303543709218502, + 0.009937492199242115, + 0.01040402427315712, + 0.0047183954156935215, + 0.003923983313143253, + 0.005654535256326199, + 0.009021400474011898, + 0.0050406064838171005, + 0.0032219323329627514, + 0.007251683622598648, + 0.0030480497516691685, + 0.034092046320438385, + 0.005211269482970238, + 0.019256891682744026, + 0.008323820307850838, + 0.007618616335093975, + 0.038471367210149765, + 0.008311515673995018, + 0.013416176661849022, + 0.0022761228028684855, + 0.002396373078227043, + 0.003861184697598219, + 0.006230109371244907, + 0.0028930671978741884, + 0.001738940249197185, + 0.002738237613812089, + 0.0033044766169041395, + 0.005297292489558458, + 0.005160122644156218, + 0.012825368903577328, + 0.007881377823650837, + 0.01215421687811613, + 0.0034301455598324537, + 0.017712309956550598, + 0.015151271596550941, + 0.006081600207835436, + 0.007158458698540926, + 0.009353341534733772, + 0.00961832795292139, + 0.004390798043459654, + 0.0047837356105446815, + 0.009345612488687038, + 0.017437327653169632, + 0.03866284713149071, + 0.014511074870824814, + 0.010491003282368183, + 0.014370176941156387, + 0.010024333372712135, + 0.04008062183856964, + 0.01147481333464384, + 0.01897813193500042, + 0.0043657501228153706, + 0.0024837893433868885, + 0.004424453712999821, + 0.0028141255024820566, + 0.002342373598366976, + 0.0040375841781497, + 0.0027999323792755604, + 0.006183984689414501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.016589319333434105 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026378588750958443 + ], + "opposite_logprob_tokens": [ + -3.651378631591797 + ], + "logprob_reward": -0.026378588750958443, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026378588750958443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21437332034111023, + 0.07469288259744644, + 0.060983747243881226, + 0.027282772585749626, + 0.10751941800117493, + 0.3004482090473175, + 0.029850030317902565, + 0.1378863900899887, + 0.01988300122320652, + 0.027080213651061058 + ], + "credit_map": [ + 0.024028535932302475, + 0.006939264014363289, + 0.09405581653118134, + 0.04737405478954315, + 0.007071556989103556, + 0.009894597344100475, + 0.013684830628335476, + 0.011324664577841759, + 0.011211003176867962, + 0.005888137500733137, + 0.011765781790018082, + 0.008298800326883793, + 0.004397786688059568, + 0.012425548397004604, + 0.008819022215902805, + 0.011886796914041042, + 0.014993415214121342, + 0.002563788555562496, + 0.011674263514578342, + 0.006117951590567827, + 0.0027869902551174164, + 0.01128414273262024, + 0.0060461764223873615, + 0.0055170184932649136, + 0.004524742718786001, + 0.0018583008786663413, + 0.006819520145654678, + 0.0033833913039416075, + 0.0021154095884412527, + 0.0021766028366982937, + 0.0029747814405709505, + 0.003430023090913892, + 0.022472864016890526, + 0.0035979272797703743, + 0.03238506615161896, + 0.008339494466781616, + 0.004716144874691963, + 0.020893065258860588, + 0.005107206758111715, + 0.010007653385400772, + 0.06828486919403076, + 0.018815897405147552, + 0.031125256791710854, + 0.04202915355563164, + 0.01511334627866745, + 0.01699623093008995, + 0.021156854927539825, + 0.08692659437656403, + 0.003229865338653326, + 0.0020536098163574934, + 0.008262362331151962, + 0.0029750110115855932, + 0.002701192395761609, + 0.002211992861703038, + 0.0026111940387636423, + 0.005804804153740406, + 0.009563811123371124, + 0.014607488177716732, + 0.017860030755400658, + 0.034872617572546005, + 0.008027161471545696, + 0.00994956772774458, + 0.022792382165789604, + 0.020213333889842033, + 0.004138148855417967, + 0.0016631510807201266, + 0.0012632007710635662, + 0.0017881713574752212, + 0.0019025654764845967, + 0.004420764744281769, + 0.0016421546461060643, + 0.003064845921471715, + 0.003059262642636895, + 0.0022945778910070658, + 0.004590492229908705, + 0.003194571239873767, + 0.0027789929881691933, + 0.0025499993935227394, + 0.0034538302570581436, + 0.005158486310392618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.026378588750958443 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902469485998154 + ], + "opposite_logprob_tokens": [ + -2.5790247917175293 + ], + "logprob_reward": -0.07902469485998154, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902469485998154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45315226912498474, + 0.06509113311767578, + 0.04174702987074852, + 0.035774845629930496, + 0.08483805507421494, + 0.031765781342983246, + 0.06264790147542953, + 0.024608176201581955, + 0.03848796710371971, + 0.1618868112564087 + ], + "credit_map": [ + 0.05968330800533295, + 0.01593293622136116, + 0.1625523418188095, + 0.07893978804349899, + 0.02268851548433304, + 0.0281752347946167, + 0.03648626059293747, + 0.048693884164094925, + 0.006927672307938337, + 0.005951299332082272, + 0.011983064003288746, + 0.011934126727283001, + 0.005024496931582689, + 0.006389694754034281, + 0.01003186870366335, + 0.0068489075638353825, + 0.007053666282445192, + 0.0022307327017188072, + 0.006491374224424362, + 0.004286801442503929, + 0.002532270038500428, + 0.004380791913717985, + 0.0046717082150280476, + 0.010099684819579124, + 0.00523787597194314, + 0.002327683847397566, + 0.006340825464576483, + 0.004449703264981508, + 0.0029408216942101717, + 0.003895289497449994, + 0.003866358892992139, + 0.006716286763548851, + 0.0185527503490448, + 0.004123153164982796, + 0.013501353561878204, + 0.009027567692101002, + 0.003611456137150526, + 0.024750858545303345, + 0.005013699177652597, + 0.006257216911762953, + 0.005372939165681601, + 0.0024551651440560818, + 0.004172425717115402, + 0.004790575243532658, + 0.0019992373418062925, + 0.002633853815495968, + 0.002542672911658883, + 0.007798912934958935, + 0.006371650844812393, + 0.007424159906804562, + 0.009309080429375172, + 0.011517558246850967, + 0.00568019226193428, + 0.004114330280572176, + 0.008515505120158195, + 0.009715423919260502, + 0.003716827370226383, + 0.002036587567999959, + 0.00583283556625247, + 0.003786752698943019, + 0.0016399333253502846, + 0.0016056634485721588, + 0.0030846577137708664, + 0.0029049201402813196, + 0.0050230929628014565, + 0.004453339613974094, + 0.0038426928222179413, + 0.00622727582231164, + 0.0038057819474488497, + 0.005175128113478422, + 0.003816448152065277, + 0.006144210230559111, + 0.02010040357708931, + 0.012072322890162468, + 0.024996476247906685, + 0.014593415893614292, + 0.014206484891474247, + 0.023001844063401222, + 0.017298826947808266, + 0.035617053508758545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.07902469485998154 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34874221682548523 + ], + "opposite_logprob_tokens": [ + -1.2237422466278076 + ], + "logprob_reward": -0.34874221682548523, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34874221682548523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4010196030139923, + 0.0856369137763977, + 0.04769014194607735, + 0.05070195719599724, + 0.17847128212451935, + 0.07795337587594986, + 0.03387439623475075, + 0.02725036069750786, + 0.03331838920712471, + 0.06408361345529556 + ], + "credit_map": [ + 0.04044825956225395, + 0.011085332371294498, + 0.1661519557237625, + 0.08722905814647675, + 0.014157187193632126, + 0.022310402244329453, + 0.0355183482170105, + 0.024119054898619652, + 0.009775698184967041, + 0.009893685579299927, + 0.015191995538771152, + 0.01359567791223526, + 0.006186190061271191, + 0.00681580975651741, + 0.013406675308942795, + 0.010771180503070354, + 0.011602225713431835, + 0.0025663557462394238, + 0.008969293907284737, + 0.005313978064805269, + 0.003362554358318448, + 0.007271073758602142, + 0.005036613438278437, + 0.0035680511500686407, + 0.006926650181412697, + 0.004525382537394762, + 0.009126648306846619, + 0.007925041019916534, + 0.0046976893208920956, + 0.003848073771223426, + 0.006809016223996878, + 0.006843454670161009, + 0.024345431476831436, + 0.01681721955537796, + 0.027668412774801254, + 0.02861016057431698, + 0.019087927415966988, + 0.01337874960154295, + 0.02279510907828808, + 0.025768261402845383, + 0.012722994200885296, + 0.00656424043700099, + 0.010310458019375801, + 0.008525103330612183, + 0.006146853789687157, + 0.006832494400441647, + 0.008621313609182835, + 0.018229912966489792, + 0.0034106713719666004, + 0.0028247598093003035, + 0.005149825010448694, + 0.005768635310232639, + 0.003325084690004587, + 0.001659788889810443, + 0.0052679176442325115, + 0.006467712577432394, + 0.0028278992976993322, + 0.0018869127379730344, + 0.00607589865103364, + 0.0041538747027516365, + 0.0015770155005156994, + 0.0010734295938163996, + 0.002457248279824853, + 0.007198084145784378, + 0.0050096712075173855, + 0.003210886614397168, + 0.003036943031474948, + 0.004179615527391434, + 0.00399490911513567, + 0.00519307516515255, + 0.004090021830052137, + 0.004603267647325993, + 0.008605442009866238, + 0.004668064881116152, + 0.008733484894037247, + 0.006617322098463774, + 0.0059827337972819805, + 0.009792434982955456, + 0.006192934233695269, + 0.013491190038621426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.34874221682548523 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991863191127777 + ], + "opposite_logprob_tokens": [ + -1.804918646812439 + ], + "logprob_reward": -0.17991863191127777, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991863191127777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2443002462387085, + 0.18482686579227448, + 0.1634463220834732, + 0.0388154461979866, + 0.11542476713657379, + 0.05354400351643562, + 0.07418100535869598, + 0.04318578168749809, + 0.05545340105891228, + 0.026822227984666824 + ], + "credit_map": [ + 0.02809958904981613, + 0.012408990412950516, + 0.07439932972192764, + 0.0409616120159626, + 0.017951536923646927, + 0.01074453629553318, + 0.037240102887153625, + 0.022494560107588768, + 0.036309368908405304, + 0.014887531287968159, + 0.020793378353118896, + 0.01737750507891178, + 0.009963673539459705, + 0.03808747977018356, + 0.021692898124456406, + 0.025715021416544914, + 0.040777161717414856, + 0.007245303597301245, + 0.030579840764403343, + 0.015056156553328037, + 0.007928525097668171, + 0.0308948066085577, + 0.015951503068208694, + 0.015013009309768677, + 0.005729861557483673, + 0.0024448204785585403, + 0.0075740390457212925, + 0.00524888327345252, + 0.002691947855055332, + 0.004545326344668865, + 0.004002536181360483, + 0.006578029599040747, + 0.02605091594159603, + 0.006533889099955559, + 0.018924366682767868, + 0.01158562395721674, + 0.006235165987163782, + 0.026888640597462654, + 0.0073094540275633335, + 0.011896707117557526, + 0.010034272447228432, + 0.004632171243429184, + 0.005566572770476341, + 0.011970906518399715, + 0.0034464180935174227, + 0.0031856114510446787, + 0.004468384198844433, + 0.010239662602543831, + 0.006696281023323536, + 0.00558868981897831, + 0.020146798342466354, + 0.009001306258141994, + 0.0053119664080441, + 0.004214830696582794, + 0.006810308899730444, + 0.016410818323493004, + 0.004206893965601921, + 0.004034043755382299, + 0.00605363491922617, + 0.008431404829025269, + 0.003458332968875766, + 0.003940522205084562, + 0.007234777323901653, + 0.005826170090585947, + 0.007646596524864435, + 0.004726385697722435, + 0.005023187957704067, + 0.008182354271411896, + 0.005159963387995958, + 0.009328392334282398, + 0.005793310236185789, + 0.009593205526471138, + 0.0025979347992688417, + 0.002415542956441641, + 0.004299595952033997, + 0.0038605681620538235, + 0.0029770347755402327, + 0.003190557472407818, + 0.0033822148106992245, + 0.004098775330930948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.17991863191127777 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, at eye level, frames the scene, the camera panning slowly and steadily across the face of a grieving adult African-American woman. She's in a park on a blustery day, the wind whipping her long, dark, intricately braided hair, some of which hang down against her cheeks. She wears a simple, dark charcoal-colored hoodie, the soft fabric pulled up at her neck. The background is a beautiful, out-of-focus bokeh of autumn leaves, soft swirls of gold and brown under the flat, diffuse light of an overcast sky. Her eyes are deep and expressive, lowered and furrowed, as she reads something off-screen, revealing a profound sadness. As the camera pans across her features, a perfectly formed tear escapes from the corner of her eye, slowly and deliberately running down her smooth cheek, glistening. Her lips, slightly parted, tremble almost imperceptibly, creating a powerful, intimate, and melancholic atmosphere of quiet heartbreak.", + "seed": 25703, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_1/index_8497_seed_25703_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2391665805131197, + "vlm_logprob": -0.2391665805131197, + "vlm_yes_no_margin": 2.125, + "vlm_token_credit": -0.2391665805131197, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02406168542802334 + ], + "opposite_logprob_tokens": [ + -4.274061679840088 + ], + "logprob_reward": -0.02406168542802334, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02406168542802334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07325330376625061, + 0.0802496150135994, + 0.14809678494930267, + 0.03265568986535072, + 0.05507076159119606, + 0.15053807199001312, + 0.041054584085941315, + 0.29043927788734436, + 0.08098449558019638, + 0.04765739664435387 + ], + "credit_map": [ + 0.006438614334911108, + 0.004145918879657984, + 0.007564864121377468, + 0.006903926841914654, + 0.012835138477385044, + 0.005627633538097143, + 0.018725737929344177, + 0.011011469177901745, + 0.014511377550661564, + 0.0025851053651422262, + 0.0061554983258247375, + 0.005878669209778309, + 0.007207600399851799, + 0.003124801442027092, + 0.01418793760240078, + 0.026598621159791946, + 0.01270539965480566, + 0.009849006310105324, + 0.026585325598716736, + 0.019670426845550537, + 0.014858073554933071, + 0.017349109053611755, + 0.0254733394831419, + 0.021606093272566795, + 0.003684069262817502, + 0.001792612369172275, + 0.0045608943328261375, + 0.0027298794593662024, + 0.004441878758370876, + 0.001987459836527705, + 0.007104854099452496, + 0.006354041397571564, + 0.005440857727080584, + 0.0031559430062770844, + 0.010043499059975147, + 0.0077088214457035065, + 0.005100504495203495, + 0.007580237928777933, + 0.008615813218057156, + 0.007425080519169569, + 0.017396355047822, + 0.00575968949124217, + 0.04404577985405922, + 0.010873774997889996, + 0.011240968480706215, + 0.03085380047559738, + 0.022447658702731133, + 0.007920040749013424, + 0.0041526793502271175, + 0.0017063039122149348, + 0.012098426930606365, + 0.005168473348021507, + 0.005680747330188751, + 0.0015034450916573405, + 0.007772558834403753, + 0.0029719483572989702, + 0.052434466779232025, + 0.011853068135678768, + 0.03497418761253357, + 0.01241588220000267, + 0.019318750128149986, + 0.08989495038986206, + 0.03002915531396866, + 0.039518803358078, + 0.007656386122107506, + 0.008209073916077614, + 0.012060112319886684, + 0.014766090549528599, + 0.01261193212121725, + 0.0053805699571967125, + 0.013109474442899227, + 0.0071908533573150635, + 0.004837624728679657, + 0.005762284155935049, + 0.007150566205382347, + 0.0063430811278522015, + 0.007135325577110052, + 0.003186871064826846, + 0.00814000703394413, + 0.005101640708744526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.02406168542802334 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018219847232103348 + ], + "opposite_logprob_tokens": [ + -4.018219947814941 + ], + "logprob_reward": -0.018219847232103348, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018219847232103348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14694732427597046, + 0.2474508136510849, + 0.07295715808868408, + 0.05279882252216339, + 0.052104149013757706, + 0.09384322911500931, + 0.11583975702524185, + 0.06610256433486938, + 0.08889152109622955, + 0.06306468695402145 + ], + "credit_map": [ + 0.012818418443202972, + 0.007031225133687258, + 0.025047238916158676, + 0.02134559489786625, + 0.019262628629803658, + 0.014289677143096924, + 0.027038022875785828, + 0.020114531740546227, + 0.031518761068582535, + 0.015590041875839233, + 0.022631153464317322, + 0.018648192286491394, + 0.03741385415196419, + 0.012371535412967205, + 0.04243076220154762, + 0.06684651970863342, + 0.0062505449168384075, + 0.0038325509522110224, + 0.016240805387496948, + 0.007071655709296465, + 0.008642151020467281, + 0.010151302441954613, + 0.012652570381760597, + 0.008115575648844242, + 0.005580518860369921, + 0.003853631904348731, + 0.0065802824683487415, + 0.003157886676490307, + 0.010224913246929646, + 0.005635232664644718, + 0.009995209984481335, + 0.0077711446210742, + 0.0037463181652128696, + 0.0035543160047382116, + 0.010179700329899788, + 0.006800111383199692, + 0.004549237433820963, + 0.005884653888642788, + 0.0064803222194314, + 0.010909485630691051, + 0.012268817983567715, + 0.005030561238527298, + 0.027807679027318954, + 0.008443288505077362, + 0.015021583996713161, + 0.004328306298702955, + 0.015512924641370773, + 0.005430062301456928, + 0.01478597242385149, + 0.004618248902261257, + 0.03427519649267197, + 0.014045000076293945, + 0.013284249231219292, + 0.004212718922644854, + 0.02123355120420456, + 0.009384815581142902, + 0.006633528973907232, + 0.0028231805190443993, + 0.018824025988578796, + 0.005381184630095959, + 0.008048608899116516, + 0.005641283467411995, + 0.010195193812251091, + 0.008555559441447258, + 0.008604207076132298, + 0.009139640256762505, + 0.012890138663351536, + 0.01567605510354042, + 0.015293606556952, + 0.006561928428709507, + 0.01387059222906828, + 0.0068553537130355835, + 0.008324118331074715, + 0.008251747116446495, + 0.007541252300143242, + 0.006331166718155146, + 0.010338294319808483, + 0.005210285075008869, + 0.011338561773300171, + 0.005729258991777897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.018219847232103348 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3487314283847809 + ], + "opposite_logprob_tokens": [ + -1.2237313985824585 + ], + "logprob_reward": -0.3487314283847809, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3487314283847809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14806801080703735, + 0.09138871729373932, + 0.1140173152089119, + 0.04898637905716896, + 0.08530087769031525, + 0.055581118911504745, + 0.07736901193857193, + 0.259073942899704, + 0.08006041496992111, + 0.04015420749783516 + ], + "credit_map": [ + 0.018502773717045784, + 0.007460256107151508, + 0.024719957262277603, + 0.019180551171302795, + 0.018318235874176025, + 0.011789427138864994, + 0.027373965829610825, + 0.020722845569252968, + 0.011716758832335472, + 0.0036056297831237316, + 0.013834169134497643, + 0.0063387928530573845, + 0.016791285946965218, + 0.0035793583374470472, + 0.014600376598536968, + 0.020922349765896797, + 0.012027537450194359, + 0.006260210182517767, + 0.012064427137374878, + 0.009709581732749939, + 0.013979635201394558, + 0.01935500279068947, + 0.023657798767089844, + 0.016963129863142967, + 0.005560960154980421, + 0.0028732300270348787, + 0.007901625707745552, + 0.004227285273373127, + 0.00960827898234129, + 0.002612797077745199, + 0.009411399252712727, + 0.006790809333324432, + 0.007933517917990685, + 0.00400133803486824, + 0.02221856266260147, + 0.009648970328271389, + 0.005958386696875095, + 0.0071556479670107365, + 0.009941799566149712, + 0.018442656844854355, + 0.007046841084957123, + 0.004100911319255829, + 0.0061023603193461895, + 0.00514482194557786, + 0.012259487994015217, + 0.005684728268533945, + 0.010529492050409317, + 0.004712479654699564, + 0.008508529514074326, + 0.004097666125744581, + 0.011758202686905861, + 0.011397447437047958, + 0.01317492127418518, + 0.0034924251958727837, + 0.01724250055849552, + 0.007697327062487602, + 0.038520872592926025, + 0.00836168322712183, + 0.06692417711019516, + 0.017039859667420387, + 0.015612761490046978, + 0.042124588042497635, + 0.03719170019030571, + 0.033298324793577194, + 0.005194599274545908, + 0.011798455379903316, + 0.008500282652676105, + 0.012868438847362995, + 0.014716323465108871, + 0.005689733661711216, + 0.012832117266952991, + 0.00846046768128872, + 0.0032437669578939676, + 0.006860936060547829, + 0.005197460763156414, + 0.0062085543759167194, + 0.005441385321319103, + 0.0029580523259937763, + 0.006482278928160667, + 0.003761772997677326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.3487314283847809 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234525203704834 + ], + "opposite_logprob_tokens": [ + -0.8984525203704834 + ], + "logprob_reward": -0.5234525203704834, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234525203704834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12360724806785583, + 0.05954597517848015, + 0.11157417297363281, + 0.04735967889428139, + 0.06363619863986969, + 0.10214676707983017, + 0.04098448529839516, + 0.34282761812210083, + 0.07476399838924408, + 0.0335538350045681 + ], + "credit_map": [ + 0.018177306279540062, + 0.00692881690338254, + 0.021716242656111717, + 0.013300689868628979, + 0.015272809192538261, + 0.007340519223362207, + 0.021534407511353493, + 0.019336452707648277, + 0.007827762514352798, + 0.005167003255337477, + 0.007938733324408531, + 0.005638015456497669, + 0.01123038586229086, + 0.0044248332269489765, + 0.008469785563647747, + 0.008849452249705791, + 0.01032007485628128, + 0.006194358225911856, + 0.01280434150248766, + 0.010205031372606754, + 0.012662303633987904, + 0.013370400294661522, + 0.03210911527276039, + 0.013908551074564457, + 0.006016057450324297, + 0.0025270909536629915, + 0.006700754165649414, + 0.004305190872400999, + 0.006907677743583918, + 0.0022922647185623646, + 0.008811184205114841, + 0.009799458086490631, + 0.005311194341629744, + 0.0029573841020464897, + 0.012913340702652931, + 0.008043801411986351, + 0.004861446563154459, + 0.0030319704674184322, + 0.010265697725117207, + 0.016251355409622192, + 0.014676795341074467, + 0.006218720693141222, + 0.021732408553361893, + 0.009729468263685703, + 0.01821502298116684, + 0.007831832394003868, + 0.016568729653954506, + 0.007173786871135235, + 0.003347370307892561, + 0.002062649466097355, + 0.006451983004808426, + 0.009150877594947815, + 0.006328019313514233, + 0.0016520705539733171, + 0.008855684660375118, + 0.0031358322594314814, + 0.05710519850254059, + 0.01564553938806057, + 0.037351351231336594, + 0.012865938246250153, + 0.024854157119989395, + 0.11798540502786636, + 0.03020344488322735, + 0.04681656137108803, + 0.005217055790126324, + 0.012219513766467571, + 0.008599681779742241, + 0.008877997286617756, + 0.015158697962760925, + 0.00615618284791708, + 0.011357076466083527, + 0.0071777887642383575, + 0.0039403061382472515, + 0.004113491624593735, + 0.004729433450847864, + 0.0059038070030510426, + 0.0032717587891966105, + 0.002364658983424306, + 0.005366671830415726, + 0.0038637053221464157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.5234525203704834 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813674211502075 + ], + "opposite_logprob_tokens": [ + -1.4063674211502075 + ], + "logprob_reward": -0.2813674211502075, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813674211502075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2061876803636551, + 0.056337323039770126, + 0.22668325901031494, + 0.013913155533373356, + 0.026959389448165894, + 0.028293246403336525, + 0.040780868381261826, + 0.3611111044883728, + 0.024762457236647606, + 0.014971512369811535 + ], + "credit_map": [ + 0.015446037985384464, + 0.012842548079788685, + 0.02138352021574974, + 0.029932847246527672, + 0.040091730654239655, + 0.028272761031985283, + 0.03135151416063309, + 0.026866724714636803, + 0.006751399952918291, + 0.0027431531343609095, + 0.007338766008615494, + 0.003926640376448631, + 0.01083958987146616, + 0.0029531170148402452, + 0.008891316130757332, + 0.01289333961904049, + 0.0215320885181427, + 0.01342196948826313, + 0.02984442561864853, + 0.01835283823311329, + 0.026693148538470268, + 0.04082633927464485, + 0.04822542145848274, + 0.02778700739145279, + 0.0011600060388445854, + 0.000873653800226748, + 0.002431244356557727, + 0.0011594961397349834, + 0.0023729437962174416, + 0.0009341678814962506, + 0.002489720704033971, + 0.002491922816261649, + 0.0029646940529346466, + 0.0018247843254357576, + 0.004681612364947796, + 0.002520020119845867, + 0.0034865119960159063, + 0.002625810680910945, + 0.005757639184594154, + 0.0030983155593276024, + 0.0024166395887732506, + 0.0026165673043578863, + 0.004118729382753372, + 0.0030899534467607737, + 0.007209735456854105, + 0.0015388812171295285, + 0.005369837395846844, + 0.0019329031929373741, + 0.00549194635823369, + 0.0013578847283497453, + 0.014329196885228157, + 0.005055589135736227, + 0.002764047123491764, + 0.0017253648256883025, + 0.0071784984320402145, + 0.0028783397283405066, + 0.037495531141757965, + 0.017179839313030243, + 0.110389344394207, + 0.02438497543334961, + 0.029830319806933403, + 0.06735096126794815, + 0.049190931022167206, + 0.02528921701014042, + 0.001883752178400755, + 0.003360804170370102, + 0.00277485023252666, + 0.003992028068751097, + 0.002992077264934778, + 0.0019126059487462044, + 0.005006188061088324, + 0.00284015154466033, + 0.0015872627263888717, + 0.0020927726291120052, + 0.0025302786380052567, + 0.002617606893181801, + 0.001896355301141739, + 0.0010142751270905137, + 0.002243628492578864, + 0.0009893320966511965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.2813674211502075 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, at eye level, frames the scene, the camera panning slowly and steadily across the face of a grieving adult African-American woman. She's in a park on a blustery day, the wind whipping her long, dark, intricately braided hair, some of which hang down against her cheeks. She wears a simple, dark charcoal-colored hoodie, the soft fabric pulled up at her neck. The background is a beautiful, out-of-focus bokeh of autumn leaves, soft swirls of gold and brown under the flat, diffuse light of an overcast sky. Her eyes are deep and expressive, lowered and furrowed, as she reads something off-screen, revealing a profound sadness. As the camera pans across her features, a perfectly formed tear escapes from the corner of her eye, slowly and deliberately running down her smooth cheek, glistening. Her lips, slightly parted, tremble almost imperceptibly, creating a powerful, intimate, and melancholic atmosphere of quiet heartbreak.", + "seed": 25704, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_1/index_8497_seed_25704_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12299786657094955, + "vlm_logprob": -0.12299786657094955, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.12299786657094955, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010875476524233818 + ], + "opposite_logprob_tokens": [ + -5.510875701904297 + ], + "logprob_reward": -0.010875476524233818, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.010875476524233818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08548452705144882, + 0.07853735983371735, + 0.08516354113817215, + 0.27913689613342285, + 0.05029747262597084, + 0.07928527146577835, + 0.04124743491411209, + 0.06858812272548676, + 0.11827747523784637, + 0.1139819473028183 + ], + "credit_map": [ + 0.007922179065644741, + 0.010992476716637611, + 0.013275743462145329, + 0.010890878736972809, + 0.008149780333042145, + 0.006587744224816561, + 0.011917298659682274, + 0.01574842445552349, + 0.009284727275371552, + 0.006634337827563286, + 0.02374916896224022, + 0.009720989502966404, + 0.0046104504726827145, + 0.004191590938717127, + 0.010277720168232918, + 0.010068370960652828, + 0.010979722253978252, + 0.021590285003185272, + 0.01789758913218975, + 0.008706534281373024, + 0.006267393007874489, + 0.0030909047927707434, + 0.008543172851204872, + 0.008087928406894207, + 0.03731667250394821, + 0.061988841742277145, + 0.04601234570145607, + 0.027762386947870255, + 0.030678139999508858, + 0.011084336787462234, + 0.03793877363204956, + 0.026355357840657234, + 0.004818066954612732, + 0.0052925823256373405, + 0.008171074092388153, + 0.0070053692907094955, + 0.006004520691931248, + 0.0030971807427704334, + 0.009884589351713657, + 0.006024084985256195, + 0.006455199792981148, + 0.005988382734358311, + 0.008768164552748203, + 0.009268371388316154, + 0.007478388492017984, + 0.010419880039989948, + 0.013993139378726482, + 0.016913743689656258, + 0.007733610924333334, + 0.003588467137888074, + 0.007147463038563728, + 0.004322282504290342, + 0.0033821361139416695, + 0.003979830536991358, + 0.006783911492675543, + 0.004309727344661951, + 0.01102948747575283, + 0.006015168968588114, + 0.009357859380543232, + 0.008621527813374996, + 0.005659426096826792, + 0.0032127571757882833, + 0.010554526932537556, + 0.01413736306130886, + 0.012475533410906792, + 0.006955313961952925, + 0.010792536661028862, + 0.016097746789455414, + 0.018799815326929092, + 0.00786609947681427, + 0.030714236199855804, + 0.01457618921995163, + 0.01066186185926199, + 0.01455382164567709, + 0.009664063341915607, + 0.033882297575473785, + 0.00781223876401782, + 0.006329596508294344, + 0.0136986980214715, + 0.01737937144935131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.010875476524233818 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029829563573002815 + ], + "opposite_logprob_tokens": [ + -3.529829502105713 + ], + "logprob_reward": -0.029829563573002815, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029829563573002815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046613022685050964, + 0.07344943284988403, + 0.08515925705432892, + 0.14878621697425842, + 0.05182650312781334, + 0.048869721591472626, + 0.019556064158678055, + 0.2004733830690384, + 0.09216584265232086, + 0.2331005334854126 + ], + "credit_map": [ + 0.004889531061053276, + 0.004577259998768568, + 0.006269730627536774, + 0.005793048068881035, + 0.005740283988416195, + 0.003935671877115965, + 0.0062984684482216835, + 0.009109033271670341, + 0.010208802297711372, + 0.010805380530655384, + 0.02297191135585308, + 0.00797417014837265, + 0.003947230521589518, + 0.0022598309442400932, + 0.008294417522847652, + 0.006987693253904581, + 0.008171413093805313, + 0.014552539214491844, + 0.01170645747333765, + 0.009599229320883751, + 0.009827516973018646, + 0.006047842558473349, + 0.012206293642520905, + 0.013047979213297367, + 0.022695306688547134, + 0.026391128078103065, + 0.04315446689724922, + 0.01413615234196186, + 0.01459679938852787, + 0.005466333590447903, + 0.011993440799415112, + 0.010352612473070621, + 0.004098572768270969, + 0.008617180399596691, + 0.007992255501449108, + 0.00882117822766304, + 0.007169974967837334, + 0.0026190667413175106, + 0.00774029828608036, + 0.004767977632582188, + 0.005378874018788338, + 0.006435976829379797, + 0.008707087486982346, + 0.004917946178466082, + 0.0043978639878332615, + 0.004617696162313223, + 0.007940600626170635, + 0.00647368049249053, + 0.0023070673923939466, + 0.001915812841616571, + 0.0029281959868967533, + 0.001959350658580661, + 0.0020286827348172665, + 0.0018165898509323597, + 0.004112202674150467, + 0.0024881637655198574, + 0.03317532315850258, + 0.012366733513772488, + 0.02552882581949234, + 0.024229586124420166, + 0.010183161124587059, + 0.01159437745809555, + 0.014981288462877274, + 0.06841409206390381, + 0.009600513614714146, + 0.006175914779305458, + 0.006724619772285223, + 0.012045437470078468, + 0.013006354682147503, + 0.012594820000231266, + 0.02307889051735401, + 0.008939296007156372, + 0.029254963621497154, + 0.028543097898364067, + 0.0606699213385582, + 0.0327741801738739, + 0.018079223111271858, + 0.01701151765882969, + 0.02332616038620472, + 0.023441482335329056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.029829563573002815 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297919601202011 + ], + "opposite_logprob_tokens": [ + -2.2379791736602783 + ], + "logprob_reward": -0.11297919601202011, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297919601202011, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06863725185394287, + 0.1097307875752449, + 0.048887699842453, + 0.1348934918642044, + 0.22299911081790924, + 0.044817887246608734, + 0.052705068141222, + 0.114168182015419, + 0.15973471105098724, + 0.043425772339105606 + ], + "credit_map": [ + 0.007578045595437288, + 0.009839005768299103, + 0.011048517189919949, + 0.00737046217545867, + 0.007097295019775629, + 0.00439049955457449, + 0.00910241436213255, + 0.012211007066071033, + 0.01147102378308773, + 0.020007649436593056, + 0.0187210775911808, + 0.014616604894399643, + 0.011178994551301003, + 0.007622268050909042, + 0.012084903195500374, + 0.014028266072273254, + 0.005061621777713299, + 0.008685644716024399, + 0.011857470497488976, + 0.004878124687820673, + 0.004510900937020779, + 0.002894825069233775, + 0.00556531036272645, + 0.00543380668386817, + 0.020724887028336525, + 0.01985732652246952, + 0.045128148049116135, + 0.0131090572103858, + 0.010931783355772495, + 0.004227352794259787, + 0.010349907912313938, + 0.010565032251179218, + 0.018011299893260002, + 0.03752532973885536, + 0.03221502527594566, + 0.03755563870072365, + 0.03248247876763344, + 0.011873715557157993, + 0.03143135830760002, + 0.021904248744249344, + 0.004145314916968346, + 0.0034822607412934303, + 0.005182900000363588, + 0.0058796764351427555, + 0.0043199555948376656, + 0.004556501749902964, + 0.008548423647880554, + 0.008702854625880718, + 0.0059129162691533566, + 0.005368921440094709, + 0.008633850142359734, + 0.005722450092434883, + 0.007645400241017342, + 0.0035177464596927166, + 0.009831261821091175, + 0.006072521675378084, + 0.024954069405794144, + 0.004285700619220734, + 0.014430055394768715, + 0.01219099573791027, + 0.0056772069074213505, + 0.009185352362692356, + 0.00926144327968359, + 0.034183356910943985, + 0.019855020567774773, + 0.009511570446193218, + 0.021697144955396652, + 0.018701601773500443, + 0.018516115844249725, + 0.02395782433450222, + 0.027812141925096512, + 0.019683292135596275, + 0.00596407987177372, + 0.005409684032201767, + 0.007104257587343454, + 0.008520533330738544, + 0.0038084411062300205, + 0.002809778554365039, + 0.0056188832968473434, + 0.004190115258097649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.11297919601202011 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17994432151317596 + ], + "opposite_logprob_tokens": [ + -1.8049442768096924 + ], + "logprob_reward": -0.17994432151317596, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17994432151317596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08920122683048248, + 0.07382816076278687, + 0.0676683858036995, + 0.0865112692117691, + 0.07421299070119858, + 0.24525754153728485, + 0.06406482309103012, + 0.043688029050827026, + 0.20544207096099854, + 0.05012552812695503 + ], + "credit_map": [ + 0.009760531596839428, + 0.011915243230760098, + 0.008695469237864017, + 0.014217948541045189, + 0.01112770289182663, + 0.006457176059484482, + 0.012454310432076454, + 0.014572842977941036, + 0.0070882802829146385, + 0.006808789446949959, + 0.014681367203593254, + 0.009087545797228813, + 0.00479359645396471, + 0.0046110306866467, + 0.010264836251735687, + 0.016492711380124092, + 0.006146515253931284, + 0.010523419827222824, + 0.009587203152477741, + 0.009061300195753574, + 0.008837702684104443, + 0.003925972618162632, + 0.008240691386163235, + 0.011345581151545048, + 0.012748499400913715, + 0.009970398619771004, + 0.033224932849407196, + 0.008717193268239498, + 0.005650578998029232, + 0.002950046444311738, + 0.006555838510394096, + 0.006693777162581682, + 0.00664669182151556, + 0.008359945379197598, + 0.01219311822205782, + 0.012520469725131989, + 0.01226937584578991, + 0.004051626194268465, + 0.010134542360901833, + 0.00803722720593214, + 0.02156917192041874, + 0.013824239373207092, + 0.022764211520552635, + 0.03125731647014618, + 0.02506554126739502, + 0.027849230915308, + 0.0461401641368866, + 0.0567876473069191, + 0.004930678755044937, + 0.004148196429014206, + 0.00717049278318882, + 0.006915070116519928, + 0.008101259358227253, + 0.007116406690329313, + 0.015829453244805336, + 0.009853266179561615, + 0.005778546445071697, + 0.0018648653058335185, + 0.006230502855032682, + 0.005285578314214945, + 0.003998999949544668, + 0.002295529004186392, + 0.005144782830029726, + 0.013089224696159363, + 0.02489699050784111, + 0.011744524352252483, + 0.017856065183877945, + 0.02928110398352146, + 0.029011983424425125, + 0.0076355054043233395, + 0.056632913649082184, + 0.02838299050927162, + 0.004822241608053446, + 0.0061859204433858395, + 0.0050261239521205425, + 0.01403042022138834, + 0.004076588433235884, + 0.0022229261230677366, + 0.006250741425901651, + 0.007510566618293524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.17994432151317596 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult African-American woman's face at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark charcoal-colored hoodie with the soft fabric pulled up at her neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's long, dark, intricately braided hair being whipped by wind against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tear escape from the corner of her eye and run down her cheek while she reads something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly and steadily across her face against a background of out-of-focus autumn leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28136077523231506 + ], + "opposite_logprob_tokens": [ + -1.4063607454299927 + ], + "logprob_reward": -0.28136077523231506, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28136077523231506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10029860585927963, + 0.08339112997055054, + 0.04650735855102539, + 0.2506544888019562, + 0.12530505657196045, + 0.10974257439374924, + 0.028335966169834137, + 0.17455092072486877, + 0.04396197944879532, + 0.03725191950798035 + ], + "credit_map": [ + 0.010994425974786282, + 0.021026981994509697, + 0.016469288617372513, + 0.011919021606445312, + 0.007197038736194372, + 0.005621197633445263, + 0.013206579722464085, + 0.013864072039723396, + 0.005304185673594475, + 0.00983193889260292, + 0.010302449576556683, + 0.01296405028551817, + 0.010534549131989479, + 0.005517188459634781, + 0.011392049491405487, + 0.017544712871313095, + 0.004181734751909971, + 0.004389539361000061, + 0.006905128713697195, + 0.005932727828621864, + 0.004318106919527054, + 0.0031959302723407745, + 0.007719957269728184, + 0.009864234365522861, + 0.03716592118144035, + 0.06146441027522087, + 0.029656607657670975, + 0.023611847311258316, + 0.03699640929698944, + 0.011803636327385902, + 0.023890163749456406, + 0.026065481826663017, + 0.008747639134526253, + 0.012574763968586922, + 0.020273057743906975, + 0.0187588632106781, + 0.01953798159956932, + 0.007733201142400503, + 0.023830829188227654, + 0.013848723843693733, + 0.007821203209459782, + 0.006367966532707214, + 0.008312064222991467, + 0.014034576714038849, + 0.011276323348283768, + 0.0076490468345582485, + 0.029960209503769875, + 0.02432117983698845, + 0.004514527972787619, + 0.0027804607525467873, + 0.003312947228550911, + 0.0028948381077498198, + 0.003791999304667115, + 0.002146160928532481, + 0.005855008959770203, + 0.0030400236137211323, + 0.039781153202056885, + 0.006665399763733149, + 0.02366594970226288, + 0.020172839984297752, + 0.007467696908861399, + 0.013075767084956169, + 0.010871916078031063, + 0.052850205451250076, + 0.007163693197071552, + 0.0027798330411314964, + 0.003948295023292303, + 0.004885357338935137, + 0.00417530769482255, + 0.005556497722864151, + 0.008065951988101006, + 0.0073870401829481125, + 0.00470975274220109, + 0.004155389033257961, + 0.0058422996662557125, + 0.005035513546317816, + 0.004230515565723181, + 0.0037538856267929077, + 0.00532700028270483, + 0.004197563044726849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.28136077523231506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_10/index_tensor([5447, 5447])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_10/index_tensor([5447, 5447])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..30a975dd75e6b75334cdd98bd72e716059983cd4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_10/index_tensor([5447, 5447])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a striking black-and-white high-dynamic range (HDR) image, an underwater close-up shot, shot from a slightly upward perspective, sways gently yet steadily in time with the rhythm of the dark current. At the center of the frame is a massive swimming crab. Illuminated by a single, intense light source from above, its carapace reveals an intricate relief of pure white highlights and deep shadows, each spike and vein striking with breathtaking sharpness. Its signature flat claws, like a pair of meticulously polished obsidian oars, gleam with the hard luster of ceramic, their jagged inner edges auguring destructive power. The giant crab secures a strand of tough seaweed with its smaller appendages. Then, with a slow, demonstrative flourish of one of its massive claws, the movement accelerates, snapping down with swift, decisive force, sinking its teeth into the seaweed's midsection. The exquisite detail allows viewers to practically hear the sharp snap of seaweed fibers being cut, the severed strands then drifting feebly and slowly into the endless darkness beyond the camera. The entire scene, rendered in stark black and white, creates a primal, cold, yet powerful atmosphere, perfectly showcasing the astonishing strength and awe-inspiring biomechanical beauty of this underwater predator.", + "seed": 16207, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_10/index_5447_seed_16207_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05687188655138016, + "vlm_logprob": -0.05687188655138016, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.05687188655138016, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04134504497051239 + ], + "opposite_logprob_tokens": [ + -3.2913451194763184 + ], + "logprob_reward": -0.04134504497051239, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04134504497051239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12201003730297089, + 0.07405445724725723, + 0.1103755459189415, + 0.10988961160182953, + 0.05179750919342041, + 0.3259396553039551, + 0.03622281178832054, + 0.0194314643740654, + 0.11762720346450806, + 0.03265174850821495 + ], + "credit_map": [ + 0.021301476284861565, + 0.013698450289666653, + 0.011722260154783726, + 0.012426944449543953, + 0.01774129457771778, + 0.008100396953523159, + 0.0213015116751194, + 0.015717700123786926, + 0.01292219664901495, + 0.007904397323727608, + 0.005656251683831215, + 0.009740100242197514, + 0.0076240901835262775, + 0.0043250201269984245, + 0.01685161143541336, + 0.009030792862176895, + 0.020474039018154144, + 0.009861817583441734, + 0.008685327135026455, + 0.011990277096629143, + 0.005882789380848408, + 0.00568047072738409, + 0.03768830746412277, + 0.010112511925399303, + 0.031823813915252686, + 0.010146573185920715, + 0.009819185361266136, + 0.008932298980653286, + 0.02200043573975563, + 0.006496825255453587, + 0.010377482511103153, + 0.010292996652424335, + 0.0072349077090620995, + 0.005060776602476835, + 0.005668135825544596, + 0.008528480306267738, + 0.005397072061896324, + 0.005510550457984209, + 0.009728115051984787, + 0.004669472109526396, + 0.08405718207359314, + 0.023396598175168037, + 0.013851962052285671, + 0.13809949159622192, + 0.011304568499326706, + 0.007350888103246689, + 0.025306839495897293, + 0.022572115063667297, + 0.006614005658775568, + 0.003710379358381033, + 0.0034512451384216547, + 0.003315563313663006, + 0.0030954524409025908, + 0.0028792894445359707, + 0.004599896725267172, + 0.00855698250234127, + 0.0031002231407910585, + 0.0028990942519158125, + 0.002201448194682598, + 0.002622493077069521, + 0.0018675520550459623, + 0.0015542736509814858, + 0.0026456587947905064, + 0.00254072155803442, + 0.018977291882038116, + 0.015811949968338013, + 0.0188981331884861, + 0.017229991033673286, + 0.008142688311636448, + 0.00993838720023632, + 0.01534751895815134, + 0.013281242921948433, + 0.005486520938575268, + 0.003315144684165716, + 0.0036335899494588375, + 0.004366885405033827, + 0.0028824324253946543, + 0.0035330690443515778, + 0.0042576841078698635, + 0.005176421254873276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.04134504497051239 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01113203726708889 + ], + "opposite_logprob_tokens": [ + -4.51113224029541 + ], + "logprob_reward": -0.01113203726708889, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01113203726708889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09153232723474503, + 0.06371605396270752, + 0.17552180588245392, + 0.04681070148944855, + 0.16489684581756592, + 0.049352776259183884, + 0.06634081900119781, + 0.2483048439025879, + 0.06532132625579834, + 0.02820252813398838 + ], + "credit_map": [ + 0.01569494791328907, + 0.009473780170083046, + 0.010076184757053852, + 0.01305854506790638, + 0.009751067496836185, + 0.008134076371788979, + 0.013659977354109287, + 0.011683749966323376, + 0.011252310127019882, + 0.008866013027727604, + 0.008660486899316311, + 0.008974631316959858, + 0.005569084081798792, + 0.005054468289017677, + 0.008393548429012299, + 0.006945515517145395, + 0.03473510220646858, + 0.012496139854192734, + 0.013036126270890236, + 0.015726789832115173, + 0.008962289430201054, + 0.010600348003208637, + 0.059176359325647354, + 0.020788652822375298, + 0.006531811784952879, + 0.00426983879879117, + 0.005154580809175968, + 0.004199891816824675, + 0.004813310690224171, + 0.005870667286217213, + 0.00614812271669507, + 0.009822475723922253, + 0.0222966019064188, + 0.01962277851998806, + 0.016123652458190918, + 0.03401198238134384, + 0.016346454620361328, + 0.014313088729977608, + 0.023030651733279228, + 0.019151628017425537, + 0.00800988171249628, + 0.005896975286304951, + 0.005699744448065758, + 0.009054279886186123, + 0.00475638173520565, + 0.003927494399249554, + 0.006718126125633717, + 0.005289891269057989, + 0.012116111814975739, + 0.0056638834066689014, + 0.004509844351559877, + 0.021864313632249832, + 0.004193942528218031, + 0.0028401443269103765, + 0.006756233517080545, + 0.008396351709961891, + 0.054245736449956894, + 0.02171497978270054, + 0.021862030029296875, + 0.07401690632104874, + 0.015227938070893288, + 0.011383860372006893, + 0.024110322818160057, + 0.025743065401911736, + 0.009361001662909985, + 0.009867371059954166, + 0.009075605310499668, + 0.010893805883824825, + 0.004320947453379631, + 0.005267476197332144, + 0.009534988552331924, + 0.0070001292042434216, + 0.003551039844751358, + 0.003272579750046134, + 0.002279652049764991, + 0.0031514635775238276, + 0.003976061940193176, + 0.0022398503497242928, + 0.004385158885270357, + 0.005346720106899738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.01113203726708889 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20160259306430817 + ], + "opposite_logprob_tokens": [ + -1.701602578163147 + ], + "logprob_reward": -0.20160259306430817, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20160259306430817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23899078369140625, + 0.10133221745491028, + 0.08286818861961365, + 0.09645584225654602, + 0.14122293889522552, + 0.15493614971637726, + 0.021192675456404686, + 0.02209656685590744, + 0.10023568570613861, + 0.04066895321011543 + ], + "credit_map": [ + 0.040313418954610825, + 0.021838827058672905, + 0.01936781033873558, + 0.038422591984272, + 0.028869658708572388, + 0.01884080469608307, + 0.038965146988630295, + 0.03237253427505493, + 0.016430724412202835, + 0.010682503692805767, + 0.007892854511737823, + 0.012118099257349968, + 0.010481514967978, + 0.008308181539177895, + 0.023460494354367256, + 0.011957842856645584, + 0.01611020229756832, + 0.0049188630655407906, + 0.005201661493629217, + 0.010098033584654331, + 0.0052435873076319695, + 0.0032296832650899887, + 0.029844004660844803, + 0.008222156204283237, + 0.029690051451325417, + 0.00886979978531599, + 0.007906980812549591, + 0.0060376389883458614, + 0.02032311074435711, + 0.005382890347391367, + 0.009053739719092846, + 0.00919162668287754, + 0.018593501299619675, + 0.015495682135224342, + 0.01203788910061121, + 0.024519190192222595, + 0.017124509438872337, + 0.010511228814721107, + 0.02861855924129486, + 0.0143223712220788, + 0.0388505645096302, + 0.011483972892165184, + 0.007379809394478798, + 0.06442330777645111, + 0.005281306803226471, + 0.004573907237499952, + 0.011923463083803654, + 0.011019826866686344, + 0.004420624580234289, + 0.0020500076934695244, + 0.0015928474022075534, + 0.003979108761996031, + 0.0016998344799503684, + 0.0019576631020754576, + 0.002264968352392316, + 0.00322762131690979, + 0.003596935886889696, + 0.003379618050530553, + 0.0031744090374559164, + 0.002669054549187422, + 0.0019003390334546566, + 0.0022100082132965326, + 0.002289393451064825, + 0.0028768074698746204, + 0.014251139014959335, + 0.016491569578647614, + 0.013328644447028637, + 0.014830084517598152, + 0.007768598850816488, + 0.00895855762064457, + 0.013240359723567963, + 0.011366733349859715, + 0.005622819997370243, + 0.004810981918126345, + 0.004179172683507204, + 0.004301333799958229, + 0.004793897736817598, + 0.003577249590307474, + 0.006409314461052418, + 0.0069741844199597836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.20160259306430817 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205407351255417 + ], + "opposite_logprob_tokens": [ + -4.264205455780029 + ], + "logprob_reward": -0.014205407351255417, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205407351255417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08715833723545074, + 0.04645630344748497, + 0.06976833194494247, + 0.2089620977640152, + 0.1093854233622551, + 0.14656518399715424, + 0.1741742491722107, + 0.03996971249580383, + 0.07674053311347961, + 0.04081980139017105 + ], + "credit_map": [ + 0.014360442757606506, + 0.00890150386840105, + 0.009495201520621777, + 0.009691310115158558, + 0.012850172817707062, + 0.0065703927539289, + 0.014372886158525944, + 0.010916434228420258, + 0.008171304129064083, + 0.0035964571870863438, + 0.004543313290923834, + 0.004968668334186077, + 0.008085102774202824, + 0.003864480182528496, + 0.008087702095508575, + 0.005139275919646025, + 0.011030509136617184, + 0.005244102329015732, + 0.006969406269490719, + 0.010321675799787045, + 0.004302052780985832, + 0.0034375442191958427, + 0.02225322276353836, + 0.006209822837263346, + 0.06810013204813004, + 0.01890815794467926, + 0.015726035460829735, + 0.013157511129975319, + 0.047415897250175476, + 0.011901190504431725, + 0.01800871454179287, + 0.015744460746645927, + 0.014735613018274307, + 0.0128400307148695, + 0.008130446076393127, + 0.019525229930877686, + 0.011214655824005604, + 0.008707123808562756, + 0.021960878744721413, + 0.01227144617587328, + 0.036429114639759064, + 0.011136853136122227, + 0.007870540022850037, + 0.05907755717635155, + 0.005741358734667301, + 0.003929757047444582, + 0.012515264563262463, + 0.00986473448574543, + 0.02499772049486637, + 0.0131567707285285, + 0.011002050712704659, + 0.0483454093337059, + 0.012737791053950787, + 0.010724029503762722, + 0.016408944502472878, + 0.0368015430867672, + 0.009262650273740292, + 0.0028541225474327803, + 0.002685428597033024, + 0.012858535163104534, + 0.0032521001994609833, + 0.0019975013565272093, + 0.0035444782115519047, + 0.003514898242428899, + 0.012136220000684261, + 0.00888594426214695, + 0.010788332670927048, + 0.009757455438375473, + 0.005696783307939768, + 0.008484511636197567, + 0.009383163414895535, + 0.011608119122684002, + 0.005273476243019104, + 0.005952662322670221, + 0.00352116534486413, + 0.0038239050190895796, + 0.0065915449522435665, + 0.00285883410833776, + 0.005654055159538984, + 0.007144156377762556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.014205407351255417 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016074350103735924 + ], + "opposite_logprob_tokens": [ + -4.141074180603027 + ], + "logprob_reward": -0.016074350103735924, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016074350103735924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15924102067947388, + 0.04668794572353363, + 0.11163730174303055, + 0.03894500061869621, + 0.1803048700094223, + 0.14854121208190918, + 0.1797952651977539, + 0.0370032973587513, + 0.04177067428827286, + 0.0560733824968338 + ], + "credit_map": [ + 0.02778138406574726, + 0.01460444275289774, + 0.012281987816095352, + 0.023360660299658775, + 0.02645653486251831, + 0.012196066789329052, + 0.023892967030405998, + 0.018666986376047134, + 0.006967250723391771, + 0.004811511840671301, + 0.005850514862686396, + 0.0068586235865950584, + 0.005288868676871061, + 0.004482370335608721, + 0.007133707404136658, + 0.00529509736225009, + 0.021855300292372704, + 0.007168653886765242, + 0.007444750051945448, + 0.012615001760423183, + 0.006638800725340843, + 0.004638063721358776, + 0.042498599737882614, + 0.008778130635619164, + 0.007896516472101212, + 0.004101844038814306, + 0.0038094264455139637, + 0.0037677143700420856, + 0.006329708267003298, + 0.0027119144797325134, + 0.005133192520588636, + 0.0051946863532066345, + 0.02546844258904457, + 0.021700501441955566, + 0.013558092527091503, + 0.029986534267663956, + 0.018908647820353508, + 0.015164191834628582, + 0.03743712604045868, + 0.01808132603764534, + 0.035301871597766876, + 0.011970160529017448, + 0.00995133351534605, + 0.055416036397218704, + 0.006611376069486141, + 0.005290026310831308, + 0.014156267046928406, + 0.009844145737588406, + 0.0431312620639801, + 0.01676834560930729, + 0.009582984261214733, + 0.06368757039308548, + 0.008963762782514095, + 0.006487745326012373, + 0.014992204494774342, + 0.016181403771042824, + 0.005895882844924927, + 0.0033922186121344566, + 0.003048434155061841, + 0.009364275261759758, + 0.005017018411308527, + 0.002320910571143031, + 0.0037821659352630377, + 0.004182393662631512, + 0.006104948464781046, + 0.006706086453050375, + 0.004726702347397804, + 0.006209979299455881, + 0.0029309606179594994, + 0.004005304537713528, + 0.006084257736802101, + 0.005002435762435198, + 0.00756119005382061, + 0.006967684719711542, + 0.004599302541464567, + 0.00457561481744051, + 0.008477749302983284, + 0.004848533775657415, + 0.007862282916903496, + 0.01118102390319109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.016074350103735924 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a striking black-and-white high-dynamic range (HDR) image, an underwater close-up shot, shot from a slightly upward perspective, sways gently yet steadily in time with the rhythm of the dark current. At the center of the frame is a massive swimming crab. Illuminated by a single, intense light source from above, its carapace reveals an intricate relief of pure white highlights and deep shadows, each spike and vein striking with breathtaking sharpness. Its signature flat claws, like a pair of meticulously polished obsidian oars, gleam with the hard luster of ceramic, their jagged inner edges auguring destructive power. The giant crab secures a strand of tough seaweed with its smaller appendages. Then, with a slow, demonstrative flourish of one of its massive claws, the movement accelerates, snapping down with swift, decisive force, sinking its teeth into the seaweed's midsection. The exquisite detail allows viewers to practically hear the sharp snap of seaweed fibers being cut, the severed strands then drifting feebly and slowly into the endless darkness beyond the camera. The entire scene, rendered in stark black and white, creates a primal, cold, yet powerful atmosphere, perfectly showcasing the astonishing strength and awe-inspiring biomechanical beauty of this underwater predator.", + "seed": 16208, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_10/index_5447_seed_16208_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05154131446033716, + "vlm_logprob": -0.05154131446033716, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.05154131446033716, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013299562968313694 + ], + "opposite_logprob_tokens": [ + -4.763299465179443 + ], + "logprob_reward": -0.013299562968313694, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.013299562968313694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.271274209022522, + 0.07557839900255203, + 0.06187259778380394, + 0.03950873389840126, + 0.05220334231853485, + 0.06851676851511002, + 0.038010697811841965, + 0.1375420093536377, + 0.19216717779636383, + 0.06332611292600632 + ], + "credit_map": [ + 0.02097853645682335, + 0.07705890387296677, + 0.021045299246907234, + 0.022088896483182907, + 0.02169264480471611, + 0.04588275030255318, + 0.01865719072520733, + 0.0438699796795845, + 0.006997422780841589, + 0.015013869851827621, + 0.00706490920856595, + 0.006156981457024813, + 0.008431315422058105, + 0.015385633334517479, + 0.00604470307007432, + 0.010483549907803535, + 0.008451291359961033, + 0.009066857397556305, + 0.0048222532495856285, + 0.005164347123354673, + 0.005530183669179678, + 0.011618787422776222, + 0.00601291935890913, + 0.011205952614545822, + 0.0030838493257761, + 0.0062323156744241714, + 0.00632741441950202, + 0.004258616827428341, + 0.004151442088186741, + 0.004791127052158117, + 0.005259010009467602, + 0.005404958501458168, + 0.0059736245311796665, + 0.004765406716614962, + 0.0075526670552790165, + 0.005667060147970915, + 0.004169100895524025, + 0.009496537037193775, + 0.006343125831335783, + 0.008235815912485123, + 0.0046358369290828705, + 0.0056084925308823586, + 0.0037102135829627514, + 0.0054814801551401615, + 0.004260790068656206, + 0.022388160228729248, + 0.0077162510715425014, + 0.014715539291501045, + 0.003934864420443773, + 0.004257144406437874, + 0.0035357552114874125, + 0.005352926906198263, + 0.003864865517243743, + 0.006771556101739407, + 0.005589923355728388, + 0.004703657701611519, + 0.007876378484070301, + 0.007568849716335535, + 0.007848975248634815, + 0.00977624487131834, + 0.02457280643284321, + 0.017154380679130554, + 0.013470890931785107, + 0.049273472279310226, + 0.0429166741669178, + 0.011249775998294353, + 0.01740148663520813, + 0.06025175005197525, + 0.008573967963457108, + 0.01557205431163311, + 0.01845557615160942, + 0.017745893448591232, + 0.00614555599167943, + 0.006654339376837015, + 0.005961812566965818, + 0.006483978126198053, + 0.0072183446027338505, + 0.011457853950560093, + 0.00958932563662529, + 0.009814897552132607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.013299562968313694 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016116110607981682 + ], + "opposite_logprob_tokens": [ + -4.141116142272949 + ], + "logprob_reward": -0.016116110607981682, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016116110607981682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14278475940227509, + 0.20655454695224762, + 0.10522320121526718, + 0.08561835438013077, + 0.04901348054409027, + 0.054605141282081604, + 0.04531886428594589, + 0.0938258096575737, + 0.10005997866392136, + 0.11699584871530533 + ], + "credit_map": [ + 0.018696671351790428, + 0.03710341826081276, + 0.011938506737351418, + 0.012424496002495289, + 0.01355453860014677, + 0.017717063426971436, + 0.011088124476373196, + 0.020261947065591812, + 0.017735684290528297, + 0.07352873682975769, + 0.0159437358379364, + 0.013774679973721504, + 0.024282071739435196, + 0.023678937926888466, + 0.015708213672041893, + 0.02190249040722847, + 0.011179096065461636, + 0.018075592815876007, + 0.008393088355660439, + 0.008480353280901909, + 0.01121683418750763, + 0.020547324791550636, + 0.007921827025711536, + 0.01940908282995224, + 0.010148276574909687, + 0.013813477009534836, + 0.014236265793442726, + 0.0072999028488993645, + 0.008351431228220463, + 0.010909823700785637, + 0.00959053635597229, + 0.01126863993704319, + 0.010443036444485188, + 0.003737054765224457, + 0.005982863716781139, + 0.004582344554364681, + 0.003740659449249506, + 0.008664385415613651, + 0.004322944208979607, + 0.0075401924550533295, + 0.0063628158532083035, + 0.00882641226053238, + 0.005672699771821499, + 0.005680573172867298, + 0.005570987705141306, + 0.007357638329267502, + 0.008680584840476513, + 0.006453429348766804, + 0.004866152070462704, + 0.005067803896963596, + 0.00464011263102293, + 0.006527337711304426, + 0.00615724315866828, + 0.006846527103334665, + 0.005797424353659153, + 0.005416268482804298, + 0.006324210669845343, + 0.0071776551194489, + 0.006387717556208372, + 0.006188735831528902, + 0.023204687982797623, + 0.011036292649805546, + 0.010232703760266304, + 0.023273801431059837, + 0.014203568920493126, + 0.005436543375253677, + 0.008373629301786423, + 0.0274741780012846, + 0.004867013078182936, + 0.010102336294949055, + 0.007507610600441694, + 0.022095099091529846, + 0.0103303836658597, + 0.012114791199564934, + 0.009601031430065632, + 0.012113926000893116, + 0.013594387099146843, + 0.02752826362848282, + 0.01554930955171585, + 0.01616375893354416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.016116110607981682 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038912534713745 + ], + "opposite_logprob_tokens": [ + -1.9103891849517822 + ], + "logprob_reward": -0.16038912534713745, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038912534713745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16986913979053497, + 0.024956976994872093, + 0.048377614468336105, + 0.05389334261417389, + 0.05055541917681694, + 0.0776994377374649, + 0.029998058453202248, + 0.047048263251781464, + 0.445748895406723, + 0.051852788776159286 + ], + "credit_map": [ + 0.01579992100596428, + 0.0464625246822834, + 0.01763225719332695, + 0.014348047785460949, + 0.01378554292023182, + 0.02311234548687935, + 0.014958970248699188, + 0.023769544437527657, + 0.0031619330402463675, + 0.006117978598922491, + 0.0021776820067316294, + 0.002273125108331442, + 0.00253569264896214, + 0.003889200510457158, + 0.001982174115255475, + 0.002819191897287965, + 0.0053717163391411304, + 0.007297551725059748, + 0.005457662045955658, + 0.0048552630469202995, + 0.0047568767331540585, + 0.008243064396083355, + 0.004479244351387024, + 0.007916240021586418, + 0.0060026757419109344, + 0.011621254496276379, + 0.00676268246024847, + 0.004235970787703991, + 0.0055243694223463535, + 0.008282541297376156, + 0.0037180883809924126, + 0.007745759561657906, + 0.0059965830296278, + 0.0054129306226968765, + 0.00600774772465229, + 0.00661982549354434, + 0.004484687000513077, + 0.005831744987517595, + 0.007252124138176441, + 0.008949780836701393, + 0.0061038946732878685, + 0.008445960469543934, + 0.009694340638816357, + 0.006591918412595987, + 0.007340996526181698, + 0.010707445442676544, + 0.012756516225636005, + 0.01605837233364582, + 0.00289511657319963, + 0.0031348583288490772, + 0.002686112653464079, + 0.003964120987802744, + 0.003124428912997246, + 0.003411954501643777, + 0.003818150144070387, + 0.006963319610804319, + 0.0034166271798312664, + 0.004071148578077555, + 0.0035294366534799337, + 0.003968236036598682, + 0.009366721846163273, + 0.01082004513591528, + 0.005728279240429401, + 0.0061477720737457275, + 0.0980730876326561, + 0.023619798943400383, + 0.04269997775554657, + 0.1476176232099533, + 0.0194931048899889, + 0.037137243896722794, + 0.037822287529706955, + 0.03928583860397339, + 0.0044404673390090466, + 0.005710149183869362, + 0.005644770804792643, + 0.00581964198499918, + 0.0065499586053192616, + 0.00996088981628418, + 0.007920597679913044, + 0.00580631336197257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.16038912534713745 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029806653037667274 + ], + "opposite_logprob_tokens": [ + -3.529806613922119 + ], + "logprob_reward": -0.029806653037667274, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029806653037667274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15700045228004456, + 0.05306832492351532, + 0.05512980371713638, + 0.039694178849458694, + 0.04110535606741905, + 0.12871158123016357, + 0.05360027402639389, + 0.29489636421203613, + 0.08419840037822723, + 0.09259530901908875 + ], + "credit_map": [ + 0.016545964404940605, + 0.05910179391503334, + 0.013307849876582623, + 0.011928439140319824, + 0.011050251312553883, + 0.017269039526581764, + 0.01245174091309309, + 0.015345362946391106, + 0.0033741078805178404, + 0.016024308279156685, + 0.004737451672554016, + 0.004197118803858757, + 0.007343865465372801, + 0.007578648626804352, + 0.004427410662174225, + 0.005385411437600851, + 0.006022926885634661, + 0.0082886116579175, + 0.006646266207098961, + 0.004690449219197035, + 0.005655319895595312, + 0.009045561775565147, + 0.005877151619642973, + 0.00890351738780737, + 0.005806322675198317, + 0.007758823689073324, + 0.006396819371730089, + 0.003408276941627264, + 0.004257212392985821, + 0.003936867229640484, + 0.0029981748666614294, + 0.00513168191537261, + 0.0055756475776433945, + 0.004496970679610968, + 0.004331488627940416, + 0.0042490847408771515, + 0.004780508577823639, + 0.007608977612107992, + 0.003944529220461845, + 0.006118149030953646, + 0.008754742331802845, + 0.01651092991232872, + 0.006112939678132534, + 0.008717676624655724, + 0.009016215801239014, + 0.040584124624729156, + 0.009908280335366726, + 0.029106661677360535, + 0.006388045381754637, + 0.005255557131022215, + 0.008469602093100548, + 0.006443261168897152, + 0.006957418750971556, + 0.007345158141106367, + 0.006769612897187471, + 0.00597161753103137, + 0.016920657828450203, + 0.016301192343235016, + 0.013058841228485107, + 0.017853369936347008, + 0.09402933716773987, + 0.027553396299481392, + 0.02148067019879818, + 0.08769889920949936, + 0.021578656509518623, + 0.006241076625883579, + 0.00572679378092289, + 0.00747314840555191, + 0.005499755498021841, + 0.007753473706543446, + 0.007939140312373638, + 0.021986357867717743, + 0.009332909248769283, + 0.012007646262645721, + 0.011800942942500114, + 0.00984079111367464, + 0.012023924849927425, + 0.014877581968903542, + 0.011898582801222801, + 0.010812923312187195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.029806653037667274 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809512034058571 + ], + "opposite_logprob_tokens": [ + -3.288095235824585 + ], + "logprob_reward": -0.03809512034058571, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809512034058571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08771652728319168, + 0.03856196254491806, + 0.07711422443389893, + 0.031395912170410156, + 0.03732699528336525, + 0.1004156693816185, + 0.02877953089773655, + 0.2583359479904175, + 0.16281121969223022, + 0.17754201591014862 + ], + "credit_map": [ + 0.008261922746896744, + 0.020108554512262344, + 0.010234066285192966, + 0.008638699539005756, + 0.008238879963755608, + 0.012208767235279083, + 0.007853894494473934, + 0.012171740643680096, + 0.0037234548944979906, + 0.009134648367762566, + 0.0026914565823972225, + 0.0033216739539057016, + 0.004256434738636017, + 0.006875623017549515, + 0.003246179549023509, + 0.005312489345669746, + 0.007434732746332884, + 0.011137408204376698, + 0.010461905971169472, + 0.006908685900270939, + 0.008201674558222294, + 0.010932628065347672, + 0.007573855575174093, + 0.014463333413004875, + 0.0033096831757575274, + 0.003338631708174944, + 0.005464497953653336, + 0.0031860540620982647, + 0.0025007068179547787, + 0.005296568386256695, + 0.0034600321669131517, + 0.004839740693569183, + 0.0038171024061739445, + 0.00456971162930131, + 0.004505799151957035, + 0.003524258267134428, + 0.003960672300308943, + 0.006583590060472488, + 0.00411700364202261, + 0.006248858757317066, + 0.008117258548736572, + 0.016364801675081253, + 0.008462697267532349, + 0.0107382507994771, + 0.010991369374096394, + 0.016061116009950638, + 0.013390197418630123, + 0.016289977356791496, + 0.0029448133427649736, + 0.0028492463752627373, + 0.0034248209558427334, + 0.003234448377043009, + 0.003633726155385375, + 0.003629337763413787, + 0.0041751316748559475, + 0.004888005554676056, + 0.01635650359094143, + 0.01533455215394497, + 0.014574461616575718, + 0.016890356317162514, + 0.07836200296878815, + 0.026071617379784584, + 0.020781900733709335, + 0.06996455788612366, + 0.048340000212192535, + 0.010951626114547253, + 0.010732720606029034, + 0.018737342208623886, + 0.009621642529964447, + 0.015429050661623478, + 0.010205893777310848, + 0.03879296034574509, + 0.016997206956148148, + 0.020545994862914085, + 0.020400317385792732, + 0.02001432701945305, + 0.02029821276664734, + 0.03255729004740715, + 0.02225557714700699, + 0.024473082274198532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.03809512034058571 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a striking black-and-white high-dynamic range (HDR) image, an underwater close-up shot, shot from a slightly upward perspective, sways gently yet steadily in time with the rhythm of the dark current. At the center of the frame is a massive swimming crab. Illuminated by a single, intense light source from above, its carapace reveals an intricate relief of pure white highlights and deep shadows, each spike and vein striking with breathtaking sharpness. Its signature flat claws, like a pair of meticulously polished obsidian oars, gleam with the hard luster of ceramic, their jagged inner edges auguring destructive power. The giant crab secures a strand of tough seaweed with its smaller appendages. Then, with a slow, demonstrative flourish of one of its massive claws, the movement accelerates, snapping down with swift, decisive force, sinking its teeth into the seaweed's midsection. The exquisite detail allows viewers to practically hear the sharp snap of seaweed fibers being cut, the severed strands then drifting feebly and slowly into the endless darkness beyond the camera. The entire scene, rendered in stark black and white, creates a primal, cold, yet powerful atmosphere, perfectly showcasing the astonishing strength and awe-inspiring biomechanical beauty of this underwater predator.", + "seed": 16209, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_10/index_5447_seed_16209_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08570314403623343, + "vlm_logprob": -0.08570314403623343, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.08570314403623343, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0463050976395607 + ], + "opposite_logprob_tokens": [ + -3.1713051795959473 + ], + "logprob_reward": -0.0463050976395607, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0463050976395607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15091559290885925, + 0.021218113601207733, + 0.015061290934681892, + 0.01845846138894558, + 0.01737135462462902, + 0.037869859486818314, + 0.06673893332481384, + 0.04218553751707077, + 0.4648529291152954, + 0.1653279960155487 + ], + "credit_map": [ + 0.019261257722973824, + 0.01769169420003891, + 0.026726625859737396, + 0.01507068332284689, + 0.02390783466398716, + 0.016978798434138298, + 0.015150783583521843, + 0.016127897426486015, + 0.002557860454544425, + 0.0024076735135167837, + 0.004282400012016296, + 0.0018307928694412112, + 0.002526921220123768, + 0.0019312634831294417, + 0.0023664222098886967, + 0.0033147786743938923, + 0.0015319858212023973, + 0.0012758933007717133, + 0.0018622428178787231, + 0.0029228939674794674, + 0.002289322903379798, + 0.0015185379888862371, + 0.0017053582705557346, + 0.0019550537690520287, + 0.0017808953998610377, + 0.0015203371876850724, + 0.002097560092806816, + 0.0024416048545390368, + 0.002207090612500906, + 0.00331379403360188, + 0.0024344297125935555, + 0.0026627464685589075, + 0.0016788452630862594, + 0.002168868202716112, + 0.0017051581526175141, + 0.0016299483831971884, + 0.0035038713831454515, + 0.0024481031578034163, + 0.0022807051427662373, + 0.001955852610990405, + 0.003201843239367008, + 0.0052077812142670155, + 0.005205865483731031, + 0.004260245244950056, + 0.00540275638923049, + 0.005667316261678934, + 0.004102915525436401, + 0.004821133799850941, + 0.005502178333699703, + 0.007491110358387232, + 0.012712945230305195, + 0.0089576356112957, + 0.006774349603801966, + 0.006474356632679701, + 0.009087705053389072, + 0.009738640859723091, + 0.007067910861223936, + 0.003605556208640337, + 0.008255484513938427, + 0.005914818495512009, + 0.0036443048156797886, + 0.0030848761089146137, + 0.004207254387438297, + 0.006405329331755638, + 0.024650173261761665, + 0.014679988846182823, + 0.15964573621749878, + 0.05055517703294754, + 0.01398634072393179, + 0.014368414878845215, + 0.017172763124108315, + 0.16979427635669708, + 0.026996299624443054, + 0.008910151198506355, + 0.024007167667150497, + 0.04607885703444481, + 0.010332908481359482, + 0.007940022274851799, + 0.020302122458815575, + 0.02076045051217079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0463050976395607 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00868625845760107 + ], + "opposite_logprob_tokens": [ + -4.758686065673828 + ], + "logprob_reward": -0.00868625845760107, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00868625845760107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.153111070394516, + 0.0835685208439827, + 0.03614982217550278, + 0.06662368774414062, + 0.03597923368215561, + 0.03022473305463791, + 0.025548744946718216, + 0.09546305239200592, + 0.3361503779888153, + 0.13718076050281525 + ], + "credit_map": [ + 0.013070104643702507, + 0.01669953763484955, + 0.030023640021681786, + 0.014549621380865574, + 0.029362080618739128, + 0.014269230887293816, + 0.02095644734799862, + 0.014180394820868969, + 0.00959018711000681, + 0.013089865446090698, + 0.013417202048003674, + 0.0074969446286559105, + 0.012944908812642097, + 0.007849318906664848, + 0.008225157856941223, + 0.010954936034977436, + 0.0036414461210370064, + 0.0030884204898029566, + 0.005139997228980064, + 0.006165936589241028, + 0.004701634868979454, + 0.003686004551127553, + 0.0053206211887300014, + 0.004405759274959564, + 0.006007353309541941, + 0.007822280749678612, + 0.004825716372579336, + 0.006320882588624954, + 0.01396440900862217, + 0.014295512810349464, + 0.006030786316841841, + 0.007356749847531319, + 0.005568911321461201, + 0.003515196032822132, + 0.004800122696906328, + 0.0038006657268851995, + 0.004248069599270821, + 0.004374257288873196, + 0.003525134641677141, + 0.006146877072751522, + 0.0026194413658231497, + 0.0025908907409757376, + 0.004694732837378979, + 0.003998299594968557, + 0.004027596674859524, + 0.004026761744171381, + 0.004479140043258667, + 0.003787873312830925, + 0.001835034810937941, + 0.0024336189962923527, + 0.0037074771244078875, + 0.0038241473957896233, + 0.002167325932532549, + 0.003203794127330184, + 0.004714549519121647, + 0.0036627966910600662, + 0.008550363592803478, + 0.006716930773109198, + 0.017009105533361435, + 0.016989802941679955, + 0.007330533117055893, + 0.01162300631403923, + 0.011282983236014843, + 0.015960324555635452, + 0.020652038976550102, + 0.015598481521010399, + 0.09936419129371643, + 0.046663910150527954, + 0.011724983341991901, + 0.014049786143004894, + 0.016472654417157173, + 0.1116243228316307, + 0.0176403746008873, + 0.007904973812401295, + 0.021492112427949905, + 0.03469868004322052, + 0.008364550769329071, + 0.007154772523790598, + 0.017605291679501534, + 0.02231999672949314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.00868625845760107 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134618103504181 + ], + "opposite_logprob_tokens": [ + -1.3134617805480957 + ], + "logprob_reward": -0.3134618103504181, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134618103504181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09109142422676086, + 0.04664509743452072, + 0.03134308382868767, + 0.026022102683782578, + 0.03161453828215599, + 0.04772801324725151, + 0.060787051916122437, + 0.05726923793554306, + 0.40601393580436707, + 0.20148546993732452 + ], + "credit_map": [ + 0.010079232975840569, + 0.010047669522464275, + 0.016728680580854416, + 0.010190769098699093, + 0.015504839830100536, + 0.007036745082587004, + 0.010731806978583336, + 0.010771680623292923, + 0.005276874173432589, + 0.006691798102110624, + 0.008825084194540977, + 0.004841975402086973, + 0.0058683292008936405, + 0.004802871495485306, + 0.004745619371533394, + 0.005592545494437218, + 0.0029942032415419817, + 0.004031683783978224, + 0.004352611023932695, + 0.006767150480300188, + 0.0032828187104314566, + 0.0031120029743760824, + 0.0031264761928468943, + 0.0036761362571269274, + 0.0026446327101439238, + 0.003196357749402523, + 0.0021187649108469486, + 0.003570214379578829, + 0.0033654605504125357, + 0.004415266215801239, + 0.0029895033221691847, + 0.003721904242411256, + 0.003047884674742818, + 0.003048307029530406, + 0.005129775032401085, + 0.003778721671551466, + 0.004286264535039663, + 0.00463343458250165, + 0.003653839463368058, + 0.004036313388496637, + 0.003574642352759838, + 0.004828541073948145, + 0.0058447495102882385, + 0.007807269226759672, + 0.006156662944704294, + 0.008087454363703728, + 0.005919733550399542, + 0.005508960224688053, + 0.0045196968130767345, + 0.006176029797643423, + 0.011264821514487267, + 0.008266691118478775, + 0.006842318922281265, + 0.006007542368024588, + 0.008901328779757023, + 0.00880861934274435, + 0.008974041789770126, + 0.004474570043385029, + 0.008362039923667908, + 0.009072631597518921, + 0.004328579176217318, + 0.006094491574913263, + 0.006264709867537022, + 0.009698169305920601, + 0.01937900111079216, + 0.018662763759493828, + 0.1441774219274521, + 0.03242826089262962, + 0.013493225909769535, + 0.014283793978393078, + 0.015876155346632004, + 0.1477133333683014, + 0.03598009794950485, + 0.010414191521704197, + 0.028659876435995102, + 0.05560010299086571, + 0.013093952089548111, + 0.008914888836443424, + 0.0206098984926939, + 0.02821246348321438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.3134618103504181 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033709973096847534 + ], + "opposite_logprob_tokens": [ + -3.40871000289917 + ], + "logprob_reward": -0.033709973096847534, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033709973096847534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.117058664560318, + 0.05416376516222954, + 0.056577738374471664, + 0.02526784874498844, + 0.058177754282951355, + 0.05206844583153725, + 0.0497698150575161, + 0.07756565511226654, + 0.3745781183242798, + 0.1347721815109253 + ], + "credit_map": [ + 0.011474903672933578, + 0.0160832479596138, + 0.016536487266421318, + 0.011466532945632935, + 0.022489111870527267, + 0.013145736418664455, + 0.014254323206841946, + 0.011608327738940716, + 0.006421592552214861, + 0.0050737434066832066, + 0.01210921909660101, + 0.005121947731822729, + 0.005154533311724663, + 0.0055934228003025055, + 0.0061256736516952515, + 0.008563636802136898, + 0.005722433794289827, + 0.004056009463965893, + 0.0056545729748904705, + 0.011391403153538704, + 0.007626520004123449, + 0.006067307200282812, + 0.008722366765141487, + 0.007337130140513182, + 0.002018363680690527, + 0.0021854485385119915, + 0.002794199623167515, + 0.002952735871076584, + 0.00413922406733036, + 0.005024296697229147, + 0.0027730786241590977, + 0.003380504436790943, + 0.0063268449157476425, + 0.0056086257100105286, + 0.0073188794776797295, + 0.006471054628491402, + 0.010852840729057789, + 0.0074809035286307335, + 0.006251953076571226, + 0.00786665640771389, + 0.004606297239661217, + 0.006808623671531677, + 0.0051353927701711655, + 0.007480020634829998, + 0.0069490945897996426, + 0.007995523512363434, + 0.007662707008421421, + 0.0054307905957102776, + 0.0045127240009605885, + 0.005344625562429428, + 0.010179786011576653, + 0.006911255419254303, + 0.003877895185723901, + 0.004888363182544708, + 0.006432454101741314, + 0.007622712757438421, + 0.007677488494664431, + 0.005905395373702049, + 0.012529727071523666, + 0.014277996495366096, + 0.0066986377350986, + 0.007302436046302319, + 0.009881657548248768, + 0.01329231821000576, + 0.026543181389570236, + 0.017002379521727562, + 0.10423273593187332, + 0.052159037441015244, + 0.01849278435111046, + 0.02324214205145836, + 0.021950645372271538, + 0.11095523834228516, + 0.019430775195360184, + 0.010341444052755833, + 0.01671179011464119, + 0.03639967367053032, + 0.008741437457501888, + 0.00978946965187788, + 0.017093248665332794, + 0.016264354810118675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.033709973096847534 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635258063673973 + ], + "opposite_logprob_tokens": [ + -3.651352643966675 + ], + "logprob_reward": -0.02635258063673973, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635258063673973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07658727467060089, + 0.0343804769217968, + 0.027383720502257347, + 0.019012225791811943, + 0.016213078051805496, + 0.02791079320013523, + 0.03768445551395416, + 0.0156860314309597, + 0.6813280582427979, + 0.0638139471411705 + ], + "credit_map": [ + 0.010834041982889175, + 0.007777562364935875, + 0.012786193750798702, + 0.007941961288452148, + 0.01175807137042284, + 0.00761334877461195, + 0.00886186771094799, + 0.00901422742754221, + 0.0034048138186335564, + 0.004925549495965242, + 0.005406162701547146, + 0.003744939574971795, + 0.005391442216932774, + 0.0036780070513486862, + 0.0035317447036504745, + 0.004297817591577768, + 0.0030529960058629513, + 0.002637851284816861, + 0.002596066566184163, + 0.0051314448937773705, + 0.0035678416024893522, + 0.0022918670438230038, + 0.004946415778249502, + 0.003159238724038005, + 0.0014878909569233656, + 0.00198471755720675, + 0.0019720098935067654, + 0.002054633805528283, + 0.003417005529627204, + 0.0036954835522919893, + 0.0020613258238881826, + 0.002339158207178116, + 0.002003848785534501, + 0.0014939835527911782, + 0.002594325225800276, + 0.0018433494260534644, + 0.0017605337779968977, + 0.002284703776240349, + 0.0020075603388249874, + 0.0022247720044106245, + 0.002831017132848501, + 0.0036701455246657133, + 0.004057903774082661, + 0.0035205380991101265, + 0.0027817904483526945, + 0.0037708256859332323, + 0.0031405813060700893, + 0.004137993324548006, + 0.0028312283102422953, + 0.00444054463878274, + 0.005855148658156395, + 0.005135959479957819, + 0.003956998232752085, + 0.004722870886325836, + 0.005726674571633339, + 0.005015031434595585, + 0.0015208646655082703, + 0.001658173161558807, + 0.0020879346411675215, + 0.0028039768803864717, + 0.0016743493033573031, + 0.0017952958587557077, + 0.0018318159272894263, + 0.002313620876520872, + 0.03413130342960358, + 0.021964648738503456, + 0.254803329706192, + 0.04938902705907822, + 0.020283667370676994, + 0.019873883575201035, + 0.019225038588047028, + 0.261657178401947, + 0.00866422988474369, + 0.005282072816044092, + 0.009212867356836796, + 0.014956907369196415, + 0.004521795082837343, + 0.004936463199555874, + 0.008330250158905983, + 0.00790935754776001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.02635258063673973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a striking black-and-white high-dynamic range (HDR) image, an underwater close-up shot, shot from a slightly upward perspective, sways gently yet steadily in time with the rhythm of the dark current. At the center of the frame is a massive swimming crab. Illuminated by a single, intense light source from above, its carapace reveals an intricate relief of pure white highlights and deep shadows, each spike and vein striking with breathtaking sharpness. Its signature flat claws, like a pair of meticulously polished obsidian oars, gleam with the hard luster of ceramic, their jagged inner edges auguring destructive power. The giant crab secures a strand of tough seaweed with its smaller appendages. Then, with a slow, demonstrative flourish of one of its massive claws, the movement accelerates, snapping down with swift, decisive force, sinking its teeth into the seaweed's midsection. The exquisite detail allows viewers to practically hear the sharp snap of seaweed fibers being cut, the severed strands then drifting feebly and slowly into the endless darkness beyond the camera. The entire scene, rendered in stark black and white, creates a primal, cold, yet powerful atmosphere, perfectly showcasing the astonishing strength and awe-inspiring biomechanical beauty of this underwater predator.", + "seed": 16210, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_10/index_5447_seed_16210_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07306914571672678, + "vlm_logprob": -0.07306914571672678, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.07306914571672678, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015062288381159306 + ], + "opposite_logprob_tokens": [ + -4.39006233215332 + ], + "logprob_reward": -0.015062288381159306, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.015062288381159306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5381525158882141, + 0.08330968022346497, + 0.05876152962446213, + 0.08054231107234955, + 0.014354176819324493, + 0.034425199031829834, + 0.07347489148378372, + 0.05283842980861664, + 0.029767632484436035, + 0.03437364101409912 + ], + "credit_map": [ + 0.031186116859316826, + 0.016175493597984314, + 0.20210868120193481, + 0.033759910613298416, + 0.029582953080534935, + 0.017675014212727547, + 0.021752536296844482, + 0.18591180443763733, + 0.006789739243686199, + 0.004085307475179434, + 0.030896345153450966, + 0.006395500153303146, + 0.011075861752033234, + 0.0035567230079323053, + 0.0039769187569618225, + 0.01653328910470009, + 0.0041671316139400005, + 0.003069828497245908, + 0.018311813473701477, + 0.0053364550694823265, + 0.004779599141329527, + 0.0027105025947093964, + 0.00310525787062943, + 0.017280949279665947, + 0.007382650393992662, + 0.005476401653140783, + 0.012486995197832584, + 0.007150726392865181, + 0.012772997841238976, + 0.005818714387714863, + 0.010633517056703568, + 0.01882031373679638, + 0.0011334881419315934, + 0.001759152626618743, + 0.0028102544602006674, + 0.0014848459977656603, + 0.002403418766334653, + 0.0010820974130183458, + 0.0018743654945865273, + 0.001806554850190878, + 0.003210916882380843, + 0.003308488056063652, + 0.0056997667998075485, + 0.005790241062641144, + 0.00472451513633132, + 0.0032202014699578285, + 0.004427723586559296, + 0.004043345805257559, + 0.006463250610977411, + 0.010875053703784943, + 0.007383381947875023, + 0.0074434829875826836, + 0.015395769849419594, + 0.0073208580724895, + 0.009459140710532665, + 0.009133960120379925, + 0.004539215937256813, + 0.007982360199093819, + 0.006019951775670052, + 0.007966260425746441, + 0.010482543148100376, + 0.004932174924761057, + 0.007372367661446333, + 0.00354355713352561, + 0.0020023034885525703, + 0.002932497765868902, + 0.003557657590135932, + 0.003974321763962507, + 0.004883775021880865, + 0.0032889374997466803, + 0.004802035633474588, + 0.004326105583459139, + 0.0035803508944809437, + 0.003601239062845707, + 0.004416043404489756, + 0.005699370987713337, + 0.004330029245465994, + 0.003758228151127696, + 0.005677337292581797, + 0.0033110459335148335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.015062288381159306 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373970836400986 + ], + "opposite_logprob_tokens": [ + -3.4087398052215576 + ], + "logprob_reward": -0.03373970836400986, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373970836400986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20554661750793457, + 0.09153825789690018, + 0.06106119230389595, + 0.15451505780220032, + 0.05668976157903671, + 0.027913792058825493, + 0.033292483538389206, + 0.08474749326705933, + 0.04415686056017876, + 0.24053841829299927 + ], + "credit_map": [ + 0.01683182269334793, + 0.021589083597064018, + 0.0222412571310997, + 0.0378316193819046, + 0.036222878843545914, + 0.013313500210642815, + 0.01943831518292427, + 0.03807817026972771, + 0.0054908073507249355, + 0.007112481165677309, + 0.020944735035300255, + 0.011028389446437359, + 0.01263001374900341, + 0.006003804039210081, + 0.00667417049407959, + 0.02165386639535427, + 0.004689380992203951, + 0.005462088622152805, + 0.017721792683005333, + 0.006236964836716652, + 0.004410420078784227, + 0.0028387210331857204, + 0.003220032434910536, + 0.0164817962795496, + 0.013226311653852463, + 0.013720802962779999, + 0.01757028140127659, + 0.019339757040143013, + 0.027818739414215088, + 0.01496153138577938, + 0.021331630647182465, + 0.026546012610197067, + 0.0044541326351463795, + 0.009312910959124565, + 0.006800305563956499, + 0.006636160425841808, + 0.01171075738966465, + 0.0036764920223504305, + 0.00707797147333622, + 0.007021042052656412, + 0.002455255016684532, + 0.004298758693039417, + 0.00389879010617733, + 0.003471455769613385, + 0.0036545516923069954, + 0.0026028947904706, + 0.0039611817337572575, + 0.0035709061194211245, + 0.0036605519708245993, + 0.0035963698755949736, + 0.0035028131678700447, + 0.00455736368894577, + 0.006714642513543367, + 0.003964282106608152, + 0.003855526214465499, + 0.0034409344661980867, + 0.0081808315590024, + 0.010502297431230545, + 0.011419096030294895, + 0.01198467519134283, + 0.012596555054187775, + 0.006547051016241312, + 0.01677953079342842, + 0.006737465504556894, + 0.0031864657066762447, + 0.00597150856629014, + 0.004363028332591057, + 0.008498463779687881, + 0.005107867997139692, + 0.004924281965941191, + 0.005562048405408859, + 0.006543201394379139, + 0.024733858183026314, + 0.04642390459775925, + 0.028125669807195663, + 0.026351606473326683, + 0.03873215988278389, + 0.02379775419831276, + 0.022472847253084183, + 0.029900621622800827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.03373970836400986 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252096563577652 + ], + "opposite_logprob_tokens": [ + -1.5020965337753296 + ], + "logprob_reward": -0.252096563577652, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252096563577652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27484941482543945, + 0.06515331566333771, + 0.19685861468315125, + 0.1423088014125824, + 0.04254824295639992, + 0.08336013555526733, + 0.05174257233738899, + 0.03517983853816986, + 0.0401858314871788, + 0.06781317293643951 + ], + "credit_map": [ + 0.0162286926060915, + 0.008797327987849712, + 0.10760732740163803, + 0.01620357111096382, + 0.012470182031393051, + 0.008584676310420036, + 0.008412596769630909, + 0.09654504805803299, + 0.006399071309715509, + 0.0021355883218348026, + 0.02630480006337166, + 0.004700149409472942, + 0.005868405103683472, + 0.0022255280055105686, + 0.003108193166553974, + 0.01441158726811409, + 0.013650973327457905, + 0.011818327941000462, + 0.06320138275623322, + 0.01974956877529621, + 0.013300219550728798, + 0.005926657002419233, + 0.014101223088800907, + 0.05511026829481125, + 0.013074168935418129, + 0.012597765773534775, + 0.016054807230830193, + 0.017750101163983345, + 0.024320200085639954, + 0.013714062049984932, + 0.019341686740517616, + 0.025456009432673454, + 0.003781752660870552, + 0.005966329947113991, + 0.005974758882075548, + 0.0034644172992557287, + 0.010082514956593513, + 0.003493865020573139, + 0.005159413907676935, + 0.004625190980732441, + 0.008024800568819046, + 0.010346771217882633, + 0.012030888348817825, + 0.01026835385710001, + 0.013753141276538372, + 0.00827053189277649, + 0.012394625693559647, + 0.008271023631095886, + 0.006868188735097647, + 0.008295601233839989, + 0.00542324548587203, + 0.005016045644879341, + 0.008475309237837791, + 0.0043235295452177525, + 0.005773393902927637, + 0.007567258086055517, + 0.004394481889903545, + 0.005406868178397417, + 0.003535493975505233, + 0.005170206073671579, + 0.005130665376782417, + 0.004574577324092388, + 0.004138365387916565, + 0.002829183591529727, + 0.0029521211981773376, + 0.003633746411651373, + 0.00446726568043232, + 0.005609380081295967, + 0.005807751324027777, + 0.006071250420063734, + 0.005574340932071209, + 0.006069974508136511, + 0.00537261925637722, + 0.0069887349382042885, + 0.012760874815285206, + 0.008790481835603714, + 0.010110517032444477, + 0.007466003764420748, + 0.008253665640950203, + 0.00807027518749237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.252096563577652 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810097277164459 + ], + "opposite_logprob_tokens": [ + -3.2881009578704834 + ], + "logprob_reward": -0.03810097277164459, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810097277164459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3334059417247772, + 0.07858019322156906, + 0.10595814883708954, + 0.13457106053829193, + 0.03669074550271034, + 0.046300195157527924, + 0.05574366822838783, + 0.08227131515741348, + 0.06568262726068497, + 0.060796063393354416 + ], + "credit_map": [ + 0.01797911524772644, + 0.01614772528409958, + 0.07373743504285812, + 0.026795078068971634, + 0.05146772414445877, + 0.014843354001641273, + 0.01814425177872181, + 0.1142912432551384, + 0.007374930661171675, + 0.00336258253082633, + 0.030052458867430687, + 0.006073498632758856, + 0.010771499015390873, + 0.0031501390039920807, + 0.003923937678337097, + 0.013871151022613049, + 0.0067419251427054405, + 0.005110227502882481, + 0.037861742079257965, + 0.008466700091958046, + 0.007405910175293684, + 0.0028323151636868715, + 0.005393602419644594, + 0.032145723700523376, + 0.013339582830667496, + 0.010702555067837238, + 0.01618395932018757, + 0.01457127183675766, + 0.022762566804885864, + 0.012670597992837429, + 0.01891106739640236, + 0.025429463014006615, + 0.0037923119962215424, + 0.0040965694934129715, + 0.006776779890060425, + 0.003263164311647415, + 0.008328604511916637, + 0.0020984874572604895, + 0.0034229441080242395, + 0.0049118841998279095, + 0.0044470313005149364, + 0.005380064249038696, + 0.007475986145436764, + 0.006879343185573816, + 0.0067774951457977295, + 0.004989264067262411, + 0.005488214548677206, + 0.004862798377871513, + 0.005539155565202236, + 0.00895052682608366, + 0.00774326641112566, + 0.005327198654413223, + 0.010937659069895744, + 0.005622380413115025, + 0.006483457982540131, + 0.005140023771673441, + 0.0076448689214885235, + 0.011523286812007427, + 0.010754220187664032, + 0.011862754821777344, + 0.013036738149821758, + 0.008861694484949112, + 0.011733067221939564, + 0.00685468316078186, + 0.005159966181963682, + 0.007445871364325285, + 0.005859990604221821, + 0.011384926736354828, + 0.008141854777932167, + 0.008898057043552399, + 0.009005160070955753, + 0.009786799550056458, + 0.005972295068204403, + 0.006205123849213123, + 0.00982282031327486, + 0.008780399337410927, + 0.009537499397993088, + 0.004630513023585081, + 0.008718563243746758, + 0.007128850091248751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.03810097277164459 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive swimming crab as the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black and white with high contrast (pure white highlights and deep shadows)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned at a slightly upward perspective looking down at the crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab use its smaller appendages to secure a strand of tough seaweed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the crab snap one of its massive claws down with swift force to cut the seaweed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346195489168167 + ], + "opposite_logprob_tokens": [ + -3.651346206665039 + ], + "logprob_reward": -0.026346195489168167, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346195489168167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.132176473736763, + 0.2537268400192261, + 0.04013682156801224, + 0.29343676567077637, + 0.01892862468957901, + 0.04690469801425934, + 0.06939369440078735, + 0.06198513135313988, + 0.03959377110004425, + 0.04371713846921921 + ], + "credit_map": [ + 0.009440888650715351, + 0.008963868953287601, + 0.021777687594294548, + 0.015155719593167305, + 0.02161273919045925, + 0.006412997376173735, + 0.009321056306362152, + 0.03949154168367386, + 0.022690458223223686, + 0.010416416451334953, + 0.09849156439304352, + 0.019545456394553185, + 0.03166966140270233, + 0.009151232428848743, + 0.01196176279336214, + 0.04980031028389931, + 0.0030979851726442575, + 0.004159897565841675, + 0.008649305440485477, + 0.006115393713116646, + 0.004214500077068806, + 0.0021905354224145412, + 0.0032523530535399914, + 0.008456856943666935, + 0.025824975222349167, + 0.024003297090530396, + 0.03328144550323486, + 0.03446393460035324, + 0.05829128995537758, + 0.026683522388339043, + 0.04040040820837021, + 0.05048791691660881, + 0.0016830404056236148, + 0.0021660916972905397, + 0.00311161857098341, + 0.001876202062703669, + 0.00455315550789237, + 0.0012332045007497072, + 0.0021011268254369497, + 0.002204187912866473, + 0.0047699701972305775, + 0.005870493594557047, + 0.005733561236411333, + 0.0072237602435052395, + 0.006890904624015093, + 0.005033088847994804, + 0.006590039003640413, + 0.004792885389178991, + 0.006321335677057505, + 0.013740049675107002, + 0.0070868018083274364, + 0.0066404761746525764, + 0.014307042583823204, + 0.005331503693014383, + 0.008834498934447765, + 0.007131993770599365, + 0.0053336904384195805, + 0.007684694603085518, + 0.0070229703560471535, + 0.010504201054573059, + 0.010041103698313236, + 0.006065544672310352, + 0.009470284916460514, + 0.005862649530172348, + 0.003265884006395936, + 0.00369657133705914, + 0.004615214187651873, + 0.0061575621366500854, + 0.004064876586198807, + 0.0056447237730026245, + 0.006038964726030827, + 0.006109980400651693, + 0.0036036469973623753, + 0.006596870254725218, + 0.00516680721193552, + 0.00492080906406045, + 0.007067910395562649, + 0.0038253930397331715, + 0.007540063001215458, + 0.004995645955204964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.026346195489168167 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_11/index_tensor([662, 662])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_11/index_tensor([662, 662])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8bd5d4c7cb870e25cb5f969189b3818c6d2c0110 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_11/index_tensor([662, 662])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle, static close-up captures the meticulous act of seasoning. A steady hand, with clean, manicured nails, hovers deliberately over a dark, heavy pot filled with a rich, simmering, deep-orange vegetable broth, from which rises a gentle stream of steam. Between thumb and index finger, a pinch of fine, earthy-green dried plant powder is held. Gently rubbing the fingers together, the powder drizzles down in a delicate, controlled pattern. The camera maintains sharp focus on the falling particles as they gently land on the broth's glistening surface, floating briefly like fine dust before beginning to dissolve and swirl into the vibrant liquid. The scene is illuminated by a warm, soft light from an off-camera source, highlighting the texture of the powder and the rich color of the broth, contrasting sharply with the dark, out-of-focus grains of the wooden table below, creating an atmosphere of concentration, tranquility, and appetizing culinary delight.", + "seed": 3900, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_11/index_662_seed_3900_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01802893904969096, + "vlm_logprob": -0.01802893904969096, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.01802893904969096, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021836912259459496 + ], + "opposite_logprob_tokens": [ + -4.896836757659912 + ], + "logprob_reward": -0.021836912259459496, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.021836912259459496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04736587405204773, + 0.644179105758667, + 0.028791408985853195, + 0.04163403809070587, + 0.04617496579885483, + 0.03541616350412369, + 0.033508796244859695, + 0.01757826656103134, + 0.03534359484910965, + 0.07000774145126343 + ], + "credit_map": [ + 0.006489903200417757, + 0.0034349996130913496, + 0.006380803883075714, + 0.0077829607762396336, + 0.006048693787306547, + 0.005365884862840176, + 0.005301366094499826, + 0.00656126718968153, + 0.08610641211271286, + 0.020841902121901512, + 0.013305452652275562, + 0.16028450429439545, + 0.03183702379465103, + 0.022816579788923264, + 0.26333168148994446, + 0.045655567198991776, + 0.004808013793081045, + 0.001365807605907321, + 0.0014623752795159817, + 0.005623104516416788, + 0.0020807036198675632, + 0.0015741863753646612, + 0.009769641794264317, + 0.0021075766999274492, + 0.0021413906943053007, + 0.0021019692067056894, + 0.002162266755476594, + 0.014433935284614563, + 0.003227280918508768, + 0.0013952591689303517, + 0.012453842908143997, + 0.0037180937360972166, + 0.0047509558498859406, + 0.0025354239623993635, + 0.0022934849839657545, + 0.011507491581141949, + 0.004747866187244654, + 0.0035170691553503275, + 0.013751748017966747, + 0.0030709258280694485, + 0.0028709417674690485, + 0.009942098520696163, + 0.0016526237595826387, + 0.00880325399339199, + 0.0030847620218992233, + 0.0026771938428282738, + 0.003927402198314667, + 0.002457886701449752, + 0.0034589774440973997, + 0.004221824929118156, + 0.002900161547586322, + 0.005960697773844004, + 0.005222131498157978, + 0.007018817123025656, + 0.003070211037993431, + 0.0016559752402827144, + 0.002173107350245118, + 0.002014192985370755, + 0.002173182088881731, + 0.003583441022783518, + 0.0017209210200235248, + 0.0010808204533532262, + 0.002749148290604353, + 0.002083453582599759, + 0.00434047682210803, + 0.004159341566264629, + 0.002360729733482003, + 0.0028199704829603434, + 0.005334350746124983, + 0.01003231480717659, + 0.0032096037175506353, + 0.0030868048779666424, + 0.009856623597443104, + 0.018786171451210976, + 0.005622913595288992, + 0.006143359467387199, + 0.012575671076774597, + 0.00658412603661418, + 0.005556164309382439, + 0.00488271052017808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.021836912259459496 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002220666501671076 + ], + "opposite_logprob_tokens": [ + -6.127220630645752 + ], + "logprob_reward": -0.002220666501671076, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002220666501671076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08096253871917725, + 0.5617567300796509, + 0.018909446895122528, + 0.01851452887058258, + 0.11618199199438095, + 0.019515270367264748, + 0.06924193352460861, + 0.02106870524585247, + 0.05682387202978134, + 0.037024952471256256 + ], + "credit_map": [ + 0.007332140114158392, + 0.008005243726074696, + 0.00415674177929759, + 0.016896940767765045, + 0.011797521263360977, + 0.005045663565397263, + 0.012700729072093964, + 0.015027569606900215, + 0.0733429566025734, + 0.0197036974132061, + 0.014120310544967651, + 0.14062359929084778, + 0.029197657480835915, + 0.01919539086520672, + 0.22583898901939392, + 0.03973418474197388, + 0.00326297408901155, + 0.0013769050128757954, + 0.001495346543379128, + 0.0029640260618180037, + 0.002297218656167388, + 0.0013976905029267073, + 0.004433159716427326, + 0.0016821286408230662, + 0.0016724776942282915, + 0.0011282088235020638, + 0.0012266389094293118, + 0.004115844611078501, + 0.0016236079391092062, + 0.0020578736439347267, + 0.004771385341882706, + 0.001918493420816958, + 0.007394585758447647, + 0.008389794267714024, + 0.007114799227565527, + 0.03541780635714531, + 0.007696525659412146, + 0.006654042750597, + 0.03771141543984413, + 0.005803038366138935, + 0.0022846513893455267, + 0.006564043462276459, + 0.0011350848944857717, + 0.00411642761901021, + 0.0012646130053326488, + 0.0014778461772948503, + 0.0014733028365299106, + 0.0011993016814813018, + 0.007331349886953831, + 0.00892854854464531, + 0.005156477447599173, + 0.012319964356720448, + 0.011673876084387302, + 0.014622477814555168, + 0.005974418018013239, + 0.003234826261177659, + 0.0024876187089830637, + 0.0016145528061315417, + 0.0025229635648429394, + 0.0026838481426239014, + 0.002076484262943268, + 0.0018662480870261788, + 0.003907569218426943, + 0.003909421153366566, + 0.0067067560739815235, + 0.007759329862892628, + 0.0042007071897387505, + 0.006137571297585964, + 0.006066531874239445, + 0.013588233850896358, + 0.006311042234301567, + 0.006053695920854807, + 0.004301326349377632, + 0.008758519776165485, + 0.0022157225757837296, + 0.0042614624835550785, + 0.005676889326423407, + 0.002378535456955433, + 0.004987508989870548, + 0.004444990307092667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.002220666501671076 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018204161897301674 + ], + "opposite_logprob_tokens": [ + -4.018204212188721 + ], + "logprob_reward": -0.018204161897301674, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018204161897301674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11691459268331528, + 0.4681473970413208, + 0.03420112654566765, + 0.1500687152147293, + 0.05777398496866226, + 0.015272341668605804, + 0.09319402277469635, + 0.014489078894257545, + 0.03148066624999046, + 0.018458101898431778 + ], + "credit_map": [ + 0.00896327942609787, + 0.013092211447656155, + 0.006089803297072649, + 0.028835570439696312, + 0.015117808245122433, + 0.010445685125887394, + 0.019013943150639534, + 0.015356292016804218, + 0.06391669064760208, + 0.014485666528344154, + 0.009834477677941322, + 0.11242558062076569, + 0.0227387435734272, + 0.01645151525735855, + 0.19302789866924286, + 0.03526686131954193, + 0.004475340712815523, + 0.0029148648027330637, + 0.0028246333822607994, + 0.005078667774796486, + 0.004729405511170626, + 0.003687317483127117, + 0.006895506288856268, + 0.0035953933838754892, + 0.008711636066436768, + 0.0037487572990357876, + 0.003311209613457322, + 0.054627660661935806, + 0.0055948966182768345, + 0.0030337641946971416, + 0.06296762079000473, + 0.008073174394667149, + 0.008819375187158585, + 0.004863782785832882, + 0.004237205255776644, + 0.009236853569746017, + 0.006576125510036945, + 0.007503047119826078, + 0.011781914159655571, + 0.004755686037242413, + 0.0016147027490660548, + 0.001274449285119772, + 0.001216525910422206, + 0.0030521065928041935, + 0.0016688425093889236, + 0.002662298269569874, + 0.002142013283446431, + 0.0016414017882198095, + 0.009247505106031895, + 0.009363017976284027, + 0.006288777571171522, + 0.02108791470527649, + 0.0123721519485116, + 0.021626166999340057, + 0.009188671596348286, + 0.004019818734377623, + 0.001950256759300828, + 0.0012221834622323513, + 0.0010606576688587666, + 0.003202781779691577, + 0.0018071559024974704, + 0.0008259351016022265, + 0.0026391029823571444, + 0.0017810051795095205, + 0.003895771224051714, + 0.0056396606378257275, + 0.0024650460109114647, + 0.004688081331551075, + 0.0034676040522754192, + 0.003807663219049573, + 0.004652352537959814, + 0.0028644874691963196, + 0.002069955226033926, + 0.004482822027057409, + 0.001116885687224567, + 0.0021422086283564568, + 0.002946825698018074, + 0.0015121489996090531, + 0.0020438064821064472, + 0.002143449615687132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.018204161897301674 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785539470613003 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.009785539470613003, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785539470613003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08763402700424194, + 0.18842776119709015, + 0.015480173751711845, + 0.16463404893875122, + 0.3135155439376831, + 0.043199751526117325, + 0.04617607966065407, + 0.038708608597517014, + 0.07707404345273972, + 0.025149932131171227 + ], + "credit_map": [ + 0.008337005041539669, + 0.007005206774920225, + 0.00515682902187109, + 0.012166082859039307, + 0.010789799503982067, + 0.008772882632911205, + 0.022323444485664368, + 0.013082779012620449, + 0.024173907935619354, + 0.008355786092579365, + 0.004797341302037239, + 0.04638155177235603, + 0.010326514951884747, + 0.007557054050266743, + 0.07363543659448624, + 0.013200163841247559, + 0.002024860819801688, + 0.001138444640673697, + 0.0018138408195227385, + 0.0022369888611137867, + 0.0017599464626982808, + 0.001868873368948698, + 0.0028964662924408913, + 0.001740752486512065, + 0.007372183725237846, + 0.005375698208808899, + 0.004821860697120428, + 0.06075676530599594, + 0.008454549126327038, + 0.005409865640103817, + 0.06374350190162659, + 0.008699626661837101, + 0.025782279670238495, + 0.02684088610112667, + 0.02006550319492817, + 0.08763886988162994, + 0.023434843868017197, + 0.02910621650516987, + 0.0845080092549324, + 0.016138942912220955, + 0.0033493589144200087, + 0.011743797920644283, + 0.0025193938054144382, + 0.011373932473361492, + 0.003455190686509013, + 0.00267038750462234, + 0.004883468616753817, + 0.0032042195089161396, + 0.004274647682905197, + 0.003413835074752569, + 0.0033376431092619896, + 0.012602223083376884, + 0.003961989656090736, + 0.010076581500470638, + 0.005786259192973375, + 0.002722900127992034, + 0.0043815369717776775, + 0.004185359459370375, + 0.004858427215367556, + 0.005175790283828974, + 0.004069117829203606, + 0.003496063407510519, + 0.006363920867443085, + 0.006178389303386211, + 0.00695180706679821, + 0.01651185192167759, + 0.0035608820617198944, + 0.012354857288300991, + 0.006209016777575016, + 0.013304505497217178, + 0.010902294889092445, + 0.007278829347342253, + 0.0028135734610259533, + 0.005648595280945301, + 0.001298153423704207, + 0.0029980912804603577, + 0.0036295109894126654, + 0.0024468128103762865, + 0.003569378051906824, + 0.0027458181139081717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.009785539470613003 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809741511940956 + ], + "opposite_logprob_tokens": [ + -3.288097381591797 + ], + "logprob_reward": -0.03809741511940956, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809741511940956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06939332187175751, + 0.7639931440353394, + 0.015587123110890388, + 0.022513989359140396, + 0.03044936992228031, + 0.04598816856741905, + 0.011257454752922058, + 0.015345747582614422, + 0.014726124703884125, + 0.010745545849204063 + ], + "credit_map": [ + 0.007081087213009596, + 0.0027112525422126055, + 0.0029599391855299473, + 0.023879054933786392, + 0.0037903720512986183, + 0.003736483631655574, + 0.02059178426861763, + 0.004643350839614868, + 0.10282527655363083, + 0.025328271090984344, + 0.015526666305959225, + 0.18872469663619995, + 0.037749193608760834, + 0.02684626169502735, + 0.31230413913726807, + 0.05468868836760521, + 0.0028496510349214077, + 0.0009234269382432103, + 0.001261053024791181, + 0.0023986361920833588, + 0.0016554304165765643, + 0.0013446774100884795, + 0.0036338157951831818, + 0.0015204339288175106, + 0.001330497907474637, + 0.0010064092930406332, + 0.000861469772644341, + 0.007161194458603859, + 0.0013306926703080535, + 0.0009342856938019395, + 0.008575896732509136, + 0.001313543994911015, + 0.00491446815431118, + 0.002700628712773323, + 0.005425185896456242, + 0.003280263626947999, + 0.004012971185147762, + 0.0029590274207293987, + 0.004669870715588331, + 0.0024869584012776613, + 0.00467650406062603, + 0.0165049247443676, + 0.0016298498958349228, + 0.012617197819054127, + 0.0026259315200150013, + 0.002508578123524785, + 0.0036513812374323606, + 0.0017738051246851683, + 0.0011244171764701605, + 0.0006705080159008503, + 0.0005880686221644282, + 0.003324591787531972, + 0.0006485652993433177, + 0.002950383350253105, + 0.0013611081521958113, + 0.0005898131057620049, + 0.0012923643225803971, + 0.001896057859994471, + 0.002280554734170437, + 0.0020373875740915537, + 0.0017713898560032248, + 0.0014431079616770148, + 0.0021121264435350895, + 0.002512760926038027, + 0.001737025217153132, + 0.001910111983306706, + 0.0009969481034204364, + 0.001465915236622095, + 0.0016808934742584825, + 0.004190467298030853, + 0.0014154304517433047, + 0.0013293334050104022, + 0.0012252364540472627, + 0.0015436505200341344, + 0.0009413431398570538, + 0.0008895621867850423, + 0.002143916906788945, + 0.0013604828855022788, + 0.0014538202667608857, + 0.0011875336058437824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.03809741511940956 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, static close-up captures the meticulous act of seasoning. A steady hand, with clean, manicured nails, hovers deliberately over a dark, heavy pot filled with a rich, simmering, deep-orange vegetable broth, from which rises a gentle stream of steam. Between thumb and index finger, a pinch of fine, earthy-green dried plant powder is held. Gently rubbing the fingers together, the powder drizzles down in a delicate, controlled pattern. The camera maintains sharp focus on the falling particles as they gently land on the broth's glistening surface, floating briefly like fine dust before beginning to dissolve and swirl into the vibrant liquid. The scene is illuminated by a warm, soft light from an off-camera source, highlighting the texture of the powder and the rich color of the broth, contrasting sharply with the dark, out-of-focus grains of the wooden table below, creating an atmosphere of concentration, tranquility, and appetizing culinary delight.", + "seed": 3901, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_11/index_662_seed_3901_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016165563464164735, + "vlm_logprob": -0.016165563464164735, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.016165563464164735, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02479737251996994 + ], + "opposite_logprob_tokens": [ + -6.524797439575195 + ], + "logprob_reward": -0.02479737251996994, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.02479737251996994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39223670959472656, + 0.10357173532247543, + 0.06239696964621544, + 0.07292962819337845, + 0.0594266839325428, + 0.06761389970779419, + 0.05695813149213791, + 0.05947038158774376, + 0.061507485806941986, + 0.06388842314481735 + ], + "credit_map": [ + 0.10163942724466324, + 0.04162219166755676, + 0.016515593975782394, + 0.040608279407024384, + 0.07106833159923553, + 0.023012632504105568, + 0.07447353005409241, + 0.023296723142266273, + 0.018186144530773163, + 0.0077627841383218765, + 0.006250209640711546, + 0.014409385621547699, + 0.011979402974247932, + 0.011726615019142628, + 0.020437179133296013, + 0.012820012867450714, + 0.00990817416459322, + 0.0053603569976985455, + 0.00601945212110877, + 0.008523261174559593, + 0.011258981190621853, + 0.005179944448173046, + 0.010737773030996323, + 0.0054090251214802265, + 0.011801720596849918, + 0.004916100762784481, + 0.005054938141256571, + 0.015863971784710884, + 0.00736550148576498, + 0.007111131679266691, + 0.015499679371714592, + 0.005316589027643204, + 0.008829895406961441, + 0.004204003605991602, + 0.005131642799824476, + 0.01050430629402399, + 0.006673368625342846, + 0.00536283990368247, + 0.013778441585600376, + 0.004942184314131737, + 0.011849446222186089, + 0.005510675720870495, + 0.0037416222039610147, + 0.009868820197880268, + 0.012623629532754421, + 0.005534878931939602, + 0.01429617591202259, + 0.004188652150332928, + 0.007784810848534107, + 0.005217743571847677, + 0.0033417236991226673, + 0.008572524413466454, + 0.010709522292017937, + 0.004992877133190632, + 0.010931416414678097, + 0.0054075149819254875, + 0.013666640035808086, + 0.008121955208480358, + 0.0030615346040576696, + 0.008235017769038677, + 0.009157351218163967, + 0.005079316906630993, + 0.00817451998591423, + 0.003974048420786858, + 0.00940555427223444, + 0.00905564147979021, + 0.004142428282648325, + 0.006742435973137617, + 0.014408579096198082, + 0.005532536190003157, + 0.00889330729842186, + 0.0033270043786615133, + 0.010544510558247566, + 0.009150190278887749, + 0.004303245339542627, + 0.00785959605127573, + 0.011156314052641392, + 0.004568060394376516, + 0.012841813266277313, + 0.0034646892454475164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.02479737251996994 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011105983518064022 + ], + "opposite_logprob_tokens": [ + -4.511106014251709 + ], + "logprob_reward": -0.011105983518064022, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011105983518064022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4474916160106659, + 0.10887369513511658, + 0.17882563173770905, + 0.03027421608567238, + 0.0436699315905571, + 0.028525765985250473, + 0.023175276815891266, + 0.07413475215435028, + 0.031386926770210266, + 0.03364221006631851 + ], + "credit_map": [ + 0.12925629317760468, + 0.01974242925643921, + 0.015187185257673264, + 0.06546646356582642, + 0.026383237913250923, + 0.017318882048130035, + 0.1552589386701584, + 0.018878180533647537, + 0.016971061006188393, + 0.007739261724054813, + 0.010340649634599686, + 0.014640433713793755, + 0.013959298841655254, + 0.010626382194459438, + 0.01850554533302784, + 0.016091061756014824, + 0.03050992079079151, + 0.013517457991838455, + 0.024011196568608284, + 0.02767479419708252, + 0.02425250969827175, + 0.013575281016528606, + 0.031072644516825676, + 0.014211832545697689, + 0.004807152785360813, + 0.002029897877946496, + 0.0016113623278215528, + 0.005938286427408457, + 0.003971044439822435, + 0.0027754984330385923, + 0.0070260618813335896, + 0.0021149124950170517, + 0.007731701713055372, + 0.0033523067831993103, + 0.0037116582971066236, + 0.008358046412467957, + 0.005481790751218796, + 0.00325050950050354, + 0.008333075791597366, + 0.003450843272730708, + 0.005204483401030302, + 0.0023142884019762278, + 0.002448484767228365, + 0.0037072303239256144, + 0.00607035169377923, + 0.002487465739250183, + 0.004420164041221142, + 0.0018732971511781216, + 0.003522573271766305, + 0.0020771618001163006, + 0.002083063358440995, + 0.0028710290789604187, + 0.005587705411016941, + 0.0015685734106227756, + 0.004039332736283541, + 0.0014258365845307708, + 0.017747754231095314, + 0.012798390351235867, + 0.0046136388555169106, + 0.008678808808326721, + 0.009795775637030602, + 0.006311536300927401, + 0.01042974554002285, + 0.0037591035943478346, + 0.0052499123848974705, + 0.0021187751553952694, + 0.0029018514323979616, + 0.005332206841558218, + 0.003518496174365282, + 0.002085867803543806, + 0.00709145562723279, + 0.003088364377617836, + 0.004788957070559263, + 0.003956719301640987, + 0.002758721588179469, + 0.0038628161419183016, + 0.007398056797683239, + 0.0026950850151479244, + 0.005788673181086779, + 0.0023931812029331923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.011105983518064022 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420446764677763 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.01420446764677763, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420446764677763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.481830894947052, + 0.08557356894016266, + 0.07022879272699356, + 0.02400157041847706, + 0.05995040014386177, + 0.045181769877672195, + 0.10446076095104218, + 0.04864770546555519, + 0.03701069951057434, + 0.043113891035318375 + ], + "credit_map": [ + 0.156328946352005, + 0.023479020223021507, + 0.011478310450911522, + 0.05812959372997284, + 0.02788109891116619, + 0.02075301669538021, + 0.15837804973125458, + 0.025402860715985298, + 0.015252572484314442, + 0.007205562200397253, + 0.006225187797099352, + 0.01004075724631548, + 0.012280887924134731, + 0.008334786631166935, + 0.018185868859291077, + 0.008047946728765965, + 0.010557380504906178, + 0.005101981107145548, + 0.0037004214245826006, + 0.012555175460875034, + 0.012483975850045681, + 0.005633024964481592, + 0.015574564225971699, + 0.004622268490493298, + 0.0034976915922015905, + 0.002352913608774543, + 0.0015869425842538476, + 0.0031076238956302404, + 0.005103003699332476, + 0.001520100748166442, + 0.005011105444282293, + 0.0018221872160211205, + 0.009903638623654842, + 0.0053899637423455715, + 0.0038155748043209314, + 0.007816684432327747, + 0.01248431857675314, + 0.00659880181774497, + 0.00928092934191227, + 0.004660491365939379, + 0.008807801641523838, + 0.005425752606242895, + 0.002392255235463381, + 0.0054677571170032024, + 0.009099524468183517, + 0.005383752752095461, + 0.006218479014933109, + 0.0023864482063800097, + 0.015056656673550606, + 0.012204541824758053, + 0.005625994876027107, + 0.010325158946216106, + 0.031469132751226425, + 0.00968873593956232, + 0.014991960488259792, + 0.005098583176732063, + 0.00843226071447134, + 0.007696045096963644, + 0.003600205760449171, + 0.007620495278388262, + 0.00595178734511137, + 0.00396844232454896, + 0.00900145061314106, + 0.002377019729465246, + 0.0072269137017428875, + 0.0036512629594653845, + 0.0024107848294079304, + 0.005110377911478281, + 0.00709690572693944, + 0.0034631043672561646, + 0.006232254672795534, + 0.0018190934788435698, + 0.0072427792474627495, + 0.0039235069416463375, + 0.0017602102598175406, + 0.004522629547864199, + 0.010149937123060226, + 0.0032191486097872257, + 0.009288009256124496, + 0.0030076690018177032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.01420446764677763 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540297582745552 + ], + "opposite_logprob_tokens": [ + -4.387540340423584 + ], + "logprob_reward": -0.012540297582745552, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540297582745552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5425059795379639, + 0.08626127243041992, + 0.040992531925439835, + 0.04025457799434662, + 0.021872354671359062, + 0.025182602927088737, + 0.10661454498767853, + 0.05964864417910576, + 0.022562773898243904, + 0.054104696959257126 + ], + "credit_map": [ + 0.17248837649822235, + 0.0331377312541008, + 0.0133349085226655, + 0.05759534612298012, + 0.04933071881532669, + 0.024491844698786736, + 0.17143017053604126, + 0.020696938037872314, + 0.01300379540771246, + 0.006331286858767271, + 0.006297019775956869, + 0.01544830109924078, + 0.00955164898186922, + 0.006583931390196085, + 0.022193998098373413, + 0.0068512908183038235, + 0.007010729052126408, + 0.00312206894159317, + 0.003369321348145604, + 0.005456508137285709, + 0.005804718006402254, + 0.004679002799093723, + 0.007996570318937302, + 0.0035536137875169516, + 0.006755420938134193, + 0.003701218403875828, + 0.002541237510740757, + 0.005949149373918772, + 0.007022402249276638, + 0.004977889358997345, + 0.006527968216687441, + 0.0027792947366833687, + 0.003847312880679965, + 0.002139049582183361, + 0.0014508680906146765, + 0.0030592763796448708, + 0.0038578540552407503, + 0.0018519345903769135, + 0.004304063506424427, + 0.0013619940727949142, + 0.005946554243564606, + 0.0022939571645110846, + 0.0013644522987306118, + 0.003194512100890279, + 0.004236703738570213, + 0.0028170249424874783, + 0.003748722840100527, + 0.0015806745504960418, + 0.015550472773611546, + 0.01162951160222292, + 0.004186379257589579, + 0.009723436087369919, + 0.03661810979247093, + 0.008288031443953514, + 0.014906884171068668, + 0.005711723119020462, + 0.009130028076469898, + 0.009803741239011288, + 0.004087660927325487, + 0.008820832706987858, + 0.008124807849526405, + 0.00577801326289773, + 0.010589640587568283, + 0.003313913941383362, + 0.0038594347424805164, + 0.002916022902354598, + 0.0020741785410791636, + 0.0015527664218097925, + 0.005971790757030249, + 0.0019013216951861978, + 0.0030346419662237167, + 0.0012526155915111303, + 0.009742979891598225, + 0.0070544215850532055, + 0.00259789964184165, + 0.004160612355917692, + 0.01495834719389677, + 0.005159304942935705, + 0.007887176238000393, + 0.002543955808505416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.012540297582745552 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018179696053266525 + ], + "opposite_logprob_tokens": [ + -4.018179893493652 + ], + "logprob_reward": -0.018179696053266525, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018179696053266525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25737088918685913, + 0.04727189242839813, + 0.10970013588666916, + 0.06872761249542236, + 0.043138522654771805, + 0.0472395084798336, + 0.060332734137773514, + 0.10571492463350296, + 0.0404977947473526, + 0.22000595927238464 + ], + "credit_map": [ + 0.07017174363136292, + 0.013864605687558651, + 0.009446382522583008, + 0.03396371006965637, + 0.017883770167827606, + 0.0161481611430645, + 0.07522421330213547, + 0.020668301731348038, + 0.0061421883292496204, + 0.004408290144056082, + 0.004393186420202255, + 0.008330649696290493, + 0.006491669453680515, + 0.004111021291464567, + 0.009554468095302582, + 0.003840420162305236, + 0.01840926706790924, + 0.009100640192627907, + 0.013901611790060997, + 0.014537373557686806, + 0.016548512503504753, + 0.0079550351947546, + 0.01804523542523384, + 0.011202456429600716, + 0.012042335234582424, + 0.006210730876773596, + 0.003442096756771207, + 0.013917040079832077, + 0.009998832829296589, + 0.006054274272173643, + 0.012557919137179852, + 0.004504379816353321, + 0.008383354172110558, + 0.0029524178244173527, + 0.0026601627469062805, + 0.00983897689729929, + 0.003987646661698818, + 0.0031859660521149635, + 0.008894113823771477, + 0.003235885873436928, + 0.00847418699413538, + 0.0041206227615475655, + 0.0024707596749067307, + 0.005745535250753164, + 0.009367736056447029, + 0.0045372373424470425, + 0.010051519609987736, + 0.0024719114881008863, + 0.009362020529806614, + 0.005696908105164766, + 0.00313010485842824, + 0.005862426478415728, + 0.015009303577244282, + 0.007475146092474461, + 0.010791195556521416, + 0.003005632199347019, + 0.027738524600863457, + 0.01631329022347927, + 0.006138407159596682, + 0.011486062780022621, + 0.016957929357886314, + 0.009124265983700752, + 0.013092206791043282, + 0.004864240065217018, + 0.008724714629352093, + 0.006691441405564547, + 0.0023005229886621237, + 0.003872556146234274, + 0.008808692917227745, + 0.0041989474557340145, + 0.0033048854675143957, + 0.0025960318744182587, + 0.03668617829680443, + 0.037735823541879654, + 0.008956992998719215, + 0.01509349700063467, + 0.058662299066782, + 0.02752169594168663, + 0.02513408288359642, + 0.010215364396572113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.018179696053266525 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, static close-up captures the meticulous act of seasoning. A steady hand, with clean, manicured nails, hovers deliberately over a dark, heavy pot filled with a rich, simmering, deep-orange vegetable broth, from which rises a gentle stream of steam. Between thumb and index finger, a pinch of fine, earthy-green dried plant powder is held. Gently rubbing the fingers together, the powder drizzles down in a delicate, controlled pattern. The camera maintains sharp focus on the falling particles as they gently land on the broth's glistening surface, floating briefly like fine dust before beginning to dissolve and swirl into the vibrant liquid. The scene is illuminated by a warm, soft light from an off-camera source, highlighting the texture of the powder and the rich color of the broth, contrasting sharply with the dark, out-of-focus grains of the wooden table below, creating an atmosphere of concentration, tranquility, and appetizing culinary delight.", + "seed": 3902, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_11/index_662_seed_3902_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.017266160063445567, + "vlm_logprob": -0.017266160063445567, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.017266160063445567, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02522813156247139 + ], + "opposite_logprob_tokens": [ + -5.400228023529053 + ], + "logprob_reward": -0.02522813156247139, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.02522813156247139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12929825484752655, + 0.0652899369597435, + 0.04412754252552986, + 0.18814198672771454, + 0.06477662920951843, + 0.21727867424488068, + 0.07416790723800659, + 0.06707552820444107, + 0.0812075212597847, + 0.06863605231046677 + ], + "credit_map": [ + 0.01022165548056364, + 0.012195824645459652, + 0.0069069936871528625, + 0.032202884554862976, + 0.012825392186641693, + 0.014699151739478111, + 0.023814665153622627, + 0.016431689262390137, + 0.011218046769499779, + 0.007613504771143198, + 0.00692447554320097, + 0.007949065417051315, + 0.007815430872142315, + 0.009691719897091389, + 0.007069018669426441, + 0.007008676417171955, + 0.005921801086515188, + 0.0034410671796649694, + 0.003847008803859353, + 0.005214793607592583, + 0.006197997368872166, + 0.006649337708950043, + 0.004689068533480167, + 0.00816646683961153, + 0.045100390911102295, + 0.019605524837970734, + 0.0066786943934857845, + 0.028408760204911232, + 0.012240943498909473, + 0.010729166679084301, + 0.05298047512769699, + 0.012398038059473038, + 0.012253116816282272, + 0.019150426611304283, + 0.0022862758487462997, + 0.010803062468767166, + 0.00629652664065361, + 0.004903963766992092, + 0.005230039823800325, + 0.003853212809190154, + 0.05427159368991852, + 0.014594010077416897, + 0.00882150698453188, + 0.05856585502624512, + 0.024012787267565727, + 0.012719460763037205, + 0.023084070533514023, + 0.021209366619586945, + 0.0065287211909890175, + 0.004486883524805307, + 0.005788450129330158, + 0.02642429806292057, + 0.007989795878529549, + 0.003722622524946928, + 0.011243187822401524, + 0.007983947172760963, + 0.009556993842124939, + 0.0060982475988566875, + 0.003987140022218227, + 0.017001593485474586, + 0.010958293452858925, + 0.005922640673816204, + 0.006708009634166956, + 0.006842607166618109, + 0.013635589741170406, + 0.010267813690006733, + 0.004535555373877287, + 0.012566437013447285, + 0.01859516091644764, + 0.00780080072581768, + 0.006865825969725847, + 0.006940329447388649, + 0.0074874963611364365, + 0.006830626633018255, + 0.0045658196322619915, + 0.005998574197292328, + 0.01848210208117962, + 0.008330412209033966, + 0.00970679521560669, + 0.007234220393002033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.02522813156247139 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004123044200241566 + ], + "opposite_logprob_tokens": [ + -5.504123210906982 + ], + "logprob_reward": -0.004123044200241566, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004123044200241566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11702457070350647, + 0.27861788868904114, + 0.054374467581510544, + 0.15676917135715485, + 0.1302267462015152, + 0.0962182953953743, + 0.09109124541282654, + 0.023997295647859573, + 0.026133403182029724, + 0.025546912103891373 + ], + "credit_map": [ + 0.02803238295018673, + 0.013932362198829651, + 0.008602641522884369, + 0.02908392809331417, + 0.007027851417660713, + 0.007475075777620077, + 0.011340553872287273, + 0.01152978464961052, + 0.05871996656060219, + 0.010493154637515545, + 0.015735184773802757, + 0.11049838364124298, + 0.009324428625404835, + 0.015639353543519974, + 0.03772251307964325, + 0.020484885200858116, + 0.009752186015248299, + 0.004994980059564114, + 0.0037827908527106047, + 0.007412398234009743, + 0.010553506202995777, + 0.009076341986656189, + 0.003352320520207286, + 0.005449943710118532, + 0.03275073319673538, + 0.007111941464245319, + 0.006085896398872137, + 0.02363435924053192, + 0.015916263684630394, + 0.011121612042188644, + 0.04865147918462753, + 0.011496875435113907, + 0.013965931721031666, + 0.036292802542448044, + 0.003968372009694576, + 0.03863828629255295, + 0.0075536794029176235, + 0.008247329853475094, + 0.01353684812784195, + 0.00802349392324686, + 0.022703014314174652, + 0.006746071390807629, + 0.004681296646595001, + 0.029659561812877655, + 0.007114699576050043, + 0.006632255855947733, + 0.010525723919272423, + 0.008155674673616886, + 0.008960301987826824, + 0.010184293612837791, + 0.004577521234750748, + 0.022072769701480865, + 0.013180405832827091, + 0.011618868447840214, + 0.012014072388410568, + 0.00848301313817501, + 0.003099605208262801, + 0.003588091116398573, + 0.0016541456570848823, + 0.002987950574606657, + 0.003949969075620174, + 0.004535580985248089, + 0.0018037310801446438, + 0.0023782209027558565, + 0.0037905701901763678, + 0.0044070822186768055, + 0.0015658034244552255, + 0.0025099844206124544, + 0.007073737680912018, + 0.0031168360728770494, + 0.0016564038814976811, + 0.0020129841286689043, + 0.004567121621221304, + 0.007783467415720224, + 0.001464195898734033, + 0.001538355485536158, + 0.004535633139312267, + 0.0023134094662964344, + 0.001284515019506216, + 0.002060214290395379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.004123044200241566 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018205096945166588 + ], + "opposite_logprob_tokens": [ + -4.018205165863037 + ], + "logprob_reward": -0.018205096945166588, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018205096945166588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07279057800769806, + 0.3248593807220459, + 0.07191872596740723, + 0.10223017632961273, + 0.09480151534080505, + 0.09756409376859665, + 0.0702773854136467, + 0.0391329787671566, + 0.0763106718659401, + 0.05011451616883278 + ], + "credit_map": [ + 0.006003046873956919, + 0.0064209504052996635, + 0.004318444989621639, + 0.016358334571123123, + 0.012944691814482212, + 0.006382671184837818, + 0.012203408405184746, + 0.008159030228853226, + 0.07122977077960968, + 0.016216877847909927, + 0.017661862075328827, + 0.11695484071969986, + 0.02394959330558777, + 0.017050020396709442, + 0.039098452776670456, + 0.022697970271110535, + 0.016446352005004883, + 0.011324453167617321, + 0.003795557888224721, + 0.006482783239334822, + 0.012117842212319374, + 0.009330476634204388, + 0.004701131954789162, + 0.007720122579485178, + 0.009401688352227211, + 0.01963358372449875, + 0.006163493730127811, + 0.0077447122894227505, + 0.022657198831439018, + 0.01675921119749546, + 0.01326802559196949, + 0.006602265872061253, + 0.021353131160140038, + 0.03435920923948288, + 0.002831757767125964, + 0.013663176447153091, + 0.00652112253010273, + 0.00513352220878005, + 0.00540527980774641, + 0.005534314550459385, + 0.01846572570502758, + 0.008048830553889275, + 0.004762987606227398, + 0.020206362009048462, + 0.018521906808018684, + 0.010822095908224583, + 0.00879229698330164, + 0.007943887263536453, + 0.007689213380217552, + 0.010086151771247387, + 0.005750617478042841, + 0.011273037642240524, + 0.012021712958812714, + 0.009657703340053558, + 0.007451369892805815, + 0.006347574293613434, + 0.0050489711575210094, + 0.004015836399048567, + 0.002372221788391471, + 0.00939816702157259, + 0.005726958159357309, + 0.0038107899017632008, + 0.005433490965515375, + 0.0033265429083257914, + 0.007611797656863928, + 0.010423751547932625, + 0.004995507188141346, + 0.006180762313306332, + 0.026534177362918854, + 0.010626702569425106, + 0.004619764629751444, + 0.005318210925906897, + 0.010981826111674309, + 0.01745695248246193, + 0.0028716844972223043, + 0.0028778882697224617, + 0.006057806313037872, + 0.004998138640075922, + 0.002321347128599882, + 0.002548874355852604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.018205096945166588 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192103132605553 + ], + "opposite_logprob_tokens": [ + -4.018192291259766 + ], + "logprob_reward": -0.018192103132605553, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018192103132605553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04747149720788002, + 0.20678065717220306, + 0.02856396697461605, + 0.03625667467713356, + 0.09522506594657898, + 0.04957586154341698, + 0.1265818476676941, + 0.05185287818312645, + 0.3242896497249603, + 0.03340187668800354 + ], + "credit_map": [ + 0.010504897683858871, + 0.008125362917780876, + 0.0029335820581763983, + 0.005652293562889099, + 0.005893032066524029, + 0.005047510843724012, + 0.004179794806987047, + 0.005135020706802607, + 0.050144921988248825, + 0.008529335260391235, + 0.011928199790418148, + 0.08306076377630234, + 0.005212265066802502, + 0.00623739417642355, + 0.02604476734995842, + 0.015622999519109726, + 0.004343333654105663, + 0.0039953445084393024, + 0.001549857435747981, + 0.0031369831413030624, + 0.0062602185644209385, + 0.003391657955944538, + 0.0021857500541955233, + 0.0037008230574429035, + 0.006229784805327654, + 0.0020740230102092028, + 0.0017348205437883735, + 0.005131988320499659, + 0.005158036015927792, + 0.003699764609336853, + 0.009530633687973022, + 0.0026976214721798897, + 0.024297021329402924, + 0.0343664176762104, + 0.0031293027568608522, + 0.010410360991954803, + 0.0061149317771196365, + 0.007180300075560808, + 0.004809296224266291, + 0.004917438607662916, + 0.011453972198069096, + 0.002489380305632949, + 0.0022937762551009655, + 0.016135036945343018, + 0.004875068552792072, + 0.0029527489095926285, + 0.005726587492972612, + 0.003649292513728142, + 0.012480438686907291, + 0.0117700956761837, + 0.00762530043721199, + 0.02250906266272068, + 0.029223453253507614, + 0.014911561273038387, + 0.01200435683131218, + 0.01605757512152195, + 0.008830017410218716, + 0.0032812091521918774, + 0.003230456728488207, + 0.01655571721494198, + 0.007492588832974434, + 0.0031406732741743326, + 0.005686879623681307, + 0.0036353361792862415, + 0.03393200784921646, + 0.048220373690128326, + 0.020862026140093803, + 0.020905952900648117, + 0.11404158174991608, + 0.04656408354640007, + 0.018257368355989456, + 0.02150627411901951, + 0.006672725081443787, + 0.009714801795780659, + 0.0016995870973914862, + 0.0016081607900559902, + 0.005114053841680288, + 0.00436200387775898, + 0.0022543964441865683, + 0.001976148923859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.018192103132605553 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020582424476742744 + ], + "opposite_logprob_tokens": [ + -3.895582437515259 + ], + "logprob_reward": -0.020582424476742744, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020582424476742744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055806711316108704, + 0.29123470187187195, + 0.020557938143610954, + 0.012406707741320133, + 0.5331332683563232, + 0.036459941416978836, + 0.024576915428042412, + 0.004623576533049345, + 0.014831704087555408, + 0.006368575152009726 + ], + "credit_map": [ + 0.01682254672050476, + 0.011705187149345875, + 0.0030519741121679544, + 0.008748789317905903, + 0.0033559638541191816, + 0.0030594728887081146, + 0.003913954831659794, + 0.0051488266326487064, + 0.06666425615549088, + 0.011460550129413605, + 0.01667870208621025, + 0.1196942999958992, + 0.006224965676665306, + 0.008264019154012203, + 0.039625559002161026, + 0.022622350603342056, + 0.0036680069752037525, + 0.001286120037548244, + 0.0009888132335618138, + 0.004304578993469477, + 0.004183568991720676, + 0.0026781607884913683, + 0.0014195548137649894, + 0.002029133727774024, + 0.0019685698207467794, + 0.001092438236810267, + 0.0008677265141159296, + 0.002082380000501871, + 0.0014787387335672975, + 0.0014394704485312104, + 0.0021169576793909073, + 0.0013604258419945836, + 0.1045340746641159, + 0.19076727330684662, + 0.010439198464155197, + 0.11889653652906418, + 0.026279224082827568, + 0.024783460423350334, + 0.038807209581136703, + 0.018626270815730095, + 0.006154775153845549, + 0.004538728389889002, + 0.0027619011234492064, + 0.005111347883939743, + 0.006939193699508905, + 0.003938897978514433, + 0.0031040015164762735, + 0.003911098465323448, + 0.003428427502512932, + 0.003370121819898486, + 0.0017192389350384474, + 0.003625997109338641, + 0.004416161682456732, + 0.003474932862445712, + 0.002089895773679018, + 0.002452139277011156, + 0.0005668221274390817, + 0.0006525726057589054, + 0.00032453055609948933, + 0.0008931511547416449, + 0.000633869378361851, + 0.0006391439237631857, + 0.00042560260044410825, + 0.0004878845647908747, + 0.0025264970026910305, + 0.003300898475572467, + 0.0007970224833115935, + 0.001028171624056995, + 0.003777541685849428, + 0.0015678689815104008, + 0.0008301542256958783, + 0.0010035489685833454, + 0.0006275345804169774, + 0.0009231502772308886, + 0.00033616070868447423, + 0.0005472908378578722, + 0.0015848164912313223, + 0.001467717345803976, + 0.0003821005520876497, + 0.00049980403855443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.020582424476742744 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, static close-up captures the meticulous act of seasoning. A steady hand, with clean, manicured nails, hovers deliberately over a dark, heavy pot filled with a rich, simmering, deep-orange vegetable broth, from which rises a gentle stream of steam. Between thumb and index finger, a pinch of fine, earthy-green dried plant powder is held. Gently rubbing the fingers together, the powder drizzles down in a delicate, controlled pattern. The camera maintains sharp focus on the falling particles as they gently land on the broth's glistening surface, floating briefly like fine dust before beginning to dissolve and swirl into the vibrant liquid. The scene is illuminated by a warm, soft light from an off-camera source, highlighting the texture of the powder and the rich color of the broth, contrasting sharply with the dark, out-of-focus grains of the wooden table below, creating an atmosphere of concentration, tranquility, and appetizing culinary delight.", + "seed": 3903, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_11/index_662_seed_3903_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.024477898422628642, + "vlm_logprob": -0.024477898422628642, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.024477898422628642, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03430221974849701 + ], + "opposite_logprob_tokens": [ + -4.534302234649658 + ], + "logprob_reward": -0.03430221974849701, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.03430221974849701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05357186123728752, + 0.03563421219587326, + 0.046575456857681274, + 0.024333490058779716, + 0.04932534694671631, + 0.2568734884262085, + 0.21775734424591064, + 0.1640496402978897, + 0.11103212088346481, + 0.04084707051515579 + ], + "credit_map": [ + 0.008098723366856575, + 0.006646318361163139, + 0.006532656028866768, + 0.004920769017189741, + 0.006632068194448948, + 0.007832883857190609, + 0.007079863455146551, + 0.005828576628118753, + 0.0048467181622982025, + 0.0050209746696054935, + 0.00395302614197135, + 0.0037291408516466618, + 0.006712228525429964, + 0.005201320629566908, + 0.0034076382871717215, + 0.002763167954981327, + 0.007411114405840635, + 0.00676302146166563, + 0.003348381957039237, + 0.0028648863080888987, + 0.012456158176064491, + 0.006384562235325575, + 0.0039292071014642715, + 0.003418123349547386, + 0.004888819996267557, + 0.002585633425042033, + 0.003595495829358697, + 0.001825452782213688, + 0.004178939387202263, + 0.003223774954676628, + 0.0021417043171823025, + 0.0018936704145744443, + 0.00857243500649929, + 0.004040454048663378, + 0.005430925637483597, + 0.003918314352631569, + 0.006692387163639069, + 0.009100088849663734, + 0.008056709542870522, + 0.0035140335094183683, + 0.09498436003923416, + 0.011202202178537846, + 0.013283755630254745, + 0.006739848293364048, + 0.019657541066408157, + 0.0908091589808464, + 0.010764851234853268, + 0.009431750513613224, + 0.023680657148361206, + 0.03888344764709473, + 0.021601848304271698, + 0.020054731518030167, + 0.04812559112906456, + 0.025802049785852432, + 0.020679915323853493, + 0.018929101526737213, + 0.021375354379415512, + 0.016773831099271774, + 0.01860540173947811, + 0.02142234891653061, + 0.021920479834079742, + 0.018996847793459892, + 0.028012357652187347, + 0.01694301702082157, + 0.022869836539030075, + 0.0166655071079731, + 0.008289237506687641, + 0.007162655238062143, + 0.023646073415875435, + 0.014145913533866405, + 0.009639116004109383, + 0.008613775484263897, + 0.005435083992779255, + 0.005506046116352081, + 0.006700103171169758, + 0.00382058322429657, + 0.006247803568840027, + 0.006356964353471994, + 0.004178621340543032, + 0.0026018666103482246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.03430221974849701 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00597974332049489 + ], + "opposite_logprob_tokens": [ + -5.130979537963867 + ], + "logprob_reward": -0.00597974332049489, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00597974332049489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11317823827266693, + 0.07095672935247421, + 0.028488004580140114, + 0.056251004338264465, + 0.01868559420108795, + 0.26756882667541504, + 0.08120261877775192, + 0.03978288546204567, + 0.14140072464942932, + 0.1824854016304016 + ], + "credit_map": [ + 0.015244534239172935, + 0.016344092786312103, + 0.011284288950264454, + 0.009119085036218166, + 0.01153489202260971, + 0.01873258501291275, + 0.017267121002078056, + 0.013651634566485882, + 0.006471243221312761, + 0.005452565848827362, + 0.013907227665185928, + 0.008216079324483871, + 0.00697437534108758, + 0.009387553669512272, + 0.012652870267629623, + 0.007894805632531643, + 0.0035554564092308283, + 0.004110062029212713, + 0.0034223217517137527, + 0.0029576413799077272, + 0.006272891070693731, + 0.0030400147661566734, + 0.0029928951989859343, + 0.0021367198787629604, + 0.01810384914278984, + 0.0031258780509233475, + 0.003797049866989255, + 0.0025374209508299828, + 0.0046175033785402775, + 0.016367990523576736, + 0.004699400160461664, + 0.003001908538863063, + 0.002273141173645854, + 0.0013486851239576936, + 0.0027247907128185034, + 0.0018228713888674974, + 0.002327590249478817, + 0.0020158360712230206, + 0.0034474243875592947, + 0.0027252547442913055, + 0.09890633076429367, + 0.008389024063944817, + 0.014288782142102718, + 0.007717209868133068, + 0.019674301147460938, + 0.09745901077985764, + 0.010365206748247147, + 0.01076896395534277, + 0.011435759253799915, + 0.010175933130085468, + 0.006668001879006624, + 0.006933680269867182, + 0.013247936964035034, + 0.009089121595025063, + 0.01649302802979946, + 0.007159150205552578, + 0.010353862307965755, + 0.004501081071794033, + 0.004011737182736397, + 0.003441195236518979, + 0.005785864777863026, + 0.007162620313465595, + 0.0023406571708619595, + 0.0021858676336705685, + 0.028255876153707504, + 0.01887170970439911, + 0.02019202709197998, + 0.014184867031872272, + 0.018943168222904205, + 0.016503948718309402, + 0.013758049346506596, + 0.01069106999784708, + 0.03007282130420208, + 0.026236718520522118, + 0.02316119335591793, + 0.01679646410048008, + 0.02753361128270626, + 0.018885761499404907, + 0.02394026704132557, + 0.015858562663197517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.00597974332049489 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029816951602697372 + ], + "opposite_logprob_tokens": [ + -3.5298168659210205 + ], + "logprob_reward": -0.029816951602697372, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029816951602697372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024344293400645256, + 0.027013178914785385, + 0.020852504298090935, + 0.009384546428918839, + 0.015651844441890717, + 0.7846841812133789, + 0.02156680077314377, + 0.017472684383392334, + 0.05671752616763115, + 0.02231244184076786 + ], + "credit_map": [ + 0.0031460507307201624, + 0.0037416955456137657, + 0.002244445262476802, + 0.002575283171609044, + 0.00302594481036067, + 0.0028673906344920397, + 0.004043037071824074, + 0.0027004461735486984, + 0.0030529615469276905, + 0.003321333322674036, + 0.003923157230019569, + 0.0028880625031888485, + 0.005225493106991053, + 0.004070057533681393, + 0.0026187007315456867, + 0.001913412124849856, + 0.0037493889685720205, + 0.0030079386197030544, + 0.001644285162910819, + 0.001518390839919448, + 0.003900450188666582, + 0.00415398832410574, + 0.0016579360235482454, + 0.0012201267527416348, + 0.0015121526084840298, + 0.001005376921966672, + 0.0011450046440586448, + 0.0006860458524897695, + 0.0015147383091971278, + 0.0011856274213641882, + 0.0015591507544741035, + 0.0007764499168843031, + 0.0034117316827178, + 0.001391950878314674, + 0.0015059823635965586, + 0.0014207343338057399, + 0.0023234980180859566, + 0.002742412267252803, + 0.0015661717625334859, + 0.001289362320676446, + 0.30016040802001953, + 0.02429938316345215, + 0.03897438198328018, + 0.01889578253030777, + 0.053995728492736816, + 0.29070302844047546, + 0.027716757729649544, + 0.02993873879313469, + 0.002736562630161643, + 0.0025873295962810516, + 0.002355493837967515, + 0.0019366804044693708, + 0.0036683930084109306, + 0.003183150663971901, + 0.003342964919283986, + 0.0017562260618433356, + 0.002457382157444954, + 0.0023796381428837776, + 0.001722279703244567, + 0.0016486691311001778, + 0.003872471395879984, + 0.002157490700483322, + 0.0015872619114816189, + 0.0016474910080432892, + 0.005043177865445614, + 0.00889480672776699, + 0.006852810736745596, + 0.00676736980676651, + 0.01199567411094904, + 0.007029770873486996, + 0.00514879310503602, + 0.00498512526974082, + 0.003435396822169423, + 0.004338032100349665, + 0.0035335798747837543, + 0.0016973904566839337, + 0.0035862179938703775, + 0.0020025442354381084, + 0.002289250958710909, + 0.0014300288166850805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.029816951602697372 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197062700986862 + ], + "opposite_logprob_tokens": [ + -4.264196872711182 + ], + "logprob_reward": -0.014197062700986862, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197062700986862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0538722462952137, + 0.05039070546627045, + 0.03509846329689026, + 0.043895404785871506, + 0.04144000634551048, + 0.4591636061668396, + 0.038506414741277695, + 0.04700962081551552, + 0.19705863296985626, + 0.03356492891907692 + ], + "credit_map": [ + 0.006598257925361395, + 0.005392617080360651, + 0.007532911375164986, + 0.007905445992946625, + 0.0059915874153375626, + 0.006968811620026827, + 0.007687795907258987, + 0.005794812925159931, + 0.006354614160954952, + 0.009415552951395512, + 0.005097588989883661, + 0.005377748049795628, + 0.010748987086117268, + 0.006251505576074123, + 0.0038892270531505346, + 0.0032554774079471827, + 0.009312460198998451, + 0.0026016058400273323, + 0.003539381083101034, + 0.002283321460708976, + 0.0035127303563058376, + 0.008881787769496441, + 0.0027218684554100037, + 0.0022453032433986664, + 0.015375614166259766, + 0.0032117823138833046, + 0.002924532862380147, + 0.0016819083830341697, + 0.003019093070179224, + 0.013616664335131645, + 0.002098094904795289, + 0.0019677162636071444, + 0.009537491016089916, + 0.002757214242592454, + 0.004301941953599453, + 0.003280077362433076, + 0.004055107478052378, + 0.009879129007458687, + 0.004876928869634867, + 0.002752117346972227, + 0.17479532957077026, + 0.01423808466643095, + 0.023077940568327904, + 0.011357911862432957, + 0.03163568675518036, + 0.17014673352241516, + 0.016462238505482674, + 0.017449650913476944, + 0.004521179012954235, + 0.0038838121108710766, + 0.004680497571825981, + 0.003869746346026659, + 0.005533120594918728, + 0.00398062402382493, + 0.00845746323466301, + 0.003579967888072133, + 0.009352774359285831, + 0.006229719612747431, + 0.004872830584645271, + 0.004556209314614534, + 0.008065445348620415, + 0.0073122247122228146, + 0.003542006714269519, + 0.0030784043483436108, + 0.019349316135048866, + 0.028181808069348335, + 0.024095695465803146, + 0.023493299260735512, + 0.04165874794125557, + 0.024745766073465347, + 0.017647363245487213, + 0.017886623740196228, + 0.005237500183284283, + 0.006486613303422928, + 0.004800428170710802, + 0.002264459151774645, + 0.005639854818582535, + 0.0032184685114771128, + 0.003316686488687992, + 0.0026009134016931057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.014197062700986862 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle static close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dark, heavy pot filled with a deep-orange vegetable broth visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a steady hand with clean, manicured nails holding a pinch of earthy-green dried plant powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the powder drizzle down in a controlled pattern and dissolve into the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark, out-of-focus wooden table illuminated by warm, soft light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809351474046707 + ], + "opposite_logprob_tokens": [ + -3.2880935668945312 + ], + "logprob_reward": -0.03809351474046707, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809351474046707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.009368202649056911, + 0.0038060140796005726, + 0.0054264068603515625, + 0.005974984727799892, + 0.007089999504387379, + 0.9329030513763428, + 0.013114295899868011, + 0.008290254510939121, + 0.006529384292662144, + 0.007497407495975494 + ], + "credit_map": [ + 0.0012256422778591514, + 0.0007660309202037752, + 0.001415587612427771, + 0.001366439857520163, + 0.0008163396851159632, + 0.0014385338872671127, + 0.0012757566291838884, + 0.0010638704989105463, + 0.0005624163895845413, + 0.00038399561890400946, + 0.00042454071808606386, + 0.0004610840405803174, + 0.0005071922205388546, + 0.00040315446676686406, + 0.0005528758047148585, + 0.0005107544711790979, + 0.0011969230836257339, + 0.0005091625498607755, + 0.000538142048753798, + 0.00040484213968738914, + 0.00047347319195978343, + 0.001337934983894229, + 0.0006205197423696518, + 0.00034540818887762725, + 0.0016412670956924558, + 0.00032581319101154804, + 0.0004846832307521254, + 0.00045482933637686074, + 0.00041059486102312803, + 0.0017678759759292006, + 0.0004145809798501432, + 0.00047533924225717783, + 0.0018176236189901829, + 0.0003204028762411326, + 0.000726949714589864, + 0.0005326094687916338, + 0.0005809794529341161, + 0.0018794267671182752, + 0.0007496518082916737, + 0.00048235495341941714, + 0.3566950857639313, + 0.028996873646974564, + 0.04637402296066284, + 0.02272859588265419, + 0.06383193284273148, + 0.3446336090564728, + 0.03399251028895378, + 0.035650406032800674, + 0.0016508325934410095, + 0.0017312406562268734, + 0.0012542925542220473, + 0.001294824993237853, + 0.0021208743564784527, + 0.001441584317944944, + 0.002376277232542634, + 0.0012443671002984047, + 0.0008149571949616075, + 0.0006038442952558398, + 0.0013684096047654748, + 0.0012168989051133394, + 0.000837560510262847, + 0.001297241891734302, + 0.0012105307541787624, + 0.0009408114710822701, + 0.0009332700865343213, + 0.0011189188808202744, + 0.0005726413801312447, + 0.0005872566835023463, + 0.0014530548360198736, + 0.0008357302285730839, + 0.0004900582716800272, + 0.0005384531687013805, + 0.0010149668669328094, + 0.0009606695384718478, + 0.0009742028196342289, + 0.0005765054957009852, + 0.0018621947383508086, + 0.001117074629291892, + 0.0004725193721242249, + 0.0005192733951844275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.03809351474046707 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_12/index_tensor([8412, 8412])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_12/index_tensor([8412, 8412])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..27773fdbe10437f3e7e24af2f452876f7f926fec --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_12/index_tensor([8412, 8412])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The low-angle macro shot begins, with the camera slowly and steadily tilting upward to depict the graceful, winding path of a slender green vine as it elegantly climbs a dark, weathered wooden trellis. The entire scene is bathed in soft, diffused morning light, illuminating the delicate, translucent hairs on the vine's newest tendrils and capturing the glow of tiny dewdrops clinging to vibrant leaves. As the camera continues its tranquil upward journey, it reveals clusters of pristine, star-shaped white flowers, their waxy petals almost glowing against the deep espresso brown veins. A soft-focus blur of lush garden foliage in the background creates a shallow depth of field, isolating the vine and its flowers. The camera's deliberate, steady movements mimic the plant's slow, patient growth, capturing the intimate and intricate dance of nature and creating an atmosphere of profound tranquility and peace.", + "seed": 1066, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_12/index_8412_seed_1066_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0879426641855389, + "vlm_logprob": -0.0879426641855389, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.0879426641855389, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19942449033260345 + ], + "opposite_logprob_tokens": [ + -1.8244245052337646 + ], + "logprob_reward": -0.19942449033260345, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.19942449033260345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30158287286758423, + 0.05981791764497757, + 0.037613190710544586, + 0.05903743952512741, + 0.2200702279806137, + 0.07904898375272751, + 0.08984921127557755, + 0.07559333741664886, + 0.0467890240252018, + 0.03059781901538372 + ], + "credit_map": [ + 0.03195931017398834, + 0.016358496621251106, + 0.1119314506649971, + 0.06564643979072571, + 0.013415452092885971, + 0.020156722515821457, + 0.019901461899280548, + 0.02221350371837616, + 0.005146347917616367, + 0.007606684695929289, + 0.008764650672674179, + 0.008706007152795792, + 0.0050539905205369, + 0.007924859412014484, + 0.008756887167692184, + 0.007858481258153915, + 0.0032775222789496183, + 0.004121285863220692, + 0.005210635252296925, + 0.0065600075758993626, + 0.0033311385195702314, + 0.0038586691953241825, + 0.005676718894392252, + 0.005577208008617163, + 0.004636713769286871, + 0.0024537728168070316, + 0.015591319650411606, + 0.00490123825147748, + 0.004149121232330799, + 0.006405457854270935, + 0.004863604437559843, + 0.016036206856369972, + 0.020523199811577797, + 0.009629089385271072, + 0.049533113837242126, + 0.021341722458600998, + 0.01953914761543274, + 0.06381236016750336, + 0.013312499038875103, + 0.02237909846007824, + 0.015880053862929344, + 0.0056168632581830025, + 0.01079231034964323, + 0.009918124414980412, + 0.009119367226958275, + 0.01280985213816166, + 0.005088486708700657, + 0.009823918342590332, + 0.022861380130052567, + 0.006343542132526636, + 0.012442345730960369, + 0.008804918266832829, + 0.00890752486884594, + 0.010703177191317081, + 0.004705215338617563, + 0.015081096440553665, + 0.023589909076690674, + 0.013243397697806358, + 0.00662603136152029, + 0.0047150771133601665, + 0.007628250867128372, + 0.00310840317979455, + 0.0038156381342560053, + 0.012866622768342495, + 0.005588162690401077, + 0.005325611215084791, + 0.008083587512373924, + 0.006846369709819555, + 0.0036144598852843046, + 0.0037705230060964823, + 0.003926356788724661, + 0.00963394995778799, + 0.004398494027554989, + 0.003975855186581612, + 0.005045031197369099, + 0.0032780850306153297, + 0.00454453332349658, + 0.0025487784296274185, + 0.002970929490402341, + 0.0038361120969057083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.19942449033260345 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16040578484535217 + ], + "opposite_logprob_tokens": [ + -1.9104057550430298 + ], + "logprob_reward": -0.16040578484535217, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16040578484535217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19837436079978943, + 0.06317304074764252, + 0.06318756192922592, + 0.03473375365138054, + 0.048785269260406494, + 0.047134436666965485, + 0.32213667035102844, + 0.07535909861326218, + 0.10317101329565048, + 0.04394477605819702 + ], + "credit_map": [ + 0.01838303916156292, + 0.009778187610208988, + 0.07731176167726517, + 0.04269720986485481, + 0.008334365673363209, + 0.015869490802288055, + 0.01015359628945589, + 0.01584673300385475, + 0.0057008168660104275, + 0.007236982695758343, + 0.013419565744698048, + 0.007478015497326851, + 0.004299428313970566, + 0.005185057874768972, + 0.008601298555731773, + 0.01125187799334526, + 0.0053537022322416306, + 0.00804169476032257, + 0.007967626675963402, + 0.011730344034731388, + 0.005973974242806435, + 0.006750408560037613, + 0.008253159001469612, + 0.009116660803556442, + 0.0026832371950149536, + 0.001918120076879859, + 0.007780091371387243, + 0.003434346290305257, + 0.002976713003590703, + 0.0040635643526911736, + 0.0040739732794463634, + 0.007803712040185928, + 0.00331369461491704, + 0.0023163098376244307, + 0.00967703852802515, + 0.0057655079290270805, + 0.005184816662222147, + 0.010074576362967491, + 0.006040957290679216, + 0.006412371061742306, + 0.008188526146113873, + 0.0036006534937769175, + 0.00549809355288744, + 0.005079754628241062, + 0.008002516813576221, + 0.007808431051671505, + 0.003848788794130087, + 0.005107675679028034, + 0.08388914167881012, + 0.021322960034012794, + 0.03938290849328041, + 0.023598024621605873, + 0.03854851797223091, + 0.044902585446834564, + 0.017006609588861465, + 0.0534859336912632, + 0.01225846353918314, + 0.012152493931353092, + 0.0151909738779068, + 0.005171260796487331, + 0.008108781650662422, + 0.00582136632874608, + 0.006367301102727652, + 0.010288462974131107, + 0.012561466544866562, + 0.014509981498122215, + 0.01628377102315426, + 0.013194378465414047, + 0.00896503310650587, + 0.00934839341789484, + 0.008770172484219074, + 0.019537827000021935, + 0.006823246832937002, + 0.005598672665655613, + 0.006185858976095915, + 0.004381503909826279, + 0.005832878407090902, + 0.005091023165732622, + 0.004089336842298508, + 0.0059422580525279045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.16040578484535217 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421128399670124 + ], + "opposite_logprob_tokens": [ + -4.264211177825928 + ], + "logprob_reward": -0.01421128399670124, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421128399670124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40077605843544006, + 0.06760302931070328, + 0.0594850555062294, + 0.050422459840774536, + 0.12019990384578705, + 0.04555097967386246, + 0.09369832277297974, + 0.023885376751422882, + 0.03767969459295273, + 0.10069908201694489 + ], + "credit_map": [ + 0.028673579916357994, + 0.017378218472003937, + 0.15990899503231049, + 0.12026795744895935, + 0.00838712602853775, + 0.02352963760495186, + 0.0208018496632576, + 0.021828705444931984, + 0.004465853795409203, + 0.008288966491818428, + 0.009115532971918583, + 0.013257703743875027, + 0.008401937782764435, + 0.007265638560056686, + 0.008215301670134068, + 0.0085920924320817, + 0.004245673306286335, + 0.006574677769094706, + 0.007240379694849253, + 0.0087354127317667, + 0.005241804290562868, + 0.009481914341449738, + 0.010458811186254025, + 0.007506386376917362, + 0.003249569097533822, + 0.0037491058465093374, + 0.01202821359038353, + 0.004464203957468271, + 0.005094743333756924, + 0.005321479868143797, + 0.004331211093813181, + 0.012183932587504387, + 0.012713019736111164, + 0.009126195684075356, + 0.018291199579834938, + 0.014393127523362637, + 0.015577666461467743, + 0.017727693542838097, + 0.01758633181452751, + 0.014784673228859901, + 0.008207529783248901, + 0.0026474108453840017, + 0.005291706882417202, + 0.00486293388530612, + 0.008044080808758736, + 0.008716183714568615, + 0.0033332142047584057, + 0.004447919316589832, + 0.023264147341251373, + 0.006480967625975609, + 0.010063023306429386, + 0.007697709836065769, + 0.013823887333273888, + 0.01272743009030819, + 0.0061859204433858395, + 0.013455239124596119, + 0.0032638362608850002, + 0.004319214727729559, + 0.0029860034119337797, + 0.0016323948511853814, + 0.00358613021671772, + 0.0019060299964621663, + 0.0023978217504918575, + 0.003793946234509349, + 0.005014118272811174, + 0.004513683263212442, + 0.0061940341256558895, + 0.004486084450036287, + 0.004603680223226547, + 0.0033799754455685616, + 0.0036748889833688736, + 0.005813229829072952, + 0.006164986174553633, + 0.015178576111793518, + 0.013727856799960136, + 0.011968137696385384, + 0.017834631726145744, + 0.006328037939965725, + 0.015072744339704514, + 0.01442410796880722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.01421128399670124 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620541125535965 + ], + "opposite_logprob_tokens": [ + -2.81205415725708 + ], + "logprob_reward": -0.0620541125535965, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620541125535965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2736756503582001, + 0.07546290755271912, + 0.04393737390637398, + 0.06875275075435638, + 0.06567960977554321, + 0.09837859123945236, + 0.09128286689519882, + 0.05538841709494591, + 0.0403093658387661, + 0.18713243305683136 + ], + "credit_map": [ + 0.04890066385269165, + 0.015005874447524548, + 0.0707802101969719, + 0.04412224516272545, + 0.01791829615831375, + 0.026400191709399223, + 0.019739530980587006, + 0.03080866113305092, + 0.005965286400169134, + 0.007868637330830097, + 0.015230621211230755, + 0.007434253580868244, + 0.007128115277737379, + 0.007037207018584013, + 0.010034036822617054, + 0.014764753170311451, + 0.003456599311903119, + 0.00438341312110424, + 0.006308576092123985, + 0.00541105167940259, + 0.003955666441470385, + 0.0077514988370239735, + 0.0052193449810147285, + 0.007451224606484175, + 0.004276260267943144, + 0.004638958722352982, + 0.017866836860775948, + 0.005804311018437147, + 0.005677003413438797, + 0.00739575270563364, + 0.0054510850459337234, + 0.01764254830777645, + 0.006955437827855349, + 0.0035389994736760855, + 0.012021183967590332, + 0.0070659443736076355, + 0.006817242596298456, + 0.01218580175191164, + 0.006793230306357145, + 0.010301767848432064, + 0.01913551241159439, + 0.007408368866890669, + 0.013376375660300255, + 0.01406856533139944, + 0.01225862093269825, + 0.01570696197450161, + 0.006573953665792942, + 0.009850229136645794, + 0.014689485542476177, + 0.010646236129105091, + 0.007965085096657276, + 0.009849241934716702, + 0.015466097742319107, + 0.014237347058951855, + 0.007796560414135456, + 0.010632817633450031, + 0.019202589988708496, + 0.005030049476772547, + 0.004793251398950815, + 0.0027138281147927046, + 0.0035611805506050587, + 0.0026416960172355175, + 0.002874162979424, + 0.014571658335626125, + 0.004198629409074783, + 0.004866555333137512, + 0.005561605095863342, + 0.004486028105020523, + 0.004854277241975069, + 0.0036204997450113297, + 0.004580303095281124, + 0.008141466416418552, + 0.010636642575263977, + 0.02943914569914341, + 0.024224670603871346, + 0.02180583029985428, + 0.03440163657069206, + 0.011439441703259945, + 0.02873728796839714, + 0.026447780430316925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0620541125535965 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036176491994410753 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.0036176491994410753, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036176491994410753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24764280021190643, + 0.06802497804164886, + 0.043574027717113495, + 0.03777659311890602, + 0.08401373028755188, + 0.2918436527252197, + 0.0782119482755661, + 0.04068995639681816, + 0.03213648498058319, + 0.07608587294816971 + ], + "credit_map": [ + 0.03393563628196716, + 0.02296941727399826, + 0.039791520684957504, + 0.08022896200418472, + 0.014607489109039307, + 0.016010425984859467, + 0.019604148343205452, + 0.02049519121646881, + 0.006423632614314556, + 0.006948059890419245, + 0.011264550499618053, + 0.008722645230591297, + 0.008075298741459846, + 0.009075629524886608, + 0.00774137768894434, + 0.00977377686649561, + 0.004971866961568594, + 0.004750327672809362, + 0.005759268067777157, + 0.004651383031159639, + 0.003178943181410432, + 0.006745181046426296, + 0.00684330565854907, + 0.006673750467598438, + 0.0033143158070743084, + 0.004068717360496521, + 0.004782244563102722, + 0.004461655858904123, + 0.0033066063188016415, + 0.005516479257494211, + 0.004838841501623392, + 0.007487730588763952, + 0.008197524584829807, + 0.005729212891310453, + 0.012439045123755932, + 0.009303019382059574, + 0.012463517487049103, + 0.011600934900343418, + 0.01287650316953659, + 0.011403964832425117, + 0.0514092743396759, + 0.019787674769759178, + 0.05114420875906944, + 0.037185635417699814, + 0.028249606490135193, + 0.053627610206604004, + 0.02034282684326172, + 0.030096780508756638, + 0.014919598586857319, + 0.006639576982706785, + 0.006433179602026939, + 0.010808411054313183, + 0.014317505992949009, + 0.007804077584296465, + 0.005962607450783253, + 0.011326978914439678, + 0.011395040899515152, + 0.003910715226083994, + 0.007414784748107195, + 0.0025348474737256765, + 0.0030871883500367403, + 0.0020298233721405268, + 0.0025435255374759436, + 0.007774024736136198, + 0.0035366970114409924, + 0.0038758323062211275, + 0.0040414175018668175, + 0.00432773819193244, + 0.0037400696892291307, + 0.002876265672966838, + 0.0034514220897108316, + 0.006287041120231152, + 0.004455171059817076, + 0.012223907746374607, + 0.00958644226193428, + 0.009101998060941696, + 0.014014624059200287, + 0.005184320267289877, + 0.011532568372786045, + 0.00998682901263237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0036176491994410753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The low-angle macro shot begins, with the camera slowly and steadily tilting upward to depict the graceful, winding path of a slender green vine as it elegantly climbs a dark, weathered wooden trellis. The entire scene is bathed in soft, diffused morning light, illuminating the delicate, translucent hairs on the vine's newest tendrils and capturing the glow of tiny dewdrops clinging to vibrant leaves. As the camera continues its tranquil upward journey, it reveals clusters of pristine, star-shaped white flowers, their waxy petals almost glowing against the deep espresso brown veins. A soft-focus blur of lush garden foliage in the background creates a shallow depth of field, isolating the vine and its flowers. The camera's deliberate, steady movements mimic the plant's slow, patient growth, capturing the intimate and intricate dance of nature and creating an atmosphere of profound tranquility and peace.", + "seed": 1067, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_12/index_8412_seed_1067_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.017820152256172152, + "vlm_logprob": -0.017820152256172152, + "vlm_yes_no_margin": 5.275, + "vlm_token_credit": -0.017820152256172152, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07169751822948456 + ], + "opposite_logprob_tokens": [ + -2.821697473526001 + ], + "logprob_reward": -0.07169751822948456, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07169751822948456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08527221530675888, + 0.1283431053161621, + 0.0715523511171341, + 0.05132293328642845, + 0.11957675963640213, + 0.12366144359111786, + 0.07288864254951477, + 0.06985564529895782, + 0.21832284331321716, + 0.05920407176017761 + ], + "credit_map": [ + 0.009181883186101913, + 0.012153259478509426, + 0.01449382584542036, + 0.010646015405654907, + 0.007823805324733257, + 0.0066748009994626045, + 0.007518051192164421, + 0.01678057201206684, + 0.02522762306034565, + 0.010368281975388527, + 0.018462486565113068, + 0.011539255268871784, + 0.01911470666527748, + 0.014481623657047749, + 0.012117122299969196, + 0.017032010480761528, + 0.009779597632586956, + 0.012189936824142933, + 0.006752965971827507, + 0.008293049409985542, + 0.007915838621556759, + 0.010660707019269466, + 0.008705481886863708, + 0.00725477235391736, + 0.00898581650108099, + 0.007513429503887892, + 0.005962899886071682, + 0.005821778438985348, + 0.007472896482795477, + 0.00756781455129385, + 0.004154813010245562, + 0.0038434851448982954, + 0.013854037038981915, + 0.009465059265494347, + 0.029445484280586243, + 0.009281565435230732, + 0.014417006634175777, + 0.01697179675102234, + 0.01359898503869772, + 0.012542827986180782, + 0.015779579058289528, + 0.014819460920989513, + 0.016723036766052246, + 0.014594885520637035, + 0.010217366740107536, + 0.02285154163837433, + 0.010647823102772236, + 0.018027745187282562, + 0.010878033004701138, + 0.007470299024134874, + 0.010146002285182476, + 0.0055157942697405815, + 0.004966406617313623, + 0.010662720538675785, + 0.007594187743961811, + 0.01565520092844963, + 0.006830330938100815, + 0.004757661372423172, + 0.011621082201600075, + 0.006443662103265524, + 0.0068920450285077095, + 0.006820144597440958, + 0.011945100501179695, + 0.014545616693794727, + 0.023987114429473877, + 0.016493137925863266, + 0.04913444072008133, + 0.03854331001639366, + 0.011982842348515987, + 0.02651136741042137, + 0.01920073665678501, + 0.03246990218758583, + 0.010126856155693531, + 0.0037994717713445425, + 0.014898810535669327, + 0.011556671001017094, + 0.0034177489578723907, + 0.004637734964489937, + 0.005402660463005304, + 0.005364119075238705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.07169751822948456 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009800413623452187 + ], + "opposite_logprob_tokens": [ + -4.634800434112549 + ], + "logprob_reward": -0.009800413623452187, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009800413623452187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0873740166425705, + 0.04556320607662201, + 0.03628727048635483, + 0.05491027608513832, + 0.0667894259095192, + 0.08423104882240295, + 0.10522031784057617, + 0.12323126941919327, + 0.3324089050292969, + 0.06398436427116394 + ], + "credit_map": [ + 0.00748342415317893, + 0.015358960255980492, + 0.012982775457203388, + 0.014673707075417042, + 0.007127252407371998, + 0.00826207548379898, + 0.010390782728791237, + 0.011095022782683372, + 0.005948316305875778, + 0.007134388200938702, + 0.005648800637573004, + 0.00776335783302784, + 0.00455712852999568, + 0.0039656213484704494, + 0.004821356851607561, + 0.005724228918552399, + 0.006740487180650234, + 0.006122677121311426, + 0.0035240924917161465, + 0.00528789684176445, + 0.003669422585517168, + 0.005400219466537237, + 0.002936622593551874, + 0.0026058475486934185, + 0.008079573512077332, + 0.009781021624803543, + 0.005394332110881805, + 0.00836870726197958, + 0.006918805185705423, + 0.007594191003590822, + 0.004446112085133791, + 0.0043275244534015656, + 0.006902643013745546, + 0.006973861251026392, + 0.013747070915997028, + 0.007121290545910597, + 0.009668363258242607, + 0.006661843974143267, + 0.008891504257917404, + 0.0068228356540203094, + 0.01333326380699873, + 0.009550877846777439, + 0.009248942136764526, + 0.007959689013659954, + 0.008078250102698803, + 0.01950058527290821, + 0.005612544249743223, + 0.010946882888674736, + 0.01727951131761074, + 0.008653802797198296, + 0.022423023357987404, + 0.00835296418517828, + 0.006218616385012865, + 0.011237761937081814, + 0.012348556891083717, + 0.0187060683965683, + 0.014498861506581306, + 0.006906737573444843, + 0.019263317808508873, + 0.010404511354863644, + 0.008618053048849106, + 0.011440658941864967, + 0.01443732064217329, + 0.037661775946617126, + 0.03491196408867836, + 0.02200569584965706, + 0.10862225294113159, + 0.04127533361315727, + 0.017571061849594116, + 0.03875131160020828, + 0.026919281110167503, + 0.042351946234703064, + 0.006566948723047972, + 0.0035898778587579727, + 0.01674639992415905, + 0.010828928090631962, + 0.003542087972164154, + 0.0054187108762562275, + 0.0036095078103244305, + 0.01368188951164484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.009800413623452187 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006392342620529234 + ], + "opposite_logprob_tokens": [ + -7.37563943862915 + ], + "logprob_reward": -0.0006392342620529234, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006392342620529234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09282421320676804, + 0.1415875256061554, + 0.16652347147464752, + 0.04726289212703705, + 0.0529232919216156, + 0.13238029181957245, + 0.07188542932271957, + 0.18455959856510162, + 0.05197109282016754, + 0.058082178235054016 + ], + "credit_map": [ + 0.009437214583158493, + 0.009685722179710865, + 0.015823015943169594, + 0.02053348533809185, + 0.00817429181188345, + 0.0075119249522686005, + 0.0063379500061273575, + 0.015320607461035252, + 0.01864287070930004, + 0.012690795585513115, + 0.023520877584815025, + 0.012550262734293938, + 0.015866585075855255, + 0.02315692789852619, + 0.01566084660589695, + 0.01949835941195488, + 0.024780040606856346, + 0.018376821652054787, + 0.02520204707980156, + 0.020297743380069733, + 0.015563854947686195, + 0.020689910277724266, + 0.01761370152235031, + 0.023999342694878578, + 0.006952529773116112, + 0.005549841560423374, + 0.006246479693800211, + 0.00944947637617588, + 0.004989373032003641, + 0.006370767951011658, + 0.003642850089818239, + 0.0040615759789943695, + 0.005212015006691217, + 0.005892174318432808, + 0.011859968304634094, + 0.005084267817437649, + 0.007413340732455254, + 0.007136465050280094, + 0.005525797139853239, + 0.004799263086169958, + 0.014490515924990177, + 0.022906586527824402, + 0.017960723489522934, + 0.013696487061679363, + 0.016173716634511948, + 0.018654294312000275, + 0.01339004747569561, + 0.01510792225599289, + 0.008929762057960033, + 0.00658900011330843, + 0.017242304980754852, + 0.006348587106913328, + 0.004867829382419586, + 0.010083923116326332, + 0.005863015539944172, + 0.011961007490754128, + 0.024427397176623344, + 0.012869244441390038, + 0.02873530052602291, + 0.01646251790225506, + 0.014148151502013206, + 0.02806670218706131, + 0.012580128386616707, + 0.047270163893699646, + 0.00578742939978838, + 0.006337171420454979, + 0.007550272159278393, + 0.010373405180871487, + 0.0036727681290358305, + 0.005014640279114246, + 0.005403113551437855, + 0.007832290604710579, + 0.006468279752880335, + 0.004620586056262255, + 0.011501247994601727, + 0.011032425798475742, + 0.005861890036612749, + 0.005517386831343174, + 0.00548826577141881, + 0.00759209506213665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.0006392342620529234 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017137139802798629 + ], + "opposite_logprob_tokens": [ + -6.376713752746582 + ], + "logprob_reward": -0.0017137139802798629, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017137139802798629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07391055673360825, + 0.05039709433913231, + 0.0404864102602005, + 0.10280463844537735, + 0.10083678364753723, + 0.10679980367422104, + 0.04779805988073349, + 0.1486763060092926, + 0.14950251579284668, + 0.17878785729408264 + ], + "credit_map": [ + 0.005829792004078627, + 0.012487906031310558, + 0.011879502795636654, + 0.015470750629901886, + 0.005971753969788551, + 0.006352758500725031, + 0.005652991123497486, + 0.010265101678669453, + 0.0049148849211633205, + 0.011788370087742805, + 0.005717978812754154, + 0.008009731769561768, + 0.0049893357791006565, + 0.002903877990320325, + 0.0038756192661821842, + 0.008197293616831303, + 0.005119489971548319, + 0.0063673751428723335, + 0.005172197241336107, + 0.004827128257602453, + 0.004029547329992056, + 0.005536277312785387, + 0.0042615169659256935, + 0.005172880366444588, + 0.016708290204405785, + 0.016910739243030548, + 0.011300508864223957, + 0.015731412917375565, + 0.011587689630687237, + 0.014766943641006947, + 0.007369212340563536, + 0.008429848589003086, + 0.009773346595466137, + 0.0076099359430372715, + 0.03136325255036354, + 0.00744930375367403, + 0.014296143315732479, + 0.010133915580809116, + 0.011112920008599758, + 0.009097964502871037, + 0.015736421570181847, + 0.010185619816184044, + 0.012394050136208534, + 0.010506370104849339, + 0.015272284857928753, + 0.02117009647190571, + 0.008956846781075, + 0.012578112073242664, + 0.0097894836217165, + 0.005908981431275606, + 0.006506933830678463, + 0.005253547336906195, + 0.0037203216925263405, + 0.005920686759054661, + 0.0035941777750849724, + 0.0071039265021681786, + 0.022466780617833138, + 0.01014890056103468, + 0.020009685307741165, + 0.013480992056429386, + 0.017063884064555168, + 0.030742498114705086, + 0.01538174320012331, + 0.01938183233141899, + 0.014271962456405163, + 0.01618395559489727, + 0.02435298264026642, + 0.02382582612335682, + 0.014084963127970695, + 0.018771057948470116, + 0.018074555322527885, + 0.019937215372920036, + 0.021790452301502228, + 0.009350165724754333, + 0.053631365299224854, + 0.023602092638611794, + 0.00981846172362566, + 0.020432589575648308, + 0.014228898100554943, + 0.02593383379280567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.0017137139802798629 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249881185591221 + ], + "opposite_logprob_tokens": [ + -5.255249977111816 + ], + "logprob_reward": -0.005249881185591221, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249881185591221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09110677987337112, + 0.10464558750391006, + 0.04970826581120491, + 0.03266926482319832, + 0.06257574260234833, + 0.04648074507713318, + 0.03110589273273945, + 0.04442960023880005, + 0.05947379022836685, + 0.4778043329715729 + ], + "credit_map": [ + 0.007875492796301842, + 0.006833635736256838, + 0.01575494185090065, + 0.011813650839030743, + 0.007985633797943592, + 0.012547772377729416, + 0.0066693988628685474, + 0.021626245230436325, + 0.013724494725465775, + 0.012570332735776901, + 0.01742028072476387, + 0.012139508500695229, + 0.011089814826846123, + 0.009696783497929573, + 0.01150487456470728, + 0.016499493271112442, + 0.006462343968451023, + 0.006008529104292393, + 0.006085252854973078, + 0.009148230776190758, + 0.004313783720135689, + 0.007027475629001856, + 0.005061439238488674, + 0.005601210054010153, + 0.00536044966429472, + 0.00561505975201726, + 0.002844808390364051, + 0.004078042693436146, + 0.00488276919350028, + 0.004487793892621994, + 0.0028041028417646885, + 0.002596238162368536, + 0.005532475188374519, + 0.006088444031774998, + 0.017467526718974113, + 0.004509999416768551, + 0.0081957271322608, + 0.007059919182211161, + 0.007216095924377441, + 0.006505552679300308, + 0.004213837906718254, + 0.007415631785988808, + 0.006602962035685778, + 0.006031867582350969, + 0.005828485824167728, + 0.004968460649251938, + 0.004425341729074717, + 0.006994155701249838, + 0.006206723861396313, + 0.004010338801890612, + 0.004379541613161564, + 0.0030889096669852734, + 0.0021883342415094376, + 0.0047180745750665665, + 0.002736281370744109, + 0.003777689067646861, + 0.007400240283459425, + 0.0025170762091875076, + 0.00787326879799366, + 0.002249572891741991, + 0.0026333217974752188, + 0.006368578877300024, + 0.0031205276027321815, + 0.012267019599676132, + 0.00675817858427763, + 0.004944066982716322, + 0.01313049253076315, + 0.010431604459881783, + 0.003953634761273861, + 0.0069856299087405205, + 0.004875004757195711, + 0.00839517917484045, + 0.07629576325416565, + 0.02728930115699768, + 0.1612931489944458, + 0.08605422079563141, + 0.014388053677976131, + 0.037249695509672165, + 0.02279416285455227, + 0.05243999883532524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.005249881185591221 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The low-angle macro shot begins, with the camera slowly and steadily tilting upward to depict the graceful, winding path of a slender green vine as it elegantly climbs a dark, weathered wooden trellis. The entire scene is bathed in soft, diffused morning light, illuminating the delicate, translucent hairs on the vine's newest tendrils and capturing the glow of tiny dewdrops clinging to vibrant leaves. As the camera continues its tranquil upward journey, it reveals clusters of pristine, star-shaped white flowers, their waxy petals almost glowing against the deep espresso brown veins. A soft-focus blur of lush garden foliage in the background creates a shallow depth of field, isolating the vine and its flowers. The camera's deliberate, steady movements mimic the plant's slow, patient growth, capturing the intimate and intricate dance of nature and creating an atmosphere of profound tranquility and peace.", + "seed": 1068, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_12/index_8412_seed_1068_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0291081128641963, + "vlm_logprob": -0.0291081128641963, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.0291081128641963, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05960816144943237 + ], + "opposite_logprob_tokens": [ + -3.059608221054077 + ], + "logprob_reward": -0.05960816144943237, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05960816144943237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07527169585227966, + 0.024126050993800163, + 0.36867576837539673, + 0.029074838384985924, + 0.10589238256216049, + 0.026501979678869247, + 0.01931096985936165, + 0.15153484046459198, + 0.17122265696525574, + 0.028388850390911102 + ], + "credit_map": [ + 0.005013725720345974, + 0.008507473394274712, + 0.006329719442874193, + 0.00969774555414915, + 0.014762552455067635, + 0.007921488955616951, + 0.014639039523899555, + 0.008399944752454758, + 0.0020717792212963104, + 0.003961259964853525, + 0.0037408440839499235, + 0.003678110893815756, + 0.002485298551619053, + 0.002962967613711953, + 0.0024643957149237394, + 0.002761393552646041, + 0.019113516435027122, + 0.009421554394066334, + 0.11434139311313629, + 0.01698199100792408, + 0.015765096992254257, + 0.029857348650693893, + 0.14721497893333435, + 0.015979880467057228, + 0.0022144776303321123, + 0.0021954490803182125, + 0.0026693660765886307, + 0.002906299429014325, + 0.004266277886927128, + 0.0021293454337865114, + 0.007362090051174164, + 0.0053315311670303345, + 0.008425894193351269, + 0.008756063878536224, + 0.005832119379192591, + 0.006693496368825436, + 0.041007984429597855, + 0.005956358276307583, + 0.009307890199124813, + 0.01991257444024086, + 0.0019125720718875527, + 0.00794279295951128, + 0.003800250357016921, + 0.002662495244294405, + 0.0023386909160763025, + 0.0013059276388958097, + 0.003941127564758062, + 0.002598121762275696, + 0.001149378134869039, + 0.0037093190476298332, + 0.0035026513505727053, + 0.0012316819047555327, + 0.0038605867885053158, + 0.0009031777153722942, + 0.0032940690871328115, + 0.0016601046081632376, + 0.010752573609352112, + 0.02103082835674286, + 0.03289082646369934, + 0.015095902606844902, + 0.015605825930833817, + 0.007858582772314548, + 0.009135713800787926, + 0.03916458040475845, + 0.02042093686759472, + 0.02541728876531124, + 0.030689595267176628, + 0.020924847573041916, + 0.03018486686050892, + 0.011500183492898941, + 0.016643952578306198, + 0.015440995804965496, + 0.005598318763077259, + 0.0027381558902561665, + 0.002714387606829405, + 0.00623130751773715, + 0.003619169117882848, + 0.0026793363504111767, + 0.0022896435111761093, + 0.0025185290724039078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.05960816144943237 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898911088705063 + ], + "opposite_logprob_tokens": [ + -2.578989028930664 + ], + "logprob_reward": -0.07898911088705063, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898911088705063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19929248094558716, + 0.0430891253054142, + 0.2514665722846985, + 0.03790215030312538, + 0.03079579584300518, + 0.0416376069188118, + 0.035885464400053024, + 0.05880310386419296, + 0.23618105053901672, + 0.064946748316288 + ], + "credit_map": [ + 0.013856371864676476, + 0.022904381155967712, + 0.016631394624710083, + 0.022891459986567497, + 0.06264176219701767, + 0.01978682354092598, + 0.023497221991419792, + 0.01708306185901165, + 0.003571454668417573, + 0.005581253673881292, + 0.007054923567920923, + 0.0061912331730127335, + 0.005668432917445898, + 0.004172412678599358, + 0.0051787537522614, + 0.005670658778399229, + 0.018454963341355324, + 0.009297700598835945, + 0.03124145418405533, + 0.014871079474687576, + 0.01193599309772253, + 0.0323810838162899, + 0.09861413389444351, + 0.034670136868953705, + 0.006712853908538818, + 0.0036553957033902407, + 0.00304060778580606, + 0.0042016333900392056, + 0.005206996574997902, + 0.0033824800048023462, + 0.005017233081161976, + 0.006684949155896902, + 0.003188334172591567, + 0.0033765772823244333, + 0.0024653577711433172, + 0.0029035969637334347, + 0.00866754725575447, + 0.002174146007746458, + 0.003862784942612052, + 0.004157452378422022, + 0.0032569679897278547, + 0.009902182035148144, + 0.005223582033067942, + 0.004522324539721012, + 0.00612949812784791, + 0.0021795127540826797, + 0.006360383238643408, + 0.004063156899064779, + 0.001975823426619172, + 0.005882483907043934, + 0.005034392699599266, + 0.002368711167946458, + 0.0038954056799411774, + 0.0018948728684335947, + 0.007054731249809265, + 0.007779046427458525, + 0.004315115045756102, + 0.009558659046888351, + 0.010302824899554253, + 0.006490010768175125, + 0.00942414253950119, + 0.003275160910561681, + 0.005751772318035364, + 0.009685419499874115, + 0.025962885469198227, + 0.03712358698248863, + 0.03162688389420509, + 0.024808116257190704, + 0.048623815178871155, + 0.018998460844159126, + 0.026818349957466125, + 0.022218948230147362, + 0.010300856083631516, + 0.005317636765539646, + 0.008653576485812664, + 0.008608829230070114, + 0.012339815497398376, + 0.006521733943372965, + 0.006081112660467625, + 0.007123184390366077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.07898911088705063 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013394916895776987 + ], + "opposite_logprob_tokens": [ + -6.626339435577393 + ], + "logprob_reward": -0.0013394916895776987, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013394916895776987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11967963725328445, + 0.05504347011446953, + 0.27151772379875183, + 0.2266828864812851, + 0.1039532870054245, + 0.02082768827676773, + 0.04550422355532646, + 0.07259193807840347, + 0.05403211712837219, + 0.030167026445269585 + ], + "credit_map": [ + 0.01008579507470131, + 0.015326470136642456, + 0.01524258404970169, + 0.015428652055561543, + 0.01887495256960392, + 0.015514658764004707, + 0.01759529672563076, + 0.01161122228950262, + 0.004614167381078005, + 0.0048262025229632854, + 0.011082318611443043, + 0.006087184883654118, + 0.009639148600399494, + 0.00648478465154767, + 0.006603371351957321, + 0.0057062916457653046, + 0.014960061758756638, + 0.016194012016057968, + 0.058417338877916336, + 0.03438709303736687, + 0.016884110867977142, + 0.022128207609057426, + 0.08933135867118835, + 0.019215546548366547, + 0.030135322362184525, + 0.011569042690098286, + 0.014374077320098877, + 0.011268407106399536, + 0.0326492004096508, + 0.017245875671505928, + 0.09270481020212173, + 0.01673615165054798, + 0.008014681749045849, + 0.010353172197937965, + 0.008820861577987671, + 0.011646266095340252, + 0.03164855018258095, + 0.005580463912338018, + 0.012199513614177704, + 0.015689769759774208, + 0.0018425858579576015, + 0.0034488756209611893, + 0.0036078987177461386, + 0.0018281980883330107, + 0.0016934608574956656, + 0.0016064546070992947, + 0.0022865573409944773, + 0.004513656720519066, + 0.0034167564008384943, + 0.008645758964121342, + 0.007982887327671051, + 0.004049535375088453, + 0.003647969802841544, + 0.003557489952072501, + 0.00456724688410759, + 0.009636573493480682, + 0.006920261774212122, + 0.011151292361319065, + 0.010297585278749466, + 0.007160068955272436, + 0.016759615391492844, + 0.005978654604405165, + 0.006188661325722933, + 0.008135799318552017, + 0.007005632389336824, + 0.007844316773116589, + 0.007004817482084036, + 0.006174458656460047, + 0.010547676123678684, + 0.003941585775464773, + 0.004859697539359331, + 0.006653935182839632, + 0.006217022892087698, + 0.0030296591576188803, + 0.0028286168817430735, + 0.0072588734328746796, + 0.004366348031908274, + 0.00193928147200495, + 0.0025578816421329975, + 0.001969343749806285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.0013394916895776987 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015118608716875315 + ], + "opposite_logprob_tokens": [ + -6.501512050628662 + ], + "logprob_reward": -0.0015118608716875315, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015118608716875315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06945417821407318, + 0.037085533142089844, + 0.05411629378795624, + 0.18957874178886414, + 0.04498359560966492, + 0.024338290095329285, + 0.015191411599516869, + 0.21545398235321045, + 0.20510470867156982, + 0.14469324052333832 + ], + "credit_map": [ + 0.005334891844540834, + 0.008923796936869621, + 0.007020331919193268, + 0.009775363840162754, + 0.012375555001199245, + 0.007685430347919464, + 0.011592227034270763, + 0.006746577098965645, + 0.004085608758032322, + 0.0028978281188756227, + 0.007603108417242765, + 0.002791645238175988, + 0.004322380758821964, + 0.004884182009845972, + 0.0036328397691249847, + 0.006867939606308937, + 0.004412712529301643, + 0.0031168481800705194, + 0.005191188305616379, + 0.00306698284111917, + 0.0033324160613119602, + 0.0067658559419214725, + 0.01947084255516529, + 0.008759444579482079, + 0.026688454672694206, + 0.009648725390434265, + 0.012557466514408588, + 0.009680530056357384, + 0.025114813819527626, + 0.013957971706986427, + 0.07871221750974655, + 0.013218559324741364, + 0.00402238592505455, + 0.004682969767600298, + 0.0019858884625136852, + 0.0031266575679183006, + 0.016130702570080757, + 0.002786981640383601, + 0.004529820755124092, + 0.007718185894191265, + 0.0018769092857837677, + 0.005516164004802704, + 0.0044868323020637035, + 0.002456182846799493, + 0.0024234186857938766, + 0.0016664924332872033, + 0.0031587525736540556, + 0.002753538778051734, + 0.0009903856553137302, + 0.0027216693852096796, + 0.0022632917389273643, + 0.0017190549988299608, + 0.0013881995109841228, + 0.0008276523440144956, + 0.00275026704184711, + 0.002530889818444848, + 0.0155830979347229, + 0.026247626170516014, + 0.04854808375239372, + 0.019807131960988045, + 0.024365004152059555, + 0.012198415584862232, + 0.011055298149585724, + 0.05764932930469513, + 0.025982938706874847, + 0.028660966083407402, + 0.0368063859641552, + 0.021649761125445366, + 0.042371708899736404, + 0.014561677351593971, + 0.01690690778195858, + 0.01816435158252716, + 0.023760110139846802, + 0.016250625252723694, + 0.012933396734297276, + 0.02358115278184414, + 0.028370190411806107, + 0.014174952171742916, + 0.012408643029630184, + 0.013214172795414925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.0015118608716875315 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091939423233271 + ], + "opposite_logprob_tokens": [ + -5.504091739654541 + ], + "logprob_reward": -0.004091939423233271, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091939423233271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2230808287858963, + 0.065814308822155, + 0.15133818984031677, + 0.1470334380865097, + 0.15974090993404388, + 0.06636686623096466, + 0.023952187970280647, + 0.09649614244699478, + 0.022178146988153458, + 0.04399893060326576 + ], + "credit_map": [ + 0.01863742433488369, + 0.022456509992480278, + 0.031872060149908066, + 0.017742210999131203, + 0.0496593713760376, + 0.02640947327017784, + 0.03456154838204384, + 0.021742234006524086, + 0.008352397009730339, + 0.004375764634460211, + 0.014391815289855003, + 0.005860297475010157, + 0.007952720858156681, + 0.007420133799314499, + 0.007077036891132593, + 0.010384143330156803, + 0.012209056876599789, + 0.01496363990008831, + 0.03326023742556572, + 0.018561551347374916, + 0.011839990504086018, + 0.017371080815792084, + 0.018476255238056183, + 0.0246563870459795, + 0.02182822674512863, + 0.007797092664986849, + 0.009381146170198917, + 0.008336745202541351, + 0.021397413685917854, + 0.011245119385421276, + 0.05680541321635246, + 0.010242274962365627, + 0.013307512737810612, + 0.013970982283353806, + 0.008407289162278175, + 0.008691132068634033, + 0.06527755409479141, + 0.0102639589458704, + 0.012494772672653198, + 0.027327707037329674, + 0.004221025854349136, + 0.015933755785226822, + 0.004800393246114254, + 0.005557433236390352, + 0.005935597699135542, + 0.005032194312661886, + 0.008331567980349064, + 0.016554899513721466, + 0.0015182627830654383, + 0.0056045809760689735, + 0.003394388360902667, + 0.0019574023317545652, + 0.002406665589660406, + 0.0018560446333140135, + 0.002822834998369217, + 0.004392008297145367, + 0.006237615831196308, + 0.01158875785768032, + 0.02204485610127449, + 0.008448288775980473, + 0.009427402168512344, + 0.004091173410415649, + 0.006146245636045933, + 0.028511803597211838, + 0.0028525423258543015, + 0.001984093338251114, + 0.0036035862285643816, + 0.0019654890056699514, + 0.004625737201422453, + 0.001922828727401793, + 0.0027389083988964558, + 0.0024849616456776857, + 0.006503296550363302, + 0.004100672900676727, + 0.008904692716896534, + 0.0028951759450137615, + 0.008784055709838867, + 0.004266385454684496, + 0.00343442359007895, + 0.005110232159495354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.004091939423233271 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The low-angle macro shot begins, with the camera slowly and steadily tilting upward to depict the graceful, winding path of a slender green vine as it elegantly climbs a dark, weathered wooden trellis. The entire scene is bathed in soft, diffused morning light, illuminating the delicate, translucent hairs on the vine's newest tendrils and capturing the glow of tiny dewdrops clinging to vibrant leaves. As the camera continues its tranquil upward journey, it reveals clusters of pristine, star-shaped white flowers, their waxy petals almost glowing against the deep espresso brown veins. A soft-focus blur of lush garden foliage in the background creates a shallow depth of field, isolating the vine and its flowers. The camera's deliberate, steady movements mimic the plant's slow, patient growth, capturing the intimate and intricate dance of nature and creating an atmosphere of profound tranquility and peace.", + "seed": 1069, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_12/index_8412_seed_1069_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01168067977996543, + "vlm_logprob": -0.01168067977996543, + "vlm_yes_no_margin": 5.55, + "vlm_token_credit": -0.01168067977996543, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.039308931678533554 + ], + "opposite_logprob_tokens": [ + -3.789309024810791 + ], + "logprob_reward": -0.039308931678533554, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.039308931678533554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0901217833161354, + 0.15023984014987946, + 0.12737947702407837, + 0.040072984993457794, + 0.09448909014463425, + 0.07362036406993866, + 0.15866902470588684, + 0.052765797823667526, + 0.1073145940899849, + 0.10532703995704651 + ], + "credit_map": [ + 0.006426827982068062, + 0.00757380947470665, + 0.014945915900170803, + 0.008425824344158173, + 0.014073316939175129, + 0.01676025241613388, + 0.01127985306084156, + 0.010635987855494022, + 0.008810301311314106, + 0.005370087921619415, + 0.05274827778339386, + 0.008914459496736526, + 0.0077231391333043575, + 0.008350512944161892, + 0.05027490109205246, + 0.008048160932958126, + 0.011428428813815117, + 0.008623301051557064, + 0.012702768668532372, + 0.026314957067370415, + 0.011797451414167881, + 0.01181807555258274, + 0.024403709918260574, + 0.020290786400437355, + 0.004762315656989813, + 0.004247425589710474, + 0.004918414168059826, + 0.005376082845032215, + 0.0049513825215399265, + 0.006267993710935116, + 0.006053165532648563, + 0.003496205899864435, + 0.01309062261134386, + 0.010381020605564117, + 0.007448985241353512, + 0.006277769338339567, + 0.011951553635299206, + 0.021306293085217476, + 0.007730443030595779, + 0.01630239374935627, + 0.006873612757772207, + 0.008791478350758553, + 0.010235195979475975, + 0.005704245995730162, + 0.005529857240617275, + 0.020224394276738167, + 0.007856491953134537, + 0.008405091241002083, + 0.015543630346655846, + 0.01751966029405594, + 0.02042684145271778, + 0.01620248146355152, + 0.010172602720558643, + 0.02333163097500801, + 0.024503586813807487, + 0.030968589708209038, + 0.0062508536502718925, + 0.00477589899674058, + 0.006350691430270672, + 0.008456600829958916, + 0.004402393475174904, + 0.00734928622841835, + 0.005049742758274078, + 0.010130329057574272, + 0.009737788699567318, + 0.008924270048737526, + 0.014406721107661724, + 0.021019164472818375, + 0.009850538335740566, + 0.012422684580087662, + 0.016106262803077698, + 0.014847165904939175, + 0.008537373505532742, + 0.004125835373997688, + 0.040756724774837494, + 0.008934118784964085, + 0.004776314366608858, + 0.006197286304086447, + 0.0039110807701945305, + 0.028088301420211792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.039308931678533554 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014218805357813835 + ], + "opposite_logprob_tokens": [ + -4.264218807220459 + ], + "logprob_reward": -0.014218805357813835, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014218805357813835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04388630390167236, + 0.14399580657482147, + 0.05959317833185196, + 0.036793749779462814, + 0.12790189683437347, + 0.12861131131649017, + 0.04423176124691963, + 0.12705321609973907, + 0.15290306508541107, + 0.13502958416938782 + ], + "credit_map": [ + 0.0024660055059939623, + 0.0031351398210972548, + 0.00974771287292242, + 0.0027126746717840433, + 0.004424711689352989, + 0.013513552956283092, + 0.0045909336768090725, + 0.003295580158010125, + 0.008622135035693645, + 0.006019141525030136, + 0.05047278478741646, + 0.01027694996446371, + 0.007613096386194229, + 0.008413980714976788, + 0.04506361857056618, + 0.007514127530157566, + 0.005908235441893339, + 0.003968864679336548, + 0.006493404973298311, + 0.011661551892757416, + 0.00535963661968708, + 0.005686114076524973, + 0.01104369293898344, + 0.009471691213548183, + 0.005069219507277012, + 0.0023736516013741493, + 0.003162823151797056, + 0.001858523115515709, + 0.0035447801928967237, + 0.00993937999010086, + 0.002743856515735388, + 0.00810152105987072, + 0.013791248202323914, + 0.011332142166793346, + 0.009294243529438972, + 0.007004345301538706, + 0.010692660696804523, + 0.03669263795018196, + 0.009846883825957775, + 0.029247762635350227, + 0.012102351523935795, + 0.013990700244903564, + 0.021709740161895752, + 0.006721906829625368, + 0.007520998828113079, + 0.042651526629924774, + 0.011328140273690224, + 0.012585966847836971, + 0.0038458893541246653, + 0.004243686329573393, + 0.007355816662311554, + 0.006167132873088121, + 0.0031243013218045235, + 0.0041332002729177475, + 0.00936979055404663, + 0.005991950631141663, + 0.014130808413028717, + 0.009532362222671509, + 0.02602345123887062, + 0.01587190479040146, + 0.008756364695727825, + 0.014228589832782745, + 0.009868940338492393, + 0.028640829026699066, + 0.014550409279763699, + 0.01538162399083376, + 0.019892435520887375, + 0.02786695957183838, + 0.015880389139056206, + 0.019392389804124832, + 0.020074767991900444, + 0.01986411027610302, + 0.010547646321356297, + 0.005212727468460798, + 0.05123467743396759, + 0.011842572130262852, + 0.007030343171209097, + 0.008118287660181522, + 0.004938540048897266, + 0.03610480949282646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.014218805357813835 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007234578370116651 + ], + "opposite_logprob_tokens": [ + -7.250723361968994 + ], + "logprob_reward": -0.0007234578370116651, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007234578370116651, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0959644615650177, + 0.04359608516097069, + 0.08771664649248123, + 0.18006934225559235, + 0.07181322574615479, + 0.08700239658355713, + 0.059061020612716675, + 0.09156622737646103, + 0.1426350325345993, + 0.1405755877494812 + ], + "credit_map": [ + 0.0085225198417902, + 0.014194853603839874, + 0.010045796632766724, + 0.017112433910369873, + 0.016005415469408035, + 0.010013616643846035, + 0.0127448420971632, + 0.007324984297156334, + 0.0035763191990554333, + 0.005208953749388456, + 0.006883102934807539, + 0.005568714812397957, + 0.003751183860003948, + 0.00536977406591177, + 0.00722472183406353, + 0.006013314705342054, + 0.014884771779179573, + 0.007656277623027563, + 0.011841744184494019, + 0.011520171537995338, + 0.010011809878051281, + 0.012390232644975185, + 0.010970711708068848, + 0.008440922014415264, + 0.03912097215652466, + 0.017512409016489983, + 0.028596559539437294, + 0.012448755092918873, + 0.022863712161779404, + 0.027956610545516014, + 0.013794481754302979, + 0.01777583174407482, + 0.0059434506110847, + 0.006107025314122438, + 0.00614358764141798, + 0.006116678472608328, + 0.006322239991277456, + 0.01736440509557724, + 0.007932428270578384, + 0.01588340476155281, + 0.009449978359043598, + 0.013161529786884785, + 0.013871119357645512, + 0.0064883301965892315, + 0.0058655994944274426, + 0.018116436898708344, + 0.009899834170937538, + 0.01014956645667553, + 0.004910817835479975, + 0.007909548468887806, + 0.010221774689853191, + 0.007229586597532034, + 0.00404696399345994, + 0.005636440124362707, + 0.00901784747838974, + 0.010088046081364155, + 0.010781479068100452, + 0.00572281563654542, + 0.017050450667738914, + 0.01088569313287735, + 0.005004654172807932, + 0.011777284555137157, + 0.008097555488348007, + 0.02224630117416382, + 0.012861134484410286, + 0.014004698023200035, + 0.03519722819328308, + 0.02067198045551777, + 0.011080487631261349, + 0.014217320829629898, + 0.008965565823018551, + 0.025636615231633186, + 0.012294488959014416, + 0.006594585254788399, + 0.04953308403491974, + 0.013460501097142696, + 0.008959162048995495, + 0.01077873446047306, + 0.0056498004123568535, + 0.033305227756500244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.0007234578370116651 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001335801207460463 + ], + "opposite_logprob_tokens": [ + -6.626335620880127 + ], + "logprob_reward": -0.001335801207460463, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001335801207460463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06418836861848831, + 0.05849010869860649, + 0.10415917634963989, + 0.052767395973205566, + 0.04234592244029045, + 0.22333158552646637, + 0.0432918407022953, + 0.18464361131191254, + 0.1195923238992691, + 0.10718968510627747 + ], + "credit_map": [ + 0.006294311955571175, + 0.008187015540897846, + 0.00842070672661066, + 0.0072075603529810905, + 0.012320476584136486, + 0.010283013805747032, + 0.006204553414136171, + 0.005270723253488541, + 0.004101297818124294, + 0.004459272604435682, + 0.014237405732274055, + 0.004499650094658136, + 0.004487978294491768, + 0.004436578135937452, + 0.017300937324762344, + 0.00496698496863246, + 0.008454550988972187, + 0.006947776768356562, + 0.009257536381483078, + 0.018218563869595528, + 0.010585491545498371, + 0.011123256757855415, + 0.023112572729587555, + 0.016459418460726738, + 0.00536767765879631, + 0.0040626684203743935, + 0.0038252610247582197, + 0.004375990014523268, + 0.005647704470902681, + 0.012645972892642021, + 0.007663071621209383, + 0.009179044514894485, + 0.00405401736497879, + 0.007546547334641218, + 0.005774264223873615, + 0.00482114776968956, + 0.005613094195723534, + 0.004913812968879938, + 0.005885747727006674, + 0.003737288061529398, + 0.02129696123301983, + 0.030884621664881706, + 0.030788259580731392, + 0.012670529074966908, + 0.016444077715277672, + 0.06846757233142853, + 0.021291295066475868, + 0.021488269791007042, + 0.004459488671272993, + 0.0040467423386871815, + 0.006480319891124964, + 0.004841939080506563, + 0.0027428537141531706, + 0.004382786341011524, + 0.007182061206549406, + 0.009155645966529846, + 0.02116195112466812, + 0.014434421434998512, + 0.03717980161309242, + 0.02201867289841175, + 0.01353436429053545, + 0.02053360641002655, + 0.014613473787903786, + 0.041167307645082474, + 0.011378232389688492, + 0.013716159388422966, + 0.021210534498095512, + 0.01788882538676262, + 0.011856799013912678, + 0.013921022415161133, + 0.01358196884393692, + 0.016038773581385612, + 0.01069035567343235, + 0.006468750070780516, + 0.025481149554252625, + 0.009562121704220772, + 0.012075463309884071, + 0.009660489857196808, + 0.006235942244529724, + 0.027015406638383865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.001335801207460463 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt upward throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender green vine climbing a dark, weathered wooden trellis visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of pristine, star-shaped white flowers with waxy petals visible on the vine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush garden foliage creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002816402819007635 + ], + "opposite_logprob_tokens": [ + -5.877816200256348 + ], + "logprob_reward": -0.002816402819007635, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002816402819007635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13348954916000366, + 0.10927072912454605, + 0.10539471358060837, + 0.05430347099900246, + 0.15125718712806702, + 0.1966453492641449, + 0.04584445804357529, + 0.07958526909351349, + 0.08903080970048904, + 0.035178497433662415 + ], + "credit_map": [ + 0.008837888017296791, + 0.009798121638596058, + 0.03117264434695244, + 0.025063080713152885, + 0.00991873163729906, + 0.025177879258990288, + 0.013181382790207863, + 0.01033979281783104, + 0.006255603861063719, + 0.005437679123133421, + 0.03198588266968727, + 0.006495522800832987, + 0.007053046487271786, + 0.007350399158895016, + 0.038048770278692245, + 0.0066438158974051476, + 0.009984239935874939, + 0.00662303576245904, + 0.01922629587352276, + 0.011376737616956234, + 0.01236246433109045, + 0.019739143550395966, + 0.014073082245886326, + 0.012009704485535622, + 0.006782941520214081, + 0.0041884006932377815, + 0.010099289938807487, + 0.0035743825137615204, + 0.004672234877943993, + 0.009822934865951538, + 0.006734420545399189, + 0.008428859524428844, + 0.016380103304982185, + 0.011490866541862488, + 0.011610413901507854, + 0.008260123431682587, + 0.012928497046232224, + 0.04333529621362686, + 0.012269305996596813, + 0.03498256206512451, + 0.019200928509235382, + 0.027070356532931328, + 0.025889968499541283, + 0.01230440754443407, + 0.013630440458655357, + 0.056360770016908646, + 0.019909678027033806, + 0.02227877825498581, + 0.004364451859146357, + 0.004014891106635332, + 0.00735330069437623, + 0.005442715249955654, + 0.002440253971144557, + 0.005436839535832405, + 0.0063352566212415695, + 0.010456743650138378, + 0.0077142538502812386, + 0.004725478123873472, + 0.019680708646774292, + 0.012414896860718727, + 0.004492685664445162, + 0.008017830550670624, + 0.005429907236248255, + 0.017109502106904984, + 0.007681912276893854, + 0.005966289900243282, + 0.01525413990020752, + 0.01938512548804283, + 0.006800721865147352, + 0.008000633679330349, + 0.011136375367641449, + 0.014805600047111511, + 0.004057866055518389, + 0.0028732125647366047, + 0.00717385858297348, + 0.004492092411965132, + 0.00387835456058383, + 0.004373863339424133, + 0.0026543387211859226, + 0.005674910265952349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.002816402819007635 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_13/index_tensor([4902, 4902])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_13/index_tensor([4902, 4902])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8c7d7a59bfa59d7d94c60a0e0572e61957eace37 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_13/index_tensor([4902, 4902])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "soft-focus, medium-length, static shot captures a behind-the-scenes rehearsal moment, rich with drama and authenticity. At the center of the frame is an enigmatic actress, clad in an intricate, gothic gown of black velvet and lace, embellished with silver thread embroidery at the hem and cuffs. Her cascading raven hair is streaked with striking strands of silver, and her face is delicately painted with smoky makeup and crimson lipstick. She is in a dimly lit rehearsal room, scattered with prop boxes and folding chairs. Spotlights slant from overhead, casting dappled light on the wooden floor. Her hands tightly grasp a massive, seemingly weighty, but in reality lightweight, prop blade. Its broad, dark silver blade is engraved with intricate runes. With the focus and power of a performer, she raises the prop sword high above her head, then, with a dramatic gesture, slams it down to the floor below. At the moment of impact, a pre-set, harmless white smoke erupted from the point of contact between the blade and the ground, swirling and spreading in all directions under the spotlight. However, at the completion of this epic moment, her expression, brimming with power just a moment before, instantly collapsed. She straightened her back, leaning back slightly, and let out a light, hearty laugh, a playful glint in her eyes, creating a captivating contrast between serious performance and relaxed authenticity.", + "seed": 13352, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_13/index_4902_seed_13352_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03255932098254562, + "vlm_logprob": -0.03255932098254562, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.03255932098254562, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07363779097795486 + ], + "opposite_logprob_tokens": [ + -2.6986377239227295 + ], + "logprob_reward": -0.07363779097795486, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07363779097795486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8064906597137451, + 0.04506092891097069, + 0.02038159780204296, + 0.026259059086441994, + 0.009816527366638184, + 0.006747932638972998, + 0.03076411597430706, + 0.011544476263225079, + 0.026406584307551384, + 0.016528243198990822 + ], + "credit_map": [ + 0.04840650409460068, + 0.0198928602039814, + 0.29493045806884766, + 0.026626799255609512, + 0.06195296719670296, + 0.019533660262823105, + 0.30174556374549866, + 0.03340176120400429, + 0.0027108185458928347, + 0.002718941308557987, + 0.01232737023383379, + 0.0037215633783489466, + 0.003943616058677435, + 0.002267336705699563, + 0.013390138745307922, + 0.003981142304837704, + 0.0018875576788559556, + 0.002654320327565074, + 0.002806642558425665, + 0.0014643706381320953, + 0.0024250636342912912, + 0.0023643041495233774, + 0.003329221159219742, + 0.003450117539614439, + 0.00230366550385952, + 0.0030504881870001554, + 0.004362146370112896, + 0.004206027369946241, + 0.0028476903680711985, + 0.0025037529412657022, + 0.003602032084017992, + 0.003383256494998932, + 0.0012314313789829612, + 0.001286907005123794, + 0.0014187326887622476, + 0.000978698837570846, + 0.0011073617497459054, + 0.0013525827089324594, + 0.0011894747149199247, + 0.001251338399015367, + 0.000699903757777065, + 0.001019682502374053, + 0.0008530545746907592, + 0.00071325758472085, + 0.000931193120777607, + 0.0007155296043492854, + 0.0008087566820904613, + 0.0010065542301163077, + 0.0032213719096034765, + 0.008355741389095783, + 0.0028152395971119404, + 0.002684524515643716, + 0.0057066334411501884, + 0.003286967519670725, + 0.0026713835541158915, + 0.0020222540479153395, + 0.0013038390316069126, + 0.0011268833186477423, + 0.0013967829290777445, + 0.0015517030842602253, + 0.0016951041761785746, + 0.001536507741548121, + 0.00157956313341856, + 0.0013540921499952674, + 0.0030807896982878447, + 0.003942009061574936, + 0.003163181245326996, + 0.0035813988652080297, + 0.0032160538248717785, + 0.003501737490296364, + 0.0026155083905905485, + 0.0033059059642255306, + 0.0015521617606282234, + 0.0024734216276556253, + 0.0023072417825460434, + 0.0011859190417453647, + 0.0025943086948245764, + 0.0018312579486519098, + 0.0021526890341192484, + 0.002431242261081934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.07363779097795486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007648587692528963 + ], + "opposite_logprob_tokens": [ + -4.882648468017578 + ], + "logprob_reward": -0.007648587692528963, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007648587692528963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7271190881729126, + 0.03392581269145012, + 0.026902910321950912, + 0.046179015189409256, + 0.01432208251208067, + 0.012862203642725945, + 0.03342578560113907, + 0.017535371705889702, + 0.03954137861728668, + 0.04818633943796158 + ], + "credit_map": [ + 0.04399070143699646, + 0.018559269607067108, + 0.2648826837539673, + 0.023927679285407066, + 0.05599556118249893, + 0.017844950780272484, + 0.2719668745994568, + 0.029951361939311028, + 0.0021312220487743616, + 0.0028839462902396917, + 0.008413635194301605, + 0.002648905385285616, + 0.0034491787664592266, + 0.0016898137982934713, + 0.009898492135107517, + 0.0028106204699724913, + 0.0016535916365683079, + 0.0026964799035340548, + 0.004575188737362623, + 0.0020834060851484537, + 0.0028104332741349936, + 0.0028367573395371437, + 0.004129145294427872, + 0.00611790781840682, + 0.0039502899162471294, + 0.009170162491500378, + 0.006182488054037094, + 0.006222112569957972, + 0.005928410682827234, + 0.003908791113644838, + 0.006634878925979137, + 0.004181882366538048, + 0.0016209547175094485, + 0.0020689729135483503, + 0.0019627753645181656, + 0.001216800301335752, + 0.001756830490194261, + 0.0020357619505375624, + 0.0015298291109502316, + 0.0021301573142409325, + 0.0014865519478917122, + 0.0012247756822034717, + 0.0017461130628362298, + 0.0011997041292488575, + 0.0014882959658280015, + 0.0018204125808551908, + 0.0016563516110181808, + 0.0022399984300136566, + 0.0035892713349312544, + 0.009134192951023579, + 0.0029293333645910025, + 0.0026425179094076157, + 0.006486020050942898, + 0.0036322087980806828, + 0.002768830629065633, + 0.0022434082347899675, + 0.00205800193361938, + 0.0019379090517759323, + 0.0018354986095800996, + 0.0023594165686517954, + 0.0024643477518111467, + 0.0026317366864532232, + 0.002324870554730296, + 0.0019235893851146102, + 0.004706281702965498, + 0.005777485202997923, + 0.004890093579888344, + 0.005238568875938654, + 0.005327913910150528, + 0.005061263218522072, + 0.0038054774049669504, + 0.004734294954687357, + 0.004900278523564339, + 0.013225873932242393, + 0.00693743210285902, + 0.003555465256795287, + 0.005708923563361168, + 0.0042290836572647095, + 0.005640783347189426, + 0.003988498356193304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.007648587692528963 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029811859130859375 + ], + "opposite_logprob_tokens": [ + -3.5298118591308594 + ], + "logprob_reward": -0.029811859130859375, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029811859130859375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6388264894485474, + 0.06800520420074463, + 0.08085101842880249, + 0.030875537544488907, + 0.016185753047466278, + 0.02623596414923668, + 0.016320785507559776, + 0.03084227815270424, + 0.04530597850680351, + 0.04655098542571068 + ], + "credit_map": [ + 0.03843960538506508, + 0.015780009329319, + 0.23148655891418457, + 0.02130137011408806, + 0.04869089275598526, + 0.016422221437096596, + 0.23968316614627838, + 0.027022695168852806, + 0.005408806726336479, + 0.011956528760492802, + 0.008371149189770222, + 0.008601192384958267, + 0.007017821539193392, + 0.005314395297318697, + 0.012820719741284847, + 0.008514597080647945, + 0.005460756365209818, + 0.008711330592632294, + 0.014738062396645546, + 0.007186210248619318, + 0.010428328067064285, + 0.007221877574920654, + 0.01302217599004507, + 0.014082284644246101, + 0.0031924389768391848, + 0.003976215608417988, + 0.004363426472991705, + 0.0057730237022042274, + 0.00302498834207654, + 0.002792859449982643, + 0.003927422687411308, + 0.003825166029855609, + 0.0020505364518612623, + 0.0021611941047012806, + 0.0017130332998931408, + 0.0016806893981993198, + 0.0017697332659736276, + 0.0020017949864268303, + 0.0023081989493221045, + 0.002500576199963689, + 0.002579446416348219, + 0.0046884543262422085, + 0.003039120463654399, + 0.0023838181514292955, + 0.004813306964933872, + 0.0025284814182668924, + 0.003088075201958418, + 0.003115263534709811, + 0.0024856661912053823, + 0.0017825175309553742, + 0.001668368000537157, + 0.0013992253225296736, + 0.0027691808063536882, + 0.0023804137017577887, + 0.0016960598295554519, + 0.002139355055987835, + 0.0036232811398804188, + 0.0038413042202591896, + 0.003629842074587941, + 0.0047241197898983955, + 0.004286159761250019, + 0.0037224895786494017, + 0.0036113846581429243, + 0.003403700655326247, + 0.004747758619487286, + 0.006764216348528862, + 0.004820678848773241, + 0.007082507945597172, + 0.005205017048865557, + 0.005110777914524078, + 0.005603932775557041, + 0.005971094593405724, + 0.004588770214468241, + 0.007537174969911575, + 0.007383641321212053, + 0.0041524991393089294, + 0.005973015446215868, + 0.004515295382589102, + 0.005105397664010525, + 0.007295193616300821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.029811859130859375 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641703985631466 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.008641703985631466, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641703985631466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4704785645008087, + 0.0394693985581398, + 0.08605509251356125, + 0.04161050170660019, + 0.022927040234208107, + 0.029560646042227745, + 0.061325546354055405, + 0.09056264907121658, + 0.06691592931747437, + 0.09109468013048172 + ], + "credit_map": [ + 0.027548234909772873, + 0.012159070000052452, + 0.1736675351858139, + 0.015857478603720665, + 0.03533656522631645, + 0.011582186445593834, + 0.17485851049423218, + 0.01946892775595188, + 0.0023771135602146387, + 0.0038427934050559998, + 0.010791794396936893, + 0.0026936789508908987, + 0.0038457047194242477, + 0.0021338791120797396, + 0.011736744083464146, + 0.002047684509307146, + 0.005930899642407894, + 0.01038374099880457, + 0.014970837160944939, + 0.007737666368484497, + 0.008655806072056293, + 0.006382448598742485, + 0.015320238657295704, + 0.016673443838953972, + 0.0032691024243831635, + 0.005834836978465319, + 0.006289211567491293, + 0.005813085939735174, + 0.0059266765601933, + 0.0032045478001236916, + 0.007541550323367119, + 0.0037314852233976126, + 0.002336797770112753, + 0.002933853305876255, + 0.0022874176502227783, + 0.0026300440076738596, + 0.0033700913190841675, + 0.001903990749269724, + 0.0049315462820231915, + 0.002533294726163149, + 0.0031240193638950586, + 0.003827012376859784, + 0.003147873328998685, + 0.002744099823758006, + 0.004108746070414782, + 0.0030648037791252136, + 0.003508123103529215, + 0.006035962142050266, + 0.007983651012182236, + 0.009773580357432365, + 0.005419946275651455, + 0.006319490727037191, + 0.009631008841097355, + 0.0073654246516525745, + 0.007540004327893257, + 0.007292426656931639, + 0.008056264370679855, + 0.012716684490442276, + 0.009309373795986176, + 0.012960009276866913, + 0.014448154717683792, + 0.008496610447764397, + 0.013991586863994598, + 0.010583953931927681, + 0.008638515137135983, + 0.009603931568562984, + 0.009271894581615925, + 0.008944137021899223, + 0.007446781732141972, + 0.0060666571371257305, + 0.007820552214980125, + 0.009123443625867367, + 0.009927017614245415, + 0.016039609909057617, + 0.017443183809518814, + 0.007931088097393513, + 0.012322464026510715, + 0.007106986828148365, + 0.00911675300449133, + 0.01120756659656763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.008641703985631466 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430566631257534 + ], + "opposite_logprob_tokens": [ + -3.1680567264556885 + ], + "logprob_reward": -0.0430566631257534, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430566631257534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4317622482776642, + 0.06867743283510208, + 0.25046801567077637, + 0.03365328535437584, + 0.04108310118317604, + 0.045733772218227386, + 0.03536829724907875, + 0.01981760933995247, + 0.02285809814929962, + 0.05057813227176666 + ], + "credit_map": [ + 0.02712339721620083, + 0.01191500574350357, + 0.15466941893100739, + 0.014479785226285458, + 0.03226344287395477, + 0.012279138900339603, + 0.15922845900058746, + 0.019803611561655998, + 0.005243104416877031, + 0.0036057636607438326, + 0.02064036764204502, + 0.004188344348222017, + 0.004733634181320667, + 0.004576482344418764, + 0.019012916833162308, + 0.006676820106804371, + 0.016160018742084503, + 0.023041538894176483, + 0.05044744908809662, + 0.021026453003287315, + 0.019420627504587173, + 0.02168700285255909, + 0.04001510143280029, + 0.05866982787847519, + 0.0024253265000879765, + 0.0036238564644008875, + 0.004958434030413628, + 0.0049515170976519585, + 0.004523762036114931, + 0.003066595643758774, + 0.005003142636269331, + 0.005100652109831572, + 0.005832651164382696, + 0.0032307864166796207, + 0.0050957053899765015, + 0.0033654775470495224, + 0.004254909697920084, + 0.006744374521076679, + 0.006286127958446741, + 0.006273068021982908, + 0.004899169318377972, + 0.005851402413100004, + 0.006437498144805431, + 0.0037390636280179024, + 0.0060336836613714695, + 0.006277431733906269, + 0.004811516497284174, + 0.007684008684009314, + 0.005082919728010893, + 0.006674453616142273, + 0.003445746609941125, + 0.0029116931837052107, + 0.005651616957038641, + 0.004822219721972942, + 0.003139267908409238, + 0.00364038092084229, + 0.002962492173537612, + 0.0020344078075140715, + 0.002442734083160758, + 0.0028274154756218195, + 0.002390797482803464, + 0.002404003869742155, + 0.0023286642972379923, + 0.0024270936846733093, + 0.0029795418959110975, + 0.0031822065357118845, + 0.0025008548982441425, + 0.002954989904537797, + 0.0030204253271222115, + 0.0022063092328608036, + 0.002855042228475213, + 0.0031587278936058283, + 0.004996359348297119, + 0.011685931123793125, + 0.009110876359045506, + 0.003977484535425901, + 0.006227579899132252, + 0.003321555908769369, + 0.0053648133762180805, + 0.005893532652407885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0430566631257534 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "soft-focus, medium-length, static shot captures a behind-the-scenes rehearsal moment, rich with drama and authenticity. At the center of the frame is an enigmatic actress, clad in an intricate, gothic gown of black velvet and lace, embellished with silver thread embroidery at the hem and cuffs. Her cascading raven hair is streaked with striking strands of silver, and her face is delicately painted with smoky makeup and crimson lipstick. She is in a dimly lit rehearsal room, scattered with prop boxes and folding chairs. Spotlights slant from overhead, casting dappled light on the wooden floor. Her hands tightly grasp a massive, seemingly weighty, but in reality lightweight, prop blade. Its broad, dark silver blade is engraved with intricate runes. With the focus and power of a performer, she raises the prop sword high above her head, then, with a dramatic gesture, slams it down to the floor below. At the moment of impact, a pre-set, harmless white smoke erupted from the point of contact between the blade and the ground, swirling and spreading in all directions under the spotlight. However, at the completion of this epic moment, her expression, brimming with power just a moment before, instantly collapsed. She straightened her back, leaning back slightly, and let out a light, hearty laugh, a playful glint in her eyes, creating a captivating contrast between serious performance and relaxed authenticity.", + "seed": 13353, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_13/index_4902_seed_13353_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02990641286596656, + "vlm_logprob": -0.02990641286596656, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.02990641286596656, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04713742062449455 + ], + "opposite_logprob_tokens": [ + -3.172137498855591 + ], + "logprob_reward": -0.04713742062449455, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04713742062449455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32098591327667236, + 0.08338532596826553, + 0.05103132873773575, + 0.09177462011575699, + 0.0727546364068985, + 0.08963519334793091, + 0.06483151763677597, + 0.05822676047682762, + 0.06864269077777863, + 0.09873204678297043 + ], + "credit_map": [ + 0.04564683884382248, + 0.04345674440264702, + 0.03569239377975464, + 0.03044402226805687, + 0.0339418426156044, + 0.03255626931786537, + 0.050890449434518814, + 0.048357363790273666, + 0.010456524789333344, + 0.013897383585572243, + 0.007166404742747545, + 0.009680650196969509, + 0.011642225086688995, + 0.008779264986515045, + 0.01438696589320898, + 0.00737590529024601, + 0.005691349506378174, + 0.006750371772795916, + 0.005045526195317507, + 0.00802831444889307, + 0.006707155145704746, + 0.005765883252024651, + 0.00723746232688427, + 0.005805265158414841, + 0.010780420154333115, + 0.01297084428369999, + 0.00968689564615488, + 0.011821139603853226, + 0.00831553153693676, + 0.015709277242422104, + 0.01333505380898714, + 0.0091554531827569, + 0.008413366042077541, + 0.009006477892398834, + 0.008944780565798283, + 0.010152879171073437, + 0.009460516273975372, + 0.009143436327576637, + 0.01095607504248619, + 0.006677104625850916, + 0.00947565771639347, + 0.01165495440363884, + 0.009918189607560635, + 0.01232586894184351, + 0.01140952855348587, + 0.00942888855934143, + 0.012931284494698048, + 0.012490821070969105, + 0.010195432230830193, + 0.010037777945399284, + 0.005941835697740316, + 0.007751368917524815, + 0.008218039758503437, + 0.0070583391934633255, + 0.00850446242839098, + 0.007124259602278471, + 0.008676307275891304, + 0.008444201201200485, + 0.008694026619195938, + 0.006149302236735821, + 0.00685573136433959, + 0.006131691392511129, + 0.006193399894982576, + 0.007082100957632065, + 0.007107708137482405, + 0.00819146167486906, + 0.007222628220915794, + 0.009357798844575882, + 0.009452583268284798, + 0.00750767532736063, + 0.01135042030364275, + 0.008452417328953743, + 0.010498903691768646, + 0.010808330960571766, + 0.011772682890295982, + 0.013790987432003021, + 0.013130000792443752, + 0.009204087778925896, + 0.01740192621946335, + 0.01212511956691742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.04713742062449455 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865340419113636 + ], + "opposite_logprob_tokens": [ + -4.75865364074707 + ], + "logprob_reward": -0.00865340419113636, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865340419113636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2692510485649109, + 0.08441759645938873, + 0.0806027352809906, + 0.10260077565908432, + 0.08006227016448975, + 0.07867227494716644, + 0.08652099221944809, + 0.06966998428106308, + 0.05951764062047005, + 0.08868467062711716 + ], + "credit_map": [ + 0.03956348076462746, + 0.036322444677352905, + 0.030915332958102226, + 0.02621867135167122, + 0.03464099019765854, + 0.025607649236917496, + 0.035359304398298264, + 0.04062319174408913, + 0.011823233217000961, + 0.014047758653759956, + 0.0095799146220088, + 0.010146424174308777, + 0.008847975172102451, + 0.010352784767746925, + 0.009661982767283916, + 0.009957526810467243, + 0.008636928163468838, + 0.01338198035955429, + 0.006595222745090723, + 0.011888144537806511, + 0.013866721652448177, + 0.00860920175909996, + 0.009967945516109467, + 0.007656589150428772, + 0.012284758500754833, + 0.015239240601658821, + 0.01192625891417265, + 0.012199988588690758, + 0.008849581703543663, + 0.015408777631819248, + 0.014536500908434391, + 0.012155677191913128, + 0.009345526807010174, + 0.010697794146835804, + 0.007773793768137693, + 0.01086942758411169, + 0.011822972446680069, + 0.011807728558778763, + 0.010478794574737549, + 0.007266233209520578, + 0.008548272773623466, + 0.010595753788948059, + 0.008231026120483875, + 0.010625403374433517, + 0.012513668276369572, + 0.009264606982469559, + 0.011214778758585453, + 0.007678760215640068, + 0.013772021979093552, + 0.014565970748662949, + 0.009862239472568035, + 0.009361866861581802, + 0.010260052047669888, + 0.00973183661699295, + 0.009778094477951527, + 0.00918891467154026, + 0.009981135837733746, + 0.008689210750162601, + 0.010940248146653175, + 0.008364830166101456, + 0.007985314354300499, + 0.006802527699619532, + 0.008015717379748821, + 0.008890998549759388, + 0.005494387354701757, + 0.008459262549877167, + 0.0077526927925646305, + 0.008345702663064003, + 0.00877147726714611, + 0.006730366498231888, + 0.006922288332134485, + 0.007041463628411293, + 0.009146143682301044, + 0.009919545613229275, + 0.010714146308600903, + 0.011583312414586544, + 0.014817965216934681, + 0.008951525203883648, + 0.01590842567384243, + 0.007643600460141897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.00865340419113636 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306408390402794 + ], + "opposite_logprob_tokens": [ + -3.1680641174316406 + ], + "logprob_reward": -0.04306408390402794, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306408390402794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2834206223487854, + 0.1758042573928833, + 0.05680454149842262, + 0.09231960773468018, + 0.057260334491729736, + 0.08234583586454391, + 0.0966932475566864, + 0.052506063133478165, + 0.06274795532226562, + 0.040097545832395554 + ], + "credit_map": [ + 0.04099678993225098, + 0.0495077520608902, + 0.03418709710240364, + 0.029194386675953865, + 0.029382815584540367, + 0.028441304340958595, + 0.03072441928088665, + 0.0409860722720623, + 0.02866177447140217, + 0.024176977574825287, + 0.020292231813073158, + 0.019623078405857086, + 0.018351180478930473, + 0.015453147701919079, + 0.02392161451280117, + 0.02532423660159111, + 0.008500497788190842, + 0.00835068617016077, + 0.006203196477144957, + 0.00621643103659153, + 0.00825210940092802, + 0.006546798162162304, + 0.00657857209444046, + 0.006156248040497303, + 0.010395608842372894, + 0.011867981404066086, + 0.009566736407577991, + 0.011710270307958126, + 0.010382314212620258, + 0.010622994042932987, + 0.012794298119843006, + 0.014979403465986252, + 0.006687032524496317, + 0.0072156996466219425, + 0.005845232401043177, + 0.008380560204386711, + 0.006935900542885065, + 0.007161287125200033, + 0.009888792410492897, + 0.005145832430571318, + 0.009523577988147736, + 0.009897801093757153, + 0.006847964599728584, + 0.01339214202016592, + 0.015120413154363632, + 0.01041574776172638, + 0.011157236061990261, + 0.005990956909954548, + 0.016144419088959694, + 0.019073784351348877, + 0.01065951306372881, + 0.011699812486767769, + 0.009684787131845951, + 0.011556035839021206, + 0.008620074018836021, + 0.009254821576178074, + 0.007189927622675896, + 0.00689270393922925, + 0.008793591521680355, + 0.006701264530420303, + 0.00779738649725914, + 0.005385053344070911, + 0.004797223024070263, + 0.0049489145167171955, + 0.005890781059861183, + 0.0069959089159965515, + 0.0074821580201387405, + 0.00847818423062563, + 0.010030600242316723, + 0.0061402334831655025, + 0.009936722926795483, + 0.007793361321091652, + 0.0042352257296442986, + 0.005085548851639032, + 0.004288174211978912, + 0.006526515353471041, + 0.0066991387866437435, + 0.004237479064613581, + 0.005735001992434263, + 0.0032904623076319695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.04306408390402794 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305494949221611 + ], + "opposite_logprob_tokens": [ + -3.1680550575256348 + ], + "logprob_reward": -0.04305494949221611, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305494949221611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1119200736284256, + 0.21000364422798157, + 0.05243847519159317, + 0.16785211861133575, + 0.039099063724279404, + 0.1442868709564209, + 0.11115328967571259, + 0.07288821041584015, + 0.06348211318254471, + 0.026876090094447136 + ], + "credit_map": [ + 0.016262007877230644, + 0.013130332343280315, + 0.014253431931138039, + 0.011437598615884781, + 0.01491633802652359, + 0.011625446379184723, + 0.018388137221336365, + 0.01190677098929882, + 0.026636317372322083, + 0.03523422032594681, + 0.015433565713465214, + 0.020853012800216675, + 0.03855693340301514, + 0.021498985588550568, + 0.0352056510746479, + 0.016584966331720352, + 0.007110638543963432, + 0.0062834895215928555, + 0.003925093915313482, + 0.007247353903949261, + 0.00905939657241106, + 0.004869794473052025, + 0.008077017031610012, + 0.00586568983271718, + 0.01780085638165474, + 0.024131091311573982, + 0.01628437079489231, + 0.029957571998238564, + 0.015337107703089714, + 0.0258424561470747, + 0.023746367543935776, + 0.01475229300558567, + 0.004372359719127417, + 0.0040368689224123955, + 0.003959067165851593, + 0.005621074698865414, + 0.007638273760676384, + 0.0038634352385997772, + 0.006131294649094343, + 0.0034766867756843567, + 0.016115082427859306, + 0.014322042465209961, + 0.009823884814977646, + 0.023181365802884102, + 0.027918260544538498, + 0.016680307686328888, + 0.023495741188526154, + 0.012750182300806046, + 0.017964687198400497, + 0.02121252939105034, + 0.011967761442065239, + 0.012245248071849346, + 0.012361254543066025, + 0.013482589274644852, + 0.010938656516373158, + 0.010980556719005108, + 0.010814669542014599, + 0.01108077634125948, + 0.011137074790894985, + 0.008018650114536285, + 0.009553894400596619, + 0.008475657552480698, + 0.005979754962027073, + 0.007827733643352985, + 0.006185328587889671, + 0.0073963613249361515, + 0.010070385411381721, + 0.009669387713074684, + 0.009033989161252975, + 0.006990025285631418, + 0.006832757964730263, + 0.0073038795962929726, + 0.0026748317759484053, + 0.0033518706914037466, + 0.0035838827025145292, + 0.0028477637097239494, + 0.005260576494038105, + 0.002264306414872408, + 0.004684150218963623, + 0.0022087078541517258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.04305494949221611 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762220611795783 + ], + "opposite_logprob_tokens": [ + -4.882622241973877 + ], + "logprob_reward": -0.00762220611795783, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762220611795783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30418282747268677, + 0.07903340458869934, + 0.08153776824474335, + 0.09295140206813812, + 0.07809098064899445, + 0.07227208465337753, + 0.07265464961528778, + 0.07123568654060364, + 0.10275442898273468, + 0.04528677463531494 + ], + "credit_map": [ + 0.041908469051122665, + 0.06206553429365158, + 0.03145444020628929, + 0.03299318253993988, + 0.02280145138502121, + 0.03869469463825226, + 0.03239186853170395, + 0.04187319427728653, + 0.014245533384382725, + 0.012861423194408417, + 0.008307860232889652, + 0.006151732988655567, + 0.01051715575158596, + 0.01418483629822731, + 0.007059136871248484, + 0.005705725401639938, + 0.009282791055738926, + 0.013907529413700104, + 0.005997879430651665, + 0.013485373929142952, + 0.009608341380953789, + 0.009857943281531334, + 0.01128778699785471, + 0.00811013299971819, + 0.011878976598381996, + 0.01774279959499836, + 0.011785999871790409, + 0.015593207441270351, + 0.007028623018413782, + 0.013165164738893509, + 0.008088482543826103, + 0.0076681459322571754, + 0.010300569236278534, + 0.013528409413993359, + 0.009386214427649975, + 0.01064389105886221, + 0.006928286515176296, + 0.010781371034681797, + 0.007284688763320446, + 0.009237555786967278, + 0.008939499966800213, + 0.012425806373357773, + 0.009925219230353832, + 0.009402686730027199, + 0.006568662356585264, + 0.009773720987141132, + 0.005497622303664684, + 0.009738866239786148, + 0.013744713738560677, + 0.014695234596729279, + 0.008068197406828403, + 0.0081234872341156, + 0.006768718361854553, + 0.008660789579153061, + 0.003411124926060438, + 0.009182385168969631, + 0.009228247217833996, + 0.01859324611723423, + 0.007914645597338676, + 0.008292741142213345, + 0.007287501823157072, + 0.007608715910464525, + 0.005233552772551775, + 0.0070770359598100185, + 0.014756458811461926, + 0.01827409118413925, + 0.0187902320176363, + 0.013436718843877316, + 0.010087399743497372, + 0.010568195953965187, + 0.005853333510458469, + 0.010988005436956882, + 0.00454909959807992, + 0.004639721941202879, + 0.006104125175625086, + 0.008952152915298939, + 0.006042120046913624, + 0.004207814112305641, + 0.00550671573728323, + 0.005285023711621761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.00762220611795783 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "soft-focus, medium-length, static shot captures a behind-the-scenes rehearsal moment, rich with drama and authenticity. At the center of the frame is an enigmatic actress, clad in an intricate, gothic gown of black velvet and lace, embellished with silver thread embroidery at the hem and cuffs. Her cascading raven hair is streaked with striking strands of silver, and her face is delicately painted with smoky makeup and crimson lipstick. She is in a dimly lit rehearsal room, scattered with prop boxes and folding chairs. Spotlights slant from overhead, casting dappled light on the wooden floor. Her hands tightly grasp a massive, seemingly weighty, but in reality lightweight, prop blade. Its broad, dark silver blade is engraved with intricate runes. With the focus and power of a performer, she raises the prop sword high above her head, then, with a dramatic gesture, slams it down to the floor below. At the moment of impact, a pre-set, harmless white smoke erupted from the point of contact between the blade and the ground, swirling and spreading in all directions under the spotlight. However, at the completion of this epic moment, her expression, brimming with power just a moment before, instantly collapsed. She straightened her back, leaning back slightly, and let out a light, hearty laugh, a playful glint in her eyes, creating a captivating contrast between serious performance and relaxed authenticity.", + "seed": 13354, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_13/index_4902_seed_13354_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02602827809751034, + "vlm_logprob": -0.02602827809751034, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.02602827809751034, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05059048905968666 + ], + "opposite_logprob_tokens": [ + -3.0505905151367188 + ], + "logprob_reward": -0.05059048905968666, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05059048905968666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16322651505470276, + 0.08911354839801788, + 0.044545646756887436, + 0.05362071841955185, + 0.07674866914749146, + 0.28500863909721375, + 0.14748536050319672, + 0.05734632909297943, + 0.042565684765577316, + 0.040338944643735886 + ], + "credit_map": [ + 0.016734259203076363, + 0.015175025910139084, + 0.03215829282999039, + 0.019895238801836967, + 0.018705179914832115, + 0.03031257353723049, + 0.01350077148526907, + 0.016745170578360558, + 0.005906878970563412, + 0.008265120908617973, + 0.014687633141875267, + 0.007603150326758623, + 0.006092519033700228, + 0.009360002353787422, + 0.010538632981479168, + 0.026659613475203514, + 0.006062633823603392, + 0.00496772862970829, + 0.006955505348742008, + 0.005787539295852184, + 0.0042938366532325745, + 0.006607830990105867, + 0.004696013871580362, + 0.005174554884433746, + 0.005924545228481293, + 0.004455192014575005, + 0.008426553569734097, + 0.007899066433310509, + 0.006964132189750671, + 0.00812425184994936, + 0.0066223773173987865, + 0.00520459795370698, + 0.008485271595418453, + 0.007565545383840799, + 0.009251599200069904, + 0.014004398137331009, + 0.010928415693342686, + 0.01050965953618288, + 0.009886020794510841, + 0.0061177522875368595, + 0.040039535611867905, + 0.01984994485974312, + 0.04047177731990814, + 0.023068243637681007, + 0.02211086079478264, + 0.03859761357307434, + 0.03781357780098915, + 0.06305709481239319, + 0.019992657005786896, + 0.011820849031209946, + 0.024160044267773628, + 0.014871452935039997, + 0.014935262501239777, + 0.02036602795124054, + 0.01921011507511139, + 0.02212894894182682, + 0.007915781810879707, + 0.006654160097241402, + 0.006715881172567606, + 0.009727448225021362, + 0.006832744926214218, + 0.006864223163574934, + 0.007324098609387875, + 0.005311992019414902, + 0.0057084569707512856, + 0.003551076166331768, + 0.007328985724598169, + 0.004252087324857712, + 0.00529636861756444, + 0.004747773986309767, + 0.005126400850713253, + 0.006554535124450922, + 0.005231862887740135, + 0.004327068105340004, + 0.006196275819092989, + 0.00443241698667407, + 0.004805696662515402, + 0.008214491419494152, + 0.0046253022737801075, + 0.00250582629814744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.05059048905968666 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009802419692277908 + ], + "opposite_logprob_tokens": [ + -4.634802341461182 + ], + "logprob_reward": -0.009802419692277908, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009802419692277908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47550323605537415, + 0.06488196551799774, + 0.03310761600732803, + 0.026599183678627014, + 0.12859703600406647, + 0.08897311985492706, + 0.043049152940511703, + 0.04344998672604561, + 0.034352533519268036, + 0.0614861324429512 + ], + "credit_map": [ + 0.0394248440861702, + 0.03202400356531143, + 0.11680587381124496, + 0.04005858674645424, + 0.030676400288939476, + 0.12715347111225128, + 0.03329020366072655, + 0.056069836020469666, + 0.004018625244498253, + 0.005403375718742609, + 0.0108376145362854, + 0.005040218122303486, + 0.004957912489771843, + 0.0065725683234632015, + 0.008019091561436653, + 0.02003256231546402, + 0.0040863556787371635, + 0.002486458048224449, + 0.006046669092029333, + 0.004151667468249798, + 0.002598605817183852, + 0.005193809978663921, + 0.003991171717643738, + 0.004552876111119986, + 0.0023450455628335476, + 0.0022707332391291857, + 0.004219364374876022, + 0.003319999435916543, + 0.0034666149877011776, + 0.0035568904131650925, + 0.003881364595144987, + 0.00353916990570724, + 0.018291020765900612, + 0.02202598750591278, + 0.012595554813742638, + 0.019220830872654915, + 0.019747581332921982, + 0.016177933663129807, + 0.012662846595048904, + 0.00787529069930315, + 0.012545608915388584, + 0.006558756809681654, + 0.012359043583273888, + 0.0069807227700948715, + 0.007302785292267799, + 0.01253981702029705, + 0.009534208104014397, + 0.021152179688215256, + 0.005510509014129639, + 0.003420209512114525, + 0.006809770129621029, + 0.004994810093194246, + 0.004584910348057747, + 0.005581239238381386, + 0.005635375156998634, + 0.006512331310659647, + 0.005793799180537462, + 0.004867893178015947, + 0.004916581325232983, + 0.005994741339236498, + 0.005050298757851124, + 0.00543954549357295, + 0.00501451687887311, + 0.006372610107064247, + 0.00469444552436471, + 0.002462824573740363, + 0.006708336994051933, + 0.0030215964652597904, + 0.004097448196262121, + 0.004281021188944578, + 0.003662700532004237, + 0.005424159578979015, + 0.007416964508593082, + 0.00815902091562748, + 0.009415990673005581, + 0.00541208079084754, + 0.00883446354418993, + 0.011592194437980652, + 0.006743809208273888, + 0.003911610227078199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.009802419692277908 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810487315058708 + ], + "opposite_logprob_tokens": [ + -3.288104772567749 + ], + "logprob_reward": -0.03810487315058708, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810487315058708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26587358117103577, + 0.11773250997066498, + 0.0303631778806448, + 0.03917543217539787, + 0.045378170907497406, + 0.2712908387184143, + 0.14399594068527222, + 0.031078578904271126, + 0.023690953850746155, + 0.03142080828547478 + ], + "credit_map": [ + 0.022876452654600143, + 0.027118118479847908, + 0.055111825466156006, + 0.023737207055091858, + 0.026489898562431335, + 0.06208677589893341, + 0.022767430171370506, + 0.025685861706733704, + 0.00828141812235117, + 0.009314987808465958, + 0.01661168411374092, + 0.0144587941467762, + 0.017514752224087715, + 0.014089989475905895, + 0.016209924593567848, + 0.021250959485769272, + 0.0029096431098878384, + 0.002864992246031761, + 0.0051818047650158405, + 0.004677819553762674, + 0.0026007858105003834, + 0.003442704677581787, + 0.003729933174327016, + 0.00495549151673913, + 0.004582130815833807, + 0.0031048767268657684, + 0.0062027582898736, + 0.005459497217088938, + 0.004680508282035589, + 0.005953692365437746, + 0.004506027325987816, + 0.004685940220952034, + 0.005208651535212994, + 0.005121092312037945, + 0.004377448931336403, + 0.00506203668192029, + 0.010571141727268696, + 0.006099398713558912, + 0.005750918295234442, + 0.0031874841079115868, + 0.03901102393865585, + 0.01797555387020111, + 0.03944816440343857, + 0.02185533381998539, + 0.020117226988077164, + 0.03777618706226349, + 0.03968837112188339, + 0.05541899800300598, + 0.018387025222182274, + 0.01146476250141859, + 0.021058468148112297, + 0.014656155370175838, + 0.015467059798538685, + 0.018575327470898628, + 0.01941877417266369, + 0.024968378245830536, + 0.004716953262686729, + 0.0020406455732882023, + 0.004159294068813324, + 0.002745125675573945, + 0.002655091928318143, + 0.005123695358633995, + 0.0036944567691534758, + 0.005943316500633955, + 0.0034343816805630922, + 0.001924328738823533, + 0.004651183262467384, + 0.002144584199413657, + 0.0021598064340651035, + 0.003206250723451376, + 0.002430166583508253, + 0.003740250365808606, + 0.0035272922832518816, + 0.0030023190192878246, + 0.005052146036177874, + 0.0034702587872743607, + 0.003921405877918005, + 0.0035237593110650778, + 0.00643491093069315, + 0.0024887127801775932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.03810487315058708 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073680594563484 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073680594563484, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073680594563484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13579347729682922, + 0.05005326494574547, + 0.13975889980793, + 0.1401493102312088, + 0.1034693792462349, + 0.22770637273788452, + 0.10791569948196411, + 0.02499079518020153, + 0.04344774782657623, + 0.02671504020690918 + ], + "credit_map": [ + 0.013086207211017609, + 0.011481119319796562, + 0.026348810642957687, + 0.014051140286028385, + 0.01289891917258501, + 0.02803095616400242, + 0.01237223669886589, + 0.01752409338951111, + 0.0037763628643006086, + 0.004769863095134497, + 0.006886019371449947, + 0.006344204768538475, + 0.004931190051138401, + 0.006087924353778362, + 0.006268847733736038, + 0.010988850146532059, + 0.01478003989905119, + 0.010866162367165089, + 0.023188285529613495, + 0.020145989954471588, + 0.013566454872488976, + 0.018032729625701904, + 0.019438613206148148, + 0.019740626215934753, + 0.014722030609846115, + 0.00865214690566063, + 0.022858653217554092, + 0.024509338662028313, + 0.009599406272172928, + 0.016931917518377304, + 0.01897585764527321, + 0.02389996126294136, + 0.013842632062733173, + 0.015103709883987904, + 0.009160944260656834, + 0.010689431801438332, + 0.01400749757885933, + 0.01470130030065775, + 0.011195671744644642, + 0.014768187887966633, + 0.031188171356916428, + 0.016144508495926857, + 0.030304407700896263, + 0.017970703542232513, + 0.01929103583097458, + 0.02860221266746521, + 0.02818620763719082, + 0.05601911619305611, + 0.014660408720374107, + 0.008220302872359753, + 0.017633680254220963, + 0.01196913979947567, + 0.011295458301901817, + 0.014953329227864742, + 0.01484155748039484, + 0.014341826550662518, + 0.003195954253897071, + 0.0024602157063782215, + 0.0033268427941948175, + 0.002346431603655219, + 0.0035498146899044514, + 0.0043272594921290874, + 0.00256548379547894, + 0.0032187930773943663, + 0.006237422581762075, + 0.004439924843609333, + 0.006314518861472607, + 0.004204446449875832, + 0.004241478629410267, + 0.00568669568747282, + 0.004801478702574968, + 0.007521779276430607, + 0.00312165729701519, + 0.0023116066586226225, + 0.003952083643525839, + 0.002822653856128454, + 0.0024396248627454042, + 0.003374753287062049, + 0.004295114893466234, + 0.004397544544190168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.011073680594563484 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020569927990436554 + ], + "opposite_logprob_tokens": [ + -3.8955700397491455 + ], + "logprob_reward": -0.020569927990436554, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020569927990436554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6942892074584961, + 0.022618459537625313, + 0.013101260177791119, + 0.046489980071783066, + 0.02670367620885372, + 0.11590798199176788, + 0.06911769509315491, + 0.004342981614172459, + 0.0033246998209506273, + 0.004104101099073887 + ], + "credit_map": [ + 0.0563339963555336, + 0.04722856357693672, + 0.17734503746032715, + 0.04824069142341614, + 0.04719315841794014, + 0.1980503499507904, + 0.045108769088983536, + 0.0747886672616005, + 0.0021594928111881018, + 0.0017944829305633903, + 0.0032276867423206568, + 0.0021382616832852364, + 0.0018326916033402085, + 0.002910032169893384, + 0.002970941597595811, + 0.005584869999438524, + 0.0018175747245550156, + 0.001081795897334814, + 0.002608377020806074, + 0.0014913398772478104, + 0.0013154629850760102, + 0.0017471665050834417, + 0.0012170516420155764, + 0.001822490943595767, + 0.005421447101980448, + 0.004033226054161787, + 0.007570215035229921, + 0.006236545741558075, + 0.004959383513778448, + 0.006040776148438454, + 0.006904952228069305, + 0.005323432385921478, + 0.0024685519747436047, + 0.0026993488427251577, + 0.0021462086588144302, + 0.003691995982080698, + 0.00498473783954978, + 0.0030908596236258745, + 0.003391367383301258, + 0.004230604972690344, + 0.016245100647211075, + 0.007941107265651226, + 0.01683223806321621, + 0.009291432797908783, + 0.009053656831383705, + 0.015411993488669395, + 0.016345815733075142, + 0.024786632508039474, + 0.009634768590331078, + 0.005464539863169193, + 0.01146579161286354, + 0.00679919496178627, + 0.0065892660059034824, + 0.00980373565107584, + 0.009030339308083057, + 0.01033005677163601, + 0.0006387159810401499, + 0.0003967954544350505, + 0.0007321559241972864, + 0.000405962549848482, + 0.00050584995187819, + 0.0007369820377789438, + 0.0004224692238494754, + 0.0005040506366640329, + 0.0004055852477904409, + 0.00023708699154667556, + 0.0006657303893007338, + 0.00030323510873131454, + 0.0003916596178896725, + 0.0005699929897673428, + 0.00035914528416469693, + 0.0003922644245903939, + 0.0004763833712786436, + 0.0002865014539565891, + 0.0007091409061104059, + 0.0005044293939135969, + 0.0004862495989073068, + 0.0005764063680544496, + 0.0007358638686127961, + 0.0003291262546554208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.020569927990436554 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "soft-focus, medium-length, static shot captures a behind-the-scenes rehearsal moment, rich with drama and authenticity. At the center of the frame is an enigmatic actress, clad in an intricate, gothic gown of black velvet and lace, embellished with silver thread embroidery at the hem and cuffs. Her cascading raven hair is streaked with striking strands of silver, and her face is delicately painted with smoky makeup and crimson lipstick. She is in a dimly lit rehearsal room, scattered with prop boxes and folding chairs. Spotlights slant from overhead, casting dappled light on the wooden floor. Her hands tightly grasp a massive, seemingly weighty, but in reality lightweight, prop blade. Its broad, dark silver blade is engraved with intricate runes. With the focus and power of a performer, she raises the prop sword high above her head, then, with a dramatic gesture, slams it down to the floor below. At the moment of impact, a pre-set, harmless white smoke erupted from the point of contact between the blade and the ground, swirling and spreading in all directions under the spotlight. However, at the completion of this epic moment, her expression, brimming with power just a moment before, instantly collapsed. She straightened her back, leaning back slightly, and let out a light, hearty laugh, a playful glint in her eyes, creating a captivating contrast between serious performance and relaxed authenticity.", + "seed": 13355, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_13/index_4902_seed_13355_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027907208167016508, + "vlm_logprob": -0.027907208167016508, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.027907208167016508, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06321738660335541 + ], + "opposite_logprob_tokens": [ + -2.8132174015045166 + ], + "logprob_reward": -0.06321738660335541, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06321738660335541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21691717207431793, + 0.07962388545274734, + 0.028453057631850243, + 0.05067194998264313, + 0.0384204275906086, + 0.026853401213884354, + 0.07533478736877441, + 0.07173267751932144, + 0.2156902253627777, + 0.1963023990392685 + ], + "credit_map": [ + 0.020665127784013748, + 0.019485509023070335, + 0.03062167763710022, + 0.030189931392669678, + 0.03205113857984543, + 0.02948802337050438, + 0.030946196988224983, + 0.02346956357359886, + 0.0067426906898617744, + 0.015130625106394291, + 0.010841784067451954, + 0.011076338589191437, + 0.009053666144609451, + 0.010832292027771473, + 0.008125115185976028, + 0.007821371778845787, + 0.002686491934582591, + 0.0029214404057711363, + 0.0030129982624202967, + 0.0029831004794687033, + 0.00492850923910737, + 0.004603761248290539, + 0.004574473015964031, + 0.002742283744737506, + 0.004430149216204882, + 0.004666783846914768, + 0.0057259369641542435, + 0.005634639412164688, + 0.007127153687179089, + 0.007367642596364021, + 0.010942114517092705, + 0.004777528811246157, + 0.005971147213131189, + 0.005229099653661251, + 0.004640466533601284, + 0.00502638379111886, + 0.004997393116354942, + 0.005090100225061178, + 0.004038249608129263, + 0.003427589312195778, + 0.004595068749040365, + 0.002868193667382002, + 0.004135290160775185, + 0.002803145907819271, + 0.0030256076715886593, + 0.003674295963719487, + 0.002875896170735359, + 0.0028759005945175886, + 0.014990925788879395, + 0.01058722473680973, + 0.008934837765991688, + 0.008762559853494167, + 0.00831361673772335, + 0.009928006678819656, + 0.0069030942395329475, + 0.006914527155458927, + 0.0054359761998057365, + 0.008355584926903248, + 0.004575744271278381, + 0.014097508043050766, + 0.013163024559617043, + 0.010328302159905434, + 0.007452935911715031, + 0.008323593996465206, + 0.02129684016108513, + 0.017747683450579643, + 0.035481084138154984, + 0.023270489647984505, + 0.042600054293870926, + 0.03324072062969208, + 0.024540826678276062, + 0.017512531951069832, + 0.014879964292049408, + 0.026933828368782997, + 0.02814612351357937, + 0.04520870000123978, + 0.019223520532250404, + 0.030027583241462708, + 0.015222485177218914, + 0.01666020229458809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.06321738660335541 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008653995580971241 + ], + "opposite_logprob_tokens": [ + -4.7586541175842285 + ], + "logprob_reward": -0.008653995580971241, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008653995580971241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19522637128829956, + 0.09308730065822601, + 0.04289614409208298, + 0.051948726177215576, + 0.04618094488978386, + 0.045042116194963455, + 0.07400473207235336, + 0.04751747101545334, + 0.24242587387561798, + 0.16167038679122925 + ], + "credit_map": [ + 0.023600341752171516, + 0.020210569724440575, + 0.031482018530368805, + 0.02180485799908638, + 0.026065964251756668, + 0.02796701341867447, + 0.0209694504737854, + 0.023126164451241493, + 0.010122664272785187, + 0.013163748197257519, + 0.013323989696800709, + 0.008720871061086655, + 0.013646653853356838, + 0.014311417937278748, + 0.009259161539375782, + 0.010538789443671703, + 0.0037779491394758224, + 0.003962521441280842, + 0.005582119803875685, + 0.00490651186555624, + 0.006660805083811283, + 0.005787162110209465, + 0.007878726348280907, + 0.004340350162237883, + 0.0059888302348554134, + 0.004269060213118792, + 0.005875634029507637, + 0.0056653679348528385, + 0.0057563334703445435, + 0.007077380083501339, + 0.010560358874499798, + 0.006755761336535215, + 0.005677135195583105, + 0.006332423072308302, + 0.008571634069085121, + 0.0051139406859874725, + 0.005257215350866318, + 0.006236022803932428, + 0.00535832904279232, + 0.003634243505075574, + 0.007589911110699177, + 0.007378195878118277, + 0.00596781587228179, + 0.004472476430237293, + 0.004088179673999548, + 0.005949720274657011, + 0.005470971576869488, + 0.004124843515455723, + 0.014044099487364292, + 0.0094026830047369, + 0.01044967770576477, + 0.00827338732779026, + 0.00789306778460741, + 0.01046775933355093, + 0.0070314486511051655, + 0.006442609243094921, + 0.005761383567005396, + 0.0063879219815135, + 0.004377124831080437, + 0.00970142986625433, + 0.00665294611826539, + 0.006462575867772102, + 0.004414686467498541, + 0.0037594055756926537, + 0.023453760892152786, + 0.021510925143957138, + 0.04158481955528259, + 0.02491217479109764, + 0.04701172187924385, + 0.037231676280498505, + 0.026782775297760963, + 0.019938016310334206, + 0.012420274317264557, + 0.02118903025984764, + 0.02381720207631588, + 0.03392564505338669, + 0.01973402127623558, + 0.02373921126127243, + 0.013192485086619854, + 0.01365252397954464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.008653995580971241 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635037526488304 + ], + "opposite_logprob_tokens": [ + -3.651350259780884 + ], + "logprob_reward": -0.02635037526488304, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635037526488304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20423325896263123, + 0.12121438980102539, + 0.11286286264657974, + 0.05424893647432327, + 0.038957301527261734, + 0.04649769514799118, + 0.03771009296178818, + 0.02772635966539383, + 0.2030106782913208, + 0.15353846549987793 + ], + "credit_map": [ + 0.019797349348664284, + 0.02809552475810051, + 0.02540403977036476, + 0.029781797900795937, + 0.031363345682621, + 0.02675459533929825, + 0.021170295774936676, + 0.021866286173462868, + 0.01128340419381857, + 0.020845457911491394, + 0.01563425362110138, + 0.01818634755909443, + 0.011934148147702217, + 0.016731854528188705, + 0.014475563541054726, + 0.01212335005402565, + 0.007974869571626186, + 0.01398947648704052, + 0.010770421475172043, + 0.011844751425087452, + 0.0218140110373497, + 0.018067741766572, + 0.01838604360818863, + 0.01001552864909172, + 0.005633852910250425, + 0.00488359946757555, + 0.007327443920075893, + 0.006393878720700741, + 0.005904844496399164, + 0.006796195171773434, + 0.008729822933673859, + 0.008579296991229057, + 0.004364499822258949, + 0.004024344962090254, + 0.00512631144374609, + 0.0037692547775804996, + 0.004720216616988182, + 0.004795067012310028, + 0.00702674500644207, + 0.005130856763571501, + 0.007249502930790186, + 0.005141470581293106, + 0.005698082037270069, + 0.00481348717585206, + 0.004084150306880474, + 0.006062888540327549, + 0.007002393715083599, + 0.006445717997848988, + 0.0038370261900126934, + 0.006661548744887114, + 0.006623002700507641, + 0.00482401717454195, + 0.004235020838677883, + 0.004903105553239584, + 0.0030845922883599997, + 0.003541773883625865, + 0.002719190204516053, + 0.003760862397029996, + 0.004424296319484711, + 0.004041621927171946, + 0.0029045003466308117, + 0.004392337519675493, + 0.0030745642725378275, + 0.0024089862126857042, + 0.020831266418099403, + 0.017708703875541687, + 0.03562769666314125, + 0.02059859409928322, + 0.0376129224896431, + 0.03124159388244152, + 0.02359182946383953, + 0.015798043459653854, + 0.01556941494345665, + 0.018157847225666046, + 0.02522379346191883, + 0.031050505116581917, + 0.011860686354339123, + 0.0234691109508276, + 0.012686249800026417, + 0.015520832501351833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.02635037526488304 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762551836669445 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.00762551836669445, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762551836669445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09464104473590851, + 0.06580635160207748, + 0.11854710429906845, + 0.20486603677272797, + 0.09894134104251862, + 0.09205933660268784, + 0.07290039211511612, + 0.06334023177623749, + 0.15356974303722382, + 0.03532841056585312 + ], + "credit_map": [ + 0.009224402718245983, + 0.00850282795727253, + 0.016548333689570427, + 0.013350875116884708, + 0.010373016819357872, + 0.013394113630056381, + 0.010797777213156223, + 0.012449700385332108, + 0.007284223567694426, + 0.007471648510545492, + 0.01015938725322485, + 0.009862948209047318, + 0.008910063654184341, + 0.010556550696492195, + 0.005301019176840782, + 0.006260508205741644, + 0.008050919510424137, + 0.014686095528304577, + 0.009260190650820732, + 0.012833664193749428, + 0.02158219926059246, + 0.01789700798690319, + 0.025571951642632484, + 0.008665078319609165, + 0.02178497053682804, + 0.021066972985863686, + 0.021629489958286285, + 0.023474622517824173, + 0.026089439168572426, + 0.028492137789726257, + 0.0432015061378479, + 0.019126903265714645, + 0.007144103292375803, + 0.005675365682691336, + 0.014133743941783905, + 0.011815882287919521, + 0.011107560247182846, + 0.01374119147658348, + 0.02373475953936577, + 0.011588729918003082, + 0.009287253953516483, + 0.00663411570712924, + 0.014212356880307198, + 0.012680383399128914, + 0.009294298477470875, + 0.012863042764365673, + 0.015292877331376076, + 0.01179500762373209, + 0.00671200267970562, + 0.014845856465399265, + 0.007797290105372667, + 0.009912927635014057, + 0.008503464050590992, + 0.010659533552825451, + 0.009136983193457127, + 0.005332330707460642, + 0.008271449245512486, + 0.006013623904436827, + 0.006165971048176289, + 0.008156532421708107, + 0.012163572944700718, + 0.009164705872535706, + 0.0076666418462991714, + 0.0057377321645617485, + 0.014964660629630089, + 0.012570913881063461, + 0.021122891455888748, + 0.017072228714823723, + 0.0299052894115448, + 0.023214200511574745, + 0.02298024483025074, + 0.011739321053028107, + 0.004057799931615591, + 0.0035476733464747667, + 0.006267784163355827, + 0.006282702088356018, + 0.0034669479355216026, + 0.0053644669242203236, + 0.003352714003995061, + 0.0029883200768381357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.00762551836669445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an actress wearing a black velvet and lace gothic gown with silver thread embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the actress have cascading raven hair streaked with silver and wear smoky makeup with crimson lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the actress holding a massive prop sword with a broad dark silver blade engraved with intricate runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the actress raising the sword above her head and then slamming it down to the floor, creating white smoke at the point of impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a dimly lit rehearsal room with prop boxes, folding chairs, and overhead spotlights casting dappled light on a wooden floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368876501917839 + ], + "opposite_logprob_tokens": [ + -3.40868878364563 + ], + "logprob_reward": -0.03368876501917839, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368876501917839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29248493909835815, + 0.14197972416877747, + 0.08395618945360184, + 0.10313781350851059, + 0.05963835120201111, + 0.051667407155036926, + 0.05259327217936516, + 0.06861024349927902, + 0.11086837947368622, + 0.035063736140728 + ], + "credit_map": [ + 0.035766568034887314, + 0.026591073721647263, + 0.05226694047451019, + 0.02838657796382904, + 0.03647550940513611, + 0.03606538847088814, + 0.03118288516998291, + 0.04574999585747719, + 0.015907933935523033, + 0.009420812129974365, + 0.03339081257581711, + 0.015662724152207375, + 0.014648349024355412, + 0.02228597179055214, + 0.010034403763711452, + 0.02062871865928173, + 0.007870624773204327, + 0.010989688336849213, + 0.009643171913921833, + 0.007887512445449829, + 0.015588836744427681, + 0.014988848008215427, + 0.010179993696510792, + 0.006807505618780851, + 0.008859905414283276, + 0.010403456166386604, + 0.011713135056197643, + 0.008930174633860588, + 0.018746430054306984, + 0.016378013417124748, + 0.019303562119603157, + 0.008803124539554119, + 0.006500062998384237, + 0.0038104583509266376, + 0.011456665582954884, + 0.005374973174184561, + 0.006730958819389343, + 0.008694943971931934, + 0.006856303196400404, + 0.010213988833129406, + 0.003863609628751874, + 0.006272233556956053, + 0.0087753776460886, + 0.005644022487103939, + 0.005470775533467531, + 0.0068945265375077724, + 0.007083148695528507, + 0.00766371376812458, + 0.004184793680906296, + 0.010080881416797638, + 0.007579608354717493, + 0.0062491814605891705, + 0.0058140503242611885, + 0.007840395905077457, + 0.006254932843148708, + 0.004589428659528494, + 0.0063867708668112755, + 0.0061908625066280365, + 0.008406956680119038, + 0.008183380588889122, + 0.013674050569534302, + 0.011192869395017624, + 0.00816834345459938, + 0.006407008972018957, + 0.013887439854443073, + 0.009481211192905903, + 0.01579335145652294, + 0.012096349149942398, + 0.0197399090975523, + 0.01699816808104515, + 0.013047047890722752, + 0.009824897162616253, + 0.0047160666435956955, + 0.005098859779536724, + 0.004548411816358566, + 0.005580640397965908, + 0.003591952845454216, + 0.005393428262323141, + 0.003195331897586584, + 0.0029390454292297363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.03368876501917839 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_14/index_tensor([5030, 5030])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_14/index_tensor([5030, 5030])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..72efbd39cb7392211cf3a4d5c8a5f28a61632cec --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_14/index_tensor([5030, 5030])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angle close-up elegantly frames the carefully arranged composition on a polished, dark walnut coffee table in a sleek, modern living room. A minimalist white ceramic bowl with a smooth, low-gloss finish sits at its center. Filled with a vibrant assortment of fresh fruit—a crisp green apple, a ripe red pear, and a small bunch of deep purple grapes—their surfaces catch a subtle, healthy sheen from a nearby light source. Juxtaposed against the bowl is a slender black remote control, its matte plastic casing and minimalist design creating a striking geometric contrast to the bowl's organic form. The entire scene is bathed in warm ambient light from a large, out-of-focus window. The soft afternoon sunlight casts long, gentle shadows across the table's polished texture, highlighting the distinct grains of the ceramic, fruit, and wood. The ambiance is one of profound tranquility and refined elegance, emphasizing the quiet harmony and deliberate contrast between natural and technological objects.", + "seed": 24532, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_14/index_5030_seed_24532_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.024787430930882693, + "vlm_logprob": -0.024787430930882693, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.024787430930882693, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025802074000239372 + ], + "opposite_logprob_tokens": [ + -4.025802135467529 + ], + "logprob_reward": -0.025802074000239372, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.025802074000239372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16376467049121857, + 0.04658660292625427, + 0.06146343797445297, + 0.06318075954914093, + 0.31160029768943787, + 0.0802757516503334, + 0.10460656881332397, + 0.04763738438487053, + 0.03770897537469864, + 0.08317551761865616 + ], + "credit_map": [ + 0.024559777230024338, + 0.008795037865638733, + 0.07199597358703613, + 0.017609769478440285, + 0.006852044258266687, + 0.011320463381707668, + 0.011618419550359249, + 0.01101318933069706, + 0.005049932282418013, + 0.005887448787689209, + 0.012517588213086128, + 0.007909600622951984, + 0.002715843729674816, + 0.004084734711796045, + 0.004127261228859425, + 0.004294194281101227, + 0.007049775216728449, + 0.013020036742091179, + 0.009117011912167072, + 0.008723001927137375, + 0.0046870047226548195, + 0.004041146021336317, + 0.009897072799503803, + 0.004928385838866234, + 0.00577910291031003, + 0.011079131625592709, + 0.010669665411114693, + 0.008634363301098347, + 0.00840718299150467, + 0.007433530408889055, + 0.004892242141067982, + 0.006285537034273148, + 0.03923993930220604, + 0.05420900508761406, + 0.029179472476243973, + 0.037371840327978134, + 0.03979821130633354, + 0.041201867163181305, + 0.03054257668554783, + 0.04005740210413933, + 0.011972333304584026, + 0.007934634573757648, + 0.009346770122647285, + 0.009336087852716446, + 0.007466928102076054, + 0.009274117648601532, + 0.012505538761615753, + 0.012439340353012085, + 0.013928398489952087, + 0.014465329237282276, + 0.015115172602236271, + 0.01055799052119255, + 0.013135984539985657, + 0.012722386047244072, + 0.010264636017382145, + 0.014416665770113468, + 0.008817853406071663, + 0.009217272512614727, + 0.005450072232633829, + 0.007503673434257507, + 0.0033941874280571938, + 0.004105760715901852, + 0.004364027176052332, + 0.004784537013620138, + 0.004366759210824966, + 0.003954538144171238, + 0.005596720147877932, + 0.00802742037922144, + 0.003419333603233099, + 0.0030189407989382744, + 0.004687950946390629, + 0.004637314472347498, + 0.010549362748861313, + 0.013982296921312809, + 0.010030457749962807, + 0.012131924740970135, + 0.00822721142321825, + 0.013500426895916462, + 0.007436901796609163, + 0.007316930685192347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.025802074000239372 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05501003563404083 + ], + "opposite_logprob_tokens": [ + -2.9300100803375244 + ], + "logprob_reward": -0.05501003563404083, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05501003563404083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13670624792575836, + 0.2409321516752243, + 0.09917957335710526, + 0.09308108687400818, + 0.20148825645446777, + 0.05173231288790703, + 0.050658371299505234, + 0.04208660498261452, + 0.03182772174477577, + 0.052307602018117905 + ], + "credit_map": [ + 0.013914165087044239, + 0.020864680409431458, + 0.019248249009251595, + 0.019371606409549713, + 0.012675256468355656, + 0.015948770567774773, + 0.01630268804728985, + 0.018380844965577126, + 0.03296627104282379, + 0.038747161626815796, + 0.05209356173872948, + 0.04385797679424286, + 0.0169731006026268, + 0.015681788325309753, + 0.02592417038977146, + 0.014688136987388134, + 0.010279323905706406, + 0.015434653498232365, + 0.019651660695672035, + 0.014001462608575821, + 0.009410647675395012, + 0.007082443218678236, + 0.011395040899515152, + 0.011924347840249538, + 0.007619867101311684, + 0.013145592994987965, + 0.014931702986359596, + 0.008961958810687065, + 0.012171484529972076, + 0.009761583991348743, + 0.017769169062376022, + 0.00871974602341652, + 0.02621592953801155, + 0.03794505074620247, + 0.01864001154899597, + 0.026798242703080177, + 0.024676712229847908, + 0.024183690547943115, + 0.018182992935180664, + 0.02484564296901226, + 0.005600349511951208, + 0.005262155085802078, + 0.0065725380554795265, + 0.008086144924163818, + 0.004627932328730822, + 0.005231446586549282, + 0.008376582525670528, + 0.007975166663527489, + 0.00686582550406456, + 0.007066697347909212, + 0.006786067970097065, + 0.006188723258674145, + 0.005473309196531773, + 0.00532765407115221, + 0.004344663582742214, + 0.00860543455928564, + 0.009156272746622562, + 0.008443436585366726, + 0.004216701723635197, + 0.006097051780670881, + 0.0034242072142660618, + 0.0030275119934231043, + 0.0035133075434714556, + 0.00420812051743269, + 0.006876314990222454, + 0.005622128956019878, + 0.0032079732045531273, + 0.003726804628968239, + 0.003229094436392188, + 0.0023703500628471375, + 0.004305610433220863, + 0.0024894466623663902, + 0.009129268117249012, + 0.008619856089353561, + 0.006190951447933912, + 0.006912906188517809, + 0.005358568392693996, + 0.004894706886261702, + 0.005139622837305069, + 0.00606172438710928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.05501003563404083 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674353027716279 + ], + "opposite_logprob_tokens": [ + -5.006743431091309 + ], + "logprob_reward": -0.00674353027716279, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674353027716279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2685506045818329, + 0.07717285305261612, + 0.11557445675134659, + 0.08056288212537766, + 0.21738265454769135, + 0.027336295694112778, + 0.044997166842222214, + 0.07202660292387009, + 0.037092071026563644, + 0.05930442363023758 + ], + "credit_map": [ + 0.03196854144334793, + 0.025521252304315567, + 0.11234040558338165, + 0.024991275742650032, + 0.01818936876952648, + 0.018517380580306053, + 0.019205553457140923, + 0.01781681925058365, + 0.009561165235936642, + 0.01838625594973564, + 0.013913842849433422, + 0.010924136266112328, + 0.005558455362915993, + 0.0071958997286856174, + 0.00717517826706171, + 0.004457920324057341, + 0.0134987598285079, + 0.028110485523939133, + 0.0170558150857687, + 0.015295060351490974, + 0.008938325569033623, + 0.006535997148603201, + 0.016332073137164116, + 0.009807948023080826, + 0.005496894475072622, + 0.010780255310237408, + 0.011377991177141666, + 0.011015391908586025, + 0.011813843622803688, + 0.013561153784394264, + 0.009748372249305248, + 0.006768983788788319, + 0.027167649939656258, + 0.04108470678329468, + 0.019863834604620934, + 0.025294093415141106, + 0.027070825919508934, + 0.027455778792500496, + 0.02171383425593376, + 0.027731923386454582, + 0.003833518363535404, + 0.0026247429195791483, + 0.0027970904484391212, + 0.0035654373932629824, + 0.005188987124711275, + 0.003712600562721491, + 0.0031602000817656517, + 0.002453716704621911, + 0.013255921192467213, + 0.009237970225512981, + 0.004115432966500521, + 0.003402753034606576, + 0.003054621396586299, + 0.003231791313737631, + 0.004422732163220644, + 0.00427594780921936, + 0.016300542280077934, + 0.014864646829664707, + 0.005896763876080513, + 0.008466736413538456, + 0.008772495202720165, + 0.00512072816491127, + 0.00629122368991375, + 0.006313467863947153, + 0.004396334756165743, + 0.005126040428876877, + 0.004905973561108112, + 0.0059589846059679985, + 0.004517876077443361, + 0.004137043375521898, + 0.004180729389190674, + 0.0038690869696438313, + 0.008870427496731281, + 0.006941759027540684, + 0.009597370401024818, + 0.006253058556467295, + 0.008545847609639168, + 0.006587222218513489, + 0.006500109564512968, + 0.006008627358824015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.00674353027716279 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197370693087578 + ], + "opposite_logprob_tokens": [ + -4.018197536468506 + ], + "logprob_reward": -0.018197370693087578, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197370693087578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14389362931251526, + 0.07823343575000763, + 0.16478653252124786, + 0.05847824364900589, + 0.23027856647968292, + 0.03402000665664673, + 0.03657195344567299, + 0.09015186131000519, + 0.08334717154502869, + 0.08023864030838013 + ], + "credit_map": [ + 0.015007163397967815, + 0.019337452948093414, + 0.025854917243123055, + 0.016150062903761864, + 0.0204313974827528, + 0.017285320907831192, + 0.016810555011034012, + 0.013016759417951107, + 0.008725493215024471, + 0.008308695629239082, + 0.012304478324949741, + 0.009331694804131985, + 0.010375803336501122, + 0.015076116658747196, + 0.008880785666406155, + 0.005230363924056292, + 0.017833979800343513, + 0.04128998517990112, + 0.02618754655122757, + 0.01842189021408558, + 0.0124172642827034, + 0.008615713566541672, + 0.026699332520365715, + 0.013320820406079292, + 0.0054358551278710365, + 0.008673099800944328, + 0.008150236681103706, + 0.007480394560843706, + 0.008530935272574425, + 0.006872605998069048, + 0.006609616335481405, + 0.0067254966124892235, + 0.025433171540498734, + 0.04113416373729706, + 0.02021258883178234, + 0.021669114008545876, + 0.04352714866399765, + 0.023385128006339073, + 0.023583661764860153, + 0.03133358433842659, + 0.007528504356741905, + 0.003944964613765478, + 0.003470035968348384, + 0.005057341884821653, + 0.0029741304460912943, + 0.004315539263188839, + 0.0036364870611578226, + 0.0030930028297007084, + 0.007181371562182903, + 0.00562131917104125, + 0.0032929605804383755, + 0.004873831290751696, + 0.0034634224139153957, + 0.004948993679136038, + 0.0030719598289579153, + 0.004118091892451048, + 0.014331251382827759, + 0.017485305666923523, + 0.009610633365809917, + 0.013308417052030563, + 0.010717951692640781, + 0.007783983834087849, + 0.008755014277994633, + 0.008159306831657887, + 0.011910418048501015, + 0.011531243100762367, + 0.011199248023331165, + 0.015370508655905724, + 0.008262783288955688, + 0.008345475420355797, + 0.009115708991885185, + 0.007611781358718872, + 0.008934476412832737, + 0.010111180134117603, + 0.009510939009487629, + 0.005455384962260723, + 0.013739638030529022, + 0.012367616407573223, + 0.011464796029031277, + 0.008654611185193062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.018197370693087578 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818414404988289 + ], + "opposite_logprob_tokens": [ + -4.018184185028076 + ], + "logprob_reward": -0.01818414404988289, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818414404988289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07502657920122147, + 0.02777145616710186, + 0.08733578771352768, + 0.055587831884622574, + 0.18224208056926727, + 0.1347130686044693, + 0.08397818356752396, + 0.11621487140655518, + 0.11531168967485428, + 0.12181846052408218 + ], + "credit_map": [ + 0.006172636523842812, + 0.011067057959735394, + 0.01937849260866642, + 0.008415605872869492, + 0.007625967729836702, + 0.00774847948923707, + 0.006655025761574507, + 0.007963317446410656, + 0.0023363945074379444, + 0.004420214798301458, + 0.006213681772351265, + 0.004024590831249952, + 0.0031261539552360773, + 0.0020796400494873524, + 0.002115227049216628, + 0.0034555522724986076, + 0.010081704705953598, + 0.018823668360710144, + 0.013928202912211418, + 0.011961880140006542, + 0.0064751203171908855, + 0.004412477370351553, + 0.015287305228412151, + 0.006365424022078514, + 0.0049905371852219105, + 0.006178641226142645, + 0.012177095748484135, + 0.006489343009889126, + 0.0067335739731788635, + 0.005053071305155754, + 0.006327411159873009, + 0.0076381610706448555, + 0.02239076979458332, + 0.031518854200839996, + 0.01750214956700802, + 0.026116156950592995, + 0.023078052327036858, + 0.020186245441436768, + 0.018290922045707703, + 0.023158930242061615, + 0.012285346165299416, + 0.015557986684143543, + 0.01984972134232521, + 0.025539813563227654, + 0.00949862226843834, + 0.010795154608786106, + 0.022621935233473778, + 0.018564488738775253, + 0.018155433237552643, + 0.014644134789705276, + 0.008873320184648037, + 0.009739936329424381, + 0.005263368133455515, + 0.007459006737917662, + 0.00974312610924244, + 0.01009985338896513, + 0.018456226214766502, + 0.01832485757768154, + 0.01606614887714386, + 0.02320406772196293, + 0.009498042985796928, + 0.008616326376795769, + 0.009163463488221169, + 0.012885737232863903, + 0.026210147887468338, + 0.023820936679840088, + 0.011268086731433868, + 0.021654238924384117, + 0.006350175943225622, + 0.00782344862818718, + 0.008708356879651546, + 0.009476304054260254, + 0.015685997903347015, + 0.011545215733349323, + 0.029441120103001595, + 0.013397588394582272, + 0.01028171181678772, + 0.014572016894817352, + 0.010886387899518013, + 0.016008419916033745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.01818414404988289 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle close-up elegantly frames the carefully arranged composition on a polished, dark walnut coffee table in a sleek, modern living room. A minimalist white ceramic bowl with a smooth, low-gloss finish sits at its center. Filled with a vibrant assortment of fresh fruit—a crisp green apple, a ripe red pear, and a small bunch of deep purple grapes—their surfaces catch a subtle, healthy sheen from a nearby light source. Juxtaposed against the bowl is a slender black remote control, its matte plastic casing and minimalist design creating a striking geometric contrast to the bowl's organic form. The entire scene is bathed in warm ambient light from a large, out-of-focus window. The soft afternoon sunlight casts long, gentle shadows across the table's polished texture, highlighting the distinct grains of the ceramic, fruit, and wood. The ambiance is one of profound tranquility and refined elegance, emphasizing the quiet harmony and deliberate contrast between natural and technological objects.", + "seed": 24533, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_14/index_5030_seed_24533_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04495456982403993, + "vlm_logprob": -0.04495456982403993, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.04495456982403993, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01570194400846958 + ], + "opposite_logprob_tokens": [ + -4.515701770782471 + ], + "logprob_reward": -0.01570194400846958, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01570194400846958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10058076679706573, + 0.05885698273777962, + 0.10042424499988556, + 0.04243602976202965, + 0.1948850303888321, + 0.3404172956943512, + 0.086585633456707, + 0.02175953797996044, + 0.02817891165614128, + 0.02587556280195713 + ], + "credit_map": [ + 0.008715360425412655, + 0.015311812050640583, + 0.010585637763142586, + 0.014858727343380451, + 0.012916501611471176, + 0.014063000679016113, + 0.009970251470804214, + 0.014159483835101128, + 0.004217550158500671, + 0.009778411127626896, + 0.0050117261707782745, + 0.005930595565587282, + 0.009958681650459766, + 0.005119659472256899, + 0.006710108369588852, + 0.012130250222980976, + 0.009043487720191479, + 0.009825221262872219, + 0.019689157605171204, + 0.012321477755904198, + 0.01070589292794466, + 0.023225795477628708, + 0.0064686876721680164, + 0.009144526906311512, + 0.004402395337820053, + 0.005201568827033043, + 0.004728829953819513, + 0.004967731889337301, + 0.005521692335605621, + 0.004941163584589958, + 0.0030081854201853275, + 0.009664461016654968, + 0.019575268030166626, + 0.02374952659010887, + 0.017862757667899132, + 0.023147284984588623, + 0.02342163771390915, + 0.021424762904644012, + 0.016721991822123528, + 0.04898180440068245, + 0.03350738063454628, + 0.0404551662504673, + 0.035231344401836395, + 0.0419275164604187, + 0.04460251331329346, + 0.05031152442097664, + 0.030885841697454453, + 0.06349601596593857, + 0.007238082587718964, + 0.00595729099586606, + 0.007320004981011152, + 0.017851628363132477, + 0.005562206264585257, + 0.018506484106183052, + 0.00931380782276392, + 0.014836132526397705, + 0.0018429248593747616, + 0.003262278623878956, + 0.0018581748008728027, + 0.0024419701658189297, + 0.003040520939975977, + 0.002491189632564783, + 0.002021185588091612, + 0.004801291506737471, + 0.0022189959418028593, + 0.003396181622520089, + 0.0036047932226210833, + 0.003220207989215851, + 0.0024471492506563663, + 0.0032919517252594233, + 0.0044694989919662476, + 0.005530133843421936, + 0.0019920438062399626, + 0.003456653794273734, + 0.0019182737451046705, + 0.0025119357742369175, + 0.003133145859465003, + 0.003128767479211092, + 0.004446687642484903, + 0.005288055632263422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.01570194400846958 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604028344154358 + ], + "opposite_logprob_tokens": [ + -1.910402774810791 + ], + "logprob_reward": -0.1604028344154358, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604028344154358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1826431304216385, + 0.09428498893976212, + 0.17210249602794647, + 0.059387121349573135, + 0.11944533884525299, + 0.040214020758867264, + 0.18369784951210022, + 0.07301255315542221, + 0.027433212846517563, + 0.04777929559350014 + ], + "credit_map": [ + 0.025798846036195755, + 0.019253622740507126, + 0.018044155091047287, + 0.02406727336347103, + 0.020537646487355232, + 0.01759161427617073, + 0.021791581064462662, + 0.03555838018655777, + 0.006942881736904383, + 0.007503029424697161, + 0.0171470046043396, + 0.009246735833585262, + 0.010498649440705776, + 0.01734405755996704, + 0.00865929014980793, + 0.016943344846367836, + 0.014819476753473282, + 0.0128053342923522, + 0.03085077740252018, + 0.023710433393716812, + 0.01753530278801918, + 0.046351972967386246, + 0.011012102477252483, + 0.015017088502645493, + 0.006488492246717215, + 0.009176784195005894, + 0.006139661651104689, + 0.007503824308514595, + 0.008590721525251865, + 0.0072324927896261215, + 0.004693894647061825, + 0.009561250917613506, + 0.013840962201356888, + 0.012337024323642254, + 0.012370424345135689, + 0.014562048017978668, + 0.013381010852754116, + 0.01338651031255722, + 0.010492228902876377, + 0.029075125232338905, + 0.004841825924813747, + 0.004933848045766354, + 0.004768391139805317, + 0.0045874291099607944, + 0.002923039486631751, + 0.004036611411720514, + 0.003991994541138411, + 0.010130882263183594, + 0.014584915712475777, + 0.0160077977925539, + 0.01397155225276947, + 0.03504687920212746, + 0.015812687575817108, + 0.04004140570759773, + 0.018374141305685043, + 0.029858464375138283, + 0.005422822665423155, + 0.00777975469827652, + 0.007834293879568577, + 0.007384801283478737, + 0.007891485467553139, + 0.009365251287817955, + 0.007659518159925938, + 0.019674621522426605, + 0.0024958003778010607, + 0.003097767010331154, + 0.002514344872906804, + 0.0026780974585562944, + 0.003957478329539299, + 0.003165756119415164, + 0.0031103347428143024, + 0.00641363300383091, + 0.004126823041588068, + 0.00432773819193244, + 0.005358902737498283, + 0.004339512437582016, + 0.003962232731282711, + 0.005351230967789888, + 0.008376363664865494, + 0.01193649135529995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.1604028344154358 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784595109522343 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.009784595109522343, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784595109522343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16315098106861115, + 0.10361826419830322, + 0.202927365899086, + 0.03450610861182213, + 0.05701470747590065, + 0.09933768957853317, + 0.19506274163722992, + 0.0736033245921135, + 0.03726191818714142, + 0.03351687639951706 + ], + "credit_map": [ + 0.015357553027570248, + 0.021904243156313896, + 0.016689499840140343, + 0.02300199866294861, + 0.022012321278452873, + 0.020825155079364777, + 0.01790737733244896, + 0.02545285038650036, + 0.015421089716255665, + 0.014983118511736393, + 0.013482285663485527, + 0.013215246610343456, + 0.012330976314842701, + 0.011268388479948044, + 0.011062182486057281, + 0.011854978278279305, + 0.013568640686571598, + 0.014062386006116867, + 0.038680385798215866, + 0.03142767772078514, + 0.020400073379278183, + 0.05488849803805351, + 0.013363497331738472, + 0.01653621904551983, + 0.0036916574463248253, + 0.004884449765086174, + 0.003841355675831437, + 0.0045189689844846725, + 0.0062653981149196625, + 0.00337812677025795, + 0.0030563753098249435, + 0.004869779106229544, + 0.007403427269309759, + 0.007628308143466711, + 0.006643233820796013, + 0.0058275191113352776, + 0.006941838655620813, + 0.005223001353442669, + 0.004498693626374006, + 0.0128486854955554, + 0.011059910990297794, + 0.011123137548565865, + 0.010362781584262848, + 0.011536970734596252, + 0.012356346473097801, + 0.013631774112582207, + 0.008525459095835686, + 0.020741309970617294, + 0.017138978466391563, + 0.01808832213282585, + 0.016607796773314476, + 0.033308275043964386, + 0.021745674312114716, + 0.03095078468322754, + 0.023044167086482048, + 0.034178752452135086, + 0.007493747863918543, + 0.010021594353020191, + 0.007777794264256954, + 0.008087649941444397, + 0.008347389288246632, + 0.00872734934091568, + 0.008157837204635143, + 0.014989964663982391, + 0.003892273176461458, + 0.004129056353121996, + 0.005362027324736118, + 0.00389108806848526, + 0.0038505487609654665, + 0.0049049281515181065, + 0.0036911554634571075, + 0.007540838792920113, + 0.0033361692912876606, + 0.004575233440846205, + 0.0036274592857807875, + 0.003124695271253586, + 0.004647668916732073, + 0.003631109371781349, + 0.0050010946579277515, + 0.005573447328060865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.009784595109522343 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012548891827464104 + ], + "opposite_logprob_tokens": [ + -4.387548923492432 + ], + "logprob_reward": -0.012548891827464104, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012548891827464104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31536707282066345, + 0.13905327022075653, + 0.09766578674316406, + 0.045985691249370575, + 0.05246630311012268, + 0.05958927795290947, + 0.11544988304376602, + 0.07823978364467621, + 0.05602021887898445, + 0.04016277194023132 + ], + "credit_map": [ + 0.04281460493803024, + 0.04308193176984787, + 0.033077601343393326, + 0.030026128515601158, + 0.060702018439769745, + 0.02462439425289631, + 0.03664182126522064, + 0.04439856857061386, + 0.013976619578897953, + 0.018252622336149216, + 0.01766606606543064, + 0.01781880110502243, + 0.017220204696059227, + 0.01644068956375122, + 0.013867175206542015, + 0.02381109818816185, + 0.007534627802670002, + 0.01298951543867588, + 0.013301340863108635, + 0.008644616231322289, + 0.015821896493434906, + 0.01130619179457426, + 0.008441844955086708, + 0.019625753164291382, + 0.004732260014861822, + 0.010520937852561474, + 0.0036855374928563833, + 0.00501627242192626, + 0.007394692394882441, + 0.0049722399562597275, + 0.0033182958140969276, + 0.006345453206449747, + 0.006923061795532703, + 0.0069928476586937904, + 0.0046415990218520164, + 0.005226398352533579, + 0.007421492133289576, + 0.006029749289155006, + 0.004629767499864101, + 0.010601392015814781, + 0.005624701734632254, + 0.008591493591666222, + 0.0061092800460755825, + 0.006958685349673033, + 0.00815233401954174, + 0.00913703441619873, + 0.003922040108591318, + 0.011093709617853165, + 0.007804626598954201, + 0.012182287871837616, + 0.010320043191313744, + 0.018934035673737526, + 0.017299938946962357, + 0.01950310543179512, + 0.011190757155418396, + 0.01821509376168251, + 0.007257912773638964, + 0.012140409089624882, + 0.009413897059857845, + 0.00883923564106226, + 0.008170885033905506, + 0.012784820050001144, + 0.006747618783265352, + 0.01288500614464283, + 0.006077948492020369, + 0.006704726722091436, + 0.005183613859117031, + 0.0052168238908052444, + 0.009082688018679619, + 0.005988651420921087, + 0.005718250293284655, + 0.01204751431941986, + 0.002964183920994401, + 0.005014973226934671, + 0.003390897298231721, + 0.004186836536973715, + 0.0051300073973834515, + 0.004895452409982681, + 0.007829692214727402, + 0.006750728935003281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.012548891827464104 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633458375930786 + ], + "opposite_logprob_tokens": [ + -3.651334524154663 + ], + "logprob_reward": -0.02633458375930786, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633458375930786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11177440732717514, + 0.05417211353778839, + 0.09424906224012375, + 0.034696854650974274, + 0.03400252014398575, + 0.05558476224541664, + 0.08703693002462387, + 0.14076152443885803, + 0.09428247809410095, + 0.2934392988681793 + ], + "credit_map": [ + 0.011647230014204979, + 0.010158070363104343, + 0.009595691226422787, + 0.01737108826637268, + 0.0165112167596817, + 0.016119223088026047, + 0.013226998038589954, + 0.017144886776804924, + 0.0037783803418278694, + 0.006187933962792158, + 0.006923256907612085, + 0.005638317670673132, + 0.007900022901594639, + 0.0064935446716845036, + 0.005543893668800592, + 0.0117067601531744, + 0.005953236483037472, + 0.005885821767151356, + 0.019334128126502037, + 0.01146525889635086, + 0.00935408752411604, + 0.02231517806649208, + 0.005495352670550346, + 0.014446001499891281, + 0.0034301162231713533, + 0.0033323767129331827, + 0.0033516669645905495, + 0.003968196455389261, + 0.005412504076957703, + 0.0044282288290560246, + 0.003566810628399253, + 0.007206957787275314, + 0.003459557890892029, + 0.0055734533816576, + 0.0034700341057032347, + 0.0037905233912169933, + 0.0035461715888231993, + 0.003456389531493187, + 0.003065780969336629, + 0.007640612777322531, + 0.0051911273039877415, + 0.006193904206156731, + 0.00510279880836606, + 0.007516553159803152, + 0.007628786843270063, + 0.007867823354899883, + 0.004955810494720936, + 0.011127961799502373, + 0.0060893152840435505, + 0.007614521775394678, + 0.006548681762069464, + 0.014721137471497059, + 0.011579749174416065, + 0.014944052323698997, + 0.008473207242786884, + 0.017066262662410736, + 0.012476994656026363, + 0.01097356341779232, + 0.013614410534501076, + 0.018047692254185677, + 0.011955030262470245, + 0.025627071037888527, + 0.01513033825904131, + 0.032936420291662216, + 0.005993506405502558, + 0.008775397203862667, + 0.008422327227890491, + 0.010569090954959393, + 0.010187707841396332, + 0.012701890431344509, + 0.014630972407758236, + 0.02300158701837063, + 0.020495392382144928, + 0.017877141013741493, + 0.02047669142484665, + 0.029907943680882454, + 0.025798242539167404, + 0.03794626146554947, + 0.05710984393954277, + 0.08382781594991684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.02633458375930786 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle close-up elegantly frames the carefully arranged composition on a polished, dark walnut coffee table in a sleek, modern living room. A minimalist white ceramic bowl with a smooth, low-gloss finish sits at its center. Filled with a vibrant assortment of fresh fruit—a crisp green apple, a ripe red pear, and a small bunch of deep purple grapes—their surfaces catch a subtle, healthy sheen from a nearby light source. Juxtaposed against the bowl is a slender black remote control, its matte plastic casing and minimalist design creating a striking geometric contrast to the bowl's organic form. The entire scene is bathed in warm ambient light from a large, out-of-focus window. The soft afternoon sunlight casts long, gentle shadows across the table's polished texture, highlighting the distinct grains of the ceramic, fruit, and wood. The ambiance is one of profound tranquility and refined elegance, emphasizing the quiet harmony and deliberate contrast between natural and technological objects.", + "seed": 24534, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_14/index_5030_seed_24534_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02209211220033467, + "vlm_logprob": -0.02209211220033467, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.02209211220033467, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012869163416326046 + ], + "opposite_logprob_tokens": [ + -4.887869358062744 + ], + "logprob_reward": -0.012869163416326046, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012869163416326046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06957170367240906, + 0.03415881469845772, + 0.04062189534306526, + 0.016855964437127113, + 0.03312275931239128, + 0.2557542026042938, + 0.04674627259373665, + 0.1393813043832779, + 0.2691144347190857, + 0.09467262774705887 + ], + "credit_map": [ + 0.007842883467674255, + 0.006151617970317602, + 0.007364584133028984, + 0.008320050314068794, + 0.014833183027803898, + 0.010736980475485325, + 0.0065940734930336475, + 0.007728339172899723, + 0.003313553286716342, + 0.002745467470958829, + 0.0029626176692545414, + 0.002989083994179964, + 0.010920656844973564, + 0.0038807301316410303, + 0.0035382655914872885, + 0.0038084457628428936, + 0.003434225916862488, + 0.00457068020477891, + 0.0036790035665035248, + 0.006650232244282961, + 0.011204883456230164, + 0.004681526683270931, + 0.0032704321201890707, + 0.0031309123151004314, + 0.0020452223252505064, + 0.0013491635909304023, + 0.0026712703984230757, + 0.0014793687732890248, + 0.0036408600863069296, + 0.0019580679945647717, + 0.0015471018850803375, + 0.002164913108572364, + 0.004643457010388374, + 0.004286784678697586, + 0.002882118569687009, + 0.0024353028275072575, + 0.005053399130702019, + 0.006420644000172615, + 0.0037815896794199944, + 0.003619469003751874, + 0.05230589583516121, + 0.012570198625326157, + 0.01029741857200861, + 0.00630186079069972, + 0.011926926672458649, + 0.1010073646903038, + 0.024506013840436935, + 0.03683853521943092, + 0.007123867981135845, + 0.00424312986433506, + 0.00448985630646348, + 0.003749560797587037, + 0.006642979569733143, + 0.01077867578715086, + 0.004204132594168186, + 0.005514075979590416, + 0.011944098398089409, + 0.02680066227912903, + 0.015059344470500946, + 0.009354479610919952, + 0.03362703323364258, + 0.0124284652993083, + 0.011600461788475513, + 0.01856677047908306, + 0.02424726076424122, + 0.04732901602983475, + 0.0285171065479517, + 0.020302969962358475, + 0.06385576725006104, + 0.022874217480421066, + 0.022206271067261696, + 0.0397818349301815, + 0.007973862811923027, + 0.012728315778076649, + 0.010432020761072636, + 0.010804462246596813, + 0.020773235708475113, + 0.01133306510746479, + 0.010249926708638668, + 0.010377753525972366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.012869163416326046 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043101079761981964 + ], + "opposite_logprob_tokens": [ + -3.1681010723114014 + ], + "logprob_reward": -0.043101079761981964, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043101079761981964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0745031014084816, + 0.04543717950582504, + 0.06234593689441681, + 0.019596394151449203, + 0.03918874263763428, + 0.26125192642211914, + 0.050434332340955734, + 0.06337378919124603, + 0.07540502399206161, + 0.30846360325813293 + ], + "credit_map": [ + 0.007856853306293488, + 0.008341806940734386, + 0.007829765789210796, + 0.01239794585853815, + 0.009836978279054165, + 0.009640861302614212, + 0.010068736039102077, + 0.008530152961611748, + 0.004293587990105152, + 0.004433819092810154, + 0.005731265991926193, + 0.00801972858607769, + 0.005606351885944605, + 0.007097968831658363, + 0.005361273884773254, + 0.004893181379884481, + 0.00603818753734231, + 0.0052604880183935165, + 0.005469856783747673, + 0.006599176209419966, + 0.0189808439463377, + 0.006517164874821901, + 0.006074147764593363, + 0.007406074088066816, + 0.0021380148828029633, + 0.002408497966825962, + 0.0024782512336969376, + 0.003685316536575556, + 0.002253606216982007, + 0.0018526682397350669, + 0.0019219628302380443, + 0.0028580771759152412, + 0.005715051665902138, + 0.004342744126915932, + 0.005403740331530571, + 0.005252679344266653, + 0.0057748774997889996, + 0.003981271758675575, + 0.0047606248408555984, + 0.003957753535360098, + 0.04394436627626419, + 0.01840130239725113, + 0.02576150931417942, + 0.013829751871526241, + 0.03333117812871933, + 0.06694377213716507, + 0.02971031703054905, + 0.029329704120755196, + 0.00625810120254755, + 0.004459558054804802, + 0.003828504355624318, + 0.00350816803984344, + 0.01126585341989994, + 0.01158497016876936, + 0.0042829494923353195, + 0.005246229935437441, + 0.00510195130482316, + 0.009697523899376392, + 0.006941418629139662, + 0.005247420631349087, + 0.015958193689584732, + 0.005987785756587982, + 0.006868682336062193, + 0.007570816203951836, + 0.006975050549954176, + 0.0075098611414432526, + 0.009221984073519707, + 0.012353857979178429, + 0.012396025471389294, + 0.009976034052670002, + 0.008256948553025723, + 0.008715261705219746, + 0.02560090273618698, + 0.04635856673121452, + 0.04873991757631302, + 0.02688729204237461, + 0.04177995026111603, + 0.03676865994930267, + 0.03074347786605358, + 0.05158481374382973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.043101079761981964 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200207905843854 + ], + "opposite_logprob_tokens": [ + -6.127200126647949 + ], + "logprob_reward": -0.002200207905843854, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200207905843854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12157700210809708, + 0.08431407064199448, + 0.026471871882677078, + 0.036919791251420975, + 0.056144095957279205, + 0.22773073613643646, + 0.09408946335315704, + 0.10167703032493591, + 0.13572506606578827, + 0.1153508648276329 + ], + "credit_map": [ + 0.011532362550497055, + 0.011794555000960827, + 0.00943784136325121, + 0.015803374350070953, + 0.02490796521306038, + 0.01884055882692337, + 0.016758576035499573, + 0.012501779943704605, + 0.008023278787732124, + 0.007087447214871645, + 0.007095703389495611, + 0.010629817843437195, + 0.016840066760778427, + 0.014052401296794415, + 0.010885161347687244, + 0.009700198657810688, + 0.002270139055326581, + 0.0023119233082979918, + 0.0026830739807337523, + 0.0029137241654098034, + 0.004448214080184698, + 0.0048036593943834305, + 0.0034669970627874136, + 0.0035741382744163275, + 0.004034529905766249, + 0.003338929731398821, + 0.00349055347032845, + 0.0038908931892365217, + 0.010913122445344925, + 0.003413224359974265, + 0.0037306821905076504, + 0.0041078547947108746, + 0.006793092470616102, + 0.006215180270373821, + 0.00418242858722806, + 0.005791085306555033, + 0.008925859816372395, + 0.007306204177439213, + 0.01113688200712204, + 0.005793359130620956, + 0.04476862773299217, + 0.011691983789205551, + 0.012014959938824177, + 0.009796461090445518, + 0.014402339234948158, + 0.08199473470449448, + 0.022291485220193863, + 0.030770165845751762, + 0.015779437497258186, + 0.008738464675843716, + 0.006707627326250076, + 0.005621503107249737, + 0.014885926619172096, + 0.024612534791231155, + 0.007084307726472616, + 0.010659671388566494, + 0.007989718578755856, + 0.011166094802320004, + 0.013043103739619255, + 0.01454854290932417, + 0.022284027189016342, + 0.008883248083293438, + 0.007947146892547607, + 0.01581514999270439, + 0.012419315055012703, + 0.025495901703834534, + 0.015465592965483665, + 0.010773117654025555, + 0.031125731766223907, + 0.010743143036961555, + 0.01093993429094553, + 0.01876233145594597, + 0.010604389943182468, + 0.014461856335401535, + 0.012196444906294346, + 0.011087506078183651, + 0.02620229311287403, + 0.014283505268394947, + 0.013209429569542408, + 0.01330544613301754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.002200207905843854 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420164667069912 + ], + "opposite_logprob_tokens": [ + -4.264201641082764 + ], + "logprob_reward": -0.01420164667069912, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420164667069912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1428861916065216, + 0.060921333730220795, + 0.048677943646907806, + 0.03986011818051338, + 0.032630763947963715, + 0.06884327530860901, + 0.07380877435207367, + 0.21172723174095154, + 0.13229221105575562, + 0.18835213780403137 + ], + "credit_map": [ + 0.01438686903566122, + 0.01420365646481514, + 0.008700964972376823, + 0.01364603079855442, + 0.04222891479730606, + 0.019882824271917343, + 0.01588769070804119, + 0.013949246145784855, + 0.005259298253804445, + 0.006405178923159838, + 0.006074034608900547, + 0.00803246907889843, + 0.014282584190368652, + 0.0076233320869505405, + 0.007453593425452709, + 0.005790841765701771, + 0.00410030921921134, + 0.00564010813832283, + 0.003833094146102667, + 0.004744899924844503, + 0.012007106095552444, + 0.00787630770355463, + 0.005274046212434769, + 0.005202073138207197, + 0.0042642103508114815, + 0.004716015420854092, + 0.004569490905851126, + 0.00676680775359273, + 0.008061638101935387, + 0.003923098556697369, + 0.0038556132931262255, + 0.00370324170216918, + 0.0042276643216609955, + 0.0035269572399556637, + 0.0031338930130004883, + 0.005229379516094923, + 0.002808018820360303, + 0.006438791286200285, + 0.0033776117488741875, + 0.003888447303324938, + 0.012680718675255775, + 0.004346199333667755, + 0.003426231676712632, + 0.003023215802386403, + 0.005713440012186766, + 0.023269452154636383, + 0.006697217933833599, + 0.00968680065125227, + 0.010044083930552006, + 0.006828974466770887, + 0.006923336070030928, + 0.006448215339332819, + 0.012035293504595757, + 0.014840051531791687, + 0.0058219535276293755, + 0.0108668627217412, + 0.016860583797097206, + 0.025325749069452286, + 0.02496786043047905, + 0.03074065037071705, + 0.053034212440252304, + 0.015780119225382805, + 0.01704103872179985, + 0.02797701396048069, + 0.009999553672969341, + 0.02408759295940399, + 0.017192505300045013, + 0.021344469860196114, + 0.025929072871804237, + 0.012543703429400921, + 0.009637253358960152, + 0.011558057740330696, + 0.02443901263177395, + 0.025401100516319275, + 0.01843656226992607, + 0.015569637529551983, + 0.02944856323301792, + 0.03680635616183281, + 0.017156483605504036, + 0.021094422787427902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.01420164667069912 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808846324682236 + ], + "opposite_logprob_tokens": [ + -3.28808856010437 + ], + "logprob_reward": -0.03808846324682236, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808846324682236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06053746119141579, + 0.046277474611997604, + 0.09543430805206299, + 0.04353562742471695, + 0.04835567623376846, + 0.2212080955505371, + 0.03420434892177582, + 0.2702644467353821, + 0.11017190665006638, + 0.07001065462827682 + ], + "credit_map": [ + 0.004886443726718426, + 0.00746638560667634, + 0.007730626035481691, + 0.00956572312861681, + 0.009011318907141685, + 0.007098463363945484, + 0.008182050660252571, + 0.006596454884856939, + 0.0040683443658053875, + 0.004968077875673771, + 0.006949651520699263, + 0.008728927932679653, + 0.005165355745702982, + 0.004671307280659676, + 0.005858045071363449, + 0.005867764353752136, + 0.008767548948526382, + 0.010424058884382248, + 0.008941862732172012, + 0.009186235256493092, + 0.025366568937897682, + 0.013123897835612297, + 0.00891954731196165, + 0.010704594664275646, + 0.004663554485887289, + 0.004253887105733156, + 0.005251368507742882, + 0.005418119020760059, + 0.010951567441225052, + 0.004051968455314636, + 0.004293886944651604, + 0.00465127918869257, + 0.005739313550293446, + 0.005100988317281008, + 0.006253047846257687, + 0.00661840895190835, + 0.004485046025365591, + 0.007833884097635746, + 0.006664561107754707, + 0.005660430993884802, + 0.02981480583548546, + 0.013187062926590443, + 0.020297160372138023, + 0.011084441095590591, + 0.022615276277065277, + 0.07193157076835632, + 0.018442843109369278, + 0.033834949135780334, + 0.0028214657213538885, + 0.006113684736192226, + 0.0032422696240246296, + 0.003857390256598592, + 0.007908415980637074, + 0.002777972025796771, + 0.003211623290553689, + 0.0042715296149253845, + 0.022449791431427002, + 0.043243926018476486, + 0.0353836826980114, + 0.05241280049085617, + 0.04157336801290512, + 0.025746464729309082, + 0.022232521325349808, + 0.0272219255566597, + 0.009597339667379856, + 0.01256707962602377, + 0.016039008274674416, + 0.0157029926776886, + 0.017622601240873337, + 0.013791529461741447, + 0.010194218717515469, + 0.014657150954008102, + 0.005503413267433643, + 0.011963746510446072, + 0.009192654862999916, + 0.010197838768362999, + 0.011506079696118832, + 0.0064431652426719666, + 0.006661095190793276, + 0.008542664349079132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.03808846324682236 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle close-up elegantly frames the carefully arranged composition on a polished, dark walnut coffee table in a sleek, modern living room. A minimalist white ceramic bowl with a smooth, low-gloss finish sits at its center. Filled with a vibrant assortment of fresh fruit—a crisp green apple, a ripe red pear, and a small bunch of deep purple grapes—their surfaces catch a subtle, healthy sheen from a nearby light source. Juxtaposed against the bowl is a slender black remote control, its matte plastic casing and minimalist design creating a striking geometric contrast to the bowl's organic form. The entire scene is bathed in warm ambient light from a large, out-of-focus window. The soft afternoon sunlight casts long, gentle shadows across the table's polished texture, highlighting the distinct grains of the ceramic, fruit, and wood. The ambiance is one of profound tranquility and refined elegance, emphasizing the quiet harmony and deliberate contrast between natural and technological objects.", + "seed": 24535, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_14/index_5030_seed_24535_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04398928713053465, + "vlm_logprob": -0.04398928713053465, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.04398928713053465, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.039604708552360535 + ], + "opposite_logprob_tokens": [ + -3.414604663848877 + ], + "logprob_reward": -0.039604708552360535, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.039604708552360535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.182216078042984, + 0.03329819440841675, + 0.11771243065595627, + 0.4459303617477417, + 0.03658505901694298, + 0.0562799908220768, + 0.04724521562457085, + 0.02398955076932907, + 0.03392159193754196, + 0.02282145991921425 + ], + "credit_map": [ + 0.018007857725024223, + 0.026648562401533127, + 0.02458333782851696, + 0.017821388319134712, + 0.025531664490699768, + 0.04363878443837166, + 0.013164812698960304, + 0.012819679453969002, + 0.004090587142854929, + 0.0032187411561608315, + 0.003270189044997096, + 0.004043557681143284, + 0.004248274024575949, + 0.005284379236400127, + 0.0030481447465717793, + 0.006094318348914385, + 0.010039303451776505, + 0.016494501382112503, + 0.020651675760746002, + 0.009780134074389935, + 0.020030483603477478, + 0.023160744458436966, + 0.008542347699403763, + 0.009013229049742222, + 0.03662813454866409, + 0.020457476377487183, + 0.12072081118822098, + 0.04811989516019821, + 0.01950068026781082, + 0.15193000435829163, + 0.033203817903995514, + 0.015369553118944168, + 0.004956508055329323, + 0.004559474531561136, + 0.004226791672408581, + 0.00410242285579443, + 0.005034014116972685, + 0.004351283423602581, + 0.003437614534050226, + 0.00591695262119174, + 0.008643601089715958, + 0.009878392331302166, + 0.005748876836150885, + 0.005487467627972364, + 0.010221405886113644, + 0.005290857050567865, + 0.006131134927272797, + 0.0048782555386424065, + 0.005555209703743458, + 0.004880403634160757, + 0.005777356214821339, + 0.005827233660966158, + 0.008099756203591824, + 0.006148707587271929, + 0.004539573099464178, + 0.006416975520551205, + 0.002655935240909457, + 0.003063764190301299, + 0.003122449852526188, + 0.002328143222257495, + 0.004183802288025618, + 0.003261936828494072, + 0.0028047130908817053, + 0.0025688065215945244, + 0.00415626959875226, + 0.005406687036156654, + 0.0034406972117722034, + 0.0033828604500740767, + 0.006834304425865412, + 0.004178975243121386, + 0.002812656806781888, + 0.0037091418635100126, + 0.0025482685305178165, + 0.003708538366481662, + 0.0017878527287393808, + 0.003365318523719907, + 0.003337422851473093, + 0.0026030822191387415, + 0.0030178255401551723, + 0.002453151158988476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.039604708552360535 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903692126274109 + ], + "opposite_logprob_tokens": [ + -2.5790369510650635 + ], + "logprob_reward": -0.07903692126274109, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903692126274109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24263978004455566, + 0.047075990587472916, + 0.09541373699903488, + 0.35702505707740784, + 0.03434789180755615, + 0.013361011631786823, + 0.045590996742248535, + 0.026021026074886322, + 0.021860754117369652, + 0.11666376888751984 + ], + "credit_map": [ + 0.014035101048648357, + 0.017210382968187332, + 0.07219179719686508, + 0.017446039244532585, + 0.0194285549223423, + 0.08268973231315613, + 0.010463004931807518, + 0.00917518138885498, + 0.0051192715764045715, + 0.006523145828396082, + 0.005295529495924711, + 0.005943679250776768, + 0.0062857381999492645, + 0.008511240594089031, + 0.004279726650565863, + 0.0051176585257053375, + 0.008406654000282288, + 0.011877382174134254, + 0.018545132130384445, + 0.008498058654367924, + 0.013608081266283989, + 0.022192146629095078, + 0.006455892231315374, + 0.005830394569784403, + 0.027636198326945305, + 0.01514038909226656, + 0.09742531180381775, + 0.0374554879963398, + 0.01524278149008751, + 0.12487815320491791, + 0.027187321335077286, + 0.012059438973665237, + 0.00535176694393158, + 0.003862364450469613, + 0.0032230447977781296, + 0.004825397860258818, + 0.003336047986522317, + 0.004257108084857464, + 0.002905450761318207, + 0.006586711388081312, + 0.001584945130161941, + 0.0019339816644787788, + 0.0016486773965880275, + 0.0013318683486431837, + 0.0023712858092039824, + 0.0014459630474448204, + 0.0012078735744580626, + 0.0018364163115620613, + 0.00574923912063241, + 0.006561495363712311, + 0.004509815480560064, + 0.006832277402281761, + 0.005457677878439426, + 0.00593174435198307, + 0.005212254822254181, + 0.005336494650691748, + 0.003228059271350503, + 0.0028961077332496643, + 0.002172156237065792, + 0.0032312420662492514, + 0.005232759285718203, + 0.0029875158797949553, + 0.0033202620688825846, + 0.0029529226012527943, + 0.002753568347543478, + 0.0031543574295938015, + 0.001996672712266445, + 0.0019697551615536213, + 0.0041493321768939495, + 0.0033145754132419825, + 0.0018562375335022807, + 0.002666253363713622, + 0.009133200161159039, + 0.01262218039482832, + 0.008833682164549828, + 0.02038642391562462, + 0.016652878373861313, + 0.018622908741235733, + 0.018766719847917557, + 0.011645786464214325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.07903692126274109 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097876638174057 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.0097876638174057, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097876638174057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4580077826976776, + 0.0908442959189415, + 0.046255845576524734, + 0.1666824221611023, + 0.05010845884680748, + 0.050982676446437836, + 0.01985280029475689, + 0.04205765202641487, + 0.025905724614858627, + 0.049302324652671814 + ], + "credit_map": [ + 0.024978872388601303, + 0.02446705289185047, + 0.1561913788318634, + 0.029529480263590813, + 0.02129805088043213, + 0.16911666095256805, + 0.01575046218931675, + 0.016675833612680435, + 0.0076337456703186035, + 0.0149885518476367, + 0.010124867781996727, + 0.009045266546308994, + 0.019402222707867622, + 0.011559627950191498, + 0.0068746814504265785, + 0.011215333826839924, + 0.005261124577373266, + 0.005980998743325472, + 0.004904554691165686, + 0.006247089710086584, + 0.00608414551243186, + 0.009518539533019066, + 0.004379978869110346, + 0.0038794137071818113, + 0.012784778140485287, + 0.007771425414830446, + 0.043104417622089386, + 0.016675980761647224, + 0.009534494951367378, + 0.058407288044691086, + 0.012725597247481346, + 0.005678441841155291, + 0.005378302652388811, + 0.00741969421505928, + 0.004542047623544931, + 0.007239479571580887, + 0.006532728672027588, + 0.004676359239965677, + 0.005766554735600948, + 0.008553295396268368, + 0.007691455539315939, + 0.007936108857393265, + 0.005507240071892738, + 0.006039801053702831, + 0.008192245848476887, + 0.005434160120785236, + 0.00519095454365015, + 0.004990710411220789, + 0.002863371977582574, + 0.0022325001191347837, + 0.0015465585747733712, + 0.002483726479113102, + 0.0031937926542013884, + 0.0026451521553099155, + 0.0021316674537956715, + 0.0027560305316001177, + 0.003819418838247657, + 0.004176161717623472, + 0.0046134223230183125, + 0.0051666260696947575, + 0.006139772478491068, + 0.006381150800734758, + 0.006887361407279968, + 0.004873739089816809, + 0.003090917831286788, + 0.003088421653956175, + 0.0019183829426765442, + 0.0028643484693020582, + 0.004288904834538698, + 0.003685368224978447, + 0.004285245202481747, + 0.002684136386960745, + 0.004924161825329065, + 0.0073904842138290405, + 0.003371035447344184, + 0.007368356920778751, + 0.007333774119615555, + 0.005662242881953716, + 0.007411404978483915, + 0.005840866826474667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0097876638174057 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549950741231441 + ], + "opposite_logprob_tokens": [ + -4.387549877166748 + ], + "logprob_reward": -0.012549950741231441, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549950741231441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28365737199783325, + 0.09487711638212204, + 0.21437974274158478, + 0.06185859441757202, + 0.06826018542051315, + 0.05861940607428551, + 0.022560931742191315, + 0.07310692965984344, + 0.02460320107638836, + 0.09807649999856949 + ], + "credit_map": [ + 0.024886617437005043, + 0.03786034882068634, + 0.0616997554898262, + 0.023410094901919365, + 0.037228357046842575, + 0.05818496644496918, + 0.019231075420975685, + 0.02115614153444767, + 0.007663071621209383, + 0.01704506389796734, + 0.00885062012821436, + 0.013340620324015617, + 0.017210256308317184, + 0.012652111239731312, + 0.010975044220685959, + 0.007140330038964748, + 0.01800922490656376, + 0.02269410900771618, + 0.04387391731142998, + 0.02227380871772766, + 0.023670025169849396, + 0.05651237443089485, + 0.012928222306072712, + 0.01441806647926569, + 0.005360601469874382, + 0.003124252427369356, + 0.01580127701163292, + 0.006960952188819647, + 0.0031747573520988226, + 0.020210521295666695, + 0.004597779363393784, + 0.002628450049087405, + 0.008102834224700928, + 0.012563967145979404, + 0.005864586681127548, + 0.009810972027480602, + 0.014365339651703835, + 0.006255196873098612, + 0.005582310725003481, + 0.005714976694434881, + 0.010030917823314667, + 0.009598850272595882, + 0.0069478717632591724, + 0.0068683684803545475, + 0.009619304910302162, + 0.005446468014270067, + 0.005519463215023279, + 0.004588166251778603, + 0.0035930154845118523, + 0.0025481085758656263, + 0.001809204462915659, + 0.0029121951665729284, + 0.002711094915866852, + 0.003722960827872157, + 0.0024699694477021694, + 0.0027943826280534267, + 0.00961883831769228, + 0.007773854769766331, + 0.0059507968835532665, + 0.008266929537057877, + 0.015470539219677448, + 0.008249206468462944, + 0.009645821526646614, + 0.008130944333970547, + 0.0031203217804431915, + 0.003543945262208581, + 0.0020075556822121143, + 0.002979474375024438, + 0.0038812083657830954, + 0.0032082705292850733, + 0.0029462052043527365, + 0.0029162191785871983, + 0.010227593593299389, + 0.015086113475263119, + 0.008394353091716766, + 0.012404879555106163, + 0.019565975293517113, + 0.010940917767584324, + 0.010631885379552841, + 0.010824782773852348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.012549950741231441 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static close-up shot of a coffee table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coffee table made of polished dark walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a minimalist white ceramic bowl with a smooth low-gloss finish visible at the center of the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bowl contain a green apple, a red pear, and a bunch of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a slender black remote control with a matte plastic casing visible next to the bowl?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896719127893448 + ], + "opposite_logprob_tokens": [ + -2.5789670944213867 + ], + "logprob_reward": -0.07896719127893448, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896719127893448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40995320677757263, + 0.08436727523803711, + 0.17043209075927734, + 0.07109445333480835, + 0.04078880324959755, + 0.09487044811248779, + 0.022925186902284622, + 0.03142067790031433, + 0.029962705448269844, + 0.04418513551354408 + ], + "credit_map": [ + 0.021285280585289, + 0.020326223224401474, + 0.1538744419813156, + 0.031063416972756386, + 0.013012401759624481, + 0.14568465948104858, + 0.012511597014963627, + 0.012195192277431488, + 0.007755615282803774, + 0.00959655363112688, + 0.007796254474669695, + 0.013763795606791973, + 0.011426406912505627, + 0.009047864936292171, + 0.015034166164696217, + 0.009946626611053944, + 0.017864422872662544, + 0.021874597296118736, + 0.02679927833378315, + 0.020939473062753677, + 0.025197576731443405, + 0.0306793674826622, + 0.01367173995822668, + 0.013405642472207546, + 0.010922417975962162, + 0.010097090154886246, + 0.006328013725578785, + 0.0072375014424324036, + 0.00899448711425066, + 0.00899723544716835, + 0.010248787701129913, + 0.008268921636044979, + 0.005467081908136606, + 0.005888047162443399, + 0.00490990187972784, + 0.005010953638702631, + 0.005329039413481951, + 0.0038062320090830326, + 0.002919500693678856, + 0.007458049338310957, + 0.009475176222622395, + 0.008010659366846085, + 0.007847331464290619, + 0.017268046736717224, + 0.011033507063984871, + 0.01124531589448452, + 0.010338760912418365, + 0.01965164579451084, + 0.0021471725776791573, + 0.0015052371891215444, + 0.0017699297750368714, + 0.004029702860862017, + 0.0016718216938897967, + 0.003039152594283223, + 0.003716216655448079, + 0.005045954138040543, + 0.0032204780727624893, + 0.003233271185308695, + 0.0035690374206751585, + 0.004177561495453119, + 0.005671952851116657, + 0.002997626317664981, + 0.005387916695326567, + 0.0031628331635147333, + 0.0031852254178375006, + 0.0038936855271458626, + 0.0025007561780512333, + 0.003330323612317443, + 0.005748271476477385, + 0.003262192476540804, + 0.003568677231669426, + 0.00447357352823019, + 0.0037389479111880064, + 0.0058039408177137375, + 0.003388188546523452, + 0.007000557146966457, + 0.006492187734693289, + 0.005777675658464432, + 0.006754035595804453, + 0.0052296011708676815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.07896719127893448 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_15/index_tensor([151, 151])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_15/index_tensor([151, 151])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ef80f2b9a5226aa533dd155c774981cfa8072644 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_15/index_tensor([151, 151])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static shot is positioned across a deserted asphalt road, its surface gleaming and rain-soaked at night. From the deep background, a pair of brilliant white LED headlights suddenly pierce the oppressive darkness, casting long, dancing beams that illuminate the fine mist hanging in the cold air and the texture of the wet road. A sleek, dark sedan appears, speeding past in a blur of motion. Its sleek, damp undercarriage reflects the fleeting, distorted world around it as its tires plow through the puddles, kicking up a dramatic, fan-shaped spray that momentarily catches the light. As quickly as it appears, the vehicle retreats, its form vanishing into the night, leaving only the intense, vibrant red glow of its twin taillights. These lights paint long, painterly streaks across the black mirror of the road, rapidly shrinking until they become two crimson pinpricks, finally vanishing, plunging the scene back into near-total darkness and a profound, mysterious atmosphere.", + "seed": 22662, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_15/index_151_seed_22662_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05232392552425154, + "vlm_logprob": -0.05232392552425154, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.05232392552425154, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05056294798851013 + ], + "opposite_logprob_tokens": [ + -6.300562858581543 + ], + "logprob_reward": -0.05056294798851013, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.05056294798851013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13387811183929443, + 0.02715134248137474, + 0.2887602746486664, + 0.1040770411491394, + 0.08138660341501236, + 0.0940714031457901, + 0.10359417647123337, + 0.067336805164814, + 0.03978082165122032, + 0.059963472187519073 + ], + "credit_map": [ + 0.01114316750317812, + 0.02478155493736267, + 0.0170546006411314, + 0.018116222694516182, + 0.018043292686343193, + 0.010498088784515858, + 0.018432093784213066, + 0.0158090777695179, + 0.0028564820531755686, + 0.0027925900649279356, + 0.00459138210862875, + 0.002615417120978236, + 0.0029554893262684345, + 0.0027249616105109453, + 0.0035659840796142817, + 0.0050490377470850945, + 0.028537720441818237, + 0.09411904215812683, + 0.042232636362314224, + 0.04922143742442131, + 0.027388479560613632, + 0.014389045536518097, + 0.0233425535261631, + 0.00952937826514244, + 0.009969362989068031, + 0.01310410350561142, + 0.013105162419378757, + 0.01699335314333439, + 0.01614089496433735, + 0.006589866708964109, + 0.013435610570013523, + 0.014738688245415688, + 0.010253376327455044, + 0.011325804516673088, + 0.01298043318092823, + 0.01537234429270029, + 0.007070604246109724, + 0.009692198596894741, + 0.009028259664773941, + 0.00566358445212245, + 0.008386204950511456, + 0.0062602851539850235, + 0.019722366705536842, + 0.01658332347869873, + 0.008608625270426273, + 0.007812203839421272, + 0.013375120237469673, + 0.013323269784450531, + 0.011620775796473026, + 0.00735473120585084, + 0.017751622945070267, + 0.011990128085017204, + 0.010493744164705276, + 0.022279666736721992, + 0.007007373962551355, + 0.015096135437488556, + 0.011261659674346447, + 0.00780745642259717, + 0.009911106899380684, + 0.009936543181538582, + 0.005997444037348032, + 0.0062227435410022736, + 0.006357009056955576, + 0.009842844679951668, + 0.006979796569794416, + 0.007461717817932367, + 0.0068189832381904125, + 0.00406991271302104, + 0.004921824671328068, + 0.0031659898813813925, + 0.003017138922587037, + 0.00334545667283237, + 0.013896972872316837, + 0.010038014501333237, + 0.011375000700354576, + 0.004304513335227966, + 0.0069623589515686035, + 0.0072023700922727585, + 0.0022806464694440365, + 0.0039035931695252657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.05056294798851013 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00043335105874575675 + ], + "opposite_logprob_tokens": [ + -7.875433444976807 + ], + "logprob_reward": -0.00043335105874575675, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.00043335105874575675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08283759653568268, + 0.02185993082821369, + 0.41704630851745605, + 0.06433209776878357, + 0.14087419211864471, + 0.05075879395008087, + 0.06266109645366669, + 0.04202824831008911, + 0.07064604014158249, + 0.04695562645792961 + ], + "credit_map": [ + 0.0069342367351055145, + 0.009622056968510151, + 0.011234117671847343, + 0.007838990539312363, + 0.01197625882923603, + 0.008774655871093273, + 0.011877961456775665, + 0.014579325914382935, + 0.0023638163693249226, + 0.003345127683132887, + 0.0031399468425661325, + 0.0019030128605663776, + 0.0020997014362365007, + 0.0018235042225569487, + 0.0028902285266667604, + 0.0042945947498083115, + 0.04413691535592079, + 0.13798342645168304, + 0.05415830388665199, + 0.08043083548545837, + 0.03540191426873207, + 0.019644254818558693, + 0.032291002571582794, + 0.012999696657061577, + 0.007937466725707054, + 0.006621510721743107, + 0.00753371836617589, + 0.008527915924787521, + 0.010626411996781826, + 0.00583853293210268, + 0.0072041964158415794, + 0.010042351670563221, + 0.011983383446931839, + 0.01681315340101719, + 0.026181137189269066, + 0.03160591423511505, + 0.014768939465284348, + 0.010736123658716679, + 0.01814616285264492, + 0.01063939742743969, + 0.00459005543962121, + 0.003843097249045968, + 0.009225798770785332, + 0.008494067005813122, + 0.005605833139270544, + 0.005684905685484409, + 0.007554076611995697, + 0.005760963074862957, + 0.006457828916609287, + 0.005853597540408373, + 0.009772569872438908, + 0.010495125316083431, + 0.007724313996732235, + 0.009398991242051125, + 0.006583469919860363, + 0.006375203374773264, + 0.007770739030092955, + 0.003958653658628464, + 0.006733554415404797, + 0.0046750218607485294, + 0.004773723427206278, + 0.004615502431988716, + 0.004854635801166296, + 0.004646419547498226, + 0.008646705187857151, + 0.013118579983711243, + 0.011274426244199276, + 0.007179912645369768, + 0.010298349894583225, + 0.008037304505705833, + 0.008073138073086739, + 0.004017631523311138, + 0.010985448956489563, + 0.004340569023042917, + 0.008121083490550518, + 0.003147843759506941, + 0.00465401541441679, + 0.0070791891776025295, + 0.0026267138309776783, + 0.006000767927616835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.00043335105874575675 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197840079665184 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.018197840079665184, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197840079665184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04457506537437439, + 0.036546822637319565, + 0.45598694682121277, + 0.0848621129989624, + 0.10749143362045288, + 0.05430750176310539, + 0.11030025035142899, + 0.06517786532640457, + 0.024985484778881073, + 0.015766574069857597 + ], + "credit_map": [ + 0.0038255946710705757, + 0.005801309831440449, + 0.006281447596848011, + 0.005266789346933365, + 0.006720114033669233, + 0.004013469908386469, + 0.007062794640660286, + 0.005603543017059565, + 0.003416465362533927, + 0.011423617601394653, + 0.003560215700417757, + 0.006202989257872105, + 0.0033937073312699795, + 0.0022226525470614433, + 0.003733130404725671, + 0.00259404256939888, + 0.05103225260972977, + 0.16999012231826782, + 0.03214304894208908, + 0.105972059071064, + 0.0395624116063118, + 0.018821027129888535, + 0.029952334240078926, + 0.008513686247169971, + 0.007257224526256323, + 0.012110192328691483, + 0.007652288302779198, + 0.014088274911046028, + 0.011458491906523705, + 0.008337192237377167, + 0.011720904149115086, + 0.012237530201673508, + 0.011784560047090054, + 0.018351906910538673, + 0.0188278965651989, + 0.01669720560312271, + 0.008765899576246738, + 0.015210594981908798, + 0.011657512746751308, + 0.006195852067321539, + 0.005141968838870525, + 0.005090632010251284, + 0.010283770971000195, + 0.010425131767988205, + 0.007039352785795927, + 0.005037426482886076, + 0.00685174111276865, + 0.004437469877302647, + 0.012831219471991062, + 0.009019635617733002, + 0.01807110197842121, + 0.01543845422565937, + 0.013458783738315105, + 0.019583938643336296, + 0.00946267694234848, + 0.012434432283043861, + 0.010137447156012058, + 0.005324702709913254, + 0.01033655647188425, + 0.007238858845084906, + 0.00677231652662158, + 0.010632512159645557, + 0.006901619490236044, + 0.007833843119442463, + 0.0028281002305448055, + 0.003754320787265897, + 0.004383965861052275, + 0.0032877775374799967, + 0.0021826960146427155, + 0.003592523979023099, + 0.0031476274598389864, + 0.0018084704643115401, + 0.0027945758774876595, + 0.002750610001385212, + 0.0028972390573471785, + 0.0013350644148886204, + 0.0019865387585014105, + 0.0017826255643740296, + 0.000986196449957788, + 0.0012337230145931244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 4 + ], + "selected_reward": -0.018197840079665184 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541710399091244 + ], + "opposite_logprob_tokens": [ + -4.387541770935059 + ], + "logprob_reward": -0.012541710399091244, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541710399091244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058320801705121994, + 0.08235964924097061, + 0.34037327766418457, + 0.09641266614198685, + 0.04362431541085243, + 0.055121611803770065, + 0.07625465840101242, + 0.11978879570960999, + 0.04923853278160095, + 0.07850570231676102 + ], + "credit_map": [ + 0.004893674980849028, + 0.010035892017185688, + 0.004981070291250944, + 0.007020170800387859, + 0.00978640653192997, + 0.007365720812231302, + 0.006627202965319157, + 0.007610656321048737, + 0.007231784053146839, + 0.02752716653048992, + 0.0076394216157495975, + 0.015413652174174786, + 0.007711049169301987, + 0.0043334453366696835, + 0.0077618700452148914, + 0.004741254262626171, + 0.039436325430870056, + 0.12598775327205658, + 0.020666904747486115, + 0.08259151130914688, + 0.028069617226719856, + 0.014317790046334267, + 0.02238122560083866, + 0.006922134198248386, + 0.007887566462159157, + 0.016794640570878983, + 0.01007465273141861, + 0.01581788994371891, + 0.01619076170027256, + 0.0076494538225233555, + 0.010928168892860413, + 0.011069525964558125, + 0.006573460064828396, + 0.009019310586154461, + 0.007528326008468866, + 0.005319558549672365, + 0.003188112750649452, + 0.004989318083971739, + 0.00341907050460577, + 0.00358715676702559, + 0.005992366932332516, + 0.00503961369395256, + 0.010514582507312298, + 0.009531697258353233, + 0.0064454954117536545, + 0.004226804710924625, + 0.007806848734617233, + 0.005564198363572359, + 0.008843502961099148, + 0.00799231231212616, + 0.011659231968224049, + 0.012182646431028843, + 0.008519167080521584, + 0.011319857090711594, + 0.006467975210398436, + 0.0092699583619833, + 0.018235689029097557, + 0.012308266945183277, + 0.01754669100046158, + 0.017340464517474174, + 0.016415253281593323, + 0.014066668227314949, + 0.01115431822836399, + 0.012721432372927666, + 0.006627938710153103, + 0.005469838157296181, + 0.008538547903299332, + 0.004152157809585333, + 0.006890788674354553, + 0.007662053219974041, + 0.005523757543414831, + 0.004373451694846153, + 0.013959439471364021, + 0.012568804435431957, + 0.011993474327027798, + 0.006624320056289434, + 0.008703922852873802, + 0.013197215273976326, + 0.004100654274225235, + 0.007357857655733824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.012541710399091244 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988377809524536 + ], + "opposite_logprob_tokens": [ + -1.8048837184906006 + ], + "logprob_reward": -0.17988377809524536, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988377809524536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06777116656303406, + 0.04921236261725426, + 0.16515175998210907, + 0.14694742858409882, + 0.1430150270462036, + 0.18753434717655182, + 0.04924359917640686, + 0.10998474061489105, + 0.04432937502861023, + 0.036810219287872314 + ], + "credit_map": [ + 0.004961404483765364, + 0.009602067060768604, + 0.007656473200768232, + 0.007896543480455875, + 0.00823940709233284, + 0.008298231288790703, + 0.008331341668963432, + 0.012785695493221283, + 0.005060248542577028, + 0.00775935361161828, + 0.00793537124991417, + 0.006072142627090216, + 0.004572055768221617, + 0.0033082927111536264, + 0.008247860707342625, + 0.006257037632167339, + 0.016826827079057693, + 0.04179555922746658, + 0.022979309782385826, + 0.0318022258579731, + 0.012610946781933308, + 0.007139638997614384, + 0.021075822412967682, + 0.010921448469161987, + 0.01658642292022705, + 0.012049928307533264, + 0.015655111521482468, + 0.01955271139740944, + 0.020754439756274223, + 0.016012681648135185, + 0.019469600170850754, + 0.02686653472483158, + 0.019972622394561768, + 0.02363712154328823, + 0.025044837966561317, + 0.022745711728930473, + 0.014510664157569408, + 0.012645317241549492, + 0.010181758552789688, + 0.01427699625492096, + 0.01575753465294838, + 0.01571570709347725, + 0.03764915093779564, + 0.03778457269072533, + 0.022305989637970924, + 0.015801647678017616, + 0.02593267522752285, + 0.01658705249428749, + 0.006173846311867237, + 0.003808951238170266, + 0.008478092029690742, + 0.0070439609698951244, + 0.0049599576741456985, + 0.009491752833127975, + 0.003485993482172489, + 0.005801043007522821, + 0.019365482032299042, + 0.013874227181077003, + 0.017621081322431564, + 0.01465864758938551, + 0.009922850877046585, + 0.01285227295011282, + 0.009686952456831932, + 0.012003218755126, + 0.009785068221390247, + 0.006353218108415604, + 0.008838564157485962, + 0.0038036706391721964, + 0.005955359432846308, + 0.004354282282292843, + 0.0025746107567101717, + 0.002664602128788829, + 0.007538979407399893, + 0.006446429528295994, + 0.006697220262140036, + 0.003541892860084772, + 0.003284280654042959, + 0.004124996718019247, + 0.0022800525184720755, + 0.0028963685035705566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.17988377809524536 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot is positioned across a deserted asphalt road, its surface gleaming and rain-soaked at night. From the deep background, a pair of brilliant white LED headlights suddenly pierce the oppressive darkness, casting long, dancing beams that illuminate the fine mist hanging in the cold air and the texture of the wet road. A sleek, dark sedan appears, speeding past in a blur of motion. Its sleek, damp undercarriage reflects the fleeting, distorted world around it as its tires plow through the puddles, kicking up a dramatic, fan-shaped spray that momentarily catches the light. As quickly as it appears, the vehicle retreats, its form vanishing into the night, leaving only the intense, vibrant red glow of its twin taillights. These lights paint long, painterly streaks across the black mirror of the road, rapidly shrinking until they become two crimson pinpricks, finally vanishing, plunging the scene back into near-total darkness and a profound, mysterious atmosphere.", + "seed": 22663, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_15/index_151_seed_22663_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09556045962963253, + "vlm_logprob": -0.09556045962963253, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.09556045962963253, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05849847570061684 + ], + "opposite_logprob_tokens": [ + -5.683498382568359 + ], + "logprob_reward": -0.05849847570061684, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.05849847570061684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10421343147754669, + 0.038599878549575806, + 0.05738571658730507, + 0.09072677046060562, + 0.15411382913589478, + 0.14012864232063293, + 0.12975318729877472, + 0.08567050844430923, + 0.08247176557779312, + 0.11693629622459412 + ], + "credit_map": [ + 0.010952461510896683, + 0.010462429374456406, + 0.01401230227202177, + 0.010041619651019573, + 0.012385450303554535, + 0.016719361767172813, + 0.013002362102270126, + 0.016637448221445084, + 0.004652678966522217, + 0.00558870192617178, + 0.006987723987549543, + 0.004102601669728756, + 0.0069108097814023495, + 0.0051336404867470264, + 0.002633375581353903, + 0.0025903480127453804, + 0.008725151419639587, + 0.0056108455173671246, + 0.012130582705140114, + 0.004570298362523317, + 0.009040753357112408, + 0.009747544303536415, + 0.004331790376454592, + 0.003228751476854086, + 0.017733968794345856, + 0.008974051102995872, + 0.0120534123852849, + 0.010303059592843056, + 0.010501095093786716, + 0.008678378537297249, + 0.007441218476742506, + 0.0150415925309062, + 0.03920658305287361, + 0.010376770049333572, + 0.013134288601577282, + 0.009594742208719254, + 0.008332091383635998, + 0.018552519381046295, + 0.012576033361256123, + 0.04234080761671066, + 0.013176320120692253, + 0.014819132164120674, + 0.023090451955795288, + 0.012570474296808243, + 0.017603669315576553, + 0.014646277762949467, + 0.021981967613101006, + 0.02224033698439598, + 0.014927303418517113, + 0.017252665013074875, + 0.014163519255816936, + 0.008116730488836765, + 0.028103800490498543, + 0.02606111206114292, + 0.010274741798639297, + 0.010853319428861141, + 0.01459585689008236, + 0.012029941193759441, + 0.014394646510481834, + 0.011759843677282333, + 0.012532016262412071, + 0.009735425934195518, + 0.004736534785479307, + 0.0058862511068582535, + 0.01426697801798582, + 0.016596879810094833, + 0.013121348805725574, + 0.007431208621710539, + 0.008253870531916618, + 0.015027083456516266, + 0.0036656346637755632, + 0.0041087595745921135, + 0.02750193513929844, + 0.01865161396563053, + 0.018677761778235435, + 0.012889798730611801, + 0.018052494153380394, + 0.014097291976213455, + 0.0029404626693576574, + 0.004124938510358334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.05849847570061684 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001388062839396298 + ], + "opposite_logprob_tokens": [ + -6.626388072967529 + ], + "logprob_reward": -0.001388062839396298, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001388062839396298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03201597183942795, + 0.042241357266902924, + 0.02411486767232418, + 0.15608614683151245, + 0.1443129926919937, + 0.11265815794467926, + 0.09760551154613495, + 0.20258133113384247, + 0.10731427371501923, + 0.0810694620013237 + ], + "credit_map": [ + 0.0033110794611275196, + 0.0028169211000204086, + 0.004755267407745123, + 0.0029445872642099857, + 0.0045747156254947186, + 0.004285356495529413, + 0.00468778470531106, + 0.0046402583830058575, + 0.007798244245350361, + 0.004809185862541199, + 0.005909849889576435, + 0.004217648413032293, + 0.005650238133966923, + 0.007332203444093466, + 0.0031243297271430492, + 0.003399657318368554, + 0.004036338068544865, + 0.001980679575353861, + 0.003798440098762512, + 0.0023682010360062122, + 0.0036555530969053507, + 0.00407477468252182, + 0.0022727339528501034, + 0.0019281473942101002, + 0.04775821790099144, + 0.010678711347281933, + 0.010462242178618908, + 0.0094687445089221, + 0.009775272570550442, + 0.015462599694728851, + 0.01290898397564888, + 0.0395713709294796, + 0.043523214757442474, + 0.008232537657022476, + 0.009203847497701645, + 0.007413529325276613, + 0.007413021754473448, + 0.009203403256833553, + 0.011596933007240295, + 0.04772650822997093, + 0.016459792852401733, + 0.011692599393427372, + 0.011849281378090382, + 0.006700619589537382, + 0.027287175878882408, + 0.013924687169492245, + 0.012096247635781765, + 0.01264775637537241, + 0.00847180001437664, + 0.01618146523833275, + 0.006071859039366245, + 0.006212538108229637, + 0.02143295668065548, + 0.022027935832738876, + 0.008840666152536869, + 0.0083662960678339, + 0.019031822681427002, + 0.035971809178590775, + 0.02863081358373165, + 0.022330915555357933, + 0.04046857729554176, + 0.028172409161925316, + 0.012095835991203785, + 0.015879157930612564, + 0.01012337300926447, + 0.0260842964053154, + 0.010026799514889717, + 0.010674328543245792, + 0.016334235668182373, + 0.02242717146873474, + 0.005886265076696873, + 0.0057577998377382755, + 0.014052978716790676, + 0.02114659734070301, + 0.007948609068989754, + 0.005745086818933487, + 0.008526246063411236, + 0.0173202957957983, + 0.00306140654720366, + 0.0032682372257113457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.001388062839396298 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020600175485014915 + ], + "opposite_logprob_tokens": [ + -3.8956000804901123 + ], + "logprob_reward": -0.020600175485014915, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020600175485014915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06780361384153366, + 0.08824387192726135, + 0.07892993092536926, + 0.13179144263267517, + 0.09188222885131836, + 0.2009929120540619, + 0.22704708576202393, + 0.041354067623615265, + 0.03878194838762283, + 0.03317287564277649 + ], + "credit_map": [ + 0.013307612389326096, + 0.007902835495769978, + 0.010393726639449596, + 0.00539717311039567, + 0.009917592629790306, + 0.01072091143578291, + 0.004901213571429253, + 0.005262548569589853, + 0.009602761827409267, + 0.009966972284018993, + 0.012615147046744823, + 0.010062731802463531, + 0.0152213703840971, + 0.012519569136202335, + 0.007429227232933044, + 0.010826089419424534, + 0.012072526849806309, + 0.0076264673843979836, + 0.019160354509949684, + 0.00665294798091054, + 0.010874406434595585, + 0.01286725141108036, + 0.005545514170080423, + 0.0041304659098386765, + 0.02285446785390377, + 0.01761636883020401, + 0.015573017299175262, + 0.021120762452483177, + 0.015511935576796532, + 0.015324927866458893, + 0.010786649771034718, + 0.013003304600715637, + 0.022681094706058502, + 0.004444582387804985, + 0.006668846122920513, + 0.0063023739494383335, + 0.0062089599668979645, + 0.014635881409049034, + 0.010095391422510147, + 0.02084510028362274, + 0.027122803032398224, + 0.02629297785460949, + 0.017671141773462296, + 0.017768459394574165, + 0.0392327643930912, + 0.03067007102072239, + 0.023257821798324585, + 0.018976883962750435, + 0.02034172974526882, + 0.04357143118977547, + 0.015079754404723644, + 0.013318447396159172, + 0.042918674647808075, + 0.06097794696688652, + 0.016338765621185303, + 0.014500334858894348, + 0.004891522694379091, + 0.0069514065980911255, + 0.004556152503937483, + 0.005069247912615538, + 0.00965122226625681, + 0.006262336391955614, + 0.001998680643737316, + 0.0019735044334083796, + 0.005376311019062996, + 0.007479209918528795, + 0.004275834653526545, + 0.005349898245185614, + 0.006841846741735935, + 0.005932734813541174, + 0.0018245071405544877, + 0.0017016044585034251, + 0.005028484854847193, + 0.00953967310488224, + 0.002678635763004422, + 0.0027383172418922186, + 0.002959229750558734, + 0.008135729469358921, + 0.0010197004303336143, + 0.0010731053771451116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.020600175485014915 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048662953078746796 + ], + "opposite_logprob_tokens": [ + -3.0486629009246826 + ], + "logprob_reward": -0.048662953078746796, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048662953078746796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031502459198236465, + 0.06287387013435364, + 0.052692659199237823, + 0.151459202170372, + 0.10248653590679169, + 0.12674987316131592, + 0.3763052225112915, + 0.034046825021505356, + 0.037887006998062134, + 0.023996300995349884 + ], + "credit_map": [ + 0.005464484915137291, + 0.0026633248198777437, + 0.004669512156397104, + 0.0025289631448686123, + 0.004233894404023886, + 0.004973598290234804, + 0.003382593160495162, + 0.0035860901698470116, + 0.010301453061401844, + 0.006370841525495052, + 0.009154380299150944, + 0.008105047978460789, + 0.008825128898024559, + 0.009199540130794048, + 0.00502470787614584, + 0.005892772693186998, + 0.00820944458246231, + 0.005066066049039364, + 0.009270348586142063, + 0.004982135258615017, + 0.008804845623672009, + 0.00819524098187685, + 0.0042219641618430614, + 0.003942615818232298, + 0.039537884294986725, + 0.014653125777840614, + 0.011030086316168308, + 0.010431230068206787, + 0.011758729815483093, + 0.017566561698913574, + 0.013575291261076927, + 0.03290628641843796, + 0.0278361476957798, + 0.00718041742220521, + 0.007313276641070843, + 0.005900352727621794, + 0.0059003522619605064, + 0.009993785060942173, + 0.0077262395061552525, + 0.03063596412539482, + 0.014051130041480064, + 0.013339972123503685, + 0.008924117311835289, + 0.014998954720795155, + 0.020620066672563553, + 0.016971513628959656, + 0.023520158603787422, + 0.014323965646326542, + 0.027772128582000732, + 0.07729798555374146, + 0.02218407765030861, + 0.02075895108282566, + 0.07722540199756622, + 0.10297246277332306, + 0.026310855522751808, + 0.02178335189819336, + 0.0039002709090709686, + 0.005498195998370647, + 0.007091815117746592, + 0.0026307106018066406, + 0.007063272874802351, + 0.003761483123525977, + 0.002068133559077978, + 0.0020329428371042013, + 0.0040630754083395, + 0.010287699289619923, + 0.005165750160813332, + 0.003601900301873684, + 0.005594911053776741, + 0.0057561565190553665, + 0.0015806708252057433, + 0.001836842973716557, + 0.003722598310559988, + 0.005851148162037134, + 0.0023761827033013105, + 0.0024462260771542788, + 0.002178523689508438, + 0.005418028216809034, + 0.0009007509797811508, + 0.001102842972613871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.048662953078746796 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486526310443878 + ], + "opposite_logprob_tokens": [ + -1.2236526012420654 + ], + "logprob_reward": -0.3486526310443878, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486526310443878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08370254188776016, + 0.02978687919676304, + 0.051212575286626816, + 0.18987442553043365, + 0.12708371877670288, + 0.14911629259586334, + 0.21835418045520782, + 0.04055938497185707, + 0.05498983711004257, + 0.055320192128419876 + ], + "credit_map": [ + 0.010600235313177109, + 0.007852958515286446, + 0.015270337462425232, + 0.006497325841337442, + 0.009248401038348675, + 0.012266064994037151, + 0.010817296802997589, + 0.011149924248456955, + 0.004753141663968563, + 0.004644651431590319, + 0.004342643544077873, + 0.0029399816412478685, + 0.0033846695441752672, + 0.004804042633622885, + 0.0019060950726270676, + 0.003011653432622552, + 0.006297003943473101, + 0.006140768527984619, + 0.009498563595116138, + 0.005195504520088434, + 0.008712802082300186, + 0.007345870137214661, + 0.0050673759542405605, + 0.002954686526209116, + 0.05874120444059372, + 0.012416557408869267, + 0.013697788119316101, + 0.009519723244011402, + 0.015240933746099472, + 0.019886761903762817, + 0.013103056699037552, + 0.04726839438080788, + 0.03691607713699341, + 0.008719351142644882, + 0.008922776207327843, + 0.00763216195628047, + 0.006454619579017162, + 0.010873149149119854, + 0.010374492965638638, + 0.03719109296798706, + 0.014617660082876682, + 0.01869453862309456, + 0.025411678478121758, + 0.016545524820685387, + 0.023314790800213814, + 0.016247915104031563, + 0.016406189650297165, + 0.017877992242574692, + 0.026118040084838867, + 0.03436587378382683, + 0.022632984444499016, + 0.012095506303012371, + 0.03923798352479935, + 0.053306832909584045, + 0.01737460494041443, + 0.013222354464232922, + 0.007198639679700136, + 0.0067178793251514435, + 0.007933489978313446, + 0.0033384663984179497, + 0.00585586205124855, + 0.005300413351505995, + 0.0020653733517974615, + 0.002149256644770503, + 0.00593512412160635, + 0.016652094200253487, + 0.005684325937181711, + 0.003891374683007598, + 0.008539510890841484, + 0.011162535287439823, + 0.0015224403468891978, + 0.0016024301294237375, + 0.0035977044608443975, + 0.020989181473851204, + 0.004481911193579435, + 0.0037446008063852787, + 0.006800670642405748, + 0.013794695027172565, + 0.000830759119708091, + 0.001080669229850173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.3486526310443878 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot is positioned across a deserted asphalt road, its surface gleaming and rain-soaked at night. From the deep background, a pair of brilliant white LED headlights suddenly pierce the oppressive darkness, casting long, dancing beams that illuminate the fine mist hanging in the cold air and the texture of the wet road. A sleek, dark sedan appears, speeding past in a blur of motion. Its sleek, damp undercarriage reflects the fleeting, distorted world around it as its tires plow through the puddles, kicking up a dramatic, fan-shaped spray that momentarily catches the light. As quickly as it appears, the vehicle retreats, its form vanishing into the night, leaving only the intense, vibrant red glow of its twin taillights. These lights paint long, painterly streaks across the black mirror of the road, rapidly shrinking until they become two crimson pinpricks, finally vanishing, plunging the scene back into near-total darkness and a profound, mysterious atmosphere.", + "seed": 22664, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_15/index_151_seed_22664_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06815358235035092, + "vlm_logprob": -0.06815358235035092, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.06815358235035092, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06371960788965225 + ], + "opposite_logprob_tokens": [ + -6.438719749450684 + ], + "logprob_reward": -0.06371960788965225, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.06371960788965225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9466570615768433, + 0.009520688094198704, + 0.006253660656511784, + 0.0040091248229146, + 0.0017053489573299885, + 0.004473838023841381, + 0.007059440948069096, + 0.007790678180754185, + 0.006402348168194294, + 0.006127814296633005 + ], + "credit_map": [ + 0.011422520503401756, + 0.017376050353050232, + 0.002569532487541437, + 0.013421626761555672, + 0.4380737245082855, + 0.43893665075302124, + 0.015096161514520645, + 0.009760755114257336, + 0.0013164299307391047, + 0.001007141312584281, + 0.000825347495265305, + 0.0004601886321324855, + 0.0017854927573353052, + 0.0018522238824516535, + 0.0013421442126855254, + 0.0009317192016169429, + 0.0005340296775102615, + 0.0003422316804062575, + 0.0004044057277496904, + 0.00030303033418022096, + 0.001113972975872457, + 0.0011614662362262607, + 0.0014295254368335009, + 0.0009649981511756778, + 0.0005758380866609514, + 0.00035671330988407135, + 0.00042218563612550497, + 0.00018722578533925116, + 0.00043729774188250303, + 0.0006092984694987535, + 0.0005728751420974731, + 0.0008476904476992786, + 0.00022118096239864826, + 0.00018517280113883317, + 0.00037115937448106706, + 8.174824324669316e-05, + 0.0001641686976654455, + 0.0002693360729608685, + 0.00010840928734978661, + 0.000304173503536731, + 0.00042109121568500996, + 0.0005270546535030007, + 0.00047210213961079717, + 0.00023309257812798023, + 0.0007201403495855629, + 0.001176372286863625, + 0.0003027551283594221, + 0.000621229293756187, + 0.0005155188264325261, + 0.0008834762265905738, + 0.000784328964073211, + 0.00070769595913589, + 0.0013260312844067812, + 0.001266306615434587, + 0.0006972670089453459, + 0.0008788161212578416, + 0.0006905172485858202, + 0.0010368645889684558, + 0.0005534780211746693, + 0.000759354792535305, + 0.0010923943482339382, + 0.0019251022022217512, + 0.0010443022474646568, + 0.0006886640912853181, + 0.0004946478875353932, + 0.0007040832424536347, + 0.0004958190838806331, + 0.00062041130149737, + 0.0012848674086853862, + 0.0012768054148182273, + 0.0007252019131556153, + 0.0008005113922990859, + 0.00047516159247606993, + 0.0007827212102711201, + 0.00040557727334089577, + 0.0005945336306467652, + 0.0009686378762125969, + 0.0011726724915206432, + 0.0010206345468759537, + 0.000707875587977469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.06371960788965225 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010941001819446683 + ], + "opposite_logprob_tokens": [ + -6.876093864440918 + ], + "logprob_reward": -0.0010941001819446683, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010941001819446683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9774916172027588, + 0.00298943929374218, + 0.0018758929800242186, + 0.003679391462355852, + 0.0015975560527294874, + 0.0031028122175484896, + 0.004761931486427784, + 0.0014418949140235782, + 0.0017580239800736308, + 0.001301341224461794 + ], + "credit_map": [ + 0.01173333078622818, + 0.017839625477790833, + 0.0025973606389015913, + 0.013797744177281857, + 0.45240312814712524, + 0.4534982144832611, + 0.015393326058983803, + 0.010228965431451797, + 0.00027301182853989303, + 0.00031468705856241286, + 0.0002358112542424351, + 0.00016356780542992055, + 0.0005081624258309603, + 0.000539630011189729, + 0.0004820016911253333, + 0.0004725674516521394, + 0.00021291262237355113, + 0.00010370558447903022, + 0.00010637398372637108, + 7.017461030045524e-05, + 0.0003297272196505219, + 0.00037103347131051123, + 0.0003771840129047632, + 0.000304781598970294, + 0.0003846020554192364, + 0.0001989475276786834, + 0.0003579843323677778, + 0.00011953046487178653, + 0.0006959026795811951, + 0.0008447798318229616, + 0.0003939035814255476, + 0.0006837411783635616, + 0.00021726028353441507, + 0.00010508636478334665, + 0.00021378025121521205, + 6.852654769318178e-05, + 0.0001787325309123844, + 0.00033462271676398814, + 0.00011736250598914921, + 0.00036218506284058094, + 0.00038743557524867356, + 0.0002706106461118907, + 0.00029156627715565264, + 0.00019558223721105605, + 0.0002807652053888887, + 0.0009283125982619822, + 0.0001725904003251344, + 0.0005759496125392616, + 0.00040556813473813236, + 0.000990995205938816, + 0.0006257128552533686, + 0.0006344730500131845, + 0.0007819596794433892, + 0.0005714237340725958, + 0.00035803637001663446, + 0.0003937631263397634, + 0.00011899158562300727, + 0.00015244766836985946, + 0.00012386686285026371, + 0.00017130631022155285, + 0.00021270560682751238, + 0.00035541903343982995, + 0.00013940448116045445, + 0.00016775353287812322, + 0.00011097572860307992, + 0.0001723304740153253, + 7.994876796146855e-05, + 0.00017265905626118183, + 0.00031868836958892643, + 0.0003345629957038909, + 0.000293027755105868, + 0.00027583102928474545, + 8.634565165266395e-05, + 0.0001119395237765275, + 6.748225860064849e-05, + 0.00011001441453117877, + 0.00026594538940116763, + 0.0003887981874868274, + 0.00011701645416906103, + 0.00015379954129457474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.0010941001819446683 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981209009885788 + ], + "opposite_logprob_tokens": [ + -3.5298120975494385 + ], + "logprob_reward": -0.02981209009885788, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981209009885788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9012603759765625, + 0.0070549738593399525, + 0.007247712463140488, + 0.0052391341887414455, + 0.003156072460114956, + 0.02500787377357483, + 0.02761969156563282, + 0.015766222029924393, + 0.0029961317777633667, + 0.004651873838156462 + ], + "credit_map": [ + 0.010770849883556366, + 0.016457222402095795, + 0.002660841913893819, + 0.012727553956210613, + 0.4172545075416565, + 0.4178309142589569, + 0.013957268558442593, + 0.00960114412009716, + 0.0008190563530661166, + 0.0006526106735691428, + 0.0007315582479350269, + 0.00048356529441662133, + 0.000991547480225563, + 0.0010679695988073945, + 0.001562780700623989, + 0.000745883968193084, + 0.0007076130132190883, + 0.0005189767689444125, + 0.0006288744625635445, + 0.00035184170701541007, + 0.001050492748618126, + 0.0011154345702379942, + 0.001246690982952714, + 0.001627787365578115, + 0.0010896010790020227, + 0.000560963002499193, + 0.0005212219548411667, + 0.00023408833658322692, + 0.000659869285300374, + 0.0007737918640486896, + 0.000492060265969485, + 0.0009075376437976956, + 0.0004998663789592683, + 0.00039341108640655875, + 0.0004050370480399579, + 0.00019220259855501354, + 0.00028547836700454354, + 0.0004766305792145431, + 0.000257207197137177, + 0.0006462386809289455, + 0.002700046868994832, + 0.0022232714109122753, + 0.0026302679907530546, + 0.0012979271123185754, + 0.0032138002570718527, + 0.0055740755051374435, + 0.0019356287084519863, + 0.0054328530095517635, + 0.0022206127177923918, + 0.0034586088731884956, + 0.002320290310308337, + 0.003705869661644101, + 0.005067883990705013, + 0.005149981938302517, + 0.002810710808262229, + 0.002885728608816862, + 0.0011414906475692987, + 0.0017166115576401353, + 0.00101217080373317, + 0.0014940949622541666, + 0.002625159453600645, + 0.0042039076797664165, + 0.0018243469530716538, + 0.0017484378768131137, + 0.00027858398971147835, + 0.000294616911560297, + 0.00026406795950606465, + 0.00038528270670212805, + 0.00035180122358724475, + 0.00046359069528989494, + 0.0005716540035791695, + 0.0003865339094772935, + 0.0003455408732406795, + 0.00041063374374061823, + 0.00025374756660312414, + 0.0004941080114804208, + 0.0009173523285426199, + 0.001141902175731957, + 0.0006382744177244604, + 0.0004503138770814985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.02981209009885788 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0205827746540308 + ], + "opposite_logprob_tokens": [ + -3.895582675933838 + ], + "logprob_reward": -0.0205827746540308, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0205827746540308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8794864416122437, + 0.006249468307942152, + 0.009530112147331238, + 0.004215881694108248, + 0.004654701333492994, + 0.030729055404663086, + 0.036288999021053314, + 0.017239879816770554, + 0.004802930634468794, + 0.006802533753216267 + ], + "credit_map": [ + 0.011033053509891033, + 0.016176432371139526, + 0.0025132743176072836, + 0.012453352101147175, + 0.4066184163093567, + 0.4071410000324249, + 0.013963363133370876, + 0.009587554261088371, + 0.0007194332429207861, + 0.0004759014118462801, + 0.000435489957453683, + 0.00029989780159667134, + 0.0015498101711273193, + 0.0015554074198007584, + 0.0007123969844542444, + 0.000501131231430918, + 0.0011827988782897592, + 0.0008154998067766428, + 0.0007010946865193546, + 0.0005361014627851546, + 0.001434443169273436, + 0.001410791533999145, + 0.001971033401787281, + 0.0014783487422391772, + 0.0007287691696546972, + 0.0003600298659875989, + 0.00038555238279514015, + 0.0001943387760547921, + 0.0006548193632625043, + 0.0007644093711860478, + 0.00042433079215697944, + 0.0007036321330815554, + 0.0005648102960549295, + 0.0003076469583902508, + 0.0005575854447670281, + 0.0002271968696732074, + 0.00041023700032383204, + 0.0008117572288028896, + 0.00041653873631730676, + 0.0013589287409558892, + 0.0037333201617002487, + 0.0027520644944161177, + 0.0031664471607655287, + 0.0017477001529186964, + 0.0029970845207571983, + 0.008234738372266293, + 0.0018971147947013378, + 0.006200586911290884, + 0.0029814511071890593, + 0.0063036405481398106, + 0.0034166513942182064, + 0.003931709099560976, + 0.0073807016015052795, + 0.0064586615189909935, + 0.002624447224661708, + 0.0031917381566017866, + 0.0014565098099410534, + 0.0018828046740964055, + 0.0014046226860955358, + 0.0027448355685919523, + 0.002981138415634632, + 0.002926253480836749, + 0.0019457171438261867, + 0.001897999201901257, + 0.0003576117742341012, + 0.0004240340495016426, + 0.0002892247575800866, + 0.0005471173790283501, + 0.0009203665540553629, + 0.0009795449441298842, + 0.0007674828520976007, + 0.0005175482365302742, + 0.0004530770529527217, + 0.0006430060020647943, + 0.00034424199839122593, + 0.0005891009350307286, + 0.0013187675504013896, + 0.0011944464640691876, + 0.0014431270537897944, + 0.0008167670457623899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0205827746540308 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22555933892726898 + ], + "opposite_logprob_tokens": [ + -1.6005593538284302 + ], + "logprob_reward": -0.22555933892726898, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22555933892726898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45355162024497986, + 0.05136876925826073, + 0.040977027267217636, + 0.017943058162927628, + 0.027548540383577347, + 0.04194483906030655, + 0.15604647994041443, + 0.13471010327339172, + 0.03397471085190773, + 0.04193494841456413 + ], + "credit_map": [ + 0.010156876407563686, + 0.014986653812229633, + 0.007375047542154789, + 0.009119482710957527, + 0.19510364532470703, + 0.19209949672222137, + 0.017012618482112885, + 0.0076977270655334, + 0.006432339549064636, + 0.006573734804987907, + 0.0041469004936516285, + 0.0035839290358126163, + 0.00764443539083004, + 0.00754678575322032, + 0.009025170467793941, + 0.0064154695719480515, + 0.004943483974784613, + 0.0031959505286067724, + 0.003216278739273548, + 0.001922368654049933, + 0.006705914624035358, + 0.006760211195796728, + 0.007429188583046198, + 0.006803626660257578, + 0.003175199730321765, + 0.0020437943749129772, + 0.0015691515291109681, + 0.0013752946397289634, + 0.0021918867714703083, + 0.002688836073502898, + 0.0019434146815910935, + 0.002955478150397539, + 0.004504441749304533, + 0.0035090807359665632, + 0.0030916850082576275, + 0.0016364791663363576, + 0.0024429121986031532, + 0.004680597689002752, + 0.00205242820084095, + 0.00563090993091464, + 0.004713528323918581, + 0.004900441970676184, + 0.003921931609511375, + 0.0028882892802357674, + 0.006467301398515701, + 0.006127548404037952, + 0.0036347992718219757, + 0.009290995076298714, + 0.012251846492290497, + 0.02909448929131031, + 0.014718360267579556, + 0.014893236570060253, + 0.027978502213954926, + 0.02781210094690323, + 0.013330521993339062, + 0.015967397019267082, + 0.008348122239112854, + 0.012041203677654266, + 0.007980002090334892, + 0.011864135041832924, + 0.02125730738043785, + 0.03976576030254364, + 0.017032232135534286, + 0.016421334818005562, + 0.0031519688200205564, + 0.003951332066208124, + 0.001946248346939683, + 0.0034726413432508707, + 0.0043996786698699, + 0.006043025758117437, + 0.0066226776689291, + 0.004387135151773691, + 0.003174746409058571, + 0.004420000594109297, + 0.0021803034469485283, + 0.004463884513825178, + 0.007017851807177067, + 0.00957847386598587, + 0.006068792659789324, + 0.005030890926718712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.22555933892726898 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot is positioned across a deserted asphalt road, its surface gleaming and rain-soaked at night. From the deep background, a pair of brilliant white LED headlights suddenly pierce the oppressive darkness, casting long, dancing beams that illuminate the fine mist hanging in the cold air and the texture of the wet road. A sleek, dark sedan appears, speeding past in a blur of motion. Its sleek, damp undercarriage reflects the fleeting, distorted world around it as its tires plow through the puddles, kicking up a dramatic, fan-shaped spray that momentarily catches the light. As quickly as it appears, the vehicle retreats, its form vanishing into the night, leaving only the intense, vibrant red glow of its twin taillights. These lights paint long, painterly streaks across the black mirror of the road, rapidly shrinking until they become two crimson pinpricks, finally vanishing, plunging the scene back into near-total darkness and a profound, mysterious atmosphere.", + "seed": 22665, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_15/index_151_seed_22665_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09333545458503067, + "vlm_logprob": -0.09333545458503067, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.09333545458503067, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0589686818420887 + ], + "opposite_logprob_tokens": [ + -5.558968544006348 + ], + "logprob_reward": -0.0589686818420887, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0589686818420887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12759973108768463, + 0.0980077013373375, + 0.06200379505753517, + 0.05853039771318436, + 0.09717831015586853, + 0.20501670241355896, + 0.08681310713291168, + 0.07043317705392838, + 0.10833828151226044, + 0.08607883751392365 + ], + "credit_map": [ + 0.014045553281903267, + 0.012215721420943737, + 0.013707716017961502, + 0.015460485592484474, + 0.016969922930002213, + 0.012209202162921429, + 0.01981036551296711, + 0.02318074367940426, + 0.012394895777106285, + 0.008878721855580807, + 0.014989867806434631, + 0.010294903069734573, + 0.009699315764009953, + 0.008513708598911762, + 0.012842086143791676, + 0.020394204184412956, + 0.010549575090408325, + 0.009490353055298328, + 0.013982507400214672, + 0.0063618808053433895, + 0.0071687279269099236, + 0.0056342813186347485, + 0.005620436277240515, + 0.0031960285268723965, + 0.004971580114215612, + 0.004572397097945213, + 0.011204452253878117, + 0.008502776734530926, + 0.010049184784293175, + 0.006385439075529575, + 0.007423423696309328, + 0.0054211425594985485, + 0.007577665150165558, + 0.010817078873515129, + 0.017541252076625824, + 0.012230115942656994, + 0.01335020363330841, + 0.01579924300312996, + 0.011093885637819767, + 0.008768858388066292, + 0.008001246489584446, + 0.008239195682108402, + 0.01254738587886095, + 0.01069968193769455, + 0.07780303806066513, + 0.07512476295232773, + 0.006697984412312508, + 0.0059033953584730625, + 0.007200885098427534, + 0.009068166837096214, + 0.015071562491357327, + 0.013392897322773933, + 0.012248707935214043, + 0.01301440130919218, + 0.009605140425264835, + 0.0072113401256501675, + 0.004993561189621687, + 0.006561261136084795, + 0.007243471220135689, + 0.00861385278403759, + 0.017737675458192825, + 0.01653766818344593, + 0.005047318991273642, + 0.0036983639001846313, + 0.006908714771270752, + 0.01378270611166954, + 0.013182337395846844, + 0.01610148698091507, + 0.025626476854085922, + 0.019680766388773918, + 0.007704973220825195, + 0.005350810009986162, + 0.0065463208593428135, + 0.01143959816545248, + 0.011512096039950848, + 0.016098108142614365, + 0.015802746638655663, + 0.015206532552838326, + 0.005738651845604181, + 0.003734779544174671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.0589686818420887 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017803546506911516 + ], + "opposite_logprob_tokens": [ + -6.3767805099487305 + ], + "logprob_reward": -0.0017803546506911516, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017803546506911516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04960726946592331, + 0.04037220776081085, + 0.03108620084822178, + 0.048708342015743256, + 0.03574589267373085, + 0.6482090353965759, + 0.06849842518568039, + 0.033355534076690674, + 0.02332441508769989, + 0.02109268680214882 + ], + "credit_map": [ + 0.006695653311908245, + 0.006831793580204248, + 0.006539138033986092, + 0.010446789674460888, + 0.004977363161742687, + 0.004724202211946249, + 0.004290519282221794, + 0.005101808812469244, + 0.00449019018560648, + 0.0042733787558972836, + 0.0048974682576954365, + 0.005412229336798191, + 0.003800786565989256, + 0.004486900754272938, + 0.004759055562317371, + 0.008252193219959736, + 0.005626042373478413, + 0.003125227754935622, + 0.00667556282132864, + 0.0024053289089351892, + 0.0029426938854157925, + 0.0035662155132740736, + 0.00384965050034225, + 0.0028954760637134314, + 0.006210506893694401, + 0.0041123745031654835, + 0.008436009287834167, + 0.004584210924804211, + 0.010789405554533005, + 0.004979389254003763, + 0.005461995489895344, + 0.004134443122893572, + 0.0028840703889727592, + 0.003742842236533761, + 0.0049382890574634075, + 0.003950275015085936, + 0.00532315531745553, + 0.008101504296064377, + 0.003856804920360446, + 0.002948947949334979, + 0.0156153729185462, + 0.011138994246721268, + 0.019007157534360886, + 0.010176634415984154, + 0.29034820199012756, + 0.28324177861213684, + 0.006203307304531336, + 0.012477504089474678, + 0.00565079553052783, + 0.0076564303599298, + 0.010626237839460373, + 0.007661906536668539, + 0.012127210386097431, + 0.01200069859623909, + 0.007739232853055, + 0.005035904701799154, + 0.0023437240161001682, + 0.002786870114505291, + 0.0036837507504969835, + 0.008875763043761253, + 0.0054915184155106544, + 0.005582723300904036, + 0.0025048411916941404, + 0.0020863409154117107, + 0.0017630349611863494, + 0.002975469222292304, + 0.0025453181006014347, + 0.002742015291005373, + 0.005943514406681061, + 0.004680116660892963, + 0.0013249489711597562, + 0.0013499944470822811, + 0.0017160201678052545, + 0.002561488188803196, + 0.0035875854082405567, + 0.001958675915375352, + 0.004447600804269314, + 0.003517724107950926, + 0.0018161024199798703, + 0.0014874886255711317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.0017803546506911516 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043067965656518936 + ], + "opposite_logprob_tokens": [ + -3.1680679321289062 + ], + "logprob_reward": -0.043067965656518936, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043067965656518936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03638564050197601, + 0.04193313419818878, + 0.0399983786046505, + 0.02336668223142624, + 0.02299426682293415, + 0.7280418872833252, + 0.04272555932402611, + 0.022335873916745186, + 0.016577590256929398, + 0.025641007348895073 + ], + "credit_map": [ + 0.004880876280367374, + 0.0031395673286169767, + 0.0063601271249353886, + 0.0045618582516908646, + 0.005214440170675516, + 0.0035903218667954206, + 0.0046018315479159355, + 0.004036616068333387, + 0.004256183747202158, + 0.005187302362173796, + 0.004229807294905186, + 0.00754203088581562, + 0.004952490795403719, + 0.0048112948425114155, + 0.005953559186309576, + 0.005000461358577013, + 0.005090881139039993, + 0.004123966675251722, + 0.006429147906601429, + 0.0022223838604986668, + 0.0058514876291155815, + 0.006730637513101101, + 0.0063586048781871796, + 0.003191267838701606, + 0.002091967035084963, + 0.002027201931923628, + 0.003361551556736231, + 0.0036833069752901793, + 0.005087592639029026, + 0.0019478293834254146, + 0.002724928315728903, + 0.002442304976284504, + 0.0018174267606809735, + 0.0027714394964277744, + 0.004436796065419912, + 0.00281048403121531, + 0.0033705534879118204, + 0.002986188745126128, + 0.002299491548910737, + 0.00250188703648746, + 0.016366178169846535, + 0.010702097788453102, + 0.02036059834063053, + 0.009658015333116055, + 0.3305813670158386, + 0.32119399309158325, + 0.005100834183394909, + 0.014078770764172077, + 0.003051126142963767, + 0.005123523995280266, + 0.00447699474170804, + 0.0060701859183609486, + 0.009885787963867188, + 0.009778707288205624, + 0.002472574356943369, + 0.0018666626419872046, + 0.001260037999600172, + 0.0023763554636389017, + 0.002244708826765418, + 0.0038449610583484173, + 0.005431707948446274, + 0.005161723587661982, + 0.0009916669223457575, + 0.0010247109457850456, + 0.0011261432664468884, + 0.0023741985205560923, + 0.0030614836141467094, + 0.0027108301874250174, + 0.0027963980101048946, + 0.002406545216217637, + 0.0010650909971445799, + 0.0010369002120569348, + 0.0018968168878927827, + 0.003649207530543208, + 0.003618604503571987, + 0.003907742910087109, + 0.004681836348026991, + 0.005191755015403032, + 0.001394299091771245, + 0.0013007433153688908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.043067965656518936 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419729832559824 + ], + "opposite_logprob_tokens": [ + -4.26419734954834 + ], + "logprob_reward": -0.01419729832559824, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419729832559824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02908610738813877, + 0.021912503987550735, + 0.023123152554035187, + 0.016939599066972733, + 0.04729050025343895, + 0.8151988387107849, + 0.01686171628534794, + 0.009197011590003967, + 0.005874600727111101, + 0.014515995979309082 + ], + "credit_map": [ + 0.0036197074223309755, + 0.0026543124113231897, + 0.003611448686569929, + 0.004087687004357576, + 0.003305939259007573, + 0.0031107815448194742, + 0.0043827835470438, + 0.004313448444008827, + 0.0030676615424454212, + 0.0030023271683603525, + 0.0025776990223675966, + 0.003758131293579936, + 0.0023061258252710104, + 0.0021541789174079895, + 0.002865006448701024, + 0.0021813728380948305, + 0.004253347869962454, + 0.0022860567551106215, + 0.004918024875223637, + 0.0036516967229545116, + 0.0018131863325834274, + 0.0021931694354861975, + 0.002459888346493244, + 0.001547781634144485, + 0.0014980389969423413, + 0.0018345407443121076, + 0.0031138595659285784, + 0.001735416240990162, + 0.003796943696215749, + 0.00144992268178612, + 0.0016002401243895292, + 0.0019106377149000764, + 0.0038089710287749767, + 0.002807397162541747, + 0.00783574115484953, + 0.002718982519581914, + 0.008357584476470947, + 0.012733767740428448, + 0.005074999760836363, + 0.003953058272600174, + 0.018157627433538437, + 0.011905812658369541, + 0.0223828312009573, + 0.010556025430560112, + 0.3706444501876831, + 0.3597767949104309, + 0.006014956161379814, + 0.01576034165918827, + 0.0011629906948655844, + 0.0018730319570749998, + 0.001994350226595998, + 0.0009986208751797676, + 0.004976753611117601, + 0.004249265417456627, + 0.0008473142515867949, + 0.0007593903574161232, + 0.0006410342757590115, + 0.0009114866843447089, + 0.0010859949979931116, + 0.002513566520065069, + 0.0014318778412416577, + 0.0014870795421302319, + 0.0005336085450835526, + 0.0005923630087636411, + 0.00042705837404355407, + 0.000898607075214386, + 0.0007516462937928736, + 0.0011549170594662428, + 0.0008643284672871232, + 0.0009264040854759514, + 0.0004289160424377769, + 0.0004227232129778713, + 0.001307858619838953, + 0.0022638377267867327, + 0.002078003715723753, + 0.002194114262238145, + 0.0024431359488517046, + 0.0027390632312744856, + 0.0008376107434742153, + 0.000652372429613024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.01419729832559824 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking across a deserted asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the road surface rain-soaked, gleaming, and reflecting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dark sedan speed past the camera, kicking up a fan-shaped spray of water from puddles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the vehicle's taillights leave long, painterly red streaks on the road before shrinking to pinpricks and vanishing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34866297245025635 + ], + "opposite_logprob_tokens": [ + -1.2236629724502563 + ], + "logprob_reward": -0.34866297245025635, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34866297245025635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030778178945183754, + 0.04222830384969711, + 0.026755746454000473, + 0.09973783046007156, + 0.055780038237571716, + 0.5274621248245239, + 0.09626566618680954, + 0.03359587863087654, + 0.043932490050792694, + 0.04346374422311783 + ], + "credit_map": [ + 0.004968629218637943, + 0.0025050684344023466, + 0.0031494805589318275, + 0.002913379343226552, + 0.004687505774199963, + 0.004128249827772379, + 0.0035698257852345705, + 0.0048560393042862415, + 0.004606528673321009, + 0.004726131912320852, + 0.004896169062703848, + 0.008084743283689022, + 0.00569235859438777, + 0.004495212342590094, + 0.005532986484467983, + 0.004194170236587524, + 0.003502586157992482, + 0.0028751250356435776, + 0.004235606174916029, + 0.0033946766052395105, + 0.0026993434876203537, + 0.0029937801882624626, + 0.004013717174530029, + 0.0030409106984734535, + 0.009158949367702007, + 0.008223749697208405, + 0.01589517667889595, + 0.015828965231776237, + 0.023275738582015038, + 0.009647534228861332, + 0.009639565832912922, + 0.008068149909377098, + 0.004082034807652235, + 0.005215429700911045, + 0.0081279706209898, + 0.0053113168105483055, + 0.00910797156393528, + 0.011199179105460644, + 0.006617540959268808, + 0.00611859280616045, + 0.012378204613924026, + 0.010623403824865818, + 0.01694975048303604, + 0.007923050783574581, + 0.23494908213615417, + 0.22763921320438385, + 0.005900816526263952, + 0.011098586022853851, + 0.0055181244388222694, + 0.008512431755661964, + 0.007992764003574848, + 0.008642210625112057, + 0.02909999154508114, + 0.027283908799290657, + 0.004908246919512749, + 0.004307989496737719, + 0.002304613124579191, + 0.002570775570347905, + 0.0037535664159804583, + 0.007056377828121185, + 0.007291377987712622, + 0.007126978132873774, + 0.0019725668244063854, + 0.001519623794592917, + 0.0028829907532781363, + 0.005569624248892069, + 0.004396309610456228, + 0.0026656531263142824, + 0.013222002424299717, + 0.010707634501159191, + 0.00274398410692811, + 0.0017442903481423855, + 0.0034838856663554907, + 0.0060134511440992355, + 0.005173773970454931, + 0.00754779577255249, + 0.007956361398100853, + 0.009043407626450062, + 0.0023721724282950163, + 0.0018728956347331405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.34866297245025635 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_2/index_tensor([8525, 8525])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_2/index_tensor([8525, 8525])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5333513b57938d87b1f0e0a220d97acb7f5f2171 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_2/index_tensor([8525, 8525])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fixed camera pans sharply to the right, capturing a striking medium shot from the side, closely following a general clad in silver armor. He rides atop a towering, pure white warhorse, its powerful steed surging at full speed, its muscular mass undulating with every stride, its long white mane and tail whipping wildly in the wind. The general wears a suit of intricately crafted, mirror-polished silver plate armor, etched with reliefs of mythical beasts, a piercing gleam in the cold light scattered by the overcast sky. He leans forward, clinging to the ornate leather saddle, a crimson cape billowing behind him like a blaze. Through the slits of his helmet's visor, adorned with an equally crimson plume, one can glimpse his taut jaw and intensely focused gaze, fixed on an enemy just beyond the frame. One hand firmly grips a long spear, its tip gleaming coldly, held nearly parallel to the ground, poised for discharge. The other grips the reins, steering his mount with pinpoint precision. The background is a vast, desolate battlefield. The camera's rapid movement blurs the mud, gravel, and wreckage of the battlefield into a dynamic, brown-gray stripe. Only the splatter of earth beneath the horse's hooves is clearly visible, imbued with an epic sense of urgency and the relentless momentum of pursuit.", + "seed": 27800, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_2/index_8525_seed_27800_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11916684778407216, + "vlm_logprob": -0.11916684778407216, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.11916684778407216, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03088817372918129 + ], + "opposite_logprob_tokens": [ + -3.780888080596924 + ], + "logprob_reward": -0.03088817372918129, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03088817372918129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08753907680511475, + 0.05268183723092079, + 0.021776538342237473, + 0.06313122808933258, + 0.18653278052806854, + 0.07595811039209366, + 0.2197190374135971, + 0.2168167233467102, + 0.02114233747124672, + 0.054702337831258774 + ], + "credit_map": [ + 0.01339650247246027, + 0.01822921633720398, + 0.011817065067589283, + 0.007855808362364769, + 0.010244801640510559, + 0.014429842121899128, + 0.006350083742290735, + 0.005215766374021769, + 0.005170334596186876, + 0.0067130704410374165, + 0.00516589405015111, + 0.005391648039221764, + 0.010520885698497295, + 0.005534884985536337, + 0.008488811552524567, + 0.005696308799088001, + 0.002725550439208746, + 0.0027199285104870796, + 0.0024373780470341444, + 0.0019912542775273323, + 0.005255552940070629, + 0.0029213109519332647, + 0.002034831792116165, + 0.0016907332465052605, + 0.007879513315856457, + 0.007158390246331692, + 0.006559798959642649, + 0.012651017867028713, + 0.008354635909199715, + 0.00719763757660985, + 0.006832592189311981, + 0.006497645750641823, + 0.05897410586476326, + 0.06025565043091774, + 0.014194320887327194, + 0.016731444746255875, + 0.011446626856923103, + 0.011997421272099018, + 0.00749126635491848, + 0.005441946443170309, + 0.009643908590078354, + 0.0098344087600708, + 0.007835811004042625, + 0.0058630649000406265, + 0.01728419028222561, + 0.0070867715403437614, + 0.009205865673720837, + 0.00920408871024847, + 0.03196260705590248, + 0.02333693578839302, + 0.029458021745085716, + 0.025392308831214905, + 0.032329343259334564, + 0.0216513741761446, + 0.02558889426290989, + 0.029999537393450737, + 0.02064274437725544, + 0.024090144783258438, + 0.030119549483060837, + 0.02381914108991623, + 0.04338641092181206, + 0.026694636791944504, + 0.029267573729157448, + 0.0187965240329504, + 0.002796862041577697, + 0.0026572325732558966, + 0.002036956837400794, + 0.001594405621290207, + 0.0036531889345496893, + 0.0030829606112092733, + 0.0029900886584073305, + 0.0023306412622332573, + 0.007503268774598837, + 0.006064720451831818, + 0.007432129234075546, + 0.005462081637233496, + 0.009422799572348595, + 0.006844050716608763, + 0.007865306921303272, + 0.00410798192024231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.03088817372918129 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31362175941467285 + ], + "opposite_logprob_tokens": [ + -1.3136217594146729 + ], + "logprob_reward": -0.31362175941467285, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31362175941467285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1233297735452652, + 0.03478136286139488, + 0.02363288216292858, + 0.029963182285428047, + 0.4214875102043152, + 0.06504496932029724, + 0.11737244576215744, + 0.08316471427679062, + 0.06289303302764893, + 0.03833014518022537 + ], + "credit_map": [ + 0.014584927819669247, + 0.01673639379441738, + 0.021255355328321457, + 0.015527758747339249, + 0.01603972539305687, + 0.020474301651120186, + 0.011077332310378551, + 0.0076339892111718655, + 0.0036936639808118343, + 0.00371447647921741, + 0.004257288295775652, + 0.005122773349285126, + 0.00605366425588727, + 0.004919465631246567, + 0.004435115959495306, + 0.0025849188677966595, + 0.0028944499790668488, + 0.0032132426276803017, + 0.0031920347828418016, + 0.0030785815324634314, + 0.004234121646732092, + 0.0028883556369692087, + 0.0020674592815339565, + 0.0020646387711167336, + 0.0037326558958739042, + 0.004424284677952528, + 0.002479909686371684, + 0.004293906502425671, + 0.003994662780314684, + 0.006165369413793087, + 0.0023804574739187956, + 0.0024919374845921993, + 0.15572458505630493, + 0.16396304965019226, + 0.02339213527739048, + 0.014936483465135098, + 0.01775297522544861, + 0.02814488671720028, + 0.010549570433795452, + 0.0070238481275737286, + 0.011854981072247028, + 0.009250955656170845, + 0.007140613626688719, + 0.006935445126146078, + 0.011450231075286865, + 0.006373882759362459, + 0.0057806046679615974, + 0.006258258130401373, + 0.02092503011226654, + 0.02702179178595543, + 0.01274901907891035, + 0.012170040979981422, + 0.014258368872106075, + 0.010656862519681454, + 0.009986800141632557, + 0.009604542516171932, + 0.007921207696199417, + 0.009242375381290913, + 0.00923397857695818, + 0.006208172533661127, + 0.01869458518922329, + 0.011240475811064243, + 0.01205921359360218, + 0.008564713411033154, + 0.008549998514354229, + 0.012375094927847385, + 0.007418866269290447, + 0.007210633251816034, + 0.006146206054836512, + 0.011449826881289482, + 0.005836517084389925, + 0.003905894700437784, + 0.004197807051241398, + 0.005929715000092983, + 0.0037073157727718353, + 0.003911075182259083, + 0.005095897242426872, + 0.005524021107703447, + 0.00543970987200737, + 0.004524610936641693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.31362175941467285 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023296110332012177 + ], + "opposite_logprob_tokens": [ + -3.7732961177825928 + ], + "logprob_reward": -0.023296110332012177, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023296110332012177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15409642457962036, + 0.040190111845731735, + 0.06665287911891937, + 0.07922062277793884, + 0.44851818680763245, + 0.039683133363723755, + 0.05610780045390129, + 0.034028321504592896, + 0.03889715299010277, + 0.042605407536029816 + ], + "credit_map": [ + 0.014989308081567287, + 0.024561043828725815, + 0.0214651208370924, + 0.01951281726360321, + 0.024836061522364616, + 0.021561160683631897, + 0.015884267166256905, + 0.011286642402410507, + 0.0064409635961055756, + 0.004910922143608332, + 0.005195684265345335, + 0.004387951921671629, + 0.00411398196592927, + 0.006806212477385998, + 0.005195213481783867, + 0.0031391791999340057, + 0.005528450943529606, + 0.008261959068477154, + 0.010036606341600418, + 0.016549738124012947, + 0.006808396428823471, + 0.006799786351621151, + 0.008624408394098282, + 0.004043536260724068, + 0.016638364642858505, + 0.010366840288043022, + 0.012285199947655201, + 0.0085736820474267, + 0.008224462158977985, + 0.010269833728671074, + 0.006213657092303038, + 0.006648581009358168, + 0.1697300523519516, + 0.17670677602291107, + 0.023622117936611176, + 0.015521359629929066, + 0.018172306939959526, + 0.02827559784054756, + 0.009515797719359398, + 0.006974147632718086, + 0.0059340354055166245, + 0.0055973161943256855, + 0.0062804389744997025, + 0.004872262477874756, + 0.006283637136220932, + 0.004781342577189207, + 0.002643651096150279, + 0.003290447872132063, + 0.013780278153717518, + 0.01726943626999855, + 0.00397050054743886, + 0.005233777221292257, + 0.0046364180743694305, + 0.004927258472889662, + 0.003709308570250869, + 0.002580823376774788, + 0.0040396880358457565, + 0.003784158267080784, + 0.0048410939052701, + 0.004220207221806049, + 0.00523449806496501, + 0.0033086712937802076, + 0.005762004293501377, + 0.002837996231392026, + 0.0046323868446052074, + 0.007056131027638912, + 0.004204185679554939, + 0.004563226830214262, + 0.004851922392845154, + 0.006315405946224928, + 0.004630471114069223, + 0.002643421059474349, + 0.005239482969045639, + 0.007343757897615433, + 0.0056033628061413765, + 0.00544682377949357, + 0.0045080604031682014, + 0.00527540035545826, + 0.0045931218191981316, + 0.004595400299876928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.023296110332012177 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024877809919416904 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024877809919416904, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024877809919416904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17036128044128418, + 0.07511439174413681, + 0.0421784482896328, + 0.11156647652387619, + 0.32538047432899475, + 0.02328481897711754, + 0.04159944877028465, + 0.07879802584648132, + 0.07016095519065857, + 0.061555732041597366 + ], + "credit_map": [ + 0.027800558134913445, + 0.030789731070399284, + 0.020706070587038994, + 0.013305221684277058, + 0.027574658393859863, + 0.02726639248430729, + 0.013929387554526329, + 0.008989267982542515, + 0.007027270272374153, + 0.008857548236846924, + 0.008317654952406883, + 0.008245008997619152, + 0.013535975478589535, + 0.008658123202621937, + 0.011472931131720543, + 0.008999872021377087, + 0.00491934921592474, + 0.005833050701767206, + 0.004631009418517351, + 0.004872075282037258, + 0.009096098132431507, + 0.005462619010359049, + 0.0027771126478910446, + 0.004587132949382067, + 0.011547750793397427, + 0.011338978074491024, + 0.012714201584458351, + 0.027951035648584366, + 0.012694748118519783, + 0.009518736973404884, + 0.011948809027671814, + 0.013852214440703392, + 0.11617451161146164, + 0.12202252447605133, + 0.019328119233250618, + 0.018252449110150337, + 0.013927887193858624, + 0.021165257319808006, + 0.009077770635485649, + 0.005431984551250935, + 0.003004347672685981, + 0.0028681079857051373, + 0.003344576107338071, + 0.0041262744925916195, + 0.002655842574313283, + 0.0023490972816944122, + 0.0019685227889567614, + 0.0029680500738322735, + 0.005143468733876944, + 0.0064750490710139275, + 0.004532902501523495, + 0.006151692010462284, + 0.006799196824431419, + 0.00392539706081152, + 0.005667438730597496, + 0.002904305001720786, + 0.006360894534736872, + 0.00848457496613264, + 0.0086368964985013, + 0.006607447750866413, + 0.018000992015004158, + 0.010767658241093159, + 0.013152818195521832, + 0.006786737125366926, + 0.009606758132576942, + 0.00897503737360239, + 0.007680061273276806, + 0.008532176725566387, + 0.007393702398985624, + 0.01149177085608244, + 0.007879259064793587, + 0.008602190762758255, + 0.006419042125344276, + 0.00799470953643322, + 0.0058677541092038155, + 0.0070147146470844746, + 0.007571571972221136, + 0.00919916108250618, + 0.008292418904602528, + 0.009196355938911438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0024877809919416904 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255404144525528 + ], + "opposite_logprob_tokens": [ + -1.6005403995513916 + ], + "logprob_reward": -0.2255404144525528, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255404144525528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10997550934553146, + 0.10239648073911667, + 0.042960941791534424, + 0.13493123650550842, + 0.25692692399024963, + 0.032790590077638626, + 0.17676816880702972, + 0.05269822105765343, + 0.06035253033041954, + 0.030199378728866577 + ], + "credit_map": [ + 0.016389641910791397, + 0.01575305126607418, + 0.019893160089850426, + 0.013756400905549526, + 0.01532555278390646, + 0.01239264477044344, + 0.008483372628688812, + 0.007981687784194946, + 0.019191410392522812, + 0.01485588401556015, + 0.009120981208980083, + 0.011992999352514744, + 0.009354976937174797, + 0.020145250484347343, + 0.009316771291196346, + 0.008418207988142967, + 0.0044340332970023155, + 0.005563986487686634, + 0.008637077175080776, + 0.007221348583698273, + 0.0055274139158427715, + 0.004948664456605911, + 0.0038191317580640316, + 0.0028092877473682165, + 0.016494503244757652, + 0.01385632622987032, + 0.016232939437031746, + 0.031489938497543335, + 0.013438978232443333, + 0.011535794474184513, + 0.014562450349330902, + 0.017320312559604645, + 0.09491808712482452, + 0.09686527401208878, + 0.013105479069054127, + 0.012125917710363865, + 0.010897158645093441, + 0.01666383258998394, + 0.006488700397312641, + 0.0058624655939638615, + 0.003596086287871003, + 0.0040603200905025005, + 0.004038158804178238, + 0.004097627475857735, + 0.006540128495544195, + 0.003177328733727336, + 0.004330062307417393, + 0.002950879279524088, + 0.04765336588025093, + 0.05919189006090164, + 0.01388533879071474, + 0.01444427389651537, + 0.01130290050059557, + 0.015845822170376778, + 0.0077228485606610775, + 0.00672173174098134, + 0.0073486994951963425, + 0.006593802943825722, + 0.007765371818095446, + 0.005860291421413422, + 0.008820703253149986, + 0.006219985894858837, + 0.006349965929985046, + 0.003739404957741499, + 0.009667307138442993, + 0.007825474254786968, + 0.004130111075937748, + 0.0071624647825956345, + 0.009044855833053589, + 0.007463743910193443, + 0.008819976821541786, + 0.006238598842173815, + 0.0046197096817195415, + 0.004152469802647829, + 0.003560430370271206, + 0.002390471054241061, + 0.004772159270942211, + 0.004179792013019323, + 0.003377180313691497, + 0.003147169016301632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.2255404144525528 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed camera pans sharply to the right, capturing a striking medium shot from the side, closely following a general clad in silver armor. He rides atop a towering, pure white warhorse, its powerful steed surging at full speed, its muscular mass undulating with every stride, its long white mane and tail whipping wildly in the wind. The general wears a suit of intricately crafted, mirror-polished silver plate armor, etched with reliefs of mythical beasts, a piercing gleam in the cold light scattered by the overcast sky. He leans forward, clinging to the ornate leather saddle, a crimson cape billowing behind him like a blaze. Through the slits of his helmet's visor, adorned with an equally crimson plume, one can glimpse his taut jaw and intensely focused gaze, fixed on an enemy just beyond the frame. One hand firmly grips a long spear, its tip gleaming coldly, held nearly parallel to the ground, poised for discharge. The other grips the reins, steering his mount with pinpoint precision. The background is a vast, desolate battlefield. The camera's rapid movement blurs the mud, gravel, and wreckage of the battlefield into a dynamic, brown-gray stripe. Only the splatter of earth beneath the horse's hooves is clearly visible, imbued with an epic sense of urgency and the relentless momentum of pursuit.", + "seed": 27801, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_2/index_8525_seed_27801_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11109748538583517, + "vlm_logprob": -0.11109748538583517, + "vlm_yes_no_margin": 3.099999988079071, + "vlm_token_credit": -0.11109748538583517, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09983035922050476 + ], + "opposite_logprob_tokens": [ + -2.474830389022827 + ], + "logprob_reward": -0.09983035922050476, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09983035922050476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5324591398239136, + 0.09659332782030106, + 0.036557864397764206, + 0.07566167414188385, + 0.037770189344882965, + 0.025657564401626587, + 0.03189334273338318, + 0.11101468652486801, + 0.03397715091705322, + 0.01841508038341999 + ], + "credit_map": [ + 0.15492062270641327, + 0.01882638782262802, + 0.02660384215414524, + 0.009937156923115253, + 0.21982420980930328, + 0.04097369685769081, + 0.01927395910024643, + 0.04209917038679123, + 0.009768908843398094, + 0.013240193948149681, + 0.00925003457814455, + 0.013064529746770859, + 0.018751652911305428, + 0.012549961917102337, + 0.012704379856586456, + 0.007263659965246916, + 0.004296962637454271, + 0.0035737191792577505, + 0.004821857903152704, + 0.0031985256355255842, + 0.004814470652490854, + 0.0038922918029129505, + 0.0064369868487119675, + 0.005523044615983963, + 0.008672437630593777, + 0.013749165460467339, + 0.007752370089292526, + 0.009054357185959816, + 0.008144446648657322, + 0.009583543986082077, + 0.006777775473892689, + 0.01192756649106741, + 0.0042539723217487335, + 0.005325412377715111, + 0.0030207003001123667, + 0.003335477551445365, + 0.006949641276150942, + 0.005461432039737701, + 0.005400558467954397, + 0.004022992216050625, + 0.00440639303997159, + 0.005111627280712128, + 0.0022074244916439056, + 0.002124490449205041, + 0.0033236618619412184, + 0.0034292431082576513, + 0.002057991223409772, + 0.002996730152517557, + 0.004127189517021179, + 0.0030535124242305756, + 0.003584515769034624, + 0.002064656699076295, + 0.006403883453458548, + 0.004927759058773518, + 0.0033788869623094797, + 0.004352935124188662, + 0.005162250250577927, + 0.0046576703898608685, + 0.0017699216259643435, + 0.0025110573042184114, + 0.04412953928112984, + 0.044483110308647156, + 0.003177017206326127, + 0.00512411305680871, + 0.0048651620745658875, + 0.003412030404433608, + 0.0037050177343189716, + 0.003847538260743022, + 0.0059056757017970085, + 0.004700968042016029, + 0.0029598232358694077, + 0.004580933600664139, + 0.0024080558214336634, + 0.0025299396365880966, + 0.003333392320200801, + 0.0021891130600124598, + 0.0031244405545294285, + 0.002174785127863288, + 0.0015827602474018931, + 0.0010725936153903604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.09983035922050476 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3488307595252991 + ], + "opposite_logprob_tokens": [ + -1.2238306999206543 + ], + "logprob_reward": -0.3488307595252991, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3488307595252991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18257859349250793, + 0.1965988427400589, + 0.19206298887729645, + 0.0909523218870163, + 0.054830554872751236, + 0.047571439296007156, + 0.041394785046577454, + 0.050173964351415634, + 0.08884934335947037, + 0.05498713254928589 + ], + "credit_map": [ + 0.03613269329071045, + 0.011980338953435421, + 0.024279987439513206, + 0.009819797240197659, + 0.03913511335849762, + 0.026313884183764458, + 0.01579001173377037, + 0.019126789644360542, + 0.01850610226392746, + 0.030476486310362816, + 0.017340434715151787, + 0.01917657069861889, + 0.0432884618639946, + 0.02997920662164688, + 0.024980755522847176, + 0.012850837782025337, + 0.01226105261594057, + 0.01728799007833004, + 0.030743436887860298, + 0.02566380798816681, + 0.02316107228398323, + 0.025145187973976135, + 0.027659349143505096, + 0.030141104012727737, + 0.009250100702047348, + 0.01678851805627346, + 0.011304808780550957, + 0.01368901040405035, + 0.00977602694183588, + 0.00963597558438778, + 0.0107790008187294, + 0.009728879667818546, + 0.004648403264582157, + 0.006241280119866133, + 0.005339098162949085, + 0.005302960053086281, + 0.009728435426950455, + 0.0077743553556501865, + 0.009115184657275677, + 0.006680837366729975, + 0.004782196134328842, + 0.01140084769576788, + 0.002861376153305173, + 0.0033211808186024427, + 0.0078078992664813995, + 0.007511601783335209, + 0.005346117075532675, + 0.00454022316262126, + 0.006206348072737455, + 0.006039129104465246, + 0.004366899840533733, + 0.0034679584205150604, + 0.006728974170982838, + 0.005072754807770252, + 0.003592636901885271, + 0.005920086055994034, + 0.005609416402876377, + 0.004900143947452307, + 0.002157286275178194, + 0.0030790017917752266, + 0.012918018735945225, + 0.011947051621973515, + 0.004234427120536566, + 0.0053286184556782246, + 0.012165226973593235, + 0.010975346900522709, + 0.008875622414052486, + 0.011271212249994278, + 0.01071570161730051, + 0.011950785294175148, + 0.00793218333274126, + 0.014963267371058464, + 0.006359301041811705, + 0.006880366243422031, + 0.010763544589281082, + 0.005013434682041407, + 0.009506452828645706, + 0.006731386296451092, + 0.004999108612537384, + 0.004733535926789045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.3488307595252991 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202351681888103 + ], + "opposite_logprob_tokens": [ + -4.264202117919922 + ], + "logprob_reward": -0.014202351681888103, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202351681888103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2324570268392563, + 0.1463029980659485, + 0.07115326076745987, + 0.12140486389398575, + 0.17383812367916107, + 0.02057667262852192, + 0.04966796562075615, + 0.13354505598545074, + 0.02020576037466526, + 0.030848274007439613 + ], + "credit_map": [ + 0.05127653479576111, + 0.01082298532128334, + 0.020812492817640305, + 0.007107933983206749, + 0.0828932523727417, + 0.016708888113498688, + 0.009875223971903324, + 0.03295969218015671, + 0.013554022647440434, + 0.0170249342918396, + 0.013695030473172665, + 0.024928126484155655, + 0.02944508194923401, + 0.018234163522720337, + 0.01730032078921795, + 0.012121321633458138, + 0.0131937051191926, + 0.008713951334357262, + 0.010677207261323929, + 0.007466723211109638, + 0.007872652262449265, + 0.007433207705616951, + 0.009412680752575397, + 0.0063831317238509655, + 0.018205301836133003, + 0.013114922679960728, + 0.02115955948829651, + 0.011391522362828255, + 0.010679432190954685, + 0.013652822934091091, + 0.011052713729441166, + 0.022148584946990013, + 0.012993878684937954, + 0.018602494150400162, + 0.018951160833239555, + 0.01296782586723566, + 0.03846509009599686, + 0.026908302679657936, + 0.0274247694760561, + 0.01752459816634655, + 0.0018146028742194176, + 0.0022585787810385227, + 0.0010607955045998096, + 0.0015586984809488058, + 0.0028055687434971333, + 0.002399124437943101, + 0.005484618246555328, + 0.003194686258211732, + 0.007414987776428461, + 0.004542055539786816, + 0.005058357957750559, + 0.0030116394627839327, + 0.007916235364973545, + 0.008193641901016235, + 0.00468906806781888, + 0.008841977454721928, + 0.005332717671990395, + 0.003919463604688644, + 0.0026109868194907904, + 0.0033043206203728914, + 0.054322000592947006, + 0.05505375564098358, + 0.003814124735072255, + 0.005187669303268194, + 0.0031990285497158766, + 0.0018129667732864618, + 0.002076775301247835, + 0.001244792016223073, + 0.003379605244845152, + 0.00480962498113513, + 0.0018216317985206842, + 0.0018613323336467147, + 0.0034431121312081814, + 0.004103861283510923, + 0.003918568138033152, + 0.0026782560162246227, + 0.007384716998785734, + 0.004637063946574926, + 0.0027823937125504017, + 0.00190030038356781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.014202351681888103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036149173974990845 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.0036149173974990845, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036149173974990845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4244999587535858, + 0.05218873172998428, + 0.0505954846739769, + 0.05335468798875809, + 0.037355028092861176, + 0.04200032353401184, + 0.020312665030360222, + 0.044294700026512146, + 0.05293256416916847, + 0.22246579825878143 + ], + "credit_map": [ + 0.11169451475143433, + 0.01511844526976347, + 0.024280743673443794, + 0.008370617404580116, + 0.16959720849990845, + 0.03129717335104942, + 0.017548048868775368, + 0.04659322649240494, + 0.005948834121227264, + 0.004677951335906982, + 0.005511495750397444, + 0.010957002639770508, + 0.006660038139671087, + 0.004545129369944334, + 0.006296421866863966, + 0.0075918640941381454, + 0.00802052952349186, + 0.0064149354584515095, + 0.006661549210548401, + 0.006105656269937754, + 0.004910533782094717, + 0.005518621299415827, + 0.005638091824948788, + 0.0073255752213299274, + 0.007362161297351122, + 0.00747256213799119, + 0.0061807227320969105, + 0.006771215237677097, + 0.004621134605258703, + 0.006360052153468132, + 0.004298398271203041, + 0.010288448072969913, + 0.005034294445067644, + 0.00785135943442583, + 0.003642007941380143, + 0.0035149059258401394, + 0.005090413149446249, + 0.004908517003059387, + 0.003829168388620019, + 0.003484366461634636, + 0.004969531204551458, + 0.009337776340544224, + 0.003966452553868294, + 0.0038699295837432146, + 0.005619301926344633, + 0.005463232286274433, + 0.004003855399787426, + 0.004770247265696526, + 0.0035855735186487436, + 0.0021831353660672903, + 0.0022733237128704786, + 0.0014597700210288167, + 0.0030128671787679195, + 0.002929086098447442, + 0.0017874527256935835, + 0.0030814586207270622, + 0.004230683669447899, + 0.004278742242604494, + 0.0024922043085098267, + 0.004055268596857786, + 0.011116567067801952, + 0.011357572861015797, + 0.004028134513646364, + 0.0027355290949344635, + 0.006369889713823795, + 0.010254750959575176, + 0.004202303476631641, + 0.012176036834716797, + 0.006544987205415964, + 0.005045461002737284, + 0.0033168471418321133, + 0.005022287834435701, + 0.025138042867183685, + 0.029744284227490425, + 0.031047746539115906, + 0.020916255190968513, + 0.04970046505331993, + 0.03279491886496544, + 0.018539810553193092, + 0.014584296382963657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.0036149173974990845 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900903910398483 + ], + "opposite_logprob_tokens": [ + -2.4640090465545654 + ], + "logprob_reward": -0.08900903910398483, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900903910398483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26664116978645325, + 0.06273006647825241, + 0.017804810777306557, + 0.025721557438373566, + 0.018374457955360413, + 0.04280257597565651, + 0.04992365837097168, + 0.47570738196372986, + 0.016920624300837517, + 0.02337365597486496 + ], + "credit_map": [ + 0.06881434470415115, + 0.009685486555099487, + 0.016277702525258064, + 0.005401003174483776, + 0.10497086495161057, + 0.020955171436071396, + 0.009941340424120426, + 0.03059527277946472, + 0.007638143375515938, + 0.009415946900844574, + 0.005820435006171465, + 0.00544998561963439, + 0.012452754192054272, + 0.007618627976626158, + 0.007045299746096134, + 0.007288880180567503, + 0.0015200241468846798, + 0.0016800375888124108, + 0.002196206711232662, + 0.0025800191797316074, + 0.0026043583638966084, + 0.0022815633565187454, + 0.002638862933963537, + 0.002303740009665489, + 0.003054945496842265, + 0.0044332146644592285, + 0.002619384555146098, + 0.0035667524207383394, + 0.0025633058976382017, + 0.0026573988143354654, + 0.0026074708439409733, + 0.004219088237732649, + 0.002284166170284152, + 0.004294784739613533, + 0.0014623154420405626, + 0.0018775167409330606, + 0.0021591882687062025, + 0.0024375265929847956, + 0.0018043193267658353, + 0.002054642653092742, + 0.003032243810594082, + 0.0063017867505550385, + 0.003915383480489254, + 0.0036678293254226446, + 0.006368433590978384, + 0.0058927228674292564, + 0.0072646779008209705, + 0.00635950081050396, + 0.006630138028413057, + 0.003378630615770817, + 0.007319023832678795, + 0.004115901421755552, + 0.009572507813572884, + 0.007352498359978199, + 0.006247956305742264, + 0.00530700758099556, + 0.020660193637013435, + 0.017812341451644897, + 0.008670107461512089, + 0.013943441212177277, + 0.19222292304039001, + 0.19370612502098083, + 0.012526736594736576, + 0.016165537759661674, + 0.002169365296140313, + 0.0022176478523761034, + 0.0012969971867278218, + 0.002476040506735444, + 0.0026067509315907955, + 0.0027268577832728624, + 0.0013951655710116029, + 0.0020318012684583664, + 0.0027403731364756823, + 0.004025321453809738, + 0.0029931673780083656, + 0.0019097330514341593, + 0.00498067494481802, + 0.002813025377690792, + 0.0020054976921528578, + 0.0019058643374592066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.08900903910398483 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed camera pans sharply to the right, capturing a striking medium shot from the side, closely following a general clad in silver armor. He rides atop a towering, pure white warhorse, its powerful steed surging at full speed, its muscular mass undulating with every stride, its long white mane and tail whipping wildly in the wind. The general wears a suit of intricately crafted, mirror-polished silver plate armor, etched with reliefs of mythical beasts, a piercing gleam in the cold light scattered by the overcast sky. He leans forward, clinging to the ornate leather saddle, a crimson cape billowing behind him like a blaze. Through the slits of his helmet's visor, adorned with an equally crimson plume, one can glimpse his taut jaw and intensely focused gaze, fixed on an enemy just beyond the frame. One hand firmly grips a long spear, its tip gleaming coldly, held nearly parallel to the ground, poised for discharge. The other grips the reins, steering his mount with pinpoint precision. The background is a vast, desolate battlefield. The camera's rapid movement blurs the mud, gravel, and wreckage of the battlefield into a dynamic, brown-gray stripe. Only the splatter of earth beneath the horse's hooves is clearly visible, imbued with an epic sense of urgency and the relentless momentum of pursuit.", + "seed": 27802, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_2/index_8525_seed_27802_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11878469651564956, + "vlm_logprob": -0.11878469651564956, + "vlm_yes_no_margin": 3.049999976158142, + "vlm_token_credit": -0.11878469651564956, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04616429656744003 + ], + "opposite_logprob_tokens": [ + -3.4211642742156982 + ], + "logprob_reward": -0.04616429656744003, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04616429656744003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14219047129154205, + 0.5024428367614746, + 0.03545378893613815, + 0.02266577072441578, + 0.01027506310492754, + 0.05216929689049721, + 0.01605328358709812, + 0.09636110812425613, + 0.06590569764375687, + 0.056482650339603424 + ], + "credit_map": [ + 0.017789017409086227, + 0.018021589145064354, + 0.013078095391392708, + 0.010842636227607727, + 0.0311623215675354, + 0.022805839776992798, + 0.015438047237694263, + 0.013052919879555702, + 0.07810419052839279, + 0.03993767499923706, + 0.04211916774511337, + 0.04284379631280899, + 0.08329007029533386, + 0.10100335627794266, + 0.07068545371294022, + 0.04445911571383476, + 0.005471572745591402, + 0.0038891825824975967, + 0.0031286091543734074, + 0.0023214693646878004, + 0.009377538226544857, + 0.004713002592325211, + 0.003651878098025918, + 0.0029005338437855244, + 0.0036057643592357635, + 0.0023720888420939445, + 0.002453242428600788, + 0.002390082459896803, + 0.003324869554489851, + 0.0034378308337181807, + 0.0018743763212114573, + 0.00320751266553998, + 0.0010967206908389926, + 0.0015453180531039834, + 0.0010995531920343637, + 0.001237650285474956, + 0.0012591831618919969, + 0.001351661398075521, + 0.0007208752795122564, + 0.0019641011022031307, + 0.007697712164372206, + 0.006308464333415031, + 0.009140633046627045, + 0.005549731198698282, + 0.005630192346870899, + 0.0089698676019907, + 0.0035326145589351654, + 0.005340081639587879, + 0.0019431815017014742, + 0.002603435656055808, + 0.0019413609988987446, + 0.0012629368575289845, + 0.0023010200820863247, + 0.0014125847956165671, + 0.0017025034176185727, + 0.002886260626837611, + 0.010733985342085361, + 0.009318528696894646, + 0.009715978987514973, + 0.010046664625406265, + 0.008995710872113705, + 0.019187426194548607, + 0.009736432693898678, + 0.018626371398568153, + 0.010121744126081467, + 0.005542317871004343, + 0.004162239842116833, + 0.005873663816601038, + 0.006058757659047842, + 0.013208585791289806, + 0.007432239130139351, + 0.013506153598427773, + 0.009030241519212723, + 0.006960237864404917, + 0.006421737838536501, + 0.008656676858663559, + 0.0041336435824632645, + 0.005218576639890671, + 0.010234479792416096, + 0.005827055778354406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.04616429656744003 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38729438185691833 + ], + "opposite_logprob_tokens": [ + -1.1372944116592407 + ], + "logprob_reward": -0.38729438185691833, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38729438185691833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21059481799602509, + 0.20370134711265564, + 0.06796237826347351, + 0.05565708503127098, + 0.050391457974910736, + 0.10024939477443695, + 0.08010034263134003, + 0.06875292211771011, + 0.08448901027441025, + 0.078101247549057 + ], + "credit_map": [ + 0.026302436366677284, + 0.028096983209252357, + 0.03443039208650589, + 0.019541271030902863, + 0.031991004943847656, + 0.025846462696790695, + 0.02549934573471546, + 0.01888691447675228, + 0.025616837665438652, + 0.036341678351163864, + 0.014094908721745014, + 0.029262563213706017, + 0.03859174624085426, + 0.0237542986869812, + 0.024445826187729836, + 0.011593494564294815, + 0.010877658613026142, + 0.009267052635550499, + 0.01341990940272808, + 0.007944448851048946, + 0.007424262352287769, + 0.008046507835388184, + 0.004532731603831053, + 0.006449810229241848, + 0.007018669042736292, + 0.009417492896318436, + 0.007122191600501537, + 0.007642995100468397, + 0.006344925612211227, + 0.006985839921981096, + 0.004112916067242622, + 0.007012055721133947, + 0.007714662700891495, + 0.0045261625200510025, + 0.006200200412422419, + 0.006141552701592445, + 0.006478325929492712, + 0.009815805591642857, + 0.0035722122993320227, + 0.005942531395703554, + 0.0156597588211298, + 0.008800430223345757, + 0.008344988338649273, + 0.012478475458920002, + 0.008248893544077873, + 0.022756580263376236, + 0.006449488457292318, + 0.017510784789919853, + 0.009171641431748867, + 0.01026840042322874, + 0.0073549444787204266, + 0.007855971343815327, + 0.012366740964353085, + 0.009720771573483944, + 0.010978532023727894, + 0.012383336201310158, + 0.008146577514708042, + 0.010035470128059387, + 0.003961842507123947, + 0.010026704519987106, + 0.009564402513206005, + 0.007758211810141802, + 0.008647417649626732, + 0.010612291283905506, + 0.008347675204277039, + 0.008963278494775295, + 0.008955879136919975, + 0.007114531937986612, + 0.009966871701180935, + 0.012587814591825008, + 0.01934388279914856, + 0.009209079667925835, + 0.011979143135249615, + 0.007069033104926348, + 0.007142597809433937, + 0.007476723752915859, + 0.008154330775141716, + 0.013653192669153214, + 0.013702457770705223, + 0.008923767134547234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.38729438185691833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981787547469139 + ], + "opposite_logprob_tokens": [ + -3.529817819595337 + ], + "logprob_reward": -0.02981787547469139, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981787547469139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17434990406036377, + 0.42806434631347656, + 0.02739928476512432, + 0.11895197629928589, + 0.032017774879932404, + 0.076021209359169, + 0.024772267788648605, + 0.0494159571826458, + 0.04359672591090202, + 0.025410588830709457 + ], + "credit_map": [ + 0.030440667644143105, + 0.018873341381549835, + 0.02371945232152939, + 0.017343293875455856, + 0.01943487673997879, + 0.035472940653562546, + 0.015705546364188194, + 0.013359763659536839, + 0.05913218855857849, + 0.03886385262012482, + 0.03941519185900688, + 0.041314929723739624, + 0.07034233212471008, + 0.06772265583276749, + 0.06643333286046982, + 0.04483984038233757, + 0.004934927448630333, + 0.0026659113354980946, + 0.003966172691434622, + 0.0017779229674488306, + 0.005527711473405361, + 0.004067267756909132, + 0.002113958587870002, + 0.0023454101756215096, + 0.016277408227324486, + 0.01363818533718586, + 0.019119376316666603, + 0.01767105795443058, + 0.016346115618944168, + 0.017062565311789513, + 0.007935578003525734, + 0.0109016764909029, + 0.004078848287463188, + 0.0030595797579735518, + 0.005856588017195463, + 0.004906105808913708, + 0.004342708270996809, + 0.003787676803767681, + 0.0021148372907191515, + 0.003871429245918989, + 0.013150052167475224, + 0.005695742089301348, + 0.00672067329287529, + 0.006740563083440065, + 0.004535038955509663, + 0.02186993695795536, + 0.0036560685839504004, + 0.013653125613927841, + 0.0034480474423617125, + 0.002891367068514228, + 0.001934690517373383, + 0.0021818606182932854, + 0.0034822600428014994, + 0.003003638470545411, + 0.0017781868809834123, + 0.006052213720977306, + 0.005572858732193708, + 0.004788433667272329, + 0.004733683541417122, + 0.004192954394966364, + 0.004335154313594103, + 0.011368467472493649, + 0.0051657771691679955, + 0.009258625097572803, + 0.00570724718272686, + 0.007237575016915798, + 0.0026737728621810675, + 0.002815189305692911, + 0.008132773451507092, + 0.007054116576910019, + 0.0056825014762580395, + 0.004293549340218306, + 0.0020528899040073156, + 0.002577246865257621, + 0.0019311647629365325, + 0.003278474323451519, + 0.002565551083534956, + 0.003916020970791578, + 0.005242778919637203, + 0.0038464602548629045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.02981787547469139 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615036141127348 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.003615036141127348, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615036141127348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12857425212860107, + 0.2556995153427124, + 0.10654114931821823, + 0.044414203613996506, + 0.0334111787378788, + 0.11636527627706528, + 0.05643586814403534, + 0.05971651151776314, + 0.13024738430976868, + 0.06859469413757324 + ], + "credit_map": [ + 0.02226187102496624, + 0.012899436056613922, + 0.026395190507173538, + 0.015146749094128609, + 0.009725811891257763, + 0.022054234519600868, + 0.010677216574549675, + 0.00941373873502016, + 0.03041757270693779, + 0.03964615613222122, + 0.01933808997273445, + 0.02498532645404339, + 0.04306686669588089, + 0.036752913147211075, + 0.03920994699001312, + 0.022282633930444717, + 0.02075727842748165, + 0.013214895501732826, + 0.011467693373560905, + 0.008940275758504868, + 0.023025348782539368, + 0.01562836393713951, + 0.006457418203353882, + 0.007049870677292347, + 0.005149842705577612, + 0.004732460714876652, + 0.007646259851753712, + 0.00467483326792717, + 0.006708432454615831, + 0.00691716093569994, + 0.0029195635579526424, + 0.005665648728609085, + 0.0045685232616961, + 0.004191496875137091, + 0.005444029811769724, + 0.0036503050941973925, + 0.00424045417457819, + 0.005247837398201227, + 0.002615994540974498, + 0.0034525389783084393, + 0.01693560555577278, + 0.016683995723724365, + 0.012039449997246265, + 0.017537541687488556, + 0.012078069150447845, + 0.015995336696505547, + 0.0092975664883852, + 0.015797704458236694, + 0.008963046595454216, + 0.007702080998569727, + 0.004956300836056471, + 0.003918072674423456, + 0.009609817527234554, + 0.006914664059877396, + 0.00504621397703886, + 0.009325669147074223, + 0.008598800748586655, + 0.004554531071335077, + 0.004147085826843977, + 0.00474503031000495, + 0.006369484588503838, + 0.015564319677650928, + 0.004849146585911512, + 0.010888111777603626, + 0.021217186003923416, + 0.011428684927523136, + 0.013651091605424881, + 0.00903747882694006, + 0.016225503757596016, + 0.0284318495541811, + 0.013229955919086933, + 0.017025617882609367, + 0.015394189395010471, + 0.006088965106755495, + 0.008283007889986038, + 0.00794167909771204, + 0.006742157507687807, + 0.009400509297847748, + 0.008965902961790562, + 0.005778273567557335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.003615036141127348 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703189253807068 + ], + "opposite_logprob_tokens": [ + -2.1270318031311035 + ], + "logprob_reward": -0.12703189253807068, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12703189253807068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.168657585978508, + 0.1390058398246765, + 0.06622636318206787, + 0.10876883566379547, + 0.08626674115657806, + 0.10654213279485703, + 0.04029913991689682, + 0.14966697990894318, + 0.09995369613170624, + 0.034612685441970825 + ], + "credit_map": [ + 0.03967159241437912, + 0.020524917170405388, + 0.011611356399953365, + 0.011410955339670181, + 0.02728053741157055, + 0.03655490651726723, + 0.010620365850627422, + 0.010982952080667019, + 0.0176853034645319, + 0.015378525480628014, + 0.011507559567689896, + 0.015360054560005665, + 0.021341849118471146, + 0.021095849573612213, + 0.020996103063225746, + 0.01564059592783451, + 0.010261714458465576, + 0.008339125663042068, + 0.010666866786777973, + 0.004720146767795086, + 0.011783872731029987, + 0.009852197021245956, + 0.005321118980646133, + 0.005281328223645687, + 0.01165611669421196, + 0.0188506618142128, + 0.012911474332213402, + 0.015043091960251331, + 0.014200320467352867, + 0.01600043661892414, + 0.009462442249059677, + 0.010644293390214443, + 0.011011113412678242, + 0.01041211187839508, + 0.008641262538731098, + 0.007544554769992828, + 0.020776033401489258, + 0.010229426436126232, + 0.005571449641138315, + 0.012080791406333447, + 0.020283745601773262, + 0.005774964578449726, + 0.003858942538499832, + 0.007940473966300488, + 0.0073664868250489235, + 0.03662513196468353, + 0.0036850960459560156, + 0.021007293835282326, + 0.004481607582420111, + 0.004960704129189253, + 0.003975792322307825, + 0.005495247896760702, + 0.005205029621720314, + 0.004216206260025501, + 0.006021418608725071, + 0.005943137686699629, + 0.01880677416920662, + 0.01635807938873768, + 0.014776035211980343, + 0.011308686807751656, + 0.012215076014399529, + 0.03304152190685272, + 0.015354306437075138, + 0.02780650369822979, + 0.009056257084012032, + 0.007231751922518015, + 0.008126022294163704, + 0.008260519243776798, + 0.009811216033995152, + 0.027839135378599167, + 0.00825671385973692, + 0.02137208916246891, + 0.004572933539748192, + 0.003941326402127743, + 0.003099992638453841, + 0.004615107551217079, + 0.0033095036633312702, + 0.005247279070317745, + 0.005472639109939337, + 0.004353903699666262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.12703189253807068 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed camera pans sharply to the right, capturing a striking medium shot from the side, closely following a general clad in silver armor. He rides atop a towering, pure white warhorse, its powerful steed surging at full speed, its muscular mass undulating with every stride, its long white mane and tail whipping wildly in the wind. The general wears a suit of intricately crafted, mirror-polished silver plate armor, etched with reliefs of mythical beasts, a piercing gleam in the cold light scattered by the overcast sky. He leans forward, clinging to the ornate leather saddle, a crimson cape billowing behind him like a blaze. Through the slits of his helmet's visor, adorned with an equally crimson plume, one can glimpse his taut jaw and intensely focused gaze, fixed on an enemy just beyond the frame. One hand firmly grips a long spear, its tip gleaming coldly, held nearly parallel to the ground, poised for discharge. The other grips the reins, steering his mount with pinpoint precision. The background is a vast, desolate battlefield. The camera's rapid movement blurs the mud, gravel, and wreckage of the battlefield into a dynamic, brown-gray stripe. Only the splatter of earth beneath the horse's hooves is clearly visible, imbued with an epic sense of urgency and the relentless momentum of pursuit.", + "seed": 27803, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_2/index_8525_seed_27803_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1607631194870919, + "vlm_logprob": -0.1607631194870919, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.1607631194870919, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1775054931640625 + ], + "opposite_logprob_tokens": [ + -1.9275054931640625 + ], + "logprob_reward": -0.1775054931640625, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1775054931640625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4176742136478424, + 0.10790977627038956, + 0.03777066990733147, + 0.03625855967402458, + 0.10453686863183975, + 0.08491773158311844, + 0.04504828527569771, + 0.03842068463563919, + 0.06328772753477097, + 0.06417550891637802 + ], + "credit_map": [ + 0.05525451898574829, + 0.07623676210641861, + 0.07430803775787354, + 0.04059452936053276, + 0.04168574512004852, + 0.04842677339911461, + 0.037838708609342575, + 0.043329108506441116, + 0.014843346551060677, + 0.01579410769045353, + 0.01169656403362751, + 0.011205245740711689, + 0.016138574108481407, + 0.01321982778608799, + 0.01522365678101778, + 0.009788453578948975, + 0.004308052361011505, + 0.0035146933514624834, + 0.005384721793234348, + 0.004693520255386829, + 0.0036034071817994118, + 0.0035773422569036484, + 0.007372280582785606, + 0.005316650960594416, + 0.005271364469081163, + 0.004161598160862923, + 0.0048383078537881374, + 0.006238345988094807, + 0.004688141867518425, + 0.0022147672716528177, + 0.00411207927390933, + 0.004733953159302473, + 0.008608449250459671, + 0.018536534160375595, + 0.013880538754165173, + 0.014590098522603512, + 0.010203775018453598, + 0.010644122958183289, + 0.015008929185569286, + 0.013064424507319927, + 0.007051047403365374, + 0.008877024054527283, + 0.005688954144716263, + 0.008272059261798859, + 0.0136268250644207, + 0.018608620390295982, + 0.01411228533834219, + 0.008680915459990501, + 0.00328773888759315, + 0.002589516807347536, + 0.0030407151207327843, + 0.0030002680141478777, + 0.007336400914937258, + 0.00958366971462965, + 0.008421005681157112, + 0.007788974326103926, + 0.002257642336189747, + 0.005268114618957043, + 0.003723554313182831, + 0.004985463339835405, + 0.004381879232823849, + 0.006798267364501953, + 0.005582073703408241, + 0.005423687864094973, + 0.006807608995586634, + 0.009300170466303825, + 0.005035866983234882, + 0.006800755858421326, + 0.008190340362489223, + 0.010073243640363216, + 0.010709474794566631, + 0.006370267830789089, + 0.009333236142992973, + 0.012185899540781975, + 0.00348988501355052, + 0.005806502886116505, + 0.008908827789127827, + 0.010312733240425587, + 0.006267843768000603, + 0.007870577275753021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.1775054931640625 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5235522985458374 + ], + "opposite_logprob_tokens": [ + -0.8985522985458374 + ], + "logprob_reward": -0.5235522985458374, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5235522985458374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17235378921031952, + 0.19359330832958221, + 0.045013342052698135, + 0.054947465658187866, + 0.06910008192062378, + 0.151910200715065, + 0.060788724571466446, + 0.035545483231544495, + 0.107220359146595, + 0.10952727496623993 + ], + "credit_map": [ + 0.030578434467315674, + 0.018904080614447594, + 0.03308357298374176, + 0.016164369881153107, + 0.015426327474415302, + 0.02415486052632332, + 0.016488507390022278, + 0.01755361445248127, + 0.016169916838407516, + 0.024564629420638084, + 0.017056051641702652, + 0.02356582134962082, + 0.018559694290161133, + 0.01703188382089138, + 0.04424230381846428, + 0.03240298479795456, + 0.00753309391438961, + 0.006423638667911291, + 0.006778768263757229, + 0.006019226275384426, + 0.003933023661375046, + 0.0025588178541511297, + 0.006252401974052191, + 0.005514367949217558, + 0.005445287562906742, + 0.004657726269215345, + 0.00811928603798151, + 0.005542472004890442, + 0.007375597488135099, + 0.00696132518351078, + 0.008871873840689659, + 0.007973892614245415, + 0.007401640061289072, + 0.010229556821286678, + 0.007601284887641668, + 0.009803993627429008, + 0.008331777527928352, + 0.007229543756693602, + 0.009941807948052883, + 0.00856047123670578, + 0.010397478938102722, + 0.017940931022167206, + 0.017395317554473877, + 0.017205428332090378, + 0.020721493288874626, + 0.024970674887299538, + 0.029021792113780975, + 0.01425707433372736, + 0.0039005570579320192, + 0.003536584787070751, + 0.003982069436460733, + 0.004850464873015881, + 0.010726588778197765, + 0.01273263804614544, + 0.010467950254678726, + 0.01059186551719904, + 0.002939491765573621, + 0.0036548518110066652, + 0.0023601010907441378, + 0.0024357042275369167, + 0.005060195457190275, + 0.008179143071174622, + 0.004178180359303951, + 0.006737814284861088, + 0.020899252966046333, + 0.016996236518025398, + 0.011019174940884113, + 0.011251280084252357, + 0.013140431605279446, + 0.011387139558792114, + 0.010842678137123585, + 0.011684161610901356, + 0.013457776047289371, + 0.019141538068652153, + 0.006762159056961536, + 0.007284026592969894, + 0.021640604361891747, + 0.024510949850082397, + 0.008169806562364101, + 0.008560393005609512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.5235522985458374 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020599357783794403 + ], + "opposite_logprob_tokens": [ + -3.895599365234375 + ], + "logprob_reward": -0.020599357783794403, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020599357783794403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2478122115135193, + 0.1778472661972046, + 0.0441841222345829, + 0.07425136119127274, + 0.10754837095737457, + 0.13989686965942383, + 0.040333811193704605, + 0.024790620431303978, + 0.08954755961894989, + 0.05378783866763115 + ], + "credit_map": [ + 0.028903162106871605, + 0.032690078020095825, + 0.04755628854036331, + 0.02335801161825657, + 0.027018465101718903, + 0.02509118802845478, + 0.026549531146883965, + 0.036645472049713135, + 0.01635153964161873, + 0.016877951100468636, + 0.020539307966828346, + 0.016099460422992706, + 0.02796691656112671, + 0.020687265321612358, + 0.03723379224538803, + 0.022091027349233627, + 0.004234985914081335, + 0.006297478452324867, + 0.006132140755653381, + 0.005050223786383867, + 0.004301500041037798, + 0.004433240741491318, + 0.0087157366797328, + 0.005018815863877535, + 0.0069797527976334095, + 0.01006143819540739, + 0.007940786890685558, + 0.010818490758538246, + 0.009827733039855957, + 0.007089156191796064, + 0.010236627422273159, + 0.011297380551695824, + 0.016923613846302032, + 0.013825247064232826, + 0.012345911003649235, + 0.012017793022096157, + 0.012960566207766533, + 0.011348066851496696, + 0.013959720730781555, + 0.01416744850575924, + 0.010314456187188625, + 0.01565845124423504, + 0.012208565138280392, + 0.017323903739452362, + 0.019740590825676918, + 0.024814492091536522, + 0.02643618732690811, + 0.01340021938085556, + 0.002988793421536684, + 0.002472114050760865, + 0.0022090349812060595, + 0.0025959417689591646, + 0.006719330325722694, + 0.00795219000428915, + 0.006172036286443472, + 0.009224372915923595, + 0.00237005646340549, + 0.004100847989320755, + 0.0017669450026005507, + 0.0021850194316357374, + 0.0033386985305696726, + 0.004705933853983879, + 0.002383486134931445, + 0.003939634654670954, + 0.011726556345820427, + 0.014706929214298725, + 0.010360178537666798, + 0.009804027155041695, + 0.011965181678533554, + 0.011044138111174107, + 0.009328014217317104, + 0.01061253622174263, + 0.00562685914337635, + 0.006706481799483299, + 0.003386084455996752, + 0.0033443125430494547, + 0.011193616315722466, + 0.012602980248630047, + 0.005691247526556253, + 0.005236255936324596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.020599357783794403 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319268717430532 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.00319268717430532, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319268717430532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1939990371465683, + 0.12082710862159729, + 0.056922994554042816, + 0.030315613374114037, + 0.045690007507801056, + 0.15019658207893372, + 0.04947950318455696, + 0.021615266799926758, + 0.15581806004047394, + 0.17513585090637207 + ], + "credit_map": [ + 0.027980124577879906, + 0.03130955621600151, + 0.03712628781795502, + 0.018833477050065994, + 0.0189854446798563, + 0.022645676508545876, + 0.019084418192505836, + 0.01803404651582241, + 0.014170840382575989, + 0.014763345010578632, + 0.023250000551342964, + 0.016682218760252, + 0.013056653551757336, + 0.01088010799139738, + 0.013278838247060776, + 0.014745105989277363, + 0.011203795671463013, + 0.008765080943703651, + 0.006307796109467745, + 0.0067358785308897495, + 0.005633527413010597, + 0.0038811517879366875, + 0.00646646972745657, + 0.007929298095405102, + 0.0032568771857768297, + 0.004534060135483742, + 0.0030872055795043707, + 0.004340345971286297, + 0.004017593339085579, + 0.0021481350995600224, + 0.004509688820689917, + 0.004421709105372429, + 0.010125945322215557, + 0.006330910138785839, + 0.005062046926468611, + 0.005680785048753023, + 0.004053278360515833, + 0.0042335400357842445, + 0.0042253886349499226, + 0.005978115368634462, + 0.011543381959199905, + 0.01706472411751747, + 0.01642744429409504, + 0.018757466226816177, + 0.019365208223462105, + 0.02399815060198307, + 0.028998518362641335, + 0.014041682705283165, + 0.0030057979747653008, + 0.0037596034817397594, + 0.003764726920053363, + 0.00450904993340373, + 0.00873303972184658, + 0.008082078769803047, + 0.009905660524964333, + 0.007719544693827629, + 0.0019095021998509765, + 0.0036560159642249346, + 0.0016311329090967774, + 0.0026716229040175676, + 0.002585899317637086, + 0.003979797940701246, + 0.001931864651851356, + 0.0032494293991476297, + 0.03286648914217949, + 0.019035791978240013, + 0.014125567860901356, + 0.019289391115307808, + 0.018927210941910744, + 0.016397112980484962, + 0.01927146501839161, + 0.015905030071735382, + 0.02088954672217369, + 0.022224707528948784, + 0.00987723283469677, + 0.012052316218614578, + 0.04097524285316467, + 0.04835624247789383, + 0.008343405090272427, + 0.01241715345531702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.00319268717430532 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a general clad in silver armor riding a white warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor mirror-polished silver plate etched with reliefs of mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse pure white with a long mane and tail whipping wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a crimson cape billowing behind the general and a crimson plume on the helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan sharply to the right while the background blurs into a dynamic brown-gray stripe?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896576076745987 + ], + "opposite_logprob_tokens": [ + -2.578965663909912 + ], + "logprob_reward": -0.07896576076745987, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896576076745987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08782508969306946, + 0.10701295733451843, + 0.03955136612057686, + 0.0639432966709137, + 0.10583705455064774, + 0.13645358383655548, + 0.1418989896774292, + 0.219835102558136, + 0.044866643846035004, + 0.05277596786618233 + ], + "credit_map": [ + 0.009307140484452248, + 0.012749386951327324, + 0.013991945423185825, + 0.0071860807947814465, + 0.008653040044009686, + 0.009875146672129631, + 0.01167686190456152, + 0.01438549067825079, + 0.012505126185715199, + 0.012065150775015354, + 0.017466386780142784, + 0.009812481701374054, + 0.016452768817543983, + 0.010687668807804585, + 0.016330720856785774, + 0.011692656204104424, + 0.003484234446659684, + 0.00402911426499486, + 0.0061782896518707275, + 0.0040763565339148045, + 0.0036807849537581205, + 0.0042702509090304375, + 0.00927113275974989, + 0.004561199806630611, + 0.007378143724054098, + 0.008759832940995693, + 0.006131377071142197, + 0.01005563884973526, + 0.0066563962027430534, + 0.005527909845113754, + 0.007868857122957706, + 0.011565137654542923, + 0.011896067298948765, + 0.011407754383981228, + 0.013684997335076332, + 0.013795020990073681, + 0.011591735295951366, + 0.011848310008645058, + 0.015897581353783607, + 0.015715589746832848, + 0.011279829777777195, + 0.017088675871491432, + 0.02095816284418106, + 0.01602018252015114, + 0.015566720627248287, + 0.01700885035097599, + 0.026895171031355858, + 0.011635983362793922, + 0.010893386788666248, + 0.011164789088070393, + 0.010612145997583866, + 0.013558963313698769, + 0.02737760730087757, + 0.021467989310622215, + 0.027859525755047798, + 0.018964575603604317, + 0.013462439179420471, + 0.028093423694372177, + 0.015142200514674187, + 0.01935221068561077, + 0.029898937791585922, + 0.04798794910311699, + 0.02387079782783985, + 0.04202713444828987, + 0.006083793938159943, + 0.009274525567889214, + 0.003305901773273945, + 0.005932115949690342, + 0.00532088428735733, + 0.005841051694005728, + 0.004541479982435703, + 0.0045668878592550755, + 0.00470494432374835, + 0.005658240988850594, + 0.0028522717766463757, + 0.0033545703627169132, + 0.011747559532523155, + 0.011929193511605263, + 0.005854469258338213, + 0.006674718577414751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.07896576076745987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_3/index_tensor([9208, 9208])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_3/index_tensor([9208, 9208])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..62336d7b4d5b537faefc308a5a6b082c6b4b1fde --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_3/index_tensor([9208, 9208])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-speed camera executes slow, arcing trajectories, capturing a breathtaking time-slice effect: an exquisite, polished gold necklace frozen in mid-air against a stark, velvety black void. The necklace itself is depicted not in motion but as a series of crisp, paused moments, with a few distinct, perfectly formed links tracing flawless arcs. This visual echo conveys the fluid, powerful movements of an unseen dancer. A striking spotlight illuminates the scene from directly above, causing the smooth, reflective gold surface to shimmer with bright, sharp highlights. The perfect trajectory, from the initial upward thrust to the moment of weightlessness at the summit to the beginning of the graceful descent, is clearly defined. The combination of shimmering gold and deep black, combined with the deconstructed movement, creates an elegant atmosphere of minimalism, intense focus, suspended time, and pure dynamic beauty.", + "seed": 19640, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_3/index_9208_seed_19640_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09513790141791105, + "vlm_logprob": -0.09513790141791105, + "vlm_yes_no_margin": 3.625000011920929, + "vlm_token_credit": -0.09513790141791105, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011474448256194592 + ], + "opposite_logprob_tokens": [ + -4.761474609375 + ], + "logprob_reward": -0.011474448256194592, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011474448256194592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06806854158639908, + 0.03465554490685463, + 0.030853595584630966, + 0.19846709072589874, + 0.09090739488601685, + 0.08325229585170746, + 0.0629599392414093, + 0.038734596222639084, + 0.16357390582561493, + 0.22852712869644165 + ], + "credit_map": [ + 0.008625427260994911, + 0.008997355587780476, + 0.008119872771203518, + 0.004815150983631611, + 0.006089804228395224, + 0.011722251772880554, + 0.008847832679748535, + 0.010850844904780388, + 0.0038938720244914293, + 0.0031474907882511616, + 0.004035995341837406, + 0.003667594864964485, + 0.0026789314579218626, + 0.006770425941795111, + 0.0035595130175352097, + 0.006901724264025688, + 0.003475841833278537, + 0.0041707283817231655, + 0.003952702973037958, + 0.005029941443353891, + 0.0029063632246106863, + 0.004988318309187889, + 0.002318138722330332, + 0.004011558368802071, + 0.020625216886401176, + 0.011096365749835968, + 0.0596606470644474, + 0.018007755279541016, + 0.009813816286623478, + 0.010525878518819809, + 0.011886289343237877, + 0.0568511039018631, + 0.012318444438278675, + 0.010863673873245716, + 0.014446509070694447, + 0.009598993696272373, + 0.00986021663993597, + 0.01240161620080471, + 0.009953107684850693, + 0.011464838869869709, + 0.010774767026305199, + 0.009241453371942043, + 0.012239188887178898, + 0.009049970656633377, + 0.009841221384704113, + 0.01339537464082241, + 0.010972158052027225, + 0.007738161366432905, + 0.010053866542875767, + 0.005772027652710676, + 0.007277736905962229, + 0.009566121734678745, + 0.005807274952530861, + 0.009505262598395348, + 0.007323152851313353, + 0.007654493674635887, + 0.004357681609690189, + 0.003799145109951496, + 0.00450705923140049, + 0.005606174003332853, + 0.0035124269779771566, + 0.007201838307082653, + 0.005069386679679155, + 0.0046808854676783085, + 0.018783176317811012, + 0.018009155988693237, + 0.014948251657187939, + 0.016968294978141785, + 0.021509578451514244, + 0.023608721792697906, + 0.019823065027594566, + 0.029923662543296814, + 0.04916239157319069, + 0.023547973483800888, + 0.021186525002121925, + 0.021891264244914055, + 0.016956496983766556, + 0.01689550280570984, + 0.023774273693561554, + 0.055112697184085846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.011474448256194592 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052642300724983215 + ], + "opposite_logprob_tokens": [ + -5.2552642822265625 + ], + "logprob_reward": -0.0052642300724983215, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052642300724983215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18609416484832764, + 0.0879792645573616, + 0.0618794746696949, + 0.059311024844646454, + 0.06485119462013245, + 0.0683756023645401, + 0.054088156670331955, + 0.05895119905471802, + 0.09278067946434021, + 0.26568928360939026 + ], + "credit_map": [ + 0.019091442227363586, + 0.02573348954319954, + 0.01959325559437275, + 0.013034224510192871, + 0.014483779668807983, + 0.03461473807692528, + 0.019942963495850563, + 0.03960025683045387, + 0.007474400103092194, + 0.008501172065734863, + 0.007481626234948635, + 0.009077073074877262, + 0.005636642687022686, + 0.01923597976565361, + 0.00589965283870697, + 0.024672703817486763, + 0.006810128688812256, + 0.008524294011294842, + 0.007059059105813503, + 0.006638893857598305, + 0.005035338457673788, + 0.010887264274060726, + 0.0060529643669724464, + 0.0108715258538723, + 0.005932110361754894, + 0.0077008274383842945, + 0.008874548599123955, + 0.010192946530878544, + 0.006814639084041119, + 0.006335761863738298, + 0.008137276396155357, + 0.005322909448295832, + 0.011149890720844269, + 0.008163519203662872, + 0.010992559604346752, + 0.00827241875231266, + 0.005928505677729845, + 0.008074060082435608, + 0.005341988522559404, + 0.0069282506592571735, + 0.011057425290346146, + 0.00782336201518774, + 0.012840316630899906, + 0.005756676662713289, + 0.0062351836822927, + 0.012048599310219288, + 0.0057339840568602085, + 0.006880046799778938, + 0.004896951839327812, + 0.003832588903605938, + 0.006487733218818903, + 0.004821887239813805, + 0.003528463887050748, + 0.01047269906848669, + 0.008426004089415073, + 0.011621826328337193, + 0.005715421866625547, + 0.0052822367288172245, + 0.00801246240735054, + 0.00859001837670803, + 0.005383089184761047, + 0.00902004074305296, + 0.0072752865962684155, + 0.009672640822827816, + 0.008777761831879616, + 0.010750165209174156, + 0.008630570024251938, + 0.009830953553318977, + 0.01600741595029831, + 0.011618522927165031, + 0.012765309773385525, + 0.014399977400898933, + 0.054366543889045715, + 0.024135787039995193, + 0.029266444966197014, + 0.02901468798518181, + 0.02039623260498047, + 0.0201994888484478, + 0.023134006187319756, + 0.06517606973648071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0052642300724983215 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206531450152397 + ], + "opposite_logprob_tokens": [ + -2.812065362930298 + ], + "logprob_reward": -0.06206531450152397, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206531450152397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08385460078716278, + 0.0444410964846611, + 0.03705448657274246, + 0.15860363841056824, + 0.049404483288526535, + 0.117959164083004, + 0.04360070079565048, + 0.08062055706977844, + 0.2739173471927643, + 0.11054390668869019 + ], + "credit_map": [ + 0.009868329390883446, + 0.010316453874111176, + 0.009649636223912239, + 0.00883716344833374, + 0.0089411037042737, + 0.014048312790691853, + 0.00839035864919424, + 0.013803238049149513, + 0.004526547156274319, + 0.004555050749331713, + 0.005117340479046106, + 0.0062854052521288395, + 0.0034209927543997765, + 0.008320978842675686, + 0.004315153229981661, + 0.007899629883468151, + 0.00425544660538435, + 0.005907527636736631, + 0.004057899117469788, + 0.0066398754715919495, + 0.003720708889886737, + 0.005349939689040184, + 0.0027339141815900803, + 0.004389178939163685, + 0.014915324747562408, + 0.008648285642266273, + 0.05108027160167694, + 0.011445478536188602, + 0.006233035121113062, + 0.008865034207701683, + 0.008008449338376522, + 0.04940775781869888, + 0.006964781787246466, + 0.005592850968241692, + 0.0083302678540349, + 0.005368346348404884, + 0.00603996217250824, + 0.005925282370299101, + 0.006430008448660374, + 0.0047529893927276134, + 0.013899734243750572, + 0.016390614211559296, + 0.012071751989424229, + 0.01738864555954933, + 0.012736942619085312, + 0.016162503510713577, + 0.014925791881978512, + 0.014383187517523766, + 0.004060506820678711, + 0.002766887191683054, + 0.005644583608955145, + 0.004217647947371006, + 0.0036758468486368656, + 0.0076368143782019615, + 0.006398742087185383, + 0.009199674241244793, + 0.007241036277264357, + 0.007179674692451954, + 0.008276586420834064, + 0.008764038793742657, + 0.008306912146508694, + 0.015434900298714638, + 0.010486814193427563, + 0.014930590987205505, + 0.03278965875506401, + 0.03032676689326763, + 0.024954376742243767, + 0.022586384788155556, + 0.029364310204982758, + 0.04712275043129921, + 0.03195876628160477, + 0.05481433495879173, + 0.02157464250922203, + 0.010064699687063694, + 0.00952314306050539, + 0.011663909070193768, + 0.010894729755818844, + 0.011395691893994808, + 0.013671536929905415, + 0.02175556682050228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -0.06206531450152397 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870932459831238 + ], + "opposite_logprob_tokens": [ + -1.1370933055877686 + ], + "logprob_reward": -0.3870932459831238, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3870932459831238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08693721890449524, + 0.04520384967327118, + 0.07657158374786377, + 0.2813809812068939, + 0.07527206838130951, + 0.06041234731674194, + 0.06178627908229828, + 0.11571897566318512, + 0.08526132255792618, + 0.11145541071891785 + ], + "credit_map": [ + 0.010089592076838017, + 0.011077827773988247, + 0.010481505654752254, + 0.0080818897113204, + 0.01031606923788786, + 0.013679556548595428, + 0.00952262245118618, + 0.013688141480088234, + 0.005261566489934921, + 0.0044755637645721436, + 0.003835233161225915, + 0.005731265991926193, + 0.0032505032140761614, + 0.008388304151594639, + 0.004692788701504469, + 0.00956861674785614, + 0.00513653177767992, + 0.007436377927660942, + 0.008625585585832596, + 0.01067084725946188, + 0.00811662059277296, + 0.013853728771209717, + 0.009447433985769749, + 0.013284445740282536, + 0.02751355804502964, + 0.016213668510317802, + 0.08487865328788757, + 0.025578249245882034, + 0.014900038950145245, + 0.015189498662948608, + 0.016912806779146194, + 0.08019450306892395, + 0.010714548639953136, + 0.011010936461389065, + 0.013044948689639568, + 0.010660958476364613, + 0.005850460845977068, + 0.009878301061689854, + 0.006049163173884153, + 0.008062738925218582, + 0.009148355573415756, + 0.006271306425333023, + 0.008334358222782612, + 0.0068841297179460526, + 0.006851555779576302, + 0.009293309412896633, + 0.00781380757689476, + 0.005815516691654921, + 0.004087404347956181, + 0.004620068706572056, + 0.007550138048827648, + 0.007519981358200312, + 0.0038742765318602324, + 0.010954376310110092, + 0.011979930102825165, + 0.011200095526874065, + 0.011842862702906132, + 0.00945951696485281, + 0.013126775622367859, + 0.014178754761815071, + 0.011212426237761974, + 0.022213945165276527, + 0.013727936893701553, + 0.019956747069954872, + 0.010228583589196205, + 0.00839370395988226, + 0.008214293979108334, + 0.007844057865440845, + 0.008374476805329323, + 0.015605973079800606, + 0.01061426755040884, + 0.01598595269024372, + 0.02391454204916954, + 0.007942198775708675, + 0.01041752751916647, + 0.011255917139351368, + 0.009002367965877056, + 0.013339857570827007, + 0.01388081256300211, + 0.021702179685235023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 9 + ], + "selected_reward": -0.3870932459831238 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097922682762146 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.0097922682762146, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097922682762146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09937699139118195, + 0.03905581682920456, + 0.050247982144355774, + 0.08690743148326874, + 0.04546597972512245, + 0.15723535418510437, + 0.07191484421491623, + 0.1047143042087555, + 0.24272167682647705, + 0.10235965251922607 + ], + "credit_map": [ + 0.011241748929023743, + 0.01535242423415184, + 0.011957565322518349, + 0.008866085670888424, + 0.011084215715527534, + 0.015946678817272186, + 0.010375499725341797, + 0.014552772976458073, + 0.004979373421519995, + 0.0037603299133479595, + 0.0033419698011130095, + 0.0029794431757181883, + 0.002702106023207307, + 0.008468035608530045, + 0.003913992550224066, + 0.008910570293664932, + 0.006501434836536646, + 0.006572210695594549, + 0.0069148121401667595, + 0.0073728980496525764, + 0.0043076565489172935, + 0.007273748051375151, + 0.004883517045527697, + 0.0064217038452625275, + 0.009196127764880657, + 0.006102600134909153, + 0.024997670203447342, + 0.006610298529267311, + 0.004998816642910242, + 0.005266891792416573, + 0.005262500140815973, + 0.02447252906858921, + 0.008558315224945545, + 0.006578480824828148, + 0.007836359553039074, + 0.004581509158015251, + 0.00415498623624444, + 0.0058677103370428085, + 0.003291146829724312, + 0.004597468767315149, + 0.018915357068181038, + 0.017486609518527985, + 0.025856230407953262, + 0.023315733298659325, + 0.014849285595119, + 0.026311388239264488, + 0.016734641045331955, + 0.013766112737357616, + 0.007672307547181845, + 0.005423013586550951, + 0.007486686110496521, + 0.00951444637030363, + 0.004996070638298988, + 0.01331211719661951, + 0.010046698153018951, + 0.013463509269058704, + 0.012078661471605301, + 0.009483353234827518, + 0.011471075005829334, + 0.01308407448232174, + 0.008953694254159927, + 0.0187697634100914, + 0.014182697981595993, + 0.01669098623096943, + 0.023352360352873802, + 0.02517309971153736, + 0.02765834890305996, + 0.022599825635552406, + 0.025576051324605942, + 0.04460407420992851, + 0.03598726540803909, + 0.037770651280879974, + 0.01812508888542652, + 0.009640328586101532, + 0.0099972914904356, + 0.012609852477908134, + 0.009462730027735233, + 0.008723173290491104, + 0.013920619152486324, + 0.019880570471286774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.0097922682762146 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed camera executes slow, arcing trajectories, capturing a breathtaking time-slice effect: an exquisite, polished gold necklace frozen in mid-air against a stark, velvety black void. The necklace itself is depicted not in motion but as a series of crisp, paused moments, with a few distinct, perfectly formed links tracing flawless arcs. This visual echo conveys the fluid, powerful movements of an unseen dancer. A striking spotlight illuminates the scene from directly above, causing the smooth, reflective gold surface to shimmer with bright, sharp highlights. The perfect trajectory, from the initial upward thrust to the moment of weightlessness at the summit to the beginning of the graceful descent, is clearly defined. The combination of shimmering gold and deep black, combined with the deconstructed movement, creates an elegant atmosphere of minimalism, intense focus, suspended time, and pure dynamic beauty.", + "seed": 19641, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_3/index_9208_seed_19641_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06447233706712723, + "vlm_logprob": -0.06447233706712723, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.06447233706712723, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010846467688679695 + ], + "opposite_logprob_tokens": [ + -5.0108466148376465 + ], + "logprob_reward": -0.010846467688679695, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.010846467688679695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046322207897901535, + 0.038861069828271866, + 0.1250147670507431, + 0.015226854011416435, + 0.06639459729194641, + 0.083487868309021, + 0.10117675364017487, + 0.04792461544275284, + 0.24658115208148956, + 0.2290101796388626 + ], + "credit_map": [ + 0.004667057655751705, + 0.00532651087269187, + 0.008698785677552223, + 0.010109532624483109, + 0.003718714229762554, + 0.0036997764836996794, + 0.004780905321240425, + 0.005320925731211901, + 0.005016493611037731, + 0.005018431227654219, + 0.00645408732816577, + 0.008127196691930294, + 0.0028922653291374445, + 0.003306231228634715, + 0.002634652191773057, + 0.005411711521446705, + 0.013129516504704952, + 0.012942234054207802, + 0.031061742454767227, + 0.011361449956893921, + 0.008703331463038921, + 0.01648556813597679, + 0.007221085950732231, + 0.024109840393066406, + 0.002466503530740738, + 0.00144115777220577, + 0.003660004585981369, + 0.0011306029045954347, + 0.0012664970709010959, + 0.0012366913724690676, + 0.001011466607451439, + 0.0030139314476400614, + 0.009616300463676453, + 0.0037926083896309137, + 0.015404868870973587, + 0.005909206811338663, + 0.0031491322442889214, + 0.005878355819731951, + 0.0094046201556921, + 0.013239501975476742, + 0.008713115006685257, + 0.0074400524608790874, + 0.019303224980831146, + 0.005419241264462471, + 0.006935418583452702, + 0.007688084151595831, + 0.008690910413861275, + 0.019297821447253227, + 0.011360595002770424, + 0.0039005957078188658, + 0.035532012581825256, + 0.006640654988586903, + 0.0033552641980350018, + 0.023752162232995033, + 0.005607834085822105, + 0.011027619242668152, + 0.007169011048972607, + 0.004294802434742451, + 0.0077837263233959675, + 0.005438270978629589, + 0.0027137333527207375, + 0.010660618543624878, + 0.003174085868522525, + 0.006690370384603739, + 0.02077736146748066, + 0.03711533173918724, + 0.035772670060396194, + 0.019891871139407158, + 0.02782062254846096, + 0.04816894605755806, + 0.01989560015499592, + 0.037138763815164566, + 0.019871747121214867, + 0.015489458106458187, + 0.06034143269062042, + 0.02707904390990734, + 0.010033966042101383, + 0.052063245326280594, + 0.012510365806519985, + 0.03162092715501785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.010846467688679695 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260198377072811 + ], + "opposite_logprob_tokens": [ + -5.255259990692139 + ], + "logprob_reward": -0.005260198377072811, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260198377072811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20492152869701385, + 0.07816044241189957, + 0.11288023740053177, + 0.02517125941812992, + 0.06241489574313164, + 0.07437381148338318, + 0.08193469047546387, + 0.08271590620279312, + 0.22460578382015228, + 0.05282140523195267 + ], + "credit_map": [ + 0.023889226838946342, + 0.022790851071476936, + 0.030998272821307182, + 0.04827475547790527, + 0.012387411668896675, + 0.016202125698328018, + 0.02008179761469364, + 0.030297091230750084, + 0.011917125433683395, + 0.006261439993977547, + 0.015177453868091106, + 0.013471923768520355, + 0.005250351969152689, + 0.005206311587244272, + 0.004858098924160004, + 0.016017738729715347, + 0.013787825591862202, + 0.009006623178720474, + 0.03136437386274338, + 0.01007937639951706, + 0.006854184903204441, + 0.009492908604443073, + 0.005862381774932146, + 0.026432562619447708, + 0.005013045854866505, + 0.0020099268294870853, + 0.007076923735439777, + 0.0015344263520091772, + 0.0015459490241482854, + 0.0013484441442415118, + 0.0020978888496756554, + 0.004544656258076429, + 0.010428677313029766, + 0.0041318368166685104, + 0.009380294941365719, + 0.006668031681329012, + 0.0035954113118350506, + 0.006501688621938229, + 0.012116341851651669, + 0.009592612273991108, + 0.013989465311169624, + 0.006068667396903038, + 0.011943915858864784, + 0.005221243016421795, + 0.00502999359741807, + 0.007123926188796759, + 0.010150574147701263, + 0.014846023172140121, + 0.009045501239597797, + 0.003382316092029214, + 0.02904064767062664, + 0.005083030089735985, + 0.002935859840363264, + 0.01868804357945919, + 0.0038731740787625313, + 0.009886115789413452, + 0.008737663738429546, + 0.005195440258830786, + 0.019736915826797485, + 0.005553852301090956, + 0.004585210233926773, + 0.021536076441407204, + 0.0055448985658586025, + 0.01182584185153246, + 0.01947174221277237, + 0.03363988175988197, + 0.03260425850749016, + 0.017979821190238, + 0.026020599529147148, + 0.04240325465798378, + 0.01886601932346821, + 0.033620208501815796, + 0.004945505876094103, + 0.006761523894965649, + 0.011178521439433098, + 0.004308605100959539, + 0.002574893878772855, + 0.013239268213510513, + 0.005058425944298506, + 0.004754664842039347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.005260198377072811 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043078016489744186 + ], + "opposite_logprob_tokens": [ + -3.1680779457092285 + ], + "logprob_reward": -0.043078016489744186, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043078016489744186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12812624871730804, + 0.10809982568025589, + 0.12414374947547913, + 0.03513230010867119, + 0.08809603750705719, + 0.17821301519870758, + 0.0702252984046936, + 0.06916394084692001, + 0.08105067908763885, + 0.11774885654449463 + ], + "credit_map": [ + 0.013777926564216614, + 0.01740199699997902, + 0.01982641965150833, + 0.027625296264886856, + 0.009082548320293427, + 0.011284928768873215, + 0.013805080205202103, + 0.015322046354413033, + 0.013102389872074127, + 0.01439457107335329, + 0.015240974724292755, + 0.016605382785201073, + 0.013726748526096344, + 0.015535837970674038, + 0.009380744770169258, + 0.01011317502707243, + 0.012889336794614792, + 0.012244119308888912, + 0.033016398549079895, + 0.011285976506769657, + 0.00819855835288763, + 0.01445767655968666, + 0.007287722546607256, + 0.024763964116573334, + 0.00518302246928215, + 0.0027843972202390432, + 0.009592845104634762, + 0.002214007778093219, + 0.0023211596999317408, + 0.0027304706163704395, + 0.002107894280925393, + 0.008198500610888004, + 0.012771696783602238, + 0.004879897926002741, + 0.019954338669776917, + 0.00857127271592617, + 0.00505045847967267, + 0.00814462173730135, + 0.01298738457262516, + 0.01573636569082737, + 0.025648508220911026, + 0.016329074278473854, + 0.035575803369283676, + 0.011675097979605198, + 0.01443332526832819, + 0.016415420919656754, + 0.019485464319586754, + 0.03865032270550728, + 0.010557943023741245, + 0.005494951270520687, + 0.014671911485493183, + 0.005363969597965479, + 0.0056157419458031654, + 0.01003292202949524, + 0.008436613716185093, + 0.010051245801150799, + 0.006073735188692808, + 0.0029003345407545567, + 0.01606881618499756, + 0.003769982373341918, + 0.002125423401594162, + 0.02472500316798687, + 0.0021394090726971626, + 0.011361237615346909, + 0.008562796749174595, + 0.012861047871410847, + 0.0113425487652421, + 0.008164306171238422, + 0.009148137643933296, + 0.012980382889509201, + 0.007479412481188774, + 0.010512047447264194, + 0.013150721788406372, + 0.008419432677328587, + 0.027547739446163177, + 0.015116396360099316, + 0.004890814423561096, + 0.021695129573345184, + 0.00666534248739481, + 0.020263278856873512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.043078016489744186 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25209471583366394 + ], + "opposite_logprob_tokens": [ + -1.5020947456359863 + ], + "logprob_reward": -0.25209471583366394, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25209471583366394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10857152938842773, + 0.057902801781892776, + 0.055072907358407974, + 0.1869775801897049, + 0.04521787911653519, + 0.09538391977548599, + 0.165695458650589, + 0.04539183899760246, + 0.11974979192018509, + 0.1200362965464592 + ], + "credit_map": [ + 0.015673518180847168, + 0.015714382752776146, + 0.017677899450063705, + 0.021537914872169495, + 0.008490489795804024, + 0.008851880207657814, + 0.010854864493012428, + 0.009770589880645275, + 0.010702120140194893, + 0.0050783539190888405, + 0.009675608016550541, + 0.01111912913620472, + 0.003662116127088666, + 0.003968385048210621, + 0.0046817585825920105, + 0.009015326388180256, + 0.00982956774532795, + 0.00476029422134161, + 0.01047268882393837, + 0.005255761556327343, + 0.003214144380763173, + 0.006084055174142122, + 0.003299868665635586, + 0.01215653121471405, + 0.02420533262193203, + 0.009418142959475517, + 0.06488475948572159, + 0.008886551484465599, + 0.006307713221758604, + 0.010186201892793179, + 0.009397692047059536, + 0.053691163659095764, + 0.007050147745758295, + 0.004253114107996225, + 0.009442267008125782, + 0.004780202638357878, + 0.0033556625712662935, + 0.002985580824315548, + 0.0060561709105968475, + 0.00729473726823926, + 0.010233866982161999, + 0.006936651188880205, + 0.023536089807748795, + 0.006433010101318359, + 0.0067162057384848595, + 0.010743160732090473, + 0.009233142249286175, + 0.021551795303821564, + 0.0203749630600214, + 0.0067285639233887196, + 0.05520186573266983, + 0.011842135339975357, + 0.005164181813597679, + 0.037812042981386185, + 0.010406697168946266, + 0.018165014684200287, + 0.007176653482019901, + 0.005044338293373585, + 0.008454493246972561, + 0.003835253184661269, + 0.002882433356717229, + 0.007886622101068497, + 0.0035793795250356197, + 0.006532668136060238, + 0.009179837070405483, + 0.015620656311511993, + 0.022053679451346397, + 0.00867511611431837, + 0.012563534080982208, + 0.022801266983151436, + 0.008280414156615734, + 0.020575284957885742, + 0.012129103764891624, + 0.006714938208460808, + 0.03152512013912201, + 0.013798539526760578, + 0.004782172851264477, + 0.024780914187431335, + 0.005355885252356529, + 0.020949620753526688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.25209471583366394 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011082286946475506 + ], + "opposite_logprob_tokens": [ + -4.511082172393799 + ], + "logprob_reward": -0.011082286946475506, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082286946475506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08447932451963425, + 0.18638308346271515, + 0.1470467746257782, + 0.11977551877498627, + 0.07754988968372345, + 0.08907738327980042, + 0.09778908640146255, + 0.021586453542113304, + 0.145803302526474, + 0.030509240925312042 + ], + "credit_map": [ + 0.01029196847230196, + 0.008458047173917294, + 0.01532079093158245, + 0.01812892034649849, + 0.0069296048022806644, + 0.008485431782901287, + 0.009202067740261555, + 0.007662491872906685, + 0.02168029360473156, + 0.026077810674905777, + 0.02385627292096615, + 0.030444679781794548, + 0.024065006524324417, + 0.027299167588353157, + 0.01573328673839569, + 0.01722653955221176, + 0.01970800571143627, + 0.014058000408113003, + 0.03255816176533699, + 0.014055364765226841, + 0.012370030395686626, + 0.017977362498641014, + 0.009554549120366573, + 0.02676529251039028, + 0.014584122225642204, + 0.0041343881748616695, + 0.044574394822120667, + 0.005437829997390509, + 0.004556622356176376, + 0.0036212224513292313, + 0.004163951613008976, + 0.038702983409166336, + 0.013920475728809834, + 0.004194408655166626, + 0.015455801039934158, + 0.006502630189061165, + 0.004964043851941824, + 0.005361804272979498, + 0.010108854621648788, + 0.01704186387360096, + 0.021489553153514862, + 0.006036495324224234, + 0.015849541872739792, + 0.006599127314984798, + 0.00476702768355608, + 0.008110493421554565, + 0.00936019979417324, + 0.016864938661456108, + 0.009828293696045876, + 0.004086475819349289, + 0.03347540646791458, + 0.007776627317070961, + 0.003788665635511279, + 0.022298689931631088, + 0.006068747956305742, + 0.01046617142856121, + 0.003298065159469843, + 0.00178374198731035, + 0.004340992774814367, + 0.0021010187920182943, + 0.0013645426370203495, + 0.0033488660119473934, + 0.001390967983752489, + 0.003958257380872965, + 0.012884673662483692, + 0.02269880846142769, + 0.020407887175679207, + 0.012685918249189854, + 0.01920606940984726, + 0.025059709325432777, + 0.011972456239163876, + 0.020887767896056175, + 0.004316833335906267, + 0.0033970512449741364, + 0.0053437985479831696, + 0.002867461647838354, + 0.0020910524763166904, + 0.0052477275021374226, + 0.002770348684862256, + 0.00447496771812439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 8 + ], + "selected_reward": -0.011082286946475506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed camera executes slow, arcing trajectories, capturing a breathtaking time-slice effect: an exquisite, polished gold necklace frozen in mid-air against a stark, velvety black void. The necklace itself is depicted not in motion but as a series of crisp, paused moments, with a few distinct, perfectly formed links tracing flawless arcs. This visual echo conveys the fluid, powerful movements of an unseen dancer. A striking spotlight illuminates the scene from directly above, causing the smooth, reflective gold surface to shimmer with bright, sharp highlights. The perfect trajectory, from the initial upward thrust to the moment of weightlessness at the summit to the beginning of the graceful descent, is clearly defined. The combination of shimmering gold and deep black, combined with the deconstructed movement, creates an elegant atmosphere of minimalism, intense focus, suspended time, and pure dynamic beauty.", + "seed": 19642, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_3/index_9208_seed_19642_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2962631996721029, + "vlm_logprob": -0.2962631996721029, + "vlm_yes_no_margin": 2.15, + "vlm_token_credit": -0.2962631996721029, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03575681149959564 + ], + "opposite_logprob_tokens": [ + -3.535756826400757 + ], + "logprob_reward": -0.03575681149959564, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03575681149959564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06805580854415894, + 0.19001071155071259, + 0.0626145675778389, + 0.11197558790445328, + 0.18899458646774292, + 0.05551667511463165, + 0.0629008337855339, + 0.12718015909194946, + 0.053863320499658585, + 0.07888773083686829 + ], + "credit_map": [ + 0.01148658711463213, + 0.007432209327816963, + 0.008885256014764309, + 0.005782109685242176, + 0.007480199448764324, + 0.009134020656347275, + 0.005472097080200911, + 0.012383333407342434, + 0.0170252975076437, + 0.05357421562075615, + 0.01993149146437645, + 0.011226095259189606, + 0.015405363403260708, + 0.013669513165950775, + 0.01128401793539524, + 0.04789471998810768, + 0.006534394808113575, + 0.006928429938852787, + 0.006173501722514629, + 0.0066893878392875195, + 0.009528200142085552, + 0.011979836039245129, + 0.0055387248285114765, + 0.009242098778486252, + 0.013391522690653801, + 0.014380354434251785, + 0.01538038719445467, + 0.009957783855497837, + 0.016481012105941772, + 0.01871739886701107, + 0.011949420906603336, + 0.011717721819877625, + 0.019266419112682343, + 0.03073917143046856, + 0.02866579219698906, + 0.01712409220635891, + 0.026501530781388283, + 0.0273449644446373, + 0.02264462597668171, + 0.016707997769117355, + 0.008613157086074352, + 0.0052244216203689575, + 0.006314354483038187, + 0.006703329738229513, + 0.007344401907175779, + 0.008092927746474743, + 0.007538887672126293, + 0.005685199052095413, + 0.01388548780232668, + 0.006291607394814491, + 0.008758888579905033, + 0.006218700669705868, + 0.005477815866470337, + 0.010628316551446915, + 0.005751152988523245, + 0.005888863932341337, + 0.014665872789919376, + 0.009113343432545662, + 0.016506999731063843, + 0.010222836397588253, + 0.009266291745007038, + 0.028285568580031395, + 0.013546033762395382, + 0.025573227554559708, + 0.0065954914316535, + 0.004413642920553684, + 0.013292393647134304, + 0.005374704021960497, + 0.004016208462417126, + 0.007572900038212538, + 0.004984901752322912, + 0.007613079156726599, + 0.006718471646308899, + 0.0064274705946445465, + 0.008458876982331276, + 0.007809789385646582, + 0.0076162549667060375, + 0.017342479899525642, + 0.006578534841537476, + 0.01793585903942585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.03575681149959564 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206889823079109 + ], + "opposite_logprob_tokens": [ + -2.8120689392089844 + ], + "logprob_reward": -0.06206889823079109, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206889823079109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10429303348064423, + 0.15338072180747986, + 0.0428641214966774, + 0.04638814926147461, + 0.04139315336942673, + 0.1410978138446808, + 0.0535445436835289, + 0.26097580790519714, + 0.020031509920954704, + 0.1360311210155487 + ], + "credit_map": [ + 0.016387132927775383, + 0.015752313658595085, + 0.008905958384275436, + 0.009548372589051723, + 0.010456682182848454, + 0.019103100523352623, + 0.009927909821271896, + 0.0142115643247962, + 0.017920494079589844, + 0.05743800103664398, + 0.00893484428524971, + 0.006471633445471525, + 0.01007656380534172, + 0.0062602548860013485, + 0.00527352886274457, + 0.04100540652871132, + 0.00408542063087225, + 0.0049307867884635925, + 0.0040453290566802025, + 0.006088730413466692, + 0.005935785360634327, + 0.006207883358001709, + 0.005941318813711405, + 0.005628864746540785, + 0.012344066984951496, + 0.004809626843780279, + 0.008020194247364998, + 0.003472581272944808, + 0.006658955477178097, + 0.005239017773419619, + 0.0030169959645718336, + 0.002826714888215065, + 0.004180608317255974, + 0.006376987788826227, + 0.0046900431625545025, + 0.003260853933170438, + 0.005791273433715105, + 0.005044772755354643, + 0.008326942101120949, + 0.0037216716445982456, + 0.02111731842160225, + 0.016207318753004074, + 0.01466951984912157, + 0.020635806024074554, + 0.016235483810305595, + 0.018228409811854362, + 0.018197407945990562, + 0.015806546434760094, + 0.008619120344519615, + 0.00409966753795743, + 0.006704043131321669, + 0.0051107024773955345, + 0.004428063984960318, + 0.007984496653079987, + 0.006570327095687389, + 0.010028119198977947, + 0.028164051473140717, + 0.019612818956375122, + 0.03409226983785629, + 0.02454431541264057, + 0.02113533765077591, + 0.0538051538169384, + 0.027773961424827576, + 0.05184788629412651, + 0.0038523913826793432, + 0.0019650987815111876, + 0.003002250101417303, + 0.0020218559075146914, + 0.0018293634057044983, + 0.002991439076140523, + 0.0016365033807232976, + 0.0027326096314936876, + 0.01707616075873375, + 0.015121505595743656, + 0.01487290021032095, + 0.021322770044207573, + 0.014705413952469826, + 0.01966020278632641, + 0.015201974660158157, + 0.018070194870233536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.06206889823079109 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134874999523163 + ], + "opposite_logprob_tokens": [ + -1.3134875297546387 + ], + "logprob_reward": -0.3134874999523163, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134874999523163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10026069730520248, + 0.19983696937561035, + 0.07618232816457748, + 0.055900439620018005, + 0.11457806080579758, + 0.09252031147480011, + 0.07054632902145386, + 0.12150786817073822, + 0.07972057908773422, + 0.08894642442464828 + ], + "credit_map": [ + 0.014526110142469406, + 0.012384546920657158, + 0.010151801630854607, + 0.011453196406364441, + 0.011024336330592632, + 0.01743214949965477, + 0.010096552781760693, + 0.013192007318139076, + 0.024498648941516876, + 0.07758598774671555, + 0.009304122999310493, + 0.006612651981413364, + 0.011313543654978275, + 0.009282364509999752, + 0.007569546811282635, + 0.05367011949419975, + 0.005910609383136034, + 0.011325646191835403, + 0.006135187111794949, + 0.012381646782159805, + 0.018040023744106293, + 0.005882014520466328, + 0.009123645722866058, + 0.007383563090115786, + 0.00790412351489067, + 0.007263675797730684, + 0.007883036509156227, + 0.005093067418783903, + 0.0073277298361063, + 0.005686955526471138, + 0.0072152926586568356, + 0.00752656627446413, + 0.010501089505851269, + 0.01453009806573391, + 0.013332145288586617, + 0.011533265933394432, + 0.016538813710212708, + 0.015103306621313095, + 0.022379156202077866, + 0.010660192929208279, + 0.011095165275037289, + 0.012729543261229992, + 0.0118807852268219, + 0.010933103039860725, + 0.011238864623010159, + 0.011970560997724533, + 0.010784512385725975, + 0.011887784115970135, + 0.008722896687686443, + 0.008608490228652954, + 0.007699303328990936, + 0.008562646806240082, + 0.007165459915995598, + 0.010727855376899242, + 0.0070293559692800045, + 0.012030325829982758, + 0.013666574843227863, + 0.008211275562644005, + 0.027154965326189995, + 0.014230874367058277, + 0.009984162636101246, + 0.02124668098986149, + 0.008891098201274872, + 0.01812223717570305, + 0.012050305493175983, + 0.0054998742416501045, + 0.013762442395091057, + 0.006196682341396809, + 0.005928831174969673, + 0.0159501601010561, + 0.006121695972979069, + 0.014210600405931473, + 0.013186409138143063, + 0.012616988271474838, + 0.0103081576526165, + 0.01646972820162773, + 0.006567538250237703, + 0.012405483052134514, + 0.008098764345049858, + 0.009293356910347939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.3134874999523163 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.053910255432129 + ], + "opposite_logprob_tokens": [ + -0.4289102852344513 + ], + "logprob_reward": -1.053910255432129, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.053910255432129, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1684080958366394, + 0.0600651390850544, + 0.05861777439713478, + 0.06958768516778946, + 0.17561282217502594, + 0.2583862543106079, + 0.025036310777068138, + 0.06093698740005493, + 0.06287520378828049, + 0.06047367677092552 + ], + "credit_map": [ + 0.027224991470575333, + 0.026903295889496803, + 0.015866028144955635, + 0.016364390030503273, + 0.017403235659003258, + 0.02552778273820877, + 0.013677502982318401, + 0.025440869852900505, + 0.009838406927883625, + 0.012772681191563606, + 0.005688079167157412, + 0.006526169367134571, + 0.005852461326867342, + 0.006933562457561493, + 0.006196650676429272, + 0.006257125176489353, + 0.0053800372406840324, + 0.011157812550663948, + 0.004735096823424101, + 0.003243440994992852, + 0.01722446084022522, + 0.0064908224157989025, + 0.004224664065986872, + 0.006161440163850784, + 0.017243433743715286, + 0.00948744174093008, + 0.011113225482404232, + 0.00644163740798831, + 0.008981111459434032, + 0.006043506320565939, + 0.005093643441796303, + 0.005183682776987553, + 0.018839463591575623, + 0.02892777882516384, + 0.025857441127300262, + 0.016108766198158264, + 0.02959078550338745, + 0.025331567972898483, + 0.01988755352795124, + 0.011069465428590775, + 0.03767561540007591, + 0.02683890238404274, + 0.028089329600334167, + 0.034631770104169846, + 0.03014327958226204, + 0.03269796445965767, + 0.03224918618798256, + 0.03606022521853447, + 0.0037681153044104576, + 0.002425708109512925, + 0.0027148716617375612, + 0.002166380174458027, + 0.0030370503664016724, + 0.003959455993026495, + 0.002354398835450411, + 0.004610329866409302, + 0.007376841735094786, + 0.004296402912586927, + 0.008321190252900124, + 0.005164803471416235, + 0.00502609321847558, + 0.013411682099103928, + 0.006784576922655106, + 0.010555397719144821, + 0.009578801691532135, + 0.004475525580346584, + 0.010868367739021778, + 0.005569310858845711, + 0.004720919765532017, + 0.010434835217893124, + 0.00498340604826808, + 0.012244041077792645, + 0.006145266350358725, + 0.005504087544977665, + 0.007407936733216047, + 0.007908806204795837, + 0.004673414397984743, + 0.012097995728254318, + 0.005632657092064619, + 0.011103518307209015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -1.053910255432129 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016092533245682716 + ], + "opposite_logprob_tokens": [ + -4.141092300415039 + ], + "logprob_reward": -0.016092533245682716, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016092533245682716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10729312896728516, + 0.09401346743106842, + 0.058227308094501495, + 0.039548829197883606, + 0.11165174841880798, + 0.1213158667087555, + 0.11381828784942627, + 0.10320582985877991, + 0.07804162800312042, + 0.17288388311862946 + ], + "credit_map": [ + 0.01755882054567337, + 0.014720624312758446, + 0.009892065078020096, + 0.012899775058031082, + 0.014314282685518265, + 0.012901649810373783, + 0.00841270200908184, + 0.016593214124441147, + 0.011157249100506306, + 0.02455195225775242, + 0.0075127980671823025, + 0.007030788343399763, + 0.007328767795115709, + 0.008514773100614548, + 0.006049308925867081, + 0.021867845207452774, + 0.005324304569512606, + 0.010089467279613018, + 0.003837720025330782, + 0.006464558187872171, + 0.011144972406327724, + 0.009313317947089672, + 0.005982053466141224, + 0.006070917006582022, + 0.0072413282468914986, + 0.004464999306946993, + 0.00799896102398634, + 0.0037481989711523056, + 0.005723833572119474, + 0.0035336671862751245, + 0.0034147067926824093, + 0.003423138288781047, + 0.010643897578120232, + 0.012432429939508438, + 0.012385542504489422, + 0.008944684639573097, + 0.017238542437553406, + 0.015163855627179146, + 0.023449433967471123, + 0.011393372900784016, + 0.015411986969411373, + 0.01913994364440441, + 0.0123406071215868, + 0.016498221084475517, + 0.015689877793192863, + 0.016580767929553986, + 0.015733864158391953, + 0.009920609183609486, + 0.015813611447811127, + 0.010940482839941978, + 0.01571539416909218, + 0.016269328072667122, + 0.009839989244937897, + 0.016105379909276962, + 0.01143710408359766, + 0.017697002738714218, + 0.009040722623467445, + 0.007482319604605436, + 0.01677543856203556, + 0.009422414936125278, + 0.008725617080926895, + 0.01971435546875, + 0.01137530617415905, + 0.020669657737016678, + 0.010338849388062954, + 0.007412251550704241, + 0.013406516052782536, + 0.00869831908494234, + 0.007134188897907734, + 0.011719897389411926, + 0.008410640992224216, + 0.01092095673084259, + 0.01800977624952793, + 0.021438026800751686, + 0.02333197183907032, + 0.027323368936777115, + 0.012610462494194508, + 0.029576236382126808, + 0.011675900779664516, + 0.02891814336180687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.016092533245682716 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "high-speed camera executes slow, arcing trajectories, capturing a breathtaking time-slice effect: an exquisite, polished gold necklace frozen in mid-air against a stark, velvety black void. The necklace itself is depicted not in motion but as a series of crisp, paused moments, with a few distinct, perfectly formed links tracing flawless arcs. This visual echo conveys the fluid, powerful movements of an unseen dancer. A striking spotlight illuminates the scene from directly above, causing the smooth, reflective gold surface to shimmer with bright, sharp highlights. The perfect trajectory, from the initial upward thrust to the moment of weightlessness at the summit to the beginning of the graceful descent, is clearly defined. The combination of shimmering gold and deep black, combined with the deconstructed movement, creates an elegant atmosphere of minimalism, intense focus, suspended time, and pure dynamic beauty.", + "seed": 19643, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_3/index_9208_seed_19643_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12247432013973594, + "vlm_logprob": -0.12247432013973594, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.12247432013973594, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023796316236257553 + ], + "opposite_logprob_tokens": [ + -3.898796319961548 + ], + "logprob_reward": -0.023796316236257553, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.023796316236257553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09521322697401047, + 0.07669559866189957, + 0.02931196428835392, + 0.038417357951402664, + 0.18893587589263916, + 0.05201910436153412, + 0.2696435749530792, + 0.0915231928229332, + 0.09649376571178436, + 0.06174635887145996 + ], + "credit_map": [ + 0.00583594711497426, + 0.009995030239224434, + 0.006725966930389404, + 0.010911846533417702, + 0.015012451447546482, + 0.021841054782271385, + 0.01110876351594925, + 0.013782176189124584, + 0.007587255444377661, + 0.008831401355564594, + 0.007637923117727041, + 0.011107577942311764, + 0.01395276840776205, + 0.011755673214793205, + 0.008269352838397026, + 0.007553648203611374, + 0.0034521573688834906, + 0.00201528612524271, + 0.003265153616666794, + 0.003719035303220153, + 0.0033176960423588753, + 0.004496423527598381, + 0.0054895346984267235, + 0.003556677373126149, + 0.003968703560531139, + 0.002444430021569133, + 0.005710734985768795, + 0.0040400829166173935, + 0.003124016337096691, + 0.008436671458184719, + 0.006306113675236702, + 0.00438660429790616, + 0.028590787202119827, + 0.01674446277320385, + 0.0407368540763855, + 0.015988238155841827, + 0.015084901824593544, + 0.03246519714593887, + 0.024808870628476143, + 0.014516563154757023, + 0.005148068070411682, + 0.00402039522305131, + 0.009222169406712055, + 0.005285339895635843, + 0.0052824174053967, + 0.009893720038235188, + 0.006476447451859713, + 0.0066905515268445015, + 0.024238580837845802, + 0.017015188932418823, + 0.05715281888842583, + 0.021385619416832924, + 0.025242749601602554, + 0.06911736726760864, + 0.014064567163586617, + 0.04142669588327408, + 0.00852456409484148, + 0.006572586949914694, + 0.02353232353925705, + 0.007290590554475784, + 0.009851880371570587, + 0.01993582211434841, + 0.009167410433292389, + 0.006648011971265078, + 0.01060077827423811, + 0.01113148219883442, + 0.015051372349262238, + 0.012017159722745419, + 0.014023186638951302, + 0.013310936279594898, + 0.013461838476359844, + 0.006897005718201399, + 0.00644639553502202, + 0.006124473176896572, + 0.008347569033503532, + 0.006668576505035162, + 0.00766255846247077, + 0.007939653471112251, + 0.012252730317413807, + 0.006304403301328421 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 8 + ], + "selected_reward": -0.023796316236257553 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792503900825977 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.009792503900825977, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792503900825977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08264085650444031, + 0.044466499239206314, + 0.1308390349149704, + 0.04266802966594696, + 0.08788343518972397, + 0.020239626988768578, + 0.40248697996139526, + 0.06700139492750168, + 0.046967219561338425, + 0.07480692118406296 + ], + "credit_map": [ + 0.007340903859585524, + 0.009814673103392124, + 0.008280937559902668, + 0.011078380048274994, + 0.010504448786377907, + 0.012907973490655422, + 0.014416981488466263, + 0.008296558633446693, + 0.004121088422834873, + 0.003090917831286788, + 0.003694170853123069, + 0.005119623150676489, + 0.007231343537569046, + 0.007652435451745987, + 0.007127848919481039, + 0.006429069209843874, + 0.009981224313378334, + 0.009922259487211704, + 0.012506425380706787, + 0.015591120347380638, + 0.015003623440861702, + 0.02510875090956688, + 0.027400415390729904, + 0.015325209125876427, + 0.0024466330651193857, + 0.0017937836237251759, + 0.0045232572592794895, + 0.003855847753584385, + 0.003735775128006935, + 0.007714335806667805, + 0.008739364333450794, + 0.009859033860266209, + 0.011001115664839745, + 0.00542086735367775, + 0.024666810408234596, + 0.005933996755629778, + 0.004250690340995789, + 0.024487575516104698, + 0.004685811698436737, + 0.007436565589159727, + 0.0014237243449315429, + 0.002000941429287195, + 0.005453748162835836, + 0.0018554524285718799, + 0.0016977014020085335, + 0.0047001619823277, + 0.001823536236770451, + 0.0012843618169426918, + 0.035857830196619034, + 0.026693405583500862, + 0.09250100702047348, + 0.03222785145044327, + 0.034479353576898575, + 0.09762106090784073, + 0.021648459136486053, + 0.06145801767706871, + 0.007348211482167244, + 0.004626901354640722, + 0.012046344578266144, + 0.007982255890965462, + 0.007867073640227318, + 0.010871363803744316, + 0.008193012326955795, + 0.008066230453550816, + 0.004343835636973381, + 0.0059664007276296616, + 0.0074563403613865376, + 0.004908920731395483, + 0.0045348578132689, + 0.0061183832585811615, + 0.008837222121655941, + 0.004801258910447359, + 0.008824308402836323, + 0.0075159803964197636, + 0.011928321793675423, + 0.005485009402036667, + 0.005030438303947449, + 0.006276112049818039, + 0.011169581674039364, + 0.018577171489596367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.009792503900825977 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867283254861832 + ], + "opposite_logprob_tokens": [ + -3.048672914505005 + ], + "logprob_reward": -0.04867283254861832, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867283254861832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07379847019910812, + 0.030837763100862503, + 0.028087671846151352, + 0.06342338770627975, + 0.15459950268268585, + 0.03086143732070923, + 0.4190973937511444, + 0.12798479199409485, + 0.04322649538516998, + 0.02808307483792305 + ], + "credit_map": [ + 0.006055065896362066, + 0.008969297632575035, + 0.006968170870095491, + 0.00913876574486494, + 0.008844122290611267, + 0.012211318127810955, + 0.011755003593862057, + 0.00985672790557146, + 0.00267806276679039, + 0.0029326369985938072, + 0.0026444089598953724, + 0.004342924803495407, + 0.005291366949677467, + 0.00506710447371006, + 0.004450110252946615, + 0.0034311464987695217, + 0.0030548160430043936, + 0.002324518049135804, + 0.0035864401143044233, + 0.0025171241723001003, + 0.0030034780502319336, + 0.004529334604740143, + 0.005434286314994097, + 0.0036376737989485264, + 0.006785326637327671, + 0.003998976666480303, + 0.011512749828398228, + 0.005974704399704933, + 0.004457208327949047, + 0.010029817931354046, + 0.010404463857412338, + 0.010260140523314476, + 0.02137378603219986, + 0.013071012683212757, + 0.027536028996109962, + 0.015559299848973751, + 0.011870726011693478, + 0.022686505690217018, + 0.026461400091648102, + 0.01604074239730835, + 0.002699400531128049, + 0.0018302240641787648, + 0.008295639418065548, + 0.0027198800817131996, + 0.0017828018171712756, + 0.007733879145234823, + 0.0025386714842170477, + 0.003260940546169877, + 0.037042036652565, + 0.027793817222118378, + 0.10575126111507416, + 0.032233208417892456, + 0.03458249196410179, + 0.10026653110980988, + 0.02227213978767395, + 0.0591559074819088, + 0.012326688505709171, + 0.009234214201569557, + 0.03263716772198677, + 0.010499196127057076, + 0.015036854892969131, + 0.025569656863808632, + 0.012191814370453358, + 0.010489190928637981, + 0.004921949468553066, + 0.004538287874311209, + 0.0060508884489536285, + 0.004907647613435984, + 0.005680745933204889, + 0.006952786352485418, + 0.006697813048958778, + 0.0034763796720653772, + 0.003299818839877844, + 0.0028075729496777058, + 0.0038468188140541315, + 0.0024776754435151815, + 0.002507636323571205, + 0.003445586422458291, + 0.004377085715532303, + 0.00532087916508317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -0.04867283254861832 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233614444732666 + ], + "opposite_logprob_tokens": [ + -0.8983614444732666 + ], + "logprob_reward": -0.5233614444732666, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233614444732666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06161246821284294, + 0.017315847799181938, + 0.021107640117406845, + 0.053649142384529114, + 0.12455286830663681, + 0.020486710593104362, + 0.5189498066902161, + 0.15649817883968353, + 0.012409184128046036, + 0.013418157584965229 + ], + "credit_map": [ + 0.005257521290332079, + 0.0075178369879722595, + 0.004636306781321764, + 0.008145476691424847, + 0.008240580558776855, + 0.010441236197948456, + 0.008819754235446453, + 0.00855375174432993, + 0.001627101213671267, + 0.0013462917413562536, + 0.0017598798731341958, + 0.002538539469242096, + 0.0026080049574375153, + 0.002702238503843546, + 0.002651338931173086, + 0.0020824538078159094, + 0.001579571166075766, + 0.0014925359282642603, + 0.0021591747645288706, + 0.0022670046892017126, + 0.002235626569017768, + 0.003884355304762721, + 0.004273151978850365, + 0.0032162200659513474, + 0.003070664359256625, + 0.0020429776050150394, + 0.006925194524228573, + 0.004805867560207844, + 0.004292474593967199, + 0.00855772290378809, + 0.012805321253836155, + 0.011148919351398945, + 0.019791441038250923, + 0.005569323897361755, + 0.03077360987663269, + 0.007588735315948725, + 0.006699000485241413, + 0.032387323677539825, + 0.009990813210606575, + 0.011752624996006489, + 0.0018424151930958033, + 0.0013379257870838046, + 0.005501911975443363, + 0.001967203337699175, + 0.001287487568333745, + 0.005145980045199394, + 0.0014649918302893639, + 0.0019387929933145642, + 0.04613524675369263, + 0.03457675874233246, + 0.12584851682186127, + 0.0398302860558033, + 0.04347875341773033, + 0.12484455108642578, + 0.02751072496175766, + 0.07672501355409622, + 0.017030268907546997, + 0.012166202068328857, + 0.032587625086307526, + 0.015492857433855534, + 0.018061533570289612, + 0.027397047728300095, + 0.018914954736828804, + 0.01484767347574234, + 0.0012252874439582229, + 0.001108260010369122, + 0.001746329595334828, + 0.0009996569715440273, + 0.001569669577293098, + 0.0020513832569122314, + 0.0021885524038225412, + 0.0015200446359813213, + 0.002102478640154004, + 0.0010692714713513851, + 0.0018447668990120292, + 0.0013290996430441737, + 0.0014591794461011887, + 0.0013068827101960778, + 0.002214669017121196, + 0.0020918091759085655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.5233614444732666 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a polished gold necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, velvety black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the necklace depicted as a series of crisp, paused moments tracing arcs rather than continuous motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spotlight coming from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the necklace's trajectory show an upward thrust, a moment of weightlessness at the summit, and the beginning of a graceful descent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006748503539711237 + ], + "opposite_logprob_tokens": [ + -5.006748676300049 + ], + "logprob_reward": -0.006748503539711237, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748503539711237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11376016587018967, + 0.043058495968580246, + 0.0318228043615818, + 0.024615300819277763, + 0.1301220804452896, + 0.06666839122772217, + 0.39912599325180054, + 0.07307389378547668, + 0.08564932644367218, + 0.032103512436151505 + ], + "credit_map": [ + 0.011506432667374611, + 0.013359990902245045, + 0.008977008983492851, + 0.014791181311011314, + 0.015138905495405197, + 0.022870495915412903, + 0.013931699097156525, + 0.013184464536607265, + 0.004472372587770224, + 0.004157593008130789, + 0.004536614753305912, + 0.005991155747324228, + 0.006817074492573738, + 0.006617284845560789, + 0.005291527137160301, + 0.005174876656383276, + 0.0038826095405966043, + 0.0030567762441933155, + 0.003047103760764003, + 0.0038868386764079332, + 0.004032391123473644, + 0.00442050164565444, + 0.0055741253308951855, + 0.00392246013507247, + 0.003329928731545806, + 0.0015798592939972878, + 0.003708421252667904, + 0.001701379893347621, + 0.001702136010862887, + 0.006078930106014013, + 0.003547599073499441, + 0.0029670484364032745, + 0.025141358375549316, + 0.008969887159764767, + 0.023914756253361702, + 0.009762550704181194, + 0.009709165431559086, + 0.022647462785243988, + 0.015753701329231262, + 0.014223207719624043, + 0.009085222147405148, + 0.004088285379111767, + 0.01788424327969551, + 0.005488728638738394, + 0.005304744001477957, + 0.015000046230852604, + 0.004619136452674866, + 0.005197993014007807, + 0.03345741331577301, + 0.028281157836318016, + 0.10642500221729279, + 0.029192421585321426, + 0.035914208739995956, + 0.09065435081720352, + 0.02248258702456951, + 0.052718862891197205, + 0.006782918237149715, + 0.006235032342374325, + 0.012470726855099201, + 0.011973883025348186, + 0.011000308208167553, + 0.010028046555817127, + 0.007624706719070673, + 0.006958275567740202, + 0.009961096569895744, + 0.009794726967811584, + 0.012259243056178093, + 0.008252269588410854, + 0.010867093689739704, + 0.014375125989317894, + 0.012359892949461937, + 0.007779889274388552, + 0.0034449873492121696, + 0.0033289126586169004, + 0.005038026720285416, + 0.003656210144981742, + 0.003036753972992301, + 0.00422168942168355, + 0.004380928352475166, + 0.0049960073083639145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.006748503539711237 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_4/index_tensor([5809, 5809])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_4/index_tensor([5809, 5809])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d8bd91a407f1be0acac5e628a5d79965bcdc766b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_4/index_tensor([5809, 5809])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, shot from a slightly upward angle, slowly but steadily advances, first glimpsing a vast, bizarre plant leaf in the foreground. Its deep, dark green, velvety surface is covered with intricate, luminous veins, resembling circuit boards, faintly luminescent. As the camera deepens, focus gradually settles on an Asian model nestled within a futuristic greenhouse. Her face is sculpted, her long, wet black hair neatly combed back. Her eyes, cold and distant, stare blankly into the camera. She wears a couture dress crafted from a liquid metal silver fabric with dramatic, architectural shoulders, making her stand out against the backdrop. Behind her, rows of stainless steel and glass vertical growth racks display a variety of exotic plants, each a psychedelic display of vibrant colors under alternating purple and magenta LED grow lights. The lights pulsated in a slow and rhythmic rhythm, periodically sweeping across her expressionless face and the glittering surface of the clothing, casting ever-changing long shadows. The entire scene was shrouded in a cold tone, and the precise coldness of technology and the ultimate beauty of luxury fashion blended at this moment, creating a fashion blockbuster atmosphere full of mystery and visual impact.", + "seed": 12047, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_4/index_5809_seed_12047_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.022695226594805717, + "vlm_logprob": -0.022695226594805717, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.022695226594805717, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011856324970722198 + ], + "opposite_logprob_tokens": [ + -4.761856555938721 + ], + "logprob_reward": -0.011856324970722198, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011856324970722198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3117935359477997, + 0.06660635769367218, + 0.07971552014350891, + 0.06928014010190964, + 0.1842813640832901, + 0.12699703872203827, + 0.03482545539736748, + 0.03683806210756302, + 0.05548792704939842, + 0.03417455777525902 + ], + "credit_map": [ + 0.03379929065704346, + 0.05837441608309746, + 0.021701620891690254, + 0.02304553985595703, + 0.020997660234570503, + 0.02838645689189434, + 0.031236207112669945, + 0.09425236284732819, + 0.008427261374890804, + 0.006961022038012743, + 0.004808611702173948, + 0.007841499522328377, + 0.004259823821485043, + 0.006404377985745668, + 0.008300709538161755, + 0.0196030642837286, + 0.008858893997967243, + 0.008775358088314533, + 0.007855299860239029, + 0.012330437079071999, + 0.006883670575916767, + 0.007853937335312366, + 0.016477802768349648, + 0.010680122300982475, + 0.009873627685010433, + 0.006535361986607313, + 0.009265310131013393, + 0.010138111189007759, + 0.009384420700371265, + 0.007237912621349096, + 0.01011624000966549, + 0.006729161832481623, + 0.054932862520217896, + 0.009096049703657627, + 0.028022246435284615, + 0.010086442343890667, + 0.011997547931969166, + 0.0538460798561573, + 0.00842114258557558, + 0.007879004813730717, + 0.010848008096218109, + 0.009033090434968472, + 0.038818009197711945, + 0.016980350017547607, + 0.011710979975759983, + 0.011291960254311562, + 0.009835589677095413, + 0.01847907155752182, + 0.0029404021333903074, + 0.007613265886902809, + 0.003515016520395875, + 0.004288730211555958, + 0.006456645205616951, + 0.002851484576240182, + 0.00455582607537508, + 0.0026040857192128897, + 0.003474664408713579, + 0.004117226228117943, + 0.006220847833901644, + 0.008135272189974785, + 0.004276486579328775, + 0.003986736293882132, + 0.0033902113791555166, + 0.0032366185914725065, + 0.004450824111700058, + 0.011090016923844814, + 0.004658326506614685, + 0.008594186045229435, + 0.008065794594585896, + 0.006165874656289816, + 0.006549281999468803, + 0.005913627799600363, + 0.004046278074383736, + 0.005663017276674509, + 0.0037049008533358574, + 0.006733683869242668, + 0.005086295306682587, + 0.0023208826314657927, + 0.004600389860570431, + 0.002019113628193736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.011856324970722198 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211048372089863 + ], + "opposite_logprob_tokens": [ + -4.264211177825928 + ], + "logprob_reward": -0.014211048372089863, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211048372089863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35391315817832947, + 0.05907293036580086, + 0.0816989466547966, + 0.08972861617803574, + 0.05435162037611008, + 0.12562324106693268, + 0.047703083604574203, + 0.05413494259119034, + 0.08576605468988419, + 0.04800739511847496 + ], + "credit_map": [ + 0.04851749911904335, + 0.04619557783007622, + 0.022634971886873245, + 0.02551688440144062, + 0.03001936338841915, + 0.02880541794002056, + 0.037600480020046234, + 0.11462296545505524, + 0.006390400696545839, + 0.006583345588296652, + 0.0037200534716248512, + 0.006281342823058367, + 0.006216664332896471, + 0.006305739749222994, + 0.008870197460055351, + 0.01470518670976162, + 0.009141712449491024, + 0.00551054859533906, + 0.008486537262797356, + 0.011959495022892952, + 0.005993620492517948, + 0.006580182816833258, + 0.022343240678310394, + 0.011683604680001736, + 0.007635558024048805, + 0.00935165211558342, + 0.010955802164971828, + 0.01661527156829834, + 0.010939172469079494, + 0.008297041989862919, + 0.016760477796196938, + 0.009173637256026268, + 0.007889417931437492, + 0.005139438901096582, + 0.0058101811446249485, + 0.006287206429988146, + 0.007108592893928289, + 0.011016358621418476, + 0.004673709627240896, + 0.006426715757697821, + 0.011917609721422195, + 0.014189112931489944, + 0.023125991225242615, + 0.017069736495614052, + 0.012437532655894756, + 0.012997514568269253, + 0.010639195330440998, + 0.02324654534459114, + 0.0042103007435798645, + 0.008464515209197998, + 0.0075058722868561745, + 0.009066803380846977, + 0.006642336491495371, + 0.0032594462390989065, + 0.005013167858123779, + 0.0035406379029154778, + 0.006298174150288105, + 0.007965398952364922, + 0.005490011535584927, + 0.007000408135354519, + 0.0077714561484754086, + 0.007697513792663813, + 0.00665489723905921, + 0.005257082637399435, + 0.005691047292202711, + 0.01583343930542469, + 0.006671115290373564, + 0.02533586323261261, + 0.01213866751641035, + 0.0053091454319655895, + 0.010552934370934963, + 0.004233842249959707, + 0.0040606665425002575, + 0.005290116183459759, + 0.010054167360067368, + 0.009518545120954514, + 0.0052574919536709785, + 0.0044949846342206, + 0.006269994657486677, + 0.0030614258721470833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.014211048372089863 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783178567886353 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.009783178567886353, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783178567886353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19978757202625275, + 0.05660250037908554, + 0.18298819661140442, + 0.027921853587031364, + 0.10168959200382233, + 0.14564143121242523, + 0.10819655656814575, + 0.06118156760931015, + 0.06149762496352196, + 0.05449312925338745 + ], + "credit_map": [ + 0.02911395952105522, + 0.02553553879261017, + 0.015585241839289665, + 0.014063136652112007, + 0.013660653494298458, + 0.017454760149121284, + 0.020401425659656525, + 0.0639728531241417, + 0.009126711636781693, + 0.006772228982299566, + 0.005082435440272093, + 0.005533549003303051, + 0.005753352772444487, + 0.006080488208681345, + 0.00731814606115222, + 0.01093558594584465, + 0.023275917395949364, + 0.029916593804955482, + 0.016900142654776573, + 0.028417833149433136, + 0.018101723864674568, + 0.024690497666597366, + 0.021455984562635422, + 0.020229503512382507, + 0.0024190472904592752, + 0.0037005115300416946, + 0.003053343389183283, + 0.004677126184105873, + 0.003953262697905302, + 0.0029534155037254095, + 0.004506430588662624, + 0.0026587159372866154, + 0.02452602982521057, + 0.006621988024562597, + 0.022665930911898613, + 0.008317465893924236, + 0.005950424820184708, + 0.018713945522904396, + 0.008092464879155159, + 0.006801331881433725, + 0.017084378749132156, + 0.01154390163719654, + 0.05041297525167465, + 0.03464953228831291, + 0.009546125307679176, + 0.006031906232237816, + 0.01084045134484768, + 0.005532165057957172, + 0.008628696203231812, + 0.011864698491990566, + 0.01692325808107853, + 0.031596723943948746, + 0.011339611373841763, + 0.00675169425085187, + 0.014839799143373966, + 0.006252075545489788, + 0.004460599273443222, + 0.008426431566476822, + 0.011564678512513638, + 0.016200073063373566, + 0.006290973164141178, + 0.004397114273160696, + 0.006152615416795015, + 0.0036890830378979445, + 0.006098732817918062, + 0.009045823477208614, + 0.005473316181451082, + 0.015334868803620338, + 0.006104066967964172, + 0.006812747567892075, + 0.007172676268965006, + 0.005455386824905872, + 0.004479461815208197, + 0.006882563699036837, + 0.008765310049057007, + 0.011030618101358414, + 0.006698933430016041, + 0.003919463604688644, + 0.00931396521627903, + 0.0034028124064207077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.009783178567886353 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076242173090577126 + ], + "opposite_logprob_tokens": [ + -4.88262414932251 + ], + "logprob_reward": -0.0076242173090577126, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076242173090577126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22804278135299683, + 0.0463426373898983, + 0.309637188911438, + 0.041096195578575134, + 0.04378261789679527, + 0.1904236227273941, + 0.05462144687771797, + 0.027517100796103477, + 0.03417566418647766, + 0.024360762909054756 + ], + "credit_map": [ + 0.02815263718366623, + 0.032876014709472656, + 0.01641480252146721, + 0.016122980043292046, + 0.013718601316213608, + 0.01814911887049675, + 0.024544937536120415, + 0.07806368172168732, + 0.008723725564777851, + 0.00678436690941453, + 0.003950172569602728, + 0.004065897781401873, + 0.005976313725113869, + 0.007706847507506609, + 0.004144642502069473, + 0.004990673623979092, + 0.04016612470149994, + 0.05015748366713524, + 0.03008878044784069, + 0.04647782817482948, + 0.029085077345371246, + 0.04376134276390076, + 0.0368429571390152, + 0.0330575592815876, + 0.0040962849743664265, + 0.005314421374350786, + 0.003991244826465845, + 0.005562928505241871, + 0.009353842586278915, + 0.005528867244720459, + 0.004061211831867695, + 0.0031873921398073435, + 0.009374662302434444, + 0.004567213822156191, + 0.012037977576255798, + 0.004136498551815748, + 0.003992476500570774, + 0.0039986660704016685, + 0.0024160922039300203, + 0.0032590290065854788, + 0.0191662460565567, + 0.009655947797000408, + 0.0653584748506546, + 0.025980183854699135, + 0.013418890535831451, + 0.015234190039336681, + 0.014980992302298546, + 0.026628702878952026, + 0.005079271737486124, + 0.006225507706403732, + 0.007416980341076851, + 0.006424239836633205, + 0.0059276968240737915, + 0.009072225540876389, + 0.008045596070587635, + 0.0064299278892576694, + 0.002688591368496418, + 0.005468061659485102, + 0.0033456268720328808, + 0.0035108753945678473, + 0.004309693351387978, + 0.0021429990883916616, + 0.003702262183651328, + 0.0023489894811064005, + 0.0029650956857949495, + 0.004171921405941248, + 0.003211934119462967, + 0.01046538632363081, + 0.00286190421320498, + 0.0035721873864531517, + 0.0035829576663672924, + 0.0033442778512835503, + 0.0026805780362337828, + 0.0029884129762649536, + 0.003785062115639448, + 0.0038439834024757147, + 0.0028812880627810955, + 0.0027448295149952173, + 0.0036863794084638357, + 0.001750229042954743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.0076242173090577126 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000136375427246 + ], + "opposite_logprob_tokens": [ + -2.6950013637542725 + ], + "logprob_reward": -0.07000136375427246, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000136375427246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31553182005882263, + 0.04484511539340019, + 0.2783108353614807, + 0.03208252415060997, + 0.07674416899681091, + 0.10037488490343094, + 0.07911909371614456, + 0.02590499445796013, + 0.03348411247134209, + 0.013602441176772118 + ], + "credit_map": [ + 0.036439504474401474, + 0.048431966453790665, + 0.021500715985894203, + 0.021757490932941437, + 0.020300576463341713, + 0.02668544463813305, + 0.0334409736096859, + 0.10697518289089203, + 0.007161607500165701, + 0.005638156086206436, + 0.005724233575165272, + 0.005311108659952879, + 0.0044167437590658665, + 0.004601925145834684, + 0.004650678485631943, + 0.007340665906667709, + 0.035096000880002975, + 0.04438282921910286, + 0.026315519586205482, + 0.04452694207429886, + 0.025540530681610107, + 0.03697751462459564, + 0.03351230174303055, + 0.031959209591150284, + 0.0036471544299274683, + 0.004991290159523487, + 0.004303041845560074, + 0.0035893497988581657, + 0.004260289017111063, + 0.004940317012369633, + 0.0032668085768818855, + 0.0030842721462249756, + 0.019008653238415718, + 0.004736337810754776, + 0.016091275960206985, + 0.005957295652478933, + 0.006920676212757826, + 0.01595872826874256, + 0.003753901459276676, + 0.004317308310419321, + 0.008211221545934677, + 0.0076141622848808765, + 0.011906898580491543, + 0.019089601933956146, + 0.00613307673484087, + 0.013960836455225945, + 0.009271090850234032, + 0.024188000708818436, + 0.007045503705739975, + 0.007916777394711971, + 0.014061002060770988, + 0.02255231700837612, + 0.006637262180447578, + 0.005946761462837458, + 0.00938057154417038, + 0.005578904878348112, + 0.0017733876593410969, + 0.0028061107732355595, + 0.006108778063207865, + 0.007484826259315014, + 0.0020873621106147766, + 0.0019003512570634484, + 0.002087649889290333, + 0.0016565296100452542, + 0.0022862167097628117, + 0.005960751324892044, + 0.00256998254917562, + 0.008474810980260372, + 0.004917872603982687, + 0.002790043130517006, + 0.00431784288957715, + 0.0021665950771421194, + 0.0018476854311302304, + 0.0015089677181094885, + 0.001574049238115549, + 0.00206558289937675, + 0.001446546521037817, + 0.002259394619613886, + 0.0016287913313135505, + 0.0012714245822280645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.07000136375427246 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a slightly upward angle, slowly but steadily advances, first glimpsing a vast, bizarre plant leaf in the foreground. Its deep, dark green, velvety surface is covered with intricate, luminous veins, resembling circuit boards, faintly luminescent. As the camera deepens, focus gradually settles on an Asian model nestled within a futuristic greenhouse. Her face is sculpted, her long, wet black hair neatly combed back. Her eyes, cold and distant, stare blankly into the camera. She wears a couture dress crafted from a liquid metal silver fabric with dramatic, architectural shoulders, making her stand out against the backdrop. Behind her, rows of stainless steel and glass vertical growth racks display a variety of exotic plants, each a psychedelic display of vibrant colors under alternating purple and magenta LED grow lights. The lights pulsated in a slow and rhythmic rhythm, periodically sweeping across her expressionless face and the glittering surface of the clothing, casting ever-changing long shadows. The entire scene was shrouded in a cold tone, and the precise coldness of technology and the ultimate beauty of luxury fashion blended at this moment, creating a fashion blockbuster atmosphere full of mystery and visual impact.", + "seed": 12048, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_4/index_5809_seed_12048_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020787741150707008, + "vlm_logprob": -0.020787741150707008, + "vlm_yes_no_margin": 4.3249999523162845, + "vlm_token_credit": -0.020787741150707008, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017307909205555916 + ], + "opposite_logprob_tokens": [ + -4.767307758331299 + ], + "logprob_reward": -0.017307909205555916, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.017307909205555916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24024876952171326, + 0.04997652396559715, + 0.04144233837723732, + 0.0228896401822567, + 0.030720287933945656, + 0.10892657935619354, + 0.2865593433380127, + 0.08886474370956421, + 0.08782302588224411, + 0.042548730969429016 + ], + "credit_map": [ + 0.0198588278144598, + 0.01835521124303341, + 0.04729620739817619, + 0.04223267734050751, + 0.022322680801153183, + 0.015373384580016136, + 0.04286627471446991, + 0.03194349631667137, + 0.004648612346500158, + 0.0069898865185678005, + 0.006776172202080488, + 0.007250060793012381, + 0.00479391124099493, + 0.007579104974865913, + 0.007103680167347193, + 0.004835093859583139, + 0.0033499300479888916, + 0.003975288942456245, + 0.009303317405283451, + 0.0076673892326653, + 0.00351160136051476, + 0.004440824501216412, + 0.005272591952234507, + 0.0039213914424180984, + 0.0020148009061813354, + 0.0027591711841523647, + 0.0035012676380574703, + 0.0029383401852101088, + 0.0036522289738059044, + 0.0033592209219932556, + 0.002338998718187213, + 0.002325613982975483, + 0.0030608815141022205, + 0.004452574998140335, + 0.0045965248718857765, + 0.00425894558429718, + 0.003044278360903263, + 0.003894287161529064, + 0.0039744907990098, + 0.003438305575400591, + 0.029824044555425644, + 0.009482272900640965, + 0.005766067188233137, + 0.022001320496201515, + 0.01077011227607727, + 0.006415392737835646, + 0.007964902557432652, + 0.01670246198773384, + 0.07558799535036087, + 0.019042082130908966, + 0.015343536622822285, + 0.11088531464338303, + 0.011849015019834042, + 0.014763842336833477, + 0.019049562513828278, + 0.020037978887557983, + 0.013894871808588505, + 0.007735937833786011, + 0.02928573451936245, + 0.009621748700737953, + 0.0124506251886487, + 0.004611922428011894, + 0.0063310591503977776, + 0.004932849667966366, + 0.00708771450445056, + 0.011003930121660233, + 0.018164467066526413, + 0.015089857392013073, + 0.00892789475619793, + 0.009662676602602005, + 0.012100279331207275, + 0.005786203779280186, + 0.005232553463429213, + 0.005962401162832975, + 0.009718598797917366, + 0.0049046715721488, + 0.0032471301965415478, + 0.003882416756823659, + 0.007074296474456787, + 0.002526665572077036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.017307909205555916 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821153610944748 + ], + "opposite_logprob_tokens": [ + -4.018211364746094 + ], + "logprob_reward": -0.01821153610944748, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01821153610944748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14558754861354828, + 0.11460389941930771, + 0.12169340252876282, + 0.03345652297139168, + 0.04907034710049629, + 0.1271272599697113, + 0.1321077197790146, + 0.21589334309101105, + 0.025737952440977097, + 0.03472207486629486 + ], + "credit_map": [ + 0.011527537368237972, + 0.014135433360934258, + 0.024267615750432014, + 0.03672652691602707, + 0.011734487488865852, + 0.010375605896115303, + 0.022979728877544403, + 0.013840613886713982, + 0.010614384897053242, + 0.018848884850740433, + 0.014905262738466263, + 0.014380879700183868, + 0.010885342955589294, + 0.01273064874112606, + 0.01476973481476307, + 0.017468756064772606, + 0.010447896085679531, + 0.005303715355694294, + 0.05079491063952446, + 0.018798263743519783, + 0.0047563668340444565, + 0.014484022743999958, + 0.009156341664493084, + 0.007951881736516953, + 0.003137874649837613, + 0.004312252160161734, + 0.00852907169610262, + 0.0034929155372083187, + 0.003972225822508335, + 0.00382808200083673, + 0.00318631948903203, + 0.0029977811500430107, + 0.004559087101370096, + 0.007350075989961624, + 0.005461167078465223, + 0.006754715461283922, + 0.0059650675393640995, + 0.005851768422871828, + 0.008447039872407913, + 0.0046814256347715855, + 0.022069906815886497, + 0.017570728436112404, + 0.008734005503356457, + 0.014837252907454967, + 0.020515503361821175, + 0.008338007144629955, + 0.008930879645049572, + 0.026130976155400276, + 0.033929672092199326, + 0.009727472439408302, + 0.007437574677169323, + 0.05004320293664932, + 0.005760074593126774, + 0.006489890627563, + 0.010551896877586842, + 0.008167929016053677, + 0.0353950634598732, + 0.02732640504837036, + 0.051636919379234314, + 0.03019869700074196, + 0.025635505095124245, + 0.013932236470282078, + 0.018461963161826134, + 0.013306565582752228, + 0.0033621955662965775, + 0.0026424468960613012, + 0.0024275314062833786, + 0.0034390282817184925, + 0.004284463357180357, + 0.003139532171189785, + 0.0036886471789330244, + 0.0027541066519916058, + 0.0036599580198526382, + 0.0036810734309256077, + 0.00584966829046607, + 0.004919335711747408, + 0.0028492019046097994, + 0.002783899661153555, + 0.008398663252592087, + 0.0025802780874073505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.01821153610944748 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739978212863207 + ], + "opposite_logprob_tokens": [ + -5.006740093231201 + ], + "logprob_reward": -0.006739978212863207, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739978212863207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05755683034658432, + 0.11353825032711029, + 0.08956941962242126, + 0.012775029987096786, + 0.05410338193178177, + 0.4841834306716919, + 0.07645120471715927, + 0.02858017012476921, + 0.06003495305776596, + 0.023207316175103188 + ], + "credit_map": [ + 0.004559854976832867, + 0.006576082203537226, + 0.007357296068221331, + 0.012773462571203709, + 0.005882875062525272, + 0.005060040857642889, + 0.01019549835473299, + 0.005151723511517048, + 0.011304196901619434, + 0.016051797196269035, + 0.014513523317873478, + 0.01796490140259266, + 0.012998160906136036, + 0.014857173897325993, + 0.011338434182107449, + 0.014510068111121655, + 0.00778244249522686, + 0.004063482396304607, + 0.037553027272224426, + 0.01415923424065113, + 0.0031225821003317833, + 0.010037507861852646, + 0.006646751891821623, + 0.006204404402524233, + 0.0014357793843373656, + 0.001851719687692821, + 0.0025339836720377207, + 0.0015405588783323765, + 0.0011125978780910373, + 0.001533161150291562, + 0.0014537718379870057, + 0.0013134583132341504, + 0.005961337126791477, + 0.008851914666593075, + 0.007578544784337282, + 0.00849408283829689, + 0.004916687961667776, + 0.005975148640573025, + 0.0064643085934221745, + 0.005861363839358091, + 0.042139001190662384, + 0.03151465952396393, + 0.02073799818754196, + 0.17410580813884735, + 0.02482604794204235, + 0.034048788249492645, + 0.03580814599990845, + 0.12100297212600708, + 0.018085751682519913, + 0.006232310086488724, + 0.004875580780208111, + 0.02100931480526924, + 0.006142539903521538, + 0.006448328495025635, + 0.006482129916548729, + 0.007175257429480553, + 0.005792609881609678, + 0.0029327936936169863, + 0.0047238231636583805, + 0.003268790664151311, + 0.0047720675356686115, + 0.0018937793793156743, + 0.0031963041983544827, + 0.0020000054500997066, + 0.005254133604466915, + 0.006287897005677223, + 0.01920158602297306, + 0.010570532642304897, + 0.005019553937017918, + 0.0031087512616068125, + 0.005448553711175919, + 0.0051439497619867325, + 0.0022077353205531836, + 0.0030014757066965103, + 0.0028848156798630953, + 0.003381005022674799, + 0.003642347641289234, + 0.0022201319225132465, + 0.0045152707025408745, + 0.00135453708935529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.006739978212863207 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735478527843952 + ], + "opposite_logprob_tokens": [ + -5.006735324859619 + ], + "logprob_reward": -0.006735478527843952, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735478527843952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058359649032354355, + 0.2063559740781784, + 0.03674110025167465, + 0.019127147272229195, + 0.05532772094011307, + 0.2437995821237564, + 0.26324430108070374, + 0.05092838034033775, + 0.01866595447063446, + 0.04745013266801834 + ], + "credit_map": [ + 0.005591448396444321, + 0.005019296891987324, + 0.013130316510796547, + 0.00737629272043705, + 0.0053796120919287205, + 0.00550876697525382, + 0.0063151526264846325, + 0.010038762353360653, + 0.023235291242599487, + 0.029368931427598, + 0.025928940623998642, + 0.023761047050356865, + 0.024525152519345284, + 0.029277972877025604, + 0.025509227067232132, + 0.024749422445893288, + 0.0032256210688501596, + 0.0032515667844563723, + 0.011423624120652676, + 0.005379262380301952, + 0.0023888791911303997, + 0.00514876376837492, + 0.002796484623104334, + 0.0031268978491425514, + 0.0016709562623873353, + 0.0021726980339735746, + 0.0034492183476686478, + 0.0025071529671549797, + 0.002344256965443492, + 0.002250532852485776, + 0.002104763174429536, + 0.002627567620947957, + 0.006904246751219034, + 0.008589182049036026, + 0.01080204825848341, + 0.007101273164153099, + 0.004309607204049826, + 0.0052994610741734505, + 0.006046307273209095, + 0.006275595165789127, + 0.025431104004383087, + 0.013931850902736187, + 0.010804779827594757, + 0.0922030657529831, + 0.007359385024756193, + 0.017228178679943085, + 0.019136788323521614, + 0.05770444497466087, + 0.06639796495437622, + 0.020753750577569008, + 0.01494236383587122, + 0.09848322719335556, + 0.013076342642307281, + 0.014035152271389961, + 0.018245531246066093, + 0.017309967428445816, + 0.004346577916294336, + 0.007483850233256817, + 0.00587253924459219, + 0.009737364947795868, + 0.009185540489852428, + 0.004368730820715427, + 0.005994920153170824, + 0.003938856068998575, + 0.0016423523193225265, + 0.0030155922286212444, + 0.0021271882578730583, + 0.0033473530784249306, + 0.0026207787450402975, + 0.0022996217012405396, + 0.001935856998898089, + 0.0016772110247984529, + 0.0050509036518633366, + 0.006792806554585695, + 0.006111857481300831, + 0.010069122537970543, + 0.0032069466542452574, + 0.006862699054181576, + 0.005823890678584576, + 0.0035319027956575155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.006735478527843952 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494380369782448 + ], + "opposite_logprob_tokens": [ + -2.929943799972534 + ], + "logprob_reward": -0.05494380369782448, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494380369782448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061723656952381134, + 0.07669517397880554, + 0.12361658364534378, + 0.024293864145874977, + 0.06635235995054245, + 0.2558042109012604, + 0.16227427124977112, + 0.18113116919994354, + 0.02304844744503498, + 0.025060264393687248 + ], + "credit_map": [ + 0.004606474190950394, + 0.00689314492046833, + 0.011475575156509876, + 0.010038384236395359, + 0.005535766948014498, + 0.006273157894611359, + 0.011128215119242668, + 0.0057729389518499374, + 0.007423451170325279, + 0.012121643871068954, + 0.010734861716628075, + 0.010220887139439583, + 0.00866827555000782, + 0.00848662480711937, + 0.008095765486359596, + 0.010943661443889141, + 0.010757076554000378, + 0.006816566921770573, + 0.046397995203733444, + 0.01973740942776203, + 0.006321542896330357, + 0.013883673585951328, + 0.010666216723620892, + 0.009036106988787651, + 0.0018060323782265186, + 0.0033443027641624212, + 0.004617564845830202, + 0.003393367165699601, + 0.0028457853477448225, + 0.003125926246866584, + 0.0024906189646571875, + 0.0026702687609940767, + 0.006878240965306759, + 0.009271749295294285, + 0.012612647376954556, + 0.007000473327934742, + 0.009640146978199482, + 0.00640782481059432, + 0.007885592989623547, + 0.006655683275312185, + 0.027310645207762718, + 0.019281527027487755, + 0.01357499323785305, + 0.0912894606590271, + 0.014575278386473656, + 0.014445903711020947, + 0.01938425377011299, + 0.055942147970199585, + 0.0408780612051487, + 0.010773791931569576, + 0.009222030639648438, + 0.06214766576886177, + 0.0077075171284377575, + 0.006929503753781319, + 0.010463573038578033, + 0.014152138493955135, + 0.03188906982541084, + 0.021100567653775215, + 0.044734831899404526, + 0.022283971309661865, + 0.023777920752763748, + 0.01143181137740612, + 0.014948873780667782, + 0.010964132845401764, + 0.002493220614269376, + 0.002329454990103841, + 0.002891118172556162, + 0.0032010965514928102, + 0.0035223597660660744, + 0.002365129068493843, + 0.003002016106620431, + 0.0032440535724163055, + 0.002786630764603615, + 0.0045173377729952335, + 0.002852723700925708, + 0.0023665567860007286, + 0.0028542065992951393, + 0.0024631889536976814, + 0.005680886097252369, + 0.0015387352323159575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.05494380369782448 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a slightly upward angle, slowly but steadily advances, first glimpsing a vast, bizarre plant leaf in the foreground. Its deep, dark green, velvety surface is covered with intricate, luminous veins, resembling circuit boards, faintly luminescent. As the camera deepens, focus gradually settles on an Asian model nestled within a futuristic greenhouse. Her face is sculpted, her long, wet black hair neatly combed back. Her eyes, cold and distant, stare blankly into the camera. She wears a couture dress crafted from a liquid metal silver fabric with dramatic, architectural shoulders, making her stand out against the backdrop. Behind her, rows of stainless steel and glass vertical growth racks display a variety of exotic plants, each a psychedelic display of vibrant colors under alternating purple and magenta LED grow lights. The lights pulsated in a slow and rhythmic rhythm, periodically sweeping across her expressionless face and the glittering surface of the clothing, casting ever-changing long shadows. The entire scene was shrouded in a cold tone, and the precise coldness of technology and the ultimate beauty of luxury fashion blended at this moment, creating a fashion blockbuster atmosphere full of mystery and visual impact.", + "seed": 12049, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_4/index_5809_seed_12049_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019139057770371438, + "vlm_logprob": -0.019139057770371438, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.019139057770371438, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010525546967983246 + ], + "opposite_logprob_tokens": [ + -5.260525703430176 + ], + "logprob_reward": -0.010525546967983246, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.010525546967983246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3815959095954895, + 0.17504927515983582, + 0.0493859201669693, + 0.05833152309060097, + 0.03566589578986168, + 0.05368981882929802, + 0.0387362502515316, + 0.1742609739303589, + 0.018551159650087357, + 0.014733296819031239 + ], + "credit_map": [ + 0.029320767149329185, + 0.12213779240846634, + 0.10676436871290207, + 0.02918696030974388, + 0.03723985701799393, + 0.016460377722978592, + 0.02715110033750534, + 0.013334674760699272, + 0.013051111251115799, + 0.010726080276072025, + 0.06225594878196716, + 0.011294104158878326, + 0.009176847524940968, + 0.008482684381306171, + 0.008318673819303513, + 0.05174381285905838, + 0.003759360406547785, + 0.003634451422840357, + 0.01628805883228779, + 0.005150365177541971, + 0.0065197208896279335, + 0.004029847215861082, + 0.006702134385704994, + 0.0033019795082509518, + 0.0167121272534132, + 0.00891741644591093, + 0.005465895868837833, + 0.003783224616199732, + 0.006291553843766451, + 0.007538933306932449, + 0.004470831248909235, + 0.0051515367813408375, + 0.0040296101942658424, + 0.004058870952576399, + 0.0036929065827280283, + 0.004301078151911497, + 0.004984878469258547, + 0.00412670848891139, + 0.006690875627100468, + 0.0037809640634804964, + 0.004635064397007227, + 0.008909477852284908, + 0.003396463580429554, + 0.005852408241480589, + 0.0038921916857361794, + 0.009222100488841534, + 0.005558657459914684, + 0.01222345419228077, + 0.003011280670762062, + 0.00732808280736208, + 0.0022606959100812674, + 0.005715617910027504, + 0.003974582068622112, + 0.003016543574631214, + 0.010212036781013012, + 0.0032174091320484877, + 0.010436584241688251, + 0.009079309180378914, + 0.05419071018695831, + 0.058861855417490005, + 0.006751668639481068, + 0.007604803889989853, + 0.01845969446003437, + 0.008876333013176918, + 0.0018626535311341286, + 0.0017904142150655389, + 0.0020414560567587614, + 0.0025663529522717, + 0.003091000020503998, + 0.0019811568781733513, + 0.0027008438482880592, + 0.0025172820314764977, + 0.0023548442404717207, + 0.0011168377241119742, + 0.0020714174024760723, + 0.001372089609503746, + 0.0013045050436630845, + 0.0026336589362472296, + 0.002090862486511469, + 0.0017890803283080459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.010525546967983246 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012554659508168697 + ], + "opposite_logprob_tokens": [ + -4.38755464553833 + ], + "logprob_reward": -0.012554659508168697, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012554659508168697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4029352068901062, + 0.08928241580724716, + 0.046831127256155014, + 0.0610753670334816, + 0.0517243891954422, + 0.04972114786505699, + 0.09303759038448334, + 0.12819194793701172, + 0.039751727133989334, + 0.03744906187057495 + ], + "credit_map": [ + 0.030910637229681015, + 0.14149366319179535, + 0.11520785093307495, + 0.026635782793164253, + 0.0429149828851223, + 0.015289543196558952, + 0.017235195264220238, + 0.013247576542198658, + 0.006834074854850769, + 0.008556404151022434, + 0.02335166372358799, + 0.006632602773606777, + 0.006970575544983149, + 0.005881361197680235, + 0.009691503830254078, + 0.0213642418384552, + 0.004774711560457945, + 0.004585266578942537, + 0.010261155664920807, + 0.0051759216003119946, + 0.0045567890629172325, + 0.004986523184925318, + 0.006734553258866072, + 0.0057562110014259815, + 0.010906792245805264, + 0.008165813982486725, + 0.00508522242307663, + 0.005142452195286751, + 0.0077111925929784775, + 0.009758476167917252, + 0.007074518129229546, + 0.00723090348765254, + 0.0054963803850114346, + 0.004714989569038153, + 0.0048352633602917194, + 0.005838810466229916, + 0.006819711998105049, + 0.006576589774340391, + 0.009248176589608192, + 0.008194473572075367, + 0.003954126499593258, + 0.005934643093496561, + 0.004058742430061102, + 0.005707444157451391, + 0.006822018418461084, + 0.006643582601100206, + 0.00711574824526906, + 0.009484846144914627, + 0.007039519492536783, + 0.01961798593401909, + 0.0059464056976139545, + 0.012770182453095913, + 0.010098312050104141, + 0.007057929877191782, + 0.023121679201722145, + 0.007385585457086563, + 0.007146973628550768, + 0.009226216934621334, + 0.022072680294513702, + 0.037801556289196014, + 0.012824581004679203, + 0.005379632115364075, + 0.023439619690179825, + 0.010300683788955212, + 0.0028821625746786594, + 0.0033947729971259832, + 0.003200245788320899, + 0.005670230835676193, + 0.00877712108194828, + 0.003887210739776492, + 0.005961699411273003, + 0.005978286266326904, + 0.00417854031547904, + 0.0036356456112116575, + 0.007712112739682198, + 0.0033881233539432287, + 0.005840002559125423, + 0.003938810434192419, + 0.004266237374395132, + 0.0044895922765135765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.012554659508168697 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639100283384323 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.004639100283384323, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639100283384323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3773091733455658, + 0.14993155002593994, + 0.0702739879488945, + 0.06606503576040268, + 0.05896386504173279, + 0.03974945470690727, + 0.05876011773943901, + 0.09449973702430725, + 0.04098586365580559, + 0.04346119984984398 + ], + "credit_map": [ + 0.03134546056389809, + 0.11598605662584305, + 0.10423747450113297, + 0.024235550314188004, + 0.046143755316734314, + 0.018051499500870705, + 0.017604628577828407, + 0.019704744219779968, + 0.019196748733520508, + 0.020682871341705322, + 0.013275282457470894, + 0.014939296059310436, + 0.01715564727783203, + 0.024456577375531197, + 0.01697809249162674, + 0.02324703522026539, + 0.007889898493885994, + 0.007379062473773956, + 0.013525137677788734, + 0.007376716006547213, + 0.008010875433683395, + 0.008032424375414848, + 0.009728712029755116, + 0.00833116378635168, + 0.018516287207603455, + 0.009352250955998898, + 0.006422099191695452, + 0.005263302009552717, + 0.006961853243410587, + 0.008933779783546925, + 0.004889082163572311, + 0.005726380739361048, + 0.004648841451853514, + 0.009720868431031704, + 0.003621176350861788, + 0.005615459755063057, + 0.00806900579482317, + 0.0066737160086631775, + 0.010727737098932266, + 0.009887062013149261, + 0.003207866568118334, + 0.0038970657624304295, + 0.0035109540913254023, + 0.00632708054035902, + 0.0043597351759672165, + 0.005271498113870621, + 0.007792262826114893, + 0.005382991395890713, + 0.004853665828704834, + 0.014938746578991413, + 0.004145125392824411, + 0.008696078322827816, + 0.007449548691511154, + 0.004673217423260212, + 0.009186415001749992, + 0.004817318636924028, + 0.006233998108655214, + 0.006861505564302206, + 0.022918233647942543, + 0.0262339785695076, + 0.007815975695848465, + 0.0056108273565769196, + 0.012938350439071655, + 0.005886862985789776, + 0.0038724690675735474, + 0.003616366069763899, + 0.0031669698655605316, + 0.007335674948990345, + 0.007079962641000748, + 0.0031837287824600935, + 0.007711705751717091, + 0.005018983036279678, + 0.005136358086019754, + 0.004760532174259424, + 0.00458221510052681, + 0.005851673427969217, + 0.006500424351543188, + 0.005488683003932238, + 0.007323647849261761, + 0.003817666321992874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.004639100283384323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059457337483763695 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.0059457337483763695, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059457337483763695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4851657450199127, + 0.07150009274482727, + 0.0456724613904953, + 0.049868352711200714, + 0.05208197608590126, + 0.026911189779639244, + 0.07947209477424622, + 0.06909094005823135, + 0.08194239437580109, + 0.03829478845000267 + ], + "credit_map": [ + 0.03957059234380722, + 0.16047506034374237, + 0.12536194920539856, + 0.03169098496437073, + 0.05614199489355087, + 0.025295434519648552, + 0.027310919016599655, + 0.019318798556923866, + 0.006270477548241615, + 0.005949271377176046, + 0.021246738731861115, + 0.004194058943539858, + 0.004655685741454363, + 0.008376861922442913, + 0.00408573355525732, + 0.016721263527870178, + 0.004426511470228434, + 0.004301717504858971, + 0.008403764106333256, + 0.004980136174708605, + 0.0076555912382900715, + 0.0056480588391423225, + 0.0048299129121005535, + 0.005426769610494375, + 0.01257286872714758, + 0.007323260884732008, + 0.0044415658339858055, + 0.0042718094773590565, + 0.005214575678110123, + 0.006079498212784529, + 0.005389160942286253, + 0.004575615748763084, + 0.0050840480253100395, + 0.00819377787411213, + 0.004124267026782036, + 0.005960862617939711, + 0.005833860486745834, + 0.005140089895576239, + 0.01050109788775444, + 0.007243973668664694, + 0.0021613931749016047, + 0.0032490461599081755, + 0.0018004438607022166, + 0.0037159398198127747, + 0.004489024635404348, + 0.003334772540256381, + 0.004480671137571335, + 0.0036799004301428795, + 0.006525931414216757, + 0.011334308423101902, + 0.003885325975716114, + 0.011613664217293262, + 0.008423739112913609, + 0.005674623418599367, + 0.02489406429231167, + 0.007120437920093536, + 0.005557470489293337, + 0.008487343788146973, + 0.015439197421073914, + 0.009622427634894848, + 0.00834652129560709, + 0.00426914868876338, + 0.012083243578672409, + 0.005285583436489105, + 0.009444517083466053, + 0.006448807660490274, + 0.008114319294691086, + 0.011519869789481163, + 0.013983463868498802, + 0.009720414876937866, + 0.009966505691409111, + 0.012744488194584846, + 0.004964595194905996, + 0.003222634783014655, + 0.010130934417247772, + 0.003554247785359621, + 0.0038891362491995096, + 0.0037746019661426544, + 0.004270672332495451, + 0.004487965721637011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.0059457337483763695 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203024834394455 + ], + "opposite_logprob_tokens": [ + -2.81203031539917 + ], + "logprob_reward": -0.06203024834394455, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203024834394455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.510989248752594, + 0.11599589884281158, + 0.030474266037344933, + 0.029848173260688782, + 0.022998059168457985, + 0.06157822161912918, + 0.0443568080663681, + 0.11267063021659851, + 0.026550360023975372, + 0.04453837871551514 + ], + "credit_map": [ + 0.03727655112743378, + 0.18542371690273285, + 0.15911482274532318, + 0.027659540995955467, + 0.052980199456214905, + 0.01352085079997778, + 0.018595043569803238, + 0.01641847752034664, + 0.009619311429560184, + 0.010710248723626137, + 0.02680092863738537, + 0.008618982508778572, + 0.014632660895586014, + 0.006707508582621813, + 0.011070368811488152, + 0.02783588506281376, + 0.0034357868134975433, + 0.0031239143572747707, + 0.0055017187260091305, + 0.003712183330208063, + 0.004113067872822285, + 0.004036490805447102, + 0.0030734085012227297, + 0.0034776912070810795, + 0.0059967245906591415, + 0.003766352077946067, + 0.0034518868196755648, + 0.002835889346897602, + 0.003346828743815422, + 0.0047007836401462555, + 0.002688521286472678, + 0.0030611837282776833, + 0.0022214262280613184, + 0.0023178604897111654, + 0.0027068702038377523, + 0.003254730487242341, + 0.0029085336718708277, + 0.002144798869267106, + 0.005052756052464247, + 0.002391082001850009, + 0.005178769584745169, + 0.008757192641496658, + 0.0038450686261057854, + 0.007146480027586222, + 0.00599751528352499, + 0.01009426824748516, + 0.00710951117798686, + 0.013449412770569324, + 0.003594051580876112, + 0.008235960267484188, + 0.0022578081116080284, + 0.006099268794059753, + 0.003535620402544737, + 0.00303836353123188, + 0.014011751860380173, + 0.0035839774645864964, + 0.007556003052741289, + 0.006612738128751516, + 0.03366430848836899, + 0.035543378442525864, + 0.004966835491359234, + 0.005777495447546244, + 0.012690250761806965, + 0.005859618540853262, + 0.002914499258622527, + 0.0022495738230645657, + 0.002957426942884922, + 0.003932930529117584, + 0.0033966992050409317, + 0.003391082165762782, + 0.004311729222536087, + 0.0033964167814701796, + 0.00484916428104043, + 0.003656398504972458, + 0.013495616614818573, + 0.0035892801824957132, + 0.004659327678382397, + 0.0036040351260453463, + 0.005833436269313097, + 0.004851117730140686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.06203024834394455 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a slightly upward angle, slowly but steadily advances, first glimpsing a vast, bizarre plant leaf in the foreground. Its deep, dark green, velvety surface is covered with intricate, luminous veins, resembling circuit boards, faintly luminescent. As the camera deepens, focus gradually settles on an Asian model nestled within a futuristic greenhouse. Her face is sculpted, her long, wet black hair neatly combed back. Her eyes, cold and distant, stare blankly into the camera. She wears a couture dress crafted from a liquid metal silver fabric with dramatic, architectural shoulders, making her stand out against the backdrop. Behind her, rows of stainless steel and glass vertical growth racks display a variety of exotic plants, each a psychedelic display of vibrant colors under alternating purple and magenta LED grow lights. The lights pulsated in a slow and rhythmic rhythm, periodically sweeping across her expressionless face and the glittering surface of the clothing, casting ever-changing long shadows. The entire scene was shrouded in a cold tone, and the precise coldness of technology and the ultimate beauty of luxury fashion blended at this moment, creating a fashion blockbuster atmosphere full of mystery and visual impact.", + "seed": 12050, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_4/index_5809_seed_12050_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.019383630622178315, + "vlm_logprob": -0.019383630622178315, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.019383630622178315, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014058134518563747 + ], + "opposite_logprob_tokens": [ + -4.3890581130981445 + ], + "logprob_reward": -0.014058134518563747, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014058134518563747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8924371004104614, + 0.020187312737107277, + 0.032940201461315155, + 0.009824528358876705, + 0.007110218051820993, + 0.01465894840657711, + 0.004130212124437094, + 0.007193543948233128, + 0.004114668816328049, + 0.007403187919408083 + ], + "credit_map": [ + 0.06914166361093521, + 0.17339546978473663, + 0.024043193086981773, + 0.07677707821130753, + 0.054380375891923904, + 0.02773679792881012, + 0.3471144139766693, + 0.11984815448522568, + 0.0023213038221001625, + 0.0020286766812205315, + 0.0023115978110581636, + 0.0017337732715532184, + 0.001837098621763289, + 0.0048286630772054195, + 0.0020168705377727747, + 0.0031093317084014416, + 0.002520380774512887, + 0.0015418125549331307, + 0.002228437690064311, + 0.011026918888092041, + 0.0012848376063629985, + 0.0022146780975162983, + 0.0030227804090827703, + 0.009100359864532948, + 0.0009119891910813749, + 0.0010197173105552793, + 0.0015821795677766204, + 0.001667653676122427, + 0.001019714050926268, + 0.0009993065614253283, + 0.0015629504341632128, + 0.0010610193712636828, + 0.0006768401945009828, + 0.0009077090653590858, + 0.0007371471147052944, + 0.0007553999894298613, + 0.0016080037457868457, + 0.0010640880791470408, + 0.0007024210644885898, + 0.000658609380479902, + 0.0010383648332208395, + 0.0058698514476418495, + 0.002170379739254713, + 0.001027447055093944, + 0.0013239093823358417, + 0.0015406697057187557, + 0.0009135750005953014, + 0.0007747520576231182, + 0.0007207854650914669, + 0.0003629834682215005, + 0.0005919414688833058, + 0.000533990329131484, + 0.0005862358957529068, + 0.0005493827047757804, + 0.00037925629294477403, + 0.00040563641232438385, + 0.0005218421574681997, + 0.0008437708020210266, + 0.0018356720684096217, + 0.0005996532854624093, + 0.0009088205988518894, + 0.00044088935828767717, + 0.0006562296184711158, + 0.0013866665540263057, + 0.0004406722728163004, + 0.00040031399112194777, + 0.0004442087665665895, + 0.000672517460770905, + 0.0006271586171351373, + 0.00032566575100645423, + 0.0007794439443387091, + 0.00042468818719498813, + 0.0006243947427719831, + 0.0005290110129863024, + 0.0011517660459503531, + 0.0007989460718818009, + 0.0011724377982318401, + 0.00046014654799364507, + 0.0009477626881562173, + 0.0017187240300700068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.014058134518563747 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549362145364285 + ], + "opposite_logprob_tokens": [ + -4.38754940032959 + ], + "logprob_reward": -0.012549362145364285, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549362145364285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1160566583275795, + 0.1094055101275444, + 0.23797230422496796, + 0.08569981902837753, + 0.11234355717897415, + 0.08064202964305878, + 0.07883869111537933, + 0.07613340765237808, + 0.05065973475575447, + 0.05224823206663132 + ], + "credit_map": [ + 0.01377298217266798, + 0.016801564022898674, + 0.015792880207300186, + 0.013695632107555866, + 0.012089955620467663, + 0.013027777895331383, + 0.01680940017104149, + 0.014066464267671108, + 0.01452525332570076, + 0.013499184511601925, + 0.012270236387848854, + 0.010232592932879925, + 0.008954609744250774, + 0.01726921834051609, + 0.014764897525310516, + 0.017889512702822685, + 0.014623722992837429, + 0.012258323840796947, + 0.015297213569283485, + 0.08822309970855713, + 0.006909274496138096, + 0.013094250112771988, + 0.017832983285188675, + 0.06973343342542648, + 0.007654773537069559, + 0.008755524642765522, + 0.012018326669931412, + 0.015245559625327587, + 0.012233147397637367, + 0.008533076383173466, + 0.013686083257198334, + 0.00757332518696785, + 0.009575719945132732, + 0.013044710271060467, + 0.012224027886986732, + 0.01560218259692192, + 0.020591236650943756, + 0.01891915127635002, + 0.011040669865906239, + 0.011345861479640007, + 0.006543864496052265, + 0.014509276486933231, + 0.013799683190882206, + 0.01011148001998663, + 0.007839198224246502, + 0.008348413743078709, + 0.011096193455159664, + 0.008393927477300167, + 0.013797314837574959, + 0.006439180579036474, + 0.009693074971437454, + 0.008896974846720695, + 0.008654332719743252, + 0.015027769841253757, + 0.009647715836763382, + 0.006682326085865498, + 0.007380959577858448, + 0.010139204561710358, + 0.016227256506681442, + 0.005609225481748581, + 0.009024566039443016, + 0.0076236724853515625, + 0.010136228986084461, + 0.009992292150855064, + 0.0072399927303195, + 0.005495118908584118, + 0.005667557008564472, + 0.00605641957372427, + 0.007503805216401815, + 0.004946340806782246, + 0.007816866040229797, + 0.005933634005486965, + 0.0069585987366735935, + 0.003665800206363201, + 0.00668904185295105, + 0.004651894327253103, + 0.007246796507388353, + 0.0037583312951028347, + 0.006929509807378054, + 0.012348259799182415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.012549362145364285 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735833827406168 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735833827406168, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735833827406168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16377055644989014, + 0.07053883373737335, + 0.10323383659124374, + 0.059904951602220535, + 0.05981264263391495, + 0.07541365176439285, + 0.07489610463380814, + 0.139842689037323, + 0.16934503614902496, + 0.08324164152145386 + ], + "credit_map": [ + 0.011978546157479286, + 0.02712705358862877, + 0.011260125786066055, + 0.014337705448269844, + 0.014675850048661232, + 0.007853200659155846, + 0.056978870183229446, + 0.01955920271575451, + 0.012210406363010406, + 0.0062652528285980225, + 0.008839706890285015, + 0.007850312627851963, + 0.006098885554820299, + 0.011548659764230251, + 0.007812623865902424, + 0.009912987239658833, + 0.0081704780459404, + 0.00488094799220562, + 0.00740034319460392, + 0.03072490356862545, + 0.003772970987483859, + 0.007296071387827396, + 0.010045758448541164, + 0.030942367389798164, + 0.004178688395768404, + 0.007328916806727648, + 0.008312406949698925, + 0.008929196745157242, + 0.011993829160928726, + 0.0052869077771902084, + 0.008714883588254452, + 0.005160124972462654, + 0.005758261773735285, + 0.009320097044110298, + 0.005362194962799549, + 0.00733925262466073, + 0.010494474321603775, + 0.008413749746978283, + 0.00842183455824852, + 0.004702776204794645, + 0.006268249824643135, + 0.008086681365966797, + 0.009559092111885548, + 0.008409322239458561, + 0.01415348518639803, + 0.006584618240594864, + 0.01617550477385521, + 0.006176697555929422, + 0.01088047493249178, + 0.006351884920150042, + 0.00789753720164299, + 0.007692210841923952, + 0.009670165367424488, + 0.010935970582067966, + 0.014768910594284534, + 0.006698948331177235, + 0.011570096015930176, + 0.022352637723088264, + 0.02652885392308235, + 0.010966507717967033, + 0.018351713195443153, + 0.010893566533923149, + 0.021619277074933052, + 0.01756003126502037, + 0.026398109272122383, + 0.022732995450496674, + 0.02092789299786091, + 0.018330655992031097, + 0.019569866359233856, + 0.016962192952632904, + 0.01997550204396248, + 0.024447832256555557, + 0.007293371483683586, + 0.0063011194579303265, + 0.008811292238533497, + 0.008686225861310959, + 0.013932289555668831, + 0.005092495586723089, + 0.01523685548454523, + 0.017887994647026062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.006735833827406168 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634967729449272 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008634967729449272, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634967729449272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.575900137424469, + 0.054685525596141815, + 0.19067147374153137, + 0.01692059077322483, + 0.015563285909593105, + 0.02977137640118599, + 0.012856853194534779, + 0.01657584309577942, + 0.03417268022894859, + 0.05288220942020416 + ], + "credit_map": [ + 0.045968662947416306, + 0.11707412451505661, + 0.015443318523466587, + 0.05003469064831734, + 0.035150591284036636, + 0.01880073733627796, + 0.220619797706604, + 0.07280824333429337, + 0.008868386968970299, + 0.00435708649456501, + 0.00655363779515028, + 0.005983851384371519, + 0.0061279903165996075, + 0.008410593494772911, + 0.005780448205769062, + 0.008603530935943127, + 0.013153009116649628, + 0.008431957103312016, + 0.012140031903982162, + 0.06768964231014252, + 0.004879872314631939, + 0.010674637742340565, + 0.015265780501067638, + 0.05843653902411461, + 0.0014406577683985233, + 0.0018956451676785946, + 0.002649476984515786, + 0.003122506896033883, + 0.0018066485645249486, + 0.00222807377576828, + 0.002181913470849395, + 0.001595668843947351, + 0.0017624963074922562, + 0.0015189035329967737, + 0.0014821791555732489, + 0.0023413551971316338, + 0.0012994041899219155, + 0.003570388536900282, + 0.0020329453982412815, + 0.0015556140569970012, + 0.002983821090310812, + 0.007470675744116306, + 0.003452049568295479, + 0.0034710217732936144, + 0.0026364552322775126, + 0.0031663060653954744, + 0.0031394469551742077, + 0.0034516016021370888, + 0.001719528459943831, + 0.0013202325208112597, + 0.0017352575669065118, + 0.001279231859371066, + 0.0022507011890411377, + 0.0017605720786377788, + 0.001567212981171906, + 0.001224116189405322, + 0.0017879479564726353, + 0.002867189235985279, + 0.0018476805416867137, + 0.001715924241580069, + 0.002919667400419712, + 0.0017503947019577026, + 0.0019211521139368415, + 0.0017658874858170748, + 0.006124165374785662, + 0.004588224459439516, + 0.004407928325235844, + 0.003148217685520649, + 0.004106850363314152, + 0.0038658399134874344, + 0.0030827249865978956, + 0.004848725628107786, + 0.004234204534441233, + 0.0041923862881958485, + 0.008904697373509407, + 0.006350390147417784, + 0.005374339409172535, + 0.0036436538212001324, + 0.006156644783914089, + 0.014025891199707985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.008634967729449272 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up starting with a view of a large plant leaf in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance slowly and steadily from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an Asian model wearing a liquid metal silver dress with architectural shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements stainless steel and glass vertical growth racks with exotic plants under purple and magenta LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by pulsating purple and magenta lights that cast long shadows on the model and her dress?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493985489010811 + ], + "opposite_logprob_tokens": [ + -2.9299397468566895 + ], + "logprob_reward": -0.05493985489010811, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493985489010811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1499357372522354, + 0.06878786534070969, + 0.26157501339912415, + 0.07001631706953049, + 0.07540183514356613, + 0.1293538510799408, + 0.046143077313899994, + 0.08370640873908997, + 0.07817725837230682, + 0.03690265864133835 + ], + "credit_map": [ + 0.009977470152080059, + 0.02279607579112053, + 0.007193923462182283, + 0.013486796990036964, + 0.010559724643826485, + 0.009254926815629005, + 0.05312255769968033, + 0.02354426123201847, + 0.01206892542541027, + 0.006513410713523626, + 0.0078119998797774315, + 0.007563116494566202, + 0.0077399094589054585, + 0.009859377518296242, + 0.0069537051022052765, + 0.010277416557073593, + 0.01705421879887581, + 0.011161629110574722, + 0.014950547367334366, + 0.09892150014638901, + 0.006410767324268818, + 0.014006128534674644, + 0.01949099823832512, + 0.07957924157381058, + 0.005801367573440075, + 0.009442286565899849, + 0.008023310452699661, + 0.009727909229695797, + 0.013482550159096718, + 0.009046703577041626, + 0.007867188192903996, + 0.006625000387430191, + 0.00707970978692174, + 0.007591932080686092, + 0.007760989014059305, + 0.01406989898532629, + 0.009633583948016167, + 0.014091660268604755, + 0.007762645371258259, + 0.007411421742290258, + 0.009488191455602646, + 0.04257028177380562, + 0.0258224718272686, + 0.00903063453733921, + 0.015250506810843945, + 0.008874719962477684, + 0.011065405793488026, + 0.007251635193824768, + 0.004862157162278891, + 0.003450837219133973, + 0.008158201351761818, + 0.00584547920152545, + 0.006380384787917137, + 0.00578857958316803, + 0.005962698254734278, + 0.005694739054888487, + 0.010878815315663815, + 0.00910184159874916, + 0.0176959540694952, + 0.006391792558133602, + 0.008237338624894619, + 0.010320521891117096, + 0.007369694299995899, + 0.013710446655750275, + 0.013270039111375809, + 0.010641336441040039, + 0.010769708082079887, + 0.00791738461703062, + 0.00901307724416256, + 0.008537818677723408, + 0.0071327961049973965, + 0.010895092971622944, + 0.003981797490268946, + 0.0031595565378665924, + 0.004523978568613529, + 0.005482872482389212, + 0.0051248627714812756, + 0.0020582673605531454, + 0.00620452594012022, + 0.006366797257214785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.05493985489010811 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_5/index_tensor([6309, 6309])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_5/index_tensor([6309, 6309])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..272b80a73dda2b973fc8fdc04eb6a5c11daa8b6a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_5/index_tensor([6309, 6309])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow-motion shot, shot from a low angle, pans right, tracking a sturdy, dark green vintage off-road vehicle as it leaps from one side of a canyon to the other. The crimson sandstone cliffs are tinged with a warm glow by the setting sun's golden afterglow, while the bottomless chasm casts long shadows below. The vehicle's massive, all-terrain tires, smeared with dried mud, rotate slowly and gracefully through the air. At its apex, the vehicle achieves a brief, weightless stillness, tracing a graceful and powerful parabola. The setting sun's rays shimmer off the finely scratched metal, casting a warm, dazzling glow and causing the expedition gear strapped to the roof rack to sway slightly. The image, dominated by warm tones, and the vast sky dotted with wisps of pink and purple clouds, creates an epic, unfettered atmosphere, perfectly embodying the romantic spirit of conquering nature.", + "seed": 6098, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_5/index_6309_seed_6098_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0307120556011796, + "vlm_logprob": -0.0307120556011796, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.0307120556011796, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010841868817806244 + ], + "opposite_logprob_tokens": [ + -5.0108418464660645 + ], + "logprob_reward": -0.010841868817806244, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.010841868817806244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1818762868642807, + 0.10923108458518982, + 0.03677010536193848, + 0.07167629897594452, + 0.1343967765569687, + 0.09163565188646317, + 0.044355474412441254, + 0.17886286973953247, + 0.07468423992395401, + 0.0765111893415451 + ], + "credit_map": [ + 0.014388729818165302, + 0.010274910368025303, + 0.02229464426636696, + 0.029987609013915062, + 0.018267178907990456, + 0.024806823581457138, + 0.040125522762537, + 0.021730881184339523, + 0.024303458631038666, + 0.008615345694124699, + 0.0056414310820400715, + 0.008004478178918362, + 0.0040100133046507835, + 0.015614096075296402, + 0.0108523890376091, + 0.032189879566431046, + 0.0037356580141931772, + 0.0031078632455319166, + 0.0049435775727033615, + 0.0038302589673548937, + 0.0020220826845616102, + 0.006992255803197622, + 0.0028486447408795357, + 0.009289762005209923, + 0.014348278753459454, + 0.005910125095397234, + 0.006060808897018433, + 0.006003199145197868, + 0.0032865721732378006, + 0.009287402965128422, + 0.004732782486826181, + 0.022047124803066254, + 0.02749834954738617, + 0.01356508582830429, + 0.013320863246917725, + 0.01223102305084467, + 0.008784310892224312, + 0.014700321480631828, + 0.008050525560975075, + 0.036246296018362045, + 0.020033683627843857, + 0.007018635980784893, + 0.007511246483772993, + 0.008438954129815102, + 0.004203010816127062, + 0.00832991674542427, + 0.006156784947961569, + 0.029943417757749557, + 0.006215536035597324, + 0.00408534100279212, + 0.005986869800835848, + 0.005240708589553833, + 0.005247806664556265, + 0.006928622256964445, + 0.0052358293905854225, + 0.005414760205894709, + 0.02293260395526886, + 0.018224814906716347, + 0.028110230341553688, + 0.01951211877167225, + 0.026569869369268417, + 0.01741664856672287, + 0.020834166556596756, + 0.02526240609586239, + 0.009318496100604534, + 0.005504816770553589, + 0.011502137407660484, + 0.011928466148674488, + 0.009967133402824402, + 0.007931621745228767, + 0.008825594559311867, + 0.009705974720418453, + 0.010822922922670841, + 0.016291653737425804, + 0.009950541891157627, + 0.007537094876170158, + 0.006581220310181379, + 0.008457904681563377, + 0.008525867946445942, + 0.008343989960849285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.010841868817806244 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890541821718216 + ], + "opposite_logprob_tokens": [ + -2.4640541076660156 + ], + "logprob_reward": -0.0890541821718216, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890541821718216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07034894078969955, + 0.015654567629098892, + 0.021886682137846947, + 0.017022328451275826, + 0.2562246024608612, + 0.20039400458335876, + 0.19133800268173218, + 0.06384960561990738, + 0.05529233440756798, + 0.10798884928226471 + ], + "credit_map": [ + 0.007847759872674942, + 0.004549829289317131, + 0.007701823487877846, + 0.00844278559088707, + 0.005892335437238216, + 0.018339205533266068, + 0.00981451477855444, + 0.007760694250464439, + 0.002012497279793024, + 0.0016303251031786203, + 0.0020193923264741898, + 0.0022630670573562384, + 0.0016840833704918623, + 0.001971085323020816, + 0.002637135796248913, + 0.0014369828859344125, + 0.0020739829633384943, + 0.0021139925811439753, + 0.0033882721327245235, + 0.002859637141227722, + 0.0016405571950599551, + 0.003827509703114629, + 0.0019720313139259815, + 0.004010702017694712, + 0.001969659700989723, + 0.001921498216688633, + 0.0023174043744802475, + 0.0022128645796328783, + 0.0015518211293965578, + 0.0020569986663758755, + 0.0017001189989969134, + 0.003291963366791606, + 0.0605417899787426, + 0.015517582185566425, + 0.021350013092160225, + 0.021913256496191025, + 0.0094058308750391, + 0.025450879707932472, + 0.014606612734496593, + 0.08743866533041, + 0.04290609806776047, + 0.013534084893763065, + 0.013131820596754551, + 0.019677897915244102, + 0.008848322555422783, + 0.01962953992187977, + 0.014392256736755371, + 0.06827401369810104, + 0.019982093945145607, + 0.019648827612400055, + 0.02641325257718563, + 0.026232603937387466, + 0.022033438086509705, + 0.023402217775583267, + 0.02830568514764309, + 0.025319913402199745, + 0.007971161976456642, + 0.004770049825310707, + 0.006077210418879986, + 0.010693283751606941, + 0.009000592865049839, + 0.006566728465259075, + 0.009650872088968754, + 0.009119712747633457, + 0.006648204755038023, + 0.004147866275161505, + 0.010706653818488121, + 0.00741185387596488, + 0.0060456362552940845, + 0.006330885924398899, + 0.006143385078758001, + 0.007857851684093475, + 0.013059842400252819, + 0.013045910745859146, + 0.01214379072189331, + 0.014882128685712814, + 0.008035728707909584, + 0.014745863154530525, + 0.015129211358726025, + 0.01694638654589653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.0890541821718216 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009798997081816196 + ], + "opposite_logprob_tokens": [ + -4.634799003601074 + ], + "logprob_reward": -0.009798997081816196, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009798997081816196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10777141898870468, + 0.04252047464251518, + 0.019791847094893456, + 0.0522480234503746, + 0.114390529692173, + 0.048267051577568054, + 0.04639405012130737, + 0.3118768632411957, + 0.13076643645763397, + 0.12597328424453735 + ], + "credit_map": [ + 0.010188917629420757, + 0.008237196132540703, + 0.011939452029764652, + 0.01693814992904663, + 0.00903591699898243, + 0.014781920239329338, + 0.019402485340833664, + 0.017247376963496208, + 0.009211646392941475, + 0.0033675103913992643, + 0.0020724048372358084, + 0.0029921475797891617, + 0.0018860335694625974, + 0.006578768137842417, + 0.0038102909456938505, + 0.012601672671735287, + 0.002081871498376131, + 0.0020201983861625195, + 0.0034184257965534925, + 0.0029251612722873688, + 0.0010237847454845905, + 0.002849423559382558, + 0.0018747623544186354, + 0.00359821948222816, + 0.011425168253481388, + 0.0037629378493875265, + 0.00324328918941319, + 0.004230131395161152, + 0.0017914972268044949, + 0.007641494274139404, + 0.002843440743163228, + 0.017310064285993576, + 0.025019194930791855, + 0.009389426559209824, + 0.01207628846168518, + 0.009299741126596928, + 0.006209174171090126, + 0.010593732818961143, + 0.007389470934867859, + 0.03441350907087326, + 0.007301537785679102, + 0.0053416285663843155, + 0.00582517497241497, + 0.007296820171177387, + 0.002741758245974779, + 0.004158372525125742, + 0.0037130839191377163, + 0.011888678185641766, + 0.005484896246343851, + 0.004361510276794434, + 0.00800892524421215, + 0.0062673999927937984, + 0.004904231987893581, + 0.006294318474829197, + 0.004719542805105448, + 0.006353221833705902, + 0.040878359228372574, + 0.03277260437607765, + 0.03388805314898491, + 0.048826415091753006, + 0.04532061889767647, + 0.027542736381292343, + 0.04320566728711128, + 0.039442405104637146, + 0.01686842553317547, + 0.013066408224403858, + 0.023010291159152985, + 0.017819104716181755, + 0.01831408031284809, + 0.011342925950884819, + 0.014706927351653576, + 0.015638282522559166, + 0.01478438451886177, + 0.01598137430846691, + 0.01322233397513628, + 0.017758535221219063, + 0.009771089069545269, + 0.01598323881626129, + 0.0183700118213892, + 0.020102310925722122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.009798997081816196 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057238109409809 + ], + "opposite_logprob_tokens": [ + -3.8955724239349365 + ], + "logprob_reward": -0.02057238109409809, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057238109409809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10499761253595352, + 0.0906790941953659, + 0.027305833995342255, + 0.03694140911102295, + 0.16069357097148895, + 0.07508134096860886, + 0.086533322930336, + 0.2201654016971588, + 0.14506931602954865, + 0.05253315344452858 + ], + "credit_map": [ + 0.012704510241746902, + 0.007911463268101215, + 0.01515941135585308, + 0.012358766049146652, + 0.011351719498634338, + 0.01504349522292614, + 0.020008092746138573, + 0.010460155084729195, + 0.019349904730916023, + 0.007572982460260391, + 0.005858226213604212, + 0.008221910335123539, + 0.003396139480173588, + 0.011758755892515182, + 0.009868340566754341, + 0.024652834981679916, + 0.0031849530059844255, + 0.0051189810037612915, + 0.0038388429675251245, + 0.004707837011665106, + 0.0018324676202610135, + 0.0024759506341069937, + 0.002811446553096175, + 0.0033353555481880903, + 0.005763855297118425, + 0.0031363186426460743, + 0.004598419182002544, + 0.00406632712110877, + 0.0020522053819149733, + 0.00514574721455574, + 0.002525902586057782, + 0.009652631357312202, + 0.03706794232130051, + 0.009729958139359951, + 0.01323297992348671, + 0.014008589088916779, + 0.005131525918841362, + 0.017948545515537262, + 0.010415639728307724, + 0.05315838381648064, + 0.014089761301875114, + 0.006341279484331608, + 0.008602127432823181, + 0.007554751820862293, + 0.0038264626637101173, + 0.007328490726649761, + 0.00551039632409811, + 0.021828066557645798, + 0.010437967255711555, + 0.008763009682297707, + 0.014398886822164059, + 0.009793451987206936, + 0.00871309731155634, + 0.01253234688192606, + 0.00874210987240076, + 0.013152455911040306, + 0.025915691629052162, + 0.021620498970150948, + 0.02401355467736721, + 0.03454630821943283, + 0.027159633114933968, + 0.02495983988046646, + 0.02929377742111683, + 0.03265611082315445, + 0.016613740473985672, + 0.013276546262204647, + 0.021850764751434326, + 0.027179859578609467, + 0.019533054903149605, + 0.013260540552437305, + 0.01611177623271942, + 0.0172430332750082, + 0.006794406101107597, + 0.006514550652354956, + 0.006485359277576208, + 0.0068643768317997456, + 0.004239855334162712, + 0.00769234960898757, + 0.005535427015274763, + 0.008406829088926315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.02057238109409809 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292848840355873 + ], + "opposite_logprob_tokens": [ + -3.7732927799224854 + ], + "logprob_reward": -0.023292848840355873, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292848840355873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13735629618167877, + 0.0223077442497015, + 0.03253161907196045, + 0.019598450511693954, + 0.14823131263256073, + 0.13363413512706757, + 0.09424524009227753, + 0.21745702624320984, + 0.06686235964298248, + 0.12777581810951233 + ], + "credit_map": [ + 0.016583142802119255, + 0.00996500812470913, + 0.015606108121573925, + 0.01639617420732975, + 0.013374796137213707, + 0.03204270452260971, + 0.01507882121950388, + 0.018309535458683968, + 0.0036846082657575607, + 0.0022214017808437347, + 0.0013357492862269282, + 0.0020436334889382124, + 0.0021485015749931335, + 0.002946657594293356, + 0.0030500709544867277, + 0.004877121653407812, + 0.003262428566813469, + 0.0027312689926475286, + 0.0036416465882211924, + 0.003951325546950102, + 0.0018001716816797853, + 0.007689657621085644, + 0.0027587772347033024, + 0.006696342956274748, + 0.001823308295570314, + 0.002506962278857827, + 0.0037239459343254566, + 0.0031308599282056093, + 0.0018741837702691555, + 0.0018983110785484314, + 0.0027115968987345695, + 0.0019292826764285564, + 0.03323687985539436, + 0.00979579333215952, + 0.01391025260090828, + 0.012586567550897598, + 0.007142448797821999, + 0.015042364597320557, + 0.009055374190211296, + 0.047461628913879395, + 0.03208012506365776, + 0.007860824465751648, + 0.010339953005313873, + 0.009633541107177734, + 0.004301686771214008, + 0.012355275452136993, + 0.008518215268850327, + 0.048544514924287796, + 0.010966282337903976, + 0.009788031689822674, + 0.01458721049129963, + 0.01308246236294508, + 0.011372853070497513, + 0.011042097583413124, + 0.0124995531514287, + 0.010906759649515152, + 0.025766735896468163, + 0.020161079242825508, + 0.021009651944041252, + 0.03557784855365753, + 0.03109140507876873, + 0.022432126104831696, + 0.03100922331213951, + 0.030408959835767746, + 0.007851352915167809, + 0.007564943749457598, + 0.011122753843665123, + 0.012895137071609497, + 0.006849328987300396, + 0.007004153449088335, + 0.005677155219018459, + 0.007897535338997841, + 0.015875667333602905, + 0.021381596103310585, + 0.01943640038371086, + 0.01573602855205536, + 0.010356617160141468, + 0.013638696633279324, + 0.015526026487350464, + 0.015824776142835617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.023292848840355873 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion shot, shot from a low angle, pans right, tracking a sturdy, dark green vintage off-road vehicle as it leaps from one side of a canyon to the other. The crimson sandstone cliffs are tinged with a warm glow by the setting sun's golden afterglow, while the bottomless chasm casts long shadows below. The vehicle's massive, all-terrain tires, smeared with dried mud, rotate slowly and gracefully through the air. At its apex, the vehicle achieves a brief, weightless stillness, tracing a graceful and powerful parabola. The setting sun's rays shimmer off the finely scratched metal, casting a warm, dazzling glow and causing the expedition gear strapped to the roof rack to sway slightly. The image, dominated by warm tones, and the vast sky dotted with wisps of pink and purple clouds, creates an epic, unfettered atmosphere, perfectly embodying the romantic spirit of conquering nature.", + "seed": 6099, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_5/index_6309_seed_6099_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02836159849539399, + "vlm_logprob": -0.02836159849539399, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.02836159849539399, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007291372865438461 + ], + "opposite_logprob_tokens": [ + -5.757291316986084 + ], + "logprob_reward": -0.007291372865438461, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007291372865438461, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28204068541526794, + 0.030575379729270935, + 0.04424164816737175, + 0.09972265362739563, + 0.06380654126405716, + 0.09418029338121414, + 0.13227634131908417, + 0.0747535303235054, + 0.05607745051383972, + 0.12232542783021927 + ], + "credit_map": [ + 0.031494140625, + 0.03722305968403816, + 0.0379863977432251, + 0.038177188485860825, + 0.04877938702702522, + 0.02970683015882969, + 0.03459283336997032, + 0.024080853909254074, + 0.0033825701102614403, + 0.004176102112978697, + 0.0046277232468128204, + 0.00415845075622201, + 0.00440342677757144, + 0.003684880444779992, + 0.0042217024601995945, + 0.0019205256830900908, + 0.004048499744385481, + 0.004156323615461588, + 0.005304612219333649, + 0.009430769830942154, + 0.0052273510955274105, + 0.008070630952715874, + 0.005042612552642822, + 0.0029608479235321283, + 0.012917561456561089, + 0.013105603866279125, + 0.01674920693039894, + 0.011831833980977535, + 0.01314629428088665, + 0.01807955652475357, + 0.007358815521001816, + 0.006533780600875616, + 0.00986295472830534, + 0.011020805686712265, + 0.008677063509821892, + 0.009496097452938557, + 0.0069563318975269794, + 0.00792023353278637, + 0.005117624066770077, + 0.004755430389195681, + 0.01911069266498089, + 0.008206083439290524, + 0.018135447055101395, + 0.011288478039205074, + 0.0071874395944178104, + 0.015317527577280998, + 0.006861742585897446, + 0.008072876371443272, + 0.01674739085137844, + 0.028656745329499245, + 0.02097432315349579, + 0.010793874971568584, + 0.017946911975741386, + 0.018425853922963142, + 0.007199151907116175, + 0.011532098986208439, + 0.011594198644161224, + 0.009657877497375011, + 0.016593024134635925, + 0.004551012068986893, + 0.011288164183497429, + 0.00774709228426218, + 0.007086779456585646, + 0.0062353783287107944, + 0.006707567255944014, + 0.006188177969306707, + 0.006561472080647945, + 0.00448752474039793, + 0.010291009210050106, + 0.006623962428420782, + 0.008119422942399979, + 0.007098314817994833, + 0.02605157159268856, + 0.019007762894034386, + 0.006794749293476343, + 0.007895327173173428, + 0.02232840284705162, + 0.012918541207909584, + 0.013874364085495472, + 0.013454708270728588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.007291372865438461 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293106526136398 + ], + "opposite_logprob_tokens": [ + -2.2379310131073 + ], + "logprob_reward": -0.11293106526136398, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293106526136398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07365753501653671, + 0.02119733765721321, + 0.025306539610028267, + 0.0399104543030262, + 0.049253568053245544, + 0.04640468582510948, + 0.4629388451576233, + 0.207848459482193, + 0.05335092917084694, + 0.02013159915804863 + ], + "credit_map": [ + 0.006876710336655378, + 0.00992558803409338, + 0.007402927149087191, + 0.010631042532622814, + 0.013239854015409946, + 0.009283655323088169, + 0.008864023722708225, + 0.007433735299855471, + 0.0016656656516715884, + 0.0026524134445935488, + 0.0026021376252174377, + 0.0025396905839443207, + 0.0037359113339334726, + 0.0030112487729638815, + 0.0029619657434523106, + 0.002028304385021329, + 0.001891275984235108, + 0.002216041786596179, + 0.0033626751974225044, + 0.004582878202199936, + 0.0039252545684576035, + 0.005916027817875147, + 0.0017991530476137996, + 0.0016132311429828405, + 0.004884633701294661, + 0.004496920853853226, + 0.00655951676890254, + 0.004319086670875549, + 0.005266940221190453, + 0.008031350560486317, + 0.0033916847314685583, + 0.0029603266157209873, + 0.01097889430820942, + 0.005982065573334694, + 0.0071076806634664536, + 0.005834248848259449, + 0.004515422508120537, + 0.007157829590141773, + 0.0032572862692177296, + 0.004420141223818064, + 0.007704136427491903, + 0.005919161718338728, + 0.009144636802375317, + 0.005291682202368975, + 0.004634204786270857, + 0.006159814540296793, + 0.002899517770856619, + 0.004651531111449003, + 0.0703011229634285, + 0.08524667471647263, + 0.12393008172512054, + 0.035478442907333374, + 0.04618315398693085, + 0.040547121316194534, + 0.026741018518805504, + 0.0345112606883049, + 0.028474746271967888, + 0.034416720271110535, + 0.032853055745363235, + 0.01734202355146408, + 0.03234299644827843, + 0.02406950294971466, + 0.019669845700263977, + 0.018679583445191383, + 0.006421629339456558, + 0.006761112716048956, + 0.005452604964375496, + 0.0033388319425284863, + 0.010942027904093266, + 0.005911191459745169, + 0.007795196957886219, + 0.0067283352836966515, + 0.0034838994033634663, + 0.002337891375645995, + 0.0014395827893167734, + 0.0014227641513571143, + 0.004700458142906427, + 0.002206510165706277, + 0.002607667353004217, + 0.0019328255439177155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.11293106526136398 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005262095481157303 + ], + "opposite_logprob_tokens": [ + -5.2552618980407715 + ], + "logprob_reward": -0.005262095481157303, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005262095481157303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2940206229686737, + 0.02348462864756584, + 0.033980969339609146, + 0.05057971924543381, + 0.05611058697104454, + 0.056355372071266174, + 0.18267609179019928, + 0.12458484619855881, + 0.04743647947907448, + 0.13077065348625183 + ], + "credit_map": [ + 0.026986416429281235, + 0.031491562724113464, + 0.0348588265478611, + 0.048103075474500656, + 0.037376146763563156, + 0.03936675563454628, + 0.05274740234017372, + 0.02309044450521469, + 0.002385658910498023, + 0.002461467171087861, + 0.0037393649108707905, + 0.0041821966879069805, + 0.0020935344509780407, + 0.0031929262913763523, + 0.0034378590062260628, + 0.001991621684283018, + 0.0032040688674896955, + 0.003143004607409239, + 0.005542347673326731, + 0.006243252661079168, + 0.004917097743600607, + 0.005283953621983528, + 0.0035867656115442514, + 0.0020604750607162714, + 0.005193146411329508, + 0.00543869799003005, + 0.01065944042056799, + 0.004829227924346924, + 0.006202448159456253, + 0.009242460131645203, + 0.0043084449134767056, + 0.0047058481723070145, + 0.007080176379531622, + 0.005438570398837328, + 0.01123015210032463, + 0.006673361174762249, + 0.007862287573516369, + 0.006843032315373421, + 0.005895667709410191, + 0.005087338387966156, + 0.011179625988006592, + 0.007199707441031933, + 0.009103777818381786, + 0.006602222565561533, + 0.00602209335193038, + 0.008106684312224388, + 0.0037945883814245462, + 0.0043466705828905106, + 0.029070956632494926, + 0.03305122256278992, + 0.044930096715688705, + 0.01322188414633274, + 0.020026562735438347, + 0.017731303349137306, + 0.010991619899868965, + 0.013652435503900051, + 0.01843411847949028, + 0.011231614276766777, + 0.02617909573018551, + 0.009508268907666206, + 0.01700921356678009, + 0.0155826760455966, + 0.01564589887857437, + 0.01099396962672472, + 0.00531168095767498, + 0.0054411375895142555, + 0.004627158399671316, + 0.004136666189879179, + 0.010357832536101341, + 0.005516392178833485, + 0.006366859655827284, + 0.005678755696862936, + 0.02761523239314556, + 0.020524196326732635, + 0.007874461822211742, + 0.00936043355613947, + 0.023205872625112534, + 0.014587595127522945, + 0.0145999351516366, + 0.013002924621105194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.005262095481157303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011070968583226204 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.011070968583226204, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011070968583226204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11138353496789932, + 0.030326826497912407, + 0.04752655699849129, + 0.09603702276945114, + 0.03620736673474312, + 0.050319962203502655, + 0.17808276414871216, + 0.052116937935352325, + 0.11197605729103088, + 0.28602293133735657 + ], + "credit_map": [ + 0.013244854286313057, + 0.009972709231078625, + 0.009498964063823223, + 0.011768805794417858, + 0.012937239371240139, + 0.019790872931480408, + 0.024960217997431755, + 0.009209866635501385, + 0.0029899179935455322, + 0.0033035632222890854, + 0.00555579736828804, + 0.004874422214925289, + 0.003769958857446909, + 0.0032223176676779985, + 0.00407125148922205, + 0.0025395979173481464, + 0.0048600598238408566, + 0.005256643984466791, + 0.007904406636953354, + 0.008559063076972961, + 0.006216587498784065, + 0.007341665681451559, + 0.00491521880030632, + 0.00247291405685246, + 0.01117856428027153, + 0.009453143924474716, + 0.014026034623384476, + 0.009006155654788017, + 0.014462733641266823, + 0.022813282907009125, + 0.006949535105377436, + 0.008147576823830605, + 0.004935163073241711, + 0.006409136578440666, + 0.005026412196457386, + 0.005858804564923048, + 0.004277646075934172, + 0.004082018043845892, + 0.003468792885541916, + 0.0021493928506970406, + 0.010152439586818218, + 0.004522767849266529, + 0.010944908484816551, + 0.004690512549132109, + 0.0036306220572441816, + 0.007011564914137125, + 0.0034207894932478666, + 0.0059463563375175, + 0.02774009294807911, + 0.03362804278731346, + 0.04453803971409798, + 0.01315745897591114, + 0.016808688640594482, + 0.017506862059235573, + 0.01071048155426979, + 0.013993091881275177, + 0.007132824510335922, + 0.008675347082316875, + 0.0075257825665175915, + 0.004028817173093557, + 0.008041709661483765, + 0.005895266309380531, + 0.0051926481537520885, + 0.005624543875455856, + 0.013867540284991264, + 0.019981827586889267, + 0.011648465879261494, + 0.007105052005499601, + 0.022203538566827774, + 0.011414185166358948, + 0.013694824650883675, + 0.012060627341270447, + 0.07443316280841827, + 0.05372460186481476, + 0.01911851391196251, + 0.014707411639392376, + 0.04370735585689545, + 0.026393979787826538, + 0.029856812208890915, + 0.024081092327833176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.011070968583226204 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252490285784006 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252490285784006, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252490285784006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0813627541065216, + 0.03625120222568512, + 0.039233822375535965, + 0.10580097883939743, + 0.0505339540541172, + 0.04114995524287224, + 0.46421369910240173, + 0.06271237879991531, + 0.024563124403357506, + 0.09417816996574402 + ], + "credit_map": [ + 0.007842838764190674, + 0.008814855478703976, + 0.011445078998804092, + 0.012931904755532742, + 0.011242230422794819, + 0.010264282114803791, + 0.010803274810314178, + 0.008018290624022484, + 0.0034048552624881268, + 0.00325946812517941, + 0.005278298631310463, + 0.005912310443818569, + 0.005222716368734837, + 0.005637718830257654, + 0.0058824592269957066, + 0.001653376268222928, + 0.0034287336748093367, + 0.0033074503298848867, + 0.0050323717296123505, + 0.004855244420468807, + 0.0062207733280956745, + 0.011395532637834549, + 0.0026135866064578295, + 0.0023801280185580254, + 0.011132021434605122, + 0.010570396669209003, + 0.015252198092639446, + 0.01119193248450756, + 0.014675186946988106, + 0.024250132963061333, + 0.00922182947397232, + 0.009507282637059689, + 0.006619631312787533, + 0.005464564077556133, + 0.008795416913926601, + 0.0059770140796899796, + 0.0076710316352546215, + 0.00638951500877738, + 0.0052890595979988575, + 0.004327716771513224, + 0.00818408839404583, + 0.005269570741802454, + 0.003362078685313463, + 0.005612865090370178, + 0.0056177475489676, + 0.007559067569673061, + 0.003158255247399211, + 0.0023862794041633606, + 0.07248828560113907, + 0.08455650508403778, + 0.12259434908628464, + 0.03393878415226936, + 0.04472687840461731, + 0.041997138410806656, + 0.027859222143888474, + 0.03605252131819725, + 0.00949805323034525, + 0.008859640918672085, + 0.00863334909081459, + 0.005178478546440601, + 0.011947677470743656, + 0.007469563279300928, + 0.005306183360517025, + 0.005819434765726328, + 0.003431465709581971, + 0.002541203750297427, + 0.005743279121816158, + 0.002396884374320507, + 0.0021573721896857023, + 0.001873726723715663, + 0.003784036496654153, + 0.0026351562701165676, + 0.02198856510221958, + 0.016079721972346306, + 0.005155558232218027, + 0.005582889541983604, + 0.01715441793203354, + 0.009669858030974865, + 0.009542147628962994, + 0.00900501199066639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -0.005252490285784006 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion shot, shot from a low angle, pans right, tracking a sturdy, dark green vintage off-road vehicle as it leaps from one side of a canyon to the other. The crimson sandstone cliffs are tinged with a warm glow by the setting sun's golden afterglow, while the bottomless chasm casts long shadows below. The vehicle's massive, all-terrain tires, smeared with dried mud, rotate slowly and gracefully through the air. At its apex, the vehicle achieves a brief, weightless stillness, tracing a graceful and powerful parabola. The setting sun's rays shimmer off the finely scratched metal, casting a warm, dazzling glow and causing the expedition gear strapped to the roof rack to sway slightly. The image, dominated by warm tones, and the vast sky dotted with wisps of pink and purple clouds, creates an epic, unfettered atmosphere, perfectly embodying the romantic spirit of conquering nature.", + "seed": 6100, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_5/index_6309_seed_6100_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.057297656685113905, + "vlm_logprob": -0.057297656685113905, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.057297656685113905, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012265728786587715 + ], + "opposite_logprob_tokens": [ + -4.637265682220459 + ], + "logprob_reward": -0.012265728786587715, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.012265728786587715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0668642446398735, + 0.0816071480512619, + 0.04010200500488281, + 0.05102735012769699, + 0.12503178417682648, + 0.09197914600372314, + 0.2598247826099396, + 0.17154064774513245, + 0.07684119790792465, + 0.03518169745802879 + ], + "credit_map": [ + 0.0044177076779305935, + 0.009121505543589592, + 0.008808506652712822, + 0.0123538663610816, + 0.010173961520195007, + 0.005010920576751232, + 0.010727059096097946, + 0.006250721402466297, + 0.00750472629442811, + 0.008239210583269596, + 0.007700477726757526, + 0.018040403723716736, + 0.005972001701593399, + 0.005001482088118792, + 0.013851157389581203, + 0.01529768854379654, + 0.004016104619950056, + 0.004503394011408091, + 0.0061857677064836025, + 0.006163256708532572, + 0.0029191516805440187, + 0.004104172810912132, + 0.005072822328656912, + 0.007137331645935774, + 0.0077265482395887375, + 0.008689597249031067, + 0.00735999783501029, + 0.004991362802684307, + 0.0046006664633750916, + 0.004751135129481554, + 0.006588953547179699, + 0.006319093052297831, + 0.013535773381590843, + 0.020838189870119095, + 0.01109746191650629, + 0.013494163751602173, + 0.016268055886030197, + 0.015284846536815166, + 0.021692441776394844, + 0.012820849195122719, + 0.007958858273923397, + 0.01591598056256771, + 0.007213957142084837, + 0.014403645880520344, + 0.012863298878073692, + 0.009597161784768105, + 0.014757505618035793, + 0.009268728084862232, + 0.0858137458562851, + 0.01660396158695221, + 0.012594539672136307, + 0.011085099540650845, + 0.014818248338997364, + 0.09321588277816772, + 0.01244096178561449, + 0.013252354227006435, + 0.017727136611938477, + 0.032892368733882904, + 0.01727045327425003, + 0.018490826711058617, + 0.022653529420495033, + 0.013752593658864498, + 0.021795185282826424, + 0.02695855125784874, + 0.010750781744718552, + 0.009419689886271954, + 0.013833122327923775, + 0.01182510145008564, + 0.004538375418633223, + 0.008125375024974346, + 0.008508777245879173, + 0.009839978069067001, + 0.003169908421114087, + 0.002486519981175661, + 0.006496043875813484, + 0.005935090593993664, + 0.0037027851212769747, + 0.003479445120319724, + 0.006106901448220015, + 0.0038050056900829077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.012265728786587715 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22555382549762726 + ], + "opposite_logprob_tokens": [ + -1.6005538702011108 + ], + "logprob_reward": -0.22555382549762726, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22555382549762726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08721380680799484, + 0.025991171598434448, + 0.02298993244767189, + 0.04528798535466194, + 0.04951546713709831, + 0.07247104495763779, + 0.6585726737976074, + 0.01509247999638319, + 0.012781781144440174, + 0.010083663277328014 + ], + "credit_map": [ + 0.004403274040669203, + 0.005140133202075958, + 0.00541281932964921, + 0.02724147029221058, + 0.003794002579525113, + 0.007107160985469818, + 0.025398513302206993, + 0.008716433309018612, + 0.001921017304994166, + 0.0022748890332877636, + 0.0034996953327208757, + 0.003156764665618539, + 0.0021493565291166306, + 0.0026056000497192144, + 0.004494692198932171, + 0.005889157298952341, + 0.0021316749043762684, + 0.0016177172074094415, + 0.0026495272759348154, + 0.002211212180554867, + 0.0019833592232316732, + 0.004673916380852461, + 0.003472586628049612, + 0.004249940160661936, + 0.004616530146449804, + 0.004374264739453793, + 0.006431667599827051, + 0.004456544294953346, + 0.004727398045361042, + 0.004040709231048822, + 0.007534544914960861, + 0.00910632498562336, + 0.003834655275568366, + 0.012186191976070404, + 0.003819782519713044, + 0.0059950086288154125, + 0.005601057317107916, + 0.004767036531120539, + 0.007677856367081404, + 0.005633876193314791, + 0.007552877999842167, + 0.009418421424925327, + 0.007059428840875626, + 0.010378984734416008, + 0.013548069633543491, + 0.008602469228208065, + 0.006277703680098057, + 0.009633088484406471, + 0.25535351037979126, + 0.025916703045368195, + 0.02068420499563217, + 0.015150482766330242, + 0.020287904888391495, + 0.27616432309150696, + 0.019617415964603424, + 0.0253981351852417, + 0.001404784619808197, + 0.0021227854304015636, + 0.001766512985341251, + 0.0021032397635281086, + 0.0016793743707239628, + 0.0017925011925399303, + 0.002009617630392313, + 0.002213663188740611, + 0.002546686679124832, + 0.0014995764940977097, + 0.0011576758697628975, + 0.0017519657267257571, + 0.00107806536834687, + 0.002363567939028144, + 0.0011400970397517085, + 0.0012441459111869335, + 0.0012548597296699882, + 0.0008654259145259857, + 0.0012753377668559551, + 0.0015425316523760557, + 0.0013069463893771172, + 0.0009416733519174159, + 0.0015874954406172037, + 0.0013093921588733792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.22555382549762726 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547596357762814 + ], + "opposite_logprob_tokens": [ + -4.387547492980957 + ], + "logprob_reward": -0.012547596357762814, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547596357762814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20665235817432404, + 0.0453486405313015, + 0.04751709848642349, + 0.04520315304398537, + 0.3042171001434326, + 0.120025135576725, + 0.08543165028095245, + 0.066618911921978, + 0.0499737374484539, + 0.029012247920036316 + ], + "credit_map": [ + 0.011159965768456459, + 0.016501177102327347, + 0.009417781606316566, + 0.06939304620027542, + 0.012745095416903496, + 0.010526555590331554, + 0.06239861622452736, + 0.014510090462863445, + 0.006446340121328831, + 0.0030157521832734346, + 0.007433872204273939, + 0.005434059072285891, + 0.003841548692435026, + 0.00426732562482357, + 0.006328888237476349, + 0.00858085136860609, + 0.005357265938073397, + 0.005089669022709131, + 0.007554229814559221, + 0.007200245745480061, + 0.0031683500856161118, + 0.005004669073969126, + 0.006329458672553301, + 0.007813205011188984, + 0.007399254012852907, + 0.007997618988156319, + 0.004569718148559332, + 0.004612103570252657, + 0.0065183150582015514, + 0.005227080546319485, + 0.004625522997230291, + 0.004253535531461239, + 0.02280435524880886, + 0.0633331760764122, + 0.0319526232779026, + 0.03149917721748352, + 0.03816646710038185, + 0.031547050923109055, + 0.047136690467596054, + 0.037777528166770935, + 0.011982305906713009, + 0.021271077916026115, + 0.015001273714005947, + 0.015406050719320774, + 0.017652979120612144, + 0.014717619866132736, + 0.012392828240990639, + 0.011600987985730171, + 0.009761339984834194, + 0.011164246127009392, + 0.008481485769152641, + 0.007112395949661732, + 0.016993753612041473, + 0.0113956518471241, + 0.011350762099027634, + 0.009172012098133564, + 0.007179652340710163, + 0.01111125573515892, + 0.004974767100065947, + 0.006235205568373203, + 0.00903062243014574, + 0.0077682239934802055, + 0.009351669810712337, + 0.01096750982105732, + 0.012110825628042221, + 0.005592186935245991, + 0.0061019668355584145, + 0.0050978646613657475, + 0.002344256965443492, + 0.009435528889298439, + 0.003996261861175299, + 0.005294847302138805, + 0.003172872355207801, + 0.0022564700338989496, + 0.00440181652083993, + 0.005475676618516445, + 0.0021740777883678675, + 0.0025572169106453657, + 0.005225502885878086, + 0.003748613875359297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.012547596357762814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026337718591094017 + ], + "opposite_logprob_tokens": [ + -3.6513376235961914 + ], + "logprob_reward": -0.026337718591094017, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026337718591094017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16227687895298004, + 0.09470334649085999, + 0.04581954702734947, + 0.04910803586244583, + 0.08317966014146805, + 0.04900573939085007, + 0.35210472345352173, + 0.09392476081848145, + 0.05688757449388504, + 0.01298974547535181 + ], + "credit_map": [ + 0.011463461443781853, + 0.018985852599143982, + 0.013063781894743443, + 0.04858507215976715, + 0.010754304938018322, + 0.008035491220653057, + 0.04297526925802231, + 0.008413647301495075, + 0.007157720159739256, + 0.009129390120506287, + 0.009367644786834717, + 0.027618037536740303, + 0.005798850208520889, + 0.007419228088110685, + 0.012418994680047035, + 0.01579347625374794, + 0.006299680098891258, + 0.006400097627192736, + 0.006083404645323753, + 0.0054758102633059025, + 0.0036992819514125586, + 0.00551233533769846, + 0.005938347894698381, + 0.00641058711335063, + 0.007475611753761768, + 0.008939588442444801, + 0.00993136316537857, + 0.005288794171065092, + 0.0042869821190834045, + 0.004935340490192175, + 0.0051702712662518024, + 0.003080085152760148, + 0.007552721071988344, + 0.023206671699881554, + 0.006296231877058744, + 0.007638521958142519, + 0.012963741086423397, + 0.009814257733523846, + 0.009063923731446266, + 0.006643586792051792, + 0.003782173153012991, + 0.00933109875768423, + 0.003890800289809704, + 0.004951191134750843, + 0.009333490394055843, + 0.005569161847233772, + 0.006269909907132387, + 0.005877911578863859, + 0.10729260742664337, + 0.025861656293272972, + 0.021938828751444817, + 0.017358412966132164, + 0.01766936667263508, + 0.11422304064035416, + 0.020223965868353844, + 0.027536854147911072, + 0.012925085611641407, + 0.01169884204864502, + 0.010032596997916698, + 0.0076208570972085, + 0.01043694093823433, + 0.018336262553930283, + 0.01084392610937357, + 0.012030255049467087, + 0.009571542963385582, + 0.00520955678075552, + 0.005586913786828518, + 0.007882582023739815, + 0.004409897141158581, + 0.009266199544072151, + 0.0070339469239115715, + 0.007926936261355877, + 0.0015539888991042972, + 0.0010722674196586013, + 0.0019298929255455732, + 0.0019041483756154776, + 0.0011064171558246017, + 0.0013581274542957544, + 0.002243875991553068, + 0.0018210271373391151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.026337718591094017 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978341419249773 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.00978341419249773, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978341419249773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18013425171375275, + 0.08081287890672684, + 0.03901207819581032, + 0.04079641401767731, + 0.07038988918066025, + 0.10279516875743866, + 0.28220200538635254, + 0.09683716297149658, + 0.055294428020715714, + 0.051725681871175766 + ], + "credit_map": [ + 0.010136187076568604, + 0.013137628324329853, + 0.010594108141958714, + 0.058988481760025024, + 0.013023965060710907, + 0.009652592241764069, + 0.05164732411503792, + 0.012953962199389935, + 0.004323809873312712, + 0.004265585448592901, + 0.006348935887217522, + 0.019109435379505157, + 0.006117071956396103, + 0.008162960410118103, + 0.012201556004583836, + 0.020283522084355354, + 0.0041987220756709576, + 0.0041472893208265305, + 0.005614790599793196, + 0.005781935527920723, + 0.0029773490969091654, + 0.004172182641923428, + 0.0051944926381111145, + 0.0069253165274858475, + 0.006454660091549158, + 0.007521837949752808, + 0.005684196017682552, + 0.0037676880601793528, + 0.004489907529205084, + 0.0043866499327123165, + 0.004736914299428463, + 0.003754560137167573, + 0.007860299199819565, + 0.014891253784298897, + 0.0062064859084784985, + 0.008253386244177818, + 0.009333109483122826, + 0.008696679957211018, + 0.008374112658202648, + 0.006774555891752243, + 0.011061623692512512, + 0.013096821494400501, + 0.012300053611397743, + 0.010694445110857487, + 0.02224159985780716, + 0.013807965442538261, + 0.01009873952716589, + 0.009493913501501083, + 0.11383159458637238, + 0.009844751097261906, + 0.0065505909733474255, + 0.005513870622962713, + 0.0067043001763522625, + 0.12260802090167999, + 0.008420398458838463, + 0.00872848741710186, + 0.01153818890452385, + 0.012889871373772621, + 0.010705407708883286, + 0.008895418606698513, + 0.012754096649587154, + 0.016381941735744476, + 0.011987214908003807, + 0.011685027740895748, + 0.011497779749333858, + 0.007044327445328236, + 0.009089178405702114, + 0.005513037089258432, + 0.0036278467159718275, + 0.0073529318906366825, + 0.004344078712165356, + 0.006825251504778862, + 0.004418453201651573, + 0.0033903736621141434, + 0.010311556048691273, + 0.009287171065807343, + 0.004828151315450668, + 0.0046515376307070255, + 0.009804174304008484, + 0.005034264642745256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.00978341419249773 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion shot, shot from a low angle, pans right, tracking a sturdy, dark green vintage off-road vehicle as it leaps from one side of a canyon to the other. The crimson sandstone cliffs are tinged with a warm glow by the setting sun's golden afterglow, while the bottomless chasm casts long shadows below. The vehicle's massive, all-terrain tires, smeared with dried mud, rotate slowly and gracefully through the air. At its apex, the vehicle achieves a brief, weightless stillness, tracing a graceful and powerful parabola. The setting sun's rays shimmer off the finely scratched metal, casting a warm, dazzling glow and causing the expedition gear strapped to the roof rack to sway slightly. The image, dominated by warm tones, and the vast sky dotted with wisps of pink and purple clouds, creates an epic, unfettered atmosphere, perfectly embodying the romantic spirit of conquering nature.", + "seed": 6101, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_5/index_6309_seed_6101_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11766732940450311, + "vlm_logprob": -0.11766732940450311, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.11766732940450311, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033071354031562805 + ], + "opposite_logprob_tokens": [ + -3.658071279525757 + ], + "logprob_reward": -0.033071354031562805, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.033071354031562805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08888398110866547, + 0.07555943727493286, + 0.04852841794490814, + 0.14927056431770325, + 0.12551045417785645, + 0.06175433471798897, + 0.06417378038167953, + 0.1094319075345993, + 0.06074492260813713, + 0.2161421924829483 + ], + "credit_map": [ + 0.010851172730326653, + 0.011881990358233452, + 0.012166466563940048, + 0.015488557517528534, + 0.008757645264267921, + 0.009125582873821259, + 0.012024256400763988, + 0.008588297292590141, + 0.00933870766311884, + 0.013535596430301666, + 0.01035730168223381, + 0.009313978254795074, + 0.008403175510466099, + 0.007665764074772596, + 0.009097493253648281, + 0.00784741248935461, + 0.007471867371350527, + 0.005941588897258043, + 0.008365868590772152, + 0.005397405941039324, + 0.004083214793354273, + 0.007421012036502361, + 0.0038271830417215824, + 0.0060202754102647305, + 0.01912751980125904, + 0.013573264703154564, + 0.013324221596121788, + 0.015769196674227715, + 0.022608637809753418, + 0.027427030727267265, + 0.019197573885321617, + 0.018243102356791496, + 0.014305422082543373, + 0.022007042542099953, + 0.014019579626619816, + 0.014239184558391571, + 0.01853894256055355, + 0.015053095296025276, + 0.010467748157680035, + 0.016879435628652573, + 0.0057787783443927765, + 0.006527370773255825, + 0.006580681540071964, + 0.007473144680261612, + 0.008989514783024788, + 0.009037143550813198, + 0.008558078669011593, + 0.008809615857899189, + 0.00939563661813736, + 0.006147937849164009, + 0.005831437185406685, + 0.006033682730048895, + 0.004062195774167776, + 0.010531852953135967, + 0.006061005871742964, + 0.016110025346279144, + 0.013149525970220566, + 0.015417372807860374, + 0.01200568675994873, + 0.01570739410817623, + 0.017779899761080742, + 0.011339998804032803, + 0.013781968504190445, + 0.01025005616247654, + 0.0075478027574718, + 0.010059861466288567, + 0.009284532628953457, + 0.008465474471449852, + 0.005278701428323984, + 0.0073658134788274765, + 0.008148380555212498, + 0.004594347905367613, + 0.020975150167942047, + 0.023875366896390915, + 0.02655073255300522, + 0.023903589695692062, + 0.01531887799501419, + 0.032182835042476654, + 0.047494273632764816, + 0.025841357186436653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.033071354031562805 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233081579208374 + ], + "opposite_logprob_tokens": [ + -0.8983081579208374 + ], + "logprob_reward": -0.5233081579208374, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233081579208374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11732742935419083, + 0.05668827146291733, + 0.24819406867027283, + 0.035960957407951355, + 0.08626887947320938, + 0.05287294462323189, + 0.1877359300851822, + 0.1460171639919281, + 0.019477790221571922, + 0.04945659637451172 + ], + "credit_map": [ + 0.014385314658284187, + 0.015322413295507431, + 0.013601365499198437, + 0.012473292648792267, + 0.008462567813694477, + 0.025545574724674225, + 0.01139355730265379, + 0.01614333502948284, + 0.00983348861336708, + 0.014174139127135277, + 0.005535473115742207, + 0.006553103681653738, + 0.005317227449268103, + 0.007535440847277641, + 0.004674416035413742, + 0.003064980497583747, + 0.032720644026994705, + 0.02377013862133026, + 0.04439794644713402, + 0.02944471500813961, + 0.01544683612883091, + 0.046366460621356964, + 0.019858086481690407, + 0.036189235746860504, + 0.0040926155634224415, + 0.005537126213312149, + 0.004143420606851578, + 0.003564300248399377, + 0.0035684993490576744, + 0.0062060230411589146, + 0.003189773764461279, + 0.0056591941975057125, + 0.00940684787929058, + 0.013096731156110764, + 0.011074161157011986, + 0.011019180528819561, + 0.011448529548943043, + 0.011063961312174797, + 0.010242157615721226, + 0.00891730934381485, + 0.005519011523574591, + 0.00727690476924181, + 0.006891995668411255, + 0.004415010567754507, + 0.005539086647331715, + 0.00982852466404438, + 0.005607916507869959, + 0.007794490549713373, + 0.022051410749554634, + 0.01433265209197998, + 0.013766690157353878, + 0.017320364713668823, + 0.01332108024507761, + 0.03590827062726021, + 0.01816626638174057, + 0.05286917835474014, + 0.01438357774168253, + 0.01591954380273819, + 0.018705589696764946, + 0.0182154830545187, + 0.01512607466429472, + 0.026807015761733055, + 0.020677166059613228, + 0.01618270017206669, + 0.002013180172070861, + 0.002900459337979555, + 0.0025092901196330786, + 0.002462490927428007, + 0.001941224909387529, + 0.0031052601989358664, + 0.0027345253620296717, + 0.0018113561673089862, + 0.004959083627909422, + 0.007341518998146057, + 0.005423649679869413, + 0.005330199841409922, + 0.004190186038613319, + 0.004047277383506298, + 0.01258506067097187, + 0.0055796122178435326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.5233081579208374 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004644795786589384 + ], + "opposite_logprob_tokens": [ + -5.379644870758057 + ], + "logprob_reward": -0.004644795786589384, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004644795786589384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06329840421676636, + 0.1365884691476822, + 0.1842801719903946, + 0.13695749640464783, + 0.16846585273742676, + 0.10183583945035934, + 0.0272251907736063, + 0.04399283602833748, + 0.04717012867331505, + 0.09018560498952866 + ], + "credit_map": [ + 0.00853687059134245, + 0.008211562409996986, + 0.007920214906334877, + 0.008492662571370602, + 0.007421445567160845, + 0.006959979422390461, + 0.007273817900568247, + 0.008481858298182487, + 0.02448159269988537, + 0.03329832851886749, + 0.013901768252253532, + 0.014735783450305462, + 0.013536106795072556, + 0.013992639258503914, + 0.011962679214775562, + 0.010679585859179497, + 0.025481905788183212, + 0.01619451493024826, + 0.030741246417164803, + 0.021259184926748276, + 0.01193336583673954, + 0.03605642914772034, + 0.014532879926264286, + 0.028080645948648453, + 0.01223469153046608, + 0.015327244997024536, + 0.02188280038535595, + 0.013362261466681957, + 0.010020490735769272, + 0.03538159653544426, + 0.010401744395494461, + 0.018346667289733887, + 0.017466314136981964, + 0.03538369759917259, + 0.01639888994395733, + 0.011051123961806297, + 0.01608586311340332, + 0.036529604345560074, + 0.012220416218042374, + 0.023329952731728554, + 0.012690338306128979, + 0.00947160180658102, + 0.017157377675175667, + 0.010907769203186035, + 0.009504783898591995, + 0.020031863823533058, + 0.010714484378695488, + 0.011357621289789677, + 0.0027890268247574568, + 0.0026973718777298927, + 0.003159070387482643, + 0.002681765239685774, + 0.002107437001541257, + 0.005888717249035835, + 0.0029538541566580534, + 0.004947950132191181, + 0.004852436948567629, + 0.00588920759037137, + 0.005355261731892824, + 0.006034135352820158, + 0.004952088464051485, + 0.006238901987671852, + 0.006244751624763012, + 0.004426056984812021, + 0.005481509957462549, + 0.008921497501432896, + 0.005179946310818195, + 0.006066948175430298, + 0.004733509849756956, + 0.007199352141469717, + 0.005853148177266121, + 0.00373422191478312, + 0.008926915936172009, + 0.0070902202278375626, + 0.011162045411765575, + 0.00760577991604805, + 0.008665424771606922, + 0.022940384224057198, + 0.010851801373064518, + 0.01294303685426712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.004644795786589384 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020573781803250313 + ], + "opposite_logprob_tokens": [ + -3.895573854446411 + ], + "logprob_reward": -0.020573781803250313, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020573781803250313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07797164469957352, + 0.07208570837974548, + 0.055005788803100586, + 0.04054217413067818, + 0.14842413365840912, + 0.09536130726337433, + 0.05310492217540741, + 0.1034570261836052, + 0.13377141952514648, + 0.22027583420276642 + ], + "credit_map": [ + 0.01065677311271429, + 0.010053289122879505, + 0.009602566249668598, + 0.010763388127088547, + 0.008000434376299381, + 0.010384361259639263, + 0.008423552848398685, + 0.010087269358336926, + 0.008814695291221142, + 0.008646962232887745, + 0.007956109009683132, + 0.011273372918367386, + 0.007629431784152985, + 0.007467310875654221, + 0.011074470356106758, + 0.009223355911672115, + 0.00811140425503254, + 0.006015125662088394, + 0.008184320293366909, + 0.007098104804754257, + 0.005682338494807482, + 0.009150364436209202, + 0.005601957906037569, + 0.005162173882126808, + 0.004566485993564129, + 0.003853773232549429, + 0.004984529688954353, + 0.004043653607368469, + 0.0037970631383359432, + 0.010101714171469212, + 0.0038133941125124693, + 0.005381562747061253, + 0.020395629107952118, + 0.023805491626262665, + 0.01383843831717968, + 0.01547640934586525, + 0.024300269782543182, + 0.021119769662618637, + 0.015350962989032269, + 0.014137164689600468, + 0.011365162208676338, + 0.012598398141562939, + 0.009118524380028248, + 0.01256390567868948, + 0.012661803513765335, + 0.013185936957597733, + 0.014079711399972439, + 0.009787864983081818, + 0.005689081270247698, + 0.00398611556738615, + 0.005687729921191931, + 0.0055471621453762054, + 0.00469540897756815, + 0.009581686928868294, + 0.010618302039802074, + 0.007299435790628195, + 0.009826590307056904, + 0.016908539459109306, + 0.01079284306615591, + 0.01607685349881649, + 0.009275704622268677, + 0.013841134496033192, + 0.017211703583598137, + 0.009523649699985981, + 0.017069773748517036, + 0.02009139209985733, + 0.01885712705552578, + 0.01848060078918934, + 0.011980791576206684, + 0.01723773404955864, + 0.019374171271920204, + 0.01067981868982315, + 0.021441221237182617, + 0.026181872934103012, + 0.028946926817297935, + 0.02327534183859825, + 0.018570367246866226, + 0.03663347661495209, + 0.0410042405128479, + 0.02422238141298294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.020573781803250313 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sturdy, dark green vintage off-road vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle shown leaping from one side of a canyon to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's massive all-terrain tires visible and smeared with dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a low angle and panning right while tracking the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dominated by crimson sandstone cliffs with expedition gear on the roof rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738557480275631 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.006738557480275631, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738557480275631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14108160138130188, + 0.08332191407680511, + 0.041748519986867905, + 0.18224318325519562, + 0.12246889621019363, + 0.055034320801496506, + 0.14304763078689575, + 0.10060834139585495, + 0.02461213618516922, + 0.10583345592021942 + ], + "credit_map": [ + 0.018696879968047142, + 0.021461397409439087, + 0.01929100975394249, + 0.01671416684985161, + 0.011853737756609917, + 0.02273428626358509, + 0.013654469512403011, + 0.01667565479874611, + 0.010997651144862175, + 0.011917867697775364, + 0.009743858128786087, + 0.011537573300302029, + 0.008061639033257961, + 0.015298503451049328, + 0.009811345487833023, + 0.0059534721076488495, + 0.006757277064025402, + 0.004784799180924892, + 0.0060159191489219666, + 0.004701092839241028, + 0.005200788378715515, + 0.004382788669317961, + 0.004812730010598898, + 0.005093125160783529, + 0.019051803275942802, + 0.02470673993229866, + 0.026961402967572212, + 0.018410075455904007, + 0.017021793872117996, + 0.035393036901950836, + 0.01966739073395729, + 0.021030955016613007, + 0.013980143703520298, + 0.0205504409968853, + 0.009682091884315014, + 0.010584027506411076, + 0.014222117140889168, + 0.023476900532841682, + 0.011930972337722778, + 0.01804221048951149, + 0.005949749145656824, + 0.008551442064344883, + 0.006707151886075735, + 0.005487519316375256, + 0.005039571318775415, + 0.009735241532325745, + 0.005576002411544323, + 0.007987640798091888, + 0.015518185682594776, + 0.011607525870203972, + 0.01435213815420866, + 0.014559466391801834, + 0.010332746431231499, + 0.027074964717030525, + 0.016117379069328308, + 0.03348521888256073, + 0.010083806701004505, + 0.013011481612920761, + 0.012900582514703274, + 0.013108259066939354, + 0.00974669773131609, + 0.018020398914813995, + 0.013495011255145073, + 0.010242100805044174, + 0.002969601424410939, + 0.0044966829009354115, + 0.004065547604113817, + 0.0034176812041550875, + 0.0020941183902323246, + 0.002707909792661667, + 0.0031612010207027197, + 0.0016993925673887134, + 0.009591706097126007, + 0.006979724392294884, + 0.014489583671092987, + 0.007161776535212994, + 0.009905489161610603, + 0.026285933330655098, + 0.015879694372415543, + 0.015539541840553284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.006738557480275631 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_6/index_tensor([3063, 3063])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_6/index_tensor([3063, 3063])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..06e7eade138322784950f63ace3dff3a18c39589 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_6/index_tensor([3063, 3063])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "horizontal tracking shot, the camera moving with a steady, fluid rhythm, closely follows a Taoist priest in a cyan robe from behind as he journeys through a lush bamboo forest shrouded in a milky white mist. His long, jet-black hair is neatly tied into a chignon, secured with an antique wooden hairpin. His loose cyan cotton and linen robe appears even more subdued in the humid air, its sleeves swaying gently behind him as he walks leisurely. Tall, upright bamboos form a natural, deep corridor, the ground carpeted with layers of withered, yellowed leaves, making it nearly silent when stepped on. The soft, scattered morning light penetrates the thin mist and dense bamboo leaves, casting ever-changing dappled light and shadow on the Taoist's back and the path ahead. He holds a white whisk in his left hand, which sways rhythmically as his body moves. His wide robe occasionally brushes against the dew-covered bamboo leaves beside him. The entire picture presents a sense of seclusion, tranquility and mystery, creating a focused and profound atmosphere of Eastern Zen.", + "seed": 31366, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_6/index_3063_seed_31366_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.019151988439261913, + "vlm_logprob": -0.019151988439261913, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.019151988439261913, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03941059112548828 + ], + "opposite_logprob_tokens": [ + -3.2894105911254883 + ], + "logprob_reward": -0.03941059112548828, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03941059112548828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2919018566608429, + 0.047561611980199814, + 0.02256961353123188, + 0.012982284650206566, + 0.045871466398239136, + 0.03167686238884926, + 0.12130215764045715, + 0.017666496336460114, + 0.022607512772083282, + 0.38586005568504333 + ], + "credit_map": [ + 0.025025101378560066, + 0.0487358421087265, + 0.02106468379497528, + 0.06380170583724976, + 0.032102715224027634, + 0.034572772681713104, + 0.039327338337898254, + 0.027271725237369537, + 0.005129707977175713, + 0.004716886673122644, + 0.004947948269546032, + 0.010639316402375698, + 0.005753589328378439, + 0.005020279437303543, + 0.007573931012302637, + 0.0037799591664224863, + 0.0035002632066607475, + 0.002663014456629753, + 0.002077842131257057, + 0.0022355676628649235, + 0.0028654527850449085, + 0.00291620334610343, + 0.0030692871659994125, + 0.003241985570639372, + 0.0018442849395796657, + 0.0013916569296270609, + 0.001768812071532011, + 0.0018296135822311044, + 0.0019350844668224454, + 0.001202135463245213, + 0.0018202726496383548, + 0.0011904261773452163, + 0.009334678761661053, + 0.004092047456651926, + 0.0062475660815835, + 0.006080396473407745, + 0.008204540237784386, + 0.005761518143117428, + 0.003652063198387623, + 0.0024986614007502794, + 0.005271799396723509, + 0.003552005160599947, + 0.003657688619568944, + 0.0026171787176281214, + 0.007851138710975647, + 0.0036018441896885633, + 0.002829620847478509, + 0.002295587444677949, + 0.027260608971118927, + 0.027934061363339424, + 0.01112388540059328, + 0.012322909198701382, + 0.015682274475693703, + 0.014885280281305313, + 0.006511774845421314, + 0.005581381265074015, + 0.002307379152625799, + 0.0020073663908988237, + 0.0012436530087143183, + 0.0028306616004556417, + 0.0031451585236936808, + 0.002528103766962886, + 0.00186412304174155, + 0.0017400530632585287, + 0.0037356349639594555, + 0.0032645640894770622, + 0.0027210579719394445, + 0.002361924620345235, + 0.003115025581791997, + 0.0030271075665950775, + 0.001960108056664467, + 0.0024220915511250496, + 0.021248525008559227, + 0.01319750864058733, + 0.020082924515008926, + 0.0067891706712543964, + 0.017117146402597427, + 0.15491652488708496, + 0.00875187199562788, + 0.1437564343214035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.03941059112548828 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979462917894125 + ], + "opposite_logprob_tokens": [ + -4.63479471206665 + ], + "logprob_reward": -0.00979462917894125, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979462917894125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39546647667884827, + 0.11010607331991196, + 0.08337006717920303, + 0.027210509404540062, + 0.04033340886235237, + 0.062004249542951584, + 0.038065001368522644, + 0.06638633459806442, + 0.039330240339040756, + 0.13772763311862946 + ], + "credit_map": [ + 0.039645180106163025, + 0.08200161159038544, + 0.044023994356393814, + 0.03605157881975174, + 0.05758390948176384, + 0.04799629747867584, + 0.035772133618593216, + 0.05239180102944374, + 0.014488902874290943, + 0.015201963484287262, + 0.008699719794094563, + 0.018125634640455246, + 0.016427846625447273, + 0.0117383673787117, + 0.015074447728693485, + 0.010349188931286335, + 0.0134664261713624, + 0.012383563444018364, + 0.010331262834370136, + 0.00723092770203948, + 0.012466860935091972, + 0.011851751245558262, + 0.0065263910219073296, + 0.009112884290516376, + 0.0035069561563432217, + 0.002858517924323678, + 0.0034734061919152737, + 0.003684408264234662, + 0.005046462640166283, + 0.0033479928970336914, + 0.003227467415854335, + 0.0020652965176850557, + 0.005134065635502338, + 0.0038436767645180225, + 0.0049453251995146275, + 0.005836359690874815, + 0.0067131961695849895, + 0.004845158196985722, + 0.005592930130660534, + 0.003422696143388748, + 0.009046719409525394, + 0.009023105725646019, + 0.008844904601573944, + 0.005119972862303257, + 0.012056333012878895, + 0.006368507165461779, + 0.006061318330466747, + 0.00548339169472456, + 0.010393938049674034, + 0.007165602874010801, + 0.002492206636816263, + 0.0029994756914675236, + 0.005795653443783522, + 0.0044015697203576565, + 0.002709996188059449, + 0.0021065580658614635, + 0.009554799646139145, + 0.00959194079041481, + 0.006067425478249788, + 0.0066828858107328415, + 0.011093712411820889, + 0.010230408981442451, + 0.006771728862076998, + 0.006393427029252052, + 0.0075166658498346806, + 0.007894590497016907, + 0.004392808768898249, + 0.0035646131727844477, + 0.004071694798767567, + 0.005277605261653662, + 0.0031583260279148817, + 0.0034539387561380863, + 0.01213126815855503, + 0.006598878651857376, + 0.009973385371267796, + 0.006491199601441622, + 0.008231563493609428, + 0.0476054772734642, + 0.004460432566702366, + 0.04223543033003807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.00979462917894125 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248458590358496 + ], + "opposite_logprob_tokens": [ + -5.255248546600342 + ], + "logprob_reward": -0.005248458590358496, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248458590358496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02819945476949215, + 0.01974707469344139, + 0.0266292504966259, + 0.011970714665949345, + 0.07008912414312363, + 0.052883297204971313, + 0.014602030627429485, + 0.012945725582540035, + 0.013408957980573177, + 0.7495242953300476 + ], + "credit_map": [ + 0.0029431304428726435, + 0.00615168921649456, + 0.002805400174111128, + 0.003859676420688629, + 0.0035045507829636335, + 0.0032589950133115053, + 0.0025045531801879406, + 0.0031714627984911203, + 0.002582478104159236, + 0.0021748768631368876, + 0.0023853585589677095, + 0.003644705982878804, + 0.002590695396065712, + 0.002419052878394723, + 0.0025017615407705307, + 0.0014481475809589028, + 0.004176526330411434, + 0.004243906121701002, + 0.003399480367079377, + 0.0023408979177474976, + 0.003780969185754657, + 0.003922875504940748, + 0.0019089572597295046, + 0.002855640836060047, + 0.002705315127968788, + 0.0013265456072986126, + 0.0022546807304024696, + 0.0012718731304630637, + 0.0011936622904613614, + 0.001802419894374907, + 0.0006981411133892834, + 0.0007180781103670597, + 0.01888892613351345, + 0.006986893713474274, + 0.011201608926057816, + 0.005479530431330204, + 0.004656235687434673, + 0.014561546966433525, + 0.0034788139164447784, + 0.004835574887692928, + 0.009339251555502415, + 0.00657281931489706, + 0.010411811992526054, + 0.003968555014580488, + 0.006753792054951191, + 0.006570660974830389, + 0.0033698303159326315, + 0.005896582268178463, + 0.003204028820618987, + 0.0038272826932370663, + 0.0014198299031704664, + 0.001065053278580308, + 0.0017914102645590901, + 0.0020919449161738157, + 0.000602923973929137, + 0.0005995588144287467, + 0.0022262567654252052, + 0.0014568913029506803, + 0.0010316960979253054, + 0.0017233456019312143, + 0.0018903532763943076, + 0.0021257197950035334, + 0.0016096278559416533, + 0.0008818357018753886, + 0.0026568255852907896, + 0.002582424320280552, + 0.001757889287546277, + 0.0012236718321219087, + 0.0012617881875485182, + 0.0017231096280738711, + 0.0007070077699609101, + 0.0014962422428652644, + 0.04058925062417984, + 0.02310233563184738, + 0.03938494622707367, + 0.013334861025214195, + 0.03375253081321716, + 0.30189526081085205, + 0.017019271850585938, + 0.28044596314430237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.005248458590358496 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627411279827356 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.007627411279827356, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627411279827356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05180532857775688, + 0.02878892794251442, + 0.015431373380124569, + 0.01830388605594635, + 0.023449918255209923, + 0.01668095774948597, + 0.05158011615276337, + 0.11732041835784912, + 0.0470106303691864, + 0.6296284794807434 + ], + "credit_map": [ + 0.0056439973413944244, + 0.010358032770454884, + 0.005288805812597275, + 0.006998088210821152, + 0.0064776293002069, + 0.005906275473535061, + 0.006504159886389971, + 0.004628336057066917, + 0.005116709973663092, + 0.0041670892387628555, + 0.003936915658414364, + 0.0028399089351296425, + 0.003034306922927499, + 0.004624673631042242, + 0.0029813172295689583, + 0.0020880065858364105, + 0.0028761615976691246, + 0.0015810420736670494, + 0.0018167231464758515, + 0.0015246851835399866, + 0.0023864894174039364, + 0.002109450986608863, + 0.001241007586941123, + 0.0018958136206492782, + 0.002691113855689764, + 0.0016059335321187973, + 0.0020019509829580784, + 0.002032024087384343, + 0.0030753612518310547, + 0.0022694976069033146, + 0.002779789734631777, + 0.001848213723860681, + 0.006155777722597122, + 0.002096864627674222, + 0.0031532244756817818, + 0.002025012858211994, + 0.0026636056136339903, + 0.004050966817885637, + 0.0018017698312178254, + 0.001502694678492844, + 0.003149711526930332, + 0.0019814693368971348, + 0.003221755614504218, + 0.001374669256620109, + 0.001914317486807704, + 0.002544191898778081, + 0.0012106767389923334, + 0.0012841656571254134, + 0.01911776140332222, + 0.010141784325242043, + 0.004514261148869991, + 0.0020562326535582542, + 0.004702985752373934, + 0.006615755148231983, + 0.001963434275239706, + 0.002467900048941374, + 0.018605269491672516, + 0.015888720750808716, + 0.011862073093652725, + 0.01624213717877865, + 0.01833268254995346, + 0.018383556976914406, + 0.007664302363991737, + 0.01034167967736721, + 0.008140966296195984, + 0.010157669894397259, + 0.006281462963670492, + 0.003513908013701439, + 0.004761354997754097, + 0.007198710460215807, + 0.002320359693840146, + 0.004636201076209545, + 0.035520605742931366, + 0.019971588626503944, + 0.032844219356775284, + 0.011047582142055035, + 0.028470462188124657, + 0.2531448006629944, + 0.014164680615067482, + 0.23446451127529144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.007627411279827356 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03367885202169418 + ], + "opposite_logprob_tokens": [ + -3.4086787700653076 + ], + "logprob_reward": -0.03367885202169418, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03367885202169418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12943318486213684, + 0.12026157230138779, + 0.04229549691081047, + 0.03710430860519409, + 0.08034069836139679, + 0.04061312973499298, + 0.06362438201904297, + 0.06757774949073792, + 0.040105607360601425, + 0.37864384055137634 + ], + "credit_map": [ + 0.026061462238430977, + 0.03106236830353737, + 0.013570505194365978, + 0.011973411776125431, + 0.010796879418194294, + 0.015685997903347015, + 0.0068475715816020966, + 0.013434993103146553, + 0.017283445224165916, + 0.014178458601236343, + 0.01462626364082098, + 0.02320760302245617, + 0.01160715613514185, + 0.01235638465732336, + 0.019099166616797447, + 0.00790309440344572, + 0.006012254860252142, + 0.0037619825452566147, + 0.0063874926418066025, + 0.004737586248666048, + 0.006677489262074232, + 0.005243490915745497, + 0.004778935573995113, + 0.004696268122643232, + 0.0059320577420294285, + 0.004043031949549913, + 0.005122387781739235, + 0.004195026587694883, + 0.006531599909067154, + 0.004061115439981222, + 0.004152467008680105, + 0.0030666240490972996, + 0.013072063215076923, + 0.008119232952594757, + 0.007812160067260265, + 0.00860709510743618, + 0.010374965146183968, + 0.010540486313402653, + 0.013621868565678596, + 0.008192826993763447, + 0.004579711705446243, + 0.004243740811944008, + 0.004686230327934027, + 0.003271352266892791, + 0.010834592394530773, + 0.004538529086858034, + 0.004670435097068548, + 0.0037885357160121202, + 0.007469130214303732, + 0.016043856739997864, + 0.007382260635495186, + 0.008054133504629135, + 0.009398538619279861, + 0.007555496878921986, + 0.004629903472959995, + 0.003091066377237439, + 0.011221876367926598, + 0.008402382954955101, + 0.006684464402496815, + 0.008231887593865395, + 0.01175861619412899, + 0.010410935617983341, + 0.006433530244976282, + 0.00443405844271183, + 0.005333154462277889, + 0.005350739695131779, + 0.00411622878164053, + 0.004887561779469252, + 0.006877628155052662, + 0.005242450162768364, + 0.00474542984738946, + 0.0035524158738553524, + 0.021417342126369476, + 0.013671543449163437, + 0.020497117191553116, + 0.012196349911391735, + 0.01805119588971138, + 0.1451834887266159, + 0.010558241978287697, + 0.13706855475902557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.03367885202169418 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "horizontal tracking shot, the camera moving with a steady, fluid rhythm, closely follows a Taoist priest in a cyan robe from behind as he journeys through a lush bamboo forest shrouded in a milky white mist. His long, jet-black hair is neatly tied into a chignon, secured with an antique wooden hairpin. His loose cyan cotton and linen robe appears even more subdued in the humid air, its sleeves swaying gently behind him as he walks leisurely. Tall, upright bamboos form a natural, deep corridor, the ground carpeted with layers of withered, yellowed leaves, making it nearly silent when stepped on. The soft, scattered morning light penetrates the thin mist and dense bamboo leaves, casting ever-changing dappled light and shadow on the Taoist's back and the path ahead. He holds a white whisk in his left hand, which sways rhythmically as his body moves. His wide robe occasionally brushes against the dew-covered bamboo leaves beside him. The entire picture presents a sense of seclusion, tranquility and mystery, creating a focused and profound atmosphere of Eastern Zen.", + "seed": 31367, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_6/index_3063_seed_31367_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.025595721416175367, + "vlm_logprob": -0.025595721416175367, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.025595721416175367, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06354208290576935 + ], + "opposite_logprob_tokens": [ + -2.813542127609253 + ], + "logprob_reward": -0.06354208290576935, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06354208290576935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26554471254348755, + 0.1695876270532608, + 0.09361079335212708, + 0.04641137272119522, + 0.05872486159205437, + 0.18691831827163696, + 0.048405252397060394, + 0.05429200828075409, + 0.041358672082424164, + 0.03514636680483818 + ], + "credit_map": [ + 0.03656001761555672, + 0.05248427391052246, + 0.04130587726831436, + 0.024011770263314247, + 0.026409626007080078, + 0.02584182471036911, + 0.022885015234351158, + 0.036046311259269714, + 0.014244021847844124, + 0.03419482335448265, + 0.03001490980386734, + 0.01685038022696972, + 0.01696353405714035, + 0.01929371990263462, + 0.017325840890407562, + 0.02070038951933384, + 0.012605938129127026, + 0.0163915753364563, + 0.012606601230800152, + 0.011203598231077194, + 0.007823912426829338, + 0.018002824857831, + 0.006773844826966524, + 0.008202495984733105, + 0.006032689940184355, + 0.00505687715485692, + 0.006622948683798313, + 0.004865544848144054, + 0.006197921931743622, + 0.004735351540148258, + 0.007190374657511711, + 0.0057096634991467, + 0.00830795057117939, + 0.005158576183021069, + 0.0064140246249735355, + 0.00749108986929059, + 0.006890606600791216, + 0.00875328853726387, + 0.010724613443017006, + 0.004984714090824127, + 0.024663731455802917, + 0.02386929839849472, + 0.017148062586784363, + 0.03163027763366699, + 0.020540360361337662, + 0.022974759340286255, + 0.03265979886054993, + 0.013432029634714127, + 0.006137004587799311, + 0.005487579386681318, + 0.005130583420395851, + 0.00596897816285491, + 0.0051697976887226105, + 0.0054231323301792145, + 0.007074868772178888, + 0.008013308979570866, + 0.006078333593904972, + 0.006942637264728546, + 0.005872227717190981, + 0.005807678680866957, + 0.007844161242246628, + 0.006035961210727692, + 0.010306360200047493, + 0.00540465023368597, + 0.004812618251889944, + 0.003906867001205683, + 0.006613392382860184, + 0.005295421928167343, + 0.005625774618238211, + 0.0040704053826630116, + 0.004923655651509762, + 0.00611053965985775, + 0.007287141866981983, + 0.0038320410531014204, + 0.004856782034039497, + 0.0027965782210230827, + 0.003729776246473193, + 0.0038526083808392286, + 0.00458786403760314, + 0.004203575663268566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.06354208290576935 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016090773046016693 + ], + "opposite_logprob_tokens": [ + -4.1410908699035645 + ], + "logprob_reward": -0.016090773046016693, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016090773046016693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29719385504722595, + 0.06436654925346375, + 0.09580481052398682, + 0.11485328525304794, + 0.037305619567632675, + 0.1120472326874733, + 0.05331593006849289, + 0.08640249073505402, + 0.07051330804824829, + 0.06819693744182587 + ], + "credit_map": [ + 0.030887356027960777, + 0.04014779254794121, + 0.03904275596141815, + 0.041761577129364014, + 0.05537484586238861, + 0.030315542593598366, + 0.0295913964509964, + 0.03007257543504238, + 0.006270186044275761, + 0.006562249734997749, + 0.006499475799500942, + 0.007321786601096392, + 0.007914260029792786, + 0.009155175648629665, + 0.008982159197330475, + 0.011661254800856113, + 0.012636158615350723, + 0.012566017918288708, + 0.016374314203858376, + 0.006697260774672031, + 0.0059152268804609776, + 0.02986997179687023, + 0.0051524825394153595, + 0.00659338291734457, + 0.013259639032185078, + 0.010055748745799065, + 0.019195904955267906, + 0.014185411855578423, + 0.012808098457753658, + 0.014076272957026958, + 0.014755385927855968, + 0.016516828909516335, + 0.006806453689932823, + 0.005485587287694216, + 0.00446271151304245, + 0.0036983955651521683, + 0.0030885401647537947, + 0.004690465983003378, + 0.004883317742496729, + 0.004190150648355484, + 0.016276588663458824, + 0.010596984066069126, + 0.012479457072913647, + 0.015513254329562187, + 0.008432252332568169, + 0.012549717910587788, + 0.016731970012187958, + 0.019467005506157875, + 0.010193086229264736, + 0.008074586279690266, + 0.007003201637417078, + 0.005147337913513184, + 0.004995431285351515, + 0.006520846392959356, + 0.004447474610060453, + 0.006933965254575014, + 0.012042438611388206, + 0.010346162132918835, + 0.009863891638815403, + 0.009905229322612286, + 0.010444195941090584, + 0.009794591926038265, + 0.012439483776688576, + 0.011566493660211563, + 0.008792266249656677, + 0.006069201976060867, + 0.009345948696136475, + 0.007685269229114056, + 0.009385839104652405, + 0.007802620064467192, + 0.010339737869799137, + 0.011092424392700195, + 0.010833521373569965, + 0.010043611750006676, + 0.012117440812289715, + 0.00565958209335804, + 0.006713261362165213, + 0.008591828867793083, + 0.005209545139223337, + 0.00902814045548439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.016090773046016693 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533940374851227 + ], + "opposite_logprob_tokens": [ + -4.387534141540527 + ], + "logprob_reward": -0.012533940374851227, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533940374851227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10793141275644302, + 0.13149641454219818, + 0.18861229717731476, + 0.0483165942132473, + 0.047752633690834045, + 0.20957721769809723, + 0.07615537196397781, + 0.08873028308153152, + 0.037201184779405594, + 0.06422656029462814 + ], + "credit_map": [ + 0.014503025449812412, + 0.018505854532122612, + 0.015649959444999695, + 0.014447913505136967, + 0.010300531052052975, + 0.010648088529706001, + 0.011075709946453571, + 0.012800334952771664, + 0.014135727658867836, + 0.028028476983308792, + 0.018246999010443687, + 0.011658716015517712, + 0.012799068354070187, + 0.02153177373111248, + 0.010086368769407272, + 0.015009290538728237, + 0.02316327951848507, + 0.02528689056634903, + 0.033170826733112335, + 0.008478617295622826, + 0.00740751763805747, + 0.07128966599702835, + 0.007498495746403933, + 0.012317005544900894, + 0.0063124834559857845, + 0.004350314848124981, + 0.007242739666253328, + 0.006316004786640406, + 0.004232619423419237, + 0.006822943687438965, + 0.005238679703325033, + 0.007800808642059565, + 0.012899111025035381, + 0.006056207697838545, + 0.00832174438983202, + 0.0034656403586268425, + 0.002732700202614069, + 0.005883206147700548, + 0.0036464850418269634, + 0.004747538361698389, + 0.02547757513821125, + 0.018053827807307243, + 0.021004600450396538, + 0.03274080529808998, + 0.014173183590173721, + 0.028805414214730263, + 0.03511808067560196, + 0.03420371934771538, + 0.015017334371805191, + 0.01129460521042347, + 0.009714571759104729, + 0.008141297847032547, + 0.0072492873296141624, + 0.008651680313050747, + 0.005970322992652655, + 0.010116268880665302, + 0.016605719923973083, + 0.011089284904301167, + 0.011034888215363026, + 0.0115109384059906, + 0.007498256396502256, + 0.010666478425264359, + 0.009770256467163563, + 0.01055445522069931, + 0.005912773311138153, + 0.007350269239395857, + 0.005203717853873968, + 0.003595613408833742, + 0.0037738748360425234, + 0.004648056346923113, + 0.0035794733557850122, + 0.003137405263260007, + 0.007664129137992859, + 0.010999745689332485, + 0.009117229841649532, + 0.006498575210571289, + 0.004148430190980434, + 0.007357888855040073, + 0.00594798568636179, + 0.012492572888731956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.012533940374851227 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538178823888302 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012538178823888302, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538178823888302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20340055227279663, + 0.14625278115272522, + 0.09014952927827835, + 0.09381545335054398, + 0.06479111313819885, + 0.051180433481931686, + 0.061155837029218674, + 0.12988334894180298, + 0.09541581571102142, + 0.06395508348941803 + ], + "credit_map": [ + 0.027984343469142914, + 0.03787121921777725, + 0.029787512496113777, + 0.01845712400972843, + 0.023167844861745834, + 0.030602624639868736, + 0.0123487813398242, + 0.02318110316991806, + 0.015706971287727356, + 0.026993421837687492, + 0.022105926647782326, + 0.015966016799211502, + 0.017550351098179817, + 0.012282028794288635, + 0.0211653970181942, + 0.014482670463621616, + 0.013520346023142338, + 0.014428956434130669, + 0.014105425216257572, + 0.006863664835691452, + 0.00577157735824585, + 0.024903714656829834, + 0.004905357491225004, + 0.005650486331433058, + 0.017299456521868706, + 0.008846866898238659, + 0.014986594207584858, + 0.009595763869583607, + 0.00978004839271307, + 0.01283765584230423, + 0.009057107381522655, + 0.011411958374083042, + 0.015855561941862106, + 0.00734346779063344, + 0.01052669994533062, + 0.005369691643863916, + 0.0047163283452391624, + 0.007607456296682358, + 0.006833855528384447, + 0.006538049783557653, + 0.006991623435169458, + 0.004552739672362804, + 0.005804102402180433, + 0.006828689482063055, + 0.0048815165646374226, + 0.009812049567699432, + 0.007706796284765005, + 0.004602920264005661, + 0.0141804413869977, + 0.00916117150336504, + 0.008654162287712097, + 0.0046755727380514145, + 0.005109849385917187, + 0.0063338554464280605, + 0.006019982509315014, + 0.0070207989774644375, + 0.02549702115356922, + 0.021871667355298996, + 0.018468502908945084, + 0.011556235142052174, + 0.013721758499741554, + 0.01416196022182703, + 0.012353425845503807, + 0.01225278154015541, + 0.014265058562159538, + 0.016132645308971405, + 0.011130108498036861, + 0.010453945957124233, + 0.00905163586139679, + 0.012749405577778816, + 0.009188610129058361, + 0.012444399297237396, + 0.011274004355072975, + 0.005528367590159178, + 0.012196728959679604, + 0.0055368822067976, + 0.004999512806534767, + 0.007624966558068991, + 0.006622200831770897, + 0.010172420181334019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.012538178823888302 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023273631930351257 + ], + "opposite_logprob_tokens": [ + -3.7732737064361572 + ], + "logprob_reward": -0.023273631930351257, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023273631930351257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1532946527004242, + 0.09261364489793777, + 0.10668821632862091, + 0.08246836811304092, + 0.03777103126049042, + 0.11691004037857056, + 0.20942047238349915, + 0.07140348851680756, + 0.07111652940511703, + 0.058313559740781784 + ], + "credit_map": [ + 0.024633610621094704, + 0.02458159066736698, + 0.020198481157422066, + 0.014325692318379879, + 0.017909282818436623, + 0.013256547972559929, + 0.015960577875375748, + 0.022428875789046288, + 0.008008867502212524, + 0.019906485453248024, + 0.017557930201292038, + 0.00866464339196682, + 0.009149130433797836, + 0.010174962691962719, + 0.011553947813808918, + 0.007597675547003746, + 0.014969726093113422, + 0.01633448339998722, + 0.01701662316918373, + 0.006211753003299236, + 0.005132513586431742, + 0.03446853905916214, + 0.0046943617053329945, + 0.007860207930207253, + 0.018025150522589684, + 0.005465454421937466, + 0.0170553270727396, + 0.009303578175604343, + 0.0070682731457054615, + 0.007805459666997194, + 0.008231543004512787, + 0.009513582102954388, + 0.006217899266630411, + 0.005087094381451607, + 0.004627391695976257, + 0.003230166854336858, + 0.002316121244803071, + 0.0064035081304609776, + 0.004349636845290661, + 0.005539213307201862, + 0.012767082080245018, + 0.014493047259747982, + 0.011394670233130455, + 0.019976992160081863, + 0.014087817631661892, + 0.010795905254781246, + 0.0211481936275959, + 0.01224633026868105, + 0.03378455713391304, + 0.022387461736798286, + 0.0219071414321661, + 0.026438666507601738, + 0.021781783550977707, + 0.02137751691043377, + 0.033117394894361496, + 0.028625963255763054, + 0.012200280092656612, + 0.008887982927262783, + 0.007749348413199186, + 0.007430930156260729, + 0.006124675273895264, + 0.007618280593305826, + 0.011971605010330677, + 0.009420380927622318, + 0.009464803151786327, + 0.007704779040068388, + 0.01128432061523199, + 0.007076875306665897, + 0.008831694722175598, + 0.007312553469091654, + 0.00896240770816803, + 0.010479096323251724, + 0.00803474709391594, + 0.006323165725916624, + 0.0067670829594135284, + 0.006101198960095644, + 0.007268940098583698, + 0.0065631805919110775, + 0.009928179904818535, + 0.007327066734433174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.023273631930351257 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "horizontal tracking shot, the camera moving with a steady, fluid rhythm, closely follows a Taoist priest in a cyan robe from behind as he journeys through a lush bamboo forest shrouded in a milky white mist. His long, jet-black hair is neatly tied into a chignon, secured with an antique wooden hairpin. His loose cyan cotton and linen robe appears even more subdued in the humid air, its sleeves swaying gently behind him as he walks leisurely. Tall, upright bamboos form a natural, deep corridor, the ground carpeted with layers of withered, yellowed leaves, making it nearly silent when stepped on. The soft, scattered morning light penetrates the thin mist and dense bamboo leaves, casting ever-changing dappled light and shadow on the Taoist's back and the path ahead. He holds a white whisk in his left hand, which sways rhythmically as his body moves. His wide robe occasionally brushes against the dew-covered bamboo leaves beside him. The entire picture presents a sense of seclusion, tranquility and mystery, creating a focused and profound atmosphere of Eastern Zen.", + "seed": 31368, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_6/index_3063_seed_31368_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.026126482151448728, + "vlm_logprob": -0.026126482151448728, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.026126482151448728, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06371792405843735 + ], + "opposite_logprob_tokens": [ + -2.813717842102051 + ], + "logprob_reward": -0.06371792405843735, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06371792405843735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10894272476434708, + 0.529293954372406, + 0.0482013076543808, + 0.051252637058496475, + 0.035399001091718674, + 0.03287272900342941, + 0.029102282598614693, + 0.030353421345353127, + 0.09839767962694168, + 0.03618425875902176 + ], + "credit_map": [ + 0.0115697355940938, + 0.015243830159306526, + 0.014459474012255669, + 0.014225033111870289, + 0.011748358607292175, + 0.018104983493685722, + 0.009325610473752022, + 0.01426569651812315, + 0.07936924695968628, + 0.03225177153944969, + 0.21871033310890198, + 0.024665724486112595, + 0.021991468966007233, + 0.08952609449625015, + 0.013961450196802616, + 0.04881783574819565, + 0.007927114143967628, + 0.003017027396708727, + 0.006455898284912109, + 0.006175286136567593, + 0.0044973473995924, + 0.007117037195712328, + 0.006187423598021269, + 0.006824173964560032, + 0.006005322560667992, + 0.0028313524089753628, + 0.005241657607257366, + 0.010204700753092766, + 0.005646490026265383, + 0.007148537784814835, + 0.008053647354245186, + 0.006120930891484022, + 0.004592906218022108, + 0.0030385300051420927, + 0.003952475264668465, + 0.004864340648055077, + 0.0045148516073822975, + 0.004742203745990992, + 0.0038904757238924503, + 0.005803218111395836, + 0.002968517364934087, + 0.003758936421945691, + 0.002722457516938448, + 0.006351250223815441, + 0.003599790623411536, + 0.004342486150562763, + 0.005539414472877979, + 0.00358987576328218, + 0.0027994466945528984, + 0.002401934936642647, + 0.0030771056190133095, + 0.004881864879280329, + 0.004858002532273531, + 0.003489662893116474, + 0.0037961937487125397, + 0.0037980705965310335, + 0.004893174860626459, + 0.0031741494312882423, + 0.0034122769720852375, + 0.00439067929983139, + 0.004145185928791761, + 0.003539902623742819, + 0.0027161624748259783, + 0.004081891383975744, + 0.022055204957723618, + 0.008365819230675697, + 0.016791274771094322, + 0.011632719077169895, + 0.006869292352348566, + 0.01077256165444851, + 0.008084259927272797, + 0.013826549984514713, + 0.005133698228746653, + 0.0031798805575817823, + 0.0025595049373805523, + 0.0050844913348555565, + 0.0035491464659571648, + 0.005244492553174496, + 0.007014978211373091, + 0.004418064374476671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.06371792405843735 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018204277381300926 + ], + "opposite_logprob_tokens": [ + -4.018204212188721 + ], + "logprob_reward": -0.018204277381300926, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018204277381300926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13746578991413116, + 0.502013087272644, + 0.05465890094637871, + 0.06302855163812637, + 0.03359029442071915, + 0.02797027677297592, + 0.02956520952284336, + 0.023016121238470078, + 0.061206888407468796, + 0.06748485565185547 + ], + "credit_map": [ + 0.016628075391054153, + 0.02549409307539463, + 0.015041148290038109, + 0.013170113787055016, + 0.01965942420065403, + 0.016851386055350304, + 0.01091412827372551, + 0.01970743201673031, + 0.07264513522386551, + 0.03064895235002041, + 0.21214525401592255, + 0.022282348945736885, + 0.017900237813591957, + 0.048681773245334625, + 0.012043459340929985, + 0.08566593378782272, + 0.004937452729791403, + 0.005075530149042606, + 0.0051316930912435055, + 0.009091440588235855, + 0.007748757489025593, + 0.008029933087527752, + 0.009247870184481144, + 0.005396231543272734, + 0.006614417303353548, + 0.003733576973900199, + 0.007506218273192644, + 0.00705079548060894, + 0.0090797059237957, + 0.013320225290954113, + 0.005113497842103243, + 0.010610119439661503, + 0.0028156714979559183, + 0.0026131393387913704, + 0.003414937062188983, + 0.006263185758143663, + 0.003587736515328288, + 0.0046485052444040775, + 0.004537454340606928, + 0.005709667690098286, + 0.003281103679910302, + 0.0035576587542891502, + 0.001900453818961978, + 0.004105505533516407, + 0.005353089887648821, + 0.0035204796586185694, + 0.003241196973249316, + 0.0030107905622571707, + 0.0030517454724758863, + 0.0036072994116693735, + 0.003158529521897435, + 0.004163084086030722, + 0.004249297082424164, + 0.0039382148534059525, + 0.00228770705871284, + 0.005109337158501148, + 0.00292911846190691, + 0.003245918545871973, + 0.0026208364870399237, + 0.0035035365726798773, + 0.0031652546022087336, + 0.0031337563414126635, + 0.0016949301352724433, + 0.0027227727696299553, + 0.015237966552376747, + 0.006721402984112501, + 0.010823527351021767, + 0.0066161262802779675, + 0.0029508855659514666, + 0.0069503202103078365, + 0.00395113555714488, + 0.007955528795719147, + 0.015315954573452473, + 0.007657591253519058, + 0.007651765365153551, + 0.008417999371886253, + 0.003547533880919218, + 0.010388923808932304, + 0.005869584623724222, + 0.008635506965219975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.018204277381300926 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775977581739426 + ], + "opposite_logprob_tokens": [ + -4.6347761154174805 + ], + "logprob_reward": -0.009775977581739426, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775977581739426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14397336542606354, + 0.2685557007789612, + 0.06165355071425438, + 0.05339323356747627, + 0.04566052556037903, + 0.11881662905216217, + 0.06056937202811241, + 0.09765999764204025, + 0.05062330514192581, + 0.09909434616565704 + ], + "credit_map": [ + 0.012730659916996956, + 0.014688658528029919, + 0.018523328006267548, + 0.014846131205558777, + 0.01487339474260807, + 0.03454393893480301, + 0.011946795508265495, + 0.021820448338985443, + 0.03709958493709564, + 0.01664966717362404, + 0.11304766684770584, + 0.012520933523774147, + 0.010163058526813984, + 0.036372099071741104, + 0.007218535989522934, + 0.03548412397503853, + 0.008214525878429413, + 0.005346069112420082, + 0.008360524661839008, + 0.008780336938798428, + 0.007061399053782225, + 0.007797003258019686, + 0.010144309140741825, + 0.005949374288320541, + 0.006885575596243143, + 0.0047376202419400215, + 0.008255008608102798, + 0.008529726415872574, + 0.007188388146460056, + 0.006909864488989115, + 0.0039627160876989365, + 0.0069243283942341805, + 0.004710891284048557, + 0.004951625131070614, + 0.005843077786266804, + 0.007105524186044931, + 0.003528582165017724, + 0.006235227454453707, + 0.003592114429920912, + 0.009693475440144539, + 0.01946079172194004, + 0.011993864551186562, + 0.012053849175572395, + 0.023297686129808426, + 0.012846159748733044, + 0.018021240830421448, + 0.007497767452150583, + 0.013645259663462639, + 0.00667992839589715, + 0.010069996118545532, + 0.009120761416852474, + 0.00755102327093482, + 0.0043441797606647015, + 0.009344381280243397, + 0.0036453106440603733, + 0.00981378648430109, + 0.014484845101833344, + 0.015825096517801285, + 0.011158466339111328, + 0.017039278522133827, + 0.009427967481315136, + 0.012581190094351768, + 0.006190109997987747, + 0.010953031480312347, + 0.008283943869173527, + 0.005557392258197069, + 0.006628683302551508, + 0.005932456348091364, + 0.004001019988209009, + 0.005927307531237602, + 0.003979792818427086, + 0.010312706232070923, + 0.02107144519686699, + 0.010763744823634624, + 0.009625278413295746, + 0.017153048887848854, + 0.005953493993729353, + 0.012335977517068386, + 0.013028296642005444, + 0.00916304811835289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.009775977581739426 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252727307379246 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.005252727307379246, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252727307379246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13218949735164642, + 0.11335115879774094, + 0.0634094700217247, + 0.15171538293361664, + 0.07294254750013351, + 0.061056531965732574, + 0.06115929037332535, + 0.04494805634021759, + 0.18239431083202362, + 0.11683379858732224 + ], + "credit_map": [ + 0.022263193503022194, + 0.01615850068628788, + 0.024963755160570145, + 0.011375660076737404, + 0.012049322947859764, + 0.021087083965539932, + 0.008710547350347042, + 0.015581403858959675, + 0.019329126924276352, + 0.007902709767222404, + 0.044225916266441345, + 0.0051906658336520195, + 0.005257613956928253, + 0.010329910553991795, + 0.0029822257347404957, + 0.018132973462343216, + 0.010280195623636246, + 0.006016366183757782, + 0.008661569096148014, + 0.007712304126471281, + 0.00938110239803791, + 0.008075747638940811, + 0.006577410735189915, + 0.006704764906316996, + 0.018902525305747986, + 0.010679698549211025, + 0.014163408428430557, + 0.018748493865132332, + 0.012523086741566658, + 0.036546289920806885, + 0.015563677996397018, + 0.024588195607066154, + 0.008195580914616585, + 0.0066962228156626225, + 0.007723516318947077, + 0.011035630479454994, + 0.007503993809223175, + 0.009190152399241924, + 0.010453565046191216, + 0.012143879197537899, + 0.008794681169092655, + 0.009904535487294197, + 0.006407565902918577, + 0.008270409889519215, + 0.00707977544516325, + 0.009047281928360462, + 0.005336381494998932, + 0.006215890869498253, + 0.007548559457063675, + 0.008533569984138012, + 0.008614413440227509, + 0.007518189959228039, + 0.00519185233861208, + 0.009044949896633625, + 0.00596209429204464, + 0.008745651692152023, + 0.005267293192446232, + 0.007748507894575596, + 0.004440010525286198, + 0.007180629763752222, + 0.004329907242208719, + 0.006727485451847315, + 0.003908045124262571, + 0.005346174351871014, + 0.0351451076567173, + 0.030090926215052605, + 0.02765360102057457, + 0.014690445736050606, + 0.011361798271536827, + 0.021304244175553322, + 0.010380497202277184, + 0.03176767751574516, + 0.011569577269256115, + 0.010053603909909725, + 0.007143838331103325, + 0.022807668894529343, + 0.010187407955527306, + 0.01530865766108036, + 0.025030914694070816, + 0.0147321168333292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.005252727307379246 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368150442838669 + ], + "opposite_logprob_tokens": [ + -3.4086813926696777 + ], + "logprob_reward": -0.03368150442838669, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368150442838669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07003751397132874, + 0.48633038997650146, + 0.11293920129537582, + 0.05641556158661842, + 0.023216402158141136, + 0.025733480229973793, + 0.03193029761314392, + 0.02480665035545826, + 0.14682938158512115, + 0.021761100739240646 + ], + "credit_map": [ + 0.015421166084706783, + 0.007559637073427439, + 0.009079129435122013, + 0.004972632043063641, + 0.007118620909750462, + 0.009517052210867405, + 0.006324138026684523, + 0.010045145638287067, + 0.08939249813556671, + 0.030267445370554924, + 0.203456312417984, + 0.020409084856510162, + 0.014830281026661396, + 0.04815157875418663, + 0.011044181883335114, + 0.06877899169921875, + 0.016674622893333435, + 0.0063263061456382275, + 0.016879430040717125, + 0.015363374724984169, + 0.011381094343960285, + 0.01833302341401577, + 0.011234352365136147, + 0.01674700900912285, + 0.005529576912522316, + 0.003461390035226941, + 0.0039314646273851395, + 0.01122494786977768, + 0.00790505949407816, + 0.007940469309687614, + 0.009829181246459484, + 0.006593470927327871, + 0.002488043624907732, + 0.0020088739693164825, + 0.0024451774079352617, + 0.004337132442742586, + 0.002680286532267928, + 0.003064085030928254, + 0.003498069243505597, + 0.0026947329752147198, + 0.0035983757115900517, + 0.002433747984468937, + 0.002455132082104683, + 0.005427427124232054, + 0.00302870268933475, + 0.003544093109667301, + 0.002512671984732151, + 0.0027333295438438654, + 0.003149523166939616, + 0.004729923792183399, + 0.00332919810898602, + 0.004752070643007755, + 0.0027582368347793818, + 0.0048224953934550285, + 0.004012598656117916, + 0.004376252181828022, + 0.0030618750024586916, + 0.0028827234636992216, + 0.002624685177579522, + 0.0047720312140882015, + 0.0032959585078060627, + 0.003280848264694214, + 0.0024151825346052647, + 0.0024733450263738632, + 0.03864678367972374, + 0.014165662229061127, + 0.028471024706959724, + 0.01549590565264225, + 0.005438131280243397, + 0.016935717314481735, + 0.007120079360902309, + 0.02055608108639717, + 0.002228465164080262, + 0.0022685970179736614, + 0.001193374628201127, + 0.0038449726998806, + 0.0022772864904254675, + 0.0027291332371532917, + 0.004349978175014257, + 0.0028692930936813354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.03368150442838669 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "horizontal tracking shot, the camera moving with a steady, fluid rhythm, closely follows a Taoist priest in a cyan robe from behind as he journeys through a lush bamboo forest shrouded in a milky white mist. His long, jet-black hair is neatly tied into a chignon, secured with an antique wooden hairpin. His loose cyan cotton and linen robe appears even more subdued in the humid air, its sleeves swaying gently behind him as he walks leisurely. Tall, upright bamboos form a natural, deep corridor, the ground carpeted with layers of withered, yellowed leaves, making it nearly silent when stepped on. The soft, scattered morning light penetrates the thin mist and dense bamboo leaves, casting ever-changing dappled light and shadow on the Taoist's back and the path ahead. He holds a white whisk in his left hand, which sways rhythmically as his body moves. His wide robe occasionally brushes against the dew-covered bamboo leaves beside him. The entire picture presents a sense of seclusion, tranquility and mystery, creating a focused and profound atmosphere of Eastern Zen.", + "seed": 31369, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_6/index_3063_seed_31369_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.028871123678982258, + "vlm_logprob": -0.028871123678982258, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.028871123678982258, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07191586494445801 + ], + "opposite_logprob_tokens": [ + -2.696915864944458 + ], + "logprob_reward": -0.07191586494445801, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07191586494445801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4918892979621887, + 0.04613928124308586, + 0.09645753353834152, + 0.05340373516082764, + 0.09117089956998825, + 0.03778942674398422, + 0.044418856501579285, + 0.02947709523141384, + 0.04865249618887901, + 0.060601308941841125 + ], + "credit_map": [ + 0.09561163932085037, + 0.07943522930145264, + 0.07106047123670578, + 0.06179286167025566, + 0.05135897547006607, + 0.07092133909463882, + 0.02905380167067051, + 0.03265497460961342, + 0.007907449267804623, + 0.005291588604450226, + 0.0043837446719408035, + 0.007283124607056379, + 0.006391275208443403, + 0.005032719578593969, + 0.004384899511933327, + 0.005464483052492142, + 0.011124057695269585, + 0.018089061602950096, + 0.01030051615089178, + 0.01718045026063919, + 0.00824112817645073, + 0.006765080615878105, + 0.00684155710041523, + 0.01791568659245968, + 0.011923983693122864, + 0.004797240719199181, + 0.007313396316021681, + 0.005900186020880938, + 0.006870825309306383, + 0.0044237058609724045, + 0.003196952398866415, + 0.008977447636425495, + 0.02164408378303051, + 0.011169841513037682, + 0.02252301201224327, + 0.007156044244766235, + 0.009177973493933678, + 0.00743878586217761, + 0.005490663927048445, + 0.0065705059096217155, + 0.003996892832219601, + 0.003610372543334961, + 0.0037011848762631416, + 0.005530768074095249, + 0.00632382882758975, + 0.004663462750613689, + 0.004944985266774893, + 0.0050179315730929375, + 0.004872988443821669, + 0.0050657764077186584, + 0.005069669336080551, + 0.006296133156865835, + 0.006266997195780277, + 0.006151200272142887, + 0.004433964379131794, + 0.0062621342949569225, + 0.0027588398661464453, + 0.002688733860850334, + 0.0025427250657230616, + 0.005245467647910118, + 0.006227678153663874, + 0.003969984129071236, + 0.003018487012013793, + 0.0030251818243414164, + 0.005405704490840435, + 0.004571062978357077, + 0.004846849944442511, + 0.007938917726278305, + 0.0076933689415454865, + 0.00953700765967369, + 0.0044019282795488834, + 0.004257657565176487, + 0.004600906278938055, + 0.003934434149414301, + 0.00557746971026063, + 0.009068895131349564, + 0.012490615248680115, + 0.009118829853832722, + 0.010118779726326466, + 0.0056913779117167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.07191586494445801 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023308804258704185 + ], + "opposite_logprob_tokens": [ + -3.773308753967285 + ], + "logprob_reward": -0.023308804258704185, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023308804258704185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30990150570869446, + 0.11275658011436462, + 0.06129889190196991, + 0.098250612616539, + 0.08975189179182053, + 0.04181428626179695, + 0.018117630854249, + 0.029101181775331497, + 0.19664514064788818, + 0.04236230626702309 + ], + "credit_map": [ + 0.07364895939826965, + 0.04602809622883797, + 0.04840625450015068, + 0.041523825377225876, + 0.023228049278259277, + 0.038067545741796494, + 0.02008790895342827, + 0.018910860642790794, + 0.013667342253029346, + 0.009453660808503628, + 0.012382986955344677, + 0.028135841712355614, + 0.015110925771296024, + 0.011348412372171879, + 0.013140983879566193, + 0.009516430087387562, + 0.006757035851478577, + 0.01233016885817051, + 0.006429347209632397, + 0.008076105266809464, + 0.0068620964884757996, + 0.004938545171171427, + 0.003531485563144088, + 0.012374106794595718, + 0.020318862050771713, + 0.01314729917794466, + 0.012829339131712914, + 0.014792516827583313, + 0.010600283741950989, + 0.006768437102437019, + 0.007795235607773066, + 0.011998632922768593, + 0.020240573212504387, + 0.006353904027491808, + 0.01715479977428913, + 0.008200504817068577, + 0.009998764842748642, + 0.008843687362968922, + 0.004109642468392849, + 0.01485002227127552, + 0.004313843324780464, + 0.005429910030215979, + 0.0054777259938418865, + 0.0048873769119381905, + 0.007443691603839397, + 0.005560866557061672, + 0.002667575841769576, + 0.00603329436853528, + 0.00248995004221797, + 0.002171907341107726, + 0.0026126750744879246, + 0.0022858695592731237, + 0.0024371652398258448, + 0.0020708651281893253, + 0.0016715462552383542, + 0.002377650933340192, + 0.00444265129044652, + 0.0031845432240515947, + 0.003853878239169717, + 0.005196113605052233, + 0.0033501191064715385, + 0.004374679643660784, + 0.0021672251168638468, + 0.002531970851123333, + 0.02039932645857334, + 0.013325618579983711, + 0.014841444790363312, + 0.025447236374020576, + 0.038056980818510056, + 0.050616856664419174, + 0.018338168039917946, + 0.015619506128132343, + 0.004047781694680452, + 0.004257249645888805, + 0.00581156462430954, + 0.006858703680336475, + 0.0058334567584097385, + 0.00607768539339304, + 0.005146392155438662, + 0.004329472780227661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.023308804258704185 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016062384471297264 + ], + "opposite_logprob_tokens": [ + -4.141062259674072 + ], + "logprob_reward": -0.016062384471297264, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016062384471297264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12432236969470978, + 0.06548049300909042, + 0.258012056350708, + 0.061910487711429596, + 0.09404805302619934, + 0.053149618208408356, + 0.07853437215089798, + 0.05454612150788307, + 0.08748536556959152, + 0.12251109629869461 + ], + "credit_map": [ + 0.01966540329158306, + 0.009870203211903572, + 0.016927527263760567, + 0.020316854119300842, + 0.020278582349419594, + 0.014601892791688442, + 0.011175314895808697, + 0.011486579664051533, + 0.00996490754187107, + 0.007386802230030298, + 0.009726555086672306, + 0.009008336812257767, + 0.01042910199612379, + 0.006875159684568644, + 0.005965359043329954, + 0.006124269217252731, + 0.033586613833904266, + 0.05790000408887863, + 0.024395685642957687, + 0.019687753170728683, + 0.019128432497382164, + 0.021792283281683922, + 0.012544997036457062, + 0.06897629052400589, + 0.012602328322827816, + 0.010271091014146805, + 0.006200612522661686, + 0.009219172410666943, + 0.007497763726860285, + 0.00435398006811738, + 0.00449181254953146, + 0.007273718249052763, + 0.015345782041549683, + 0.009206305257976055, + 0.02394418604671955, + 0.00892751943320036, + 0.012415316887199879, + 0.008587691932916641, + 0.00634174607694149, + 0.009279496967792511, + 0.007891654036939144, + 0.004807511810213327, + 0.007445632945746183, + 0.004415361676365137, + 0.006192181259393692, + 0.012076479382812977, + 0.00408314261585474, + 0.006237648893147707, + 0.014538287185132504, + 0.009319689124822617, + 0.014288847334682941, + 0.007859489880502224, + 0.008977913297712803, + 0.0094053465873003, + 0.005506092216819525, + 0.008638699539005756, + 0.008427707478404045, + 0.005889425985515118, + 0.007781504653394222, + 0.006424123886972666, + 0.006293227430433035, + 0.00827516708523035, + 0.004434550181031227, + 0.007020407821983099, + 0.012971448712050915, + 0.01049006450921297, + 0.01143624447286129, + 0.012872769497334957, + 0.009884791448712349, + 0.013289051130414009, + 0.008239777758717537, + 0.008301211521029472, + 0.015588751062750816, + 0.01731698215007782, + 0.016464168205857277, + 0.02143716625869274, + 0.011530313640832901, + 0.014337548986077309, + 0.012744979001581669, + 0.013091174885630608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.016062384471297264 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736070849001408 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.006736070849001408, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736070849001408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20987825095653534, + 0.08635015040636063, + 0.14045342803001404, + 0.19199219346046448, + 0.09235475212335587, + 0.05634118616580963, + 0.032984890043735504, + 0.044676780700683594, + 0.07768599689006805, + 0.06728241592645645 + ], + "credit_map": [ + 0.05027920752763748, + 0.032310906797647476, + 0.033583443611860275, + 0.02724321186542511, + 0.01523735374212265, + 0.018737515434622765, + 0.015079235658049583, + 0.017407365143299103, + 0.013684121891856194, + 0.010570080950856209, + 0.01054344605654478, + 0.014418545179069042, + 0.011767380870878696, + 0.01092915702611208, + 0.007146600633859634, + 0.007290818728506565, + 0.016027279198169708, + 0.018580622971057892, + 0.013221136294305325, + 0.0255805104970932, + 0.014174772426486015, + 0.010448010638356209, + 0.01527902390807867, + 0.027142060920596123, + 0.04618849977850914, + 0.019808176904916763, + 0.02978166751563549, + 0.024567507207393646, + 0.02576262690126896, + 0.015453017316758633, + 0.013740857131779194, + 0.0166898425668478, + 0.025362422689795494, + 0.008616551756858826, + 0.02142230048775673, + 0.0069827004335820675, + 0.007554399780929089, + 0.007837281562387943, + 0.004916604608297348, + 0.009662488475441933, + 0.007157573010772467, + 0.005201821681112051, + 0.008641519583761692, + 0.0052154273726046085, + 0.006352085620164871, + 0.012288140133023262, + 0.004315790720283985, + 0.0071688247844576836, + 0.007159411907196045, + 0.003937308676540852, + 0.004738691728562117, + 0.0031107112299650908, + 0.003468654118478298, + 0.004524183459579945, + 0.002346460474655032, + 0.003699470078572631, + 0.007646095473319292, + 0.006283176597207785, + 0.005856064613908529, + 0.004608531016856432, + 0.00533656869083643, + 0.006917883176356554, + 0.004662689287215471, + 0.0033657734747976065, + 0.008683777414262295, + 0.005744472146034241, + 0.006524537689983845, + 0.011970619671046734, + 0.011986937373876572, + 0.020186256617307663, + 0.006385622080415487, + 0.006203776691108942, + 0.006690607406198978, + 0.009328445419669151, + 0.009983057156205177, + 0.009301419369876385, + 0.00865805335342884, + 0.008759274147450924, + 0.004568541422486305, + 0.009993020445108414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.006736070849001408 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Taoist priest wearing a cyan robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the priest's hair long, jet-black, tied in a chignon, and secured with an antique wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a lush bamboo forest with a ground covered in withered yellowed leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shrouded in a milky white mist with soft, scattered morning light creating dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a horizontal tracking shot moving steadily from behind the priest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026332493871450424 + ], + "opposite_logprob_tokens": [ + -3.651332378387451 + ], + "logprob_reward": -0.026332493871450424, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026332493871450424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0946144163608551, + 0.05835557356476784, + 0.271035373210907, + 0.09953375160694122, + 0.03542976826429367, + 0.049636561423540115, + 0.049179740250110626, + 0.07009149342775345, + 0.226393923163414, + 0.04572942107915878 + ], + "credit_map": [ + 0.013617108575999737, + 0.009463158436119556, + 0.009358588606119156, + 0.013987896032631397, + 0.016129983589053154, + 0.010397270321846008, + 0.011441652663052082, + 0.01021876186132431, + 0.006909962277859449, + 0.006297542713582516, + 0.006204304751008749, + 0.014392415061593056, + 0.007881926372647285, + 0.005224299617111683, + 0.004088175017386675, + 0.007356947287917137, + 0.029406890273094177, + 0.06777369976043701, + 0.024399206042289734, + 0.02271408587694168, + 0.02394021302461624, + 0.020800244063138962, + 0.015634125098586082, + 0.06636693328619003, + 0.021932166069746017, + 0.009571184404194355, + 0.01358663197606802, + 0.011833896860480309, + 0.010704100131988525, + 0.011227825656533241, + 0.008922271430492401, + 0.01175567414611578, + 0.008003061637282372, + 0.0031127063557505608, + 0.004775781184434891, + 0.002852463163435459, + 0.004007627256214619, + 0.005299813579767942, + 0.0019901879131793976, + 0.005388124380260706, + 0.005105020012706518, + 0.0050391219556331635, + 0.005675226449966431, + 0.00551029946655035, + 0.008955913595855236, + 0.008808779530227184, + 0.00492273923009634, + 0.005619458854198456, + 0.0050816722214221954, + 0.005423055030405521, + 0.0065390607342123985, + 0.006771298591047525, + 0.007326492108404636, + 0.006743373349308968, + 0.004946691449731588, + 0.006348094902932644, + 0.0074646007269620895, + 0.008354322984814644, + 0.00913984328508377, + 0.01186398509889841, + 0.011380032636225224, + 0.00862465426325798, + 0.006979742553085089, + 0.006284313276410103, + 0.021942265331745148, + 0.017549734562635422, + 0.01871057227253914, + 0.04614126309752464, + 0.033207785338163376, + 0.042150743305683136, + 0.025475185364484787, + 0.021216372027993202, + 0.005368029698729515, + 0.007609553635120392, + 0.0034858689177781343, + 0.005284021142870188, + 0.006230450700968504, + 0.006250186823308468, + 0.0038031490985304117, + 0.0076981643214821815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.026332493871450424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_7/index_tensor([5665, 5665])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_7/index_tensor([5665, 5665])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..359e44e48546071f3fffabebaea39b1ae1683819 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_7/index_tensor([5665, 5665])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static close-up captures a hyper-realistic CG rendering of a bright yellow handheld sprayer, carelessly tossed onto a weathered stone patio. Faint scuffs and a touch of dirt mark the sprayer's smooth plastic exterior, evidence of past use. A thick, viscous, vibrant green liquid gushes forth in a steady stream from a black nozzle, not splashing but clinging tenaciously to the rough, uneven surface of the gray and beige stone slabs. This unnatural fluid fizzes and bubbles with vibrant energy, its surface constantly fractured as tiny, shimmering spheres form, expand, and burst, releasing wisps of light vapor. Bathed in the warm golden glow of the afternoon sun, the scene casts long, soft shadows, capturing the translucent quality of the bubbling liquid, making it appear to radiate an eerie internal energy across the cool stone, creating a deeply unsettling yet visually captivating atmosphere.", + "seed": 22730, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_7/index_5665_seed_22730_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05373496543616056, + "vlm_logprob": -0.05373496543616056, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.05373496543616056, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1837664693593979 + ], + "opposite_logprob_tokens": [ + -1.808766484260559 + ], + "logprob_reward": -0.1837664693593979, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1837664693593979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33803990483283997, + 0.16085024178028107, + 0.10487790405750275, + 0.05598776414990425, + 0.061954740434885025, + 0.0695875883102417, + 0.06930562108755112, + 0.056030742824077606, + 0.06732530891895294, + 0.016040241345763206 + ], + "credit_map": [ + 0.039853453636169434, + 0.046909116208553314, + 0.03925749287009239, + 0.05738304555416107, + 0.04405372589826584, + 0.04616887494921684, + 0.03478000685572624, + 0.02963416464626789, + 0.0187092088162899, + 0.02168208733201027, + 0.016109181568026543, + 0.016114257276058197, + 0.031454768031835556, + 0.02145354449748993, + 0.01718871295452118, + 0.01813848316669464, + 0.011977762915194035, + 0.012635232880711555, + 0.010544950142502785, + 0.01100629661232233, + 0.020430337637662888, + 0.014290518127381802, + 0.012585650198161602, + 0.011407160200178623, + 0.0061416043899953365, + 0.004939014092087746, + 0.009053305722773075, + 0.011675785295665264, + 0.004409720189869404, + 0.007588271051645279, + 0.006438104901462793, + 0.005741951521486044, + 0.0076864431612193584, + 0.00696163484826684, + 0.006807093042880297, + 0.005930047482252121, + 0.009529872797429562, + 0.009117485024034977, + 0.011518286541104317, + 0.004403884056955576, + 0.010243295691907406, + 0.007277289405465126, + 0.006271892227232456, + 0.005833019968122244, + 0.012135456316173077, + 0.010188586078584194, + 0.010635399259626865, + 0.0070026470348238945, + 0.008712264709174633, + 0.009566759690642357, + 0.006324842572212219, + 0.00993445236235857, + 0.010817988775670528, + 0.01050443947315216, + 0.008073455654084682, + 0.00537141552194953, + 0.005898255854845047, + 0.005307473707944155, + 0.005124097689986229, + 0.010221073403954506, + 0.008374406024813652, + 0.007340162992477417, + 0.008698459714651108, + 0.00506680877879262, + 0.00539805181324482, + 0.006739273201674223, + 0.0036346991546452045, + 0.010179122909903526, + 0.007509895134717226, + 0.009731817990541458, + 0.01642525941133499, + 0.007707188371568918, + 0.0018416274106130004, + 0.00187768938485533, + 0.001254522823728621, + 0.0027075987309217453, + 0.0017425118712708354, + 0.002375048352405429, + 0.0026240202132612467, + 0.0016172225587069988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.1837664693593979 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810613602399826 + ], + "opposite_logprob_tokens": [ + -3.2881062030792236 + ], + "logprob_reward": -0.03810613602399826, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810613602399826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40811794996261597, + 0.1441243439912796, + 0.05503720045089722, + 0.04444069415330887, + 0.038895171135663986, + 0.06231395900249481, + 0.1253063976764679, + 0.04791465401649475, + 0.03460577130317688, + 0.03924379497766495 + ], + "credit_map": [ + 0.0782257542014122, + 0.043222084641456604, + 0.06194959580898285, + 0.054201044142246246, + 0.029609443619847298, + 0.06007325276732445, + 0.03549030423164368, + 0.0453464575111866, + 0.015569283626973629, + 0.01646711677312851, + 0.015892209485173225, + 0.017391832545399666, + 0.024663351476192474, + 0.021079299971461296, + 0.016117015853524208, + 0.016944246366620064, + 0.007365132216364145, + 0.007210267707705498, + 0.006004089489579201, + 0.008156918920576572, + 0.0077673885971307755, + 0.007664532866328955, + 0.005258707795292139, + 0.005610161926597357, + 0.004419046454131603, + 0.003661596216261387, + 0.007386393845081329, + 0.007608366198837757, + 0.003758253064006567, + 0.006136543117463589, + 0.00601758249104023, + 0.005452913697808981, + 0.00495136808604002, + 0.004579510074108839, + 0.004731777589768171, + 0.0046136220917105675, + 0.004698833916336298, + 0.005376879125833511, + 0.005158923100680113, + 0.004784257151186466, + 0.00796053372323513, + 0.005298038013279438, + 0.008465243503451347, + 0.006284882314503193, + 0.007393199950456619, + 0.009348985739052296, + 0.009885311126708984, + 0.007677762769162655, + 0.018882524222135544, + 0.013071466237306595, + 0.008754675276577473, + 0.014373229816555977, + 0.013357308693230152, + 0.01816627010703087, + 0.021912559866905212, + 0.016788365319371223, + 0.006046521011739969, + 0.005249958019703627, + 0.004034806042909622, + 0.005996311083436012, + 0.005200715735554695, + 0.006841585040092468, + 0.009769299998879433, + 0.004775459878146648, + 0.002158717019483447, + 0.0025332989171147346, + 0.0021005887538194656, + 0.006843203213065863, + 0.0023144721053540707, + 0.00540021201595664, + 0.010386209934949875, + 0.002869072137400508, + 0.0029960651881992817, + 0.0029800347983837128, + 0.0037119395565241575, + 0.012159288860857487, + 0.0020089854951947927, + 0.0061643957160413265, + 0.007210251409560442, + 0.0020128334872424603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.03810613602399826 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072474420070648 + ], + "opposite_logprob_tokens": [ + -4.1410722732543945 + ], + "logprob_reward": -0.016072474420070648, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072474420070648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19137394428253174, + 0.11919455230236053, + 0.10037647932767868, + 0.11324969679117203, + 0.03143764287233353, + 0.036303043365478516, + 0.06688906997442245, + 0.07177779823541641, + 0.2408362179994583, + 0.0285615436732769 + ], + "credit_map": [ + 0.018559524789452553, + 0.018217604607343674, + 0.022945798933506012, + 0.028890978544950485, + 0.034473590552806854, + 0.025632064789533615, + 0.025561144575476646, + 0.017093222588300705, + 0.018835781142115593, + 0.017896587029099464, + 0.010324958711862564, + 0.010752702131867409, + 0.02386205829679966, + 0.0172316562384367, + 0.01100445631891489, + 0.009286352433264256, + 0.015049353241920471, + 0.009340020827949047, + 0.012810925021767616, + 0.013662395067512989, + 0.00970978569239378, + 0.014568657614290714, + 0.014216843992471695, + 0.011018499732017517, + 0.009498458355665207, + 0.009502485394477844, + 0.017245346680283546, + 0.022511359304189682, + 0.013772641308605671, + 0.01578821800649166, + 0.01344961766153574, + 0.011481566354632378, + 0.003303786041215062, + 0.0051030651666224, + 0.0035103661939501762, + 0.0034123726654797792, + 0.0052894726395606995, + 0.004559800028800964, + 0.003753976197913289, + 0.0025048034731298685, + 0.00365994474850595, + 0.0041056289337575436, + 0.0034376883413642645, + 0.004060553852468729, + 0.006133502349257469, + 0.005342385731637478, + 0.005903515964746475, + 0.003659820416942239, + 0.008373741991817951, + 0.007034328766167164, + 0.005319997202605009, + 0.008132697083055973, + 0.011498943902552128, + 0.009517318569123745, + 0.0088651729747653, + 0.008146866224706173, + 0.005079076159745455, + 0.005200486164540052, + 0.0067686475813388824, + 0.01656324416399002, + 0.0066735572181642056, + 0.010688746348023415, + 0.013602335005998611, + 0.007201703265309334, + 0.023867929354310036, + 0.024802275002002716, + 0.013722131960093975, + 0.03539688140153885, + 0.030851952731609344, + 0.034322772175073624, + 0.050938013941049576, + 0.026934266090393066, + 0.002540449844673276, + 0.0022344293538480997, + 0.0023966443259269, + 0.007248526439070702, + 0.0024769078008830547, + 0.004186071455478668, + 0.005589737556874752, + 0.0018887786427512765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.016072474420070648 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546890415251255 + ], + "opposite_logprob_tokens": [ + -4.387547016143799 + ], + "logprob_reward": -0.012546890415251255, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546890415251255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44998109340667725, + 0.16796115040779114, + 0.17993348836898804, + 0.03253530338406563, + 0.030820835381746292, + 0.027422046288847923, + 0.02271401323378086, + 0.032106079161167145, + 0.0261237695813179, + 0.030402226373553276 + ], + "credit_map": [ + 0.06303327530622482, + 0.03747963905334473, + 0.06873893737792969, + 0.06956126540899277, + 0.0491672120988369, + 0.06587132066488266, + 0.04862865060567856, + 0.04750082641839981, + 0.013807994313538074, + 0.028386542573571205, + 0.015010307542979717, + 0.011861560866236687, + 0.04462544620037079, + 0.026246000081300735, + 0.014069495722651482, + 0.01395380962640047, + 0.026833994314074516, + 0.01782497577369213, + 0.023181678727269173, + 0.025555403903126717, + 0.019552011042833328, + 0.026472093537449837, + 0.020620865747332573, + 0.019892489537596703, + 0.00280950590968132, + 0.003989675082266331, + 0.0031433962285518646, + 0.0036536799743771553, + 0.006589030846953392, + 0.004614408127963543, + 0.004565530922263861, + 0.00317008001729846, + 0.0031267059966921806, + 0.003893394721671939, + 0.0030100818257778883, + 0.004454358480870724, + 0.004468149039894342, + 0.004462451674044132, + 0.004346214700490236, + 0.003059481969103217, + 0.0041147684678435326, + 0.0024504056200385094, + 0.0029850490391254425, + 0.004988973494619131, + 0.0035523639526218176, + 0.00412222882732749, + 0.0025840417947620153, + 0.0026242155581712723, + 0.0025816247798502445, + 0.0021917372941970825, + 0.0017092860070988536, + 0.002415054477751255, + 0.0026168557815253735, + 0.0033450473565608263, + 0.004493155982345343, + 0.003361254232004285, + 0.002556808292865753, + 0.002676289062947035, + 0.00296579347923398, + 0.006228665355592966, + 0.0025790082290768623, + 0.004724100232124329, + 0.0075388639234006405, + 0.0028365529142320156, + 0.0027245786041021347, + 0.0026396180037409067, + 0.0020075044594705105, + 0.004950460512191057, + 0.0032842722721397877, + 0.0038892000447958708, + 0.00399903254583478, + 0.0026291057001799345, + 0.0016596878413110971, + 0.002796205459162593, + 0.002680719830095768, + 0.006699323654174805, + 0.002362067811191082, + 0.004320871084928513, + 0.007006962783634663, + 0.0028763890732079744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.012546890415251255 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818285696208477 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.01818285696208477, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818285696208477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4036679267883301, + 0.101430743932724, + 0.08201219886541367, + 0.0686550885438919, + 0.07328744232654572, + 0.058130063116550446, + 0.05169185250997543, + 0.039477214217185974, + 0.07973162084817886, + 0.04191587492823601 + ], + "credit_map": [ + 0.05711458995938301, + 0.03303574025630951, + 0.05623894929885864, + 0.07706921547651291, + 0.03297116979956627, + 0.054011907428503036, + 0.058879394084215164, + 0.03434694930911064, + 0.012555352412164211, + 0.011256932280957699, + 0.01197081245481968, + 0.012680025771260262, + 0.016723377630114555, + 0.014310984872281551, + 0.011303603649139404, + 0.010629648342728615, + 0.010048285126686096, + 0.012073240242898464, + 0.009777838364243507, + 0.010480665601789951, + 0.011975470930337906, + 0.01049829088151455, + 0.008815701119601727, + 0.008342703804373741, + 0.0073891738429665565, + 0.007767245173454285, + 0.005517140030860901, + 0.01102365367114544, + 0.012147982604801655, + 0.009887044318020344, + 0.009395685978233814, + 0.005527162458747625, + 0.01177864521741867, + 0.00854475237429142, + 0.006943879649043083, + 0.007667975500226021, + 0.010950454510748386, + 0.010717451572418213, + 0.011731205508112907, + 0.0049530742689967155, + 0.008376164361834526, + 0.006338032893836498, + 0.005716864950954914, + 0.00647761719301343, + 0.008846891112625599, + 0.008336320519447327, + 0.00855991244316101, + 0.00547825638204813, + 0.007539501413702965, + 0.0056012519635260105, + 0.0045027779415249825, + 0.00419028103351593, + 0.0053294044919312, + 0.007809086702764034, + 0.009654383175075054, + 0.0070651620626449585, + 0.0037979381158947945, + 0.0034260349348187447, + 0.004558926448225975, + 0.004627109970897436, + 0.004898520652204752, + 0.0052892668172717094, + 0.00814504362642765, + 0.004734373185783625, + 0.005507064983248711, + 0.0069272806867957115, + 0.005399641580879688, + 0.01348001416772604, + 0.006956086494028568, + 0.011648849584162235, + 0.021237362176179886, + 0.008575311861932278, + 0.0028248284943401814, + 0.004093147814273834, + 0.003749492345377803, + 0.010708707384765148, + 0.004419045988470316, + 0.0057057165540754795, + 0.005719390232115984, + 0.004695541225373745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01818285696208477 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static close-up captures a hyper-realistic CG rendering of a bright yellow handheld sprayer, carelessly tossed onto a weathered stone patio. Faint scuffs and a touch of dirt mark the sprayer's smooth plastic exterior, evidence of past use. A thick, viscous, vibrant green liquid gushes forth in a steady stream from a black nozzle, not splashing but clinging tenaciously to the rough, uneven surface of the gray and beige stone slabs. This unnatural fluid fizzes and bubbles with vibrant energy, its surface constantly fractured as tiny, shimmering spheres form, expand, and burst, releasing wisps of light vapor. Bathed in the warm golden glow of the afternoon sun, the scene casts long, soft shadows, capturing the translucent quality of the bubbling liquid, making it appear to radiate an eerie internal energy across the cool stone, creating a deeply unsettling yet visually captivating atmosphere.", + "seed": 22731, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_7/index_5665_seed_22731_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.013274103961884975, + "vlm_logprob": -0.013274103961884975, + "vlm_yes_no_margin": 4.4499999523162845, + "vlm_token_credit": -0.013274103961884975, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006843225099146366 + ], + "opposite_logprob_tokens": [ + -5.381843090057373 + ], + "logprob_reward": -0.006843225099146366, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.006843225099146366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25885504484176636, + 0.10896354913711548, + 0.062250375747680664, + 0.04878760129213333, + 0.054442428052425385, + 0.1230306476354599, + 0.15363843739032745, + 0.09386871010065079, + 0.04058452323079109, + 0.05557875335216522 + ], + "credit_map": [ + 0.02574462629854679, + 0.015504368580877781, + 0.04943173751235008, + 0.046881575137376785, + 0.018716961145401, + 0.037349116057157516, + 0.022654522210359573, + 0.042572133243083954, + 0.01103364396840334, + 0.005840403959155083, + 0.012786322273314, + 0.007140427362173796, + 0.01259097084403038, + 0.01814485900104046, + 0.010387456975877285, + 0.03103945590555668, + 0.007710916921496391, + 0.005945858079940081, + 0.007278798148036003, + 0.007044358644634485, + 0.007961727678775787, + 0.009271805174648762, + 0.00620160810649395, + 0.010835304856300354, + 0.005676805507391691, + 0.004433091264218092, + 0.004905755165964365, + 0.005998526234179735, + 0.005800472106784582, + 0.007218174636363983, + 0.005816133692860603, + 0.008938641287386417, + 0.006412390619516373, + 0.0066161141730844975, + 0.0067298198118805885, + 0.004834929946810007, + 0.006463589146733284, + 0.007417816203087568, + 0.006864944938570261, + 0.009102828800678253, + 0.010864844545722008, + 0.01336706057190895, + 0.013013849034905434, + 0.013024535030126572, + 0.014666730538010597, + 0.016227364540100098, + 0.02772270143032074, + 0.014143559150397778, + 0.013583872467279434, + 0.01795482635498047, + 0.01672310382127762, + 0.012633607722818851, + 0.021596504375338554, + 0.02119910903275013, + 0.029505303129553795, + 0.020442111417651176, + 0.008516140282154083, + 0.007155591621994972, + 0.015412772074341774, + 0.009399102069437504, + 0.012295897118747234, + 0.013567360118031502, + 0.010815669782459736, + 0.016706177964806557, + 0.0053906976245343685, + 0.0033626200165599585, + 0.006516969762742519, + 0.0034143293742090464, + 0.004628434311598539, + 0.005576701369136572, + 0.005576234310865402, + 0.006118535529822111, + 0.0069823903031647205, + 0.003294045338407159, + 0.01750563457608223, + 0.005712202284485102, + 0.004688661079853773, + 0.007537226192653179, + 0.00438108341768384, + 0.005477510392665863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.006843225099146366 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081579141318798 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081579141318798, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081579141318798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1920718252658844, + 0.1489347368478775, + 0.18763522803783417, + 0.05307205393910408, + 0.01659036986529827, + 0.06407687813043594, + 0.13517743349075317, + 0.08937735110521317, + 0.07625993341207504, + 0.03680422902107239 + ], + "credit_map": [ + 0.027206210419535637, + 0.00869892816990614, + 0.048678915947675705, + 0.03277435153722763, + 0.008760899305343628, + 0.024277808144688606, + 0.015035194344818592, + 0.02663951739668846, + 0.014243525452911854, + 0.013364133425056934, + 0.036082036793231964, + 0.01685870625078678, + 0.01533811166882515, + 0.021797265857458115, + 0.014460175298154354, + 0.01679077371954918, + 0.029783666133880615, + 0.014711172319948673, + 0.03508562222123146, + 0.026054970920085907, + 0.017245542258024216, + 0.02643011324107647, + 0.016485771164298058, + 0.021838374435901642, + 0.00861817505210638, + 0.00391888152807951, + 0.006006540730595589, + 0.004575655329972506, + 0.005901557859033346, + 0.008208145387470722, + 0.004992935340851545, + 0.010850165039300919, + 0.0023890447337180376, + 0.0019048693357035518, + 0.0017864005640149117, + 0.0017028875881806016, + 0.002007971052080393, + 0.0021873963996767998, + 0.002109591383486986, + 0.002502207877114415, + 0.008373457938432693, + 0.003939399961382151, + 0.012705220840871334, + 0.01113937608897686, + 0.004640635568648577, + 0.007779113948345184, + 0.009608571417629719, + 0.005891099106520414, + 0.017435507848858833, + 0.012700499035418034, + 0.02762731723487377, + 0.01146150752902031, + 0.01375976949930191, + 0.016246158629655838, + 0.01999879628419876, + 0.015947870910167694, + 0.015591763891279697, + 0.006346158683300018, + 0.018986912444233894, + 0.015570429153740406, + 0.005750525742769241, + 0.010972809977829456, + 0.007523516193032265, + 0.008635233156383038, + 0.013594391755759716, + 0.0049474285915493965, + 0.019734226167201996, + 0.01062734518200159, + 0.004160911776125431, + 0.01107469480484724, + 0.0054708197712898254, + 0.006650114431977272, + 0.0055748941376805305, + 0.0034289814066141844, + 0.0060554975643754005, + 0.00488703791052103, + 0.004106127191334963, + 0.005180858075618744, + 0.0034258540254086256, + 0.004144976381212473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.011081579141318798 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067663207650185 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.016067663207650185, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067663207650185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15299998223781586, + 0.10082590579986572, + 0.06314148753881454, + 0.06417415291070938, + 0.18015512824058533, + 0.11306048929691315, + 0.08089369535446167, + 0.09811398386955261, + 0.06383652240037918, + 0.08279869705438614 + ], + "credit_map": [ + 0.01993100345134735, + 0.008514605462551117, + 0.026713021099567413, + 0.03202997148036957, + 0.008863227441906929, + 0.020957596600055695, + 0.01679103635251522, + 0.019199514761567116, + 0.013227702118456364, + 0.009465498849749565, + 0.021157575771212578, + 0.01010893378406763, + 0.009322956204414368, + 0.015044113621115685, + 0.00999094545841217, + 0.012508179992437363, + 0.007672916632145643, + 0.007140268571674824, + 0.006632765289396048, + 0.006115742027759552, + 0.007877163589000702, + 0.008995051495730877, + 0.008685854263603687, + 0.01002172939479351, + 0.006492826156318188, + 0.003966853488236666, + 0.0073057799600064754, + 0.005777852144092321, + 0.006006730254739523, + 0.010178657248616219, + 0.006054839119315147, + 0.018390610814094543, + 0.02422342263162136, + 0.02772464044392109, + 0.015746360644698143, + 0.013358949683606625, + 0.022886233404278755, + 0.026860935613512993, + 0.021387197077274323, + 0.027967384085059166, + 0.013127131387591362, + 0.007798313163220882, + 0.023082895204424858, + 0.015929698944091797, + 0.008675700053572655, + 0.014981936663389206, + 0.01787298172712326, + 0.011591830290853977, + 0.008768264204263687, + 0.008976374752819538, + 0.01290079578757286, + 0.009987539611756802, + 0.009018008597195148, + 0.010645944625139236, + 0.010785977356135845, + 0.009810788556933403, + 0.010501515120267868, + 0.008862334303557873, + 0.018873654305934906, + 0.01530345156788826, + 0.010170357301831245, + 0.01232595182955265, + 0.012107105925679207, + 0.009969620034098625, + 0.009172864258289337, + 0.003774433396756649, + 0.018460646271705627, + 0.007845885120332241, + 0.004775331821292639, + 0.009463449940085411, + 0.005390349309891462, + 0.004953561816364527, + 0.011508123949170113, + 0.004900272469967604, + 0.027523396536707878, + 0.008232186548411846, + 0.005309989210218191, + 0.011592967435717583, + 0.0059521254152059555, + 0.007779634092003107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.016067663207650185 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419894304126501 + ], + "opposite_logprob_tokens": [ + -4.2641987800598145 + ], + "logprob_reward": -0.01419894304126501, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419894304126501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2094053477048874, + 0.16021771728992462, + 0.05203964188694954, + 0.032665595412254333, + 0.023201551288366318, + 0.04639090225100517, + 0.14932385087013245, + 0.11288366466760635, + 0.05975939705967903, + 0.1541122943162918 + ], + "credit_map": [ + 0.02149970829486847, + 0.015280019491910934, + 0.03709357976913452, + 0.04467536509037018, + 0.012554510496556759, + 0.028473271057009697, + 0.02484910562634468, + 0.024979809299111366, + 0.01993033103644848, + 0.008823584765195847, + 0.05056676268577576, + 0.017955295741558075, + 0.009418563917279243, + 0.024473227560520172, + 0.010470387525856495, + 0.018579581752419472, + 0.006921766791492701, + 0.005300973076373339, + 0.005214665085077286, + 0.005939135327935219, + 0.005922633223235607, + 0.007364093326032162, + 0.007564732804894447, + 0.0078116427175700665, + 0.0029806960374116898, + 0.003233497031033039, + 0.0041842153295874596, + 0.003374205669388175, + 0.004431516397744417, + 0.004851801786571741, + 0.004961253609508276, + 0.004648411646485329, + 0.002675669500604272, + 0.0018338499357923865, + 0.0035511532332748175, + 0.0034047244116663933, + 0.0026841622311621904, + 0.003113345243036747, + 0.0028735182713717222, + 0.0030651320703327656, + 0.005875356961041689, + 0.003171172458678484, + 0.010058141313493252, + 0.007307803258299828, + 0.0039568087086081505, + 0.005932536907494068, + 0.005538820754736662, + 0.0045502642169594765, + 0.015627216547727585, + 0.014255661517381668, + 0.030438967049121857, + 0.015400856733322144, + 0.015462042763829231, + 0.018683766946196556, + 0.021827733144164085, + 0.01762760989367962, + 0.013080921024084091, + 0.011330028995871544, + 0.018253345042467117, + 0.021323902532458305, + 0.01040708925575018, + 0.013411168940365314, + 0.01415185071527958, + 0.010925372131168842, + 0.008125610649585724, + 0.004054494202136993, + 0.01682090200483799, + 0.006078262347728014, + 0.004792863968759775, + 0.008564420975744724, + 0.006004127208143473, + 0.0053187208250164986, + 0.023739004507660866, + 0.011104976758360863, + 0.048050060868263245, + 0.014861240983009338, + 0.009188410826027393, + 0.02087019570171833, + 0.01206220407038927, + 0.014236222952604294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.01419894304126501 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018179109320044518 + ], + "opposite_logprob_tokens": [ + -4.018178939819336 + ], + "logprob_reward": -0.018179109320044518, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018179109320044518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23548655211925507, + 0.05230903998017311, + 0.056984152644872665, + 0.041663236916065216, + 0.05203551799058914, + 0.0646628737449646, + 0.14305056631565094, + 0.23710820078849792, + 0.040386561304330826, + 0.07631324976682663 + ], + "credit_map": [ + 0.03514285013079643, + 0.01788610778748989, + 0.04784271866083145, + 0.043776702135801315, + 0.0184557493776083, + 0.03397210314869881, + 0.01656498573720455, + 0.021845348179340363, + 0.00799589417874813, + 0.005065837409347296, + 0.009750514291226864, + 0.0065191020257771015, + 0.004716000985354185, + 0.006886881776154041, + 0.004446332808583975, + 0.006928481627255678, + 0.007074601948261261, + 0.006583866663277149, + 0.007318099495023489, + 0.005429744720458984, + 0.007368745282292366, + 0.007674296386539936, + 0.007204983849078417, + 0.008329814299941063, + 0.00555919436737895, + 0.002506343415006995, + 0.00672651594504714, + 0.005979642271995544, + 0.005173260346055031, + 0.0056314063258469105, + 0.004961758852005005, + 0.005125121679157019, + 0.007675024680793285, + 0.0066083152778446674, + 0.006448684725910425, + 0.004116611555218697, + 0.005623803939670324, + 0.007636625319719315, + 0.00471334857866168, + 0.009213104844093323, + 0.006342662964016199, + 0.005453880410641432, + 0.009630158543586731, + 0.007164857815951109, + 0.008153281174600124, + 0.008460704237222672, + 0.011249478906393051, + 0.008207854814827442, + 0.015323527157306671, + 0.013052848167717457, + 0.03155703470110893, + 0.011814785189926624, + 0.01605992764234543, + 0.018464410677552223, + 0.01974191516637802, + 0.017036130651831627, + 0.031496357172727585, + 0.013282428495585918, + 0.07041686773300171, + 0.034581370651721954, + 0.017655322328209877, + 0.03270426765084267, + 0.019771451130509377, + 0.017200149595737457, + 0.004267459269613028, + 0.004437684081494808, + 0.005018854513764381, + 0.005107259843498468, + 0.0044456650502979755, + 0.005364591721445322, + 0.004973876290023327, + 0.0067711784504354, + 0.006567712873220444, + 0.005857564974576235, + 0.008130260743200779, + 0.0059954384341835976, + 0.013020041398704052, + 0.012490499764680862, + 0.01543431542813778, + 0.008817424066364765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.018179109320044518 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static close-up captures a hyper-realistic CG rendering of a bright yellow handheld sprayer, carelessly tossed onto a weathered stone patio. Faint scuffs and a touch of dirt mark the sprayer's smooth plastic exterior, evidence of past use. A thick, viscous, vibrant green liquid gushes forth in a steady stream from a black nozzle, not splashing but clinging tenaciously to the rough, uneven surface of the gray and beige stone slabs. This unnatural fluid fizzes and bubbles with vibrant energy, its surface constantly fractured as tiny, shimmering spheres form, expand, and burst, releasing wisps of light vapor. Bathed in the warm golden glow of the afternoon sun, the scene casts long, soft shadows, capturing the translucent quality of the bubbling liquid, making it appear to radiate an eerie internal energy across the cool stone, creating a deeply unsettling yet visually captivating atmosphere.", + "seed": 22732, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_7/index_5665_seed_22732_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.034985079802572726, + "vlm_logprob": -0.034985079802572726, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.034985079802572726, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010063623078167439 + ], + "opposite_logprob_tokens": [ + -5.135063648223877 + ], + "logprob_reward": -0.010063623078167439, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.010063623078167439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2406131625175476, + 0.10608617961406708, + 0.03797459602355957, + 0.09722671657800674, + 0.12491153180599213, + 0.031492650508880615, + 0.08665157854557037, + 0.055374667048454285, + 0.05288873612880707, + 0.16678014397621155 + ], + "credit_map": [ + 0.025082828477025032, + 0.02341197058558464, + 0.038402315229177475, + 0.03623439371585846, + 0.026393605396151543, + 0.03142477199435234, + 0.033179979771375656, + 0.02648329548537731, + 0.0095828827470541, + 0.012373214587569237, + 0.017513493075966835, + 0.01911015622317791, + 0.009020955301821232, + 0.009998081251978874, + 0.01729636825621128, + 0.011191029101610184, + 0.003200200153514743, + 0.004662114195525646, + 0.005369867663830519, + 0.004320759326219559, + 0.005506434477865696, + 0.0043559898622334, + 0.0061778416857123375, + 0.004381389357149601, + 0.0075202826410532, + 0.006411749869585037, + 0.009730061516165733, + 0.009955995716154575, + 0.01110018603503704, + 0.02342691645026207, + 0.012988104484975338, + 0.01609342359006405, + 0.013166116550564766, + 0.011900669895112514, + 0.01816859468817711, + 0.011614621616899967, + 0.01952216774225235, + 0.024533161893486977, + 0.010676701553165913, + 0.015329493209719658, + 0.0031464386265724897, + 0.0031986949034035206, + 0.0035241798032075167, + 0.002923166612163186, + 0.0043169911950826645, + 0.0038713933899998665, + 0.005316025111824274, + 0.005195763893425465, + 0.007885965518653393, + 0.011294292286038399, + 0.011169970035552979, + 0.010517862625420094, + 0.015129808336496353, + 0.014292171224951744, + 0.009550251998007298, + 0.006811257917433977, + 0.0048840902745723724, + 0.004151980392634869, + 0.007082832977175713, + 0.00826472882181406, + 0.00787506252527237, + 0.006182444281876087, + 0.00712915975600481, + 0.009804368019104004, + 0.00587099976837635, + 0.004029127303510904, + 0.012889697216451168, + 0.008321257308125496, + 0.004914058372378349, + 0.0061015780083835125, + 0.006119840312749147, + 0.0046421741135418415, + 0.00984097458422184, + 0.0073890723288059235, + 0.01715533807873726, + 0.02080174908041954, + 0.011991468258202076, + 0.01635320484638214, + 0.055500637739896774, + 0.027747690677642822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.010063623078167439 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193509429693222 + ], + "opposite_logprob_tokens": [ + -4.01819372177124 + ], + "logprob_reward": -0.018193509429693222, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193509429693222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.298561692237854, + 0.024770313873887062, + 0.008906825445592403, + 0.044769614934921265, + 0.4555068910121918, + 0.016796348616480827, + 0.033332619816064835, + 0.02521473541855812, + 0.02224685810506344, + 0.06989404559135437 + ], + "credit_map": [ + 0.0398104153573513, + 0.018203958868980408, + 0.08142029494047165, + 0.02454056218266487, + 0.020241841673851013, + 0.038596004247665405, + 0.029412947595119476, + 0.046335719525814056, + 0.002219788497313857, + 0.0023185210302472115, + 0.0036308057606220245, + 0.0033641094341874123, + 0.0025306970346719027, + 0.003414338920265436, + 0.003955848049372435, + 0.00333620747551322, + 0.0007565601263195276, + 0.0011402831878513098, + 0.0012910207733511925, + 0.0011251837713643909, + 0.0010040979832410812, + 0.0010855626314878464, + 0.0014888810692355037, + 0.0010152366012334824, + 0.0056196036748588085, + 0.002257017185911536, + 0.006406387779861689, + 0.0028320953715592623, + 0.004789054859429598, + 0.010335429571568966, + 0.003674076870083809, + 0.008855953812599182, + 0.039235346019268036, + 0.027820289134979248, + 0.04738539457321167, + 0.02420378103852272, + 0.07668579369783401, + 0.11175265908241272, + 0.04139383137226105, + 0.08702979981899261, + 0.0019398542353883386, + 0.0014969148905947804, + 0.0016729655908420682, + 0.001370299723930657, + 0.002976839430630207, + 0.0031028673984110355, + 0.0021860876586288214, + 0.0020505208522081375, + 0.0030945874750614166, + 0.004562986083328724, + 0.0030557652935385704, + 0.004750518128275871, + 0.005550043657422066, + 0.004514205735176802, + 0.004113968927413225, + 0.003690549870952964, + 0.0022914567962288857, + 0.0019346883054822683, + 0.0029779120814055204, + 0.004884281195700169, + 0.0020359924528747797, + 0.002814890816807747, + 0.0036618730518966913, + 0.004613643046468496, + 0.00200070277787745, + 0.0015230508288368583, + 0.0043481625616550446, + 0.0037203305400907993, + 0.002169253770262003, + 0.003198865568265319, + 0.0032063075341284275, + 0.0020801860373467207, + 0.004922776482999325, + 0.004839874804019928, + 0.00969801377505064, + 0.009145546704530716, + 0.004249922465533018, + 0.00659073144197464, + 0.020693400874733925, + 0.009753784164786339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.018193509429693222 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048655007034540176 + ], + "opposite_logprob_tokens": [ + -3.0486550331115723 + ], + "logprob_reward": -0.048655007034540176, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048655007034540176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20448249578475952, + 0.07028820365667343, + 0.07083632797002792, + 0.137993723154068, + 0.19466516375541687, + 0.032045941799879074, + 0.136556476354599, + 0.037781137973070145, + 0.053492020815610886, + 0.06185849383473396 + ], + "credit_map": [ + 0.022919336333870888, + 0.023561500012874603, + 0.022030794993042946, + 0.028602901846170425, + 0.016523223370313644, + 0.02628011256456375, + 0.042174696922302246, + 0.022389942780137062, + 0.005915485788136721, + 0.0072519611567258835, + 0.013289710506796837, + 0.01676071062684059, + 0.00632968544960022, + 0.007478434592485428, + 0.007470881100744009, + 0.005791342817246914, + 0.012774860486388206, + 0.008771928958594799, + 0.020931009203195572, + 0.008388141170144081, + 0.004314846359193325, + 0.005018726922571659, + 0.005830781534314156, + 0.00480604125186801, + 0.02028588019311428, + 0.01284714974462986, + 0.016771433874964714, + 0.012232419103384018, + 0.01572931744158268, + 0.019461117684841156, + 0.02589024417102337, + 0.014776168391108513, + 0.01821434684097767, + 0.011587420478463173, + 0.02316085621714592, + 0.011702065356075764, + 0.0321323536336422, + 0.043828077614307404, + 0.018145883455872536, + 0.03589416295289993, + 0.0037161235231906176, + 0.003068318124860525, + 0.0038745494093745947, + 0.0029300283640623093, + 0.004748707637190819, + 0.004460261203348637, + 0.004598502069711685, + 0.004649453330785036, + 0.013017592020332813, + 0.010703062638640404, + 0.02146165445446968, + 0.018041854724287987, + 0.017924010753631592, + 0.015459561720490456, + 0.016679415479302406, + 0.023269323632121086, + 0.0028253020718693733, + 0.0032501721289008856, + 0.0052189393900334835, + 0.006403316278010607, + 0.0036340877413749695, + 0.003135708859190345, + 0.00661926856264472, + 0.006694343872368336, + 0.004583525937050581, + 0.004414794966578484, + 0.007596734911203384, + 0.008085922338068485, + 0.007401118520647287, + 0.009565312415361404, + 0.006324992515146732, + 0.005519624799489975, + 0.005646119359880686, + 0.005433434154838324, + 0.008253322914242744, + 0.009872481226921082, + 0.004971017129719257, + 0.01168899703770876, + 0.009700434282422066, + 0.006292693316936493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.048655007034540176 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043069567531347275 + ], + "opposite_logprob_tokens": [ + -3.16806960105896 + ], + "logprob_reward": -0.043069567531347275, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043069567531347275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3327162265777588, + 0.05326388031244278, + 0.08974508941173553, + 0.09162639081478119, + 0.10239002108573914, + 0.031982459127902985, + 0.11435050517320633, + 0.057874489575624466, + 0.043504782021045685, + 0.08254613727331161 + ], + "credit_map": [ + 0.048626165837049484, + 0.017817825078964233, + 0.09835818409919739, + 0.026482371613383293, + 0.018837859854102135, + 0.031595438718795776, + 0.0389038547873497, + 0.05209454521536827, + 0.005785917863249779, + 0.005517142359167337, + 0.008552010171115398, + 0.006537941750138998, + 0.00465197442099452, + 0.005505382549017668, + 0.00997452437877655, + 0.006738986819982529, + 0.016410037875175476, + 0.008043212816119194, + 0.02824568562209606, + 0.007029128726571798, + 0.005585399456322193, + 0.006813602987676859, + 0.0092807961627841, + 0.008337218314409256, + 0.012601868249475956, + 0.011145606637001038, + 0.009209664538502693, + 0.010611005127429962, + 0.012883185409009457, + 0.016514388844370842, + 0.011189984157681465, + 0.007470688316971064, + 0.009539066813886166, + 0.006161033641546965, + 0.011122805066406727, + 0.0053450544364750385, + 0.01791396550834179, + 0.023787323385477066, + 0.010197916999459267, + 0.018322855234146118, + 0.003042506752535701, + 0.002258929656818509, + 0.003076522145420313, + 0.0034803326707333326, + 0.004661524202674627, + 0.005241809878498316, + 0.005788607522845268, + 0.004432228393852711, + 0.009138884954154491, + 0.006939623970538378, + 0.01675916649401188, + 0.01593799516558647, + 0.016013862565159798, + 0.009328637272119522, + 0.01805756241083145, + 0.0221747774630785, + 0.005482751410454512, + 0.0040031298995018005, + 0.0078460443764925, + 0.011080359108746052, + 0.006120364647358656, + 0.005127113312482834, + 0.00985987950116396, + 0.008354850113391876, + 0.004319342784583569, + 0.003535522148013115, + 0.006430990528315306, + 0.006753688678145409, + 0.006462736986577511, + 0.006613942328840494, + 0.005185907706618309, + 0.0042026531882584095, + 0.005017921328544617, + 0.004856784362345934, + 0.010123812593519688, + 0.012815472669899464, + 0.007066825404763222, + 0.01697075366973877, + 0.013293945230543613, + 0.012400620616972446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.043069567531347275 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054943691939115524 + ], + "opposite_logprob_tokens": [ + -2.929943799972534 + ], + "logprob_reward": -0.054943691939115524, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054943691939115524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11907821893692017, + 0.07927883416414261, + 0.057244278490543365, + 0.14681150019168854, + 0.042413465678691864, + 0.06584198772907257, + 0.20754234492778778, + 0.027611853554844856, + 0.0526086762547493, + 0.2015688270330429 + ], + "credit_map": [ + 0.008626727387309074, + 0.008993932045996189, + 0.013966197147965431, + 0.015020915307104588, + 0.00981178879737854, + 0.022617952898144722, + 0.023697981610894203, + 0.016342725604772568, + 0.008648212067782879, + 0.0073105222545564175, + 0.011025245301425457, + 0.01564076356589794, + 0.0052324943244457245, + 0.00903247855603695, + 0.012364253401756287, + 0.010024859569966793, + 0.01051732525229454, + 0.0077105918899178505, + 0.010658093728125095, + 0.006654008291661739, + 0.004352868534624577, + 0.004932950250804424, + 0.005238172132521868, + 0.007180265616625547, + 0.01704203523695469, + 0.01456436887383461, + 0.01779591292142868, + 0.012972782365977764, + 0.01583000086247921, + 0.0230645090341568, + 0.02203056402504444, + 0.023511333391070366, + 0.003010314889252186, + 0.003380203852429986, + 0.003992112819105387, + 0.0035210642963647842, + 0.006739208009094, + 0.010279466398060322, + 0.0038524975534528494, + 0.007638596463948488, + 0.00961936917155981, + 0.006050515919923782, + 0.008995636366307735, + 0.005686403717845678, + 0.008041351102292538, + 0.006163330748677254, + 0.009322045370936394, + 0.011963333003222942, + 0.021074797958135605, + 0.030252836644649506, + 0.024793118238449097, + 0.02409827895462513, + 0.03646938130259514, + 0.031138701364398003, + 0.023388849571347237, + 0.016326379030942917, + 0.0026842711959034204, + 0.002975833835080266, + 0.0035527257714420557, + 0.003996916115283966, + 0.003499418729916215, + 0.0030876710079610348, + 0.004202038515359163, + 0.0036129802465438843, + 0.008031411096453667, + 0.005507612135261297, + 0.008442659862339497, + 0.007200146093964577, + 0.0053678229451179504, + 0.006624950096011162, + 0.006415080279111862, + 0.005018992815166712, + 0.02882089465856552, + 0.0339161679148674, + 0.02787960134446621, + 0.028402667492628098, + 0.017207946628332138, + 0.014445741660892963, + 0.030789632350206375, + 0.020106179639697075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 3 + ], + "selected_reward": -0.054943691939115524 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static close-up captures a hyper-realistic CG rendering of a bright yellow handheld sprayer, carelessly tossed onto a weathered stone patio. Faint scuffs and a touch of dirt mark the sprayer's smooth plastic exterior, evidence of past use. A thick, viscous, vibrant green liquid gushes forth in a steady stream from a black nozzle, not splashing but clinging tenaciously to the rough, uneven surface of the gray and beige stone slabs. This unnatural fluid fizzes and bubbles with vibrant energy, its surface constantly fractured as tiny, shimmering spheres form, expand, and burst, releasing wisps of light vapor. Bathed in the warm golden glow of the afternoon sun, the scene casts long, soft shadows, capturing the translucent quality of the bubbling liquid, making it appear to radiate an eerie internal energy across the cool stone, creating a deeply unsettling yet visually captivating atmosphere.", + "seed": 22733, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_7/index_5665_seed_22733_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03667556643486023, + "vlm_logprob": -0.03667556643486023, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.03667556643486023, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02999836392700672 + ], + "opposite_logprob_tokens": [ + -3.654998302459717 + ], + "logprob_reward": -0.02999836392700672, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02999836392700672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19285805523395538, + 0.09023882448673248, + 0.21261821687221527, + 0.03242126479744911, + 0.29633429646492004, + 0.02626202069222927, + 0.061311230063438416, + 0.01680154725909233, + 0.04358551278710365, + 0.02756902389228344 + ], + "credit_map": [ + 0.028524091467261314, + 0.015263994224369526, + 0.07821493595838547, + 0.012358556501567364, + 0.008189676329493523, + 0.010426432825624943, + 0.02362571284174919, + 0.01625465787947178, + 0.019963694736361504, + 0.006250896491110325, + 0.032170847058296204, + 0.005757825914770365, + 0.0038852912839502096, + 0.006809690501540899, + 0.008464760147035122, + 0.006935807876288891, + 0.03335518762469292, + 0.07179319113492966, + 0.010377847589552402, + 0.024930143728852272, + 0.011203373782336712, + 0.013476058840751648, + 0.012090080417692661, + 0.035392336547374725, + 0.004772300831973553, + 0.0035070416051894426, + 0.003618067828938365, + 0.0024673673324286938, + 0.003941326402127743, + 0.0049347649328410625, + 0.002787704113870859, + 0.006392693612724543, + 0.017711970955133438, + 0.014772639609873295, + 0.007065786514431238, + 0.061650700867176056, + 0.013212652876973152, + 0.010794080793857574, + 0.12454032897949219, + 0.04658612236380577, + 0.002813116880133748, + 0.0034609190188348293, + 0.002207498298957944, + 0.004473363980650902, + 0.004788307938724756, + 0.0033278614282608032, + 0.0030308240093290806, + 0.002160129602998495, + 0.006263103801757097, + 0.008019795641303062, + 0.008453505113720894, + 0.009779606945812702, + 0.005922343116253614, + 0.006578611675649881, + 0.009548407047986984, + 0.006745853461325169, + 0.0014525895239785314, + 0.0021683413069695234, + 0.002459227340295911, + 0.002394570969045162, + 0.0018141146283596754, + 0.0018771766917780042, + 0.002707373583689332, + 0.001928152283653617, + 0.0036652283743023872, + 0.005915686953812838, + 0.004089148249477148, + 0.013583426363766193, + 0.00487402593716979, + 0.0031625917181372643, + 0.0061068288050591946, + 0.0021885752212256193, + 0.002334657358005643, + 0.0039335014298558235, + 0.0034601876977831125, + 0.004416193813085556, + 0.002724930178374052, + 0.003197553800418973, + 0.0046283891424536705, + 0.002873610705137253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.02999836392700672 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003437727689743 + ], + "opposite_logprob_tokens": [ + -2.6950342655181885 + ], + "logprob_reward": -0.07003437727689743, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003437727689743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2973032295703888, + 0.018973704427480698, + 0.058194734156131744, + 0.11043748259544373, + 0.30180925130844116, + 0.0688573494553566, + 0.026401283219456673, + 0.03460805118083954, + 0.037357501685619354, + 0.046057358384132385 + ], + "credit_map": [ + 0.04408351331949234, + 0.016073815524578094, + 0.10925060510635376, + 0.022748686373233795, + 0.010903219692409039, + 0.010838893242180347, + 0.06851695477962494, + 0.014887569472193718, + 0.00202700262889266, + 0.002782612107694149, + 0.001401921734213829, + 0.0030191324185580015, + 0.0019696184899657965, + 0.0019585785921663046, + 0.0027013756334781647, + 0.0031134665478020906, + 0.008921032771468163, + 0.019549360498785973, + 0.0022394179832190275, + 0.006896347273141146, + 0.003007040359079838, + 0.004385682754218578, + 0.0036890488117933273, + 0.00950680486857891, + 0.02640705741941929, + 0.00952148623764515, + 0.008788538165390491, + 0.008454570546746254, + 0.0038427652325481176, + 0.008309585973620415, + 0.01620480976998806, + 0.02890867367386818, + 0.019787104800343513, + 0.016144219785928726, + 0.007718899752944708, + 0.07238534837961197, + 0.01330501027405262, + 0.010904936119914055, + 0.12721854448318481, + 0.034345224499702454, + 0.00683730049058795, + 0.0065633757039904594, + 0.007431492675095797, + 0.011299611069262028, + 0.009150940924882889, + 0.011722961440682411, + 0.00843566469848156, + 0.007416006177663803, + 0.0028510510455816984, + 0.003292220877483487, + 0.004544573836028576, + 0.0035752675030380487, + 0.003207030240446329, + 0.0036665694788098335, + 0.0021829172037541866, + 0.0030816562939435244, + 0.0034914007410407066, + 0.004563230089843273, + 0.006336935330182314, + 0.003935744520276785, + 0.0038576265797019005, + 0.0033570656087249517, + 0.005652216728776693, + 0.0034138336777687073, + 0.003453872399404645, + 0.003341656643897295, + 0.0037845256738364697, + 0.004818610846996307, + 0.0035323449410498142, + 0.009941859170794487, + 0.004068481270223856, + 0.004416154231876135, + 0.003266270738095045, + 0.005663314834237099, + 0.005109194200485945, + 0.007226669695228338, + 0.005301488097757101, + 0.00650689797475934, + 0.008023804053664207, + 0.004959718324244022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.07003437727689743 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635037526488304 + ], + "opposite_logprob_tokens": [ + -3.651350259780884 + ], + "logprob_reward": -0.02635037526488304, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635037526488304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11715583503246307, + 0.01370406337082386, + 0.06267062574625015, + 0.34582197666168213, + 0.2272186428308487, + 0.06197734549641609, + 0.06983793526887894, + 0.02048048935830593, + 0.04946206137537956, + 0.031671032309532166 + ], + "credit_map": [ + 0.021206369623541832, + 0.00668683834373951, + 0.03203347325325012, + 0.018893921747803688, + 0.005939879920333624, + 0.00842119101434946, + 0.015049358829855919, + 0.008924806490540504, + 0.0012550096726045012, + 0.001300307223573327, + 0.002151111140847206, + 0.0018167265225201845, + 0.0011844944674521685, + 0.0015881413128226995, + 0.002081260783597827, + 0.0023270114324986935, + 0.011688840575516224, + 0.02400851622223854, + 0.002613542601466179, + 0.007075492758303881, + 0.0031514589209109545, + 0.0042769149877130985, + 0.002782111056149006, + 0.007073746994137764, + 0.06851258873939514, + 0.029783960431814194, + 0.02167513221502304, + 0.04783070832490921, + 0.014406668953597546, + 0.02667178399860859, + 0.05557049438357353, + 0.08137063682079315, + 0.015753332525491714, + 0.012412031181156635, + 0.007197462022304535, + 0.054525818675756454, + 0.010370276868343353, + 0.011355222202837467, + 0.0930991992354393, + 0.022505292668938637, + 0.00506439246237278, + 0.005946315824985504, + 0.006946878507733345, + 0.007778035011142492, + 0.008672070689499378, + 0.014207832515239716, + 0.005950636230409145, + 0.007411185186356306, + 0.007263397332280874, + 0.00913811195641756, + 0.013481923379004002, + 0.009164665825664997, + 0.009139024652540684, + 0.007907574996352196, + 0.007092148531228304, + 0.006651098839938641, + 0.0019897203892469406, + 0.002603459171950817, + 0.003142643254250288, + 0.003189207287505269, + 0.0019268859177827835, + 0.0026099635288119316, + 0.002657541073858738, + 0.002361070131883025, + 0.00414759898558259, + 0.006462938152253628, + 0.004375404678285122, + 0.0141591252759099, + 0.0057131764478981495, + 0.005478000268340111, + 0.006269667763262987, + 0.002856147475540638, + 0.0030090999789536, + 0.004154268652200699, + 0.004339744336903095, + 0.0036564047913998365, + 0.005698504392057657, + 0.0031334327068179846, + 0.004170405678451061, + 0.0035091694444417953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.02635037526488304 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023300303146243095 + ], + "opposite_logprob_tokens": [ + -3.7733004093170166 + ], + "logprob_reward": -0.023300303146243095, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023300303146243095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14504975080490112, + 0.058882858604192734, + 0.05588879808783531, + 0.1507491171360016, + 0.16263055801391602, + 0.04018295556306839, + 0.10611778497695923, + 0.0647231861948967, + 0.07655184715986252, + 0.13922317326068878 + ], + "credit_map": [ + 0.012453378178179264, + 0.009117540903389454, + 0.007797683589160442, + 0.03514588996767998, + 0.010560371913015842, + 0.009768227115273476, + 0.04929988086223602, + 0.010906766168773174, + 0.013024909421801567, + 0.0034273271448910236, + 0.023143930360674858, + 0.004185875877737999, + 0.002039998071268201, + 0.0024486654438078403, + 0.005401602014899254, + 0.005210550036281347, + 0.008383466862142086, + 0.009208551608026028, + 0.004375644959509373, + 0.007255093194544315, + 0.004092660266906023, + 0.004916276782751083, + 0.009531033225357533, + 0.008126073516905308, + 0.026462048292160034, + 0.014340192079544067, + 0.011701736599206924, + 0.018080182373523712, + 0.008578646928071976, + 0.013912291266024113, + 0.02146388776600361, + 0.03621013090014458, + 0.010350091382861137, + 0.007934102788567543, + 0.003657510271295905, + 0.03681880608201027, + 0.0072419289499521255, + 0.004732062574476004, + 0.06987465173006058, + 0.022021392360329628, + 0.0034488618839532137, + 0.00424060458317399, + 0.0049704909324646, + 0.005895308218896389, + 0.005576804745942354, + 0.00809924490749836, + 0.003943916875869036, + 0.004007724113762379, + 0.010470832698047161, + 0.014642074704170227, + 0.018078887835144997, + 0.01775790937244892, + 0.013840854167938232, + 0.010437475517392159, + 0.01146428007632494, + 0.009425475262105465, + 0.008086378686130047, + 0.01001682598143816, + 0.011250256560742855, + 0.01242805551737547, + 0.0071401698514819145, + 0.004402061924338341, + 0.00653683440759778, + 0.004862601403146982, + 0.006299192551523447, + 0.009861215949058533, + 0.007018426898866892, + 0.021751733496785164, + 0.00917927734553814, + 0.008255732245743275, + 0.009791286662220955, + 0.004394974559545517, + 0.013097523711621761, + 0.019318558275699615, + 0.01713518612086773, + 0.016264773905277252, + 0.02910173125565052, + 0.011761718429625034, + 0.018004128709435463, + 0.014539533294737339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.023300303146243095 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright yellow handheld sprayer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sprayer positioned on a weathered stone patio with gray and beige slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, viscous, vibrant green liquid gush from a black nozzle in a steady stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the green liquid fizz, bubble, and form tiny shimmering spheres that burst into light vapor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a low-angle static close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369441255927086 + ], + "opposite_logprob_tokens": [ + -3.4086945056915283 + ], + "logprob_reward": -0.03369441255927086, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369441255927086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05985341966152191, + 0.06415249407291412, + 0.07768268138170242, + 0.17303168773651123, + 0.06147456169128418, + 0.03608998283743858, + 0.056417398154735565, + 0.06974044442176819, + 0.37015214562416077, + 0.03140517324209213 + ], + "credit_map": [ + 0.005900898482650518, + 0.005433171987533569, + 0.0036224224604666233, + 0.01130856666713953, + 0.007880805060267448, + 0.007535106502473354, + 0.013938844203948975, + 0.0042336005717515945, + 0.01798044890165329, + 0.005067965015769005, + 0.02415982447564602, + 0.0038889136631041765, + 0.0017827521078288555, + 0.002622878411784768, + 0.005326705053448677, + 0.0033230057451874018, + 0.0158611498773098, + 0.02818303368985653, + 0.005149524193257093, + 0.006855885963886976, + 0.004975010175257921, + 0.0034781538415700197, + 0.007701625116169453, + 0.005478294100612402, + 0.04228340461850166, + 0.01560547947883606, + 0.018583767116069794, + 0.013035469688475132, + 0.00648927828297019, + 0.01364077627658844, + 0.02420624904334545, + 0.03918725997209549, + 0.0037512423004955053, + 0.00333760236389935, + 0.002008519833907485, + 0.015086719766259193, + 0.0031969856936484575, + 0.0030371975153684616, + 0.024361705407500267, + 0.006694589741528034, + 0.0037943527568131685, + 0.0047721946612000465, + 0.003856855910271406, + 0.006593370344489813, + 0.0065422519110143185, + 0.0037195496261119843, + 0.004078180994838476, + 0.0027332280296832323, + 0.005808176472783089, + 0.007519358769059181, + 0.009078903123736382, + 0.007701236288994551, + 0.006382448133081198, + 0.00583798810839653, + 0.007444176357239485, + 0.006645113229751587, + 0.006652491632848978, + 0.01033787615597248, + 0.01279539056122303, + 0.012767941690981388, + 0.008667988702654839, + 0.004745493642985821, + 0.007483090739697218, + 0.006290177349001169, + 0.032686542719602585, + 0.052880510687828064, + 0.031848736107349396, + 0.11647256463766098, + 0.049176137894392014, + 0.02120545692741871, + 0.05009126290678978, + 0.015790944918990135, + 0.003030736232176423, + 0.004761453252285719, + 0.004437318071722984, + 0.004872927907854319, + 0.0034941858612000942, + 0.0043622651137411594, + 0.0033039262052625418, + 0.003142361994832754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.03369441255927086 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_8/index_tensor([433, 433])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_8/index_tensor([433, 433])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3664966c69739d98d31d6110eeef27ba42016685 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_8/index_tensor([433, 433])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot with a handheld camera, this high-angle extreme close-up, slightly trembling and with a palpable sense of anticipation, focuses on an ornate red envelope as it drops into the frame. The envelope, made of thick, finely textured vermilion paper, is adorned with an intricately embossed gold leaf pattern featuring swirling dragons, shimmering under the warm spotlight overhead. Below, a vast pool of shimmering black silk awaits, its surface a soft, undulating fold that catches the light, creating an abyss of deep, velvety shadows with a liquid sheen. The envelope deliberately falls downward, its corner first making contact with the silk, sending subtle, slow-motion ripples across the dark surface. As the luxurious black fabric yields, it continues its steady and graceful descent, the rich crimson paper disappearing as the black, flowing folds approach. Capturing the final moment as the last corner of the shimmering gold leaf is engulfed by the silk, the scene leaves a profound sense of luxury, refinement, and a quietly dramatic ending.", + "seed": 18010, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_8/index_433_seed_18010_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06521401749923825, + "vlm_logprob": -0.06521401749923825, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.06521401749923825, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04474518075585365 + ], + "opposite_logprob_tokens": [ + -3.4197452068328857 + ], + "logprob_reward": -0.04474518075585365, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04474518075585365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10831435769796371, + 0.050804171711206436, + 0.04503802955150604, + 0.06683675944805145, + 0.05848349630832672, + 0.0534500777721405, + 0.044367510825395584, + 0.08116083592176437, + 0.3445092439651489, + 0.14703547954559326 + ], + "credit_map": [ + 0.025071002542972565, + 0.010874750092625618, + 0.02469492331147194, + 0.012024447321891785, + 0.0061261942610144615, + 0.011555135250091553, + 0.008585219271481037, + 0.009382694028317928, + 0.007645501289516687, + 0.004857840482145548, + 0.009694916196167469, + 0.006847039330750704, + 0.0048535666428506374, + 0.005902030970901251, + 0.006114151794463396, + 0.004889125004410744, + 0.007021323312073946, + 0.005439532920718193, + 0.006307018455117941, + 0.004492388106882572, + 0.00479018222540617, + 0.005944355390965939, + 0.004266535397619009, + 0.006776695605367422, + 0.010308032855391502, + 0.009657971560955048, + 0.007569856476038694, + 0.0074224877171218395, + 0.0066220397129654884, + 0.009995395317673683, + 0.004836583975702524, + 0.01042439416050911, + 0.011953925713896751, + 0.004603619687259197, + 0.005295855458825827, + 0.005683241877704859, + 0.004427850712090731, + 0.01249406673014164, + 0.006588979158550501, + 0.007435962557792664, + 0.00706457206979394, + 0.0054824515245854855, + 0.006831462495028973, + 0.006449751555919647, + 0.004514907021075487, + 0.0069906883873045444, + 0.007343091536313295, + 0.008773159235715866, + 0.006683675106614828, + 0.004360490012913942, + 0.0031298554968088865, + 0.006279969122260809, + 0.00431034853681922, + 0.007375092711299658, + 0.004788794554769993, + 0.00743928924202919, + 0.011232622899115086, + 0.009316355921328068, + 0.01020839437842369, + 0.007915971800684929, + 0.0066351518034935, + 0.011612090282142162, + 0.007786547765135765, + 0.016453707590699196, + 0.056484583765268326, + 0.053431253880262375, + 0.034602805972099304, + 0.04481436684727669, + 0.024997681379318237, + 0.05634742230176926, + 0.022829629480838776, + 0.051001522690057755, + 0.018250910565257072, + 0.01379366498440504, + 0.02086573839187622, + 0.015764178708195686, + 0.013757742941379547, + 0.018344786018133163, + 0.023561939597129822, + 0.02269653230905533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.04474518075585365 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016099337488412857 + ], + "opposite_logprob_tokens": [ + -4.141099452972412 + ], + "logprob_reward": -0.016099337488412857, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016099337488412857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1620585024356842, + 0.10729077458381653, + 0.03140295296907425, + 0.06459071487188339, + 0.061733875423669815, + 0.033191364258527756, + 0.07156572490930557, + 0.1437939554452896, + 0.1450733095407486, + 0.17929880321025848 + ], + "credit_map": [ + 0.036331672221422195, + 0.01525126677006483, + 0.030628899112343788, + 0.013485721312463284, + 0.012689901515841484, + 0.024484168738126755, + 0.014624962583184242, + 0.014561919495463371, + 0.01940898969769478, + 0.008597943931818008, + 0.02063056081533432, + 0.013367509469389915, + 0.00792424101382494, + 0.010625844821333885, + 0.014506204053759575, + 0.012229486368596554, + 0.004992043599486351, + 0.00232700165361166, + 0.004434024915099144, + 0.00300351670011878, + 0.004690277855843306, + 0.004588560666888952, + 0.002759812865406275, + 0.004607712849974632, + 0.011729424819350243, + 0.008438813500106335, + 0.00805173721164465, + 0.007995132356882095, + 0.006022694054991007, + 0.011452586390078068, + 0.004954751115292311, + 0.005945578683167696, + 0.008266720920801163, + 0.005809902213513851, + 0.007411975879222155, + 0.005874232389032841, + 0.005931763909757137, + 0.009806660003960133, + 0.00614058505743742, + 0.012492034584283829, + 0.0038641972932964563, + 0.0048410287126898766, + 0.0032957452349364758, + 0.0039189280942082405, + 0.0038772134575992823, + 0.006298902444541454, + 0.002203842857852578, + 0.0048915063962340355, + 0.010034027509391308, + 0.013656790368258953, + 0.005984009709209204, + 0.009034759365022182, + 0.0073519861325621605, + 0.011511147022247314, + 0.005430247168987989, + 0.008562752045691013, + 0.019787004217505455, + 0.01935470663011074, + 0.015452275983989239, + 0.015130254440009594, + 0.013220198452472687, + 0.024059994146227837, + 0.010849927552044392, + 0.025939591228961945, + 0.023080909624695778, + 0.017601128667593002, + 0.014795797877013683, + 0.0197189562022686, + 0.00995026808232069, + 0.02510027028620243, + 0.011411448940634727, + 0.02341453731060028, + 0.02663220465183258, + 0.01768980361521244, + 0.026641370728611946, + 0.017656993120908737, + 0.012299906462430954, + 0.029011504724621773, + 0.012694559060037136, + 0.036672454327344894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.016099337488412857 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370075300335884 + ], + "opposite_logprob_tokens": [ + -3.408700704574585 + ], + "logprob_reward": -0.03370075300335884, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370075300335884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23225924372673035, + 0.08460399508476257, + 0.07621218264102936, + 0.04541885107755661, + 0.024803142994642258, + 0.055386755615472794, + 0.06672165542840958, + 0.13152889907360077, + 0.09989132732152939, + 0.18317387998104095 + ], + "credit_map": [ + 0.06007925420999527, + 0.014748113229870796, + 0.06422845274209976, + 0.022873224690556526, + 0.011316949501633644, + 0.019641581922769547, + 0.021968333050608635, + 0.017403364181518555, + 0.015523471869528294, + 0.007584040053188801, + 0.015750644728541374, + 0.010139291174709797, + 0.010177538730204105, + 0.008154953829944134, + 0.008578430861234665, + 0.008695629425346851, + 0.012829053215682507, + 0.010177572257816792, + 0.010039221495389938, + 0.004512202925980091, + 0.013275839388370514, + 0.007761992514133453, + 0.005819507408887148, + 0.011796800419688225, + 0.008692297153174877, + 0.008875845931470394, + 0.0041509666480124, + 0.003961538430303335, + 0.006194285582751036, + 0.007465794682502747, + 0.0024190363474190235, + 0.0036590907257050276, + 0.004876643884927034, + 0.0025242820847779512, + 0.0028631272725760937, + 0.002627725014463067, + 0.0027083498425781727, + 0.003969431854784489, + 0.003156549297273159, + 0.002077034441754222, + 0.008471383713185787, + 0.009189755655825138, + 0.006261531729251146, + 0.00634834123775363, + 0.007052036002278328, + 0.009229079820215702, + 0.0037822609301656485, + 0.005052371881902218, + 0.008887790143489838, + 0.013291852548718452, + 0.005909336730837822, + 0.007815960794687271, + 0.006754919420927763, + 0.011288588866591454, + 0.004880272783339024, + 0.007892941124737263, + 0.017151188105344772, + 0.019042007625102997, + 0.01383676752448082, + 0.014159607701003551, + 0.012211035937070847, + 0.02128790132701397, + 0.012094144709408283, + 0.021746262907981873, + 0.02020900882780552, + 0.009226342663168907, + 0.009469378739595413, + 0.009680068120360374, + 0.005485887639224529, + 0.014553182758390903, + 0.007801239378750324, + 0.023466233164072037, + 0.022541720420122147, + 0.01873597875237465, + 0.032717298716306686, + 0.018777411431074142, + 0.011931410990655422, + 0.026719721034169197, + 0.01807577535510063, + 0.03367457166314125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.03370075300335884 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557398676872253 + ], + "opposite_logprob_tokens": [ + -1.600574016571045 + ], + "logprob_reward": -0.22557398676872253, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557398676872253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08807380497455597, + 0.04947342351078987, + 0.11367203295230865, + 0.05271052196621895, + 0.04602205753326416, + 0.039873190224170685, + 0.07445811480283737, + 0.08848980069160461, + 0.3232080936431885, + 0.12401895970106125 + ], + "credit_map": [ + 0.019060255959630013, + 0.007418797817081213, + 0.02138037048280239, + 0.00765386363491416, + 0.007207178510725498, + 0.010687961243093014, + 0.007593511138111353, + 0.007071865722537041, + 0.009505419991910458, + 0.002995801391080022, + 0.00923271756619215, + 0.00668965233489871, + 0.0031711275223642588, + 0.004996332339942455, + 0.005780926439911127, + 0.007101444993168116, + 0.026550373062491417, + 0.013863516971468925, + 0.01780163310468197, + 0.0071456520818173885, + 0.019142653793096542, + 0.010658069513738155, + 0.008834290318191051, + 0.009675850160419941, + 0.007421216461807489, + 0.010319974273443222, + 0.006446232087910175, + 0.005095309112221003, + 0.007341052405536175, + 0.006676052697002888, + 0.0031457054428756237, + 0.006264980416744947, + 0.006408942863345146, + 0.004872583318501711, + 0.005340537987649441, + 0.005723697133362293, + 0.005149835720658302, + 0.006115747150033712, + 0.00823982059955597, + 0.004170896019786596, + 0.0074651772156357765, + 0.005138401873409748, + 0.0032629123888909817, + 0.004371057264506817, + 0.004355090670287609, + 0.0057811373844742775, + 0.0037713265046477318, + 0.0057280887849628925, + 0.01121586561203003, + 0.011002515442669392, + 0.006336810998618603, + 0.00672541931271553, + 0.008527466095983982, + 0.01321318093687296, + 0.006172875873744488, + 0.01126397866755724, + 0.011941619217395782, + 0.011454900726675987, + 0.010447909124195576, + 0.009438875131309032, + 0.008878779597580433, + 0.01609722711145878, + 0.006479223258793354, + 0.01375127024948597, + 0.05689975991845131, + 0.04382097348570824, + 0.030052995309233665, + 0.03746933117508888, + 0.022240370512008667, + 0.0532376691699028, + 0.02066405490040779, + 0.05882294103503227, + 0.01655878685414791, + 0.008596201427280903, + 0.028626300394535065, + 0.012403937987983227, + 0.008858641609549522, + 0.015408660285174847, + 0.013690277934074402, + 0.019876152276992798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.22557398676872253 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950829479843378 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950829479843378, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950829479843378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1060013696551323, + 0.038641881197690964, + 0.04218016937375069, + 0.0430944487452507, + 0.008779028430581093, + 0.022037982940673828, + 0.032216090708971024, + 0.04134431481361389, + 0.30630946159362793, + 0.35939523577690125 + ], + "credit_map": [ + 0.019406210631132126, + 0.012078361585736275, + 0.021733686327934265, + 0.011631423607468605, + 0.008561648428440094, + 0.012226833961904049, + 0.009395362809300423, + 0.010967844165861607, + 0.007569995243102312, + 0.005068029277026653, + 0.007743136025965214, + 0.003113291459158063, + 0.003589924657717347, + 0.003794959280639887, + 0.004529766738414764, + 0.003232780611142516, + 0.007956283167004585, + 0.0048439996317029, + 0.007480686996132135, + 0.0031603146344423294, + 0.006230936385691166, + 0.004554490093141794, + 0.0033999986480921507, + 0.0045534600503742695, + 0.006243568379431963, + 0.0064638350158929825, + 0.00448893615975976, + 0.004661214537918568, + 0.005768818315118551, + 0.005792286712676287, + 0.0034933967981487513, + 0.006182392593473196, + 0.001883210032247007, + 0.0009303734987042844, + 0.0010323026217520237, + 0.0008505044388584793, + 0.0006409261259250343, + 0.0016667396994307637, + 0.0009355669608339667, + 0.0008394051110371947, + 0.0029988689348101616, + 0.002778660273179412, + 0.0022693572100251913, + 0.0024027822073549032, + 0.002932474948465824, + 0.003469361923635006, + 0.0022501861676573753, + 0.00293629034422338, + 0.004910299554467201, + 0.003764224238693714, + 0.002828544471412897, + 0.004481016658246517, + 0.0028536857571452856, + 0.005110675003379583, + 0.0033223049249500036, + 0.00494533684104681, + 0.005373884458094835, + 0.006268193479627371, + 0.004770889412611723, + 0.004307759925723076, + 0.003937254659831524, + 0.00656682625412941, + 0.003908806946128607, + 0.006210700608789921, + 0.052598997950553894, + 0.03907322511076927, + 0.030146801844239235, + 0.03420635685324669, + 0.023151196539402008, + 0.05243357643485069, + 0.022440826520323753, + 0.05225849524140358, + 0.04477819800376892, + 0.025944754481315613, + 0.05870196595788002, + 0.03096744231879711, + 0.03655432164669037, + 0.04583420976996422, + 0.04991873726248741, + 0.06669560074806213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.005950829479843378 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with a handheld camera, this high-angle extreme close-up, slightly trembling and with a palpable sense of anticipation, focuses on an ornate red envelope as it drops into the frame. The envelope, made of thick, finely textured vermilion paper, is adorned with an intricately embossed gold leaf pattern featuring swirling dragons, shimmering under the warm spotlight overhead. Below, a vast pool of shimmering black silk awaits, its surface a soft, undulating fold that catches the light, creating an abyss of deep, velvety shadows with a liquid sheen. The envelope deliberately falls downward, its corner first making contact with the silk, sending subtle, slow-motion ripples across the dark surface. As the luxurious black fabric yields, it continues its steady and graceful descent, the rich crimson paper disappearing as the black, flowing folds approach. Capturing the final moment as the last corner of the shimmering gold leaf is engulfed by the silk, the scene leaves a profound sense of luxury, refinement, and a quietly dramatic ending.", + "seed": 18011, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_8/index_433_seed_18011_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06423268392682076, + "vlm_logprob": -0.06423268392682076, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.06423268392682076, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06278331577777863 + ], + "opposite_logprob_tokens": [ + -3.0627832412719727 + ], + "logprob_reward": -0.06278331577777863, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06278331577777863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19502344727516174, + 0.09002215415239334, + 0.050813429057598114, + 0.06819572299718857, + 0.1290670782327652, + 0.04015342518687248, + 0.24966418743133545, + 0.07487887144088745, + 0.06197391077876091, + 0.04020780324935913 + ], + "credit_map": [ + 0.030725309625267982, + 0.030552227050065994, + 0.026040038093924522, + 0.030475346371531487, + 0.02167639136314392, + 0.016428641974925995, + 0.01924576610326767, + 0.019879743456840515, + 0.017125925049185753, + 0.017092019319534302, + 0.012825602665543556, + 0.011448252946138382, + 0.010667041875422001, + 0.007141566835343838, + 0.007719749119132757, + 0.006002001464366913, + 0.0066902791149914265, + 0.0055168126709759235, + 0.00674911355599761, + 0.008229725062847137, + 0.006739807780832052, + 0.005442087072879076, + 0.005503962282091379, + 0.005941643845289946, + 0.008833364583551884, + 0.004220922477543354, + 0.007854294031858444, + 0.010295477695763111, + 0.012048193253576756, + 0.008404096588492393, + 0.009927690029144287, + 0.00661168247461319, + 0.011641617864370346, + 0.006166419945657253, + 0.01968466490507126, + 0.01352397259324789, + 0.022546108812093735, + 0.015395435504615307, + 0.019768495112657547, + 0.020340366289019585, + 0.004801096394658089, + 0.0031788970809429884, + 0.0053605870343744755, + 0.006065791938453913, + 0.004872632212936878, + 0.003939034882932901, + 0.004689546301960945, + 0.007245837710797787, + 0.03678315505385399, + 0.01878056488931179, + 0.03362126275897026, + 0.0345388762652874, + 0.040516678243875504, + 0.015992505475878716, + 0.034970980137586594, + 0.03446017950773239, + 0.012757575139403343, + 0.0058600930497050285, + 0.009680351242423058, + 0.010325483977794647, + 0.009669151157140732, + 0.005963971372693777, + 0.010933487676084042, + 0.009688761085271835, + 0.010034635663032532, + 0.004776000045239925, + 0.011902336031198502, + 0.009363977238535881, + 0.007427881937474012, + 0.004629926756024361, + 0.007036850322037935, + 0.0068023111671209335, + 0.006481502205133438, + 0.004376807250082493, + 0.006034352350980043, + 0.0062794736586511135, + 0.003300692653283477, + 0.0033052910584956408, + 0.004822879564017057, + 0.005606809165328741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.06278331577777863 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033734291791915894 + ], + "opposite_logprob_tokens": [ + -3.4087343215942383 + ], + "logprob_reward": -0.033734291791915894, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033734291791915894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3519144058227539, + 0.06252007931470871, + 0.08892808854579926, + 0.12301552295684814, + 0.09769932180643082, + 0.04321194067597389, + 0.07352708280086517, + 0.06064266338944435, + 0.055289376527071, + 0.043251559138298035 + ], + "credit_map": [ + 0.05798519402742386, + 0.053428322076797485, + 0.04716016352176666, + 0.047528985887765884, + 0.04061103239655495, + 0.027573969215154648, + 0.035832539200782776, + 0.041794151067733765, + 0.00741594098508358, + 0.006516049616038799, + 0.007053188513964415, + 0.007731525227427483, + 0.014566299505531788, + 0.0072395214810967445, + 0.0062474035657942295, + 0.005750145297497511, + 0.01154413539916277, + 0.008188948035240173, + 0.014564553275704384, + 0.020405691117048264, + 0.009125599637627602, + 0.00919486302882433, + 0.007726746611297131, + 0.008177547715604305, + 0.018788866698741913, + 0.009434656240046024, + 0.018195169046521187, + 0.019129354506731033, + 0.02214282937347889, + 0.011546450667083263, + 0.012187247164547443, + 0.011590936221182346, + 0.01142867561429739, + 0.006698651239275932, + 0.014676596038043499, + 0.012451864778995514, + 0.01865132339298725, + 0.008136079646646976, + 0.014906177297234535, + 0.010749947279691696, + 0.006219496950507164, + 0.003539236029610038, + 0.0052532656118273735, + 0.005768387112766504, + 0.007136711850762367, + 0.003932125400751829, + 0.006893058307468891, + 0.004469654988497496, + 0.011413060128688812, + 0.004520531743764877, + 0.01041239034384489, + 0.010563435964286327, + 0.010564133524894714, + 0.005047425162047148, + 0.012152398936450481, + 0.008853696286678314, + 0.009815897792577744, + 0.004738405346870422, + 0.007691848091781139, + 0.007566716987639666, + 0.006387643050402403, + 0.004028117284178734, + 0.008700760081410408, + 0.011713269166648388, + 0.009864061139523983, + 0.004895500838756561, + 0.00884439330548048, + 0.009195574559271336, + 0.005909268278628588, + 0.0038914079777896404, + 0.006931648124009371, + 0.005757514387369156, + 0.00770382396876812, + 0.005529211834073067, + 0.006335800047963858, + 0.006944896187633276, + 0.004156582057476044, + 0.003418826265260577, + 0.004601173102855682, + 0.004561242181807756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.033734291791915894 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370121121406555 + ], + "opposite_logprob_tokens": [ + -3.408701181411743 + ], + "logprob_reward": -0.03370121121406555, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370121121406555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40041449666023254, + 0.09185045212507248, + 0.04854312166571617, + 0.05744825303554535, + 0.053663477301597595, + 0.09555698186159134, + 0.061441197991371155, + 0.06579581648111343, + 0.06311202049255371, + 0.06217411160469055 + ], + "credit_map": [ + 0.06971495598554611, + 0.046235088258981705, + 0.0567801408469677, + 0.05410733073949814, + 0.05982432886958122, + 0.03376159071922302, + 0.03879156708717346, + 0.04119952768087387, + 0.015908118337392807, + 0.014255478978157043, + 0.014192602597177029, + 0.01229819469153881, + 0.012123587541282177, + 0.007879712618887424, + 0.008638930507004261, + 0.006553837098181248, + 0.006376379169523716, + 0.00437817582860589, + 0.007755754981189966, + 0.011878997087478638, + 0.004050244111567736, + 0.004119433928281069, + 0.004616140853613615, + 0.005367995705455542, + 0.00849173590540886, + 0.004284122493118048, + 0.00788524653762579, + 0.011734253726899624, + 0.0058818659745156765, + 0.006807856261730194, + 0.005724787712097168, + 0.006638387218117714, + 0.006615208927541971, + 0.003977010492235422, + 0.007345701567828655, + 0.007657244335860014, + 0.007906797342002392, + 0.006580883637070656, + 0.007812503725290298, + 0.005768134258687496, + 0.011759163811802864, + 0.006100599654018879, + 0.012690054252743721, + 0.014137349091470242, + 0.013677702285349369, + 0.007207405287772417, + 0.013167345896363258, + 0.016817372292280197, + 0.00833496730774641, + 0.0052246456034481525, + 0.008086362853646278, + 0.007803072687238455, + 0.01075176615267992, + 0.004851353354752064, + 0.00894181802868843, + 0.00744721619412303, + 0.011151243932545185, + 0.0049776420928537846, + 0.007544221822172403, + 0.010945437476038933, + 0.006766705773770809, + 0.005086692050099373, + 0.008585687726736069, + 0.01073819026350975, + 0.011854718439280987, + 0.0071416390128433704, + 0.011745454743504524, + 0.008863307535648346, + 0.006509770639240742, + 0.004939362872391939, + 0.006552629638463259, + 0.005505148321390152, + 0.0105018625035882, + 0.00704918522387743, + 0.008641092106699944, + 0.009831869974732399, + 0.005163631401956081, + 0.004177660681307316, + 0.007314077578485012, + 0.009494737721979618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.03370121121406555 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798684448003769 + ], + "opposite_logprob_tokens": [ + -1.804868459701538 + ], + "logprob_reward": -0.1798684448003769, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798684448003769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2997410297393799, + 0.13878430426120758, + 0.05405379831790924, + 0.07721151411533356, + 0.038213927298784256, + 0.05758021026849747, + 0.13436667621135712, + 0.059675730764865875, + 0.09786306321620941, + 0.042509790509939194 + ], + "credit_map": [ + 0.04476296529173851, + 0.04403187707066536, + 0.04328708350658417, + 0.038271497935056686, + 0.032815463840961456, + 0.02697204425930977, + 0.03214450925588608, + 0.03745555132627487, + 0.021789195016026497, + 0.01485358364880085, + 0.01951051503419876, + 0.015170268714427948, + 0.027065830305218697, + 0.013214698061347008, + 0.017444392666220665, + 0.009735791943967342, + 0.009631341323256493, + 0.006270469631999731, + 0.007802464999258518, + 0.008969348855316639, + 0.006100181024521589, + 0.004375769291073084, + 0.0054062483832240105, + 0.005497970152646303, + 0.008719624020159245, + 0.0039924876764416695, + 0.009667367674410343, + 0.014013498090207577, + 0.01206787396222353, + 0.01105951052159071, + 0.008759209886193275, + 0.008931934833526611, + 0.004808966536074877, + 0.0025459553580731153, + 0.004866077098995447, + 0.005752909928560257, + 0.005805596709251404, + 0.003702996764332056, + 0.004862433299422264, + 0.005868986248970032, + 0.006268588826060295, + 0.0032020919024944305, + 0.007560682017356157, + 0.006204396020621061, + 0.009766287170350552, + 0.004032348748296499, + 0.0075204502791166306, + 0.013025358319282532, + 0.022372668609023094, + 0.0059629082679748535, + 0.016618171706795692, + 0.0163014754652977, + 0.017260339111089706, + 0.007622729521244764, + 0.025416923686861992, + 0.022811448201537132, + 0.008752278983592987, + 0.003822408150881529, + 0.007743378169834614, + 0.007007469888776541, + 0.010674110613763332, + 0.0043962751515209675, + 0.009939078241586685, + 0.00734072457998991, + 0.014438210986554623, + 0.01074955053627491, + 0.014733042567968369, + 0.016714472323656082, + 0.00918631348758936, + 0.007832802832126617, + 0.01299744751304388, + 0.011211209930479527, + 0.006066037807613611, + 0.0034890484530478716, + 0.005459747742861509, + 0.007061629090458155, + 0.003921419847756624, + 0.003083728486672044, + 0.0070675271563231945, + 0.006360644940286875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.1798684448003769 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076156049966812 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011076156049966812, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076156049966812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3207446038722992, + 0.15085595846176147, + 0.035999950021505356, + 0.06143168732523918, + 0.14361628890037537, + 0.06738802790641785, + 0.07040845602750778, + 0.06588214635848999, + 0.04306549206376076, + 0.04060739651322365 + ], + "credit_map": [ + 0.05084874853491783, + 0.03945032134652138, + 0.047352567315101624, + 0.0435398630797863, + 0.04443620890378952, + 0.029320422559976578, + 0.027778157964348793, + 0.038018327206373215, + 0.021463090553879738, + 0.024137847125530243, + 0.020577600225806236, + 0.03103836625814438, + 0.024255679920315742, + 0.010072015225887299, + 0.011886097490787506, + 0.007425257470458746, + 0.004915356636047363, + 0.003437204984948039, + 0.005153266247361898, + 0.007170248311012983, + 0.004284301307052374, + 0.003615008667111397, + 0.003396041225641966, + 0.004028520546853542, + 0.010272379964590073, + 0.005617964547127485, + 0.007828062400221825, + 0.008533481508493423, + 0.008476072922348976, + 0.0074704959988594055, + 0.007017680909484625, + 0.006215554196387529, + 0.01277226023375988, + 0.007230579387396574, + 0.021614767611026764, + 0.01623610407114029, + 0.025380117818713188, + 0.016749588772654533, + 0.0214595515280962, + 0.022173333913087845, + 0.009428080171346664, + 0.0061268992722034454, + 0.008359003812074661, + 0.008740662597119808, + 0.00931463297456503, + 0.004357185680419207, + 0.008809038437902927, + 0.01225251890718937, + 0.009507223032414913, + 0.004199672024697065, + 0.009295797906816006, + 0.011576981283724308, + 0.010084877721965313, + 0.004249144345521927, + 0.013706778176128864, + 0.0077879782766103745, + 0.011061032302677631, + 0.0049333651550114155, + 0.007386685814708471, + 0.009919914416968822, + 0.007671449799090624, + 0.005161468405276537, + 0.00937756896018982, + 0.01037065964192152, + 0.007096134591847658, + 0.00323014915920794, + 0.007011933717876673, + 0.006504688411951065, + 0.005663522984832525, + 0.0030875380616635084, + 0.005028342828154564, + 0.005443183705210686, + 0.006329997908324003, + 0.004313841927796602, + 0.005234652664512396, + 0.0064291395246982574, + 0.0029011934529989958, + 0.0031127939000725746, + 0.006746005266904831, + 0.005539771169424057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.011076156049966812 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with a handheld camera, this high-angle extreme close-up, slightly trembling and with a palpable sense of anticipation, focuses on an ornate red envelope as it drops into the frame. The envelope, made of thick, finely textured vermilion paper, is adorned with an intricately embossed gold leaf pattern featuring swirling dragons, shimmering under the warm spotlight overhead. Below, a vast pool of shimmering black silk awaits, its surface a soft, undulating fold that catches the light, creating an abyss of deep, velvety shadows with a liquid sheen. The envelope deliberately falls downward, its corner first making contact with the silk, sending subtle, slow-motion ripples across the dark surface. As the luxurious black fabric yields, it continues its steady and graceful descent, the rich crimson paper disappearing as the black, flowing folds approach. Capturing the final moment as the last corner of the shimmering gold leaf is engulfed by the silk, the scene leaves a profound sense of luxury, refinement, and a quietly dramatic ending.", + "seed": 18012, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_8/index_433_seed_18012_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.050560207664966585, + "vlm_logprob": -0.050560207664966585, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.050560207664966585, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.052780766040086746 + ], + "opposite_logprob_tokens": [ + -3.177780866622925 + ], + "logprob_reward": -0.052780766040086746, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.052780766040086746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08076499402523041, + 0.019367465749382973, + 0.05327905714511871, + 0.044803228229284286, + 0.02020537108182907, + 0.031155120581388474, + 0.1990184783935547, + 0.3805631399154663, + 0.11405064910650253, + 0.0567924864590168 + ], + "credit_map": [ + 0.005955778062343597, + 0.010100698098540306, + 0.013043939135968685, + 0.010786929167807102, + 0.00534029770642519, + 0.008174719288945198, + 0.015672452747821808, + 0.011690184473991394, + 0.0022830632515251637, + 0.002836846513673663, + 0.0032773520797491074, + 0.002686427440494299, + 0.001990126445889473, + 0.0019652151968330145, + 0.0024957519490271807, + 0.0018326813587918878, + 0.010215077549219131, + 0.006491268519312143, + 0.009257074445486069, + 0.004251559264957905, + 0.005552581045776606, + 0.005346708465367556, + 0.004891027696430683, + 0.0072737568989396095, + 0.006886723451316357, + 0.004910813644528389, + 0.0059999702498316765, + 0.006923735607415438, + 0.005501076113432646, + 0.006213323678821325, + 0.004868740681558847, + 0.003498845500871539, + 0.003014251822605729, + 0.002474795328453183, + 0.0035002748481929302, + 0.0023388536646962166, + 0.002015202771872282, + 0.0015310884919017553, + 0.0019501404603943229, + 0.0033807626459747553, + 0.004545930773019791, + 0.004194342531263828, + 0.004371714312583208, + 0.0035043649841099977, + 0.0037870644591748714, + 0.002859684405848384, + 0.0036236699670553207, + 0.004268351476639509, + 0.07741906493902206, + 0.010757273994386196, + 0.0718507468700409, + 0.00861380435526371, + 0.008568240329623222, + 0.006664834916591644, + 0.005965046584606171, + 0.009179468266665936, + 0.03297556936740875, + 0.02403990551829338, + 0.12860800325870514, + 0.016283631324768066, + 0.022238848730921745, + 0.1320931315422058, + 0.008874130435287952, + 0.01544991135597229, + 0.020994404330849648, + 0.009782373905181885, + 0.025605961680412292, + 0.007818625308573246, + 0.019029395654797554, + 0.015160244889557362, + 0.008522181771695614, + 0.007137461565434933, + 0.007631033658981323, + 0.005190017633140087, + 0.0065537854097783566, + 0.009326038882136345, + 0.005572365131229162, + 0.010374918580055237, + 0.0064557879231870174, + 0.005688536446541548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.052780766040086746 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038135625422000885 + ], + "opposite_logprob_tokens": [ + -3.288135528564453 + ], + "logprob_reward": -0.038135625422000885, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038135625422000885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1517782062292099, + 0.07017248123884201, + 0.1473381519317627, + 0.08738163113594055, + 0.06112271174788475, + 0.04724913090467453, + 0.039705194532871246, + 0.27169016003608704, + 0.03673696145415306, + 0.08682531863451004 + ], + "credit_map": [ + 0.011707000434398651, + 0.011350167915225029, + 0.024823933839797974, + 0.021002328023314476, + 0.009947773069143295, + 0.018522247672080994, + 0.03143391013145447, + 0.022990863770246506, + 0.007016581483185291, + 0.01273298729211092, + 0.008988273330032825, + 0.008967348374426365, + 0.00815440621227026, + 0.00944325141608715, + 0.008214493282139301, + 0.0066551403142511845, + 0.019368914887309074, + 0.01934777945280075, + 0.022646009922027588, + 0.012572795152664185, + 0.01703956350684166, + 0.014322836883366108, + 0.01572718657553196, + 0.02631308324635029, + 0.014743135310709476, + 0.007343487348407507, + 0.0147553076967597, + 0.011079508811235428, + 0.009640985168516636, + 0.01006810087710619, + 0.010671735741198063, + 0.009079374372959137, + 0.008986590430140495, + 0.007339793723076582, + 0.008054730482399464, + 0.011304939165711403, + 0.006045179441571236, + 0.005476250313222408, + 0.009147913195192814, + 0.004767318721860647, + 0.007133409380912781, + 0.006634412799030542, + 0.007070088293403387, + 0.004919315222650766, + 0.006160442251712084, + 0.0042876689694821835, + 0.005535820033401251, + 0.005507975816726685, + 0.008036757819354534, + 0.0042840116657316685, + 0.00813087448477745, + 0.004309898242354393, + 0.006044209934771061, + 0.002157158451154828, + 0.0032767546363174915, + 0.003465531161054969, + 0.020420163869857788, + 0.015169641003012657, + 0.091013602912426, + 0.011735852807760239, + 0.011051459237933159, + 0.10478711873292923, + 0.005291726905852556, + 0.012220596894621849, + 0.01109253615140915, + 0.0018080832669511437, + 0.012987938709557056, + 0.002024357672780752, + 0.0017885881243273616, + 0.0037879920564591885, + 0.0014753642026335, + 0.0017721033655107021, + 0.011230610311031342, + 0.012595362961292267, + 0.010291575454175472, + 0.01337585598230362, + 0.008163384161889553, + 0.014378753490746021, + 0.009189046919345856, + 0.007600739132612944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.038135625422000885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796700924634933 + ], + "opposite_logprob_tokens": [ + -3.529796600341797 + ], + "logprob_reward": -0.029796700924634933, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796700924634933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09289310872554779, + 0.03205529600381851, + 0.10293165594339371, + 0.1801474392414093, + 0.1298966109752655, + 0.04056892916560173, + 0.19224148988723755, + 0.05119595676660538, + 0.11233548074960709, + 0.06573405861854553 + ], + "credit_map": [ + 0.008638111874461174, + 0.0115382494404912, + 0.013929805718362331, + 0.011174383573234081, + 0.006839018315076828, + 0.01422208920121193, + 0.01369931735098362, + 0.012852131389081478, + 0.004347873851656914, + 0.0048471251502633095, + 0.004975112620741129, + 0.004282698966562748, + 0.0033150140661746264, + 0.0032422218937426805, + 0.0036654712166637182, + 0.0033797782380133867, + 0.015273327939212322, + 0.01554205734282732, + 0.01610557734966278, + 0.008864649571478367, + 0.008849114179611206, + 0.009233961813151836, + 0.008918745443224907, + 0.02014422044157982, + 0.030661283060908318, + 0.014149435795843601, + 0.030388791114091873, + 0.018301380798220634, + 0.019193682819604874, + 0.025730492547154427, + 0.023686982691287994, + 0.01803538389503956, + 0.024566777050495148, + 0.016584882512688637, + 0.02968653291463852, + 0.01127320621162653, + 0.007957503199577332, + 0.008103366009891033, + 0.01087920367717743, + 0.02084514871239662, + 0.00727524608373642, + 0.006843592971563339, + 0.006488586775958538, + 0.005731366574764252, + 0.0046533942222595215, + 0.0028275453951209784, + 0.0028983773663640022, + 0.003850822104141116, + 0.07406201213598251, + 0.010666262358427048, + 0.06806715577840805, + 0.008654664270579815, + 0.010068285278975964, + 0.006693040486425161, + 0.006151031237095594, + 0.00787903182208538, + 0.004986916668713093, + 0.007124697789549828, + 0.01340858917683363, + 0.00424661161378026, + 0.004017943050712347, + 0.01154634915292263, + 0.0028374698013067245, + 0.0030273767188191414, + 0.02350788563489914, + 0.009625197388231754, + 0.02118905447423458, + 0.010068170726299286, + 0.018806520849466324, + 0.010670959949493408, + 0.01061851717531681, + 0.00784916803240776, + 0.009193653240799904, + 0.006001455243676901, + 0.006759935524314642, + 0.009171448647975922, + 0.006169654428958893, + 0.012607358396053314, + 0.008309081196784973, + 0.007521477062255144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.029796700924634933 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902169018983841 + ], + "opposite_logprob_tokens": [ + -2.464021682739258 + ], + "logprob_reward": -0.08902169018983841, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902169018983841, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14200326800346375, + 0.02536197565495968, + 0.08321922272443771, + 0.1214861050248146, + 0.1566799432039261, + 0.01928332820534706, + 0.07934556156396866, + 0.22607648372650146, + 0.05834358558058739, + 0.08820050209760666 + ], + "credit_map": [ + 0.010052465833723545, + 0.013075117953121662, + 0.02176317758858204, + 0.019267696887254715, + 0.010839751921594143, + 0.013714390806853771, + 0.03298800811171532, + 0.020302684977650642, + 0.0028551756404340267, + 0.0025842562317848206, + 0.003504415974020958, + 0.0023763123899698257, + 0.0024469951167702675, + 0.0033059364650398493, + 0.003631451167166233, + 0.004657435230910778, + 0.012630781158804893, + 0.011823922395706177, + 0.013088059611618519, + 0.007079427596181631, + 0.005979187786579132, + 0.007294028531759977, + 0.008073682896792889, + 0.017250139266252518, + 0.01852213218808174, + 0.01091589406132698, + 0.018231136724352837, + 0.01383937057107687, + 0.014631696976721287, + 0.018822018057107925, + 0.013642191886901855, + 0.012881680391728878, + 0.02943451516330242, + 0.020496373996138573, + 0.03686490282416344, + 0.013301700353622437, + 0.015303398482501507, + 0.009772198274731636, + 0.011704660952091217, + 0.019802218303084373, + 0.002568103140220046, + 0.0024061016738414764, + 0.0029973292257636786, + 0.0032797781750559807, + 0.0017263205954805017, + 0.0017052514012902975, + 0.0015341502148658037, + 0.0030662950593978167, + 0.031238852068781853, + 0.003957577981054783, + 0.028658414259552956, + 0.005142596550285816, + 0.0023273152764886618, + 0.002444881247356534, + 0.00250995927490294, + 0.0030659777112305164, + 0.016604868695139885, + 0.012653051875531673, + 0.0751781240105629, + 0.009464163333177567, + 0.009169456548988819, + 0.0887899100780487, + 0.004191076382994652, + 0.010025851428508759, + 0.007381882052868605, + 0.0060340785421431065, + 0.008441363461315632, + 0.005599276162683964, + 0.011324483901262283, + 0.008713848888874054, + 0.005572163034230471, + 0.005276496987789869, + 0.010646011680364609, + 0.009754907339811325, + 0.011535394005477428, + 0.01132703572511673, + 0.009339598938822746, + 0.01637851446866989, + 0.009872042573988438, + 0.009347010403871536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.08902169018983841 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306625574827194 + ], + "opposite_logprob_tokens": [ + -3.1680662631988525 + ], + "logprob_reward": -0.04306625574827194, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306625574827194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11846208572387695, + 0.054411232471466064, + 0.050627049058675766, + 0.04607103019952774, + 0.04459221288561821, + 0.03607974946498871, + 0.15453636646270752, + 0.24356941878795624, + 0.15899202227592468, + 0.09265884011983871 + ], + "credit_map": [ + 0.009132394567131996, + 0.010778778232634068, + 0.02053365670144558, + 0.01649080216884613, + 0.0098187867552042, + 0.013188445009291172, + 0.023851534351706505, + 0.014667687937617302, + 0.006111595779657364, + 0.00964184571057558, + 0.007413005456328392, + 0.0056137279607355595, + 0.006206050049513578, + 0.007504109293222427, + 0.006601708475500345, + 0.005319193471223116, + 0.006930903065949678, + 0.007635682821273804, + 0.007159436121582985, + 0.0056881774216890335, + 0.005426786839962006, + 0.005081194452941418, + 0.005642763804644346, + 0.007062103599309921, + 0.007789822295308113, + 0.0048055932857096195, + 0.007219557184726, + 0.007922637276351452, + 0.004401619546115398, + 0.005385653581470251, + 0.0051515474915504456, + 0.003394596977159381, + 0.007149789482355118, + 0.0052789002656936646, + 0.007344339974224567, + 0.007126740179955959, + 0.003779676975682378, + 0.0030435658991336823, + 0.005131484009325504, + 0.005737714469432831, + 0.006220389623194933, + 0.005980724934488535, + 0.005698743741959333, + 0.0038295916747301817, + 0.005264783278107643, + 0.0029134005308151245, + 0.0027454106602817774, + 0.0034267057199031115, + 0.06126389279961586, + 0.007792260963469744, + 0.05580148845911026, + 0.006236168090254068, + 0.008063255809247494, + 0.004930778406560421, + 0.0043454356491565704, + 0.006103096529841423, + 0.01776006445288658, + 0.01769435778260231, + 0.06909862160682678, + 0.011889724060893059, + 0.016931846737861633, + 0.0877951830625534, + 0.010465919971466064, + 0.011933686211705208, + 0.034987371414899826, + 0.013384290970861912, + 0.03896578401327133, + 0.011210350319743156, + 0.02290349081158638, + 0.01848946325480938, + 0.010055303573608398, + 0.008995972573757172, + 0.012761888094246387, + 0.010223996825516224, + 0.010100194253027439, + 0.012030540965497494, + 0.009132258594036102, + 0.018242105841636658, + 0.011508025228977203, + 0.00865983497351408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.04306625574827194 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with a handheld camera, this high-angle extreme close-up, slightly trembling and with a palpable sense of anticipation, focuses on an ornate red envelope as it drops into the frame. The envelope, made of thick, finely textured vermilion paper, is adorned with an intricately embossed gold leaf pattern featuring swirling dragons, shimmering under the warm spotlight overhead. Below, a vast pool of shimmering black silk awaits, its surface a soft, undulating fold that catches the light, creating an abyss of deep, velvety shadows with a liquid sheen. The envelope deliberately falls downward, its corner first making contact with the silk, sending subtle, slow-motion ripples across the dark surface. As the luxurious black fabric yields, it continues its steady and graceful descent, the rich crimson paper disappearing as the black, flowing folds approach. Capturing the final moment as the last corner of the shimmering gold leaf is engulfed by the silk, the scene leaves a profound sense of luxury, refinement, and a quietly dramatic ending.", + "seed": 18013, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_8/index_433_seed_18013_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03056946974247694, + "vlm_logprob": -0.03056946974247694, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.03056946974247694, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03168122470378876 + ], + "opposite_logprob_tokens": [ + -3.9066812992095947 + ], + "logprob_reward": -0.03168122470378876, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03168122470378876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2736887037754059, + 0.08159254491329193, + 0.08773234486579895, + 0.0502854585647583, + 0.04074831306934357, + 0.06160425394773483, + 0.043068669736385345, + 0.2003260850906372, + 0.09879308938980103, + 0.06216052919626236 + ], + "credit_map": [ + 0.0335514098405838, + 0.025208143517374992, + 0.03773069381713867, + 0.029533114284276962, + 0.0596967376768589, + 0.04399644955992699, + 0.021585410460829735, + 0.02238675206899643, + 0.011111206375062466, + 0.01028844527900219, + 0.013575845398008823, + 0.012629829347133636, + 0.007514588534832001, + 0.008127243258059025, + 0.007352433167397976, + 0.010992959141731262, + 0.009573615156114101, + 0.014504611492156982, + 0.009392482228577137, + 0.007496290374547243, + 0.018552366644144058, + 0.00861149001866579, + 0.010072862729430199, + 0.009528618305921555, + 0.005483352579176426, + 0.006347115151584148, + 0.00648210383951664, + 0.007405589334666729, + 0.004466253798455, + 0.0032725343480706215, + 0.00812914315611124, + 0.008699365891516209, + 0.0041632093489170074, + 0.004407181870192289, + 0.006280038971453905, + 0.004060804843902588, + 0.005719858221709728, + 0.004183822777122259, + 0.004899040795862675, + 0.00703435530886054, + 0.004737201612442732, + 0.007282592821866274, + 0.010513966903090477, + 0.004279710352420807, + 0.008803827688097954, + 0.0045134685933589935, + 0.00906680803745985, + 0.012406676076352596, + 0.005166375543922186, + 0.003456612816080451, + 0.006959841586649418, + 0.0044409288093447685, + 0.0033820602111518383, + 0.0033984999172389507, + 0.008371027186512947, + 0.007893328554928303, + 0.012563898228108883, + 0.014496790245175362, + 0.0554313100874424, + 0.009936928749084473, + 0.015308634378015995, + 0.01085581909865141, + 0.031423620879650116, + 0.050309088081121445, + 0.009950850158929825, + 0.014371496625244617, + 0.018933754414319992, + 0.009640599600970745, + 0.01226811669766903, + 0.009786706417798996, + 0.005749918986111879, + 0.01809164509177208, + 0.004397551529109478, + 0.013308241963386536, + 0.0037901178002357483, + 0.009514151141047478, + 0.015271428972482681, + 0.0029781877528876066, + 0.006442233920097351, + 0.00645861541852355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.03168122470378876 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012560899369418621 + ], + "opposite_logprob_tokens": [ + -4.387560844421387 + ], + "logprob_reward": -0.012560899369418621, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012560899369418621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11374524235725403, + 0.10825634747743607, + 0.1595354825258255, + 0.03552386909723282, + 0.01548283826559782, + 0.03224443271756172, + 0.03601614758372307, + 0.04002070799469948, + 0.13976417481899261, + 0.3194108009338379 + ], + "credit_map": [ + 0.011536187492311, + 0.012952124699950218, + 0.011394325643777847, + 0.013210438191890717, + 0.028170160949230194, + 0.02010733261704445, + 0.007498838473111391, + 0.00887583289295435, + 0.017177550122141838, + 0.013475798070430756, + 0.018437547609210014, + 0.014094105921685696, + 0.01110641285777092, + 0.01107646245509386, + 0.00923838745802641, + 0.013650079257786274, + 0.01822199486196041, + 0.037658896297216415, + 0.03723546490073204, + 0.01149226538836956, + 0.016704261302947998, + 0.007345721125602722, + 0.011008575558662415, + 0.01986829750239849, + 0.003869208274409175, + 0.002688697073608637, + 0.004494273103773594, + 0.005160324741154909, + 0.0033285748213529587, + 0.002482878975570202, + 0.005978624802082777, + 0.007521288003772497, + 0.0018585757352411747, + 0.0019006290240213275, + 0.0022408084478229284, + 0.0015989352250471711, + 0.0022316966205835342, + 0.0016393400728702545, + 0.0015457511181011796, + 0.0024671016726642847, + 0.0025281694252043962, + 0.003711169585585594, + 0.005257710348814726, + 0.002523215487599373, + 0.005654965061694384, + 0.0027521688025444746, + 0.0034605918917804956, + 0.006356445606797934, + 0.0033278262708336115, + 0.005178576335310936, + 0.005207995418459177, + 0.0031731543131172657, + 0.0050468347035348415, + 0.003433508099988103, + 0.005689787678420544, + 0.004958464298397303, + 0.004237803164869547, + 0.005410440731793642, + 0.0064753200858831406, + 0.004358931910246611, + 0.005777672864496708, + 0.004939383361488581, + 0.004219642840325832, + 0.004601509775966406, + 0.009736692532896996, + 0.012447248212993145, + 0.039856377989053726, + 0.007973209954798222, + 0.011867980472743511, + 0.010275584645569324, + 0.01264992170035839, + 0.034957148134708405, + 0.01849202997982502, + 0.061188653111457825, + 0.01549313310533762, + 0.046673860400915146, + 0.10236097127199173, + 0.015101623721420765, + 0.03179018199443817, + 0.028310352936387062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.012560899369418621 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023286210373044014 + ], + "opposite_logprob_tokens": [ + -3.7732861042022705 + ], + "logprob_reward": -0.023286210373044014, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023286210373044014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5217814445495605, + 0.07192817330360413, + 0.24878151714801788, + 0.022245541214942932, + 0.01242130994796753, + 0.024384193122386932, + 0.019621295854449272, + 0.02380279079079628, + 0.034978218376636505, + 0.02005554921925068 + ], + "credit_map": [ + 0.0698397308588028, + 0.048866480588912964, + 0.06805824488401413, + 0.07159300148487091, + 0.08625505864620209, + 0.07035966217517853, + 0.053830552846193314, + 0.05297865718603134, + 0.009396473877131939, + 0.00892464816570282, + 0.011341916397213936, + 0.010441496968269348, + 0.0074821035377681255, + 0.008921121247112751, + 0.006458884570747614, + 0.008961512707173824, + 0.023214196786284447, + 0.0643741562962532, + 0.04533809795975685, + 0.016291966661810875, + 0.04071749374270439, + 0.014634309336543083, + 0.01527654193341732, + 0.028934728354215622, + 0.0027510132640600204, + 0.002701323013752699, + 0.0025161111261695623, + 0.0024851339403539896, + 0.002002944238483906, + 0.001467659487389028, + 0.004027490504086018, + 0.004293864127248526, + 0.001555922208353877, + 0.0010278956033289433, + 0.0019057703902944922, + 0.0019359696889296174, + 0.0010746833868324757, + 0.001438505481928587, + 0.0016009852988645434, + 0.0018815764924511313, + 0.0025916090235114098, + 0.0028373978566378355, + 0.0036963983438909054, + 0.003985524643212557, + 0.0030453396029770374, + 0.002081945538520813, + 0.0023346105590462685, + 0.003811365459114313, + 0.0012468347558751702, + 0.0025443050544708967, + 0.0031003367621451616, + 0.0013988395221531391, + 0.0037641541566699743, + 0.0015528236981481314, + 0.002448138315230608, + 0.0035658604465425014, + 0.002567316871136427, + 0.0023101726546883583, + 0.004809234756976366, + 0.0021221470087766647, + 0.0023808511905372143, + 0.002572645666077733, + 0.003174183191731572, + 0.0038662368897348642, + 0.0021230324637144804, + 0.00524855125695467, + 0.006219998002052307, + 0.003423916408792138, + 0.005023252684623003, + 0.003111819038167596, + 0.004311008844524622, + 0.005516635254025459, + 0.0014786813408136368, + 0.004229440353810787, + 0.0012340006651356816, + 0.0026799167972058058, + 0.005813955329358578, + 0.0014560278505086899, + 0.0014673840487375855, + 0.0016961408546194434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.023286210373044014 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062027446925640106 + ], + "opposite_logprob_tokens": [ + -2.8120274543762207 + ], + "logprob_reward": -0.062027446925640106, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062027446925640106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5308725237846375, + 0.08743344992399216, + 0.0793662741780281, + 0.019391629844903946, + 0.0250198133289814, + 0.05845170095562935, + 0.03021422028541565, + 0.03845611959695816, + 0.10598362237215042, + 0.02481066808104515 + ], + "credit_map": [ + 0.06442607939243317, + 0.05210576951503754, + 0.06582748889923096, + 0.06263568252325058, + 0.11016982793807983, + 0.07823564857244492, + 0.04661644995212555, + 0.05085557699203491, + 0.01816842518746853, + 0.013048148714005947, + 0.013980520889163017, + 0.010530294850468636, + 0.00966991763561964, + 0.008043021894991398, + 0.006519917398691177, + 0.007473200559616089, + 0.009584185667335987, + 0.015775829553604126, + 0.019116433337330818, + 0.006331536918878555, + 0.010598971508443356, + 0.0046714539639651775, + 0.004730004817247391, + 0.008557857945561409, + 0.002374421339482069, + 0.0022375164553523064, + 0.0023955826181918383, + 0.0019973847083747387, + 0.0024528182111680508, + 0.002165381796658039, + 0.00314215081743896, + 0.0026263752952218056, + 0.0022505694068968296, + 0.00296371360309422, + 0.003320972202345729, + 0.003555508330464363, + 0.0033478124532848597, + 0.0027090960647910833, + 0.0031038231682032347, + 0.003768316237255931, + 0.007949387654662132, + 0.004655813332647085, + 0.009956036694347858, + 0.012224855832755566, + 0.005315809510648251, + 0.004087393172085285, + 0.006075053010135889, + 0.008187349885702133, + 0.0034142418298870325, + 0.005135228857398033, + 0.004867054522037506, + 0.003183723194524646, + 0.0036721625365316868, + 0.002455903682857752, + 0.0030829589813947678, + 0.004402945749461651, + 0.0032631061039865017, + 0.004545294679701328, + 0.0067217410542070866, + 0.003170411568135023, + 0.005345327313989401, + 0.003916596528142691, + 0.005524027161300182, + 0.005969615653157234, + 0.009971456602215767, + 0.019357455894351006, + 0.010200594551861286, + 0.01505650021135807, + 0.017186207696795464, + 0.010502828285098076, + 0.014381766319274902, + 0.009326821193099022, + 0.0019079234916716814, + 0.0034972389694303274, + 0.0023348722606897354, + 0.004114831797778606, + 0.006307203322649002, + 0.001789060770533979, + 0.002661985345184803, + 0.0021975520066916943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.062027446925640106 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle extreme close-up with a handheld, slightly trembling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an ornate red envelope made of thick, finely textured vermilion paper with an embossed gold leaf pattern of swirling dragons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast pool of shimmering black silk with soft, undulating folds and deep, velvety shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the envelope fall downward corner-first into the silk, creating slow-motion ripples as it descends?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm spotlight overhead that makes the gold leaf pattern shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023291567340493202 + ], + "opposite_logprob_tokens": [ + -3.77329158782959 + ], + "logprob_reward": -0.023291567340493202, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023291567340493202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19308491051197052, + 0.18750202655792236, + 0.07316295057535172, + 0.02727358229458332, + 0.048836782574653625, + 0.05608086287975311, + 0.03259822726249695, + 0.13538162410259247, + 0.18258577585220337, + 0.06349331140518188 + ], + "credit_map": [ + 0.023571817204356194, + 0.023243913426995277, + 0.022349035367369652, + 0.01807502657175064, + 0.041878409683704376, + 0.032371725887060165, + 0.01590065099298954, + 0.015694336965680122, + 0.03025205433368683, + 0.019015874713659286, + 0.028899729251861572, + 0.02494647353887558, + 0.02204647846519947, + 0.026008306071162224, + 0.015014040283858776, + 0.021319057792425156, + 0.006521554663777351, + 0.011023590341210365, + 0.01275098416954279, + 0.006070524919778109, + 0.015135339461266994, + 0.006837684195488691, + 0.005208042450249195, + 0.00961522851139307, + 0.002134348265826702, + 0.00344366324134171, + 0.003953433129936457, + 0.002497795969247818, + 0.0037790315691381693, + 0.0028655771166086197, + 0.0030633227434009314, + 0.005536412354558706, + 0.004778166767209768, + 0.005101406946778297, + 0.006668514106422663, + 0.005200460087507963, + 0.0067202867940068245, + 0.004993850830942392, + 0.007102686446160078, + 0.008271414786577225, + 0.003916220273822546, + 0.008432664908468723, + 0.008676659315824509, + 0.0047820801846683025, + 0.009228953160345554, + 0.00523296557366848, + 0.006545631214976311, + 0.009265687316656113, + 0.002375771291553974, + 0.0032823393121361732, + 0.005128347314894199, + 0.002568461000919342, + 0.0044591110199689865, + 0.004188202787190676, + 0.005145864095538855, + 0.005450131371617317, + 0.008434507064521313, + 0.011319541372358799, + 0.030776383355259895, + 0.009454523213207722, + 0.012795157730579376, + 0.010837024077773094, + 0.023202331736683846, + 0.028562158346176147, + 0.01591983251273632, + 0.017772207036614418, + 0.04528554528951645, + 0.010012109763920307, + 0.015645073726773262, + 0.017052071169018745, + 0.019564734771847725, + 0.041334204375743866, + 0.006202676799148321, + 0.009775407612323761, + 0.004568879492580891, + 0.007419349625706673, + 0.014294154942035675, + 0.004909352399408817, + 0.008763507939875126, + 0.007559981197118759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.023291567340493202 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_9/index_tensor([5497, 5497])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_9/index_tensor([5497, 5497])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7128a67f874c93af5b4e6c1196fc7a1e7fb2779c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_9/index_tensor([5497, 5497])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, low-angle pan sweeps across a vast colony of pillar corals, towering against the murky backdrop of the deep ocean. A single, focused, and intensely dramatic sidelight, like a stage spotlight, sweeps steadily across the frame. As the beam grazes the pale, ivory-colored coral's surface, composed of countless tubular skeletons, it sharply and intimately depicts every intricate groove, tiny pore, and undulating texture of life. The light casts deep, sharply defined, inky shadows on the backlit coral, creating a stark contrast with the illuminated surface, sculpting this underwater wonder from the darkness. Tiny suspended marine particles shimmer in the beam, heightening the sense of underwater volume. The entire image is imbued with Annie Leibovitz's signature dramatic tension and iconography, depicting the silent coral colony as a living portrait imbued with majestic power and quiet beauty.", + "seed": 12797, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_9/index_5497_seed_12797_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08286791183054447, + "vlm_logprob": -0.08286791183054447, + "vlm_yes_no_margin": 2.675000023841858, + "vlm_token_credit": -0.08286791183054447, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15228812396526337 + ], + "opposite_logprob_tokens": [ + -2.0272881984710693 + ], + "logprob_reward": -0.15228812396526337, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.15228812396526337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05635443329811096, + 0.16432951390743256, + 0.1035911962389946, + 0.039858072996139526, + 0.15044771134853363, + 0.09649426490068436, + 0.0750402882695198, + 0.04245947673916817, + 0.15331239998340607, + 0.11811261624097824 + ], + "credit_map": [ + 0.007586606778204441, + 0.008372275158762932, + 0.005707663483917713, + 0.009967844001948833, + 0.006798658985644579, + 0.0051071918569505215, + 0.008174239657819271, + 0.004639956634491682, + 0.010408114641904831, + 0.016220321878790855, + 0.035998255014419556, + 0.06371204555034637, + 0.01019429974257946, + 0.008120201528072357, + 0.011498107574880123, + 0.008178172633051872, + 0.01181437075138092, + 0.02295602671802044, + 0.02088146284222603, + 0.017721589654684067, + 0.011239800602197647, + 0.005867642816156149, + 0.0066047897562384605, + 0.006505514029413462, + 0.003688099095597863, + 0.0052290852181613445, + 0.00684499554336071, + 0.006481857504695654, + 0.005488258320838213, + 0.003966388758271933, + 0.004480129573494196, + 0.0036792606115341187, + 0.021959597244858742, + 0.025434110313653946, + 0.0123756667599082, + 0.014365170150995255, + 0.017180077731609344, + 0.011907767504453659, + 0.03589504957199097, + 0.011330264620482922, + 0.014588269405066967, + 0.01191330049186945, + 0.0061308047734200954, + 0.006274575833231211, + 0.008022161200642586, + 0.007512309588491917, + 0.031784482300281525, + 0.01026836410164833, + 0.005944103933870792, + 0.01021368894726038, + 0.01057721022516489, + 0.010471250861883163, + 0.00920058973133564, + 0.008358674123883247, + 0.011508185416460037, + 0.008766583167016506, + 0.0034178956411778927, + 0.006165161728858948, + 0.004314274527132511, + 0.008597027510404587, + 0.004836027044802904, + 0.005130433477461338, + 0.005171238444745541, + 0.004827420227229595, + 0.01096330676227808, + 0.035349033772945404, + 0.013414393179118633, + 0.02971062809228897, + 0.017061885446310043, + 0.011846490204334259, + 0.020899837836623192, + 0.014066820964217186, + 0.010384926572442055, + 0.02444150485098362, + 0.011567547917366028, + 0.02430027909576893, + 0.012248996645212173, + 0.009168935008347034, + 0.017134813591837883, + 0.008865606971085072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 4 + ], + "selected_reward": -0.15228812396526337 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372311219573021 + ], + "opposite_logprob_tokens": [ + -3.4087231159210205 + ], + "logprob_reward": -0.03372311219573021, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372311219573021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25278186798095703, + 0.16055838763713837, + 0.14517542719841003, + 0.05589570105075836, + 0.05168847367167473, + 0.06549186259508133, + 0.07341022789478302, + 0.127689391374588, + 0.02172723039984703, + 0.045581381767988205 + ], + "credit_map": [ + 0.016549231484532356, + 0.03520728275179863, + 0.03792579099535942, + 0.08443903923034668, + 0.019916124641895294, + 0.015828801319003105, + 0.02735600434243679, + 0.01555961836129427, + 0.013754105195403099, + 0.03350535035133362, + 0.01077018678188324, + 0.043225813657045364, + 0.015538345091044903, + 0.017360232770442963, + 0.014813779853284359, + 0.01159058976918459, + 0.017740368843078613, + 0.020003927871584892, + 0.014130504801869392, + 0.03257002681493759, + 0.02616032026708126, + 0.012426749803125858, + 0.013772175647318363, + 0.008371356874704361, + 0.003975254949182272, + 0.005133842583745718, + 0.015241919085383415, + 0.012066727504134178, + 0.004473668988794088, + 0.003985648043453693, + 0.003660827409476042, + 0.007357816211879253, + 0.0056168315932154655, + 0.0072792829014360905, + 0.005109600257128477, + 0.006941085681319237, + 0.006041469983756542, + 0.005292187910526991, + 0.0109124556183815, + 0.004495564382523298, + 0.010493534617125988, + 0.010742937214672565, + 0.004218650981783867, + 0.005318518728017807, + 0.00759878009557724, + 0.005393214989453554, + 0.015766410157084465, + 0.005959821864962578, + 0.00515136169269681, + 0.009611115790903568, + 0.006044778041541576, + 0.009137558750808239, + 0.015202794224023819, + 0.011188442818820477, + 0.00945788249373436, + 0.007616301532834768, + 0.007611474487930536, + 0.0229113157838583, + 0.009503870271146297, + 0.019336363300681114, + 0.023701364174485207, + 0.016101017594337463, + 0.014901189133524895, + 0.013622819446027279, + 0.0012397984974086285, + 0.002531563164666295, + 0.0027547013014554977, + 0.0036321531515568495, + 0.002898941980674863, + 0.0040258727967739105, + 0.0022100929636508226, + 0.0024341088719666004, + 0.0035538068041205406, + 0.007506207097321749, + 0.0060178241692483425, + 0.0073188431560993195, + 0.006662944331765175, + 0.0038083354011178017, + 0.006605563685297966, + 0.004107861779630184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03372311219573021 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294150352478027 + ], + "opposite_logprob_tokens": [ + -2.2379415035247803 + ], + "logprob_reward": -0.11294150352478027, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294150352478027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1452137678861618, + 0.19276656210422516, + 0.1027546301484108, + 0.043736156076192856, + 0.07938707619905472, + 0.13124465942382812, + 0.08456700295209885, + 0.07462907582521439, + 0.04682500287890434, + 0.09887611120939255 + ], + "credit_map": [ + 0.015137315727770329, + 0.014350894838571548, + 0.0261186845600605, + 0.038321200758218765, + 0.012598490342497826, + 0.008415661752223969, + 0.018745074048638344, + 0.011526438407599926, + 0.01183334644883871, + 0.014486578293144703, + 0.048645954579114914, + 0.0778459832072258, + 0.01113910786807537, + 0.0077013676054775715, + 0.012049664743244648, + 0.00906454399228096, + 0.009432212449610233, + 0.027269484475255013, + 0.014861591160297394, + 0.016710294410586357, + 0.011288870126008987, + 0.007126048672944307, + 0.009780475869774818, + 0.006285652983933687, + 0.005432679783552885, + 0.008030935190618038, + 0.0060393912717700005, + 0.004723239224404097, + 0.006681334227323532, + 0.004908531904220581, + 0.0036071049980819225, + 0.0043129390105605125, + 0.0150302117690444, + 0.008300450630486012, + 0.005066820420324802, + 0.005565776955336332, + 0.006633087061345577, + 0.004761374555528164, + 0.027056843042373657, + 0.0069725122302770615, + 0.012046059593558311, + 0.02043299563229084, + 0.00915239192545414, + 0.009824657812714577, + 0.017163120210170746, + 0.012504469603300095, + 0.0336080938577652, + 0.016512878239154816, + 0.007614952977746725, + 0.014463978819549084, + 0.011918684467673302, + 0.011004356667399406, + 0.012769772671163082, + 0.008756096474826336, + 0.009018980897963047, + 0.009020176716148853, + 0.004825702402740717, + 0.017234990373253822, + 0.008023042231798172, + 0.010007522068917751, + 0.010859565809369087, + 0.008397893980145454, + 0.00814040843397379, + 0.007139954715967178, + 0.003422909416258335, + 0.00948534905910492, + 0.006208828650414944, + 0.0063824765384197235, + 0.006064532324671745, + 0.005277285818010569, + 0.0057460470125079155, + 0.0042375740595161915, + 0.009660927578806877, + 0.017486492171883583, + 0.010999340564012527, + 0.022160835564136505, + 0.010651475749909878, + 0.007907586172223091, + 0.01215052604675293, + 0.007858927361667156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.11294150352478027 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026361754164099693 + ], + "opposite_logprob_tokens": [ + -3.6513617038726807 + ], + "logprob_reward": -0.026361754164099693, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026361754164099693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12419767677783966, + 0.26718100905418396, + 0.07949747145175934, + 0.06351561844348907, + 0.21521104872226715, + 0.15328486263751984, + 0.03689093142747879, + 0.022918226197361946, + 0.02462184801697731, + 0.012681316584348679 + ], + "credit_map": [ + 0.009103002026677132, + 0.034466132521629333, + 0.012037470936775208, + 0.03410295024514198, + 0.00864042155444622, + 0.00554549740627408, + 0.012263448908925056, + 0.008038745261728764, + 0.01733499951660633, + 0.02035657875239849, + 0.05578499659895897, + 0.11538363993167877, + 0.01625210978090763, + 0.013387019746005535, + 0.016646478325128555, + 0.01203516498208046, + 0.010731259360909462, + 0.010171649046242237, + 0.00804560910910368, + 0.01355418935418129, + 0.01720169559121132, + 0.007934893481433392, + 0.006743210833519697, + 0.005114965606480837, + 0.007498109247535467, + 0.012549281120300293, + 0.005862928461283445, + 0.008822844363749027, + 0.009273757226765156, + 0.007115735672414303, + 0.006908627226948738, + 0.005484337918460369, + 0.04545215144753456, + 0.0235598087310791, + 0.013163883239030838, + 0.013101350516080856, + 0.019022436812520027, + 0.01208219863474369, + 0.06750208139419556, + 0.021327147260308266, + 0.02227761410176754, + 0.01484491303563118, + 0.010289446450769901, + 0.008171925321221352, + 0.01076025702059269, + 0.008233520202338696, + 0.05990617349743843, + 0.018801024183630943, + 0.002837174339219928, + 0.005315945018082857, + 0.005826058331876993, + 0.0046320524998009205, + 0.0038734632544219494, + 0.004781187511980534, + 0.004908604081720114, + 0.004716443829238415, + 0.0014727972447872162, + 0.004296583589166403, + 0.002296845195814967, + 0.00373972300440073, + 0.003013859037309885, + 0.002531569916754961, + 0.003178861690685153, + 0.0023879841901361942, + 0.0014388756826519966, + 0.005947227124124765, + 0.0018921279115602374, + 0.004338379018008709, + 0.002346325432881713, + 0.0029516161885112524, + 0.0032359459437429905, + 0.002471351996064186, + 0.0010302060982212424, + 0.001580444280989468, + 0.0016325587639585137, + 0.002568648662418127, + 0.0015977268340066075, + 0.0011460465611889958, + 0.0019435035064816475, + 0.0011821822263300419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.026361754164099693 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902506530284882 + ], + "opposite_logprob_tokens": [ + -2.4640250205993652 + ], + "logprob_reward": -0.08902506530284882, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902506530284882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12738244235515594, + 0.141306534409523, + 0.21608281135559082, + 0.0797208845615387, + 0.10886768996715546, + 0.1500122845172882, + 0.04860677570104599, + 0.0332755483686924, + 0.03505348414182663, + 0.05969151481986046 + ], + "credit_map": [ + 0.008547588251531124, + 0.01786244846880436, + 0.029664764180779457, + 0.039058003574609756, + 0.007776681333780289, + 0.006695932243019342, + 0.00898655317723751, + 0.008790460415184498, + 0.010660446248948574, + 0.018865585327148438, + 0.026782691478729248, + 0.04541241377592087, + 0.013572901487350464, + 0.007586338557302952, + 0.011885278858244419, + 0.006540877278894186, + 0.01883482187986374, + 0.013048220425844193, + 0.055772654712200165, + 0.0778377503156662, + 0.018553713336586952, + 0.009217403829097748, + 0.009821565821766853, + 0.012996682897210121, + 0.006242919247597456, + 0.007105169352144003, + 0.020658385008573532, + 0.01823691464960575, + 0.008998888544738293, + 0.005500635132193565, + 0.004995239432901144, + 0.00798273365944624, + 0.01570112630724907, + 0.015589411370456219, + 0.007656967267394066, + 0.011796713806688786, + 0.010741395875811577, + 0.007859033532440662, + 0.030619753524661064, + 0.008903288282454014, + 0.02867080830037594, + 0.01391507126390934, + 0.007730383425951004, + 0.007354005239903927, + 0.01104250643402338, + 0.007878610864281654, + 0.05586185306310654, + 0.017559053376317024, + 0.0033781288657337427, + 0.012081757187843323, + 0.005849732551723719, + 0.005445683375000954, + 0.006851883605122566, + 0.003544224426150322, + 0.0064055160619318485, + 0.005049849860370159, + 0.002436612965539098, + 0.00695680920034647, + 0.003025445155799389, + 0.005177708342671394, + 0.0056030647829174995, + 0.003437580307945609, + 0.0035259765572845936, + 0.0031123484950512648, + 0.002487372839823365, + 0.0075097521767020226, + 0.003295031376183033, + 0.005567973479628563, + 0.005505543202161789, + 0.0037962887436151505, + 0.004156795796006918, + 0.002734723500907421, + 0.004902827553451061, + 0.00835150945931673, + 0.008286753669381142, + 0.01132175698876381, + 0.007014978211373091, + 0.005082711577415466, + 0.009649738669395447, + 0.005081238690763712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.08902506530284882 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle pan sweeps across a vast colony of pillar corals, towering against the murky backdrop of the deep ocean. A single, focused, and intensely dramatic sidelight, like a stage spotlight, sweeps steadily across the frame. As the beam grazes the pale, ivory-colored coral's surface, composed of countless tubular skeletons, it sharply and intimately depicts every intricate groove, tiny pore, and undulating texture of life. The light casts deep, sharply defined, inky shadows on the backlit coral, creating a stark contrast with the illuminated surface, sculpting this underwater wonder from the darkness. Tiny suspended marine particles shimmer in the beam, heightening the sense of underwater volume. The entire image is imbued with Annie Leibovitz's signature dramatic tension and iconography, depicting the silent coral colony as a living portrait imbued with majestic power and quiet beauty.", + "seed": 12798, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_9/index_5497_seed_12798_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04002230614423752, + "vlm_logprob": -0.04002230614423752, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.04002230614423752, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08747735619544983 + ], + "opposite_logprob_tokens": [ + -2.587477445602417 + ], + "logprob_reward": -0.08747735619544983, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08747735619544983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19896942377090454, + 0.07244528830051422, + 0.05730675160884857, + 0.04192524775862694, + 0.03573651611804962, + 0.055636655539274216, + 0.051320984959602356, + 0.09735384583473206, + 0.2151559591293335, + 0.17414934933185577 + ], + "credit_map": [ + 0.023516850546002388, + 0.0377022884786129, + 0.02143402211368084, + 0.022909121587872505, + 0.02085128426551819, + 0.018058279529213905, + 0.02813788317143917, + 0.02635970152914524, + 0.010668560862541199, + 0.007522797677665949, + 0.010320613160729408, + 0.00761388149112463, + 0.008083605207502842, + 0.009513492695987225, + 0.01120235025882721, + 0.007519985549151897, + 0.0083163408562541, + 0.007274568546563387, + 0.006179237272590399, + 0.007034516893327236, + 0.007338912226259708, + 0.0067940205335617065, + 0.00682729110121727, + 0.0075418660417199135, + 0.005412624217569828, + 0.00600013742223382, + 0.005874812602996826, + 0.006224510725587606, + 0.005312779918313026, + 0.0052792225033044815, + 0.0036280937492847443, + 0.0041930656880140305, + 0.004591343924403191, + 0.004673421382904053, + 0.003675708081573248, + 0.004142899997532368, + 0.00475243479013443, + 0.00425370829179883, + 0.005130236502736807, + 0.004516763612627983, + 0.004225211683660746, + 0.005084101110696793, + 0.008071434684097767, + 0.0065174284391105175, + 0.005962057039141655, + 0.013325081206858158, + 0.006026768125593662, + 0.006424577906727791, + 0.006343527231365442, + 0.007901858538389206, + 0.005644960328936577, + 0.00626368960365653, + 0.005970458034425974, + 0.005268934648483992, + 0.006438903976231813, + 0.007488654926419258, + 0.011395822279155254, + 0.011244217865169048, + 0.015498816035687923, + 0.009122751653194427, + 0.007976029999554157, + 0.024083247408270836, + 0.007697292137891054, + 0.010335668921470642, + 0.026776516810059547, + 0.06204806640744209, + 0.03016500361263752, + 0.007951788604259491, + 0.010248098522424698, + 0.06227686628699303, + 0.009668487124145031, + 0.006021138746291399, + 0.013113181106746197, + 0.012922482565045357, + 0.02816120907664299, + 0.028728565201163292, + 0.019284971058368683, + 0.023356251418590546, + 0.028410794213414192, + 0.02017190121114254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.08747735619544983 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255878061056137 + ], + "opposite_logprob_tokens": [ + -4.387558937072754 + ], + "logprob_reward": -0.01255878061056137, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255878061056137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34951356053352356, + 0.0717858299612999, + 0.059220995754003525, + 0.03589390963315964, + 0.08339427411556244, + 0.087351955473423, + 0.030230076983571053, + 0.08102268725633621, + 0.14954929053783417, + 0.05203745514154434 + ], + "credit_map": [ + 0.037602197378873825, + 0.056155573576688766, + 0.0348542220890522, + 0.04077882319688797, + 0.04421941190958023, + 0.046292368322610855, + 0.04719948768615723, + 0.042411454021930695, + 0.009121819399297237, + 0.0072107394225895405, + 0.008286247961223125, + 0.006393565330654383, + 0.010771564207971096, + 0.012492685578763485, + 0.010467247106134892, + 0.007041963282972574, + 0.009900502860546112, + 0.006398160010576248, + 0.008273180574178696, + 0.006777207367122173, + 0.0055426498875021935, + 0.008536584675312042, + 0.007668613456189632, + 0.006124095059931278, + 0.004961549770087004, + 0.0038848884869366884, + 0.004825085401535034, + 0.003570753149688244, + 0.0035242640879005194, + 0.006135505624115467, + 0.004167534876614809, + 0.004824324045330286, + 0.00875017698854208, + 0.008020798675715923, + 0.012862936593592167, + 0.005957492161542177, + 0.011014312505722046, + 0.020722689107060432, + 0.007908488623797894, + 0.008157381787896156, + 0.010331888683140278, + 0.007883421145379543, + 0.009080599993467331, + 0.014364323578774929, + 0.00966548640280962, + 0.010136702097952366, + 0.013249830342829227, + 0.012639699503779411, + 0.0033620791509747505, + 0.004831591621041298, + 0.0026510730385780334, + 0.00276814098469913, + 0.003094002604484558, + 0.003971445839852095, + 0.004247497301548719, + 0.005304247140884399, + 0.0091834906488657, + 0.009678835980594158, + 0.01058768667280674, + 0.008735990151762962, + 0.007843997329473495, + 0.009921513497829437, + 0.009900389239192009, + 0.015170784667134285, + 0.014114705845713615, + 0.04007939249277115, + 0.02057241089642048, + 0.007239723112434149, + 0.007086341734975576, + 0.04397718608379364, + 0.009149868972599506, + 0.00732965674251318, + 0.004062917549163103, + 0.004269543569535017, + 0.010652564465999603, + 0.008723752573132515, + 0.003399363486096263, + 0.007863882929086685, + 0.008099180646240711, + 0.004966247361153364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.01255878061056137 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033707551658153534 + ], + "opposite_logprob_tokens": [ + -3.408707618713379 + ], + "logprob_reward": -0.033707551658153534, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033707551658153534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10069701075553894, + 0.060827597975730896, + 0.04341379180550575, + 0.023372836410999298, + 0.03927459195256233, + 0.11378584802150726, + 0.030940547585487366, + 0.03821912035346031, + 0.3565881550312042, + 0.1928805410861969 + ], + "credit_map": [ + 0.013114292174577713, + 0.016793768852949142, + 0.010515237227082253, + 0.010143952444195747, + 0.010669402778148651, + 0.01065328810364008, + 0.015583465807139874, + 0.013223598711192608, + 0.009645790793001652, + 0.007642720360308886, + 0.009591673500835896, + 0.006677124183624983, + 0.0048385364934802055, + 0.008389418944716454, + 0.00801095925271511, + 0.00603137444704771, + 0.006291778292506933, + 0.005101953167468309, + 0.005912423599511385, + 0.004596702754497528, + 0.004316792357712984, + 0.006684860680252314, + 0.005411864724010229, + 0.005097416695207357, + 0.003256752621382475, + 0.0033781116362661123, + 0.003712679026648402, + 0.003062690608203411, + 0.0024214230943471193, + 0.003027775324881077, + 0.002488541416823864, + 0.002024862915277481, + 0.004409360233694315, + 0.00518319895491004, + 0.0037003722973167896, + 0.004894907586276531, + 0.005426580552011728, + 0.003900541691109538, + 0.006460377015173435, + 0.005299252457916737, + 0.011619976721704006, + 0.00992300733923912, + 0.013375052250921726, + 0.01497387234121561, + 0.0136404512450099, + 0.02232356369495392, + 0.014446618966758251, + 0.013483302667737007, + 0.004458350129425526, + 0.004346807952970266, + 0.003338027745485306, + 0.0031691978219896555, + 0.003124767914414406, + 0.0033988531213253736, + 0.003915997222065926, + 0.005188546143472195, + 0.00603396026417613, + 0.0039581917226314545, + 0.006124275736510754, + 0.0023909409064799547, + 0.003400493413209915, + 0.010321383364498615, + 0.002779793692752719, + 0.0032100852113217115, + 0.03268183395266533, + 0.10058178007602692, + 0.05471183732151985, + 0.012366934679448605, + 0.013950522989034653, + 0.11687681823968887, + 0.01424136571586132, + 0.01117704901844263, + 0.015270342119038105, + 0.021960904821753502, + 0.032372284680604935, + 0.029092442244291306, + 0.014732979238033295, + 0.03360016644001007, + 0.026214804500341415, + 0.019636621698737144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.033707551658153534 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043083611875772476 + ], + "opposite_logprob_tokens": [ + -3.168083667755127 + ], + "logprob_reward": -0.043083611875772476, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043083611875772476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15493907034397125, + 0.10621543228626251, + 0.12483753263950348, + 0.04921828955411911, + 0.08362416923046112, + 0.07111635059118271, + 0.04693280905485153, + 0.06846596300601959, + 0.11474892497062683, + 0.17990148067474365 + ], + "credit_map": [ + 0.02328389510512352, + 0.02813635766506195, + 0.022121762856841087, + 0.015924379229545593, + 0.014925824478268623, + 0.011680479161441326, + 0.01879163272678852, + 0.020074736326932907, + 0.011597150936722755, + 0.013028380461037159, + 0.01093644741922617, + 0.009679046459496021, + 0.01611522026360035, + 0.01256855484098196, + 0.016537999734282494, + 0.015752620995044708, + 0.02049052156507969, + 0.014003284275531769, + 0.01594739593565464, + 0.016531897708773613, + 0.009084639139473438, + 0.016791485249996185, + 0.013912040740251541, + 0.01807626336812973, + 0.006369913928210735, + 0.006307513918727636, + 0.00675938231870532, + 0.00706655764952302, + 0.005069854203611612, + 0.008295228704810143, + 0.0045099882408976555, + 0.004839853383600712, + 0.011093937791883945, + 0.008382542058825493, + 0.01312873512506485, + 0.008598494343459606, + 0.007569707930088043, + 0.021672388538718224, + 0.0059821391478180885, + 0.007196223828941584, + 0.007222574669867754, + 0.0054862587712705135, + 0.009101555682718754, + 0.009321475401520729, + 0.006288295611739159, + 0.01633262448012829, + 0.008057876490056515, + 0.009305687621235847, + 0.00689298938959837, + 0.007874099537730217, + 0.005545399617403746, + 0.004318522289395332, + 0.00458007026463747, + 0.004798765294253826, + 0.005028218030929565, + 0.007894743233919144, + 0.011607366614043713, + 0.008270373567938805, + 0.00766330910846591, + 0.007185045164078474, + 0.006811285857111216, + 0.008564108051359653, + 0.008609354496002197, + 0.009755118750035763, + 0.012180080637335777, + 0.023622674867510796, + 0.018158089369535446, + 0.008412334136664867, + 0.007387590128928423, + 0.023487260565161705, + 0.011209890246391296, + 0.010291000828146935, + 0.017285706475377083, + 0.016890672966837883, + 0.03635897487401962, + 0.03249955177307129, + 0.013896184042096138, + 0.022859932854771614, + 0.025119362398982048, + 0.014991097152233124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.043083611875772476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328423038125038 + ], + "opposite_logprob_tokens": [ + -3.7732841968536377 + ], + "logprob_reward": -0.02328423038125038, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328423038125038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17813105881214142, + 0.07444615662097931, + 0.03960885480046272, + 0.041031837463378906, + 0.030345939099788666, + 0.22851192951202393, + 0.06196850538253784, + 0.03763128072023392, + 0.18413390219211578, + 0.12419048696756363 + ], + "credit_map": [ + 0.017587630078196526, + 0.018213223665952682, + 0.02062809467315674, + 0.020805586129426956, + 0.020015515387058258, + 0.023037608712911606, + 0.03198136016726494, + 0.025862039998173714, + 0.010358169674873352, + 0.008498815819621086, + 0.008624649606645107, + 0.009617148898541927, + 0.008188827894628048, + 0.00949912890791893, + 0.011655932292342186, + 0.00800348725169897, + 0.005683903116732836, + 0.004961904603987932, + 0.004862828645855188, + 0.0044156345538794994, + 0.0032253297977149487, + 0.004688991699367762, + 0.005781830754131079, + 0.005988430697470903, + 0.004888813942670822, + 0.004321592394262552, + 0.00580216059461236, + 0.003671246347948909, + 0.0039004518184810877, + 0.006575381848961115, + 0.005595623515546322, + 0.006276567932218313, + 0.004147650673985481, + 0.004006687551736832, + 0.0029452296439558268, + 0.0031943228095769882, + 0.003860973985865712, + 0.0028796682599931955, + 0.005028687883168459, + 0.004282716196030378, + 0.01775285229086876, + 0.018534468486905098, + 0.02863389626145363, + 0.03306213766336441, + 0.02061018906533718, + 0.03434716910123825, + 0.039238542318344116, + 0.03633265942335129, + 0.007797878701239824, + 0.007407207973301411, + 0.006736484356224537, + 0.006533651612699032, + 0.006720772944390774, + 0.006294370163232088, + 0.012363066896796227, + 0.008115073665976524, + 0.004680813290178776, + 0.005389639642089605, + 0.00467727379873395, + 0.004346437286585569, + 0.0037014943081885576, + 0.004316958133131266, + 0.004521314054727554, + 0.005997349973767996, + 0.01895836554467678, + 0.052354492247104645, + 0.025980032980442047, + 0.006159831769764423, + 0.007670569233596325, + 0.057335373014211655, + 0.008673394098877907, + 0.007001844234764576, + 0.013120605610311031, + 0.012248693034052849, + 0.025083081796765327, + 0.015687284991145134, + 0.009755675680935383, + 0.02382615953683853, + 0.01648394763469696, + 0.007985033094882965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.02328423038125038 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle pan sweeps across a vast colony of pillar corals, towering against the murky backdrop of the deep ocean. A single, focused, and intensely dramatic sidelight, like a stage spotlight, sweeps steadily across the frame. As the beam grazes the pale, ivory-colored coral's surface, composed of countless tubular skeletons, it sharply and intimately depicts every intricate groove, tiny pore, and undulating texture of life. The light casts deep, sharply defined, inky shadows on the backlit coral, creating a stark contrast with the illuminated surface, sculpting this underwater wonder from the darkness. Tiny suspended marine particles shimmer in the beam, heightening the sense of underwater volume. The entire image is imbued with Annie Leibovitz's signature dramatic tension and iconography, depicting the silent coral colony as a living portrait imbued with majestic power and quiet beauty.", + "seed": 12799, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_9/index_5497_seed_12799_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04240181408822537, + "vlm_logprob": -0.04240181408822537, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.04240181408822537, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0707026943564415 + ], + "opposite_logprob_tokens": [ + -2.8207027912139893 + ], + "logprob_reward": -0.0707026943564415, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0707026943564415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09760488569736481, + 0.08530755341053009, + 0.1111343652009964, + 0.057026490569114685, + 0.05454065650701523, + 0.08387967199087143, + 0.08361560106277466, + 0.202889546751976, + 0.0780307725071907, + 0.1459704339504242 + ], + "credit_map": [ + 0.012311301194131374, + 0.007939230650663376, + 0.010290170088410378, + 0.009525747038424015, + 0.012777808122336864, + 0.017269164323806763, + 0.01506716012954712, + 0.012424303218722343, + 0.005684437230229378, + 0.005247300956398249, + 0.0058761462569236755, + 0.006947264540940523, + 0.007080522831529379, + 0.008133008144795895, + 0.02370503731071949, + 0.022633833810687065, + 0.007492513861507177, + 0.013102966360747814, + 0.015738246962428093, + 0.01567300781607628, + 0.014786765910685062, + 0.015268124639987946, + 0.014578176662325859, + 0.014494556002318859, + 0.006292723584920168, + 0.0036830587778240442, + 0.009501868858933449, + 0.007872422225773335, + 0.006224681623280048, + 0.007135413587093353, + 0.008593759499490261, + 0.007722560781985521, + 0.005380970425903797, + 0.004786656238138676, + 0.007649970706552267, + 0.0066122752614319324, + 0.004167057573795319, + 0.004775035195052624, + 0.012554545886814594, + 0.00861414335668087, + 0.005912089720368385, + 0.005679984111338854, + 0.011605791747570038, + 0.013495365157723427, + 0.006457134149968624, + 0.010575268417596817, + 0.017862593755126, + 0.01229144912213087, + 0.004464052151888609, + 0.008220345713198185, + 0.011297805234789848, + 0.012731154449284077, + 0.0065252576023340225, + 0.0067915115505456924, + 0.022570548579096794, + 0.011014929041266441, + 0.011769218370318413, + 0.01286966260522604, + 0.027281736955046654, + 0.03814908117055893, + 0.014123035594820976, + 0.024285849183797836, + 0.04440454766154289, + 0.030006425455212593, + 0.008227848447859287, + 0.009433316066861153, + 0.011088532395660877, + 0.012209541164338589, + 0.005958640482276678, + 0.009612473659217358, + 0.011374070309102535, + 0.010126348584890366, + 0.017054148018360138, + 0.012446483597159386, + 0.01473164465278387, + 0.029951827600598335, + 0.011182970367372036, + 0.019422059878706932, + 0.021081900224089622, + 0.020099403336644173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.0707026943564415 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309217259287834 + ], + "opposite_logprob_tokens": [ + -3.1680922508239746 + ], + "logprob_reward": -0.04309217259287834, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309217259287834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11265495419502258, + 0.2585162818431854, + 0.16020703315734863, + 0.07551400363445282, + 0.057917773723602295, + 0.07210029661655426, + 0.11106844991445541, + 0.03471234068274498, + 0.03489301726222038, + 0.08241590112447739 + ], + "credit_map": [ + 0.012513404712080956, + 0.009197277016937733, + 0.014853586442768574, + 0.014067036099731922, + 0.013317994773387909, + 0.019470090046525, + 0.015827633440494537, + 0.01340792328119278, + 0.011665798723697662, + 0.007075925823301077, + 0.010755566880106926, + 0.012724760919809341, + 0.012888448312878609, + 0.017621668055653572, + 0.09548673778772354, + 0.09029736369848251, + 0.01498936116695404, + 0.015564404428005219, + 0.02408594638109207, + 0.020894702523946762, + 0.015432232990860939, + 0.0183874424546957, + 0.02716962993144989, + 0.02368330769240856, + 0.005798319820314646, + 0.006785072386264801, + 0.011920641176402569, + 0.012485193088650703, + 0.007929466664791107, + 0.009904607199132442, + 0.010824905708432198, + 0.009865796193480492, + 0.006076819263398647, + 0.008137241937220097, + 0.009609604254364967, + 0.005901555065065622, + 0.004336617421358824, + 0.006170128937810659, + 0.00984219741076231, + 0.00784360896795988, + 0.008050232194364071, + 0.006775223184376955, + 0.009958678856492043, + 0.007752229925245047, + 0.006568309385329485, + 0.009178536012768745, + 0.013169472105801105, + 0.010647610761225224, + 0.011377937160432339, + 0.012423793785274029, + 0.016612188890576363, + 0.015850374475121498, + 0.008490555919706821, + 0.009698866866528988, + 0.021952973678708076, + 0.014661752618849277, + 0.003932297695428133, + 0.0033285373356193304, + 0.004869113210588694, + 0.0050873118452727795, + 0.002810761099681258, + 0.0037062224000692368, + 0.006213007494807243, + 0.004765085410326719, + 0.003352648112922907, + 0.004717809613794088, + 0.005419659893959761, + 0.005120407789945602, + 0.0030397181399166584, + 0.003921845927834511, + 0.004637851379811764, + 0.004683072213083506, + 0.007657760288566351, + 0.007693861611187458, + 0.010518118739128113, + 0.016866162419319153, + 0.004806802608072758, + 0.010812481865286827, + 0.010841164737939835, + 0.01321954745799303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.04309217259287834 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009796753525733948 + ], + "opposite_logprob_tokens": [ + -4.634796619415283 + ], + "logprob_reward": -0.009796753525733948, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009796753525733948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06596869230270386, + 0.15518730878829956, + 0.041973959654569626, + 0.042744386941194534, + 0.04989122226834297, + 0.0784810334444046, + 0.08105827867984772, + 0.20175711810588837, + 0.1489511877298355, + 0.1339867264032364 + ], + "credit_map": [ + 0.0061676776967942715, + 0.005793564021587372, + 0.007428668439388275, + 0.008800913579761982, + 0.008657548576593399, + 0.010725943371653557, + 0.010162115097045898, + 0.008232271298766136, + 0.00978037528693676, + 0.0033471619244664907, + 0.005139798391610384, + 0.005927127320319414, + 0.005873722489923239, + 0.010231103748083115, + 0.05863100662827492, + 0.05625702813267708, + 0.004776929039508104, + 0.003261807607486844, + 0.004769396502524614, + 0.005427536554634571, + 0.004322014283388853, + 0.00460159033536911, + 0.007574703544378281, + 0.007239988073706627, + 0.005365676246583462, + 0.00385400652885437, + 0.007078594993799925, + 0.004651072435081005, + 0.0040812124498188496, + 0.005409787409007549, + 0.0064860377460718155, + 0.00581800052896142, + 0.004575997591018677, + 0.005091245751827955, + 0.006263498216867447, + 0.006243829149752855, + 0.0056006889790296555, + 0.006196798291057348, + 0.008725088089704514, + 0.007194081321358681, + 0.007261213846504688, + 0.0064431228674948215, + 0.009164211340248585, + 0.010381084866821766, + 0.008912232704460621, + 0.009133872576057911, + 0.016451383009552956, + 0.010733922012150288, + 0.005411560647189617, + 0.00797031819820404, + 0.011655841022729874, + 0.013133996166288853, + 0.0062717352993786335, + 0.00861962977796793, + 0.017993396148085594, + 0.010001813992857933, + 0.012870297767221928, + 0.015444512479007244, + 0.02577928826212883, + 0.037664931267499924, + 0.015778034925460815, + 0.026371514424681664, + 0.037678010761737823, + 0.03017054684460163, + 0.010820278897881508, + 0.011959164403378963, + 0.019379843026399612, + 0.024637354537844658, + 0.013337859883904457, + 0.01915420964360237, + 0.02738705649971962, + 0.022275440394878387, + 0.009707808494567871, + 0.010708469897508621, + 0.02210313268005848, + 0.02331935614347458, + 0.008283299393951893, + 0.017005247995257378, + 0.020065313205122948, + 0.022794123739004135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.009796753525733948 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062074389308691025 + ], + "opposite_logprob_tokens": [ + -2.8120744228363037 + ], + "logprob_reward": -0.062074389308691025, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062074389308691025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11365239322185516, + 0.14683681726455688, + 0.11946392804384232, + 0.07332540303468704, + 0.17150196433067322, + 0.04444139450788498, + 0.04925854876637459, + 0.13168491423130035, + 0.07778956741094589, + 0.07204505801200867 + ], + "credit_map": [ + 0.01224190928041935, + 0.007898353971540928, + 0.012024518102407455, + 0.010192367248237133, + 0.011778904125094414, + 0.014916318468749523, + 0.023109322413802147, + 0.02149070054292679, + 0.013213662430644035, + 0.008185680955648422, + 0.012315243482589722, + 0.010499872267246246, + 0.010856732726097107, + 0.01185748539865017, + 0.04059774428606033, + 0.03931040316820145, + 0.009850255213677883, + 0.008819330483675003, + 0.01618078351020813, + 0.01128362212330103, + 0.013857794925570488, + 0.022174183279275894, + 0.020560694858431816, + 0.01673726737499237, + 0.006704919971525669, + 0.0051413909532129765, + 0.008234555833041668, + 0.013418098911643028, + 0.008207138627767563, + 0.010833642445504665, + 0.010369001887738705, + 0.010416657663881779, + 0.01917940564453602, + 0.025174565613269806, + 0.03871471807360649, + 0.018807876855134964, + 0.009536318480968475, + 0.016973285004496574, + 0.016210457310080528, + 0.026905354112386703, + 0.004176993854343891, + 0.003414345206692815, + 0.00552871311083436, + 0.005461357533931732, + 0.0040619755163788795, + 0.0049217757768929005, + 0.01015209499746561, + 0.006724140141159296, + 0.00816639605909586, + 0.005206947680562735, + 0.005889831110835075, + 0.0065168109722435474, + 0.004677458666265011, + 0.005135780666023493, + 0.00711338734254241, + 0.006551940459758043, + 0.007728027179837227, + 0.008325979113578796, + 0.01615229994058609, + 0.020094074308872223, + 0.008032158017158508, + 0.013539337553083897, + 0.03618719056248665, + 0.021625855937600136, + 0.00495160510763526, + 0.005436423700302839, + 0.008229091763496399, + 0.013073418289422989, + 0.006820847280323505, + 0.012377998791635036, + 0.015418234281241894, + 0.011481943540275097, + 0.005314568057656288, + 0.00657171756029129, + 0.010152817703783512, + 0.014233946800231934, + 0.004501266870647669, + 0.00853220745921135, + 0.010519551113247871, + 0.012218988500535488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.062074389308691025 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634306065738201 + ], + "opposite_logprob_tokens": [ + -3.6513431072235107 + ], + "logprob_reward": -0.02634306065738201, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634306065738201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056567270308732986, + 0.11405328661203384, + 0.051998935639858246, + 0.021969199180603027, + 0.04522036761045456, + 0.031144997105002403, + 0.07343316078186035, + 0.1995529681444168, + 0.1094672903418541, + 0.2965925335884094 + ], + "credit_map": [ + 0.006249857135117054, + 0.0051975860260427, + 0.006355201825499535, + 0.008443144150078297, + 0.006575842387974262, + 0.00708454055711627, + 0.008969413116574287, + 0.007691678125411272, + 0.007929263636469841, + 0.00523694371804595, + 0.008775381371378899, + 0.009759224951267242, + 0.008222302421927452, + 0.008536702021956444, + 0.0334155410528183, + 0.03217792138457298, + 0.005289128515869379, + 0.006397759076207876, + 0.008579343557357788, + 0.006558454595506191, + 0.006267065182328224, + 0.006299810018390417, + 0.006148132961243391, + 0.0064592426642775536, + 0.002311449032276869, + 0.0029993820935487747, + 0.00332808680832386, + 0.0033799433149397373, + 0.002816905966028571, + 0.002372351009398699, + 0.002268292475491762, + 0.0024927863851189613, + 0.0040032691322267056, + 0.003929274156689644, + 0.007781466469168663, + 0.006097913254052401, + 0.0033811922185122967, + 0.0062748705968260765, + 0.007426470518112183, + 0.006325908005237579, + 0.0026181053835898638, + 0.0037734038196504116, + 0.0030014922376722097, + 0.004434894770383835, + 0.0027223885990679264, + 0.0031940650660544634, + 0.006937543861567974, + 0.004463100805878639, + 0.00790751725435257, + 0.00561547139659524, + 0.01069146953523159, + 0.010580153204500675, + 0.0052940016612410545, + 0.007503274362534285, + 0.015237618237733841, + 0.010603652335703373, + 0.016857463866472244, + 0.017259864136576653, + 0.032705552875995636, + 0.027670782059431076, + 0.014605686068534851, + 0.025899656116962433, + 0.03970223665237427, + 0.024851715192198753, + 0.00801049079746008, + 0.009679322130978107, + 0.016684433445334435, + 0.017532076686620712, + 0.00963122583925724, + 0.013915116898715496, + 0.01859593205153942, + 0.015418673865497112, + 0.018940968438982964, + 0.02446177788078785, + 0.05101904273033142, + 0.05249647796154022, + 0.019277423620224, + 0.03435301035642624, + 0.04828525334596634, + 0.047758549451828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.02634306065738201 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle pan sweeps across a vast colony of pillar corals, towering against the murky backdrop of the deep ocean. A single, focused, and intensely dramatic sidelight, like a stage spotlight, sweeps steadily across the frame. As the beam grazes the pale, ivory-colored coral's surface, composed of countless tubular skeletons, it sharply and intimately depicts every intricate groove, tiny pore, and undulating texture of life. The light casts deep, sharply defined, inky shadows on the backlit coral, creating a stark contrast with the illuminated surface, sculpting this underwater wonder from the darkness. Tiny suspended marine particles shimmer in the beam, heightening the sense of underwater volume. The entire image is imbued with Annie Leibovitz's signature dramatic tension and iconography, depicting the silent coral colony as a living portrait imbued with majestic power and quiet beauty.", + "seed": 12800, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000072/group_9/index_5497_seed_12800_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03805170115083456, + "vlm_logprob": -0.03805170115083456, + "vlm_yes_no_margin": 3.674999952316284, + "vlm_token_credit": -0.03805170115083456, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.058999475091695786 + ], + "opposite_logprob_tokens": [ + -3.183999538421631 + ], + "logprob_reward": -0.058999475091695786, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.058999475091695786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15593408048152924, + 0.32568445801734924, + 0.09181153774261475, + 0.02317962795495987, + 0.031568389385938644, + 0.03036617673933506, + 0.2132440060377121, + 0.06079937517642975, + 0.030593497678637505, + 0.036818768829107285 + ], + "credit_map": [ + 0.0375952422618866, + 0.010056786239147186, + 0.010650668293237686, + 0.03301815688610077, + 0.007922363467514515, + 0.014025099575519562, + 0.018531804904341698, + 0.024133963510394096, + 0.04914705082774162, + 0.030328666791319847, + 0.029672302305698395, + 0.04685509204864502, + 0.018830984830856323, + 0.02543836459517479, + 0.05072536692023277, + 0.07468666136264801, + 0.011705469340085983, + 0.01081089861690998, + 0.01212860457599163, + 0.023941511288285255, + 0.008515477180480957, + 0.004969780333340168, + 0.010911192744970322, + 0.008828610181808472, + 0.0027120010927319527, + 0.0033758601639419794, + 0.0031712993513792753, + 0.006136111915111542, + 0.0020741650369018316, + 0.0014852583408355713, + 0.0021612360142171383, + 0.0020636986009776592, + 0.0037708012387156487, + 0.0037108950782567263, + 0.004866426344960928, + 0.006662361323833466, + 0.0017798793269321322, + 0.0020298673771321774, + 0.005477281287312508, + 0.0032708789221942425, + 0.0035380078479647636, + 0.0025958074256777763, + 0.0029838846530765295, + 0.007333502173423767, + 0.002541591878980398, + 0.002619770122691989, + 0.004958223085850477, + 0.003795391181483865, + 0.020389730110764503, + 0.013447473756968975, + 0.011384259909391403, + 0.04194735363125801, + 0.01677411049604416, + 0.012744269333779812, + 0.07073353230953217, + 0.02582329511642456, + 0.006945643108338118, + 0.0035787983797490597, + 0.002978493692353368, + 0.006589778698980808, + 0.003188903210684657, + 0.0064147706143558025, + 0.015660470351576805, + 0.015442518517374992, + 0.0022317455150187016, + 0.003737156745046377, + 0.0038332385011017323, + 0.004662208259105682, + 0.003340082708746195, + 0.005465701688081026, + 0.003814560594037175, + 0.0035088087897747755, + 0.003712109290063381, + 0.0028559581842273474, + 0.00485515920445323, + 0.005654545035213232, + 0.00331681314855814, + 0.006605667993426323, + 0.005472698714584112, + 0.004345824010670185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.058999475091695786 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018219612538814545 + ], + "opposite_logprob_tokens": [ + -4.018219470977783 + ], + "logprob_reward": -0.018219612538814545, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018219612538814545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40222740173339844, + 0.30773648619651794, + 0.04645320773124695, + 0.0178300179541111, + 0.015500253066420555, + 0.013043134473264217, + 0.0807638168334961, + 0.03441367298364639, + 0.05082608014345169, + 0.031205931678414345 + ], + "credit_map": [ + 0.10404883325099945, + 0.02598465420305729, + 0.023302234709262848, + 0.07439372688531876, + 0.015154278837144375, + 0.03604535385966301, + 0.04522092267870903, + 0.07807739824056625, + 0.05941179767251015, + 0.0245346799492836, + 0.023040547966957092, + 0.023237209767103195, + 0.013722801581025124, + 0.018572352826595306, + 0.054977670311927795, + 0.09023943543434143, + 0.0059730582870543, + 0.005844390485435724, + 0.00646223546937108, + 0.012355776503682137, + 0.003957874607294798, + 0.002967468462884426, + 0.004880168475210667, + 0.004012236371636391, + 0.00197050953283906, + 0.002216863678768277, + 0.0024617912713438272, + 0.00543397618457675, + 0.001160166342742741, + 0.0009213133016601205, + 0.002034706063568592, + 0.001630689948797226, + 0.0024746751878410578, + 0.002698376541957259, + 0.002964657498523593, + 0.002146640559658408, + 0.0010516069596633315, + 0.0009165895753540099, + 0.001784616382792592, + 0.0014630912337452173, + 0.002410315442830324, + 0.0013351351954042912, + 0.0020931861363351345, + 0.0019853375852108, + 0.0010131308808922768, + 0.0011184284230694175, + 0.001764923450537026, + 0.001322676776908338, + 0.020185546949505806, + 0.006236450280994177, + 0.004981379024684429, + 0.008537803776562214, + 0.0071398233994841576, + 0.009711268357932568, + 0.007835193537175655, + 0.01613635942339897, + 0.0041598426178097725, + 0.006632709875702858, + 0.002839823253452778, + 0.00427401764318347, + 0.0046777124516665936, + 0.004379597492516041, + 0.004349514842033386, + 0.0031004564370959997, + 0.00470897788181901, + 0.002602756256237626, + 0.007987277582287788, + 0.008981425315141678, + 0.003216423327103257, + 0.007451345678418875, + 0.006613881327211857, + 0.009263996966183186, + 0.002391174202784896, + 0.002680085599422455, + 0.004095870535820723, + 0.005558257922530174, + 0.004210909828543663, + 0.00457763671875, + 0.004547282587736845, + 0.003144713817164302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.018219612538814545 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979462917894125 + ], + "opposite_logprob_tokens": [ + -4.63479471206665 + ], + "logprob_reward": -0.00979462917894125, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979462917894125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08627775311470032, + 0.5009918212890625, + 0.0333159938454628, + 0.017249921336770058, + 0.035730864852666855, + 0.078712597489357, + 0.0811448022723198, + 0.030789745971560478, + 0.0954594612121582, + 0.040327031165361404 + ], + "credit_map": [ + 0.020266743376851082, + 0.004754635039716959, + 0.004702017642557621, + 0.008763056248426437, + 0.004576936364173889, + 0.011052146553993225, + 0.0076670036651194096, + 0.024495216086506844, + 0.06937284022569656, + 0.0549757294356823, + 0.05149936303496361, + 0.08434348553419113, + 0.03275243565440178, + 0.043341729789972305, + 0.07001464068889618, + 0.09469159692525864, + 0.004749697633087635, + 0.003908318933099508, + 0.0046319495886564255, + 0.008824610151350498, + 0.0027695901226252317, + 0.002254936844110489, + 0.0033229393884539604, + 0.0028539514169096947, + 0.00196897448040545, + 0.002697208197787404, + 0.002843859139829874, + 0.0048509882763028145, + 0.0013100922806188464, + 0.00109459413215518, + 0.0010412696283310652, + 0.0014429343864321709, + 0.003998498897999525, + 0.004839596804231405, + 0.005779874511063099, + 0.006432790774852037, + 0.002380325458943844, + 0.0025643305853009224, + 0.006093906704336405, + 0.003641539951786399, + 0.009797239676117897, + 0.006221481133252382, + 0.007372443564236164, + 0.013429606333374977, + 0.008702711202204227, + 0.007654130924493074, + 0.014425057917833328, + 0.011109922081232071, + 0.015045166946947575, + 0.004776695743203163, + 0.0035666076000779867, + 0.01515284739434719, + 0.0035446910187602043, + 0.006225924473255873, + 0.025486627593636513, + 0.007346239406615496, + 0.003774181241169572, + 0.0017379710916429758, + 0.0015464745229110122, + 0.004519907757639885, + 0.0013989980798214674, + 0.0017085877479985356, + 0.009134650230407715, + 0.006968977861106396, + 0.010085384361445904, + 0.011264573782682419, + 0.008617859333753586, + 0.01120962854474783, + 0.009584271349012852, + 0.022649342194199562, + 0.009239246137440205, + 0.012809153646230698, + 0.0038527948781847954, + 0.0026614516973495483, + 0.004914798308163881, + 0.0062017617747187614, + 0.003156515769660473, + 0.00842239335179329, + 0.0057563260197639465, + 0.005360990762710571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.00979462917894125 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210343360900879 + ], + "opposite_logprob_tokens": [ + -4.264210224151611 + ], + "logprob_reward": -0.014210343360900879, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210343360900879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11555603891611099, + 0.3948194086551666, + 0.09544163942337036, + 0.04449765011668205, + 0.03809879347681999, + 0.0703587457537651, + 0.11244519054889679, + 0.06120621785521507, + 0.04476084187626839, + 0.022815469652414322 + ], + "credit_map": [ + 0.023911133408546448, + 0.007750560063868761, + 0.007693947292864323, + 0.009574010036885738, + 0.007038375828415155, + 0.016356047242879868, + 0.009221294894814491, + 0.03401067852973938, + 0.05978887155652046, + 0.041662152856588364, + 0.03956901654601097, + 0.05949804559350014, + 0.02609149180352688, + 0.031478822231292725, + 0.05668218061327934, + 0.0800488293170929, + 0.011722421273589134, + 0.011090085841715336, + 0.012347749434411526, + 0.027756107971072197, + 0.006679890677332878, + 0.0057379030622541904, + 0.011692970991134644, + 0.008414511568844318, + 0.005158044397830963, + 0.005618405062705278, + 0.006532006897032261, + 0.011379830539226532, + 0.0027575765270739794, + 0.0031222603283822536, + 0.005083993077278137, + 0.00484553724527359, + 0.004648192320019007, + 0.006083366926759481, + 0.006360309198498726, + 0.008630470372736454, + 0.002598219783976674, + 0.0024928066413849592, + 0.003677008207887411, + 0.003608424449339509, + 0.009867649525403976, + 0.006261792033910751, + 0.010372675955295563, + 0.008519730530679226, + 0.0045152995735406876, + 0.005479439161717892, + 0.013961350545287132, + 0.0113808149471879, + 0.00952176470309496, + 0.0036192648112773895, + 0.004513934720307589, + 0.023631563410162926, + 0.004596623592078686, + 0.0071015739813447, + 0.04108002036809921, + 0.018380453810095787, + 0.007288353517651558, + 0.004231323022395372, + 0.0027439489495009184, + 0.00853083934634924, + 0.003981000278145075, + 0.005055360496044159, + 0.016563761979341507, + 0.012811636552214622, + 0.004329717718064785, + 0.0037296877708286047, + 0.005484079010784626, + 0.006147339008748531, + 0.00399758480489254, + 0.01054470706731081, + 0.004793178290128708, + 0.005734550300985575, + 0.002149401931092143, + 0.0023637190461158752, + 0.0016666162991896272, + 0.0029705639462918043, + 0.002396235242486, + 0.006111934315413237, + 0.0021448838524520397, + 0.0030121176969259977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.014210343360900879 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, low-angle pan across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a vast colony of pillar corals with pale, ivory-colored tubular skeletons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single, focused sidelight sweeping steadily across the frame like a stage spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting create deep, sharply defined, inky shadows on the backlit coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny suspended marine particles visible shimmering in the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903444558382034 + ], + "opposite_logprob_tokens": [ + -2.4640345573425293 + ], + "logprob_reward": -0.08903444558382034, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903444558382034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32837045192718506, + 0.1466539055109024, + 0.10649939626455307, + 0.04272289574146271, + 0.027098244056105614, + 0.031466949731111526, + 0.160158172249794, + 0.0777151882648468, + 0.047805193811655045, + 0.03150969371199608 + ], + "credit_map": [ + 0.05805079638957977, + 0.033221688121557236, + 0.024235857650637627, + 0.07284557819366455, + 0.03870056942105293, + 0.039585795253515244, + 0.0396183617413044, + 0.022111790254712105, + 0.021499644964933395, + 0.018496399745345116, + 0.015528210438787937, + 0.020036058500409126, + 0.010895981453359127, + 0.012465630657970905, + 0.020328553393483162, + 0.027403434738516808, + 0.010503554716706276, + 0.013594383373856544, + 0.01687043160200119, + 0.02821282297372818, + 0.0076095787808299065, + 0.007044089026749134, + 0.012471050024032593, + 0.010193492285907269, + 0.004425130784511566, + 0.005239124875515699, + 0.005942027550190687, + 0.01256297342479229, + 0.002996105933561921, + 0.0023017532657831907, + 0.005959965288639069, + 0.003295816481113434, + 0.002790028927847743, + 0.0037198602221906185, + 0.004027115181088448, + 0.005080102942883968, + 0.0017407081322744489, + 0.0017411693697795272, + 0.0049153720028698444, + 0.0030838900711387396, + 0.003999096807092428, + 0.002232656581327319, + 0.0036287265829741955, + 0.008141634985804558, + 0.001627057557925582, + 0.0021170752588659525, + 0.005143697839230299, + 0.004577005282044411, + 0.037168949842453, + 0.010449061170220375, + 0.008861465379595757, + 0.02373751439154148, + 0.009564148262143135, + 0.020068610087037086, + 0.032902792096138, + 0.017405623570084572, + 0.00726304529234767, + 0.0043279449455440044, + 0.0060296570882201195, + 0.013851706869900227, + 0.005775750614702702, + 0.006796718575060368, + 0.020606527104973793, + 0.013063834980130196, + 0.0032113322522491217, + 0.006567961070686579, + 0.006272885948419571, + 0.010000077076256275, + 0.005036032758653164, + 0.005020794924348593, + 0.006946605164557695, + 0.004749504383653402, + 0.003523733699694276, + 0.003995127975940704, + 0.0037557538598775864, + 0.004406267777085304, + 0.0038297169376164675, + 0.005392066203057766, + 0.0030669516418129206, + 0.003540071891620755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.08903444558382034 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_0/index_tensor([2217, 2217])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_0/index_tensor([2217, 2217])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..99e199e6cb1e0acf86148195af48df2cc6ad8764 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_0/index_tensor([2217, 2217])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot smoothly pans to the right, following a strange procession through a minimalist, cartoon-like world composed of crisp colors and flat blocks of color. The centerpiece of the scene is a large, furry, red, car-shaped pillow, covered in soft, fluffy down, boasting exaggerated round white headlights and a smiling black grille. It sits solemnly on a simple wooden trailer. A group of short, cartoon-like figures in uniform gray overalls carefully escort it. Divided into two groups, some in the front tug at the trailer's ropes, while others push cautiously from behind. Their steps are synchronized and cautious, their faces bearing a serious, focused expression that stands out from the scene. The background is a flat, blocky palette of soft mint green ground and azure sky, dotted in the distance by a few geometric, dark green trees. The entire scene is bathed in bright, even, shadowless light. However, this cartoon purity is broken by a strong cinematic texture: the entire picture is full of obvious film grain, and details are deliberately sharpened, which gives a childlike scene a surreal, documentary-like solemnity and creates a sense of absurd and realistic drama.", + "seed": 24417, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_0/index_2217_seed_24417_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12659206436946988, + "vlm_logprob": -0.12659206436946988, + "vlm_yes_no_margin": 3.000000023841858, + "vlm_token_credit": -0.12659206436946988, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3504576086997986 + ], + "opposite_logprob_tokens": [ + -1.3504576683044434 + ], + "logprob_reward": -0.3504576086997986, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3504576086997986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06986196339130402, + 0.05706625059247017, + 0.07332836836576462, + 0.09645877033472061, + 0.1862909197807312, + 0.1235480085015297, + 0.06350017338991165, + 0.1328951120376587, + 0.07686993479728699, + 0.12018046528100967 + ], + "credit_map": [ + 0.012157924473285675, + 0.005994569510221481, + 0.018774298951029778, + 0.006838735193014145, + 0.005282629746943712, + 0.006366128101944923, + 0.006137475371360779, + 0.008310200646519661, + 0.005745161324739456, + 0.00528311263769865, + 0.012749651446938515, + 0.005917079281061888, + 0.004139542579650879, + 0.01048161182552576, + 0.00444509694352746, + 0.008304998278617859, + 0.011159912683069706, + 0.006933484226465225, + 0.020512379705905914, + 0.008667710237205029, + 0.007576039992272854, + 0.007847052998840809, + 0.005882729310542345, + 0.004749061074107885, + 0.015207706950604916, + 0.008550481870770454, + 0.017050815746188164, + 0.010787328705191612, + 0.00970951933413744, + 0.013704887591302395, + 0.009376456029713154, + 0.012071570381522179, + 0.03385547176003456, + 0.015583558939397335, + 0.031018082052469254, + 0.02236088179051876, + 0.016606265679001808, + 0.031383268535137177, + 0.014400792308151722, + 0.021082604303956032, + 0.012238908559083939, + 0.012573746964335442, + 0.013903842307627201, + 0.024721331894397736, + 0.013133916072547436, + 0.015385829843580723, + 0.012631287798285484, + 0.018959153443574905, + 0.005625519901514053, + 0.009303553961217403, + 0.009506048634648323, + 0.018715329468250275, + 0.004103207960724831, + 0.005930558778345585, + 0.005203897599130869, + 0.005112054757773876, + 0.015522472560405731, + 0.05122368037700653, + 0.008905651047825813, + 0.011052776128053665, + 0.027451014146208763, + 0.006302907131612301, + 0.007390435319393873, + 0.005046163685619831, + 0.005190833937376738, + 0.01628553494811058, + 0.005876053124666214, + 0.005494624376296997, + 0.012131121009588242, + 0.008089309558272362, + 0.011712473817169666, + 0.012089980766177177, + 0.013715495355427265, + 0.014921661466360092, + 0.010100645013153553, + 0.014837286435067654, + 0.01139888260513544, + 0.014733336865901947, + 0.016467463225126266, + 0.024005690589547157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.3504576086997986 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005966945551335812 + ], + "opposite_logprob_tokens": [ + -5.130967140197754 + ], + "logprob_reward": -0.005966945551335812, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005966945551335812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08936569094657898, + 0.053114887326955795, + 0.09243117272853851, + 0.07391884177923203, + 0.14675112068653107, + 0.039503224194049835, + 0.1210222914814949, + 0.20365020632743835, + 0.03803364560008049, + 0.14220893383026123 + ], + "credit_map": [ + 0.011544871143996716, + 0.009792966768145561, + 0.013267827220261097, + 0.011813769116997719, + 0.008237756788730621, + 0.012828920967876911, + 0.011483684182167053, + 0.010395892895758152, + 0.004514475353062153, + 0.005002198740839958, + 0.01139750424772501, + 0.005330284126102924, + 0.00616782670840621, + 0.008584704250097275, + 0.005856212228536606, + 0.006261684466153383, + 0.010592247359454632, + 0.015993798151612282, + 0.012868951074779034, + 0.011956948786973953, + 0.012240193784236908, + 0.010981948114931583, + 0.008180146105587482, + 0.009616944007575512, + 0.01053713820874691, + 0.0064434814266860485, + 0.01005291473120451, + 0.007209350820630789, + 0.007956658490002155, + 0.016813719645142555, + 0.006373210344463587, + 0.00853236485272646, + 0.02297385223209858, + 0.013836582191288471, + 0.027808744460344315, + 0.019279401749372482, + 0.014214077033102512, + 0.018531689420342445, + 0.013172438368201256, + 0.016934331506490707, + 0.005520277190953493, + 0.007211328949779272, + 0.00524187134578824, + 0.005103229079395533, + 0.0032248811330646276, + 0.005015225149691105, + 0.003456595353782177, + 0.004729813430458307, + 0.009839457459747791, + 0.018866879865527153, + 0.010442531667649746, + 0.05040140077471733, + 0.004302519839257002, + 0.014286712743341923, + 0.005581571254879236, + 0.0073012192733585835, + 0.03372837230563164, + 0.054592300206422806, + 0.03282706066966057, + 0.022866277024149895, + 0.032505035400390625, + 0.008395989425480366, + 0.01038403995335102, + 0.00835113413631916, + 0.0029907068237662315, + 0.004395685158669949, + 0.0053515383042395115, + 0.003671172773465514, + 0.004463277291506529, + 0.004357390105724335, + 0.006341405212879181, + 0.006462472956627607, + 0.012122420594096184, + 0.023974111303687096, + 0.01605932042002678, + 0.01634339801967144, + 0.01587720960378647, + 0.013178067281842232, + 0.02441249042749405, + 0.020241910591721535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 9 + ], + "selected_reward": -0.005966945551335812 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278928935527802 + ], + "opposite_logprob_tokens": [ + -2.017789363861084 + ], + "logprob_reward": -0.14278928935527802, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14278928935527802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07871948927640915, + 0.07893785834312439, + 0.1250840723514557, + 0.08564320206642151, + 0.06047987565398216, + 0.10198775678873062, + 0.06920354813337326, + 0.16128164529800415, + 0.1295507252216339, + 0.10911184549331665 + ], + "credit_map": [ + 0.009498986415565014, + 0.007148373872041702, + 0.021404223516583443, + 0.006279837340116501, + 0.005923965014517307, + 0.010955866426229477, + 0.007496469188481569, + 0.010011769831180573, + 0.00979303102940321, + 0.006255785934627056, + 0.019793163985013962, + 0.006317357532680035, + 0.004086766391992569, + 0.011892598122358322, + 0.005984507501125336, + 0.014814650639891624, + 0.018623685464262962, + 0.020139694213867188, + 0.029975377023220062, + 0.016611745581030846, + 0.0084224883466959, + 0.014170579612255096, + 0.008507921360433102, + 0.008632580749690533, + 0.012202814221382141, + 0.008395954966545105, + 0.011887452565133572, + 0.008205168880522251, + 0.007234553340822458, + 0.02443726919591427, + 0.005171859636902809, + 0.008108130656182766, + 0.007477949373424053, + 0.005579682998359203, + 0.013464957475662231, + 0.008888169191777706, + 0.006067944224923849, + 0.007106178440153599, + 0.004422725643962622, + 0.0074722724966704845, + 0.01518268883228302, + 0.008023816160857677, + 0.009845140390098095, + 0.023573366925120354, + 0.010366523638367653, + 0.01221891026943922, + 0.009187715128064156, + 0.013589601032435894, + 0.0060256477445364, + 0.008036755956709385, + 0.009348686784505844, + 0.007941482588648796, + 0.006150500383228064, + 0.019912397488951683, + 0.00395891722291708, + 0.007829157635569572, + 0.023601572960615158, + 0.01768341101706028, + 0.042719852179288864, + 0.029848597943782806, + 0.02291751280426979, + 0.0073441024869680405, + 0.010079624131321907, + 0.007086966652423143, + 0.02327408269047737, + 0.014853653497993946, + 0.030904358252882957, + 0.01987074315547943, + 0.013769838027656078, + 0.006549933925271034, + 0.013726728968322277, + 0.006601389963179827, + 0.018951566889882088, + 0.016210392117500305, + 0.014456536620855331, + 0.013979549519717693, + 0.012153075076639652, + 0.010222283191978931, + 0.012489346787333488, + 0.01064910739660263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.14278928935527802 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00672908453270793 + ], + "opposite_logprob_tokens": [ + -5.0067291259765625 + ], + "logprob_reward": -0.00672908453270793, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00672908453270793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14696942269802094, + 0.1242297813296318, + 0.05340287834405899, + 0.05580815300345421, + 0.07543390244245529, + 0.2081868201494217, + 0.05514947324991226, + 0.11964070051908493, + 0.04017835110425949, + 0.1210005134344101 + ], + "credit_map": [ + 0.017500920221209526, + 0.013330555520951748, + 0.03038964420557022, + 0.019061630591750145, + 0.01344610657542944, + 0.019743185490369797, + 0.0186707004904747, + 0.014826679602265358, + 0.024635687470436096, + 0.00852554477751255, + 0.02563312277197838, + 0.01080418936908245, + 0.008302350528538227, + 0.01846504583954811, + 0.008335267193615437, + 0.019528573378920555, + 0.008483001962304115, + 0.007135991007089615, + 0.012209633365273476, + 0.005366017110645771, + 0.00515896687284112, + 0.005663286428898573, + 0.003384547308087349, + 0.0060014380142092705, + 0.007631014566868544, + 0.0062479679472744465, + 0.009117412380874157, + 0.007768212351948023, + 0.00479807797819376, + 0.012087608687579632, + 0.004048084374517202, + 0.004109776113182306, + 0.01430024579167366, + 0.007570503279566765, + 0.017867719754576683, + 0.009423153474926949, + 0.006537555251270533, + 0.007003343664109707, + 0.005830650683492422, + 0.006900731474161148, + 0.033385369926691055, + 0.04209516569972038, + 0.03493906930088997, + 0.04038603976368904, + 0.014015568420290947, + 0.015589455142617226, + 0.011270694434642792, + 0.016505466774106026, + 0.005406644660979509, + 0.005753643345087767, + 0.006012648809701204, + 0.0172922033816576, + 0.0034182157833129168, + 0.00796264223754406, + 0.0039032769855111837, + 0.005400201305747032, + 0.026339847594499588, + 0.027527693659067154, + 0.020132578909397125, + 0.013345261104404926, + 0.01688935048878193, + 0.004738595336675644, + 0.005628885235637426, + 0.0050384788773953915, + 0.004998513963073492, + 0.006234612315893173, + 0.0058425310999155045, + 0.005839718505740166, + 0.0076067522168159485, + 0.0025739329867064953, + 0.004335979465395212, + 0.0027463098522275686, + 0.010711152106523514, + 0.02166658453643322, + 0.012564270757138729, + 0.018878433853387833, + 0.012684259563684464, + 0.011503361165523529, + 0.016033772379159927, + 0.016958674415946007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.00672908453270793 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12701739370822906 + ], + "opposite_logprob_tokens": [ + -2.1270174980163574 + ], + "logprob_reward": -0.12701739370822906, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12701739370822906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20242083072662354, + 0.12299894541501999, + 0.08290692418813705, + 0.04175454378128052, + 0.06485626101493835, + 0.12988820672035217, + 0.04018000513315201, + 0.10018378496170044, + 0.0939527079463005, + 0.12085775285959244 + ], + "credit_map": [ + 0.02865312620997429, + 0.017520952969789505, + 0.03370266407728195, + 0.030874431133270264, + 0.016132818534970284, + 0.021014025434851646, + 0.025582557544112206, + 0.02894025854766369, + 0.020896749570965767, + 0.008104781620204449, + 0.031159598380327225, + 0.012061701156198978, + 0.006745871622115374, + 0.014349168166518211, + 0.012091076001524925, + 0.017589999362826347, + 0.014822187833487988, + 0.009414365515112877, + 0.01610252633690834, + 0.00886616576462984, + 0.006060218438506126, + 0.011257441714406013, + 0.005639392416924238, + 0.010744618251919746, + 0.006370895076543093, + 0.005581763572990894, + 0.007044087164103985, + 0.006554188206791878, + 0.0038702762685716152, + 0.005779155530035496, + 0.003341064555570483, + 0.0032131150364875793, + 0.010314646176993847, + 0.006984819658100605, + 0.015994764864444733, + 0.008598881773650646, + 0.0051860446110367775, + 0.006835949141532183, + 0.004946662113070488, + 0.005994485225528479, + 0.018420159816741943, + 0.011334685608744621, + 0.016850892454385757, + 0.034218404442071915, + 0.009774673730134964, + 0.012491505593061447, + 0.009996138513088226, + 0.016801737248897552, + 0.0035865092650055885, + 0.0057626692578196526, + 0.005023977719247341, + 0.006085178814828396, + 0.004224738571792841, + 0.009956317022442818, + 0.00231127068400383, + 0.0032293477561324835, + 0.024600287899374962, + 0.027215229347348213, + 0.016138626262545586, + 0.009323367848992348, + 0.010351045057177544, + 0.0034773466177284718, + 0.004092409275472164, + 0.004985472187399864, + 0.008277708664536476, + 0.027828164398670197, + 0.008731952868402004, + 0.007505316752940416, + 0.015850642696022987, + 0.0072676194831728935, + 0.010239868424832821, + 0.008251434192061424, + 0.012497445568442345, + 0.01655571348965168, + 0.014341660775244236, + 0.01751473732292652, + 0.012330603785812855, + 0.012173479422926903, + 0.014873053878545761, + 0.020571062341332436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.12701739370822906 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot smoothly pans to the right, following a strange procession through a minimalist, cartoon-like world composed of crisp colors and flat blocks of color. The centerpiece of the scene is a large, furry, red, car-shaped pillow, covered in soft, fluffy down, boasting exaggerated round white headlights and a smiling black grille. It sits solemnly on a simple wooden trailer. A group of short, cartoon-like figures in uniform gray overalls carefully escort it. Divided into two groups, some in the front tug at the trailer's ropes, while others push cautiously from behind. Their steps are synchronized and cautious, their faces bearing a serious, focused expression that stands out from the scene. The background is a flat, blocky palette of soft mint green ground and azure sky, dotted in the distance by a few geometric, dark green trees. The entire scene is bathed in bright, even, shadowless light. However, this cartoon purity is broken by a strong cinematic texture: the entire picture is full of obvious film grain, and details are deliberately sharpened, which gives a childlike scene a surreal, documentary-like solemnity and creates a sense of absurd and realistic drama.", + "seed": 24418, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_0/index_2217_seed_24418_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09250000822357833, + "vlm_logprob": -0.09250000822357833, + "vlm_yes_no_margin": 3.350000023841858, + "vlm_token_credit": -0.09250000822357833, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13096968829631805 + ], + "opposite_logprob_tokens": [ + -2.505969762802124 + ], + "logprob_reward": -0.13096968829631805, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.13096968829631805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08564969152212143, + 0.10367735475301743, + 0.10883304476737976, + 0.18835420906543732, + 0.10035651177167892, + 0.03692450746893883, + 0.04318138211965561, + 0.12527520954608917, + 0.15446794033050537, + 0.053280144929885864 + ], + "credit_map": [ + 0.008028583601117134, + 0.009226915426552296, + 0.01221379078924656, + 0.009879227727651596, + 0.006940394174307585, + 0.019574828445911407, + 0.010042661800980568, + 0.009743292815983295, + 0.009782386012375355, + 0.013136926107108593, + 0.013672304339706898, + 0.015426685102283955, + 0.01803787238895893, + 0.007072982843965292, + 0.0159913320094347, + 0.010556870140135288, + 0.017345169559121132, + 0.007821005769073963, + 0.02405427396297455, + 0.009534723125398159, + 0.007647290825843811, + 0.013692423701286316, + 0.014271005056798458, + 0.014467163011431694, + 0.02607366256415844, + 0.01405902486294508, + 0.02859244868159294, + 0.02263428084552288, + 0.029110781848430634, + 0.019681746140122414, + 0.021053817123174667, + 0.02714846469461918, + 0.011514219455420971, + 0.008332621306180954, + 0.01154144573956728, + 0.011782667599618435, + 0.010200693272054195, + 0.021127328276634216, + 0.014454527758061886, + 0.011403013952076435, + 0.007066982798278332, + 0.004015316721051931, + 0.004761924967169762, + 0.003842829028144479, + 0.005698558408766985, + 0.0023087402805685997, + 0.003930121660232544, + 0.0053000375628471375, + 0.00688660703599453, + 0.004535196349024773, + 0.007867501117289066, + 0.004213481210172176, + 0.006605666130781174, + 0.0034076236188411713, + 0.004728973843157291, + 0.0049363370053470135, + 0.01923844777047634, + 0.007208374794572592, + 0.031581539660692215, + 0.009477769955992699, + 0.014866126701235771, + 0.015320107340812683, + 0.012203141115605831, + 0.015379715710878372, + 0.009851840324699879, + 0.00868266448378563, + 0.0103041036054492, + 0.027435993775725365, + 0.02420745976269245, + 0.02685888484120369, + 0.02422151528298855, + 0.02290547452867031, + 0.006249225698411465, + 0.003726128488779068, + 0.008376560173928738, + 0.005720140878111124, + 0.011375103145837784, + 0.007496459875255823, + 0.006244034972041845, + 0.0040924944914877415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.13096968829631805 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005269922316074371 + ], + "opposite_logprob_tokens": [ + -5.255270004272461 + ], + "logprob_reward": -0.005269922316074371, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005269922316074371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07627388834953308, + 0.05716315284371376, + 0.06457160413265228, + 0.039945170283317566, + 0.06338359415531158, + 0.40598154067993164, + 0.0800655260682106, + 0.10960999876260757, + 0.04189499467611313, + 0.061110492795705795 + ], + "credit_map": [ + 0.007574086543172598, + 0.006533975247293711, + 0.01886589080095291, + 0.008055332116782665, + 0.005906611680984497, + 0.015373533591628075, + 0.004564423114061356, + 0.009400034323334694, + 0.005666179116815329, + 0.006726243533194065, + 0.008416758850216866, + 0.00917838979512453, + 0.008442716673016548, + 0.00468596164137125, + 0.0073045906610786915, + 0.006742317695170641, + 0.007007506210356951, + 0.005426493473351002, + 0.014686164446175098, + 0.006374460645020008, + 0.007918423973023891, + 0.008487091399729252, + 0.006784421391785145, + 0.007887046784162521, + 0.006982875522226095, + 0.0026900023221969604, + 0.007115358952432871, + 0.0033894984517246485, + 0.005186507944017649, + 0.004337199963629246, + 0.00496729975566268, + 0.005276429001241922, + 0.006597800645977259, + 0.005674542393535376, + 0.010641721077263355, + 0.008787843398749828, + 0.009236391633749008, + 0.006742159370332956, + 0.008504783734679222, + 0.007198350038379431, + 0.07352237403392792, + 0.027483057230710983, + 0.04970605671405792, + 0.03331765905022621, + 0.04478323087096214, + 0.026875970885157585, + 0.048020876944065094, + 0.10227233171463013, + 0.006745156366378069, + 0.006690447684377432, + 0.012060754001140594, + 0.00789770856499672, + 0.01272246241569519, + 0.010942477732896805, + 0.009523050859570503, + 0.013483481481671333, + 0.013866129331290722, + 0.007655267138034105, + 0.022140562534332275, + 0.010629489086568356, + 0.015308916568756104, + 0.012976025231182575, + 0.011272056959569454, + 0.015761561691761017, + 0.0034078077878803015, + 0.0024195832666009665, + 0.004787967074662447, + 0.0063638207502663136, + 0.00596275320276618, + 0.007253262680023909, + 0.004761532414704561, + 0.0069382693618535995, + 0.005541122052818537, + 0.004827081225812435, + 0.013589895330369473, + 0.006637170910835266, + 0.009911620989441872, + 0.007990407757461071, + 0.0064169797115027905, + 0.006196221336722374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.005269922316074371 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279238879680634 + ], + "opposite_logprob_tokens": [ + -2.0177924633026123 + ], + "logprob_reward": -0.14279238879680634, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14279238879680634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19335104525089264, + 0.06167350709438324, + 0.12907984852790833, + 0.031003186479210854, + 0.06426363438367844, + 0.11845635622739792, + 0.07486411929130554, + 0.14583934843540192, + 0.07118166238069534, + 0.11028726398944855 + ], + "credit_map": [ + 0.01736748032271862, + 0.009318873286247253, + 0.04757656529545784, + 0.01566731557250023, + 0.009953035041689873, + 0.05225900188088417, + 0.014813103713095188, + 0.026395652443170547, + 0.005606988910585642, + 0.0047936649061739445, + 0.010966610163450241, + 0.006304894108325243, + 0.008844394236803055, + 0.012178486213088036, + 0.006135961040854454, + 0.006842504255473614, + 0.017894592136144638, + 0.007879171520471573, + 0.03480961546301842, + 0.012857023626565933, + 0.009901697747409344, + 0.017411988228559494, + 0.00886663980782032, + 0.019459104165434837, + 0.003858278039842844, + 0.0019687707535922527, + 0.005166955757886171, + 0.0037775482051074505, + 0.003259007353335619, + 0.0032154887448996305, + 0.005444331094622612, + 0.004312805365771055, + 0.009207663126289845, + 0.005915283691138029, + 0.009198424406349659, + 0.004807738121598959, + 0.004614964593201876, + 0.018152551725506783, + 0.0062299068085849285, + 0.006137096788734198, + 0.0085478900000453, + 0.008629966527223587, + 0.01973532885313034, + 0.009612295776605606, + 0.013955028727650642, + 0.008518083021044731, + 0.012926710769534111, + 0.03653104230761528, + 0.009332709945738316, + 0.008039535954594612, + 0.016386853531003, + 0.006070703733712435, + 0.010255048982799053, + 0.007028382737189531, + 0.00824693776667118, + 0.009503945708274841, + 0.02016771398484707, + 0.011981655843555927, + 0.034378957003355026, + 0.012830707244575024, + 0.019366594031453133, + 0.016620395705103874, + 0.010882020927965641, + 0.019611302763223648, + 0.008798387832939625, + 0.004525362513959408, + 0.014461805112659931, + 0.0120451170951128, + 0.008222038857638836, + 0.007161908317357302, + 0.008150330744683743, + 0.007816702127456665, + 0.015455537475645542, + 0.009562915191054344, + 0.025554930791258812, + 0.01135467179119587, + 0.01302392315119505, + 0.017005780711770058, + 0.008560674265027046, + 0.009768824093043804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.14279238879680634 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003611353924497962 + ], + "opposite_logprob_tokens": [ + -5.6286115646362305 + ], + "logprob_reward": -0.003611353924497962, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003611353924497962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08078724145889282, + 0.123049795627594, + 0.27781426906585693, + 0.125127375125885, + 0.029561998322606087, + 0.07106959819793701, + 0.045376311987638474, + 0.06828607618808746, + 0.07765671610832214, + 0.1012706309556961 + ], + "credit_map": [ + 0.008146071806550026, + 0.005791402887552977, + 0.011880448088049889, + 0.008876499719917774, + 0.005350449122488499, + 0.02375604212284088, + 0.007008098531514406, + 0.009978225454688072, + 0.013694027438759804, + 0.011633369140326977, + 0.02045155130326748, + 0.013961437158286572, + 0.01515890471637249, + 0.017748575657606125, + 0.015823910012841225, + 0.014578014612197876, + 0.04911178722977638, + 0.019859904423356056, + 0.07210630923509598, + 0.016440700739622116, + 0.017088742926716805, + 0.04205064848065376, + 0.022032244130969048, + 0.03912391513586044, + 0.015016831457614899, + 0.010374120436608791, + 0.0202906783670187, + 0.012361126951873302, + 0.009536756202578545, + 0.016202392056584358, + 0.013981609605252743, + 0.027363844215869904, + 0.0025162240490317345, + 0.0039292387664318085, + 0.0038700783625245094, + 0.0034219371154904366, + 0.0032545498106628656, + 0.0023156884126365185, + 0.005632712040096521, + 0.004621568135917187, + 0.011802498251199722, + 0.0060812123119831085, + 0.007066499907523394, + 0.007451421581208706, + 0.014848434366285801, + 0.005061794072389603, + 0.01112104021012783, + 0.00763669004663825, + 0.0061379410326480865, + 0.00403241952881217, + 0.008848677389323711, + 0.002993215573951602, + 0.006352314725518227, + 0.0053626541048288345, + 0.004459911957383156, + 0.007189176045358181, + 0.008714580908417702, + 0.003780737053602934, + 0.013095364905893803, + 0.005714050959795713, + 0.009459474124014378, + 0.011397134512662888, + 0.007026491221040487, + 0.009098239243030548, + 0.0054456936195492744, + 0.005945436656475067, + 0.01055340189486742, + 0.012900984846055508, + 0.010615699924528599, + 0.013767854310572147, + 0.009077556431293488, + 0.00935009028762579, + 0.007516829762607813, + 0.007789299823343754, + 0.021979114040732384, + 0.011197155341506004, + 0.01803179271519184, + 0.013773677870631218, + 0.011560827493667603, + 0.009421930648386478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.003611353924497962 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17985668778419495 + ], + "opposite_logprob_tokens": [ + -1.8048566579818726 + ], + "logprob_reward": -0.17985668778419495, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17985668778419495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13927069306373596, + 0.05855927988886833, + 0.11034203320741653, + 0.0865846574306488, + 0.042109351605176926, + 0.15526707470417023, + 0.050201416015625, + 0.12447823584079742, + 0.07796499878168106, + 0.15522229671478271 + ], + "credit_map": [ + 0.013973773457109928, + 0.01026216708123684, + 0.019471002742648125, + 0.01877709850668907, + 0.00995450560003519, + 0.03315245732665062, + 0.0170749444514513, + 0.016604743897914886, + 0.006762257777154446, + 0.007850995287299156, + 0.006336061749607325, + 0.007578077260404825, + 0.01023704931139946, + 0.0046027167700231075, + 0.009311249479651451, + 0.0058808717876672745, + 0.016260020434856415, + 0.008832288905978203, + 0.02565533109009266, + 0.008587264455854893, + 0.01132582500576973, + 0.015909232199192047, + 0.00991186685860157, + 0.013860204257071018, + 0.010905570350587368, + 0.004749104380607605, + 0.017111850902438164, + 0.008719823323190212, + 0.008622124791145325, + 0.0101102814078331, + 0.008916120044887066, + 0.01744978316128254, + 0.0049804276786744595, + 0.0025612812023609877, + 0.009425841271877289, + 0.0057061887346208096, + 0.0038767503574490547, + 0.0049082254990935326, + 0.005538965109735727, + 0.005111671984195709, + 0.02182399481534958, + 0.013499938882887363, + 0.015935903415083885, + 0.018438376486301422, + 0.03340408578515053, + 0.010467159561812878, + 0.02594771608710289, + 0.015749892219901085, + 0.003879316383972764, + 0.004770167637616396, + 0.009816682897508144, + 0.0033392361365258694, + 0.006552744656801224, + 0.007729844655841589, + 0.00608564680442214, + 0.008027777075767517, + 0.01663779281079769, + 0.007095714099705219, + 0.027817610651254654, + 0.008410973474383354, + 0.01952400989830494, + 0.016449518501758575, + 0.008863861672580242, + 0.019678756594657898, + 0.006830753292888403, + 0.006049400195479393, + 0.015048381872475147, + 0.012434075586497784, + 0.008772044442594051, + 0.008280014619231224, + 0.009496322833001614, + 0.011054006405174732, + 0.022258352488279343, + 0.0107297133654356, + 0.027117280289530754, + 0.01463721040636301, + 0.030415929853916168, + 0.02164996974170208, + 0.014647397212684155, + 0.013766449876129627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.17985668778419495 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot smoothly pans to the right, following a strange procession through a minimalist, cartoon-like world composed of crisp colors and flat blocks of color. The centerpiece of the scene is a large, furry, red, car-shaped pillow, covered in soft, fluffy down, boasting exaggerated round white headlights and a smiling black grille. It sits solemnly on a simple wooden trailer. A group of short, cartoon-like figures in uniform gray overalls carefully escort it. Divided into two groups, some in the front tug at the trailer's ropes, while others push cautiously from behind. Their steps are synchronized and cautious, their faces bearing a serious, focused expression that stands out from the scene. The background is a flat, blocky palette of soft mint green ground and azure sky, dotted in the distance by a few geometric, dark green trees. The entire scene is bathed in bright, even, shadowless light. However, this cartoon purity is broken by a strong cinematic texture: the entire picture is full of obvious film grain, and details are deliberately sharpened, which gives a childlike scene a surreal, documentary-like solemnity and creates a sense of absurd and realistic drama.", + "seed": 24419, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_0/index_2217_seed_24419_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07531000678427517, + "vlm_logprob": -0.07531000678427517, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.07531000678427517, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15575531125068665 + ], + "opposite_logprob_tokens": [ + -2.1557552814483643 + ], + "logprob_reward": -0.15575531125068665, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.15575531125068665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1467025876045227, + 0.03431686386466026, + 0.04651854187250137, + 0.029863758012652397, + 0.06525716185569763, + 0.2597372829914093, + 0.19994977116584778, + 0.06846166402101517, + 0.06749904155731201, + 0.08169331401586533 + ], + "credit_map": [ + 0.010117098689079285, + 0.014600291848182678, + 0.01907919906079769, + 0.039836786687374115, + 0.008857919834554195, + 0.014444866217672825, + 0.023630840703845024, + 0.016135595738887787, + 0.006102010142058134, + 0.0027086532209068537, + 0.00424975948408246, + 0.003224191488698125, + 0.0024118106812238693, + 0.008227553218603134, + 0.00262002763338387, + 0.004772859159857035, + 0.009574334137141705, + 0.0053663672879338264, + 0.007548265624791384, + 0.004750723950564861, + 0.0032247519120573997, + 0.006429246626794338, + 0.004029499366879463, + 0.005595354363322258, + 0.004443826619535685, + 0.0017337027238681912, + 0.00608309218659997, + 0.0030489687342196703, + 0.002256000181660056, + 0.0031665749847888947, + 0.0023261436726897955, + 0.00680545112118125, + 0.006581559311598539, + 0.004602978937327862, + 0.004958556033670902, + 0.01798211969435215, + 0.0034055537544190884, + 0.00886605679988861, + 0.007730776444077492, + 0.011129563674330711, + 0.035776395350694656, + 0.018924303352832794, + 0.0357501283288002, + 0.035646043717861176, + 0.02183467335999012, + 0.04063895717263222, + 0.02538507618010044, + 0.04578172788023949, + 0.01628737896680832, + 0.017951836809515953, + 0.020153803750872612, + 0.0804515853524208, + 0.01066290121525526, + 0.010047264397144318, + 0.02872338891029358, + 0.0156716275960207, + 0.014311346225440502, + 0.004961913451552391, + 0.009767208248376846, + 0.006978919263929129, + 0.006275298539549112, + 0.006464752368628979, + 0.006044128909707069, + 0.013658098876476288, + 0.007347624283283949, + 0.008521948009729385, + 0.013551045209169388, + 0.012557710520923138, + 0.005089940503239632, + 0.007246536202728748, + 0.0066697681322693825, + 0.006514477543532848, + 0.009271165356040001, + 0.010088604874908924, + 0.013749237172305584, + 0.012696396559476852, + 0.004972417373210192, + 0.007317439187318087, + 0.011063472367823124, + 0.01253458671271801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.15575531125068665 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003630002262070775 + ], + "opposite_logprob_tokens": [ + -5.6286301612854 + ], + "logprob_reward": -0.003630002262070775, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003630002262070775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13421598076820374, + 0.05566303804516792, + 0.045516714453697205, + 0.043493956327438354, + 0.06495318561792374, + 0.18238386511802673, + 0.308800607919693, + 0.03645423427224159, + 0.030292581766843796, + 0.09822580963373184 + ], + "credit_map": [ + 0.01431324053555727, + 0.019802389666438103, + 0.017116259783506393, + 0.016723785549402237, + 0.012292008846998215, + 0.0140739930793643, + 0.020979518070816994, + 0.018914805725216866, + 0.008426394313573837, + 0.007049601059406996, + 0.005607775412499905, + 0.008792410604655743, + 0.0057618808932602406, + 0.006260150112211704, + 0.008298260159790516, + 0.0054665678180754185, + 0.008496012538671494, + 0.00409564608708024, + 0.007143769413232803, + 0.0039049212355166674, + 0.0033636861480772495, + 0.005397868808358908, + 0.004692517686635256, + 0.008422299288213253, + 0.005671761929988861, + 0.0026204593013972044, + 0.007836404256522655, + 0.005397515837103128, + 0.002526280004531145, + 0.004532155115157366, + 0.003976901993155479, + 0.010932485572993755, + 0.011524355039000511, + 0.004819037392735481, + 0.007545784581452608, + 0.0173178743571043, + 0.004011035431176424, + 0.005038764793425798, + 0.008772151544690132, + 0.005924183409661055, + 0.023659514263272285, + 0.01451873779296875, + 0.021415628492832184, + 0.027544094249606133, + 0.020884929224848747, + 0.02472301758825779, + 0.020708410069346428, + 0.02892955020070076, + 0.023222092539072037, + 0.025566453114151955, + 0.025269264355301857, + 0.13383953273296356, + 0.013880953192710876, + 0.009290283545851707, + 0.047653838992118835, + 0.03007819876074791, + 0.005169987678527832, + 0.004001345951110125, + 0.005769767332822084, + 0.008626515977084637, + 0.0025937892496585846, + 0.0027922235894948244, + 0.004027660470455885, + 0.00347294588573277, + 0.00393905583769083, + 0.0032735210843384266, + 0.004737212788313627, + 0.006886569317430258, + 0.0022005874197930098, + 0.003214093390852213, + 0.003471642266958952, + 0.002569902455434203, + 0.012618739157915115, + 0.011506089009344578, + 0.009527689777314663, + 0.01324863638728857, + 0.012870303355157375, + 0.014674751088023186, + 0.010968908667564392, + 0.012810704298317432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.003630002262070775 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368818759918213 + ], + "opposite_logprob_tokens": [ + -3.4086880683898926 + ], + "logprob_reward": -0.03368818759918213, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368818759918213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30063629150390625, + 0.03811148926615715, + 0.04034585505723953, + 0.03230723738670349, + 0.06987186521291733, + 0.09411495923995972, + 0.17156320810317993, + 0.0518176443874836, + 0.10009455680847168, + 0.10113697499036789 + ], + "credit_map": [ + 0.02355840615928173, + 0.029445286840200424, + 0.04361314699053764, + 0.08409786969423294, + 0.017933769151568413, + 0.027997413650155067, + 0.04578980430960655, + 0.028200572356581688, + 0.006846252828836441, + 0.0021230827551335096, + 0.0031181538943201303, + 0.0020799534395337105, + 0.002755108056589961, + 0.009986277669668198, + 0.002547177253291011, + 0.008655477315187454, + 0.006949419621378183, + 0.0029402514919638634, + 0.008381097577512264, + 0.0037151603028178215, + 0.002385882195085287, + 0.0044470583088696, + 0.0034252821933478117, + 0.008101697079837322, + 0.004369218833744526, + 0.0029407087713479996, + 0.006390493828803301, + 0.003652924671769142, + 0.001720789005048573, + 0.0034507436212152243, + 0.0022116880863904953, + 0.007570668589323759, + 0.011098450981080532, + 0.0057199131697416306, + 0.008157113566994667, + 0.0191358495503664, + 0.003678401466459036, + 0.006257441360503435, + 0.009401874616742134, + 0.006422810722142458, + 0.012082604691386223, + 0.00719272019341588, + 0.011996585875749588, + 0.013972719199955463, + 0.010883006267249584, + 0.012967292219400406, + 0.01001566369086504, + 0.015004361048340797, + 0.01165208499878645, + 0.013624469749629498, + 0.012873128987848759, + 0.07301969081163406, + 0.007837309502065182, + 0.007089763414114714, + 0.026004675775766373, + 0.019462065771222115, + 0.006675089709460735, + 0.004781001713126898, + 0.009139005094766617, + 0.010530045256018639, + 0.0026426727417856455, + 0.004462042357772589, + 0.004914406221359968, + 0.00867337454110384, + 0.009788076393306255, + 0.012129154987633228, + 0.017801083624362946, + 0.023206843063235283, + 0.006926055531948805, + 0.00803208164870739, + 0.009902717545628548, + 0.012308529578149319, + 0.011924540624022484, + 0.014694947749376297, + 0.016423676162958145, + 0.017696211114525795, + 0.007216045167297125, + 0.007488321978598833, + 0.014258217997848988, + 0.011435002088546753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.03368818759918213 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036120666190981865 + ], + "opposite_logprob_tokens": [ + -5.628612041473389 + ], + "logprob_reward": -0.0036120666190981865, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036120666190981865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12510254979133606, + 0.09504097700119019, + 0.07043562084436417, + 0.12110529839992523, + 0.12411854416131973, + 0.14099860191345215, + 0.1407892256975174, + 0.07315735518932343, + 0.058837540447711945, + 0.05041433870792389 + ], + "credit_map": [ + 0.012043586000800133, + 0.01805517077445984, + 0.015849586576223373, + 0.02336297556757927, + 0.012360690161585808, + 0.012488218955695629, + 0.013549714349210262, + 0.017392609268426895, + 0.014447547495365143, + 0.007144161034375429, + 0.010769863612949848, + 0.009210777468979359, + 0.007169329561293125, + 0.024324538186192513, + 0.0069937980733811855, + 0.01498096901923418, + 0.014006773941218853, + 0.006851941347122192, + 0.011708826757967472, + 0.007079699542373419, + 0.00409951526671648, + 0.008775562979280949, + 0.004954550415277481, + 0.012958750128746033, + 0.014450978487730026, + 0.0064236922189593315, + 0.028520923107862473, + 0.012647837400436401, + 0.005923416465520859, + 0.01412010658532381, + 0.008254826068878174, + 0.030763516202569008, + 0.015635138377547264, + 0.007934750989079475, + 0.011522004380822182, + 0.025165311992168427, + 0.009443357586860657, + 0.02361653372645378, + 0.011691083200275898, + 0.019110355526208878, + 0.020474741235375404, + 0.010850896127521992, + 0.017755083739757538, + 0.019629748538136482, + 0.012113286182284355, + 0.017557108774781227, + 0.01542729139328003, + 0.027190450578927994, + 0.01107260026037693, + 0.011478791944682598, + 0.012649678625166416, + 0.04970970004796982, + 0.013684917241334915, + 0.011153947561979294, + 0.020055843517184258, + 0.01098374743014574, + 0.007762513123452663, + 0.006295542232692242, + 0.009816357865929604, + 0.013893979601562023, + 0.005411974620074034, + 0.007371590472757816, + 0.009009693749248981, + 0.013595701195299625, + 0.007292129565030336, + 0.006269827019423246, + 0.011276799254119396, + 0.010938000865280628, + 0.004434039816260338, + 0.0062222592532634735, + 0.006360058672726154, + 0.006044422276318073, + 0.006822721101343632, + 0.006103893741965294, + 0.00751694617792964, + 0.00958639569580555, + 0.0037340745329856873, + 0.003474139142781496, + 0.007358690723776817, + 0.005817476660013199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.0036120666190981865 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986446619033813 + ], + "opposite_logprob_tokens": [ + -1.8048644065856934 + ], + "logprob_reward": -0.17986446619033813, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986446619033813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5895610451698303, + 0.05304503068327904, + 0.0232548788189888, + 0.01949751377105713, + 0.061661671847105026, + 0.048364825546741486, + 0.1290850192308426, + 0.03284322842955589, + 0.01275121234357357, + 0.029935605823993683 + ], + "credit_map": [ + 0.0476367324590683, + 0.06242264807224274, + 0.0792396292090416, + 0.1687784194946289, + 0.035953398793935776, + 0.054507434368133545, + 0.08704488724470139, + 0.05397787317633629, + 0.008975966833531857, + 0.003055891487747431, + 0.0056166816502809525, + 0.0036450650077313185, + 0.003574699629098177, + 0.011859014630317688, + 0.004725425038486719, + 0.011592286638915539, + 0.004376159515231848, + 0.002129045082256198, + 0.004194121807813644, + 0.0024414958897978067, + 0.0016711867647245526, + 0.0024142011534422636, + 0.0019889110699295998, + 0.004039756488054991, + 0.0025204834528267384, + 0.001912106410600245, + 0.0039400365203619, + 0.0025533316656947136, + 0.0016432998236268759, + 0.0021216163877397776, + 0.0015639991033822298, + 0.00324264052323997, + 0.007260131649672985, + 0.004784103948622942, + 0.004962956067174673, + 0.02059377171099186, + 0.003504832973703742, + 0.00385666242800653, + 0.008573745377361774, + 0.008125467225909233, + 0.005504310596734285, + 0.0035577332600951195, + 0.00817176140844822, + 0.006925017107278109, + 0.004722492769360542, + 0.005287963896989822, + 0.005630235653370619, + 0.008565310388803482, + 0.012984172441065311, + 0.010641414672136307, + 0.014970575459301472, + 0.05008024349808693, + 0.005997714586555958, + 0.005526646971702576, + 0.018314022570848465, + 0.010570231825113297, + 0.006321284919977188, + 0.0022199677769094706, + 0.004718436859548092, + 0.003833663184195757, + 0.0023324680514633656, + 0.003691960358992219, + 0.0031399803701788187, + 0.006585469003766775, + 0.0026193319354206324, + 0.0014069428434595466, + 0.0021130635868757963, + 0.0019946694374084473, + 0.0009763606358319521, + 0.0012179642217233777, + 0.0011904002167284489, + 0.0012324793497100472, + 0.0033178837038576603, + 0.0026672258973121643, + 0.0037954887375235558, + 0.004330745432525873, + 0.002164551056921482, + 0.003414300736039877, + 0.004031555727124214, + 0.006213854532688856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.17986446619033813 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot smoothly pans to the right, following a strange procession through a minimalist, cartoon-like world composed of crisp colors and flat blocks of color. The centerpiece of the scene is a large, furry, red, car-shaped pillow, covered in soft, fluffy down, boasting exaggerated round white headlights and a smiling black grille. It sits solemnly on a simple wooden trailer. A group of short, cartoon-like figures in uniform gray overalls carefully escort it. Divided into two groups, some in the front tug at the trailer's ropes, while others push cautiously from behind. Their steps are synchronized and cautious, their faces bearing a serious, focused expression that stands out from the scene. The background is a flat, blocky palette of soft mint green ground and azure sky, dotted in the distance by a few geometric, dark green trees. The entire scene is bathed in bright, even, shadowless light. However, this cartoon purity is broken by a strong cinematic texture: the entire picture is full of obvious film grain, and details are deliberately sharpened, which gives a childlike scene a surreal, documentary-like solemnity and creates a sense of absurd and realistic drama.", + "seed": 24420, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_0/index_2217_seed_24420_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10838277251459658, + "vlm_logprob": -0.10838277251459658, + "vlm_yes_no_margin": 3.3499999761581423, + "vlm_token_credit": -0.10838277251459658, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2652547359466553 + ], + "opposite_logprob_tokens": [ + -1.6402547359466553 + ], + "logprob_reward": -0.2652547359466553, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2652547359466553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17603600025177002, + 0.04882766678929329, + 0.026289040222764015, + 0.11469503492116928, + 0.05844071879982948, + 0.23322302103042603, + 0.09004399925470352, + 0.021561892703175545, + 0.18391387164592743, + 0.04696881026029587 + ], + "credit_map": [ + 0.018074585124850273, + 0.01807139813899994, + 0.026730412617325783, + 0.012777612544596195, + 0.01431904174387455, + 0.0314304418861866, + 0.023554518818855286, + 0.031077975407242775, + 0.007201449945569038, + 0.003808389650657773, + 0.011522823944687843, + 0.004502525087445974, + 0.00314683117903769, + 0.004737638868391514, + 0.005758063867688179, + 0.008149941451847553, + 0.004239353816956282, + 0.0024493811652064323, + 0.003930477425456047, + 0.0027582638431340456, + 0.002581894863396883, + 0.0030881587881594896, + 0.0037219137884676456, + 0.003519597230479121, + 0.010088724084198475, + 0.027777619659900665, + 0.013806070201098919, + 0.015148354694247246, + 0.01463906466960907, + 0.010493157431483269, + 0.012347230687737465, + 0.010394813492894173, + 0.006279506720602512, + 0.00531153567135334, + 0.008524419739842415, + 0.006309255957603455, + 0.005204818677157164, + 0.00658575352281332, + 0.006844026502221823, + 0.01338139921426773, + 0.027889642864465714, + 0.021225426346063614, + 0.038491182029247284, + 0.025420978665351868, + 0.02442515641450882, + 0.02370752952992916, + 0.024145672097802162, + 0.04791741818189621, + 0.01926218532025814, + 0.007055209018290043, + 0.016366735100746155, + 0.006676001008599997, + 0.010108034126460552, + 0.01331300474703312, + 0.006400450132787228, + 0.010862375609576702, + 0.0026598626282066107, + 0.0022627967409789562, + 0.0030640484765172005, + 0.0032189255580306053, + 0.003440045053139329, + 0.0023236190900206566, + 0.0020990658085793257, + 0.0024935288820415735, + 0.01615607738494873, + 0.0059554991312325, + 0.0171605683863163, + 0.01158881839364767, + 0.0076082744635641575, + 0.05234471336007118, + 0.009659665636718273, + 0.06344024091959, + 0.00808646995574236, + 0.004077232908457518, + 0.00828445516526699, + 0.01166299358010292, + 0.0034057144075632095, + 0.0026584432926028967, + 0.004066421650350094, + 0.004727075342088938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.2652547359466553 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022166224662214518 + ], + "opposite_logprob_tokens": [ + -6.127216815948486 + ], + "logprob_reward": -0.0022166224662214518, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022166224662214518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08078980445861816, + 0.09261889010667801, + 0.048192914575338364, + 0.08261287957429886, + 0.04431437700986862, + 0.05779324099421501, + 0.09573324769735336, + 0.1333436369895935, + 0.25577667355537415, + 0.10882431268692017 + ], + "credit_map": [ + 0.013466037809848785, + 0.013110609725117683, + 0.01245351042598486, + 0.006321405991911888, + 0.00809452123939991, + 0.011003974825143814, + 0.00836673378944397, + 0.007973013445734978, + 0.011044487357139587, + 0.01389376912266016, + 0.01440419815480709, + 0.007098217029124498, + 0.012961150147020817, + 0.013057754375040531, + 0.008474293164908886, + 0.011685023084282875, + 0.006172826047986746, + 0.006559786852449179, + 0.007235727738589048, + 0.004727947060018778, + 0.007025782018899918, + 0.006311433389782906, + 0.005406106822192669, + 0.004753303248435259, + 0.008641372434794903, + 0.01734534464776516, + 0.010444137267768383, + 0.009462629444897175, + 0.0075146351009607315, + 0.009895565919578075, + 0.007977970875799656, + 0.011331231333315372, + 0.007060594391077757, + 0.004140685312449932, + 0.006327377632260323, + 0.004898882936686277, + 0.004175910260528326, + 0.004448810592293739, + 0.004488049540668726, + 0.00877406820654869, + 0.004980179015547037, + 0.009759626351296902, + 0.006074994802474976, + 0.00798545591533184, + 0.00889001227915287, + 0.00571935810148716, + 0.0061394087970256805, + 0.008244211785495281, + 0.010477575473487377, + 0.008186733350157738, + 0.013476766645908356, + 0.007373428903520107, + 0.014955220744013786, + 0.01892705075442791, + 0.009290152229368687, + 0.0130463233217597, + 0.014679689891636372, + 0.0045070038177073, + 0.012241708114743233, + 0.006881303619593382, + 0.0072519718669354916, + 0.04741625115275383, + 0.006593592464923859, + 0.033772118389606476, + 0.024207640439271927, + 0.011136019602417946, + 0.01775355450809002, + 0.013848341070115566, + 0.015819424763321877, + 0.08929608762264252, + 0.011417992413043976, + 0.0722975805401802, + 0.00772875314578414, + 0.00991699006408453, + 0.012704999186098576, + 0.022615788504481316, + 0.010873484425246716, + 0.009699211455881596, + 0.02129746600985527, + 0.013987615704536438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.0022166224662214518 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279280602931976 + ], + "opposite_logprob_tokens": [ + -2.0177927017211914 + ], + "logprob_reward": -0.14279280602931976, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14279280602931976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28161072731018066, + 0.22332985699176788, + 0.10301558673381805, + 0.04417947307229042, + 0.03632402792572975, + 0.03728404641151428, + 0.06118648126721382, + 0.1451238989830017, + 0.029006998986005783, + 0.038938917219638824 + ], + "credit_map": [ + 0.028582781553268433, + 0.031517885625362396, + 0.042478278279304504, + 0.028255823999643326, + 0.02596217766404152, + 0.045365333557128906, + 0.03899285942316055, + 0.04045557975769043, + 0.019970227032899857, + 0.042599599808454514, + 0.04051029309630394, + 0.01724274270236492, + 0.022559182718396187, + 0.025880908593535423, + 0.02564602717757225, + 0.028920847922563553, + 0.019507793709635735, + 0.008407024666666985, + 0.022395316511392593, + 0.008049368858337402, + 0.006936254445463419, + 0.011415572836995125, + 0.009263426996767521, + 0.01704082451760769, + 0.0053107161074876785, + 0.008477245457470417, + 0.005987139884382486, + 0.0058502680622041225, + 0.004735167603939772, + 0.003964806441217661, + 0.004531382583081722, + 0.005322742275893688, + 0.002548107411712408, + 0.004626444540917873, + 0.0030098333954811096, + 0.0036470878403633833, + 0.004445099737495184, + 0.004449603147804737, + 0.003997089806944132, + 0.009600759483873844, + 0.005589346867054701, + 0.003795535070821643, + 0.006232188083231449, + 0.0045068534091115, + 0.004750288091599941, + 0.0036991636734455824, + 0.0032643331214785576, + 0.005446333438158035, + 0.006502012722194195, + 0.006563237868249416, + 0.011197025887668133, + 0.0065674809738993645, + 0.006923259701579809, + 0.009604173712432384, + 0.003817854216322303, + 0.010011428967118263, + 0.020401768386363983, + 0.0071370904333889484, + 0.022174397483468056, + 0.00869111530482769, + 0.00960508268326521, + 0.03850506991147995, + 0.007349796127527952, + 0.03125957399606705, + 0.0031243611592799425, + 0.0025591617450118065, + 0.004555347841233015, + 0.0030447831377387047, + 0.004186491947621107, + 0.0033877152018249035, + 0.001915588742122054, + 0.006233546417206526, + 0.004621521569788456, + 0.0028982816729694605, + 0.0052998908795416355, + 0.008207286708056927, + 0.004258918110281229, + 0.0034201948437839746, + 0.004314425867050886, + 0.005918397568166256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.14279280602931976 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004632930271327496 + ], + "opposite_logprob_tokens": [ + -5.379632949829102 + ], + "logprob_reward": -0.004632930271327496, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004632930271327496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2159474939107895, + 0.07187477499246597, + 0.0563681423664093, + 0.05573757737874985, + 0.038981594145298004, + 0.12129969149827957, + 0.09277499467134476, + 0.07816392183303833, + 0.19822050631046295, + 0.07063134014606476 + ], + "credit_map": [ + 0.014988495968282223, + 0.033072274178266525, + 0.026696331799030304, + 0.02512776106595993, + 0.016423771157860756, + 0.03449697047472, + 0.030699897557497025, + 0.0344419926404953, + 0.007407791912555695, + 0.010148406960070133, + 0.011895785108208656, + 0.005151338409632444, + 0.006909055635333061, + 0.009185601957142353, + 0.007349501363933086, + 0.013827286660671234, + 0.007240781560540199, + 0.006440300494432449, + 0.0062410649843513966, + 0.005994702689349651, + 0.007095614448189735, + 0.006554975640028715, + 0.00786365382373333, + 0.008937045931816101, + 0.006239055190235376, + 0.00872713327407837, + 0.0066163064911961555, + 0.008174845017492771, + 0.004405362065881491, + 0.006389061454683542, + 0.006048883311450481, + 0.00913692731410265, + 0.00473435502499342, + 0.004836909472942352, + 0.006370522081851959, + 0.004430471919476986, + 0.0036341187078505754, + 0.0037395055405795574, + 0.003660494228824973, + 0.007575215306133032, + 0.01394579652696848, + 0.013759230263531208, + 0.018779033794999123, + 0.01345820538699627, + 0.01273394376039505, + 0.013465669006109238, + 0.01135247852653265, + 0.023805323988199234, + 0.0215441957116127, + 0.00618455745279789, + 0.01551691722124815, + 0.005707997363060713, + 0.0073831272311508656, + 0.014928016811609268, + 0.006226800847798586, + 0.015283375978469849, + 0.00971051026135683, + 0.003266864689067006, + 0.013303108513355255, + 0.0046585239470005035, + 0.005708049517124891, + 0.02178247645497322, + 0.004092046990990639, + 0.015642335638403893, + 0.01773929037153721, + 0.007661944720894098, + 0.021922096610069275, + 0.01265398133546114, + 0.00899563916027546, + 0.04926906153559685, + 0.011554667726159096, + 0.0684238076210022, + 0.008561531081795692, + 0.004899024032056332, + 0.020049145445227623, + 0.009692005813121796, + 0.005143200512975454, + 0.008170544169843197, + 0.005442716646939516, + 0.008673164993524551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.004632930271327496 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and panning smoothly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a large, red, car-shaped pillow with fluffy down, white headlights, and a smiling black grille?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there cartoon-like figures in gray overalls pulling ropes in front and pushing from behind the trailer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of a flat mint green ground, azure sky, and distant dark green geometric trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a strong film grain texture with deliberately sharpened details?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12701676785945892 + ], + "opposite_logprob_tokens": [ + -2.12701678276062 + ], + "logprob_reward": -0.12701676785945892, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12701676785945892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2890801727771759, + 0.044169507920742035, + 0.033699966967105865, + 0.029957178980112076, + 0.047616373747587204, + 0.06497403979301453, + 0.05240993574261665, + 0.07883062213659286, + 0.27270978689193726, + 0.0865524411201477 + ], + "credit_map": [ + 0.021434837952256203, + 0.0411943756043911, + 0.03486507758498192, + 0.027451496571302414, + 0.023381194099783897, + 0.0531904362142086, + 0.038738787174224854, + 0.04882393032312393, + 0.005132858641445637, + 0.006797227077186108, + 0.008510026149451733, + 0.003190492745488882, + 0.004362101666629314, + 0.00544986966997385, + 0.004382067825645208, + 0.006344858556985855, + 0.004362789914011955, + 0.005350941326469183, + 0.004515699110925198, + 0.0040908618830144405, + 0.0035306066274642944, + 0.0034703316632658243, + 0.004256286192685366, + 0.004122448153793812, + 0.002869121264666319, + 0.004315792117267847, + 0.005008662585169077, + 0.0034767307806760073, + 0.0027922894805669785, + 0.0030841731932014227, + 0.003319479525089264, + 0.005090928170830011, + 0.005533169489353895, + 0.003817363642156124, + 0.006163313984870911, + 0.005049633327871561, + 0.004620636347681284, + 0.00564802810549736, + 0.005538672674447298, + 0.011245552450418472, + 0.005031932611018419, + 0.008675036951899529, + 0.00674808444455266, + 0.008424983359873295, + 0.01195873785763979, + 0.008745066821575165, + 0.007986155338585377, + 0.007404039613902569, + 0.005417133681476116, + 0.005819070618599653, + 0.00790185946971178, + 0.005437352694571018, + 0.009126145392656326, + 0.0064180949702858925, + 0.0047756703570485115, + 0.0075146048329770565, + 0.013816460967063904, + 0.003899388713762164, + 0.015956955030560493, + 0.005836214404553175, + 0.006260926369577646, + 0.016039716079831123, + 0.004341660998761654, + 0.012679291889071465, + 0.02530251257121563, + 0.007946616038680077, + 0.02379130944609642, + 0.014120091684162617, + 0.011313749477267265, + 0.08894454687833786, + 0.013697615824639797, + 0.08759333193302155, + 0.007723803166300058, + 0.008539612405002117, + 0.01142630260437727, + 0.024540353566408157, + 0.007161078974604607, + 0.006398524157702923, + 0.011137633584439754, + 0.00962512195110321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.12701676785945892 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_1/index_tensor([1908, 1908])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_1/index_tensor([1908, 1908])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..544860f2d15d57d3a00979c205a36eb9d1f9976c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_1/index_tensor([1908, 1908])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a low angle, slowly pulls back to reveal a wider scene, following the passenger's movements. At the center of the frame is a sleek, vibrantly colored cartoon yellow taxi. Its mirror-like surface perfectly reflects the surrounding light, while the \"TAXI\" lightbox on the roof emits a soft white glow. As the taxi's rear door swings gracefully upward, a young woman alights. She's wearing a stylish dark blue trench coat, the material subtly shimmering in the shifting light, and her long, curly chestnut hair sways gently with her movements. One foot cautiously touches the wet asphalt, creating a tiny ripple illuminated by the neon light. Then, she calmly steps out of the taxi, planted on the sidewalk's edge, a look of fresh-faced curiosity and wonder on her face. She stepped into a futuristic, bustling metropolis at night. Towering skyscrapers were dotted with giant, flashing digital billboards, and cascading neon signs painted the streets a vibrant sea of deep blue, psychedelic purple, and fiery pink. Complex, dynamic light and shadow flowed across the sleek yellow body of a cartoon taxi, casting ever-changing specks of color on the passenger's windbreaker. The resulting scene, a sophisticated, Pixar-quality 3D animation, was imbued with the urban energy, technological sophistication, and a touch of the dreamlike wonder of a first visit to the big city.", + "seed": 7360, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_1/index_1908_seed_7360_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1202339076437056, + "vlm_logprob": -0.1202339076437056, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.1202339076437056, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01926485262811184 + ], + "opposite_logprob_tokens": [ + -5.0192646980285645 + ], + "logprob_reward": -0.01926485262811184, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.01926485262811184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13317617774009705, + 0.0473271980881691, + 0.04665498062968254, + 0.040173038840293884, + 0.04709301516413689, + 0.1042768657207489, + 0.15087778866291046, + 0.03992793336510658, + 0.042996253818273544, + 0.34749674797058105 + ], + "credit_map": [ + 0.042901426553726196, + 0.017537597566843033, + 0.008927911520004272, + 0.006398387718945742, + 0.007449323311448097, + 0.013433056883513927, + 0.009137884713709354, + 0.027390584349632263, + 0.01626778207719326, + 0.005894810892641544, + 0.002259934553876519, + 0.001973650651052594, + 0.001995423110201955, + 0.005551379173994064, + 0.004241752903908491, + 0.009142464026808739, + 0.014488271437585354, + 0.007581742946058512, + 0.0038408595137298107, + 0.002671420807018876, + 0.0030742816161364317, + 0.005528104957193136, + 0.0049773529171943665, + 0.004492946434766054, + 0.006064936984330416, + 0.009857499040663242, + 0.003805695101618767, + 0.0033526604529470205, + 0.0041318172588944435, + 0.0048246607184410095, + 0.005127643700689077, + 0.0030081255827099085, + 0.006106080953031778, + 0.004504017997533083, + 0.00583685515448451, + 0.009321408346295357, + 0.003972396720200777, + 0.004847459029406309, + 0.007260932121425867, + 0.005243862979114056, + 0.009068980813026428, + 0.008892270736396313, + 0.02081180177628994, + 0.011954384855926037, + 0.01617172546684742, + 0.017804544419050217, + 0.012200091034173965, + 0.007373059168457985, + 0.01351061649620533, + 0.0142269441857934, + 0.04300171881914139, + 0.016005827113986015, + 0.017731759697198868, + 0.012823227792978287, + 0.014445281587541103, + 0.019132409244775772, + 0.004439473617821932, + 0.004822581075131893, + 0.007041778415441513, + 0.0048784310929477215, + 0.0057258992455899715, + 0.006252975668758154, + 0.004077427554875612, + 0.0026893671602010727, + 0.005447932984679937, + 0.005409878212958574, + 0.006540341302752495, + 0.007617384660989046, + 0.006456718314439058, + 0.004257687833160162, + 0.00354886706918478, + 0.0037174425087869167, + 0.0587083175778389, + 0.05000375583767891, + 0.04601375758647919, + 0.07523830235004425, + 0.04948737472295761, + 0.019332092255353928, + 0.029280226677656174, + 0.019432934001088142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.01926485262811184 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031582415103912 + ], + "opposite_logprob_tokens": [ + -2.350315809249878 + ], + "logprob_reward": -0.10031582415103912, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031582415103912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06660713255405426, + 0.05200197175145149, + 0.02527589537203312, + 0.037321340292692184, + 0.12040570378303528, + 0.3258845806121826, + 0.2203039824962616, + 0.021292610093951225, + 0.02137484773993492, + 0.10953196883201599 + ], + "credit_map": [ + 0.011614051647484303, + 0.009642786346375942, + 0.007000288460403681, + 0.006478079129010439, + 0.0061514172703027725, + 0.008444060571491718, + 0.005797060206532478, + 0.011479390785098076, + 0.004409239161759615, + 0.00464288517832756, + 0.004240983631461859, + 0.003907925449311733, + 0.004665336571633816, + 0.017678331583738327, + 0.005843144375830889, + 0.006614124868065119, + 0.005634299013763666, + 0.005776432808488607, + 0.0016059188637882471, + 0.0020621123258024454, + 0.002089886227622628, + 0.002097618766129017, + 0.0031712972559034824, + 0.0028383315075188875, + 0.004825705196708441, + 0.007932428270578384, + 0.003767110873013735, + 0.0040319934487342834, + 0.003940609749406576, + 0.004312650766223669, + 0.00512862391769886, + 0.003382214577868581, + 0.014740216545760632, + 0.011548248119652271, + 0.018035750836133957, + 0.024280894547700882, + 0.00868696253746748, + 0.0133795952424407, + 0.019029494374990463, + 0.010704537853598595, + 0.026409821584820747, + 0.040440723299980164, + 0.08412022143602371, + 0.03789143264293671, + 0.02811591327190399, + 0.06379169970750809, + 0.023042792454361916, + 0.022071978077292442, + 0.022102635353803635, + 0.02231360226869583, + 0.06296798586845398, + 0.02054637484252453, + 0.025954395532608032, + 0.017103346064686775, + 0.023172136396169662, + 0.0261435117572546, + 0.002071420894935727, + 0.006034412421286106, + 0.0021418193355202675, + 0.001851645647548139, + 0.0026925017591565847, + 0.0016589814331382513, + 0.00135194999165833, + 0.0034898784942924976, + 0.002451719017699361, + 0.003204463515430689, + 0.0022255803924053907, + 0.0029459844809025526, + 0.003713418496772647, + 0.0036523446906358004, + 0.0016432645497843623, + 0.0015380698023363948, + 0.01642574928700924, + 0.016665352508425713, + 0.01376685407012701, + 0.023617755621671677, + 0.013779534958302975, + 0.007714343722909689, + 0.010149741545319557, + 0.007412629202008247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.10031582415103912 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950829479843378 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950829479843378, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950829479843378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0718492716550827, + 0.056785039603710175, + 0.04065892472863197, + 0.0969378873705864, + 0.058936480432748795, + 0.16563352942466736, + 0.21765556931495667, + 0.05741775035858154, + 0.09494898468255997, + 0.1391766369342804 + ], + "credit_map": [ + 0.019157888367772102, + 0.010625917464494705, + 0.005157338455319405, + 0.005817904137074947, + 0.004304816480726004, + 0.007263238541781902, + 0.006006238050758839, + 0.013515923172235489, + 0.01000191830098629, + 0.00670004915446043, + 0.0032264215406030416, + 0.0026565990410745144, + 0.004903578199446201, + 0.01617012731730938, + 0.0053201583214104176, + 0.007806182838976383, + 0.00911601260304451, + 0.011402915231883526, + 0.0022084314841777086, + 0.0030108701903373003, + 0.0040061441250145435, + 0.003123584669083357, + 0.004581436514854431, + 0.003209528047591448, + 0.011854056268930435, + 0.029807401821017265, + 0.009320704266428947, + 0.00608227401971817, + 0.007559692021459341, + 0.011846913024783134, + 0.014814476482570171, + 0.005652368068695068, + 0.009601977653801441, + 0.006222167052328587, + 0.008767402730882168, + 0.009535265155136585, + 0.005290154833346605, + 0.007780145853757858, + 0.006370913237333298, + 0.005368447862565517, + 0.011408718302845955, + 0.011033790186047554, + 0.05327988788485527, + 0.017897551879286766, + 0.012361324392259121, + 0.04439600557088852, + 0.007237561047077179, + 0.008018678985536098, + 0.022999636828899384, + 0.02142592892050743, + 0.0703565925359726, + 0.016288483515381813, + 0.02855231799185276, + 0.016824238002300262, + 0.01423565298318863, + 0.026972701773047447, + 0.0047546750865876675, + 0.004999825730919838, + 0.004610615782439709, + 0.0038894861936569214, + 0.01832525245845318, + 0.005999090615659952, + 0.0036768238060176373, + 0.01116197183728218, + 0.0143636679276824, + 0.0139723876491189, + 0.009090774692595005, + 0.00837415549904108, + 0.023365242406725883, + 0.012623677030205727, + 0.005935329012572765, + 0.007223744876682758, + 0.023532390594482422, + 0.018148982897400856, + 0.01805751584470272, + 0.03241204097867012, + 0.01863783411681652, + 0.008098133839666843, + 0.011761813424527645, + 0.008527922444045544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.005950829479843378 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270252764225006 + ], + "opposite_logprob_tokens": [ + -2.1270253658294678 + ], + "logprob_reward": -0.1270252764225006, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270252764225006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08862917125225067, + 0.07521144300699234, + 0.043523382395505905, + 0.03765730559825897, + 0.08801854401826859, + 0.4175069332122803, + 0.12331537902355194, + 0.03238647058606148, + 0.01839287020266056, + 0.07535843551158905 + ], + "credit_map": [ + 0.019727099686861038, + 0.010164623148739338, + 0.0060705989599227905, + 0.0056672971695661545, + 0.007669713348150253, + 0.0176366176456213, + 0.006259420420974493, + 0.015433802269399166, + 0.007206372916698456, + 0.008914373815059662, + 0.008424162864685059, + 0.008011762984097004, + 0.0068326774053275585, + 0.015898048877716064, + 0.010077599436044693, + 0.009846446104347706, + 0.007782332599163055, + 0.013529176823794842, + 0.0030109775252640247, + 0.00336186564527452, + 0.003326933365315199, + 0.0027464316226541996, + 0.005528322421014309, + 0.004237343091517687, + 0.0050669340416789055, + 0.006168609485030174, + 0.0036720260977745056, + 0.0033421397674828768, + 0.004606686066836119, + 0.005249313544481993, + 0.005766007583588362, + 0.003785589709877968, + 0.012925073504447937, + 0.008517002686858177, + 0.013718594796955585, + 0.013637258671224117, + 0.007765667978674173, + 0.011822490952908993, + 0.012043217197060585, + 0.007589235436171293, + 0.029568970203399658, + 0.02670578844845295, + 0.1305595338344574, + 0.046546585857868195, + 0.040383923798799515, + 0.11248524487018585, + 0.015341089107096195, + 0.015915829688310623, + 0.013710970059037209, + 0.013188989832997322, + 0.037323154509067535, + 0.009944389574229717, + 0.01613147370517254, + 0.009528613649308681, + 0.006757223047316074, + 0.016730565577745438, + 0.0027632308192551136, + 0.010883002541959286, + 0.0019102327059954405, + 0.00306510366499424, + 0.006858988665044308, + 0.002673237817361951, + 0.0020200104918330908, + 0.002212666207924485, + 0.001786242239177227, + 0.002195179695263505, + 0.0018327757716178894, + 0.001897629932500422, + 0.003379232482984662, + 0.005065837409347296, + 0.0011500061955302954, + 0.0010859653120860457, + 0.012053594924509525, + 0.012329338118433952, + 0.007848762907087803, + 0.014418446458876133, + 0.011733779683709145, + 0.005147829186171293, + 0.006265066098421812, + 0.005561613943427801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.1270252764225006 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486127555370331 + ], + "opposite_logprob_tokens": [ + -1.2236127853393555 + ], + "logprob_reward": -0.3486127555370331, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486127555370331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10476494580507278, + 0.03995908796787262, + 0.07510711997747421, + 0.030651859939098358, + 0.07500340789556503, + 0.24163131415843964, + 0.1727169305086136, + 0.03611883893609047, + 0.028810610994696617, + 0.19523589313030243 + ], + "credit_map": [ + 0.01806519739329815, + 0.012337503023445606, + 0.007500063627958298, + 0.007929784245789051, + 0.009156549349427223, + 0.02576575055718422, + 0.008167891763150692, + 0.01584220491349697, + 0.007430662866681814, + 0.005937875714153051, + 0.0026467982679605484, + 0.002878316678106785, + 0.0033989434596151114, + 0.007676685694605112, + 0.0034934685099869967, + 0.006496339105069637, + 0.019816266372799873, + 0.01743408851325512, + 0.004850758705288172, + 0.0053862277418375015, + 0.004847467876970768, + 0.005121155641973019, + 0.009995612315833569, + 0.007655540015548468, + 0.004455495625734329, + 0.005193708930164576, + 0.003574149217456579, + 0.003173593897372484, + 0.003117902670055628, + 0.004495831206440926, + 0.0042021446861326694, + 0.0024390348698943853, + 0.007848060689866543, + 0.0053999233059585094, + 0.013581983745098114, + 0.016022933647036552, + 0.00497341388836503, + 0.007527440320700407, + 0.012327546253800392, + 0.007322111167013645, + 0.016064267605543137, + 0.015504312701523304, + 0.07844021916389465, + 0.026510288938879967, + 0.017434248700737953, + 0.06876899302005768, + 0.008242885582149029, + 0.010666091926395893, + 0.017534589394927025, + 0.016847198829054832, + 0.05466524511575699, + 0.013366373255848885, + 0.022255148738622665, + 0.013835484161973, + 0.01231176033616066, + 0.021901147440075874, + 0.0034012754913419485, + 0.010688049718737602, + 0.0036754566244781017, + 0.0033123712055385113, + 0.007486860267817974, + 0.0027002161368727684, + 0.00260699400678277, + 0.002247618278488517, + 0.0030279341153800488, + 0.0036331601440906525, + 0.002780838403850794, + 0.004297506529837847, + 0.004946658853441477, + 0.005430886056274176, + 0.0023803480435162783, + 0.002313279779627919, + 0.03430737555027008, + 0.027697967365384102, + 0.024639727547764778, + 0.043051011860370636, + 0.027332881465554237, + 0.010981233790516853, + 0.0163344144821167, + 0.010891268961131573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.3486127555370331 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a low angle, slowly pulls back to reveal a wider scene, following the passenger's movements. At the center of the frame is a sleek, vibrantly colored cartoon yellow taxi. Its mirror-like surface perfectly reflects the surrounding light, while the \"TAXI\" lightbox on the roof emits a soft white glow. As the taxi's rear door swings gracefully upward, a young woman alights. She's wearing a stylish dark blue trench coat, the material subtly shimmering in the shifting light, and her long, curly chestnut hair sways gently with her movements. One foot cautiously touches the wet asphalt, creating a tiny ripple illuminated by the neon light. Then, she calmly steps out of the taxi, planted on the sidewalk's edge, a look of fresh-faced curiosity and wonder on her face. She stepped into a futuristic, bustling metropolis at night. Towering skyscrapers were dotted with giant, flashing digital billboards, and cascading neon signs painted the streets a vibrant sea of deep blue, psychedelic purple, and fiery pink. Complex, dynamic light and shadow flowed across the sleek yellow body of a cartoon taxi, casting ever-changing specks of color on the passenger's windbreaker. The resulting scene, a sophisticated, Pixar-quality 3D animation, was imbued with the urban energy, technological sophistication, and a touch of the dreamlike wonder of a first visit to the big city.", + "seed": 7361, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_1/index_1908_seed_7361_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15801517609506846, + "vlm_logprob": -0.15801517609506846, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.15801517609506846, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022783884778618813 + ], + "opposite_logprob_tokens": [ + -6.1477837562561035 + ], + "logprob_reward": -0.022783884778618813, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.022783884778618813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1402195543050766, + 0.06628554314374924, + 0.33343860507011414, + 0.06069940701127052, + 0.06188840791583061, + 0.03851873055100441, + 0.13193513453006744, + 0.0326511524617672, + 0.07868301868438721, + 0.05568046495318413 + ], + "credit_map": [ + 0.015290671959519386, + 0.015331970527768135, + 0.021790945902466774, + 0.027121003717184067, + 0.01308022253215313, + 0.011142264120280743, + 0.017400596290826797, + 0.01906188204884529, + 0.007461448200047016, + 0.005038601811975241, + 0.009820129722356796, + 0.008279578760266304, + 0.0053024147637188435, + 0.015340986661612988, + 0.006537805311381817, + 0.008504579775035381, + 0.012700767256319523, + 0.02355141006410122, + 0.006576919928193092, + 0.06024179980158806, + 0.07543497532606125, + 0.017971038818359375, + 0.11287881433963776, + 0.024082863703370094, + 0.0059638177044689655, + 0.005508525762706995, + 0.007507802452892065, + 0.007415523752570152, + 0.011520525440573692, + 0.008004766888916492, + 0.006546769291162491, + 0.008231674320995808, + 0.006093773525208235, + 0.011879038996994495, + 0.004524000454694033, + 0.007598682306706905, + 0.008081784471869469, + 0.0062028709799051285, + 0.008253928273916245, + 0.009254327975213528, + 0.0032889703288674355, + 0.002085213316604495, + 0.007573988288640976, + 0.0042332676239311695, + 0.0032214592210948467, + 0.007018634118139744, + 0.003965584561228752, + 0.0071316128596663475, + 0.008563775569200516, + 0.048292551189661026, + 0.005781345535069704, + 0.005423278082162142, + 0.014911162666976452, + 0.03283581882715225, + 0.006645578891038895, + 0.00948162842541933, + 0.005032799672335386, + 0.003318153554573655, + 0.0033912414219230413, + 0.0037805624306201935, + 0.004717002157121897, + 0.003759134793654084, + 0.004801030270755291, + 0.0038512293249368668, + 0.009589278139173985, + 0.007248838432133198, + 0.008677595295011997, + 0.0069478717632591724, + 0.010004236362874508, + 0.009608379565179348, + 0.01054922491312027, + 0.016057591885328293, + 0.006603309419006109, + 0.006138138007372618, + 0.004144251812249422, + 0.005917677190154791, + 0.009175722487270832, + 0.008913927711546421, + 0.009048261679708958, + 0.005739179439842701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.022783884778618813 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635977976024151 + ], + "opposite_logprob_tokens": [ + -3.651359796524048 + ], + "logprob_reward": -0.02635977976024151, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635977976024151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07665762305259705, + 0.030614934861660004, + 0.3075505495071411, + 0.02346337027847767, + 0.015835532918572426, + 0.03167891874909401, + 0.3503524363040924, + 0.030879104509949684, + 0.0590628981590271, + 0.07390465587377548 + ], + "credit_map": [ + 0.011225174181163311, + 0.008192181587219238, + 0.008022022433578968, + 0.013472856022417545, + 0.008009166456758976, + 0.009257297031581402, + 0.01025379728525877, + 0.008225121535360813, + 0.0030420389957726, + 0.0024918466806411743, + 0.004333049058914185, + 0.003783390624448657, + 0.0023809082340449095, + 0.0058729881420731544, + 0.004186904989182949, + 0.004523808602243662, + 0.011721737682819366, + 0.01858220249414444, + 0.004878872539848089, + 0.055087290704250336, + 0.06612477451562881, + 0.015385926701128483, + 0.10945995897054672, + 0.026309793815016747, + 0.002129861619323492, + 0.0020841278601437807, + 0.0023249643854796886, + 0.003615926019847393, + 0.003250619163736701, + 0.002487648744136095, + 0.0032136503141373396, + 0.00435657100751996, + 0.0014915529172867537, + 0.0015636010793969035, + 0.000959946250077337, + 0.0017904113046824932, + 0.0022307932376861572, + 0.0027589001692831516, + 0.002352521289139986, + 0.0026878071948885918, + 0.004552728962153196, + 0.004292431753128767, + 0.006833226419985294, + 0.0034440995659679174, + 0.0030958110000938177, + 0.002861496526747942, + 0.0036875419318675995, + 0.0029115816578269005, + 0.025702599436044693, + 0.1490766853094101, + 0.012645694427192211, + 0.013997478410601616, + 0.020889414474368095, + 0.0767316147685051, + 0.012819112278521061, + 0.03848982974886894, + 0.004726245068013668, + 0.005984472576528788, + 0.0032863812521100044, + 0.0026244886685162783, + 0.0045799799263477325, + 0.004029933828860521, + 0.0031120898202061653, + 0.002535514999181032, + 0.005703102797269821, + 0.006467668805271387, + 0.005712187383323908, + 0.0049902088940143585, + 0.008113449439406395, + 0.00889890268445015, + 0.01348053477704525, + 0.005696842446923256, + 0.005292196292430162, + 0.006391200702637434, + 0.004347596783190966, + 0.00511549785733223, + 0.026305781677365303, + 0.008473871275782585, + 0.006818895693868399, + 0.011159619316458702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.02635977976024151 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256284959614277 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256284959614277, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256284959614277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1110810786485672, + 0.1810733675956726, + 0.25519880652427673, + 0.05646773427724838, + 0.021805213764309883, + 0.023265458643436432, + 0.2478661984205246, + 0.023287056013941765, + 0.04433990269899368, + 0.03561512380838394 + ], + "credit_map": [ + 0.011995857581496239, + 0.011942618526518345, + 0.012860657647252083, + 0.01669255830347538, + 0.017134014517068863, + 0.015146229416131973, + 0.016073880717158318, + 0.009235265664756298, + 0.030757684260606766, + 0.01858091726899147, + 0.027087019756436348, + 0.022213563323020935, + 0.017489219084382057, + 0.029158059507608414, + 0.017916589975357056, + 0.017870333045721054, + 0.02627391181886196, + 0.01262922678142786, + 0.009667127393186092, + 0.05571203678846359, + 0.014560623094439507, + 0.02437412366271019, + 0.08764030039310455, + 0.02434149943292141, + 0.006247317418456078, + 0.005082504823803902, + 0.004698649048805237, + 0.004479337017983198, + 0.010333758778870106, + 0.013762702234089375, + 0.006609923206269741, + 0.005253549665212631, + 0.002187162172049284, + 0.002601280575618148, + 0.0020616103429347277, + 0.002684452570974827, + 0.00239953538402915, + 0.004632715601474047, + 0.001680199522525072, + 0.003558260854333639, + 0.0029940034728497267, + 0.0018182137282565236, + 0.003423248650506139, + 0.0019431894179433584, + 0.0032133942004293203, + 0.0032906397245824337, + 0.003309683408588171, + 0.0032730884850025177, + 0.01698712259531021, + 0.10599300265312195, + 0.009788984432816505, + 0.009148084558546543, + 0.014165056869387627, + 0.06803828477859497, + 0.008601713925600052, + 0.015143964439630508, + 0.0036821430549025536, + 0.0036863707937300205, + 0.0017402907833456993, + 0.0021164068020880222, + 0.003722632071003318, + 0.002952402690425515, + 0.002348812995478511, + 0.0030380007810890675, + 0.012033906765282154, + 0.005898790899664164, + 0.0032426933757960796, + 0.0029123821295797825, + 0.004772338084876537, + 0.0038394613657146692, + 0.006767362356185913, + 0.004872972145676613, + 0.0032052434980869293, + 0.003657681867480278, + 0.0024430719204247, + 0.0028256478253751993, + 0.010046588256955147, + 0.004778568632900715, + 0.0023429354187101126, + 0.006315389648079872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.005256284959614277 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870486319065094 + ], + "opposite_logprob_tokens": [ + -1.137048602104187 + ], + "logprob_reward": -0.3870486319065094, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870486319065094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062333039939403534, + 0.03630582243204117, + 0.5286610722541809, + 0.011089003644883633, + 0.011788971722126007, + 0.013259771279990673, + 0.11377563327550888, + 0.03826336935162544, + 0.04531635716557503, + 0.13920702040195465 + ], + "credit_map": [ + 0.010387273505330086, + 0.004978993441909552, + 0.010359598323702812, + 0.008329191245138645, + 0.005600528791546822, + 0.006791180465370417, + 0.009621512144804, + 0.006264760624617338, + 0.004492016974836588, + 0.003911777399480343, + 0.00461864797398448, + 0.005164606962352991, + 0.0030250349082052708, + 0.005773013923317194, + 0.005451448727399111, + 0.0038692776579409838, + 0.02568889781832695, + 0.03537312150001526, + 0.009312325157225132, + 0.08300378173589706, + 0.12849128246307373, + 0.026912933215498924, + 0.17761445045471191, + 0.04226427525281906, + 0.001012321561574936, + 0.0010638884268701077, + 0.0012821453856304288, + 0.0011345117818564177, + 0.001961349742487073, + 0.002109444234520197, + 0.0012053590035066009, + 0.0013199836248531938, + 0.0013022981584072113, + 0.0012158692115917802, + 0.001197835081256926, + 0.0015699046198278666, + 0.000914607779122889, + 0.0017451482126489282, + 0.001319073373451829, + 0.002524235285818577, + 0.0011516950326040387, + 0.0011081116972491145, + 0.003910696599632502, + 0.001317304908297956, + 0.0013261548010632396, + 0.001932054408825934, + 0.0014660295564681292, + 0.001047724625095725, + 0.007523227483034134, + 0.045185040682554245, + 0.004687173757702112, + 0.004496095236390829, + 0.009590019471943378, + 0.03303715959191322, + 0.004711752757430077, + 0.00454516289755702, + 0.006494942586869001, + 0.01088077574968338, + 0.0034216810017824173, + 0.003660927526652813, + 0.0043245358392596245, + 0.003007144434377551, + 0.0034137640614062548, + 0.003059596987441182, + 0.0057899863459169865, + 0.006580956745892763, + 0.002507814671844244, + 0.0031547918915748596, + 0.006636780686676502, + 0.010037415660917759, + 0.005033695138990879, + 0.005574916489422321, + 0.009719819761812687, + 0.009003234095871449, + 0.006829776335507631, + 0.005563977174460888, + 0.05662794038653374, + 0.01823088526725769, + 0.008944732137024403, + 0.0242866612970829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.3870486319065094 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486272990703583 + ], + "opposite_logprob_tokens": [ + -1.2236273288726807 + ], + "logprob_reward": -0.3486272990703583, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486272990703583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03147285059094429, + 0.02181605063378811, + 0.6984389424324036, + 0.11890935152769089, + 0.01895756646990776, + 0.03517472371459007, + 0.029158225283026695, + 0.012305931188166142, + 0.011433533392846584, + 0.02233288064599037 + ], + "credit_map": [ + 0.003621545620262623, + 0.003107118420302868, + 0.004404641687870026, + 0.005257213953882456, + 0.0029710368253290653, + 0.0039559234865009785, + 0.004775775596499443, + 0.0033795947674661875, + 0.003594527952373028, + 0.0024159089662134647, + 0.0028585901018232107, + 0.0026210607029497623, + 0.0018459951970726252, + 0.003935006447136402, + 0.0024048322811722755, + 0.0021401280537247658, + 0.02814408391714096, + 0.04427415132522583, + 0.01147401425987482, + 0.11824400722980499, + 0.15890493988990784, + 0.03736491873860359, + 0.2424584925174713, + 0.05757433548569679, + 0.010119005106389523, + 0.010086150839924812, + 0.010124342516064644, + 0.011955244466662407, + 0.02211553417146206, + 0.020009156316518784, + 0.020956743508577347, + 0.013543186709284782, + 0.0015821884153410792, + 0.002248752163723111, + 0.0028658390510827303, + 0.0027203685604035854, + 0.002195690292865038, + 0.0018129524542018771, + 0.002448473358526826, + 0.0030833007767796516, + 0.003540756879374385, + 0.007389216683804989, + 0.004735003225505352, + 0.0033905364107340574, + 0.003938960377126932, + 0.00448858505114913, + 0.00287497416138649, + 0.0048166909255087376, + 0.0017707281513139606, + 0.0046077207662165165, + 0.0012068673968315125, + 0.0018478602869436145, + 0.002095684176310897, + 0.006378110498189926, + 0.002151977038010955, + 0.00909927673637867, + 0.0018834968795999885, + 0.0027043551672250032, + 0.000900263141375035, + 0.0011882784310728312, + 0.0013515760656446218, + 0.002250304678454995, + 0.0010115951299667358, + 0.0010160612873733044, + 0.0021177390590310097, + 0.0011228699004277587, + 0.0007426125230267644, + 0.0006952497642487288, + 0.002065629232674837, + 0.0014750980772078037, + 0.0019347098423168063, + 0.0012796249939128757, + 0.002282459754496813, + 0.002231268212199211, + 0.001505125081166625, + 0.0015482648741453886, + 0.007034383248537779, + 0.002766340272501111, + 0.0017878826474770904, + 0.0031771587673574686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.3486272990703583 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a low angle, slowly pulls back to reveal a wider scene, following the passenger's movements. At the center of the frame is a sleek, vibrantly colored cartoon yellow taxi. Its mirror-like surface perfectly reflects the surrounding light, while the \"TAXI\" lightbox on the roof emits a soft white glow. As the taxi's rear door swings gracefully upward, a young woman alights. She's wearing a stylish dark blue trench coat, the material subtly shimmering in the shifting light, and her long, curly chestnut hair sways gently with her movements. One foot cautiously touches the wet asphalt, creating a tiny ripple illuminated by the neon light. Then, she calmly steps out of the taxi, planted on the sidewalk's edge, a look of fresh-faced curiosity and wonder on her face. She stepped into a futuristic, bustling metropolis at night. Towering skyscrapers were dotted with giant, flashing digital billboards, and cascading neon signs painted the streets a vibrant sea of deep blue, psychedelic purple, and fiery pink. Complex, dynamic light and shadow flowed across the sleek yellow body of a cartoon taxi, casting ever-changing specks of color on the passenger's windbreaker. The resulting scene, a sophisticated, Pixar-quality 3D animation, was imbued with the urban energy, technological sophistication, and a touch of the dreamlike wonder of a first visit to the big city.", + "seed": 7362, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_1/index_1908_seed_7362_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1367468498647213, + "vlm_logprob": -0.1367468498647213, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.1367468498647213, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04044090956449509 + ], + "opposite_logprob_tokens": [ + -3.665440797805786 + ], + "logprob_reward": -0.04044090956449509, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.04044090956449509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08179996907711029, + 0.1567416936159134, + 0.08553650975227356, + 0.06957916170358658, + 0.08212213963270187, + 0.05482031777501106, + 0.039333947002887726, + 0.0874224305152893, + 0.2561478316783905, + 0.08649598807096481 + ], + "credit_map": [ + 0.007751874625682831, + 0.008914136327803135, + 0.01064041256904602, + 0.010227725841104984, + 0.010969027876853943, + 0.008726869709789753, + 0.011250324547290802, + 0.01331960130482912, + 0.00784845370799303, + 0.0060072531923651695, + 0.05996134504675865, + 0.035892363637685776, + 0.006405639462172985, + 0.0227219108492136, + 0.009215964935719967, + 0.008688777685165405, + 0.009245253168046474, + 0.0059811631217598915, + 0.022641237825155258, + 0.01443810760974884, + 0.007316527422517538, + 0.005814261268824339, + 0.0093489745631814, + 0.010750995017588139, + 0.008405642583966255, + 0.004290392622351646, + 0.006847138050943613, + 0.01352615561336279, + 0.0069626811891794205, + 0.009399614296853542, + 0.013545927591621876, + 0.006601609755307436, + 0.00832882709801197, + 0.005121951922774315, + 0.00991313811391592, + 0.009037068113684654, + 0.01965387724339962, + 0.009314393624663353, + 0.012248001992702484, + 0.008504888974130154, + 0.00856397021561861, + 0.005620602518320084, + 0.006479984614998102, + 0.005957807879894972, + 0.008900374174118042, + 0.006226809229701757, + 0.006454554386436939, + 0.00661621754989028, + 0.004127838648855686, + 0.0031755201052874327, + 0.0035735834389925003, + 0.004353814292699099, + 0.007427997887134552, + 0.004069381859153509, + 0.007986234501004219, + 0.004619578830897808, + 0.009433918632566929, + 0.006342634093016386, + 0.00993706751614809, + 0.006803687661886215, + 0.010732555761933327, + 0.013764863833785057, + 0.012683444656431675, + 0.01772427186369896, + 0.019530871883034706, + 0.009338540956377983, + 0.005491218063980341, + 0.005847558379173279, + 0.0128105990588665, + 0.1017199158668518, + 0.015596269629895687, + 0.08581286668777466, + 0.005971269682049751, + 0.005121384747326374, + 0.0023978203535079956, + 0.002685536863282323, + 0.004850361961871386, + 0.028857136145234108, + 0.008101439103484154, + 0.02851104363799095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.04044090956449509 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549759604036808 + ], + "opposite_logprob_tokens": [ + -2.929975986480713 + ], + "logprob_reward": -0.0549759604036808, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549759604036808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07093267142772675, + 0.0786038264632225, + 0.28449490666389465, + 0.13939771056175232, + 0.055842239409685135, + 0.09659112989902496, + 0.024176839739084244, + 0.049539364874362946, + 0.1009911522269249, + 0.09943019598722458 + ], + "credit_map": [ + 0.00816770177334547, + 0.004892799537628889, + 0.009494963102042675, + 0.010416814126074314, + 0.005274372640997171, + 0.008254779502749443, + 0.013561272993683815, + 0.01086996216326952, + 0.005319137126207352, + 0.004070438910275698, + 0.020924843847751617, + 0.02415100485086441, + 0.004448509309440851, + 0.00958099402487278, + 0.005211981013417244, + 0.004896915517747402, + 0.016486020758748055, + 0.01145312748849392, + 0.0959797129034996, + 0.09438005834817886, + 0.018415823578834534, + 0.015150447376072407, + 0.022104158997535706, + 0.010525540448725224, + 0.009255916811525822, + 0.009726137854158878, + 0.01583978347480297, + 0.015405247919261456, + 0.0355556420981884, + 0.01817362941801548, + 0.019095297902822495, + 0.016346056014299393, + 0.00471744267269969, + 0.003950837068259716, + 0.007366884965449572, + 0.006835401523858309, + 0.010179038159549236, + 0.006212263368070126, + 0.008105567656457424, + 0.008474807254970074, + 0.010994622483849525, + 0.005651533603668213, + 0.007232825271785259, + 0.0218088086694479, + 0.007536391261965036, + 0.02661127597093582, + 0.007413604762405157, + 0.009342067874968052, + 0.003136158687993884, + 0.002287960145622492, + 0.002495117951184511, + 0.0018679896602407098, + 0.0032549600582569838, + 0.004331476055085659, + 0.003369302488863468, + 0.0034338748082518578, + 0.004610986914485693, + 0.005378534086048603, + 0.005103775765746832, + 0.0031723561696708202, + 0.007399463094770908, + 0.0056272889487445354, + 0.007845173589885235, + 0.010401790030300617, + 0.0076463534496724606, + 0.0036589063238352537, + 0.0030245757661759853, + 0.0026918582152575254, + 0.005955733358860016, + 0.0389983095228672, + 0.006167616695165634, + 0.032847799360752106, + 0.007070691790431738, + 0.004512448329478502, + 0.001946261036209762, + 0.0028554052114486694, + 0.005545461550354958, + 0.03409487381577492, + 0.00965126883238554, + 0.03375378996133804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.0549759604036808 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813761532306671 + ], + "opposite_logprob_tokens": [ + -1.4063761234283447 + ], + "logprob_reward": -0.2813761532306671, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813761532306671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22352981567382812, + 0.11397193372249603, + 0.15566937625408173, + 0.08792518079280853, + 0.02894098125398159, + 0.15214043855667114, + 0.019559333100914955, + 0.04157673940062523, + 0.10486089438199997, + 0.07182536274194717 + ], + "credit_map": [ + 0.028879350051283836, + 0.023362111300230026, + 0.027369990944862366, + 0.03159918636083603, + 0.02725907415151596, + 0.02449030429124832, + 0.026653867214918137, + 0.033915914595127106, + 0.006096552591770887, + 0.004735466558486223, + 0.041381023824214935, + 0.03199320659041405, + 0.0054565113969147205, + 0.011651820503175259, + 0.005799171980470419, + 0.006858167704194784, + 0.0046841795556247234, + 0.003746940754354, + 0.06170188635587692, + 0.06095890700817108, + 0.006383057218044996, + 0.005243722815066576, + 0.008342097513377666, + 0.004608582239598036, + 0.01592329889535904, + 0.006625007838010788, + 0.01082809828221798, + 0.013610674999654293, + 0.008076870813965797, + 0.013364454731345177, + 0.007069567684084177, + 0.012427199631929398, + 0.00273051206022501, + 0.002479786518961191, + 0.003951002843677998, + 0.0037429092917591333, + 0.005136198829859495, + 0.003666164120659232, + 0.003690525656566024, + 0.003543880535289645, + 0.012636358849704266, + 0.009768078103661537, + 0.012224367819726467, + 0.03134414181113243, + 0.02033305913209915, + 0.03614594787359238, + 0.015103193931281567, + 0.014585275202989578, + 0.0024257067125290632, + 0.0013761586742475629, + 0.0022303631994873285, + 0.0017095641233026981, + 0.0020689074881374836, + 0.004301644396036863, + 0.002419589553028345, + 0.0030273974407464266, + 0.006650857161730528, + 0.0062028500251472, + 0.0033223049249500036, + 0.0035002483054995537, + 0.005311666987836361, + 0.007289308123290539, + 0.004274807404726744, + 0.0050246925093233585, + 0.007902780547738075, + 0.004384104162454605, + 0.002414821647107601, + 0.002172989770770073, + 0.004499110393226147, + 0.04253868758678436, + 0.005650944542139769, + 0.0352974496781826, + 0.004453095607459545, + 0.0033833179622888565, + 0.0013663371792063117, + 0.0014036818174645305, + 0.002441640943288803, + 0.02707863226532936, + 0.004692788235843182, + 0.027005860581994057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.2813761532306671 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270408183336258 + ], + "opposite_logprob_tokens": [ + -2.1270408630371094 + ], + "logprob_reward": -0.1270408183336258, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270408183336258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059108857065439224, + 0.21837326884269714, + 0.39231207966804504, + 0.049302808940410614, + 0.03851217404007912, + 0.13364553451538086, + 0.07580623030662537, + 0.012138822115957737, + 0.007355025038123131, + 0.013445217162370682 + ], + "credit_map": [ + 0.005978552624583244, + 0.0049753570929169655, + 0.007462077774107456, + 0.009884114377200603, + 0.006608747411519289, + 0.007077322341501713, + 0.009621751494705677, + 0.007500938139855862, + 0.007744562346488237, + 0.005064990371465683, + 0.08683665841817856, + 0.08282420784235, + 0.007464746944606304, + 0.007273704279214144, + 0.009010545909404755, + 0.012153835035860538, + 0.008595775812864304, + 0.0077524082735180855, + 0.16297978162765503, + 0.16211819648742676, + 0.011631988920271397, + 0.010696965269744396, + 0.01721702702343464, + 0.01131991297006607, + 0.0066410028375685215, + 0.0036215754225850105, + 0.006808690261095762, + 0.005094226449728012, + 0.0049012587405741215, + 0.00985716748982668, + 0.006859647575765848, + 0.005519238766282797, + 0.0035438251215964556, + 0.0026884705293923616, + 0.0039135473780334, + 0.0056092506274580956, + 0.006342658773064613, + 0.004632583353668451, + 0.004958778154104948, + 0.006823062896728516, + 0.014719042927026749, + 0.008294631727039814, + 0.0104781249538064, + 0.02918749488890171, + 0.012132667936384678, + 0.034126415848731995, + 0.011951019987463951, + 0.01275613158941269, + 0.007868031039834023, + 0.006548834033310413, + 0.006190734449774027, + 0.008167941123247147, + 0.014152811840176582, + 0.009776434861123562, + 0.016802119091153145, + 0.006299320608377457, + 0.0008037515799514949, + 0.0008366359397768974, + 0.0008440123638138175, + 0.0010530189611017704, + 0.0017027752473950386, + 0.0018313172040507197, + 0.003180472180247307, + 0.001886838930658996, + 0.0005470736650750041, + 0.00029266224009916186, + 0.00024461784050799906, + 0.00024382940318901092, + 0.000579707557335496, + 0.002587876981124282, + 0.0005839167279191315, + 0.002275340724736452, + 0.0007449638796970248, + 0.0008892728365026414, + 0.0005363078089430928, + 0.0006943171028979123, + 0.001485840999521315, + 0.0036312530282884836, + 0.0020807269029319286, + 0.003382534021511674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.1270408183336258 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799004077911377 + ], + "opposite_logprob_tokens": [ + -1.8049004077911377 + ], + "logprob_reward": -0.1799004077911377, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799004077911377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45013371109962463, + 0.11366286128759384, + 0.14305329322814941, + 0.05675593018531799, + 0.033295292407274246, + 0.031384870409965515, + 0.026461249217391014, + 0.04377011954784393, + 0.01992136612534523, + 0.08156130462884903 + ], + "credit_map": [ + 0.052291158586740494, + 0.04958471655845642, + 0.07540597021579742, + 0.06471005827188492, + 0.06476931273937225, + 0.03772313520312309, + 0.06349454075098038, + 0.04215480014681816, + 0.009000095538794994, + 0.007464440539479256, + 0.019030606374144554, + 0.020882142707705498, + 0.008467675186693668, + 0.02757926471531391, + 0.011470713652670383, + 0.009767916053533554, + 0.004323950968682766, + 0.004417789168655872, + 0.0514785572886467, + 0.054468512535095215, + 0.006088155321776867, + 0.006085074506700039, + 0.008600007742643356, + 0.007591244764626026, + 0.006716227158904076, + 0.004778436850756407, + 0.005787660833448172, + 0.013126435689628124, + 0.005165405105799437, + 0.007448253221809864, + 0.007439634762704372, + 0.006293877959251404, + 0.003210681490600109, + 0.0028173616155982018, + 0.0036725483369082212, + 0.005049802828580141, + 0.004995625000447035, + 0.006345235742628574, + 0.0030407316517084837, + 0.004163309000432491, + 0.003096248023211956, + 0.002022906905040145, + 0.0027578198350965977, + 0.007778716739267111, + 0.0031646450515836477, + 0.0070658219046890736, + 0.0024658169131726027, + 0.0030328971333801746, + 0.003507730783894658, + 0.0019071056740358472, + 0.001555685419589281, + 0.0014972335193306208, + 0.0035043905954807997, + 0.007025128696113825, + 0.003495024284347892, + 0.003968952689319849, + 0.0043585593812167645, + 0.006135952193289995, + 0.004371340852230787, + 0.0035292203538119793, + 0.005694647319614887, + 0.007092209532856941, + 0.006210897583514452, + 0.00637729000300169, + 0.0019173116888850927, + 0.004139000084251165, + 0.0015996298752725124, + 0.0015108524821698666, + 0.001983162248507142, + 0.00362827954813838, + 0.002222259994596243, + 0.0029208699706941843, + 0.005179347936064005, + 0.0037059499882161617, + 0.0013074502348899841, + 0.0015143913915380836, + 0.0026698969304561615, + 0.031247811391949654, + 0.0049535660073161125, + 0.030982889235019684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.1799004077911377 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a low angle, slowly pulls back to reveal a wider scene, following the passenger's movements. At the center of the frame is a sleek, vibrantly colored cartoon yellow taxi. Its mirror-like surface perfectly reflects the surrounding light, while the \"TAXI\" lightbox on the roof emits a soft white glow. As the taxi's rear door swings gracefully upward, a young woman alights. She's wearing a stylish dark blue trench coat, the material subtly shimmering in the shifting light, and her long, curly chestnut hair sways gently with her movements. One foot cautiously touches the wet asphalt, creating a tiny ripple illuminated by the neon light. Then, she calmly steps out of the taxi, planted on the sidewalk's edge, a look of fresh-faced curiosity and wonder on her face. She stepped into a futuristic, bustling metropolis at night. Towering skyscrapers were dotted with giant, flashing digital billboards, and cascading neon signs painted the streets a vibrant sea of deep blue, psychedelic purple, and fiery pink. Complex, dynamic light and shadow flowed across the sleek yellow body of a cartoon taxi, casting ever-changing specks of color on the passenger's windbreaker. The resulting scene, a sophisticated, Pixar-quality 3D animation, was imbued with the urban energy, technological sophistication, and a touch of the dreamlike wonder of a first visit to the big city.", + "seed": 7363, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_1/index_1908_seed_7363_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10783027932047844, + "vlm_logprob": -0.10783027932047844, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.10783027932047844, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016691084951162338 + ], + "opposite_logprob_tokens": [ + -6.016691207885742 + ], + "logprob_reward": -0.016691084951162338, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.016691084951162338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2113533616065979, + 0.09831276535987854, + 0.06836465746164322, + 0.08282001316547394, + 0.09086441248655319, + 0.09119868278503418, + 0.06429776549339294, + 0.12050750106573105, + 0.07116025686264038, + 0.10112058371305466 + ], + "credit_map": [ + 0.03700265660881996, + 0.0434674434363842, + 0.021985186263918877, + 0.025942262262105942, + 0.023572593927383423, + 0.015672067180275917, + 0.019618511199951172, + 0.024092625826597214, + 0.010358721949160099, + 0.015518463216722012, + 0.012920254841446877, + 0.01320140715688467, + 0.011678004637360573, + 0.011689137667417526, + 0.011930622160434723, + 0.01101614534854889, + 0.00743075180798769, + 0.006613587494939566, + 0.01214953325688839, + 0.008660029619932175, + 0.0077706268057227135, + 0.00833451934158802, + 0.007839376106858253, + 0.009566228836774826, + 0.006769957486540079, + 0.00876904558390379, + 0.013469217345118523, + 0.012582109309732914, + 0.015034285373985767, + 0.0059965187683701515, + 0.008191008120775223, + 0.01200786978006363, + 0.00926982518285513, + 0.008313097059726715, + 0.02151903696358204, + 0.014203650876879692, + 0.0067184665240347385, + 0.006104584783315659, + 0.008624897338449955, + 0.016110852360725403, + 0.015321336686611176, + 0.008806407451629639, + 0.020455583930015564, + 0.018587233498692513, + 0.005963290575891733, + 0.005933863576501608, + 0.007958002388477325, + 0.008172962814569473, + 0.010716043412685394, + 0.007034304551780224, + 0.010001024231314659, + 0.014306564815342426, + 0.005791948176920414, + 0.004808057565242052, + 0.005107927601784468, + 0.006531893275678158, + 0.016340117901563644, + 0.026902617886662483, + 0.013921794481575489, + 0.021675337105989456, + 0.011370559222996235, + 0.012262637726962566, + 0.008688039146363735, + 0.009346387349069118, + 0.013081243261694908, + 0.008800564333796501, + 0.008235588669776917, + 0.014164764434099197, + 0.007813974283635616, + 0.007297627627849579, + 0.004800980444997549, + 0.0069655003026127815, + 0.020109670236706734, + 0.011600956320762634, + 0.012033228762447834, + 0.007280046120285988, + 0.018431929871439934, + 0.012785965576767921, + 0.008883146569132805, + 0.00999563280493021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.016691084951162338 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054969415068626404 + ], + "opposite_logprob_tokens": [ + -2.929969310760498 + ], + "logprob_reward": -0.054969415068626404, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054969415068626404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10291921347379684, + 0.04680182784795761, + 0.079975925385952, + 0.19186697900295258, + 0.07562804222106934, + 0.05106222257018089, + 0.05776514112949371, + 0.272748202085495, + 0.08841198682785034, + 0.03282046690583229 + ], + "credit_map": [ + 0.018688131123781204, + 0.017904097214341164, + 0.015017341822385788, + 0.012318556196987629, + 0.010462475940585136, + 0.007889431901276112, + 0.009742014110088348, + 0.010897164233028889, + 0.005486269481480122, + 0.005669937934726477, + 0.007242676801979542, + 0.008817440830171108, + 0.004931354895234108, + 0.0046578627079725266, + 0.004300201311707497, + 0.005696085747331381, + 0.009751364588737488, + 0.015343275852501392, + 0.01225682720541954, + 0.010826007463037968, + 0.009374214336276054, + 0.0067518833093345165, + 0.006082865409553051, + 0.009589483961462975, + 0.01250460371375084, + 0.02063356712460518, + 0.018551046028733253, + 0.024755604565143585, + 0.034848589450120926, + 0.030196886509656906, + 0.01870424672961235, + 0.031672440469264984, + 0.0065194303169846535, + 0.007667280733585358, + 0.012720974162220955, + 0.008328041061758995, + 0.009070269763469696, + 0.00586021738126874, + 0.009534664452075958, + 0.015927162021398544, + 0.007777976803481579, + 0.006225036922842264, + 0.009745613671839237, + 0.007511358708143234, + 0.0052543212659657, + 0.005360142327845097, + 0.004370753187686205, + 0.004817015025764704, + 0.010983938351273537, + 0.006701516453176737, + 0.007925010286271572, + 0.007524834480136633, + 0.006465364713221788, + 0.007444249000400305, + 0.0045611015520989895, + 0.006159125827252865, + 0.04159044846892357, + 0.06530969589948654, + 0.03549481928348541, + 0.039294611662626266, + 0.019185606390237808, + 0.0324111208319664, + 0.016896719112992287, + 0.02256515808403492, + 0.02194887213408947, + 0.014109961688518524, + 0.015403431840240955, + 0.013987469486892223, + 0.0073753888718783855, + 0.00785455945879221, + 0.00322863576002419, + 0.004503671545535326, + 0.006049636285752058, + 0.004595335107296705, + 0.0038713098037987947, + 0.004247110802680254, + 0.004550027195364237, + 0.003307310864329338, + 0.0026906421408057213, + 0.0035090947058051825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.054969415068626404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950592458248138 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950592458248138, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950592458248138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2504749596118927, + 0.16059339046478271, + 0.08354858309030533, + 0.08496768772602081, + 0.09254130721092224, + 0.04481789097189903, + 0.057789821177721024, + 0.06687711179256439, + 0.04596814513206482, + 0.11242111027240753 + ], + "credit_map": [ + 0.041671134531497955, + 0.05000058561563492, + 0.02814893238246441, + 0.031098613515496254, + 0.027786217629909515, + 0.018955575302243233, + 0.023911790922284126, + 0.028902120888233185, + 0.01742461882531643, + 0.02434437908232212, + 0.02328159101307392, + 0.02483278140425682, + 0.01695367321372032, + 0.01645548827946186, + 0.018544232472777367, + 0.018756616860628128, + 0.01200750470161438, + 0.008280382491648197, + 0.014238881878554821, + 0.00951892789453268, + 0.008811397477984428, + 0.009522291831672192, + 0.00863172672688961, + 0.012537464499473572, + 0.00871274247765541, + 0.00945022888481617, + 0.011453946121037006, + 0.010429663583636284, + 0.012052873149514198, + 0.009958356618881226, + 0.007510397117584944, + 0.015399482101202011, + 0.010526400990784168, + 0.007193402852863073, + 0.024673165753483772, + 0.010152834467589855, + 0.009252105839550495, + 0.006584384478628635, + 0.007939650677144527, + 0.016219360753893852, + 0.009426873177289963, + 0.0038611143827438354, + 0.01116347685456276, + 0.007458045147359371, + 0.003358688438311219, + 0.0030318633653223515, + 0.003240736201405525, + 0.003277095500379801, + 0.008555018343031406, + 0.005736319348216057, + 0.009453917853534222, + 0.010667027905583382, + 0.0050790682435035706, + 0.007385293021798134, + 0.004625030793249607, + 0.006288143340498209, + 0.01673862896859646, + 0.011537319049239159, + 0.008811219595372677, + 0.007139914203435183, + 0.004915434401482344, + 0.008382550440728664, + 0.003812242764979601, + 0.005539798177778721, + 0.007906110025942326, + 0.009405963122844696, + 0.00520595908164978, + 0.007002609316259623, + 0.004764470271766186, + 0.005718872416764498, + 0.0025424850173294544, + 0.003421676578000188, + 0.02323155105113983, + 0.00986932311207056, + 0.012657233513891697, + 0.008621295914053917, + 0.021064406260848045, + 0.01355522871017456, + 0.010553508996963501, + 0.012868568301200867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.005950592458248138 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291467398405075 + ], + "opposite_logprob_tokens": [ + -2.237914562225342 + ], + "logprob_reward": -0.11291467398405075, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291467398405075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07647819817066193, + 0.1078716591000557, + 0.09241565316915512, + 0.08875340223312378, + 0.13059407472610474, + 0.053368277847766876, + 0.13204817473888397, + 0.13759683072566986, + 0.14731793105602264, + 0.033555880188941956 + ], + "credit_map": [ + 0.010365867987275124, + 0.006736562587320805, + 0.013793251477181911, + 0.011289813555777073, + 0.007163550239056349, + 0.005753498058766127, + 0.010985516011714935, + 0.010390129871666431, + 0.01564229093492031, + 0.013440212234854698, + 0.012990005314350128, + 0.020862339064478874, + 0.010235212743282318, + 0.010416252538561821, + 0.011275033466517925, + 0.013010300695896149, + 0.011165032163262367, + 0.01447670441120863, + 0.014215732924640179, + 0.013960644602775574, + 0.01037590578198433, + 0.00809958390891552, + 0.008486137725412846, + 0.011635902337729931, + 0.007928646169602871, + 0.010032478719949722, + 0.011931885033845901, + 0.01186889223754406, + 0.01308982353657484, + 0.010335972532629967, + 0.01245457399636507, + 0.011111123487353325, + 0.011759170331060886, + 0.01108863577246666, + 0.023561649024486542, + 0.016279958188533783, + 0.01738908141851425, + 0.01190933771431446, + 0.014560875482857227, + 0.024045363068580627, + 0.005776943638920784, + 0.0039576259441673756, + 0.009881476871669292, + 0.01338654663413763, + 0.0047673070803284645, + 0.004246547352522612, + 0.006235003471374512, + 0.00511682266369462, + 0.016823366284370422, + 0.020849697291851044, + 0.016920028254389763, + 0.026706963777542114, + 0.01441892422735691, + 0.0131340641528368, + 0.01023979764431715, + 0.012955332174897194, + 0.02500978112220764, + 0.019373254850506783, + 0.025847280398011208, + 0.0156874880194664, + 0.016298338770866394, + 0.016580121591687202, + 0.008556954562664032, + 0.01024358905851841, + 0.024481458589434624, + 0.028648732230067253, + 0.020136134698987007, + 0.018758567050099373, + 0.016962390393018723, + 0.013884109444916248, + 0.00981201883405447, + 0.014634495601058006, + 0.006364530883729458, + 0.003550749970600009, + 0.004355139099061489, + 0.003917320165783167, + 0.004380008205771446, + 0.0033095006365329027, + 0.002823749789968133, + 0.004854877013713121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.11291467398405075 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrantly colored cartoon yellow taxi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a young woman wearing a dark blue trench coat exit the taxi through the rear door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic metropolis at night with towering skyscrapers and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the taxi's roof lightbox emitting a soft white glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34862563014030457 + ], + "opposite_logprob_tokens": [ + -1.223625659942627 + ], + "logprob_reward": -0.34862563014030457, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34862563014030457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10655224323272705, + 0.10616263747215271, + 0.04803108051419258, + 0.10095586627721786, + 0.04835330322384834, + 0.10166166722774506, + 0.03570633754134178, + 0.06618032604455948, + 0.045462653040885925, + 0.3409338593482971 + ], + "credit_map": [ + 0.01310974545776844, + 0.012235026806592941, + 0.014891425147652626, + 0.013830860145390034, + 0.010536378249526024, + 0.011104696430265903, + 0.015826186165213585, + 0.015017935074865818, + 0.012341555207967758, + 0.016938434913754463, + 0.010244078934192657, + 0.01668683812022209, + 0.012857387773692608, + 0.009087245911359787, + 0.01382173877209425, + 0.014185361564159393, + 0.006795467808842659, + 0.007528490386903286, + 0.007076031994074583, + 0.005436969920992851, + 0.0057432567700743675, + 0.004985195118933916, + 0.0047567314468324184, + 0.0057089379988610744, + 0.008076761849224567, + 0.010717308148741722, + 0.011871474795043468, + 0.017688000574707985, + 0.015507173724472523, + 0.009569498710334301, + 0.015472931787371635, + 0.012052717618644238, + 0.004213409963995218, + 0.005363886710256338, + 0.003635072847828269, + 0.007790237665176392, + 0.007725312374532223, + 0.0052845110185444355, + 0.007236860226839781, + 0.00710401451215148, + 0.01630437932908535, + 0.013584913685917854, + 0.02076495997607708, + 0.017079461365938187, + 0.01056598499417305, + 0.0072403582744300365, + 0.008976567536592484, + 0.007145043928176165, + 0.004543944727629423, + 0.0034843774046748877, + 0.004951956681907177, + 0.008870107121765614, + 0.003360255155712366, + 0.0025688305031508207, + 0.0027771848253905773, + 0.005149682052433491, + 0.010802124626934528, + 0.013448424637317657, + 0.011614833027124405, + 0.009033845737576485, + 0.0053610424511134624, + 0.007916427217423916, + 0.0035075629130005836, + 0.004496067296713591, + 0.0054102991707623005, + 0.013607887551188469, + 0.005243431311100721, + 0.007199457380920649, + 0.004818696063011885, + 0.0040088617242872715, + 0.0024483941961079836, + 0.0027256268076598644, + 0.10752727091312408, + 0.03841947391629219, + 0.04668300226330757, + 0.02678978070616722, + 0.04084334895014763, + 0.04371101036667824, + 0.018594760447740555, + 0.018365202471613884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.34862563014030457 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_10/index_tensor([9373, 9373])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_10/index_tensor([9373, 9373])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..950f3a3a5eb74e9cbecbf2fc4314cda4b7f214c9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_10/index_tensor([9373, 9373])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, tight medium shot depicts the massive trunk of a majestic nanmu tree, its growth captured in a slow, mesmerizing time-lapse. The camera remains perfectly still, allowing the eye to absorb the intricate golden bark—a complex tapestry of deep vertical cracks and a delicate network of fissures, with a subtle, waxy sheen. As time passes, the massive trunk thickens noticeably and almost imperceptibly, its surface expanding and reorganizing in a silent testament to the passing seasons. A vast, soft-edged shadow, cast from the dense canopy above, drifts steadily and deliberately through the woods, its slow flight highlighting and obscuring rich textures in a rhythmic, hypnotic dance. The scene is bathed in the warm, dappled light of the forest, creating a dominant hue of golden yellow and earthy brown, contrasting sharply with the cool, deep tones of the shifting shadows. The overall atmosphere is one of profound calm and contemplation, a contemplative and visually rich study of silent, powerful growth.", + "seed": 7345, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_10/index_9373_seed_7345_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3720579557179008, + "vlm_logprob": -0.3720579557179008, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.3720579557179008, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021299084648489952 + ], + "opposite_logprob_tokens": [ + -4.146298885345459 + ], + "logprob_reward": -0.021299084648489952, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.021299084648489952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09196623414754868, + 0.07942431420087814, + 0.021588226780295372, + 0.08207854628562927, + 0.03355240821838379, + 0.019612032920122147, + 0.03145606443285942, + 0.028815338388085365, + 0.11107475310564041, + 0.5004320740699768 + ], + "credit_map": [ + 0.019929608330130577, + 0.00950597133487463, + 0.016646571457386017, + 0.007253674790263176, + 0.01055917888879776, + 0.010164650157094002, + 0.007781216874718666, + 0.010125355795025826, + 0.008566674776375294, + 0.0070377979427576065, + 0.008731449022889137, + 0.009993346408009529, + 0.011950012296438217, + 0.00873266439884901, + 0.013836177065968513, + 0.010576194152235985, + 0.0023351628333330154, + 0.002310999669134617, + 0.0025127441622316837, + 0.002938798861578107, + 0.001997043378651142, + 0.004427243024110794, + 0.002334535587579012, + 0.002731697866693139, + 0.0071749440394341946, + 0.025065544992685318, + 0.005676713772118092, + 0.006420882418751717, + 0.009540618397295475, + 0.004914910066872835, + 0.007655664812773466, + 0.015629272907972336, + 0.0034198721405118704, + 0.005101836286485195, + 0.0031357035040855408, + 0.0032399878837168217, + 0.0035310350358486176, + 0.004830813501030207, + 0.004134948831051588, + 0.006158210337162018, + 0.0018311664462089539, + 0.0032991706393659115, + 0.0026791724376380444, + 0.0022827396169304848, + 0.0021281049121171236, + 0.0024327505379915237, + 0.003510482609272003, + 0.0014484458370134234, + 0.004370836541056633, + 0.003024191362783313, + 0.006779336370527744, + 0.0031671838369220495, + 0.003274043556302786, + 0.004635849967598915, + 0.0036678181495517492, + 0.002536807442083955, + 0.004571968223899603, + 0.0024116646964102983, + 0.005409564822912216, + 0.003547027939930558, + 0.0029938106890767813, + 0.004240216221660376, + 0.0028140165377408266, + 0.002827068790793419, + 0.008059175685048103, + 0.018922442570328712, + 0.018702734261751175, + 0.017999086529016495, + 0.009473484009504318, + 0.010351764038205147, + 0.017615390941500664, + 0.009950673207640648, + 0.038721781224012375, + 0.034183401614427567, + 0.07571244984865189, + 0.025972304865717888, + 0.029987569898366928, + 0.05746832862496376, + 0.058441441506147385, + 0.17994481325149536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.021299084648489952 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.974295973777771 + ], + "opposite_logprob_tokens": [ + -0.4742959439754486 + ], + "logprob_reward": -0.974295973777771, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.974295973777771, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1960640698671341, + 0.13704754412174225, + 0.027280466631054878, + 0.12007921934127808, + 0.10399382561445236, + 0.0285982433706522, + 0.030637886375188828, + 0.06538496911525726, + 0.15875093638896942, + 0.13216280937194824 + ], + "credit_map": [ + 0.02863120287656784, + 0.01582115702331066, + 0.05827440693974495, + 0.015214387327432632, + 0.012631216086447239, + 0.02036074362695217, + 0.013577382080256939, + 0.03155358508229256, + 0.01107033435255289, + 0.010108783841133118, + 0.018400751054286957, + 0.02715168707072735, + 0.019672710448503494, + 0.015592815354466438, + 0.019418606534600258, + 0.01563185453414917, + 0.0034613395109772682, + 0.002988555235788226, + 0.0032469455618411303, + 0.004234788473695517, + 0.0030211086850613356, + 0.0035482279490679502, + 0.0038133179768919945, + 0.0029661841690540314, + 0.008664955385029316, + 0.031308554112911224, + 0.006701159756630659, + 0.009936612099409103, + 0.01693801023066044, + 0.008597669191658497, + 0.012261784635484219, + 0.02567048743367195, + 0.008608508855104446, + 0.025418497622013092, + 0.007778502535074949, + 0.013126634992659092, + 0.013832801952958107, + 0.012907013297080994, + 0.011989746242761612, + 0.010332128033041954, + 0.004880815744400024, + 0.003931762184947729, + 0.0029202711302787066, + 0.003138786880299449, + 0.002721150405704975, + 0.004005815833806992, + 0.0035546827130019665, + 0.0034449612721800804, + 0.002277197316288948, + 0.0061339884996414185, + 0.0042741005308926105, + 0.003392499405890703, + 0.0038188269827514887, + 0.004091502167284489, + 0.0037360121496021748, + 0.002913761418312788, + 0.010293000377714634, + 0.0069216410629451275, + 0.011216643266379833, + 0.009209955111145973, + 0.005983353592455387, + 0.008425064384937286, + 0.007459286134690046, + 0.005876022856682539, + 0.015734700486063957, + 0.019017526879906654, + 0.016294946894049644, + 0.019523413851857185, + 0.018872469663619995, + 0.024692142382264137, + 0.025075692683458328, + 0.019540049135684967, + 0.014426532201468945, + 0.0222527626901865, + 0.01007839385420084, + 0.012715629301965237, + 0.019349684938788414, + 0.015416616573929787, + 0.01852150820195675, + 0.019401704892516136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.974295973777771 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812529891729355 + ], + "opposite_logprob_tokens": [ + -3.2881252765655518 + ], + "logprob_reward": -0.03812529891729355, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812529891729355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07326187193393707, + 0.04605511575937271, + 0.044903047382831573, + 0.04408634826540947, + 0.022680187597870827, + 0.020486539229750633, + 0.048946280032396317, + 0.04747997969388962, + 0.5663562417030334, + 0.08574442565441132 + ], + "credit_map": [ + 0.01760271191596985, + 0.0060545289888978004, + 0.013994131237268448, + 0.003907958511263132, + 0.005007031839340925, + 0.007671641185879707, + 0.006098918151110411, + 0.012924945913255215, + 0.004412524402141571, + 0.003660462098196149, + 0.005698913708329201, + 0.01091914251446724, + 0.004984590224921703, + 0.004568380769342184, + 0.006132876500487328, + 0.005678222980350256, + 0.0034937409218400717, + 0.0033403090201318264, + 0.004908980336040258, + 0.007490723859518766, + 0.006301555782556534, + 0.005219278857111931, + 0.010165111161768436, + 0.003983344417065382, + 0.0025817444548010826, + 0.009800507687032223, + 0.0023719388991594315, + 0.0036329312715679407, + 0.007178005296736956, + 0.0030730455182492733, + 0.005195149686187506, + 0.01025302056223154, + 0.0016836783615872264, + 0.005176248028874397, + 0.0016328739002346992, + 0.003019625786691904, + 0.0024411685299128294, + 0.0020690690726041794, + 0.003839653916656971, + 0.0028178682550787926, + 0.00196851696819067, + 0.00254275044426322, + 0.002939448691904545, + 0.0024060106370598078, + 0.0023625728208571672, + 0.002406521001830697, + 0.00359616382047534, + 0.002264553215354681, + 0.0040882909670472145, + 0.009913971647620201, + 0.006229266989976168, + 0.006196120288223028, + 0.008130955509841442, + 0.00535691948607564, + 0.005506273824721575, + 0.0035244780592620373, + 0.005253144074231386, + 0.005438840948045254, + 0.007060867268592119, + 0.008214354515075684, + 0.004688171204179525, + 0.006654120981693268, + 0.006628843955695629, + 0.0035416320897638798, + 0.052549395710229874, + 0.07505419850349426, + 0.06319839507341385, + 0.08126223087310791, + 0.06321799010038376, + 0.06593996286392212, + 0.1181073933839798, + 0.04702664166688919, + 0.010564352385699749, + 0.012373583391308784, + 0.011166535317897797, + 0.011351102031767368, + 0.010651234537363052, + 0.006967105437070131, + 0.010078875347971916, + 0.01259162463247776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.03812529891729355 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003906917118001729 + ], + "opposite_logprob_tokens": [ + -7.875390529632568 + ], + "logprob_reward": -0.0003906917118001729, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.0003906917118001729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12948033213615417, + 0.1162300705909729, + 0.028981288895010948, + 0.1382572203874588, + 0.07257704436779022, + 0.034211136400699615, + 0.03214222192764282, + 0.04330431669950485, + 0.1725003868341446, + 0.23231597244739532 + ], + "credit_map": [ + 0.02503136917948723, + 0.013921096920967102, + 0.024244248867034912, + 0.01938924379646778, + 0.013630421832203865, + 0.011309041641652584, + 0.010279200971126556, + 0.011675719171762466, + 0.01177790854126215, + 0.008311128243803978, + 0.013284832239151001, + 0.026081165298819542, + 0.014396903105080128, + 0.013546531088650227, + 0.015498730354011059, + 0.013332866132259369, + 0.002195750828832388, + 0.0025873142294585705, + 0.003240190213546157, + 0.004395213909447193, + 0.004515433683991432, + 0.004072121810168028, + 0.0055823298171162605, + 0.0023929341696202755, + 0.010031179524958134, + 0.04180428758263588, + 0.005911277141422033, + 0.00871384795755148, + 0.02044869028031826, + 0.008233525790274143, + 0.010789226740598679, + 0.03232519328594208, + 0.005788417998701334, + 0.019393669441342354, + 0.005629366729408503, + 0.009294570423662663, + 0.008162363432347775, + 0.008804628625512123, + 0.008826149627566338, + 0.006677884142845869, + 0.003333362517878413, + 0.0060936324298381805, + 0.0034079563338309526, + 0.0035322715993970633, + 0.003450847463682294, + 0.005159738007932901, + 0.004834604915231466, + 0.004398723598569632, + 0.0030167754739522934, + 0.005931604653596878, + 0.004308972507715225, + 0.0037681437097489834, + 0.004495162982493639, + 0.004368991125375032, + 0.003500391496345401, + 0.0027521802112460136, + 0.005299027543514967, + 0.00453398609533906, + 0.007759769447147846, + 0.008446375839412212, + 0.004143937490880489, + 0.004302157089114189, + 0.0055968803353607655, + 0.0032221830915659666, + 0.016543127596378326, + 0.021960796788334846, + 0.01989126019179821, + 0.02042456716299057, + 0.01925063133239746, + 0.019812926650047302, + 0.040791261941194534, + 0.013825811445713043, + 0.018691347911953926, + 0.019202368333935738, + 0.035044100135564804, + 0.013164260424673557, + 0.019281774759292603, + 0.025179317221045494, + 0.026933718472719193, + 0.07481908798217773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.0003906917118001729 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261787295341492 + ], + "opposite_logprob_tokens": [ + -0.5761787295341492 + ], + "logprob_reward": -0.8261787295341492, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261787295341492, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0643744096159935, + 0.09736276417970657, + 0.040427424013614655, + 0.1060183122754097, + 0.03262699395418167, + 0.03136172890663147, + 0.1498785763978958, + 0.05316074937582016, + 0.2770319879055023, + 0.14775706827640533 + ], + "credit_map": [ + 0.014349660836160183, + 0.00470508448779583, + 0.017917625606060028, + 0.004553264006972313, + 0.0052451323717832565, + 0.005649006459861994, + 0.005729008000344038, + 0.006225623656064272, + 0.009681534953415394, + 0.007582783233374357, + 0.010656539350748062, + 0.021188342943787575, + 0.013150261715054512, + 0.009265471249818802, + 0.014733475632965565, + 0.011104357428848743, + 0.00410905946046114, + 0.003457054030150175, + 0.004067644942551851, + 0.005759786814451218, + 0.005847158841788769, + 0.006790652871131897, + 0.006936226971447468, + 0.0034598365891724825, + 0.00865156389772892, + 0.03185972198843956, + 0.007134721614420414, + 0.006955548655241728, + 0.011528591625392437, + 0.006318149622529745, + 0.0077499933540821075, + 0.02582002803683281, + 0.0023792211432009935, + 0.007223003543913364, + 0.003019248601049185, + 0.00365053815767169, + 0.004107429180294275, + 0.0038659670390188694, + 0.004186039324849844, + 0.004195545334368944, + 0.0031765131279826164, + 0.004913195036351681, + 0.004547764081507921, + 0.00354769523255527, + 0.0029046260751783848, + 0.00410083681344986, + 0.005468911956995726, + 0.002702186116948724, + 0.02128099463880062, + 0.014900241047143936, + 0.03087466023862362, + 0.01775054633617401, + 0.016399629414081573, + 0.015699751675128937, + 0.019634027034044266, + 0.013338737189769745, + 0.0067990561947226524, + 0.005361460614949465, + 0.009139600209891796, + 0.007699606474488974, + 0.004862739238888025, + 0.007248684298247099, + 0.005973690189421177, + 0.0060759130865335464, + 0.02222765050828457, + 0.03838125616312027, + 0.04385776445269585, + 0.040071431547403336, + 0.026888158172369003, + 0.027605026960372925, + 0.05888286605477333, + 0.019117839634418488, + 0.017968816682696342, + 0.024310346692800522, + 0.014227992855012417, + 0.012860780581831932, + 0.023511309176683426, + 0.013819107785820961, + 0.019641321152448654, + 0.021417388692498207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.8261787295341492 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, tight medium shot depicts the massive trunk of a majestic nanmu tree, its growth captured in a slow, mesmerizing time-lapse. The camera remains perfectly still, allowing the eye to absorb the intricate golden bark—a complex tapestry of deep vertical cracks and a delicate network of fissures, with a subtle, waxy sheen. As time passes, the massive trunk thickens noticeably and almost imperceptibly, its surface expanding and reorganizing in a silent testament to the passing seasons. A vast, soft-edged shadow, cast from the dense canopy above, drifts steadily and deliberately through the woods, its slow flight highlighting and obscuring rich textures in a rhythmic, hypnotic dance. The scene is bathed in the warm, dappled light of the forest, creating a dominant hue of golden yellow and earthy brown, contrasting sharply with the cool, deep tones of the shifting shadows. The overall atmosphere is one of profound calm and contemplation, a contemplative and visually rich study of silent, powerful growth.", + "seed": 7346, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_10/index_9373_seed_7346_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.276450591799221, + "vlm_logprob": -0.276450591799221, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.276450591799221, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009272960014641285 + ], + "opposite_logprob_tokens": [ + -5.384273052215576 + ], + "logprob_reward": -0.009272960014641285, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.009272960014641285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16005820035934448, + 0.0790853425860405, + 0.06538774073123932, + 0.42420169711112976, + 0.03501411899924278, + 0.03676733002066612, + 0.023591767996549606, + 0.10688825696706772, + 0.02971571311354637, + 0.03928981348872185 + ], + "credit_map": [ + 0.02580283023416996, + 0.012559891678392887, + 0.018208039924502373, + 0.017620131373405457, + 0.01908979006111622, + 0.03450525552034378, + 0.014497732743620872, + 0.017774514853954315, + 0.011409681290388107, + 0.01208101212978363, + 0.011702403426170349, + 0.009407789446413517, + 0.008450362831354141, + 0.010601885616779327, + 0.006662972737103701, + 0.008769232779741287, + 0.008360443636775017, + 0.013562090694904327, + 0.009766013361513615, + 0.006807796191424131, + 0.010023116134107113, + 0.006540623959153891, + 0.005501171108335257, + 0.004826482385396957, + 0.0556982196867466, + 0.16523624956607819, + 0.02019251510500908, + 0.021907063201069832, + 0.11234869062900543, + 0.016485806554555893, + 0.0204890426248312, + 0.011844162829220295, + 0.003727167146280408, + 0.003775795688852668, + 0.0038936850614845753, + 0.00394342839717865, + 0.004720739554613829, + 0.006263578776270151, + 0.004469000734388828, + 0.0042207203805446625, + 0.003815180854871869, + 0.004419736098498106, + 0.004891295451670885, + 0.003546212101355195, + 0.003934609703719616, + 0.004503823351114988, + 0.006701818201690912, + 0.004954655189067125, + 0.0031226070132106543, + 0.003637957852333784, + 0.00212139543145895, + 0.00271543744020164, + 0.0025655964855104685, + 0.0046375822275877, + 0.002640838036313653, + 0.0021503539755940437, + 0.025520388036966324, + 0.009400465525686741, + 0.005488353781402111, + 0.0068373531103134155, + 0.006270184181630611, + 0.01063967403024435, + 0.011896420270204544, + 0.030835414305329323, + 0.0060172551311552525, + 0.0032241118606179953, + 0.0018986533395946026, + 0.004213642328977585, + 0.0024246887769550085, + 0.0040346961468458176, + 0.004889748990535736, + 0.003012913977727294, + 0.005999618209898472, + 0.005325980484485626, + 0.003876072820276022, + 0.00407356396317482, + 0.007062666118144989, + 0.004235340282320976, + 0.0049882326275110245, + 0.003728339448571205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.009272960014641285 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233572125434875 + ], + "opposite_logprob_tokens": [ + -0.8983572125434875 + ], + "logprob_reward": -0.5233572125434875, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233572125434875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11946094781160355, + 0.10940823704004288, + 0.12580250203609467, + 0.056515686213970184, + 0.04157152399420738, + 0.07778237760066986, + 0.052090637385845184, + 0.24394458532333374, + 0.05366962030529976, + 0.11975385993719101 + ], + "credit_map": [ + 0.01942119188606739, + 0.011712532490491867, + 0.013304515741765499, + 0.01459644827991724, + 0.021654924377799034, + 0.017963357269763947, + 0.010513347573578358, + 0.010294628329575062, + 0.012978080660104752, + 0.020347191020846367, + 0.012250959873199463, + 0.01153573114424944, + 0.017459958791732788, + 0.009420596994459629, + 0.01628926955163479, + 0.009126455523073673, + 0.016265347599983215, + 0.021962938830256462, + 0.015622609294950962, + 0.01687287911772728, + 0.01769712008535862, + 0.014067883603274822, + 0.011128668673336506, + 0.012185068801045418, + 0.0069037145003676414, + 0.018557284027338028, + 0.004016325809061527, + 0.003948438446968794, + 0.012666633352637291, + 0.0045876712538301945, + 0.0034253548365086317, + 0.0024102688767015934, + 0.0056188078597188, + 0.0049046906642615795, + 0.008254225365817547, + 0.005668663419783115, + 0.004553616512566805, + 0.005721724592149258, + 0.004494538530707359, + 0.0023552619386464357, + 0.008765105158090591, + 0.008632746525108814, + 0.015086641535162926, + 0.00910925306379795, + 0.009069963358342648, + 0.008450226858258247, + 0.011572370305657387, + 0.007096080109477043, + 0.004493658896535635, + 0.007592032663524151, + 0.00524000683799386, + 0.007126554846763611, + 0.0075671058148145676, + 0.005934468004852533, + 0.009795637801289558, + 0.004341179504990578, + 0.06890837848186493, + 0.01920199953019619, + 0.014170280657708645, + 0.011638421565294266, + 0.01138453371822834, + 0.01978786662220955, + 0.02062813565135002, + 0.07822498679161072, + 0.005384036805480719, + 0.007264773361384869, + 0.005561142694205046, + 0.008085285313427448, + 0.005606011021882296, + 0.006787601392716169, + 0.011148220859467983, + 0.0038325537461787462, + 0.009113880805671215, + 0.019157497212290764, + 0.009889965876936913, + 0.021413691341876984, + 0.01982455886900425, + 0.012435901910066605, + 0.019363150000572205, + 0.008555215783417225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.5233572125434875 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302864283323288 + ], + "opposite_logprob_tokens": [ + -3.7733027935028076 + ], + "logprob_reward": -0.023302864283323288, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302864283323288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06914567202329636, + 0.1485949009656906, + 0.05469641461968422, + 0.192909374833107, + 0.1237550601363182, + 0.029352013021707535, + 0.018902767449617386, + 0.0950273871421814, + 0.0179178174585104, + 0.24969862401485443 + ], + "credit_map": [ + 0.011928648687899113, + 0.008157068863511086, + 0.009227396920323372, + 0.008881470188498497, + 0.01390482671558857, + 0.006927581503987312, + 0.005408792290836573, + 0.004709886386990547, + 0.019485341385006905, + 0.021715929731726646, + 0.016479704529047012, + 0.024904072284698486, + 0.02296956442296505, + 0.0120738809928298, + 0.01865447871387005, + 0.012311925180256367, + 0.008149528875946999, + 0.01043130736798048, + 0.005279490724205971, + 0.006601580418646336, + 0.007373823318630457, + 0.005978173576295376, + 0.006906147114932537, + 0.003976361360400915, + 0.024576283991336823, + 0.073765330016613, + 0.011453144252300262, + 0.010850798338651657, + 0.050026558339595795, + 0.007306225132197142, + 0.009526005014777184, + 0.005405035801231861, + 0.017630252987146378, + 0.012300584465265274, + 0.02847198024392128, + 0.013632607646286488, + 0.012061544694006443, + 0.017850695177912712, + 0.010617570951581001, + 0.011189821176230907, + 0.0026975006330758333, + 0.0033407839946448803, + 0.004500084090977907, + 0.0036894476506859064, + 0.004357949364930391, + 0.003308434970676899, + 0.004413453862071037, + 0.0030443600844591856, + 0.002049850532785058, + 0.003031705506145954, + 0.0024596338625997305, + 0.002309973817318678, + 0.0023294503334909678, + 0.002031164476647973, + 0.0029981019906699657, + 0.0016928863478824496, + 0.02265823818743229, + 0.007812159601598978, + 0.008619643747806549, + 0.00556084793061018, + 0.005644159857183695, + 0.011934897862374783, + 0.008261444047093391, + 0.024535998702049255, + 0.0036111006047576666, + 0.0015073131071403623, + 0.0025973001029342413, + 0.0018531985115259886, + 0.001280744792893529, + 0.0022308335173875093, + 0.0028512142598628998, + 0.001986113376915455, + 0.018583185970783234, + 0.04311693087220192, + 0.012837073765695095, + 0.07621458917856216, + 0.025249728932976723, + 0.017019880935549736, + 0.0318019725382328, + 0.02487526834011078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.023302864283323288 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00014685506175737828 + ], + "opposite_logprob_tokens": [ + -8.875146865844727 + ], + "logprob_reward": -0.00014685506175737828, + "yes_no_margin_reward": 8.875, + "token_credit_reward": -0.00014685506175737828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11226508766412735, + 0.07624594122171402, + 0.10912790149450302, + 0.21935135126113892, + 0.05357363820075989, + 0.031650397926568985, + 0.04541364312171936, + 0.2011261284351349, + 0.06801247596740723, + 0.08323344588279724 + ], + "credit_map": [ + 0.0176327433437109, + 0.00954398512840271, + 0.01672440953552723, + 0.013865785673260689, + 0.012820745818316936, + 0.019094128161668777, + 0.010820663534104824, + 0.011762626469135284, + 0.010450601577758789, + 0.012098977342247963, + 0.00947478972375393, + 0.007966864854097366, + 0.009559517726302147, + 0.010143094696104527, + 0.008893590420484543, + 0.007658503483980894, + 0.014924638904631138, + 0.014517465606331825, + 0.011591272428631783, + 0.01297687366604805, + 0.014901760034263134, + 0.016917632892727852, + 0.01144535094499588, + 0.01185290515422821, + 0.028314195573329926, + 0.08515363931655884, + 0.010803569108247757, + 0.011582955718040466, + 0.057914696633815765, + 0.008605868555605412, + 0.010560811497271061, + 0.006415603682398796, + 0.007161141838878393, + 0.0077950553968548775, + 0.007955033332109451, + 0.004395616706460714, + 0.005938164424151182, + 0.008431030437350273, + 0.0051852683536708355, + 0.006712324917316437, + 0.0035608357284218073, + 0.0035420882049947977, + 0.004491361323744059, + 0.0032123676501214504, + 0.0034450709354132414, + 0.00497273076325655, + 0.005306857638061047, + 0.0031190866138786077, + 0.00502049271017313, + 0.006709157954901457, + 0.0038833545986562967, + 0.0045265317894518375, + 0.005213047843426466, + 0.007478954270482063, + 0.007673186250030994, + 0.004908915609121323, + 0.05411234498023987, + 0.014356188476085663, + 0.010750088840723038, + 0.011119076982140541, + 0.00950480904430151, + 0.018828047439455986, + 0.017826076596975327, + 0.06462949514389038, + 0.00784704927355051, + 0.008578320033848286, + 0.006973889656364918, + 0.010900157503783703, + 0.006004021503031254, + 0.006894782651215792, + 0.016077004373073578, + 0.00473724864423275, + 0.009616564959287643, + 0.014026114717125893, + 0.006573867052793503, + 0.012912469916045666, + 0.013350599445402622, + 0.008522641845047474, + 0.011771906167268753, + 0.006459283642470837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.00014685506175737828 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261730670928955 + ], + "opposite_logprob_tokens": [ + -0.5761730670928955 + ], + "logprob_reward": -0.8261730670928955, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261730670928955, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.031783174723386765, + 0.061632268130779266, + 0.027970006689429283, + 0.2606651484966278, + 0.08208002150058746, + 0.033455848693847656, + 0.05540205165743828, + 0.3553932011127472, + 0.0178542323410511, + 0.07376407831907272 + ], + "credit_map": [ + 0.0052419123239815235, + 0.002809093566611409, + 0.006784193683415651, + 0.0029233722016215324, + 0.003675782587379217, + 0.004407620523124933, + 0.003368793986737728, + 0.0025724072474986315, + 0.007207584101706743, + 0.009215264581143856, + 0.007220636587589979, + 0.007170003838837147, + 0.008512369357049465, + 0.007173897232860327, + 0.01020118873566389, + 0.004931324627250433, + 0.003660490270704031, + 0.004702984821051359, + 0.003567512845620513, + 0.003020005766302347, + 0.004160742741078138, + 0.003507218323647976, + 0.002348940819501877, + 0.0030021113343536854, + 0.032875362783670425, + 0.09865177422761917, + 0.015098292380571365, + 0.015692654997110367, + 0.06705952435731888, + 0.010032350197434425, + 0.013518012128770351, + 0.007737172767519951, + 0.011708030477166176, + 0.009516938589513302, + 0.009291301481425762, + 0.0067944652400910854, + 0.008928406052291393, + 0.015903998166322708, + 0.008816490881145, + 0.011120391078293324, + 0.004356193356215954, + 0.0031213623005896807, + 0.008722866885364056, + 0.0032925636041909456, + 0.0041831680573523045, + 0.004749440588057041, + 0.0030333721078932285, + 0.0019968824926763773, + 0.00713999941945076, + 0.010207677260041237, + 0.006494287401437759, + 0.00470479391515255, + 0.0066070291213691235, + 0.00750422989949584, + 0.00750455167144537, + 0.005239479709416628, + 0.0981212630867958, + 0.03358092159032822, + 0.018285144120454788, + 0.01779865100979805, + 0.017535168677568436, + 0.0276629701256752, + 0.033339723944664, + 0.1090693324804306, + 0.002095848089084029, + 0.002355808624997735, + 0.001658453606069088, + 0.0026063770055770874, + 0.0018227669643238187, + 0.0017510928446426988, + 0.003980874549597502, + 0.0015830090269446373, + 0.005512435454875231, + 0.011642605997622013, + 0.0057464404962956905, + 0.013396135531365871, + 0.012305482290685177, + 0.0077112410217523575, + 0.01159656047821045, + 0.0058531854301691055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.8261730670928955 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, tight medium shot depicts the massive trunk of a majestic nanmu tree, its growth captured in a slow, mesmerizing time-lapse. The camera remains perfectly still, allowing the eye to absorb the intricate golden bark—a complex tapestry of deep vertical cracks and a delicate network of fissures, with a subtle, waxy sheen. As time passes, the massive trunk thickens noticeably and almost imperceptibly, its surface expanding and reorganizing in a silent testament to the passing seasons. A vast, soft-edged shadow, cast from the dense canopy above, drifts steadily and deliberately through the woods, its slow flight highlighting and obscuring rich textures in a rhythmic, hypnotic dance. The scene is bathed in the warm, dappled light of the forest, creating a dominant hue of golden yellow and earthy brown, contrasting sharply with the cool, deep tones of the shifting shadows. The overall atmosphere is one of profound calm and contemplation, a contemplative and visually rich study of silent, powerful growth.", + "seed": 7347, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_10/index_9373_seed_7347_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.32751193552976476, + "vlm_logprob": -0.32751193552976476, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.32751193552976476, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0482485331594944 + ], + "opposite_logprob_tokens": [ + -3.173248529434204 + ], + "logprob_reward": -0.0482485331594944, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0482485331594944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6146044135093689, + 0.05069350451231003, + 0.08395351469516754, + 0.08302922546863556, + 0.03111368417739868, + 0.018418919295072556, + 0.03186196833848953, + 0.023613620549440384, + 0.016665278002619743, + 0.04604590684175491 + ], + "credit_map": [ + 0.07654479146003723, + 0.04811989143490791, + 0.084296315908432, + 0.09032107144594193, + 0.10615631937980652, + 0.047432757914066315, + 0.10636834055185318, + 0.05536489933729172, + 0.0058527979999780655, + 0.006434448063373566, + 0.0069452556781470776, + 0.007812844589352608, + 0.005433683283627033, + 0.004693979397416115, + 0.007798612117767334, + 0.005721881054341793, + 0.009951289743185043, + 0.0051217940635979176, + 0.009936102665960789, + 0.005382899660617113, + 0.022577820345759392, + 0.012786388397216797, + 0.011412960477173328, + 0.006784252822399139, + 0.007750750053673983, + 0.00338146835565567, + 0.012209692038595676, + 0.0044230199418962, + 0.012369341216981411, + 0.02855289727449417, + 0.006958955433219671, + 0.007383101619780064, + 0.004379560239613056, + 0.0026132790371775627, + 0.003280766075477004, + 0.004602307453751564, + 0.0035675312392413616, + 0.004583981819450855, + 0.004680807702243328, + 0.0034054480493068695, + 0.00381729518994689, + 0.0018834456568583846, + 0.0020323495846241713, + 0.002105451188981533, + 0.0017062920378521085, + 0.001469285343773663, + 0.0027614186983555555, + 0.0026433817110955715, + 0.004467819817364216, + 0.002949465997517109, + 0.0032669510692358017, + 0.004266317468136549, + 0.0030942687299102545, + 0.004362915176898241, + 0.005188078619539738, + 0.004266148898750544, + 0.002458915812894702, + 0.0018929195357486606, + 0.00247320462949574, + 0.003368023782968521, + 0.002309581730514765, + 0.0035011363215744495, + 0.004146331921219826, + 0.003463504835963249, + 0.0026862553786486387, + 0.0014873011969029903, + 0.0026397884357720613, + 0.0016929283738136292, + 0.0016811207169666886, + 0.001818115939386189, + 0.0024188526440411806, + 0.002240914385765791, + 0.0076835425570607185, + 0.0046315426006913185, + 0.007177772931754589, + 0.005416055675595999, + 0.005734889768064022, + 0.004571627825498581, + 0.005693646613508463, + 0.0051368242129683495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0482485331594944 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742961525917053 + ], + "opposite_logprob_tokens": [ + -0.4742961525917053 + ], + "logprob_reward": -0.9742961525917053, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742961525917053, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.35838592052459717, + 0.08602822571992874, + 0.10674522817134857, + 0.21581204235553741, + 0.04905063658952713, + 0.040419403463602066, + 0.03296714648604393, + 0.06294601410627365, + 0.02918785810470581, + 0.018457556143403053 + ], + "credit_map": [ + 0.026897238567471504, + 0.023986803367733955, + 0.04544484615325928, + 0.04544628784060478, + 0.07902263104915619, + 0.02783282846212387, + 0.07416531443595886, + 0.03558995574712753, + 0.01721731759607792, + 0.008351263590157032, + 0.013571901246905327, + 0.007647536229342222, + 0.011715248227119446, + 0.006062934175133705, + 0.012318558990955353, + 0.009143467992544174, + 0.010332576930522919, + 0.006473834160715342, + 0.011625987477600574, + 0.006265475880354643, + 0.02984306775033474, + 0.01971021294593811, + 0.014267589896917343, + 0.008226488716900349, + 0.020256487652659416, + 0.008852888830006123, + 0.03355215862393379, + 0.009859461337327957, + 0.034007422626018524, + 0.0784413069486618, + 0.011449980549514294, + 0.019392335787415504, + 0.009972321800887585, + 0.004659151658415794, + 0.005399193149060011, + 0.006017505191266537, + 0.004593351390212774, + 0.006229736376553774, + 0.00678100623190403, + 0.005398367531597614, + 0.008205972611904144, + 0.0037925983779132366, + 0.004699258599430323, + 0.005211510229855776, + 0.005007416009902954, + 0.0034175883047282696, + 0.005726514849811792, + 0.004358548671007156, + 0.004082850180566311, + 0.0031087324023246765, + 0.0042942967265844345, + 0.004605924244970083, + 0.004291119985282421, + 0.003610537853091955, + 0.0053384192287921906, + 0.0036352660972625017, + 0.007036538794636726, + 0.00558566814288497, + 0.007136057130992413, + 0.007754179649055004, + 0.008194574154913425, + 0.007848537527024746, + 0.010144144296646118, + 0.009246314875781536, + 0.005684633739292622, + 0.003182479878887534, + 0.004522640258073807, + 0.0035155625082552433, + 0.002393579576164484, + 0.003107435302808881, + 0.002822358626872301, + 0.0039591677486896515, + 0.0035091431345790625, + 0.002216238994151354, + 0.0023269280791282654, + 0.002289039082825184, + 0.0014657177962362766, + 0.002179139293730259, + 0.0020355305168777704, + 0.0024358180817216635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.9742961525917053 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811405599117279 + ], + "opposite_logprob_tokens": [ + -3.288114070892334 + ], + "logprob_reward": -0.03811405599117279, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811405599117279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13101083040237427, + 0.05739576369524002, + 0.022472333163022995, + 0.12528833746910095, + 0.05183999240398407, + 0.14291851222515106, + 0.12306158989667892, + 0.25443777441978455, + 0.02733510173857212, + 0.06423977762460709 + ], + "credit_map": [ + 0.01544734463095665, + 0.010866736061871052, + 0.018776575103402138, + 0.014741174876689911, + 0.023407476022839546, + 0.012729437090456486, + 0.02094477228820324, + 0.014097299426794052, + 0.007367000915110111, + 0.005849742796272039, + 0.008502726443111897, + 0.00606452627107501, + 0.009890553541481495, + 0.0031111303251236677, + 0.011333555914461613, + 0.005276528187096119, + 0.002738222246989608, + 0.0017256139544770122, + 0.0028794542886316776, + 0.002131265588104725, + 0.0034124183002859354, + 0.003347069025039673, + 0.003426520386710763, + 0.0028117685578763485, + 0.011149957776069641, + 0.005132800899446011, + 0.01846826821565628, + 0.005216372199356556, + 0.020596563816070557, + 0.04603048413991928, + 0.0068700858391821384, + 0.01182380598038435, + 0.00808131042867899, + 0.004397601820528507, + 0.010532062500715256, + 0.004769620019942522, + 0.007447342853993177, + 0.0038238249253481627, + 0.006629402749240398, + 0.006158830597996712, + 0.028554854914546013, + 0.012355181388556957, + 0.01699022762477398, + 0.017139500007033348, + 0.022318461909890175, + 0.012484953738749027, + 0.01698046550154686, + 0.016094863414764404, + 0.014675935730338097, + 0.012509656138718128, + 0.012730023823678493, + 0.015291235409677029, + 0.0165893342345953, + 0.015113876201212406, + 0.021104993298649788, + 0.015046535059809685, + 0.031111901625990868, + 0.022878067567944527, + 0.03139670193195343, + 0.024766728281974792, + 0.023407869040966034, + 0.033266372978687286, + 0.046831175684928894, + 0.04077897220849991, + 0.004236728884279728, + 0.003056739689782262, + 0.0036999830044806004, + 0.0024507581256330013, + 0.0031331395730376244, + 0.0028068232350051403, + 0.0029271470848470926, + 0.005023782607167959, + 0.009624595753848553, + 0.005929676815867424, + 0.00933090876787901, + 0.0075563788414001465, + 0.008435333147644997, + 0.0065727983601391315, + 0.008310530334711075, + 0.008479563519358635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.03811405599117279 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007231004419736564 + ], + "opposite_logprob_tokens": [ + -7.250722885131836 + ], + "logprob_reward": -0.0007231004419736564, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007231004419736564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4052561819553375, + 0.058092474937438965, + 0.04826837033033371, + 0.08453136682510376, + 0.08348386734724045, + 0.13820940256118774, + 0.04849289730191231, + 0.07279942184686661, + 0.027189921587705612, + 0.0336761437356472 + ], + "credit_map": [ + 0.05481215566396713, + 0.031504590064287186, + 0.05006241053342819, + 0.045314136892557144, + 0.0770740732550621, + 0.030285296961665154, + 0.08016633242368698, + 0.03603716567158699, + 0.012172446586191654, + 0.006535244174301624, + 0.009109226986765862, + 0.006956762168556452, + 0.0053493245504796505, + 0.005636942572891712, + 0.006156797520816326, + 0.0061757308430969715, + 0.0064447177574038506, + 0.004760374315083027, + 0.0059509375132620335, + 0.005759820342063904, + 0.006059388630092144, + 0.005510661285370588, + 0.007844137959182262, + 0.005938326939940453, + 0.009340004995465279, + 0.004288203082978725, + 0.007968182675540447, + 0.005054541397839785, + 0.01793849654495716, + 0.024547800421714783, + 0.008136211894452572, + 0.007257925346493721, + 0.014177192933857441, + 0.006589079275727272, + 0.017070060595870018, + 0.008568027056753635, + 0.011681955307722092, + 0.006058515980839729, + 0.009605592116713524, + 0.009733435697853565, + 0.02492588572204113, + 0.012638573534786701, + 0.016420938074588776, + 0.020340867340564728, + 0.019377050921320915, + 0.012593141756951809, + 0.017151396721601486, + 0.014761538244783878, + 0.00836697593331337, + 0.004285448230803013, + 0.005825844593346119, + 0.006844526156783104, + 0.00418761745095253, + 0.0053219604305922985, + 0.007843945175409317, + 0.005816572345793247, + 0.008630928583443165, + 0.005809764843434095, + 0.00845242664217949, + 0.008563248440623283, + 0.006877074018120766, + 0.00982705969363451, + 0.014065285213291645, + 0.01057362835854292, + 0.005206689238548279, + 0.0022541030775755644, + 0.005854468327015638, + 0.0028648204170167446, + 0.001816601725295186, + 0.0022316796239465475, + 0.003277835203334689, + 0.0036837225779891014, + 0.00422686105594039, + 0.00348666706122458, + 0.007490281015634537, + 0.004924371372908354, + 0.0026862702798098326, + 0.003203200874850154, + 0.003363567404448986, + 0.0042949216440320015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0007231004419736564 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761778354644775 + ], + "opposite_logprob_tokens": [ + -0.8261778354644775 + ], + "logprob_reward": -0.5761778354644775, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761778354644775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1486758142709732, + 0.054531387984752655, + 0.13855743408203125, + 0.10597043484449387, + 0.06925104558467865, + 0.2545570135116577, + 0.053307194262742996, + 0.04784635454416275, + 0.0708494633436203, + 0.05645382031798363 + ], + "credit_map": [ + 0.024547461420297623, + 0.0156608447432518, + 0.01978529430925846, + 0.022598152980208397, + 0.016641277819871902, + 0.010122445411980152, + 0.023228755220770836, + 0.016091587021946907, + 0.00938027910888195, + 0.004702871199697256, + 0.008417042903602123, + 0.005242439452558756, + 0.007558838929980993, + 0.005203516222536564, + 0.008135240525007248, + 0.00589115871116519, + 0.013074507005512714, + 0.007319665513932705, + 0.01364978775382042, + 0.007213654462248087, + 0.03881906718015671, + 0.03035852313041687, + 0.017790010198950768, + 0.010332218371331692, + 0.019490743055939674, + 0.0057953521609306335, + 0.014306299388408661, + 0.007485616486519575, + 0.010873888619244099, + 0.026897195726633072, + 0.011883818544447422, + 0.009237518534064293, + 0.008438793942332268, + 0.006468815263360739, + 0.011573820374906063, + 0.007649783045053482, + 0.010982404462993145, + 0.0056450399570167065, + 0.010672749020159245, + 0.00781963113695383, + 0.04548793286085129, + 0.025482455268502235, + 0.02813779190182686, + 0.03668613359332085, + 0.03081967681646347, + 0.024201516062021255, + 0.03777894377708435, + 0.02596256509423256, + 0.007829568348824978, + 0.004150845110416412, + 0.006186098325997591, + 0.007654116023331881, + 0.005218945909291506, + 0.007834214717149734, + 0.007915455847978592, + 0.006517950911074877, + 0.008426924236118793, + 0.004646058194339275, + 0.0069413152523338795, + 0.0058523486368358135, + 0.00375558715313673, + 0.004548964090645313, + 0.0067357635125517845, + 0.006939388811588287, + 0.012354688718914986, + 0.008726191706955433, + 0.012376387603580952, + 0.007692061830312014, + 0.00481989374384284, + 0.005438300780951977, + 0.0096040153875947, + 0.009837918914854527, + 0.010527987964451313, + 0.0070571149699389935, + 0.006531147751957178, + 0.007984877564013004, + 0.004992547444999218, + 0.006175331771373749, + 0.006911463104188442, + 0.006273350212723017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.5761778354644775 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tight medium shot depicts the massive trunk of a majestic nanmu tree, its growth captured in a slow, mesmerizing time-lapse. The camera remains perfectly still, allowing the eye to absorb the intricate golden bark—a complex tapestry of deep vertical cracks and a delicate network of fissures, with a subtle, waxy sheen. As time passes, the massive trunk thickens noticeably and almost imperceptibly, its surface expanding and reorganizing in a silent testament to the passing seasons. A vast, soft-edged shadow, cast from the dense canopy above, drifts steadily and deliberately through the woods, its slow flight highlighting and obscuring rich textures in a rhythmic, hypnotic dance. The scene is bathed in the warm, dappled light of the forest, creating a dominant hue of golden yellow and earthy brown, contrasting sharply with the cool, deep tones of the shifting shadows. The overall atmosphere is one of profound calm and contemplation, a contemplative and visually rich study of silent, powerful growth.", + "seed": 7348, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_10/index_9373_seed_7348_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2609583277371712, + "vlm_logprob": -0.2609583277371712, + "vlm_yes_no_margin": 3.624999988079071, + "vlm_token_credit": -0.2609583277371712, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010349651798605919 + ], + "opposite_logprob_tokens": [ + -5.010349750518799 + ], + "logprob_reward": -0.010349651798605919, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.010349651798605919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04295782744884491, + 0.07944604009389877, + 0.0506160631775856, + 0.18173137307167053, + 0.4209751784801483, + 0.08569439500570297, + 0.020893296226859093, + 0.05639145150780678, + 0.012293695472180843, + 0.04900067672133446 + ], + "credit_map": [ + 0.0033495693933218718, + 0.008317449130117893, + 0.004082084633409977, + 0.006744324695318937, + 0.008437513373792171, + 0.004304167814552784, + 0.005092661362141371, + 0.002630053088068962, + 0.008365240879356861, + 0.012198137119412422, + 0.006775649730116129, + 0.014513352885842323, + 0.007125543430447578, + 0.016306811943650246, + 0.008336341008543968, + 0.0058249481953680515, + 0.0036985918413847685, + 0.010168935172259808, + 0.003473225049674511, + 0.006853824481368065, + 0.007177472580224276, + 0.008163917809724808, + 0.005823390558362007, + 0.005256705451756716, + 0.018086155876517296, + 0.018870964646339417, + 0.02712836302816868, + 0.029810532927513123, + 0.022495174780488014, + 0.027816101908683777, + 0.018228258937597275, + 0.019295811653137207, + 0.05895377695560455, + 0.0420854315161705, + 0.05776582658290863, + 0.0486777126789093, + 0.04205260053277016, + 0.09042143821716309, + 0.030959894880652428, + 0.05005848780274391, + 0.00915487390011549, + 0.012083036825060844, + 0.008157060481607914, + 0.007595708128064871, + 0.006997298449277878, + 0.022741304710507393, + 0.006020317319780588, + 0.012944791465997696, + 0.002393503673374653, + 0.0032879274804145098, + 0.0026150124613195658, + 0.002127248328179121, + 0.003400867572054267, + 0.002437879331409931, + 0.0020187217742204666, + 0.0026121337432414293, + 0.006981678772717714, + 0.007884802296757698, + 0.004135000519454479, + 0.005325003527104855, + 0.006577921565622091, + 0.01036657951772213, + 0.005660857539623976, + 0.009459604509174824, + 0.00145565380807966, + 0.0013361107558012009, + 0.0017707220977172256, + 0.0019173531327396631, + 0.0010440554469823837, + 0.0017043636180460453, + 0.0013826913200318813, + 0.001682743546552956, + 0.0038670580834150314, + 0.005135065410286188, + 0.0036550890654325485, + 0.012918494641780853, + 0.0050867716781795025, + 0.006519775837659836, + 0.005610299296677113, + 0.006208118051290512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.010349651798605919 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870998024940491 + ], + "opposite_logprob_tokens": [ + -1.1370997428894043 + ], + "logprob_reward": -0.3870998024940491, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.3870998024940491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2645045220851898, + 0.08919113874435425, + 0.05946886166930199, + 0.15021799504756927, + 0.07161445170640945, + 0.13194847106933594, + 0.05406320095062256, + 0.10794312506914139, + 0.01276023592799902, + 0.05828799307346344 + ], + "credit_map": [ + 0.011090891435742378, + 0.10780321806669235, + 0.00747367599979043, + 0.08257809281349182, + 0.023100268095731735, + 0.007969249039888382, + 0.01811646856367588, + 0.006372676696628332, + 0.0073775602504611015, + 0.020205652341246605, + 0.00718540558591485, + 0.01788814179599285, + 0.012258290313184261, + 0.009472488425672054, + 0.0095544857904315, + 0.005249124485999346, + 0.004311166238039732, + 0.009586557745933533, + 0.0047555118799209595, + 0.015509416349232197, + 0.006719154305756092, + 0.005338281858712435, + 0.00904291495680809, + 0.004205862991511822, + 0.011538426391780376, + 0.02888374589383602, + 0.016046840697526932, + 0.025966685265302658, + 0.018552925437688828, + 0.01525841373950243, + 0.013235013000667095, + 0.020735975354909897, + 0.006050017196685076, + 0.008366161026060581, + 0.005666723940521479, + 0.02121659368276596, + 0.0105422493070364, + 0.0077233947813510895, + 0.006292137783020735, + 0.005757179576903582, + 0.015344511717557907, + 0.025884835049510002, + 0.01589025743305683, + 0.01358499750494957, + 0.013330499641597271, + 0.022671498358249664, + 0.008483516052365303, + 0.01675836369395256, + 0.007379932329058647, + 0.007122849114239216, + 0.0075784414075315, + 0.005242513492703438, + 0.005347914528101683, + 0.007365434430539608, + 0.006634771358221769, + 0.007391354534775019, + 0.008242754265666008, + 0.02250463142991066, + 0.006123577244579792, + 0.011717562563717365, + 0.01321119349449873, + 0.015668541193008423, + 0.010083241388201714, + 0.02039162628352642, + 0.0015638609183952212, + 0.001054871012456715, + 0.002031957730650902, + 0.002104782732203603, + 0.000886423687916249, + 0.0019374004332348704, + 0.0015661247307434678, + 0.001614815671928227, + 0.006020559463649988, + 0.006417249329388142, + 0.006784460507333279, + 0.012971354648470879, + 0.0064442274160683155, + 0.007922772318124771, + 0.005739771295338869, + 0.00598760275170207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.3870998024940491 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648086339235306 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.008648086339235306, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648086339235306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08633526414632797, + 0.029178043827414513, + 0.04726483300328255, + 0.07146908342838287, + 0.0727311298251152, + 0.09029292315244675, + 0.2504768669605255, + 0.04166162386536598, + 0.036899153143167496, + 0.2736910283565521 + ], + "credit_map": [ + 0.00484498543664813, + 0.02421863004565239, + 0.004587995819747448, + 0.028464755043387413, + 0.007061445619910955, + 0.0050246999599039555, + 0.007970988750457764, + 0.004161756485700607, + 0.002160553587600589, + 0.0045371581800282, + 0.0027814009226858616, + 0.0061386339366436005, + 0.004330282565206289, + 0.002687709406018257, + 0.004246834199875593, + 0.0022954707965254784, + 0.004970171023160219, + 0.006165012717247009, + 0.004555186256766319, + 0.00939533393830061, + 0.005359615199267864, + 0.007766394410282373, + 0.005681149661540985, + 0.0033719707280397415, + 0.004366451874375343, + 0.014952048659324646, + 0.003426470560953021, + 0.008486054837703705, + 0.00870115403085947, + 0.009785458445549011, + 0.005969900172203779, + 0.015781540423631668, + 0.007277009543031454, + 0.008945398963987827, + 0.006066710222512484, + 0.01680515706539154, + 0.010916091501712799, + 0.008858287706971169, + 0.00812461320310831, + 0.005737867206335068, + 0.008743347600102425, + 0.023722490295767784, + 0.006248126272112131, + 0.010763396508991718, + 0.011464639566838741, + 0.009767277166247368, + 0.005706714000552893, + 0.013876930810511112, + 0.03179749473929405, + 0.037682000547647476, + 0.03250189125537872, + 0.026995046064257622, + 0.023105133324861526, + 0.03661699593067169, + 0.03156942129135132, + 0.030208898708224297, + 0.0044202255085110664, + 0.004294870886951685, + 0.0039410158060491085, + 0.004738449119031429, + 0.0050445739179849625, + 0.007321329787373543, + 0.004723277874290943, + 0.007177884224802256, + 0.00448204530403018, + 0.004040103871375322, + 0.005169265903532505, + 0.004601369611918926, + 0.002988226944580674, + 0.006365315057337284, + 0.00328788161277771, + 0.005964945070445538, + 0.027666211128234863, + 0.029383772984147072, + 0.02654227614402771, + 0.06384184211492538, + 0.03155697137117386, + 0.03778231516480446, + 0.02767120860517025, + 0.029246442019939423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.008648086339235306 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00034505134681239724 + ], + "opposite_logprob_tokens": [ + -8.000345230102539 + ], + "logprob_reward": -0.00034505134681239724, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.00034505134681239724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4340968132019043, + 0.14598725736141205, + 0.03987500071525574, + 0.16817329823970795, + 0.05051232874393463, + 0.05746623873710632, + 0.03708528354763985, + 0.02652793750166893, + 0.014539583586156368, + 0.025736257433891296 + ], + "credit_map": [ + 0.0169813334941864, + 0.1803896129131317, + 0.011487041600048542, + 0.13626553118228912, + 0.03989933803677559, + 0.011023370549082756, + 0.028844157233834267, + 0.00920646172016859, + 0.014806177467107773, + 0.026372632011771202, + 0.01178332231938839, + 0.02305630035698414, + 0.014733830466866493, + 0.029640045017004013, + 0.015174218453466892, + 0.010420736856758595, + 0.0030262237414717674, + 0.005548774730414152, + 0.003271915018558502, + 0.008891526609659195, + 0.004580316599458456, + 0.006230904255062342, + 0.005522516090422869, + 0.002802822506055236, + 0.01228789146989584, + 0.033887822180986404, + 0.012636941857635975, + 0.022418662905693054, + 0.021665163338184357, + 0.018759261816740036, + 0.01389593631029129, + 0.03262161463499069, + 0.005629268474876881, + 0.005778137594461441, + 0.004669398069381714, + 0.009050714783370495, + 0.006909061688929796, + 0.008006070740520954, + 0.00530096935108304, + 0.00516870804131031, + 0.006719361525028944, + 0.004921152722090483, + 0.007644583471119404, + 0.005879195872694254, + 0.004748578183352947, + 0.01521807350218296, + 0.004577436484396458, + 0.007757858373224735, + 0.004357584752142429, + 0.006506753619760275, + 0.005360580049455166, + 0.004113662056624889, + 0.003352539846673608, + 0.004932638723403215, + 0.0039768703281879425, + 0.004484649747610092, + 0.0033150517847388983, + 0.0025862669572234154, + 0.0019335235701873899, + 0.0030319548677653074, + 0.0031996897887438536, + 0.004759480711072683, + 0.003158529056236148, + 0.004543440882116556, + 0.0012702594976872206, + 0.0017756049055606127, + 0.0016885562799870968, + 0.0024215744342654943, + 0.0011866672430187464, + 0.0025078561156988144, + 0.0016478818142786622, + 0.0020411843433976173, + 0.002817165572196245, + 0.0031491995323449373, + 0.0033142296597361565, + 0.003907745238393545, + 0.0024386269506067038, + 0.0038991563487797976, + 0.0029568832833319902, + 0.0032532503828406334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.00034505134681239724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tree trunk with golden bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tree trunk visibly thicken and expand over time?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large soft-edged shadow drifting steadily across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dominant color palette composed of golden yellow and earthy brown tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983490467071533 + ], + "opposite_logprob_tokens": [ + -0.5233490467071533 + ], + "logprob_reward": -0.8983490467071533, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983490467071533, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3464229702949524, + 0.06339574605226517, + 0.045748766511678696, + 0.12020067125558853, + 0.13031424582004547, + 0.07257988303899765, + 0.07927399128675461, + 0.04113264009356499, + 0.016681792214512825, + 0.08424930274486542 + ], + "credit_map": [ + 0.01474514789879322, + 0.1366375982761383, + 0.009995760396122932, + 0.11010254174470901, + 0.02867552451789379, + 0.01224703062325716, + 0.0248099397867918, + 0.009209430776536465, + 0.005834093317389488, + 0.012020502239465714, + 0.004474870394915342, + 0.011014855466783047, + 0.007140758913010359, + 0.01239808276295662, + 0.006082296837121248, + 0.0044302851893007755, + 0.0035942879039794207, + 0.009068354964256287, + 0.003242459148168564, + 0.004630562849342823, + 0.006854095496237278, + 0.0072935461066663265, + 0.004483404569327831, + 0.006582050118595362, + 0.008178151212632656, + 0.027188582345843315, + 0.008088360540568829, + 0.017045196145772934, + 0.01617019809782505, + 0.012156431563198566, + 0.008625066839158535, + 0.022748684510588646, + 0.015747617930173874, + 0.013359281234443188, + 0.014878233894705772, + 0.02358798123896122, + 0.015829401090741158, + 0.02198920026421547, + 0.011332182213664055, + 0.013590349815785885, + 0.00986070092767477, + 0.014112299308180809, + 0.007206633687019348, + 0.007775475271046162, + 0.007324859965592623, + 0.012838910333812237, + 0.004095717798918486, + 0.009365283884108067, + 0.009287851862609386, + 0.010978701524436474, + 0.009783526882529259, + 0.007726065814495087, + 0.008413752540946007, + 0.013113928958773613, + 0.01022416353225708, + 0.00974600575864315, + 0.00425501074641943, + 0.00624057836830616, + 0.003054363653063774, + 0.00297910044901073, + 0.004666375927627087, + 0.008399242535233498, + 0.004938157740980387, + 0.006599809508770704, + 0.0018729433650150895, + 0.00115508574526757, + 0.002026355592533946, + 0.0023058054503053427, + 0.0020634860266000032, + 0.0021795008797198534, + 0.0027072082739323378, + 0.0023714066483080387, + 0.008898859843611717, + 0.009136194363236427, + 0.008933452889323235, + 0.019186239689588547, + 0.010106334462761879, + 0.011535914614796638, + 0.008051883429288864, + 0.008400417864322662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.8983490467071533 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_11/index_tensor([2781, 2781])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_11/index_tensor([2781, 2781])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8063b2398c0089c07fd6f387cfaba0a5c35b54d3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_11/index_tensor([2781, 2781])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera sweeps across a vast, futuristic financial city skyline in a smooth, 360-degree panorama, adhering strictly to a high-contrast, monochrome, dark aesthetic. Against a pitch-black, starless backdrop, countless towering skyscrapers pierce the sky with twisted, asymmetrical geometric structures. Their crystalline facades, composed of polished obsidian and cool metal alloys, reflect a sharp, icy light from a single internal light source. Beneath these majestic structures, a symbolic data torrent of countless glowing binary codes and abstract data graphics surges and recedes like a vast digital tide. This dazzling, bright white torrent carves a dynamic path through the profound darkness, its rhythmic ebb casting ever-shifting, elongated shadows across the vast architectural surfaces, as if the lifeblood of the entire city is being slowly drained away. The resulting visual spectacle is cold, solemn, and imbued with an epic quality.", + "seed": 31923, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_11/index_2781_seed_31923_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3484421018511057, + "vlm_logprob": -0.3484421018511057, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.3484421018511057, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4109630584716797 + ], + "opposite_logprob_tokens": [ + -0.2859630286693573 + ], + "logprob_reward": -1.4109630584716797, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4109630584716797, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20672737061977386, + 0.04492276906967163, + 0.0961344763636589, + 0.1599777787923813, + 0.10769752413034439, + 0.1036296933889389, + 0.07059923559427261, + 0.03802036866545677, + 0.10178671032190323, + 0.07050403207540512 + ], + "credit_map": [ + 0.02255966328084469, + 0.022706689313054085, + 0.019999369978904724, + 0.020308008417487144, + 0.05317159369587898, + 0.030911453068256378, + 0.01708626002073288, + 0.019984345883131027, + 0.00665599200874567, + 0.0045702806673944, + 0.004692393355071545, + 0.002912009134888649, + 0.010504156351089478, + 0.007679454050958157, + 0.004962846636772156, + 0.0029456361662596464, + 0.015478389337658882, + 0.013650299981236458, + 0.008045921102166176, + 0.006366507615894079, + 0.01615375094115734, + 0.015375854447484016, + 0.007646593730896711, + 0.013417162001132965, + 0.013499993830919266, + 0.02995644509792328, + 0.017721131443977356, + 0.011387145146727562, + 0.022760506719350815, + 0.026900598779320717, + 0.01292072981595993, + 0.02483123168349266, + 0.0121285580098629, + 0.019436683505773544, + 0.010498038493096828, + 0.010193334892392159, + 0.014832427725195885, + 0.02015158161520958, + 0.009359423071146011, + 0.011097473092377186, + 0.015873165801167488, + 0.01258092187345028, + 0.014772447757422924, + 0.010548313148319721, + 0.01208076998591423, + 0.011404651217162609, + 0.009794551879167557, + 0.016574859619140625, + 0.009682172909379005, + 0.010052040219306946, + 0.005778910126537085, + 0.004955058451741934, + 0.011885415762662888, + 0.006999969482421875, + 0.009107496589422226, + 0.01213817112147808, + 0.004143643658608198, + 0.004082763101905584, + 0.004231145605444908, + 0.007442744914442301, + 0.006736654322594404, + 0.0038118953816592693, + 0.0033502005971968174, + 0.004221320152282715, + 0.010383822955191135, + 0.013531913980841637, + 0.027852365747094154, + 0.010683254338800907, + 0.011732065118849277, + 0.006373407319188118, + 0.013601819053292274, + 0.007628070656210184, + 0.016453079879283905, + 0.006550218909978867, + 0.01432408019900322, + 0.0055297305807471275, + 0.004894156940281391, + 0.005592536646872759, + 0.011069625616073608, + 0.006090603768825531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -1.4109630584716797 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033726684749126434 + ], + "opposite_logprob_tokens": [ + -3.408726692199707 + ], + "logprob_reward": -0.033726684749126434, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033726684749126434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1287621557712555, + 0.039895541965961456, + 0.2545939087867737, + 0.14489580690860748, + 0.07144589722156525, + 0.056774236261844635, + 0.12020213901996613, + 0.07249155640602112, + 0.04852134734392166, + 0.06241739168763161 + ], + "credit_map": [ + 0.013497821055352688, + 0.014891783706843853, + 0.010031341575086117, + 0.01059605460613966, + 0.036159805953502655, + 0.022442257031798363, + 0.010144348256289959, + 0.010998740792274475, + 0.004599744454026222, + 0.004949572496116161, + 0.0034729912877082825, + 0.0024902895092964172, + 0.010606001131236553, + 0.006416507065296173, + 0.004333600867539644, + 0.0030268370173871517, + 0.022375866770744324, + 0.05314908176660538, + 0.014474770054221153, + 0.01105158869177103, + 0.04250139370560646, + 0.07194498926401138, + 0.016847163438796997, + 0.02224905602633953, + 0.013301498256623745, + 0.020734878256917, + 0.018010854721069336, + 0.011632249690592289, + 0.020611818879842758, + 0.021042877808213234, + 0.012919747270643711, + 0.02664187364280224, + 0.007672440726310015, + 0.012619683519005775, + 0.006372741889208555, + 0.007878029718995094, + 0.010391213931143284, + 0.012361141853034496, + 0.006968305446207523, + 0.0071823373436927795, + 0.00801246240735054, + 0.005924230441451073, + 0.008195357397198677, + 0.006260494701564312, + 0.007246788125485182, + 0.004901416599750519, + 0.005801632534712553, + 0.010431847535073757, + 0.017163999378681183, + 0.020468086004257202, + 0.011195431463420391, + 0.009882580488920212, + 0.020619267597794533, + 0.014149942435324192, + 0.011182527989149094, + 0.015540302731096745, + 0.006926535628736019, + 0.010027371346950531, + 0.006262773182243109, + 0.008431884460151196, + 0.01891999877989292, + 0.012529434636235237, + 0.004846482537686825, + 0.004547065123915672, + 0.007025671191513538, + 0.009348283521831036, + 0.006094029173254967, + 0.004663018509745598, + 0.00668485788628459, + 0.006479415111243725, + 0.004317354876548052, + 0.003908715210855007, + 0.01582830213010311, + 0.010966612957417965, + 0.0054544731974601746, + 0.003967948257923126, + 0.00588532630354166, + 0.0061832694336771965, + 0.008613530546426773, + 0.005517925601452589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.033726684749126434 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296844482421875 + ], + "opposite_logprob_tokens": [ + -2.2379684448242188 + ], + "logprob_reward": -0.11296844482421875, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296844482421875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11476633697748184, + 0.10560055077075958, + 0.1536053717136383, + 0.07186038047075272, + 0.043785128742456436, + 0.07846498489379883, + 0.05540591478347778, + 0.05787933990359306, + 0.03912881389260292, + 0.27950319647789 + ], + "credit_map": [ + 0.01436527818441391, + 0.012758547440171242, + 0.009450693614780903, + 0.0112483324483037, + 0.021501261740922928, + 0.024668177589774132, + 0.008504180237650871, + 0.012269861996173859, + 0.01675598695874214, + 0.014107671566307545, + 0.009658196941018105, + 0.00940706580877304, + 0.019692247733473778, + 0.018990136682987213, + 0.009108101017773151, + 0.007881148718297482, + 0.018523966893553734, + 0.021343519911170006, + 0.013596842996776104, + 0.011821392923593521, + 0.02021072618663311, + 0.03254330903291702, + 0.017040489241480827, + 0.018525121733546257, + 0.004658971913158894, + 0.022840550169348717, + 0.003971865866333246, + 0.004030323121696711, + 0.010178061202168465, + 0.018194768577814102, + 0.004093626979738474, + 0.0038922165986150503, + 0.00494592497125268, + 0.010051730088889599, + 0.003413684666156769, + 0.0033938686829060316, + 0.00602251710370183, + 0.007970813661813736, + 0.0032416542526334524, + 0.004744937177747488, + 0.008087142370641232, + 0.013510667718946934, + 0.0026812227442860603, + 0.00636633113026619, + 0.020380521193146706, + 0.007063745986670256, + 0.0035938136279582977, + 0.01678154058754444, + 0.006254046689718962, + 0.00692406203597784, + 0.0033003247808665037, + 0.004823898430913687, + 0.005244427360594273, + 0.005487023387104273, + 0.009518458507955074, + 0.01385367289185524, + 0.0048307678662240505, + 0.006868853233754635, + 0.008103109896183014, + 0.014175022952258587, + 0.005660160444676876, + 0.00536824157461524, + 0.00563292670994997, + 0.007240256294608116, + 0.0037222097162157297, + 0.00994698517024517, + 0.005460473243147135, + 0.003403788199648261, + 0.0041054324246943, + 0.0028364460449665785, + 0.007476866245269775, + 0.0021766144782304764, + 0.04321933537721634, + 0.023054972290992737, + 0.03389687091112137, + 0.018438391387462616, + 0.019362477585673332, + 0.02025873400270939, + 0.07496342062950134, + 0.04630899056792259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.11296844482421875 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639456048607826 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.004639456048607826, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639456048607826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1836659014225006, + 0.09410569816827774, + 0.11760436743497849, + 0.07511624693870544, + 0.11418648809194565, + 0.11678679287433624, + 0.04707634821534157, + 0.06766801327466965, + 0.12558774650096893, + 0.05820242688059807 + ], + "credit_map": [ + 0.019701600074768066, + 0.017229817807674408, + 0.015981853008270264, + 0.020765064284205437, + 0.044106997549533844, + 0.027029814198613167, + 0.01792861893773079, + 0.02092212624847889, + 0.011685722507536411, + 0.013818702660501003, + 0.00927185732871294, + 0.006486023310571909, + 0.017631087452173233, + 0.021116355434060097, + 0.007608724758028984, + 0.006487222854048014, + 0.014219827018678188, + 0.025547239929437637, + 0.007565407548099756, + 0.005245366133749485, + 0.016245925799012184, + 0.02842017449438572, + 0.006345798261463642, + 0.014014629647135735, + 0.006797561421990395, + 0.018309924751520157, + 0.007379069458693266, + 0.007716393563896418, + 0.011005863547325134, + 0.009498223662376404, + 0.006855329964309931, + 0.007553871255367994, + 0.011189432814717293, + 0.022319188341498375, + 0.010126546025276184, + 0.00972027238458395, + 0.017375867813825607, + 0.022977979853749275, + 0.010568968951702118, + 0.009908230975270271, + 0.017003623768687248, + 0.017448969185352325, + 0.009957955218851566, + 0.011386623606085777, + 0.02154664508998394, + 0.009299821220338345, + 0.008208004757761955, + 0.021935157477855682, + 0.007314667571336031, + 0.0071540712378919125, + 0.005356078501790762, + 0.004429607652127743, + 0.006181266624480486, + 0.004773937631398439, + 0.0054696835577487946, + 0.006397038698196411, + 0.005703520029783249, + 0.006456948816776276, + 0.009081888012588024, + 0.016741439700126648, + 0.007683099247515202, + 0.006913220509886742, + 0.006630445830523968, + 0.00845744926482439, + 0.012786869890987873, + 0.033654987812042236, + 0.015076973475515842, + 0.008289558812975883, + 0.01653328165411949, + 0.007500350475311279, + 0.0238127950578928, + 0.007932919077575207, + 0.008473671972751617, + 0.0052244532853364944, + 0.010559027083218098, + 0.005291721783578396, + 0.004221622832119465, + 0.004959704354405403, + 0.0096969585865736, + 0.00977526605129242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.004639456048607826 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991286516189575 + ], + "opposite_logprob_tokens": [ + -1.804912805557251 + ], + "logprob_reward": -0.17991286516189575, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991286516189575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18797218799591064, + 0.08346149325370789, + 0.08923038840293884, + 0.054911255836486816, + 0.10345378518104553, + 0.08236420899629593, + 0.05098169669508934, + 0.044244978576898575, + 0.0903773382306099, + 0.21300259232521057 + ], + "credit_map": [ + 0.022557111456990242, + 0.024318715557456017, + 0.017188971862196922, + 0.030806874856352806, + 0.03290826454758644, + 0.026025783270597458, + 0.017170023173093796, + 0.016996435821056366, + 0.011474426835775375, + 0.008467776700854301, + 0.008520746603608131, + 0.007804218214005232, + 0.013715502806007862, + 0.023640034720301628, + 0.004905736539512873, + 0.0049330489709973335, + 0.012033267877995968, + 0.010213417932391167, + 0.008953397162258625, + 0.007505626417696476, + 0.01309114508330822, + 0.015888679772615433, + 0.009993152692914009, + 0.011551695875823498, + 0.006412276532500982, + 0.007645418867468834, + 0.004993587266653776, + 0.005019759759306908, + 0.00681340554729104, + 0.008449053391814232, + 0.006512404419481754, + 0.009065347723662853, + 0.014134762808680534, + 0.020966054871678352, + 0.009885982610285282, + 0.008269033394753933, + 0.016023511067032814, + 0.016655845567584038, + 0.008031082339584827, + 0.009487509727478027, + 0.01135820522904396, + 0.011508440598845482, + 0.0038274149410426617, + 0.007195730227977037, + 0.01858626864850521, + 0.009155647829174995, + 0.004864324815571308, + 0.0158681720495224, + 0.007621574215590954, + 0.008930148556828499, + 0.006041422951966524, + 0.004178894683718681, + 0.00803709588944912, + 0.0057411352172493935, + 0.005047807469964027, + 0.005383617710322142, + 0.003964198753237724, + 0.006008945405483246, + 0.005489588249474764, + 0.007666929624974728, + 0.007926187478005886, + 0.00471498304978013, + 0.003786534769460559, + 0.0046876100823283195, + 0.009963582269847393, + 0.025059182196855545, + 0.012493272311985493, + 0.0067170048132538795, + 0.00979718379676342, + 0.006070989184081554, + 0.015603410080075264, + 0.00467271264642477, + 0.05617166683077812, + 0.02476474642753601, + 0.018098415806889534, + 0.01579293981194496, + 0.013187657110393047, + 0.01781601272523403, + 0.044676102697849274, + 0.022495046257972717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.17991286516189575 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera sweeps across a vast, futuristic financial city skyline in a smooth, 360-degree panorama, adhering strictly to a high-contrast, monochrome, dark aesthetic. Against a pitch-black, starless backdrop, countless towering skyscrapers pierce the sky with twisted, asymmetrical geometric structures. Their crystalline facades, composed of polished obsidian and cool metal alloys, reflect a sharp, icy light from a single internal light source. Beneath these majestic structures, a symbolic data torrent of countless glowing binary codes and abstract data graphics surges and recedes like a vast digital tide. This dazzling, bright white torrent carves a dynamic path through the profound darkness, its rhythmic ebb casting ever-shifting, elongated shadows across the vast architectural surfaces, as if the lifeblood of the entire city is being slowly drained away. The resulting visual spectacle is cold, solemn, and imbued with an epic quality.", + "seed": 31924, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_11/index_2781_seed_31924_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.29340128134936094, + "vlm_logprob": -0.29340128134936094, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.29340128134936094, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9849225282669067 + ], + "opposite_logprob_tokens": [ + -0.48492252826690674 + ], + "logprob_reward": -0.9849225282669067, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9849225282669067, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05857902392745018, + 0.14652591943740845, + 0.02500917762517929, + 0.03882792219519615, + 0.08540525287389755, + 0.11930230259895325, + 0.3720482885837555, + 0.059249669313430786, + 0.06614035367965698, + 0.028912048786878586 + ], + "credit_map": [ + 0.008242580108344555, + 0.007384458556771278, + 0.006015240214765072, + 0.011091538704931736, + 0.007043330930173397, + 0.006141387391835451, + 0.006586229894310236, + 0.006074254866689444, + 0.013396378606557846, + 0.0540875643491745, + 0.00470003904774785, + 0.004366221837699413, + 0.05415469408035278, + 0.006804208736866713, + 0.004522350616753101, + 0.004494459368288517, + 0.003541984362527728, + 0.0032903223764151335, + 0.0042007057927548885, + 0.0025733427610248327, + 0.0032017501071095467, + 0.0032100260723382235, + 0.0032069184817373753, + 0.001784128719009459, + 0.0067255557514727116, + 0.006390554830431938, + 0.0035059351939707994, + 0.006112333852797747, + 0.003937885630875826, + 0.005899216514080763, + 0.0035645749885588884, + 0.002691865898668766, + 0.01769755221903324, + 0.015022274106740952, + 0.007438360713422298, + 0.010984398424625397, + 0.00795790832489729, + 0.012783710844814777, + 0.007439472246915102, + 0.006081576459109783, + 0.011750700883567333, + 0.03432314842939377, + 0.030628245323896408, + 0.012703693471848965, + 0.009550673887133598, + 0.007381958421319723, + 0.007375238463282585, + 0.005588637199252844, + 0.033827390521764755, + 0.08479505032300949, + 0.018466832116246223, + 0.11553552746772766, + 0.020715923979878426, + 0.026483818888664246, + 0.055311307311058044, + 0.016912441700696945, + 0.007536700461059809, + 0.008277682587504387, + 0.007847040891647339, + 0.005945722572505474, + 0.011593349277973175, + 0.007612885441631079, + 0.00563204986974597, + 0.0048042358830571175, + 0.005144059658050537, + 0.00950587633997202, + 0.0075240358710289, + 0.011319204233586788, + 0.008834711275994778, + 0.00832311250269413, + 0.006352354772388935, + 0.009136996231973171, + 0.0031900175381451845, + 0.002508406760171056, + 0.004015768878161907, + 0.005040785763412714, + 0.0033563924953341484, + 0.0034276244696229696, + 0.003006651531904936, + 0.004366401582956314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.9849225282669067 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023307057097554207 + ], + "opposite_logprob_tokens": [ + -3.7733070850372314 + ], + "logprob_reward": -0.023307057097554207, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023307057097554207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22715209424495697, + 0.3092546761035919, + 0.09197721630334854, + 0.07772617042064667, + 0.04794977232813835, + 0.03715914487838745, + 0.04348088428378105, + 0.11507157981395721, + 0.026813434436917305, + 0.02341505140066147 + ], + "credit_map": [ + 0.028050579130649567, + 0.0382196344435215, + 0.024286996573209763, + 0.021791527047753334, + 0.02842685393989086, + 0.03604673966765404, + 0.02455425076186657, + 0.02577551268041134, + 0.024754520505666733, + 0.120732381939888, + 0.008863608352839947, + 0.007098036352545023, + 0.11632534116506577, + 0.013453432358801365, + 0.00928584672510624, + 0.008741523139178753, + 0.013393517583608627, + 0.01521697361022234, + 0.018568167462944984, + 0.007285787723958492, + 0.012834123335778713, + 0.010974492877721786, + 0.0077082570642232895, + 0.005995896644890308, + 0.013877364806830883, + 0.012366168200969696, + 0.008473954163491726, + 0.005949606187641621, + 0.009133804589509964, + 0.015076277777552605, + 0.006120888516306877, + 0.00672810897231102, + 0.008508979342877865, + 0.007695261854678392, + 0.004887492395937443, + 0.00673955213278532, + 0.005038590636104345, + 0.00597262242808938, + 0.0046621644869446754, + 0.004445116501301527, + 0.006003010552376509, + 0.004719847813248634, + 0.006791156716644764, + 0.0027788381557911634, + 0.006724874023348093, + 0.00339354295283556, + 0.004724777769297361, + 0.002023101784288883, + 0.0038681281730532646, + 0.004701059311628342, + 0.00256442092359066, + 0.013803278096020222, + 0.0023925260175019503, + 0.0033822469413280487, + 0.010885179974138737, + 0.0018840490374714136, + 0.013889865949749947, + 0.013179040513932705, + 0.013398824259638786, + 0.018208034336566925, + 0.01805170625448227, + 0.013460438698530197, + 0.012133887968957424, + 0.012749796733260155, + 0.0023229706566780806, + 0.003368349513038993, + 0.0032401718199253082, + 0.005612239707261324, + 0.0031899854075163603, + 0.0031344785820692778, + 0.0023679290898144245, + 0.0035773110575973988, + 0.002353612333536148, + 0.0028761497233062983, + 0.0026114031206816435, + 0.0037945222575217485, + 0.0024167553056031466, + 0.002625325694680214, + 0.002961580641567707, + 0.003775703487917781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.023307057097554207 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033921897411346 + ], + "opposite_logprob_tokens": [ + -2.35033917427063 + ], + "logprob_reward": -0.10033921897411346, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033921897411346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07808016985654831, + 0.47140446305274963, + 0.017944686114788055, + 0.012020918540656567, + 0.05822356790304184, + 0.2321673035621643, + 0.02056587114930153, + 0.08447974175214767, + 0.015667835250496864, + 0.009445377625524998 + ], + "credit_map": [ + 0.011910614557564259, + 0.012881653383374214, + 0.007586566265672445, + 0.011708729900419712, + 0.009026430547237396, + 0.007608811371028423, + 0.009805849753320217, + 0.0075515094213187695, + 0.03755424544215202, + 0.1876385360956192, + 0.011912208981812, + 0.00901725608855486, + 0.1840338110923767, + 0.019079195335507393, + 0.009984148666262627, + 0.012185046449303627, + 0.002724046353250742, + 0.002690523164346814, + 0.0019855983555316925, + 0.0012706804554909468, + 0.0023131542839109898, + 0.0032866341061890125, + 0.002233887556940317, + 0.0014401638181880116, + 0.0016731563955545425, + 0.0016631553880870342, + 0.0019260323606431484, + 0.001423206296749413, + 0.0011883638799190521, + 0.001932988059706986, + 0.0010963224340230227, + 0.0011176937259733677, + 0.008200488053262234, + 0.013434115797281265, + 0.006308963522315025, + 0.009495377540588379, + 0.005131185986101627, + 0.006230901461094618, + 0.005137060768902302, + 0.004285475704818964, + 0.029148448258638382, + 0.050596341490745544, + 0.03493404760956764, + 0.025601375848054886, + 0.0392606295645237, + 0.016137290745973587, + 0.0264738816767931, + 0.010015287436544895, + 0.0025209474842995405, + 0.00598506722599268, + 0.002197089372202754, + 0.0017545196460559964, + 0.0027793149929493666, + 0.0026007562410086393, + 0.001279049669392407, + 0.0014491259353235364, + 0.009324511513113976, + 0.005231605377048254, + 0.01588084176182747, + 0.01694430783390999, + 0.005170509219169617, + 0.008744003251194954, + 0.014620710164308548, + 0.008563251234591007, + 0.00190100132022053, + 0.0016088238917291164, + 0.0019009923562407494, + 0.003144172951579094, + 0.001961346482858062, + 0.0021435164380818605, + 0.0015653457958251238, + 0.0014426364796236157, + 0.0007702425355091691, + 0.0009035480907186866, + 0.001890762010589242, + 0.0015903033781796694, + 0.0010742662707343698, + 0.0011278461897745728, + 0.001458795159123838, + 0.0006296145147643983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.10033921897411346 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782706387341022 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.009782706387341022, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782706387341022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16014248132705688, + 0.056166261434555054, + 0.19247037172317505, + 0.023886656388640404, + 0.07438349723815918, + 0.1376766413450241, + 0.09830553084611893, + 0.1669028103351593, + 0.0527222715318203, + 0.037343475967645645 + ], + "credit_map": [ + 0.01969132199883461, + 0.020424434915184975, + 0.02112884446978569, + 0.024482235312461853, + 0.015138194896280766, + 0.020228195935487747, + 0.02014508657157421, + 0.018904170021414757, + 0.005758303217589855, + 0.017138376832008362, + 0.004788215272128582, + 0.004018794279545546, + 0.015251376666128635, + 0.003195458557456732, + 0.0030106406193226576, + 0.003005098784342408, + 0.026192251592874527, + 0.03298605978488922, + 0.03607112169265747, + 0.017814142629504204, + 0.022675499320030212, + 0.018898339942097664, + 0.0247552078217268, + 0.013077753596007824, + 0.005147096235305071, + 0.003786094719544053, + 0.0021400656551122665, + 0.0023143815342336893, + 0.002483152551576495, + 0.0037261692341417074, + 0.0021271647419780493, + 0.0021625321824103594, + 0.017634201794862747, + 0.014829523861408234, + 0.005057359114289284, + 0.007847730070352554, + 0.007208194583654404, + 0.009150512516498566, + 0.006181026808917522, + 0.006474950350821018, + 0.012689927592873573, + 0.035960737615823746, + 0.03657287359237671, + 0.014530841261148453, + 0.014152600429952145, + 0.008564207702875137, + 0.009010355919599533, + 0.006195101886987686, + 0.01182482298463583, + 0.025739045813679695, + 0.013219491578638554, + 0.022168224677443504, + 0.00621508713811636, + 0.00851355493068695, + 0.005757815204560757, + 0.00486748619005084, + 0.017332473769783974, + 0.026019984856247902, + 0.025928368791937828, + 0.02645188942551613, + 0.017539774999022484, + 0.018735378980636597, + 0.019496383145451546, + 0.015398564748466015, + 0.004236800596117973, + 0.006379484198987484, + 0.006645574700087309, + 0.01387774758040905, + 0.006973009090870619, + 0.006842898670583963, + 0.004049163311719894, + 0.0037175966426730156, + 0.0037934572901576757, + 0.004086266737431288, + 0.006946984212845564, + 0.006546406541019678, + 0.004350544884800911, + 0.0051672812551259995, + 0.0032267540227621794, + 0.0032257800921797752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.009782706387341022 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486548960208893 + ], + "opposite_logprob_tokens": [ + -1.223654866218567 + ], + "logprob_reward": -0.3486548960208893, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486548960208893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05629558861255646, + 0.5849116444587708, + 0.051351625472307205, + 0.019417518749833107, + 0.032373666763305664, + 0.07913094013929367, + 0.0388508141040802, + 0.09358692914247513, + 0.018702780827879906, + 0.025378501042723656 + ], + "credit_map": [ + 0.01078229583799839, + 0.0072547672316432, + 0.008719542063772678, + 0.006026189308613539, + 0.00768774701282382, + 0.0062796021811664104, + 0.004217867739498615, + 0.005327578634023666, + 0.04633401334285736, + 0.23402249813079834, + 0.013912558555603027, + 0.011269486509263515, + 0.22869251668453217, + 0.02418036013841629, + 0.011849505826830864, + 0.01465064100921154, + 0.0059396494179964066, + 0.009006397798657417, + 0.00906809139996767, + 0.004691700451076031, + 0.005962300579994917, + 0.004598572850227356, + 0.00781821645796299, + 0.004266696982085705, + 0.003012592438608408, + 0.0030516497790813446, + 0.002096337964758277, + 0.002547159558162093, + 0.0018356111831963062, + 0.0030925218015909195, + 0.0019016704754903913, + 0.0018799769459292293, + 0.004647185560315847, + 0.00496965367347002, + 0.003658028319478035, + 0.0032298702280968428, + 0.004444542340934277, + 0.005311067681759596, + 0.003859855467453599, + 0.0022534606978297234, + 0.006649091374129057, + 0.01785122975707054, + 0.023062771186232567, + 0.006232203915715218, + 0.009211950935423374, + 0.004436507821083069, + 0.007956295274198055, + 0.00373089243657887, + 0.00403331033885479, + 0.008403036743402481, + 0.0018550391541793942, + 0.006505321711301804, + 0.00931457057595253, + 0.0034922317136079073, + 0.0036012500058859587, + 0.0016460528131574392, + 0.007949477061629295, + 0.009318390861153603, + 0.018642045557498932, + 0.01949857361614704, + 0.010844914242625237, + 0.010464200749993324, + 0.009207594208419323, + 0.007661739364266396, + 0.0019045009976252913, + 0.002754488028585911, + 0.003355371067300439, + 0.0024164938367903233, + 0.002155437134206295, + 0.002492450177669525, + 0.0019336150726303458, + 0.0016904238145798445, + 0.00309198466129601, + 0.003080028109252453, + 0.003118589986115694, + 0.0031537539325654507, + 0.004205075092613697, + 0.002950359834358096, + 0.002429964952170849, + 0.0033487447071820498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 5 + ], + "selected_reward": -0.3486548960208893 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera sweeps across a vast, futuristic financial city skyline in a smooth, 360-degree panorama, adhering strictly to a high-contrast, monochrome, dark aesthetic. Against a pitch-black, starless backdrop, countless towering skyscrapers pierce the sky with twisted, asymmetrical geometric structures. Their crystalline facades, composed of polished obsidian and cool metal alloys, reflect a sharp, icy light from a single internal light source. Beneath these majestic structures, a symbolic data torrent of countless glowing binary codes and abstract data graphics surges and recedes like a vast digital tide. This dazzling, bright white torrent carves a dynamic path through the profound darkness, its rhythmic ebb casting ever-shifting, elongated shadows across the vast architectural surfaces, as if the lifeblood of the entire city is being slowly drained away. The resulting visual spectacle is cold, solemn, and imbued with an epic quality.", + "seed": 31925, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_11/index_2781_seed_31925_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.23893690975382925, + "vlm_logprob": -0.23893690975382925, + "vlm_yes_no_margin": 2.349999952316284, + "vlm_token_credit": -0.23893690975382925, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6377865076065063 + ], + "opposite_logprob_tokens": [ + -0.7627865076065063 + ], + "logprob_reward": -0.6377865076065063, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6377865076065063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18697071075439453, + 0.02722129411995411, + 0.15846991539001465, + 0.04848826304078102, + 0.02863866277039051, + 0.02874159626662731, + 0.4410565197467804, + 0.044807251542806625, + 0.015048709698021412, + 0.020557120442390442 + ], + "credit_map": [ + 0.01362982764840126, + 0.012879207730293274, + 0.053199052810668945, + 0.01665618270635605, + 0.05168047174811363, + 0.01132568996399641, + 0.012750977650284767, + 0.01484929583966732, + 0.002012082375586033, + 0.0028003051411360502, + 0.0034058585297316313, + 0.0048855505883693695, + 0.0026631176006048918, + 0.0038552386686205864, + 0.003705529263243079, + 0.0038936117198318243, + 0.05508729815483093, + 0.028542203828692436, + 0.005936611909419298, + 0.006546498276293278, + 0.00916852243244648, + 0.010114070028066635, + 0.00812392495572567, + 0.034950774163007736, + 0.005885701160877943, + 0.006931901443749666, + 0.006269353907555342, + 0.004758257884532213, + 0.005682969931513071, + 0.005397016182541847, + 0.007498749997466803, + 0.006064311601221561, + 0.003456622129306197, + 0.003186911577358842, + 0.004164924845099449, + 0.003346274374052882, + 0.0035790663678199053, + 0.0034675500355660915, + 0.004987581633031368, + 0.0024497308768332005, + 0.0036400791723281145, + 0.0027043004520237446, + 0.0031710988841950893, + 0.0027403312269598246, + 0.003124393755570054, + 0.0036151933018118143, + 0.006021867971867323, + 0.0037243312690407038, + 0.035956885665655136, + 0.16578270494937897, + 0.022492142394185066, + 0.01998080313205719, + 0.03354993090033531, + 0.05738257244229317, + 0.010239594615995884, + 0.09567185491323471, + 0.0054819416254758835, + 0.006350971292704344, + 0.007249553222209215, + 0.0029667839407920837, + 0.007928491570055485, + 0.004888804163783789, + 0.0036753490567207336, + 0.006265357602387667, + 0.002701306249946356, + 0.0017416776390746236, + 0.0011618384160101414, + 0.0012915381230413914, + 0.0019653234630823135, + 0.0024885765742510557, + 0.002353191375732422, + 0.0013452573912218213, + 0.0023468555882573128, + 0.0031039221212267876, + 0.0016041708877310157, + 0.0022684410214424133, + 0.0024899954441934824, + 0.0029121139086782932, + 0.002909604227170348, + 0.0029220152646303177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.6377865076065063 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819959655404091 + ], + "opposite_logprob_tokens": [ + -4.018199443817139 + ], + "logprob_reward": -0.01819959655404091, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819959655404091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07231192290782928, + 0.03535384684801102, + 0.13999654352664948, + 0.3497961461544037, + 0.06760929524898529, + 0.01923956349492073, + 0.1423768252134323, + 0.12246730178594589, + 0.01987048238515854, + 0.03097803331911564 + ], + "credit_map": [ + 0.005802535451948643, + 0.009056421928107738, + 0.009014738723635674, + 0.010087099857628345, + 0.012374702841043472, + 0.0076688905246555805, + 0.00803377851843834, + 0.010273758322000504, + 0.0034232228063046932, + 0.005684971809387207, + 0.003816476557403803, + 0.006863189395517111, + 0.003865073202177882, + 0.0051863593980669975, + 0.0032215584069490433, + 0.003292997367680073, + 0.04726226255297661, + 0.03920578956604004, + 0.004072188399732113, + 0.005695073865354061, + 0.00756937125697732, + 0.009785209782421589, + 0.0058754971250891685, + 0.02053116075694561, + 0.029891379177570343, + 0.04918156564235687, + 0.04193359240889549, + 0.0287028718739748, + 0.0394584983587265, + 0.06303386390209198, + 0.03607437387108803, + 0.061519984155893326, + 0.009336396120488644, + 0.0068856473080813885, + 0.015677345916628838, + 0.00660358602181077, + 0.009055917151272297, + 0.006143578793853521, + 0.009380368515849113, + 0.004526454024016857, + 0.0020325512159615755, + 0.002263195114210248, + 0.00206541595980525, + 0.0019045774824917316, + 0.003506968729197979, + 0.0024694146122783422, + 0.002582838758826256, + 0.0024146034847944975, + 0.011947989463806152, + 0.053015220910310745, + 0.007630531210452318, + 0.00677069928497076, + 0.010554923675954342, + 0.018563613295555115, + 0.0038615046069025993, + 0.03003234602510929, + 0.013362604193389416, + 0.016327865421772003, + 0.012377473525702953, + 0.010943666100502014, + 0.020136674866080284, + 0.022947421297430992, + 0.01582697220146656, + 0.01054463256150484, + 0.004119986668229103, + 0.0027104667387902737, + 0.001460836618207395, + 0.001837205607444048, + 0.0026536814402788877, + 0.003035050816833973, + 0.0023889276199042797, + 0.0016643271083012223, + 0.0030102392192929983, + 0.0056751929223537445, + 0.0016948669217526913, + 0.0027283011004328728, + 0.005378393456339836, + 0.005485787522047758, + 0.00409551290795207, + 0.0029097378719598055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.01819959655404091 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813384532928467 + ], + "opposite_logprob_tokens": [ + -1.4063384532928467 + ], + "logprob_reward": -0.2813384532928467, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813384532928467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15348464250564575, + 0.04776296392083168, + 0.12103675305843353, + 0.24754992127418518, + 0.11573261022567749, + 0.06632249057292938, + 0.0694945678114891, + 0.06041964516043663, + 0.04079494997859001, + 0.07740147411823273 + ], + "credit_map": [ + 0.012053967453539371, + 0.012882990762591362, + 0.037708830088377, + 0.012234579771757126, + 0.043236032128334045, + 0.0108603835105896, + 0.01123708114027977, + 0.013270777650177479, + 0.0036423406563699245, + 0.004667534027248621, + 0.005092574283480644, + 0.0074483975768089294, + 0.003923787269741297, + 0.007795825134962797, + 0.006726896855980158, + 0.008465605787932873, + 0.043017446994781494, + 0.03728875517845154, + 0.003957511857151985, + 0.004432884510606527, + 0.006715430412441492, + 0.00839932356029749, + 0.004862222820520401, + 0.012363170273602009, + 0.019895514473319054, + 0.03432627022266388, + 0.03250119835138321, + 0.022716151550412178, + 0.027740545570850372, + 0.0404779352247715, + 0.0259325560182333, + 0.043959762901067734, + 0.012032490223646164, + 0.013361971825361252, + 0.014173369854688644, + 0.015538083389401436, + 0.01824452169239521, + 0.014415787532925606, + 0.01782410405576229, + 0.010142283514142036, + 0.009012446738779545, + 0.009239890612661839, + 0.008390570990741253, + 0.006414860021322966, + 0.009712696075439453, + 0.00867937970906496, + 0.006158141419291496, + 0.008714510127902031, + 0.0053313919343054295, + 0.022857319563627243, + 0.0048763686791062355, + 0.004048652481287718, + 0.006093467585742474, + 0.008119175210595131, + 0.003839687677100301, + 0.014328506775200367, + 0.009487025439739227, + 0.009255745448172092, + 0.009079597890377045, + 0.004412750713527203, + 0.00901220366358757, + 0.006104097701609135, + 0.0045027947053313255, + 0.008565426804125309, + 0.004418179392814636, + 0.011073012836277485, + 0.001973876729607582, + 0.0026677269488573074, + 0.007319409400224686, + 0.0033693592995405197, + 0.004443710204213858, + 0.005529672373086214, + 0.008538572117686272, + 0.0077852183021605015, + 0.0053787836804986, + 0.008110374212265015, + 0.00887139793485403, + 0.009309070184826851, + 0.015865949913859367, + 0.013542107306420803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.2813384532928467 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252490285784006 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252490285784006, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252490285784006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16424241662025452, + 0.03842097148299217, + 0.1535559743642807, + 0.2573409080505371, + 0.047870934009552, + 0.026862857863307, + 0.08820116519927979, + 0.05476950854063034, + 0.06933505833148956, + 0.09940014034509659 + ], + "credit_map": [ + 0.010954455472528934, + 0.010389471426606178, + 0.0494367890059948, + 0.011254029348492622, + 0.05181567743420601, + 0.010847079567611217, + 0.010157201439142227, + 0.009387720376253128, + 0.0023562705609947443, + 0.004025954753160477, + 0.004412337671965361, + 0.008615851402282715, + 0.0031136777251958847, + 0.00449290731921792, + 0.005535692907869816, + 0.005868279375135899, + 0.053332213312387466, + 0.03631772845983505, + 0.005114514380693436, + 0.005645982455462217, + 0.0088425911962986, + 0.009995649568736553, + 0.0079250643029809, + 0.02638223022222519, + 0.01922314241528511, + 0.03185661882162094, + 0.034052617847919464, + 0.023675870150327682, + 0.02842555195093155, + 0.04355146363377571, + 0.030034739524126053, + 0.0465209074318409, + 0.005293277557939291, + 0.005976600106805563, + 0.006912895943969488, + 0.005734320264309645, + 0.006274535320699215, + 0.006255611777305603, + 0.007267605047672987, + 0.00415608799085021, + 0.004215005785226822, + 0.0032018276397138834, + 0.0028227916918694973, + 0.0019423702033236623, + 0.005023550242185593, + 0.0037745388690382242, + 0.0031886708457022905, + 0.0026941043324768543, + 0.008085004985332489, + 0.030677974224090576, + 0.005410103593021631, + 0.004946511704474688, + 0.007093439809978008, + 0.011321906931698322, + 0.006422465667128563, + 0.01424375455826521, + 0.005648849532008171, + 0.009009410627186298, + 0.0074172052554786205, + 0.0022802117746323347, + 0.012450525537133217, + 0.008152705617249012, + 0.0029620311688631773, + 0.006848568096756935, + 0.007687885779887438, + 0.007463307119905949, + 0.004919272381812334, + 0.009031165391206741, + 0.010543297976255417, + 0.009011421352624893, + 0.010212108492851257, + 0.01046659704297781, + 0.012955408543348312, + 0.009068354032933712, + 0.005126316100358963, + 0.007030862383544445, + 0.012147377245128155, + 0.011978955008089542, + 0.02061833254992962, + 0.020474540069699287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.005252490285784006 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25210750102996826 + ], + "opposite_logprob_tokens": [ + -1.5021075010299683 + ], + "logprob_reward": -0.25210750102996826, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25210750102996826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30216071009635925, + 0.039674509316682816, + 0.05160300433635712, + 0.15213623642921448, + 0.054777681827545166, + 0.07903849333524704, + 0.20201946794986725, + 0.047593116760253906, + 0.03692926466464996, + 0.0340675450861454 + ], + "credit_map": [ + 0.021286243572831154, + 0.02086379937827587, + 0.08942277729511261, + 0.01972062513232231, + 0.09821242839097977, + 0.01881987601518631, + 0.016382068395614624, + 0.017452877014875412, + 0.0028509292751550674, + 0.00532557861879468, + 0.005022608209401369, + 0.005547298118472099, + 0.003570904955267906, + 0.007598489988595247, + 0.0045692771673202515, + 0.005189420189708471, + 0.010303630493581295, + 0.012567759491503239, + 0.005083817522972822, + 0.005282812286168337, + 0.004098283126950264, + 0.004903858061879873, + 0.005089562851935625, + 0.004273270722478628, + 0.012790527194738388, + 0.022384952753782272, + 0.018389830365777016, + 0.01255465392023325, + 0.01757647469639778, + 0.025472722947597504, + 0.016032669693231583, + 0.026934392750263214, + 0.009711641818284988, + 0.005097271408885717, + 0.01097374688833952, + 0.0060036443173885345, + 0.006249610800296068, + 0.006009670440107584, + 0.005564030259847641, + 0.0051680575124919415, + 0.01013782899826765, + 0.008940321393311024, + 0.009973805397748947, + 0.008103543892502785, + 0.0116443345323205, + 0.00824744999408722, + 0.012478108517825603, + 0.00951309036463499, + 0.015852635726332664, + 0.06909039616584778, + 0.010432818904519081, + 0.011712529696524143, + 0.01618378609418869, + 0.023593241348862648, + 0.01516830176115036, + 0.03998573124408722, + 0.007051191292703152, + 0.005466714967042208, + 0.005898226518183947, + 0.004343066830188036, + 0.00545461755245924, + 0.00777485640719533, + 0.00690371124073863, + 0.0047007217071950436, + 0.005597259383648634, + 0.005111272446811199, + 0.0023587250616401434, + 0.004183126147836447, + 0.0055693211033940315, + 0.00466918246820569, + 0.00541767617687583, + 0.004022697452455759, + 0.0033506948966532946, + 0.004694960545748472, + 0.0016456994926556945, + 0.002850747900083661, + 0.004462088458240032, + 0.004238103050738573, + 0.008239807561039925, + 0.004585438407957554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.25210750102996826 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera sweeps across a vast, futuristic financial city skyline in a smooth, 360-degree panorama, adhering strictly to a high-contrast, monochrome, dark aesthetic. Against a pitch-black, starless backdrop, countless towering skyscrapers pierce the sky with twisted, asymmetrical geometric structures. Their crystalline facades, composed of polished obsidian and cool metal alloys, reflect a sharp, icy light from a single internal light source. Beneath these majestic structures, a symbolic data torrent of countless glowing binary codes and abstract data graphics surges and recedes like a vast digital tide. This dazzling, bright white torrent carves a dynamic path through the profound darkness, its rhythmic ebb casting ever-shifting, elongated shadows across the vast architectural surfaces, as if the lifeblood of the entire city is being slowly drained away. The resulting visual spectacle is cold, solemn, and imbued with an epic quality.", + "seed": 31926, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_11/index_2781_seed_31926_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3691290348768234, + "vlm_logprob": -0.3691290348768234, + "vlm_yes_no_margin": 1.975, + "vlm_token_credit": -0.3691290348768234, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4154711961746216 + ], + "opposite_logprob_tokens": [ + -0.2904711961746216 + ], + "logprob_reward": -1.4154711961746216, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4154711961746216, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07860149443149567, + 0.04260800778865814, + 0.09675457328557968, + 0.161521315574646, + 0.14464768767356873, + 0.026974992826581, + 0.0560566708445549, + 0.06828124076128006, + 0.24897341430187225, + 0.07558056712150574 + ], + "credit_map": [ + 0.011558876372873783, + 0.013662471435964108, + 0.0068346248008310795, + 0.007614390458911657, + 0.012521196156740189, + 0.009397072717547417, + 0.0057428446598351, + 0.011270026676356792, + 0.005472251679748297, + 0.00898163951933384, + 0.003769121365621686, + 0.0038348441012203693, + 0.006688393652439117, + 0.00599622679874301, + 0.002405848354101181, + 0.005459681153297424, + 0.016454268246889114, + 0.015232374891638756, + 0.00857824832201004, + 0.0065440163016319275, + 0.01268440019339323, + 0.011985122226178646, + 0.008221902884542942, + 0.017054248601198196, + 0.025095000863075256, + 0.021079232916235924, + 0.0234347153455019, + 0.018365338444709778, + 0.011735550127923489, + 0.02021166868507862, + 0.027181336656212807, + 0.014418501406908035, + 0.027551403269171715, + 0.011012480594217777, + 0.02152259089052677, + 0.015593957155942917, + 0.010438394732773304, + 0.012645442970097065, + 0.016501259058713913, + 0.029382169246673584, + 0.005453805904835463, + 0.0025327419862151146, + 0.0036371981259435415, + 0.002569034229964018, + 0.0042589157819747925, + 0.0031821110751479864, + 0.0023327977396547794, + 0.0030083900783210993, + 0.010176295414566994, + 0.007010905537754297, + 0.0068207294680178165, + 0.004062562715262175, + 0.012769399210810661, + 0.006083571817725897, + 0.004688164684921503, + 0.0044450475834310055, + 0.00989304855465889, + 0.0063447230495512486, + 0.008126725442707539, + 0.006164342164993286, + 0.010010034777224064, + 0.00974539015442133, + 0.007499203085899353, + 0.010497787036001682, + 0.04057282954454422, + 0.026147183030843735, + 0.0675586387515068, + 0.0301156435161829, + 0.01828262209892273, + 0.0257906224578619, + 0.016548603773117065, + 0.023957306519150734, + 0.010858849622309208, + 0.0077660768292844296, + 0.009959922172129154, + 0.009495003148913383, + 0.01129100937396288, + 0.017800912261009216, + 0.004468823783099651, + 0.0039399792440235615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 4 + ], + "selected_reward": -1.4154711961746216 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906803280115128 + ], + "opposite_logprob_tokens": [ + -2.4640679359436035 + ], + "logprob_reward": -0.08906803280115128, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906803280115128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1047717034816742, + 0.046282198280096054, + 0.10137093812227249, + 0.2971845269203186, + 0.20450431108474731, + 0.01776498556137085, + 0.03773469477891922, + 0.028332315385341644, + 0.07131180912256241, + 0.09074243158102036 + ], + "credit_map": [ + 0.01498602144420147, + 0.01772543415427208, + 0.007664202246814966, + 0.008165596053004265, + 0.017861729487776756, + 0.012278110720217228, + 0.008743316866457462, + 0.017347291111946106, + 0.005537883378565311, + 0.011805332265794277, + 0.004487111233174801, + 0.0026068876031786203, + 0.006783363875001669, + 0.0065858936868608, + 0.0024191064294427633, + 0.006056623067706823, + 0.021741308271884918, + 0.015490736812353134, + 0.008889220654964447, + 0.008418551646173, + 0.015535098500549793, + 0.013395778834819794, + 0.0063379304483532906, + 0.011562318541109562, + 0.04922660440206528, + 0.04861893504858017, + 0.03155415132641792, + 0.028963690623641014, + 0.023420697078108788, + 0.03985409066081047, + 0.04927534610033035, + 0.026271002367138863, + 0.029867416247725487, + 0.02919146977365017, + 0.018226411193609238, + 0.024032680317759514, + 0.04256048798561096, + 0.029126066714525223, + 0.015616904944181442, + 0.015882886946201324, + 0.0036607927177101374, + 0.0014049084857106209, + 0.002359777921810746, + 0.001242576981894672, + 0.0022071758285164833, + 0.002277820371091366, + 0.001986035145819187, + 0.002625897992402315, + 0.006978542543947697, + 0.004799848422408104, + 0.004154003690928221, + 0.0024995063431560993, + 0.008724484592676163, + 0.005291538313031197, + 0.0031348448246717453, + 0.002151928376406431, + 0.004515551030635834, + 0.0027919202111661434, + 0.0034278479870408773, + 0.0026702184695750475, + 0.004196000285446644, + 0.003804005915299058, + 0.0030054189264774323, + 0.003921353258192539, + 0.011235861107707024, + 0.007078425493091345, + 0.017740242183208466, + 0.007196687627583742, + 0.004501542076468468, + 0.0072767529636621475, + 0.0065198736265301704, + 0.009762423112988472, + 0.016535500064492226, + 0.005153130739927292, + 0.010053114034235477, + 0.005852834787219763, + 0.01655563898384571, + 0.028671110048890114, + 0.004710438661277294, + 0.003210661467164755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.08906803280115128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868157580494881 + ], + "opposite_logprob_tokens": [ + -3.0486814975738525 + ], + "logprob_reward": -0.04868157580494881, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868157580494881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2076474130153656, + 0.07211683690547943, + 0.15152256190776825, + 0.14803260564804077, + 0.11892655491828918, + 0.03385903313755989, + 0.042398709803819656, + 0.06259164214134216, + 0.08859357237815857, + 0.07431109249591827 + ], + "credit_map": [ + 0.029387466609477997, + 0.03136212006211281, + 0.014602404087781906, + 0.01705498807132244, + 0.03194684162735939, + 0.03344285115599632, + 0.01938074640929699, + 0.030469980090856552, + 0.008645621128380299, + 0.01323492731899023, + 0.007282033562660217, + 0.005867402534931898, + 0.010730646550655365, + 0.010732547380030155, + 0.00519851129502058, + 0.0104251429438591, + 0.022542597725987434, + 0.032255180180072784, + 0.01525097619742155, + 0.01152169518172741, + 0.018548255786299706, + 0.013774785213172436, + 0.00990720558911562, + 0.027721857652068138, + 0.024571554735302925, + 0.028868913650512695, + 0.014696970582008362, + 0.013545232824981213, + 0.012817960232496262, + 0.018690235912799835, + 0.024624736979603767, + 0.010216997005045414, + 0.018154369667172432, + 0.012935045175254345, + 0.016401851549744606, + 0.009843002073466778, + 0.018371401354670525, + 0.015993406996130943, + 0.011169797740876675, + 0.016057677567005157, + 0.005008283536881208, + 0.004904777742922306, + 0.0035721075255423784, + 0.0022963949013501406, + 0.007608999032527208, + 0.00493715750053525, + 0.003264941740781069, + 0.002266369527205825, + 0.006755875889211893, + 0.004526101518422365, + 0.008337085135281086, + 0.004071420524269342, + 0.007827778346836567, + 0.004545629024505615, + 0.0038866782560944557, + 0.0024481373839080334, + 0.008695384487509727, + 0.005336145404726267, + 0.008841597475111485, + 0.00599490525200963, + 0.009173020720481873, + 0.011143609881401062, + 0.006036975421011448, + 0.007370001170784235, + 0.01786397397518158, + 0.006823357194662094, + 0.025788795202970505, + 0.010662171058356762, + 0.006702458020299673, + 0.009303947910666466, + 0.0053836749866604805, + 0.006065188907086849, + 0.013789852149784565, + 0.006317913066595793, + 0.005687830038368702, + 0.0063408976420760155, + 0.012929735705256462, + 0.02352554351091385, + 0.00307524879463017, + 0.0026440632063895464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.04868157580494881 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080164462327957 + ], + "opposite_logprob_tokens": [ + -4.511080265045166 + ], + "logprob_reward": -0.011080164462327957, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080164462327957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10993100702762604, + 0.0392848365008831, + 0.08773630112409592, + 0.11156551539897919, + 0.11047855764627457, + 0.02577778697013855, + 0.040961939841508865, + 0.11033028364181519, + 0.11993342638015747, + 0.24400028586387634 + ], + "credit_map": [ + 0.031923454254865646, + 0.015099221840500832, + 0.007565068081021309, + 0.008877442218363285, + 0.010882953181862831, + 0.011679797433316708, + 0.007970225997269154, + 0.015932856127619743, + 0.006580025888979435, + 0.006959602236747742, + 0.0045576575212180614, + 0.004412910435348749, + 0.004588827956467867, + 0.004410403780639172, + 0.002693201880902052, + 0.005082216113805771, + 0.008657845668494701, + 0.01780039072036743, + 0.01087357010692358, + 0.008716019801795483, + 0.009649387560784817, + 0.009042412973940372, + 0.00899104680866003, + 0.014005640521645546, + 0.017252549529075623, + 0.01327471248805523, + 0.016298934817314148, + 0.0145430201664567, + 0.00791423674672842, + 0.012544343248009682, + 0.020424330607056618, + 0.009313399903476238, + 0.01771906577050686, + 0.013198567554354668, + 0.017282120883464813, + 0.010961627587676048, + 0.011895163916051388, + 0.01760573871433735, + 0.011231268756091595, + 0.010585024021565914, + 0.004227549768984318, + 0.0034629625733941793, + 0.00425606919452548, + 0.00195226538926363, + 0.0035423394292593002, + 0.0032290127128362656, + 0.0030844074208289385, + 0.002023183275014162, + 0.007477720733731985, + 0.005490015726536512, + 0.004960203543305397, + 0.002820231718942523, + 0.009255368262529373, + 0.004451470915228128, + 0.0034814823884516954, + 0.0030254500452429056, + 0.01456134021282196, + 0.011707999743521214, + 0.01742779277265072, + 0.01052266824990511, + 0.015213798731565475, + 0.022681161761283875, + 0.00952970515936613, + 0.008685825392603874, + 0.022657999768853188, + 0.012078479863703251, + 0.038339681923389435, + 0.012413037940859795, + 0.007172247860580683, + 0.012471803463995457, + 0.006231340114027262, + 0.008568849414587021, + 0.041350699961185455, + 0.017016848549246788, + 0.032362669706344604, + 0.021867331117391586, + 0.03960130736231804, + 0.06899522244930267, + 0.012150474824011326, + 0.010655749589204788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.011080164462327957 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera movement a smooth 360-degree panorama sweeping across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the skyscrapers depicted as towering structures with twisted, asymmetrical geometric shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual aesthetic high-contrast, monochrome, and dark with a pitch-black, starless backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a glowing torrent of binary codes and abstract data graphics surging beneath the skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crystalline facades of the buildings reflect a sharp, icy light from a single internal source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813442051410675 + ], + "opposite_logprob_tokens": [ + -1.4063441753387451 + ], + "logprob_reward": -0.2813442051410675, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813442051410675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12635116279125214, + 0.045109279453754425, + 0.09890811890363693, + 0.17161880433559418, + 0.23294387757778168, + 0.026119614019989967, + 0.05259579047560692, + 0.07094474881887436, + 0.0637892559170723, + 0.11161942034959793 + ], + "credit_map": [ + 0.015354246832430363, + 0.019691959023475647, + 0.013856112957000732, + 0.01369093544781208, + 0.021349240094423294, + 0.014627828262746334, + 0.009909212589263916, + 0.017871610820293427, + 0.00676051527261734, + 0.007894830778241158, + 0.0056061032228171825, + 0.0035833243746310472, + 0.006061115302145481, + 0.006049312185496092, + 0.003205021610483527, + 0.005949051585048437, + 0.020309632644057274, + 0.014518462121486664, + 0.00810871459543705, + 0.005797075107693672, + 0.016515467315912247, + 0.013796054758131504, + 0.007638490758836269, + 0.012224203906953335, + 0.031559694558382034, + 0.03394712135195732, + 0.014805424027144909, + 0.018238943070173264, + 0.016970176249742508, + 0.019648093730211258, + 0.02572707273066044, + 0.010722264647483826, + 0.03360191360116005, + 0.028417624533176422, + 0.019767191261053085, + 0.024671273306012154, + 0.04079148545861244, + 0.029204709455370903, + 0.020580658689141273, + 0.035909008234739304, + 0.004212833940982819, + 0.002676929347217083, + 0.0044431621208786964, + 0.0032164272852241993, + 0.0027422932907938957, + 0.002823017770424485, + 0.003225098829716444, + 0.0027798498049378395, + 0.007656432222574949, + 0.006011241115629673, + 0.009759098291397095, + 0.005886624101549387, + 0.008878948166966438, + 0.005607220344245434, + 0.004692233167588711, + 0.004103986546397209, + 0.010160421021282673, + 0.007673752028495073, + 0.006751008331775665, + 0.008675565011799335, + 0.007846822030842304, + 0.010873842053115368, + 0.008533431217074394, + 0.010429895482957363, + 0.010907018557190895, + 0.006504288408905268, + 0.013958996161818504, + 0.007739287801086903, + 0.005161798559129238, + 0.009700870141386986, + 0.005219535902142525, + 0.004597457591444254, + 0.01341371051967144, + 0.013799253851175308, + 0.012508966028690338, + 0.017109889537096024, + 0.018167728558182716, + 0.022091571241617203, + 0.008783262223005295, + 0.00574503606185317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.2813442051410675 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_12/index_tensor([7386, 7386])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_12/index_tensor([7386, 7386])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cb9c48c10c1586f7e50b79eaca8604414cecb0c3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_12/index_tensor([7386, 7386])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a slightly overhead angle, the camera slowly pans to the right, following the movement. On the smooth lacquer surface of a dark, finely grained antique wooden table, scattered haphazardly are a string of pearls of varying sizes, the remains of a broken necklace, emitting a warm, milky white sheen. A warm yellow table lamp, shimmering obliquely from the side, casts a bright spot of light on the curved surface of each pearl. A slender, fair female hand, nails painted crimson, enters the frame with a decisive and fluid gesture. Without hesitation, it sweeps all the pearls towards the edge of the table, from left to right. Instantly, the pearls cascade down like a swarm of startled creatures, crashing onto the polished dark hardwood floor below with a crisp, dense sound. They then roll in all directions, eventually disappearing into the shadows of the frame. The entire scene, dimly lit, creates a dramatic atmosphere of elegance, decisiveness, and a touch of destructive beauty.", + "seed": 8836, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_12/index_7386_seed_8836_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.041769695654511454, + "vlm_logprob": -0.041769695654511454, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.041769695654511454, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10248130559921265 + ], + "opposite_logprob_tokens": [ + -3.1024813652038574 + ], + "logprob_reward": -0.10248130559921265, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.10248130559921265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055211734026670456, + 0.02470184862613678, + 0.014532262459397316, + 0.041705891489982605, + 0.03985174745321274, + 0.1619664877653122, + 0.17534120380878448, + 0.3815224766731262, + 0.07130466401576996, + 0.0338616669178009 + ], + "credit_map": [ + 0.011987085454165936, + 0.006654498632997274, + 0.008152436465024948, + 0.008239317685365677, + 0.0044211032800376415, + 0.003976372070610523, + 0.007077349815517664, + 0.004703565966337919, + 0.0027607844676822424, + 0.004947304725646973, + 0.003954071085900068, + 0.004342728294432163, + 0.002108747372403741, + 0.001805590814910829, + 0.002915437798947096, + 0.0018671838333830237, + 0.0015781980473548174, + 0.001562452525831759, + 0.001891419873572886, + 0.0020279958844184875, + 0.0014871630119159818, + 0.0029552695341408253, + 0.001409718650393188, + 0.0016200444661080837, + 0.005394271574914455, + 0.005727970506995916, + 0.005141329951584339, + 0.004697727505117655, + 0.004102733451873064, + 0.007007562555372715, + 0.004414830356836319, + 0.005219464655965567, + 0.004174353554844856, + 0.0060477773658931255, + 0.006039467640221119, + 0.006205038167536259, + 0.003610632847994566, + 0.0060132392682135105, + 0.00395723432302475, + 0.0038040010258555412, + 0.024683266878128052, + 0.014509727247059345, + 0.017350729554891586, + 0.015742985531687737, + 0.01339048333466053, + 0.029442323371767998, + 0.02309008687734604, + 0.023756882175803185, + 0.010820958763360977, + 0.018827402964234352, + 0.018603460863232613, + 0.00781521201133728, + 0.035937294363975525, + 0.0458265021443367, + 0.018163854256272316, + 0.01934652402997017, + 0.0044178348034620285, + 0.007694595027714968, + 0.012395456433296204, + 0.0034574575256556273, + 0.16003702580928802, + 0.16304276883602142, + 0.013833158649504185, + 0.016644177958369255, + 0.007749040611088276, + 0.005080095026642084, + 0.006973327603191137, + 0.00493391091004014, + 0.01164530124515295, + 0.014304579235613346, + 0.013908956199884415, + 0.0067094494588673115, + 0.0031937845051288605, + 0.0023121885024011135, + 0.003356942208483815, + 0.002056690864264965, + 0.004867169074714184, + 0.004871570039540529, + 0.009683771058917046, + 0.0035195511300116777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.10248130559921265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014232203364372253 + ], + "opposite_logprob_tokens": [ + -4.264232158660889 + ], + "logprob_reward": -0.014232203364372253, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014232203364372253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12577234208583832, + 0.030181456357240677, + 0.037048809230327606, + 0.08243367820978165, + 0.15817515552043915, + 0.10852488875389099, + 0.11854156106710434, + 0.20131857693195343, + 0.10721561312675476, + 0.030787885189056396 + ], + "credit_map": [ + 0.014027267694473267, + 0.014484258368611336, + 0.01907982863485813, + 0.021057119593024254, + 0.016366366297006607, + 0.012611699290573597, + 0.019911739975214005, + 0.008234059438109398, + 0.0026383353397250175, + 0.0038936391938477755, + 0.0042115203104913235, + 0.005511609837412834, + 0.0034025053028017282, + 0.0028942679055035114, + 0.005048407707363367, + 0.0025811700616031885, + 0.006061254534870386, + 0.005030165892094374, + 0.00579383410513401, + 0.005157520994544029, + 0.00400278577581048, + 0.0048093535006046295, + 0.003233346389606595, + 0.002960548037663102, + 0.011964571662247181, + 0.01811933144927025, + 0.011087937280535698, + 0.012063601985573769, + 0.010422158055007458, + 0.006433887407183647, + 0.0064036305993795395, + 0.005938566755503416, + 0.023719901219010353, + 0.03423694893717766, + 0.026783479377627373, + 0.02006782963871956, + 0.010904856026172638, + 0.017679931595921516, + 0.008340161293745041, + 0.01644204929471016, + 0.010250409133732319, + 0.009546528570353985, + 0.013696481473743916, + 0.01305572409182787, + 0.01168380118906498, + 0.018841423094272614, + 0.016898902133107185, + 0.014551625587046146, + 0.008050047792494297, + 0.018109843134880066, + 0.007957886904478073, + 0.005540365353226662, + 0.025222815573215485, + 0.02878100983798504, + 0.013183403760194778, + 0.011696193367242813, + 0.004431550856679678, + 0.0045802355743944645, + 0.010413520969450474, + 0.002448504092171788, + 0.0844101682305336, + 0.07659422606229782, + 0.010137028992176056, + 0.00830333773046732, + 0.010654066689312458, + 0.0071197617799043655, + 0.01057974062860012, + 0.006511378567665815, + 0.01654665730893612, + 0.022857585921883583, + 0.02148863859474659, + 0.011457783170044422, + 0.0018501210724934936, + 0.0030485792085528374, + 0.002012247685343027, + 0.0015428663464263082, + 0.005449449643492699, + 0.006097964942455292, + 0.007879186421632767, + 0.0029074677731841803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.014232203364372253 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865182563662529 + ], + "opposite_logprob_tokens": [ + -3.048651933670044 + ], + "logprob_reward": -0.04865182563662529, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865182563662529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15211233496665955, + 0.024945201352238655, + 0.03396229073405266, + 0.056954991072416306, + 0.052220914512872696, + 0.07919285446405411, + 0.10740362107753754, + 0.4450671672821045, + 0.03416737914085388, + 0.013973192311823368 + ], + "credit_map": [ + 0.02058512344956398, + 0.020148906856775284, + 0.020845189690589905, + 0.027248764410614967, + 0.011486129835247993, + 0.015034792944788933, + 0.019429491832852364, + 0.01733393780887127, + 0.002205762779340148, + 0.0026887175627052784, + 0.003093868028372526, + 0.003406140021979809, + 0.003297394374385476, + 0.0028286189772188663, + 0.004815093707293272, + 0.002609607297927141, + 0.004064091481268406, + 0.0043472228571772575, + 0.004759977571666241, + 0.0034039027523249388, + 0.003643646603450179, + 0.00692442711442709, + 0.003378857159987092, + 0.0034401663579046726, + 0.008012760430574417, + 0.008672931231558323, + 0.007136503700166941, + 0.008221099153161049, + 0.006779786664992571, + 0.005667134653776884, + 0.005945319309830666, + 0.006519456394016743, + 0.008079820312559605, + 0.011639607138931751, + 0.00945497676730156, + 0.0053193653002381325, + 0.0035522333346307278, + 0.005490737967193127, + 0.0032247432973235846, + 0.005459432490170002, + 0.009555835276842117, + 0.006808590609580278, + 0.009138242341578007, + 0.008050958625972271, + 0.007292268797755241, + 0.015638301149010658, + 0.011131434701383114, + 0.011577222496271133, + 0.004715437535196543, + 0.008425804786384106, + 0.006275673862546682, + 0.0039818985387682915, + 0.026608482003211975, + 0.035308241844177246, + 0.008178435266017914, + 0.01390964537858963, + 0.004779694601893425, + 0.008505034260451794, + 0.005928476341068745, + 0.003908936399966478, + 0.19801321625709534, + 0.18915435671806335, + 0.018423255532979965, + 0.016354212537407875, + 0.0036935999523848295, + 0.002477044705301523, + 0.0031271926127374172, + 0.002459597773849964, + 0.005696834996342659, + 0.005928753409534693, + 0.007532516028732061, + 0.0032518403604626656, + 0.0007909807027317584, + 0.0007313957321457565, + 0.0009900066070258617, + 0.00061290233861655, + 0.001983500551432371, + 0.0027945428155362606, + 0.004370664246380329, + 0.0016991986194625497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.04865182563662529 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784122928977013 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784122928977013, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784122928977013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03462165221571922, + 0.01581474579870701, + 0.03230131044983864, + 0.019982080906629562, + 0.06977756321430206, + 0.12819711863994598, + 0.19396761059761047, + 0.20622824132442474, + 0.26817721128463745, + 0.030932441353797913 + ], + "credit_map": [ + 0.004528082441538572, + 0.0035555115900933743, + 0.00596191082149744, + 0.007359824143350124, + 0.003737398888915777, + 0.0025118577759712934, + 0.004498220980167389, + 0.002468843711540103, + 0.0015854953089728951, + 0.0025329296477138996, + 0.0020146288443356752, + 0.0024848105385899544, + 0.001999617787078023, + 0.0021759136579930782, + 0.0016476373421028256, + 0.0013737115077674389, + 0.004534668289124966, + 0.0038724937476217747, + 0.004333749879151583, + 0.004822940099984407, + 0.0037319306284189224, + 0.004717748612165451, + 0.003050080966204405, + 0.0032377021852880716, + 0.002661305945366621, + 0.003022577613592148, + 0.0030063313897699118, + 0.0023558929096907377, + 0.002297126455232501, + 0.0024318096693605185, + 0.0021408754400908947, + 0.002066161250695586, + 0.008898821659386158, + 0.014649533666670322, + 0.008344233967363834, + 0.008504273369908333, + 0.005945119541138411, + 0.008480648510158062, + 0.00764394411817193, + 0.007310989312827587, + 0.01600814051926136, + 0.00821717269718647, + 0.02137097902595997, + 0.01628420688211918, + 0.006705449894070625, + 0.025526931509375572, + 0.015505571849644184, + 0.0185786671936512, + 0.012460407800972462, + 0.02136375568807125, + 0.016148095950484276, + 0.007997073233127594, + 0.03702211007475853, + 0.05056002736091614, + 0.02074868604540825, + 0.02766745537519455, + 0.004566372372210026, + 0.00703806895762682, + 0.005780097562819719, + 0.002397021744400263, + 0.08838535845279694, + 0.07591884583234787, + 0.012835818342864513, + 0.009306668303906918, + 0.02672515995800495, + 0.01799243502318859, + 0.025691833347082138, + 0.017646364867687225, + 0.04337227717041969, + 0.044124048203229904, + 0.06691879034042358, + 0.02570629119873047, + 0.0016899238107725978, + 0.001472552539780736, + 0.0019335015676915646, + 0.0017964272992685437, + 0.005959090776741505, + 0.00562398973852396, + 0.009384914301335812, + 0.0030720399226993322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.009784122928977013 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033699020743370056 + ], + "opposite_logprob_tokens": [ + -3.4086990356445312 + ], + "logprob_reward": -0.033699020743370056, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033699020743370056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058444783091545105, + 0.037961699068546295, + 0.047846149653196335, + 0.02211662568151951, + 0.05797826498746872, + 0.08860868215560913, + 0.2973237931728363, + 0.16630376875400543, + 0.19093284010887146, + 0.0324833020567894 + ], + "credit_map": [ + 0.007288987748324871, + 0.007038881070911884, + 0.008321781642735004, + 0.008379338309168816, + 0.008317598141729832, + 0.00490216352045536, + 0.010065054520964622, + 0.00413098419085145, + 0.004066541325300932, + 0.0068666948936879635, + 0.005729977507144213, + 0.007519623264670372, + 0.004001963883638382, + 0.0034344468731433153, + 0.003617603797465563, + 0.0027248540427535772, + 0.005657503847032785, + 0.0056314365938305855, + 0.005823968444019556, + 0.005867484025657177, + 0.005290733650326729, + 0.00803006999194622, + 0.006135907024145126, + 0.005409054458141327, + 0.002768842736259103, + 0.0020934802014380693, + 0.003167936112731695, + 0.0023822530638426542, + 0.0024009966291487217, + 0.0038026615511626005, + 0.002620114479213953, + 0.002880343934521079, + 0.005786426831036806, + 0.004772826097905636, + 0.009315717965364456, + 0.010688375681638718, + 0.005254659801721573, + 0.009256655350327492, + 0.006258014589548111, + 0.006645599380135536, + 0.01065080240368843, + 0.009346839040517807, + 0.012659817934036255, + 0.009923207573592663, + 0.008772679604589939, + 0.013507518917322159, + 0.01217922754585743, + 0.011568601243197918, + 0.018131164833903313, + 0.03550385683774948, + 0.027671393007040024, + 0.012794671580195427, + 0.05451110750436783, + 0.0787673145532608, + 0.025806572288274765, + 0.04413777217268944, + 0.009541995823383331, + 0.009833313524723053, + 0.02844041958451271, + 0.004958042874932289, + 0.046472206711769104, + 0.04480747506022453, + 0.011829795315861702, + 0.010420547798275948, + 0.017958389595150948, + 0.012446803040802479, + 0.018558036535978317, + 0.010505604557693005, + 0.030098114162683487, + 0.04573330655694008, + 0.03176039084792137, + 0.023872215300798416, + 0.0019568572752177715, + 0.001987102208659053, + 0.0021586983930319548, + 0.0014818633208051324, + 0.00675869919359684, + 0.006841890048235655, + 0.007266133092343807, + 0.004032065160572529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.033699020743370056 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, the camera slowly pans to the right, following the movement. On the smooth lacquer surface of a dark, finely grained antique wooden table, scattered haphazardly are a string of pearls of varying sizes, the remains of a broken necklace, emitting a warm, milky white sheen. A warm yellow table lamp, shimmering obliquely from the side, casts a bright spot of light on the curved surface of each pearl. A slender, fair female hand, nails painted crimson, enters the frame with a decisive and fluid gesture. Without hesitation, it sweeps all the pearls towards the edge of the table, from left to right. Instantly, the pearls cascade down like a swarm of startled creatures, crashing onto the polished dark hardwood floor below with a crisp, dense sound. They then roll in all directions, eventually disappearing into the shadows of the frame. The entire scene, dimly lit, creates a dramatic atmosphere of elegance, decisiveness, and a touch of destructive beauty.", + "seed": 8837, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_12/index_7386_seed_8837_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04101863838732243, + "vlm_logprob": -0.04101863838732243, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.04101863838732243, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07624103873968124 + ], + "opposite_logprob_tokens": [ + -3.2012410163879395 + ], + "logprob_reward": -0.07624103873968124, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.07624103873968124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07659528404474258, + 0.060210518538951874, + 0.08279944956302643, + 0.04220017045736313, + 0.11555839329957962, + 0.14244310557842255, + 0.12246113270521164, + 0.22117982804775238, + 0.05982818454504013, + 0.07672395557165146 + ], + "credit_map": [ + 0.008411220274865627, + 0.00865989550948143, + 0.009958505630493164, + 0.008498668670654297, + 0.008415905758738518, + 0.011985228396952152, + 0.007065344601869583, + 0.013600518926978111, + 0.007068573031574488, + 0.006242475006729364, + 0.00806968379765749, + 0.01031890045851469, + 0.006094386335462332, + 0.005522708874195814, + 0.008195601403713226, + 0.00869818776845932, + 0.009908405132591724, + 0.008564619347453117, + 0.010793876834213734, + 0.0072435238398611546, + 0.010651357471942902, + 0.01082794088870287, + 0.010819880291819572, + 0.013989850878715515, + 0.00825201254338026, + 0.00440398370847106, + 0.007915280759334564, + 0.0034044890198856592, + 0.003916940186172724, + 0.005576616618782282, + 0.0034741791896522045, + 0.005256671458482742, + 0.010621418245136738, + 0.007145578507333994, + 0.015583167783915997, + 0.027087511494755745, + 0.010296816006302834, + 0.013196087442338467, + 0.015062374994158745, + 0.016565440222620964, + 0.01103703211992979, + 0.005986934062093496, + 0.030797170475125313, + 0.01758718676865101, + 0.015147811733186245, + 0.027383936569094658, + 0.01953106001019478, + 0.014971968717873096, + 0.008561408147215843, + 0.00818432867527008, + 0.020975762978196144, + 0.017271099612116814, + 0.011228607036173344, + 0.018917160108685493, + 0.019147098064422607, + 0.018175672739744186, + 0.036519862711429596, + 0.03179911524057388, + 0.03584815561771393, + 0.017711229622364044, + 0.029286088421940804, + 0.023348890244960785, + 0.02945020981132984, + 0.017216281965374947, + 0.005460980348289013, + 0.007357520051300526, + 0.008543867617845535, + 0.0061850352212786674, + 0.011189418844878674, + 0.009317263029515743, + 0.007231967989355326, + 0.004542127717286348, + 0.008496485650539398, + 0.005275264382362366, + 0.008311429992318153, + 0.0054201362654566765, + 0.011742223985493183, + 0.02508559823036194, + 0.005861356388777494, + 0.006531460210680962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.07624103873968124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009811745956540108 + ], + "opposite_logprob_tokens": [ + -4.634811878204346 + ], + "logprob_reward": -0.009811745956540108, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009811745956540108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10127221792936325, + 0.19312350451946259, + 0.24344900250434875, + 0.025373881682753563, + 0.03687762841582298, + 0.15743085741996765, + 0.05904416739940643, + 0.05044420436024666, + 0.02905498817563057, + 0.10392946749925613 + ], + "credit_map": [ + 0.008344857953488827, + 0.010014912113547325, + 0.016645513474941254, + 0.012208630330860615, + 0.012528283521533012, + 0.009978325106203556, + 0.014711976051330566, + 0.016839735209941864, + 0.015607251785695553, + 0.013726242817938328, + 0.02817443385720253, + 0.031140975654125214, + 0.02415943518280983, + 0.021232066676020622, + 0.025884905830025673, + 0.033198218792676926, + 0.030854593962430954, + 0.02720584347844124, + 0.030707482248544693, + 0.02099091000854969, + 0.03257139027118683, + 0.03336919844150543, + 0.027903689071536064, + 0.03984591364860535, + 0.0028330280911177397, + 0.002855859464034438, + 0.003297345479950309, + 0.002189628081396222, + 0.005507672671228647, + 0.0036427080631256104, + 0.002216431312263012, + 0.002831211080774665, + 0.0030398049857467413, + 0.0032063580583781004, + 0.006582559552043676, + 0.005070478189736605, + 0.0035239350982010365, + 0.0041437880136072636, + 0.004896438214927912, + 0.0064142681658267975, + 0.008308894000947475, + 0.006325224880129099, + 0.03847939148545265, + 0.021259251981973648, + 0.011837400496006012, + 0.026927195489406586, + 0.023856589570641518, + 0.020436933264136314, + 0.006644028704613447, + 0.0029976542573422194, + 0.016253596171736717, + 0.008154099807143211, + 0.006688996218144894, + 0.006060059182345867, + 0.00615589739754796, + 0.006089838221669197, + 0.009322977624833584, + 0.006225546356290579, + 0.006141242571175098, + 0.006612830795347691, + 0.005018296185880899, + 0.0049153901636600494, + 0.007517456542700529, + 0.00469046737998724, + 0.002227307530120015, + 0.0046540009789168835, + 0.004022382665425539, + 0.002938496181741357, + 0.0060395305044949055, + 0.004501048941165209, + 0.002752083819359541, + 0.001920140697620809, + 0.01189827173948288, + 0.006142381113022566, + 0.009797425009310246, + 0.004264896269887686, + 0.017881687730550766, + 0.03750786930322647, + 0.005376431625336409, + 0.011060521937906742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.009811745956540108 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370317071676254 + ], + "opposite_logprob_tokens": [ + -3.408703088760376 + ], + "logprob_reward": -0.03370317071676254, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370317071676254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34550225734710693, + 0.14850392937660217, + 0.06975150853395462, + 0.04652568697929382, + 0.03533777594566345, + 0.10380978882312775, + 0.05163925886154175, + 0.03510100767016411, + 0.04386427626013756, + 0.11996456980705261 + ], + "credit_map": [ + 0.034406956285238266, + 0.03145824745297432, + 0.05776055157184601, + 0.03349161148071289, + 0.033016592264175415, + 0.05028977990150452, + 0.04385439306497574, + 0.061224132776260376, + 0.0076595330610871315, + 0.006001044064760208, + 0.04226211830973625, + 0.013234054669737816, + 0.008813922293484211, + 0.010108985006809235, + 0.01113210804760456, + 0.04929216206073761, + 0.007343906443566084, + 0.009913971647620201, + 0.007238942198455334, + 0.004575680010020733, + 0.015747446566820145, + 0.01119477953761816, + 0.006111640948802233, + 0.007625137455761433, + 0.005976146552711725, + 0.004189562983810902, + 0.009235689416527748, + 0.0037247075233608484, + 0.005035252310335636, + 0.010781521908938885, + 0.0036373657640069723, + 0.003945440053939819, + 0.0038362275809049606, + 0.002102719387039542, + 0.004394832532852888, + 0.008585724979639053, + 0.0032093997579067945, + 0.0035368180833756924, + 0.005025272723287344, + 0.00464677857235074, + 0.004026331473141909, + 0.0024444549344480038, + 0.04023845121264458, + 0.015291518531739712, + 0.003061646828427911, + 0.005686252377927303, + 0.019507108256220818, + 0.013554027304053307, + 0.004831761587411165, + 0.0039566620253026485, + 0.012766378931701183, + 0.00730404257774353, + 0.005393660627305508, + 0.006135822739452124, + 0.004980890545994043, + 0.006270041689276695, + 0.003709268756210804, + 0.004858018830418587, + 0.004165388643741608, + 0.00562730198726058, + 0.003407652722671628, + 0.0031618706416338682, + 0.006713053211569786, + 0.003458455204963684, + 0.003163886023685336, + 0.006698316894471645, + 0.007344692479819059, + 0.003745128633454442, + 0.007807963527739048, + 0.007199978455901146, + 0.0043916357681155205, + 0.0035126737784594297, + 0.013405206613242626, + 0.006951177027076483, + 0.011202746070921421, + 0.004977765027433634, + 0.02131076343357563, + 0.044929202646017075, + 0.005643618293106556, + 0.01154409535229206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.03370317071676254 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023293431848287582 + ], + "opposite_logprob_tokens": [ + -3.7732934951782227 + ], + "logprob_reward": -0.023293431848287582, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023293431848287582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12428124248981476, + 0.09223940968513489, + 0.07418103516101837, + 0.05742941424250603, + 0.05936221405863762, + 0.06184028834104538, + 0.12463212013244629, + 0.22034966945648193, + 0.14674699306488037, + 0.03893759846687317 + ], + "credit_map": [ + 0.01103871688246727, + 0.01381785236299038, + 0.02240770310163498, + 0.014555004425346851, + 0.010705086402595043, + 0.01286461390554905, + 0.01714295521378517, + 0.021749306470155716, + 0.005202094092965126, + 0.004647915717214346, + 0.023094091564416885, + 0.010578107088804245, + 0.0068352362141013145, + 0.006792784668505192, + 0.008694205433130264, + 0.026394972577691078, + 0.007847967557609081, + 0.007525437045842409, + 0.00844288244843483, + 0.0069613088853657246, + 0.010933725163340569, + 0.01139494962990284, + 0.009323461912572384, + 0.011751295998692513, + 0.005301757715642452, + 0.0044783130288124084, + 0.014396916143596172, + 0.0048944419249892235, + 0.007117636036127806, + 0.006872488185763359, + 0.00527858966961503, + 0.009089267812669277, + 0.00498181814327836, + 0.004012012388557196, + 0.009806963615119457, + 0.009442384354770184, + 0.008180377073585987, + 0.008602624759078026, + 0.007448017131537199, + 0.006888015195727348, + 0.005094326566904783, + 0.0028225078713148832, + 0.016611434519290924, + 0.0080257598310709, + 0.006043065804988146, + 0.007176942192018032, + 0.00816691666841507, + 0.007899336516857147, + 0.013539518229663372, + 0.005636293441057205, + 0.0333385095000267, + 0.017503177747130394, + 0.015560976229608059, + 0.013326184824109077, + 0.014870413579046726, + 0.010857042856514454, + 0.03042956441640854, + 0.029739979654550552, + 0.023977816104888916, + 0.03705945238471031, + 0.023899372667074203, + 0.021754490211606026, + 0.036137957125902176, + 0.017351048067212105, + 0.010695177130401134, + 0.02262336015701294, + 0.01780340075492859, + 0.013435800559818745, + 0.02742217294871807, + 0.021189013496041298, + 0.01936953328549862, + 0.0142085375264287, + 0.0037572775036096573, + 0.0030791140161454678, + 0.003834170289337635, + 0.005218638107180595, + 0.006138196215033531, + 0.008850420825183392, + 0.0033049462363123894, + 0.004754834342747927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.023293431848287582 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204380467534065 + ], + "opposite_logprob_tokens": [ + -2.8120439052581787 + ], + "logprob_reward": -0.06204380467534065, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204380467534065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10333409160375595, + 0.10078265517950058, + 0.09222631901502609, + 0.0724649503827095, + 0.10691596567630768, + 0.24441833794116974, + 0.06700465828180313, + 0.07096875458955765, + 0.07315737754106522, + 0.06872689723968506 + ], + "credit_map": [ + 0.013401146978139877, + 0.00787382572889328, + 0.016206463798880577, + 0.008054312318563461, + 0.009168421849608421, + 0.015574260614812374, + 0.013731252402067184, + 0.019324397668242455, + 0.0055327205918729305, + 0.004423464182764292, + 0.027723655104637146, + 0.009529522620141506, + 0.006548814941197634, + 0.007583427242934704, + 0.007855319418013096, + 0.031585730612277985, + 0.008648517541587353, + 0.012650827877223492, + 0.0110095189884305, + 0.0065878392197191715, + 0.019667716696858406, + 0.013958871364593506, + 0.008819152601063251, + 0.01088387705385685, + 0.00723957410082221, + 0.004643396008759737, + 0.018211478367447853, + 0.005740346387028694, + 0.009978817775845528, + 0.01369424443691969, + 0.005339148920029402, + 0.007617942988872528, + 0.00818308349698782, + 0.007678642403334379, + 0.013047099113464355, + 0.023329969495534897, + 0.012599628418684006, + 0.01573152467608452, + 0.012486154213547707, + 0.013859869912266731, + 0.010097119957208633, + 0.0058809127658605576, + 0.09475120902061462, + 0.034364938735961914, + 0.007177706342190504, + 0.013707887381315231, + 0.04322412610054016, + 0.03521445766091347, + 0.004883567336946726, + 0.0053068893030285835, + 0.013342412188649178, + 0.0075240652076900005, + 0.007081456016749144, + 0.008559695445001125, + 0.010956620797514915, + 0.00934994500130415, + 0.007819993421435356, + 0.008699167519807816, + 0.012676792219281197, + 0.01120257843285799, + 0.005652078427374363, + 0.007094519212841988, + 0.009368163533508778, + 0.008455470204353333, + 0.004679636098444462, + 0.014281676150858402, + 0.010676713660359383, + 0.006274610757827759, + 0.01416705921292305, + 0.011412656866014004, + 0.007186725735664368, + 0.004478298593312502, + 0.006972279399633408, + 0.0037597285117954016, + 0.006472706329077482, + 0.004238591995090246, + 0.011987102217972279, + 0.02541051246225834, + 0.0035302306059747934, + 0.006355745252221823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.06204380467534065 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, the camera slowly pans to the right, following the movement. On the smooth lacquer surface of a dark, finely grained antique wooden table, scattered haphazardly are a string of pearls of varying sizes, the remains of a broken necklace, emitting a warm, milky white sheen. A warm yellow table lamp, shimmering obliquely from the side, casts a bright spot of light on the curved surface of each pearl. A slender, fair female hand, nails painted crimson, enters the frame with a decisive and fluid gesture. Without hesitation, it sweeps all the pearls towards the edge of the table, from left to right. Instantly, the pearls cascade down like a swarm of startled creatures, crashing onto the polished dark hardwood floor below with a crisp, dense sound. They then roll in all directions, eventually disappearing into the shadows of the frame. The entire scene, dimly lit, creates a dramatic atmosphere of elegance, decisiveness, and a touch of destructive beauty.", + "seed": 8838, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_12/index_7386_seed_8838_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.049456368759274484, + "vlm_logprob": -0.049456368759274484, + "vlm_yes_no_margin": 3.299999952316284, + "vlm_token_credit": -0.049456368759274484, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1030779778957367 + ], + "opposite_logprob_tokens": [ + -2.7280778884887695 + ], + "logprob_reward": -0.1030779778957367, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.1030779778957367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2120531052350998, + 0.1218457743525505, + 0.08367945998907089, + 0.03681951388716698, + 0.13939370214939117, + 0.07574347406625748, + 0.052177175879478455, + 0.13299565017223358, + 0.09238756448030472, + 0.052904631942510605 + ], + "credit_map": [ + 0.017135504633188248, + 0.01532441470772028, + 0.01771591790020466, + 0.026550322771072388, + 0.014398942701518536, + 0.012982543557882309, + 0.0573744960129261, + 0.050570953637361526, + 0.004684830084443092, + 0.004476354457437992, + 0.005221166182309389, + 0.024886658415198326, + 0.007631084881722927, + 0.004533505532890558, + 0.046734798699617386, + 0.02367737516760826, + 0.004692246206104755, + 0.006623482797294855, + 0.008095001801848412, + 0.010670573450624943, + 0.009318264201283455, + 0.005752378609031439, + 0.019599558785557747, + 0.018927957862615585, + 0.0033242946956306696, + 0.005270165856927633, + 0.0050618089735507965, + 0.002818190259858966, + 0.005107292905449867, + 0.004748227074742317, + 0.005441391374915838, + 0.005048139952123165, + 0.00877773854881525, + 0.01262807846069336, + 0.017318785190582275, + 0.01466391608119011, + 0.01451923605054617, + 0.010273031890392303, + 0.03267817571759224, + 0.028534745797514915, + 0.008609716780483723, + 0.009798368439078331, + 0.009444552473723888, + 0.007654462940990925, + 0.008352112956345081, + 0.011894754134118557, + 0.0074308584444224834, + 0.012558653019368649, + 0.005354202818125486, + 0.008396835066378117, + 0.011556369252502918, + 0.006312411278486252, + 0.004631762858480215, + 0.0061814188957214355, + 0.004179599229246378, + 0.005564582999795675, + 0.01184568926692009, + 0.00962120108306408, + 0.021618677303195, + 0.012601829133927822, + 0.00993434526026249, + 0.012513500638306141, + 0.028056690469384193, + 0.02680371142923832, + 0.008714436553418636, + 0.010547088459134102, + 0.011799748986959457, + 0.009673716500401497, + 0.012871651910245419, + 0.012134358286857605, + 0.017591767013072968, + 0.009054800495505333, + 0.0036711785942316055, + 0.00588219054043293, + 0.005799062084406614, + 0.007868960499763489, + 0.009769568219780922, + 0.008671878837049007, + 0.007997370325028896, + 0.003244423307478428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.1030779778957367 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01822207123041153 + ], + "opposite_logprob_tokens": [ + -4.018221855163574 + ], + "logprob_reward": -0.01822207123041153, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01822207123041153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4273938238620758, + 0.15467946231365204, + 0.04157975688576698, + 0.04872588440775871, + 0.039495762437582016, + 0.020179612562060356, + 0.03396196290850639, + 0.045641086995601654, + 0.08877066522836685, + 0.09957204014062881 + ], + "credit_map": [ + 0.019458770751953125, + 0.017245588824152946, + 0.027633992955088615, + 0.04571039229631424, + 0.02316536195576191, + 0.01922892965376377, + 0.14409837126731873, + 0.1308523565530777, + 0.0070115793496370316, + 0.008868497796356678, + 0.010892349295318127, + 0.030428791418671608, + 0.009141824208199978, + 0.008557583205401897, + 0.05400749295949936, + 0.02577131986618042, + 0.002736611058935523, + 0.004454261157661676, + 0.004482031334191561, + 0.006498394533991814, + 0.0061705950647592545, + 0.004816296044737101, + 0.006066089496016502, + 0.00635547237470746, + 0.004660361912101507, + 0.004741697572171688, + 0.0060495734214782715, + 0.003995226696133614, + 0.006828714162111282, + 0.006070159375667572, + 0.010025810450315475, + 0.0063543315045535564, + 0.0024653763975948095, + 0.0037541266065090895, + 0.004093331750482321, + 0.004214759450405836, + 0.0041536386124789715, + 0.0030240374617278576, + 0.009047831408679485, + 0.008742657490074635, + 0.0016385397175326943, + 0.0018265799153596163, + 0.003395591862499714, + 0.002596519887447357, + 0.001579355914145708, + 0.002408925909548998, + 0.002738539595156908, + 0.003995557315647602, + 0.004462664481252432, + 0.004620121326297522, + 0.00737680122256279, + 0.0034703414421528578, + 0.003296045819297433, + 0.004007136914879084, + 0.0024012841749936342, + 0.00432756170630455, + 0.0026673886459320784, + 0.0018214046722278, + 0.004429863765835762, + 0.00554680498316884, + 0.003220628947019577, + 0.0038316131103783846, + 0.011714383028447628, + 0.012408992275595665, + 0.007551384158432484, + 0.011798996478319168, + 0.008833804167807102, + 0.009853729978203773, + 0.012376613914966583, + 0.014902463182806969, + 0.014421767555177212, + 0.009031896479427814, + 0.004876377992331982, + 0.011357780545949936, + 0.012959721498191357, + 0.013113865628838539, + 0.019971324130892754, + 0.01648593135178089, + 0.015303699299693108, + 0.005503326654434204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.01822207123041153 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495531111955643 + ], + "opposite_logprob_tokens": [ + -2.929955244064331 + ], + "logprob_reward": -0.05495531111955643, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495531111955643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10529527068138123, + 0.1486508697271347, + 0.04812053218483925, + 0.12919901311397552, + 0.049011316150426865, + 0.04733775183558464, + 0.05093364417552948, + 0.0603940412402153, + 0.16536785662174225, + 0.19568972289562225 + ], + "credit_map": [ + 0.010713692754507065, + 0.012618115171790123, + 0.014748405665159225, + 0.013311531394720078, + 0.016504405066370964, + 0.011853374540805817, + 0.012746179476380348, + 0.012799572199583054, + 0.0068293362855911255, + 0.005099788308143616, + 0.007492725271731615, + 0.02937437780201435, + 0.007479486055672169, + 0.005148768424987793, + 0.05745488405227661, + 0.02977151796221733, + 0.004054262302815914, + 0.004691755399107933, + 0.004845744930207729, + 0.005822562146931887, + 0.0037809263449162245, + 0.0035783189814537764, + 0.01161020714789629, + 0.009736757725477219, + 0.015355833806097507, + 0.012273003347218037, + 0.015290416777133942, + 0.013344028033316135, + 0.020461248233914375, + 0.018996814265847206, + 0.017670338973402977, + 0.01580733060836792, + 0.00537651963531971, + 0.005945886950939894, + 0.00544226448982954, + 0.004131406545639038, + 0.006952859926968813, + 0.003787751542404294, + 0.010744601488113403, + 0.006630032788962126, + 0.004606856033205986, + 0.004408878739923239, + 0.009095351211726665, + 0.00683167576789856, + 0.003952445462346077, + 0.005194453056901693, + 0.005413143429905176, + 0.00783495046198368, + 0.005678513087332249, + 0.008847317658364773, + 0.008536521345376968, + 0.005978281144052744, + 0.005495766177773476, + 0.005350883584469557, + 0.003802504623308778, + 0.007243860978633165, + 0.007173740305006504, + 0.005894921720027924, + 0.0073203714564442635, + 0.010230602696537971, + 0.004502967465668917, + 0.007950699888169765, + 0.009080823510885239, + 0.008239919319748878, + 0.010492297820746899, + 0.021037433296442032, + 0.01265819650143385, + 0.01586681604385376, + 0.020813999697566032, + 0.027822768315672874, + 0.0320991687476635, + 0.024577176198363304, + 0.012827465310692787, + 0.01847650110721588, + 0.019920870661735535, + 0.0282284002751112, + 0.038253381848335266, + 0.03438786044716835, + 0.02915259823203087, + 0.014442650601267815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.05495531111955643 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073880717158318 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073880717158318, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073880717158318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20069490373134613, + 0.042816102504730225, + 0.029420608654618263, + 0.02672025002539158, + 0.05436239391565323, + 0.03377271816134453, + 0.03916771337389946, + 0.1292196363210678, + 0.32474228739738464, + 0.11908341199159622 + ], + "credit_map": [ + 0.006400220561772585, + 0.0075539411045610905, + 0.007067427970468998, + 0.015589217655360699, + 0.009541645646095276, + 0.006932221818715334, + 0.07635428011417389, + 0.07125597447156906, + 0.0027102744206786156, + 0.0036736063193529844, + 0.0037232486065477133, + 0.005905941128730774, + 0.004128213971853256, + 0.002762424061074853, + 0.011757665313780308, + 0.008154729381203651, + 0.0017052805051207542, + 0.0026498856022953987, + 0.0031892077531665564, + 0.003757795551791787, + 0.0031251877080649137, + 0.002831435529515147, + 0.006594767794013023, + 0.005567049607634544, + 0.002619076520204544, + 0.00449092872440815, + 0.00381271168589592, + 0.0025520380586385727, + 0.003840927965939045, + 0.003423643996939063, + 0.0027863665018230677, + 0.003194559132680297, + 0.003793503390625119, + 0.005079768132418394, + 0.005704012233763933, + 0.004878109786659479, + 0.005838577635586262, + 0.004941191058605909, + 0.012959842570126057, + 0.011167391203343868, + 0.00414081709459424, + 0.004549361765384674, + 0.004715492948889732, + 0.00459351995959878, + 0.0036803174298256636, + 0.003805084154009819, + 0.0035062357783317566, + 0.004781892988830805, + 0.004055541008710861, + 0.0045115831308066845, + 0.008534596301615238, + 0.005223108921200037, + 0.0036524825263768435, + 0.0033897473476827145, + 0.004659502301365137, + 0.005141156725585461, + 0.009670461528003216, + 0.008812440559267998, + 0.0123194744810462, + 0.014812063425779343, + 0.0077460757456719875, + 0.009518858976662159, + 0.03322504088282585, + 0.03311522677540779, + 0.02019684575498104, + 0.06233089417219162, + 0.03798326849937439, + 0.04127950221300125, + 0.05009833350777626, + 0.04075755551457405, + 0.05078476294875145, + 0.02131112478673458, + 0.011814557947218418, + 0.016972549259662628, + 0.014875954017043114, + 0.010998818092048168, + 0.020505724474787712, + 0.018100788816809654, + 0.018142879009246826, + 0.0076721422374248505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.016073880717158318 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054952602833509445 + ], + "opposite_logprob_tokens": [ + -2.929952621459961 + ], + "logprob_reward": -0.054952602833509445, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054952602833509445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12215818464756012, + 0.1868685930967331, + 0.05780581757426262, + 0.117235466837883, + 0.035014308989048004, + 0.03622220456600189, + 0.04094839096069336, + 0.11065813153982162, + 0.21491177380084991, + 0.07817711681127548 + ], + "credit_map": [ + 0.00436103530228138, + 0.003969856537878513, + 0.007308007217943668, + 0.01401496771723032, + 0.008625668473541737, + 0.004612683318555355, + 0.041663385927677155, + 0.037602584809064865, + 0.007205937057733536, + 0.0052577960304915905, + 0.008700557984411716, + 0.03474581614136696, + 0.010492493398487568, + 0.006539886351674795, + 0.07356218993663788, + 0.04036391153931618, + 0.003092904109507799, + 0.004269614350050688, + 0.0055396705865859985, + 0.00770248007029295, + 0.005865714978426695, + 0.004839135333895683, + 0.014424184337258339, + 0.012072116136550903, + 0.012867990881204605, + 0.009699635207653046, + 0.015196134336292744, + 0.012061238288879395, + 0.021516578271985054, + 0.01744183339178562, + 0.014411034993827343, + 0.014041022397577763, + 0.0044403173960745335, + 0.0043855104595422745, + 0.006211664993315935, + 0.003222360974177718, + 0.005572035443037748, + 0.0033215421717613935, + 0.003951054532080889, + 0.0039098234847188, + 0.003076342400163412, + 0.004395433235913515, + 0.006904188543558121, + 0.005297692026942968, + 0.003866500686854124, + 0.0032352269627153873, + 0.0033868392929434776, + 0.006059978157281876, + 0.003764915280044079, + 0.004842998925596476, + 0.0072358171455562115, + 0.00537889776751399, + 0.00443916954100132, + 0.005362203344702721, + 0.00499311275780201, + 0.0049312785267829895, + 0.006975620985031128, + 0.008230148814618587, + 0.011194867081940174, + 0.013944928534328938, + 0.007786705158650875, + 0.00935195293277502, + 0.026193400844931602, + 0.02698051929473877, + 0.013632653281092644, + 0.022714698687195778, + 0.01554726343601942, + 0.025602150708436966, + 0.031954649835824966, + 0.03512267768383026, + 0.055722590535879135, + 0.01461509708315134, + 0.00433707470074296, + 0.007119184825569391, + 0.007883384823799133, + 0.009320981800556183, + 0.015235473401844501, + 0.017747050151228905, + 0.010378365404903889, + 0.006155603099614382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.054952602833509445 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, the camera slowly pans to the right, following the movement. On the smooth lacquer surface of a dark, finely grained antique wooden table, scattered haphazardly are a string of pearls of varying sizes, the remains of a broken necklace, emitting a warm, milky white sheen. A warm yellow table lamp, shimmering obliquely from the side, casts a bright spot of light on the curved surface of each pearl. A slender, fair female hand, nails painted crimson, enters the frame with a decisive and fluid gesture. Without hesitation, it sweeps all the pearls towards the edge of the table, from left to right. Instantly, the pearls cascade down like a swarm of startled creatures, crashing onto the polished dark hardwood floor below with a crisp, dense sound. They then roll in all directions, eventually disappearing into the shadows of the frame. The entire scene, dimly lit, creates a dramatic atmosphere of elegance, decisiveness, and a touch of destructive beauty.", + "seed": 8839, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_12/index_7386_seed_8839_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03178681116551161, + "vlm_logprob": -0.03178681116551161, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.03178681116551161, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03421029448509216 + ], + "opposite_logprob_tokens": [ + -4.159210205078125 + ], + "logprob_reward": -0.03421029448509216, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.03421029448509216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28771594166755676, + 0.11426471173763275, + 0.012598571367561817, + 0.021912764757871628, + 0.03917821869254112, + 0.3335523307323456, + 0.04767099767923355, + 0.0558602437376976, + 0.04072757065296173, + 0.04651856794953346 + ], + "credit_map": [ + 0.034729331731796265, + 0.03070446103811264, + 0.03436068072915077, + 0.03142908588051796, + 0.035121530294418335, + 0.03254421427845955, + 0.042081307619810104, + 0.04674536734819412, + 0.014916699379682541, + 0.009318938478827477, + 0.01557066384702921, + 0.011023790575563908, + 0.010757043957710266, + 0.017757773399353027, + 0.02288999781012535, + 0.01202982198446989, + 0.0012998158344998956, + 0.0019878714811056852, + 0.0011760586639866233, + 0.0011051413603127003, + 0.0013964823447167873, + 0.0019194044871255755, + 0.0015419188421219587, + 0.0021718807984143496, + 0.0021779797971248627, + 0.002773054875433445, + 0.0023009246215224266, + 0.0025089161936193705, + 0.004237274639308453, + 0.002988067688420415, + 0.0028609714936465025, + 0.002065579406917095, + 0.00366988149471581, + 0.005418608896434307, + 0.004560206551104784, + 0.005382150411605835, + 0.007628472521901131, + 0.005105822347104549, + 0.0042750900611281395, + 0.0031379908323287964, + 0.01684069074690342, + 0.013397677801549435, + 0.1139611080288887, + 0.1329609900712967, + 0.007209782022982836, + 0.031855612993240356, + 0.008925280533730984, + 0.008401228114962578, + 0.007642692886292934, + 0.008471554145216942, + 0.006251722574234009, + 0.003166635287925601, + 0.007377448491752148, + 0.004778990987688303, + 0.004683396779000759, + 0.005298562813550234, + 0.00575066776946187, + 0.010390914045274258, + 0.006148920860141516, + 0.0053075579926371574, + 0.005914550274610519, + 0.007265164982527494, + 0.007488951086997986, + 0.007593526970595121, + 0.0031371447257697582, + 0.0033399041276425123, + 0.003292388515546918, + 0.006185963749885559, + 0.007965905591845512, + 0.005433422978967428, + 0.005853796377778053, + 0.005519047845155001, + 0.002642503008246422, + 0.0038409715052694082, + 0.00887676328420639, + 0.004462531767785549, + 0.008719045668840408, + 0.006262826267629862, + 0.005209559109061956, + 0.0065043712966144085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.03421029448509216 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011097141541540623 + ], + "opposite_logprob_tokens": [ + -4.511096954345703 + ], + "logprob_reward": -0.011097141541540623, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011097141541540623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2878531813621521, + 0.0749523714184761, + 0.04389369487762451, + 0.05429534614086151, + 0.03678150847554207, + 0.3034350275993347, + 0.08609126508235931, + 0.037108954042196274, + 0.04412844032049179, + 0.03146025910973549 + ], + "credit_map": [ + 0.03692445158958435, + 0.03573978692293167, + 0.03704756125807762, + 0.02702365815639496, + 0.041519492864608765, + 0.035931844264268875, + 0.03437940031290054, + 0.03928699344396591, + 0.006260619033128023, + 0.004845519550144672, + 0.009351486340165138, + 0.006212668027728796, + 0.006562955677509308, + 0.01523509994149208, + 0.014359705150127411, + 0.012124315835535526, + 0.004824751056730747, + 0.0058453381061553955, + 0.0051197572611272335, + 0.004148583393543959, + 0.007287178188562393, + 0.005896951071918011, + 0.005122008733451366, + 0.005649126600474119, + 0.004978757351636887, + 0.005325395613908768, + 0.004081280902028084, + 0.0067421370185911655, + 0.012424302287399769, + 0.007810176350176334, + 0.007950903847813606, + 0.004982390906661749, + 0.004625235218554735, + 0.005934194196015596, + 0.003492717631161213, + 0.004000147804617882, + 0.0069926162250339985, + 0.004685248248279095, + 0.004176540300250053, + 0.0028748088516294956, + 0.017008235678076744, + 0.012388135306537151, + 0.10389447957277298, + 0.1210937350988388, + 0.005775294732302427, + 0.028511937707662582, + 0.008434813469648361, + 0.006328393239527941, + 0.010097441263496876, + 0.01207365095615387, + 0.011866237036883831, + 0.006596652325242758, + 0.016110043972730637, + 0.008767824620008469, + 0.007322229444980621, + 0.01325718779116869, + 0.0035982646513730288, + 0.005003524478524923, + 0.0038303937762975693, + 0.002801802009344101, + 0.003766523441299796, + 0.010017029941082, + 0.0030757884960621595, + 0.005015628878027201, + 0.003396421205252409, + 0.002814771141856909, + 0.003317415714263916, + 0.006779468152672052, + 0.00768700847402215, + 0.005241553299129009, + 0.008642229251563549, + 0.006249571219086647, + 0.0022052363492548466, + 0.003540663281455636, + 0.004987848456948996, + 0.004675555042922497, + 0.004157163668423891, + 0.0037180082872509956, + 0.0043175602331757545, + 0.00385822425596416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.011097141541540623 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495023354887962 + ], + "opposite_logprob_tokens": [ + -2.92995023727417 + ], + "logprob_reward": -0.05495023354887962, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495023354887962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19843925535678864, + 0.13598580658435822, + 0.02917042002081871, + 0.030297622084617615, + 0.1181061640381813, + 0.2471380978822708, + 0.11175975203514099, + 0.045995235443115234, + 0.04340456426143646, + 0.03970310837030411 + ], + "credit_map": [ + 0.02406608872115612, + 0.021845510229468346, + 0.023211099207401276, + 0.021988585591316223, + 0.024291766807436943, + 0.02249746210873127, + 0.0284334197640419, + 0.03210531547665596, + 0.015732048079371452, + 0.010342412628233433, + 0.016574067994952202, + 0.013555081561207771, + 0.01366044208407402, + 0.01585541106760502, + 0.031647343188524246, + 0.01861898973584175, + 0.003014695830643177, + 0.0029090314637869596, + 0.003547544125467539, + 0.0021241663489490747, + 0.006653048563748598, + 0.004399658180773258, + 0.0031130961142480373, + 0.00340917962603271, + 0.00299816089682281, + 0.0033934449311345816, + 0.0022677842061966658, + 0.004481567069888115, + 0.00549897737801075, + 0.00441284803673625, + 0.004552766215056181, + 0.0026920728851109743, + 0.016161400824785233, + 0.021781397983431816, + 0.011238214559853077, + 0.009854758158326149, + 0.0218953900039196, + 0.015028517693281174, + 0.012667977251112461, + 0.009478507563471794, + 0.011271863244473934, + 0.010535798966884613, + 0.08771474659442902, + 0.0994945839047432, + 0.004109042230993509, + 0.021236417815089226, + 0.007850156165659428, + 0.004925491288304329, + 0.013328144326806068, + 0.015176668763160706, + 0.015108413994312286, + 0.008829685859382153, + 0.021365078166127205, + 0.012490355409681797, + 0.00904521532356739, + 0.01641618087887764, + 0.004258415661752224, + 0.006052626762539148, + 0.003939512185752392, + 0.0043366821482777596, + 0.006740962620824575, + 0.007714141625910997, + 0.007823118008673191, + 0.0051297773607075214, + 0.0029556849040091038, + 0.002800947055220604, + 0.0021204298827797174, + 0.007536084856837988, + 0.004595767240971327, + 0.005234466399997473, + 0.013264710083603859, + 0.004896474536508322, + 0.0026445614639669657, + 0.0020287304650992155, + 0.006846535950899124, + 0.005548117682337761, + 0.005848123226314783, + 0.005914817564189434, + 0.00492522120475769, + 0.005947000812739134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.05495023354887962 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581724122166634 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581724122166634, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581724122166634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12640166282653809, + 0.03307577967643738, + 0.04703306779265404, + 0.07782379537820816, + 0.0572480671107769, + 0.3974781930446625, + 0.10982827097177505, + 0.0777086466550827, + 0.04412931576371193, + 0.02927314303815365 + ], + "credit_map": [ + 0.018190573900938034, + 0.013664180412888527, + 0.015840595588088036, + 0.013988870196044445, + 0.01351290475577116, + 0.01428277138620615, + 0.016225337982177734, + 0.020696425810456276, + 0.003907963167876005, + 0.0030352796893566847, + 0.004681768361479044, + 0.0027768658474087715, + 0.0037141116335988045, + 0.00467404630035162, + 0.006115592550486326, + 0.004170150961726904, + 0.0065740495920181274, + 0.00647309422492981, + 0.00610970938578248, + 0.0041980259120464325, + 0.006006661802530289, + 0.006797690410166979, + 0.0055044712498784065, + 0.005369365680962801, + 0.008726686239242554, + 0.006358369719237089, + 0.006191558670252562, + 0.010993083007633686, + 0.015953341498970985, + 0.010679757222533226, + 0.011869322508573532, + 0.007051676511764526, + 0.006982443854212761, + 0.01049251202493906, + 0.005781522020697594, + 0.005210878327488899, + 0.010910592041909695, + 0.00710521312430501, + 0.005962612573057413, + 0.00480229128152132, + 0.019090555608272552, + 0.015562577173113823, + 0.13951362669467926, + 0.1602032482624054, + 0.00678549287840724, + 0.03622400760650635, + 0.011352472007274628, + 0.008746221661567688, + 0.011979270726442337, + 0.017423618584871292, + 0.014906100928783417, + 0.008654119446873665, + 0.022208984941244125, + 0.0095975361764431, + 0.00976751558482647, + 0.015291121788322926, + 0.007389219943434, + 0.009457694366574287, + 0.004647401627153158, + 0.005193617194890976, + 0.007043763995170593, + 0.020263899117708206, + 0.009800760075449944, + 0.013912287540733814, + 0.004024484194815159, + 0.003881592769175768, + 0.004643076565116644, + 0.004979059565812349, + 0.009524228051304817, + 0.0051529002375900745, + 0.00487867183983326, + 0.00704529881477356, + 0.002567051211372018, + 0.0023475599009543657, + 0.004969906993210316, + 0.002624992746859789, + 0.0046097030863165855, + 0.004297755192965269, + 0.0035777029115706682, + 0.004278469365090132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.020581724122166634 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a smooth, dark, finely grained antique wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are scattered pearls of varying sizes and a broken necklace visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender fair female hand with crimson nail polish enter the frame and sweep the pearls from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls cascade down from the table onto the dark hardwood floor and roll into the shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right following the movement of the hand and pearls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038094662129879 + ], + "opposite_logprob_tokens": [ + -3.2880947589874268 + ], + "logprob_reward": -0.038094662129879, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038094662129879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.034009069204330444, + 0.05828994885087013, + 0.013109383173286915, + 0.029070233926177025, + 0.06207524985074997, + 0.6599081754684448, + 0.06177307665348053, + 0.030643202364444733, + 0.031943149864673615, + 0.019178487360477448 + ], + "credit_map": [ + 0.005320631433278322, + 0.003982915543019772, + 0.0042125810869038105, + 0.0035239672288298607, + 0.0039354851469397545, + 0.003928354475647211, + 0.0035940567031502724, + 0.005511080380529165, + 0.007364936172962189, + 0.00472725834697485, + 0.00707950908690691, + 0.005901230499148369, + 0.005948148667812347, + 0.007121633738279343, + 0.012838916853070259, + 0.007308321073651314, + 0.001401726040057838, + 0.0014199455035850406, + 0.0015597684541717172, + 0.0009840653510764241, + 0.002870198804885149, + 0.00183566112536937, + 0.001456549041904509, + 0.0015814702492207289, + 0.003394728060811758, + 0.0028694316279143095, + 0.0027872505597770214, + 0.0042160083539783955, + 0.0045513976365327835, + 0.003847386222332716, + 0.0048001473769545555, + 0.002603887114673853, + 0.008103757165372372, + 0.009133324958384037, + 0.005887100473046303, + 0.006674281321465969, + 0.011167916469275951, + 0.009131611324846745, + 0.006884878035634756, + 0.005092385690659285, + 0.024877680465579033, + 0.02796427719295025, + 0.23776698112487793, + 0.2680903971195221, + 0.010963177308440208, + 0.057047247886657715, + 0.01989409141242504, + 0.013304328545928001, + 0.00617469334974885, + 0.009269225411117077, + 0.009120246395468712, + 0.004585864022374153, + 0.012041079811751842, + 0.006214409600943327, + 0.006547650322318077, + 0.007819910533726215, + 0.0045024072751402855, + 0.00456531485542655, + 0.0027997030410915613, + 0.0031312094070017338, + 0.0030064366292208433, + 0.0041199056431651115, + 0.0037223934195935726, + 0.004795834422111511, + 0.0026494800113141537, + 0.0032385881058871746, + 0.0022280632983893156, + 0.003345285076647997, + 0.007454811595380306, + 0.004223235417157412, + 0.004262947477400303, + 0.004540739580988884, + 0.0013498709304258227, + 0.001344500808045268, + 0.003061279421672225, + 0.002216385444626212, + 0.0034438464790582657, + 0.0027157277800142765, + 0.002222056034952402, + 0.0028248210437595844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.038094662129879 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_13/index_tensor([2049, 2049])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_13/index_tensor([2049, 2049])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..563f675d8264c3e4459d26a98db93dd47ea7dfa3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_13/index_tensor([2049, 2049])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fluid single shot begins from a slightly upward angle, the camera steadily moving forward, following the steps of a silver-haired young man as he walks down a narrow, damp back alley. The scene is shrouded in a chilling blue hue. The mottled brick walls on either side are stained with water, and the puddles on the ground reflect a cold light. A single, old-fashioned orange streetlight at the alley's entrance casts a cone of warm light, slicing through the deep blue of the night like a sharp sword, creating a striking visual contrast between warm and cold. The young man's short, distinctly silver hair stands up like icy spikes. The warm light highlights the sharp contours of his hair and profile, while his front is largely immersed in a cold blue shadow. A pair of eagle-like eyes gleam fiercely within the shadows. Clad in a worn black leather motorcycle jacket, dark jeans, and heavy combat boots, he walks with a steady, menacing gait, closing in on his cornered opponent. When he finally stopped, his body leaned forward slightly, invading the other person's personal space, his jawline tense, his lips silently opening and closing, as if he was questioning in a very low but powerful voice. The whole picture was full of dramatic tension and a tense atmosphere that was about to break out.", + "seed": 10810, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_13/index_2049_seed_10810_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.025469908607192338, + "vlm_logprob": -0.025469908607192338, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.025469908607192338, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05183224752545357 + ], + "opposite_logprob_tokens": [ + -3.0518321990966797 + ], + "logprob_reward": -0.05183224752545357, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05183224752545357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07724342495203018, + 0.034921810030937195, + 0.03392580524086952, + 0.04882212355732918, + 0.20241306722164154, + 0.0708242729306221, + 0.1199420765042305, + 0.18811176717281342, + 0.06525013595819473, + 0.15854550898075104 + ], + "credit_map": [ + 0.006956553552299738, + 0.00751624908298254, + 0.007485410198569298, + 0.00783360656350851, + 0.0069987173192203045, + 0.018095713108778, + 0.00610552541911602, + 0.01625165157020092, + 0.004064200446009636, + 0.0030980752781033516, + 0.0046231551095843315, + 0.003041223157197237, + 0.004070346709340811, + 0.0037130615673959255, + 0.006299230735749006, + 0.006012523081153631, + 0.0033582805190235376, + 0.0026209845673292875, + 0.004694753792136908, + 0.003448821371421218, + 0.006076582241803408, + 0.003219432896003127, + 0.006455319933593273, + 0.004051628522574902, + 0.004733830690383911, + 0.0049334438517689705, + 0.0055667622946202755, + 0.00604218291118741, + 0.006375929806381464, + 0.006575440056622028, + 0.008482998237013817, + 0.006111538968980312, + 0.01567942276597023, + 0.020470110699534416, + 0.02048390917479992, + 0.018936891108751297, + 0.05533241480588913, + 0.018973305821418762, + 0.029527509585022926, + 0.02300950326025486, + 0.006752204615622759, + 0.008119143545627594, + 0.00663095572963357, + 0.006275512743741274, + 0.015486634336411953, + 0.010177315212786198, + 0.010206378065049648, + 0.007176131475716829, + 0.008965040557086468, + 0.013585822656750679, + 0.008984042331576347, + 0.013462111353874207, + 0.025148281827569008, + 0.016422709450125694, + 0.0167362280189991, + 0.01663784123957157, + 0.02689141221344471, + 0.019446294754743576, + 0.021872691810131073, + 0.02106095850467682, + 0.029359782114624977, + 0.020299753174185753, + 0.0225457064807415, + 0.026635171845555305, + 0.007282348349690437, + 0.007051448803395033, + 0.01146035734564066, + 0.006993293296545744, + 0.008441674523055553, + 0.008050665259361267, + 0.010042588226497173, + 0.0059277648106217384, + 0.012795867398381233, + 0.01562540791928768, + 0.014092585071921349, + 0.016702722758054733, + 0.02869955636560917, + 0.030386460945010185, + 0.02244388312101364, + 0.017799027264118195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.05183224752545357 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001049801823683083 + ], + "opposite_logprob_tokens": [ + -6.876049995422363 + ], + "logprob_reward": -0.001049801823683083, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001049801823683083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17863819003105164, + 0.13654325902462006, + 0.2030019760131836, + 0.10623927414417267, + 0.04019820690155029, + 0.08339464664459229, + 0.06485976278781891, + 0.05268922448158264, + 0.03226730227470398, + 0.10216817259788513 + ], + "credit_map": [ + 0.03375714272260666, + 0.01735602505505085, + 0.02458089217543602, + 0.017472539097070694, + 0.01615993119776249, + 0.030807923525571823, + 0.01598675735294819, + 0.02251698449254036, + 0.02477799728512764, + 0.00961472187191248, + 0.022551154717803, + 0.01056305319070816, + 0.015958841890096664, + 0.02181067503988743, + 0.012529660016298294, + 0.018737155944108963, + 0.03042484074831009, + 0.035970211029052734, + 0.03157377988100052, + 0.01817817986011505, + 0.020341429859399796, + 0.012967832386493683, + 0.023826133459806442, + 0.029719581827521324, + 0.01592225953936577, + 0.009683254174888134, + 0.01833248697221279, + 0.00980781577527523, + 0.012267091311514378, + 0.014552020467817783, + 0.016069814562797546, + 0.009604529477655888, + 0.003819073783233762, + 0.00514020724222064, + 0.005479900632053614, + 0.00393859064206481, + 0.007873805239796638, + 0.0050454954616725445, + 0.00531576806679368, + 0.003585370723158121, + 0.008880803361535072, + 0.00830765813589096, + 0.008187482133507729, + 0.010888876393437386, + 0.010795063339173794, + 0.009973789565265179, + 0.015144762583076954, + 0.011216208338737488, + 0.009333809837698936, + 0.008594482205808163, + 0.006107495632022619, + 0.0074987816624343395, + 0.007399875670671463, + 0.008435884490609169, + 0.008427008986473083, + 0.009062422439455986, + 0.009241227991878986, + 0.005363762844353914, + 0.007635656278580427, + 0.005116638261824846, + 0.006625032052397728, + 0.0057647619396448135, + 0.006272480357438326, + 0.0066696638241410255, + 0.004531004931777716, + 0.0032748491503298283, + 0.007624784484505653, + 0.002781942253932357, + 0.003139858366921544, + 0.0048258183524012566, + 0.003158966079354286, + 0.002930077724158764, + 0.00976709183305502, + 0.017237305641174316, + 0.010742071084678173, + 0.009843108244240284, + 0.014850745908915997, + 0.011771777644753456, + 0.013543390668928623, + 0.014412673190236092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.001049801823683083 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864376410841942 + ], + "opposite_logprob_tokens": [ + -3.0486438274383545 + ], + "logprob_reward": -0.04864376410841942, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864376410841942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17451895773410797, + 0.09547671675682068, + 0.10344269126653671, + 0.17435631155967712, + 0.064595527946949, + 0.06946004927158356, + 0.1377646028995514, + 0.0698329359292984, + 0.041770368814468384, + 0.06878186017274857 + ], + "credit_map": [ + 0.0201392974704504, + 0.015495658852159977, + 0.019509676843881607, + 0.017618808895349503, + 0.016775980591773987, + 0.03017255663871765, + 0.013509041629731655, + 0.04129793122410774, + 0.009148920886218548, + 0.00712973577901721, + 0.013539382256567478, + 0.006321808788925409, + 0.013179964385926723, + 0.01263487245887518, + 0.01000079046934843, + 0.023521240800619125, + 0.008304080925881863, + 0.01099511981010437, + 0.017564481124281883, + 0.011931146495044231, + 0.016064370051026344, + 0.009050375781953335, + 0.015414872206747532, + 0.014118241146206856, + 0.019312696531414986, + 0.030408242717385292, + 0.023692328482866287, + 0.014610790647566319, + 0.036761123687028885, + 0.016776878386735916, + 0.02029801532626152, + 0.012496214359998703, + 0.007552880793809891, + 0.007367854937911034, + 0.006988356355577707, + 0.00716342031955719, + 0.013353055343031883, + 0.0075772725977003574, + 0.00760669307783246, + 0.006985990796238184, + 0.006059012841433287, + 0.005595584399998188, + 0.00491573940962553, + 0.008723430335521698, + 0.008393274620175362, + 0.0066982777789235115, + 0.016225161030888557, + 0.012849567458033562, + 0.020748788490891457, + 0.015986662358045578, + 0.012207383289933205, + 0.01496274396777153, + 0.01320682279765606, + 0.0186723992228508, + 0.017299402505159378, + 0.02468038536608219, + 0.005304577294737101, + 0.0064393444918096066, + 0.010829050093889236, + 0.00749854976311326, + 0.010159929282963276, + 0.010149920359253883, + 0.009189595468342304, + 0.010261965915560722, + 0.0038440716452896595, + 0.005819409154355526, + 0.0062270076014101505, + 0.004943398758769035, + 0.005345574580132961, + 0.005451129283756018, + 0.005311461631208658, + 0.004828319884836674, + 0.0064983246847987175, + 0.00908378604799509, + 0.007786192465573549, + 0.006633936893194914, + 0.009948943741619587, + 0.010342269204556942, + 0.009605002589523792, + 0.008883400820195675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.04864376410841942 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254269111901522 + ], + "opposite_logprob_tokens": [ + -5.25525426864624 + ], + "logprob_reward": -0.005254269111901522, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254269111901522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0840589627623558, + 0.12061715126037598, + 0.08237336575984955, + 0.0999845489859581, + 0.0891551747918129, + 0.05104941874742508, + 0.21551598608493805, + 0.10596060752868652, + 0.03940638154745102, + 0.11187836527824402 + ], + "credit_map": [ + 0.007155693601816893, + 0.008630438707768917, + 0.008548443205654621, + 0.007935116067528725, + 0.009962175041437149, + 0.021402470767498016, + 0.005507128778845072, + 0.014917491935193539, + 0.012942378409206867, + 0.010008675046265125, + 0.0181205403059721, + 0.01291163545101881, + 0.01722113788127899, + 0.012921207584440708, + 0.025177491828799248, + 0.011314089410007, + 0.006528822239488363, + 0.015401521697640419, + 0.01103141624480486, + 0.007008059415966272, + 0.011020069941878319, + 0.0096864253282547, + 0.011464057490229607, + 0.010232995264232159, + 0.01185757014900446, + 0.011501832865178585, + 0.013270166702568531, + 0.008648470044136047, + 0.015808558091521263, + 0.01582149975001812, + 0.013986398465931416, + 0.009090051054954529, + 0.006448436062783003, + 0.008996647782623768, + 0.007317269686609507, + 0.008682978339493275, + 0.02158930152654648, + 0.01031674537807703, + 0.013046047650277615, + 0.01275775395333767, + 0.004382168408483267, + 0.004687125328928232, + 0.005045104771852493, + 0.008332963101565838, + 0.006007209420204163, + 0.009155015461146832, + 0.007688918150961399, + 0.005750914569944143, + 0.023664455860853195, + 0.023237543180584908, + 0.01889490894973278, + 0.028473762795329094, + 0.03396898880600929, + 0.028814833611249924, + 0.02488120272755623, + 0.033580292016267776, + 0.014670459553599358, + 0.011250545270740986, + 0.017656629905104637, + 0.00928043294698, + 0.017518721520900726, + 0.013481313362717628, + 0.01099028903990984, + 0.011112215928733349, + 0.005124131217598915, + 0.0038869911804795265, + 0.006942110136151314, + 0.003419511718675494, + 0.00403605354949832, + 0.005689067766070366, + 0.005407158751040697, + 0.004901358857750893, + 0.010150042362511158, + 0.014148108661174774, + 0.008434119634330273, + 0.010828081518411636, + 0.020315876230597496, + 0.016382262110710144, + 0.01789025031030178, + 0.013729620724916458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 9 + ], + "selected_reward": -0.005254269111901522 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020569460466504097 + ], + "opposite_logprob_tokens": [ + -3.8955695629119873 + ], + "logprob_reward": -0.020569460466504097, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020569460466504097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10644534230232239, + 0.112875796854496, + 0.17434878647327423, + 0.0816083699464798, + 0.06857641041278839, + 0.10966407507658005, + 0.20175811648368835, + 0.052075691521167755, + 0.03861461579799652, + 0.05403279885649681 + ], + "credit_map": [ + 0.022057896479964256, + 0.008284330368041992, + 0.015421627089381218, + 0.01347906794399023, + 0.010471992194652557, + 0.015688443556427956, + 0.007856333628296852, + 0.013185645453631878, + 0.010020270012319088, + 0.009031779132783413, + 0.017335450276732445, + 0.010813887231051922, + 0.02381536364555359, + 0.009945884346961975, + 0.02081035077571869, + 0.011102807708084583, + 0.024485284462571144, + 0.030037306249141693, + 0.02369442582130432, + 0.019273487851023674, + 0.017694048583507538, + 0.01914435625076294, + 0.022680507972836494, + 0.017339367419481277, + 0.0056679765693843365, + 0.009340406395494938, + 0.009858024306595325, + 0.009544569067656994, + 0.012010795064270496, + 0.015001390129327774, + 0.011390598490834236, + 0.008794602937996387, + 0.0057752132415771484, + 0.009934036992490292, + 0.0064446707256138325, + 0.006197908893227577, + 0.015458513051271439, + 0.008671196177601814, + 0.009083074517548084, + 0.007011795416474342, + 0.008136114105582237, + 0.013091904111206532, + 0.0077961948700249195, + 0.015619461424648762, + 0.01933896169066429, + 0.009619415737688541, + 0.02184157818555832, + 0.01422044262290001, + 0.023023372516036034, + 0.020118162035942078, + 0.017912408336997032, + 0.019954554736614227, + 0.023536760360002518, + 0.03621137887239456, + 0.02549166977405548, + 0.03550979867577553, + 0.005689989775419235, + 0.004849471151828766, + 0.006652709562331438, + 0.0049133943393826485, + 0.006032000295817852, + 0.008563847281038761, + 0.006654425524175167, + 0.0087198531255126, + 0.005061792675405741, + 0.004360358230769634, + 0.006494161672890186, + 0.003836805932223797, + 0.005118330474942923, + 0.005685102194547653, + 0.00446982029825449, + 0.0035882426891475916, + 0.0059144459664821625, + 0.005619116593152285, + 0.006591708865016699, + 0.005080781411379576, + 0.008551378734409809, + 0.0061899772845208645, + 0.009725378826260567, + 0.006360010709613562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.020569460466504097 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid single shot begins from a slightly upward angle, the camera steadily moving forward, following the steps of a silver-haired young man as he walks down a narrow, damp back alley. The scene is shrouded in a chilling blue hue. The mottled brick walls on either side are stained with water, and the puddles on the ground reflect a cold light. A single, old-fashioned orange streetlight at the alley's entrance casts a cone of warm light, slicing through the deep blue of the night like a sharp sword, creating a striking visual contrast between warm and cold. The young man's short, distinctly silver hair stands up like icy spikes. The warm light highlights the sharp contours of his hair and profile, while his front is largely immersed in a cold blue shadow. A pair of eagle-like eyes gleam fiercely within the shadows. Clad in a worn black leather motorcycle jacket, dark jeans, and heavy combat boots, he walks with a steady, menacing gait, closing in on his cornered opponent. When he finally stopped, his body leaned forward slightly, invading the other person's personal space, his jawline tense, his lips silently opening and closing, as if he was questioning in a very low but powerful voice. The whole picture was full of dramatic tension and a tense atmosphere that was about to break out.", + "seed": 10811, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_13/index_2049_seed_10811_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02073831909801811, + "vlm_logprob": -0.02073831909801811, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.02073831909801811, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.053298428654670715 + ], + "opposite_logprob_tokens": [ + -3.0532984733581543 + ], + "logprob_reward": -0.053298428654670715, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.053298428654670715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07375163584947586, + 0.017312902957201004, + 0.014304117299616337, + 0.022138457745313644, + 0.018511652946472168, + 0.04084505885839462, + 0.02605229802429676, + 0.18890783190727234, + 0.16146357357501984, + 0.43671250343322754 + ], + "credit_map": [ + 0.008264523930847645, + 0.0059730373322963715, + 0.012682405300438404, + 0.00544208288192749, + 0.004335383418947458, + 0.01510391104966402, + 0.003228552406653762, + 0.018721740692853928, + 0.0016768237110227346, + 0.002339872531592846, + 0.002404215279966593, + 0.0029001403599977493, + 0.0015163019997999072, + 0.0016645754221826792, + 0.001405055751092732, + 0.003405919298529625, + 0.003946289885789156, + 0.0015171980485320091, + 0.002511423546820879, + 0.0013208019081503153, + 0.0008765378734096885, + 0.0012476721312850714, + 0.0014859018847346306, + 0.0013982921373099089, + 0.0027278398629277945, + 0.0031840812880545855, + 0.002917500678449869, + 0.0029509880114346743, + 0.0019894151482731104, + 0.002289936412125826, + 0.0020045991986989975, + 0.0040740966796875, + 0.0038388781249523163, + 0.00179652264341712, + 0.0028059917967766523, + 0.0021866450551897287, + 0.0016116718761622906, + 0.002115360228344798, + 0.0019769123755395412, + 0.002179669914767146, + 0.005753748584538698, + 0.003884701756760478, + 0.005973228719085455, + 0.003779173828661442, + 0.004453596193343401, + 0.0048912810161709785, + 0.006866095121949911, + 0.005243232008069754, + 0.0050004166550934315, + 0.001772640272974968, + 0.0038451363798230886, + 0.0041052005253732204, + 0.0024238694459199905, + 0.0026147214230149984, + 0.0035721284803003073, + 0.0027181850746273994, + 0.018309710547327995, + 0.008224444463849068, + 0.061283741146326065, + 0.06871733069419861, + 0.0055856951512396336, + 0.006984895095229149, + 0.012706568464636803, + 0.007095453329384327, + 0.023942189291119576, + 0.01687435433268547, + 0.02304317243397236, + 0.027273425832390785, + 0.017001783475279808, + 0.019508007913827896, + 0.020764706656336784, + 0.013055935502052307, + 0.04389012232422829, + 0.03699299693107605, + 0.047549180686473846, + 0.057058416306972504, + 0.08499383181333542, + 0.051029618829488754, + 0.07588407397270203, + 0.039314255118370056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.053298428654670715 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019527667900547385 + ], + "opposite_logprob_tokens": [ + -6.251952648162842 + ], + "logprob_reward": -0.0019527667900547385, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019527667900547385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1753048449754715, + 0.117482028901577, + 0.10650911182165146, + 0.06173483654856682, + 0.10478208214044571, + 0.047275930643081665, + 0.048423610627651215, + 0.1488259881734848, + 0.11822877824306488, + 0.07143279165029526 + ], + "credit_map": [ + 0.018086988478899002, + 0.01484943088144064, + 0.021206112578511238, + 0.01917297951877117, + 0.009787380695343018, + 0.036254290491342545, + 0.014341255649924278, + 0.041606396436691284, + 0.020123086869716644, + 0.011302215978503227, + 0.019265824928879738, + 0.015040110796689987, + 0.008646593429148197, + 0.01074291206896305, + 0.008949363604187965, + 0.023411918431520462, + 0.02999790385365486, + 0.009914839640259743, + 0.017358984798192978, + 0.010303082875907421, + 0.0065881493501365185, + 0.011271767318248749, + 0.00788537785410881, + 0.013189012184739113, + 0.011064274236559868, + 0.006316372659057379, + 0.008770165964961052, + 0.007356751710176468, + 0.008148822002112865, + 0.004963585175573826, + 0.008150228299200535, + 0.006964638363569975, + 0.016339752823114395, + 0.010587036609649658, + 0.013768796809017658, + 0.01656358130276203, + 0.0063536809757351875, + 0.013584665954113007, + 0.013863837346434593, + 0.013720731250941753, + 0.010501988232135773, + 0.003538009012117982, + 0.007848075591027737, + 0.004749665502458811, + 0.004519772715866566, + 0.003912466578185558, + 0.0070578353479504585, + 0.005148119293153286, + 0.013389771804213524, + 0.002952978014945984, + 0.00754045182839036, + 0.004883017390966415, + 0.0033043990842998028, + 0.004893864039331675, + 0.0046147932298481464, + 0.006844335235655308, + 0.013053377158939838, + 0.0062420605681836605, + 0.04833977669477463, + 0.05697306990623474, + 0.00367533415555954, + 0.004399018827825785, + 0.01077697891741991, + 0.00536637008190155, + 0.016523247584700584, + 0.011169605888426304, + 0.01689852587878704, + 0.018967537209391594, + 0.013215302489697933, + 0.01562904193997383, + 0.015578517690300941, + 0.0102470014244318, + 0.007881269790232182, + 0.005542243830859661, + 0.008698254823684692, + 0.00931495800614357, + 0.014498013071715832, + 0.005029736086726189, + 0.01256504189223051, + 0.00790327787399292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.0019527667900547385 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015160269103944302 + ], + "opposite_logprob_tokens": [ + -6.501515865325928 + ], + "logprob_reward": -0.0015160269103944302, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015160269103944302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.183360755443573, + 0.10390687733888626, + 0.10504208505153656, + 0.0957539975643158, + 0.0442734956741333, + 0.12277190387248993, + 0.03462158143520355, + 0.08576099574565887, + 0.17864058911800385, + 0.04586770012974739 + ], + "credit_map": [ + 0.014538855291903019, + 0.012134875170886517, + 0.0369785912334919, + 0.013084173202514648, + 0.014560217969119549, + 0.054772648960351944, + 0.014450287446379662, + 0.022841118276119232, + 0.01346564944833517, + 0.010781793855130672, + 0.012172386050224304, + 0.008146124891936779, + 0.016915086656808853, + 0.010153919458389282, + 0.011873318813741207, + 0.020398607477545738, + 0.020066943019628525, + 0.015473747625946999, + 0.014019927941262722, + 0.0098491245880723, + 0.014744891785085201, + 0.01330307126045227, + 0.008024770766496658, + 0.009559622034430504, + 0.009101017378270626, + 0.013208146207034588, + 0.012773559428751469, + 0.014070370234549046, + 0.014296676963567734, + 0.008172913454473019, + 0.00901711080223322, + 0.015114212408661842, + 0.006358789745718241, + 0.004230191931128502, + 0.006337317172437906, + 0.007254829630255699, + 0.004819447174668312, + 0.003167608752846718, + 0.007860774174332619, + 0.004244539421051741, + 0.018191570416092873, + 0.00849829986691475, + 0.01703754998743534, + 0.010416710749268532, + 0.02198917418718338, + 0.014831789769232273, + 0.020610386505723, + 0.011196437291800976, + 0.008380834944546223, + 0.0020321407355368137, + 0.00522661441937089, + 0.004124579019844532, + 0.0035159119870513678, + 0.0030513021629303694, + 0.004077706020325422, + 0.004212494008243084, + 0.007895015180110931, + 0.0038666168693453074, + 0.026975905522704124, + 0.031012488529086113, + 0.0031579937785863876, + 0.002962638856843114, + 0.005850350018590689, + 0.004039989318698645, + 0.025115597993135452, + 0.013653326779603958, + 0.025772638618946075, + 0.024658655747771263, + 0.019604545086622238, + 0.02372799441218376, + 0.026974128559231758, + 0.019133733585476875, + 0.005345732904970646, + 0.006810074206441641, + 0.005507935304194689, + 0.005478628445416689, + 0.007147792726755142, + 0.0040717218071222305, + 0.006967711728066206, + 0.004538109991699457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.0015160269103944302 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059118263423443 + ], + "opposite_logprob_tokens": [ + -3.8955912590026855 + ], + "logprob_reward": -0.02059118263423443, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059118263423443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1269996017217636, + 0.04906120151281357, + 0.04799731820821762, + 0.06864459812641144, + 0.08447398990392685, + 0.08126824349164963, + 0.03926520422101021, + 0.18583190441131592, + 0.18224680423736572, + 0.1342112123966217 + ], + "credit_map": [ + 0.012174621224403381, + 0.009410816244781017, + 0.019804323092103004, + 0.009814941324293613, + 0.006242750212550163, + 0.031609389930963516, + 0.008330711163580418, + 0.029612042009830475, + 0.005287685431540012, + 0.0053681880235672, + 0.006591711193323135, + 0.0033644738141447306, + 0.006133393384516239, + 0.004830850288271904, + 0.00665983697399497, + 0.010825060307979584, + 0.008419161662459373, + 0.005912366788834333, + 0.006067871581763029, + 0.004548696801066399, + 0.0066652558743953705, + 0.005240204278379679, + 0.005509399808943272, + 0.00563435722142458, + 0.009281669743359089, + 0.008638441562652588, + 0.009949151426553726, + 0.008422933518886566, + 0.008726994507014751, + 0.00581721356138587, + 0.008207641541957855, + 0.009600548073649406, + 0.009650730527937412, + 0.007581982295960188, + 0.01155251543968916, + 0.014516174793243408, + 0.0070426007732748985, + 0.005979705136269331, + 0.019719082862138748, + 0.008431199006736279, + 0.01000764686614275, + 0.00855058990418911, + 0.01119780633598566, + 0.008904076181352139, + 0.0070395320653915405, + 0.011469339951872826, + 0.013017881661653519, + 0.011081364937126637, + 0.004704652354121208, + 0.0026768608950078487, + 0.006060396786779165, + 0.004757948685437441, + 0.004803065676242113, + 0.0027130034286528826, + 0.0046690683811903, + 0.008880206383764744, + 0.01629239320755005, + 0.009368238970637321, + 0.05202046036720276, + 0.06443571299314499, + 0.008114280179142952, + 0.008352946490049362, + 0.013344155624508858, + 0.013903713785111904, + 0.027136707678437233, + 0.018591204658150673, + 0.025364931672811508, + 0.03007328324019909, + 0.019747771322727203, + 0.021710501983761787, + 0.02449118345975876, + 0.015131204389035702, + 0.013435021974146366, + 0.008245863951742649, + 0.01768481731414795, + 0.021699920296669006, + 0.02508656121790409, + 0.00773282628506422, + 0.02943110652267933, + 0.010895091108977795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.02059118263423443 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026333190500736237 + ], + "opposite_logprob_tokens": [ + -3.6513330936431885 + ], + "logprob_reward": -0.026333190500736237, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026333190500736237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14837902784347534, + 0.06531157344579697, + 0.04188619926571846, + 0.12921598553657532, + 0.10159537941217422, + 0.04910014197230339, + 0.06597498059272766, + 0.1801309734582901, + 0.06724642217159271, + 0.1511593610048294 + ], + "credit_map": [ + 0.015498645603656769, + 0.008227499201893806, + 0.030055638402700424, + 0.012136450037360191, + 0.00784697849303484, + 0.03796835243701935, + 0.008965510874986649, + 0.02767994813621044, + 0.009873200207948685, + 0.005446708295494318, + 0.01035480946302414, + 0.004879205487668514, + 0.00670922826975584, + 0.006523275747895241, + 0.004841061308979988, + 0.016684086993336678, + 0.008937853388488293, + 0.004026441369205713, + 0.006118356250226498, + 0.004425437655299902, + 0.004021205939352512, + 0.0039933412335813046, + 0.004208214581012726, + 0.006155349314212799, + 0.01855088770389557, + 0.018472569063305855, + 0.018026255071163177, + 0.015306235291063786, + 0.011316063813865185, + 0.01262761652469635, + 0.017429478466510773, + 0.017486877739429474, + 0.010229988023638725, + 0.009131032973527908, + 0.013777771033346653, + 0.01776541955769062, + 0.008216286078095436, + 0.007400984410196543, + 0.0252833254635334, + 0.009790569543838501, + 0.00598570192232728, + 0.0051653217524290085, + 0.0066519733518362045, + 0.006094738841056824, + 0.005655281245708466, + 0.005782361142337322, + 0.008097345940768719, + 0.005667415447533131, + 0.01291823759675026, + 0.0052977995947003365, + 0.009786522015929222, + 0.006972585804760456, + 0.00838816910982132, + 0.00541087007150054, + 0.00810988713055849, + 0.009090907871723175, + 0.019858265295624733, + 0.011804400943219662, + 0.050946470350027084, + 0.05833115428686142, + 0.008487370796501637, + 0.007953161373734474, + 0.011928506195545197, + 0.010821648873388767, + 0.0075366320088505745, + 0.008590314537286758, + 0.008359336294233799, + 0.008651751093566418, + 0.007835828699171543, + 0.006630617659538984, + 0.01181159820407629, + 0.007830342277884483, + 0.01456017792224884, + 0.011784935370087624, + 0.01799117587506771, + 0.023050455376505852, + 0.02694173902273178, + 0.011128904297947884, + 0.03100697696208954, + 0.014694983139634132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.026333190500736237 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid single shot begins from a slightly upward angle, the camera steadily moving forward, following the steps of a silver-haired young man as he walks down a narrow, damp back alley. The scene is shrouded in a chilling blue hue. The mottled brick walls on either side are stained with water, and the puddles on the ground reflect a cold light. A single, old-fashioned orange streetlight at the alley's entrance casts a cone of warm light, slicing through the deep blue of the night like a sharp sword, creating a striking visual contrast between warm and cold. The young man's short, distinctly silver hair stands up like icy spikes. The warm light highlights the sharp contours of his hair and profile, while his front is largely immersed in a cold blue shadow. A pair of eagle-like eyes gleam fiercely within the shadows. Clad in a worn black leather motorcycle jacket, dark jeans, and heavy combat boots, he walks with a steady, menacing gait, closing in on his cornered opponent. When he finally stopped, his body leaned forward slightly, invading the other person's personal space, his jawline tense, his lips silently opening and closing, as if he was questioning in a very low but powerful voice. The whole picture was full of dramatic tension and a tense atmosphere that was about to break out.", + "seed": 10812, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_13/index_2049_seed_10812_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019692787993699314, + "vlm_logprob": -0.019692787993699314, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.019692787993699314, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04624420404434204 + ], + "opposite_logprob_tokens": [ + -3.1712441444396973 + ], + "logprob_reward": -0.04624420404434204, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04624420404434204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06253937631845474, + 0.07387466728687286, + 0.03856978937983513, + 0.028240812942385674, + 0.07332762330770493, + 0.06336098909378052, + 0.204341322183609, + 0.0856463611125946, + 0.2097960263490677, + 0.160302996635437 + ], + "credit_map": [ + 0.009625262580811977, + 0.003401088761165738, + 0.009854421950876713, + 0.00803503580391407, + 0.004550031386315823, + 0.010015545412898064, + 0.005862576421350241, + 0.011195414699614048, + 0.006645828019827604, + 0.0028875262942165136, + 0.008735272102057934, + 0.008774767629802227, + 0.005871616769582033, + 0.013346642255783081, + 0.01410458143800497, + 0.013508431613445282, + 0.004553922452032566, + 0.0034094429574906826, + 0.004970391280949116, + 0.004348498769104481, + 0.0026456443592906, + 0.008428098633885384, + 0.0034167903941124678, + 0.006797000765800476, + 0.002690536668524146, + 0.0025673978962004185, + 0.002985952189192176, + 0.003994116093963385, + 0.0019744394812732935, + 0.005907791666686535, + 0.0037948512472212315, + 0.004325727932155132, + 0.012030377052724361, + 0.006425765808671713, + 0.01890498585999012, + 0.009183810092508793, + 0.004145568702369928, + 0.009460635483264923, + 0.005200686398893595, + 0.007975798100233078, + 0.008226725272834301, + 0.010722054168581963, + 0.008811664767563343, + 0.007299148477613926, + 0.0049389926716685295, + 0.007679244503378868, + 0.008279773406684399, + 0.007403380703181028, + 0.018833531066775322, + 0.059026170521974564, + 0.021289534866809845, + 0.023656992241740227, + 0.012040684930980206, + 0.02392624318599701, + 0.016348592936992645, + 0.02921956777572632, + 0.006563903763890266, + 0.012184574268758297, + 0.01596897840499878, + 0.011935538612306118, + 0.009149841032922268, + 0.012152595445513725, + 0.0096671087667346, + 0.008023818954825401, + 0.0181969553232193, + 0.03491469472646713, + 0.02388981357216835, + 0.04612697660923004, + 0.028681468218564987, + 0.019745344296097755, + 0.019209153950214386, + 0.019031614065170288, + 0.014861900359392166, + 0.028943266719579697, + 0.024223048239946365, + 0.022511791437864304, + 0.019166026264429092, + 0.019514985382556915, + 0.01815124973654747, + 0.01293072197586298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.04624420404434204 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017188312485814095 + ], + "opposite_logprob_tokens": [ + -6.376718997955322 + ], + "logprob_reward": -0.0017188312485814095, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017188312485814095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3684796392917633, + 0.11234956234693527, + 0.07288728654384613, + 0.045407675206661224, + 0.05625006556510925, + 0.025741271674633026, + 0.030478032305836678, + 0.19311444461345673, + 0.05130835250020027, + 0.043983686715364456 + ], + "credit_map": [ + 0.03970775753259659, + 0.014973497949540615, + 0.04420848190784454, + 0.05389311909675598, + 0.021734893321990967, + 0.07872853428125381, + 0.06159992143511772, + 0.05363340303301811, + 0.012122641317546368, + 0.0053053912706673145, + 0.015744296833872795, + 0.015082025900483131, + 0.009939799085259438, + 0.01921618916094303, + 0.0188076663762331, + 0.016131551936268806, + 0.007077512796968222, + 0.005383757874369621, + 0.00911687221378088, + 0.010507174767553806, + 0.00530505133792758, + 0.013035883195698261, + 0.010450816713273525, + 0.012010223232209682, + 0.0033972382079809904, + 0.0026171449571847916, + 0.004281933885067701, + 0.0033689476549625397, + 0.00446750782430172, + 0.014020174741744995, + 0.003409298602491617, + 0.009845430962741375, + 0.008494720794260502, + 0.005256549920886755, + 0.011192087084054947, + 0.006648092996329069, + 0.004911551717668772, + 0.007682030089199543, + 0.00518893264234066, + 0.006876098923385143, + 0.003095389809459448, + 0.003647557459771633, + 0.004195064771920443, + 0.003028252860531211, + 0.0015576971927657723, + 0.0030209701508283615, + 0.0028632760513573885, + 0.004333063960075378, + 0.0025080558843910694, + 0.007782389409840107, + 0.0032660826109349728, + 0.003336625872179866, + 0.001736838254146278, + 0.0033064314629882574, + 0.003487091278657317, + 0.005054518114775419, + 0.017423279583454132, + 0.023795295506715775, + 0.024469120427966118, + 0.04741700366139412, + 0.012784283608198166, + 0.025537623092532158, + 0.022694913670420647, + 0.018992912024259567, + 0.0048838029615581036, + 0.00916275754570961, + 0.005114069674164057, + 0.01065379660576582, + 0.005555881652981043, + 0.0051734657026827335, + 0.0049333092756569386, + 0.005831270478665829, + 0.003670267527922988, + 0.0062690796330571175, + 0.006437055300921202, + 0.005023992154747248, + 0.00477919215336442, + 0.006028390023857355, + 0.007800607476383448, + 0.003975100349634886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0017188312485814095 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046380325220525265 + ], + "opposite_logprob_tokens": [ + -5.379638195037842 + ], + "logprob_reward": -0.0046380325220525265, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046380325220525265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12834927439689636, + 0.07467734813690186, + 0.13352470099925995, + 0.07833275198936462, + 0.0922398716211319, + 0.06042889505624771, + 0.11422023922204971, + 0.18666665256023407, + 0.06246085837483406, + 0.06909945607185364 + ], + "credit_map": [ + 0.017783990129828453, + 0.008964232169091702, + 0.014218988828361034, + 0.017014142125844955, + 0.014956148341298103, + 0.021408485248684883, + 0.014876524917781353, + 0.019126763567328453, + 0.006876883562654257, + 0.0048109195195138454, + 0.0098051642999053, + 0.009760882705450058, + 0.00785934366285801, + 0.011311969719827175, + 0.012965771369636059, + 0.011286416091024876, + 0.01496928371489048, + 0.009331739507615566, + 0.015931710600852966, + 0.010535984300076962, + 0.0175471194088459, + 0.024070896208286285, + 0.017795376479625702, + 0.023342585191130638, + 0.0038211368955671787, + 0.007620260119438171, + 0.008298711851239204, + 0.009357615374028683, + 0.0067601255141198635, + 0.01871872879564762, + 0.008341548964381218, + 0.015414624474942684, + 0.015105086378753185, + 0.006596007850021124, + 0.019461486488580704, + 0.008723495528101921, + 0.008111909963190556, + 0.016386326402425766, + 0.0072743846103549, + 0.010581175796687603, + 0.007786538451910019, + 0.007957698777318, + 0.01049261074513197, + 0.0066222017630934715, + 0.004045671317726374, + 0.007901500910520554, + 0.006091700401157141, + 0.009530971758067608, + 0.014894360676407814, + 0.010016359388828278, + 0.009173923172056675, + 0.015164410695433617, + 0.010926969349384308, + 0.01573203317821026, + 0.01331457681953907, + 0.02499760501086712, + 0.01709994487464428, + 0.02373112179338932, + 0.020468350499868393, + 0.04927820339798927, + 0.010274874046444893, + 0.027522416785359383, + 0.021929366514086723, + 0.01636236533522606, + 0.005758877843618393, + 0.008388186804950237, + 0.00886171218007803, + 0.008494410663843155, + 0.008244631811976433, + 0.008629174903035164, + 0.005282294005155563, + 0.008801568299531937, + 0.00684813316911459, + 0.009363804943859577, + 0.011543921194970608, + 0.009280986152589321, + 0.007015633396804333, + 0.008401298895478249, + 0.009429769590497017, + 0.007215905003249645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.0046380325220525265 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016076931729912758 + ], + "opposite_logprob_tokens": [ + -4.141077041625977 + ], + "logprob_reward": -0.016076931729912758, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016076931729912758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0302482508122921, + 0.018052145838737488, + 0.037150271236896515, + 0.022328021004796028, + 0.12406476587057114, + 0.044736575335264206, + 0.06918148696422577, + 0.5710514187812805, + 0.04521352797746658, + 0.03797347843647003 + ], + "credit_map": [ + 0.0036714388988912106, + 0.001902701216749847, + 0.004124966915696859, + 0.005800978280603886, + 0.0024450018536299467, + 0.0053543271496891975, + 0.0030168392695486546, + 0.003931999206542969, + 0.0016554901376366615, + 0.0008974549127742648, + 0.002186772646382451, + 0.002101001562550664, + 0.0017281833570450544, + 0.0029691332019865513, + 0.0036632458213716745, + 0.0028508633840829134, + 0.0027088818605989218, + 0.003504664171487093, + 0.0028649023734033108, + 0.0062516601756215096, + 0.0037028652150183916, + 0.006858456879854202, + 0.005361872259527445, + 0.005896969232708216, + 0.0017456042114645243, + 0.0016906338278204203, + 0.004051977768540382, + 0.002299309242516756, + 0.0009348131716251373, + 0.005622244905680418, + 0.0018649576231837273, + 0.00411848071962595, + 0.02004130557179451, + 0.0110732177272439, + 0.031355876475572586, + 0.019926365464925766, + 0.005068751983344555, + 0.009945644065737724, + 0.013447894714772701, + 0.013205708004534245, + 0.006971293594688177, + 0.007307713385671377, + 0.008410172536969185, + 0.004893722012639046, + 0.0026966750156134367, + 0.003943047020584345, + 0.006168061401695013, + 0.004345891531556845, + 0.009591849520802498, + 0.007592883892357349, + 0.006211104802787304, + 0.010997533798217773, + 0.0025536443572491407, + 0.007814627140760422, + 0.004696457646787167, + 0.01972338743507862, + 0.04870787635445595, + 0.06797982007265091, + 0.08122295886278152, + 0.1343078762292862, + 0.03720719739794731, + 0.076856330037117, + 0.06700503826141357, + 0.057764388620853424, + 0.004184721037745476, + 0.009518114849925041, + 0.004986734129488468, + 0.008352466858923435, + 0.007296700030565262, + 0.0039632245898246765, + 0.004231142811477184, + 0.0026804213412106037, + 0.003208903595805168, + 0.004774030763655901, + 0.005639429669827223, + 0.005663183052092791, + 0.004295403137803078, + 0.0048543415032327175, + 0.00566859683021903, + 0.0038695926778018475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.016076931729912758 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029785940423607826 + ], + "opposite_logprob_tokens": [ + -3.5297858715057373 + ], + "logprob_reward": -0.029785940423607826, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029785940423607826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0947476476430893, + 0.06564423441886902, + 0.086261086165905, + 0.07587723433971405, + 0.21689832210540771, + 0.03921141102910042, + 0.17027828097343445, + 0.10850198566913605, + 0.08740635216236115, + 0.05517348647117615 + ], + "credit_map": [ + 0.012142518535256386, + 0.005239015445113182, + 0.011282566003501415, + 0.011322538368403912, + 0.009478597901761532, + 0.01759365387260914, + 0.01285465806722641, + 0.014834101311862469, + 0.004972836468368769, + 0.0030750990845263004, + 0.008135517127811909, + 0.008382970467209816, + 0.005455929320305586, + 0.010706204921007156, + 0.013637276366353035, + 0.01127840206027031, + 0.006743123289197683, + 0.009637628681957722, + 0.00761024747043848, + 0.016079405322670937, + 0.006307270377874374, + 0.014220815151929855, + 0.013359605334699154, + 0.012302988208830357, + 0.005361920688301325, + 0.00657033221796155, + 0.006090475246310234, + 0.008095571771264076, + 0.006258788518607616, + 0.021112145856022835, + 0.007189680822193623, + 0.015198319219052792, + 0.04030301794409752, + 0.023175865411758423, + 0.06303425878286362, + 0.02005770243704319, + 0.01633869856595993, + 0.019799184054136276, + 0.014211016707122326, + 0.01997857168316841, + 0.003925716504454613, + 0.005138780921697617, + 0.005204957444220781, + 0.005174159538000822, + 0.0022892048582434654, + 0.005366907455027103, + 0.004920674487948418, + 0.007191011216491461, + 0.01740824617445469, + 0.0449007973074913, + 0.01214180514216423, + 0.020638225600123405, + 0.01624194160103798, + 0.01720692776143551, + 0.016575941815972328, + 0.025164376944303513, + 0.007752724923193455, + 0.01305884774774313, + 0.02654929831624031, + 0.015412774868309498, + 0.007929563522338867, + 0.012548552826046944, + 0.013114450499415398, + 0.01213577575981617, + 0.009838023222982883, + 0.013854065909981728, + 0.010627026669681072, + 0.021328572183847427, + 0.007869967259466648, + 0.008555841632187366, + 0.009156005457043648, + 0.006176844704896212, + 0.004681132268160582, + 0.010771415196359158, + 0.009384866803884506, + 0.007785920053720474, + 0.005582078360021114, + 0.0049881706945598125, + 0.005216680467128754, + 0.006763221230357885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.029785940423607826 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid single shot begins from a slightly upward angle, the camera steadily moving forward, following the steps of a silver-haired young man as he walks down a narrow, damp back alley. The scene is shrouded in a chilling blue hue. The mottled brick walls on either side are stained with water, and the puddles on the ground reflect a cold light. A single, old-fashioned orange streetlight at the alley's entrance casts a cone of warm light, slicing through the deep blue of the night like a sharp sword, creating a striking visual contrast between warm and cold. The young man's short, distinctly silver hair stands up like icy spikes. The warm light highlights the sharp contours of his hair and profile, while his front is largely immersed in a cold blue shadow. A pair of eagle-like eyes gleam fiercely within the shadows. Clad in a worn black leather motorcycle jacket, dark jeans, and heavy combat boots, he walks with a steady, menacing gait, closing in on his cornered opponent. When he finally stopped, his body leaned forward slightly, invading the other person's personal space, his jawline tense, his lips silently opening and closing, as if he was questioning in a very low but powerful voice. The whole picture was full of dramatic tension and a tense atmosphere that was about to break out.", + "seed": 10813, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_13/index_2049_seed_10813_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04305107044056058, + "vlm_logprob": -0.04305107044056058, + "vlm_yes_no_margin": 3.924999976158142, + "vlm_token_credit": -0.04305107044056058, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15020620822906494 + ], + "opposite_logprob_tokens": [ + -2.0252060890197754 + ], + "logprob_reward": -0.15020620822906494, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.15020620822906494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02798926830291748, + 0.13493652641773224, + 0.016488201916217804, + 0.07794373482465744, + 0.05008964240550995, + 0.03394748643040657, + 0.027321824803948402, + 0.02959511987864971, + 0.04021125286817551, + 0.5614769458770752 + ], + "credit_map": [ + 0.0015495014376938343, + 0.0033858485985547304, + 0.0026067933067679405, + 0.003699600463733077, + 0.004403979983180761, + 0.004264335613697767, + 0.003424963681027293, + 0.004654244054108858, + 0.0025554331950843334, + 0.009950816631317139, + 0.0035340830218046904, + 0.00788898766040802, + 0.002451853593811393, + 0.05387139320373535, + 0.003893193555995822, + 0.05079076439142227, + 0.0014383434318006039, + 0.0022128799464553595, + 0.00234615127556026, + 0.002503484021872282, + 0.002103611594066024, + 0.002134472131729126, + 0.0019669784232974052, + 0.0017822808586061, + 0.0038408427499234676, + 0.008416143245995045, + 0.003882575547322631, + 0.0038701400626450777, + 0.004004150163382292, + 0.024198384955525398, + 0.003650924190878868, + 0.02608056738972664, + 0.0031402376480400562, + 0.005038893315941095, + 0.004862730856984854, + 0.004183700308203697, + 0.008120875805616379, + 0.009327255189418793, + 0.0032418030314147472, + 0.012174148112535477, + 0.004607811104506254, + 0.00417995173484087, + 0.008496802300214767, + 0.0033551687374711037, + 0.002315720310434699, + 0.0052507962100207806, + 0.00314173917286098, + 0.002599496627226472, + 0.0046439869329333305, + 0.002502326387912035, + 0.003301505232229829, + 0.002716836053878069, + 0.003179785329848528, + 0.005047236569225788, + 0.003061146941035986, + 0.0028690025210380554, + 0.0023065863642841578, + 0.0018347942968830466, + 0.004005902446806431, + 0.002444877754896879, + 0.005390775389969349, + 0.005943420343101025, + 0.003451768308877945, + 0.004216996021568775, + 0.004686387721449137, + 0.0021272418089210987, + 0.0034346794709563255, + 0.0029322404880076647, + 0.007431614212691784, + 0.008318865671753883, + 0.004528941120952368, + 0.00675128074362874, + 0.03990069031715393, + 0.014462911523878574, + 0.032566700130701065, + 0.016413068398833275, + 0.03447942063212395, + 0.06760289520025253, + 0.22009456157684326, + 0.13595670461654663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.15020620822906494 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004111409652978182 + ], + "opposite_logprob_tokens": [ + -5.504111289978027 + ], + "logprob_reward": -0.004111409652978182, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004111409652978182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0609750896692276, + 0.046341728419065475, + 0.024160850793123245, + 0.02815646305680275, + 0.016596617177128792, + 0.015041252598166466, + 0.04935680329799652, + 0.07324901968240738, + 0.038482267409563065, + 0.6476399898529053 + ], + "credit_map": [ + 0.0036537921987473965, + 0.0069962176494300365, + 0.0077123092487454414, + 0.011998600326478481, + 0.004339336883276701, + 0.013392714783549309, + 0.00488445395603776, + 0.007997655309736729, + 0.001576894079335034, + 0.0035671780351549387, + 0.0019457325106486678, + 0.003145248396322131, + 0.001318486756645143, + 0.017033692449331284, + 0.0016709617339074612, + 0.016083532944321632, + 0.0027983617037534714, + 0.00273312721401453, + 0.0023721957113593817, + 0.00394913274794817, + 0.002431368688121438, + 0.004738936200737953, + 0.002543219132348895, + 0.0025945075321942568, + 0.0021603971254080534, + 0.005016649607568979, + 0.004770857281982899, + 0.002904082415625453, + 0.0025001796893775463, + 0.005193020217120647, + 0.002563993213698268, + 0.003047277918085456, + 0.0020048965234309435, + 0.0024406523443758488, + 0.0023550817277282476, + 0.0017417407361790538, + 0.002135274466127157, + 0.0022243093699216843, + 0.0009941586758941412, + 0.0027005027513951063, + 0.0021724896505475044, + 0.0016977800987660885, + 0.002958958502858877, + 0.0019151188898831606, + 0.0010946354595944285, + 0.0023612151853740215, + 0.0011384142562747002, + 0.0017026402056217194, + 0.011100043542683125, + 0.006768265273422003, + 0.008443803526461124, + 0.0035333437845110893, + 0.004235307220369577, + 0.006775012705475092, + 0.0052429563365876675, + 0.003258064389228821, + 0.009677867405116558, + 0.005100650247186422, + 0.011288672685623169, + 0.005562251433730125, + 0.007236518897116184, + 0.009985417127609253, + 0.009571235626935959, + 0.0148263955488801, + 0.002430737018585205, + 0.001513029565103352, + 0.003145379712805152, + 0.0016128862043842673, + 0.0044240099377930164, + 0.01163940317928791, + 0.004207088612020016, + 0.009509727358818054, + 0.04623674973845482, + 0.016622550785541534, + 0.03715936839580536, + 0.01885540969669819, + 0.037738166749477386, + 0.07617061585187912, + 0.25520890951156616, + 0.15964818000793457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.004111409652978182 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810246288776398 + ], + "opposite_logprob_tokens": [ + -3.288102388381958 + ], + "logprob_reward": -0.03810246288776398, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810246288776398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023276645690202713, + 0.057744305580854416, + 0.012151329778134823, + 0.019544867798686028, + 0.012308046221733093, + 0.0117065804079175, + 0.10388021171092987, + 0.0945877656340599, + 0.03506692498922348, + 0.6297333836555481 + ], + "credit_map": [ + 0.0017839253414422274, + 0.0030642563942819834, + 0.0028784044552594423, + 0.0035567237064242363, + 0.0031396248377859592, + 0.004497244488447905, + 0.0022750282660126686, + 0.0020814365707337856, + 0.0019132621819153428, + 0.004282590467482805, + 0.0020177438855171204, + 0.0030430848710238934, + 0.0018175352597609162, + 0.022154681384563446, + 0.0015486531192436814, + 0.020966749638319016, + 0.0010273887310177088, + 0.0014702508924528956, + 0.0015126056969165802, + 0.0015733294421806931, + 0.0010769369546324015, + 0.0026704566553235054, + 0.0013856899458914995, + 0.0014346709940582514, + 0.0024436030071228743, + 0.0022492860443890095, + 0.002370898611843586, + 0.0017967966850847006, + 0.0016787122003734112, + 0.004332748241722584, + 0.0014838511124253273, + 0.0031889681704342365, + 0.0011918841628357768, + 0.0012785721337422729, + 0.0019084359519183636, + 0.000851105956826359, + 0.0010573315666988492, + 0.0027552323881536722, + 0.0008358980994671583, + 0.002429584739729762, + 0.0012907901545986533, + 0.0012634668964892626, + 0.0020194256212562323, + 0.0014184991596266627, + 0.0012808449100703, + 0.0018423731671646237, + 0.0011832903837785125, + 0.0014078886015340686, + 0.013290710747241974, + 0.023579221218824387, + 0.01417966466397047, + 0.009153759106993675, + 0.009811082854866982, + 0.017733270302414894, + 0.009072929620742798, + 0.007059562485665083, + 0.012858695350587368, + 0.005790500435978174, + 0.017805427312850952, + 0.005477132275700569, + 0.006566003896296024, + 0.014643114991486073, + 0.011980587616562843, + 0.019466295838356018, + 0.005293597001582384, + 0.0018688665004447103, + 0.004158904775977135, + 0.0018590586259961128, + 0.003437361679971218, + 0.00936963502317667, + 0.0057185497134923935, + 0.00336094805970788, + 0.04440529644489288, + 0.016357455402612686, + 0.03694714233279228, + 0.01986016519367695, + 0.03515545651316643, + 0.07224913686513901, + 0.24852034449577332, + 0.15623833239078522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.03810246288776398 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647141046822071 + ], + "opposite_logprob_tokens": [ + -4.7586469650268555 + ], + "logprob_reward": -0.008647141046822071, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647141046822071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09265334159135818, + 0.13968615233898163, + 0.09048640727996826, + 0.10774735361337662, + 0.10728630423545837, + 0.039298247545957565, + 0.11278431862592697, + 0.06956922262907028, + 0.1281297653913498, + 0.11235882341861725 + ], + "credit_map": [ + 0.007239229511469603, + 0.011482455767691135, + 0.009637676179409027, + 0.009516211226582527, + 0.014724855311214924, + 0.01292145624756813, + 0.01249171607196331, + 0.014639741741120815, + 0.0028464351780712605, + 0.010519210249185562, + 0.00421333871781826, + 0.008715406060218811, + 0.0035020478535443544, + 0.05447305366396904, + 0.004056980833411217, + 0.05135967209935188, + 0.005656936671584845, + 0.010685297660529613, + 0.013679658994078636, + 0.012999745085835457, + 0.01535931695252657, + 0.007271424867212772, + 0.015546823851764202, + 0.009287206456065178, + 0.007300439290702343, + 0.014006748795509338, + 0.007765571586787701, + 0.009485945105552673, + 0.006799214519560337, + 0.025070982053875923, + 0.01002172939479351, + 0.027296723797917366, + 0.008253660053014755, + 0.010979714803397655, + 0.011659427545964718, + 0.00966190081089735, + 0.025676174089312553, + 0.014483979903161526, + 0.009970762766897678, + 0.016600683331489563, + 0.005193226505070925, + 0.0055603389628231525, + 0.0056473445147275925, + 0.005041828844696283, + 0.004585928749293089, + 0.004208624828606844, + 0.005142697133123875, + 0.0039182621985673904, + 0.015136322006583214, + 0.022577019408345222, + 0.01689550280570984, + 0.00976091530174017, + 0.009728686884045601, + 0.019892597571015358, + 0.010309959761798382, + 0.008483308367431164, + 0.00904194638133049, + 0.004026463720947504, + 0.00837786216288805, + 0.005889580585062504, + 0.0066534713841974735, + 0.010197085328400135, + 0.008968709968030453, + 0.016414104029536247, + 0.019550170749425888, + 0.004765864461660385, + 0.01510117668658495, + 0.005054184701293707, + 0.010262195020914078, + 0.01594960317015648, + 0.023139936849474907, + 0.034306637942790985, + 0.008957796730101109, + 0.0031081167981028557, + 0.0063509694300591946, + 0.004749857820570469, + 0.006994821131229401, + 0.012830061838030815, + 0.04192616418004036, + 0.027441037818789482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -0.008647141046822071 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with distinctly silver hair styled like icy spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, damp back alley with mottled brick walls and puddles reflecting cold light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single old-fashioned orange streetlight at the alley entrance casting a cone of warm light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a worn black leather motorcycle jacket, dark jeans, and heavy combat boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly upward angle and move steadily forward following the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188130386173725 + ], + "opposite_logprob_tokens": [ + -4.264188289642334 + ], + "logprob_reward": -0.014188130386173725, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188130386173725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09071093797683716, + 0.2620205879211426, + 0.03162512555718422, + 0.029481222853064537, + 0.07533431798219681, + 0.054064925760030746, + 0.22724288702011108, + 0.06734933704137802, + 0.08057127147912979, + 0.0815994143486023 + ], + "credit_map": [ + 0.004761086776852608, + 0.00965526420623064, + 0.009876706637442112, + 0.010712412185966969, + 0.012842931784689426, + 0.020100707188248634, + 0.009478995576500893, + 0.013282831758260727, + 0.005114527884870768, + 0.019357141107320786, + 0.006909138057380915, + 0.013753844425082207, + 0.005479095969349146, + 0.10612919181585312, + 0.006292055360972881, + 0.09898556023836136, + 0.0023231033701449633, + 0.004097490105777979, + 0.00450562871992588, + 0.004126124549657106, + 0.003574354574084282, + 0.004120056051760912, + 0.005864694714546204, + 0.0030136713758111, + 0.002463687676936388, + 0.0038157098460942507, + 0.0035797390155494213, + 0.003037197282537818, + 0.003596108639612794, + 0.005161052569746971, + 0.00288603943772614, + 0.004941687453538179, + 0.005055878311395645, + 0.008292311802506447, + 0.008389480412006378, + 0.006858986336737871, + 0.014146234840154648, + 0.014338827691972256, + 0.006983149331063032, + 0.011269443668425083, + 0.004522848874330521, + 0.0064169238321483135, + 0.008036152459681034, + 0.009795131161808968, + 0.005007264204323292, + 0.010743111371994019, + 0.005264733452349901, + 0.004278757143765688, + 0.050464730709791183, + 0.02725222520530224, + 0.033963341265916824, + 0.018393956124782562, + 0.014077133499085903, + 0.042111143469810486, + 0.02296597883105278, + 0.01801435835659504, + 0.008143380284309387, + 0.0033065301831811666, + 0.01255861483514309, + 0.004365252330899239, + 0.005480872467160225, + 0.012104175984859467, + 0.008227009326219559, + 0.013163503259420395, + 0.007338507100939751, + 0.003749958472326398, + 0.008427005261182785, + 0.0023720390163362026, + 0.007760444190353155, + 0.018797647207975388, + 0.009417255409061909, + 0.022708404809236526, + 0.0073502762243151665, + 0.003663106821477413, + 0.007548180408775806, + 0.005482580047100782, + 0.008054547011852264, + 0.010528516955673695, + 0.02511151134967804, + 0.013860691338777542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.014188130386173725 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_14/index_tensor([8056, 8056])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_14/index_tensor([8056, 8056])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b584b90c13790fc6dc254f7bdb3da640f13f4e41 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_14/index_tensor([8056, 8056])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Extreme close-up static shots, rendered with hyper-real digital clarity, capture the scene with scientific precision. Within a simple, transparent drinking glass, set against a softly blurred neutral background, lies a conical mound of ultra-fine, vibrant, electric-green barley powder, its texture so fine that individual grains are visible. Bright, even studio lighting illuminates the glass, casting no harsh shadows and making the colors pop. A perfectly spherical drop of crystal-clear water descends with deliberate slowness, its surface reflecting the light. Upon contact with the highest peak of the powder, the fine particles instantly seethe in a chain reaction. As the vibrant green transforms into a rich, dark, almost emerald liquid, the mound visibly collapses inward, and the liquid immediately begins to swirl with methodical precision, creating an elegant, captivating vortex that pulls the remaining dry powder into the deepening green swirl, creating a focused and hypnotic atmosphere of satisfying transformation.", + "seed": 24140, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_14/index_8056_seed_24140_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0334192369133234, + "vlm_logprob": -0.0334192369133234, + "vlm_yes_no_margin": 3.474999952316284, + "vlm_token_credit": -0.0334192369133234, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030450478196144104 + ], + "opposite_logprob_tokens": [ + -3.6554505825042725 + ], + "logprob_reward": -0.030450478196144104, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.030450478196144104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31237635016441345, + 0.14409461617469788, + 0.042826537042856216, + 0.1405724436044693, + 0.03347174450755119, + 0.07240575551986694, + 0.03054691106081009, + 0.077431820333004, + 0.08788272738456726, + 0.05839115008711815 + ], + "credit_map": [ + 0.03217725455760956, + 0.044995155185461044, + 0.03301636129617691, + 0.029067426919937134, + 0.04193628206849098, + 0.04302145540714264, + 0.049142975360155106, + 0.03901940584182739, + 0.015805566683411598, + 0.01716163381934166, + 0.015504816547036171, + 0.014925297349691391, + 0.02420242875814438, + 0.02425503544509411, + 0.01959400810301304, + 0.012645806185901165, + 0.006154863629490137, + 0.004854137077927589, + 0.00444105314090848, + 0.004796035122126341, + 0.004756406880915165, + 0.006934431381523609, + 0.006813148036599159, + 0.004076455254107714, + 0.02575664035975933, + 0.015709342435002327, + 0.01598808914422989, + 0.01271749846637249, + 0.011143842712044716, + 0.022742340341210365, + 0.020667122676968575, + 0.015847565606236458, + 0.003787316381931305, + 0.0041727339848876, + 0.0032392467837780714, + 0.00442300271242857, + 0.005260915495455265, + 0.0044176010414958, + 0.0054773325100541115, + 0.002693592570722103, + 0.006829241290688515, + 0.014659537002444267, + 0.007609246764332056, + 0.007219661958515644, + 0.01093901228159666, + 0.010840442031621933, + 0.007148014847189188, + 0.007160590495914221, + 0.0038007062394171953, + 0.00470036780461669, + 0.0030407134909182787, + 0.004496508277952671, + 0.002939454512670636, + 0.004185271915048361, + 0.0033764794934540987, + 0.004007407929748297, + 0.009871939197182655, + 0.0064932601526379585, + 0.008507666178047657, + 0.018185952678322792, + 0.0054778787307441235, + 0.013492578640580177, + 0.008233251981437206, + 0.007169288117438555, + 0.008837921544909477, + 0.012789610773324966, + 0.008833467960357666, + 0.012320143170654774, + 0.011378911323845387, + 0.012606393545866013, + 0.01029355451464653, + 0.010822721756994724, + 0.005941885989159346, + 0.009013916365802288, + 0.004068921785801649, + 0.010473376139998436, + 0.008396896533668041, + 0.007689739111810923, + 0.007392001338303089, + 0.005414409097284079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.030450478196144104 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029813364148139954 + ], + "opposite_logprob_tokens": [ + -3.529813289642334 + ], + "logprob_reward": -0.029813364148139954, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029813364148139954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28605905175209045, + 0.07661914080381393, + 0.0795537605881691, + 0.11812768131494522, + 0.09919389337301254, + 0.07194291800260544, + 0.035463735461235046, + 0.06638680398464203, + 0.09761718660593033, + 0.0690358504652977 + ], + "credit_map": [ + 0.03172220289707184, + 0.03912805765867233, + 0.029280461370944977, + 0.0303356796503067, + 0.03415665403008461, + 0.03847634419798851, + 0.046937692910432816, + 0.03602194786071777, + 0.009999281726777554, + 0.007962710224092007, + 0.010841618292033672, + 0.009190034121274948, + 0.008516468107700348, + 0.011586353182792664, + 0.01139114797115326, + 0.007131521590054035, + 0.008799038827419281, + 0.009229186922311783, + 0.007667528465390205, + 0.011235346086323261, + 0.008593124337494373, + 0.013295135460793972, + 0.013438256457448006, + 0.007296139374375343, + 0.021399416029453278, + 0.013061609119176865, + 0.013812006451189518, + 0.010707502253353596, + 0.011475923471152782, + 0.018590116873383522, + 0.016199173405766487, + 0.012881926260888577, + 0.01218529511243105, + 0.019675057381391525, + 0.004604267422109842, + 0.007482072804123163, + 0.02389460802078247, + 0.01583297736942768, + 0.010109464637935162, + 0.005410144571214914, + 0.008257722482085228, + 0.012035905383527279, + 0.0063632214441895485, + 0.009273460134863853, + 0.012716925702989101, + 0.0107109434902668, + 0.007227512076497078, + 0.005357217509299517, + 0.004581209272146225, + 0.007532026153057814, + 0.0026170287746936083, + 0.0037098380271345377, + 0.005990983918309212, + 0.0050620222464203835, + 0.0028320334386080503, + 0.0031385940965265036, + 0.007959384471178055, + 0.012589597143232822, + 0.005298312287777662, + 0.009777985513210297, + 0.008344681933522224, + 0.008879221975803375, + 0.006848668679594994, + 0.006688945461064577, + 0.010153167881071568, + 0.0127181988209486, + 0.008747848682105541, + 0.014999516308307648, + 0.014438243582844734, + 0.014131355099380016, + 0.012296043336391449, + 0.010132803581655025, + 0.009115046821534634, + 0.007443600334227085, + 0.006571018137037754, + 0.012923761270940304, + 0.006799813825637102, + 0.008602486923336983, + 0.009241105988621712, + 0.00833901297301054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.029813364148139954 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033693719655275345 + ], + "opposite_logprob_tokens": [ + -3.408693790435791 + ], + "logprob_reward": -0.033693719655275345, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033693719655275345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2089824229478836, + 0.05343743413686752, + 0.05749848857522011, + 0.06783580780029297, + 0.1399371176958084, + 0.12926971912384033, + 0.12865948677062988, + 0.052320435643196106, + 0.06531838327646255, + 0.09674070030450821 + ], + "credit_map": [ + 0.03728852793574333, + 0.019471675157546997, + 0.02213960513472557, + 0.02267736755311489, + 0.019812552258372307, + 0.034737132489681244, + 0.029757946729660034, + 0.023097626864910126, + 0.009360560216009617, + 0.004398559685796499, + 0.005242675077170134, + 0.007347811944782734, + 0.0061009908095002174, + 0.008608612231910229, + 0.007747376337647438, + 0.0046308464370667934, + 0.00651586102321744, + 0.006155761890113354, + 0.005713807884603739, + 0.00801872368901968, + 0.006708559114485979, + 0.007930985651910305, + 0.01168924942612648, + 0.004765539895743132, + 0.010054666548967361, + 0.005721602588891983, + 0.005829937756061554, + 0.00813691969960928, + 0.010336222127079964, + 0.00801914930343628, + 0.014071253128349781, + 0.00566606130450964, + 0.017512347549200058, + 0.021569639444351196, + 0.006258232519030571, + 0.013782455585896969, + 0.029745617881417274, + 0.021342216059565544, + 0.019523927941918373, + 0.010202686302363873, + 0.01186493132263422, + 0.0202261283993721, + 0.011856469325721264, + 0.016936739906668663, + 0.022197416052222252, + 0.018905648961663246, + 0.016099268570542336, + 0.011183115653693676, + 0.015000690706074238, + 0.01929485611617565, + 0.015238826163113117, + 0.017178775742650032, + 0.012056941166520119, + 0.016944974660873413, + 0.01593155413866043, + 0.01701286807656288, + 0.007982118986546993, + 0.009186293929815292, + 0.005027395207434893, + 0.006297779735177755, + 0.006677365396171808, + 0.007547029759734869, + 0.005061646457761526, + 0.004540806636214256, + 0.0073751225136220455, + 0.014035462401807308, + 0.004203478340059519, + 0.005541374906897545, + 0.015184211544692516, + 0.010426788590848446, + 0.003752612741664052, + 0.004799331538379192, + 0.015338399447500706, + 0.013402908109128475, + 0.00965309888124466, + 0.013243253342807293, + 0.011243082582950592, + 0.011619964614510536, + 0.012392140924930573, + 0.009847857989370823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.033693719655275345 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495565012097359 + ], + "opposite_logprob_tokens": [ + -2.9299557209014893 + ], + "logprob_reward": -0.05495565012097359, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495565012097359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10452442616224289, + 0.0986066535115242, + 0.060294318944215775, + 0.16419802606105804, + 0.08301649242639542, + 0.10662951320409775, + 0.07141943275928497, + 0.11617832630872726, + 0.09615089744329453, + 0.0989818125963211 + ], + "credit_map": [ + 0.015393553301692009, + 0.011911673471331596, + 0.009588520042598248, + 0.01572038047015667, + 0.01121546234935522, + 0.015132579021155834, + 0.0165473110973835, + 0.00901496410369873, + 0.011425638571381569, + 0.010480747558176517, + 0.007940292358398438, + 0.011752489022910595, + 0.016181282699108124, + 0.01369934156537056, + 0.018210405483841896, + 0.008916483260691166, + 0.006757160648703575, + 0.006189278792589903, + 0.005614570342004299, + 0.009417073801159859, + 0.007361068390309811, + 0.007996530272066593, + 0.011987632140517235, + 0.004971015267074108, + 0.0335426852107048, + 0.01543271355330944, + 0.016503525897860527, + 0.012719080783426762, + 0.02095651812851429, + 0.029014023020863533, + 0.018594752997159958, + 0.017434757202863693, + 0.009510792791843414, + 0.010580410249531269, + 0.006458892486989498, + 0.010053561069071293, + 0.01456383429467678, + 0.011382007040083408, + 0.013185055926442146, + 0.0072819534689188, + 0.011230605654418468, + 0.016073424369096756, + 0.01270168274641037, + 0.012951087206602097, + 0.010532203130424023, + 0.016143200919032097, + 0.011737599037587643, + 0.015259725973010063, + 0.009158301167190075, + 0.007238882593810558, + 0.00935159157961607, + 0.008748314343392849, + 0.005911458749324083, + 0.010960263200104237, + 0.010072254575788975, + 0.009978383779525757, + 0.015576804056763649, + 0.01288137398660183, + 0.010868164710700512, + 0.021283496171236038, + 0.012489475309848785, + 0.019256044179201126, + 0.01222960278391838, + 0.011593395844101906, + 0.011208264157176018, + 0.01431406568735838, + 0.008121861144900322, + 0.01328540313988924, + 0.012474912218749523, + 0.013930078595876694, + 0.011008433066308498, + 0.01180789340287447, + 0.013320441357791424, + 0.01031351275742054, + 0.005395997781306505, + 0.017720084637403488, + 0.013641662895679474, + 0.014202232472598553, + 0.01544246543198824, + 0.008945439010858536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.05495565012097359 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182972446084023 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.018182972446084023, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018182972446084023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22652557492256165, + 0.1278490573167801, + 0.13473102450370789, + 0.08585658669471741, + 0.1938602775335312, + 0.03755446523427963, + 0.023418063297867775, + 0.0395788811147213, + 0.04246552288532257, + 0.08816064149141312 + ], + "credit_map": [ + 0.022181428968906403, + 0.03083912841975689, + 0.024350352585315704, + 0.02102985605597496, + 0.03011257015168667, + 0.03116133250296116, + 0.036867816001176834, + 0.02998306229710579, + 0.01761457324028015, + 0.011601588688790798, + 0.008600279688835144, + 0.017580823972821236, + 0.017125533893704414, + 0.02231646329164505, + 0.021929902955889702, + 0.01107986643910408, + 0.023694701492786407, + 0.012807830236852169, + 0.01226236205548048, + 0.01626238413155079, + 0.013780640438199043, + 0.023218052461743355, + 0.018375512212514877, + 0.0143295181915164, + 0.01595435105264187, + 0.007766035385429859, + 0.008419085294008255, + 0.008533204905688763, + 0.011349872685968876, + 0.01402029674500227, + 0.012685018591582775, + 0.007128710392862558, + 0.02403760515153408, + 0.02906624786555767, + 0.013472325168550014, + 0.021159308031201363, + 0.03888281062245369, + 0.028702087700366974, + 0.022720064967870712, + 0.01581980660557747, + 0.004448858555406332, + 0.0037584546953439713, + 0.004934830591082573, + 0.0043678488582372665, + 0.0037121728528290987, + 0.005843558814376593, + 0.006120349746197462, + 0.004368382506072521, + 0.0025867237709462643, + 0.0027138760779052973, + 0.0027055933605879545, + 0.003809729591012001, + 0.0022865578066557646, + 0.0035998031962662935, + 0.0029385783709585667, + 0.0027771966997534037, + 0.004572962410748005, + 0.006283846218138933, + 0.0034035551361739635, + 0.007621191442012787, + 0.0036935643292963505, + 0.005686758551746607, + 0.004762421362102032, + 0.003554577473551035, + 0.00580616993829608, + 0.006536603439599276, + 0.003061541123315692, + 0.004797479137778282, + 0.006598141975700855, + 0.006558939348906279, + 0.0038528458680957556, + 0.005253791809082031, + 0.012216375209391117, + 0.01174488477408886, + 0.007112511433660984, + 0.017277605831623077, + 0.009852811694145203, + 0.009882079437375069, + 0.011603981256484985, + 0.008470377884805202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.018182972446084023 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Extreme close-up static shots, rendered with hyper-real digital clarity, capture the scene with scientific precision. Within a simple, transparent drinking glass, set against a softly blurred neutral background, lies a conical mound of ultra-fine, vibrant, electric-green barley powder, its texture so fine that individual grains are visible. Bright, even studio lighting illuminates the glass, casting no harsh shadows and making the colors pop. A perfectly spherical drop of crystal-clear water descends with deliberate slowness, its surface reflecting the light. Upon contact with the highest peak of the powder, the fine particles instantly seethe in a chain reaction. As the vibrant green transforms into a rich, dark, almost emerald liquid, the mound visibly collapses inward, and the liquid immediately begins to swirl with methodical precision, creating an elegant, captivating vortex that pulls the remaining dry powder into the deepening green swirl, creating a focused and hypnotic atmosphere of satisfying transformation.", + "seed": 24141, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_14/index_8056_seed_24141_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023145885672420263, + "vlm_logprob": -0.023145885672420263, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.023145885672420263, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03578568249940872 + ], + "opposite_logprob_tokens": [ + -3.535785675048828 + ], + "logprob_reward": -0.03578568249940872, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03578568249940872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07271832972764969, + 0.029664471745491028, + 0.2096215933561325, + 0.16621620953083038, + 0.1290218085050583, + 0.06313040107488632, + 0.08592104911804199, + 0.07457901537418365, + 0.12471150606870651, + 0.04441562667489052 + ], + "credit_map": [ + 0.0056755635887384415, + 0.007375375833362341, + 0.010807473212480545, + 0.012824458070099354, + 0.005231727845966816, + 0.004428026266396046, + 0.01676834560930729, + 0.009607361629605293, + 0.0018210032721981406, + 0.004223086405545473, + 0.005682678893208504, + 0.004770743194967508, + 0.0021310418378561735, + 0.0012967967195436358, + 0.006342716980725527, + 0.0033964046742767096, + 0.010999868623912334, + 0.032063573598861694, + 0.039055705070495605, + 0.03145129233598709, + 0.01506013236939907, + 0.0090912114828825, + 0.05125945433974266, + 0.020640337839722633, + 0.01203928329050541, + 0.016308574005961418, + 0.03894167020916939, + 0.043992139399051666, + 0.009840935468673706, + 0.006188444793224335, + 0.026982299983501434, + 0.011922848410904408, + 0.011957509443163872, + 0.014519806019961834, + 0.016956955194473267, + 0.025711189955472946, + 0.010456603951752186, + 0.008281040005385876, + 0.023476112633943558, + 0.017662592232227325, + 0.012932541780173779, + 0.0055435458198189735, + 0.011895240284502506, + 0.005680114962160587, + 0.0056845154613256454, + 0.010266199707984924, + 0.004551562946289778, + 0.006576681509613991, + 0.01070216204971075, + 0.011799723841249943, + 0.00901854783296585, + 0.00910627469420433, + 0.014905760996043682, + 0.012002886272966862, + 0.00613743718713522, + 0.012248256243765354, + 0.004604383837431669, + 0.024103349074721336, + 0.022046497091650963, + 0.0052299522794783115, + 0.005789395421743393, + 0.005461337044835091, + 0.003081169445067644, + 0.00426293071359396, + 0.012347061187028885, + 0.012126525864005089, + 0.009694065898656845, + 0.016222819685935974, + 0.015551074407994747, + 0.033671967685222626, + 0.014502222649753094, + 0.010595763102173805, + 0.0046258182264864445, + 0.004700665362179279, + 0.0057615856640040874, + 0.007318370044231415, + 0.008809330873191357, + 0.0037182248197495937, + 0.004952196963131428, + 0.0045294370502233505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.03578568249940872 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008655531331896782 + ], + "opposite_logprob_tokens": [ + -4.758655548095703 + ], + "logprob_reward": -0.008655531331896782, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008655531331896782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12031702697277069, + 0.054744791239500046, + 0.040907278656959534, + 0.17542584240436554, + 0.044200699776411057, + 0.11617526412010193, + 0.14502030611038208, + 0.07626255601644516, + 0.0678098127245903, + 0.15913644433021545 + ], + "credit_map": [ + 0.008415249176323414, + 0.014397894963622093, + 0.021496670320630074, + 0.01897599548101425, + 0.007290184497833252, + 0.006763446144759655, + 0.02914201095700264, + 0.013835576362907887, + 0.002902374370023608, + 0.006566180381923914, + 0.011519496329128742, + 0.008083511143922806, + 0.00337293092161417, + 0.002342707011848688, + 0.0139186205342412, + 0.006038968451321125, + 0.0024205856025218964, + 0.004718118812888861, + 0.00794326700270176, + 0.005451631732285023, + 0.0026301778852939606, + 0.00211706617847085, + 0.0087520070374012, + 0.006874425336718559, + 0.018863189965486526, + 0.04283076152205467, + 0.022860577329993248, + 0.019473373889923096, + 0.029122576117515564, + 0.012385610491037369, + 0.018149619922041893, + 0.011740123853087425, + 0.004896909464150667, + 0.005646450910717249, + 0.005221143364906311, + 0.007514222525060177, + 0.004376825410872698, + 0.0032094046473503113, + 0.006732385605573654, + 0.006603360176086426, + 0.019234884530305862, + 0.010487452149391174, + 0.021242480725049973, + 0.010367002338171005, + 0.016978967934846878, + 0.013195161707699299, + 0.010000012814998627, + 0.014669293537735939, + 0.033236488699913025, + 0.0139799565076828, + 0.02247057668864727, + 0.009649718180298805, + 0.02762012369930744, + 0.015321971848607063, + 0.011715074069797993, + 0.011026395484805107, + 0.006067910697311163, + 0.01275515928864479, + 0.01699097454547882, + 0.005376369226723909, + 0.007953854277729988, + 0.00862615555524826, + 0.006489443127065897, + 0.012002681382000446, + 0.005689308047294617, + 0.0075189839117228985, + 0.0061044939793646336, + 0.01079457439482212, + 0.008062043227255344, + 0.011886095628142357, + 0.012569648213684559, + 0.005184662528336048, + 0.014134409837424755, + 0.020388798788189888, + 0.021856654435396194, + 0.02470431849360466, + 0.03926650807261467, + 0.012792866677045822, + 0.014006289653480053, + 0.011986603029072285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 6 + ], + "selected_reward": -0.008655531331896782 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058802917599678 + ], + "opposite_logprob_tokens": [ + -3.895587921142578 + ], + "logprob_reward": -0.02058802917599678, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058802917599678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09434144198894501, + 0.027609869837760925, + 0.13226744532585144, + 0.03242005780339241, + 0.09524468332529068, + 0.1148173063993454, + 0.10631385445594788, + 0.08229850232601166, + 0.26811307668685913, + 0.046573713421821594 + ], + "credit_map": [ + 0.005348552949726582, + 0.011368692852556705, + 0.01857135444879532, + 0.01442903745919466, + 0.0059956437908113, + 0.004215520806610584, + 0.024196626618504524, + 0.010216021910309792, + 0.001794900861568749, + 0.00350887980312109, + 0.005239791236817837, + 0.004231917206197977, + 0.002014016266912222, + 0.0017189764184877276, + 0.006378630641847849, + 0.0027227618265897036, + 0.0078727500513196, + 0.02233160473406315, + 0.023380115628242493, + 0.020236847922205925, + 0.010151291266083717, + 0.005731961224228144, + 0.029745256528258324, + 0.0128176249563694, + 0.0038687768392264843, + 0.005462649744004011, + 0.004792968742549419, + 0.004281376488506794, + 0.003757807658985257, + 0.0020864109974354506, + 0.004227290395647287, + 0.003942778799682856, + 0.00654126750305295, + 0.014526627957820892, + 0.011865748092532158, + 0.02330910973250866, + 0.008018034510314465, + 0.005202579777687788, + 0.014160986058413982, + 0.011620338074862957, + 0.024989338591694832, + 0.01032325066626072, + 0.02034289948642254, + 0.009991566650569439, + 0.009425435215234756, + 0.015910957008600235, + 0.011724933050572872, + 0.01210893876850605, + 0.010181976482272148, + 0.015154294669628143, + 0.014461208134889603, + 0.010990221053361893, + 0.018715864047408104, + 0.014735055156052113, + 0.008750458247959614, + 0.013324778527021408, + 0.00819524098187685, + 0.012356036342680454, + 0.01667044498026371, + 0.01158113218843937, + 0.010067509487271309, + 0.0075888121500611305, + 0.007108258083462715, + 0.008731071837246418, + 0.019671952351927757, + 0.023521574214100838, + 0.015525455586612225, + 0.021697869524359703, + 0.03493349999189377, + 0.08351746201515198, + 0.03242715448141098, + 0.03681812062859535, + 0.004678680561482906, + 0.004555277526378632, + 0.005908828694373369, + 0.00909214373677969, + 0.011097640730440617, + 0.0035068339202553034, + 0.004570759367197752, + 0.0031635507475584745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.02058802917599678 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630960550159216 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630960550159216, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630960550159216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03366253897547722, + 0.022056205198168755, + 0.09036493301391602, + 0.05702701956033707, + 0.08133833855390549, + 0.06459664553403854, + 0.25602737069129944, + 0.061899758875370026, + 0.29254642128944397, + 0.04048074781894684 + ], + "credit_map": [ + 0.0019031575648114085, + 0.004012477584183216, + 0.006623690016567707, + 0.005428537726402283, + 0.002056033816188574, + 0.0018506668275222182, + 0.007441217545419931, + 0.004346759524196386, + 0.001585861318744719, + 0.003159439191222191, + 0.003553547663614154, + 0.0033371185418218374, + 0.0015375319635495543, + 0.0011474270140752196, + 0.004400000441819429, + 0.0033352815080434084, + 0.005171117838472128, + 0.013972954824566841, + 0.016394581645727158, + 0.013661118224263191, + 0.006829315796494484, + 0.004066589288413525, + 0.021693332120776176, + 0.00857592560350895, + 0.006341998931020498, + 0.009706323966383934, + 0.00943631213158369, + 0.01179448887705803, + 0.00569793488830328, + 0.004199398681521416, + 0.006290460471063852, + 0.003560105338692665, + 0.006512431427836418, + 0.010936236009001732, + 0.009694594889879227, + 0.020685261115431786, + 0.006906717550009489, + 0.004545200616121292, + 0.011525913141667843, + 0.010531987063586712, + 0.013779489323496819, + 0.005918918177485466, + 0.01225923839956522, + 0.005297730211168528, + 0.00566039327532053, + 0.009817936457693577, + 0.0049424730241298676, + 0.0069204713217914104, + 0.047660306096076965, + 0.031235000118613243, + 0.03646673262119293, + 0.014596400782465935, + 0.04594830423593521, + 0.043442219495773315, + 0.01321109663695097, + 0.02346731722354889, + 0.00428032735362649, + 0.01719735562801361, + 0.01378539390861988, + 0.00468424241989851, + 0.008559571579098701, + 0.0045765177346765995, + 0.004638330545276403, + 0.004178022034466267, + 0.01941707916557789, + 0.02585338056087494, + 0.027828317135572433, + 0.05713442340493202, + 0.028646770864725113, + 0.04230072721838951, + 0.04483500123023987, + 0.04653076454997063, + 0.0032475267071276903, + 0.00472458079457283, + 0.005716689396649599, + 0.010286437347531319, + 0.006388637702912092, + 0.002885926980525255, + 0.003551185829564929, + 0.003679764922708273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 2 + ], + "selected_reward": -0.007630960550159216 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043069224804639816 + ], + "opposite_logprob_tokens": [ + -3.1680691242218018 + ], + "logprob_reward": -0.043069224804639816, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043069224804639816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12372773140668869, + 0.1564561277627945, + 0.07208000123500824, + 0.07589057087898254, + 0.1000068187713623, + 0.15677295625209808, + 0.08577508479356766, + 0.10161874443292618, + 0.06604549288749695, + 0.06162645295262337 + ], + "credit_map": [ + 0.013904014602303505, + 0.011789158917963505, + 0.019582726061344147, + 0.018739959225058556, + 0.009144844487309456, + 0.008450491353869438, + 0.027612168341875076, + 0.014504390768706799, + 0.010610845871269703, + 0.0197149571031332, + 0.03153622895479202, + 0.02235698327422142, + 0.010658986866474152, + 0.006530240643769503, + 0.03662865236401558, + 0.018419262021780014, + 0.005561225116252899, + 0.0083411680534482, + 0.013112775050103664, + 0.011274716816842556, + 0.005341102369129658, + 0.0036474845837801695, + 0.0156968105584383, + 0.00910472497344017, + 0.0033125479239970446, + 0.011757184751331806, + 0.015351653099060059, + 0.01809556968510151, + 0.0069696661084890366, + 0.0031431440729647875, + 0.011784350499510765, + 0.005476467311382294, + 0.007447538431733847, + 0.010572191327810287, + 0.010882463306188583, + 0.021662095561623573, + 0.010832339525222778, + 0.008547131903469563, + 0.013171997852623463, + 0.016891077160835266, + 0.02694256417453289, + 0.016304634511470795, + 0.020261574536561966, + 0.01622021570801735, + 0.019161799922585487, + 0.02018156833946705, + 0.017525508999824524, + 0.020175088196992874, + 0.01741218939423561, + 0.00892944261431694, + 0.012366955168545246, + 0.0045422655530273914, + 0.0139261269941926, + 0.013829508796334267, + 0.005953109823167324, + 0.008815493434667587, + 0.010354612953960896, + 0.026357408612966537, + 0.010260961018502712, + 0.009491555392742157, + 0.014602978713810444, + 0.014170872047543526, + 0.007572909817099571, + 0.008807455189526081, + 0.006020461209118366, + 0.00635558133944869, + 0.005061098374426365, + 0.00868234969675541, + 0.00729408347979188, + 0.01615409180521965, + 0.008426496759057045, + 0.008051340468227863, + 0.004690730012953281, + 0.0073705376125872135, + 0.009130161255598068, + 0.016130248084664345, + 0.009371780790388584, + 0.005000190809369087, + 0.0052353814244270325, + 0.004697432741522789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.043069224804639816 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Extreme close-up static shots, rendered with hyper-real digital clarity, capture the scene with scientific precision. Within a simple, transparent drinking glass, set against a softly blurred neutral background, lies a conical mound of ultra-fine, vibrant, electric-green barley powder, its texture so fine that individual grains are visible. Bright, even studio lighting illuminates the glass, casting no harsh shadows and making the colors pop. A perfectly spherical drop of crystal-clear water descends with deliberate slowness, its surface reflecting the light. Upon contact with the highest peak of the powder, the fine particles instantly seethe in a chain reaction. As the vibrant green transforms into a rich, dark, almost emerald liquid, the mound visibly collapses inward, and the liquid immediately begins to swirl with methodical precision, creating an elegant, captivating vortex that pulls the remaining dry powder into the deepening green swirl, creating a focused and hypnotic atmosphere of satisfying transformation.", + "seed": 24142, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_14/index_8056_seed_24142_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03786709904670715, + "vlm_logprob": -0.03786709904670715, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.03786709904670715, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07526367157697678 + ], + "opposite_logprob_tokens": [ + -2.700263738632202 + ], + "logprob_reward": -0.07526367157697678, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07526367157697678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1171087697148323, + 0.1077098473906517, + 0.06966381520032883, + 0.07225970923900604, + 0.10549546033143997, + 0.0666658952832222, + 0.1792854368686676, + 0.14310899376869202, + 0.12552721798419952, + 0.013174906373023987 + ], + "credit_map": [ + 0.014470549300312996, + 0.009276246652007103, + 0.007890826091170311, + 0.011364461854100227, + 0.029270656406879425, + 0.01955106481909752, + 0.01714835874736309, + 0.00813659094274044, + 0.01550485473126173, + 0.017221299931406975, + 0.009961671195924282, + 0.012294033542275429, + 0.01287966687232256, + 0.011220308020710945, + 0.019307326525449753, + 0.009320676326751709, + 0.006924866698682308, + 0.005704483482986689, + 0.004690872505307198, + 0.007503123953938484, + 0.014059759676456451, + 0.017914529889822006, + 0.007357349153608084, + 0.005508829839527607, + 0.009522353298962116, + 0.006800093222409487, + 0.00380536075681448, + 0.00985023844987154, + 0.01290968433022499, + 0.010404327884316444, + 0.014225286431610584, + 0.004742356017231941, + 0.015371664427220821, + 0.012040193192660809, + 0.009852462448179722, + 0.015186313539743423, + 0.015131563879549503, + 0.016553467139601707, + 0.011855171993374825, + 0.009504606947302818, + 0.010079297237098217, + 0.007119498681277037, + 0.0064820656552910805, + 0.007666648365557194, + 0.011856098659336567, + 0.00867640320211649, + 0.007515660487115383, + 0.00727021275088191, + 0.015832066535949707, + 0.017098689451813698, + 0.020153885707259178, + 0.03618841618299484, + 0.020315829664468765, + 0.018254190683364868, + 0.037777408957481384, + 0.01366493385285139, + 0.023821108043193817, + 0.007596487645059824, + 0.0071691712364554405, + 0.013994374312460423, + 0.026345496997237206, + 0.04377923905849457, + 0.011733668856322765, + 0.008669436909258366, + 0.014155708253383636, + 0.014756967313587666, + 0.012923197820782661, + 0.02480127103626728, + 0.009404804557561874, + 0.009302976541221142, + 0.0150372963398695, + 0.025144992396235466, + 0.0014598198467865586, + 0.0023422709200531244, + 0.0016671285266056657, + 0.001914770808070898, + 0.0019458044553175569, + 0.0009396927198395133, + 0.0014634582912549376, + 0.0014419592916965485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.07526367157697678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060367725789547 + ], + "opposite_logprob_tokens": [ + -3.895603656768799 + ], + "logprob_reward": -0.02060367725789547, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060367725789547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10847319662570953, + 0.05436054989695549, + 0.06320231407880783, + 0.13025352358818054, + 0.05282480642199516, + 0.024367326870560646, + 0.19952768087387085, + 0.29497599601745605, + 0.05283462256193161, + 0.01917998120188713 + ], + "credit_map": [ + 0.009419132024049759, + 0.006509021855890751, + 0.003736390732228756, + 0.005219184793531895, + 0.040912970900535583, + 0.030582353472709656, + 0.007089178543537855, + 0.005004976410418749, + 0.0068638138473033905, + 0.00676344521343708, + 0.005172864533960819, + 0.011529750190675259, + 0.006115839816629887, + 0.005047561600804329, + 0.009106026031076908, + 0.0037612468004226685, + 0.010081498883664608, + 0.005119109060615301, + 0.006257134955376387, + 0.008315403014421463, + 0.008057888597249985, + 0.012197905220091343, + 0.008191845379769802, + 0.004981531295925379, + 0.014646653085947037, + 0.013951811008155346, + 0.007177993655204773, + 0.016357915475964546, + 0.03270414471626282, + 0.016370898112654686, + 0.019586799666285515, + 0.009457307867705822, + 0.005418044980615377, + 0.005476979538798332, + 0.004953911527991295, + 0.007151068188250065, + 0.009824277833104134, + 0.0058351606130599976, + 0.00962788425385952, + 0.004537481814622879, + 0.0021447832696139812, + 0.0029425546526908875, + 0.002104533137753606, + 0.0025965047534555197, + 0.005488819442689419, + 0.0027801678515970707, + 0.0037978901527822018, + 0.0025120757054537535, + 0.015279099345207214, + 0.018414869904518127, + 0.023296324536204338, + 0.04231113940477371, + 0.02142409421503544, + 0.01951518841087818, + 0.044697120785713196, + 0.01458986010402441, + 0.03394988179206848, + 0.01758042722940445, + 0.039738111197948456, + 0.05606578290462494, + 0.02292788214981556, + 0.041636593639850616, + 0.057893890887498856, + 0.02518344111740589, + 0.005875458940863609, + 0.008386475034058094, + 0.005279358476400375, + 0.007625455968081951, + 0.008620332926511765, + 0.00486149312928319, + 0.006797900423407555, + 0.005388150457292795, + 0.0019529053242877126, + 0.002961718011647463, + 0.0029623969458043575, + 0.002672479720786214, + 0.002564181573688984, + 0.0016624194104224443, + 0.0021427760366350412, + 0.002261105226352811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.02060367725789547 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058732882142067 + ], + "opposite_logprob_tokens": [ + -3.89558744430542 + ], + "logprob_reward": -0.02058732882142067, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058732882142067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08663284033536911, + 0.06192164123058319, + 0.08296596258878708, + 0.07514062523841858, + 0.06781725585460663, + 0.04576073959469795, + 0.2391132265329361, + 0.21321134269237518, + 0.07539913058280945, + 0.052037231624126434 + ], + "credit_map": [ + 0.01178999338299036, + 0.009077112190425396, + 0.007195118349045515, + 0.013608233071863651, + 0.015851380303502083, + 0.009141475893557072, + 0.013263726606965065, + 0.00670580891892314, + 0.006220325827598572, + 0.005989301484078169, + 0.004838651046156883, + 0.010819390416145325, + 0.008494582958519459, + 0.0065077622421085835, + 0.012525704689323902, + 0.006525923032313585, + 0.008910780772566795, + 0.006298039108514786, + 0.006840528454631567, + 0.011576555669307709, + 0.01758952997624874, + 0.016799230128526688, + 0.008578484877943993, + 0.006372817326337099, + 0.010575927793979645, + 0.0072065494023263454, + 0.004916305188089609, + 0.008616066537797451, + 0.014681997708976269, + 0.012183853425085545, + 0.011973780579864979, + 0.0049861520528793335, + 0.008274015039205551, + 0.008439639583230019, + 0.00817823689430952, + 0.009614396840333939, + 0.010889207012951374, + 0.008192197419703007, + 0.009833833202719688, + 0.004395733587443829, + 0.005925000179558992, + 0.005654708482325077, + 0.005222756881266832, + 0.007285237777978182, + 0.007177319377660751, + 0.004604538902640343, + 0.005211524665355682, + 0.004679656587541103, + 0.02139381878077984, + 0.020601024851202965, + 0.030167710036039352, + 0.05604234337806702, + 0.01963958702981472, + 0.022522209212183952, + 0.04708979278802872, + 0.021656764671206474, + 0.02526530623435974, + 0.015237579122185707, + 0.04133515805006027, + 0.027827009558677673, + 0.021499861031770706, + 0.027701793238520622, + 0.023397346958518028, + 0.030947303399443626, + 0.00910993479192257, + 0.007003941107541323, + 0.005212067626416683, + 0.01954369992017746, + 0.00790221244096756, + 0.005317701958119869, + 0.014705735258758068, + 0.006603849586099386, + 0.00548820523545146, + 0.007920206524431705, + 0.00651227543130517, + 0.00753032136708498, + 0.008819048292934895, + 0.0037429798394441605, + 0.006125184241682291, + 0.005899012554436922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.02058732882142067 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980896644294262 + ], + "opposite_logprob_tokens": [ + -3.52980899810791 + ], + "logprob_reward": -0.02980896644294262, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980896644294262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14353010058403015, + 0.0594380684196949, + 0.05773399770259857, + 0.04600100591778755, + 0.0895732119679451, + 0.04181376099586487, + 0.26720884442329407, + 0.2134798765182495, + 0.032147303223609924, + 0.04907377436757088 + ], + "credit_map": [ + 0.011866660788655281, + 0.010210790671408176, + 0.007658807095140219, + 0.015008433721959591, + 0.04241965711116791, + 0.031790267676115036, + 0.01459599006921053, + 0.009979513473808765, + 0.006425346713513136, + 0.0059560127556324005, + 0.009317759424448013, + 0.008573301136493683, + 0.00867966003715992, + 0.005920651368796825, + 0.009717498905956745, + 0.0048478408716619015, + 0.007163673173636198, + 0.004361639730632305, + 0.006477536633610725, + 0.010346665047109127, + 0.006795068271458149, + 0.009478778578341007, + 0.008709646761417389, + 0.004400992766022682, + 0.004295787774026394, + 0.0037465025670826435, + 0.003645069897174835, + 0.008146786130964756, + 0.009676573798060417, + 0.00535208685323596, + 0.008127251639962196, + 0.003010951913893223, + 0.012268243357539177, + 0.010332404635846615, + 0.008160276338458061, + 0.011383959092199802, + 0.013940119184553623, + 0.014184588566422462, + 0.011429935693740845, + 0.00787369254976511, + 0.00408928794786334, + 0.004056563135236502, + 0.004383186809718609, + 0.004725063219666481, + 0.0077203745022416115, + 0.00456094229593873, + 0.006930118426680565, + 0.00534822978079319, + 0.02147025801241398, + 0.021576272323727608, + 0.03310313820838928, + 0.06481565535068512, + 0.021273484453558922, + 0.024016669020056725, + 0.05898473784327507, + 0.02196863293647766, + 0.01858326978981495, + 0.01318586990237236, + 0.028282463550567627, + 0.04404815658926964, + 0.024078406393527985, + 0.019013095647096634, + 0.04170360043644905, + 0.024585040286183357, + 0.003325045108795166, + 0.003044326324015856, + 0.0024637903552502394, + 0.00914198998361826, + 0.0032801541965454817, + 0.0017776014283299446, + 0.006316053215414286, + 0.0027983428444713354, + 0.004868097137659788, + 0.004467269405722618, + 0.008621163666248322, + 0.00926303118467331, + 0.004360842518508434, + 0.004277403932064772, + 0.006435396149754524, + 0.006780568975955248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.02980896644294262 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307185113430023 + ], + "opposite_logprob_tokens": [ + -3.168071746826172 + ], + "logprob_reward": -0.04307185113430023, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307185113430023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18162821233272552, + 0.14975453913211823, + 0.07662612199783325, + 0.09857872873544693, + 0.08728567510843277, + 0.04456771910190582, + 0.06509369611740112, + 0.18860392272472382, + 0.09186208248138428, + 0.01599939353764057 + ], + "credit_map": [ + 0.020438512787222862, + 0.021827541291713715, + 0.01108766533434391, + 0.019620174542069435, + 0.04258235916495323, + 0.031148478388786316, + 0.022937364876270294, + 0.011986096389591694, + 0.017280826345086098, + 0.014375035651028156, + 0.01892218552529812, + 0.023937063291668892, + 0.016499394550919533, + 0.018465077504515648, + 0.027329472824931145, + 0.012945475056767464, + 0.010161022655665874, + 0.007589530199766159, + 0.0076048593036830425, + 0.010104686953127384, + 0.015488279983401299, + 0.008204663172364235, + 0.010926726274192333, + 0.006546346936374903, + 0.01471112109720707, + 0.009888138622045517, + 0.006624702364206314, + 0.012625595554709435, + 0.024657147005200386, + 0.012640135362744331, + 0.0117397615686059, + 0.005692115053534508, + 0.01004866138100624, + 0.009486963972449303, + 0.009425965137779713, + 0.0145719014108181, + 0.011592772789299488, + 0.00973707064986229, + 0.0137841971591115, + 0.008638134226202965, + 0.00496751070022583, + 0.00408159801736474, + 0.005989531520754099, + 0.005727983079850674, + 0.008886118419468403, + 0.0053107948042452335, + 0.004949496127665043, + 0.004654679447412491, + 0.006356991361826658, + 0.006841885391622782, + 0.009769024327397346, + 0.012209653854370117, + 0.008602830581367016, + 0.006393717136234045, + 0.008029316551983356, + 0.0068902671337127686, + 0.030711930245161057, + 0.006913383956998587, + 0.007437455467879772, + 0.028586169704794884, + 0.02365386299788952, + 0.052601929754018784, + 0.025510216131806374, + 0.013188965618610382, + 0.007386869750916958, + 0.00953318364918232, + 0.006787819787859917, + 0.023140298202633858, + 0.011316576972603798, + 0.006647319532930851, + 0.01897401176393032, + 0.008075990714132786, + 0.0015641803620383143, + 0.0022023203782737255, + 0.0018929534126073122, + 0.0019780208822339773, + 0.0028091268613934517, + 0.0012407168978825212, + 0.0020239620935171843, + 0.002288110088557005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.04307185113430023 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Extreme close-up static shots, rendered with hyper-real digital clarity, capture the scene with scientific precision. Within a simple, transparent drinking glass, set against a softly blurred neutral background, lies a conical mound of ultra-fine, vibrant, electric-green barley powder, its texture so fine that individual grains are visible. Bright, even studio lighting illuminates the glass, casting no harsh shadows and making the colors pop. A perfectly spherical drop of crystal-clear water descends with deliberate slowness, its surface reflecting the light. Upon contact with the highest peak of the powder, the fine particles instantly seethe in a chain reaction. As the vibrant green transforms into a rich, dark, almost emerald liquid, the mound visibly collapses inward, and the liquid immediately begins to swirl with methodical precision, creating an elegant, captivating vortex that pulls the remaining dry powder into the deepening green swirl, creating a focused and hypnotic atmosphere of satisfying transformation.", + "seed": 24143, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_14/index_8056_seed_24143_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02786192437633872, + "vlm_logprob": -0.02786192437633872, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.02786192437633872, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03391695395112038 + ], + "opposite_logprob_tokens": [ + -3.53391695022583 + ], + "logprob_reward": -0.03391695395112038, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03391695395112038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8199754357337952, + 0.017197685316205025, + 0.019603395834565163, + 0.01867499016225338, + 0.011188848875463009, + 0.01752343773841858, + 0.03459176793694496, + 0.01524482574313879, + 0.03344384580850601, + 0.012555807828903198 + ], + "credit_map": [ + 0.26969003677368164, + 0.09718729555606842, + 0.028407609090209007, + 0.014687979593873024, + 0.02919095754623413, + 0.33256852626800537, + 0.021843789145350456, + 0.026399189606308937, + 0.0023129472974687815, + 0.00215800479054451, + 0.0023960433900356293, + 0.002408213447779417, + 0.001739712548442185, + 0.002774953143671155, + 0.0021451935172080994, + 0.0012626174138858914, + 0.003733773482963443, + 0.002725128550082445, + 0.002618380356580019, + 0.0011777571635320783, + 0.0025865209754556417, + 0.0034271744079887867, + 0.002004280686378479, + 0.0013303799787536263, + 0.0027877839747816324, + 0.0024186146911233664, + 0.0019827838987112045, + 0.0013682876015082002, + 0.001842708676122129, + 0.006010646000504494, + 0.0011181493755429983, + 0.001146015478298068, + 0.00135285675060004, + 0.0020241544116288424, + 0.0019297508988529444, + 0.0009949356317520142, + 0.001556865405291319, + 0.0012128932867199183, + 0.0011097791139036417, + 0.0010076129110530019, + 0.0011914924252778292, + 0.002687053522095084, + 0.0038248817436397076, + 0.0016670211916789412, + 0.003733172081410885, + 0.001080524642020464, + 0.0020037302747368813, + 0.0013355622068047523, + 0.0030429784674197435, + 0.006821016315370798, + 0.006112707778811455, + 0.0028971994761377573, + 0.0043848236091434956, + 0.00471818633377552, + 0.0030785815324634314, + 0.003536275355145335, + 0.0016468148678541183, + 0.0016588824801146984, + 0.0030066187027841806, + 0.0019112678710371256, + 0.0019045750377699733, + 0.0015624351799488068, + 0.0016041788039729, + 0.0019500517519190907, + 0.002338946098461747, + 0.002448938088491559, + 0.010110354982316494, + 0.009818361140787601, + 0.0024685736279934645, + 0.0018000603886321187, + 0.0016997314523905516, + 0.002758878981694579, + 0.0008093543583527207, + 0.0008441073005087674, + 0.004837713669985533, + 0.002845287322998047, + 0.0009672193555161357, + 0.0005977748660370708, + 0.0006110443500801921, + 0.0010433067800477147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.03391695395112038 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067511084489524364 + ], + "opposite_logprob_tokens": [ + -5.00675106048584 + ], + "logprob_reward": -0.0067511084489524364, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067511084489524364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2100454568862915, + 0.10870644450187683, + 0.03790721669793129, + 0.12281349301338196, + 0.06033767759799957, + 0.13672040402889252, + 0.1738252341747284, + 0.06533487886190414, + 0.042209986597299576, + 0.042099207639694214 + ], + "credit_map": [ + 0.06379280984401703, + 0.02533087506890297, + 0.009145036339759827, + 0.007224593311548233, + 0.007899947464466095, + 0.08199967443943024, + 0.007683948148041964, + 0.006968577858060598, + 0.027637183666229248, + 0.012017834931612015, + 0.007683781441301107, + 0.005577084142714739, + 0.006233712192624807, + 0.03678406774997711, + 0.007464057300239801, + 0.0053087277337908745, + 0.006848576944321394, + 0.007702838163822889, + 0.004046390298753977, + 0.0027947700582444668, + 0.006778755225241184, + 0.004821993410587311, + 0.0028583710081875324, + 0.002055524382740259, + 0.024785835295915604, + 0.010724833235144615, + 0.00881635770201683, + 0.005928815808147192, + 0.010088427923619747, + 0.04835159704089165, + 0.006634914316236973, + 0.007482712622731924, + 0.009292079135775566, + 0.009409908205270767, + 0.008062151260674, + 0.0051445262506604195, + 0.006912632845342159, + 0.009080150164663792, + 0.005790386348962784, + 0.006645839661359787, + 0.010611580684781075, + 0.012362099252641201, + 0.034115590155124664, + 0.006627077702432871, + 0.04807038605213165, + 0.010740917176008224, + 0.007964490912854671, + 0.006228259764611721, + 0.018036076799035072, + 0.04744410142302513, + 0.015781879425048828, + 0.017592797055840492, + 0.02364977076649666, + 0.017672790214419365, + 0.01832462288439274, + 0.015323189087212086, + 0.007223567925393581, + 0.004874580539762974, + 0.01073024794459343, + 0.00876580085605383, + 0.008253215812146664, + 0.007005853112787008, + 0.010516376234591007, + 0.007965236902236938, + 0.005149995442479849, + 0.004002576228231192, + 0.007451880723237991, + 0.008190830238163471, + 0.0038846496026962996, + 0.003097901353612542, + 0.0029021804220974445, + 0.007529971655458212, + 0.004122063517570496, + 0.0034065840300172567, + 0.004845442716032267, + 0.012522364035248756, + 0.003656176384538412, + 0.003796585137024522, + 0.004415132571011782, + 0.005334862507879734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0067511084489524364 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543123215436935 + ], + "opposite_logprob_tokens": [ + -4.387543201446533 + ], + "logprob_reward": -0.012543123215436935, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543123215436935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3771185576915741, + 0.06099238246679306, + 0.017718974500894547, + 0.047799985855817795, + 0.054172221571207047, + 0.2675043046474457, + 0.05206076055765152, + 0.05817065387964249, + 0.01947415992617607, + 0.04498795419931412 + ], + "credit_map": [ + 0.11977045983076096, + 0.04178453981876373, + 0.016122499480843544, + 0.014520537108182907, + 0.013945046812295914, + 0.14446419477462769, + 0.013195006176829338, + 0.013316281139850616, + 0.01666538417339325, + 0.0075018866918981075, + 0.003677177242934704, + 0.0028457539156079292, + 0.002639623824506998, + 0.022799398750066757, + 0.0028103392105549574, + 0.0020528233144432306, + 0.003226018510758877, + 0.002912591677159071, + 0.0019548290874809027, + 0.001591106760315597, + 0.002604876644909382, + 0.0021216932218521833, + 0.0019939017947763205, + 0.0013139566872268915, + 0.008395317010581493, + 0.00529871741309762, + 0.0036187402438372374, + 0.0024052828084677458, + 0.00541169848293066, + 0.017701217904686928, + 0.0020531436894088984, + 0.00291586946696043, + 0.006855306215584278, + 0.00893403310328722, + 0.006037845741957426, + 0.007369596976786852, + 0.006487902719527483, + 0.0057611060328781605, + 0.007170349825173616, + 0.005556081887334585, + 0.023375030606985092, + 0.01983465813100338, + 0.056819330900907516, + 0.013157731853425503, + 0.1036892831325531, + 0.022182224318385124, + 0.01519840583205223, + 0.013247611932456493, + 0.005303326528519392, + 0.008552293293178082, + 0.007615877315402031, + 0.004314571153372526, + 0.007101657334715128, + 0.008252966217696667, + 0.005757671315222979, + 0.005162398796528578, + 0.006596534512937069, + 0.007174279540777206, + 0.007275579497218132, + 0.007065565790981054, + 0.007812249939888716, + 0.007021908648312092, + 0.00751151330769062, + 0.007713024504482746, + 0.0012868075864389539, + 0.0026537575758993626, + 0.0036109175998717546, + 0.005528340116143227, + 0.0017722557531669736, + 0.0010107886046171188, + 0.0014138970291242003, + 0.002197395544499159, + 0.002840888686478138, + 0.002215647604316473, + 0.010800953023135662, + 0.016275329515337944, + 0.002618250669911504, + 0.0030094801913946867, + 0.003170928917825222, + 0.004056474193930626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.012543123215436935 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073059290647507 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.016073059290647507, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073059290647507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6179732084274292, + 0.0816301703453064, + 0.007957291789352894, + 0.02034829370677471, + 0.013208054937422276, + 0.032890088856220245, + 0.04790544509887695, + 0.04139339178800583, + 0.04258844628930092, + 0.09410557150840759 + ], + "credit_map": [ + 0.20254233479499817, + 0.07436913251876831, + 0.021211983636021614, + 0.010210158303380013, + 0.021950282156467438, + 0.25165310502052307, + 0.016211915761232376, + 0.019824279472231865, + 0.023073159158229828, + 0.010098142549395561, + 0.004587084054946899, + 0.0025702500715851784, + 0.0033975050318986177, + 0.03186508268117905, + 0.003087334567680955, + 0.0029516108334064484, + 0.0010515425819903612, + 0.001169936847873032, + 0.001205529784783721, + 0.0007714987732470036, + 0.001200409489683807, + 0.0009388528997078538, + 0.0009297853684984148, + 0.0006897360435687006, + 0.002979687415063381, + 0.0032543237321078777, + 0.0018806589068844914, + 0.0017547027673572302, + 0.0024808431044220924, + 0.0055222357623279095, + 0.001282591954804957, + 0.001193250878714025, + 0.0018330676248297095, + 0.0018502097809687257, + 0.0017328901449218392, + 0.0014741268241778016, + 0.0015480784932151437, + 0.0016582530224695802, + 0.0017234324477612972, + 0.0013879963662475348, + 0.0026061467360705137, + 0.00325038586743176, + 0.006138102151453495, + 0.004568659700453281, + 0.005424825940281153, + 0.0027720320504158735, + 0.0049024298787117004, + 0.003227504901587963, + 0.004406902007758617, + 0.009636133909225464, + 0.009319149889051914, + 0.0032404966186732054, + 0.006752359215170145, + 0.004860765766352415, + 0.005423585884273052, + 0.004266052506864071, + 0.0044191754423081875, + 0.003640390932559967, + 0.005803920794278383, + 0.004361922387033701, + 0.005759530700743198, + 0.004571492318063974, + 0.007032657042145729, + 0.005804304964840412, + 0.002338198246434331, + 0.002434053458273411, + 0.015504827722907066, + 0.014305712655186653, + 0.0019588745199143887, + 0.0016105613904073834, + 0.001997229177504778, + 0.0024389890022575855, + 0.004538837354630232, + 0.0035167455207556486, + 0.036908965557813644, + 0.03508158400654793, + 0.0035760069731622934, + 0.00296670850366354, + 0.0038390131667256355, + 0.003677709959447384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.016073059290647507 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a conical mound of ultra-fine, electric-green barley powder inside a transparent drinking glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a perfectly spherical drop of crystal-clear water descend slowly and make contact with the highest peak of the powder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Upon contact, do the fine particles instantly seethe in a chain reaction, transforming the vibrant green powder into a rich, dark emerald liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid mound visibly collapse inward and immediately begin to swirl with methodical precision into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred neutral color with bright, even studio lighting that casts no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002537697553635 + ], + "opposite_logprob_tokens": [ + -2.6950254440307617 + ], + "logprob_reward": -0.07002537697553635, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002537697553635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2378305196762085, + 0.3058040738105774, + 0.03298070281744003, + 0.08425188064575195, + 0.03478745371103287, + 0.09719438850879669, + 0.0840054303407669, + 0.045985519886016846, + 0.02899872325360775, + 0.048161230981349945 + ], + "credit_map": [ + 0.07518478482961655, + 0.027424070984125137, + 0.009483610279858112, + 0.007505805231630802, + 0.009304603561758995, + 0.0923132449388504, + 0.008011575788259506, + 0.008602837100625038, + 0.049492258578538895, + 0.03511027619242668, + 0.03317530080676079, + 0.03773792088031769, + 0.020301654934883118, + 0.06997191160917282, + 0.04000306874513626, + 0.020011698827147484, + 0.004852354526519775, + 0.005697282962501049, + 0.00484489556401968, + 0.002781623275950551, + 0.005247793160378933, + 0.0041146534495055676, + 0.0032726384233683348, + 0.0021694612223654985, + 0.014371889643371105, + 0.008547080680727959, + 0.008678905665874481, + 0.004145164042711258, + 0.00841515138745308, + 0.03167785704135895, + 0.0037714058998972178, + 0.0046444302424788475, + 0.0055053383111953735, + 0.004264891147613525, + 0.004939910490065813, + 0.0036251810379326344, + 0.00470450334250927, + 0.004599754698574543, + 0.003868402913212776, + 0.0032794689759612083, + 0.006007656455039978, + 0.007136759348213673, + 0.02822624333202839, + 0.007017363328486681, + 0.02967882715165615, + 0.006887400522828102, + 0.005973770748823881, + 0.00626636715605855, + 0.007594133261591196, + 0.014904690906405449, + 0.01371660828590393, + 0.005758649669587612, + 0.01039901003241539, + 0.015755515545606613, + 0.007615591399371624, + 0.008261228911578655, + 0.004588673822581768, + 0.005593638401478529, + 0.005432256497442722, + 0.0062969946302473545, + 0.006478732451796532, + 0.0053487843833863735, + 0.005610531195998192, + 0.006635907106101513, + 0.0019271125784143806, + 0.006298759952187538, + 0.004628568422049284, + 0.004300544038414955, + 0.004772275686264038, + 0.0018588545499369502, + 0.00306399492546916, + 0.0021486133337020874, + 0.0035636876709759235, + 0.004769519437104464, + 0.01327226310968399, + 0.011314967647194862, + 0.005583389662206173, + 0.0028902485501021147, + 0.0036987492348998785, + 0.0030684068333357573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.07002537697553635 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_15/index_tensor([4060, 4060])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_15/index_tensor([4060, 4060])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..792e1af504fc17a8be177759e514972261a484e8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_15/index_tensor([4060, 4060])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sharp, shallow depth-of-field close-up focuses squarely on the awed face of a young woman dressed in traditional shrine maiden garb. She stands quietly at the wooden steps of an ancient shrine, her long, flowing, raven-black hair held back by a simple white washi headband, strands brushing against her fair cheeks in the night breeze. The cool moonlight frames her delicate profile. Her head tilted slightly, her lips parted in wonder, her eyes, as clear as deep pools, gaze at the abyss of the night sky. Within these eyes, a magnificent meteor shower is perfectly reflected: countless fleeting, slender, brilliant arcs of silver light streak across her pupils, slightly dilated with concentration, as if the entire galaxy were ablaze in her gaze. Her pristine white Chihaya coat and crimson hakama take on a solemn yet soft quality in the dim light, while the shrine's architectural outlines and hanging stone lanterns behind her are completely blurred, transforming into a dreamy, softly shimmering backdrop composed of warm yellow dots. The entire image, shrouded in profound tranquility, exudes a sacred and ethereal beauty of classical Oriental beauty, capturing a poetic moment of dialogue between humanity and the universe.", + "seed": 5018, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_15/index_4060_seed_5018_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.016278811683878302, + "vlm_logprob": -0.016278811683878302, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.016278811683878302, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011757600121200085 + ], + "opposite_logprob_tokens": [ + -4.8867573738098145 + ], + "logprob_reward": -0.011757600121200085, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.011757600121200085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3530263900756836, + 0.047290313988924026, + 0.04856952279806137, + 0.0683855265378952, + 0.06097705662250519, + 0.0425385944545269, + 0.0428939051926136, + 0.07661881297826767, + 0.10810168832540512, + 0.15159815549850464 + ], + "credit_map": [ + 0.04123338311910629, + 0.031920406967401505, + 0.11531679332256317, + 0.017722448334097862, + 0.025726420804858208, + 0.03402600437402725, + 0.012446220964193344, + 0.07463472336530685, + 0.0061977701261639595, + 0.004004818852990866, + 0.004689228720963001, + 0.00604267930611968, + 0.007572194095700979, + 0.006635025609284639, + 0.006985840387642384, + 0.005162762477993965, + 0.004401832353323698, + 0.0060829222202301025, + 0.006642101798206568, + 0.004675827920436859, + 0.010207880288362503, + 0.006974095944315195, + 0.005743703804910183, + 0.0038411577697843313, + 0.00721339788287878, + 0.008941089734435081, + 0.005823554005473852, + 0.005934935994446278, + 0.0076221139170229435, + 0.009828929789364338, + 0.00976313091814518, + 0.013258378021419048, + 0.007168610580265522, + 0.007618358824402094, + 0.009268662892282009, + 0.005512065254151821, + 0.005876488517969847, + 0.008541645482182503, + 0.005138050764799118, + 0.011853175237774849, + 0.00550079345703125, + 0.006178323179483414, + 0.009060020558536053, + 0.002819924848154187, + 0.005679870955646038, + 0.006564296316355467, + 0.0030334831681102514, + 0.0037018812727183104, + 0.0037136587779968977, + 0.006082464940845966, + 0.008005190640687943, + 0.004179502837359905, + 0.007119735237210989, + 0.007022270932793617, + 0.0033441567793488503, + 0.003426922485232353, + 0.005129820201545954, + 0.011701641604304314, + 0.012125014327466488, + 0.00759788416326046, + 0.008530470542609692, + 0.013469216413795948, + 0.005598021671175957, + 0.012466740794479847, + 0.013119814917445183, + 0.0115043418481946, + 0.009807066060602665, + 0.01597430557012558, + 0.020397931337356567, + 0.013143911957740784, + 0.013934831134974957, + 0.010219485498964787, + 0.017178788781166077, + 0.006983808241784573, + 0.020197704434394836, + 0.015524529851973057, + 0.03678654134273529, + 0.019955120980739594, + 0.025845196098089218, + 0.00912646483629942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.011757600121200085 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02332068420946598 + ], + "opposite_logprob_tokens": [ + -3.7733206748962402 + ], + "logprob_reward": -0.02332068420946598, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02332068420946598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4262368381023407, + 0.06996329873800278, + 0.04839823767542839, + 0.13807222247123718, + 0.14613935351371765, + 0.029829371720552444, + 0.018735786899924278, + 0.05419919639825821, + 0.04234949126839638, + 0.02607620134949684 + ], + "credit_map": [ + 0.05633629858493805, + 0.02127000316977501, + 0.1481216698884964, + 0.026118867099285126, + 0.019439276307821274, + 0.03032255358994007, + 0.019495926797389984, + 0.10513224452733994, + 0.009579582139849663, + 0.00885884091258049, + 0.009217755869030952, + 0.007037687581032515, + 0.006880763918161392, + 0.009602031670510769, + 0.009266563691198826, + 0.009520069696009159, + 0.007092825602740049, + 0.005893038585782051, + 0.0065937405452132225, + 0.00353445578366518, + 0.008462413214147091, + 0.007571229245513678, + 0.004191381391137838, + 0.0050591519102454185, + 0.013919600285589695, + 0.010816531255841255, + 0.013721375726163387, + 0.01897404156625271, + 0.03280064836144447, + 0.01924043335020542, + 0.013028559274971485, + 0.015571031719446182, + 0.026920771226286888, + 0.023749997839331627, + 0.023944104090332985, + 0.013112415559589863, + 0.012526958249509335, + 0.024173522368073463, + 0.00951599795371294, + 0.012195593677461147, + 0.004445889964699745, + 0.005099916830658913, + 0.0037006742786616087, + 0.002347653266042471, + 0.0044267731718719006, + 0.004457044415175915, + 0.0031541623175144196, + 0.0021972586400806904, + 0.0018511234084144235, + 0.0027525790501385927, + 0.004014898091554642, + 0.0017727937083691359, + 0.0021179006434977055, + 0.003242687787860632, + 0.001354233012534678, + 0.0016295708483085036, + 0.004261043854057789, + 0.01350064855068922, + 0.006972789764404297, + 0.0032385822851210833, + 0.0027019872795790434, + 0.017003629356622696, + 0.0020779480691999197, + 0.004442567937076092, + 0.005574687849730253, + 0.006699215155094862, + 0.005304892547428608, + 0.004454459063708782, + 0.005603319965302944, + 0.005410745739936829, + 0.004393914248794317, + 0.004908259492367506, + 0.003547962987795472, + 0.0022693302016705275, + 0.004185777623206377, + 0.002315543359145522, + 0.005422225687652826, + 0.0035076874773949385, + 0.0026752115227282047, + 0.002152464585378766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.02332068420946598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978801865130663 + ], + "opposite_logprob_tokens": [ + -4.6347880363464355 + ], + "logprob_reward": -0.00978801865130663, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978801865130663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45285099744796753, + 0.06091737747192383, + 0.053013838827610016, + 0.1447378247976303, + 0.06292130053043365, + 0.0468003936111927, + 0.02317318692803383, + 0.02876889891922474, + 0.051464904099702835, + 0.07535131275653839 + ], + "credit_map": [ + 0.08255608379840851, + 0.053720422089099884, + 0.14293445646762848, + 0.024370674043893814, + 0.026428667828440666, + 0.05095882713794708, + 0.016849614679813385, + 0.055032212287187576, + 0.00864837784320116, + 0.008609363809227943, + 0.010449737310409546, + 0.006149913650006056, + 0.005323435179889202, + 0.009320482611656189, + 0.005196776241064072, + 0.007219284772872925, + 0.006430746987462044, + 0.006265255156904459, + 0.0065882946364581585, + 0.004437580704689026, + 0.011525548063218594, + 0.0078163156285882, + 0.004766727332025766, + 0.005183365661650896, + 0.014883812516927719, + 0.011310881935060024, + 0.015339558012783527, + 0.017874551936984062, + 0.03381394222378731, + 0.02111157588660717, + 0.012797629460692406, + 0.01760585978627205, + 0.010213425382971764, + 0.0075838142074644566, + 0.011226879432797432, + 0.004616699647158384, + 0.005266118329018354, + 0.010410790331661701, + 0.0055607883259654045, + 0.008042784407734871, + 0.007513662800192833, + 0.007326179649680853, + 0.007428656332194805, + 0.003403088776394725, + 0.006381658371537924, + 0.007082891650497913, + 0.003620077623054385, + 0.004044175613671541, + 0.0028317004907876253, + 0.0035004247911274433, + 0.003873980138450861, + 0.0014688349328935146, + 0.003083655843511224, + 0.0055474485270679, + 0.001210699905641377, + 0.001656441716477275, + 0.002592134289443493, + 0.005013741552829742, + 0.00376152410171926, + 0.0027915749233216047, + 0.003921797499060631, + 0.0047650523483753204, + 0.001909074024297297, + 0.0040139988996088505, + 0.005963888950645924, + 0.006319999694824219, + 0.006399767939001322, + 0.007026220206171274, + 0.010078373365104198, + 0.006515826098620892, + 0.004887121729552746, + 0.004273703787475824, + 0.007978536188602448, + 0.005126163829118013, + 0.009320175275206566, + 0.007913417182862759, + 0.017384743317961693, + 0.00927474070340395, + 0.012722906656563282, + 0.0056306165643036366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.00978801865130663 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002823891816660762 + ], + "opposite_logprob_tokens": [ + -5.877823829650879 + ], + "logprob_reward": -0.002823891816660762, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002823891816660762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28683987259864807, + 0.05645469203591347, + 0.05093987286090851, + 0.05953843519091606, + 0.16960254311561584, + 0.09550276398658752, + 0.06265042722225189, + 0.08280237019062042, + 0.04715897887945175, + 0.08851001411676407 + ], + "credit_map": [ + 0.01955444924533367, + 0.03998236358165741, + 0.0812809020280838, + 0.01483946293592453, + 0.013465708121657372, + 0.03525863215327263, + 0.008503276854753494, + 0.07395510375499725, + 0.006809378042817116, + 0.0073230634443461895, + 0.008576205931603909, + 0.007085643708705902, + 0.006349475122988224, + 0.008086278103291988, + 0.0038196491077542305, + 0.008405000902712345, + 0.006190240848809481, + 0.01033853180706501, + 0.005239017773419619, + 0.004347176756709814, + 0.006488758604973555, + 0.008644024841487408, + 0.0053322468884289265, + 0.00435987813398242, + 0.00700353505089879, + 0.007915985770523548, + 0.009201574139297009, + 0.004814018961042166, + 0.008083723485469818, + 0.009473535232245922, + 0.0050301020964980125, + 0.008015966042876244, + 0.034769438207149506, + 0.026460492983460426, + 0.02596190571784973, + 0.01601707562804222, + 0.011038334108889103, + 0.028577599674463272, + 0.008302517235279083, + 0.01847519911825657, + 0.027988940477371216, + 0.014877463690936565, + 0.009174143895506859, + 0.004848515149205923, + 0.006439764518290758, + 0.015541094355285168, + 0.005961046554148197, + 0.010671798139810562, + 0.0066750673577189445, + 0.011256028898060322, + 0.009673653170466423, + 0.005129663273692131, + 0.007291710004210472, + 0.013264310546219349, + 0.0048135071992874146, + 0.004546490032225847, + 0.009089067578315735, + 0.013012498617172241, + 0.016185186803340912, + 0.0067900074645876884, + 0.0077510010451078415, + 0.014468181878328323, + 0.003844805993139744, + 0.01166162732988596, + 0.007490232121199369, + 0.0077813733369112015, + 0.005897789727896452, + 0.004239929374307394, + 0.006751293316483498, + 0.006326490081846714, + 0.004095752723515034, + 0.0045761242508888245, + 0.011530774645507336, + 0.013218148611485958, + 0.018314706161618233, + 0.008775914087891579, + 0.010042844340205193, + 0.010927656665444374, + 0.009415533393621445, + 0.006284439470618963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.002823891816660762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370386362075806 + ], + "opposite_logprob_tokens": [ + -3.4087038040161133 + ], + "logprob_reward": -0.03370386362075806, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370386362075806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3931065499782562, + 0.05374762788414955, + 0.06015533208847046, + 0.054618313908576965, + 0.0702289342880249, + 0.024536283686757088, + 0.03467273712158203, + 0.10411147773265839, + 0.06152163818478584, + 0.14330103993415833 + ], + "credit_map": [ + 0.0896676629781723, + 0.046165864914655685, + 0.0782211497426033, + 0.02864379994571209, + 0.027758216485381126, + 0.04365739971399307, + 0.022631747648119926, + 0.056360743939876556, + 0.007129074074327946, + 0.005966255906969309, + 0.007520874962210655, + 0.008023040369153023, + 0.007026820909231901, + 0.00749672157689929, + 0.006208684761077166, + 0.004376155789941549, + 0.010521354153752327, + 0.009471405297517776, + 0.010984569787979126, + 0.0038190113846212626, + 0.0052305846475064754, + 0.010124320164322853, + 0.0038437533657997847, + 0.00616033049300313, + 0.006962867453694344, + 0.00610716175287962, + 0.0073366835713386536, + 0.006596891209483147, + 0.007094888016581535, + 0.007105302065610886, + 0.0051146261394023895, + 0.008299892768263817, + 0.011707874946296215, + 0.013344570063054562, + 0.0074323988519608974, + 0.006014157552272081, + 0.009334259666502476, + 0.01091333944350481, + 0.0049681589007377625, + 0.006514174863696098, + 0.00503500085324049, + 0.0032162065617740154, + 0.0028347240295261145, + 0.0019403230398893356, + 0.0036609387025237083, + 0.0035046276170760393, + 0.002354075200855732, + 0.001990386750549078, + 0.0041067153215408325, + 0.0058048744685947895, + 0.005827843677252531, + 0.002582210348919034, + 0.005111307837069035, + 0.006291499827057123, + 0.00230508204549551, + 0.0026432033628225327, + 0.0146323312073946, + 0.017922352999448776, + 0.01755652204155922, + 0.008537565357983112, + 0.011360601522028446, + 0.01698901876807213, + 0.005895085167139769, + 0.011218007653951645, + 0.009347365237772465, + 0.011413837783038616, + 0.008590497076511383, + 0.006782554090023041, + 0.007978086359798908, + 0.00835471972823143, + 0.005421736743301153, + 0.0036328413989394903, + 0.0291866734623909, + 0.012911967933177948, + 0.03289800137281418, + 0.010423392057418823, + 0.016866467893123627, + 0.02105526067316532, + 0.008726421743631363, + 0.011232851073145866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.03370386362075806 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sharp, shallow depth-of-field close-up focuses squarely on the awed face of a young woman dressed in traditional shrine maiden garb. She stands quietly at the wooden steps of an ancient shrine, her long, flowing, raven-black hair held back by a simple white washi headband, strands brushing against her fair cheeks in the night breeze. The cool moonlight frames her delicate profile. Her head tilted slightly, her lips parted in wonder, her eyes, as clear as deep pools, gaze at the abyss of the night sky. Within these eyes, a magnificent meteor shower is perfectly reflected: countless fleeting, slender, brilliant arcs of silver light streak across her pupils, slightly dilated with concentration, as if the entire galaxy were ablaze in her gaze. Her pristine white Chihaya coat and crimson hakama take on a solemn yet soft quality in the dim light, while the shrine's architectural outlines and hanging stone lanterns behind her are completely blurred, transforming into a dreamy, softly shimmering backdrop composed of warm yellow dots. The entire image, shrouded in profound tranquility, exudes a sacred and ethereal beauty of classical Oriental beauty, capturing a poetic moment of dialogue between humanity and the universe.", + "seed": 5019, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_15/index_4060_seed_5019_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.018629429722204806, + "vlm_logprob": -0.018629429722204806, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.018629429722204806, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012007102370262146 + ], + "opposite_logprob_tokens": [ + -6.137007236480713 + ], + "logprob_reward": -0.012007102370262146, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.012007102370262146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.142510324716568, + 0.03733323514461517, + 0.06763375550508499, + 0.06521647423505783, + 0.13454650342464447, + 0.10546860098838806, + 0.1962948888540268, + 0.060647230595350266, + 0.06751946359872818, + 0.12282952666282654 + ], + "credit_map": [ + 0.012017937377095222, + 0.018224645406007767, + 0.01700661890208721, + 0.01953141763806343, + 0.020378660410642624, + 0.02098678983747959, + 0.015828479081392288, + 0.01853577420115471, + 0.003742413828149438, + 0.005047160666435957, + 0.006053658667951822, + 0.0037293757777661085, + 0.004302236717194319, + 0.0053791506215929985, + 0.004485622048377991, + 0.00459361681714654, + 0.005797270219773054, + 0.005136882420629263, + 0.022027157247066498, + 0.004168249666690826, + 0.007511579431593418, + 0.0039042560383677483, + 0.0032953524496406317, + 0.015793008729815483, + 0.006799955852329731, + 0.009769615717232227, + 0.011915097944438457, + 0.005468590650707483, + 0.010454335249960423, + 0.004977071192115545, + 0.007434823550283909, + 0.00839698500931263, + 0.025189552456140518, + 0.017992213368415833, + 0.013563188724219799, + 0.00826224498450756, + 0.012678936123847961, + 0.013366205617785454, + 0.01669427752494812, + 0.026799874380230904, + 0.011077092960476875, + 0.014035860076546669, + 0.012350286357104778, + 0.011393295601010323, + 0.0125493835657835, + 0.01670638844370842, + 0.014111113734543324, + 0.013245168142020702, + 0.012105987407267094, + 0.021769417449831963, + 0.03315164893865585, + 0.036679014563560486, + 0.026298763230443, + 0.016992177814245224, + 0.02488631382584572, + 0.024411538615822792, + 0.005281651392579079, + 0.006070562172681093, + 0.011126015335321426, + 0.009662567637860775, + 0.008762876503169537, + 0.00743838120251894, + 0.007723941467702389, + 0.004581230226904154, + 0.007513453718274832, + 0.011113285087049007, + 0.00999496877193451, + 0.008141672238707542, + 0.006702165585011244, + 0.01040820125490427, + 0.008378482423722744, + 0.005267228931188583, + 0.008619146421551704, + 0.02600683458149433, + 0.016751987859606743, + 0.017997413873672485, + 0.0161328986287117, + 0.010659388266503811, + 0.018468640744686127, + 0.008193198591470718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.012007102370262146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868089407682419 + ], + "opposite_logprob_tokens": [ + -3.0486807823181152 + ], + "logprob_reward": -0.04868089407682419, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868089407682419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24776127934455872, + 0.1519882082939148, + 0.06612329930067062, + 0.09103412181138992, + 0.05912836268544197, + 0.07428834587335587, + 0.1431250125169754, + 0.0730881318449974, + 0.05738084390759468, + 0.036082375794649124 + ], + "credit_map": [ + 0.01408577710390091, + 0.020746029913425446, + 0.02597305178642273, + 0.07370397448539734, + 0.017954768612980843, + 0.022760160267353058, + 0.02967659942805767, + 0.042860936373472214, + 0.009057775139808655, + 0.02460041083395481, + 0.021896280348300934, + 0.023371277377009392, + 0.00956069678068161, + 0.01449523400515318, + 0.02257939986884594, + 0.02642713487148285, + 0.004728842526674271, + 0.0057224128395318985, + 0.021675841882824898, + 0.0047640493139624596, + 0.008064474910497665, + 0.003009694628417492, + 0.003927409183233976, + 0.014230579137802124, + 0.010856694541871548, + 0.008910712786018848, + 0.01928425393998623, + 0.006476246751844883, + 0.010885858908295631, + 0.009173554368317127, + 0.00771764200180769, + 0.017729150131344795, + 0.008375235833227634, + 0.007718443870544434, + 0.008401352912187576, + 0.004419846460223198, + 0.005886617116630077, + 0.007526320870965719, + 0.008725892752408981, + 0.00807464960962534, + 0.005931305233389139, + 0.009817865677177906, + 0.011226626113057137, + 0.007487030699849129, + 0.008345626294612885, + 0.01116868481040001, + 0.009367345832288265, + 0.010943860746920109, + 0.00859223585575819, + 0.019007423892617226, + 0.021952103823423386, + 0.024845795705914497, + 0.02165108546614647, + 0.012038882821798325, + 0.017738712951540947, + 0.017298778519034386, + 0.007001576479524374, + 0.00809544324874878, + 0.01102521177381277, + 0.008823482319712639, + 0.013993934728205204, + 0.007581479847431183, + 0.009978796355426311, + 0.006588204763829708, + 0.006019496358931065, + 0.011272558942437172, + 0.00703531876206398, + 0.007329882588237524, + 0.006236088462173939, + 0.0047813779674470425, + 0.009045295417308807, + 0.005660824477672577, + 0.003492361633107066, + 0.0068731047213077545, + 0.005032029934227467, + 0.0044500394724309444, + 0.00558574590831995, + 0.0034282018896192312, + 0.004546287935227156, + 0.0026746047660708427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.04868089407682419 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003620974952355027 + ], + "opposite_logprob_tokens": [ + -5.6286211013793945 + ], + "logprob_reward": -0.003620974952355027, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003620974952355027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1260783076286316, + 0.09747380763292313, + 0.09419553726911545, + 0.09195853769779205, + 0.09170640259981155, + 0.18241678178310394, + 0.06514431536197662, + 0.09782170504331589, + 0.055354829877614975, + 0.09784982353448868 + ], + "credit_map": [ + 0.011191866360604763, + 0.016958650201559067, + 0.016609488055109978, + 0.015089976601302624, + 0.012481468729674816, + 0.019604381173849106, + 0.014219340868294239, + 0.0199231319129467, + 0.011245841160416603, + 0.015773890540003777, + 0.012340866029262543, + 0.012901676818728447, + 0.008077970705926418, + 0.01343558356165886, + 0.012155620381236076, + 0.011542359367012978, + 0.006019533611834049, + 0.005449560936540365, + 0.03323952853679657, + 0.005515108350664377, + 0.011572957038879395, + 0.006222371477633715, + 0.0038639691192656755, + 0.022312507033348083, + 0.012102183885872364, + 0.014738623984158039, + 0.01114830281585455, + 0.007472958415746689, + 0.017063995823264122, + 0.009927657432854176, + 0.009665203280746937, + 0.009839603677392006, + 0.009756611660122871, + 0.011052665300667286, + 0.009937766939401627, + 0.005406201351433992, + 0.009165615774691105, + 0.015214769169688225, + 0.014977408573031425, + 0.01619536429643631, + 0.014186977408826351, + 0.022271232679486275, + 0.017759300768375397, + 0.0225188210606575, + 0.03239883482456207, + 0.027085142210125923, + 0.027890458703041077, + 0.018306011334061623, + 0.006126929074525833, + 0.007450981996953487, + 0.010169352404773235, + 0.007488244213163853, + 0.007489129435271025, + 0.013238168321549892, + 0.0078066145069897175, + 0.00537489540874958, + 0.007830324582755566, + 0.012827165424823761, + 0.012888717465102673, + 0.011500474996864796, + 0.018709957599639893, + 0.011942703276872635, + 0.01506743859499693, + 0.007054921705275774, + 0.005855840630829334, + 0.005755212157964706, + 0.007772312965244055, + 0.007559670601040125, + 0.005935730412602425, + 0.007604156155139208, + 0.007207480724900961, + 0.0076644220389425755, + 0.007617987226694822, + 0.021947981789708138, + 0.012762779369950294, + 0.014872481115162373, + 0.011276005767285824, + 0.00838313065469265, + 0.014890973456203938, + 0.006098481360822916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.003620974952355027 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002493607811629772 + ], + "opposite_logprob_tokens": [ + -6.002493381500244 + ], + "logprob_reward": -0.002493607811629772, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002493607811629772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15977616608142853, + 0.09722623974084854, + 0.06284000724554062, + 0.09096194058656693, + 0.10118474066257477, + 0.11436871439218521, + 0.21027813851833344, + 0.03095039539039135, + 0.05625254660844803, + 0.07616108655929565 + ], + "credit_map": [ + 0.00781625509262085, + 0.017589909955859184, + 0.020444344729185104, + 0.03469689562916756, + 0.015519794076681137, + 0.01912713423371315, + 0.018129335716366768, + 0.02645249478518963, + 0.01255723088979721, + 0.013953045010566711, + 0.014248283579945564, + 0.011770534329116344, + 0.007237862329930067, + 0.014056495390832424, + 0.010270862840116024, + 0.013131916522979736, + 0.006734324619174004, + 0.006335204001516104, + 0.01372531894594431, + 0.0037585506215691566, + 0.008702411316335201, + 0.011190804652869701, + 0.003872858127579093, + 0.008520525880157948, + 0.010751843452453613, + 0.008203662931919098, + 0.021585185080766678, + 0.0060770753771066666, + 0.009672766551375389, + 0.008891900070011616, + 0.006206710357218981, + 0.019572800025343895, + 0.016205018386244774, + 0.015937386080622673, + 0.010442731902003288, + 0.008234167471528053, + 0.011368026956915855, + 0.010942387394607067, + 0.017546653747558594, + 0.010508374311029911, + 0.01004308182746172, + 0.014927292242646217, + 0.015089772641658783, + 0.014461240731179714, + 0.012361247092485428, + 0.016874151304364204, + 0.015630941838026047, + 0.01498098112642765, + 0.01511548925191164, + 0.021450309082865715, + 0.03686409443616867, + 0.03305782005190849, + 0.028047025203704834, + 0.027920063585042953, + 0.022848108783364296, + 0.024975229054689407, + 0.00356479175388813, + 0.0039028762839734554, + 0.004958552774041891, + 0.0035809502005577087, + 0.00412420928478241, + 0.004652899689972401, + 0.003916259855031967, + 0.002249857410788536, + 0.005579675547778606, + 0.011318438686430454, + 0.006872676312923431, + 0.007528210058808327, + 0.006627445574849844, + 0.005184517707675695, + 0.008529708720743656, + 0.004611876793205738, + 0.005251942668110132, + 0.017559295520186424, + 0.010272358544170856, + 0.011248339898884296, + 0.008251012302935123, + 0.006389404647052288, + 0.011918480508029461, + 0.0052702585235238075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.002493607811629772 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634456939995289 + ], + "opposite_logprob_tokens": [ + -3.6513445377349854 + ], + "logprob_reward": -0.02634456939995289, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634456939995289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15480157732963562, + 0.08377321064472198, + 0.10984928905963898, + 0.07529965043067932, + 0.16683140397071838, + 0.0877905860543251, + 0.20825618505477905, + 0.04054190218448639, + 0.04488442838191986, + 0.027971792966127396 + ], + "credit_map": [ + 0.009133566170930862, + 0.01942737214267254, + 0.020435592159628868, + 0.0319230891764164, + 0.012918679974973202, + 0.01956452615559101, + 0.017979519441723824, + 0.023419223725795746, + 0.010323255322873592, + 0.014553366228938103, + 0.012618827633559704, + 0.009721361100673676, + 0.006022852845489979, + 0.0120756970718503, + 0.00784280989319086, + 0.010615040548145771, + 0.007492510601878166, + 0.010713957250118256, + 0.026197658851742744, + 0.007483631372451782, + 0.013114719651639462, + 0.01476237177848816, + 0.006779320538043976, + 0.02330511063337326, + 0.012248302809894085, + 0.008790098130702972, + 0.013912866823375225, + 0.006461482960730791, + 0.005523055791854858, + 0.010194176807999611, + 0.005861029028892517, + 0.012308643199503422, + 0.025483647361397743, + 0.016859034076333046, + 0.018445031717419624, + 0.008814114145934582, + 0.013101944699883461, + 0.019304007291793823, + 0.028964795172214508, + 0.035858847200870514, + 0.008457750082015991, + 0.012596417218446732, + 0.011283794417977333, + 0.009470202028751373, + 0.010651051066815853, + 0.012792132794857025, + 0.0108071593567729, + 0.011732086539268494, + 0.017692673951387405, + 0.017100144177675247, + 0.037774018943309784, + 0.021242041140794754, + 0.01998666115105152, + 0.05184520035982132, + 0.022585909813642502, + 0.020029539242386818, + 0.003586011240258813, + 0.006232148967683315, + 0.005858484655618668, + 0.005456342827528715, + 0.005442233756184578, + 0.005643284879624844, + 0.004342178348451853, + 0.003981216344982386, + 0.005456462502479553, + 0.006905215792357922, + 0.0061051673255860806, + 0.007510548457503319, + 0.003890023799613118, + 0.0057225301861763, + 0.004917628597468138, + 0.004376853816211224, + 0.002699896227568388, + 0.004921544808894396, + 0.004071776755154133, + 0.004418281372636557, + 0.0033654693979769945, + 0.0022878290619701147, + 0.0038412646390497684, + 0.002365728607401252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.02634456939995289 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sharp, shallow depth-of-field close-up focuses squarely on the awed face of a young woman dressed in traditional shrine maiden garb. She stands quietly at the wooden steps of an ancient shrine, her long, flowing, raven-black hair held back by a simple white washi headband, strands brushing against her fair cheeks in the night breeze. The cool moonlight frames her delicate profile. Her head tilted slightly, her lips parted in wonder, her eyes, as clear as deep pools, gaze at the abyss of the night sky. Within these eyes, a magnificent meteor shower is perfectly reflected: countless fleeting, slender, brilliant arcs of silver light streak across her pupils, slightly dilated with concentration, as if the entire galaxy were ablaze in her gaze. Her pristine white Chihaya coat and crimson hakama take on a solemn yet soft quality in the dim light, while the shrine's architectural outlines and hanging stone lanterns behind her are completely blurred, transforming into a dreamy, softly shimmering backdrop composed of warm yellow dots. The entire image, shrouded in profound tranquility, exudes a sacred and ethereal beauty of classical Oriental beauty, capturing a poetic moment of dialogue between humanity and the universe.", + "seed": 5020, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_15/index_4060_seed_5020_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01869109468534589, + "vlm_logprob": -0.01869109468534589, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.01869109468534589, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010485557839274406 + ], + "opposite_logprob_tokens": [ + -5.885485649108887 + ], + "logprob_reward": -0.010485557839274406, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.010485557839274406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10253193229436874, + 0.0506172850728035, + 0.04469596967101097, + 0.051467034965753555, + 0.07533235847949982, + 0.034132055938243866, + 0.3282742202281952, + 0.25162473320961, + 0.03295571729540825, + 0.028368687257170677 + ], + "credit_map": [ + 0.01587088406085968, + 0.012470363639295101, + 0.023942114785313606, + 0.011447501368820667, + 0.01424960047006607, + 0.00780301820486784, + 0.007915823720395565, + 0.008832629770040512, + 0.0065589789301157, + 0.006349907722324133, + 0.013031181879341602, + 0.004812818020582199, + 0.006701044738292694, + 0.005952899344265461, + 0.0038649807684123516, + 0.0033454722724854946, + 0.005740493070334196, + 0.006387924309819937, + 0.005923551972955465, + 0.005783432628959417, + 0.006798374000936747, + 0.004151089116930962, + 0.006943144835531712, + 0.002967960899695754, + 0.008603318594396114, + 0.00645162770524621, + 0.006991745438426733, + 0.00499893631786108, + 0.004649822134524584, + 0.009418352507054806, + 0.005059342365711927, + 0.005293890833854675, + 0.010990677401423454, + 0.008783282712101936, + 0.02462565340101719, + 0.008120830170810223, + 0.005900849122554064, + 0.007507708854973316, + 0.003989755641669035, + 0.005413599777966738, + 0.004716678522527218, + 0.006755623035132885, + 0.006118809338659048, + 0.004263468086719513, + 0.0030181175097823143, + 0.00429883087053895, + 0.0026606665924191475, + 0.0022998605854809284, + 0.05439724773168564, + 0.035058218985795975, + 0.1118624359369278, + 0.04787348583340645, + 0.014667307958006859, + 0.03377818688750267, + 0.014130879193544388, + 0.016506468877196312, + 0.06126341596245766, + 0.018005365505814552, + 0.10397157818078995, + 0.02621074765920639, + 0.007938697934150696, + 0.013435265980660915, + 0.010613827966153622, + 0.010185839608311653, + 0.00426171300932765, + 0.004584027919918299, + 0.004491040017455816, + 0.003955550026148558, + 0.004702060483396053, + 0.003598062554374337, + 0.004425217863172293, + 0.0029380491469055414, + 0.00650310143828392, + 0.003136545652523637, + 0.0038236461114138365, + 0.0025799681898206472, + 0.0037442268803715706, + 0.002795075997710228, + 0.003956796135753393, + 0.0018293255707249045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.010485557839274406 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026366515085101128 + ], + "opposite_logprob_tokens": [ + -3.6513664722442627 + ], + "logprob_reward": -0.026366515085101128, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026366515085101128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13314786553382874, + 0.07008885592222214, + 0.027669619768857956, + 0.019890988245606422, + 0.017517628148198128, + 0.018543262034654617, + 0.34144994616508484, + 0.33525350689888, + 0.017945269122719765, + 0.018493102863430977 + ], + "credit_map": [ + 0.012664251960814, + 0.011318372562527657, + 0.022258302196860313, + 0.01212986744940281, + 0.02206990122795105, + 0.012326431460678577, + 0.029126852750778198, + 0.011253884062170982, + 0.00806240551173687, + 0.006997545249760151, + 0.014860423281788826, + 0.006779742892831564, + 0.011302979663014412, + 0.006032484117895365, + 0.01151968538761139, + 0.004533585626631975, + 0.003627469763159752, + 0.004128494765609503, + 0.003494015894830227, + 0.004193566739559174, + 0.004306078888475895, + 0.0029710354283452034, + 0.0033359802328050137, + 0.0016129766590893269, + 0.0034572617150843143, + 0.0019188852747902274, + 0.0026035578921437263, + 0.0022846367210149765, + 0.0023496216163039207, + 0.0032853574957698584, + 0.0018771591130644083, + 0.002114507369697094, + 0.0024785620626062155, + 0.0019011482363566756, + 0.002358827507123351, + 0.0020668311044573784, + 0.002726396545767784, + 0.0027141482569277287, + 0.0017257138388231397, + 0.001545997685752809, + 0.001470572897233069, + 0.0014610736398026347, + 0.0030726015102118254, + 0.003957462031394243, + 0.0019035683944821358, + 0.0024056502152234316, + 0.002994200913235545, + 0.0012781309196725488, + 0.07726828008890152, + 0.033920884132385254, + 0.11485929787158966, + 0.03061457723379135, + 0.01474760565906763, + 0.041643109172582626, + 0.01413988322019577, + 0.014256276190280914, + 0.07301559299230576, + 0.027303466573357582, + 0.13862842321395874, + 0.038232095539569855, + 0.01199428178369999, + 0.018423844128847122, + 0.012065875343978405, + 0.015589919872581959, + 0.0016036891611292958, + 0.001969186821952462, + 0.0022450315300375223, + 0.0032669876236468554, + 0.002627431880682707, + 0.001731869881041348, + 0.002992125228047371, + 0.0015089462976902723, + 0.001773259835317731, + 0.002405352657660842, + 0.0027331295423209667, + 0.00246360688470304, + 0.0025844282936304808, + 0.0014691492542624474, + 0.0036358789075165987, + 0.0014282953925430775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.026366515085101128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545124627649784 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012545124627649784, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545124627649784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13982807099819183, + 0.020348718389868736, + 0.021620186045765877, + 0.013921697624027729, + 0.01783815585076809, + 0.02616853639483452, + 0.7115700244903564, + 0.025897039100527763, + 0.00933817494660616, + 0.01346934586763382 + ], + "credit_map": [ + 0.026537520810961723, + 0.013097196817398071, + 0.04077262803912163, + 0.012045149691402912, + 0.012241525575518608, + 0.009877406060695648, + 0.014548034407198429, + 0.010708596557378769, + 0.002044543158262968, + 0.0025406901258975267, + 0.003155091777443886, + 0.0027233967557549477, + 0.003941302187740803, + 0.001646446529775858, + 0.002765593584626913, + 0.0015316548524424434, + 0.003090378362685442, + 0.0027936233673244715, + 0.002486816141754389, + 0.0025940346531569958, + 0.004118842072784901, + 0.0015883581945672631, + 0.0034502067137509584, + 0.001497922115959227, + 0.0023613593075424433, + 0.001533913309685886, + 0.0017951222835108638, + 0.0018049010541290045, + 0.0015371204353868961, + 0.0019072545692324638, + 0.0014978719409555197, + 0.0014841535594314337, + 0.002537726191803813, + 0.002177102956920862, + 0.0031369354110211134, + 0.0018493386451154947, + 0.001765173627063632, + 0.0032715233974158764, + 0.0015074623515829444, + 0.001592893386259675, + 0.003205694258213043, + 0.0026017704512923956, + 0.004217392299324274, + 0.0026867021806538105, + 0.0016268164617940784, + 0.007027169223874807, + 0.001810986315831542, + 0.002992004854604602, + 0.10119283199310303, + 0.046287257224321365, + 0.2920469641685486, + 0.10647439956665039, + 0.026893261820077896, + 0.08554963767528534, + 0.020779337733983994, + 0.032346319407224655, + 0.005397787783294916, + 0.0023925674613565207, + 0.00871888268738985, + 0.002626163186505437, + 0.0018933381652459502, + 0.0022480187471956015, + 0.0014665162889286876, + 0.0011537637328729033, + 0.001215654774568975, + 0.0014611980877816677, + 0.0011453278129920363, + 0.0013541218359023333, + 0.001629192614927888, + 0.0007160621462389827, + 0.001060844399034977, + 0.0007557733333669603, + 0.0019010769901797175, + 0.0018102810718119144, + 0.002441280521452427, + 0.0016351278172805905, + 0.0022246791049838066, + 0.0010891483398154378, + 0.0012626287061721087, + 0.0011051221517845988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.012545124627649784 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955924745649099 + ], + "opposite_logprob_tokens": [ + -5.130955696105957 + ], + "logprob_reward": -0.005955924745649099, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955924745649099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07483558356761932, + 0.03934802860021591, + 0.02648819237947464, + 0.03585595637559891, + 0.07305467873811722, + 0.09064684808254242, + 0.577929675579071, + 0.032134536653757095, + 0.019222170114517212, + 0.030484270304441452 + ], + "credit_map": [ + 0.01396920345723629, + 0.008413692004978657, + 0.010117171332240105, + 0.009246435947716236, + 0.012332365848124027, + 0.007538509089499712, + 0.007150835823267698, + 0.006067374721169472, + 0.004200650844722986, + 0.005237423814833164, + 0.008618754334747791, + 0.004304343834519386, + 0.006727981846779585, + 0.004295721650123596, + 0.003401355817914009, + 0.0025617980863898993, + 0.003902345197275281, + 0.004072622396051884, + 0.0034225459676235914, + 0.0031937630847096443, + 0.004499631933867931, + 0.0020155946258455515, + 0.003814704017713666, + 0.001566984341479838, + 0.005656643770635128, + 0.005662214942276478, + 0.005154038313776255, + 0.003694557584822178, + 0.00419299490749836, + 0.006762718316167593, + 0.002001075306907296, + 0.0027317102067172527, + 0.010476755909621716, + 0.007992499507963657, + 0.02630312740802765, + 0.008836418390274048, + 0.003967927768826485, + 0.006885400507599115, + 0.004215883556753397, + 0.004376666154712439, + 0.010063395835459232, + 0.009995011612772942, + 0.00979068037122488, + 0.012818502262234688, + 0.008547287434339523, + 0.02377588488161564, + 0.009878618642687798, + 0.0057774693705141544, + 0.0814538225531578, + 0.037800151854753494, + 0.23582899570465088, + 0.07678377628326416, + 0.022180670872330666, + 0.08028755336999893, + 0.017061561346054077, + 0.02653319574892521, + 0.00810183770954609, + 0.0026702990289777517, + 0.00859841052442789, + 0.0034126807004213333, + 0.002059903461486101, + 0.003355396678671241, + 0.0017132622888311744, + 0.002222746843472123, + 0.0028390379156917334, + 0.0023789098486304283, + 0.0034985679667443037, + 0.002421577228233218, + 0.002259600441902876, + 0.0016335246618837118, + 0.0025828522630035877, + 0.001608103048056364, + 0.004028402268886566, + 0.0030155740678310394, + 0.004860123619437218, + 0.004451490473002195, + 0.005816417280584574, + 0.0022089574486017227, + 0.0036438305396586657, + 0.0024594743736088276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.005955924745649099 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810235112905502 + ], + "opposite_logprob_tokens": [ + -3.288102388381958 + ], + "logprob_reward": -0.03810235112905502, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810235112905502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19834168255329132, + 0.030451279133558273, + 0.018403127789497375, + 0.01861317828297615, + 0.04069282487034798, + 0.049688272178173065, + 0.4562229812145233, + 0.1539202332496643, + 0.013126521371304989, + 0.02053983137011528 + ], + "credit_map": [ + 0.028614427894353867, + 0.024598250165581703, + 0.058044880628585815, + 0.021722545847296715, + 0.018629342317581177, + 0.017276525497436523, + 0.013234012760221958, + 0.016221707686781883, + 0.003167180111631751, + 0.0048913853242993355, + 0.005545474588871002, + 0.0028342106379568577, + 0.005670973565429449, + 0.00291504361666739, + 0.003210705239325762, + 0.0022163051180541515, + 0.002017123391851783, + 0.0027566105127334595, + 0.0037857331335544586, + 0.0020134025253355503, + 0.003135473933070898, + 0.0018237652257084846, + 0.0017018584767356515, + 0.0011691602412611246, + 0.0029356444720178843, + 0.002117475727573037, + 0.002197728492319584, + 0.002204367658123374, + 0.002817252418026328, + 0.001880586496554315, + 0.002877962775528431, + 0.001582160941325128, + 0.005870747845619917, + 0.004506728146225214, + 0.012617487460374832, + 0.004784198943525553, + 0.0027571404352784157, + 0.004113357979804277, + 0.003103237831965089, + 0.0029399257618933916, + 0.006585969589650631, + 0.008350697346031666, + 0.0064255367033183575, + 0.008191470056772232, + 0.004616508726030588, + 0.006237843539565802, + 0.005592511035501957, + 0.003687736112624407, + 0.08954028785228729, + 0.043259404599666595, + 0.16033251583576202, + 0.05434706434607506, + 0.02442195452749729, + 0.03918973729014397, + 0.02006145752966404, + 0.025070568546652794, + 0.038030754774808884, + 0.010563756339251995, + 0.06392071396112442, + 0.016333751380443573, + 0.004263230599462986, + 0.008908411487936974, + 0.004837168846279383, + 0.007062445394694805, + 0.002511722967028618, + 0.0022267582826316357, + 0.001883133896626532, + 0.0011783677618950605, + 0.001626635086722672, + 0.0010424180654808879, + 0.0013501105131581426, + 0.001307375030592084, + 0.0029151716735213995, + 0.0025086617097258568, + 0.0027459219563752413, + 0.0025379196740686893, + 0.003048422047868371, + 0.002017136663198471, + 0.00246881484054029, + 0.0022977814078330994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.03810235112905502 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sharp, shallow depth-of-field close-up focuses squarely on the awed face of a young woman dressed in traditional shrine maiden garb. She stands quietly at the wooden steps of an ancient shrine, her long, flowing, raven-black hair held back by a simple white washi headband, strands brushing against her fair cheeks in the night breeze. The cool moonlight frames her delicate profile. Her head tilted slightly, her lips parted in wonder, her eyes, as clear as deep pools, gaze at the abyss of the night sky. Within these eyes, a magnificent meteor shower is perfectly reflected: countless fleeting, slender, brilliant arcs of silver light streak across her pupils, slightly dilated with concentration, as if the entire galaxy were ablaze in her gaze. Her pristine white Chihaya coat and crimson hakama take on a solemn yet soft quality in the dim light, while the shrine's architectural outlines and hanging stone lanterns behind her are completely blurred, transforming into a dreamy, softly shimmering backdrop composed of warm yellow dots. The entire image, shrouded in profound tranquility, exudes a sacred and ethereal beauty of classical Oriental beauty, capturing a poetic moment of dialogue between humanity and the universe.", + "seed": 5021, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_15/index_4060_seed_5021_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02985905148088932, + "vlm_logprob": -0.02985905148088932, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.02985905148088932, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01732899807393551 + ], + "opposite_logprob_tokens": [ + -4.767329216003418 + ], + "logprob_reward": -0.01732899807393551, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01732899807393551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14024503529071808, + 0.1824580729007721, + 0.042612191289663315, + 0.035746559500694275, + 0.09593087434768677, + 0.09183935821056366, + 0.03639029711484909, + 0.04752584919333458, + 0.09421499818563461, + 0.2330368012189865 + ], + "credit_map": [ + 0.01855410262942314, + 0.015755096450448036, + 0.02836325392127037, + 0.00944166723638773, + 0.017158934846520424, + 0.02079133875668049, + 0.013787118718028069, + 0.01639353111386299, + 0.015230772085487843, + 0.010719209909439087, + 0.01639072597026825, + 0.00975336879491806, + 0.02318676747381687, + 0.03502188250422478, + 0.038761742413043976, + 0.03339359536767006, + 0.004054712597280741, + 0.004871475975960493, + 0.0061309137381613255, + 0.004222510848194361, + 0.005911312997341156, + 0.006054600700736046, + 0.005188279785215855, + 0.0061783865094184875, + 0.004194342996925116, + 0.0035353931598365307, + 0.0048001548275351524, + 0.004631740972399712, + 0.00447158282622695, + 0.005103814881294966, + 0.004834678489714861, + 0.0041748518124222755, + 0.014753847382962704, + 0.011214441619813442, + 0.009409427642822266, + 0.00802349392324686, + 0.020520363003015518, + 0.010834443382918835, + 0.008651511743664742, + 0.012523346580564976, + 0.009655093774199486, + 0.0035389973782002926, + 0.004149931017309427, + 0.005598928313702345, + 0.007569415029138327, + 0.031322870403528214, + 0.006608958821743727, + 0.023395169526338577, + 0.0026630989741533995, + 0.0038239837158471346, + 0.003338774899020791, + 0.004064591601490974, + 0.007932869717478752, + 0.0053855059668421745, + 0.003907158970832825, + 0.0052743093110620975, + 0.005936446599662304, + 0.004134799353778362, + 0.007066169288009405, + 0.004306926857680082, + 0.006739157252013683, + 0.008424374274909496, + 0.0054715219885110855, + 0.005446454510092735, + 0.01159195601940155, + 0.010847581550478935, + 0.00841216929256916, + 0.006132944021373987, + 0.02300761453807354, + 0.012860599905252457, + 0.011227373965084553, + 0.010134763084352016, + 0.02178158611059189, + 0.022138983011245728, + 0.028504246845841408, + 0.029851479455828667, + 0.044814713299274445, + 0.022045137360692024, + 0.025527583435177803, + 0.03837306797504425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.01732899807393551 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499017611145973 + ], + "opposite_logprob_tokens": [ + -2.929990291595459 + ], + "logprob_reward": -0.05499017611145973, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499017611145973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19376373291015625, + 0.04672667756676674, + 0.056572075933218, + 0.11590253561735153, + 0.12205657362937927, + 0.11994624882936478, + 0.05211123451590538, + 0.030750328674912453, + 0.07084403187036514, + 0.19132663309574127 + ], + "credit_map": [ + 0.029481416568160057, + 0.02249489165842533, + 0.028307398781180382, + 0.01417575217783451, + 0.03136204183101654, + 0.02039165236055851, + 0.01770281419157982, + 0.029847748577594757, + 0.006209642626345158, + 0.005268674809485674, + 0.004425733350217342, + 0.003789302194491029, + 0.008908379822969437, + 0.0062493737787008286, + 0.005776202771812677, + 0.006099361926317215, + 0.006694029550999403, + 0.00654330151155591, + 0.00930524617433548, + 0.006507869344204664, + 0.005991189740598202, + 0.006367743480950594, + 0.006946271285414696, + 0.00821642018854618, + 0.012854655273258686, + 0.014157080091536045, + 0.01253432221710682, + 0.009769824333488941, + 0.01259511150419712, + 0.024444375187158585, + 0.012893821112811565, + 0.016653330996632576, + 0.014449390582740307, + 0.01620844379067421, + 0.011829927563667297, + 0.011067033745348454, + 0.019038323312997818, + 0.015603533945977688, + 0.01491584349423647, + 0.01894408091902733, + 0.010966517962515354, + 0.011502090841531754, + 0.012216843664646149, + 0.009292836301028728, + 0.010838274843990803, + 0.023644227534532547, + 0.009539609774947166, + 0.031945839524269104, + 0.004206592217087746, + 0.00787342619150877, + 0.004269744735211134, + 0.006310656201094389, + 0.011052974499762058, + 0.007194876205176115, + 0.005092156585305929, + 0.006110806483775377, + 0.002833399921655655, + 0.0033580888994038105, + 0.0027701235376298428, + 0.004367148969322443, + 0.004656264558434486, + 0.0040621645748615265, + 0.0039056839887052774, + 0.004797448869794607, + 0.004555773455649614, + 0.008882511407136917, + 0.007058217190206051, + 0.009620590135455132, + 0.012671822682023048, + 0.00837826170027256, + 0.011464955285191536, + 0.008211891166865826, + 0.011367052793502808, + 0.028535855934023857, + 0.030447155237197876, + 0.03706110641360283, + 0.017489369958639145, + 0.014694919809699059, + 0.023225562646985054, + 0.028505589812994003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.05499017611145973 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029807347804307938 + ], + "opposite_logprob_tokens": [ + -3.5298073291778564 + ], + "logprob_reward": -0.029807347804307938, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807347804307938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22121292352676392, + 0.13821850717067719, + 0.05896560847759247, + 0.12275960296392441, + 0.12745323777198792, + 0.03196974843740463, + 0.04339922219514847, + 0.022988663986325264, + 0.16042594611644745, + 0.07260651141405106 + ], + "credit_map": [ + 0.028391307219862938, + 0.023065706714987755, + 0.03264100104570389, + 0.013590618036687374, + 0.030198704451322556, + 0.024399960413575172, + 0.020382625982165337, + 0.048542991280555725, + 0.02063434012234211, + 0.012191176414489746, + 0.014515919610857964, + 0.0075056604109704494, + 0.01125209592282772, + 0.029918856918811798, + 0.015769628807902336, + 0.026430832222104073, + 0.007329064421355724, + 0.0074760690331459045, + 0.007044939789921045, + 0.008089784532785416, + 0.007208990398794413, + 0.008995503187179565, + 0.006050049792975187, + 0.006771203596144915, + 0.016318973153829575, + 0.011527689173817635, + 0.015181341208517551, + 0.013031481765210629, + 0.020517785102128983, + 0.01725279726088047, + 0.013225761242210865, + 0.01570376381278038, + 0.019303366541862488, + 0.021632911637425423, + 0.012091717682778835, + 0.01431113388389349, + 0.01678762026131153, + 0.014766310341656208, + 0.010720647871494293, + 0.01783953234553337, + 0.0042060501873493195, + 0.004569771233946085, + 0.0030453454237431288, + 0.004031964112073183, + 0.004589007701724768, + 0.004386201500892639, + 0.003341933246701956, + 0.0037994745653122663, + 0.0042923828586936, + 0.005045273806899786, + 0.0039609940722584724, + 0.006199512630701065, + 0.007280503865331411, + 0.006308060605078936, + 0.005156660918146372, + 0.0051558371633291245, + 0.002377638127654791, + 0.002015533624216914, + 0.0029598418623209, + 0.0030261091887950897, + 0.0038286170456558466, + 0.003736992832273245, + 0.0023595725651830435, + 0.0026843587402254343, + 0.01422259584069252, + 0.019716178998351097, + 0.01254518423229456, + 0.011228862218558788, + 0.04421541467308998, + 0.019967179745435715, + 0.01972036622464657, + 0.018810179084539413, + 0.006411567330360413, + 0.007588699460029602, + 0.009364010766148567, + 0.012284735217690468, + 0.007831785827875137, + 0.008053486235439777, + 0.009385760873556137, + 0.011686470359563828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.029807347804307938 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409882515668869 + ], + "opposite_logprob_tokens": [ + -5.504098892211914 + ], + "logprob_reward": -0.00409882515668869, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409882515668869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10011143237352371, + 0.0958443433046341, + 0.0888240784406662, + 0.12588535249233246, + 0.057079050689935684, + 0.07644706964492798, + 0.06606005132198334, + 0.07202352583408356, + 0.24004794657230377, + 0.07767712324857712 + ], + "credit_map": [ + 0.013019584119319916, + 0.009180640801787376, + 0.014099590480327606, + 0.0074983024969697, + 0.017340589314699173, + 0.012242093682289124, + 0.011901493184268475, + 0.014829139225184917, + 0.007283409126102924, + 0.0054862177930772305, + 0.007673036772757769, + 0.0076872180216014385, + 0.016361825168132782, + 0.015726016834378242, + 0.02144918031990528, + 0.014177442528307438, + 0.009445684030652046, + 0.00932781770825386, + 0.01521350722759962, + 0.011356362141668797, + 0.010718616656959057, + 0.01168186217546463, + 0.010048828087747097, + 0.01103140413761139, + 0.011033865623176098, + 0.01049075648188591, + 0.013756445609033108, + 0.013499335385859013, + 0.01893443614244461, + 0.018063921481370926, + 0.025369539856910706, + 0.014737045392394066, + 0.007012012414634228, + 0.0069972495548427105, + 0.004699474200606346, + 0.004920714534819126, + 0.010733108967542648, + 0.0071098958142101765, + 0.005596107803285122, + 0.010010486468672752, + 0.007522860076278448, + 0.004014322068542242, + 0.0068437461741268635, + 0.005923241376876831, + 0.009117442183196545, + 0.015937652438879013, + 0.0066639394499361515, + 0.02042386680841446, + 0.005517447367310524, + 0.009193568490445614, + 0.00641280971467495, + 0.008276480250060558, + 0.01026524044573307, + 0.009345768019557, + 0.006482096388936043, + 0.01056663878262043, + 0.009613546542823315, + 0.005090240854769945, + 0.011874979361891747, + 0.0049599697813391685, + 0.010178899392485619, + 0.014267811551690102, + 0.007184252608567476, + 0.008853824809193611, + 0.047944873571395874, + 0.0193144790828228, + 0.01838916912674904, + 0.013633004389703274, + 0.04909504950046539, + 0.03993353247642517, + 0.02197982370853424, + 0.029758000746369362, + 0.006473881658166647, + 0.00864200945943594, + 0.011179964058101177, + 0.012833159416913986, + 0.007845116779208183, + 0.009032683447003365, + 0.007868604734539986, + 0.013801704160869122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.00409882515668869 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp, shallow depth-of-field close-up focusing on a young woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing traditional shrine maiden garb consisting of a white Chihaya coat and crimson hakama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long, raven-black hair held back by a simple white washi headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background elements, including the shrine steps and stone lanterns, completely blurred into a backdrop of warm yellow dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant arcs of silver light reflecting as a meteor shower within the woman's eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306991025805473 + ], + "opposite_logprob_tokens": [ + -3.168069839477539 + ], + "logprob_reward": -0.04306991025805473, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306991025805473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1306210160255432, + 0.15201853215694427, + 0.05047433823347092, + 0.09880998730659485, + 0.08632461726665497, + 0.11105503886938095, + 0.04734091833233833, + 0.041844457387924194, + 0.1771158128976822, + 0.10439523309469223 + ], + "credit_map": [ + 0.019065173342823982, + 0.013018359430134296, + 0.021704239770770073, + 0.008286163210868835, + 0.014326617121696472, + 0.01996411196887493, + 0.014513283967971802, + 0.019743064418435097, + 0.011304412968456745, + 0.01136802788823843, + 0.012087271548807621, + 0.009143617935478687, + 0.016549957916140556, + 0.03142472356557846, + 0.028869597241282463, + 0.031270936131477356, + 0.005750034470111132, + 0.006702627055346966, + 0.006236228626221418, + 0.007512805052101612, + 0.006103991996496916, + 0.007696932647377253, + 0.005000176373869181, + 0.0054715415462851524, + 0.012338245287537575, + 0.013341160491108894, + 0.014401904307305813, + 0.012462007813155651, + 0.010272514075040817, + 0.011741863563656807, + 0.013154803775250912, + 0.011097489856183529, + 0.015668608248233795, + 0.00795450434088707, + 0.010242010466754436, + 0.00651285145431757, + 0.01260328758507967, + 0.010886622592806816, + 0.009227722883224487, + 0.013229011557996273, + 0.010827778838574886, + 0.005842179991304874, + 0.011114767752587795, + 0.009469347074627876, + 0.013203775510191917, + 0.02198931574821472, + 0.010049725882709026, + 0.02855815552175045, + 0.005121820606291294, + 0.00638611102476716, + 0.005058744456619024, + 0.006100095342844725, + 0.007336783688515425, + 0.006379259284585714, + 0.00417027622461319, + 0.006787827238440514, + 0.004297683481127024, + 0.004156168084591627, + 0.004483974073082209, + 0.00457123713567853, + 0.00661379424855113, + 0.0074576144106686115, + 0.005805708467960358, + 0.0044582802802324295, + 0.020427407696843147, + 0.018449360504746437, + 0.017989277839660645, + 0.012993201613426208, + 0.041453856974840164, + 0.02096705138683319, + 0.025176936760544777, + 0.019658716395497322, + 0.008185259066522121, + 0.010968716815114021, + 0.016702568158507347, + 0.016751274466514587, + 0.009526874870061874, + 0.013315163552761078, + 0.011566943489015102, + 0.0173784326761961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.04306991025805473 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_2/index_tensor([5772, 5772])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_2/index_tensor([5772, 5772])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ff6ffe7ab81b5d7c3e2c044f55fc5d8d0349c43b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_2/index_tensor([5772, 5772])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The scene unfolds in a medium close-up, with the camera beginning with a slow, deliberate push-in on a lone white candle, its candle showing delicate drips of cooling wax as it serves as the sole source of illumination in the pitch-black room. Its single flame trembles erratically, a teardrop of vibrant orange and yellow flame, its core a core of intense, incandescent light. This wild, unpredictable flickering casts vast, dancing shadows that wriggle and twist across the richly textured dark surfaces of the wooden table and the near background. As the camera inexorably approaches, focus settles on the flame itself, its ethereal, smoky tip licking at the oppressive darkness. The push-in continues, moving through the body of the flame, dissolving into an extreme close-up of the blackened fiber wick, its tip emitting a bright, incandescent red-orange flash that gently pulsates with heat, creating an atmosphere of profound solitude and concentration, where warm, dynamic light engages in a silent, brutal battle against a deep, enveloping darkness.", + "seed": 9759, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_2/index_5772_seed_9759_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05521104054059833, + "vlm_logprob": -0.05521104054059833, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.05521104054059833, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010121921077370644 + ], + "opposite_logprob_tokens": [ + -4.885121822357178 + ], + "logprob_reward": -0.010121921077370644, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010121921077370644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027556760236620903, + 0.15144507586956024, + 0.029605869203805923, + 0.030847623944282532, + 0.582375168800354, + 0.04023214429616928, + 0.04062807187438011, + 0.025792233645915985, + 0.04737124964594841, + 0.024145716801285744 + ], + "credit_map": [ + 0.002024295972660184, + 0.005398990586400032, + 0.0046111103147268295, + 0.003616570495069027, + 0.0020341393537819386, + 0.004231014754623175, + 0.0023592987563461065, + 0.003281342564150691, + 0.007658112794160843, + 0.004168433602899313, + 0.05993717908859253, + 0.00487235002219677, + 0.0030332275200635195, + 0.0591166727244854, + 0.0023180702701210976, + 0.010341045446693897, + 0.002205979311838746, + 0.0020049549639225006, + 0.009460138157010078, + 0.0019189226441085339, + 0.0009648089180700481, + 0.008889730088412762, + 0.0011209551012143493, + 0.0030403819400817156, + 0.0018264586105942726, + 0.003392178798094392, + 0.009591283276677132, + 0.0026937369257211685, + 0.0006734810885973275, + 0.009965327568352222, + 0.0009552289848215878, + 0.0017499313689768314, + 0.03833562135696411, + 0.012641740031540394, + 0.24295471608638763, + 0.02317993715405464, + 0.0070722452364861965, + 0.23625043034553528, + 0.008888526819646358, + 0.013052034191787243, + 0.00336679769679904, + 0.004618469160050154, + 0.00970090925693512, + 0.0038848521653562784, + 0.0013199326349422336, + 0.010445092804729939, + 0.0019304255256429315, + 0.0049656713381409645, + 0.0037871592212468386, + 0.005706569645553827, + 0.006219254340976477, + 0.00705921184271574, + 0.0017787989927455783, + 0.0037947315722703934, + 0.002997991628944874, + 0.009284360334277153, + 0.004592780955135822, + 0.002393964910879731, + 0.007723830174654722, + 0.0026436378248035908, + 0.000977915246039629, + 0.00217811344191432, + 0.001043054391629994, + 0.004238939844071865, + 0.00807318277657032, + 0.00361101352609694, + 0.013832254335284233, + 0.004540721885859966, + 0.002061043167486787, + 0.00377675355412066, + 0.002045146655291319, + 0.009431136772036552, + 0.0054735722951591015, + 0.001320323208346963, + 0.007905946113169193, + 0.0015441764844581485, + 0.0012942267348989844, + 0.0018364667193964124, + 0.0014033286133781075, + 0.003367679426446557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 8 + ], + "selected_reward": -0.010121921077370644 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590949803590775 + ], + "opposite_logprob_tokens": [ + -3.8955910205841064 + ], + "logprob_reward": -0.020590949803590775, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590949803590775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02728881873190403, + 0.15254051983356476, + 0.05487691983580589, + 0.10845424979925156, + 0.4609687626361847, + 0.034810010343790054, + 0.057529743760824203, + 0.07721620053052902, + 0.01603441871702671, + 0.010280375368893147 + ], + "credit_map": [ + 0.0018191190902143717, + 0.002995941787958145, + 0.006224885117262602, + 0.0044362954795360565, + 0.001193093485198915, + 0.005221869330853224, + 0.0013942891964688897, + 0.004003323148936033, + 0.007477546576410532, + 0.004596681799739599, + 0.059919483959674835, + 0.006485098972916603, + 0.0030500441789627075, + 0.06046874076128006, + 0.0027052524965256453, + 0.007837655022740364, + 0.004292586352676153, + 0.01298031210899353, + 0.0109309833496809, + 0.00805013906210661, + 0.002443076344206929, + 0.008496441878378391, + 0.002069350564852357, + 0.005614022724330425, + 0.005329834762960672, + 0.004985028877854347, + 0.041208427399396896, + 0.005601579323410988, + 0.0015359618701040745, + 0.04205666109919548, + 0.0020634245593100786, + 0.005673314444720745, + 0.03066728450357914, + 0.011472955346107483, + 0.19212506711483002, + 0.01936635747551918, + 0.005762362852692604, + 0.18404309451580048, + 0.0069174254313111305, + 0.010614176280796528, + 0.004263431299477816, + 0.005303055979311466, + 0.00589725049212575, + 0.004136383067816496, + 0.0016334664542227983, + 0.0054487199522554874, + 0.002328249393031001, + 0.005799451377242804, + 0.005563241429626942, + 0.008710728026926517, + 0.0088387131690979, + 0.007095489650964737, + 0.003728016745299101, + 0.005663855001330376, + 0.0034947101958096027, + 0.014434983022511005, + 0.008252384141087532, + 0.009286854416131973, + 0.013862733729183674, + 0.014256205409765244, + 0.005091328639537096, + 0.0061919656582176685, + 0.005896601360291243, + 0.014378117397427559, + 0.002206280827522278, + 0.0019024488283321261, + 0.003373975632712245, + 0.0021546746138483286, + 0.0009215091704390943, + 0.0012633013539016247, + 0.0010851254919543862, + 0.0031271015759557486, + 0.002261051442474127, + 0.0006952676922082901, + 0.003113917075097561, + 0.0011465022107586265, + 0.000818652450107038, + 0.000657290976960212, + 0.0007908012485131621, + 0.0007968907011672854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.020590949803590775 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22564192116260529 + ], + "opposite_logprob_tokens": [ + -1.6006419658660889 + ], + "logprob_reward": -0.22564192116260529, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22564192116260529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055443137884140015, + 0.01582086831331253, + 0.01715635508298874, + 0.06127959489822388, + 0.7759563326835632, + 0.012976118363440037, + 0.00904874224215746, + 0.03179530054330826, + 0.016279973089694977, + 0.004243641626089811 + ], + "credit_map": [ + 0.0038929630536586046, + 0.006108451168984175, + 0.014529330655932426, + 0.00656194007024169, + 0.0024229090195149183, + 0.013387473300099373, + 0.002539323642849922, + 0.006000740919262171, + 0.0010320213623344898, + 0.001289067673496902, + 0.005139148328453302, + 0.0015702750533819199, + 0.00040877488208934665, + 0.005109146237373352, + 0.0003971039841417223, + 0.000875328725669533, + 0.0010112968739122152, + 0.0012490127701312304, + 0.006257497239857912, + 0.000846722221467644, + 0.00034677007352001965, + 0.005678809247910976, + 0.0003121198096778244, + 0.0014541265554726124, + 0.0035522987600415945, + 0.003942180424928665, + 0.02214522287249565, + 0.003674355335533619, + 0.0010559498332440853, + 0.02092369832098484, + 0.0012764339335262775, + 0.004709452390670776, + 0.05161837115883827, + 0.017659064382314682, + 0.3240252435207367, + 0.031465042382478714, + 0.009626436047255993, + 0.31244024634361267, + 0.011966544203460217, + 0.01715528592467308, + 0.0017510791076347232, + 0.0017764949006959796, + 0.0020245250780135393, + 0.001792952069081366, + 0.0006887337076477706, + 0.0014880988746881485, + 0.0011328932596370578, + 0.002321339678019285, + 0.0010171723552048206, + 0.0015412024222314358, + 0.001531377900391817, + 0.0014833505265414715, + 0.0005028228624723852, + 0.0007607025327160954, + 0.0007555433548986912, + 0.0014565690653398633, + 0.003479065140709281, + 0.0036237307358533144, + 0.0054357112385332584, + 0.005339544266462326, + 0.0017822531517595053, + 0.0026149863842874765, + 0.002033322351053357, + 0.0074866837821900845, + 0.002394568407908082, + 0.0012889050412923098, + 0.004185643047094345, + 0.001750434865243733, + 0.0008650270174257457, + 0.0013309119967743754, + 0.0007038303883746266, + 0.003760650521144271, + 0.0009071808308362961, + 0.00040310961776413023, + 0.0009496656712144613, + 0.0003042909374926239, + 0.00023975854855962098, + 0.0003354069485794753, + 0.00021572371770162135, + 0.0008885050192475319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.22564192116260529 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015163839561864734 + ], + "opposite_logprob_tokens": [ + -6.501516342163086 + ], + "logprob_reward": -0.0015163839561864734, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015163839561864734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09067104011774063, + 0.041358817368745804, + 0.1701923906803131, + 0.08672837913036346, + 0.3711268901824951, + 0.061969321221113205, + 0.0649123564362526, + 0.02256598137319088, + 0.0702122375369072, + 0.020262550562620163 + ], + "credit_map": [ + 0.008886388503015041, + 0.013492023572325706, + 0.016121484339237213, + 0.014909773133695126, + 0.005844536703079939, + 0.01342121884226799, + 0.007525821682065725, + 0.010469787754118443, + 0.002368345856666565, + 0.0036648246459662914, + 0.012005295604467392, + 0.0033767574932426214, + 0.0020975754596292973, + 0.012306447140872478, + 0.002298449631780386, + 0.003241121768951416, + 0.009290352463722229, + 0.024698615074157715, + 0.04979267343878746, + 0.01747443526983261, + 0.00462439376860857, + 0.04649730026721954, + 0.004995697643607855, + 0.012818916700780392, + 0.004645194858312607, + 0.0046629090793430805, + 0.030624421313405037, + 0.004725698847323656, + 0.001646992634050548, + 0.03259798884391785, + 0.0021037699189037085, + 0.005721404682844877, + 0.024430440738797188, + 0.008617411367595196, + 0.15005798637866974, + 0.015792211517691612, + 0.005016214679926634, + 0.15199005603790283, + 0.005911419168114662, + 0.009311164729297161, + 0.005721655208617449, + 0.00952958595007658, + 0.01168869249522686, + 0.007933752611279488, + 0.0021952087990939617, + 0.012379460036754608, + 0.0032882061786949635, + 0.009232756681740284, + 0.006864939816296101, + 0.009768744930624962, + 0.00993337295949459, + 0.010092314332723618, + 0.0038127547595649958, + 0.00561405997723341, + 0.0044535081833601, + 0.014372667297720909, + 0.0027828598394989967, + 0.002781455172225833, + 0.0033553631510585546, + 0.00248932046815753, + 0.001455843448638916, + 0.0019354835385456681, + 0.00155170657671988, + 0.00621394906193018, + 0.008983444422483444, + 0.005533198360353708, + 0.014575249515473843, + 0.007759607397019863, + 0.0037396016996353865, + 0.006244543474167585, + 0.003040055278688669, + 0.020336534827947617, + 0.003731201635673642, + 0.00223613646812737, + 0.004186528269201517, + 0.0021471779327839613, + 0.0022547319531440735, + 0.0014566368190571666, + 0.0012241726508364081, + 0.0030259648337960243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.0015163839561864734 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184026703238487 + ], + "opposite_logprob_tokens": [ + -4.018184185028076 + ], + "logprob_reward": -0.018184026703238487, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184026703238487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03235297650098801, + 0.02773241512477398, + 0.00949664693325758, + 0.02000725455582142, + 0.8141038417816162, + 0.022110844030976295, + 0.03806380555033684, + 0.01493115909397602, + 0.011074564419686794, + 0.010126586072146893 + ], + "credit_map": [ + 0.001945733092725277, + 0.002678799210116267, + 0.009369412437081337, + 0.0019303008448332548, + 0.0012554693967103958, + 0.009146793745458126, + 0.0014523270074278116, + 0.0045741405338048935, + 0.001367388991639018, + 0.0014238766161724925, + 0.010025030933320522, + 0.001800279482267797, + 0.0007344974437728524, + 0.009955108165740967, + 0.0007338331779465079, + 0.0016924007795751095, + 0.0007031919085420668, + 0.0017205564072355628, + 0.0018085315823554993, + 0.0009990575490519404, + 0.00041381875053048134, + 0.0015469164354726672, + 0.0003748388262465596, + 0.0019297357648611069, + 0.0011098366230726242, + 0.0012578672030940652, + 0.0069828643463552, + 0.0011234440607950091, + 0.00032453425228595734, + 0.006640132516622543, + 0.00046158721670508385, + 0.0021069892682135105, + 0.05463552474975586, + 0.017766527831554413, + 0.3399330973625183, + 0.032740361988544464, + 0.009939001873135567, + 0.3279809057712555, + 0.012412117794156075, + 0.018696319311857224, + 0.002062269253656268, + 0.003402703907340765, + 0.005513007286936045, + 0.0027986750937998295, + 0.0008590108482167125, + 0.00372383208014071, + 0.000997638562694192, + 0.002753708977252245, + 0.004432023037225008, + 0.006796861067414284, + 0.006525976117700338, + 0.006904411129653454, + 0.0020035100169479847, + 0.0030112259555608034, + 0.003588411957025528, + 0.004801381845027208, + 0.0020878538489341736, + 0.0013976171612739563, + 0.003716690931469202, + 0.0014951634220778942, + 0.0010165553539991379, + 0.0014159965794533491, + 0.0010388311930000782, + 0.002762450836598873, + 0.0019685321021825075, + 0.0008371347212232649, + 0.003002840094268322, + 0.0009549404494464397, + 0.0006223049131222069, + 0.001037128153257072, + 0.0005107199540361762, + 0.002140963450074196, + 0.0019152802415192127, + 0.000594245851971209, + 0.003375864587724209, + 0.0008175417315214872, + 0.0006674070027656853, + 0.0006408351473510265, + 0.0007872036658227444, + 0.0013282076688483357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.018184026703238487 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene unfolds in a medium close-up, with the camera beginning with a slow, deliberate push-in on a lone white candle, its candle showing delicate drips of cooling wax as it serves as the sole source of illumination in the pitch-black room. Its single flame trembles erratically, a teardrop of vibrant orange and yellow flame, its core a core of intense, incandescent light. This wild, unpredictable flickering casts vast, dancing shadows that wriggle and twist across the richly textured dark surfaces of the wooden table and the near background. As the camera inexorably approaches, focus settles on the flame itself, its ethereal, smoky tip licking at the oppressive darkness. The push-in continues, moving through the body of the flame, dissolving into an extreme close-up of the blackened fiber wick, its tip emitting a bright, incandescent red-orange flash that gently pulsates with heat, creating an atmosphere of profound solitude and concentration, where warm, dynamic light engages in a silent, brutal battle against a deep, enveloping darkness.", + "seed": 9760, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_2/index_5772_seed_9760_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03553556150291115, + "vlm_logprob": -0.03553556150291115, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.03553556150291115, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012611988931894302 + ], + "opposite_logprob_tokens": [ + -4.6376118659973145 + ], + "logprob_reward": -0.012611988931894302, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.012611988931894302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.079103484749794, + 0.028014512732625008, + 0.030553139746189117, + 0.03128824010491371, + 0.038997750729322433, + 0.018378138542175293, + 0.042360030114650726, + 0.48794057965278625, + 0.0659007653594017, + 0.1774633675813675 + ], + "credit_map": [ + 0.009896956384181976, + 0.006309180986136198, + 0.013790776021778584, + 0.007658333517611027, + 0.005342308897525072, + 0.010501233860850334, + 0.014945928007364273, + 0.010658771730959415, + 0.00323501811362803, + 0.0015756398206576705, + 0.004003110807389021, + 0.00539372768253088, + 0.0016138992505148053, + 0.0033680845517665148, + 0.0053957486525177956, + 0.00342928315512836, + 0.0028066805098205805, + 0.0024889588821679354, + 0.005173278972506523, + 0.004085238557308912, + 0.0028274166397750378, + 0.003743899753317237, + 0.004705207422375679, + 0.004722459241747856, + 0.0030252139549702406, + 0.002174850320443511, + 0.0073154098354279995, + 0.005141646135598421, + 0.0017991196364164352, + 0.0038451587315648794, + 0.003914720844477415, + 0.004072121344506741, + 0.0028061429038643837, + 0.0037184564862400293, + 0.007645699195563793, + 0.005381474271416664, + 0.002960094017907977, + 0.004795810207724571, + 0.006013702135533094, + 0.005676370579749346, + 0.0017313776770606637, + 0.0014112225035205483, + 0.0029780189506709576, + 0.0032900692895054817, + 0.0013685065787285566, + 0.0025374609977006912, + 0.0023818144109100103, + 0.002679667668417096, + 0.00347866490483284, + 0.00435974495485425, + 0.005936922039836645, + 0.002992900088429451, + 0.0037336971145123243, + 0.009076339192688465, + 0.0028415548149496317, + 0.00994020514190197, + 0.05592825263738632, + 0.07251691073179245, + 0.07009004056453705, + 0.05415971577167511, + 0.06188542768359184, + 0.06418803334236145, + 0.03548828139901161, + 0.07368392497301102, + 0.009710229001939297, + 0.0053551639430224895, + 0.011012986302375793, + 0.006541107315570116, + 0.006345118395984173, + 0.00953628495335579, + 0.008757631294429302, + 0.008642246015369892, + 0.02802198752760887, + 0.009731217287480831, + 0.03336052969098091, + 0.016039688140153885, + 0.01607971079647541, + 0.02187228947877884, + 0.023969128727912903, + 0.028388826176524162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.012611988931894302 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002926617860794 + ], + "opposite_logprob_tokens": [ + -2.6950292587280273 + ], + "logprob_reward": -0.07002926617860794, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002926617860794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10738231241703033, + 0.07213813066482544, + 0.10076289623975754, + 0.07991147041320801, + 0.07405005395412445, + 0.07201527804136276, + 0.1130826324224472, + 0.12966148555278778, + 0.20034071803092957, + 0.05065499246120453 + ], + "credit_map": [ + 0.013921364210546017, + 0.013887997716665268, + 0.012121323496103287, + 0.012859683483839035, + 0.008407478220760822, + 0.014310739003121853, + 0.016455426812171936, + 0.015418305061757565, + 0.008983949199318886, + 0.004815629217773676, + 0.011272274889051914, + 0.012027524411678314, + 0.004693589638918638, + 0.009454307146370411, + 0.009186764247715473, + 0.01170408632606268, + 0.007857649587094784, + 0.00648857606574893, + 0.013523620553314686, + 0.01476272102445364, + 0.008374757133424282, + 0.012587260454893112, + 0.024251725524663925, + 0.01291658915579319, + 0.008092235773801804, + 0.004386047832667828, + 0.019148586317896843, + 0.012577305547893047, + 0.004613728262484074, + 0.010170193389058113, + 0.010762957856059074, + 0.010160421952605247, + 0.0069880494847893715, + 0.008291991427540779, + 0.014298184774816036, + 0.01157474610954523, + 0.005945471115410328, + 0.008424900472164154, + 0.009502273052930832, + 0.00902444776147604, + 0.007052387576550245, + 0.005505211651325226, + 0.015496032312512398, + 0.015690932050347328, + 0.003401557682082057, + 0.008847995661199093, + 0.006140868179500103, + 0.00988029595464468, + 0.012611418962478638, + 0.010726875625550747, + 0.025724271312355995, + 0.009627344086766243, + 0.008375326171517372, + 0.017959700897336006, + 0.006888467352837324, + 0.02116922102868557, + 0.01529686339199543, + 0.020215213298797607, + 0.01678563468158245, + 0.013485956937074661, + 0.017418593168258667, + 0.01476797740906477, + 0.012332296930253506, + 0.019358955323696136, + 0.02187097817659378, + 0.01669633574783802, + 0.019642779603600502, + 0.02283875085413456, + 0.01572442054748535, + 0.03492457792162895, + 0.02801329456269741, + 0.04062957316637039, + 0.007548066321760416, + 0.0037928400561213493, + 0.0097691984847188, + 0.0028740011621266603, + 0.004049432929605246, + 0.008770184591412544, + 0.002629173221066594, + 0.011222093366086483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.07002926617860794 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499401316046715 + ], + "opposite_logprob_tokens": [ + -2.9299941062927246 + ], + "logprob_reward": -0.05499401316046715, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499401316046715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10000060498714447, + 0.2192295789718628, + 0.04234405234456062, + 0.05526765435934067, + 0.0671427920460701, + 0.055322885513305664, + 0.07577721774578094, + 0.19348007440567017, + 0.041737549006938934, + 0.14969757199287415 + ], + "credit_map": [ + 0.009527324698865414, + 0.00915255956351757, + 0.014896320179104805, + 0.013455569744110107, + 0.009597533382475376, + 0.010755390860140324, + 0.02204936183989048, + 0.010566558688879013, + 0.03400641307234764, + 0.015300683677196503, + 0.027698498219251633, + 0.045182038098573685, + 0.012020310387015343, + 0.02941085211932659, + 0.030230507254600525, + 0.025380294770002365, + 0.0041465479880571365, + 0.003644127631559968, + 0.005766578949987888, + 0.004474940709769726, + 0.003351066028699279, + 0.0054762959480285645, + 0.008135809563100338, + 0.0073486920446157455, + 0.005517569836229086, + 0.002968804445117712, + 0.014428820461034775, + 0.009194862097501755, + 0.00238260836340487, + 0.008031457662582397, + 0.004402229096740484, + 0.008341306820511818, + 0.005070237442851067, + 0.00601176405325532, + 0.008134075440466404, + 0.006563758011907339, + 0.0028737010434269905, + 0.01839536428451538, + 0.005954212509095669, + 0.014139688573777676, + 0.007181202527135611, + 0.004017080646008253, + 0.01322956196963787, + 0.009137497283518314, + 0.0020053996704518795, + 0.008002697490155697, + 0.004337658174335957, + 0.00741179846227169, + 0.009789797477424145, + 0.011920186690986156, + 0.010833857581019402, + 0.006748542655259371, + 0.007694203872233629, + 0.010923841968178749, + 0.006729851011186838, + 0.011136949062347412, + 0.023474955931305885, + 0.029611116275191307, + 0.027660803869366646, + 0.021251454949378967, + 0.024181094020605087, + 0.024341195821762085, + 0.012543216347694397, + 0.030416250228881836, + 0.003229669760912657, + 0.003437282284721732, + 0.005514898337423801, + 0.003634063294157386, + 0.0038632494397461414, + 0.007311482913792133, + 0.005368177313357592, + 0.009378728456795216, + 0.02269468829035759, + 0.00955631211400032, + 0.037448473274707794, + 0.011426075361669064, + 0.008115268312394619, + 0.02835049107670784, + 0.010656465776264668, + 0.021449808031320572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 9 + ], + "selected_reward": -0.05499401316046715 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019443194614723325 + ], + "opposite_logprob_tokens": [ + -6.251944541931152 + ], + "logprob_reward": -0.0019443194614723325, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019443194614723325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10805130749940872, + 0.12559738755226135, + 0.04728972911834717, + 0.08030045032501221, + 0.03849291801452637, + 0.06386107206344604, + 0.10559537261724472, + 0.16145752370357513, + 0.08512400090694427, + 0.18423017859458923 + ], + "credit_map": [ + 0.012290183454751968, + 0.010572333820164204, + 0.013858947902917862, + 0.015320383943617344, + 0.009090885519981384, + 0.015360090881586075, + 0.01695553958415985, + 0.014602948911488056, + 0.01697971299290657, + 0.008633679710328579, + 0.020268432796001434, + 0.012097722850739956, + 0.011319748125970364, + 0.017972124740481377, + 0.019251374527812004, + 0.01907457783818245, + 0.00449933810159564, + 0.0038582398556172848, + 0.008075469173491001, + 0.006340905092656612, + 0.003509699134156108, + 0.0068631283938884735, + 0.006175312213599682, + 0.007967637851834297, + 0.006138231139630079, + 0.0051358407363295555, + 0.012239673174917698, + 0.009834772907197475, + 0.004842904396355152, + 0.011574365198612213, + 0.014875871129333973, + 0.0156587865203619, + 0.004134134389460087, + 0.00667582405731082, + 0.004905935376882553, + 0.00552849005907774, + 0.0023369973059743643, + 0.005288047716021538, + 0.004383795894682407, + 0.005239694844931364, + 0.009049960412085056, + 0.004442104138433933, + 0.010712334886193275, + 0.009879200719296932, + 0.003266694489866495, + 0.009986599907279015, + 0.004986649379134178, + 0.011537526734173298, + 0.008890340104699135, + 0.013945101760327816, + 0.014420157298445702, + 0.010313034057617188, + 0.012562987394630909, + 0.016389211639761925, + 0.010600283741950989, + 0.018474262207746506, + 0.015666581690311432, + 0.017686352133750916, + 0.030714720487594604, + 0.01432016771286726, + 0.014486454427242279, + 0.018604667857289314, + 0.010236961767077446, + 0.0397416315972805, + 0.009214190766215324, + 0.005881182383745909, + 0.010503281839191914, + 0.006263018120080233, + 0.007441937457770109, + 0.014228880405426025, + 0.010720335878431797, + 0.020871169865131378, + 0.03361939638853073, + 0.010217061266303062, + 0.037098199129104614, + 0.011038769967854023, + 0.013846072368323803, + 0.02224935032427311, + 0.013858010061085224, + 0.04230332002043724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.0019443194614723325 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809821978211403 + ], + "opposite_logprob_tokens": [ + -3.2880983352661133 + ], + "logprob_reward": -0.03809821978211403, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809821978211403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08521897345781326, + 0.05163482204079628, + 0.04937836527824402, + 0.03475995734333992, + 0.05017875134944916, + 0.042568664997816086, + 0.09610433131456375, + 0.35196128487586975, + 0.1752966195344925, + 0.06289822608232498 + ], + "credit_map": [ + 0.010840301401913166, + 0.010742498561739922, + 0.011107493191957474, + 0.010529021732509136, + 0.007372364867478609, + 0.010458555072546005, + 0.015272684395313263, + 0.0088960537686944, + 0.00650467025116086, + 0.004260726273059845, + 0.006786881014704704, + 0.008762919344007969, + 0.0035589488688856363, + 0.0069803837686777115, + 0.007274806033819914, + 0.007505491841584444, + 0.004016504157334566, + 0.0033260101918131113, + 0.0074453577399253845, + 0.00732768140733242, + 0.003974877297878265, + 0.006363030523061752, + 0.009985940530896187, + 0.0069389683194458485, + 0.0032532596960663795, + 0.001922696945257485, + 0.00849576573818922, + 0.004897110629826784, + 0.001996885985136032, + 0.00455634668469429, + 0.0032446898985654116, + 0.006393201649188995, + 0.004765934310853481, + 0.005233445670455694, + 0.0053530726581811905, + 0.007901639677584171, + 0.004170066677033901, + 0.005984398070722818, + 0.010222528129816055, + 0.006547664292156696, + 0.004271507728844881, + 0.004529077559709549, + 0.006826975382864475, + 0.006198253482580185, + 0.0026797091122716665, + 0.0062103671953082085, + 0.004976886324584484, + 0.006875889375805855, + 0.009773745201528072, + 0.0114113949239254, + 0.017161773517727852, + 0.008772007189691067, + 0.009870370849967003, + 0.01402298454195261, + 0.008447853848338127, + 0.016644198447465897, + 0.04245469346642494, + 0.05415717512369156, + 0.05118979141116142, + 0.03913818672299385, + 0.04575277492403984, + 0.04427125304937363, + 0.02387968637049198, + 0.05111774802207947, + 0.020109184086322784, + 0.011487875133752823, + 0.028730211779475212, + 0.01620168797671795, + 0.01453358307480812, + 0.025708122178912163, + 0.0258370041847229, + 0.03268894925713539, + 0.009501234628260136, + 0.003799636848270893, + 0.011760517954826355, + 0.004709827713668346, + 0.006569351069629192, + 0.008212375454604626, + 0.007272914983332157, + 0.011072362773120403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.03809821978211403 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene unfolds in a medium close-up, with the camera beginning with a slow, deliberate push-in on a lone white candle, its candle showing delicate drips of cooling wax as it serves as the sole source of illumination in the pitch-black room. Its single flame trembles erratically, a teardrop of vibrant orange and yellow flame, its core a core of intense, incandescent light. This wild, unpredictable flickering casts vast, dancing shadows that wriggle and twist across the richly textured dark surfaces of the wooden table and the near background. As the camera inexorably approaches, focus settles on the flame itself, its ethereal, smoky tip licking at the oppressive darkness. The push-in continues, moving through the body of the flame, dissolving into an extreme close-up of the blackened fiber wick, its tip emitting a bright, incandescent red-orange flash that gently pulsates with heat, creating an atmosphere of profound solitude and concentration, where warm, dynamic light engages in a silent, brutal battle against a deep, enveloping darkness.", + "seed": 9761, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_2/index_5772_seed_9761_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.014242700021713971, + "vlm_logprob": -0.014242700021713971, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.014242700021713971, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00887154508382082 + ], + "opposite_logprob_tokens": [ + -5.258871555328369 + ], + "logprob_reward": -0.00887154508382082, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00887154508382082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17977596819400787, + 0.15471623837947845, + 0.13451170921325684, + 0.04718814790248871, + 0.05980081111192703, + 0.07564830780029297, + 0.08886061608791351, + 0.09215942770242691, + 0.07688946276903152, + 0.09044929593801498 + ], + "credit_map": [ + 0.02220749668776989, + 0.016148021444678307, + 0.02574586495757103, + 0.022531993687152863, + 0.020339656621217728, + 0.027079785242676735, + 0.01597891002893448, + 0.029744239524006844, + 0.015888622030615807, + 0.013755674473941326, + 0.025298509746789932, + 0.020531823858618736, + 0.014812417328357697, + 0.021456751972436905, + 0.018379246816039085, + 0.024593189358711243, + 0.015435770153999329, + 0.01451756339520216, + 0.019623877480626106, + 0.010969731956720352, + 0.016417019069194794, + 0.02008514292538166, + 0.01694638468325138, + 0.020516211166977882, + 0.0043482123874127865, + 0.003344222204759717, + 0.008051459677517414, + 0.005689919460564852, + 0.004065534565597773, + 0.006810594350099564, + 0.008016872219741344, + 0.006861330941319466, + 0.004925561603158712, + 0.004607864189893007, + 0.00926660280674696, + 0.007899055257439613, + 0.004762645345181227, + 0.009360896423459053, + 0.010396047495305538, + 0.008582141250371933, + 0.008762230165302753, + 0.006971157621592283, + 0.013224422931671143, + 0.007229005917906761, + 0.00604834733530879, + 0.01060935016721487, + 0.011918510310351849, + 0.01088529173284769, + 0.015381962060928345, + 0.005297516472637653, + 0.009430096484720707, + 0.006948554888367653, + 0.004639452323317528, + 0.015184232033789158, + 0.010402091778814793, + 0.02157670632004738, + 0.008489650674164295, + 0.010784686543047428, + 0.010414035990834236, + 0.010589942336082458, + 0.016631105914711952, + 0.010783923789858818, + 0.009672868065536022, + 0.014793211594223976, + 0.006674736272543669, + 0.006321200169622898, + 0.0110862972214818, + 0.009256858378648758, + 0.015711544081568718, + 0.00807529129087925, + 0.009133056737482548, + 0.010630475357174873, + 0.010639901272952557, + 0.004421961959451437, + 0.007800176739692688, + 0.00525463093072176, + 0.008342343382537365, + 0.022290220484137535, + 0.011235472746193409, + 0.020464584231376648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.00887154508382082 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016077987849712372 + ], + "opposite_logprob_tokens": [ + -4.141077995300293 + ], + "logprob_reward": -0.016077987849712372, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016077987849712372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1375022828578949, + 0.04020101949572563, + 0.06327913701534271, + 0.08893115818500519, + 0.08904528617858887, + 0.2429029941558838, + 0.09515797346830368, + 0.15763235092163086, + 0.04012823477387428, + 0.04521951079368591 + ], + "credit_map": [ + 0.016328230500221252, + 0.013168116100132465, + 0.02026013657450676, + 0.013946650549769402, + 0.016018252819776535, + 0.02341078221797943, + 0.011136876419186592, + 0.023233234882354736, + 0.004928179085254669, + 0.003437320701777935, + 0.006287571508437395, + 0.00548416655510664, + 0.00404014578089118, + 0.006348765920847654, + 0.004401446785777807, + 0.005273429211229086, + 0.004637857899069786, + 0.004755338188260794, + 0.011640215292572975, + 0.006639048922806978, + 0.005967497359961271, + 0.01185675710439682, + 0.010056309401988983, + 0.00772611889988184, + 0.011452486738562584, + 0.007040992844849825, + 0.012002638541162014, + 0.008748555555939674, + 0.008646806702017784, + 0.014157763682305813, + 0.011848696507513523, + 0.015033225528895855, + 0.008919073268771172, + 0.007359311915934086, + 0.014242194592952728, + 0.010980126447975636, + 0.00772028649225831, + 0.014448406174778938, + 0.010474060662090778, + 0.0149018419906497, + 0.023666536435484886, + 0.023150138556957245, + 0.051238756626844406, + 0.024958597496151924, + 0.022108931094408035, + 0.035490840673446655, + 0.02584954723715782, + 0.0364396907389164, + 0.01191010233014822, + 0.0095508499071002, + 0.008768572472035885, + 0.010393045842647552, + 0.007319054566323757, + 0.01387842744588852, + 0.012760055251419544, + 0.020577874034643173, + 0.011182419024407864, + 0.022014444693922997, + 0.01657133921980858, + 0.016675738617777824, + 0.034448109567165375, + 0.01890144683420658, + 0.014819336123764515, + 0.023019522428512573, + 0.006347841117531061, + 0.0025053394492715597, + 0.005748773459345102, + 0.0028377813287079334, + 0.004249683581292629, + 0.0055349296890199184, + 0.003487139707431197, + 0.009416751563549042, + 0.004587467759847641, + 0.0018327281577512622, + 0.006083295214921236, + 0.002370673231780529, + 0.003736473387107253, + 0.008791310712695122, + 0.007030656095594168, + 0.01078691054135561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.016077987849712372 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018204044550657272 + ], + "opposite_logprob_tokens": [ + -4.018204212188721 + ], + "logprob_reward": -0.018204044550657272, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018204044550657272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29127922654151917, + 0.14809751510620117, + 0.09745118021965027, + 0.026813089847564697, + 0.04998849704861641, + 0.04104701802134514, + 0.07182422280311584, + 0.06233213096857071, + 0.032898951321840286, + 0.17826813459396362 + ], + "credit_map": [ + 0.03559081256389618, + 0.02705118991434574, + 0.03897946700453758, + 0.03327620401978493, + 0.035694729536771774, + 0.04755063354969025, + 0.023187506943941116, + 0.04994869977235794, + 0.01330035924911499, + 0.014240109361708164, + 0.024237966164946556, + 0.020741315558552742, + 0.015495051629841328, + 0.02066243626177311, + 0.018890980631113052, + 0.02052929997444153, + 0.008897398598492146, + 0.008393104188144207, + 0.01779482513666153, + 0.009629779495298862, + 0.009855988435447216, + 0.014716899022459984, + 0.013360124081373215, + 0.014803068712353706, + 0.0025682852137833834, + 0.0019551359582692385, + 0.005105450749397278, + 0.0027698036283254623, + 0.0021194510627537966, + 0.0036690908018499613, + 0.003790481248870492, + 0.004835390951484442, + 0.00502995727583766, + 0.004233826417475939, + 0.008221869356930256, + 0.006369355600327253, + 0.004329517018049955, + 0.007522116880863905, + 0.006902506109327078, + 0.007379349321126938, + 0.0037307722959667444, + 0.003730660304427147, + 0.005763864144682884, + 0.00523742102086544, + 0.003658218076452613, + 0.006351345684379339, + 0.004199055954813957, + 0.008375680074095726, + 0.007202443201094866, + 0.008815878070890903, + 0.006856669671833515, + 0.009963549673557281, + 0.006799902766942978, + 0.009165630675852299, + 0.010122542269527912, + 0.012897610664367676, + 0.007633067201822996, + 0.008316553197801113, + 0.005718320608139038, + 0.005811026785522699, + 0.008621915243566036, + 0.008157419040799141, + 0.006818288471549749, + 0.011255543678998947, + 0.0024855199735611677, + 0.0032025305554270744, + 0.004917627200484276, + 0.004308158997446299, + 0.007329440675675869, + 0.003471029456704855, + 0.003638774622231722, + 0.0035458693746477365, + 0.009869670495390892, + 0.004058435093611479, + 0.042284514755010605, + 0.012245147489011288, + 0.00941633153706789, + 0.02172803319990635, + 0.06472913920879364, + 0.013936865143477917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.018204044550657272 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017190692014992237 + ], + "opposite_logprob_tokens": [ + -6.376718997955322 + ], + "logprob_reward": -0.0017190692014992237, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017190692014992237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07409407943487167, + 0.1497374325990677, + 0.09623969346284866, + 0.04326428845524788, + 0.056024715304374695, + 0.08038917928934097, + 0.1840806007385254, + 0.09568384289741516, + 0.08784319460391998, + 0.13264304399490356 + ], + "credit_map": [ + 0.005926677491515875, + 0.008760073222219944, + 0.010011429898440838, + 0.009559936821460724, + 0.008611011318862438, + 0.014059950597584248, + 0.007497869897633791, + 0.009667125530540943, + 0.01854681223630905, + 0.013047567568719387, + 0.023603038862347603, + 0.019620604813098907, + 0.014636838808655739, + 0.020657788962125778, + 0.01641738973557949, + 0.023207388818264008, + 0.011951114982366562, + 0.006932342890650034, + 0.015517759136855602, + 0.008980651386082172, + 0.0077827065251767635, + 0.014692342840135098, + 0.013944534584879875, + 0.016438229009509087, + 0.0044204494915902615, + 0.0032123506534844637, + 0.010556287132203579, + 0.002555645303800702, + 0.002986079314723611, + 0.009026559069752693, + 0.0029605224262923002, + 0.00754639133810997, + 0.005682749208062887, + 0.004404453095048666, + 0.0094629917293787, + 0.006389852613210678, + 0.00511607201769948, + 0.010012740269303322, + 0.006816512439399958, + 0.008139334619045258, + 0.006089362781494856, + 0.010161413811147213, + 0.01231918204575777, + 0.012813684530556202, + 0.009101640433073044, + 0.009564354084432125, + 0.011083743534982204, + 0.009255792014300823, + 0.021427255123853683, + 0.022541727870702744, + 0.022004319354891777, + 0.02776096761226654, + 0.02154393307864666, + 0.022199437022209167, + 0.021367158740758896, + 0.025235788896679878, + 0.012176728807389736, + 0.012830613180994987, + 0.008692099712789059, + 0.010149115696549416, + 0.01514698937535286, + 0.012842411175370216, + 0.008226736448705196, + 0.015619128942489624, + 0.006931987591087818, + 0.01098257303237915, + 0.008710723370313644, + 0.012640183791518211, + 0.020252108573913574, + 0.009275296702980995, + 0.010257813148200512, + 0.008792494423687458, + 0.009677228517830372, + 0.005080068949609995, + 0.02441578358411789, + 0.008298700675368309, + 0.00729731610044837, + 0.026040609925985336, + 0.03902041167020798, + 0.012812895700335503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 9 + ], + "selected_reward": -0.0017190692014992237 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340853422880173 + ], + "opposite_logprob_tokens": [ + -3.651340961456299 + ], + "logprob_reward": -0.026340853422880173, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340853422880173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.311022013425827, + 0.09095315635204315, + 0.047539521008729935, + 0.05522000789642334, + 0.02683386206626892, + 0.1427019089460373, + 0.0991106852889061, + 0.06002014875411987, + 0.06243165582418442, + 0.10416707396507263 + ], + "credit_map": [ + 0.036576565355062485, + 0.029579751193523407, + 0.046618521213531494, + 0.036132268607616425, + 0.03588774800300598, + 0.04972333461046219, + 0.02495410293340683, + 0.05154969170689583, + 0.011863462626934052, + 0.007669624872505665, + 0.014739918522536755, + 0.009874816983938217, + 0.008801720105111599, + 0.01374293677508831, + 0.009793197736144066, + 0.014467472210526466, + 0.006540268659591675, + 0.003214603988453746, + 0.007204569410532713, + 0.003915155306458473, + 0.00437747361138463, + 0.007750235963612795, + 0.004836983047425747, + 0.009700223803520203, + 0.004919954109936953, + 0.005267168395221233, + 0.009662020951509476, + 0.006908178795129061, + 0.005957181099802256, + 0.00856667011976242, + 0.007152683101594448, + 0.0067861489951610565, + 0.0024442391004413366, + 0.0022390265949070454, + 0.004563000053167343, + 0.0032694891560822725, + 0.002509678015485406, + 0.004033169709146023, + 0.0038346943911165, + 0.003940562717616558, + 0.012363567017018795, + 0.012620625086128712, + 0.030274638906121254, + 0.015211445279419422, + 0.015566892921924591, + 0.02040936052799225, + 0.014381203800439835, + 0.021874157711863518, + 0.013132276944816113, + 0.008735568262636662, + 0.009882012382149696, + 0.01049443893134594, + 0.007126244716346264, + 0.0148738082498312, + 0.013481340371072292, + 0.021384987980127335, + 0.0044273450039327145, + 0.006853246595710516, + 0.009764636866748333, + 0.0063182818703353405, + 0.011638625524938107, + 0.007689923979341984, + 0.005898341536521912, + 0.007429738994687796, + 0.008067872375249863, + 0.005422038491815329, + 0.007653210312128067, + 0.007367607671767473, + 0.006927442271262407, + 0.008295073173940182, + 0.006288144737482071, + 0.01241026446223259, + 0.008610988967120647, + 0.004088392946869135, + 0.013815291225910187, + 0.007531792391091585, + 0.007011201232671738, + 0.02275572344660759, + 0.023972289636731148, + 0.016381388530135155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.026340853422880173 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene unfolds in a medium close-up, with the camera beginning with a slow, deliberate push-in on a lone white candle, its candle showing delicate drips of cooling wax as it serves as the sole source of illumination in the pitch-black room. Its single flame trembles erratically, a teardrop of vibrant orange and yellow flame, its core a core of intense, incandescent light. This wild, unpredictable flickering casts vast, dancing shadows that wriggle and twist across the richly textured dark surfaces of the wooden table and the near background. As the camera inexorably approaches, focus settles on the flame itself, its ethereal, smoky tip licking at the oppressive darkness. The push-in continues, moving through the body of the flame, dissolving into an extreme close-up of the blackened fiber wick, its tip emitting a bright, incandescent red-orange flash that gently pulsates with heat, creating an atmosphere of profound solitude and concentration, where warm, dynamic light engages in a silent, brutal battle against a deep, enveloping darkness.", + "seed": 9762, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_2/index_5772_seed_9762_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03686156463809311, + "vlm_logprob": -0.03686156463809311, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.03686156463809311, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0222710482776165 + ], + "opposite_logprob_tokens": [ + -4.022271156311035 + ], + "logprob_reward": -0.0222710482776165, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0222710482776165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12081345915794373, + 0.05507616326212883, + 0.06077038124203682, + 0.0566067099571228, + 0.20553825795650482, + 0.15723036229610443, + 0.14060351252555847, + 0.12738177180290222, + 0.050649695098400116, + 0.025329714640975 + ], + "credit_map": [ + 0.014012576080858707, + 0.008147007785737514, + 0.014519283547997475, + 0.011550317518413067, + 0.013490279205143452, + 0.016876552253961563, + 0.022857321426272392, + 0.019360117614269257, + 0.004876589402556419, + 0.00402456521987915, + 0.008276810869574547, + 0.00551665760576725, + 0.007210943382233381, + 0.009423919953405857, + 0.006824835203588009, + 0.008921840228140354, + 0.004468970000743866, + 0.006512037944048643, + 0.009807298891246319, + 0.008120797574520111, + 0.008159522898495197, + 0.00794718787074089, + 0.00780504010617733, + 0.007949522696435452, + 0.005584956146776676, + 0.006926636211574078, + 0.005552067421376705, + 0.004288756288588047, + 0.00903914775699377, + 0.007996682077646255, + 0.00586191238835454, + 0.011356553062796593, + 0.016944142058491707, + 0.04960259795188904, + 0.020594226196408272, + 0.014210865832865238, + 0.034548886120319366, + 0.028629817068576813, + 0.01828642003238201, + 0.02272130735218525, + 0.0184563547372818, + 0.0180460624396801, + 0.03584771230816841, + 0.011125551536679268, + 0.015565067529678345, + 0.023622354492545128, + 0.012698497623205185, + 0.02186875231564045, + 0.01492591854184866, + 0.012808042578399181, + 0.01605968363583088, + 0.013598148711025715, + 0.00628293352201581, + 0.02740616910159588, + 0.012893388979136944, + 0.03662923723459244, + 0.011566135101020336, + 0.005418766289949417, + 0.036059148609638214, + 0.009942350909113884, + 0.006443350575864315, + 0.03753038868308067, + 0.007685365155339241, + 0.012736264616250992, + 0.00538382725790143, + 0.0034634890034794807, + 0.008813473396003246, + 0.00475957291200757, + 0.004203214310109615, + 0.009281765669584274, + 0.006393704563379288, + 0.00835064984858036, + 0.005765617825090885, + 0.0017518053064122796, + 0.003895130939781666, + 0.002265053102746606, + 0.0015645003877580166, + 0.0042334855534136295, + 0.0024603044148534536, + 0.0033938162960112095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.0222710482776165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496196821331978 + ], + "opposite_logprob_tokens": [ + -2.929961919784546 + ], + "logprob_reward": -0.05496196821331978, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496196821331978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07545758783817291, + 0.1830548793077469, + 0.069609135389328, + 0.07263407856225967, + 0.04762076959013939, + 0.21081718802452087, + 0.06005173921585083, + 0.18679702281951904, + 0.0699131190776825, + 0.0240444615483284 + ], + "credit_map": [ + 0.00834977999329567, + 0.004104532767087221, + 0.01491798646748066, + 0.006566202733665705, + 0.008862498216331005, + 0.010761232115328312, + 0.007778085768222809, + 0.014117267914116383, + 0.01193864643573761, + 0.010709114372730255, + 0.04858773201704025, + 0.015191398561000824, + 0.02329467609524727, + 0.02276510000228882, + 0.027274683117866516, + 0.02329353056848049, + 0.006100363098084927, + 0.007083269767463207, + 0.014343684539198875, + 0.007899518124759197, + 0.00768773490563035, + 0.010050592944025993, + 0.0064239841885864735, + 0.010019990615546703, + 0.010420400649309158, + 0.007429599296301603, + 0.007642379496246576, + 0.010933888144791126, + 0.009351479820907116, + 0.009840202517807484, + 0.007692823652178049, + 0.0093233035877347, + 0.003586387261748314, + 0.00357698742300272, + 0.006244541145861149, + 0.006166927050799131, + 0.00452715577557683, + 0.007932054810225964, + 0.009420247748494148, + 0.006166473496705294, + 0.017836960032582283, + 0.01605847105383873, + 0.03346329182386398, + 0.025073431432247162, + 0.022182157263159752, + 0.03673121705651283, + 0.02375825121998787, + 0.03571341931819916, + 0.006068216171115637, + 0.011325056664645672, + 0.008095923811197281, + 0.008893144316971302, + 0.004249776713550091, + 0.006972658913582563, + 0.007474414072930813, + 0.006972553674131632, + 0.012741991318762302, + 0.008350527845323086, + 0.061695974320173264, + 0.01185611542314291, + 0.009180134162306786, + 0.057478196918964386, + 0.010254300199449062, + 0.015239781700074673, + 0.007810738403350115, + 0.004840545821934938, + 0.012445439584553242, + 0.009237327612936497, + 0.005199925508350134, + 0.011124426499009132, + 0.008334167301654816, + 0.010920547880232334, + 0.005231027491390705, + 0.0016434924909844995, + 0.0038268042262643576, + 0.002439146861433983, + 0.002107804175466299, + 0.0036407546140253544, + 0.0017109329346567392, + 0.0034444998018443584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.05496196821331978 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901664823293686 + ], + "opposite_logprob_tokens": [ + -2.57901668548584 + ], + "logprob_reward": -0.07901664823293686, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901664823293686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07292617112398148, + 0.1446821391582489, + 0.04347675293684006, + 0.0941670835018158, + 0.03181488439440727, + 0.3286169171333313, + 0.07653272151947021, + 0.07843643426895142, + 0.06321000307798386, + 0.06613688915967941 + ], + "credit_map": [ + 0.006573040969669819, + 0.0034446860663592815, + 0.011922497302293777, + 0.007099309470504522, + 0.00942473579198122, + 0.011336655355989933, + 0.009817813523113728, + 0.013307429850101471, + 0.009885447099804878, + 0.007203329354524612, + 0.04372689500451088, + 0.011751171201467514, + 0.015628062188625336, + 0.017997123301029205, + 0.0208083838224411, + 0.017681725323200226, + 0.003957598004490137, + 0.003965839743614197, + 0.006443944294005632, + 0.0053211539052426815, + 0.005849543027579784, + 0.006676511839032173, + 0.004174366127699614, + 0.0070877973921597, + 0.00871182605624199, + 0.01279351394623518, + 0.009138648398220539, + 0.006593227852135897, + 0.015471826307475567, + 0.014301241375505924, + 0.01091382559388876, + 0.016242971643805504, + 0.0027097833808511496, + 0.00420245248824358, + 0.0054320599883794785, + 0.0033290365245193243, + 0.0037308274768292904, + 0.0047370558604598045, + 0.003799575613811612, + 0.003874093759804964, + 0.022282227873802185, + 0.025037448853254318, + 0.028277264907956123, + 0.04903256520628929, + 0.04794849455356598, + 0.04666559025645256, + 0.0613059364259243, + 0.04806738346815109, + 0.006972766015678644, + 0.006595233920961618, + 0.008716879412531853, + 0.007945225574076176, + 0.003316324669867754, + 0.016352085396647453, + 0.008070860058069229, + 0.018563350662589073, + 0.010388044640421867, + 0.006753549911081791, + 0.014325560070574284, + 0.00824953243136406, + 0.005472993943840265, + 0.011061838828027248, + 0.00857480801641941, + 0.013610114343464375, + 0.010889890603721142, + 0.006367089226841927, + 0.012397308833897114, + 0.006096647121012211, + 0.0033142815809696913, + 0.010377675294876099, + 0.0037246854044497013, + 0.010042428970336914, + 0.016472669318318367, + 0.003891776781529188, + 0.011373860761523247, + 0.005554641131311655, + 0.00372749543748796, + 0.011242003180086613, + 0.0050998348742723465, + 0.008774608373641968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.07901664823293686 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001718117156997323 + ], + "opposite_logprob_tokens": [ + -6.376718044281006 + ], + "logprob_reward": -0.001718117156997323, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001718117156997323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07850603014230728, + 0.22851325571537018, + 0.028921781107783318, + 0.04731811210513115, + 0.06132303178310394, + 0.1259165108203888, + 0.14737369120121002, + 0.10792778432369232, + 0.051273103803396225, + 0.1229267418384552 + ], + "credit_map": [ + 0.011630555614829063, + 0.005122426897287369, + 0.013040144927799702, + 0.007197377271950245, + 0.010168411768972874, + 0.011440679430961609, + 0.011126061901450157, + 0.00878036767244339, + 0.020876454189419746, + 0.011521385982632637, + 0.07585091888904572, + 0.0186834204941988, + 0.022621095180511475, + 0.029347503557801247, + 0.023581648245453835, + 0.02603081241250038, + 0.003323339857161045, + 0.0034350950736552477, + 0.005480380263179541, + 0.0037898316513746977, + 0.0029900497756898403, + 0.0033103604800999165, + 0.002883032662793994, + 0.0037096887826919556, + 0.005176410544663668, + 0.006201484706252813, + 0.004249030724167824, + 0.0059852683916687965, + 0.0066344463266432285, + 0.006170209962874651, + 0.00471164844930172, + 0.00818960927426815, + 0.005021700169891119, + 0.007234344724565744, + 0.005354311317205429, + 0.007355771958827972, + 0.008273809216916561, + 0.010171618312597275, + 0.008068813011050224, + 0.00984265562146902, + 0.01297149620950222, + 0.011986213736236095, + 0.025786763057112694, + 0.011654088273644447, + 0.012468208558857441, + 0.01990235410630703, + 0.011328286491334438, + 0.019819099456071854, + 0.015273204073309898, + 0.013085976243019104, + 0.016339506953954697, + 0.015585404820740223, + 0.006821908988058567, + 0.027928326278924942, + 0.015175705775618553, + 0.03716364502906799, + 0.00900979246944189, + 0.005102538038045168, + 0.03103962354362011, + 0.008311837911605835, + 0.009342479519546032, + 0.02881905622780323, + 0.006514785811305046, + 0.009787658229470253, + 0.006602675653994083, + 0.004805508069694042, + 0.007713568862527609, + 0.00532521354034543, + 0.004873418714851141, + 0.00764048658311367, + 0.005016976967453957, + 0.009295251220464706, + 0.018069174140691757, + 0.005774311721324921, + 0.02028902806341648, + 0.00965720135718584, + 0.005671361926943064, + 0.03183959051966667, + 0.010926361195743084, + 0.02069970779120922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 5 + ], + "selected_reward": -0.001718117156997323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium close-up of a lone white candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, deliberate push-in movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the candle the sole source of illumination in a pitch-black room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame exhibit a vibrant orange and yellow color with a core of intense incandescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move through the flame to end in an extreme close-up of the blackened wick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340041309595108 + ], + "opposite_logprob_tokens": [ + -3.6513400077819824 + ], + "logprob_reward": -0.026340041309595108, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340041309595108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04704431816935539, + 0.2974413335323334, + 0.04211157560348511, + 0.0336725115776062, + 0.0762348622083664, + 0.10657478868961334, + 0.06526318192481995, + 0.09741079807281494, + 0.0886056125164032, + 0.1456409990787506 + ], + "credit_map": [ + 0.005091672297567129, + 0.002753563690930605, + 0.012671418488025665, + 0.00374857266433537, + 0.005654745735228062, + 0.006452571135014296, + 0.004727064166218042, + 0.005944706965237856, + 0.02121310867369175, + 0.015285165049135685, + 0.09203805774450302, + 0.025517551228404045, + 0.032741449773311615, + 0.036820925772190094, + 0.0397934690117836, + 0.03403162583708763, + 0.0038447529077529907, + 0.003275342984125018, + 0.008654503151774406, + 0.004598758649080992, + 0.005075087305158377, + 0.006018526386469603, + 0.003648858517408371, + 0.006995746400207281, + 0.002987639745697379, + 0.003711797995492816, + 0.003368169069290161, + 0.0025016693398356438, + 0.004849039018154144, + 0.005992883816361427, + 0.00255369720980525, + 0.007707614451646805, + 0.008854209445416927, + 0.013730519451200962, + 0.012154768221080303, + 0.004867231007665396, + 0.009370507672429085, + 0.0110328309237957, + 0.00476981233805418, + 0.011454976163804531, + 0.008274232968688011, + 0.00856251735240221, + 0.012141224928200245, + 0.014375361613929272, + 0.013348850421607494, + 0.016403265297412872, + 0.01689225248992443, + 0.01657707989215851, + 0.004252597223967314, + 0.006032173987478018, + 0.00515780970454216, + 0.005439585540443659, + 0.0032177651301026344, + 0.01866094209253788, + 0.004762585274875164, + 0.01773972250521183, + 0.00938607007265091, + 0.009502586908638477, + 0.01845163106918335, + 0.009749218821525574, + 0.011874545365571976, + 0.018864544108510017, + 0.009257499128580093, + 0.010324706323444843, + 0.011214837431907654, + 0.007489843759685755, + 0.014033480547368526, + 0.009489421732723713, + 0.005726263392716646, + 0.016500679776072502, + 0.008389344438910484, + 0.01576174609363079, + 0.017069078981876373, + 0.006692929193377495, + 0.022451486438512802, + 0.01397474855184555, + 0.009387214668095112, + 0.032833654433488846, + 0.014478287659585476, + 0.02875361405313015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 5 + ], + "selected_reward": -0.026340041309595108 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_3/index_tensor([4841, 4841])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_3/index_tensor([4841, 4841])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..03942c58cfd896dc7993158198aaf6cac7afa987 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_3/index_tensor([4841, 4841])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly advancing wide shot, transitioning from a distant shot to a medium-long shot, reveals a magnificent mountaintop shimmering in the morning light. Within the frame, a rolling sea of clouds resembles a boundless white ocean, engulfing the mountainsides, with only a few dark blue peaks emerging like isolated islands. The sky shimmers delicately from deep indigo to pale white, with a warm orange-pink halo already permeating the horizon. On this secluded rocky platform, a Taoist priest, clad in a flowing off-white, wide-sleeved linen robe and a delicate bun held in place by an ancient wooden hairpin, stands with his back to the rising sun. He wields a peachwood sword, its blade a warm reddish-brown, engraved with faint, ancient runes. With his legs slightly bent and his center of gravity lowered, he dances with a steady, deliberate rhythm, the hem of his robe and its wide cuffs flowing gracefully with his movements. His movements were not swift but imbued with a restrained sense of power. Each turn, arm raise, and sword swing was precise and continuous. With a deft twist of his wrist, the tip of the peach wood sword traced clear, brilliant streaks of golden light through the cool air. These streaks seemed composed of countless luminous particles, lingering briefly in the air, forming graceful arcs and symbols before slowly dissipating like stardust. The soft morning light cast a faint golden rim around his silhouette, complementing the radiance of the sword tip. Against a cold background, the entire scene was punctuated by the warm glow of the rising sun and the golden radiance of the sword tip, creating an epic atmosphere of sacredness, tranquility, and the mystical power of the East.", + "seed": 21632, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_3/index_4841_seed_21632_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.030762510281056164, + "vlm_logprob": -0.030762510281056164, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.030762510281056164, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012597156688570976 + ], + "opposite_logprob_tokens": [ + -4.51259708404541 + ], + "logprob_reward": -0.012597156688570976, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.012597156688570976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39945363998413086, + 0.06253870576620102, + 0.11613951623439789, + 0.04651585966348648, + 0.026836350560188293, + 0.049919918179512024, + 0.0869101881980896, + 0.09273983538150787, + 0.054460518062114716, + 0.06448554992675781 + ], + "credit_map": [ + 0.11918456107378006, + 0.04097340255975723, + 0.14708247780799866, + 0.034816231578588486, + 0.009317472577095032, + 0.013204986229538918, + 0.01898381859064102, + 0.015890685841441154, + 0.009157150983810425, + 0.0075202276930212975, + 0.006840935908257961, + 0.008819656446576118, + 0.0062374514527618885, + 0.008690809831023216, + 0.007892069406807423, + 0.0073803989216685295, + 0.00999758392572403, + 0.029531307518482208, + 0.005587256513535976, + 0.007942781783640385, + 0.00925707258284092, + 0.03632412850856781, + 0.012020628899335861, + 0.005478753242641687, + 0.007345257326960564, + 0.00522212591022253, + 0.0049198889173567295, + 0.008205939084291458, + 0.0051689548417925835, + 0.00567478546872735, + 0.004231458064168692, + 0.005747449118643999, + 0.00412619998678565, + 0.0028341077268123627, + 0.002668988425284624, + 0.0035252731759101152, + 0.003577225375920534, + 0.0028452216647565365, + 0.0033046803437173367, + 0.003954652231186628, + 0.00586076732724905, + 0.0057840957306325436, + 0.005247244145721197, + 0.006268576253205538, + 0.005078363232314587, + 0.006758221425116062, + 0.006985229440033436, + 0.007937421090900898, + 0.017020806670188904, + 0.011152810417115688, + 0.012863468378782272, + 0.007358947303146124, + 0.007499960716813803, + 0.01192645262926817, + 0.008804497309029102, + 0.010283243842422962, + 0.023289235308766365, + 0.017284531146287918, + 0.011345066130161285, + 0.005901389755308628, + 0.005823386833071709, + 0.015312579460442066, + 0.006927666254341602, + 0.00685598561540246, + 0.010295715183019638, + 0.007659949362277985, + 0.00877627357840538, + 0.00583897577598691, + 0.004457684233784676, + 0.008508156053721905, + 0.005113312974572182, + 0.003810448106378317, + 0.00812477245926857, + 0.008239672519266605, + 0.00926823541522026, + 0.007622289936989546, + 0.006583884824067354, + 0.008431456051766872, + 0.010001213289797306, + 0.006214024033397436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.012597156688570976 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018207088112831116 + ], + "opposite_logprob_tokens": [ + -4.01820707321167 + ], + "logprob_reward": -0.018207088112831116, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018207088112831116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4418097138404846, + 0.044874679297208786, + 0.1813834309577942, + 0.0882936492562294, + 0.014548728242516518, + 0.03687330707907677, + 0.0401957705616951, + 0.0726870596408844, + 0.03647455945611, + 0.042859163135290146 + ], + "credit_map": [ + 0.07969590276479721, + 0.047557681798934937, + 0.06458241492509842, + 0.02813102863729, + 0.033509958535432816, + 0.07055424153804779, + 0.04781583324074745, + 0.06996262073516846, + 0.005313246976584196, + 0.0071649751625955105, + 0.003473918419331312, + 0.005091591738164425, + 0.005178495310246944, + 0.008568874560296535, + 0.00656227907165885, + 0.003521295264363289, + 0.01306125707924366, + 0.029145462438464165, + 0.01074570044875145, + 0.02610078640282154, + 0.023896409198641777, + 0.030162625014781952, + 0.039472177624702454, + 0.008799000643193722, + 0.013179007917642593, + 0.01126609742641449, + 0.009502647444605827, + 0.02039642073214054, + 0.009728568606078625, + 0.008054214529693127, + 0.009595806710422039, + 0.006570876110345125, + 0.0024457103572785854, + 0.0017186484765261412, + 0.00195558974519372, + 0.001690479344688356, + 0.001800874131731689, + 0.0015187280951067805, + 0.002155895112082362, + 0.0012628008844330907, + 0.0067932796664536, + 0.0043684132397174835, + 0.003308587707579136, + 0.004590809345245361, + 0.003907802980393171, + 0.004336124751716852, + 0.005413563922047615, + 0.004154720809310675, + 0.004549549892544746, + 0.005367872770875692, + 0.0034918540623039007, + 0.005672883708029985, + 0.006271635182201862, + 0.0046182433143258095, + 0.0069594732485711575, + 0.0032642532605677843, + 0.014793053269386292, + 0.011785010807216167, + 0.005260791163891554, + 0.00822358950972557, + 0.006742189638316631, + 0.008942296728491783, + 0.008051684126257896, + 0.008888434618711472, + 0.004195242654532194, + 0.004366322420537472, + 0.003917641472071409, + 0.008297229185700417, + 0.0042389146983623505, + 0.003939066082239151, + 0.004702482372522354, + 0.0028176582418382168, + 0.004972049966454506, + 0.006802576128393412, + 0.0029964535497128963, + 0.0035000850912183523, + 0.007166146766394377, + 0.005226893816143274, + 0.008829455822706223, + 0.0033654964063316584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.018207088112831116 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500157177448273 + ], + "opposite_logprob_tokens": [ + -2.9300014972686768 + ], + "logprob_reward": -0.05500157177448273, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500157177448273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2948457896709442, + 0.04657343775033951, + 0.08205071836709976, + 0.06459981203079224, + 0.027334537357091904, + 0.04719855263829231, + 0.061337679624557495, + 0.3114277422428131, + 0.017761312425136566, + 0.04687045142054558 + ], + "credit_map": [ + 0.0643884688615799, + 0.027177536860108376, + 0.07309140264987946, + 0.02496175281703472, + 0.014216032810509205, + 0.03425491973757744, + 0.021671216934919357, + 0.03508446738123894, + 0.007344779092818499, + 0.006891357246786356, + 0.004047878552228212, + 0.006905088201165199, + 0.004630815237760544, + 0.00625929981470108, + 0.007481429725885391, + 0.003012786153703928, + 0.009821328334510326, + 0.011712413281202316, + 0.005779580678790808, + 0.009075142443180084, + 0.012499582953751087, + 0.010508827865123749, + 0.01750403828918934, + 0.0051498026587069035, + 0.00683040963485837, + 0.007296656258404255, + 0.007904089987277985, + 0.01409878209233284, + 0.007773067336529493, + 0.0064211501739919186, + 0.008062867447733879, + 0.006212787237018347, + 0.003243738552555442, + 0.0033985483460128307, + 0.002584816887974739, + 0.00344742089509964, + 0.0032783816568553448, + 0.0029324039351195097, + 0.004171003587543964, + 0.004278223030269146, + 0.007576436270028353, + 0.0057914829812943935, + 0.004955394193530083, + 0.005590734072029591, + 0.004055928438901901, + 0.006246763281524181, + 0.006780680269002914, + 0.006201133131980896, + 0.010759847238659859, + 0.007561864797025919, + 0.006158038508147001, + 0.006403309293091297, + 0.006669438444077969, + 0.008628681302070618, + 0.008546153083443642, + 0.006610349752008915, + 0.03625141456723213, + 0.03640132024884224, + 0.05828043445944786, + 0.02781979739665985, + 0.02528981864452362, + 0.03619031980633736, + 0.043393295258283615, + 0.04780132323503494, + 0.002695700852200389, + 0.0022260043770074844, + 0.001544101512990892, + 0.0026262535247951746, + 0.0020125105511397123, + 0.0017663984326645732, + 0.0030450078193098307, + 0.0018453339580446482, + 0.004689447116106749, + 0.005004493054002523, + 0.003993095364421606, + 0.005971368867903948, + 0.0072287567891180515, + 0.0052739642560482025, + 0.009588593617081642, + 0.005120731890201569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.05500157177448273 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948815029114485 + ], + "opposite_logprob_tokens": [ + -5.130949020385742 + ], + "logprob_reward": -0.005948815029114485, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948815029114485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26429644227027893, + 0.06412946432828903, + 0.029921522364020348, + 0.06414496898651123, + 0.045204438269138336, + 0.08796997368335724, + 0.0737471804022789, + 0.22831587493419647, + 0.08453019708395004, + 0.05774000287055969 + ], + "credit_map": [ + 0.07530288398265839, + 0.02626090496778488, + 0.09410098940134048, + 0.021363209933042526, + 0.008195297792553902, + 0.011710402555763721, + 0.015659941360354424, + 0.01170277874916792, + 0.007410354446619749, + 0.007118469104170799, + 0.005493855103850365, + 0.009599225595593452, + 0.010519156232476234, + 0.009362942539155483, + 0.009286525659263134, + 0.005338924936950207, + 0.0031067265663295984, + 0.004624996799975634, + 0.002084471518173814, + 0.0026133665814995766, + 0.0036797705106437206, + 0.0052856337279081345, + 0.005584964528679848, + 0.0029415853787213564, + 0.009054360911250114, + 0.008738625794649124, + 0.007741753943264484, + 0.012671604752540588, + 0.008093023672699928, + 0.0057020168751478195, + 0.007566242013126612, + 0.00457732891663909, + 0.010309785604476929, + 0.005753607489168644, + 0.003506965935230255, + 0.004782671108841896, + 0.00637136772274971, + 0.006451851688325405, + 0.00434647873044014, + 0.003681702772155404, + 0.01401153951883316, + 0.008287640288472176, + 0.009739134460687637, + 0.009035644121468067, + 0.01003739982843399, + 0.014239015989005566, + 0.01186386775225401, + 0.010755721479654312, + 0.009069295600056648, + 0.011382154189050198, + 0.005486255511641502, + 0.010101004503667355, + 0.008450755849480629, + 0.00887786690145731, + 0.015311111696064472, + 0.005068726371973753, + 0.05105516314506531, + 0.03826619312167168, + 0.031199881806969643, + 0.01766321063041687, + 0.011847364716231823, + 0.03868304193019867, + 0.01898360066115856, + 0.02061738446354866, + 0.0070205749943852425, + 0.009780663065612316, + 0.008300920948386192, + 0.020811647176742554, + 0.008460895158350468, + 0.00901256874203682, + 0.014375777915120125, + 0.006767142564058304, + 0.007664035074412823, + 0.00836255680769682, + 0.004088836256414652, + 0.004624775145202875, + 0.008893975988030434, + 0.006176232825964689, + 0.012649995274841785, + 0.00527958944439888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.005948815029114485 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062057919800281525 + ], + "opposite_logprob_tokens": [ + -2.8120579719543457 + ], + "logprob_reward": -0.062057919800281525, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062057919800281525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12759506702423096, + 0.05876120552420616, + 0.05318940430879593, + 0.13379181921482086, + 0.12401986122131348, + 0.0583118237555027, + 0.12027740478515625, + 0.2875666916370392, + 0.0179042536765337, + 0.018582483753561974 + ], + "credit_map": [ + 0.01803542673587799, + 0.014194604009389877, + 0.023290513083338737, + 0.01327784825116396, + 0.008210331201553345, + 0.019529007375240326, + 0.015547072514891624, + 0.01551027037203312, + 0.006152051500976086, + 0.006202060263603926, + 0.00647911848500371, + 0.008569818921387196, + 0.009447725489735603, + 0.005076115485280752, + 0.010593216866254807, + 0.006241099908947945, + 0.008295278996229172, + 0.009436333552002907, + 0.004568921402096748, + 0.004999775905162096, + 0.005900734569877386, + 0.009208743460476398, + 0.006733671762049198, + 0.0040459465235471725, + 0.018013617023825645, + 0.01663678139448166, + 0.013964930549263954, + 0.03156105428934097, + 0.014925364404916763, + 0.012825725600123405, + 0.01803533546626568, + 0.007829012349247932, + 0.036223337054252625, + 0.013996290042996407, + 0.010235579684376717, + 0.009405416436493397, + 0.011699654161930084, + 0.02010972984135151, + 0.008813225664198399, + 0.01353661809116602, + 0.013353931717574596, + 0.0071778204292058945, + 0.0057541439309716225, + 0.004787920042872429, + 0.005656353197991848, + 0.007066614460200071, + 0.007897326722741127, + 0.006617717444896698, + 0.018868878483772278, + 0.014513400383293629, + 0.011572125367820263, + 0.012841020710766315, + 0.0144716277718544, + 0.014656179584562778, + 0.0198527779430151, + 0.013501398265361786, + 0.07177232950925827, + 0.048821888864040375, + 0.02190854586660862, + 0.01797405071556568, + 0.028315335512161255, + 0.0429401732981205, + 0.030439842492341995, + 0.025394534692168236, + 0.0021964486222714186, + 0.0021955305710434914, + 0.0024971719831228256, + 0.00274035707116127, + 0.0024444120936095715, + 0.001775207812897861, + 0.002761244075372815, + 0.0012938810978084803, + 0.002092349110171199, + 0.003089958569034934, + 0.0021527402568608522, + 0.0017125108279287815, + 0.002289434662088752, + 0.0029420244973152876, + 0.0026756711304187775, + 0.0016277933027595282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.062057919800281525 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing wide shot, transitioning from a distant shot to a medium-long shot, reveals a magnificent mountaintop shimmering in the morning light. Within the frame, a rolling sea of clouds resembles a boundless white ocean, engulfing the mountainsides, with only a few dark blue peaks emerging like isolated islands. The sky shimmers delicately from deep indigo to pale white, with a warm orange-pink halo already permeating the horizon. On this secluded rocky platform, a Taoist priest, clad in a flowing off-white, wide-sleeved linen robe and a delicate bun held in place by an ancient wooden hairpin, stands with his back to the rising sun. He wields a peachwood sword, its blade a warm reddish-brown, engraved with faint, ancient runes. With his legs slightly bent and his center of gravity lowered, he dances with a steady, deliberate rhythm, the hem of his robe and its wide cuffs flowing gracefully with his movements. His movements were not swift but imbued with a restrained sense of power. Each turn, arm raise, and sword swing was precise and continuous. With a deft twist of his wrist, the tip of the peach wood sword traced clear, brilliant streaks of golden light through the cool air. These streaks seemed composed of countless luminous particles, lingering briefly in the air, forming graceful arcs and symbols before slowly dissipating like stardust. The soft morning light cast a faint golden rim around his silhouette, complementing the radiance of the sword tip. Against a cold background, the entire scene was punctuated by the warm glow of the rising sun and the golden radiance of the sword tip, creating an epic atmosphere of sacredness, tranquility, and the mystical power of the East.", + "seed": 21633, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_3/index_4841_seed_21633_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.021009888965636492, + "vlm_logprob": -0.021009888965636492, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.021009888965636492, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014754111878573895 + ], + "opposite_logprob_tokens": [ + -4.389754295349121 + ], + "logprob_reward": -0.014754111878573895, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014754111878573895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11772159487009048, + 0.045663632452487946, + 0.03732464835047722, + 0.051534853875637054, + 0.04626050218939781, + 0.20845378935337067, + 0.05356011167168617, + 0.1230672299861908, + 0.2606782913208008, + 0.055735331028699875 + ], + "credit_map": [ + 0.016982942819595337, + 0.0214706901460886, + 0.016122566536068916, + 0.012867243029177189, + 0.011752202175557613, + 0.008839725516736507, + 0.014463823288679123, + 0.0152223976328969, + 0.005594277288764715, + 0.007219481281936169, + 0.006344222463667393, + 0.007084947545081377, + 0.00413434999063611, + 0.0031244594138115644, + 0.006374020595103502, + 0.005787879694253206, + 0.0039498526602983475, + 0.0053148046135902405, + 0.007984207011759281, + 0.006312072742730379, + 0.0038339123129844666, + 0.003086138516664505, + 0.0031684774439781904, + 0.003675183979794383, + 0.005587665364146233, + 0.007069610990583897, + 0.00969669595360756, + 0.006209860555827618, + 0.006204613950103521, + 0.005485722795128822, + 0.0052649290300905704, + 0.006015754770487547, + 0.009501305408775806, + 0.00916764885187149, + 0.0068070064298808575, + 0.005387177225202322, + 0.002294936217367649, + 0.00453155068680644, + 0.004399018827825785, + 0.004171860869973898, + 0.06223508343100548, + 0.03093104064464569, + 0.04526891931891441, + 0.010176017880439758, + 0.012198197655379772, + 0.01346712838858366, + 0.015553683042526245, + 0.018623704090714455, + 0.009011482819914818, + 0.011113694868981838, + 0.007673317100852728, + 0.004173798020929098, + 0.004161208402365446, + 0.005649508908390999, + 0.00630869111046195, + 0.005468410439789295, + 0.02196706086397171, + 0.029465872794389725, + 0.02125655859708786, + 0.011378156952559948, + 0.008955537341535091, + 0.011863642372190952, + 0.010558370500802994, + 0.007622030097991228, + 0.04509937763214111, + 0.05623526871204376, + 0.050394557416439056, + 0.02023840695619583, + 0.02243199199438095, + 0.02305643819272518, + 0.022769136354327202, + 0.02045312151312828, + 0.009647419676184654, + 0.007514120079576969, + 0.008966424502432346, + 0.0044354586862027645, + 0.004969888366758823, + 0.0074310158379375935, + 0.005222904961556196, + 0.007548099849373102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.014754111878573895 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014212459325790405 + ], + "opposite_logprob_tokens": [ + -4.264212608337402 + ], + "logprob_reward": -0.014212459325790405, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014212459325790405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21369408071041107, + 0.055978622287511826, + 0.0648341104388237, + 0.05886410176753998, + 0.12198903411626816, + 0.10401776432991028, + 0.08505178242921829, + 0.07477222383022308, + 0.06647922098636627, + 0.15431910753250122 + ], + "credit_map": [ + 0.027018185704946518, + 0.04205438494682312, + 0.02730652131140232, + 0.03360912576317787, + 0.019330503419041634, + 0.013741092756390572, + 0.024294443428516388, + 0.0263398177921772, + 0.005913469009101391, + 0.006628041621297598, + 0.007170690223574638, + 0.005920963361859322, + 0.00704975938424468, + 0.007592486683279276, + 0.00778780085965991, + 0.007915415801107883, + 0.005776936654001474, + 0.007680748123675585, + 0.009506581351161003, + 0.013223770074546337, + 0.007450265344232321, + 0.005578475072979927, + 0.009027530439198017, + 0.006589805707335472, + 0.005712633486837149, + 0.005494758952409029, + 0.007338698487728834, + 0.004870608448982239, + 0.009861387312412262, + 0.013262768276035786, + 0.007254842668771744, + 0.005068402737379074, + 0.019138634204864502, + 0.013322933577001095, + 0.01598300412297249, + 0.010625652968883514, + 0.019887728616595268, + 0.019558683037757874, + 0.013712601736187935, + 0.00975980143994093, + 0.03142879530787468, + 0.013967057690024376, + 0.008792705833911896, + 0.004406566731631756, + 0.008962868712842464, + 0.01170297246426344, + 0.00830648373812437, + 0.016450311988592148, + 0.011552575044333935, + 0.007871005684137344, + 0.006559487897902727, + 0.00903223641216755, + 0.014274446293711662, + 0.007683697156608105, + 0.02140558511018753, + 0.006672749761492014, + 0.016470907256007195, + 0.009150061756372452, + 0.01111825555562973, + 0.005268717184662819, + 0.0068069505505263805, + 0.012472633272409439, + 0.009254982694983482, + 0.004229715093970299, + 0.00935385376214981, + 0.005656723398715258, + 0.006477747578173876, + 0.008283534087240696, + 0.007402955554425716, + 0.013460356742143631, + 0.008461939170956612, + 0.007382113952189684, + 0.02035479061305523, + 0.024655679240822792, + 0.01933620125055313, + 0.01735866442322731, + 0.019373813644051552, + 0.017558908089995384, + 0.023935366421937943, + 0.01174567174166441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.014212459325790405 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029821693897247314 + ], + "opposite_logprob_tokens": [ + -3.5298216342926025 + ], + "logprob_reward": -0.029821693897247314, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029821693897247314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1239924356341362, + 0.0602128766477108, + 0.04742361232638359, + 0.10259997099637985, + 0.15766683220863342, + 0.07177449017763138, + 0.05972695350646973, + 0.15031881630420685, + 0.1895008385181427, + 0.03678317368030548 + ], + "credit_map": [ + 0.014201948419213295, + 0.020530426874756813, + 0.016606450080871582, + 0.016125131398439407, + 0.011628557927906513, + 0.01207338459789753, + 0.021580420434474945, + 0.011246126145124435, + 0.006509925704449415, + 0.007487916387617588, + 0.00762941176071763, + 0.010421245358884335, + 0.0071235341019928455, + 0.00478356285020709, + 0.009837215766310692, + 0.006420062854886055, + 0.004783775191754103, + 0.005567748565226793, + 0.006260613910853863, + 0.005938306916505098, + 0.005777551792562008, + 0.007189245894551277, + 0.007255354896187782, + 0.004651012364774942, + 0.009889332577586174, + 0.01049330085515976, + 0.016262201592326164, + 0.011250212788581848, + 0.012484039179980755, + 0.015035414136946201, + 0.014472723007202148, + 0.0127127505838871, + 0.026835964992642403, + 0.017065856605768204, + 0.017726775258779526, + 0.014406015165150166, + 0.020530348643660545, + 0.02729467861354351, + 0.021353628486394882, + 0.01245356909930706, + 0.018666427582502365, + 0.00762699730694294, + 0.00929590780287981, + 0.004148007370531559, + 0.009346405044198036, + 0.005000314209610224, + 0.009133582934737206, + 0.008556842803955078, + 0.008014941588044167, + 0.006975019350647926, + 0.006039731204509735, + 0.004870464559644461, + 0.00935453362762928, + 0.0074883392080664635, + 0.012859022244811058, + 0.00412490451708436, + 0.034129004925489426, + 0.014800718985497952, + 0.029870135709643364, + 0.00874693226069212, + 0.009306415915489197, + 0.028460927307605743, + 0.01596054434776306, + 0.00904413964599371, + 0.029286092147231102, + 0.022400690242648125, + 0.026445135474205017, + 0.013865812681615353, + 0.018636498600244522, + 0.039524663239717484, + 0.015512746758759022, + 0.02382921241223812, + 0.006167601328343153, + 0.003752429038286209, + 0.005092148669064045, + 0.003094858257099986, + 0.00565964262932539, + 0.005465132184326649, + 0.004118409473448992, + 0.0034329514019191265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.029821693897247314 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319446949288249 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.00319446949288249, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319446949288249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10511747747659683, + 0.08103390038013458, + 0.03559107333421707, + 0.10497161746025085, + 0.05305857211351395, + 0.05488656088709831, + 0.06830942630767822, + 0.2122548222541809, + 0.18934616446495056, + 0.09543034434318542 + ], + "credit_map": [ + 0.011631174013018608, + 0.011232908815145493, + 0.012924778275191784, + 0.01830611377954483, + 0.011136935092508793, + 0.009585387073457241, + 0.01836300827562809, + 0.011937177740037441, + 0.007205411791801453, + 0.009472561068832874, + 0.010248997248709202, + 0.012623753398656845, + 0.008082103915512562, + 0.005748520605266094, + 0.016772190108895302, + 0.010880360379815102, + 0.00394772831350565, + 0.005269614513963461, + 0.004344069864600897, + 0.00533332210034132, + 0.0047576711513102055, + 0.0029968139715492725, + 0.004520305432379246, + 0.004421548452228308, + 0.011322631500661373, + 0.00946071743965149, + 0.026691993698477745, + 0.013673280365765095, + 0.010269786231219769, + 0.013214179314672947, + 0.011582878418266773, + 0.008756153285503387, + 0.006678324192762375, + 0.010360741056501865, + 0.006956503726541996, + 0.006058407016098499, + 0.005703755654394627, + 0.007330342661589384, + 0.005338500719517469, + 0.004631996154785156, + 0.007083899807184935, + 0.009812836535274982, + 0.009286033920943737, + 0.00382983754388988, + 0.006017616018652916, + 0.0065653277561068535, + 0.0070200115442276, + 0.005270997993648052, + 0.012031960301101208, + 0.010250023566186428, + 0.008628737181425095, + 0.007542451843619347, + 0.006340349093079567, + 0.007443761453032494, + 0.008990388363599777, + 0.007081750780344009, + 0.03228441998362541, + 0.018774842843413353, + 0.024213524535298347, + 0.016101954504847527, + 0.021002227440476418, + 0.05093162879347801, + 0.03384772315621376, + 0.015098496340215206, + 0.03162241727113724, + 0.020660320296883583, + 0.025426829233765602, + 0.013809260912239552, + 0.018523087725043297, + 0.03909372165799141, + 0.0187598317861557, + 0.0214507058262825, + 0.014525902457535267, + 0.019448263570666313, + 0.013723877258598804, + 0.0078087737783789635, + 0.009156472980976105, + 0.01248298492282629, + 0.010074496269226074, + 0.008209578692913055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.00319446949288249 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043066710233688354 + ], + "opposite_logprob_tokens": [ + -3.1680667400360107 + ], + "logprob_reward": -0.043066710233688354, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043066710233688354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11098916083574295, + 0.036211222410202026, + 0.03350694477558136, + 0.03742535784840584, + 0.04600295051932335, + 0.053959108889102936, + 0.04236230254173279, + 0.07179145514965057, + 0.510452389717102, + 0.057299111038446426 + ], + "credit_map": [ + 0.011936311610043049, + 0.015836495906114578, + 0.01469628419727087, + 0.012165492400527, + 0.011618980206549168, + 0.012402568943798542, + 0.02194192074239254, + 0.010391106829047203, + 0.0043540336191654205, + 0.0051327780820429325, + 0.004781877156347036, + 0.003717567306011915, + 0.004389586858451366, + 0.005278426688164473, + 0.0046195974573493, + 0.003937352914363146, + 0.0037427167408168316, + 0.0028330760542303324, + 0.00395707692950964, + 0.003824555315077305, + 0.004853545222431421, + 0.005312879104167223, + 0.004874911159276962, + 0.004108184948563576, + 0.0033596372231841087, + 0.005392315331846476, + 0.0053164223209023476, + 0.005833004135638475, + 0.00410144217312336, + 0.004375996068120003, + 0.005302981473505497, + 0.003743561217561364, + 0.009638911113142967, + 0.004249725490808487, + 0.005673921201378107, + 0.004490625578910112, + 0.004078986588865519, + 0.0071104709059000015, + 0.004051519557833672, + 0.0067087882198393345, + 0.012766704894602299, + 0.007548932917416096, + 0.0064450837671756744, + 0.0026322309859097004, + 0.006598555948585272, + 0.0054277293384075165, + 0.00691354600712657, + 0.005626321770250797, + 0.006593768484890461, + 0.0055031622759997845, + 0.00471173319965601, + 0.005576591473072767, + 0.005007966887205839, + 0.004534663632512093, + 0.006942122709006071, + 0.003492291085422039, + 0.014358646236360073, + 0.0077680665999650955, + 0.011023101396858692, + 0.005741267930716276, + 0.007341667544096708, + 0.012072522193193436, + 0.008618227206170559, + 0.004867957904934883, + 0.08382458239793777, + 0.050331808626651764, + 0.07649781554937363, + 0.035724349319934845, + 0.050383128225803375, + 0.10090383142232895, + 0.043035123497247696, + 0.06975176185369492, + 0.007442330941557884, + 0.006004173308610916, + 0.006131127942353487, + 0.008082032203674316, + 0.007812588475644588, + 0.004983679391443729, + 0.012023866176605225, + 0.004819315858185291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.043066710233688354 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing wide shot, transitioning from a distant shot to a medium-long shot, reveals a magnificent mountaintop shimmering in the morning light. Within the frame, a rolling sea of clouds resembles a boundless white ocean, engulfing the mountainsides, with only a few dark blue peaks emerging like isolated islands. The sky shimmers delicately from deep indigo to pale white, with a warm orange-pink halo already permeating the horizon. On this secluded rocky platform, a Taoist priest, clad in a flowing off-white, wide-sleeved linen robe and a delicate bun held in place by an ancient wooden hairpin, stands with his back to the rising sun. He wields a peachwood sword, its blade a warm reddish-brown, engraved with faint, ancient runes. With his legs slightly bent and his center of gravity lowered, he dances with a steady, deliberate rhythm, the hem of his robe and its wide cuffs flowing gracefully with his movements. His movements were not swift but imbued with a restrained sense of power. Each turn, arm raise, and sword swing was precise and continuous. With a deft twist of his wrist, the tip of the peach wood sword traced clear, brilliant streaks of golden light through the cool air. These streaks seemed composed of countless luminous particles, lingering briefly in the air, forming graceful arcs and symbols before slowly dissipating like stardust. The soft morning light cast a faint golden rim around his silhouette, complementing the radiance of the sword tip. Against a cold background, the entire scene was punctuated by the warm glow of the rising sun and the golden radiance of the sword tip, creating an epic atmosphere of sacredness, tranquility, and the mystical power of the East.", + "seed": 21634, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_3/index_4841_seed_21634_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.015442375466227532, + "vlm_logprob": -0.015442375466227532, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.015442375466227532, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006440478377044201 + ], + "opposite_logprob_tokens": [ + -5.25644063949585 + ], + "logprob_reward": -0.006440478377044201, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.006440478377044201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21925880014896393, + 0.09019791334867477, + 0.02485932782292366, + 0.035491976886987686, + 0.061307959258556366, + 0.21435914933681488, + 0.1538364589214325, + 0.02997099608182907, + 0.07837068289518356, + 0.09234676510095596 + ], + "credit_map": [ + 0.022459862753748894, + 0.019743328914046288, + 0.03294186294078827, + 0.020968690514564514, + 0.02787434123456478, + 0.03314102441072464, + 0.031747013330459595, + 0.030382679775357246, + 0.006888458505272865, + 0.00995153933763504, + 0.00904285628348589, + 0.007239070255309343, + 0.019255466759204865, + 0.013311637565493584, + 0.016555234789848328, + 0.007953650318086147, + 0.0019681949634104967, + 0.00250103697180748, + 0.002741259289905429, + 0.002744081662967801, + 0.004309165757149458, + 0.0038237967528402805, + 0.0042783706448972225, + 0.0024934227112680674, + 0.0035595004446804523, + 0.003468411508947611, + 0.003447527764365077, + 0.0036297368351370096, + 0.0040080351755023, + 0.005545126274228096, + 0.008862868882715702, + 0.0029707716312259436, + 0.005369492806494236, + 0.010803105309605598, + 0.005709121003746986, + 0.0074330526404082775, + 0.008002778515219688, + 0.008585446514189243, + 0.00883924774825573, + 0.006565713789314032, + 0.023361433297395706, + 0.046801336109638214, + 0.020381072536110878, + 0.02912045456469059, + 0.024535944685339928, + 0.032485637813806534, + 0.01833869144320488, + 0.019334569573402405, + 0.016107290983200073, + 0.026003357023000717, + 0.008283961564302444, + 0.03167214244604111, + 0.01778489165008068, + 0.01959162950515747, + 0.024951718747615814, + 0.009441482834517956, + 0.003471049014478922, + 0.003055659821256995, + 0.0028836047276854515, + 0.003646113211289048, + 0.003580548334866762, + 0.003953666891902685, + 0.0058267368003726006, + 0.0035536179784685373, + 0.008780239149928093, + 0.006883186288177967, + 0.011362220160663128, + 0.010419653728604317, + 0.009231404401361942, + 0.010270438157022, + 0.012830501422286034, + 0.008593033999204636, + 0.00821748562157154, + 0.0064210244454443455, + 0.009409540332853794, + 0.01626269891858101, + 0.010210509411990643, + 0.01190955936908722, + 0.012731321156024933, + 0.017184626311063766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.006440478377044201 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651631884276867 + ], + "opposite_logprob_tokens": [ + -4.7586517333984375 + ], + "logprob_reward": -0.008651631884276867, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651631884276867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16222307085990906, + 0.05077434331178665, + 0.032612625509500504, + 0.03891553729772568, + 0.15033239126205444, + 0.11744555085897446, + 0.2648027241230011, + 0.042072225362062454, + 0.06962134689092636, + 0.07120014727115631 + ], + "credit_map": [ + 0.01566198468208313, + 0.01336513552814722, + 0.018164321780204773, + 0.025319188833236694, + 0.015781650319695473, + 0.022840073332190514, + 0.03146964684128761, + 0.01962108351290226, + 0.005268834065645933, + 0.0054911114275455475, + 0.005103406496345997, + 0.007565067149698734, + 0.005606474354863167, + 0.006402073428034782, + 0.00918753445148468, + 0.0061498479917645454, + 0.0029830855783075094, + 0.0029389786068350077, + 0.003315247595310211, + 0.004223392810672522, + 0.005766137968748808, + 0.004391687922179699, + 0.005456697661429644, + 0.003537398763000965, + 0.0049585457891225815, + 0.0041381241753697395, + 0.004244354087859392, + 0.003834442002698779, + 0.003928045742213726, + 0.005640804301947355, + 0.008379201404750347, + 0.0037920232862234116, + 0.01237642765045166, + 0.012974979355931282, + 0.008159340359270573, + 0.01759752258658409, + 0.03458376228809357, + 0.024471646174788475, + 0.03088957443833351, + 0.009279152378439903, + 0.015389601700007915, + 0.02326587215065956, + 0.014766762964427471, + 0.01078055053949356, + 0.012752576731145382, + 0.01604936271905899, + 0.008860175497829914, + 0.015580649487674236, + 0.026572085916996002, + 0.06341533362865448, + 0.012896937318146229, + 0.050537336617708206, + 0.02288719080388546, + 0.04136126488447189, + 0.032836079597473145, + 0.014296512119472027, + 0.00494195194914937, + 0.004145685583353043, + 0.0027720232028514147, + 0.00409166282042861, + 0.0059931459836661816, + 0.007568661589175463, + 0.00871982704848051, + 0.003839271143078804, + 0.0055163707584142685, + 0.008341306820511818, + 0.004278303124010563, + 0.005815297365188599, + 0.007968464866280556, + 0.012007257901132107, + 0.018937496468424797, + 0.006756857503205538, + 0.010699235834181309, + 0.006737589370459318, + 0.006663460750132799, + 0.006959020625799894, + 0.00989128090441227, + 0.011623657308518887, + 0.011597749777138233, + 0.0070281559601426125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.008651631884276867 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026366746053099632 + ], + "opposite_logprob_tokens": [ + -3.651366710662842 + ], + "logprob_reward": -0.026366746053099632, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026366746053099632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06968770921230316, + 0.02904372289776802, + 0.04141758754849434, + 0.050978440791368484, + 0.11492288112640381, + 0.22490639984607697, + 0.2799158990383148, + 0.05020692199468613, + 0.058711305260658264, + 0.0802091658115387 + ], + "credit_map": [ + 0.006578437983989716, + 0.006180062424391508, + 0.008156543597579002, + 0.008180402219295502, + 0.008533425629138947, + 0.009989495389163494, + 0.012983594089746475, + 0.009085744619369507, + 0.003003382822498679, + 0.0033760846126824617, + 0.0019637562800198793, + 0.0030932980589568615, + 0.0035579942632466555, + 0.004755730740725994, + 0.006505089346319437, + 0.002788383513689041, + 0.005200293380767107, + 0.003686117473989725, + 0.003980549052357674, + 0.004225138574838638, + 0.005556854419410229, + 0.006362478248775005, + 0.008044658228754997, + 0.004361490253359079, + 0.004358736332505941, + 0.0033509982749819756, + 0.00851317960768938, + 0.005984755232930183, + 0.006622531451284885, + 0.007176531013101339, + 0.009242304600775242, + 0.005729395896196365, + 0.013076863251626492, + 0.012493500486016273, + 0.009333858266472816, + 0.014365234412252903, + 0.01796959526836872, + 0.01626432128250599, + 0.022209342569112778, + 0.009210154414176941, + 0.020996451377868652, + 0.05714119225740433, + 0.015791699290275574, + 0.026710782200098038, + 0.030198827385902405, + 0.036137208342552185, + 0.02008892595767975, + 0.017841288819909096, + 0.025093553587794304, + 0.07269967347383499, + 0.012632099911570549, + 0.05169752612709999, + 0.021955294534564018, + 0.04628831148147583, + 0.03400804474949837, + 0.01554136723279953, + 0.004289691336452961, + 0.0036965084727853537, + 0.0031621039379388094, + 0.004667711444199085, + 0.008417083881795406, + 0.01126309484243393, + 0.009667009115219116, + 0.005043713841587305, + 0.005361509509384632, + 0.005370395258069038, + 0.0032389038242399693, + 0.004525273106992245, + 0.0076944767497479916, + 0.011894791387021542, + 0.015074514783918858, + 0.005551436450332403, + 0.008166570216417313, + 0.007165278773754835, + 0.0061034997925162315, + 0.0054560462012887, + 0.012350043281912804, + 0.01387088093906641, + 0.013386169448494911, + 0.013710672967135906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.026366746053099632 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948103964328766 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005948103964328766, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948103964328766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07320146262645721, + 0.03277949243783951, + 0.01800917275249958, + 0.024267027154564857, + 0.01964164525270462, + 0.4142487943172455, + 0.33722126483917236, + 0.04198579862713814, + 0.01837226375937462, + 0.020272990688681602 + ], + "credit_map": [ + 0.008119854144752026, + 0.008635428734123707, + 0.005395485088229179, + 0.005857404321432114, + 0.00957888551056385, + 0.012690020725131035, + 0.017481757327914238, + 0.005442633293569088, + 0.003510342678055167, + 0.0031171964947134256, + 0.003522282699123025, + 0.0034717111848294735, + 0.004806996323168278, + 0.005070499610155821, + 0.00651299674063921, + 0.002767468336969614, + 0.0014177641132846475, + 0.0014116071397438645, + 0.0018190492410212755, + 0.002414815593510866, + 0.0020855299662798643, + 0.0025389529764652252, + 0.004520144313573837, + 0.001801312668249011, + 0.002269491320475936, + 0.0017889492446556687, + 0.0027076022233814, + 0.002912516938522458, + 0.0034541734494268894, + 0.0034603364765644073, + 0.005008803680539131, + 0.0026651558000594378, + 0.0025999306235462427, + 0.0037382107693701982, + 0.0017042921390384436, + 0.0022311287466436625, + 0.0020091317128390074, + 0.002800716320052743, + 0.002831402700394392, + 0.0017268338706344366, + 0.04328253120183945, + 0.0966179370880127, + 0.038914795964956284, + 0.05049720034003258, + 0.048278775066137314, + 0.06215939298272133, + 0.033635303378105164, + 0.04086289927363396, + 0.031152918934822083, + 0.08091142773628235, + 0.01629384234547615, + 0.06418310850858688, + 0.03087967447936535, + 0.05222449079155922, + 0.04241315647959709, + 0.019162677228450775, + 0.0039008602034300566, + 0.003867598483338952, + 0.003159309271723032, + 0.005037089344114065, + 0.0048326244577765465, + 0.007013312075287104, + 0.010850774124264717, + 0.003324235323816538, + 0.001613658620044589, + 0.0013092136941850185, + 0.0010098244529217482, + 0.0015344881685450673, + 0.0036504711024463177, + 0.0037669443991035223, + 0.0036102039739489555, + 0.0018774614436551929, + 0.0019144245889037848, + 0.0014381720684468746, + 0.0014669956872239709, + 0.001798624056391418, + 0.002512182341888547, + 0.0033267613034695387, + 0.0025691117625683546, + 0.005246720276772976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.005948103964328766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980491705238819 + ], + "opposite_logprob_tokens": [ + -3.5298049449920654 + ], + "logprob_reward": -0.02980491705238819, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980491705238819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12873795628547668, + 0.04552833363413811, + 0.038195110857486725, + 0.031589996069669724, + 0.048851240426301956, + 0.14337453246116638, + 0.38890540599823, + 0.04487873241305351, + 0.08217482268810272, + 0.0477638877928257 + ], + "credit_map": [ + 0.01631152629852295, + 0.00919842254370451, + 0.011632358655333519, + 0.016178423538804054, + 0.018112551420927048, + 0.01859450526535511, + 0.02380445785820484, + 0.014905702322721481, + 0.006121803540736437, + 0.004032114986330271, + 0.003520836355164647, + 0.004106384702026844, + 0.00752154178917408, + 0.007468753028661013, + 0.007523971609771252, + 0.005232925061136484, + 0.0038323476910591125, + 0.0033190955873578787, + 0.004118563141673803, + 0.00410571089014411, + 0.005838403943926096, + 0.005872265435755253, + 0.007045164704322815, + 0.004063554108142853, + 0.0028789672069251537, + 0.0024704416282474995, + 0.002366097643971443, + 0.0035742928739637136, + 0.003745976835489273, + 0.005142302718013525, + 0.008749532513320446, + 0.0026623818557709455, + 0.005174179561436176, + 0.004715307615697384, + 0.004558739252388477, + 0.006671062204986811, + 0.0065991985611617565, + 0.0067562502808868885, + 0.009596900083124638, + 0.00477959867566824, + 0.01757155731320381, + 0.02715948596596718, + 0.018488958477973938, + 0.017345566302537918, + 0.01440958771854639, + 0.018970191478729248, + 0.011659779585897923, + 0.01776939630508423, + 0.03336575999855995, + 0.10976909101009369, + 0.017497045919299126, + 0.0630359947681427, + 0.03169085085391998, + 0.06670244038105011, + 0.04713623225688934, + 0.01970798149704933, + 0.003956950735300779, + 0.0050697652623057365, + 0.0037856355775147676, + 0.00626167980954051, + 0.005822341423481703, + 0.00798195879906416, + 0.007859863340854645, + 0.004140531178563833, + 0.007415450178086758, + 0.00802974496036768, + 0.004631953779608011, + 0.00633509736508131, + 0.010343294590711594, + 0.015483938157558441, + 0.02207229658961296, + 0.007863042876124382, + 0.005297356750816107, + 0.003356526605784893, + 0.004033091478049755, + 0.0058389948680996895, + 0.007667220663279295, + 0.008901302702724934, + 0.0068650213070213795, + 0.00580437108874321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.02980491705238819 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing wide shot, transitioning from a distant shot to a medium-long shot, reveals a magnificent mountaintop shimmering in the morning light. Within the frame, a rolling sea of clouds resembles a boundless white ocean, engulfing the mountainsides, with only a few dark blue peaks emerging like isolated islands. The sky shimmers delicately from deep indigo to pale white, with a warm orange-pink halo already permeating the horizon. On this secluded rocky platform, a Taoist priest, clad in a flowing off-white, wide-sleeved linen robe and a delicate bun held in place by an ancient wooden hairpin, stands with his back to the rising sun. He wields a peachwood sword, its blade a warm reddish-brown, engraved with faint, ancient runes. With his legs slightly bent and his center of gravity lowered, he dances with a steady, deliberate rhythm, the hem of his robe and its wide cuffs flowing gracefully with his movements. His movements were not swift but imbued with a restrained sense of power. Each turn, arm raise, and sword swing was precise and continuous. With a deft twist of his wrist, the tip of the peach wood sword traced clear, brilliant streaks of golden light through the cool air. These streaks seemed composed of countless luminous particles, lingering briefly in the air, forming graceful arcs and symbols before slowly dissipating like stardust. The soft morning light cast a faint golden rim around his silhouette, complementing the radiance of the sword tip. Against a cold background, the entire scene was punctuated by the warm glow of the rising sun and the golden radiance of the sword tip, creating an epic atmosphere of sacredness, tranquility, and the mystical power of the East.", + "seed": 21635, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_3/index_4841_seed_21635_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.035017131734639405, + "vlm_logprob": -0.035017131734639405, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.035017131734639405, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015036571770906448 + ], + "opposite_logprob_tokens": [ + -4.390036582946777 + ], + "logprob_reward": -0.015036571770906448, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.015036571770906448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14138706028461456, + 0.10107171535491943, + 0.0621667355298996, + 0.1973959505558014, + 0.04961613193154335, + 0.19195622205734253, + 0.10824687033891678, + 0.08537021279335022, + 0.03149518743157387, + 0.03129388019442558 + ], + "credit_map": [ + 0.014745669439435005, + 0.01813080906867981, + 0.01495386753231287, + 0.019721943885087967, + 0.012408558279275894, + 0.018480202183127403, + 0.02325792796909809, + 0.019688082858920097, + 0.013061008416116238, + 0.011275466531515121, + 0.011291979812085629, + 0.02344260737299919, + 0.008540313690900803, + 0.00817197933793068, + 0.012565894983708858, + 0.012722463347017765, + 0.007503667380660772, + 0.007879081182181835, + 0.008597024716436863, + 0.00934706162661314, + 0.0044873300939798355, + 0.008219568058848381, + 0.009440726600587368, + 0.006692273076623678, + 0.019457440823316574, + 0.040186330676078796, + 0.015924477949738503, + 0.021750980988144875, + 0.026466429233551025, + 0.022595250979065895, + 0.031816039234399796, + 0.01919899880886078, + 0.005010763183236122, + 0.0030533447861671448, + 0.00511509133502841, + 0.004503512289375067, + 0.007953627035021782, + 0.0071680727414786816, + 0.010275054723024368, + 0.006536664441227913, + 0.026454897597432137, + 0.013508986681699753, + 0.013913353905081749, + 0.01968013122677803, + 0.03450324386358261, + 0.028472766280174255, + 0.02189072221517563, + 0.03353213146328926, + 0.012499922886490822, + 0.005892659537494183, + 0.008742344565689564, + 0.01723329909145832, + 0.013018219731748104, + 0.018856242299079895, + 0.014916127547621727, + 0.01708805561065674, + 0.009230704978108406, + 0.006957326550036669, + 0.008061015978455544, + 0.007145360112190247, + 0.012893510051071644, + 0.025557275861501694, + 0.0064486293122172356, + 0.009076393209397793, + 0.004746054299175739, + 0.0035110944882035255, + 0.003264177590608597, + 0.0033089141361415386, + 0.004669776652008295, + 0.0034809629432857037, + 0.005523094907402992, + 0.0029911152087152004, + 0.0068084923550486565, + 0.002309144241735339, + 0.003457738319411874, + 0.0027648084796965122, + 0.003968698438256979, + 0.004708539694547653, + 0.0023411696311086416, + 0.00493528600782156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.015036571770906448 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0233065914362669 + ], + "opposite_logprob_tokens": [ + -3.7733066082000732 + ], + "logprob_reward": -0.0233065914362669, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0233065914362669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16001112759113312, + 0.08058369159698486, + 0.09489447623491287, + 0.05478272959589958, + 0.047155410051345825, + 0.2921496033668518, + 0.0504496693611145, + 0.07870916277170181, + 0.05384115129709244, + 0.08742302656173706 + ], + "credit_map": [ + 0.01578887365758419, + 0.0238737054169178, + 0.012193986214697361, + 0.01514233835041523, + 0.018507380038499832, + 0.018565837293863297, + 0.03598928824067116, + 0.019949719309806824, + 0.010035465471446514, + 0.011341230943799019, + 0.005663339979946613, + 0.0095219686627388, + 0.013234245590865612, + 0.009081480093300343, + 0.013802354224026203, + 0.007903603836894035, + 0.009446214884519577, + 0.009749024175107479, + 0.009546265006065369, + 0.010154950432479382, + 0.009506724774837494, + 0.012325389310717583, + 0.02642989158630371, + 0.00773602444678545, + 0.005625681485980749, + 0.009887199848890305, + 0.006004838738590479, + 0.007033279165625572, + 0.006475290283560753, + 0.00634805578738451, + 0.006988856475800276, + 0.006419528741389513, + 0.00397036736831069, + 0.004526984412223101, + 0.006221756339073181, + 0.005050522740930319, + 0.00551915168762207, + 0.006400017533451319, + 0.009790153242647648, + 0.005676456727087498, + 0.03928229585289955, + 0.019214456900954247, + 0.020128317177295685, + 0.033044297248125076, + 0.05359415337443352, + 0.047624289989471436, + 0.031070424243807793, + 0.04819134622812271, + 0.00397531921043992, + 0.005280633922666311, + 0.009153973311185837, + 0.008250645361840725, + 0.005717116873711348, + 0.004306844435632229, + 0.008001488633453846, + 0.005763649009168148, + 0.007498821243643761, + 0.008831357583403587, + 0.007418772671371698, + 0.010910497047007084, + 0.011138121597468853, + 0.016376586630940437, + 0.009154104627668858, + 0.00738090556114912, + 0.004724114667624235, + 0.006837720982730389, + 0.009220863692462444, + 0.007754333317279816, + 0.006253407336771488, + 0.0064825620502233505, + 0.008348342031240463, + 0.0042198128066957, + 0.015502536669373512, + 0.006051418837159872, + 0.008282698690891266, + 0.007863761857151985, + 0.019391121342778206, + 0.00909150205552578, + 0.005829984322190285, + 0.015409995801746845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.0233065914362669 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005516439676285 + ], + "opposite_logprob_tokens": [ + -2.6950552463531494 + ], + "logprob_reward": -0.07005516439676285, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005516439676285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17364899814128876, + 0.10276462882757187, + 0.03372494503855705, + 0.0603441558778286, + 0.04384578764438629, + 0.3701651692390442, + 0.03596033155918121, + 0.048197004944086075, + 0.05700436234474182, + 0.07434467226266861 + ], + "credit_map": [ + 0.015709320083260536, + 0.024774180725216866, + 0.01490548625588417, + 0.020008636638522148, + 0.02102767676115036, + 0.027557648718357086, + 0.02993519976735115, + 0.019730830565094948, + 0.013227035291492939, + 0.014001652598381042, + 0.005729014053940773, + 0.010171713307499886, + 0.013465739786624908, + 0.01506335660815239, + 0.018913373351097107, + 0.012192737311124802, + 0.004180120304226875, + 0.003821037942543626, + 0.003960611764341593, + 0.004334421828389168, + 0.004596944898366928, + 0.005148483440279961, + 0.004399671219289303, + 0.0032836534082889557, + 0.005229775793850422, + 0.012853685766458511, + 0.0062711830250918865, + 0.006934919394552708, + 0.008352012373507023, + 0.007512150332331657, + 0.006648215930908918, + 0.006542210467159748, + 0.005469713360071182, + 0.005518586840480566, + 0.004585610702633858, + 0.0071987006813287735, + 0.005459818989038467, + 0.005367578007280827, + 0.005243056453764439, + 0.005002718418836594, + 0.05503230169415474, + 0.0287320539355278, + 0.027141613885760307, + 0.041469115763902664, + 0.059189703315496445, + 0.05008070915937424, + 0.04183385893702507, + 0.06668578833341599, + 0.0019838507287204266, + 0.0027471871580928564, + 0.005358714144676924, + 0.005806419998407364, + 0.004415290895849466, + 0.003738587722182274, + 0.007472791243344545, + 0.004437484312802553, + 0.007451870013028383, + 0.0038569211028516293, + 0.003547315951436758, + 0.005163608118891716, + 0.006304693408310413, + 0.01020326092839241, + 0.004412890877574682, + 0.0072564431466162205, + 0.005988459102809429, + 0.004860634915530682, + 0.006344212219119072, + 0.004886263515800238, + 0.007102200761437416, + 0.012375257909297943, + 0.007533437106758356, + 0.007913894020020962, + 0.023978346958756447, + 0.003952222876250744, + 0.007302219979465008, + 0.005888448562473059, + 0.008953846991062164, + 0.008403533138334751, + 0.004531227983534336, + 0.01133482065051794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.07005516439676285 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638150800019503 + ], + "opposite_logprob_tokens": [ + -5.379638195037842 + ], + "logprob_reward": -0.004638150800019503, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638150800019503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15529806911945343, + 0.08230999112129211, + 0.04512840136885643, + 0.05912328138947487, + 0.09447197616100311, + 0.25241819024086, + 0.053510796278715134, + 0.11501368135213852, + 0.05072906240820885, + 0.09199650585651398 + ], + "credit_map": [ + 0.01660723239183426, + 0.021247442811727524, + 0.013292807154357433, + 0.017653003334999084, + 0.020633796229958534, + 0.02331162616610527, + 0.02615940384566784, + 0.01639275997877121, + 0.00870297010987997, + 0.010609124787151814, + 0.005600962322205305, + 0.008636104874312878, + 0.00820989441126585, + 0.012925123795866966, + 0.018854500725865364, + 0.008771313354372978, + 0.005487811751663685, + 0.005026964005082846, + 0.0061727166175842285, + 0.004241051152348518, + 0.005052081774920225, + 0.006421615369617939, + 0.008290804922580719, + 0.004435354378074408, + 0.00769026717171073, + 0.007958735339343548, + 0.007887262850999832, + 0.007040209136903286, + 0.008477920666337013, + 0.006265441887080669, + 0.0073684691451489925, + 0.006434977520257235, + 0.006695568561553955, + 0.008299182169139385, + 0.0183951947838068, + 0.008631356991827488, + 0.011498684994876385, + 0.011391214095056057, + 0.01675974763929844, + 0.012801031582057476, + 0.023842357099056244, + 0.015294118784368038, + 0.018782051280140877, + 0.029015779495239258, + 0.062103450298309326, + 0.06332478672266006, + 0.016111472621560097, + 0.023944193497300148, + 0.005283167585730553, + 0.0036634248681366444, + 0.005380864720791578, + 0.008203981444239616, + 0.00501998420804739, + 0.007277068216353655, + 0.011707271449267864, + 0.006975035648792982, + 0.010458569973707199, + 0.010754954069852829, + 0.013761424459517002, + 0.015369555912911892, + 0.015504693612456322, + 0.024169284850358963, + 0.012122186832129955, + 0.012873017229139805, + 0.005535673815757036, + 0.004376683849841356, + 0.005528775975108147, + 0.0047399066388607025, + 0.007169781718403101, + 0.00677279569208622, + 0.00930038932710886, + 0.007305053528398275, + 0.011804040521383286, + 0.0058152577839791775, + 0.008192847482860088, + 0.007522095460444689, + 0.024362588301301003, + 0.01572742871940136, + 0.005913999862968922, + 0.012658241204917431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.004638150800019503 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded rocky mountaintop with a sea of clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Taoist priest visible wearing an off-white wide-sleeved linen robe and a bun with a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest holding a peachwood sword with a reddish-brown blade engraved with faint ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the priest perform slow, deliberate dance movements with his back to the rising sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the sword swings create visible streaks of golden light composed of luminous particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204918026924133 + ], + "opposite_logprob_tokens": [ + -2.812049150466919 + ], + "logprob_reward": -0.06204918026924133, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204918026924133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1287851631641388, + 0.07609103620052338, + 0.060797978192567825, + 0.08372297883033752, + 0.04935272037982941, + 0.31035733222961426, + 0.05820389837026596, + 0.06674201041460037, + 0.10419884324073792, + 0.061748065054416656 + ], + "credit_map": [ + 0.01156758051365614, + 0.01842671073973179, + 0.009535611607134342, + 0.011249637231230736, + 0.020274290814995766, + 0.018392402678728104, + 0.019509457051753998, + 0.019829468801617622, + 0.009411235339939594, + 0.009719375520944595, + 0.004656536970287561, + 0.009322406724095345, + 0.010190295055508614, + 0.009214225225150585, + 0.016374533995985985, + 0.0072024245746433735, + 0.004242184106260538, + 0.006324259098619223, + 0.004811489954590797, + 0.006802798248827457, + 0.010149935260415077, + 0.008859744295477867, + 0.014932924881577492, + 0.0046746390871703625, + 0.007638343144208193, + 0.01720678061246872, + 0.007475717458873987, + 0.009478289633989334, + 0.009806175716221333, + 0.009998329915106297, + 0.014125456102192402, + 0.007993871346116066, + 0.003769662231206894, + 0.0035381661728024483, + 0.003702392801642418, + 0.004688591230660677, + 0.006031353026628494, + 0.007219583261758089, + 0.014558504335582256, + 0.005844464059919119, + 0.04550066962838173, + 0.023647254332900047, + 0.021254533901810646, + 0.03747977688908577, + 0.04507865011692047, + 0.04435654729604721, + 0.03548387065529823, + 0.05755601078271866, + 0.007431510835886002, + 0.003681181464344263, + 0.006699788384139538, + 0.007934433408081532, + 0.007810639683157206, + 0.009766953065991402, + 0.008024686947464943, + 0.006854700855910778, + 0.006904610898345709, + 0.005528672598302364, + 0.007750788237899542, + 0.009894086979329586, + 0.0073508913628757, + 0.010130171664059162, + 0.007775438018143177, + 0.011407347396016121, + 0.015999091789126396, + 0.008189140819013119, + 0.013028301298618317, + 0.00797975156456232, + 0.012486201710999012, + 0.01930907368659973, + 0.013029374182224274, + 0.01417790912091732, + 0.014458714053034782, + 0.0034626706037670374, + 0.00619027903303504, + 0.004113370552659035, + 0.007012121379375458, + 0.013193994760513306, + 0.0037548604886978865, + 0.00956205278635025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.06204918026924133 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_4/index_tensor([6907, 6907])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_4/index_tensor([6907, 6907])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..866b498573800e7256306587dec7cb060b0cbef2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_4/index_tensor([6907, 6907])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle shot, hovering like a drone, focuses on a stretch of fine, freshly wet sand. In the bright, soft afternoon sunlight, the wet sand takes on a deep caramel hue, shimmering with shimmering waves. Several depressions hold crystal-clear water, reflecting the azure sky. At the center of the frame, two crabs of stark contrasting colors engage in a fierce duel: one, clad in a fiery scarlet carapace, its massive, disproportionate claws tipped with ivory white, appear exceptionally sharp; the other, with a sapphire blue shell speckled with yellow, gleams metallic in the sunlight. They lower their bodies, each holding its formidable claws high, cautiously circling each other, then suddenly lunge forward, their massive claws clashing in mid-air with a resounding snap. They kept pushing and shoving in the struggle, and every strong swing would throw tiny sand and water droplets into the air. The whole scene was full of primitive tension and tense natural drama, creating an atmosphere of vitality that was both cruel and magnificent.", + "seed": 24302, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_4/index_6907_seed_24302_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05940507724881172, + "vlm_logprob": -0.05940507724881172, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.05940507724881172, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09173925966024399 + ], + "opposite_logprob_tokens": [ + -2.8417391777038574 + ], + "logprob_reward": -0.09173925966024399, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.09173925966024399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0879460796713829, + 0.03628874942660332, + 0.01798795908689499, + 0.4083430767059326, + 0.039110831916332245, + 0.028393281623721123, + 0.29052817821502686, + 0.02334529720246792, + 0.016080236062407494, + 0.05197631195187569 + ], + "credit_map": [ + 0.020103853195905685, + 0.007811392657458782, + 0.010412635281682014, + 0.0058996486477553844, + 0.008278414607048035, + 0.017929360270500183, + 0.008444154635071754, + 0.009066620841622353, + 0.007206416688859463, + 0.004431797191500664, + 0.005643412936478853, + 0.0045655472204089165, + 0.003344703931361437, + 0.004522860981523991, + 0.0029349615797400475, + 0.003639047034084797, + 0.0028772891964763403, + 0.0016643264098092914, + 0.002798638539388776, + 0.002172265201807022, + 0.0024853243958204985, + 0.002185647375881672, + 0.0018302315147593617, + 0.001974238082766533, + 0.07489944994449615, + 0.036204781383275986, + 0.08202052116394043, + 0.038795653730630875, + 0.03668287768959999, + 0.07857631146907806, + 0.02258807234466076, + 0.03857541084289551, + 0.006475086789578199, + 0.006969498936086893, + 0.00535234622657299, + 0.005457083694636822, + 0.004024173133075237, + 0.0036040511913597584, + 0.002757893642410636, + 0.00447069900110364, + 0.004045640118420124, + 0.002853712998330593, + 0.0037846241611987352, + 0.004257637076079845, + 0.0028533856384456158, + 0.0028954416047781706, + 0.003322869073599577, + 0.004379970952868462, + 0.04518891125917435, + 0.02201695181429386, + 0.04496419429779053, + 0.02194896712899208, + 0.030562423169612885, + 0.06953299045562744, + 0.015296489000320435, + 0.04101725295186043, + 0.003686955664306879, + 0.0018776123179122806, + 0.004291422199457884, + 0.002262386493384838, + 0.002703326754271984, + 0.0024526265915483236, + 0.002064169617369771, + 0.004006796050816774, + 0.00424461904913187, + 0.001722242683172226, + 0.003017997834831476, + 0.0012498646974563599, + 0.0012138976017013192, + 0.0016015747096389532, + 0.0010441805934533477, + 0.001985859591513872, + 0.009044237434864044, + 0.007125041447579861, + 0.00724569009616971, + 0.006210531108081341, + 0.006533064879477024, + 0.004928710870444775, + 0.005732309073209763, + 0.005156729370355606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.09173925966024399 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018213175237178802 + ], + "opposite_logprob_tokens": [ + -4.018213272094727 + ], + "logprob_reward": -0.018213175237178802, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018213175237178802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16563162207603455, + 0.01618598960340023, + 0.03734525293111801, + 0.3837049901485443, + 0.02657955139875412, + 0.04523793235421181, + 0.19762027263641357, + 0.034977324306964874, + 0.024239463731646538, + 0.06847765296697617 + ], + "credit_map": [ + 0.028374457731842995, + 0.011429019272327423, + 0.017477750778198242, + 0.007128897123038769, + 0.011710233055055141, + 0.049140725284814835, + 0.01011398434638977, + 0.03025656007230282, + 0.0032883412204682827, + 0.0018427735194563866, + 0.002111600711941719, + 0.00162844592705369, + 0.0021907216869294643, + 0.0017015696503221989, + 0.0020237499848008156, + 0.0013987863203510642, + 0.0066216266714036465, + 0.0037912591360509396, + 0.006484339479357004, + 0.0036228399258106947, + 0.005090029910206795, + 0.003979588858783245, + 0.004023843444883823, + 0.0037317266687750816, + 0.06685550510883331, + 0.031434133648872375, + 0.0739477202296257, + 0.03683749958872795, + 0.035386987030506134, + 0.0798785388469696, + 0.022939037531614304, + 0.036425549536943436, + 0.008758999407291412, + 0.0020691766403615475, + 0.008130792528390884, + 0.0014004585100337863, + 0.0013185414718464017, + 0.0015130077954381704, + 0.0018939526053145528, + 0.0014946237206459045, + 0.010879193432629108, + 0.005794219207018614, + 0.00765822222456336, + 0.004020715598016977, + 0.004178820643573999, + 0.0035501790698617697, + 0.0039007102604955435, + 0.005255871918052435, + 0.03214453160762787, + 0.015903517603874207, + 0.032672569155693054, + 0.015760531648993492, + 0.0192252229899168, + 0.04482976719737053, + 0.009414894506335258, + 0.02766922116279602, + 0.006351937539875507, + 0.002676525618880987, + 0.006209142971783876, + 0.002904432127252221, + 0.0032338134478777647, + 0.005758274346590042, + 0.00258243503049016, + 0.00526076415553689, + 0.003788919886574149, + 0.00222420459613204, + 0.0031142313964664936, + 0.0020998704712837934, + 0.003049603896215558, + 0.0033262933138757944, + 0.004021414089947939, + 0.0026149265468120575, + 0.008709060959517956, + 0.0074372668750584126, + 0.012315671890974045, + 0.00768296979367733, + 0.006888118106871843, + 0.008319420740008354, + 0.0090042008087039, + 0.008120943792164326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.018213175237178802 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048666927963495255 + ], + "opposite_logprob_tokens": [ + -3.0486669540405273 + ], + "logprob_reward": -0.048666927963495255, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048666927963495255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18856468796730042, + 0.08127102255821228, + 0.05287883058190346, + 0.14104890823364258, + 0.11470236629247665, + 0.039022352546453476, + 0.10771463066339493, + 0.14161139726638794, + 0.05836564302444458, + 0.07482011616230011 + ], + "credit_map": [ + 0.027174927294254303, + 0.0135739641264081, + 0.02129225619137287, + 0.01025291346013546, + 0.016996799036860466, + 0.05491960793733597, + 0.011605123989284039, + 0.03274911642074585, + 0.016465459018945694, + 0.009362843818962574, + 0.012356707826256752, + 0.008756986819207668, + 0.008416485041379929, + 0.007207969669252634, + 0.010416875593364239, + 0.008287709206342697, + 0.006780625320971012, + 0.00458300719037652, + 0.007508350070565939, + 0.006319727748632431, + 0.006188546773046255, + 0.007458808831870556, + 0.006155074108392, + 0.007884693332016468, + 0.021182138472795486, + 0.012561898678541183, + 0.024262983351945877, + 0.0151313915848732, + 0.012907320633530617, + 0.02292526513338089, + 0.01091739721596241, + 0.021160529926419258, + 0.035601332783699036, + 0.013048856519162655, + 0.03524909168481827, + 0.009370625019073486, + 0.007611691951751709, + 0.005715090315788984, + 0.0037873315159231424, + 0.004318349063396454, + 0.007394115440547466, + 0.006461970042437315, + 0.00614159507676959, + 0.004889444913715124, + 0.0031876920256763697, + 0.00305172405205667, + 0.003213591407984495, + 0.004682222381234169, + 0.018366599455475807, + 0.009424109011888504, + 0.01492357812821865, + 0.006843860726803541, + 0.01063983328640461, + 0.025810392573475838, + 0.006388773210346699, + 0.01531748753041029, + 0.025000888854265213, + 0.01594538986682892, + 0.02002153918147087, + 0.011121902614831924, + 0.016670046374201775, + 0.024318931624293327, + 0.009954926557838917, + 0.018577784299850464, + 0.009646086022257805, + 0.0049818032421171665, + 0.0074583617970347404, + 0.0059692515060305595, + 0.0073754205368459225, + 0.005857497453689575, + 0.010992070659995079, + 0.00608515739440918, + 0.018073229119181633, + 0.007831264287233353, + 0.01226822193711996, + 0.006125335115939379, + 0.006375960074365139, + 0.008859246037900448, + 0.007810432463884354, + 0.007476433180272579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.048666927963495255 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810475766658783 + ], + "opposite_logprob_tokens": [ + -3.288104772567749 + ], + "logprob_reward": -0.03810475766658783, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810475766658783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11124544590711594, + 0.05907253548502922, + 0.041177064180374146, + 0.12727504968643188, + 0.17164817452430725, + 0.0458628386259079, + 0.20569704473018646, + 0.1288989633321762, + 0.049833498895168304, + 0.05928937718272209 + ], + "credit_map": [ + 0.019307300448417664, + 0.012070356868207455, + 0.01431144867092371, + 0.01026749424636364, + 0.010340302251279354, + 0.01578925922513008, + 0.010447478853166103, + 0.01871180720627308, + 0.01063456293195486, + 0.008346343412995338, + 0.009033683687448502, + 0.0061718011274933815, + 0.007839297875761986, + 0.0053754085674881935, + 0.0064647020772099495, + 0.00520673580467701, + 0.005777990911155939, + 0.004153551999479532, + 0.006260388996452093, + 0.004629284143447876, + 0.005329231731593609, + 0.004794223699718714, + 0.004773322027176619, + 0.005459066014736891, + 0.02114582434296608, + 0.010622121393680573, + 0.0216076597571373, + 0.012260465882718563, + 0.012388741597533226, + 0.020742418244481087, + 0.010449368506669998, + 0.018058456480503082, + 0.05899452418088913, + 0.013885091990232468, + 0.060480326414108276, + 0.011563760228455067, + 0.009724101051688194, + 0.007608938496559858, + 0.005099175497889519, + 0.004292243160307407, + 0.00960922334343195, + 0.009384953416883945, + 0.007094986736774445, + 0.006047953385859728, + 0.003338045906275511, + 0.003390355734154582, + 0.0029011659789830446, + 0.004096153657883406, + 0.03298712149262428, + 0.015750374644994736, + 0.030836962163448334, + 0.014858988113701344, + 0.021343719214200974, + 0.04964045062661171, + 0.011502404697239399, + 0.028777029365301132, + 0.019949907436966896, + 0.01382298581302166, + 0.01851229928433895, + 0.009783945046365261, + 0.016251947730779648, + 0.025137821212410927, + 0.0092011122033, + 0.016238946467638016, + 0.007591929752379656, + 0.00423326063901186, + 0.006261251866817474, + 0.005131682846695185, + 0.006540156900882721, + 0.004557812586426735, + 0.010639687068760395, + 0.004877716302871704, + 0.00966038927435875, + 0.008081505075097084, + 0.008120451122522354, + 0.007167808711528778, + 0.006521852221339941, + 0.007071906700730324, + 0.004892406053841114, + 0.007773053832352161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -0.03810475766658783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030126571655273 + ], + "opposite_logprob_tokens": [ + -2.3503012657165527 + ], + "logprob_reward": -0.10030126571655273, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030126571655273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1528061479330063, + 0.030255410820245743, + 0.04459814727306366, + 0.25551921129226685, + 0.08020264655351639, + 0.056897904723882675, + 0.20617952942848206, + 0.019983282312750816, + 0.06685496866703033, + 0.08670274168252945 + ], + "credit_map": [ + 0.020840156823396683, + 0.01024025771766901, + 0.01807977631688118, + 0.007415011525154114, + 0.011160322465002537, + 0.04962673783302307, + 0.00812224019318819, + 0.027321644127368927, + 0.006720400881022215, + 0.004642071668058634, + 0.004755616653710604, + 0.002132162218913436, + 0.003072545863687992, + 0.003345766104757786, + 0.003012614091858268, + 0.0025742340367287397, + 0.0057192095555365086, + 0.003559911623597145, + 0.0068108076229691505, + 0.006054496858268976, + 0.005219042766839266, + 0.00503145344555378, + 0.005873037967830896, + 0.006330188363790512, + 0.0501803494989872, + 0.022384611889719963, + 0.04755430668592453, + 0.023774951696395874, + 0.021724790334701538, + 0.045601386576890945, + 0.01971166953444481, + 0.024587130174040794, + 0.030821651220321655, + 0.003994532395154238, + 0.03029009699821472, + 0.003667467273771763, + 0.003006273414939642, + 0.004060443490743637, + 0.0018955236300826073, + 0.002466662786900997, + 0.011597179807722569, + 0.008515691384673119, + 0.010571111924946308, + 0.006440052762627602, + 0.004952499642968178, + 0.0044217705726623535, + 0.00430062972009182, + 0.006098967511206865, + 0.0330265536904335, + 0.014985472895205021, + 0.03611231967806816, + 0.017227038741111755, + 0.020119519904255867, + 0.04675675928592682, + 0.01001928374171257, + 0.027932574972510338, + 0.003969964571297169, + 0.0020895679481327534, + 0.004566176328808069, + 0.0016996456542983651, + 0.0021205467637628317, + 0.0017574394587427378, + 0.001726023736409843, + 0.002053917618468404, + 0.009333653375506401, + 0.005201028194278479, + 0.009704931639134884, + 0.0045031034387648106, + 0.00889587216079235, + 0.006334580481052399, + 0.014036559499800205, + 0.008845239877700806, + 0.018725117668509483, + 0.010187945328652859, + 0.01446613110601902, + 0.00962077733129263, + 0.008651242591440678, + 0.009669282473623753, + 0.006956096272915602, + 0.008426145650446415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.10030126571655273 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle shot, hovering like a drone, focuses on a stretch of fine, freshly wet sand. In the bright, soft afternoon sunlight, the wet sand takes on a deep caramel hue, shimmering with shimmering waves. Several depressions hold crystal-clear water, reflecting the azure sky. At the center of the frame, two crabs of stark contrasting colors engage in a fierce duel: one, clad in a fiery scarlet carapace, its massive, disproportionate claws tipped with ivory white, appear exceptionally sharp; the other, with a sapphire blue shell speckled with yellow, gleams metallic in the sunlight. They lower their bodies, each holding its formidable claws high, cautiously circling each other, then suddenly lunge forward, their massive claws clashing in mid-air with a resounding snap. They kept pushing and shoving in the struggle, and every strong swing would throw tiny sand and water droplets into the air. The whole scene was full of primitive tension and tense natural drama, creating an atmosphere of vitality that was both cruel and magnificent.", + "seed": 24303, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_4/index_6907_seed_24303_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06012509800493717, + "vlm_logprob": -0.06012509800493717, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.06012509800493717, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12630058825016022 + ], + "opposite_logprob_tokens": [ + -2.626300573348999 + ], + "logprob_reward": -0.12630058825016022, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.12630058825016022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08196227997541428, + 0.30753955245018005, + 0.0658629909157753, + 0.02030292898416519, + 0.35493725538253784, + 0.020536143332719803, + 0.027832161635160446, + 0.05425938591361046, + 0.03312801942229271, + 0.033639244735240936 + ], + "credit_map": [ + 0.015476252883672714, + 0.007737272884696722, + 0.012177543714642525, + 0.0052877869457006454, + 0.005252121016383171, + 0.012621674686670303, + 0.006627509370446205, + 0.016782116144895554, + 0.08199456334114075, + 0.01163925975561142, + 0.027446484193205833, + 0.009011448360979557, + 0.01338219828903675, + 0.13880406320095062, + 0.008549164980649948, + 0.016712380573153496, + 0.016312940046191216, + 0.003465075511485338, + 0.005727492738515139, + 0.0036453029606491327, + 0.0038413074798882008, + 0.024902140721678734, + 0.0028435413260012865, + 0.005125189200043678, + 0.0022846143692731857, + 0.0022725043818354607, + 0.0025527470279484987, + 0.002253900747746229, + 0.0019616475328803062, + 0.0029730075038969517, + 0.0019283830188214779, + 0.0040761251002550125, + 0.10673046112060547, + 0.013792430981993675, + 0.01601356454193592, + 0.006691290065646172, + 0.01553017646074295, + 0.16215822100639343, + 0.008733444847166538, + 0.025287684053182602, + 0.0025952132418751717, + 0.001700923196040094, + 0.002436934970319271, + 0.0026871212758123875, + 0.002216046443209052, + 0.002639187965542078, + 0.002666351618245244, + 0.0035943652037531137, + 0.003835475305095315, + 0.0031084769871085882, + 0.003616417059674859, + 0.003084590658545494, + 0.004648856818675995, + 0.002681201556697488, + 0.0024179613683372736, + 0.004439183510839939, + 0.005568859167397022, + 0.0069100745022296906, + 0.006594620179384947, + 0.01251030433923006, + 0.006393927615135908, + 0.004984954837709665, + 0.006153875030577183, + 0.0051427725702524185, + 0.0061848037876188755, + 0.004715988412499428, + 0.0049383435398340225, + 0.003487236564978957, + 0.0051251621916890144, + 0.003561656689271331, + 0.0023563362192362547, + 0.002758488990366459, + 0.007965712808072567, + 0.00515504227951169, + 0.004384374246001244, + 0.0031946727540344, + 0.0027578191366046667, + 0.0034954254515469074, + 0.0031289777252823114, + 0.003557218937203288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.12630058825016022 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865730457007885 + ], + "opposite_logprob_tokens": [ + -4.758657455444336 + ], + "logprob_reward": -0.00865730457007885, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865730457007885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13357746601104736, + 0.04158272221684456, + 0.03698499873280525, + 0.03373580425977707, + 0.6153998374938965, + 0.01767723262310028, + 0.029414650052785873, + 0.01612219400703907, + 0.05358414724469185, + 0.02192094177007675 + ], + "credit_map": [ + 0.02297123521566391, + 0.013106727041304111, + 0.023449625819921494, + 0.011140826158225536, + 0.013889273628592491, + 0.026937708258628845, + 0.01161124836653471, + 0.010470828972756863, + 0.008127610199153423, + 0.002391144633293152, + 0.005103380884975195, + 0.0024598147720098495, + 0.0026885245461016893, + 0.014548717066645622, + 0.0022782557643949986, + 0.003985279239714146, + 0.006993195973336697, + 0.0026673544198274612, + 0.0035654990933835506, + 0.0033012668136507273, + 0.004529846366494894, + 0.009003550745546818, + 0.00343360286206007, + 0.003490688279271126, + 0.003769416594877839, + 0.0045888181775808334, + 0.0043585882522165775, + 0.004056721460074186, + 0.004650886636227369, + 0.0035686632618308067, + 0.0039058777038007975, + 0.0048368386924266815, + 0.18453475832939148, + 0.02390667423605919, + 0.02791713736951351, + 0.01157632190734148, + 0.028776448220014572, + 0.2792145311832428, + 0.014729329384863377, + 0.04474463686347008, + 0.0021401357371360064, + 0.0017349018016830087, + 0.002565058646723628, + 0.0018123772460967302, + 0.0028604469262063503, + 0.002424095058813691, + 0.0015096727292984724, + 0.0026305478531867266, + 0.0039623454213142395, + 0.0035232838708907366, + 0.0031029372476041317, + 0.0037573925219476223, + 0.0039035233203321695, + 0.0027366667054593563, + 0.004109980072826147, + 0.004318523686379194, + 0.0020414169412106276, + 0.002045681932941079, + 0.0021772778127342463, + 0.002756725996732712, + 0.0016680752160027623, + 0.0017153116641566157, + 0.001954170409590006, + 0.001763535663485527, + 0.008654528297483921, + 0.006093886680901051, + 0.007020954973995686, + 0.005981327500194311, + 0.006374528631567955, + 0.007932520471513271, + 0.005578454118221998, + 0.005947950761765242, + 0.00382040673866868, + 0.00239266874268651, + 0.002688175067305565, + 0.002963193692266941, + 0.0017149497289210558, + 0.0018059501890093088, + 0.002581345848739147, + 0.003954252693802118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.00865730457007885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062059711664915085 + ], + "opposite_logprob_tokens": [ + -2.8120596408843994 + ], + "logprob_reward": -0.062059711664915085, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062059711664915085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22826619446277618, + 0.12336714565753937, + 0.04142262786626816, + 0.029352419078350067, + 0.3296866714954376, + 0.08260346204042435, + 0.03512362763285637, + 0.045761410146951675, + 0.03517043590545654, + 0.04924605414271355 + ], + "credit_map": [ + 0.026535633951425552, + 0.04229496791958809, + 0.030540192499756813, + 0.03142332658171654, + 0.01713089644908905, + 0.029893599450588226, + 0.030792295932769775, + 0.01965528540313244, + 0.029434671625494957, + 0.00973670743405819, + 0.010127316229045391, + 0.005259504076093435, + 0.007047751918435097, + 0.048726294189691544, + 0.00512324133887887, + 0.007911665365099907, + 0.007516887504607439, + 0.00426208833232522, + 0.0035529814194887877, + 0.0033978528808802366, + 0.004506622441112995, + 0.008683701045811176, + 0.005101621616631746, + 0.0044008707627654076, + 0.0037899473682045937, + 0.0030154495034366846, + 0.0035025025717914104, + 0.0044722664169967175, + 0.002558446954935789, + 0.004221276845782995, + 0.0028675557114183903, + 0.00492497393861413, + 0.09315280616283417, + 0.01845342479646206, + 0.0173798780888319, + 0.00968506932258606, + 0.019341081380844116, + 0.13992366194725037, + 0.00827945489436388, + 0.023471273481845856, + 0.010074224323034286, + 0.006890405435115099, + 0.012302634306252003, + 0.007550246547907591, + 0.011843349784612656, + 0.010470856912434101, + 0.009589597582817078, + 0.013882149010896683, + 0.006519765593111515, + 0.004652435891330242, + 0.003819621168076992, + 0.004388077650219202, + 0.0051606870256364346, + 0.0027374704368412495, + 0.0036036826204508543, + 0.00424188794568181, + 0.004362461157143116, + 0.00570279685780406, + 0.005847297143191099, + 0.010610288009047508, + 0.004868492949754, + 0.004160219803452492, + 0.006000026129186153, + 0.004209826234728098, + 0.005910264328122139, + 0.004174059722572565, + 0.004707311280071735, + 0.0037590356077998877, + 0.003926954697817564, + 0.004029867704957724, + 0.003117291023954749, + 0.0055456506088376045, + 0.009122285060584545, + 0.007166139781475067, + 0.007197396829724312, + 0.0037588481791317463, + 0.004201500676572323, + 0.005922619719058275, + 0.004842248745262623, + 0.007035014219582081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.062059711664915085 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866306856274605 + ], + "opposite_logprob_tokens": [ + -3.0486631393432617 + ], + "logprob_reward": -0.04866306856274605, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866306856274605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1243259459733963, + 0.19500625133514404, + 0.03781010955572128, + 0.034555863589048386, + 0.3878297507762909, + 0.04862501099705696, + 0.02514808252453804, + 0.03278488293290138, + 0.017960388213396072, + 0.09595376998186111 + ], + "credit_map": [ + 0.01672043837606907, + 0.012817053124308586, + 0.01599848084151745, + 0.013234279118478298, + 0.010438828729093075, + 0.01820223219692707, + 0.02005082555115223, + 0.016863804310560226, + 0.050157900899648666, + 0.008783213794231415, + 0.01826968975365162, + 0.007584400940686464, + 0.008126549422740936, + 0.08227338641881943, + 0.00726603576913476, + 0.012545046396553516, + 0.0073115285485982895, + 0.0033602484036237, + 0.0031935889273881912, + 0.002731389831751585, + 0.005149603821337223, + 0.008990277536213398, + 0.003749249270185828, + 0.003324219025671482, + 0.005959556438028812, + 0.005969345103949308, + 0.003992429003119469, + 0.004873521625995636, + 0.002619261620566249, + 0.004973975475877523, + 0.00277020875364542, + 0.0033975609112530947, + 0.11033883690834045, + 0.017849694937467575, + 0.018254511058330536, + 0.008313174359500408, + 0.02829732932150364, + 0.164901465177536, + 0.010997384786605835, + 0.02887731045484543, + 0.006935714278370142, + 0.007731215562671423, + 0.006758476607501507, + 0.0047398158349096775, + 0.007642524316906929, + 0.005444480571895838, + 0.0037717055529356003, + 0.005601074080914259, + 0.0032672567758709192, + 0.003663967363536358, + 0.0034250628668814898, + 0.0027857832610607147, + 0.0029973131604492664, + 0.0028204945847392082, + 0.0023888477589935064, + 0.003799352329224348, + 0.0030687188263982534, + 0.004142537247389555, + 0.004442035220563412, + 0.006849620491266251, + 0.0039368486031889915, + 0.0030223564244806767, + 0.0040053692646324635, + 0.003317394061014056, + 0.0021254047751426697, + 0.0020548354368656874, + 0.002696060109883547, + 0.00220073782838881, + 0.0030344719998538494, + 0.0014115110971033573, + 0.0020180719438940287, + 0.0024192926939576864, + 0.01912798546254635, + 0.012482765130698681, + 0.012617923319339752, + 0.007812025491148233, + 0.007274889852851629, + 0.011550421826541424, + 0.00802893377840519, + 0.01705881580710411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.04866306856274605 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494481697678566 + ], + "opposite_logprob_tokens": [ + -2.9299447536468506 + ], + "logprob_reward": -0.05494481697678566, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494481697678566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06974922120571136, + 0.2000126838684082, + 0.07534808665513992, + 0.06614150106906891, + 0.38486185669898987, + 0.03704012930393219, + 0.030012236908078194, + 0.062044281512498856, + 0.037826523184776306, + 0.0369635708630085 + ], + "credit_map": [ + 0.013039696030318737, + 0.008486256934702396, + 0.010156839154660702, + 0.006325682159513235, + 0.0062435283325612545, + 0.012858225964009762, + 0.0050908904522657394, + 0.007548091001808643, + 0.0518687479197979, + 0.008495467714965343, + 0.017311695963144302, + 0.006509748287498951, + 0.007993160746991634, + 0.08827503770589828, + 0.007240876089781523, + 0.012317929416894913, + 0.01932205632328987, + 0.0037012703251093626, + 0.006382325664162636, + 0.004039239138364792, + 0.0039827520959079266, + 0.02807159349322319, + 0.0034573955927044153, + 0.006391439121216536, + 0.007694698870182037, + 0.008074615150690079, + 0.009134272113442421, + 0.006188470404595137, + 0.006510406732559204, + 0.007393836509436369, + 0.0064543732441961765, + 0.014690813608467579, + 0.1164000928401947, + 0.014169907197356224, + 0.01687159202992916, + 0.007394059095531702, + 0.01676269620656967, + 0.17541782557964325, + 0.00945920217782259, + 0.028386441990733147, + 0.004635446704924107, + 0.003948873374611139, + 0.006945949979126453, + 0.004328310023993254, + 0.004240515176206827, + 0.005015266593545675, + 0.0037759996484965086, + 0.00414976105093956, + 0.003977484535425901, + 0.0034776541870087385, + 0.003093387233093381, + 0.0031718588434159756, + 0.006589984986931086, + 0.0026529193855822086, + 0.004183211363852024, + 0.0028657326474785805, + 0.007003874517977238, + 0.008664305321872234, + 0.008750666864216328, + 0.009627466090023518, + 0.008611835539340973, + 0.005961300805211067, + 0.0070818630047142506, + 0.00634295679628849, + 0.005434076767414808, + 0.00460640387609601, + 0.00479560112580657, + 0.004419301170855761, + 0.005078022833913565, + 0.006032145116478205, + 0.003471826668828726, + 0.003989137709140778, + 0.007172317709773779, + 0.00518453773111105, + 0.0054716612212359905, + 0.0025582301896065474, + 0.003144564339891076, + 0.004317770712077618, + 0.0037757691461592913, + 0.005338712129741907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.05494481697678566 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle shot, hovering like a drone, focuses on a stretch of fine, freshly wet sand. In the bright, soft afternoon sunlight, the wet sand takes on a deep caramel hue, shimmering with shimmering waves. Several depressions hold crystal-clear water, reflecting the azure sky. At the center of the frame, two crabs of stark contrasting colors engage in a fierce duel: one, clad in a fiery scarlet carapace, its massive, disproportionate claws tipped with ivory white, appear exceptionally sharp; the other, with a sapphire blue shell speckled with yellow, gleams metallic in the sunlight. They lower their bodies, each holding its formidable claws high, cautiously circling each other, then suddenly lunge forward, their massive claws clashing in mid-air with a resounding snap. They kept pushing and shoving in the struggle, and every strong swing would throw tiny sand and water droplets into the air. The whole scene was full of primitive tension and tense natural drama, creating an atmosphere of vitality that was both cruel and magnificent.", + "seed": 24304, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_4/index_6907_seed_24304_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05895472802221775, + "vlm_logprob": -0.05895472802221775, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.05895472802221775, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13871927559375763 + ], + "opposite_logprob_tokens": [ + -2.763719320297241 + ], + "logprob_reward": -0.13871927559375763, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.13871927559375763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15902204811573029, + 0.09638118743896484, + 0.1473938524723053, + 0.07890348136425018, + 0.09878426790237427, + 0.11306820809841156, + 0.10972688347101212, + 0.029651472344994545, + 0.08575224131345749, + 0.08131635934114456 + ], + "credit_map": [ + 0.024507518857717514, + 0.02005673013627529, + 0.017436781898140907, + 0.0176425501704216, + 0.030686799436807632, + 0.01910838484764099, + 0.012150716967880726, + 0.017432566732168198, + 0.01386770885437727, + 0.010215400718152523, + 0.008902202360332012, + 0.014438704587519169, + 0.016241872683167458, + 0.010833453387022018, + 0.011951393447816372, + 0.009930453263223171, + 0.027620559558272362, + 0.01563091017305851, + 0.010412188246846199, + 0.010183775797486305, + 0.025658752769231796, + 0.03402066230773926, + 0.013442927040159702, + 0.010424069128930569, + 0.011307577602565289, + 0.011305569671094418, + 0.008311527781188488, + 0.012334643863141537, + 0.010376117192208767, + 0.009515802375972271, + 0.00810590572655201, + 0.0076463413424789906, + 0.011264236643910408, + 0.013760900124907494, + 0.010473014786839485, + 0.016122890636324883, + 0.010301010683178902, + 0.013303877785801888, + 0.011501532047986984, + 0.012056805193424225, + 0.016787851229310036, + 0.014503725804388523, + 0.015424405224621296, + 0.013950093649327755, + 0.010784484446048737, + 0.012900476343929768, + 0.015082432888448238, + 0.01363474689424038, + 0.013759936206042767, + 0.014685846865177155, + 0.011568567715585232, + 0.00942046009004116, + 0.011890776455402374, + 0.012998620979487896, + 0.02391006052494049, + 0.011492620222270489, + 0.0043893917463719845, + 0.003913085907697678, + 0.003968545701354742, + 0.00302805844694376, + 0.003376638749614358, + 0.003976232837885618, + 0.003761557163670659, + 0.0032379631884396076, + 0.022745897993445396, + 0.010915146209299564, + 0.01227937638759613, + 0.005830321926623583, + 0.006923486012965441, + 0.013184340670704842, + 0.005494647193700075, + 0.00837902445346117, + 0.010248558595776558, + 0.010232413187623024, + 0.012434796430170536, + 0.008344191126525402, + 0.010605749674141407, + 0.012841065414249897, + 0.008570346981287003, + 0.008039237931370735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.13871927559375763 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020612435415387154 + ], + "opposite_logprob_tokens": [ + -3.8956124782562256 + ], + "logprob_reward": -0.020612435415387154, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020612435415387154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19804802536964417, + 0.17283649742603302, + 0.11666256934404373, + 0.10188252478837967, + 0.08341330289840698, + 0.08833441883325577, + 0.049886055290699005, + 0.06850656867027283, + 0.05624207481741905, + 0.06418801099061966 + ], + "credit_map": [ + 0.036896392703056335, + 0.02179931104183197, + 0.014218246564269066, + 0.016571933403611183, + 0.04878930747509003, + 0.027969250455498695, + 0.01136417780071497, + 0.020439371466636658, + 0.029243307188153267, + 0.014209614135324955, + 0.022395728155970573, + 0.02690732292830944, + 0.02632797695696354, + 0.017066843807697296, + 0.01649424247443676, + 0.020191431045532227, + 0.015492644160985947, + 0.01165736559778452, + 0.007997412234544754, + 0.010797268711030483, + 0.03076792322099209, + 0.020906733348965645, + 0.009556618519127369, + 0.009486573748290539, + 0.01613464578986168, + 0.012924490496516228, + 0.014635744504630566, + 0.008650499396026134, + 0.015522398054599762, + 0.013762298971414566, + 0.012877978384494781, + 0.007374459411948919, + 0.012975888326764107, + 0.009789799340069294, + 0.01248476654291153, + 0.010082065127789974, + 0.009158388711512089, + 0.01186511479318142, + 0.009442941285669804, + 0.007614323869347572, + 0.00977769959717989, + 0.007465029135346413, + 0.014173552393913269, + 0.009995510801672935, + 0.008588346652686596, + 0.012726725079119205, + 0.01572638936340809, + 0.009881152771413326, + 0.005506435409188271, + 0.004391317721456289, + 0.008278879337012768, + 0.00439448794350028, + 0.0051699671894311905, + 0.005624857265502214, + 0.010788985528051853, + 0.005731115583330393, + 0.007142108865082264, + 0.0072703189216554165, + 0.014410508796572685, + 0.007665209472179413, + 0.006012224592268467, + 0.009259411133825779, + 0.011131630279123783, + 0.00561514450237155, + 0.009955394081771374, + 0.0060116080567240715, + 0.009569740854203701, + 0.0055257962085306644, + 0.004533902741968632, + 0.007291293703019619, + 0.005558631848543882, + 0.007795699406415224, + 0.007108159828931093, + 0.007059507071971893, + 0.007122097071260214, + 0.009829486720263958, + 0.010736186988651752, + 0.007601186167448759, + 0.01032086368650198, + 0.004410517867654562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.020612435415387154 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812805563211441 + ], + "opposite_logprob_tokens": [ + -3.288128137588501 + ], + "logprob_reward": -0.03812805563211441, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812805563211441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11995776742696762, + 0.11491841077804565, + 0.377483069896698, + 0.07080869376659393, + 0.07570318132638931, + 0.050952136516571045, + 0.03777100890874863, + 0.028049657121300697, + 0.07633426040410995, + 0.048021771013736725 + ], + "credit_map": [ + 0.014748864807188511, + 0.011492114514112473, + 0.012544603087008, + 0.0173336248844862, + 0.02591574192047119, + 0.013524745590984821, + 0.007568132597953081, + 0.016829954460263252, + 0.021967042237520218, + 0.010346757248044014, + 0.008769278414547443, + 0.01031981036067009, + 0.028152382001280785, + 0.014576906338334084, + 0.011440390720963478, + 0.00934585276991129, + 0.03552159667015076, + 0.045580875128507614, + 0.04694822058081627, + 0.08428676426410675, + 0.03276226669549942, + 0.06866157799959183, + 0.03552412986755371, + 0.028197655454277992, + 0.01466590166091919, + 0.007604433223605156, + 0.010628725402057171, + 0.006039836443960667, + 0.011083191260695457, + 0.009472401812672615, + 0.0063677444122731686, + 0.004946461878716946, + 0.011611497029662132, + 0.005607908125966787, + 0.010680342093110085, + 0.010233418084681034, + 0.006895096506923437, + 0.01099866721779108, + 0.011490906588733196, + 0.008185350336134434, + 0.006725397426635027, + 0.003901373129338026, + 0.008650673553347588, + 0.005764674860984087, + 0.004056551493704319, + 0.008237304165959358, + 0.008073157630860806, + 0.005543002858757973, + 0.003655393607914448, + 0.004125108011066914, + 0.00506423506885767, + 0.0036723187658935785, + 0.0037593788001686335, + 0.004339477512985468, + 0.008574228733778, + 0.004580870736390352, + 0.004396084696054459, + 0.0028582436498254538, + 0.003175960620865226, + 0.002792051760479808, + 0.002689902437850833, + 0.0028442959301173687, + 0.006511038169264793, + 0.0027820803225040436, + 0.012587414123117924, + 0.0066695972345769405, + 0.014657747931778431, + 0.006902500055730343, + 0.006602296605706215, + 0.010007031261920929, + 0.007604571990668774, + 0.011303109116852283, + 0.006399098318070173, + 0.006103042047470808, + 0.006021466571837664, + 0.008404799737036228, + 0.00458049401640892, + 0.00610010139644146, + 0.00588472094386816, + 0.0045280493795871735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.03812805563211441 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866760969161987 + ], + "opposite_logprob_tokens": [ + -3.0486676692962646 + ], + "logprob_reward": -0.04866760969161987, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866760969161987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1291627585887909, + 0.10089068859815598, + 0.15086080133914948, + 0.08489957451820374, + 0.13533563911914825, + 0.08451797068119049, + 0.07988347113132477, + 0.054464712738990784, + 0.07663478702306747, + 0.10334966331720352 + ], + "credit_map": [ + 0.022040601819753647, + 0.014594853855669498, + 0.011878536082804203, + 0.010459013283252716, + 0.03275927156209946, + 0.017238205298781395, + 0.007149118930101395, + 0.01304314099252224, + 0.021436382085084915, + 0.00912430789321661, + 0.009069709107279778, + 0.013049792498350143, + 0.016085373237729073, + 0.011674287728965282, + 0.011182175017893314, + 0.009268657304346561, + 0.014991426840424538, + 0.017274633049964905, + 0.015907760709524155, + 0.025829214602708817, + 0.029753493145108223, + 0.024812966585159302, + 0.012263184413313866, + 0.01002811174839735, + 0.019263673573732376, + 0.008818776346743107, + 0.013178695924580097, + 0.006500537041574717, + 0.011238672770559788, + 0.012001910246908665, + 0.008252222090959549, + 0.005645086523145437, + 0.020098915323615074, + 0.009005938656628132, + 0.019253946840763092, + 0.02045648731291294, + 0.010484321974217892, + 0.01889920048415661, + 0.02136043831706047, + 0.0157763808965683, + 0.009356939233839512, + 0.007904788479208946, + 0.011951697058975697, + 0.00991835817694664, + 0.0073744188994169235, + 0.012039214372634888, + 0.01604648120701313, + 0.009926071390509605, + 0.007934561930596828, + 0.008212661370635033, + 0.009381543844938278, + 0.007224337197840214, + 0.01041520107537508, + 0.008119206875562668, + 0.01943991892039776, + 0.009156035259366035, + 0.0067428830079734325, + 0.005731496028602123, + 0.009303940460085869, + 0.004945182241499424, + 0.004317505285143852, + 0.006183920428156853, + 0.01156377699226141, + 0.005676002241671085, + 0.011057296767830849, + 0.006148683372884989, + 0.015302103012800217, + 0.007260565645992756, + 0.006994742900133133, + 0.010327745229005814, + 0.007474782411009073, + 0.0120688546448946, + 0.01351187564432621, + 0.011707103811204433, + 0.015366833657026291, + 0.014603327959775925, + 0.013467199169099331, + 0.014968029223382473, + 0.01291187759488821, + 0.006813419982790947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.04866760969161987 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048646263778209686 + ], + "opposite_logprob_tokens": [ + -3.0486462116241455 + ], + "logprob_reward": -0.048646263778209686, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048646263778209686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13019110262393951, + 0.15453100204467773, + 0.3369438350200653, + 0.08604555577039719, + 0.0458945631980896, + 0.07065007835626602, + 0.033252816647291183, + 0.03821185231208801, + 0.04678216576576233, + 0.05749708041548729 + ], + "credit_map": [ + 0.023102330043911934, + 0.014961873181164265, + 0.012538653798401356, + 0.010302786715328693, + 0.029656222090125084, + 0.01759011298418045, + 0.009338079951703548, + 0.012701033614575863, + 0.042339298874139786, + 0.014876803383231163, + 0.014164324849843979, + 0.015063106082379818, + 0.02633797563612461, + 0.01914268732070923, + 0.010574079118669033, + 0.012032706290483475, + 0.04258158430457115, + 0.035652875900268555, + 0.025173937901854515, + 0.026349056512117386, + 0.10276629030704498, + 0.050558239221572876, + 0.029023433104157448, + 0.024838389828801155, + 0.015972688794136047, + 0.010439288802444935, + 0.012446670792996883, + 0.00840185210108757, + 0.011847218498587608, + 0.01066020131111145, + 0.00919080525636673, + 0.007086823228746653, + 0.00636221282184124, + 0.005211126524955034, + 0.006839489098638296, + 0.0058560981415212154, + 0.004793180618435144, + 0.006573862861841917, + 0.005559922195971012, + 0.004698667675256729, + 0.012451756745576859, + 0.007662096992135048, + 0.005895897280424833, + 0.01161725539714098, + 0.007827448658645153, + 0.008171762339770794, + 0.00981703121215105, + 0.007206820882856846, + 0.004174161236733198, + 0.004118587356060743, + 0.005724537186324596, + 0.0037568064872175455, + 0.0026801915373653173, + 0.0044392733834683895, + 0.004272259771823883, + 0.0040869987569749355, + 0.005300260614603758, + 0.0041532963514328, + 0.007034637033939362, + 0.002992367371916771, + 0.005469386465847492, + 0.005351453088223934, + 0.004216659814119339, + 0.0036937855184078217, + 0.0075347344391047955, + 0.005267556291073561, + 0.006470792926847935, + 0.00540988938882947, + 0.0035409219563007355, + 0.00602106936275959, + 0.006770511157810688, + 0.005766687914729118, + 0.007221235893666744, + 0.005543429404497147, + 0.008786972612142563, + 0.005858584772795439, + 0.00794941931962967, + 0.009545033797621727, + 0.006727966479957104, + 0.005864436272531748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.048646263778209686 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle shot, hovering like a drone, focuses on a stretch of fine, freshly wet sand. In the bright, soft afternoon sunlight, the wet sand takes on a deep caramel hue, shimmering with shimmering waves. Several depressions hold crystal-clear water, reflecting the azure sky. At the center of the frame, two crabs of stark contrasting colors engage in a fierce duel: one, clad in a fiery scarlet carapace, its massive, disproportionate claws tipped with ivory white, appear exceptionally sharp; the other, with a sapphire blue shell speckled with yellow, gleams metallic in the sunlight. They lower their bodies, each holding its formidable claws high, cautiously circling each other, then suddenly lunge forward, their massive claws clashing in mid-air with a resounding snap. They kept pushing and shoving in the struggle, and every strong swing would throw tiny sand and water droplets into the air. The whole scene was full of primitive tension and tense natural drama, creating an atmosphere of vitality that was both cruel and magnificent.", + "seed": 24305, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_4/index_6907_seed_24305_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04431265294551849, + "vlm_logprob": -0.04431265294551849, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.04431265294551849, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1455834060907364 + ], + "opposite_logprob_tokens": [ + -2.270583391189575 + ], + "logprob_reward": -0.1455834060907364, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1455834060907364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07764064520597458, + 0.02729642577469349, + 0.03569154813885689, + 0.0679185539484024, + 0.1955747902393341, + 0.048016034066677094, + 0.32215824723243713, + 0.0357862152159214, + 0.08721453696489334, + 0.10270298272371292 + ], + "credit_map": [ + 0.009809550829231739, + 0.0066976952366530895, + 0.008531437255442142, + 0.015804212540388107, + 0.009851356036961079, + 0.008801446296274662, + 0.006784351076930761, + 0.011360598728060722, + 0.004639291670173407, + 0.002308021066710353, + 0.0028153816238045692, + 0.0023963621351867914, + 0.004737709648907185, + 0.003936918918043375, + 0.003007960971444845, + 0.003454779041931033, + 0.00445201201364398, + 0.003369713667780161, + 0.0041328067891299725, + 0.004697585478425026, + 0.005038404371589422, + 0.004483743105083704, + 0.005683881230652332, + 0.0038334045093506575, + 0.02032887376844883, + 0.005862127058207989, + 0.017155852168798447, + 0.006472669541835785, + 0.004667285829782486, + 0.004855702631175518, + 0.004244121722877026, + 0.004331924952566624, + 0.02132805436849594, + 0.01643882319331169, + 0.037580665200948715, + 0.030360549688339233, + 0.018741441890597343, + 0.020221678540110588, + 0.02702520415186882, + 0.023878373205661774, + 0.008409609086811543, + 0.006201537791639566, + 0.009489256888628006, + 0.0074240476824343204, + 0.004690119065344334, + 0.005489668808877468, + 0.0028561854269355536, + 0.0034556121099740267, + 0.1245989128947258, + 0.02154216356575489, + 0.09876304864883423, + 0.019457347691059113, + 0.014192021451890469, + 0.018692336976528168, + 0.011667429469525814, + 0.013244973495602608, + 0.006130588706582785, + 0.0053153508342802525, + 0.005297002382576466, + 0.005054872948676348, + 0.0038982152473181486, + 0.002525141928344965, + 0.004490706138312817, + 0.0030743384268134832, + 0.012700927443802357, + 0.014539793133735657, + 0.014137410558760166, + 0.007890873588621616, + 0.008937804959714413, + 0.013601182959973812, + 0.007024869322776794, + 0.008381681516766548, + 0.019007792696356773, + 0.013397589325904846, + 0.016541695222258568, + 0.009949593804776669, + 0.008388792164623737, + 0.015568861737847328, + 0.009771156124770641, + 0.010077504441142082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -0.1455834060907364 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014218922704458237 + ], + "opposite_logprob_tokens": [ + -4.264218807220459 + ], + "logprob_reward": -0.014218922704458237, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014218922704458237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08928364515304565, + 0.07823126018047333, + 0.15764856338500977, + 0.08780860155820847, + 0.08969106525182724, + 0.10303197056055069, + 0.09987287223339081, + 0.05405447259545326, + 0.14893557131290436, + 0.09144189953804016 + ], + "credit_map": [ + 0.01468559168279171, + 0.013607700355350971, + 0.008448801934719086, + 0.009901045821607113, + 0.010397649370133877, + 0.011297784745693207, + 0.008743933402001858, + 0.0122011573985219, + 0.009162822738289833, + 0.006025474518537521, + 0.011942803859710693, + 0.012220711447298527, + 0.009974496439099312, + 0.009105075150728226, + 0.011237162165343761, + 0.008562731556594372, + 0.019320355728268623, + 0.013921123929321766, + 0.016603542491793633, + 0.017664514482021332, + 0.020208217203617096, + 0.028431478887796402, + 0.02383132092654705, + 0.017668044194579124, + 0.026704609394073486, + 0.005972758866846561, + 0.02237579971551895, + 0.007443042006343603, + 0.0059392270632088184, + 0.007174628786742687, + 0.006569403689354658, + 0.005629139952361584, + 0.01066664233803749, + 0.006903608329594135, + 0.015779532492160797, + 0.014089704491198063, + 0.008485439233481884, + 0.010918262414634228, + 0.012135818600654602, + 0.010712070390582085, + 0.015906980261206627, + 0.014094569720327854, + 0.018701327964663506, + 0.01230345107614994, + 0.01605413667857647, + 0.01211530901491642, + 0.0065780868753790855, + 0.007278122007846832, + 0.028399625793099403, + 0.009223656728863716, + 0.022074125707149506, + 0.006252289284020662, + 0.008708547800779343, + 0.010503008961677551, + 0.005544462241232395, + 0.009167172946035862, + 0.006636904086917639, + 0.006489301100373268, + 0.007666336838155985, + 0.008728760294616222, + 0.005935333669185638, + 0.005120474845170975, + 0.0071886442601680756, + 0.006288726814091206, + 0.022186096757650375, + 0.02740296721458435, + 0.023302201181650162, + 0.014384466223418713, + 0.0162770114839077, + 0.019903698936104774, + 0.011404833756387234, + 0.014074317179620266, + 0.013424486853182316, + 0.012620709836483002, + 0.01525773387402296, + 0.00997147150337696, + 0.007531765382736921, + 0.016020553186535835, + 0.008460232056677341, + 0.008154957555234432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 5 + ], + "selected_reward": -0.014218922704458237 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596787333488464 + ], + "opposite_logprob_tokens": [ + -3.895596742630005 + ], + "logprob_reward": -0.020596787333488464, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596787333488464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22700445353984833, + 0.09047254920005798, + 0.12215720862150192, + 0.09082558751106262, + 0.05684371665120125, + 0.038542553782463074, + 0.15976868569850922, + 0.050387218594551086, + 0.04238813370466232, + 0.12160980701446533 + ], + "credit_map": [ + 0.03135690465569496, + 0.030611945316195488, + 0.029528550803661346, + 0.04235854372382164, + 0.022772500291466713, + 0.027472643181681633, + 0.01817866787314415, + 0.024724720045924187, + 0.01933545246720314, + 0.01028070505708456, + 0.009838847443461418, + 0.00920200813561678, + 0.012002373114228249, + 0.012938877567648888, + 0.008356641046702862, + 0.008517657406628132, + 0.01706046238541603, + 0.013628014363348484, + 0.012114117853343487, + 0.011700914241373539, + 0.016511430963873863, + 0.020355040207505226, + 0.018387962132692337, + 0.012399277649819851, + 0.031764183193445206, + 0.005211392883211374, + 0.025621216744184494, + 0.006437303964048624, + 0.0057333000004291534, + 0.00626315176486969, + 0.0037603708915412426, + 0.006034678779542446, + 0.007713228464126587, + 0.007770138792693615, + 0.009184802882373333, + 0.008017130196094513, + 0.0062591200694441795, + 0.004367063753306866, + 0.006929825060069561, + 0.006602417677640915, + 0.007243135944008827, + 0.004192461725324392, + 0.0073184543289244175, + 0.004423991777002811, + 0.0052228327840566635, + 0.0044809794053435326, + 0.00227370229549706, + 0.0033870041370391846, + 0.0627753734588623, + 0.009528557769954205, + 0.050332508981227875, + 0.009034549817442894, + 0.008449443615972996, + 0.008184527978301048, + 0.006215323228389025, + 0.005248405504971743, + 0.009212343022227287, + 0.009105686098337173, + 0.007972260005772114, + 0.0058733681216835976, + 0.0036874590441584587, + 0.003830966306850314, + 0.004628651775419712, + 0.006076493766158819, + 0.007716149091720581, + 0.006162305362522602, + 0.007603172678500414, + 0.003443219931796193, + 0.003294163616374135, + 0.007347590755671263, + 0.0027344857808202505, + 0.0040870532393455505, + 0.021425500512123108, + 0.017662158235907555, + 0.019059989601373672, + 0.009939209558069706, + 0.009981470182538033, + 0.024698302149772644, + 0.008788137696683407, + 0.010055048391222954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.020596787333488464 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584410056471825 + ], + "opposite_logprob_tokens": [ + -3.8955843448638916 + ], + "logprob_reward": -0.020584410056471825, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584410056471825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1077975407242775, + 0.055417250841856, + 0.13501359522342682, + 0.12349597364664078, + 0.030985161662101746, + 0.1548221856355667, + 0.10423460602760315, + 0.060897428542375565, + 0.08426404744386673, + 0.14307217299938202 + ], + "credit_map": [ + 0.016726313158869743, + 0.016488678753376007, + 0.014673870988190174, + 0.01742224395275116, + 0.012570632621645927, + 0.015430367551743984, + 0.0072816782630980015, + 0.007203761488199234, + 0.012769180350005627, + 0.00459168991073966, + 0.006111038848757744, + 0.00640323618426919, + 0.007546458393335342, + 0.00690966984257102, + 0.006015091668814421, + 0.005070890299975872, + 0.020718680694699287, + 0.01107745710760355, + 0.013211923651397228, + 0.012219615280628204, + 0.01923292502760887, + 0.0263237114995718, + 0.019284849986433983, + 0.012944445013999939, + 0.04219542816281319, + 0.007358003873378038, + 0.03270798176527023, + 0.008930418640375137, + 0.008127300068736076, + 0.009040080942213535, + 0.006370052695274353, + 0.00876670703291893, + 0.0036796433851122856, + 0.004671076778322458, + 0.0050608497112989426, + 0.00366626912727952, + 0.0027350622694939375, + 0.004142819903790951, + 0.0035141324624419212, + 0.0035153075587004423, + 0.02594703435897827, + 0.016666630282998085, + 0.030364591628313065, + 0.018947098404169083, + 0.020739739760756493, + 0.020366011187434196, + 0.008276943117380142, + 0.013514140620827675, + 0.01902269385755062, + 0.01292606070637703, + 0.014534831047058105, + 0.012728678993880749, + 0.011195902712643147, + 0.01303913164883852, + 0.00794404186308384, + 0.012843267060816288, + 0.013996021822094917, + 0.01156650111079216, + 0.009241201914846897, + 0.006586098577827215, + 0.005533737130463123, + 0.0035748861264437437, + 0.005360486917197704, + 0.00503849983215332, + 0.01251857727766037, + 0.015578284859657288, + 0.01326976902782917, + 0.008284265175461769, + 0.009650451131165028, + 0.01096999179571867, + 0.006204071920365095, + 0.007788649294525385, + 0.023460185155272484, + 0.020643651485443115, + 0.022412464022636414, + 0.01259180810302496, + 0.013225309550762177, + 0.03185294568538666, + 0.008961793966591358, + 0.009924017824232578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.020584410056471825 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and high-angle, hovering like a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two crabs visible in the center of the frame engaging in a duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one crab clad in a fiery scarlet carapace with massive claws tipped in ivory white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other crab featuring a sapphire blue shell speckled with yellow that gleams metallically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the crabs lunge forward and clash their claws in mid-air, throwing sand and water droplets into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579738542437553 + ], + "opposite_logprob_tokens": [ + -3.8955798149108887 + ], + "logprob_reward": -0.020579738542437553, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579738542437553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061586715281009674, + 0.04761543869972229, + 0.027847908437252045, + 0.08092565089464188, + 0.09775110334157944, + 0.15919266641139984, + 0.2556367516517639, + 0.16455568373203278, + 0.044925376772880554, + 0.05996274948120117 + ], + "credit_map": [ + 0.008161638863384724, + 0.0065742670558393, + 0.007018634118139744, + 0.01451416127383709, + 0.006766201928257942, + 0.007536471821367741, + 0.005451099947094917, + 0.005564235150814056, + 0.008847786113619804, + 0.005226218607276678, + 0.003984034061431885, + 0.005167306400835514, + 0.007636871188879013, + 0.006099318619817495, + 0.006701985374093056, + 0.003951915539801121, + 0.004409811459481716, + 0.002708657179027796, + 0.0036480273120105267, + 0.004289067815989256, + 0.002630353206768632, + 0.005010009743273258, + 0.0031277542002499104, + 0.0020242268219590187, + 0.019493939355015755, + 0.0072363754734396935, + 0.016996081918478012, + 0.009309166111052036, + 0.007645679172128439, + 0.006761830300092697, + 0.0071324072778224945, + 0.006350160576403141, + 0.01401427760720253, + 0.010230855084955692, + 0.01868722029030323, + 0.01542823389172554, + 0.008713025599718094, + 0.008218459784984589, + 0.010956067591905594, + 0.01150295976549387, + 0.025383206084370613, + 0.01984371989965439, + 0.03139110282063484, + 0.01957087777554989, + 0.021695608273148537, + 0.01950446516275406, + 0.009344642981886864, + 0.01245904341340065, + 0.09451720863580704, + 0.017744898796081543, + 0.07357452064752579, + 0.018812937662005424, + 0.011557553894817829, + 0.01622723415493965, + 0.010925661772489548, + 0.012276727706193924, + 0.03814540430903435, + 0.031091419979929924, + 0.027252037078142166, + 0.014925950206816196, + 0.013357443735003471, + 0.012386471033096313, + 0.015457979403436184, + 0.011938969604671001, + 0.005739293061196804, + 0.008141785860061646, + 0.007047856692224741, + 0.004640062339603901, + 0.005056444089859724, + 0.006331316661089659, + 0.0034024815540760756, + 0.004566130228340626, + 0.009863020852208138, + 0.007996080443263054, + 0.010568464174866676, + 0.005652691703289747, + 0.004571306519210339, + 0.012405282817780972, + 0.004163018893450499, + 0.004742880817502737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.020579738542437553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_5/index_tensor([1767, 1767])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_5/index_tensor([1767, 1767])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ea09e5f35c97aaa6d3f544f58d484ebdf4983fab --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_5/index_tensor([1767, 1767])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In one unbroken shot, the camera glides forward, slowly and ghostly, from a low angle, through the center aisle of the vast, smoke-filled Legislative Council chamber, a scene shrouded in a stark, dystopian sci-fi aesthetic. The imposing, brutalist concrete building, with its towering columns and vaulted ceilings, is almost entirely shrouded in deep shadows and a thick, synthetic haze that drifts lazily through the cavernous space. On either side of the camera's path, rows of spacious, empty, high-backed seats, constructed of cool, dark metal alloy and draped with worn black synthetic fabric, are arranged in quiet, orderly tiers. The only illumination comes from several monolithic screens embedded in the room's walls; they flicker unsteadily, silently displaying stark, graphic propaganda—bold, abstract symbols and faceless figures rendered in a harsh palette of scarlet and pristine white. This intermittent, cold digital light cast long, dancing shadows that crawled across the empty seats and polished stone floors, creating a scene of oppressive, silencing power and emphasizing the profound, tomblike silence of the abandoned halls of authority.", + "seed": 16653, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_5/index_1767_seed_16653_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01357613792642951, + "vlm_logprob": -0.01357613792642951, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.01357613792642951, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0365731343626976 + ], + "opposite_logprob_tokens": [ + -4.1615729331970215 + ], + "logprob_reward": -0.0365731343626976, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0365731343626976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2705518305301666, + 0.06580068171024323, + 0.08256065100431442, + 0.12049883604049683, + 0.13328032195568085, + 0.044559545814991, + 0.06790675967931747, + 0.061904389411211014, + 0.03412488102912903, + 0.11881207674741745 + ], + "credit_map": [ + 0.021374080330133438, + 0.05881787836551666, + 0.020600106567144394, + 0.0362880565226078, + 0.034185852855443954, + 0.030781976878643036, + 0.044388193637132645, + 0.02411569654941559, + 0.007509333547204733, + 0.008819624781608582, + 0.009192587807774544, + 0.011856750585138798, + 0.006819657515734434, + 0.007121858652681112, + 0.0076575265266001225, + 0.00682334927842021, + 0.007554478012025356, + 0.014419754035770893, + 0.00933653861284256, + 0.01286276150494814, + 0.007753034587949514, + 0.012336242944002151, + 0.009589256718754768, + 0.008708592504262924, + 0.011125716380774975, + 0.016923148185014725, + 0.0278000645339489, + 0.018110712990164757, + 0.009912398643791676, + 0.013952103443443775, + 0.007458653301000595, + 0.015216046944260597, + 0.005823093932121992, + 0.014205672778189182, + 0.034151431173086166, + 0.007930784486234188, + 0.051988113671541214, + 0.008716396056115627, + 0.006332812365144491, + 0.00413203239440918, + 0.004127637483179569, + 0.007318073883652687, + 0.007065640762448311, + 0.00857478566467762, + 0.0029657906852662563, + 0.004826613236218691, + 0.004601955879479647, + 0.005079052411019802, + 0.005805619060993195, + 0.009746621362864971, + 0.011087873950600624, + 0.008460803888738155, + 0.006437066942453384, + 0.007659615483134985, + 0.00523716863244772, + 0.013472001999616623, + 0.004773965571075678, + 0.006620092783123255, + 0.009417506866157055, + 0.008407261222600937, + 0.00533995172008872, + 0.0066028768196702, + 0.0059785908088088036, + 0.014764155261218548, + 0.0022300693672150373, + 0.00360851944424212, + 0.00505096884444356, + 0.005600904580205679, + 0.00321607175283134, + 0.004798364359885454, + 0.003821174381300807, + 0.005798812955617905, + 0.007031023968011141, + 0.008313866332173347, + 0.01970990188419819, + 0.038707707077264786, + 0.005707993172109127, + 0.010018603876233101, + 0.013994832523167133, + 0.015328160487115383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0365731343626976 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005267906002700329 + ], + "opposite_logprob_tokens": [ + -5.255268096923828 + ], + "logprob_reward": -0.005267906002700329, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005267906002700329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2178122103214264, + 0.08689785748720169, + 0.09925402700901031, + 0.05439203977584839, + 0.07619576156139374, + 0.05577774718403816, + 0.157257542014122, + 0.1336512565612793, + 0.05878753215074539, + 0.05997408553957939 + ], + "credit_map": [ + 0.018225830048322678, + 0.022968092933297157, + 0.019641336053609848, + 0.023965569213032722, + 0.02802312932908535, + 0.061114851385354996, + 0.01881358027458191, + 0.02505979686975479, + 0.00813749898225069, + 0.01008620671927929, + 0.010840218514204025, + 0.017192663624882698, + 0.00868781004101038, + 0.010786223225295544, + 0.01118430681526661, + 0.00998292863368988, + 0.0068410723470151424, + 0.013979380019009113, + 0.011558574624359608, + 0.019623395055532455, + 0.010572882369160652, + 0.016632163897156715, + 0.010966142639517784, + 0.009080411866307259, + 0.00443921098485589, + 0.007868802174925804, + 0.00873799528926611, + 0.009361442178487778, + 0.0058357249945402145, + 0.00791937392205, + 0.004193329717963934, + 0.006036154460161924, + 0.0063798013143241405, + 0.01089167408645153, + 0.010982564650475979, + 0.01600370556116104, + 0.010081181302666664, + 0.007425091695040464, + 0.008046443574130535, + 0.006385292857885361, + 0.004292997065931559, + 0.009866244159638882, + 0.008556295186281204, + 0.011104854755103588, + 0.006202102638781071, + 0.002567003248259425, + 0.006710173562169075, + 0.006478074472397566, + 0.016042007133364677, + 0.019351333379745483, + 0.024097124114632607, + 0.03010684624314308, + 0.011542053893208504, + 0.019099591299891472, + 0.01569480635225773, + 0.02132377400994301, + 0.009008650667965412, + 0.012780968099832535, + 0.018699219450354576, + 0.019222496077418327, + 0.008586202748119831, + 0.01880277507007122, + 0.015857841819524765, + 0.030693097040057182, + 0.004855234641581774, + 0.007662205025553703, + 0.008831681683659554, + 0.010563046671450138, + 0.0044669657945632935, + 0.006087241228669882, + 0.00635659322142601, + 0.009964559227228165, + 0.003999033942818642, + 0.008344058878719807, + 0.010422592051327229, + 0.011485233902931213, + 0.003834377508610487, + 0.006117217242717743, + 0.006510097533464432, + 0.00926146935671568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.005267906002700329 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638394996523857 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638394996523857, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638394996523857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38321617245674133, + 0.07326829433441162, + 0.07392013818025589, + 0.04573948308825493, + 0.055346228182315826, + 0.045125458389520645, + 0.08797094225883484, + 0.11908050626516342, + 0.03617389127612114, + 0.08015885949134827 + ], + "credit_map": [ + 0.024005308747291565, + 0.03573114424943924, + 0.023354176431894302, + 0.029415413737297058, + 0.05695761367678642, + 0.1426340639591217, + 0.0305059552192688, + 0.04061248153448105, + 0.007485515903681517, + 0.009071662090718746, + 0.011575204320251942, + 0.014178694225847721, + 0.0061116693541407585, + 0.009472515434026718, + 0.008409928530454636, + 0.006963097490370274, + 0.0056580305099487305, + 0.009869962930679321, + 0.009592438116669655, + 0.018631957471370697, + 0.00631828373298049, + 0.007916239090263844, + 0.008175836876034737, + 0.007757391780614853, + 0.004582038149237633, + 0.004877830855548382, + 0.009192053228616714, + 0.005451136268675327, + 0.005028665065765381, + 0.007813672535121441, + 0.003340179566293955, + 0.00545390835031867, + 0.0038672066293656826, + 0.006912179756909609, + 0.011096645146608353, + 0.005710489582270384, + 0.017402971163392067, + 0.003366197692230344, + 0.0036911282222718, + 0.0032994116190820932, + 0.003347928635776043, + 0.009396610781550407, + 0.0067743961699306965, + 0.009091772139072418, + 0.0037655981723219156, + 0.002732621505856514, + 0.00482208002358675, + 0.0051944502629339695, + 0.011503812856972218, + 0.011132081970572472, + 0.014281206764280796, + 0.013243394903838634, + 0.007993512786924839, + 0.013351010158658028, + 0.006442585494369268, + 0.01002333965152502, + 0.009362326934933662, + 0.011715046130120754, + 0.017114786431193352, + 0.020297693088650703, + 0.01023796945810318, + 0.011299622245132923, + 0.013013198971748352, + 0.02603987418115139, + 0.002607970731332898, + 0.0034213941544294357, + 0.0053294566459953785, + 0.004339578095823526, + 0.00290418672375381, + 0.005863720551133156, + 0.0036228529643267393, + 0.00808473490178585, + 0.0073280660435557365, + 0.010603513568639755, + 0.014319113455712795, + 0.01978987827897072, + 0.005107674282044172, + 0.007744255475699902, + 0.00706173712387681, + 0.008204629644751549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.008638394996523857 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778101928532124 + ], + "opposite_logprob_tokens": [ + -4.634778022766113 + ], + "logprob_reward": -0.009778101928532124, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778101928532124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1503455638885498, + 0.10409189760684967, + 0.21892738342285156, + 0.11423097550868988, + 0.04980093240737915, + 0.035982646048069, + 0.08773932605981827, + 0.10560350865125656, + 0.04707789048552513, + 0.08619987964630127 + ], + "credit_map": [ + 0.010879096575081348, + 0.020899862051010132, + 0.01312075462192297, + 0.023793183267116547, + 0.019952615723013878, + 0.027513442561030388, + 0.020594574511051178, + 0.013592052273452282, + 0.01231896597892046, + 0.012231012806296349, + 0.018328428268432617, + 0.014042017981410027, + 0.011768168769776821, + 0.014982528984546661, + 0.010078220628201962, + 0.010342561639845371, + 0.012435953132808208, + 0.03147405758500099, + 0.02546355500817299, + 0.037076517939567566, + 0.02244899608194828, + 0.04582995921373367, + 0.023654567077755928, + 0.020543767139315605, + 0.010311554186046124, + 0.010917792096734047, + 0.0296927560120821, + 0.014280921779572964, + 0.01037059910595417, + 0.01346583291888237, + 0.006879907567054033, + 0.01831160858273506, + 0.004284304566681385, + 0.0064802272245287895, + 0.008090000599622726, + 0.00603845389559865, + 0.007237331476062536, + 0.008970280177891254, + 0.0030592724215239286, + 0.00564106460660696, + 0.003748849034309387, + 0.0050291502848267555, + 0.005584608297795057, + 0.005365460179746151, + 0.00403588404878974, + 0.005111184436827898, + 0.0030212178826332092, + 0.004086290951818228, + 0.007386640179902315, + 0.009174496866762638, + 0.013974201865494251, + 0.011813351884484291, + 0.0072355191223323345, + 0.01734289713203907, + 0.009336142800748348, + 0.011476071551442146, + 0.005968589335680008, + 0.007759582716971636, + 0.016486959531903267, + 0.01873661018908024, + 0.006293003913015127, + 0.011931286193430424, + 0.012021228671073914, + 0.026406247168779373, + 0.002475969260558486, + 0.004354710690677166, + 0.00692096259444952, + 0.006713248789310455, + 0.005040633026510477, + 0.009861952625215054, + 0.0034031211398541927, + 0.008307290263473988, + 0.005603279452770948, + 0.007369521539658308, + 0.014345568604767323, + 0.028147924691438675, + 0.003615407273173332, + 0.0064163520000875, + 0.009585860185325146, + 0.011115974746644497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.009778101928532124 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762315234169364 + ], + "opposite_logprob_tokens": [ + -4.882623195648193 + ], + "logprob_reward": -0.00762315234169364, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762315234169364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32757821679115295, + 0.06142600253224373, + 0.0605681948363781, + 0.05966467782855034, + 0.1549927294254303, + 0.028835278004407883, + 0.04776047542691231, + 0.17441661655902863, + 0.02202773280441761, + 0.06273001432418823 + ], + "credit_map": [ + 0.03015938028693199, + 0.044167060405015945, + 0.02950442023575306, + 0.03655888885259628, + 0.05042874440550804, + 0.0696244090795517, + 0.04036048427224159, + 0.026774831116199493, + 0.005219982471317053, + 0.00716971093788743, + 0.01037756260484457, + 0.014560001902282238, + 0.00640990911051631, + 0.0059675914235413074, + 0.0062483916990458965, + 0.005472849123179913, + 0.006070650648325682, + 0.012027458287775517, + 0.008172689005732536, + 0.006059540435671806, + 0.00549102108925581, + 0.01403441745787859, + 0.00390643859282136, + 0.004805981181561947, + 0.005743077956140041, + 0.008526413701474667, + 0.012173593044281006, + 0.01372616644948721, + 0.00453418493270874, + 0.004960210062563419, + 0.004130842629820108, + 0.0058701918460428715, + 0.009853734634816647, + 0.02264973148703575, + 0.031589072197675705, + 0.01765817031264305, + 0.050588786602020264, + 0.007859239354729652, + 0.00823451578617096, + 0.006559466477483511, + 0.0019903844222426414, + 0.005108338315039873, + 0.004299148451536894, + 0.007318839896470308, + 0.0023347472306340933, + 0.0016566216945648193, + 0.003260604804381728, + 0.002866593888029456, + 0.0031009679660201073, + 0.007394692394882441, + 0.00753778638318181, + 0.0076783704571425915, + 0.005684040952473879, + 0.0029242997989058495, + 0.006497199647128582, + 0.00694311922416091, + 0.011683432385325432, + 0.012847074307501316, + 0.02676880545914173, + 0.032722607254981995, + 0.01321793906390667, + 0.021746870130300522, + 0.017361383885145187, + 0.0380684994161129, + 0.0017886633286252618, + 0.0029074957128614187, + 0.003546828404068947, + 0.003908244892954826, + 0.0017614978132769465, + 0.0026912663597613573, + 0.003096676431596279, + 0.0023270610254257917, + 0.0032173520885407925, + 0.004819698631763458, + 0.010098748840391636, + 0.01865619607269764, + 0.003789555048570037, + 0.006108390167355537, + 0.008748356252908707, + 0.007291723508387804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.00762315234169364 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In one unbroken shot, the camera glides forward, slowly and ghostly, from a low angle, through the center aisle of the vast, smoke-filled Legislative Council chamber, a scene shrouded in a stark, dystopian sci-fi aesthetic. The imposing, brutalist concrete building, with its towering columns and vaulted ceilings, is almost entirely shrouded in deep shadows and a thick, synthetic haze that drifts lazily through the cavernous space. On either side of the camera's path, rows of spacious, empty, high-backed seats, constructed of cool, dark metal alloy and draped with worn black synthetic fabric, are arranged in quiet, orderly tiers. The only illumination comes from several monolithic screens embedded in the room's walls; they flicker unsteadily, silently displaying stark, graphic propaganda—bold, abstract symbols and faceless figures rendered in a harsh palette of scarlet and pristine white. This intermittent, cold digital light cast long, dancing shadows that crawled across the empty seats and polished stone floors, creating a scene of oppressive, silencing power and emphasizing the profound, tomblike silence of the abandoned halls of authority.", + "seed": 16654, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_5/index_1767_seed_16654_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.007469447515904904, + "vlm_logprob": -0.007469447515904904, + "vlm_yes_no_margin": 5.275, + "vlm_token_credit": -0.007469447515904904, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006836239714175463 + ], + "opposite_logprob_tokens": [ + -6.131836414337158 + ], + "logprob_reward": -0.006836239714175463, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.006836239714175463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09832984209060669, + 0.0965263620018959, + 0.14450186491012573, + 0.1985832303762436, + 0.031212691217660904, + 0.10801294445991516, + 0.05702837556600571, + 0.0883808359503746, + 0.028322840109467506, + 0.14910104870796204 + ], + "credit_map": [ + 0.011788196861743927, + 0.014829784631729126, + 0.012369963340461254, + 0.009865577332675457, + 0.016887495294213295, + 0.011270307935774326, + 0.010797124356031418, + 0.01052139513194561, + 0.010287612676620483, + 0.015841469168663025, + 0.014371599070727825, + 0.008012731559574604, + 0.011997231282293797, + 0.012706214562058449, + 0.012273477390408516, + 0.011036025360226631, + 0.007744865026324987, + 0.008721890859305859, + 0.04587375000119209, + 0.006464090198278427, + 0.007211789488792419, + 0.011358474381268024, + 0.005855399649590254, + 0.051271598786115646, + 0.009024628438055515, + 0.004212604835629463, + 0.0759773701429367, + 0.00577994529157877, + 0.0066519733518362045, + 0.008963175117969513, + 0.0040345992892980576, + 0.08393893390893936, + 0.0028189874719828367, + 0.004371853545308113, + 0.004082005936652422, + 0.0034143913071602583, + 0.00396046182140708, + 0.004773139487951994, + 0.0034690124448388815, + 0.004322840366512537, + 0.013257870450615883, + 0.01583739183843136, + 0.011815243400633335, + 0.005582342855632305, + 0.029223380610346794, + 0.018103105947375298, + 0.007366866804659367, + 0.006826740223914385, + 0.009114392101764679, + 0.009183956310153008, + 0.006933221127837896, + 0.0036999026779085398, + 0.0071100154891610146, + 0.0067703877575695515, + 0.0063127377070486546, + 0.007903759367763996, + 0.010519010946154594, + 0.007801729720085859, + 0.006183602847158909, + 0.006796274799853563, + 0.013329153880476952, + 0.026253238320350647, + 0.005724742077291012, + 0.011773082427680492, + 0.0055633168667554855, + 0.0030524316243827343, + 0.002797589171677828, + 0.0038282331079244614, + 0.0031941926572471857, + 0.003806239692494273, + 0.0025667091831564903, + 0.0035141243133693933, + 0.014181154780089855, + 0.012688255868852139, + 0.023252831771969795, + 0.04162348806858063, + 0.010071349330246449, + 0.008836686611175537, + 0.013894484378397465, + 0.024552790448069572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 2 + ], + "selected_reward": -0.006836239714175463 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744240876287222 + ], + "opposite_logprob_tokens": [ + -5.006744384765625 + ], + "logprob_reward": -0.006744240876287222, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744240876287222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09289693087339401, + 0.03867555409669876, + 0.22647148370742798, + 0.209618479013443, + 0.050296682864427567, + 0.06079382449388504, + 0.04986460134387016, + 0.03335132077336311, + 0.04862990975379944, + 0.18940117955207825 + ], + "credit_map": [ + 0.012309307232499123, + 0.017591426149010658, + 0.009896621108055115, + 0.009669262915849686, + 0.01235273014754057, + 0.011327346786856651, + 0.009308263659477234, + 0.01044197753071785, + 0.0033062149304896593, + 0.003860753495246172, + 0.007924593985080719, + 0.004085334949195385, + 0.0022078403271734715, + 0.006875593215227127, + 0.005221267696470022, + 0.005193960852921009, + 0.011198037303984165, + 0.008571877144277096, + 0.08325664699077606, + 0.008923357352614403, + 0.0072609554044902325, + 0.020520387217402458, + 0.00837949849665165, + 0.07836073637008667, + 0.008626770228147507, + 0.0036639391910284758, + 0.08169903606176376, + 0.006575615610927343, + 0.006421064026653767, + 0.00941502582281828, + 0.003908670507371426, + 0.08930838108062744, + 0.004852595739066601, + 0.007105302065610886, + 0.007352068088948727, + 0.005044238176196814, + 0.00489447545260191, + 0.007614301983267069, + 0.005346514750272036, + 0.008087190799415112, + 0.006548822857439518, + 0.012822656892240047, + 0.006030849646776915, + 0.0029408626724034548, + 0.015353432856500149, + 0.008502121083438396, + 0.0038626401219516993, + 0.004732441622763872, + 0.012770089320838451, + 0.009312096051871777, + 0.004676369018852711, + 0.002735698828473687, + 0.003733682679012418, + 0.007815804332494736, + 0.004073211923241615, + 0.004747654311358929, + 0.003124037990346551, + 0.002794307190924883, + 0.0033268912229686975, + 0.004985286854207516, + 0.004044497385621071, + 0.007043750025331974, + 0.0038714995607733727, + 0.004161052405834198, + 0.008557567372918129, + 0.006285875104367733, + 0.0047621470876038074, + 0.005134863778948784, + 0.004441982600837946, + 0.00409594876691699, + 0.008309600874781609, + 0.00704192603006959, + 0.020794562995433807, + 0.027786964550614357, + 0.023325840011239052, + 0.04650872200727463, + 0.017376355826854706, + 0.015160055831074715, + 0.014470863156020641, + 0.023977817967534065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.006744240876287222 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003613373264670372 + ], + "opposite_logprob_tokens": [ + -5.628613471984863 + ], + "logprob_reward": -0.003613373264670372, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613373264670372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09324432164430618, + 0.06641150265932083, + 0.07612177729606628, + 0.17207229137420654, + 0.038518887013196945, + 0.03530212491750717, + 0.04294555261731148, + 0.1960231363773346, + 0.06292319297790527, + 0.2164372205734253 + ], + "credit_map": [ + 0.012254868634045124, + 0.009199151769280434, + 0.012874442152678967, + 0.011403866112232208, + 0.012498544529080391, + 0.011114968918263912, + 0.010294412262737751, + 0.013604073785245419, + 0.006389341317117214, + 0.011042830534279346, + 0.0078115565702319145, + 0.007133173756301403, + 0.006968584842979908, + 0.008210153318941593, + 0.01203578058630228, + 0.006820084992796183, + 0.007866735570132732, + 0.007484807632863522, + 0.010897132568061352, + 0.005092021077871323, + 0.008229604922235012, + 0.014054498635232449, + 0.0067482320591807365, + 0.01574874296784401, + 0.00693746842443943, + 0.004439293406903744, + 0.06516338884830475, + 0.006808726117014885, + 0.005249368958175182, + 0.008213851600885391, + 0.004318739287555218, + 0.07094147056341171, + 0.0036487907636910677, + 0.004854134283959866, + 0.004857221618294716, + 0.003753549885004759, + 0.004757874179631472, + 0.0077659557573497295, + 0.0030775433406233788, + 0.005803817417472601, + 0.0044103660620749, + 0.0064887432381510735, + 0.0038765775971114635, + 0.002874907571822405, + 0.006340608466416597, + 0.004465821664780378, + 0.002907962305471301, + 0.003937138710170984, + 0.009640742093324661, + 0.006540381815284491, + 0.004241767339408398, + 0.0027374497149139643, + 0.005284316372126341, + 0.005348797421902418, + 0.00427468866109848, + 0.004877407103776932, + 0.016743918880820274, + 0.020601801574230194, + 0.02206283062696457, + 0.030587313696742058, + 0.019470172002911568, + 0.01952192932367325, + 0.03761659190058708, + 0.029418569058179855, + 0.0101963235065341, + 0.006348947994410992, + 0.006561079062521458, + 0.00820367131382227, + 0.007742912974208593, + 0.006550251040607691, + 0.009273283183574677, + 0.008046724833548069, + 0.018024129793047905, + 0.029701877385377884, + 0.024153655394911766, + 0.06345624476671219, + 0.01708158478140831, + 0.01653025858104229, + 0.016830848529934883, + 0.030658619478344917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.003613373264670372 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091820679605007 + ], + "opposite_logprob_tokens": [ + -5.504091739654541 + ], + "logprob_reward": -0.004091820679605007, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091820679605007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11092100292444229, + 0.08305947482585907, + 0.042531028389930725, + 0.39495721459388733, + 0.02552892453968525, + 0.03927680850028992, + 0.07349192351102829, + 0.04362476244568825, + 0.03433986008167267, + 0.15226897597312927 + ], + "credit_map": [ + 0.012879355810582638, + 0.011453327722847462, + 0.015925845131278038, + 0.012771164067089558, + 0.019283592700958252, + 0.014334089122712612, + 0.010557360015809536, + 0.01371628325432539, + 0.006997581571340561, + 0.010647227987647057, + 0.013876601122319698, + 0.009343503043055534, + 0.008615527302026749, + 0.009430770762264729, + 0.012990131042897701, + 0.011158131062984467, + 0.00464178342372179, + 0.004652098286896944, + 0.00738875288516283, + 0.0034200118388980627, + 0.0031908517703413963, + 0.006339238025248051, + 0.004169754683971405, + 0.008728540502488613, + 0.016801079735159874, + 0.007246678229421377, + 0.15377016365528107, + 0.011525973677635193, + 0.011914624832570553, + 0.01715080253779888, + 0.007287857588380575, + 0.16926003992557526, + 0.0026840551290661097, + 0.0028191928286105394, + 0.003999726381152868, + 0.0022952069994062185, + 0.0027234184090048075, + 0.0047720386646687984, + 0.0022852530237287283, + 0.003950035199522972, + 0.00483767269179225, + 0.009505712427198887, + 0.003905436024069786, + 0.0034119333140552044, + 0.005811120383441448, + 0.0044035231694579124, + 0.0030516600236296654, + 0.004349753260612488, + 0.011203278787434101, + 0.013387423940002918, + 0.00663685193285346, + 0.004673893563449383, + 0.013525229878723621, + 0.01171825174242258, + 0.0054022166877985, + 0.006944781634956598, + 0.004049372393637896, + 0.005281672812998295, + 0.005336624104529619, + 0.006733029615134001, + 0.0053494032472372055, + 0.005832596216350794, + 0.005054188892245293, + 0.005987878423184156, + 0.005722999572753906, + 0.0029370267875492573, + 0.005001678131520748, + 0.0038405407685786486, + 0.003439334686845541, + 0.004433588124811649, + 0.004147940780967474, + 0.004816755652427673, + 0.014050924219191074, + 0.021316763013601303, + 0.015637710690498352, + 0.036505069583654404, + 0.015187067911028862, + 0.013823681510984898, + 0.011582847684621811, + 0.024164922535419464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.004091820679605007 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016061563044786453 + ], + "opposite_logprob_tokens": [ + -4.141061782836914 + ], + "logprob_reward": -0.016061563044786453, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016061563044786453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13324031233787537, + 0.05235162377357483, + 0.03355986252427101, + 0.3826037049293518, + 0.1782168596982956, + 0.061846379190683365, + 0.03357560932636261, + 0.033286113291978836, + 0.03561367839574814, + 0.05570589751005173 + ], + "credit_map": [ + 0.015693679451942444, + 0.01220979355275631, + 0.019430406391620636, + 0.020913736894726753, + 0.015658462420105934, + 0.015134981833398342, + 0.016523519530892372, + 0.0176757350564003, + 0.006088297814130783, + 0.007778244558721781, + 0.007512352894991636, + 0.006402826402336359, + 0.0046659475192427635, + 0.00640936940908432, + 0.007326098158955574, + 0.006168484687805176, + 0.002936713630333543, + 0.0029374398291110992, + 0.004249195102602243, + 0.0038053153548389673, + 0.003288410836830735, + 0.005956380628049374, + 0.005419809836894274, + 0.004966596607118845, + 0.01624760590493679, + 0.006119768135249615, + 0.14924956858158112, + 0.012607925571501255, + 0.010888423770666122, + 0.016464466229081154, + 0.007598558906465769, + 0.16342738270759583, + 0.015644218772649765, + 0.02687360905110836, + 0.02467348240315914, + 0.020548969507217407, + 0.02359582856297493, + 0.02977905422449112, + 0.016014862805604935, + 0.021086817607283592, + 0.007170903496444225, + 0.012957729399204254, + 0.00555887958034873, + 0.003067109966650605, + 0.01642461121082306, + 0.007386672310531139, + 0.0040399921126663685, + 0.005240487400442362, + 0.004272439982742071, + 0.006637518294155598, + 0.003228939836844802, + 0.0035589784383773804, + 0.0036170308012515306, + 0.005033586174249649, + 0.0048405565321445465, + 0.0023865618277341127, + 0.0028742272406816483, + 0.00285260914824903, + 0.003908821847289801, + 0.004998069256544113, + 0.0048652091063559055, + 0.00327459373511374, + 0.004977962467819452, + 0.005534619092941284, + 0.0041439589112997055, + 0.00317703140899539, + 0.0036250425036996603, + 0.005150948651134968, + 0.003995513543486595, + 0.0032793288119137287, + 0.007075001951307058, + 0.005166852381080389, + 0.004601786844432354, + 0.0052664815448224545, + 0.006637002341449261, + 0.018282031640410423, + 0.003997186664491892, + 0.0032434328459203243, + 0.004763446748256683, + 0.008914529345929623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.016061563044786453 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In one unbroken shot, the camera glides forward, slowly and ghostly, from a low angle, through the center aisle of the vast, smoke-filled Legislative Council chamber, a scene shrouded in a stark, dystopian sci-fi aesthetic. The imposing, brutalist concrete building, with its towering columns and vaulted ceilings, is almost entirely shrouded in deep shadows and a thick, synthetic haze that drifts lazily through the cavernous space. On either side of the camera's path, rows of spacious, empty, high-backed seats, constructed of cool, dark metal alloy and draped with worn black synthetic fabric, are arranged in quiet, orderly tiers. The only illumination comes from several monolithic screens embedded in the room's walls; they flicker unsteadily, silently displaying stark, graphic propaganda—bold, abstract symbols and faceless figures rendered in a harsh palette of scarlet and pristine white. This intermittent, cold digital light cast long, dancing shadows that crawled across the empty seats and polished stone floors, creating a scene of oppressive, silencing power and emphasizing the profound, tomblike silence of the abandoned halls of authority.", + "seed": 16655, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_5/index_1767_seed_16655_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.015464187413454056, + "vlm_logprob": -0.015464187413454056, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.015464187413454056, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017008861526846886 + ], + "opposite_logprob_tokens": [ + -5.1420087814331055 + ], + "logprob_reward": -0.017008861526846886, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.017008861526846886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2920040190219879, + 0.22855360805988312, + 0.08688386529684067, + 0.030063847079873085, + 0.0407877117395401, + 0.07028072327375412, + 0.053366705775260925, + 0.04180333763360977, + 0.058103617280721664, + 0.09815260022878647 + ], + "credit_map": [ + 0.019556716084480286, + 0.056474655866622925, + 0.012023013085126877, + 0.015894314274191856, + 0.09626176953315735, + 0.016112012788653374, + 0.015011167153716087, + 0.06067034602165222, + 0.017488598823547363, + 0.06788512319326401, + 0.05033339932560921, + 0.020198971033096313, + 0.023355800658464432, + 0.01876855455338955, + 0.012571762315928936, + 0.01795138418674469, + 0.011097786016762257, + 0.012386134825646877, + 0.010134827345609665, + 0.009053574874997139, + 0.011676415801048279, + 0.01611204445362091, + 0.008577811531722546, + 0.007845263928174973, + 0.003810417139902711, + 0.003737505991011858, + 0.002055402612313628, + 0.002483692718669772, + 0.004797748290002346, + 0.005651037208735943, + 0.0045661935582757, + 0.0029618467669934034, + 0.004080239683389664, + 0.005523210857063532, + 0.005007895641028881, + 0.004172401037067175, + 0.004935113713145256, + 0.0058286371640861034, + 0.0033066526520997286, + 0.007933555170893669, + 0.010813220404088497, + 0.020796779543161392, + 0.005887346342206001, + 0.006283915136009455, + 0.008299760520458221, + 0.009488550946116447, + 0.004863172769546509, + 0.003847969463095069, + 0.012103292159736156, + 0.003848389256745577, + 0.0033580262679606676, + 0.003519591875374317, + 0.009494469501078129, + 0.014856009744107723, + 0.0029879179783165455, + 0.0031990071292966604, + 0.004694750998169184, + 0.00825386494398117, + 0.004296419210731983, + 0.00446113059297204, + 0.005345986224710941, + 0.005873674992471933, + 0.004221417475491762, + 0.004656089469790459, + 0.005630028899759054, + 0.012225844897329807, + 0.006376736797392368, + 0.005057275760918856, + 0.007711958140134811, + 0.007695385720580816, + 0.00528050446882844, + 0.00812587607651949, + 0.005163018126040697, + 0.005145946517586708, + 0.004717255476862192, + 0.004218237474560738, + 0.03206329792737961, + 0.03226564824581146, + 0.008081848733127117, + 0.006497345399111509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.017008861526846886 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005958769004791975 + ], + "opposite_logprob_tokens": [ + -5.130958557128906 + ], + "logprob_reward": -0.005958769004791975, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958769004791975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5217459797859192, + 0.15902943909168243, + 0.03268039971590042, + 0.06203041970729828, + 0.023867998272180557, + 0.02491530030965805, + 0.02227562665939331, + 0.018256811425089836, + 0.0977572649717331, + 0.03744073957204819 + ], + "credit_map": [ + 0.03537766635417938, + 0.0902676060795784, + 0.01622188463807106, + 0.017904112115502357, + 0.19180694222450256, + 0.029452862218022346, + 0.019765127450227737, + 0.12094980478286743, + 0.016133004799485207, + 0.050994306802749634, + 0.022020742297172546, + 0.012652027420699596, + 0.027261894196271896, + 0.013133428059518337, + 0.00716581242159009, + 0.009668231941759586, + 0.0030786339193582535, + 0.007965960539877415, + 0.005839390214532614, + 0.003076625522226095, + 0.004156237468123436, + 0.0039672343991696835, + 0.003017715411260724, + 0.0015786058502271771, + 0.007453609257936478, + 0.016977587714791298, + 0.0032579628750681877, + 0.006483161821961403, + 0.007776937447488308, + 0.010852554813027382, + 0.00488319294527173, + 0.004345415625721216, + 0.003255365416407585, + 0.0038128651212900877, + 0.003146444447338581, + 0.0017866017296910286, + 0.00378067372366786, + 0.003929409198462963, + 0.0016597247449681163, + 0.0024969172663986683, + 0.003750702366232872, + 0.004542679991573095, + 0.0028242303524166346, + 0.0020044390112161636, + 0.003973674029111862, + 0.004136553965508938, + 0.0016062184004113078, + 0.0020768053364008665, + 0.0029268034268170595, + 0.005499953869730234, + 0.0013897090684622526, + 0.002307495567947626, + 0.003966901917010546, + 0.003285936312749982, + 0.0015574515564367175, + 0.0013413761043921113, + 0.0024686804972589016, + 0.002403598977252841, + 0.0020786100067198277, + 0.0019353682873770595, + 0.003349117236211896, + 0.0023138250689953566, + 0.001523517188616097, + 0.002184094861149788, + 0.010320776142179966, + 0.01541188359260559, + 0.012307275086641312, + 0.008468949235975742, + 0.01462320052087307, + 0.013410655781626701, + 0.010005156509578228, + 0.013209378346800804, + 0.0018377038650214672, + 0.002739234594628215, + 0.0015053347451612353, + 0.001611247775144875, + 0.012678069062530994, + 0.0129928607493639, + 0.0018251240253448486, + 0.002251169178634882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.005958769004791975 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627056445926428 + ], + "opposite_logprob_tokens": [ + -4.882627010345459 + ], + "logprob_reward": -0.007627056445926428, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627056445926428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5191487669944763, + 0.06340152025222778, + 0.06234712153673172, + 0.11013247072696686, + 0.042091984301805496, + 0.023674137890338898, + 0.046966396272182465, + 0.029195725917816162, + 0.05787089839577675, + 0.04517093673348427 + ], + "credit_map": [ + 0.03646627441048622, + 0.09000002592802048, + 0.012512268498539925, + 0.018863840028643608, + 0.19028307497501373, + 0.029137199744582176, + 0.01822618395090103, + 0.12365984916687012, + 0.00582485506311059, + 0.014015302993357182, + 0.007737905718386173, + 0.006380964070558548, + 0.013937334530055523, + 0.006881093140691519, + 0.004859523847699165, + 0.003764538327232003, + 0.006159426178783178, + 0.014842561446130276, + 0.005246937274932861, + 0.0057072509080171585, + 0.00870050024241209, + 0.011288678273558617, + 0.005644578021019697, + 0.00475718779489398, + 0.011756290681660175, + 0.021437272429466248, + 0.006151643581688404, + 0.011544803157448769, + 0.019633380696177483, + 0.01859574019908905, + 0.012245789170265198, + 0.00876754429191351, + 0.005836522206664085, + 0.004709777422249317, + 0.0032768866512924433, + 0.002828708617016673, + 0.009389204904437065, + 0.006861906033009291, + 0.0041762953624129295, + 0.005012680776417255, + 0.004168700892478228, + 0.005217917263507843, + 0.0021700167562812567, + 0.0025718335527926683, + 0.002167782047763467, + 0.0034503827337175608, + 0.002489055274054408, + 0.0014384459936991334, + 0.009144728071987629, + 0.006545801181346178, + 0.002384434686973691, + 0.005392716731876135, + 0.00745865423232317, + 0.010022168047726154, + 0.003583772573620081, + 0.0024341214448213577, + 0.003600300755351782, + 0.004274163860827684, + 0.003079903544858098, + 0.003475796664133668, + 0.004187202081084251, + 0.004133460111916065, + 0.0031347305048257113, + 0.003310164902359247, + 0.0056124161928892136, + 0.008435322903096676, + 0.007467103190720081, + 0.005635165609419346, + 0.008248620666563511, + 0.008304208517074585, + 0.006578970234841108, + 0.007589081302285194, + 0.002291348995640874, + 0.003382909344509244, + 0.0018358626402914524, + 0.0021720537915825844, + 0.015298792161047459, + 0.015384708531200886, + 0.00250949221663177, + 0.0022957639303058386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.007627056445926428 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863721314817667 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.00863721314817667, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863721314817667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.620151698589325, + 0.05380626767873764, + 0.015404684469103813, + 0.02447207272052765, + 0.027489934116601944, + 0.058395255357027054, + 0.030717970803380013, + 0.02535640262067318, + 0.11499086767435074, + 0.029214804992079735 + ], + "credit_map": [ + 0.04281536117196083, + 0.10804052650928497, + 0.013061698526144028, + 0.021141966804862022, + 0.23099467158317566, + 0.03207015246152878, + 0.020668597891926765, + 0.15135879814624786, + 0.005328964442014694, + 0.00980395544320345, + 0.008333703503012657, + 0.003867632942274213, + 0.009110188111662865, + 0.00917023979127407, + 0.003856776747852564, + 0.004334805998951197, + 0.0019358632853254676, + 0.0021417082753032446, + 0.00310345902107656, + 0.0011360117932781577, + 0.0019977842457592487, + 0.0018264014506712556, + 0.0016156058991327882, + 0.0016478500328958035, + 0.0032482678070664406, + 0.006216052453964949, + 0.001617780071683228, + 0.0022010516840964556, + 0.0031622210517525673, + 0.003965876065194607, + 0.0021660446655005217, + 0.0018947779899463058, + 0.0036205423530191183, + 0.002774070482701063, + 0.005331562366336584, + 0.001938832807354629, + 0.004143757279962301, + 0.0044290125370025635, + 0.001853237277828157, + 0.0033989199437201023, + 0.01023029163479805, + 0.012686114758253098, + 0.00645039277151227, + 0.0066039483062922955, + 0.006401710212230682, + 0.008078920654952526, + 0.004541569389402866, + 0.0034023073967546225, + 0.005188675131648779, + 0.004526454024016857, + 0.0017304206266999245, + 0.0033672386780381203, + 0.005779613275080919, + 0.005749293137341738, + 0.0023316338192671537, + 0.0020446411799639463, + 0.0026515740901231766, + 0.004222750198096037, + 0.00336343701928854, + 0.0029523514676839113, + 0.002831622725352645, + 0.004040071740746498, + 0.0019187830621376634, + 0.003375813364982605, + 0.012850911356508732, + 0.021643029525876045, + 0.011827213689684868, + 0.011812656186521053, + 0.013807293027639389, + 0.01877599209547043, + 0.010811660438776016, + 0.013462116941809654, + 0.0016446682857349515, + 0.0026639061979949474, + 0.0013723381562158465, + 0.00112484535202384, + 0.00905078835785389, + 0.009522408246994019, + 0.0019320392748340964, + 0.0019038099562749267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.00863721314817667 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808903694152832 + ], + "opposite_logprob_tokens": [ + -3.2880890369415283 + ], + "logprob_reward": -0.03808903694152832, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808903694152832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.441297709941864, + 0.08406047523021698, + 0.032160744071006775, + 0.07946696132421494, + 0.0690460130572319, + 0.034891899675130844, + 0.04271123558282852, + 0.048818621784448624, + 0.14601482450962067, + 0.021531453356146812 + ], + "credit_map": [ + 0.029697036370635033, + 0.07672901451587677, + 0.010217142291367054, + 0.01457519643008709, + 0.1627669334411621, + 0.025457745417952538, + 0.015471807681024075, + 0.10638284683227539, + 0.011825557798147202, + 0.01806294545531273, + 0.02078884467482567, + 0.00652699451893568, + 0.00808177050203085, + 0.00941227562725544, + 0.0045566861517727375, + 0.004805402364581823, + 0.0039192684926092625, + 0.006209236104041338, + 0.003716869978234172, + 0.0034743621945381165, + 0.00444961478933692, + 0.00513811269775033, + 0.0036235314328223467, + 0.001629748847335577, + 0.009235472418367863, + 0.018208326771855354, + 0.005216477904468775, + 0.0068174805492162704, + 0.011226160451769829, + 0.011938357725739479, + 0.008604842238128185, + 0.008219846524298191, + 0.009473654441535473, + 0.010050099343061447, + 0.007422072812914848, + 0.004872418008744717, + 0.013013576157391071, + 0.010465146042406559, + 0.005848592612892389, + 0.007900451309978962, + 0.005080960225313902, + 0.005640905816107988, + 0.004144193604588509, + 0.0036962677258998156, + 0.004807586316019297, + 0.005262518767267466, + 0.0035789075773209333, + 0.002680560341104865, + 0.006943707354366779, + 0.007585126906633377, + 0.0030865087173879147, + 0.003951525315642357, + 0.007389602717012167, + 0.007275361567735672, + 0.004008586518466473, + 0.002470819279551506, + 0.005713876336812973, + 0.007041288539767265, + 0.005631461273878813, + 0.006030608434230089, + 0.006085961125791073, + 0.006942942272871733, + 0.005090957973152399, + 0.0062815300188958645, + 0.013074015267193317, + 0.026610197499394417, + 0.01646333560347557, + 0.011030320078134537, + 0.025076745077967644, + 0.01954738050699234, + 0.014890956692397594, + 0.019321884959936142, + 0.0018102864269167185, + 0.0028200368396937847, + 0.0017393984599038959, + 0.00240432214923203, + 0.0038235413376241922, + 0.004204371012747288, + 0.002342048566788435, + 0.0023874486796557903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.03808903694152832 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In one unbroken shot, the camera glides forward, slowly and ghostly, from a low angle, through the center aisle of the vast, smoke-filled Legislative Council chamber, a scene shrouded in a stark, dystopian sci-fi aesthetic. The imposing, brutalist concrete building, with its towering columns and vaulted ceilings, is almost entirely shrouded in deep shadows and a thick, synthetic haze that drifts lazily through the cavernous space. On either side of the camera's path, rows of spacious, empty, high-backed seats, constructed of cool, dark metal alloy and draped with worn black synthetic fabric, are arranged in quiet, orderly tiers. The only illumination comes from several monolithic screens embedded in the room's walls; they flicker unsteadily, silently displaying stark, graphic propaganda—bold, abstract symbols and faceless figures rendered in a harsh palette of scarlet and pristine white. This intermittent, cold digital light cast long, dancing shadows that crawled across the empty seats and polished stone floors, creating a scene of oppressive, silencing power and emphasizing the profound, tomblike silence of the abandoned halls of authority.", + "seed": 16656, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_5/index_1767_seed_16656_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.013176652975380421, + "vlm_logprob": -0.013176652975380421, + "vlm_yes_no_margin": 4.974999952316284, + "vlm_token_credit": -0.013176652975380421, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.036307960748672485 + ], + "opposite_logprob_tokens": [ + -4.0363078117370605 + ], + "logprob_reward": -0.036307960748672485, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.036307960748672485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.236177459359169, + 0.14989037811756134, + 0.0690818801522255, + 0.034792959690093994, + 0.12990432977676392, + 0.053595758974552155, + 0.1012149229645729, + 0.0993727520108223, + 0.06643794476985931, + 0.05953160673379898 + ], + "credit_map": [ + 0.026010531932115555, + 0.02572675794363022, + 0.035774774849414825, + 0.03332895413041115, + 0.03252208232879639, + 0.027164075523614883, + 0.028643229976296425, + 0.027007078751921654, + 0.014174637384712696, + 0.009230646304786205, + 0.015537547878921032, + 0.01704747974872589, + 0.026382695883512497, + 0.02061009220778942, + 0.026336733251810074, + 0.02057056874036789, + 0.004954300355166197, + 0.004575956612825394, + 0.01025136187672615, + 0.006116347387433052, + 0.011138569563627243, + 0.011844150722026825, + 0.010278986766934395, + 0.009922216646373272, + 0.003736654296517372, + 0.003377303946763277, + 0.0034701800905168056, + 0.003407954005524516, + 0.004958342760801315, + 0.005855110008269548, + 0.005093581974506378, + 0.004893837030977011, + 0.008938752114772797, + 0.029433632269501686, + 0.011079620569944382, + 0.010320709086954594, + 0.01864445023238659, + 0.008311858400702477, + 0.010011016391217709, + 0.033164288848638535, + 0.0060909162275493145, + 0.005742833483964205, + 0.010251333005726337, + 0.007653046865016222, + 0.005954797845333815, + 0.0055350568145513535, + 0.007982981391251087, + 0.004384798463433981, + 0.009130025282502174, + 0.015034367330372334, + 0.008212700486183167, + 0.010791532695293427, + 0.01543288305401802, + 0.020359259098768234, + 0.012406272813677788, + 0.009847894310951233, + 0.01553471852093935, + 0.012808119878172874, + 0.01452964823693037, + 0.009103172458708286, + 0.01210864633321762, + 0.012209946289658546, + 0.012822537682950497, + 0.0102559644728899, + 0.01270026434212923, + 0.011292320676147938, + 0.008395236916840076, + 0.004686332773417234, + 0.00932281743735075, + 0.007394443266093731, + 0.00572527851909399, + 0.006921256426721811, + 0.00607961043715477, + 0.009251135401427746, + 0.006856156047433615, + 0.003886127844452858, + 0.012044795788824558, + 0.008092347532510757, + 0.007929823361337185, + 0.0053916145116090775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.036307960748672485 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005263874307274818 + ], + "opposite_logprob_tokens": [ + -5.255263805389404 + ], + "logprob_reward": -0.005263874307274818, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005263874307274818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1165151447057724, + 0.057043835520744324, + 0.05567578226327896, + 0.02942243032157421, + 0.09306102246046066, + 0.06788717210292816, + 0.09777919948101044, + 0.2770194411277771, + 0.09907325357198715, + 0.10652273893356323 + ], + "credit_map": [ + 0.013548997230827808, + 0.012113542295992374, + 0.011866417713463306, + 0.011065686121582985, + 0.024182744324207306, + 0.013592476025223732, + 0.015573380514979362, + 0.014571907930076122, + 0.005315558519214392, + 0.004791636485606432, + 0.006250698119401932, + 0.00725165382027626, + 0.005652188789099455, + 0.007267524488270283, + 0.012633563950657845, + 0.007881011813879013, + 0.004328636918216944, + 0.003949025645852089, + 0.006171345245093107, + 0.006215899717062712, + 0.009329549968242645, + 0.0098180016502738, + 0.008488260209560394, + 0.00737506290897727, + 0.002275993349030614, + 0.0029232779052108526, + 0.0030076676048338413, + 0.002908258233219385, + 0.005017745774239302, + 0.004959399811923504, + 0.00301057449541986, + 0.005319512914866209, + 0.012298859655857086, + 0.009520634077489376, + 0.013501289300620556, + 0.011136381886899471, + 0.014055801555514336, + 0.011022189632058144, + 0.012823885306715965, + 0.008701976388692856, + 0.006896974518895149, + 0.01131479348987341, + 0.008136944845318794, + 0.008958949707448483, + 0.007407448720186949, + 0.008343951776623726, + 0.009939959272742271, + 0.006888150703161955, + 0.009776546619832516, + 0.01664954610168934, + 0.008511099964380264, + 0.011347060091793537, + 0.012866846285760403, + 0.01166043896228075, + 0.009470301680266857, + 0.017497355118393898, + 0.025104813277721405, + 0.04315515607595444, + 0.04749506711959839, + 0.026063628494739532, + 0.04403558000922203, + 0.02387155219912529, + 0.047022465616464615, + 0.020271167159080505, + 0.010515781119465828, + 0.012876411899924278, + 0.01554422453045845, + 0.010644163936376572, + 0.010080345906317234, + 0.008373452350497246, + 0.01602838933467865, + 0.015010482631623745, + 0.01316335890442133, + 0.013081815093755722, + 0.013372964225709438, + 0.009791342541575432, + 0.01509128324687481, + 0.01488999929279089, + 0.015949612483382225, + 0.01118235383182764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.005263874307274818 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015160269103944302 + ], + "opposite_logprob_tokens": [ + -6.501515865325928 + ], + "logprob_reward": -0.0015160269103944302, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015160269103944302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34184348583221436, + 0.06842382997274399, + 0.04065226763486862, + 0.09998326748609543, + 0.10398902744054794, + 0.0574718713760376, + 0.11593819409608841, + 0.031856026500463486, + 0.046961963176727295, + 0.09288004785776138 + ], + "credit_map": [ + 0.03157580643892288, + 0.044291138648986816, + 0.05194685235619545, + 0.049719277769327164, + 0.04405985400080681, + 0.03365626186132431, + 0.051873352378606796, + 0.03472095727920532, + 0.006334519479423761, + 0.006685476750135422, + 0.0091285714879632, + 0.009199609979987144, + 0.007948837243020535, + 0.007466554641723633, + 0.012554354034364223, + 0.009105905890464783, + 0.0037990729324519634, + 0.0032196822576224804, + 0.0037431702949106693, + 0.004058456048369408, + 0.008211039006710052, + 0.004980383440852165, + 0.007137273903936148, + 0.005503188818693161, + 0.009742758236825466, + 0.015400107018649578, + 0.006234943401068449, + 0.011704212054610252, + 0.013398502953350544, + 0.013984491117298603, + 0.01426009926944971, + 0.015258151106536388, + 0.01232033409178257, + 0.008758080191910267, + 0.015664272010326385, + 0.014637302607297897, + 0.01260717585682869, + 0.012452545575797558, + 0.017041079699993134, + 0.010508237406611443, + 0.005064106080681086, + 0.010668579488992691, + 0.006175725255161524, + 0.005811805836856365, + 0.007580853067338467, + 0.009189944714307785, + 0.006969497539103031, + 0.0060113645158708096, + 0.006039755884557962, + 0.0162857323884964, + 0.008322805166244507, + 0.008102444000542164, + 0.019002441316843033, + 0.026170169934630394, + 0.01441738661378622, + 0.017597459256649017, + 0.00567956268787384, + 0.00394261023029685, + 0.0034024016931653023, + 0.0024651074782013893, + 0.0026276963762938976, + 0.004227445460855961, + 0.006262012291699648, + 0.003249191679060459, + 0.0065287756733596325, + 0.005284570157527924, + 0.005660227034240961, + 0.0025186219718307257, + 0.010137438774108887, + 0.006828728131949902, + 0.004465607460588217, + 0.005537991411983967, + 0.007986071519553661, + 0.010433465242385864, + 0.008463584817945957, + 0.007737382315099239, + 0.02182326465845108, + 0.018352238461375237, + 0.009717294946312904, + 0.008366753347218037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0015160269103944302 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735715549439192 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735715549439192, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735715549439192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13051363825798035, + 0.09558311104774475, + 0.19056889414787292, + 0.14593295753002167, + 0.07979287952184677, + 0.12484651803970337, + 0.04659821093082428, + 0.04704757034778595, + 0.03891497477889061, + 0.10020118951797485 + ], + "credit_map": [ + 0.01524269673973322, + 0.019283387809991837, + 0.0131888622418046, + 0.015886375680565834, + 0.021266911178827286, + 0.01650458388030529, + 0.013453558087348938, + 0.015687257051467896, + 0.00788097269833088, + 0.006909932475537062, + 0.010231167078018188, + 0.01086016558110714, + 0.014813468791544437, + 0.009499497711658478, + 0.021029839292168617, + 0.014358065091073513, + 0.01570214331150055, + 0.011320716701447964, + 0.017944538965821266, + 0.016523044556379318, + 0.03128928691148758, + 0.029192423447966576, + 0.03696553036570549, + 0.0316312238574028, + 0.010237349197268486, + 0.01535558607429266, + 0.008153700269758701, + 0.011735926382243633, + 0.02662479877471924, + 0.03132222965359688, + 0.019924674183130264, + 0.02257869392633438, + 0.006985153537243605, + 0.012243700213730335, + 0.008067887276411057, + 0.008408029563724995, + 0.011890850029885769, + 0.009767943993210793, + 0.010254506021738052, + 0.012174810282886028, + 0.013148298487067223, + 0.015034019947052002, + 0.00888819433748722, + 0.008538713678717613, + 0.0171707384288311, + 0.03636236861348152, + 0.013789401389658451, + 0.011914785951375961, + 0.0057978034019470215, + 0.005881345830857754, + 0.004566102288663387, + 0.0069308956153690815, + 0.004016859922558069, + 0.00852703396230936, + 0.00553022138774395, + 0.005347945261746645, + 0.007951405830681324, + 0.005507417488843203, + 0.007594666443765163, + 0.004841127898544073, + 0.004975076764822006, + 0.005815152544528246, + 0.006172641180455685, + 0.004190084524452686, + 0.004204979166388512, + 0.0037779260892421007, + 0.003968329168856144, + 0.0029293568804860115, + 0.007620163261890411, + 0.008094951510429382, + 0.004108447581529617, + 0.004210821818560362, + 0.014358550310134888, + 0.025602100417017937, + 0.009514887817203999, + 0.0065656304359436035, + 0.0142229488119483, + 0.012193458154797554, + 0.007427765987813473, + 0.010315842926502228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.006735715549439192 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving forward through the center aisle of a vast chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of empty, high-backed seats made of dark metal alloy and draped with black fabric visible on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment filled with a thick, synthetic haze and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do monolithic screens on the walls display flickering propaganda with scarlet and white abstract symbols and faceless figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a brutalist concrete building with towering columns and vaulted ceilings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016059687361121178 + ], + "opposite_logprob_tokens": [ + -4.141059875488281 + ], + "logprob_reward": -0.016059687361121178, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016059687361121178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10977708548307419, + 0.05368452146649361, + 0.06598114967346191, + 0.0671817883849144, + 0.045552752912044525, + 0.3153545558452606, + 0.09910357743501663, + 0.05027955770492554, + 0.12256152182817459, + 0.0705234631896019 + ], + "credit_map": [ + 0.01090134959667921, + 0.010030961595475674, + 0.01811710000038147, + 0.013959667645394802, + 0.016840601339936256, + 0.011420461349189281, + 0.014926902949810028, + 0.01358004566282034, + 0.005431387107819319, + 0.005128621123731136, + 0.007826698012650013, + 0.006215602625161409, + 0.00760157173499465, + 0.005721050780266523, + 0.00876596663147211, + 0.006993625313043594, + 0.005769740790128708, + 0.0043603768572211266, + 0.006754118017852306, + 0.006493913475424051, + 0.009358592331409454, + 0.00787407997995615, + 0.015231623314321041, + 0.010138708166778088, + 0.003774580778554082, + 0.008623150177299976, + 0.0063668605871498585, + 0.006547663826495409, + 0.009037303738296032, + 0.013038438744843006, + 0.008932418189942837, + 0.010861380025744438, + 0.00485580088570714, + 0.003676261054351926, + 0.0066095213405787945, + 0.007049955427646637, + 0.004664489533752203, + 0.00443307077512145, + 0.007781811524182558, + 0.006481844000518322, + 0.03659764304757118, + 0.06724942475557327, + 0.040535490959882736, + 0.024463549256324768, + 0.0468473918735981, + 0.032395023852586746, + 0.039850253611803055, + 0.02741580829024315, + 0.013470417819917202, + 0.0099712535738945, + 0.008942078799009323, + 0.016674965620040894, + 0.012380851432681084, + 0.01421347726136446, + 0.012559705413877964, + 0.01089082844555378, + 0.004990823566913605, + 0.007261293940246105, + 0.0075843543745577335, + 0.0045789568684995174, + 0.005500584840774536, + 0.006499327719211578, + 0.008478114381432533, + 0.005386108998209238, + 0.011872085742652416, + 0.011322425678372383, + 0.011900849640369415, + 0.006346082780510187, + 0.0316731259226799, + 0.026677045971155167, + 0.009963210672140121, + 0.012806701473891735, + 0.00716396002098918, + 0.008497289381921291, + 0.007281650323420763, + 0.0065477038733661175, + 0.015188020654022694, + 0.009837128221988678, + 0.0067994361743330956, + 0.00920828152447939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.016059687361121178 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_6/index_tensor([443, 443])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_6/index_tensor([443, 443])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3f532d6f2640b56634b93a1bc5f4fddd97f60c85 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_6/index_tensor([443, 443])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot begins in close-up, then slowly and steadily pulls back to a medium shot, initially focusing on a pair of calloused, yet remarkably steady hands. Carefully, these hands grip a fine chisel, gleaming with a cold metallic sheen. With a steady and deliberate rhythm, the tip of the knife glides and slices across a naturally formed, rough-surfaced dead branch. Fine wood chips gently swirl and drift as the knife moves, and the graceful curve of a bird's wing and the outline of its head are gradually freed from the wood. As the camera pulls away, we see the full extent of the craftsman: an Asian man with graying hair tied in a small bun, wearing a simple dark blue cotton-linen shirt with sleeves rolled up to the elbows. His face wears an expression of almost meditative concentration. He sits in a sun-drenched studio, where the warm afternoon sun slants through a large floor-to-ceiling window, creating distinct beams of light in the air and illuminating the dancing dust particles. In the background, wooden shelves are cluttered with a variety of carving tools, unfinished works, and various woods. The entire scene is bathed in a warm golden hue, highlighting the natural grain of the wood and the wrinkles on the artisans' skin, creating an artistic atmosphere of tranquility, concentration, and creative beauty.", + "seed": 23493, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_6/index_443_seed_23493_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03857993111014366, + "vlm_logprob": -0.03857993111014366, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.03857993111014366, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03877457231283188 + ], + "opposite_logprob_tokens": [ + -4.038774490356445 + ], + "logprob_reward": -0.03877457231283188, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.03877457231283188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18036527931690216, + 0.08610997349023819, + 0.18181264400482178, + 0.030036257579922676, + 0.03651094809174538, + 0.09683442860841751, + 0.0732312947511673, + 0.199292853474617, + 0.0774679183959961, + 0.038338448852300644 + ], + "credit_map": [ + 0.026657193899154663, + 0.017538322135806084, + 0.020776424556970596, + 0.010947071015834808, + 0.030143167823553085, + 0.020150568336248398, + 0.017364097759127617, + 0.03678841516375542, + 0.01233971118927002, + 0.013058989308774471, + 0.008720635436475277, + 0.009246356785297394, + 0.00942396279424429, + 0.009344900026917458, + 0.008866502903401852, + 0.015108904801309109, + 0.021485310047864914, + 0.014249262399971485, + 0.02700037881731987, + 0.018937973305583, + 0.03583313524723053, + 0.01812204159796238, + 0.024725526571273804, + 0.021459005773067474, + 0.0041168262250721455, + 0.002721041440963745, + 0.0035447923000901937, + 0.0035468880087137222, + 0.004344387445598841, + 0.0043103317730128765, + 0.003518472658470273, + 0.00393351586535573, + 0.0050044916570186615, + 0.006912560667842627, + 0.003884875914081931, + 0.004937333520501852, + 0.0040227994322776794, + 0.0030438145622611046, + 0.0042531173676252365, + 0.004451951943337917, + 0.020343899726867676, + 0.007243716157972813, + 0.006388576701283455, + 0.011486492119729519, + 0.006759709678590298, + 0.024857252836227417, + 0.005503607913851738, + 0.0142511622980237, + 0.01032076496630907, + 0.007513617165386677, + 0.009735548868775368, + 0.015076125971972942, + 0.004764390643686056, + 0.008110372349619865, + 0.006005394738167524, + 0.011705074459314346, + 0.05011659115552902, + 0.033037394285202026, + 0.011323854327201843, + 0.007529088295996189, + 0.02139236591756344, + 0.04085010662674904, + 0.011033468879759312, + 0.024009963497519493, + 0.016426118090748787, + 0.014612332917749882, + 0.007986243814229965, + 0.0073026190511882305, + 0.008464734070003033, + 0.0077309939078986645, + 0.0064902580343186855, + 0.008454608730971813, + 0.006606229580938816, + 0.005960228852927685, + 0.004061568062752485, + 0.0028586529660969973, + 0.006267826072871685, + 0.003518457990139723, + 0.004771154839545488, + 0.0042943269945681095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.03877457231283188 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04872142896056175 + ], + "opposite_logprob_tokens": [ + -3.0487213134765625 + ], + "logprob_reward": -0.04872142896056175, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04872142896056175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2693590223789215, + 0.07944726943969727, + 0.10358282178640366, + 0.024709371849894524, + 0.04953886941075325, + 0.19733808934688568, + 0.03568105772137642, + 0.1622430831193924, + 0.051364220678806305, + 0.026736170053482056 + ], + "credit_map": [ + 0.03051121160387993, + 0.07369228452444077, + 0.01927170343697071, + 0.017140906304121017, + 0.023261550813913345, + 0.027753571048378944, + 0.018741754814982414, + 0.05898605287075043, + 0.017672697082161903, + 0.006008862983435392, + 0.005269923713058233, + 0.004077954217791557, + 0.006987214554101229, + 0.006934222765266895, + 0.008312943391501904, + 0.024183453992009163, + 0.016013428568840027, + 0.01355363242328167, + 0.01843886263668537, + 0.009984400123357773, + 0.010012737475335598, + 0.011196594685316086, + 0.008931710384786129, + 0.015451454557478428, + 0.004128960892558098, + 0.0021047743503004313, + 0.0031745994929224253, + 0.002995749469846487, + 0.002733095083385706, + 0.003496152348816395, + 0.0027986408676952124, + 0.0032773991115391254, + 0.006336466409265995, + 0.005788283888250589, + 0.005694487132132053, + 0.00940545741468668, + 0.003421533852815628, + 0.006808266043663025, + 0.003997840452939272, + 0.008086537942290306, + 0.039030738174915314, + 0.015707390382885933, + 0.014718831516802311, + 0.026783933863043785, + 0.009052778594195843, + 0.053646545857191086, + 0.011861436069011688, + 0.026536429300904274, + 0.003946181386709213, + 0.0035420875065028667, + 0.004528485704213381, + 0.005042856093496084, + 0.0036647494416683912, + 0.004368413705378771, + 0.004937480203807354, + 0.005650802981108427, + 0.045709941536188126, + 0.030103158205747604, + 0.010558228939771652, + 0.007942048832774162, + 0.008374045602977276, + 0.03685422241687775, + 0.007151402998715639, + 0.015550032258033752, + 0.007763865869492292, + 0.007644833065569401, + 0.007514886558055878, + 0.007161925546824932, + 0.0051980880089104176, + 0.004838787019252777, + 0.004158019088208675, + 0.007083815988153219, + 0.003831746056675911, + 0.003055717796087265, + 0.004692274145781994, + 0.002205398865044117, + 0.0038924820255488157, + 0.002727865008637309, + 0.0031872079707682133, + 0.003143477952107787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.04872142896056175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0380970723927021 + ], + "opposite_logprob_tokens": [ + -3.2880971431732178 + ], + "logprob_reward": -0.0380970723927021, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0380970723927021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2619471549987793, + 0.040614958852529526, + 0.08367617428302765, + 0.06269444525241852, + 0.05221204459667206, + 0.32981452345848083, + 0.04733625799417496, + 0.06921467930078506, + 0.03093789517879486, + 0.02155180647969246 + ], + "credit_map": [ + 0.0763375535607338, + 0.057403676211833954, + 0.008831932209432125, + 0.007177446968853474, + 0.032052867114543915, + 0.017884260043501854, + 0.011006857268512249, + 0.051252592355012894, + 0.007550508715212345, + 0.003928560763597488, + 0.005332341883331537, + 0.004564551170915365, + 0.0037254774942994118, + 0.004437560681253672, + 0.004590875003486872, + 0.006485081743448973, + 0.009544293396174908, + 0.009858272969722748, + 0.008083532564342022, + 0.007922197692096233, + 0.018002083525061607, + 0.010684508830308914, + 0.009622488170862198, + 0.009958799928426743, + 0.009211341850459576, + 0.004423817154020071, + 0.011950382962822914, + 0.004918673541396856, + 0.006396466866135597, + 0.008180168457329273, + 0.008433090522885323, + 0.009180505760014057, + 0.008325743488967419, + 0.00558383297175169, + 0.0074454075656831264, + 0.009812969714403152, + 0.0037401735316962004, + 0.005126959644258022, + 0.005245462525635958, + 0.006931496784090996, + 0.08745938539505005, + 0.024782296270132065, + 0.008989723399281502, + 0.015104240737855434, + 0.013372554443776608, + 0.11665788292884827, + 0.009048202075064182, + 0.05440021678805351, + 0.005874487105756998, + 0.004733356647193432, + 0.006231355015188456, + 0.004992109723389149, + 0.004282334353774786, + 0.006931112613528967, + 0.0034499431494623423, + 0.010841558687388897, + 0.013857563026249409, + 0.010217536240816116, + 0.010273858904838562, + 0.009101239033043385, + 0.005673457868397236, + 0.008960030972957611, + 0.00482880137860775, + 0.006302193272858858, + 0.006144470535218716, + 0.003420524299144745, + 0.0039031219203025103, + 0.0034363556187599897, + 0.0023516984656453133, + 0.0034923481289297342, + 0.0024852075148373842, + 0.005704171489924192, + 0.002438710303977132, + 0.0020896706264466047, + 0.0033906977623701096, + 0.002468053251504898, + 0.003983383532613516, + 0.0024881623685359955, + 0.002366194501519203, + 0.0023269352968782187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.0380970723927021 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062056463211774826 + ], + "opposite_logprob_tokens": [ + -2.812056541442871 + ], + "logprob_reward": -0.062056463211774826, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062056463211774826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062105827033519745, + 0.038117170333862305, + 0.053230226039886475, + 0.020829597488045692, + 0.06477140635251999, + 0.12917110323905945, + 0.07665718346834183, + 0.480340838432312, + 0.03970729559659958, + 0.035069361329078674 + ], + "credit_map": [ + 0.020630067214369774, + 0.005356287118047476, + 0.0047762407921254635, + 0.003860896686092019, + 0.0038006065879017115, + 0.0033842462580651045, + 0.004899731837213039, + 0.015397747047245502, + 0.005812702234834433, + 0.0035327645018696785, + 0.004469998646527529, + 0.003218414494767785, + 0.003353168023750186, + 0.005513359792530537, + 0.004131364170461893, + 0.008085398934781551, + 0.007858291268348694, + 0.0040296041406691074, + 0.010082137770950794, + 0.00476992828771472, + 0.005162341985851526, + 0.004964814987033606, + 0.007593412883579731, + 0.008769689127802849, + 0.0029454699251800776, + 0.00209422898478806, + 0.0014344718074426055, + 0.0022670249454677105, + 0.003646140219643712, + 0.0029806396923959255, + 0.0022795903496444225, + 0.003182030515745282, + 0.009680578485131264, + 0.007046709768474102, + 0.005269843619316816, + 0.0077201067470014095, + 0.012061466462910175, + 0.006478006020188332, + 0.006881797686219215, + 0.009632901288568974, + 0.022265402600169182, + 0.009554625488817692, + 0.009857500903308392, + 0.021606899797916412, + 0.010354002937674522, + 0.026629270985722542, + 0.009493691846728325, + 0.019409704953432083, + 0.00592435896396637, + 0.008011545985937119, + 0.008859438821673393, + 0.00709304865449667, + 0.00639805244281888, + 0.013086157850921154, + 0.008362498134374619, + 0.018922079354524612, + 0.1200837641954422, + 0.07580514252185822, + 0.038692399859428406, + 0.02555752731859684, + 0.03214308246970177, + 0.12072258442640305, + 0.024522928521037102, + 0.04281339794397354, + 0.004503084812313318, + 0.0039240275509655476, + 0.005409589968621731, + 0.0037020205054432154, + 0.006715075112879276, + 0.005287724547088146, + 0.0037045064382255077, + 0.006461264565587044, + 0.0029531384352594614, + 0.0021162459161132574, + 0.003998349886387587, + 0.002282849047333002, + 0.007297617848962545, + 0.004956674762070179, + 0.00625614495947957, + 0.005208340473473072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.062056463211774826 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250118672847748 + ], + "opposite_logprob_tokens": [ + -5.255249977111816 + ], + "logprob_reward": -0.005250118672847748, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250118672847748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18030068278312683, + 0.059408038854599, + 0.04436887428164482, + 0.014160430058836937, + 0.027285240590572357, + 0.10365283489227295, + 0.04498976841568947, + 0.3646749258041382, + 0.09060029685497284, + 0.07055892050266266 + ], + "credit_map": [ + 0.04909894987940788, + 0.058073703199625015, + 0.008001161739230156, + 0.008939324878156185, + 0.01579040102660656, + 0.011699005030095577, + 0.010048899799585342, + 0.01864922046661377, + 0.01617930829524994, + 0.004267858806997538, + 0.0036201642360538244, + 0.00434832600876689, + 0.002319915918633342, + 0.004298201762139797, + 0.002720048651099205, + 0.02165420912206173, + 0.0078806821256876, + 0.0046510533429682255, + 0.009692630730569363, + 0.0038335155695676804, + 0.002597615821287036, + 0.003722555236890912, + 0.005460516549646854, + 0.006530298851430416, + 0.0022412054240703583, + 0.0014829981373623013, + 0.0020696285646408796, + 0.0012026905314996839, + 0.001975269755348563, + 0.0018659413326531649, + 0.0016300830757245421, + 0.0016926121897995472, + 0.003805013606324792, + 0.00418855482712388, + 0.003621199633926153, + 0.004689676221460104, + 0.002079418860375881, + 0.0024887300096452236, + 0.0027779063675552607, + 0.003634735709056258, + 0.024668293073773384, + 0.0078748082742095, + 0.0045150876976549625, + 0.009284170344471931, + 0.004583805799484253, + 0.03315167501568794, + 0.0035864112433046103, + 0.015988577157258987, + 0.007745269685983658, + 0.006229891441762447, + 0.006086053792387247, + 0.006005115807056427, + 0.0033469246700406075, + 0.0047988928854465485, + 0.004510839004069567, + 0.006266773212701082, + 0.10676240175962448, + 0.06724395602941513, + 0.022936614230275154, + 0.00871435645967722, + 0.016164522618055344, + 0.09858434647321701, + 0.011885281652212143, + 0.03238338604569435, + 0.011016503907740116, + 0.010566335171461105, + 0.02697920985519886, + 0.013894312083721161, + 0.0052082291804254055, + 0.008944930508732796, + 0.004802477080374956, + 0.009188300929963589, + 0.00875682570040226, + 0.011790207587182522, + 0.012609812431037426, + 0.010713721625506878, + 0.008435063995420933, + 0.007452949415892363, + 0.006143826525658369, + 0.004656504839658737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.005250118672847748 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins in close-up, then slowly and steadily pulls back to a medium shot, initially focusing on a pair of calloused, yet remarkably steady hands. Carefully, these hands grip a fine chisel, gleaming with a cold metallic sheen. With a steady and deliberate rhythm, the tip of the knife glides and slices across a naturally formed, rough-surfaced dead branch. Fine wood chips gently swirl and drift as the knife moves, and the graceful curve of a bird's wing and the outline of its head are gradually freed from the wood. As the camera pulls away, we see the full extent of the craftsman: an Asian man with graying hair tied in a small bun, wearing a simple dark blue cotton-linen shirt with sleeves rolled up to the elbows. His face wears an expression of almost meditative concentration. He sits in a sun-drenched studio, where the warm afternoon sun slants through a large floor-to-ceiling window, creating distinct beams of light in the air and illuminating the dancing dust particles. In the background, wooden shelves are cluttered with a variety of carving tools, unfinished works, and various woods. The entire scene is bathed in a warm golden hue, highlighting the natural grain of the wood and the wrinkles on the artisans' skin, creating an artistic atmosphere of tranquility, concentration, and creative beauty.", + "seed": 23494, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_6/index_443_seed_23494_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023441795026883483, + "vlm_logprob": -0.023441795026883483, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.023441795026883483, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030018260702490807 + ], + "opposite_logprob_tokens": [ + -5.030018329620361 + ], + "logprob_reward": -0.030018260702490807, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.030018260702490807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09705986082553864, + 0.2927074730396271, + 0.04207603260874748, + 0.05918996036052704, + 0.042456623166799545, + 0.119708351790905, + 0.11425165832042694, + 0.14453686773777008, + 0.039302434772253036, + 0.04871073737740517 + ], + "credit_map": [ + 0.012030944228172302, + 0.011193675920367241, + 0.01248414721339941, + 0.01693720929324627, + 0.006141013000160456, + 0.011089703068137169, + 0.015266042202711105, + 0.011917121708393097, + 0.011303834617137909, + 0.010792586952447891, + 0.008966543711721897, + 0.00920908059924841, + 0.015815356746315956, + 0.11473146080970764, + 0.10604164004325867, + 0.0158469770103693, + 0.0033661015331745148, + 0.0067463768646121025, + 0.00413070572540164, + 0.00418701209127903, + 0.005140404682606459, + 0.004266050644218922, + 0.006304421927779913, + 0.007934960536658764, + 0.005967586766928434, + 0.007621742784976959, + 0.0053996252827346325, + 0.006675658747553825, + 0.011690108105540276, + 0.008014504797756672, + 0.005691390950232744, + 0.008129341527819633, + 0.004541815724223852, + 0.003381013637408614, + 0.003892584703862667, + 0.004261201713234186, + 0.007012964691966772, + 0.006565236020833254, + 0.006136010400950909, + 0.0066657946445047855, + 0.01223010290414095, + 0.00913823302835226, + 0.016487443819642067, + 0.013015164993703365, + 0.018653497099876404, + 0.019797123968601227, + 0.011609423905611038, + 0.018777359277009964, + 0.015325166285037994, + 0.009192831814289093, + 0.011792262084782124, + 0.006054306402802467, + 0.018495962023735046, + 0.014359362423419952, + 0.011577815748751163, + 0.0274539515376091, + 0.011969414539635181, + 0.010177139192819595, + 0.02061298117041588, + 0.014834141358733177, + 0.01673048362135887, + 0.01536390371620655, + 0.029604505747556686, + 0.025244301185011864, + 0.0038572403136640787, + 0.004736350849270821, + 0.0045414334163069725, + 0.0039535327814519405, + 0.0057310438714921474, + 0.00331106036901474, + 0.007513674907386303, + 0.005658099427819252, + 0.004732947330921888, + 0.004163319244980812, + 0.0077374824322760105, + 0.007059826981276274, + 0.006528161000460386, + 0.00524807209149003, + 0.007369729224592447, + 0.005871200934052467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 5 + ], + "selected_reward": -0.030018260702490807 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038142282515764236 + ], + "opposite_logprob_tokens": [ + -3.288142204284668 + ], + "logprob_reward": -0.038142282515764236, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038142282515764236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1674216240644455, + 0.16473060846328735, + 0.07544369995594025, + 0.40157559514045715, + 0.03388664871454239, + 0.032072875648736954, + 0.056184232234954834, + 0.021180162206292152, + 0.01804940402507782, + 0.02945515140891075 + ], + "credit_map": [ + 0.017741957679390907, + 0.01551058329641819, + 0.021922102198004723, + 0.022307531908154488, + 0.013182557187974453, + 0.03333994746208191, + 0.015639804303646088, + 0.027777135372161865, + 0.007069337647408247, + 0.0048636142164468765, + 0.005733349360525608, + 0.005004120524972677, + 0.008978779427707195, + 0.06261458247900009, + 0.0601307637989521, + 0.010336067527532578, + 0.0049020773731172085, + 0.004611747805029154, + 0.004956327844411135, + 0.006105051841586828, + 0.0184162650257349, + 0.011980824172496796, + 0.012633650563657284, + 0.011837761849164963, + 0.03803546726703644, + 0.06551719456911087, + 0.05085480958223343, + 0.054496895521879196, + 0.0473945215344429, + 0.04714058339595795, + 0.05908486247062683, + 0.039051249623298645, + 0.0025170641019940376, + 0.0029268949292600155, + 0.0028824200853705406, + 0.005095226224511862, + 0.005958947818726301, + 0.0052344584837555885, + 0.005767226684838533, + 0.003504411317408085, + 0.0028430940583348274, + 0.0022117618937045336, + 0.002922440180554986, + 0.0020333724096417427, + 0.005791754461824894, + 0.0052186246030032635, + 0.004954241681843996, + 0.00609758822247386, + 0.009415270760655403, + 0.004404657054692507, + 0.005828557536005974, + 0.002203174401074648, + 0.004857062827795744, + 0.00669664004817605, + 0.004590230993926525, + 0.018188638612627983, + 0.0023278698790818453, + 0.001580001087859273, + 0.004724186845123768, + 0.0021044122986495495, + 0.002693958580493927, + 0.002361136022955179, + 0.0022492804564535618, + 0.003139317501336336, + 0.0017411881126463413, + 0.0015539589803665876, + 0.0020025933627039194, + 0.0015226545510813594, + 0.002212675753980875, + 0.0017644185572862625, + 0.004321695771068335, + 0.002930218819528818, + 0.0032851137220859528, + 0.0025829493533819914, + 0.006364087574183941, + 0.0036901950370520353, + 0.004413021262735128, + 0.0024689347483217716, + 0.002694763708859682, + 0.003956087864935398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.038142282515764236 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796700924634933 + ], + "opposite_logprob_tokens": [ + -3.529796600341797 + ], + "logprob_reward": -0.029796700924634933, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796700924634933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08960510790348053, + 0.6242681741714478, + 0.036092888563871384, + 0.07391205430030823, + 0.02281617745757103, + 0.042805664241313934, + 0.017582736909389496, + 0.02240641601383686, + 0.01793903484940529, + 0.052571747452020645 + ], + "credit_map": [ + 0.008671093732118607, + 0.007186477538198233, + 0.01056778896600008, + 0.007569896522909403, + 0.007077097427099943, + 0.01656360924243927, + 0.014008336700499058, + 0.0179608054459095, + 0.02209979109466076, + 0.013266971334815025, + 0.015677912160754204, + 0.017143450677394867, + 0.033224333077669144, + 0.24964360892772675, + 0.23598143458366394, + 0.03723069280385971, + 0.003767260815948248, + 0.0037078741006553173, + 0.003010013373568654, + 0.003326273523271084, + 0.00685355206951499, + 0.004663595464080572, + 0.005437837447971106, + 0.005326480604708195, + 0.0074417851865291595, + 0.01047285832464695, + 0.00746297650039196, + 0.010567027144134045, + 0.01071910373866558, + 0.009565545246005058, + 0.01037058886140585, + 0.00731216324493289, + 0.002076623495668173, + 0.001908714766614139, + 0.0030124804470688105, + 0.002959922421723604, + 0.0038999072276055813, + 0.003405987285077572, + 0.0032671086955815554, + 0.0022854330018162727, + 0.0037576009053736925, + 0.00293293921276927, + 0.004162715282291174, + 0.004433199763298035, + 0.006921900901943445, + 0.006276431027799845, + 0.008477303199470043, + 0.00584357650950551, + 0.002941045444458723, + 0.0012780452379956841, + 0.001985775073990226, + 0.0008998717530630529, + 0.0012207061517983675, + 0.0018686165567487478, + 0.001724882167764008, + 0.005663794931024313, + 0.0014772254507988691, + 0.0018123494228348136, + 0.00450004730373621, + 0.002066149841994047, + 0.0026844569947570562, + 0.0021413664799183607, + 0.004388948902487755, + 0.003335873130708933, + 0.002239811234176159, + 0.001622433657757938, + 0.0020734991412609816, + 0.0018553045811131597, + 0.0020240081939846277, + 0.0011770945275202394, + 0.004099742043763399, + 0.0028471408877521753, + 0.005786634515970945, + 0.00511180330067873, + 0.016203347593545914, + 0.007997673004865646, + 0.004941884893923998, + 0.004085289314389229, + 0.003575432812795043, + 0.004869684111326933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.029796700924634933 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016063440591096878 + ], + "opposite_logprob_tokens": [ + -4.141063213348389 + ], + "logprob_reward": -0.016063440591096878, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016063440591096878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03493587300181389, + 0.8000900149345398, + 0.0447157584130764, + 0.012273251079022884, + 0.020886415615677834, + 0.01821509376168251, + 0.008532377891242504, + 0.01481793262064457, + 0.014760567806661129, + 0.030772803351283073 + ], + "credit_map": [ + 0.002167974365875125, + 0.0023725582286715508, + 0.002184065990149975, + 0.0038338275626301765, + 0.0017062932020053267, + 0.011730656959116459, + 0.0074654193595051765, + 0.0034750704653561115, + 0.028156641870737076, + 0.01662767305970192, + 0.019028401002287865, + 0.02147255465388298, + 0.04308962821960449, + 0.32031190395355225, + 0.304343044757843, + 0.04706006869673729, + 0.0032781055197119713, + 0.005927569232881069, + 0.003758449573069811, + 0.0040415385738015175, + 0.011330029927194118, + 0.006729901768267155, + 0.005977860186249018, + 0.003672298276796937, + 0.0012362838024273515, + 0.0013335268013179302, + 0.0009515599813312292, + 0.0013050641864538193, + 0.0025568762794137, + 0.0016258703544735909, + 0.0019096586620435119, + 0.001354409847408533, + 0.0017561576096341014, + 0.00146630615927279, + 0.0017147971084341407, + 0.003137787338346243, + 0.0035207513719797134, + 0.003264632308855653, + 0.003411567537114024, + 0.002614413620904088, + 0.0015373053029179573, + 0.001204044558107853, + 0.001448088907636702, + 0.001235822681337595, + 0.004085793625563383, + 0.0033376181963831186, + 0.0027164150960743427, + 0.0026500041130930185, + 0.0007779322331771255, + 0.0007650416810065508, + 0.0014655061531811953, + 0.0005950412014499307, + 0.001558572519570589, + 0.0011760630877688527, + 0.0011146176839247346, + 0.0010796021670103073, + 0.0016029183752834797, + 0.0014643536414951086, + 0.002612706273794174, + 0.0012545683421194553, + 0.0027862570714205503, + 0.0016206017462536693, + 0.0016952978912740946, + 0.001781228929758072, + 0.0016176592325791717, + 0.0010493656154721975, + 0.0012681055814027786, + 0.0015055217081680894, + 0.002398611744865775, + 0.001107219373807311, + 0.003413507482036948, + 0.002400575904175639, + 0.003559474367648363, + 0.0024237202014774084, + 0.0030251338612288237, + 0.003685650648549199, + 0.007815972901880741, + 0.0030706380493938923, + 0.0032743369229137897, + 0.0039178733713924885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.016063440591096878 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00318829040043056 + ], + "opposite_logprob_tokens": [ + -5.753188133239746 + ], + "logprob_reward": -0.00318829040043056, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00318829040043056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10286969691514969, + 0.21447937190532684, + 0.046137019991874695, + 0.06796894967556, + 0.04204954206943512, + 0.07876866310834885, + 0.14149291813373566, + 0.09275393933057785, + 0.09333325177431107, + 0.12014658004045486 + ], + "credit_map": [ + 0.011011766269803047, + 0.006861696019768715, + 0.009972775354981422, + 0.00824045017361641, + 0.0059174965135753155, + 0.030456483364105225, + 0.01729389652609825, + 0.013115146197378635, + 0.007658431306481361, + 0.00826612114906311, + 0.008885113522410393, + 0.00692851422354579, + 0.010702650994062424, + 0.08175861835479736, + 0.07802547514438629, + 0.012254452332854271, + 0.0047384584322571754, + 0.004836620297282934, + 0.003994092810899019, + 0.00420443806797266, + 0.008734394796192646, + 0.0062943631783127785, + 0.006073804572224617, + 0.007260847836732864, + 0.004981901030987501, + 0.008416750468313694, + 0.011133497580885887, + 0.009321054443717003, + 0.00823922548443079, + 0.008507085032761097, + 0.011381015181541443, + 0.005988426972180605, + 0.005194546654820442, + 0.0027927313931286335, + 0.006066474132239819, + 0.005943791940808296, + 0.0056943949311971664, + 0.00610853498801589, + 0.005779760889708996, + 0.004469305742532015, + 0.013278570026159286, + 0.006359323859214783, + 0.012367083691060543, + 0.006606840528547764, + 0.008154312148690224, + 0.010211929678916931, + 0.010741399601101875, + 0.011049207299947739, + 0.026632357388734818, + 0.00992653053253889, + 0.015005948022007942, + 0.003887266619130969, + 0.005128256045281887, + 0.018392834812402725, + 0.009381975047290325, + 0.05313775688409805, + 0.00921846367418766, + 0.009813282638788223, + 0.023558281362056732, + 0.009729389101266861, + 0.006594456732273102, + 0.005810397677123547, + 0.010958442464470863, + 0.017071237787604332, + 0.009685812518000603, + 0.01202414184808731, + 0.01754671335220337, + 0.008679945021867752, + 0.013460611924529076, + 0.008484854362905025, + 0.012116060592234135, + 0.011335119605064392, + 0.013974899426102638, + 0.015088093467056751, + 0.03427547961473465, + 0.014618203975260258, + 0.010722018778324127, + 0.009640445932745934, + 0.010467622429132462, + 0.01135983131825924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -0.00318829040043056 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins in close-up, then slowly and steadily pulls back to a medium shot, initially focusing on a pair of calloused, yet remarkably steady hands. Carefully, these hands grip a fine chisel, gleaming with a cold metallic sheen. With a steady and deliberate rhythm, the tip of the knife glides and slices across a naturally formed, rough-surfaced dead branch. Fine wood chips gently swirl and drift as the knife moves, and the graceful curve of a bird's wing and the outline of its head are gradually freed from the wood. As the camera pulls away, we see the full extent of the craftsman: an Asian man with graying hair tied in a small bun, wearing a simple dark blue cotton-linen shirt with sleeves rolled up to the elbows. His face wears an expression of almost meditative concentration. He sits in a sun-drenched studio, where the warm afternoon sun slants through a large floor-to-ceiling window, creating distinct beams of light in the air and illuminating the dancing dust particles. In the background, wooden shelves are cluttered with a variety of carving tools, unfinished works, and various woods. The entire scene is bathed in a warm golden hue, highlighting the natural grain of the wood and the wrinkles on the artisans' skin, creating an artistic atmosphere of tranquility, concentration, and creative beauty.", + "seed": 23495, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_6/index_443_seed_23495_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028362207254394887, + "vlm_logprob": -0.028362207254394887, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.028362207254394887, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03069143183529377 + ], + "opposite_logprob_tokens": [ + -4.405691623687744 + ], + "logprob_reward": -0.03069143183529377, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.03069143183529377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2280196100473404, + 0.11277446150779724, + 0.10484891384840012, + 0.0872347429394722, + 0.10011864453554153, + 0.19871442019939423, + 0.032728713005781174, + 0.05225822702050209, + 0.0311407633125782, + 0.052161529660224915 + ], + "credit_map": [ + 0.026017501950263977, + 0.02198532596230507, + 0.03947419300675392, + 0.044904179871082306, + 0.019902924075722694, + 0.020501624792814255, + 0.03775791451334953, + 0.017475947737693787, + 0.01292289886623621, + 0.007670653983950615, + 0.01457263808697462, + 0.03711312264204025, + 0.007036207243800163, + 0.006941674742847681, + 0.018695969134569168, + 0.00782129354774952, + 0.013370668515563011, + 0.01228150725364685, + 0.020817939192056656, + 0.02047899179160595, + 0.007475610356777906, + 0.0051376367919147015, + 0.0178745836019516, + 0.007411976810544729, + 0.006099848076701164, + 0.015096931718289852, + 0.014440599828958511, + 0.019803626462817192, + 0.01159780751913786, + 0.007487430702894926, + 0.006293293088674545, + 0.006415197160094976, + 0.008410343900322914, + 0.015813563019037247, + 0.011450842954218388, + 0.020886432379484177, + 0.010742018930613995, + 0.010621014051139355, + 0.013505037873983383, + 0.008689395152032375, + 0.011448441073298454, + 0.013207212090492249, + 0.037831805646419525, + 0.0654940977692604, + 0.01130345743149519, + 0.036585357040166855, + 0.013215011917054653, + 0.00962905678898096, + 0.0029409737326204777, + 0.002259022556245327, + 0.005411751568317413, + 0.003490305505692959, + 0.006022811867296696, + 0.005640803370624781, + 0.0030525210313498974, + 0.003910520114004612, + 0.007987722754478455, + 0.004107625223696232, + 0.005421945825219154, + 0.005436442792415619, + 0.009401887655258179, + 0.004471798427402973, + 0.007951028645038605, + 0.007479778025299311, + 0.003416522406041622, + 0.0030596156138926744, + 0.00294909137301147, + 0.004008992109447718, + 0.007521718740463257, + 0.002512571169063449, + 0.004968372639268637, + 0.002703877864405513, + 0.011838906444609165, + 0.005758324172347784, + 0.011792848818004131, + 0.002250141464173794, + 0.003361410927027464, + 0.01216870080679655, + 0.002642739797011018, + 0.002348459791392088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.03069143183529377 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207573413848877 + ], + "opposite_logprob_tokens": [ + -2.812075614929199 + ], + "logprob_reward": -0.06207573413848877, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207573413848877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24120871722698212, + 0.0703861266374588, + 0.26860544085502625, + 0.04330035299062729, + 0.035452019423246384, + 0.2412949502468109, + 0.023350616917014122, + 0.02154666744172573, + 0.04051285982131958, + 0.01434226892888546 + ], + "credit_map": [ + 0.021422216668725014, + 0.029575040563941002, + 0.039090316742658615, + 0.05444122850894928, + 0.01877111755311489, + 0.02169065922498703, + 0.04039424657821655, + 0.015823878347873688, + 0.010890491306781769, + 0.004500758834183216, + 0.006590350531041622, + 0.021729830652475357, + 0.00546080619096756, + 0.004618873354047537, + 0.009570920839905739, + 0.007024086080491543, + 0.028708970174193382, + 0.025834107771515846, + 0.06335441023111343, + 0.07790577411651611, + 0.011946571990847588, + 0.009902207180857658, + 0.03837168589234352, + 0.012581728398799896, + 0.0032930714078247547, + 0.006949893664568663, + 0.00723792240023613, + 0.00854627788066864, + 0.005515760742127895, + 0.005031989887356758, + 0.0033773486502468586, + 0.003348086029291153, + 0.003457817016169429, + 0.007158332969993353, + 0.003057950409129262, + 0.006089820526540279, + 0.0036612467374652624, + 0.003857978153973818, + 0.004251015372574329, + 0.00391785753890872, + 0.010558453388512135, + 0.01313109789043665, + 0.08785591274499893, + 0.02832973189651966, + 0.00980000663548708, + 0.07065451890230179, + 0.013168465346097946, + 0.007796790916472673, + 0.0029003270901739597, + 0.002295935992151499, + 0.005722136236727238, + 0.002958388300612569, + 0.002779551548883319, + 0.002526841824874282, + 0.002230837708339095, + 0.0019365978660061955, + 0.00366170983761549, + 0.0018509207293391228, + 0.002935208845883608, + 0.0024637803435325623, + 0.003541397862136364, + 0.0017773493891581893, + 0.0032352444250136614, + 0.002081055659800768, + 0.0046067265793681145, + 0.004420343320816755, + 0.00391462491825223, + 0.004488599486649036, + 0.006609461735934019, + 0.003542958525940776, + 0.0057651628740131855, + 0.007164984010159969, + 0.002500206930562854, + 0.0019049388356506824, + 0.001432672725059092, + 0.0011029005981981754, + 0.002202486153692007, + 0.0027864326257258654, + 0.0014270568499341607, + 0.0009855746757239103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.06207573413848877 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369118645787239 + ], + "opposite_logprob_tokens": [ + -3.408691167831421 + ], + "logprob_reward": -0.03369118645787239, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369118645787239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2628667950630188, + 0.1621779501438141, + 0.24437005817890167, + 0.05536893755197525, + 0.0380953848361969, + 0.0703427791595459, + 0.0418318510055542, + 0.030210131779313087, + 0.08031068742275238, + 0.014425394125282764 + ], + "credit_map": [ + 0.02630024589598179, + 0.02524031512439251, + 0.05013640597462654, + 0.0550413504242897, + 0.025965658947825432, + 0.027528934180736542, + 0.036970119923353195, + 0.015683738514780998, + 0.01705304905772209, + 0.011058536358177662, + 0.02604040503501892, + 0.0592583492398262, + 0.007600817363709211, + 0.009468381293118, + 0.022761031985282898, + 0.00893737655133009, + 0.031255993992090225, + 0.028506554663181305, + 0.052524957805871964, + 0.043799348175525665, + 0.016955696046352386, + 0.011909447610378265, + 0.04464149847626686, + 0.014776562340557575, + 0.004524010233581066, + 0.009039663709700108, + 0.00882571842521429, + 0.011029284447431564, + 0.008122955448925495, + 0.0049447868950665, + 0.00400376133620739, + 0.004878755193203688, + 0.0025069057010114193, + 0.006964999716728926, + 0.004476056434214115, + 0.006317122373729944, + 0.004359547980129719, + 0.0040879822336137295, + 0.0060142152942717075, + 0.0033685562666505575, + 0.005531683564186096, + 0.005945227108895779, + 0.008796964772045612, + 0.01177323330193758, + 0.007467436604201794, + 0.014520489610731602, + 0.006664896849542856, + 0.009642845019698143, + 0.005559882149100304, + 0.004002560395747423, + 0.008535865694284439, + 0.005270357243716717, + 0.0034844984766095877, + 0.0042164926417171955, + 0.004596235696226358, + 0.006165956147015095, + 0.005008273292332888, + 0.0024613027926534414, + 0.0029204552993178368, + 0.003642292693257332, + 0.005221404600888491, + 0.002606960479170084, + 0.00438799848780036, + 0.003961443435400724, + 0.013115844689309597, + 0.009134001098573208, + 0.010571892373263836, + 0.009431611746549606, + 0.01121089793741703, + 0.00743268895894289, + 0.008234483189880848, + 0.011179273948073387, + 0.001890763407573104, + 0.0026308970991522074, + 0.001519414596259594, + 0.0015011228388175368, + 0.002264512935653329, + 0.00173159537371248, + 0.0018437975086271763, + 0.0010432912968099117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.03369118645787239 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537707574665546 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537707574665546, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537707574665546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12575462460517883, + 0.12635137140750885, + 0.05426720529794693, + 0.048101164400577545, + 0.04560794681310654, + 0.19507843255996704, + 0.06555018573999405, + 0.14804618060588837, + 0.14547638595104218, + 0.045766498893499374 + ], + "credit_map": [ + 0.012820105068385601, + 0.016199255362153053, + 0.020054830238223076, + 0.03045899234712124, + 0.012330424040555954, + 0.008378640748560429, + 0.016462206840515137, + 0.009050161577761173, + 0.0201468076556921, + 0.009772061370313168, + 0.022284960374236107, + 0.016466770321130753, + 0.01075571309775114, + 0.009397530928254128, + 0.01881055347621441, + 0.0187169648706913, + 0.009002315811812878, + 0.006396781653165817, + 0.007793358527123928, + 0.009670116007328033, + 0.006726228632032871, + 0.004714358597993851, + 0.006524163298308849, + 0.003439880209043622, + 0.003926802892237902, + 0.008090352639555931, + 0.00950137060135603, + 0.00814684759825468, + 0.006109245587140322, + 0.004995716270059347, + 0.0034344829618930817, + 0.003896343056112528, + 0.003381329821422696, + 0.007202289067208767, + 0.005256007891148329, + 0.0102682588621974, + 0.004689513240009546, + 0.004259111825376749, + 0.006430244538933039, + 0.004121194127947092, + 0.008688881061971188, + 0.010122915729880333, + 0.07867974787950516, + 0.03455430269241333, + 0.005444867070764303, + 0.04207925125956535, + 0.009407016448676586, + 0.00610146252438426, + 0.004900373052805662, + 0.005215628072619438, + 0.006483444478362799, + 0.006506226025521755, + 0.014717639423906803, + 0.014957834035158157, + 0.0055735232308506966, + 0.007195519749075174, + 0.013566229492425919, + 0.012027572840452194, + 0.011026918888092041, + 0.01695641502737999, + 0.03278439864516258, + 0.012632416561245918, + 0.027412477880716324, + 0.0216397512704134, + 0.01177711971104145, + 0.016048278659582138, + 0.011839990504086018, + 0.014926166273653507, + 0.031255919486284256, + 0.01776374690234661, + 0.016992297023534775, + 0.02487286925315857, + 0.007175136357545853, + 0.0058132861740887165, + 0.011181153357028961, + 0.003777043428272009, + 0.006447895430028439, + 0.003132194047793746, + 0.0051455991342663765, + 0.003094190964475274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.012537707574665546 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002814976265653968 + ], + "opposite_logprob_tokens": [ + -5.877814769744873 + ], + "logprob_reward": -0.002814976265653968, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002814976265653968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1316930055618286, + 0.20350967347621918, + 0.08269645273685455, + 0.03548889979720116, + 0.02945256046950817, + 0.186595156788826, + 0.09983939677476883, + 0.07391669601202011, + 0.08498579263687134, + 0.0718223974108696 + ], + "credit_map": [ + 0.013567733578383923, + 0.014740602113306522, + 0.02216256782412529, + 0.027392162010073662, + 0.011214708909392357, + 0.010213183239102364, + 0.022527428343892097, + 0.009874608367681503, + 0.024711906909942627, + 0.01324041374027729, + 0.028887737542390823, + 0.0745277926325798, + 0.009555446915328503, + 0.008773867972195148, + 0.032168373465538025, + 0.011644135229289532, + 0.00905167032033205, + 0.007428556215018034, + 0.019360922276973724, + 0.019879227504134178, + 0.006108751520514488, + 0.003098219633102417, + 0.012671178206801414, + 0.005097922869026661, + 0.0027838258538395166, + 0.0060833222232759, + 0.005893356632441282, + 0.008166482672095299, + 0.0038628948386758566, + 0.0033741334918886423, + 0.00265170494094491, + 0.002673177281394601, + 0.002967454958707094, + 0.004724887199699879, + 0.0037022095639258623, + 0.005588672123849392, + 0.0026243424508720636, + 0.0027565625496208668, + 0.0034887290094047785, + 0.0035997016821056604, + 0.007748703472316265, + 0.010125222615897655, + 0.06962724030017853, + 0.023115480318665504, + 0.006583265028893948, + 0.05343189090490341, + 0.010161389596760273, + 0.005801951512694359, + 0.01161176711320877, + 0.008282588794827461, + 0.017120692878961563, + 0.010418654419481754, + 0.015537736006081104, + 0.01995657943189144, + 0.009401853196322918, + 0.007509525865316391, + 0.009859049692749977, + 0.00885318499058485, + 0.01435320358723402, + 0.007639692164957523, + 0.010873126797378063, + 0.0066834972240030766, + 0.009666258469223976, + 0.005988680757582188, + 0.015986816957592964, + 0.012177848257124424, + 0.012958911247551441, + 0.009551946073770523, + 0.006833762861788273, + 0.013306017965078354, + 0.007046522106975317, + 0.007123959716409445, + 0.013104517012834549, + 0.009144121780991554, + 0.024484021589159966, + 0.00623607961460948, + 0.0059718783013522625, + 0.005569330882281065, + 0.004329914227128029, + 0.0029825225938111544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.002814976265653968 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins in close-up, then slowly and steadily pulls back to a medium shot, initially focusing on a pair of calloused, yet remarkably steady hands. Carefully, these hands grip a fine chisel, gleaming with a cold metallic sheen. With a steady and deliberate rhythm, the tip of the knife glides and slices across a naturally formed, rough-surfaced dead branch. Fine wood chips gently swirl and drift as the knife moves, and the graceful curve of a bird's wing and the outline of its head are gradually freed from the wood. As the camera pulls away, we see the full extent of the craftsman: an Asian man with graying hair tied in a small bun, wearing a simple dark blue cotton-linen shirt with sleeves rolled up to the elbows. His face wears an expression of almost meditative concentration. He sits in a sun-drenched studio, where the warm afternoon sun slants through a large floor-to-ceiling window, creating distinct beams of light in the air and illuminating the dancing dust particles. In the background, wooden shelves are cluttered with a variety of carving tools, unfinished works, and various woods. The entire scene is bathed in a warm golden hue, highlighting the natural grain of the wood and the wrinkles on the artisans' skin, creating an artistic atmosphere of tranquility, concentration, and creative beauty.", + "seed": 23496, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_6/index_443_seed_23496_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.018259232118725776, + "vlm_logprob": -0.018259232118725776, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.018259232118725776, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020178280770778656 + ], + "opposite_logprob_tokens": [ + -5.520178318023682 + ], + "logprob_reward": -0.020178280770778656, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.020178280770778656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.154694601893425, + 0.08261780440807343, + 0.09780940413475037, + 0.08767005801200867, + 0.10332486033439636, + 0.10544860363006592, + 0.12481418251991272, + 0.06216317042708397, + 0.11688625067472458, + 0.06457104533910751 + ], + "credit_map": [ + 0.02426246553659439, + 0.01107318140566349, + 0.015264313668012619, + 0.020673882216215134, + 0.01633378118276596, + 0.025533752515912056, + 0.017519844695925713, + 0.024033401161432266, + 0.010183434933423996, + 0.0057071722112596035, + 0.007685121614485979, + 0.013042225502431393, + 0.01335942093282938, + 0.011345439590513706, + 0.010215663351118565, + 0.01107933558523655, + 0.007418894208967686, + 0.009479016996920109, + 0.009182178415358067, + 0.01777845062315464, + 0.012914677150547504, + 0.013822939246892929, + 0.0104390112683177, + 0.01677425019443035, + 0.007431421894580126, + 0.008239153772592545, + 0.008500779047608376, + 0.013543860986828804, + 0.013021367602050304, + 0.014704737812280655, + 0.00937211886048317, + 0.012856623157858849, + 0.014052024111151695, + 0.011432311497628689, + 0.013287469744682312, + 0.01355350948870182, + 0.012431505136191845, + 0.015137620270252228, + 0.01029205322265625, + 0.01313837431371212, + 0.009547904133796692, + 0.015891065821051598, + 0.00866627786308527, + 0.010609081946313381, + 0.016730323433876038, + 0.018825098872184753, + 0.011450037360191345, + 0.01372881792485714, + 0.011870448477566242, + 0.015881892293691635, + 0.017810892313718796, + 0.01371302455663681, + 0.016575198620557785, + 0.01680808514356613, + 0.013999749906361103, + 0.018154894933104515, + 0.007460758090019226, + 0.006316494196653366, + 0.008321334607899189, + 0.005410651210695505, + 0.009776282124221325, + 0.010357103310525417, + 0.006063802633434534, + 0.008456746116280556, + 0.009276612661778927, + 0.009221530519425869, + 0.01370156928896904, + 0.013854040764272213, + 0.018052784726023674, + 0.015434697270393372, + 0.023565661162137985, + 0.01377936638891697, + 0.005753566976636648, + 0.00632481649518013, + 0.008982609957456589, + 0.005427234806120396, + 0.01200201828032732, + 0.010197367519140244, + 0.008023842237889767, + 0.007859589532017708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.020178280770778656 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029846802353858948 + ], + "opposite_logprob_tokens": [ + -3.5298469066619873 + ], + "logprob_reward": -0.029846802353858948, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029846802353858948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3427249789237976, + 0.053601112216711044, + 0.13830351829528809, + 0.05569271370768547, + 0.1314697414636612, + 0.0676734447479248, + 0.05844701826572418, + 0.06002478301525116, + 0.053724683821201324, + 0.038338012993335724 + ], + "credit_map": [ + 0.024026526138186455, + 0.019064895808696747, + 0.022180184721946716, + 0.09126761555671692, + 0.05049871280789375, + 0.05612283945083618, + 0.03626025468111038, + 0.04330391436815262, + 0.008436731062829494, + 0.00454887468367815, + 0.005841794889420271, + 0.005214745178818703, + 0.008473673835396767, + 0.006894029676914215, + 0.006046139635145664, + 0.008145119063556194, + 0.011783815920352936, + 0.011996178887784481, + 0.009621766395866871, + 0.022644925862550735, + 0.016835922375321388, + 0.020838439464569092, + 0.01664152182638645, + 0.027940943837165833, + 0.005847640335559845, + 0.005555647891014814, + 0.007146606687456369, + 0.008194491267204285, + 0.007762521505355835, + 0.008403800427913666, + 0.00548543082550168, + 0.007296570111066103, + 0.014373383484780788, + 0.010611517354846, + 0.023950714617967606, + 0.019562194123864174, + 0.013287581503391266, + 0.021447783336043358, + 0.009501541964709759, + 0.018735019490122795, + 0.0059092119336128235, + 0.006488748360425234, + 0.006293312646448612, + 0.011880839243531227, + 0.010678422637283802, + 0.010202880017459393, + 0.006454102694988251, + 0.00976592767983675, + 0.005876939278095961, + 0.007124403491616249, + 0.009546174667775631, + 0.0045087942853569984, + 0.007629933301359415, + 0.00862794741988182, + 0.006510349456220865, + 0.008622473105788231, + 0.00424071354791522, + 0.004902829881757498, + 0.007944833487272263, + 0.0044266050681471825, + 0.010037485510110855, + 0.013705973513424397, + 0.007610478904098272, + 0.007155856117606163, + 0.0031532724387943745, + 0.004387926775962114, + 0.0065780384466052055, + 0.005977989640086889, + 0.012212340719997883, + 0.00764632411301136, + 0.0076345414854586124, + 0.006134245544672012, + 0.004096880555152893, + 0.004647535737603903, + 0.006698460318148136, + 0.0025679331738501787, + 0.0075146108865737915, + 0.004815469961613417, + 0.004954927135258913, + 0.0030421928968280554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.029846802353858948 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057752013206482 + ], + "opposite_logprob_tokens": [ + -3.8955774307250977 + ], + "logprob_reward": -0.02057752013206482, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057752013206482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10218524932861328, + 0.08557184040546417, + 0.16502808034420013, + 0.029098453000187874, + 0.19037684798240662, + 0.2735641300678253, + 0.04142028093338013, + 0.02685588039457798, + 0.041359055787324905, + 0.04454021528363228 + ], + "credit_map": [ + 0.008286834694445133, + 0.008697432465851307, + 0.006108790636062622, + 0.00945486780256033, + 0.014902078546583652, + 0.024714943021535873, + 0.011614121496677399, + 0.018406184390187263, + 0.00877199787646532, + 0.004586472641676664, + 0.008201481774449348, + 0.009047649800777435, + 0.01637829653918743, + 0.008409978821873665, + 0.013761655427515507, + 0.016414301469922066, + 0.030104853212833405, + 0.011753411032259464, + 0.01935066655278206, + 0.03148689121007919, + 0.019036617130041122, + 0.02481112815439701, + 0.013053115457296371, + 0.015431403182446957, + 0.002280480694025755, + 0.0029166073072701693, + 0.0037060249596834183, + 0.004198484588414431, + 0.0030660151969641447, + 0.004606439732015133, + 0.0030022088903933764, + 0.0053221918642520905, + 0.027870429679751396, + 0.014434999786317348, + 0.037491172552108765, + 0.019606024026870728, + 0.016357216984033585, + 0.03283641114830971, + 0.013660654425621033, + 0.02811993472278118, + 0.036261603236198425, + 0.029257548972964287, + 0.024096252396702766, + 0.038499124348163605, + 0.0537167564034462, + 0.04100479185581207, + 0.023470010608434677, + 0.027258018031716347, + 0.0036805474665015936, + 0.004139807540923357, + 0.005132138729095459, + 0.004328782670199871, + 0.004089668393135071, + 0.005124930292367935, + 0.00875563733279705, + 0.006168769672513008, + 0.0032550932373851538, + 0.0021125804632902145, + 0.003099798923358321, + 0.0023480835370719433, + 0.00445198267698288, + 0.005470144562423229, + 0.002752928528934717, + 0.003365268697962165, + 0.004031266551464796, + 0.003471770090982318, + 0.005051009356975555, + 0.005924959667026997, + 0.005852083675563335, + 0.004683289676904678, + 0.008788828738033772, + 0.003555849427357316, + 0.0037771358620375395, + 0.006601195316761732, + 0.009536940604448318, + 0.002490619197487831, + 0.006862714886665344, + 0.007236687932163477, + 0.005111830774694681, + 0.002923091873526573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.02057752013206482 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606050878763199 + ], + "opposite_logprob_tokens": [ + -4.1410603523254395 + ], + "logprob_reward": -0.01606050878763199, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606050878763199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049670711159706116, + 0.038725804537534714, + 0.026479359716176987, + 0.031267501413822174, + 0.10230866074562073, + 0.1711118370294571, + 0.07244572788476944, + 0.18541088700294495, + 0.1336900144815445, + 0.1888895183801651 + ], + "credit_map": [ + 0.005356559529900551, + 0.0031481930054724216, + 0.005545055493712425, + 0.004088196903467178, + 0.007574328221380711, + 0.009775888174772263, + 0.007443527691066265, + 0.006738961674273014, + 0.007429314777255058, + 0.002041017170995474, + 0.0045337192714214325, + 0.00334661896340549, + 0.005233566276729107, + 0.004517241846770048, + 0.0045214504934847355, + 0.007102873641997576, + 0.0028427334036678076, + 0.0020575777161866426, + 0.003081762231886387, + 0.003986065275967121, + 0.0036794173065572977, + 0.004710320383310318, + 0.0018882901640608907, + 0.004233191721141338, + 0.002715560607612133, + 0.0026124194264411926, + 0.002981295809149742, + 0.004747237078845501, + 0.004611093085259199, + 0.0049054608680307865, + 0.0037959895562380552, + 0.00489844148978591, + 0.016007402911782265, + 0.009298515506088734, + 0.015474017709493637, + 0.014760789461433887, + 0.01143613737076521, + 0.015463020652532578, + 0.007491419091820717, + 0.012377358973026276, + 0.01509262528270483, + 0.019656863063573837, + 0.01919582672417164, + 0.020766787230968475, + 0.03450029343366623, + 0.0269270408898592, + 0.016241544857621193, + 0.01873083785176277, + 0.005616598762571812, + 0.007885701954364777, + 0.007292426191270351, + 0.007891467772424221, + 0.007999560795724392, + 0.009803877212107182, + 0.014397864229977131, + 0.011558230966329575, + 0.019748076796531677, + 0.020726297050714493, + 0.02493421919643879, + 0.015095152892172337, + 0.026598863303661346, + 0.03331451863050461, + 0.025559144094586372, + 0.0194346122443676, + 0.006620252970606089, + 0.010424607433378696, + 0.01419212855398655, + 0.016670387238264084, + 0.03735407069325447, + 0.012384971603751183, + 0.022200308740139008, + 0.013843276537954807, + 0.015165939927101135, + 0.012496959418058395, + 0.019542813301086426, + 0.013000847771763802, + 0.04500993341207504, + 0.031918495893478394, + 0.026941027492284775, + 0.024813495576381683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.01606050878763199 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot that slowly pulls back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands shown as calloused, steady, and gripping a fine chisel with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted as the chisel slicing a rough-surfaced dead branch to reveal a bird's wing and head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the craftsman an Asian man with graying hair in a bun, wearing a dark blue cotton-linen shirt with rolled-up sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set in a sun-drenched studio with warm golden light beams, dust particles, and wooden shelves in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633048549294472 + ], + "opposite_logprob_tokens": [ + -5.379632949829102 + ], + "logprob_reward": -0.004633048549294472, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633048549294472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17296439409255981, + 0.06459666043519974, + 0.021542517468333244, + 0.02182806096971035, + 0.05132705718278885, + 0.11093136668205261, + 0.08785289525985718, + 0.09785772114992142, + 0.21563029289245605, + 0.15546901524066925 + ], + "credit_map": [ + 0.010766963474452496, + 0.009885424748063087, + 0.009073009714484215, + 0.028311068192124367, + 0.03142944350838661, + 0.038387615233659744, + 0.01899251714348793, + 0.02611834928393364, + 0.009350699372589588, + 0.0043164934031665325, + 0.006225447170436382, + 0.008912025019526482, + 0.008213570341467857, + 0.006717613898217678, + 0.009192261844873428, + 0.01166855450719595, + 0.0028803758323192596, + 0.0015580208273604512, + 0.0027684858068823814, + 0.0036613792181015015, + 0.0026299396995455027, + 0.002394024981185794, + 0.002018615370616317, + 0.003631674451753497, + 0.0026168115437030792, + 0.002137324307113886, + 0.0024897188413888216, + 0.0027420446276664734, + 0.002912677824497223, + 0.0032360865734517574, + 0.002078128047287464, + 0.003615268738940358, + 0.007374875247478485, + 0.005106688477098942, + 0.006685050204396248, + 0.00577674200758338, + 0.004160086624324322, + 0.008456151001155376, + 0.005495072342455387, + 0.008272393606603146, + 0.009026064537465572, + 0.01435006968677044, + 0.012051433324813843, + 0.011749276891350746, + 0.027143562212586403, + 0.01831524260342121, + 0.009501558728516102, + 0.008794154971837997, + 0.007155626080930233, + 0.010142583400011063, + 0.01313929632306099, + 0.008638988249003887, + 0.009154467843472958, + 0.012728276662528515, + 0.014147381298244, + 0.012746271677315235, + 0.01238916628062725, + 0.008601635694503784, + 0.012885420583188534, + 0.007342941593378782, + 0.01690811477601528, + 0.01637141779065132, + 0.012992491945624352, + 0.010366534814238548, + 0.02052200213074684, + 0.021022213622927666, + 0.03391470015048981, + 0.025821013376116753, + 0.03064984828233719, + 0.025218531489372253, + 0.037002068012952805, + 0.02147991769015789, + 0.016033682972192764, + 0.022750234231352806, + 0.03740551695227623, + 0.009271831251680851, + 0.013179415836930275, + 0.024271992966532707, + 0.018604082986712456, + 0.013952265493571758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.004633048549294472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_7/index_tensor([9727, 9727])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_7/index_tensor([9727, 9727])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c4a6808b69a717b647d9ae33e2d26a63b64cece6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_7/index_tensor([9727, 9727])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle, ultra-wide shot captures a sprawling futuristic metropolis at night, its rain-slicked streets like a dark mirror reflecting the spectacle unfolding below. A fleet of enormous, towering robots slowly and gracefully winds its way through a canyon-like avenue formed by massive black glass and steel skyscrapers. Their sleek, vaguely humanoid forms are constructed from matte black alloy that absorbs ambient light. Yet, their entire frames are traced by a complex network of vibrant blue and white lights that rhythmically pulsate, illuminating their path and casting long, shifting shadows. Each massive step descends with silent, ponderous weight, creating noticeable ripples in the puddles of water on the asphalt. The damp ground perfectly reflects the robots' lighting and the ever-changing colors of the city's towering holographic advertisements, which flicker silently in the background. The scene is immersed in a cool cyberpunk color palette of deep blue, electronic magentas, and pure white, creating an atmosphere of silent, awe-inspiring grandeur and technological mystery.", + "seed": 18006, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_7/index_9727_seed_18006_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03581126760691404, + "vlm_logprob": -0.03581126760691404, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.03581126760691404, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011247680522501469 + ], + "opposite_logprob_tokens": [ + -4.886247634887695 + ], + "logprob_reward": -0.011247680522501469, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.011247680522501469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32525748014450073, + 0.09795573353767395, + 0.04806729406118393, + 0.09853368997573853, + 0.05902533605694771, + 0.06942956149578094, + 0.04161805287003517, + 0.033667437732219696, + 0.13318412005901337, + 0.0932612344622612 + ], + "credit_map": [ + 0.02552178129553795, + 0.024604134261608124, + 0.02407381869852543, + 0.03609008342027664, + 0.02599339932203293, + 0.06440328061580658, + 0.036516718566417694, + 0.08805429190397263, + 0.014989642426371574, + 0.011424717493355274, + 0.007363489363342524, + 0.014168026857078075, + 0.009052655659615993, + 0.0083910021930933, + 0.019092539325356483, + 0.013473665341734886, + 0.007474330719560385, + 0.004630476236343384, + 0.0036045559681952, + 0.00742145674303174, + 0.005122291389852762, + 0.006501592695713043, + 0.008605778217315674, + 0.004706817679107189, + 0.01182591449469328, + 0.010628285817801952, + 0.01174122467637062, + 0.014316207729279995, + 0.00854759942740202, + 0.011798382736742496, + 0.019057078287005424, + 0.010619008913636208, + 0.012182526290416718, + 0.004313955549150705, + 0.005189821589738131, + 0.010767263360321522, + 0.0046655419282615185, + 0.007971865125000477, + 0.006353107281029224, + 0.007581261917948723, + 0.0076102521270513535, + 0.007928572595119476, + 0.00745160598307848, + 0.012908765114843845, + 0.007659778464585543, + 0.006586302537471056, + 0.012230188585817814, + 0.007054109591990709, + 0.006223526783287525, + 0.0030229215044528246, + 0.005722644738852978, + 0.0048495386727154255, + 0.0051734247244894505, + 0.0046867807395756245, + 0.0060998001135885715, + 0.005839422810822725, + 0.003285276470705867, + 0.0030233038123697042, + 0.004238452296704054, + 0.005384412128478289, + 0.0027404918801039457, + 0.0067055667750537395, + 0.004483826458454132, + 0.003806111868470907, + 0.024691415950655937, + 0.007680215872824192, + 0.011600090190768242, + 0.007772855926305056, + 0.007871368899941444, + 0.04920857772231102, + 0.016645407304167747, + 0.007714206352829933, + 0.010220644064247608, + 0.004752262029796839, + 0.010167980566620827, + 0.0071877846494317055, + 0.00760240713134408, + 0.027188127860426903, + 0.02062966302037239, + 0.005512373987585306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.011247680522501469 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020600641146302223 + ], + "opposite_logprob_tokens": [ + -3.8956005573272705 + ], + "logprob_reward": -0.020600641146302223, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020600641146302223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3005804419517517, + 0.1799112856388092, + 0.04671149700880051, + 0.1389114111661911, + 0.03664880245923996, + 0.0595957450568676, + 0.04604656994342804, + 0.03584517911076546, + 0.040491390973329544, + 0.11525765061378479 + ], + "credit_map": [ + 0.022859474644064903, + 0.022673068568110466, + 0.020507942885160446, + 0.028389349579811096, + 0.022038280963897705, + 0.03151698037981987, + 0.06566262990236282, + 0.0869327187538147, + 0.027010813355445862, + 0.0298655666410923, + 0.01419628132134676, + 0.022548168897628784, + 0.021443720906972885, + 0.017784062772989273, + 0.02651902474462986, + 0.020543642342090607, + 0.0071613541804254055, + 0.004716379102319479, + 0.004705275874584913, + 0.005866418592631817, + 0.005330136511474848, + 0.006002094596624374, + 0.00709170987829566, + 0.005838130600750446, + 0.015796223655343056, + 0.009735412895679474, + 0.015839865431189537, + 0.016399450600147247, + 0.01649383082985878, + 0.0165972001850605, + 0.032683972269296646, + 0.015365460887551308, + 0.004227697383612394, + 0.002226081909611821, + 0.0056816828437149525, + 0.0046528419479727745, + 0.004429427906870842, + 0.004345893394201994, + 0.004388101864606142, + 0.006697074975818396, + 0.007177785504609346, + 0.0047413622960448265, + 0.007712215650826693, + 0.01335248351097107, + 0.004823276773095131, + 0.00677872821688652, + 0.009015207178890705, + 0.005994689650833607, + 0.0052204253152012825, + 0.0036294590681791306, + 0.005778039339929819, + 0.0053319926373660564, + 0.007095306646078825, + 0.005622964818030596, + 0.007009818684309721, + 0.006358564831316471, + 0.0029569489415735006, + 0.0025540455244481564, + 0.0059517743065953255, + 0.0053985691629350185, + 0.004043539986014366, + 0.005213208496570587, + 0.004706595558673143, + 0.005020498298108578, + 0.005267789121717215, + 0.0032826168462634087, + 0.004989671520888805, + 0.0036278855986893177, + 0.006263395771384239, + 0.0091328090056777, + 0.00485267024487257, + 0.003074557287618518, + 0.012455394491553307, + 0.00786413624882698, + 0.014478012919425964, + 0.01622207649052143, + 0.010029155761003494, + 0.025020480155944824, + 0.022981660440564156, + 0.006206733174622059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.020600641146302223 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818905957043171 + ], + "opposite_logprob_tokens": [ + -4.018188953399658 + ], + "logprob_reward": -0.01818905957043171, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818905957043171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1783483773469925, + 0.075662761926651, + 0.09199701994657516, + 0.08006229996681213, + 0.041345350444316864, + 0.09638712555170059, + 0.08533060550689697, + 0.11228927969932556, + 0.1351354867219925, + 0.1034417524933815 + ], + "credit_map": [ + 0.020289350301027298, + 0.01705821044743061, + 0.01716689206659794, + 0.02386533096432686, + 0.018778925761580467, + 0.03545932099223137, + 0.021777600049972534, + 0.023952748626470566, + 0.01290375366806984, + 0.0058686211705207825, + 0.007361449766904116, + 0.00824833381921053, + 0.00613012257963419, + 0.008928151801228523, + 0.016854602843523026, + 0.009367723017930984, + 0.009453194215893745, + 0.007226439192891121, + 0.011321129277348518, + 0.01165208499878645, + 0.01375795528292656, + 0.01189217995852232, + 0.01376153714954853, + 0.012932485900819302, + 0.010438312776386738, + 0.011468030512332916, + 0.006867361254990101, + 0.013613023795187473, + 0.007305437698960304, + 0.008737791329622269, + 0.013066955842077732, + 0.00856537651270628, + 0.004691852256655693, + 0.004415056202560663, + 0.0034731635823845863, + 0.005372261628508568, + 0.007811347488313913, + 0.00590206403285265, + 0.005074243526905775, + 0.0046053603291511536, + 0.013075149618089199, + 0.008869113400578499, + 0.01071621011942625, + 0.015509319491684437, + 0.011651395820081234, + 0.0109894759953022, + 0.013585688546299934, + 0.011990758590400219, + 0.008249395526945591, + 0.008286691270768642, + 0.009082293137907982, + 0.009411421604454517, + 0.019292645156383514, + 0.011151555925607681, + 0.011989319697022438, + 0.007867271080613136, + 0.009347604587674141, + 0.011549003422260284, + 0.013461686670780182, + 0.012377516366541386, + 0.023428522050380707, + 0.02016288973391056, + 0.012934617698192596, + 0.009027425199747086, + 0.02873908542096615, + 0.007595484144985676, + 0.00914929248392582, + 0.007110150996595621, + 0.013432384468615055, + 0.046562936156988144, + 0.01307767629623413, + 0.00946846418082714, + 0.014856123365461826, + 0.006845925003290176, + 0.01122995000332594, + 0.012844272889196873, + 0.01221928745508194, + 0.019060276448726654, + 0.02066437341272831, + 0.005721533205360174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.01818905957043171 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707892060279846 + ], + "opposite_logprob_tokens": [ + -2.1270790100097656 + ], + "logprob_reward": -0.12707892060279846, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707892060279846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13555623590946198, + 0.1499280333518982, + 0.06254436075687408, + 0.22292710840702057, + 0.018090717494487762, + 0.07663962990045547, + 0.02209431491792202, + 0.02850380539894104, + 0.052172593772411346, + 0.23154319822788239 + ], + "credit_map": [ + 0.0147868562489748, + 0.011361056938767433, + 0.017700618132948875, + 0.015840057283639908, + 0.0176247525960207, + 0.013886993750929832, + 0.023139476776123047, + 0.02121642231941223, + 0.023720208555459976, + 0.0201729629188776, + 0.011251356452703476, + 0.01756332255899906, + 0.017558982595801353, + 0.020346764475107193, + 0.024073775857686996, + 0.015240658074617386, + 0.008174818940460682, + 0.004965144209563732, + 0.008725959807634354, + 0.007874232716858387, + 0.008210906758904457, + 0.0076476819813251495, + 0.009054780937731266, + 0.007890833541750908, + 0.024979600682854652, + 0.019908584654331207, + 0.022212505340576172, + 0.027079377323389053, + 0.02798335626721382, + 0.025970853865146637, + 0.05243288725614548, + 0.022359924390912056, + 0.0020942064002156258, + 0.0017837941413745284, + 0.002842629561200738, + 0.001936031156219542, + 0.0019527742406353354, + 0.0023783287033438683, + 0.0028419552836567163, + 0.00226099812425673, + 0.010684269480407238, + 0.007271873764693737, + 0.010203135199844837, + 0.012547112070024014, + 0.007548546884208918, + 0.007845845073461533, + 0.012186850421130657, + 0.008352001197636127, + 0.0028891998808830976, + 0.00194273772649467, + 0.0028212261386215687, + 0.0024978744331747293, + 0.003356592496857047, + 0.0026606516912579536, + 0.0037206728011369705, + 0.002205359051004052, + 0.0026403579395264387, + 0.002937444718554616, + 0.0047493912279605865, + 0.004034769721329212, + 0.002393623348325491, + 0.005124520044773817, + 0.0036068265326321125, + 0.0030168734956532717, + 0.009368141181766987, + 0.005102782044559717, + 0.004528054501861334, + 0.005183378234505653, + 0.00605660630390048, + 0.012018932029604912, + 0.005980554968118668, + 0.003934144973754883, + 0.031038640066981316, + 0.016442621126770973, + 0.025825029239058495, + 0.03534390404820442, + 0.022300923243165016, + 0.04511723667383194, + 0.04319560527801514, + 0.012279240414500237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.12707892060279846 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001940036192536354 + ], + "opposite_logprob_tokens": [ + -6.2519402503967285 + ], + "logprob_reward": -0.001940036192536354, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001940036192536354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31208863854408264, + 0.07830652594566345, + 0.05136192962527275, + 0.06931784749031067, + 0.04430440440773964, + 0.09449303895235062, + 0.04862743988633156, + 0.049029137939214706, + 0.15917076170444489, + 0.09330027550458908 + ], + "credit_map": [ + 0.02496427297592163, + 0.03220069035887718, + 0.02479863353073597, + 0.028517432510852814, + 0.03476141393184662, + 0.042260173708200455, + 0.03771553188562393, + 0.086870476603508, + 0.0084975715726614, + 0.012382226064801216, + 0.0071425228379666805, + 0.010403292253613472, + 0.008684434927999973, + 0.008550671860575676, + 0.013787789270281792, + 0.008858016692101955, + 0.008048992604017258, + 0.004560759756714106, + 0.004514330066740513, + 0.007096387445926666, + 0.007044002879410982, + 0.007070737425237894, + 0.007676753681153059, + 0.00534996809437871, + 0.0074570998549461365, + 0.007754837162792683, + 0.007611044682562351, + 0.010219963267445564, + 0.007271544076502323, + 0.00808791071176529, + 0.014407113194465637, + 0.006508332211524248, + 0.007165180519223213, + 0.003920621704310179, + 0.0038419654592871666, + 0.005618503782898188, + 0.0073922122828662395, + 0.0058035943657159805, + 0.005967710167169571, + 0.004594611003994942, + 0.01480274461209774, + 0.0091560622677207, + 0.010436050593852997, + 0.013796810992062092, + 0.01595957949757576, + 0.01010073907673359, + 0.010968462564051151, + 0.009272591210901737, + 0.006093980278819799, + 0.004491622559726238, + 0.0058544292114675045, + 0.005978131666779518, + 0.008854744955897331, + 0.005713366903364658, + 0.006404663901776075, + 0.0052364966832101345, + 0.006031192373484373, + 0.005126377567648888, + 0.0055358451791107655, + 0.006504690740257502, + 0.006306234281510115, + 0.009494228288531303, + 0.006475315894931555, + 0.003555252682417631, + 0.0373653843998909, + 0.011160289868712425, + 0.008748852647840977, + 0.011374490335583687, + 0.011389561928808689, + 0.054703809320926666, + 0.014234554953873158, + 0.01019381731748581, + 0.013474944047629833, + 0.007155349012464285, + 0.013870365917682648, + 0.01182700414210558, + 0.007814452983438969, + 0.013065964914858341, + 0.021509772166609764, + 0.004582416266202927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.001940036192536354 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, ultra-wide shot captures a sprawling futuristic metropolis at night, its rain-slicked streets like a dark mirror reflecting the spectacle unfolding below. A fleet of enormous, towering robots slowly and gracefully winds its way through a canyon-like avenue formed by massive black glass and steel skyscrapers. Their sleek, vaguely humanoid forms are constructed from matte black alloy that absorbs ambient light. Yet, their entire frames are traced by a complex network of vibrant blue and white lights that rhythmically pulsate, illuminating their path and casting long, shifting shadows. Each massive step descends with silent, ponderous weight, creating noticeable ripples in the puddles of water on the asphalt. The damp ground perfectly reflects the robots' lighting and the ever-changing colors of the city's towering holographic advertisements, which flicker silently in the background. The scene is immersed in a cool cyberpunk color palette of deep blue, electronic magentas, and pure white, creating an atmosphere of silent, awe-inspiring grandeur and technological mystery.", + "seed": 18007, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_7/index_9727_seed_18007_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.021509449754375964, + "vlm_logprob": -0.021509449754375964, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.021509449754375964, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015738792717456818 + ], + "opposite_logprob_tokens": [ + -4.3907389640808105 + ], + "logprob_reward": -0.015738792717456818, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.015738792717456818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11312904208898544, + 0.027870338410139084, + 0.20251712203025818, + 0.3097836375236511, + 0.054443348199129105, + 0.07152067869901657, + 0.08261244744062424, + 0.05373488739132881, + 0.03732113540172577, + 0.047067414969205856 + ], + "credit_map": [ + 0.01072343997657299, + 0.010381164029240608, + 0.018253374844789505, + 0.023907843977212906, + 0.009240022860467434, + 0.013471852988004684, + 0.017413733527064323, + 0.009737609885632992, + 0.004516438115388155, + 0.002305047120898962, + 0.004085750784724951, + 0.004068445414304733, + 0.002912299707531929, + 0.004202313255518675, + 0.0030607234220951796, + 0.002719318028539419, + 0.010254927910864353, + 0.007974646985530853, + 0.009863938204944134, + 0.007479080930352211, + 0.011361334472894669, + 0.07518475502729416, + 0.06794145703315735, + 0.012456988915801048, + 0.039778467267751694, + 0.012996768578886986, + 0.022466981783509254, + 0.1037834882736206, + 0.01438214536756277, + 0.015266046859323978, + 0.08724762499332428, + 0.013862106017768383, + 0.008953726850450039, + 0.008090517483651638, + 0.011852294206619263, + 0.006868869997560978, + 0.0036768552381545305, + 0.005188873969018459, + 0.0031180549412965775, + 0.006694154813885689, + 0.009944461286067963, + 0.00845512468367815, + 0.012655352242290974, + 0.009469568729400635, + 0.004690075293183327, + 0.00915150623768568, + 0.007382308132946491, + 0.00977227371186018, + 0.010123427957296371, + 0.014223670586943626, + 0.0111395213752985, + 0.008217277005314827, + 0.007532824762165546, + 0.01237641554325819, + 0.008798152208328247, + 0.010201151482760906, + 0.009411850944161415, + 0.0077799418941140175, + 0.008988023735582829, + 0.005737064406275749, + 0.006010253913700581, + 0.005941806826740503, + 0.005528978072106838, + 0.004336967132985592, + 0.0015687572304159403, + 0.00629265746101737, + 0.0035352110862731934, + 0.01304990891367197, + 0.0046029360964894295, + 0.0024597388692200184, + 0.0036276287864893675, + 0.002184296725317836, + 0.0030240900814533234, + 0.005450003314763308, + 0.0036208981182426214, + 0.008905027061700821, + 0.0052733696065843105, + 0.005716875195503235, + 0.008238205686211586, + 0.006838949862867594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.015738792717456818 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210226014256477 + ], + "opposite_logprob_tokens": [ + -4.264210224151611 + ], + "logprob_reward": -0.014210226014256477, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210226014256477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2194279283285141, + 0.11294207721948624, + 0.07516633719205856, + 0.04481089860200882, + 0.078824982047081, + 0.07376954704523087, + 0.13859963417053223, + 0.028786296024918556, + 0.06408898532390594, + 0.16358330845832825 + ], + "credit_map": [ + 0.01831694319844246, + 0.02062881551682949, + 0.03201741352677345, + 0.05067431181669235, + 0.01713075488805771, + 0.030863426625728607, + 0.03301365673542023, + 0.016782624647021294, + 0.011847141198813915, + 0.011784607544541359, + 0.015641607344150543, + 0.021810872480273247, + 0.013642843812704086, + 0.013294135220348835, + 0.013533171266317368, + 0.01138770580291748, + 0.008704646490514278, + 0.006523866206407547, + 0.010453037917613983, + 0.011907628737390041, + 0.006196134258061647, + 0.011176771484315395, + 0.010512708686292171, + 0.009691544808447361, + 0.007232964970171452, + 0.004397031385451555, + 0.005278145428746939, + 0.008403421379625797, + 0.003593121888116002, + 0.004557435866445303, + 0.004937442019581795, + 0.006411338225007057, + 0.011381478980183601, + 0.01167832687497139, + 0.010714732110500336, + 0.009800776839256287, + 0.006468925625085831, + 0.01076030358672142, + 0.008426124230027199, + 0.009594318456947803, + 0.01056825090199709, + 0.011389466002583504, + 0.012063547968864441, + 0.008695892989635468, + 0.007779012434184551, + 0.006447657011449337, + 0.007251468952745199, + 0.00957425870001316, + 0.028732918202877045, + 0.02241075225174427, + 0.0207691490650177, + 0.0085346270352602, + 0.014139831997454166, + 0.0179318655282259, + 0.010280135087668896, + 0.015800362452864647, + 0.005291405599564314, + 0.004150091670453548, + 0.003413455095142126, + 0.003748971037566662, + 0.00375514873303473, + 0.002348737558349967, + 0.0038846845272928476, + 0.002193805295974016, + 0.0036182014737278223, + 0.011460709385573864, + 0.009181126020848751, + 0.010875756852328777, + 0.00988574419170618, + 0.004780859220772982, + 0.00949634239077568, + 0.004790256265550852, + 0.018237367272377014, + 0.015619884245097637, + 0.013685944490134716, + 0.03176119551062584, + 0.021770626306533813, + 0.017799386754631996, + 0.030277788639068604, + 0.014431116171181202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.014210226014256477 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740925367921591 + ], + "opposite_logprob_tokens": [ + -5.006741046905518 + ], + "logprob_reward": -0.006740925367921591, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740925367921591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08990750461816788, + 0.10066941380500793, + 0.06504688411951065, + 0.060874901711940765, + 0.05910685658454895, + 0.10171262919902802, + 0.1615997552871704, + 0.028953853994607925, + 0.020345164462924004, + 0.31178298592567444 + ], + "credit_map": [ + 0.00831648614257574, + 0.007877127267420292, + 0.014443960040807724, + 0.017449988052248955, + 0.007410360500216484, + 0.011620230041444302, + 0.01379929669201374, + 0.008990053087472916, + 0.005899387411773205, + 0.010082214139401913, + 0.01001250371336937, + 0.012551477178931236, + 0.010658785700798035, + 0.0254205372184515, + 0.012834970839321613, + 0.013209536671638489, + 0.005991592071950436, + 0.004518862813711166, + 0.00730289937928319, + 0.0077136182226240635, + 0.0038407552056014538, + 0.016920890659093857, + 0.012428810819983482, + 0.006329452618956566, + 0.0076105971820652485, + 0.003600682131946087, + 0.0053185466676950455, + 0.01865149661898613, + 0.0035379386972635984, + 0.004703798331320286, + 0.013837632723152637, + 0.0036142112221568823, + 0.008807425387203693, + 0.008982470259070396, + 0.005842949263751507, + 0.007215946912765503, + 0.00845346599817276, + 0.005942064803093672, + 0.006262089125812054, + 0.007600448559969664, + 0.016073528677225113, + 0.015248438343405724, + 0.015457306988537312, + 0.012049552984535694, + 0.010985241271555424, + 0.009498837403953075, + 0.009714018553495407, + 0.01268570963293314, + 0.025219259783625603, + 0.02296561934053898, + 0.02333337627351284, + 0.009577643126249313, + 0.021163582801818848, + 0.030402513220906258, + 0.011781004257500172, + 0.01715673692524433, + 0.00471041863784194, + 0.0042626564390957355, + 0.003829608205705881, + 0.0029631906654685736, + 0.00406274339184165, + 0.003276858013123274, + 0.0030442969873547554, + 0.0028040797915309668, + 0.0013207645388320088, + 0.0030686301179230213, + 0.0024866205640137196, + 0.004413770046085119, + 0.002409921260550618, + 0.0020281511824578047, + 0.003247361397370696, + 0.0013699454721063375, + 0.033798202872276306, + 0.024769287556409836, + 0.04219780117273331, + 0.05458156019449234, + 0.03483707830309868, + 0.03729618340730667, + 0.04967030882835388, + 0.034632548689842224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.006740925367921591 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004260271787643 + ], + "opposite_logprob_tokens": [ + -2.695042610168457 + ], + "logprob_reward": -0.07004260271787643, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004260271787643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059799641370773315, + 0.09158574789762497, + 0.07112614810466766, + 0.06592658162117004, + 0.03335040062665939, + 0.03816178813576698, + 0.13337381184101105, + 0.009701641276478767, + 0.03050621598958969, + 0.46646803617477417 + ], + "credit_map": [ + 0.004999048076570034, + 0.005976190324872732, + 0.009230079129338264, + 0.013503928668797016, + 0.005019850563257933, + 0.004655853845179081, + 0.0111989201977849, + 0.005215775687247515, + 0.0045171258971095085, + 0.009142368100583553, + 0.008160112425684929, + 0.00982390996068716, + 0.009064963087439537, + 0.02597120776772499, + 0.013123825192451477, + 0.011782245710492134, + 0.008853648789227009, + 0.0054807160049676895, + 0.010499827563762665, + 0.008446432650089264, + 0.00690406933426857, + 0.012179833836853504, + 0.009413053281605244, + 0.009348565712571144, + 0.00882459431886673, + 0.003171599470078945, + 0.004737634211778641, + 0.022506210952997208, + 0.003028459381312132, + 0.0035292014945298433, + 0.017363859340548515, + 0.002765027107670903, + 0.004443368408828974, + 0.00454613147303462, + 0.007491312921047211, + 0.004436236806213856, + 0.0021068186033517122, + 0.003859621472656727, + 0.0025083343498408794, + 0.003958576358854771, + 0.006818320136517286, + 0.006560103502124548, + 0.006316679995507002, + 0.004100850783288479, + 0.003414302133023739, + 0.0030657309107482433, + 0.0028308159671723843, + 0.005054983776062727, + 0.02692369930446148, + 0.017336459830403328, + 0.020457711070775986, + 0.008961423300206661, + 0.012536305002868176, + 0.01671483926475048, + 0.010388542897999287, + 0.020054837688803673, + 0.0013599122175946832, + 0.0012874555541202426, + 0.0016148383729159832, + 0.0010393228149041533, + 0.0014011895982548594, + 0.0007492920849472284, + 0.0014010912273079157, + 0.000848540454171598, + 0.0017281947657465935, + 0.0049764481373131275, + 0.0049591814167797565, + 0.005794299766421318, + 0.004152496811002493, + 0.0025554136373102665, + 0.003949293401092291, + 0.002390891080722213, + 0.04584813490509987, + 0.0340014211833477, + 0.06200535222887993, + 0.07946628332138062, + 0.05229775980114937, + 0.0595889650285244, + 0.07726482301950455, + 0.05599532276391983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.07004260271787643 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008147019543685019 + ], + "opposite_logprob_tokens": [ + -7.125814914703369 + ], + "logprob_reward": -0.0008147019543685019, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008147019543685019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10635935515165329, + 0.05757917836308479, + 0.07331087440252304, + 0.11311279237270355, + 0.04138420149683952, + 0.07918372005224228, + 0.0854359045624733, + 0.026372792199254036, + 0.024604544043540955, + 0.3926566541194916 + ], + "credit_map": [ + 0.013723987154662609, + 0.009924228303134441, + 0.014545032754540443, + 0.014869527891278267, + 0.010342505760490894, + 0.015962105244398117, + 0.011392391286790371, + 0.01559958141297102, + 0.004652245435863733, + 0.004357782192528248, + 0.00665220757946372, + 0.008090177550911903, + 0.006660476326942444, + 0.012408713810145855, + 0.007139408029615879, + 0.00761816743761301, + 0.006166530307382345, + 0.006468841340392828, + 0.0056848651729524136, + 0.0069820028729736805, + 0.0054093399085104465, + 0.018579140305519104, + 0.019185418263077736, + 0.004834736231714487, + 0.014637802727520466, + 0.00500534800812602, + 0.00805663876235485, + 0.03732580319046974, + 0.005673830863088369, + 0.007182570639997721, + 0.0299716517329216, + 0.005259140394628048, + 0.006660577841103077, + 0.005685760173946619, + 0.005672963336110115, + 0.005954381078481674, + 0.003579366486519575, + 0.006018121726810932, + 0.0036425814032554626, + 0.004170452244579792, + 0.013289110735058784, + 0.006773274391889572, + 0.014737733639776707, + 0.012141074053943157, + 0.006374978460371494, + 0.006952633615583181, + 0.009028105065226555, + 0.009886806830763817, + 0.010554858483374119, + 0.014763327315449715, + 0.011197439394891262, + 0.009007793851196766, + 0.011223219335079193, + 0.010919177904725075, + 0.00918775424361229, + 0.008582334965467453, + 0.003457485930994153, + 0.002878678496927023, + 0.005566930398344994, + 0.0035992467310279608, + 0.002816627500578761, + 0.0019848390948027372, + 0.003604413941502571, + 0.0024645687080919743, + 0.001476980629377067, + 0.0035886650439351797, + 0.00412548566237092, + 0.0049292417243123055, + 0.0029337622690945864, + 0.0027346613351255655, + 0.003011211985722184, + 0.0018045363249257207, + 0.04404290020465851, + 0.02912559174001217, + 0.05050202086567879, + 0.06892228871583939, + 0.04543226212263107, + 0.0487387515604496, + 0.06388453394174576, + 0.042008284479379654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.0008147019543685019 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, ultra-wide shot captures a sprawling futuristic metropolis at night, its rain-slicked streets like a dark mirror reflecting the spectacle unfolding below. A fleet of enormous, towering robots slowly and gracefully winds its way through a canyon-like avenue formed by massive black glass and steel skyscrapers. Their sleek, vaguely humanoid forms are constructed from matte black alloy that absorbs ambient light. Yet, their entire frames are traced by a complex network of vibrant blue and white lights that rhythmically pulsate, illuminating their path and casting long, shifting shadows. Each massive step descends with silent, ponderous weight, creating noticeable ripples in the puddles of water on the asphalt. The damp ground perfectly reflects the robots' lighting and the ever-changing colors of the city's towering holographic advertisements, which flicker silently in the background. The scene is immersed in a cool cyberpunk color palette of deep blue, electronic magentas, and pure white, creating an atmosphere of silent, awe-inspiring grandeur and technological mystery.", + "seed": 18008, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_7/index_9727_seed_18008_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14092969857156276, + "vlm_logprob": -0.14092969857156276, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.14092969857156276, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23349891602993011 + ], + "opposite_logprob_tokens": [ + -1.6084989309310913 + ], + "logprob_reward": -0.23349891602993011, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23349891602993011, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21512216329574585, + 0.2630040645599365, + 0.0664505586028099, + 0.0664745420217514, + 0.0754900723695755, + 0.029901236295700073, + 0.08086257427930832, + 0.06165849789977074, + 0.10924817621707916, + 0.03178808093070984 + ], + "credit_map": [ + 0.0346500463783741, + 0.012843070551753044, + 0.026190685108304024, + 0.034173306077718735, + 0.018968340009450912, + 0.021104736253619194, + 0.042501457035541534, + 0.02469053864479065, + 0.030692458152770996, + 0.01659943163394928, + 0.02485775575041771, + 0.043843355029821396, + 0.02523035742342472, + 0.05262042582035065, + 0.03704988956451416, + 0.03211042657494545, + 0.010952415876090527, + 0.007544619496911764, + 0.010110268369317055, + 0.009088700637221336, + 0.0058479029685258865, + 0.006953107193112373, + 0.006594462785869837, + 0.009359086863696575, + 0.006696722470223904, + 0.005333588924258947, + 0.009537458419799805, + 0.007616833318024874, + 0.007409520912915468, + 0.0064761764369904995, + 0.010901400819420815, + 0.012502843514084816, + 0.0038563699927181005, + 0.0044631450437009335, + 0.018461789935827255, + 0.00503232516348362, + 0.003922692965716124, + 0.0034244658891111612, + 0.028416141867637634, + 0.007913151755928993, + 0.002615179866552353, + 0.004368484951555729, + 0.0034196563065052032, + 0.005851211957633495, + 0.0022608020808547735, + 0.002920478815212846, + 0.0038795890286564827, + 0.0045858388766646385, + 0.007877861149609089, + 0.007986912503838539, + 0.015873612836003304, + 0.008944609202444553, + 0.009479126892983913, + 0.006518091540783644, + 0.016219809651374817, + 0.007962558418512344, + 0.0077992891892790794, + 0.006404565181583166, + 0.0062021175399422646, + 0.007017855532467365, + 0.008189069107174873, + 0.009295876137912273, + 0.00869216863065958, + 0.008057561703026295, + 0.01065472885966301, + 0.0077395960688591, + 0.010804668068885803, + 0.014237728901207447, + 0.020085560157895088, + 0.014857756905257702, + 0.016446085646748543, + 0.014422064647078514, + 0.003518106881529093, + 0.0034152623265981674, + 0.005074086599051952, + 0.002805307973176241, + 0.0030285369139164686, + 0.005956044886261225, + 0.0031077766325324774, + 0.004882962442934513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.23349891602993011 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02984067052602768 + ], + "opposite_logprob_tokens": [ + -3.5298407077789307 + ], + "logprob_reward": -0.02984067052602768, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02984067052602768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16304980218410492, + 0.3404490053653717, + 0.04869774729013443, + 0.03624453395605087, + 0.039603523910045624, + 0.07728995382785797, + 0.07179012149572372, + 0.04757829010486603, + 0.16027382016181946, + 0.015023119747638702 + ], + "credit_map": [ + 0.030342664569616318, + 0.011216556653380394, + 0.017463980242609978, + 0.024747289717197418, + 0.01668854057788849, + 0.01639869436621666, + 0.02949254773557186, + 0.016699543222784996, + 0.04317193478345871, + 0.020732220262289047, + 0.033310454338788986, + 0.055900782346725464, + 0.031641170382499695, + 0.0688045397400856, + 0.04314139857888222, + 0.04374650865793228, + 0.007632127497345209, + 0.005550091620534658, + 0.00934111699461937, + 0.007241301238536835, + 0.0048638638108968735, + 0.004553128499537706, + 0.004554082173854113, + 0.0049620335921645164, + 0.00292331725358963, + 0.003798927878960967, + 0.005517284385859966, + 0.004193542059510946, + 0.0033573678229004145, + 0.004831255879253149, + 0.006113871932029724, + 0.005508971866220236, + 0.003154915291815996, + 0.003808298846706748, + 0.008695106022059917, + 0.0034198081120848656, + 0.0026573294308036566, + 0.003798507619649172, + 0.009572396986186504, + 0.004497162532061338, + 0.007791922427713871, + 0.01358178909868002, + 0.00883463490754366, + 0.01613958179950714, + 0.0055320714600384235, + 0.0047218408435583115, + 0.010158812627196312, + 0.01052930299192667, + 0.006147501990199089, + 0.004685133695602417, + 0.017139004543423653, + 0.008188065141439438, + 0.006137797608971596, + 0.0033915338572114706, + 0.02014017291367054, + 0.005960911978036165, + 0.00501032080501318, + 0.005288999993354082, + 0.005013741552829742, + 0.006220834795385599, + 0.00726341363042593, + 0.00691779050976038, + 0.005634346045553684, + 0.006228842306882143, + 0.015267303213477135, + 0.012098063714802265, + 0.014724524691700935, + 0.02125970833003521, + 0.030969273298978806, + 0.019268566742539406, + 0.024128058925271034, + 0.022558320313692093, + 0.0015773976920172572, + 0.0010411425027996302, + 0.0026985621079802513, + 0.0011680802563205361, + 0.001741421059705317, + 0.0023636878468096256, + 0.00201359367929399, + 0.0024192354176193476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.02984067052602768 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647494949400425 + ], + "opposite_logprob_tokens": [ + -4.758647441864014 + ], + "logprob_reward": -0.008647494949400425, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647494949400425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1251068115234375, + 0.34155023097991943, + 0.11263686418533325, + 0.08547253161668777, + 0.07034210860729218, + 0.06924960762262344, + 0.06982824206352234, + 0.02328755147755146, + 0.0364910326898098, + 0.06603507697582245 + ], + "credit_map": [ + 0.013760357163846493, + 0.007711746264249086, + 0.015656733885407448, + 0.022019416093826294, + 0.013362273573875427, + 0.010845155455172062, + 0.026791799813508987, + 0.014959324151277542, + 0.042945198714733124, + 0.02108270861208439, + 0.032092247158288956, + 0.056534543633461, + 0.028094148263335228, + 0.06481239944696426, + 0.03936150670051575, + 0.056627485901117325, + 0.015937473624944687, + 0.011232729069888592, + 0.013628638349473476, + 0.014895714819431305, + 0.018018823117017746, + 0.015493995510041714, + 0.011787397786974907, + 0.011642090044915676, + 0.008117888122797012, + 0.010262745432555676, + 0.010821839794516563, + 0.010300076566636562, + 0.007964248768985271, + 0.008973860181868076, + 0.015670999884605408, + 0.013360870070755482, + 0.006309228017926216, + 0.009143716655671597, + 0.012598550878465176, + 0.007412726059556007, + 0.006526286713778973, + 0.004882588516920805, + 0.016177508980035782, + 0.007291504647582769, + 0.005776581354439259, + 0.01020054705440998, + 0.007254469208419323, + 0.016714245080947876, + 0.008115611039102077, + 0.005637793801724911, + 0.00794235523790121, + 0.007607998792082071, + 0.008367260918021202, + 0.008527151308953762, + 0.015452231280505657, + 0.009821241721510887, + 0.00442934175953269, + 0.00354628823697567, + 0.013510595075786114, + 0.006174133159220219, + 0.0024162016343325377, + 0.0028684688732028008, + 0.002633594674989581, + 0.0029115371871739626, + 0.0036584974732249975, + 0.0030280367936939, + 0.0030184860806912184, + 0.0027527285274118185, + 0.004834338556975126, + 0.003866363549605012, + 0.006721163168549538, + 0.003964604344218969, + 0.003979092463850975, + 0.00374359218403697, + 0.004719744902104139, + 0.004662133287638426, + 0.007062425836920738, + 0.006999302189797163, + 0.007450018543750048, + 0.005653108470141888, + 0.008301167748868465, + 0.009870278649032116, + 0.00927116721868515, + 0.011427609249949455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.008647494949400425 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42904162406921387 + ], + "opposite_logprob_tokens": [ + -1.0540416240692139 + ], + "logprob_reward": -0.42904162406921387, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42904162406921387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15745463967323303, + 0.5801412463188171, + 0.04458408057689667, + 0.028105346485972404, + 0.04512444883584976, + 0.022069575265049934, + 0.02510053478181362, + 0.02874496579170227, + 0.054447583854198456, + 0.014227617532014847 + ], + "credit_map": [ + 0.023133542388677597, + 0.010095177218317986, + 0.017737558111548424, + 0.025510627776384354, + 0.016470516100525856, + 0.014784286729991436, + 0.03150755167007446, + 0.01821536011993885, + 0.07365500181913376, + 0.03655098378658295, + 0.052926547825336456, + 0.09739868342876434, + 0.04973996803164482, + 0.11910281330347061, + 0.07073941826820374, + 0.08002777397632599, + 0.0060148099437355995, + 0.003187342081218958, + 0.0052545382641255856, + 0.005861211102455854, + 0.005623217206448317, + 0.005713826976716518, + 0.006681837607175112, + 0.006247290875762701, + 0.0028303845319896936, + 0.0025780024006962776, + 0.004349566996097565, + 0.003478346625342965, + 0.0024049896746873856, + 0.0029310425743460655, + 0.004756844602525234, + 0.004776166286319494, + 0.0022992207668721676, + 0.002322794869542122, + 0.010192755609750748, + 0.0032848140690475702, + 0.002124733291566372, + 0.0025781418662518263, + 0.016907252371311188, + 0.00541473226621747, + 0.0027766346465796232, + 0.0038111673202365637, + 0.0026146729942411184, + 0.0042876615189015865, + 0.0018410475458949804, + 0.0014157344121485949, + 0.0026850728318095207, + 0.0026375818997621536, + 0.002935351338237524, + 0.001992171863093972, + 0.005010100547224283, + 0.00360319041647017, + 0.0023332247510552406, + 0.0014779226621612906, + 0.004881270229816437, + 0.002867300296202302, + 0.0030205780640244484, + 0.0030434844084084034, + 0.0029801344498991966, + 0.003404218005016446, + 0.00467407563701272, + 0.004202747717499733, + 0.0035831178538501263, + 0.00383660476654768, + 0.005519806407392025, + 0.004949551075696945, + 0.006103107705712318, + 0.00833006203174591, + 0.008421177044510841, + 0.006106276996433735, + 0.006623812019824982, + 0.008393784984946251, + 0.0014079406391829252, + 0.001059993519447744, + 0.0023048550356179476, + 0.0012042049784213305, + 0.0013438141904771328, + 0.0022999297361820936, + 0.0024452300276607275, + 0.0021616469603031874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.42904162406921387 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003619787283241749 + ], + "opposite_logprob_tokens": [ + -5.62861967086792 + ], + "logprob_reward": -0.003619787283241749, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003619787283241749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11735229939222336, + 0.1903315633535385, + 0.040152233093976974, + 0.17729327082633972, + 0.23240970075130463, + 0.04657987505197525, + 0.04226856678724289, + 0.038456935435533524, + 0.061039622873067856, + 0.05411597713828087 + ], + "credit_map": [ + 0.009397446177899837, + 0.0054910387843847275, + 0.01215416844934225, + 0.02037491276860237, + 0.012968515045940876, + 0.010019541718065739, + 0.024949902668595314, + 0.02199677750468254, + 0.023694869130849838, + 0.010320480912923813, + 0.017965691164135933, + 0.03298771753907204, + 0.016106892377138138, + 0.03787463158369064, + 0.021740105003118515, + 0.0296411644667387, + 0.005581341218203306, + 0.0031485557556152344, + 0.005172887817025185, + 0.006246908102184534, + 0.004654293414205313, + 0.006984733510762453, + 0.0032348346430808306, + 0.00512867933139205, + 0.015684641897678375, + 0.01150205172598362, + 0.021481938660144806, + 0.019320504739880562, + 0.019888928160071373, + 0.019353235140442848, + 0.03025716543197632, + 0.03980480879545212, + 0.014155598357319832, + 0.00867160689085722, + 0.042187657207250595, + 0.015709316357970238, + 0.011404384858906269, + 0.022353745996952057, + 0.08407747000455856, + 0.03384992107748985, + 0.004254926927387714, + 0.005564210005104542, + 0.005094892345368862, + 0.00722495885565877, + 0.004622611682862043, + 0.006786877289414406, + 0.004647610709071159, + 0.008383788168430328, + 0.0035532552283257246, + 0.003495020791888237, + 0.008936584927141666, + 0.004339197650551796, + 0.004528431687504053, + 0.0026414678432047367, + 0.010059517808258533, + 0.004715087357908487, + 0.004375088028609753, + 0.004439481068402529, + 0.003494545118883252, + 0.004675582982599735, + 0.005472247954457998, + 0.0058151837438344955, + 0.005100372713059187, + 0.005084431730210781, + 0.00604295264929533, + 0.00465113390237093, + 0.007650887593626976, + 0.006960063707083464, + 0.008392446674406528, + 0.009029103443026543, + 0.010822249576449394, + 0.007490784395486116, + 0.005566688720136881, + 0.0072890277951955795, + 0.009352569468319416, + 0.006357820704579353, + 0.004924141336232424, + 0.007805423811078072, + 0.0058817374520003796, + 0.006938565988093615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.003619787283241749 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, ultra-wide shot captures a sprawling futuristic metropolis at night, its rain-slicked streets like a dark mirror reflecting the spectacle unfolding below. A fleet of enormous, towering robots slowly and gracefully winds its way through a canyon-like avenue formed by massive black glass and steel skyscrapers. Their sleek, vaguely humanoid forms are constructed from matte black alloy that absorbs ambient light. Yet, their entire frames are traced by a complex network of vibrant blue and white lights that rhythmically pulsate, illuminating their path and casting long, shifting shadows. Each massive step descends with silent, ponderous weight, creating noticeable ripples in the puddles of water on the asphalt. The damp ground perfectly reflects the robots' lighting and the ever-changing colors of the city's towering holographic advertisements, which flicker silently in the background. The scene is immersed in a cool cyberpunk color palette of deep blue, electronic magentas, and pure white, creating an atmosphere of silent, awe-inspiring grandeur and technological mystery.", + "seed": 18009, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_7/index_9727_seed_18009_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09919742443598807, + "vlm_logprob": -0.09919742443598807, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.09919742443598807, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13637501001358032 + ], + "opposite_logprob_tokens": [ + -2.1363749504089355 + ], + "logprob_reward": -0.13637501001358032, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13637501001358032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2559458017349243, + 0.09366478025913239, + 0.060912951827049255, + 0.042264435440301895, + 0.03309836983680725, + 0.07194120436906815, + 0.03137374669313431, + 0.11541274935007095, + 0.21644467115402222, + 0.07894127070903778 + ], + "credit_map": [ + 0.02988041751086712, + 0.028226885944604874, + 0.03257976099848747, + 0.02563072182238102, + 0.03817868232727051, + 0.023015370592474937, + 0.04129474237561226, + 0.03713923692703247, + 0.0139577342197299, + 0.007503197528421879, + 0.013706863857805729, + 0.015853440389037132, + 0.011036904528737068, + 0.007710321340709925, + 0.011398233473300934, + 0.01249808631837368, + 0.00850999727845192, + 0.005560243967920542, + 0.008685574866831303, + 0.007864028215408325, + 0.006262382958084345, + 0.0048774899914860725, + 0.009779932908713818, + 0.00937330536544323, + 0.004378573969006538, + 0.004942202940583229, + 0.005626007914543152, + 0.006336812861263752, + 0.005058171693235636, + 0.0047639235854148865, + 0.006765452213585377, + 0.004393292590975761, + 0.005575233604758978, + 0.004061657004058361, + 0.004037972539663315, + 0.0044045038521289825, + 0.00411263620480895, + 0.0033247226383537054, + 0.0036292693112045527, + 0.003952377941459417, + 0.006784248631447554, + 0.006723887287080288, + 0.008766253478825092, + 0.007250750903040171, + 0.010147294960916042, + 0.009718777611851692, + 0.010240062139928341, + 0.012309938669204712, + 0.0031081801280379295, + 0.003521411446854472, + 0.0036056656390428543, + 0.0039361026138067245, + 0.0032889549620449543, + 0.004445887636393309, + 0.004549210425466299, + 0.004918335471302271, + 0.013212059624493122, + 0.014042280614376068, + 0.011820987798273563, + 0.014802813529968262, + 0.016460439190268517, + 0.010479659773409367, + 0.016760997474193573, + 0.017833519726991653, + 0.027754899114370346, + 0.023559194058179855, + 0.02293320558965206, + 0.0310650784522295, + 0.023017175495624542, + 0.02551194280385971, + 0.03545162454247475, + 0.02715155854821205, + 0.008615406230092049, + 0.008105101063847542, + 0.008815798908472061, + 0.011639731004834175, + 0.006893448065966368, + 0.005978000815957785, + 0.017288044095039368, + 0.011605747044086456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.13637501001358032 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026376847177743912 + ], + "opposite_logprob_tokens": [ + -3.651376962661743 + ], + "logprob_reward": -0.026376847177743912, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026376847177743912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2573784291744232, + 0.16389018297195435, + 0.08071611821651459, + 0.076120525598526, + 0.03828071430325508, + 0.04818139597773552, + 0.03975756838917732, + 0.07009739428758621, + 0.14245755970478058, + 0.0831201821565628 + ], + "credit_map": [ + 0.02914164960384369, + 0.024642491713166237, + 0.03277796879410744, + 0.02647615782916546, + 0.0407322458922863, + 0.02205551601946354, + 0.04234202951192856, + 0.03921034559607506, + 0.024124328047037125, + 0.013654755428433418, + 0.023554490879178047, + 0.023173397406935692, + 0.02119770646095276, + 0.015510335564613342, + 0.020570240914821625, + 0.022104917094111443, + 0.011614705435931683, + 0.007519413251429796, + 0.011766801588237286, + 0.012508703395724297, + 0.008772926405072212, + 0.006174600683152676, + 0.012545191682875156, + 0.009813779965043068, + 0.009474264457821846, + 0.009079618379473686, + 0.009206261485815048, + 0.010916627943515778, + 0.010151880793273449, + 0.008816159330308437, + 0.012232614681124687, + 0.006243106443434954, + 0.006168413907289505, + 0.004227038938552141, + 0.00474543496966362, + 0.005211365409195423, + 0.004522358998656273, + 0.003432197729125619, + 0.0048158299177885056, + 0.0051580760627985, + 0.0055172559805214405, + 0.00446728803217411, + 0.0058431969955563545, + 0.007174111902713776, + 0.006144342944025993, + 0.005830030422657728, + 0.005612031556665897, + 0.007593138143420219, + 0.004332817159593105, + 0.003421695437282324, + 0.004528702702373266, + 0.006226310972124338, + 0.005574143957346678, + 0.004598744213581085, + 0.005042086821049452, + 0.006033068988472223, + 0.009738292545080185, + 0.0061249504797160625, + 0.007553235627710819, + 0.011185002513229847, + 0.00834778230637312, + 0.00840919092297554, + 0.008141075260937214, + 0.010597857646644115, + 0.01838817447423935, + 0.013989104889333248, + 0.015330956317484379, + 0.01731724664568901, + 0.01779049076139927, + 0.021850431337952614, + 0.02239503152668476, + 0.015396127477288246, + 0.008836453780531883, + 0.006668880116194487, + 0.0095726503059268, + 0.010988044552505016, + 0.006975363940000534, + 0.0076904441229999065, + 0.018887685611844063, + 0.01350066065788269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.026376847177743912 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607775315642357 + ], + "opposite_logprob_tokens": [ + -4.141077518463135 + ], + "logprob_reward": -0.01607775315642357, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607775315642357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2142542451620102, + 0.07317143678665161, + 0.07764926552772522, + 0.08448485285043716, + 0.05391823872923851, + 0.026416359469294548, + 0.06536708027124405, + 0.14255335927009583, + 0.17167358100414276, + 0.09051167219877243 + ], + "credit_map": [ + 0.030231140553951263, + 0.03202245011925697, + 0.0281438697129488, + 0.019918372854590416, + 0.0271774772554636, + 0.02738027647137642, + 0.025246212258934975, + 0.024134427309036255, + 0.011569426394999027, + 0.007854973897337914, + 0.00999407283961773, + 0.010621345601975918, + 0.006764041725546122, + 0.009355698712170124, + 0.008845790289342403, + 0.008166074752807617, + 0.0167783685028553, + 0.008460509590804577, + 0.010004405863583088, + 0.0083089554682374, + 0.007443178445100784, + 0.007087088655680418, + 0.011179548688232899, + 0.008387201465666294, + 0.008671576157212257, + 0.011099540628492832, + 0.010224193334579468, + 0.012452946044504642, + 0.011768859811127186, + 0.008495477959513664, + 0.012059710919857025, + 0.009712539613246918, + 0.005207611247897148, + 0.0056981537491083145, + 0.006817723624408245, + 0.007083307020366192, + 0.00807779747992754, + 0.00757882883772254, + 0.006173116620630026, + 0.0072816903702914715, + 0.0027621572371572256, + 0.002847271040081978, + 0.003338053822517395, + 0.004156284965574741, + 0.0034549206029623747, + 0.003435177728533745, + 0.002888844581320882, + 0.0035336455330252647, + 0.008343715220689774, + 0.007716675288975239, + 0.00868870597332716, + 0.007735531311482191, + 0.007278993260115385, + 0.008491241373121738, + 0.00746499327942729, + 0.00964721292257309, + 0.015135335735976696, + 0.020639516413211823, + 0.016594674438238144, + 0.02201666310429573, + 0.015197810716927052, + 0.016063950955867767, + 0.018904849886894226, + 0.018000544980168343, + 0.02615271508693695, + 0.023882407695055008, + 0.018281441181898117, + 0.025664422661066055, + 0.014993335120379925, + 0.013445815071463585, + 0.02753680944442749, + 0.021716618910431862, + 0.007537038065493107, + 0.011649644933640957, + 0.011034153401851654, + 0.019368570297956467, + 0.007713764440268278, + 0.007705342955887318, + 0.01469170581549406, + 0.010811437852680683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.01607775315642357 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31353986263275146 + ], + "opposite_logprob_tokens": [ + -1.3135398626327515 + ], + "logprob_reward": -0.31353986263275146, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31353986263275146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31230616569519043, + 0.1004963368177414, + 0.08323270827531815, + 0.14401499927043915, + 0.0428614616394043, + 0.06441584974527359, + 0.04747377708554268, + 0.0614616833627224, + 0.08591185510158539, + 0.05782515928149223 + ], + "credit_map": [ + 0.030293459072709084, + 0.029927048832178116, + 0.037697095423936844, + 0.038556914776563644, + 0.0650634914636612, + 0.025112556293606758, + 0.05192664638161659, + 0.0337289497256279, + 0.013624204322695732, + 0.00760621065273881, + 0.0162150040268898, + 0.013858061283826828, + 0.012523441575467587, + 0.008054401725530624, + 0.013827123679220676, + 0.014787891879677773, + 0.011428344063460827, + 0.0063127027824521065, + 0.011576777324080467, + 0.013489200733602047, + 0.010526969097554684, + 0.008080553263425827, + 0.013616165146231651, + 0.008201994933187962, + 0.015907326713204384, + 0.017902961000800133, + 0.018659979104995728, + 0.01759684830904007, + 0.01975405402481556, + 0.015639523044228554, + 0.0253777913749218, + 0.013176490552723408, + 0.005344655364751816, + 0.004538093227893114, + 0.005330328829586506, + 0.005766680929809809, + 0.005702456459403038, + 0.00384505488909781, + 0.006238865200430155, + 0.0060953255742788315, + 0.007246722467243671, + 0.004173837136477232, + 0.008192724548280239, + 0.010482937097549438, + 0.008890892378985882, + 0.006781924981623888, + 0.009369476698338985, + 0.00927732978016138, + 0.0058369399048388, + 0.004579858388751745, + 0.006070766132324934, + 0.0067126271314918995, + 0.005749296396970749, + 0.005068469326943159, + 0.00634922506287694, + 0.007106591947376728, + 0.008866343647241592, + 0.006798797287046909, + 0.00635767774656415, + 0.009555558674037457, + 0.006317382212728262, + 0.0054276916198432446, + 0.007035182788968086, + 0.01110304705798626, + 0.012591691687703133, + 0.008199435658752918, + 0.009378118440508842, + 0.011285590007901192, + 0.008838498964905739, + 0.009954994544386864, + 0.015493585728108883, + 0.010169931687414646, + 0.0061498950235545635, + 0.007941788993775845, + 0.006791288964450359, + 0.009041907265782356, + 0.0041563124395906925, + 0.0050330013036727905, + 0.00965038500726223, + 0.009060575626790524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.31353986263275146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle, ultra-wide view of a futuristic metropolis at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous, towering robots with sleek, vaguely humanoid forms moving through a canyon-like avenue of black glass and steel skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the robots constructed from matte black alloy with a complex network of vibrant blue and white lights tracing their frames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the robots move slowly and gracefully, creating ripples in puddles on the rain-slicked streets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a cool cyberpunk color palette of deep blue, electronic magentas, and pure white with holographic advertisements in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036176491994410753 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.0036176491994410753, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036176491994410753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1643330156803131, + 0.09323342144489288, + 0.04741799086332321, + 0.04347853735089302, + 0.028703514486551285, + 0.03549846261739731, + 0.07777979969978333, + 0.07494503259658813, + 0.10247596353292465, + 0.3321342468261719 + ], + "credit_map": [ + 0.012390058487653732, + 0.014039257541298866, + 0.01987704448401928, + 0.025887733325362206, + 0.023825587704777718, + 0.012350999750196934, + 0.03725937753915787, + 0.018702981993556023, + 0.012999371625483036, + 0.006425325758755207, + 0.01552051305770874, + 0.011160745285451412, + 0.00900171883404255, + 0.012355877086520195, + 0.011639130301773548, + 0.014130748808383942, + 0.007957376539707184, + 0.0056282333098351955, + 0.005333102308213711, + 0.007420386653393507, + 0.005005952436476946, + 0.0048157330602407455, + 0.006674523930996656, + 0.004582684487104416, + 0.005196361802518368, + 0.005808650050312281, + 0.004969110246747732, + 0.005884015001356602, + 0.004680303391069174, + 0.00581518467515707, + 0.006551880855113268, + 0.004573035053908825, + 0.004176925402134657, + 0.0031710201874375343, + 0.003478669561445713, + 0.0032852899748831987, + 0.004467157647013664, + 0.003538321703672409, + 0.003114143619313836, + 0.0034719889517873526, + 0.0039538247510790825, + 0.0029121404513716698, + 0.0045638009905815125, + 0.004144934006035328, + 0.003647653153166175, + 0.005003894679248333, + 0.004656309261918068, + 0.006615910213440657, + 0.009797937236726284, + 0.009606817737221718, + 0.010262756608426571, + 0.008486492559313774, + 0.008244737051427364, + 0.010272002778947353, + 0.009557945653796196, + 0.011551115661859512, + 0.007410100195556879, + 0.008806965313851833, + 0.008390972390770912, + 0.009802508167922497, + 0.010047463700175285, + 0.009951240383088589, + 0.007992264814674854, + 0.012543519027531147, + 0.014801757410168648, + 0.009335080161690712, + 0.01073923148214817, + 0.011166371405124664, + 0.010479766875505447, + 0.010824540629982948, + 0.021846266463398933, + 0.013282953761518002, + 0.03315633162856102, + 0.023541022092103958, + 0.035548366606235504, + 0.04339553415775299, + 0.030601467937231064, + 0.0476948581635952, + 0.071889229118824, + 0.04630744829773903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0036176491994410753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_8/index_tensor([5029, 5029])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_8/index_tensor([5029, 5029])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5ed758f059546eb5376dc0def9deb14bb43bc0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_8/index_tensor([5029, 5029])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium-long shot, richly hand-painted in a watercolor style, captures a small, iron-hulled, rifle-powered boat sailing across the vast, tranquil Weishan Lake from a slightly sideways, horizontal angle. The boat's hull is a weathered, dark gray-green, speckled with rust along its edges and seams, reflecting a dull metallic sheen under the soft, diffuse light. With a long, slightly hoarse whistle, the boat begins to steadily decelerate. The V-shaped white waves previously stirred up by the bow gradually soften, transforming into delicate ripples that spread outward. At this point, the entire frame is accompanied by a slight flicker, like that of old film, as the camera begins a slow and precise shift in focus: in the foreground, the crystal-clear ripples on the water gradually lose focus, becoming blurred and dreamy, while the camera's focus steadily shifts backward, finally settling sharply on an old flag at the stern, gently fluttering in the breeze. The bright red color of the flag has faded, revealing a dull ochre color. The edges are worn and tattered, with fine tassel-like lines. Against the backdrop of the gray-blue lake and sky, it creates a tranquil, nostalgic and slightly melancholic poetic atmosphere, like a moving ink painting carrying the memories of the years.", + "seed": 1739, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_8/index_5029_seed_1739_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.21220025084912778, + "vlm_logprob": -0.21220025084912778, + "vlm_yes_no_margin": 2.125, + "vlm_token_credit": -0.21220025084912778, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.7306922674179077 + ], + "opposite_logprob_tokens": [ + -0.8556922674179077 + ], + "logprob_reward": -0.7306922674179077, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.7306922674179077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025063075125217438, + 0.014480473473668098, + 0.008655035868287086, + 0.01903526671230793, + 0.05342191830277443, + 0.058559104800224304, + 0.6903622150421143, + 0.03077959455549717, + 0.08618007600307465, + 0.013463239185512066 + ], + "credit_map": [ + 0.0015728181460872293, + 0.003044670447707176, + 0.0019132777815684676, + 0.002743677468970418, + 0.003418105421587825, + 0.005167994182556868, + 0.003357109148055315, + 0.0038454218301922083, + 0.0011146035976707935, + 0.001901439973153174, + 0.0010997977806255221, + 0.0011892660986632109, + 0.0019213490886613727, + 0.0024889507330954075, + 0.0015790503239259124, + 0.003186014713719487, + 0.0007040465716272593, + 0.0012286100536584854, + 0.0008920911932364106, + 0.0016739582642912865, + 0.0011097310343757272, + 0.0012003572192043066, + 0.000931123155169189, + 0.0009151179692707956, + 0.0019426755607128143, + 0.0020592832006514072, + 0.0028044614009559155, + 0.0023924887645989656, + 0.0037359546404331923, + 0.0034383712336421013, + 0.0009778152452781796, + 0.001684214686974883, + 0.0028764954768121243, + 0.0027654487639665604, + 0.002694558585062623, + 0.0023320692125707865, + 0.018399668857455254, + 0.017663469538092613, + 0.0032471618615090847, + 0.0034430462401360273, + 0.0012406714959070086, + 0.0018022399162873626, + 0.001959514804184437, + 0.0012809309409931302, + 0.023495377972722054, + 0.02436981536448002, + 0.0014519165270030499, + 0.002958641154691577, + 0.02157786302268505, + 0.022988222539424896, + 0.018972307443618774, + 0.01312373485416174, + 0.29958197474479675, + 0.2724947929382324, + 0.011987969279289246, + 0.029635360464453697, + 0.0028919780161231756, + 0.0034857699647545815, + 0.004608331713825464, + 0.004435548093169928, + 0.003889671294018626, + 0.003472016891464591, + 0.004873537924140692, + 0.00312274182215333, + 0.007399530149996281, + 0.011317646130919456, + 0.012058264575898647, + 0.013856700621545315, + 0.010199037380516529, + 0.008732838556170464, + 0.012739866971969604, + 0.009876188822090626, + 0.0016923410585150123, + 0.0025616041384637356, + 0.002008823910728097, + 0.0014685398200526834, + 0.0015744162956252694, + 0.0011641699820756912, + 0.001653673010878265, + 0.001339671784080565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.7306922674179077 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029849348589777946 + ], + "opposite_logprob_tokens": [ + -3.5298492908477783 + ], + "logprob_reward": -0.029849348589777946, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029849348589777946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07244573533535004, + 0.027312112972140312, + 0.0711878091096878, + 0.06465352326631546, + 0.036425963044166565, + 0.06371699273586273, + 0.4933440685272217, + 0.04026254639029503, + 0.0825342908501625, + 0.048116911202669144 + ], + "credit_map": [ + 0.010698719881474972, + 0.005712227430194616, + 0.008798218332231045, + 0.0040337457321584225, + 0.011900767683982849, + 0.01623624935746193, + 0.00506659084931016, + 0.009999219328165054, + 0.0020691491663455963, + 0.0038102215621620417, + 0.002909785369411111, + 0.0026339052710682154, + 0.004445546772330999, + 0.004523951094597578, + 0.0027721934020519257, + 0.004147360101342201, + 0.006465900223702192, + 0.008352573961019516, + 0.007169801741838455, + 0.012161820195615292, + 0.010578341782093048, + 0.012800493277609348, + 0.006146952509880066, + 0.0075119296088814735, + 0.0037899280432611704, + 0.010850724764168262, + 0.0057558100670576096, + 0.011215316131711006, + 0.011352360248565674, + 0.010671894997358322, + 0.00569496862590313, + 0.005322521086782217, + 0.0043389149941504, + 0.0035325963981449604, + 0.003810026915743947, + 0.0022448371164500713, + 0.007795420940965414, + 0.00716685876250267, + 0.0030654463917016983, + 0.004471858963370323, + 0.006576548796147108, + 0.004361560568213463, + 0.005485839676111937, + 0.0024232501164078712, + 0.011347443796694279, + 0.021749190986156464, + 0.0031210838351398706, + 0.008652076125144958, + 0.015371919609606266, + 0.01721850223839283, + 0.013789198361337185, + 0.009736831299960613, + 0.21370664238929749, + 0.1908111274242401, + 0.009011701680719852, + 0.02369816228747368, + 0.004015600308775902, + 0.0036472335923463106, + 0.009225117042660713, + 0.004297000356018543, + 0.00377648021094501, + 0.0054238406009972095, + 0.004388430621474981, + 0.005488845985382795, + 0.00759885273873806, + 0.009892062284052372, + 0.01031820010393858, + 0.01395795214921236, + 0.008274012245237827, + 0.010915497317910194, + 0.012693344615399837, + 0.008884367533028126, + 0.008686792105436325, + 0.004490608815103769, + 0.008997977711260319, + 0.004408219829201698, + 0.0051568918861448765, + 0.005254465155303478, + 0.005865860730409622, + 0.005256093107163906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.029849348589777946 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700526013970375 + ], + "opposite_logprob_tokens": [ + -2.6950526237487793 + ], + "logprob_reward": -0.0700526013970375, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700526013970375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03175840526819229, + 0.03894579038023949, + 0.018915895372629166, + 0.05147736892104149, + 0.08386830240488052, + 0.07686176896095276, + 0.575981616973877, + 0.03435494378209114, + 0.06184626370668411, + 0.02598966658115387 + ], + "credit_map": [ + 0.0033206066582351923, + 0.00284007890149951, + 0.002323924796655774, + 0.0034180290531367064, + 0.004176093731075525, + 0.004537191241979599, + 0.003970551770180464, + 0.007171927951276302, + 0.002910385839641094, + 0.0044442168436944485, + 0.003791557624936104, + 0.004549538251012564, + 0.004818728659301996, + 0.005224872846156359, + 0.00448389258235693, + 0.008722593076527119, + 0.0014815477188676596, + 0.001988558564335108, + 0.0014265720965340734, + 0.002180254552513361, + 0.002755418885499239, + 0.0033962451852858067, + 0.0015842434950172901, + 0.004103055223822594, + 0.007669633720070124, + 0.005999199114739895, + 0.007926837541162968, + 0.005312767345458269, + 0.009557447396218777, + 0.007227029651403427, + 0.002710767323151231, + 0.005073685199022293, + 0.006287600379437208, + 0.006053353659808636, + 0.00885030534118414, + 0.0031466686632484198, + 0.024001194164156914, + 0.018682703375816345, + 0.00492995698004961, + 0.011916516348719597, + 0.006104957312345505, + 0.004239603411406279, + 0.006349090486764908, + 0.002943896222859621, + 0.02390250563621521, + 0.024172507226467133, + 0.0034420101437717676, + 0.005707195028662682, + 0.014133592136204243, + 0.019953837618231773, + 0.011269671842455864, + 0.011632506735622883, + 0.25404804944992065, + 0.23397862911224365, + 0.009872078895568848, + 0.021093206480145454, + 0.0028689829632639885, + 0.0030523897148668766, + 0.004808107390999794, + 0.0040093655698001385, + 0.00371445226483047, + 0.004930105526000261, + 0.0077934167347848415, + 0.0031781219877302647, + 0.006547162309288979, + 0.011571717448532581, + 0.005826467648148537, + 0.009658142924308777, + 0.007537372410297394, + 0.006584195885807276, + 0.007007361855357885, + 0.007113841827958822, + 0.003540279809385538, + 0.002016098704189062, + 0.0035507914144545794, + 0.003548993030562997, + 0.0034282226115465164, + 0.0034013984259217978, + 0.0038778295274823904, + 0.0026260516606271267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.0700526013970375 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002126425504684 + ], + "opposite_logprob_tokens": [ + -2.695021152496338 + ], + "logprob_reward": -0.07002126425504684, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002126425504684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018444711342453957, + 0.008125336840748787, + 0.006975999567657709, + 0.00783531740307808, + 0.03463369235396385, + 0.01943574845790863, + 0.7540274262428284, + 0.020811980590224266, + 0.08776935189962387, + 0.04194045066833496 + ], + "credit_map": [ + 0.0015360278775915504, + 0.0016108752461150289, + 0.0013212559279054403, + 0.0011194174876436591, + 0.003201432293280959, + 0.005329848267138004, + 0.001642674091272056, + 0.0026831794530153275, + 0.0005960325943306088, + 0.0008493581553921103, + 0.000589229166507721, + 0.000705529295373708, + 0.0016524451784789562, + 0.0017514316132292151, + 0.0007958344067446887, + 0.0011854757321998477, + 0.0005695539293810725, + 0.0007505983812734485, + 0.0008018831722438335, + 0.0011907706502825022, + 0.0008335140882991254, + 0.0011474566999822855, + 0.0006548601668328047, + 0.0010273625375702977, + 0.0010239535477012396, + 0.0009519000886939466, + 0.0012943309266120195, + 0.0005979553097859025, + 0.0012629906414076686, + 0.0011874680640175939, + 0.000556361279450357, + 0.0009603574289940298, + 0.0031562531366944313, + 0.0022449358366429806, + 0.003675921820104122, + 0.0011004105908796191, + 0.011650859378278255, + 0.008740573190152645, + 0.00183052197098732, + 0.002234214683994651, + 0.0021093229297548532, + 0.0018306550337001681, + 0.0022004684433341026, + 0.0011343240039423108, + 0.005325519014149904, + 0.002982932608574629, + 0.0012741463724523783, + 0.0025783784221857786, + 0.022341249510645866, + 0.02597571723163128, + 0.01959039270877838, + 0.014911018311977386, + 0.32862746715545654, + 0.29640820622444153, + 0.01348967757076025, + 0.032683659344911575, + 0.002277121413499117, + 0.002153697656467557, + 0.004840082488954067, + 0.0022963806986808777, + 0.00279467785730958, + 0.0025652062613517046, + 0.0020373468287289143, + 0.0018474665703251958, + 0.008045557886362076, + 0.011230860836803913, + 0.011601844802498817, + 0.015038883313536644, + 0.009668296203017235, + 0.010998246259987354, + 0.012972711585462093, + 0.008212937973439693, + 0.005984522867947817, + 0.003939422778785229, + 0.005918647162616253, + 0.0041395812295377254, + 0.006545545998960733, + 0.00565072288736701, + 0.005826916079968214, + 0.003935088403522968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.07002126425504684 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038577258586884 + ], + "opposite_logprob_tokens": [ + -1.9103857278823853 + ], + "logprob_reward": -0.16038577258586884, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038577258586884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.016583789139986038, + 0.007281166035681963, + 0.011602498590946198, + 0.010039038956165314, + 0.06341839581727982, + 0.04521379992365837, + 0.7894957661628723, + 0.013578358106315136, + 0.027230851352214813, + 0.015556289814412594 + ], + "credit_map": [ + 0.0015251010190695524, + 0.0019546281546354294, + 0.0019946256652474403, + 0.0015756692737340927, + 0.0020114269573241472, + 0.003204540815204382, + 0.0016865251818671823, + 0.0026312731206417084, + 0.000539685133844614, + 0.0008420634549111128, + 0.0005009964224882424, + 0.0004941414226777852, + 0.0015215096063911915, + 0.001998486462980509, + 0.0004972456954419613, + 0.0008870386518537998, + 0.0011876055505126715, + 0.0013514576712623239, + 0.001408170792274177, + 0.0019641430117189884, + 0.0016165835550054908, + 0.0016390662640333176, + 0.0009607266983948648, + 0.0014747455716133118, + 0.0009874440729618073, + 0.0008215370471589267, + 0.0009349509491585195, + 0.001087502227164805, + 0.0018000975251197815, + 0.002188745653256774, + 0.0008043997804634273, + 0.0014143632724881172, + 0.0025460098404437304, + 0.004126229789108038, + 0.002765486016869545, + 0.0021123625338077545, + 0.02425272949039936, + 0.021081238985061646, + 0.0035341575276106596, + 0.0030001853592693806, + 0.002704286715015769, + 0.0032785425428301096, + 0.003369225887581706, + 0.0016488245455548167, + 0.01638486422598362, + 0.011028644628822803, + 0.002396228024736047, + 0.004403182305395603, + 0.023166503757238388, + 0.02784317545592785, + 0.01952839270234108, + 0.015026792883872986, + 0.3445105254650116, + 0.30897629261016846, + 0.014918049797415733, + 0.035526104271411896, + 0.0011542144929990172, + 0.0015158557798713446, + 0.0021764636039733887, + 0.001640924601815641, + 0.001755452947691083, + 0.0018008348997682333, + 0.0017148444894701242, + 0.0018197690369561315, + 0.0031398823484778404, + 0.00523066520690918, + 0.003976570907980204, + 0.0038309062365442514, + 0.003105943789705634, + 0.0025047562085092068, + 0.0028523497749119997, + 0.002589777810499072, + 0.0020975410006940365, + 0.0014138722326606512, + 0.002122619654983282, + 0.0023516619112342596, + 0.001813922543078661, + 0.0023432779125869274, + 0.0015367615269497037, + 0.0018766341963782907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.16038577258586884 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium-long shot, richly hand-painted in a watercolor style, captures a small, iron-hulled, rifle-powered boat sailing across the vast, tranquil Weishan Lake from a slightly sideways, horizontal angle. The boat's hull is a weathered, dark gray-green, speckled with rust along its edges and seams, reflecting a dull metallic sheen under the soft, diffuse light. With a long, slightly hoarse whistle, the boat begins to steadily decelerate. The V-shaped white waves previously stirred up by the bow gradually soften, transforming into delicate ripples that spread outward. At this point, the entire frame is accompanied by a slight flicker, like that of old film, as the camera begins a slow and precise shift in focus: in the foreground, the crystal-clear ripples on the water gradually lose focus, becoming blurred and dreamy, while the camera's focus steadily shifts backward, finally settling sharply on an old flag at the stern, gently fluttering in the breeze. The bright red color of the flag has faded, revealing a dull ochre color. The edges are worn and tattered, with fine tassel-like lines. Against the backdrop of the gray-blue lake and sky, it creates a tranquil, nostalgic and slightly melancholic poetic atmosphere, like a moving ink painting carrying the memories of the years.", + "seed": 1740, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_8/index_5029_seed_1740_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.18079816922545433, + "vlm_logprob": -0.18079816922545433, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.18079816922545433, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6058980226516724 + ], + "opposite_logprob_tokens": [ + -0.9808980226516724 + ], + "logprob_reward": -0.6058980226516724, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.6058980226516724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11318454146385193, + 0.053343866020441055, + 0.27091288566589355, + 0.03765105456113815, + 0.12711980938911438, + 0.05567135661840439, + 0.10832348465919495, + 0.11463318765163422, + 0.05835384130477905, + 0.06080595403909683 + ], + "credit_map": [ + 0.011479458771646023, + 0.008307943120598793, + 0.018890783190727234, + 0.009411961771547794, + 0.020212896168231964, + 0.02666103094816208, + 0.00722261518239975, + 0.010997854173183441, + 0.004551639314740896, + 0.009470131248235703, + 0.00973221193999052, + 0.006229359656572342, + 0.007848678156733513, + 0.007547196000814438, + 0.0028003077022731304, + 0.005164338741451502, + 0.031653132289648056, + 0.04978831857442856, + 0.03592522814869881, + 0.0382811464369297, + 0.04363314062356949, + 0.028725555166602135, + 0.02176531031727791, + 0.021141046658158302, + 0.004446915816515684, + 0.00608887430280447, + 0.004053458105772734, + 0.003830666420981288, + 0.005560309626162052, + 0.004391262773424387, + 0.003617737675085664, + 0.0056618270464241505, + 0.02056768536567688, + 0.011072598397731781, + 0.020911261439323425, + 0.012541158124804497, + 0.01815454475581646, + 0.012958620674908161, + 0.013105529360473156, + 0.01780841313302517, + 0.008083341643214226, + 0.005253813229501247, + 0.01068660244345665, + 0.0057959407567977905, + 0.006167291197925806, + 0.004385645966976881, + 0.008259669877588749, + 0.007039050105959177, + 0.01636442355811596, + 0.01598132774233818, + 0.016801349818706512, + 0.014261044561862946, + 0.012016243301331997, + 0.009318547323346138, + 0.014542073011398315, + 0.009038480930030346, + 0.028886185958981514, + 0.008995371870696545, + 0.029974564909934998, + 0.015506100840866566, + 0.004802674986422062, + 0.007539140060544014, + 0.008202414959669113, + 0.01072673499584198, + 0.007668909151107073, + 0.0056502725929021835, + 0.008198274299502373, + 0.013533143326640129, + 0.005347855854779482, + 0.006380131002515554, + 0.004410095047205687, + 0.00716516375541687, + 0.006569023709744215, + 0.00526061886921525, + 0.00783171784132719, + 0.010042046196758747, + 0.0059140934608876705, + 0.012939603067934513, + 0.005129941273480654, + 0.007118907757103443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.6058980226516724 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020624931901693344 + ], + "opposite_logprob_tokens": [ + -3.895624876022339 + ], + "logprob_reward": -0.020624931901693344, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020624931901693344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16321831941604614, + 0.07366294413805008, + 0.15799745917320251, + 0.04038939252495766, + 0.12333188205957413, + 0.07235553115606308, + 0.04283884912729263, + 0.07898837327957153, + 0.18138891458511353, + 0.0658283531665802 + ], + "credit_map": [ + 0.013460761867463589, + 0.015045078471302986, + 0.028141150251030922, + 0.015601571649312973, + 0.02573215588927269, + 0.03689608722925186, + 0.010443289764225483, + 0.017898214980959892, + 0.008648859336972237, + 0.012386373244225979, + 0.009022686630487442, + 0.010047266259789467, + 0.010217848233878613, + 0.011444192379713058, + 0.004910232033580542, + 0.0069854892790317535, + 0.017612755298614502, + 0.02482166327536106, + 0.018720263615250587, + 0.017100801691412926, + 0.024378672242164612, + 0.03178586810827255, + 0.008023365400731564, + 0.015554085373878479, + 0.004424003418534994, + 0.004703726153820753, + 0.0045073130168020725, + 0.005462498404085636, + 0.006358199752867222, + 0.005836072377860546, + 0.003324029967188835, + 0.0057735503651201725, + 0.019825149327516556, + 0.011476981453597546, + 0.029790444299578667, + 0.008891640231013298, + 0.018947390839457512, + 0.012531876564025879, + 0.011415816843509674, + 0.010452574118971825, + 0.00900941900908947, + 0.008230417035520077, + 0.008900966495275497, + 0.005529713351279497, + 0.011166581884026527, + 0.00820181518793106, + 0.014412564225494862, + 0.006904056295752525, + 0.007730966899544001, + 0.006954598240554333, + 0.005979202222079039, + 0.0056827436201274395, + 0.004155613947659731, + 0.0039051000494509935, + 0.00426218006759882, + 0.004168443847447634, + 0.011764169670641422, + 0.012905115261673927, + 0.01044559571892023, + 0.009255597367882729, + 0.00669005885720253, + 0.011101381853222847, + 0.007844684645533562, + 0.00898177083581686, + 0.021807152777910233, + 0.0192780252546072, + 0.029709501191973686, + 0.04097241908311844, + 0.01790757104754448, + 0.01957549899816513, + 0.01570766791701317, + 0.016431087628006935, + 0.00705852871760726, + 0.0064022294245660305, + 0.00873532984405756, + 0.00850733183324337, + 0.008595134131610394, + 0.013550509698688984, + 0.005178103689104319, + 0.007801190949976444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.020624931901693344 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900860905647278 + ], + "opposite_logprob_tokens": [ + -2.5790085792541504 + ], + "logprob_reward": -0.07900860905647278, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900860905647278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0720270425081253, + 0.028736909851431847, + 0.04638310521841049, + 0.04859457165002823, + 0.1615784913301468, + 0.23859287798404694, + 0.12043564021587372, + 0.12084273993968964, + 0.1170717403292656, + 0.04573685675859451 + ], + "credit_map": [ + 0.0049280086532235146, + 0.006566179450601339, + 0.012281988747417927, + 0.005315608344972134, + 0.013524181209504604, + 0.014035318978130817, + 0.005131862126290798, + 0.010243899188935757, + 0.002632428891956806, + 0.002795338397845626, + 0.006425542291253805, + 0.0027418266981840134, + 0.0041146245785057545, + 0.004288732539862394, + 0.0018330860184505582, + 0.0039053333457559347, + 0.006454938091337681, + 0.004727539140731096, + 0.006769008468836546, + 0.0032704449258744717, + 0.006822654511779547, + 0.012020126916468143, + 0.002215505577623844, + 0.004102888982743025, + 0.0055673508904874325, + 0.0057659088633954525, + 0.009320721961557865, + 0.005283593665808439, + 0.007961108349263668, + 0.006732413079589605, + 0.002606765367090702, + 0.0053567118011415005, + 0.02766856551170349, + 0.014741643331944942, + 0.03406250476837158, + 0.009261920116841793, + 0.030010327696800232, + 0.017464499920606613, + 0.012717305682599545, + 0.015651730820536613, + 0.03239622712135315, + 0.020277509465813637, + 0.040004912763834, + 0.016761429607868195, + 0.03807694837450981, + 0.025378013029694557, + 0.03978857025504112, + 0.02590927481651306, + 0.023343946784734726, + 0.016250619664788246, + 0.021213781088590622, + 0.016384338960051537, + 0.010165253654122353, + 0.010105212219059467, + 0.012759649194777012, + 0.010212833993136883, + 0.01849750243127346, + 0.013355337083339691, + 0.017021022737026215, + 0.014842060394585133, + 0.010414757765829563, + 0.016438234597444534, + 0.013577567413449287, + 0.016696253791451454, + 0.013760119676589966, + 0.012982760556042194, + 0.017840063199400902, + 0.021235836669802666, + 0.01330222561955452, + 0.014787506312131882, + 0.010605270974338055, + 0.012557954527437687, + 0.0044877477921545506, + 0.0036837717052549124, + 0.006023441441357136, + 0.006072009447962046, + 0.0064548603259027, + 0.009575607255101204, + 0.004349559545516968, + 0.005089858081191778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.07900860905647278 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809443116188049 + ], + "opposite_logprob_tokens": [ + -3.2880945205688477 + ], + "logprob_reward": -0.03809443116188049, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809443116188049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05301165580749512, + 0.0325656421482563, + 0.07898815721273422, + 0.05363881215453148, + 0.1628238409757614, + 0.09841813147068024, + 0.06708298623561859, + 0.1827595978975296, + 0.17469052970409393, + 0.09602072834968567 + ], + "credit_map": [ + 0.004319322761148214, + 0.008930554613471031, + 0.006582866422832012, + 0.005891802720725536, + 0.007896192371845245, + 0.010225359350442886, + 0.0034417081624269485, + 0.005723841954022646, + 0.003367837518453598, + 0.004190394654870033, + 0.0064346385188400745, + 0.003513944800943136, + 0.004462243523448706, + 0.0054900809191167355, + 0.0019923155196011066, + 0.003114179940894246, + 0.007934220135211945, + 0.010514141991734505, + 0.008617299608886242, + 0.007611242588609457, + 0.011977804824709892, + 0.019238336011767387, + 0.004553080536425114, + 0.008542023599147797, + 0.005789847578853369, + 0.0060312943533062935, + 0.007658993825316429, + 0.00556325726211071, + 0.009056679904460907, + 0.009946219623088837, + 0.003193424316123128, + 0.006399091333150864, + 0.026212625205516815, + 0.019050702452659607, + 0.037088826298713684, + 0.013369748368859291, + 0.02379886619746685, + 0.015687955543398857, + 0.01418609730899334, + 0.013429012149572372, + 0.01242771651595831, + 0.00730161415413022, + 0.01650485023856163, + 0.007732298690825701, + 0.015140322037041187, + 0.012728346511721611, + 0.014476324431598186, + 0.01210664864629507, + 0.012481522746384144, + 0.0075803836807608604, + 0.014260094612836838, + 0.008480466902256012, + 0.005127782467752695, + 0.0055528064258396626, + 0.0067947739735245705, + 0.006805155426263809, + 0.029413923621177673, + 0.02209349535405636, + 0.03367242217063904, + 0.030848979949951172, + 0.011385485529899597, + 0.01803339086472988, + 0.016621071845293045, + 0.020690811797976494, + 0.02053551748394966, + 0.014769745990633965, + 0.026568040251731873, + 0.03524935990571976, + 0.016375737264752388, + 0.020744463428854942, + 0.015675220638513565, + 0.02477242983877659, + 0.009958060458302498, + 0.008284507319331169, + 0.011566298082470894, + 0.016279758885502815, + 0.013413773849606514, + 0.017058560624718666, + 0.009049017913639545, + 0.010410744696855545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.03809443116188049 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036485135555267 + ], + "opposite_logprob_tokens": [ + -1.9103648662567139 + ], + "logprob_reward": -0.16036485135555267, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036485135555267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06834731251001358, + 0.09344487637281418, + 0.061791859567165375, + 0.03435320779681206, + 0.09275482594966888, + 0.05786483362317085, + 0.18785324692726135, + 0.19510869681835175, + 0.16995114088058472, + 0.03852998837828636 + ], + "credit_map": [ + 0.00733434222638607, + 0.009164268150925636, + 0.009514986537396908, + 0.00790596567094326, + 0.00870796013623476, + 0.01075046043843031, + 0.005718030501157045, + 0.009251304902136326, + 0.008134221658110619, + 0.00934300571680069, + 0.02108205109834671, + 0.00852259248495102, + 0.014032754115760326, + 0.01689615473151207, + 0.004626780282706022, + 0.010807326063513756, + 0.00935481209307909, + 0.006163439247757196, + 0.006841063033789396, + 0.004869531374424696, + 0.011030048131942749, + 0.01444205641746521, + 0.0028500554617494345, + 0.00624085683375597, + 0.0037222467362880707, + 0.002858550753444433, + 0.004402895923703909, + 0.003526050364598632, + 0.0052024987526237965, + 0.006764065939933062, + 0.002914656652137637, + 0.004962243139743805, + 0.011929173022508621, + 0.012499316595494747, + 0.01394696906208992, + 0.009353981353342533, + 0.01563122496008873, + 0.00832659937441349, + 0.009320281445980072, + 0.011747284792363644, + 0.01035201083868742, + 0.005630368832498789, + 0.013730043545365334, + 0.004234760068356991, + 0.0057960073463618755, + 0.0037368927150964737, + 0.0075440374203026295, + 0.006840722169727087, + 0.03310443460941315, + 0.03209470584988594, + 0.03809593245387077, + 0.019726576283574104, + 0.01344921626150608, + 0.020039334893226624, + 0.01709456369280815, + 0.014248484745621681, + 0.029400696977972984, + 0.031125221401453018, + 0.023033825680613518, + 0.02545313350856304, + 0.01695341430604458, + 0.02548893168568611, + 0.02006474696099758, + 0.02358873561024666, + 0.022726166993379593, + 0.022096853703260422, + 0.026026766747236252, + 0.03922843933105469, + 0.016501029953360558, + 0.018444238230586052, + 0.013378480449318886, + 0.01154918223619461, + 0.00499583687633276, + 0.0030155801214277744, + 0.005797921679913998, + 0.004938005469739437, + 0.0035432095173746347, + 0.007509400136768818, + 0.0036617412697523832, + 0.005068297032266855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.16036485135555267 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium-long shot, richly hand-painted in a watercolor style, captures a small, iron-hulled, rifle-powered boat sailing across the vast, tranquil Weishan Lake from a slightly sideways, horizontal angle. The boat's hull is a weathered, dark gray-green, speckled with rust along its edges and seams, reflecting a dull metallic sheen under the soft, diffuse light. With a long, slightly hoarse whistle, the boat begins to steadily decelerate. The V-shaped white waves previously stirred up by the bow gradually soften, transforming into delicate ripples that spread outward. At this point, the entire frame is accompanied by a slight flicker, like that of old film, as the camera begins a slow and precise shift in focus: in the foreground, the crystal-clear ripples on the water gradually lose focus, becoming blurred and dreamy, while the camera's focus steadily shifts backward, finally settling sharply on an old flag at the stern, gently fluttering in the breeze. The bright red color of the flag has faded, revealing a dull ochre color. The edges are worn and tattered, with fine tassel-like lines. Against the backdrop of the gray-blue lake and sky, it creates a tranquil, nostalgic and slightly melancholic poetic atmosphere, like a moving ink painting carrying the memories of the years.", + "seed": 1741, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_8/index_5029_seed_1741_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.15548057947307825, + "vlm_logprob": -0.15548057947307825, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.15548057947307825, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5385305881500244 + ], + "opposite_logprob_tokens": [ + -1.0385305881500244 + ], + "logprob_reward": -0.5385305881500244, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.5385305881500244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13260234892368317, + 0.12451925128698349, + 0.047307077795267105, + 0.16790616512298584, + 0.052253127098083496, + 0.23331685364246368, + 0.0739080011844635, + 0.06062590330839157, + 0.059562839567661285, + 0.04799838364124298 + ], + "credit_map": [ + 0.03376028314232826, + 0.011214527301490307, + 0.026593325659632683, + 0.009720438160002232, + 0.015222937799990177, + 0.016489749774336815, + 0.008244303055107594, + 0.01135678868740797, + 0.012623042799532413, + 0.016907138749957085, + 0.015248321928083897, + 0.009145259857177734, + 0.03044460527598858, + 0.013317352160811424, + 0.010471435263752937, + 0.016362100839614868, + 0.01098841056227684, + 0.005161290522664785, + 0.0061151813715696335, + 0.004220115020871162, + 0.005589085631072521, + 0.0055678426288068295, + 0.003767366986721754, + 0.005897785071283579, + 0.018268469721078873, + 0.023498063907027245, + 0.023941416293382645, + 0.015382934361696243, + 0.024540124461054802, + 0.020910227671265602, + 0.02105928771197796, + 0.02030564285814762, + 0.006645692512392998, + 0.005606465507298708, + 0.0076427278108894825, + 0.006234149914234877, + 0.004016454331576824, + 0.005777978338301182, + 0.004891061689704657, + 0.011438596062362194, + 0.034978847950696945, + 0.022105779498815536, + 0.07295981794595718, + 0.06438474357128143, + 0.00666928943246603, + 0.011751396581530571, + 0.01085616648197174, + 0.009610816836357117, + 0.006749316584318876, + 0.010511746630072594, + 0.009000059217214584, + 0.013489299453794956, + 0.006746027152985334, + 0.006838515866547823, + 0.013184353709220886, + 0.007388681638985872, + 0.007243443746119738, + 0.007033395115286112, + 0.007517421618103981, + 0.005227795802056789, + 0.0063919322565197945, + 0.009739878587424755, + 0.007592237554490566, + 0.009879800491034985, + 0.009807012043893337, + 0.011026926338672638, + 0.007039094343781471, + 0.006957693491131067, + 0.008516511879861355, + 0.0061828866600990295, + 0.005201792810112238, + 0.004830924328416586, + 0.004122113808989525, + 0.006436699070036411, + 0.007164057809859514, + 0.010061879642307758, + 0.004208824131637812, + 0.006182813551276922, + 0.005794994067400694, + 0.004027002025395632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.5385305881500244 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372841328382492 + ], + "opposite_logprob_tokens": [ + -3.4087283611297607 + ], + "logprob_reward": -0.03372841328382492, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372841328382492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20387054979801178, + 0.1804048866033554, + 0.09708798676729202, + 0.10582081973552704, + 0.02123568207025528, + 0.2063860297203064, + 0.06813956052064896, + 0.01590222865343094, + 0.05936351418495178, + 0.04178876802325249 + ], + "credit_map": [ + 0.037037644535303116, + 0.02639714814722538, + 0.02755313366651535, + 0.021080801263451576, + 0.03264463320374489, + 0.02221674472093582, + 0.013148204423487186, + 0.02379222959280014, + 0.017751362174749374, + 0.02168934978544712, + 0.02620135433971882, + 0.024728862568736076, + 0.02307148464024067, + 0.021282698959112167, + 0.0201368760317564, + 0.02554289624094963, + 0.01553952693939209, + 0.01123722456395626, + 0.012115650810301304, + 0.009826984256505966, + 0.012930135242640972, + 0.015189481899142265, + 0.007907899096608162, + 0.012341085821390152, + 0.010486684739589691, + 0.013040418736636639, + 0.01593145728111267, + 0.00914257112890482, + 0.019225621595978737, + 0.011585474014282227, + 0.011934205889701843, + 0.014474384486675262, + 0.0019342054147273302, + 0.002514482010155916, + 0.003142167115584016, + 0.0033288991544395685, + 0.002190169645473361, + 0.0033120412845164537, + 0.002408466534689069, + 0.002405251143500209, + 0.031125839799642563, + 0.020514318719506264, + 0.06394045799970627, + 0.058519795536994934, + 0.00566581217572093, + 0.010545984841883183, + 0.007529747672379017, + 0.008544066920876503, + 0.005686720367521048, + 0.009521671570837498, + 0.00805817823857069, + 0.011360248550772667, + 0.006720123812556267, + 0.006499584764242172, + 0.013115325942635536, + 0.007177713327109814, + 0.0018491727532818913, + 0.0018745552515611053, + 0.0013612335314974189, + 0.0016940644709393382, + 0.0017381752841174603, + 0.0022274767979979515, + 0.0024740241933614016, + 0.00268352753482759, + 0.008737537078559399, + 0.0074098785407841206, + 0.006585957948118448, + 0.008033128455281258, + 0.00711016682907939, + 0.0073395646177232265, + 0.004917929880321026, + 0.009229353629052639, + 0.002611595904454589, + 0.003077723318710923, + 0.006934934761375189, + 0.0060126869939267635, + 0.0034269047901034355, + 0.010324887931346893, + 0.00490611232817173, + 0.0044939229264855385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.03372841328382492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033709511160850525 + ], + "opposite_logprob_tokens": [ + -3.4087095260620117 + ], + "logprob_reward": -0.033709511160850525, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033709511160850525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07718188315629959, + 0.44397199153900146, + 0.09295736253261566, + 0.14161840081214905, + 0.0303190927952528, + 0.05012941733002663, + 0.027980798855423927, + 0.04143267869949341, + 0.05211424455046654, + 0.042294085025787354 + ], + "credit_map": [ + 0.014824744313955307, + 0.010220097377896309, + 0.010468235239386559, + 0.006598716136068106, + 0.008150644600391388, + 0.008429575711488724, + 0.005259234923869371, + 0.013230633921921253, + 0.041837964206933975, + 0.06841372698545456, + 0.05040021613240242, + 0.0359317846596241, + 0.10435064882040024, + 0.06113375723361969, + 0.025929879397153854, + 0.055974025279283524, + 0.01672564446926117, + 0.009463938884437084, + 0.011523386463522911, + 0.008548097684979439, + 0.009479879401624203, + 0.015253505669534206, + 0.007745042443275452, + 0.0142178600654006, + 0.01464434340596199, + 0.0134042426943779, + 0.019985409453511238, + 0.015009765513241291, + 0.02041015774011612, + 0.01692093163728714, + 0.01655045710504055, + 0.024693084880709648, + 0.003348225262016058, + 0.0034669209271669388, + 0.004733875393867493, + 0.0038501641247421503, + 0.004176655784249306, + 0.004048691131174564, + 0.002983539365231991, + 0.0037110212724655867, + 0.006552115082740784, + 0.006391320377588272, + 0.01021728664636612, + 0.011335354298353195, + 0.0035667475312948227, + 0.004199331160634756, + 0.0038170877378433943, + 0.004050172865390778, + 0.0027573395054787397, + 0.005468433722853661, + 0.00303087430074811, + 0.005836925935000181, + 0.002596469596028328, + 0.003260228084400296, + 0.0028742714785039425, + 0.002156256465241313, + 0.005411749240010977, + 0.005834820214658976, + 0.0036704223603010178, + 0.003736932761967182, + 0.00655724573880434, + 0.005803102161735296, + 0.005106609780341387, + 0.005311798769980669, + 0.008232328109443188, + 0.006542946211993694, + 0.0055854953825473785, + 0.007960053160786629, + 0.005714328959584236, + 0.005334123037755489, + 0.004231927450746298, + 0.008513038046658039, + 0.004374357871711254, + 0.004758348688483238, + 0.00529452646151185, + 0.0064978632144629955, + 0.005467346403747797, + 0.006963334511965513, + 0.005206383764743805, + 0.0037319210823625326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.033709511160850525 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076745577156544 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.011076745577156544, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076745577156544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10299656540155411, + 0.3850518763065338, + 0.040691033005714417, + 0.023363342508673668, + 0.026341121643781662, + 0.2513141334056854, + 0.033628545701503754, + 0.045126259326934814, + 0.0490647554397583, + 0.04242236539721489 + ], + "credit_map": [ + 0.015036706812679768, + 0.011874926276504993, + 0.013725065626204014, + 0.013191749341785908, + 0.015077970921993256, + 0.012842495925724506, + 0.007435799576342106, + 0.013811863027513027, + 0.036478109657764435, + 0.0613984614610672, + 0.043983228504657745, + 0.030653152614831924, + 0.09004957973957062, + 0.05313321575522423, + 0.022242410108447075, + 0.047113724052906036, + 0.006712831556797028, + 0.0062345112673938274, + 0.004745038691908121, + 0.004748701583594084, + 0.004205787088721991, + 0.004871485289186239, + 0.0042308634147048, + 0.004941820632666349, + 0.0035601395647972822, + 0.0028866815846413374, + 0.0031762232538312674, + 0.003377024782821536, + 0.003107886528596282, + 0.0028416879940778017, + 0.0023004035465419292, + 0.0021132950205355883, + 0.0033960086293518543, + 0.002350025111809373, + 0.004089666996151209, + 0.0029142422135919333, + 0.00474056089296937, + 0.0033529491629451513, + 0.002477599773555994, + 0.003020069096237421, + 0.04132656380534172, + 0.02223856747150421, + 0.08304934948682785, + 0.06478860974311829, + 0.007172539364546537, + 0.013427811674773693, + 0.010486875660717487, + 0.008823827840387821, + 0.003374264342710376, + 0.006102137267589569, + 0.003592838067561388, + 0.006025549024343491, + 0.0025401522871106863, + 0.0030139354057610035, + 0.006084315944463015, + 0.002895358018577099, + 0.005030670668929815, + 0.007100326474756002, + 0.00584826385602355, + 0.004412956535816193, + 0.006330918986350298, + 0.00465049734339118, + 0.00530218007043004, + 0.006450447253882885, + 0.008341677486896515, + 0.0060690767131745815, + 0.005745144560933113, + 0.0065003931522369385, + 0.007171320263296366, + 0.0047248429618775845, + 0.003889916930347681, + 0.006622391287237406, + 0.0030707891564816236, + 0.003569156164303422, + 0.0070431893691420555, + 0.006510385777801275, + 0.0035380912013351917, + 0.009315158240497112, + 0.00505342660471797, + 0.0043221707455813885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.011076745577156544 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035763919353485 + ], + "opposite_logprob_tokens": [ + -1.9103575944900513 + ], + "logprob_reward": -0.16035763919353485, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035763919353485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06081448495388031, + 0.2864156663417816, + 0.04071417450904846, + 0.04177464172244072, + 0.03584449738264084, + 0.3878919184207916, + 0.04537443816661835, + 0.0354444719851017, + 0.03496174141764641, + 0.030763892456889153 + ], + "credit_map": [ + 0.00687235314399004, + 0.006243146490305662, + 0.009473422542214394, + 0.007442985195666552, + 0.007566129323095083, + 0.007905559614300728, + 0.003791608614847064, + 0.011519285850226879, + 0.028312241658568382, + 0.03634963184595108, + 0.035492103546857834, + 0.027474429458379745, + 0.057528309524059296, + 0.0404391884803772, + 0.01717192493379116, + 0.043647851794958115, + 0.00911793578416109, + 0.0032459720969200134, + 0.005488668102771044, + 0.002975601004436612, + 0.004540438763797283, + 0.005352916195988655, + 0.003558333497494459, + 0.006434311158955097, + 0.004158987198024988, + 0.005125748459249735, + 0.006332471501082182, + 0.006353330798447132, + 0.004093199037015438, + 0.004199368413537741, + 0.003471875563263893, + 0.008039664477109909, + 0.00516551686450839, + 0.0038943111430853605, + 0.005138196982443333, + 0.0029772294219583273, + 0.005596213508397341, + 0.006036515813320875, + 0.0026985493022948503, + 0.004337967839092016, + 0.06118633225560188, + 0.03519706800580025, + 0.12238600105047226, + 0.10527251660823822, + 0.0112858721986413, + 0.020350687205791473, + 0.015451643615961075, + 0.01676182635128498, + 0.00450993562117219, + 0.007878913544118404, + 0.005561857484281063, + 0.008113224990665913, + 0.0036301410291343927, + 0.0031049770768731833, + 0.008221963420510292, + 0.004353429190814495, + 0.004096238873898983, + 0.005038305651396513, + 0.004607276525348425, + 0.0030639884062111378, + 0.006318948231637478, + 0.003994188271462917, + 0.003999663982540369, + 0.004325861111283302, + 0.0065871053375303745, + 0.004517816472798586, + 0.004258692730218172, + 0.004049642477184534, + 0.0037590228021144867, + 0.004301753360778093, + 0.003644032869488001, + 0.0038436735048890114, + 0.005149567499756813, + 0.004282965324819088, + 0.004072247073054314, + 0.00575696025043726, + 0.0040023536421358585, + 0.002733005443587899, + 0.0027982499450445175, + 0.001968544442206621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.16035763919353485 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium-long shot, richly hand-painted in a watercolor style, captures a small, iron-hulled, rifle-powered boat sailing across the vast, tranquil Weishan Lake from a slightly sideways, horizontal angle. The boat's hull is a weathered, dark gray-green, speckled with rust along its edges and seams, reflecting a dull metallic sheen under the soft, diffuse light. With a long, slightly hoarse whistle, the boat begins to steadily decelerate. The V-shaped white waves previously stirred up by the bow gradually soften, transforming into delicate ripples that spread outward. At this point, the entire frame is accompanied by a slight flicker, like that of old film, as the camera begins a slow and precise shift in focus: in the foreground, the crystal-clear ripples on the water gradually lose focus, becoming blurred and dreamy, while the camera's focus steadily shifts backward, finally settling sharply on an old flag at the stern, gently fluttering in the breeze. The bright red color of the flag has faded, revealing a dull ochre color. The edges are worn and tattered, with fine tassel-like lines. Against the backdrop of the gray-blue lake and sky, it creates a tranquil, nostalgic and slightly melancholic poetic atmosphere, like a moving ink painting carrying the memories of the years.", + "seed": 1742, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_8/index_5029_seed_1742_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1573646642267704, + "vlm_logprob": -0.1573646642267704, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.1573646642267704, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4886355400085449 + ], + "opposite_logprob_tokens": [ + -1.113635540008545 + ], + "logprob_reward": -0.4886355400085449, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4886355400085449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1250738501548767, + 0.1662423014640808, + 0.03645747900009155, + 0.26779159903526306, + 0.06789223849773407, + 0.0750657171010971, + 0.02810162864625454, + 0.05271264165639877, + 0.0958421602845192, + 0.08482042700052261 + ], + "credit_map": [ + 0.011952842585742474, + 0.010669833980500698, + 0.012241456657648087, + 0.015626685693860054, + 0.00886731781065464, + 0.015410433523356915, + 0.02635294571518898, + 0.02395232766866684, + 0.014680405147373676, + 0.009099649265408516, + 0.05471557006239891, + 0.01228056289255619, + 0.012579020112752914, + 0.011715018190443516, + 0.016636932268738747, + 0.034535132348537445, + 0.0035240058787167072, + 0.004655158147215843, + 0.004171284381300211, + 0.004893741570413113, + 0.004636713769286871, + 0.007665600627660751, + 0.002945700893178582, + 0.003965269774198532, + 0.006704134400933981, + 0.01010691374540329, + 0.010376660153269768, + 0.025090821087360382, + 0.012858494184911251, + 0.012960582040250301, + 0.0884970873594284, + 0.10119685530662537, + 0.0050760796293616295, + 0.004406983032822609, + 0.01024129893630743, + 0.007438769564032555, + 0.0074025532230734825, + 0.007980279624462128, + 0.012326586991548538, + 0.013019679114222527, + 0.004886862821877003, + 0.0058140140026807785, + 0.01647946797311306, + 0.010141929611563683, + 0.005168870557099581, + 0.0035465818364173174, + 0.009276057593524456, + 0.019751926884055138, + 0.0029613268561661243, + 0.002481585368514061, + 0.004476842004805803, + 0.003691983874887228, + 0.0024135620333254337, + 0.0024766218848526478, + 0.00374139123596251, + 0.005858310963958502, + 0.008392567746341228, + 0.0077452813275158405, + 0.00670921104028821, + 0.008417898789048195, + 0.0030621637124568224, + 0.003726405557245016, + 0.0050081717781722546, + 0.009650936350226402, + 0.008484261110424995, + 0.004524845164269209, + 0.030103877186775208, + 0.02949802577495575, + 0.0029359867330640554, + 0.00818189512938261, + 0.00530984764918685, + 0.006803412456065416, + 0.010878708213567734, + 0.021633004769682884, + 0.008791725151240826, + 0.014568785205483437, + 0.004335478413850069, + 0.005663783755153418, + 0.007115127053111792, + 0.011833807453513145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.4886355400085449 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029849348589777946 + ], + "opposite_logprob_tokens": [ + -3.5298492908477783 + ], + "logprob_reward": -0.029849348589777946, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029849348589777946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1334403157234192, + 0.08462591469287872, + 0.04358663409948349, + 0.43861833214759827, + 0.050852399319410324, + 0.02460263855755329, + 0.0217069610953331, + 0.09299200773239136, + 0.05060670152306557, + 0.058968063443899155 + ], + "credit_map": [ + 0.014316866174340248, + 0.01534435898065567, + 0.015022525563836098, + 0.017395678907632828, + 0.011166236363351345, + 0.02115856669843197, + 0.02042408287525177, + 0.018611997365951538, + 0.009833629243075848, + 0.008626689203083515, + 0.013551265001296997, + 0.0064833760261535645, + 0.012525750324130058, + 0.013314567506313324, + 0.010297512635588646, + 0.009993127547204494, + 0.0045907264575362206, + 0.006322948727756739, + 0.005221300758421421, + 0.007056571543216705, + 0.0032897824421525, + 0.004902195185422897, + 0.00630792323499918, + 0.005895184352993965, + 0.009242834523320198, + 0.011554732918739319, + 0.014237952418625355, + 0.039806220680475235, + 0.01609734818339348, + 0.02054126374423504, + 0.15483936667442322, + 0.1722986102104187, + 0.003886294784024358, + 0.004639570135623217, + 0.007747239898890257, + 0.009986380115151405, + 0.005076124332845211, + 0.0043656122870743275, + 0.007771342061460018, + 0.007379836868494749, + 0.0020502060651779175, + 0.0017760859336704016, + 0.005012795329093933, + 0.0025980437640100718, + 0.0017260677414014935, + 0.0016241985140368342, + 0.002942045219242573, + 0.006873194593936205, + 0.0028712458442896605, + 0.0012531530810520053, + 0.003600268391892314, + 0.002843688940629363, + 0.001254372764378786, + 0.0019223530543968081, + 0.0025531237479299307, + 0.005408757366240025, + 0.011156441643834114, + 0.014022166840732098, + 0.011610559187829494, + 0.013069085776805878, + 0.009013560600578785, + 0.005887462757527828, + 0.009403988718986511, + 0.01882874220609665, + 0.005406832322478294, + 0.0031850032974034548, + 0.01395172718912363, + 0.011630205437541008, + 0.001974428305402398, + 0.005981793627142906, + 0.0025256185326725245, + 0.005951095838099718, + 0.00702352449297905, + 0.009562079794704914, + 0.008335917256772518, + 0.012731803581118584, + 0.0034535573795437813, + 0.0043680728413164616, + 0.006177070084959269, + 0.007316035684198141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.029849348589777946 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207853555679321 + ], + "opposite_logprob_tokens": [ + -2.8120784759521484 + ], + "logprob_reward": -0.06207853555679321, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207853555679321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1497923582792282, + 0.12217941135168076, + 0.07054165005683899, + 0.2917262315750122, + 0.060584958642721176, + 0.04792385548353195, + 0.032741859555244446, + 0.058661796152591705, + 0.08328897505998611, + 0.08255890011787415 + ], + "credit_map": [ + 0.021591678261756897, + 0.014526954852044582, + 0.01871008798480034, + 0.01571682281792164, + 0.014068877324461937, + 0.018682831898331642, + 0.02546958066523075, + 0.021025512367486954, + 0.0100707383826375, + 0.008158328011631966, + 0.029653314501047134, + 0.009143456816673279, + 0.011425135657191277, + 0.012734735384583473, + 0.014359638094902039, + 0.026634065434336662, + 0.00729832099750638, + 0.008649983443319798, + 0.00745053868740797, + 0.0072852857410907745, + 0.008284524083137512, + 0.012923037633299828, + 0.008965341374278069, + 0.00968462135642767, + 0.009872755035758018, + 0.017724843695759773, + 0.017763204872608185, + 0.028249239549040794, + 0.01248204242438078, + 0.016550391912460327, + 0.08894435316324234, + 0.10013940185308456, + 0.003891808446496725, + 0.002662187209352851, + 0.007906688377261162, + 0.006899281404912472, + 0.005528388079255819, + 0.00787123292684555, + 0.01329451147466898, + 0.012530860491096973, + 0.005378303118050098, + 0.0029583184514194727, + 0.007903377525508404, + 0.005500956904143095, + 0.0038305427879095078, + 0.003621708368882537, + 0.0069709960371255875, + 0.011759651824831963, + 0.004676592070609331, + 0.002667507156729698, + 0.005946674849838018, + 0.0041772122494876385, + 0.002091724658384919, + 0.002532424870878458, + 0.003898129565641284, + 0.006751595064997673, + 0.006994249764829874, + 0.0049773212522268295, + 0.009492469020187855, + 0.010853970423340797, + 0.0037664996925741434, + 0.0062775020487606525, + 0.005807748530060053, + 0.010492036119103432, + 0.007393974810838699, + 0.005856695119291544, + 0.024812648072838783, + 0.023221002891659737, + 0.0030902454163879156, + 0.005750644952058792, + 0.004788163583725691, + 0.008375600911676884, + 0.010767065919935703, + 0.01122211292386055, + 0.00948425568640232, + 0.015352144837379456, + 0.00558152049779892, + 0.009347853250801563, + 0.009617681615054607, + 0.011186259798705578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.06207853555679321 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026344453915953636 + ], + "opposite_logprob_tokens": [ + -3.6513445377349854 + ], + "logprob_reward": -0.026344453915953636, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026344453915953636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09712303429841995, + 0.15385761857032776, + 0.08978195488452911, + 0.30780094861984253, + 0.09387722611427307, + 0.05647048354148865, + 0.03247900307178497, + 0.11458888649940491, + 0.021689800545573235, + 0.03233104199171066 + ], + "credit_map": [ + 0.00962384045124054, + 0.0075157396495342255, + 0.011931994929909706, + 0.00897760409861803, + 0.006049502175301313, + 0.011266225017607212, + 0.024196209385991096, + 0.01756192184984684, + 0.010791768319904804, + 0.009279632940888405, + 0.04359639808535576, + 0.009480634704232216, + 0.014726121909916401, + 0.015547044575214386, + 0.013312146067619324, + 0.03712388128042221, + 0.011689688079059124, + 0.01007800828665495, + 0.010338647291064262, + 0.010348068550229073, + 0.009680739603936672, + 0.015196605585515499, + 0.011112608015537262, + 0.011337594129145145, + 0.02169191464781761, + 0.03369617462158203, + 0.02380228601396084, + 0.030473629012703896, + 0.01126858964562416, + 0.011665655300021172, + 0.08716484159231186, + 0.08803785592317581, + 0.0072187636978924274, + 0.008194378577172756, + 0.015236571431159973, + 0.015116017311811447, + 0.008807922713458538, + 0.008776946924626827, + 0.015214809216558933, + 0.01531181763857603, + 0.003976223990321159, + 0.004548774566501379, + 0.009368179365992546, + 0.008813475258648396, + 0.004149306565523148, + 0.0033345473930239677, + 0.009660927578806877, + 0.012619046494364738, + 0.004410935565829277, + 0.0036096672993153334, + 0.004708481021225452, + 0.005478703882545233, + 0.0023326855152845383, + 0.0025525062810629606, + 0.0038537592627108097, + 0.005532265640795231, + 0.013003842905163765, + 0.017657112330198288, + 0.013738823123276234, + 0.0157859418541193, + 0.010274241678416729, + 0.009480181150138378, + 0.011084078811109066, + 0.023564670234918594, + 0.002712382236495614, + 0.0022826516069471836, + 0.003649364924058318, + 0.004881720989942551, + 0.0013727470068261027, + 0.0016792915994301438, + 0.0024402712006121874, + 0.0026713688857853413, + 0.0035764460917562246, + 0.004824147559702396, + 0.0038961307145655155, + 0.0048025138676166534, + 0.002762326505035162, + 0.0036967338528484106, + 0.004251064267009497, + 0.004521682858467102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.026344453915953636 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a small, iron-hulled, rifle-powered boat sailing on Weishan Lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull dark gray-green with visible rust speckles and a dull metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boat steadily decelerate while the V-shaped waves soften into delicate ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera focus shift from blurred foreground ripples to a sharp focus on an old, faded red flag at the stern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video accompanied by a slight flicker resembling old film?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799154430627823 + ], + "opposite_logprob_tokens": [ + -1.804915428161621 + ], + "logprob_reward": -0.1799154430627823, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799154430627823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14643974602222443, + 0.10588886588811874, + 0.05570618435740471, + 0.4651798903942108, + 0.053918104618787766, + 0.029770221561193466, + 0.01871958002448082, + 0.08419977128505707, + 0.021385518833994865, + 0.01879209280014038 + ], + "credit_map": [ + 0.013445144519209862, + 0.011900893412530422, + 0.017811452969908714, + 0.01763254962861538, + 0.008245688863098621, + 0.015280275605618954, + 0.03415831923484802, + 0.027965420857071877, + 0.008785427547991276, + 0.004254471976310015, + 0.03655108064413071, + 0.008227373473346233, + 0.004739667288959026, + 0.00688047194853425, + 0.007726642768830061, + 0.02872372791171074, + 0.005558736156672239, + 0.005726504139602184, + 0.006215444300323725, + 0.007299911696463823, + 0.006125487852841616, + 0.010482501238584518, + 0.008133997209370136, + 0.006163604091852903, + 0.02339225448668003, + 0.03741893544793129, + 0.030627969652414322, + 0.045091595500707626, + 0.013476855121552944, + 0.013176369480788708, + 0.14609141647815704, + 0.15590447187423706, + 0.003825257532298565, + 0.004740139935165644, + 0.008055035024881363, + 0.011175470426678658, + 0.004828099627047777, + 0.005427410826086998, + 0.0073521011509001255, + 0.008514588698744774, + 0.00209126272238791, + 0.001330242957919836, + 0.004926573019474745, + 0.006077924277633429, + 0.002206207485869527, + 0.001767667941749096, + 0.004273033235222101, + 0.007097308058291674, + 0.002879965817555785, + 0.0012191830901429057, + 0.003492620773613453, + 0.00265685748308897, + 0.0009089770610444248, + 0.001184237189590931, + 0.002051019575446844, + 0.004326718859374523, + 0.009273871779441833, + 0.013262566179037094, + 0.009106216952204704, + 0.01325166691094637, + 0.008187140338122845, + 0.007419869769364595, + 0.009269751608371735, + 0.014428689144551754, + 0.0025764089077711105, + 0.0030891180504113436, + 0.004133882932364941, + 0.00388909038156271, + 0.0010599484667181969, + 0.00147489202208817, + 0.0018435512902215123, + 0.0033186268992722034, + 0.002013742458075285, + 0.0025035336147993803, + 0.0027170151006430387, + 0.003594174049794674, + 0.0009437691769562662, + 0.0011518836254253983, + 0.0023308484815061092, + 0.0035371240228414536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.1799154430627823 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_9/index_tensor([3875, 3875])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_9/index_tensor([3875, 3875])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6da89bdd0c824661d74f4c16e3a0ee0ccd534c72 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_9/index_tensor([3875, 3875])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "seamless 360-degree tracking close-up unfolds smoothly around a curious sparrow perched on a weathered, ancient stone wall. The sparrow, small and plump, with feathers a complex interweaving of earthy browns, soft grays, and sharp black stripes, tilts its head from time to time, its dark, shiny eyes gleaming with inquiry. The wall it stands on is itself a tapestry of natural textures, composed of irregularly shaped stones. Vast swathes of vibrant green and deep emerald moss cling tenaciously to the rough surface, while clumps of delicate ferns sprout from the damp crevices. Steadying itself on its delicate legs and claws, the sparrow executes a series of quick, light hops along the wall's uneven top, pausing intermittently to delicately peck at a patch of moss with its tiny beak or peer curiously into a deep crevice between the stones, completely absorbed in its exploration. As the camera continues to rotate smoothly, soft, diffused sunlight is evenly distributed across the entire scene, constantly revealing new angles and microscopic details of the stones and moss, creating a dynamic yet tranquil background. The entire image is filled with a peaceful, natural atmosphere, capturing an intimate and fascinating moment of pure beauty and curiosity in nature.", + "seed": 11714, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_9/index_3875_seed_11714_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2360213004052639, + "vlm_logprob": -0.2360213004052639, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.2360213004052639, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.525785505771637 + ], + "opposite_logprob_tokens": [ + -0.900785505771637 + ], + "logprob_reward": -0.525785505771637, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.525785505771637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15207119286060333, + 0.03132639080286026, + 0.1266736388206482, + 0.06094641238451004, + 0.26121824979782104, + 0.0431639701128006, + 0.035048410296440125, + 0.02182779461145401, + 0.03318484500050545, + 0.23453910648822784 + ], + "credit_map": [ + 0.012860419228672981, + 0.019399505108594894, + 0.017327746376395226, + 0.03254710137844086, + 0.022088076919317245, + 0.01513308472931385, + 0.014825020916759968, + 0.017890233546495438, + 0.0028445005882531404, + 0.004215976223349571, + 0.003149560187011957, + 0.004420440178364515, + 0.005292578600347042, + 0.0035893043968826532, + 0.00410279119387269, + 0.0037112408317625523, + 0.012929324992001057, + 0.014921806752681732, + 0.02217048406600952, + 0.014841219410300255, + 0.012286728248000145, + 0.020793914794921875, + 0.007627378683537245, + 0.021102769300341606, + 0.006920600309967995, + 0.006570588797330856, + 0.00931891705840826, + 0.008521711453795433, + 0.003594537265598774, + 0.01134930644184351, + 0.004767234902828932, + 0.009903518483042717, + 0.04398268833756447, + 0.04091999679803848, + 0.06471358239650726, + 0.025237593799829483, + 0.025340519845485687, + 0.014662330038845539, + 0.015535940416157246, + 0.030825579538941383, + 0.0056715006940066814, + 0.004835380241274834, + 0.0051987855695188046, + 0.005206447560340166, + 0.008228546939790249, + 0.0050151655450463295, + 0.0044372910633683205, + 0.004570850636810064, + 0.0045795137993991375, + 0.003006346058100462, + 0.0036041883286088705, + 0.004270154517143965, + 0.004467639606446028, + 0.007562071084976196, + 0.0029251736123114824, + 0.00463332561776042, + 0.0025960044004023075, + 0.002524884417653084, + 0.0034557932522147894, + 0.0028292841743677855, + 0.002064730040729046, + 0.0028003097977489233, + 0.0028444991912692785, + 0.002712290035560727, + 0.004425923805683851, + 0.0029114075005054474, + 0.006032126490026712, + 0.0030768285505473614, + 0.004876185208559036, + 0.003731365082785487, + 0.0034233015030622482, + 0.004707707092165947, + 0.01817144826054573, + 0.010265825316309929, + 0.09604904055595398, + 0.013757253997027874, + 0.011228211224079132, + 0.0199282169342041, + 0.008427673950791359, + 0.05671143904328346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.525785505771637 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081814765930176 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081814765930176, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081814765930176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0626586526632309, + 0.02237420529127121, + 0.052252136170864105, + 0.04439253732562065, + 0.15664461255073547, + 0.08054979145526886, + 0.03946664556860924, + 0.017603058367967606, + 0.11057305335998535, + 0.41348522901535034 + ], + "credit_map": [ + 0.0053278598934412, + 0.007475510239601135, + 0.007567453198134899, + 0.01081875804811716, + 0.009700329974293709, + 0.00784499105066061, + 0.00715654669329524, + 0.006767197512090206, + 0.0020394569728523493, + 0.0031437124125659466, + 0.002282862551510334, + 0.00551376398652792, + 0.0024459841661155224, + 0.0020945819560438395, + 0.0020683256443589926, + 0.0027855169028043747, + 0.004401694517582655, + 0.006212451495230198, + 0.008383299224078655, + 0.006010203622281551, + 0.005994989536702633, + 0.008476583287119865, + 0.004584422335028648, + 0.008188490755856037, + 0.004730640910565853, + 0.005897726863622665, + 0.006041462533175945, + 0.007551524788141251, + 0.0032600380945950747, + 0.007735488936305046, + 0.003855322953313589, + 0.005320334807038307, + 0.028827032074332237, + 0.023345405235886574, + 0.04397987574338913, + 0.015151060186326504, + 0.008969124406576157, + 0.008232709020376205, + 0.008392058312892914, + 0.019747352227568626, + 0.012997505255043507, + 0.006559920962899923, + 0.017422568053007126, + 0.008624663576483727, + 0.008072148077189922, + 0.007264678832143545, + 0.0053381118923425674, + 0.014270197600126266, + 0.005792880430817604, + 0.00429051648825407, + 0.0046456968411803246, + 0.008352723903954029, + 0.0029079928062856197, + 0.005095588508993387, + 0.003535487921908498, + 0.00484575517475605, + 0.0021196086890995502, + 0.0020934203639626503, + 0.0025403809268027544, + 0.002203678246587515, + 0.0022424752824008465, + 0.0018348192097619176, + 0.002098777564242482, + 0.0024698989000171423, + 0.008843546733260155, + 0.0073667773976922035, + 0.025753309950232506, + 0.009154323488473892, + 0.01401399727910757, + 0.013862923718988895, + 0.008163891732692719, + 0.023414289578795433, + 0.03385859727859497, + 0.014904416166245937, + 0.17508971691131592, + 0.022800829261541367, + 0.012527129612863064, + 0.03574301674962044, + 0.013463581912219524, + 0.1050979495048523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.011081814765930176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525260902941227 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.00525260902941227, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525260902941227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10921426117420197, + 0.06147272512316704, + 0.11927513778209686, + 0.1512652188539505, + 0.12413295358419418, + 0.07678588479757309, + 0.14696449041366577, + 0.059697434306144714, + 0.029801107943058014, + 0.12139078974723816 + ], + "credit_map": [ + 0.01120968908071518, + 0.014545968733727932, + 0.011541287414729595, + 0.028676092624664307, + 0.009690076112747192, + 0.009887213818728924, + 0.009495391510426998, + 0.01416854653507471, + 0.005373795051127672, + 0.007652871776372194, + 0.006767161190509796, + 0.010753672569990158, + 0.010187763720750809, + 0.006624429486691952, + 0.006508563179522753, + 0.007604469545185566, + 0.012647018767893314, + 0.01501979399472475, + 0.025186551734805107, + 0.019196627661585808, + 0.011221940629184246, + 0.011907114647328854, + 0.011700330302119255, + 0.012395759113132954, + 0.01906219869852066, + 0.01847977191209793, + 0.017414765432476997, + 0.021510332822799683, + 0.014110959134995937, + 0.031055552884936333, + 0.009308913722634315, + 0.020322727039456367, + 0.02226843312382698, + 0.017113687470555305, + 0.030578026548027992, + 0.01169040147215128, + 0.012790274806320667, + 0.007081589195877314, + 0.007200906984508038, + 0.015409636311233044, + 0.012013153173029423, + 0.0065940264612436295, + 0.013268358074128628, + 0.007508923765271902, + 0.01181038562208414, + 0.00847869087010622, + 0.0061486768536269665, + 0.010963669046759605, + 0.018659207969903946, + 0.013933857902884483, + 0.014998747035861015, + 0.0199684239923954, + 0.020687256008386612, + 0.027489233762025833, + 0.012318805791437626, + 0.018908968195319176, + 0.010047253221273422, + 0.006927650421857834, + 0.008308648131787777, + 0.007054849527776241, + 0.005149010103195906, + 0.009612574242055416, + 0.005688773933798075, + 0.006908670533448458, + 0.003028636099770665, + 0.0029171938076615334, + 0.0049708569422364235, + 0.0034128197003155947, + 0.0040783048607409, + 0.004425605293363333, + 0.0037860642187297344, + 0.0031816288828849792, + 0.009433265775442123, + 0.006938783917576075, + 0.04374902322888374, + 0.008978081867098808, + 0.008261458948254585, + 0.010940886102616787, + 0.005803062580525875, + 0.027286231517791748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.00525260902941227 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248339846730232 + ], + "opposite_logprob_tokens": [ + -5.255248546600342 + ], + "logprob_reward": -0.005248339846730232, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248339846730232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09742988646030426, + 0.0611155740916729, + 0.17033250629901886, + 0.10981226712465286, + 0.12628957629203796, + 0.14376944303512573, + 0.033407062292099, + 0.04588548466563225, + 0.15920905768871307, + 0.052749212831258774 + ], + "credit_map": [ + 0.008383459411561489, + 0.014093488454818726, + 0.010916653089225292, + 0.017757253721356392, + 0.016731223091483116, + 0.008701276034116745, + 0.01169665902853012, + 0.009149858728051186, + 0.006023263093084097, + 0.00846767332404852, + 0.0055230711586773396, + 0.01296436507254839, + 0.007100374437868595, + 0.006215363275259733, + 0.006421539466828108, + 0.008399917744100094, + 0.01748453639447689, + 0.028951425105333328, + 0.022977005690336227, + 0.023821713402867317, + 0.0175612960010767, + 0.021122297272086143, + 0.021987494081258774, + 0.01642671972513199, + 0.01314546912908554, + 0.009866134263575077, + 0.015389365144073963, + 0.01343782152980566, + 0.008758273907005787, + 0.01942342519760132, + 0.009947070851922035, + 0.019844695925712585, + 0.019086018204689026, + 0.01714158058166504, + 0.02163086086511612, + 0.011877397075295448, + 0.020591501146554947, + 0.010745818726718426, + 0.009667576290667057, + 0.015548812225461006, + 0.015171336941421032, + 0.01586264744400978, + 0.020032498985528946, + 0.015144228003919125, + 0.029190469533205032, + 0.014176478609442711, + 0.015502681024372578, + 0.018689095973968506, + 0.004546168260276318, + 0.005163728259503841, + 0.003246587235480547, + 0.005596410017460585, + 0.003468171926215291, + 0.004454954992979765, + 0.003432417754083872, + 0.0034986212849617004, + 0.006246951408684254, + 0.0051219710148870945, + 0.005765071604400873, + 0.005478793289512396, + 0.00484052300453186, + 0.006945458706468344, + 0.005462544970214367, + 0.006024164613336325, + 0.013128138147294521, + 0.013465282507240772, + 0.031026240438222885, + 0.015525983646512032, + 0.024164246395230293, + 0.019751518964767456, + 0.012945457361638546, + 0.029202185571193695, + 0.00677471375092864, + 0.004725719802081585, + 0.006723193451762199, + 0.00475001847371459, + 0.009206067770719528, + 0.008109943009912968, + 0.0041424245573580265, + 0.008317125961184502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 5 + ], + "selected_reward": -0.005248339846730232 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327382326126099 + ], + "opposite_logprob_tokens": [ + -0.7577382326126099 + ], + "logprob_reward": -0.6327382326126099, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327382326126099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09768068045377731, + 0.038738545030355453, + 0.06328202039003372, + 0.10297441482543945, + 0.06393351405858994, + 0.13520509004592896, + 0.02813120372593403, + 0.01844526082277298, + 0.07816892862319946, + 0.37344032526016235 + ], + "credit_map": [ + 0.008290198631584644, + 0.012829347513616085, + 0.01051709521561861, + 0.0187047328799963, + 0.012895156629383564, + 0.010291215032339096, + 0.011584565974771976, + 0.01256837509572506, + 0.0037287806626409292, + 0.0054124570451676846, + 0.004089137073606253, + 0.006303280591964722, + 0.005844929255545139, + 0.005002229008823633, + 0.003605993464589119, + 0.004751737229526043, + 0.007838004268705845, + 0.012572037987411022, + 0.009885425679385662, + 0.00950558390468359, + 0.004691924899816513, + 0.00757288932800293, + 0.004076408222317696, + 0.0071397507563233376, + 0.012495402246713638, + 0.010332610458135605, + 0.015119102783501148, + 0.012254740111529827, + 0.008500511758029461, + 0.018673403188586235, + 0.008070115931332111, + 0.017528526484966278, + 0.009777595289051533, + 0.007798678707331419, + 0.010988953523337841, + 0.007585866842418909, + 0.007579389493912458, + 0.00580615596845746, + 0.006479516159743071, + 0.007917363196611404, + 0.014985169284045696, + 0.015547730959951878, + 0.015581193380057812, + 0.015285865403711796, + 0.03102070838212967, + 0.012940538115799427, + 0.015257674269378185, + 0.014586216770112514, + 0.005457643885165453, + 0.0028765895403921604, + 0.0032101899851113558, + 0.005934085231274366, + 0.0020613595843315125, + 0.0033819773234426975, + 0.0020066106226295233, + 0.003202747320756316, + 0.0022562099620699883, + 0.002293570665642619, + 0.0025610753800719976, + 0.0019740681163966656, + 0.0016383365727961063, + 0.0025919550098478794, + 0.002142533892765641, + 0.002987511223182082, + 0.006557969376444817, + 0.007034896407276392, + 0.01319023035466671, + 0.009604932740330696, + 0.011282744817435741, + 0.009782692417502403, + 0.00885994266718626, + 0.011855521239340305, + 0.02954339049756527, + 0.016184421256184578, + 0.15514251589775085, + 0.02243596501648426, + 0.012290623039007187, + 0.03152087703347206, + 0.013137863017618656, + 0.09318464994430542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 3 + ], + "selected_reward": -0.6327382326126099 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "seamless 360-degree tracking close-up unfolds smoothly around a curious sparrow perched on a weathered, ancient stone wall. The sparrow, small and plump, with feathers a complex interweaving of earthy browns, soft grays, and sharp black stripes, tilts its head from time to time, its dark, shiny eyes gleaming with inquiry. The wall it stands on is itself a tapestry of natural textures, composed of irregularly shaped stones. Vast swathes of vibrant green and deep emerald moss cling tenaciously to the rough surface, while clumps of delicate ferns sprout from the damp crevices. Steadying itself on its delicate legs and claws, the sparrow executes a series of quick, light hops along the wall's uneven top, pausing intermittently to delicately peck at a patch of moss with its tiny beak or peer curiously into a deep crevice between the stones, completely absorbed in its exploration. As the camera continues to rotate smoothly, soft, diffused sunlight is evenly distributed across the entire scene, constantly revealing new angles and microscopic details of the stones and moss, creating a dynamic yet tranquil background. The entire image is filled with a peaceful, natural atmosphere, capturing an intimate and fascinating moment of pure beauty and curiosity in nature.", + "seed": 11715, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_9/index_3875_seed_11715_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.18385501997545362, + "vlm_logprob": -0.18385501997545362, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.18385501997545362, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1304258555173874 + ], + "opposite_logprob_tokens": [ + -2.1304259300231934 + ], + "logprob_reward": -0.1304258555173874, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1304258555173874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12207991629838943, + 0.06020099297165871, + 0.22795656323432922, + 0.0701352059841156, + 0.03293813019990921, + 0.0559956356883049, + 0.19762752950191498, + 0.12627968192100525, + 0.07635781168937683, + 0.03042859397828579 + ], + "credit_map": [ + 0.020468654111027718, + 0.011026273481547832, + 0.011289631016552448, + 0.020928746089339256, + 0.017043594270944595, + 0.012898818589746952, + 0.011799560859799385, + 0.016624627634882927, + 0.010405352339148521, + 0.005021465476602316, + 0.0075006247498095036, + 0.005475752055644989, + 0.008549320511519909, + 0.008943446911871433, + 0.006747240666300058, + 0.007557794451713562, + 0.04127455875277519, + 0.013399085029959679, + 0.016171537339687347, + 0.020387720316648483, + 0.033065732568502426, + 0.04095972701907158, + 0.021561643108725548, + 0.04113656282424927, + 0.0049196877516806126, + 0.005967769306153059, + 0.0057005141861736774, + 0.008600092492997646, + 0.014261807315051556, + 0.006775185465812683, + 0.013992957770824432, + 0.009917192161083221, + 0.0023612447548657656, + 0.004193977452814579, + 0.004066585563123226, + 0.0034545205999165773, + 0.005235129501670599, + 0.003380606649443507, + 0.005800697021186352, + 0.00444537028670311, + 0.0031739496625959873, + 0.005078775342553854, + 0.004007668700069189, + 0.010065101087093353, + 0.007894615642726421, + 0.009009583853185177, + 0.008991752751171589, + 0.007774184923619032, + 0.0046141077764332294, + 0.005241152830421925, + 0.005904568359255791, + 0.008480173535645008, + 0.007505368907004595, + 0.08456618338823318, + 0.01001154724508524, + 0.07130441069602966, + 0.009035373106598854, + 0.012342983856797218, + 0.014418809674680233, + 0.01376058254390955, + 0.015374096110463142, + 0.018518736585974693, + 0.018288206309080124, + 0.024540897458791733, + 0.006519784685224295, + 0.005697058513760567, + 0.009209878742694855, + 0.013960880227386951, + 0.011417029425501823, + 0.00868930108845234, + 0.013088076375424862, + 0.007775801233947277, + 0.0027189371176064014, + 0.004745983984321356, + 0.003367731114849448, + 0.00396366510540247, + 0.0026520637329667807, + 0.0044054193422198296, + 0.0039855074137449265, + 0.004589286632835865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.1304258555173874 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071535646915436 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.016071535646915436, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071535646915436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03557794168591499, + 0.017788492143154144, + 0.10886204987764359, + 0.06844022125005722, + 0.08349171280860901, + 0.06517386436462402, + 0.44466936588287354, + 0.04348626360297203, + 0.03795647993683815, + 0.09455360472202301 + ], + "credit_map": [ + 0.006622148212045431, + 0.003012266708537936, + 0.004760533571243286, + 0.004388600122183561, + 0.005110281053930521, + 0.003626676741987467, + 0.003450016025453806, + 0.004607417155057192, + 0.0028549269773066044, + 0.001946598058566451, + 0.0020812239963561296, + 0.001813868060708046, + 0.0022963550873100758, + 0.0020525259897112846, + 0.0018980354070663452, + 0.002844956936314702, + 0.019141357392072678, + 0.0064542703330516815, + 0.009606856852769852, + 0.010845913551747799, + 0.010822012089192867, + 0.01936492696404457, + 0.010138829238712788, + 0.02248786762356758, + 0.006244785152375698, + 0.005410375073552132, + 0.007892273366451263, + 0.01183925848454237, + 0.006385371554642916, + 0.009775926358997822, + 0.011434216983616352, + 0.00945801381021738, + 0.009522185660898685, + 0.01547041442245245, + 0.011456597596406937, + 0.010956312529742718, + 0.007355074863880873, + 0.008925412781536579, + 0.008862622082233429, + 0.010943086817860603, + 0.006570563651621342, + 0.011113261803984642, + 0.007822701707482338, + 0.00842830166220665, + 0.003848534543067217, + 0.008466976694762707, + 0.008384267799556255, + 0.010539255104959011, + 0.014973641373217106, + 0.015053129754960537, + 0.018330397084355354, + 0.019173510372638702, + 0.012553064152598381, + 0.18404516577720642, + 0.022712018340826035, + 0.1578284502029419, + 0.0023954925127327442, + 0.002869064686819911, + 0.00426793796941638, + 0.0049536097794771194, + 0.005248747766017914, + 0.00736552057787776, + 0.00584038533270359, + 0.010545500554144382, + 0.0025364255998283625, + 0.003857014235109091, + 0.005796448793262243, + 0.0059039234183728695, + 0.004794933367520571, + 0.004767036531120539, + 0.0053687868639826775, + 0.004931907169520855, + 0.005715914070606232, + 0.010767990723252296, + 0.013326357118785381, + 0.020884687080979347, + 0.006280974950641394, + 0.012903846800327301, + 0.009629749692976475, + 0.015044072642922401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 9 + ], + "selected_reward": -0.016071535646915436 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786011651158333 + ], + "opposite_logprob_tokens": [ + -4.634786128997803 + ], + "logprob_reward": -0.009786011651158333, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786011651158333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0712527185678482, + 0.04856101796030998, + 0.1504826843738556, + 0.030965983867645264, + 0.028050484135746956, + 0.05145219713449478, + 0.41119733452796936, + 0.08053598552942276, + 0.049020152539014816, + 0.07848145067691803 + ], + "credit_map": [ + 0.012491107918322086, + 0.007158818654716015, + 0.006741047836840153, + 0.012834408320486546, + 0.0077676549553871155, + 0.008101404644548893, + 0.008408527821302414, + 0.007749755866825581, + 0.00805491954088211, + 0.0050054071471095085, + 0.004351379815489054, + 0.004101215861737728, + 0.004467631224542856, + 0.007686331868171692, + 0.005134885665029287, + 0.009759249165654182, + 0.02860066667199135, + 0.008697960525751114, + 0.010722012259066105, + 0.013064054772257805, + 0.017972692847251892, + 0.03032984398305416, + 0.013467022217810154, + 0.027628442272543907, + 0.0027160614263266325, + 0.0019873050041496754, + 0.004365457221865654, + 0.004797908011823893, + 0.003896402195096016, + 0.0035699408035725355, + 0.005225394386798143, + 0.00440751388669014, + 0.0022897524759173393, + 0.0035243332386016846, + 0.004302429035305977, + 0.0026918211951851845, + 0.0038337334990501404, + 0.0028101038187742233, + 0.004452706314623356, + 0.004145605023950338, + 0.004224670585244894, + 0.005268815439194441, + 0.0035878231283277273, + 0.006012819707393646, + 0.0054466198198497295, + 0.011829451657831669, + 0.0073961834423244, + 0.007685815915465355, + 0.010908969677984715, + 0.013277022168040276, + 0.01433180458843708, + 0.017365068197250366, + 0.011624133214354515, + 0.17320027947425842, + 0.02190779522061348, + 0.14858226478099823, + 0.004729766398668289, + 0.005977087654173374, + 0.008635804988443851, + 0.010052686557173729, + 0.0077546206302940845, + 0.013236946426331997, + 0.00951748713850975, + 0.020631585270166397, + 0.003819876117631793, + 0.0045900787226855755, + 0.008300197310745716, + 0.009819860570132732, + 0.005459303967654705, + 0.006348594091832638, + 0.0060256244614720345, + 0.0046566142700612545, + 0.005286257714033127, + 0.011925771832466125, + 0.009447340853512287, + 0.01736830733716488, + 0.006839512847363949, + 0.012101486325263977, + 0.007840503938496113, + 0.007672272156924009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.009786011651158333 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248576868325472 + ], + "opposite_logprob_tokens": [ + -5.255248546600342 + ], + "logprob_reward": -0.005248576868325472, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248576868325472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04338199272751808, + 0.03846108913421631, + 0.06811010092496872, + 0.037867140024900436, + 0.0676143541932106, + 0.04305925592780113, + 0.5647341012954712, + 0.05953773856163025, + 0.033188264816999435, + 0.04404592514038086 + ], + "credit_map": [ + 0.010880755260586739, + 0.00317802163772285, + 0.003643198637291789, + 0.0030762224923819304, + 0.006119417026638985, + 0.0053843059577047825, + 0.00423828512430191, + 0.0068617919459939, + 0.006433016154915094, + 0.004856483545154333, + 0.004074496682733297, + 0.0032947654835879803, + 0.00631728395819664, + 0.0034589385613799095, + 0.004860529210418463, + 0.005165573675185442, + 0.010587336495518684, + 0.004389986861497164, + 0.0044851601123809814, + 0.005745352245867252, + 0.007124254945665598, + 0.015388916246592999, + 0.008613969199359417, + 0.0117751220241189, + 0.003027877537533641, + 0.0037648805882781744, + 0.0031997787300497293, + 0.00572483753785491, + 0.005729887168854475, + 0.005436984822154045, + 0.006185419857501984, + 0.004797469824552536, + 0.005632658954709768, + 0.005987145006656647, + 0.007040372584015131, + 0.009091484360396862, + 0.007819597609341145, + 0.01032777689397335, + 0.011298997327685356, + 0.010416324250400066, + 0.0036275263410061598, + 0.00407627085223794, + 0.003485623048618436, + 0.008404663763940334, + 0.0034341055434197187, + 0.0039046118035912514, + 0.008756567724049091, + 0.007369887083768845, + 0.012676523067057133, + 0.01693609543144703, + 0.013425453566014767, + 0.02133391797542572, + 0.01612074114382267, + 0.2432943731546402, + 0.028925903141498566, + 0.21202105283737183, + 0.006898153573274612, + 0.004856547806411982, + 0.011417103931307793, + 0.00602384191006422, + 0.00552661158144474, + 0.009689689613878727, + 0.007811876013875008, + 0.007313912268728018, + 0.003380612703040242, + 0.0041604782454669476, + 0.0033462275750935078, + 0.0036248688120394945, + 0.005713526625186205, + 0.0036781502421945333, + 0.005040150135755539, + 0.004244247451424599, + 0.003173466771841049, + 0.006325033959001303, + 0.005826437380164862, + 0.010001384653151035, + 0.0028838657308369875, + 0.005597714800387621, + 0.004676766227930784, + 0.005561255384236574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.005248576868325472 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7577431201934814 + ], + "opposite_logprob_tokens": [ + -0.6327431201934814 + ], + "logprob_reward": -0.7577431201934814, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7577431201934814, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04918202757835388, + 0.036611203104257584, + 0.05586511641740799, + 0.03407290577888489, + 0.07369382679462433, + 0.03594152256846428, + 0.5869328379631042, + 0.04807419702410698, + 0.04980260506272316, + 0.0298237893730402 + ], + "credit_map": [ + 0.006715708877891302, + 0.0035832433495670557, + 0.00529001047834754, + 0.00874712597578764, + 0.004446856211870909, + 0.003941886592656374, + 0.003998213913291693, + 0.012458981946110725, + 0.005618591792881489, + 0.0037838274147361517, + 0.003967783413827419, + 0.0036826191935688257, + 0.004999575670808554, + 0.004828850273042917, + 0.004110043402761221, + 0.005619911011308432, + 0.009760207496583462, + 0.0032516997307538986, + 0.005337569396942854, + 0.005526382476091385, + 0.005701783113181591, + 0.01204803679138422, + 0.0064097256399691105, + 0.007829711772501469, + 0.0029612050857394934, + 0.0035190267954021692, + 0.003208465874195099, + 0.004909950774163008, + 0.004369986243546009, + 0.005218921694904566, + 0.005335832480341196, + 0.004549514967948198, + 0.006003346759825945, + 0.008952365256845951, + 0.010569537989795208, + 0.008277620188891888, + 0.010640869848430157, + 0.008178898133337498, + 0.009773604571819305, + 0.01129758171737194, + 0.0031251143664121628, + 0.0031964892987161875, + 0.0033988968934863806, + 0.0065573533065617085, + 0.003711910452693701, + 0.004251393489539623, + 0.006042215973138809, + 0.00565814645960927, + 0.012517672032117844, + 0.01708117127418518, + 0.013690580613911152, + 0.02290652133524418, + 0.014793419279158115, + 0.25490593910217285, + 0.029611382633447647, + 0.22142617404460907, + 0.005691523663699627, + 0.0042742411606013775, + 0.009831946343183517, + 0.0049506183713674545, + 0.003934752661734819, + 0.007653726730495691, + 0.004877147264778614, + 0.006860238965600729, + 0.003952730447053909, + 0.005378889851272106, + 0.005985993891954422, + 0.00638619577512145, + 0.00781427975744009, + 0.005556964315474033, + 0.007917986251413822, + 0.006809566635638475, + 0.0020711144898086786, + 0.003579219337552786, + 0.004239496309310198, + 0.0060124583542346954, + 0.0021198850590735674, + 0.00383745227009058, + 0.0026341727934777737, + 0.005329990293830633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.7577431201934814 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "seamless 360-degree tracking close-up unfolds smoothly around a curious sparrow perched on a weathered, ancient stone wall. The sparrow, small and plump, with feathers a complex interweaving of earthy browns, soft grays, and sharp black stripes, tilts its head from time to time, its dark, shiny eyes gleaming with inquiry. The wall it stands on is itself a tapestry of natural textures, composed of irregularly shaped stones. Vast swathes of vibrant green and deep emerald moss cling tenaciously to the rough surface, while clumps of delicate ferns sprout from the damp crevices. Steadying itself on its delicate legs and claws, the sparrow executes a series of quick, light hops along the wall's uneven top, pausing intermittently to delicately peck at a patch of moss with its tiny beak or peer curiously into a deep crevice between the stones, completely absorbed in its exploration. As the camera continues to rotate smoothly, soft, diffused sunlight is evenly distributed across the entire scene, constantly revealing new angles and microscopic details of the stones and moss, creating a dynamic yet tranquil background. The entire image is filled with a peaceful, natural atmosphere, capturing an intimate and fascinating moment of pure beauty and curiosity in nature.", + "seed": 11716, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_9/index_3875_seed_11716_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09948384668678045, + "vlm_logprob": -0.09948384668678045, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.09948384668678045, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010116138495504856 + ], + "opposite_logprob_tokens": [ + -4.885116100311279 + ], + "logprob_reward": -0.010116138495504856, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010116138495504856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09477904438972473, + 0.17917540669441223, + 0.14931626617908478, + 0.08161178231239319, + 0.07500360906124115, + 0.03231009840965271, + 0.06928077340126038, + 0.015437477268278599, + 0.2702987492084503, + 0.032786775380373 + ], + "credit_map": [ + 0.008185075595974922, + 0.012235566973686218, + 0.009906581602990627, + 0.015880122780799866, + 0.012868586927652359, + 0.012884917668998241, + 0.013635986484587193, + 0.009182203561067581, + 0.02303141914308071, + 0.034010495990514755, + 0.015479620546102524, + 0.027775850147008896, + 0.023525912314653397, + 0.022393710911273956, + 0.020774811506271362, + 0.012183582410216331, + 0.009081555530428886, + 0.007879436947405338, + 0.007256664801388979, + 0.043311044573783875, + 0.013966490514576435, + 0.057004664093256, + 0.0056212968192994595, + 0.0051951054483652115, + 0.012000974267721176, + 0.010849887505173683, + 0.006201650947332382, + 0.0076800258830189705, + 0.0198417529463768, + 0.009596656076610088, + 0.007749056909233332, + 0.007691775914281607, + 0.005040268879383802, + 0.0070871044881641865, + 0.017841974273324013, + 0.004511063452810049, + 0.027291905134916306, + 0.007526008877903223, + 0.002858106978237629, + 0.0028471697587519884, + 0.004905750043690205, + 0.006661861669272184, + 0.006301684305071831, + 0.003214962314814329, + 0.002915490185841918, + 0.0030305793043226004, + 0.002847569528967142, + 0.0024321977980434895, + 0.004642792046070099, + 0.011463623493909836, + 0.012253561057150364, + 0.012091198936104774, + 0.0074141258373856544, + 0.008550699800252914, + 0.005777696147561073, + 0.007087070029228926, + 0.00157118309289217, + 0.001775891287252307, + 0.0028173206374049187, + 0.0020156721584498882, + 0.0023668771609663963, + 0.00205309153534472, + 0.0015050345100462437, + 0.0013324066530913115, + 0.017382225021719933, + 0.007178930100053549, + 0.006538779009133577, + 0.005146585870534182, + 0.11417026072740555, + 0.10423894226551056, + 0.006716172676533461, + 0.008926844224333763, + 0.003178249578922987, + 0.004299079068005085, + 0.0042700376361608505, + 0.0039383224211633205, + 0.008271068334579468, + 0.0038559504318982363, + 0.0024635735899209976, + 0.002510493155568838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 2 + ], + "selected_reward": -0.010116138495504856 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463874451816082 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.00463874451816082, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463874451816082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0786602720618248, + 0.03651563450694084, + 0.12377448379993439, + 0.07664282619953156, + 0.09006982296705246, + 0.07136892527341843, + 0.09850970655679703, + 0.025842992588877678, + 0.2599409520626068, + 0.13867439329624176 + ], + "credit_map": [ + 0.005387631710618734, + 0.008288929238915443, + 0.009044752456247807, + 0.011346864514052868, + 0.013582608662545681, + 0.01209009438753128, + 0.011133641935884953, + 0.007785753812640905, + 0.0039224461652338505, + 0.004900532774627209, + 0.006053783465176821, + 0.006071137264370918, + 0.003528493456542492, + 0.0047541530802845955, + 0.003763437271118164, + 0.0035216519609093666, + 0.007739923428744078, + 0.006878757383674383, + 0.006237855181097984, + 0.0370793342590332, + 0.010229626670479774, + 0.04464602842926979, + 0.004101543687283993, + 0.006861400790512562, + 0.009043607860803604, + 0.009154586121439934, + 0.01620480790734291, + 0.008741546422243118, + 0.007470844313502312, + 0.011636180803179741, + 0.007574758492410183, + 0.006816491950303316, + 0.0061837234534323215, + 0.009633747860789299, + 0.019653910771012306, + 0.005802534986287355, + 0.03298736363649368, + 0.008878307417035103, + 0.0031337463296949863, + 0.0037964903749525547, + 0.01005922257900238, + 0.01520050410181284, + 0.009702644310891628, + 0.0119889285415411, + 0.006481877528131008, + 0.006340047810226679, + 0.006695163901895285, + 0.004900544881820679, + 0.006225870922207832, + 0.016086837276816368, + 0.017047075554728508, + 0.017724376171827316, + 0.010707752779126167, + 0.01170111820101738, + 0.00853896513581276, + 0.010477706789970398, + 0.002333075739443302, + 0.0028113124426454306, + 0.004022006876766682, + 0.0038417447358369827, + 0.004812938626855612, + 0.0030487957410514355, + 0.002373799216002226, + 0.0025993199087679386, + 0.014867456629872322, + 0.0067828199826180935, + 0.008839783258736134, + 0.004958559293299913, + 0.1057552918791771, + 0.10358553379774094, + 0.006931188516318798, + 0.0082203084602952, + 0.010977636091411114, + 0.0092869121581316, + 0.03207174316048622, + 0.024054666981101036, + 0.016637489199638367, + 0.018405165523290634, + 0.013994230888783932, + 0.013246556743979454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.00463874451816082 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031916177831590176 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.0031916177831590176, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031916177831590176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09583692997694016, + 0.07459823042154312, + 0.04638452082872391, + 0.09694328904151917, + 0.28628623485565186, + 0.07787785679101944, + 0.14674854278564453, + 0.04570877179503441, + 0.058747097849845886, + 0.07086855173110962 + ], + "credit_map": [ + 0.007469751872122288, + 0.007140525616705418, + 0.014723890461027622, + 0.009820343926548958, + 0.01724839396774769, + 0.016482925042510033, + 0.011016777716577053, + 0.011934321373701096, + 0.0072476002387702465, + 0.012151326052844524, + 0.009234443306922913, + 0.015960415825247765, + 0.006665172055363655, + 0.00780856516212225, + 0.006381046958267689, + 0.009149659425020218, + 0.0032852410804480314, + 0.006980272941291332, + 0.006665857974439859, + 0.005421793553978205, + 0.005627155769616365, + 0.009662264958024025, + 0.004433928988873959, + 0.004308005794882774, + 0.01375780813395977, + 0.011181404814124107, + 0.02120250277221203, + 0.012075986713171005, + 0.011482222937047482, + 0.011760086752474308, + 0.00820080190896988, + 0.007282466161996126, + 0.02053828537464142, + 0.03945949301123619, + 0.06218086555600166, + 0.017456047236919403, + 0.09704630076885223, + 0.024991851300001144, + 0.010619515553116798, + 0.013993860222399235, + 0.010629763826727867, + 0.012864775024354458, + 0.013893247582018375, + 0.007442008703947067, + 0.008467566221952438, + 0.00836595892906189, + 0.008693209849298, + 0.007521322928369045, + 0.009435350075364113, + 0.02103152498602867, + 0.0249797273427248, + 0.026452170684933662, + 0.016798067837953568, + 0.016633495688438416, + 0.01232909131795168, + 0.019089117646217346, + 0.003783351741731167, + 0.0061937314458191395, + 0.004498072434216738, + 0.004092354793101549, + 0.009813368320465088, + 0.009835651144385338, + 0.004097695462405682, + 0.003394543658941984, + 0.007274878676980734, + 0.0031100548803806305, + 0.0037862590979784727, + 0.004347680602222681, + 0.020418629050254822, + 0.010332032106816769, + 0.0036732966545969248, + 0.005804268643260002, + 0.008692207746207714, + 0.005900735035538673, + 0.013584519736468792, + 0.0063117449171841145, + 0.012924927286803722, + 0.009058563970029354, + 0.005265322979539633, + 0.009130534715950489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.0031916177831590176 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052475095726549625 + ], + "opposite_logprob_tokens": [ + -5.255247592926025 + ], + "logprob_reward": -0.0052475095726549625, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052475095726549625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08515740931034088, + 0.03738672658801079, + 0.29014548659324646, + 0.04495171085000038, + 0.09185054898262024, + 0.04383786395192146, + 0.06927056610584259, + 0.030102696269750595, + 0.2830319106578827, + 0.024265063926577568 + ], + "credit_map": [ + 0.00633947504684329, + 0.010794202797114849, + 0.00933685153722763, + 0.01211517583578825, + 0.01784122921526432, + 0.012587126344442368, + 0.009168860502541065, + 0.00697448942810297, + 0.005657501053065062, + 0.006066369358450174, + 0.005069925915449858, + 0.004578359890729189, + 0.00435446435585618, + 0.003926942124962807, + 0.004308817442506552, + 0.00342435110360384, + 0.018390119075775146, + 0.013533203862607479, + 0.009463892318308353, + 0.09137745946645737, + 0.02454565092921257, + 0.11136758327484131, + 0.008213900960981846, + 0.01325368881225586, + 0.007238902151584625, + 0.006145336199551821, + 0.004158548079431057, + 0.0050635868683457375, + 0.00880453921854496, + 0.005090399645268917, + 0.003829692490398884, + 0.0046207113191485405, + 0.0063642761670053005, + 0.007487023714929819, + 0.021403316408395767, + 0.005745430011302233, + 0.03381148725748062, + 0.009646248072385788, + 0.0035426386166363955, + 0.00385013734921813, + 0.004874380771070719, + 0.005078244023025036, + 0.00782647542655468, + 0.00718735670670867, + 0.005497820675373077, + 0.0040330118499696255, + 0.0042697107419371605, + 0.0050708698108792305, + 0.004609901458024979, + 0.009721812792122364, + 0.012929817661643028, + 0.012314998544752598, + 0.007559130899608135, + 0.008027622476220131, + 0.006098973099142313, + 0.008008311502635479, + 0.002610255731269717, + 0.00393071910366416, + 0.005137280561029911, + 0.004698371514678001, + 0.003515715943649411, + 0.004135418683290482, + 0.003355764551088214, + 0.0027191736735403538, + 0.01610873080790043, + 0.0072752502746880054, + 0.008938999846577644, + 0.006031371653079987, + 0.11414292454719543, + 0.1136334016919136, + 0.007645021192729473, + 0.009256245568394661, + 0.00210683885961771, + 0.0034661751706153154, + 0.0032530229073017836, + 0.0017328535905107856, + 0.0071351248770952225, + 0.002690942259505391, + 0.002166061196476221, + 0.001714045531116426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.0052475095726549625 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742252230644226 + ], + "opposite_logprob_tokens": [ + -0.9742252230644226 + ], + "logprob_reward": -0.4742252230644226, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742252230644226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06542840600013733, + 0.04115116596221924, + 0.2681952714920044, + 0.0670996606349945, + 0.053494181483983994, + 0.039492715150117874, + 0.02032713033258915, + 0.013070348650217056, + 0.4091840386390686, + 0.022557061165571213 + ], + "credit_map": [ + 0.004996159579604864, + 0.00674604345113039, + 0.007965841330587864, + 0.009215138852596283, + 0.011618779972195625, + 0.009360176511108875, + 0.008794588968157768, + 0.006731676869094372, + 0.0058115697465837, + 0.007086017169058323, + 0.005851464346051216, + 0.00483767781406641, + 0.0046205781400203705, + 0.004387629684060812, + 0.004647566005587578, + 0.003908663522452116, + 0.017390165477991104, + 0.013860261999070644, + 0.009556732140481472, + 0.08360441029071808, + 0.022053539752960205, + 0.1014401763677597, + 0.007501722779124975, + 0.012788270600140095, + 0.009276211261749268, + 0.008221684023737907, + 0.011952279135584831, + 0.009256730787456036, + 0.010819475166499615, + 0.006595090497285128, + 0.005947399418801069, + 0.005030792206525803, + 0.004425013903528452, + 0.003734870348125696, + 0.012026221491396427, + 0.0027960899751633406, + 0.020167410373687744, + 0.00626749824732542, + 0.002030725125223398, + 0.002046352718025446, + 0.005028968211263418, + 0.00850458163768053, + 0.004340244457125664, + 0.007573992479592562, + 0.004400112200528383, + 0.003498126519843936, + 0.002983083948493004, + 0.0031636045314371586, + 0.0016438611783087254, + 0.0037262141704559326, + 0.00219042063690722, + 0.0031719240359961987, + 0.00312623823992908, + 0.0031659321393817663, + 0.0018679824424907565, + 0.0014345587696880102, + 0.0008850928861647844, + 0.0013556970516219735, + 0.002252999460324645, + 0.002146817510947585, + 0.002278219675645232, + 0.0016697176033630967, + 0.0013751956867054105, + 0.0011066087754443288, + 0.021372294053435326, + 0.008983817882835865, + 0.015039089135825634, + 0.007420724257826805, + 0.150255486369133, + 0.1748121678829193, + 0.008858545683324337, + 0.02244190312922001, + 0.0027414800133556128, + 0.004729339387267828, + 0.0031026708893477917, + 0.0018312976462766528, + 0.0029052142053842545, + 0.0028625410050153732, + 0.002144269645214081, + 0.00224024779163301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.4742252230644226 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "seamless 360-degree tracking close-up unfolds smoothly around a curious sparrow perched on a weathered, ancient stone wall. The sparrow, small and plump, with feathers a complex interweaving of earthy browns, soft grays, and sharp black stripes, tilts its head from time to time, its dark, shiny eyes gleaming with inquiry. The wall it stands on is itself a tapestry of natural textures, composed of irregularly shaped stones. Vast swathes of vibrant green and deep emerald moss cling tenaciously to the rough surface, while clumps of delicate ferns sprout from the damp crevices. Steadying itself on its delicate legs and claws, the sparrow executes a series of quick, light hops along the wall's uneven top, pausing intermittently to delicately peck at a patch of moss with its tiny beak or peer curiously into a deep crevice between the stones, completely absorbed in its exploration. As the camera continues to rotate smoothly, soft, diffused sunlight is evenly distributed across the entire scene, constantly revealing new angles and microscopic details of the stones and moss, creating a dynamic yet tranquil background. The entire image is filled with a peaceful, natural atmosphere, capturing an intimate and fascinating moment of pure beauty and curiosity in nature.", + "seed": 11717, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000078/group_9/index_3875_seed_11717_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16347633306868375, + "vlm_logprob": -0.16347633306868375, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.16347633306868375, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003861472476273775 + ], + "opposite_logprob_tokens": [ + -6.878861427307129 + ], + "logprob_reward": -0.003861472476273775, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.003861472476273775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09420684725046158, + 0.1104603111743927, + 0.04137599468231201, + 0.3909808397293091, + 0.03444119170308113, + 0.08092278242111206, + 0.03913717344403267, + 0.07124438881874084, + 0.03917187452316284, + 0.09805857390165329 + ], + "credit_map": [ + 0.011257128790020943, + 0.011455395258963108, + 0.008408327586948872, + 0.014290845021605492, + 0.01936061680316925, + 0.009693311527371407, + 0.009089640341699123, + 0.010651580058038235, + 0.012575656175613403, + 0.011242546141147614, + 0.021452555432915688, + 0.013866081833839417, + 0.017016949132084846, + 0.01023236196488142, + 0.014177207835018635, + 0.009896948002278805, + 0.003694358514621854, + 0.004439583979547024, + 0.006523348391056061, + 0.006933138240128756, + 0.006923007313162088, + 0.0026699083391577005, + 0.005737996194511652, + 0.004454655572772026, + 0.03656219318509102, + 0.03763682022690773, + 0.06398899853229523, + 0.050503119826316833, + 0.05813301354646683, + 0.05791532248258591, + 0.041205085813999176, + 0.04503626003861427, + 0.0027329055592417717, + 0.003128474112600088, + 0.005939847324043512, + 0.00556012662127614, + 0.0036080621648579836, + 0.004640805535018444, + 0.0044745649211108685, + 0.00435640336945653, + 0.007655220106244087, + 0.005489473231136799, + 0.017682645469903946, + 0.013127154670655727, + 0.007413215469568968, + 0.013354440219700336, + 0.009747459553182125, + 0.006453170906752348, + 0.004206143319606781, + 0.003428086405619979, + 0.006121654994785786, + 0.006149566266685724, + 0.004280883353203535, + 0.0073919580318033695, + 0.0029763570055365562, + 0.004582523833960295, + 0.009169456548988819, + 0.0071463496424257755, + 0.009333032183349133, + 0.012979034334421158, + 0.009426478296518326, + 0.007760144304484129, + 0.007923389784991741, + 0.0075065032579004765, + 0.005744370631873608, + 0.004602080676704645, + 0.007110437378287315, + 0.0048940181732177734, + 0.0044134133495390415, + 0.005513845477253199, + 0.0034095400478690863, + 0.003484168089926243, + 0.012062369845807552, + 0.01643584668636322, + 0.01256851851940155, + 0.012691214680671692, + 0.01591731794178486, + 0.009264999069273472, + 0.012098057195544243, + 0.007020241115242243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.003861472476273775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486539825797081 + ], + "opposite_logprob_tokens": [ + -3.048654079437256 + ], + "logprob_reward": -0.0486539825797081, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486539825797081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03840410336852074, + 0.025668762624263763, + 0.061779215931892395, + 0.6727914214134216, + 0.01621185429394245, + 0.028614304959774017, + 0.01492974441498518, + 0.049440328031778336, + 0.03947789594531059, + 0.052682340145111084 + ], + "credit_map": [ + 0.005414040293544531, + 0.0047216834500432014, + 0.004536597058176994, + 0.004082739353179932, + 0.005481484346091747, + 0.0036500971764326096, + 0.0048287189565598965, + 0.005688747391104698, + 0.0037030500825494528, + 0.0021951987873762846, + 0.004766370635479689, + 0.0038525769487023354, + 0.0025551316794008017, + 0.0020243017934262753, + 0.003368666395545006, + 0.003203469095751643, + 0.005750855430960655, + 0.006837622262537479, + 0.011440673843026161, + 0.008967303670942783, + 0.009177947416901588, + 0.004137452691793442, + 0.009507125243544579, + 0.005960241891443729, + 0.0635819137096405, + 0.06814426183700562, + 0.10385926067829132, + 0.07961047440767288, + 0.11193522065877914, + 0.085667185485363, + 0.07175067812204361, + 0.08824245631694794, + 0.0016480745980516076, + 0.0017971189226955175, + 0.0025476128794252872, + 0.002226513344794512, + 0.0010216837981715798, + 0.001747269881889224, + 0.0021580797620117664, + 0.0030655027367174625, + 0.0037928069941699505, + 0.001959742745384574, + 0.005946697201579809, + 0.004641317762434483, + 0.002312131691724062, + 0.004348997958004475, + 0.0034243417903780937, + 0.00218827067874372, + 0.0018371311016380787, + 0.0009448743076063693, + 0.002605853835120797, + 0.0013198499800637364, + 0.001017348957248032, + 0.0031125713139772415, + 0.0013097269693389535, + 0.002782389521598816, + 0.006366336718201637, + 0.005711179226636887, + 0.007939707487821579, + 0.008446627296507359, + 0.006091676652431488, + 0.005587237887084484, + 0.005194670055061579, + 0.004102899227291346, + 0.007100752089172602, + 0.004238767083734274, + 0.008263172581791878, + 0.003992985002696514, + 0.0026430159341543913, + 0.006503733340650797, + 0.0034243694972246885, + 0.003311101347208023, + 0.005827849265187979, + 0.007208510302007198, + 0.007839111611247063, + 0.010565286502242088, + 0.004527627490460873, + 0.004333734977990389, + 0.00933597981929779, + 0.00304424692876637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.0486539825797081 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024895647075027227 + ], + "opposite_logprob_tokens": [ + -6.0024895668029785 + ], + "logprob_reward": -0.0024895647075027227, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024895647075027227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09412769228219986, + 0.04479363188147545, + 0.13828621804714203, + 0.3629566431045532, + 0.04806351289153099, + 0.06520497798919678, + 0.05692397803068161, + 0.05282963439822197, + 0.08070562034845352, + 0.05610806122422218 + ], + "credit_map": [ + 0.013173224404454231, + 0.00840039923787117, + 0.011592534370720387, + 0.017872823402285576, + 0.011509493924677372, + 0.009504977613687515, + 0.008995897136628628, + 0.013078343123197556, + 0.007119312882423401, + 0.005745258182287216, + 0.00581079488620162, + 0.00733053358271718, + 0.004583694506436586, + 0.0034448641818016768, + 0.005825625732541084, + 0.004933546297252178, + 0.012604864314198494, + 0.014511694200336933, + 0.026617497205734253, + 0.014564314857125282, + 0.016939179971814156, + 0.00851431768387556, + 0.027650238946080208, + 0.01688411273062229, + 0.03471029922366142, + 0.03735687956213951, + 0.05300414189696312, + 0.04770462587475777, + 0.05714577063918114, + 0.04814179986715317, + 0.037207964807748795, + 0.047685157507658005, + 0.005621120799332857, + 0.006929873954504728, + 0.006267746444791555, + 0.007450254634022713, + 0.004174487199634314, + 0.005409485660493374, + 0.005669956561177969, + 0.006540588568896055, + 0.009223084896802902, + 0.00518931495025754, + 0.010573193430900574, + 0.011119506321847439, + 0.007452517747879028, + 0.00557085033506155, + 0.006592851597815752, + 0.009483655914664268, + 0.005994538776576519, + 0.00454355962574482, + 0.0077493502758443356, + 0.008407358080148697, + 0.003809915157034993, + 0.009201334789395332, + 0.005745359696447849, + 0.011472560465335846, + 0.0071535054594278336, + 0.0059858039021492004, + 0.009750842116773129, + 0.005759021732956171, + 0.007034145295619965, + 0.005933613982051611, + 0.006495124194771051, + 0.004717577714473009, + 0.01688309572637081, + 0.007643391378223896, + 0.019916296005249023, + 0.008071991614997387, + 0.004965899512171745, + 0.00901679415255785, + 0.0059483665972948074, + 0.0082597890868783, + 0.004877287428826094, + 0.006831145845353603, + 0.00866784155368805, + 0.010921511799097061, + 0.0052867066115140915, + 0.0057881902903318405, + 0.00975092500448227, + 0.003984448034316301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0024895647075027227 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633048549294472 + ], + "opposite_logprob_tokens": [ + -5.379632949829102 + ], + "logprob_reward": -0.004633048549294472, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633048549294472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06697361916303635, + 0.03286769986152649, + 0.14751115441322327, + 0.20321987569332123, + 0.04998010769486427, + 0.1583506464958191, + 0.05595404654741287, + 0.11451414227485657, + 0.13129840791225433, + 0.03933022916316986 + ], + "credit_map": [ + 0.009369109757244587, + 0.006352932658046484, + 0.009821904823184013, + 0.013254263438284397, + 0.005793726537376642, + 0.0065254876390099525, + 0.006250082049518824, + 0.009606109000742435, + 0.003826655214652419, + 0.0024421107955276966, + 0.006257242523133755, + 0.004865054041147232, + 0.0044553885236382484, + 0.0029823852237313986, + 0.004303861875087023, + 0.0037350032944232225, + 0.013282555155456066, + 0.015802232548594475, + 0.027180220931768417, + 0.01782369427382946, + 0.019393371418118477, + 0.010409487411379814, + 0.025596225634217262, + 0.018023362383246422, + 0.017441492527723312, + 0.019042545929551125, + 0.03433563932776451, + 0.03289968892931938, + 0.017126627266407013, + 0.04081794247031212, + 0.02066114731132984, + 0.020894793793559074, + 0.005134681239724159, + 0.00471080606803298, + 0.00818695966154337, + 0.006759084761142731, + 0.005242917686700821, + 0.0062425886280834675, + 0.006440155673772097, + 0.007262916304171085, + 0.020412147045135498, + 0.012321313843131065, + 0.03044755756855011, + 0.02522953227162361, + 0.013627911917865276, + 0.02293364331126213, + 0.018439721316099167, + 0.014938837848603725, + 0.00813334435224533, + 0.004848468117415905, + 0.010403567925095558, + 0.0062050968408584595, + 0.005887703970074654, + 0.008696641772985458, + 0.004380639176815748, + 0.007398584391921759, + 0.015416538342833519, + 0.012325250543653965, + 0.014106262475252151, + 0.02093726582825184, + 0.0157022587954998, + 0.013539595529437065, + 0.012114708311855793, + 0.010372262448072433, + 0.024841463193297386, + 0.01145431399345398, + 0.031763020902872086, + 0.01205615233629942, + 0.008882452733814716, + 0.014683393761515617, + 0.009251861833035946, + 0.01836574450135231, + 0.005876400973647833, + 0.00560725899413228, + 0.0059370058588683605, + 0.005587663967162371, + 0.004225133918225765, + 0.003902009455487132, + 0.00477914372459054, + 0.0034156115725636482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.004633048549294472 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small, plump sparrow with earthy brown, soft gray, and black striped feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sparrow perched on and hopping along a weathered, ancient stone wall composed of irregularly shaped stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the stone wall and crevices covered with vibrant green and deep emerald moss and delicate ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow perform actions such as tilting its head, hopping, pecking at moss, and peering into crevices?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth 360-degree tracking rotation around the sparrow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7577435970306396 + ], + "opposite_logprob_tokens": [ + -0.6327435970306396 + ], + "logprob_reward": -0.7577435970306396, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7577435970306396, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05248529464006424, + 0.05777518078684807, + 0.057035066187381744, + 0.24920418858528137, + 0.03845343738794327, + 0.16071584820747375, + 0.09368646889925003, + 0.10987164080142975, + 0.14242425560951233, + 0.038348592817783356 + ], + "credit_map": [ + 0.007760908454656601, + 0.005754445679485798, + 0.008892012760043144, + 0.008130171336233616, + 0.006806388963013887, + 0.004288962576538324, + 0.004622752778232098, + 0.006229651626199484, + 0.006835532374680042, + 0.006834729108959436, + 0.010299176909029484, + 0.007673710584640503, + 0.005940950475633144, + 0.004607210401445627, + 0.008167608641088009, + 0.007416261825710535, + 0.006580763962119818, + 0.008144601248204708, + 0.007921758107841015, + 0.0044450280256569386, + 0.013059121556580067, + 0.003841337049379945, + 0.0068297116085886955, + 0.006212740205228329, + 0.027314642444252968, + 0.029546305537223816, + 0.030150795355439186, + 0.026801330968737602, + 0.0454055555164814, + 0.034514158964157104, + 0.02663448266685009, + 0.028836918994784355, + 0.0039521255530416965, + 0.0034210020676255226, + 0.007430896162986755, + 0.006290087476372719, + 0.003084154101088643, + 0.004712671972811222, + 0.004046243615448475, + 0.00551625806838274, + 0.021333977580070496, + 0.0128356097266078, + 0.034702327102422714, + 0.023052578791975975, + 0.013002507388591766, + 0.026720160618424416, + 0.014879284426569939, + 0.014189404435455799, + 0.011554638855159283, + 0.008451213128864765, + 0.016605883836746216, + 0.016983263194561005, + 0.009225165471434593, + 0.015842819586396217, + 0.006630564574152231, + 0.00839291326701641, + 0.015163865871727467, + 0.01052833255380392, + 0.013202156871557236, + 0.0208621583878994, + 0.015956593677401543, + 0.012414527125656605, + 0.012113428674638271, + 0.009630570188164711, + 0.028399406000971794, + 0.011688965372741222, + 0.038499027490615845, + 0.011450833640992641, + 0.011562615633010864, + 0.013789738528430462, + 0.00937049649655819, + 0.01766318269073963, + 0.004885730799287558, + 0.0052735102362930775, + 0.007883096113801003, + 0.00400113919749856, + 0.004326146561652422, + 0.003689057659357786, + 0.005360059905797243, + 0.0029298511799424887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.7577435970306396 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_0/index_tensor([8016, 8016])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_0/index_tensor([8016, 8016])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a52029f5c9f818ba4791b5946d323f81468e0326 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_0/index_tensor([8016, 8016])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "first-person perspective shot from inside a rocking carriage, moving forward at eye level, perfectly simulates the tangible sensation of a bumpy ride. The camera gently sways and pitches in time with the carriage's rhythm. Peering out through the rustic wooden window frames, one sees the narrow streets of an ancient village, paved entirely with weathered cobblestones of varying sizes. The streets are lined with a patchwork of wooden and stone houses, their walls covered in deep green moss. Ancient tiled roofs are stained by age, and occasionally wisps of smoke can be seen rising from stone chimneys. The warm, golden afternoon sun slants across the scene, casting a soft glow over the scene, lengthening the shadows of the buildings and making even tiny particles of dust in the air visible. The entire picture presents a remarkable retro film texture, with warm and saturated colors, obvious graininess and occasional slight lens flare caused by direct sunlight. As the carriage slowly moves forward, the scenery in front of you rhythmically rises and falls, creating a tranquil, nostalgic and immersive experience, as if a dusty memory is slowly being awakened.", + "seed": 1341, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_0/index_8016_seed_1341_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0968420366756618, + "vlm_logprob": -0.0968420366756618, + "vlm_yes_no_margin": 2.924999976158142, + "vlm_token_credit": -0.0968420366756618, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08655781298875809 + ], + "opposite_logprob_tokens": [ + -2.5865578651428223 + ], + "logprob_reward": -0.08655781298875809, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08655781298875809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09832671284675598, + 0.15583214163780212, + 0.10205428302288055, + 0.04835965111851692, + 0.09769264608621597, + 0.09893426299095154, + 0.07147075235843658, + 0.059908948838710785, + 0.0775771290063858, + 0.18984349071979523 + ], + "credit_map": [ + 0.0106400391086936, + 0.014980164356529713, + 0.011476966552436352, + 0.01689937524497509, + 0.009733295999467373, + 0.009340017102658749, + 0.013088220730423927, + 0.012168625369668007, + 0.036552973091602325, + 0.021726904436945915, + 0.017321007326245308, + 0.04409682750701904, + 0.006170788314193487, + 0.006383200641721487, + 0.013514354825019836, + 0.010066078044474125, + 0.015577258542180061, + 0.013290961273014545, + 0.014287276193499565, + 0.015404041856527328, + 0.009412125684320927, + 0.009911426343023777, + 0.016000665724277496, + 0.008170515298843384, + 0.009158114902675152, + 0.0075891283340752125, + 0.011415605433285236, + 0.005921080242842436, + 0.00485113263130188, + 0.0034246183931827545, + 0.0038571215700358152, + 0.002142840065062046, + 0.015148293226957321, + 0.013171550817787647, + 0.012637514621019363, + 0.011405730620026588, + 0.011217540130019188, + 0.0074990433640778065, + 0.017325038090348244, + 0.009287920780479908, + 0.016914645209908485, + 0.01220785640180111, + 0.015353899449110031, + 0.012986318208277225, + 0.013780876062810421, + 0.00809132494032383, + 0.012826199643313885, + 0.006773132365196943, + 0.007045927923172712, + 0.0062948851846158504, + 0.007977047935128212, + 0.010201363824307919, + 0.008388862013816833, + 0.013019908219575882, + 0.010330481454730034, + 0.008212265558540821, + 0.0044428338296711445, + 0.00353953056037426, + 0.011480595916509628, + 0.010090189054608345, + 0.006825178861618042, + 0.006490461062639952, + 0.008920707739889622, + 0.008119448088109493, + 0.0036268839612603188, + 0.005502151325345039, + 0.008939467370510101, + 0.018765365704894066, + 0.004242671187967062, + 0.004360745660960674, + 0.021186336874961853, + 0.010953499935567379, + 0.015727700665593147, + 0.018209995701909065, + 0.026472942903637886, + 0.026369484141469002, + 0.024451272562146187, + 0.02179819345474243, + 0.03824763372540474, + 0.018566232174634933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.08655781298875809 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005975240375846624 + ], + "opposite_logprob_tokens": [ + -5.130975246429443 + ], + "logprob_reward": -0.005975240375846624, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005975240375846624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14473964273929596, + 0.11641979217529297, + 0.14574512839317322, + 0.049592554569244385, + 0.09773264080286026, + 0.15730051696300507, + 0.037327900528907776, + 0.026179445907473564, + 0.07214757055044174, + 0.1528148502111435 + ], + "credit_map": [ + 0.012410224415361881, + 0.021999100223183632, + 0.017744801938533783, + 0.019191423431038857, + 0.012929156422615051, + 0.012948471121490002, + 0.03015589900314808, + 0.017360564321279526, + 0.02882578782737255, + 0.015469846315681934, + 0.015403348952531815, + 0.028605971485376358, + 0.004752117209136486, + 0.007001696620136499, + 0.008492791093885899, + 0.007868234068155289, + 0.02389690838754177, + 0.017796026542782784, + 0.020386453717947006, + 0.021557144820690155, + 0.011964823119342327, + 0.01755097135901451, + 0.021096883341670036, + 0.011495922692120075, + 0.008507967926561832, + 0.005555297248065472, + 0.007190349977463484, + 0.006296234671026468, + 0.006056445650756359, + 0.007009137887507677, + 0.0058225300163030624, + 0.0031545928213745356, + 0.019244030117988586, + 0.012699595652520657, + 0.01230909675359726, + 0.010771387256681919, + 0.007106989622116089, + 0.007222356740385294, + 0.020052731037139893, + 0.008326451294124126, + 0.033025339245796204, + 0.01945771463215351, + 0.02811523526906967, + 0.026031844317913055, + 0.012783394195139408, + 0.01239713840186596, + 0.015165829099714756, + 0.010324030183255672, + 0.004409186542034149, + 0.0029871182050555944, + 0.006233828607946634, + 0.005475016310811043, + 0.0042809732258319855, + 0.005408360622823238, + 0.004908079747110605, + 0.0036253370344638824, + 0.00219513732008636, + 0.002944972598925233, + 0.00545487878844142, + 0.004512817598879337, + 0.003238873789086938, + 0.0025727148167788982, + 0.0023581215646117926, + 0.0029019303619861603, + 0.0029889268334954977, + 0.0064943572506308556, + 0.007788797374814749, + 0.013815298676490784, + 0.006070054601877928, + 0.005655850749462843, + 0.017695503309369087, + 0.011638778261840343, + 0.01218979712575674, + 0.013854301534593105, + 0.021935490891337395, + 0.02112807333469391, + 0.01862545870244503, + 0.019194986671209335, + 0.02952227182686329, + 0.016364458948373795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.005975240375846624 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279921352863312 + ], + "opposite_logprob_tokens": [ + -2.017799139022827 + ], + "logprob_reward": -0.14279921352863312, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14279921352863312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12729045748710632, + 0.13604223728179932, + 0.29923805594444275, + 0.06294766813516617, + 0.1044234037399292, + 0.08041637390851974, + 0.021062646061182022, + 0.025872787460684776, + 0.03066400997340679, + 0.11204231530427933 + ], + "credit_map": [ + 0.011923064477741718, + 0.021308014169335365, + 0.010273103602230549, + 0.023211779072880745, + 0.00974985770881176, + 0.013655349612236023, + 0.027668755501508713, + 0.00950053334236145, + 0.03721243143081665, + 0.020286958664655685, + 0.01556359976530075, + 0.03588174283504486, + 0.003792591392993927, + 0.00473268935456872, + 0.01258791796863079, + 0.005984294693917036, + 0.04963405802845955, + 0.04339584708213806, + 0.040324434638023376, + 0.043117303401231766, + 0.0235364381223917, + 0.03960781544446945, + 0.035948269069194794, + 0.023673901334404945, + 0.011082901619374752, + 0.007357473485171795, + 0.012748096138238907, + 0.0097552090883255, + 0.0066802590154111385, + 0.005634878296405077, + 0.006552462931722403, + 0.0031363843008875847, + 0.02948646806180477, + 0.013658049516379833, + 0.014672686345875263, + 0.01092531904578209, + 0.006959118414670229, + 0.009437273256480694, + 0.012975756078958511, + 0.006308733485639095, + 0.015202213078737259, + 0.01139969564974308, + 0.016286524012684822, + 0.014284777455031872, + 0.005478190258145332, + 0.005322595592588186, + 0.007220371160656214, + 0.005222007166594267, + 0.002377418102696538, + 0.0018484462052583694, + 0.003758738050237298, + 0.004098726436495781, + 0.0025615496560931206, + 0.001776959397830069, + 0.00308053195476532, + 0.0015602753264829516, + 0.0024008939508348703, + 0.0030927429907023907, + 0.005052548833191395, + 0.005234111566096544, + 0.002722277771681547, + 0.0019317042315378785, + 0.002642980543896556, + 0.0027955251280218363, + 0.0023195352405309677, + 0.003885390469804406, + 0.005082567688077688, + 0.004907355178147554, + 0.0030906323809176683, + 0.004122327547520399, + 0.0031210461165755987, + 0.004135152790695429, + 0.007728857919573784, + 0.010064991191029549, + 0.015620582737028599, + 0.015809539705514908, + 0.014224703423678875, + 0.016085023060441017, + 0.021651314571499825, + 0.010857302695512772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.14279921352863312 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255936861038208 + ], + "opposite_logprob_tokens": [ + -1.6005936861038208 + ], + "logprob_reward": -0.2255936861038208, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255936861038208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12837974727153778, + 0.19373765587806702, + 0.07334364950656891, + 0.1862616389989853, + 0.18547941744327545, + 0.06116991862654686, + 0.02550012245774269, + 0.03356289491057396, + 0.05240153893828392, + 0.06016339361667633 + ], + "credit_map": [ + 0.01922118477523327, + 0.021633697673678398, + 0.01786663569509983, + 0.022639678791165352, + 0.008455001749098301, + 0.009279177524149418, + 0.01970524713397026, + 0.009579140692949295, + 0.05079089477658272, + 0.028135187923908234, + 0.02194950543344021, + 0.04766552522778511, + 0.007286077365279198, + 0.007044742349535227, + 0.019843382760882378, + 0.011022360064089298, + 0.013437055051326752, + 0.014727316796779633, + 0.010248269885778427, + 0.00984890665858984, + 0.006032815668731928, + 0.007049920037388802, + 0.00631480198353529, + 0.005684567149728537, + 0.030473463237285614, + 0.028163105249404907, + 0.04128067567944527, + 0.02248190902173519, + 0.019583959132432938, + 0.016943350434303284, + 0.017651651054620743, + 0.009683527983725071, + 0.05042204633355141, + 0.028308968991041183, + 0.026518993079662323, + 0.022328879684209824, + 0.009120102971792221, + 0.014941123314201832, + 0.022973772138357162, + 0.010865564458072186, + 0.010606564581394196, + 0.008146182633936405, + 0.011207003146409988, + 0.010462632402777672, + 0.00489464309066534, + 0.003968055825680494, + 0.007383580785244703, + 0.0045012664049863815, + 0.004075853154063225, + 0.0016737484838813543, + 0.003977022599428892, + 0.003297345945611596, + 0.0030971188098192215, + 0.0027675735764205456, + 0.0030488476622849703, + 0.00356261502020061, + 0.0027164132334291935, + 0.004034795332700014, + 0.00732245622202754, + 0.005573086906224489, + 0.0031331211794167757, + 0.002690376015380025, + 0.004198479000478983, + 0.0038941737730056047, + 0.0033842353150248528, + 0.0072382246144115925, + 0.005969890858978033, + 0.008103526197373867, + 0.005675804801285267, + 0.007403762545436621, + 0.007981759496033192, + 0.006644337438046932, + 0.006282202433794737, + 0.011536956764757633, + 0.008215072564780712, + 0.007904134690761566, + 0.007853942923247814, + 0.005521418061107397, + 0.006873420439660549, + 0.005976255517452955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.2255936861038208 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328423038125038 + ], + "opposite_logprob_tokens": [ + -3.7732841968536377 + ], + "logprob_reward": -0.02328423038125038, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328423038125038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.128754660487175, + 0.13130617141723633, + 0.09127437323331833, + 0.058747921139001846, + 0.05556362867355347, + 0.12334389239549637, + 0.037743791937828064, + 0.07384144514799118, + 0.2516181468963623, + 0.047806013375520706 + ], + "credit_map": [ + 0.019339771941304207, + 0.0198488999158144, + 0.01161193661391735, + 0.017551269382238388, + 0.01266139093786478, + 0.012922623194754124, + 0.025705967098474503, + 0.00911280419677496, + 0.02452465519309044, + 0.020335102453827858, + 0.018001744523644447, + 0.03344973176717758, + 0.005602049175649881, + 0.007643021643161774, + 0.01279442198574543, + 0.008955446071922779, + 0.013332139700651169, + 0.010508066043257713, + 0.013799348846077919, + 0.012683326378464699, + 0.009326759725809097, + 0.009488594718277454, + 0.013633477501571178, + 0.008502664044499397, + 0.01042413990944624, + 0.007309918291866779, + 0.010583815164864063, + 0.007394448854029179, + 0.006540665868669748, + 0.005272991023957729, + 0.00709239998832345, + 0.004129544831812382, + 0.009139691479504108, + 0.00644255056977272, + 0.00722136814147234, + 0.005967767909169197, + 0.007259782403707504, + 0.005543830804526806, + 0.009609689004719257, + 0.004378945101052523, + 0.024844160303473473, + 0.017091939225792885, + 0.020207270979881287, + 0.020177317783236504, + 0.010094267316162586, + 0.009067834354937077, + 0.012622510083019733, + 0.0092385932803154, + 0.006285151932388544, + 0.002869111718609929, + 0.006537257693707943, + 0.007554495707154274, + 0.0029230425134301186, + 0.0047857314348220825, + 0.004091967362910509, + 0.0026970328763127327, + 0.005571577697992325, + 0.008039191365242004, + 0.012701391242444515, + 0.01048301625996828, + 0.006704109255224466, + 0.006979455705732107, + 0.007469737436622381, + 0.015892965719103813, + 0.01152171939611435, + 0.016183121129870415, + 0.032033342868089676, + 0.0696934387087822, + 0.013654910959303379, + 0.017397616058588028, + 0.06292755901813507, + 0.028206441551446915, + 0.004830455407500267, + 0.00896045845001936, + 0.006968473084270954, + 0.007263696752488613, + 0.005925280973315239, + 0.003996625542640686, + 0.007094604894518852, + 0.0027664138469845057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.02328423038125038 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "first-person perspective shot from inside a rocking carriage, moving forward at eye level, perfectly simulates the tangible sensation of a bumpy ride. The camera gently sways and pitches in time with the carriage's rhythm. Peering out through the rustic wooden window frames, one sees the narrow streets of an ancient village, paved entirely with weathered cobblestones of varying sizes. The streets are lined with a patchwork of wooden and stone houses, their walls covered in deep green moss. Ancient tiled roofs are stained by age, and occasionally wisps of smoke can be seen rising from stone chimneys. The warm, golden afternoon sun slants across the scene, casting a soft glow over the scene, lengthening the shadows of the buildings and making even tiny particles of dust in the air visible. The entire picture presents a remarkable retro film texture, with warm and saturated colors, obvious graininess and occasional slight lens flare caused by direct sunlight. As the carriage slowly moves forward, the scenery in front of you rhythmically rises and falls, creating a tranquil, nostalgic and immersive experience, as if a dusty memory is slowly being awakened.", + "seed": 1342, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_0/index_8016_seed_1342_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.038920373236760494, + "vlm_logprob": -0.038920373236760494, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.038920373236760494, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06088443472981453 + ], + "opposite_logprob_tokens": [ + -2.935884475708008 + ], + "logprob_reward": -0.06088443472981453, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06088443472981453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12158187478780746, + 0.15872839093208313, + 0.08145401626825333, + 0.08524172753095627, + 0.046480096876621246, + 0.10691145062446594, + 0.08056837320327759, + 0.16148295998573303, + 0.06439782679080963, + 0.09315331280231476 + ], + "credit_map": [ + 0.011488695628941059, + 0.01193263754248619, + 0.02044956386089325, + 0.01388414017856121, + 0.01896839402616024, + 0.014672728255391121, + 0.013130933046340942, + 0.017054783180356026, + 0.02852376364171505, + 0.0242075826972723, + 0.023029858246445656, + 0.01470754574984312, + 0.016288205981254578, + 0.0286838561296463, + 0.013554655015468597, + 0.009732907637953758, + 0.010082048363983631, + 0.011370137333869934, + 0.01804453134536743, + 0.012094993144273758, + 0.007220201659947634, + 0.00879833847284317, + 0.007122877053916454, + 0.006720885168761015, + 0.008236577734351158, + 0.010105359368026257, + 0.0076715280301868916, + 0.021439535543322563, + 0.010869824327528477, + 0.009495973587036133, + 0.007310314103960991, + 0.010112606920301914, + 0.003366864984855056, + 0.007891968823969364, + 0.006984657607972622, + 0.0052443877793848515, + 0.0052910977974534035, + 0.006657444406300783, + 0.006419171113520861, + 0.004624499008059502, + 0.01038991566747427, + 0.019694805145263672, + 0.014238690957427025, + 0.011788391508162022, + 0.013678298331797123, + 0.011285371147096157, + 0.01141449436545372, + 0.014421478845179081, + 0.005361203104257584, + 0.004778718575835228, + 0.013789902441203594, + 0.016603771597146988, + 0.005904956720769405, + 0.012422964908182621, + 0.010617333464324474, + 0.011089514940977097, + 0.011460020206868649, + 0.012230221182107925, + 0.03493543341755867, + 0.018272042274475098, + 0.012792483903467655, + 0.022397028282284737, + 0.012577018700540066, + 0.03681870549917221, + 0.004805129487067461, + 0.004900829866528511, + 0.006255710031837225, + 0.008199119009077549, + 0.00714531447738409, + 0.010039012879133224, + 0.008348933421075344, + 0.014703773893415928, + 0.0076514785178005695, + 0.008102331310510635, + 0.013356008566915989, + 0.013750825077295303, + 0.012129489332437515, + 0.01151463482528925, + 0.01642829366028309, + 0.010220248252153397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.06088443472981453 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002841603709384799 + ], + "opposite_logprob_tokens": [ + -5.877841472625732 + ], + "logprob_reward": -0.002841603709384799, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002841603709384799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07714159786701202, + 0.10526320338249207, + 0.05974726378917694, + 0.12819445133209229, + 0.0354582704603672, + 0.06552781909704208, + 0.06533101946115494, + 0.3175885081291199, + 0.10335462540388107, + 0.04239330813288689 + ], + "credit_map": [ + 0.0063261310569942, + 0.009187711402773857, + 0.008716479875147343, + 0.00893284473568201, + 0.008153199218213558, + 0.010384642519056797, + 0.009575004689395428, + 0.01586558111011982, + 0.01661531627178192, + 0.011111854575574398, + 0.007918083108961582, + 0.009846235625445843, + 0.007087971083819866, + 0.03063926473259926, + 0.015739290043711662, + 0.006305179558694363, + 0.008258476853370667, + 0.00878169946372509, + 0.013959013856947422, + 0.00872803758829832, + 0.005590290296822786, + 0.006172845605760813, + 0.004056262318044901, + 0.0042006284929811954, + 0.008812082931399345, + 0.010435453616082668, + 0.012731305323541164, + 0.04824739694595337, + 0.01457414124161005, + 0.0075514852069318295, + 0.008263719268143177, + 0.0175788514316082, + 0.002748955972492695, + 0.005022268742322922, + 0.004489988554269075, + 0.004124379716813564, + 0.0039774272590875626, + 0.005721122957766056, + 0.005368121434003115, + 0.004006003960967064, + 0.005185471847653389, + 0.00828295573592186, + 0.01144612766802311, + 0.011846442706882954, + 0.007625900208950043, + 0.008171130903065205, + 0.006925408262759447, + 0.006044378038495779, + 0.005972784478217363, + 0.006212071981281042, + 0.008237982168793678, + 0.008772392757236958, + 0.005049466621130705, + 0.01222880370914936, + 0.008454732596874237, + 0.010402779094874859, + 0.013695534318685532, + 0.006289773620665073, + 0.10877592116594315, + 0.016973871737718582, + 0.011367999017238617, + 0.0262079406529665, + 0.011932586319744587, + 0.12234483659267426, + 0.007208550814539194, + 0.006451573688536882, + 0.01090899296104908, + 0.012877385132014751, + 0.008616771548986435, + 0.015282717533409595, + 0.014604478143155575, + 0.027404146268963814, + 0.0031559208873659372, + 0.003125883871689439, + 0.006182973273098469, + 0.005652276799082756, + 0.00547122722491622, + 0.006097002420574427, + 0.00615491671487689, + 0.006553103215992451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.002841603709384799 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369913622736931 + ], + "opposite_logprob_tokens": [ + -3.4086990356445312 + ], + "logprob_reward": -0.03369913622736931, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369913622736931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08761647343635559, + 0.11921285837888718, + 0.09606438130140305, + 0.23832298815250397, + 0.026776764541864395, + 0.053592294454574585, + 0.06407792121171951, + 0.22140443325042725, + 0.05799917131662369, + 0.034932684153318405 + ], + "credit_map": [ + 0.0077701956033706665, + 0.009941566735506058, + 0.016669701784849167, + 0.01093397568911314, + 0.011194273829460144, + 0.011886048130691051, + 0.009017355740070343, + 0.010203367099165916, + 0.012752467766404152, + 0.01000052597373724, + 0.010927814990282059, + 0.013833907432854176, + 0.010552960447967052, + 0.034427884966135025, + 0.016265349462628365, + 0.010451965034008026, + 0.010673291981220245, + 0.01233229786157608, + 0.01957813836634159, + 0.0177135206758976, + 0.007769155781716108, + 0.007533255498856306, + 0.009159019216895103, + 0.011305712163448334, + 0.012876706197857857, + 0.014299227856099606, + 0.022897332906723022, + 0.09738673269748688, + 0.023356968536973, + 0.012079262174665928, + 0.017296597361564636, + 0.038130179047584534, + 0.002377965021878481, + 0.00321379816159606, + 0.0029552148189395666, + 0.005150866694748402, + 0.002534502884373069, + 0.0035163795109838247, + 0.003076045075431466, + 0.003951996099203825, + 0.003866808023303747, + 0.00789555348455906, + 0.00864592008292675, + 0.0068445028737187386, + 0.0074712191708385944, + 0.005546056665480137, + 0.0071326144970953465, + 0.006189622916281223, + 0.004395333584398031, + 0.005502307787537575, + 0.007328172214329243, + 0.01001584343612194, + 0.0047469609417021275, + 0.01204700767993927, + 0.007632649037986994, + 0.012409654445946217, + 0.009461253881454468, + 0.007888106629252434, + 0.0647805705666542, + 0.01667795330286026, + 0.012018164619803429, + 0.027492348104715347, + 0.013666993007063866, + 0.06941907107830048, + 0.004727769643068314, + 0.0037375788670033216, + 0.007107329089194536, + 0.006438352167606354, + 0.0057596247643232346, + 0.007361036259680986, + 0.008343148976564407, + 0.01452433317899704, + 0.0029522187542170286, + 0.0027222572825849056, + 0.004270054865628481, + 0.0040522185154259205, + 0.005117799621075392, + 0.004927417729049921, + 0.00562590267509222, + 0.005264819134026766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "selected_reward": -0.03369913622736931 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899594306945801 + ], + "opposite_logprob_tokens": [ + -2.578995943069458 + ], + "logprob_reward": -0.07899594306945801, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899594306945801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03314713388681412, + 0.04182881489396095, + 0.015344029292464256, + 0.03363638371229172, + 0.02587120421230793, + 0.10265759378671646, + 0.08315739035606384, + 0.6117186546325684, + 0.02934725396335125, + 0.02329154871404171 + ], + "credit_map": [ + 0.0030563718173652887, + 0.003378257853910327, + 0.008038493804633617, + 0.003416795050725341, + 0.0026677364949136972, + 0.005322424229234457, + 0.0027764092665165663, + 0.004490647930651903, + 0.004245830699801445, + 0.003339167917147279, + 0.0035873972810804844, + 0.004216440487653017, + 0.0033317399211227894, + 0.012168345972895622, + 0.00702780531719327, + 0.003912086598575115, + 0.0027240405324846506, + 0.0024264114908874035, + 0.0029575072694569826, + 0.001817735261283815, + 0.0015018184203654528, + 0.0011172564700245857, + 0.001493359450250864, + 0.0013059013290330768, + 0.002566517796367407, + 0.002963096834719181, + 0.004399215802550316, + 0.010497881099581718, + 0.0026892335154116154, + 0.0036057846155017614, + 0.0029808664694428444, + 0.003933790139853954, + 0.00213921838440001, + 0.0026643131859600544, + 0.0032968176528811455, + 0.0043436684645712376, + 0.0027822814881801605, + 0.003214654279872775, + 0.0037581047508865595, + 0.0036721464712172747, + 0.007610177621245384, + 0.015122995711863041, + 0.016792185604572296, + 0.013167729601264, + 0.013986703008413315, + 0.011571136303246021, + 0.01172440405935049, + 0.012682265602052212, + 0.006161204539239407, + 0.007110374979674816, + 0.010838265530765057, + 0.011228608898818493, + 0.008991142734885216, + 0.012664585374295712, + 0.011650007218122482, + 0.014513206668198109, + 0.024390239268541336, + 0.009588020853698254, + 0.2221493273973465, + 0.030209848657250404, + 0.012734897434711456, + 0.04305883124470711, + 0.021796733140945435, + 0.24779075384140015, + 0.0020793890580534935, + 0.0015298211947083473, + 0.003719154978170991, + 0.0052971127443015575, + 0.0018793841591104865, + 0.005902778822928667, + 0.00420996081084013, + 0.004729655105620623, + 0.002074583899229765, + 0.0016177281504496932, + 0.003540035104379058, + 0.002305082743987441, + 0.002911393530666828, + 0.0036735490430146456, + 0.003374408697709441, + 0.003794768126681447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.07899594306945801 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818074844777584 + ], + "opposite_logprob_tokens": [ + -4.018180847167969 + ], + "logprob_reward": -0.01818074844777584, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818074844777584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08487315475940704, + 0.03545359894633293, + 0.044520750641822815, + 0.10236199200153351, + 0.02234688401222229, + 0.05127189680933952, + 0.0853429064154625, + 0.5368465185165405, + 0.013602341525256634, + 0.023379946127533913 + ], + "credit_map": [ + 0.006503036245703697, + 0.007092658895999193, + 0.015974048525094986, + 0.010616072453558445, + 0.008577238768339157, + 0.013577609322965145, + 0.00940284226089716, + 0.013129662722349167, + 0.004513844847679138, + 0.004435468465089798, + 0.004354530014097691, + 0.0032424333039671183, + 0.003409790340811014, + 0.008057338185608387, + 0.0051793986931443214, + 0.002260800451040268, + 0.0052564493380486965, + 0.006046617403626442, + 0.010584142990410328, + 0.006686717737466097, + 0.0038255564868450165, + 0.0040045445784926414, + 0.0038650562055408955, + 0.004251666367053986, + 0.006512538064271212, + 0.007295595016330481, + 0.010143257677555084, + 0.03865137696266174, + 0.00943045038729906, + 0.00620499812066555, + 0.009010707959532738, + 0.015113083645701408, + 0.0018006630707532167, + 0.002535137115046382, + 0.0030044817831367254, + 0.002510067541152239, + 0.0017282769549638033, + 0.0025282660499215126, + 0.003892152104526758, + 0.0043478417210280895, + 0.003413103288039565, + 0.00900483038276434, + 0.007991594262421131, + 0.006447252351790667, + 0.0072201918810606, + 0.005917922593653202, + 0.005476330406963825, + 0.005800674203783274, + 0.005054936278611422, + 0.0051541258580982685, + 0.011949943378567696, + 0.015986373648047447, + 0.0061061447486281395, + 0.009206922724843025, + 0.013686355203390121, + 0.018198110163211823, + 0.0229386817663908, + 0.022055407986044884, + 0.1716780662536621, + 0.032692164182662964, + 0.018742891028523445, + 0.04777970910072327, + 0.026051536202430725, + 0.19490808248519897, + 0.000975694740191102, + 0.000968033098615706, + 0.0019062497885897756, + 0.002816856140270829, + 0.0014776383759453893, + 0.0020246978383511305, + 0.0016943670343607664, + 0.0017388061387464404, + 0.00201811664737761, + 0.0015233209123834968, + 0.0036473844666033983, + 0.0028398684225976467, + 0.0025761714205145836, + 0.0036812694743275642, + 0.0034171738661825657, + 0.0036766440607607365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.01818074844777584 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "first-person perspective shot from inside a rocking carriage, moving forward at eye level, perfectly simulates the tangible sensation of a bumpy ride. The camera gently sways and pitches in time with the carriage's rhythm. Peering out through the rustic wooden window frames, one sees the narrow streets of an ancient village, paved entirely with weathered cobblestones of varying sizes. The streets are lined with a patchwork of wooden and stone houses, their walls covered in deep green moss. Ancient tiled roofs are stained by age, and occasionally wisps of smoke can be seen rising from stone chimneys. The warm, golden afternoon sun slants across the scene, casting a soft glow over the scene, lengthening the shadows of the buildings and making even tiny particles of dust in the air visible. The entire picture presents a remarkable retro film texture, with warm and saturated colors, obvious graininess and occasional slight lens flare caused by direct sunlight. As the carriage slowly moves forward, the scenery in front of you rhythmically rises and falls, creating a tranquil, nostalgic and immersive experience, as if a dusty memory is slowly being awakened.", + "seed": 1343, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_0/index_8016_seed_1343_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.060866395384073256, + "vlm_logprob": -0.060866395384073256, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.060866395384073256, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02281569503247738 + ], + "opposite_logprob_tokens": [ + -4.147815704345703 + ], + "logprob_reward": -0.02281569503247738, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02281569503247738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17682576179504395, + 0.05388987436890602, + 0.026706963777542114, + 0.05618473142385483, + 0.033445168286561966, + 0.06363178044557571, + 0.19906377792358398, + 0.1255691945552826, + 0.20024563372135162, + 0.06443704664707184 + ], + "credit_map": [ + 0.01873226836323738, + 0.018087025731801987, + 0.023853762075304985, + 0.014045213349163532, + 0.02297591045498848, + 0.031130824238061905, + 0.01997997798025608, + 0.028020774945616722, + 0.007542230188846588, + 0.006072174292057753, + 0.0077841030433773994, + 0.004662462510168552, + 0.004617306403815746, + 0.007215776946395636, + 0.006655855569988489, + 0.00933996681123972, + 0.003554175840690732, + 0.003261844627559185, + 0.003909741062670946, + 0.002588185016065836, + 0.001989200245589018, + 0.0021822566632181406, + 0.0031487946398556232, + 0.006072766147553921, + 0.004165916703641415, + 0.004075011238455772, + 0.007453569211065769, + 0.019041184335947037, + 0.0031939903274178505, + 0.0029261561576277018, + 0.008707355707883835, + 0.006621548440307379, + 0.003435774240642786, + 0.00320060015656054, + 0.004783255513757467, + 0.0029830692801624537, + 0.005187462083995342, + 0.005233595613390207, + 0.0036497199907898903, + 0.004971692804247141, + 0.004955274052917957, + 0.005983063485473394, + 0.006382269784808159, + 0.007127418648451567, + 0.009942105039954185, + 0.007743301335722208, + 0.008836806751787663, + 0.012661542743444443, + 0.012331032194197178, + 0.01594502292573452, + 0.05531672015786171, + 0.0127195343375206, + 0.011195310391485691, + 0.01251226756721735, + 0.02347775734961033, + 0.05556614696979523, + 0.007717757485806942, + 0.012234460562467575, + 0.014384247362613678, + 0.011565010994672775, + 0.017438974231481552, + 0.030153263360261917, + 0.01587834767997265, + 0.016197139397263527, + 0.013221390545368195, + 0.013997365720570087, + 0.03477252647280693, + 0.019442664459347725, + 0.02007845975458622, + 0.03870360180735588, + 0.030624650418758392, + 0.029404977336525917, + 0.008387270383536816, + 0.01099011767655611, + 0.008625425398349762, + 0.006225756835192442, + 0.007757970597594976, + 0.006182336248457432, + 0.0068712299689650536, + 0.009396945126354694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.02281569503247738 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004112359136343002 + ], + "opposite_logprob_tokens": [ + -5.504112243652344 + ], + "logprob_reward": -0.004112359136343002, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004112359136343002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1839086413383484, + 0.07878173142671585, + 0.06398414075374603, + 0.07359065860509872, + 0.07353198528289795, + 0.11170153319835663, + 0.23771381378173828, + 0.06634741276502609, + 0.07087596505880356, + 0.0395641103386879 + ], + "credit_map": [ + 0.02162039279937744, + 0.016555704176425934, + 0.030199693515896797, + 0.011968323029577732, + 0.016680020838975906, + 0.02587023563683033, + 0.028034962713718414, + 0.03297930583357811, + 0.009917673654854298, + 0.010290302336215973, + 0.012947634793817997, + 0.008551931008696556, + 0.007453722879290581, + 0.010157888755202293, + 0.006596407853066921, + 0.012866172008216381, + 0.005568185821175575, + 0.006818830035626888, + 0.006905105896294117, + 0.018207574263215065, + 0.0039494396187365055, + 0.004381408449262381, + 0.007169060409069061, + 0.010984532535076141, + 0.005962138529866934, + 0.0059757111594080925, + 0.007819919846951962, + 0.019816549494862556, + 0.005861788056790829, + 0.0066329315304756165, + 0.010560543276369572, + 0.01096107717603445, + 0.007271826732903719, + 0.0070279380306601524, + 0.012286072596907616, + 0.0075157112441957, + 0.0087712612003088, + 0.009812030009925365, + 0.008764371275901794, + 0.012082775123417377, + 0.008836019784212112, + 0.01168600283563137, + 0.01572517305612564, + 0.009896752424538136, + 0.018903736025094986, + 0.010550867766141891, + 0.015836047008633614, + 0.020266935229301453, + 0.014517512172460556, + 0.018756769597530365, + 0.06262964755296707, + 0.016261659562587738, + 0.016001861542463303, + 0.017323551699519157, + 0.028361527249217033, + 0.06386128813028336, + 0.004719854332506657, + 0.005927043501287699, + 0.005841502454131842, + 0.007310028653591871, + 0.009846191853284836, + 0.016388537362217903, + 0.008113828487694263, + 0.008200429379940033, + 0.006333248224109411, + 0.00813099555671215, + 0.014060472138226032, + 0.006795926950871944, + 0.006837878841906786, + 0.007194109261035919, + 0.01029935572296381, + 0.011223974637687206, + 0.006377204786986113, + 0.006011875346302986, + 0.00545551348477602, + 0.002607853850349784, + 0.0035522826947271824, + 0.004184227902442217, + 0.004752098582684994, + 0.006623053923249245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.004112359136343002 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705247104167938 + ], + "opposite_logprob_tokens": [ + -2.1270525455474854 + ], + "logprob_reward": -0.12705247104167938, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705247104167938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21444015204906464, + 0.06332752853631973, + 0.1541646420955658, + 0.1138761043548584, + 0.07725304365158081, + 0.05067747086286545, + 0.10081259906291962, + 0.05992435663938522, + 0.0821889340877533, + 0.08333508670330048 + ], + "credit_map": [ + 0.02306017093360424, + 0.02088281698524952, + 0.029565982520580292, + 0.023102635517716408, + 0.027994947507977486, + 0.027475977316498756, + 0.024759763851761818, + 0.03759785369038582, + 0.007679032161831856, + 0.00563921220600605, + 0.011350564658641815, + 0.007181983441114426, + 0.00562711525708437, + 0.010235031135380268, + 0.006489383988082409, + 0.009125204756855965, + 0.012436499819159508, + 0.01411797758191824, + 0.01792905293405056, + 0.04838276654481888, + 0.008530844002962112, + 0.009929511696100235, + 0.019489755854010582, + 0.02334822714328766, + 0.011620383709669113, + 0.009440627880394459, + 0.011262024752795696, + 0.03176487982273102, + 0.014444640837609768, + 0.006510597188025713, + 0.009941685944795609, + 0.018891271203756332, + 0.007201561238616705, + 0.006990580353885889, + 0.012966057285666466, + 0.006786057725548744, + 0.010105426423251629, + 0.008460095152258873, + 0.009001253172755241, + 0.01574201136827469, + 0.005541203077882528, + 0.007790400646626949, + 0.006549336481839418, + 0.004364828113466501, + 0.006074982695281506, + 0.007118380628526211, + 0.004857364110648632, + 0.00838097557425499, + 0.004976433701813221, + 0.005076265893876553, + 0.026094457134604454, + 0.0059776706621050835, + 0.0069867572747170925, + 0.007535201963037252, + 0.014537930488586426, + 0.029627887532114983, + 0.004902014043182135, + 0.006204965524375439, + 0.00860991608351469, + 0.006133183836936951, + 0.009471005760133266, + 0.010417584329843521, + 0.006810508202761412, + 0.007375182118266821, + 0.011780636385083199, + 0.014196728356182575, + 0.012571337632834911, + 0.006952507421374321, + 0.010151175782084465, + 0.007930228486657143, + 0.006825495511293411, + 0.011780824512243271, + 0.011799578554928303, + 0.012813633307814598, + 0.01171770878136158, + 0.005609214771538973, + 0.008211815729737282, + 0.009002800099551678, + 0.009922138415277004, + 0.014258196577429771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.12705247104167938 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707094848155975 + ], + "opposite_logprob_tokens": [ + -2.127070903778076 + ], + "logprob_reward": -0.12707094848155975, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707094848155975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14499148726463318, + 0.07466508448123932, + 0.11587199568748474, + 0.11175312101840973, + 0.022830918431282043, + 0.051201216876506805, + 0.18306075036525726, + 0.07099667191505432, + 0.04769128933548927, + 0.1769375056028366 + ], + "credit_map": [ + 0.01513897255063057, + 0.017890166491270065, + 0.01900695078074932, + 0.011537835001945496, + 0.01417444460093975, + 0.022422857582569122, + 0.022155676037073135, + 0.022664586082100868, + 0.010175329633057117, + 0.008384658955037594, + 0.013084174133837223, + 0.007943526841700077, + 0.0059632048942148685, + 0.009108680300414562, + 0.007953529246151447, + 0.012051980011165142, + 0.013358918949961662, + 0.015401293523609638, + 0.020991330966353416, + 0.013896355405449867, + 0.010289309546351433, + 0.014374143444001675, + 0.011164029128849506, + 0.016396619379520416, + 0.014010247774422169, + 0.014031295664608479, + 0.010643729008734226, + 0.020616421476006508, + 0.018906323239207268, + 0.008729028515517712, + 0.011705677956342697, + 0.013110395520925522, + 0.002060460392385721, + 0.0017295102588832378, + 0.003016240894794464, + 0.002129126340150833, + 0.003679696237668395, + 0.00337452394887805, + 0.002777331741526723, + 0.004064029548317194, + 0.0037008989602327347, + 0.0073471395298838615, + 0.006198052316904068, + 0.005593232344835997, + 0.007801605388522148, + 0.006318660452961922, + 0.005324732977896929, + 0.00891689769923687, + 0.008741648867726326, + 0.008051414974033833, + 0.047344811260700226, + 0.009306209161877632, + 0.008633783087134361, + 0.014744453132152557, + 0.02844577096402645, + 0.05779266357421875, + 0.005243683233857155, + 0.006344165187329054, + 0.010315832681953907, + 0.0077354926615953445, + 0.00991777889430523, + 0.014499243348836899, + 0.008520008996129036, + 0.008420460857450962, + 0.005676115397363901, + 0.005147496238350868, + 0.0067128585651516914, + 0.0046746390871703625, + 0.005038154311478138, + 0.006605018861591816, + 0.007108632940798998, + 0.006728373933583498, + 0.034756358712911606, + 0.023862915113568306, + 0.03199763968586922, + 0.012669139541685581, + 0.01327507197856903, + 0.017711272463202477, + 0.014907981269061565, + 0.02775714173913002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.12707094848155975 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328050322830677 + ], + "opposite_logprob_tokens": [ + -3.773280620574951 + ], + "logprob_reward": -0.02328050322830677, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328050322830677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24193213880062103, + 0.07956558465957642, + 0.06429295986890793, + 0.06737486273050308, + 0.07279810309410095, + 0.04480874538421631, + 0.2854117751121521, + 0.05915038660168648, + 0.03211371228098869, + 0.052551645785570145 + ], + "credit_map": [ + 0.031150493770837784, + 0.02706771157681942, + 0.03716538846492767, + 0.024073852226138115, + 0.026662206277251244, + 0.029352402314543724, + 0.02596983127295971, + 0.040490273386240005, + 0.010984780266880989, + 0.009748200885951519, + 0.013829628005623817, + 0.007288017310202122, + 0.008286425843834877, + 0.011088399216532707, + 0.007203691639006138, + 0.01113644614815712, + 0.006130392197519541, + 0.009025533683598042, + 0.01093654241412878, + 0.011189941316843033, + 0.004582355264574289, + 0.00694748479872942, + 0.005453501362353563, + 0.010027209296822548, + 0.004226082935929298, + 0.0048048049211502075, + 0.004627783317118883, + 0.017691118642687798, + 0.003000575117766857, + 0.00431241886690259, + 0.006034570746123791, + 0.02267751470208168, + 0.00772035913541913, + 0.005877790041267872, + 0.011586572043597698, + 0.006824805401265621, + 0.006024753674864769, + 0.008765396662056446, + 0.010304901748895645, + 0.015693532302975655, + 0.003925727214664221, + 0.006069546565413475, + 0.004851033911108971, + 0.00425393832847476, + 0.006505805533379316, + 0.005309208296239376, + 0.006050077732652426, + 0.007843410596251488, + 0.025272611528635025, + 0.03616597130894661, + 0.0354105606675148, + 0.03238464891910553, + 0.02701311744749546, + 0.033059246838092804, + 0.039641425013542175, + 0.05646420642733574, + 0.004350061062723398, + 0.005050421692430973, + 0.004912227857857943, + 0.005701316520571709, + 0.007853190414607525, + 0.012163489125669003, + 0.009554073214530945, + 0.00956561230123043, + 0.003395206294953823, + 0.0032282418105751276, + 0.0038282049354165792, + 0.0032502098474651575, + 0.0032174279913306236, + 0.004370738752186298, + 0.0058137038722634315, + 0.005009980406612158, + 0.007004506886005402, + 0.005168522708117962, + 0.00547775300219655, + 0.004468600731343031, + 0.004607012961059809, + 0.006381941959261894, + 0.009084279648959637, + 0.010359030216932297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.02328050322830677 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "first-person perspective shot from inside a rocking carriage, moving forward at eye level, perfectly simulates the tangible sensation of a bumpy ride. The camera gently sways and pitches in time with the carriage's rhythm. Peering out through the rustic wooden window frames, one sees the narrow streets of an ancient village, paved entirely with weathered cobblestones of varying sizes. The streets are lined with a patchwork of wooden and stone houses, their walls covered in deep green moss. Ancient tiled roofs are stained by age, and occasionally wisps of smoke can be seen rising from stone chimneys. The warm, golden afternoon sun slants across the scene, casting a soft glow over the scene, lengthening the shadows of the buildings and making even tiny particles of dust in the air visible. The entire picture presents a remarkable retro film texture, with warm and saturated colors, obvious graininess and occasional slight lens flare caused by direct sunlight. As the carriage slowly moves forward, the scenery in front of you rhythmically rises and falls, creating a tranquil, nostalgic and immersive experience, as if a dusty memory is slowly being awakened.", + "seed": 1344, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_0/index_8016_seed_1344_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05413776114583015, + "vlm_logprob": -0.05413776114583015, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.05413776114583015, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04043278098106384 + ], + "opposite_logprob_tokens": [ + -3.4154326915740967 + ], + "logprob_reward": -0.04043278098106384, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04043278098106384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14843326807022095, + 0.05099378526210785, + 0.14085055887699127, + 0.1673397570848465, + 0.06258723139762878, + 0.03240662068128586, + 0.05577950179576874, + 0.2085338681936264, + 0.04680964723229408, + 0.08626579493284225 + ], + "credit_map": [ + 0.017216240987181664, + 0.01334032230079174, + 0.02771764248609543, + 0.03006053902208805, + 0.008850536309182644, + 0.013877395540475845, + 0.015722835436463356, + 0.02164774388074875, + 0.006886171642690897, + 0.00946483202278614, + 0.007310252636671066, + 0.00836747046560049, + 0.0037802194710820913, + 0.0051374430768191814, + 0.006370160728693008, + 0.0036772361490875483, + 0.0178686510771513, + 0.027761787176132202, + 0.011284909211099148, + 0.01767510548233986, + 0.02045312710106373, + 0.015504728071391582, + 0.018710972741246223, + 0.011591272428631783, + 0.013928357511758804, + 0.01053874846547842, + 0.042920153588056564, + 0.043271519243717194, + 0.01628659851849079, + 0.011428863741457462, + 0.017027541995048523, + 0.011937967501580715, + 0.009527451358735561, + 0.004122923593968153, + 0.007969406433403492, + 0.00527794798836112, + 0.007630314212292433, + 0.005401763133704662, + 0.011674330569803715, + 0.010983092710375786, + 0.006331998854875565, + 0.0019993751775473356, + 0.005410162266343832, + 0.002682213671505451, + 0.003155753016471863, + 0.004783793818205595, + 0.005215806420892477, + 0.002827512100338936, + 0.009400778450071812, + 0.004123216029256582, + 0.011071284301578999, + 0.007073734421283007, + 0.0044363681226968765, + 0.006497448310256004, + 0.009079846553504467, + 0.004096816293895245, + 0.06863842904567719, + 0.010412381961941719, + 0.07776152342557907, + 0.013656430877745152, + 0.006022618152201176, + 0.009104199707508087, + 0.012661713175475597, + 0.010276553221046925, + 0.0044277808628976345, + 0.004024113994091749, + 0.006789365317672491, + 0.008752419613301754, + 0.003766916459426284, + 0.004965940024703741, + 0.006948928814381361, + 0.007134179584681988, + 0.00633628387004137, + 0.006864350754767656, + 0.008607101626694202, + 0.014153667725622654, + 0.008133495226502419, + 0.011298703029751778, + 0.01442340575158596, + 0.01644878089427948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.04043278098106384 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032120561227202415 + ], + "opposite_logprob_tokens": [ + -5.753211975097656 + ], + "logprob_reward": -0.0032120561227202415, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032120561227202415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2472425103187561, + 0.07538101077079773, + 0.09617560356855392, + 0.08900037407875061, + 0.07460141181945801, + 0.12330616265535355, + 0.059899237006902695, + 0.1032668948173523, + 0.057928889989852905, + 0.0731978490948677 + ], + "credit_map": [ + 0.022172752767801285, + 0.01761234737932682, + 0.05221743881702423, + 0.0306285060942173, + 0.0350787527859211, + 0.027306511998176575, + 0.02578892931342125, + 0.0364372693002224, + 0.007552799303084612, + 0.01365066971629858, + 0.012407139874994755, + 0.011605177074670792, + 0.0070532881654798985, + 0.00843534991145134, + 0.006193665787577629, + 0.008482920937240124, + 0.012979337945580482, + 0.013317483477294445, + 0.010233132168650627, + 0.013754637911915779, + 0.011532320640981197, + 0.011534362100064754, + 0.010629204101860523, + 0.012195127084851265, + 0.005214405246078968, + 0.007273230701684952, + 0.015514479018747807, + 0.021999051794409752, + 0.008650789968669415, + 0.007710207253694534, + 0.014762970618903637, + 0.007875241339206696, + 0.010756880044937134, + 0.0036694828886538744, + 0.006375735625624657, + 0.00541703961789608, + 0.008316469378769398, + 0.006508530583232641, + 0.015141383744776249, + 0.018415888771414757, + 0.023188045248389244, + 0.00945691205561161, + 0.011617098934948444, + 0.011015611700713634, + 0.010024920105934143, + 0.02691756747663021, + 0.020400024950504303, + 0.01068598311394453, + 0.008597630076110363, + 0.003828464774414897, + 0.010754694230854511, + 0.006078534293919802, + 0.00545567087829113, + 0.006777357775717974, + 0.006910833064466715, + 0.011496051214635372, + 0.03525222837924957, + 0.004634852521121502, + 0.03708074614405632, + 0.006656369660049677, + 0.00393493939191103, + 0.005200433544814587, + 0.005304894410073757, + 0.005202434491366148, + 0.006011732388287783, + 0.006349682807922363, + 0.01000099815428257, + 0.008349823765456676, + 0.004560794681310654, + 0.006414239760488272, + 0.008213100954890251, + 0.008028515614569187, + 0.006756315007805824, + 0.011840790510177612, + 0.007674625143408775, + 0.01271028071641922, + 0.007549162954092026, + 0.0069432067684829235, + 0.009959331713616848, + 0.009764140471816063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.0032120561227202415 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002703845500946 + ], + "opposite_logprob_tokens": [ + -2.6950271129608154 + ], + "logprob_reward": -0.07002703845500946, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002703845500946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10265390574932098, + 0.03820376470685005, + 0.35929742455482483, + 0.07421653717756271, + 0.07036294043064117, + 0.1050938293337822, + 0.06616178154945374, + 0.08608658611774445, + 0.04589336737990379, + 0.052029889076948166 + ], + "credit_map": [ + 0.011669677682220936, + 0.009496725164353848, + 0.019568294286727905, + 0.024329278618097305, + 0.008284327574074268, + 0.007439589127898216, + 0.011604052037000656, + 0.010261952877044678, + 0.003863729303702712, + 0.003893704153597355, + 0.008818290196359158, + 0.00751878647133708, + 0.0023339414037764072, + 0.0038846780080348253, + 0.00331278913654387, + 0.004577845800668001, + 0.0561082623898983, + 0.05618845671415329, + 0.0252483319491148, + 0.047794125974178314, + 0.05339054390788078, + 0.03149070218205452, + 0.04955557361245155, + 0.03952139988541603, + 0.007692642509937286, + 0.008055036887526512, + 0.014898649416863918, + 0.010209692642092705, + 0.01049902755767107, + 0.005989097058773041, + 0.010052520781755447, + 0.006819870322942734, + 0.004812528844922781, + 0.005899087525904179, + 0.01759244129061699, + 0.00502636656165123, + 0.010538223199546337, + 0.00563054671511054, + 0.014082803390920162, + 0.0067809405736625195, + 0.024880068376660347, + 0.006704797502607107, + 0.009684726595878601, + 0.007493979297578335, + 0.007346601225435734, + 0.027673285454511642, + 0.01439959928393364, + 0.00691076647490263, + 0.010394701734185219, + 0.008081050589680672, + 0.010262723080813885, + 0.011952093802392483, + 0.005524409469217062, + 0.0050619496032595634, + 0.010057522915303707, + 0.004827328026294708, + 0.029477892443537712, + 0.003159615909680724, + 0.033104442059993744, + 0.005330425221472979, + 0.002849467098712921, + 0.004139983560889959, + 0.004118781536817551, + 0.0039059692062437534, + 0.008840366266667843, + 0.004628350492566824, + 0.0060293348506093025, + 0.007423521485179663, + 0.0035818745382130146, + 0.006209990009665489, + 0.00562495831400156, + 0.003554965602234006, + 0.003255840390920639, + 0.004404315259307623, + 0.004137286450713873, + 0.010334580205380917, + 0.0053819469176232815, + 0.006909520830959082, + 0.009173988364636898, + 0.008432406932115555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.07002703845500946 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428251564502716 + ], + "opposite_logprob_tokens": [ + -2.017825126647949 + ], + "logprob_reward": -0.1428251564502716, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428251564502716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14506743848323822, + 0.08102477341890335, + 0.10978838801383972, + 0.05818840488791466, + 0.11940714716911316, + 0.19718964397907257, + 0.07226322591304779, + 0.043088771402835846, + 0.111418217420578, + 0.0625639408826828 + ], + "credit_map": [ + 0.014290956780314445, + 0.010205315425992012, + 0.036518581211566925, + 0.0176899004727602, + 0.015461866743862629, + 0.014368323609232903, + 0.009026688523590565, + 0.02750580385327339, + 0.006421446334570646, + 0.006665615830570459, + 0.019881287589669228, + 0.01526612788438797, + 0.005592611618340015, + 0.012473640032112598, + 0.005587249528616667, + 0.009136798791587353, + 0.015450356528162956, + 0.01597430557012558, + 0.008545076474547386, + 0.014102154411375523, + 0.01464544702321291, + 0.011211578734219074, + 0.014545702375471592, + 0.015313777141273022, + 0.006774354260414839, + 0.006349904462695122, + 0.009253970347344875, + 0.006694717798382044, + 0.007923142053186893, + 0.006571819074451923, + 0.007884707301855087, + 0.006735790520906448, + 0.01163359172642231, + 0.01239483430981636, + 0.01582861877977848, + 0.013967460952699184, + 0.020571360364556313, + 0.014204508624970913, + 0.018298104405403137, + 0.012508663348853588, + 0.023506570607423782, + 0.024811219424009323, + 0.016782084479928017, + 0.024531397968530655, + 0.024702342227101326, + 0.020743442699313164, + 0.03957630693912506, + 0.022536294534802437, + 0.012499352917075157, + 0.007115029729902744, + 0.01370347011834383, + 0.010643689893186092, + 0.005389525555074215, + 0.007263137027621269, + 0.010396840050816536, + 0.005252182483673096, + 0.008457514457404613, + 0.002880771877244115, + 0.012710771523416042, + 0.0029850974678993225, + 0.002693097572773695, + 0.004725716542452574, + 0.0037537228781729937, + 0.004882080014795065, + 0.012532621622085571, + 0.01397336833178997, + 0.013075387105345726, + 0.022367972880601883, + 0.008823647163808346, + 0.019778601825237274, + 0.010691781528294086, + 0.010174838826060295, + 0.0064582196064293385, + 0.005150310229510069, + 0.0082468306645751, + 0.00858568400144577, + 0.005175834987312555, + 0.013183130882680416, + 0.008521823212504387, + 0.007242106832563877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.1428251564502716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a first-person perspective inside a rocking carriage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera sway and pitch rhythmically to simulate a bumpy ride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the streets paved with weathered cobblestones of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the house walls covered in deep green moss with ancient tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a retro film texture with warm, saturated colors, obvious graininess, and lens flare?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191773720085621 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.014191773720085621, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191773720085621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1671583354473114, + 0.23164023458957672, + 0.07623901963233948, + 0.08969014883041382, + 0.1417834311723709, + 0.04829317331314087, + 0.04982680827379227, + 0.034160204231739044, + 0.08145450055599213, + 0.07975412160158157 + ], + "credit_map": [ + 0.018173566088080406, + 0.01483552623540163, + 0.03531385213136673, + 0.024706581607460976, + 0.0155775286257267, + 0.020260903984308243, + 0.019044967368245125, + 0.019245414063334465, + 0.01687159389257431, + 0.014483841136097908, + 0.054069213569164276, + 0.03972477838397026, + 0.022024011239409447, + 0.03169867396354675, + 0.02189946547150612, + 0.030868684872984886, + 0.010237454436719418, + 0.013376648537814617, + 0.008535073138773441, + 0.010288072749972343, + 0.007793070748448372, + 0.010155562311410904, + 0.009141051210463047, + 0.006712094880640507, + 0.005987430457025766, + 0.006516446825116873, + 0.023585528135299683, + 0.025398071855306625, + 0.007055378518998623, + 0.005540638696402311, + 0.009563179686665535, + 0.006043486297130585, + 0.031722426414489746, + 0.008004304021596909, + 0.04707253351807594, + 0.01218228880316019, + 0.006235966924577951, + 0.005913610570132732, + 0.013974778354167938, + 0.01667754538357258, + 0.01157590001821518, + 0.0025531977880746126, + 0.005220439750701189, + 0.003167548915371299, + 0.004256206098943949, + 0.013221954926848412, + 0.004452131222933531, + 0.003845801344141364, + 0.007421600632369518, + 0.0033580090384930372, + 0.012771760113537312, + 0.004110593814402819, + 0.004200286231935024, + 0.0075780367478728294, + 0.004978169221431017, + 0.005408362485468388, + 0.010251123458147049, + 0.0028156908228993416, + 0.009460993111133575, + 0.0022640509996563196, + 0.0015795871149748564, + 0.002031879499554634, + 0.0026937993243336678, + 0.0030630819965153933, + 0.009147990494966507, + 0.009438632987439632, + 0.010176271200180054, + 0.015183376148343086, + 0.007615364156663418, + 0.010040942579507828, + 0.010391771793365479, + 0.009460157714784145, + 0.007307931315153837, + 0.009391790255904198, + 0.00813743844628334, + 0.009751575067639351, + 0.009478282183408737, + 0.012560563161969185, + 0.011163771152496338, + 0.011962774209678173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.014191773720085621 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_1/index_tensor([8491, 8491])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_1/index_tensor([8491, 8491])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..207550885ae52f828cb4890d6537080322d97fbe --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_1/index_tensor([8491, 8491])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, initially static, focuses on a piece of yellow talisman paper suspended against a dark, blurred background. Its texture is rough, and upon it, intricate, twisting, and powerfully drawn mystical runes in vibrant vermilion red. Suddenly, without warning, a fiery red dot ignites at the center of the paper. Immediately, a brilliant orange-red flame bursts forth, ravenously devouring the paper. The edges of the paper rapidly curl and char under the licking of the flames, while the red runes glow with an extraordinary brilliance in the firelight, as if instantly animated. Then, as if lifted up by an invisible force, the entire burning talisman paper spins and struggles upward. The camera rapidly tilts upward, closely tracking its movement, capturing the dynamic moment as it leaves behind a trail of dancing, flickering black embers. The image is characterized by a stark contrast of light and shadow, with the flames acting as the sole, dynamic source of light, casting a violent, dancing shadow against the pitch-black background, creating a tense atmosphere imbued with a sense of mysterious ritual and the eruption of supernatural power.", + "seed": 15680, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_1/index_8491_seed_15680_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.012249076180160046, + "vlm_logprob": -0.012249076180160046, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.012249076180160046, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010597973130643368 + ], + "opposite_logprob_tokens": [ + -5.385598182678223 + ], + "logprob_reward": -0.010597973130643368, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.010597973130643368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17993395030498505, + 0.032955337315797806, + 0.0678115114569664, + 0.07985538244247437, + 0.16534750163555145, + 0.04674440622329712, + 0.061808016151189804, + 0.08980432152748108, + 0.06792191416025162, + 0.2078176885843277 + ], + "credit_map": [ + 0.01687365397810936, + 0.021282117813825607, + 0.013225331902503967, + 0.014957486651837826, + 0.02547546662390232, + 0.028127623721957207, + 0.022680796682834625, + 0.03731146082282066, + 0.0031748495530337095, + 0.0031824286561459303, + 0.002883867360651493, + 0.0030475882813334465, + 0.0045269811525940895, + 0.005651731509715319, + 0.004342108033597469, + 0.006145780440419912, + 0.008316694758832455, + 0.0064308722503483295, + 0.005369242746382952, + 0.006203248165547848, + 0.005097652785480022, + 0.013414829038083553, + 0.005802404135465622, + 0.017176557332277298, + 0.0072978343814611435, + 0.009482280351221561, + 0.007253775838762522, + 0.008210430853068829, + 0.007213745731860399, + 0.013936072587966919, + 0.007142836228013039, + 0.019318396225571632, + 0.011455673724412918, + 0.012613598257303238, + 0.01707231067121029, + 0.014558259397745132, + 0.012861628085374832, + 0.03416058048605919, + 0.016164006665349007, + 0.04646144434809685, + 0.0036514110397547483, + 0.004818346817046404, + 0.004741877783089876, + 0.004101981874555349, + 0.005320985335856676, + 0.008838973939418793, + 0.0037996871396899223, + 0.011471141129732132, + 0.0036196112632751465, + 0.0025800324510782957, + 0.0042679947800934315, + 0.0029239878058433533, + 0.002865932183340192, + 0.01670556515455246, + 0.005876876413822174, + 0.02296801283955574, + 0.014369039796292782, + 0.01190048549324274, + 0.014428960159420967, + 0.01018680352717638, + 0.009267954155802727, + 0.00869892816990614, + 0.009394141845405102, + 0.011558005586266518, + 0.010217264294624329, + 0.005120416637510061, + 0.0062293680384755135, + 0.01020723581314087, + 0.007755097467452288, + 0.010178418830037117, + 0.009901660494506359, + 0.008312450721859932, + 0.01920962706208229, + 0.024165403097867966, + 0.034093715250492096, + 0.02284320257604122, + 0.0344088114798069, + 0.027418017387390137, + 0.020433079451322556, + 0.025245817378163338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.010597973130643368 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032107490114867687 + ], + "opposite_logprob_tokens": [ + -5.753210544586182 + ], + "logprob_reward": -0.0032107490114867687, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032107490114867687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4100927412509918, + 0.07454818487167358, + 0.03215523436665535, + 0.06589820981025696, + 0.12925982475280762, + 0.06382785737514496, + 0.03748461231589317, + 0.052550431340932846, + 0.06016039848327637, + 0.07402254641056061 + ], + "credit_map": [ + 0.029541729018092155, + 0.06037221848964691, + 0.02413378283381462, + 0.03749114274978638, + 0.06902322173118591, + 0.06485949456691742, + 0.05890890955924988, + 0.0657622218132019, + 0.005505026318132877, + 0.007497701328247786, + 0.006192006170749664, + 0.008120367303490639, + 0.010255105793476105, + 0.012740213423967361, + 0.012107545509934425, + 0.012130221351981163, + 0.003342405194416642, + 0.00255044037476182, + 0.0033919429406523705, + 0.003229545196518302, + 0.0028906113002449274, + 0.005851134192198515, + 0.0032703750766813755, + 0.00762878218665719, + 0.006802213843911886, + 0.008023194968700409, + 0.006314584519714117, + 0.008403710089623928, + 0.005991910118609667, + 0.010689032264053822, + 0.005879508331418037, + 0.013794058002531528, + 0.011364033445715904, + 0.015000130981206894, + 0.01774449273943901, + 0.015012486837804317, + 0.011514715850353241, + 0.021777713671326637, + 0.012343139387667179, + 0.02450312115252018, + 0.004449207801371813, + 0.01025624293833971, + 0.005499015562236309, + 0.006370189134031534, + 0.005932412575930357, + 0.011603029444813728, + 0.005494144279509783, + 0.014223617501556873, + 0.0027612759731709957, + 0.00563549529761076, + 0.0033167230430990458, + 0.004540976136922836, + 0.003605534788221121, + 0.006046752445399761, + 0.005092034116387367, + 0.006485820282250643, + 0.007159773260354996, + 0.007620315533131361, + 0.008440972305834293, + 0.006741642020642757, + 0.005138305481523275, + 0.004673477727919817, + 0.005783203057944775, + 0.006992744747549295, + 0.013503177091479301, + 0.004444536752998829, + 0.009308482520282269, + 0.007424565497785807, + 0.003905955236405134, + 0.006822964642196894, + 0.005682742223143578, + 0.009067974053323269, + 0.007144962437450886, + 0.007102516014128923, + 0.009481113404035568, + 0.00971605721861124, + 0.010103371925652027, + 0.011269740760326385, + 0.004974433220922947, + 0.014230350963771343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0032107490114867687 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001717998180538416 + ], + "opposite_logprob_tokens": [ + -6.376718044281006 + ], + "logprob_reward": -0.001717998180538416, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001717998180538416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08388693630695343, + 0.039835426956415176, + 0.033593498170375824, + 0.04459709674119949, + 0.11183740198612213, + 0.14495372772216797, + 0.03482174500823021, + 0.05383313074707985, + 0.05453013256192207, + 0.39811089634895325 + ], + "credit_map": [ + 0.00925737340003252, + 0.005730221047997475, + 0.008339975029230118, + 0.005638440605252981, + 0.0038145037833601236, + 0.018792415037751198, + 0.005178913939744234, + 0.027135096490383148, + 0.004348925314843655, + 0.004539939574897289, + 0.004537784028798342, + 0.003322668606415391, + 0.005749307572841644, + 0.006623614579439163, + 0.0038480572402477264, + 0.006865133065730333, + 0.0034646582789719105, + 0.0027224640361964703, + 0.003843883750960231, + 0.0031937824096530676, + 0.0025387455243617296, + 0.006788705009967089, + 0.0027844018768519163, + 0.008256858214735985, + 0.0037418378051370382, + 0.006155830807983875, + 0.0037615522742271423, + 0.0051006111316382885, + 0.003685825504362583, + 0.007801522500813007, + 0.0036840892862528563, + 0.010665826499462128, + 0.00997193343937397, + 0.011989092454314232, + 0.015777435153722763, + 0.011331290006637573, + 0.011371713131666183, + 0.01861490309238434, + 0.010677402839064598, + 0.022103630006313324, + 0.011693399399518967, + 0.017283804714679718, + 0.01330247987061739, + 0.011465239338576794, + 0.01556185819208622, + 0.030130047351121902, + 0.01096884161233902, + 0.03454805538058281, + 0.002803558250889182, + 0.0036850031465291977, + 0.005465064197778702, + 0.004222281277179718, + 0.0026363551151007414, + 0.005524756386876106, + 0.004351773764938116, + 0.0061329505406320095, + 0.008459102362394333, + 0.005147238727658987, + 0.007718648295849562, + 0.006009179167449474, + 0.006328917108476162, + 0.006370062008500099, + 0.0072646671906113625, + 0.006535313557833433, + 0.007727338466793299, + 0.006320200394839048, + 0.0069349673576653, + 0.005386096425354481, + 0.005610235035419464, + 0.006049522664397955, + 0.00812850147485733, + 0.008373270742595196, + 0.03445732221007347, + 0.04922442138195038, + 0.048945117741823196, + 0.051792316138744354, + 0.07630576938390732, + 0.049361228942871094, + 0.04546337574720383, + 0.042561355978250504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.001717998180538416 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628476247191429 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628476247191429, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628476247191429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07016883045434952, + 0.05509817600250244, + 0.05049728974699974, + 0.037681709975004196, + 0.08197398483753204, + 0.12051858752965927, + 0.049514345824718475, + 0.19840262830257416, + 0.06307706236839294, + 0.27306732535362244 + ], + "credit_map": [ + 0.009996537119150162, + 0.005310999229550362, + 0.0066144587472081184, + 0.006116834003478289, + 0.0037102247588336468, + 0.017247838899493217, + 0.004139884375035763, + 0.01703205704689026, + 0.006233528256416321, + 0.004856666550040245, + 0.00647088885307312, + 0.005091439466923475, + 0.003306480823084712, + 0.012003696523606777, + 0.005391256418079138, + 0.011744214221835136, + 0.006077589932829142, + 0.0052041709423065186, + 0.004737601149827242, + 0.005716927349567413, + 0.005088475067168474, + 0.008498793467879295, + 0.005796576850116253, + 0.009377156384289265, + 0.003416478866711259, + 0.0051781246438622475, + 0.003437555395066738, + 0.0035788321401923895, + 0.0037680091336369514, + 0.007254511117935181, + 0.0036096512340009212, + 0.00743854371830821, + 0.006597528699785471, + 0.007914945483207703, + 0.009073253720998764, + 0.009390147402882576, + 0.00868531409651041, + 0.01583278737962246, + 0.006703386083245277, + 0.017776621505618095, + 0.006226843222975731, + 0.009111188352108002, + 0.006039079744368792, + 0.00937145296484232, + 0.007068568374961615, + 0.02892703004181385, + 0.012661277316510677, + 0.041113149374723434, + 0.0040288823656737804, + 0.004894013050943613, + 0.005116008687764406, + 0.005106307100504637, + 0.0034774772357195616, + 0.00961596891283989, + 0.00551370345056057, + 0.011761986650526524, + 0.04111817851662636, + 0.023898694664239883, + 0.03069659322500229, + 0.022735167294740677, + 0.015729479491710663, + 0.01869121752679348, + 0.01763421855866909, + 0.027899060398340225, + 0.008649535477161407, + 0.0038276046980172396, + 0.007765090558677912, + 0.009133681654930115, + 0.006049505900591612, + 0.00878470204770565, + 0.010433714836835861, + 0.008433232083916664, + 0.025640562176704407, + 0.032985858619213104, + 0.04196092113852501, + 0.02957702986896038, + 0.042683202773332596, + 0.03754206746816635, + 0.03145570307970047, + 0.031221991404891014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.007628476247191429 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809018433094025 + ], + "opposite_logprob_tokens": [ + -3.288090229034424 + ], + "logprob_reward": -0.03809018433094025, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809018433094025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042132262140512466, + 0.04218209907412529, + 0.11231855303049088, + 0.05731865391135216, + 0.07111743092536926, + 0.15551471710205078, + 0.042218130081892014, + 0.31105396151542664, + 0.08092208206653595, + 0.085222028195858 + ], + "credit_map": [ + 0.004215848632156849, + 0.0035344327334314585, + 0.005315608810633421, + 0.003936342429369688, + 0.0035740670282393694, + 0.008442891761660576, + 0.00424680020660162, + 0.00886627659201622, + 0.0044963848777115345, + 0.00319362198933959, + 0.004818854853510857, + 0.0033583908807486296, + 0.002687874250113964, + 0.009288618341088295, + 0.004832977429032326, + 0.009505381807684898, + 0.012944241985678673, + 0.008393369615077972, + 0.01203412376344204, + 0.008803267031908035, + 0.009604970924556255, + 0.023841973394155502, + 0.00952885951846838, + 0.027167750522494316, + 0.005414182785898447, + 0.006284250877797604, + 0.007208176422864199, + 0.005361203569918871, + 0.004216174129396677, + 0.010721593163907528, + 0.004422795493155718, + 0.013690279796719551, + 0.004235553089529276, + 0.007838401943445206, + 0.007017737720161676, + 0.007484177127480507, + 0.005728508811444044, + 0.014738427475094795, + 0.00582141475751996, + 0.01825321838259697, + 0.010755243711173534, + 0.016009846702218056, + 0.008777146227657795, + 0.013240250758826733, + 0.011384627781808376, + 0.03329604119062424, + 0.017836155369877815, + 0.044215425848960876, + 0.003138030180707574, + 0.0041249399073421955, + 0.004168480169028044, + 0.004044432193040848, + 0.003838113509118557, + 0.008537599816918373, + 0.0037720848340541124, + 0.010594450868666172, + 0.06295360624790192, + 0.037028584629297256, + 0.04566173255443573, + 0.03673454374074936, + 0.02575233206152916, + 0.02976432628929615, + 0.02982601523399353, + 0.043332867324352264, + 0.012074551545083523, + 0.009360719472169876, + 0.010088875889778137, + 0.009850786067545414, + 0.007812900468707085, + 0.010026982985436916, + 0.010366681963205338, + 0.011340586468577385, + 0.008800804615020752, + 0.010011396370828152, + 0.014985964633524418, + 0.011801758781075478, + 0.008920272812247276, + 0.01123614888638258, + 0.008096265606582165, + 0.011369435116648674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.03809018433094025 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, initially static, focuses on a piece of yellow talisman paper suspended against a dark, blurred background. Its texture is rough, and upon it, intricate, twisting, and powerfully drawn mystical runes in vibrant vermilion red. Suddenly, without warning, a fiery red dot ignites at the center of the paper. Immediately, a brilliant orange-red flame bursts forth, ravenously devouring the paper. The edges of the paper rapidly curl and char under the licking of the flames, while the red runes glow with an extraordinary brilliance in the firelight, as if instantly animated. Then, as if lifted up by an invisible force, the entire burning talisman paper spins and struggles upward. The camera rapidly tilts upward, closely tracking its movement, capturing the dynamic moment as it leaves behind a trail of dancing, flickering black embers. The image is characterized by a stark contrast of light and shadow, with the flames acting as the sole, dynamic source of light, casting a violent, dancing shadow against the pitch-black background, creating a tense atmosphere imbued with a sense of mysterious ritual and the eruption of supernatural power.", + "seed": 15681, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_1/index_8491_seed_15681_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04020326954778284, + "vlm_logprob": -0.04020326954778284, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.04020326954778284, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0077173179015517235 + ], + "opposite_logprob_tokens": [ + -5.507717132568359 + ], + "logprob_reward": -0.0077173179015517235, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0077173179015517235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05945216491818428, + 0.03255263343453407, + 0.02392548881471157, + 0.08809614926576614, + 0.03570639714598656, + 0.0319291315972805, + 0.035466939210891724, + 0.02856074646115303, + 0.5424034595489502, + 0.1219068095088005 + ], + "credit_map": [ + 0.005535465199500322, + 0.011009913869202137, + 0.008150829002261162, + 0.005371833220124245, + 0.00614436948671937, + 0.005922986660152674, + 0.007455263752490282, + 0.009861505590379238, + 0.0039488207548856735, + 0.006557459477335215, + 0.0034680471289902925, + 0.004063136409968138, + 0.002922872779890895, + 0.0040661669336259365, + 0.003910061437636614, + 0.0036160729359835386, + 0.0022163987159729004, + 0.003014899557456374, + 0.0038188013713806868, + 0.0021425148006528616, + 0.0015693261520937085, + 0.0029317205771803856, + 0.0029284763149917126, + 0.005303350277245045, + 0.0032272054813802242, + 0.004745494574308395, + 0.003304754151031375, + 0.0023127892054617405, + 0.010505473241209984, + 0.03493119031190872, + 0.0030797028448432684, + 0.02598954550921917, + 0.0030804313719272614, + 0.004534702282398939, + 0.004811322316527367, + 0.002740988042205572, + 0.006951941642910242, + 0.0030517950654029846, + 0.0032861297950148582, + 0.007249091286212206, + 0.0028169266879558563, + 0.005556440446525812, + 0.0028459737077355385, + 0.0028737331740558147, + 0.003806990571320057, + 0.0026757565792649984, + 0.007519364822655916, + 0.0038339465390890837, + 0.004578264430165291, + 0.004494565073400736, + 0.004940177779644728, + 0.00410539610311389, + 0.004347082693129778, + 0.004489601589739323, + 0.0037199503276497126, + 0.0047919051721692085, + 0.0026806106325238943, + 0.0025811409577727318, + 0.005687321536242962, + 0.0037646268028765917, + 0.0034278053790330887, + 0.005314546637237072, + 0.002039162674918771, + 0.0030655364971607924, + 0.040136925876140594, + 0.04321574047207832, + 0.21872878074645996, + 0.0179490614682436, + 0.11984217911958694, + 0.038722191005945206, + 0.022000711411237717, + 0.04180789366364479, + 0.005745444446802139, + 0.020835742354393005, + 0.014565320685505867, + 0.018658222630620003, + 0.02168416790664196, + 0.0199330635368824, + 0.011453337036073208, + 0.009031515568494797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.0077173179015517235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028330450877547264 + ], + "opposite_logprob_tokens": [ + -5.877832889556885 + ], + "logprob_reward": -0.0028330450877547264, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028330450877547264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08972478657960892, + 0.040443334728479385, + 0.34913793206214905, + 0.11636226624250412, + 0.022876063361763954, + 0.021729296073317528, + 0.02040555141866207, + 0.07338976114988327, + 0.22377517819404602, + 0.04215580224990845 + ], + "credit_map": [ + 0.008647790178656578, + 0.01425733882933855, + 0.010142664425075054, + 0.009277844801545143, + 0.009119539521634579, + 0.010988143272697926, + 0.010939136147499084, + 0.016352331265807152, + 0.00486541585996747, + 0.006145252380520105, + 0.0064359549432992935, + 0.004701313562691212, + 0.004078656435012817, + 0.0046248892322182655, + 0.006498339120298624, + 0.003093513660132885, + 0.031845822930336, + 0.04026119410991669, + 0.04259375482797623, + 0.030919602140784264, + 0.022305574268102646, + 0.06430503726005554, + 0.0418875627219677, + 0.07501939684152603, + 0.004004733171314001, + 0.006590178702026606, + 0.003641448449343443, + 0.0035317956935614347, + 0.014418144710361958, + 0.04323963448405266, + 0.004059777595102787, + 0.036876555532217026, + 0.0018848661566153169, + 0.0044024428352713585, + 0.0027156658470630646, + 0.0018222385551780462, + 0.004408110398799181, + 0.0014412158634513617, + 0.002373805269598961, + 0.0038277183193713427, + 0.001722989953123033, + 0.0033049732446670532, + 0.0017772995634004474, + 0.0019748329650610685, + 0.002737105591222644, + 0.0020210477523505688, + 0.005468157585710287, + 0.002722889883443713, + 0.0024277446791529655, + 0.003740378189831972, + 0.002362377941608429, + 0.002688042353838682, + 0.0024501194711774588, + 0.002157752402126789, + 0.002368292538449168, + 0.0022108445409685373, + 0.007424534764140844, + 0.006390847265720367, + 0.018404757604002953, + 0.007737999781966209, + 0.005882542580366135, + 0.01663394831120968, + 0.004748170729726553, + 0.006166953593492508, + 0.016176048666238785, + 0.022143177688121796, + 0.08239113539457321, + 0.01115654781460762, + 0.045109812170267105, + 0.022229669615626335, + 0.010755978524684906, + 0.013812793418765068, + 0.002286049071699381, + 0.007851770147681236, + 0.007237827870994806, + 0.004902700427919626, + 0.007263844832777977, + 0.005445779301226139, + 0.003727043280377984, + 0.00344078685157001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.0028330450877547264 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008143446175381541 + ], + "opposite_logprob_tokens": [ + -7.125814437866211 + ], + "logprob_reward": -0.0008143446175381541, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008143446175381541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024532945826649666, + 0.0060382140800356865, + 0.05456170067191124, + 0.08521310985088348, + 0.008261755108833313, + 0.006429044064134359, + 0.007324600126594305, + 0.018663574010133743, + 0.7704278826713562, + 0.01854727230966091 + ], + "credit_map": [ + 0.002643004758283496, + 0.004775143228471279, + 0.0025150699075311422, + 0.003198039485141635, + 0.003069839673116803, + 0.0024131142999976873, + 0.002743640448898077, + 0.003175092628225684, + 0.0007044518715701997, + 0.0005702115595340729, + 0.0012843699660152197, + 0.0007420090842060745, + 0.0005581422592513263, + 0.000774221436586231, + 0.0008269978570751846, + 0.0005778102786280215, + 0.004781460855156183, + 0.006897094659507275, + 0.006207393482327461, + 0.005229823291301727, + 0.003376150503754616, + 0.008977783843874931, + 0.0065348208881914616, + 0.01255717035382986, + 0.0034675467759370804, + 0.0049354927614331245, + 0.0028399841394275427, + 0.0023815648164600134, + 0.010345892049372196, + 0.03365502879023552, + 0.002712908899411559, + 0.024874694645404816, + 0.000773602572735399, + 0.001173243043012917, + 0.0010587300639599562, + 0.001023405697196722, + 0.0011110069463029504, + 0.0008751978166401386, + 0.001140162581577897, + 0.001106406794860959, + 0.0005010021268390119, + 0.0009929946390911937, + 0.0005506753223016858, + 0.0006590585107915103, + 0.0008402215898968279, + 0.0007002688944339752, + 0.0015582038322463632, + 0.0006266189157031476, + 0.0007145751733332872, + 0.0015668157720938325, + 0.0009292420581914485, + 0.0010405931388959289, + 0.0006513865664601326, + 0.0008107932517305017, + 0.0009463864844292402, + 0.000664807390421629, + 0.001738492283038795, + 0.001565204351209104, + 0.004847620148211718, + 0.001888152095489204, + 0.0013261402491480112, + 0.004520594608038664, + 0.001350313308648765, + 0.0014270570827648044, + 0.058104682713747025, + 0.06347166746854782, + 0.3077273368835449, + 0.025296088308095932, + 0.1672569066286087, + 0.05245833471417427, + 0.027267111465334892, + 0.0688457041978836, + 0.001339745125733316, + 0.003398460103198886, + 0.002949845278635621, + 0.0020837411284446716, + 0.0026317888405174017, + 0.0027499301359057426, + 0.0017439915100112557, + 0.001649769488722086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.0008143446175381541 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781407192349434 + ], + "opposite_logprob_tokens": [ + -4.634781360626221 + ], + "logprob_reward": -0.009781407192349434, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781407192349434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0037972612772136927, + 0.0018066936172544956, + 0.010974926874041557, + 0.008590463548898697, + 0.002972321817651391, + 0.003601125907152891, + 0.008365229703485966, + 0.024498000741004944, + 0.918009877204895, + 0.017384136095643044 + ], + "credit_map": [ + 0.0003586517123039812, + 0.0004556986677926034, + 0.00048356910701841116, + 0.0004222100251354277, + 0.00029735671705566347, + 0.0005370412254706025, + 0.0004276386462152004, + 0.0008150950307026505, + 0.00021307187853381038, + 0.0002220531168859452, + 0.0002469326136633754, + 0.00022187734430190176, + 0.00019802755559794605, + 0.0002163867757190019, + 0.00023557597887702286, + 0.00025276836822740734, + 0.0009253047755919397, + 0.0012849653139710426, + 0.0013688262552022934, + 0.0009972502011805773, + 0.0007113460451364517, + 0.0020785522647202015, + 0.0013036741875112057, + 0.002305008238181472, + 0.00038310146192088723, + 0.00045641345786862075, + 0.00060255330754444, + 0.00027062444132752717, + 0.0011759052285924554, + 0.003524801693856716, + 0.0003468622453510761, + 0.0018302013631910086, + 0.0002452420012559742, + 0.00046652855235151947, + 0.00035526190185919404, + 0.00031680939719080925, + 0.0002887866576202214, + 0.0004151507746428251, + 0.0005047040758654475, + 0.00037983834045007825, + 0.000302960368571803, + 0.0007392576662823558, + 0.00030484970193356276, + 0.0003644842654466629, + 0.00049104995559901, + 0.00033993771648965776, + 0.0006693066097795963, + 0.0003892796812579036, + 0.0012245222460478544, + 0.0013164920965209603, + 0.0009376987582072616, + 0.0009946272475644946, + 0.0008376689511351287, + 0.0008396710036322474, + 0.0010282722068950534, + 0.0011862767860293388, + 0.002540480811148882, + 0.0027841648552566767, + 0.004556262865662575, + 0.002994864247739315, + 0.002123757964000106, + 0.004027540795505047, + 0.002304122317582369, + 0.003166807582601905, + 0.06969230622053146, + 0.0755218043923378, + 0.3661311864852905, + 0.029510481283068657, + 0.19907480478286743, + 0.06047607213258743, + 0.033021874725818634, + 0.08458133786916733, + 0.0010201401310041547, + 0.003132758429273963, + 0.002419266849756241, + 0.0022033669520169497, + 0.0026406131219118834, + 0.00322099681943655, + 0.0015615384327247739, + 0.001185456058010459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.009781407192349434 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987023293972015 + ], + "opposite_logprob_tokens": [ + -1.8048702478408813 + ], + "logprob_reward": -0.17987023293972015, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987023293972015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031023560091853142, + 0.008162139914929867, + 0.05158822610974312, + 0.07853784412145615, + 0.023043086752295494, + 0.025654364377260208, + 0.022896090522408485, + 0.1601431965827942, + 0.3050391972064972, + 0.2939123511314392 + ], + "credit_map": [ + 0.0029638276901096106, + 0.003950695041567087, + 0.0033013804350048304, + 0.0035725028719753027, + 0.0024988865479826927, + 0.004177904687821865, + 0.004105636849999428, + 0.00645272433757782, + 0.0011257908772677183, + 0.0009107703808695078, + 0.0011924289865419269, + 0.0011293657589703798, + 0.0009121476905420423, + 0.0007706350297667086, + 0.0012792842462658882, + 0.0008417172357439995, + 0.004858955275267363, + 0.006107962224632502, + 0.006227784790098667, + 0.004748418927192688, + 0.003129919059574604, + 0.0094009879976511, + 0.006068756803870201, + 0.011045441031455994, + 0.00680119451135397, + 0.0065987794660031796, + 0.006734016351401806, + 0.0058954739943146706, + 0.010911600664258003, + 0.01506543904542923, + 0.0062118717469275, + 0.020319458097219467, + 0.0015573179116472602, + 0.0023187173064798117, + 0.002951858565211296, + 0.002053431700915098, + 0.005018417723476887, + 0.00221161893568933, + 0.0029808501712977886, + 0.003950873855501413, + 0.002337956801056862, + 0.004844461567699909, + 0.00247291405685246, + 0.0024514575488865376, + 0.0030166106298565865, + 0.0030746576376259327, + 0.004164963029325008, + 0.0032913428731262684, + 0.0036839155945926905, + 0.0030872616916894913, + 0.0028366141486912966, + 0.002755040768533945, + 0.0021956833079457283, + 0.0027702467050403357, + 0.002595816273242235, + 0.0029715094715356827, + 0.018523795530200005, + 0.017095638439059258, + 0.02407616749405861, + 0.015555557794868946, + 0.017225774005055428, + 0.020898405462503433, + 0.01859852485358715, + 0.02816932462155819, + 0.023205291479825974, + 0.02100185677409172, + 0.11759066581726074, + 0.011098655872046947, + 0.0647842213511467, + 0.022991623729467392, + 0.011771371588110924, + 0.03259548917412758, + 0.018206343054771423, + 0.06493605673313141, + 0.03590709716081619, + 0.04053638130426407, + 0.052636757493019104, + 0.03580908104777336, + 0.026396475732326508, + 0.01948416791856289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.17987023293972015 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, initially static, focuses on a piece of yellow talisman paper suspended against a dark, blurred background. Its texture is rough, and upon it, intricate, twisting, and powerfully drawn mystical runes in vibrant vermilion red. Suddenly, without warning, a fiery red dot ignites at the center of the paper. Immediately, a brilliant orange-red flame bursts forth, ravenously devouring the paper. The edges of the paper rapidly curl and char under the licking of the flames, while the red runes glow with an extraordinary brilliance in the firelight, as if instantly animated. Then, as if lifted up by an invisible force, the entire burning talisman paper spins and struggles upward. The camera rapidly tilts upward, closely tracking its movement, capturing the dynamic moment as it leaves behind a trail of dancing, flickering black embers. The image is characterized by a stark contrast of light and shadow, with the flames acting as the sole, dynamic source of light, casting a violent, dancing shadow against the pitch-black background, creating a tense atmosphere imbued with a sense of mysterious ritual and the eruption of supernatural power.", + "seed": 15682, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_1/index_8491_seed_15682_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03956486666575074, + "vlm_logprob": -0.03956486666575074, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.03956486666575074, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007669053506106138 + ], + "opposite_logprob_tokens": [ + -6.382668972015381 + ], + "logprob_reward": -0.007669053506106138, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.007669053506106138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30928197503089905, + 0.0537465363740921, + 0.041560377925634384, + 0.06536109745502472, + 0.0809021145105362, + 0.04901568591594696, + 0.07680290937423706, + 0.1162296012043953, + 0.06264447420835495, + 0.14445528388023376 + ], + "credit_map": [ + 0.03213313966989517, + 0.05343960225582123, + 0.04147402569651604, + 0.04922850802540779, + 0.026325898244976997, + 0.032684165984392166, + 0.03600245714187622, + 0.0379941500723362, + 0.005566764157265425, + 0.00783116277307272, + 0.008092799223959446, + 0.009407689794898033, + 0.005571491084992886, + 0.005867285188287497, + 0.006724108010530472, + 0.004685231018811464, + 0.004064954351633787, + 0.005932262167334557, + 0.008997146971523762, + 0.006993006449192762, + 0.0041575804352760315, + 0.002844145754352212, + 0.005109827965497971, + 0.0034614482428878546, + 0.007967068813741207, + 0.010568328201770782, + 0.009713700041174889, + 0.015437938272953033, + 0.005787784233689308, + 0.004935496021062136, + 0.005043081473559141, + 0.005907695274800062, + 0.00867204274982214, + 0.018794558942317963, + 0.010666748508810997, + 0.014465557411313057, + 0.007141719106584787, + 0.006087945308536291, + 0.007154669612646103, + 0.007918862625956535, + 0.005087774246931076, + 0.00908477883785963, + 0.006010835058987141, + 0.007058474700897932, + 0.004744900390505791, + 0.006884749513119459, + 0.0038835247978568077, + 0.006260645110160112, + 0.012298278510570526, + 0.015441852621734142, + 0.01739838719367981, + 0.008913780562579632, + 0.005934101063758135, + 0.006216145120561123, + 0.0059198783710598946, + 0.004680481739342213, + 0.010553822852671146, + 0.03059478849172592, + 0.020312558859586716, + 0.01932654343545437, + 0.012713950127363205, + 0.007839455269277096, + 0.006684309337288141, + 0.00820416770875454, + 0.006841809488832951, + 0.007145519368350506, + 0.010048752650618553, + 0.010770987719297409, + 0.008638002909719944, + 0.005335066933184862, + 0.00928127858787775, + 0.00458305049687624, + 0.01030885148793459, + 0.02236032299697399, + 0.021643485873937607, + 0.019132433459162712, + 0.017957856878638268, + 0.013739308342337608, + 0.023731088265776634, + 0.015581931918859482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.007669053506106138 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019516960019245744 + ], + "opposite_logprob_tokens": [ + -6.251951694488525 + ], + "logprob_reward": -0.0019516960019245744, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019516960019245744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1895730346441269, + 0.13728965818881989, + 0.04458554461598396, + 0.17462711036205292, + 0.09601007401943207, + 0.0783524140715599, + 0.05651845410466194, + 0.09096450358629227, + 0.05485918000340462, + 0.07722006738185883 + ], + "credit_map": [ + 0.01766228675842285, + 0.02252766862511635, + 0.03671470284461975, + 0.029599880799651146, + 0.016017084941267967, + 0.01723860390484333, + 0.02635776624083519, + 0.023455029353499413, + 0.011914599686861038, + 0.021331144496798515, + 0.019068783149123192, + 0.02512725070118904, + 0.015445638447999954, + 0.011606782674789429, + 0.018770264461636543, + 0.014025173150002956, + 0.0050642602145671844, + 0.008038808591663837, + 0.009223954752087593, + 0.008386274799704552, + 0.0037468820810317993, + 0.0024917321279644966, + 0.004346955101937056, + 0.0032866739202290773, + 0.016197241842746735, + 0.024624746292829514, + 0.02117348648607731, + 0.01820218563079834, + 0.03150191158056259, + 0.024773769080638885, + 0.021285802125930786, + 0.016867948696017265, + 0.012645797803997993, + 0.024725686758756638, + 0.010889925062656403, + 0.013784278184175491, + 0.008912907913327217, + 0.00793677382171154, + 0.009015749208629131, + 0.008098955266177654, + 0.008162851445376873, + 0.016249792650341988, + 0.012522776611149311, + 0.014613011851906776, + 0.007113853469491005, + 0.006861351430416107, + 0.007076870184391737, + 0.005751898977905512, + 0.008013191632926464, + 0.007252413779497147, + 0.012056836858391762, + 0.012280370108783245, + 0.0041391500271856785, + 0.004191211424767971, + 0.0048306779935956, + 0.0037545973900705576, + 0.00842527486383915, + 0.018029071390628815, + 0.012353678233921528, + 0.01417160127311945, + 0.014807095751166344, + 0.010845193639397621, + 0.007465984206646681, + 0.004866594914346933, + 0.0042983596213161945, + 0.009130510501563549, + 0.008521230891346931, + 0.007517017424106598, + 0.014597875066101551, + 0.00289725000038743, + 0.00550844008103013, + 0.002388496184721589, + 0.005606483668088913, + 0.011899705976247787, + 0.010539653711020947, + 0.010819838382303715, + 0.013461507856845856, + 0.005927417427301407, + 0.011019980534911156, + 0.00794546864926815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0019516960019245744 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015148365637287498 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015148365637287498, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015148365637287498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14396549761295319, + 0.03194050490856171, + 0.04447408765554428, + 0.1598903238773346, + 0.11875033378601074, + 0.10753999650478363, + 0.10842546075582504, + 0.11355910450220108, + 0.10294659435749054, + 0.06850812584161758 + ], + "credit_map": [ + 0.013823792338371277, + 0.015443463809788227, + 0.020025022327899933, + 0.022562900558114052, + 0.014171368442475796, + 0.017580481246113777, + 0.02041345089673996, + 0.019945012405514717, + 0.0028163280803710222, + 0.005368565674871206, + 0.006582838483154774, + 0.006176090333610773, + 0.0026862802915275097, + 0.0020852109882980585, + 0.003692182246595621, + 0.00253300741314888, + 0.005157193168997765, + 0.0076530505903065205, + 0.007895250804722309, + 0.01007300615310669, + 0.0034986233804374933, + 0.0024631437845528126, + 0.004361459985375404, + 0.0033723553642630577, + 0.01578189618885517, + 0.0314054861664772, + 0.022105690091848373, + 0.020039722323417664, + 0.015929924324154854, + 0.016935350373387337, + 0.018034078180789948, + 0.019658176228404045, + 0.017311232164502144, + 0.03454955667257309, + 0.014123577624559402, + 0.01711275801062584, + 0.010929105803370476, + 0.00806321483105421, + 0.008992193266749382, + 0.0076686916872859, + 0.01167361717671156, + 0.01753423549234867, + 0.021655837073922157, + 0.015296075493097305, + 0.012699409388005733, + 0.01144635584205389, + 0.009524146094918251, + 0.0077103134244680405, + 0.017423413693904877, + 0.017342165112495422, + 0.01718173921108246, + 0.02693254128098488, + 0.00686381570994854, + 0.009005021303892136, + 0.00783954281359911, + 0.00583722535520792, + 0.009166577830910683, + 0.022123344242572784, + 0.016479961574077606, + 0.016650334000587463, + 0.018799414858222008, + 0.010962927713990211, + 0.01178518682718277, + 0.0075913541950285435, + 0.0069405315443873405, + 0.018933244049549103, + 0.017365187406539917, + 0.019576476886868477, + 0.01798628829419613, + 0.00579884834587574, + 0.011977467685937881, + 0.004368546884506941, + 0.004086171742528677, + 0.009120117872953415, + 0.009015250019729137, + 0.011022951453924179, + 0.010597321204841137, + 0.006476593669503927, + 0.009729394689202309, + 0.008460324257612228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.0015148365637287498 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634979411959648 + ], + "opposite_logprob_tokens": [ + -3.6513497829437256 + ], + "logprob_reward": -0.02634979411959648, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634979411959648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11527720093727112, + 0.0453481562435627, + 0.03395973518490791, + 0.12267943471670151, + 0.07153116166591644, + 0.08580494672060013, + 0.11529341340065002, + 0.09947054833173752, + 0.161790668964386, + 0.14884477853775024 + ], + "credit_map": [ + 0.013943224214017391, + 0.017183545976877213, + 0.017339026555418968, + 0.019673550501465797, + 0.009890474379062653, + 0.012516440823674202, + 0.013008535839617252, + 0.011722392402589321, + 0.004593381658196449, + 0.0066857850179076195, + 0.009335639886558056, + 0.007320299278944731, + 0.0036088498309254646, + 0.003751752432435751, + 0.0062499199993908405, + 0.0038025283720344305, + 0.003565286286175251, + 0.006047020200639963, + 0.008054457604885101, + 0.005813999101519585, + 0.003057061228901148, + 0.0021325023844838142, + 0.0026515687350183725, + 0.002637840574607253, + 0.011203918606042862, + 0.02268245443701744, + 0.01696496270596981, + 0.01656956784427166, + 0.018256280571222305, + 0.014543729834258556, + 0.011364663951098919, + 0.011093850247561932, + 0.006503789685666561, + 0.012172949500381947, + 0.009177741594612598, + 0.011321833357214928, + 0.0059838686138391495, + 0.009099126793444157, + 0.010336057282984257, + 0.006935792043805122, + 0.006975224241614342, + 0.01308676227927208, + 0.012917439453303814, + 0.012665076181292534, + 0.010261524468660355, + 0.012213725596666336, + 0.008438427932560444, + 0.009246760979294777, + 0.01579403504729271, + 0.015592224895954132, + 0.02790604718029499, + 0.02779126912355423, + 0.006759209092706442, + 0.006834793835878372, + 0.008288856595754623, + 0.006326983217149973, + 0.007283370476216078, + 0.024771634489297867, + 0.014225759543478489, + 0.016213102266192436, + 0.011914501897990704, + 0.00722912373021245, + 0.00996227003633976, + 0.007870794273912907, + 0.015531781129539013, + 0.026918785646557808, + 0.02648809552192688, + 0.02500881254673004, + 0.025996802374720573, + 0.010960178449749947, + 0.021149568259716034, + 0.00973664503544569, + 0.009617126546800137, + 0.01963098905980587, + 0.026547592133283615, + 0.023355133831501007, + 0.02269711345434189, + 0.015194647014141083, + 0.019197560846805573, + 0.012604606337845325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.02634979411959648 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033895313739777 + ], + "opposite_logprob_tokens": [ + -1.9103389978408813 + ], + "logprob_reward": -0.16033895313739777, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033895313739777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08489351719617844, + 0.02695900946855545, + 0.01755381003022194, + 0.10913390666246414, + 0.2301572561264038, + 0.06534881889820099, + 0.04592021182179451, + 0.11115240305662155, + 0.21434378623962402, + 0.09453728795051575 + ], + "credit_map": [ + 0.007955996319651604, + 0.012980892322957516, + 0.013514047488570213, + 0.014669199474155903, + 0.007951103150844574, + 0.007549986243247986, + 0.010839242488145828, + 0.009433039464056492, + 0.0026013324968516827, + 0.003557630581781268, + 0.004878576844930649, + 0.00494415033608675, + 0.0029710044618695974, + 0.0027033996302634478, + 0.0028921442572027445, + 0.0024107671342790127, + 0.0018247116822749376, + 0.0027766814455389977, + 0.003764090593904257, + 0.0043239290826022625, + 0.0014041783288121223, + 0.0009065101621672511, + 0.0015343724517151713, + 0.0010193344205617905, + 0.010187662206590176, + 0.02033500000834465, + 0.014468776993453503, + 0.020106205716729164, + 0.013773524202406406, + 0.010799103416502476, + 0.009847995825111866, + 0.009615626186132431, + 0.02929871343076229, + 0.06633298844099045, + 0.026347707957029343, + 0.035918377339839935, + 0.021793125197291374, + 0.014803517609834671, + 0.017145924270153046, + 0.01851688139140606, + 0.005774049088358879, + 0.011328207328915596, + 0.008539934642612934, + 0.010930702090263367, + 0.00738368509337306, + 0.008913721889257431, + 0.006297459360212088, + 0.006181048694998026, + 0.007034584879875183, + 0.008603034541010857, + 0.008204104378819466, + 0.007184110581874847, + 0.004123638849705458, + 0.003888725535944104, + 0.0038997759111225605, + 0.002982232952490449, + 0.010939047671854496, + 0.022901736199855804, + 0.014068585820496082, + 0.01698080264031887, + 0.015357539057731628, + 0.013166826218366623, + 0.008371497504413128, + 0.009366360493004322, + 0.018589144572615623, + 0.03327096253633499, + 0.035197604447603226, + 0.026732247322797775, + 0.055650245398283005, + 0.009813185781240463, + 0.024144239723682404, + 0.010946134105324745, + 0.005810068920254707, + 0.01587323099374771, + 0.016466937959194183, + 0.01672426424920559, + 0.017266442999243736, + 0.007061289623379707, + 0.008968672715127468, + 0.006366366986185312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.16033895313739777 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, initially static, focuses on a piece of yellow talisman paper suspended against a dark, blurred background. Its texture is rough, and upon it, intricate, twisting, and powerfully drawn mystical runes in vibrant vermilion red. Suddenly, without warning, a fiery red dot ignites at the center of the paper. Immediately, a brilliant orange-red flame bursts forth, ravenously devouring the paper. The edges of the paper rapidly curl and char under the licking of the flames, while the red runes glow with an extraordinary brilliance in the firelight, as if instantly animated. Then, as if lifted up by an invisible force, the entire burning talisman paper spins and struggles upward. The camera rapidly tilts upward, closely tracking its movement, capturing the dynamic moment as it leaves behind a trail of dancing, flickering black embers. The image is characterized by a stark contrast of light and shadow, with the flames acting as the sole, dynamic source of light, casting a violent, dancing shadow against the pitch-black background, creating a tense atmosphere imbued with a sense of mysterious ritual and the eruption of supernatural power.", + "seed": 15683, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_1/index_8491_seed_15683_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.015022567007690668, + "vlm_logprob": -0.015022567007690668, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.015022567007690668, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016519788652658463 + ], + "opposite_logprob_tokens": [ + -5.016520023345947 + ], + "logprob_reward": -0.016519788652658463, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.016519788652658463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.403530478477478, + 0.05737966299057007, + 0.08565238863229752, + 0.08828330785036087, + 0.04535491392016411, + 0.03106478601694107, + 0.035528913140296936, + 0.031693488359451294, + 0.08470562845468521, + 0.1368064135313034 + ], + "credit_map": [ + 0.01587245985865593, + 0.013901319354772568, + 0.01904793083667755, + 0.010278650559484959, + 0.01946241408586502, + 0.1439228355884552, + 0.023759756237268448, + 0.15728513896465302, + 0.0062509868294000626, + 0.006575069855898619, + 0.007609672844409943, + 0.005114204715937376, + 0.011248762719333172, + 0.005600378382951021, + 0.00854997057467699, + 0.006430627312511206, + 0.007153471931815147, + 0.011420565657317638, + 0.01792910508811474, + 0.011455575935542583, + 0.012366659939289093, + 0.006184146273881197, + 0.010423670522868633, + 0.008719196543097496, + 0.007972277700901031, + 0.014655203558504581, + 0.014876600354909897, + 0.011780236847698689, + 0.014224023558199406, + 0.0063833435997366905, + 0.010804472491145134, + 0.0075871641747653484, + 0.007520446088165045, + 0.0052296495996415615, + 0.005402769893407822, + 0.005886916071176529, + 0.005036299582570791, + 0.005250436719506979, + 0.005905980709940195, + 0.005122420843690634, + 0.002514546038582921, + 0.003117355750873685, + 0.0036229470279067755, + 0.003555371891707182, + 0.004057008307427168, + 0.00470011867582798, + 0.005176999140530825, + 0.004320445004850626, + 0.004473437555134296, + 0.00786313321441412, + 0.004643253982067108, + 0.0033853845670819283, + 0.004057629965245724, + 0.0036182727199047804, + 0.00430145813152194, + 0.003186348592862487, + 0.004358699545264244, + 0.0024377936497330666, + 0.008656932972371578, + 0.004036460537463427, + 0.0034900037571787834, + 0.0029658223502337933, + 0.0035649603232741356, + 0.002182821510359645, + 0.009493113495409489, + 0.008157474920153618, + 0.009782489389181137, + 0.010940625332295895, + 0.006420561112463474, + 0.0066002290695905685, + 0.02220960147678852, + 0.011101546697318554, + 0.009451721794903278, + 0.01982048712670803, + 0.03139019384980202, + 0.024060936644673347, + 0.009880159050226212, + 0.016791583970189095, + 0.010333680547773838, + 0.015077668242156506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.016519788652658463 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005969671066850424 + ], + "opposite_logprob_tokens": [ + -5.130969524383545 + ], + "logprob_reward": -0.005969671066850424, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005969671066850424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10947650671005249, + 0.03480254486203194, + 0.06215200200676918, + 0.07088679820299149, + 0.019740531221032143, + 0.03371085599064827, + 0.03474331647157669, + 0.06134195625782013, + 0.06430673599243164, + 0.5088388323783875 + ], + "credit_map": [ + 0.003362063318490982, + 0.004925079643726349, + 0.005475468002259731, + 0.004700574092566967, + 0.006105375476181507, + 0.038821104913949966, + 0.0061110686510801315, + 0.03997576981782913, + 0.0022863345220685005, + 0.00547028286382556, + 0.0038103193510323763, + 0.0033106249757111073, + 0.009465038776397705, + 0.0028545144014060497, + 0.004138552583754063, + 0.0034668755251914263, + 0.0029469840228557587, + 0.010398121550679207, + 0.00852007046341896, + 0.009968534111976624, + 0.013712147250771523, + 0.0039087990298867226, + 0.008755565620958805, + 0.003941771574318409, + 0.0048975530080497265, + 0.01272624172270298, + 0.008752267807722092, + 0.014226427301764488, + 0.009895932860672474, + 0.005249088164418936, + 0.009367023594677448, + 0.00577226048335433, + 0.002331043593585491, + 0.002351756440475583, + 0.002906671492382884, + 0.0028920096810907125, + 0.0019750786013901234, + 0.0019486091332510114, + 0.0032219705171883106, + 0.002113389316946268, + 0.003082211362197995, + 0.0033798818476498127, + 0.0034495401196181774, + 0.003219080623239279, + 0.005726235453039408, + 0.005017366260290146, + 0.004972951021045446, + 0.0048635876737535, + 0.003930045757442713, + 0.005544380750507116, + 0.0041963993571698666, + 0.003528145607560873, + 0.0036266811657696962, + 0.0041995421051979065, + 0.0052191512659192085, + 0.004498965572565794, + 0.006945009808987379, + 0.005754626356065273, + 0.009045382030308247, + 0.007466989103704691, + 0.010533506982028484, + 0.006735814735293388, + 0.009383569471538067, + 0.005477048456668854, + 0.0062835970893502235, + 0.004511660430580378, + 0.006312192417681217, + 0.008850417099893093, + 0.005216045770794153, + 0.006921655032783747, + 0.0185422170907259, + 0.00766894593834877, + 0.018030643463134766, + 0.059537868946790695, + 0.01719360239803791, + 0.012252574786543846, + 0.005781868007034063, + 0.21407021582126617, + 0.010145161300897598, + 0.17182683944702148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.005969671066850424 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036196685396134853 + ], + "opposite_logprob_tokens": [ + -5.62861967086792 + ], + "logprob_reward": -0.0036196685396134853, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036196685396134853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16372837126255035, + 0.06608635932207108, + 0.08485648781061172, + 0.05609813332557678, + 0.042070284485816956, + 0.055419597774744034, + 0.10190733522176743, + 0.04793078824877739, + 0.31083670258522034, + 0.0710659772157669 + ], + "credit_map": [ + 0.010199657641351223, + 0.012924810871481895, + 0.014034997671842575, + 0.007142886985093355, + 0.015351234935224056, + 0.04769367352128029, + 0.009908569045364857, + 0.04647252336144447, + 0.00682627409696579, + 0.008124873042106628, + 0.009741164743900299, + 0.006672329269349575, + 0.012596781365573406, + 0.006248437333852053, + 0.008958039805293083, + 0.00691845640540123, + 0.00552587304264307, + 0.012715645134449005, + 0.009741025976836681, + 0.011804718524217606, + 0.01678103394806385, + 0.009214500896632671, + 0.01276713889092207, + 0.006306544411927462, + 0.0058111888356506824, + 0.007948766462504864, + 0.007768500130623579, + 0.007992596365511417, + 0.009068531915545464, + 0.004583609756082296, + 0.006359931547194719, + 0.006565001327544451, + 0.005944629665464163, + 0.005685870535671711, + 0.006809774320572615, + 0.005371734965592623, + 0.002888313727453351, + 0.0047806124202907085, + 0.004458395764231682, + 0.006130949594080448, + 0.0045124986208975315, + 0.005292325746268034, + 0.005564708262681961, + 0.008304531686007977, + 0.010279623791575432, + 0.005560624413192272, + 0.010954036377370358, + 0.004951245617121458, + 0.014434022828936577, + 0.02001241222023964, + 0.01122105773538351, + 0.01033378578722477, + 0.010810624808073044, + 0.013035956770181656, + 0.009974800981581211, + 0.0120846601203084, + 0.006201352924108505, + 0.00405513821169734, + 0.006715500261634588, + 0.005116458050906658, + 0.007752198725938797, + 0.005257113836705685, + 0.007351541891694069, + 0.005481477826833725, + 0.031190810725092888, + 0.024753820151090622, + 0.03279554843902588, + 0.046673599630594254, + 0.02717847004532814, + 0.027606436982750893, + 0.08485417813062668, + 0.035783831030130386, + 0.0048652333207428455, + 0.011353865265846252, + 0.009278335608541965, + 0.011494919657707214, + 0.005815580952912569, + 0.011127226054668427, + 0.005518579855561256, + 0.01161222718656063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.0036196685396134853 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950355436652899 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950355436652899, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950355436652899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05190398544073105, + 0.011781909503042698, + 0.015114915557205677, + 0.053317222744226456, + 0.009250740520656109, + 0.020276017487049103, + 0.017401479184627533, + 0.03414367139339447, + 0.08185158669948578, + 0.7049585580825806 + ], + "credit_map": [ + 0.0035267930943518877, + 0.0026634528767317533, + 0.0032321724575012922, + 0.0017483679112046957, + 0.0035498985089361668, + 0.01657327078282833, + 0.0045175570994615555, + 0.016092471778392792, + 0.0009344802820123732, + 0.0017080920515581965, + 0.0021791348699480295, + 0.0008385125547647476, + 0.001977799693122506, + 0.0010410392424091697, + 0.0011262535117566586, + 0.001976596424356103, + 0.0010464825900271535, + 0.002280906541272998, + 0.0015561791369691491, + 0.003273882670328021, + 0.0022918819449841976, + 0.001221038750372827, + 0.0023758402094244957, + 0.0010687025496736169, + 0.004126089159399271, + 0.011374436318874359, + 0.0064234365709125996, + 0.008488736115396023, + 0.008649395778775215, + 0.0038725058548152447, + 0.006402818486094475, + 0.003979803062975407, + 0.00117413979023695, + 0.0008607780328020453, + 0.0012606632662937045, + 0.0018119843443855643, + 0.0009178210748359561, + 0.00093834288418293, + 0.0013539701467379928, + 0.0009330408647656441, + 0.0020258070435374975, + 0.001972184283658862, + 0.002434366848319769, + 0.0023003853857517242, + 0.0029865591786801815, + 0.002262665657326579, + 0.0038710939697921276, + 0.0024229527916759253, + 0.002003550063818693, + 0.002248496748507023, + 0.0021563530899584293, + 0.0019006315851584077, + 0.0017893871990963817, + 0.0022671795450150967, + 0.0025951433926820755, + 0.002440738258883357, + 0.0045477477833628654, + 0.003650355851277709, + 0.004650214221328497, + 0.005093655548989773, + 0.004787105601280928, + 0.0038717384450137615, + 0.0043541062623262405, + 0.0031887453515082598, + 0.010329928249120712, + 0.008554910309612751, + 0.009198424406349659, + 0.012731953524053097, + 0.005618126597255468, + 0.004277192521840334, + 0.020051486790180206, + 0.011089559644460678, + 0.02585024945437908, + 0.07913573831319809, + 0.022857530042529106, + 0.01647254079580307, + 0.008410481736063957, + 0.29647982120513916, + 0.01411498710513115, + 0.24163711071014404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.005950355436652899 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a piece of yellow talisman paper with rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are intricate twisting mystical runes drawn in vibrant vermilion red on the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fiery red dot ignite at the center of the paper followed by an orange-red flame that curls and chars the edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the burning paper spin and struggle upward as if lifted by an invisible force while leaving a trail of black embers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rapidly tilt upward to track the paper's movement against a dark, blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305335134267807 + ], + "opposite_logprob_tokens": [ + -3.168053388595581 + ], + "logprob_reward": -0.04305335134267807, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305335134267807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020436841994524002, + 0.006668437737971544, + 0.006090404465794563, + 0.011917322874069214, + 0.004190071485936642, + 0.0065795425325632095, + 0.0070490967482328415, + 0.005346468649804592, + 0.014710589312016964, + 0.9170112013816833 + ], + "credit_map": [ + 0.0012481547892093658, + 0.00179491366725415, + 0.001463916851207614, + 0.0008196410490199924, + 0.0018418997060507536, + 0.005868540145456791, + 0.0016181954415515065, + 0.00578157976269722, + 0.0005825425032526255, + 0.0010213643545284867, + 0.0008358772029168904, + 0.0006364160217344761, + 0.0017018594080582261, + 0.00050271803047508, + 0.0006853415397927165, + 0.0007023182697594166, + 0.0006194608868099749, + 0.000956260075327009, + 0.0007613450870849192, + 0.0010363331530243158, + 0.0008071605116128922, + 0.000601523439399898, + 0.0007837709854356945, + 0.0005245492793619633, + 0.0007962574600242078, + 0.00231318362057209, + 0.001299142837524414, + 0.0020410879515111446, + 0.0018821912817656994, + 0.000901171937584877, + 0.0018155649304389954, + 0.0008687219815328717, + 0.00040387859917245805, + 0.0005167550989426672, + 0.0006524921045638621, + 0.0005137633415870368, + 0.00044308151700533926, + 0.0004525359545368701, + 0.0007238775724545121, + 0.0004836872685700655, + 0.0008885156712494791, + 0.0007256632088683546, + 0.0007405707146972418, + 0.0007140678353607655, + 0.0008370117284357548, + 0.000724239565897733, + 0.0010926320683211088, + 0.0008568409830331802, + 0.001023080083541572, + 0.0009602909558452666, + 0.0007996331551112235, + 0.0006834791856817901, + 0.0007103186217136681, + 0.0010632768971845508, + 0.0008965280721895397, + 0.0009124890784732997, + 0.0006603210931643844, + 0.00046134856529533863, + 0.0010510887950658798, + 0.0007734622922725976, + 0.0006397451506927609, + 0.000521932786796242, + 0.0008168048807419837, + 0.0004217643290758133, + 0.0016595714259892702, + 0.0016329046338796616, + 0.0021852324716746807, + 0.002241556067019701, + 0.001256274525076151, + 0.0011745770461857319, + 0.0029865733813494444, + 0.0015738982474431396, + 0.03290533646941185, + 0.1034313514828682, + 0.027355140075087547, + 0.01693408004939556, + 0.010067024268209934, + 0.38995736837387085, + 0.018083179369568825, + 0.3182776868343353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.04305335134267807 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_10/index_tensor([7475, 7475])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_10/index_tensor([7475, 7475])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2a477c331358143f6496b30fa93081e171c62328 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_10/index_tensor([7475, 7475])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "macro close-up shot from a level angle draws the viewer's gaze to a tranquil corner of the forest floor. A ray of warm afternoon sunlight penetrates the dense canopy, casting a soft beam that illuminates the center of the scene. The camera focuses on a young fawn, its light brown fur speckled with white appearing exceptionally soft and fluffy in the light. Its large, obsidian eyes gleam with innocence. Cautiously, it approaches a rotting log covered in emerald green moss. Atop this decaying log, a massive, fan-shaped, crimson Ganoderma lucidum thrives. Its surface is as smooth as lacquer, its edges gleaming with a golden sheen, revealing distinct rings of growth. The fawn's moist black nose gently touches the edge of the mushroom, and then, with its soft lips and tiny teeth, it delicately and tentatively nibbles at a corner of the cap, its movements filled with innocence and concentration. The entire background of the picture is blurred by the shallow depth of field into a dreamy green and brown halo. Rich details, saturated warm tones and soft light and shadow are intertwined, giving this quiet and vital moment a thick and delicate texture and atmosphere like a classical oil painting.", + "seed": 13411, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_10/index_7475_seed_13411_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02443294257391244, + "vlm_logprob": -0.02443294257391244, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.02443294257391244, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05615842342376709 + ], + "opposite_logprob_tokens": [ + -2.9311585426330566 + ], + "logprob_reward": -0.05615842342376709, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05615842342376709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33160534501075745, + 0.054328057914972305, + 0.02516677789390087, + 0.02217770740389824, + 0.025332478806376457, + 0.055691909044981, + 0.05214909091591835, + 0.036518145352602005, + 0.2056899517774582, + 0.19134053587913513 + ], + "credit_map": [ + 0.02325003407895565, + 0.041463643312454224, + 0.03883666172623634, + 0.038872115314006805, + 0.03645507991313934, + 0.04226924106478691, + 0.06320367753505707, + 0.04725490137934685, + 0.005619517527520657, + 0.00629773736000061, + 0.006727712228894234, + 0.006358277518302202, + 0.005409479606896639, + 0.006672096438705921, + 0.005788412876427174, + 0.011454824358224869, + 0.0035291502717882395, + 0.0028778205160051584, + 0.0029590954072773457, + 0.0022116745822131634, + 0.002761684823781252, + 0.0036210266407579184, + 0.0037290905602276325, + 0.0034772350918501616, + 0.0039885942824184895, + 0.002192505169659853, + 0.0026330717373639345, + 0.002230859361588955, + 0.0013237098464742303, + 0.0033013734500855207, + 0.002463751705363393, + 0.0040438417345285416, + 0.005100839305669069, + 0.0029123034328222275, + 0.003439861349761486, + 0.001726042479276657, + 0.002202589064836502, + 0.003657808294519782, + 0.0025983289815485477, + 0.0036947045009583235, + 0.005305956117808819, + 0.012942963279783726, + 0.004029444418847561, + 0.004651089198887348, + 0.009623981080949306, + 0.009160999208688736, + 0.0048513393849134445, + 0.005126137286424637, + 0.006758897099643946, + 0.009945055469870567, + 0.006355978548526764, + 0.003926984034478664, + 0.007112070918083191, + 0.007704166229814291, + 0.004322053398936987, + 0.006023883353918791, + 0.005335616879165173, + 0.005850957240909338, + 0.004685464780777693, + 0.002963112201541662, + 0.004457402508705854, + 0.005160363391041756, + 0.0035232885275036097, + 0.004541940055787563, + 0.025320949032902718, + 0.01994636468589306, + 0.02978919818997383, + 0.01442610751837492, + 0.03739580512046814, + 0.03370990604162216, + 0.014474724419414997, + 0.030626889318227768, + 0.02360476739704609, + 0.030880799517035484, + 0.024156466126441956, + 0.019942210987210274, + 0.02600444108247757, + 0.023012610152363777, + 0.015401906333863735, + 0.028337350115180016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.05615842342376709 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020606013014912605 + ], + "opposite_logprob_tokens": [ + -3.89560604095459 + ], + "logprob_reward": -0.020606013014912605, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020606013014912605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2933160364627838, + 0.09116757661104202, + 0.15608814358711243, + 0.08523650467395782, + 0.07293757796287537, + 0.0698242112994194, + 0.06003229320049286, + 0.04568304866552353, + 0.09276558458805084, + 0.03294895589351654 + ], + "credit_map": [ + 0.016467567533254623, + 0.032306741923093796, + 0.02756437286734581, + 0.05430545285344124, + 0.048506028950214386, + 0.05695417895913124, + 0.03408362716436386, + 0.023128099739551544, + 0.008853740058839321, + 0.007997151464223862, + 0.010666804388165474, + 0.008263974450528622, + 0.009139715693891048, + 0.010573250241577625, + 0.015077388845384121, + 0.02059556171298027, + 0.02077374793589115, + 0.011593508534133434, + 0.020213982090353966, + 0.015584132634103298, + 0.0125889228656888, + 0.021093076094985008, + 0.027117254212498665, + 0.02712353691458702, + 0.008914198726415634, + 0.005160405300557613, + 0.00959188211709261, + 0.006633987184613943, + 0.005690783262252808, + 0.014383750036358833, + 0.010927807539701462, + 0.02393369935452938, + 0.006520284339785576, + 0.004530046600848436, + 0.00995673518627882, + 0.0038607127498835325, + 0.004811086226254702, + 0.01331953052431345, + 0.005383615382015705, + 0.024555571377277374, + 0.007449377328157425, + 0.0045892540365457535, + 0.00755669642239809, + 0.0037124345544725657, + 0.005307105835527182, + 0.013007279485464096, + 0.00795197393745184, + 0.020250098779797554, + 0.005600085482001305, + 0.0071206046268343925, + 0.004569936078041792, + 0.006942401640117168, + 0.007916624657809734, + 0.009067784994840622, + 0.008383502252399921, + 0.010431360453367233, + 0.004273167811334133, + 0.005294944159686565, + 0.004467211663722992, + 0.003842836245894432, + 0.006465437356382608, + 0.006470766384154558, + 0.009511356242001057, + 0.005357332527637482, + 0.01139928586781025, + 0.007319409400224686, + 0.013264241628348827, + 0.006544425152242184, + 0.012503793463110924, + 0.015110094100236893, + 0.01104808785021305, + 0.015576254576444626, + 0.004463539458811283, + 0.0029473425820469856, + 0.0036379520315676928, + 0.003213010262697935, + 0.003068049205467105, + 0.0041823917999863625, + 0.00424809567630291, + 0.007188576273620129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.020606013014912605 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956517532467842 + ], + "opposite_logprob_tokens": [ + -5.130956649780273 + ], + "logprob_reward": -0.005956517532467842, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956517532467842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2676890790462494, + 0.12648837268352509, + 0.11284787952899933, + 0.05023988336324692, + 0.04678305238485336, + 0.04887329787015915, + 0.03833165764808655, + 0.08026956766843796, + 0.12957638502120972, + 0.09890078753232956 + ], + "credit_map": [ + 0.020860571414232254, + 0.0336821973323822, + 0.028134116902947426, + 0.03271741047501564, + 0.03426988795399666, + 0.04398651793599129, + 0.04391644895076752, + 0.030121952295303345, + 0.01661895401775837, + 0.011651895008981228, + 0.012670998461544514, + 0.012201362289488316, + 0.012875626794993877, + 0.015869205817580223, + 0.02122514322400093, + 0.023375194519758224, + 0.008701752871274948, + 0.009267478249967098, + 0.011291936971247196, + 0.01220499724149704, + 0.00875884760171175, + 0.016231980174779892, + 0.0301792211830616, + 0.01621166616678238, + 0.005192552227526903, + 0.0032515088096261024, + 0.005160519387573004, + 0.004900062456727028, + 0.004884129855781794, + 0.007289595436304808, + 0.009800871834158897, + 0.009760644286870956, + 0.005315765738487244, + 0.007441991474479437, + 0.005433857440948486, + 0.005641818977892399, + 0.004908625036478043, + 0.006405862048268318, + 0.0053631155751645565, + 0.006272014230489731, + 0.006864514201879501, + 0.005075099412351847, + 0.00429501524195075, + 0.003947865683585405, + 0.006158816162496805, + 0.007293274626135826, + 0.00733005627989769, + 0.007908657193183899, + 0.005355921108275652, + 0.005616762675344944, + 0.003166674403473735, + 0.0041301073506474495, + 0.004846951924264431, + 0.004854296334087849, + 0.0046160463243722916, + 0.005744897294789553, + 0.009333999827504158, + 0.008134658448398113, + 0.01062117237597704, + 0.007155910599976778, + 0.013458065688610077, + 0.010896909050643444, + 0.009666240774095058, + 0.011002613231539726, + 0.01285192184150219, + 0.012744187377393246, + 0.014385281130671501, + 0.010191169567406178, + 0.02358745038509369, + 0.020644797012209892, + 0.015674393624067307, + 0.019497182220220566, + 0.014278856106102467, + 0.009761422872543335, + 0.010344141162931919, + 0.012975498102605343, + 0.010320124216377735, + 0.012223590165376663, + 0.013578496873378754, + 0.01541865337640047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.005956517532467842 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810521960258484 + ], + "opposite_logprob_tokens": [ + -3.2881052494049072 + ], + "logprob_reward": -0.03810521960258484, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810521960258484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19865252077579498, + 0.12422310560941696, + 0.05361662060022354, + 0.10936015844345093, + 0.017877178266644478, + 0.04186386615037918, + 0.10724547505378723, + 0.04023032635450363, + 0.17260748147964478, + 0.13432328402996063 + ], + "credit_map": [ + 0.013607623986899853, + 0.021897748112678528, + 0.02153829298913479, + 0.022843467071652412, + 0.028724053874611855, + 0.03275265544652939, + 0.034568481147289276, + 0.022720186039805412, + 0.015863751992583275, + 0.017086079344153404, + 0.012200825847685337, + 0.015218638814985752, + 0.014199163764715195, + 0.014977137558162212, + 0.01576364040374756, + 0.018913863226771355, + 0.0070841070264577866, + 0.005320187192410231, + 0.005861235316842794, + 0.0064603956416249275, + 0.0049421172589063644, + 0.007781316060572863, + 0.00958574190735817, + 0.00658152112737298, + 0.0123904375359416, + 0.01407938078045845, + 0.01199012715369463, + 0.011327533051371574, + 0.009909987449645996, + 0.015774168074131012, + 0.01788386143743992, + 0.016004659235477448, + 0.0015723748365417123, + 0.0025340598076581955, + 0.0018995529972016811, + 0.0016629834426566958, + 0.002041490050032735, + 0.002965726191177964, + 0.0018510944209992886, + 0.0033498958218842745, + 0.0033419260289520025, + 0.0055712307803332806, + 0.0027302114758640528, + 0.0041548023000359535, + 0.00611609173938632, + 0.006566019728779793, + 0.005866148043423891, + 0.007517436519265175, + 0.018371421843767166, + 0.015693848952651024, + 0.0144592160359025, + 0.008094053715467453, + 0.014621349051594734, + 0.0153878815472126, + 0.008408385328948498, + 0.012209323234856129, + 0.004690289497375488, + 0.0044300914742052555, + 0.0043050264939665794, + 0.0031689205206930637, + 0.0051954565569758415, + 0.005527029745280743, + 0.007599326316267252, + 0.00531418714672327, + 0.01932455226778984, + 0.01571952924132347, + 0.02081667073071003, + 0.012829951010644436, + 0.032616693526506424, + 0.027861155569553375, + 0.015982581302523613, + 0.02745635062456131, + 0.018200112506747246, + 0.019000910222530365, + 0.017867447808384895, + 0.014478595927357674, + 0.018344929441809654, + 0.016225220635533333, + 0.012004801072180271, + 0.018201259896159172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.03810521960258484 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013385392958298326 + ], + "opposite_logprob_tokens": [ + -6.626338481903076 + ], + "logprob_reward": -0.0013385392958298326, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013385392958298326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29208695888519287, + 0.04410579055547714, + 0.058161891996860504, + 0.02135772816836834, + 0.031382251530885696, + 0.04117215797305107, + 0.07820342481136322, + 0.11322940140962601, + 0.21817027032375336, + 0.10213018953800201 + ], + "credit_map": [ + 0.019730297848582268, + 0.027641264721751213, + 0.035200148820877075, + 0.04396994039416313, + 0.03995892032980919, + 0.042329054325819016, + 0.05070311203598976, + 0.032554205507040024, + 0.004383973777294159, + 0.004726012237370014, + 0.005267067346721888, + 0.006610010284930468, + 0.004270188510417938, + 0.005417200271040201, + 0.0071606324054300785, + 0.006270706187933683, + 0.007323288358747959, + 0.0047868769615888596, + 0.005641142372041941, + 0.0063486951403319836, + 0.004610009025782347, + 0.008136932738125324, + 0.011162950657308102, + 0.010151995345950127, + 0.0027238107286393642, + 0.0021826003212481737, + 0.002579390536993742, + 0.0020362602081149817, + 0.00186476344242692, + 0.0033637501765042543, + 0.002123915823176503, + 0.004483236465603113, + 0.007087404374033213, + 0.002938001649454236, + 0.004199417773634195, + 0.0022543056402355433, + 0.0026232388336211443, + 0.004628143738955259, + 0.0029594909865409136, + 0.004692248068749905, + 0.004089064430445433, + 0.00823465920984745, + 0.003639067290350795, + 0.003115743398666382, + 0.004790568724274635, + 0.007149914279580116, + 0.004106413573026657, + 0.00604673009365797, + 0.00823252648115158, + 0.019343918189406395, + 0.006468928884714842, + 0.006936113815754652, + 0.0107799107208848, + 0.012130005285143852, + 0.006129444111138582, + 0.008182581514120102, + 0.011972957290709019, + 0.013046542182564735, + 0.013709373772144318, + 0.009733702056109905, + 0.020697150379419327, + 0.015487764962017536, + 0.013376720249652863, + 0.015205194242298603, + 0.021726613864302635, + 0.018261639401316643, + 0.024885116145014763, + 0.017020398750901222, + 0.039868541061878204, + 0.035647857934236526, + 0.02061309479176998, + 0.04014700651168823, + 0.013223029673099518, + 0.01574939675629139, + 0.013331455178558826, + 0.011144183576107025, + 0.013699496164917946, + 0.012159706093370914, + 0.008141027763485909, + 0.014681898057460785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.0013385392958298326 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "macro close-up shot from a level angle draws the viewer's gaze to a tranquil corner of the forest floor. A ray of warm afternoon sunlight penetrates the dense canopy, casting a soft beam that illuminates the center of the scene. The camera focuses on a young fawn, its light brown fur speckled with white appearing exceptionally soft and fluffy in the light. Its large, obsidian eyes gleam with innocence. Cautiously, it approaches a rotting log covered in emerald green moss. Atop this decaying log, a massive, fan-shaped, crimson Ganoderma lucidum thrives. Its surface is as smooth as lacquer, its edges gleaming with a golden sheen, revealing distinct rings of growth. The fawn's moist black nose gently touches the edge of the mushroom, and then, with its soft lips and tiny teeth, it delicately and tentatively nibbles at a corner of the cap, its movements filled with innocence and concentration. The entire background of the picture is blurred by the shallow depth of field into a dreamy green and brown halo. Rich details, saturated warm tones and soft light and shadow are intertwined, giving this quiet and vital moment a thick and delicate texture and atmosphere like a classical oil painting.", + "seed": 13412, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_10/index_7475_seed_13412_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01721502486616373, + "vlm_logprob": -0.01721502486616373, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.01721502486616373, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06323875486850739 + ], + "opposite_logprob_tokens": [ + -2.8132388591766357 + ], + "logprob_reward": -0.06323875486850739, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06323875486850739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1930568665266037, + 0.16690844297409058, + 0.036952801048755646, + 0.2507486343383789, + 0.1296650618314743, + 0.11680576950311661, + 0.030930135399103165, + 0.019391175359487534, + 0.0206978227943182, + 0.034843314439058304 + ], + "credit_map": [ + 0.024740077555179596, + 0.027515774592757225, + 0.03055546060204506, + 0.027918115258216858, + 0.03489168733358383, + 0.01419546827673912, + 0.018321489915251732, + 0.014918793924152851, + 0.01382444053888321, + 0.018852917477488518, + 0.015393689274787903, + 0.020058097317814827, + 0.027560414746403694, + 0.015812629833817482, + 0.03411237522959709, + 0.02129388228058815, + 0.0056748175993561745, + 0.0028402768075466156, + 0.007307068444788456, + 0.004447586834430695, + 0.002862986410036683, + 0.004972707014530897, + 0.005059120710939169, + 0.0037882383912801743, + 0.008671109564602375, + 0.006426245905458927, + 0.10096867382526398, + 0.010264914482831955, + 0.00946925487369299, + 0.10368628799915314, + 0.004761758726090193, + 0.0065003931522369385, + 0.008892307989299297, + 0.016955753788352013, + 0.015076009556651115, + 0.012692439369857311, + 0.034496013075113297, + 0.018412107601761818, + 0.013090771622955799, + 0.010049661621451378, + 0.0074892351403832436, + 0.013767119497060776, + 0.015811938792467117, + 0.01939530111849308, + 0.016359902918338776, + 0.01932232268154621, + 0.011172889731824398, + 0.013487058691680431, + 0.0023798232432454824, + 0.004479498136788607, + 0.0037213091272860765, + 0.004024793393909931, + 0.005076930392533541, + 0.004247201606631279, + 0.0037706159055233, + 0.003229965455830097, + 0.0019468922400847077, + 0.0031098255421966314, + 0.00362385343760252, + 0.001844488549977541, + 0.002330299699679017, + 0.0026125165168195963, + 0.0015582736814394593, + 0.002365024061873555, + 0.0014721907209604979, + 0.0018115498824045062, + 0.0037413081154227257, + 0.0028528275433927774, + 0.0023111163172870874, + 0.0024204233195632696, + 0.003885726211592555, + 0.0022026796359568834, + 0.004394079092890024, + 0.004271392710506916, + 0.004108793567866087, + 0.004413815215229988, + 0.005848933011293411, + 0.003993954975157976, + 0.003978785593062639, + 0.0038335626013576984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.06323875486850739 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007644920144230127 + ], + "opposite_logprob_tokens": [ + -4.882645130157471 + ], + "logprob_reward": -0.007644920144230127, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007644920144230127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06815452873706818, + 0.057984087616205215, + 0.036173198372125626, + 0.3167279362678528, + 0.04991504177451134, + 0.06503048539161682, + 0.06684872508049011, + 0.03369119390845299, + 0.0817401334643364, + 0.22373469173908234 + ], + "credit_map": [ + 0.007596781477332115, + 0.008815078996121883, + 0.010416326113045216, + 0.011615879833698273, + 0.010099304839968681, + 0.005037163849920034, + 0.00838173646479845, + 0.00619225949048996, + 0.006261673290282488, + 0.00775409210473299, + 0.005868521053344011, + 0.006582027301192284, + 0.007115147542208433, + 0.006702069193124771, + 0.009785347618162632, + 0.00791520532220602, + 0.004337330814450979, + 0.0022948405239731073, + 0.008813043124973774, + 0.0029533375054597855, + 0.0028872231487184763, + 0.009038236923515797, + 0.0028951214626431465, + 0.002954061608761549, + 0.010403686203062534, + 0.007611373905092478, + 0.12840965390205383, + 0.01339242048561573, + 0.01201584655791521, + 0.13236108422279358, + 0.006126198917627335, + 0.006407697219401598, + 0.003976836334913969, + 0.008334990590810776, + 0.007322446443140507, + 0.005321206524968147, + 0.006701374892145395, + 0.005922966171056032, + 0.005971159785985947, + 0.0063640642911195755, + 0.003925067838281393, + 0.008099481463432312, + 0.008414514362812042, + 0.01100901048630476, + 0.008243540301918983, + 0.010279644280672073, + 0.006647743750363588, + 0.00841148104518652, + 0.006959529593586922, + 0.006811183411628008, + 0.011191662400960922, + 0.008241147734224796, + 0.008755901828408241, + 0.010517521761357784, + 0.00542097445577383, + 0.008950804360210896, + 0.0036735846661031246, + 0.005358540918678045, + 0.004875998478382826, + 0.0035424039233475924, + 0.004577420186251402, + 0.004598059691488743, + 0.0034413260873407125, + 0.0036238597240298986, + 0.00606686482205987, + 0.00645072665065527, + 0.013259626924991608, + 0.010921807028353214, + 0.011692062951624393, + 0.01004544086754322, + 0.015092997811734676, + 0.008210602216422558, + 0.01636640913784504, + 0.01165399607270956, + 0.020585697144269943, + 0.022201983258128166, + 0.016256986185908318, + 0.029429873451590538, + 0.07583336532115936, + 0.03140636533498764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.007644920144230127 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744122598320246 + ], + "opposite_logprob_tokens": [ + -5.006743907928467 + ], + "logprob_reward": -0.006744122598320246, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744122598320246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04569244757294655, + 0.06110232695937157, + 0.03241964057087898, + 0.14190317690372467, + 0.0434417761862278, + 0.12150849401950836, + 0.08536873757839203, + 0.030296839773654938, + 0.3208889961242676, + 0.11737754940986633 + ], + "credit_map": [ + 0.003486241912469268, + 0.004336988553404808, + 0.005761534906923771, + 0.006863982882350683, + 0.006240265443921089, + 0.005929330829530954, + 0.0065071615390479565, + 0.006566946394741535, + 0.009732883423566818, + 0.006338819861412048, + 0.009439549408853054, + 0.006437535863369703, + 0.008498444221913815, + 0.007030141074210405, + 0.007594545371830463, + 0.006030413322150707, + 0.004657622426748276, + 0.0025780806317925453, + 0.0053853741846978664, + 0.003140780609101057, + 0.0024565833155065775, + 0.005300940480083227, + 0.00479253102093935, + 0.004107734654098749, + 0.0054940832778811455, + 0.003574968548491597, + 0.05640462040901184, + 0.007535181473940611, + 0.005290305707603693, + 0.05686027556657791, + 0.003573697293177247, + 0.0031700439285486937, + 0.0031546533573418856, + 0.006277026142925024, + 0.00486217625439167, + 0.0041727665811777115, + 0.005307825747877359, + 0.005068001803010702, + 0.0066667282953858376, + 0.007932599633932114, + 0.006965226028114557, + 0.013683353550732136, + 0.016196412965655327, + 0.019791722297668457, + 0.016100401058793068, + 0.019444260746240616, + 0.012600875459611416, + 0.016726244240999222, + 0.0056997076608240604, + 0.009454408660531044, + 0.010900481604039669, + 0.01175608765333891, + 0.012688905000686646, + 0.011115332134068012, + 0.012827527709305286, + 0.01092628762125969, + 0.0029283708427101374, + 0.004653048235923052, + 0.004333650227636099, + 0.002971609588712454, + 0.0038208281621336937, + 0.004194965120404959, + 0.004395098891109228, + 0.0029992691706866026, + 0.022439610213041306, + 0.02692164108157158, + 0.05469846352934837, + 0.0502033494412899, + 0.03564558923244476, + 0.03613339737057686, + 0.06343124061822891, + 0.03141571953892708, + 0.008968685753643513, + 0.008134232833981514, + 0.010202341713011265, + 0.013881382532417774, + 0.010061687789857388, + 0.015015803277492523, + 0.0373365618288517, + 0.013776860199868679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -0.006744122598320246 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631433662027121 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631433662027121, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631433662027121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11775398254394531, + 0.10979893058538437, + 0.1163051426410675, + 0.20060475170612335, + 0.10326473414897919, + 0.14114753901958466, + 0.04292569309473038, + 0.03284943476319313, + 0.09298480302095413, + 0.04236500710248947 + ], + "credit_map": [ + 0.014437745325267315, + 0.017575837671756744, + 0.0194176584482193, + 0.015612518414855003, + 0.02096129208803177, + 0.008477798663079739, + 0.011800440959632397, + 0.009470692835748196, + 0.009875459596514702, + 0.012096491642296314, + 0.010331333614885807, + 0.012878235429525375, + 0.02002960629761219, + 0.009594910778105259, + 0.021279113367199898, + 0.013713779859244823, + 0.014890950173139572, + 0.010390933603048325, + 0.02133476547896862, + 0.010221380740404129, + 0.013103187084197998, + 0.016793331131339073, + 0.019184602424502373, + 0.010385986417531967, + 0.007769158110022545, + 0.007980676367878914, + 0.07263673841953278, + 0.008842331357300282, + 0.01279661525040865, + 0.07480145245790482, + 0.007734236773103476, + 0.008043531328439713, + 0.007380844093859196, + 0.01653304696083069, + 0.011306511238217354, + 0.009676378220319748, + 0.023088427260518074, + 0.013496564701199532, + 0.011266874149441719, + 0.010516084730625153, + 0.008907784707844257, + 0.016012635082006454, + 0.019027888774871826, + 0.023361865431070328, + 0.019324762746691704, + 0.022676892578601837, + 0.014287987723946571, + 0.017547717317938805, + 0.0027104930486530066, + 0.004894710145890713, + 0.005460577085614204, + 0.004996675997972488, + 0.0063599287532269955, + 0.006294489372521639, + 0.005714464001357555, + 0.006494356784969568, + 0.00309031642973423, + 0.004112737253308296, + 0.006832440849393606, + 0.003086469369009137, + 0.0037626016419380903, + 0.004577640909701586, + 0.003787865862250328, + 0.003599361749365926, + 0.006020665634423494, + 0.009980378672480583, + 0.012460489757359028, + 0.01979854516685009, + 0.013069627806544304, + 0.01067840401083231, + 0.012734792195260525, + 0.008241903968155384, + 0.0038727312348783016, + 0.004574597347527742, + 0.004073392134159803, + 0.004086509346961975, + 0.006699859630316496, + 0.0050341966561973095, + 0.00868594367057085, + 0.005337781738489866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.007631433662027121 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008158930577337742 + ], + "opposite_logprob_tokens": [ + -7.1258158683776855 + ], + "logprob_reward": -0.0008158930577337742, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008158930577337742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14512032270431519, + 0.09874136000871658, + 0.04556483030319214, + 0.38479068875312805, + 0.06909807026386261, + 0.11312044411897659, + 0.03939596191048622, + 0.024152392521500587, + 0.04088899493217468, + 0.039126962423324585 + ], + "credit_map": [ + 0.01585901528596878, + 0.02109689638018608, + 0.027250591665506363, + 0.020126821473240852, + 0.020183812826871872, + 0.013839617371559143, + 0.013364318758249283, + 0.013399257324635983, + 0.009878582321107388, + 0.008940910920500755, + 0.009077802300453186, + 0.013842892833054066, + 0.0132168959826231, + 0.01101172436028719, + 0.020347557961940765, + 0.012424997054040432, + 0.005750209093093872, + 0.003317416412755847, + 0.009394883178174496, + 0.004664380103349686, + 0.0034182709641754627, + 0.00815250538289547, + 0.005489223171025515, + 0.005377941764891148, + 0.013040868565440178, + 0.0082682054489851, + 0.15976455807685852, + 0.016006138175725937, + 0.01137873250991106, + 0.1631746143102646, + 0.004555841907858849, + 0.008601722307503223, + 0.004892166703939438, + 0.009933711029589176, + 0.007717057131230831, + 0.007117109838873148, + 0.015838827937841415, + 0.00904966052621603, + 0.007479982916265726, + 0.007069552317261696, + 0.0072741699405014515, + 0.013391111046075821, + 0.016393786296248436, + 0.019947685301303864, + 0.014566665515303612, + 0.01870720274746418, + 0.010337083600461483, + 0.01250273734331131, + 0.0044440352357923985, + 0.0032447653356939554, + 0.007742349989712238, + 0.00470219599083066, + 0.004429223481565714, + 0.006745525635778904, + 0.003924424760043621, + 0.004163440316915512, + 0.002294649137184024, + 0.004763966426253319, + 0.0030337332282215357, + 0.0025865030474960804, + 0.003317303257063031, + 0.0034933092538267374, + 0.0018183343345299363, + 0.0028445939533412457, + 0.0032880022190511227, + 0.004702981095761061, + 0.00716411042958498, + 0.005907213315367699, + 0.00427152868360281, + 0.005094427149742842, + 0.006730563938617706, + 0.003730170428752899, + 0.003959414549171925, + 0.00370579375885427, + 0.004999702796339989, + 0.0038102036342024803, + 0.004991801921278238, + 0.005105762742459774, + 0.006236824207007885, + 0.006317459046840668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.0008158930577337742 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "macro close-up shot from a level angle draws the viewer's gaze to a tranquil corner of the forest floor. A ray of warm afternoon sunlight penetrates the dense canopy, casting a soft beam that illuminates the center of the scene. The camera focuses on a young fawn, its light brown fur speckled with white appearing exceptionally soft and fluffy in the light. Its large, obsidian eyes gleam with innocence. Cautiously, it approaches a rotting log covered in emerald green moss. Atop this decaying log, a massive, fan-shaped, crimson Ganoderma lucidum thrives. Its surface is as smooth as lacquer, its edges gleaming with a golden sheen, revealing distinct rings of growth. The fawn's moist black nose gently touches the edge of the mushroom, and then, with its soft lips and tiny teeth, it delicately and tentatively nibbles at a corner of the cap, its movements filled with innocence and concentration. The entire background of the picture is blurred by the shallow depth of field into a dreamy green and brown halo. Rich details, saturated warm tones and soft light and shadow are intertwined, giving this quiet and vital moment a thick and delicate texture and atmosphere like a classical oil painting.", + "seed": 13413, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_10/index_7475_seed_13413_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.021466715016867966, + "vlm_logprob": -0.021466715016867966, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.021466715016867966, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.049738120287656784 + ], + "opposite_logprob_tokens": [ + -3.0497381687164307 + ], + "logprob_reward": -0.049738120287656784, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.049738120287656784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11146418005228043, + 0.07707930356264114, + 0.21475034952163696, + 0.0945025086402893, + 0.07005324214696884, + 0.06629626452922821, + 0.09614777565002441, + 0.12402714788913727, + 0.06454141438007355, + 0.08113782107830048 + ], + "credit_map": [ + 0.010668167844414711, + 0.012718014419078827, + 0.013290838338434696, + 0.011685843579471111, + 0.019060978665947914, + 0.011881493031978607, + 0.0200532004237175, + 0.01210565771907568, + 0.005724666174501181, + 0.010200496762990952, + 0.008044297806918621, + 0.008319465443491936, + 0.01142529584467411, + 0.01064154040068388, + 0.01289888471364975, + 0.009824661538004875, + 0.017999067902565002, + 0.027896620333194733, + 0.03398841992020607, + 0.03125395253300667, + 0.02758374810218811, + 0.02286662720143795, + 0.02099086530506611, + 0.03217105567455292, + 0.010739791207015514, + 0.012941681779921055, + 0.009628867730498314, + 0.014672464691102505, + 0.015302023850381374, + 0.010078109800815582, + 0.00837418157607317, + 0.01276539359241724, + 0.005882728844881058, + 0.012796062976121902, + 0.007424693088978529, + 0.011358221992850304, + 0.008904699236154556, + 0.008646159432828426, + 0.00847036112099886, + 0.006570319179445505, + 0.006366454064846039, + 0.011564445681869984, + 0.008359875530004501, + 0.006746639497578144, + 0.010410996153950691, + 0.010318361222743988, + 0.005743591580539942, + 0.006785905919969082, + 0.013503076508641243, + 0.012184650637209415, + 0.014638925902545452, + 0.00805448368191719, + 0.015504521317780018, + 0.011344386264681816, + 0.007066350430250168, + 0.013851393945515156, + 0.021080022677779198, + 0.016968922689557076, + 0.0230720192193985, + 0.01007886417210102, + 0.012482663616538048, + 0.01370986644178629, + 0.011539741419255733, + 0.015095064416527748, + 0.011584567837417126, + 0.00701142055913806, + 0.009609872475266457, + 0.006243359297513962, + 0.007471576798707247, + 0.007743730675429106, + 0.005841380916535854, + 0.009035512804985046, + 0.011513547971844673, + 0.011593306437134743, + 0.011387958191335201, + 0.0067413486540317535, + 0.009289821609854698, + 0.009444357827305794, + 0.005850270856171846, + 0.015317218378186226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.049738120287656784 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821165345609188 + ], + "opposite_logprob_tokens": [ + -4.018211841583252 + ], + "logprob_reward": -0.01821165345609188, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01821165345609188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06046035513281822, + 0.029498597607016563, + 0.07723735272884369, + 0.042585887014865875, + 0.06775277107954025, + 0.13303084671497345, + 0.04639068990945816, + 0.25584474205970764, + 0.05122462287545204, + 0.23597410321235657 + ], + "credit_map": [ + 0.00638379855081439, + 0.006960812024772167, + 0.007078440859913826, + 0.005523674190044403, + 0.008652514778077602, + 0.00733382860198617, + 0.012499367818236351, + 0.006027915049344301, + 0.002858724445104599, + 0.0033315413165837526, + 0.00444426853209734, + 0.003077230416238308, + 0.004138236399739981, + 0.0031843322794884443, + 0.004786351695656776, + 0.00367791298776865, + 0.005545197520405054, + 0.009068635292351246, + 0.013585787266492844, + 0.008877800777554512, + 0.010907386429607868, + 0.0072611309587955475, + 0.009148210287094116, + 0.012843204662203789, + 0.004985280800610781, + 0.005001775920391083, + 0.006996142212301493, + 0.00523538189008832, + 0.006065752357244492, + 0.004698929842561483, + 0.004489344544708729, + 0.005113278049975634, + 0.007155236322432756, + 0.008637967519462109, + 0.009992335923016071, + 0.005723378621041775, + 0.005197407212108374, + 0.019079511985182762, + 0.004703992046415806, + 0.007262940052896738, + 0.012135150842368603, + 0.022258561104536057, + 0.013296155259013176, + 0.01219995878636837, + 0.019581202417612076, + 0.020255083218216896, + 0.01535903848707676, + 0.01794569194316864, + 0.006052505224943161, + 0.0049492670223116875, + 0.008045780472457409, + 0.0044025550596416, + 0.005432543810456991, + 0.004840285051614046, + 0.007160373497754335, + 0.005507376976311207, + 0.03938271850347519, + 0.03901321440935135, + 0.04614109545946121, + 0.020852744579315186, + 0.02965015172958374, + 0.02829231694340706, + 0.022629298269748688, + 0.029883213341236115, + 0.00853357370942831, + 0.00413016602396965, + 0.007846756838262081, + 0.00897208135575056, + 0.004715587478131056, + 0.005232139024883509, + 0.006406712345778942, + 0.005387607961893082, + 0.025855684652924538, + 0.017741309478878975, + 0.033309370279312134, + 0.01270970143377781, + 0.014576424844563007, + 0.03283433988690376, + 0.02270362712442875, + 0.07624364644289017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.01821165345609188 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008650568313896656 + ], + "opposite_logprob_tokens": [ + -4.758650779724121 + ], + "logprob_reward": -0.008650568313896656, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008650568313896656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10260841995477676, + 0.05711138993501663, + 0.12790508568286896, + 0.06577711552381516, + 0.05338941514492035, + 0.1797652244567871, + 0.08365776389837265, + 0.13762015104293823, + 0.07557239383459091, + 0.11659302562475204 + ], + "credit_map": [ + 0.011197998188436031, + 0.009618745185434818, + 0.009325684048235416, + 0.009695074521005154, + 0.01693081669509411, + 0.011666857637465, + 0.022700412198901176, + 0.011472831480205059, + 0.005175102036446333, + 0.006658290047198534, + 0.008258981630206108, + 0.0056696804240345955, + 0.009684746153652668, + 0.0063245827332139015, + 0.009025271050632, + 0.006314735859632492, + 0.01026097871363163, + 0.013254704885184765, + 0.02174016647040844, + 0.014140695333480835, + 0.016169125214219093, + 0.011376637034118176, + 0.018134165555238724, + 0.022828614339232445, + 0.0068338774144649506, + 0.009191928431391716, + 0.006735958158969879, + 0.008396447636187077, + 0.011589368805289268, + 0.0069269961677491665, + 0.00889784749597311, + 0.007204694673418999, + 0.004501567222177982, + 0.00831886101514101, + 0.006321309134364128, + 0.008267657831311226, + 0.006301145069301128, + 0.006703589111566544, + 0.007087207864969969, + 0.005888072773814201, + 0.014732510782778263, + 0.025438180193305016, + 0.017670750617980957, + 0.029017845168709755, + 0.03072115033864975, + 0.02092701941728592, + 0.02459033392369747, + 0.01666744239628315, + 0.010202242061495781, + 0.007664288394153118, + 0.013611231930553913, + 0.008621400222182274, + 0.008783096447587013, + 0.01167907565832138, + 0.012814005836844444, + 0.010282422415912151, + 0.02279532328248024, + 0.016430692747235298, + 0.027757572010159492, + 0.011757654137909412, + 0.012935428880155087, + 0.018567100167274475, + 0.012671534903347492, + 0.01470484584569931, + 0.012923717498779297, + 0.009580149315297604, + 0.00888501014560461, + 0.010178645141422749, + 0.009486659429967403, + 0.00751829519867897, + 0.008163733407855034, + 0.00883618276566267, + 0.017964841797947884, + 0.012085139751434326, + 0.01696876436471939, + 0.013932154513895512, + 0.011224202811717987, + 0.01263431180268526, + 0.009716721251606941, + 0.022066889330744743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.008650568313896656 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029809430241584778 + ], + "opposite_logprob_tokens": [ + -3.5298094749450684 + ], + "logprob_reward": -0.029809430241584778, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029809430241584778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19452543556690216, + 0.068768709897995, + 0.14206545054912567, + 0.2567887306213379, + 0.06468520313501358, + 0.04841602221131325, + 0.06983309984207153, + 0.04014264792203903, + 0.06798633933067322, + 0.04678838327527046 + ], + "credit_map": [ + 0.019609274342656136, + 0.020047694444656372, + 0.030521854758262634, + 0.016284378245472908, + 0.023620907217264175, + 0.024692777544260025, + 0.04140954464673996, + 0.0183390025049448, + 0.006880578119307756, + 0.0071101924404501915, + 0.007697972934693098, + 0.007582462392747402, + 0.00917248334735632, + 0.008074614219367504, + 0.01100054569542408, + 0.011249870993196964, + 0.011162427254021168, + 0.017903346568346024, + 0.02470734715461731, + 0.013978103175759315, + 0.018327539786696434, + 0.01459327433258295, + 0.017393749207258224, + 0.023999664932489395, + 0.020908793434500694, + 0.02665131911635399, + 0.03465156629681587, + 0.03568144515156746, + 0.03517748787999153, + 0.028690172359347343, + 0.04501822218298912, + 0.03000970184803009, + 0.005654424894601107, + 0.009452029131352901, + 0.00808875821530819, + 0.007566072046756744, + 0.009637994691729546, + 0.008919021114706993, + 0.00869849044829607, + 0.006668418645858765, + 0.005628657527267933, + 0.007614981848746538, + 0.005233080126345158, + 0.0044065420515835285, + 0.006366947200149298, + 0.006127398926764727, + 0.00530272675678134, + 0.007735688239336014, + 0.00718557508662343, + 0.009899068623781204, + 0.008758310228586197, + 0.004847399424761534, + 0.01720951870083809, + 0.008068062365055084, + 0.007062316872179508, + 0.006802851799875498, + 0.005834667943418026, + 0.005787849426269531, + 0.004834643565118313, + 0.004071790259331465, + 0.0036951955407857895, + 0.0060098436661064625, + 0.0035429697018116713, + 0.006365687120705843, + 0.012259585782885551, + 0.008108371868729591, + 0.007171790115535259, + 0.006552737206220627, + 0.009554482996463776, + 0.007422014605253935, + 0.00856421235948801, + 0.008353138342499733, + 0.004646566696465015, + 0.007894660346210003, + 0.005813031457364559, + 0.005576937925070524, + 0.006291924510151148, + 0.0057811615988612175, + 0.004276466090232134, + 0.006507636979222298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.029809430241584778 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009238027851097286 + ], + "opposite_logprob_tokens": [ + -7.0009236335754395 + ], + "logprob_reward": -0.0009238027851097286, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009238027851097286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11707489937543869, + 0.09649339318275452, + 0.1629868745803833, + 0.029274147003889084, + 0.07038388401269913, + 0.0384269580245018, + 0.11362957954406738, + 0.18061578273773193, + 0.07868008315563202, + 0.11243439465761185 + ], + "credit_map": [ + 0.014061037451028824, + 0.010227196849882603, + 0.023013507947325706, + 0.011694834567606449, + 0.017287997528910637, + 0.0117349186912179, + 0.01639738865196705, + 0.012658016756176949, + 0.008198446594178677, + 0.011401982046663761, + 0.01333655882626772, + 0.008786140941083431, + 0.01646735705435276, + 0.011115779168903828, + 0.014669745229184628, + 0.012517384253442287, + 0.011698318645358086, + 0.01595846377313137, + 0.02775074541568756, + 0.021008310839533806, + 0.018936961889266968, + 0.015455400571227074, + 0.022749673575162888, + 0.029429011046886444, + 0.0029457106720656157, + 0.0031817362178117037, + 0.002670824062079191, + 0.0034984974190592766, + 0.003721213201060891, + 0.0033775591291487217, + 0.0066871801391243935, + 0.003191425697878003, + 0.007380892988294363, + 0.00977238081395626, + 0.00971834547817707, + 0.009504352696239948, + 0.009453339502215385, + 0.007734364829957485, + 0.007839334197342396, + 0.00898087676614523, + 0.005003379192203283, + 0.004423920530825853, + 0.005229939240962267, + 0.004732489585876465, + 0.003782370127737522, + 0.0073027703911066055, + 0.003858649404719472, + 0.0040934388525784016, + 0.013847613707184792, + 0.012965800240635872, + 0.01339265052229166, + 0.009856795892119408, + 0.023622173815965652, + 0.010848108679056168, + 0.017019134014844894, + 0.01207730732858181, + 0.02982550486922264, + 0.028866108506917953, + 0.027092060074210167, + 0.014326044358313084, + 0.0247492752969265, + 0.017929429188370705, + 0.016128620132803917, + 0.02169874869287014, + 0.01429983600974083, + 0.010086053051054478, + 0.008365505374968052, + 0.006486931350082159, + 0.01113960798829794, + 0.008631780743598938, + 0.010301131755113602, + 0.00936923734843731, + 0.010922075249254704, + 0.013184238225221634, + 0.013496823608875275, + 0.008409307338297367, + 0.010541989468038082, + 0.015890970826148987, + 0.00946069322526455, + 0.030528290197253227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.0009238027851097286 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "macro close-up shot from a level angle draws the viewer's gaze to a tranquil corner of the forest floor. A ray of warm afternoon sunlight penetrates the dense canopy, casting a soft beam that illuminates the center of the scene. The camera focuses on a young fawn, its light brown fur speckled with white appearing exceptionally soft and fluffy in the light. Its large, obsidian eyes gleam with innocence. Cautiously, it approaches a rotting log covered in emerald green moss. Atop this decaying log, a massive, fan-shaped, crimson Ganoderma lucidum thrives. Its surface is as smooth as lacquer, its edges gleaming with a golden sheen, revealing distinct rings of growth. The fawn's moist black nose gently touches the edge of the mushroom, and then, with its soft lips and tiny teeth, it delicately and tentatively nibbles at a corner of the cap, its movements filled with innocence and concentration. The entire background of the picture is blurred by the shallow depth of field into a dreamy green and brown halo. Rich details, saturated warm tones and soft light and shadow are intertwined, giving this quiet and vital moment a thick and delicate texture and atmosphere like a classical oil painting.", + "seed": 13414, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_10/index_7475_seed_13414_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.014089907449670137, + "vlm_logprob": -0.014089907449670137, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.014089907449670137, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027182530611753464 + ], + "opposite_logprob_tokens": [ + -3.6521825790405273 + ], + "logprob_reward": -0.027182530611753464, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.027182530611753464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41455498337745667, + 0.03736623004078865, + 0.06761956214904785, + 0.0929221585392952, + 0.1077546626329422, + 0.05103960633277893, + 0.04718572273850441, + 0.025243861600756645, + 0.08644240349531174, + 0.06987085938453674 + ], + "credit_map": [ + 0.021857894957065582, + 0.02343267761170864, + 0.027493569999933243, + 0.027349654585123062, + 0.1379203051328659, + 0.1399436742067337, + 0.016056954860687256, + 0.02050020918250084, + 0.0025298334658145905, + 0.0023423635866492987, + 0.004732324741780758, + 0.0032936176285147667, + 0.007713897153735161, + 0.006668795831501484, + 0.0049770791083574295, + 0.005108315031975508, + 0.005311285611242056, + 0.009985656477510929, + 0.006861523259431124, + 0.011677545495331287, + 0.01034444011747837, + 0.009698493406176567, + 0.007088661193847656, + 0.006651947274804115, + 0.007489305455237627, + 0.011394228786230087, + 0.011503951624035835, + 0.010158410295844078, + 0.019179416820406914, + 0.016570700332522392, + 0.009654128924012184, + 0.006972003262490034, + 0.009318285621702671, + 0.0167026836425066, + 0.015004878863692284, + 0.01496871653944254, + 0.015885408967733383, + 0.020102685317397118, + 0.007655482739210129, + 0.008116512559354305, + 0.005519415717571974, + 0.007411560975015163, + 0.005147942807525396, + 0.005682890769094229, + 0.008005740121006966, + 0.0073279719799757, + 0.005024803802371025, + 0.006919275037944317, + 0.0055563910864293575, + 0.007844078354537487, + 0.005577644798904657, + 0.005709212739020586, + 0.008260848931968212, + 0.005188089795410633, + 0.00501123396679759, + 0.0040382156148552895, + 0.002440630691125989, + 0.003765459405258298, + 0.0028311177156865597, + 0.0024224394001066685, + 0.0065123168751597404, + 0.003458707593381405, + 0.001732963020913303, + 0.0020802223589271307, + 0.012384637258946896, + 0.011427669785916805, + 0.009400459937751293, + 0.011817649938166142, + 0.014659712091088295, + 0.010335245169699192, + 0.005754745565354824, + 0.010662281885743141, + 0.008516916073858738, + 0.006104848347604275, + 0.010655008256435394, + 0.008684330619871616, + 0.007463147398084402, + 0.009558230638504028, + 0.00811864249408245, + 0.010769727639853954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.027182530611753464 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008653995580971241 + ], + "opposite_logprob_tokens": [ + -4.7586541175842285 + ], + "logprob_reward": -0.008653995580971241, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008653995580971241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3585489094257355, + 0.10777401924133301, + 0.034413132816553116, + 0.06636067479848862, + 0.03820595517754555, + 0.0454367958009243, + 0.028966858983039856, + 0.0686655044555664, + 0.0999942347407341, + 0.1516338437795639 + ], + "credit_map": [ + 0.01670820079743862, + 0.020878266543149948, + 0.02575077675282955, + 0.018444770947098732, + 0.11735718697309494, + 0.12286850810050964, + 0.018018366768956184, + 0.0185228418558836, + 0.007673405110836029, + 0.009397829882800579, + 0.0175942312926054, + 0.009638609364628792, + 0.023238644003868103, + 0.025324223563075066, + 0.007554279640316963, + 0.007352800574153662, + 0.0031505257356911898, + 0.0036790084559470415, + 0.004142445977777243, + 0.006258909124881029, + 0.004122404847294092, + 0.003845742205157876, + 0.004825834650546312, + 0.004388262517750263, + 0.00550225842744112, + 0.006738495081663132, + 0.008424925617873669, + 0.009060117416083813, + 0.011699369177222252, + 0.01260384451597929, + 0.0058795684017241, + 0.0064520989544689655, + 0.004289258737117052, + 0.005035659298300743, + 0.00474545918405056, + 0.003826625645160675, + 0.00541885569691658, + 0.004700490273535252, + 0.0045607006177306175, + 0.005628905259072781, + 0.005638748407363892, + 0.005350180435925722, + 0.005412304773926735, + 0.004886061884462833, + 0.006245494820177555, + 0.0074768359772861, + 0.00448773754760623, + 0.0059394328854978085, + 0.004157669842243195, + 0.004003006499260664, + 0.002724653808400035, + 0.0030942116864025593, + 0.004972231574356556, + 0.0037931990809738636, + 0.0032914779148995876, + 0.0029304074123501778, + 0.00700950575992465, + 0.008023975417017937, + 0.008515901863574982, + 0.007871773093938828, + 0.01265527494251728, + 0.008880848065018654, + 0.005913782399147749, + 0.0097944475710392, + 0.014218080788850784, + 0.013832827098667622, + 0.009268231689929962, + 0.01383218914270401, + 0.01763789728283882, + 0.01123742200434208, + 0.009927676059305668, + 0.01003990974277258, + 0.018325598910450935, + 0.012769289314746857, + 0.023998038843274117, + 0.02393634058535099, + 0.016972966492176056, + 0.022823523730039597, + 0.012826916761696339, + 0.01998117007315159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.008653995580971241 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036216876469552517 + ], + "opposite_logprob_tokens": [ + -5.628621578216553 + ], + "logprob_reward": -0.0036216876469552517, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036216876469552517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19607548415660858, + 0.11534086614847183, + 0.04086901247501373, + 0.06871906667947769, + 0.193610280752182, + 0.07323335856199265, + 0.06236722320318222, + 0.03948202356696129, + 0.12819211184978485, + 0.08211055397987366 + ], + "credit_map": [ + 0.01768270879983902, + 0.017203785479068756, + 0.01730562560260296, + 0.018191279843449593, + 0.038980551064014435, + 0.03829209506511688, + 0.02711532637476921, + 0.021304132416844368, + 0.009740016423165798, + 0.009264637716114521, + 0.01961105316877365, + 0.012595336884260178, + 0.019542163237929344, + 0.02074866183102131, + 0.012711802497506142, + 0.011127201840281487, + 0.003132255282253027, + 0.0033909277990460396, + 0.006531490944325924, + 0.00638450775295496, + 0.007135306019335985, + 0.00545323034748435, + 0.0037050913088023663, + 0.005136209074407816, + 0.005048682447522879, + 0.0047727664932608604, + 0.007907559163868427, + 0.009526778012514114, + 0.011308105662465096, + 0.010248132981359959, + 0.011506420560181141, + 0.008400630205869675, + 0.016764534637331963, + 0.023212026804685593, + 0.024487335234880447, + 0.029245316982269287, + 0.03030894137918949, + 0.028966812416911125, + 0.022706681862473488, + 0.01791864074766636, + 0.010827800258994102, + 0.0071262698620557785, + 0.006976427976042032, + 0.009931249544024467, + 0.0073624104261398315, + 0.00923572015017271, + 0.009399501606822014, + 0.012373982928693295, + 0.010327943600714207, + 0.005494900047779083, + 0.005028694402426481, + 0.005687003489583731, + 0.0076019009575247765, + 0.00771632045507431, + 0.006728312000632286, + 0.013782156631350517, + 0.0048659988678991795, + 0.004258384462445974, + 0.0068489680998027325, + 0.006079657934606075, + 0.003522705752402544, + 0.0058482009917497635, + 0.004964218009263277, + 0.003093893639743328, + 0.018869740888476372, + 0.018475310876965523, + 0.013101238757371902, + 0.015631090849637985, + 0.024506185203790665, + 0.015090874396264553, + 0.010759335942566395, + 0.011758353561162949, + 0.0088090430945158, + 0.007750192657113075, + 0.012467418797314167, + 0.011876039206981659, + 0.009982890449464321, + 0.011751142330467701, + 0.00966733694076538, + 0.009806502610445023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.0036216876469552517 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029808156192302704 + ], + "opposite_logprob_tokens": [ + -3.5298080444335938 + ], + "logprob_reward": -0.029808156192302704, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029808156192302704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1449378877878189, + 0.08300462365150452, + 0.14481189846992493, + 0.20196221768856049, + 0.1955968588590622, + 0.05008579045534134, + 0.042653631418943405, + 0.04053553566336632, + 0.06462091952562332, + 0.031790658831596375 + ], + "credit_map": [ + 0.01881500519812107, + 0.015894755721092224, + 0.01330980658531189, + 0.020287826657295227, + 0.02626720629632473, + 0.017626849934458733, + 0.01549611147493124, + 0.01724032685160637, + 0.008828200399875641, + 0.010876304470002651, + 0.010343593545258045, + 0.008668394759297371, + 0.012156848795711994, + 0.014765220694243908, + 0.008899410255253315, + 0.008466656319797039, + 0.010070174932479858, + 0.016314836218953133, + 0.01751977950334549, + 0.031964898109436035, + 0.023888085037469864, + 0.01880512945353985, + 0.013266491703689098, + 0.012982519343495369, + 0.016239942982792854, + 0.01950426585972309, + 0.026058675721287727, + 0.02838815562427044, + 0.03576255962252617, + 0.04291630908846855, + 0.017164837568998337, + 0.015927471220493317, + 0.016268828883767128, + 0.02641322836279869, + 0.02463284134864807, + 0.03307189419865608, + 0.028931500390172005, + 0.028740689158439636, + 0.020864417776465416, + 0.016673460602760315, + 0.006434353534132242, + 0.006493622902780771, + 0.004943534731864929, + 0.006153568625450134, + 0.005536206066608429, + 0.00735861249268055, + 0.0045600817538797855, + 0.008605810813605785, + 0.005811724346131086, + 0.0038019318599253893, + 0.003975956700742245, + 0.005062510259449482, + 0.003994541708379984, + 0.005199624691158533, + 0.006006430834531784, + 0.008800911717116833, + 0.005298062227666378, + 0.005453750491142273, + 0.00308346189558506, + 0.005872170440852642, + 0.0068163820542395115, + 0.005627746693789959, + 0.004115985706448555, + 0.0042679752223193645, + 0.009726714342832565, + 0.006273076869547367, + 0.007139348890632391, + 0.008459619246423244, + 0.008426940068602562, + 0.007731202989816666, + 0.004216320812702179, + 0.012647690251469612, + 0.003383950563147664, + 0.003032510867342353, + 0.0034500479232519865, + 0.0069096204824745655, + 0.00399087043479085, + 0.0045537385158240795, + 0.0028354425448924303, + 0.003634476801380515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.029808156192302704 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a young fawn with light brown fur speckled with white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a rotting log covered in emerald green moss in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, fan-shaped, crimson Ganoderma lucidum mushroom growing on the log with a smooth lacquer-like surface and golden sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fawn gently touch the mushroom with its nose and then delicately nibble at the cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred by a shallow depth of field into a green and brown halo?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011831672163680196 + ], + "opposite_logprob_tokens": [ + -6.751183032989502 + ], + "logprob_reward": -0.0011831672163680196, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011831672163680196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35401129722595215, + 0.10026835650205612, + 0.1398105025291443, + 0.0420784093439579, + 0.04430903121829033, + 0.047587066888809204, + 0.07270818948745728, + 0.025824235752224922, + 0.07707106322050095, + 0.09633191674947739 + ], + "credit_map": [ + 0.025506358593702316, + 0.022900471463799477, + 0.03522057831287384, + 0.021296365186572075, + 0.10319437086582184, + 0.108043372631073, + 0.01564716547727585, + 0.022202573716640472, + 0.006969686131924391, + 0.005861368961632252, + 0.015823177993297577, + 0.0064800032414495945, + 0.02389407902956009, + 0.026774704456329346, + 0.007650784216821194, + 0.006814535707235336, + 0.012622901238501072, + 0.009377151727676392, + 0.017355158925056458, + 0.0219406895339489, + 0.017034975811839104, + 0.016309408470988274, + 0.0186651349067688, + 0.026505056768655777, + 0.0031889465171843767, + 0.005040837451815605, + 0.005597135052084923, + 0.004603130277246237, + 0.009009831584990025, + 0.008774026297032833, + 0.003389695892110467, + 0.002474800916388631, + 0.004000111483037472, + 0.004795256070792675, + 0.005384128540754318, + 0.005021391902118921, + 0.00699472613632679, + 0.008207517676055431, + 0.004381077364087105, + 0.0055248187854886055, + 0.0039510601200163364, + 0.004676260985434055, + 0.005681862123310566, + 0.006208956707268953, + 0.007926159538328648, + 0.006870492827147245, + 0.0054695261642336845, + 0.006802740972489119, + 0.008013807237148285, + 0.008571138605475426, + 0.007423712871968746, + 0.009209837764501572, + 0.006746526341885328, + 0.008489320985972881, + 0.008772679604589939, + 0.015481153503060341, + 0.003599422750994563, + 0.0037528264801949263, + 0.0032969671301543713, + 0.0029979790560901165, + 0.003104608040302992, + 0.0032654746901243925, + 0.0031601430382579565, + 0.0026468124706298113, + 0.010378901846706867, + 0.011929457075893879, + 0.007476010825484991, + 0.009909405373036861, + 0.013698632828891277, + 0.008550819009542465, + 0.006112218834459782, + 0.009015603922307491, + 0.014282314106822014, + 0.0071774129755795, + 0.020870819687843323, + 0.014547972939908504, + 0.007312229834496975, + 0.015534510836005211, + 0.007533154916018248, + 0.009073483757674694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0011831672163680196 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_11/index_tensor([6525, 6525])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_11/index_tensor([6525, 6525])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..21364ea169e6e49ed26ea2c7504f19d1ee1ba49e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_11/index_tensor([6525, 6525])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme cinematic macro shot, rendered in ultra-high definition, slowly zooms in on the interior of a gold-lipped pearl oyster, set against a deep, out-of-focus, abyssal black background. At the center of the frame, a lustrous pearl, its surface a creamy, milky white. Its nearly spherical shape catches soft, directed light, revealing a deep inner glow with subtle hints of rose and gold. The oyster's iridescent mantle gently nestles around the pearl, a delicate fleshy tissue that seems to breathe with slow, rhythmic pulsations, its folds delicately caressing the pearl's surface. The mantle's surface shimmers with a hypnotic, oily iridescence, its colors shifting and undulating delicately across a spectrum of malachite green, soft violet, and warm gold. The camera's deliberate movements capture the breathtaking play of light within the surrounding mother-of-pearl – a landscape of polished, undulating nacre, where every microscopic ridge and crystalline layer refracts light, creating a living universe teeming with color and texture. The focused, tranquil atmosphere is one of profound natural wonder, revealing the slow, beautiful creation of gemstones in stunning detail.", + "seed": 14261, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_11/index_6525_seed_14261_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0315929957665503, + "vlm_logprob": -0.0315929957665503, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.0315929957665503, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013934116810560226 + ], + "opposite_logprob_tokens": [ + -4.513934135437012 + ], + "logprob_reward": -0.013934116810560226, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013934116810560226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11914974451065063, + 0.08687291294336319, + 0.02756214141845703, + 0.041351981461048126, + 0.440901517868042, + 0.036027517169713974, + 0.0400044210255146, + 0.024159660562872887, + 0.14635007083415985, + 0.03762005269527435 + ], + "credit_map": [ + 0.011519270949065685, + 0.01252585556358099, + 0.01754557155072689, + 0.02502337656915188, + 0.009967470541596413, + 0.013623770326375961, + 0.01303369551897049, + 0.01591072976589203, + 0.009252368472516537, + 0.010214872658252716, + 0.011698619462549686, + 0.010469482280313969, + 0.01105382852256298, + 0.012496795505285263, + 0.010308719240128994, + 0.011378227733075619, + 0.0023789198603481054, + 0.002419430064037442, + 0.003592037595808506, + 0.007186713628470898, + 0.0027139822486788034, + 0.0031013470143079758, + 0.003175073768943548, + 0.002994638867676258, + 0.0031979463528841734, + 0.00247386796399951, + 0.0068448809906840324, + 0.010005711577832699, + 0.002444254467263818, + 0.008708122186362743, + 0.0037887098733335733, + 0.00388848758302629, + 0.017194027081131935, + 0.015373707748949528, + 0.022235816344618797, + 0.022996697574853897, + 0.1651524007320404, + 0.034758467227220535, + 0.016919652000069618, + 0.1462707221508026, + 0.003186620771884918, + 0.0026645672041922808, + 0.004464463796466589, + 0.009950723499059677, + 0.0048834760673344135, + 0.002881755353882909, + 0.003264117054641247, + 0.0047317929565906525, + 0.003724172944203019, + 0.0030895641539245844, + 0.006529252044856548, + 0.00700876209884882, + 0.005423291586339474, + 0.0041707963682711124, + 0.005624294746667147, + 0.004434282425791025, + 0.001864181482233107, + 0.0026057600043714046, + 0.0029018898494541645, + 0.003978821914643049, + 0.0027377791702747345, + 0.003914790228009224, + 0.0025204571429640055, + 0.0036359825171530247, + 0.01000264286994934, + 0.014873393811285496, + 0.02346673421561718, + 0.019906338304281235, + 0.011876688338816166, + 0.029643557965755463, + 0.012021397240459919, + 0.02455933578312397, + 0.0031261087860912085, + 0.0025874373968690634, + 0.0041877515614032745, + 0.004780470859259367, + 0.002964828861877322, + 0.010291790589690208, + 0.0031020361930131912, + 0.006579627748578787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.013934116810560226 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981880120933056 + ], + "opposite_logprob_tokens": [ + -3.5298187732696533 + ], + "logprob_reward": -0.02981880120933056, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981880120933056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08903968334197998, + 0.057294812053442, + 0.054120589047670364, + 0.07002897560596466, + 0.09762600064277649, + 0.07464979588985443, + 0.043604347854852676, + 0.025944175198674202, + 0.14491528272628784, + 0.34277626872062683 + ], + "credit_map": [ + 0.009171073324978352, + 0.009549673646688461, + 0.011403418146073818, + 0.017029883340001106, + 0.009784598834812641, + 0.007611055858433247, + 0.010835744440555573, + 0.013654248788952827, + 0.005099022760987282, + 0.0062209852039813995, + 0.007388669066131115, + 0.009717997163534164, + 0.004520648159086704, + 0.005896190647035837, + 0.007505928166210651, + 0.010945375077426434, + 0.008144984021782875, + 0.005175688769668341, + 0.0073290313594043255, + 0.007964078336954117, + 0.005142382346093655, + 0.006642918102443218, + 0.00655756052583456, + 0.007163946982473135, + 0.005648065824061632, + 0.004109879024326801, + 0.010392292402684689, + 0.014203937724232674, + 0.005575110204517841, + 0.01904335618019104, + 0.004461606964468956, + 0.006594731006771326, + 0.0039940993301570415, + 0.003836155403405428, + 0.004632235039025545, + 0.005636059679090977, + 0.034109316766262054, + 0.008633061312139034, + 0.004348090849816799, + 0.03243699297308922, + 0.007977471686899662, + 0.006275949068367481, + 0.009595844894647598, + 0.015598449856042862, + 0.011339813470840454, + 0.006533674895763397, + 0.007988899946212769, + 0.009339699521660805, + 0.0026431509759277105, + 0.0030484094750136137, + 0.004891359247267246, + 0.01465926505625248, + 0.003952714614570141, + 0.004195632413029671, + 0.004781921394169331, + 0.005431899800896645, + 0.0018519710283726454, + 0.002728269901126623, + 0.0034091570414602757, + 0.006032153498381376, + 0.0023003206588327885, + 0.003408640855923295, + 0.0027034992817789316, + 0.0035101640969514847, + 0.011721108108758926, + 0.016649819910526276, + 0.01648620516061783, + 0.018320197239518166, + 0.011986998841166496, + 0.03299633786082268, + 0.014081533066928387, + 0.022673098370432854, + 0.02594737708568573, + 0.028385519981384277, + 0.02994641475379467, + 0.038071054965257645, + 0.03746559098362923, + 0.09972018003463745, + 0.025897465646266937, + 0.057342708110809326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.02981880120933056 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028309053741395473 + ], + "opposite_logprob_tokens": [ + -5.877830982208252 + ], + "logprob_reward": -0.0028309053741395473, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028309053741395473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11443634331226349, + 0.057495128363370895, + 0.028423087671399117, + 0.16143468022346497, + 0.21766318380832672, + 0.05433265119791031, + 0.0357775054872036, + 0.05776945874094963, + 0.14965757727622986, + 0.12301038950681686 + ], + "credit_map": [ + 0.01598663628101349, + 0.017017913982272148, + 0.015204149298369884, + 0.024953439831733704, + 0.009423957206308842, + 0.010414804331958294, + 0.012317752465605736, + 0.009117700159549713, + 0.005286342930048704, + 0.005964567419141531, + 0.009643087163567543, + 0.013513999059796333, + 0.004978644195944071, + 0.005215882323682308, + 0.005375825800001621, + 0.007516783662140369, + 0.0032605775631964207, + 0.003350545885041356, + 0.004770656116306782, + 0.006329625379294157, + 0.002227338030934334, + 0.0027838549576699734, + 0.00285166222602129, + 0.0028488291427493095, + 0.007518076803535223, + 0.008352703414857388, + 0.022001024335622787, + 0.05809691175818443, + 0.0073104375042021275, + 0.03617006912827492, + 0.010404951870441437, + 0.011580513790249825, + 0.008418546058237553, + 0.007965614087879658, + 0.01077898871153593, + 0.011582200415432453, + 0.08049528300762177, + 0.017802586778998375, + 0.007900120690464973, + 0.07271984219551086, + 0.004460870288312435, + 0.004859456326812506, + 0.009102522395551205, + 0.0064802332781255245, + 0.00873525533825159, + 0.0073921820148825645, + 0.0063226609490811825, + 0.006979475263506174, + 0.003597396658733487, + 0.0031559087801724672, + 0.006327483337372541, + 0.006538067944347858, + 0.003871987108141184, + 0.003921856638044119, + 0.004476895555853844, + 0.003887908300384879, + 0.004651063121855259, + 0.005571711342781782, + 0.007916520349681377, + 0.014437398873269558, + 0.005640463903546333, + 0.00607090862467885, + 0.006717968732118607, + 0.00676342798396945, + 0.010724724270403385, + 0.016482267528772354, + 0.01607206091284752, + 0.013590686954557896, + 0.011086853221058846, + 0.03958671912550926, + 0.013925834558904171, + 0.0281884353607893, + 0.011603917926549911, + 0.012225622311234474, + 0.01261416357010603, + 0.015247754752635956, + 0.0140165314078331, + 0.02931431494653225, + 0.010491791181266308, + 0.017496298998594284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.0028309053741395473 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080753989517689 + ], + "opposite_logprob_tokens": [ + -4.511080741882324 + ], + "logprob_reward": -0.011080753989517689, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080753989517689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13889144361019135, + 0.07899978756904602, + 0.05464303493499756, + 0.28702521324157715, + 0.08794844150543213, + 0.04948864132165909, + 0.03544178605079651, + 0.07805798202753067, + 0.1112758219242096, + 0.07822781056165695 + ], + "credit_map": [ + 0.01409720629453659, + 0.017369121313095093, + 0.020240377634763718, + 0.02590462751686573, + 0.012565777637064457, + 0.012432788498699665, + 0.017184708267450333, + 0.019096840173006058, + 0.007190750911831856, + 0.01015153992921114, + 0.011684993281960487, + 0.015024008229374886, + 0.008144139312207699, + 0.010050540789961815, + 0.006707669701427221, + 0.010046142153441906, + 0.004901723936200142, + 0.005145121365785599, + 0.010486309416592121, + 0.012919503264129162, + 0.004178238566964865, + 0.005806681700050831, + 0.005245463456958532, + 0.005959993693977594, + 0.01377616822719574, + 0.014100583270192146, + 0.04312995448708534, + 0.10317200422286987, + 0.012820162810385227, + 0.06351256370544434, + 0.01841275952756405, + 0.01810101792216301, + 0.004014500882476568, + 0.004784656688570976, + 0.006428833119571209, + 0.00658116489648819, + 0.030223578214645386, + 0.0058172596618533134, + 0.004935037810355425, + 0.025163406506180763, + 0.0035186195746064186, + 0.003940041642636061, + 0.006751564331352711, + 0.014888105913996696, + 0.004559429362416267, + 0.004378455691039562, + 0.005162496119737625, + 0.006289925891906023, + 0.003861261997371912, + 0.0031948769465088844, + 0.0054498761892318726, + 0.005714900325983763, + 0.00470531964674592, + 0.005395684391260147, + 0.0029920493252575397, + 0.004127817694097757, + 0.008106839843094349, + 0.010274638421833515, + 0.010786198079586029, + 0.013328575529158115, + 0.0081204604357481, + 0.010491165332496166, + 0.008059508167207241, + 0.008890598081052303, + 0.008338777348399162, + 0.01043478213250637, + 0.01229873113334179, + 0.00902875792235136, + 0.008407440036535263, + 0.031128190457820892, + 0.010369532741606236, + 0.021269608289003372, + 0.006374466232955456, + 0.008884232491254807, + 0.007956461980938911, + 0.00992185901850462, + 0.009130725637078285, + 0.016994841396808624, + 0.006797886919230223, + 0.012167341075837612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.011080753989517689 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030040144920349 + ], + "opposite_logprob_tokens": [ + -2.3503003120422363 + ], + "logprob_reward": -0.10030040144920349, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030040144920349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08676887303590775, + 0.06599200516939163, + 0.07044096291065216, + 0.1435551643371582, + 0.22469308972358704, + 0.04635228216648102, + 0.06440667808055878, + 0.019337540492415428, + 0.2379506230354309, + 0.04050283506512642 + ], + "credit_map": [ + 0.007050162181258202, + 0.013309996575117111, + 0.011470401659607887, + 0.018886316567659378, + 0.00806637667119503, + 0.00954726804047823, + 0.007773416116833687, + 0.010664933361113071, + 0.005771406460553408, + 0.010786525905132294, + 0.008835980668663979, + 0.009333953261375427, + 0.006401986349374056, + 0.006965754088014364, + 0.007596631534397602, + 0.010299771092832088, + 0.010297579690814018, + 0.00865861028432846, + 0.010571308434009552, + 0.012318318709731102, + 0.006373291369527578, + 0.007287637796252966, + 0.007018489763140678, + 0.007915726862847805, + 0.007193982135504484, + 0.008192917332053185, + 0.018010346218943596, + 0.048697128891944885, + 0.00730903260409832, + 0.0345434844493866, + 0.009317363612353802, + 0.010290904901921749, + 0.008664119988679886, + 0.00775701692327857, + 0.011606778018176556, + 0.012954875826835632, + 0.08228696882724762, + 0.018573932349681854, + 0.007916685193777084, + 0.07493270933628082, + 0.003736685961484909, + 0.004418158438056707, + 0.006162955891340971, + 0.011148313991725445, + 0.005812038201838732, + 0.0046313111670315266, + 0.004298301413655281, + 0.006144518963992596, + 0.005813575349748135, + 0.0052810111083090305, + 0.00868726521730423, + 0.010554632171988487, + 0.009128491394221783, + 0.009664383716881275, + 0.005958135239779949, + 0.009319184347987175, + 0.0018481521401554346, + 0.0019715989474207163, + 0.0024495357647538185, + 0.0035026438999921083, + 0.00206867977976799, + 0.0030427370220422745, + 0.0021091774106025696, + 0.0023450145963579416, + 0.01808883808553219, + 0.017931103706359863, + 0.033281806856393814, + 0.028875110670924187, + 0.021206937730312347, + 0.05531826615333557, + 0.02050107903778553, + 0.04274747893214226, + 0.0032698807772248983, + 0.003675447078421712, + 0.005160927306860685, + 0.004524768330156803, + 0.0033821710385382175, + 0.01019416656345129, + 0.0032814641017466784, + 0.007014011964201927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.10030040144920349 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme cinematic macro shot, rendered in ultra-high definition, slowly zooms in on the interior of a gold-lipped pearl oyster, set against a deep, out-of-focus, abyssal black background. At the center of the frame, a lustrous pearl, its surface a creamy, milky white. Its nearly spherical shape catches soft, directed light, revealing a deep inner glow with subtle hints of rose and gold. The oyster's iridescent mantle gently nestles around the pearl, a delicate fleshy tissue that seems to breathe with slow, rhythmic pulsations, its folds delicately caressing the pearl's surface. The mantle's surface shimmers with a hypnotic, oily iridescence, its colors shifting and undulating delicately across a spectrum of malachite green, soft violet, and warm gold. The camera's deliberate movements capture the breathtaking play of light within the surrounding mother-of-pearl – a landscape of polished, undulating nacre, where every microscopic ridge and crystalline layer refracts light, creating a living universe teeming with color and texture. The focused, tranquil atmosphere is one of profound natural wonder, revealing the slow, beautiful creation of gemstones in stunning detail.", + "seed": 14262, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_11/index_6525_seed_14262_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.031141353014390916, + "vlm_logprob": -0.031141353014390916, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.031141353014390916, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009897679090499878 + ], + "opposite_logprob_tokens": [ + -5.259897708892822 + ], + "logprob_reward": -0.009897679090499878, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.009897679090499878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07981671392917633, + 0.03055112436413765, + 0.26712578535079956, + 0.04538213089108467, + 0.019749604165554047, + 0.01946866512298584, + 0.09353777766227722, + 0.04372730478644371, + 0.3092004954814911, + 0.0914403647184372 + ], + "credit_map": [ + 0.006333419121801853, + 0.01845184713602066, + 0.010300739668309689, + 0.008984588086605072, + 0.007966073229908943, + 0.010050447657704353, + 0.006626697722822428, + 0.011102897115051746, + 0.0022473528515547514, + 0.003510002978146076, + 0.0041420008055865765, + 0.005566815380007029, + 0.003665864933282137, + 0.003012640168890357, + 0.0033536984119564295, + 0.005052750464528799, + 0.014043423347175121, + 0.013962098397314548, + 0.10113511979579926, + 0.01613672263920307, + 0.011561097577214241, + 0.08114926517009735, + 0.016077149659395218, + 0.01306093204766512, + 0.0038088473957031965, + 0.009633135050535202, + 0.005883706733584404, + 0.005486657842993736, + 0.006748317275196314, + 0.004253796301782131, + 0.005117746535688639, + 0.0044499230571091175, + 0.0020268161315470934, + 0.002699808683246374, + 0.0021420111879706383, + 0.002500934759154916, + 0.0024442956782877445, + 0.0030279303900897503, + 0.002820305060595274, + 0.0020875015761703253, + 0.002180128823965788, + 0.002480119001120329, + 0.0018770579481497407, + 0.0021657319739460945, + 0.0021128563676029444, + 0.0023072445765137672, + 0.0023608237970620394, + 0.003984702285379171, + 0.008223638869822025, + 0.005200037267059088, + 0.030011476948857307, + 0.007678083144128323, + 0.005942558869719505, + 0.011608612723648548, + 0.007762619759887457, + 0.017110753804445267, + 0.004415442701429129, + 0.0070259105414152145, + 0.004647796507924795, + 0.006308574229478836, + 0.005400015972554684, + 0.004182677250355482, + 0.005870091263204813, + 0.005876794457435608, + 0.02137487567961216, + 0.01990228332579136, + 0.01838633231818676, + 0.028628110885620117, + 0.01883014664053917, + 0.05541975423693657, + 0.12248262763023376, + 0.024176377803087234, + 0.010014921426773071, + 0.00882297195494175, + 0.01709308847784996, + 0.010566174052655697, + 0.009846238419413567, + 0.008554964326322079, + 0.014362071640789509, + 0.012179934419691563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 6 + ], + "selected_reward": -0.009897679090499878 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370881825685501 + ], + "opposite_logprob_tokens": [ + -3.4087088108062744 + ], + "logprob_reward": -0.03370881825685501, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370881825685501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05950918048620224, + 0.010358789935708046, + 0.5853588581085205, + 0.01181549672037363, + 0.02589157037436962, + 0.02208040840923786, + 0.060901857912540436, + 0.07719395309686661, + 0.09260254353284836, + 0.05428731441497803 + ], + "credit_map": [ + 0.005606271792203188, + 0.014115648344159126, + 0.006436627358198166, + 0.005443584639579058, + 0.008213551715016365, + 0.007608444429934025, + 0.005565210245549679, + 0.006519845221191645, + 0.0007754981052130461, + 0.0014443030813708901, + 0.0008866769494488835, + 0.002580094849690795, + 0.0012333207996562123, + 0.001063722069375217, + 0.0011725513031706214, + 0.0012026230106130242, + 0.030651479959487915, + 0.027307316660881042, + 0.23541155457496643, + 0.030606914311647415, + 0.02079831063747406, + 0.1802162379026413, + 0.03207540139555931, + 0.028291651979088783, + 0.0009431910002604127, + 0.00115112098865211, + 0.0019096796168014407, + 0.001079100533388555, + 0.001336281537078321, + 0.0012405384331941605, + 0.002516289008781314, + 0.001639297348447144, + 0.002115541370585561, + 0.0028617526404559612, + 0.004336967598646879, + 0.0027380092069506645, + 0.0031652129255235195, + 0.0032034849282354116, + 0.004118162207305431, + 0.0033524425234645605, + 0.0014462098479270935, + 0.002066499786451459, + 0.008967479690909386, + 0.0016131092561408877, + 0.000827139476314187, + 0.0037489579990506172, + 0.0017394352471455932, + 0.0016715782694518566, + 0.004957515746355057, + 0.0027660566847771406, + 0.022396648302674294, + 0.004951078444719315, + 0.002909968141466379, + 0.007099317852407694, + 0.004730434622615576, + 0.011090840213000774, + 0.00771655747666955, + 0.005284353159368038, + 0.008757936768233776, + 0.011857316829264164, + 0.005876126233488321, + 0.013973332941532135, + 0.008975289762020111, + 0.01475304365158081, + 0.0066766999661922455, + 0.005533657036721706, + 0.008709035813808441, + 0.0064496248960494995, + 0.00653057312592864, + 0.01657659187912941, + 0.03366785869002342, + 0.008458508178591728, + 0.006850416772067547, + 0.004502250347286463, + 0.009633271023631096, + 0.0058852266520261765, + 0.006300822366029024, + 0.005522310733795166, + 0.007144734263420105, + 0.008448285982012749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.03370881825685501 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007285801111720502 + ], + "opposite_logprob_tokens": [ + -7.250728607177734 + ], + "logprob_reward": -0.0007285801111720502, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007285801111720502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0807647779583931, + 0.03814898058772087, + 0.30242714285850525, + 0.05771135538816452, + 0.03162393346428871, + 0.06853769719600677, + 0.037274070084095, + 0.06184937432408333, + 0.2209736406803131, + 0.10068906098604202 + ], + "credit_map": [ + 0.0061131310649216175, + 0.02026067115366459, + 0.009163766168057919, + 0.007793867029249668, + 0.008198883384466171, + 0.014597279019653797, + 0.008372019976377487, + 0.0062651620246469975, + 0.0035067920107394457, + 0.005546854343265295, + 0.005568769760429859, + 0.003999579697847366, + 0.00477211456745863, + 0.005570541601628065, + 0.004441858734935522, + 0.004742469172924757, + 0.016350002959370613, + 0.015606490895152092, + 0.11921151727437973, + 0.01596447080373764, + 0.011450630612671375, + 0.09230383485555649, + 0.016570305451750755, + 0.014969895593822002, + 0.004616812337189913, + 0.010776865296065807, + 0.007128814235329628, + 0.005644277203828096, + 0.0077229710295796394, + 0.005890078842639923, + 0.009877604432404041, + 0.006053931079804897, + 0.0030669833067804575, + 0.004326930269598961, + 0.004703506827354431, + 0.0037285881116986275, + 0.003660811809822917, + 0.005589203909039497, + 0.003699854016304016, + 0.0028480568435043097, + 0.003975311294198036, + 0.014048310928046703, + 0.026557276025414467, + 0.005532620474696159, + 0.0037351257633417845, + 0.004636919125914574, + 0.004393761046230793, + 0.005658366251736879, + 0.005209960043430328, + 0.0036279952619224787, + 0.005585683509707451, + 0.0036253295838832855, + 0.0027183068450540304, + 0.00572167569771409, + 0.0026118983514606953, + 0.008173220790922642, + 0.005463615525513887, + 0.006309879943728447, + 0.007612076122313738, + 0.008529060520231724, + 0.006404966581612825, + 0.011754821985960007, + 0.006276090629398823, + 0.00949886254966259, + 0.015528698451817036, + 0.019346779212355614, + 0.012682548724114895, + 0.024003950878977776, + 0.016675474122166634, + 0.0328100249171257, + 0.08174590021371841, + 0.01818026788532734, + 0.009427778422832489, + 0.008640715852379799, + 0.011966053396463394, + 0.01320522092282772, + 0.011516723781824112, + 0.009913860820233822, + 0.019215084612369537, + 0.016803622245788574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.0007285801111720502 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075920425355434 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011075920425355434, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075920425355434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05302455648779869, + 0.020877525210380554, + 0.46959707140922546, + 0.018322385847568512, + 0.0322127491235733, + 0.05467090383172035, + 0.12090733647346497, + 0.026836493983864784, + 0.15461236238479614, + 0.04893859475851059 + ], + "credit_map": [ + 0.003931436687707901, + 0.00994417630136013, + 0.007486613001674414, + 0.006701827049255371, + 0.006301012821495533, + 0.006595234386622906, + 0.007810726761817932, + 0.0042535290122032166, + 0.002080756239593029, + 0.0026818052865564823, + 0.0021552436519414186, + 0.002845147391781211, + 0.002955747302621603, + 0.0024112092796713114, + 0.002828295575454831, + 0.0029193207155913115, + 0.024514086544513702, + 0.02281567081809044, + 0.1878749579191208, + 0.02486925572156906, + 0.016232410445809364, + 0.14573369920253754, + 0.025366593152284622, + 0.022190390154719353, + 0.0014639198780059814, + 0.0026753831189125776, + 0.0028774638194590807, + 0.0017802194925025105, + 0.0022961078211665154, + 0.0017091362969949841, + 0.002591605531051755, + 0.0029285515192896128, + 0.0025504555087536573, + 0.004095788579434156, + 0.004594461992383003, + 0.003609144128859043, + 0.003639346221461892, + 0.0036510860081762075, + 0.00688541354611516, + 0.0031870512757450342, + 0.003794660558924079, + 0.007530293893069029, + 0.020030302926898003, + 0.003796681761741638, + 0.003498806618154049, + 0.007893579080700874, + 0.004647946450859308, + 0.0034786327742040157, + 0.010320345871150494, + 0.00743749039247632, + 0.04202273115515709, + 0.00977636780589819, + 0.006999543868005276, + 0.013858078047633171, + 0.009655864909291267, + 0.020836923271417618, + 0.002602407243102789, + 0.0022751728538423777, + 0.003715568222105503, + 0.004367581568658352, + 0.003093505511060357, + 0.0053319670259952545, + 0.002331970725208521, + 0.003118320368230343, + 0.011520025320351124, + 0.01707671582698822, + 0.015406482852995396, + 0.020457813516259193, + 0.016824036836624146, + 0.013919433578848839, + 0.04349304735660553, + 0.015914807096123695, + 0.004009487107396126, + 0.004904481116682291, + 0.004233955405652523, + 0.006237454246729612, + 0.004610976204276085, + 0.005135176237672567, + 0.01146461721509695, + 0.008342452347278595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.011075920425355434 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1002957671880722 + ], + "opposite_logprob_tokens": [ + -2.3502957820892334 + ], + "logprob_reward": -0.1002957671880722, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1002957671880722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03981518745422363, + 0.02414611168205738, + 0.2511356770992279, + 0.03254377841949463, + 0.04266570881009102, + 0.022369863465428352, + 0.08705884218215942, + 0.03535662218928337, + 0.4233464300632477, + 0.0415617935359478 + ], + "credit_map": [ + 0.002969976980239153, + 0.0049945637583732605, + 0.0038772362750023603, + 0.0026081250980496407, + 0.003867257619276643, + 0.011570635251700878, + 0.0026683304458856583, + 0.007259058766067028, + 0.0018849654588848352, + 0.005984255578368902, + 0.0026540718972682953, + 0.0027586035430431366, + 0.0038927928544580936, + 0.002592366188764572, + 0.0022031988482922316, + 0.002175858709961176, + 0.013387659564614296, + 0.013228636234998703, + 0.09852657467126846, + 0.012959950603544712, + 0.010051687248051167, + 0.07675033062696457, + 0.01321704313158989, + 0.013013781979680061, + 0.0024689873680472374, + 0.0054532615467906, + 0.005465555004775524, + 0.003097501350566745, + 0.0038213932421058416, + 0.002994456561282277, + 0.00420328788459301, + 0.00503933522850275, + 0.0031892673578113317, + 0.005468000192195177, + 0.011103327386081219, + 0.004548260476440191, + 0.004575580824166536, + 0.005117517430335283, + 0.004168170038610697, + 0.004495583008974791, + 0.0015418629627674818, + 0.002851804019883275, + 0.007570090238004923, + 0.0016685806913301349, + 0.0013156614731997252, + 0.003422135254368186, + 0.0015780735993757844, + 0.002421655459329486, + 0.006994130555540323, + 0.006293428596109152, + 0.02927362732589245, + 0.007511167321354151, + 0.0052795675583183765, + 0.014333948493003845, + 0.007147012744098902, + 0.01022595539689064, + 0.004010360222309828, + 0.004951228387653828, + 0.005040672607719898, + 0.005077898968011141, + 0.004040180239826441, + 0.005956723820418119, + 0.003063517389819026, + 0.0032160391565412283, + 0.03248879685997963, + 0.05317388102412224, + 0.03949573636054993, + 0.051041129976511, + 0.04912861809134483, + 0.07167480885982513, + 0.07418597489595413, + 0.05215750262141228, + 0.0049759200774133205, + 0.004523209296166897, + 0.007970021106302738, + 0.004459974821656942, + 0.004546532407402992, + 0.0038433969020843506, + 0.005881489720195532, + 0.005361251998692751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 6 + ], + "selected_reward": -0.1002957671880722 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme cinematic macro shot, rendered in ultra-high definition, slowly zooms in on the interior of a gold-lipped pearl oyster, set against a deep, out-of-focus, abyssal black background. At the center of the frame, a lustrous pearl, its surface a creamy, milky white. Its nearly spherical shape catches soft, directed light, revealing a deep inner glow with subtle hints of rose and gold. The oyster's iridescent mantle gently nestles around the pearl, a delicate fleshy tissue that seems to breathe with slow, rhythmic pulsations, its folds delicately caressing the pearl's surface. The mantle's surface shimmers with a hypnotic, oily iridescence, its colors shifting and undulating delicately across a spectrum of malachite green, soft violet, and warm gold. The camera's deliberate movements capture the breathtaking play of light within the surrounding mother-of-pearl – a landscape of polished, undulating nacre, where every microscopic ridge and crystalline layer refracts light, creating a living universe teeming with color and texture. The focused, tranquil atmosphere is one of profound natural wonder, revealing the slow, beautiful creation of gemstones in stunning detail.", + "seed": 14263, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_11/index_6525_seed_14263_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06625441284850239, + "vlm_logprob": -0.06625441284850239, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.06625441284850239, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020177112892270088 + ], + "opposite_logprob_tokens": [ + -4.270176887512207 + ], + "logprob_reward": -0.020177112892270088, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.020177112892270088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11190713942050934, + 0.07557238638401031, + 0.03820019215345383, + 0.025328025221824646, + 0.040224164724349976, + 0.04237464815378189, + 0.25499385595321655, + 0.13665959239006042, + 0.24308617413043976, + 0.031653814017772675 + ], + "credit_map": [ + 0.021662520244717598, + 0.020142486318945885, + 0.009190421551465988, + 0.01024475134909153, + 0.012232715263962746, + 0.011926607228815556, + 0.014433838427066803, + 0.012073811143636703, + 0.012432672083377838, + 0.013156340457499027, + 0.006064356304705143, + 0.006484450772404671, + 0.008791868574917316, + 0.009960194118320942, + 0.012161617167294025, + 0.006520895287394524, + 0.005940242670476437, + 0.0041901241056621075, + 0.0039409757591784, + 0.006307270843535662, + 0.003153868718072772, + 0.005122576840221882, + 0.005688879638910294, + 0.0038562575355172157, + 0.0036326569970697165, + 0.003444513538852334, + 0.0024204463697969913, + 0.003529890673235059, + 0.0029237717390060425, + 0.004684239625930786, + 0.0030583387706428766, + 0.0016341697191819549, + 0.0064764851704239845, + 0.007240880746394396, + 0.00437135761603713, + 0.0038936228957027197, + 0.005087109282612801, + 0.004648619797080755, + 0.004532685037702322, + 0.003973407205194235, + 0.0069360677152872086, + 0.004946965258568525, + 0.005361683201044798, + 0.006270377431064844, + 0.00431791739538312, + 0.004979387391358614, + 0.005135000683367252, + 0.004427253268659115, + 0.05002989247441292, + 0.025051072239875793, + 0.018186289817094803, + 0.022470060735940933, + 0.03819049894809723, + 0.041793473064899445, + 0.03970973193645477, + 0.019562849774956703, + 0.020909877493977547, + 0.013413496315479279, + 0.023740405216813087, + 0.01603076606988907, + 0.016365783289074898, + 0.01577603630721569, + 0.015351017005741596, + 0.015072216279804707, + 0.03160187602043152, + 0.028736000880599022, + 0.03525703027844429, + 0.02972630225121975, + 0.03315453603863716, + 0.030523987486958504, + 0.024782633408904076, + 0.02930382639169693, + 0.004157552029937506, + 0.0035622308496385813, + 0.004920915700495243, + 0.00447705527767539, + 0.0038706595078110695, + 0.003916293848305941, + 0.0036283957306295633, + 0.0031207113061100245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.020177112892270088 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038739681243896 + ], + "opposite_logprob_tokens": [ + -1.910387396812439 + ], + "logprob_reward": -0.16038739681243896, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038739681243896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12467612326145172, + 0.07106844335794449, + 0.03716310113668442, + 0.05257788673043251, + 0.0593847818672657, + 0.040002405643463135, + 0.3783421814441681, + 0.1670323610305786, + 0.048636820167303085, + 0.021115917712450027 + ], + "credit_map": [ + 0.026897339150309563, + 0.016497693955898285, + 0.010520228184759617, + 0.012659832835197449, + 0.01084411796182394, + 0.017810309305787086, + 0.018760550767183304, + 0.010686046443879604, + 0.015004908666014671, + 0.013181219808757305, + 0.005285671912133694, + 0.0057965051382780075, + 0.008664590306580067, + 0.01040746457874775, + 0.007181803695857525, + 0.005546282976865768, + 0.006200852803885937, + 0.0056929378770291805, + 0.004661676473915577, + 0.004025946371257305, + 0.004326826427131891, + 0.0052550723776221275, + 0.003547844709828496, + 0.0034519443288445473, + 0.010045390576124191, + 0.008200537413358688, + 0.0039421864785254, + 0.004705915227532387, + 0.006840329617261887, + 0.008686386048793793, + 0.005593796260654926, + 0.004563340917229652, + 0.010163985192775726, + 0.006619160994887352, + 0.004941025748848915, + 0.0072096665389835835, + 0.007105154450982809, + 0.012199395336210728, + 0.00649444293230772, + 0.004651953931897879, + 0.007717510219663382, + 0.005275023635476828, + 0.0040804482996463776, + 0.00518658384680748, + 0.004229943733662367, + 0.005423438735306263, + 0.004429971799254417, + 0.003659485373646021, + 0.07543124258518219, + 0.03832710534334183, + 0.02935696206986904, + 0.03282872214913368, + 0.059211257845163345, + 0.05830405652523041, + 0.05678635090589523, + 0.028096480295062065, + 0.026825107634067535, + 0.02261255495250225, + 0.021711088716983795, + 0.01824374683201313, + 0.015761451795697212, + 0.0181178729981184, + 0.027198996394872665, + 0.01656154729425907, + 0.007395897060632706, + 0.00637871352955699, + 0.005959653295576572, + 0.004572974517941475, + 0.006687103305011988, + 0.006481177639216185, + 0.006965871900320053, + 0.004195427987724543, + 0.002647077664732933, + 0.0026642109733074903, + 0.0032457986380904913, + 0.0029267813079059124, + 0.0023009758442640305, + 0.002660782542079687, + 0.0024180489126592875, + 0.0022522418294101954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.16038739681243896 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001957288011908531 + ], + "opposite_logprob_tokens": [ + -6.251957416534424 + ], + "logprob_reward": -0.001957288011908531, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001957288011908531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33586689829826355, + 0.09698571264743805, + 0.04436348378658295, + 0.031101295724511147, + 0.032498233020305634, + 0.04978087171912193, + 0.16977337002754211, + 0.11095866560935974, + 0.07851528376340866, + 0.050156280398368835 + ], + "credit_map": [ + 0.05573665350675583, + 0.039218463003635406, + 0.02525651641190052, + 0.0359070710837841, + 0.029050584882497787, + 0.056001994758844376, + 0.057952556759119034, + 0.03674302622675896, + 0.01627511903643608, + 0.008639011532068253, + 0.008672554977238178, + 0.013964823447167873, + 0.011699602007865906, + 0.01319124735891819, + 0.014895811676979065, + 0.009647546336054802, + 0.006847017910331488, + 0.004772518295794725, + 0.005907644052058458, + 0.00602286821231246, + 0.005421583075076342, + 0.004606804810464382, + 0.005819346755743027, + 0.004965701140463352, + 0.004185993690043688, + 0.004494329448789358, + 0.0027050559874624014, + 0.003956041764467955, + 0.004487868398427963, + 0.00410970626398921, + 0.004441344644874334, + 0.002720955293625593, + 0.0060035246424376965, + 0.004163188394159079, + 0.0032527954317629337, + 0.003862536745145917, + 0.00322295562364161, + 0.0037412599194794893, + 0.004421034827828407, + 0.0038309376686811447, + 0.008297057822346687, + 0.004914294462651014, + 0.0064749950543046, + 0.007543371059000492, + 0.004847328644245863, + 0.006263036280870438, + 0.005785355344414711, + 0.005655434913933277, + 0.03263905271887779, + 0.016866156831383705, + 0.011362564750015736, + 0.016870638355612755, + 0.024724405258893967, + 0.026582907885313034, + 0.027199767529964447, + 0.013527859933674335, + 0.017046824097633362, + 0.012427657842636108, + 0.012760244309902191, + 0.014450070448219776, + 0.008483939804136753, + 0.012924425303936005, + 0.018533064052462578, + 0.01433243416249752, + 0.01002101693302393, + 0.010088921524584293, + 0.011150631122291088, + 0.008567314594984055, + 0.011576646007597446, + 0.010165788233280182, + 0.009652478620409966, + 0.007292488589882851, + 0.006550834979861975, + 0.005301894154399633, + 0.007216871716082096, + 0.007745614740997553, + 0.006024649366736412, + 0.005517462734133005, + 0.006665595807135105, + 0.005133357830345631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.001957288011908531 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953317973762751 + ], + "opposite_logprob_tokens": [ + -5.130953311920166 + ], + "logprob_reward": -0.005953317973762751, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953317973762751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1963665783405304, + 0.05967085808515549, + 0.05197574943304062, + 0.07677966356277466, + 0.045197248458862305, + 0.04847734421491623, + 0.21119974553585052, + 0.17187054455280304, + 0.07045037299394608, + 0.06801193952560425 + ], + "credit_map": [ + 0.03566513955593109, + 0.02084778994321823, + 0.01603352651000023, + 0.025337716564536095, + 0.01814029924571514, + 0.03197534382343292, + 0.03373178839683533, + 0.01463494822382927, + 0.009092292748391628, + 0.010955643840134144, + 0.004200007300823927, + 0.007632257416844368, + 0.006643272936344147, + 0.0072007752023637295, + 0.008370626717805862, + 0.005575977265834808, + 0.011135702021420002, + 0.008392661809921265, + 0.004286966752260923, + 0.0040025049820542336, + 0.006050500553101301, + 0.006442919839173555, + 0.007699772249907255, + 0.003964714705944061, + 0.015482489950954914, + 0.009928579442203045, + 0.006240352522581816, + 0.007340616080909967, + 0.009088435210287571, + 0.012182149104773998, + 0.010205741971731186, + 0.006311296485364437, + 0.007061999291181564, + 0.007322659250348806, + 0.004222087096422911, + 0.0050508007407188416, + 0.005596559029072523, + 0.005541135091334581, + 0.006551709026098251, + 0.003850294044241309, + 0.00804071594029665, + 0.005496099125593901, + 0.006480899639427662, + 0.005956739187240601, + 0.006150528788566589, + 0.005527028348296881, + 0.005451741628348827, + 0.00537358270958066, + 0.042921457439661026, + 0.02271546609699726, + 0.01676006056368351, + 0.020004626363515854, + 0.03058805875480175, + 0.0293611828237772, + 0.03190243989229202, + 0.01694643311202526, + 0.027392815798521042, + 0.016957785934209824, + 0.023971354588866234, + 0.025565778836607933, + 0.01936178095638752, + 0.022096820175647736, + 0.017763599753379822, + 0.018760589882731438, + 0.009558242745697498, + 0.010136479511857033, + 0.008343372493982315, + 0.0072892457246780396, + 0.008763648569583893, + 0.008553444407880306, + 0.009358842857182026, + 0.00844709575176239, + 0.008491276763379574, + 0.00791371800005436, + 0.009490211494266987, + 0.007056127768009901, + 0.012463657185435295, + 0.009090766310691833, + 0.00784992054104805, + 0.005656254477798939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.005953317973762751 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279694855213165 + ], + "opposite_logprob_tokens": [ + -2.0177969932556152 + ], + "logprob_reward": -0.14279694855213165, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14279694855213165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15226206183433533, + 0.051975131034851074, + 0.032678887248039246, + 0.06658589839935303, + 0.04724443331360817, + 0.09151046723127365, + 0.09812850505113602, + 0.3025500178337097, + 0.038582198321819305, + 0.11848241090774536 + ], + "credit_map": [ + 0.023935770615935326, + 0.01953149400651455, + 0.012682604603469372, + 0.018679305911064148, + 0.014719090424478054, + 0.02436111867427826, + 0.020518774166703224, + 0.017833901569247246, + 0.008366771973669529, + 0.006815860513597727, + 0.005116969347000122, + 0.006075186189264059, + 0.007218246813863516, + 0.007948649115860462, + 0.005817186087369919, + 0.004616260062903166, + 0.005646105855703354, + 0.0052621373906731606, + 0.0024281865917146206, + 0.003806196851655841, + 0.005198150873184204, + 0.003915791865438223, + 0.004016932565718889, + 0.002405382925644517, + 0.009889504872262478, + 0.007064440753310919, + 0.007770292926579714, + 0.010890519246459007, + 0.009107283316552639, + 0.008779310621321201, + 0.007233795709908009, + 0.0058507490903139114, + 0.00694205891340971, + 0.006874514743685722, + 0.006242229137569666, + 0.00736478203907609, + 0.004671397153288126, + 0.005082021001726389, + 0.005829316098242998, + 0.004238113295286894, + 0.015103471465408802, + 0.010755905881524086, + 0.01068804133683443, + 0.013805896043777466, + 0.009742457419633865, + 0.011788299307227135, + 0.00971092190593481, + 0.009915470145642757, + 0.016354290768504143, + 0.009428604505956173, + 0.008822400122880936, + 0.01011823583394289, + 0.013982413336634636, + 0.017692333087325096, + 0.014122958295047283, + 0.0076072728261351585, + 0.04972340911626816, + 0.029282309114933014, + 0.03335190564393997, + 0.032268986105918884, + 0.03389233350753784, + 0.037845104932785034, + 0.05038752406835556, + 0.035798460245132446, + 0.005689731799066067, + 0.004857752937823534, + 0.005066125653684139, + 0.004469907842576504, + 0.005023157224059105, + 0.004276582971215248, + 0.004953209776431322, + 0.004245728719979525, + 0.014794321730732918, + 0.01609344594180584, + 0.016274889931082726, + 0.015483157709240913, + 0.017496159300208092, + 0.015856238082051277, + 0.011235801503062248, + 0.011248398572206497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.14279694855213165 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme cinematic macro shot, rendered in ultra-high definition, slowly zooms in on the interior of a gold-lipped pearl oyster, set against a deep, out-of-focus, abyssal black background. At the center of the frame, a lustrous pearl, its surface a creamy, milky white. Its nearly spherical shape catches soft, directed light, revealing a deep inner glow with subtle hints of rose and gold. The oyster's iridescent mantle gently nestles around the pearl, a delicate fleshy tissue that seems to breathe with slow, rhythmic pulsations, its folds delicately caressing the pearl's surface. The mantle's surface shimmers with a hypnotic, oily iridescence, its colors shifting and undulating delicately across a spectrum of malachite green, soft violet, and warm gold. The camera's deliberate movements capture the breathtaking play of light within the surrounding mother-of-pearl – a landscape of polished, undulating nacre, where every microscopic ridge and crystalline layer refracts light, creating a living universe teeming with color and texture. The focused, tranquil atmosphere is one of profound natural wonder, revealing the slow, beautiful creation of gemstones in stunning detail.", + "seed": 14264, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_11/index_6525_seed_14264_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05283922269009054, + "vlm_logprob": -0.05283922269009054, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.05283922269009054, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009907475672662258 + ], + "opposite_logprob_tokens": [ + -5.259907245635986 + ], + "logprob_reward": -0.009907475672662258, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.009907475672662258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11020715534687042, + 0.1209961324930191, + 0.07751542329788208, + 0.0368569977581501, + 0.03639077767729759, + 0.1303647756576538, + 0.15046852827072144, + 0.12288632988929749, + 0.11273876577615738, + 0.10157513618469238 + ], + "credit_map": [ + 0.016069237142801285, + 0.008633705787360668, + 0.021204376593232155, + 0.01230283360928297, + 0.006098250392824411, + 0.008332070894539356, + 0.011754936538636684, + 0.02581174299120903, + 0.01681782677769661, + 0.005987243726849556, + 0.02638203464448452, + 0.014121916145086288, + 0.006246963981539011, + 0.00893899891525507, + 0.011376157402992249, + 0.031124981120228767, + 0.009870843961834908, + 0.008275371976196766, + 0.012839383445680141, + 0.009734131395816803, + 0.011814028024673462, + 0.007108353544026613, + 0.009095669724047184, + 0.00877764355391264, + 0.0038264549802988768, + 0.005286155268549919, + 0.004455316346138716, + 0.007085558492690325, + 0.0034654918126761913, + 0.003954944666475058, + 0.00404093973338604, + 0.004742132965475321, + 0.0040830885991454124, + 0.0035754884593188763, + 0.005040125455707312, + 0.0052867550402879715, + 0.002860755193978548, + 0.004374257288873196, + 0.005565156694501638, + 0.0056051514111459255, + 0.01422579400241375, + 0.011411015875637531, + 0.030279401689767838, + 0.027636490762233734, + 0.008904241025447845, + 0.00929240882396698, + 0.017084874212741852, + 0.011530548334121704, + 0.014100099913775921, + 0.032062314450740814, + 0.01175991352647543, + 0.02004348486661911, + 0.0093384999781847, + 0.015816209837794304, + 0.02606651745736599, + 0.021281490102410316, + 0.009433180093765259, + 0.010108181275427341, + 0.012454131618142128, + 0.026426024734973907, + 0.015539588406682014, + 0.019887324422597885, + 0.01489932555705309, + 0.014138574711978436, + 0.009534802287817001, + 0.02774094231426716, + 0.008478442206978798, + 0.02351032942533493, + 0.008057021535933018, + 0.00910171214491129, + 0.016012175008654594, + 0.010303344577550888, + 0.012326342985033989, + 0.006087454967200756, + 0.006990028079599142, + 0.031401701271533966, + 0.006244909483939409, + 0.0103375893086195, + 0.01854463294148445, + 0.009642479941248894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.009907475672662258 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003770768642426 + ], + "opposite_logprob_tokens": [ + -2.695037603378296 + ], + "logprob_reward": -0.07003770768642426, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003770768642426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25958865880966187, + 0.13867630064487457, + 0.09796827286481857, + 0.031160950660705566, + 0.023961449041962624, + 0.022422274574637413, + 0.06130899116396904, + 0.12559418380260468, + 0.041226375848054886, + 0.1980925053358078 + ], + "credit_map": [ + 0.03550656512379646, + 0.017391514033079147, + 0.034660324454307556, + 0.033571772277355194, + 0.017886413261294365, + 0.01943865790963173, + 0.035705581307411194, + 0.06542783230543137, + 0.018371716141700745, + 0.00813406240195036, + 0.020197153091430664, + 0.012672877870500088, + 0.007922829128801823, + 0.011514612473547459, + 0.014268518425524235, + 0.04559453949332237, + 0.010207809507846832, + 0.010763418860733509, + 0.01696406863629818, + 0.009479538537561893, + 0.010985489003360271, + 0.008866225369274616, + 0.012863853946328163, + 0.017837869003415108, + 0.0029058947693556547, + 0.00346859241835773, + 0.0035369829274713993, + 0.005239780992269516, + 0.0029425551183521748, + 0.004317086655646563, + 0.0033825128339231014, + 0.0053675477392971516, + 0.0027016866952180862, + 0.0026178117841482162, + 0.0030932719819247723, + 0.0043347119353711605, + 0.0019476644229143858, + 0.002495825756341219, + 0.0035561465192586184, + 0.0032143311109393835, + 0.0023666080087423325, + 0.001766488770954311, + 0.006215011700987816, + 0.0034453070256859064, + 0.0013485756935551763, + 0.0017712060362100601, + 0.003559503238648176, + 0.0019495758460834622, + 0.007098120171576738, + 0.0056927939876914024, + 0.008438223972916603, + 0.008623176254332066, + 0.0057957167737185955, + 0.009553153067827225, + 0.009366603568196297, + 0.006741207093000412, + 0.01346579845994711, + 0.012703217566013336, + 0.017928091809153557, + 0.018503978848457336, + 0.011815913021564484, + 0.016640298068523407, + 0.013739670626819134, + 0.02079722471535206, + 0.0036456931848078966, + 0.009052311070263386, + 0.002965030260384083, + 0.008948009461164474, + 0.003643265925347805, + 0.006100254599004984, + 0.004071831237524748, + 0.002799980342388153, + 0.029479218646883965, + 0.011355897411704063, + 0.013073249720036983, + 0.07702402770519257, + 0.0064346096478402615, + 0.013692663051187992, + 0.03774522617459297, + 0.009287625551223755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.07003770768642426 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001191859133541584 + ], + "opposite_logprob_tokens": [ + -6.751192092895508 + ], + "logprob_reward": -0.001191859133541584, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001191859133541584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20779412984848022, + 0.12408457696437836, + 0.10256383568048477, + 0.10156372934579849, + 0.06393741071224213, + 0.1597377359867096, + 0.06105583533644676, + 0.028079334646463394, + 0.04983700439333916, + 0.10134639590978622 + ], + "credit_map": [ + 0.02298837900161743, + 0.009755889885127544, + 0.020008036866784096, + 0.024800557643175125, + 0.012576653622090816, + 0.018907656893134117, + 0.04523087665438652, + 0.05352608859539032, + 0.017296433448791504, + 0.0075212703086435795, + 0.014328859746456146, + 0.0116525674238801, + 0.007615822367370129, + 0.011735562235116959, + 0.011658135801553726, + 0.04227593168616295, + 0.008674845099449158, + 0.012845374643802643, + 0.01293351873755455, + 0.011301281861960888, + 0.01929977908730507, + 0.008612185716629028, + 0.016419043764472008, + 0.012477803975343704, + 0.013185491785407066, + 0.00922323763370514, + 0.01448790542781353, + 0.016901832073926926, + 0.008210079744458199, + 0.008407644927501678, + 0.011954400688409805, + 0.0191931314766407, + 0.00759412907063961, + 0.005805091001093388, + 0.008477023802697659, + 0.014977513812482357, + 0.004892594646662474, + 0.005640189163386822, + 0.01037969533354044, + 0.006171171087771654, + 0.01668822020292282, + 0.014192648231983185, + 0.023080652579665184, + 0.037409983575344086, + 0.012558880262076855, + 0.013320576399564743, + 0.0247020460665226, + 0.017784731462597847, + 0.006074467208236456, + 0.011842980049550533, + 0.005455531645566225, + 0.007430884055793285, + 0.004117283970117569, + 0.0067113246768713, + 0.010190573520958424, + 0.009232793003320694, + 0.0028985189273953438, + 0.0037056580185890198, + 0.0033734857570379972, + 0.004192870110273361, + 0.0028433995321393013, + 0.0038150653708726168, + 0.0030989646911621094, + 0.004151370842009783, + 0.004956408403813839, + 0.007661147508770227, + 0.004632617812603712, + 0.009260876104235649, + 0.004544361028820276, + 0.006862437352538109, + 0.005560303572565317, + 0.006358856335282326, + 0.008569451980292797, + 0.00799558311700821, + 0.01106965634971857, + 0.023129627108573914, + 0.006756008602678776, + 0.01080433838069439, + 0.019498353824019432, + 0.01352338120341301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.001191859133541584 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003194113029167056 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.003194113029167056, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003194113029167056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12548883259296417, + 0.1429406702518463, + 0.20626258850097656, + 0.048398297280073166, + 0.05467066541314125, + 0.08469727635383606, + 0.024074915796518326, + 0.0836460217833519, + 0.08320222795009613, + 0.14661847054958344 + ], + "credit_map": [ + 0.018125927075743675, + 0.011141317896544933, + 0.022925768047571182, + 0.016478249803185463, + 0.009292359463870525, + 0.010414906777441502, + 0.013770119287073612, + 0.023340191692113876, + 0.024814879521727562, + 0.007982742972671986, + 0.04392116889357567, + 0.014959382824599743, + 0.006538113113492727, + 0.007641870528459549, + 0.014422816224396229, + 0.022659700363874435, + 0.025504644960165024, + 0.01713198609650135, + 0.022765086963772774, + 0.023258542641997337, + 0.018722794950008392, + 0.019247129559516907, + 0.025008130818605423, + 0.0546242780983448, + 0.005674963351339102, + 0.004269696772098541, + 0.011859487742185593, + 0.0072493222542107105, + 0.003170300740748644, + 0.004478898365050554, + 0.005562096368521452, + 0.006133532617241144, + 0.005807822570204735, + 0.005673430860042572, + 0.006599221378564835, + 0.0118157509714365, + 0.005031068343669176, + 0.005428318399935961, + 0.00876868236809969, + 0.005546368192881346, + 0.01196569949388504, + 0.005347437225282192, + 0.02678264118731022, + 0.012446637265384197, + 0.00500244926661253, + 0.0051215109415352345, + 0.01120710838586092, + 0.006823789328336716, + 0.003081769682466984, + 0.0022679620888084173, + 0.0037163239903748035, + 0.003773312782868743, + 0.0017108723986893892, + 0.0025415061973035336, + 0.003123605391010642, + 0.0038595618680119514, + 0.01029446441680193, + 0.008539566770195961, + 0.010792423970997334, + 0.013007397763431072, + 0.009964991360902786, + 0.010268268175423145, + 0.009949849918484688, + 0.010829059407114983, + 0.006387998815625906, + 0.007412318605929613, + 0.010460381396114826, + 0.016185199841856956, + 0.009868117049336433, + 0.01145483460277319, + 0.011180383153259754, + 0.010252993553876877, + 0.026333866640925407, + 0.010269402526319027, + 0.012383418157696724, + 0.044107191264629364, + 0.0070403399877250195, + 0.011400409042835236, + 0.02318195067346096, + 0.011901887133717537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.003194113029167056 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of the interior of a gold-lipped pearl oyster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, out-of-focus, and abyssal black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pearl creamy, milky white, nearly spherical, and showing a deep inner glow with hints of rose and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oyster's mantle shimmer with iridescence shifting between malachite green, soft violet, and warm gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986495792865753 + ], + "opposite_logprob_tokens": [ + -1.8048650026321411 + ], + "logprob_reward": -0.17986495792865753, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986495792865753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1340060979127884, + 0.03334721177816391, + 0.017561787739396095, + 0.03460310399532318, + 0.022919153794646263, + 0.0637616366147995, + 0.05957966297864914, + 0.056573253124952316, + 0.06443070620298386, + 0.5132173895835876 + ], + "credit_map": [ + 0.0194480512291193, + 0.013791004195809364, + 0.028088022023439407, + 0.01722508668899536, + 0.010510246269404888, + 0.01135143544524908, + 0.014178884215652943, + 0.019413379952311516, + 0.004508081823587418, + 0.0020009276922792196, + 0.0038373665884137154, + 0.004795609507709742, + 0.002352121053263545, + 0.0021854983642697334, + 0.004321240354329348, + 0.009346366859972477, + 0.0022631308529525995, + 0.001557873678393662, + 0.004249261692166328, + 0.0021297764033079147, + 0.0016176477074623108, + 0.0012820787960663438, + 0.0020920319948345423, + 0.0023699866142123938, + 0.003923294600099325, + 0.0038762714248150587, + 0.00491549214348197, + 0.004824321251362562, + 0.0024332490283995867, + 0.0036329878494143486, + 0.004029319621622562, + 0.006968169007450342, + 0.002540928078815341, + 0.0020433603785932064, + 0.004834576975554228, + 0.0049342443235218525, + 0.0018604039214551449, + 0.0020035244524478912, + 0.002614769618958235, + 0.002087346976622939, + 0.008158624172210693, + 0.004157327581197023, + 0.023225286975502968, + 0.00971777830272913, + 0.0029176909010857344, + 0.0024503921158611774, + 0.009126621298491955, + 0.004007913637906313, + 0.007934262976050377, + 0.005825507454574108, + 0.00630759634077549, + 0.007505303248763084, + 0.004590800032019615, + 0.008201928809285164, + 0.007870791479945183, + 0.011343469843268394, + 0.004802589304745197, + 0.0070546893402934074, + 0.007920610718429089, + 0.00898619182407856, + 0.005364158656448126, + 0.005959635134786367, + 0.007334443274885416, + 0.009150938130915165, + 0.004808509722352028, + 0.01180196087807417, + 0.005055244080722332, + 0.012032372877001762, + 0.005566015373915434, + 0.010017548687756062, + 0.008257586508989334, + 0.006891467142850161, + 0.07653484493494034, + 0.04170478135347366, + 0.036561593413352966, + 0.17469161748886108, + 0.01887001283466816, + 0.03391098231077194, + 0.09538194537162781, + 0.03556161746382713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.17986495792865753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_12/index_tensor([5924, 5924])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_12/index_tensor([5924, 5924])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b00746a47548ae2f308c3151bea571f74a61826f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_12/index_tensor([5924, 5924])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly descending, gently circling drone shot offers awe-inspiring views of a vast golden lotus throne floating in an abstract void. Constructed not of smooth gold, the throne reveals the texture of gilded metal weathered by time. Its surface is riddled with minute cracks and mottled colors. Simple, powerful lines outline each petal, imbued with the allure of Dunhuang murals. The throne's massive petals, layered one after another, steadily and gracefully unfold with a solemn and leisurely rhythm. As the innermost petals unfold, the compassionate figure of Guanyin, seated at the throne's center, is fully revealed. She clad in a pure, immaculate white silk robe, its folds flowing smoothly and simply. She sits in a serene lotus posture, her face serene and compassionate. Her eyes, slightly downcast, seem to survey the world, a faint, transcendent smile playing upon her lips. The entire scene is bathed in a soft, sacred golden light emanating from the lotus and Guanyin herself. This light casts gentle shadows across the mottled golden surface, highlighting its ancient texture. The backdrop, rather than the real world, features stylized auspicious cloud patterns composed of mineral pigments such as ochre, cinnabar, and malachite green. This perfectly replicates the weathered aesthetic of millennia-old murals, creating an atmosphere of solemnity, tranquility, and sacredness that transcends time and space.", + "seed": 18519, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_12/index_5924_seed_18519_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07060855496674776, + "vlm_logprob": -0.07060855496674776, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.07060855496674776, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20968875288963318 + ], + "opposite_logprob_tokens": [ + -1.7096887826919556 + ], + "logprob_reward": -0.20968875288963318, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20968875288963318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014404544606804848, + 0.009439143352210522, + 0.008497819304466248, + 0.04073832556605339, + 0.019791139289736748, + 0.8630654811859131, + 0.009405624121427536, + 0.014555883593857288, + 0.008282039314508438, + 0.011819998733699322 + ], + "credit_map": [ + 0.001973534468561411, + 0.001855434151366353, + 0.001404029200784862, + 0.0026320957113057375, + 0.0012553317937999964, + 0.0018699007341638207, + 0.0012842167634516954, + 0.002130001550540328, + 0.0017005604458972812, + 0.0012739819940179586, + 0.0006810381892137229, + 0.0009367825696244836, + 0.0009297741344198585, + 0.001848954358138144, + 0.0010510723805055022, + 0.0010169794550165534, + 0.0011951049091294408, + 0.001320985029451549, + 0.0007304351893253624, + 0.0008691954426467419, + 0.0010234143119305372, + 0.001556862611323595, + 0.0008709273533895612, + 0.0009308941080234945, + 0.003239464247599244, + 0.0018725375412032008, + 0.001388276694342494, + 0.0019620959646999836, + 0.0016294855158776045, + 0.01586904563009739, + 0.0018122921464964747, + 0.01296512596309185, + 0.006303807254880667, + 0.0015537143917754292, + 0.0016182370018213987, + 0.0012444683816283941, + 0.001447469461709261, + 0.004088339861482382, + 0.001419072155840695, + 0.0021160305477678776, + 0.04280874878168106, + 0.023264897987246513, + 0.009081563912332058, + 0.021824490278959274, + 0.3507024943828583, + 0.36986562609672546, + 0.017862919718027115, + 0.027654753997921944, + 0.0006707301363348961, + 0.0008005158160813153, + 0.0006334874196909368, + 0.0011654775589704514, + 0.0023946226574480534, + 0.0022955762688070536, + 0.0006393110379576683, + 0.0008059032261371613, + 0.0023148932959884405, + 0.0019439351744949818, + 0.0014529715990647674, + 0.0019472823478281498, + 0.0018721442902460694, + 0.0022863890044391155, + 0.0012555744033306837, + 0.0014826927799731493, + 0.0008915149956010282, + 0.0010114104952663183, + 0.0008271699189208448, + 0.001485633198171854, + 0.0009103014599531889, + 0.001233161659911275, + 0.0010794036788865924, + 0.0008434440242126584, + 0.0021034199744462967, + 0.0009793714853003621, + 0.0032041443046182394, + 0.000790979596786201, + 0.0009364548604935408, + 0.0015343440463766456, + 0.0005999049753881991, + 0.001671380246989429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.20968875288963318 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292584985494614 + ], + "opposite_logprob_tokens": [ + -2.2379257678985596 + ], + "logprob_reward": -0.11292584985494614, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292584985494614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13625624775886536, + 0.022747589275240898, + 0.028370724990963936, + 0.07270408421754837, + 0.05544114112854004, + 0.4975907802581787, + 0.021700801327824593, + 0.10707283765077591, + 0.0299887303262949, + 0.028127044439315796 + ], + "credit_map": [ + 0.014362392015755177, + 0.012264106422662735, + 0.014376177452504635, + 0.012593022547662258, + 0.011299656704068184, + 0.03094913624227047, + 0.013048524968326092, + 0.027363231405615807, + 0.003045216901227832, + 0.003251477610319853, + 0.002319859806448221, + 0.002056894823908806, + 0.0023225771728903055, + 0.005158039275556803, + 0.002056712983176112, + 0.002536810701712966, + 0.003928679507225752, + 0.003055407665669918, + 0.0040784492157399654, + 0.0031678779050707817, + 0.00255145994015038, + 0.005505794193595648, + 0.0032801611814647913, + 0.002802897710353136, + 0.00533861480653286, + 0.002581981010735035, + 0.0017001698724925518, + 0.0034762625582516193, + 0.002697242423892021, + 0.02986886166036129, + 0.003318958915770054, + 0.02372199483215809, + 0.00767146609723568, + 0.0032698188442736864, + 0.003581531345844269, + 0.0035165417939424515, + 0.004677398130297661, + 0.019013654440641403, + 0.002945106476545334, + 0.010765626095235348, + 0.024410400539636612, + 0.013468298129737377, + 0.005614380817860365, + 0.0129695488139987, + 0.20303016901016235, + 0.2115495204925537, + 0.010324104689061642, + 0.016224350780248642, + 0.0022466378286480904, + 0.0022203035186976194, + 0.0024682742077857256, + 0.0029623841401189566, + 0.003032564651221037, + 0.004769982770085335, + 0.0018713101744651794, + 0.002129344502463937, + 0.00815974548459053, + 0.006926469970494509, + 0.005060127470642328, + 0.0072813136503100395, + 0.034286148846149445, + 0.034877147525548935, + 0.005717959720641375, + 0.004763917997479439, + 0.0033423115964978933, + 0.0023656091652810574, + 0.003779646707698703, + 0.003179239109158516, + 0.004691585432738066, + 0.006744210608303547, + 0.0021505814511328936, + 0.0037355460226535797, + 0.002469644183292985, + 0.002770985011011362, + 0.004778039176017046, + 0.0023567418102174997, + 0.004940918646752834, + 0.0035882298834621906, + 0.002270953729748726, + 0.004951531067490578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.11292584985494614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069423407316208 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.016069423407316208, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069423407316208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049496959894895554, + 0.010373054072260857, + 0.009309702552855015, + 0.04020833596587181, + 0.0420060008764267, + 0.7777890563011169, + 0.01393461599946022, + 0.010865950025618076, + 0.02128436416387558, + 0.024731986224651337 + ], + "credit_map": [ + 0.0060119107365608215, + 0.00681895250454545, + 0.005653054919093847, + 0.0043539246544241905, + 0.0062673273496329784, + 0.00871917512267828, + 0.005388657096773386, + 0.006283957045525312, + 0.0015937428688630462, + 0.0009446979966014624, + 0.0009647101978771389, + 0.001139066880568862, + 0.0008257674635387957, + 0.0028384122997522354, + 0.000870999414473772, + 0.001195657067000866, + 0.0016201288672164083, + 0.0011163087328895926, + 0.0007107029086910188, + 0.0009023137972690165, + 0.0010634801583364606, + 0.0016365788178518414, + 0.0012543844059109688, + 0.0010058039333671331, + 0.0038143207784742117, + 0.0018683772068470716, + 0.0013474597362801433, + 0.00203963671810925, + 0.0015637855976819992, + 0.01527426391839981, + 0.0015959154115989804, + 0.012704573571681976, + 0.007674010936170816, + 0.001932749175466597, + 0.0024988956283777952, + 0.0021642372012138367, + 0.0023525534197688103, + 0.015076831914484501, + 0.0019571140874177217, + 0.00834961049258709, + 0.03795072063803673, + 0.02145863138139248, + 0.008359292522072792, + 0.020141547545790672, + 0.3151327669620514, + 0.3336680829524994, + 0.01607513055205345, + 0.02500287815928459, + 0.001223903615027666, + 0.0010726662585511804, + 0.0009223908418789506, + 0.001357354922220111, + 0.003323835553601384, + 0.0037819298449903727, + 0.0009176546009257436, + 0.0013348800130188465, + 0.0017077350057661533, + 0.00133914640173316, + 0.0008257150184363127, + 0.0014770259149372578, + 0.0014681464526802301, + 0.001925081480294466, + 0.0010917445179075003, + 0.0010313554666936398, + 0.0032327400986105204, + 0.0020925155840814114, + 0.001903329393826425, + 0.0027442581485956907, + 0.00262580718845129, + 0.004550119861960411, + 0.002193621825426817, + 0.0019419706659391522, + 0.0031838780269026756, + 0.0024417866952717304, + 0.005765416659414768, + 0.0019544323440641165, + 0.0036207609809935093, + 0.0027868275064975023, + 0.0014391980366781354, + 0.003539685858413577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.016069423407316208 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734294816851616 + ], + "opposite_logprob_tokens": [ + -5.006734371185303 + ], + "logprob_reward": -0.006734294816851616, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734294816851616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028142783790826797, + 0.011159163899719715, + 0.009621650911867619, + 0.027802156284451485, + 0.01749078929424286, + 0.7587721347808838, + 0.027338329702615738, + 0.05828157067298889, + 0.021149804815649986, + 0.04024164006114006 + ], + "credit_map": [ + 0.0038125552237033844, + 0.0028561267536133528, + 0.003373085055500269, + 0.002396267605945468, + 0.003272333648055792, + 0.005120046902447939, + 0.003964471165090799, + 0.0033478946425020695, + 0.0019516865722835064, + 0.0010406294604763389, + 0.001005770405754447, + 0.001215467811562121, + 0.001093200291506946, + 0.0025412696413695812, + 0.0011033185292035341, + 0.0012078207219019532, + 0.0017591826617717743, + 0.001440404448658228, + 0.0009037707350216806, + 0.0009579964098520577, + 0.001005890197120607, + 0.0016884716460481286, + 0.0008255788125097752, + 0.001040355651639402, + 0.0025180510710924864, + 0.001731697004288435, + 0.00100168667268008, + 0.001335094915702939, + 0.0010911058634519577, + 0.010755765251815319, + 0.001333187916316092, + 0.008035566657781601, + 0.005603346507996321, + 0.001758126774802804, + 0.0014191445661708713, + 0.0011795698665082455, + 0.0016615408239886165, + 0.002652157796546817, + 0.0014661286259070039, + 0.0017507747979834676, + 0.040118057280778885, + 0.022769248113036156, + 0.010106800124049187, + 0.019461510702967644, + 0.3114006817340851, + 0.3146127462387085, + 0.016175001859664917, + 0.02412804402410984, + 0.0018421950517222285, + 0.0023700189776718616, + 0.0014774402370676398, + 0.004918768536299467, + 0.005946111865341663, + 0.00633615767583251, + 0.0024495800025761127, + 0.0019980526994913816, + 0.003237310564145446, + 0.0031383251771330833, + 0.0017562337452545762, + 0.0035868112463504076, + 0.02096283994615078, + 0.020169569179415703, + 0.0027164299972355366, + 0.0027140481397509575, + 0.0025014712009578943, + 0.0024286683183163404, + 0.002768489997833967, + 0.0033640945330262184, + 0.0029218527488410473, + 0.003340010764077306, + 0.0018855654634535313, + 0.0019396486459299922, + 0.004314961843192577, + 0.006871514488011599, + 0.004603646695613861, + 0.005552164278924465, + 0.005310270935297012, + 0.005128099583089352, + 0.003215290606021881, + 0.005245693027973175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.006734294816851616 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624453864991665 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624453864991665, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624453864991665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16498439013957977, + 0.0849907249212265, + 0.023550672456622124, + 0.11618942022323608, + 0.09725077450275421, + 0.3410165011882782, + 0.06547529995441437, + 0.029801923781633377, + 0.03881765529513359, + 0.03792262077331543 + ], + "credit_map": [ + 0.022526448592543602, + 0.018303556367754936, + 0.01858590729534626, + 0.017804879695177078, + 0.018410054966807365, + 0.028252342715859413, + 0.024119626730680466, + 0.01698158122599125, + 0.011511269025504589, + 0.008092845790088177, + 0.010280726477503777, + 0.010385106317698956, + 0.008311916142702103, + 0.018142228946089745, + 0.008941863663494587, + 0.009324762970209122, + 0.0036227977834641933, + 0.0025270795449614525, + 0.002314947545528412, + 0.0021674640011042356, + 0.002439808566123247, + 0.005228239577263594, + 0.002230751095339656, + 0.0030195845756679773, + 0.010388164781033993, + 0.007660222239792347, + 0.003936772234737873, + 0.006194971967488527, + 0.005637286230921745, + 0.042465075850486755, + 0.005219859536737204, + 0.03468706086277962, + 0.017614958807826042, + 0.006084367632865906, + 0.007134813349694014, + 0.0068579548969864845, + 0.0066290986724197865, + 0.029246428981423378, + 0.005590126384049654, + 0.018093029037117958, + 0.02008303999900818, + 0.012065736576914787, + 0.00511795561760664, + 0.009602741338312626, + 0.14101527631282806, + 0.1348726898431778, + 0.007791805546730757, + 0.010467268526554108, + 0.005371249746531248, + 0.0036725648678839207, + 0.005130353849381208, + 0.007185711525380611, + 0.017010977491736412, + 0.01615596003830433, + 0.004066545516252518, + 0.0068819355219602585, + 0.004711282905191183, + 0.004162278957664967, + 0.002305193804204464, + 0.0037545287050306797, + 0.004997709300369024, + 0.004816113505512476, + 0.0027361055836081505, + 0.0023187107872217894, + 0.004338317085057497, + 0.007126726675778627, + 0.004862502217292786, + 0.0050420695915818214, + 0.004897558130323887, + 0.005204011686146259, + 0.003996329847723246, + 0.0033501379657536745, + 0.0037854888942092657, + 0.00447107246145606, + 0.006923672743141651, + 0.0038312592078000307, + 0.00498103816062212, + 0.0041677518747746944, + 0.0042591821402311325, + 0.005503155756741762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.007624453864991665 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly descending, gently circling drone shot offers awe-inspiring views of a vast golden lotus throne floating in an abstract void. Constructed not of smooth gold, the throne reveals the texture of gilded metal weathered by time. Its surface is riddled with minute cracks and mottled colors. Simple, powerful lines outline each petal, imbued with the allure of Dunhuang murals. The throne's massive petals, layered one after another, steadily and gracefully unfold with a solemn and leisurely rhythm. As the innermost petals unfold, the compassionate figure of Guanyin, seated at the throne's center, is fully revealed. She clad in a pure, immaculate white silk robe, its folds flowing smoothly and simply. She sits in a serene lotus posture, her face serene and compassionate. Her eyes, slightly downcast, seem to survey the world, a faint, transcendent smile playing upon her lips. The entire scene is bathed in a soft, sacred golden light emanating from the lotus and Guanyin herself. This light casts gentle shadows across the mottled golden surface, highlighting its ancient texture. The backdrop, rather than the real world, features stylized auspicious cloud patterns composed of mineral pigments such as ochre, cinnabar, and malachite green. This perfectly replicates the weathered aesthetic of millennia-old murals, creating an atmosphere of solemnity, tranquility, and sacredness that transcends time and space.", + "seed": 18520, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_12/index_5924_seed_18520_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.155780597217381, + "vlm_logprob": -0.155780597217381, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.155780597217381, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09854939579963684 + ], + "opposite_logprob_tokens": [ + -2.4735493659973145 + ], + "logprob_reward": -0.09854939579963684, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09854939579963684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11331716179847717, + 0.02300644852221012, + 0.048288892954587936, + 0.023458128795027733, + 0.09426601231098175, + 0.07361303269863129, + 0.05939481407403946, + 0.11279170215129852, + 0.0822286307811737, + 0.3696351647377014 + ], + "credit_map": [ + 0.011928142979741096, + 0.0080031156539917, + 0.017184069380164146, + 0.007986905984580517, + 0.008648956194519997, + 0.026527686044573784, + 0.008069201372563839, + 0.024969080463051796, + 0.0022009508684277534, + 0.002779665170237422, + 0.0034787775948643684, + 0.0027314210310578346, + 0.0026298612356185913, + 0.0024130339734256268, + 0.0027529094368219376, + 0.004019830841571093, + 0.00794872734695673, + 0.0035657440312206745, + 0.010512777604162693, + 0.0040926155634224415, + 0.0038055842742323875, + 0.006246011704206467, + 0.003249045927077532, + 0.00886838510632515, + 0.0024998446460813284, + 0.00403461093083024, + 0.002605325309559703, + 0.0035358541645109653, + 0.0038199061527848244, + 0.002726604463532567, + 0.0019765892066061497, + 0.0022593929897993803, + 0.014406401664018631, + 0.011524251662194729, + 0.018760452046990395, + 0.007348848041146994, + 0.009631107561290264, + 0.01209260243922472, + 0.007858079858124256, + 0.012644268572330475, + 0.0130525603890419, + 0.00702752685174346, + 0.007328928913921118, + 0.007504358887672424, + 0.010524150915443897, + 0.014876261353492737, + 0.005655982065945864, + 0.0076432544738054276, + 0.01474379189312458, + 0.00595733430236578, + 0.003763132728636265, + 0.004602077417075634, + 0.008758822456002235, + 0.013377878814935684, + 0.004183382727205753, + 0.0040083942003548145, + 0.013861976563930511, + 0.007612587418407202, + 0.005838405340909958, + 0.005242022220045328, + 0.009142241440713406, + 0.028366927057504654, + 0.007939038798213005, + 0.034788500517606735, + 0.009124393574893475, + 0.013853502459824085, + 0.008403723128139973, + 0.010454664938151836, + 0.011387895792722702, + 0.010883288457989693, + 0.008423087187111378, + 0.009698077104985714, + 0.03867160156369209, + 0.027903858572244644, + 0.061675384640693665, + 0.026620736345648766, + 0.03100406564772129, + 0.08838317543268204, + 0.030674852430820465, + 0.06470148265361786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.09854939579963684 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.632765531539917 + ], + "opposite_logprob_tokens": [ + -0.757765531539917 + ], + "logprob_reward": -0.632765531539917, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.632765531539917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06312012672424316, + 0.025730539113283157, + 0.016264328733086586, + 0.020197447389364243, + 0.02701156958937645, + 0.05148819461464882, + 0.04941428452730179, + 0.4883367121219635, + 0.05496751517057419, + 0.20346921682357788 + ], + "credit_map": [ + 0.006341839209198952, + 0.0055825356394052505, + 0.008056104183197021, + 0.004364872816950083, + 0.005931994877755642, + 0.013915098272264004, + 0.005746958777308464, + 0.013180725276470184, + 0.003700901521369815, + 0.0027270440477877855, + 0.0038161093834787607, + 0.0022578046191483736, + 0.0032877642661333084, + 0.0030062412843108177, + 0.002535368548706174, + 0.004399308934807777, + 0.002252141246572137, + 0.001469723996706307, + 0.0025230967439711094, + 0.0015646638348698616, + 0.0025787607301026583, + 0.002222496084868908, + 0.0019252903293818235, + 0.001728158094920218, + 0.0021532641258090734, + 0.003663803217932582, + 0.002404024824500084, + 0.004003103356808424, + 0.002077486366033554, + 0.0021969855297356844, + 0.001834805472753942, + 0.0018639754271134734, + 0.0032452652230858803, + 0.0035275339614599943, + 0.0038221869617700577, + 0.003398501081392169, + 0.0031892978586256504, + 0.0034287346061319113, + 0.0022472848650068045, + 0.004152767360210419, + 0.008804097771644592, + 0.004837187007069588, + 0.00490413187071681, + 0.004875916987657547, + 0.008246613666415215, + 0.00980352982878685, + 0.004392600152641535, + 0.005624118726700544, + 0.006983322557061911, + 0.0051918732933700085, + 0.007128068711608648, + 0.003096612635999918, + 0.005276887211948633, + 0.0075310952961444855, + 0.00478534959256649, + 0.00942107755690813, + 0.033796194940805435, + 0.017030557617545128, + 0.017061620950698853, + 0.009028040803968906, + 0.01589912734925747, + 0.1836501806974411, + 0.015664106234908104, + 0.19620691239833832, + 0.007147605065256357, + 0.008700210601091385, + 0.007969172671437263, + 0.0038988934829831123, + 0.007594495080411434, + 0.007468304131180048, + 0.00454952334985137, + 0.007639314979314804, + 0.021147413179278374, + 0.015736239030957222, + 0.030853575095534325, + 0.0142618203535676, + 0.0172511525452137, + 0.04860670119524002, + 0.016685642302036285, + 0.038926683366298676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.632765531539917 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980387583374977 + ], + "opposite_logprob_tokens": [ + -3.529803991317749 + ], + "logprob_reward": -0.02980387583374977, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980387583374977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08257776498794556, + 0.04046115651726723, + 0.05064079165458679, + 0.024545198306441307, + 0.050981517881155014, + 0.06424040347337723, + 0.15000168979167938, + 0.20680667459964752, + 0.1201559379696846, + 0.20958879590034485 + ], + "credit_map": [ + 0.004491190891712904, + 0.005013211164623499, + 0.00777074322104454, + 0.0034362943843007088, + 0.0060816118493676186, + 0.0243191160261631, + 0.0038606117013841867, + 0.027604982256889343, + 0.005024307407438755, + 0.005604222882539034, + 0.005259715020656586, + 0.0034076115116477013, + 0.005095839034765959, + 0.004550337325781584, + 0.004374157637357712, + 0.007144965697079897, + 0.007262873463332653, + 0.004512254614382982, + 0.009677731432020664, + 0.004168026149272919, + 0.004850081168115139, + 0.006385859102010727, + 0.0035527432337403297, + 0.01023122575134039, + 0.0019433554261922836, + 0.004059892147779465, + 0.002534759696573019, + 0.004189175087958574, + 0.003981311805546284, + 0.003469826653599739, + 0.00257582264021039, + 0.0017910570604726672, + 0.005138535983860493, + 0.005728636868298054, + 0.010590014979243279, + 0.005507674999535084, + 0.006198881193995476, + 0.007010235916823149, + 0.004177239257842302, + 0.006630301475524902, + 0.014413431286811829, + 0.005994552280753851, + 0.00507754972204566, + 0.0061728088185191154, + 0.006577720399945974, + 0.015523310750722885, + 0.004435533192008734, + 0.006045493297278881, + 0.039789829403162, + 0.013369009830057621, + 0.00927033368498087, + 0.010324095375835896, + 0.018559984862804413, + 0.037513550370931625, + 0.009240725077688694, + 0.01193416491150856, + 0.018585950136184692, + 0.010962149128317833, + 0.011121350340545177, + 0.00915178656578064, + 0.015879828482866287, + 0.06934734433889389, + 0.01154304575175047, + 0.06021523103117943, + 0.012518039904534817, + 0.017469922080636024, + 0.015538698062300682, + 0.012803598307073116, + 0.016448281705379486, + 0.015173321589827538, + 0.010614359751343727, + 0.01958971470594406, + 0.022168977186083794, + 0.015363048762083054, + 0.03639085218310356, + 0.014658122323453426, + 0.017391599714756012, + 0.049295857548713684, + 0.017563721165060997, + 0.036756597459316254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.02980387583374977 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525094848126173 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.00525094848126173, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525094848126173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15916593372821808, + 0.028617126867175102, + 0.018175236880779266, + 0.03615901619195938, + 0.038364872336387634, + 0.062537781894207, + 0.06389036774635315, + 0.11283372342586517, + 0.04578135535120964, + 0.43447455763816833 + ], + "credit_map": [ + 0.006196971051394939, + 0.006525014527142048, + 0.008848840370774269, + 0.005004932638257742, + 0.0040346356108784676, + 0.06204718351364136, + 0.006031472701579332, + 0.0604768842458725, + 0.0031692783813923597, + 0.0033098456915467978, + 0.003581531345844269, + 0.002918230602517724, + 0.005841183941811323, + 0.003477023681625724, + 0.003025023266673088, + 0.003295008558779955, + 0.002869419753551483, + 0.0016614423366263509, + 0.003898992668837309, + 0.0014319795882329345, + 0.002264074282720685, + 0.0024221681524068117, + 0.0014030259335413575, + 0.0022241331171244383, + 0.003639830043539405, + 0.0059622302651405334, + 0.004880725871771574, + 0.005656547844409943, + 0.004622127395123243, + 0.004638256970793009, + 0.0027919826097786427, + 0.003967313561588526, + 0.004232740495353937, + 0.004517391324043274, + 0.007992101833224297, + 0.0032253002282232046, + 0.006149108521640301, + 0.004997401963919401, + 0.003414274426177144, + 0.0038365558721125126, + 0.007595718838274479, + 0.008648023940622807, + 0.006474265828728676, + 0.0075899772346019745, + 0.010090209543704987, + 0.010989204980432987, + 0.0069624134339392185, + 0.0041879648342728615, + 0.02001681923866272, + 0.0036119057331234217, + 0.0034141179639846087, + 0.0024688083212822676, + 0.004573083017021418, + 0.021139392629265785, + 0.0031222787220031023, + 0.0055439667776227, + 0.01235963311046362, + 0.0055539486929774284, + 0.0063225082121789455, + 0.003731131786480546, + 0.007393297273665667, + 0.04108661413192749, + 0.005271368660032749, + 0.031115226447582245, + 0.006874768994748592, + 0.007811666000634432, + 0.006518400274217129, + 0.00280310632660985, + 0.006261960603296757, + 0.0054582469165325165, + 0.0029414244927465916, + 0.007111780345439911, + 0.04869130253791809, + 0.034639060497283936, + 0.0713106021285057, + 0.032051678746938705, + 0.03534826263785362, + 0.10408096760511398, + 0.03524724766612053, + 0.07310543209314346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.00525094848126173 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533234432339668 + ], + "opposite_logprob_tokens": [ + -4.387533187866211 + ], + "logprob_reward": -0.012533234432339668, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533234432339668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13996513187885284, + 0.034614283591508865, + 0.009070845320820808, + 0.015334364026784897, + 0.015272376127541065, + 0.017085479572415352, + 0.04772420972585678, + 0.4122144281864166, + 0.0299687460064888, + 0.2787501811981201 + ], + "credit_map": [ + 0.007241563871502876, + 0.005016649141907692, + 0.010264210402965546, + 0.003951105289161205, + 0.003963783383369446, + 0.050673190504312515, + 0.005782059859484434, + 0.05307256430387497, + 0.004540235735476017, + 0.0029416214674711227, + 0.0070344205014407635, + 0.0030619765166193247, + 0.003467498579993844, + 0.00609438493847847, + 0.0031440346501767635, + 0.004330112133175135, + 0.001610163482837379, + 0.000814705912489444, + 0.0017719822935760021, + 0.0009150349069386721, + 0.0005829893634654582, + 0.0012124202912673354, + 0.0008160059805959463, + 0.0013475428568199277, + 0.0015300428494811058, + 0.0024205974768847227, + 0.002569349016994238, + 0.0025420847814530134, + 0.0017665182240307331, + 0.0018066306365653872, + 0.0012863376177847385, + 0.0014128019101917744, + 0.002446036087349057, + 0.002087400294840336, + 0.002865305170416832, + 0.0011117116082459688, + 0.0015017137629911304, + 0.001850124797783792, + 0.0012846285244449973, + 0.0021254552993923426, + 0.0019342516316100955, + 0.0025329284835606813, + 0.001965394476428628, + 0.001867305370979011, + 0.0031666227150708437, + 0.002347905421629548, + 0.0016065682284533978, + 0.0016645019641146064, + 0.011418122798204422, + 0.003875491674989462, + 0.004171920474618673, + 0.002385030733421445, + 0.004018144682049751, + 0.011665524914860725, + 0.003361230017617345, + 0.006828742101788521, + 0.040449656546115875, + 0.013311418704688549, + 0.017149005085229874, + 0.006851073354482651, + 0.010362486355006695, + 0.14411498606204987, + 0.012194886803627014, + 0.16778089106082916, + 0.004193142522126436, + 0.005134362727403641, + 0.004700355231761932, + 0.002243593567982316, + 0.003509591333568096, + 0.003413924016058445, + 0.0019109216518700123, + 0.004862854722887278, + 0.02995278500020504, + 0.02174334041774273, + 0.04574429243803024, + 0.019911762326955795, + 0.02317344956099987, + 0.06697524338960648, + 0.022813916206359863, + 0.048435378819704056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.012533234432339668 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly descending, gently circling drone shot offers awe-inspiring views of a vast golden lotus throne floating in an abstract void. Constructed not of smooth gold, the throne reveals the texture of gilded metal weathered by time. Its surface is riddled with minute cracks and mottled colors. Simple, powerful lines outline each petal, imbued with the allure of Dunhuang murals. The throne's massive petals, layered one after another, steadily and gracefully unfold with a solemn and leisurely rhythm. As the innermost petals unfold, the compassionate figure of Guanyin, seated at the throne's center, is fully revealed. She clad in a pure, immaculate white silk robe, its folds flowing smoothly and simply. She sits in a serene lotus posture, her face serene and compassionate. Her eyes, slightly downcast, seem to survey the world, a faint, transcendent smile playing upon her lips. The entire scene is bathed in a soft, sacred golden light emanating from the lotus and Guanyin herself. This light casts gentle shadows across the mottled golden surface, highlighting its ancient texture. The backdrop, rather than the real world, features stylized auspicious cloud patterns composed of mineral pigments such as ochre, cinnabar, and malachite green. This perfectly replicates the weathered aesthetic of millennia-old murals, creating an atmosphere of solemnity, tranquility, and sacredness that transcends time and space.", + "seed": 18521, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_12/index_5924_seed_18521_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04880835572257638, + "vlm_logprob": -0.04880835572257638, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.04880835572257638, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05972709134221077 + ], + "opposite_logprob_tokens": [ + -3.059727191925049 + ], + "logprob_reward": -0.05972709134221077, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05972709134221077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11810479313135147, + 0.17783714830875397, + 0.0767190232872963, + 0.07595989853143692, + 0.07058783620595932, + 0.13198736310005188, + 0.11087372899055481, + 0.04492814466357231, + 0.08899913728237152, + 0.10400290042161942 + ], + "credit_map": [ + 0.010874166153371334, + 0.008421150036156178, + 0.012428460642695427, + 0.009543655440211296, + 0.015168564394116402, + 0.021597880870103836, + 0.01915610581636429, + 0.02091480977833271, + 0.00845077820122242, + 0.00986767653375864, + 0.03715820983052254, + 0.016241250559687614, + 0.012583615258336067, + 0.03959118202328682, + 0.030408797785639763, + 0.02353563718497753, + 0.005786567460745573, + 0.006011602934449911, + 0.008523676544427872, + 0.012956737540662289, + 0.007665066514164209, + 0.008103178814053535, + 0.013465890660881996, + 0.014206294901669025, + 0.0072179255075752735, + 0.009716999717056751, + 0.008125693537294865, + 0.015179000794887543, + 0.009023928083479404, + 0.008276421576738358, + 0.00939255952835083, + 0.009027370251715183, + 0.007701749913394451, + 0.007819305174052715, + 0.008766427636146545, + 0.009708856232464314, + 0.007625522557646036, + 0.009632092900574207, + 0.009968954138457775, + 0.009364932775497437, + 0.013467696495354176, + 0.017240485176444054, + 0.020722195506095886, + 0.01641414687037468, + 0.013480626977980137, + 0.01530724298208952, + 0.017393937334418297, + 0.017961038276553154, + 0.017264587804675102, + 0.010572300292551517, + 0.01788809709250927, + 0.010506057180464268, + 0.009433595463633537, + 0.013682963326573372, + 0.012920825742185116, + 0.01860530488193035, + 0.007454971317201853, + 0.004998997785151005, + 0.004986785817891359, + 0.004304006230086088, + 0.005413687322288752, + 0.007025027647614479, + 0.004141503479331732, + 0.006603165064007044, + 0.022670337930321693, + 0.007067227736115456, + 0.010678664781153202, + 0.007764344569295645, + 0.008729731664061546, + 0.014311411418020725, + 0.005871235858649015, + 0.011906185187399387, + 0.013756929896771908, + 0.007261528633534908, + 0.016258804127573967, + 0.008377740159630775, + 0.010036441497504711, + 0.025670427829027176, + 0.00836939923465252, + 0.014271623454988003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.05972709134221077 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280128479003906 + ], + "opposite_logprob_tokens": [ + -2.017801284790039 + ], + "logprob_reward": -0.14280128479003906, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14280128479003906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08162900060415268, + 0.2126801311969757, + 0.18444477021694183, + 0.05819950997829437, + 0.03932558000087738, + 0.14053881168365479, + 0.04753810539841652, + 0.07172786444425583, + 0.07944547384977341, + 0.08447074145078659 + ], + "credit_map": [ + 0.006353951059281826, + 0.007614682894200087, + 0.007968062534928322, + 0.008426010608673096, + 0.009827153757214546, + 0.009730549529194832, + 0.01532116997987032, + 0.016387419775128365, + 0.01680135168135166, + 0.0190721582621336, + 0.01817155070602894, + 0.030751952901482582, + 0.019730310887098312, + 0.032857559621334076, + 0.037853583693504333, + 0.037441689521074295, + 0.016996199265122414, + 0.0143059017136693, + 0.019629759714007378, + 0.029372157528996468, + 0.015642428770661354, + 0.01703346148133278, + 0.03585055470466614, + 0.03561430051922798, + 0.0041886367835104465, + 0.008862143382430077, + 0.006973670795559883, + 0.008482259698212147, + 0.005807187873870134, + 0.0070410799235105515, + 0.008112676441669464, + 0.008731857873499393, + 0.004231883678585291, + 0.004478651098906994, + 0.005230896640568972, + 0.00927544105798006, + 0.0034569355193525553, + 0.003882810240611434, + 0.0043913149274885654, + 0.0043776449747383595, + 0.013656373135745525, + 0.009620826691389084, + 0.01966172643005848, + 0.025550419464707375, + 0.010183650068938732, + 0.0176343135535717, + 0.025761237367987633, + 0.018470263108611107, + 0.007237724494189024, + 0.004902656190097332, + 0.00630567129701376, + 0.006069466937333345, + 0.006343525368720293, + 0.004113021306693554, + 0.006200534291565418, + 0.006365503184497356, + 0.017354512587189674, + 0.005612069275230169, + 0.010053553618490696, + 0.004822570364922285, + 0.006003640126436949, + 0.011514454148709774, + 0.0058937398716807365, + 0.010473323054611683, + 0.020179415121674538, + 0.006107689812779427, + 0.010433643124997616, + 0.0065025584772229195, + 0.00804195273667574, + 0.010135388933122158, + 0.005656445864588022, + 0.012388378381729126, + 0.010259009897708893, + 0.006763041485100985, + 0.012365785427391529, + 0.009325498715043068, + 0.008788851089775562, + 0.01923982799053192, + 0.008546054363250732, + 0.009182671085000038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.14280128479003906 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020586278289556503 + ], + "opposite_logprob_tokens": [ + -3.8955862522125244 + ], + "logprob_reward": -0.020586278289556503, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020586278289556503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08794881403446198, + 0.06818028539419174, + 0.16202059388160706, + 0.13568459451198578, + 0.07265334576368332, + 0.10509441792964935, + 0.10974752902984619, + 0.10955861955881119, + 0.07807295769453049, + 0.07103879749774933 + ], + "credit_map": [ + 0.007406103424727917, + 0.0094229681417346, + 0.007547426503151655, + 0.013604482635855675, + 0.01011549774557352, + 0.010199954733252525, + 0.013947200030088425, + 0.015705175697803497, + 0.006416240707039833, + 0.006105921231210232, + 0.011869978159666061, + 0.011334894225001335, + 0.005273982882499695, + 0.00988049153238535, + 0.009747731499373913, + 0.007551044225692749, + 0.013983633369207382, + 0.021309485659003258, + 0.018345333635807037, + 0.025332124903798103, + 0.0195982176810503, + 0.014312840066850185, + 0.02226894162595272, + 0.026870008558034897, + 0.015611576847732067, + 0.014274816028773785, + 0.016323024407029152, + 0.02341165952384472, + 0.012581996619701385, + 0.01878068968653679, + 0.015517755411565304, + 0.0191830825060606, + 0.005858114920556545, + 0.008038457483053207, + 0.010125823318958282, + 0.01212314609438181, + 0.00699509447440505, + 0.006048574112355709, + 0.012236730195581913, + 0.011227407492697239, + 0.009069644846022129, + 0.011406587436795235, + 0.018763680011034012, + 0.01415847148746252, + 0.010002795606851578, + 0.010542291216552258, + 0.015347151085734367, + 0.015803789719939232, + 0.013908535242080688, + 0.014405445195734501, + 0.013199501670897007, + 0.013410113751888275, + 0.015827273949980736, + 0.012055989354848862, + 0.013272181153297424, + 0.013668481260538101, + 0.020729808136820793, + 0.01195481140166521, + 0.017276190221309662, + 0.008744855411350727, + 0.013730956241488457, + 0.01374730747193098, + 0.008516787551343441, + 0.014857898466289043, + 0.017066916450858116, + 0.007382318843156099, + 0.012890457175672054, + 0.008140906691551208, + 0.008334178477525711, + 0.007569635286927223, + 0.006977943237870932, + 0.009710599668323994, + 0.011537367478013039, + 0.009055805392563343, + 0.009130221791565418, + 0.00912600290030241, + 0.0087271798402071, + 0.008792147971689701, + 0.007551939692348242, + 0.007118132431060076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 6 + ], + "selected_reward": -0.020586278289556503 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193184673786163 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014193184673786163, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193184673786163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09299556910991669, + 0.04545566067099571, + 0.04165670648217201, + 0.06156564876437187, + 0.037843864411115646, + 0.09787127375602722, + 0.11437956243753433, + 0.0786762535572052, + 0.2658604383468628, + 0.16369502246379852 + ], + "credit_map": [ + 0.005991239566355944, + 0.006706205662339926, + 0.007407919969409704, + 0.010467871092259884, + 0.013455546461045742, + 0.011912691406905651, + 0.025065546855330467, + 0.011988550424575806, + 0.0033388296142220497, + 0.004681069869548082, + 0.006421572063118219, + 0.005469174589961767, + 0.005274449475109577, + 0.007663743570446968, + 0.006557089742273092, + 0.006049732677638531, + 0.0032002131920307875, + 0.004381324164569378, + 0.005193457007408142, + 0.00843602791428566, + 0.004368429537862539, + 0.002985546365380287, + 0.004905336536467075, + 0.008186370134353638, + 0.006282011047005653, + 0.01078400481492281, + 0.006427674554288387, + 0.008425923995673656, + 0.006898900028318167, + 0.007975396700203419, + 0.007803310174494982, + 0.006968426052480936, + 0.003084424650296569, + 0.003861929988488555, + 0.005911427550017834, + 0.005320130381733179, + 0.004424162674695253, + 0.0035877544432878494, + 0.0071494085714221, + 0.004504628013819456, + 0.01123830582946539, + 0.009518417529761791, + 0.012460960075259209, + 0.010352722369134426, + 0.009854099713265896, + 0.020428521558642387, + 0.013978889212012291, + 0.010039354674518108, + 0.012994954362511635, + 0.014018659479916096, + 0.016938352957367897, + 0.013937772251665592, + 0.01528643723577261, + 0.011470165103673935, + 0.012788670137524605, + 0.01694454625248909, + 0.012250572443008423, + 0.009808961302042007, + 0.011862063780426979, + 0.0077609531581401825, + 0.011208966374397278, + 0.008839823305606842, + 0.006092204246670008, + 0.010852710343897343, + 0.05828305333852768, + 0.029518214985728264, + 0.0345931276679039, + 0.024834292009472847, + 0.027046244591474533, + 0.03764094412326813, + 0.021151220425963402, + 0.0327933132648468, + 0.021760351955890656, + 0.014325841329991817, + 0.021846560761332512, + 0.025234106928110123, + 0.01572580635547638, + 0.0292202141135931, + 0.01717195473611355, + 0.018410179764032364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.014193184673786163 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067339395172894 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.0067339395172894, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067339395172894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2248048633337021, + 0.03313527628779411, + 0.04838017374277115, + 0.06681233644485474, + 0.03844676539301872, + 0.07546483725309372, + 0.04456836357712746, + 0.17949487268924713, + 0.11752460151910782, + 0.17136788368225098 + ], + "credit_map": [ + 0.015684621408581734, + 0.014189395122230053, + 0.01967521384358406, + 0.021101737394928932, + 0.03496706858277321, + 0.028687572106719017, + 0.05554114654660225, + 0.03495809808373451, + 0.0023604638408869505, + 0.0028784824535250664, + 0.005337976850569248, + 0.004185389261692762, + 0.0036012837663292885, + 0.006689231377094984, + 0.0040023429319262505, + 0.004080102778971195, + 0.004428883083164692, + 0.0051360707730054855, + 0.006696127355098724, + 0.007302037440240383, + 0.005350446328520775, + 0.0040875752456486225, + 0.006893960293382406, + 0.008485072292387486, + 0.00536672119051218, + 0.011521536856889725, + 0.0075109731405973434, + 0.010159486904740334, + 0.0069336071610450745, + 0.0068190195597708225, + 0.009388964623212814, + 0.009112028405070305, + 0.0028301505371928215, + 0.0035681233275681734, + 0.005217057187110186, + 0.008525150828063488, + 0.0036007778253406286, + 0.0031433850526809692, + 0.006554353982210159, + 0.005007768515497446, + 0.010128934867680073, + 0.0077939811162650585, + 0.008150913752615452, + 0.009218846447765827, + 0.007474064361304045, + 0.01445258129388094, + 0.010191892273724079, + 0.00805362407118082, + 0.006664634216576815, + 0.005182870198041201, + 0.007449937518686056, + 0.004330562427639961, + 0.004036352504044771, + 0.005399012938141823, + 0.005827080924063921, + 0.0056779151782393456, + 0.04623419791460037, + 0.015881372615695, + 0.02717999368906021, + 0.009366756305098534, + 0.013501977548003197, + 0.03505297377705574, + 0.010861455462872982, + 0.02141614444553852, + 0.02433643490076065, + 0.010389236733317375, + 0.014579052105545998, + 0.008477184921503067, + 0.010619061067700386, + 0.027337556704878807, + 0.007810161914676428, + 0.013975913636386395, + 0.029829557985067368, + 0.0073906234465539455, + 0.026899810880422592, + 0.008615979924798012, + 0.013612798415124416, + 0.05480482801795006, + 0.007239330559968948, + 0.022974951192736626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.0067339395172894 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly descending, gently circling drone shot offers awe-inspiring views of a vast golden lotus throne floating in an abstract void. Constructed not of smooth gold, the throne reveals the texture of gilded metal weathered by time. Its surface is riddled with minute cracks and mottled colors. Simple, powerful lines outline each petal, imbued with the allure of Dunhuang murals. The throne's massive petals, layered one after another, steadily and gracefully unfold with a solemn and leisurely rhythm. As the innermost petals unfold, the compassionate figure of Guanyin, seated at the throne's center, is fully revealed. She clad in a pure, immaculate white silk robe, its folds flowing smoothly and simply. She sits in a serene lotus posture, her face serene and compassionate. Her eyes, slightly downcast, seem to survey the world, a faint, transcendent smile playing upon her lips. The entire scene is bathed in a soft, sacred golden light emanating from the lotus and Guanyin herself. This light casts gentle shadows across the mottled golden surface, highlighting its ancient texture. The backdrop, rather than the real world, features stylized auspicious cloud patterns composed of mineral pigments such as ochre, cinnabar, and malachite green. This perfectly replicates the weathered aesthetic of millennia-old murals, creating an atmosphere of solemnity, tranquility, and sacredness that transcends time and space.", + "seed": 18522, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_12/index_5924_seed_18522_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03421905525028705, + "vlm_logprob": -0.03421905525028705, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.03421905525028705, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.056329600512981415 + ], + "opposite_logprob_tokens": [ + -3.0563294887542725 + ], + "logprob_reward": -0.056329600512981415, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.056329600512981415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10030191391706467, + 0.5229614973068237, + 0.02400553785264492, + 0.023331638425588608, + 0.03028830699622631, + 0.04435106739401817, + 0.05173235386610031, + 0.06675511598587036, + 0.06815146654844284, + 0.06812117993831635 + ], + "credit_map": [ + 0.02528657391667366, + 0.006911614444106817, + 0.005139237269759178, + 0.017305251210927963, + 0.00702661694958806, + 0.007800032384693623, + 0.02513274922966957, + 0.0056998333893716335, + 0.025486880913376808, + 0.015960583463311195, + 0.0103842131793499, + 0.1590845137834549, + 0.02099742367863655, + 0.01535367127507925, + 0.23711735010147095, + 0.03857678920030594, + 0.0022962044458836317, + 0.003106774529442191, + 0.0029042987152934074, + 0.003858122043311596, + 0.0023429631255567074, + 0.003512286115437746, + 0.0026932123582810163, + 0.003291673492640257, + 0.002986038802191615, + 0.002011287258937955, + 0.002978596370667219, + 0.0030624375212937593, + 0.0021936828270554543, + 0.0036581172607839108, + 0.002585452049970627, + 0.0038560235407203436, + 0.0018269805004820228, + 0.004286830313503742, + 0.006256941240280867, + 0.009712045080959797, + 0.0022581240627914667, + 0.0017962296260520816, + 0.002505313837900758, + 0.0016458388417959213, + 0.00704013230279088, + 0.004460064694285393, + 0.0046578459441661835, + 0.007447965443134308, + 0.004869462456554174, + 0.00582884019240737, + 0.006114654708653688, + 0.003932095132768154, + 0.004739781375974417, + 0.006941881030797958, + 0.0056356689892709255, + 0.008141916245222092, + 0.007481173146516085, + 0.005682394839823246, + 0.007124259136617184, + 0.005985272582620382, + 0.0064330752938985825, + 0.01014464721083641, + 0.007646486628800631, + 0.009113818407058716, + 0.008644386194646358, + 0.00646009249612689, + 0.008624051697552204, + 0.009688552469015121, + 0.0030987674836069345, + 0.005506140645593405, + 0.005775342229753733, + 0.021887315437197685, + 0.0053636664524674416, + 0.004698572214692831, + 0.009400193579494953, + 0.012421461753547192, + 0.006449604872614145, + 0.008237193338572979, + 0.008290168829262257, + 0.008124650456011295, + 0.011973625980317593, + 0.007074994500726461, + 0.007826832123100758, + 0.010144106112420559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.056329600512981415 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031560808420181 + ], + "opposite_logprob_tokens": [ + -2.350315570831299 + ], + "logprob_reward": -0.10031560808420181, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031560808420181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12421844154596329, + 0.12309816479682922, + 0.03625992685556412, + 0.09524551779031754, + 0.16892026364803314, + 0.06767214834690094, + 0.12397214770317078, + 0.046950943768024445, + 0.10936354845762253, + 0.10429885983467102 + ], + "credit_map": [ + 0.015383616089820862, + 0.011435030959546566, + 0.013240937143564224, + 0.01898433454334736, + 0.016706714406609535, + 0.016450852155685425, + 0.02158944681286812, + 0.01042751595377922, + 0.008774296380579472, + 0.009013394825160503, + 0.006438302807509899, + 0.017042944207787514, + 0.011931818909943104, + 0.018077462911605835, + 0.03482985869050026, + 0.01699007861316204, + 0.0049102106131613255, + 0.002804019721224904, + 0.005201790016144514, + 0.006473843939602375, + 0.003145224880427122, + 0.003774140030145645, + 0.004055076278746128, + 0.005895621608942747, + 0.011083334684371948, + 0.011708047240972519, + 0.011656227521598339, + 0.010078458115458488, + 0.01019999198615551, + 0.018179243430495262, + 0.008953030221164227, + 0.013387185521423817, + 0.0091784056276083, + 0.009448372758924961, + 0.05950284004211426, + 0.0619422011077404, + 0.004655830562114716, + 0.005885240156203508, + 0.009864450432360172, + 0.008442932739853859, + 0.007818766869604588, + 0.006284918636083603, + 0.008577200584113598, + 0.009565753862261772, + 0.009134098887443542, + 0.010054431855678558, + 0.009431131184101105, + 0.006805846467614174, + 0.007979314774274826, + 0.017766382545232773, + 0.01597859151661396, + 0.022554414346814156, + 0.017462173476815224, + 0.011381104588508606, + 0.017610033974051476, + 0.01324012503027916, + 0.004487349186092615, + 0.0056291487999260426, + 0.005756710190325975, + 0.008220057003200054, + 0.007139118853956461, + 0.003278617514297366, + 0.007190803997218609, + 0.005249137990176678, + 0.004104828927665949, + 0.009196039289236069, + 0.006956661120057106, + 0.038748305290937424, + 0.009239349514245987, + 0.005496736615896225, + 0.017392456531524658, + 0.018229179084300995, + 0.006516557186841965, + 0.011828357353806496, + 0.014407221227884293, + 0.011226636357605457, + 0.017775120213627815, + 0.008797833696007729, + 0.012708458118140697, + 0.02103867568075657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.10031560808420181 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638749830424786 + ], + "opposite_logprob_tokens": [ + -4.758638858795166 + ], + "logprob_reward": -0.008638749830424786, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638749830424786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09825535863637924, + 0.3090516924858093, + 0.03205928951501846, + 0.024996735155582428, + 0.13921964168548584, + 0.04973273724317551, + 0.05375552177429199, + 0.054608188569545746, + 0.01914089359343052, + 0.21918000280857086 + ], + "credit_map": [ + 0.03315611928701401, + 0.007133635226637125, + 0.006764869671314955, + 0.010241599753499031, + 0.00798986479640007, + 0.012989520095288754, + 0.012452403083443642, + 0.007527329493314028, + 0.014505540952086449, + 0.00990489311516285, + 0.005726536735892296, + 0.10244567692279816, + 0.014436408877372742, + 0.009184068068861961, + 0.13801783323287964, + 0.014830689877271652, + 0.0042577642016112804, + 0.004463640041649342, + 0.0030908763874322176, + 0.005418823100626469, + 0.004217897541821003, + 0.0034237706568092108, + 0.0034205259289592505, + 0.003765987465158105, + 0.003424630733206868, + 0.0027046191971749067, + 0.0027489070780575275, + 0.004291800782084465, + 0.002897171303629875, + 0.0032421580981463194, + 0.002733660163357854, + 0.002953785937279463, + 0.007237717509269714, + 0.011023814789950848, + 0.0436379611492157, + 0.051501888781785965, + 0.004961406346410513, + 0.00669352849945426, + 0.007916825823485851, + 0.006246483419090509, + 0.005133860744535923, + 0.006939215119928122, + 0.0052102417685091496, + 0.009680213406682014, + 0.007352815940976143, + 0.005348046775907278, + 0.006276203319430351, + 0.0037921362090855837, + 0.0049234721809625626, + 0.00792224332690239, + 0.00716327503323555, + 0.00901313778012991, + 0.007623027078807354, + 0.006631817668676376, + 0.00665457034483552, + 0.0038239751011133194, + 0.005086110904812813, + 0.008429007604718208, + 0.0070299915969371796, + 0.008493762463331223, + 0.007251938804984093, + 0.003934556618332863, + 0.006498902570456266, + 0.007883910089731216, + 0.0013921771897003055, + 0.001800520927645266, + 0.0018910549115389585, + 0.0028611335437744856, + 0.0031266442965716124, + 0.003058204660192132, + 0.0018733376637101173, + 0.0031378178391605616, + 0.025884263217449188, + 0.022942284122109413, + 0.03139114752411842, + 0.02983826957643032, + 0.0305417999625206, + 0.03056901879608631, + 0.02518458664417267, + 0.02282859943807125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 4 + ], + "selected_reward": -0.008638749830424786 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615511115640402 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.003615511115640402, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615511115640402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05541815608739853, + 0.25848403573036194, + 0.026043962687253952, + 0.06436453014612198, + 0.3201845586299896, + 0.02007349021732807, + 0.02323070541024208, + 0.02228904329240322, + 0.051592789590358734, + 0.1583186835050583 + ], + "credit_map": [ + 0.007435846142470837, + 0.004357350990176201, + 0.0025571780279278755, + 0.011660546064376831, + 0.0041954535990953445, + 0.004951975308358669, + 0.016272827982902527, + 0.003986983094364405, + 0.011086419224739075, + 0.008286692202091217, + 0.006214930675923824, + 0.0898618996143341, + 0.010199862532317638, + 0.007509770803153515, + 0.1133703961968422, + 0.011954083107411861, + 0.0025844634510576725, + 0.004546125419437885, + 0.002609826158732176, + 0.0031689642928540707, + 0.003733509685844183, + 0.003790630027651787, + 0.0030188101809471846, + 0.0025916381273418665, + 0.006669227499514818, + 0.005465263035148382, + 0.00748187443241477, + 0.007180546876043081, + 0.006039661820977926, + 0.014372912235558033, + 0.006419789511710405, + 0.01073526218533516, + 0.01581980474293232, + 0.03129328787326813, + 0.09793245047330856, + 0.12593483924865723, + 0.008274901658296585, + 0.008930264972150326, + 0.019048959016799927, + 0.012950063683092594, + 0.0029757542070001364, + 0.002652605064213276, + 0.002167177153751254, + 0.0037126310635358095, + 0.003015045542269945, + 0.00216731452383101, + 0.0019965830724686384, + 0.001386381802149117, + 0.0019250926561653614, + 0.004348811227828264, + 0.003087036544457078, + 0.003475810866802931, + 0.003554742783308029, + 0.001991072902455926, + 0.0028240757528692484, + 0.002024063840508461, + 0.002245598239824176, + 0.0023973151110112667, + 0.0022487761452794075, + 0.002996021881699562, + 0.0042493101209402084, + 0.0016207185108214617, + 0.004500833339989185, + 0.002030470408499241, + 0.002240898786112666, + 0.004433661233633757, + 0.0035381459165364504, + 0.01892523653805256, + 0.0039050860796123743, + 0.00441109761595726, + 0.007591653149574995, + 0.006547016557306051, + 0.019937142729759216, + 0.015333762392401695, + 0.023985717445611954, + 0.02164076268672943, + 0.01820344850420952, + 0.022422242909669876, + 0.01900821179151535, + 0.01778741180896759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.003615511115640402 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slowly descending, gently circling drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast golden lotus throne with a weathered, gilded metal texture featuring minute cracks and mottled colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the throne's massive petals unfold steadily and gracefully in a layered sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure of Guanyin seated in a serene lotus posture at the throne's center, wearing a pure white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden light with a backdrop of stylized auspicious cloud patterns in ochre, cinnabar, and malachite green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002195806708186865 + ], + "opposite_logprob_tokens": [ + -6.127195835113525 + ], + "logprob_reward": -0.002195806708186865, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002195806708186865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18027567863464355, + 0.23536545038223267, + 0.02548280358314514, + 0.04408537223935127, + 0.07366245985031128, + 0.03781607374548912, + 0.056081030517816544, + 0.07319603115320206, + 0.026831360533833504, + 0.24720367789268494 + ], + "credit_map": [ + 0.027776263654232025, + 0.017050106078386307, + 0.008424143306910992, + 0.03358989208936691, + 0.015570444986224174, + 0.02095642313361168, + 0.04314917325973511, + 0.013759247958660126, + 0.010979585349559784, + 0.008262621238827705, + 0.0052620344795286655, + 0.07987043261528015, + 0.008626291528344154, + 0.006905043497681618, + 0.10392012447118759, + 0.011539329774677753, + 0.0032799532637000084, + 0.004315277095884085, + 0.0028201264794915915, + 0.002598137129098177, + 0.0037683467380702496, + 0.002992786932736635, + 0.0026834183372557163, + 0.0030247587710618973, + 0.004488389007747173, + 0.0032226289622485638, + 0.005128263030201197, + 0.004969259724020958, + 0.004542514681816101, + 0.00967445969581604, + 0.005189386662095785, + 0.006870473735034466, + 0.005319081246852875, + 0.005316910799592733, + 0.02135685458779335, + 0.023051081225275993, + 0.0040391902439296246, + 0.00533962482586503, + 0.004583373200148344, + 0.00465634698048234, + 0.005288813728839159, + 0.004464824218302965, + 0.003774851094931364, + 0.0064232237637043, + 0.0036285051610320807, + 0.00537105230614543, + 0.004398665390908718, + 0.004466136917471886, + 0.0046958280727267265, + 0.006061688531190157, + 0.00805822852998972, + 0.007219519466161728, + 0.005277600139379501, + 0.009087208658456802, + 0.008934062905609608, + 0.006746901199221611, + 0.00692009786143899, + 0.010598236694931984, + 0.009935358539223671, + 0.010196869261562824, + 0.009524264372885227, + 0.006736484356224537, + 0.007801084313541651, + 0.011483638547360897, + 0.0021573954727500677, + 0.002399769378826022, + 0.0028311656787991524, + 0.004484814591705799, + 0.003426242619752884, + 0.004402387421578169, + 0.004107804037630558, + 0.003021782962605357, + 0.02378254570066929, + 0.025602905079722404, + 0.030751962214708328, + 0.0343555212020874, + 0.035427648574113846, + 0.03201412782073021, + 0.03021325170993805, + 0.0350557379424572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.002195806708186865 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_13/index_tensor([1544, 1544])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_13/index_tensor([1544, 1544])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..530049a2cf1dd773469062c9c6f758c39a14fe6f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_13/index_tensor([1544, 1544])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a striking wide-angle shot, then the camera zooms in and out at breakneck speed, simulating the dynamic contraction of all visual elements toward the center of the frame. Within this fractured, surreal dimensional space, countless floating geometric fragments, twisted urban debris, and nebulous streaks of light, initially suspended in a pitch-black void, are now violently pulled and deformed by an invisible, immense gravitational force, transforming into dazzling streams of light that converge relentlessly toward the young man at the absolute center of the frame. The young man, clad in a tight, matte, dark blue combat suit with futuristic lines, leans back slightly, arms outstretched, his short, jet-black hair flowing in an invisible current. His eyes are closed, his pale face betraying the pain and fortitude of enduring the immense energy impact. Finally, all matter and energy collapsed violently within a ten-thousandth of a second, condensing into a hexagonal crystal embedded under the skin of his chest. The crystal was beating rhythmically, steadily and powerfully, and each beat emitted a soft and sacred blue glow. This light penetrated his combat uniform, not only illuminating his resolute face, but also providing the only source of light for this dark space that had just undergone a devastating reconstruction, creating an epic atmosphere of sacredness, rebirth and full of unknown power.", + "seed": 3862, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_13/index_1544_seed_3862_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09750743680633604, + "vlm_logprob": -0.09750743680633604, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.09750743680633604, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22810596227645874 + ], + "opposite_logprob_tokens": [ + -1.6031060218811035 + ], + "logprob_reward": -0.22810596227645874, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22810596227645874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2716447710990906, + 0.04536543786525726, + 0.24878524243831635, + 0.04518762230873108, + 0.03419192135334015, + 0.08290746062994003, + 0.12804420292377472, + 0.04224972799420357, + 0.0672331377863884, + 0.03439056873321533 + ], + "credit_map": [ + 0.020828111097216606, + 0.027225561439990997, + 0.027604127302765846, + 0.0310105811804533, + 0.033456951379776, + 0.045047856867313385, + 0.04407825693488121, + 0.04239331930875778, + 0.00555008789524436, + 0.006530421320348978, + 0.003941827453672886, + 0.0048083155415952206, + 0.00706464983522892, + 0.007262055762112141, + 0.004888187162578106, + 0.005319888237863779, + 0.010278420522809029, + 0.007820934057235718, + 0.00688833836466074, + 0.005171129945665598, + 0.10206066071987152, + 0.10347770899534225, + 0.007325444370508194, + 0.005762581247836351, + 0.004941798746585846, + 0.0057197860442101955, + 0.006613847799599171, + 0.006147892214357853, + 0.005922621116042137, + 0.006462214980274439, + 0.00427643908187747, + 0.005103019531816244, + 0.003740427317097783, + 0.0038128229789435863, + 0.0069949631579220295, + 0.0036135653499513865, + 0.003577290801331401, + 0.005325310863554478, + 0.003615420311689377, + 0.0035121175460517406, + 0.007593116723001003, + 0.0095340171828866, + 0.006760541815310717, + 0.013370214961469173, + 0.012729763053357601, + 0.0071103074587881565, + 0.012100791558623314, + 0.013708705082535744, + 0.00514059467241168, + 0.0065222205594182014, + 0.004894591402262449, + 0.0040853628888726234, + 0.05207967013120651, + 0.0061011845245957375, + 0.03985286131501198, + 0.009367704391479492, + 0.0036885710433125496, + 0.005161256063729525, + 0.00684477761387825, + 0.003921738360077143, + 0.0034837329294532537, + 0.006896370556205511, + 0.00792281236499548, + 0.004330463241785765, + 0.006774908863008022, + 0.012839682400226593, + 0.005054236389696598, + 0.005802879575639963, + 0.0054153152741491795, + 0.009239526465535164, + 0.013310683891177177, + 0.00879590306431055, + 0.004049076233059168, + 0.003931698389351368, + 0.003891672007739544, + 0.0032436002511531115, + 0.004204874392598867, + 0.005569275468587875, + 0.0036199004389345646, + 0.00588047131896019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.22810596227645874 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.142761692404747 + ], + "opposite_logprob_tokens": [ + -2.017761707305908 + ], + "logprob_reward": -0.142761692404747, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.142761692404747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19061197340488434, + 0.07748961448669434, + 0.22250650823116302, + 0.08440420031547546, + 0.0648864358663559, + 0.11668922007083893, + 0.08220192790031433, + 0.07092206925153732, + 0.04221145436167717, + 0.04807654395699501 + ], + "credit_map": [ + 0.01474432647228241, + 0.018650712445378304, + 0.019577285274863243, + 0.022838929668068886, + 0.02343600243330002, + 0.03097120299935341, + 0.03792237117886543, + 0.022471146658062935, + 0.013847295194864273, + 0.008020860143005848, + 0.005671525374054909, + 0.007362420205026865, + 0.015436426736414433, + 0.01312453392893076, + 0.007046760991215706, + 0.006979787722229958, + 0.01089630275964737, + 0.009165388531982899, + 0.009760716930031776, + 0.010942192748188972, + 0.08154860138893127, + 0.08314456790685654, + 0.009952444583177567, + 0.007096282206475735, + 0.010049644857645035, + 0.011984854936599731, + 0.009277086704969406, + 0.008490263484418392, + 0.00877144280821085, + 0.013803903013467789, + 0.011974419467151165, + 0.010052578523755074, + 0.00998328160494566, + 0.007807046640664339, + 0.008648542687296867, + 0.00777336023747921, + 0.007963634096086025, + 0.0063094752840697765, + 0.008401716127991676, + 0.007999386638402939, + 0.010081158019602299, + 0.011052923277020454, + 0.01152877788990736, + 0.02299037203192711, + 0.01771167293190956, + 0.009893183596432209, + 0.01908169500529766, + 0.014349440112709999, + 0.006793578155338764, + 0.01078861951828003, + 0.007662385236471891, + 0.008473041467368603, + 0.01898014172911644, + 0.005997965112328529, + 0.01679925061762333, + 0.006706943269819021, + 0.007410992868244648, + 0.008188999257981777, + 0.007829657755792141, + 0.005717872176319361, + 0.008111583068966866, + 0.010599746368825436, + 0.015894122421741486, + 0.007169090677052736, + 0.004125941079109907, + 0.012169538997113705, + 0.003220112295821309, + 0.003906345926225185, + 0.004782824777066708, + 0.005122118163853884, + 0.004955813754349947, + 0.003928758203983307, + 0.005059202201664448, + 0.005824708845466375, + 0.005566577892750502, + 0.00681839557364583, + 0.0044162035919725895, + 0.007359177805483341, + 0.006131910718977451, + 0.006900368724018335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.142761692404747 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036145609337836504 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.0036145609337836504, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036145609337836504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08758176118135452, + 0.12822261452674866, + 0.2591521143913269, + 0.07512100785970688, + 0.09029654413461685, + 0.0851445123553276, + 0.10192941129207611, + 0.06379146873950958, + 0.04140704125165939, + 0.0673534944653511 + ], + "credit_map": [ + 0.0070770313031971455, + 0.010557024739682674, + 0.007049211300909519, + 0.008767233230173588, + 0.0106978639960289, + 0.015435507521033287, + 0.014875487424433231, + 0.013122408650815487, + 0.024370137602090836, + 0.015561805106699467, + 0.011785292066633701, + 0.016675114631652832, + 0.01874319650232792, + 0.020598778501152992, + 0.010884496383368969, + 0.00960379745811224, + 0.015054325573146343, + 0.01043704804033041, + 0.010073842480778694, + 0.0113600492477417, + 0.09574685245752335, + 0.0951458290219307, + 0.012089835479855537, + 0.00924434419721365, + 0.01040932908654213, + 0.010609607212245464, + 0.01143121998757124, + 0.009927592240273952, + 0.00870496779680252, + 0.010768149048089981, + 0.007615959271788597, + 0.005654179025441408, + 0.01583249680697918, + 0.01043891441076994, + 0.014085125178098679, + 0.011063741520047188, + 0.010355915874242783, + 0.00918454211205244, + 0.006960914470255375, + 0.012374897487461567, + 0.00731251435354352, + 0.00819956324994564, + 0.008018532767891884, + 0.014581648632884026, + 0.014065686613321304, + 0.007710214238613844, + 0.013282816857099533, + 0.011973533779382706, + 0.005978547502309084, + 0.009067106992006302, + 0.007693329360336065, + 0.007581688463687897, + 0.03290196880698204, + 0.0075872913002967834, + 0.02180836722254753, + 0.009311112575232983, + 0.008595935069024563, + 0.01089153066277504, + 0.007107036653906107, + 0.005293393507599831, + 0.006233999505639076, + 0.007187436800450087, + 0.010318580083549023, + 0.00816356111317873, + 0.005267329514026642, + 0.009420922957360744, + 0.0031558985356241465, + 0.0031286217272281647, + 0.0045250109396874905, + 0.00499862851575017, + 0.005408453289419413, + 0.005502175074070692, + 0.010349276475608349, + 0.006842863280326128, + 0.007950976490974426, + 0.006219212431460619, + 0.006168339867144823, + 0.012436507269740105, + 0.006190444342792034, + 0.011195873841643333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.0036145609337836504 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043054379522800446 + ], + "opposite_logprob_tokens": [ + -3.1680543422698975 + ], + "logprob_reward": -0.043054379522800446, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043054379522800446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09020865708589554, + 0.06503546237945557, + 0.07771462947130203, + 0.0766347125172615, + 0.11092408001422882, + 0.125181183218956, + 0.23954834043979645, + 0.12497945874929428, + 0.04606010019779205, + 0.0437132902443409 + ], + "credit_map": [ + 0.008587971329689026, + 0.011449036188423634, + 0.010507727041840553, + 0.008845534175634384, + 0.012827010825276375, + 0.013517312705516815, + 0.013850496150553226, + 0.010623574256896973, + 0.008718309924006462, + 0.006545699667185545, + 0.004794469568878412, + 0.006552014499902725, + 0.013516763225197792, + 0.011380203068256378, + 0.006534880958497524, + 0.00699312798678875, + 0.009673980996012688, + 0.005298160947859287, + 0.005696007050573826, + 0.007757868617773056, + 0.017010565847158432, + 0.01980477198958397, + 0.007520527578890324, + 0.004952747840434313, + 0.008041714318096638, + 0.01042622048407793, + 0.009273359552025795, + 0.007655161898583174, + 0.012234872207045555, + 0.01328296959400177, + 0.0094017768278718, + 0.006318647880107164, + 0.01488052774220705, + 0.011480312794446945, + 0.014925255440175533, + 0.013947546482086182, + 0.010664304718375206, + 0.013398146256804466, + 0.011403944343328476, + 0.020224053412675858, + 0.011898713186383247, + 0.015754997730255127, + 0.012039611116051674, + 0.01611250452697277, + 0.019484547898173332, + 0.013640539720654488, + 0.018131814897060394, + 0.01811845973134041, + 0.012850845232605934, + 0.01678384095430374, + 0.012280721217393875, + 0.01003317628055811, + 0.09199246764183044, + 0.012030375190079212, + 0.061629995703697205, + 0.021946949884295464, + 0.016235211864113808, + 0.03185484930872917, + 0.012028861790895462, + 0.01342596672475338, + 0.01107768528163433, + 0.011522308923304081, + 0.01500607281923294, + 0.013828524388372898, + 0.004299411084502935, + 0.007683061994612217, + 0.004563329741358757, + 0.005433040205389261, + 0.004588946234434843, + 0.007596881128847599, + 0.007086613215506077, + 0.004808821715414524, + 0.0054414644837379456, + 0.006451795808970928, + 0.004386116750538349, + 0.006603943184018135, + 0.00492863729596138, + 0.006908946670591831, + 0.0035702234599739313, + 0.005422169342637062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.043054379522800446 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000058889389038 + ], + "opposite_logprob_tokens": [ + -2.695000648498535 + ], + "logprob_reward": -0.07000058889389038, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000058889389038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0918789952993393, + 0.029664497822523117, + 0.1826125532388687, + 0.030098330229520798, + 0.055345188826322556, + 0.21011023223400116, + 0.2841338813304901, + 0.025117836892604828, + 0.04907948523759842, + 0.04195898398756981 + ], + "credit_map": [ + 0.008984335698187351, + 0.010405556298792362, + 0.01110624335706234, + 0.010788783431053162, + 0.010502696968615055, + 0.012532012537121773, + 0.01774020306766033, + 0.00981915183365345, + 0.0025809991639107466, + 0.0023212258238345385, + 0.002525338437408209, + 0.003435588674619794, + 0.0032592767383903265, + 0.00462811766192317, + 0.0069232177920639515, + 0.003990734927356243, + 0.008273827843368053, + 0.007371320389211178, + 0.005623386707156897, + 0.006648496724665165, + 0.07115830481052399, + 0.07146627455949783, + 0.007063888479024172, + 0.005007036961615086, + 0.004192288499325514, + 0.0030614552088081837, + 0.004097203724086285, + 0.0041001904755830765, + 0.003202952677384019, + 0.00463834498077631, + 0.0033466063905507326, + 0.0034592878073453903, + 0.009503166191279888, + 0.009293668903410435, + 0.007587209809571505, + 0.007060435600578785, + 0.005534425377845764, + 0.0058970823884010315, + 0.004581323359161615, + 0.005887875333428383, + 0.01884038746356964, + 0.023045668378472328, + 0.018466202542185783, + 0.030010182410478592, + 0.027860574424266815, + 0.018489299342036247, + 0.030960867181420326, + 0.042437050491571426, + 0.013050134293735027, + 0.016796937212347984, + 0.011197800748050213, + 0.009653852321207523, + 0.11184150725603104, + 0.01616678014397621, + 0.0838141068816185, + 0.02161277085542679, + 0.0023898889776319265, + 0.0038620620034635067, + 0.0031374371610581875, + 0.0026534060016274452, + 0.002911030314862728, + 0.0031922373455017805, + 0.003476980607956648, + 0.0034947930835187435, + 0.005813583265990019, + 0.00976630300283432, + 0.0036414468195289373, + 0.0044967057183384895, + 0.0036118391435593367, + 0.0060501703992486, + 0.010735049843788147, + 0.004964387975633144, + 0.006624981760978699, + 0.007574708666652441, + 0.004490137565881014, + 0.003959171008318663, + 0.005185740068554878, + 0.004174999427050352, + 0.00316980155184865, + 0.006779441609978676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.07000058889389038 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a striking wide-angle shot, then the camera zooms in and out at breakneck speed, simulating the dynamic contraction of all visual elements toward the center of the frame. Within this fractured, surreal dimensional space, countless floating geometric fragments, twisted urban debris, and nebulous streaks of light, initially suspended in a pitch-black void, are now violently pulled and deformed by an invisible, immense gravitational force, transforming into dazzling streams of light that converge relentlessly toward the young man at the absolute center of the frame. The young man, clad in a tight, matte, dark blue combat suit with futuristic lines, leans back slightly, arms outstretched, his short, jet-black hair flowing in an invisible current. His eyes are closed, his pale face betraying the pain and fortitude of enduring the immense energy impact. Finally, all matter and energy collapsed violently within a ten-thousandth of a second, condensing into a hexagonal crystal embedded under the skin of his chest. The crystal was beating rhythmically, steadily and powerfully, and each beat emitted a soft and sacred blue glow. This light penetrated his combat uniform, not only illuminating his resolute face, but also providing the only source of light for this dark space that had just undergone a devastating reconstruction, creating an epic atmosphere of sacredness, rebirth and full of unknown power.", + "seed": 3863, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_13/index_1544_seed_3863_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12208924274891615, + "vlm_logprob": -0.12208924274891615, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.12208924274891615, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4314940869808197 + ], + "opposite_logprob_tokens": [ + -1.056494116783142 + ], + "logprob_reward": -0.4314940869808197, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4314940869808197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2327680140733719, + 0.18038815259933472, + 0.18321125209331512, + 0.09360171109437943, + 0.12152741104364395, + 0.06294510513544083, + 0.023923300206661224, + 0.05079803243279457, + 0.023055152967572212, + 0.02778179757297039 + ], + "credit_map": [ + 0.02453492023050785, + 0.0335402749478817, + 0.03842279314994812, + 0.034586917608976364, + 0.026556789875030518, + 0.02844098024070263, + 0.022862862795591354, + 0.023822490125894547, + 0.024189073592424393, + 0.020195327699184418, + 0.01828092522919178, + 0.017001375555992126, + 0.019097674638032913, + 0.024385446682572365, + 0.022319847717881203, + 0.03491848334670067, + 0.02087448164820671, + 0.011185011826455593, + 0.016933482140302658, + 0.008275660686194897, + 0.011531984433531761, + 0.05198908969759941, + 0.00928808655589819, + 0.05313346907496452, + 0.008103431202471256, + 0.0175916850566864, + 0.013853337615728378, + 0.009956604801118374, + 0.013395431451499462, + 0.010692593641579151, + 0.00996407214552164, + 0.010044565424323082, + 0.014679995365440845, + 0.012205688282847404, + 0.031949419528245926, + 0.009544719941914082, + 0.018652701750397682, + 0.014088553376495838, + 0.011164899915456772, + 0.009241433814167976, + 0.009485793299973011, + 0.011145571246743202, + 0.007684372831135988, + 0.007957961410284042, + 0.008642797358334064, + 0.0075587439350783825, + 0.005522914230823517, + 0.004946954548358917, + 0.004339188802987337, + 0.004026367329061031, + 0.0022818895522505045, + 0.002076714066788554, + 0.004620712250471115, + 0.0024666637182235718, + 0.002087924163788557, + 0.002023841952905059, + 0.005972350016236305, + 0.008109253831207752, + 0.005227111745625734, + 0.005476096179336309, + 0.007257848046720028, + 0.008444058708846569, + 0.005537822842597961, + 0.004773496650159359, + 0.0033721388317644596, + 0.00457513565197587, + 0.0032426698599010706, + 0.0017830064753070474, + 0.0033381504472345114, + 0.0026266812346875668, + 0.0020651998929679394, + 0.0020521720871329308, + 0.004483961965888739, + 0.004188593942672014, + 0.003755338257178664, + 0.0019151446176692843, + 0.0022654677741229534, + 0.0043739923276007175, + 0.0024646336678415537, + 0.004334666766226292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.4314940869808197 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000014185905457 + ], + "opposite_logprob_tokens": [ + -2.695000171661377 + ], + "logprob_reward": -0.07000014185905457, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000014185905457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.144540935754776, + 0.08343987166881561, + 0.10497507452964783, + 0.09766069799661636, + 0.1236419826745987, + 0.04372033476829529, + 0.0693361833691597, + 0.11750608682632446, + 0.04360516741871834, + 0.1715736836194992 + ], + "credit_map": [ + 0.009950011968612671, + 0.0167709831148386, + 0.019670473411679268, + 0.019466876983642578, + 0.013434803113341331, + 0.015229620039463043, + 0.016273731365799904, + 0.03374442458152771, + 0.009706363081932068, + 0.00809984840452671, + 0.009630745276808739, + 0.006190381478518248, + 0.00987189169973135, + 0.01285814493894577, + 0.010049557313323021, + 0.017032932490110397, + 0.015445283614099026, + 0.009931170381605625, + 0.01097202394157648, + 0.008440252393484116, + 0.00992318894714117, + 0.024717025458812714, + 0.008839559741318226, + 0.016706567257642746, + 0.010324884206056595, + 0.016896650195121765, + 0.01378719974309206, + 0.012213609181344509, + 0.012008395977318287, + 0.009394221007823944, + 0.013985533267259598, + 0.009050185792148113, + 0.018336711451411247, + 0.015202745795249939, + 0.040183182805776596, + 0.011883929371833801, + 0.012085366062819958, + 0.009588710963726044, + 0.007047948427498341, + 0.009313377551734447, + 0.006133546587079763, + 0.007082534022629261, + 0.004467384424060583, + 0.006090510170906782, + 0.006335666868835688, + 0.005676461849361658, + 0.004133777227252722, + 0.003800449427217245, + 0.009966380894184113, + 0.011824402958154678, + 0.009335123002529144, + 0.006123999133706093, + 0.010259807109832764, + 0.00572877936065197, + 0.007160922046750784, + 0.008936765603721142, + 0.011682078242301941, + 0.014416687190532684, + 0.01474979892373085, + 0.014062171801924706, + 0.016013644635677338, + 0.020453210920095444, + 0.012722013518214226, + 0.013406471349298954, + 0.006368654780089855, + 0.007159218657761812, + 0.005167813505977392, + 0.00368049880489707, + 0.005911423824727535, + 0.006010272074490786, + 0.004854630213230848, + 0.004452653229236603, + 0.02727426402270794, + 0.028193505480885506, + 0.023111101239919662, + 0.010641987435519695, + 0.014357642270624638, + 0.026440635323524475, + 0.018807411193847656, + 0.02274712175130844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.07000014185905457 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252371542155743 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252371542155743, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252371542155743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1889401525259018, + 0.07951200008392334, + 0.06365707516670227, + 0.09269454330205917, + 0.154864102602005, + 0.03986033797264099, + 0.085130475461483, + 0.09404055774211884, + 0.062053341418504715, + 0.1392473727464676 + ], + "credit_map": [ + 0.015531739220023155, + 0.028285376727581024, + 0.02391928993165493, + 0.028734220191836357, + 0.019665932282805443, + 0.0192585326731205, + 0.019133316352963448, + 0.034411750733852386, + 0.010835785418748856, + 0.009620008990168571, + 0.009127889759838581, + 0.008072597905993462, + 0.010649380274116993, + 0.008457967080175877, + 0.008919532410800457, + 0.013828840106725693, + 0.0061899409629404545, + 0.009463903494179249, + 0.010688447393476963, + 0.009061835706233978, + 0.006731527391821146, + 0.007013313937932253, + 0.007956690154969692, + 0.006551419850438833, + 0.008295705541968346, + 0.01752788946032524, + 0.014300903305411339, + 0.010435700416564941, + 0.008903720416128635, + 0.013731660321354866, + 0.007799102459102869, + 0.011699860915541649, + 0.02769649401307106, + 0.027538437396287918, + 0.04432680457830429, + 0.0119799068197608, + 0.013379188254475594, + 0.011776390485465527, + 0.009627983905375004, + 0.008538886904716492, + 0.0055711655877530575, + 0.005285848863422871, + 0.004906950518488884, + 0.00437931576743722, + 0.005197100806981325, + 0.005386779084801674, + 0.0025150077417492867, + 0.006618166342377663, + 0.014361686073243618, + 0.015145515091717243, + 0.013657313771545887, + 0.00895881000906229, + 0.011861010454595089, + 0.007804659195244312, + 0.005273148417472839, + 0.00806833803653717, + 0.010050286538898945, + 0.020982233807444572, + 0.011461485177278519, + 0.0102155189961195, + 0.014061294496059418, + 0.009894546121358871, + 0.00765153206884861, + 0.0097236642614007, + 0.008660669438540936, + 0.009737412445247173, + 0.012700439430773258, + 0.006259200628846884, + 0.00741186086088419, + 0.00798281840980053, + 0.0040571424178779125, + 0.005243795458227396, + 0.023917876183986664, + 0.030883068218827248, + 0.01525186374783516, + 0.008650646544992924, + 0.014200408942997456, + 0.019763702526688576, + 0.011908246204257011, + 0.01467155758291483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.005252371542155743 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369591012597084 + ], + "opposite_logprob_tokens": [ + -3.408695936203003 + ], + "logprob_reward": -0.03369591012597084, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369591012597084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11352132260799408, + 0.025914179161190987, + 0.057477809488773346, + 0.064100481569767, + 0.03870820626616478, + 0.04213476553559303, + 0.11966325342655182, + 0.20800691843032837, + 0.026816820725798607, + 0.30365628004074097 + ], + "credit_map": [ + 0.012186062522232533, + 0.01800020970404148, + 0.018002431839704514, + 0.018093246966600418, + 0.010553710162639618, + 0.011442771181464195, + 0.011670266278088093, + 0.01357261836528778, + 0.003688720054924488, + 0.0031532461289316416, + 0.0031850726809352636, + 0.0024959035217761993, + 0.002660354832187295, + 0.0036846213042736053, + 0.0029777269810438156, + 0.004068530164659023, + 0.0067917583510279655, + 0.004516254179179668, + 0.007134169805794954, + 0.003918161150068045, + 0.0049344454891979694, + 0.013232652097940445, + 0.004044584464281797, + 0.012905778363347054, + 0.006322719622403383, + 0.012346685864031315, + 0.008039789274334908, + 0.008089537732303143, + 0.006929055321961641, + 0.007059193681925535, + 0.006844291463494301, + 0.008469205349683762, + 0.004890067037194967, + 0.008310487493872643, + 0.007184866350144148, + 0.0038834991864860058, + 0.004331056959927082, + 0.0028837628196924925, + 0.004094826988875866, + 0.0031296361703425646, + 0.005810791626572609, + 0.006817796267569065, + 0.006488367915153503, + 0.0043174088932573795, + 0.005943413358181715, + 0.005757056642323732, + 0.0024463238660246134, + 0.004553601611405611, + 0.016388509422540665, + 0.01673135720193386, + 0.026137441396713257, + 0.009684327989816666, + 0.0175172071903944, + 0.012854876928031445, + 0.010005260817706585, + 0.01034427061676979, + 0.021947884932160378, + 0.0185305867344141, + 0.028795389458537102, + 0.02172466367483139, + 0.02809991128742695, + 0.041005030274391174, + 0.02243781089782715, + 0.02546563372015953, + 0.003800326958298683, + 0.003940373659133911, + 0.005016495008021593, + 0.003283928846940398, + 0.003136074636131525, + 0.003100197995081544, + 0.0024117608554661274, + 0.0021276576444506645, + 0.04777436703443527, + 0.05968492478132248, + 0.03761468082666397, + 0.02052823267877102, + 0.02861747331917286, + 0.04303980991244316, + 0.032692328095436096, + 0.03370445594191551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.03369591012597084 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700037032365799 + ], + "opposite_logprob_tokens": [ + -2.6950037479400635 + ], + "logprob_reward": -0.0700037032365799, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700037032365799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11996520310640335, + 0.18068428337574005, + 0.21212053298950195, + 0.07831678539514542, + 0.13976024091243744, + 0.041543398052453995, + 0.07330045849084854, + 0.0678035318851471, + 0.03849143907427788, + 0.04801405593752861 + ], + "credit_map": [ + 0.013032199814915657, + 0.016378341242671013, + 0.021165145561099052, + 0.012438533827662468, + 0.012393616139888763, + 0.013770642690360546, + 0.011617080308496952, + 0.019169647246599197, + 0.020933059975504875, + 0.019885296002030373, + 0.020850418135523796, + 0.019420918077230453, + 0.017196299508213997, + 0.02414507046341896, + 0.026833737269043922, + 0.03141949698328972, + 0.024355391040444374, + 0.01443922147154808, + 0.024421365931630135, + 0.0129462955519557, + 0.014397191815078259, + 0.05317384749650955, + 0.016979388892650604, + 0.051407843828201294, + 0.0057824947871267796, + 0.01795322075486183, + 0.010318056680262089, + 0.008173814974725246, + 0.008480780757963657, + 0.009349333122372627, + 0.009897771291434765, + 0.00836132001131773, + 0.01568710058927536, + 0.016806604340672493, + 0.03780033066868782, + 0.014161229133605957, + 0.016706248745322227, + 0.013810183852910995, + 0.012134511023759842, + 0.012654034420847893, + 0.0059558493085205555, + 0.006329379975795746, + 0.004890686832368374, + 0.004728496074676514, + 0.005362319760024548, + 0.005348744336515665, + 0.0038143154233694077, + 0.005113611929118633, + 0.010879039764404297, + 0.01275395043194294, + 0.015304344706237316, + 0.005473489407449961, + 0.010400062426924706, + 0.00560494652017951, + 0.006834516767412424, + 0.006050118710845709, + 0.007319582626223564, + 0.008331447839736938, + 0.006830827798694372, + 0.005557570606470108, + 0.010318910703063011, + 0.012301137670874596, + 0.010468472726643085, + 0.0066755847074091434, + 0.005808098241686821, + 0.008426191285252571, + 0.004863651469349861, + 0.0035603775177150965, + 0.004522740840911865, + 0.004850244615226984, + 0.003491874085739255, + 0.002968264278024435, + 0.007000562734901905, + 0.007552243769168854, + 0.006224777549505234, + 0.0034820360597223043, + 0.005236861761659384, + 0.006990515161305666, + 0.004311878699809313, + 0.007215187419205904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.0700037032365799 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a striking wide-angle shot, then the camera zooms in and out at breakneck speed, simulating the dynamic contraction of all visual elements toward the center of the frame. Within this fractured, surreal dimensional space, countless floating geometric fragments, twisted urban debris, and nebulous streaks of light, initially suspended in a pitch-black void, are now violently pulled and deformed by an invisible, immense gravitational force, transforming into dazzling streams of light that converge relentlessly toward the young man at the absolute center of the frame. The young man, clad in a tight, matte, dark blue combat suit with futuristic lines, leans back slightly, arms outstretched, his short, jet-black hair flowing in an invisible current. His eyes are closed, his pale face betraying the pain and fortitude of enduring the immense energy impact. Finally, all matter and energy collapsed violently within a ten-thousandth of a second, condensing into a hexagonal crystal embedded under the skin of his chest. The crystal was beating rhythmically, steadily and powerfully, and each beat emitted a soft and sacred blue glow. This light penetrated his combat uniform, not only illuminating his resolute face, but also providing the only source of light for this dark space that had just undergone a devastating reconstruction, creating an epic atmosphere of sacredness, rebirth and full of unknown power.", + "seed": 3864, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_13/index_1544_seed_3864_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.047858170606195925, + "vlm_logprob": -0.047858170606195925, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.047858170606195925, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10275252908468246 + ], + "opposite_logprob_tokens": [ + -2.352752447128296 + ], + "logprob_reward": -0.10275252908468246, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10275252908468246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2430170476436615, + 0.15030421316623688, + 0.04614033177495003, + 0.06046729162335396, + 0.24459753930568695, + 0.13265123963356018, + 0.05327720195055008, + 0.028044767677783966, + 0.015699036419391632, + 0.025801362469792366 + ], + "credit_map": [ + 0.02741541899740696, + 0.05855172127485275, + 0.030568638816475868, + 0.028013547882437706, + 0.021821413189172745, + 0.032623447477817535, + 0.019485577940940857, + 0.024537239223718643, + 0.02652127481997013, + 0.01173248142004013, + 0.01712687872350216, + 0.010916542261838913, + 0.012077823281288147, + 0.03258957713842392, + 0.008770104497671127, + 0.030569525435566902, + 0.004731412045657635, + 0.00634699035435915, + 0.006830612663179636, + 0.006643330212682486, + 0.005918324459344149, + 0.004968844819813967, + 0.006025714334100485, + 0.004675101023167372, + 0.007556655444204807, + 0.0077478461898863316, + 0.005610049702227116, + 0.005631215404719114, + 0.005678271409124136, + 0.017421673983335495, + 0.006597265601158142, + 0.004224308300763369, + 0.031122995540499687, + 0.023731645196676254, + 0.01789398305118084, + 0.029753560200333595, + 0.0258681308478117, + 0.04218340292572975, + 0.04962645471096039, + 0.02441735565662384, + 0.014936406165361404, + 0.013669108971953392, + 0.025834038853645325, + 0.01494732964783907, + 0.017853014171123505, + 0.012874645181000233, + 0.015100210905075073, + 0.017436455935239792, + 0.006648086942732334, + 0.003758010221645236, + 0.005046951584517956, + 0.006976660341024399, + 0.0065053654834628105, + 0.008088938891887665, + 0.007443140260875225, + 0.008810041472315788, + 0.004149571526795626, + 0.002295905724167824, + 0.0036843481939285994, + 0.0035904934629797935, + 0.0027752798050642014, + 0.004845506977289915, + 0.002663380466401577, + 0.004040278494358063, + 0.0030654584988951683, + 0.0016637661028653383, + 0.001837616553530097, + 0.0012806992745026946, + 0.0018724300898611546, + 0.0032189562916755676, + 0.0012773509370163083, + 0.0014827579725533724, + 0.003245388623327017, + 0.0017900466918945312, + 0.0037340393755584955, + 0.004080526065081358, + 0.002786864759400487, + 0.003217374673113227, + 0.0037193454336375, + 0.0032277738209813833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.10275252908468246 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062029238790273666 + ], + "opposite_logprob_tokens": [ + -2.8120291233062744 + ], + "logprob_reward": -0.062029238790273666, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062029238790273666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1156291589140892, + 0.11270460486412048, + 0.053291503340005875, + 0.06251108646392822, + 0.22686296701431274, + 0.04162539169192314, + 0.10401598364114761, + 0.10571613907814026, + 0.055784232914447784, + 0.12185892462730408 + ], + "credit_map": [ + 0.01447886973619461, + 0.02382749691605568, + 0.01277509331703186, + 0.014976650476455688, + 0.010759078897535801, + 0.016085054725408554, + 0.011828338727355003, + 0.010898576118052006, + 0.03063879907131195, + 0.00853982288390398, + 0.007305952254682779, + 0.005745879840105772, + 0.009658213704824448, + 0.034753989428281784, + 0.007054508198052645, + 0.009007441811263561, + 0.007060528267174959, + 0.005671871826052666, + 0.006202221382409334, + 0.0073839896358549595, + 0.007136958185583353, + 0.007700677495449781, + 0.005987996235489845, + 0.006147262640297413, + 0.008548960089683533, + 0.008247089572250843, + 0.004314962308853865, + 0.005072291940450668, + 0.01108494307845831, + 0.014009803533554077, + 0.00642819469794631, + 0.004804837983101606, + 0.026823941618204117, + 0.02317097783088684, + 0.017185799777507782, + 0.033113084733486176, + 0.02834721840918064, + 0.03385869413614273, + 0.04466293752193451, + 0.019700312986969948, + 0.004545761272311211, + 0.004225610289722681, + 0.009123709052801132, + 0.004475221503525972, + 0.007265087217092514, + 0.004440415184944868, + 0.0037871345411986113, + 0.0037624528631567955, + 0.017319029197096825, + 0.007726356852799654, + 0.00929697323590517, + 0.009075737558305264, + 0.011221480555832386, + 0.025703204795718193, + 0.013360686600208282, + 0.010312519036233425, + 0.014539999887347221, + 0.010946330614387989, + 0.012231243774294853, + 0.011902847327291965, + 0.021489271894097328, + 0.00993525143712759, + 0.012465453706681728, + 0.01220573578029871, + 0.008678706362843513, + 0.003485070774331689, + 0.00983518362045288, + 0.008077310398221016, + 0.005830526351928711, + 0.00862649455666542, + 0.005609393585473299, + 0.005641544703394175, + 0.013941200450062752, + 0.00905484240502119, + 0.015537410043179989, + 0.01992676593363285, + 0.01091716717928648, + 0.013030244037508965, + 0.017247134819626808, + 0.02220415510237217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.062029238790273666 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780699387192726 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780699387192726, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780699387192726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09414812922477722, + 0.08438986539840698, + 0.10710928589105606, + 0.0863020196557045, + 0.08389320969581604, + 0.05500450357794762, + 0.12430045008659363, + 0.09873751550912857, + 0.0611334852874279, + 0.20498155057430267 + ], + "credit_map": [ + 0.012364315800368786, + 0.016751088201999664, + 0.010609731078147888, + 0.009936130605638027, + 0.010167303495109081, + 0.01279685739427805, + 0.009128140285611153, + 0.012394559569656849, + 0.01899660937488079, + 0.004981113597750664, + 0.007720747962594032, + 0.006827902514487505, + 0.005122306756675243, + 0.024422455579042435, + 0.007114189676940441, + 0.009204533882439137, + 0.01207824144512415, + 0.011649232357740402, + 0.013926778919994831, + 0.017816515639424324, + 0.012713679112493992, + 0.012402983382344246, + 0.016047155484557152, + 0.010474706999957561, + 0.01247416902333498, + 0.014075778424739838, + 0.006490134634077549, + 0.010307864286005497, + 0.015805790200829506, + 0.008748339489102364, + 0.01200888678431511, + 0.006391051225364208, + 0.01351778395473957, + 0.008298857137560844, + 0.009483300149440765, + 0.010997657664120197, + 0.008853883482515812, + 0.009882735088467598, + 0.012196819297969341, + 0.010662169195711613, + 0.006734809838235378, + 0.0057491520419716835, + 0.007345132529735565, + 0.006942287087440491, + 0.00885805580765009, + 0.005819248966872692, + 0.006486678496003151, + 0.007069136016070843, + 0.014691258780658245, + 0.008444725535809994, + 0.015928413718938828, + 0.013465547002851963, + 0.019476771354675293, + 0.016541946679353714, + 0.01677684485912323, + 0.018974943086504936, + 0.01494662556797266, + 0.010896752588450909, + 0.011919376440346241, + 0.009137981571257114, + 0.020702291280031204, + 0.009725317358970642, + 0.009529775008559227, + 0.01187939289957285, + 0.007010418921709061, + 0.004455039743334055, + 0.012452949769794941, + 0.0072501832619309425, + 0.006380752194672823, + 0.00847233273088932, + 0.006568331271409988, + 0.008543477393686771, + 0.02332199178636074, + 0.017233066260814667, + 0.038820214569568634, + 0.02038705348968506, + 0.02741417847573757, + 0.027758341282606125, + 0.018201919272542, + 0.03184477239847183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.009780699387192726 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781762026250362 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.009781762026250362, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781762026250362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09790560603141785, + 0.07945066690444946, + 0.2753320634365082, + 0.044097453355789185, + 0.06604238599538803, + 0.05661511421203613, + 0.05001659691333771, + 0.0325402170419693, + 0.12630325555801392, + 0.17169669270515442 + ], + "credit_map": [ + 0.0134020010009408, + 0.017070531845092773, + 0.011275647208094597, + 0.010652372613549232, + 0.011940822936594486, + 0.014977601356804371, + 0.010523354634642601, + 0.008063281886279583, + 0.009874396957457066, + 0.008771423250436783, + 0.010886416770517826, + 0.009135995991528034, + 0.005642827600240707, + 0.011448096483945847, + 0.009499834850430489, + 0.014191676862537861, + 0.02753731794655323, + 0.025991328060626984, + 0.03286997601389885, + 0.05372011661529541, + 0.0332723893225193, + 0.032491087913513184, + 0.04484451189637184, + 0.02460533194243908, + 0.005924465600401163, + 0.005916492082178593, + 0.005562514066696167, + 0.004221697803586721, + 0.006813269108533859, + 0.006990500725805759, + 0.005415136925876141, + 0.0032533761113882065, + 0.009601711295545101, + 0.008890597149729729, + 0.008730726316571236, + 0.009517457336187363, + 0.006212624721229076, + 0.005872364621609449, + 0.00978371873497963, + 0.007433189079165459, + 0.007871306501328945, + 0.0080414405092597, + 0.006550790276378393, + 0.005992876831442118, + 0.009660580195486546, + 0.006462686695158482, + 0.007125675678253174, + 0.004909761715680361, + 0.008938203565776348, + 0.0030947099439799786, + 0.004596110433340073, + 0.005396742839366198, + 0.004396077711135149, + 0.01264773029834032, + 0.005860364995896816, + 0.005086658988147974, + 0.004959802608937025, + 0.003481004387140274, + 0.0030301667284220457, + 0.005064547993242741, + 0.006048906594514847, + 0.0035668695345520973, + 0.003629119135439396, + 0.00275979982689023, + 0.018001614138484, + 0.01002754271030426, + 0.02330172061920166, + 0.01829175092279911, + 0.012803804129362106, + 0.015446033328771591, + 0.014680947177112103, + 0.013749834150075912, + 0.024963900446891785, + 0.011316009797155857, + 0.023965537548065186, + 0.027269616723060608, + 0.013630841858685017, + 0.020928815007209778, + 0.023835722357034683, + 0.025786248967051506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.009781762026250362 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054946623742580414 + ], + "opposite_logprob_tokens": [ + -2.9299466609954834 + ], + "logprob_reward": -0.054946623742580414, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054946623742580414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07398735731840134, + 0.1591491997241974, + 0.17816485464572906, + 0.08155101537704468, + 0.03708897903561592, + 0.09593313187360764, + 0.07751644402742386, + 0.07107853144407272, + 0.08923935145139694, + 0.13629114627838135 + ], + "credit_map": [ + 0.007080869749188423, + 0.008397511206567287, + 0.008879492990672588, + 0.009507497772574425, + 0.007786300498992205, + 0.010560207068920135, + 0.009302479214966297, + 0.012473003007471561, + 0.029545890167355537, + 0.019371209666132927, + 0.008801505900919437, + 0.01775994338095188, + 0.012612058781087399, + 0.03845648095011711, + 0.01907466910779476, + 0.01352743525058031, + 0.01616843417286873, + 0.017941396683454514, + 0.02268175594508648, + 0.03653424605727196, + 0.018785379827022552, + 0.01726595126092434, + 0.033378005027770996, + 0.015409691259264946, + 0.010688203386962414, + 0.007205725647509098, + 0.012796318158507347, + 0.008482849225401878, + 0.007780616171658039, + 0.016302814707159996, + 0.011707241646945477, + 0.006587242241948843, + 0.005826364737004042, + 0.0034176521003246307, + 0.006322088185697794, + 0.0038215727545320988, + 0.0038329120725393295, + 0.005029820837080479, + 0.0038111249450594187, + 0.005027445498853922, + 0.011022915132343769, + 0.010909251868724823, + 0.011193071492016315, + 0.012104036286473274, + 0.01952425017952919, + 0.008862684480845928, + 0.012319917790591717, + 0.009997007437050343, + 0.011086734011769295, + 0.004991753958165646, + 0.01099978294223547, + 0.008822225965559483, + 0.01108804065734148, + 0.013450536876916885, + 0.0077858492732048035, + 0.009291522204875946, + 0.007738569285720587, + 0.006915424484759569, + 0.010847522877156734, + 0.007782180327922106, + 0.013307642191648483, + 0.01003771461546421, + 0.006846131756901741, + 0.0076033491641283035, + 0.014697624370455742, + 0.0072678448632359505, + 0.013873050920665264, + 0.011038639582693577, + 0.009760187938809395, + 0.01500236801803112, + 0.008753067813813686, + 0.008846565149724483, + 0.01560216210782528, + 0.009776216931641102, + 0.030403707176446915, + 0.011450869962573051, + 0.012705178931355476, + 0.015223821625113487, + 0.012598950415849686, + 0.028530243784189224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.054946623742580414 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a striking wide-angle shot, then the camera zooms in and out at breakneck speed, simulating the dynamic contraction of all visual elements toward the center of the frame. Within this fractured, surreal dimensional space, countless floating geometric fragments, twisted urban debris, and nebulous streaks of light, initially suspended in a pitch-black void, are now violently pulled and deformed by an invisible, immense gravitational force, transforming into dazzling streams of light that converge relentlessly toward the young man at the absolute center of the frame. The young man, clad in a tight, matte, dark blue combat suit with futuristic lines, leans back slightly, arms outstretched, his short, jet-black hair flowing in an invisible current. His eyes are closed, his pale face betraying the pain and fortitude of enduring the immense energy impact. Finally, all matter and energy collapsed violently within a ten-thousandth of a second, condensing into a hexagonal crystal embedded under the skin of his chest. The crystal was beating rhythmically, steadily and powerfully, and each beat emitted a soft and sacred blue glow. This light penetrated his combat uniform, not only illuminating his resolute face, but also providing the only source of light for this dark space that had just undergone a devastating reconstruction, creating an epic atmosphere of sacredness, rebirth and full of unknown power.", + "seed": 3865, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_13/index_1544_seed_3865_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.025213223416358232, + "vlm_logprob": -0.025213223416358232, + "vlm_yes_no_margin": 4.024999952316284, + "vlm_token_credit": -0.025213223416358232, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.045567531138658524 + ], + "opposite_logprob_tokens": [ + -3.170567512512207 + ], + "logprob_reward": -0.045567531138658524, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.045567531138658524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17173117399215698, + 0.08577200770378113, + 0.1980663239955902, + 0.09017528593540192, + 0.081167072057724, + 0.05583050101995468, + 0.056835249066352844, + 0.06091110035777092, + 0.12979662418365479, + 0.06971469521522522 + ], + "credit_map": [ + 0.02095760405063629, + 0.01483867783099413, + 0.027571383863687515, + 0.027511874213814735, + 0.018409287557005882, + 0.01144329085946083, + 0.035461362451314926, + 0.015537691302597523, + 0.007403598167002201, + 0.009981208480894566, + 0.008960481733083725, + 0.011314928531646729, + 0.01088003721088171, + 0.010257662273943424, + 0.010350906290113926, + 0.016623178496956825, + 0.02111014910042286, + 0.022129643708467484, + 0.032293856143951416, + 0.026237623766064644, + 0.03867905214428902, + 0.01891297474503517, + 0.012250574305653572, + 0.026452450081706047, + 0.010462715290486813, + 0.015368764288723469, + 0.00702162180095911, + 0.004119421821087599, + 0.01998147927224636, + 0.01692996732890606, + 0.007560885511338711, + 0.008730428293347359, + 0.007963225245475769, + 0.014813795685768127, + 0.008550131693482399, + 0.014562512747943401, + 0.009735198691487312, + 0.009540625847876072, + 0.010359380394220352, + 0.005642199423164129, + 0.006074214819818735, + 0.009178598411381245, + 0.0048813596367836, + 0.005637090187519789, + 0.006924140267074108, + 0.008420819416642189, + 0.006968275643885136, + 0.007745998911559582, + 0.006806409452110529, + 0.00889138039201498, + 0.006568651646375656, + 0.006250149570405483, + 0.006178262643516064, + 0.004735726397484541, + 0.009266159497201443, + 0.008138508535921574, + 0.006516330875456333, + 0.005338013172149658, + 0.006451309658586979, + 0.006922086700797081, + 0.00819894578307867, + 0.009225726127624512, + 0.008435677736997604, + 0.009823012165725231, + 0.009230290539562702, + 0.006080904044210911, + 0.009763943031430244, + 0.008424049243330956, + 0.03511185199022293, + 0.0359574593603611, + 0.012410908006131649, + 0.012817214243113995, + 0.00750736566260457, + 0.008295014500617981, + 0.008533491753041744, + 0.009351503103971481, + 0.005440826993435621, + 0.011019294150173664, + 0.009126993827521801, + 0.010440203361213207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.045567531138658524 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188124522566795 + ], + "opposite_logprob_tokens": [ + -4.018187999725342 + ], + "logprob_reward": -0.018188124522566795, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018188124522566795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08608337491750717, + 0.029720120131969452, + 0.03520895540714264, + 0.04225588217377663, + 0.052561625838279724, + 0.14541907608509064, + 0.12332803010940552, + 0.17950214445590973, + 0.24395021796226501, + 0.061970558017492294 + ], + "credit_map": [ + 0.011312698945403099, + 0.011467144824564457, + 0.020119715481996536, + 0.00861282367259264, + 0.010504767298698425, + 0.00603927718475461, + 0.009107743389904499, + 0.008919215761125088, + 0.002821952337399125, + 0.0029879019130021334, + 0.005380850750952959, + 0.005161084700375795, + 0.0026629671920090914, + 0.002543420996516943, + 0.0030507908668369055, + 0.005111151374876499, + 0.0036710568238049746, + 0.004817930515855551, + 0.004833436105400324, + 0.00456175347790122, + 0.004877573344856501, + 0.003840986406430602, + 0.003734811209142208, + 0.004871409386396408, + 0.0044066268019378185, + 0.007296941243112087, + 0.0037649075966328382, + 0.0031605158001184464, + 0.005598241928964853, + 0.010614769533276558, + 0.0028419604059308767, + 0.004571918863803148, + 0.006863140966743231, + 0.007871213369071484, + 0.006057656370103359, + 0.007125247269868851, + 0.00700039928779006, + 0.005447085946798325, + 0.006485689897090197, + 0.005711197387427092, + 0.018199779093265533, + 0.019817911088466644, + 0.014309992082417011, + 0.0177683774381876, + 0.019139720126986504, + 0.01583590917289257, + 0.02147665247321129, + 0.018870731815695763, + 0.013555229641497135, + 0.016043657436966896, + 0.015886882320046425, + 0.012901915237307549, + 0.013747277669608593, + 0.012100011110305786, + 0.022105595096945763, + 0.016987459734082222, + 0.015463934279978275, + 0.025415631011128426, + 0.02397351898252964, + 0.03130465745925903, + 0.020080335438251495, + 0.01912977173924446, + 0.026985296979546547, + 0.017148999497294426, + 0.01959160342812538, + 0.014744576066732407, + 0.03123203106224537, + 0.038537390530109406, + 0.028643302619457245, + 0.03602220118045807, + 0.052691586315631866, + 0.02248753048479557, + 0.00879328791052103, + 0.007596240844577551, + 0.007022844161838293, + 0.009030879475176334, + 0.005990775767713785, + 0.007263720501214266, + 0.009095801040530205, + 0.007177014835178852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.018188124522566795 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003192449454218149 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.003192449454218149, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003192449454218149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043770160526037216, + 0.01666976325213909, + 0.019213352352380753, + 0.015892749652266502, + 0.05636996775865555, + 0.07016728818416595, + 0.12258557975292206, + 0.41474127769470215, + 0.03469841554760933, + 0.2058914303779602 + ], + "credit_map": [ + 0.006068740040063858, + 0.007135327905416489, + 0.008839311078190804, + 0.005639120936393738, + 0.004063378553837538, + 0.0026857098564505577, + 0.006002138834446669, + 0.00333643751218915, + 0.0015286867273971438, + 0.0019415440037846565, + 0.0018830026965588331, + 0.002271257573738694, + 0.0022971865255385637, + 0.0020193203818053007, + 0.001625416218303144, + 0.0031033519189804792, + 0.0022709749173372984, + 0.002325501060113311, + 0.002637682482600212, + 0.0029353792779147625, + 0.0020388197153806686, + 0.0021878851111978292, + 0.002361901802942157, + 0.002455211477354169, + 0.0021993073169142008, + 0.0024451464414596558, + 0.0021819802932441235, + 0.0010230548214167356, + 0.0020984322763979435, + 0.002843925030902028, + 0.0012401248095557094, + 0.001860779826529324, + 0.012062788009643555, + 0.008794581517577171, + 0.0043943352065980434, + 0.006010198500007391, + 0.007512534502893686, + 0.007809402421116829, + 0.004495920147746801, + 0.005290214437991381, + 0.007752073463052511, + 0.011182326823472977, + 0.008344984613358974, + 0.00919559970498085, + 0.009046949446201324, + 0.007590205408632755, + 0.009442491456866264, + 0.007612660061568022, + 0.011832354590296745, + 0.01437543798238039, + 0.01679525524377823, + 0.012350357137620449, + 0.011666316539049149, + 0.009654157795011997, + 0.0263555608689785, + 0.01955614984035492, + 0.039322223514318466, + 0.040703512728214264, + 0.046659957617521286, + 0.05251554399728775, + 0.05258575826883316, + 0.06705458462238312, + 0.053454142063856125, + 0.06244559586048126, + 0.0022266353480517864, + 0.002922693733125925, + 0.003711065510287881, + 0.006091872230172157, + 0.003842843696475029, + 0.005453239660710096, + 0.00623789057135582, + 0.004212176892906427, + 0.016264023259282112, + 0.013757345266640186, + 0.02752402238547802, + 0.033804915845394135, + 0.017627524212002754, + 0.03211064264178276, + 0.03411383181810379, + 0.030689144507050514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.003192449454218149 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066374257206917 + ], + "opposite_logprob_tokens": [ + -4.141066551208496 + ], + "logprob_reward": -0.016066374257206917, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066374257206917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01744086667895317, + 0.01056504063308239, + 0.01669285073876381, + 0.009599672630429268, + 0.011735990643501282, + 0.035301849246025085, + 0.03308216109871864, + 0.02276935800909996, + 0.04189341515302658, + 0.8009187579154968 + ], + "credit_map": [ + 0.0021476494148373604, + 0.003305295715108514, + 0.0027908135671168566, + 0.002076758537441492, + 0.0021003740839660168, + 0.0015129661187529564, + 0.0016973207239061594, + 0.0018096868880093098, + 0.0015534228878095746, + 0.0012719384394586086, + 0.0011021981481462717, + 0.001370040001347661, + 0.0008842468378134072, + 0.0009269359870813787, + 0.0010427910601720214, + 0.0024134668055921793, + 0.0018254047026857734, + 0.0017285689245909452, + 0.0018077221466228366, + 0.0021363929845392704, + 0.00209035468287766, + 0.0012875980464741588, + 0.00387355568818748, + 0.0019432520493865013, + 0.0011283174389973283, + 0.0015435623936355114, + 0.0009699801448732615, + 0.0007541872910223901, + 0.001833435264416039, + 0.0014045616844668984, + 0.0009815279627218843, + 0.000984100392088294, + 0.002064033644273877, + 0.0018718187930062413, + 0.0015462151495739818, + 0.0009983094641938806, + 0.0014749850379303098, + 0.001831947942264378, + 0.00106395676266402, + 0.0008847237913869321, + 0.004948833957314491, + 0.004583952482789755, + 0.0038887422997504473, + 0.00446701655164361, + 0.004723313730210066, + 0.003703926457092166, + 0.005452876444905996, + 0.0035331854596734047, + 0.0038228111807256937, + 0.005677756853401661, + 0.0035457387566566467, + 0.003460736246779561, + 0.0037961043417453766, + 0.0024838214740157127, + 0.005364171229302883, + 0.004931021016091108, + 0.002760774688795209, + 0.002656032796949148, + 0.002530457219108939, + 0.0031858482398092747, + 0.0028382812160998583, + 0.002855028724297881, + 0.002909949515014887, + 0.003032985609024763, + 0.0038298091385513544, + 0.0038838551845401525, + 0.0046064406633377075, + 0.008476631715893745, + 0.004052056465297937, + 0.005907767917960882, + 0.0065038674511015415, + 0.004632987547665834, + 0.06671889126300812, + 0.048983484506607056, + 0.1077619418501854, + 0.12846748530864716, + 0.06455008685588837, + 0.12599769234657288, + 0.12783023715019226, + 0.13060897588729858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.016066374257206917 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a pitch-black void containing floating geometric fragments and twisted urban debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man visible at the absolute center of the frame wearing a tight, matte, dark blue combat suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the young man have short, jet-black hair, closed eyes, and an outstretched posture while leaning back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a hexagonal crystal visible embedded under the skin of the young man's chest emitting a soft blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute rapid zooms in and out to simulate dynamic contraction of visual elements toward the center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305163770914078 + ], + "opposite_logprob_tokens": [ + -3.1680517196655273 + ], + "logprob_reward": -0.04305163770914078, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305163770914078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19156865775585175, + 0.0540144108235836, + 0.06406474113464355, + 0.030903682112693787, + 0.05065542459487915, + 0.06186937913298607, + 0.06761704385280609, + 0.33852362632751465, + 0.052564214915037155, + 0.08821888267993927 + ], + "credit_map": [ + 0.025293972343206406, + 0.028478389605879784, + 0.025659147650003433, + 0.032566145062446594, + 0.013868996873497963, + 0.017491567879915237, + 0.01850159652531147, + 0.029708828777074814, + 0.004211280960589647, + 0.006809941492974758, + 0.010613721795380116, + 0.008004490286111832, + 0.006467081140726805, + 0.005929308943450451, + 0.0056920964270830154, + 0.006286489777266979, + 0.004531295970082283, + 0.006542488932609558, + 0.0063494606874883175, + 0.017137499526143074, + 0.010589489713311195, + 0.003796142293140292, + 0.007687803357839584, + 0.0074305543676018715, + 0.003245172556489706, + 0.004785323515534401, + 0.0033358680084347725, + 0.0021863961592316628, + 0.005124588962644339, + 0.004943148698657751, + 0.003500809893012047, + 0.0037823752500116825, + 0.006138709373772144, + 0.007645113859325647, + 0.005810924340039492, + 0.00891825184226036, + 0.0062029724940657616, + 0.004743256140500307, + 0.0069612483493983746, + 0.004234947729855776, + 0.008297191001474857, + 0.010896963998675346, + 0.006340000778436661, + 0.007504155393689871, + 0.007107678800821304, + 0.0066331420093774796, + 0.008560379967093468, + 0.006529861129820347, + 0.009141464717686176, + 0.00922301784157753, + 0.0077971662394702435, + 0.007187813054770231, + 0.009006834588944912, + 0.005768985021859407, + 0.010501952841877937, + 0.008989808149635792, + 0.03452693670988083, + 0.03419730067253113, + 0.039621833711862564, + 0.0435548759996891, + 0.043574295938014984, + 0.05018196627497673, + 0.043057650327682495, + 0.04980877786874771, + 0.004821456968784332, + 0.0050712889060378075, + 0.006602274253964424, + 0.008161311969161034, + 0.008327199146151543, + 0.008774396032094955, + 0.006208750884979963, + 0.004597535822540522, + 0.007989192381501198, + 0.006254766136407852, + 0.011441693641245365, + 0.013340211473405361, + 0.00765596516430378, + 0.013064687140285969, + 0.014498288743197918, + 0.013974077999591827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.04305163770914078 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_14/index_tensor([40, 40])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_14/index_tensor([40, 40])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..610befae7959412d94fdb263139a3fbd5c4b080a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_14/index_tensor([40, 40])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, horizontal, medium shot, shot from behind, captures a man with slightly disheveled gray hair, wearing a dark brown tweed jacket and a beige wool sweater. He sits alone in a high-backed leather armchair, his back to the camera, facing a large arched window. Outside, dusk falls, and a gentle rain silently drizzles across the glass. His study is surrounded by towering, dark wooden bookshelves, crammed with ancient books of varying thickness, bound in leather or cloth. Some books are even haphazardly stacked on the floor and on the corners of the table. The man leans forward slightly, his elbows resting on his knees, his hands clasped before him, his head slightly lowered, his entire body almost motionless, completely absorbed in his thoughts. The cool twilight filtering in from the window serves as the primary light source, softly outlining his shoulders and head. Farther back, a small desk lamp casts a warm, orange glow, illuminating the gilded lettering on the book spines. The entire scene is dominated by deep wood tones and cool blue, dotted with warm lights, creating an academic atmosphere of extreme tranquility, concentration, contemplation, and a hint of melancholy.", + "seed": 28268, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_14/index_40_seed_28268_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03687990582548082, + "vlm_logprob": -0.03687990582548082, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.03687990582548082, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008876270614564419 + ], + "opposite_logprob_tokens": [ + -5.258876323699951 + ], + "logprob_reward": -0.008876270614564419, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008876270614564419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3312782049179077, + 0.13163261115550995, + 0.09197533875703812, + 0.07667068392038345, + 0.051259782165288925, + 0.055529046803712845, + 0.0608178973197937, + 0.059716418385505676, + 0.062483206391334534, + 0.07863681763410568 + ], + "credit_map": [ + 0.09029769152402878, + 0.025231290608644485, + 0.04087583348155022, + 0.05281439796090126, + 0.02558288723230362, + 0.026241345331072807, + 0.037563908845186234, + 0.03267085552215576, + 0.03204375505447388, + 0.008536932989954948, + 0.023031599819660187, + 0.02690458483994007, + 0.006645245011895895, + 0.009552674368023872, + 0.012996574863791466, + 0.011921252124011517, + 0.01958918571472168, + 0.006607488729059696, + 0.01591692492365837, + 0.01802690699696541, + 0.006275178398936987, + 0.006538283079862595, + 0.009707766585052013, + 0.009313603863120079, + 0.01673050783574581, + 0.005257178097963333, + 0.01025294978171587, + 0.012008024379611015, + 0.007405542302876711, + 0.006075143814086914, + 0.009231277741491795, + 0.009710053913295269, + 0.010381309315562248, + 0.004958226811140776, + 0.007281818892806768, + 0.00912037305533886, + 0.0038855362217873335, + 0.004695471841841936, + 0.006070030853152275, + 0.004867014940828085, + 0.008528819307684898, + 0.004698524251580238, + 0.007878485135734081, + 0.010365312919020653, + 0.005685090087354183, + 0.00411719037219882, + 0.007920395582914352, + 0.006335227284580469, + 0.011060531251132488, + 0.004396175965666771, + 0.011619647033512592, + 0.012467429041862488, + 0.004047892987728119, + 0.004334669094532728, + 0.007540018763393164, + 0.0053515308536589146, + 0.01481879223138094, + 0.0054217251017689705, + 0.007529492489993572, + 0.010477851144969463, + 0.005704247858375311, + 0.0045879315584897995, + 0.005594595335423946, + 0.0055817849934101105, + 0.016751619055867195, + 0.0038406294770538807, + 0.009263461455702782, + 0.010241379961371422, + 0.004921694751828909, + 0.0037649895530194044, + 0.007616383023560047, + 0.006083051208406687, + 0.018552130088210106, + 0.007574163377285004, + 0.010682016611099243, + 0.011538083665072918, + 0.005604649428278208, + 0.0065069664269685745, + 0.009935370646417141, + 0.0082434406504035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.008876270614564419 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062040552496910095 + ], + "opposite_logprob_tokens": [ + -2.8120405673980713 + ], + "logprob_reward": -0.062040552496910095, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062040552496910095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4164485037326813, + 0.11685691773891449, + 0.09092721343040466, + 0.04971592500805855, + 0.06712733954191208, + 0.04752437770366669, + 0.03732950985431671, + 0.061868567019701004, + 0.047151997685432434, + 0.06504963338375092 + ], + "credit_map": [ + 0.03643862158060074, + 0.03309618681669235, + 0.06670069694519043, + 0.07576934248209, + 0.0464402511715889, + 0.03460409864783287, + 0.057813066989183426, + 0.06558626890182495, + 0.0094002615660429, + 0.01103849895298481, + 0.015628615394234657, + 0.019452514126896858, + 0.013932720758020878, + 0.010288768447935581, + 0.016860738396644592, + 0.02025480568408966, + 0.007844883017241955, + 0.007639415096491575, + 0.012870953418314457, + 0.016955040395259857, + 0.011414222419261932, + 0.008450967259705067, + 0.012647600844502449, + 0.013104140758514404, + 0.0052171628922224045, + 0.0038329993840306997, + 0.0076472461223602295, + 0.009454488754272461, + 0.00603232579305768, + 0.004063804633915424, + 0.006357481237500906, + 0.007110423408448696, + 0.007320363540202379, + 0.00618312181904912, + 0.0071593220345675945, + 0.010518449358642101, + 0.009453336708247662, + 0.007444256916642189, + 0.008504649624228477, + 0.01054383721202612, + 0.0055987099185585976, + 0.004249845631420612, + 0.004974781069904566, + 0.007749912329018116, + 0.006612473167479038, + 0.004120808094739914, + 0.006731582339853048, + 0.007486269809305668, + 0.005265865474939346, + 0.002646737266331911, + 0.007337955757975578, + 0.0077696372754871845, + 0.0037756669335067272, + 0.0025447309017181396, + 0.004250759724527597, + 0.003738157683983445, + 0.006727033760398626, + 0.004835848696529865, + 0.008756288327276707, + 0.010828283615410328, + 0.008199863135814667, + 0.005639484617859125, + 0.007919295690953732, + 0.008962470106780529, + 0.006437386851757765, + 0.003743437584489584, + 0.004600072745233774, + 0.007045023143291473, + 0.006719164550304413, + 0.0041997358202934265, + 0.0072465138509869576, + 0.0071606626734137535, + 0.008009422570466995, + 0.004889378789812326, + 0.008558775298297405, + 0.010854237712919712, + 0.008743547834455967, + 0.005190469790250063, + 0.010379892773926258, + 0.00842391699552536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.062040552496910095 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028156896587461233 + ], + "opposite_logprob_tokens": [ + -5.8778157234191895 + ], + "logprob_reward": -0.0028156896587461233, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028156896587461233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40611302852630615, + 0.13647305965423584, + 0.07697787135839462, + 0.04680224135518074, + 0.046514153480529785, + 0.0602082833647728, + 0.04159730672836304, + 0.04568136855959892, + 0.06691967695951462, + 0.07271301001310349 + ], + "credit_map": [ + 0.06703179329633713, + 0.030183756723999977, + 0.06334248930215836, + 0.0683128833770752, + 0.03657691553235054, + 0.020659280940890312, + 0.053434595465660095, + 0.06657133251428604, + 0.026746949180960655, + 0.012795696035027504, + 0.014808913692831993, + 0.020109085366129875, + 0.012064368464052677, + 0.00889615435153246, + 0.021119356155395508, + 0.019932523369789124, + 0.014224683865904808, + 0.0070472098886966705, + 0.009121917188167572, + 0.011476111598312855, + 0.007249513175338507, + 0.004264890681952238, + 0.012600250542163849, + 0.01099329348653555, + 0.006922105792909861, + 0.0037300216499716043, + 0.005195398814976215, + 0.006945738103240728, + 0.004793835338205099, + 0.0030770564917474985, + 0.007883649319410324, + 0.00825443398207426, + 0.008767145685851574, + 0.0032004525419324636, + 0.006155043840408325, + 0.006719792727380991, + 0.003959139343351126, + 0.003178380662575364, + 0.007730891462415457, + 0.006803308147937059, + 0.012950708158314228, + 0.004710662178695202, + 0.00953950546681881, + 0.011490637436509132, + 0.004297842271625996, + 0.004253124352544546, + 0.0067118383012712, + 0.00625396054238081, + 0.008080767467617989, + 0.0030696135945618153, + 0.004265664145350456, + 0.006260423921048641, + 0.003288016887381673, + 0.002848862437531352, + 0.00659001711755991, + 0.0071939402259886265, + 0.011398838832974434, + 0.004236134700477123, + 0.005292275454849005, + 0.006986804306507111, + 0.0035183150321245193, + 0.003939908463507891, + 0.004506339319050312, + 0.005802754312753677, + 0.01589162088930607, + 0.00456590810790658, + 0.011407814919948578, + 0.012785852886736393, + 0.004020212218165398, + 0.004289289470762014, + 0.007255460135638714, + 0.0067035178653895855, + 0.01620047725737095, + 0.006639032159000635, + 0.011646339669823647, + 0.013190253637731075, + 0.0038412283174693584, + 0.005647118203341961, + 0.007872556336224079, + 0.007676003035157919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0028156896587461233 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048648420721292496 + ], + "opposite_logprob_tokens": [ + -3.0486483573913574 + ], + "logprob_reward": -0.048648420721292496, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048648420721292496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36172157526016235, + 0.1445377916097641, + 0.10434327274560928, + 0.0633121058344841, + 0.09745630621910095, + 0.06731648743152618, + 0.04917128011584282, + 0.04379427060484886, + 0.0295022614300251, + 0.038844648748636246 + ], + "credit_map": [ + 0.07703164219856262, + 0.029813088476657867, + 0.04939847066998482, + 0.06757672131061554, + 0.030244629830121994, + 0.023386172950267792, + 0.044584762305021286, + 0.03968607261776924, + 0.042446382343769073, + 0.010338591411709785, + 0.018896088004112244, + 0.023466194048523903, + 0.010479419492185116, + 0.010115353390574455, + 0.014309379272162914, + 0.014486389234662056, + 0.02587718889117241, + 0.009669479914009571, + 0.013152756728231907, + 0.015370219945907593, + 0.01133981253951788, + 0.008435976691544056, + 0.010883883573114872, + 0.009613958187401295, + 0.015810148790478706, + 0.005802023224532604, + 0.00581643870100379, + 0.00907223578542471, + 0.006429937668144703, + 0.006632830947637558, + 0.00720714358612895, + 0.006541344802826643, + 0.018637266010046005, + 0.00982593558728695, + 0.012087482027709484, + 0.013847376219928265, + 0.010984987951815128, + 0.009864519350230694, + 0.010624579153954983, + 0.011584158055484295, + 0.008826224133372307, + 0.007019215729087591, + 0.00873485952615738, + 0.011199536733329296, + 0.006813175044953823, + 0.009598334319889545, + 0.00804121233522892, + 0.007083922158926725, + 0.009883038699626923, + 0.003927030134946108, + 0.00882845651358366, + 0.009897517040371895, + 0.004172245040535927, + 0.003855237504467368, + 0.004303483758121729, + 0.004304271191358566, + 0.009904522448778152, + 0.004609710071235895, + 0.005604258738458157, + 0.006181116681545973, + 0.004098838660866022, + 0.004459836054593325, + 0.004525751806795597, + 0.004410236608237028, + 0.006871131248772144, + 0.0025655345525592566, + 0.0035755818244069815, + 0.004397219512611628, + 0.002941934857517481, + 0.002585233421996236, + 0.0035074162296950817, + 0.0030582118779420853, + 0.00779887055978179, + 0.003986984491348267, + 0.0037796287797391415, + 0.006093012634664774, + 0.004205096513032913, + 0.0038832512218505144, + 0.005057172384113073, + 0.004040633328258991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.048648420721292496 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06201859563589096 + ], + "opposite_logprob_tokens": [ + -2.812018632888794 + ], + "logprob_reward": -0.06201859563589096, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06201859563589096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2659536600112915, + 0.07938069105148315, + 0.06161791458725929, + 0.045436177402734756, + 0.1263052076101303, + 0.15917961299419403, + 0.03223441168665886, + 0.05096057057380676, + 0.0353480726480484, + 0.14358368515968323 + ], + "credit_map": [ + 0.055222656577825546, + 0.024098219349980354, + 0.026310574263334274, + 0.035062968730926514, + 0.02342386357486248, + 0.02370106242597103, + 0.03551610931754112, + 0.042618196457624435, + 0.00842237751930952, + 0.008282640017569065, + 0.009593399241566658, + 0.013752875849604607, + 0.0068465908989310265, + 0.0072561767883598804, + 0.011361113749444485, + 0.01386551558971405, + 0.007516155950725079, + 0.006521770264953375, + 0.007660572417080402, + 0.009458593092858791, + 0.0067675914615392685, + 0.0052240341901779175, + 0.009083353914320469, + 0.009385840967297554, + 0.007051027845591307, + 0.004435372073203325, + 0.006687176413834095, + 0.007628369610756636, + 0.004973581060767174, + 0.0031110330019146204, + 0.006132392678409815, + 0.0054172249510884285, + 0.03568606451153755, + 0.010447490029036999, + 0.014050822705030441, + 0.019808193668723106, + 0.011262810789048672, + 0.009436443448066711, + 0.010683069936931133, + 0.014930308796465397, + 0.05056462809443474, + 0.011715398170053959, + 0.018251920118927956, + 0.026648329570889473, + 0.012238254770636559, + 0.010319054126739502, + 0.011969183571636677, + 0.01747284084558487, + 0.008384743705391884, + 0.0024005838204175234, + 0.0033676184248179197, + 0.005078782327473164, + 0.002467714948579669, + 0.002328320173546672, + 0.004322730470448732, + 0.0038839182816445827, + 0.013312083669006824, + 0.0035613684449344873, + 0.005605365615338087, + 0.007720751687884331, + 0.005427262745797634, + 0.004476672969758511, + 0.005434252321720123, + 0.0054228100925683975, + 0.006134445313364267, + 0.0030627569649368525, + 0.003394945990294218, + 0.00521653238683939, + 0.004850170109421015, + 0.0031138802878558636, + 0.004866233095526695, + 0.004709107801318169, + 0.03772473707795143, + 0.011942846700549126, + 0.016920797526836395, + 0.025401664897799492, + 0.013033627532422543, + 0.010908369906246662, + 0.013057142496109009, + 0.014594503678381443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.06201859563589096 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, horizontal, medium shot, shot from behind, captures a man with slightly disheveled gray hair, wearing a dark brown tweed jacket and a beige wool sweater. He sits alone in a high-backed leather armchair, his back to the camera, facing a large arched window. Outside, dusk falls, and a gentle rain silently drizzles across the glass. His study is surrounded by towering, dark wooden bookshelves, crammed with ancient books of varying thickness, bound in leather or cloth. Some books are even haphazardly stacked on the floor and on the corners of the table. The man leans forward slightly, his elbows resting on his knees, his hands clasped before him, his head slightly lowered, his entire body almost motionless, completely absorbed in his thoughts. The cool twilight filtering in from the window serves as the primary light source, softly outlining his shoulders and head. Farther back, a small desk lamp casts a warm, orange glow, illuminating the gilded lettering on the book spines. The entire scene is dominated by deep wood tones and cool blue, dotted with warm lights, creating an academic atmosphere of extreme tranquility, concentration, contemplation, and a hint of melancholy.", + "seed": 28269, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_14/index_40_seed_28269_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03979582991451025, + "vlm_logprob": -0.03979582991451025, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.03979582991451025, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012293282896280289 + ], + "opposite_logprob_tokens": [ + -5.387293338775635 + ], + "logprob_reward": -0.012293282896280289, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.012293282896280289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24539047479629517, + 0.17460136115550995, + 0.08787678927183151, + 0.06802850216627121, + 0.0921546220779419, + 0.05154034495353699, + 0.043253008276224136, + 0.048919644206762314, + 0.0650857612490654, + 0.12314943969249725 + ], + "credit_map": [ + 0.03178414702415466, + 0.01416183914989233, + 0.0244259275496006, + 0.0318429209291935, + 0.021810419857501984, + 0.03311654180288315, + 0.05262589454650879, + 0.03562280163168907, + 0.018045641481876373, + 0.006773687899112701, + 0.020229298621416092, + 0.02519727498292923, + 0.02626720257103443, + 0.017967838793992996, + 0.03371937572956085, + 0.026401037350296974, + 0.012908714823424816, + 0.00437513180077076, + 0.013536296784877777, + 0.008872109465301037, + 0.01103669311851263, + 0.012366389855742455, + 0.012865323573350906, + 0.011916130781173706, + 0.012301781214773655, + 0.0037264481652528048, + 0.011864811182022095, + 0.005731032229959965, + 0.006190089508891106, + 0.007316964212805033, + 0.009855080395936966, + 0.011042299680411816, + 0.011034050025045872, + 0.0037618097849190235, + 0.012253524735569954, + 0.010887439362704754, + 0.008166218176484108, + 0.01873609609901905, + 0.015341199934482574, + 0.011974290013313293, + 0.009099684655666351, + 0.002321251668035984, + 0.006978983990848064, + 0.006318656262010336, + 0.0045373509638011456, + 0.0059105511754751205, + 0.008978668600320816, + 0.007395201828330755, + 0.008972521871328354, + 0.002569017466157675, + 0.005137085448950529, + 0.0047498117201030254, + 0.0038357460871338844, + 0.004164897836744785, + 0.006623893044888973, + 0.007200036663562059, + 0.00845196284353733, + 0.002888823626562953, + 0.0047669499181210995, + 0.00482993945479393, + 0.005432228557765484, + 0.008018814958631992, + 0.007774289697408676, + 0.006756635382771492, + 0.012587393634021282, + 0.003133912105113268, + 0.008178550750017166, + 0.006813731975853443, + 0.004787015728652477, + 0.008846212178468704, + 0.009006427600979805, + 0.011732518672943115, + 0.013064938597381115, + 0.005509345792233944, + 0.018875787034630775, + 0.014710702933371067, + 0.016311785206198692, + 0.015069345943629742, + 0.023943085223436356, + 0.01566445082426071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.012293282896280289 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497528240084648 + ], + "opposite_logprob_tokens": [ + -2.9299752712249756 + ], + "logprob_reward": -0.05497528240084648, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497528240084648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23177729547023773, + 0.10726957023143768, + 0.14497160911560059, + 0.11506353318691254, + 0.07216192781925201, + 0.043910134583711624, + 0.06294605880975723, + 0.1073826253414154, + 0.05257449671626091, + 0.06194275617599487 + ], + "credit_map": [ + 0.02762136422097683, + 0.012354935519397259, + 0.032315053045749664, + 0.037430476397275925, + 0.024615023285150528, + 0.03543766587972641, + 0.032661136239767075, + 0.029341667890548706, + 0.010777924209833145, + 0.004287734627723694, + 0.014203295111656189, + 0.019461004063487053, + 0.011520070023834705, + 0.01554191205650568, + 0.01779434084892273, + 0.013683310709893703, + 0.023973925039172173, + 0.007313434034585953, + 0.022094281390309334, + 0.01364914234727621, + 0.01720600575208664, + 0.025525934994220734, + 0.015351665206253529, + 0.01985722966492176, + 0.01942415162920952, + 0.00634297588840127, + 0.01718214899301529, + 0.011370549909770489, + 0.012183375656604767, + 0.02017825096845627, + 0.012722300365567207, + 0.015659790486097336, + 0.00858334545046091, + 0.002669961890205741, + 0.011265730485320091, + 0.007501465268433094, + 0.005972273647785187, + 0.014636751264333725, + 0.010320083238184452, + 0.011212317273020744, + 0.006621103268116713, + 0.002242195652797818, + 0.006687878165394068, + 0.005053476430475712, + 0.004266039934009314, + 0.008364325389266014, + 0.004259809385985136, + 0.00641530891880393, + 0.014572269283235073, + 0.004044766072183847, + 0.008528520353138447, + 0.0051728649996221066, + 0.005342482123523951, + 0.009438294917345047, + 0.006571386940777302, + 0.009275481104850769, + 0.011622765101492405, + 0.0034239140804857016, + 0.009022712707519531, + 0.012458117678761482, + 0.012893137522041798, + 0.018965143710374832, + 0.023975014686584473, + 0.015021824277937412, + 0.006336732301861048, + 0.0020430050790309906, + 0.006945142988115549, + 0.00668293097987771, + 0.007828188128769398, + 0.006189514417201281, + 0.009484181180596352, + 0.00706480722874403, + 0.008579072542488575, + 0.0026291916146874428, + 0.011289162561297417, + 0.006705294828861952, + 0.00578314159065485, + 0.008187818340957165, + 0.0094946613535285, + 0.009274415671825409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.05497528240084648 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628357969224453 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628357969224453, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628357969224453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2076416313648224, + 0.11930108815431595, + 0.11592525988817215, + 0.0950285792350769, + 0.07974930852651596, + 0.0743887722492218, + 0.041226256638765335, + 0.03067929483950138, + 0.050526563078165054, + 0.18553318083286285 + ], + "credit_map": [ + 0.02780299261212349, + 0.009815793484449387, + 0.029413308948278427, + 0.028411230072379112, + 0.021463964134454727, + 0.039488062262535095, + 0.026786116883158684, + 0.02446017786860466, + 0.017392219975590706, + 0.0059802355244755745, + 0.02419731579720974, + 0.012839503586292267, + 0.015596827492117882, + 0.01617431454360485, + 0.013630630448460579, + 0.01349005475640297, + 0.014829225838184357, + 0.005026642698794603, + 0.013532997108995914, + 0.012166953645646572, + 0.014982232823967934, + 0.016625234857201576, + 0.019202983006834984, + 0.019558999687433243, + 0.012737682089209557, + 0.004724991973489523, + 0.013962513767182827, + 0.010206533595919609, + 0.010542985051870346, + 0.01533878967165947, + 0.013878058642148972, + 0.013637031428515911, + 0.009649891406297684, + 0.002980817575007677, + 0.011944678612053394, + 0.006765562109649181, + 0.009066293016076088, + 0.014075662940740585, + 0.01245881337672472, + 0.012807593680918217, + 0.02005745656788349, + 0.004024244844913483, + 0.00935107097029686, + 0.004859922453761101, + 0.005397124215960503, + 0.0120832035318017, + 0.0067998734302818775, + 0.011815880425274372, + 0.005874843802303076, + 0.0022758750710636377, + 0.005410688929259777, + 0.005863001570105553, + 0.005101153161376715, + 0.003751463023945689, + 0.006657973863184452, + 0.0062912628054618835, + 0.0038607707247138023, + 0.001168489339761436, + 0.006090446375310421, + 0.0024504917673766613, + 0.0037194888573139906, + 0.004636602476239204, + 0.004011351149529219, + 0.004741658456623554, + 0.006255574058741331, + 0.00248486315831542, + 0.009194417856633663, + 0.004897468723356724, + 0.00465422123670578, + 0.006261749658733606, + 0.00766783906146884, + 0.009110434912145138, + 0.016443829983472824, + 0.008167007006704807, + 0.02124943770468235, + 0.02660413086414337, + 0.034478578716516495, + 0.02196311019361019, + 0.03446695953607559, + 0.02216014824807644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.007628357969224453 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205097213387489 + ], + "opposite_logprob_tokens": [ + -2.8120510578155518 + ], + "logprob_reward": -0.06205097213387489, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205097213387489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21387626230716705, + 0.10686583817005157, + 0.09454041719436646, + 0.17127250134944916, + 0.08582112938165665, + 0.1172056719660759, + 0.05659838393330574, + 0.04757813364267349, + 0.034347858279943466, + 0.07189378887414932 + ], + "credit_map": [ + 0.031957294791936874, + 0.011893972754478455, + 0.03916288912296295, + 0.021177874878048897, + 0.021323537454009056, + 0.030736684799194336, + 0.025986159220337868, + 0.03163783997297287, + 0.014433966018259525, + 0.005171256605535746, + 0.013604172505438328, + 0.016665179282426834, + 0.012872828170657158, + 0.01601797714829445, + 0.01386578194797039, + 0.014234669506549835, + 0.013054981827735901, + 0.006853913888335228, + 0.01303312461823225, + 0.013391617685556412, + 0.009781657718122005, + 0.009456886909902096, + 0.013443706557154655, + 0.015524535439908504, + 0.031976133584976196, + 0.010582038201391697, + 0.02338368073105812, + 0.02063373662531376, + 0.014717561192810535, + 0.020867355167865753, + 0.024723464623093605, + 0.024388516321778297, + 0.014405418187379837, + 0.004392544273287058, + 0.012934396043419838, + 0.00928844790905714, + 0.00842753704637289, + 0.011082105338573456, + 0.011203726753592491, + 0.014086958952248096, + 0.03242109715938568, + 0.007452238351106644, + 0.015086108818650246, + 0.008741911500692368, + 0.008739852346479893, + 0.012334548868238926, + 0.010419368743896484, + 0.022010549902915955, + 0.011328386142849922, + 0.003495939774438739, + 0.00819180253893137, + 0.007412686478346586, + 0.004035593010485172, + 0.005533527582883835, + 0.0075841122306883335, + 0.009016335941851139, + 0.00665205018594861, + 0.002397604053840041, + 0.00782601349055767, + 0.005686620716005564, + 0.0037619394715875387, + 0.007117952685803175, + 0.0064298175275325775, + 0.007706135511398315, + 0.0034539764747023582, + 0.0015490198275074363, + 0.00401688227429986, + 0.003177063539624214, + 0.0033905671443790197, + 0.007308052387088537, + 0.005657656583935022, + 0.00579464016482234, + 0.007483913097530603, + 0.0026593059301376343, + 0.013916082680225372, + 0.010780448094010353, + 0.006220406852662563, + 0.009120711125433445, + 0.009438837878406048, + 0.012274082750082016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.06205097213387489 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062031254172325134 + ], + "opposite_logprob_tokens": [ + -2.8120312690734863 + ], + "logprob_reward": -0.062031254172325134, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062031254172325134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29922550916671753, + 0.15033450722694397, + 0.13162119686603546, + 0.05981811508536339, + 0.04876650497317314, + 0.05339536815881729, + 0.08857876807451248, + 0.02242196723818779, + 0.055129505693912506, + 0.09070857614278793 + ], + "credit_map": [ + 0.03769230842590332, + 0.016895610839128494, + 0.04750403016805649, + 0.04549337923526764, + 0.02816419117152691, + 0.03461560606956482, + 0.048054080456495285, + 0.04080629721283913, + 0.020774787291884422, + 0.007059046998620033, + 0.033074215054512024, + 0.01843627169728279, + 0.01357210986316204, + 0.015219438821077347, + 0.01994287595152855, + 0.022255757823586464, + 0.020024072378873825, + 0.006431116256862879, + 0.023862827569246292, + 0.011068745516240597, + 0.017102213576436043, + 0.016756510362029076, + 0.01582767814397812, + 0.02054804004728794, + 0.011331661604344845, + 0.003718140535056591, + 0.008052115328609943, + 0.005178023595362902, + 0.006082364358007908, + 0.010174666531383991, + 0.0061258901841938496, + 0.009155252017080784, + 0.006310691125690937, + 0.0017384873935952783, + 0.008242219686508179, + 0.004216277971863747, + 0.003987301606684923, + 0.010210503824055195, + 0.006910710595548153, + 0.007150312885642052, + 0.006961399223655462, + 0.002423491096124053, + 0.009123730473220348, + 0.006156406830996275, + 0.005189703311771154, + 0.007455761078745127, + 0.007249367423355579, + 0.0088355066254735, + 0.011362307704985142, + 0.00309654138982296, + 0.01571049913764, + 0.00900187436491251, + 0.013255366124212742, + 0.009689603932201862, + 0.012358256615698338, + 0.014104319736361504, + 0.003078893292695284, + 0.0012214401504024863, + 0.0033586216159164906, + 0.0020718544255942106, + 0.0034150255378335714, + 0.002753518521785736, + 0.0031380008440464735, + 0.0033846122678369284, + 0.007954791188240051, + 0.002331729279831052, + 0.006604020018130541, + 0.00493056233972311, + 0.005369555205106735, + 0.008707753382623196, + 0.007723814807832241, + 0.01150728389620781, + 0.01153589878231287, + 0.003778439946472645, + 0.014704602770507336, + 0.01199126336723566, + 0.009088532999157906, + 0.010862607508897781, + 0.013072382658720016, + 0.015674851834774017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.062031254172325134 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, horizontal, medium shot, shot from behind, captures a man with slightly disheveled gray hair, wearing a dark brown tweed jacket and a beige wool sweater. He sits alone in a high-backed leather armchair, his back to the camera, facing a large arched window. Outside, dusk falls, and a gentle rain silently drizzles across the glass. His study is surrounded by towering, dark wooden bookshelves, crammed with ancient books of varying thickness, bound in leather or cloth. Some books are even haphazardly stacked on the floor and on the corners of the table. The man leans forward slightly, his elbows resting on his knees, his hands clasped before him, his head slightly lowered, his entire body almost motionless, completely absorbed in his thoughts. The cool twilight filtering in from the window serves as the primary light source, softly outlining his shoulders and head. Farther back, a small desk lamp casts a warm, orange glow, illuminating the gilded lettering on the book spines. The entire scene is dominated by deep wood tones and cool blue, dotted with warm lights, creating an academic atmosphere of extreme tranquility, concentration, contemplation, and a hint of melancholy.", + "seed": 28270, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_14/index_40_seed_28270_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05104600489139557, + "vlm_logprob": -0.05104600489139557, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.05104600489139557, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014373810961842537 + ], + "opposite_logprob_tokens": [ + -5.014373779296875 + ], + "logprob_reward": -0.014373810961842537, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.014373810961842537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09883470088243484, + 0.11542314291000366, + 0.05684937164187431, + 0.0540197491645813, + 0.10368648916482925, + 0.07980547100305557, + 0.09655372053384781, + 0.14164911210536957, + 0.18355509638786316, + 0.06962315738201141 + ], + "credit_map": [ + 0.005996947176754475, + 0.007799541112035513, + 0.008621862158179283, + 0.014745925553143024, + 0.010969131253659725, + 0.009041317738592625, + 0.021471664309501648, + 0.020188303664326668, + 0.009004856459796429, + 0.009911859408020973, + 0.011841789819300175, + 0.019915543496608734, + 0.013608074747025967, + 0.018484121188521385, + 0.017986221238970757, + 0.01467067375779152, + 0.004723906982690096, + 0.003421040251851082, + 0.005264125764369965, + 0.010172771289944649, + 0.008336110040545464, + 0.00417653052136302, + 0.013309869915246964, + 0.007445013150572777, + 0.004359349142760038, + 0.0042702918872237206, + 0.006399826612323523, + 0.008849598467350006, + 0.007164819166064262, + 0.0045060510747134686, + 0.010168463923037052, + 0.008301344700157642, + 0.009220507927238941, + 0.007115654647350311, + 0.013909798115491867, + 0.015434308908879757, + 0.014991496689617634, + 0.016317661851644516, + 0.015530685894191265, + 0.011166369542479515, + 0.0047267284244298935, + 0.006570412311702967, + 0.012436418794095516, + 0.010522265918552876, + 0.0175605658441782, + 0.008322841487824917, + 0.012687994167208672, + 0.0069782426580786705, + 0.008825136348605156, + 0.008080928586423397, + 0.008925786241889, + 0.01267920807003975, + 0.02169664017856121, + 0.010842550545930862, + 0.014672362245619297, + 0.010831112042069435, + 0.013311590999364853, + 0.01001998782157898, + 0.02045862376689911, + 0.02736016735434532, + 0.024756602942943573, + 0.012670060619711876, + 0.019945595413446426, + 0.013126467354595661, + 0.018281595781445503, + 0.013485359959304333, + 0.03206605091691017, + 0.024938585236668587, + 0.03232239559292793, + 0.022077539935708046, + 0.025837503373622894, + 0.014546060003340244, + 0.012730035930871964, + 0.006351142190396786, + 0.006303077097982168, + 0.0066197593696415424, + 0.011780316941440105, + 0.00871867872774601, + 0.010216951370239258, + 0.006903196219354868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.014373810961842537 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295033991336823 + ], + "opposite_logprob_tokens": [ + -2.237950325012207 + ], + "logprob_reward": -0.11295033991336823, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295033991336823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15305742621421814, + 0.08145200461149216, + 0.057772889733314514, + 0.053434066474437714, + 0.14759361743927002, + 0.08689650893211365, + 0.09373094141483307, + 0.15251390635967255, + 0.13452987372875214, + 0.03901880234479904 + ], + "credit_map": [ + 0.012881220318377018, + 0.010505342856049538, + 0.020164914429187775, + 0.023076511919498444, + 0.019325321540236473, + 0.02196977473795414, + 0.024149490520358086, + 0.02098485454916954, + 0.005832422990351915, + 0.0058054374530911446, + 0.008346145041286945, + 0.013758190907537937, + 0.012601228430867195, + 0.009996656328439713, + 0.01427573710680008, + 0.010836184024810791, + 0.003673950908705592, + 0.0035887621343135834, + 0.005867077503353357, + 0.007615054026246071, + 0.012021762318909168, + 0.007114686071872711, + 0.011363581754267216, + 0.00652801850810647, + 0.004849506542086601, + 0.003713303478434682, + 0.006590334698557854, + 0.006319806445389986, + 0.008875825442373753, + 0.005235753022134304, + 0.011006886139512062, + 0.006842650473117828, + 0.009655077941715717, + 0.010213227942585945, + 0.020939549431204796, + 0.01936214230954647, + 0.020143212750554085, + 0.02017471194267273, + 0.028658494353294373, + 0.018447211012244225, + 0.006174447014927864, + 0.007109487894922495, + 0.012535089626908302, + 0.009245012886822224, + 0.019780660048127174, + 0.00999430101364851, + 0.014739862643182278, + 0.007317648734897375, + 0.010701851919293404, + 0.007051364053040743, + 0.009861169382929802, + 0.01164313405752182, + 0.022299597039818764, + 0.012148491106927395, + 0.012328483164310455, + 0.007696849759668112, + 0.015715474262833595, + 0.011855602264404297, + 0.014698917046189308, + 0.021633069962263107, + 0.03664826974272728, + 0.02074861153960228, + 0.019351104274392128, + 0.011862856335937977, + 0.01551680639386177, + 0.009718554094433784, + 0.020309165120124817, + 0.01759721152484417, + 0.022711636498570442, + 0.01688234880566597, + 0.01917896792292595, + 0.012615183368325233, + 0.005903608165681362, + 0.002833268139511347, + 0.004525987431406975, + 0.006645267363637686, + 0.005177382379770279, + 0.003613439155742526, + 0.006613117177039385, + 0.0037067297380417585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.11295033991336823 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539708986878395 + ], + "opposite_logprob_tokens": [ + -4.387539863586426 + ], + "logprob_reward": -0.012539708986878395, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539708986878395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050622906535863876, + 0.09709226340055466, + 0.06438104808330536, + 0.04027437046170235, + 0.38917720317840576, + 0.03753819316625595, + 0.03619717061519623, + 0.06271715462207794, + 0.10868260264396667, + 0.11331704258918762 + ], + "credit_map": [ + 0.003702073823660612, + 0.0038665253669023514, + 0.005419743712991476, + 0.006553789600729942, + 0.005592976231127977, + 0.005754332523792982, + 0.010020572692155838, + 0.009712889790534973, + 0.005368772428482771, + 0.006372353993356228, + 0.011067281477153301, + 0.019462384283542633, + 0.013007293455302715, + 0.008404937572777271, + 0.019398987293243408, + 0.014010252431035042, + 0.004200301598757505, + 0.004192013759166002, + 0.006282408256083727, + 0.009487929753959179, + 0.013802741654217243, + 0.006514449138194323, + 0.012533389031887054, + 0.007367810700088739, + 0.0036429984029382467, + 0.0024279183708131313, + 0.003194875782355666, + 0.0055657499469816685, + 0.007189162541180849, + 0.0026895226910710335, + 0.010231194086372852, + 0.00533294677734375, + 0.025530356913805008, + 0.02709520235657692, + 0.04645441845059395, + 0.05569728463888168, + 0.0621141642332077, + 0.05607346445322037, + 0.06901311874389648, + 0.04719918221235275, + 0.0025094386655837297, + 0.00259003066457808, + 0.007385327015072107, + 0.004602115601301193, + 0.007890080101788044, + 0.004040171857923269, + 0.005477777682244778, + 0.0030432529747486115, + 0.003187253838405013, + 0.002599874511361122, + 0.003199364058673382, + 0.003987644333392382, + 0.01005206722766161, + 0.004569802898913622, + 0.0051854457706213, + 0.003415717976167798, + 0.011646760627627373, + 0.0053433384746313095, + 0.0070433239452540874, + 0.010465403087437153, + 0.007279046345502138, + 0.006170789245516062, + 0.009409341961145401, + 0.005359153728932142, + 0.016843188554048538, + 0.008779489435255527, + 0.014941285364329815, + 0.012571325525641441, + 0.016260890290141106, + 0.01596270315349102, + 0.01533723808825016, + 0.00798648502677679, + 0.013236726634204388, + 0.008905471302568913, + 0.012202603742480278, + 0.00943217147141695, + 0.014959893189370632, + 0.020893530920147896, + 0.019874207675457, + 0.013812441378831863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.012539708986878395 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348400861024857 + ], + "opposite_logprob_tokens": [ + -3.651348352432251 + ], + "logprob_reward": -0.026348400861024857, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348400861024857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19056524336338043, + 0.30666935443878174, + 0.10903707146644592, + 0.08705011755228043, + 0.07728319615125656, + 0.042132437229156494, + 0.024703865870833397, + 0.028221191838383675, + 0.056960221379995346, + 0.0773773118853569 + ], + "credit_map": [ + 0.01942385919392109, + 0.015931658446788788, + 0.02045578882098198, + 0.025664230808615685, + 0.021462418138980865, + 0.022817201912403107, + 0.03374844044446945, + 0.031061628833413124, + 0.0204845629632473, + 0.036841247230768204, + 0.024728339165449142, + 0.05293196067214012, + 0.037941981106996536, + 0.027054840698838234, + 0.06431443244218826, + 0.042371973395347595, + 0.010982168838381767, + 0.011057056486606598, + 0.014169968664646149, + 0.01810424029827118, + 0.013211998157203197, + 0.009552466683089733, + 0.01727876625955105, + 0.0146804079413414, + 0.011048296466469765, + 0.00531464209780097, + 0.01101148221641779, + 0.010508229956030846, + 0.009557327255606651, + 0.008955645374953747, + 0.015015705488622189, + 0.015638791024684906, + 0.005481763277202845, + 0.004627848509699106, + 0.014590393751859665, + 0.012751891277730465, + 0.008523530326783657, + 0.00856956746429205, + 0.014142013154923916, + 0.008596185594797134, + 0.0026314291171729565, + 0.002652308437973261, + 0.008705483749508858, + 0.0044027529656887054, + 0.007565416395664215, + 0.006254931446164846, + 0.007194323930889368, + 0.002725787227973342, + 0.00207812013104558, + 0.0015286823036149144, + 0.00533676240593195, + 0.002757252659648657, + 0.0034599376376718283, + 0.002519422210752964, + 0.004530895967036486, + 0.002492789179086685, + 0.004003453534096479, + 0.001967255026102066, + 0.0035667074844241142, + 0.004170756321400404, + 0.0032378945033997297, + 0.0033723050728440285, + 0.004311908967792988, + 0.003590908832848072, + 0.006648471113294363, + 0.004221342504024506, + 0.008039608597755432, + 0.008302310481667519, + 0.007927760481834412, + 0.008250725455582142, + 0.008314663544297218, + 0.005255329422652721, + 0.007516510784626007, + 0.00500227278098464, + 0.008652789518237114, + 0.007116100285202265, + 0.010583820752799511, + 0.015919838100671768, + 0.013595599681138992, + 0.008990377187728882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.026348400861024857 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901776373386383 + ], + "opposite_logprob_tokens": [ + -2.464017868041992 + ], + "logprob_reward": -0.08901776373386383, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901776373386383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12204910069704056, + 0.11683231592178345, + 0.11968644708395004, + 0.0756913274526596, + 0.3208198845386505, + 0.08061644434928894, + 0.040818192064762115, + 0.023796986788511276, + 0.04093314707279205, + 0.05875613912940025 + ], + "credit_map": [ + 0.01293760072439909, + 0.008591675199568272, + 0.016414785757660866, + 0.014740563929080963, + 0.01275341585278511, + 0.016015971079468727, + 0.022442134097218513, + 0.018152957782149315, + 0.00887320563197136, + 0.010721406899392605, + 0.014813728630542755, + 0.02610161527991295, + 0.00997451413422823, + 0.012481077574193478, + 0.01742532104253769, + 0.016441455110907555, + 0.010914317332208157, + 0.006579208187758923, + 0.014181751757860184, + 0.020878689363598824, + 0.01880054920911789, + 0.010034462437033653, + 0.02378842607140541, + 0.014509053900837898, + 0.005378093104809523, + 0.006050840951502323, + 0.006482102908194065, + 0.008703895844519138, + 0.015369724482297897, + 0.007000762969255447, + 0.0172779131680727, + 0.009427995420992374, + 0.02186294086277485, + 0.021542126312851906, + 0.03976057097315788, + 0.049070581793785095, + 0.047350361943244934, + 0.04355715587735176, + 0.05753256008028984, + 0.040143609046936035, + 0.00967797264456749, + 0.005630314815789461, + 0.0193984042853117, + 0.009130205027759075, + 0.007680954411625862, + 0.00850820541381836, + 0.012270444072782993, + 0.008319949731230736, + 0.003400991437956691, + 0.003763890825212002, + 0.006363066844642162, + 0.006792648695409298, + 0.005523348227143288, + 0.004869162105023861, + 0.005662544164806604, + 0.004442543722689152, + 0.0017855055630207062, + 0.0013435104629024863, + 0.0037081053014844656, + 0.00536477193236351, + 0.003413059748709202, + 0.0021133909467607737, + 0.0035494593903422356, + 0.002519184025004506, + 0.004478634335100651, + 0.003078488865867257, + 0.0067764464765787125, + 0.004206900950521231, + 0.006991449743509293, + 0.006291243713349104, + 0.0058639170601964, + 0.00324607128277421, + 0.006141225341707468, + 0.005146430339664221, + 0.007029889617115259, + 0.005296159535646439, + 0.013436710461974144, + 0.00842300709336996, + 0.008412198163568974, + 0.0048705218359827995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.08901776373386383 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, horizontal, medium shot, shot from behind, captures a man with slightly disheveled gray hair, wearing a dark brown tweed jacket and a beige wool sweater. He sits alone in a high-backed leather armchair, his back to the camera, facing a large arched window. Outside, dusk falls, and a gentle rain silently drizzles across the glass. His study is surrounded by towering, dark wooden bookshelves, crammed with ancient books of varying thickness, bound in leather or cloth. Some books are even haphazardly stacked on the floor and on the corners of the table. The man leans forward slightly, his elbows resting on his knees, his hands clasped before him, his head slightly lowered, his entire body almost motionless, completely absorbed in his thoughts. The cool twilight filtering in from the window serves as the primary light source, softly outlining his shoulders and head. Farther back, a small desk lamp casts a warm, orange glow, illuminating the gilded lettering on the book spines. The entire scene is dominated by deep wood tones and cool blue, dotted with warm lights, creating an academic atmosphere of extreme tranquility, concentration, contemplation, and a hint of melancholy.", + "seed": 28271, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_14/index_40_seed_28271_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.049247647263109685, + "vlm_logprob": -0.049247647263109685, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.049247647263109685, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015263068489730358 + ], + "opposite_logprob_tokens": [ + -4.890263080596924 + ], + "logprob_reward": -0.015263068489730358, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.015263068489730358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09902939945459366, + 0.13121773302555084, + 0.06829533725976944, + 0.08098810166120529, + 0.03220813348889351, + 0.06139688938856125, + 0.2359326332807541, + 0.1343051940202713, + 0.06309188902378082, + 0.09353466331958771 + ], + "credit_map": [ + 0.010898573324084282, + 0.006682648789137602, + 0.011116161942481995, + 0.010697039775550365, + 0.006169109605252743, + 0.016965823248028755, + 0.023897267878055573, + 0.012602764181792736, + 0.008970571681857109, + 0.012329154647886753, + 0.018033703789114952, + 0.027910016477108, + 0.007093074265867472, + 0.019362248480319977, + 0.022871559485793114, + 0.01464739628136158, + 0.005410614889115095, + 0.00595121318474412, + 0.005339155439287424, + 0.008739938028156757, + 0.002900807885453105, + 0.009323877282440662, + 0.02368594892323017, + 0.006943780928850174, + 0.008910183794796467, + 0.007622759323567152, + 0.007470821961760521, + 0.01186707615852356, + 0.0035305877681821585, + 0.010865181684494019, + 0.020823048427700996, + 0.009898442775011063, + 0.002116954419761896, + 0.003423892194405198, + 0.0038503746036440134, + 0.007930588908493519, + 0.0015085480408743024, + 0.003596940077841282, + 0.006756828632205725, + 0.0030240043997764587, + 0.005232419352978468, + 0.007002717349678278, + 0.0062584299594163895, + 0.009365350939333439, + 0.004576091654598713, + 0.009304016828536987, + 0.011172990314662457, + 0.008484870195388794, + 0.04675213247537613, + 0.012969041243195534, + 0.015267224982380867, + 0.01954929530620575, + 0.01041876059025526, + 0.02332158572971821, + 0.08062520623207092, + 0.027029376477003098, + 0.019195912405848503, + 0.013335918076336384, + 0.019719956442713737, + 0.020759643986821175, + 0.007941302843391895, + 0.016076788306236267, + 0.024493351578712463, + 0.012782310135662556, + 0.011113193817436695, + 0.0038216570392251015, + 0.005850070621818304, + 0.009191819466650486, + 0.0029543747659772635, + 0.003371327882632613, + 0.019101936370134354, + 0.0076875099912285805, + 0.010300344787538052, + 0.005614728666841984, + 0.009242989122867584, + 0.012415740638971329, + 0.004901153966784477, + 0.005921881180256605, + 0.030176561325788498, + 0.014961257576942444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.015263068489730358 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032077878713608 + ], + "opposite_logprob_tokens": [ + -2.350320816040039 + ], + "logprob_reward": -0.10032077878713608, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032077878713608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22393232583999634, + 0.08701855689287186, + 0.06931405514478683, + 0.10059590637683868, + 0.04079315811395645, + 0.07477633655071259, + 0.11826179921627045, + 0.1310754418373108, + 0.08130380511283875, + 0.07292862981557846 + ], + "credit_map": [ + 0.03359225392341614, + 0.01813449151813984, + 0.02160704880952835, + 0.0406881682574749, + 0.015157366171479225, + 0.038445595651865005, + 0.027658501639962196, + 0.028648896142840385, + 0.006981012411415577, + 0.00888687465339899, + 0.011235078796744347, + 0.012333300895988941, + 0.00557790556922555, + 0.013955104164779186, + 0.019682036712765694, + 0.008367247879505157, + 0.007685377262532711, + 0.00474134087562561, + 0.00546943349763751, + 0.010842016898095608, + 0.0041124396957457066, + 0.014703026041388512, + 0.011340407654643059, + 0.010420012287795544, + 0.00672077527269721, + 0.010792827233672142, + 0.010470043867826462, + 0.023298189043998718, + 0.005105979274958372, + 0.014429379254579544, + 0.018982145935297012, + 0.010796568356454372, + 0.00235561840236187, + 0.0030044002924114466, + 0.003939271438866854, + 0.007377649191766977, + 0.002304419409483671, + 0.007750923279672861, + 0.009295200929045677, + 0.004765675403177738, + 0.006462045479565859, + 0.008070711977779865, + 0.008427810855209827, + 0.008739563636481762, + 0.004987239837646484, + 0.011395063251256943, + 0.017733732238411903, + 0.008960172533988953, + 0.01338033378124237, + 0.008950254879891872, + 0.010164573788642883, + 0.01522049494087696, + 0.006209938786923885, + 0.014726702123880386, + 0.03273788467049599, + 0.016871612519025803, + 0.021411867812275887, + 0.012907969765365124, + 0.018243899568915367, + 0.019585875794291496, + 0.007083486765623093, + 0.013310051523149014, + 0.02518918737769127, + 0.01334311068058014, + 0.01597015932202339, + 0.004923065192997456, + 0.007623546756803989, + 0.007974648848176003, + 0.004628588445484638, + 0.004663773346692324, + 0.025825416669249535, + 0.009694612585008144, + 0.007120391353964806, + 0.006728112231940031, + 0.006563230883330107, + 0.01449071429669857, + 0.006209954619407654, + 0.006996790412813425, + 0.012978209182620049, + 0.011841226369142532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.10032077878713608 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978388637304306 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.00978388637304306, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978388637304306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31537380814552307, + 0.061455365270376205, + 0.06358403712511063, + 0.08539008349180222, + 0.05757065862417221, + 0.1514413058757782, + 0.0646805614233017, + 0.07159259915351868, + 0.0324338935315609, + 0.0964776948094368 + ], + "credit_map": [ + 0.029228784143924713, + 0.015539002604782581, + 0.019277632236480713, + 0.02797957882285118, + 0.011237534694373608, + 0.04357080161571503, + 0.12455645948648453, + 0.04398399963974953, + 0.00528754061087966, + 0.007261928636580706, + 0.007911760360002518, + 0.009460588917136192, + 0.005555180367082357, + 0.008316206745803356, + 0.012220251373946667, + 0.005441906861960888, + 0.0038251199293881655, + 0.006392129231244326, + 0.005102284252643585, + 0.006204887758940458, + 0.003835159819573164, + 0.01444376539438963, + 0.015336456708610058, + 0.008444237522780895, + 0.0074045113287866116, + 0.008685803972184658, + 0.007344971410930157, + 0.013507706113159657, + 0.0043444037437438965, + 0.011728978715837002, + 0.023257868364453316, + 0.009115836583077908, + 0.003567206673324108, + 0.004822163376957178, + 0.004953149240463972, + 0.014055580832064152, + 0.0029504788108170033, + 0.008052580989897251, + 0.013638440519571304, + 0.005531058646738529, + 0.012609409168362617, + 0.013832147233188152, + 0.014023493975400925, + 0.020027263090014458, + 0.007531852461397648, + 0.02050633355975151, + 0.0440523661673069, + 0.01885843835771084, + 0.007920983247458935, + 0.0051657832227647305, + 0.006628090515732765, + 0.011593490839004517, + 0.004156972747296095, + 0.00830941740423441, + 0.01278784591704607, + 0.00811797846108675, + 0.011591761372983456, + 0.007784485351294279, + 0.009009485132992268, + 0.010092953220009804, + 0.00383018865250051, + 0.008275462314486504, + 0.01380111649632454, + 0.007207150105386972, + 0.004563570022583008, + 0.002110038883984089, + 0.004181514028459787, + 0.006124725565314293, + 0.0020939067471772432, + 0.00308501860126853, + 0.006515704095363617, + 0.003759414190426469, + 0.008077767677605152, + 0.006601438857614994, + 0.01018384750932455, + 0.013983973301947117, + 0.0061885882169008255, + 0.006789704319089651, + 0.02859131246805191, + 0.01606106571853161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.00978388637304306 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058347500860691 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058347500860691, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058347500860691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3778325319290161, + 0.08625611662864685, + 0.03656693920493126, + 0.0493227019906044, + 0.02945401333272457, + 0.09051824361085892, + 0.22631339728832245, + 0.04213818162679672, + 0.01616252027451992, + 0.0454353429377079 + ], + "credit_map": [ + 0.030544644221663475, + 0.014136356301605701, + 0.023173637688159943, + 0.03795624151825905, + 0.014435076154768467, + 0.051068082451820374, + 0.15565720200538635, + 0.05086129158735275, + 0.005219282116740942, + 0.009554284624755383, + 0.008625945076346397, + 0.015694206580519676, + 0.005489070899784565, + 0.01510639674961567, + 0.01692197285592556, + 0.009644960053265095, + 0.0026830467395484447, + 0.002994680078700185, + 0.003702778834849596, + 0.005812251009047031, + 0.003114014631137252, + 0.006845253054052591, + 0.006464395206421614, + 0.004950521048158407, + 0.0032975568901747465, + 0.003871689550578594, + 0.004799782298505306, + 0.006568580865859985, + 0.003213317599147558, + 0.011282586492598057, + 0.011074066162109375, + 0.005215120036154985, + 0.001980094937607646, + 0.002398452255874872, + 0.0029844408854842186, + 0.00448349816724658, + 0.0018223387887701392, + 0.0053732688538730145, + 0.007404612377285957, + 0.0030073069501668215, + 0.007529170718044043, + 0.008191048167645931, + 0.008106241934001446, + 0.017969615757465363, + 0.0053393044508993626, + 0.011662427335977554, + 0.019374476745724678, + 0.012345956638455391, + 0.046694133430719376, + 0.012947323732078075, + 0.013523834757506847, + 0.01612236350774765, + 0.00822692085057497, + 0.019045192748308182, + 0.08355691283941269, + 0.02619672194123268, + 0.004476167727261782, + 0.005719929002225399, + 0.0052013099193573, + 0.0071541909128427505, + 0.002322537824511528, + 0.005759216845035553, + 0.007505776826292276, + 0.0039990562945604324, + 0.0016745510511100292, + 0.001226027961820364, + 0.0025459507014602423, + 0.0024000639095902443, + 0.0009086932404898107, + 0.0014142014551907778, + 0.003717777319252491, + 0.002275254810228944, + 0.004837758839130402, + 0.0028118002228438854, + 0.0057700579054653645, + 0.010464741848409176, + 0.0028971086721867323, + 0.004510023631155491, + 0.008557776920497417, + 0.005586072336882353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.02058347500860691 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, horizontal, medium shot taken from behind the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark brown tweed jacket and a beige wool sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man sitting alone in a high-backed leather armchair with his back to the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool twilight from a large arched window and a warm orange desk lamp in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the bookshelves filled with ancient books, some of which are stacked on the floor and table corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028702765703201 + ], + "opposite_logprob_tokens": [ + -2.3502869606018066 + ], + "logprob_reward": -0.10028702765703201, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028702765703201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5941377878189087, + 0.040827587246894836, + 0.05690266191959381, + 0.039234988391399384, + 0.02532295323908329, + 0.055888235569000244, + 0.055649060755968094, + 0.038987793028354645, + 0.038676321506500244, + 0.05437260866165161 + ], + "credit_map": [ + 0.05971440672874451, + 0.022456366568803787, + 0.03294428810477257, + 0.05232022702693939, + 0.020245861262083054, + 0.08119070529937744, + 0.24293752014636993, + 0.0823284238576889, + 0.002987317740917206, + 0.004145897924900055, + 0.004260542802512646, + 0.00648622028529644, + 0.0029508336447179317, + 0.005105736665427685, + 0.009533791802823544, + 0.005357248708605766, + 0.00474349083378911, + 0.0034353327937424183, + 0.004461035598069429, + 0.005766159854829311, + 0.00283916131593287, + 0.014101123437285423, + 0.012379316613078117, + 0.009177039377391338, + 0.0020062043331563473, + 0.0025771749205887318, + 0.003665090538561344, + 0.004451355431228876, + 0.0021887377370148897, + 0.010876286774873734, + 0.008173927664756775, + 0.005296213086694479, + 0.0013444260694086552, + 0.003142333123832941, + 0.002508478704839945, + 0.0037483477499336004, + 0.001996516715735197, + 0.005210785195231438, + 0.00385493878275156, + 0.0035171271301805973, + 0.005166870076209307, + 0.005929327104240656, + 0.0045655956491827965, + 0.0111218998208642, + 0.0034874468110501766, + 0.006252177990972996, + 0.01300769578665495, + 0.006357224192470312, + 0.011454230174422264, + 0.003461680142208934, + 0.0039027174934744835, + 0.0046637337654829025, + 0.0023990324698388577, + 0.005220166873186827, + 0.017533984035253525, + 0.007013517897576094, + 0.0042072259820997715, + 0.004253263119608164, + 0.005167939700186253, + 0.006519102491438389, + 0.002718031406402588, + 0.0053812358528375626, + 0.005963497329503298, + 0.004777494817972183, + 0.003848495427519083, + 0.002807430224493146, + 0.005243973340839148, + 0.008229375816881657, + 0.0025929315015673637, + 0.004810061771422625, + 0.005856338422745466, + 0.005287718493491411, + 0.0033926006872206926, + 0.0043044365011155605, + 0.00622635567560792, + 0.01067365799099207, + 0.003829707158729434, + 0.005705773364752531, + 0.012934616766870022, + 0.007305463310331106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.10028702765703201 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_15/index_tensor([2536, 2536])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_15/index_tensor([2536, 2536])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..db7a44c2c9df5f2e37ecb132222eecaa4229d9a9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_15/index_tensor([2536, 2536])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, shot from a level angle, slowly pans forward, focusing on a wrinkled elderly man. Beneath his sparse silver-white hair lie gentle, deep eyes. The wrinkles on his forehead and around his eyes, as deep as the furrows of time, bear witness to the passage of time. He sits serenely in a room bathed in afternoon sunlight, wearing a worn, rough, dark gray sweater. Soft, warm light slants in from the side windows, illuminating the dust particles dancing in the air and casting long, blurred shadows on the wooden bookshelves behind him. Steadily and calmly, he slowly extends his hand, veined and covered in age spots and fine lines, toward the camera. His knuckles are slightly thickened from years of hard work, and the lines on his palm are clearly visible, as if reaching out to the audience on screen or offering a silent message. The entire picture presents a warm and faded retro film texture, with low color saturation and slight film grain. The main colors are amber and dark brown, creating a tranquil, nostalgic and story-filled emotional atmosphere.", + "seed": 13022, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_15/index_2536_seed_13022_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01110949944704771, + "vlm_logprob": -0.01110949944704771, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.01110949944704771, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005341069307178259 + ], + "opposite_logprob_tokens": [ + -6.380341053009033 + ], + "logprob_reward": -0.005341069307178259, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.005341069307178259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07695458084344864, + 0.3017694056034088, + 0.02805970050394535, + 0.06652762740850449, + 0.03879908472299576, + 0.0688619539141655, + 0.0873064249753952, + 0.12705673277378082, + 0.15692684054374695, + 0.04773756116628647 + ], + "credit_map": [ + 0.009354040957987309, + 0.012835720553994179, + 0.008871291764080524, + 0.006372236181050539, + 0.01624613255262375, + 0.005212386138737202, + 0.007489912677556276, + 0.010572865605354309, + 0.12280761450529099, + 0.027028845623135567, + 0.02230621874332428, + 0.008175883442163467, + 0.018104292452335358, + 0.017532048746943474, + 0.017050648108124733, + 0.06876389682292938, + 0.0030064848251640797, + 0.0032709285151213408, + 0.0034146099351346493, + 0.0024270780850201845, + 0.003962972201406956, + 0.003215362783521414, + 0.004530651960521936, + 0.004231615923345089, + 0.009582915343344212, + 0.008273410610854626, + 0.011467383243143559, + 0.005959906615316868, + 0.0077460771426558495, + 0.004981665406376123, + 0.005451607517898083, + 0.01306467317044735, + 0.0039043568540364504, + 0.0034331262577325106, + 0.004166020080447197, + 0.004127284977585077, + 0.005902259144932032, + 0.004604398272931576, + 0.006258765701204538, + 0.0064028785564005375, + 0.006238036323338747, + 0.008669760078191757, + 0.006213576532900333, + 0.006395823787897825, + 0.013390996493399143, + 0.00980345904827118, + 0.008039787411689758, + 0.010110525414347649, + 0.00928187370300293, + 0.011261691339313984, + 0.009386050514876842, + 0.009764507412910461, + 0.01304995734244585, + 0.009569601155817509, + 0.013977712951600552, + 0.011015038937330246, + 0.017244888469576836, + 0.018190210685133934, + 0.013497519306838512, + 0.019011223688721657, + 0.012272809632122517, + 0.012572634033858776, + 0.02041802555322647, + 0.013849439099431038, + 0.02197028324007988, + 0.019285814836621284, + 0.014119983650743961, + 0.014393750578165054, + 0.02062893658876419, + 0.016868237406015396, + 0.030105555430054665, + 0.019554296508431435, + 0.010329820215702057, + 0.005186415743082762, + 0.0033624402713030577, + 0.0035642902366816998, + 0.006196902133524418, + 0.006588699761778116, + 0.0061908625066280365, + 0.00631813844665885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.005341069307178259 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018214697018265724 + ], + "opposite_logprob_tokens": [ + -4.018214702606201 + ], + "logprob_reward": -0.018214697018265724, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018214697018265724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09803469479084015, + 0.38925057649612427, + 0.11432623118162155, + 0.037544842809438705, + 0.03757186606526375, + 0.054678842425346375, + 0.03821471706032753, + 0.1187092736363411, + 0.06366579234600067, + 0.04800320789217949 + ], + "credit_map": [ + 0.016250809654593468, + 0.013878347352147102, + 0.010256604291498661, + 0.0061361175030469894, + 0.019706251099705696, + 0.008310242556035519, + 0.007902347482740879, + 0.015593966469168663, + 0.15893672406673431, + 0.0319267138838768, + 0.02987818233668804, + 0.012915350496768951, + 0.022048350423574448, + 0.022615982219576836, + 0.023801984265446663, + 0.08712725341320038, + 0.012649794109165668, + 0.015511853620409966, + 0.015572021715342999, + 0.010655843652784824, + 0.017842045053839684, + 0.014174415729939938, + 0.011823982931673527, + 0.016096271574497223, + 0.005530815105885267, + 0.003686922136694193, + 0.0050139171071350574, + 0.004009878728538752, + 0.007321497891098261, + 0.0028414896223694086, + 0.004219109658151865, + 0.004921213258057833, + 0.004711959045380354, + 0.0035773697309195995, + 0.003004212863743305, + 0.0033491405192762613, + 0.007728642784059048, + 0.005314390640705824, + 0.004062057938426733, + 0.005824089515954256, + 0.0058888765051960945, + 0.006550686899572611, + 0.004513624124228954, + 0.005207686685025692, + 0.00986112467944622, + 0.006625445559620857, + 0.0071609606966376305, + 0.008870434947311878, + 0.0035029235295951366, + 0.004741746000945568, + 0.004370909184217453, + 0.0049967337399721146, + 0.005743689369410276, + 0.0037152026779949665, + 0.00576033815741539, + 0.00538317346945405, + 0.011945106089115143, + 0.022737938910722733, + 0.013257304206490517, + 0.01739436201751232, + 0.01405410561710596, + 0.009614883922040462, + 0.014233493246138096, + 0.015472069382667542, + 0.009194612503051758, + 0.005399545654654503, + 0.005870102439075708, + 0.010059760883450508, + 0.008542285300791264, + 0.0064309570007026196, + 0.011462457478046417, + 0.006706062238663435, + 0.008320586755871773, + 0.0035934140905737877, + 0.004317880142480135, + 0.00486663356423378, + 0.006076447665691376, + 0.007440652698278427, + 0.005787451285868883, + 0.00760013610124588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 2 + ], + "selected_reward": -0.018214697018265724 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017239484004676342 + ], + "opposite_logprob_tokens": [ + -6.376723766326904 + ], + "logprob_reward": -0.0017239484004676342, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017239484004676342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05174770951271057, + 0.06627315282821655, + 0.06285642832517624, + 0.14474233984947205, + 0.16205741465091705, + 0.12776005268096924, + 0.15944570302963257, + 0.06960388273000717, + 0.0897199958562851, + 0.06579330563545227 + ], + "credit_map": [ + 0.007927941158413887, + 0.009349449537694454, + 0.005660161841660738, + 0.004553619772195816, + 0.009139963425695896, + 0.0031743550207465887, + 0.004314439836889505, + 0.007627779617905617, + 0.024683156982064247, + 0.005125132389366627, + 0.004801258910447359, + 0.0028723275754600763, + 0.005336178932338953, + 0.005121133755892515, + 0.005184269975870848, + 0.013149695470929146, + 0.008336586877703667, + 0.007490446791052818, + 0.006395981181412935, + 0.006308289244771004, + 0.008623701520264149, + 0.004802938550710678, + 0.012003662064671516, + 0.008894827216863632, + 0.021620314568281174, + 0.012610973790287971, + 0.02468138001859188, + 0.014642982743680477, + 0.022001897916197777, + 0.009665139950811863, + 0.020192403346300125, + 0.019327249377965927, + 0.013174929656088352, + 0.0160385575145483, + 0.01304271537810564, + 0.016034448519349098, + 0.02355829067528248, + 0.019639965146780014, + 0.03261086717247963, + 0.02795761451125145, + 0.01217668317258358, + 0.015522906556725502, + 0.012336908839643002, + 0.011510657146573067, + 0.028032217174768448, + 0.012605825439095497, + 0.01815279759466648, + 0.017422057688236237, + 0.014236188493669033, + 0.016287140548229218, + 0.014745168387889862, + 0.026597436517477036, + 0.023773716762661934, + 0.014786546118557453, + 0.027308613061904907, + 0.02171090431511402, + 0.00981223490089178, + 0.009971176274120808, + 0.008124385960400105, + 0.011293009854853153, + 0.007104134187102318, + 0.005939136724919081, + 0.009216711856424809, + 0.008143092505633831, + 0.01402293425053358, + 0.0071111987344920635, + 0.0057046920992434025, + 0.012484992854297161, + 0.012565277516841888, + 0.011715324595570564, + 0.015257470309734344, + 0.010858108289539814, + 0.008979695849120617, + 0.005919142626225948, + 0.0064615122973918915, + 0.007230074144899845, + 0.006789284758269787, + 0.01078714244067669, + 0.008590197190642357, + 0.011036256328225136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.0017239484004676342 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606719382107258 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.01606719382107258, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606719382107258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20253872871398926, + 0.11641549319028854, + 0.08645405620336533, + 0.07305239886045456, + 0.05370020493865013, + 0.08518843352794647, + 0.05787791311740875, + 0.18997955322265625, + 0.08520262688398361, + 0.0495905838906765 + ], + "credit_map": [ + 0.02569357492029667, + 0.042932845652103424, + 0.030193308368325233, + 0.016078781336545944, + 0.023395126685500145, + 0.018240872770547867, + 0.014032606035470963, + 0.03197162598371506, + 0.040297843515872955, + 0.012499540112912655, + 0.009793935343623161, + 0.004689861088991165, + 0.008262564428150654, + 0.009166080504655838, + 0.008291042409837246, + 0.023414619266986847, + 0.008255980908870697, + 0.011744830757379532, + 0.010198017582297325, + 0.007354258559644222, + 0.016223078593611717, + 0.013531623408198357, + 0.007091502659022808, + 0.012054773978888988, + 0.00947161577641964, + 0.00925481878221035, + 0.015467335470020771, + 0.004762082826346159, + 0.010184336453676224, + 0.00471215695142746, + 0.005186252761632204, + 0.014013798907399178, + 0.008665825240314007, + 0.0055824280716478825, + 0.007923773489892483, + 0.00398498447611928, + 0.006840722169727087, + 0.005849743727594614, + 0.006610556971281767, + 0.008242174051702023, + 0.011203146539628506, + 0.012930693104863167, + 0.010124395601451397, + 0.00908683706074953, + 0.01123296283185482, + 0.009197339415550232, + 0.008768710307776928, + 0.01264435239136219, + 0.005200615152716637, + 0.006101436447352171, + 0.004655114375054836, + 0.006313762627542019, + 0.010105358436703682, + 0.012205664999783039, + 0.004343175794929266, + 0.008952787145972252, + 0.02547491155564785, + 0.0352993868291378, + 0.020240236073732376, + 0.03249362111091614, + 0.017083099111914635, + 0.01495500560849905, + 0.023287329822778702, + 0.02114596590399742, + 0.01226048357784748, + 0.009732372127473354, + 0.0074248588643968105, + 0.0076356609351933, + 0.011685817502439022, + 0.010162959806621075, + 0.016193032264709473, + 0.010107440873980522, + 0.0071436818689107895, + 0.00382864847779274, + 0.00449478579685092, + 0.004777878988534212, + 0.004833689890801907, + 0.008962357416749, + 0.006711871828883886, + 0.00883767195045948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.01606719382107258 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200588688254356 + ], + "opposite_logprob_tokens": [ + -4.264200687408447 + ], + "logprob_reward": -0.014200588688254356, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200588688254356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07090886682271957, + 0.09907541424036026, + 0.058451127260923386, + 0.037937916815280914, + 0.04379962757229805, + 0.20956574380397797, + 0.07975920289754868, + 0.17595979571342468, + 0.06792090088129044, + 0.15662142634391785 + ], + "credit_map": [ + 0.006826288532465696, + 0.013109012506902218, + 0.00803319364786148, + 0.007290421053767204, + 0.015324492007493973, + 0.004221659153699875, + 0.006902995053678751, + 0.009200803935527802, + 0.029095875099301338, + 0.012401500716805458, + 0.009525288827717304, + 0.005336053669452667, + 0.01205176766961813, + 0.0062318542040884495, + 0.006914771627634764, + 0.01751829870045185, + 0.013514810241758823, + 0.008311723358929157, + 0.006558775436133146, + 0.004371712915599346, + 0.006617050152271986, + 0.004467064049094915, + 0.004180091433227062, + 0.010429903864860535, + 0.0044507528655231, + 0.0037278346717357635, + 0.0048333643935620785, + 0.0042993915267288685, + 0.006305828224867582, + 0.0024026939645409584, + 0.005392719525843859, + 0.006525331176817417, + 0.0046347021125257015, + 0.005394508130848408, + 0.004963772837072611, + 0.004377407021820545, + 0.008291313424706459, + 0.004537657368928194, + 0.004839963745325804, + 0.006760301999747753, + 0.028945108875632286, + 0.02792091853916645, + 0.02292463928461075, + 0.02306079864501953, + 0.028407637029886246, + 0.02279442548751831, + 0.025768399238586426, + 0.02974381484091282, + 0.00952571164816618, + 0.007687662728130817, + 0.007395998574793339, + 0.010191960260272026, + 0.008044334128499031, + 0.01539603155106306, + 0.010592077858746052, + 0.010925422422587872, + 0.022524487227201462, + 0.032464124262332916, + 0.024405304342508316, + 0.02613263577222824, + 0.018765676766633987, + 0.011142411269247532, + 0.017883555963635445, + 0.022641604766249657, + 0.014874966815114021, + 0.007176869083195925, + 0.0063894446939229965, + 0.007587130647152662, + 0.00678587518632412, + 0.006617526989430189, + 0.009540626779198647, + 0.00894845835864544, + 0.03217258304357529, + 0.012507440522313118, + 0.01588423177599907, + 0.012878531590104103, + 0.016943970695137978, + 0.019695553928613663, + 0.02683078870177269, + 0.019708318635821342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.014200588688254356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a level angle, slowly pans forward, focusing on a wrinkled elderly man. Beneath his sparse silver-white hair lie gentle, deep eyes. The wrinkles on his forehead and around his eyes, as deep as the furrows of time, bear witness to the passage of time. He sits serenely in a room bathed in afternoon sunlight, wearing a worn, rough, dark gray sweater. Soft, warm light slants in from the side windows, illuminating the dust particles dancing in the air and casting long, blurred shadows on the wooden bookshelves behind him. Steadily and calmly, he slowly extends his hand, veined and covered in age spots and fine lines, toward the camera. His knuckles are slightly thickened from years of hard work, and the lines on his palm are clearly visible, as if reaching out to the audience on screen or offering a silent message. The entire picture presents a warm and faded retro film texture, with low color saturation and slight film grain. The main colors are amber and dark brown, creating a tranquil, nostalgic and story-filled emotional atmosphere.", + "seed": 13023, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_15/index_2536_seed_13023_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.029722878709435463, + "vlm_logprob": -0.029722878709435463, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.029722878709435463, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005405808798968792 + ], + "opposite_logprob_tokens": [ + -5.755405902862549 + ], + "logprob_reward": -0.005405808798968792, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.005405808798968792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19228313863277435, + 0.1370207667350769, + 0.04972822964191437, + 0.08324985206127167, + 0.07958173006772995, + 0.07931199669837952, + 0.08775990456342697, + 0.07540556788444519, + 0.10887540876865387, + 0.10678338259458542 + ], + "credit_map": [ + 0.05374976992607117, + 0.021824797615408897, + 0.03265893831849098, + 0.02134627290070057, + 0.012947343289852142, + 0.01493997685611248, + 0.020139513537287712, + 0.01467653177678585, + 0.03990820795297623, + 0.011890925467014313, + 0.018961841240525246, + 0.011924021877348423, + 0.00836362224072218, + 0.019916238263249397, + 0.0157917533069849, + 0.010264167562127113, + 0.007056418806314468, + 0.007913577370345592, + 0.004402979742735624, + 0.0068255201913416386, + 0.00420545507222414, + 0.005786196794360876, + 0.006133737042546272, + 0.007404351606965065, + 0.019644467160105705, + 0.010361350141465664, + 0.016081465408205986, + 0.006956698838621378, + 0.004052877426147461, + 0.004852395970374346, + 0.013941301964223385, + 0.007359304465353489, + 0.018825501203536987, + 0.01275284681469202, + 0.010545804165303707, + 0.008712918497622013, + 0.004937963094562292, + 0.010537420399487019, + 0.006737538147717714, + 0.006531737744808197, + 0.007966297678649426, + 0.02200312539935112, + 0.011100280098617077, + 0.012263358570635319, + 0.007670891005545855, + 0.0057862745597958565, + 0.007512819953262806, + 0.005008955020457506, + 0.008683295920491219, + 0.016357600688934326, + 0.010037575848400593, + 0.014555764384567738, + 0.007659594528377056, + 0.00745274405926466, + 0.017156856134533882, + 0.005856481846421957, + 0.011055966839194298, + 0.008850534446537495, + 0.01251825038343668, + 0.012508600018918514, + 0.005414455663412809, + 0.006215732078999281, + 0.012139285914599895, + 0.006702743470668793, + 0.03090634196996689, + 0.011969367042183876, + 0.03216114267706871, + 0.01273445226252079, + 0.006021635141223669, + 0.005784433335065842, + 0.005135669838637114, + 0.0041623711585998535, + 0.030438149347901344, + 0.0056928424164652824, + 0.0392800010740757, + 0.0163316261023283, + 0.0032045904081314802, + 0.0046902867034077644, + 0.003751654177904129, + 0.003394249128177762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.005405808798968792 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330845594406128 + ], + "opposite_logprob_tokens": [ + -3.773308515548706 + ], + "logprob_reward": -0.02330845594406128, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330845594406128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2575620412826538, + 0.05406099930405617, + 0.25011926889419556, + 0.07980727404356003, + 0.045282989740371704, + 0.0342111811041832, + 0.06251592189073563, + 0.07384849339723587, + 0.07539965957403183, + 0.06719216704368591 + ], + "credit_map": [ + 0.02115611918270588, + 0.02844829484820366, + 0.03206536918878555, + 0.03489920496940613, + 0.022045334801077843, + 0.02367967925965786, + 0.04681583493947983, + 0.04845219478011131, + 0.011029690504074097, + 0.007300831843167543, + 0.007651778403669596, + 0.0066697984002530575, + 0.0036041822750121355, + 0.007479446940124035, + 0.0060611688531935215, + 0.004264101851731539, + 0.03294393792748451, + 0.028045015409588814, + 0.023213695734739304, + 0.026145050302147865, + 0.017518300563097, + 0.03994321823120117, + 0.035929445177316666, + 0.046380579471588135, + 0.013592054136097431, + 0.010438889265060425, + 0.009942933917045593, + 0.00975787453353405, + 0.004812406376004219, + 0.006430907640606165, + 0.01341312751173973, + 0.011419081129133701, + 0.011915432289242744, + 0.008369289338588715, + 0.005522583145648241, + 0.005726013332605362, + 0.0029561726842075586, + 0.004127054009586573, + 0.00404746737331152, + 0.002618978498503566, + 0.003370773047208786, + 0.006815525237470865, + 0.004301158711314201, + 0.003783141728490591, + 0.0030489852651953697, + 0.0037719830870628357, + 0.003984593786299229, + 0.005135021638125181, + 0.008090262301266193, + 0.012485455721616745, + 0.008438610471785069, + 0.011022369377315044, + 0.005132356658577919, + 0.005387363489717245, + 0.008180586621165276, + 0.003778918879106641, + 0.007198156323283911, + 0.012074964120984077, + 0.01051690336316824, + 0.013449443504214287, + 0.007092667743563652, + 0.007826615124940872, + 0.011522717773914337, + 0.004167019389569759, + 0.01679224707186222, + 0.012059023603796959, + 0.014364966191351414, + 0.010885356925427914, + 0.004635852295905352, + 0.0052123297937214375, + 0.007684870157390833, + 0.003765018191188574, + 0.01680452562868595, + 0.006504559889435768, + 0.019562803208827972, + 0.008894107304513454, + 0.003813313553109765, + 0.003495537443086505, + 0.004589310381561518, + 0.0035280075389891863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.02330845594406128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787428192794323 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.009787428192794323, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787428192794323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07997261732816696, + 0.022384872660040855, + 0.024086199700832367, + 0.04580666869878769, + 0.07124242931604385, + 0.04446634650230408, + 0.15431956946849823, + 0.17568077147006989, + 0.19503018260002136, + 0.18701042234897614 + ], + "credit_map": [ + 0.020377429202198982, + 0.011195999570190907, + 0.009714487008750439, + 0.007850331254303455, + 0.005171613767743111, + 0.010507212020456791, + 0.0067658559419214725, + 0.008389684371650219, + 0.004065399058163166, + 0.002889293245971203, + 0.0030937001574784517, + 0.0026614617090672255, + 0.001518111675977707, + 0.002926008542999625, + 0.00304507021792233, + 0.002185829682275653, + 0.0033628810197114944, + 0.002916440600529313, + 0.0025632993783801794, + 0.0025467530358582735, + 0.0014654966071248055, + 0.004860417917370796, + 0.004476621281355619, + 0.0018942904425784945, + 0.011313741095364094, + 0.005278385244309902, + 0.008478487841784954, + 0.0036429036408662796, + 0.002007796661928296, + 0.003217420307919383, + 0.007597844582051039, + 0.004270086530596018, + 0.020308135077357292, + 0.010523772798478603, + 0.009095339104533195, + 0.008665502071380615, + 0.003630132880061865, + 0.007479768246412277, + 0.00626713503152132, + 0.005272644571959972, + 0.0054173520766198635, + 0.008968346752226353, + 0.007639688905328512, + 0.005601898301392794, + 0.0046044266782701015, + 0.0038102383259683847, + 0.005420636851340532, + 0.0030037551186978817, + 0.018142128363251686, + 0.027431348338723183, + 0.017893880605697632, + 0.027028826996684074, + 0.01255245041102171, + 0.011514253914356232, + 0.02984238974750042, + 0.009914291091263294, + 0.02222173847258091, + 0.032114915549755096, + 0.023564253002405167, + 0.03383719548583031, + 0.0169938113540411, + 0.018075881525874138, + 0.01633024401962757, + 0.012542741373181343, + 0.05350663885474205, + 0.018677467480301857, + 0.051923952996730804, + 0.02392636239528656, + 0.01022857055068016, + 0.01173113938421011, + 0.013209538534283638, + 0.011826501227915287, + 0.0554308257997036, + 0.012371301651000977, + 0.06417962163686752, + 0.02585991472005844, + 0.0058404733426868916, + 0.0074794357642531395, + 0.009121568873524666, + 0.00672728568315506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.009787428192794323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032692551612854 + ], + "opposite_logprob_tokens": [ + -2.3503270149230957 + ], + "logprob_reward": -0.10032692551612854, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032692551612854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23468029499053955, + 0.30520787835121155, + 0.05478501692414284, + 0.1352614015340805, + 0.050524596124887466, + 0.06667850911617279, + 0.028361845761537552, + 0.04526445269584656, + 0.032127730548381805, + 0.04710829257965088 + ], + "credit_map": [ + 0.05316835641860962, + 0.026727715507149696, + 0.027146944776177406, + 0.022690214216709137, + 0.017865225672721863, + 0.021961798891425133, + 0.0324479416012764, + 0.03267207369208336, + 0.09336159378290176, + 0.0242142453789711, + 0.04832369089126587, + 0.023986048996448517, + 0.01729578524827957, + 0.04289163276553154, + 0.036463432013988495, + 0.018671423196792603, + 0.006688671186566353, + 0.0055871400982141495, + 0.006700064521282911, + 0.007242943625897169, + 0.0033109572250396013, + 0.012017447501420975, + 0.009315935894846916, + 0.00392185989767313, + 0.030661530792713165, + 0.018871372565627098, + 0.02763342671096325, + 0.011744248680770397, + 0.005766660440713167, + 0.007903342135250568, + 0.023715557530522346, + 0.008965249173343182, + 0.01313796080648899, + 0.00799603946506977, + 0.0063610803335905075, + 0.006146397907286882, + 0.00294996565207839, + 0.005375569220632315, + 0.004264510236680508, + 0.004293073434382677, + 0.006849050056189299, + 0.013404086232185364, + 0.008149344474077225, + 0.009595793671905994, + 0.006179548799991608, + 0.0076714991591870785, + 0.007282239384949207, + 0.007546953856945038, + 0.0027786388527601957, + 0.003572516841813922, + 0.003866989864036441, + 0.004860324785113335, + 0.00255668256431818, + 0.0023144828155636787, + 0.006716850679367781, + 0.0016953609883785248, + 0.006239467300474644, + 0.0072129229083657265, + 0.0059043169021606445, + 0.007128073368221521, + 0.003230125643312931, + 0.004410055465996265, + 0.007622697856277227, + 0.0035167953465133905, + 0.00753947626799345, + 0.004205137491226196, + 0.007021857425570488, + 0.00453872699290514, + 0.002185774501413107, + 0.002130331937223673, + 0.00243460014462471, + 0.002071827184408903, + 0.012409229762852192, + 0.0036336458288133144, + 0.015600156038999557, + 0.006986450869590044, + 0.0020041554234921932, + 0.0020171399228274822, + 0.0022729113698005676, + 0.0021846015006303787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.10032692551612854 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978577509522438 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.00978577509522438, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978577509522438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26929089426994324, + 0.0392339713871479, + 0.03129804879426956, + 0.03811643645167351, + 0.08132773637771606, + 0.10377434641122818, + 0.0736895352602005, + 0.09339313954114914, + 0.1735990047454834, + 0.09627684205770493 + ], + "credit_map": [ + 0.030746275559067726, + 0.03019038215279579, + 0.03717290237545967, + 0.034682389348745346, + 0.022714652121067047, + 0.029193535447120667, + 0.044066764414310455, + 0.04052400588989258, + 0.008039639331400394, + 0.0045200614258646965, + 0.005083680618554354, + 0.005473189055919647, + 0.002740590600296855, + 0.003841083962470293, + 0.004697420634329319, + 0.004838306922465563, + 0.005621881689876318, + 0.0029519926756620407, + 0.003707288531586528, + 0.00397835485637188, + 0.0020926434081047773, + 0.004797345958650112, + 0.004743911791592836, + 0.0034046319779008627, + 0.009113620966672897, + 0.004556016996502876, + 0.006628861650824547, + 0.005026674829423428, + 0.002191514242440462, + 0.0017772363498806953, + 0.006178997922688723, + 0.0026435162872076035, + 0.021057793870568275, + 0.012893840670585632, + 0.01058960147202015, + 0.009831858798861504, + 0.004570084158331156, + 0.010526116006076336, + 0.006186425220221281, + 0.00567201804369688, + 0.011336765252053738, + 0.02846818044781685, + 0.01603693515062332, + 0.01756855845451355, + 0.0076828752644360065, + 0.005622947588562965, + 0.010760274715721607, + 0.006297815591096878, + 0.008210494183003902, + 0.013367130421102047, + 0.009566210210323334, + 0.012200755998492241, + 0.006966313812881708, + 0.007258768659085035, + 0.011157547123730183, + 0.00496232183650136, + 0.013959024101495743, + 0.015767721459269524, + 0.012497786432504654, + 0.014829592779278755, + 0.008634975180029869, + 0.011182626709342003, + 0.01107704360038042, + 0.005444372072815895, + 0.031514909118413925, + 0.030214931815862656, + 0.029116913676261902, + 0.026718784123659134, + 0.01412404328584671, + 0.012056002393364906, + 0.018228016793727875, + 0.011625414714217186, + 0.01181106548756361, + 0.014231964945793152, + 0.021128937602043152, + 0.016128025949001312, + 0.008020173758268356, + 0.009273060597479343, + 0.009792391210794449, + 0.005891225766390562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.00978577509522438 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a level angle, slowly pans forward, focusing on a wrinkled elderly man. Beneath his sparse silver-white hair lie gentle, deep eyes. The wrinkles on his forehead and around his eyes, as deep as the furrows of time, bear witness to the passage of time. He sits serenely in a room bathed in afternoon sunlight, wearing a worn, rough, dark gray sweater. Soft, warm light slants in from the side windows, illuminating the dust particles dancing in the air and casting long, blurred shadows on the wooden bookshelves behind him. Steadily and calmly, he slowly extends his hand, veined and covered in age spots and fine lines, toward the camera. His knuckles are slightly thickened from years of hard work, and the lines on his palm are clearly visible, as if reaching out to the audience on screen or offering a silent message. The entire picture presents a warm and faded retro film texture, with low color saturation and slight film grain. The main colors are amber and dark brown, creating a tranquil, nostalgic and story-filled emotional atmosphere.", + "seed": 13024, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_15/index_2536_seed_13024_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019252557959407567, + "vlm_logprob": -0.019252557959407567, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.019252557959407567, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00531735410913825 + ], + "opposite_logprob_tokens": [ + -5.880317211151123 + ], + "logprob_reward": -0.00531735410913825, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.00531735410913825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1487548053264618, + 0.08464288711547852, + 0.04798723757266998, + 0.13375237584114075, + 0.08219102770090103, + 0.0475396029651165, + 0.15297776460647583, + 0.07429701834917068, + 0.1602928638458252, + 0.06756438314914703 + ], + "credit_map": [ + 0.02885996364057064, + 0.022128386422991753, + 0.015577875077724457, + 0.013384453020989895, + 0.01943330280482769, + 0.015882102772593498, + 0.01158476248383522, + 0.021903974935412407, + 0.023019591346383095, + 0.009817028418183327, + 0.013558318838477135, + 0.010671092197299004, + 0.004852441139519215, + 0.006135938223451376, + 0.005609680898487568, + 0.01097880955785513, + 0.007664535194635391, + 0.006503732409328222, + 0.008521854877471924, + 0.0077039687894284725, + 0.003865298582240939, + 0.0029257347341626883, + 0.0036022881977260113, + 0.0071998294442892075, + 0.014336111955344677, + 0.020941119641065598, + 0.02281956747174263, + 0.021109825000166893, + 0.01494382694363594, + 0.010360509157180786, + 0.010664871893823147, + 0.018576547503471375, + 0.008272986859083176, + 0.014477239921689034, + 0.01152217946946621, + 0.01305042952299118, + 0.00922340340912342, + 0.005433943122625351, + 0.007590696681290865, + 0.01262015663087368, + 0.0047623757272958755, + 0.008652010932564735, + 0.0064834789372980595, + 0.006201806012541056, + 0.004790334962308407, + 0.00508154695853591, + 0.004556701984256506, + 0.007011353969573975, + 0.016527697443962097, + 0.02271684817969799, + 0.033766403794288635, + 0.02403298020362854, + 0.012093810364603996, + 0.008922887966036797, + 0.012696885503828526, + 0.022220272570848465, + 0.008479439653456211, + 0.012524865567684174, + 0.012020177207887173, + 0.008640947751700878, + 0.005536939483135939, + 0.006066479720175266, + 0.008948560804128647, + 0.01207960955798626, + 0.023683244362473488, + 0.027724003419280052, + 0.029134979471564293, + 0.020624972879886627, + 0.011529597453773022, + 0.008755390532314777, + 0.014760980382561684, + 0.024079708382487297, + 0.009497620165348053, + 0.008946006186306477, + 0.012121746316552162, + 0.008074858225882053, + 0.006431845482438803, + 0.009622909128665924, + 0.004176337271928787, + 0.008693070150911808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.00531735410913825 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370962664484978 + ], + "opposite_logprob_tokens": [ + -3.4087095260620117 + ], + "logprob_reward": -0.03370962664484978, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370962664484978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12239143252372742, + 0.14413194358348846, + 0.17186041176319122, + 0.08625458180904388, + 0.08485403656959534, + 0.04307058826088905, + 0.04616279900074005, + 0.08522412925958633, + 0.044618718326091766, + 0.171431303024292 + ], + "credit_map": [ + 0.031201401725411415, + 0.010470821522176266, + 0.017222503200173378, + 0.018852701410651207, + 0.00991650391370058, + 0.009877924807369709, + 0.0077960798516869545, + 0.01705349236726761, + 0.03853783756494522, + 0.014322771690785885, + 0.01980462484061718, + 0.020898081362247467, + 0.008058382198214531, + 0.009710152633488178, + 0.011803158558905125, + 0.0209969375282526, + 0.04679142311215401, + 0.016273263841867447, + 0.02850145660340786, + 0.02330349013209343, + 0.010603268630802631, + 0.012051321566104889, + 0.011222743429243565, + 0.023113444447517395, + 0.014926119707524776, + 0.011754532344639301, + 0.012021954171359539, + 0.010033083148300648, + 0.008584222756326199, + 0.009875253774225712, + 0.006648381240665913, + 0.012411030009388924, + 0.009412835352122784, + 0.015304411761462688, + 0.016547007486224174, + 0.012906737625598907, + 0.007275126408785582, + 0.005397303029894829, + 0.005570854526013136, + 0.012439761310815811, + 0.004642532207071781, + 0.0059285215102136135, + 0.007630146108567715, + 0.007940946146845818, + 0.0035512172617018223, + 0.004097839817404747, + 0.002683446276932955, + 0.00659593753516674, + 0.005285587627440691, + 0.006653564516454935, + 0.00943029671907425, + 0.005504269618541002, + 0.004211165010929108, + 0.0036124519538134336, + 0.004672096576541662, + 0.006793366279453039, + 0.010724649764597416, + 0.015823960304260254, + 0.012884334661066532, + 0.008887778967618942, + 0.007920844480395317, + 0.009075719863176346, + 0.007496186997741461, + 0.012410653755068779, + 0.005604568403214216, + 0.00573741365224123, + 0.005915977992117405, + 0.006297590676695108, + 0.004065179266035557, + 0.007449836935847998, + 0.004073947202414274, + 0.005474206525832415, + 0.022849705070257187, + 0.03242795914411545, + 0.028068596497178078, + 0.022153588011860847, + 0.013520886190235615, + 0.016712527722120285, + 0.014105075970292091, + 0.02159295789897442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.03370962664484978 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584410056471825 + ], + "opposite_logprob_tokens": [ + -3.8955843448638916 + ], + "logprob_reward": -0.020584410056471825, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584410056471825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04515887051820755, + 0.029306620359420776, + 0.04457850754261017, + 0.23007096350193024, + 0.08587081730365753, + 0.050274789333343506, + 0.29096224904060364, + 0.025067178532481194, + 0.03338474780321121, + 0.16532522439956665 + ], + "credit_map": [ + 0.006487926933914423, + 0.00459061237052083, + 0.0060907332226634026, + 0.0063404212705791, + 0.003808115841820836, + 0.008120378479361534, + 0.003922808915376663, + 0.00579787464812398, + 0.0062874481081962585, + 0.0031083098147064447, + 0.0044665210880339146, + 0.003642103634774685, + 0.002149311825633049, + 0.002936909906566143, + 0.0023443708196282387, + 0.004371645860373974, + 0.008884896524250507, + 0.005556928459554911, + 0.006426211446523666, + 0.005983143579214811, + 0.0036368670407682657, + 0.003750766161829233, + 0.004222766496241093, + 0.006116925738751888, + 0.049390874803066254, + 0.02662338875234127, + 0.03002431057393551, + 0.026375556364655495, + 0.023442361503839493, + 0.026417681947350502, + 0.017793672159314156, + 0.03000311739742756, + 0.009078295901417732, + 0.010062647052109241, + 0.01886564865708351, + 0.014620542526245117, + 0.006738098803907633, + 0.005300505552440882, + 0.008752922527492046, + 0.012452160008251667, + 0.00464024068787694, + 0.007587783504277468, + 0.010704923421144485, + 0.008347699418663979, + 0.0037772657815366983, + 0.0030764418188482523, + 0.003858087584376335, + 0.008282351307570934, + 0.031056459993124008, + 0.03986382111907005, + 0.0736820325255394, + 0.05167783051729202, + 0.019358975812792778, + 0.012860746122896671, + 0.0231403186917305, + 0.03932207077741623, + 0.003311163280159235, + 0.004364020191133022, + 0.002963865175843239, + 0.003849499626085162, + 0.0023192549124360085, + 0.0018653909210115671, + 0.002755939494818449, + 0.0036380442325025797, + 0.0045761060900986195, + 0.005727825686335564, + 0.004773193970322609, + 0.0036957482807338238, + 0.003205912420526147, + 0.0034818712156265974, + 0.003042858559638262, + 0.0048812320455908775, + 0.026211757212877274, + 0.04163007810711861, + 0.028138060122728348, + 0.015072236768901348, + 0.013285676017403603, + 0.00879127811640501, + 0.010541528463363647, + 0.021654605865478516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -0.020584410056471825 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577870309352875 + ], + "opposite_logprob_tokens": [ + -3.895577907562256 + ], + "logprob_reward": -0.020577870309352875, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577870309352875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08136855065822601, + 0.21708916127681732, + 0.03968837857246399, + 0.2552867531776428, + 0.04786737263202667, + 0.03854139521718025, + 0.12990589439868927, + 0.03831196576356888, + 0.06252788752317429, + 0.08941260725259781 + ], + "credit_map": [ + 0.013707411475479603, + 0.011003554798662663, + 0.013459869660437107, + 0.011184589937329292, + 0.008450289256870747, + 0.006434950046241283, + 0.005984290968626738, + 0.011143594980239868, + 0.06359802931547165, + 0.019488975405693054, + 0.0357927605509758, + 0.02727850340306759, + 0.009911789558827877, + 0.015258110128343105, + 0.016066396608948708, + 0.02969459258019924, + 0.009693527594208717, + 0.004201805684715509, + 0.00654213223606348, + 0.0049325511790812016, + 0.003307109232991934, + 0.002941432176157832, + 0.0027118513826280832, + 0.005357966758310795, + 0.05569390952587128, + 0.029430393129587173, + 0.0244329571723938, + 0.028686150908470154, + 0.02891537919640541, + 0.035732779651880264, + 0.01930876448750496, + 0.03308642655611038, + 0.005437379237264395, + 0.006114295683801174, + 0.009388715028762817, + 0.006965375505387783, + 0.003570564789697528, + 0.005091791041195393, + 0.0043931808322668076, + 0.006906069815158844, + 0.0034458546433597803, + 0.006833079736679792, + 0.005895103327929974, + 0.005360618699342012, + 0.0035830994602292776, + 0.005146781448274851, + 0.003146410919725895, + 0.005130445118993521, + 0.013552220538258553, + 0.015362431295216084, + 0.034837041050195694, + 0.02249707654118538, + 0.008600901812314987, + 0.006645648740231991, + 0.01060427911579609, + 0.017806295305490494, + 0.004320022650063038, + 0.005090147256851196, + 0.0036728600971400738, + 0.0038287912029772997, + 0.003849199041724205, + 0.008706835098564625, + 0.003758789272978902, + 0.005085320677608252, + 0.008510139770805836, + 0.008931844495236874, + 0.009119508787989616, + 0.00586625374853611, + 0.004887404851615429, + 0.010703850537538528, + 0.005914285313338041, + 0.008594597689807415, + 0.010261822491884232, + 0.016282513737678528, + 0.01054733619093895, + 0.0069969124160707, + 0.008976833894848824, + 0.01803763397037983, + 0.007449204567819834, + 0.010860356502234936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 6 + ], + "selected_reward": -0.020577870309352875 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073528677225113 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073528677225113, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073528677225113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10897000133991241, + 0.043394967913627625, + 0.03867313638329506, + 0.14381355047225952, + 0.0920301228761673, + 0.13003137707710266, + 0.23498599231243134, + 0.05462604761123657, + 0.034266337752342224, + 0.11920853704214096 + ], + "credit_map": [ + 0.011651454493403435, + 0.014133697375655174, + 0.01363673247396946, + 0.015190238133072853, + 0.014669139869511127, + 0.016017848625779152, + 0.009250884875655174, + 0.014419998973608017, + 0.008068249560892582, + 0.005075165070593357, + 0.005928926169872284, + 0.005659110844135284, + 0.0031967847608029842, + 0.006602685432881117, + 0.0033090531360358, + 0.005554988514631987, + 0.008298670873045921, + 0.00530086737126112, + 0.004372813738882542, + 0.004691042471677065, + 0.0029900302179157734, + 0.0036711059510707855, + 0.004136263858526945, + 0.005212338641285896, + 0.026195982471108437, + 0.016711808741092682, + 0.016704941168427467, + 0.01859850250184536, + 0.019906582310795784, + 0.01674473099410534, + 0.011159498244524002, + 0.017791498452425003, + 0.00986213143914938, + 0.011973232962191105, + 0.01997300796210766, + 0.01351625844836235, + 0.007350035011768341, + 0.005948827601969242, + 0.00945130456238985, + 0.01395531464368105, + 0.010916768573224545, + 0.01907162182033062, + 0.027834642678499222, + 0.02347755990922451, + 0.008578743785619736, + 0.00936388410627842, + 0.008984751999378204, + 0.021803393959999084, + 0.024102047085762024, + 0.031998034566640854, + 0.05613334849476814, + 0.04295917972922325, + 0.017170652747154236, + 0.01229104120284319, + 0.01841210573911667, + 0.03191955015063286, + 0.008194436319172382, + 0.007389141246676445, + 0.008843093179166317, + 0.006644946523010731, + 0.003611016320064664, + 0.005203309003263712, + 0.00623980350792408, + 0.008500299416482449, + 0.005278045777231455, + 0.004782048519700766, + 0.005325507838279009, + 0.003500475315377116, + 0.0029694936238229275, + 0.003933279309421778, + 0.003097853157669306, + 0.005379629321396351, + 0.017939038574695587, + 0.028786877170205116, + 0.020044974982738495, + 0.010964423418045044, + 0.009211396798491478, + 0.0075616552494466305, + 0.00871156994253397, + 0.015988599509000778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.016073528677225113 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a level angle, slowly pans forward, focusing on a wrinkled elderly man. Beneath his sparse silver-white hair lie gentle, deep eyes. The wrinkles on his forehead and around his eyes, as deep as the furrows of time, bear witness to the passage of time. He sits serenely in a room bathed in afternoon sunlight, wearing a worn, rough, dark gray sweater. Soft, warm light slants in from the side windows, illuminating the dust particles dancing in the air and casting long, blurred shadows on the wooden bookshelves behind him. Steadily and calmly, he slowly extends his hand, veined and covered in age spots and fine lines, toward the camera. His knuckles are slightly thickened from years of hard work, and the lines on his palm are clearly visible, as if reaching out to the audience on screen or offering a silent message. The entire picture presents a warm and faded retro film texture, with low color saturation and slight film grain. The main colors are amber and dark brown, creating a tranquil, nostalgic and story-filled emotional atmosphere.", + "seed": 13025, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_15/index_2536_seed_13025_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01493010288104415, + "vlm_logprob": -0.01493010288104415, + "vlm_yes_no_margin": 4.4499999523162845, + "vlm_token_credit": -0.01493010288104415, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020161692053079605 + ], + "opposite_logprob_tokens": [ + -4.1451616287231445 + ], + "logprob_reward": -0.020161692053079605, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.020161692053079605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06634373217821121, + 0.03730548545718193, + 0.0778379961848259, + 0.28465110063552856, + 0.125062957406044, + 0.11229953914880753, + 0.048388127237558365, + 0.14006929099559784, + 0.061555419117212296, + 0.04648639261722565 + ], + "credit_map": [ + 0.004679783713072538, + 0.007878551259636879, + 0.005956570152193308, + 0.010475929826498032, + 0.010994572192430496, + 0.008567577227950096, + 0.010651965625584126, + 0.007138777058571577, + 0.007538839243352413, + 0.004843998234719038, + 0.005428550764918327, + 0.004305014852434397, + 0.003185965120792389, + 0.0036675857845693827, + 0.00419286172837019, + 0.004142667632550001, + 0.007658719550818205, + 0.008785003796219826, + 0.0059796287678182125, + 0.010706204921007156, + 0.0097574507817626, + 0.006867782212793827, + 0.01584525965154171, + 0.012237942777574062, + 0.02494555152952671, + 0.06642138957977295, + 0.020262915641069412, + 0.03677864000201225, + 0.03590010106563568, + 0.018901968374848366, + 0.05253880098462105, + 0.02890169434249401, + 0.024718958884477615, + 0.018873583525419235, + 0.018506653606891632, + 0.010025696828961372, + 0.009763361886143684, + 0.006967698223888874, + 0.021949024870991707, + 0.014257971197366714, + 0.01081964559853077, + 0.011982488445937634, + 0.012640347704291344, + 0.008384576998651028, + 0.020713981240987778, + 0.014181465841829777, + 0.014647289179265499, + 0.018929732963442802, + 0.004160046111792326, + 0.005503599997609854, + 0.004872745368629694, + 0.005535901058465242, + 0.009280473925173283, + 0.008037984371185303, + 0.00603825505822897, + 0.0049591162241995335, + 0.00897185318171978, + 0.01610364392399788, + 0.008891282603144646, + 0.008933687582612038, + 0.031162630766630173, + 0.024208903312683105, + 0.010562467388808727, + 0.03123481571674347, + 0.004830922931432724, + 0.010129017755389214, + 0.005089940968900919, + 0.004344006069004536, + 0.010881814174354076, + 0.011561076156795025, + 0.003526399377733469, + 0.011192237958312035, + 0.004087046254426241, + 0.006182106677442789, + 0.006279286462813616, + 0.007259063888341188, + 0.005888459272682667, + 0.0057704565115273, + 0.005820572376251221, + 0.005199396051466465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.020161692053079605 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331392839550972 + ], + "opposite_logprob_tokens": [ + -3.7733139991760254 + ], + "logprob_reward": -0.02331392839550972, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331392839550972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08266555517911911, + 0.0440269336104393, + 0.11392706632614136, + 0.19749075174331665, + 0.07782503962516785, + 0.05342114716768265, + 0.09405415505170822, + 0.2381044626235962, + 0.05865639075636864, + 0.03982844203710556 + ], + "credit_map": [ + 0.008362339809536934, + 0.009807653725147247, + 0.006108580157160759, + 0.009339655749499798, + 0.013172190636396408, + 0.009973817504942417, + 0.013745480217039585, + 0.012155840173363686, + 0.006316314451396465, + 0.004850282333791256, + 0.004640498198568821, + 0.004934401251375675, + 0.0064895437099039555, + 0.004329448565840721, + 0.00640412513166666, + 0.006062324624508619, + 0.01220913976430893, + 0.01213379018008709, + 0.011253110133111477, + 0.020878437906503677, + 0.01252716314047575, + 0.010333235375583172, + 0.019689476117491722, + 0.014902726747095585, + 0.016676979139447212, + 0.04425634816288948, + 0.015518414787948132, + 0.03049791045486927, + 0.028214138001203537, + 0.014399411156773567, + 0.02963610365986824, + 0.01829146407544613, + 0.016947001218795776, + 0.011428050696849823, + 0.008871533907949924, + 0.006070911884307861, + 0.009532345458865166, + 0.0043894643895328045, + 0.012693656608462334, + 0.00789208896458149, + 0.00765237445011735, + 0.0073418449610471725, + 0.007042264565825462, + 0.003594709327444434, + 0.0066423723474144936, + 0.007233554497361183, + 0.00707580428570509, + 0.0068382276222109795, + 0.006130330730229616, + 0.00828044954687357, + 0.009622124023735523, + 0.009008876979351044, + 0.023294420912861824, + 0.015536308288574219, + 0.012194176204502583, + 0.009987475350499153, + 0.011711105704307556, + 0.01585295796394348, + 0.009351666085422039, + 0.007727210875600576, + 0.08003319054841995, + 0.022552426904439926, + 0.01184256561100483, + 0.07903336733579636, + 0.004635017830878496, + 0.009116186760365963, + 0.006678049452602863, + 0.004029749892652035, + 0.011293597519397736, + 0.012177865020930767, + 0.0038148073945194483, + 0.006911123171448708, + 0.00461875693872571, + 0.005911113228648901, + 0.006007223855704069, + 0.006111358758062124, + 0.004547975957393646, + 0.004655918572098017, + 0.003916718531399965, + 0.004059379920363426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.02331392839550972 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003202668856829405 + ], + "opposite_logprob_tokens": [ + -5.753202438354492 + ], + "logprob_reward": -0.003202668856829405, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003202668856829405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050471287220716476, + 0.04230299964547157, + 0.024172276258468628, + 0.03400953486561775, + 0.10833562165498734, + 0.047858819365501404, + 0.20547960698604584, + 0.1686868667602539, + 0.2311483472585678, + 0.08753464370965958 + ], + "credit_map": [ + 0.008177646435797215, + 0.008565280586481094, + 0.0033621545881032944, + 0.004759426694363356, + 0.005635556299239397, + 0.007996014319360256, + 0.006176265887916088, + 0.005798943806439638, + 0.004187870305031538, + 0.007984376512467861, + 0.004562184680253267, + 0.004744702018797398, + 0.005117973778396845, + 0.0029581086710095406, + 0.005885837133973837, + 0.006861947942525148, + 0.003409650409594178, + 0.0035823765210807323, + 0.0016700404230505228, + 0.0029852204024791718, + 0.002850631019100547, + 0.002051598858088255, + 0.004026395734399557, + 0.0035963624250143766, + 0.003688531694933772, + 0.004286668729037046, + 0.00320665561594069, + 0.0052717518992722034, + 0.004414418246597052, + 0.003430761396884918, + 0.005953675601631403, + 0.0037570707499980927, + 0.021074295043945312, + 0.01551844459027052, + 0.012501537799835205, + 0.010709796100854874, + 0.013616018928587437, + 0.005225527565926313, + 0.017292646691203117, + 0.012397351674735546, + 0.005258761812001467, + 0.007067042402923107, + 0.004575666505843401, + 0.004530574195086956, + 0.005903745535761118, + 0.0047268979251384735, + 0.006732078734785318, + 0.009064052253961563, + 0.017130106687545776, + 0.01792444847524166, + 0.02352806366980076, + 0.028818955644965172, + 0.046877965331077576, + 0.02381030097603798, + 0.02570020593702793, + 0.02168956957757473, + 0.019127132371068, + 0.019544141367077827, + 0.013337812386453152, + 0.019426120445132256, + 0.03206503018736839, + 0.023180974647402763, + 0.014683295972645283, + 0.02732235938310623, + 0.016813717782497406, + 0.04617531970143318, + 0.025093799456954002, + 0.01726827584207058, + 0.04470391571521759, + 0.04824042320251465, + 0.013811458833515644, + 0.01904142089188099, + 0.006796849425882101, + 0.009359339252114296, + 0.009535301476716995, + 0.011230308562517166, + 0.013415043242275715, + 0.012801348231732845, + 0.01068242359906435, + 0.013714030385017395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.003202668856829405 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781643748283386 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.009781643748283386, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781643748283386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07950703054666519, + 0.0375252440571785, + 0.10953008383512497, + 0.07466461509466171, + 0.12468528747558594, + 0.28008025884628296, + 0.05720385164022446, + 0.1304265856742859, + 0.054685674607753754, + 0.051691390573978424 + ], + "credit_map": [ + 0.014150711707770824, + 0.010805296711623669, + 0.0077830720692873, + 0.008929576724767685, + 0.010660359635949135, + 0.011642652563750744, + 0.008139380253851414, + 0.007395981810986996, + 0.0065834796987473965, + 0.005261983722448349, + 0.004866736009716988, + 0.004766255617141724, + 0.00354429567232728, + 0.003403738373890519, + 0.0050964695401489735, + 0.0040022870525717735, + 0.025403078645467758, + 0.00920860841870308, + 0.0193636491894722, + 0.013170457445085049, + 0.011040744371712208, + 0.007547674234956503, + 0.014864685013890266, + 0.008931183256208897, + 0.005816935561597347, + 0.015262654982507229, + 0.006902175955474377, + 0.00980034563690424, + 0.010038696229457855, + 0.006693666335195303, + 0.013503843918442726, + 0.006646295543760061, + 0.021739350631833076, + 0.016972316429018974, + 0.01312036719173193, + 0.011506975628435612, + 0.01965675875544548, + 0.009176455438137054, + 0.01698027364909649, + 0.015532790683209896, + 0.040950533002614975, + 0.0445072166621685, + 0.039829086512327194, + 0.022009246051311493, + 0.03772086277604103, + 0.022839466109871864, + 0.041002701967954636, + 0.03122115507721901, + 0.00847723800688982, + 0.007877934724092484, + 0.006833422463387251, + 0.006605879869312048, + 0.007279965095221996, + 0.005787283647805452, + 0.008232944644987583, + 0.006109186448156834, + 0.006703418679535389, + 0.011936494149267673, + 0.005869903601706028, + 0.006867995020002127, + 0.03919345140457153, + 0.016581347212195396, + 0.008160267025232315, + 0.03511371836066246, + 0.004109186120331287, + 0.00913322251290083, + 0.005398291628807783, + 0.004281292669475079, + 0.011232297867536545, + 0.010022683069109917, + 0.0032788810785859823, + 0.007229818031191826, + 0.006162041332572699, + 0.006939397659152746, + 0.006556549109518528, + 0.005705921910703182, + 0.006848958786576986, + 0.006843909155577421, + 0.0032861593645066023, + 0.00934845581650734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.009781643748283386 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a wrinkled elderly man with sparse silver-white hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a worn, rough, dark gray sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slowly extend his hand toward the camera, showing age spots and fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a room with wooden bookshelves in the background, bathed in afternoon sunlight with visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a warm, faded retro film texture with low color saturation and slight film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819058135151863 + ], + "opposite_logprob_tokens": [ + -4.018190383911133 + ], + "logprob_reward": -0.01819058135151863, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819058135151863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07770109176635742, + 0.03278864175081253, + 0.029338225722312927, + 0.08997951447963715, + 0.272613525390625, + 0.1254265308380127, + 0.07463205605745316, + 0.15006372332572937, + 0.09330399334430695, + 0.054152701050043106 + ], + "credit_map": [ + 0.012559734284877777, + 0.01002987939864397, + 0.004898781888186932, + 0.0065047862008214, + 0.011287868022918701, + 0.014696449972689152, + 0.010346113704144955, + 0.007377479691058397, + 0.007290452253073454, + 0.00441376306116581, + 0.004945987835526466, + 0.0031528878025710583, + 0.0035696974955499172, + 0.0028006560169160366, + 0.0033183186315000057, + 0.003296878421679139, + 0.003005238017067313, + 0.0034306917805224657, + 0.0033068580087274313, + 0.0035464668180793524, + 0.005179860629141331, + 0.0028093860018998384, + 0.004555586259812117, + 0.0035041377414017916, + 0.007125097326934338, + 0.0196953434497118, + 0.006204275414347649, + 0.011530113406479359, + 0.012313535436987877, + 0.006576227489858866, + 0.0161102544516325, + 0.010424675419926643, + 0.05477411672472954, + 0.04155656322836876, + 0.03777507320046425, + 0.025880981236696243, + 0.020132120698690414, + 0.014479042962193489, + 0.04700031504034996, + 0.031015312299132347, + 0.011787313967943192, + 0.019452448934316635, + 0.01661190390586853, + 0.00962576363235712, + 0.02176050841808319, + 0.012174221687018871, + 0.017389776185154915, + 0.01662459596991539, + 0.005460578482598066, + 0.007339030038565397, + 0.008324368856847286, + 0.006848957855254412, + 0.018015732988715172, + 0.010348699055612087, + 0.00977291725575924, + 0.008521771989762783, + 0.0062397923320531845, + 0.009683013893663883, + 0.007046649698168039, + 0.007816634140908718, + 0.050931524485349655, + 0.014080540277063847, + 0.006730175577104092, + 0.04753539338707924, + 0.006423283834010363, + 0.016246633604168892, + 0.008469895459711552, + 0.00857274979352951, + 0.01757158152759075, + 0.020220838487148285, + 0.007462679408490658, + 0.008336328901350498, + 0.004758581984788179, + 0.005871880333870649, + 0.0062011475674808025, + 0.007084627635776997, + 0.007227135822176933, + 0.007151906378567219, + 0.005018461029976606, + 0.010838957503437996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.01819058135151863 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_2/index_tensor([8559, 8559])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_2/index_tensor([8559, 8559])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..253b4e211fb346c02263d67fc665e9c88cfd577f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_2/index_tensor([8559, 8559])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, extreme wide shot surveys a vast, sun-scorched desert plain under the harsh, white-out light of the midday sun. The parched, cracked earth stretches to distant, hazy mountains, and the air shimmers with a fiery glow. Suddenly, the ground begins to buckle and crack violently, a spiderweb of matter ripping outward from its center. With immense, relentless force, a massive, rectangular monolith rises from the earth, its craggy, weathered granite surface betraying immense age. Its catastrophic ascent shatters the surrounding terrain, hurling chunks of rock and soil into the pale blue sky. A dense, roiling maelstrom of pulverized rock and ochre dust rapidly expands, tumbling upward and outward, engulfing the base of the slab and unleashing a powerful, apocalyptic, unleashing of primordial earthly forces.", + "seed": 15971, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_2/index_8559_seed_15971_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0157231506658718, + "vlm_logprob": -0.0157231506658718, + "vlm_yes_no_margin": 5.4, + "vlm_token_credit": -0.0157231506658718, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06871397793292999 + ], + "opposite_logprob_tokens": [ + -2.818713903427124 + ], + "logprob_reward": -0.06871397793292999, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06871397793292999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47096097469329834, + 0.08790290355682373, + 0.04378291964530945, + 0.06087600067257881, + 0.036843303591012955, + 0.08180960267782211, + 0.036839958280324936, + 0.13230346143245697, + 0.02442706935107708, + 0.024253714829683304 + ], + "credit_map": [ + 0.03985855355858803, + 0.02163538709282875, + 0.020952139049768448, + 0.009734570048749447, + 0.17662815749645233, + 0.18220841884613037, + 0.009262208826839924, + 0.010681542567908764, + 0.007433174643665552, + 0.01403513178229332, + 0.010652557015419006, + 0.01214255765080452, + 0.0141235226765275, + 0.013004068285226822, + 0.007184980437159538, + 0.00932690966874361, + 0.003249896690249443, + 0.005421891342848539, + 0.009040617384016514, + 0.003435639664530754, + 0.007097829133272171, + 0.004322574008256197, + 0.002533831400796771, + 0.00868063885718584, + 0.003913369495421648, + 0.009227321483194828, + 0.005740354768931866, + 0.014806404709815979, + 0.015535471960902214, + 0.004525850061327219, + 0.003638241905719042, + 0.0034889879170805216, + 0.003987891599535942, + 0.0057741389609873295, + 0.006054526660591364, + 0.0056902687065303326, + 0.004542991053313017, + 0.003359077963978052, + 0.00419554952532053, + 0.0032388600520789623, + 0.004949057940393686, + 0.013892426155507565, + 0.009239238686859608, + 0.009557122364640236, + 0.009725261479616165, + 0.007298002485185862, + 0.018666058778762817, + 0.008482434786856174, + 0.0030945097096264362, + 0.004747156985104084, + 0.006967515684664249, + 0.004480792209506035, + 0.004148758947849274, + 0.0038151429034769535, + 0.0036521742586046457, + 0.005933905020356178, + 0.012540415860712528, + 0.018546750769019127, + 0.02447439543902874, + 0.014938775449991226, + 0.023958008736371994, + 0.0072145015001297, + 0.020054513588547707, + 0.01057610847055912, + 0.0022629008162766695, + 0.0035303684417158365, + 0.004249638877809048, + 0.0027217005845159292, + 0.0024242987856268883, + 0.001667447853833437, + 0.0038962955586612225, + 0.0036744195967912674, + 0.00203178939409554, + 0.0031731731723994017, + 0.0032924057450145483, + 0.002126390114426613, + 0.002345944056287408, + 0.0035561819095164537, + 0.005048657767474651, + 0.002679172670468688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.06871397793292999 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001522692502476275 + ], + "opposite_logprob_tokens": [ + -6.501522541046143 + ], + "logprob_reward": -0.001522692502476275, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001522692502476275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49478819966316223, + 0.10885556787252426, + 0.03487295284867287, + 0.07109634578227997, + 0.028560420498251915, + 0.03586546331644058, + 0.040659189224243164, + 0.06422417610883713, + 0.04515155032277107, + 0.07592616230249405 + ], + "credit_map": [ + 0.07119952887296677, + 0.044268399477005005, + 0.033811524510383606, + 0.025592464953660965, + 0.135552778840065, + 0.14397555589675903, + 0.01748080737888813, + 0.022907119244337082, + 0.008992849849164486, + 0.01428037416189909, + 0.015447682701051235, + 0.021194132044911385, + 0.009027800522744656, + 0.018240218982100487, + 0.008334293961524963, + 0.013338220305740833, + 0.0029676456470042467, + 0.0032579326070845127, + 0.007248484063893557, + 0.0032066020648926497, + 0.00417314050719142, + 0.004978935234248638, + 0.004016125109046698, + 0.005024088080972433, + 0.004605122841894627, + 0.007874766364693642, + 0.0072699724696576595, + 0.01916668750345707, + 0.01742984727025032, + 0.0037064794450998306, + 0.006797373294830322, + 0.004246096592396498, + 0.002634667092934251, + 0.004604600835591555, + 0.004327648784965277, + 0.0046891113743186, + 0.005060379393398762, + 0.0021020120475441217, + 0.002665939275175333, + 0.002476061461493373, + 0.0019994713366031647, + 0.009568894281983376, + 0.0039009754545986652, + 0.0037980496417731047, + 0.004098546225577593, + 0.00292060780338943, + 0.0064706094563007355, + 0.003108313074335456, + 0.003480890765786171, + 0.0064386180602014065, + 0.0067328279837965965, + 0.004432136658579111, + 0.005836356896907091, + 0.003632064675912261, + 0.004215361084789038, + 0.005890932399779558, + 0.005729028023779392, + 0.009588513523340225, + 0.011242487467825413, + 0.0075668152421712875, + 0.013610518537461758, + 0.003336676862090826, + 0.007064003963023424, + 0.0060861362144351006, + 0.00463235005736351, + 0.0060489000752568245, + 0.006321161054074764, + 0.00470403116196394, + 0.005554964765906334, + 0.0035163138527423143, + 0.007922627031803131, + 0.006451200693845749, + 0.0059787980280816555, + 0.010539056733250618, + 0.011231006123125553, + 0.007004069164395332, + 0.008026622235774994, + 0.009776335209608078, + 0.014767765067517757, + 0.008602509275078773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.001522692502476275 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028234163764864206 + ], + "opposite_logprob_tokens": [ + -5.877823352813721 + ], + "logprob_reward": -0.0028234163764864206, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028234163764864206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31556856632232666, + 0.1031469777226448, + 0.17731362581253052, + 0.023632703348994255, + 0.03291787952184677, + 0.04369453340768814, + 0.020190536975860596, + 0.09773293882608414, + 0.06962981075048447, + 0.1161724403500557 + ], + "credit_map": [ + 0.013580607250332832, + 0.016335222870111465, + 0.012523991987109184, + 0.01040563452988863, + 0.12479943037033081, + 0.12298157811164856, + 0.006979739293456078, + 0.007962373085319996, + 0.009529790841042995, + 0.018399279564619064, + 0.012441097758710384, + 0.012547771446406841, + 0.01771676540374756, + 0.014907492324709892, + 0.007707345765084028, + 0.009897437877953053, + 0.008276082575321198, + 0.005529258865863085, + 0.06652951240539551, + 0.012631095014512539, + 0.007550741545855999, + 0.007825884036719799, + 0.010061333887279034, + 0.058909714221954346, + 0.002151931868866086, + 0.006086006294935942, + 0.0026049381121993065, + 0.0030773102771490812, + 0.004321912303566933, + 0.0020845483522862196, + 0.0014860326191410422, + 0.001820022240281105, + 0.0020145971793681383, + 0.005333065055310726, + 0.0075620487332344055, + 0.006122115068137646, + 0.005140781868249178, + 0.002241504145786166, + 0.0023281231988221407, + 0.0021756435744464397, + 0.0028079606126993895, + 0.008413475938141346, + 0.005311384331434965, + 0.006051470525562763, + 0.0044930302537977695, + 0.002848528092727065, + 0.009475139901041985, + 0.004293540492653847, + 0.001431082608178258, + 0.002050225855782628, + 0.0035344266798347235, + 0.0023803790099918842, + 0.003946606535464525, + 0.0016517661279067397, + 0.0029244350735098124, + 0.0022716147359460592, + 0.007995753549039364, + 0.015804531052708626, + 0.01647614687681198, + 0.011330227367579937, + 0.020970311015844345, + 0.005432770121842623, + 0.011042308062314987, + 0.00868089310824871, + 0.005351280327886343, + 0.014767373912036419, + 0.006758132949471474, + 0.007686654571443796, + 0.012685545720160007, + 0.003716646693646908, + 0.012273138388991356, + 0.0063910409808158875, + 0.010208397172391415, + 0.028315875679254532, + 0.008779691532254219, + 0.007805887144058943, + 0.01967223733663559, + 0.013735598884522915, + 0.016830673441290855, + 0.010824090801179409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.0028234163764864206 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019402741454541683 + ], + "opposite_logprob_tokens": [ + -6.2519402503967285 + ], + "logprob_reward": -0.0019402741454541683, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019402741454541683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0754329040646553, + 0.11499087512493134, + 0.08098921924829483, + 0.04328591004014015, + 0.03514746576547623, + 0.064956434071064, + 0.043040577322244644, + 0.30322328209877014, + 0.14740417897701263, + 0.09152916073799133 + ], + "credit_map": [ + 0.014921031892299652, + 0.013701926916837692, + 0.010760615579783916, + 0.008245275355875492, + 0.008432106114923954, + 0.008096808567643166, + 0.0056854174472391605, + 0.005589725915342569, + 0.010304752737283707, + 0.02339884266257286, + 0.01075646560639143, + 0.015366570092737675, + 0.01452262420207262, + 0.018908536061644554, + 0.010306120850145817, + 0.011426961049437523, + 0.003898490685969591, + 0.004456187132745981, + 0.027692416682839394, + 0.005868301261216402, + 0.0068809580989181995, + 0.004269914701581001, + 0.004632170312106609, + 0.023290786892175674, + 0.0033498515840619802, + 0.005573154427111149, + 0.003519053105264902, + 0.009342200122773647, + 0.011932216584682465, + 0.004062902182340622, + 0.002756283152848482, + 0.002750248881056905, + 0.0025689948815852404, + 0.006920434068888426, + 0.0068480777554214, + 0.0042391614988446236, + 0.006629184819757938, + 0.0025083196815103292, + 0.0030260789208114147, + 0.0024072150699794292, + 0.003201467916369438, + 0.01564740389585495, + 0.007114442531019449, + 0.008139673620462418, + 0.0076498230919241905, + 0.004081103019416332, + 0.012886040844023228, + 0.006236477289348841, + 0.0034050329122692347, + 0.005339208524674177, + 0.0061973389238119125, + 0.005760216154158115, + 0.010181905701756477, + 0.002893600380048156, + 0.004935309290885925, + 0.004327964503318071, + 0.03044743649661541, + 0.039929620921611786, + 0.058250658214092255, + 0.03377416729927063, + 0.05073856934905052, + 0.015957674011588097, + 0.04953543841838837, + 0.02458970621228218, + 0.015184429474174976, + 0.026732632890343666, + 0.019370192661881447, + 0.019270623102784157, + 0.020501205697655678, + 0.009937828406691551, + 0.019930832087993622, + 0.016476433724164963, + 0.008280163630843163, + 0.01059362106025219, + 0.011303485371172428, + 0.008278222754597664, + 0.008102389983832836, + 0.017128588631749153, + 0.01799720712006092, + 0.009845484979450703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.0019402741454541683 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036153923720121384 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.0036153923720121384, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036153923720121384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18503551185131073, + 0.02895135059952736, + 0.1121862456202507, + 0.017864584922790527, + 0.05045243352651596, + 0.0968669131398201, + 0.03977625444531441, + 0.335213303565979, + 0.058824703097343445, + 0.0748286321759224 + ], + "credit_map": [ + 0.012809165753424168, + 0.010551112703979015, + 0.007082867436110973, + 0.006388187408447266, + 0.06903672218322754, + 0.06978606432676315, + 0.004311388358473778, + 0.005070046056061983, + 0.003186935791745782, + 0.005954424384981394, + 0.0029405816458165646, + 0.003348709549754858, + 0.004703117068856955, + 0.0037463160697370768, + 0.0027830174658447504, + 0.0022882530465722084, + 0.0055273063480854034, + 0.0027635684236884117, + 0.04370589926838875, + 0.007916872389614582, + 0.004095274023711681, + 0.004776934161782265, + 0.005410197656601667, + 0.03799021616578102, + 0.001322745461948216, + 0.0028224363923072815, + 0.0015252428129315376, + 0.0036898665130138397, + 0.0034833094105124474, + 0.0018911219667643309, + 0.0018386059673503041, + 0.0012912607053294778, + 0.00534319132566452, + 0.01023909728974104, + 0.008450767025351524, + 0.005478611681610346, + 0.007181731052696705, + 0.004321052692830563, + 0.005004743114113808, + 0.0044332463294267654, + 0.00454715546220541, + 0.022144567221403122, + 0.010647796094417572, + 0.01194318663328886, + 0.010846571065485477, + 0.00601179851219058, + 0.021303419023752213, + 0.009422430768609047, + 0.003693251172080636, + 0.004558858461678028, + 0.0061597093008458614, + 0.00403058435767889, + 0.005001800134778023, + 0.003554765833541751, + 0.006259426940232515, + 0.006517861504107714, + 0.030728822574019432, + 0.04542525112628937, + 0.06373956054449081, + 0.03782803565263748, + 0.0630887821316719, + 0.018937943503260612, + 0.047665245831012726, + 0.027799731120467186, + 0.005036310292780399, + 0.006011215504258871, + 0.010570649057626724, + 0.007414016406983137, + 0.007527939975261688, + 0.004001314286142588, + 0.010547845624387264, + 0.007715418469160795, + 0.006332367658615112, + 0.009340590797364712, + 0.009248225018382072, + 0.008775279857218266, + 0.007640628144145012, + 0.010319427587091923, + 0.01478545367717743, + 0.008386674337089062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.0036153923720121384 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, extreme wide shot surveys a vast, sun-scorched desert plain under the harsh, white-out light of the midday sun. The parched, cracked earth stretches to distant, hazy mountains, and the air shimmers with a fiery glow. Suddenly, the ground begins to buckle and crack violently, a spiderweb of matter ripping outward from its center. With immense, relentless force, a massive, rectangular monolith rises from the earth, its craggy, weathered granite surface betraying immense age. Its catastrophic ascent shatters the surrounding terrain, hurling chunks of rock and soil into the pale blue sky. A dense, roiling maelstrom of pulverized rock and ochre dust rapidly expands, tumbling upward and outward, engulfing the base of the slab and unleashing a powerful, apocalyptic, unleashing of primordial earthly forces.", + "seed": 15972, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_2/index_8559_seed_15972_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.014920767466537654, + "vlm_logprob": -0.014920767466537654, + "vlm_yes_no_margin": 5.45, + "vlm_token_credit": -0.014920767466537654, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06347305327653885 + ], + "opposite_logprob_tokens": [ + -2.9384729862213135 + ], + "logprob_reward": -0.06347305327653885, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06347305327653885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09424358606338501, + 0.06978500634431839, + 0.0476142093539238, + 0.05092634633183479, + 0.03890771418809891, + 0.2619352638721466, + 0.12387765198945999, + 0.20946520566940308, + 0.0401647686958313, + 0.06308028101921082 + ], + "credit_map": [ + 0.00887156743556261, + 0.010479392483830452, + 0.009253954514861107, + 0.013848050497472286, + 0.008285468444228172, + 0.009008598513901234, + 0.02158288098871708, + 0.012913675978779793, + 0.006259229965507984, + 0.00617922842502594, + 0.013171326369047165, + 0.012034113518893719, + 0.009643365629017353, + 0.007244332227855921, + 0.007960627786815166, + 0.007292779162526131, + 0.0050557986833155155, + 0.004020072519779205, + 0.006104782689362764, + 0.005420488305389881, + 0.007800211664289236, + 0.00936896912753582, + 0.005036468617618084, + 0.004807420540601015, + 0.005170321092009544, + 0.006235555279999971, + 0.006493495311588049, + 0.006757121533155441, + 0.0070804329589009285, + 0.0060131424106657505, + 0.006765746045857668, + 0.006410531233996153, + 0.0030673553701490164, + 0.0032791767735034227, + 0.00548505550250411, + 0.0036951377987861633, + 0.003918295726180077, + 0.009093939326703548, + 0.004951026290655136, + 0.005417726468294859, + 0.020917750895023346, + 0.10796141624450684, + 0.0073090060614049435, + 0.011012270115315914, + 0.0206553116440773, + 0.07785128802061081, + 0.009675470180809498, + 0.006552731152623892, + 0.01836647279560566, + 0.025437474250793457, + 0.00944177620112896, + 0.011221951805055141, + 0.02043772116303444, + 0.022164389491081238, + 0.008458188734948635, + 0.008349677547812462, + 0.030250944197177887, + 0.04108482599258423, + 0.014425245113670826, + 0.02078075520694256, + 0.035831108689308167, + 0.025496693328022957, + 0.016169793903827667, + 0.025425825268030167, + 0.0046012066304683685, + 0.007713729981333017, + 0.004330575931817293, + 0.004784984979778528, + 0.00598944490775466, + 0.005724444054067135, + 0.003333801869302988, + 0.003686582436785102, + 0.008114606142044067, + 0.00864206999540329, + 0.009144588373601437, + 0.005905123893171549, + 0.007274624425917864, + 0.008222502656280994, + 0.006421792786568403, + 0.009354970417916775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.06347305327653885 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010531361913308501 + ], + "opposite_logprob_tokens": [ + -6.876053333282471 + ], + "logprob_reward": -0.0010531361913308501, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010531361913308501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15724341571331024, + 0.09390461444854736, + 0.03517835959792137, + 0.07135391235351562, + 0.05445243790745735, + 0.11212436109781265, + 0.15828491747379303, + 0.10446646064519882, + 0.15213645994663239, + 0.06085505709052086 + ], + "credit_map": [ + 0.01187098864465952, + 0.01118768472224474, + 0.01751798577606678, + 0.021876152604818344, + 0.026567475870251656, + 0.02998330444097519, + 0.02025970257818699, + 0.01798010990023613, + 0.008292801678180695, + 0.01099742017686367, + 0.009922651574015617, + 0.013166879303753376, + 0.02184666506946087, + 0.014150493778288364, + 0.007811603602021933, + 0.007716100197285414, + 0.0025850473903119564, + 0.003659514943137765, + 0.005874378141015768, + 0.004661643411964178, + 0.005609574262052774, + 0.005855240859091282, + 0.0033175034914165735, + 0.0036154561676084995, + 0.011166995391249657, + 0.00678041810169816, + 0.008264503441751003, + 0.007997828535735607, + 0.006737162359058857, + 0.01263890229165554, + 0.008418369106948376, + 0.009349734522402287, + 0.004424941726028919, + 0.004697917494922876, + 0.007990291342139244, + 0.0047333273105323315, + 0.006453070789575577, + 0.01124813687056303, + 0.007759439758956432, + 0.007145313546061516, + 0.007415670435875654, + 0.03676539286971092, + 0.007441749330610037, + 0.006743798032402992, + 0.012771236710250378, + 0.026279496029019356, + 0.009256808087229729, + 0.005450207740068436, + 0.023354018107056618, + 0.032657988369464874, + 0.011712661944329739, + 0.01452037412673235, + 0.02647135779261589, + 0.02797563374042511, + 0.011043618433177471, + 0.010549275204539299, + 0.014290881343185902, + 0.019495531916618347, + 0.009562436491250992, + 0.012312301434576511, + 0.015523712150752544, + 0.011681408621370792, + 0.008704687468707561, + 0.012895504012703896, + 0.01729670912027359, + 0.018482523038983345, + 0.023667821660637856, + 0.021808357909321785, + 0.014359495602548122, + 0.02503691427409649, + 0.014786834828555584, + 0.016697809100151062, + 0.007233115844428539, + 0.007105898577719927, + 0.010277727618813515, + 0.0073622954078018665, + 0.005995668470859528, + 0.005664465948939323, + 0.0065376306883990765, + 0.010678254067897797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.0010531361913308501 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410214951261878 + ], + "opposite_logprob_tokens": [ + -5.5041022300720215 + ], + "logprob_reward": -0.00410214951261878, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410214951261878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06127442792057991, + 0.042676158249378204, + 0.03347005322575569, + 0.01269385777413845, + 0.017427796497941017, + 0.1022711843252182, + 0.022615278139710426, + 0.5871334075927734, + 0.07191295176744461, + 0.04852481931447983 + ], + "credit_map": [ + 0.00579798175022006, + 0.005154732149094343, + 0.0069513400085270405, + 0.0059692999348044395, + 0.010316155850887299, + 0.012748412787914276, + 0.007185475900769234, + 0.007151034660637379, + 0.003684369381517172, + 0.0049260458908975124, + 0.004471529275178909, + 0.005486497189849615, + 0.009438924491405487, + 0.006716684903949499, + 0.0038404646329581738, + 0.004111648537218571, + 0.0038928152061998844, + 0.002652572700753808, + 0.005670438054949045, + 0.004723262973129749, + 0.005154195707291365, + 0.00496148644015193, + 0.0027383463457226753, + 0.003676940919831395, + 0.001465407432988286, + 0.0015893876552581787, + 0.002029780764132738, + 0.0015910824295133352, + 0.0016562098171561956, + 0.0013511953875422478, + 0.0014794343151152134, + 0.0015313614858314395, + 0.0017857523635029793, + 0.0016653187340125442, + 0.003028697334229946, + 0.0014140369603410363, + 0.0018589216051623225, + 0.0038853499572724104, + 0.001997199608013034, + 0.0017925221472978592, + 0.005481056869029999, + 0.035043660551309586, + 0.005483120214194059, + 0.005459963344037533, + 0.009127393364906311, + 0.029832666739821434, + 0.00743288267403841, + 0.004410445224493742, + 0.003628962906077504, + 0.004005596973001957, + 0.0017650979571044445, + 0.002894229954108596, + 0.0024512577801942825, + 0.004437166731804609, + 0.001541414181701839, + 0.0018915545661002398, + 0.07624462246894836, + 0.1274113953113556, + 0.04345056042075157, + 0.058735091239213943, + 0.10837560892105103, + 0.06516434997320175, + 0.04274999350309372, + 0.06500177830457687, + 0.0077776531688869, + 0.007947136647999287, + 0.011370651423931122, + 0.011085771955549717, + 0.0066336351446807384, + 0.011145113967359066, + 0.007191897369921207, + 0.008761103264987469, + 0.006770234555006027, + 0.007397180888801813, + 0.006533072330057621, + 0.0047549111768603325, + 0.005408796016126871, + 0.005241651087999344, + 0.005078226327896118, + 0.007340752519667149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.00410214951261878 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013368725776672363 + ], + "opposite_logprob_tokens": [ + -6.626337051391602 + ], + "logprob_reward": -0.0013368725776672363, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013368725776672363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10207676887512207, + 0.03530508279800415, + 0.05138545110821724, + 0.020092148333787918, + 0.05020510405302048, + 0.12478313595056534, + 0.13605770468711853, + 0.24620670080184937, + 0.18176935613155365, + 0.052118465304374695 + ], + "credit_map": [ + 0.008243118412792683, + 0.011324848048388958, + 0.009283370338380337, + 0.012755473144352436, + 0.012045514769852161, + 0.01341440249234438, + 0.017153669148683548, + 0.017856379970908165, + 0.0027267667464911938, + 0.003520875470712781, + 0.004585746210068464, + 0.005717750638723373, + 0.006544166244566441, + 0.005137321073561907, + 0.003302584867924452, + 0.003769876668229699, + 0.005246181972324848, + 0.0037207971327006817, + 0.009407307021319866, + 0.007946046069264412, + 0.007421680260449648, + 0.006570706143975258, + 0.004653588403016329, + 0.006419152021408081, + 0.0026601043064147234, + 0.0023769917897880077, + 0.0025666027795523405, + 0.002184727694839239, + 0.0031073549762368202, + 0.0022852683905512094, + 0.0025283668655902147, + 0.002382732927799225, + 0.004140508361160755, + 0.004595935810357332, + 0.00831290427595377, + 0.005221059545874596, + 0.005925895180553198, + 0.006899987813085318, + 0.007486064452677965, + 0.007622753735631704, + 0.007800230756402016, + 0.03811468556523323, + 0.008399068377912045, + 0.007734977174550295, + 0.01504501048475504, + 0.027407852932810783, + 0.013477189466357231, + 0.006804133765399456, + 0.02020716853439808, + 0.028392566367983818, + 0.010502936318516731, + 0.010904577560722828, + 0.02351558953523636, + 0.023814452812075615, + 0.009570937603712082, + 0.00914948433637619, + 0.030630236491560936, + 0.04971415176987648, + 0.023574089631438255, + 0.028426511213183403, + 0.040294673293828964, + 0.024610258638858795, + 0.02086036652326584, + 0.028096430003643036, + 0.02002882771193981, + 0.022407567128539085, + 0.027808889746665955, + 0.02790234610438347, + 0.01691318489611149, + 0.027282552793622017, + 0.018850406631827354, + 0.020575596019625664, + 0.007456403225660324, + 0.00668693520128727, + 0.007365940138697624, + 0.00599557813256979, + 0.0049455189146101475, + 0.007047464605420828, + 0.005470849107950926, + 0.00714977877214551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.0013368725776672363 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046386257745325565 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.0046386257745325565, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046386257745325565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06954925507307053, + 0.04252523183822632, + 0.038230471312999725, + 0.013587060384452343, + 0.012490964494645596, + 0.0482976995408535, + 0.1402243673801422, + 0.5165185332298279, + 0.04717429727315903, + 0.0714021697640419 + ], + "credit_map": [ + 0.0061728693544864655, + 0.006042486522346735, + 0.0063973646610975266, + 0.00824967585504055, + 0.012044504284858704, + 0.014389988034963608, + 0.008936189115047455, + 0.00731618283316493, + 0.004121383652091026, + 0.004088800400495529, + 0.005156429950147867, + 0.007259640842676163, + 0.007031767629086971, + 0.006148878950625658, + 0.0038239203859120607, + 0.004894411191344261, + 0.004215432796627283, + 0.0036531283985823393, + 0.007407786324620247, + 0.005705714225769043, + 0.004866380710154772, + 0.004021350760012865, + 0.0036239877808839083, + 0.00473669171333313, + 0.0013082388322800398, + 0.0017663359176367521, + 0.0016940225614234805, + 0.0015479816356673837, + 0.002016212325543165, + 0.0016438303282484412, + 0.0019753845408558846, + 0.0016350538935512304, + 0.001213939511217177, + 0.0015100667951628566, + 0.0022662191186100245, + 0.0012705750996246934, + 0.0014404394896700978, + 0.0014287367230281234, + 0.001846907427534461, + 0.001514080329798162, + 0.005046243313699961, + 0.018696991726756096, + 0.002581847831606865, + 0.0025783732999116182, + 0.0054183476604521275, + 0.008528724312782288, + 0.0036838531959801912, + 0.0017633159877732396, + 0.02667316235601902, + 0.02589712291955948, + 0.008950658142566681, + 0.010251739993691444, + 0.02072807401418686, + 0.032265644520521164, + 0.007506020367145538, + 0.007951940409839153, + 0.06754367053508759, + 0.11131187528371811, + 0.03848865255713463, + 0.05317841097712517, + 0.09380145370960236, + 0.05745220184326172, + 0.03788524121046066, + 0.05685701593756676, + 0.005949538201093674, + 0.00984180998057127, + 0.005109601654112339, + 0.0052848514169454575, + 0.006330958567559719, + 0.007047549355775118, + 0.004261920228600502, + 0.003348067868500948, + 0.009560765698552132, + 0.007685214281082153, + 0.011144786141812801, + 0.008942992426455021, + 0.0072952029295265675, + 0.007988808676600456, + 0.008907798677682877, + 0.009876597672700882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.0046386257745325565 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, extreme wide shot surveys a vast, sun-scorched desert plain under the harsh, white-out light of the midday sun. The parched, cracked earth stretches to distant, hazy mountains, and the air shimmers with a fiery glow. Suddenly, the ground begins to buckle and crack violently, a spiderweb of matter ripping outward from its center. With immense, relentless force, a massive, rectangular monolith rises from the earth, its craggy, weathered granite surface betraying immense age. Its catastrophic ascent shatters the surrounding terrain, hurling chunks of rock and soil into the pale blue sky. A dense, roiling maelstrom of pulverized rock and ochre dust rapidly expands, tumbling upward and outward, engulfing the base of the slab and unleashing a powerful, apocalyptic, unleashing of primordial earthly forces.", + "seed": 15973, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_2/index_8559_seed_15973_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04331601054873317, + "vlm_logprob": -0.04331601054873317, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.04331601054873317, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17206987738609314 + ], + "opposite_logprob_tokens": [ + -1.9220699071884155 + ], + "logprob_reward": -0.17206987738609314, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17206987738609314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1835610717535019, + 0.05303594470024109, + 0.07321049273014069, + 0.10207824409008026, + 0.049585651606321335, + 0.09365785121917725, + 0.04109003767371178, + 0.05198846012353897, + 0.2614094913005829, + 0.09038276970386505 + ], + "credit_map": [ + 0.016942370682954788, + 0.020857399329543114, + 0.016557898372411728, + 0.015959417447447777, + 0.05149620398879051, + 0.029256753623485565, + 0.011458645574748516, + 0.021032370626926422, + 0.007414286956191063, + 0.004649963229894638, + 0.004123812075704336, + 0.0038846516981720924, + 0.01385475043207407, + 0.009547443129122257, + 0.003612678498029709, + 0.005948362872004509, + 0.012288675643503666, + 0.009472609497606754, + 0.0067893462255597115, + 0.007400942035019398, + 0.009904076345264912, + 0.01402756292372942, + 0.006185123696923256, + 0.007142157759517431, + 0.013501537963747978, + 0.008482447825372219, + 0.007560447324067354, + 0.009639140218496323, + 0.02009548619389534, + 0.02280963957309723, + 0.010606121271848679, + 0.00938341859728098, + 0.005029449239373207, + 0.008098672144114971, + 0.005566082429140806, + 0.006696103606373072, + 0.0054429080337285995, + 0.005958558991551399, + 0.007105338387191296, + 0.005688541103154421, + 0.006866666954010725, + 0.0069492231123149395, + 0.010500824078917503, + 0.013327674008905888, + 0.008410933427512646, + 0.004654530435800552, + 0.031232304871082306, + 0.011715693399310112, + 0.0031982401851564646, + 0.0040367660112679005, + 0.005509390961378813, + 0.009066212922334671, + 0.0047308108769357204, + 0.0027852077037096024, + 0.006899786647409201, + 0.0048636216670274734, + 0.003594779409468174, + 0.006358197890222073, + 0.00899041723459959, + 0.008994270116090775, + 0.00789432879537344, + 0.004052224103361368, + 0.0068455711007118225, + 0.005258668679744005, + 0.030854351818561554, + 0.03571566194295883, + 0.027859153226017952, + 0.030176028609275818, + 0.03680099546909332, + 0.03652627766132355, + 0.036647386848926544, + 0.02682962268590927, + 0.016996869817376137, + 0.010525098070502281, + 0.0072660669684410095, + 0.006314052268862724, + 0.0235099196434021, + 0.0075237383134663105, + 0.010042703710496426, + 0.008204318583011627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.17206987738609314 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015265013789758086 + ], + "opposite_logprob_tokens": [ + -6.501526355743408 + ], + "logprob_reward": -0.0015265013789758086, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015265013789758086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3488147258758545, + 0.06670398265123367, + 0.08079968392848969, + 0.05118449404835701, + 0.07120119780302048, + 0.046554334461688995, + 0.034708112478256226, + 0.0597270168364048, + 0.1927676647901535, + 0.047538794577121735 + ], + "credit_map": [ + 0.043290335685014725, + 0.049577392637729645, + 0.030910704284906387, + 0.03461112827062607, + 0.07554550468921661, + 0.05340675264596939, + 0.022464977577328682, + 0.03900792822241783, + 0.007776170969009399, + 0.00814689602702856, + 0.007145363837480545, + 0.007925035431981087, + 0.00975289661437273, + 0.010174091905355453, + 0.0055421264842152596, + 0.010241400450468063, + 0.013414710760116577, + 0.009994050487875938, + 0.00938073918223381, + 0.008404548279941082, + 0.009801207110285759, + 0.015454613603651524, + 0.00600647134706378, + 0.008343341760337353, + 0.006988118402659893, + 0.003741962369531393, + 0.006320044863969088, + 0.006212309468537569, + 0.006698017939925194, + 0.00949010904878378, + 0.006898802705109119, + 0.004835130181163549, + 0.006434230599552393, + 0.007226717658340931, + 0.005059957969933748, + 0.023488277569413185, + 0.003942125476896763, + 0.004496896639466286, + 0.013522189110517502, + 0.007030809763818979, + 0.0035755313001573086, + 0.005011172965168953, + 0.0061892736703157425, + 0.007294714916497469, + 0.004665487911552191, + 0.003866804763674736, + 0.0067510767839848995, + 0.009200274012982845, + 0.0029288812074810266, + 0.003687339136376977, + 0.006314829923212528, + 0.005707724019885063, + 0.0038624517619609833, + 0.002928625326603651, + 0.004856611602008343, + 0.004421650432050228, + 0.006167476996779442, + 0.004517419263720512, + 0.007090134546160698, + 0.006572594400495291, + 0.009515570476651192, + 0.003728891257196665, + 0.017106816172599792, + 0.00502811511978507, + 0.021219657734036446, + 0.027230525389313698, + 0.019983571022748947, + 0.021515950560569763, + 0.024699395522475243, + 0.024087684229016304, + 0.0331435464322567, + 0.020887315273284912, + 0.007116428576409817, + 0.0063085295259952545, + 0.004693513736128807, + 0.0035695114638656378, + 0.010872195474803448, + 0.003955196589231491, + 0.005770673044025898, + 0.005252745468169451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.0015265013789758086 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046457452699542046 + ], + "opposite_logprob_tokens": [ + -5.379645824432373 + ], + "logprob_reward": -0.0046457452699542046, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046457452699542046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45131614804267883, + 0.04671616479754448, + 0.05510050430893898, + 0.030124377459287643, + 0.06610588729381561, + 0.045843303203582764, + 0.064754918217659, + 0.05397247150540352, + 0.14507663249969482, + 0.040989652276039124 + ], + "credit_map": [ + 0.058272834867239, + 0.062360089272260666, + 0.04491494968533516, + 0.03312067687511444, + 0.09030873328447342, + 0.08694971352815628, + 0.02648044191300869, + 0.04890870302915573, + 0.005181880202144384, + 0.0054220412857830524, + 0.006056167185306549, + 0.006886353250592947, + 0.005753831006586552, + 0.006774598732590675, + 0.005237532779574394, + 0.0054037612862885, + 0.006128591950982809, + 0.008511155843734741, + 0.0053347391076385975, + 0.0061114197596907616, + 0.007341628894209862, + 0.010266068391501904, + 0.005480653140693903, + 0.0059262486174702644, + 0.0024261625949293375, + 0.002843249123543501, + 0.003847783198580146, + 0.004178406670689583, + 0.0051073781214654446, + 0.004238140769302845, + 0.005449908785521984, + 0.0020333481952548027, + 0.004328874405473471, + 0.006721630226820707, + 0.004751980770379305, + 0.019915418699383736, + 0.00488988496363163, + 0.006125612184405327, + 0.013278443366289139, + 0.006094042211771011, + 0.004541119560599327, + 0.00423200661316514, + 0.004378417041152716, + 0.0058326516300439835, + 0.003689773380756378, + 0.0042577628046274185, + 0.012549607083201408, + 0.006361966021358967, + 0.005673267412930727, + 0.00929157342761755, + 0.007458201609551907, + 0.011132653802633286, + 0.010705489665269852, + 0.006398720666766167, + 0.007763449568301439, + 0.006331561598926783, + 0.0046525984071195126, + 0.004722645506262779, + 0.007381031289696693, + 0.0073638432659208775, + 0.007275396957993507, + 0.003795773722231388, + 0.013845734298229218, + 0.0049354503862559795, + 0.015950772911310196, + 0.020405065268278122, + 0.013224397785961628, + 0.01786237396299839, + 0.019604433327913284, + 0.0193992480635643, + 0.02139013260602951, + 0.01724020391702652, + 0.005904024001210928, + 0.0038988336455076933, + 0.003763194428756833, + 0.00429568812251091, + 0.008427348919212818, + 0.0035240079741925, + 0.004699296783655882, + 0.006477257702499628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.0046457452699542046 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640405531972647 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.004640405531972647, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640405531972647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1284528225660324, + 0.12271909415721893, + 0.12928007543087006, + 0.09929012507200241, + 0.03341164067387581, + 0.02735568769276142, + 0.05274641886353493, + 0.1977171152830124, + 0.13490581512451172, + 0.07412119209766388 + ], + "credit_map": [ + 0.016127239912748337, + 0.01828392967581749, + 0.013023690320551395, + 0.009717808105051517, + 0.02063385210931301, + 0.02372482791543007, + 0.010184749029576778, + 0.016756724566221237, + 0.012271687388420105, + 0.011946289800107479, + 0.013561829924583435, + 0.020215826109051704, + 0.019136741757392883, + 0.016537124291062355, + 0.012058751657605171, + 0.016990836709737778, + 0.01752905175089836, + 0.01789732836186886, + 0.01114539336413145, + 0.01353373657912016, + 0.016565505415201187, + 0.028366612270474434, + 0.011475909501314163, + 0.012766536325216293, + 0.008029620163142681, + 0.010019331239163876, + 0.0096151577308774, + 0.01160398032516241, + 0.020967986434698105, + 0.018618687987327576, + 0.013706264086067677, + 0.006729098968207836, + 0.002387427259236574, + 0.0036799502559006214, + 0.00367514556273818, + 0.007088100537657738, + 0.002825598232448101, + 0.0034722955897450447, + 0.006671161390841007, + 0.0036119644064456224, + 0.002105533378198743, + 0.0027859932743012905, + 0.0034174825996160507, + 0.0033408759627491236, + 0.003006461076438427, + 0.002865967806428671, + 0.0054999846033751965, + 0.004333389922976494, + 0.004122315440326929, + 0.004752617795020342, + 0.009534734301269054, + 0.007940360344946384, + 0.0064151412807404995, + 0.004980097990483046, + 0.008301738649606705, + 0.006699410267174244, + 0.022090092301368713, + 0.013035118579864502, + 0.024882547557353973, + 0.02172103524208069, + 0.02400178648531437, + 0.011609806679189205, + 0.06368675827980042, + 0.016689972952008247, + 0.011984705924987793, + 0.017085373401641846, + 0.012941209599375725, + 0.01730780489742756, + 0.01818779669702053, + 0.015126696787774563, + 0.027808543294668198, + 0.014463677071034908, + 0.01723211258649826, + 0.007733410689979792, + 0.004495449364185333, + 0.004432200454175472, + 0.022488318383693695, + 0.005940670613199472, + 0.006399461533874273, + 0.005399574991315603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.004640405531972647 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033697523176670074 + ], + "opposite_logprob_tokens": [ + -3.4086976051330566 + ], + "logprob_reward": -0.033697523176670074, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033697523176670074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1473129242658615, + 0.07343387603759766, + 0.2541359066963196, + 0.05391212925314903, + 0.0695655569434166, + 0.028752999380230904, + 0.026379114016890526, + 0.08589214831590652, + 0.11098412424325943, + 0.14963127672672272 + ], + "credit_map": [ + 0.01879444159567356, + 0.020644741132855415, + 0.014099610969424248, + 0.011034985072910786, + 0.027194099500775337, + 0.032054126262664795, + 0.010417718440294266, + 0.01307318639010191, + 0.007248631212860346, + 0.0064570363610982895, + 0.007410632446408272, + 0.006179322022944689, + 0.01642511412501335, + 0.015796253457665443, + 0.00383606250397861, + 0.010080823674798012, + 0.029339447617530823, + 0.04042529687285423, + 0.02333899959921837, + 0.03137131780385971, + 0.031501464545726776, + 0.046189650893211365, + 0.026247547939419746, + 0.025722166523337364, + 0.007017113268375397, + 0.004021984525024891, + 0.004808987956494093, + 0.006303854286670685, + 0.00853632390499115, + 0.010951049625873566, + 0.006246591452509165, + 0.006026218645274639, + 0.007219911552965641, + 0.0047295233234763145, + 0.004615641199052334, + 0.02416059374809265, + 0.0029433961026370525, + 0.0042603760957717896, + 0.015007655136287212, + 0.0066284541971981525, + 0.002407758729532361, + 0.003202266525477171, + 0.0026685132179409266, + 0.004045654088258743, + 0.002433583838865161, + 0.002843610942363739, + 0.00562261464074254, + 0.005528997164219618, + 0.0030897518154233694, + 0.003156989347189665, + 0.004113255999982357, + 0.00416655233129859, + 0.003761082189157605, + 0.002405675360932946, + 0.003122382564470172, + 0.0025634211488068104, + 0.012500532902777195, + 0.004391607362776995, + 0.00574010843411088, + 0.008209197781980038, + 0.010785575024783611, + 0.0027613185811787844, + 0.033523645251989365, + 0.007980156689882278, + 0.014323389157652855, + 0.012389843352138996, + 0.015646519139409065, + 0.01186168659478426, + 0.015955286100506783, + 0.01685846596956253, + 0.012419040314853191, + 0.0115298917517066, + 0.02989654615521431, + 0.020241592079401016, + 0.008736691437661648, + 0.008512970991432667, + 0.04454952850937843, + 0.012997659854590893, + 0.013754173181951046, + 0.010942107066512108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.033697523176670074 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, extreme wide shot surveys a vast, sun-scorched desert plain under the harsh, white-out light of the midday sun. The parched, cracked earth stretches to distant, hazy mountains, and the air shimmers with a fiery glow. Suddenly, the ground begins to buckle and crack violently, a spiderweb of matter ripping outward from its center. With immense, relentless force, a massive, rectangular monolith rises from the earth, its craggy, weathered granite surface betraying immense age. Its catastrophic ascent shatters the surrounding terrain, hurling chunks of rock and soil into the pale blue sky. A dense, roiling maelstrom of pulverized rock and ochre dust rapidly expands, tumbling upward and outward, engulfing the base of the slab and unleashing a powerful, apocalyptic, unleashing of primordial earthly forces.", + "seed": 15974, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_2/index_8559_seed_15974_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.013765753153711558, + "vlm_logprob": -0.013765753153711558, + "vlm_yes_no_margin": 5.725, + "vlm_token_credit": -0.013765753153711558, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.059650614857673645 + ], + "opposite_logprob_tokens": [ + -3.0596506595611572 + ], + "logprob_reward": -0.059650614857673645, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.059650614857673645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11445560306310654, + 0.06633349508047104, + 0.0761827751994133, + 0.06584075093269348, + 0.01717344857752323, + 0.01599479652941227, + 0.032740268856287, + 0.096522256731987, + 0.05271293222904205, + 0.4620437026023865 + ], + "credit_map": [ + 0.010790452361106873, + 0.012278524227440357, + 0.01935015432536602, + 0.01999393105506897, + 0.011653912253677845, + 0.011791842058300972, + 0.013195820152759552, + 0.015400958247482777, + 0.006070438772439957, + 0.014392473734915257, + 0.0036953140515834093, + 0.0053490749560296535, + 0.015942934900522232, + 0.00727264117449522, + 0.007801630534231663, + 0.005808990448713303, + 0.005979611538350582, + 0.014503183774650097, + 0.008180230855941772, + 0.011716925539076328, + 0.012085200287401676, + 0.005842133425176144, + 0.010334833525121212, + 0.0075406599789857864, + 0.005111804232001305, + 0.005374044645577669, + 0.0040696426294744015, + 0.003497558878734708, + 0.02161470614373684, + 0.009596796706318855, + 0.004905696492642164, + 0.011670495383441448, + 0.002256452338770032, + 0.001759219914674759, + 0.0019101090729236603, + 0.0026180220302194357, + 0.002177236834540963, + 0.0019972079899162054, + 0.0024341444950550795, + 0.0020210554357618093, + 0.002445600926876068, + 0.0018230214482173324, + 0.0017399019561707973, + 0.0013894805451855063, + 0.002483863616362214, + 0.0017278199084103107, + 0.002567842835560441, + 0.0018172638956457376, + 0.0023624831810593605, + 0.003177511040121317, + 0.003021583193913102, + 0.0030452159699052572, + 0.007797892205417156, + 0.006443403195589781, + 0.003346024313941598, + 0.0035461559891700745, + 0.007739100139588118, + 0.006690173409879208, + 0.008152628317475319, + 0.0053239320404827595, + 0.029671842232346535, + 0.0222103763371706, + 0.009435877203941345, + 0.007298331707715988, + 0.005300827790051699, + 0.010256425477564335, + 0.004768881015479565, + 0.004794815089553595, + 0.0054046763107180595, + 0.007327561266720295, + 0.008142976090312004, + 0.006716770585626364, + 0.014404200948774815, + 0.021248498931527138, + 0.048407651484012604, + 0.1459505707025528, + 0.10567409545183182, + 0.052073877304792404, + 0.06088817119598389, + 0.013396638445556164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.059650614857673645 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006466205231845379 + ], + "opposite_logprob_tokens": [ + -7.375646591186523 + ], + "logprob_reward": -0.0006466205231845379, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006466205231845379, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26783376932144165, + 0.04053047299385071, + 0.05986282229423523, + 0.11892936378717422, + 0.01839539408683777, + 0.015084287151694298, + 0.04654119536280632, + 0.020274652168154716, + 0.051615845412015915, + 0.36093220114707947 + ], + "credit_map": [ + 0.020978812128305435, + 0.03261614590883255, + 0.04957617074251175, + 0.03586956113576889, + 0.03475460410118103, + 0.027135128155350685, + 0.03678043931722641, + 0.030122892931103706, + 0.004450105130672455, + 0.005702026188373566, + 0.0037987162359058857, + 0.004356789402663708, + 0.00572992954403162, + 0.00610440643504262, + 0.005678270477801561, + 0.004710228648036718, + 0.005511316936463118, + 0.01595725491642952, + 0.004700922407209873, + 0.004847399890422821, + 0.012973101809620857, + 0.006093773525208235, + 0.005394581239670515, + 0.004384469240903854, + 0.008824891410768032, + 0.007678828202188015, + 0.004681965336203575, + 0.004254854284226894, + 0.0441797710955143, + 0.02538393996655941, + 0.005250910762697458, + 0.018674205988645554, + 0.001766819041222334, + 0.0022000365424901247, + 0.0024649787228554487, + 0.002102365717291832, + 0.002260772977024317, + 0.002115311799570918, + 0.0027256004977971315, + 0.002759508788585663, + 0.0019263281719759107, + 0.0016376231797039509, + 0.001182335545308888, + 0.0019365099724382162, + 0.002430735621601343, + 0.0022528590634465218, + 0.0016120949294418097, + 0.002105800900608301, + 0.004786630626767874, + 0.004126961808651686, + 0.002709585940465331, + 0.0022272660862654448, + 0.012519300915300846, + 0.0125119024887681, + 0.0033957173582166433, + 0.004263831302523613, + 0.0021737180650234222, + 0.002617830643430352, + 0.002254290971904993, + 0.0018634910229593515, + 0.0039566815830767155, + 0.002676105359569192, + 0.0025185502599924803, + 0.0022139863576740026, + 0.00399100873619318, + 0.010652225464582443, + 0.004827592521905899, + 0.006532884202897549, + 0.006364595610648394, + 0.0038982888218015432, + 0.008370981551706791, + 0.006978269200772047, + 0.011086666025221348, + 0.01657325029373169, + 0.03813907131552696, + 0.1163204163312912, + 0.08089026808738708, + 0.039401352405548096, + 0.048013437539339066, + 0.010507725179195404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.0006466205231845379 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041004871018230915 + ], + "opposite_logprob_tokens": [ + -5.504100322723389 + ], + "logprob_reward": -0.0041004871018230915, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041004871018230915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14714577794075012, + 0.03884818032383919, + 0.2131357491016388, + 0.1421416848897934, + 0.050749052315950394, + 0.049777623265981674, + 0.04894045367836952, + 0.03012922964990139, + 0.08827434480190277, + 0.1908578872680664 + ], + "credit_map": [ + 0.009420900605618954, + 0.01627306453883648, + 0.02096264809370041, + 0.029507452622056007, + 0.014635402709245682, + 0.014123940840363503, + 0.027443978935480118, + 0.01477839145809412, + 0.005314272828400135, + 0.0054813772439956665, + 0.0038515848573297262, + 0.004223514813929796, + 0.004660065285861492, + 0.00462342519313097, + 0.005020491313189268, + 0.005673449020832777, + 0.01829209178686142, + 0.05822896957397461, + 0.013254652731120586, + 0.019529473036527634, + 0.06383044272661209, + 0.013639404438436031, + 0.01496527623385191, + 0.011395446956157684, + 0.01586569845676422, + 0.008763023652136326, + 0.005933261942118406, + 0.0051911333575844765, + 0.041426341980695724, + 0.0492972694337368, + 0.004683435894548893, + 0.010981522500514984, + 0.005790867377072573, + 0.005948692560195923, + 0.006316904444247484, + 0.0068979389034211636, + 0.006019162014126778, + 0.00662252027541399, + 0.007648524362593889, + 0.005504439119249582, + 0.004431934095919132, + 0.007245425600558519, + 0.0046889143995940685, + 0.00739516643807292, + 0.008447890169918537, + 0.004921790678054094, + 0.007074940949678421, + 0.0055715590715408325, + 0.004363575950264931, + 0.005384455434978008, + 0.005463667679578066, + 0.008469580672681332, + 0.007023313548415899, + 0.0048622493632137775, + 0.007679774425923824, + 0.005693834275007248, + 0.002190749393776059, + 0.004249211400747299, + 0.003797050565481186, + 0.0023194439709186554, + 0.004996981471776962, + 0.004432188346982002, + 0.004120890982449055, + 0.0040227132849395275, + 0.010715574957430363, + 0.016489028930664062, + 0.009271617978811264, + 0.011959264054894447, + 0.008036386221647263, + 0.006133845075964928, + 0.013667064718902111, + 0.012001561932265759, + 0.012292662635445595, + 0.02593458816409111, + 0.020252933725714684, + 0.04050862416625023, + 0.03714408725500107, + 0.018138019368052483, + 0.025276586413383484, + 0.011310381814837456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.0041004871018230915 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008142255246639252 + ], + "opposite_logprob_tokens": [ + -7.125814437866211 + ], + "logprob_reward": -0.0008142255246639252, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008142255246639252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06848122924566269, + 0.011981074698269367, + 0.03557004779577255, + 0.1352158933877945, + 0.030637482181191444, + 0.020550517365336418, + 0.0499526709318161, + 0.029967283830046654, + 0.08267825841903687, + 0.5349655151367188 + ], + "credit_map": [ + 0.006452311296015978, + 0.008648992516100407, + 0.017406664788722992, + 0.005023969803005457, + 0.008266685530543327, + 0.007848342880606651, + 0.005036795046180487, + 0.009797472506761551, + 0.0012387109454721212, + 0.0015775938518345356, + 0.001465639565140009, + 0.0008307854295708239, + 0.0016094580059871078, + 0.0018657008185982704, + 0.0014463277766481042, + 0.0019468575483188033, + 0.0028158046770840883, + 0.008562668226659298, + 0.0029861442744731903, + 0.0035650632344186306, + 0.009118882939219475, + 0.003018961288034916, + 0.0031548822298645973, + 0.002347636967897415, + 0.01429319940507412, + 0.008527101948857307, + 0.005468982737511396, + 0.005763653665781021, + 0.04081527143716812, + 0.04529435560107231, + 0.005276304669678211, + 0.009777038358151913, + 0.0038344007916748524, + 0.0035607379395514727, + 0.003318841103464365, + 0.0034264163114130497, + 0.0041870735585689545, + 0.003705371171236038, + 0.0044171810150146484, + 0.004187460523098707, + 0.002384219318628311, + 0.002924310276284814, + 0.002133928472176194, + 0.003068298799917102, + 0.003042479744181037, + 0.0016739654820412397, + 0.0025639007799327374, + 0.0027594135608524084, + 0.007191882934421301, + 0.0056451670825481415, + 0.004337341990321875, + 0.004481821786612272, + 0.010839171707630157, + 0.00960561353713274, + 0.0047140903770923615, + 0.003137583378702402, + 0.0025229586753994226, + 0.005768562667071819, + 0.003464299254119396, + 0.003144618822261691, + 0.003743790788576007, + 0.002510590711608529, + 0.004892366472631693, + 0.00392009736970067, + 0.008478758856654167, + 0.011232045479118824, + 0.009440536610782146, + 0.013069829903542995, + 0.008024794980883598, + 0.006635111756622791, + 0.012691531330347061, + 0.013105648569762707, + 0.01666789874434471, + 0.024905700236558914, + 0.05767321214079857, + 0.17720924317836761, + 0.10896127671003342, + 0.05263519659638405, + 0.07667241990566254, + 0.02024056389927864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.0008142255246639252 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, low-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, sun-scorched desert plain with cracked earth and distant hazy mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, rectangular monolith with a craggy, weathered granite surface rise from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the rising monolith cause the ground to buckle, crack, and hurl chunks of rock and soil into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dense, roiling maelstrom of pulverized rock and ochre dust expand and engulf the base of the monolith?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036168177612125874 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.0036168177612125874, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036168177612125874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03660477697849274, + 0.013650412671267986, + 0.022969599813222885, + 0.03658280894160271, + 0.016687991097569466, + 0.012085363268852234, + 0.016777733340859413, + 0.039801985025405884, + 0.051976703107357025, + 0.7528626322746277 + ], + "credit_map": [ + 0.00331680360250175, + 0.004418864380568266, + 0.00850751530379057, + 0.0038252128288149834, + 0.003529260866343975, + 0.003514884738251567, + 0.004693722352385521, + 0.004798512440174818, + 0.0021733848843723536, + 0.0024495390243828297, + 0.0010094776516780257, + 0.0010916965547949076, + 0.0031689051538705826, + 0.0017961330013349652, + 0.0009357644594274461, + 0.0010255122324451804, + 0.0017647859640419483, + 0.005330117419362068, + 0.0015307676512748003, + 0.002680656500160694, + 0.00579602038487792, + 0.001590566011145711, + 0.002751520136371255, + 0.0015251677250489593, + 0.0033758957870304585, + 0.0036144766490906477, + 0.002776700770482421, + 0.002502254443243146, + 0.008630432188510895, + 0.009973322041332722, + 0.0028856690041720867, + 0.002824056427925825, + 0.0019268011674284935, + 0.002420807955786586, + 0.0017710590036585927, + 0.0021403476130217314, + 0.0017352188006043434, + 0.002512429840862751, + 0.00202920101583004, + 0.002152127679437399, + 0.0017713455017656088, + 0.001500152749940753, + 0.0015250496799126267, + 0.0014890138991177082, + 0.0014342948561534286, + 0.0011176500702276826, + 0.001631245599128306, + 0.0016166104469448328, + 0.001786816050298512, + 0.0014429453294724226, + 0.0017321414779871702, + 0.0022779509890824556, + 0.00221897941082716, + 0.0018444294109940529, + 0.002421312965452671, + 0.0030531578231602907, + 0.003947741352021694, + 0.0046387636102736, + 0.005513324867933989, + 0.004431686829775572, + 0.003870287211611867, + 0.004066995345056057, + 0.006767644546926022, + 0.006565544288605452, + 0.0051977364346385, + 0.009700778871774673, + 0.005547282751649618, + 0.0077184513211250305, + 0.004535967484116554, + 0.003846353618428111, + 0.00838921032845974, + 0.007040920201689005, + 0.023680005222558975, + 0.03669215366244316, + 0.08180700242519379, + 0.24611473083496094, + 0.15097317099571228, + 0.0739060565829277, + 0.10519452393054962, + 0.034495025873184204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.0036168177612125874 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_3/index_tensor([4199, 4199])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_3/index_tensor([4199, 4199])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6b65d3d8e1575b0a8e212f6e0da1c1c102b1990a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_3/index_tensor([4199, 4199])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angled medium shot focuses on a pristine white ceramic mold with a classic grooved ring design, resting on a warm, rustic wooden kitchen countertop. The mold's rim is filled with a vibrant, translucent red gel. Its surface, initially smooth and mirror-like, casts soft shadows and creates a subtle glow on the ceramic in the bright, natural daylight streaming in from the off-camera window. Suddenly, an invisible vibration travels through the countertop, causing the mold to begin a rapid, subtle sway. This movement immediately sets the gel into a mesmerizing vibration; it sways violently, its smooth surface catching the light as it rocks back and forth within the confines of the ceramic. Then, the vibrant movement gradually and gracefully subsides, the waves becoming smaller and slower, until the gel settles with a final, subtle ripple, returning to a perfect, glass-like stillness, creating an atmosphere of focused, almost scientific, quiet observation.", + "seed": 1326, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_3/index_4199_seed_1326_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1855344370007515, + "vlm_logprob": -0.1855344370007515, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.1855344370007515, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05291871353983879 + ], + "opposite_logprob_tokens": [ + -3.5529186725616455 + ], + "logprob_reward": -0.05291871353983879, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.05291871353983879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30297642946243286, + 0.0816558375954628, + 0.08751227706670761, + 0.048834092915058136, + 0.09530454874038696, + 0.07063111662864685, + 0.12670382857322693, + 0.0552901066839695, + 0.11292106658220291, + 0.018170684576034546 + ], + "credit_map": [ + 0.04196842387318611, + 0.017239060252904892, + 0.06354732811450958, + 0.03453083708882332, + 0.023275746032595634, + 0.02095112018287182, + 0.06342394649982452, + 0.038039952516555786, + 0.009630233980715275, + 0.0037814441602677107, + 0.013768048956990242, + 0.011602049693465233, + 0.005528497975319624, + 0.004698744509369135, + 0.02134666219353676, + 0.011300161480903625, + 0.005588130094110966, + 0.006963725201785564, + 0.009414293803274632, + 0.01168022584170103, + 0.009214231744408607, + 0.005347489379346371, + 0.025596601888537407, + 0.013707572594285011, + 0.006607856135815382, + 0.004685599822551012, + 0.008561110123991966, + 0.005467243492603302, + 0.004682109225541353, + 0.0030082236044108868, + 0.010032269172370434, + 0.005789679009467363, + 0.01233612559735775, + 0.00885524321347475, + 0.020415302366018295, + 0.011427699588239193, + 0.010378174483776093, + 0.00870983675122261, + 0.012173282913863659, + 0.011008883826434612, + 0.012525519356131554, + 0.007269560359418392, + 0.012085506692528725, + 0.007572035305202007, + 0.00957169197499752, + 0.005686626769602299, + 0.007738839834928513, + 0.00818133819848299, + 0.01860141009092331, + 0.008581675589084625, + 0.028583137318491936, + 0.013586198911070824, + 0.013591364957392216, + 0.008788367733359337, + 0.02056599035859108, + 0.014405686408281326, + 0.0097681675106287, + 0.003991938196122646, + 0.012173512950539589, + 0.005713137332350016, + 0.005778892897069454, + 0.00443080672994256, + 0.00717891613021493, + 0.006254734471440315, + 0.011714153923094273, + 0.0065161301754415035, + 0.01871930994093418, + 0.014444676227867603, + 0.010279219597578049, + 0.007154176477342844, + 0.029720192775130272, + 0.014373214915394783, + 0.0027301309164613485, + 0.0015887500485405326, + 0.003603041172027588, + 0.0022690091282129288, + 0.0014926143921911716, + 0.0016519244527444243, + 0.0026979222893714905, + 0.002137292642146349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.05291871353983879 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009818474762141705 + ], + "opposite_logprob_tokens": [ + -4.6348185539245605 + ], + "logprob_reward": -0.009818474762141705, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009818474762141705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18006177246570587, + 0.14388464391231537, + 0.23482614755630493, + 0.03921856731176376, + 0.0886346772313118, + 0.02854088321328163, + 0.10442503541707993, + 0.04875002056360245, + 0.09585518389940262, + 0.03580307215452194 + ], + "credit_map": [ + 0.02286047302186489, + 0.012567534111440182, + 0.044192470610141754, + 0.027832699939608574, + 0.018677720800042152, + 0.014723626896739006, + 0.021108675748109818, + 0.01809857226908207, + 0.016043594107031822, + 0.007526668254286051, + 0.021751247346401215, + 0.018937639892101288, + 0.00904562696814537, + 0.009458337910473347, + 0.040152356028556824, + 0.020969172939658165, + 0.028553463518619537, + 0.016182612627744675, + 0.05198848247528076, + 0.030964119359850883, + 0.015681585296988487, + 0.0176288690418005, + 0.04480708763003349, + 0.029019935056567192, + 0.004284116439521313, + 0.0027592703700065613, + 0.005244231782853603, + 0.007329376880079508, + 0.0033403479028493166, + 0.004193432629108429, + 0.007572861388325691, + 0.004494930151849985, + 0.01105737779289484, + 0.006775490939617157, + 0.02210545353591442, + 0.011099527589976788, + 0.008322447538375854, + 0.007342292927205563, + 0.011763307265937328, + 0.010168788954615593, + 0.0032444638200104237, + 0.002080962061882019, + 0.004702312406152487, + 0.0033435162622481585, + 0.0021503528114408255, + 0.002253282582387328, + 0.0068420832976698875, + 0.003923910204321146, + 0.012317168526351452, + 0.006331672426313162, + 0.01778206042945385, + 0.015931997448205948, + 0.00967385433614254, + 0.007869602181017399, + 0.02165091782808304, + 0.012867762707173824, + 0.0053421552293002605, + 0.004633918404579163, + 0.006676285061985254, + 0.004906667862087488, + 0.006298881955444813, + 0.0033267384860664606, + 0.010728774592280388, + 0.006836600601673126, + 0.00636263657361269, + 0.004832501523196697, + 0.008773714303970337, + 0.016541289165616035, + 0.008791584521532059, + 0.005167214199900627, + 0.03091367706656456, + 0.014472570270299911, + 0.004946880508214235, + 0.0032950458116829395, + 0.007987053133547306, + 0.0035821539349853992, + 0.0035844489466398954, + 0.00340624013915658, + 0.005140305031090975, + 0.00386094325222075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.009818474762141705 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108983252197504 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.01108983252197504, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108983252197504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22172005474567413, + 0.05059013515710831, + 0.3559896647930145, + 0.05312905088067055, + 0.042761508375406265, + 0.018001383170485497, + 0.045032646507024765, + 0.08854812383651733, + 0.0902508944272995, + 0.03397655859589577 + ], + "credit_map": [ + 0.03734491392970085, + 0.015726745128631592, + 0.04893898218870163, + 0.02850484661757946, + 0.02446037530899048, + 0.02009393461048603, + 0.022037910297513008, + 0.02461233362555504, + 0.005453881807625294, + 0.008279575034976006, + 0.006173619069159031, + 0.004531670827418566, + 0.008946304209530354, + 0.0036075536627322435, + 0.0073310611769557, + 0.006266471929848194, + 0.04089468717575073, + 0.018027642741799355, + 0.053634047508239746, + 0.04375477880239487, + 0.030629919841885567, + 0.021335754543542862, + 0.09653028845787048, + 0.05118256434798241, + 0.00768462335690856, + 0.003622569376602769, + 0.01067022979259491, + 0.005529300309717655, + 0.0054211728274822235, + 0.004588067997246981, + 0.00907085370272398, + 0.00654223095625639, + 0.005606277845799923, + 0.00474008871242404, + 0.007824941538274288, + 0.004609961528331041, + 0.006272788159549236, + 0.003618343733251095, + 0.0053058504126966, + 0.004783256445080042, + 0.0014443506952375174, + 0.0012220264179632068, + 0.0017970763146877289, + 0.0030189401004463434, + 0.0020020075608044863, + 0.001407682546414435, + 0.0047265952453017235, + 0.00238270522095263, + 0.008135008625686169, + 0.004066179972141981, + 0.005702233873307705, + 0.0037134247832000256, + 0.005056409630924463, + 0.0032517656218260527, + 0.008709358982741833, + 0.00639826525002718, + 0.014544703997671604, + 0.008749538101255894, + 0.014849765226244926, + 0.008750992827117443, + 0.010623440146446228, + 0.006485785823315382, + 0.01401086151599884, + 0.010533041320741177, + 0.008574506267905235, + 0.005147418938577175, + 0.01351387333124876, + 0.011653445661067963, + 0.008925083093345165, + 0.005254748277366161, + 0.025095537304878235, + 0.01208628248423338, + 0.00429162522777915, + 0.0014186586486175656, + 0.005437374580651522, + 0.0038520721718668938, + 0.002283488167449832, + 0.0021938395220786333, + 0.00947217084467411, + 0.005027328617870808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.01108983252197504 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.693359375 + ], + "opposite_logprob_tokens": [ + -0.693359375 + ], + "logprob_reward": -0.693359375, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.693359375, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17116448283195496, + 0.056820645928382874, + 0.10762042552232742, + 0.04526377469301224, + 0.04970528185367584, + 0.041291192173957825, + 0.15069034695625305, + 0.06715551763772964, + 0.2887164354324341, + 0.021571865305304527 + ], + "credit_map": [ + 0.019277088344097137, + 0.01651887595653534, + 0.03150147572159767, + 0.020523149520158768, + 0.021794235333800316, + 0.013802660629153252, + 0.029082508757710457, + 0.018664494156837463, + 0.006101618520915508, + 0.004999383352696896, + 0.009362710639834404, + 0.007485927082598209, + 0.006335003301501274, + 0.003634836059063673, + 0.012249906547367573, + 0.00665126321837306, + 0.009442964568734169, + 0.00466612633317709, + 0.013106896542012691, + 0.01538635604083538, + 0.009188896045088768, + 0.006243237294256687, + 0.03338530287146568, + 0.016200648620724678, + 0.007032650522887707, + 0.003282255260273814, + 0.008371395990252495, + 0.0046747708693146706, + 0.004466070793569088, + 0.004094352480024099, + 0.00795670785009861, + 0.005385575350373983, + 0.005348799750208855, + 0.0031753943767398596, + 0.008473028428852558, + 0.006625728216022253, + 0.004252375569194555, + 0.003700982080772519, + 0.011608155444264412, + 0.006520821712911129, + 0.0034848288632929325, + 0.0034181729424744844, + 0.004162612836807966, + 0.005710792262107134, + 0.004783932119607925, + 0.0022629329469054937, + 0.01120962668210268, + 0.006258295848965645, + 0.01327028963714838, + 0.006503457203507423, + 0.017611561343073845, + 0.021354936063289642, + 0.013660726137459278, + 0.008074918761849403, + 0.04758608713746071, + 0.022628391161561012, + 0.007148294243961573, + 0.005458938889205456, + 0.010345669463276863, + 0.007558048702776432, + 0.008535275235772133, + 0.005290381610393524, + 0.013893981464207172, + 0.008924932219088078, + 0.020037854090332985, + 0.014798731543123722, + 0.029308291152119637, + 0.04536288231611252, + 0.028493881225585938, + 0.015561291016638279, + 0.09215255081653595, + 0.0430009663105011, + 0.0024058674462139606, + 0.0014929650351405144, + 0.0028003579936921597, + 0.0028267595916986465, + 0.0020291393157094717, + 0.0015296959318220615, + 0.005468572955578566, + 0.003018507733941078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.693359375 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16048578917980194 + ], + "opposite_logprob_tokens": [ + -1.9104857444763184 + ], + "logprob_reward": -0.16048578917980194, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16048578917980194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1941954493522644, + 0.07442396879196167, + 0.04511262848973274, + 0.12018796801567078, + 0.32318079471588135, + 0.030333654955029488, + 0.12055040895938873, + 0.06528681516647339, + 0.01081346906721592, + 0.015914849936962128 + ], + "credit_map": [ + 0.031613416969776154, + 0.013220365159213543, + 0.03889830410480499, + 0.020096678286790848, + 0.016168564558029175, + 0.013374215923249722, + 0.03755012899637222, + 0.023273779079318047, + 0.01111980713903904, + 0.007387758232653141, + 0.016391567885875702, + 0.00794208887964487, + 0.009182612411677837, + 0.0054887160658836365, + 0.00831763818860054, + 0.008593777194619179, + 0.005542031489312649, + 0.0038721447344869375, + 0.007047867402434349, + 0.005507091525942087, + 0.005159290973097086, + 0.0030160231981426477, + 0.009097307920455933, + 0.005870872642844915, + 0.01559224259108305, + 0.009193609468638897, + 0.0309181809425354, + 0.015205219388008118, + 0.010629432275891304, + 0.010078717954456806, + 0.014985691756010056, + 0.013584872707724571, + 0.03944528102874756, + 0.023186182603240013, + 0.07838697731494904, + 0.03722328320145607, + 0.02733117900788784, + 0.024989129975438118, + 0.055674053728580475, + 0.03694472834467888, + 0.0028985049575567245, + 0.003225035732612014, + 0.0031462276820093393, + 0.0032472554594278336, + 0.004682803992182016, + 0.0018836859380826354, + 0.0070290397852659225, + 0.004221100825816393, + 0.010673115029931068, + 0.006427227519452572, + 0.01106082834303379, + 0.017233198508620262, + 0.009133340790867805, + 0.006849874276667833, + 0.038690563291311264, + 0.020482253283262253, + 0.007075991947203875, + 0.005811079405248165, + 0.010719318874180317, + 0.0070123085752129555, + 0.006863192189484835, + 0.004225736018270254, + 0.014643914997577667, + 0.008935278281569481, + 0.0013824277557432652, + 0.0009269680012948811, + 0.0013209932949393988, + 0.0011768452823162079, + 0.0014571795472875237, + 0.0006921848398633301, + 0.0023382592480629683, + 0.001518610748462379, + 0.0016217816155403852, + 0.0012979765888303518, + 0.00257017370313406, + 0.0019432185217738152, + 0.0016797386342659593, + 0.00102215725928545, + 0.003819042118266225, + 0.001960762543603778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.16048578917980194 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot focuses on a pristine white ceramic mold with a classic grooved ring design, resting on a warm, rustic wooden kitchen countertop. The mold's rim is filled with a vibrant, translucent red gel. Its surface, initially smooth and mirror-like, casts soft shadows and creates a subtle glow on the ceramic in the bright, natural daylight streaming in from the off-camera window. Suddenly, an invisible vibration travels through the countertop, causing the mold to begin a rapid, subtle sway. This movement immediately sets the gel into a mesmerizing vibration; it sways violently, its smooth surface catching the light as it rocks back and forth within the confines of the ceramic. Then, the vibrant movement gradually and gracefully subsides, the waves becoming smaller and slower, until the gel settles with a final, subtle ripple, returning to a perfect, glass-like stillness, creating an atmosphere of focused, almost scientific, quiet observation.", + "seed": 1327, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_3/index_4199_seed_1327_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15828267131000756, + "vlm_logprob": -0.15828267131000756, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.15828267131000756, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0478256493806839 + ], + "opposite_logprob_tokens": [ + -4.297825813293457 + ], + "logprob_reward": -0.0478256493806839, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0478256493806839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09076759964227676, + 0.08015576004981995, + 0.02468234673142433, + 0.049156419932842255, + 0.11823509633541107, + 0.23298832774162292, + 0.06560638546943665, + 0.12235903739929199, + 0.15968634188175201, + 0.05636272206902504 + ], + "credit_map": [ + 0.010525792837142944, + 0.007042326498776674, + 0.009811010211706161, + 0.01108554471284151, + 0.01426276471465826, + 0.003968018107116222, + 0.02006950043141842, + 0.014002639800310135, + 0.017189115285873413, + 0.00448249327018857, + 0.013901636004447937, + 0.004183247685432434, + 0.007927333936095238, + 0.0050037638284265995, + 0.0061791022308170795, + 0.02128906175494194, + 0.004505496006458998, + 0.0019466773374006152, + 0.004516063258051872, + 0.001679105102084577, + 0.003154815174639225, + 0.0018128262599930167, + 0.002345234388485551, + 0.004722129087895155, + 0.009761609137058258, + 0.003959674388170242, + 0.010600443929433823, + 0.0032821858767420053, + 0.004480281379073858, + 0.004172511398792267, + 0.0050491224974393845, + 0.007850589230656624, + 0.01459769532084465, + 0.007798669394105673, + 0.01714833453297615, + 0.01448796410113573, + 0.013780689798295498, + 0.005531014874577522, + 0.025964414700865746, + 0.01892632059752941, + 0.028959160670638084, + 0.01660139299929142, + 0.02436644583940506, + 0.027914835140109062, + 0.03704901412129402, + 0.012423280626535416, + 0.047193411737680435, + 0.03848078474402428, + 0.009449038654565811, + 0.0051298015750944614, + 0.009525244124233723, + 0.004261626396328211, + 0.0077416375279426575, + 0.0061737955547869205, + 0.007330893538892269, + 0.015994353219866753, + 0.02173926867544651, + 0.008108374662697315, + 0.01841382496058941, + 0.008219719864428043, + 0.019842272624373436, + 0.007213117089122534, + 0.013959269039332867, + 0.02486318349838257, + 0.021607838571071625, + 0.012998881749808788, + 0.019238080829381943, + 0.013901556842029095, + 0.03811410069465637, + 0.009756321087479591, + 0.01591421850025654, + 0.02815534546971321, + 0.009062682278454304, + 0.004380150232464075, + 0.008651583455502987, + 0.0038958208169788122, + 0.009090847335755825, + 0.006109021604061127, + 0.005799363367259502, + 0.0093732550740242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.0478256493806839 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008685313165187836 + ], + "opposite_logprob_tokens": [ + -4.758685111999512 + ], + "logprob_reward": -0.008685313165187836, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008685313165187836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17341428995132446, + 0.21508347988128662, + 0.05549733713269234, + 0.048202987760305405, + 0.038180187344551086, + 0.1334158182144165, + 0.04753793403506279, + 0.08373460173606873, + 0.09007326513528824, + 0.11486010253429413 + ], + "credit_map": [ + 0.037467800080776215, + 0.01106918416917324, + 0.033506520092487335, + 0.007129190489649773, + 0.01505516842007637, + 0.014156402088701725, + 0.011478252708911896, + 0.043551769107580185, + 0.06172799691557884, + 0.00918483640998602, + 0.025742927566170692, + 0.006589682772755623, + 0.012295342050492764, + 0.016261175274848938, + 0.009079990908503532, + 0.07420153170824051, + 0.008892598561942577, + 0.004174598027020693, + 0.006161162164062262, + 0.0039032024797052145, + 0.005109126213937998, + 0.004896433092653751, + 0.006921615917235613, + 0.01543859951198101, + 0.007972313091158867, + 0.003262634389102459, + 0.0038496118504554033, + 0.003415079787373543, + 0.008321615867316723, + 0.0036586467176675797, + 0.006290754768997431, + 0.01143233384937048, + 0.0045515079982578754, + 0.0035088344011455774, + 0.005080295726656914, + 0.003438537009060383, + 0.0069303978234529495, + 0.004123200662434101, + 0.004487119615077972, + 0.00606029387563467, + 0.021261870861053467, + 0.008244113996624947, + 0.03172556310892105, + 0.00884915515780449, + 0.015989823266863823, + 0.008797199465334415, + 0.014208218082785606, + 0.02433987520635128, + 0.0056013986468315125, + 0.0040842401795089245, + 0.007367286831140518, + 0.004045982379466295, + 0.007963518612086773, + 0.0038689342327415943, + 0.006579113658517599, + 0.008027457632124424, + 0.011745443567633629, + 0.006345412693917751, + 0.013796861283481121, + 0.006939601618796587, + 0.016467737033963203, + 0.008030964061617851, + 0.008032609708607197, + 0.012375970371067524, + 0.011081992648541927, + 0.007937892340123653, + 0.009200223721563816, + 0.008453629910945892, + 0.023135924711823463, + 0.0055675869807600975, + 0.009711791761219501, + 0.01498422957956791, + 0.017390862107276917, + 0.010999291203916073, + 0.016807368025183678, + 0.009325523860752583, + 0.019816458225250244, + 0.013700737617909908, + 0.011068368330597878, + 0.015751488506793976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.008685313165187836 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009797108359634876 + ], + "opposite_logprob_tokens": [ + -4.634797096252441 + ], + "logprob_reward": -0.009797108359634876, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009797108359634876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08088133484125137, + 0.07929082214832306, + 0.026245756074786186, + 0.028099825605750084, + 0.28426069021224976, + 0.23369644582271576, + 0.015982966870069504, + 0.14779970049858093, + 0.042268116027116776, + 0.0614742636680603 + ], + "credit_map": [ + 0.014380414970219135, + 0.005086660850793123, + 0.009404466487467289, + 0.007418861147016287, + 0.012594388797879219, + 0.006184288766235113, + 0.011568574234843254, + 0.014243688434362411, + 0.017581017687916756, + 0.004036299418658018, + 0.011356822215020657, + 0.004379512742161751, + 0.009257223457098007, + 0.004726932384073734, + 0.0066582635045051575, + 0.02129475772380829, + 0.0037015785928815603, + 0.0018607763340696692, + 0.005079203750938177, + 0.0018458714475855231, + 0.0038142476696521044, + 0.0015167436795309186, + 0.004238098859786987, + 0.004189237020909786, + 0.005732686724513769, + 0.0019486148376017809, + 0.0029572900384664536, + 0.001530878827907145, + 0.00415235897526145, + 0.0022765023168176413, + 0.0030717060435563326, + 0.006429790053516626, + 0.04395214840769768, + 0.016867024824023247, + 0.04684513434767723, + 0.025053951889276505, + 0.0293422881513834, + 0.018427889794111252, + 0.047801800072193146, + 0.055970463901758194, + 0.036607854068279266, + 0.016227243468165398, + 0.025463247671723366, + 0.015853943303227425, + 0.0549604706466198, + 0.01601923443377018, + 0.028117382898926735, + 0.040447112172842026, + 0.0030331385787576437, + 0.0011028997832909226, + 0.002603645669296384, + 0.0008213170804083347, + 0.001765612163580954, + 0.0014825609978288412, + 0.0014007657300680876, + 0.0037730270996689796, + 0.02405143529176712, + 0.0103959571570158, + 0.014318662695586681, + 0.013598437421023846, + 0.026330625638365746, + 0.008806037716567516, + 0.021432621404528618, + 0.028865940868854523, + 0.00819752924144268, + 0.0031492391135543585, + 0.005716556683182716, + 0.002859990345314145, + 0.006856484804302454, + 0.0031839744187891483, + 0.003424155293032527, + 0.008880190551280975, + 0.009872461669147015, + 0.004661243408918381, + 0.010340151377022266, + 0.004659971222281456, + 0.009872065857052803, + 0.00499076209962368, + 0.006644452456384897, + 0.010433162562549114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.009797108359634876 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233277082443237 + ], + "opposite_logprob_tokens": [ + -0.8983277082443237 + ], + "logprob_reward": -0.5233277082443237, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233277082443237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04713369905948639, + 0.024548247456550598, + 0.028162984177470207, + 0.018630104139447212, + 0.21956023573875427, + 0.20973117649555206, + 0.027728557586669922, + 0.23274315893650055, + 0.02771371230483055, + 0.16404815018177032 + ], + "credit_map": [ + 0.007235940080136061, + 0.004475450608879328, + 0.007893436588346958, + 0.0037696019280701876, + 0.008475090377032757, + 0.003953759092837572, + 0.004107463173568249, + 0.007222960237413645, + 0.003583799349144101, + 0.001391954836435616, + 0.005771296098828316, + 0.0024822624400258064, + 0.00327476323582232, + 0.0017987563041970134, + 0.003028871025890112, + 0.003216543234884739, + 0.004516499117016792, + 0.0017739252652972937, + 0.00545112881809473, + 0.001982595771551132, + 0.004632176365703344, + 0.001387114287354052, + 0.0038895183242857456, + 0.004530024249106646, + 0.0024051263462752104, + 0.0014967835741117597, + 0.0022711975034326315, + 0.0020741652697324753, + 0.002374030649662018, + 0.0017823823727667332, + 0.003028721548616886, + 0.0031976974569261074, + 0.023733196780085564, + 0.01798619143664837, + 0.020359646528959274, + 0.02850857749581337, + 0.035141173750162125, + 0.01278175599873066, + 0.05096098408102989, + 0.03008871152997017, + 0.023373190313577652, + 0.015463070012629032, + 0.017837364226579666, + 0.02789541520178318, + 0.03154643252491951, + 0.009115711785852909, + 0.05135555937886238, + 0.03314441815018654, + 0.005276803858578205, + 0.0019667872693389654, + 0.003372028237208724, + 0.0015104877529665828, + 0.003998334053903818, + 0.0022327874321490526, + 0.003021705662831664, + 0.006349625065922737, + 0.04542415216565132, + 0.014280667528510094, + 0.028411947190761566, + 0.016782622784376144, + 0.033378999680280685, + 0.0130965206772089, + 0.027424732223153114, + 0.053943511098623276, + 0.006014001555740833, + 0.001750611118040979, + 0.0025933205615729094, + 0.00141426851041615, + 0.004760683048516512, + 0.001704795635305345, + 0.0018463146407157183, + 0.007629717234522104, + 0.021386906504631042, + 0.01343406829982996, + 0.022887544706463814, + 0.012428459711372852, + 0.03338077664375305, + 0.017008591443300247, + 0.018092621117830276, + 0.025429196655750275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.5233277082443237 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20177757740020752 + ], + "opposite_logprob_tokens": [ + -1.7017775774002075 + ], + "logprob_reward": -0.20177757740020752, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20177757740020752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26164641976356506, + 0.07475941628217697, + 0.04311072453856468, + 0.04500895366072655, + 0.1039929911494255, + 0.05604834109544754, + 0.06988328695297241, + 0.23375527560710907, + 0.06977041065692902, + 0.04202408343553543 + ], + "credit_map": [ + 0.04995284602046013, + 0.013202695176005363, + 0.028138546273112297, + 0.02529507502913475, + 0.03610428795218468, + 0.012988235801458359, + 0.049105025827884674, + 0.046859726309776306, + 0.01069260947406292, + 0.005499171558767557, + 0.012021336704492569, + 0.007249764632433653, + 0.006259610876441002, + 0.0051767597906291485, + 0.01137758418917656, + 0.016482578590512276, + 0.008189226500689983, + 0.0032218133565038443, + 0.009943202137947083, + 0.0025635145138949156, + 0.00360294827260077, + 0.005149521864950657, + 0.0036686218809336424, + 0.006771877408027649, + 0.009544702246785164, + 0.00382487615570426, + 0.009885046631097794, + 0.0030145824421197176, + 0.0038195333909243345, + 0.004145052284002304, + 0.0042841872200369835, + 0.006490971427410841, + 0.01577789895236492, + 0.0075838626362383366, + 0.0102745546028018, + 0.010091105476021767, + 0.012377592734992504, + 0.006176005117595196, + 0.018681656569242477, + 0.023030312731862068, + 0.010416684672236443, + 0.003923117648810148, + 0.005566087551414967, + 0.0031048832461237907, + 0.010761646553874016, + 0.003419779473915696, + 0.006867574993520975, + 0.011988567188382149, + 0.01039630826562643, + 0.0047592176124453545, + 0.01618458889424801, + 0.005795755423605442, + 0.009814773686230183, + 0.007519253063946962, + 0.006338454782962799, + 0.00907493382692337, + 0.0428713783621788, + 0.014638200402259827, + 0.03139428049325943, + 0.017531558871269226, + 0.03685471788048744, + 0.013714573346078396, + 0.02727581560611725, + 0.04947476089000702, + 0.008114826865494251, + 0.006062414962798357, + 0.008168498054146767, + 0.006566022057086229, + 0.017581844702363014, + 0.004407850559800863, + 0.00787338986992836, + 0.010995565913617611, + 0.008010970428586006, + 0.002837579930201173, + 0.004329223185777664, + 0.0025329759810119867, + 0.006183648481965065, + 0.004768238868564367, + 0.0037314780056476593, + 0.009629969485104084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.20177757740020752 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot focuses on a pristine white ceramic mold with a classic grooved ring design, resting on a warm, rustic wooden kitchen countertop. The mold's rim is filled with a vibrant, translucent red gel. Its surface, initially smooth and mirror-like, casts soft shadows and creates a subtle glow on the ceramic in the bright, natural daylight streaming in from the off-camera window. Suddenly, an invisible vibration travels through the countertop, causing the mold to begin a rapid, subtle sway. This movement immediately sets the gel into a mesmerizing vibration; it sways violently, its smooth surface catching the light as it rocks back and forth within the confines of the ceramic. Then, the vibrant movement gradually and gracefully subsides, the waves becoming smaller and slower, until the gel settles with a final, subtle ripple, returning to a perfect, glass-like stillness, creating an atmosphere of focused, almost scientific, quiet observation.", + "seed": 1328, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_3/index_4199_seed_1328_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12330666882917285, + "vlm_logprob": -0.12330666882917285, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.12330666882917285, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03664921596646309 + ], + "opposite_logprob_tokens": [ + -4.161649227142334 + ], + "logprob_reward": -0.03664921596646309, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.03664921596646309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2824801504611969, + 0.11366124451160431, + 0.02705850638449192, + 0.04138552397489548, + 0.07171772420406342, + 0.3085453510284424, + 0.05870150402188301, + 0.03215956315398216, + 0.04098998010158539, + 0.023300454020500183 + ], + "credit_map": [ + 0.0358729250729084, + 0.03178668022155762, + 0.05827127397060394, + 0.03436509147286415, + 0.019993692636489868, + 0.029095908626914024, + 0.03588738292455673, + 0.037207167595624924, + 0.017786450684070587, + 0.012465315870940685, + 0.01862460933625698, + 0.012599176727235317, + 0.008937854319810867, + 0.015932999551296234, + 0.013919981196522713, + 0.013394856825470924, + 0.0035273772664368153, + 0.0033273445442318916, + 0.005320930387824774, + 0.0037121588829904795, + 0.0022341462317854166, + 0.0035779529716819525, + 0.002577603328973055, + 0.002780991606414318, + 0.006057003512978554, + 0.004547988064587116, + 0.008472008630633354, + 0.004134621005505323, + 0.004207055550068617, + 0.0053019411861896515, + 0.005095846951007843, + 0.0035690616350620985, + 0.011152132414281368, + 0.007940636947751045, + 0.012325217016041279, + 0.008307578973472118, + 0.005435900762677193, + 0.009824441745877266, + 0.00839269906282425, + 0.008339113555848598, + 0.050616104155778885, + 0.03482897952198982, + 0.0584774874150753, + 0.025626011192798615, + 0.022197311744093895, + 0.043780528008937836, + 0.03157481551170349, + 0.041444119065999985, + 0.009219774976372719, + 0.008650426752865314, + 0.012257600203156471, + 0.006002736743539572, + 0.004913431126624346, + 0.004598996136337519, + 0.006112460047006607, + 0.006946077570319176, + 0.004781581461429596, + 0.0038799275644123554, + 0.0056113009341061115, + 0.0035869241692125797, + 0.003353483509272337, + 0.0031393454410135746, + 0.004252909682691097, + 0.0035540906246751547, + 0.005155123770236969, + 0.004639791324734688, + 0.006234835367649794, + 0.006441427394747734, + 0.004603379871696234, + 0.0036578266881406307, + 0.005388940684497356, + 0.004868656862527132, + 0.002668580040335655, + 0.0031207320280373096, + 0.0036216082517057657, + 0.0042113629169762135, + 0.0023758667521178722, + 0.0015001812716946006, + 0.002906613517552614, + 0.002895510522648692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.03664921596646309 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026404596865177155 + ], + "opposite_logprob_tokens": [ + -3.651404619216919 + ], + "logprob_reward": -0.026404596865177155, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026404596865177155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24563060700893402, + 0.10591452568769455, + 0.14098215103149414, + 0.0426618792116642, + 0.054164886474609375, + 0.10876801609992981, + 0.07112771272659302, + 0.09147785604000092, + 0.0716044157743454, + 0.06766799837350845 + ], + "credit_map": [ + 0.034216027706861496, + 0.020291486755013466, + 0.0451369509100914, + 0.028714807704091072, + 0.02479773573577404, + 0.029629239812493324, + 0.029612751677632332, + 0.03323160484433174, + 0.01496898289769888, + 0.010176446288824081, + 0.01644079200923443, + 0.015124513767659664, + 0.009968495927751064, + 0.009118124842643738, + 0.01627487875521183, + 0.013842292129993439, + 0.02119162119925022, + 0.014770366251468658, + 0.017503688111901283, + 0.023068787530064583, + 0.012723162770271301, + 0.0187015850096941, + 0.019281670451164246, + 0.013741266913712025, + 0.005137208383530378, + 0.005650029052048922, + 0.008573927916586399, + 0.005492459051311016, + 0.004156552255153656, + 0.004362322855740786, + 0.004962626378983259, + 0.004326752852648497, + 0.007080680225044489, + 0.00603260425850749, + 0.009920458309352398, + 0.007696964778006077, + 0.0037083541974425316, + 0.005250577814877033, + 0.007759572006762028, + 0.006715674884617329, + 0.01738727279007435, + 0.012529423460364342, + 0.018642155453562737, + 0.00940739456564188, + 0.008188669569790363, + 0.015434307046234608, + 0.012426174245774746, + 0.014752616174519062, + 0.006438907701522112, + 0.01093950867652893, + 0.011011146008968353, + 0.011524870991706848, + 0.007528476882725954, + 0.006110722199082375, + 0.009481496177613735, + 0.008092587813735008, + 0.015866084024310112, + 0.009363681077957153, + 0.01636534370481968, + 0.009479437954723835, + 0.00770104955881834, + 0.01609526202082634, + 0.007664911914616823, + 0.008942083455622196, + 0.007757497951388359, + 0.008645057678222656, + 0.012002188712358475, + 0.009998100809752941, + 0.007183379027992487, + 0.007895821705460548, + 0.009028024040162563, + 0.00909434910863638, + 0.011987448669970036, + 0.007241441402584314, + 0.007793821394443512, + 0.00842239335179329, + 0.004956115037202835, + 0.011522858403623104, + 0.008006765507161617, + 0.007737158797681332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.026404596865177155 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005962797906249762 + ], + "opposite_logprob_tokens": [ + -5.13096284866333 + ], + "logprob_reward": -0.005962797906249762, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005962797906249762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.162388876080513, + 0.07697145640850067, + 0.036861639469861984, + 0.12014349550008774, + 0.042914822697639465, + 0.35742858052253723, + 0.07748163491487503, + 0.06261978298425674, + 0.03513023629784584, + 0.028059499338269234 + ], + "credit_map": [ + 0.02125190943479538, + 0.01802949793636799, + 0.026815107092261314, + 0.022226089611649513, + 0.01596706546843052, + 0.019399531185626984, + 0.019933123141527176, + 0.018766529858112335, + 0.017113562673330307, + 0.007267650216817856, + 0.00638905493542552, + 0.005892714951187372, + 0.008345003239810467, + 0.014278947375714779, + 0.009508294053375721, + 0.008176220580935478, + 0.0058167497627437115, + 0.0028491460252553225, + 0.005477425176650286, + 0.004240775015205145, + 0.0042726704850792885, + 0.004693595692515373, + 0.005094151943922043, + 0.004417121410369873, + 0.015518972650170326, + 0.014767581596970558, + 0.027501674368977547, + 0.014855061657726765, + 0.00790349766612053, + 0.010074355639517307, + 0.014330757781863213, + 0.015191580168902874, + 0.007056675851345062, + 0.004026577342301607, + 0.008092152886092663, + 0.005235468968749046, + 0.004938085097819567, + 0.0033529645297676325, + 0.005722559057176113, + 0.00449033547192812, + 0.05524618923664093, + 0.04119474068284035, + 0.07348786294460297, + 0.031345680356025696, + 0.024977080523967743, + 0.04634326323866844, + 0.03775005787611008, + 0.04708367586135864, + 0.010729231871664524, + 0.010246005840599537, + 0.011343173682689667, + 0.010038480162620544, + 0.007506529334932566, + 0.009110935032367706, + 0.009344086982309818, + 0.009163182228803635, + 0.007660806644707918, + 0.008338800631463528, + 0.010472374968230724, + 0.009160377085208893, + 0.006048816721886396, + 0.006949198432266712, + 0.007409220561385155, + 0.006580180488526821, + 0.0042131333611905575, + 0.0043005929328501225, + 0.005951884668320417, + 0.006159912794828415, + 0.0029984309803694487, + 0.004053393378853798, + 0.0035644120071083307, + 0.0038884703535586596, + 0.0053593856282532215, + 0.002945441287010908, + 0.0044729155488312244, + 0.0022974847815930843, + 0.0023957528173923492, + 0.00518585043027997, + 0.0026607366744428873, + 0.0027419310063123703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.005962797906249762 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870507478713989 + ], + "opposite_logprob_tokens": [ + -1.137050747871399 + ], + "logprob_reward": -0.3870507478713989, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870507478713989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21735486388206482, + 0.08203107863664627, + 0.11452557891607285, + 0.04301692545413971, + 0.08992975950241089, + 0.263763427734375, + 0.03698541224002838, + 0.053847577422857285, + 0.058192890137434006, + 0.0403524711728096 + ], + "credit_map": [ + 0.030676288530230522, + 0.029226550832390785, + 0.02675621397793293, + 0.03680155798792839, + 0.016103176400065422, + 0.022226158529520035, + 0.03370263800024986, + 0.021862298250198364, + 0.014588057063519955, + 0.0077555519528687, + 0.009260591119527817, + 0.009408071637153625, + 0.007981942966580391, + 0.01237355824559927, + 0.011351054534316063, + 0.009312254376709461, + 0.017342112958431244, + 0.013374139554798603, + 0.016101384535431862, + 0.01452645193785429, + 0.009651930071413517, + 0.01862911693751812, + 0.012937734834849834, + 0.011962715536355972, + 0.005100945010781288, + 0.005296880844980478, + 0.010066921822726727, + 0.004886273294687271, + 0.0035779380705207586, + 0.0041001541540026665, + 0.004863762762397528, + 0.005124053917825222, + 0.01265252660959959, + 0.012345116585493088, + 0.014128201641142368, + 0.009836978279054165, + 0.011419900692999363, + 0.008246231824159622, + 0.012097124010324478, + 0.009203682653605938, + 0.044685591012239456, + 0.02832462638616562, + 0.049095675349235535, + 0.020772669464349747, + 0.01908615417778492, + 0.0409025102853775, + 0.02494131401181221, + 0.03595492243766785, + 0.004762039054185152, + 0.00475721200928092, + 0.00444468529894948, + 0.005372137762606144, + 0.004531283862888813, + 0.0036331957671791315, + 0.004424763843417168, + 0.005060098133981228, + 0.009286218322813511, + 0.0075005474500358105, + 0.008071057498455048, + 0.0057160924188792706, + 0.004473286680877209, + 0.006603133864700794, + 0.006234974134713411, + 0.005962270777672529, + 0.007083043456077576, + 0.00658373674377799, + 0.009563163854181767, + 0.010715868324041367, + 0.005621511954814196, + 0.006023851223289967, + 0.007998769171535969, + 0.004602954722940922, + 0.005638478323817253, + 0.004885710775852203, + 0.005659361835569143, + 0.006529810838401318, + 0.004319366998970509, + 0.003218545578420162, + 0.005628528539091349, + 0.004472674336284399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.3870507478713989 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16046598553657532 + ], + "opposite_logprob_tokens": [ + -1.910465955734253 + ], + "logprob_reward": -0.16046598553657532, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16046598553657532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16429021954536438, + 0.03597632423043251, + 0.0096987159922719, + 0.11199336498975754, + 0.1008794903755188, + 0.4524834156036377, + 0.06759748607873917, + 0.046948738396167755, + 0.007522806525230408, + 0.002609501127153635 + ], + "credit_map": [ + 0.024325337260961533, + 0.018087511882185936, + 0.03258097916841507, + 0.017949674278497696, + 0.012350822798907757, + 0.020223280414938927, + 0.01991184800863266, + 0.018860768526792526, + 0.006125432904809713, + 0.003649708116427064, + 0.006197340786457062, + 0.003535100258886814, + 0.003038693219423294, + 0.005492871161550283, + 0.0036492743529379368, + 0.004287900403141975, + 0.0011789482086896896, + 0.0011862422106787562, + 0.002106612315401435, + 0.0011070651235058904, + 0.0007682378054596484, + 0.0007167223957367241, + 0.0013694921508431435, + 0.0012653946178033948, + 0.014552331529557705, + 0.013697698712348938, + 0.0268327035009861, + 0.01362709328532219, + 0.007490785792469978, + 0.008033421821892262, + 0.013524848967790604, + 0.01423446275293827, + 0.013178338296711445, + 0.012789514847099781, + 0.021194228902459145, + 0.013597458600997925, + 0.0058540464378893375, + 0.008602690882980824, + 0.013975637964904308, + 0.01168756652623415, + 0.07022664695978165, + 0.05091867595911026, + 0.09073284268379211, + 0.039917636662721634, + 0.031890638172626495, + 0.060441721230745316, + 0.04770999774336815, + 0.0606452114880085, + 0.0081058070063591, + 0.007096488028764725, + 0.01042084489017725, + 0.01169490721076727, + 0.00631870049983263, + 0.008454582653939724, + 0.009058402851223946, + 0.0064477454870939255, + 0.0057805259712040424, + 0.005926371086388826, + 0.007753469981253147, + 0.006919101811945438, + 0.004915519617497921, + 0.0054607391357421875, + 0.005371361970901489, + 0.004821649752557278, + 0.0009911208180710673, + 0.0007982440292835236, + 0.0015415303641930223, + 0.0011222350876778364, + 0.0007044983794912696, + 0.0008397545316256583, + 0.0007861001649871469, + 0.0007393229170702398, + 0.000384457700420171, + 0.0003211243310943246, + 0.0005610079970210791, + 0.00031979740015231073, + 0.00020238725119270384, + 0.00023415837495122105, + 0.0003330608014948666, + 0.00025350722717121243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.16046598553657532 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot focuses on a pristine white ceramic mold with a classic grooved ring design, resting on a warm, rustic wooden kitchen countertop. The mold's rim is filled with a vibrant, translucent red gel. Its surface, initially smooth and mirror-like, casts soft shadows and creates a subtle glow on the ceramic in the bright, natural daylight streaming in from the off-camera window. Suddenly, an invisible vibration travels through the countertop, causing the mold to begin a rapid, subtle sway. This movement immediately sets the gel into a mesmerizing vibration; it sways violently, its smooth surface catching the light as it rocks back and forth within the confines of the ceramic. Then, the vibrant movement gradually and gracefully subsides, the waves becoming smaller and slower, until the gel settles with a final, subtle ripple, returning to a perfect, glass-like stillness, creating an atmosphere of focused, almost scientific, quiet observation.", + "seed": 1329, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_3/index_4199_seed_1329_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15178346987813712, + "vlm_logprob": -0.15178346987813712, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.15178346987813712, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04680928215384483 + ], + "opposite_logprob_tokens": [ + -3.671809196472168 + ], + "logprob_reward": -0.04680928215384483, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.04680928215384483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05648435652256012, + 0.012799195945262909, + 0.008450099267065525, + 0.004732644185423851, + 0.13565312325954437, + 0.006609708536416292, + 0.07426296919584274, + 0.6741131544113159, + 0.006794179789721966, + 0.02010050043463707 + ], + "credit_map": [ + 0.005693943705409765, + 0.0021264045499265194, + 0.017612969502806664, + 0.0023777829483151436, + 0.01601293496787548, + 0.002489815466105938, + 0.003981877118349075, + 0.006188634783029556, + 0.0018622362986207008, + 0.0007341175805777311, + 0.0029928467702120543, + 0.0009497877326793969, + 0.0019307790789753199, + 0.0009683916578069329, + 0.0017028776928782463, + 0.0016581604722887278, + 0.001391913741827011, + 0.0005715616280212998, + 0.0016405065543949604, + 0.0007123834802769125, + 0.0015723685501143336, + 0.0005656111752614379, + 0.0009870523354038596, + 0.0010087019763886929, + 0.0008156939875334501, + 0.00029420919599942863, + 0.0009024309692904353, + 0.0003154506266582757, + 0.0006257324712350965, + 0.00043828689376823604, + 0.0007179694366641343, + 0.0006228708080016077, + 0.009280488826334476, + 0.003700821427628398, + 0.04465102404356003, + 0.004339687526226044, + 0.04946097359061241, + 0.00499021215364337, + 0.009554388001561165, + 0.009675540961325169, + 0.0009742818656377494, + 0.00031985060195438564, + 0.0011450760066509247, + 0.00042126275366172194, + 0.001988049363717437, + 0.0004316441772971302, + 0.0006610268028452992, + 0.0006685176631435752, + 0.005186747293919325, + 0.0021208394318819046, + 0.025871075689792633, + 0.0025796126574277878, + 0.027927596122026443, + 0.002956641837954521, + 0.002634926699101925, + 0.004985543433576822, + 0.05081767216324806, + 0.016420241445302963, + 0.2326173633337021, + 0.021598588675260544, + 0.2546876072883606, + 0.029042687267065048, + 0.024284737184643745, + 0.0446443147957325, + 0.0005494754295796156, + 0.0003371488710399717, + 0.0019938352052122355, + 0.0004725268518086523, + 0.0019733759108930826, + 0.00039179998566396534, + 0.00034228109871037304, + 0.0007337366114370525, + 0.0015984246274456382, + 0.0007080016657710075, + 0.006468879524618387, + 0.0007498024497181177, + 0.007323084864765406, + 0.001020035590045154, + 0.0009060365264303982, + 0.0013262368738651276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.04680928215384483 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0206448994576931 + ], + "opposite_logprob_tokens": [ + -3.8956449031829834 + ], + "logprob_reward": -0.0206448994576931, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0206448994576931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11881062388420105, + 0.03270239755511284, + 0.01622132398188114, + 0.009848515503108501, + 0.21009762585163116, + 0.007948744110763073, + 0.033635467290878296, + 0.5485531091690063, + 0.005180815700441599, + 0.017001420259475708 + ], + "credit_map": [ + 0.009480021893978119, + 0.004754190798848867, + 0.03644360229372978, + 0.0047421506606042385, + 0.04155873507261276, + 0.00494141923263669, + 0.008814835920929909, + 0.008075659163296223, + 0.00456185033544898, + 0.0019318689592182636, + 0.007526189088821411, + 0.0022283263970166445, + 0.005343839060515165, + 0.0022582905367016792, + 0.004922725725919008, + 0.003929303027689457, + 0.0024227502290159464, + 0.0012906392803415656, + 0.003796543227508664, + 0.0014212209498509765, + 0.0021018509287387133, + 0.000980958342552185, + 0.0021568406373262405, + 0.002050519920885563, + 0.0007993251783773303, + 0.0007160176755860448, + 0.002703525824472308, + 0.0005976383690722287, + 0.002531988313421607, + 0.0006367810419760644, + 0.0009293532930314541, + 0.0009338848758488894, + 0.014333400875329971, + 0.005604146979749203, + 0.06978928297758102, + 0.0064422357827425, + 0.07667336612939835, + 0.007653464563190937, + 0.014529055915772915, + 0.015072667971253395, + 0.0007087235571816564, + 0.00040957320015877485, + 0.0022225319407880306, + 0.0005769411218352616, + 0.002100968500599265, + 0.00048655248247087, + 0.0006715192575939, + 0.0007719336426816881, + 0.0022001173347234726, + 0.0010104862740263343, + 0.01174991950392723, + 0.0012696841731667519, + 0.011737573891878128, + 0.0013436013832688332, + 0.0018381951376795769, + 0.0024858901742845774, + 0.04110994189977646, + 0.01339161116629839, + 0.1897842288017273, + 0.017434880137443542, + 0.20694896578788757, + 0.023480506613850594, + 0.019908078014850616, + 0.03649485483765602, + 0.000686155108269304, + 0.0004631568444892764, + 0.0008047589799389243, + 0.0006641700747422874, + 0.0008856425993144512, + 0.00041403478826396167, + 0.0006196920876391232, + 0.0006432048976421356, + 0.0019410441163927317, + 0.0010907206451520324, + 0.0037609029095619917, + 0.0009774345671758056, + 0.005740098189562559, + 0.0011937677627429366, + 0.0009200596250593662, + 0.0013773904647678137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.0206448994576931 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007641489617526531 + ], + "opposite_logprob_tokens": [ + -4.882641315460205 + ], + "logprob_reward": -0.007641489617526531, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007641489617526531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0548781082034111, + 0.020794514566659927, + 0.03146412596106529, + 0.018729928880929947, + 0.38023844361305237, + 0.011527263559401035, + 0.07188120484352112, + 0.36299604177474976, + 0.02591162919998169, + 0.021578803658485413 + ], + "credit_map": [ + 0.005335656926035881, + 0.0025053631979972124, + 0.014956225641071796, + 0.0030103290919214487, + 0.01743284985423088, + 0.0029325345531105995, + 0.003157359780743718, + 0.00554778752848506, + 0.00309560215100646, + 0.0017272901022806764, + 0.0035337063018232584, + 0.001421640394255519, + 0.0037321739364415407, + 0.001366487587802112, + 0.003892346052452922, + 0.0020252682734280825, + 0.0038526407442986965, + 0.002201052848249674, + 0.005713304504752159, + 0.001976753119379282, + 0.007658842951059341, + 0.0016226951265707612, + 0.0055377427488565445, + 0.002901093801483512, + 0.0033329613506793976, + 0.0015167942037805915, + 0.0032055284827947617, + 0.0012051844969391823, + 0.0038577772211283445, + 0.0012697489000856876, + 0.002243262715637684, + 0.0020986723247915506, + 0.025096341967582703, + 0.010316664353013039, + 0.12718714773654938, + 0.012253593653440475, + 0.1381659209728241, + 0.013793634250760078, + 0.02622362971305847, + 0.027201524004340172, + 0.0016273877117782831, + 0.0007610124303027987, + 0.0036743299569934607, + 0.0010043669026345015, + 0.0015684429090470076, + 0.0007056095637381077, + 0.000692154688294977, + 0.0014939586399123073, + 0.004369541071355343, + 0.0018107503419741988, + 0.025645531713962555, + 0.0024830561596900225, + 0.02563301846385002, + 0.002798069268465042, + 0.004046471323817968, + 0.005094767082482576, + 0.027358317747712135, + 0.008827985264360905, + 0.1252216398715973, + 0.011632152833044529, + 0.13685692846775055, + 0.01557331532239914, + 0.013531127944588661, + 0.023994557559490204, + 0.0022026407532393932, + 0.0009850480128079653, + 0.00791103858500719, + 0.0010640788823366165, + 0.009404460899531841, + 0.0014056242071092129, + 0.001067199045792222, + 0.001871539861895144, + 0.0019454953726381063, + 0.001461442792788148, + 0.005477407481521368, + 0.0012420725543051958, + 0.006982498802244663, + 0.0013062796788290143, + 0.001347032724879682, + 0.0018165742512792349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.007641489617526531 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233287811279297 + ], + "opposite_logprob_tokens": [ + -0.8983287811279297 + ], + "logprob_reward": -0.5233287811279297, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233287811279297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02795117162168026, + 0.017882468178868294, + 0.0165263544768095, + 0.00941111333668232, + 0.5340437889099121, + 0.004274094942957163, + 0.05530582740902901, + 0.2776506841182709, + 0.015445072203874588, + 0.04150950163602829 + ], + "credit_map": [ + 0.0032946693245321512, + 0.0018364964053034782, + 0.004203907214105129, + 0.0027469396591186523, + 0.006720470264554024, + 0.00236343196593225, + 0.0032915689516812563, + 0.0034936852753162384, + 0.002431502565741539, + 0.0008815084584057331, + 0.0032865663524717093, + 0.0017811516299843788, + 0.0025485989172011614, + 0.001376996049657464, + 0.002630939008668065, + 0.0029452042654156685, + 0.0020864352118223906, + 0.0010307237971574068, + 0.0037431051023304462, + 0.0012202603975310922, + 0.0031579677015542984, + 0.0010294622043147683, + 0.0020309332758188248, + 0.0022274653892964125, + 0.0013180221430957317, + 0.0006229987484402955, + 0.0014848461141809821, + 0.0008251483668573201, + 0.0013443834614008665, + 0.0006695984629914165, + 0.0018683269154280424, + 0.0012777892407029867, + 0.03482215106487274, + 0.014508802443742752, + 0.17948147654533386, + 0.017030632123351097, + 0.19428229331970215, + 0.019296687096357346, + 0.036425087600946426, + 0.0381966233253479, + 0.0006720030796714127, + 0.00028309153276495636, + 0.0007112415041774511, + 0.00033563748002052307, + 0.0007829692913219333, + 0.0003406173491384834, + 0.0006239792564883828, + 0.0005245550419203937, + 0.003735660808160901, + 0.001708661438897252, + 0.019605077803134918, + 0.0019653437193483114, + 0.020038489252328873, + 0.0022227063309401274, + 0.0019894696306437254, + 0.004040414933115244, + 0.020659945905208588, + 0.007067830767482519, + 0.09665902704000473, + 0.008878711611032486, + 0.10415852069854736, + 0.011684578843414783, + 0.009898190386593342, + 0.018643874675035477, + 0.0013375665294006467, + 0.0009538737358525395, + 0.004419898614287376, + 0.000989660038612783, + 0.004490791354328394, + 0.0009183462243527174, + 0.0010157638462260365, + 0.00131917092949152, + 0.0035121929831802845, + 0.0017712562112137675, + 0.012286226265132427, + 0.0018914674874395132, + 0.014840526506304741, + 0.0022580891381949186, + 0.0018811222398653626, + 0.003068619640544057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.5233287811279297 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pristine white ceramic mold with a classic grooved ring design resting on a warm, rustic wooden kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mold's rim filled with a vibrant, translucent red gel that is initially smooth and mirror-like?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gel undergo a sequence of rapid violent swaying followed by a gradual subsidence into a final subtle ripple before returning to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright natural daylight streaming in from an off-camera window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604928970336914 + ], + "opposite_logprob_tokens": [ + -1.9104928970336914 + ], + "logprob_reward": -0.1604928970336914, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604928970336914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014641420915722847, + 0.003646373748779297, + 0.0038612522184848785, + 0.0023569308687001467, + 0.0548008531332016, + 0.00824046228080988, + 0.021338049322366714, + 0.8868845105171204, + 0.003605685429647565, + 0.0006244945107027888 + ], + "credit_map": [ + 0.0016955625033006072, + 0.0005546233733184636, + 0.005618471652269363, + 0.0006167346728034317, + 0.0037972237914800644, + 0.0005257418961264193, + 0.00048142235027626157, + 0.0013516421895474195, + 0.0007545428234152496, + 0.0001880854688351974, + 0.0011407871497794986, + 0.00026381833595223725, + 0.00039253936847671866, + 0.00021237967303022742, + 0.00028951046988368034, + 0.00040471047395840287, + 0.0005907408194616437, + 0.00026596232783049345, + 0.000609168375376612, + 0.0003368500038050115, + 0.0008594382670708001, + 0.0002297786413691938, + 0.0005162485176697373, + 0.0004530656151473522, + 0.0003548706881701946, + 0.00016766239423304796, + 0.0006548729143105447, + 0.00011970458581345156, + 0.00035546786966733634, + 0.00015404475561808795, + 0.00031289036269299686, + 0.00023741740733385086, + 0.003694562241435051, + 0.0015061760786920786, + 0.018111057579517365, + 0.0017980302218347788, + 0.019861329346895218, + 0.002039309125393629, + 0.0038996990770101547, + 0.0038906894624233246, + 0.0007842769264243543, + 0.0004265674506314099, + 0.0029647822957485914, + 0.00047381228068843484, + 0.002069134032353759, + 0.0003429441712796688, + 0.00032451789593324065, + 0.0008544271695427597, + 0.0015708006685599685, + 0.0006228174897842109, + 0.007262715138494968, + 0.0007206835434772074, + 0.008078664541244507, + 0.0008745770901441574, + 0.0007483224617317319, + 0.0014594699023291469, + 0.0665799155831337, + 0.021633490920066833, + 0.30695950984954834, + 0.028400376439094543, + 0.33452507853507996, + 0.03820743039250374, + 0.03170427307486534, + 0.058874405920505524, + 0.00031177213531918824, + 0.00013434277207124978, + 0.0011106999590992928, + 0.0001447367831133306, + 0.0013239451218396425, + 0.0001850027620093897, + 0.0001328698272118345, + 0.0002623161126393825, + 5.067185702500865e-05, + 3.974627543357201e-05, + 0.00017413207388017327, + 5.436499850475229e-05, + 0.00015609960246365517, + 3.663367897388525e-05, + 4.167974111624062e-05, + 7.116625783964992e-05 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.1604928970336914 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_4/index_tensor([5350, 5350])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_4/index_tensor([5350, 5350])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c941378195a719652c4158ec3dad9851b5c7070a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_4/index_tensor([5350, 5350])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly low-angle, extreme wide shot captures a breathtaking panorama rendered in a beautiful watercolor animation style. A lone high school girl, resembling Ichika Miyamizu with her short, dark brown bob hairstyle, stands in quiet contemplation on a windswept hilltop amidst the grass. Clad in her dark school uniform, she forms a delicate silhouette against the magnificent celestial backdrop. Above her, the aurora borealis dances across the vast, star-studded night sky, forming a vibrant, shimmering ribbon of green and purple. These celestial lights gracefully ripple and swirl, their brilliant colors softly seeping into the sky's deep indigo canvas, creating diffused painterly edges, as if wet paint were spreading across textured paper. The otherworldly glow of the aurora casts a cool, magical glow on the mountaintop and over the girl's shoulders, creating a profound atmosphere of quiet wonder and solitary awe. A gentle, cool breeze gently ruffles her hair and the hem of her skirt, the only subtle movement on the ground, contrasting sharply with the dynamic, silent scene unfolding in the sky.", + "seed": 1406, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_4/index_5350_seed_1406_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.16845205635763705, + "vlm_logprob": -0.16845205635763705, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.16845205635763705, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7079541683197021 + ], + "opposite_logprob_tokens": [ + -0.7079541683197021 + ], + "logprob_reward": -0.7079541683197021, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7079541683197021, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17133204638957977, + 0.057278841733932495, + 0.05913209170103073, + 0.14283862709999084, + 0.0669877752661705, + 0.1304806023836136, + 0.1807464063167572, + 0.11820194870233536, + 0.034758295863866806, + 0.03824342414736748 + ], + "credit_map": [ + 0.017628101631999016, + 0.0195116326212883, + 0.020603150129318237, + 0.030584851279854774, + 0.022129876539111137, + 0.02111629955470562, + 0.027301454916596413, + 0.012456666678190231, + 0.004940595477819443, + 0.013493052683770657, + 0.009764607064425945, + 0.006606833077967167, + 0.006447428837418556, + 0.006150702480226755, + 0.006354459561407566, + 0.003521151840686798, + 0.007245141081511974, + 0.009326915256679058, + 0.010279683396220207, + 0.007154358085244894, + 0.005397678352892399, + 0.008140559308230877, + 0.007276033516973257, + 0.004311718512326479, + 0.01388944499194622, + 0.025196529924869537, + 0.01496063731610775, + 0.018276769667863846, + 0.01361776515841484, + 0.02455318532884121, + 0.019795382395386696, + 0.012548903934657574, + 0.0069869267754256725, + 0.010921662673354149, + 0.011464321985840797, + 0.007529365364462137, + 0.007038038689643145, + 0.0072794463485479355, + 0.010708802379667759, + 0.005059205926954746, + 0.01203810703009367, + 0.013556817546486855, + 0.020483603700995445, + 0.011676148511469364, + 0.011824879795312881, + 0.014691533520817757, + 0.033259496092796326, + 0.012950011529028416, + 0.020463716238737106, + 0.017386717721819878, + 0.022878918796777725, + 0.02079213224351406, + 0.015025799162685871, + 0.019337553530931473, + 0.04254159331321716, + 0.022319955751299858, + 0.011946850456297398, + 0.012207445688545704, + 0.014185638166964054, + 0.01399088092148304, + 0.01215537078678608, + 0.013544936664402485, + 0.028036393225193024, + 0.012134420685470104, + 0.003499504178762436, + 0.003837218042463064, + 0.004632962867617607, + 0.0046400572173297405, + 0.004858823027461767, + 0.005479191895574331, + 0.005315228831022978, + 0.0024953079409897327, + 0.003601853270083666, + 0.0026629322674125433, + 0.005510926712304354, + 0.006093418225646019, + 0.00521960249170661, + 0.00744994031265378, + 0.004592709243297577, + 0.0031120372004806995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.7079541683197021 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263705775141716 + ], + "opposite_logprob_tokens": [ + -3.6513705253601074 + ], + "logprob_reward": -0.0263705775141716, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263705775141716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11116451770067215, + 0.08779219537973404, + 0.058694563806056976, + 0.14161473512649536, + 0.06570645421743393, + 0.0834287777543068, + 0.16260388493537903, + 0.19809262454509735, + 0.04095304384827614, + 0.049949198961257935 + ], + "credit_map": [ + 0.008772476576268673, + 0.018599679693579674, + 0.01390383206307888, + 0.02101955935359001, + 0.015725618228316307, + 0.012649892829358578, + 0.012716179713606834, + 0.007777282036840916, + 0.007629521656781435, + 0.016664545983076096, + 0.01515510305762291, + 0.010516566224396229, + 0.012805268168449402, + 0.009762709029018879, + 0.010908504948019981, + 0.0043499767780303955, + 0.007990787737071514, + 0.00873082596808672, + 0.0068800062872469425, + 0.008476775139570236, + 0.005981693509966135, + 0.008394548669457436, + 0.006599334999918938, + 0.005640590097755194, + 0.013853668235242367, + 0.02494332380592823, + 0.015270340256392956, + 0.02301637828350067, + 0.01593627780675888, + 0.02238776721060276, + 0.013645515777170658, + 0.012561459094285965, + 0.006046772934496403, + 0.010234872810542583, + 0.010160130448639393, + 0.009901282377541065, + 0.0058196368627250195, + 0.0056869033724069595, + 0.011329716071486473, + 0.0065271384082734585, + 0.009805861860513687, + 0.011103787459433079, + 0.012619021348655224, + 0.013782908208668232, + 0.007077652029693127, + 0.00747331976890564, + 0.013276259414851665, + 0.008289970457553864, + 0.015986625105142593, + 0.015953553840517998, + 0.02990524098277092, + 0.0131582235917449, + 0.015061823651194572, + 0.02755451574921608, + 0.029737967997789383, + 0.01524594146758318, + 0.019376296550035477, + 0.01840152218937874, + 0.03541262447834015, + 0.03179919347167015, + 0.018090685829520226, + 0.015581389889121056, + 0.03907864913344383, + 0.020352251827716827, + 0.0045003741979599, + 0.003910934552550316, + 0.004833502229303122, + 0.005175171419978142, + 0.005710597615689039, + 0.006388336420059204, + 0.006338214501738548, + 0.004095911048352718, + 0.005064734257757664, + 0.005041460506618023, + 0.013311284594237804, + 0.004798839800059795, + 0.004200195427983999, + 0.007417420856654644, + 0.005567769054323435, + 0.004547493066638708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.0263705775141716 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307207837700844 + ], + "opposite_logprob_tokens": [ + -3.168071985244751 + ], + "logprob_reward": -0.04307207837700844, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307207837700844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11072264611721039, + 0.04239780455827713, + 0.052645277231931686, + 0.09873444586992264, + 0.06900307536125183, + 0.11525221914052963, + 0.17766475677490234, + 0.23859566450119019, + 0.04203738644719124, + 0.052946701645851135 + ], + "credit_map": [ + 0.011122992262244225, + 0.015717191621661186, + 0.012117769569158554, + 0.021532472223043442, + 0.011756004765629768, + 0.011383925564587116, + 0.016780799254775047, + 0.010311492718756199, + 0.0033598372247070074, + 0.010747534222900867, + 0.007170988712459803, + 0.004475925117731094, + 0.005351223982870579, + 0.0035152663476765156, + 0.005023578647524118, + 0.0027534516993910074, + 0.004156833980232477, + 0.003947235643863678, + 0.0087107690051198, + 0.006155729293823242, + 0.008133652620017529, + 0.004377604927867651, + 0.012636333703994751, + 0.004527114797383547, + 0.0071877059526741505, + 0.014432298950850964, + 0.013364903628826141, + 0.01067419070750475, + 0.00967838428914547, + 0.012206310406327248, + 0.020428815856575966, + 0.01076184120029211, + 0.004354400094598532, + 0.007751214783638716, + 0.010684258304536343, + 0.0077023692429065704, + 0.007054533809423447, + 0.004079102538526058, + 0.018925532698631287, + 0.00845166202634573, + 0.00919575896114111, + 0.009229883551597595, + 0.018116643652319908, + 0.014315861277282238, + 0.006706353276968002, + 0.006672982592135668, + 0.03484239801764488, + 0.016172347590327263, + 0.015266058035194874, + 0.018112417310476303, + 0.03291713818907738, + 0.013668549247086048, + 0.017717968672513962, + 0.027601012960076332, + 0.036043956875801086, + 0.01633765734732151, + 0.020987844094634056, + 0.019923610612750053, + 0.038725223392248154, + 0.03675468638539314, + 0.022568728774785995, + 0.01740717887878418, + 0.05912652611732483, + 0.023101860657334328, + 0.00536483246833086, + 0.004322279244661331, + 0.00688962871208787, + 0.00516137620434165, + 0.006171811372041702, + 0.004364393651485443, + 0.006626646965742111, + 0.003136415034532547, + 0.005200076848268509, + 0.0026853266172111034, + 0.013374767266213894, + 0.004951565060764551, + 0.0091015063226223, + 0.00812511332333088, + 0.0059638358652591705, + 0.0035445131361484528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.04307207837700844 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818423556163907 + ], + "opposite_logprob_tokens": [ + -5.877818584442139 + ], + "logprob_reward": -0.002818423556163907, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818423556163907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13296709954738617, + 0.08124514669179916, + 0.12781356275081635, + 0.12581199407577515, + 0.10436966270208359, + 0.05272558704018593, + 0.11102811247110367, + 0.13043096661567688, + 0.09035613387823105, + 0.043251700699329376 + ], + "credit_map": [ + 0.014538304880261421, + 0.030083835124969482, + 0.01696784794330597, + 0.021490037441253662, + 0.014993255026638508, + 0.014471612870693207, + 0.011332152411341667, + 0.009090040810406208, + 0.0052873678505420685, + 0.026933874934911728, + 0.01745712012052536, + 0.006364546250551939, + 0.00854765810072422, + 0.009403130039572716, + 0.00474033085629344, + 0.0025111159775406122, + 0.02010084129869938, + 0.01769191399216652, + 0.016335636377334595, + 0.02552206814289093, + 0.01367289386689663, + 0.015322876162827015, + 0.010208643972873688, + 0.008958683349192142, + 0.015758849680423737, + 0.017575176432728767, + 0.018443213775753975, + 0.02311781421303749, + 0.011987769976258278, + 0.014463141560554504, + 0.012984453700482845, + 0.011481573805212975, + 0.007574834395200014, + 0.018500616773962975, + 0.016071347519755363, + 0.016007138416171074, + 0.008855035528540611, + 0.007743594236671925, + 0.017604224383831024, + 0.012012869119644165, + 0.005580070894211531, + 0.006634070072323084, + 0.006505690049380064, + 0.004940134007483721, + 0.00399428978562355, + 0.005439985077828169, + 0.011608134023845196, + 0.008023211732506752, + 0.009956754744052887, + 0.011250615119934082, + 0.016063639894127846, + 0.009063279256224632, + 0.0110706752166152, + 0.021345065906643867, + 0.0192972794175148, + 0.012980801053345203, + 0.014223913662135601, + 0.012735632248222828, + 0.01755221001803875, + 0.014768818393349648, + 0.013745809905230999, + 0.01610974222421646, + 0.029179556295275688, + 0.012115275487303734, + 0.010127034038305283, + 0.008716735057532787, + 0.02029799483716488, + 0.009018437936902046, + 0.014578616246581078, + 0.009522936306893826, + 0.011752254329621792, + 0.006342124193906784, + 0.0034204539842903614, + 0.004792611114680767, + 0.009347534738481045, + 0.005129778757691383, + 0.00656740739941597, + 0.005286583676934242, + 0.0053878785111010075, + 0.00331944995559752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.002818423556163907 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062045034021139145 + ], + "opposite_logprob_tokens": [ + -2.812045097351074 + ], + "logprob_reward": -0.062045034021139145, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062045034021139145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06041037663817406, + 0.05592604726552963, + 0.04285109043121338, + 0.04354936257004738, + 0.03765571862459183, + 0.08690083771944046, + 0.11367952078580856, + 0.4097769260406494, + 0.10850070416927338, + 0.04074938967823982 + ], + "credit_map": [ + 0.005794375669211149, + 0.0069242920726537704, + 0.007229367736726999, + 0.012057904154062271, + 0.006136904004961252, + 0.005790519993752241, + 0.011085216887295246, + 0.0053918021731078625, + 0.003608623519539833, + 0.0189539585262537, + 0.011848763562738895, + 0.0034139195922762156, + 0.005401719361543655, + 0.005986085161566734, + 0.0050004152581095695, + 0.0017125658923760056, + 0.004361391067504883, + 0.0033352866303175688, + 0.006525600794702768, + 0.0054559833370149136, + 0.0036183088086545467, + 0.0034795000683516264, + 0.011808548122644424, + 0.00426647812128067, + 0.004817798733711243, + 0.006338353734463453, + 0.006857284344732761, + 0.00723178731277585, + 0.004292892757803202, + 0.004594869911670685, + 0.006525107193738222, + 0.002891272772103548, + 0.0035652024671435356, + 0.004743664059787989, + 0.005354739259928465, + 0.004754707217216492, + 0.0032715722918510437, + 0.0034281143452972174, + 0.0078612444922328, + 0.004676477517932653, + 0.00840357132256031, + 0.007337683346122503, + 0.013725114986300468, + 0.010908662341535091, + 0.0045560807920992374, + 0.006263545248657465, + 0.02671961858868599, + 0.008986573666334152, + 0.010399852879345417, + 0.009739622473716736, + 0.016006791964173317, + 0.011269878596067429, + 0.009030652232468128, + 0.018134471029043198, + 0.028655434027314186, + 0.010442831553518772, + 0.03546430915594101, + 0.0313725508749485, + 0.06826604902744293, + 0.06676281243562698, + 0.03734102100133896, + 0.028211021795868874, + 0.09918462485074997, + 0.043174561113119125, + 0.010873747058212757, + 0.009147152304649353, + 0.025697268545627594, + 0.012680835090577602, + 0.017602430656552315, + 0.010588694363832474, + 0.015974361449480057, + 0.005936227738857269, + 0.0028788908384740353, + 0.003296520793810487, + 0.007265785709023476, + 0.003579970682039857, + 0.007660471368581057, + 0.0066566127352416515, + 0.005386679898947477, + 0.004024461843073368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.062045034021139145 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle, extreme wide shot captures a breathtaking panorama rendered in a beautiful watercolor animation style. A lone high school girl, resembling Ichika Miyamizu with her short, dark brown bob hairstyle, stands in quiet contemplation on a windswept hilltop amidst the grass. Clad in her dark school uniform, she forms a delicate silhouette against the magnificent celestial backdrop. Above her, the aurora borealis dances across the vast, star-studded night sky, forming a vibrant, shimmering ribbon of green and purple. These celestial lights gracefully ripple and swirl, their brilliant colors softly seeping into the sky's deep indigo canvas, creating diffused painterly edges, as if wet paint were spreading across textured paper. The otherworldly glow of the aurora casts a cool, magical glow on the mountaintop and over the girl's shoulders, creating a profound atmosphere of quiet wonder and solitary awe. A gentle, cool breeze gently ruffles her hair and the hem of her skirt, the only subtle movement on the ground, contrasting sharply with the dynamic, silent scene unfolding in the sky.", + "seed": 1407, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_4/index_5350_seed_1407_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2177420352352783, + "vlm_logprob": -0.2177420352352783, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.2177420352352783, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9068967700004578 + ], + "opposite_logprob_tokens": [ + -0.5318967700004578 + ], + "logprob_reward": -0.9068967700004578, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9068967700004578, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07219479233026505, + 0.06614159792661667, + 0.08787383884191513, + 0.3053101599216461, + 0.1618865728378296, + 0.06993871182203293, + 0.08397095650434494, + 0.022417563945055008, + 0.03606926277279854, + 0.09419649839401245 + ], + "credit_map": [ + 0.006908770650625229, + 0.01258655171841383, + 0.006593702826648951, + 0.007933792658150196, + 0.007935736328363419, + 0.009230722673237324, + 0.009143691509962082, + 0.011861818842589855, + 0.007878891192376614, + 0.011468257755041122, + 0.006736745126545429, + 0.008071141317486763, + 0.0062063392251729965, + 0.008480322547256947, + 0.0071577453054487705, + 0.010142156854271889, + 0.01424490287899971, + 0.0062007056549191475, + 0.00692408625036478, + 0.007980210706591606, + 0.006222826428711414, + 0.007530622649937868, + 0.014576535671949387, + 0.02419395186007023, + 0.062140703201293945, + 0.03239727392792702, + 0.06976368278265, + 0.037657253444194794, + 0.046885646879673004, + 0.03179338201880455, + 0.012735569849610329, + 0.011936644092202187, + 0.05100866034626961, + 0.013110318221151829, + 0.050265852361917496, + 0.019467251375317574, + 0.007640428841114044, + 0.009895634837448597, + 0.00457220571115613, + 0.005926234647631645, + 0.01404709555208683, + 0.007391412276774645, + 0.01634093001484871, + 0.008537995629012585, + 0.005349429789930582, + 0.005429335869848728, + 0.00541862053796649, + 0.007423891220241785, + 0.009730524383485317, + 0.009026998654007912, + 0.010982663370668888, + 0.013694590888917446, + 0.010443931445479393, + 0.011790584772825241, + 0.00825703889131546, + 0.010044624097645283, + 0.0026773165445774794, + 0.004496273584663868, + 0.0020866000559180975, + 0.0035029512364417315, + 0.0026260840240865946, + 0.0019141955999657512, + 0.0026065013371407986, + 0.002507641213014722, + 0.0039543104358017445, + 0.0073183393105864525, + 0.003442120738327503, + 0.005938318092375994, + 0.004298253916203976, + 0.003598699113354087, + 0.004078456200659275, + 0.0034407644998282194, + 0.01133712287992239, + 0.015737103298306465, + 0.01220590528100729, + 0.016213057562708855, + 0.012644304893910885, + 0.010767963714897633, + 0.007715513464063406, + 0.007575525902211666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.9068967700004578 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029813595116138458 + ], + "opposite_logprob_tokens": [ + -3.529813528060913 + ], + "logprob_reward": -0.029813595116138458, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029813595116138458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29692214727401733, + 0.12201400846242905, + 0.03219373896718025, + 0.1954910308122635, + 0.12164350599050522, + 0.03924274444580078, + 0.050972383469343185, + 0.07532474398612976, + 0.03169922158122063, + 0.03449650853872299 + ], + "credit_map": [ + 0.024030080065131187, + 0.04264169931411743, + 0.0333775095641613, + 0.032981354743242264, + 0.03835843876004219, + 0.06013532727956772, + 0.029512984678149223, + 0.03588474914431572, + 0.010564271360635757, + 0.022158820182085037, + 0.009569906629621983, + 0.014948181807994843, + 0.014984317123889923, + 0.018274199217557907, + 0.015406020916998386, + 0.016108276322484016, + 0.0034896191209554672, + 0.0037995825987309217, + 0.004844164941459894, + 0.0040438007563352585, + 0.002941296435892582, + 0.0027921840082854033, + 0.004910013172775507, + 0.005373077932745218, + 0.02246200107038021, + 0.03633802384138107, + 0.016367146745324135, + 0.020149599760770798, + 0.04403233900666237, + 0.028753574937582016, + 0.016077920794487, + 0.011310410685837269, + 0.036003392189741135, + 0.008603636175394058, + 0.035370681434869766, + 0.014546987600624561, + 0.006542663089931011, + 0.00830422155559063, + 0.005775805097073317, + 0.006496102083474398, + 0.005318677984178066, + 0.004770423751324415, + 0.0032434677705168724, + 0.004866184201091528, + 0.005938821472227573, + 0.007652810774743557, + 0.0033995467238128185, + 0.004052806179970503, + 0.005724638234823942, + 0.010192789137363434, + 0.006046182010322809, + 0.008671856485307217, + 0.006351672578603029, + 0.003924905322492123, + 0.004667456261813641, + 0.00539288017898798, + 0.013009793125092983, + 0.008088611997663975, + 0.00655110040679574, + 0.011012913659214973, + 0.009371732361614704, + 0.007813083939254284, + 0.00982692651450634, + 0.00965057872235775, + 0.005020450800657272, + 0.0035753631964325905, + 0.0032125250436365604, + 0.004689960740506649, + 0.003877936163917184, + 0.0034703388810157776, + 0.00455462047830224, + 0.003298026043921709, + 0.004003420006483793, + 0.0065900650806725025, + 0.00416909996420145, + 0.005380640737712383, + 0.00407065125182271, + 0.004059326369315386, + 0.003880786942318082, + 0.0023425172548741102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.029813595116138458 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904458582401276 + ], + "opposite_logprob_tokens": [ + -2.4640445709228516 + ], + "logprob_reward": -0.08904458582401276, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904458582401276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15521131455898285, + 0.02623133361339569, + 0.08627896010875702, + 0.025199009105563164, + 0.12906648218631744, + 0.5052662491798401, + 0.024438994005322456, + 0.012818482704460621, + 0.017000727355480194, + 0.01848844811320305 + ], + "credit_map": [ + 0.011371172964572906, + 0.015997886657714844, + 0.01560670230537653, + 0.01692003570497036, + 0.028109023347496986, + 0.03199245035648346, + 0.01566714607179165, + 0.01954689621925354, + 0.0028460638131946325, + 0.005148356780409813, + 0.0033068815246224403, + 0.0031728472094982862, + 0.0026268186047673225, + 0.0033282944932579994, + 0.0029928553849458694, + 0.0028092158026993275, + 0.01412135735154152, + 0.005431485362350941, + 0.00715482234954834, + 0.007963261567056179, + 0.00545034883543849, + 0.007790186908096075, + 0.014734983444213867, + 0.023632517084479332, + 0.0032622623257339, + 0.003970257937908173, + 0.0028533681761473417, + 0.0037262924015522003, + 0.0030339534860104322, + 0.003505613189190626, + 0.0022199798841029406, + 0.0026272819377481937, + 0.0334276482462883, + 0.011804414913058281, + 0.030122866854071617, + 0.01316885557025671, + 0.013143481686711311, + 0.011986575089395046, + 0.008732850663363934, + 0.00667978823184967, + 0.16166134178638458, + 0.022726599127054214, + 0.17290335893630981, + 0.06624817103147507, + 0.030060751363635063, + 0.018681485205888748, + 0.013940987177193165, + 0.01904354616999626, + 0.0030294570606201887, + 0.003494668984785676, + 0.0025398964062333107, + 0.0040369234047830105, + 0.003032719250768423, + 0.003383407834917307, + 0.0029006563127040863, + 0.002021264052018523, + 0.001594739267602563, + 0.0017039352096617222, + 0.001343987649306655, + 0.0016338912537321448, + 0.0013701418647542596, + 0.0017932761693373322, + 0.001654022024013102, + 0.0017244892660528421, + 0.0020448733121156693, + 0.0020776453893631697, + 0.0015826303279027343, + 0.002838331740349531, + 0.0017000865191221237, + 0.0018903686432167888, + 0.0024952259846031666, + 0.002371566602960229, + 0.0024458810221403837, + 0.0021097692660987377, + 0.002503742463886738, + 0.002049574861302972, + 0.0020802796352654696, + 0.0020220321603119373, + 0.0021364118438214064, + 0.003140757093206048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.08904458582401276 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009190387791022658 + ], + "opposite_logprob_tokens": [ + -7.000918865203857 + ], + "logprob_reward": -0.0009190387791022658, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009190387791022658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037893928587436676, + 0.08654924482107162, + 0.03867008537054062, + 0.2235632836818695, + 0.09532754868268967, + 0.4120175242424011, + 0.04332725331187248, + 0.010857487097382545, + 0.02645201049745083, + 0.02534165233373642 + ], + "credit_map": [ + 0.003084292635321617, + 0.006264110095798969, + 0.0030662978533655405, + 0.004446492996066809, + 0.008769669570028782, + 0.004875554237514734, + 0.004065657965838909, + 0.0033218543976545334, + 0.009394575841724873, + 0.015019228681921959, + 0.010783887468278408, + 0.011068691499531269, + 0.008436924777925014, + 0.01416324544698, + 0.0077823675237596035, + 0.009900320321321487, + 0.005416336935013533, + 0.004565452691167593, + 0.004384497180581093, + 0.004409179091453552, + 0.0028053191490471363, + 0.0038735466077923775, + 0.004623310174793005, + 0.008592439815402031, + 0.06039413809776306, + 0.016293130815029144, + 0.06795431673526764, + 0.03046593815088272, + 0.01977073960006237, + 0.015693210065364838, + 0.006062197033315897, + 0.006929612252861261, + 0.02755833975970745, + 0.009611648507416248, + 0.024283573031425476, + 0.010208368301391602, + 0.007701820228248835, + 0.007782367989420891, + 0.00381208723410964, + 0.004369334317743778, + 0.13077840209007263, + 0.01969386450946331, + 0.14017216861248016, + 0.05435189604759216, + 0.024332353845238686, + 0.0163036547601223, + 0.011035840027034283, + 0.015349337831139565, + 0.005420504603534937, + 0.006187314633280039, + 0.004786994773894548, + 0.006982672959566116, + 0.0049387626349925995, + 0.006516423542052507, + 0.00443590572103858, + 0.00405867537483573, + 0.0015428046463057399, + 0.0014432467287406325, + 0.0019041933119297028, + 0.0012060446897521615, + 0.001280917669646442, + 0.0012218371266499162, + 0.0011289286194369197, + 0.001129515003412962, + 0.0036370165180414915, + 0.0038357803132385015, + 0.0025447930674999952, + 0.004423791542649269, + 0.0041001057252287865, + 0.0036996244452893734, + 0.0021981203462928534, + 0.0020127790048718452, + 0.0032680758740752935, + 0.0041838535107672215, + 0.0030630812980234623, + 0.0040843854658305645, + 0.003565259277820587, + 0.002976702991873026, + 0.0020970231853425503, + 0.002103270962834358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.0009190387791022658 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620361864566803 + ], + "opposite_logprob_tokens": [ + -2.8120362758636475 + ], + "logprob_reward": -0.0620361864566803, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620361864566803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06436807662248611, + 0.04262583330273628, + 0.025364799425005913, + 0.028723809868097305, + 0.10671289265155792, + 0.6443119049072266, + 0.02933870628476143, + 0.017048994079232216, + 0.02366132102906704, + 0.01784370094537735 + ], + "credit_map": [ + 0.005842659156769514, + 0.007536450866609812, + 0.009266933426260948, + 0.007185364607721567, + 0.010810483247041702, + 0.010106734000146389, + 0.0066075194627046585, + 0.007011924404650927, + 0.0046247015707194805, + 0.008365661837160587, + 0.005880267824977636, + 0.0052687195129692554, + 0.004476252943277359, + 0.006693345960229635, + 0.003606037702411413, + 0.0037108450196683407, + 0.002988781314343214, + 0.004193526692688465, + 0.002868280978873372, + 0.0025827360805124044, + 0.002105303341522813, + 0.0033915836829692125, + 0.003025803016498685, + 0.004208781290799379, + 0.005392709746956825, + 0.0028252776246517897, + 0.006058651953935623, + 0.003525321837514639, + 0.004195853136479855, + 0.0032450086437165737, + 0.0016906687524169683, + 0.001790315262041986, + 0.026692872866988182, + 0.012317330576479435, + 0.02219824306666851, + 0.011023598723113537, + 0.008329972624778748, + 0.009303644299507141, + 0.008510329760611057, + 0.00833689235150814, + 0.20906399190425873, + 0.02794707752764225, + 0.22307035326957703, + 0.0843176543712616, + 0.038890015333890915, + 0.02212991565465927, + 0.016498293727636337, + 0.022394590079784393, + 0.003519788384437561, + 0.004991430789232254, + 0.0037511910777539015, + 0.004646705463528633, + 0.003613285953179002, + 0.0035641915164887905, + 0.0029188974294811487, + 0.002333211712539196, + 0.0026004938408732414, + 0.002424695296213031, + 0.0017051061149686575, + 0.0022532660514116287, + 0.0020715976133942604, + 0.001752982148900628, + 0.002170056337490678, + 0.002070796210318804, + 0.002563840476796031, + 0.002848982810974121, + 0.002665083622559905, + 0.0045114378444850445, + 0.0021740272641181946, + 0.002389408415183425, + 0.003621634328737855, + 0.002886903937906027, + 0.002247377997264266, + 0.002059218706563115, + 0.0018100248416885734, + 0.0027840693946927786, + 0.0016349246725440025, + 0.001634271233342588, + 0.0032552178017795086, + 0.002418594202026725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.0620361864566803 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle, extreme wide shot captures a breathtaking panorama rendered in a beautiful watercolor animation style. A lone high school girl, resembling Ichika Miyamizu with her short, dark brown bob hairstyle, stands in quiet contemplation on a windswept hilltop amidst the grass. Clad in her dark school uniform, she forms a delicate silhouette against the magnificent celestial backdrop. Above her, the aurora borealis dances across the vast, star-studded night sky, forming a vibrant, shimmering ribbon of green and purple. These celestial lights gracefully ripple and swirl, their brilliant colors softly seeping into the sky's deep indigo canvas, creating diffused painterly edges, as if wet paint were spreading across textured paper. The otherworldly glow of the aurora casts a cool, magical glow on the mountaintop and over the girl's shoulders, creating a profound atmosphere of quiet wonder and solitary awe. A gentle, cool breeze gently ruffles her hair and the hem of her skirt, the only subtle movement on the ground, contrasting sharply with the dynamic, silent scene unfolding in the sky.", + "seed": 1408, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_4/index_5350_seed_1408_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.22549103270284832, + "vlm_logprob": -0.22549103270284832, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.22549103270284832, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9864751100540161 + ], + "opposite_logprob_tokens": [ + -0.4864751100540161 + ], + "logprob_reward": -0.9864751100540161, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9864751100540161, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20338371396064758, + 0.045120734721422195, + 0.03126893937587738, + 0.1450919359922409, + 0.06234394758939743, + 0.06358326971530914, + 0.28060877323150635, + 0.10898805409669876, + 0.036366432905197144, + 0.02324414812028408 + ], + "credit_map": [ + 0.01866237260401249, + 0.01744467206299305, + 0.034036751836538315, + 0.01949034444987774, + 0.02966965362429619, + 0.03305331990122795, + 0.025989633053541183, + 0.02503696084022522, + 0.00626047607511282, + 0.005211467854678631, + 0.005481875501573086, + 0.0050966753624379635, + 0.005664377007633448, + 0.005123618058860302, + 0.005372127052396536, + 0.006910115946084261, + 0.005091089755296707, + 0.0024953328538686037, + 0.004471285734325647, + 0.003917186986654997, + 0.004184307064861059, + 0.003977084532380104, + 0.003665590425953269, + 0.0034670631866902113, + 0.03718462958931923, + 0.008715542033314705, + 0.01278473250567913, + 0.008179979398846626, + 0.04046949744224548, + 0.01898503303527832, + 0.011747600510716438, + 0.007024910766631365, + 0.0059430464170873165, + 0.00544329546391964, + 0.009327036328613758, + 0.00783905852586031, + 0.00815337523818016, + 0.009583218023180962, + 0.007811167277395725, + 0.008243749849498272, + 0.009138991124927998, + 0.005935365799814463, + 0.010318776592612267, + 0.006299009546637535, + 0.008558463305234909, + 0.008469071239233017, + 0.008077233098447323, + 0.006786358542740345, + 0.021923746913671494, + 0.018882205709815025, + 0.01893460936844349, + 0.0073919594287872314, + 0.08569205552339554, + 0.10597579926252365, + 0.009204164147377014, + 0.012604247778654099, + 0.010250807739794254, + 0.01097201369702816, + 0.015746893361210823, + 0.015148699283599854, + 0.018834806978702545, + 0.012793570756912231, + 0.01189752109348774, + 0.013343747705221176, + 0.0030834136996418238, + 0.0061935679987072945, + 0.004608297720551491, + 0.006306006573140621, + 0.0035986071452498436, + 0.005404915194958448, + 0.0025608742143958807, + 0.004610748030245304, + 0.0023798677138984203, + 0.003272661939263344, + 0.002470447914674878, + 0.0037263543345034122, + 0.0031108304392546415, + 0.003014498855918646, + 0.0016891458071768284, + 0.003580341348424554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.9864751100540161 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020591532811522484 + ], + "opposite_logprob_tokens": [ + -3.8955914974212646 + ], + "logprob_reward": -0.020591532811522484, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020591532811522484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15234583616256714, + 0.09181592613458633, + 0.03829231113195419, + 0.06060231849551201, + 0.03219262883067131, + 0.03681834414601326, + 0.4483630359172821, + 0.10859307646751404, + 0.010010497644543648, + 0.020966019481420517 + ], + "credit_map": [ + 0.01733994483947754, + 0.015743382275104523, + 0.0175639558583498, + 0.015185743570327759, + 0.02039422281086445, + 0.016027549281716347, + 0.015766805037856102, + 0.03432422876358032, + 0.015474152751266956, + 0.009391514584422112, + 0.010643075220286846, + 0.010991073213517666, + 0.013022537343204021, + 0.009024117141962051, + 0.011560148559510708, + 0.011709311977028847, + 0.004949788562953472, + 0.0036102321464568377, + 0.0052554854191839695, + 0.0032836783211678267, + 0.00726489769294858, + 0.0059959664940834045, + 0.004698201082646847, + 0.003234061412513256, + 0.015170733444392681, + 0.004125342704355717, + 0.00546431727707386, + 0.0028887710068374872, + 0.016905179247260094, + 0.0074022612534463406, + 0.004750712309032679, + 0.0038950012531131506, + 0.0039411429315805435, + 0.0032979801762849092, + 0.0031857071444392204, + 0.0036692230496555567, + 0.006668137852102518, + 0.004169927444308996, + 0.004131389316171408, + 0.003129124641418457, + 0.004597662948071957, + 0.004117763601243496, + 0.0036598192527890205, + 0.004027277231216431, + 0.005245981737971306, + 0.004049404989928007, + 0.00500443996861577, + 0.00611599488183856, + 0.0333654060959816, + 0.02653607167303562, + 0.022320952266454697, + 0.013370363041758537, + 0.14501385390758514, + 0.16744716465473175, + 0.016676252707839012, + 0.023632990196347237, + 0.010813252069056034, + 0.010083490051329136, + 0.018093453720211983, + 0.013355747796595097, + 0.01874420791864395, + 0.012889747507870197, + 0.010296327993273735, + 0.01431685034185648, + 0.0008566773030906916, + 0.0010987033601850271, + 0.0011072988854721189, + 0.0011780830100178719, + 0.001593118067830801, + 0.0014441213570535183, + 0.00142858293838799, + 0.0013039123732596636, + 0.0018038321286439896, + 0.00267320335842669, + 0.0028938159812241793, + 0.004304723348468542, + 0.0029294616542756557, + 0.0027726914267987013, + 0.0013237070525065064, + 0.0022645851131528616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.020591532811522484 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048665110021829605 + ], + "opposite_logprob_tokens": [ + -3.0486650466918945 + ], + "logprob_reward": -0.048665110021829605, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048665110021829605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2357216477394104, + 0.05820106714963913, + 0.03640774264931679, + 0.14662685990333557, + 0.11170192807912827, + 0.10068570077419281, + 0.1955747753381729, + 0.033232539892196655, + 0.022064700722694397, + 0.05978301540017128 + ], + "credit_map": [ + 0.022295931354165077, + 0.02659487910568714, + 0.036315277218818665, + 0.026495421305298805, + 0.0390186607837677, + 0.030097026377916336, + 0.03278642147779465, + 0.022118037566542625, + 0.008457271382212639, + 0.0076369112357497215, + 0.011015139520168304, + 0.009277886711061, + 0.0066804345697164536, + 0.005799391772598028, + 0.005400762893259525, + 0.0039332713931798935, + 0.005218853708356619, + 0.006086213514208794, + 0.004427929874509573, + 0.007790729403495789, + 0.0029917315114289522, + 0.0030722147785127163, + 0.004119642078876495, + 0.0027004312723875046, + 0.020160332322120667, + 0.028444573283195496, + 0.012938239611685276, + 0.018950266763567924, + 0.02214353159070015, + 0.014955970458686352, + 0.01786077208817005, + 0.011173181235790253, + 0.009665016084909439, + 0.008728807792067528, + 0.014408279210329056, + 0.01235383003950119, + 0.015614944510161877, + 0.009324342012405396, + 0.029120583087205887, + 0.012486129067838192, + 0.012460102327167988, + 0.00948430597782135, + 0.015687722712755203, + 0.009877149015665054, + 0.020644675940275192, + 0.011274860240519047, + 0.013873145915567875, + 0.0073837400414049625, + 0.01729997992515564, + 0.020802153274416924, + 0.025948762893676758, + 0.01678018644452095, + 0.0303940512239933, + 0.05002807453274727, + 0.01505240611732006, + 0.019269157201051712, + 0.002477337373420596, + 0.00461828988045454, + 0.005107158329337835, + 0.0056223333813250065, + 0.003318051341921091, + 0.004532201215624809, + 0.002906190697103739, + 0.004650981165468693, + 0.001957405125722289, + 0.0024324278347194195, + 0.0021105040796101093, + 0.003360204165801406, + 0.003521813079714775, + 0.003234891453757882, + 0.0021112109534442425, + 0.0033362433314323425, + 0.00574373546987772, + 0.007197644095867872, + 0.012015147134661674, + 0.00847635418176651, + 0.007298577576875687, + 0.007222545798867941, + 0.00641463790088892, + 0.0054143741726875305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.048665110021829605 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017119289841502905 + ], + "opposite_logprob_tokens": [ + -6.376711845397949 + ], + "logprob_reward": -0.0017119289841502905, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017119289841502905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34428870677948, + 0.058571189641952515, + 0.035758670419454575, + 0.11545582860708237, + 0.11537598073482513, + 0.10204705595970154, + 0.09012166410684586, + 0.03902002424001694, + 0.047265369445085526, + 0.052095528692007065 + ], + "credit_map": [ + 0.03194679319858551, + 0.031734805554151535, + 0.04358181357383728, + 0.02784617431461811, + 0.0679149404168129, + 0.04776907339692116, + 0.03395857661962509, + 0.05953652411699295, + 0.009015480056405067, + 0.006552682723850012, + 0.008618682622909546, + 0.006363143678754568, + 0.008578247390687466, + 0.008434860967099667, + 0.004926625173538923, + 0.006081465166062117, + 0.005587409250438213, + 0.0032835800666362047, + 0.004514600615948439, + 0.004323509521782398, + 0.004400868434458971, + 0.004466509446501732, + 0.004193922504782677, + 0.004988271743059158, + 0.028571292757987976, + 0.011380335316061974, + 0.009383241645991802, + 0.007782943081110716, + 0.03157755360007286, + 0.014017482288181782, + 0.0064291805028915405, + 0.006313793361186981, + 0.010019841603934765, + 0.00880441814661026, + 0.01640540175139904, + 0.013254552148282528, + 0.015839410945773125, + 0.015113427303731441, + 0.019345862790942192, + 0.016593072563409805, + 0.018827339634299278, + 0.0072029042057693005, + 0.012810801155865192, + 0.0075661358423531055, + 0.014732904732227325, + 0.018105629831552505, + 0.011595213785767555, + 0.011206124909222126, + 0.009262359701097012, + 0.008714639581739902, + 0.010448182933032513, + 0.007864374667406082, + 0.013714154250919819, + 0.025889310985803604, + 0.006141498684883118, + 0.008087147027254105, + 0.0035463091917335987, + 0.00431916443631053, + 0.005816509947180748, + 0.004740588366985321, + 0.006738440599292517, + 0.005963871721178293, + 0.002793398452922702, + 0.005101743154227734, + 0.0037288980092853308, + 0.005321769043803215, + 0.008566833101212978, + 0.0057435245253145695, + 0.005786283407360315, + 0.007313807029277086, + 0.004256690386682749, + 0.0065475632436573505, + 0.004260021261870861, + 0.006588355638086796, + 0.0056675211526453495, + 0.008299951441586018, + 0.00939080398529768, + 0.008057385683059692, + 0.00496262963861227, + 0.0048688556998968124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.0017119289841502905 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001148164272308 + ], + "opposite_logprob_tokens": [ + -2.6950113773345947 + ], + "logprob_reward": -0.07001148164272308, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001148164272308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0844060406088829, + 0.058661263436079025, + 0.036705877631902695, + 0.07448159158229828, + 0.0803060382604599, + 0.059354472905397415, + 0.4629215598106384, + 0.04691212996840477, + 0.05672667548060417, + 0.039524372667074203 + ], + "credit_map": [ + 0.010174840688705444, + 0.011316441930830479, + 0.010890521109104156, + 0.011698251590132713, + 0.01130298338830471, + 0.010115660727024078, + 0.007848141714930534, + 0.011059202253818512, + 0.008298661559820175, + 0.005986896343529224, + 0.008039034903049469, + 0.008634069934487343, + 0.009269692935049534, + 0.0061713107861578465, + 0.005206941161304712, + 0.007054654881358147, + 0.004758726805448532, + 0.004036980681121349, + 0.0038091125898063183, + 0.004748302977532148, + 0.005329751409590244, + 0.004937971942126751, + 0.005048006772994995, + 0.004037021659314632, + 0.009006998501718044, + 0.012450475245714188, + 0.007267926353961229, + 0.010682008229196072, + 0.008411413058638573, + 0.007966539822518826, + 0.011519757099449635, + 0.007176476065069437, + 0.00894896686077118, + 0.005676026456058025, + 0.010620977729558945, + 0.0075151799246668816, + 0.01101785060018301, + 0.00862113106995821, + 0.01675734668970108, + 0.011148565448820591, + 0.006475332658737898, + 0.007215641904622316, + 0.007779422216117382, + 0.006922577973455191, + 0.012612143531441689, + 0.007499664556235075, + 0.007052190136164427, + 0.0037974982988089323, + 0.03892846032977104, + 0.025855934247374535, + 0.025224070996046066, + 0.008690898306667805, + 0.15238572657108307, + 0.18210673332214355, + 0.010947519913315773, + 0.018782228231430054, + 0.003685161005705595, + 0.004534129984676838, + 0.008738433010876179, + 0.006567978300154209, + 0.0048706503584980965, + 0.005511322058737278, + 0.006719859316945076, + 0.006284597795456648, + 0.005174926482141018, + 0.005207027308642864, + 0.01138212252408266, + 0.007021105848252773, + 0.006207636091858149, + 0.010544613935053349, + 0.005145280621945858, + 0.006043958477675915, + 0.003401267109438777, + 0.004978530574589968, + 0.008031981065869331, + 0.006474989466369152, + 0.003592599183320999, + 0.0047438908368349075, + 0.004197434056550264, + 0.004103681072592735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.07001148164272308 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle, extreme wide shot captures a breathtaking panorama rendered in a beautiful watercolor animation style. A lone high school girl, resembling Ichika Miyamizu with her short, dark brown bob hairstyle, stands in quiet contemplation on a windswept hilltop amidst the grass. Clad in her dark school uniform, she forms a delicate silhouette against the magnificent celestial backdrop. Above her, the aurora borealis dances across the vast, star-studded night sky, forming a vibrant, shimmering ribbon of green and purple. These celestial lights gracefully ripple and swirl, their brilliant colors softly seeping into the sky's deep indigo canvas, creating diffused painterly edges, as if wet paint were spreading across textured paper. The otherworldly glow of the aurora casts a cool, magical glow on the mountaintop and over the girl's shoulders, creating a profound atmosphere of quiet wonder and solitary awe. A gentle, cool breeze gently ruffles her hair and the hem of her skirt, the only subtle movement on the ground, contrasting sharply with the dynamic, silent scene unfolding in the sky.", + "seed": 1409, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_4/index_5350_seed_1409_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2739280310459435, + "vlm_logprob": -0.2739280310459435, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.2739280310459435, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2305350303649902 + ], + "opposite_logprob_tokens": [ + -0.35553503036499023 + ], + "logprob_reward": -1.2305350303649902, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2305350303649902, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.392964243888855, + 0.057957567274570465, + 0.202463760972023, + 0.02642957493662834, + 0.03494853898882866, + 0.10241533070802689, + 0.04168599843978882, + 0.05081852898001671, + 0.05784565210342407, + 0.03247085586190224 + ], + "credit_map": [ + 0.04392518475651741, + 0.13103587925434113, + 0.0247498732060194, + 0.02688770927488804, + 0.023774873465299606, + 0.10353163629770279, + 0.019655561074614525, + 0.019403476268053055, + 0.007085883524268866, + 0.01580621488392353, + 0.010176758281886578, + 0.006386183202266693, + 0.0035812342539429665, + 0.00542167853564024, + 0.0041860309429466724, + 0.005313574336469173, + 0.01990954950451851, + 0.0444440059363842, + 0.07097181677818298, + 0.025272853672504425, + 0.009894339367747307, + 0.01075792033225298, + 0.012142629362642765, + 0.009070634841918945, + 0.0030467163305729628, + 0.0042588659562170506, + 0.003139818087220192, + 0.002644972177222371, + 0.002301903208717704, + 0.0054274024441838264, + 0.0030852400232106447, + 0.0025246525183320045, + 0.00861926469951868, + 0.004868008196353912, + 0.002581088338047266, + 0.003053846303373575, + 0.0021890890784561634, + 0.006891659926623106, + 0.003577370895072818, + 0.0031682054977864027, + 0.02240375615656376, + 0.014327827841043472, + 0.009037363342940807, + 0.010746747255325317, + 0.009410141035914421, + 0.01403977070003748, + 0.014339970424771309, + 0.008109742775559425, + 0.006808917503803968, + 0.008624697104096413, + 0.0059792897664010525, + 0.004623175598680973, + 0.003936902154237032, + 0.0038940217345952988, + 0.005266853142529726, + 0.0025521349161863327, + 0.01395621057599783, + 0.006692417431622744, + 0.005765341222286224, + 0.00558637036010623, + 0.003947271499782801, + 0.004739520605653524, + 0.006002889014780521, + 0.004128503613173962, + 0.004356744233518839, + 0.0031063230708241463, + 0.017537394538521767, + 0.006835483014583588, + 0.002740498399361968, + 0.017802560701966286, + 0.002817674772813916, + 0.0026489642914384604, + 0.003005168167874217, + 0.008008758537471294, + 0.0029522692784667015, + 0.003225309308618307, + 0.0035853111185133457, + 0.003801740240305662, + 0.005398749839514494, + 0.0024935463443398476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -1.2305350303649902 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038108889013528824 + ], + "opposite_logprob_tokens": [ + -3.2881088256835938 + ], + "logprob_reward": -0.038108889013528824, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038108889013528824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1546664834022522, + 0.06231096386909485, + 0.2288345843553543, + 0.06187843531370163, + 0.052679914981126785, + 0.14841137826442719, + 0.059327367693185806, + 0.15270842611789703, + 0.05941842868924141, + 0.019763996824622154 + ], + "credit_map": [ + 0.013982360251247883, + 0.03704741969704628, + 0.01687699556350708, + 0.018550055101513863, + 0.010883359238505363, + 0.027764417231082916, + 0.016517169773578644, + 0.013044709339737892, + 0.011178756132721901, + 0.016473622992634773, + 0.005988874938338995, + 0.006261056754738092, + 0.00506650609895587, + 0.006005673203617334, + 0.005906872451305389, + 0.005429601762443781, + 0.02617548033595085, + 0.08744235336780548, + 0.030375121161341667, + 0.015595771372318268, + 0.012931551784276962, + 0.036906275898218155, + 0.010749413631856441, + 0.00865860003978014, + 0.005417402368038893, + 0.006389870308339596, + 0.012893774546682835, + 0.0071345181204378605, + 0.004170663189142942, + 0.01339589525014162, + 0.005359935108572245, + 0.007116374559700489, + 0.013644885271787643, + 0.00761824194341898, + 0.006745904218405485, + 0.005283094011247158, + 0.0032866126857697964, + 0.008255411870777607, + 0.0034300272818654776, + 0.00441573653370142, + 0.026807095855474472, + 0.03892600163817406, + 0.012198277749121189, + 0.012887257151305676, + 0.01287284679710865, + 0.0203759353607893, + 0.012068345211446285, + 0.01227562502026558, + 0.008036410436034203, + 0.011090798303484917, + 0.007997922599315643, + 0.009456060826778412, + 0.005927744787186384, + 0.004966918844729662, + 0.007093654479831457, + 0.004757857881486416, + 0.027835708111524582, + 0.017378808930516243, + 0.014432180672883987, + 0.016291821375489235, + 0.018132269382476807, + 0.0145554319024086, + 0.03446361795067787, + 0.009618578478693962, + 0.004139524884521961, + 0.0036927119363099337, + 0.019146880134940147, + 0.008297602646052837, + 0.0025489437393844128, + 0.016518032178282738, + 0.0026964719872921705, + 0.0023782618809491396, + 0.0016006360528990626, + 0.004466892685741186, + 0.0017681020544841886, + 0.0016513089649379253, + 0.0023571904748678207, + 0.002266167663037777, + 0.0032151970081031322, + 0.0024385021533817053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.038108889013528824 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205175817012787 + ], + "opposite_logprob_tokens": [ + -2.812051773071289 + ], + "logprob_reward": -0.06205175817012787, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205175817012787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21331262588500977, + 0.05333033576607704, + 0.2553707957267761, + 0.03644074872136116, + 0.1289091557264328, + 0.0990038812160492, + 0.07899603247642517, + 0.0650370866060257, + 0.0427180677652359, + 0.026881378144025803 + ], + "credit_map": [ + 0.024440336972475052, + 0.06837063282728195, + 0.03461100161075592, + 0.018580392003059387, + 0.012862810865044594, + 0.03196926414966583, + 0.011370663531124592, + 0.011107470840215683, + 0.006709192879498005, + 0.011955943889915943, + 0.006635802797973156, + 0.007006662432104349, + 0.003978849388659, + 0.0051017217338085175, + 0.006195782218128443, + 0.00574636971578002, + 0.01818864606320858, + 0.08560929447412491, + 0.08527812361717224, + 0.0245925635099411, + 0.012747772969305515, + 0.012864438816905022, + 0.007897621020674706, + 0.00819228496402502, + 0.004251059144735336, + 0.004532774444669485, + 0.003890498774126172, + 0.004433115478605032, + 0.0044252765364944935, + 0.0032588730100542307, + 0.006591266021132469, + 0.005057881120592356, + 0.04279931262135506, + 0.008531974628567696, + 0.006476412061601877, + 0.006921944674104452, + 0.0061052883975207806, + 0.042909007519483566, + 0.010261633433401585, + 0.004903570748865604, + 0.014048817567527294, + 0.014407308772206306, + 0.013328738510608673, + 0.019540967419743538, + 0.007985338568687439, + 0.008905330672860146, + 0.012577909976243973, + 0.008209452964365482, + 0.013620935380458832, + 0.00920187309384346, + 0.009109877049922943, + 0.009668534621596336, + 0.00781454611569643, + 0.008309832774102688, + 0.015994492918252945, + 0.0052759237587451935, + 0.009610454551875591, + 0.007492819335311651, + 0.0064714751206338406, + 0.007613170426338911, + 0.007690254133194685, + 0.0070381429977715015, + 0.01421264000236988, + 0.00490811700001359, + 0.002885773777961731, + 0.0030546311754733324, + 0.01177141536027193, + 0.005468106362968683, + 0.0023059509694576263, + 0.010436923243105412, + 0.004098572768270969, + 0.00269668223336339, + 0.003062336938455701, + 0.0038807697128504515, + 0.002555038081482053, + 0.002431524218991399, + 0.0030337220523506403, + 0.002338235732167959, + 0.00577788008376956, + 0.003801864804700017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.06205175817012787 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248221103101969 + ], + "opposite_logprob_tokens": [ + -5.255248069763184 + ], + "logprob_reward": -0.005248221103101969, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248221103101969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4420197606086731, + 0.10274757444858551, + 0.10927202552556992, + 0.02509259060025215, + 0.0474538616836071, + 0.13163743913173676, + 0.042265865951776505, + 0.037960235029459, + 0.014329636469483376, + 0.04722098261117935 + ], + "credit_map": [ + 0.04509780555963516, + 0.14334790408611298, + 0.024597546085715294, + 0.02649075910449028, + 0.02726942114531994, + 0.13680744171142578, + 0.0187847800552845, + 0.019624115899205208, + 0.013122082687914371, + 0.018449854105710983, + 0.012847085483372211, + 0.013892597518861294, + 0.0067370859906077385, + 0.013390063308179379, + 0.009735255502164364, + 0.014573548920452595, + 0.010443612933158875, + 0.027154557406902313, + 0.038963742554187775, + 0.012927836738526821, + 0.004843368194997311, + 0.005130543373525143, + 0.004910991992801428, + 0.004897371865808964, + 0.0026398850604891777, + 0.0032594387885183096, + 0.0029252199456095695, + 0.003243745770305395, + 0.003133009420707822, + 0.0022585156839340925, + 0.0038286312483251095, + 0.0038041421212255955, + 0.014407834969460964, + 0.004434584174305201, + 0.0030275059398263693, + 0.002370088826864958, + 0.002195865148678422, + 0.015520205721259117, + 0.0035147403832525015, + 0.0019830362871289253, + 0.02365800365805626, + 0.025121761485934258, + 0.013146317563951015, + 0.017958806827664375, + 0.008863354101777077, + 0.016682879999279976, + 0.011681322008371353, + 0.014524992555379868, + 0.007749785669147968, + 0.008539420552551746, + 0.0038454155437648296, + 0.0037411649245768785, + 0.0034578419290483, + 0.005362183321267366, + 0.006933670025318861, + 0.002636383520439267, + 0.0061004976741969585, + 0.003923911601305008, + 0.0039995573461055756, + 0.004729383625090122, + 0.0043139527551829815, + 0.003592507680878043, + 0.008554530330002308, + 0.0027458961121737957, + 0.0012877446133643389, + 0.0022324975579977036, + 0.0029576080851256847, + 0.0019100649515166879, + 0.0012551432009786367, + 0.001309820101596415, + 0.0015260990476235747, + 0.0018506592605262995, + 0.00439969077706337, + 0.015343363396823406, + 0.0035268706269562244, + 0.002480399562045932, + 0.003415964310988784, + 0.012225058861076832, + 0.002947943052276969, + 0.002881692722439766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.005248221103101969 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as an extreme wide shot capturing a panoramic view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a watercolor animation with diffused painterly edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a high school girl with short dark brown bob hair standing on a windswept hilltop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a night sky featuring the aurora borealis in green and purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there visible movement of the girl's hair and skirt due to a gentle breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336962565779686 + ], + "opposite_logprob_tokens": [ + -3.408696174621582 + ], + "logprob_reward": -0.0336962565779686, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336962565779686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23448732495307922, + 0.07711608707904816, + 0.12767569720745087, + 0.05828854441642761, + 0.09194374084472656, + 0.07050015777349472, + 0.1475372165441513, + 0.06363412737846375, + 0.035893186926841736, + 0.09292392432689667 + ], + "credit_map": [ + 0.022278837859630585, + 0.07014177739620209, + 0.01562885008752346, + 0.01813218556344509, + 0.013413849286735058, + 0.07252980023622513, + 0.010180904529988766, + 0.012181132100522518, + 0.011928902938961983, + 0.018933884799480438, + 0.007421334274113178, + 0.008332710713148117, + 0.005324722733348608, + 0.00917828269302845, + 0.008035733364522457, + 0.007960520684719086, + 0.013809840194880962, + 0.04613926261663437, + 0.016683004796504974, + 0.007308743428438902, + 0.008149995468556881, + 0.02124672755599022, + 0.008657175116240978, + 0.005680937319993973, + 0.005335004068911076, + 0.0057151042856276035, + 0.014409077353775501, + 0.007060114294290543, + 0.0036043785512447357, + 0.014139068312942982, + 0.004141426645219326, + 0.003884370904415846, + 0.01793481968343258, + 0.007241346407681704, + 0.012789069674909115, + 0.006953904405236244, + 0.0053426725789904594, + 0.027746785432100296, + 0.009506715461611748, + 0.004428428132086992, + 0.01255783997476101, + 0.017177658155560493, + 0.00548507459461689, + 0.006366495508700609, + 0.005119444336742163, + 0.011371939443051815, + 0.005777348764240742, + 0.006644357927143574, + 0.03430477902293205, + 0.014955722726881504, + 0.014862379059195518, + 0.01646917127072811, + 0.013180412352085114, + 0.01835527829825878, + 0.026691973209381104, + 0.008717494085431099, + 0.010114885866641998, + 0.00831661093980074, + 0.006855289917439222, + 0.007606452330946922, + 0.007061080541461706, + 0.006560811772942543, + 0.011451056227087975, + 0.005667936056852341, + 0.004903891123831272, + 0.0053423671051859856, + 0.006086120381951332, + 0.004431008826941252, + 0.003195033874362707, + 0.004155359696596861, + 0.004153471905738115, + 0.0036259335465729237, + 0.008850662037730217, + 0.025159208104014397, + 0.006809129845350981, + 0.006932571530342102, + 0.007977628149092197, + 0.014183475635945797, + 0.013863099738955498, + 0.009148146025836468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.0336962565779686 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_5/index_tensor([1245, 1245])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_5/index_tensor([1245, 1245])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..243db7dbccd3bc88911e019d87fdc1f805975b69 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_5/index_tensor([1245, 1245])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful, low-angle, sweeping shot plunges us into a world of powerful, hand-drawn sketches constructed entirely of charcoal and graphite. Amidst this abstract waterscape of wild brushstrokes and smudged ink, two butterfly carp, each with enormous, silky fins, engage in a fierce duel. One is pitch black, its scales outlined with fine crosshatching; the other is pure white, its bright areas highlighted by bold white space, creating a stark visual contrast. The camera follows them in a tense, rotating pattern, capturing every moment as they chase, spiral, and collide. Their vast, flowing fins trace graceful yet deadly arcs with each swing, and every violent collision stirs up splashes of ink and broken lines across the frame, as if their struggle is composing the surrounding environment in real time. The high-contrast light and shadow effects are extremely emphasized through heavy shadows and sharp white spaces. There is not a trace of superfluous color in the entire scene, only the pure confrontation of black, white and gray, creating a dynamic beauty that is both tense and magnificent as well as full of primitive explosive power.", + "seed": 5035, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_5/index_1245_seed_5035_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.019225051067769526, + "vlm_logprob": -0.019225051067769526, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.019225051067769526, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017034059390425682 + ], + "opposite_logprob_tokens": [ + -4.26703405380249 + ], + "logprob_reward": -0.017034059390425682, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017034059390425682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2733818292617798, + 0.11180015653371811, + 0.04059547558426857, + 0.03858739882707596, + 0.04613104090094566, + 0.03091953694820404, + 0.025865819305181503, + 0.14732179045677185, + 0.18338127434253693, + 0.10201564431190491 + ], + "credit_map": [ + 0.04393008351325989, + 0.020147409290075302, + 0.028183942660689354, + 0.02038189396262169, + 0.02219097688794136, + 0.042202387005090714, + 0.035544343292713165, + 0.060800809413194656, + 0.012083323672413826, + 0.013759959489107132, + 0.00825747475028038, + 0.022156154736876488, + 0.012508676387369633, + 0.00950144324451685, + 0.016603929921984673, + 0.01692919060587883, + 0.004104729276150465, + 0.004758457187563181, + 0.004641946870833635, + 0.006392397917807102, + 0.003910517320036888, + 0.005138086620718241, + 0.005915909074246883, + 0.0057334317825734615, + 0.003292747074738145, + 0.004583778791129589, + 0.0035977426450699568, + 0.0066951969638466835, + 0.005018122959882021, + 0.004100552294403315, + 0.004060419276356697, + 0.007238839752972126, + 0.008548025041818619, + 0.0057899607345461845, + 0.008248706348240376, + 0.004913091193884611, + 0.0038639805279672146, + 0.0043506259098649025, + 0.0064077056013047695, + 0.004008943680673838, + 0.0038318682927638292, + 0.004631525836884975, + 0.003691313788294792, + 0.004049698822200298, + 0.0032721590250730515, + 0.002624582266435027, + 0.004160931333899498, + 0.004657457582652569, + 0.0037471482064574957, + 0.002077896613627672, + 0.0029356181621551514, + 0.003953989129513502, + 0.0022657529916614294, + 0.003211516886949539, + 0.003339879447594285, + 0.0043340204283595085, + 0.013401167467236519, + 0.011560237966477871, + 0.014071560464799404, + 0.02146785333752632, + 0.010900452733039856, + 0.04566971957683563, + 0.011289243586361408, + 0.01896156184375286, + 0.010354802012443542, + 0.011379586532711983, + 0.012648769654333591, + 0.009847060777246952, + 0.011112879030406475, + 0.06159428879618645, + 0.05149919539690018, + 0.014944685623049736, + 0.007524107117205858, + 0.010630713775753975, + 0.009083058685064316, + 0.014808280393481255, + 0.01199339423328638, + 0.01817331835627556, + 0.014844473451375961, + 0.014958303421735764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.017034059390425682 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007646813057363033 + ], + "opposite_logprob_tokens": [ + -4.8826470375061035 + ], + "logprob_reward": -0.007646813057363033, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007646813057363033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10251953452825546, + 0.10734868794679642, + 0.1984352320432663, + 0.0842929482460022, + 0.06909773498773575, + 0.03687426820397377, + 0.02118910849094391, + 0.06070193275809288, + 0.0845983549952507, + 0.23494213819503784 + ], + "credit_map": [ + 0.02052980475127697, + 0.007631797809153795, + 0.013254278339445591, + 0.009844661690294743, + 0.010514222085475922, + 0.018261363729834557, + 0.01178409717977047, + 0.010699312202632427, + 0.006801014766097069, + 0.007791641168296337, + 0.007675016764551401, + 0.03372467681765556, + 0.006908765994012356, + 0.010615872219204903, + 0.02036161534488201, + 0.013470076024532318, + 0.008124127984046936, + 0.007340734824538231, + 0.006699822377413511, + 0.06720151752233505, + 0.0064959595911204815, + 0.005417059175670147, + 0.08072778582572937, + 0.01642824150621891, + 0.007008717395365238, + 0.010318572632968426, + 0.010204596444964409, + 0.01363166980445385, + 0.00747127877548337, + 0.010013965889811516, + 0.009519070386886597, + 0.01612507365643978, + 0.005077051930129528, + 0.008240224793553352, + 0.008980844169855118, + 0.012171448208391666, + 0.01129644364118576, + 0.007823852822184563, + 0.008864082396030426, + 0.006643784232437611, + 0.005813966505229473, + 0.006852677557617426, + 0.004025094676762819, + 0.004733681678771973, + 0.0045896489173173904, + 0.002702650846913457, + 0.0031832559034228325, + 0.004973291885107756, + 0.0031268708407878876, + 0.00198345142416656, + 0.0023424539249390364, + 0.003083721734583378, + 0.002140671480447054, + 0.001732182689011097, + 0.0027647868264466524, + 0.004014968406409025, + 0.005646276753395796, + 0.007611386943608522, + 0.005615184549242258, + 0.006489903666079044, + 0.006439410150051117, + 0.00740255881100893, + 0.011785478331148624, + 0.009711739607155323, + 0.00461962167173624, + 0.005557528231292963, + 0.00589409563690424, + 0.003793170442804694, + 0.004030620213598013, + 0.02683274820446968, + 0.026871955022215843, + 0.006998616270720959, + 0.020255550742149353, + 0.023854542523622513, + 0.022694449871778488, + 0.03770935535430908, + 0.029977701604366302, + 0.03637552261352539, + 0.028519906103610992, + 0.03555510938167572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.007646813057363033 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067780554294586 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.016067780554294586, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067780554294586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10621262341737747, + 0.13803213834762573, + 0.165727436542511, + 0.05168578028678894, + 0.04447658732533455, + 0.025431813672184944, + 0.06519819796085358, + 0.10506317764520645, + 0.18324311077594757, + 0.11492916941642761 + ], + "credit_map": [ + 0.017266983166337013, + 0.00673804571852088, + 0.01193286757916212, + 0.0057977731339633465, + 0.00811050646007061, + 0.0221722349524498, + 0.013606742955744267, + 0.020587464794516563, + 0.01395443920046091, + 0.012369535863399506, + 0.010155095718801022, + 0.025170616805553436, + 0.012795851565897465, + 0.016229720786213875, + 0.02011907286942005, + 0.02723781205713749, + 0.006705268286168575, + 0.007813149131834507, + 0.007655072491616011, + 0.04673228785395622, + 0.007924431003630161, + 0.006774301175028086, + 0.06344126909971237, + 0.018681660294532776, + 0.0042960611172020435, + 0.0074753607623279095, + 0.0031151054427027702, + 0.005853873677551746, + 0.004913111682981253, + 0.005700874608010054, + 0.010616124607622623, + 0.009715266525745392, + 0.007332795765250921, + 0.004615469370037317, + 0.008258714340627193, + 0.005950321443378925, + 0.0028782628942281008, + 0.004231196828186512, + 0.0073781865648925304, + 0.0038316380232572556, + 0.0038324648048728704, + 0.005285714752972126, + 0.0027285730466246605, + 0.0033685241360217333, + 0.002889904659241438, + 0.001755388337187469, + 0.0025299990084022284, + 0.0030412429478019476, + 0.011823920533061028, + 0.005094862077385187, + 0.0081477714702487, + 0.007167333737015724, + 0.007123981136828661, + 0.007799108512699604, + 0.0062432074919342995, + 0.011798013001680374, + 0.008851218968629837, + 0.007714330684393644, + 0.010653845965862274, + 0.004576023668050766, + 0.008166477084159851, + 0.02327539585530758, + 0.006963165942579508, + 0.03486271947622299, + 0.011433577165007591, + 0.009468808770179749, + 0.0133203798905015, + 0.01046296302229166, + 0.007964332588016987, + 0.04991203173995018, + 0.06091546639800072, + 0.019765552133321762, + 0.010174475610256195, + 0.012376135215163231, + 0.011516474187374115, + 0.020892580971121788, + 0.014335408806800842, + 0.014753272756934166, + 0.014133911579847336, + 0.01674690656363964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.016067780554294586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732044741511345 + ], + "opposite_logprob_tokens": [ + -5.006731986999512 + ], + "logprob_reward": -0.006732044741511345, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732044741511345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1516474038362503, + 0.0816643163561821, + 0.14942286908626556, + 0.05250861868262291, + 0.05520173907279968, + 0.023924576118588448, + 0.027141885831952095, + 0.14008697867393494, + 0.24729043245315552, + 0.07111120969057083 + ], + "credit_map": [ + 0.02782352827489376, + 0.00908657442778349, + 0.01900533027946949, + 0.01508223544806242, + 0.011961142532527447, + 0.021779995411634445, + 0.017252320423722267, + 0.029656264930963516, + 0.007462676148861647, + 0.00631926953792572, + 0.007229591254144907, + 0.018977077677845955, + 0.00775489816442132, + 0.008330805227160454, + 0.012437238357961178, + 0.013152755796909332, + 0.008755148388445377, + 0.012475715950131416, + 0.01314542070031166, + 0.02205301821231842, + 0.014243227429687977, + 0.011463167145848274, + 0.04019198566675186, + 0.027095183730125427, + 0.004986800719052553, + 0.006334239616990089, + 0.004246864467859268, + 0.004902264568954706, + 0.00433658342808485, + 0.006160814315080643, + 0.011732100509107113, + 0.009808949194848537, + 0.008417525328695774, + 0.0057476419024169445, + 0.008998420089483261, + 0.009083976037800312, + 0.005682535003870726, + 0.0042170509696006775, + 0.009547345340251923, + 0.003507243702188134, + 0.004119363613426685, + 0.0028394286055117846, + 0.002933615120127797, + 0.0031141869258135557, + 0.002402056008577347, + 0.0017311375122517347, + 0.003680654801428318, + 0.0031041321344673634, + 0.0043741329573094845, + 0.0023894207552075386, + 0.003244561143219471, + 0.0030528209172189236, + 0.0024418351240456104, + 0.002998492680490017, + 0.0033929303754121065, + 0.005247690714895725, + 0.010232040658593178, + 0.009757274761795998, + 0.011975512839853764, + 0.020721865817904472, + 0.010839303024113178, + 0.040880728513002396, + 0.013144842348992825, + 0.0225354116410017, + 0.013713992200791836, + 0.010152380913496017, + 0.015455673448741436, + 0.010964788496494293, + 0.012381681241095066, + 0.08044923096895218, + 0.08333239704370499, + 0.020840290933847427, + 0.005506210960447788, + 0.00779198482632637, + 0.00810409989207983, + 0.012686085887253284, + 0.00722455233335495, + 0.009062723256647587, + 0.008123255334794521, + 0.012612298130989075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.006732044741511345 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864455759525299 + ], + "opposite_logprob_tokens": [ + -3.048644542694092 + ], + "logprob_reward": -0.04864455759525299, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864455759525299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09385508298873901, + 0.08198666572570801, + 0.1551460176706314, + 0.07504266500473022, + 0.06181581690907478, + 0.04109831899404526, + 0.024999547749757767, + 0.11768566817045212, + 0.16201519966125488, + 0.18635500967502594 + ], + "credit_map": [ + 0.01864633522927761, + 0.0073179262690246105, + 0.012270822189748287, + 0.008401521481573582, + 0.00724153034389019, + 0.014475924894213676, + 0.010335714556276798, + 0.015165310353040695, + 0.005920149851590395, + 0.005095831584185362, + 0.006173386704176664, + 0.023673677816987038, + 0.007568794302642345, + 0.007889773696660995, + 0.013795102946460247, + 0.011869954876601696, + 0.012507219798862934, + 0.019015630707144737, + 0.014376673847436905, + 0.022200442850589752, + 0.014521542005240917, + 0.013389047235250473, + 0.025868168100714684, + 0.03326728940010071, + 0.008674515411257744, + 0.011884900741279125, + 0.007366984151303768, + 0.008376501500606537, + 0.007165458053350449, + 0.008613084442913532, + 0.009900229051709175, + 0.01306099258363247, + 0.007353898137807846, + 0.007873406633734703, + 0.010149236768484116, + 0.011122718453407288, + 0.00593460351228714, + 0.004657573066651821, + 0.009512403048574924, + 0.005211979150772095, + 0.006231706123799086, + 0.0047066849656403065, + 0.004705557134002447, + 0.006363667082041502, + 0.005261643324047327, + 0.0033051304053515196, + 0.005815570242702961, + 0.004708361346274614, + 0.003465337213128805, + 0.002326525514945388, + 0.0029802382923662663, + 0.0029949003364890814, + 0.002989259548485279, + 0.0026186699979007244, + 0.004831988830119371, + 0.002792628016322851, + 0.008704562671482563, + 0.007655577268451452, + 0.010497807525098324, + 0.007251132279634476, + 0.0086089838296175, + 0.03323376178741455, + 0.005910398438572884, + 0.03582344949245453, + 0.010019716806709766, + 0.007029014639556408, + 0.009437955915927887, + 0.007457049563527107, + 0.008303669281303883, + 0.05039319023489952, + 0.05590704828500748, + 0.013467561453580856, + 0.015353904105722904, + 0.018500979989767075, + 0.019805926829576492, + 0.03361188620328903, + 0.019015435129404068, + 0.026208113878965378, + 0.022930193692445755, + 0.03092857263982296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.04864455759525299 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, low-angle, sweeping shot plunges us into a world of powerful, hand-drawn sketches constructed entirely of charcoal and graphite. Amidst this abstract waterscape of wild brushstrokes and smudged ink, two butterfly carp, each with enormous, silky fins, engage in a fierce duel. One is pitch black, its scales outlined with fine crosshatching; the other is pure white, its bright areas highlighted by bold white space, creating a stark visual contrast. The camera follows them in a tense, rotating pattern, capturing every moment as they chase, spiral, and collide. Their vast, flowing fins trace graceful yet deadly arcs with each swing, and every violent collision stirs up splashes of ink and broken lines across the frame, as if their struggle is composing the surrounding environment in real time. The high-contrast light and shadow effects are extremely emphasized through heavy shadows and sharp white spaces. There is not a trace of superfluous color in the entire scene, only the pure confrontation of black, white and gray, creating a dynamic beauty that is both tense and magnificent as well as full of primitive explosive power.", + "seed": 5036, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_5/index_1245_seed_5036_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.013087409175932407, + "vlm_logprob": -0.013087409175932407, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.013087409175932407, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015744425356388092 + ], + "opposite_logprob_tokens": [ + -4.390744209289551 + ], + "logprob_reward": -0.015744425356388092, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.015744425356388092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22586214542388916, + 0.061633821576833725, + 0.06320592015981674, + 0.2524748146533966, + 0.08147072792053223, + 0.049480702728033066, + 0.06557872891426086, + 0.026564612984657288, + 0.09992991387844086, + 0.07379866391420364 + ], + "credit_map": [ + 0.03146745637059212, + 0.018795166164636612, + 0.028158625587821007, + 0.02009117230772972, + 0.0274068433791399, + 0.025469351559877396, + 0.021485907956957817, + 0.05298760533332825, + 0.0047005838714540005, + 0.009018254466354847, + 0.007730078883469105, + 0.006659131031483412, + 0.00663318345323205, + 0.007139537483453751, + 0.009781299158930779, + 0.009971747174859047, + 0.007318201474845409, + 0.0045606764033436775, + 0.007756669539958239, + 0.009653775952756405, + 0.006830384489148855, + 0.009682347998023033, + 0.006279055029153824, + 0.011124800890684128, + 0.019779641181230545, + 0.012881340458989143, + 0.015675069764256477, + 0.05942459776997566, + 0.01606821082532406, + 0.02467941865324974, + 0.03227723017334938, + 0.07168927043676376, + 0.010285530239343643, + 0.005733320955187082, + 0.014314062893390656, + 0.009301253594458103, + 0.007329883985221386, + 0.010873310267925262, + 0.011736561544239521, + 0.011896793730556965, + 0.005251973867416382, + 0.005040345247834921, + 0.004922622814774513, + 0.006565649528056383, + 0.005710817873477936, + 0.005287733394652605, + 0.007250390946865082, + 0.00945116113871336, + 0.007420228328555822, + 0.00694494042545557, + 0.00697770481929183, + 0.008634594269096851, + 0.008237014524638653, + 0.009797821752727032, + 0.008786742575466633, + 0.008779670111835003, + 0.003103540977463126, + 0.0029906549025326967, + 0.0039639947935938835, + 0.002829519798979163, + 0.0035404381342232227, + 0.0037013378459960222, + 0.0026093206834048033, + 0.0038258039858192205, + 0.008973812684416771, + 0.0152699314057827, + 0.01012533437460661, + 0.011601713486015797, + 0.013942182995378971, + 0.009465212933719158, + 0.017440348863601685, + 0.013111366890370846, + 0.009598989970982075, + 0.007114322856068611, + 0.006921596359461546, + 0.005617128219455481, + 0.012721695005893707, + 0.01009376160800457, + 0.011166589334607124, + 0.010564563795924187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.015744425356388092 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007647404447197914 + ], + "opposite_logprob_tokens": [ + -4.882647514343262 + ], + "logprob_reward": -0.007647404447197914, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007647404447197914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07532958686351776, + 0.046749383211135864, + 0.04845220595598221, + 0.21679610013961792, + 0.4324931204319, + 0.023798923939466476, + 0.03803681582212448, + 0.02823876030743122, + 0.03562799096107483, + 0.05447717756032944 + ], + "credit_map": [ + 0.010377591475844383, + 0.006468957290053368, + 0.010545141994953156, + 0.007239312399178743, + 0.009997365064918995, + 0.008239064365625381, + 0.006336242891848087, + 0.016125908121466637, + 0.004889327567070723, + 0.007836206816136837, + 0.0045836749486625195, + 0.00494037102907896, + 0.004489596467465162, + 0.005655448418110609, + 0.005304830148816109, + 0.009049925021827221, + 0.006394551135599613, + 0.005134543403983116, + 0.005384897347539663, + 0.00539170578122139, + 0.0060059912502765656, + 0.007065492682158947, + 0.0071135093457996845, + 0.005961512215435505, + 0.01608428545296192, + 0.01279729325324297, + 0.0163673534989357, + 0.044025320559740067, + 0.014851673506200314, + 0.021092727780342102, + 0.035675130784511566, + 0.05590229853987694, + 0.025575675070285797, + 0.019361821934580803, + 0.04458967596292496, + 0.12832611799240112, + 0.04455148056149483, + 0.11371063441038132, + 0.04192376509308815, + 0.01445392332971096, + 0.0036188545636832714, + 0.0026759079191833735, + 0.00217603356577456, + 0.002953465562313795, + 0.0020778293255716562, + 0.0031596433836966753, + 0.0026346854865550995, + 0.004502500407397747, + 0.004059273283928633, + 0.004127194173634052, + 0.00434008426964283, + 0.00816552247852087, + 0.003839672775939107, + 0.005174198187887669, + 0.003970799967646599, + 0.0043600695207715034, + 0.0033399255480617285, + 0.0027460234705358744, + 0.006173565052449703, + 0.00313887232914567, + 0.002991385292261839, + 0.003530813381075859, + 0.0027832069899886847, + 0.0035349661484360695, + 0.0036644102074205875, + 0.0060026999562978745, + 0.004168947692960501, + 0.0033005664590746164, + 0.004778427071869373, + 0.003029242856428027, + 0.005216080229729414, + 0.005467613693326712, + 0.007981187663972378, + 0.006002406589686871, + 0.005746546201407909, + 0.0041011180728673935, + 0.009237448684871197, + 0.006847895681858063, + 0.007749330252408981, + 0.006811241619288921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.007647404447197914 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107556652277708 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.01107556652277708, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107556652277708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06817440688610077, + 0.03548172488808632, + 0.049670714884996414, + 0.060585372149944305, + 0.6022593975067139, + 0.044022999703884125, + 0.03369157761335373, + 0.029399771243333817, + 0.046773772686719894, + 0.029940227046608925 + ], + "credit_map": [ + 0.010685435496270657, + 0.005803556181490421, + 0.007612265646457672, + 0.007117005065083504, + 0.008639595471322536, + 0.007417804561555386, + 0.007347255013883114, + 0.013551485724747181, + 0.0049714078195393085, + 0.003847929183393717, + 0.004495750181376934, + 0.004489734768867493, + 0.002819654531776905, + 0.0041028354316949844, + 0.004691589158028364, + 0.006062823347747326, + 0.005866509862244129, + 0.004083422943949699, + 0.0064817145466804504, + 0.006469069980084896, + 0.004904754459857941, + 0.0064943814650177956, + 0.00622823741286993, + 0.009142624214291573, + 0.005039227195084095, + 0.003122534602880478, + 0.004604340065270662, + 0.013827577233314514, + 0.004584081936627626, + 0.004268092103302479, + 0.008555212058126926, + 0.016584310680627823, + 0.03481977432966232, + 0.02632651850581169, + 0.06017131730914116, + 0.17608174681663513, + 0.06581980735063553, + 0.16411639750003815, + 0.05505509674549103, + 0.019868798553943634, + 0.005179409869015217, + 0.0035325558856129646, + 0.008252175524830818, + 0.0044145844876766205, + 0.00455118902027607, + 0.004772547632455826, + 0.00482926657423377, + 0.008491270244121552, + 0.003906335681676865, + 0.0038564391434192657, + 0.00470450334250927, + 0.004055032040923834, + 0.005075881257653236, + 0.003970786929130554, + 0.003071481827646494, + 0.005051115062087774, + 0.003273892216384411, + 0.0026692713145166636, + 0.00508160050958395, + 0.003150861244648695, + 0.0033134203404188156, + 0.004238821566104889, + 0.003339480608701706, + 0.004332424607127905, + 0.004703494254499674, + 0.006408311892300844, + 0.005304362624883652, + 0.005300408229231834, + 0.0052807084284722805, + 0.004727570805698633, + 0.007203887216746807, + 0.007845028303563595, + 0.00471260491758585, + 0.0032855067402124405, + 0.002707631094381213, + 0.0033955497201532125, + 0.0034615120384842157, + 0.0032263363245874643, + 0.00497073819860816, + 0.004180346615612507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.01107556652277708 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635065793991089 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004635065793991089, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635065793991089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05340287461876869, + 0.05311219021677971, + 0.10435790568590164, + 0.03885042294859886, + 0.5498276352882385, + 0.05016301944851875, + 0.06366395205259323, + 0.018627328798174858, + 0.034271687269210815, + 0.03372306749224663 + ], + "credit_map": [ + 0.006601417902857065, + 0.003341120667755604, + 0.008657548576593399, + 0.0070420196279883385, + 0.005809989757835865, + 0.008022936061024666, + 0.004578179679811001, + 0.009349657222628593, + 0.006844694260507822, + 0.004723664373159409, + 0.008707870729267597, + 0.00567853357642889, + 0.006932972930371761, + 0.007555901538580656, + 0.006223809439688921, + 0.006444740574806929, + 0.012659819796681404, + 0.011964138597249985, + 0.01519519742578268, + 0.015747344121336937, + 0.010554560460150242, + 0.01575876772403717, + 0.008103194646537304, + 0.014374883845448494, + 0.0039840200915932655, + 0.004728384781628847, + 0.004308488219976425, + 0.004980064462870359, + 0.004834142047911882, + 0.005174002610146999, + 0.0059542469680309296, + 0.004887073300778866, + 0.03188431262969971, + 0.02503274753689766, + 0.0548439621925354, + 0.16254787147045135, + 0.0581696480512619, + 0.14727285504341125, + 0.051651082932949066, + 0.018425103276968002, + 0.008814732544124126, + 0.006645976100116968, + 0.0036518520209938288, + 0.00480914069339633, + 0.0035815939772874117, + 0.00687369192019105, + 0.005703553557395935, + 0.01008247584104538, + 0.008210411295294762, + 0.007053878158330917, + 0.01002388633787632, + 0.010849273763597012, + 0.008035901933908463, + 0.0061574955470860004, + 0.007488107308745384, + 0.00584499491378665, + 0.002250098390504718, + 0.0018677802290767431, + 0.0035298236180096865, + 0.0028865544591099024, + 0.0017429497092962265, + 0.0027209969703108072, + 0.0015594691503793001, + 0.002069656504318118, + 0.0034400131553411484, + 0.004918032791465521, + 0.003907680045813322, + 0.003942529205232859, + 0.005091418046504259, + 0.0030390736646950245, + 0.005351953208446503, + 0.004580986220389605, + 0.004629711154848337, + 0.003731844713911414, + 0.003359660040587187, + 0.002561404136940837, + 0.005104871466755867, + 0.004691063892096281, + 0.003979955799877644, + 0.00566455302760005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.004635065793991089 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633458375930786 + ], + "opposite_logprob_tokens": [ + -3.651334524154663 + ], + "logprob_reward": -0.02633458375930786, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633458375930786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08288541436195374, + 0.08370605856180191, + 0.07214584201574326, + 0.1379653960466385, + 0.3476477563381195, + 0.04374750703573227, + 0.055191121995449066, + 0.04591117426753044, + 0.06444693356752396, + 0.06635275483131409 + ], + "credit_map": [ + 0.010418643243610859, + 0.005201747640967369, + 0.013330895453691483, + 0.008000584319233894, + 0.0084471944719553, + 0.009413962252438068, + 0.007331253029406071, + 0.02074112929403782, + 0.014747479930520058, + 0.007692680694162846, + 0.009932785294950008, + 0.006087008863687515, + 0.008054893463850021, + 0.011682211421430111, + 0.0063936118967831135, + 0.019115392118692398, + 0.008927693590521812, + 0.008980710059404373, + 0.009618598036468029, + 0.011603799648582935, + 0.006309253629297018, + 0.012144248001277447, + 0.004877294879406691, + 0.009684249758720398, + 0.013147215358912945, + 0.008847711607813835, + 0.016270579770207405, + 0.02975502796471119, + 0.009594092145562172, + 0.01068115048110485, + 0.016945024952292442, + 0.03272460401058197, + 0.020024748519062996, + 0.0159237589687109, + 0.03396419435739517, + 0.09855329990386963, + 0.03850651904940605, + 0.09498841315507889, + 0.031685393303632736, + 0.014001421630382538, + 0.006131390575319529, + 0.004477724898606539, + 0.005335379391908646, + 0.005606017541140318, + 0.003949628211557865, + 0.006688544526696205, + 0.0054151988588273525, + 0.006143625359982252, + 0.007973620668053627, + 0.005755099933594465, + 0.009234718047082424, + 0.008812787011265755, + 0.004720422904938459, + 0.006960317492485046, + 0.005819006823003292, + 0.005915149115025997, + 0.005693668965250254, + 0.00450492138043046, + 0.00780651206150651, + 0.0075894370675086975, + 0.004003916401416063, + 0.006995142437517643, + 0.003447786206379533, + 0.0058697909116744995, + 0.006142755970358849, + 0.010413113981485367, + 0.006150704342871904, + 0.006904464680701494, + 0.00921070296317339, + 0.0067099216394126415, + 0.011543097905814648, + 0.007372172549366951, + 0.008842462673783302, + 0.007595603354275227, + 0.007294494193047285, + 0.005378235597163439, + 0.009856017306447029, + 0.009182654321193695, + 0.00936911441385746, + 0.008834179490804672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.02633458375930786 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, low-angle, sweeping shot plunges us into a world of powerful, hand-drawn sketches constructed entirely of charcoal and graphite. Amidst this abstract waterscape of wild brushstrokes and smudged ink, two butterfly carp, each with enormous, silky fins, engage in a fierce duel. One is pitch black, its scales outlined with fine crosshatching; the other is pure white, its bright areas highlighted by bold white space, creating a stark visual contrast. The camera follows them in a tense, rotating pattern, capturing every moment as they chase, spiral, and collide. Their vast, flowing fins trace graceful yet deadly arcs with each swing, and every violent collision stirs up splashes of ink and broken lines across the frame, as if their struggle is composing the surrounding environment in real time. The high-contrast light and shadow effects are extremely emphasized through heavy shadows and sharp white spaces. There is not a trace of superfluous color in the entire scene, only the pure confrontation of black, white and gray, creating a dynamic beauty that is both tense and magnificent as well as full of primitive explosive power.", + "seed": 5037, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_5/index_1245_seed_5037_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02553176339715719, + "vlm_logprob": -0.02553176339715719, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.02553176339715719, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.051892463117837906 + ], + "opposite_logprob_tokens": [ + -3.0518925189971924 + ], + "logprob_reward": -0.051892463117837906, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.051892463117837906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21689586341381073, + 0.08854148536920547, + 0.11336500942707062, + 0.05519188940525055, + 0.12949787080287933, + 0.04928524047136307, + 0.061953987926244736, + 0.08218549937009811, + 0.07198358327150345, + 0.13109955191612244 + ], + "credit_map": [ + 0.029698416590690613, + 0.024876639246940613, + 0.017018085345625877, + 0.022791313007473946, + 0.04305028170347214, + 0.025498833507299423, + 0.029859598726034164, + 0.024102717638015747, + 0.019153928384184837, + 0.006720234174281359, + 0.006121085491031408, + 0.005417431239038706, + 0.030273502692580223, + 0.004857009742408991, + 0.010381738655269146, + 0.005616564303636551, + 0.02177889086306095, + 0.01282082125544548, + 0.015721522271633148, + 0.009524806402623653, + 0.02061566524207592, + 0.013220922090113163, + 0.0103653809055686, + 0.009317017160356045, + 0.009072757326066494, + 0.005021112971007824, + 0.008069075644016266, + 0.00371545716188848, + 0.006760045886039734, + 0.013045979663729668, + 0.004353016149252653, + 0.005154450889676809, + 0.012460882775485516, + 0.016387904062867165, + 0.025073658674955368, + 0.013349728658795357, + 0.018896188586950302, + 0.012206552550196648, + 0.017302947118878365, + 0.013820014894008636, + 0.006344081833958626, + 0.005896362941712141, + 0.010067066177725792, + 0.004959624260663986, + 0.0052089025266468525, + 0.004902930930256844, + 0.006131801288574934, + 0.005774475168436766, + 0.008259466849267483, + 0.005079947412014008, + 0.006387790199369192, + 0.004226112272590399, + 0.005879683420062065, + 0.014590252190828323, + 0.0056770616210997105, + 0.011853679083287716, + 0.019355474039912224, + 0.006032485514879227, + 0.01133112981915474, + 0.005954221822321415, + 0.00680143665522337, + 0.010555938817560673, + 0.006134808529168367, + 0.01602001301944256, + 0.007454371079802513, + 0.008626921102404594, + 0.008015722036361694, + 0.0074801635928452015, + 0.00754175940528512, + 0.017145633697509766, + 0.007928017526865005, + 0.007790995296090841, + 0.013663549907505512, + 0.015418273396790028, + 0.011397304944694042, + 0.008717192336916924, + 0.012190165929496288, + 0.01978740654885769, + 0.007403473369777203, + 0.04252220317721367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.051892463117837906 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006758213043212891 + ], + "opposite_logprob_tokens": [ + -5.006758213043213 + ], + "logprob_reward": -0.006758213043212891, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006758213043212891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07645105570554733, + 0.04674878716468811, + 0.21486826241016388, + 0.02826419100165367, + 0.17917509377002716, + 0.03908409923315048, + 0.03625083342194557, + 0.11315477639436722, + 0.05260027199983597, + 0.2134026736021042 + ], + "credit_map": [ + 0.013741347938776016, + 0.008390096947550774, + 0.00746690621599555, + 0.009666837751865387, + 0.008852412924170494, + 0.01003846526145935, + 0.00795728713274002, + 0.010337702929973602, + 0.013216166757047176, + 0.0045756506733596325, + 0.005242597311735153, + 0.005224514286965132, + 0.008056535385549068, + 0.003609968814998865, + 0.0031571933068335056, + 0.0036661578342318535, + 0.03769964724779129, + 0.02106322906911373, + 0.04122341796755791, + 0.022259781137108803, + 0.02075488120317459, + 0.030743207782506943, + 0.01731162890791893, + 0.02381245605647564, + 0.0038998927921056747, + 0.0020735380239784718, + 0.0043703727424144745, + 0.002143903635442257, + 0.003414297476410866, + 0.007403709925711155, + 0.0022291699424386024, + 0.0027293055318295956, + 0.020374130457639694, + 0.01573379524052143, + 0.03299952298402786, + 0.01096700131893158, + 0.016468310728669167, + 0.01903156191110611, + 0.0254985298961401, + 0.038102246820926666, + 0.004281454719603062, + 0.00557967321947217, + 0.006655932869762182, + 0.004908406175673008, + 0.004442746285349131, + 0.0038994373753666878, + 0.004584144800901413, + 0.004732304718345404, + 0.003928078338503838, + 0.0023913735058158636, + 0.0032649768982082605, + 0.00244509126059711, + 0.003003067336976528, + 0.008385375142097473, + 0.0035708046052604914, + 0.009262066334486008, + 0.017343558371067047, + 0.008547354489564896, + 0.016330748796463013, + 0.00748269259929657, + 0.009421062655746937, + 0.015820780768990517, + 0.011492077261209488, + 0.02671649679541588, + 0.0097800949588418, + 0.006169059779495001, + 0.0061657181940972805, + 0.0051248688250780106, + 0.0065220934338867664, + 0.00830956269055605, + 0.004869033582508564, + 0.005659845657646656, + 0.022164061665534973, + 0.03149817883968353, + 0.017234625294804573, + 0.016993138939142227, + 0.02070505917072296, + 0.031498923897743225, + 0.013912652619183064, + 0.05939604714512825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.006758213043212891 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328970469534397 + ], + "opposite_logprob_tokens": [ + -3.773289680480957 + ], + "logprob_reward": -0.02328970469534397, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328970469534397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1118931993842125, + 0.1890629082918167, + 0.13291993737220764, + 0.0350894033908844, + 0.09774914383888245, + 0.032403044402599335, + 0.04187478497624397, + 0.2516151964664459, + 0.04287208616733551, + 0.06452023983001709 + ], + "credit_map": [ + 0.026738911867141724, + 0.011894474737346172, + 0.010413958691060543, + 0.009504412300884724, + 0.015286746434867382, + 0.013909686356782913, + 0.013606969267129898, + 0.01053804811090231, + 0.05813264846801758, + 0.011540858075022697, + 0.01081075519323349, + 0.012266553938388824, + 0.05967407673597336, + 0.009565032087266445, + 0.016954729333519936, + 0.010118287056684494, + 0.027988657355308533, + 0.011852538213133812, + 0.024827180430293083, + 0.012381019070744514, + 0.013437913730740547, + 0.0169150922447443, + 0.010962393134832382, + 0.01455515343695879, + 0.0046881819143891335, + 0.0038897308986634016, + 0.004767751321196556, + 0.002741561969742179, + 0.006667151581496, + 0.004121087491512299, + 0.004179968032985926, + 0.004033970646560192, + 0.012148795649409294, + 0.009790812619030476, + 0.024048397317528725, + 0.011222625151276588, + 0.0067046149633824825, + 0.00895027443766594, + 0.012815545313060284, + 0.012068088166415691, + 0.0032315447460860014, + 0.0047151739709079266, + 0.005758491810411215, + 0.0036109902430325747, + 0.0032238380517810583, + 0.0029624984599649906, + 0.004273606929928064, + 0.004626901354640722, + 0.006233365274965763, + 0.0037284940481185913, + 0.00520360330119729, + 0.0037607743870466948, + 0.0042515164241194725, + 0.006585468538105488, + 0.004978449549525976, + 0.007133117411285639, + 0.03454438969492912, + 0.020595835521817207, + 0.03245322033762932, + 0.017209764569997787, + 0.03173824027180672, + 0.04239635169506073, + 0.017165089026093483, + 0.05551232025027275, + 0.007279635407030582, + 0.004916522651910782, + 0.004660121165215969, + 0.003701371606439352, + 0.006158806849271059, + 0.007206412963569164, + 0.00352184334769845, + 0.00542737590149045, + 0.00718894274905324, + 0.007442408706992865, + 0.006260930560529232, + 0.003912501037120819, + 0.00630832789465785, + 0.009434698149561882, + 0.005377332679927349, + 0.018595105037093163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 2 + ], + "selected_reward": -0.02328970469534397 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628594525158405 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628594525158405, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628594525158405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08884050697088242, + 0.17217174172401428, + 0.08830047398805618, + 0.1625005304813385, + 0.06602054089307785, + 0.06385500729084015, + 0.06699320673942566, + 0.07597436755895615, + 0.05125880613923073, + 0.16408474743366241 + ], + "credit_map": [ + 0.01592034474015236, + 0.00717757036909461, + 0.008129122667014599, + 0.00893708597868681, + 0.010566678829491138, + 0.017042435705661774, + 0.011134512722492218, + 0.00993276759982109, + 0.04766454920172691, + 0.011628405191004276, + 0.010914728976786137, + 0.010732671245932579, + 0.0591021366417408, + 0.009087445214390755, + 0.012465173378586769, + 0.010576650500297546, + 0.016490893438458443, + 0.009867139160633087, + 0.010186197236180305, + 0.010343853384256363, + 0.012837090529501438, + 0.013445677235722542, + 0.008074434474110603, + 0.007055194117128849, + 0.03061348758637905, + 0.01736285910010338, + 0.022997569292783737, + 0.009971392340958118, + 0.015339967794716358, + 0.038914572447538376, + 0.011712058447301388, + 0.015588637441396713, + 0.007226912304759026, + 0.008419569581747055, + 0.009029426611959934, + 0.006644923705607653, + 0.008337236940860748, + 0.00806367676705122, + 0.009435574524104595, + 0.008863226510584354, + 0.006500748451799154, + 0.0093750124797225, + 0.009215138852596283, + 0.005749587435275316, + 0.005967315286397934, + 0.005140473134815693, + 0.008212860673666, + 0.01369388122111559, + 0.009713277220726013, + 0.0056958431378006935, + 0.0074516721069812775, + 0.004224055912345648, + 0.0067183710634708405, + 0.017165210098028183, + 0.006258886773139238, + 0.009765896014869213, + 0.014564266428351402, + 0.006851401645690203, + 0.010271668434143066, + 0.006728316657245159, + 0.007029684726148844, + 0.01131616160273552, + 0.006490934174507856, + 0.012721937149763107, + 0.009900860488414764, + 0.005355786066502333, + 0.0061605037190020084, + 0.0037455714773386717, + 0.006665416527539492, + 0.009357724338769913, + 0.003012290457263589, + 0.007060652133077383, + 0.017000360414385796, + 0.01829688996076584, + 0.015374936163425446, + 0.009678150527179241, + 0.014323456212878227, + 0.024052539840340614, + 0.0104737039655447, + 0.054884713143110275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 3 + ], + "selected_reward": -0.007628594525158405 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808984160423279 + ], + "opposite_logprob_tokens": [ + -3.2880897521972656 + ], + "logprob_reward": -0.03808984160423279, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808984160423279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08139859139919281, + 0.34380003809928894, + 0.06471915543079376, + 0.05914681777358055, + 0.13545899093151093, + 0.065602608025074, + 0.031162181869149208, + 0.12773437798023224, + 0.028168436139822006, + 0.06280877441167831 + ], + "credit_map": [ + 0.012388801202178001, + 0.00725963618606329, + 0.007194128818809986, + 0.007347635459154844, + 0.008792036212980747, + 0.015622922219336033, + 0.010408619418740273, + 0.012384818866848946, + 0.10358025133609772, + 0.021644968539476395, + 0.019374657422304153, + 0.021033799275755882, + 0.11349120736122131, + 0.017486341297626495, + 0.02975691854953766, + 0.017431912943720818, + 0.011220731772482395, + 0.006808329373598099, + 0.010464714840054512, + 0.008700803853571415, + 0.007960123009979725, + 0.006074636708945036, + 0.007232052739709616, + 0.0062577626667916775, + 0.00785212405025959, + 0.0052011325024068356, + 0.010452690534293652, + 0.004622543230652809, + 0.007829390466213226, + 0.009252883493900299, + 0.007812702096998692, + 0.0061233495362102985, + 0.011885630898177624, + 0.01320010144263506, + 0.023161808028817177, + 0.010338307358324528, + 0.019622500985860825, + 0.020244143903255463, + 0.02110706828534603, + 0.015899430960416794, + 0.006584562826901674, + 0.00935404933989048, + 0.009162613190710545, + 0.00604925025254488, + 0.007752552628517151, + 0.009610336273908615, + 0.00948211457580328, + 0.007607127074152231, + 0.0041069346480071545, + 0.0018453876255080104, + 0.0036476701498031616, + 0.0019954778254032135, + 0.002833163132891059, + 0.0092854592949152, + 0.0022514096926897764, + 0.00519667798653245, + 0.026609953492879868, + 0.009014982730150223, + 0.016148641705513, + 0.007286261301487684, + 0.014228727668523788, + 0.019732290878891945, + 0.008831923827528954, + 0.025881605222821236, + 0.003107727039605379, + 0.0031702236738055944, + 0.0026284363120794296, + 0.002917470643296838, + 0.0036535391118377447, + 0.007386308629065752, + 0.00226415297947824, + 0.003040577284991741, + 0.0065366351045668125, + 0.007331428583711386, + 0.006828311365097761, + 0.0041010091081261635, + 0.0064003062434494495, + 0.008169084787368774, + 0.0074242595583200455, + 0.016017744317650795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.03808984160423279 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful, low-angle, sweeping shot plunges us into a world of powerful, hand-drawn sketches constructed entirely of charcoal and graphite. Amidst this abstract waterscape of wild brushstrokes and smudged ink, two butterfly carp, each with enormous, silky fins, engage in a fierce duel. One is pitch black, its scales outlined with fine crosshatching; the other is pure white, its bright areas highlighted by bold white space, creating a stark visual contrast. The camera follows them in a tense, rotating pattern, capturing every moment as they chase, spiral, and collide. Their vast, flowing fins trace graceful yet deadly arcs with each swing, and every violent collision stirs up splashes of ink and broken lines across the frame, as if their struggle is composing the surrounding environment in real time. The high-contrast light and shadow effects are extremely emphasized through heavy shadows and sharp white spaces. There is not a trace of superfluous color in the entire scene, only the pure confrontation of black, white and gray, creating a dynamic beauty that is both tense and magnificent as well as full of primitive explosive power.", + "seed": 5038, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_5/index_1245_seed_5038_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.017917009256780147, + "vlm_logprob": -0.017917009256780147, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.017917009256780147, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01045759953558445 + ], + "opposite_logprob_tokens": [ + -4.885457515716553 + ], + "logprob_reward": -0.01045759953558445, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01045759953558445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11720054596662521, + 0.11841488629579544, + 0.08999688178300858, + 0.167671337723732, + 0.053209926933050156, + 0.06517456471920013, + 0.12817054986953735, + 0.04537041857838631, + 0.04172562435269356, + 0.1730651706457138 + ], + "credit_map": [ + 0.014145693741738796, + 0.013912922702729702, + 0.010202488861978054, + 0.012561147101223469, + 0.01660134643316269, + 0.01695115678012371, + 0.014155695214867592, + 0.018670106306672096, + 0.016732295975089073, + 0.007715753745287657, + 0.021172448992729187, + 0.005613045301288366, + 0.018739761784672737, + 0.009123803116381168, + 0.02552318386733532, + 0.013794596306979656, + 0.015663065016269684, + 0.007720175199210644, + 0.019939282909035683, + 0.0068579548969864845, + 0.007109286729246378, + 0.011094932444393635, + 0.009083845652639866, + 0.012528343126177788, + 0.015549076721072197, + 0.00856328196823597, + 0.014843210577964783, + 0.011772660538554192, + 0.038836605846881866, + 0.01564038172364235, + 0.040380481630563736, + 0.022085653617978096, + 0.008252358995378017, + 0.005293271970003843, + 0.007309565320611, + 0.007923044264316559, + 0.005727561190724373, + 0.005142719950526953, + 0.006627483293414116, + 0.006933924742043018, + 0.006448457017540932, + 0.004737600218504667, + 0.009544284082949162, + 0.007045821286737919, + 0.011067637242376804, + 0.008754822425544262, + 0.006937077734619379, + 0.01063887495547533, + 0.016993500292301178, + 0.021297475323081017, + 0.010919072665274143, + 0.015163811855018139, + 0.01164793036878109, + 0.012997349724173546, + 0.016606302931904793, + 0.02254512533545494, + 0.004251610022038221, + 0.0047891270369291306, + 0.0059760557487607, + 0.005184924229979515, + 0.004535216372460127, + 0.007102335337549448, + 0.006992628797888756, + 0.006538527086377144, + 0.005827595479786396, + 0.0053038871847093105, + 0.004305941052734852, + 0.004737309645861387, + 0.005834602285176516, + 0.003997155465185642, + 0.005562519188970327, + 0.006156619172543287, + 0.018842926248908043, + 0.015561110340058804, + 0.01903689093887806, + 0.020035233348608017, + 0.016478756442666054, + 0.02585584670305252, + 0.02653469331562519, + 0.030719710513949394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.01045759953558445 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009798170067369938 + ], + "opposite_logprob_tokens": [ + -4.634798049926758 + ], + "logprob_reward": -0.009798170067369938, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009798170067369938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21721549332141876, + 0.08870775252580643, + 0.12912842631340027, + 0.13098405301570892, + 0.042096011340618134, + 0.10459231585264206, + 0.050412073731422424, + 0.10387405008077621, + 0.05468427389860153, + 0.07830556482076645 + ], + "credit_map": [ + 0.01832045055925846, + 0.029387831687927246, + 0.015918616205453873, + 0.01673957332968712, + 0.03852631151676178, + 0.045444801449775696, + 0.02646339125931263, + 0.02641451545059681, + 0.014670878648757935, + 0.010722796432673931, + 0.009563815779983997, + 0.008111811242997646, + 0.012478847056627274, + 0.008559869602322578, + 0.013351251371204853, + 0.011248487047851086, + 0.023566793650388718, + 0.00981729757040739, + 0.025723055005073547, + 0.008340388536453247, + 0.012465937994420528, + 0.017664650455117226, + 0.01080507691949606, + 0.020745227113366127, + 0.01434115506708622, + 0.011197465471923351, + 0.01490903552621603, + 0.011897058226168156, + 0.01523505337536335, + 0.013133808970451355, + 0.027381855994462967, + 0.02288861945271492, + 0.007505437824875116, + 0.003911126405000687, + 0.006722904741764069, + 0.00537052983418107, + 0.0036797658540308475, + 0.00426521897315979, + 0.00537076685577631, + 0.005270263645797968, + 0.009352610446512699, + 0.008480801247060299, + 0.01881861314177513, + 0.01313258521258831, + 0.00966205820441246, + 0.01603008806705475, + 0.01289080735296011, + 0.016224751248955727, + 0.004910338670015335, + 0.013687263242900372, + 0.005275573115795851, + 0.00606079027056694, + 0.0054277717135846615, + 0.004660343285650015, + 0.00515705905854702, + 0.005232932511717081, + 0.008436143398284912, + 0.008864549919962883, + 0.014756234362721443, + 0.012557058595120907, + 0.008623645640909672, + 0.014583171345293522, + 0.013103512115776539, + 0.022949738427996635, + 0.007372286636382341, + 0.006872445810586214, + 0.005084186792373657, + 0.007059809751808643, + 0.007623020093888044, + 0.004841781221330166, + 0.007272930815815926, + 0.008557812310755253, + 0.007865163497626781, + 0.005423594731837511, + 0.010097367689013481, + 0.013517995364964008, + 0.007545256521552801, + 0.009557169862091541, + 0.012131981551647186, + 0.012167030945420265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.009798170067369938 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160650834441185 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.0160650834441185, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160650834441185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17964990437030792, + 0.08566919714212418, + 0.08732502162456512, + 0.06454745680093765, + 0.08172774314880371, + 0.05857907980680466, + 0.17087894678115845, + 0.15197871625423431, + 0.05271294713020325, + 0.06693097949028015 + ], + "credit_map": [ + 0.015141562558710575, + 0.01970275677740574, + 0.012513358145952225, + 0.020704273134469986, + 0.030445948243141174, + 0.029269486665725708, + 0.02443096786737442, + 0.02744155377149582, + 0.012112488970160484, + 0.007594752591103315, + 0.014307419769465923, + 0.005617675371468067, + 0.015299721620976925, + 0.008994351141154766, + 0.014495547860860825, + 0.007247239351272583, + 0.009489549323916435, + 0.007720859255641699, + 0.013290992937982082, + 0.005529771558940411, + 0.013567879796028137, + 0.012652314268052578, + 0.01400707010179758, + 0.011066590435802937, + 0.0072275977581739426, + 0.004813989158719778, + 0.007385582197457552, + 0.005636637099087238, + 0.010859748348593712, + 0.010032260790467262, + 0.00999150425195694, + 0.008600141853094101, + 0.01970748044550419, + 0.006482457742094994, + 0.016508935019373894, + 0.009354311972856522, + 0.005174548830837011, + 0.007869197987020016, + 0.007708321325480938, + 0.008922490291297436, + 0.0040129367262125015, + 0.004081753082573414, + 0.01202355232089758, + 0.005981191527098417, + 0.004310349002480507, + 0.006535737309604883, + 0.007050183601677418, + 0.014583378098905087, + 0.017864322289824486, + 0.0399542860686779, + 0.01724248006939888, + 0.016840511932969093, + 0.01495036855340004, + 0.016242383047938347, + 0.01708478294312954, + 0.030699806287884712, + 0.012821425683796406, + 0.009899218566715717, + 0.027274450287222862, + 0.012323872186243534, + 0.010872879065573215, + 0.03254850208759308, + 0.0193000640720129, + 0.026938293129205704, + 0.007410360034555197, + 0.00805526040494442, + 0.006742925848811865, + 0.004740392789244652, + 0.0059028510004282, + 0.004498210735619068, + 0.00765241077169776, + 0.007710534613579512, + 0.007490193005651236, + 0.006263074465095997, + 0.006455475930124521, + 0.007705312222242355, + 0.008073274046182632, + 0.00965243298560381, + 0.011218572035431862, + 0.010072645731270313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0160650834441185 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004632692784070969 + ], + "opposite_logprob_tokens": [ + -5.379632472991943 + ], + "logprob_reward": -0.004632692784070969, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004632692784070969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07980336993932724, + 0.2555953562259674, + 0.08265813440084457, + 0.13058984279632568, + 0.0566425584256649, + 0.16049720346927643, + 0.05816435441374779, + 0.0561017245054245, + 0.054691411554813385, + 0.0652560293674469 + ], + "credit_map": [ + 0.010835682041943073, + 0.011184952221810818, + 0.0064145782962441444, + 0.008286050520837307, + 0.010573326610028744, + 0.011336339637637138, + 0.010963814333081245, + 0.010208625346422195, + 0.04011038318276405, + 0.008672947995364666, + 0.06966110318899155, + 0.0080907316878438, + 0.03393268212676048, + 0.015292789787054062, + 0.06654929369688034, + 0.01328542921692133, + 0.009344109334051609, + 0.011099759489297867, + 0.013396731577813625, + 0.007968177087605, + 0.010252430103719234, + 0.009361688047647476, + 0.013228482566773891, + 0.008006758987903595, + 0.011470712721347809, + 0.007978283800184727, + 0.009533770382404327, + 0.007878784090280533, + 0.034330934286117554, + 0.015053576789796352, + 0.032034825533628464, + 0.012308948673307896, + 0.013777701184153557, + 0.0049953097477555275, + 0.010963073000311852, + 0.00865934044122696, + 0.00409128051251173, + 0.0037172248121351004, + 0.00359209137968719, + 0.006846535950899124, + 0.012256848625838757, + 0.009239591658115387, + 0.024335535243153572, + 0.016255591064691544, + 0.026496483013033867, + 0.02220139093697071, + 0.017833057790994644, + 0.03187870606780052, + 0.006639132276177406, + 0.015141492709517479, + 0.006247736979275942, + 0.007854392752051353, + 0.005363240838050842, + 0.005414326675236225, + 0.007612123154103756, + 0.003891903907060623, + 0.00583096407353878, + 0.007295624352991581, + 0.006627541035413742, + 0.006584700662642717, + 0.0049591585993766785, + 0.006533193867653608, + 0.007367066107690334, + 0.010903478600084782, + 0.00554056977853179, + 0.006276839412748814, + 0.006552073173224926, + 0.007186812814325094, + 0.007630613632500172, + 0.005435483064502478, + 0.006705679930746555, + 0.009363343007862568, + 0.006367258261889219, + 0.005813571624457836, + 0.0076310643926262856, + 0.008174823597073555, + 0.006007491145282984, + 0.008980280719697475, + 0.011585024185478687, + 0.01069651823490858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.004632692784070969 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main subjects depicted as butterfly carp with enormous, silky fins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one carp pitch black with fine crosshatching outlines and the other pure white with bold white space highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene constructed entirely of charcoal and graphite sketches with an abstract waterscape background of wild brushstrokes and smudged ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the carp engage in a fierce duel involving chasing, spiraling, and colliding while stirring up splashes of ink and broken lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the subjects in a tense, rotating pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863150045275688 + ], + "opposite_logprob_tokens": [ + -3.048631429672241 + ], + "logprob_reward": -0.04863150045275688, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863150045275688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08346670866012573, + 0.22877970337867737, + 0.06360436230897903, + 0.07140129059553146, + 0.05572008341550827, + 0.32556310296058655, + 0.04323377460241318, + 0.03489993140101433, + 0.06791291385889053, + 0.025418108329176903 + ], + "credit_map": [ + 0.009157601743936539, + 0.009980622678995132, + 0.007736076135188341, + 0.009452326223254204, + 0.00979358796030283, + 0.013784737326204777, + 0.01244965847581625, + 0.011112096719443798, + 0.032417118549346924, + 0.010211069136857986, + 0.05408649519085884, + 0.007819331251084805, + 0.03437972813844681, + 0.021071894094347954, + 0.05306413769721985, + 0.01572992466390133, + 0.009194038808345795, + 0.007892988622188568, + 0.008579721674323082, + 0.006159598473459482, + 0.007878397591412067, + 0.0077426075004041195, + 0.008457853458821774, + 0.00769914872944355, + 0.005324504803866148, + 0.004518615081906319, + 0.006255790125578642, + 0.003947379533201456, + 0.017023814842104912, + 0.011229138821363449, + 0.012902825139462948, + 0.010199216194450855, + 0.008228042162954807, + 0.005785041023045778, + 0.008711779490113258, + 0.009882201440632343, + 0.0045229424722492695, + 0.005215416196733713, + 0.006588179152458906, + 0.006786481477320194, + 0.02339732088148594, + 0.021253829821944237, + 0.05155280977487564, + 0.02590453065931797, + 0.08934050053358078, + 0.032627757638692856, + 0.02847268618643284, + 0.05301368236541748, + 0.00488865002989769, + 0.011268487200140953, + 0.004371377639472485, + 0.005428551230579615, + 0.003932297229766846, + 0.003908909857273102, + 0.005754036828875542, + 0.0036814636550843716, + 0.003605819074437022, + 0.002432066248729825, + 0.004993381444364786, + 0.0038360280450433493, + 0.0029158329125493765, + 0.0062544336542487144, + 0.005603682715445757, + 0.005258687771856785, + 0.009378123097121716, + 0.01003717165440321, + 0.007005989085882902, + 0.008156408555805683, + 0.007102919742465019, + 0.005872590467333794, + 0.0072103226557374, + 0.01314939372241497, + 0.004866683389991522, + 0.0020625940524041653, + 0.0028634651098400354, + 0.0021357920486479998, + 0.0026144948787987232, + 0.0035454407334327698, + 0.004686581902205944, + 0.0026430569123476744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.04863150045275688 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_6/index_tensor([9675, 9675])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_6/index_tensor([9675, 9675])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..01e5da4977ab2ed6b76faa74aac16fc31a3dd18e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_6/index_tensor([9675, 9675])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight close-up initially depicts an antique hardcover book, impossibly suspended in mid-air. Bound in crimson leather with a worn texture and delicate gold-leaf filigree on its spine, the book performs a rhythmic, self-propelled acrobatic movement, tumbling back and forth and gracefully rotating on its vertical axis. With each rotation, its yellowed-edged pages open and close, emitting a soft, trembling whisper amidst the profound silence. The camera then executes a swift, smooth pullback, rapidly zooming out to reveal the book's true context. It is a solitary object, floating in the central aisle of a vast, empty, cathedral-like library. Towering dark mahogany bookshelves, filled with countless books, stretch into the shadows of the vaulted ceiling. Golden afternoon sunlight penetrates the tall, arched windows, illuminating dancing specks of dust and casting long, dramatic shadows across the polished wooden floor. The rapid exposition shifts the scene from one of magical curiosity to one of profound loneliness and quiet wonder, emphasizing a single point of movement within a vast, static, and silent world.", + "seed": 11505, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_6/index_9675_seed_11505_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13204916790127755, + "vlm_logprob": -0.13204916790127755, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.13204916790127755, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02794131264090538 + ], + "opposite_logprob_tokens": [ + -3.7779412269592285 + ], + "logprob_reward": -0.02794131264090538, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02794131264090538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14305053651332855, + 0.08987194299697876, + 0.15165922045707703, + 0.0715978667140007, + 0.043430257588624954, + 0.06756535172462463, + 0.048583246767520905, + 0.21342390775680542, + 0.16078954935073853, + 0.010028144344687462 + ], + "credit_map": [ + 0.014378017745912075, + 0.019104735925793648, + 0.0148715665563941, + 0.02096247673034668, + 0.019145367667078972, + 0.01966526173055172, + 0.014172473922371864, + 0.020750630646944046, + 0.011634374968707561, + 0.009633286856114864, + 0.012232775799930096, + 0.016242636367678642, + 0.010927182622253895, + 0.008143052458763123, + 0.008394087664783001, + 0.012664548121392727, + 0.02322816289961338, + 0.012514357455074787, + 0.018710756674408913, + 0.017791081219911575, + 0.020877627655863762, + 0.017975280061364174, + 0.022187581285834312, + 0.018374362960457802, + 0.00756527204066515, + 0.0072836074978113174, + 0.011077802628278732, + 0.005780481267720461, + 0.008403340354561806, + 0.00714067742228508, + 0.008108763955533504, + 0.01623792015016079, + 0.005367981735616922, + 0.005180057138204575, + 0.006361739244312048, + 0.003870636923238635, + 0.004615530837327242, + 0.004843413829803467, + 0.007446661125868559, + 0.005744231399148703, + 0.011144915595650673, + 0.006877563428133726, + 0.009648076258599758, + 0.007266591303050518, + 0.00790203083306551, + 0.004848558455705643, + 0.00927405059337616, + 0.01060356292873621, + 0.0060190907679498196, + 0.005628540180623531, + 0.007906903512775898, + 0.005385599099099636, + 0.005943913012742996, + 0.006383172702044249, + 0.004446875303983688, + 0.006869153585284948, + 0.040598202496767044, + 0.013443824835121632, + 0.03535320237278938, + 0.0208140816539526, + 0.020403001457452774, + 0.03302375599741936, + 0.024694334715604782, + 0.025093499571084976, + 0.019180290400981903, + 0.011583920568227768, + 0.022520890459418297, + 0.02632405050098896, + 0.022046158090233803, + 0.018396783620119095, + 0.023155787959694862, + 0.017581671476364136, + 0.001680835266597569, + 0.0007088464335538447, + 0.0016082830261439085, + 0.0010098472703248262, + 0.001248220563866198, + 0.0015273622702807188, + 0.0009859808487817645, + 0.0012587691890075803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.02794131264090538 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020598305389285088 + ], + "opposite_logprob_tokens": [ + -3.8955984115600586 + ], + "logprob_reward": -0.020598305389285088, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020598305389285088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05612197145819664, + 0.06719377636909485, + 0.12182946503162384, + 0.16583485901355743, + 0.11404421925544739, + 0.1894483119249344, + 0.15045885741710663, + 0.03083191066980362, + 0.03414088115096092, + 0.07009577006101608 + ], + "credit_map": [ + 0.00714890519157052, + 0.010576360858976841, + 0.006702982354909182, + 0.00895723607391119, + 0.006996248383074999, + 0.005871855188161135, + 0.005159575026482344, + 0.004708812106400728, + 0.009838655591011047, + 0.008083384484052658, + 0.008039015345275402, + 0.01257268525660038, + 0.006908718030899763, + 0.006277549080550671, + 0.006774493958801031, + 0.008699273690581322, + 0.014414848759770393, + 0.008259741589426994, + 0.017599718645215034, + 0.017070520669221878, + 0.013887383043766022, + 0.01630731113255024, + 0.015884799882769585, + 0.018405145034193993, + 0.020528635010123253, + 0.020072652027010918, + 0.02368154376745224, + 0.011864233762025833, + 0.017682190984487534, + 0.01498307753354311, + 0.023819558322429657, + 0.03320297598838806, + 0.013870187103748322, + 0.010493197478353977, + 0.01714077591896057, + 0.010570762678980827, + 0.009141250513494015, + 0.013427682220935822, + 0.02495977096259594, + 0.014440584927797318, + 0.02565452642738819, + 0.0129292793571949, + 0.02873408794403076, + 0.02049139142036438, + 0.024965845048427582, + 0.011375653557479382, + 0.03233373165130615, + 0.03296378254890442, + 0.018277766183018684, + 0.019340919330716133, + 0.024289265275001526, + 0.019689252600073814, + 0.018760597333312035, + 0.014074470847845078, + 0.019554713740944862, + 0.016471870243549347, + 0.004624408669769764, + 0.0026359460316598415, + 0.0049254875630140305, + 0.0026088235899806023, + 0.00458022253587842, + 0.004481434356421232, + 0.0029907957650721073, + 0.0039847916923463345, + 0.005468050483614206, + 0.0018658721819519997, + 0.0050636520609259605, + 0.004214414395391941, + 0.004003708250820637, + 0.0050875525921583176, + 0.0037088135723024607, + 0.004728817380964756, + 0.010340407490730286, + 0.0054792980663478374, + 0.011894925497472286, + 0.007650208659470081, + 0.006772228050976992, + 0.008702312596142292, + 0.007833913899958134, + 0.011422473937273026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.020598305389285088 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205892935395241 + ], + "opposite_logprob_tokens": [ + -2.812058925628662 + ], + "logprob_reward": -0.06205892935395241, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205892935395241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2078775316476822, + 0.11766694486141205, + 0.2146429866552353, + 0.1406111866235733, + 0.1080419048666954, + 0.05771812051534653, + 0.05989956483244896, + 0.047502048313617706, + 0.02476961724460125, + 0.021270083263516426 + ], + "credit_map": [ + 0.016400638967752457, + 0.05093660205602646, + 0.022670704871416092, + 0.0500296875834465, + 0.021732546389102936, + 0.015350432135164738, + 0.018393516540527344, + 0.012363407760858536, + 0.013073284178972244, + 0.010846522636711597, + 0.018607018515467644, + 0.02096197009086609, + 0.01294796634465456, + 0.007664082106202841, + 0.009928074665367603, + 0.023638024926185608, + 0.032462798058986664, + 0.016534999012947083, + 0.027109231799840927, + 0.024896368384361267, + 0.03019520826637745, + 0.025274178013205528, + 0.03029501810669899, + 0.027875186875462532, + 0.014663665555417538, + 0.027695205062627792, + 0.019259896129369736, + 0.008617988787591457, + 0.020657271146774292, + 0.013223346322774887, + 0.022625865414738655, + 0.013867948204278946, + 0.011326217092573643, + 0.010302982293069363, + 0.016014257445931435, + 0.014101951383054256, + 0.007737616077065468, + 0.019032394513487816, + 0.01776534505188465, + 0.01176113449037075, + 0.008379293605685234, + 0.004153999965637922, + 0.011140020564198494, + 0.007373281754553318, + 0.006218057591468096, + 0.003559458302333951, + 0.008349217474460602, + 0.008544789627194405, + 0.007978549227118492, + 0.006145072169601917, + 0.011079879477620125, + 0.007022655103355646, + 0.006750359199941158, + 0.0059801326133310795, + 0.006902128458023071, + 0.008040784858167171, + 0.008045924827456474, + 0.0027390350587666035, + 0.007569102104753256, + 0.005210503004491329, + 0.005190873518586159, + 0.006607227027416229, + 0.004811463877558708, + 0.0073279207572340965, + 0.004170800559222698, + 0.0014271835098043084, + 0.0039747003465890884, + 0.0031066082883626223, + 0.0029842695221304893, + 0.0021055173128843307, + 0.003427392104640603, + 0.0035731452517211437, + 0.0027338415384292603, + 0.0014018487418070436, + 0.002664456842467189, + 0.003011968918144703, + 0.0024214142467826605, + 0.0023681833408772945, + 0.0028116321191191673, + 0.0038567364681512117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.06205892935395241 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233067870140076 + ], + "opposite_logprob_tokens": [ + -0.8983067870140076 + ], + "logprob_reward": -0.5233067870140076, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233067870140076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08330907672643661, + 0.05500326678156853, + 0.08952121436595917, + 0.09351865202188492, + 0.07452493160963058, + 0.19033801555633545, + 0.09429185092449188, + 0.22449351847171783, + 0.04848506301641464, + 0.04651438444852829 + ], + "credit_map": [ + 0.007387020159512758, + 0.00900053046643734, + 0.009271110408008099, + 0.00859223771840334, + 0.013100801035761833, + 0.011562003754079342, + 0.00909571535885334, + 0.015299663878977299, + 0.007395039312541485, + 0.004500475246459246, + 0.007004734594374895, + 0.007519197184592485, + 0.005928540602326393, + 0.0071664233691990376, + 0.005660138558596373, + 0.009828721173107624, + 0.01284225843846798, + 0.007813741452991962, + 0.011456793174147606, + 0.008713502436876297, + 0.014696767553687096, + 0.011242696084082127, + 0.010468259453773499, + 0.012287200428545475, + 0.017673179507255554, + 0.008487638086080551, + 0.014841903932392597, + 0.007042829878628254, + 0.010101454332470894, + 0.010629476048052311, + 0.012525209225714207, + 0.012216961942613125, + 0.008072810247540474, + 0.008424683474004269, + 0.011588326655328274, + 0.00932710524648428, + 0.007062404882162809, + 0.007620819378644228, + 0.012369354255497456, + 0.010059425607323647, + 0.039270371198654175, + 0.01679549179971218, + 0.034117233008146286, + 0.022399786859750748, + 0.01575334370136261, + 0.016429301351308823, + 0.021887250244617462, + 0.02368524670600891, + 0.011310417205095291, + 0.010493619367480278, + 0.01573711633682251, + 0.011570620350539684, + 0.01171241793781519, + 0.006542070768773556, + 0.013608948327600956, + 0.013316645286977291, + 0.026131903752684593, + 0.018909862264990807, + 0.03127491474151611, + 0.03634140267968178, + 0.030694086104631424, + 0.023835092782974243, + 0.03131379932165146, + 0.025992456823587418, + 0.005914983805269003, + 0.002482442185282707, + 0.006544994655996561, + 0.007052800618112087, + 0.006665926426649094, + 0.006362630054354668, + 0.007605130784213543, + 0.005856158211827278, + 0.006852025166153908, + 0.0032447774428874254, + 0.006686266511678696, + 0.005640523973852396, + 0.006376486737281084, + 0.004870344884693623, + 0.0052460310980677605, + 0.007597930263727903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.5233067870140076 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340505108237267 + ], + "opposite_logprob_tokens": [ + -3.6513404846191406 + ], + "logprob_reward": -0.026340505108237267, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340505108237267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19906379282474518, + 0.06102737784385681, + 0.11953360587358475, + 0.10483324527740479, + 0.0621614083647728, + 0.17807908356189728, + 0.07700411230325699, + 0.10080396384000778, + 0.05080920085310936, + 0.04668421670794487 + ], + "credit_map": [ + 0.014281573705375195, + 0.044635459780693054, + 0.023745708167552948, + 0.048995859920978546, + 0.02302025444805622, + 0.014544211328029633, + 0.017342302948236465, + 0.012498420663177967, + 0.0128693338483572, + 0.004889823962002993, + 0.014319888316094875, + 0.007983332499861717, + 0.005504016298800707, + 0.0048272437416017056, + 0.003881626296788454, + 0.0067521147429943085, + 0.02011536806821823, + 0.009808586910367012, + 0.014860063791275024, + 0.015270846895873547, + 0.015171403996646404, + 0.012466802261769772, + 0.017659593373537064, + 0.014180939644575119, + 0.019194815307855606, + 0.008328280411660671, + 0.014935764484107494, + 0.007058572024106979, + 0.010058422572910786, + 0.015273848548531532, + 0.010781004093587399, + 0.019202537834644318, + 0.0073403045535087585, + 0.009738985449075699, + 0.0086294524371624, + 0.007773217279464006, + 0.00526813417673111, + 0.008672117255628109, + 0.008041700348258018, + 0.006697496399283409, + 0.03551928326487541, + 0.01325138472020626, + 0.02732282690703869, + 0.01789894700050354, + 0.01978043094277382, + 0.01347143854945898, + 0.027687881141901016, + 0.02314687706530094, + 0.009667756035923958, + 0.005828820168972015, + 0.012700915336608887, + 0.0065936376340687275, + 0.008811372332274914, + 0.010118918493390083, + 0.0096571184694767, + 0.013625571504235268, + 0.01126126665621996, + 0.009293210692703724, + 0.015298698097467422, + 0.017070867121219635, + 0.015252195298671722, + 0.00783898588269949, + 0.014671633951365948, + 0.01011709775775671, + 0.007826592773199081, + 0.003294567344710231, + 0.007720202673226595, + 0.007079390808939934, + 0.005372862331569195, + 0.008037867024540901, + 0.0052449144423007965, + 0.006232804618775845, + 0.006851953454315662, + 0.0034348000772297382, + 0.00561471376568079, + 0.007110788952559233, + 0.00609424477443099, + 0.00600182730704546, + 0.005854442249983549, + 0.005721447989344597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.026340505108237267 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight close-up initially depicts an antique hardcover book, impossibly suspended in mid-air. Bound in crimson leather with a worn texture and delicate gold-leaf filigree on its spine, the book performs a rhythmic, self-propelled acrobatic movement, tumbling back and forth and gracefully rotating on its vertical axis. With each rotation, its yellowed-edged pages open and close, emitting a soft, trembling whisper amidst the profound silence. The camera then executes a swift, smooth pullback, rapidly zooming out to reveal the book's true context. It is a solitary object, floating in the central aisle of a vast, empty, cathedral-like library. Towering dark mahogany bookshelves, filled with countless books, stretch into the shadows of the vaulted ceiling. Golden afternoon sunlight penetrates the tall, arched windows, illuminating dancing specks of dust and casting long, dramatic shadows across the polished wooden floor. The rapid exposition shifts the scene from one of magical curiosity to one of profound loneliness and quiet wonder, emphasizing a single point of movement within a vast, static, and silent world.", + "seed": 11506, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_6/index_9675_seed_11506_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.027484660409390925, + "vlm_logprob": -0.027484660409390925, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.027484660409390925, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027675462886691093 + ], + "opposite_logprob_tokens": [ + -3.6526753902435303 + ], + "logprob_reward": -0.027675462886691093, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.027675462886691093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5900087952613831, + 0.03603706136345863, + 0.0366063229739666, + 0.04095733165740967, + 0.03562375530600548, + 0.052769165486097336, + 0.03604265674948692, + 0.03174770995974541, + 0.03982491046190262, + 0.10038234293460846 + ], + "credit_map": [ + 0.04204706847667694, + 0.058662306517362595, + 0.10381784290075302, + 0.03523370996117592, + 0.08384323865175247, + 0.18569891154766083, + 0.039736807346343994, + 0.04096890613436699, + 0.0033988608047366142, + 0.004839489236474037, + 0.005774210207164288, + 0.003162106964737177, + 0.00581287732347846, + 0.003913624212145805, + 0.0038187040481716394, + 0.005317189265042543, + 0.006059623789042234, + 0.005748297553509474, + 0.004283759742975235, + 0.0034808763302862644, + 0.0030911907088011503, + 0.00399891659617424, + 0.004223525058478117, + 0.005720133427530527, + 0.00827098824083805, + 0.005325626116245985, + 0.006757077295333147, + 0.004471257328987122, + 0.0029447502456605434, + 0.005706578493118286, + 0.0034282056149095297, + 0.004052847158163786, + 0.01271923165768385, + 0.004803342744708061, + 0.006328387185931206, + 0.0027008396573364735, + 0.0018033260712400079, + 0.00245351018384099, + 0.0020267846994102, + 0.00278833438642323, + 0.01680688001215458, + 0.005468846298754215, + 0.009362980723381042, + 0.007742275018244982, + 0.002640240592882037, + 0.0028315121307969093, + 0.0027316403575241566, + 0.005184795241802931, + 0.004961955826729536, + 0.004328563809394836, + 0.005386354401707649, + 0.004290145821869373, + 0.004131963476538658, + 0.004602511879056692, + 0.0038085351698100567, + 0.004532627295702696, + 0.005116492509841919, + 0.003703567897900939, + 0.0043251123279333115, + 0.005442655645310879, + 0.003557774005457759, + 0.002335858065634966, + 0.0028161327354609966, + 0.0044501167722046375, + 0.005801051389425993, + 0.004429493099451065, + 0.007911665365099907, + 0.005841939710080624, + 0.004855228587985039, + 0.002741569420322776, + 0.0030673916917294264, + 0.005176567938178778, + 0.01924709416925907, + 0.009655263274908066, + 0.011262902058660984, + 0.007746974937617779, + 0.016158318147063255, + 0.004222792107611895, + 0.014435328543186188, + 0.01765366457402706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.027675462886691093 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088063940405846 + ], + "opposite_logprob_tokens": [ + -4.511087894439697 + ], + "logprob_reward": -0.011088063940405846, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088063940405846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13753348588943481, + 0.04635278880596161, + 0.09140682965517044, + 0.08410907536745071, + 0.07608205080032349, + 0.035179462283849716, + 0.10096894204616547, + 0.10157830268144608, + 0.06784117221832275, + 0.25894784927368164 + ], + "credit_map": [ + 0.010594906285405159, + 0.009119785390794277, + 0.027524622157216072, + 0.007706655189394951, + 0.017045872285962105, + 0.0503377839922905, + 0.007481051608920097, + 0.007722820155322552, + 0.0038123847916722298, + 0.00620828103274107, + 0.006063370034098625, + 0.005681704264134169, + 0.007783675566315651, + 0.0045452299527823925, + 0.004967307206243277, + 0.007290837354958057, + 0.014135535806417465, + 0.013677004724740982, + 0.009761592373251915, + 0.007985604926943779, + 0.01740073226392269, + 0.009768325835466385, + 0.008918792940676212, + 0.00975924450904131, + 0.019051026552915573, + 0.011404979042708874, + 0.01415832620114088, + 0.006448653060942888, + 0.004826612304896116, + 0.012463509105145931, + 0.007543283514678478, + 0.008212684653699398, + 0.012888479977846146, + 0.009420210495591164, + 0.012300265952944756, + 0.008458214811980724, + 0.01121652964502573, + 0.00849559623748064, + 0.005109638907015324, + 0.008193115703761578, + 0.00648011639714241, + 0.004794316831976175, + 0.005582898389548063, + 0.006058459170162678, + 0.0035570510663092136, + 0.0022895114962011576, + 0.0031381023582071066, + 0.0032790065743029118, + 0.018799887970089912, + 0.011508500203490257, + 0.010999071411788464, + 0.012627176009118557, + 0.010044421069324017, + 0.012358954176306725, + 0.011321984231472015, + 0.013308953493833542, + 0.01997869461774826, + 0.01320509985089302, + 0.018001863732933998, + 0.013615692034363747, + 0.007455326151102781, + 0.007432898972183466, + 0.008255788125097752, + 0.0136329410597682, + 0.00917865987867117, + 0.012445230968296528, + 0.006946052890270948, + 0.0077468217350542545, + 0.009957931935787201, + 0.005205570720136166, + 0.008300353772938251, + 0.008060554973781109, + 0.06838809698820114, + 0.023273691534996033, + 0.04089109227061272, + 0.023085765540599823, + 0.016981614753603935, + 0.015520056709647179, + 0.02633620612323284, + 0.04447134584188461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.011088063940405846 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306533932685852 + ], + "opposite_logprob_tokens": [ + -3.168065309524536 + ], + "logprob_reward": -0.04306533932685852, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306533932685852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2672758996486664, + 0.11105524003505707, + 0.20723150670528412, + 0.04207409918308258, + 0.030304625630378723, + 0.03361129015684128, + 0.06251035630702972, + 0.08749522268772125, + 0.03866902366280556, + 0.119772769510746 + ], + "credit_map": [ + 0.01732702925801277, + 0.0098385876044631, + 0.05916512385010719, + 0.011264042928814888, + 0.032206110656261444, + 0.10486549139022827, + 0.015777623280882835, + 0.016831887885928154, + 0.011075768619775772, + 0.017203867435455322, + 0.012373288162052631, + 0.014041813090443611, + 0.016590459272265434, + 0.014004654251039028, + 0.012175481766462326, + 0.013589913956820965, + 0.03164716437458992, + 0.025666000321507454, + 0.02615533210337162, + 0.01646064780652523, + 0.04623698443174362, + 0.02335076592862606, + 0.017559975385665894, + 0.020154649391770363, + 0.006657755002379417, + 0.0038262007292360067, + 0.005742054432630539, + 0.00679242517799139, + 0.004882744047790766, + 0.006938403472304344, + 0.0036869440227746964, + 0.003547576954588294, + 0.009124109521508217, + 0.004323798231780529, + 0.00504541490226984, + 0.002348115900531411, + 0.002090532798320055, + 0.002654911484569311, + 0.00217384728603065, + 0.002543899929150939, + 0.006865654606372118, + 0.003811820177361369, + 0.005585297476500273, + 0.004125338513404131, + 0.003161490662023425, + 0.003975244704633951, + 0.002926221350207925, + 0.0031602263916283846, + 0.012540090829133987, + 0.0062323687598109245, + 0.008195459842681885, + 0.010642731562256813, + 0.007839972153306007, + 0.004587559029459953, + 0.005203246604651213, + 0.007268932182341814, + 0.01084977202117443, + 0.011812255717813969, + 0.01010102964937687, + 0.015334781259298325, + 0.011261415667831898, + 0.006658504251390696, + 0.010804298333823681, + 0.010673172771930695, + 0.005172418896108866, + 0.006883553229272366, + 0.0052784513682127, + 0.004468495491892099, + 0.005133030004799366, + 0.002244272269308567, + 0.00486050546169281, + 0.00462829926982522, + 0.033984724432229996, + 0.01014526467770338, + 0.01836561970412731, + 0.010538571514189243, + 0.008059876970946789, + 0.0064114760607481, + 0.011574940755963326, + 0.020692303776741028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.04306533932685852 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306214302778244 + ], + "opposite_logprob_tokens": [ + -3.168062210083008 + ], + "logprob_reward": -0.04306214302778244, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306214302778244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08570819348096848, + 0.03635738790035248, + 0.04149448499083519, + 0.05220016837120056, + 0.08351348340511322, + 0.06320376694202423, + 0.07150740176439285, + 0.11774976551532745, + 0.076645627617836, + 0.3716197609901428 + ], + "credit_map": [ + 0.0059116254560649395, + 0.006859572604298592, + 0.015573833137750626, + 0.006555511616170406, + 0.010732659138739109, + 0.026027146726846695, + 0.005834870506078005, + 0.008212974295020103, + 0.004318124148994684, + 0.00376105890609324, + 0.004506333731114864, + 0.004812696482986212, + 0.004680786747485399, + 0.003342334646731615, + 0.004065167624503374, + 0.006870886776596308, + 0.006536985747516155, + 0.004529107827693224, + 0.005946602672338486, + 0.002597304992377758, + 0.0077271792106330395, + 0.0051241531036794186, + 0.003842918900772929, + 0.005190232768654823, + 0.012847666628658772, + 0.006397380959242582, + 0.009152236394584179, + 0.0052933949045836926, + 0.003923463635146618, + 0.006233124062418938, + 0.004009523894637823, + 0.004343375097960234, + 0.02219821885228157, + 0.010016866959631443, + 0.013885542750358582, + 0.007981727831065655, + 0.010134246200323105, + 0.007078734692186117, + 0.005126170348376036, + 0.007091977167874575, + 0.011131549254059792, + 0.009344127960503101, + 0.010280740447342396, + 0.008912483230233192, + 0.0072347987443208694, + 0.004138090647757053, + 0.006066580303013325, + 0.0060953968204557896, + 0.013060084544122219, + 0.005799644626677036, + 0.009965884499251842, + 0.009166542440652847, + 0.004429594613611698, + 0.00947075616568327, + 0.008553357794880867, + 0.011061535216867924, + 0.02469065599143505, + 0.015502064488828182, + 0.02062944509088993, + 0.019015789031982422, + 0.0056885285302996635, + 0.008949056267738342, + 0.008184894919395447, + 0.015089328400790691, + 0.011243270710110664, + 0.01377900131046772, + 0.010242980904877186, + 0.008940288797020912, + 0.008827040903270245, + 0.005244888830929995, + 0.008426693268120289, + 0.0099414624273777, + 0.09835083037614822, + 0.0334310345351696, + 0.05816907063126564, + 0.031422048807144165, + 0.02534346468746662, + 0.021552901715040207, + 0.03782540559768677, + 0.06552501022815704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.04306214302778244 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012532292865216732 + ], + "opposite_logprob_tokens": [ + -4.3875322341918945 + ], + "logprob_reward": -0.012532292865216732, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532292865216732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13321909308433533, + 0.044552281498909, + 0.03949772194027901, + 0.07664390653371811, + 0.01861521787941456, + 0.07645915448665619, + 0.09431741386651993, + 0.08277139812707901, + 0.05175803601741791, + 0.3821657598018646 + ], + "credit_map": [ + 0.011649321764707565, + 0.0074272071942687035, + 0.025710579007864, + 0.006753463763743639, + 0.016989151015877724, + 0.04759930446743965, + 0.008265577256679535, + 0.0088244853541255, + 0.00587815186008811, + 0.0061704181134700775, + 0.007394068408757448, + 0.004672341514378786, + 0.007379185874015093, + 0.004936457145959139, + 0.0038363987114280462, + 0.004285258706659079, + 0.006062774453312159, + 0.00582968769595027, + 0.007169688586145639, + 0.0032939044758677483, + 0.004742923192679882, + 0.004227547440677881, + 0.0031593278981745243, + 0.005011865869164467, + 0.017359066754579544, + 0.010550123639404774, + 0.013364902697503567, + 0.0074574765749275684, + 0.004680452402681112, + 0.009069358929991722, + 0.006925986614078283, + 0.007236542645841837, + 0.0034673127811402082, + 0.002630115021020174, + 0.0024998337030410767, + 0.002682139165699482, + 0.0016368344658985734, + 0.0018922891467809677, + 0.0019136768532916903, + 0.0018930155783891678, + 0.02517019957304001, + 0.007764095906168222, + 0.013239688239991665, + 0.013268261216580868, + 0.0030826274305582047, + 0.005079307127743959, + 0.003537310753017664, + 0.005317662376910448, + 0.0193696990609169, + 0.011552643030881882, + 0.010728131048381329, + 0.013330038636922836, + 0.009209342300891876, + 0.009533250704407692, + 0.00868192408233881, + 0.011912382207810879, + 0.015657732263207436, + 0.008861241862177849, + 0.014480599202215672, + 0.01527081523090601, + 0.004426259081810713, + 0.005389578640460968, + 0.008342407643795013, + 0.010342767462134361, + 0.007847759872674942, + 0.009057470597326756, + 0.007463691756129265, + 0.007446483708918095, + 0.005670912563800812, + 0.0023457773495465517, + 0.005083942320197821, + 0.006841996684670448, + 0.11775536090135574, + 0.027193671092391014, + 0.05668142810463905, + 0.03328726440668106, + 0.01810724101960659, + 0.01583837904036045, + 0.03784583881497383, + 0.07545655220746994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.012532292865216732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight close-up initially depicts an antique hardcover book, impossibly suspended in mid-air. Bound in crimson leather with a worn texture and delicate gold-leaf filigree on its spine, the book performs a rhythmic, self-propelled acrobatic movement, tumbling back and forth and gracefully rotating on its vertical axis. With each rotation, its yellowed-edged pages open and close, emitting a soft, trembling whisper amidst the profound silence. The camera then executes a swift, smooth pullback, rapidly zooming out to reveal the book's true context. It is a solitary object, floating in the central aisle of a vast, empty, cathedral-like library. Towering dark mahogany bookshelves, filled with countless books, stretch into the shadows of the vaulted ceiling. Golden afternoon sunlight penetrates the tall, arched windows, illuminating dancing specks of dust and casting long, dramatic shadows across the polished wooden floor. The rapid exposition shifts the scene from one of magical curiosity to one of profound loneliness and quiet wonder, emphasizing a single point of movement within a vast, static, and silent world.", + "seed": 11507, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_6/index_9675_seed_11507_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08641816098242998, + "vlm_logprob": -0.08641816098242998, + "vlm_yes_no_margin": 3.1249999523162844, + "vlm_token_credit": -0.08641816098242998, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18369662761688232 + ], + "opposite_logprob_tokens": [ + -1.8086966276168823 + ], + "logprob_reward": -0.18369662761688232, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18369662761688232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2896675169467926, + 0.07677196711301804, + 0.07259557396173477, + 0.196718230843544, + 0.08968795835971832, + 0.08119112253189087, + 0.07180718332529068, + 0.026526592671871185, + 0.05719448998570442, + 0.03783933073282242 + ], + "credit_map": [ + 0.020437682047486305, + 0.0326060988008976, + 0.026826493442058563, + 0.026571223512291908, + 0.06358573585748672, + 0.021838655695319176, + 0.05988384410738945, + 0.037917811423540115, + 0.010659397579729557, + 0.007779848761856556, + 0.013691387139260769, + 0.0092431316152215, + 0.009515590034425259, + 0.008118027821183205, + 0.00853045005351305, + 0.009234133176505566, + 0.013344863429665565, + 0.006789443548768759, + 0.015323052182793617, + 0.006336030550301075, + 0.004845526535063982, + 0.0040632495656609535, + 0.005380570888519287, + 0.016512839123606682, + 0.01213387493044138, + 0.023616552352905273, + 0.0457710437476635, + 0.03057265840470791, + 0.009524411521852016, + 0.026153577491641045, + 0.014321605674922466, + 0.03462450951337814, + 0.01600746065378189, + 0.01252842415124178, + 0.013363110832870007, + 0.0074242097325623035, + 0.008274290710687637, + 0.011814549565315247, + 0.00879090279340744, + 0.011485010385513306, + 0.013296701945364475, + 0.01844540797173977, + 0.012753997929394245, + 0.007297138683497906, + 0.007351174484938383, + 0.008937325328588486, + 0.006443870719522238, + 0.006665500812232494, + 0.008546057157218456, + 0.018643870949745178, + 0.01003209874033928, + 0.009715007618069649, + 0.005930828396230936, + 0.0036571905948221684, + 0.009508193470537663, + 0.005773933604359627, + 0.00417804392054677, + 0.004369932226836681, + 0.003848457243293524, + 0.004079221747815609, + 0.0022659983951598406, + 0.0026634293608367443, + 0.0020303833298385143, + 0.0030911241192370653, + 0.008897163905203342, + 0.004458294250071049, + 0.008569160476326942, + 0.007566267624497414, + 0.004567903000861406, + 0.005674530752003193, + 0.009742881171405315, + 0.007718292530626059, + 0.0034477117005735636, + 0.003691420890390873, + 0.005493928678333759, + 0.003423103829845786, + 0.003962485585361719, + 0.004554636310786009, + 0.009284607134759426, + 0.003981436137109995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.18369662761688232 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821727119386196 + ], + "opposite_logprob_tokens": [ + -4.018217086791992 + ], + "logprob_reward": -0.01821727119386196, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01821727119386196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07206086069345474, + 0.10388463735580444, + 0.16015860438346863, + 0.09472721070051193, + 0.07248653471469879, + 0.17785875499248505, + 0.1715632826089859, + 0.03330482915043831, + 0.058146968483924866, + 0.055808328092098236 + ], + "credit_map": [ + 0.005497491452842951, + 0.012080078013241291, + 0.007610268425196409, + 0.008824915625154972, + 0.009902417659759521, + 0.00752871111035347, + 0.011124332435429096, + 0.009492647834122181, + 0.013669717125594616, + 0.009850256145000458, + 0.014460122212767601, + 0.01435590535402298, + 0.010640555061399937, + 0.01216434221714735, + 0.015554063953459263, + 0.013189674355089664, + 0.022731609642505646, + 0.023793360218405724, + 0.02173749729990959, + 0.020490409806370735, + 0.020064348354935646, + 0.012421094812452793, + 0.016983622685074806, + 0.021936671808362007, + 0.006068849936127663, + 0.00680293096229434, + 0.017705364152789116, + 0.01010138913989067, + 0.010810033418238163, + 0.01077409740537405, + 0.011928479187190533, + 0.02053607814013958, + 0.01510616671293974, + 0.010788332670927048, + 0.008860824629664421, + 0.007330640219151974, + 0.005179738160222769, + 0.009057089686393738, + 0.0072815800085663795, + 0.008882161229848862, + 0.023306461051106453, + 0.03762715682387352, + 0.02722245454788208, + 0.017112765461206436, + 0.014077093452215195, + 0.012187751941382885, + 0.02648935094475746, + 0.01983571983873844, + 0.03053232841193676, + 0.05577114224433899, + 0.02300431579351425, + 0.020290475338697433, + 0.011521542444825172, + 0.0063902391120791435, + 0.014186276122927666, + 0.009866955690085888, + 0.007104963064193726, + 0.004090236034244299, + 0.0048724086955189705, + 0.003968148026615381, + 0.0025689024478197098, + 0.0031461052130907774, + 0.003651305800303817, + 0.003902759635820985, + 0.010742872953414917, + 0.005704366136342287, + 0.008244787342846394, + 0.006419118959456682, + 0.005126313306391239, + 0.0058352090418338776, + 0.00727409403771162, + 0.008800206705927849, + 0.004798148293048143, + 0.00665273517370224, + 0.007368548773229122, + 0.006116777192801237, + 0.006400176789611578, + 0.007435921113938093, + 0.01010203454643488, + 0.006933987140655518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.01821727119386196 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016081858426332474 + ], + "opposite_logprob_tokens": [ + -4.141081809997559 + ], + "logprob_reward": -0.016081858426332474, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081858426332474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15222033858299255, + 0.1970239132642746, + 0.07119672000408173, + 0.2232169359922409, + 0.06542371213436127, + 0.07091939449310303, + 0.07301297783851624, + 0.020762957632541656, + 0.05730361118912697, + 0.06891946494579315 + ], + "credit_map": [ + 0.011684115044772625, + 0.01940622180700302, + 0.014266014099121094, + 0.01636330597102642, + 0.021749991923570633, + 0.01477110292762518, + 0.030188120901584625, + 0.023791471496224403, + 0.0264807790517807, + 0.014989650808274746, + 0.030478863045573235, + 0.01827254891395569, + 0.02530018426477909, + 0.01738595962524414, + 0.033766474574804306, + 0.030349448323249817, + 0.011716562323272228, + 0.008453600108623505, + 0.011570140719413757, + 0.006286896765232086, + 0.01105254516005516, + 0.005131729878485203, + 0.004839876666665077, + 0.012145370244979858, + 0.011183804832398891, + 0.014113581739366055, + 0.042623694986104965, + 0.030201297253370285, + 0.019891411066055298, + 0.013917740434408188, + 0.03305095434188843, + 0.05823444575071335, + 0.008820422925055027, + 0.010020457208156586, + 0.008745892904698849, + 0.006878640968352556, + 0.00601598434150219, + 0.008808226324617863, + 0.008125119842588902, + 0.008008970879018307, + 0.009362269192934036, + 0.014440104365348816, + 0.01013989094644785, + 0.008779789321124554, + 0.00555829145014286, + 0.006160570774227381, + 0.010751095600426197, + 0.005727381911128759, + 0.014800363220274448, + 0.015260497108101845, + 0.01160145178437233, + 0.009953695349395275, + 0.0046427566558122635, + 0.004846656695008278, + 0.006111671682447195, + 0.0057958862744271755, + 0.003545154817402363, + 0.002469368278980255, + 0.0027522272430360317, + 0.003423952031880617, + 0.0016459875041618943, + 0.001961996778845787, + 0.0020204929169267416, + 0.0029437777120620012, + 0.007401743903756142, + 0.004873329307883978, + 0.009160879999399185, + 0.0061076185666024685, + 0.005977147724479437, + 0.006429803092032671, + 0.009232399053871632, + 0.008120687678456306, + 0.006010111421346664, + 0.006114252842962742, + 0.007749793119728565, + 0.006266626063734293, + 0.008727861568331718, + 0.00784599781036377, + 0.017674902454018593, + 0.008529920130968094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.016081858426332474 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155863463878632 + ], + "opposite_logprob_tokens": [ + -1.7015585899353027 + ], + "logprob_reward": -0.20155863463878632, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155863463878632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08086961507797241, + 0.10430647432804108, + 0.08804347366094589, + 0.11960020661354065, + 0.056220803409814835, + 0.18149711191654205, + 0.11636444181203842, + 0.06425221264362335, + 0.14220820367336273, + 0.04663746803998947 + ], + "credit_map": [ + 0.0063313571736216545, + 0.01337487529963255, + 0.007982966490089893, + 0.00681186793372035, + 0.018420513719320297, + 0.008460350334644318, + 0.010369746014475822, + 0.009117924608290195, + 0.010969738475978374, + 0.007179793901741505, + 0.016550222411751747, + 0.008468464016914368, + 0.011740585789084435, + 0.012488450855016708, + 0.02314232662320137, + 0.013766879215836525, + 0.016644185408949852, + 0.009308955632150173, + 0.016002757474780083, + 0.006361871957778931, + 0.006978315766900778, + 0.005298463627696037, + 0.00551478611305356, + 0.021934133023023605, + 0.009268282912671566, + 0.014454460702836514, + 0.022069761529564857, + 0.012941832654178143, + 0.0071274228394031525, + 0.024437112733721733, + 0.009148533456027508, + 0.020152797922492027, + 0.006975823547691107, + 0.009404049254953861, + 0.007336630020290613, + 0.00739965308457613, + 0.004986367654055357, + 0.00723479175940156, + 0.007159143220633268, + 0.005724337883293629, + 0.025836028158664703, + 0.03138359636068344, + 0.024352431297302246, + 0.016585823148489, + 0.013798849657177925, + 0.02682403288781643, + 0.017854992300271988, + 0.024861346930265427, + 0.022363051772117615, + 0.017900066450238228, + 0.017919441685080528, + 0.016149673610925674, + 0.009109273552894592, + 0.013590346090495586, + 0.011781947687268257, + 0.007550628390163183, + 0.01055896282196045, + 0.007421697489917278, + 0.008861685171723366, + 0.0076948851346969604, + 0.006536394357681274, + 0.007079451344907284, + 0.006140557583421469, + 0.009958576411008835, + 0.02523101679980755, + 0.01132273767143488, + 0.022120309993624687, + 0.013344553299248219, + 0.018658675253391266, + 0.01634148135781288, + 0.01726469025015831, + 0.017924737185239792, + 0.004496849607676268, + 0.006789896171540022, + 0.0068186204880476, + 0.007683078292757273, + 0.004064606968313456, + 0.005137989297509193, + 0.00568076828494668, + 0.005965657997876406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.20155863463878632 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536413036286831 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536413036286831, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536413036286831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16900579631328583, + 0.08130677789449692, + 0.0671774372458458, + 0.06274435669183731, + 0.04429851844906807, + 0.08688314259052277, + 0.17492271959781647, + 0.09683695435523987, + 0.13666526973247528, + 0.08015897870063782 + ], + "credit_map": [ + 0.012853304855525494, + 0.019680244848132133, + 0.01724383793771267, + 0.016327019780874252, + 0.03373708203434944, + 0.015490725636482239, + 0.029319291934370995, + 0.02435428835451603, + 0.011224878020584583, + 0.008276771754026413, + 0.014013104140758514, + 0.008072170428931713, + 0.009890617802739143, + 0.008401956409215927, + 0.011724157258868217, + 0.00970312487334013, + 0.015603871084749699, + 0.004729165695607662, + 0.014168360270559788, + 0.005395501386374235, + 0.0033110943622887135, + 0.00413883849978447, + 0.0049999733455479145, + 0.014830635860562325, + 0.0076612853445112705, + 0.00545825669541955, + 0.011563318781554699, + 0.00827931147068739, + 0.0033955806866288185, + 0.009450011886656284, + 0.004692249931395054, + 0.01224434468895197, + 0.0062262737192213535, + 0.007219903636723757, + 0.006250900682061911, + 0.005029592197388411, + 0.0036275233142077923, + 0.007304949685931206, + 0.003900800831615925, + 0.00473857531324029, + 0.015315715223550797, + 0.0181693434715271, + 0.013469281606376171, + 0.01100990641862154, + 0.005491393618285656, + 0.005925131496042013, + 0.00959523394703865, + 0.00790714006870985, + 0.034943655133247375, + 0.06184939667582512, + 0.025271650403738022, + 0.020010942593216896, + 0.010670007206499577, + 0.007012834306806326, + 0.008237648755311966, + 0.006926584057509899, + 0.01246380340307951, + 0.009973062202334404, + 0.013334192335605621, + 0.014104320667684078, + 0.012340654619038105, + 0.00896033737808466, + 0.013585480861365795, + 0.012075105682015419, + 0.016739124432206154, + 0.011366098187863827, + 0.02064044214785099, + 0.015568426810204983, + 0.015377863310277462, + 0.014345552772283554, + 0.02179158665239811, + 0.020836176350712776, + 0.009158320724964142, + 0.009647511877119541, + 0.01169264130294323, + 0.008474706672132015, + 0.007161916699260473, + 0.012996792793273926, + 0.011286289431154728, + 0.009740796871483326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.012536413036286831 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight close-up initially depicts an antique hardcover book, impossibly suspended in mid-air. Bound in crimson leather with a worn texture and delicate gold-leaf filigree on its spine, the book performs a rhythmic, self-propelled acrobatic movement, tumbling back and forth and gracefully rotating on its vertical axis. With each rotation, its yellowed-edged pages open and close, emitting a soft, trembling whisper amidst the profound silence. The camera then executes a swift, smooth pullback, rapidly zooming out to reveal the book's true context. It is a solitary object, floating in the central aisle of a vast, empty, cathedral-like library. Towering dark mahogany bookshelves, filled with countless books, stretch into the shadows of the vaulted ceiling. Golden afternoon sunlight penetrates the tall, arched windows, illuminating dancing specks of dust and casting long, dramatic shadows across the polished wooden floor. The rapid exposition shifts the scene from one of magical curiosity to one of profound loneliness and quiet wonder, emphasizing a single point of movement within a vast, static, and silent world.", + "seed": 11508, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_6/index_9675_seed_11508_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.037678617238998416, + "vlm_logprob": -0.037678617238998416, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.037678617238998416, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05112174153327942 + ], + "opposite_logprob_tokens": [ + -3.051121711730957 + ], + "logprob_reward": -0.05112174153327942, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05112174153327942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2473490834236145, + 0.042700331658124924, + 0.08731820434331894, + 0.1522361934185028, + 0.32047879695892334, + 0.0408746600151062, + 0.041112858802080154, + 0.03137744963169098, + 0.025803856551647186, + 0.010748582892119884 + ], + "credit_map": [ + 0.019546788185834885, + 0.027271853759884834, + 0.058947835117578506, + 0.031069695949554443, + 0.02542443946003914, + 0.025332830846309662, + 0.025153080001473427, + 0.034602563828229904, + 0.004638578277081251, + 0.003994223196059465, + 0.008893546648323536, + 0.004853804595768452, + 0.003131685545668006, + 0.004526326432824135, + 0.006153057795017958, + 0.006509110331535339, + 0.007610067259520292, + 0.005467420443892479, + 0.013776816427707672, + 0.006656944751739502, + 0.004578264430165291, + 0.012316403910517693, + 0.012974665500223637, + 0.02393762394785881, + 0.01154237985610962, + 0.01661009155213833, + 0.022876350209116936, + 0.017026877030730247, + 0.018554823473095894, + 0.02260696142911911, + 0.022280262783169746, + 0.020738445222377777, + 0.0366496704518795, + 0.015350941568613052, + 0.10620180517435074, + 0.09458567202091217, + 0.013469488359987736, + 0.026628393679857254, + 0.016598744317889214, + 0.010994059965014458, + 0.008667340502142906, + 0.004133783746510744, + 0.006790746934711933, + 0.009402128867805004, + 0.0016367885982617736, + 0.0033421264961361885, + 0.002245265292003751, + 0.004656483884900808, + 0.00562636461108923, + 0.007101460825651884, + 0.005494360346347094, + 0.006302237045019865, + 0.004470548126846552, + 0.0028735778760164976, + 0.006034740246832371, + 0.0032095685601234436, + 0.004856928717344999, + 0.004282413981854916, + 0.004243947099894285, + 0.003359410911798477, + 0.004749279003590345, + 0.0017062623519450426, + 0.005493884440511465, + 0.002685325685888529, + 0.004337280988693237, + 0.0012250356376171112, + 0.010061745531857014, + 0.005181307438760996, + 0.0011594723910093307, + 0.0013936386676505208, + 0.001252502202987671, + 0.0011928738094866276, + 0.001516931690275669, + 0.0010614022612571716, + 0.0015263467794284225, + 0.001324970624409616, + 0.0013752959202975035, + 0.0014328877441585064, + 0.0010227974271401763, + 0.0014879509108141065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.05112174153327942 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016089951619505882 + ], + "opposite_logprob_tokens": [ + -4.141089916229248 + ], + "logprob_reward": -0.016089951619505882, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016089951619505882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15210650861263275, + 0.04843056574463844, + 0.1896796077489853, + 0.039661817252635956, + 0.17987090349197388, + 0.1056714579463005, + 0.07067306339740753, + 0.06391895562410355, + 0.08800298720598221, + 0.061984118074178696 + ], + "credit_map": [ + 0.010896299965679646, + 0.018599502742290497, + 0.039904627948999405, + 0.019374899566173553, + 0.015601493418216705, + 0.0130582544952631, + 0.012583362869918346, + 0.022088071331381798, + 0.002523717936128378, + 0.0024841551203280687, + 0.011112963780760765, + 0.004943881183862686, + 0.003063902724534273, + 0.0053741480223834515, + 0.0063211326487362385, + 0.0126066654920578, + 0.01772260293364525, + 0.016611138358712196, + 0.03622686117887497, + 0.01771119050681591, + 0.020502638071775436, + 0.028693746775388718, + 0.033737506717443466, + 0.018473926931619644, + 0.004241312388330698, + 0.005897775292396545, + 0.006317142862826586, + 0.003985540010035038, + 0.0038500099908560514, + 0.005014633759856224, + 0.005949367303401232, + 0.004406037740409374, + 0.015302366577088833, + 0.00807890947908163, + 0.055742599070072174, + 0.05427411571145058, + 0.011918067932128906, + 0.01440630853176117, + 0.009078366681933403, + 0.011070173233747482, + 0.015747053548693657, + 0.009625371545553207, + 0.017064178362488747, + 0.026130646467208862, + 0.006581027526408434, + 0.009588995948433876, + 0.0072823804803192616, + 0.01365180779248476, + 0.009290092624723911, + 0.011548209935426712, + 0.008973806165158749, + 0.009754355996847153, + 0.008192579261958599, + 0.006908820942044258, + 0.009613433852791786, + 0.006391766015440226, + 0.008881541900336742, + 0.009004727937281132, + 0.011039220727980137, + 0.007958394475281239, + 0.004749613348394632, + 0.007486248854547739, + 0.00742239598184824, + 0.007376811932772398, + 0.01711447909474373, + 0.007775460369884968, + 0.011834181845188141, + 0.009778783656656742, + 0.009234358556568623, + 0.014440403319895267, + 0.010936561971902847, + 0.006888758856803179, + 0.010183259844779968, + 0.0064233411103487015, + 0.0063879843801259995, + 0.007046414539217949, + 0.006174187175929546, + 0.007726862095296383, + 0.008536252193152905, + 0.009505818597972393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.016089951619505882 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062051087617874146 + ], + "opposite_logprob_tokens": [ + -2.8120510578155518 + ], + "logprob_reward": -0.062051087617874146, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062051087617874146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1086379885673523, + 0.06304875761270523, + 0.11540519446134567, + 0.10881185531616211, + 0.2541361451148987, + 0.1440320461988449, + 0.06935276091098785, + 0.07527413219213486, + 0.035794079303741455, + 0.025507057085633278 + ], + "credit_map": [ + 0.007768051233142614, + 0.011737862601876259, + 0.016454584896564484, + 0.015753012150526047, + 0.018697993829846382, + 0.01257311087101698, + 0.014342878013849258, + 0.011310496367514133, + 0.006775110960006714, + 0.006116724107414484, + 0.012766525149345398, + 0.0068439096212387085, + 0.004537273198366165, + 0.0051097325049340725, + 0.008260006085038185, + 0.012639479711651802, + 0.010067650116980076, + 0.008442191407084465, + 0.023401152342557907, + 0.011104123666882515, + 0.015067403204739094, + 0.009911934845149517, + 0.025040898472070694, + 0.012369836680591106, + 0.006910663563758135, + 0.010097869671881199, + 0.01580001600086689, + 0.013223975896835327, + 0.011859375983476639, + 0.015632769092917442, + 0.01944316178560257, + 0.01584402471780777, + 0.026342686265707016, + 0.014393756166100502, + 0.08165232092142105, + 0.07232993096113205, + 0.015937741845846176, + 0.019945163279771805, + 0.014085602015256882, + 0.009448929689824581, + 0.022722339257597923, + 0.014394918456673622, + 0.025901775807142258, + 0.04479232430458069, + 0.005737840197980404, + 0.011634058319032192, + 0.008038857020437717, + 0.010809924453496933, + 0.010390929877758026, + 0.011193658225238323, + 0.009870150126516819, + 0.009216554462909698, + 0.00760933943092823, + 0.0064462581649422646, + 0.010046551935374737, + 0.004579322878271341, + 0.007694950792938471, + 0.011454854160547256, + 0.0070175593718886375, + 0.009073946624994278, + 0.011715204454958439, + 0.0068291109055280685, + 0.01625872403383255, + 0.005229786969721317, + 0.006446548737585545, + 0.0027882622089236975, + 0.0045718420296907425, + 0.005415244959294796, + 0.003169445553794503, + 0.004294154699891806, + 0.006056293845176697, + 0.003052287269383669, + 0.004664294421672821, + 0.0027837944217026234, + 0.003091776045039296, + 0.0027104374021291733, + 0.002704234328120947, + 0.0031721123959869146, + 0.0031158500351011753, + 0.00326455757021904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.062051087617874146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306533932685852 + ], + "opposite_logprob_tokens": [ + -3.168065309524536 + ], + "logprob_reward": -0.04306533932685852, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306533932685852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1374642252922058, + 0.04149897024035454, + 0.13830120861530304, + 0.11518359184265137, + 0.2157641351222992, + 0.06232057511806488, + 0.10927220433950424, + 0.05582057684659958, + 0.07903324067592621, + 0.04534126818180084 + ], + "credit_map": [ + 0.00838480331003666, + 0.013690545223653316, + 0.03385606035590172, + 0.019123297184705734, + 0.01545861829072237, + 0.012163878418505192, + 0.01313177589327097, + 0.021655244752764702, + 0.003981131128966808, + 0.0030222786590456963, + 0.008461271412670612, + 0.006175855174660683, + 0.0024769555311650038, + 0.00407314533367753, + 0.004566393792629242, + 0.008741938509047031, + 0.015643225982785225, + 0.010083741508424282, + 0.03543403744697571, + 0.005946733057498932, + 0.009128131903707981, + 0.013314074836671352, + 0.021032171323895454, + 0.02771908789873123, + 0.008134225383400917, + 0.010743468068540096, + 0.017205217853188515, + 0.013175562024116516, + 0.013606615364551544, + 0.01545831561088562, + 0.02008637599647045, + 0.016773801296949387, + 0.01889779232442379, + 0.009235376492142677, + 0.07713781297206879, + 0.06998652964830399, + 0.007865116000175476, + 0.017523864284157753, + 0.009037974290549755, + 0.00607965886592865, + 0.010534846223890781, + 0.007826785556972027, + 0.01048255804926157, + 0.015802403911948204, + 0.0030187484808266163, + 0.0065672858618199825, + 0.003156224964186549, + 0.004931719973683357, + 0.013075713068246841, + 0.014922544360160828, + 0.016949577257037163, + 0.01263861358165741, + 0.01069676224142313, + 0.014538086019456387, + 0.014222430065274239, + 0.012228483334183693, + 0.007722164038568735, + 0.007694901432842016, + 0.010082917287945747, + 0.00902275275439024, + 0.004360297694802284, + 0.005469490773975849, + 0.007531453389674425, + 0.003936596680432558, + 0.012031701393425465, + 0.006937788799405098, + 0.010516209527850151, + 0.01071491651237011, + 0.008097619749605656, + 0.013216299004852772, + 0.010240006260573864, + 0.007278708275407553, + 0.006895179394632578, + 0.0040599810890853405, + 0.007000272627919912, + 0.006355384364724159, + 0.004318976309150457, + 0.004921684507280588, + 0.004784330725669861, + 0.007005458697676659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.04306533932685852 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an antique hardcover book bound in crimson leather with gold-leaf filigree on its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the book perform a rhythmic acrobatic movement involving tumbling and rotating on its vertical axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the book's yellowed-edged pages open and close with each rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift, smooth pullback from a tight close-up to reveal a vast, empty cathedral-like library?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a library with towering dark mahogany bookshelves, arched windows, and golden afternoon sunlight illuminating dust specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064966097474098 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.016064966097474098, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064966097474098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08515921235084534, + 0.07404758781194687, + 0.10098230838775635, + 0.10430549085140228, + 0.08106166869401932, + 0.1257142871618271, + 0.07037075608968735, + 0.11387158185243607, + 0.09328388422727585, + 0.15120317041873932 + ], + "credit_map": [ + 0.009280717000365257, + 0.008869534358382225, + 0.017759894952178, + 0.009869751520454884, + 0.00688420794904232, + 0.008194084279239178, + 0.010381441563367844, + 0.01391958724707365, + 0.006340994965285063, + 0.009143865667283535, + 0.01757483184337616, + 0.006742482539266348, + 0.005623774603009224, + 0.007294499780982733, + 0.006422052159905434, + 0.014905084855854511, + 0.011807798407971859, + 0.010390388779342175, + 0.019815554842352867, + 0.009278588928282261, + 0.00968868937343359, + 0.010100425221025944, + 0.01661944016814232, + 0.013281431049108505, + 0.009043562225997448, + 0.012992201372981071, + 0.014795958995819092, + 0.010407589375972748, + 0.01148069929331541, + 0.015939675271511078, + 0.016021380200982094, + 0.013624425046145916, + 0.009383447468280792, + 0.005214277654886246, + 0.015801016241312027, + 0.017643455415964127, + 0.007503893692046404, + 0.008445987477898598, + 0.00767850736156106, + 0.009391086176037788, + 0.028674427419900894, + 0.014121497981250286, + 0.020438753068447113, + 0.026193397119641304, + 0.005004389677196741, + 0.01168401725590229, + 0.006192624568939209, + 0.013405189849436283, + 0.015674805268645287, + 0.010381413623690605, + 0.010666427202522755, + 0.00818421971052885, + 0.006505450699478388, + 0.007099345792084932, + 0.006468425039201975, + 0.005390665959566832, + 0.022323962301015854, + 0.021796230226755142, + 0.01667790114879608, + 0.015517216175794601, + 0.009972833096981049, + 0.009133185259997845, + 0.009497877210378647, + 0.008952380158007145, + 0.011744282208383083, + 0.007398167625069618, + 0.02316255494952202, + 0.02136625535786152, + 0.008362495340406895, + 0.010709620080888271, + 0.005775102414190769, + 0.004765411838889122, + 0.04801460728049278, + 0.019504712894558907, + 0.020643556490540504, + 0.011107809841632843, + 0.014004993252456188, + 0.017262380570173264, + 0.009628343395888805, + 0.011036763899028301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.016064966097474098 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_7/index_tensor([8960, 8960])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_7/index_tensor([8960, 8960])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..64bec0debc5a4e2941362326376f3d5bea06394e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_7/index_tensor([8960, 8960])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic low-angle shot rapidly tracks a massive black bull as it explosively charges forward across a sunny plain. Its dark hide gleams in the harsh, direct sunlight, its dense, powerful muscles ripple, its entire form rendered in a slight motion blur to emphasize the sheer speed of its charge. Thick, white steam erupts violently from its flared nostrils, a stark contrast to its black form. With its head lowered, its massive, sleek horns aimed forward, its hooves dig into the dry, ochre earth, kicking up a massive cloud of dust and small pebbles that tumble and swirl in its wake. The scene is filled with primal power, the bull's dark, muscular form contrasting sharply with the bright, chaotic explosions of dust, creating an intense and visceral atmosphere of untamed, explosive violence.", + "seed": 28160, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_7/index_8960_seed_28160_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.00921729365363717, + "vlm_logprob": -0.00921729365363717, + "vlm_yes_no_margin": 5.4, + "vlm_token_credit": -0.00921729365363717, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01211864035576582 + ], + "opposite_logprob_tokens": [ + -6.887118816375732 + ], + "logprob_reward": -0.01211864035576582, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.01211864035576582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1565873622894287, + 0.08652911335229874, + 0.04527895152568817, + 0.06575509160757065, + 0.17542940378189087, + 0.04113226756453514, + 0.08322674036026001, + 0.06858253479003906, + 0.16567043960094452, + 0.11180807650089264 + ], + "credit_map": [ + 0.03204543888568878, + 0.009750144556164742, + 0.0162655059248209, + 0.007715100888162851, + 0.00867677666246891, + 0.02565549500286579, + 0.009627691470086575, + 0.04685121402144432, + 0.017436617985367775, + 0.008723405189812183, + 0.01752566732466221, + 0.006586004514247179, + 0.00792760867625475, + 0.013734949752688408, + 0.006492448039352894, + 0.008102412335574627, + 0.007766468450427055, + 0.003520185360684991, + 0.008601033128798008, + 0.002844297094270587, + 0.006666661240160465, + 0.00614472059533, + 0.004557471722364426, + 0.005178114399313927, + 0.011563525535166264, + 0.010282495059072971, + 0.007224987726658583, + 0.0052313292399048805, + 0.00946188997477293, + 0.010718008503317833, + 0.004010939504951239, + 0.0072619179263710976, + 0.009203189052641392, + 0.012499364092946053, + 0.027814393863081932, + 0.005428689066320658, + 0.006982137449085712, + 0.06962809711694717, + 0.005276941228657961, + 0.03859659284353256, + 0.005776433739811182, + 0.004832637030631304, + 0.008495450951159, + 0.003679579123854637, + 0.0051769656129181385, + 0.0067144725471735, + 0.002377281663939357, + 0.004079445730894804, + 0.006672638468444347, + 0.008295819163322449, + 0.009164084680378437, + 0.005446407478302717, + 0.008269856683909893, + 0.026015611365437508, + 0.005181209649890661, + 0.014181115664541721, + 0.0072005996480584145, + 0.011157448403537273, + 0.00862844754010439, + 0.009078611619770527, + 0.006204649806022644, + 0.00976981408894062, + 0.00835026428103447, + 0.008192697539925575, + 0.013127310201525688, + 0.022959398105740547, + 0.03865973278880119, + 0.020479952916502953, + 0.0235220305621624, + 0.014779802411794662, + 0.017024148255586624, + 0.015118055045604706, + 0.009607695043087006, + 0.01335272565484047, + 0.012250667437911034, + 0.026075566187500954, + 0.01127990335226059, + 0.009562958031892776, + 0.01950768381357193, + 0.010170873254537582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.01211864035576582 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820041425526142 + ], + "opposite_logprob_tokens": [ + -4.018200397491455 + ], + "logprob_reward": -0.01820041425526142, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820041425526142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19778306782245636, + 0.11599259078502655, + 0.12546099722385406, + 0.033615272492170334, + 0.12638847529888153, + 0.05963866412639618, + 0.0315980464220047, + 0.14319825172424316, + 0.08149078488349915, + 0.08483387529850006 + ], + "credit_map": [ + 0.04143761470913887, + 0.013801423832774162, + 0.0355960950255394, + 0.012758608907461166, + 0.009287410415709019, + 0.02353043109178543, + 0.009886002168059349, + 0.05148547887802124, + 0.019817590713500977, + 0.015139993280172348, + 0.019268594682216644, + 0.009591269306838512, + 0.012236591428518295, + 0.018044382333755493, + 0.010308638215065002, + 0.011585533618927002, + 0.02490273304283619, + 0.020785609260201454, + 0.017561383545398712, + 0.00857496727257967, + 0.013548802584409714, + 0.019568301737308502, + 0.010254495777189732, + 0.010264710523188114, + 0.0039406935684382915, + 0.0041690790094435215, + 0.004094365518540144, + 0.0033127653878182173, + 0.0036818059161305428, + 0.007062956690788269, + 0.0025931329000741243, + 0.004760473966598511, + 0.013095879927277565, + 0.01450486108660698, + 0.01954059489071369, + 0.00476709334179759, + 0.004997280891984701, + 0.04437286779284477, + 0.003123720409348607, + 0.02198617160320282, + 0.005983924493193626, + 0.006525363773107529, + 0.011044328100979328, + 0.008835354819893837, + 0.005858737975358963, + 0.008345124311745167, + 0.006212302949279547, + 0.006833530962467194, + 0.004431948997080326, + 0.0039034991059452295, + 0.002996121533215046, + 0.0026734613347798586, + 0.003304335754364729, + 0.007848668843507767, + 0.0019569415599107742, + 0.004483070690184832, + 0.01249707955867052, + 0.022331448271870613, + 0.015292157419025898, + 0.011986939236521721, + 0.015979332849383354, + 0.030576586723327637, + 0.014436543919146061, + 0.020098155364394188, + 0.005803847685456276, + 0.008662529289722443, + 0.019465595483779907, + 0.01080700196325779, + 0.010808574967086315, + 0.007578974589705467, + 0.010238529182970524, + 0.008125736378133297, + 0.007492122706025839, + 0.012548261322081089, + 0.010266607627272606, + 0.015975307673215866, + 0.009039415046572685, + 0.009730990044772625, + 0.012155262753367424, + 0.007625909056514502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.01820041425526142 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024903970770537853 + ], + "opposite_logprob_tokens": [ + -6.002490520477295 + ], + "logprob_reward": -0.0024903970770537853, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024903970770537853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25846800208091736, + 0.07570883631706238, + 0.029272593557834625, + 0.03858058527112007, + 0.0361664742231369, + 0.1293490082025528, + 0.10529229789972305, + 0.0677502378821373, + 0.032601192593574524, + 0.22681072354316711 + ], + "credit_map": [ + 0.0893210768699646, + 0.015813764184713364, + 0.019813090562820435, + 0.006716165225952864, + 0.008936461992561817, + 0.02560172788798809, + 0.008373121730983257, + 0.08389260619878769, + 0.015723293647170067, + 0.008950266987085342, + 0.013713574968278408, + 0.005534949246793985, + 0.006448494270443916, + 0.012615599669516087, + 0.005840252619236708, + 0.006882412359118462, + 0.004605816677212715, + 0.0029409152921289206, + 0.004776221700012684, + 0.00241809687577188, + 0.003365087555721402, + 0.0037892882246524096, + 0.0026422005612403154, + 0.004734970163553953, + 0.0074242353439331055, + 0.004264326300472021, + 0.006110980175435543, + 0.002749549224972725, + 0.004490333143621683, + 0.007271510548889637, + 0.0022745938040316105, + 0.003995059989392757, + 0.006479653529822826, + 0.008115071803331375, + 0.005273542366921902, + 0.0035514438059180975, + 0.0029953240882605314, + 0.005506257992237806, + 0.002177747432142496, + 0.002067434834316373, + 0.014347807504236698, + 0.0074919103644788265, + 0.019009023904800415, + 0.007466190494596958, + 0.015019497834146023, + 0.042969055473804474, + 0.0049077169969677925, + 0.01813782937824726, + 0.010450542904436588, + 0.012967498041689396, + 0.010125754401087761, + 0.0066856518387794495, + 0.011733521707355976, + 0.030653953552246094, + 0.006024490110576153, + 0.016650892794132233, + 0.00668096449226141, + 0.008490662090480328, + 0.006028197705745697, + 0.0060933176428079605, + 0.0058927410282194614, + 0.015886137261986732, + 0.006402015220373869, + 0.012276208028197289, + 0.0030909578781574965, + 0.003547399304807186, + 0.0037817375268787146, + 0.004430357366800308, + 0.005432430654764175, + 0.005282523110508919, + 0.004707061219960451, + 0.00232872786000371, + 0.015873029828071594, + 0.01361711323261261, + 0.026983069255948067, + 0.0454728826880455, + 0.018751703202724457, + 0.015929827466607094, + 0.07438004016876221, + 0.015803074464201927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0024903970770537853 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639340288937092 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.008639340288937092, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639340288937092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2798885107040405, + 0.0986860990524292, + 0.09493409097194672, + 0.054662011563777924, + 0.2048249989748001, + 0.028194542974233627, + 0.037252649664878845, + 0.0580277256667614, + 0.057375963777303696, + 0.08615338802337646 + ], + "credit_map": [ + 0.04206180199980736, + 0.021724650636315346, + 0.05360528081655502, + 0.019472556188702583, + 0.016178235411643982, + 0.03605686128139496, + 0.019687065854668617, + 0.0711020678281784, + 0.019679121673107147, + 0.014821858145296574, + 0.017794858664274216, + 0.007589962799102068, + 0.007756664417684078, + 0.014827189967036247, + 0.006930066738277674, + 0.009286394342780113, + 0.015049885027110577, + 0.010296398773789406, + 0.020164722576737404, + 0.006543660536408424, + 0.01003437489271164, + 0.014553547836840153, + 0.007174364291131496, + 0.01111714355647564, + 0.009463397786021233, + 0.006804508622735739, + 0.007187817711383104, + 0.004329929128289223, + 0.004309084732085466, + 0.01047748513519764, + 0.0030716084875166416, + 0.009018182754516602, + 0.010205177590250969, + 0.00911027193069458, + 0.03405354917049408, + 0.00495420815423131, + 0.005399743560701609, + 0.08813877403736115, + 0.005872802808880806, + 0.04709050431847572, + 0.0033644249197095633, + 0.0026637057308107615, + 0.004545475356280804, + 0.0018977036233991385, + 0.003170749871060252, + 0.007261710707098246, + 0.0020633649546653032, + 0.0032274092081934214, + 0.004323666449636221, + 0.005987226963043213, + 0.0036049680784344673, + 0.0031314773950725794, + 0.00719306617975235, + 0.006715438328683376, + 0.002348901005461812, + 0.00394790805876255, + 0.006114807911217213, + 0.009043836034834385, + 0.005856952164322138, + 0.005445281974971294, + 0.006063721142709255, + 0.010370430536568165, + 0.0061278631910681725, + 0.009004839695990086, + 0.00576150743290782, + 0.008833673782646656, + 0.008050107397139072, + 0.00714680552482605, + 0.010831859894096851, + 0.006159984041005373, + 0.0068763429298996925, + 0.003715692786499858, + 0.006656748708337545, + 0.008268988691270351, + 0.009828727692365646, + 0.01633824035525322, + 0.005174585618078709, + 0.005378869362175465, + 0.029021093621850014, + 0.005486139561980963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.008639340288937092 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637676291167736 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637676291167736, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637676291167736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15653394162654877, + 0.07805408537387848, + 0.02699556201696396, + 0.03279481828212738, + 0.22119055688381195, + 0.10428175330162048, + 0.061402566730976105, + 0.07487696409225464, + 0.06698239594697952, + 0.17688727378845215 + ], + "credit_map": [ + 0.02814514748752117, + 0.01212347112596035, + 0.020882654935121536, + 0.010125093162059784, + 0.009955640882253647, + 0.024140160530805588, + 0.010763587430119514, + 0.04039818048477173, + 0.013108387589454651, + 0.012211115099489689, + 0.013335521332919598, + 0.004805194213986397, + 0.008430996909737587, + 0.01291273720562458, + 0.005829363130033016, + 0.007420772686600685, + 0.003643433563411236, + 0.0031665852293372154, + 0.004562282469123602, + 0.002073695883154869, + 0.003279042197391391, + 0.003897234331816435, + 0.0030664291698485613, + 0.0033068587072193623, + 0.005923403427004814, + 0.0038303076289594173, + 0.007104484364390373, + 0.003678490873426199, + 0.0022383553441613913, + 0.0050279186107218266, + 0.001675497624091804, + 0.003316359594464302, + 0.010995863936841488, + 0.010028942488133907, + 0.0372052825987339, + 0.007262516301125288, + 0.00841614045202732, + 0.09184254705905914, + 0.009905809536576271, + 0.0455334410071373, + 0.010395723395049572, + 0.00907688308507204, + 0.015566280111670494, + 0.01221950352191925, + 0.012601625174283981, + 0.025936026126146317, + 0.007156483829021454, + 0.01132922898977995, + 0.008576137945055962, + 0.009949786588549614, + 0.006282244808971882, + 0.005646985024213791, + 0.0117462994530797, + 0.009835213422775269, + 0.0032561873085796833, + 0.006109714042395353, + 0.00717185577377677, + 0.01397091057151556, + 0.009677708148956299, + 0.010249756276607513, + 0.008730163797736168, + 0.007282623089849949, + 0.01139336172491312, + 0.006400587968528271, + 0.004555383697152138, + 0.00688486872240901, + 0.01612403616309166, + 0.008940068073570728, + 0.008970667608082294, + 0.006344677414745092, + 0.008206052705645561, + 0.006956643424928188, + 0.006878932937979698, + 0.020076466724276543, + 0.012764301151037216, + 0.05872899293899536, + 0.008786812424659729, + 0.007517437916249037, + 0.052032701671123505, + 0.01010163314640522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.004637676291167736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle shot rapidly tracks a massive black bull as it explosively charges forward across a sunny plain. Its dark hide gleams in the harsh, direct sunlight, its dense, powerful muscles ripple, its entire form rendered in a slight motion blur to emphasize the sheer speed of its charge. Thick, white steam erupts violently from its flared nostrils, a stark contrast to its black form. With its head lowered, its massive, sleek horns aimed forward, its hooves dig into the dry, ochre earth, kicking up a massive cloud of dust and small pebbles that tumble and swirl in its wake. The scene is filled with primal power, the bull's dark, muscular form contrasting sharply with the bright, chaotic explosions of dust, creating an intense and visceral atmosphere of untamed, explosive violence.", + "seed": 28161, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_7/index_8960_seed_28161_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01289488966576755, + "vlm_logprob": -0.01289488966576755, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.01289488966576755, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010598916560411453 + ], + "opposite_logprob_tokens": [ + -5.135599136352539 + ], + "logprob_reward": -0.010598916560411453, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.010598916560411453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2450457215309143, + 0.061411697417497635, + 0.06881923973560333, + 0.04465486481785774, + 0.20671220123767853, + 0.13994334638118744, + 0.10068374872207642, + 0.05885317549109459, + 0.05108338221907616, + 0.022792674601078033 + ], + "credit_map": [ + 0.024023674428462982, + 0.03151394799351692, + 0.04615551233291626, + 0.03547726199030876, + 0.023223331198096275, + 0.04331095889210701, + 0.018943989649415016, + 0.02239704132080078, + 0.006599931046366692, + 0.008567784912884235, + 0.008800109848380089, + 0.005443025380373001, + 0.004184409510344267, + 0.01054423302412033, + 0.004834193270653486, + 0.012438014149665833, + 0.008688628673553467, + 0.010767756029963493, + 0.007369408383965492, + 0.008266068063676357, + 0.010277663357555866, + 0.006235957145690918, + 0.008642038330435753, + 0.008571719750761986, + 0.007162027060985565, + 0.005880034528672695, + 0.005899827927350998, + 0.007250761613249779, + 0.003785262815654278, + 0.004586134105920792, + 0.0057641929015517235, + 0.0043266243301332, + 0.01711493916809559, + 0.016628950834274292, + 0.05250745266675949, + 0.013353991322219372, + 0.016022861003875732, + 0.028175469487905502, + 0.016630662605166435, + 0.046277888119220734, + 0.014124292880296707, + 0.00954167265444994, + 0.041915323585271835, + 0.010319516994059086, + 0.009153432212769985, + 0.021754182875156403, + 0.011486189439892769, + 0.02164873480796814, + 0.011475015431642532, + 0.012848060578107834, + 0.018052326515316963, + 0.009791970252990723, + 0.010321495123207569, + 0.015004051849246025, + 0.007007376290857792, + 0.016183461993932724, + 0.00787772424519062, + 0.007361438125371933, + 0.011083806864917278, + 0.006789418403059244, + 0.006053299177438021, + 0.005729407072067261, + 0.005669818725436926, + 0.008288265205919743, + 0.0074942996725440025, + 0.006676407065242529, + 0.009644225239753723, + 0.004967273212969303, + 0.004914239048957825, + 0.006608005613088608, + 0.004586802329868078, + 0.006192131899297237, + 0.0019789531361311674, + 0.004099983721971512, + 0.0021641571074724197, + 0.0019843687769025564, + 0.004363188520073891, + 0.002235414693132043, + 0.003551037982106209, + 0.0024155694991350174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.010598916560411453 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029819726943969727 + ], + "opposite_logprob_tokens": [ + -3.5298197269439697 + ], + "logprob_reward": -0.029819726943969727, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029819726943969727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17891016602516174, + 0.14635901153087616, + 0.05993141978979111, + 0.0680064782500267, + 0.17677107453346252, + 0.17231965065002441, + 0.05409296602010727, + 0.04596199840307236, + 0.06307146698236465, + 0.03457575663924217 + ], + "credit_map": [ + 0.017569946125149727, + 0.02925162762403488, + 0.032966919243335724, + 0.02012491598725319, + 0.022113898769021034, + 0.028761200606822968, + 0.01387044321745634, + 0.014251220971345901, + 0.014839978888630867, + 0.015032429248094559, + 0.024638649076223373, + 0.013436798006296158, + 0.008709720335900784, + 0.026898857206106186, + 0.012078455649316311, + 0.030724136158823967, + 0.005903428420424461, + 0.007254993077367544, + 0.006900051608681679, + 0.007655388675630093, + 0.007208588998764753, + 0.007120369467884302, + 0.00729061383754015, + 0.010597985237836838, + 0.008114387281239033, + 0.005068624392151833, + 0.020788932219147682, + 0.005218938458710909, + 0.0045291050337255, + 0.004354711621999741, + 0.004819141700863838, + 0.015112632885575294, + 0.014065111055970192, + 0.02154962159693241, + 0.03158562630414963, + 0.014091437682509422, + 0.027568629011511803, + 0.017044734209775925, + 0.016726508736610413, + 0.03413941711187363, + 0.01781451515853405, + 0.011108147911727428, + 0.053160134702920914, + 0.0117537472397089, + 0.009610096924006939, + 0.02591058611869812, + 0.015636220574378967, + 0.027326196432113647, + 0.00437575951218605, + 0.006397655233740807, + 0.010194423608481884, + 0.0070491936057806015, + 0.004984358791261911, + 0.00431025167927146, + 0.003984922543168068, + 0.012796401977539062, + 0.004593007266521454, + 0.009720650501549244, + 0.004754035267978907, + 0.006256715394556522, + 0.006018531043082476, + 0.0038894754834473133, + 0.005149258300662041, + 0.00558032700791955, + 0.013446074910461903, + 0.00950043648481369, + 0.010411798022687435, + 0.005421987269073725, + 0.0063362871296703815, + 0.006240937393158674, + 0.005248239263892174, + 0.006465711165219545, + 0.0038942222017794847, + 0.004986356943845749, + 0.006293435115367174, + 0.005124822724610567, + 0.004982260055840015, + 0.002814728766679764, + 0.0033834348432719707, + 0.003096497617661953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.029819726943969727 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107780635356903 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.01107780635356903, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107780635356903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1920575350522995, + 0.11322372406721115, + 0.027816900983452797, + 0.06316158175468445, + 0.04236165061593056, + 0.3482229709625244, + 0.16632647812366486, + 0.02478310652077198, + 0.011868017725646496, + 0.010178079828619957 + ], + "credit_map": [ + 0.037929147481918335, + 0.01730278506875038, + 0.0408199168741703, + 0.016280801966786385, + 0.012354251928627491, + 0.03602306917309761, + 0.008451917208731174, + 0.022895630449056625, + 0.010710439644753933, + 0.018220514059066772, + 0.014906959608197212, + 0.021333487704396248, + 0.0143683897331357, + 0.009766348637640476, + 0.012659905478358269, + 0.011257672682404518, + 0.005769465118646622, + 0.003604272613301873, + 0.003024136880412698, + 0.0023314624559134245, + 0.0029896290507167578, + 0.004132739268243313, + 0.00241595390252769, + 0.0035492428578436375, + 0.007341020740568638, + 0.003260179189965129, + 0.024048153311014175, + 0.0030606016516685486, + 0.002738677430897951, + 0.002702259225770831, + 0.0026047842111438513, + 0.017405904829502106, + 0.0034822740126401186, + 0.0035119042731821537, + 0.01035773940384388, + 0.0028753564693033695, + 0.004385787062346935, + 0.0051436941139400005, + 0.0031267129816114902, + 0.009478181600570679, + 0.03654094785451889, + 0.023583335801959038, + 0.10672563314437866, + 0.023693755269050598, + 0.020722664892673492, + 0.051167625933885574, + 0.030646758154034615, + 0.05514223501086235, + 0.018398726359009743, + 0.01846029981970787, + 0.030029654502868652, + 0.017701921984553337, + 0.014964754693210125, + 0.02399040013551712, + 0.016176611185073853, + 0.02660411037504673, + 0.004425719380378723, + 0.003302055411040783, + 0.004916962701827288, + 0.0027631877455860376, + 0.0019483715295791626, + 0.0023948305752128363, + 0.0015502219321206212, + 0.0034817589912563562, + 0.0019124140962958336, + 0.0014746906235814095, + 0.0016803287435323, + 0.0013434102293103933, + 0.0012259114300832152, + 0.0012153317220509052, + 0.0010210045147687197, + 0.001994927180930972, + 0.0018771769246086478, + 0.0015072335954755545, + 0.0014399363426491618, + 0.0007121817325241864, + 0.0016824300400912762, + 0.0007519118371419609, + 0.0013027039822191, + 0.0009045048500411212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.01107780635356903 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784122928977013 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784122928977013, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784122928977013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13343986868858337, + 0.37604737281799316, + 0.039056677371263504, + 0.04280071705579758, + 0.0537671260535717, + 0.04171883687376976, + 0.03612932562828064, + 0.2376822531223297, + 0.022772861644625664, + 0.01658495143055916 + ], + "credit_map": [ + 0.017240028828382492, + 0.0180987399071455, + 0.018784215673804283, + 0.02408306673169136, + 0.011474087834358215, + 0.017781412228941917, + 0.011903168633580208, + 0.014075151644647121, + 0.036715805530548096, + 0.05732150375843048, + 0.04737225919961929, + 0.05155239626765251, + 0.03308175504207611, + 0.05483686923980713, + 0.03148960322141647, + 0.06367717683315277, + 0.006403766106814146, + 0.0036969431675970554, + 0.004898098297417164, + 0.005051347892731428, + 0.004301775246858597, + 0.005444088019430637, + 0.003102259710431099, + 0.006158401258289814, + 0.003970639314502478, + 0.002959900302812457, + 0.013101517222821712, + 0.0022735458333045244, + 0.00411939574405551, + 0.003095641965046525, + 0.0034306165762245655, + 0.009849460795521736, + 0.005047143436968327, + 0.006598213221877813, + 0.012593797408044338, + 0.004937402904033661, + 0.007185498718172312, + 0.005565795581787825, + 0.004043084569275379, + 0.007796192076057196, + 0.004832773003727198, + 0.003301705699414015, + 0.011447845958173275, + 0.00284684170037508, + 0.0037034337874501944, + 0.005581917241215706, + 0.002766236662864685, + 0.007238082587718964, + 0.0042370380833745, + 0.0034740271512418985, + 0.006171141751110554, + 0.002668480621650815, + 0.0035744390916079283, + 0.00561080826446414, + 0.004764982033520937, + 0.005628407001495361, + 0.02529235929250717, + 0.031204121187329292, + 0.04487670585513115, + 0.027291711419820786, + 0.02533668838441372, + 0.027809251099824905, + 0.018001019954681396, + 0.03787040337920189, + 0.0038434097077697515, + 0.0038832116406410933, + 0.0034889117814600468, + 0.0021483558230102062, + 0.002503983909264207, + 0.002334281802177429, + 0.001855846494436264, + 0.0027148595545440912, + 0.002132303547114134, + 0.0023355011362582445, + 0.002015738980844617, + 0.002096207346767187, + 0.002569334814324975, + 0.0015022094594314694, + 0.001991787226870656, + 0.0019418695010244846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.009784122928977013 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193875541910529 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.003193875541910529, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193875541910529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18585634231567383, + 0.06415192782878876, + 0.05008598789572716, + 0.04004959389567375, + 0.20577825605869293, + 0.20613594353199005, + 0.08227036893367767, + 0.08322552591562271, + 0.041424043476581573, + 0.04102204740047455 + ], + "credit_map": [ + 0.037241991609334946, + 0.023056533187627792, + 0.026367269456386566, + 0.02081305906176567, + 0.013985148631036282, + 0.027294427156448364, + 0.01143826637417078, + 0.025659656152129173, + 0.00785822607576847, + 0.013405691832304, + 0.0078064026311039925, + 0.009987058117985725, + 0.00714149372652173, + 0.005711241625249386, + 0.0049986522644758224, + 0.007243158295750618, + 0.008933384902775288, + 0.004563083406537771, + 0.0057641793973743916, + 0.006194121669977903, + 0.004484463483095169, + 0.008258206769824028, + 0.003579096170142293, + 0.00830945186316967, + 0.006252242252230644, + 0.00535078439861536, + 0.006411158014088869, + 0.005667545832693577, + 0.003595465561375022, + 0.003239021170884371, + 0.0046540494076907635, + 0.004879325162619352, + 0.025572003796696663, + 0.019989028573036194, + 0.04640018567442894, + 0.016044124960899353, + 0.015926672145724297, + 0.02804267778992653, + 0.017325596883893013, + 0.0364779494702816, + 0.021809956058859825, + 0.012955429963767529, + 0.06334149837493896, + 0.01444205641746521, + 0.011717526242136955, + 0.02998221665620804, + 0.018784131854772568, + 0.03310312703251839, + 0.008523891679942608, + 0.008218678645789623, + 0.013775402680039406, + 0.01152156013995409, + 0.00834222137928009, + 0.008079318329691887, + 0.00919431634247303, + 0.014614979736506939, + 0.0101272352039814, + 0.008572224527597427, + 0.01605856604874134, + 0.007791312411427498, + 0.010342778638005257, + 0.008949494920670986, + 0.007303006947040558, + 0.01408089604228735, + 0.009440318681299686, + 0.004269037861377001, + 0.007542149629443884, + 0.0036515218671411276, + 0.003929306287318468, + 0.00410642521455884, + 0.0037994959857314825, + 0.004685787484049797, + 0.004593661054968834, + 0.006810024380683899, + 0.005259300582110882, + 0.004787585698068142, + 0.007799518760293722, + 0.0037170995492488146, + 0.0048929983749985695, + 0.0031618596985936165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.003193875541910529 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle shot rapidly tracks a massive black bull as it explosively charges forward across a sunny plain. Its dark hide gleams in the harsh, direct sunlight, its dense, powerful muscles ripple, its entire form rendered in a slight motion blur to emphasize the sheer speed of its charge. Thick, white steam erupts violently from its flared nostrils, a stark contrast to its black form. With its head lowered, its massive, sleek horns aimed forward, its hooves dig into the dry, ochre earth, kicking up a massive cloud of dust and small pebbles that tumble and swirl in its wake. The scene is filled with primal power, the bull's dark, muscular form contrasting sharply with the bright, chaotic explosions of dust, creating an intense and visceral atmosphere of untamed, explosive violence.", + "seed": 28162, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_7/index_8960_seed_28162_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.006058515608310699, + "vlm_logprob": -0.006058515608310699, + "vlm_yes_no_margin": 5.65, + "vlm_token_credit": -0.006058515608310699, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00676448829472065 + ], + "opposite_logprob_tokens": [ + -7.1317644119262695 + ], + "logprob_reward": -0.00676448829472065, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.00676448829472065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14376525580883026, + 0.05987069010734558, + 0.10708832740783691, + 0.09814389795064926, + 0.04179249331355095, + 0.12453587353229523, + 0.03965933248400688, + 0.14973123371601105, + 0.18476226925849915, + 0.05065067112445831 + ], + "credit_map": [ + 0.028198417276144028, + 0.013876608572900295, + 0.029458170756697655, + 0.025037216022610664, + 0.01478571817278862, + 0.010374067351222038, + 0.01005642767995596, + 0.01197862159460783, + 0.01462600752711296, + 0.008772753179073334, + 0.008809740655124187, + 0.006886748131364584, + 0.00570487417280674, + 0.004084576852619648, + 0.004164016339927912, + 0.00682197418063879, + 0.014190156012773514, + 0.017658963799476624, + 0.009409070014953613, + 0.012656491249799728, + 0.013505508191883564, + 0.015591014176607132, + 0.01025630533695221, + 0.013820821419358253, + 0.011129376478493214, + 0.004746148828417063, + 0.03422032669186592, + 0.004155603237450123, + 0.003784686326980591, + 0.014340169727802277, + 0.0028583859093487263, + 0.022909201681613922, + 0.0046985275112092495, + 0.006446007639169693, + 0.004325037822127342, + 0.005290135741233826, + 0.008135654032230377, + 0.0037188981659710407, + 0.0034647760912775993, + 0.005713455379009247, + 0.025051498785614967, + 0.014163910411298275, + 0.020248444750905037, + 0.009948560036718845, + 0.011791064403951168, + 0.008073678240180016, + 0.00737467547878623, + 0.027884043753147125, + 0.005584522616118193, + 0.0031134996097534895, + 0.006734166294336319, + 0.004499495029449463, + 0.004090180154889822, + 0.004248033743351698, + 0.0030412725172936916, + 0.008348163217306137, + 0.022000152617692947, + 0.018031183630228043, + 0.022771013900637627, + 0.02083507552742958, + 0.01723986491560936, + 0.010520672425627708, + 0.015102152712643147, + 0.02323111519217491, + 0.05376836284995079, + 0.007982553914189339, + 0.06134200096130371, + 0.008422065526247025, + 0.011305253952741623, + 0.013633279129862785, + 0.009247886948287487, + 0.019060872495174408, + 0.007547029294073582, + 0.010496486909687519, + 0.008313861675560474, + 0.005065163131803274, + 0.005015450995415449, + 0.005418876186013222, + 0.004218141548335552, + 0.004575663711875677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.00676448829472065 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076417261734604836 + ], + "opposite_logprob_tokens": [ + -4.882641792297363 + ], + "logprob_reward": -0.0076417261734604836, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076417261734604836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13628949224948883, + 0.1080414354801178, + 0.06634223461151123, + 0.04147666320204735, + 0.11052154004573822, + 0.0721123218536377, + 0.1562725305557251, + 0.062191300094127655, + 0.18478287756443024, + 0.06196962669491768 + ], + "credit_map": [ + 0.02449619397521019, + 0.015681058168411255, + 0.021562153473496437, + 0.01949779875576496, + 0.015428713522851467, + 0.014182145707309246, + 0.0084720179438591, + 0.016969414427876472, + 0.016173996031284332, + 0.016563912853598595, + 0.01166447252035141, + 0.01571551524102688, + 0.016249217092990875, + 0.010523593984544277, + 0.007478659972548485, + 0.013672069646418095, + 0.008472285233438015, + 0.011002904735505581, + 0.005857287906110287, + 0.008196254260838032, + 0.008333101868629456, + 0.010563811287283897, + 0.0058919996954500675, + 0.00802458543330431, + 0.006759826093912125, + 0.0038198940455913544, + 0.0099325031042099, + 0.0017403990495949984, + 0.0023341819178313017, + 0.006008690223097801, + 0.0014414058532565832, + 0.009439760819077492, + 0.013575424440205097, + 0.013065185397863388, + 0.014881689101457596, + 0.008438715711236, + 0.01766648143529892, + 0.00905856117606163, + 0.006242496892809868, + 0.027592986822128296, + 0.007910413667559624, + 0.009265211410820484, + 0.012566053308546543, + 0.008682727813720703, + 0.01137162372469902, + 0.004711543675512075, + 0.007378693670034409, + 0.0102260522544384, + 0.025110624730587006, + 0.013941936194896698, + 0.025277525186538696, + 0.015484971925616264, + 0.018489539623260498, + 0.016546539962291718, + 0.01541215181350708, + 0.026009242981672287, + 0.007652158848941326, + 0.007100704591721296, + 0.013166835531592369, + 0.008743452839553356, + 0.006452605128288269, + 0.005084630101919174, + 0.005050226580351591, + 0.008940684609115124, + 0.05026455968618393, + 0.009109013713896275, + 0.05403190478682518, + 0.00919566210359335, + 0.010644877329468727, + 0.01838614046573639, + 0.010740294121205807, + 0.02241041697561741, + 0.008955749683082104, + 0.014158044941723347, + 0.008997567929327488, + 0.006778369657695293, + 0.006693460512906313, + 0.007404692005366087, + 0.0055709402076900005, + 0.0034108024556189775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.0076417261734604836 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002489326987415552 + ], + "opposite_logprob_tokens": [ + -6.00248908996582 + ], + "logprob_reward": -0.002489326987415552, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002489326987415552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41554421186447144, + 0.07002566009759903, + 0.04670390859246254, + 0.030499225482344627, + 0.045711830258369446, + 0.04597185179591179, + 0.09616541862487793, + 0.06637831032276154, + 0.09958834946155548, + 0.08341126888990402 + ], + "credit_map": [ + 0.0790243074297905, + 0.049653053283691406, + 0.06258092075586319, + 0.056902650743722916, + 0.04457999765872955, + 0.04483339190483093, + 0.02841131016612053, + 0.04955858737230301, + 0.015152051113545895, + 0.010295163840055466, + 0.009274807758629322, + 0.008069531060755253, + 0.005572685040533543, + 0.006236272398382425, + 0.0044833580031991005, + 0.010941796004772186, + 0.007709542755037546, + 0.00587631156668067, + 0.005384748335927725, + 0.0040090507827699184, + 0.004763117525726557, + 0.006301103625446558, + 0.004048020113259554, + 0.008612018078565598, + 0.0035360867623239756, + 0.0027359058149158955, + 0.009385963901877403, + 0.0012597710592672229, + 0.0016142763197422028, + 0.002862208290025592, + 0.0008524292497895658, + 0.008252587169408798, + 0.005827320273965597, + 0.005150452256202698, + 0.006125875748693943, + 0.0037733428180217743, + 0.00729608116671443, + 0.003213537158444524, + 0.00240280176512897, + 0.011922426521778107, + 0.007206425536423922, + 0.005074703600257635, + 0.008006214164197445, + 0.003095963504165411, + 0.006395889911800623, + 0.003163035726174712, + 0.0029397897887974977, + 0.010089835152029991, + 0.010818976908922195, + 0.007006385829299688, + 0.017722297459840775, + 0.012586056254804134, + 0.010847578756511211, + 0.01103059109300375, + 0.01354064792394638, + 0.012612897902727127, + 0.0074372501112520695, + 0.008685973472893238, + 0.016764935106039047, + 0.007354138419032097, + 0.007130361162126064, + 0.0038179296534508467, + 0.005656021647155285, + 0.009531702846288681, + 0.02926471456885338, + 0.005420872941613197, + 0.037364184856414795, + 0.004872722551226616, + 0.004447412211447954, + 0.006260187830775976, + 0.0040051257237792015, + 0.007953125983476639, + 0.012854749336838722, + 0.02198965847492218, + 0.007276736665517092, + 0.010959942825138569, + 0.007296730764210224, + 0.012385413981974125, + 0.005880897864699364, + 0.004767144098877907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.002489326987415552 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978152547031641 + ], + "opposite_logprob_tokens": [ + -4.634781360626221 + ], + "logprob_reward": -0.00978152547031641, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978152547031641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10221441835165024, + 0.08576201647520065, + 0.06176239997148514, + 0.1568351536989212, + 0.02070169895887375, + 0.12786448001861572, + 0.03671684488654137, + 0.07478146255016327, + 0.25481459498405457, + 0.07854682952165604 + ], + "credit_map": [ + 0.016636239364743233, + 0.012475019320845604, + 0.014070975594222546, + 0.01333205122500658, + 0.01490398496389389, + 0.011267206631600857, + 0.009495319798588753, + 0.010033626109361649, + 0.017305370420217514, + 0.012067158706486225, + 0.010522980242967606, + 0.009195382706820965, + 0.013438637368381023, + 0.0071157244965434074, + 0.006967280060052872, + 0.009149491786956787, + 0.00804379303008318, + 0.01001710630953312, + 0.007175080478191376, + 0.008179137483239174, + 0.0062962789088487625, + 0.01038711704313755, + 0.004216975066810846, + 0.007446919567883015, + 0.01328517496585846, + 0.003951866179704666, + 0.0627671480178833, + 0.003733567427843809, + 0.0030524253379553556, + 0.017822006717324257, + 0.002291815122589469, + 0.04993116483092308, + 0.002351585775613785, + 0.0037592221051454544, + 0.0037378286942839622, + 0.002130576642230153, + 0.0025654914788901806, + 0.001792686409316957, + 0.0013988317223265767, + 0.0029654770623892546, + 0.03209178149700165, + 0.013688774779438972, + 0.01763894036412239, + 0.008859831839799881, + 0.011690465733408928, + 0.008236944675445557, + 0.006268639117479324, + 0.029389111325144768, + 0.003257492557168007, + 0.0044076936319470406, + 0.006133949384093285, + 0.0036156463902443647, + 0.005694482941180468, + 0.005315255373716354, + 0.002707844600081444, + 0.005584483500570059, + 0.008869427256286144, + 0.010087096132338047, + 0.0175960510969162, + 0.008263288997113705, + 0.008819419890642166, + 0.005160414148122072, + 0.006953050848096609, + 0.009032717905938625, + 0.04950734227895737, + 0.029604516923427582, + 0.0767156183719635, + 0.021394779905676842, + 0.013852886855602264, + 0.02075193263590336, + 0.018621914088726044, + 0.024365633726119995, + 0.010670872405171394, + 0.020609473809599876, + 0.008894369937479496, + 0.009576435200870037, + 0.008624207228422165, + 0.009519493207335472, + 0.005886699538677931, + 0.004765291232615709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -0.00978152547031641 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615511115640402 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.003615511115640402, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615511115640402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18754643201828003, + 0.15747272968292236, + 0.043647509068250656, + 0.09866450726985931, + 0.05415860563516617, + 0.06731469184160233, + 0.07793319225311279, + 0.07127070426940918, + 0.15861612558364868, + 0.08337551355361938 + ], + "credit_map": [ + 0.03430735319852829, + 0.023297034204006195, + 0.02608594298362732, + 0.020953256636857986, + 0.025190114974975586, + 0.01966867968440056, + 0.01471460610628128, + 0.023329444229602814, + 0.027689779177308083, + 0.02444455586373806, + 0.01818029023706913, + 0.019922038540244102, + 0.024368515238165855, + 0.012446824461221695, + 0.010410211980342865, + 0.02001051977276802, + 0.005703669972717762, + 0.005279181525111198, + 0.005782196763902903, + 0.004362633917480707, + 0.003926670644432306, + 0.004211087245494127, + 0.0035092956386506557, + 0.01087277289479971, + 0.0166286863386631, + 0.007279000245034695, + 0.02655210718512535, + 0.004215969704091549, + 0.005548997316509485, + 0.017360812053084373, + 0.003266045358031988, + 0.01781289465725422, + 0.008026600815355778, + 0.005640511400997639, + 0.00903611071407795, + 0.005435463041067123, + 0.00838738027960062, + 0.004210293292999268, + 0.003951037768274546, + 0.009471212513744831, + 0.014061592519283295, + 0.007535888347774744, + 0.010824870318174362, + 0.005646406672894955, + 0.009782860986888409, + 0.006200741045176983, + 0.004529292695224285, + 0.00873304158449173, + 0.00974513404071331, + 0.006929009221494198, + 0.012733044102787971, + 0.011594279669225216, + 0.009170785546302795, + 0.009585964493453503, + 0.00691878842189908, + 0.011256188154220581, + 0.008494873531162739, + 0.007022051140666008, + 0.011365463957190514, + 0.007876988500356674, + 0.00692372489720583, + 0.00995179358869791, + 0.0076586659997701645, + 0.011977146379649639, + 0.040733207017183304, + 0.009859276935458183, + 0.04738835245370865, + 0.008610404096543789, + 0.009066767990589142, + 0.011535080149769783, + 0.0115220220759511, + 0.01990101858973503, + 0.010972284711897373, + 0.01032751053571701, + 0.009084505029022694, + 0.005352500826120377, + 0.019954292103648186, + 0.013687960803508759, + 0.00796229112893343, + 0.006034164689481258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.003615511115640402 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle shot rapidly tracks a massive black bull as it explosively charges forward across a sunny plain. Its dark hide gleams in the harsh, direct sunlight, its dense, powerful muscles ripple, its entire form rendered in a slight motion blur to emphasize the sheer speed of its charge. Thick, white steam erupts violently from its flared nostrils, a stark contrast to its black form. With its head lowered, its massive, sleek horns aimed forward, its hooves dig into the dry, ochre earth, kicking up a massive cloud of dust and small pebbles that tumble and swirl in its wake. The scene is filled with primal power, the bull's dark, muscular form contrasting sharply with the bright, chaotic explosions of dust, creating an intense and visceral atmosphere of untamed, explosive violence.", + "seed": 28163, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_7/index_8960_seed_28163_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0034641455160453914, + "vlm_logprob": -0.0034641455160453914, + "vlm_yes_no_margin": 6.325, + "vlm_token_credit": -0.0034641455160453914, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005552939139306545 + ], + "opposite_logprob_tokens": [ + -8.130553245544434 + ], + "logprob_reward": -0.005552939139306545, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.005552939139306545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11522876471281052, + 0.04827563837170601, + 0.03519817069172859, + 0.04228862375020981, + 0.4863399565219879, + 0.10242989659309387, + 0.05328067019581795, + 0.031592030078172684, + 0.06490626931190491, + 0.020459961146116257 + ], + "credit_map": [ + 0.008514853194355965, + 0.0073905508033931255, + 0.009538745507597923, + 0.006796560250222683, + 0.008816352114081383, + 0.0231948122382164, + 0.01832675375044346, + 0.03265013545751572, + 0.003830754430964589, + 0.005483898334205151, + 0.005777382291853428, + 0.005330590531229973, + 0.006882425397634506, + 0.00681894738227129, + 0.006099268794059753, + 0.00805237703025341, + 0.0036051415372639894, + 0.0052117337472736835, + 0.00392887694761157, + 0.0020404760725796223, + 0.004416666459292173, + 0.005663295742124319, + 0.004731741268187761, + 0.00560024194419384, + 0.003948170691728592, + 0.005437822081148624, + 0.0035414425656199455, + 0.005035726819187403, + 0.005287306383252144, + 0.010205945931375027, + 0.003821705700829625, + 0.0050105080008506775, + 0.02416488528251648, + 0.06460033357143402, + 0.02722725085914135, + 0.013562140055000782, + 0.1353735327720642, + 0.06356991082429886, + 0.025576461106538773, + 0.13226544857025146, + 0.009182998910546303, + 0.021489888429641724, + 0.005168023984879255, + 0.004683121107518673, + 0.015783753246068954, + 0.012569043785333633, + 0.008384566754102707, + 0.02516850456595421, + 0.005056056194007397, + 0.007167911622673273, + 0.005060543771833181, + 0.005114783998578787, + 0.0075350129045546055, + 0.006804291624575853, + 0.00709906592965126, + 0.009443007409572601, + 0.002257147105410695, + 0.0034535436425358057, + 0.002371364738792181, + 0.002965832594782114, + 0.0036483590956777334, + 0.0077657550573349, + 0.0036982714664191008, + 0.005431756842881441, + 0.008179501630365849, + 0.009155943058431149, + 0.009132257662713528, + 0.007344698999077082, + 0.00844890158623457, + 0.00852784514427185, + 0.007141447626054287, + 0.006975677330046892, + 0.00159757724031806, + 0.0015459565911442041, + 0.0027418590616434813, + 0.0018026448087766767, + 0.002573342528194189, + 0.003480927087366581, + 0.0031900855246931314, + 0.0035275693517178297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.005552939139306545 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00362489465624094 + ], + "opposite_logprob_tokens": [ + -5.62862491607666 + ], + "logprob_reward": -0.00362489465624094, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00362489465624094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14339379966259003, + 0.025969192385673523, + 0.03542175516486168, + 0.037458498030900955, + 0.5278162360191345, + 0.04509172961115837, + 0.044550515711307526, + 0.05332463979721069, + 0.018569309264421463, + 0.06840429455041885 + ], + "credit_map": [ + 0.012420564889907837, + 0.014533924870193005, + 0.016912013292312622, + 0.010694184340536594, + 0.0095464326441288, + 0.02364557795226574, + 0.02762000449001789, + 0.028021102771162987, + 0.001861737808212638, + 0.0025239423848688602, + 0.00379798817448318, + 0.00223477347753942, + 0.0033528681378811598, + 0.003232549875974655, + 0.004056836944073439, + 0.004908495582640171, + 0.0031507627572864294, + 0.005404395051300526, + 0.0035795988515019417, + 0.0027895707171410322, + 0.003321496769785881, + 0.006338464096188545, + 0.004249054938554764, + 0.006588412448763847, + 0.0029374985024333, + 0.004805105738341808, + 0.0046741957776248455, + 0.00506106810644269, + 0.003992751240730286, + 0.007641386240720749, + 0.0034399928990751505, + 0.004906500689685345, + 0.024447482079267502, + 0.06694266945123672, + 0.025223711505532265, + 0.009796873666346073, + 0.1464710831642151, + 0.0744052454829216, + 0.02594710700213909, + 0.15458205342292786, + 0.004106810316443443, + 0.007716902997344732, + 0.0031432867981493473, + 0.0029102391563355923, + 0.005479860585182905, + 0.007237858604639769, + 0.0038759915623813868, + 0.010620777495205402, + 0.005561482161283493, + 0.006306836381554604, + 0.00529244402423501, + 0.0044593787752091885, + 0.005635334178805351, + 0.0070424508303403854, + 0.00408095633611083, + 0.006171632558107376, + 0.0030566465575248003, + 0.004869203083217144, + 0.004225610289722681, + 0.0044652852229774, + 0.006583991926163435, + 0.013815665617585182, + 0.006880589295178652, + 0.009427645243704319, + 0.0020414977334439754, + 0.0023741191253066063, + 0.002668044064193964, + 0.0019632030744105577, + 0.0020768020767718554, + 0.002456812420859933, + 0.0029596579261124134, + 0.002029173308983445, + 0.00508846202865243, + 0.005208990070968866, + 0.009319698438048363, + 0.005174449644982815, + 0.008900140412151814, + 0.011782975867390633, + 0.010532627813518047, + 0.01239694468677044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.00362489465624094 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001712523982860148 + ], + "opposite_logprob_tokens": [ + -6.376712322235107 + ], + "logprob_reward": -0.001712523982860148, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001712523982860148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.143373042345047, + 0.06249556690454483, + 0.046179238706827164, + 0.04178354889154434, + 0.3559630513191223, + 0.04525614157319069, + 0.11346248537302017, + 0.03664889931678772, + 0.13421466946601868, + 0.02062331512570381 + ], + "credit_map": [ + 0.019832322373986244, + 0.01833834871649742, + 0.011556018143892288, + 0.012618829496204853, + 0.013829482719302177, + 0.02240748144686222, + 0.024031251668930054, + 0.020759308710694313, + 0.004839801695197821, + 0.005136267747730017, + 0.012008773162961006, + 0.007685698103159666, + 0.008770773187279701, + 0.010031137615442276, + 0.007720085326582193, + 0.006303034257143736, + 0.0036146726924926043, + 0.004756282549351454, + 0.003622285556048155, + 0.0024669927079230547, + 0.004493101499974728, + 0.010323098860681057, + 0.0058601065538823605, + 0.011042699217796326, + 0.003142768982797861, + 0.005311863962560892, + 0.004718304146081209, + 0.004529564641416073, + 0.003761912928894162, + 0.009803731925785542, + 0.005279989913105965, + 0.005235414020717144, + 0.016735272482037544, + 0.04526902735233307, + 0.01749982312321663, + 0.007358197588473558, + 0.09879427403211594, + 0.0495123527944088, + 0.017832191661000252, + 0.10296192765235901, + 0.0038349044043570757, + 0.009497174061834812, + 0.0036549584474414587, + 0.0022001962643116713, + 0.006633755285292864, + 0.005615720059722662, + 0.0034695968497544527, + 0.010349838063120842, + 0.017532959580421448, + 0.015985626727342606, + 0.014246998354792595, + 0.009735449217259884, + 0.011838122271001339, + 0.01683623529970646, + 0.011133750900626183, + 0.016153346747159958, + 0.0026037339121103287, + 0.004328913986682892, + 0.0023656776174902916, + 0.0027558128349483013, + 0.003971803467720747, + 0.0067517408169806, + 0.005161180626600981, + 0.008710037916898727, + 0.009989691898226738, + 0.011932408437132835, + 0.011058816686272621, + 0.011903444305062294, + 0.0112595921382308, + 0.020364515483379364, + 0.029825283214449883, + 0.02788093313574791, + 0.002195856301113963, + 0.0016195521457120776, + 0.0030932363588362932, + 0.0014334589941427112, + 0.00207695085555315, + 0.0034059016034007072, + 0.0033184834755957127, + 0.003479874925687909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.001712523982860148 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028165215626358986 + ], + "opposite_logprob_tokens": [ + -5.877816677093506 + ], + "logprob_reward": -0.0028165215626358986, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028165215626358986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.115611232817173, + 0.011892194859683514, + 0.016716763377189636, + 0.01950395666062832, + 0.6236611008644104, + 0.056664153933525085, + 0.11058319360017776, + 0.02112523652613163, + 0.015248900279402733, + 0.008993229828774929 + ], + "credit_map": [ + 0.007273107301443815, + 0.0089685944840312, + 0.008282041177153587, + 0.005840584635734558, + 0.007343219593167305, + 0.02394205704331398, + 0.018931403756141663, + 0.035030219703912735, + 0.0010128041030839086, + 0.0011030687019228935, + 0.001603208133019507, + 0.0012351832119747996, + 0.0018595235887914896, + 0.0016591758467257023, + 0.001808873494155705, + 0.0016103582456707954, + 0.0012311753816902637, + 0.0017487361328676343, + 0.0017978824907913804, + 0.0011352060828357935, + 0.0015909578651189804, + 0.003901915391907096, + 0.002875275444239378, + 0.00243561458773911, + 0.0024628329556435347, + 0.0025795612018555403, + 0.0026085611898452044, + 0.0031103023793548346, + 0.002278565429151058, + 0.002408245112746954, + 0.0018247906118631363, + 0.002231098245829344, + 0.028322461992502213, + 0.07707473635673523, + 0.028645560145378113, + 0.010815066285431385, + 0.17080967128276825, + 0.09068495780229568, + 0.03053215518593788, + 0.1867765486240387, + 0.005279067438095808, + 0.011364425532519817, + 0.003778740530833602, + 0.0030935381073504686, + 0.008417598903179169, + 0.009446351788938046, + 0.00425339862704277, + 0.011031029745936394, + 0.01475647184997797, + 0.018876906484365463, + 0.010938305407762527, + 0.010276172310113907, + 0.018470723181962967, + 0.015553544275462627, + 0.010087202303111553, + 0.011623872444033623, + 0.0012196862371638417, + 0.002012112410739064, + 0.0012764664134010673, + 0.0012012068182229996, + 0.0017044594278559089, + 0.006464496720582247, + 0.002430870896205306, + 0.004815939348191023, + 0.001435384969227016, + 0.0018670405261218548, + 0.0018875051755458117, + 0.0016951735597103834, + 0.001337119610980153, + 0.002116992836818099, + 0.0025878162123262882, + 0.002321867737919092, + 0.0006459586438722908, + 0.0005867593572475016, + 0.0015682397643104196, + 0.0007759880973026156, + 0.0011133255902677774, + 0.0013931590365245938, + 0.0014421367086470127, + 0.0014676627470180392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.0028165215626358986 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive black bull with dense, powerful muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bull performing an explosive charge forward with its head lowered and horns aimed forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a sunny plain with dry, ochre earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible effects of thick white steam erupting from the bull's flared nostrils and a massive cloud of dust kicked up by its hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dramatic low-angle view with rapid camera tracking and slight motion blur on the bull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003613848239183426 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.003613848239183426, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613848239183426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050606612116098404, + 0.02222738228738308, + 0.031159432604908943, + 0.06027550622820854, + 0.589507520198822, + 0.06633028388023376, + 0.06185690686106682, + 0.06360753625631332, + 0.0176167543977499, + 0.036812059581279755 + ], + "credit_map": [ + 0.005406671203672886, + 0.005627841223031282, + 0.004198640119284391, + 0.0036488035693764687, + 0.004765674006193876, + 0.00838958378881216, + 0.0067812721244990826, + 0.011788129806518555, + 0.001554270857013762, + 0.0016558176139369607, + 0.0036299547646194696, + 0.0020844663958996534, + 0.003364419098943472, + 0.002800922840833664, + 0.003815403440967202, + 0.0033221275079995394, + 0.003142458153888583, + 0.0035478973295539618, + 0.0032332243863493204, + 0.002033020369708538, + 0.004058860242366791, + 0.007420471869409084, + 0.003074491862207651, + 0.004649010952562094, + 0.006608966272324324, + 0.00804070383310318, + 0.00667422404512763, + 0.005232558585703373, + 0.0063460823148489, + 0.01107795350253582, + 0.004721399862319231, + 0.011573614552617073, + 0.027030959725379944, + 0.07478276640176773, + 0.02876734361052513, + 0.011640791781246662, + 0.16365794837474823, + 0.08243360370397568, + 0.029349220916628838, + 0.17184486985206604, + 0.007735913619399071, + 0.014098694548010826, + 0.006434332113713026, + 0.004607736598700285, + 0.010348779149353504, + 0.006736386101692915, + 0.005502232350409031, + 0.010866214521229267, + 0.0055573731660842896, + 0.009250102564692497, + 0.0049539245665073395, + 0.007057371083647013, + 0.009519624523818493, + 0.008022613823413849, + 0.00720046553760767, + 0.01029543112963438, + 0.003936303313821554, + 0.006977028213441372, + 0.00430439505726099, + 0.0051129404455423355, + 0.005948052275925875, + 0.01864931918680668, + 0.0068276384845376015, + 0.011851862072944641, + 0.002417481504380703, + 0.0020913416519761086, + 0.0025445702485740185, + 0.0030599639285355806, + 0.0022205775603652, + 0.001953848171979189, + 0.00192393793258816, + 0.0014050346799194813, + 0.0031116893514990807, + 0.0031621898524463177, + 0.004738400690257549, + 0.002991075161844492, + 0.004843817558139563, + 0.0062024290673434734, + 0.004943069536238909, + 0.006819389294832945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.003613848239183426 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_8/index_tensor([9815, 9815])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_8/index_tensor([9815, 9815])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fd02247bc659cc7674d3e5d168876e2e0b84ce38 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_8/index_tensor([9815, 9815])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Viewed from a first-person perspective, the camera executes a slow, continuous, and tranquil forward drift beneath the crystal-clear surface of the Great Barrier Reef. Rendered in a vibrant, impressionistic style, the entire scene transforms the underwater world into a living, breathing painting. Gentle ripples on the water above act like a dynamic lens, fragmenting the brilliant sunlight into a constant dance of shimmering, corrosive patterns that punctuate the seascape below. Coral structures lose their clear biological definition, dissolving into a breathtaking mosaic of textured colors and shifting light; vast fields of staghorn and brain coral appear as thick, visible splotches and swirls of soft pinks, deep violets, and warm oranges. Schools of fish are depicted not as individual creatures but as fleeting blurs and swift, ephemeral streaks of pure pigment. Flashes of electric blue and canary yellow dart across the turquoise water like the quick, deliberate strokes of a painter's brush. The atmosphere is one of profound tranquility and dreamlike beauty, where the boundaries between light, water, and life blur into a beautiful, ever-shifting synthesis of movement and color.", + "seed": 28844, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_8/index_9815_seed_28844_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08653108328580857, + "vlm_logprob": -0.08653108328580857, + "vlm_yes_no_margin": 4.524999952316284, + "vlm_token_credit": -0.08653108328580857, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02921740710735321 + ], + "opposite_logprob_tokens": [ + -4.02921724319458 + ], + "logprob_reward": -0.02921740710735321, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.02921740710735321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10858415067195892, + 0.4202979803085327, + 0.1126561164855957, + 0.06473235785961151, + 0.011869216337800026, + 0.013184952549636364, + 0.015121908858418465, + 0.06790687143802643, + 0.12751297652721405, + 0.058133501559495926 + ], + "credit_map": [ + 0.021130619570612907, + 0.009407536126673222, + 0.015559816733002663, + 0.007697432301938534, + 0.01975402422249317, + 0.02111821062862873, + 0.006866424344480038, + 0.007050080690532923, + 0.09169892966747284, + 0.17194077372550964, + 0.010167124681174755, + 0.007876871153712273, + 0.026685774326324463, + 0.02353532612323761, + 0.027746835723519325, + 0.060646314173936844, + 0.02733542025089264, + 0.0172621700912714, + 0.009556795470416546, + 0.005451452452689409, + 0.012325611896812916, + 0.015709616243839264, + 0.010200703516602516, + 0.014814338646829128, + 0.019542386755347252, + 0.021038299426436424, + 0.00254668015986681, + 0.0017183261225000024, + 0.01028505526483059, + 0.003677068278193474, + 0.002661711536347866, + 0.0032628215849399567, + 0.001866727601736784, + 0.0020766884554177523, + 0.0009282463579438627, + 0.0007942152442410588, + 0.0013037382159382105, + 0.0019370458321645856, + 0.0015985985519364476, + 0.0013639547396451235, + 0.0018902049632743, + 0.0016463174251839519, + 0.0015254743630066514, + 0.001547585241496563, + 0.0024193134158849716, + 0.001797272707335651, + 0.0012559868628159165, + 0.0011027967557311058, + 0.0021426337771117687, + 0.0020415950566530228, + 0.0011678338050842285, + 0.0016258344985544682, + 0.0034826681949198246, + 0.0019989125430583954, + 0.0012386433081701398, + 0.0014237858122214675, + 0.018236691132187843, + 0.0197826586663723, + 0.003172619268298149, + 0.0030926542822271585, + 0.008021136745810509, + 0.008119752630591393, + 0.0028261663392186165, + 0.004655187018215656, + 0.04663922265172005, + 0.04959296062588692, + 0.00346966739743948, + 0.003845488652586937, + 0.007100556045770645, + 0.008219784125685692, + 0.004478297661989927, + 0.004166996572166681, + 0.017289238050580025, + 0.020464876666665077, + 0.002878420287743211, + 0.003463854780420661, + 0.00399866234511137, + 0.004349210765212774, + 0.0034697605296969414, + 0.002219475107267499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.02921740710735321 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.387020081281662 + ], + "opposite_logprob_tokens": [ + -1.1370201110839844 + ], + "logprob_reward": -0.387020081281662, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.387020081281662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03552156686782837, + 0.5417529940605164, + 0.10019230097532272, + 0.15803900361061096, + 0.00932458695024252, + 0.007039420306682587, + 0.037865180522203445, + 0.023000290617346764, + 0.06254404783248901, + 0.024720553308725357 + ], + "credit_map": [ + 0.004751694854348898, + 0.003850943874567747, + 0.004947556182742119, + 0.003857301315292716, + 0.006185832899063826, + 0.006335940212011337, + 0.002771970583125949, + 0.002820331137627363, + 0.12205076962709427, + 0.22695931792259216, + 0.01048695296049118, + 0.00984224583953619, + 0.03479759767651558, + 0.030554259195923805, + 0.03269767388701439, + 0.07436425983905792, + 0.027780836448073387, + 0.014021250419318676, + 0.005272503010928631, + 0.003256472060456872, + 0.009660485200583935, + 0.010462779551744461, + 0.01095405500382185, + 0.018783917650580406, + 0.03960932791233063, + 0.049084827303886414, + 0.006489295978099108, + 0.0064535182900726795, + 0.02330835722386837, + 0.010701891034841537, + 0.009098105132579803, + 0.013293697498738766, + 0.0019696808885782957, + 0.00157542887609452, + 0.001003689132630825, + 0.0008086197194643319, + 0.0007822794141247869, + 0.0013908924302086234, + 0.0009789043106138706, + 0.0008150924113579094, + 0.0010470467386767268, + 0.001020908122882247, + 0.0007259869598783553, + 0.00082913733785972, + 0.0012353776255622506, + 0.0006593148573301733, + 0.0008838794310577214, + 0.0006377696408890188, + 0.004082051571458578, + 0.004823033697903156, + 0.0034143950324505568, + 0.004092530347406864, + 0.006326098460704088, + 0.004708191845566034, + 0.005164274014532566, + 0.005254607181996107, + 0.007274278439581394, + 0.007980666123330593, + 0.0009876209078356624, + 0.0010525129036977887, + 0.0018367754528298974, + 0.0016422455664724112, + 0.0010858396999537945, + 0.001140353037044406, + 0.023010792210698128, + 0.02650035172700882, + 0.0013101533986628056, + 0.0016166139394044876, + 0.0038237276021391153, + 0.003123797010630369, + 0.0018851383356377482, + 0.0012734764022752643, + 0.007153150625526905, + 0.008250519633293152, + 0.0010685373563319445, + 0.0016630565514788032, + 0.0020976183004677296, + 0.0017429395811632276, + 0.0016811807872727513, + 0.001063553150743246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.387020081281662 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010461101774126291 + ], + "opposite_logprob_tokens": [ + -6.876046180725098 + ], + "logprob_reward": -0.0010461101774126291, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010461101774126291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046400558203458786, + 0.4324740171432495, + 0.047407299280166626, + 0.05290750414133072, + 0.008880003355443478, + 0.009123622439801693, + 0.0345827117562294, + 0.02588840015232563, + 0.31911441683769226, + 0.023221449926495552 + ], + "credit_map": [ + 0.00676929485052824, + 0.004145843908190727, + 0.007096089422702789, + 0.004322831518948078, + 0.00707652885466814, + 0.009513664990663528, + 0.004091172944754362, + 0.003385131945833564, + 0.09517956525087357, + 0.18004803359508514, + 0.008554491214454174, + 0.007055389694869518, + 0.028145018965005875, + 0.023783328011631966, + 0.02705990895628929, + 0.06264828145503998, + 0.007050927262753248, + 0.005957817658782005, + 0.004552945494651794, + 0.00325483875349164, + 0.008424244821071625, + 0.00854339636862278, + 0.0052137672901153564, + 0.00440936116501689, + 0.015784800052642822, + 0.016483966261148453, + 0.0023327572271227837, + 0.0017653771210461855, + 0.007628639228641987, + 0.0042878128588199615, + 0.0022442175541073084, + 0.002379931742325425, + 0.0013236147351562977, + 0.0012616883032023907, + 0.0015072051901370287, + 0.0009966911748051643, + 0.0010255518136546016, + 0.0009670405997894704, + 0.000983854872174561, + 0.000814356782939285, + 0.001186348032206297, + 0.0011663793120533228, + 0.000989676220342517, + 0.0010649197502061725, + 0.0011570975184440613, + 0.0011489680036902428, + 0.0015491950325667858, + 0.0008610385120846331, + 0.004046259913593531, + 0.004782169591635466, + 0.0028784554451704025, + 0.0034395239781588316, + 0.005888518877327442, + 0.005485784728080034, + 0.004042652901262045, + 0.004019349813461304, + 0.004040219821035862, + 0.004891456104815006, + 0.0020327623933553696, + 0.002735869260504842, + 0.002863283036276698, + 0.0033816988579928875, + 0.0023769517429172993, + 0.0035661577712744474, + 0.12306741625070572, + 0.137017622590065, + 0.005608804523944855, + 0.00575812766328454, + 0.014004088938236237, + 0.020053258165717125, + 0.00912674330174923, + 0.004478325601667166, + 0.004236781038343906, + 0.005065594799816608, + 0.0016017660964280367, + 0.002240155590698123, + 0.002923316555097699, + 0.0022636321373283863, + 0.0030952501110732555, + 0.0017949527828022838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.0010461101774126291 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193889684975147 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014193889684975147, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193889684975147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04322412982583046, + 0.7463715076446533, + 0.02477410063147545, + 0.05644041672348976, + 0.01083622220903635, + 0.0092876385897398, + 0.010816098190844059, + 0.018527330830693245, + 0.07132072746753693, + 0.008401893079280853 + ], + "credit_map": [ + 0.005228933412581682, + 0.0038826316595077515, + 0.005838494747877121, + 0.003836150048300624, + 0.007536967284977436, + 0.010876629501581192, + 0.0028280364349484444, + 0.00319628044962883, + 0.1681199073791504, + 0.3128964304924011, + 0.013964547775685787, + 0.011808549985289574, + 0.04792921617627144, + 0.04112827032804489, + 0.045676399022340775, + 0.10484808683395386, + 0.0031843530014157295, + 0.0039299833588302135, + 0.0023892074823379517, + 0.002265287097543478, + 0.004756005015224218, + 0.003816469805315137, + 0.0026570444460958242, + 0.0017757475143298507, + 0.014442430809140205, + 0.017620759084820747, + 0.0021320655941963196, + 0.00195898674428463, + 0.008975102566182613, + 0.0038874768652021885, + 0.002990756183862686, + 0.004432835150510073, + 0.0018835397204384208, + 0.00186059158295393, + 0.0012322061229497194, + 0.001059234607964754, + 0.0010195060167461634, + 0.0013766268966719508, + 0.0013824458001181483, + 0.0010220699477940798, + 0.0013961497461423278, + 0.0017451867461204529, + 0.0009843299631029367, + 0.0006785138393752277, + 0.0015792930498719215, + 0.0010514607420191169, + 0.0010981099912896752, + 0.0007545932894572616, + 0.0017838254570960999, + 0.0017381072975695133, + 0.0009082061005756259, + 0.00126793107483536, + 0.0016540357610210776, + 0.0015534720150753856, + 0.0009891019435599446, + 0.0009214166202582419, + 0.0031510277185589075, + 0.002938777208328247, + 0.0014173579402267933, + 0.0016421015607193112, + 0.002593271667137742, + 0.003857267089188099, + 0.0011415814515203238, + 0.0017859438667073846, + 0.027499234303832054, + 0.03044303134083748, + 0.0013877495657652617, + 0.0013907814864069223, + 0.0030526481568813324, + 0.004467992577701807, + 0.001929659047164023, + 0.0011496184160932899, + 0.0012481645680963993, + 0.0010931296274065971, + 0.0005992674268782139, + 0.0008283103234134614, + 0.001114536658860743, + 0.001238372060470283, + 0.0012750299647450447, + 0.0010050819255411625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.014193889684975147 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001177928177639842 + ], + "opposite_logprob_tokens": [ + -6.751177787780762 + ], + "logprob_reward": -0.001177928177639842, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001177928177639842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062157124280929565, + 0.5451271533966064, + 0.03981049358844757, + 0.03984793275594711, + 0.018434124067425728, + 0.01159595139324665, + 0.042225200682878494, + 0.04694760963320732, + 0.1689470261335373, + 0.024907400831580162 + ], + "credit_map": [ + 0.010428376495838165, + 0.006310056429356337, + 0.01009341049939394, + 0.004521084949374199, + 0.01205222588032484, + 0.010794999077916145, + 0.003927427809685469, + 0.004029542673379183, + 0.12185736745595932, + 0.2275780588388443, + 0.010725934989750385, + 0.00903048925101757, + 0.03500143811106682, + 0.029690219089388847, + 0.0341099388897419, + 0.07713375240564346, + 0.007969324477016926, + 0.004770428407937288, + 0.002638375386595726, + 0.002029995433986187, + 0.0067602647468447685, + 0.00564039871096611, + 0.004184759687632322, + 0.005816948600113392, + 0.009620293974876404, + 0.010687540285289288, + 0.0029656458646059036, + 0.0017817113548517227, + 0.0066954512149095535, + 0.0034123039804399014, + 0.0019127356354147196, + 0.0027722513768821955, + 0.003919840790331364, + 0.0029101292602717876, + 0.0027730586007237434, + 0.001720030908472836, + 0.0017874246696010232, + 0.002229534089565277, + 0.0015564148779958487, + 0.0015376911032944918, + 0.0018789871828630567, + 0.001877751317806542, + 0.0011325468076393008, + 0.0010904503287747502, + 0.0017290478572249413, + 0.0013047392712906003, + 0.0016541715012863278, + 0.0009282574173994362, + 0.005401262082159519, + 0.006168256979435682, + 0.0031363251619040966, + 0.003237940836697817, + 0.008834575302898884, + 0.007064413744956255, + 0.003932856488972902, + 0.0044495705515146255, + 0.010921425186097622, + 0.01172117330133915, + 0.0029544064309448004, + 0.003036226611584425, + 0.005510225426405668, + 0.006858489941805601, + 0.0025509309489279985, + 0.0033947320189327, + 0.0653441771864891, + 0.07323025166988373, + 0.0029459393117576838, + 0.003435141872614622, + 0.007994793355464935, + 0.008060689084231853, + 0.005337510723620653, + 0.0025985136162489653, + 0.006892569828778505, + 0.00809738039970398, + 0.0009838355472311378, + 0.0019316405523568392, + 0.0024904811289161444, + 0.002154012443497777, + 0.0014321872731670737, + 0.000925293832551688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.001177928177639842 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Viewed from a first-person perspective, the camera executes a slow, continuous, and tranquil forward drift beneath the crystal-clear surface of the Great Barrier Reef. Rendered in a vibrant, impressionistic style, the entire scene transforms the underwater world into a living, breathing painting. Gentle ripples on the water above act like a dynamic lens, fragmenting the brilliant sunlight into a constant dance of shimmering, corrosive patterns that punctuate the seascape below. Coral structures lose their clear biological definition, dissolving into a breathtaking mosaic of textured colors and shifting light; vast fields of staghorn and brain coral appear as thick, visible splotches and swirls of soft pinks, deep violets, and warm oranges. Schools of fish are depicted not as individual creatures but as fleeting blurs and swift, ephemeral streaks of pure pigment. Flashes of electric blue and canary yellow dart across the turquoise water like the quick, deliberate strokes of a painter's brush. The atmosphere is one of profound tranquility and dreamlike beauty, where the boundaries between light, water, and life blur into a beautiful, ever-shifting synthesis of movement and color.", + "seed": 28845, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_8/index_9815_seed_28845_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09098543187137693, + "vlm_logprob": -0.09098543187137693, + "vlm_yes_no_margin": 4.225000011920929, + "vlm_token_credit": -0.09098543187137693, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09123622626066208 + ], + "opposite_logprob_tokens": [ + -2.591236114501953 + ], + "logprob_reward": -0.09123622626066208, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09123622626066208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6531416773796082, + 0.032741934061050415, + 0.017837828025221825, + 0.03220008313655853, + 0.021622978150844574, + 0.17807285487651825, + 0.017838355153799057, + 0.014744031243026257, + 0.026753613725304604, + 0.005046686623245478 + ], + "credit_map": [ + 0.0619768463075161, + 0.25278839468955994, + 0.030744845047593117, + 0.025225983932614326, + 0.16965125501155853, + 0.05730178579688072, + 0.030387692153453827, + 0.025064829736948013, + 0.002880997722968459, + 0.009261592291295528, + 0.0037543107755482197, + 0.0034552831202745438, + 0.005358213558793068, + 0.0033877091482281685, + 0.00294533371925354, + 0.001698488718830049, + 0.001388361444696784, + 0.00302877277135849, + 0.0020750609692186117, + 0.0018617365276440978, + 0.0033087763004004955, + 0.0018779498059302568, + 0.0026348086539655924, + 0.0016623595729470253, + 0.0032157807145267725, + 0.003535191062837839, + 0.0030183966737240553, + 0.0028596383053809404, + 0.0065503609366714954, + 0.0058410088531672955, + 0.0035402867943048477, + 0.0036394158378243446, + 0.002772918436676264, + 0.0027261001523584127, + 0.0031138015910983086, + 0.002845335053279996, + 0.0028898667078465223, + 0.002306436887010932, + 0.002749837702140212, + 0.0022186802234500647, + 0.0089793149381876, + 0.06459552049636841, + 0.004376543220132589, + 0.006308631505817175, + 0.07633540034294128, + 0.008884354494512081, + 0.005326656624674797, + 0.00326640997081995, + 0.0021160123869776726, + 0.003516649128869176, + 0.0017630946822464466, + 0.0021857700776308775, + 0.003104844596236944, + 0.0021260231733322144, + 0.001758239115588367, + 0.0012677196646109223, + 0.0013770340010523796, + 0.0012216672766953707, + 0.0016038522589951754, + 0.0027920533902943134, + 0.0018432429060339928, + 0.0015439020935446024, + 0.002898784354329109, + 0.0014634934486821294, + 0.0020889812149107456, + 0.0023798441980034113, + 0.004129047971218824, + 0.00454562995582819, + 0.0029089360032230616, + 0.0029009049758315086, + 0.005389566998928785, + 0.0024106998462229967, + 0.00047594791976734996, + 0.0006181956850923598, + 0.0007189179304987192, + 0.0006599011830985546, + 0.0007578622899018228, + 0.0005222606123425066, + 0.0007352794636972249, + 0.0005583205493167043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.09123622626066208 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3485862612724304 + ], + "opposite_logprob_tokens": [ + -1.2235863208770752 + ], + "logprob_reward": -0.3485862612724304, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.3485862612724304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8168850541114807, + 0.020649686455726624, + 0.00886925496160984, + 0.022233737632632256, + 0.014477918855845928, + 0.08296211808919907, + 0.008845667354762554, + 0.011327419430017471, + 0.009428032673895359, + 0.004321154672652483 + ], + "credit_map": [ + 0.06369081139564514, + 0.37413811683654785, + 0.011165949515998363, + 0.008497552014887333, + 0.2906397581100464, + 0.046219415962696075, + 0.010868600569665432, + 0.011664814315736294, + 0.0014795544557273388, + 0.006761528085917234, + 0.0015794544015079737, + 0.0015627688262611628, + 0.005215630866587162, + 0.0019404520280659199, + 0.001089358702301979, + 0.001020939671434462, + 0.0017886689165607095, + 0.0008733369177207351, + 0.000646416621748358, + 0.0009502422763034701, + 0.0022500876802951097, + 0.0008881977410055697, + 0.0006506455829367042, + 0.0008216587593778968, + 0.0018068685894832015, + 0.0025138012133538723, + 0.002752638189122081, + 0.0026644382160156965, + 0.0029769167304039, + 0.00208379328250885, + 0.00482224440202117, + 0.0026130350306630135, + 0.0018105613999068737, + 0.001582399127073586, + 0.0013484222581610084, + 0.0015376907540485263, + 0.002734886948019266, + 0.0016563981771469116, + 0.002384115243330598, + 0.001423444366082549, + 0.004362310748547316, + 0.029535146430134773, + 0.002109060762450099, + 0.002692870795726776, + 0.03525671735405922, + 0.004303985275328159, + 0.002803595270961523, + 0.0018984292401000857, + 0.0007378350128419697, + 0.0010657323291525245, + 0.0012589666293933988, + 0.0016077811596915126, + 0.0011068587191402912, + 0.0008515931549482048, + 0.0013308763736858964, + 0.0008860238012857735, + 0.0011878353543579578, + 0.001065758871845901, + 0.002100655809044838, + 0.0016005606157705188, + 0.001173678901977837, + 0.0013118607457727194, + 0.0016057976754382253, + 0.001281271455809474, + 0.0007484467932954431, + 0.0011804463574662805, + 0.0013057523174211383, + 0.0015536367427557707, + 0.0013198895612731576, + 0.00088421197142452, + 0.0015822320710867643, + 0.0008534174412488937, + 0.00033072216319851577, + 0.0004701555008068681, + 0.0007750894874334335, + 0.0007773026591166854, + 0.0005485204746946692, + 0.00045105675235390663, + 0.0005675334250554442, + 0.0004007741517852992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.3485862612724304 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001049920916557312 + ], + "opposite_logprob_tokens": [ + -6.876049995422363 + ], + "logprob_reward": -0.001049920916557312, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001049920916557312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7849618196487427, + 0.02145654708147049, + 0.01751059852540493, + 0.03253628686070442, + 0.031813669949769974, + 0.06825689226388931, + 0.011881266720592976, + 0.01183356437832117, + 0.011060706339776516, + 0.00868869386613369 + ], + "credit_map": [ + 0.06303897500038147, + 0.3575899004936218, + 0.012158365920186043, + 0.00984118226915598, + 0.2726646661758423, + 0.04672800749540329, + 0.01091908197849989, + 0.012021581642329693, + 0.002020658692345023, + 0.005796853452920914, + 0.0015508240321651101, + 0.0021992542315274477, + 0.0036929443012923002, + 0.002175974892452359, + 0.002082236809656024, + 0.0019378004362806678, + 0.003609933191910386, + 0.0025450510438531637, + 0.001215126016177237, + 0.0009219229104928672, + 0.005416805390268564, + 0.0016679995460435748, + 0.0011180117726325989, + 0.001015747431665659, + 0.0034829499199986458, + 0.0035588047467172146, + 0.002849574200809002, + 0.0029213365633040667, + 0.006918960250914097, + 0.004958955105394125, + 0.0038112797774374485, + 0.004034426063299179, + 0.003045331919565797, + 0.0036284092348068953, + 0.004304430913180113, + 0.004858714994043112, + 0.005444422364234924, + 0.0034131365828216076, + 0.004365750588476658, + 0.0027534733526408672, + 0.004776704590767622, + 0.018919266760349274, + 0.003561954712495208, + 0.003482294734567404, + 0.025307632982730865, + 0.004102595150470734, + 0.00399349071085453, + 0.0041129570454359055, + 0.001168124726973474, + 0.002131089800968766, + 0.0012265132972970605, + 0.001893007312901318, + 0.0020281989127397537, + 0.0010947758564725518, + 0.0013509959680959582, + 0.0009885611943900585, + 0.0010902326321229339, + 0.0012059552827849984, + 0.0018941713497042656, + 0.0015198953915387392, + 0.001679042587056756, + 0.0016097304178401828, + 0.0016476346645504236, + 0.0011869024019688368, + 0.0010833607520908117, + 0.001191705116070807, + 0.001801293925382197, + 0.0013856810983270407, + 0.0017758558969944715, + 0.0012696126941591501, + 0.0016228273743763566, + 0.0009303691913373768, + 0.000683165155351162, + 0.0010293582454323769, + 0.0011739860055968165, + 0.0014177921693772078, + 0.001091461512260139, + 0.0007477539475075901, + 0.001464779139496386, + 0.0010803969344124198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.001049920916557312 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254265196621418 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.01254265196621418, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254265196621418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7505300641059875, + 0.019051194190979004, + 0.032611310482025146, + 0.011847301386296749, + 0.03019496239721775, + 0.10994352400302887, + 0.011566554196178913, + 0.011046133935451508, + 0.015066922642290592, + 0.008142012171447277 + ], + "credit_map": [ + 0.06421739608049393, + 0.324544757604599, + 0.017140226438641548, + 0.01640772633254528, + 0.2372816503047943, + 0.051924705505371094, + 0.016648119315505028, + 0.022365473210811615, + 0.003638630034402013, + 0.0043649738654494286, + 0.001787833054549992, + 0.001311398926191032, + 0.0016735369572415948, + 0.0028559642378240824, + 0.001931376988068223, + 0.0014874801272526383, + 0.005253069568425417, + 0.0038910710718482733, + 0.003247588174417615, + 0.0029504660051316023, + 0.009103766642510891, + 0.002886549336835742, + 0.0037762410938739777, + 0.0015025613829493523, + 0.0012422050349414349, + 0.0017593035008758307, + 0.001004306017421186, + 0.0011121967108920217, + 0.0030762972310185432, + 0.0015076928539201617, + 0.0012283643009141088, + 0.0009169357945211232, + 0.002976406831294298, + 0.0027150572277605534, + 0.004327018745243549, + 0.005159228108823299, + 0.0047722370363771915, + 0.0032882734667509794, + 0.004216787405312061, + 0.002739951014518738, + 0.006076496094465256, + 0.0386398509144783, + 0.003409851109609008, + 0.0038795792497694492, + 0.046764422208070755, + 0.005765083711594343, + 0.0032342581544071436, + 0.0021739855874329805, + 0.0011977334506809711, + 0.0018816464580595493, + 0.0014312155544757843, + 0.0015639831544831395, + 0.00165378802921623, + 0.0011612746166065335, + 0.0015665729297325015, + 0.0011103394208475947, + 0.0009183729416690767, + 0.0012666962575167418, + 0.0013236983213573694, + 0.0017250183736905456, + 0.001227544853463769, + 0.0010515268659219146, + 0.0015460598515346646, + 0.0019872162956744432, + 0.0010960459476336837, + 0.0017786035314202309, + 0.0024983163457363844, + 0.0022529384586960077, + 0.0023210514336824417, + 0.001555100898258388, + 0.0019229124300181866, + 0.0016419535968452692, + 0.0007465563248842955, + 0.0008730336558073759, + 0.001227478845976293, + 0.0010318882996216416, + 0.0012852228246629238, + 0.0008433760376647115, + 0.0009897256968542933, + 0.0011447301367297769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.01254265196621418 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015120989410206676 + ], + "opposite_logprob_tokens": [ + -6.501512050628662 + ], + "logprob_reward": -0.0015120989410206676, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015120989410206676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6093429327011108, + 0.053149785846471786, + 0.044940367341041565, + 0.03430077061057091, + 0.019334018230438232, + 0.16979241371154785, + 0.015463807620108128, + 0.013490037061274052, + 0.03095712885260582, + 0.009228763170540333 + ], + "credit_map": [ + 0.04580816999077797, + 0.275622695684433, + 0.00932697020471096, + 0.007619103882461786, + 0.2181335836648941, + 0.03292057290673256, + 0.01080166082829237, + 0.00911016296595335, + 0.005326997488737106, + 0.01635173335671425, + 0.004870567005127668, + 0.00509848166257143, + 0.007951128296554089, + 0.005724928341805935, + 0.005109018180519342, + 0.0027169310487806797, + 0.01143510453402996, + 0.005771574564278126, + 0.0012471979716792703, + 0.0011485650902613997, + 0.018024956807494164, + 0.00464975880458951, + 0.0014549311017617583, + 0.0012082753237336874, + 0.003895403817296028, + 0.004209624603390694, + 0.0024199532344937325, + 0.002265315502882004, + 0.009493795223534107, + 0.00505334697663784, + 0.0032005817629396915, + 0.0037627459969371557, + 0.0023087395820766687, + 0.0023992147762328386, + 0.002965371124446392, + 0.0022678410168737173, + 0.0035233490634709597, + 0.0024619947653263807, + 0.0019439634634181857, + 0.001463544205762446, + 0.009662671014666557, + 0.05960517376661301, + 0.003781844163313508, + 0.0058338032104074955, + 0.0721709355711937, + 0.008777743205428123, + 0.00622726883739233, + 0.0037329529877752066, + 0.001480243168771267, + 0.002469839761033654, + 0.002130210166797042, + 0.003320131218060851, + 0.001506182481534779, + 0.0013154370244592428, + 0.0018780408427119255, + 0.0013637224910780787, + 0.000992600223980844, + 0.0013240913394838572, + 0.002183333970606327, + 0.002219397109001875, + 0.0015165965305641294, + 0.0012506353668868542, + 0.0023529210593551397, + 0.0016504605300724506, + 0.002479005139321089, + 0.003312088083475828, + 0.0038696439005434513, + 0.0047882553189992905, + 0.004231538623571396, + 0.0033887578174471855, + 0.0061964658088982105, + 0.0026913725305348635, + 0.0008758768672123551, + 0.0012150828260928392, + 0.0012322753900662065, + 0.000947757565882057, + 0.001551171881146729, + 0.001038393471390009, + 0.001051658415235579, + 0.0013165472773835063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0015120989410206676 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Viewed from a first-person perspective, the camera executes a slow, continuous, and tranquil forward drift beneath the crystal-clear surface of the Great Barrier Reef. Rendered in a vibrant, impressionistic style, the entire scene transforms the underwater world into a living, breathing painting. Gentle ripples on the water above act like a dynamic lens, fragmenting the brilliant sunlight into a constant dance of shimmering, corrosive patterns that punctuate the seascape below. Coral structures lose their clear biological definition, dissolving into a breathtaking mosaic of textured colors and shifting light; vast fields of staghorn and brain coral appear as thick, visible splotches and swirls of soft pinks, deep violets, and warm oranges. Schools of fish are depicted not as individual creatures but as fleeting blurs and swift, ephemeral streaks of pure pigment. Flashes of electric blue and canary yellow dart across the turquoise water like the quick, deliberate strokes of a painter's brush. The atmosphere is one of profound tranquility and dreamlike beauty, where the boundaries between light, water, and life blur into a beautiful, ever-shifting synthesis of movement and color.", + "seed": 28846, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_8/index_9815_seed_28846_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07270482915919274, + "vlm_logprob": -0.07270482915919274, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.07270482915919274, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05893462896347046 + ], + "opposite_logprob_tokens": [ + -3.1839346885681152 + ], + "logprob_reward": -0.05893462896347046, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05893462896347046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11933691799640656, + 0.046475403010845184, + 0.23821769654750824, + 0.14289279282093048, + 0.09442564100027084, + 0.03962784633040428, + 0.036109425127506256, + 0.08731483668088913, + 0.025480711832642555, + 0.17011870443820953 + ], + "credit_map": [ + 0.01083630695939064, + 0.02277945540845394, + 0.012886510230600834, + 0.016080765053629875, + 0.02400115877389908, + 0.013551387004554272, + 0.009611363522708416, + 0.009589971974492073, + 0.007083254866302013, + 0.007879305630922318, + 0.005040405783802271, + 0.004726534243673086, + 0.006975627038627863, + 0.005506542511284351, + 0.005367593374103308, + 0.0038961393292993307, + 0.032373104244470596, + 0.05568021535873413, + 0.0177573524415493, + 0.018488474190235138, + 0.03557925671339035, + 0.04487869143486023, + 0.018286865204572678, + 0.01517374999821186, + 0.016579292714595795, + 0.03503400832414627, + 0.009860430844128132, + 0.011611935682594776, + 0.022985516116023064, + 0.026322750374674797, + 0.009809463284909725, + 0.01068939920514822, + 0.011577966623008251, + 0.02238485962152481, + 0.007230304181575775, + 0.00722343847155571, + 0.02047697827219963, + 0.012157097458839417, + 0.007558232638984919, + 0.005816769320517778, + 0.003903484670445323, + 0.007201868109405041, + 0.004930080380290747, + 0.004778036382049322, + 0.006204408127814531, + 0.005508446134626865, + 0.0039354050531983376, + 0.0031661151442676783, + 0.0033134918194264174, + 0.00673602893948555, + 0.0036069038324058056, + 0.0037895815912634134, + 0.005013874266296625, + 0.006171178538352251, + 0.003849579719826579, + 0.00362878805026412, + 0.010068261064589024, + 0.010594603605568409, + 0.010908761993050575, + 0.010642683133482933, + 0.014209618791937828, + 0.014463948085904121, + 0.00823905784636736, + 0.00818790402263403, + 0.0027164435014128685, + 0.0035046858247369528, + 0.003168300725519657, + 0.0018865057500079274, + 0.004646861460059881, + 0.004247288219630718, + 0.0027803361881524324, + 0.002530290512368083, + 0.013366005383431911, + 0.02612503618001938, + 0.013645727187395096, + 0.013984277844429016, + 0.05328276380896568, + 0.030282747000455856, + 0.01179446466267109, + 0.007637670263648033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 3 + ], + "selected_reward": -0.05893462896347046 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28130730986595154 + ], + "opposite_logprob_tokens": [ + -1.406307339668274 + ], + "logprob_reward": -0.28130730986595154, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28130730986595154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14862975478172302, + 0.07954379171133041, + 0.028897477313876152, + 0.04312838241457939, + 0.44821667671203613, + 0.013895567506551743, + 0.02419915236532688, + 0.06816320866346359, + 0.07051423192024231, + 0.07481172680854797 + ], + "credit_map": [ + 0.0187835730612278, + 0.02074616216123104, + 0.012251128442585468, + 0.012341548688709736, + 0.04000125080347061, + 0.022212808951735497, + 0.011663719080388546, + 0.010629571974277496, + 0.012866745702922344, + 0.013735340908169746, + 0.007556742522865534, + 0.00767068937420845, + 0.017173150554299355, + 0.009280730970203876, + 0.006329468451440334, + 0.004930919036269188, + 0.003186204005032778, + 0.00473837461322546, + 0.0030727332923561335, + 0.0029508064035326242, + 0.006532494910061359, + 0.003734384663403034, + 0.0025130549911409616, + 0.0021694262977689505, + 0.005240709520876408, + 0.005911370273679495, + 0.00440753111615777, + 0.004123332444578409, + 0.006140211131423712, + 0.010144265368580818, + 0.0035090362653136253, + 0.003651929320767522, + 0.0496479794383049, + 0.07594142854213715, + 0.04268317669630051, + 0.049144893884658813, + 0.07364355027675629, + 0.0707138255238533, + 0.038677845150232315, + 0.04776396602392197, + 0.001350470818579197, + 0.0028501006308943033, + 0.0017096444498747587, + 0.0015544499037787318, + 0.0014995550736784935, + 0.0021132463589310646, + 0.0015696169575676322, + 0.0012484830804169178, + 0.0018590801628306508, + 0.004140688572078943, + 0.002299713436514139, + 0.003195761237293482, + 0.004287434741854668, + 0.0035289383959025145, + 0.0024369913153350353, + 0.0024505455512553453, + 0.01064668595790863, + 0.006274799816310406, + 0.00952449906617403, + 0.007669712416827679, + 0.012338436208665371, + 0.008276662789285183, + 0.006208925042301416, + 0.007223488297313452, + 0.006609811447560787, + 0.00510012311860919, + 0.011224531568586826, + 0.009664702229201794, + 0.007980260998010635, + 0.007599896285682917, + 0.010404566302895546, + 0.011930336244404316, + 0.007649702951312065, + 0.007902677170932293, + 0.010140978731215, + 0.007402128539979458, + 0.016358477994799614, + 0.009774434380233288, + 0.00785006769001484, + 0.007733258884400129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.28130730986595154 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015209070406854153 + ], + "opposite_logprob_tokens": [ + -6.501521110534668 + ], + "logprob_reward": -0.0015209070406854153, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015209070406854153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18462532758712769, + 0.036107514053583145, + 0.04120437428355217, + 0.07032587379217148, + 0.4369310140609741, + 0.029950307682156563, + 0.024662114679813385, + 0.024665653705596924, + 0.03775753825902939, + 0.11377029865980148 + ], + "credit_map": [ + 0.015205317176878452, + 0.04793478548526764, + 0.006998458411544561, + 0.00831274688243866, + 0.06968968361616135, + 0.021486982703208923, + 0.006778353359550238, + 0.008218974806368351, + 0.003808057401329279, + 0.006536014378070831, + 0.0036162715405225754, + 0.0036799951922148466, + 0.006125385407358408, + 0.003351605962961912, + 0.0047529032453894615, + 0.004237278830260038, + 0.005753230769187212, + 0.007155008614063263, + 0.0036704600788652897, + 0.004024476278573275, + 0.0071735386736691, + 0.006503494922071695, + 0.003732984885573387, + 0.003191181691363454, + 0.007462561596184969, + 0.01670522801578045, + 0.004230604972690344, + 0.005383431911468506, + 0.012289884500205517, + 0.013988366350531578, + 0.004638122394680977, + 0.005627675447613001, + 0.052078742533922195, + 0.07531438022851944, + 0.03986475616693497, + 0.04819300025701523, + 0.06910242140293121, + 0.06906792521476746, + 0.03881893679499626, + 0.04449082538485527, + 0.0034518945030868053, + 0.0066284663043916225, + 0.0040128473192453384, + 0.001815686235204339, + 0.004055246710777283, + 0.004470286890864372, + 0.0025341922882944345, + 0.0029816869646310806, + 0.002008795505389571, + 0.004956773016601801, + 0.0023844700772315264, + 0.0024996085558086634, + 0.004319549538195133, + 0.0031465415377169847, + 0.0027224437799304724, + 0.002623931737616658, + 0.0026032403111457825, + 0.004259057808667421, + 0.0028069589752703905, + 0.0022936270106583834, + 0.0034310652408748865, + 0.0049787284806370735, + 0.002238454530015588, + 0.0020545206498354673, + 0.0036316681653261185, + 0.00248918728902936, + 0.006054895464330912, + 0.005097927525639534, + 0.0043868413195014, + 0.003533340059220791, + 0.005857750773429871, + 0.006705929059535265, + 0.009359239600598812, + 0.01411485020071268, + 0.009352422319352627, + 0.011940741911530495, + 0.02963174693286419, + 0.020925480872392654, + 0.01069137454032898, + 0.007754444144666195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.0015209070406854153 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0205826573073864 + ], + "opposite_logprob_tokens": [ + -3.895582675933838 + ], + "logprob_reward": -0.0205826573073864, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0205826573073864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24425284564495087, + 0.030618805438280106, + 0.04725682735443115, + 0.06210704520344734, + 0.43988052010536194, + 0.02703380398452282, + 0.023747771978378296, + 0.034937258809804916, + 0.01775878667831421, + 0.0724063515663147 + ], + "credit_map": [ + 0.019166002050042152, + 0.07120995223522186, + 0.010721192695200443, + 0.01017600018531084, + 0.09268910437822342, + 0.021377818658947945, + 0.008428925648331642, + 0.010483850724995136, + 0.005229107104241848, + 0.004334369208663702, + 0.0022561755031347275, + 0.0031303642317652702, + 0.005376875400543213, + 0.003192371688783169, + 0.003958446905016899, + 0.0031410958617925644, + 0.00662144273519516, + 0.009875686839222908, + 0.004641223698854446, + 0.004290665965527296, + 0.006925033405423164, + 0.008043305017054081, + 0.003908122424036264, + 0.002951348666101694, + 0.007348234765231609, + 0.015332725830376148, + 0.0041336482390761375, + 0.004746588412672281, + 0.010942238382995129, + 0.011261491104960442, + 0.004123759921640158, + 0.004218359477818012, + 0.051505666226148605, + 0.08462873101234436, + 0.038673631846904755, + 0.04214534908533096, + 0.06763800978660583, + 0.07033083587884903, + 0.037100743502378464, + 0.04785752668976784, + 0.0028955326415598392, + 0.005846803076565266, + 0.003244875930249691, + 0.002017732011154294, + 0.0034644599072635174, + 0.005232459399849176, + 0.0022155167534947395, + 0.0021164233330637217, + 0.0019383805338293314, + 0.00458782073110342, + 0.0022999360226094723, + 0.002504415810108185, + 0.0037396387197077274, + 0.0035600061528384686, + 0.0026090340688824654, + 0.0025085401721298695, + 0.004562124609947205, + 0.004292261321097612, + 0.005009894724935293, + 0.004651265684515238, + 0.0049606082029640675, + 0.004625631030648947, + 0.0032005554530769587, + 0.003634918946772814, + 0.002390607027336955, + 0.001937350258231163, + 0.002432028530165553, + 0.0016056980239227414, + 0.0027403151616454124, + 0.0018449483904987574, + 0.002674766816198826, + 0.0021330725867301226, + 0.0057887667790055275, + 0.008150794543325901, + 0.008034253492951393, + 0.008369156159460545, + 0.018203895539045334, + 0.009961821138858795, + 0.008719809353351593, + 0.005177849438041449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.0205826573073864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001178642618469894 + ], + "opposite_logprob_tokens": [ + -6.751178741455078 + ], + "logprob_reward": -0.001178642618469894, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001178642618469894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20387427508831024, + 0.04103060066699982, + 0.05074737221002579, + 0.1779506802558899, + 0.3026179373264313, + 0.058368828147649765, + 0.034878235310316086, + 0.025558875873684883, + 0.05639832094311714, + 0.04857482761144638 + ], + "credit_map": [ + 0.014966695569455624, + 0.06351494044065475, + 0.0106691624969244, + 0.009056571871042252, + 0.07340766489505768, + 0.01679297164082527, + 0.00748132448643446, + 0.007984963245689869, + 0.00678427517414093, + 0.00841238722205162, + 0.0032534082420170307, + 0.0037125875242054462, + 0.00805242545902729, + 0.004060852341353893, + 0.0031754274386912584, + 0.0035792384296655655, + 0.007053664419800043, + 0.009055142290890217, + 0.0050941877998411655, + 0.005842892453074455, + 0.006078304257243872, + 0.00855051726102829, + 0.004901897627860308, + 0.004170771222561598, + 0.02103468030691147, + 0.041551362723112106, + 0.012578734196722507, + 0.014790748246014118, + 0.02847439795732498, + 0.03409789130091667, + 0.01194286160171032, + 0.013480025343596935, + 0.036099668592214584, + 0.050599824637174606, + 0.029876822605729103, + 0.03389904275536537, + 0.04296746477484703, + 0.04738054797053337, + 0.030493637546896935, + 0.03130093961954117, + 0.006992808543145657, + 0.012825366109609604, + 0.00576605461537838, + 0.004349519498646259, + 0.007787003181874752, + 0.010386343114078045, + 0.0051750424318015575, + 0.005086697172373533, + 0.0029144948348402977, + 0.007087509613484144, + 0.004026907961815596, + 0.0039171865209937096, + 0.00419580889865756, + 0.004386535380035639, + 0.004148712381720543, + 0.0042010825127363205, + 0.003386612981557846, + 0.0038895828183740377, + 0.0025929694529622793, + 0.002289721742272377, + 0.0041367425583302975, + 0.004738204181194305, + 0.002066404325887561, + 0.0024586403742432594, + 0.005532724782824516, + 0.004166206810623407, + 0.009744025766849518, + 0.007863654755055904, + 0.00633968086913228, + 0.005031015258282423, + 0.008902806788682938, + 0.00881821010261774, + 0.004546329844743013, + 0.007861155085265636, + 0.004761677701026201, + 0.005404253955930471, + 0.007107576355338097, + 0.007920985110104084, + 0.005848669912666082, + 0.005124181974679232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.001178642618469894 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Viewed from a first-person perspective, the camera executes a slow, continuous, and tranquil forward drift beneath the crystal-clear surface of the Great Barrier Reef. Rendered in a vibrant, impressionistic style, the entire scene transforms the underwater world into a living, breathing painting. Gentle ripples on the water above act like a dynamic lens, fragmenting the brilliant sunlight into a constant dance of shimmering, corrosive patterns that punctuate the seascape below. Coral structures lose their clear biological definition, dissolving into a breathtaking mosaic of textured colors and shifting light; vast fields of staghorn and brain coral appear as thick, visible splotches and swirls of soft pinks, deep violets, and warm oranges. Schools of fish are depicted not as individual creatures but as fleeting blurs and swift, ephemeral streaks of pure pigment. Flashes of electric blue and canary yellow dart across the turquoise water like the quick, deliberate strokes of a painter's brush. The atmosphere is one of profound tranquility and dreamlike beauty, where the boundaries between light, water, and life blur into a beautiful, ever-shifting synthesis of movement and color.", + "seed": 28847, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_8/index_9815_seed_28847_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1114779950818047, + "vlm_logprob": -0.1114779950818047, + "vlm_yes_no_margin": 4.399999994039535, + "vlm_token_credit": -0.1114779950818047, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07434698194265366 + ], + "opposite_logprob_tokens": [ + -2.8243470191955566 + ], + "logprob_reward": -0.07434698194265366, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07434698194265366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11272791028022766, + 0.030739396810531616, + 0.5652308464050293, + 0.01968228444457054, + 0.025080908089876175, + 0.020586352795362473, + 0.10921142995357513, + 0.05997969210147858, + 0.04100381210446358, + 0.015757400542497635 + ], + "credit_map": [ + 0.027229243889451027, + 0.009160296060144901, + 0.005147006828337908, + 0.005552204791456461, + 0.014043597504496574, + 0.03800038620829582, + 0.007936329580843449, + 0.005658844485878944, + 0.004894486628472805, + 0.007361426949501038, + 0.002089825924485922, + 0.002054478507488966, + 0.006194206885993481, + 0.0034094613511115313, + 0.002684438368305564, + 0.002051070798188448, + 0.1228800117969513, + 0.041757356375455856, + 0.021305538713932037, + 0.01586383581161499, + 0.24909374117851257, + 0.07973218709230423, + 0.01733221299946308, + 0.017265940085053444, + 0.0031495457515120506, + 0.0021114591509103775, + 0.002251280238851905, + 0.0012820350239053369, + 0.004476274363696575, + 0.003156807739287615, + 0.0019331795629113913, + 0.0013217012165114284, + 0.0030412168707698584, + 0.003660891205072403, + 0.003825362306088209, + 0.003699844004586339, + 0.002737263450399041, + 0.002140713855624199, + 0.003377430373802781, + 0.0025981850922107697, + 0.002608501585200429, + 0.004522046074271202, + 0.0017216582782566547, + 0.0021658933255821466, + 0.002851356053724885, + 0.002825334668159485, + 0.0021585188806056976, + 0.0017330431146547198, + 0.007776323240250349, + 0.023375369608402252, + 0.010343499481678009, + 0.006898671854287386, + 0.032453980296850204, + 0.01278661284595728, + 0.009919109754264355, + 0.005657853092998266, + 0.00617721863090992, + 0.007578578311949968, + 0.003909633960574865, + 0.013293573632836342, + 0.010529151186347008, + 0.0063405283726751804, + 0.007513629272580147, + 0.004637373145669699, + 0.006808241363614798, + 0.005449314601719379, + 0.002823011949658394, + 0.00613307673484087, + 0.0028823327738791704, + 0.0021880550775676966, + 0.010853016749024391, + 0.0038667612243443727, + 0.002165230456739664, + 0.0018531236564740539, + 0.0017358134500682354, + 0.0011306001106277108, + 0.0022115646861493587, + 0.004136070609092712, + 0.0013901726342737675, + 0.0011348237749189138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.07434698194265366 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742174446582794 + ], + "opposite_logprob_tokens": [ + -0.974217414855957 + ], + "logprob_reward": -0.4742174446582794, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.4742174446582794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04752900451421738, + 0.10303263366222382, + 0.2888983190059662, + 0.022041963413357735, + 0.02357536181807518, + 0.13096660375595093, + 0.26175230741500854, + 0.04095059633255005, + 0.03735522925853729, + 0.04389801248908043 + ], + "credit_map": [ + 0.009513894096016884, + 0.003988152835518122, + 0.004957384895533323, + 0.0031531956046819687, + 0.003291155444458127, + 0.012815766036510468, + 0.005691357888281345, + 0.0041180988773703575, + 0.02167508937418461, + 0.018160097301006317, + 0.004864857066422701, + 0.004907734226435423, + 0.028073502704501152, + 0.012019435875117779, + 0.006719983648508787, + 0.006611930672079325, + 0.0627310499548912, + 0.019691070541739464, + 0.013890656642615795, + 0.008969791233539581, + 0.12403940409421921, + 0.041493404656648636, + 0.009394154883921146, + 0.008688781410455704, + 0.005371769890189171, + 0.002877536928281188, + 0.002273677382618189, + 0.0014666374772787094, + 0.004291755612939596, + 0.0025092666037380695, + 0.0018407246097922325, + 0.0014105946756899357, + 0.0030113838147372007, + 0.0036357794888317585, + 0.0028038506861776114, + 0.003120267763733864, + 0.0023704208433628082, + 0.0021850713528692722, + 0.00364035670645535, + 0.002808231860399246, + 0.018376324325799942, + 0.02698039449751377, + 0.009984261356294155, + 0.014463583938777447, + 0.020993297919631004, + 0.020902689546346664, + 0.00940823182463646, + 0.00985781941562891, + 0.016436973586678505, + 0.07728654146194458, + 0.012454896233975887, + 0.012006777338683605, + 0.09018351137638092, + 0.022878173738718033, + 0.019904671236872673, + 0.010600754991173744, + 0.0039226822555065155, + 0.006004562601447105, + 0.002916814759373665, + 0.007198085077106953, + 0.009262831881642342, + 0.00466896640136838, + 0.004354839213192463, + 0.0026218160055577755, + 0.0062149870209395885, + 0.004281823989003897, + 0.0025516944006085396, + 0.005898676812648773, + 0.0022383490577340126, + 0.0020903602708131075, + 0.0102771557867527, + 0.003802182152867317, + 0.0066591231152415276, + 0.005300544667989016, + 0.004716966301202774, + 0.004142732359468937, + 0.004480569623410702, + 0.009616068564355373, + 0.005444055423140526, + 0.0035379482433199883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 5 + ], + "selected_reward": -0.4742174446582794 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010488491971045732 + ], + "opposite_logprob_tokens": [ + -6.876049041748047 + ], + "logprob_reward": -0.0010488491971045732, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010488491971045732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18494991958141327, + 0.18237440288066864, + 0.06064481660723686, + 0.10371418297290802, + 0.0402742400765419, + 0.03454184904694557, + 0.07458972930908203, + 0.1753043830394745, + 0.08406507968902588, + 0.059541452676057816 + ], + "credit_map": [ + 0.05616430938243866, + 0.013932724483311176, + 0.007600536569952965, + 0.006729916203767061, + 0.01182481087744236, + 0.07298988848924637, + 0.008468272164463997, + 0.0072394367307424545, + 0.04392498359084129, + 0.02767082117497921, + 0.00773851340636611, + 0.00598440645262599, + 0.06436409801244736, + 0.015673836693167686, + 0.008972390554845333, + 0.008045335300266743, + 0.00912147294729948, + 0.005420920439064503, + 0.00675912806764245, + 0.004562443122267723, + 0.017227930948138237, + 0.009392686188220978, + 0.004436448682099581, + 0.003723778761923313, + 0.021662171930074692, + 0.01695600524544716, + 0.010921260342001915, + 0.007584653794765472, + 0.012564465403556824, + 0.01738729141652584, + 0.008637025952339172, + 0.00800129771232605, + 0.006407707929611206, + 0.006651680450886488, + 0.00394799979403615, + 0.003732205368578434, + 0.007304678205400705, + 0.004558355547487736, + 0.0041490234434604645, + 0.0035225823521614075, + 0.003964673727750778, + 0.006336364895105362, + 0.004645978566259146, + 0.00411646394059062, + 0.005431588273495436, + 0.004038421902805567, + 0.002857926767319441, + 0.003150424687191844, + 0.007378351408988237, + 0.0076035670936107635, + 0.008136974647641182, + 0.008718203753232956, + 0.012751193717122078, + 0.009914347901940346, + 0.013443734496831894, + 0.0066433437168598175, + 0.016750343143939972, + 0.022250419482588768, + 0.01024159137159586, + 0.040670957416296005, + 0.03627338632941246, + 0.01635410077869892, + 0.019483694806694984, + 0.013279871083796024, + 0.010299666784703732, + 0.012743417173624039, + 0.008124840445816517, + 0.010078493505716324, + 0.012358484789729118, + 0.011080224066972733, + 0.01222569402307272, + 0.007154247257858515, + 0.008401433005928993, + 0.006899598054587841, + 0.005583155434578657, + 0.006489480379968882, + 0.009586376138031483, + 0.009859306737780571, + 0.008681681007146835, + 0.004040417727082968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0010488491971045732 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736662704497576 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.006736662704497576, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736662704497576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10610228776931763, + 0.10997257381677628, + 0.1953030377626419, + 0.075131356716156, + 0.08831033855676651, + 0.050085704773664474, + 0.2046065479516983, + 0.04171443730592728, + 0.08013050258159637, + 0.04864325374364853 + ], + "credit_map": [ + 0.015193360857665539, + 0.013627913780510426, + 0.007931874133646488, + 0.009829070419073105, + 0.020245227962732315, + 0.016791561618447304, + 0.01450908649712801, + 0.007974185980856419, + 0.015333815477788448, + 0.026074115186929703, + 0.00915990024805069, + 0.008424270898103714, + 0.01981222815811634, + 0.011295726522803307, + 0.011503441259264946, + 0.008369076065719128, + 0.03160606324672699, + 0.015565330162644386, + 0.019794128835201263, + 0.010940192267298698, + 0.06619832664728165, + 0.023497024551033974, + 0.012399456463754177, + 0.015302525833249092, + 0.010833856649696827, + 0.009007732383906841, + 0.008283399976789951, + 0.007842178456485271, + 0.008939817547798157, + 0.01540107000619173, + 0.008772382512688637, + 0.006050920579582453, + 0.011335528455674648, + 0.012261994183063507, + 0.012772803194820881, + 0.013285839930176735, + 0.00865534320473671, + 0.0073085990734398365, + 0.01254184078425169, + 0.010148385539650917, + 0.0069591328501701355, + 0.01161060482263565, + 0.003550854977220297, + 0.005634556990116835, + 0.007944248616695404, + 0.007001063786447048, + 0.003607311751693487, + 0.003777931211516261, + 0.014754947274923325, + 0.030194345861673355, + 0.01466258056461811, + 0.023487474769353867, + 0.044246722012758255, + 0.021833211183547974, + 0.036982182413339615, + 0.018445085734128952, + 0.0035950099118053913, + 0.004588049370795488, + 0.0033570730593055487, + 0.005539160221815109, + 0.010023334994912148, + 0.0058228932321071625, + 0.005572434049099684, + 0.003216484095901251, + 0.013436770997941494, + 0.008324100635945797, + 0.005635480862110853, + 0.012541905976831913, + 0.0045126378536224365, + 0.003676115535199642, + 0.0240799430757761, + 0.007923539727926254, + 0.0075414348393678665, + 0.006578715052455664, + 0.0051908548921346664, + 0.005794225726276636, + 0.0042042359709739685, + 0.009818854741752148, + 0.005757124163210392, + 0.0037578074261546135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.006736662704497576 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective first-person, drifting forward beneath the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in the Great Barrier Reef with visible coral structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coral structures rendered as a mosaic of soft pinks, deep violets, and warm oranges rather than clear biological forms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of fish depicted as fleeting blurs and swift streaks of pigment rather than individual creatures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style impressionistic, featuring shimmering patterns of light and color that fragment the sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010400369064882398 + ], + "opposite_logprob_tokens": [ + -6.876039981842041 + ], + "logprob_reward": -0.0010400369064882398, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010400369064882398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18684129416942596, + 0.06996537744998932, + 0.25473466515541077, + 0.05044722929596901, + 0.021142341196537018, + 0.08345908671617508, + 0.07480812072753906, + 0.07533114403486252, + 0.11667878925800323, + 0.06659197062253952 + ], + "credit_map": [ + 0.04268572852015495, + 0.015404252335429192, + 0.011094252578914165, + 0.008176302537322044, + 0.02365906909108162, + 0.06241658329963684, + 0.013458477333188057, + 0.009946626611053944, + 0.014126673340797424, + 0.009404540061950684, + 0.0045019411481916904, + 0.0034911916591227055, + 0.02010585553944111, + 0.007179310545325279, + 0.005177316255867481, + 0.005978541914373636, + 0.052420150488615036, + 0.018010307103395462, + 0.029399581253528595, + 0.011082571931183338, + 0.08117838203907013, + 0.03242157772183418, + 0.01087180245667696, + 0.019350284710526466, + 0.007940673269331455, + 0.006431323941797018, + 0.004794083070009947, + 0.005130880046635866, + 0.007229577749967575, + 0.009654952213168144, + 0.005497048143297434, + 0.0037686892319470644, + 0.002621403895318508, + 0.0029608190525323153, + 0.0029701495077461004, + 0.0018270497675985098, + 0.0026990892365574837, + 0.0038940415252000093, + 0.0019014051649719477, + 0.0022683818824589252, + 0.012762322090566158, + 0.016140349209308624, + 0.0066746934317052364, + 0.009350809268653393, + 0.014545775018632412, + 0.01327704731374979, + 0.005092853680253029, + 0.005615223199129105, + 0.006577870808541775, + 0.01663978025317192, + 0.006676137913018465, + 0.007720237597823143, + 0.016347816213965416, + 0.00794301275163889, + 0.007729572243988514, + 0.005173688754439354, + 0.00691894581541419, + 0.010020180605351925, + 0.004687729757279158, + 0.0158782247453928, + 0.01779879257082939, + 0.006840662099421024, + 0.00795198604464531, + 0.005234620533883572, + 0.019133413210511208, + 0.015490182675421238, + 0.009004181250929832, + 0.016308456659317017, + 0.011085135862231255, + 0.009919213131070137, + 0.02497245743870735, + 0.010765735059976578, + 0.009570915251970291, + 0.008103882893919945, + 0.007377782836556435, + 0.006212411914020777, + 0.00726716173812747, + 0.013985589146614075, + 0.008486971259117126, + 0.005587251391261816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.0010400369064882398 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_9/index_tensor([188, 188])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_9/index_tensor([188, 188])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8c9194f790bf5006b6b05883f137dfeed8a06694 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_9/index_tensor([188, 188])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning high-angle aerial photograph, as if taken from a drone, slowly tracks a lone mountain biker as he navigates a winding, narrow dirt path carved into the ridge of a vast, rugged mountain range. The low, late afternoon sun casts a warm golden hue over the epic landscape, casting long, dramatic shadows into the deep canyons below and illuminating the jagged, sun-scorched peaks in the distance. The riders, petite yet distinctive, clad in vibrant red windbreakers and black helmets, lean deliberately into the curves of the reddish-brown path, their bodies mere specks of determined movement. Their full-suspension mountain bikes, with their thick, knobby tires, steadily kick up a plume of light-colored ochre dust, capturing the whisper of golden light that follows them. The camera maintains its wide perspective, emphasizing the vastness of the wilderness and the solitary nature of the journey, creating a powerful atmosphere of awe, freedom, and the quiet endurance of humanity against the vastness of nature.", + "seed": 5572, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_9/index_188_seed_5572_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10627487068995833, + "vlm_logprob": -0.10627487068995833, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.10627487068995833, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038427963852882385 + ], + "opposite_logprob_tokens": [ + -3.5384280681610107 + ], + "logprob_reward": -0.038427963852882385, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.038427963852882385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2074861228466034, + 0.1453237235546112, + 0.15392880141735077, + 0.06086592748761177, + 0.09010548144578934, + 0.07430863380432129, + 0.13207431137561798, + 0.03407052159309387, + 0.0526728555560112, + 0.04916359484195709 + ], + "credit_map": [ + 0.01741897501051426, + 0.01506184134632349, + 0.027446405962109566, + 0.046061091125011444, + 0.01655218005180359, + 0.02032102458178997, + 0.04166121780872345, + 0.022963399067521095, + 0.009562830440700054, + 0.007568784523755312, + 0.022358261048793793, + 0.043091025203466415, + 0.00843198224902153, + 0.010211236774921417, + 0.0217224583029747, + 0.02237713523209095, + 0.01602596789598465, + 0.018119709566235542, + 0.02280137501657009, + 0.022284051403403282, + 0.021496200934052467, + 0.013612600974738598, + 0.019574617967009544, + 0.02001427672803402, + 0.005652733147144318, + 0.007239567115902901, + 0.007109328638762236, + 0.006704962812364101, + 0.013498460873961449, + 0.007686697877943516, + 0.008180605247616768, + 0.0047935741022229195, + 0.007495517376810312, + 0.007328598294407129, + 0.009515165351331234, + 0.014731091447174549, + 0.018747111782431602, + 0.010494914837181568, + 0.012099016457796097, + 0.009694065898656845, + 0.006645860616117716, + 0.009568311274051666, + 0.010032756254076958, + 0.008312717080116272, + 0.01802654191851616, + 0.008582607842981815, + 0.006709581706672907, + 0.006430258974432945, + 0.008707979694008827, + 0.007875337265431881, + 0.014003215357661247, + 0.004124595317989588, + 0.03678068891167641, + 0.04988908767700195, + 0.004089144989848137, + 0.006604256574064493, + 0.004011185839772224, + 0.004798565991222858, + 0.005125546362251043, + 0.003938155714422464, + 0.004759469069540501, + 0.006090383045375347, + 0.0028012001421302557, + 0.0025460170581936836, + 0.004496283829212189, + 0.006041466724127531, + 0.003508360590785742, + 0.005889587569981813, + 0.009635239839553833, + 0.013344403356313705, + 0.005155594553798437, + 0.004601924680173397, + 0.0038920408114790916, + 0.0035676059778779745, + 0.003899415722116828, + 0.00560270668938756, + 0.01322928722947836, + 0.014073239639401436, + 0.0025931173004209995, + 0.0023061814717948437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.038427963852882385 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812185674905777 + ], + "opposite_logprob_tokens": [ + -3.2881219387054443 + ], + "logprob_reward": -0.03812185674905777, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812185674905777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4610661268234253, + 0.05771542340517044, + 0.11983595788478851, + 0.07862038165330887, + 0.04226108267903328, + 0.02166161872446537, + 0.11260142922401428, + 0.029088139533996582, + 0.033632755279541016, + 0.043517112731933594 + ], + "credit_map": [ + 0.03842943534255028, + 0.031239939853549004, + 0.05669372156262398, + 0.07670079916715622, + 0.06987001746892929, + 0.05451192706823349, + 0.09900347888469696, + 0.034616779536008835, + 0.004803944379091263, + 0.003532531438395381, + 0.00869040098041296, + 0.015486002899706364, + 0.0040390691719949245, + 0.004484573379158974, + 0.009335602633655071, + 0.007343298755586147, + 0.011545103043317795, + 0.0136384516954422, + 0.021140294149518013, + 0.015762491151690483, + 0.014104348607361317, + 0.013051686808466911, + 0.01411338709294796, + 0.01648019440472126, + 0.007155735976994038, + 0.010427141562104225, + 0.011749058030545712, + 0.011765500530600548, + 0.009984871372580528, + 0.006996653974056244, + 0.011830235831439495, + 0.008711184374988079, + 0.0036019764374941587, + 0.005970378406345844, + 0.008791067637503147, + 0.006103226449340582, + 0.004523119889199734, + 0.004033475182950497, + 0.005018085241317749, + 0.004219754133373499, + 0.0021189923863857985, + 0.0022468960378319025, + 0.003350225044414401, + 0.0024080073926597834, + 0.004194740206003189, + 0.0028519551269710064, + 0.002555321902036667, + 0.0019354810938239098, + 0.00523443054407835, + 0.006102441344410181, + 0.018237167969346046, + 0.0032429478596895933, + 0.026168495416641235, + 0.045747365802526474, + 0.002834531245753169, + 0.005034047644585371, + 0.003290896536782384, + 0.003942613024264574, + 0.004618194419890642, + 0.0024974856059998274, + 0.004044895060360432, + 0.00538579560816288, + 0.0029479009099304676, + 0.0023603576701134443, + 0.003633287036791444, + 0.003689573844894767, + 0.0031198011711239815, + 0.0033058200497180223, + 0.005609180312603712, + 0.00741981528699398, + 0.003764621913433075, + 0.0030906524043530226, + 0.005754714831709862, + 0.004673713352531195, + 0.0053517199121415615, + 0.005564937833696604, + 0.006628981325775385, + 0.008238304406404495, + 0.00332990032620728, + 0.003974843304604292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.03812185674905777 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42897891998291016 + ], + "opposite_logprob_tokens": [ + -1.0539789199829102 + ], + "logprob_reward": -0.42897891998291016, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42897891998291016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2940225601196289, + 0.09976587444543839, + 0.08391162753105164, + 0.11483664065599442, + 0.043174296617507935, + 0.07798529416322708, + 0.11884191632270813, + 0.05494948849081993, + 0.06486433744430542, + 0.047647979110479355 + ], + "credit_map": [ + 0.02904917672276497, + 0.01916692592203617, + 0.033489927649497986, + 0.05439802259206772, + 0.036172788590192795, + 0.03959761559963226, + 0.05089552700519562, + 0.031252577900886536, + 0.010715119540691376, + 0.006073512602597475, + 0.013580325059592724, + 0.023744963109493256, + 0.007959947921335697, + 0.009146062657237053, + 0.013892491348087788, + 0.014653450809419155, + 0.009701183065772057, + 0.008767829276621342, + 0.015244759619235992, + 0.009923099540174007, + 0.010355358943343163, + 0.008030785247683525, + 0.01177632063627243, + 0.010112293995916843, + 0.011072801426053047, + 0.010144929401576519, + 0.017079509794712067, + 0.024377165362238884, + 0.011755116283893585, + 0.01438229065388441, + 0.012031280435621738, + 0.01399354636669159, + 0.0039321850053966045, + 0.005257750395685434, + 0.009038261137902737, + 0.005527033936232328, + 0.0054745119996368885, + 0.004174262750893831, + 0.005935899913311005, + 0.003834390314295888, + 0.0071580372750759125, + 0.007291941903531551, + 0.01042999792844057, + 0.00834068562835455, + 0.01680973917245865, + 0.010567241348326206, + 0.008886672556400299, + 0.008500983007252216, + 0.02179543301463127, + 0.015418248251080513, + 0.01504103560000658, + 0.007379484828561544, + 0.02048218995332718, + 0.016363028436899185, + 0.0059506855905056, + 0.016411807388067245, + 0.0058126854710280895, + 0.00536335026845336, + 0.005958663299679756, + 0.00468997098505497, + 0.008940322324633598, + 0.015396260656416416, + 0.004003941547125578, + 0.004784297663718462, + 0.005025949794799089, + 0.008991663344204426, + 0.0057620699517428875, + 0.006744901649653912, + 0.013420100323855877, + 0.015843261033296585, + 0.004964856896549463, + 0.004111532121896744, + 0.005144686903804541, + 0.0046448842622339725, + 0.006940975785255432, + 0.007706291973590851, + 0.0068669249303638935, + 0.008632978424429893, + 0.003507225541397929, + 0.004204011522233486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.42897891998291016 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257233511656523 + ], + "opposite_logprob_tokens": [ + -5.2552571296691895 + ], + "logprob_reward": -0.005257233511656523, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257233511656523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0831531509757042, + 0.020446525886654854, + 0.03799556568264961, + 0.08353646099567413, + 0.04178986698389053, + 0.021518927067518234, + 0.6505200862884521, + 0.021579086780548096, + 0.014574284665286541, + 0.02488606981933117 + ], + "credit_map": [ + 0.004819364752620459, + 0.004405202344059944, + 0.011674054898321629, + 0.019875504076480865, + 0.00674532912671566, + 0.00534693943336606, + 0.02375836856663227, + 0.00652839383110404, + 0.001375854597426951, + 0.001170142786577344, + 0.003069101832807064, + 0.006030283402651548, + 0.0011826278641819954, + 0.0015616555465385318, + 0.0029135190416127443, + 0.0031433403491973877, + 0.0028576357290148735, + 0.003889089450240135, + 0.010649509727954865, + 0.005453018471598625, + 0.0036044022999703884, + 0.005252222064882517, + 0.0031002420000731945, + 0.003189449431374669, + 0.008564064279198647, + 0.00788075290620327, + 0.014233042486011982, + 0.014557651244103909, + 0.008803615346550941, + 0.008291603997349739, + 0.01213520672172308, + 0.009070519357919693, + 0.00385834788903594, + 0.0034521003253757954, + 0.006605273112654686, + 0.006157785654067993, + 0.007353584282100201, + 0.003967211116105318, + 0.006194807589054108, + 0.004200753755867481, + 0.0019220520043745637, + 0.003941262606531382, + 0.0026615734677761793, + 0.0025765830650925636, + 0.0038060820661485195, + 0.0026293129194527864, + 0.0022874942515045404, + 0.0016945659881457686, + 0.02614413946866989, + 0.031039804220199585, + 0.08724893629550934, + 0.021498827263712883, + 0.17239366471767426, + 0.26072147488594055, + 0.02089414931833744, + 0.03057907149195671, + 0.0023432292509824038, + 0.0029915771447122097, + 0.0027232945431023836, + 0.002261481713503599, + 0.00352623057551682, + 0.004951645620167255, + 0.0012889151694253087, + 0.0014927120646461844, + 0.0010450839763507247, + 0.0015453329542651772, + 0.0013371646637097, + 0.0018434056546539068, + 0.0026656591799110174, + 0.003151518991217017, + 0.002005391288548708, + 0.0009807272581383586, + 0.0031639118678867817, + 0.0019054135773330927, + 0.001960123423486948, + 0.0021627030801028013, + 0.005777533166110516, + 0.006884740199893713, + 0.0015153440181165934, + 0.001516300835646689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.005257233511656523 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588379353284836 + ], + "opposite_logprob_tokens": [ + -3.8955883979797363 + ], + "logprob_reward": -0.020588379353284836, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588379353284836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1507481187582016, + 0.044609829783439636, + 0.20185278356075287, + 0.029343990609049797, + 0.05799025297164917, + 0.057778824120759964, + 0.23067590594291687, + 0.09460220485925674, + 0.0839715525507927, + 0.04842657595872879 + ], + "credit_map": [ + 0.008743757382035255, + 0.007349669001996517, + 0.023792820051312447, + 0.04015515372157097, + 0.010503157041966915, + 0.01167935412377119, + 0.03568028658628464, + 0.012843919917941093, + 0.0038691957015544176, + 0.0032909363508224487, + 0.006113701965659857, + 0.010370329022407532, + 0.0034832835663110018, + 0.006805091630667448, + 0.0051780082285404205, + 0.005499283317476511, + 0.01499970257282257, + 0.015897084027528763, + 0.05072801932692528, + 0.028796199709177017, + 0.02168290875852108, + 0.02889878861606121, + 0.021229028701782227, + 0.01962105557322502, + 0.0028982765506953, + 0.003267111489549279, + 0.0045335181057453156, + 0.004444052930921316, + 0.003487985348328948, + 0.002363205421715975, + 0.004505613818764687, + 0.0038442290388047695, + 0.005913461558520794, + 0.004032730124890804, + 0.006975858937948942, + 0.00792861171066761, + 0.011038847267627716, + 0.006729993969202042, + 0.007094725500792265, + 0.008276022039353848, + 0.005989592522382736, + 0.009210626594722271, + 0.008830943144857883, + 0.006759276147931814, + 0.010691150091588497, + 0.006104720756411552, + 0.005669509060680866, + 0.0045230090618133545, + 0.014483919367194176, + 0.013920524157583714, + 0.02642059698700905, + 0.008046546950936317, + 0.06262455135583878, + 0.08685213327407837, + 0.007396775297820568, + 0.010930844582617283, + 0.010268359445035458, + 0.010186944156885147, + 0.016397492960095406, + 0.012891251593828201, + 0.013414165005087852, + 0.01425683032721281, + 0.010204696096479893, + 0.006982459221035242, + 0.007306663319468498, + 0.008210370317101479, + 0.00895728264003992, + 0.009404141455888748, + 0.014480572193861008, + 0.018522948026657104, + 0.01005439180880785, + 0.0070351846516132355, + 0.007489366456866264, + 0.00541320163756609, + 0.004136121831834316, + 0.004354771692305803, + 0.009174779988825321, + 0.010029682889580727, + 0.0035883975215256214, + 0.004240254405885935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.020588379353284836 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning high-angle aerial photograph, as if taken from a drone, slowly tracks a lone mountain biker as he navigates a winding, narrow dirt path carved into the ridge of a vast, rugged mountain range. The low, late afternoon sun casts a warm golden hue over the epic landscape, casting long, dramatic shadows into the deep canyons below and illuminating the jagged, sun-scorched peaks in the distance. The riders, petite yet distinctive, clad in vibrant red windbreakers and black helmets, lean deliberately into the curves of the reddish-brown path, their bodies mere specks of determined movement. Their full-suspension mountain bikes, with their thick, knobby tires, steadily kick up a plume of light-colored ochre dust, capturing the whisper of golden light that follows them. The camera maintains its wide perspective, emphasizing the vastness of the wilderness and the solitary nature of the journey, creating a powerful atmosphere of awe, freedom, and the quiet endurance of humanity against the vastness of nature.", + "seed": 5573, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_9/index_188_seed_5573_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09833540813997388, + "vlm_logprob": -0.09833540813997388, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.09833540813997388, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013885797932744026 + ], + "opposite_logprob_tokens": [ + -4.638885974884033 + ], + "logprob_reward": -0.013885797932744026, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013885797932744026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2951759696006775, + 0.06898271292448044, + 0.04186404123902321, + 0.14362837374210358, + 0.07793357968330383, + 0.1139233410358429, + 0.012176983058452606, + 0.1846083551645279, + 0.037798792123794556, + 0.02390783093869686 + ], + "credit_map": [ + 0.03503890335559845, + 0.0359562411904335, + 0.025060173124074936, + 0.04116907715797424, + 0.03764071688055992, + 0.049478087574243546, + 0.030724404379725456, + 0.04010836035013199, + 0.008684328757226467, + 0.008485551923513412, + 0.004278790205717087, + 0.0046587372198700905, + 0.020310135558247566, + 0.012471714057028294, + 0.004824584815651178, + 0.005268877372145653, + 0.003961238544434309, + 0.004312953911721706, + 0.0044083683751523495, + 0.006697212811559439, + 0.004547738004475832, + 0.004926342982798815, + 0.007779042236506939, + 0.00523114250972867, + 0.023538438603281975, + 0.01263453159481287, + 0.00672959303483367, + 0.010325750336050987, + 0.008582530543208122, + 0.05508021265268326, + 0.014982905238866806, + 0.011754404753446579, + 0.007454912643879652, + 0.011859608814120293, + 0.007066444028168917, + 0.009918572381138802, + 0.005230099428445101, + 0.01915465109050274, + 0.009304714389145374, + 0.007944573648273945, + 0.006227179896086454, + 0.007937928661704063, + 0.0058836727403104305, + 0.006680442485958338, + 0.0361466109752655, + 0.03995884209871292, + 0.006392097566276789, + 0.004696565680205822, + 0.0013502076035365462, + 0.0012825861340388656, + 0.0011097150854766369, + 0.0013759569264948368, + 0.002353114075958729, + 0.0016175564378499985, + 0.0012847504112869501, + 0.0018030961509793997, + 0.0287280585616827, + 0.008197893388569355, + 0.010377169586718082, + 0.010823974385857582, + 0.039496712386608124, + 0.06929470598697662, + 0.00794872548431158, + 0.00974111445248127, + 0.0038882822263985872, + 0.006740320939570665, + 0.003937884233891964, + 0.004517499357461929, + 0.004499420057982206, + 0.007014235947281122, + 0.003664169693365693, + 0.003536980366334319, + 0.003080712864175439, + 0.003680009627714753, + 0.003456065198406577, + 0.00295180338434875, + 0.002720109187066555, + 0.002407390158623457, + 0.002854358870536089, + 0.002757382346317172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.013885797932744026 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059924043715 + ], + "opposite_logprob_tokens": [ + -3.895599126815796 + ], + "logprob_reward": -0.02059924043715, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059924043715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0891440212726593, + 0.04838348552584648, + 0.06216500699520111, + 0.07086025178432465, + 0.04516493156552315, + 0.3920816481113434, + 0.015825187787413597, + 0.15521028637886047, + 0.02074521966278553, + 0.10041999816894531 + ], + "credit_map": [ + 0.01128305122256279, + 0.010383933782577515, + 0.004989142995327711, + 0.0072031160816550255, + 0.014774029143154621, + 0.024377023801207542, + 0.007052465341985226, + 0.009081262163817883, + 0.005962561350315809, + 0.006655405275523663, + 0.004064399749040604, + 0.004840599838644266, + 0.009219437837600708, + 0.0057497331872582436, + 0.007104198914021254, + 0.004787146579474211, + 0.006016336847096682, + 0.00696423789486289, + 0.006732375826686621, + 0.009118200279772282, + 0.0073278918862342834, + 0.006516645662486553, + 0.011384735815227032, + 0.008104581385850906, + 0.010392065159976482, + 0.007062211632728577, + 0.003221649443730712, + 0.004821143113076687, + 0.004908538423478603, + 0.027951395139098167, + 0.007281733676791191, + 0.005221512168645859, + 0.004843706265091896, + 0.006029630545526743, + 0.004469267092645168, + 0.00865055713802576, + 0.005426603369414806, + 0.005606136284768581, + 0.005755063612014055, + 0.004383966792374849, + 0.02218671515583992, + 0.026482034474611282, + 0.02013050764799118, + 0.01865607313811779, + 0.1231684610247612, + 0.14515134692192078, + 0.01958196610212326, + 0.016724536195397377, + 0.0017965759616345167, + 0.0017246630741283298, + 0.0017109331674873829, + 0.0030332733877003193, + 0.002736185910180211, + 0.0014074159553274512, + 0.0016665385337546468, + 0.001749602030031383, + 0.02399960532784462, + 0.005619558040052652, + 0.007728924509137869, + 0.008984299376606941, + 0.03533823788166046, + 0.060286931693553925, + 0.005816306918859482, + 0.007436420768499374, + 0.0014888737350702286, + 0.0030977569986134768, + 0.0035533281043171883, + 0.0029575733933597803, + 0.0022571177687495947, + 0.0028825688641518354, + 0.002522096037864685, + 0.0019859059248119593, + 0.008601276203989983, + 0.013598519377410412, + 0.009967153891921043, + 0.013172059319913387, + 0.012268886901438236, + 0.014171515591442585, + 0.014502221718430519, + 0.014138362370431423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.02059924043715 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42897167801856995 + ], + "opposite_logprob_tokens": [ + -1.0539716482162476 + ], + "logprob_reward": -0.42897167801856995, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42897167801856995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14612752199172974, + 0.028694946318864822, + 0.13594606518745422, + 0.03114972822368145, + 0.107145756483078, + 0.19281132519245148, + 0.026819715276360512, + 0.2717493176460266, + 0.01969287358224392, + 0.03986278176307678 + ], + "credit_map": [ + 0.017161957919597626, + 0.015156345441937447, + 0.012641949579119682, + 0.016671206802129745, + 0.021729428321123123, + 0.030785204842686653, + 0.013204528950154781, + 0.018776895478367805, + 0.002761024981737137, + 0.003960023634135723, + 0.0036329319700598717, + 0.004084590822458267, + 0.004120582714676857, + 0.0026795798912644386, + 0.004078506026417017, + 0.003377706976607442, + 0.012813570909202099, + 0.016588328406214714, + 0.014992966316640377, + 0.02417677640914917, + 0.01355059165507555, + 0.011759612709283829, + 0.025539835914969444, + 0.016524381935596466, + 0.006720803212374449, + 0.010336381383240223, + 0.002241786802187562, + 0.002800487680360675, + 0.0034302333369851112, + 0.0019438748713582754, + 0.001969784963876009, + 0.0017063763225451112, + 0.0094061978161335, + 0.010948225855827332, + 0.010942937806248665, + 0.018403857946395874, + 0.009107735939323902, + 0.02183648943901062, + 0.015158114023506641, + 0.011342196725308895, + 0.01222308725118637, + 0.0140037527307868, + 0.009807652793824673, + 0.009183796122670174, + 0.06066098436713219, + 0.06935476511716843, + 0.009748606011271477, + 0.007828675210475922, + 0.0026376477908343077, + 0.003923076670616865, + 0.002267891075462103, + 0.0053623514249920845, + 0.0027919767890125513, + 0.0030096997506916523, + 0.003051195992156863, + 0.0037758774124085903, + 0.04352373257279396, + 0.009235582314431667, + 0.012883991934359074, + 0.013726607896387577, + 0.06282726675271988, + 0.10869327932596207, + 0.007948082871735096, + 0.012910773046314716, + 0.0018316622590646148, + 0.005363758187741041, + 0.0020070266909897327, + 0.0018912021769210696, + 0.0018913699313998222, + 0.0021541740279644728, + 0.0015279975486919284, + 0.0030256828758865595, + 0.005136873107403517, + 0.005133936647325754, + 0.004382318817079067, + 0.005721203982830048, + 0.004406343679875135, + 0.004839332774281502, + 0.004770633764564991, + 0.005472138524055481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.42897167801856995 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007634746376425028 + ], + "opposite_logprob_tokens": [ + -4.88263463973999 + ], + "logprob_reward": -0.007634746376425028, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007634746376425028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08040868490934372, + 0.06725093722343445, + 0.09927835315465927, + 0.030336076393723488, + 0.033037085086107254, + 0.10461802780628204, + 0.03257731720805168, + 0.3435824513435364, + 0.04876148700714111, + 0.16014952957630157 + ], + "credit_map": [ + 0.007278122939169407, + 0.010525713674724102, + 0.006062989588826895, + 0.007069443818181753, + 0.014602879993617535, + 0.014907426200807095, + 0.00735778734087944, + 0.012604324147105217, + 0.008955439552664757, + 0.012742571532726288, + 0.0046883984468877316, + 0.007500899955630302, + 0.014775324612855911, + 0.007553453091531992, + 0.0056625897996127605, + 0.005372268613427877, + 0.010608608834445477, + 0.013969548977911472, + 0.01028925459831953, + 0.014924632385373116, + 0.0113760307431221, + 0.014079742133617401, + 0.013634616509079933, + 0.010395929217338562, + 0.004934260621666908, + 0.005404278635978699, + 0.0021111650858074427, + 0.0031820640433579683, + 0.0030270167626440525, + 0.005705453455448151, + 0.0026732783298939466, + 0.0032985613215714693, + 0.005468007642775774, + 0.006994948722422123, + 0.002503288211300969, + 0.002928032772615552, + 0.003965477459132671, + 0.005627367179840803, + 0.0027007353492081165, + 0.0028492261189967394, + 0.005938668269664049, + 0.0070213498547673225, + 0.005450666416436434, + 0.0047461832873523235, + 0.03385277837514877, + 0.03786393627524376, + 0.0055412426590919495, + 0.004203201737254858, + 0.004483022727072239, + 0.004266039002686739, + 0.0032292166724801064, + 0.005053888540714979, + 0.005356106441468, + 0.003890545107424259, + 0.003598583862185478, + 0.002699918346479535, + 0.05315428227186203, + 0.013331468217074871, + 0.016876664012670517, + 0.019792864099144936, + 0.0782763883471489, + 0.133681982755661, + 0.01201721467077732, + 0.016451619565486908, + 0.003747418290004134, + 0.010026560164988041, + 0.006410850677639246, + 0.006533355452120304, + 0.005913316272199154, + 0.006276764441281557, + 0.004921645391732454, + 0.004931579809635878, + 0.014687835238873959, + 0.02553189918398857, + 0.021657435223460197, + 0.02183157205581665, + 0.02008489891886711, + 0.02342700958251953, + 0.015278983861207962, + 0.017649918794631958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.007634746376425028 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585577934980392 + ], + "opposite_logprob_tokens": [ + -3.895585536956787 + ], + "logprob_reward": -0.020585577934980392, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585577934980392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056413453072309494, + 0.03367527201771736, + 0.0412493571639061, + 0.3132513463497162, + 0.04129868000745773, + 0.15042448043823242, + 0.014247766695916653, + 0.29804083704948425, + 0.024896884337067604, + 0.026501916348934174 + ], + "credit_map": [ + 0.006515022832900286, + 0.009120237082242966, + 0.004776735790073872, + 0.00746099604293704, + 0.008133928291499615, + 0.007318606600165367, + 0.007234082091599703, + 0.00585384713485837, + 0.004754540044814348, + 0.004621300380676985, + 0.0032532978802919388, + 0.003622259944677353, + 0.008259675465524197, + 0.003721388289704919, + 0.0028560450300574303, + 0.0025867647491395473, + 0.0040037198923528194, + 0.004999905358999968, + 0.0038216602988541126, + 0.006377299781888723, + 0.004638333339244127, + 0.003708586795255542, + 0.008246365003287792, + 0.0054534850642085075, + 0.05434355139732361, + 0.02735198847949505, + 0.01416422426700592, + 0.02008516900241375, + 0.016950631514191628, + 0.12663069367408752, + 0.03230421617627144, + 0.021420883014798164, + 0.005040113348513842, + 0.008197586052119732, + 0.003700782312080264, + 0.002463605487719178, + 0.0021095257252454758, + 0.013905582018196583, + 0.0037214718759059906, + 0.0021600171457976103, + 0.008457942865788937, + 0.01067628525197506, + 0.007330026943236589, + 0.007176246494054794, + 0.04803850129246712, + 0.05489864945411682, + 0.007581545040011406, + 0.006265286821871996, + 0.001262269215658307, + 0.0013075186870992184, + 0.0014530885964632034, + 0.003091971157118678, + 0.002211130689829588, + 0.001470993272960186, + 0.0015976967988535762, + 0.0018530982779338956, + 0.04745050519704819, + 0.01053775567561388, + 0.013805469498038292, + 0.01351767498999834, + 0.06923598051071167, + 0.11982894688844681, + 0.009092623367905617, + 0.01457186695188284, + 0.0020218626596033573, + 0.004538388457149267, + 0.0025778240524232388, + 0.00260104495100677, + 0.002988611115142703, + 0.006267975550144911, + 0.0017810373101383448, + 0.002120140939950943, + 0.0021214825101196766, + 0.004044972360134125, + 0.005205344408750534, + 0.0037804085295647383, + 0.0018705263501033187, + 0.0021505861077457666, + 0.004285627976059914, + 0.0030429663602262735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.020585577934980392 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning high-angle aerial photograph, as if taken from a drone, slowly tracks a lone mountain biker as he navigates a winding, narrow dirt path carved into the ridge of a vast, rugged mountain range. The low, late afternoon sun casts a warm golden hue over the epic landscape, casting long, dramatic shadows into the deep canyons below and illuminating the jagged, sun-scorched peaks in the distance. The riders, petite yet distinctive, clad in vibrant red windbreakers and black helmets, lean deliberately into the curves of the reddish-brown path, their bodies mere specks of determined movement. Their full-suspension mountain bikes, with their thick, knobby tires, steadily kick up a plume of light-colored ochre dust, capturing the whisper of golden light that follows them. The camera maintains its wide perspective, emphasizing the vastness of the wilderness and the solitary nature of the journey, creating a powerful atmosphere of awe, freedom, and the quiet endurance of humanity against the vastness of nature.", + "seed": 5574, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_9/index_188_seed_5574_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09685258036479354, + "vlm_logprob": -0.09685258036479354, + "vlm_yes_no_margin": 3.1000000119209288, + "vlm_token_credit": -0.09685258036479354, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06347641348838806 + ], + "opposite_logprob_tokens": [ + -2.938476324081421 + ], + "logprob_reward": -0.06347641348838806, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06347641348838806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16620245575904846, + 0.04093213006854057, + 0.08664121478796005, + 0.3798401355743408, + 0.03078347072005272, + 0.03321906924247742, + 0.03723471611738205, + 0.08122847229242325, + 0.0478900671005249, + 0.09602825343608856 + ], + "credit_map": [ + 0.0231670793145895, + 0.015810497105121613, + 0.01841907948255539, + 0.015518919564783573, + 0.03648226335644722, + 0.026267506182193756, + 0.017171397805213928, + 0.013365707360208035, + 0.006235250737518072, + 0.004712476395070553, + 0.004875378217548132, + 0.004247158300131559, + 0.006171971559524536, + 0.005859491880983114, + 0.0032958618830889463, + 0.0055345408618450165, + 0.011181876063346863, + 0.01341409794986248, + 0.0068815345875918865, + 0.010095157660543919, + 0.01199121680110693, + 0.01268744096159935, + 0.008711038157343864, + 0.011678852140903473, + 0.0996723473072052, + 0.04001450166106224, + 0.012198169715702534, + 0.01317319180816412, + 0.022234808653593063, + 0.15682590007781982, + 0.009434121660888195, + 0.026287104934453964, + 0.0039238170720636845, + 0.0038426609244197607, + 0.003926028497517109, + 0.0033468925394117832, + 0.005125205032527447, + 0.003606774378567934, + 0.0035172668285667896, + 0.0034948266111314297, + 0.004479850642383099, + 0.00424048351123929, + 0.0022592884488403797, + 0.003869859967380762, + 0.0032306716311722994, + 0.0032859421335160732, + 0.006729967426508665, + 0.00512300431728363, + 0.004836747422814369, + 0.0047387173399329185, + 0.0034435573033988476, + 0.003390655154362321, + 0.005551875103265047, + 0.0045211780816316605, + 0.004949354566633701, + 0.005802631843835115, + 0.005759282037615776, + 0.006346949841827154, + 0.004801434464752674, + 0.004576046951115131, + 0.02251167967915535, + 0.013042996637523174, + 0.00921742431819439, + 0.01497266162186861, + 0.0037262742407619953, + 0.0053122988902032375, + 0.00583213847130537, + 0.006473949179053307, + 0.004531787242740393, + 0.0054002003744244576, + 0.010138573125004768, + 0.006474844180047512, + 0.008707491680979729, + 0.013329231180250645, + 0.007117231842130423, + 0.010643075220286846, + 0.016250131651759148, + 0.011464258655905724, + 0.015914209187030792, + 0.012602625414729118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.06347641348838806 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038122545927762985 + ], + "opposite_logprob_tokens": [ + -3.2881226539611816 + ], + "logprob_reward": -0.038122545927762985, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038122545927762985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.095243439078331, + 0.04472341388463974, + 0.24937967956066132, + 0.3936200737953186, + 0.07258744537830353, + 0.023098653182387352, + 0.02197495847940445, + 0.03177971392869949, + 0.01565035991370678, + 0.051942259073257446 + ], + "credit_map": [ + 0.012757005169987679, + 0.008916111662983894, + 0.009738254360854626, + 0.008413553237915039, + 0.026497626677155495, + 0.01705552451312542, + 0.0055807484313845634, + 0.0062846108339726925, + 0.005191456526517868, + 0.005572645924985409, + 0.004949783906340599, + 0.00530093302950263, + 0.007212274707853794, + 0.00569448247551918, + 0.005031239241361618, + 0.005770599003881216, + 0.02986866608262062, + 0.047792550176382065, + 0.0182813610881567, + 0.02964765951037407, + 0.04815361276268959, + 0.035673145204782486, + 0.020897023379802704, + 0.019065670669078827, + 0.08850831538438797, + 0.04482296481728554, + 0.014782424084842205, + 0.022443868219852448, + 0.031572602689266205, + 0.14968843758106232, + 0.013847255147993565, + 0.02795419842004776, + 0.00811704620718956, + 0.01552998274564743, + 0.009388630278408527, + 0.006984105799347162, + 0.007879359647631645, + 0.008539601229131222, + 0.0067171757109463215, + 0.009431542828679085, + 0.0028686528094112873, + 0.002816891996189952, + 0.0018616823945194483, + 0.002482400741428137, + 0.002548061078414321, + 0.0021597209852188826, + 0.004650517366826534, + 0.003710726508870721, + 0.0035321759060025215, + 0.0018776424694806337, + 0.0016401085304096341, + 0.002109148306772113, + 0.0037154420278966427, + 0.0031666180584579706, + 0.0022030838299542665, + 0.0037307399325072765, + 0.0029791209381073713, + 0.004835248924791813, + 0.0033362878020852804, + 0.002675574505701661, + 0.005191108211874962, + 0.005062918644398451, + 0.004282359965145588, + 0.0034170958679169416, + 0.001904366072267294, + 0.0015949064400047064, + 0.0013899720506742597, + 0.001612866879440844, + 0.0015934868715703487, + 0.0026018102653324604, + 0.0028241551481187344, + 0.0021287957206368446, + 0.0044257668778300285, + 0.006872424390166998, + 0.003972281701862812, + 0.008007040247321129, + 0.00802752748131752, + 0.005522046238183975, + 0.008787669241428375, + 0.006327503360807896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.038122545927762985 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486754298210144 + ], + "opposite_logprob_tokens": [ + -1.2236754894256592 + ], + "logprob_reward": -0.3486754298210144, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.3486754298210144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051573995500802994, + 0.02603202871978283, + 0.07596193999052048, + 0.7279540300369263, + 0.027863649651408195, + 0.021130403503775597, + 0.020418651401996613, + 0.018089085817337036, + 0.010622267611324787, + 0.020353879779577255 + ], + "credit_map": [ + 0.005621952470391989, + 0.0035847865510731936, + 0.004181770142167807, + 0.005210888106375933, + 0.010358407162129879, + 0.005778761580586433, + 0.00996172521263361, + 0.006875714752823114, + 0.004161243326961994, + 0.002815427491441369, + 0.003265694947913289, + 0.0025939655024558306, + 0.0034106303937733173, + 0.004143830388784409, + 0.0025605205446481705, + 0.003080719383433461, + 0.008725045248866081, + 0.01138785295188427, + 0.004614792764186859, + 0.006823545787483454, + 0.018783725798130035, + 0.012478304095566273, + 0.006423806771636009, + 0.006724873092025518, + 0.18950031697750092, + 0.0778425931930542, + 0.023647140711545944, + 0.028102122247219086, + 0.04386954754590988, + 0.29766005277633667, + 0.018750183284282684, + 0.048582181334495544, + 0.0039397659711539745, + 0.005772067233920097, + 0.002083668950945139, + 0.0020819329656660557, + 0.004203649237751961, + 0.0035938723012804985, + 0.003012451808899641, + 0.0031762421131134033, + 0.0037722827401012182, + 0.004646061919629574, + 0.0011184599716216326, + 0.0011207733768969774, + 0.0032446382101625204, + 0.0030652035493403673, + 0.0019886880181729794, + 0.002174297347664833, + 0.0035407315008342266, + 0.002304809633642435, + 0.0016604044940322638, + 0.002112875459715724, + 0.0021780752576887608, + 0.002393084578216076, + 0.002463757526129484, + 0.003764914581552148, + 0.0013818948063999414, + 0.0013713771477341652, + 0.0011079689720645547, + 0.0016433431301265955, + 0.005182452965527773, + 0.0028504827059805393, + 0.0018640273483470082, + 0.002687540603801608, + 0.0012757257791236043, + 0.0016381979221478105, + 0.0011890977621078491, + 0.0008615489932708442, + 0.0010822262847796082, + 0.0014323558425530791, + 0.00178623478859663, + 0.0013568809954449534, + 0.0015642819926142693, + 0.002130168955773115, + 0.001393694430589676, + 0.00414314353838563, + 0.0022580395452678204, + 0.0025233905762434006, + 0.0039926799945533276, + 0.002348483307287097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.3486754298210144 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636047434061766 + ], + "opposite_logprob_tokens": [ + -4.882636070251465 + ], + "logprob_reward": -0.007636047434061766, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636047434061766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1574748456478119, + 0.04822111874818802, + 0.21671162545681, + 0.4277452528476715, + 0.020370880141854286, + 0.026169022545218468, + 0.042504072189331055, + 0.021112103015184402, + 0.012597543187439442, + 0.027093496173620224 + ], + "credit_map": [ + 0.019954323768615723, + 0.014190673828125, + 0.01292828656733036, + 0.014643676578998566, + 0.03488895669579506, + 0.029578711837530136, + 0.019473640248179436, + 0.011816583573818207, + 0.008436636067926884, + 0.005335970316082239, + 0.005095181055366993, + 0.004091236274689436, + 0.006486870348453522, + 0.008037075400352478, + 0.0043883915059268475, + 0.006349759176373482, + 0.027798818424344063, + 0.041169919073581696, + 0.017445074394345284, + 0.028652813285589218, + 0.0317079983651638, + 0.028335172683000565, + 0.02126789465546608, + 0.02033393457531929, + 0.1150350272655487, + 0.04446692392230034, + 0.01395757682621479, + 0.014569933526217937, + 0.02562633901834488, + 0.1769840270280838, + 0.009978407062590122, + 0.02712704986333847, + 0.0050561618991196156, + 0.003067675046622753, + 0.0017152196960523725, + 0.0018576629227027297, + 0.0030521631706506014, + 0.002428085543215275, + 0.0015067760832607746, + 0.0016871370607987046, + 0.004052839707583189, + 0.004672907758504152, + 0.0019223532872274518, + 0.0021685417741537094, + 0.003003699006512761, + 0.003079022979363799, + 0.0022847859654575586, + 0.004984872415661812, + 0.006268692668527365, + 0.005890937987715006, + 0.0027071689255535603, + 0.0024876189418137074, + 0.00882750004529953, + 0.00532929040491581, + 0.004534036852419376, + 0.00645882822573185, + 0.0026617893017828465, + 0.002661807695403695, + 0.0011636283015832305, + 0.001811791444197297, + 0.00431816978380084, + 0.0029943964909762144, + 0.0031814556568861008, + 0.002319064922630787, + 0.0014893701300024986, + 0.0014422598760575056, + 0.001156575744971633, + 0.001806239248253405, + 0.0011751361889764667, + 0.0017292248085141182, + 0.0021229328121989965, + 0.0016758049605414271, + 0.0036563370376825333, + 0.0038872305303812027, + 0.0017489122692495584, + 0.003303670324385166, + 0.0036576162092387676, + 0.0035080467350780964, + 0.004345491994172335, + 0.002986192237585783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.007636047434061766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635246515274048 + ], + "opposite_logprob_tokens": [ + -3.6513524055480957 + ], + "logprob_reward": -0.02635246515274048, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635246515274048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1295553594827652, + 0.06333520263433456, + 0.07151594758033752, + 0.22660470008850098, + 0.07351676374673843, + 0.08389772474765778, + 0.09277046471834183, + 0.06550923734903336, + 0.13150554895401, + 0.06178905814886093 + ], + "credit_map": [ + 0.016610363498330116, + 0.012264751829206944, + 0.013309258036315441, + 0.013535534963011742, + 0.03382537513971329, + 0.02118532918393612, + 0.012014072388410568, + 0.006810680963099003, + 0.00933105032891035, + 0.005967655684798956, + 0.00662227114662528, + 0.0055414289236068726, + 0.007732560392469168, + 0.007654738612473011, + 0.006734571419656277, + 0.013750922866165638, + 0.008814076893031597, + 0.010175969451665878, + 0.004615092650055885, + 0.0063540260307490826, + 0.02045374922454357, + 0.011137431487441063, + 0.005857225973159075, + 0.00410836935043335, + 0.06199101731181145, + 0.023133449256420135, + 0.006412012968212366, + 0.006083737127482891, + 0.014502100646495819, + 0.09477315098047256, + 0.005735937040299177, + 0.01397328544408083, + 0.008018139749765396, + 0.021630780771374702, + 0.005576432216912508, + 0.0055729481391608715, + 0.009690281935036182, + 0.010723436251282692, + 0.006812927313148975, + 0.005491822026669979, + 0.011129344813525677, + 0.012893118895590305, + 0.0058035580441355705, + 0.007686952129006386, + 0.009793111123144627, + 0.00858716294169426, + 0.016748761758208275, + 0.011255716904997826, + 0.009948925115168095, + 0.0068028774112463, + 0.007776565384119749, + 0.006643375847488642, + 0.013195609673857689, + 0.01037913653999567, + 0.014506270177662373, + 0.02351769432425499, + 0.005457154009491205, + 0.007744739763438702, + 0.006605339236557484, + 0.007817830890417099, + 0.01072703953832388, + 0.00882736686617136, + 0.010328136384487152, + 0.008001632988452911, + 0.01235940121114254, + 0.013864217326045036, + 0.019035639241337776, + 0.01812122017145157, + 0.011326519772410393, + 0.016075946390628815, + 0.02456136979162693, + 0.0161612369120121, + 0.005430317483842373, + 0.008438399992883205, + 0.005162585526704788, + 0.008386009372770786, + 0.009603622369468212, + 0.007398987654596567, + 0.010476435534656048, + 0.006892702076584101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.02635246515274048 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning high-angle aerial photograph, as if taken from a drone, slowly tracks a lone mountain biker as he navigates a winding, narrow dirt path carved into the ridge of a vast, rugged mountain range. The low, late afternoon sun casts a warm golden hue over the epic landscape, casting long, dramatic shadows into the deep canyons below and illuminating the jagged, sun-scorched peaks in the distance. The riders, petite yet distinctive, clad in vibrant red windbreakers and black helmets, lean deliberately into the curves of the reddish-brown path, their bodies mere specks of determined movement. Their full-suspension mountain bikes, with their thick, knobby tires, steadily kick up a plume of light-colored ochre dust, capturing the whisper of golden light that follows them. The camera maintains its wide perspective, emphasizing the vastness of the wilderness and the solitary nature of the journey, creating a powerful atmosphere of awe, freedom, and the quiet endurance of humanity against the vastness of nature.", + "seed": 5575, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000096/group_9/index_188_seed_5575_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1049356371164322, + "vlm_logprob": -0.1049356371164322, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.1049356371164322, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012686500325798988 + ], + "opposite_logprob_tokens": [ + -5.012686729431152 + ], + "logprob_reward": -0.012686500325798988, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.012686500325798988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17199991643428802, + 0.04260576516389847, + 0.46537676453590393, + 0.03292742744088173, + 0.026423025876283646, + 0.018958959728479385, + 0.027919460088014603, + 0.0205525029450655, + 0.14463740587234497, + 0.04859873652458191 + ], + "credit_map": [ + 0.025571005418896675, + 0.021281294524669647, + 0.014439618214964867, + 0.0259668231010437, + 0.014363549649715424, + 0.01781408116221428, + 0.012434474192559719, + 0.040129080414772034, + 0.005617016926407814, + 0.005908186547458172, + 0.0052320281974971294, + 0.005894230213016272, + 0.005906979087740183, + 0.004418450407683849, + 0.004009506665170193, + 0.005619369447231293, + 0.08883683383464813, + 0.029369335621595383, + 0.04364946112036705, + 0.16983814537525177, + 0.0406792052090168, + 0.030939437448978424, + 0.02624516747891903, + 0.03581921011209488, + 0.002776104025542736, + 0.004119912628084421, + 0.003988593351095915, + 0.006173631642013788, + 0.005102319177240133, + 0.005076914094388485, + 0.002374851144850254, + 0.0033151034731417894, + 0.0031267704907804728, + 0.004210945684462786, + 0.0032425899989902973, + 0.003124228911474347, + 0.003358864225447178, + 0.0023266486823558807, + 0.0028470109682530165, + 0.004185968078672886, + 0.0026197396218776703, + 0.002673288807272911, + 0.0025902416091412306, + 0.0031709535978734493, + 0.0014576877001672983, + 0.0015335088828578591, + 0.00310585368424654, + 0.0018076880369335413, + 0.0033391760662198067, + 0.003578614443540573, + 0.003981680143624544, + 0.0036244369111955166, + 0.0026095155626535416, + 0.003930320497602224, + 0.003292999230325222, + 0.0035627183970063925, + 0.0025891470722854137, + 0.002670665504410863, + 0.0023755175061523914, + 0.0028788784984499216, + 0.002364310435950756, + 0.0018025782192125916, + 0.003519883146509528, + 0.0023515254724770784, + 0.01618928276002407, + 0.01484412606805563, + 0.018170827999711037, + 0.018993893638253212, + 0.01679118536412716, + 0.01562965288758278, + 0.03053184226155281, + 0.013486603274941444, + 0.005458246450871229, + 0.00369977462105453, + 0.004987337160855532, + 0.0050490498542785645, + 0.00378405605442822, + 0.010757314041256905, + 0.00898467842489481, + 0.005878283642232418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.012686500325798988 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208345673978329 + ], + "opposite_logprob_tokens": [ + -4.2642083168029785 + ], + "logprob_reward": -0.014208345673978329, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208345673978329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031446754932403564, + 0.019489174708724022, + 0.08480954170227051, + 0.013845132663846016, + 0.015425032936036587, + 0.024860944598913193, + 0.016088103875517845, + 0.08237242698669434, + 0.6325809359550476, + 0.0790819600224495 + ], + "credit_map": [ + 0.004075007513165474, + 0.004629495088011026, + 0.003657258814200759, + 0.004486686550080776, + 0.0032723613549023867, + 0.00366546050645411, + 0.0035952695179730654, + 0.004065215587615967, + 0.002063904656097293, + 0.0032263395842164755, + 0.002827657153829932, + 0.0029175723902881145, + 0.002961453516036272, + 0.0015212483704090118, + 0.0020207625348120928, + 0.001950237201526761, + 0.014880222268402576, + 0.005139176268130541, + 0.008565916679799557, + 0.03085050918161869, + 0.009070603176951408, + 0.005399984773248434, + 0.0049117072485387325, + 0.0059914193116128445, + 0.0011283084750175476, + 0.0022535466123372316, + 0.0027771403547376394, + 0.0025666907895356417, + 0.0016796467825770378, + 0.0014942751731723547, + 0.0009863930754363537, + 0.0009591313428245485, + 0.001567641505971551, + 0.002855072496458888, + 0.0026775046717375517, + 0.002306619193404913, + 0.001481744460761547, + 0.0013644665013998747, + 0.0016070473939180374, + 0.0015649363631382585, + 0.0022155954502522945, + 0.0031038112938404083, + 0.0034273976925760508, + 0.0062010628171265125, + 0.0017654673429206014, + 0.002290274715051055, + 0.0038057025521993637, + 0.0020516342483460903, + 0.0019083254737779498, + 0.0019002273911610246, + 0.00218013278208673, + 0.002403220161795616, + 0.001490953378379345, + 0.0023303653579205275, + 0.0024511839728802443, + 0.0014236947754397988, + 0.008561637252569199, + 0.01220762450248003, + 0.00593420397490263, + 0.007810086011886597, + 0.011123593896627426, + 0.00872041005641222, + 0.01951042003929615, + 0.008504451252520084, + 0.06221309304237366, + 0.06284669786691666, + 0.07271450757980347, + 0.08773157000541687, + 0.08196589350700378, + 0.058165330439805984, + 0.15164852142333984, + 0.05529528856277466, + 0.007932235486805439, + 0.006346937734633684, + 0.009097295813262463, + 0.008081593550741673, + 0.006178238429129124, + 0.013552389107644558, + 0.020116416737437248, + 0.007776850368827581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.014208345673978329 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47437989711761475 + ], + "opposite_logprob_tokens": [ + -0.9743798971176147 + ], + "logprob_reward": -0.47437989711761475, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47437989711761475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2223442643880844, + 0.03257690370082855, + 0.07473941147327423, + 0.08437921851873398, + 0.08865173161029816, + 0.04036331549286842, + 0.052223965525627136, + 0.04449150711297989, + 0.1908130943775177, + 0.16941653192043304 + ], + "credit_map": [ + 0.037370577454566956, + 0.022851722314953804, + 0.0170518159866333, + 0.034263975918293, + 0.020395906642079353, + 0.020069226622581482, + 0.01285584643483162, + 0.05748520419001579, + 0.004804810043424368, + 0.003541727317497134, + 0.0036483868025243282, + 0.004093511495739222, + 0.005269445013254881, + 0.003950429614633322, + 0.003210853785276413, + 0.004057738464325666, + 0.009150443598628044, + 0.008145458996295929, + 0.007870715111494064, + 0.010319309309124947, + 0.019099777564406395, + 0.007345275022089481, + 0.004202370997518301, + 0.00860606785863638, + 0.008179827593266964, + 0.007988285273313522, + 0.014584630727767944, + 0.022213581949472427, + 0.008469165302813053, + 0.010112393647432327, + 0.00684088421985507, + 0.005990458186715841, + 0.007727360352873802, + 0.01530113909393549, + 0.01657922938466072, + 0.015952277928590775, + 0.01003479678183794, + 0.005441909655928612, + 0.008643410168588161, + 0.008971607312560081, + 0.003945059608668089, + 0.005785445217043161, + 0.005096224136650562, + 0.006868612486869097, + 0.0035626287572085857, + 0.002796364016830921, + 0.009260528720915318, + 0.0030484546441584826, + 0.004743108060210943, + 0.008512435480952263, + 0.007239180151373148, + 0.0088174007833004, + 0.0045053912326693535, + 0.0067514036782085896, + 0.005149639677256346, + 0.006505407392978668, + 0.005398142617195845, + 0.006671191658824682, + 0.004396270960569382, + 0.005460852291435003, + 0.005423619877547026, + 0.006147458218038082, + 0.006710434798151255, + 0.004283537156879902, + 0.02005700394511223, + 0.01651097647845745, + 0.020777881145477295, + 0.02811281569302082, + 0.01936684362590313, + 0.022221533581614494, + 0.04575510695576668, + 0.0180109441280365, + 0.013035306707024574, + 0.014142405241727829, + 0.01455118041485548, + 0.015578819438815117, + 0.014862381853163242, + 0.02835923805832863, + 0.04745452105998993, + 0.02143269032239914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.47437989711761475 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028198501095175743 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.0028198501095175743, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028198501095175743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23899959027767181, + 0.08587487787008286, + 0.531670093536377, + 0.011472777463495731, + 0.0224834606051445, + 0.030561190098524094, + 0.017266085371375084, + 0.020288273692131042, + 0.018484631553292274, + 0.02289898507297039 + ], + "credit_map": [ + 0.030045611783862114, + 0.034300368279218674, + 0.013685540296137333, + 0.04786407947540283, + 0.020291578024625778, + 0.03314375877380371, + 0.013726932927966118, + 0.04594173654913902, + 0.013692210428416729, + 0.009990205988287926, + 0.009811419993638992, + 0.01294885016977787, + 0.012278230860829353, + 0.011438777670264244, + 0.007189012598246336, + 0.008526180870831013, + 0.09605031460523605, + 0.03472714126110077, + 0.05097387358546257, + 0.19237199425697327, + 0.05031411349773407, + 0.035726215690374374, + 0.031020168215036392, + 0.040486305952072144, + 0.0011303715873509645, + 0.0016568731516599655, + 0.0014049053424969316, + 0.0021682351361960173, + 0.001626966055482626, + 0.00183869490865618, + 0.000642543425783515, + 0.001004189602099359, + 0.0021790515165776014, + 0.003136174753308296, + 0.003170086070895195, + 0.003796264762058854, + 0.0018218824407085776, + 0.002226183656603098, + 0.0024717736523598433, + 0.003682047827169299, + 0.003495087381452322, + 0.0028822405729442835, + 0.006576377432793379, + 0.008307283744215965, + 0.0018169283866882324, + 0.0018449468770995736, + 0.002057347446680069, + 0.003580980934202671, + 0.0019414162961766124, + 0.0022567862179130316, + 0.0025390491355210543, + 0.0025245589204132557, + 0.0016099180793389678, + 0.0017562505090609193, + 0.0029024635441601276, + 0.00173564616125077, + 0.0028262652922421694, + 0.0026091637555509806, + 0.001833154703490436, + 0.0022188627626746893, + 0.00230201194062829, + 0.002112393965944648, + 0.004155535716563463, + 0.0022308884654194117, + 0.0022243517450988293, + 0.0022347946651279926, + 0.001566407736390829, + 0.0015912508824840188, + 0.0031178416684269905, + 0.0022333944216370583, + 0.003162648994475603, + 0.0023539417888969183, + 0.0016230344772338867, + 0.0017187825869768858, + 0.002394072711467743, + 0.002444248180836439, + 0.0025536343455314636, + 0.0023634901735931635, + 0.0068031600676476955, + 0.0029985650908201933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0028198501095175743 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle aerial perspective, as if taken from a drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone mountain biker visible navigating a winding, narrow dirt path on a mountain ridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the rider and bike wearing vibrant red windbreakers and black helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low, late afternoon sun casting a warm golden hue and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly track the biker while maintaining a wide perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583592355251312 + ], + "opposite_logprob_tokens": [ + -3.8955836296081543 + ], + "logprob_reward": -0.020583592355251312, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583592355251312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11529436707496643, + 0.0486052967607975, + 0.22880589962005615, + 0.029234373942017555, + 0.022177092730998993, + 0.026628144085407257, + 0.015564589761197567, + 0.09804916381835938, + 0.3448620140552521, + 0.07077906280755997 + ], + "credit_map": [ + 0.016453146934509277, + 0.01365014910697937, + 0.008617023006081581, + 0.01981057971715927, + 0.011903706006705761, + 0.01562914438545704, + 0.008678177371621132, + 0.02055244892835617, + 0.0056933388113975525, + 0.007694851141422987, + 0.0071060191839933395, + 0.006994253024458885, + 0.0067206318490207195, + 0.004034160170704126, + 0.00510744471102953, + 0.005254602525383234, + 0.04376567527651787, + 0.015021095983684063, + 0.020318131893873215, + 0.07968100160360336, + 0.01949823461472988, + 0.017102759331464767, + 0.012742545455694199, + 0.020676443353295326, + 0.0024401939008384943, + 0.0031913621351122856, + 0.0034787808544933796, + 0.0061837369576096535, + 0.004237132612615824, + 0.005170900374650955, + 0.0016188514418900013, + 0.0029134154319763184, + 0.001949378172867, + 0.0021251721773296595, + 0.004270576406270266, + 0.003371004480868578, + 0.0014930632896721363, + 0.002406506100669503, + 0.0023262216709554195, + 0.00423516845330596, + 0.0025335527025163174, + 0.002984191756695509, + 0.005365512799471617, + 0.006578720174729824, + 0.0016955784521996975, + 0.0019041030900552869, + 0.003187700407579541, + 0.002378785517066717, + 0.0019322449807077646, + 0.0018098613945767283, + 0.0024852589704096317, + 0.001858962350524962, + 0.0015727992868050933, + 0.0021567572839558125, + 0.0022575624752789736, + 0.0014911422040313482, + 0.010510500520467758, + 0.012651454657316208, + 0.007418641820549965, + 0.00919717364013195, + 0.012309776619076729, + 0.00921513419598341, + 0.026095421984791756, + 0.010651059448719025, + 0.03514740616083145, + 0.03375941887497902, + 0.041360124945640564, + 0.04581539332866669, + 0.04275159165263176, + 0.03154769912362099, + 0.08345479518175125, + 0.031025603413581848, + 0.0057416134513914585, + 0.0052457391284406185, + 0.005351574160158634, + 0.009084598161280155, + 0.00763770891353488, + 0.008435631170868874, + 0.019754312932491302, + 0.009527886286377907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.020583592355251312 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_0/index_tensor([6375, 6375])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_0/index_tensor([6375, 6375])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1b54ccc2d23a32f88aeabd791565e7a5d3b88d26 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_0/index_tensor([6375, 6375])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful shot begins with a close-up overhead shot of a distorted face. The Six-Fingered Qin Demon raises his head, roaring silently into the swirling blood-red clouds. His long, untidy black hair whips wildly in the unearthly wind, every muscle on his face tensed with utter fury, veins bulging from his forehead. The camera then smoothly pans upward and back, revealing his entire form—clothed in a tattered dark crimson robe, he stands alone atop the ruins of a vast ancient building. His gaze is frantic, as his right hand, rumored to have an extra finger, frantically strums the strings of an ancient black guqin before him at a speed barely tangible. Each plucking unleashes visible, blade-like, white sonic waves. These ripples of energy tear through the surrounding air, instantly severing the flying dust and debris, a perfect illustration of the aesthetics of destructive violence. As the camera continues to rise, a despairing panorama unfolds before his eyes: the entire mountaintop beneath his feet is collapsing, huge rocks and broken ground are constantly falling into the bottomless dark abyss below, and the whole world is heading towards destruction in his furious piano music, creating a grand and tragic atmosphere of the coming of doomsday.", + "seed": 22232, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_0/index_6375_seed_22232_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03510430604219437, + "vlm_logprob": -0.03510430604219437, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.03510430604219437, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04586203023791313 + ], + "opposite_logprob_tokens": [ + -3.1708619594573975 + ], + "logprob_reward": -0.04586203023791313, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04586203023791313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09260164946317673, + 0.11476731300354004, + 0.08004393428564072, + 0.09446756541728973, + 0.04772663488984108, + 0.09401647746562958, + 0.12686611711978912, + 0.17146077752113342, + 0.08013884723186493, + 0.09791062772274017 + ], + "credit_map": [ + 0.012057540006935596, + 0.01267382875084877, + 0.0068294391967356205, + 0.012126103043556213, + 0.01556764543056488, + 0.010517320595681667, + 0.011497239582240582, + 0.011332549154758453, + 0.01651068590581417, + 0.010228992439806461, + 0.016667842864990234, + 0.014252547174692154, + 0.013146837241947651, + 0.012586892582476139, + 0.01427122950553894, + 0.017102288082242012, + 0.012193981558084488, + 0.007438845932483673, + 0.008091860450804234, + 0.00872149970382452, + 0.010223370976746082, + 0.016793938353657722, + 0.007780743762850761, + 0.008799699135124683, + 0.011090412735939026, + 0.00774467783048749, + 0.010091929696500301, + 0.011274435557425022, + 0.011479347944259644, + 0.01220120582729578, + 0.017190929502248764, + 0.013394626788794994, + 0.003498399630188942, + 0.005635319277644157, + 0.006719077471643686, + 0.004808907862752676, + 0.007630548439919949, + 0.007434970699250698, + 0.005636636167764664, + 0.006362779065966606, + 0.004979295190423727, + 0.005599379073828459, + 0.007027078419923782, + 0.004867430776357651, + 0.006321622524410486, + 0.030689628794789314, + 0.005272261798381805, + 0.029259786009788513, + 0.008874516934156418, + 0.020664388313889503, + 0.0106283500790596, + 0.008203449659049511, + 0.010810526087880135, + 0.034397147595882416, + 0.006005764007568359, + 0.027281979098916054, + 0.015908747911453247, + 0.014724893495440483, + 0.04236002638936043, + 0.01758294552564621, + 0.014236140064895153, + 0.03276419639587402, + 0.012517486698925495, + 0.021366355940699577, + 0.006486434489488602, + 0.011471082456409931, + 0.010354386642575264, + 0.011720207519829273, + 0.01265752874314785, + 0.010370303876698017, + 0.008329818956553936, + 0.008749102242290974, + 0.005127998068928719, + 0.008377267979085445, + 0.014766093343496323, + 0.01464836485683918, + 0.016387291252613068, + 0.007029329426586628, + 0.012083939276635647, + 0.019490353763103485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.04586203023791313 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023301467299461365 + ], + "opposite_logprob_tokens": [ + -3.773301362991333 + ], + "logprob_reward": -0.023301467299461365, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023301467299461365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0630665197968483, + 0.23306596279144287, + 0.028725462034344673, + 0.015093271620571613, + 0.02120763435959816, + 0.11613211035728455, + 0.06218387931585312, + 0.28421610593795776, + 0.10165265947580338, + 0.07465633004903793 + ], + "credit_map": [ + 0.007580541539937258, + 0.006862131413072348, + 0.008145707659423351, + 0.010525739751756191, + 0.007645297329872847, + 0.0069807907566428185, + 0.008482600562274456, + 0.0068437159061431885, + 0.018767226487398148, + 0.02492576465010643, + 0.01863139122724533, + 0.046358004212379456, + 0.027025865390896797, + 0.026256756857037544, + 0.028723619878292084, + 0.042377348989248276, + 0.0035001966170966625, + 0.0037498860619962215, + 0.0034713682252913713, + 0.0038161040283739567, + 0.0039106798358261585, + 0.0044563389383256435, + 0.003340427065268159, + 0.0024804642889648676, + 0.0012808102183043957, + 0.0017713481793180108, + 0.0017981429118663073, + 0.0015962067991495132, + 0.0023821089416742325, + 0.002433938905596733, + 0.001405753311701119, + 0.0024249637499451637, + 0.0018284067045897245, + 0.00308147631585598, + 0.0024633705615997314, + 0.001997103216126561, + 0.0035838503390550613, + 0.0027142944745719433, + 0.0029074745252728462, + 0.002631660783663392, + 0.0057259611785411835, + 0.015070311725139618, + 0.006615848280489445, + 0.004640510305762291, + 0.007898538373410702, + 0.0373905748128891, + 0.0030026286840438843, + 0.03578775003552437, + 0.007248536217957735, + 0.015775689855217934, + 0.007372490130364895, + 0.004015602171421051, + 0.01049345638602972, + 0.008223170414566994, + 0.003537212498486042, + 0.005517731420695782, + 0.02875705063343048, + 0.027164176106452942, + 0.08127857744693756, + 0.033851947635412216, + 0.02658236026763916, + 0.035790957510471344, + 0.022911084815859795, + 0.027879994362592697, + 0.007143677677959204, + 0.008960907347500324, + 0.021239085122942924, + 0.012470591813325882, + 0.01355232298374176, + 0.012236673384904861, + 0.01089879497885704, + 0.015150615945458412, + 0.005412882659584284, + 0.013540247455239296, + 0.007741267327219248, + 0.01159769855439663, + 0.013092679902911186, + 0.006041842512786388, + 0.00997567642480135, + 0.007254042662680149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.023301467299461365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818133518099785 + ], + "opposite_logprob_tokens": [ + -4.018181324005127 + ], + "logprob_reward": -0.01818133518099785, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818133518099785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06597123295068741, + 0.08557244390249252, + 0.07661938667297363, + 0.042828407138586044, + 0.05396633967757225, + 0.11178463697433472, + 0.17323261499404907, + 0.07569283992052078, + 0.14386358857154846, + 0.17046859860420227 + ], + "credit_map": [ + 0.007259724661707878, + 0.0077661327086389065, + 0.0041633001528680325, + 0.01099788025021553, + 0.008589021861553192, + 0.006549265701323748, + 0.012778396718204021, + 0.007867506705224514, + 0.01009947620332241, + 0.009227268397808075, + 0.008597013540565968, + 0.014095424674451351, + 0.009162284433841705, + 0.009300272911787033, + 0.012474113143980503, + 0.012616584077477455, + 0.009740347974002361, + 0.013547571375966072, + 0.006291941273957491, + 0.006763522978872061, + 0.010120409540832043, + 0.0154064716771245, + 0.005638648755848408, + 0.009110466577112675, + 0.00349844410084188, + 0.006943340878933668, + 0.0033071849029511213, + 0.004960644990205765, + 0.00794216338545084, + 0.006045330781489611, + 0.005603037774562836, + 0.004528256598860025, + 0.004059013444930315, + 0.005955762695521116, + 0.01408852357417345, + 0.0039018504321575165, + 0.006804642267525196, + 0.007973100058734417, + 0.005991144571453333, + 0.0051922984421253204, + 0.0058294013142585754, + 0.007673380896449089, + 0.006983037572354078, + 0.005949764046818018, + 0.008905638940632343, + 0.03726252540946007, + 0.0053194076754152775, + 0.03386146202683449, + 0.013175008818507195, + 0.028282899409532547, + 0.009158620610833168, + 0.008302162401378155, + 0.01811508648097515, + 0.051539789885282516, + 0.005160896573215723, + 0.039498135447502136, + 0.00600471394136548, + 0.005248679779469967, + 0.011236532591283321, + 0.006613872479647398, + 0.005236258264631033, + 0.02212674543261528, + 0.004821895156055689, + 0.014404135756194592, + 0.010646313428878784, + 0.012516817077994347, + 0.0340498648583889, + 0.018028466030955315, + 0.017862459644675255, + 0.014691105112433434, + 0.01231626607477665, + 0.023752281442284584, + 0.010941863991320133, + 0.025139737874269485, + 0.01611497439444065, + 0.013877368532121181, + 0.03601188585162163, + 0.011252177879214287, + 0.015252995304763317, + 0.04187757894396782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.01818133518099785 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000414282083511 + ], + "opposite_logprob_tokens": [ + -2.6950042247772217 + ], + "logprob_reward": -0.07000414282083511, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000414282083511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06156191602349281, + 0.03733973577618599, + 0.018586428835988045, + 0.022675905376672745, + 0.030384879559278488, + 0.0802869126200676, + 0.24508439004421234, + 0.09735064953565598, + 0.30777299404144287, + 0.09895619004964828 + ], + "credit_map": [ + 0.0066224681213498116, + 0.006572321522980928, + 0.006641658488661051, + 0.01090503390878439, + 0.007590966764837503, + 0.0062562949024140835, + 0.010556780733168125, + 0.006416392512619495, + 0.007624961901456118, + 0.0026109726168215275, + 0.007016249932348728, + 0.0047853668220341206, + 0.0029590290505439043, + 0.005151428282260895, + 0.0036088814958930016, + 0.0035828480031341314, + 0.002169741317629814, + 0.0022915364243090153, + 0.0028159795328974724, + 0.0020695910789072514, + 0.0024227979592978954, + 0.002500526374205947, + 0.001964957220479846, + 0.0023512989282608032, + 0.004739069379866123, + 0.002430078573524952, + 0.0026392692234367132, + 0.0018123779445886612, + 0.002114791888743639, + 0.0035538682714104652, + 0.001975322375074029, + 0.0034111281856894493, + 0.0026981798000633717, + 0.0034864589106291533, + 0.0060014137998223305, + 0.0027702858205884695, + 0.0024882573634386063, + 0.005155143793672323, + 0.0023224749602377415, + 0.005462664179503918, + 0.004531181417405605, + 0.015704523772001266, + 0.0058680931106209755, + 0.006426284555345774, + 0.010686593130230904, + 0.017719093710184097, + 0.0048078629188239574, + 0.01454327441751957, + 0.014554177410900593, + 0.022648682817816734, + 0.013873409479856491, + 0.011631087400019169, + 0.020418375730514526, + 0.09082164615392685, + 0.006354783661663532, + 0.06478223204612732, + 0.009097486734390259, + 0.012037200853228569, + 0.020555691793560982, + 0.011655514128506184, + 0.011603015474975109, + 0.012284199707210064, + 0.009249914437532425, + 0.010867626406252384, + 0.023752538487315178, + 0.026454802602529526, + 0.07025963813066483, + 0.04048949480056763, + 0.03575592115521431, + 0.035573333501815796, + 0.02380193956196308, + 0.05168531835079193, + 0.008973383344709873, + 0.016256706789135933, + 0.00964705552905798, + 0.009733553044497967, + 0.017401868477463722, + 0.011565585620701313, + 0.012037107720971107, + 0.013340931385755539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.07000414282083511 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018172554671764374 + ], + "opposite_logprob_tokens": [ + -4.018172740936279 + ], + "logprob_reward": -0.018172554671764374, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018172554671764374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0499538779258728, + 0.06337292492389679, + 0.028831908479332924, + 0.01954009011387825, + 0.024533778429031372, + 0.2448335736989975, + 0.2043185830116272, + 0.07914300262928009, + 0.15628276765346527, + 0.12918952107429504 + ], + "credit_map": [ + 0.0060304985381662846, + 0.00506163202226162, + 0.005836171563714743, + 0.007763800676912069, + 0.005595841445028782, + 0.0072230082005262375, + 0.007527605164796114, + 0.004915317520499229, + 0.011417501606047153, + 0.0055465116165578365, + 0.011232406832277775, + 0.008517025038599968, + 0.004887428134679794, + 0.00771284569054842, + 0.007654281333088875, + 0.006404915824532509, + 0.004190679639577866, + 0.003141961758956313, + 0.0026582786813378334, + 0.0023954270873218775, + 0.0032886427361518145, + 0.006462136749178171, + 0.0030450939666479826, + 0.003649687161669135, + 0.002518882043659687, + 0.0025339119601994753, + 0.002484518103301525, + 0.002062432235106826, + 0.0029737818986177444, + 0.0022363970056176186, + 0.0019827682990580797, + 0.0027473957743495703, + 0.00223330850712955, + 0.0026131998747587204, + 0.007157489657402039, + 0.0014837442431598902, + 0.0020492251496762037, + 0.004013716243207455, + 0.0017150972271338105, + 0.003267996944487095, + 0.012932592071592808, + 0.03128350153565407, + 0.013852902688086033, + 0.009902683086693287, + 0.01650000736117363, + 0.07793150842189789, + 0.007177041377872229, + 0.07525329291820526, + 0.011367525905370712, + 0.015788322314620018, + 0.013210766017436981, + 0.009463014081120491, + 0.015130882151424885, + 0.07463306188583374, + 0.007473815232515335, + 0.057251181453466415, + 0.007712986320257187, + 0.007835634984076023, + 0.018414627760648727, + 0.008791196160018444, + 0.007727320771664381, + 0.012936640530824661, + 0.00726684695109725, + 0.008457738906145096, + 0.012106884270906448, + 0.017342373728752136, + 0.025995658710598946, + 0.019697511568665504, + 0.019757959991693497, + 0.018909627571702003, + 0.01813151128590107, + 0.024341220036149025, + 0.010706410743296146, + 0.018707487732172012, + 0.012150256894528866, + 0.012090102769434452, + 0.02312343753874302, + 0.010523514822125435, + 0.015012639574706554, + 0.026875648647546768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.018172554671764374 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful shot begins with a close-up overhead shot of a distorted face. The Six-Fingered Qin Demon raises his head, roaring silently into the swirling blood-red clouds. His long, untidy black hair whips wildly in the unearthly wind, every muscle on his face tensed with utter fury, veins bulging from his forehead. The camera then smoothly pans upward and back, revealing his entire form—clothed in a tattered dark crimson robe, he stands alone atop the ruins of a vast ancient building. His gaze is frantic, as his right hand, rumored to have an extra finger, frantically strums the strings of an ancient black guqin before him at a speed barely tangible. Each plucking unleashes visible, blade-like, white sonic waves. These ripples of energy tear through the surrounding air, instantly severing the flying dust and debris, a perfect illustration of the aesthetics of destructive violence. As the camera continues to rise, a despairing panorama unfolds before his eyes: the entire mountaintop beneath his feet is collapsing, huge rocks and broken ground are constantly falling into the bottomless dark abyss below, and the whole world is heading towards destruction in his furious piano music, creating a grand and tragic atmosphere of the coming of doomsday.", + "seed": 22233, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_0/index_6375_seed_22233_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.028946377988904714, + "vlm_logprob": -0.028946377988904714, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.028946377988904714, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.039826855063438416 + ], + "opposite_logprob_tokens": [ + -3.2898268699645996 + ], + "logprob_reward": -0.039826855063438416, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.039826855063438416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16572710871696472, + 0.6096091866493225, + 0.05372793227434158, + 0.022926144301891327, + 0.030600281432271004, + 0.037140876054763794, + 0.024563487619161606, + 0.012214195914566517, + 0.026177261024713516, + 0.01731351763010025 + ], + "credit_map": [ + 0.010466163977980614, + 0.008048756048083305, + 0.046611104160547256, + 0.015657149255275726, + 0.009973235428333282, + 0.009086892940104008, + 0.050577398389577866, + 0.015306408517062664, + 0.05852997675538063, + 0.06019793823361397, + 0.056373488157987595, + 0.07819709181785583, + 0.10628625005483627, + 0.05348522216081619, + 0.1308368295431137, + 0.06570235639810562, + 0.004212388768792152, + 0.0049858856946229935, + 0.007019130978733301, + 0.010645227506756783, + 0.0076682972721755505, + 0.0054437266662716866, + 0.007503871340304613, + 0.006249402184039354, + 0.0022417595610022545, + 0.0029634551610797644, + 0.0016380198067054152, + 0.003753483295440674, + 0.0032343710772693157, + 0.0034772309008985758, + 0.002954235067591071, + 0.002663589781150222, + 0.0035442847292870283, + 0.003081261646002531, + 0.005022967699915171, + 0.0038070986047387123, + 0.0034419617149978876, + 0.004808505065739155, + 0.0035457678604871035, + 0.003348434576764703, + 0.005706723313778639, + 0.003076741239055991, + 0.005742412526160479, + 0.002843305468559265, + 0.0032331899274140596, + 0.006621610838919878, + 0.005529794376343489, + 0.004387098364531994, + 0.002392409136518836, + 0.002871386706829071, + 0.0028680984396487474, + 0.0033682682551443577, + 0.0018739141523838043, + 0.003947490826249123, + 0.0033218569587916136, + 0.003920062445104122, + 0.002441291231662035, + 0.001112767611630261, + 0.0010253366781398654, + 0.0017097860109061003, + 0.0012496105628088117, + 0.0014781826175749302, + 0.0016342147719115019, + 0.0015630061971023679, + 0.0029954586643725634, + 0.003573824418708682, + 0.0026764969807118177, + 0.0030151386745274067, + 0.003599062329158187, + 0.002654006704688072, + 0.0035065223928540945, + 0.004156752023845911, + 0.00155049713794142, + 0.0019069131230935454, + 0.002333286916837096, + 0.002813370432704687, + 0.0033675602171570063, + 0.0015952428802847862, + 0.0017382166115567088, + 0.002008429728448391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.039826855063438416 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647999536246061 + ], + "opposite_logprob_tokens": [ + -5.379648208618164 + ], + "logprob_reward": -0.004647999536246061, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647999536246061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3854701519012451, + 0.09025321900844574, + 0.10326945036649704, + 0.08809961378574371, + 0.08357764780521393, + 0.029739368706941605, + 0.07144510746002197, + 0.06768905371427536, + 0.03584260120987892, + 0.04461381211876869 + ], + "credit_map": [ + 0.031051667407155037, + 0.01244445238262415, + 0.11450690776109695, + 0.03741869702935219, + 0.020886074751615524, + 0.018499797210097313, + 0.11409992724657059, + 0.03656259551644325, + 0.008423958905041218, + 0.009622734040021896, + 0.008620924316346645, + 0.011536743491888046, + 0.014886053279042244, + 0.010562600567936897, + 0.017501158639788628, + 0.009099042043089867, + 0.011931611225008965, + 0.007977990433573723, + 0.011508957482874393, + 0.01123169157654047, + 0.015047339722514153, + 0.021890932694077492, + 0.012435470707714558, + 0.011245455592870712, + 0.007680089678615332, + 0.015014315955340862, + 0.0057722218334674835, + 0.01354084350168705, + 0.012191272340714931, + 0.01227384340018034, + 0.013765724375844002, + 0.007861303165555, + 0.010229182429611683, + 0.009098314680159092, + 0.012879925779998302, + 0.008049913682043552, + 0.008793434128165245, + 0.013736831024289131, + 0.007219941355288029, + 0.013570109382271767, + 0.0030018596444278955, + 0.0020052914042025805, + 0.004917457699775696, + 0.004469961393624544, + 0.002253297483548522, + 0.002806369448080659, + 0.006022016983479261, + 0.0042631132528185844, + 0.007818209938704967, + 0.008553869090974331, + 0.009181609377264977, + 0.008810956962406635, + 0.006502761039882898, + 0.013025270774960518, + 0.007748115342110395, + 0.009804317727684975, + 0.009244905784726143, + 0.008413608185946941, + 0.007978332228958607, + 0.008938560262322426, + 0.008998523466289043, + 0.008016438223421574, + 0.008031259290874004, + 0.008067435584962368, + 0.0035576445516198874, + 0.00566288223490119, + 0.004163963720202446, + 0.0039003847632557154, + 0.006925374735146761, + 0.0036879596300423145, + 0.003875957801938057, + 0.004068434704095125, + 0.005358356982469559, + 0.004908038303256035, + 0.005518113728612661, + 0.006140857469290495, + 0.00561068719252944, + 0.004967567045241594, + 0.0071051158010959625, + 0.0050050741992890835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.004647999536246061 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195887371897697 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014195887371897697, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195887371897697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27074873447418213, + 0.31386473774909973, + 0.10642703622579575, + 0.04325157776474953, + 0.05390400066971779, + 0.04867178574204445, + 0.080252505838871, + 0.024573568254709244, + 0.03167801350355148, + 0.026628008112311363 + ], + "credit_map": [ + 0.021893780678510666, + 0.01906345598399639, + 0.07276155054569244, + 0.026550788432359695, + 0.016065534204244614, + 0.014313176274299622, + 0.07316350191831589, + 0.026936959475278854, + 0.03078879788517952, + 0.03422456979751587, + 0.02841886878013611, + 0.04070707783102989, + 0.05738718807697296, + 0.027506932616233826, + 0.061712417751550674, + 0.03311888128519058, + 0.009438230656087399, + 0.015071656554937363, + 0.014484209939837456, + 0.013413695618510246, + 0.015085776336491108, + 0.013374910689890385, + 0.0121754240244627, + 0.01338313426822424, + 0.003862924873828888, + 0.006734590977430344, + 0.0034204241819679737, + 0.007678940426558256, + 0.006231812294572592, + 0.004402146674692631, + 0.006744838785380125, + 0.004175900015980005, + 0.0069331699050962925, + 0.004540706519037485, + 0.008043128997087479, + 0.005957767833024263, + 0.006684035994112492, + 0.009272294119000435, + 0.006171999964863062, + 0.006300895940512419, + 0.008985387161374092, + 0.0050551556050777435, + 0.006026196759194136, + 0.00456876028329134, + 0.005359061062335968, + 0.0089274225756526, + 0.003920416347682476, + 0.005829386878758669, + 0.007575363852083683, + 0.008892307057976723, + 0.012014019303023815, + 0.014470609836280346, + 0.008692612871527672, + 0.008488930761814117, + 0.013512439094483852, + 0.00660622026771307, + 0.0037679646629840136, + 0.002701736520975828, + 0.003581201657652855, + 0.00287582422606647, + 0.002931303111836314, + 0.0026795435696840286, + 0.0030219245236366987, + 0.00301407091319561, + 0.00354552548378706, + 0.0039397249929606915, + 0.003345494857057929, + 0.0027640233747661114, + 0.005190686322748661, + 0.004195332061499357, + 0.0034996909089386463, + 0.005197534803301096, + 0.0036176845896989107, + 0.0029410223942250013, + 0.002280566841363907, + 0.0034896277356892824, + 0.004603599663823843, + 0.0031378567218780518, + 0.002477167872712016, + 0.004080480430275202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.014195887371897697 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000603526830673 + ], + "opposite_logprob_tokens": [ + -2.6950061321258545 + ], + "logprob_reward": -0.07000603526830673, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000603526830673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04466497525572777, + 0.458963543176651, + 0.14384448528289795, + 0.030711520463228226, + 0.03075387142598629, + 0.039234962314367294, + 0.13214699923992157, + 0.03551824390888214, + 0.0350959487259388, + 0.0490654781460762 + ], + "credit_map": [ + 0.0022496844176203012, + 0.006862923037260771, + 0.008378139697015285, + 0.0057611470110714436, + 0.00317871174775064, + 0.0028729478362947702, + 0.010401446372270584, + 0.004959967453032732, + 0.04340895637869835, + 0.04287232831120491, + 0.04096684977412224, + 0.06371480226516724, + 0.07528330385684967, + 0.03943006694316864, + 0.10103849321603775, + 0.05224870890378952, + 0.008642653003334999, + 0.019175993278622627, + 0.012362132780253887, + 0.033008381724357605, + 0.01556823868304491, + 0.013451311737298965, + 0.02948770299553871, + 0.01214805617928505, + 0.002718426287174225, + 0.002401498379185796, + 0.0014788865810260177, + 0.0044450946152210236, + 0.0024720330256968737, + 0.009362869895994663, + 0.003518227022141218, + 0.00431448221206665, + 0.005440654698759317, + 0.002268821932375431, + 0.0056054494343698025, + 0.0030583885964006186, + 0.0025063171051442623, + 0.0045796833001077175, + 0.00263862288556993, + 0.004655930679291487, + 0.00635400740429759, + 0.0029975271318107843, + 0.0060994308441877365, + 0.003173271892592311, + 0.0033520497381687164, + 0.006384822074323893, + 0.005567930173128843, + 0.005305921193212271, + 0.010901075787842274, + 0.00849379226565361, + 0.012048458680510521, + 0.011790622025728226, + 0.007295173592865467, + 0.03670299053192139, + 0.012772049754858017, + 0.03214282914996147, + 0.005990745034068823, + 0.004444997292011976, + 0.0050481767393648624, + 0.004224443342536688, + 0.0049808393232524395, + 0.004003601614385843, + 0.0036470117047429085, + 0.0031784235034137964, + 0.0037777770776301622, + 0.004585828632116318, + 0.004044389817863703, + 0.004215980414301157, + 0.005601349752396345, + 0.0031886151991784573, + 0.0060861012898385525, + 0.0035959044471383095, + 0.005000019446015358, + 0.00587442284449935, + 0.005141316447407007, + 0.008217070251703262, + 0.007696080021560192, + 0.004579451400786638, + 0.008805220015347004, + 0.003751893062144518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 6 + ], + "selected_reward": -0.07000603526830673 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016055112704634666 + ], + "opposite_logprob_tokens": [ + -4.141055107116699 + ], + "logprob_reward": -0.016055112704634666, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016055112704634666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40956270694732666, + 0.06049385666847229, + 0.10393934696912766, + 0.05906187370419502, + 0.07513874024152756, + 0.04441601783037186, + 0.07516536861658096, + 0.05857580527663231, + 0.0712636187672615, + 0.042382705956697464 + ], + "credit_map": [ + 0.03420419618487358, + 0.01540465373545885, + 0.1235436499118805, + 0.03648564964532852, + 0.022381804883480072, + 0.01703895814716816, + 0.12231650203466415, + 0.038187239319086075, + 0.005685559008270502, + 0.007961161434650421, + 0.0076307617127895355, + 0.010258989408612251, + 0.008614408783614635, + 0.008535346947610378, + 0.006426801439374685, + 0.005380825139582157, + 0.006246672943234444, + 0.009966432116925716, + 0.016988348215818405, + 0.02245534211397171, + 0.01101432740688324, + 0.008412430062890053, + 0.01883092150092125, + 0.01002486702054739, + 0.0058858455158770084, + 0.0036223959177732468, + 0.0024549535010010004, + 0.006285862531512976, + 0.004792467691004276, + 0.021220333874225616, + 0.005361555144190788, + 0.009438460692763329, + 0.010592243634164333, + 0.009493516758084297, + 0.009467379190027714, + 0.008012871257960796, + 0.010074585676193237, + 0.008622662164270878, + 0.007701575290411711, + 0.011173905804753304, + 0.005725616589188576, + 0.003824559273198247, + 0.006492533255368471, + 0.004742275923490524, + 0.004173848312348127, + 0.006399052683264017, + 0.007789469324052334, + 0.0052686589770019054, + 0.008515939116477966, + 0.010053071193397045, + 0.010803506709635258, + 0.012829063460230827, + 0.007415717002004385, + 0.006318169645965099, + 0.011960162781178951, + 0.007269741967320442, + 0.007068514823913574, + 0.007321627344936132, + 0.007692166604101658, + 0.009191161021590233, + 0.0076143485493958, + 0.006096083205193281, + 0.008203057572245598, + 0.005388846155256033, + 0.005604903679341078, + 0.008557300083339214, + 0.007910152897238731, + 0.011708607897162437, + 0.009669740684330463, + 0.0075639099813997746, + 0.012604953721165657, + 0.007644052151590586, + 0.005962648428976536, + 0.004225487820804119, + 0.00691163819283247, + 0.0047401306219398975, + 0.005121445748955011, + 0.00462270388379693, + 0.004463216755539179, + 0.006335437763482332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.016055112704634666 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful shot begins with a close-up overhead shot of a distorted face. The Six-Fingered Qin Demon raises his head, roaring silently into the swirling blood-red clouds. His long, untidy black hair whips wildly in the unearthly wind, every muscle on his face tensed with utter fury, veins bulging from his forehead. The camera then smoothly pans upward and back, revealing his entire form—clothed in a tattered dark crimson robe, he stands alone atop the ruins of a vast ancient building. His gaze is frantic, as his right hand, rumored to have an extra finger, frantically strums the strings of an ancient black guqin before him at a speed barely tangible. Each plucking unleashes visible, blade-like, white sonic waves. These ripples of energy tear through the surrounding air, instantly severing the flying dust and debris, a perfect illustration of the aesthetics of destructive violence. As the camera continues to rise, a despairing panorama unfolds before his eyes: the entire mountaintop beneath his feet is collapsing, huge rocks and broken ground are constantly falling into the bottomless dark abyss below, and the whole world is heading towards destruction in his furious piano music, creating a grand and tragic atmosphere of the coming of doomsday.", + "seed": 22234, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_0/index_6375_seed_22234_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028923108614981173, + "vlm_logprob": -0.028923108614981173, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.028923108614981173, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0459294356405735 + ], + "opposite_logprob_tokens": [ + -3.170929431915283 + ], + "logprob_reward": -0.0459294356405735, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0459294356405735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16141878068447113, + 0.05927131697535515, + 0.3188678026199341, + 0.11897817999124527, + 0.03060254268348217, + 0.02577192708849907, + 0.05597661808133125, + 0.06185650825500488, + 0.0800551101565361, + 0.08720121532678604 + ], + "credit_map": [ + 0.016403978690505028, + 0.014097274281084538, + 0.022962942719459534, + 0.017245173454284668, + 0.01862359419465065, + 0.023548226803541183, + 0.02529696747660637, + 0.023240623995661736, + 0.006321740336716175, + 0.005517425946891308, + 0.007179517298936844, + 0.007056230679154396, + 0.005953324493020773, + 0.013166454620659351, + 0.008558554574847221, + 0.0055180625058710575, + 0.03722721338272095, + 0.020723562687635422, + 0.04446881264448166, + 0.030808256939053535, + 0.02712574414908886, + 0.04679107666015625, + 0.051053017377853394, + 0.060670118778944016, + 0.012844951823353767, + 0.01406735647469759, + 0.014956123195588589, + 0.01245385967195034, + 0.02261122316122055, + 0.015691859647631645, + 0.015396676026284695, + 0.010956118814647198, + 0.0040000737644732, + 0.0026383420918136835, + 0.0034181082155555487, + 0.0034161109942942858, + 0.005128851626068354, + 0.00382631435059011, + 0.004975059535354376, + 0.0031996804755181074, + 0.0034201378002762794, + 0.0030541089363396168, + 0.002213439904153347, + 0.0028645284473896027, + 0.004554801620543003, + 0.0032053946051746607, + 0.004097724333405495, + 0.002361792605370283, + 0.005975149106234312, + 0.006715191528201103, + 0.007693775463849306, + 0.006041414104402065, + 0.008517760783433914, + 0.00839332863688469, + 0.006955754477530718, + 0.005684239324182272, + 0.005479899235069752, + 0.007737908512353897, + 0.007976195774972439, + 0.005378217436373234, + 0.013055047020316124, + 0.004804555792361498, + 0.009512333199381828, + 0.00791234988719225, + 0.008718636818230152, + 0.009883320890367031, + 0.013290048576891422, + 0.007210577838122845, + 0.011888445354998112, + 0.007855097763240337, + 0.00742944935336709, + 0.013779538683593273, + 0.006510475650429726, + 0.007269468158483505, + 0.020525062456727028, + 0.006842615082859993, + 0.0059019592590630054, + 0.007853180170059204, + 0.009929094463586807, + 0.022369353100657463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.0459294356405735 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764338206499815 + ], + "opposite_logprob_tokens": [ + -4.882643222808838 + ], + "logprob_reward": -0.00764338206499815, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764338206499815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08287971466779709, + 0.15534326434135437, + 0.06064426153898239, + 0.06781474500894547, + 0.040689390152692795, + 0.05290781706571579, + 0.05190705507993698, + 0.04299929365515709, + 0.17415928840637207, + 0.27065518498420715 + ], + "credit_map": [ + 0.007350302301347256, + 0.007363830227404833, + 0.02106134034693241, + 0.007188855670392513, + 0.005822631064802408, + 0.02077433094382286, + 0.00725890276953578, + 0.006059509236365557, + 0.012441140599548817, + 0.013662365265190601, + 0.016690321266651154, + 0.02189488150179386, + 0.013644857332110405, + 0.02395668439567089, + 0.027867423370480537, + 0.025185585021972656, + 0.00560259073972702, + 0.004855786915868521, + 0.005699840374290943, + 0.008414710871875286, + 0.005856019910424948, + 0.008891931734979153, + 0.015129238367080688, + 0.006194134242832661, + 0.007367669139057398, + 0.00784983579069376, + 0.008704440668225288, + 0.007505356799811125, + 0.012690488249063492, + 0.009341571480035782, + 0.008220450952649117, + 0.0061349207535386086, + 0.005431758239865303, + 0.003365758340805769, + 0.005321037955582142, + 0.005096906796097755, + 0.006866724696010351, + 0.0049944776110351086, + 0.005678681191056967, + 0.00393404345959425, + 0.006854078732430935, + 0.004978755488991737, + 0.005937591195106506, + 0.006574769504368305, + 0.006615377031266689, + 0.007058657240122557, + 0.008427531458437443, + 0.006461051758378744, + 0.006716378498822451, + 0.006552041973918676, + 0.007231217343360186, + 0.004734501708298922, + 0.008893623948097229, + 0.006536208558827639, + 0.006170202512294054, + 0.005072873085737228, + 0.005817006342113018, + 0.004142026416957378, + 0.005934214685112238, + 0.0047028567641973495, + 0.005513584241271019, + 0.006979374215006828, + 0.004275606945157051, + 0.005634621251374483, + 0.016744723543524742, + 0.019707350060343742, + 0.021411294117569923, + 0.041254907846450806, + 0.015072396025061607, + 0.010782850906252861, + 0.024077754467725754, + 0.025108011439442635, + 0.015043259598314762, + 0.015285271219909191, + 0.08419530838727951, + 0.01665480062365532, + 0.015064463950693607, + 0.015388085506856441, + 0.022313321009278297, + 0.08671065419912338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.00764338206499815 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018194444477558136 + ], + "opposite_logprob_tokens": [ + -4.018194675445557 + ], + "logprob_reward": -0.018194444477558136, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018194444477558136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19815793633460999, + 0.05986698716878891, + 0.041286975145339966, + 0.0851951614022255, + 0.1151476800441742, + 0.022845124825835228, + 0.05626766011118889, + 0.03191160783171654, + 0.13523158431053162, + 0.25408926606178284 + ], + "credit_map": [ + 0.013871146366000175, + 0.013093597255647182, + 0.055495504289865494, + 0.009993541985750198, + 0.015445723198354244, + 0.06375904381275177, + 0.01094997301697731, + 0.015549431554973125, + 0.004051122814416885, + 0.005912231747061014, + 0.006260495167225599, + 0.008349181152880192, + 0.006094350479543209, + 0.009435509331524372, + 0.010056568309664726, + 0.009707536548376083, + 0.0042243157513439655, + 0.0031354730017483234, + 0.005817159079015255, + 0.0028506452217698097, + 0.004006532020866871, + 0.007347117178142071, + 0.005242020823061466, + 0.008663715794682503, + 0.011021276004612446, + 0.010894609615206718, + 0.013548743911087513, + 0.009592538699507713, + 0.013485996052622795, + 0.009347253479063511, + 0.009678340516984463, + 0.007626411970704794, + 0.00930892676115036, + 0.014465651474893093, + 0.007723370101302862, + 0.013224168680608273, + 0.02573532611131668, + 0.016094321385025978, + 0.01869194209575653, + 0.009903976693749428, + 0.0028866599313914776, + 0.0025329566560685635, + 0.002171571832150221, + 0.0022363855969160795, + 0.003929116763174534, + 0.002978229895234108, + 0.004190653562545776, + 0.0019195514032617211, + 0.005842579994350672, + 0.006125596351921558, + 0.01177955325692892, + 0.006365097127854824, + 0.007688697427511215, + 0.005979812238365412, + 0.0055939871817827225, + 0.006892339326441288, + 0.002497059991583228, + 0.002314857440069318, + 0.0037305401638150215, + 0.003104353090748191, + 0.007949402555823326, + 0.0034911988768726587, + 0.004497813992202282, + 0.004326385911554098, + 0.016150828450918198, + 0.017959188669919968, + 0.019739840179681778, + 0.021255960687994957, + 0.01716904528439045, + 0.012076355516910553, + 0.01143339928239584, + 0.019446978345513344, + 0.013355421833693981, + 0.012266110628843307, + 0.08606090396642685, + 0.01171199418604374, + 0.00862670037895441, + 0.014463108964264393, + 0.01864629052579403, + 0.08895877003669739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.018194444477558136 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980063669383526 + ], + "opposite_logprob_tokens": [ + -3.5298006534576416 + ], + "logprob_reward": -0.02980063669383526, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980063669383526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18926429748535156, + 0.035486675798892975, + 0.05308832973241806, + 0.07374942302703857, + 0.16723819077014923, + 0.024046871811151505, + 0.0998716652393341, + 0.02162092551589012, + 0.23331616818904877, + 0.10231740027666092 + ], + "credit_map": [ + 0.016486182808876038, + 0.014050411991775036, + 0.054208848625421524, + 0.010140867903828621, + 0.013227257877588272, + 0.05934948846697807, + 0.008375733159482479, + 0.013425514101982117, + 0.003139440668746829, + 0.004060748033225536, + 0.004612473770976067, + 0.003741188906133175, + 0.003965511452406645, + 0.0076470086351037025, + 0.0042970809154212475, + 0.004023225046694279, + 0.004656159784644842, + 0.0035097934305667877, + 0.007119269575923681, + 0.0057175434194505215, + 0.0048939138650894165, + 0.0075822328217327595, + 0.009809139184653759, + 0.009800281375646591, + 0.009997533634305, + 0.009728542529046535, + 0.011670177802443504, + 0.005717466585338116, + 0.011362554505467415, + 0.009974738582968712, + 0.0068136462941765785, + 0.008484764955937862, + 0.02525145374238491, + 0.017289526760578156, + 0.025578539818525314, + 0.015458685345947742, + 0.025104019790887833, + 0.018832966685295105, + 0.018958184868097305, + 0.020764818415045738, + 0.002526220865547657, + 0.0025675231590867043, + 0.003313262015581131, + 0.003040604991838336, + 0.0038691589143127203, + 0.0024749243166297674, + 0.003537530545145273, + 0.0027176483999937773, + 0.011731230653822422, + 0.007174362428486347, + 0.023277590051293373, + 0.008985547348856926, + 0.010621699504554272, + 0.010116656310856342, + 0.009037557989358902, + 0.01892702840268612, + 0.0019627087749540806, + 0.0017684772610664368, + 0.0030367658473551273, + 0.003459595376625657, + 0.0023080292157828808, + 0.0025477518793195486, + 0.0031415801495313644, + 0.0033960193395614624, + 0.017860716208815575, + 0.0228080116212368, + 0.02907092496752739, + 0.05002165585756302, + 0.02645915187895298, + 0.01686919294297695, + 0.04119079187512398, + 0.029035739600658417, + 0.006969267036765814, + 0.006706579122692347, + 0.034251224249601364, + 0.006992181297391653, + 0.005664059892296791, + 0.005464115180075169, + 0.008172450587153435, + 0.028097528964281082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.02980063669383526 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043047644197940826 + ], + "opposite_logprob_tokens": [ + -3.1680476665496826 + ], + "logprob_reward": -0.043047644197940826, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043047644197940826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08971147239208221, + 0.12666338682174683, + 0.015534758567810059, + 0.019825216382741928, + 0.06148244068026543, + 0.02169119380414486, + 0.03136257454752922, + 0.029433531686663628, + 0.06402285397052765, + 0.5402726531028748 + ], + "credit_map": [ + 0.009424989111721516, + 0.012607023119926453, + 0.010660151951014996, + 0.0076241688802838326, + 0.013254018500447273, + 0.015463359653949738, + 0.009313588961958885, + 0.011364171281456947, + 0.009497743099927902, + 0.011031362228095531, + 0.013578006066381931, + 0.017625123262405396, + 0.013657520525157452, + 0.01947729103267193, + 0.02223573625087738, + 0.01956060342490673, + 0.002137519419193268, + 0.0013928588014096022, + 0.0013768757926300168, + 0.0010544941760599613, + 0.0022379604633897543, + 0.0023350336123257875, + 0.0017037134384736419, + 0.0032963028643280268, + 0.0028716190718114376, + 0.0020708581432700157, + 0.002765370300039649, + 0.0018424656009301543, + 0.0028705112636089325, + 0.0023639285936951637, + 0.002609890652820468, + 0.0024305719416588545, + 0.006642573978751898, + 0.007611078675836325, + 0.006148225162178278, + 0.006052505224943161, + 0.011431615799665451, + 0.007892720401287079, + 0.010503395460546017, + 0.005200323183089495, + 0.002723870100453496, + 0.003053958062082529, + 0.0018769580638036132, + 0.002219071611762047, + 0.003606390440836549, + 0.003007178194820881, + 0.0030815477948635817, + 0.00212221872061491, + 0.004342148546129465, + 0.003939678426831961, + 0.004536367021501064, + 0.003146841423586011, + 0.004086833912879229, + 0.003569338470697403, + 0.0031955542508512735, + 0.00454581156373024, + 0.0032101478427648544, + 0.0023762520868331194, + 0.0037565447855740786, + 0.00403477530926466, + 0.005208945833146572, + 0.0025543239898979664, + 0.004418262280523777, + 0.003874279325827956, + 0.006875769700855017, + 0.006839571055024862, + 0.012588954530656338, + 0.005371845792979002, + 0.0065940069034695625, + 0.004673938266932964, + 0.0069163464941084385, + 0.014162418432533741, + 0.032351505011320114, + 0.029284024611115456, + 0.168925940990448, + 0.02872801572084427, + 0.018693091347813606, + 0.03615749999880791, + 0.03957507014274597, + 0.18655745685100555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.043047644197940826 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful shot begins with a close-up overhead shot of a distorted face. The Six-Fingered Qin Demon raises his head, roaring silently into the swirling blood-red clouds. His long, untidy black hair whips wildly in the unearthly wind, every muscle on his face tensed with utter fury, veins bulging from his forehead. The camera then smoothly pans upward and back, revealing his entire form—clothed in a tattered dark crimson robe, he stands alone atop the ruins of a vast ancient building. His gaze is frantic, as his right hand, rumored to have an extra finger, frantically strums the strings of an ancient black guqin before him at a speed barely tangible. Each plucking unleashes visible, blade-like, white sonic waves. These ripples of energy tear through the surrounding air, instantly severing the flying dust and debris, a perfect illustration of the aesthetics of destructive violence. As the camera continues to rise, a despairing panorama unfolds before his eyes: the entire mountaintop beneath his feet is collapsing, huge rocks and broken ground are constantly falling into the bottomless dark abyss below, and the whole world is heading towards destruction in his furious piano music, creating a grand and tragic atmosphere of the coming of doomsday.", + "seed": 22235, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_0/index_6375_seed_22235_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.039811845682561395, + "vlm_logprob": -0.039811845682561395, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.039811845682561395, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04971282556653023 + ], + "opposite_logprob_tokens": [ + -3.049712896347046 + ], + "logprob_reward": -0.04971282556653023, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04971282556653023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09469185769557953, + 0.05158955603837967, + 0.05896106734871864, + 0.030717212706804276, + 0.7082241177558899, + 0.008186518214643002, + 0.008539602160453796, + 0.017366761341691017, + 0.013094834052026272, + 0.008628416806459427 + ], + "credit_map": [ + 0.011114021763205528, + 0.003280114382505417, + 0.00638118851929903, + 0.0037248204462230206, + 0.004170161206275225, + 0.034724172204732895, + 0.004786238074302673, + 0.026511143893003464, + 0.006401499267667532, + 0.006074844393879175, + 0.004172436892986298, + 0.008105705492198467, + 0.004091228358447552, + 0.009130718186497688, + 0.006507923826575279, + 0.0071052005514502525, + 0.004513048101216555, + 0.008764470927417278, + 0.00699447188526392, + 0.008291423320770264, + 0.006696834694594145, + 0.006485058926045895, + 0.011028827168047428, + 0.006186926271766424, + 0.002786427503451705, + 0.004564677365124226, + 0.004283139482140541, + 0.0034609735012054443, + 0.0038689624052494764, + 0.005163500551134348, + 0.003560106735676527, + 0.003029425861313939, + 0.11969846487045288, + 0.04135623201727867, + 0.12377218902111053, + 0.02219441346824169, + 0.04277815297245979, + 0.30955421924591064, + 0.01645882986485958, + 0.032411616295576096, + 0.0008414153126068413, + 0.0006129380781203508, + 0.001167587237432599, + 0.0009109263774007559, + 0.0013446413213387132, + 0.001140950946137309, + 0.001388757722452283, + 0.0007793015684001148, + 0.0007198783569037914, + 0.0008480606484226882, + 0.0011043434496968985, + 0.0012496171984821558, + 0.001316442503593862, + 0.0008569200872443616, + 0.0014473908813670278, + 0.0009969496168196201, + 0.0015301264356821775, + 0.0016698921099305153, + 0.0022581235971301794, + 0.0020126819144934416, + 0.00222559692338109, + 0.0018458524718880653, + 0.0038732956163585186, + 0.0019511922728270292, + 0.0010351183591410518, + 0.0013022503117099404, + 0.0015081524616107345, + 0.001874291687272489, + 0.0030808935407549143, + 0.0012448937632143497, + 0.0017273585544899106, + 0.0013218759559094906, + 0.0012941183522343636, + 0.001296631875447929, + 0.0010914822341874242, + 0.0007086816476657987, + 0.0013945576502010226, + 0.0012699104845523834, + 0.0008228569058701396, + 0.0007501784129999578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.04971282556653023 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006747082807123661 + ], + "opposite_logprob_tokens": [ + -5.006747245788574 + ], + "logprob_reward": -0.006747082807123661, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747082807123661, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06432594358921051, + 0.011919355019927025, + 0.021147621795535088, + 0.02585741877555847, + 0.8174381852149963, + 0.01149123627692461, + 0.023072820156812668, + 0.012831995263695717, + 0.006367378402501345, + 0.00554802268743515 + ], + "credit_map": [ + 0.009132049046456814, + 0.0019487477838993073, + 0.0044971853494644165, + 0.001992691308259964, + 0.002009740797802806, + 0.020055025815963745, + 0.0025760680437088013, + 0.022114437073469162, + 0.0015008453046903014, + 0.0010862681083381176, + 0.0019743831362575293, + 0.001229903195053339, + 0.001210082438774407, + 0.0014428958529606462, + 0.0017570131458342075, + 0.0017179641872644424, + 0.002277206163853407, + 0.003792558563873172, + 0.00231454661116004, + 0.0026261864695698023, + 0.0024175073485821486, + 0.0025085636880248785, + 0.003224722808226943, + 0.0019863301422446966, + 0.0038767133373767138, + 0.002305080648511648, + 0.00370761938393116, + 0.0020389824640005827, + 0.0022360750008374453, + 0.008096886798739433, + 0.001815845724195242, + 0.0017802161164581776, + 0.11812597513198853, + 0.04534918814897537, + 0.16383469104766846, + 0.02632037177681923, + 0.04941977187991142, + 0.3574840724468231, + 0.019030645489692688, + 0.03787347674369812, + 0.001125968061387539, + 0.0009954579873010516, + 0.001225945889018476, + 0.0018415013328194618, + 0.0018178002210333943, + 0.0013797357678413391, + 0.002076502423733473, + 0.0010283248266205192, + 0.0016751332441344857, + 0.0022150969598442316, + 0.0031175725162029266, + 0.0023582889698445797, + 0.005565883591771126, + 0.002405214589089155, + 0.0026410019490867853, + 0.0030946286860853434, + 0.0011657460127025843, + 0.0012013523373752832, + 0.0015599812613800168, + 0.0015348483575507998, + 0.0016802770551294088, + 0.001597768161445856, + 0.0023096767254173756, + 0.0017823450034484267, + 0.0010289686033502221, + 0.0009359816904179752, + 0.0006230262224562466, + 0.0008490273030474782, + 0.000931891961954534, + 0.0007791886455379426, + 0.0005910387262701988, + 0.0006282555987127125, + 0.0006658807396888733, + 0.0007699016132391989, + 0.0009247335256077349, + 0.0005731657147407532, + 0.00041986670112237334, + 0.000777380948420614, + 0.0005128498305566609, + 0.0009042436722666025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.006747082807123661 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058417536318302 + ], + "opposite_logprob_tokens": [ + -3.8955841064453125 + ], + "logprob_reward": -0.02058417536318302, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058417536318302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.088438019156456, + 0.014456280507147312, + 0.026567265391349792, + 0.03799324110150337, + 0.7589591145515442, + 0.009994111955165863, + 0.006694905925542116, + 0.02780921570956707, + 0.015631984919309616, + 0.01345592550933361 + ], + "credit_map": [ + 0.016325198113918304, + 0.0027372052427381277, + 0.006754112429916859, + 0.0029324733186513186, + 0.0029141942504793406, + 0.02343016304075718, + 0.0033269282430410385, + 0.03001774474978447, + 0.0016328879864886403, + 0.0019539198838174343, + 0.002147551393136382, + 0.0011170095531269908, + 0.0015092819230630994, + 0.0022070766426622868, + 0.0017235362902283669, + 0.0021650169510394335, + 0.0024221038911491632, + 0.00277147488668561, + 0.0031470905523747206, + 0.003519632387906313, + 0.0031655733473598957, + 0.002915049437433481, + 0.004719947464764118, + 0.003906394354999065, + 0.004051953088492155, + 0.005427008029073477, + 0.003712383331730962, + 0.005105085205286741, + 0.004163558129221201, + 0.00849453266710043, + 0.004259863868355751, + 0.0027788577135652304, + 0.12956887483596802, + 0.03348634019494057, + 0.13641144335269928, + 0.022708170115947723, + 0.04619560018181801, + 0.33876052498817444, + 0.017350278794765472, + 0.03447785973548889, + 0.001145996619015932, + 0.0007732450612820685, + 0.0008890180033631623, + 0.001280851080082357, + 0.0019841494504362345, + 0.0016068285331130028, + 0.0014596572145819664, + 0.0008543665171600878, + 0.0006545110954903066, + 0.0007486081449314952, + 0.0008128822082653642, + 0.0007121952367015183, + 0.0012577420566231012, + 0.0007479189080186188, + 0.0011732849525287747, + 0.0005877633229829371, + 0.00215738033875823, + 0.0025857877917587757, + 0.0036469860933721066, + 0.0030631942208856344, + 0.004025254864245653, + 0.002741458360105753, + 0.0061135259456932545, + 0.0034756276290863752, + 0.0022238786332309246, + 0.00221628206782043, + 0.0014412876917049289, + 0.0018156013684347272, + 0.00298335961997509, + 0.0018368343589827418, + 0.0017640955047681928, + 0.0013506457908079028, + 0.001640744274482131, + 0.0017898622900247574, + 0.0016033730935305357, + 0.0014630494406446815, + 0.002518954686820507, + 0.0018263570964336395, + 0.0013774833641946316, + 0.0012361012632027268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.02058417536318302 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897115498781204 + ], + "opposite_logprob_tokens": [ + -2.5789711475372314 + ], + "logprob_reward": -0.07897115498781204, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897115498781204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2897493839263916, + 0.12899750471115112, + 0.07178297638893127, + 0.07496225833892822, + 0.2673877477645874, + 0.02944740280508995, + 0.046785298734903336, + 0.039338018745183945, + 0.021277522668242455, + 0.030271949246525764 + ], + "credit_map": [ + 0.029373012483119965, + 0.008803080767393112, + 0.020691273733973503, + 0.008295939303934574, + 0.008950325660407543, + 0.11030002683401108, + 0.010960339568555355, + 0.09237537533044815, + 0.016707882285118103, + 0.012109411880373955, + 0.024620074778795242, + 0.011915680021047592, + 0.006894715595990419, + 0.01885892078280449, + 0.010838703252375126, + 0.027052108198404312, + 0.0077471621334552765, + 0.007252632640302181, + 0.014974544756114483, + 0.008283933624625206, + 0.005908985622227192, + 0.008020760491490364, + 0.011818761005997658, + 0.007776194252073765, + 0.008638602681457996, + 0.009251399897038937, + 0.010265279561281204, + 0.009168559685349464, + 0.007641679607331753, + 0.019113898277282715, + 0.004724751692265272, + 0.0061580827459692955, + 0.030653811991214752, + 0.035242706537246704, + 0.047333307564258575, + 0.013739182613790035, + 0.019946027547121048, + 0.09530222415924072, + 0.008923348039388657, + 0.01624711975455284, + 0.0028446647338569164, + 0.0026037918869405985, + 0.002920289058238268, + 0.0040595438331365585, + 0.004929845687001944, + 0.003590994980186224, + 0.005042817909270525, + 0.003455453785136342, + 0.004334561061114073, + 0.004794999025762081, + 0.006433522328734398, + 0.005431458819657564, + 0.010155187919735909, + 0.0041909124702215195, + 0.005296825896948576, + 0.006147829350084066, + 0.0048714904114604, + 0.0048341164365410805, + 0.006545057985931635, + 0.0036382281687110662, + 0.0035490607842803, + 0.004962950479239225, + 0.003986922092735767, + 0.006950189359486103, + 0.0022511309944093227, + 0.0018615844892337918, + 0.0021163038909435272, + 0.004082683473825455, + 0.0031432965770363808, + 0.0023944727145135403, + 0.0029823631048202515, + 0.0024456861428916454, + 0.0046539814211428165, + 0.00630210991948843, + 0.0036406368017196655, + 0.002338955644518137, + 0.0031543662771582603, + 0.004316041246056557, + 0.0026555482763797045, + 0.0032103119883686304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.07897115498781204 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the character depicted as a Six-Fingered Qin Demon with long, untidy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a tattered dark crimson robe and standing atop the ruins of an ancient building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character playing an ancient black guqin with his right hand, which has an extra finger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible, blade-like white sonic waves generated by the guqin that sever flying dust and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up overhead shot and then smoothly pan upward and back to reveal the collapsing mountaintop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043043989688158035 + ], + "opposite_logprob_tokens": [ + -3.168044090270996 + ], + "logprob_reward": -0.043043989688158035, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043043989688158035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07121098041534424, + 0.018846992403268814, + 0.022731292992830276, + 0.0451648123562336, + 0.7162373065948486, + 0.026517337188124657, + 0.05928092077374458, + 0.02536645717918873, + 0.005125111434608698, + 0.009518777951598167 + ], + "credit_map": [ + 0.015023624524474144, + 0.00537497503682971, + 0.00795876607298851, + 0.00481223501265049, + 0.004283807706087828, + 0.021850505843758583, + 0.0033031196799129248, + 0.008603942580521107, + 0.002266326919198036, + 0.0021678837947547436, + 0.002233169972896576, + 0.0016394690610468388, + 0.002027250360697508, + 0.0027361009269952774, + 0.003075476037338376, + 0.0027013143990188837, + 0.0028433366678655148, + 0.0017205496551468968, + 0.003924706485122442, + 0.004397352691739798, + 0.0019294139929115772, + 0.002298123436048627, + 0.00237649935297668, + 0.003241310827434063, + 0.0060683488845825195, + 0.00641697458922863, + 0.0052503603510558605, + 0.005822190083563328, + 0.0057144369930028915, + 0.006953836418688297, + 0.004510870203375816, + 0.004427794367074966, + 0.09423528611660004, + 0.0469750314950943, + 0.14902861416339874, + 0.026009296998381615, + 0.04282577708363533, + 0.3052404522895813, + 0.017288099974393845, + 0.0346347838640213, + 0.003050713799893856, + 0.0025970914866775274, + 0.003193061565980315, + 0.003984780050814152, + 0.003687006887048483, + 0.0033791454043239355, + 0.004310428164899349, + 0.002315111458301544, + 0.004143337719142437, + 0.005609179846942425, + 0.007847059518098831, + 0.006344757974147797, + 0.014125621877610683, + 0.006227351259440184, + 0.0070677464827895164, + 0.007915864698588848, + 0.0022863312624394894, + 0.0026976261287927628, + 0.003026962513104081, + 0.003249909495934844, + 0.003418766660615802, + 0.002697634743526578, + 0.00430193729698658, + 0.0036872881464660168, + 0.0005921359406784177, + 0.0005998609703965485, + 0.000578535080421716, + 0.0005511163617484272, + 0.0010897924657911062, + 0.0004351249081082642, + 0.0005948840989731252, + 0.000683661550283432, + 0.0010787536157295108, + 0.0011537580285221338, + 0.0009813180658966303, + 0.0011963131837546825, + 0.0018647446995601058, + 0.001278841751627624, + 0.001036116387695074, + 0.0009289325680583715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.043043989688158035 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_1/index_tensor([996, 996])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_1/index_tensor([996, 996])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0953b999e8e9ad39c37c5922d5345eda46e304cd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_1/index_tensor([996, 996])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, intimate close-up, angled slightly downward, captures a pair of slender hands methodically arranging a section of a towering, dark mahogany bookshelf. The skin of the hands is pale and smooth, the movements cautious and gentle. One hand carefully pushes a thick, crimson, leather-bound volume, its spine showing the slight cracks of age and embossed with unremarkable traces of faded gold leaf filigree, until its edge is perfectly flush with the adjacent one. Warm, soft light, as if from a nearby blackout lamp, casts a focused golden glow, highlighting the rich texture of the leather spines—from deep forest green to worn brown—and creating a soft sheen on the manicured fingernails. As the first hand holds the book in place, the fingers of the other delicately trace along the top edge of the aligned volume, ensuring a perfectly flat surface. The background is a soft-focus blur of more books and dark wood, enhancing the shallow depth of field and focusing the viewer's attention on the meticulous action, evoking an atmosphere of tranquility, concentration, and scholarship, imbued with the scent of old paper and leather.", + "seed": 23523, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_1/index_996_seed_23523_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021496809320524335, + "vlm_logprob": -0.021496809320524335, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.021496809320524335, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01851174421608448 + ], + "opposite_logprob_tokens": [ + -5.143511772155762 + ], + "logprob_reward": -0.01851174421608448, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01851174421608448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21173301339149475, + 0.09062517434358597, + 0.06959887593984604, + 0.12498287111520767, + 0.06393767893314362, + 0.05117183178663254, + 0.06831110268831253, + 0.11715903133153915, + 0.10208044946193695, + 0.10039989650249481 + ], + "credit_map": [ + 0.021767210215330124, + 0.02699020877480507, + 0.019908709451556206, + 0.0240232665091753, + 0.024849677458405495, + 0.03219873085618019, + 0.025654224678874016, + 0.036341000348329544, + 0.009119844995439053, + 0.010782442055642605, + 0.010112463496625423, + 0.007534383330494165, + 0.01109867263585329, + 0.021867379546165466, + 0.008125098422169685, + 0.01198489312082529, + 0.00892787054181099, + 0.008175677619874477, + 0.010735948570072651, + 0.006000836845487356, + 0.0107651362195611, + 0.006789218168705702, + 0.008873979561030865, + 0.009330214001238346, + 0.01743294671177864, + 0.008337441831827164, + 0.018112806603312492, + 0.007762422319501638, + 0.013465430587530136, + 0.027913924306631088, + 0.008658007718622684, + 0.02329990081489086, + 0.011860357597470284, + 0.007019861601293087, + 0.008495229296386242, + 0.006764724850654602, + 0.0058506871573626995, + 0.008511797524988651, + 0.005048638675361872, + 0.010386388748884201, + 0.009510593488812447, + 0.004699341021478176, + 0.007074352819472551, + 0.006022186949849129, + 0.005866014864295721, + 0.005150377284735441, + 0.004422848578542471, + 0.008426123298704624, + 0.01219807006418705, + 0.009840834885835648, + 0.007701289840042591, + 0.005849522538483143, + 0.0041098264046013355, + 0.007624499034136534, + 0.009061016142368317, + 0.011926045641303062, + 0.016758019104599953, + 0.01199521217495203, + 0.02487107366323471, + 0.007236719597131014, + 0.012164250016212463, + 0.016251837834715843, + 0.009586134925484657, + 0.018295789137482643, + 0.015317130833864212, + 0.013526438735425472, + 0.017636310309171677, + 0.007717845495790243, + 0.009454200975596905, + 0.014436821453273296, + 0.012042391113936901, + 0.011949318461120129, + 0.009625774808228016, + 0.012834981083869934, + 0.0135155338793993, + 0.008958771824836731, + 0.01905572973191738, + 0.009476082399487495, + 0.014135297387838364, + 0.012797730974853039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.01851174421608448 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029854901134967804 + ], + "opposite_logprob_tokens": [ + -3.5298550128936768 + ], + "logprob_reward": -0.029854901134967804, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029854901134967804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12245667725801468, + 0.10769584029912949, + 0.07154487073421478, + 0.11125285923480988, + 0.1289520412683487, + 0.08423688262701035, + 0.02728106826543808, + 0.1708589643239975, + 0.08053934574127197, + 0.09518147259950638 + ], + "credit_map": [ + 0.009512502700090408, + 0.01684742420911789, + 0.011407731100916862, + 0.01214657537639141, + 0.015447835437953472, + 0.01855292357504368, + 0.014678016304969788, + 0.023863673210144043, + 0.01280926913022995, + 0.012344731949269772, + 0.019512994214892387, + 0.011442977003753185, + 0.011888711713254452, + 0.020822083577513695, + 0.009951519779860973, + 0.008923547342419624, + 0.009844773449003696, + 0.00803639180958271, + 0.010750005953013897, + 0.006504491437226534, + 0.011254696175456047, + 0.006798362359404564, + 0.008302257396280766, + 0.01005389355123043, + 0.017466112971305847, + 0.007911169901490211, + 0.018596608191728592, + 0.009251012466847897, + 0.010112632066011429, + 0.016565309837460518, + 0.008836532942950726, + 0.022513482719659805, + 0.022662272676825523, + 0.00849839486181736, + 0.021130239591002464, + 0.007971901446580887, + 0.011397846974432468, + 0.017807675525546074, + 0.008436013013124466, + 0.031047698110342026, + 0.016454827040433884, + 0.00794706866145134, + 0.010578921996057034, + 0.007779631298035383, + 0.009664565324783325, + 0.007284386549144983, + 0.010889414697885513, + 0.013638066127896309, + 0.004846879746764898, + 0.0033535282127559185, + 0.0027068976778537035, + 0.0022686482407152653, + 0.0021808447781950235, + 0.003551139496266842, + 0.0025543675292283297, + 0.005818761419504881, + 0.02320205420255661, + 0.016680685803294182, + 0.028221603482961655, + 0.010006540454924107, + 0.01213021855801344, + 0.03009120747447014, + 0.01593237742781639, + 0.03459428623318672, + 0.010728597640991211, + 0.01301033329218626, + 0.009530910290777683, + 0.006044951267540455, + 0.009037933312356472, + 0.010635226033627987, + 0.010215329006314278, + 0.011336062103509903, + 0.01124708354473114, + 0.012668807059526443, + 0.009807503782212734, + 0.009074898436665535, + 0.01789521612226963, + 0.01195121556520462, + 0.008963430300354958, + 0.013573317788541317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.029854901134967804 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048666927963495255 + ], + "opposite_logprob_tokens": [ + -3.0486669540405273 + ], + "logprob_reward": -0.048666927963495255, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048666927963495255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.134339839220047, + 0.09376540780067444, + 0.12175076454877853, + 0.082758828997612, + 0.1058313250541687, + 0.0850900262594223, + 0.11320840567350388, + 0.0853181928396225, + 0.10114655643701553, + 0.07679063826799393 + ], + "credit_map": [ + 0.013368587009608746, + 0.017106784507632256, + 0.015477403067052364, + 0.020559361204504967, + 0.014745323918759823, + 0.017993832007050514, + 0.013490176759660244, + 0.021598367020487785, + 0.011484618298709393, + 0.009992331266403198, + 0.012013478204607964, + 0.008647935464978218, + 0.012728847563266754, + 0.014319722540676594, + 0.011882225051522255, + 0.012696243822574615, + 0.015941573306918144, + 0.01820654235780239, + 0.018170662224292755, + 0.011127844452857971, + 0.016148535534739494, + 0.01084689237177372, + 0.011973854154348373, + 0.019334858283400536, + 0.015758471563458443, + 0.006184353493154049, + 0.011384360492229462, + 0.007532728370279074, + 0.008088445290923119, + 0.01125797163695097, + 0.007578824646770954, + 0.014973673038184643, + 0.02149861678481102, + 0.006915980484336615, + 0.020989207550883293, + 0.0068458872847259045, + 0.00705658970400691, + 0.013789891265332699, + 0.005786232650279999, + 0.022948915138840675, + 0.02318553440272808, + 0.010090549476444721, + 0.010858642868697643, + 0.010070296004414558, + 0.008347438648343086, + 0.006129459012299776, + 0.006969539914280176, + 0.009438573382794857, + 0.018237827345728874, + 0.00973348505795002, + 0.011969812214374542, + 0.007468739990144968, + 0.007689884398132563, + 0.012281729839742184, + 0.016454128548502922, + 0.0293728057295084, + 0.011981736868619919, + 0.0076865339651703835, + 0.02007444016635418, + 0.006405654828995466, + 0.010975100100040436, + 0.008926036767661572, + 0.007989483885467052, + 0.011279209516942501, + 0.015483738854527473, + 0.021328773349523544, + 0.015059204772114754, + 0.007444487418979406, + 0.010453146882355213, + 0.01276983693242073, + 0.008247768506407738, + 0.01035960204899311, + 0.0075212265364825726, + 0.009432688355445862, + 0.00949865486472845, + 0.006193093955516815, + 0.015326418913900852, + 0.008941007778048515, + 0.008366462774574757, + 0.011511089280247688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.048666927963495255 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002823535120114684 + ], + "opposite_logprob_tokens": [ + -5.877823352813721 + ], + "logprob_reward": -0.002823535120114684, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002823535120114684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21158704161643982, + 0.19672022759914398, + 0.07123304903507233, + 0.07428398728370667, + 0.0898590162396431, + 0.08602160215377808, + 0.06572867184877396, + 0.10877028852701187, + 0.057103365659713745, + 0.03869274631142616 + ], + "credit_map": [ + 0.022665027529001236, + 0.026625556871294975, + 0.01733814924955368, + 0.02721487171947956, + 0.03627682849764824, + 0.02738223783671856, + 0.025314774364233017, + 0.028769606724381447, + 0.02089853771030903, + 0.02016315422952175, + 0.03199175000190735, + 0.020248547196388245, + 0.02587163634598255, + 0.03733358532190323, + 0.02273484691977501, + 0.01747817173600197, + 0.010767954401671886, + 0.006731345783919096, + 0.013493926264345646, + 0.005954519379884005, + 0.008799047209322453, + 0.007225740235298872, + 0.008380964398384094, + 0.00987955555319786, + 0.011644343845546246, + 0.007002117112278938, + 0.00914580374956131, + 0.007228102069348097, + 0.00887627899646759, + 0.007646279875189066, + 0.006838460918515921, + 0.015902603045105934, + 0.016794400289654732, + 0.0052590095438063145, + 0.01569330133497715, + 0.0064683216623961926, + 0.006755105219781399, + 0.011704511009156704, + 0.0072554051876068115, + 0.01992896758019924, + 0.020671676844358444, + 0.007814434356987476, + 0.010404287837445736, + 0.00835167896002531, + 0.009290294721722603, + 0.006608335766941309, + 0.009538277052342892, + 0.013342620804905891, + 0.012631765566766262, + 0.005255694501101971, + 0.008230301551520824, + 0.004414257127791643, + 0.004542844835668802, + 0.011133803054690361, + 0.007599452044814825, + 0.011920548975467682, + 0.017038218677043915, + 0.019024090841412544, + 0.01732117496430874, + 0.0061132293194532394, + 0.009170778095722198, + 0.01416934747248888, + 0.008692607283592224, + 0.01724085584282875, + 0.0059889391995966434, + 0.00889852736145258, + 0.008663906715810299, + 0.00765374256297946, + 0.00853925384581089, + 0.004994574468582869, + 0.00635205302387476, + 0.006012372672557831, + 0.0038664487656205893, + 0.005784486886113882, + 0.004419050645083189, + 0.0026815261226147413, + 0.005459629464894533, + 0.005395196843892336, + 0.005099216941744089, + 0.005987195298075676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.002823535120114684 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626938167959452 + ], + "opposite_logprob_tokens": [ + -4.882627010345459 + ], + "logprob_reward": -0.007626938167959452, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626938167959452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1469467133283615, + 0.08383247256278992, + 0.048007309436798096, + 0.09860549122095108, + 0.06349898874759674, + 0.12773674726486206, + 0.21534821391105652, + 0.07365556806325912, + 0.0921170711517334, + 0.05025140568614006 + ], + "credit_map": [ + 0.018765559419989586, + 0.01799672469496727, + 0.012230885215103626, + 0.019193455576896667, + 0.021369226276874542, + 0.01803223229944706, + 0.02078385278582573, + 0.01857479102909565, + 0.010370721109211445, + 0.008151334710419178, + 0.010230720974504948, + 0.006843941286206245, + 0.011383561417460442, + 0.010528611950576305, + 0.009862266480922699, + 0.016461309045553207, + 0.006018664687871933, + 0.005595194175839424, + 0.0064974562264978886, + 0.005862798076122999, + 0.0061353836208581924, + 0.005161713808774948, + 0.005733096972107887, + 0.007003002800047398, + 0.021370090544223785, + 0.0071476357989013195, + 0.014615912921726704, + 0.008433663286268711, + 0.009979493916034698, + 0.015964074060320854, + 0.008136697113513947, + 0.012957921251654625, + 0.009688296355307102, + 0.005459526553750038, + 0.008359369821846485, + 0.007592393085360527, + 0.004747915081679821, + 0.0065133883617818356, + 0.004338349215686321, + 0.016799749806523323, + 0.018848225474357605, + 0.012799388729035854, + 0.0168096125125885, + 0.01257660984992981, + 0.0148856732994318, + 0.010317818261682987, + 0.024510540068149567, + 0.01698887348175049, + 0.040946654975414276, + 0.035683102905750275, + 0.034960657358169556, + 0.019601618871092796, + 0.015951190143823624, + 0.027811335399746895, + 0.020874034613370895, + 0.01951962523162365, + 0.012565885670483112, + 0.010690678842365742, + 0.010079546831548214, + 0.005601430777460337, + 0.008632528595626354, + 0.010074738413095474, + 0.006361800245940685, + 0.009648963809013367, + 0.013342234306037426, + 0.010078639723360538, + 0.021463846787810326, + 0.008648836985230446, + 0.010091368108987808, + 0.01118953712284565, + 0.007296978961676359, + 0.010005632415413857, + 0.005130598321557045, + 0.00776318833231926, + 0.00618754094466567, + 0.003906875848770142, + 0.006752630230039358, + 0.006090857088565826, + 0.007373121101409197, + 0.007046591490507126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.007626938167959452 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, intimate close-up, angled slightly downward, captures a pair of slender hands methodically arranging a section of a towering, dark mahogany bookshelf. The skin of the hands is pale and smooth, the movements cautious and gentle. One hand carefully pushes a thick, crimson, leather-bound volume, its spine showing the slight cracks of age and embossed with unremarkable traces of faded gold leaf filigree, until its edge is perfectly flush with the adjacent one. Warm, soft light, as if from a nearby blackout lamp, casts a focused golden glow, highlighting the rich texture of the leather spines—from deep forest green to worn brown—and creating a soft sheen on the manicured fingernails. As the first hand holds the book in place, the fingers of the other delicately trace along the top edge of the aligned volume, ensuring a perfectly flat surface. The background is a soft-focus blur of more books and dark wood, enhancing the shallow depth of field and focusing the viewer's attention on the meticulous action, evoking an atmosphere of tranquility, concentration, and scholarship, imbued with the scent of old paper and leather.", + "seed": 23524, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_1/index_996_seed_23524_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016298588272184132, + "vlm_logprob": -0.016298588272184132, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.016298588272184132, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015408990904688835 + ], + "opposite_logprob_tokens": [ + -4.765408992767334 + ], + "logprob_reward": -0.015408990904688835, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.015408990904688835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12184952199459076, + 0.16218149662017822, + 0.07447967678308487, + 0.03725812956690788, + 0.05675794184207916, + 0.06207592412829399, + 0.07086142152547836, + 0.06272962689399719, + 0.18616333603858948, + 0.1656428873538971 + ], + "credit_map": [ + 0.009280937723815441, + 0.020558036863803864, + 0.012761070393025875, + 0.016385041177272797, + 0.014680560678243637, + 0.01388243492692709, + 0.015829680487513542, + 0.01847175695002079, + 0.014170407317578793, + 0.0214614849537611, + 0.011425599455833435, + 0.010824625380337238, + 0.041151612997055054, + 0.02666350267827511, + 0.01800527423620224, + 0.018478993326425552, + 0.005909521598368883, + 0.009953097440302372, + 0.0066679539158940315, + 0.007260849699378014, + 0.009781531989574432, + 0.0069740284234285355, + 0.015255209058523178, + 0.012677484191954136, + 0.00527770584449172, + 0.0044084652327001095, + 0.0036126223858445883, + 0.005306147504597902, + 0.0047651007771492004, + 0.004848339129239321, + 0.004259463865309954, + 0.004780283663421869, + 0.005416397470980883, + 0.004975262563675642, + 0.004807499703019857, + 0.010281966999173164, + 0.006035682745277882, + 0.008992358110845089, + 0.006207181606441736, + 0.010041600093245506, + 0.007398503366857767, + 0.005814630072563887, + 0.004536030348390341, + 0.00937869492918253, + 0.006566185504198074, + 0.01050132792443037, + 0.005545520689338446, + 0.012335038743913174, + 0.009325611405074596, + 0.008235426619648933, + 0.005339703056961298, + 0.010302217677235603, + 0.012356285005807877, + 0.007740413304418325, + 0.008846606127917767, + 0.008715156465768814, + 0.005393247120082378, + 0.00912205409258604, + 0.0055744703859090805, + 0.01394138764590025, + 0.009781933389604092, + 0.0061133005656301975, + 0.006546953693032265, + 0.006256284192204475, + 0.021224452182650566, + 0.027110328897833824, + 0.012186577543616295, + 0.019024331122636795, + 0.028052248060703278, + 0.025854626670479774, + 0.024270880967378616, + 0.02843991108238697, + 0.01841871254146099, + 0.019093239679932594, + 0.012122144922614098, + 0.018359823152422905, + 0.03186529874801636, + 0.017745476216077805, + 0.019474077969789505, + 0.028564119711518288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.015408990904688835 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012598569504916668 + ], + "opposite_logprob_tokens": [ + -4.387598514556885 + ], + "logprob_reward": -0.012598569504916668, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012598569504916668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08548075705766678, + 0.16443553566932678, + 0.08630464226007462, + 0.04728882014751434, + 0.1713433712720871, + 0.054318297654390335, + 0.058105017989873886, + 0.20686055719852448, + 0.05732763931155205, + 0.06853535771369934 + ], + "credit_map": [ + 0.007556143682450056, + 0.010885204188525677, + 0.00845707394182682, + 0.012506085447967052, + 0.012661765329539776, + 0.011430389247834682, + 0.00890337210148573, + 0.013080714270472527, + 0.011376539245247841, + 0.021375516429543495, + 0.011197981424629688, + 0.015297535806894302, + 0.04182051867246628, + 0.027284035459160805, + 0.016989251598715782, + 0.019094126299023628, + 0.005347237456589937, + 0.009203714318573475, + 0.005460877902805805, + 0.009215638041496277, + 0.012735916301608086, + 0.010709655471146107, + 0.019981788471341133, + 0.013649802654981613, + 0.006174951791763306, + 0.004153289366513491, + 0.003139375476166606, + 0.007066711783409119, + 0.007324866950511932, + 0.006701141595840454, + 0.004708793479949236, + 0.008019689470529556, + 0.019961029291152954, + 0.019506489858031273, + 0.014348682016134262, + 0.02736319974064827, + 0.020058520138263702, + 0.02338472753763199, + 0.023140551522374153, + 0.0235801599919796, + 0.00558639969676733, + 0.00549955852329731, + 0.004100124817341566, + 0.005962755996733904, + 0.007752429693937302, + 0.009392044506967068, + 0.007334391586482525, + 0.00869059283286333, + 0.0069136531092226505, + 0.0073275030590593815, + 0.004351280629634857, + 0.005642478354275227, + 0.01094338484108448, + 0.006881676148623228, + 0.008313959464430809, + 0.007731080520898104, + 0.023055020719766617, + 0.03319045156240463, + 0.013300537131726742, + 0.012572147883474827, + 0.046593572944402695, + 0.030519520863890648, + 0.020628679543733597, + 0.027000628411769867, + 0.00525473989546299, + 0.008262158371508121, + 0.00408584950491786, + 0.006793827749788761, + 0.01032347697764635, + 0.007018677890300751, + 0.008704638108611107, + 0.006884270813316107, + 0.0066535682417452335, + 0.006302772555500269, + 0.003480553161352873, + 0.004741191864013672, + 0.012033491395413876, + 0.0118021909147501, + 0.008448180742561817, + 0.01507340557873249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.012598569504916668 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810877725481987 + ], + "opposite_logprob_tokens": [ + -3.2881088256835938 + ], + "logprob_reward": -0.03810877725481987, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810877725481987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12184689939022064, + 0.0967758372426033, + 0.08819882571697235, + 0.09049993008375168, + 0.04767864942550659, + 0.04173894599080086, + 0.07928179949522018, + 0.2545637786388397, + 0.08731328696012497, + 0.09210207313299179 + ], + "credit_map": [ + 0.008838345296680927, + 0.01968255639076233, + 0.013883629813790321, + 0.018102873116731644, + 0.010364505462348461, + 0.014510498382151127, + 0.013289667665958405, + 0.023174822330474854, + 0.006738012190908194, + 0.015860622748732567, + 0.009273119270801544, + 0.010761013254523277, + 0.022254236042499542, + 0.011441643349826336, + 0.010592306964099407, + 0.009854870848357677, + 0.004162883851677179, + 0.007525003049522638, + 0.004690272733569145, + 0.008001203648746014, + 0.008365602232515812, + 0.008927376940846443, + 0.026065051555633545, + 0.02046142891049385, + 0.009951344691216946, + 0.01058061234652996, + 0.007112196646630764, + 0.009579996578395367, + 0.018410326912999153, + 0.010001475922763348, + 0.013706878758966923, + 0.011157089844346046, + 0.005129566416144371, + 0.00504232756793499, + 0.0032583456486463547, + 0.007646557874977589, + 0.006821323186159134, + 0.005780813284218311, + 0.00837200228124857, + 0.0056277066469192505, + 0.0068495976738631725, + 0.0036525328177958727, + 0.0036383899860084057, + 0.004098591394722462, + 0.004415404051542282, + 0.0073128617368638515, + 0.0037798124831169844, + 0.007991751655936241, + 0.006215503439307213, + 0.011515512131154537, + 0.005691601429134607, + 0.009227737784385681, + 0.018247203901410103, + 0.009948722086846828, + 0.00863142590969801, + 0.009804091416299343, + 0.030506933107972145, + 0.03801868110895157, + 0.016693362966179848, + 0.025656472891569138, + 0.05325135588645935, + 0.03496994823217392, + 0.022113298997282982, + 0.03335370495915413, + 0.008786914870142937, + 0.009647815488278866, + 0.006419166456907988, + 0.009278020821511745, + 0.012156151235103607, + 0.014342213980853558, + 0.010704278014600277, + 0.015978721901774406, + 0.00952896848320961, + 0.012513069435954094, + 0.006831148639321327, + 0.008685135282576084, + 0.019758358597755432, + 0.009779412299394608, + 0.01158197596669197, + 0.013423996046185493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.03810877725481987 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741754245012999 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741754245012999, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741754245012999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15342439711093903, + 0.19526122510433197, + 0.06987696141004562, + 0.08224594593048096, + 0.13089725375175476, + 0.05944674089550972, + 0.08379890024662018, + 0.08234421163797379, + 0.03533973544836044, + 0.10736463963985443 + ], + "credit_map": [ + 0.012828911654651165, + 0.023690493777394295, + 0.018475335091352463, + 0.024251479655504227, + 0.0204409658908844, + 0.021418139338493347, + 0.014607835561037064, + 0.01771124079823494, + 0.017895037308335304, + 0.025949861854314804, + 0.015488795004785061, + 0.013979224488139153, + 0.047360751777887344, + 0.033043988049030304, + 0.02035638689994812, + 0.021187176927924156, + 0.0077476706355810165, + 0.008653406985104084, + 0.006566635798662901, + 0.007269436959177256, + 0.009113800711929798, + 0.0069388654083013535, + 0.01406931970268488, + 0.009517824277281761, + 0.010210967622697353, + 0.01132535096257925, + 0.007696740794926882, + 0.009735946543514729, + 0.012850369326770306, + 0.010051378048956394, + 0.009774914011359215, + 0.010600273497402668, + 0.02962755225598812, + 0.01143169030547142, + 0.010580705478787422, + 0.013346015475690365, + 0.012908551841974258, + 0.02154092863202095, + 0.01225173007696867, + 0.019210079684853554, + 0.008042678236961365, + 0.005769836716353893, + 0.005244027357548475, + 0.006363490130752325, + 0.00964555237442255, + 0.010329638607800007, + 0.005427184514701366, + 0.008624332956969738, + 0.00931585393846035, + 0.01117730513215065, + 0.007521132007241249, + 0.00862588919699192, + 0.016468433663249016, + 0.01189365517348051, + 0.008524429053068161, + 0.01027219370007515, + 0.008927933871746063, + 0.011969512328505516, + 0.007742650341242552, + 0.013189483433961868, + 0.012017094530165195, + 0.007917075417935848, + 0.010630170814692974, + 0.009950298815965652, + 0.0035316459834575653, + 0.004578165709972382, + 0.0029715360142290592, + 0.004521925933659077, + 0.0047518000937998295, + 0.005660309921950102, + 0.004882956854999065, + 0.004441395867615938, + 0.013943972997367382, + 0.012376406230032444, + 0.008011546917259693, + 0.011028613895177841, + 0.017838554456830025, + 0.014613544568419456, + 0.01270844042301178, + 0.016843562945723534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.006741754245012999 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634849451482296 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008634849451482296, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634849451482296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1269882619380951, + 0.23144924640655518, + 0.08645021915435791, + 0.1129215657711029, + 0.05018201470375061, + 0.03851279616355896, + 0.047950029373168945, + 0.09931222349405289, + 0.14321750402450562, + 0.06301621347665787 + ], + "credit_map": [ + 0.012650089338421822, + 0.018513735383749008, + 0.009647517465054989, + 0.016955096274614334, + 0.014530804008245468, + 0.020975397899746895, + 0.016395796090364456, + 0.017319800332188606, + 0.021810809150338173, + 0.024998949840664864, + 0.01882399618625641, + 0.017194600775837898, + 0.056972626596689224, + 0.03927353024482727, + 0.02543710358440876, + 0.026937611401081085, + 0.006332815624773502, + 0.011441352777183056, + 0.006246617063879967, + 0.00740351527929306, + 0.010031870566308498, + 0.007930921390652657, + 0.021919872611761093, + 0.015143248252570629, + 0.007118347566574812, + 0.01094383466988802, + 0.009291495196521282, + 0.018171684816479683, + 0.020485904067754745, + 0.012930619530379772, + 0.021267157047986984, + 0.012712515890598297, + 0.005363759119063616, + 0.005109932739287615, + 0.004912531469017267, + 0.007120940368622541, + 0.00719020701944828, + 0.007319397758692503, + 0.007513818331062794, + 0.005651424638926983, + 0.006699255667626858, + 0.004933946765959263, + 0.002697896910831332, + 0.005561359226703644, + 0.004341786727309227, + 0.004837075714021921, + 0.0044829645194113255, + 0.0049585080705583096, + 0.005323132500052452, + 0.004673736169934273, + 0.003124904353171587, + 0.006292786914855242, + 0.006432146765291691, + 0.008190038613975048, + 0.0058706970885396, + 0.008042581379413605, + 0.011775730177760124, + 0.018102368339896202, + 0.006812913343310356, + 0.0057196347042918205, + 0.024964842945337296, + 0.01241412851959467, + 0.008660840801894665, + 0.010861745104193687, + 0.01633201539516449, + 0.029596997424960136, + 0.012795943766832352, + 0.01309742871671915, + 0.027479499578475952, + 0.013488820753991604, + 0.016320673748850822, + 0.014106104150414467, + 0.006442475598305464, + 0.006429270375519991, + 0.005823602434247732, + 0.005306120030581951, + 0.010121453553438187, + 0.011420044116675854, + 0.007253617979586124, + 0.01021962147206068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.008634849451482296 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, intimate close-up, angled slightly downward, captures a pair of slender hands methodically arranging a section of a towering, dark mahogany bookshelf. The skin of the hands is pale and smooth, the movements cautious and gentle. One hand carefully pushes a thick, crimson, leather-bound volume, its spine showing the slight cracks of age and embossed with unremarkable traces of faded gold leaf filigree, until its edge is perfectly flush with the adjacent one. Warm, soft light, as if from a nearby blackout lamp, casts a focused golden glow, highlighting the rich texture of the leather spines—from deep forest green to worn brown—and creating a soft sheen on the manicured fingernails. As the first hand holds the book in place, the fingers of the other delicately trace along the top edge of the aligned volume, ensuring a perfectly flat surface. The background is a soft-focus blur of more books and dark wood, enhancing the shallow depth of field and focusing the viewer's attention on the meticulous action, evoking an atmosphere of tranquility, concentration, and scholarship, imbued with the scent of old paper and leather.", + "seed": 23525, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_1/index_996_seed_23525_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01568845668807626, + "vlm_logprob": -0.01568845668807626, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.01568845668807626, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013310973532497883 + ], + "opposite_logprob_tokens": [ + -5.38831090927124 + ], + "logprob_reward": -0.013310973532497883, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.013310973532497883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10065782815217972, + 0.05136936902999878, + 0.11411544680595398, + 0.05817847326397896, + 0.0848778784275055, + 0.07775911688804626, + 0.12426388263702393, + 0.12216804176568985, + 0.1441354900598526, + 0.1224745586514473 + ], + "credit_map": [ + 0.010006113909184933, + 0.02198171615600586, + 0.008306396193802357, + 0.011041663587093353, + 0.010264295153319836, + 0.014978921040892601, + 0.007411451544612646, + 0.016667267307639122, + 0.005478021688759327, + 0.007356521673500538, + 0.007540060207247734, + 0.004299402702599764, + 0.005719476379454136, + 0.007203143090009689, + 0.005705040879547596, + 0.008067705668509007, + 0.004183663520962, + 0.007703481707721949, + 0.0034941413905471563, + 0.003614068031311035, + 0.0031900934409350157, + 0.045108187943696976, + 0.0046210261061787605, + 0.04220077767968178, + 0.005636400543153286, + 0.004737354815006256, + 0.005178163293749094, + 0.0027984976768493652, + 0.004146247170865536, + 0.015851039439439774, + 0.006006136536598206, + 0.01382463425397873, + 0.005008391570299864, + 0.004629108589142561, + 0.010918380692601204, + 0.004700849764049053, + 0.006548612378537655, + 0.02748083509504795, + 0.0069490838795900345, + 0.018642619252204895, + 0.005105173215270042, + 0.0038996131625026464, + 0.015573092736303806, + 0.009758499450981617, + 0.006874804385006428, + 0.009472215548157692, + 0.010470965877175331, + 0.01660475879907608, + 0.006789787672460079, + 0.010836251080036163, + 0.015148628503084183, + 0.02426076866686344, + 0.014066995121538639, + 0.01918119564652443, + 0.01758497580885887, + 0.016395287588238716, + 0.011771857738494873, + 0.017772294580936432, + 0.017315085977315903, + 0.013043313287198544, + 0.01624138280749321, + 0.015740862116217613, + 0.014611417427659035, + 0.015671828761696815, + 0.011106053367257118, + 0.01293773204088211, + 0.012118128128349781, + 0.016648512333631516, + 0.030397459864616394, + 0.02358386665582657, + 0.019575072452425957, + 0.01776866428554058, + 0.01593346893787384, + 0.011363361030817032, + 0.010521265678107738, + 0.01697627641260624, + 0.013211308978497982, + 0.018686335533857346, + 0.017847755923867226, + 0.017934782430529594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.013310973532497883 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014239490032196045 + ], + "opposite_logprob_tokens": [ + -4.264239311218262 + ], + "logprob_reward": -0.014239490032196045, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014239490032196045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3094482123851776, + 0.057913489639759064, + 0.15790198743343353, + 0.036691077053546906, + 0.0718858540058136, + 0.053786810487508774, + 0.08919999748468399, + 0.0870342031121254, + 0.08506526052951813, + 0.0510731004178524 + ], + "credit_map": [ + 0.02723618410527706, + 0.0515260323882103, + 0.02318393439054489, + 0.04466238245368004, + 0.04375256970524788, + 0.03568211570382118, + 0.03636208921670914, + 0.04704289510846138, + 0.00497333612293005, + 0.006962064187973738, + 0.005498296115547419, + 0.005954396910965443, + 0.0054981340654194355, + 0.010888364166021347, + 0.006996528245508671, + 0.0111423684284091, + 0.005425248295068741, + 0.004747752565890551, + 0.004250118974596262, + 0.0030634861905127764, + 0.0032561724074184895, + 0.06954056024551392, + 0.004112370777875185, + 0.06350629031658173, + 0.003663251642137766, + 0.0032889831345528364, + 0.005265721585601568, + 0.0026345530059188604, + 0.003671490354463458, + 0.008055895566940308, + 0.003245712025091052, + 0.00686546741053462, + 0.004032539203763008, + 0.0061415196396410465, + 0.005698263645172119, + 0.006529605481773615, + 0.007208093069493771, + 0.011390931904315948, + 0.021019570529460907, + 0.009865328669548035, + 0.004042576998472214, + 0.004810255952179432, + 0.0075119719840586185, + 0.006949213333427906, + 0.004387221299111843, + 0.006317500956356525, + 0.012172248214483261, + 0.0075958180241286755, + 0.004528076853603125, + 0.006559889763593674, + 0.008836719207465649, + 0.015603919513523579, + 0.009259232319891453, + 0.013427493162453175, + 0.02165272831916809, + 0.009331927634775639, + 0.0062946449033916, + 0.015967385843396187, + 0.012051154859364033, + 0.009979515336453915, + 0.010060107335448265, + 0.012921437621116638, + 0.010134217329323292, + 0.009625740349292755, + 0.005370937753468752, + 0.007444047834724188, + 0.00666374247521162, + 0.013965181075036526, + 0.012988217175006866, + 0.012275394052267075, + 0.013102716766297817, + 0.013255015015602112, + 0.007782125845551491, + 0.005180357489734888, + 0.004111711401492357, + 0.006783300545066595, + 0.005342625547200441, + 0.007203574292361736, + 0.008232722990214825, + 0.006436680909246206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.014239490032196045 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033712852746248245 + ], + "opposite_logprob_tokens": [ + -3.408712863922119 + ], + "logprob_reward": -0.033712852746248245, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033712852746248245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11442706733942032, + 0.033184923231601715, + 0.35678619146347046, + 0.03928912803530693, + 0.08060291409492493, + 0.052419357001781464, + 0.14461758732795715, + 0.04568459466099739, + 0.09127961099147797, + 0.041708655655384064 + ], + "credit_map": [ + 0.012261537835001945, + 0.014522382989525795, + 0.007559799589216709, + 0.013256563805043697, + 0.013308492489159107, + 0.01630397140979767, + 0.014256155118346214, + 0.02295815758407116, + 0.003200795268639922, + 0.0060122050344944, + 0.003122337395325303, + 0.0029119320679455996, + 0.003588121384382248, + 0.00469196168705821, + 0.003717520274221897, + 0.005940049886703491, + 0.011075657792389393, + 0.008663292974233627, + 0.01049045380204916, + 0.007789153140038252, + 0.008789287880063057, + 0.15557868778705597, + 0.011385275050997734, + 0.1430143564939499, + 0.00420330511406064, + 0.006207927595824003, + 0.004433573689311743, + 0.0029178240802139044, + 0.005297338590025902, + 0.005518263205885887, + 0.005669548176229, + 0.0050413464196026325, + 0.005628224462270737, + 0.00693622836843133, + 0.008393662981688976, + 0.0067177265882492065, + 0.008169295266270638, + 0.014406991191208363, + 0.0196513831615448, + 0.010699405334889889, + 0.003628662321716547, + 0.0029330134857445955, + 0.010266820900142193, + 0.00518030347302556, + 0.0050959838554263115, + 0.006282478105276823, + 0.008020575158298016, + 0.011011517606675625, + 0.008717140182852745, + 0.011583825573325157, + 0.01806679740548134, + 0.026633096858859062, + 0.016713321208953857, + 0.022188615053892136, + 0.026832392439246178, + 0.013882379047572613, + 0.0041446019895374775, + 0.0059120007790625095, + 0.005986835341900587, + 0.007360520772635937, + 0.004863003734499216, + 0.00624838238582015, + 0.005984796676784754, + 0.005184451583772898, + 0.007342588622123003, + 0.010885076597332954, + 0.009961407631635666, + 0.008334080688655376, + 0.01749270223081112, + 0.014193336479365826, + 0.012140128761529922, + 0.010930284857749939, + 0.006027000956237316, + 0.005608442705124617, + 0.00404652114957571, + 0.006082926411181688, + 0.0033175372518599033, + 0.006100127007812262, + 0.0049021802842617035, + 0.005623916629701853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.033712852746248245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004644202534109354 + ], + "opposite_logprob_tokens": [ + -5.379644393920898 + ], + "logprob_reward": -0.004644202534109354, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004644202534109354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10450129210948944, + 0.07955867052078247, + 0.10127884894609451, + 0.14431415498256683, + 0.180741548538208, + 0.0943840891122818, + 0.07516220957040787, + 0.08173251152038574, + 0.06335410475730896, + 0.07497259229421616 + ], + "credit_map": [ + 0.011004561558365822, + 0.01641332358121872, + 0.008274073712527752, + 0.011534574441611767, + 0.010633870027959347, + 0.016406286507844925, + 0.011247453279793262, + 0.018987150862812996, + 0.006610948126763105, + 0.011200792156159878, + 0.010480179451406002, + 0.006946133449673653, + 0.008448174223303795, + 0.014471059665083885, + 0.007499973755329847, + 0.013901407830417156, + 0.012131020426750183, + 0.013331355527043343, + 0.005505627021193504, + 0.0059275515377521515, + 0.006332649849355221, + 0.027449071407318115, + 0.006458616349846125, + 0.02414296381175518, + 0.010354739613831043, + 0.009807432070374489, + 0.008539022877812386, + 0.007098591886460781, + 0.009343662299215794, + 0.04684566706418991, + 0.011510300450026989, + 0.04081474989652634, + 0.010767899453639984, + 0.01230971422046423, + 0.016312681138515472, + 0.00952420849353075, + 0.011117901653051376, + 0.05165598914027214, + 0.029216786846518517, + 0.03983638063073158, + 0.0055459365248680115, + 0.00516402255743742, + 0.01617620140314102, + 0.01039664726704359, + 0.007307203020900488, + 0.011915566399693489, + 0.015301071107387543, + 0.022577444091439247, + 0.0037646712735295296, + 0.004964780993759632, + 0.0073499735444784164, + 0.020095758140087128, + 0.007524931337684393, + 0.012870877981185913, + 0.008883814327418804, + 0.009707407094538212, + 0.0047853682190179825, + 0.013556737452745438, + 0.010507911443710327, + 0.011546489782631397, + 0.009906281717121601, + 0.010627001523971558, + 0.010802343487739563, + 0.010000384412705898, + 0.004120469093322754, + 0.007615397218614817, + 0.006499903276562691, + 0.00773980375379324, + 0.007633601315319538, + 0.009136461652815342, + 0.010018182918429375, + 0.010590290650725365, + 0.015266411937773228, + 0.010230091400444508, + 0.007477463223040104, + 0.009177938103675842, + 0.006162541452795267, + 0.0103531489148736, + 0.007958229631185532, + 0.008346769027411938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.004644202534109354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534764595329762 + ], + "opposite_logprob_tokens": [ + -4.3875346183776855 + ], + "logprob_reward": -0.012534764595329762, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534764595329762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08190162479877472, + 0.05698680132627487, + 0.20891039073467255, + 0.07979818433523178, + 0.14835447072982788, + 0.08918088674545288, + 0.12278847396373749, + 0.0724937692284584, + 0.0672006830573082, + 0.07238474488258362 + ], + "credit_map": [ + 0.007770191412419081, + 0.01505327969789505, + 0.007786357309669256, + 0.011615252122282982, + 0.009609593078494072, + 0.00943452026695013, + 0.008715968579053879, + 0.011916449293494225, + 0.005395231768488884, + 0.007369630038738251, + 0.00796938594430685, + 0.006027068942785263, + 0.005842200480401516, + 0.009431160055100918, + 0.006188183557242155, + 0.008763937279582024, + 0.008916187100112438, + 0.008370871655642986, + 0.0070505994372069836, + 0.004656045697629452, + 0.0049241031520068645, + 0.08666296303272247, + 0.007556319236755371, + 0.08077328652143478, + 0.006155930459499359, + 0.010543374344706535, + 0.006948074791580439, + 0.0051294839940965176, + 0.005897738039493561, + 0.021323058754205704, + 0.005430795252323151, + 0.018369723111391068, + 0.00975977536290884, + 0.010621888563036919, + 0.013301664963364601, + 0.00783306173980236, + 0.005618923809379339, + 0.042596157640218735, + 0.026841064915060997, + 0.03178191930055618, + 0.0064874934032559395, + 0.006108093541115522, + 0.015270191244781017, + 0.01910298317670822, + 0.00779676903039217, + 0.009348046034574509, + 0.013296475633978844, + 0.011770823039114475, + 0.008784070611000061, + 0.012843457981944084, + 0.017804546281695366, + 0.017805125564336777, + 0.015034576877951622, + 0.01748017407953739, + 0.01921526901423931, + 0.013821251690387726, + 0.005356526933610439, + 0.009588775224983692, + 0.0070387134328484535, + 0.0089655676856637, + 0.009337699972093105, + 0.009664926677942276, + 0.011888369917869568, + 0.010653176344931126, + 0.00622620340436697, + 0.011557183228433132, + 0.007741488516330719, + 0.007735914550721645, + 0.008638174273073673, + 0.009188621304929256, + 0.00819327775388956, + 0.007919815368950367, + 0.01627795398235321, + 0.010018945671617985, + 0.006316610146313906, + 0.007334229536354542, + 0.006189546547830105, + 0.009764964692294598, + 0.00827060453593731, + 0.008211886510252953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -0.012534764595329762 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, intimate close-up, angled slightly downward, captures a pair of slender hands methodically arranging a section of a towering, dark mahogany bookshelf. The skin of the hands is pale and smooth, the movements cautious and gentle. One hand carefully pushes a thick, crimson, leather-bound volume, its spine showing the slight cracks of age and embossed with unremarkable traces of faded gold leaf filigree, until its edge is perfectly flush with the adjacent one. Warm, soft light, as if from a nearby blackout lamp, casts a focused golden glow, highlighting the rich texture of the leather spines—from deep forest green to worn brown—and creating a soft sheen on the manicured fingernails. As the first hand holds the book in place, the fingers of the other delicately trace along the top edge of the aligned volume, ensuring a perfectly flat surface. The background is a soft-focus blur of more books and dark wood, enhancing the shallow depth of field and focusing the viewer's attention on the meticulous action, evoking an atmosphere of tranquility, concentration, and scholarship, imbued with the scent of old paper and leather.", + "seed": 23526, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_1/index_996_seed_23526_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.010566114936955273, + "vlm_logprob": -0.010566114936955273, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.010566114936955273, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009567243978381157 + ], + "opposite_logprob_tokens": [ + -5.8845672607421875 + ], + "logprob_reward": -0.009567243978381157, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.009567243978381157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12665408849716187, + 0.0959043875336647, + 0.04780266061425209, + 0.0469110831618309, + 0.08932717889547348, + 0.07567957043647766, + 0.10759252309799194, + 0.20226165652275085, + 0.12416937947273254, + 0.08369747549295425 + ], + "credit_map": [ + 0.009759635664522648, + 0.017526814714074135, + 0.015618206933140755, + 0.013701842166483402, + 0.01693546213209629, + 0.023256903514266014, + 0.01287887990474701, + 0.016976360231637955, + 0.010135963559150696, + 0.01087490376085043, + 0.015215450897812843, + 0.017238158732652664, + 0.008009298704564571, + 0.010794427245855331, + 0.011641633696854115, + 0.011994552798569202, + 0.0037918630987405777, + 0.00773665914312005, + 0.005977013148367405, + 0.00737219350412488, + 0.004288058262318373, + 0.00832610484212637, + 0.0039801327511668205, + 0.006330637726932764, + 0.004948274698108435, + 0.005525608547031879, + 0.0061254664324223995, + 0.007645867299288511, + 0.006783558055758476, + 0.00515535194426775, + 0.0040404596365988255, + 0.006686500273644924, + 0.012468439526855946, + 0.011339899152517319, + 0.007622054778039455, + 0.010124915279448032, + 0.01748812571167946, + 0.012089747004210949, + 0.006074124481528997, + 0.012119874358177185, + 0.008109967224299908, + 0.009080055169761181, + 0.011632915586233139, + 0.00796753354370594, + 0.014231355860829353, + 0.00835210457444191, + 0.005744314286857843, + 0.010561322793364525, + 0.013188297860324383, + 0.00709477998316288, + 0.009980831295251846, + 0.012396560981869698, + 0.01541842333972454, + 0.024759560823440552, + 0.007244765292853117, + 0.017509307712316513, + 0.02234923094511032, + 0.013404139317572117, + 0.011317601427435875, + 0.016460472717881203, + 0.05013512447476387, + 0.04537329822778702, + 0.013685645535588264, + 0.029536135494709015, + 0.009358488023281097, + 0.009700578637421131, + 0.008688059635460377, + 0.013762926682829857, + 0.012139638885855675, + 0.030178721994161606, + 0.020991051569581032, + 0.019349917769432068, + 0.008131717331707478, + 0.019789312034845352, + 0.007059205323457718, + 0.00952161941677332, + 0.0070630647242069244, + 0.00856361910700798, + 0.008171100169420242, + 0.015397834591567516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.009567243978381157 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0161100123077631 + ], + "opposite_logprob_tokens": [ + -4.141109943389893 + ], + "logprob_reward": -0.0161100123077631, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0161100123077631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06000299006700516, + 0.027046164497733116, + 0.021694809198379517, + 0.0290187057107687, + 0.07558871805667877, + 0.07089205831289291, + 0.06484480947256088, + 0.33594298362731934, + 0.07724721729755402, + 0.23772156238555908 + ], + "credit_map": [ + 0.0042482661083340645, + 0.0065017943270504475, + 0.005113317631185055, + 0.004312317818403244, + 0.010554099455475807, + 0.011434940621256828, + 0.0047209179028868675, + 0.013117331080138683, + 0.0023472080938518047, + 0.00427259411662817, + 0.004227547440677881, + 0.002610882744193077, + 0.0029460720252245665, + 0.004651458002626896, + 0.002653609961271286, + 0.003336793277412653, + 0.0026084338314831257, + 0.0019294342491775751, + 0.003239614190533757, + 0.0023622470907866955, + 0.002793010324239731, + 0.003923052456229925, + 0.0016265478916466236, + 0.0032124684657901525, + 0.0027039044070988894, + 0.002971624955534935, + 0.003081217175349593, + 0.004863088950514793, + 0.005873359274119139, + 0.003685690462589264, + 0.00237018964253366, + 0.0034696292132139206, + 0.00980729516595602, + 0.00796339102089405, + 0.00658183079212904, + 0.01148530375212431, + 0.01229681633412838, + 0.012158246710896492, + 0.004941234830766916, + 0.010354593396186829, + 0.00739839905872941, + 0.006251294165849686, + 0.00981493853032589, + 0.008528889156877995, + 0.01934463158249855, + 0.006373512092977762, + 0.004237147513777018, + 0.00894323829561472, + 0.007210343610495329, + 0.006140151061117649, + 0.006697187665849924, + 0.0076591623947024345, + 0.013417246751487255, + 0.008653038181364536, + 0.005775950849056244, + 0.009291724301874638, + 0.02384992502629757, + 0.01831303909420967, + 0.022880520671606064, + 0.03315167874097824, + 0.0768062174320221, + 0.08326313644647598, + 0.02423708513379097, + 0.053441356867551804, + 0.005021247547119856, + 0.005859688390046358, + 0.00737796351313591, + 0.008874238468706608, + 0.015417546033859253, + 0.011011666618287563, + 0.012565750628709793, + 0.011119108647108078, + 0.013777916319668293, + 0.07476654648780823, + 0.009442615322768688, + 0.023730739951133728, + 0.019091298803687096, + 0.016360003501176834, + 0.019327830523252487, + 0.061224598437547684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.0161100123077631 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192220479249954 + ], + "opposite_logprob_tokens": [ + -4.018192291259766 + ], + "logprob_reward": -0.018192220479249954, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018192220479249954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1537102311849594, + 0.05155009403824806, + 0.04505259916186333, + 0.04563462734222412, + 0.05875379964709282, + 0.11849517375230789, + 0.025199024006724358, + 0.049941517412662506, + 0.16502165794372559, + 0.28664126992225647 + ], + "credit_map": [ + 0.01710410788655281, + 0.013917352072894573, + 0.023139141499996185, + 0.02891174517571926, + 0.015611615963280201, + 0.01551525853574276, + 0.02192276157438755, + 0.017588265240192413, + 0.004358404316008091, + 0.007978764362633228, + 0.008119890466332436, + 0.007399131543934345, + 0.003802874591201544, + 0.007681414019316435, + 0.005587469786405563, + 0.00662215193733573, + 0.00451319757848978, + 0.006750414147973061, + 0.0064751687459647655, + 0.004301583860069513, + 0.005780065432190895, + 0.0073724305257201195, + 0.003433407051488757, + 0.006426332518458366, + 0.0030200164765119553, + 0.003725856076925993, + 0.004864460323005915, + 0.004795832559466362, + 0.011175881139934063, + 0.007384589407593012, + 0.0035318045411258936, + 0.007136190310120583, + 0.008563337847590446, + 0.00602182000875473, + 0.00680078798905015, + 0.006031072232872248, + 0.00724475784227252, + 0.011334620416164398, + 0.004734862130135298, + 0.008022550493478775, + 0.012501838617026806, + 0.00800263974815607, + 0.014041303656995296, + 0.014674943871796131, + 0.029068848118185997, + 0.015458964742720127, + 0.00741172581911087, + 0.017334919422864914, + 0.0023178355768322945, + 0.0014576615067198873, + 0.0016029338585212827, + 0.0026222809683531523, + 0.003457197453826666, + 0.007815496996045113, + 0.0015353119233623147, + 0.0043903072364628315, + 0.004209531005471945, + 0.0032698730938136578, + 0.004149560816586018, + 0.00390032771974802, + 0.010821285657584667, + 0.012555422261357307, + 0.003597796196117997, + 0.007437725085765123, + 0.018627064302563667, + 0.014419298619031906, + 0.008795840665698051, + 0.01508236676454544, + 0.015969250351190567, + 0.04740172624588013, + 0.017116574570536613, + 0.02760954573750496, + 0.015196139924228191, + 0.09950312972068787, + 0.008949369192123413, + 0.02231307327747345, + 0.019624989479780197, + 0.01877346821129322, + 0.019110726192593575, + 0.08317039161920547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.018192220479249954 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013381821336224675 + ], + "opposite_logprob_tokens": [ + -6.626338005065918 + ], + "logprob_reward": -0.0013381821336224675, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013381821336224675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16016729176044464, + 0.05961400270462036, + 0.0544184148311615, + 0.058187298476696014, + 0.06329183280467987, + 0.0486847385764122, + 0.04956100136041641, + 0.06616324931383133, + 0.1887858808040619, + 0.25112631916999817 + ], + "credit_map": [ + 0.013780592009425163, + 0.020063186064362526, + 0.025013402104377747, + 0.0155527638271451, + 0.018652915954589844, + 0.029811426997184753, + 0.011011478491127491, + 0.026281526312232018, + 0.005843173246830702, + 0.009181797504425049, + 0.006433508358895779, + 0.010537521913647652, + 0.00838399026542902, + 0.006103423424065113, + 0.005402667447924614, + 0.0077279191464185715, + 0.007848158478736877, + 0.007509117946028709, + 0.008852735161781311, + 0.006116223521530628, + 0.0077080074697732925, + 0.006098648998886347, + 0.004246021155267954, + 0.0060394988395273685, + 0.005696739070117474, + 0.004816635046154261, + 0.00510540883988142, + 0.00615690415725112, + 0.012758967466652393, + 0.011017822660505772, + 0.004155759233981371, + 0.008479060605168343, + 0.010090073570609093, + 0.005379143171012402, + 0.006882688496261835, + 0.009523245505988598, + 0.010668987408280373, + 0.00777021236717701, + 0.003311034757643938, + 0.009666451252996922, + 0.006937700789421797, + 0.0035177250392735004, + 0.00488133542239666, + 0.007356996648013592, + 0.011777977459132671, + 0.0053222207352519035, + 0.002140667289495468, + 0.00675011333078146, + 0.005727244075387716, + 0.004098701290786266, + 0.005300526972860098, + 0.005702335853129625, + 0.00911407358944416, + 0.007528460584580898, + 0.0045996312983334064, + 0.007490027695894241, + 0.0083188870921731, + 0.005349881015717983, + 0.005555351264774799, + 0.006282234564423561, + 0.014668019488453865, + 0.010904930531978607, + 0.007230560760945082, + 0.007853380404412746, + 0.0232899971306324, + 0.021625524386763573, + 0.02817363850772381, + 0.03623843938112259, + 0.017777713015675545, + 0.019742704927921295, + 0.021226031705737114, + 0.020711828023195267, + 0.014279915019869804, + 0.08642079681158066, + 0.009272633120417595, + 0.021102894097566605, + 0.017617905512452126, + 0.015434246510267258, + 0.016921542584896088, + 0.0700763687491417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0013381821336224675 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of slender, pale hands visible arranging books on a dark mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand pushing a thick, crimson, leather-bound volume with a spine showing cracks and faded gold leaf filigree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the other hand delicately trace along the top edge of the aligned volume to ensure a flat surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting warm and soft, casting a focused golden glow that highlights the texture of the leather spines and manicured fingernails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, intimate close-up angled slightly downward with a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622915785759687 + ], + "opposite_logprob_tokens": [ + -4.882622718811035 + ], + "logprob_reward": -0.007622915785759687, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622915785759687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08202694356441498, + 0.043718431144952774, + 0.04666434973478317, + 0.08589410781860352, + 0.07151176035404205, + 0.057965509593486786, + 0.05363943800330162, + 0.19595341384410858, + 0.047340862452983856, + 0.3152851462364197 + ], + "credit_map": [ + 0.006321866996586323, + 0.006452462170273066, + 0.013815096579492092, + 0.006249185185879469, + 0.01272683683782816, + 0.013138294219970703, + 0.008363365195691586, + 0.014959840103983879, + 0.004702589940279722, + 0.0040758224204182625, + 0.00724537530913949, + 0.004755658563226461, + 0.006256043910980225, + 0.007128472439944744, + 0.004104591906070709, + 0.005449878983199596, + 0.0052027334459125996, + 0.0065611121244728565, + 0.009056059643626213, + 0.004826194141060114, + 0.006156340707093477, + 0.005839362740516663, + 0.00379610201343894, + 0.00522644491866231, + 0.013131367973983288, + 0.008912740275263786, + 0.008484532125294209, + 0.009594901464879513, + 0.011552160605788231, + 0.01595151238143444, + 0.00691832322627306, + 0.011348573490977287, + 0.008736701682209969, + 0.00930153951048851, + 0.009112573228776455, + 0.008240900933742523, + 0.010169110260903835, + 0.007917472161352634, + 0.007655108347535133, + 0.010378357023000717, + 0.006269791629165411, + 0.004913254175335169, + 0.0064052799716591835, + 0.007629879750311375, + 0.012367225252091885, + 0.0059415618889033794, + 0.0050812833942472935, + 0.009357236325740814, + 0.00742513919249177, + 0.004232886712998152, + 0.005867379251867533, + 0.007719242945313454, + 0.006303716916590929, + 0.010001803748309612, + 0.0044506690464913845, + 0.007638599257916212, + 0.0144804073497653, + 0.016668686643242836, + 0.022036293521523476, + 0.030185842886567116, + 0.027101268991827965, + 0.03545646369457245, + 0.02146328054368496, + 0.028561176732182503, + 0.0039857798255980015, + 0.004875426646322012, + 0.004763422533869743, + 0.006741291377693415, + 0.007593173999339342, + 0.006967946421355009, + 0.006232398562133312, + 0.006181421689689159, + 0.020327381789684296, + 0.08705693483352661, + 0.01596350595355034, + 0.03834444284439087, + 0.030180713161826134, + 0.020368510857224464, + 0.030971068888902664, + 0.0720725730061531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.007622915785759687 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_10/index_tensor([3990, 3990])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_10/index_tensor([3990, 3990])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3407544d46ee7eaddff167cb1832627e3bc1f088 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_10/index_tensor([3990, 3990])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures a shrub rendered in a captivating sketch-style ink and watercolor, its forms swaying gently in the breeze. The shrub is covered in fluffy, pom-pom-like flowers, each a vibrant canary-yellow globe of delicate, silken petals. Delicate, expressive black ink lines loosely define the shapes of the graceful, vine-like branches and sparse green leaves, adding a graphic quality to the organic forms. The entire plant sways rhythmically back and forth, and this subtle movement causes each flower cluster to subtly tremble. The artistry is highlighted by the visible soft colors, where the translucent yellow watercolor meets the green of the leaves, creating a beautiful wet-on-wet effect. The background, an abstract palette of pale green and earthy brown, focuses entirely on the shrub's gentle movement. The scene is illuminated by a soft, diffuse light, enhancing the brightness of the colors and conveying an atmosphere of tranquility, peace, and artistic refinement.", + "seed": 581, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_10/index_3990_seed_581_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05086653418838978, + "vlm_logprob": -0.05086653418838978, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.05086653418838978, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005016953218728304 + ], + "opposite_logprob_tokens": [ + -5.505016803741455 + ], + "logprob_reward": -0.005016953218728304, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005016953218728304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07620198279619217, + 0.027515767142176628, + 0.015411026775836945, + 0.023998025804758072, + 0.04869203642010689, + 0.400474488735199, + 0.020498227328062057, + 0.23516058921813965, + 0.055146194994449615, + 0.09690166264772415 + ], + "credit_map": [ + 0.008755980990827084, + 0.00695745088160038, + 0.00805966928601265, + 0.00997966155409813, + 0.008517203852534294, + 0.011567372828722, + 0.009534427896142006, + 0.012830215506255627, + 0.0033117595594376326, + 0.002363980747759342, + 0.003934639971703291, + 0.003040764480829239, + 0.003006537677720189, + 0.0041396645829081535, + 0.0032619743142277002, + 0.00445644510909915, + 0.0028221250977367163, + 0.0011566663160920143, + 0.0014569920022040606, + 0.001752151525579393, + 0.0015620335470885038, + 0.002231256105005741, + 0.0019299171399325132, + 0.002499884460121393, + 0.002790143946185708, + 0.0024448141921311617, + 0.002652403200045228, + 0.001916229142807424, + 0.0014664416667073965, + 0.006173847708851099, + 0.0015828456962481141, + 0.004971299786120653, + 0.0051238336600363255, + 0.0032036437187343836, + 0.004075796343386173, + 0.0030662885401397943, + 0.0032939184457063675, + 0.01572134904563427, + 0.0033478233963251114, + 0.010859386995434761, + 0.015695758163928986, + 0.01052254531532526, + 0.02694788947701454, + 0.015195571817457676, + 0.013264195993542671, + 0.16964569687843323, + 0.009896200150251389, + 0.13930661976337433, + 0.002924768254160881, + 0.0023054617922753096, + 0.003679107641801238, + 0.0026892866007983685, + 0.0013815198326483369, + 0.002534157130867243, + 0.002331417752429843, + 0.002652510302141309, + 0.008280259557068348, + 0.004536458756774664, + 0.01454218290746212, + 0.005486681591719389, + 0.007401145063340664, + 0.09490153938531876, + 0.005960024427622557, + 0.09405229985713959, + 0.005185899790376425, + 0.004160306416451931, + 0.006214857567101717, + 0.004383056424558163, + 0.004267186392098665, + 0.012797224335372448, + 0.0041517652571201324, + 0.013985899277031422, + 0.008245428092777729, + 0.002893803408369422, + 0.027349844574928284, + 0.0045498511753976345, + 0.004523570649325848, + 0.036940619349479675, + 0.004222828894853592, + 0.008175717666745186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.005016953218728304 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381055623292923 + ], + "opposite_logprob_tokens": [ + -3.2881054878234863 + ], + "logprob_reward": -0.0381055623292923, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381055623292923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05177178978919983, + 0.042188093066215515, + 0.054828815162181854, + 0.020169982686638832, + 0.14359287917613983, + 0.08478712290525436, + 0.06220034882426262, + 0.15114733576774597, + 0.3026304841041565, + 0.08668308705091476 + ], + "credit_map": [ + 0.005458845756947994, + 0.006819070316851139, + 0.00608572643250227, + 0.008431334979832172, + 0.0047665489837527275, + 0.006657607853412628, + 0.006058041006326675, + 0.007494622375816107, + 0.005888367537409067, + 0.002760072937235236, + 0.004298647399991751, + 0.004788343794643879, + 0.0026748802047222853, + 0.007671209052205086, + 0.004271394573152065, + 0.009835184551775455, + 0.00932483933866024, + 0.0036702819634228945, + 0.005023702513426542, + 0.006707337684929371, + 0.00558900460600853, + 0.011959274299442768, + 0.004195210989564657, + 0.008359165862202644, + 0.002562736626714468, + 0.0036901026032865047, + 0.002130120526999235, + 0.00267096352763474, + 0.0017875912599265575, + 0.0026950316969305277, + 0.0017560875276103616, + 0.0028773508965969086, + 0.023769231513142586, + 0.007630092557519674, + 0.022546615451574326, + 0.010416124016046524, + 0.015264879912137985, + 0.031737539917230606, + 0.009425003081560135, + 0.022803403437137604, + 0.003484145738184452, + 0.002460436662659049, + 0.005957281216979027, + 0.0032652842346578836, + 0.0027127403300255537, + 0.034991614520549774, + 0.0024417638778686523, + 0.02947387285530567, + 0.009345886297523975, + 0.006238394416868687, + 0.008297301828861237, + 0.007672977168112993, + 0.005205573048442602, + 0.011179216206073761, + 0.006812477950006723, + 0.007448531687259674, + 0.010684830136597157, + 0.0054248301312327385, + 0.01847740449011326, + 0.005751195829361677, + 0.0048939818516373634, + 0.048884324729442596, + 0.0051572248339653015, + 0.051873546093702316, + 0.026723599061369896, + 0.009018308483064175, + 0.05701008439064026, + 0.015689976513385773, + 0.015273839235305786, + 0.11040118336677551, + 0.011278956197202206, + 0.057234544306993484, + 0.010081361047923565, + 0.0028810319490730762, + 0.020304439589381218, + 0.004426538944244385, + 0.004514115396887064, + 0.031596384942531586, + 0.005012917798012495, + 0.007866300642490387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.0381055623292923 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896047085523605 + ], + "opposite_logprob_tokens": [ + -2.578960418701172 + ], + "logprob_reward": -0.07896047085523605, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896047085523605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07920090109109879, + 0.02480125240981579, + 0.04966580495238304, + 0.033185191452503204, + 0.06027233973145485, + 0.39174920320510864, + 0.04476441070437431, + 0.15024256706237793, + 0.08468571305274963, + 0.08143262565135956 + ], + "credit_map": [ + 0.00900433212518692, + 0.010163390077650547, + 0.008666903711855412, + 0.013413303531706333, + 0.008246391080319881, + 0.009918682277202606, + 0.009142397902905941, + 0.010645493865013123, + 0.0031852370593696833, + 0.0025502447970211506, + 0.0027863287832587957, + 0.0029198480769991875, + 0.0026585818268358707, + 0.0037753332871943712, + 0.002407957799732685, + 0.004517721943557262, + 0.005819642450660467, + 0.0026231100782752037, + 0.004899202845990658, + 0.003688254626467824, + 0.0037892740219831467, + 0.013033014722168446, + 0.002765656914561987, + 0.013047645799815655, + 0.004527909215539694, + 0.0026385134551674128, + 0.004218440502882004, + 0.0035601358395069838, + 0.0040245335549116135, + 0.005296352785080671, + 0.003302064025774598, + 0.00561723904684186, + 0.007376655470579863, + 0.004667692817747593, + 0.008150172419846058, + 0.005952238105237484, + 0.006479916162788868, + 0.010047982446849346, + 0.006528391968458891, + 0.011069287545979023, + 0.014469166286289692, + 0.010509044863283634, + 0.026979787275195122, + 0.014443905092775822, + 0.012121710926294327, + 0.16646303236484528, + 0.00979707296937704, + 0.13696545362472534, + 0.006659436039626598, + 0.007292045745998621, + 0.006006530951708555, + 0.005326474085450172, + 0.0029411688446998596, + 0.005913223139941692, + 0.005383342504501343, + 0.005242185201495886, + 0.007032972760498524, + 0.0042619784362614155, + 0.012635807506740093, + 0.004844378679990768, + 0.005297699943184853, + 0.05564073473215103, + 0.004787611775100231, + 0.05574136599898338, + 0.010525842197239399, + 0.005485848523676395, + 0.02016272395849228, + 0.00572759285569191, + 0.0052416459657251835, + 0.02377243898808956, + 0.004865747876465321, + 0.008903865702450275, + 0.00817521195858717, + 0.003396096406504512, + 0.02212260663509369, + 0.003251780057325959, + 0.004784582182765007, + 0.02660590037703514, + 0.00543084554374218, + 0.007665589917451143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.07896047085523605 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005245612468570471 + ], + "opposite_logprob_tokens": [ + -5.255245685577393 + ], + "logprob_reward": -0.005245612468570471, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005245612468570471, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04676980897784233, + 0.057928912341594696, + 0.06322871893644333, + 0.0977281704545021, + 0.031037768349051476, + 0.18195559084415436, + 0.04884769767522812, + 0.22330021858215332, + 0.19713439047336578, + 0.05206869915127754 + ], + "credit_map": [ + 0.006184758152812719, + 0.005345282144844532, + 0.006157261319458485, + 0.005588208325207233, + 0.003455128287896514, + 0.006208708975464106, + 0.005931158550083637, + 0.007899302057921886, + 0.011272205039858818, + 0.00459116930142045, + 0.005345947574824095, + 0.004357119090855122, + 0.00545066362246871, + 0.00957738421857357, + 0.003946810960769653, + 0.01338761206716299, + 0.01018440816551447, + 0.0036618534941226244, + 0.006080013699829578, + 0.00799156166613102, + 0.006707374472171068, + 0.008760364726185799, + 0.00640837149694562, + 0.013434767723083496, + 0.013048826716840267, + 0.011046620085835457, + 0.009875917807221413, + 0.007938621565699577, + 0.007545220199972391, + 0.028491578996181488, + 0.004375847522169352, + 0.015405543148517609, + 0.0034655919298529625, + 0.004089738242328167, + 0.0030117055866867304, + 0.0031911188270896673, + 0.003310485975816846, + 0.007038937881588936, + 0.0026469763834029436, + 0.004283211659640074, + 0.008804536424577236, + 0.005618750117719173, + 0.013341366313397884, + 0.007603958249092102, + 0.0070385863073170185, + 0.07444465905427933, + 0.004935984965413809, + 0.06016775593161583, + 0.005608673673123121, + 0.00460286159068346, + 0.009609653614461422, + 0.004086425062268972, + 0.0030835652723908424, + 0.009145132265985012, + 0.003411645069718361, + 0.009299742057919502, + 0.009258822537958622, + 0.004972836934030056, + 0.01625017635524273, + 0.005992162507027388, + 0.0067914510145783424, + 0.08497153222560883, + 0.006223898846656084, + 0.0888393223285675, + 0.018122628331184387, + 0.005868226755410433, + 0.0371512696146965, + 0.010145102627575397, + 0.009735186584293842, + 0.07076472043991089, + 0.006822061724960804, + 0.03852519020438194, + 0.005839164834469557, + 0.0022213051561266184, + 0.013836665078997612, + 0.002928879577666521, + 0.0028515695594251156, + 0.016060063615441322, + 0.0023252638056874275, + 0.0060057854279875755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.005245612468570471 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12700407207012177 + ], + "opposite_logprob_tokens": [ + -2.1270041465759277 + ], + "logprob_reward": -0.12700407207012177, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12700407207012177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05910397320985794, + 0.05167863517999649, + 0.044486526399850845, + 0.04382991045713425, + 0.06791619956493378, + 0.21542707085609436, + 0.037333495914936066, + 0.24926835298538208, + 0.1867363303899765, + 0.0442194789648056 + ], + "credit_map": [ + 0.007419726345688105, + 0.005833330564200878, + 0.005084099713712931, + 0.005622526165097952, + 0.00875992700457573, + 0.011194263584911823, + 0.0047909049317240715, + 0.010399198159575462, + 0.007834198884665966, + 0.0030268728733062744, + 0.005174907390028238, + 0.0046041375026106834, + 0.006056307349354029, + 0.009272384457290173, + 0.004988997709006071, + 0.010720835067331791, + 0.008503476157784462, + 0.0023678792640566826, + 0.004105030559003353, + 0.004580319859087467, + 0.004336575977504253, + 0.007017642259597778, + 0.004080291371792555, + 0.009495317935943604, + 0.006792040541768074, + 0.0037054128479212523, + 0.004241786897182465, + 0.0036759839858859777, + 0.002990860491991043, + 0.01165340468287468, + 0.0023809417616575956, + 0.008389485068619251, + 0.0058078342117369175, + 0.004115267191082239, + 0.0061509935185313225, + 0.0036512643564492464, + 0.004473262000828981, + 0.02049412950873375, + 0.0036581892054528, + 0.019565263763070107, + 0.008561389520764351, + 0.006784756202250719, + 0.014928563497960567, + 0.008503335528075695, + 0.007493302691727877, + 0.09107796847820282, + 0.005441334564238787, + 0.0726364254951477, + 0.0050109922885894775, + 0.002797092776745558, + 0.005951495375484228, + 0.003294846508651972, + 0.0027967230416834354, + 0.007808358874171972, + 0.004178779199719429, + 0.005495209246873856, + 0.008633875288069248, + 0.004661740735173225, + 0.017253557220101357, + 0.005763278342783451, + 0.007545110769569874, + 0.09981083124876022, + 0.005824823398143053, + 0.09977515786886215, + 0.021509937942028046, + 0.008952693082392216, + 0.042923375964164734, + 0.010738215409219265, + 0.012272863648831844, + 0.05736720189452171, + 0.012768412008881569, + 0.02020364999771118, + 0.004532008431851864, + 0.002347109606489539, + 0.010801736265420914, + 0.0023745675571262836, + 0.0030799624510109425, + 0.01388455554842949, + 0.002870283555239439, + 0.004329259507358074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.12700407207012177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a shrub rendered in a captivating sketch-style ink and watercolor, its forms swaying gently in the breeze. The shrub is covered in fluffy, pom-pom-like flowers, each a vibrant canary-yellow globe of delicate, silken petals. Delicate, expressive black ink lines loosely define the shapes of the graceful, vine-like branches and sparse green leaves, adding a graphic quality to the organic forms. The entire plant sways rhythmically back and forth, and this subtle movement causes each flower cluster to subtly tremble. The artistry is highlighted by the visible soft colors, where the translucent yellow watercolor meets the green of the leaves, creating a beautiful wet-on-wet effect. The background, an abstract palette of pale green and earthy brown, focuses entirely on the shrub's gentle movement. The scene is illuminated by a soft, diffuse light, enhancing the brightness of the colors and conveying an atmosphere of tranquility, peace, and artistic refinement.", + "seed": 582, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_10/index_3990_seed_582_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06509141884744167, + "vlm_logprob": -0.06509141884744167, + "vlm_yes_no_margin": 3.4000000238418577, + "vlm_token_credit": -0.06509141884744167, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006159844808280468 + ], + "opposite_logprob_tokens": [ + -5.381159782409668 + ], + "logprob_reward": -0.006159844808280468, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.006159844808280468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2804018557071686, + 0.0632738322019577, + 0.05499826744198799, + 0.08656254410743713, + 0.058526258915662766, + 0.03863552585244179, + 0.02531563676893711, + 0.10146862268447876, + 0.16322174668312073, + 0.12759573757648468 + ], + "credit_map": [ + 0.04002094268798828, + 0.04157333821058273, + 0.048960115760564804, + 0.02586383931338787, + 0.02726103365421295, + 0.0334194079041481, + 0.03983893245458603, + 0.023464243859052658, + 0.010023335926234722, + 0.0065871174447238445, + 0.015916908159852028, + 0.005511417053639889, + 0.007183062378317118, + 0.007364504039287567, + 0.005962308030575514, + 0.004725174978375435, + 0.00638398015871644, + 0.008447893895208836, + 0.007727993186563253, + 0.00788954645395279, + 0.0060548437759280205, + 0.004456948023289442, + 0.010463348589837551, + 0.0035737145226448774, + 0.014927612617611885, + 0.009295890107750893, + 0.024074066430330276, + 0.008505351841449738, + 0.005786951165646315, + 0.00546029070392251, + 0.010459154844284058, + 0.008053231053054333, + 0.010812130756676197, + 0.006752221845090389, + 0.018225910142064095, + 0.005233755800873041, + 0.0037224588450044394, + 0.0037377839908003807, + 0.004952848888933659, + 0.005089147947728634, + 0.005667613819241524, + 0.005010190419852734, + 0.0066704717464745045, + 0.0038486230187118053, + 0.003702681278809905, + 0.00340463500469923, + 0.004515505861490965, + 0.0058158026076853275, + 0.0037508548703044653, + 0.0026922542601823807, + 0.0033487174659967422, + 0.0023570419289171696, + 0.003026622114703059, + 0.003301643766462803, + 0.0030899373814463615, + 0.003748564748093486, + 0.015626346692442894, + 0.014011676423251629, + 0.010683423839509487, + 0.009767232462763786, + 0.014904731884598732, + 0.012974061071872711, + 0.013683288358151913, + 0.009817867539823055, + 0.026335379108786583, + 0.02693428471684456, + 0.01220400258898735, + 0.013479428365826607, + 0.030904332175850868, + 0.02115822210907936, + 0.016829879954457283, + 0.015376221388578415, + 0.02022230252623558, + 0.022457974031567574, + 0.0068657840602099895, + 0.009259078651666641, + 0.029136493802070618, + 0.01770017296075821, + 0.01398457307368517, + 0.00796935148537159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.006159844808280468 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428181231021881 + ], + "opposite_logprob_tokens": [ + -2.0178182125091553 + ], + "logprob_reward": -0.1428181231021881, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1428181231021881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28573697805404663, + 0.06831248104572296, + 0.11372996866703033, + 0.06402907520532608, + 0.08278661966323853, + 0.09644581377506256, + 0.018808308988809586, + 0.06515811383724213, + 0.09596089273691177, + 0.10903166979551315 + ], + "credit_map": [ + 0.04522806033492088, + 0.04023439437150955, + 0.07257849723100662, + 0.027272718027234077, + 0.017663970589637756, + 0.026593664661049843, + 0.03205881267786026, + 0.024106865748763084, + 0.009802713990211487, + 0.010240374132990837, + 0.014498145319521427, + 0.0069933864288032055, + 0.007832786068320274, + 0.005987809970974922, + 0.007247751113027334, + 0.005709516815841198, + 0.01506003737449646, + 0.013982487842440605, + 0.020865214988589287, + 0.01675187237560749, + 0.011292236857116222, + 0.008662251755595207, + 0.017833946272730827, + 0.009281931445002556, + 0.01064777560532093, + 0.00579233979806304, + 0.015809431672096252, + 0.0076875281520187855, + 0.00482191052287817, + 0.004898315295577049, + 0.007968828082084656, + 0.00640295073390007, + 0.013564648106694221, + 0.011809391900897026, + 0.018427738919854164, + 0.0094060730189085, + 0.006509993690997362, + 0.0063691893592476845, + 0.009417809545993805, + 0.007281781639903784, + 0.013484238646924496, + 0.011965159326791763, + 0.01857518032193184, + 0.008145902305841446, + 0.011590782552957535, + 0.007782392669469118, + 0.010001275688409805, + 0.014900903217494488, + 0.0030858858954161406, + 0.0022797149140387774, + 0.0024964939802885056, + 0.0016379698645323515, + 0.002889032941311598, + 0.0023947861045598984, + 0.0017539366381242871, + 0.002270490862429142, + 0.009220845997333527, + 0.007465345319360495, + 0.013804350048303604, + 0.007896854542195797, + 0.0058558639138937, + 0.006043901201337576, + 0.006520380266010761, + 0.008350579999387264, + 0.014931317418813705, + 0.010110567323863506, + 0.015698347240686417, + 0.008781730197370052, + 0.009087101556360722, + 0.01879698969423771, + 0.0088949054479599, + 0.009659943170845509, + 0.017602305859327316, + 0.016936976462602615, + 0.012828776612877846, + 0.006829926744103432, + 0.0227529127150774, + 0.014958841726183891, + 0.009916812181472778, + 0.007205121219158173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.1428181231021881 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494651198387146 + ], + "opposite_logprob_tokens": [ + -2.9299464225769043 + ], + "logprob_reward": -0.05494651198387146, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494651198387146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08302625268697739, + 0.07377489656209946, + 0.09312313050031662, + 0.0797104761004448, + 0.027014413848519325, + 0.015824414789676666, + 0.06495445221662521, + 0.11722392588853836, + 0.39674970507621765, + 0.04859837517142296 + ], + "credit_map": [ + 0.0126593466848135, + 0.01005066279321909, + 0.01076509989798069, + 0.010338319465517998, + 0.008356833830475807, + 0.009272580966353416, + 0.012894359417259693, + 0.008689054287970066, + 0.01137269102036953, + 0.005979978013783693, + 0.01554220449179411, + 0.006883260793983936, + 0.010419076308608055, + 0.009070757776498795, + 0.008678421378135681, + 0.005828504916280508, + 0.017976483330130577, + 0.005899427458643913, + 0.03576088696718216, + 0.007716478779911995, + 0.003890242427587509, + 0.004844832234084606, + 0.0047373464331030846, + 0.012297429144382477, + 0.01124113705009222, + 0.011668954975903034, + 0.015711955726146698, + 0.011522633023560047, + 0.006696629803627729, + 0.006444575730711222, + 0.009763306006789207, + 0.006661278661340475, + 0.004375752527266741, + 0.0031418560538440943, + 0.005251921247690916, + 0.0024578527081757784, + 0.003027071710675955, + 0.002764446195214987, + 0.0032921815291047096, + 0.0027033300139009953, + 0.0023160502314567566, + 0.001595618319697678, + 0.0027000473346561193, + 0.001979104708880186, + 0.0017204709583893418, + 0.0015074394177645445, + 0.0019270827760919929, + 0.002078601159155369, + 0.009636078029870987, + 0.009490746073424816, + 0.008512514643371105, + 0.005235154181718826, + 0.012185451574623585, + 0.006451355293393135, + 0.00875419843941927, + 0.00468895398080349, + 0.01643064059317112, + 0.01965896226465702, + 0.012474370189011097, + 0.016137197613716125, + 0.01934453472495079, + 0.010464629158377647, + 0.01431421097368002, + 0.008399383164942265, + 0.06844733655452728, + 0.06057968735694885, + 0.0337827131152153, + 0.031718239188194275, + 0.07413142919540405, + 0.05193544924259186, + 0.04140511527657509, + 0.03474970534443855, + 0.007441241759806871, + 0.006703243125230074, + 0.007042564917355776, + 0.0034167657140642405, + 0.010177205316722393, + 0.005802145693451166, + 0.004746810998767614, + 0.003268393687903881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.05494651198387146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008631895296275616 + ], + "opposite_logprob_tokens": [ + -4.758631706237793 + ], + "logprob_reward": -0.008631895296275616, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008631895296275616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13268202543258667, + 0.14320354163646698, + 0.03196055442094803, + 0.07844844460487366, + 0.08660022914409637, + 0.14072459936141968, + 0.08683313429355621, + 0.025559352710843086, + 0.09931695461273193, + 0.17467118799686432 + ], + "credit_map": [ + 0.021837886422872543, + 0.018916312605142593, + 0.03285134583711624, + 0.014323295094072819, + 0.009002100676298141, + 0.013709276914596558, + 0.01144355721771717, + 0.010598262771964073, + 0.030817167833447456, + 0.012041576206684113, + 0.04729212075471878, + 0.011173753067851067, + 0.008315314538776875, + 0.007975470274686813, + 0.009912952780723572, + 0.01567518338561058, + 0.004759557545185089, + 0.0026085982099175453, + 0.008952335454523563, + 0.0032021801453083754, + 0.0023826658725738525, + 0.002215593121945858, + 0.004545451141893864, + 0.003294173162430525, + 0.010854179039597511, + 0.00782499648630619, + 0.02078515850007534, + 0.009696940891444683, + 0.007798888720571995, + 0.007566571235656738, + 0.006035126280039549, + 0.007886582985520363, + 0.016025414690375328, + 0.005856901407241821, + 0.028805244714021683, + 0.008575638756155968, + 0.005210282746702433, + 0.005482967477291822, + 0.008243846707046032, + 0.008399924263358116, + 0.022787658497691154, + 0.02359236404299736, + 0.02504871040582657, + 0.012566071934998035, + 0.01961897499859333, + 0.012783289887011051, + 0.012594924308359623, + 0.011732601560652256, + 0.013280902989208698, + 0.011330021545290947, + 0.010690642520785332, + 0.007386145647615194, + 0.014740659855306149, + 0.009544613771140575, + 0.008441201411187649, + 0.011418953537940979, + 0.004132349044084549, + 0.0033882379066199064, + 0.004073161166161299, + 0.003041575662791729, + 0.0024929349310696125, + 0.0022830539382994175, + 0.0035852619912475348, + 0.0025627792347222567, + 0.015563206747174263, + 0.01640966534614563, + 0.008432857692241669, + 0.008946807123720646, + 0.017390115186572075, + 0.01227575819939375, + 0.010619042441248894, + 0.009679500944912434, + 0.036060892045497894, + 0.011615738272666931, + 0.07077213376760483, + 0.013234823010861874, + 0.00594939524307847, + 0.008669950067996979, + 0.0102838184684515, + 0.01808444783091545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.008631895296275616 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290071904659271 + ], + "opposite_logprob_tokens": [ + -2.237900733947754 + ], + "logprob_reward": -0.11290071904659271, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290071904659271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11802756786346436, + 0.05981893837451935, + 0.09889984130859375, + 0.20889237523078918, + 0.05050754174590111, + 0.03670649975538254, + 0.06614486873149872, + 0.06562177836894989, + 0.2483173906803131, + 0.04706313833594322 + ], + "credit_map": [ + 0.018578246235847473, + 0.017574161291122437, + 0.02100059948861599, + 0.011993338353931904, + 0.009129284881055355, + 0.011998021975159645, + 0.015150899067521095, + 0.012603020295500755, + 0.009849339723587036, + 0.007761466316878796, + 0.011382294818758965, + 0.006552282255142927, + 0.004591514356434345, + 0.00722871208563447, + 0.006896392907947302, + 0.005556933116167784, + 0.01879674382507801, + 0.0055887396447360516, + 0.03685545548796654, + 0.0069703273475170135, + 0.004261435475200415, + 0.004682134836912155, + 0.009580419398844242, + 0.012164582498371601, + 0.030458157882094383, + 0.01862695999443531, + 0.04482458159327507, + 0.024313541129231453, + 0.020683206617832184, + 0.021016472950577736, + 0.028046797960996628, + 0.020922645926475525, + 0.009266762994229794, + 0.0044545335695147514, + 0.015798045322299004, + 0.004907018505036831, + 0.003169149160385132, + 0.0033918360713869333, + 0.004380693659186363, + 0.00513949990272522, + 0.005497327074408531, + 0.004896883387118578, + 0.007136212661862373, + 0.0037668836303055286, + 0.0028809390496462584, + 0.0033621678594499826, + 0.004247127566486597, + 0.004918959923088551, + 0.010640934109687805, + 0.0071128481067717075, + 0.014622257091104984, + 0.0056969961151480675, + 0.009729092009365559, + 0.006447476334869862, + 0.007085631601512432, + 0.0048096319660544395, + 0.009306137450039387, + 0.007945115678012371, + 0.008774353191256523, + 0.008401526138186455, + 0.008453130722045898, + 0.0082058971747756, + 0.006295146886259317, + 0.00824047438800335, + 0.04233164340257645, + 0.03752421960234642, + 0.021146954968571663, + 0.020581072196364403, + 0.045462943613529205, + 0.03274551033973694, + 0.026846421882510185, + 0.021678637713193893, + 0.00748995179310441, + 0.006539470050483942, + 0.01046073529869318, + 0.0031828812789171934, + 0.007286262698471546, + 0.004588738549500704, + 0.0037848472129553556, + 0.0037302514538168907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.11290071904659271 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a shrub rendered in a captivating sketch-style ink and watercolor, its forms swaying gently in the breeze. The shrub is covered in fluffy, pom-pom-like flowers, each a vibrant canary-yellow globe of delicate, silken petals. Delicate, expressive black ink lines loosely define the shapes of the graceful, vine-like branches and sparse green leaves, adding a graphic quality to the organic forms. The entire plant sways rhythmically back and forth, and this subtle movement causes each flower cluster to subtly tremble. The artistry is highlighted by the visible soft colors, where the translucent yellow watercolor meets the green of the leaves, creating a beautiful wet-on-wet effect. The background, an abstract palette of pale green and earthy brown, focuses entirely on the shrub's gentle movement. The scene is illuminated by a soft, diffuse light, enhancing the brightness of the colors and conveying an atmosphere of tranquility, peace, and artistic refinement.", + "seed": 583, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_10/index_3990_seed_583_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05737860640510917, + "vlm_logprob": -0.05737860640510917, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.05737860640510917, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00490592559799552 + ], + "opposite_logprob_tokens": [ + -5.754905700683594 + ], + "logprob_reward": -0.00490592559799552, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00490592559799552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10076998919248581, + 0.06665629893541336, + 0.023453205823898315, + 0.013467342592775822, + 0.018184300512075424, + 0.1961807757616043, + 0.18221688270568848, + 0.11937057226896286, + 0.24065975844860077, + 0.03904086351394653 + ], + "credit_map": [ + 0.019386861473321915, + 0.010051527991890907, + 0.015722200274467468, + 0.01070417184382677, + 0.008034306578338146, + 0.015595629811286926, + 0.007946697995066643, + 0.013328589498996735, + 0.00825134664773941, + 0.009024668484926224, + 0.012908133678138256, + 0.005921748001128435, + 0.004926108755171299, + 0.012309475801885128, + 0.004865023773163557, + 0.008449791930615902, + 0.0033230583649128675, + 0.002202036324888468, + 0.004919004160910845, + 0.0022139744833111763, + 0.0018819317920133471, + 0.003896906040608883, + 0.0021531349048018456, + 0.0028631589375436306, + 0.002672244096174836, + 0.0012726604472845793, + 0.0020835348404943943, + 0.001179771381430328, + 0.001411502598784864, + 0.0023747493978589773, + 0.0009976375149562955, + 0.001475243247114122, + 0.0031659347005188465, + 0.0014488021843135357, + 0.003604512196034193, + 0.0014656574930995703, + 0.0018000849522650242, + 0.002640429185703397, + 0.0017184545285999775, + 0.002340425970032811, + 0.055614762008190155, + 0.017249835655093193, + 0.026373621076345444, + 0.009454146027565002, + 0.008820569142699242, + 0.020307421684265137, + 0.010138596408069134, + 0.04822184145450592, + 0.03618370369076729, + 0.01115669310092926, + 0.036723535507917404, + 0.007450232747942209, + 0.014346802607178688, + 0.017735015600919724, + 0.007681694813072681, + 0.05093919858336449, + 0.015076179057359695, + 0.007927916012704372, + 0.034283705055713654, + 0.010273183695971966, + 0.010640479624271393, + 0.010813185945153236, + 0.0077532012946903706, + 0.02260272018611431, + 0.06557849794626236, + 0.014185380190610886, + 0.08933518081903458, + 0.009091520681977272, + 0.010859171859920025, + 0.015979409217834473, + 0.008211626671254635, + 0.027418959885835648, + 0.005510475020855665, + 0.0029082857072353363, + 0.006902134511619806, + 0.002677140524610877, + 0.0025249910540878773, + 0.004180933814495802, + 0.0027096925769001245, + 0.011627210304141045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.00490592559799552 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029809197410941124 + ], + "opposite_logprob_tokens": [ + -3.5298092365264893 + ], + "logprob_reward": -0.029809197410941124, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029809197410941124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2763473093509674, + 0.10107875615358353, + 0.032271385192871094, + 0.02302018739283085, + 0.025705300271511078, + 0.2671099603176117, + 0.08673283457756042, + 0.0856316015124321, + 0.054988760501146317, + 0.04711383581161499 + ], + "credit_map": [ + 0.06382079422473907, + 0.028932111337780952, + 0.03742395341396332, + 0.0271877720952034, + 0.019051063805818558, + 0.04728822410106659, + 0.019348187372088432, + 0.03329521417617798, + 0.015952251851558685, + 0.0063928826712071896, + 0.029426898807287216, + 0.00774753512814641, + 0.005296406801789999, + 0.012388883158564568, + 0.007150489371269941, + 0.01672341674566269, + 0.004101291764527559, + 0.0030671649146825075, + 0.005909586790949106, + 0.002407986205071211, + 0.0027841900009661913, + 0.006848016753792763, + 0.0022134464234113693, + 0.004939701873809099, + 0.004294482059776783, + 0.001973748439922929, + 0.0049714744091033936, + 0.0017388267442584038, + 0.0015003331936895847, + 0.0033073192462325096, + 0.0019457598682492971, + 0.00328824482858181, + 0.004503083415329456, + 0.0015254764584824443, + 0.008027642965316772, + 0.002388057764619589, + 0.001607003971002996, + 0.00272359955124557, + 0.001417122082784772, + 0.0035133177880197763, + 0.0856103003025055, + 0.015896601602435112, + 0.08030574023723602, + 0.012939861044287682, + 0.008086493238806725, + 0.023849189281463623, + 0.009207134135067463, + 0.031214630231261253, + 0.0278010331094265, + 0.0054436782374978065, + 0.008605735376477242, + 0.0033215899020433426, + 0.004495250526815653, + 0.009866726584732533, + 0.003800271777436137, + 0.02339855767786503, + 0.011299156583845615, + 0.00939431693404913, + 0.01842460036277771, + 0.010876930318772793, + 0.007855468429625034, + 0.01024559698998928, + 0.006689595989882946, + 0.010845944285392761, + 0.009902046993374825, + 0.006628711707890034, + 0.016321543604135513, + 0.0036940902937203646, + 0.005812642630189657, + 0.0039622243493795395, + 0.0026169349439442158, + 0.006050571799278259, + 0.007072253152728081, + 0.003660226473584771, + 0.008051454089581966, + 0.003039828734472394, + 0.0041680485010147095, + 0.00539217796176672, + 0.002679648343473673, + 0.013050196692347527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.029809197410941124 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304364696145058 + ], + "opposite_logprob_tokens": [ + -3.168043613433838 + ], + "logprob_reward": -0.04304364696145058, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304364696145058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09079559892416, + 0.05321256071329117, + 0.0325222909450531, + 0.02326146885752678, + 0.018775317817926407, + 0.16018564999103546, + 0.1746809333562851, + 0.24149180948734283, + 0.11164011806249619, + 0.09343424439430237 + ], + "credit_map": [ + 0.018818095326423645, + 0.008218500763177872, + 0.013823903165757656, + 0.009044425562024117, + 0.00637791445478797, + 0.016258059069514275, + 0.006235331296920776, + 0.012019370682537556, + 0.008263264782726765, + 0.00538378581404686, + 0.010804119519889355, + 0.0056528812274336815, + 0.0033739216160029173, + 0.007787785027176142, + 0.0037991118151694536, + 0.008147697895765305, + 0.0042889658361673355, + 0.0034578789491206408, + 0.005312094930559397, + 0.002474965760484338, + 0.002833191305398941, + 0.0066378312185406685, + 0.0027148195076733828, + 0.004802545066922903, + 0.004650298971682787, + 0.0019711623899638653, + 0.004717632196843624, + 0.001851221197284758, + 0.0018668350530788302, + 0.003293773625046015, + 0.0015907378401607275, + 0.0033198073506355286, + 0.0024012394715100527, + 0.0024424006696790457, + 0.0025819323491305113, + 0.0018796938238665462, + 0.0021586597431451082, + 0.0029046148993074894, + 0.0018231977010145783, + 0.002583578694611788, + 0.030460625886917114, + 0.008510091342031956, + 0.04009789228439331, + 0.004008341580629349, + 0.005108950659632683, + 0.01121730450540781, + 0.008951990865170956, + 0.051830437034368515, + 0.037517283111810684, + 0.01877090148627758, + 0.019764574244618416, + 0.007181959226727486, + 0.018361283466219902, + 0.02294062450528145, + 0.008738517761230469, + 0.04140578210353851, + 0.029624758288264275, + 0.020489517599344254, + 0.07289756089448929, + 0.02100484073162079, + 0.020200803875923157, + 0.02271798625588417, + 0.015429539605975151, + 0.03912680223584175, + 0.02793949656188488, + 0.007959013804793358, + 0.0352790541946888, + 0.005433008074760437, + 0.009963544085621834, + 0.010137715376913548, + 0.004491266794502735, + 0.010437028482556343, + 0.011997463181614876, + 0.0081567307934165, + 0.016092246398329735, + 0.0038289001677185297, + 0.00904155895113945, + 0.011821171268820763, + 0.005319191142916679, + 0.027176981791853905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.04304364696145058 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007619958370923996 + ], + "opposite_logprob_tokens": [ + -4.882619857788086 + ], + "logprob_reward": -0.007619958370923996, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007619958370923996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05094226449728012, + 0.08008618652820587, + 0.1580205261707306, + 0.02586028166115284, + 0.0282965786755085, + 0.1015252098441124, + 0.06423434615135193, + 0.042726125568151474, + 0.4269014298915863, + 0.021407118067145348 + ], + "credit_map": [ + 0.009696558117866516, + 0.004995435010641813, + 0.0074529219418764114, + 0.004167571663856506, + 0.0036904767621308565, + 0.011483876965939999, + 0.003939102403819561, + 0.005516314413398504, + 0.015415171161293983, + 0.007287151180207729, + 0.01856106147170067, + 0.006654783617705107, + 0.0040822201408445835, + 0.010166669264435768, + 0.006795903667807579, + 0.011123218573629856, + 0.024319889023900032, + 0.011249171569943428, + 0.04890737682580948, + 0.009416882880032063, + 0.008207209408283234, + 0.01987530291080475, + 0.015025471337139606, + 0.021019212901592255, + 0.004165321122854948, + 0.0023593222722411156, + 0.005406234413385391, + 0.002262202324345708, + 0.002364083658903837, + 0.004386401269584894, + 0.0021569691598415375, + 0.0027597472071647644, + 0.004738341085612774, + 0.002103483071550727, + 0.008173303678631783, + 0.002650743816047907, + 0.0017710308311507106, + 0.0028939039912074804, + 0.002157165901735425, + 0.0038086031563580036, + 0.02781529538333416, + 0.008523311465978622, + 0.028842095285654068, + 0.005179704166948795, + 0.0045793317258358, + 0.01142860110849142, + 0.0038639630656689405, + 0.011292887851595879, + 0.017082950100302696, + 0.0039847432635724545, + 0.009384899400174618, + 0.0021850108169019222, + 0.004493136424571276, + 0.005993402563035488, + 0.0024230368435382843, + 0.018687160685658455, + 0.005282253492623568, + 0.003005210543051362, + 0.012635797262191772, + 0.003011607099324465, + 0.0033714037854224443, + 0.003924134187400341, + 0.0025225416757166386, + 0.00897317472845316, + 0.12068028748035431, + 0.019296795129776, + 0.166193887591362, + 0.014039888046681881, + 0.012116730213165283, + 0.027924399822950363, + 0.01197839342057705, + 0.05467100813984871, + 0.0031971363350749016, + 0.0018943178001791239, + 0.004285308998078108, + 0.0012681769439950585, + 0.0015725569101050496, + 0.0025251691695302725, + 0.0011814945610240102, + 0.005482954438775778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.007619958370923996 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20151430368423462 + ], + "opposite_logprob_tokens": [ + -1.7015142440795898 + ], + "logprob_reward": -0.20151430368423462, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20151430368423462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10361883789300919, + 0.030750764533877373, + 0.04410342872142792, + 0.08702737838029861, + 0.018894122913479805, + 0.2284419983625412, + 0.17650724947452545, + 0.07095466554164886, + 0.18778479099273682, + 0.051916785538196564 + ], + "credit_map": [ + 0.020068908110260963, + 0.010055390186607838, + 0.018885448575019836, + 0.009904753416776657, + 0.006842389702796936, + 0.016303693875670433, + 0.007644444238394499, + 0.013913808390498161, + 0.004950706381350756, + 0.005085458047688007, + 0.00433344766497612, + 0.0027306939009577036, + 0.0024760367814451456, + 0.004965952597558498, + 0.002261154819279909, + 0.003947315271943808, + 0.006909195799380541, + 0.0028171231970191, + 0.012174401432275772, + 0.003196656471118331, + 0.002773528452962637, + 0.005936376750469208, + 0.004754913505166769, + 0.005541236139833927, + 0.01586049608886242, + 0.007596930023282766, + 0.018938817083835602, + 0.007694227620959282, + 0.007388230413198471, + 0.013683098368346691, + 0.006411142647266388, + 0.009454438462853432, + 0.0029847039841115475, + 0.0023094310890883207, + 0.0043824720196425915, + 0.0017970012268051505, + 0.0015659226337447762, + 0.0022678386885672808, + 0.0016175692435353994, + 0.0019691847264766693, + 0.04689725115895271, + 0.01167468260973692, + 0.0559084378182888, + 0.005542561877518892, + 0.005586644168943167, + 0.01655779965221882, + 0.012172127142548561, + 0.0741024911403656, + 0.03361763060092926, + 0.016841549426317215, + 0.029958218336105347, + 0.0074511063285171986, + 0.014267963357269764, + 0.01792910508811474, + 0.00886851828545332, + 0.04757315665483475, + 0.008892732672393322, + 0.009071782231330872, + 0.011431924998760223, + 0.008441012352705002, + 0.008060920052230358, + 0.010935434140264988, + 0.006341223139315844, + 0.007779632695019245, + 0.04750677943229675, + 0.01839277893304825, + 0.06582020968198776, + 0.0072348411194980145, + 0.01600460521876812, + 0.012805378064513206, + 0.0063497102819383144, + 0.013670481741428375, + 0.007300094701349735, + 0.005868620704859495, + 0.008439596742391586, + 0.0032147434540092945, + 0.003965345211327076, + 0.0062562525272369385, + 0.004016423597931862, + 0.01285570953041315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.20151430368423462 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a shrub rendered in a captivating sketch-style ink and watercolor, its forms swaying gently in the breeze. The shrub is covered in fluffy, pom-pom-like flowers, each a vibrant canary-yellow globe of delicate, silken petals. Delicate, expressive black ink lines loosely define the shapes of the graceful, vine-like branches and sparse green leaves, adding a graphic quality to the organic forms. The entire plant sways rhythmically back and forth, and this subtle movement causes each flower cluster to subtly tremble. The artistry is highlighted by the visible soft colors, where the translucent yellow watercolor meets the green of the leaves, creating a beautiful wet-on-wet effect. The background, an abstract palette of pale green and earthy brown, focuses entirely on the shrub's gentle movement. The scene is illuminated by a soft, diffuse light, enhancing the brightness of the colors and conveying an atmosphere of tranquility, peace, and artistic refinement.", + "seed": 584, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_10/index_3990_seed_584_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.058950027683749795, + "vlm_logprob": -0.058950027683749795, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.058950027683749795, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006589588243514299 + ], + "opposite_logprob_tokens": [ + -5.381589412689209 + ], + "logprob_reward": -0.006589588243514299, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.006589588243514299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13276566565036774, + 0.16071034967899323, + 0.06263797730207443, + 0.1317560225725174, + 0.06407655030488968, + 0.09603091329336166, + 0.13691096007823944, + 0.04927103966474533, + 0.07864627242088318, + 0.08719422668218613 + ], + "credit_map": [ + 0.01149996742606163, + 0.017550436779856682, + 0.014052904210984707, + 0.019833339378237724, + 0.02257964201271534, + 0.015969203785061836, + 0.012155808508396149, + 0.019124364480376244, + 0.015606852248311043, + 0.012793210335075855, + 0.013627367094159126, + 0.01180163025856018, + 0.011603971011936665, + 0.03367961198091507, + 0.012050818651914597, + 0.04954688623547554, + 0.005577963776886463, + 0.005300893913954496, + 0.0073319352231919765, + 0.006690075621008873, + 0.006329473108053207, + 0.008403605781495571, + 0.007471996825188398, + 0.015532033517956734, + 0.01546756736934185, + 0.014987734146416187, + 0.014019059017300606, + 0.013663174584507942, + 0.020110897719860077, + 0.014789632521569729, + 0.013613696210086346, + 0.025104260072112083, + 0.0052623264491558075, + 0.005690432619303465, + 0.006602398119866848, + 0.010627892799675465, + 0.0074769072234630585, + 0.007088937796652317, + 0.010996121913194656, + 0.010331531055271626, + 0.00915363896638155, + 0.006648269947618246, + 0.00843380019068718, + 0.008068445138633251, + 0.009370902553200722, + 0.02018243446946144, + 0.009643948636949062, + 0.02452947571873665, + 0.016032617539167404, + 0.016319043934345245, + 0.015886003151535988, + 0.01135533768683672, + 0.022213585674762726, + 0.016984140500426292, + 0.014411942102015018, + 0.023708296939730644, + 0.005410342942923307, + 0.005513426847755909, + 0.005498548969626427, + 0.006157621741294861, + 0.005610414315015078, + 0.00593723775818944, + 0.005808892194181681, + 0.009334554895758629, + 0.008195377886295319, + 0.0072386800311505795, + 0.0077440375462174416, + 0.0058972835540771484, + 0.009144037961959839, + 0.010949317365884781, + 0.009057722054421902, + 0.020419815555214882, + 0.008347616530954838, + 0.007583353668451309, + 0.008766941726207733, + 0.004348320420831442, + 0.0069050234742462635, + 0.013468672521412373, + 0.004740532953292131, + 0.03303376957774162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.006589588243514299 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043078359216451645 + ], + "opposite_logprob_tokens": [ + -3.1680784225463867 + ], + "logprob_reward": -0.043078359216451645, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043078359216451645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3040948510169983, + 0.21916696429252625, + 0.02924368903040886, + 0.06983595341444016, + 0.04371992498636246, + 0.09058278053998947, + 0.0867624282836914, + 0.02678794041275978, + 0.011350015178322792, + 0.11845538765192032 + ], + "credit_map": [ + 0.02593117021024227, + 0.026709526777267456, + 0.02512747421860695, + 0.024985823780298233, + 0.051024120301008224, + 0.05210460349917412, + 0.03615279868245125, + 0.06205934286117554, + 0.022140922024846077, + 0.015304028987884521, + 0.02081446535885334, + 0.013592648319900036, + 0.01604892499744892, + 0.03959105908870697, + 0.012649686075747013, + 0.07902523875236511, + 0.0033576886635273695, + 0.002778401831164956, + 0.0037917913869023323, + 0.004231918603181839, + 0.0030035271774977446, + 0.003204087959602475, + 0.0037264954298734665, + 0.005149781238287687, + 0.0070786140859127045, + 0.012409619055688381, + 0.005909509025514126, + 0.007659691385924816, + 0.011267060413956642, + 0.007341308519244194, + 0.008169678971171379, + 0.01000047568231821, + 0.0038025658577680588, + 0.0032205714378505945, + 0.004441430792212486, + 0.006791478488594294, + 0.003976551350206137, + 0.005541233811527491, + 0.007304457947611809, + 0.008641639724373817, + 0.008134827949106693, + 0.006418234668672085, + 0.008534029126167297, + 0.00530253117904067, + 0.007055750582367182, + 0.01871371455490589, + 0.006402555853128433, + 0.030021140351891518, + 0.009656545706093311, + 0.01319073885679245, + 0.008482027798891068, + 0.006883989553898573, + 0.018201135098934174, + 0.008854988031089306, + 0.01006590761244297, + 0.011427103541791439, + 0.0028503381181508303, + 0.004172069486230612, + 0.0030423072166740894, + 0.00368058611638844, + 0.0032199120614677668, + 0.002913252217695117, + 0.003554021241143346, + 0.00335545651614666, + 0.0011021895334124565, + 0.0007233492797240615, + 0.0011932130437344313, + 0.0009442248265258968, + 0.0007478281622752547, + 0.0023809524718672037, + 0.0011355996830388904, + 0.0031226580031216145, + 0.01193385198712349, + 0.009322321973741055, + 0.011777788400650024, + 0.00592401809990406, + 0.00940245296806097, + 0.01943790167570114, + 0.006286362186074257, + 0.044370707124471664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.043078359216451645 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987202107906342 + ], + "opposite_logprob_tokens": [ + -1.8048720359802246 + ], + "logprob_reward": -0.17987202107906342, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987202107906342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.464830219745636, + 0.08840380609035492, + 0.043205637484788895, + 0.1323462873697281, + 0.05059855431318283, + 0.05010823905467987, + 0.06635735929012299, + 0.04855046793818474, + 0.027034088969230652, + 0.028565283864736557 + ], + "credit_map": [ + 0.044078100472688675, + 0.05365612730383873, + 0.04747374728322029, + 0.04487285017967224, + 0.08930177986621857, + 0.05350986495614052, + 0.05991881340742111, + 0.07201898097991943, + 0.00933089293539524, + 0.008169877342879772, + 0.011011290363967419, + 0.007864928804337978, + 0.008397532626986504, + 0.020456233993172646, + 0.006820077076554298, + 0.016352973878383636, + 0.004632709082216024, + 0.0036984772887080908, + 0.00548209622502327, + 0.004206660203635693, + 0.003982976078987122, + 0.007531705312430859, + 0.005162287037819624, + 0.008508728817105293, + 0.014072100631892681, + 0.013886631466448307, + 0.011973694898188114, + 0.015276472084224224, + 0.011691492982208729, + 0.025292789563536644, + 0.01372129749506712, + 0.02643180638551712, + 0.0045223962515592575, + 0.005383273120969534, + 0.005887968000024557, + 0.007104482967406511, + 0.006057173013687134, + 0.006114103831350803, + 0.007261342369019985, + 0.00826781615614891, + 0.004870058968663216, + 0.004081204999238253, + 0.004194267094135284, + 0.004200325347483158, + 0.007092164363712072, + 0.008911143988370895, + 0.006560646928846836, + 0.010198433883488178, + 0.007149302866309881, + 0.00991287175565958, + 0.0051888409070670605, + 0.00425033550709486, + 0.012923176400363445, + 0.008421618491411209, + 0.005760460626333952, + 0.012750763446092606, + 0.004717031493782997, + 0.004257441032677889, + 0.005646051838994026, + 0.005855043884366751, + 0.004533601924777031, + 0.00919810775667429, + 0.005998855922371149, + 0.008344337344169617, + 0.002447630511596799, + 0.0030053756199777126, + 0.0024819255340844393, + 0.0029492501635104418, + 0.0022781111765652895, + 0.0041877031326293945, + 0.0029287408106029034, + 0.006755355279892683, + 0.00290047493763268, + 0.002092809183523059, + 0.002587220398709178, + 0.001802597544156015, + 0.0024826026055961847, + 0.005090480204671621, + 0.0017273621633648872, + 0.009881739504635334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.17987202107906342 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003188884584233165 + ], + "opposite_logprob_tokens": [ + -5.7531890869140625 + ], + "logprob_reward": -0.003188884584233165, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003188884584233165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1812949776649475, + 0.35607683658599854, + 0.08200287073850632, + 0.03824077546596527, + 0.0711597129702568, + 0.08599555492401123, + 0.026690935716032982, + 0.017057349905371666, + 0.021617718040943146, + 0.11986327171325684 + ], + "credit_map": [ + 0.02106497809290886, + 0.0220218263566494, + 0.01748998463153839, + 0.01263352856040001, + 0.034157395362854004, + 0.024590644985437393, + 0.021558597683906555, + 0.02777804620563984, + 0.04187723621726036, + 0.024273810908198357, + 0.02967004105448723, + 0.02033575437963009, + 0.03152856603264809, + 0.0597710981965065, + 0.029176466166973114, + 0.119443878531456, + 0.008038334548473358, + 0.009824051521718502, + 0.008065782487392426, + 0.0064037335105240345, + 0.008824504911899567, + 0.01095994096249342, + 0.008749697357416153, + 0.021136825904250145, + 0.0032819248735904694, + 0.005807151086628437, + 0.003203604370355606, + 0.004346259403973818, + 0.006883225869387388, + 0.004613147117197514, + 0.004748015198856592, + 0.005357448477298021, + 0.008156931959092617, + 0.006561923306435347, + 0.00746888667345047, + 0.005119457375258207, + 0.0069137499667704105, + 0.009324430488049984, + 0.008200843818485737, + 0.019413497298955917, + 0.011884177103638649, + 0.005805684719234705, + 0.008316503837704659, + 0.005620115902274847, + 0.008768345229327679, + 0.016057470813393593, + 0.007707482669502497, + 0.021835777908563614, + 0.002933553420007229, + 0.0031138353515416384, + 0.003234384348616004, + 0.0024269288405776024, + 0.004198339767754078, + 0.0029354910366237164, + 0.0031191417947411537, + 0.004729262553155422, + 0.0026991437189280987, + 0.0015492405509576201, + 0.0020547318272292614, + 0.001991077559068799, + 0.001582365483045578, + 0.002274380996823311, + 0.0014634798280894756, + 0.0034429309889674187, + 0.0021586802322417498, + 0.0018696794286370277, + 0.002048012800514698, + 0.001723825465887785, + 0.002853651065379381, + 0.0038451338186860085, + 0.003267639083787799, + 0.0038510989397764206, + 0.012628364376723766, + 0.010317344218492508, + 0.01181500032544136, + 0.007326764985918999, + 0.007048561703413725, + 0.016136372461915016, + 0.011674320325255394, + 0.04291655495762825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.003188884584233165 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shrub covered in fluffy, pom-pom-like canary-yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a sketch-style ink and watercolor with visible black ink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shrub sway gently back and forth while the flowers tremble subtly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an abstract palette of pale green and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202128529548645 + ], + "opposite_logprob_tokens": [ + -2.812021255493164 + ], + "logprob_reward": -0.06202128529548645, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202128529548645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16083024442195892, + 0.1417497992515564, + 0.07539312541484833, + 0.1544654667377472, + 0.04108031839132309, + 0.09861935675144196, + 0.06863251328468323, + 0.15768907964229584, + 0.03084873966872692, + 0.07069139927625656 + ], + "credit_map": [ + 0.01991059072315693, + 0.016069386154413223, + 0.01526085939258337, + 0.013275520876049995, + 0.026902971789240837, + 0.029382506385445595, + 0.01491709053516388, + 0.02511131763458252, + 0.01879454217851162, + 0.011894715018570423, + 0.01946367137134075, + 0.01112134288996458, + 0.016537124291062355, + 0.024430913850665092, + 0.014200801961123943, + 0.025306686758995056, + 0.007765908259898424, + 0.008986281231045723, + 0.007727900519967079, + 0.007205454166978598, + 0.006708664353936911, + 0.013530547730624676, + 0.006231983192265034, + 0.017236383631825447, + 0.015965504571795464, + 0.021176936104893684, + 0.01494750939309597, + 0.017733603715896606, + 0.022229475900530815, + 0.02216300182044506, + 0.017989763990044594, + 0.0222596637904644, + 0.0032064723782241344, + 0.005742509383708239, + 0.003460572799667716, + 0.0039366441778838634, + 0.007169446907937527, + 0.005427906289696693, + 0.004496967885643244, + 0.007639797870069742, + 0.009399757720530033, + 0.006501859985291958, + 0.009067405015230179, + 0.008952103555202484, + 0.01249756384640932, + 0.022950617596507072, + 0.00937542412430048, + 0.019874626770615578, + 0.009194047190248966, + 0.009054584428668022, + 0.010487345978617668, + 0.0059951539151370525, + 0.007960581220686436, + 0.007064279168844223, + 0.008479149080812931, + 0.010397370904684067, + 0.020768718793988228, + 0.018840529024600983, + 0.01854524202644825, + 0.01660328172147274, + 0.019729146733880043, + 0.021409617736935616, + 0.01370147243142128, + 0.028091080486774445, + 0.0033323352690786123, + 0.003278696443885565, + 0.0030141607858240604, + 0.0030161805916577578, + 0.0032081531826406717, + 0.00447995075955987, + 0.004161222837865353, + 0.006358037702739239, + 0.006613757461309433, + 0.005817442201077938, + 0.009003376588225365, + 0.00479555269703269, + 0.006806116551160812, + 0.010914805345237255, + 0.006207481492310762, + 0.020532866939902306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.06202128529548645 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_11/index_tensor([6074, 6074])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_11/index_tensor([6074, 6074])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b57f3c38bb70adace68b785c754aec689409bbfb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_11/index_tensor([6074, 6074])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the surreal and emotive style of Makoto Shinkai, a low-angle tracking shot follows Shun Kazama, a determined young man, as he strides unwaveringly forward. He wears a dark, high-collared school uniform, the fabric appearing damp and subtly shimmering under intricate lighting. The entire scene is bathed in the iconic golden hour light, casting long shadows and infusing the atmosphere with a warm, nostalgic glow. As the camera smoothly slides behind and to the side of him, dramatic anamorphic lens flares streak horizontally across the screen, created by the headlights of passing cars and glowing streetlights. The backdrop is a rain-slicked city street at dusk, every detail rendered with photorealistic precision; the wet asphalt perfectly reflects the vibrant neon signs of storefronts and the deep blue of the twilight sky, creating a world of idyllic beauty. Shun walks with a steady and purposeful pace, his foot occasionally splashing into a shallow puddle, sending a cloud of glowing droplets into the air. His expression is intensely focused, his gaze fixed on an unseen point ahead, embodying a powerful sense of forward momentum and a quiet, emotional gravity.", + "seed": 3268, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_11/index_6074_seed_3268_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06338755413889885, + "vlm_logprob": -0.06338755413889885, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.06338755413889885, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07609213143587112 + ], + "opposite_logprob_tokens": [ + -2.82609224319458 + ], + "logprob_reward": -0.07609213143587112, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07609213143587112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08255670219659805, + 0.2804613411426544, + 0.13134580850601196, + 0.12003926187753677, + 0.06497175246477127, + 0.062293391674757004, + 0.034520331770181656, + 0.0996885746717453, + 0.06935176253318787, + 0.05477109178900719 + ], + "credit_map": [ + 0.007380939554423094, + 0.007162300404161215, + 0.010466180741786957, + 0.01972898840904236, + 0.008785337209701538, + 0.009048903360962868, + 0.012926491908729076, + 0.007057553622871637, + 0.0729801133275032, + 0.02330404706299305, + 0.013443971984088421, + 0.018628833815455437, + 0.019260453060269356, + 0.01346403919160366, + 0.10270360112190247, + 0.016676276922225952, + 0.010651439428329468, + 0.013018542900681496, + 0.011918099597096443, + 0.024669377133250237, + 0.02731562778353691, + 0.007529763970524073, + 0.02353774383664131, + 0.012705205008387566, + 0.01663399487733841, + 0.014487219974398613, + 0.018998976796865463, + 0.009875169955193996, + 0.01905485801398754, + 0.009975148364901543, + 0.012486360967159271, + 0.018527524545788765, + 0.004034892655909061, + 0.004756734240800142, + 0.005141680594533682, + 0.0057360464707016945, + 0.01477990485727787, + 0.004695161711424589, + 0.016503479331731796, + 0.009323840029537678, + 0.006631408352404833, + 0.007649099919945002, + 0.00473928963765502, + 0.0061271172016859055, + 0.012600504793226719, + 0.008670169860124588, + 0.00975121557712555, + 0.0061245812103152275, + 0.004693073686212301, + 0.003782127285376191, + 0.001982141053304076, + 0.004298869054764509, + 0.005545090883970261, + 0.002403814811259508, + 0.006836902815848589, + 0.004978312645107508, + 0.013336879201233387, + 0.014292282983660698, + 0.009018272161483765, + 0.01247374527156353, + 0.010669529438018799, + 0.01834770478308201, + 0.009211573749780655, + 0.012338568456470966, + 0.009345746599137783, + 0.01083952747285366, + 0.008834166452288628, + 0.0164655651897192, + 0.005522524937987328, + 0.008456474170088768, + 0.005518408492207527, + 0.004369345493614674, + 0.004857126623392105, + 0.006927600596100092, + 0.009534027427434921, + 0.010940817184746265, + 0.0050040604546666145, + 0.007774900645017624, + 0.005052174441516399, + 0.004680377431213856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.07609213143587112 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17995059490203857 + ], + "opposite_logprob_tokens": [ + -1.8049505949020386 + ], + "logprob_reward": -0.17995059490203857, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17995059490203857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07449585199356079, + 0.3753041625022888, + 0.07761268317699432, + 0.13459177315235138, + 0.02303500846028328, + 0.022853896021842957, + 0.037601303309202194, + 0.09159842133522034, + 0.07065717875957489, + 0.09224975109100342 + ], + "credit_map": [ + 0.004911653231829405, + 0.005902546923607588, + 0.006590403616428375, + 0.021665746346116066, + 0.0072709787636995316, + 0.005019376054406166, + 0.015635890886187553, + 0.007499247323721647, + 0.0753077045083046, + 0.029065970331430435, + 0.021917320787906647, + 0.03111182153224945, + 0.03417074680328369, + 0.021499251946806908, + 0.13448984920978546, + 0.027741456404328346, + 0.00729347812011838, + 0.010953503660857677, + 0.007673910818994045, + 0.010987164452672005, + 0.0128159299492836, + 0.006954524666070938, + 0.01293760072439909, + 0.007996559143066406, + 0.023559795692563057, + 0.015418183989822865, + 0.023456981405615807, + 0.012896383181214333, + 0.01545551884919405, + 0.010599249973893166, + 0.016451135277748108, + 0.016754519194364548, + 0.002058523939922452, + 0.001818414544686675, + 0.002804046729579568, + 0.002506714081391692, + 0.003591137006878853, + 0.002521384973078966, + 0.004223075229674578, + 0.0035117107909172773, + 0.0025116754695773125, + 0.002366651315242052, + 0.0021182408090680838, + 0.0035955458879470825, + 0.00416880939155817, + 0.0030065651517361403, + 0.0031893078703433275, + 0.0018970969831570983, + 0.003588635940104723, + 0.008276980370283127, + 0.002384015591815114, + 0.004712289664894342, + 0.0061372146010398865, + 0.002458377741277218, + 0.005597861483693123, + 0.004445922560989857, + 0.011778056621551514, + 0.012189188040792942, + 0.007955383509397507, + 0.013666032813489437, + 0.010038190521299839, + 0.016765166074037552, + 0.008716055192053318, + 0.010490341112017632, + 0.006518350448459387, + 0.013987099751830101, + 0.006930076517164707, + 0.012058654800057411, + 0.005706408526748419, + 0.008977685123682022, + 0.007237800396978855, + 0.009241091087460518, + 0.007560172118246555, + 0.011823211796581745, + 0.013687041588127613, + 0.01692149229347706, + 0.008120512589812279, + 0.01102382317185402, + 0.010799259878695011, + 0.012314235791563988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 9 + ], + "selected_reward": -0.17995059490203857 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200683578848839 + ], + "opposite_logprob_tokens": [ + -6.127200603485107 + ], + "logprob_reward": -0.002200683578848839, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200683578848839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039553478360176086, + 0.38771384954452515, + 0.08874595910310745, + 0.2570398151874542, + 0.028881654143333435, + 0.04153844714164734, + 0.013978245668113232, + 0.08598270267248154, + 0.025703759863972664, + 0.030862124636769295 + ], + "credit_map": [ + 0.004237772896885872, + 0.0034590072464197874, + 0.004812734201550484, + 0.008754831738770008, + 0.004055000375956297, + 0.005450144875794649, + 0.005965619347989559, + 0.0028183639515191317, + 0.12723080813884735, + 0.03704576566815376, + 0.025613445788621902, + 0.021381793543696404, + 0.01910320669412613, + 0.01256825216114521, + 0.12413828819990158, + 0.02063223160803318, + 0.01295964140444994, + 0.015468057245016098, + 0.011010202579200268, + 0.010584354400634766, + 0.01109112985432148, + 0.00671354029327631, + 0.013978255912661552, + 0.006940770428627729, + 0.049485236406326294, + 0.03127558156847954, + 0.05040184035897255, + 0.02535553276538849, + 0.02302360162138939, + 0.017800960689783096, + 0.03372848033905029, + 0.025968577712774277, + 0.0035535001661628485, + 0.0041411444544792175, + 0.0045113894157111645, + 0.003986932337284088, + 0.0030583019834011793, + 0.002707586158066988, + 0.0034730553161352873, + 0.003449744312092662, + 0.005663787014782429, + 0.004723452962934971, + 0.004761544521898031, + 0.0037980107590556145, + 0.007595214061439037, + 0.005027002189308405, + 0.005764462053775787, + 0.00420497078448534, + 0.0018582780612632632, + 0.0015957687282934785, + 0.0010931013384833932, + 0.0015925129409879446, + 0.002327734837308526, + 0.001547065912745893, + 0.00222030165605247, + 0.0017434805631637573, + 0.011303993873298168, + 0.012302526272833347, + 0.007682495750486851, + 0.012138930149376392, + 0.008715915493667126, + 0.014852370135486126, + 0.0082401093095541, + 0.010746356099843979, + 0.0024788095615804195, + 0.005726550240069628, + 0.002679946832358837, + 0.004467393271625042, + 0.002573157660663128, + 0.0031539767514914274, + 0.00274356035515666, + 0.001880364608950913, + 0.0029733343981206417, + 0.004297052975744009, + 0.005622968543320894, + 0.006338732782751322, + 0.002483571181073785, + 0.0040261633694171906, + 0.002829704899340868, + 0.0022905932273715734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.002200683578848839 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811485692858696 + ], + "opposite_logprob_tokens": [ + -3.2881147861480713 + ], + "logprob_reward": -0.03811485692858696, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811485692858696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08082037419080734, + 0.32154861092567444, + 0.06595733016729355, + 0.09723927080631256, + 0.05134708806872368, + 0.16095241904258728, + 0.03585813194513321, + 0.11257008463144302, + 0.029424935579299927, + 0.04428178817033768 + ], + "credit_map": [ + 0.006797497160732746, + 0.007975952699780464, + 0.00805702991783619, + 0.017530295997858047, + 0.0097450977191329, + 0.007190643344074488, + 0.01407774630934, + 0.00944611057639122, + 0.10242181271314621, + 0.0312457587569952, + 0.0211596991866827, + 0.019256308674812317, + 0.017780376598238945, + 0.013078550808131695, + 0.09939059615135193, + 0.017215494066476822, + 0.0071428255178034306, + 0.006645592860877514, + 0.009603847749531269, + 0.009250731207430363, + 0.011436901055276394, + 0.006342739798128605, + 0.010401624254882336, + 0.005133067257702351, + 0.009687593206763268, + 0.011523386463522911, + 0.015795888379216194, + 0.011534381657838821, + 0.009938745759427547, + 0.0070065101608633995, + 0.01363429520279169, + 0.018118474632501602, + 0.007698933128267527, + 0.006887665018439293, + 0.007245373446494341, + 0.0062963045202195644, + 0.00489051965996623, + 0.0053652431815862656, + 0.0065824915654957294, + 0.006380556151270866, + 0.01845634914934635, + 0.01875031180679798, + 0.020533615723252296, + 0.027391748502850533, + 0.023548850789666176, + 0.021892303600907326, + 0.01931149512529373, + 0.011067739687860012, + 0.0035138705279678106, + 0.003849470755085349, + 0.002664677333086729, + 0.006814148742705584, + 0.005029635038226843, + 0.004339905455708504, + 0.005626675207167864, + 0.0040197488851845264, + 0.013659648597240448, + 0.015588412992656231, + 0.010281774215400219, + 0.014797339215874672, + 0.016023846343159676, + 0.01772012934088707, + 0.010387740097939968, + 0.014111185446381569, + 0.003200442297384143, + 0.004197052679955959, + 0.0038982718251645565, + 0.003156984457746148, + 0.0028864990454167128, + 0.005332400556653738, + 0.0029306404758244753, + 0.003822643542662263, + 0.0054376679472625256, + 0.006323264911770821, + 0.0086292065680027, + 0.011062217876315117, + 0.0030007201712578535, + 0.0035289409570395947, + 0.003483912907540798, + 0.0028158538043498993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.03811485692858696 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057950384914875 + ], + "opposite_logprob_tokens": [ + -3.8955795764923096 + ], + "logprob_reward": -0.02057950384914875, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057950384914875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08225277811288834, + 0.2150970697402954, + 0.05288446694612503, + 0.1340741515159607, + 0.046613264828920364, + 0.02040627785027027, + 0.025480952113866806, + 0.33435967564582825, + 0.05349245294928551, + 0.03533896803855896 + ], + "credit_map": [ + 0.006657252553850412, + 0.006507229525595903, + 0.008610357530415058, + 0.021344762295484543, + 0.008269641548395157, + 0.005889188032597303, + 0.017732229083776474, + 0.0072421119548380375, + 0.05480845272541046, + 0.01782391406595707, + 0.009562037885189056, + 0.013993550091981888, + 0.015308023430407047, + 0.010769559070467949, + 0.0792652815580368, + 0.013566241599619389, + 0.006450772285461426, + 0.00784370955079794, + 0.004779525566846132, + 0.007401001639664173, + 0.00808028969913721, + 0.004306077491492033, + 0.009429074823856354, + 0.0045940084382891655, + 0.01969539374113083, + 0.015107393264770508, + 0.021427176892757416, + 0.013509147800505161, + 0.014173449017107487, + 0.010354087688028812, + 0.017044585198163986, + 0.022762905806303024, + 0.003061879426240921, + 0.00294350809417665, + 0.0028599011711776257, + 0.003548472886905074, + 0.011331146582961082, + 0.003157653147354722, + 0.013012005016207695, + 0.006698695942759514, + 0.002060156548395753, + 0.0018859851406887174, + 0.002668729517608881, + 0.0015445533208549023, + 0.0037177607882767916, + 0.0028244233690202236, + 0.003332623280584812, + 0.0023720429744571447, + 0.0027145198546350002, + 0.002955768723040819, + 0.002220000373199582, + 0.0027072778902947903, + 0.004914391785860062, + 0.0027765503618866205, + 0.0040013729594647884, + 0.003191065276041627, + 0.046248894184827805, + 0.04694153741002083, + 0.030456319451332092, + 0.04569540172815323, + 0.03334798663854599, + 0.05888031795620918, + 0.0314134880900383, + 0.04137571156024933, + 0.008100125938653946, + 0.0083305099979043, + 0.006996862590312958, + 0.005217359866946936, + 0.005287733860313892, + 0.00841082539409399, + 0.004644469358026981, + 0.006504559423774481, + 0.0034686920698732138, + 0.005541646853089333, + 0.0058150514960289, + 0.006258884910494089, + 0.0033711136784404516, + 0.004660836886614561, + 0.003046253230422735, + 0.0031764868181198835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.02057950384914875 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the surreal and emotive style of Makoto Shinkai, a low-angle tracking shot follows Shun Kazama, a determined young man, as he strides unwaveringly forward. He wears a dark, high-collared school uniform, the fabric appearing damp and subtly shimmering under intricate lighting. The entire scene is bathed in the iconic golden hour light, casting long shadows and infusing the atmosphere with a warm, nostalgic glow. As the camera smoothly slides behind and to the side of him, dramatic anamorphic lens flares streak horizontally across the screen, created by the headlights of passing cars and glowing streetlights. The backdrop is a rain-slicked city street at dusk, every detail rendered with photorealistic precision; the wet asphalt perfectly reflects the vibrant neon signs of storefronts and the deep blue of the twilight sky, creating a world of idyllic beauty. Shun walks with a steady and purposeful pace, his foot occasionally splashing into a shallow puddle, sending a cloud of glowing droplets into the air. His expression is intensely focused, his gaze fixed on an unseen point ahead, embodying a powerful sense of forward momentum and a quiet, emotional gravity.", + "seed": 3269, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_11/index_6074_seed_3269_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.038194318395107986, + "vlm_logprob": -0.038194318395107986, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.038194318395107986, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024698752909898758 + ], + "opposite_logprob_tokens": [ + -4.149698734283447 + ], + "logprob_reward": -0.024698752909898758, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.024698752909898758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051283322274684906, + 0.0356593057513237, + 0.4281250238418579, + 0.19944992661476135, + 0.06705229729413986, + 0.08799030631780624, + 0.06598343700170517, + 0.02644718438386917, + 0.02050025574862957, + 0.017508957535028458 + ], + "credit_map": [ + 0.008169109001755714, + 0.006559945177286863, + 0.006769730243831873, + 0.00504315784201026, + 0.005081170704215765, + 0.006472154054790735, + 0.0055827796459198, + 0.007605275604873896, + 0.002939322032034397, + 0.0067618307657539845, + 0.0023487412836402655, + 0.004289119504392147, + 0.005830392707139254, + 0.004050574265420437, + 0.005919354967772961, + 0.003519969992339611, + 0.029403407126665115, + 0.0243288055062294, + 0.15115317702293396, + 0.02576262503862381, + 0.017712203785777092, + 0.012173186056315899, + 0.14099106192588806, + 0.02660060115158558, + 0.006456547416746616, + 0.006564090494066477, + 0.008400500752031803, + 0.08250905573368073, + 0.0070000700652599335, + 0.006235557608306408, + 0.07416173815727234, + 0.00812235102057457, + 0.005655497312545776, + 0.00541964964941144, + 0.013187471777200699, + 0.01227810513228178, + 0.006168783642351627, + 0.00750372651964426, + 0.007851256988942623, + 0.008987807668745518, + 0.007628185208886862, + 0.00861372146755457, + 0.011366596445441246, + 0.02241172455251217, + 0.0052041420713067055, + 0.005134952254593372, + 0.01366544608026743, + 0.013965542428195477, + 0.008528552018105984, + 0.006979457102715969, + 0.004601269960403442, + 0.006968968082219362, + 0.006218248046934605, + 0.015147220343351364, + 0.008626331575214863, + 0.008913391269743443, + 0.0024558824952691793, + 0.0017674636328592896, + 0.003384219715371728, + 0.004039181396365166, + 0.004124136175960302, + 0.0025395997799932957, + 0.00534061947837472, + 0.0027960832230746746, + 0.002531124744564295, + 0.001979868160560727, + 0.002825692528858781, + 0.0030786970164626837, + 0.0013445754302665591, + 0.002173866145312786, + 0.003653724445030093, + 0.002912706695497036, + 0.001652369974181056, + 0.0012767085572704673, + 0.003059893846511841, + 0.004016425460577011, + 0.0016624131239950657, + 0.0013382077449932694, + 0.001962264534085989, + 0.0025406726635992527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.024698752909898758 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296641826629639 + ], + "opposite_logprob_tokens": [ + -2.237966537475586 + ], + "logprob_reward": -0.11296641826629639, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296641826629639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09657545387744904, + 0.08896077424287796, + 0.03603231906890869, + 0.2193102091550827, + 0.1259395182132721, + 0.10533791035413742, + 0.18495149910449982, + 0.057299014180898666, + 0.044511329382658005, + 0.04108194634318352 + ], + "credit_map": [ + 0.013563559390604496, + 0.011530124582350254, + 0.012518907897174358, + 0.012192199937999249, + 0.010218399576842785, + 0.009662875905632973, + 0.009520224295556545, + 0.017369164153933525, + 0.008171272464096546, + 0.015149864368140697, + 0.007747545838356018, + 0.009146804921329021, + 0.013042852282524109, + 0.012926333583891392, + 0.013466659002006054, + 0.009309442713856697, + 0.006886556278914213, + 0.005149126052856445, + 0.0055819484405219555, + 0.002907764632254839, + 0.003609044710174203, + 0.003039916278794408, + 0.004362250678241253, + 0.004495711997151375, + 0.011598989367485046, + 0.008784527890384197, + 0.021143702790141106, + 0.07737090438604355, + 0.008049949072301388, + 0.007937794551253319, + 0.07051850855350494, + 0.013905825093388557, + 0.010418208315968513, + 0.011122625321149826, + 0.02509412169456482, + 0.0228731632232666, + 0.008525928482413292, + 0.015105620957911015, + 0.016330260783433914, + 0.01646958664059639, + 0.01189979538321495, + 0.011078275740146637, + 0.011211064644157887, + 0.026163076981902122, + 0.0061558885499835014, + 0.007191758137196302, + 0.01734001934528351, + 0.014298032969236374, + 0.021494274958968163, + 0.015197688713669777, + 0.01010159682482481, + 0.020386943593621254, + 0.014168229885399342, + 0.05024351179599762, + 0.03129852935671806, + 0.02206072211265564, + 0.005114021245390177, + 0.006717017386108637, + 0.008912495337426662, + 0.005965196993201971, + 0.01235171128064394, + 0.008946717716753483, + 0.0061362674459815025, + 0.003155587473884225, + 0.003426759038120508, + 0.003989253658801317, + 0.006362734362483025, + 0.00819617323577404, + 0.003123008646070957, + 0.005094213876873255, + 0.007425802294164896, + 0.00689338892698288, + 0.004356126766651869, + 0.004027415532618761, + 0.007330655120313168, + 0.010363690555095673, + 0.004087365698069334, + 0.0024769746232777834, + 0.004202826879918575, + 0.004236891400068998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.11296641826629639 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673760985955596 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.00673760985955596, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673760985955596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056490276008844376, + 0.03288942947983742, + 0.5871205925941467, + 0.09395856410264969, + 0.10050782561302185, + 0.04325168579816818, + 0.025808393955230713, + 0.02366107888519764, + 0.021589728072285652, + 0.014722423627972603 + ], + "credit_map": [ + 0.010479451157152653, + 0.006781443487852812, + 0.005172086413949728, + 0.004781774710863829, + 0.006373807787895203, + 0.006641253363341093, + 0.006533246487379074, + 0.009727214463055134, + 0.0033492804504930973, + 0.0038376571610569954, + 0.003172586439177394, + 0.003204634413123131, + 0.004624394699931145, + 0.0036785209085792303, + 0.006949068512767553, + 0.004073287360370159, + 0.04249688982963562, + 0.03245348483324051, + 0.1971389800310135, + 0.03681158274412155, + 0.027710149064660072, + 0.018063753843307495, + 0.19572049379348755, + 0.036725252866744995, + 0.008990206755697727, + 0.005723786540329456, + 0.017625097185373306, + 0.01747712679207325, + 0.00608482863754034, + 0.0072594271041452885, + 0.016655128449201584, + 0.014142962172627449, + 0.00923256017267704, + 0.006520984228700399, + 0.021476268768310547, + 0.017281483858823776, + 0.006327473092824221, + 0.011981925927102566, + 0.012579599395394325, + 0.015107528306543827, + 0.00723301712423563, + 0.004895191174000502, + 0.004104463383555412, + 0.008143721148371696, + 0.0027353395707905293, + 0.003674579318612814, + 0.007225988898426294, + 0.005239385645836592, + 0.002739141695201397, + 0.004168706946074963, + 0.0023698629811406136, + 0.0028225979767739773, + 0.0031617346685379744, + 0.004547519143670797, + 0.0034024189226329327, + 0.0025964120868593454, + 0.0020531818736344576, + 0.0017378965858370066, + 0.004056298173964024, + 0.003024727338925004, + 0.0049780248664319515, + 0.003320972668007016, + 0.0026695223059505224, + 0.0018204556545242667, + 0.0017886973218992352, + 0.0019992245361208916, + 0.003069781931117177, + 0.003889543702825904, + 0.0019079777412116528, + 0.0020325230434536934, + 0.0031622026581317186, + 0.003739778883755207, + 0.0012983470223844051, + 0.0015851071802899241, + 0.002008302602916956, + 0.0021970404777675867, + 0.002237090142443776, + 0.0014471313916146755, + 0.0018053172389045358, + 0.002144087105989456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.00673760985955596 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329086884856224 + ], + "opposite_logprob_tokens": [ + -3.7732908725738525 + ], + "logprob_reward": -0.02329086884856224, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329086884856224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16910994052886963, + 0.025831105187535286, + 0.42627769708633423, + 0.06021905690431595, + 0.12445355951786041, + 0.06534740328788757, + 0.020357351750135422, + 0.024296827614307404, + 0.05144963040947914, + 0.0326574370265007 + ], + "credit_map": [ + 0.026101121678948402, + 0.0228506401181221, + 0.017328275367617607, + 0.015321497805416584, + 0.014974649995565414, + 0.01908167451620102, + 0.01753120683133602, + 0.03592085465788841, + 0.003031263593584299, + 0.0037166723050177097, + 0.0022453563287854195, + 0.00296991690993309, + 0.003191147930920124, + 0.0029507409781217575, + 0.005559153854846954, + 0.0021668504923582077, + 0.029737813398241997, + 0.024400264024734497, + 0.1509699821472168, + 0.025812942534685135, + 0.017318809404969215, + 0.01361802127212286, + 0.13863036036491394, + 0.025789480656385422, + 0.0064671519212424755, + 0.004111176356673241, + 0.008226265199482441, + 0.011943468824028969, + 0.004462725017219782, + 0.006493146996945143, + 0.010388392023742199, + 0.00812672171741724, + 0.010308885015547276, + 0.009473124518990517, + 0.02609463594853878, + 0.023945312947034836, + 0.0080221276730299, + 0.01469510793685913, + 0.014907329343259335, + 0.0170070119202137, + 0.010667099617421627, + 0.006805106997489929, + 0.006358916871249676, + 0.013132293708622456, + 0.004030318930745125, + 0.005056959576904774, + 0.011270392686128616, + 0.008026308380067348, + 0.0014177279081195593, + 0.002347754081711173, + 0.0015395144000649452, + 0.003102861577644944, + 0.003159730928018689, + 0.00219645444303751, + 0.0036998188588768244, + 0.0028934874571859837, + 0.0027003686409443617, + 0.0018225753447040915, + 0.00457068532705307, + 0.003430739277973771, + 0.0028003694023936987, + 0.0027696206234395504, + 0.004148904234170914, + 0.0020535620860755444, + 0.004242690280079842, + 0.0042154365219175816, + 0.0064240507781505585, + 0.006765559781342745, + 0.004908130969852209, + 0.00500292144715786, + 0.006872677709907293, + 0.01301815826445818, + 0.003124609123915434, + 0.0020098944660276175, + 0.004670839756727219, + 0.0037762620486319065, + 0.0053408341482281685, + 0.003695237683132291, + 0.004183745011687279, + 0.005856012925505638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.02329086884856224 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023277942091226578 + ], + "opposite_logprob_tokens": [ + -3.773277997970581 + ], + "logprob_reward": -0.023277942091226578, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023277942091226578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10113466531038284, + 0.04560788720846176, + 0.09409268945455551, + 0.2200678586959839, + 0.14882048964500427, + 0.06682461500167847, + 0.15206454694271088, + 0.04938407242298126, + 0.08330685645341873, + 0.038696322590112686 + ], + "credit_map": [ + 0.012768753804266453, + 0.011945168487727642, + 0.013606815598905087, + 0.012451259419322014, + 0.007545743137598038, + 0.009926678612828255, + 0.01644875295460224, + 0.01644149050116539, + 0.004929724149405956, + 0.005963942501693964, + 0.0033984347246587276, + 0.0077661024406552315, + 0.008098078891634941, + 0.004727709107100964, + 0.006679937709122896, + 0.004043958615511656, + 0.008554558269679546, + 0.006901980377733707, + 0.025791427120566368, + 0.007041980978101492, + 0.005709128454327583, + 0.004115231800824404, + 0.029328657314181328, + 0.0066497293300926685, + 0.016431083902716637, + 0.012097959406673908, + 0.028201887384057045, + 0.06261399388313293, + 0.011389192193746567, + 0.011224327608942986, + 0.05667528882622719, + 0.0214341152459383, + 0.01108116377145052, + 0.012037440203130245, + 0.028994983062148094, + 0.029909130185842514, + 0.01005680114030838, + 0.01697711832821369, + 0.02015366032719612, + 0.019610196352005005, + 0.008468195796012878, + 0.007360285613685846, + 0.006827166769653559, + 0.012983306311070919, + 0.004879843909293413, + 0.005980207584798336, + 0.009710618294775486, + 0.010614992119371891, + 0.01157078891992569, + 0.03301806375384331, + 0.011269654147326946, + 0.024997279047966003, + 0.017793752253055573, + 0.012368290685117245, + 0.021817322820425034, + 0.01922939345240593, + 0.005076565779745579, + 0.008210481144487858, + 0.0069529814645648, + 0.006635445170104504, + 0.006374388001859188, + 0.004949721973389387, + 0.00698679918423295, + 0.004197689238935709, + 0.00817975401878357, + 0.006559308152645826, + 0.013323108665645123, + 0.016470452770590782, + 0.004090158734470606, + 0.008222057484090328, + 0.014668912626802921, + 0.011793099343776703, + 0.004634407814592123, + 0.0036860695108771324, + 0.007430230733007193, + 0.007736233063042164, + 0.004989941604435444, + 0.0029117027297616005, + 0.0038111668545752764, + 0.003496569814160466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.023277942091226578 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the surreal and emotive style of Makoto Shinkai, a low-angle tracking shot follows Shun Kazama, a determined young man, as he strides unwaveringly forward. He wears a dark, high-collared school uniform, the fabric appearing damp and subtly shimmering under intricate lighting. The entire scene is bathed in the iconic golden hour light, casting long shadows and infusing the atmosphere with a warm, nostalgic glow. As the camera smoothly slides behind and to the side of him, dramatic anamorphic lens flares streak horizontally across the screen, created by the headlights of passing cars and glowing streetlights. The backdrop is a rain-slicked city street at dusk, every detail rendered with photorealistic precision; the wet asphalt perfectly reflects the vibrant neon signs of storefronts and the deep blue of the twilight sky, creating a world of idyllic beauty. Shun walks with a steady and purposeful pace, his foot occasionally splashing into a shallow puddle, sending a cloud of glowing droplets into the air. His expression is intensely focused, his gaze fixed on an unseen point ahead, embodying a powerful sense of forward momentum and a quiet, emotional gravity.", + "seed": 3270, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_11/index_6074_seed_3270_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05529851044993848, + "vlm_logprob": -0.05529851044993848, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.05529851044993848, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030985258519649506 + ], + "opposite_logprob_tokens": [ + -3.6559853553771973 + ], + "logprob_reward": -0.030985258519649506, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.030985258519649506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0648355558514595, + 0.02376449666917324, + 0.09581026434898376, + 0.04731092229485512, + 0.030552243813872337, + 0.08580842614173889, + 0.5465013980865479, + 0.04467587545514107, + 0.039350900799036026, + 0.021389998495578766 + ], + "credit_map": [ + 0.0063757929019629955, + 0.004374496638774872, + 0.00614139623939991, + 0.008831162005662918, + 0.008132046088576317, + 0.010882647708058357, + 0.009270175360143185, + 0.01082783192396164, + 0.002158462069928646, + 0.0021968658547848463, + 0.0016052209539338946, + 0.0028399834409356117, + 0.0035247434861958027, + 0.003395155305042863, + 0.003174265380948782, + 0.004869798198342323, + 0.005599534139037132, + 0.006091015879064798, + 0.005106869153678417, + 0.012987532652914524, + 0.013395302928984165, + 0.012219895608723164, + 0.030041826888918877, + 0.010368281975388527, + 0.0025593435857445, + 0.0037732564378529787, + 0.0026535866782069206, + 0.0058575826697051525, + 0.00748911127448082, + 0.008432058617472649, + 0.007489222101867199, + 0.009056759998202324, + 0.00186845229472965, + 0.002951333299279213, + 0.0021173192653805017, + 0.0060033597983419895, + 0.004780411720275879, + 0.004373025149106979, + 0.005846284795552492, + 0.002612052019685507, + 0.0096583915874362, + 0.01318917702883482, + 0.01221645250916481, + 0.012392022646963596, + 0.00833506416529417, + 0.011286080814898014, + 0.010219446383416653, + 0.008511779829859734, + 0.025960173457860947, + 0.02790897525846958, + 0.22463220357894897, + 0.19794127345085144, + 0.008170662447810173, + 0.011912108398973942, + 0.02274664305150509, + 0.0272293109446764, + 0.005335968919098377, + 0.0048713115975260735, + 0.007786134723573923, + 0.007472075056284666, + 0.007093410473316908, + 0.002980035962536931, + 0.006308439187705517, + 0.002828494179993868, + 0.0032647093757987022, + 0.003929414786398411, + 0.004095445387065411, + 0.006360420025885105, + 0.005726507864892483, + 0.007623387034982443, + 0.004817544482648373, + 0.0035334692802280188, + 0.0020186954643577337, + 0.002645451109856367, + 0.003346371231600642, + 0.0024926133919507265, + 0.004258179105818272, + 0.0011341689387336373, + 0.003105653915554285, + 0.0023888631258159876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.030985258519649506 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992530763149261 + ], + "opposite_logprob_tokens": [ + -1.8049253225326538 + ], + "logprob_reward": -0.17992530763149261, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992530763149261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03900116682052612, + 0.021483249962329865, + 0.05250825732946396, + 0.023725023493170738, + 0.03524879366159439, + 0.129165381193161, + 0.5484987497329712, + 0.04083023965358734, + 0.1000068411231041, + 0.00953237246721983 + ], + "credit_map": [ + 0.0039959801360964775, + 0.0044583044946193695, + 0.0026532167103141546, + 0.0053758686408400536, + 0.005956178996711969, + 0.005939943250268698, + 0.005426748190075159, + 0.0051949238404631615, + 0.0015946460189297795, + 0.002026618691161275, + 0.0013555955374613404, + 0.0019856607541441917, + 0.0027679726481437683, + 0.0034894223790615797, + 0.0027816975489258766, + 0.00548163428902626, + 0.006217623129487038, + 0.005374976899474859, + 0.008089090697467327, + 0.0049615055322647095, + 0.005586511921137571, + 0.007477192673832178, + 0.00788792222738266, + 0.006913428660482168, + 0.0016071991994976997, + 0.0026081043761223555, + 0.00210403255186975, + 0.003836998250335455, + 0.003209545277059078, + 0.003504086984321475, + 0.003038116730749607, + 0.003816938726231456, + 0.0026294183917343616, + 0.0053723533637821674, + 0.0030636084266006947, + 0.005936779081821442, + 0.004280338529497385, + 0.004522997885942459, + 0.00528072752058506, + 0.004162563942372799, + 0.015015942044556141, + 0.01941858045756817, + 0.02000647410750389, + 0.01924458146095276, + 0.011350814253091812, + 0.015576080419123173, + 0.0160995926707983, + 0.012453307397663593, + 0.026744728907942772, + 0.02857971377670765, + 0.2241368591785431, + 0.19859787821769714, + 0.008930136449635029, + 0.011568824760615826, + 0.02329206094145775, + 0.026648523285984993, + 0.00584798539057374, + 0.004671250935643911, + 0.00851345807313919, + 0.008449569344520569, + 0.003583161626011133, + 0.002627398818731308, + 0.0044729942455887794, + 0.0026644179597496986, + 0.008953304030001163, + 0.010929938405752182, + 0.010539147071540356, + 0.01677124947309494, + 0.00881303008645773, + 0.024379806593060493, + 0.01004011183977127, + 0.009580242447555065, + 0.000673591741360724, + 0.0012609796831384301, + 0.0020700811874121428, + 0.0010064479429274797, + 0.002038457663729787, + 0.0005043447599746287, + 0.0011851144954562187, + 0.0007933539454825222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.17992530763149261 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019432486733421683 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.0019432486733421683, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019432486733421683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04747524484992027, + 0.014439399354159832, + 0.032009489834308624, + 0.042099080979824066, + 0.02024519257247448, + 0.11117605119943619, + 0.6742596626281738, + 0.01685774326324463, + 0.015668801963329315, + 0.025769295170903206 + ], + "credit_map": [ + 0.004309773910790682, + 0.007871145382523537, + 0.0061253756284713745, + 0.005201125517487526, + 0.004232943989336491, + 0.007657979615032673, + 0.005155237857252359, + 0.0069216592237353325, + 0.0011034198105335236, + 0.0024358341470360756, + 0.0008371834992431104, + 0.0017129448242485523, + 0.0022937061730772257, + 0.002011560369282961, + 0.0023212293162941933, + 0.0017235218547284603, + 0.004268735647201538, + 0.002557840198278427, + 0.006481862161308527, + 0.002369393827393651, + 0.003682472510263324, + 0.004769167862832546, + 0.003734182333573699, + 0.004145835526287556, + 0.004100746009498835, + 0.005213596858084202, + 0.007167428266257048, + 0.0037275473587214947, + 0.004322393797338009, + 0.0061520664021372795, + 0.005545223131775856, + 0.0058700768277049065, + 0.002153202658519149, + 0.0034382587764412165, + 0.0022231026086956263, + 0.0018037393456324935, + 0.0015382375568151474, + 0.003302893368527293, + 0.002240855013951659, + 0.0035449040587991476, + 0.01403204258531332, + 0.019422437995672226, + 0.01566648855805397, + 0.016077982261776924, + 0.010390560142695904, + 0.013973429799079895, + 0.011398682370781898, + 0.010214430280029774, + 0.03308697044849396, + 0.03043886460363865, + 0.2843928635120392, + 0.22324621677398682, + 0.010529975406825542, + 0.013889549300074577, + 0.02633683755993843, + 0.05233842134475708, + 0.0024676001630723476, + 0.0014917771331965923, + 0.003772081108763814, + 0.0023475121706724167, + 0.0027457010000944138, + 0.000858859159052372, + 0.0020085859578102827, + 0.0011656268034130335, + 0.0015012503135949373, + 0.0017293711425736547, + 0.002288619289174676, + 0.0021644337102770805, + 0.0013385581551119685, + 0.003423376940190792, + 0.001836331794038415, + 0.0013868605019524693, + 0.004545920062810183, + 0.0026255003176629543, + 0.0027007912285625935, + 0.002744274912402034, + 0.0025543563533574343, + 0.001615585875697434, + 0.005429388955235481, + 0.003553475718945265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.0019432486733421683 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306694120168686 + ], + "opposite_logprob_tokens": [ + -3.16806697845459 + ], + "logprob_reward": -0.04306694120168686, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306694120168686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026856929063796997, + 0.010911892168223858, + 0.01671844907104969, + 0.012279878370463848, + 0.009206033311784267, + 0.01569751277565956, + 0.8840681314468384, + 0.009353489615023136, + 0.006745849270373583, + 0.008161813952028751 + ], + "credit_map": [ + 0.0027224929071962833, + 0.002755584195256233, + 0.002617079997435212, + 0.003250119974836707, + 0.0025260960683226585, + 0.004523728974163532, + 0.003913890104740858, + 0.0045479373075068, + 0.0011949692852795124, + 0.0010817216243594885, + 0.001075486303307116, + 0.0012342986883595586, + 0.001307088416069746, + 0.0016180360689759254, + 0.0016052656574174762, + 0.001795026590116322, + 0.0014435467310249805, + 0.002002248540520668, + 0.001608682214282453, + 0.0014928949531167746, + 0.0026033269241452217, + 0.0021158820018172264, + 0.003727146191522479, + 0.0017247205832973123, + 0.0012356771621853113, + 0.00120553991291672, + 0.001189030590467155, + 0.0014328763354569674, + 0.001473984564654529, + 0.001807739376090467, + 0.002453003078699112, + 0.0014820266515016556, + 0.0009703772375360131, + 0.0014200083678588271, + 0.0006220162031240761, + 0.0009550598915666342, + 0.0011214162223041058, + 0.0012739883968606591, + 0.0014924351125955582, + 0.0013507328694686294, + 0.0025283340364694595, + 0.0038760837633162737, + 0.0017783889779821038, + 0.002368543529883027, + 0.001376533880829811, + 0.0010323150781914592, + 0.0015125054633244872, + 0.0012248087441548705, + 0.04265321046113968, + 0.04384806752204895, + 0.36543914675712585, + 0.3128342032432556, + 0.013363062404096127, + 0.01856009103357792, + 0.03645217418670654, + 0.05091818794608116, + 0.0014301702613011003, + 0.0008177523850463331, + 0.0019163412507623434, + 0.0017022219253703952, + 0.0009963888442143798, + 0.0005365919205360115, + 0.0013992397580295801, + 0.0005547840264625847, + 0.0007406873046420515, + 0.0007130230660550296, + 0.0009795231744647026, + 0.0008894175989553332, + 0.0007579107768833637, + 0.0010551969753578305, + 0.0008882891270332038, + 0.0007218013633973897, + 0.0006265481933951378, + 0.0011654425179585814, + 0.0021190044935792685, + 0.0007125803967937827, + 0.0012459784047678113, + 0.0005876107607036829, + 0.0009510598029009998, + 0.0007535887416452169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.04306694120168686 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020571796223521233 + ], + "opposite_logprob_tokens": [ + -3.895571708679199 + ], + "logprob_reward": -0.020571796223521233, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020571796223521233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030573682859539986, + 0.022105474025011063, + 0.019390597939491272, + 0.015811916440725327, + 0.02682383544743061, + 0.03072740137577057, + 0.7744536399841309, + 0.026951737701892853, + 0.04018297418951988, + 0.012978778220713139 + ], + "credit_map": [ + 0.003238463308662176, + 0.003674012143164873, + 0.002032280433923006, + 0.004087445326149464, + 0.0035464998800307512, + 0.005957133136689663, + 0.0038024699315428734, + 0.004235381726175547, + 0.0016578268259763718, + 0.0015800659311935306, + 0.0019611555617302656, + 0.0026424892712384462, + 0.003982417285442352, + 0.002983840648084879, + 0.004303071182221174, + 0.0029946104623377323, + 0.0015081113670021296, + 0.0018569676904007792, + 0.00155614351388067, + 0.0013097062474116683, + 0.004420632962137461, + 0.0026231214869767427, + 0.0039622588083148, + 0.0021536583080887794, + 0.0011638551950454712, + 0.0017286401707679033, + 0.0012615916784852743, + 0.0023369642440229654, + 0.0022833221592009068, + 0.0025915540754795074, + 0.002450727391988039, + 0.001995262922719121, + 0.0019550961442291737, + 0.004008518531918526, + 0.0024488181807100773, + 0.004810779821127653, + 0.003256777999922633, + 0.0035560729447752237, + 0.004534527659416199, + 0.00225324765779078, + 0.004099879879504442, + 0.004650848917663097, + 0.004505195189267397, + 0.003703903406858444, + 0.0037701898254454136, + 0.003276870818808675, + 0.0036932886578142643, + 0.0030272258445620537, + 0.03826293721795082, + 0.0339786596596241, + 0.3289138376712799, + 0.2541971504688263, + 0.011743700131773949, + 0.015631232410669327, + 0.029422570019960403, + 0.062303610146045685, + 0.0028998786583542824, + 0.0024414616636931896, + 0.00316581130027771, + 0.005113034043461084, + 0.003941719885915518, + 0.0012949644587934017, + 0.005665338132530451, + 0.0024295311886817217, + 0.004190763924270868, + 0.005095337983220816, + 0.005448794923722744, + 0.005383780226111412, + 0.004384116735309362, + 0.006758841685950756, + 0.005389916244894266, + 0.0035314271226525307, + 0.0026026407722383738, + 0.001824039500206709, + 0.0016583192627876997, + 0.0012665257090702653, + 0.001309921033680439, + 0.0008364251116290689, + 0.001581630320288241, + 0.0018992783734574914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.020571796223521233 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the surreal and emotive style of Makoto Shinkai, a low-angle tracking shot follows Shun Kazama, a determined young man, as he strides unwaveringly forward. He wears a dark, high-collared school uniform, the fabric appearing damp and subtly shimmering under intricate lighting. The entire scene is bathed in the iconic golden hour light, casting long shadows and infusing the atmosphere with a warm, nostalgic glow. As the camera smoothly slides behind and to the side of him, dramatic anamorphic lens flares streak horizontally across the screen, created by the headlights of passing cars and glowing streetlights. The backdrop is a rain-slicked city street at dusk, every detail rendered with photorealistic precision; the wet asphalt perfectly reflects the vibrant neon signs of storefronts and the deep blue of the twilight sky, creating a world of idyllic beauty. Shun walks with a steady and purposeful pace, his foot occasionally splashing into a shallow puddle, sending a cloud of glowing droplets into the air. His expression is intensely focused, his gaze fixed on an unseen point ahead, embodying a powerful sense of forward momentum and a quiet, emotional gravity.", + "seed": 3271, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_11/index_6074_seed_3271_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05471463240683079, + "vlm_logprob": -0.05471463240683079, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.05471463240683079, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038411788642406464 + ], + "opposite_logprob_tokens": [ + -3.538411855697632 + ], + "logprob_reward": -0.038411788642406464, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.038411788642406464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17806683480739594, + 0.07744163274765015, + 0.08094565570354462, + 0.10066129267215729, + 0.08291548490524292, + 0.12022014707326889, + 0.07976069301366806, + 0.10451958328485489, + 0.1048305556178093, + 0.07063809782266617 + ], + "credit_map": [ + 0.018007628619670868, + 0.016993530094623566, + 0.024238765239715576, + 0.025883443653583527, + 0.02028813026845455, + 0.01795370690524578, + 0.023363547399640083, + 0.031338077038526535, + 0.005737306550145149, + 0.010529709048569202, + 0.019714444875717163, + 0.008572285994887352, + 0.011688102036714554, + 0.006468878127634525, + 0.009709612466394901, + 0.00502130389213562, + 0.00556773878633976, + 0.008116202428936958, + 0.00758174667134881, + 0.006499108392745256, + 0.0136045441031456, + 0.01442760694772005, + 0.01359735056757927, + 0.011551366187632084, + 0.008082622662186623, + 0.005773862823843956, + 0.006881531793624163, + 0.007578612770885229, + 0.031393472105264664, + 0.012340865097939968, + 0.020282452926039696, + 0.008327874355018139, + 0.004755108151584864, + 0.004964833147823811, + 0.003577333642169833, + 0.005638140253722668, + 0.0038417428731918335, + 0.004802332725375891, + 0.02414834313094616, + 0.03118765540421009, + 0.01120892632752657, + 0.0154672646895051, + 0.014458790421485901, + 0.01553895603865385, + 0.019364718347787857, + 0.012791266664862633, + 0.017472010105848312, + 0.013918224722146988, + 0.006079720798879862, + 0.00859761331230402, + 0.006155067123472691, + 0.015962589532136917, + 0.010631276294589043, + 0.005539150908589363, + 0.020935866981744766, + 0.0058594075962901115, + 0.008648182265460491, + 0.007633145432919264, + 0.010978005826473236, + 0.017869986593723297, + 0.01757650449872017, + 0.006697698030620813, + 0.025106076151132584, + 0.010009987279772758, + 0.027885904535651207, + 0.015457590110599995, + 0.010923121124505997, + 0.016104206442832947, + 0.0058453865349292755, + 0.007950699888169765, + 0.011798780411481857, + 0.00886488240212202, + 0.006115318275988102, + 0.009431666694581509, + 0.007187359035015106, + 0.01151250023394823, + 0.008702252991497517, + 0.006311824545264244, + 0.012817120179533958, + 0.008560060523450375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.038411788642406464 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603979617357254 + ], + "opposite_logprob_tokens": [ + -1.910398006439209 + ], + "logprob_reward": -0.1603979617357254, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603979617357254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06347879767417908, + 0.09596714377403259, + 0.09435588866472244, + 0.05624322593212128, + 0.22631001472473145, + 0.10030551999807358, + 0.07673367857933044, + 0.05224782973527908, + 0.08633337914943695, + 0.1480245292186737 + ], + "credit_map": [ + 0.008221548050642014, + 0.005114458035677671, + 0.012073644436895847, + 0.007620200049132109, + 0.005305288825184107, + 0.00840151496231556, + 0.00851671863347292, + 0.008225425146520138, + 0.007219970226287842, + 0.015792831778526306, + 0.018260115757584572, + 0.01018825639039278, + 0.01674012281000614, + 0.008199109695851803, + 0.012378070503473282, + 0.007188664749264717, + 0.00860377587378025, + 0.016556579619646072, + 0.007780133746564388, + 0.012647148221731186, + 0.015061709098517895, + 0.011382069438695908, + 0.01133111771196127, + 0.0109933540225029, + 0.010442161932587624, + 0.007651332765817642, + 0.004627245012670755, + 0.0064584556967020035, + 0.008521022275090218, + 0.005665910430252552, + 0.006793685257434845, + 0.0060834139585494995, + 0.011242873035371304, + 0.015958799049258232, + 0.01168274600058794, + 0.013991180807352066, + 0.025274211540818214, + 0.01688886433839798, + 0.060114193707704544, + 0.07115714251995087, + 0.010650303214788437, + 0.013344240374863148, + 0.0120615279302001, + 0.015118755400180817, + 0.009868979454040527, + 0.010338700376451015, + 0.015369842760264874, + 0.013553174212574959, + 0.0076395864598453045, + 0.006581046152859926, + 0.007206964306533337, + 0.013783277943730354, + 0.009776296094059944, + 0.006827407982200384, + 0.019280744716525078, + 0.005638359114527702, + 0.004758286289870739, + 0.00363999349065125, + 0.005107918754220009, + 0.008347501046955585, + 0.006356700323522091, + 0.010624928399920464, + 0.00933772511780262, + 0.004074776545166969, + 0.007505066227167845, + 0.008039398118853569, + 0.00998107623308897, + 0.012973247095942497, + 0.009811013005673885, + 0.006690801586955786, + 0.016113484278321266, + 0.015219291672110558, + 0.008709502406418324, + 0.013980153016746044, + 0.008949071168899536, + 0.030819516628980637, + 0.014570843428373337, + 0.011827000416815281, + 0.04178938642144203, + 0.017379067838191986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.1603979617357254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002821514382958412 + ], + "opposite_logprob_tokens": [ + -5.877821445465088 + ], + "logprob_reward": -0.002821514382958412, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002821514382958412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09429151564836502, + 0.06272181123495102, + 0.07194559276103973, + 0.09157145768404007, + 0.11674907803535461, + 0.08221149444580078, + 0.1347024142742157, + 0.10560982674360275, + 0.1688595414161682, + 0.0713372752070427 + ], + "credit_map": [ + 0.013463848270475864, + 0.010966992005705833, + 0.012145491316914558, + 0.01775682345032692, + 0.009277849458158016, + 0.009591566398739815, + 0.013518773019313812, + 0.007570184767246246, + 0.0044167181476950645, + 0.009058092720806599, + 0.007052807603031397, + 0.006603338290005922, + 0.013382024131715298, + 0.009532758966088295, + 0.007876332849264145, + 0.0047997357323765755, + 0.004863830283284187, + 0.009344048798084259, + 0.009205145761370659, + 0.012411314994096756, + 0.010587811470031738, + 0.007021036464720964, + 0.009229540824890137, + 0.009282865561544895, + 0.008145621977746487, + 0.006238152272999287, + 0.0045691137202084064, + 0.006553441286087036, + 0.02980867773294449, + 0.013178783468902111, + 0.016068311408162117, + 0.007009363733232021, + 0.008755635470151901, + 0.008500399999320507, + 0.010229877196252346, + 0.007461425382643938, + 0.011953825131058693, + 0.008784293197095394, + 0.028478462249040604, + 0.03258516639471054, + 0.008425300940871239, + 0.010817919857800007, + 0.009813707321882248, + 0.010421285405755043, + 0.011551626957952976, + 0.009583944454789162, + 0.012520719319581985, + 0.009076991118490696, + 0.011301721446216106, + 0.012057840824127197, + 0.010459491051733494, + 0.026643285527825356, + 0.01413237676024437, + 0.013073078356683254, + 0.03771379217505455, + 0.009320845827460289, + 0.013688042759895325, + 0.009189833886921406, + 0.018515419214963913, + 0.014552677981555462, + 0.013280851766467094, + 0.01873432658612728, + 0.009705978445708752, + 0.007942700758576393, + 0.03703969344496727, + 0.03133703023195267, + 0.01953156851232052, + 0.027386503294110298, + 0.007992798462510109, + 0.013604707084596157, + 0.017176710069179535, + 0.014790528453886509, + 0.008549888618290424, + 0.013340787962079048, + 0.006175370421260595, + 0.012398788705468178, + 0.008692892268300056, + 0.007252223324030638, + 0.010215381160378456, + 0.0047119492664933205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.002821514382958412 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048664771020412445 + ], + "opposite_logprob_tokens": [ + -3.0486648082733154 + ], + "logprob_reward": -0.048664771020412445, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048664771020412445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14835171401500702, + 0.09101426601409912, + 0.15824387967586517, + 0.1678823083639145, + 0.19110070168972015, + 0.023918267339468002, + 0.044026169925928116, + 0.0683988556265831, + 0.0775488018989563, + 0.029515016824007034 + ], + "credit_map": [ + 0.013874044641852379, + 0.01616217941045761, + 0.019231021404266357, + 0.02157656103372574, + 0.017856871709227562, + 0.026154085993766785, + 0.017522914335131645, + 0.01597403548657894, + 0.0054903412237763405, + 0.014780380763113499, + 0.009158818982541561, + 0.010148127563297749, + 0.018783563748002052, + 0.011983293108642101, + 0.013171815313398838, + 0.007497924380004406, + 0.013177060522139072, + 0.016398942098021507, + 0.016262298449873924, + 0.01846596784889698, + 0.021380973979830742, + 0.029095396399497986, + 0.02190007083117962, + 0.021563176065683365, + 0.01535270269960165, + 0.011604824103415012, + 0.012364663183689117, + 0.015614084899425507, + 0.04897157847881317, + 0.02091546729207039, + 0.027966661378741264, + 0.01509233471006155, + 0.016936734318733215, + 0.016913393512368202, + 0.02413853257894516, + 0.016226062551140785, + 0.013906328938901424, + 0.011466596275568008, + 0.043322812765836716, + 0.04819025844335556, + 0.002140589989721775, + 0.0029401862993836403, + 0.0027491291984915733, + 0.00354462000541389, + 0.003259895369410515, + 0.0023942652624100447, + 0.0036794801708310843, + 0.0032101026736199856, + 0.002735361224040389, + 0.00423329696059227, + 0.0036583899054676294, + 0.011009679175913334, + 0.0033067951444536448, + 0.003024181816726923, + 0.012888409197330475, + 0.0031700616236776114, + 0.0058733681216835976, + 0.006056919693946838, + 0.007026868872344494, + 0.010861709713935852, + 0.008594171144068241, + 0.016519060358405113, + 0.008977613411843777, + 0.004489149898290634, + 0.015539168380200863, + 0.00996831152588129, + 0.007289350964128971, + 0.01342692319303751, + 0.006860490422695875, + 0.0067556193098425865, + 0.011636048555374146, + 0.006072889547795057, + 0.0026062994729727507, + 0.003081436036154628, + 0.0017431378364562988, + 0.004518119152635336, + 0.004128522705286741, + 0.0030553119722753763, + 0.00657018693163991, + 0.0038120034150779247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.048664771020412445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing a dark, high-collared school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character walking forward with a steady pace while splashing into a shallow puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rain-slicked city street at dusk with neon signs and a twilight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature dramatic anamorphic lens flares streaking horizontally across the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a smooth low-angle tracking shot that slides behind and to the side of the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023277126252651215 + ], + "opposite_logprob_tokens": [ + -3.7732770442962646 + ], + "logprob_reward": -0.023277126252651215, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023277126252651215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11563235521316528, + 0.17147164046764374, + 0.07344204932451248, + 0.06483853608369827, + 0.13558465242385864, + 0.10266156494617462, + 0.06993944197893143, + 0.04210542514920235, + 0.08563916385173798, + 0.13868510723114014 + ], + "credit_map": [ + 0.008994725532829762, + 0.009950460866093636, + 0.012529685162007809, + 0.02099660411477089, + 0.011140609160065651, + 0.00946699921041727, + 0.017887214198708534, + 0.024666063487529755, + 0.012671370059251785, + 0.0262205358594656, + 0.016644759103655815, + 0.021981334313750267, + 0.03749154880642891, + 0.015179825946688652, + 0.026729002594947815, + 0.01455327495932579, + 0.005179365165531635, + 0.006128726061433554, + 0.007909956388175488, + 0.0068703508004546165, + 0.008291694335639477, + 0.013707412406802177, + 0.014319361187517643, + 0.01103518158197403, + 0.014944425784051418, + 0.004975493531674147, + 0.00757017033174634, + 0.0062269773334264755, + 0.007524693850427866, + 0.006925930734723806, + 0.008236309513449669, + 0.008434535004198551, + 0.006903530098497868, + 0.010124181397259235, + 0.006993627175688744, + 0.007605373859405518, + 0.011637270450592041, + 0.008627466857433319, + 0.038181815296411514, + 0.0455113984644413, + 0.009339929558336735, + 0.013307635672390461, + 0.013762371614575386, + 0.013148600235581398, + 0.014447090215981007, + 0.00996114406734705, + 0.014440745115280151, + 0.014254061505198479, + 0.0038217694964259863, + 0.0038346515502780676, + 0.002743036486208439, + 0.0210843738168478, + 0.0052140881307423115, + 0.002722304780036211, + 0.02672002650797367, + 0.0037991960998624563, + 0.004355780780315399, + 0.0029698293656110764, + 0.004276650957763195, + 0.007123211864382029, + 0.00506006320938468, + 0.007642223499715328, + 0.0070535968989133835, + 0.0036240764893591404, + 0.016220824792981148, + 0.007924417965114117, + 0.007960088551044464, + 0.014063646085560322, + 0.012456457130610943, + 0.006358481477946043, + 0.015709875151515007, + 0.004945383872836828, + 0.009927314706146717, + 0.01338084228336811, + 0.007864328101277351, + 0.027836402878165245, + 0.020816750824451447, + 0.010004451498389244, + 0.03940623253583908, + 0.009448801167309284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 4 + ], + "selected_reward": -0.023277126252651215 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_12/index_tensor([9232, 9232])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_12/index_tensor([9232, 9232])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fa38f8ad03717f9f75f0b3acd874859845d4c47e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_12/index_tensor([9232, 9232])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a military archival-style shot with a signature vintage film texture, featuring slight color desaturation and noticeable grain, the camera pans in a fluid, unbroken 360-degree pan, closely following a missile hurtling through the sky at high speed. The missile's sleek gray-white metal hull and tail stabilizers reflect a cool, austere luster in the harsh sunlight. A dense, swirling column of white smoke erupts from its tail, leaving a striking, persistent trail across the vast pale blue sky. As the camera rotates at a constant speed, the viewer's perspective shifts around the missile, while the horizon and sparse clouds in the background rotate accordingly, creating a powerful sense of motion and an immersive sense of speed. In the final stage of the surround shooting, the camera angle is exactly aimed at the direction of the missile's advance, capturing it rushing towards a distant, vaguely outlined aerial target with unstoppable momentum and precision. Finally, at the decisive moment when the missile contacts the target, the picture is occupied by a violent and silent explosion. A dazzling orange-red fireball suddenly bursts out, and then turns into a black smoke mixed with metal fragments, slowly dissipating in the air. The whole process is full of calm and shocking destructive aesthetics.", + "seed": 25731, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_12/index_9232_seed_25731_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3044964429922402, + "vlm_logprob": -0.3044964429922402, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.3044964429922402, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034429844468832016 + ], + "opposite_logprob_tokens": [ + -3.5344297885894775 + ], + "logprob_reward": -0.034429844468832016, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.034429844468832016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07881366461515427, + 0.2561275362968445, + 0.10123282670974731, + 0.03657810762524605, + 0.1275169849395752, + 0.042974185198545456, + 0.06409930437803268, + 0.06652838736772537, + 0.07018832117319107, + 0.15594060719013214 + ], + "credit_map": [ + 0.010551739484071732, + 0.007918181829154491, + 0.009998166002333164, + 0.010606061667203903, + 0.011091816239058971, + 0.012625807896256447, + 0.007500512059777975, + 0.00852138176560402, + 0.026394987478852272, + 0.027927201241254807, + 0.027242006734013557, + 0.027307365089654922, + 0.033805955201387405, + 0.05337300896644592, + 0.02254694513976574, + 0.03753006085753441, + 0.008799037896096706, + 0.009110894054174423, + 0.010566511191427708, + 0.007494274061173201, + 0.013257618993520737, + 0.02924647368490696, + 0.010897252708673477, + 0.011860765516757965, + 0.003125280374661088, + 0.004757771734148264, + 0.003961838316172361, + 0.004204429220408201, + 0.006573006045073271, + 0.005142796318978071, + 0.0028992416337132454, + 0.0059137423522770405, + 0.012397225946187973, + 0.014341749250888824, + 0.016232147812843323, + 0.01328864973038435, + 0.0195242241024971, + 0.01657472737133503, + 0.013488908298313618, + 0.02166934683918953, + 0.0053751165978610516, + 0.006084976252168417, + 0.005985107272863388, + 0.004555420018732548, + 0.004382047336548567, + 0.006480775307863951, + 0.0033490294590592384, + 0.006761712953448296, + 0.00852656364440918, + 0.009775799699127674, + 0.006311224307864904, + 0.006642667111009359, + 0.006842657458037138, + 0.007941662333905697, + 0.008373375982046127, + 0.009685357101261616, + 0.009089424274861813, + 0.007966665551066399, + 0.009154282510280609, + 0.011632615700364113, + 0.00702047860249877, + 0.007206199690699577, + 0.007017475552856922, + 0.007441249676048756, + 0.0056383018381893635, + 0.005830816458910704, + 0.008913831785321236, + 0.012705609202384949, + 0.007733388803899288, + 0.008089405484497547, + 0.009628436528146267, + 0.011648539453744888, + 0.013664133846759796, + 0.015496955253183842, + 0.021564871072769165, + 0.01724131964147091, + 0.02035396359860897, + 0.014863017946481705, + 0.027169914916157722, + 0.025586429983377457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 4 + ], + "selected_reward": -0.034429844468832016 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296215653419495 + ], + "opposite_logprob_tokens": [ + -2.237962245941162 + ], + "logprob_reward": -0.11296215653419495, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296215653419495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08301873505115509, + 0.21387174725532532, + 0.07366970181465149, + 0.08317475765943527, + 0.038097232580184937, + 0.04910406097769737, + 0.2326563000679016, + 0.0695418119430542, + 0.09100908041000366, + 0.06585653871297836 + ], + "credit_map": [ + 0.015103762038052082, + 0.009419906884431839, + 0.011685987003147602, + 0.008871684782207012, + 0.011762517504394054, + 0.009667801670730114, + 0.007559029385447502, + 0.008948054164648056, + 0.02068241499364376, + 0.023930655792355537, + 0.024090858176350594, + 0.02103947103023529, + 0.03087039478123188, + 0.04658038914203644, + 0.01864520087838173, + 0.02803238295018673, + 0.006854809354990721, + 0.006355721037834883, + 0.007191394921392202, + 0.005991300102323294, + 0.010293245315551758, + 0.020439686253666878, + 0.00689829234033823, + 0.009645265527069569, + 0.008442548103630543, + 0.006342618260532618, + 0.010971731506288052, + 0.00952820759266615, + 0.009400143288075924, + 0.010724804364144802, + 0.008508220314979553, + 0.019256500527262688, + 0.003654315834864974, + 0.004787153098732233, + 0.0041873459704220295, + 0.004461981821805239, + 0.004875221289694309, + 0.004770482424646616, + 0.0046574417501688, + 0.0067032938823103905, + 0.004959458485245705, + 0.007029515691101551, + 0.007013896945863962, + 0.007962730713188648, + 0.004062971565872431, + 0.00621073367074132, + 0.004075086209923029, + 0.007789671886712313, + 0.0359138660132885, + 0.03786296024918556, + 0.02542947605252266, + 0.01819971576333046, + 0.03581896796822548, + 0.03855501115322113, + 0.022427404299378395, + 0.01844893954694271, + 0.005936348345130682, + 0.010738692246377468, + 0.00853334367275238, + 0.0063705178909003735, + 0.010895984247326851, + 0.010149846784770489, + 0.008716429583728313, + 0.008200659416615963, + 0.008206937462091446, + 0.0066927410662174225, + 0.018555471673607826, + 0.02138783596456051, + 0.007084758952260017, + 0.011017659679055214, + 0.010328168980777264, + 0.007735516410320997, + 0.009809235110878944, + 0.0059509482234716415, + 0.013060121797025204, + 0.011619010008871555, + 0.007024221587926149, + 0.006042683031409979, + 0.0055580404587090015, + 0.006792283616960049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 8 + ], + "selected_reward": -0.11296215653419495 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3133878707885742 + ], + "opposite_logprob_tokens": [ + -0.31338784098625183 + ], + "logprob_reward": -1.3133878707885742, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3133878707885742, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07126972079277039, + 0.23606999218463898, + 0.04872547835111618, + 0.040629301220178604, + 0.039687998592853546, + 0.07935342192649841, + 0.13986288011074066, + 0.1833120733499527, + 0.10919263958930969, + 0.05189654231071472 + ], + "credit_map": [ + 0.010094203986227512, + 0.0073661501519382, + 0.008334361016750336, + 0.006743110716342926, + 0.010388529859483242, + 0.012935483828186989, + 0.007013692054897547, + 0.008394189178943634, + 0.02225382812321186, + 0.025467341765761375, + 0.025134585797786713, + 0.0263675507158041, + 0.033642228692770004, + 0.047201164066791534, + 0.020991243422031403, + 0.03501203656196594, + 0.0031726856250315905, + 0.0041479370556771755, + 0.004948108457028866, + 0.0038544288836419582, + 0.005954009015113115, + 0.015390357933938503, + 0.006328014191240072, + 0.004929929506033659, + 0.0037916547153145075, + 0.005606574006378651, + 0.005106565076857805, + 0.005861163139343262, + 0.006851332727819681, + 0.005491478834301233, + 0.0036572499666363, + 0.004263279028236866, + 0.0034591704607009888, + 0.005134530831128359, + 0.004644830245524645, + 0.0037182033993303776, + 0.005651668179780245, + 0.005447687581181526, + 0.0046704430133104324, + 0.0069614602252841, + 0.008505703881382942, + 0.011186500079929829, + 0.010569903999567032, + 0.011303141713142395, + 0.008443805389106274, + 0.010424036532640457, + 0.009060201235115528, + 0.009860118851065636, + 0.019503237679600716, + 0.024163734167814255, + 0.01631215773522854, + 0.011396721005439758, + 0.0206192247569561, + 0.021133070811629295, + 0.013417412526905537, + 0.01331731304526329, + 0.02214670367538929, + 0.022153954952955246, + 0.030796952545642853, + 0.016822030767798424, + 0.02534383349120617, + 0.018893428146839142, + 0.023606764152646065, + 0.023548394441604614, + 0.008791324682533741, + 0.009291257709264755, + 0.012362856417894363, + 0.013472023420035839, + 0.014238160103559494, + 0.013369113206863403, + 0.013429815880954266, + 0.02423807419836521, + 0.006896666716784239, + 0.003975937142968178, + 0.010079729370772839, + 0.008486350066959858, + 0.006384390406310558, + 0.00517207570374012, + 0.005675463937222958, + 0.005225926171988249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -1.3133878707885742 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496591702103615 + ], + "opposite_logprob_tokens": [ + -2.9299659729003906 + ], + "logprob_reward": -0.05496591702103615, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496591702103615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03227394446730614, + 0.07454406470060349, + 0.09176190197467804, + 0.027739305049180984, + 0.030214210972189903, + 0.0414486788213253, + 0.2639566957950592, + 0.05742286518216133, + 0.11639809608459473, + 0.2642402946949005 + ], + "credit_map": [ + 0.002686962718144059, + 0.002890908159315586, + 0.003087586723268032, + 0.002861473010852933, + 0.005981470458209515, + 0.007953131571412086, + 0.003360281465575099, + 0.003452130826190114, + 0.006619131192564964, + 0.008698313497006893, + 0.008214154280722141, + 0.008271259255707264, + 0.01099958922713995, + 0.01594141125679016, + 0.006622819695621729, + 0.009177383035421371, + 0.007219440303742886, + 0.008187945000827312, + 0.009481027722358704, + 0.00755559466779232, + 0.012484438717365265, + 0.027203761041164398, + 0.008759169839322567, + 0.010870534926652908, + 0.002414851449429989, + 0.0029461225494742393, + 0.0032239535357803106, + 0.004576453939080238, + 0.00423405459150672, + 0.0033636142034083605, + 0.0031232291366904974, + 0.00385702564381063, + 0.002760540693998337, + 0.004383400082588196, + 0.003777124220505357, + 0.0033052123617380857, + 0.003621236188337207, + 0.004108248744159937, + 0.00352575839497149, + 0.004732691217213869, + 0.004829905927181244, + 0.005917998496443033, + 0.004636989440768957, + 0.003781113773584366, + 0.004835838917642832, + 0.006480008363723755, + 0.003583682933822274, + 0.007383143063634634, + 0.0350792221724987, + 0.05107679218053818, + 0.02353975549340248, + 0.01959792897105217, + 0.042359061539173126, + 0.039375804364681244, + 0.028354374691843987, + 0.024573728442192078, + 0.005112790036946535, + 0.007229737937450409, + 0.008317675441503525, + 0.007972137071192265, + 0.007527161855250597, + 0.0072024185210466385, + 0.006641536485403776, + 0.007419408764690161, + 0.009755427949130535, + 0.010630168952047825, + 0.017044929787516594, + 0.020444385707378387, + 0.010186671279370785, + 0.012641675770282745, + 0.016256961971521378, + 0.019437866285443306, + 0.024865731596946716, + 0.021482344716787338, + 0.03756709769368172, + 0.03780069574713707, + 0.03555271774530411, + 0.02751319482922554, + 0.03434896469116211, + 0.04510952904820442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.05496591702103615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736426148563623 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.006736426148563623, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736426148563623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0891595408320427, + 0.2338114231824875, + 0.031158743426203728, + 0.041378360241651535, + 0.027790918946266174, + 0.05011224374175072, + 0.2686169147491455, + 0.0802745670080185, + 0.0966460108757019, + 0.08105132728815079 + ], + "credit_map": [ + 0.010855492204427719, + 0.008307873271405697, + 0.010197064839303493, + 0.00643916055560112, + 0.015621558763086796, + 0.01408339012414217, + 0.009462774731218815, + 0.014192218892276287, + 0.02140865847468376, + 0.02616911195218563, + 0.025017086416482925, + 0.024916663765907288, + 0.03382851555943489, + 0.050838254392147064, + 0.02177659422159195, + 0.02985651232302189, + 0.00326199340634048, + 0.0032042204402387142, + 0.003666201140731573, + 0.0029160233680158854, + 0.004338931292295456, + 0.006256689317524433, + 0.0038181410636752844, + 0.0036965401377528906, + 0.0038582319393754005, + 0.003739184234291315, + 0.005745869595557451, + 0.00544562516734004, + 0.006222632713615894, + 0.005295444745570421, + 0.004866134840995073, + 0.0062052286230027676, + 0.00267282547429204, + 0.003698766464367509, + 0.0034352208022028208, + 0.0026017562486231327, + 0.005466706119477749, + 0.004154481925070286, + 0.002652406459674239, + 0.003108751028776169, + 0.0048249936662614346, + 0.007292679511010647, + 0.007193650119006634, + 0.008207980543375015, + 0.005291384644806385, + 0.005401551723480225, + 0.005191868171095848, + 0.006708128843456507, + 0.037367045879364014, + 0.05099208652973175, + 0.02593112923204899, + 0.02087261527776718, + 0.04392723739147186, + 0.041769757866859436, + 0.02515472285449505, + 0.022602304816246033, + 0.00979615654796362, + 0.010009123012423515, + 0.014214091002941132, + 0.00807062815874815, + 0.011826646514236927, + 0.009224878624081612, + 0.008859359659254551, + 0.008273674175143242, + 0.00847802683711052, + 0.006230365019291639, + 0.017049256712198257, + 0.016972994431853294, + 0.007740458473563194, + 0.012580064125359058, + 0.01065612118691206, + 0.016938718035817146, + 0.009721196256577969, + 0.006544894073158503, + 0.014935850165784359, + 0.012610669247806072, + 0.012586986646056175, + 0.007249004207551479, + 0.008361504413187504, + 0.009041215293109417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 8 + ], + "selected_reward": -0.006736426148563623 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a military archival-style shot with a signature vintage film texture, featuring slight color desaturation and noticeable grain, the camera pans in a fluid, unbroken 360-degree pan, closely following a missile hurtling through the sky at high speed. The missile's sleek gray-white metal hull and tail stabilizers reflect a cool, austere luster in the harsh sunlight. A dense, swirling column of white smoke erupts from its tail, leaving a striking, persistent trail across the vast pale blue sky. As the camera rotates at a constant speed, the viewer's perspective shifts around the missile, while the horizon and sparse clouds in the background rotate accordingly, creating a powerful sense of motion and an immersive sense of speed. In the final stage of the surround shooting, the camera angle is exactly aimed at the direction of the missile's advance, capturing it rushing towards a distant, vaguely outlined aerial target with unstoppable momentum and precision. Finally, at the decisive moment when the missile contacts the target, the picture is occupied by a violent and silent explosion. A dazzling orange-red fireball suddenly bursts out, and then turns into a black smoke mixed with metal fragments, slowly dissipating in the air. The whole process is full of calm and shocking destructive aesthetics.", + "seed": 25732, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_12/index_9232_seed_25732_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3039174632402137, + "vlm_logprob": -0.3039174632402137, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.3039174632402137, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04079494625329971 + ], + "opposite_logprob_tokens": [ + -3.540794849395752 + ], + "logprob_reward": -0.04079494625329971, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.04079494625329971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18947406113147736, + 0.04902592673897743, + 0.08406174927949905, + 0.3213091194629669, + 0.04651116579771042, + 0.03968141973018646, + 0.18158645927906036, + 0.033250387758016586, + 0.02991059236228466, + 0.0251891128718853 + ], + "credit_map": [ + 0.02366984821856022, + 0.029170311987400055, + 0.021279050037264824, + 0.01949905790388584, + 0.025446468964219093, + 0.021026600152254105, + 0.019835002720355988, + 0.029547739773988724, + 0.002967318519949913, + 0.0077510252594947815, + 0.003381275339052081, + 0.002447765087708831, + 0.016505194827914238, + 0.008180462755262852, + 0.0032188664190471172, + 0.0045740255154669285, + 0.003143695881590247, + 0.03738703206181526, + 0.0018716868944466114, + 0.0012144538341090083, + 0.023632101714611053, + 0.008576001040637493, + 0.0014720959588885307, + 0.006764687597751617, + 0.018944911658763885, + 0.053397346287965775, + 0.007920988835394382, + 0.006774591747671366, + 0.12542560696601868, + 0.07687501609325409, + 0.00661329785361886, + 0.025357404723763466, + 0.0023190462961792946, + 0.017509417608380318, + 0.003120244713500142, + 0.0022053092252463102, + 0.012439755722880363, + 0.004712675232440233, + 0.0017819389468058944, + 0.0024227797985076904, + 0.0029755299910902977, + 0.00845651887357235, + 0.003694243961945176, + 0.0032656004186719656, + 0.011071672663092613, + 0.004639848135411739, + 0.0023995493538677692, + 0.003178459359332919, + 0.016065867617726326, + 0.026082750409841537, + 0.024727875366806984, + 0.021712614223361015, + 0.019271131604909897, + 0.037639815360307693, + 0.012016728520393372, + 0.02406967431306839, + 0.0028760789427906275, + 0.003581006545573473, + 0.004891523625701666, + 0.006230930797755718, + 0.004040813073515892, + 0.004418619908392429, + 0.003319017356261611, + 0.003892400534823537, + 0.005353322718292475, + 0.0040055266581475735, + 0.0036953056696802378, + 0.0024406055454164743, + 0.0030257555190473795, + 0.004675411153584719, + 0.001977684907615185, + 0.004736983682960272, + 0.002360797021538019, + 0.0022529573179781437, + 0.004151416942477226, + 0.005147193092852831, + 0.0021158268209546804, + 0.0024804265704005957, + 0.002868673065677285, + 0.0038118246011435986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.04079494625329971 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761876702308655 + ], + "opposite_logprob_tokens": [ + -0.8261876702308655 + ], + "logprob_reward": -0.5761876702308655, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761876702308655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05433201044797897, + 0.03296518698334694, + 0.05211992934346199, + 0.6038476824760437, + 0.01768425479531288, + 0.027392664924263954, + 0.03737383708357811, + 0.04086253419518471, + 0.046210434287786484, + 0.08721140027046204 + ], + "credit_map": [ + 0.0061832633800804615, + 0.012305510230362415, + 0.006438057869672775, + 0.004451090935617685, + 0.005125315394252539, + 0.0071057602763175964, + 0.006363323889672756, + 0.006359694059938192, + 0.002314917743206024, + 0.007004901301115751, + 0.004010926466435194, + 0.0013059501070529222, + 0.010214696638286114, + 0.0036684575024992228, + 0.001731844968162477, + 0.0027134940028190613, + 0.002059805439785123, + 0.021709255874156952, + 0.00171088392380625, + 0.0017368268454447389, + 0.015602692030370235, + 0.004836350213736296, + 0.0016428661765530705, + 0.002821255475282669, + 0.03895954787731171, + 0.144364595413208, + 0.020799115300178528, + 0.01809142529964447, + 0.1427796632051468, + 0.1890980750322342, + 0.013943447731435299, + 0.03581187129020691, + 0.0012495183618739247, + 0.004401420243084431, + 0.001699339714832604, + 0.0015781630063429475, + 0.0036357229109853506, + 0.0017610200447961688, + 0.0014125107554718852, + 0.0019465633668005466, + 0.002243868773803115, + 0.008018944412469864, + 0.00270592188462615, + 0.0016733892261981964, + 0.005242529790848494, + 0.0036960358265787363, + 0.001654100720770657, + 0.002157878363505006, + 0.003072039457038045, + 0.005459991749376059, + 0.0057357135228812695, + 0.004572800826281309, + 0.0035881358198821545, + 0.00684748450294137, + 0.0031152134761214256, + 0.00498245982453227, + 0.002850649179890752, + 0.0058432347141206264, + 0.004540208727121353, + 0.005956581328064203, + 0.006105290725827217, + 0.005539408419281244, + 0.004091568291187286, + 0.005935597699135542, + 0.006775131449103355, + 0.005622405093163252, + 0.0071821389719843864, + 0.005754021927714348, + 0.003906669095158577, + 0.005569701548665762, + 0.004251954611390829, + 0.007148414384573698, + 0.009137569926679134, + 0.009084007702767849, + 0.014800113625824451, + 0.021293768659234047, + 0.008936712518334389, + 0.008205374702811241, + 0.008268431760370731, + 0.007485433481633663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.5761876702308655 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.898271918296814 + ], + "opposite_logprob_tokens": [ + -0.523271918296814 + ], + "logprob_reward": -0.898271918296814, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.898271918296814, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08147017657756805, + 0.10137208551168442, + 0.03609670326113701, + 0.03700525686144829, + 0.07700813561677933, + 0.14872339367866516, + 0.08359289169311523, + 0.20260171592235565, + 0.12178690731525421, + 0.11034274101257324 + ], + "credit_map": [ + 0.010240153409540653, + 0.01136108674108982, + 0.010379497893154621, + 0.008586019277572632, + 0.00975350383669138, + 0.008987420238554478, + 0.009837961755692959, + 0.012324533425271511, + 0.006636928766965866, + 0.015898363664746284, + 0.008276968263089657, + 0.006407005246728659, + 0.029839474707841873, + 0.019265318289399147, + 0.006189424078911543, + 0.00885859876871109, + 0.001493128133006394, + 0.013839103281497955, + 0.00166799605358392, + 0.0016748695634305477, + 0.010214663110673428, + 0.003988086245954037, + 0.0018789160531014204, + 0.001339938840828836, + 0.0030303008388727903, + 0.00875373650342226, + 0.002755458001047373, + 0.0024521241430193186, + 0.005677569191902876, + 0.00831117108464241, + 0.0028363235760480165, + 0.003188576316460967, + 0.003935839049518108, + 0.032223064452409744, + 0.003748557763174176, + 0.0024486035108566284, + 0.019828028976917267, + 0.01007223129272461, + 0.0019285900052636862, + 0.0028232168406248093, + 0.005672638304531574, + 0.056333027780056, + 0.0063275559805333614, + 0.004835471976548433, + 0.053779516369104385, + 0.010258794762194157, + 0.004621349740773439, + 0.006895050406455994, + 0.006462431978434324, + 0.01168737094849348, + 0.014755785465240479, + 0.010073446668684483, + 0.008063767105340958, + 0.011892232112586498, + 0.009034700691699982, + 0.011623161844909191, + 0.01564949005842209, + 0.026182830333709717, + 0.019098859280347824, + 0.025704285129904747, + 0.019126741215586662, + 0.02912161499261856, + 0.021400071680545807, + 0.04631783440709114, + 0.021602416411042213, + 0.015920141711831093, + 0.01640569046139717, + 0.012762627564370632, + 0.014049867168068886, + 0.013797463849186897, + 0.013268690556287766, + 0.013980007730424404, + 0.011280999518930912, + 0.010533425956964493, + 0.015111139975488186, + 0.013965412974357605, + 0.013152862899005413, + 0.015694236382842064, + 0.012800280004739761, + 0.017804374918341637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.898271918296814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000719645875506103 + ], + "opposite_logprob_tokens": [ + -7.2507195472717285 + ], + "logprob_reward": -0.000719645875506103, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.000719645875506103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11455019563436508, + 0.09094953536987305, + 0.055533118546009064, + 0.15260019898414612, + 0.09514760226011276, + 0.08995230495929718, + 0.05451899394392967, + 0.1797797530889511, + 0.11214097589254379, + 0.05482733994722366 + ], + "credit_map": [ + 0.012835369445383549, + 0.017992330715060234, + 0.012504358775913715, + 0.010804926976561546, + 0.013716294430196285, + 0.014762399718165398, + 0.013669654726982117, + 0.018264856189489365, + 0.0056087723933160305, + 0.014848683029413223, + 0.006925055757164955, + 0.005643283948302269, + 0.028249455615878105, + 0.016386711969971657, + 0.005640754941850901, + 0.007646822836250067, + 0.0022250402253121138, + 0.02271387353539467, + 0.0022002658806741238, + 0.0019161388045176864, + 0.01713424175977707, + 0.0058279577642679214, + 0.0018885075114667416, + 0.0016270935302600265, + 0.00961729884147644, + 0.03430194407701492, + 0.006540993694216013, + 0.005227185320109129, + 0.03579272702336311, + 0.04744800925254822, + 0.004401016514748335, + 0.009271013550460339, + 0.0046685729175806046, + 0.04084087535738945, + 0.004168962128460407, + 0.0024450579658150673, + 0.02479691430926323, + 0.013058707118034363, + 0.0019610777962952852, + 0.0032074253540486097, + 0.0034969986882060766, + 0.03301583603024483, + 0.00427815830335021, + 0.003382506314665079, + 0.03209823742508888, + 0.006423548795282841, + 0.0035607044119387865, + 0.003696318017318845, + 0.004484825301915407, + 0.007390095852315426, + 0.009037064388394356, + 0.006206220015883446, + 0.005721376743167639, + 0.006186929531395435, + 0.006443413905799389, + 0.00904906541109085, + 0.014118527993559837, + 0.024475613608956337, + 0.015341066755354404, + 0.021426130086183548, + 0.018756676465272903, + 0.026614727452397346, + 0.019041715189814568, + 0.04000530391931534, + 0.0202587079256773, + 0.013751255348324776, + 0.015430090948939323, + 0.01155148632824421, + 0.013778412714600563, + 0.014122921973466873, + 0.011009849607944489, + 0.012238250114023685, + 0.006193150766193867, + 0.005937305279076099, + 0.009147163480520248, + 0.01155428308993578, + 0.004447729326784611, + 0.007098546717315912, + 0.004711896646767855, + 0.005737267434597015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.000719645875506103 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036131355445832014 + ], + "opposite_logprob_tokens": [ + -5.628612995147705 + ], + "logprob_reward": -0.0036131355445832014, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036131355445832014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3310258090496063, + 0.018897779285907745, + 0.022802291437983513, + 0.44285252690315247, + 0.013998620212078094, + 0.04012772813439369, + 0.024882575497031212, + 0.05538227781653404, + 0.02079983986914158, + 0.02923055738210678 + ], + "credit_map": [ + 0.03921087831258774, + 0.04958510771393776, + 0.04217782989144325, + 0.03337525576353073, + 0.04401528462767601, + 0.03664446622133255, + 0.032956503331661224, + 0.05306047573685646, + 0.0018215356394648552, + 0.002998366951942444, + 0.0027392779011279345, + 0.0014792598085477948, + 0.0033652184065431356, + 0.0029213111847639084, + 0.001383390394039452, + 0.0021894180681556463, + 0.0011356777977198362, + 0.008963453583419323, + 0.0010302166920155287, + 0.0012451977236196399, + 0.005842151120305061, + 0.002427763305604458, + 0.0008076127851381898, + 0.0013502168003469706, + 0.02801675535738468, + 0.12548251450061798, + 0.017345314845442772, + 0.0151747427880764, + 0.06640904396772385, + 0.14532777667045593, + 0.010182308033108711, + 0.0349140390753746, + 0.001476776204071939, + 0.002258376916870475, + 0.0019477450987324119, + 0.0014736668672412634, + 0.0017715145368129015, + 0.002174341818317771, + 0.0010468028485774994, + 0.0018493946408852935, + 0.0014286600053310394, + 0.01622822694480419, + 0.001171091222204268, + 0.0010033990256488323, + 0.016037726774811745, + 0.0022343576420098543, + 0.0009603145299479365, + 0.0010639481479302049, + 0.0019891890697181225, + 0.003636703360825777, + 0.0048212455585598946, + 0.002994279144331813, + 0.0023139207623898983, + 0.0040146238170564175, + 0.0024244494270533323, + 0.0026881617959588766, + 0.00451823603361845, + 0.00885673239827156, + 0.004225192591547966, + 0.006040552631020546, + 0.005846506915986538, + 0.009079517796635628, + 0.005880000535398722, + 0.010935533791780472, + 0.0032369752880185843, + 0.0032533854246139526, + 0.0025532818399369717, + 0.002048708964139223, + 0.002512019593268633, + 0.0026991930790245533, + 0.0020838286727666855, + 0.0024124449118971825, + 0.0030046922620385885, + 0.0028242282569408417, + 0.004498339723795652, + 0.00545772397890687, + 0.0033830825705081224, + 0.0033731884323060513, + 0.002911446150392294, + 0.0037778548430651426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.0036131355445832014 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a military archival-style shot with a signature vintage film texture, featuring slight color desaturation and noticeable grain, the camera pans in a fluid, unbroken 360-degree pan, closely following a missile hurtling through the sky at high speed. The missile's sleek gray-white metal hull and tail stabilizers reflect a cool, austere luster in the harsh sunlight. A dense, swirling column of white smoke erupts from its tail, leaving a striking, persistent trail across the vast pale blue sky. As the camera rotates at a constant speed, the viewer's perspective shifts around the missile, while the horizon and sparse clouds in the background rotate accordingly, creating a powerful sense of motion and an immersive sense of speed. In the final stage of the surround shooting, the camera angle is exactly aimed at the direction of the missile's advance, capturing it rushing towards a distant, vaguely outlined aerial target with unstoppable momentum and precision. Finally, at the decisive moment when the missile contacts the target, the picture is occupied by a violent and silent explosion. A dazzling orange-red fireball suddenly bursts out, and then turns into a black smoke mixed with metal fragments, slowly dissipating in the air. The whole process is full of calm and shocking destructive aesthetics.", + "seed": 25733, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_12/index_9232_seed_25733_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.5049558146158233, + "vlm_logprob": -0.5049558146158233, + "vlm_yes_no_margin": 1.8, + "vlm_token_credit": -0.5049558146158233, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18529541790485382 + ], + "opposite_logprob_tokens": [ + -1.8102954626083374 + ], + "logprob_reward": -0.18529541790485382, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18529541790485382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05026033893227577, + 0.045335181057453156, + 0.02032439410686493, + 0.031958311796188354, + 0.10488316416740417, + 0.30931222438812256, + 0.1195589005947113, + 0.06266804784536362, + 0.19843456149101257, + 0.0572647750377655 + ], + "credit_map": [ + 0.0031336997635662556, + 0.004960390739142895, + 0.00479912431910634, + 0.006313679739832878, + 0.009000735357403755, + 0.01188393123447895, + 0.004444892983883619, + 0.005723893642425537, + 0.002629461931064725, + 0.004408626351505518, + 0.004190865438431501, + 0.006552938371896744, + 0.008758957497775555, + 0.007980260998010635, + 0.005627406295388937, + 0.005186674650758505, + 0.0019445763900876045, + 0.0024437292013317347, + 0.0026289215311408043, + 0.0026150622870773077, + 0.003307206090539694, + 0.0031015134882181883, + 0.002848016796633601, + 0.0014353729784488678, + 0.002176113659515977, + 0.003664021147415042, + 0.00382172130048275, + 0.004311456345021725, + 0.004550847224891186, + 0.006785696372389793, + 0.003455311758443713, + 0.0031931488774716854, + 0.00876312144100666, + 0.010965627618134022, + 0.015585734508931637, + 0.019923312589526176, + 0.014567933976650238, + 0.011952648870646954, + 0.013740015216171741, + 0.009384791366755962, + 0.039592597633600235, + 0.01644541136920452, + 0.12815873324871063, + 0.024324502795934677, + 0.007381600793451071, + 0.030582370236516, + 0.02483726106584072, + 0.03798979893326759, + 0.009012396447360516, + 0.0179377943277359, + 0.023254361003637314, + 0.014018770307302475, + 0.009790593758225441, + 0.010442332364618778, + 0.013817362487316132, + 0.02128530852496624, + 0.005878228694200516, + 0.007938303984701633, + 0.006765015423297882, + 0.006171656772494316, + 0.016630569472908974, + 0.011503891088068485, + 0.0040808492340147495, + 0.0036995387636125088, + 0.00898507609963417, + 0.0723593533039093, + 0.051614511758089066, + 0.013323859311640263, + 0.026743384078145027, + 0.011899608187377453, + 0.007546871434897184, + 0.005961948074400425, + 0.0038582405541092157, + 0.006700059864670038, + 0.006955488119274378, + 0.008362365886569023, + 0.012092889286577702, + 0.00523426104336977, + 0.009336503222584724, + 0.004724977072328329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.18529541790485382 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159731805324554 + ], + "opposite_logprob_tokens": [ + -1.7015973329544067 + ], + "logprob_reward": -0.20159731805324554, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159731805324554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1333332359790802, + 0.027055159211158752, + 0.01572210155427456, + 0.03495117649435997, + 0.10188937187194824, + 0.2986160218715668, + 0.05219209939241409, + 0.10081576555967331, + 0.19095180928707123, + 0.04447324201464653 + ], + "credit_map": [ + 0.0141634251922369, + 0.01862061209976673, + 0.01640085130929947, + 0.020023448392748833, + 0.016833266243338585, + 0.017043957486748695, + 0.016329728066921234, + 0.013917955569922924, + 0.0023946459405124187, + 0.002241549314931035, + 0.0032027973793447018, + 0.004543888848274946, + 0.0037656391505151987, + 0.0046484884805977345, + 0.0030032391659915447, + 0.003254910232499242, + 0.001345211872830987, + 0.0016005306970328093, + 0.0024507648777216673, + 0.0021632236894220114, + 0.0025102407671511173, + 0.00220519513823092, + 0.0016789620276540518, + 0.0017679709708318114, + 0.0032543279230594635, + 0.003932022023946047, + 0.004330561030656099, + 0.007400714326649904, + 0.004427158273756504, + 0.004986123647540808, + 0.003972009755671024, + 0.0026482578832656145, + 0.007946250960230827, + 0.0068532926961779594, + 0.028384381905198097, + 0.016657231375575066, + 0.007626510690897703, + 0.015617200173437595, + 0.0076119862496852875, + 0.01119251735508442, + 0.02913511171936989, + 0.017070485278964043, + 0.11993242055177689, + 0.022392522543668747, + 0.009542074985802174, + 0.04746833071112633, + 0.021538378670811653, + 0.031536705791950226, + 0.0037233633920550346, + 0.007794724777340889, + 0.010981163010001183, + 0.006041830871254206, + 0.003990808967500925, + 0.0045891148038208485, + 0.005549374967813492, + 0.009521719068288803, + 0.005846373736858368, + 0.027954740449786186, + 0.0028856557328253984, + 0.0030721721705049276, + 0.04356611520051956, + 0.010839028283953667, + 0.004010715521872044, + 0.002640962367877364, + 0.008174835704267025, + 0.06944409012794495, + 0.056051019579172134, + 0.012509450316429138, + 0.023058148100972176, + 0.011037053540349007, + 0.005194526631385088, + 0.005482680629938841, + 0.003826020285487175, + 0.006271700374782085, + 0.005057207774370909, + 0.005847650114446878, + 0.006369701586663723, + 0.005505604203790426, + 0.0043973131105303764, + 0.007198045961558819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.20159731805324554 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9102938175201416 + ], + "opposite_logprob_tokens": [ + -0.16029387712478638 + ], + "logprob_reward": -1.9102938175201416, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9102938175201416, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04881930723786354, + 0.036532338708639145, + 0.0852474570274353, + 0.046803027391433716, + 0.121792733669281, + 0.12321778386831284, + 0.059036169201135635, + 0.3409937024116516, + 0.10068324953317642, + 0.0368742011487484 + ], + "credit_map": [ + 0.004242117051035166, + 0.004856874700635672, + 0.007220477797091007, + 0.005884242244064808, + 0.009977761656045914, + 0.005758034531027079, + 0.006071930751204491, + 0.004807867109775543, + 0.0027475582901388407, + 0.0033943592570722103, + 0.0041155279614031315, + 0.005704354960471392, + 0.00577090959995985, + 0.005632832180708647, + 0.004976234398782253, + 0.00419055949896574, + 0.005047002341598272, + 0.010845614597201347, + 0.024054938927292824, + 0.010906034149229527, + 0.009426700882613659, + 0.011329418048262596, + 0.006912535056471825, + 0.006725209299474955, + 0.0026430359575897455, + 0.006055979523807764, + 0.005383979994803667, + 0.005823331885039806, + 0.006652077194303274, + 0.009683550335466862, + 0.005128106102347374, + 0.005432965699583292, + 0.010600241832435131, + 0.009034974500536919, + 0.030665691941976547, + 0.020113136619329453, + 0.008592682890594006, + 0.019154328852891922, + 0.008835786022245884, + 0.014795894734561443, + 0.009115826338529587, + 0.009989175945520401, + 0.036920201033353806, + 0.012598872184753418, + 0.009013419970870018, + 0.02028966322541237, + 0.015980172902345657, + 0.009310459718108177, + 0.005625173915177584, + 0.0070274509489536285, + 0.012203102000057697, + 0.006951313465833664, + 0.0055511887185275555, + 0.004773072432726622, + 0.007698857691138983, + 0.009206012822687626, + 0.026544297114014626, + 0.0701860710978508, + 0.015765011310577393, + 0.015990126878023148, + 0.13628284633159637, + 0.04816707596182823, + 0.016965677961707115, + 0.01109260693192482, + 0.0055203489027917385, + 0.02813642844557762, + 0.018978385254740715, + 0.010962574742734432, + 0.013150230050086975, + 0.009059354662895203, + 0.008500450290739536, + 0.006375471595674753, + 0.001924248063005507, + 0.0036266171373426914, + 0.005038448143750429, + 0.006393367424607277, + 0.004584787413477898, + 0.004389866255223751, + 0.004746291320770979, + 0.006170572247356176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -1.9102938175201416 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22565105557441711 + ], + "opposite_logprob_tokens": [ + -1.6006510257720947 + ], + "logprob_reward": -0.22565105557441711, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22565105557441711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03155495971441269, + 0.06619173288345337, + 0.08492593467235565, + 0.08998813480138779, + 0.13702607154846191, + 0.15320543944835663, + 0.024336282163858414, + 0.11286501586437225, + 0.19275343418121338, + 0.10715300589799881 + ], + "credit_map": [ + 0.0034726103767752647, + 0.0030358019284904003, + 0.0033604942727833986, + 0.003996442537754774, + 0.005081694573163986, + 0.005730613134801388, + 0.003752400865778327, + 0.0031249045860022306, + 0.004651554860174656, + 0.005606122314929962, + 0.007301275618374348, + 0.011038723401725292, + 0.00887372437864542, + 0.009877932257950306, + 0.00939808040857315, + 0.009444324299693108, + 0.004156117793172598, + 0.00881611742079258, + 0.03082859143614769, + 0.012871012091636658, + 0.006508558988571167, + 0.007114696782082319, + 0.008229619823396206, + 0.006401218939572573, + 0.005193847697228193, + 0.009604867547750473, + 0.010402453131973743, + 0.01305041927844286, + 0.01129210740327835, + 0.01832793653011322, + 0.009653463028371334, + 0.012463041581213474, + 0.010390839539468288, + 0.01266341283917427, + 0.030590660870075226, + 0.0230441614985466, + 0.015095604583621025, + 0.021680263802409172, + 0.011976286768913269, + 0.011584838852286339, + 0.020014408975839615, + 0.008311714977025986, + 0.06277337670326233, + 0.012532567605376244, + 0.004243175499141216, + 0.014048985205590725, + 0.01229736115783453, + 0.01898385025560856, + 0.001744015607982874, + 0.003727199509739876, + 0.004199886694550514, + 0.0030412001069635153, + 0.001983585534617305, + 0.0022911548148840666, + 0.003134138183668256, + 0.004215099848806858, + 0.013391785323619843, + 0.014443770982325077, + 0.009136470966041088, + 0.00861662533134222, + 0.033595964312553406, + 0.023119498044252396, + 0.005407680757343769, + 0.005153229460120201, + 0.007844367995858192, + 0.06137719377875328, + 0.05263269320130348, + 0.015336459502577782, + 0.02041507326066494, + 0.01362621784210205, + 0.01189281977713108, + 0.009628614410758018, + 0.008055740036070347, + 0.011168358847498894, + 0.012085877358913422, + 0.01156712882220745, + 0.024929344654083252, + 0.010180596262216568, + 0.01863321289420128, + 0.010532744228839874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.22565105557441711 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019414640264585614 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019414640264585614, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019414640264585614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07191352546215057, + 0.025186706334352493, + 0.07415656745433807, + 0.05320509895682335, + 0.09217504411935806, + 0.20718391239643097, + 0.0459042526781559, + 0.18547633290290833, + 0.19329415261745453, + 0.05150440335273743 + ], + "credit_map": [ + 0.005791184492409229, + 0.008217760361731052, + 0.009264834225177765, + 0.010235191322863102, + 0.014075236395001411, + 0.010031264275312424, + 0.006701608654111624, + 0.007596448063850403, + 0.0023275339044630527, + 0.002641692291945219, + 0.0036626928485929966, + 0.003669669618830085, + 0.003601104486733675, + 0.003485281951725483, + 0.003402448957785964, + 0.0023962820414453745, + 0.004445566795766354, + 0.008145525120198727, + 0.02038572169840336, + 0.009759069420397282, + 0.008573963306844234, + 0.009582067839801311, + 0.007173092104494572, + 0.006091562565416098, + 0.0034356543328613043, + 0.006272023543715477, + 0.006077013444155455, + 0.008735321462154388, + 0.007481090724468231, + 0.00949095468968153, + 0.006171572953462601, + 0.005541463848203421, + 0.009450648911297321, + 0.007050406653434038, + 0.021017752587795258, + 0.014338013716042042, + 0.006854361854493618, + 0.015152120031416416, + 0.007223806343972683, + 0.011087940074503422, + 0.02438345178961754, + 0.011552618816494942, + 0.08111140131950378, + 0.01662435568869114, + 0.007842836901545525, + 0.02323990687727928, + 0.019317585974931717, + 0.023111743852496147, + 0.0032914974726736546, + 0.005385170690715313, + 0.009115378372371197, + 0.005725809372961521, + 0.004750929307192564, + 0.004758207593113184, + 0.005731259472668171, + 0.007146000396460295, + 0.014363909140229225, + 0.03522893041372299, + 0.01005660742521286, + 0.010228514671325684, + 0.06890697777271271, + 0.02676231972873211, + 0.01210545189678669, + 0.00782362837344408, + 0.009334547445178032, + 0.0689350813627243, + 0.04743504896759987, + 0.01400731597095728, + 0.026376858353614807, + 0.01163692120462656, + 0.007893136702477932, + 0.007675259839743376, + 0.0037096883170306683, + 0.005731792654842138, + 0.006429164670407772, + 0.009459376335144043, + 0.00620156479999423, + 0.005817343946546316, + 0.00837338250130415, + 0.005782091058790684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.0019414640264585614 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a military archival-style shot with a signature vintage film texture, featuring slight color desaturation and noticeable grain, the camera pans in a fluid, unbroken 360-degree pan, closely following a missile hurtling through the sky at high speed. The missile's sleek gray-white metal hull and tail stabilizers reflect a cool, austere luster in the harsh sunlight. A dense, swirling column of white smoke erupts from its tail, leaving a striking, persistent trail across the vast pale blue sky. As the camera rotates at a constant speed, the viewer's perspective shifts around the missile, while the horizon and sparse clouds in the background rotate accordingly, creating a powerful sense of motion and an immersive sense of speed. In the final stage of the surround shooting, the camera angle is exactly aimed at the direction of the missile's advance, capturing it rushing towards a distant, vaguely outlined aerial target with unstoppable momentum and precision. Finally, at the decisive moment when the missile contacts the target, the picture is occupied by a violent and silent explosion. A dazzling orange-red fireball suddenly bursts out, and then turns into a black smoke mixed with metal fragments, slowly dissipating in the air. The whole process is full of calm and shocking destructive aesthetics.", + "seed": 25734, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_12/index_9232_seed_25734_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.44212328367866577, + "vlm_logprob": -0.44212328367866577, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.44212328367866577, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022798335179686546 + ], + "opposite_logprob_tokens": [ + -4.147798538208008 + ], + "logprob_reward": -0.022798335179686546, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.022798335179686546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2569500207901001, + 0.13100340962409973, + 0.02446417696774006, + 0.12586529552936554, + 0.03471403568983078, + 0.028282752260565758, + 0.2279813289642334, + 0.07561859488487244, + 0.034794166684150696, + 0.06032619625329971 + ], + "credit_map": [ + 0.01057707890868187, + 0.08000271767377853, + 0.0956866443157196, + 0.022943789139389992, + 0.0209264624863863, + 0.011588405817747116, + 0.005778418388217688, + 0.009446509182453156, + 0.0056920419447124004, + 0.013188951648771763, + 0.023790881037712097, + 0.008577550761401653, + 0.014692221768200397, + 0.04454382508993149, + 0.009391025640070438, + 0.011126895435154438, + 0.001195609220303595, + 0.004273163620382547, + 0.004110176581889391, + 0.0018567658262327313, + 0.001865254482254386, + 0.006566921249032021, + 0.001667492906562984, + 0.0029287906363606453, + 0.005251357331871986, + 0.005738987121731043, + 0.03977468982338905, + 0.007081549149006605, + 0.008332286961376667, + 0.039860162883996964, + 0.007696041837334633, + 0.012130217626690865, + 0.003176072146743536, + 0.003596503986045718, + 0.004928475711494684, + 0.00272228941321373, + 0.005875828675925732, + 0.005263646133244038, + 0.00524861691519618, + 0.0039025985170155764, + 0.0038439687341451645, + 0.0035492144525051117, + 0.0036859738174825907, + 0.003249207278713584, + 0.004975302144885063, + 0.002771812491118908, + 0.003835198236629367, + 0.0023720732424408197, + 0.023412629961967468, + 0.02181471884250641, + 0.04458135738968849, + 0.06628037989139557, + 0.012571467086672783, + 0.007299428340047598, + 0.035115208476781845, + 0.01690613478422165, + 0.012582391500473022, + 0.014980845153331757, + 0.012185119092464447, + 0.010086050257086754, + 0.005198261234909296, + 0.006753827445209026, + 0.005477588158100843, + 0.008354502730071545, + 0.0026518451049923897, + 0.0039908261969685555, + 0.005203121807426214, + 0.004641164559870958, + 0.0036147614009678364, + 0.005864082835614681, + 0.004752421285957098, + 0.004075942095369101, + 0.005202638451009989, + 0.004964090883731842, + 0.009119099006056786, + 0.01329902932047844, + 0.0057236128486692905, + 0.010138367302715778, + 0.00455109030008316, + 0.007328262086957693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.022798335179686546 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761808156967163 + ], + "opposite_logprob_tokens": [ + -0.8261808156967163 + ], + "logprob_reward": -0.5761808156967163, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761808156967163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.821437418460846, + 0.021932102739810944, + 0.008077088743448257, + 0.007580725010484457, + 0.006070001050829887, + 0.02369103394448757, + 0.036471687257289886, + 0.04130417853593826, + 0.02541644126176834, + 0.00801923032850027 + ], + "credit_map": [ + 0.030682742595672607, + 0.27589550614356995, + 0.29210126399993896, + 0.06889129430055618, + 0.07683129608631134, + 0.03576171398162842, + 0.014521640725433826, + 0.026751985773444176, + 0.0010354547994211316, + 0.001874698675237596, + 0.003719308879226446, + 0.0023357803001999855, + 0.0022704319562762976, + 0.004050540272146463, + 0.0032327494118362665, + 0.0034131400752812624, + 0.0006617699982598424, + 0.0015766393626108766, + 0.001258535892702639, + 0.000761220813728869, + 0.0006740984390489757, + 0.0016982938395813107, + 0.0005910260952077806, + 0.0008555050590075552, + 0.00046518805902451277, + 0.0006535938591696322, + 0.0018748173024505377, + 0.0008176153642125428, + 0.000720798154361546, + 0.001302001066505909, + 0.0007956058252602816, + 0.0009511055541224778, + 0.0006673629395663738, + 0.0006741822580806911, + 0.0007605516002513468, + 0.0008604598697274923, + 0.0008299417095258832, + 0.0006639099447056651, + 0.0007152567850425839, + 0.0008983358275145292, + 0.002625476336106658, + 0.0020446910057216883, + 0.004783911630511284, + 0.005919057875871658, + 0.0021079732105135918, + 0.0019259982509538531, + 0.0017146776663139462, + 0.0025692509952932596, + 0.0035586131270974874, + 0.0032001386862248182, + 0.006985727231949568, + 0.010098041966557503, + 0.002170354127883911, + 0.0011422957759350538, + 0.006295157130807638, + 0.0030213608406484127, + 0.007594299502670765, + 0.008965224027633667, + 0.008985547348856926, + 0.004663459490984678, + 0.00268831430003047, + 0.002709538908675313, + 0.00255759060382843, + 0.003140208078548312, + 0.001995811704546213, + 0.002884780755266547, + 0.0040453956462442875, + 0.0030249496921896935, + 0.001992352306842804, + 0.003296169452369213, + 0.0025566236581653357, + 0.005620358511805534, + 0.0008925828151404858, + 0.0006617244216613472, + 0.001090112142264843, + 0.0011875901836901903, + 0.0006956573924981058, + 0.0012023932067677379, + 0.0007866233354434371, + 0.0015025469474494457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.5761808156967163 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6004987955093384 + ], + "opposite_logprob_tokens": [ + -0.22549882531166077 + ], + "logprob_reward": -1.6004987955093384, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6004987955093384, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.47017669677734375, + 0.05506521463394165, + 0.01476197224110365, + 0.09674974530935287, + 0.03880967199802399, + 0.023425674065947533, + 0.10314792394638062, + 0.13707295060157776, + 0.02260243333876133, + 0.03818774223327637 + ], + "credit_map": [ + 0.01703503541648388, + 0.15944182872772217, + 0.16033779084682465, + 0.03866608813405037, + 0.04803142324090004, + 0.02174941822886467, + 0.009103038348257542, + 0.01581208035349846, + 0.0025019149761646986, + 0.005717006977647543, + 0.009307092987000942, + 0.0038026082329452038, + 0.006041947286576033, + 0.017689647153019905, + 0.004452353343367577, + 0.005552641581743956, + 0.0011072003981098533, + 0.0014928476884961128, + 0.002757295034825802, + 0.0011411424493417144, + 0.0014217508723959327, + 0.0032662178855389357, + 0.0012088919756934047, + 0.002366625238209963, + 0.004095206502825022, + 0.004017492290586233, + 0.035652339458465576, + 0.004261836409568787, + 0.004827914293855429, + 0.03526605665683746, + 0.0033193742856383324, + 0.005309522617608309, + 0.0037763856817036867, + 0.004138828720897436, + 0.005612278822809458, + 0.0027606866788119078, + 0.006169136147946119, + 0.004799157381057739, + 0.006140068173408508, + 0.00541312899440527, + 0.002388602588325739, + 0.002686787862330675, + 0.00399421202018857, + 0.003215618897229433, + 0.003432493656873703, + 0.0036239682231098413, + 0.00140468031167984, + 0.0026793116703629494, + 0.011006313376128674, + 0.008509665727615356, + 0.019870953634381294, + 0.030126577243208885, + 0.005697854794561863, + 0.0035877740010619164, + 0.017259828746318817, + 0.0070889596827328205, + 0.028837233781814575, + 0.03554997965693474, + 0.024810707196593285, + 0.01329852920025587, + 0.00942150130867958, + 0.007931067608296871, + 0.007535723969340324, + 0.00968820508569479, + 0.0025999511126428843, + 0.0023414327297359705, + 0.0033313464373350143, + 0.0025299852713942528, + 0.0018216363387182355, + 0.0036805588752031326, + 0.0017838914645835757, + 0.004513631109148264, + 0.002628024434670806, + 0.0031651209574192762, + 0.005762683693319559, + 0.00985147338360548, + 0.0038351300172507763, + 0.003948155324906111, + 0.003677397035062313, + 0.005319755524396896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -1.6004987955093384 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649858646094799 + ], + "opposite_logprob_tokens": [ + -4.758649826049805 + ], + "logprob_reward": -0.008649858646094799, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649858646094799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3668684959411621, + 0.010694990865886211, + 0.009628157131373882, + 0.03315959498286247, + 0.016469288617372513, + 0.02489636279642582, + 0.14552965760231018, + 0.09037915617227554, + 0.03132754936814308, + 0.2710467278957367 + ], + "credit_map": [ + 0.013721026480197906, + 0.12023935467004776, + 0.13152582943439484, + 0.03137015178799629, + 0.03163653612136841, + 0.01669929176568985, + 0.008823649026453495, + 0.012852690182626247, + 0.0007200175314210355, + 0.001769459224306047, + 0.0013348915381357074, + 0.0012657602783292532, + 0.0011789571726694703, + 0.00234623602591455, + 0.0007618244853802025, + 0.0013178460067138076, + 0.0006344284047372639, + 0.0010833467822521925, + 0.002287063980475068, + 0.0009723725961521268, + 0.0011011611204594374, + 0.0014710476389154792, + 0.0009485863265581429, + 0.0011301509803161025, + 0.0018987585790455341, + 0.002446686150506139, + 0.010846844874322414, + 0.001957469619810581, + 0.0027172896079719067, + 0.008776321075856686, + 0.002360104350373149, + 0.002156120026484132, + 0.0015767946606501937, + 0.001550624961964786, + 0.0024096572306007147, + 0.0019227404845878482, + 0.0019085075473412871, + 0.0024974013213068247, + 0.0021272096782922745, + 0.0024763555265963078, + 0.003196616191416979, + 0.0026733921840786934, + 0.003511847695335746, + 0.005880223121494055, + 0.002849012613296509, + 0.002307622227817774, + 0.002058793790638447, + 0.002418854273855686, + 0.015299861319363117, + 0.01374264620244503, + 0.028456179425120354, + 0.04239463433623314, + 0.007450365461409092, + 0.004668591544032097, + 0.02318110316991806, + 0.01033627800643444, + 0.011961414478719234, + 0.015589247457683086, + 0.015994152054190636, + 0.011041071265935898, + 0.007953404448926449, + 0.009625875391066074, + 0.007799353450536728, + 0.010414643213152885, + 0.002590279560536146, + 0.0034540107008069754, + 0.005075582768768072, + 0.0046799927949905396, + 0.0031117924954742193, + 0.004995861556380987, + 0.0034149654675275087, + 0.004005066119134426, + 0.02422919124364853, + 0.02414071187376976, + 0.03482462465763092, + 0.05744235962629318, + 0.025511354207992554, + 0.05223606154322624, + 0.022929424419999123, + 0.02973303012549877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.008649858646094799 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a missile with a sleek gray-white metal hull and tail stabilizers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video shot with a vintage film texture, slight color desaturation, and noticeable grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan around the missile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video end with a violent explosion producing an orange-red fireball that turns into black smoke and metal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pale blue sky with sparse clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002488613361492753 + ], + "opposite_logprob_tokens": [ + -6.002488613128662 + ], + "logprob_reward": -0.002488613361492753, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002488613361492753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7489291429519653, + 0.039933476597070694, + 0.013361778110265732, + 0.056497201323509216, + 0.009406461380422115, + 0.029108690097928047, + 0.03234274685382843, + 0.032452575862407684, + 0.014822795055806637, + 0.02314516343176365 + ], + "credit_map": [ + 0.027280034497380257, + 0.25614601373672485, + 0.2562197148799896, + 0.06139933690428734, + 0.07801967859268188, + 0.0334913432598114, + 0.012955766171216965, + 0.023417238146066666, + 0.001647073426283896, + 0.0027261630166321993, + 0.010283557698130608, + 0.003029882675036788, + 0.004007694311439991, + 0.012634417973458767, + 0.0023866533301770687, + 0.0032180321868509054, + 0.0009203817462548614, + 0.0018430696800351143, + 0.0030710850842297077, + 0.001127162599004805, + 0.0011771220015361905, + 0.0034079281613230705, + 0.0009805732406675816, + 0.0008344544912688434, + 0.0024204771034419537, + 0.0020471359603106976, + 0.02210601046681404, + 0.002355801174417138, + 0.0024447643663734198, + 0.020949039608240128, + 0.0019061296479776502, + 0.0022678328678011894, + 0.0009992476552724838, + 0.0008363996748812497, + 0.001335799228399992, + 0.0007566105341538787, + 0.001338853850029409, + 0.0013884168583899736, + 0.0012473156675696373, + 0.0015038170386105776, + 0.0030092098750174046, + 0.0018141272012144327, + 0.006109146401286125, + 0.007068254519253969, + 0.0027215455193072557, + 0.002503942232578993, + 0.002636019606143236, + 0.003246442647650838, + 0.003204016713425517, + 0.0030947658233344555, + 0.005985202733427286, + 0.008773230016231537, + 0.0019921709317713976, + 0.0011507581220939755, + 0.005531943868845701, + 0.0026106543373316526, + 0.00494402414187789, + 0.006203154101967812, + 0.006945588160306215, + 0.002639597514644265, + 0.0025922630447894335, + 0.002564245369285345, + 0.003393584629520774, + 0.003170114243403077, + 0.0013558213831856847, + 0.0015227015828713775, + 0.002099866745993495, + 0.0021161448676139116, + 0.0016000091563910246, + 0.002258926397189498, + 0.0020430306904017925, + 0.0018262919038534164, + 0.0015908993082121015, + 0.0019067485118284822, + 0.0031775818206369877, + 0.005934333894401789, + 0.0019212323240935802, + 0.003477102378383279, + 0.0021214077714830637, + 0.0030158564914017916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.002488613361492753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_13/index_tensor([3251, 3251])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_13/index_tensor([3251, 3251])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4889ebde4f2137985caa4a9c9912340e23df6691 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_13/index_tensor([3251, 3251])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with an intimate close-up, rendered in the warm, faded sepia tones of a realistic old photograph. A small ginger cat, its fur a soft down, clings tenaciously to the thick, cream-colored knit of its owner's sweater. With its large, crystalline blue eyes, the kitten gazes forward with innocent trust, its miniature pink nose twitching slightly, its paws gently hooking into the wool. Then, the camera slowly and steadily pulls back, a deliberate and gentle movement that gradually expands the frame. The image shows the kitten resting securely on a person's chest, their hands tenderly cradling it. As the camera continues to retreat, the soft-focus scene resolutely unveils an old-fashioned photography studio, bathed in soft, diffuse light from a large, invisible window that casts long, gentle shadows. In the background, an old bellows camera, made of polished dark wood with brass fittings, stands on a tall wooden tripod, its lens trained directly on the couple, ready to capture this moment. The entire scene is imbued with a quiet, nostalgic atmosphere, evoking a sense of timeless tenderness and captured memories.", + "seed": 30212, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_13/index_3251_seed_30212_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.016266017826274036, + "vlm_logprob": -0.016266017826274036, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.016266017826274036, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00777989299967885 + ], + "opposite_logprob_tokens": [ + -5.007780075073242 + ], + "logprob_reward": -0.00777989299967885, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00777989299967885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2439100295305252, + 0.08762147277593613, + 0.1697908341884613, + 0.10414908081293106, + 0.07256167382001877, + 0.056541673839092255, + 0.08559757471084595, + 0.0416836179792881, + 0.03531888127326965, + 0.10282520204782486 + ], + "credit_map": [ + 0.026207320392131805, + 0.01989196427166462, + 0.054578747600317, + 0.02395571768283844, + 0.022619623690843582, + 0.05129446089267731, + 0.020686481148004532, + 0.024675702676177025, + 0.011577178724110126, + 0.012985852546989918, + 0.0089142806828022, + 0.009352477267384529, + 0.008092581294476986, + 0.009445510804653168, + 0.013434394262731075, + 0.013819197192788124, + 0.012170681729912758, + 0.018155662342905998, + 0.016005437821149826, + 0.02564370073378086, + 0.029286295175552368, + 0.027483554556965828, + 0.023027727380394936, + 0.01801777444779873, + 0.005340130999684334, + 0.012130619026720524, + 0.011858945712447166, + 0.008506660349667072, + 0.01525900699198246, + 0.007992320694029331, + 0.022635899484157562, + 0.020425496622920036, + 0.007134472951292992, + 0.008149314671754837, + 0.012964452616870403, + 0.0078679658472538, + 0.009134845808148384, + 0.00649747159332037, + 0.011047938838601112, + 0.009765198454260826, + 0.004238475579768419, + 0.005609962623566389, + 0.013040184043347836, + 0.008651894517242908, + 0.006452362984418869, + 0.0035591134801506996, + 0.004471622407436371, + 0.010518054477870464, + 0.009341512806713581, + 0.01457787211984396, + 0.015023883432149887, + 0.014632709324359894, + 0.006546445190906525, + 0.00789947435259819, + 0.006557702552527189, + 0.011017970740795135, + 0.0035633521620184183, + 0.005216890946030617, + 0.0043304986320436, + 0.0041237869299948215, + 0.006537222303450108, + 0.004508120473474264, + 0.0072587826289236546, + 0.006144959479570389, + 0.0031754213850945234, + 0.004592376761138439, + 0.0062467013485729694, + 0.007163799833506346, + 0.004366917535662651, + 0.0032237295527011156, + 0.002965206280350685, + 0.0035847302060574293, + 0.006767574232071638, + 0.0047799646854400635, + 0.014263986609876156, + 0.008144227787852287, + 0.011409571394324303, + 0.02059171162545681, + 0.009883372113108635, + 0.02698478288948536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.00777989299967885 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003200054634362459 + ], + "opposite_logprob_tokens": [ + -5.753200054168701 + ], + "logprob_reward": -0.003200054634362459, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003200054634362459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37264829874038696, + 0.05597829073667526, + 0.0754890963435173, + 0.045805368572473526, + 0.03249196335673332, + 0.07431674003601074, + 0.046712782233953476, + 0.05826064944267273, + 0.07624053955078125, + 0.16205629706382751 + ], + "credit_map": [ + 0.026139259338378906, + 0.033373091369867325, + 0.02550043910741806, + 0.043820541352033615, + 0.032469116151332855, + 0.09801671653985977, + 0.03354697301983833, + 0.0797821506857872, + 0.007774951867759228, + 0.008788230828940868, + 0.006756159011274576, + 0.006397457327693701, + 0.005158589221537113, + 0.006134313996881247, + 0.009323370642960072, + 0.0056452201679348946, + 0.007686799857765436, + 0.008877492509782314, + 0.009727862663567066, + 0.009032834321260452, + 0.009707508608698845, + 0.009954397566616535, + 0.01043009851127863, + 0.010072104632854462, + 0.003652255516499281, + 0.007250410038977861, + 0.00724157365038991, + 0.004265368916094303, + 0.005656401626765728, + 0.003764157881960273, + 0.006617886479943991, + 0.007357317488640547, + 0.003435489721596241, + 0.003504921915009618, + 0.004020102322101593, + 0.0041129011660814285, + 0.0031534405425190926, + 0.0034254014026373625, + 0.006298329681158066, + 0.004541375674307346, + 0.004674759227782488, + 0.007389246951788664, + 0.00838808249682188, + 0.007406089920550585, + 0.011243847198784351, + 0.006711204536259174, + 0.014756433665752411, + 0.013747076503932476, + 0.0052907089702785015, + 0.0051461514085531235, + 0.008006804622709751, + 0.008076069876551628, + 0.005795871373265982, + 0.002814154140651226, + 0.0059971981681883335, + 0.005585821811109781, + 0.006886078976094723, + 0.005490906536579132, + 0.008773338980972767, + 0.00785430520772934, + 0.007009836845099926, + 0.004815499763935804, + 0.006927302572876215, + 0.010503382422029972, + 0.0048300824128091335, + 0.003930352162569761, + 0.009706830605864525, + 0.019306324422359467, + 0.006030665244907141, + 0.009473126381635666, + 0.007363222539424896, + 0.01559993252158165, + 0.009503770619630814, + 0.0061726877465844154, + 0.02644042670726776, + 0.012282012030482292, + 0.015993613749742508, + 0.032059866935014725, + 0.015618175268173218, + 0.04398574307560921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.003200054634362459 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808169439435005 + ], + "opposite_logprob_tokens": [ + -3.288081645965576 + ], + "logprob_reward": -0.03808169439435005, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808169439435005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3126769959926605, + 0.1870620995759964, + 0.1768811196088791, + 0.07742757350206375, + 0.04834916442632675, + 0.04264475405216217, + 0.04103117063641548, + 0.04909670352935791, + 0.030740054324269295, + 0.03409039229154587 + ], + "credit_map": [ + 0.022799907252192497, + 0.03265093266963959, + 0.047329131513834, + 0.030184954404830933, + 0.023398205637931824, + 0.08625521510839462, + 0.020926132798194885, + 0.04913250729441643, + 0.0169186033308506, + 0.017189152538776398, + 0.030541477724909782, + 0.027811726555228233, + 0.026523059234023094, + 0.016675077378749847, + 0.028361767530441284, + 0.023041237145662308, + 0.01567562110722065, + 0.01865721121430397, + 0.017905307933688164, + 0.021774131804704666, + 0.028613558039069176, + 0.024316666647791862, + 0.02746761403977871, + 0.022470999509096146, + 0.009147951379418373, + 0.015263662673532963, + 0.014589663594961166, + 0.010179754346609116, + 0.0074332160875201225, + 0.006459347903728485, + 0.007551875431090593, + 0.006802103947848082, + 0.006564721465110779, + 0.005251031834632158, + 0.007829921320080757, + 0.0064683519303798676, + 0.004420844838023186, + 0.004999465309083462, + 0.005445586517453194, + 0.007369239814579487, + 0.003902086755260825, + 0.0026655509136617184, + 0.00993707962334156, + 0.005056269932538271, + 0.005426394287496805, + 0.003176298923790455, + 0.005969100631773472, + 0.006511973682790995, + 0.003953167237341404, + 0.007303509395569563, + 0.0071099745109677315, + 0.007281547877937555, + 0.003102867631241679, + 0.0041386219672858715, + 0.0029489141888916492, + 0.005192567594349384, + 0.003926154226064682, + 0.006267634220421314, + 0.0059890286065638065, + 0.006479879375547171, + 0.007123950868844986, + 0.0060494947247207165, + 0.005842387676239014, + 0.007418177090585232, + 0.002209989121183753, + 0.00312896934337914, + 0.004552559461444616, + 0.008237075060606003, + 0.0031064411159604788, + 0.0023408238776028156, + 0.0023162937723100185, + 0.004847902804613113, + 0.0036115520633757114, + 0.004298387560993433, + 0.004939809441566467, + 0.006917567458003759, + 0.003965423908084631, + 0.0024822864215821028, + 0.0029079800006002188, + 0.004967386834323406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03808169439435005 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024852838832885027 + ], + "opposite_logprob_tokens": [ + -6.002485275268555 + ], + "logprob_reward": -0.0024852838832885027, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024852838832885027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12975375354290009, + 0.08225037902593613, + 0.0956660583615303, + 0.19741259515285492, + 0.1145872101187706, + 0.02291949652135372, + 0.13107547163963318, + 0.10477481782436371, + 0.04811329022049904, + 0.07344697415828705 + ], + "credit_map": [ + 0.009913293644785881, + 0.014321000315248966, + 0.018893936648964882, + 0.0223322045058012, + 0.009943689219653606, + 0.01679847203195095, + 0.01760691963136196, + 0.019944235682487488, + 0.009207317605614662, + 0.012391826137900352, + 0.012540829367935658, + 0.012797662988305092, + 0.008005991578102112, + 0.009162691421806812, + 0.009630294516682625, + 0.00851376075297594, + 0.011967324651777744, + 0.009145532734692097, + 0.011844667606055737, + 0.011144889518618584, + 0.013940583914518356, + 0.01293496135622263, + 0.013264685869216919, + 0.011423418298363686, + 0.023060381412506104, + 0.038111887872219086, + 0.03916934132575989, + 0.02906213514506817, + 0.01922132447361946, + 0.017911219969391823, + 0.0176719781011343, + 0.013204340822994709, + 0.01606578752398491, + 0.012136277742683887, + 0.02381264790892601, + 0.015103145502507687, + 0.008828902617096901, + 0.00804454367607832, + 0.009095856919884682, + 0.021500051021575928, + 0.0027032482903450727, + 0.003756102407351136, + 0.0025460775941610336, + 0.0032710253726691008, + 0.0030213117133826017, + 0.002203173004090786, + 0.0026523408014327288, + 0.002766217803582549, + 0.014957749284803867, + 0.02286096289753914, + 0.025155264884233475, + 0.024595551192760468, + 0.009002406150102615, + 0.010165444575250149, + 0.007453680969774723, + 0.016884412616491318, + 0.012718652375042439, + 0.012525581754744053, + 0.02006213180720806, + 0.016776280477643013, + 0.009267181158065796, + 0.006235514301806688, + 0.008205491118133068, + 0.018983988091349602, + 0.0037065453361719847, + 0.00383717636577785, + 0.00761744100600481, + 0.013901207596063614, + 0.004119414370507002, + 0.002914133481681347, + 0.002966175088658929, + 0.009051196277141571, + 0.006273389328271151, + 0.005329889245331287, + 0.01076216995716095, + 0.008364378474652767, + 0.00582120381295681, + 0.012288722209632397, + 0.006805516313761473, + 0.017801698297262192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.0024852838832885027 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029783163219690323 + ], + "opposite_logprob_tokens": [ + -3.529783248901367 + ], + "logprob_reward": -0.029783163219690323, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029783163219690323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31226488947868347, + 0.08643825352191925, + 0.08019225299358368, + 0.08187255263328552, + 0.10551004111766815, + 0.06920814514160156, + 0.11897817254066467, + 0.04508546367287636, + 0.024702034890651703, + 0.07574818283319473 + ], + "credit_map": [ + 0.027210697531700134, + 0.023615170270204544, + 0.05313608795404434, + 0.02144770510494709, + 0.02039158158004284, + 0.09484376013278961, + 0.02056717686355114, + 0.051052745431661606, + 0.01018031220883131, + 0.009068298153579235, + 0.015034799464046955, + 0.012619616463780403, + 0.008889256976544857, + 0.010008828714489937, + 0.0106210857629776, + 0.010016048327088356, + 0.006356159690767527, + 0.010003789328038692, + 0.00687201414257288, + 0.008494739420711994, + 0.013229930773377419, + 0.010229788720607758, + 0.014387458562850952, + 0.010618367232382298, + 0.01109644677489996, + 0.011888922192156315, + 0.015895850956439972, + 0.010438859462738037, + 0.006807561032474041, + 0.006191307678818703, + 0.007846656255424023, + 0.011706951074302197, + 0.010527173988521099, + 0.013799457810819149, + 0.02173743024468422, + 0.011455878615379333, + 0.007992666214704514, + 0.007314249873161316, + 0.009111984632909298, + 0.023571200668811798, + 0.005156087223440409, + 0.005355124361813068, + 0.016735048964619637, + 0.01081329770386219, + 0.008614440448582172, + 0.005269164685159922, + 0.00611149100586772, + 0.011153485625982285, + 0.014369850978255272, + 0.02224533073604107, + 0.01946139894425869, + 0.019516095519065857, + 0.008733091875910759, + 0.010447187349200249, + 0.008011111058294773, + 0.016194108873605728, + 0.004953066818416119, + 0.006716650910675526, + 0.007834390737116337, + 0.006590057630091906, + 0.005100989248603582, + 0.0028497155290097, + 0.004199727438390255, + 0.006840866524726152, + 0.0027658178005367517, + 0.0022860190365463495, + 0.0035841956268996, + 0.0028433415573090315, + 0.0023671172093600035, + 0.002552477177232504, + 0.002319504739716649, + 0.005983560346066952, + 0.005808213260024786, + 0.0041142902337014675, + 0.011096540838479996, + 0.00601420970633626, + 0.00734699098393321, + 0.016549112275242805, + 0.007890122011303902, + 0.016928700730204582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.029783163219690323 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with an intimate close-up, rendered in the warm, faded sepia tones of a realistic old photograph. A small ginger cat, its fur a soft down, clings tenaciously to the thick, cream-colored knit of its owner's sweater. With its large, crystalline blue eyes, the kitten gazes forward with innocent trust, its miniature pink nose twitching slightly, its paws gently hooking into the wool. Then, the camera slowly and steadily pulls back, a deliberate and gentle movement that gradually expands the frame. The image shows the kitten resting securely on a person's chest, their hands tenderly cradling it. As the camera continues to retreat, the soft-focus scene resolutely unveils an old-fashioned photography studio, bathed in soft, diffuse light from a large, invisible window that casts long, gentle shadows. In the background, an old bellows camera, made of polished dark wood with brass fittings, stands on a tall wooden tripod, its lens trained directly on the couple, ready to capture this moment. The entire scene is imbued with a quiet, nostalgic atmosphere, evoking a sense of timeless tenderness and captured memories.", + "seed": 30213, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_13/index_3251_seed_30213_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.028893844969570637, + "vlm_logprob": -0.028893844969570637, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.028893844969570637, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028525054454803467 + ], + "opposite_logprob_tokens": [ + -3.6535251140594482 + ], + "logprob_reward": -0.028525054454803467, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.028525054454803467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1433386653661728, + 0.09032586216926575, + 0.08472079038619995, + 0.07126922905445099, + 0.1261291652917862, + 0.12036880105733871, + 0.11453370749950409, + 0.05196624621748924, + 0.08918682485818863, + 0.10816073417663574 + ], + "credit_map": [ + 0.01825418882071972, + 0.015338408760726452, + 0.013671855442225933, + 0.021644387394189835, + 0.017018236219882965, + 0.010708099231123924, + 0.02888878993690014, + 0.017814697697758675, + 0.011935679242014885, + 0.012237746268510818, + 0.009995055384933949, + 0.012810603715479374, + 0.01030262652784586, + 0.013104195706546307, + 0.013254931196570396, + 0.006685023196041584, + 0.008500666357576847, + 0.01264664251357317, + 0.008473887108266354, + 0.009759563952684402, + 0.01853329874575138, + 0.012787933461368084, + 0.008258681744337082, + 0.005760115571320057, + 0.019005021080374718, + 0.006719924509525299, + 0.008343799971044064, + 0.008133064024150372, + 0.006533449981361628, + 0.008561749011278152, + 0.006897394545376301, + 0.007074827328324318, + 0.011538537219166756, + 0.00883305724710226, + 0.012539349496364594, + 0.01152035128325224, + 0.020510094240307808, + 0.015124508179724216, + 0.01755671575665474, + 0.02850654534995556, + 0.010273322463035583, + 0.01377213466912508, + 0.015505589544773102, + 0.009154550731182098, + 0.014830105938017368, + 0.014394486322999, + 0.016450360417366028, + 0.025988247245550156, + 0.011394429951906204, + 0.010180583223700523, + 0.012661037966609001, + 0.012584628537297249, + 0.011245827190577984, + 0.025035317987203598, + 0.013903769664466381, + 0.017528105527162552, + 0.007042677141726017, + 0.007287140004336834, + 0.00656426977366209, + 0.006252639461308718, + 0.004895977675914764, + 0.007922006770968437, + 0.005301236175000668, + 0.006700293626636267, + 0.006709121633321047, + 0.011235740967094898, + 0.010104281827807426, + 0.00887860357761383, + 0.012092232704162598, + 0.010531344451010227, + 0.009353697299957275, + 0.020281808450818062, + 0.009862481616437435, + 0.014092981815338135, + 0.012554426677525043, + 0.012676992453634739, + 0.01151945348829031, + 0.0113550228998065, + 0.015212254598736763, + 0.020887121558189392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.028525054454803467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646194823086262 + ], + "opposite_logprob_tokens": [ + -4.758646011352539 + ], + "logprob_reward": -0.008646194823086262, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646194823086262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10934299975633621, + 0.04157562553882599, + 0.2113976925611496, + 0.023681243881583214, + 0.06450340151786804, + 0.04309025779366493, + 0.07077539712190628, + 0.3241008222103119, + 0.04948147013783455, + 0.06205110624432564 + ], + "credit_map": [ + 0.013994375243782997, + 0.008936911821365356, + 0.01796112023293972, + 0.01755593903362751, + 0.01180698536336422, + 0.007721713278442621, + 0.014563011936843395, + 0.016802942380309105, + 0.006623824127018452, + 0.004813648760318756, + 0.004750128369778395, + 0.005069215316325426, + 0.0048400769010186195, + 0.0058150566183030605, + 0.005171334370970726, + 0.004492337815463543, + 0.022914305329322815, + 0.015173721127212048, + 0.015439799055457115, + 0.012219255790114403, + 0.06996720284223557, + 0.04978375881910324, + 0.013868113048374653, + 0.01203154493123293, + 0.0032597072422504425, + 0.002338652964681387, + 0.0030883660074323416, + 0.0037482185289263725, + 0.0027142262551933527, + 0.002635505050420761, + 0.0029691336676478386, + 0.002927434630692005, + 0.00755850737914443, + 0.010580385103821754, + 0.0072278412990272045, + 0.008001805283129215, + 0.00658152811229229, + 0.00571068050339818, + 0.007576349191367626, + 0.011266305111348629, + 0.005657935049384832, + 0.009173883125185966, + 0.0044244443997740746, + 0.005544312298297882, + 0.0038600030820816755, + 0.004524499177932739, + 0.005035912152379751, + 0.004869266413152218, + 0.00893232598900795, + 0.005810493603348732, + 0.0065616765059530735, + 0.006979210767894983, + 0.007647346239537001, + 0.017246980220079422, + 0.008810656145215034, + 0.008786708116531372, + 0.03241201117634773, + 0.06466187536716461, + 0.043425172567367554, + 0.038765355944633484, + 0.022370463237166405, + 0.02285958267748356, + 0.04436067119240761, + 0.05524568632245064, + 0.00492877559736371, + 0.0053313784301280975, + 0.007222133222967386, + 0.007178359664976597, + 0.007059267722070217, + 0.005477930884808302, + 0.005131055135279894, + 0.007152569945901632, + 0.005694745108485222, + 0.005506596062332392, + 0.008343461900949478, + 0.009569397196173668, + 0.009305757470428944, + 0.0059410687536001205, + 0.007198133040219545, + 0.01049194484949112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.008646194823086262 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038085367530584335 + ], + "opposite_logprob_tokens": [ + -3.288085460662842 + ], + "logprob_reward": -0.038085367530584335, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038085367530584335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07483144849538803, + 0.08345624059438705, + 0.09720679372549057, + 0.05339760333299637, + 0.13865864276885986, + 0.23480963706970215, + 0.19927281141281128, + 0.05763823539018631, + 0.041279520839452744, + 0.019449051469564438 + ], + "credit_map": [ + 0.011355108581483364, + 0.009341886267066002, + 0.010615465231239796, + 0.006948369555175304, + 0.007596906740218401, + 0.007744035217911005, + 0.007712617516517639, + 0.013517074286937714, + 0.01170385628938675, + 0.0121034886687994, + 0.009801401756703854, + 0.013830190524458885, + 0.008102533407509327, + 0.012632271274924278, + 0.008167936466634274, + 0.007114579435437918, + 0.009628061205148697, + 0.010156387463212013, + 0.01122021023184061, + 0.013252343982458115, + 0.025475306436419487, + 0.012869743630290031, + 0.009498093277215958, + 0.005106666125357151, + 0.010281456634402275, + 0.005462484434247017, + 0.006356974132359028, + 0.008993130177259445, + 0.004329897928982973, + 0.004951199050992727, + 0.005915801040828228, + 0.007106664590537548, + 0.014246771112084389, + 0.02597677707672119, + 0.01631791703402996, + 0.017876075580716133, + 0.013354785740375519, + 0.015837129205465317, + 0.01560566108673811, + 0.019443543627858162, + 0.018074918538331985, + 0.022127117961645126, + 0.02998063713312149, + 0.01906845159828663, + 0.03102763183414936, + 0.018318982794880867, + 0.024075573310256004, + 0.07213633507490158, + 0.02701674960553646, + 0.018960053101181984, + 0.02368289791047573, + 0.018456852063536644, + 0.022235242649912834, + 0.033185236155986786, + 0.026485946029424667, + 0.02924986742436886, + 0.005108124576508999, + 0.009913449175655842, + 0.00665880786255002, + 0.006951267831027508, + 0.006000330671668053, + 0.005631591193377972, + 0.00844897422939539, + 0.008925695903599262, + 0.00574065325781703, + 0.004129910841584206, + 0.00742242019623518, + 0.006820436101406813, + 0.003755115671083331, + 0.003944874741137028, + 0.003835841780528426, + 0.005630274303257465, + 0.0018382789567112923, + 0.0021058947313576937, + 0.0026817985344678164, + 0.0021247989498078823, + 0.0029220415744930506, + 0.0022480071056634188, + 0.0021664625965058804, + 0.0033617711160331964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.038085367530584335 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575884729623795 + ], + "opposite_logprob_tokens": [ + -3.895576000213623 + ], + "logprob_reward": -0.020575884729623795, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575884729623795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051477521657943726, + 0.04325258731842041, + 0.04610658809542656, + 0.035296835005283356, + 0.05317837372422218, + 0.09903069585561752, + 0.28338998556137085, + 0.13078303635120392, + 0.10161174833774567, + 0.15587258338928223 + ], + "credit_map": [ + 0.008435705676674843, + 0.006077205762267113, + 0.009047338739037514, + 0.00512042036280036, + 0.005200578831136227, + 0.0049277679063379765, + 0.005150331184267998, + 0.0075181759893894196, + 0.008641420863568783, + 0.004699003882706165, + 0.004914259072393179, + 0.005035297479480505, + 0.004878311883658171, + 0.005480037536472082, + 0.004973094444721937, + 0.004631167743355036, + 0.006355641409754753, + 0.004033228848129511, + 0.0036427988670766354, + 0.004757747519761324, + 0.009794667363166809, + 0.008318407461047173, + 0.0043540094047784805, + 0.004850088153034449, + 0.005557405762374401, + 0.004564099945127964, + 0.0036580096930265427, + 0.003954123705625534, + 0.003930838778614998, + 0.004294874612241983, + 0.004920513369143009, + 0.004416970536112785, + 0.005679298657923937, + 0.010288435034453869, + 0.005687200464308262, + 0.0034917977172881365, + 0.006010199431329966, + 0.008723423816263676, + 0.004815224092453718, + 0.008482796140015125, + 0.008329939097166061, + 0.01052868738770485, + 0.011580652557313442, + 0.008322428911924362, + 0.011883683502674103, + 0.015001125633716583, + 0.014075241982936859, + 0.019308937713503838, + 0.02706184610724449, + 0.020068461075425148, + 0.0271090567111969, + 0.022431205958127975, + 0.03438762202858925, + 0.07945124804973602, + 0.03559448570013046, + 0.037286095321178436, + 0.016725147143006325, + 0.018701624125242233, + 0.015582223422825336, + 0.011698859743773937, + 0.012087773531675339, + 0.024405285716056824, + 0.012797062285244465, + 0.01878506876528263, + 0.00823286734521389, + 0.010081097483634949, + 0.011981821618974209, + 0.008558614179491997, + 0.016013195738196373, + 0.01529943011701107, + 0.009919706732034683, + 0.021525023505091667, + 0.01201833225786686, + 0.01256992295384407, + 0.012204966507852077, + 0.012851675041019917, + 0.03351140767335892, + 0.022848162800073624, + 0.019536271691322327, + 0.030331850051879883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.020575884729623795 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048636723309755325 + ], + "opposite_logprob_tokens": [ + -3.0486366748809814 + ], + "logprob_reward": -0.048636723309755325, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048636723309755325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23581042885780334, + 0.04649578407406807, + 0.21873846650123596, + 0.05431945621967316, + 0.16634656488895416, + 0.05320773273706436, + 0.054769884794950485, + 0.03768153488636017, + 0.07264730334281921, + 0.05998287349939346 + ], + "credit_map": [ + 0.025492839515209198, + 0.03088746964931488, + 0.026074616238474846, + 0.041025713086128235, + 0.029267771169543266, + 0.016975436359643936, + 0.03338073194026947, + 0.032705843448638916, + 0.00910835713148117, + 0.004995360970497131, + 0.00540384603664279, + 0.0062105185352265835, + 0.005756956525146961, + 0.005065320059657097, + 0.005498635116964579, + 0.004456787835806608, + 0.028547663241624832, + 0.020696287974715233, + 0.02529251202940941, + 0.02154848724603653, + 0.05508795008063316, + 0.03333064168691635, + 0.019639866426587105, + 0.014595065265893936, + 0.008558531291782856, + 0.005961834918707609, + 0.0059506092220544815, + 0.009132858365774155, + 0.0052514695562422276, + 0.006736586336046457, + 0.006293910555541515, + 0.006433655042201281, + 0.01742372289299965, + 0.02119913138449192, + 0.015035131946206093, + 0.014795060269534588, + 0.01988324336707592, + 0.02337299846112728, + 0.020076587796211243, + 0.03456069529056549, + 0.005685022566467524, + 0.007753835525363684, + 0.0053437817841768265, + 0.006465804297477007, + 0.005295907147228718, + 0.00711081363260746, + 0.00752683449536562, + 0.008025732822716236, + 0.007787167094647884, + 0.007489723619073629, + 0.005108716897666454, + 0.005210544914007187, + 0.00690850056707859, + 0.00935231801122427, + 0.0051341853104531765, + 0.007778732106089592, + 0.007313990034162998, + 0.004925273358821869, + 0.004121839068830013, + 0.0036502820439636707, + 0.003938804846256971, + 0.004418531898409128, + 0.004691580776125193, + 0.0046212333254516125, + 0.0070931692607700825, + 0.007729297038167715, + 0.010362640954554081, + 0.01203315518796444, + 0.008391286246478558, + 0.0050231399945914745, + 0.010440096259117126, + 0.011574523523449898, + 0.0052256276831030846, + 0.007124283351004124, + 0.0063461665995419025, + 0.007856840267777443, + 0.0065666851587593555, + 0.006403261795639992, + 0.0091498838737607, + 0.011310122907161713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.048636723309755325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with an intimate close-up, rendered in the warm, faded sepia tones of a realistic old photograph. A small ginger cat, its fur a soft down, clings tenaciously to the thick, cream-colored knit of its owner's sweater. With its large, crystalline blue eyes, the kitten gazes forward with innocent trust, its miniature pink nose twitching slightly, its paws gently hooking into the wool. Then, the camera slowly and steadily pulls back, a deliberate and gentle movement that gradually expands the frame. The image shows the kitten resting securely on a person's chest, their hands tenderly cradling it. As the camera continues to retreat, the soft-focus scene resolutely unveils an old-fashioned photography studio, bathed in soft, diffuse light from a large, invisible window that casts long, gentle shadows. In the background, an old bellows camera, made of polished dark wood with brass fittings, stands on a tall wooden tripod, its lens trained directly on the couple, ready to capture this moment. The entire scene is imbued with a quiet, nostalgic atmosphere, evoking a sense of timeless tenderness and captured memories.", + "seed": 30214, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_13/index_3251_seed_30214_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.026444204431027175, + "vlm_logprob": -0.026444204431027175, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.026444204431027175, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.045246053487062454 + ], + "opposite_logprob_tokens": [ + -3.170246124267578 + ], + "logprob_reward": -0.045246053487062454, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.045246053487062454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16351871192455292, + 0.09119556844234467, + 0.06955495476722717, + 0.15436603128910065, + 0.060320574790239334, + 0.11602840572595596, + 0.11857227981090546, + 0.047589097172021866, + 0.13406631350517273, + 0.044788043946027756 + ], + "credit_map": [ + 0.017177151516079903, + 0.025414206087589264, + 0.02211514115333557, + 0.017177417874336243, + 0.016594043001532555, + 0.019763410091400146, + 0.028807053342461586, + 0.016470296308398247, + 0.011549343355000019, + 0.014197997748851776, + 0.010436744429171085, + 0.009744704701006413, + 0.007513235788792372, + 0.012708715163171291, + 0.013647426851093769, + 0.011397399008274078, + 0.0052547273226082325, + 0.012273970991373062, + 0.0062202392145991325, + 0.00669634435325861, + 0.007741862908005714, + 0.014917450025677681, + 0.0075131854973733425, + 0.008937172591686249, + 0.017071718350052834, + 0.029694871976971626, + 0.014772002585232258, + 0.0197148434817791, + 0.017136329784989357, + 0.019747726619243622, + 0.01699831709265709, + 0.019230235368013382, + 0.010138914920389652, + 0.007131833583116531, + 0.008767951279878616, + 0.0074080671183764935, + 0.005634088534861803, + 0.0060726371593773365, + 0.006576072424650192, + 0.008591010235249996, + 0.011508279480040073, + 0.010244573466479778, + 0.017822472378611565, + 0.022865211591124535, + 0.014989054761826992, + 0.01157717127352953, + 0.011599909514188766, + 0.015421736985445023, + 0.008094573393464088, + 0.007889378815889359, + 0.00791449286043644, + 0.009670671075582504, + 0.011758220382034779, + 0.031158532947301865, + 0.015402034856379032, + 0.026684381067752838, + 0.0056389812380075455, + 0.005647548008710146, + 0.005514743737876415, + 0.004807355348020792, + 0.005354119930416346, + 0.007765704300254583, + 0.004747016355395317, + 0.008113625459372997, + 0.005959928967058659, + 0.007027596700936556, + 0.010713270865380764, + 0.006166582461446524, + 0.007736205123364925, + 0.04423503950238228, + 0.006567271891981363, + 0.045660413801670074, + 0.003426668467000127, + 0.003799940226599574, + 0.004866648465394974, + 0.004229199141263962, + 0.0036083662416785955, + 0.010170440189540386, + 0.005168854724615812, + 0.009517927654087543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.045246053487062454 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410571089014411 + ], + "opposite_logprob_tokens": [ + -5.504105567932129 + ], + "logprob_reward": -0.00410571089014411, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410571089014411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1245822086930275, + 0.09197602421045303, + 0.19562207162380219, + 0.15859755873680115, + 0.03660565987229347, + 0.04984869435429573, + 0.03152242302894592, + 0.18789970874786377, + 0.0880306139588356, + 0.0353151373565197 + ], + "credit_map": [ + 0.01327152643352747, + 0.013601398095488548, + 0.012749041430652142, + 0.014040418900549412, + 0.014626742340624332, + 0.019805969670414925, + 0.023470381274819374, + 0.013016707263886929, + 0.005313824396580458, + 0.011115356348454952, + 0.006310417782515287, + 0.009462421759963036, + 0.0092335669323802, + 0.02069787122309208, + 0.015967296436429024, + 0.013875260949134827, + 0.01026399340480566, + 0.016964303329586983, + 0.011395051144063473, + 0.016529520973563194, + 0.01915268786251545, + 0.05909864977002144, + 0.03336526080965996, + 0.028852583840489388, + 0.010797788389027119, + 0.024499651044607162, + 0.011334408074617386, + 0.017589829862117767, + 0.018571292981505394, + 0.026929261162877083, + 0.025686398148536682, + 0.023188918828964233, + 0.0041367048397660255, + 0.003865102306008339, + 0.0032128109596669674, + 0.007054754998534918, + 0.003282044315710664, + 0.005221786443144083, + 0.004828245844691992, + 0.0050042071379721165, + 0.004672956187278032, + 0.004221139010041952, + 0.0070836544036865234, + 0.007975003682076931, + 0.006629772949963808, + 0.007251937873661518, + 0.005388727877289057, + 0.006625498179346323, + 0.0034289942122995853, + 0.0048271771520376205, + 0.003581010038033128, + 0.0021133339032530785, + 0.0038567970041185617, + 0.005424736067652702, + 0.002920266706496477, + 0.005370102822780609, + 0.008307150565087795, + 0.007295516319572926, + 0.006370868533849716, + 0.004695229232311249, + 0.006436143536120653, + 0.0716000497341156, + 0.008316943421959877, + 0.0748777985572815, + 0.0026900849770754576, + 0.002259134314954281, + 0.004241274669766426, + 0.004086244851350784, + 0.004386167507618666, + 0.032974280416965485, + 0.003930387552827597, + 0.03346302732825279, + 0.002459015231579542, + 0.0021695494651794434, + 0.002827694173902273, + 0.0035680264700204134, + 0.004663399886339903, + 0.007251671049743891, + 0.005744635127484798, + 0.006631144322454929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.00410571089014411 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625991944223642 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007625991944223642, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625991944223642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11764215677976608, + 0.10186588019132614, + 0.11704494804143906, + 0.21615168452262878, + 0.091334767639637, + 0.1337808072566986, + 0.036604080349206924, + 0.08342786133289337, + 0.06851323693990707, + 0.03363459184765816 + ], + "credit_map": [ + 0.009747345931828022, + 0.014674687758088112, + 0.013221883215010166, + 0.013409611769020557, + 0.016114836558699608, + 0.021693697199225426, + 0.0159995649009943, + 0.012780524790287018, + 0.005461702588945627, + 0.023056596517562866, + 0.00637331698089838, + 0.013169770129024982, + 0.008395344950258732, + 0.015420724637806416, + 0.015494704246520996, + 0.014493713155388832, + 0.00669829873368144, + 0.013234985992312431, + 0.006959019228816032, + 0.01545463316142559, + 0.013971422798931599, + 0.02835417166352272, + 0.016045473515987396, + 0.01632693223655224, + 0.018479570746421814, + 0.04250847175717354, + 0.018452642485499382, + 0.02791677415370941, + 0.025265218690037727, + 0.02851691097021103, + 0.02641306072473526, + 0.028599023818969727, + 0.013684130273759365, + 0.008459780365228653, + 0.012406907044351101, + 0.009549795649945736, + 0.010690080001950264, + 0.013551200740039349, + 0.011631707660853863, + 0.011361166834831238, + 0.013663915917277336, + 0.013072645291686058, + 0.02187291905283928, + 0.013972611166536808, + 0.014596864581108093, + 0.019448518753051758, + 0.014093151316046715, + 0.023060180246829987, + 0.004115611780434847, + 0.003079312387853861, + 0.00420262198895216, + 0.0048133376985788345, + 0.002969104563817382, + 0.006133607588708401, + 0.004107039421796799, + 0.007183443754911423, + 0.003203802974894643, + 0.0036183209158480167, + 0.0051421853713691235, + 0.004799982998520136, + 0.007130250334739685, + 0.026403462514281273, + 0.007214091252535582, + 0.02591576799750328, + 0.0035862440709024668, + 0.004433139227330685, + 0.00773931248113513, + 0.005017654038965702, + 0.005318954586982727, + 0.018555648624897003, + 0.005630847532302141, + 0.018231432884931564, + 0.003422546898946166, + 0.003192045260220766, + 0.003085300326347351, + 0.0032221137080341578, + 0.0031256761867552996, + 0.00729442248120904, + 0.0037595091853290796, + 0.006532975938171148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.007625991944223642 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248458590358496 + ], + "opposite_logprob_tokens": [ + -5.255248546600342 + ], + "logprob_reward": -0.005248458590358496, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248458590358496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05826985836029053, + 0.09818468987941742, + 0.040654346346855164, + 0.26502591371536255, + 0.05539432168006897, + 0.06806216388940811, + 0.05874185264110565, + 0.17144140601158142, + 0.11658534407615662, + 0.06764017045497894 + ], + "credit_map": [ + 0.00394261721521616, + 0.007275748532265425, + 0.004446454346179962, + 0.0086228521540761, + 0.00714410375803709, + 0.010270688682794571, + 0.009619091637432575, + 0.006948298774659634, + 0.005269451532512903, + 0.018253600224852562, + 0.006106800865381956, + 0.011699097231030464, + 0.009966064244508743, + 0.015440845862030983, + 0.017560357227921486, + 0.013888475485146046, + 0.002479136921465397, + 0.004022909794002771, + 0.0030401453841477633, + 0.003887958126142621, + 0.0041000922210514545, + 0.011272105388343334, + 0.005882082507014275, + 0.00596991553902626, + 0.024248726665973663, + 0.05351855978369713, + 0.021564843133091927, + 0.037354521453380585, + 0.03274569287896156, + 0.027457574382424355, + 0.031244920566678047, + 0.03689108043909073, + 0.006793470121920109, + 0.0067565301433205605, + 0.005040412303060293, + 0.011434115469455719, + 0.005638742353767157, + 0.0067680939100682735, + 0.006242448464035988, + 0.006720514036715031, + 0.0064241038635373116, + 0.007606895640492439, + 0.009944451041519642, + 0.011335284449160099, + 0.00821442250162363, + 0.007966243661940098, + 0.006812400184571743, + 0.00975836906582117, + 0.005650038365274668, + 0.00485844723880291, + 0.0074511026032269, + 0.004469366744160652, + 0.004438663367182016, + 0.012260896153748035, + 0.005384404212236404, + 0.01422892976552248, + 0.007122475653886795, + 0.009631012566387653, + 0.009413386695086956, + 0.008041027933359146, + 0.008726583793759346, + 0.06058965250849724, + 0.008208592422306538, + 0.059708669781684875, + 0.0055539170280098915, + 0.00478475634008646, + 0.010086878202855587, + 0.005631399340927601, + 0.005678464658558369, + 0.03989611938595772, + 0.0049193562008440495, + 0.0400344580411911, + 0.006543135270476341, + 0.007178593892604113, + 0.009272214956581593, + 0.009095318615436554, + 0.008735480718314648, + 0.009228870272636414, + 0.009438702836632729, + 0.008147858083248138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.005248458590358496 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999480724334717 + ], + "opposite_logprob_tokens": [ + -2.6949949264526367 + ], + "logprob_reward": -0.06999480724334717, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999480724334717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09325867146253586, + 0.07519790530204773, + 0.03353056311607361, + 0.2625153362751007, + 0.08817467093467712, + 0.05366286262869835, + 0.044858671724796295, + 0.06576547771692276, + 0.2356557995080948, + 0.047380074858665466 + ], + "credit_map": [ + 0.005885294638574123, + 0.014326246455311775, + 0.009141760878264904, + 0.012165265157818794, + 0.010123117826879025, + 0.013078458607196808, + 0.01788710616528988, + 0.010651422664523125, + 0.008616549894213676, + 0.012522159144282341, + 0.01103057898581028, + 0.009884649887681007, + 0.0061658937484025955, + 0.0076643479987978935, + 0.010360338725149632, + 0.00895338412374258, + 0.0018581337062641978, + 0.005932920146733522, + 0.0027296298649162054, + 0.004039432387799025, + 0.005203548353165388, + 0.005200068932026625, + 0.004515735898166895, + 0.004051093943417072, + 0.02511637471616268, + 0.053756266832351685, + 0.021399496123194695, + 0.03464463725686073, + 0.030888082459568977, + 0.02986181527376175, + 0.030764859169721603, + 0.036083824932575226, + 0.009921659715473652, + 0.01321451272815466, + 0.010168541222810745, + 0.01177302934229374, + 0.010875307954847813, + 0.011567775160074234, + 0.010170452296733856, + 0.010483389720320702, + 0.0067382086999714375, + 0.0058019827120006084, + 0.008780176751315594, + 0.009238392114639282, + 0.0051764193922281265, + 0.005928973201662302, + 0.003905133344233036, + 0.008093575946986675, + 0.0035784179344773293, + 0.0034782257862389088, + 0.004821042995899916, + 0.003885351587086916, + 0.005434826482087374, + 0.00976902712136507, + 0.005576739087700844, + 0.00831503514200449, + 0.004274563863873482, + 0.003144695656374097, + 0.0038709838408976793, + 0.0028189646545797586, + 0.0034294691868126392, + 0.021618688479065895, + 0.004803316667675972, + 0.021804790943861008, + 0.00690700626000762, + 0.005360469222068787, + 0.006788524799048901, + 0.004927205853164196, + 0.0066566187888383865, + 0.09898258000612259, + 0.0056426976807415485, + 0.10039068758487701, + 0.004765263292938471, + 0.004231572151184082, + 0.005672023165971041, + 0.005756209138780832, + 0.004665641579777002, + 0.007608742453157902, + 0.005862525664269924, + 0.008818100206553936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.06999480724334717 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with an intimate close-up, rendered in the warm, faded sepia tones of a realistic old photograph. A small ginger cat, its fur a soft down, clings tenaciously to the thick, cream-colored knit of its owner's sweater. With its large, crystalline blue eyes, the kitten gazes forward with innocent trust, its miniature pink nose twitching slightly, its paws gently hooking into the wool. Then, the camera slowly and steadily pulls back, a deliberate and gentle movement that gradually expands the frame. The image shows the kitten resting securely on a person's chest, their hands tenderly cradling it. As the camera continues to retreat, the soft-focus scene resolutely unveils an old-fashioned photography studio, bathed in soft, diffuse light from a large, invisible window that casts long, gentle shadows. In the background, an old bellows camera, made of polished dark wood with brass fittings, stands on a tall wooden tripod, its lens trained directly on the couple, ready to capture this moment. The entire scene is imbued with a quiet, nostalgic atmosphere, evoking a sense of timeless tenderness and captured memories.", + "seed": 30215, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_13/index_3251_seed_30215_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.018518637400120497, + "vlm_logprob": -0.018518637400120497, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.018518637400120497, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009343350306153297 + ], + "opposite_logprob_tokens": [ + -4.884343147277832 + ], + "logprob_reward": -0.009343350306153297, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.009343350306153297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3115817606449127, + 0.12955330312252045, + 0.09172970056533813, + 0.07663009315729141, + 0.09267248213291168, + 0.07634744793176651, + 0.04260997846722603, + 0.05166148394346237, + 0.060179632157087326, + 0.06703414022922516 + ], + "credit_map": [ + 0.026056209579110146, + 0.025717495009303093, + 0.039371307939291, + 0.0333663709461689, + 0.054333899170160294, + 0.02751871570944786, + 0.061951007694005966, + 0.043266747146844864, + 0.020870137959718704, + 0.015850400552153587, + 0.015967469662427902, + 0.018308570608496666, + 0.020932521671056747, + 0.010119852609932423, + 0.015942569822072983, + 0.011561785824596882, + 0.011042525991797447, + 0.007224653847515583, + 0.011451919563114643, + 0.009673776105046272, + 0.0175256859511137, + 0.00735674099996686, + 0.01666826754808426, + 0.010786127299070358, + 0.009919836185872555, + 0.009154715575277805, + 0.008225343190133572, + 0.011793053708970547, + 0.009708408266305923, + 0.006060952786356211, + 0.011147472076117992, + 0.010620315559208393, + 0.015206574462354183, + 0.012341375462710857, + 0.011963513679802418, + 0.013250325806438923, + 0.016430366784334183, + 0.004997743293642998, + 0.012506498023867607, + 0.005976085551083088, + 0.013809112831950188, + 0.012768370099365711, + 0.006239361595362425, + 0.009928233921527863, + 0.011487393639981747, + 0.005101005081087351, + 0.00924579706043005, + 0.007768175099045038, + 0.004602139815688133, + 0.0043658060021698475, + 0.004616625141352415, + 0.005578141659498215, + 0.004740172531455755, + 0.004324349109083414, + 0.008546249009668827, + 0.005836495663970709, + 0.009274855256080627, + 0.006126365624368191, + 0.007116576191037893, + 0.006791482213884592, + 0.005196805577725172, + 0.004945721011608839, + 0.00583231495693326, + 0.006377363111823797, + 0.007148077245801687, + 0.006375021766871214, + 0.008622642606496811, + 0.006605582311749458, + 0.009641607291996479, + 0.006527645047754049, + 0.0075842333026230335, + 0.007674821186810732, + 0.005342771764844656, + 0.0047859991900622845, + 0.008503948338329792, + 0.007259939797222614, + 0.010298413224518299, + 0.009648805484175682, + 0.011249341070652008, + 0.009944920428097248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.009343350306153297 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744240876287222 + ], + "opposite_logprob_tokens": [ + -5.006744384765625 + ], + "logprob_reward": -0.006744240876287222, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744240876287222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20292015373706818, + 0.10548457503318787, + 0.09107962250709534, + 0.13519737124443054, + 0.06101606786251068, + 0.059087127447128296, + 0.08597645908594131, + 0.07156684249639511, + 0.09437034279108047, + 0.0933014303445816 + ], + "credit_map": [ + 0.01589840278029442, + 0.026159802451729774, + 0.02692258171737194, + 0.02060581184923649, + 0.03293631225824356, + 0.01946975104510784, + 0.040302980691194534, + 0.020624516531825066, + 0.00617661327123642, + 0.006760127376765013, + 0.01004470232874155, + 0.010913798585534096, + 0.027297165244817734, + 0.010670812800526619, + 0.022296397015452385, + 0.011324967257678509, + 0.012665674090385437, + 0.009985010139644146, + 0.008680170401930809, + 0.008007381111383438, + 0.011931267566978931, + 0.014685048721730709, + 0.012784343212842941, + 0.012340736575424671, + 0.018091190606355667, + 0.015848791226744652, + 0.013339769095182419, + 0.026755137369036674, + 0.016897480934858322, + 0.010014315135776997, + 0.018705157563090324, + 0.015545534901320934, + 0.012010904960334301, + 0.006071796175092459, + 0.009364459663629532, + 0.006521671544760466, + 0.009786133654415607, + 0.005087440833449364, + 0.007729869801551104, + 0.00444379635155201, + 0.0095896627753973, + 0.007714278995990753, + 0.008504894562065601, + 0.007143640425056219, + 0.007474022451788187, + 0.005966850556433201, + 0.00677874032407999, + 0.005915042012929916, + 0.00940447673201561, + 0.009826547466218472, + 0.012625624425709248, + 0.011686326935887337, + 0.009189576841890812, + 0.008480940014123917, + 0.014007086865603924, + 0.010755891911685467, + 0.007033847272396088, + 0.009022770449519157, + 0.015133058652281761, + 0.013365722261369228, + 0.005367794539779425, + 0.004676744807511568, + 0.007049534935504198, + 0.009917381219565868, + 0.009449011646211147, + 0.017203152179718018, + 0.00983294378966093, + 0.014059361070394516, + 0.010427115485072136, + 0.010094190947711468, + 0.009533988311886787, + 0.013770584017038345, + 0.008154571056365967, + 0.007760540582239628, + 0.009435254149138927, + 0.009106501005589962, + 0.012088034301996231, + 0.01636917144060135, + 0.016910890117287636, + 0.013476476073265076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.006744240876287222 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029790107160806656 + ], + "opposite_logprob_tokens": [ + -3.529790163040161 + ], + "logprob_reward": -0.029790107160806656, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029790107160806656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14773470163345337, + 0.20928454399108887, + 0.07432878762483597, + 0.15303188562393188, + 0.15085479617118835, + 0.05673505365848541, + 0.0448983795940876, + 0.08271832764148712, + 0.04991605505347252, + 0.030497489497065544 + ], + "credit_map": [ + 0.009744557552039623, + 0.02148519828915596, + 0.01565101183950901, + 0.01688777655363083, + 0.022069279104471207, + 0.0099985022097826, + 0.029556959867477417, + 0.022341413423419, + 0.018287986516952515, + 0.017440371215343475, + 0.0215532798320055, + 0.014887635596096516, + 0.05016498267650604, + 0.02675318904221058, + 0.03618693724274635, + 0.02401016093790531, + 0.008549260906875134, + 0.006449192296713591, + 0.00976447481662035, + 0.010514031164348125, + 0.009964563883841038, + 0.006455664522945881, + 0.014893143437802792, + 0.007738461252301931, + 0.01930520124733448, + 0.018122412264347076, + 0.01858633942902088, + 0.023384004831314087, + 0.02118556573987007, + 0.010846046730875969, + 0.02343943528831005, + 0.01816287450492382, + 0.0274836253374815, + 0.017773134633898735, + 0.019563674926757812, + 0.019849298521876335, + 0.018969574943184853, + 0.009914995171129704, + 0.01873116008937359, + 0.01856933906674385, + 0.008365793153643608, + 0.011014174669981003, + 0.006840310525149107, + 0.008016908541321754, + 0.007980144582688808, + 0.003364383243024349, + 0.00594714330509305, + 0.005206197500228882, + 0.005454721860587597, + 0.0032827795948833227, + 0.005243627820163965, + 0.0070048109628260136, + 0.005669483914971352, + 0.005888874642550945, + 0.006868969649076462, + 0.005485112313181162, + 0.011333491653203964, + 0.00958737451583147, + 0.011952053755521774, + 0.01069825142621994, + 0.010470309294760227, + 0.007089571561664343, + 0.008570106700062752, + 0.013017164543271065, + 0.004775989800691605, + 0.004550166893750429, + 0.00907843466848135, + 0.005014351103454828, + 0.006405624561011791, + 0.006167850457131863, + 0.006414303556084633, + 0.007509332150220871, + 0.0028896303847432137, + 0.0029035168699920177, + 0.0050054858438670635, + 0.0036040181294083595, + 0.004075727425515652, + 0.004116669297218323, + 0.004025515168905258, + 0.003876925678923726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.029790107160806656 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634140715003014 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.008634140715003014, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634140715003014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17232397198677063, + 0.12561039626598358, + 0.09939242154359818, + 0.14871054887771606, + 0.06235430762171745, + 0.075301893055439, + 0.1058850809931755, + 0.03999309241771698, + 0.07506737112998962, + 0.09536093473434448 + ], + "credit_map": [ + 0.012419251725077629, + 0.02610708773136139, + 0.021126754581928253, + 0.023672033101320267, + 0.02242289111018181, + 0.015989307314157486, + 0.02927950769662857, + 0.021307136863470078, + 0.00876757875084877, + 0.009844939224421978, + 0.013935643248260021, + 0.010579806752502918, + 0.032001469284296036, + 0.014232918620109558, + 0.02342699095606804, + 0.012821056880056858, + 0.015108427964150906, + 0.009203375317156315, + 0.010301973670721054, + 0.008920427411794662, + 0.018683111295104027, + 0.008676756173372269, + 0.015791740268468857, + 0.01270661223679781, + 0.023478319868445396, + 0.017589937895536423, + 0.017878612503409386, + 0.019678646698594093, + 0.01701366901397705, + 0.013526623137295246, + 0.018489733338356018, + 0.021055003628134727, + 0.010313762351870537, + 0.00902114249765873, + 0.009042334742844105, + 0.008618403226137161, + 0.008281206712126732, + 0.0037285147700458765, + 0.008406619541347027, + 0.0049423277378082275, + 0.013937578536570072, + 0.010257250629365444, + 0.0068489257246255875, + 0.008718294091522694, + 0.011031299829483032, + 0.006211731117218733, + 0.009616541676223278, + 0.008680273778736591, + 0.02289586514234543, + 0.010274163447320461, + 0.01732020266354084, + 0.009874868206679821, + 0.008864672854542732, + 0.008564953692257404, + 0.013468913733959198, + 0.014621436595916748, + 0.004718870390206575, + 0.005078242626041174, + 0.006665593944489956, + 0.005824402440339327, + 0.003908345941454172, + 0.0041460637003183365, + 0.004018903709948063, + 0.005632669664919376, + 0.007973907515406609, + 0.009970908053219318, + 0.016109298914670944, + 0.009163296781480312, + 0.006295633502304554, + 0.006678484845906496, + 0.009340639226138592, + 0.009535200893878937, + 0.008741511031985283, + 0.015301532112061977, + 0.017837025225162506, + 0.016874413937330246, + 0.006080728489905596, + 0.007051194552332163, + 0.010502857156097889, + 0.012971673160791397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.008634140715003014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in warm, faded sepia tones resembling an old photograph?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small ginger cat with soft down fur clinging to a thick, cream-colored knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to expand the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an old-fashioned photography studio containing a dark wood bellows camera on a wooden tripod?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffuse light casting long, gentle shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038081347942352295 + ], + "opposite_logprob_tokens": [ + -3.288081407546997 + ], + "logprob_reward": -0.038081347942352295, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038081347942352295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35599225759506226, + 0.07281136512756348, + 0.0600140206515789, + 0.09101199358701706, + 0.11301568895578384, + 0.0426449254155159, + 0.07744687050580978, + 0.03680587559938431, + 0.08601482957601547, + 0.0642421767115593 + ], + "credit_map": [ + 0.033530715852975845, + 0.05142050236463547, + 0.040486931800842285, + 0.04694259911775589, + 0.04857373237609863, + 0.029062194749712944, + 0.06230030953884125, + 0.04367528483271599, + 0.01347809936851263, + 0.008707204833626747, + 0.008374919183552265, + 0.008064663968980312, + 0.008505076169967651, + 0.009077249094843864, + 0.010491854511201382, + 0.006112304981797934, + 0.011316737160086632, + 0.006259322632104158, + 0.006313195452094078, + 0.006417732685804367, + 0.007372292689979076, + 0.005672134459018707, + 0.008570324629545212, + 0.008092286065220833, + 0.014052510261535645, + 0.008366347290575504, + 0.010575813241302967, + 0.009938539005815983, + 0.012678529135882854, + 0.007602599449455738, + 0.01338769868016243, + 0.014409961178898811, + 0.017341306433081627, + 0.014480317011475563, + 0.013878865167498589, + 0.01835246942937374, + 0.018282314762473106, + 0.006273823790252209, + 0.01612241193652153, + 0.00828417669981718, + 0.005157778039574623, + 0.0064879837445914745, + 0.0036794894840568304, + 0.006039926316589117, + 0.006045377813279629, + 0.003478418570011854, + 0.004408926237374544, + 0.007347029633820057, + 0.009446599520742893, + 0.009065182879567146, + 0.012251443229615688, + 0.011253499425947666, + 0.007712365128099918, + 0.00514667434617877, + 0.009987859055399895, + 0.012583255767822266, + 0.003276623785495758, + 0.003257768927142024, + 0.006737260147929192, + 0.006298914086073637, + 0.0030867839232087135, + 0.003528510220348835, + 0.0044357613660395145, + 0.006184255238622427, + 0.008952461183071136, + 0.008267223834991455, + 0.02089867554605007, + 0.008938469924032688, + 0.012713167816400528, + 0.006239746231585741, + 0.00999389123171568, + 0.010011207312345505, + 0.004829490557312965, + 0.008417923003435135, + 0.01450461708009243, + 0.009973225183784962, + 0.004737949464470148, + 0.006780174560844898, + 0.006474731490015984, + 0.008524070493876934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.038081347942352295 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_14/index_tensor([9256, 9256])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_14/index_tensor([9256, 9256])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..64630fa1d8d3903ecb89fe6236f67b4d666bad65 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_14/index_tensor([9256, 9256])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the style of a sports documentary, a low-angle, high-speed camera tracking shot, held close to the ground, captures in extreme slow motion the moment a professional motorcyclist conquers a precipitous mountain curve. Radiating the golden glow of the late afternoon sun, the athlete, clad in an aerodynamic black leather racing suit, its form-fitting leather embellished with striking red and white stripes, leans toward the apex at an improbable angle, his knees, covered in heavily worn sliders, almost parallel to the rough, dark gray asphalt. He controls a matte gray modern racing car, its lines sharp and its suspension compressed to the extreme by the immense centrifugal force. As the motorcycle leans to its limit, the motorcycle's metal footpegs and the athlete's knee sliders scrape against the road, instantly sending forth a shower of dazzling sparks. Under the extreme slow-motion particle effects, each spark is clearly visible. Like blazing miniature comets, they splash backwards with bright orange-yellow trails, briefly illuminating the sidewall of the rear tire that is spinning at high speed and slightly deformed by pressure, while stirring up a small amount of rubber and dust particles suspended in the warm air. The whole picture is full of the ultimate beauty of power, concentration and speed, creating a heart-pounding atmosphere of tension and awe.", + "seed": 19465, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_14/index_9256_seed_19465_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.029948740173131226, + "vlm_logprob": -0.029948740173131226, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.029948740173131226, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017394840717315674 + ], + "opposite_logprob_tokens": [ + -4.26739501953125 + ], + "logprob_reward": -0.017394840717315674, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017394840717315674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1201988011598587, + 0.19024604558944702, + 0.06996748596429825, + 0.07543990761041641, + 0.14785155653953552, + 0.06339432299137115, + 0.10130830854177475, + 0.09450556337833405, + 0.07816633582115173, + 0.05892170965671539 + ], + "credit_map": [ + 0.013067895546555519, + 0.014207315631210804, + 0.014272150583565235, + 0.012927791103720665, + 0.016596416011452675, + 0.01618022285401821, + 0.013554390519857407, + 0.019392624497413635, + 0.014903957955539227, + 0.01940896548330784, + 0.014623790979385376, + 0.04223935678601265, + 0.017714079469442368, + 0.04412485659122467, + 0.017278067767620087, + 0.019952964037656784, + 0.012533213011920452, + 0.008910778909921646, + 0.007421273272484541, + 0.008520033210515976, + 0.008096692152321339, + 0.009144091978669167, + 0.006690910551697016, + 0.008650487288832664, + 0.01094778347760439, + 0.009141994640231133, + 0.008812167681753635, + 0.00818724650889635, + 0.008531859144568443, + 0.008912891149520874, + 0.00836783368140459, + 0.012538129463791847, + 0.015816275030374527, + 0.02608652226626873, + 0.016310609877109528, + 0.020954731851816177, + 0.022332387045025826, + 0.011462985537946224, + 0.016803765669465065, + 0.018084285780787468, + 0.008529206737875938, + 0.0048321387730538845, + 0.010901224799454212, + 0.009235461242496967, + 0.007507139816880226, + 0.008026494644582272, + 0.007631679065525532, + 0.006730976514518261, + 0.00972703006118536, + 0.009430915117263794, + 0.01745479740202427, + 0.019396714866161346, + 0.013829154893755913, + 0.008589300327003002, + 0.013013984076678753, + 0.009866411797702312, + 0.007188841234892607, + 0.012807535007596016, + 0.011345377191901207, + 0.020892595872282982, + 0.009735587984323502, + 0.010777951218187809, + 0.012477430514991283, + 0.009280248545110226, + 0.009501202031970024, + 0.010600333102047443, + 0.008631018921732903, + 0.008764771744608879, + 0.01687699556350708, + 0.005926242098212242, + 0.008324719034135342, + 0.009541044011712074, + 0.00444523710757494, + 0.006428866647183895, + 0.004520983435213566, + 0.004662118386477232, + 0.01181021984666586, + 0.00926175992935896, + 0.010125098749995232, + 0.007667424622923136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.017394840717315674 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032272338867188 + ], + "opposite_logprob_tokens": [ + -2.350322723388672 + ], + "logprob_reward": -0.10032272338867188, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032272338867188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1324072778224945, + 0.09027189016342163, + 0.043433062732219696, + 0.09337613731622696, + 0.16588278114795685, + 0.08307778090238571, + 0.10044098645448685, + 0.15432564914226532, + 0.10951415449380875, + 0.02727026492357254 + ], + "credit_map": [ + 0.029515216127038002, + 0.01729251816868782, + 0.008458879776299, + 0.011665605939924717, + 0.02246830053627491, + 0.015501085668802261, + 0.011681649833917618, + 0.015824023634195328, + 0.009022608399391174, + 0.01065296120941639, + 0.011183062568306923, + 0.012483358383178711, + 0.010979563929140568, + 0.015709776431322098, + 0.01265376340597868, + 0.007586798630654812, + 0.0037609501741826534, + 0.0043256934732198715, + 0.005358046852052212, + 0.003540735226124525, + 0.004576507024466991, + 0.006802654359489679, + 0.006393472198396921, + 0.008675002492964268, + 0.006565274205058813, + 0.005583826452493668, + 0.013542926870286465, + 0.005320143420249224, + 0.010881326161324978, + 0.0240208487957716, + 0.011563770473003387, + 0.015898022800683975, + 0.017897874116897583, + 0.033693067729473114, + 0.01797168329358101, + 0.02087285742163658, + 0.022579500451683998, + 0.016246607527136803, + 0.01807272620499134, + 0.01854845881462097, + 0.010543843731284142, + 0.009685634635388851, + 0.015747375786304474, + 0.007137000560760498, + 0.00971394032239914, + 0.008388414978981018, + 0.009821193292737007, + 0.01204038318246603, + 0.015682924538850784, + 0.009286493062973022, + 0.015371003188192844, + 0.014265850186347961, + 0.013398716226220131, + 0.009496049955487251, + 0.012077763676643372, + 0.010862196795642376, + 0.01259406004101038, + 0.025316128507256508, + 0.017858345061540604, + 0.026408739387989044, + 0.02346987836062908, + 0.017684347927570343, + 0.01730240508913994, + 0.01369174662977457, + 0.014823528937995434, + 0.014592275954782963, + 0.014193943701684475, + 0.01066764909774065, + 0.023654911667108536, + 0.009350710548460484, + 0.011565489694476128, + 0.010665646754205227, + 0.0019178123911842704, + 0.0036777921486645937, + 0.0025155055336654186, + 0.0022411178797483444, + 0.007069431710988283, + 0.004030325450003147, + 0.0030383807606995106, + 0.0027799003291875124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.10032272338867188 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579621195793152 + ], + "opposite_logprob_tokens": [ + -3.8955795764923096 + ], + "logprob_reward": -0.020579621195793152, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579621195793152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0689336284995079, + 0.133525088429451, + 0.047732818871736526, + 0.10682830959558487, + 0.053384117782115936, + 0.11590096354484558, + 0.10213790833950043, + 0.08321008831262589, + 0.10304269939661026, + 0.18530438840389252 + ], + "credit_map": [ + 0.008273713290691376, + 0.008451182395219803, + 0.007083156146109104, + 0.007610453758388758, + 0.011238574050366879, + 0.0086432546377182, + 0.008841882459819317, + 0.008791409432888031, + 0.013755517080426216, + 0.01462620496749878, + 0.017238328233361244, + 0.027141772210597992, + 0.011444371193647385, + 0.020766884088516235, + 0.014281445182859898, + 0.014270558021962643, + 0.004478881135582924, + 0.005047667771577835, + 0.005188743118196726, + 0.006016717292368412, + 0.009109259583055973, + 0.0065651037730276585, + 0.005779014900326729, + 0.005547432694584131, + 0.010481053963303566, + 0.009416413493454456, + 0.013631602749228477, + 0.010282442905008793, + 0.01448578480631113, + 0.021327054128050804, + 0.012590538710355759, + 0.01461341418325901, + 0.007359880022704601, + 0.010729226283729076, + 0.006082028616219759, + 0.00898824818432331, + 0.005136558320373297, + 0.004472354892641306, + 0.004799768328666687, + 0.005816055461764336, + 0.01223527267575264, + 0.013201933354139328, + 0.017152506858110428, + 0.009925032034516335, + 0.019314268603920937, + 0.014951881021261215, + 0.013944230042397976, + 0.01517584640532732, + 0.014753691852092743, + 0.013967513106763363, + 0.012881774455308914, + 0.016567640006542206, + 0.011440346017479897, + 0.010346485301852226, + 0.01228283904492855, + 0.009897616691887379, + 0.006913620512932539, + 0.008995766751468182, + 0.009289284236729145, + 0.016514962539076805, + 0.013071393594145775, + 0.009901740588247776, + 0.010498200543224812, + 0.00802511814981699, + 0.01078209187835455, + 0.010021122172474861, + 0.01170885656028986, + 0.010193778201937675, + 0.022734075784683228, + 0.017255762591958046, + 0.011799927800893784, + 0.008547089993953705, + 0.018938006833195686, + 0.024284740909934044, + 0.015703899785876274, + 0.014538217335939407, + 0.03747858107089996, + 0.02570478431880474, + 0.028042711317539215, + 0.02061343938112259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.020579621195793152 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028147385455667973 + ], + "opposite_logprob_tokens": [ + -5.877814769744873 + ], + "logprob_reward": -0.0028147385455667973, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028147385455667973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11595099419355392, + 0.12049399316310883, + 0.05533425882458687, + 0.09106690436601639, + 0.07806987315416336, + 0.08766595274209976, + 0.15378141403198242, + 0.07900960743427277, + 0.10307970643043518, + 0.1155472993850708 + ], + "credit_map": [ + 0.015310905873775482, + 0.011710704304277897, + 0.009375984780490398, + 0.01174331083893776, + 0.023421378806233406, + 0.015718523412942886, + 0.014629948884248734, + 0.014040232636034489, + 0.012570352293550968, + 0.009642655961215496, + 0.020649278536438942, + 0.02542542852461338, + 0.010080644860863686, + 0.019232043996453285, + 0.01067302841693163, + 0.01222055684775114, + 0.005577400326728821, + 0.0078038969077169895, + 0.006674314383417368, + 0.006122804246842861, + 0.007168835960328579, + 0.007246024440973997, + 0.007319496013224125, + 0.007421488407999277, + 0.012866362929344177, + 0.010493445210158825, + 0.010142666287720203, + 0.014469891786575317, + 0.008333977311849594, + 0.011269102804362774, + 0.009262694045901299, + 0.014228769578039646, + 0.008153746835887432, + 0.01929110661149025, + 0.00739166047424078, + 0.011294812895357609, + 0.009595478884875774, + 0.005773826502263546, + 0.009511098265647888, + 0.007058148272335529, + 0.009891077876091003, + 0.00895240530371666, + 0.008457805961370468, + 0.008596333675086498, + 0.013531164266169071, + 0.015619167126715183, + 0.013430251739919186, + 0.009187747724354267, + 0.018970422446727753, + 0.012058710679411888, + 0.04469544440507889, + 0.037079211324453354, + 0.009545273147523403, + 0.010364691726863384, + 0.012767421081662178, + 0.00830024667084217, + 0.006259756162762642, + 0.008091036230325699, + 0.0087643563747406, + 0.01773393526673317, + 0.010890595614910126, + 0.00864203181117773, + 0.010330846533179283, + 0.008297047577798367, + 0.012559527531266212, + 0.013424822129309177, + 0.01321189571171999, + 0.012423473410308361, + 0.015145765617489815, + 0.011455172672867775, + 0.012848141603171825, + 0.012010904029011726, + 0.010677976533770561, + 0.016513099893927574, + 0.008798205293715, + 0.009046544320881367, + 0.028818974271416664, + 0.016954006627202034, + 0.012999641709029675, + 0.011738848872482777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.0028147385455667973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863177701830864 + ], + "opposite_logprob_tokens": [ + -4.758631706237793 + ], + "logprob_reward": -0.00863177701830864, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863177701830864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11178939789533615, + 0.11624390631914139, + 0.06783594936132431, + 0.08114006370306015, + 0.09010764956474304, + 0.06617943942546844, + 0.24606896936893463, + 0.09187227487564087, + 0.0653175339102745, + 0.06344479322433472 + ], + "credit_map": [ + 0.018127718940377235, + 0.010957376100122929, + 0.009477492421865463, + 0.010088772512972355, + 0.01767887733876705, + 0.015156837180256844, + 0.012521211989223957, + 0.01778111234307289, + 0.011340414173901081, + 0.012669563293457031, + 0.02061629854142666, + 0.019882503896951675, + 0.009351341985166073, + 0.023223331198096275, + 0.009000969119369984, + 0.010159485973417759, + 0.012372389435768127, + 0.008002905175089836, + 0.008974893018603325, + 0.00815640203654766, + 0.007092464715242386, + 0.01081851590424776, + 0.0060563720762729645, + 0.006362011656165123, + 0.007496765814721584, + 0.005859896074980497, + 0.013650206848978996, + 0.005253033712506294, + 0.007358741480857134, + 0.02080756053328514, + 0.0092911496758461, + 0.011422708630561829, + 0.008314964361488819, + 0.011001215316355228, + 0.01014302670955658, + 0.008793429471552372, + 0.01164539810270071, + 0.011674501933157444, + 0.011495526880025864, + 0.017039578408002853, + 0.0060717519372701645, + 0.006861486937850714, + 0.005955358501523733, + 0.0073530483059585094, + 0.011370443738996983, + 0.009043462574481964, + 0.009916025213897228, + 0.009607864543795586, + 0.018474716693162918, + 0.014217445626854897, + 0.059382833540439606, + 0.0864514634013176, + 0.027361437678337097, + 0.013657940551638603, + 0.015877127647399902, + 0.010646024718880653, + 0.0072303456254303455, + 0.013126102276146412, + 0.009625883772969246, + 0.016473012045025826, + 0.010571220889687538, + 0.011694164015352726, + 0.010709952563047409, + 0.01244159322232008, + 0.007434239611029625, + 0.007832146249711514, + 0.00937000010162592, + 0.00608971668407321, + 0.009427892044186592, + 0.008520503528416157, + 0.0074616689234972, + 0.009181363508105278, + 0.007213477976620197, + 0.007320426404476166, + 0.005488932132720947, + 0.00539435213431716, + 0.01463033352047205, + 0.011087235994637012, + 0.006615221034735441, + 0.005694811698049307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.00863177701830864 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of a sports documentary, a low-angle, high-speed camera tracking shot, held close to the ground, captures in extreme slow motion the moment a professional motorcyclist conquers a precipitous mountain curve. Radiating the golden glow of the late afternoon sun, the athlete, clad in an aerodynamic black leather racing suit, its form-fitting leather embellished with striking red and white stripes, leans toward the apex at an improbable angle, his knees, covered in heavily worn sliders, almost parallel to the rough, dark gray asphalt. He controls a matte gray modern racing car, its lines sharp and its suspension compressed to the extreme by the immense centrifugal force. As the motorcycle leans to its limit, the motorcycle's metal footpegs and the athlete's knee sliders scrape against the road, instantly sending forth a shower of dazzling sparks. Under the extreme slow-motion particle effects, each spark is clearly visible. Like blazing miniature comets, they splash backwards with bright orange-yellow trails, briefly illuminating the sidewall of the rear tire that is spinning at high speed and slightly deformed by pressure, while stirring up a small amount of rubber and dust particles suspended in the warm air. The whole picture is full of the ultimate beauty of power, concentration and speed, creating a heart-pounding atmosphere of tension and awe.", + "seed": 19466, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_14/index_9256_seed_19466_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11034523202106357, + "vlm_logprob": -0.11034523202106357, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.11034523202106357, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35143232345581055 + ], + "opposite_logprob_tokens": [ + -1.2264323234558105 + ], + "logprob_reward": -0.35143232345581055, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35143232345581055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25580188632011414, + 0.04371267557144165, + 0.03791011869907379, + 0.02934325486421585, + 0.02520139515399933, + 0.31779950857162476, + 0.03630393370985985, + 0.038382094353437424, + 0.1074739620089531, + 0.10807118564844131 + ], + "credit_map": [ + 0.02545505203306675, + 0.03807199001312256, + 0.05374689772725105, + 0.04085393622517586, + 0.022926952689886093, + 0.020512616261839867, + 0.03349355608224869, + 0.020740900188684464, + 0.007617773953825235, + 0.005282143130898476, + 0.003943366929888725, + 0.0055853221565485, + 0.0050043207593262196, + 0.007164251059293747, + 0.0036221921909600496, + 0.0054933056235313416, + 0.004665979649871588, + 0.006338079925626516, + 0.004523118492215872, + 0.003108900971710682, + 0.005765543319284916, + 0.004454915411770344, + 0.004433127585798502, + 0.004620456136763096, + 0.0030019942205399275, + 0.0041751554235816, + 0.0040430063381791115, + 0.0034263101406395435, + 0.0036813467741012573, + 0.0032012874726206064, + 0.004660297650843859, + 0.0031538582406938076, + 0.0035553432535380125, + 0.004013214726001024, + 0.002448893152177334, + 0.002966727362945676, + 0.002818998182192445, + 0.0035180170089006424, + 0.003420314285904169, + 0.0024598881136626005, + 0.02567797712981701, + 0.017190558835864067, + 0.12238796055316925, + 0.021996958181262016, + 0.017832791432738304, + 0.08678906410932541, + 0.015455085784196854, + 0.010469138622283936, + 0.0033499340061098337, + 0.0050692311488091946, + 0.004359950311481953, + 0.006737831514328718, + 0.004052778240293264, + 0.004128964617848396, + 0.004252796992659569, + 0.004352447111159563, + 0.0062541174702346325, + 0.005308423191308975, + 0.00500688748434186, + 0.004906618967652321, + 0.004381938371807337, + 0.0038783210329711437, + 0.004681696649640799, + 0.003964090719819069, + 0.01852044276893139, + 0.021969426423311234, + 0.013942117802798748, + 0.00846896506845951, + 0.015597578138113022, + 0.006494276225566864, + 0.016702286899089813, + 0.005778873339295387, + 0.032393861562013626, + 0.030136413872241974, + 0.0046342951245605946, + 0.005522375460714102, + 0.012702876701951027, + 0.006232487037777901, + 0.012381115928292274, + 0.004067760892212391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.35143232345581055 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17993058264255524 + ], + "opposite_logprob_tokens": [ + -1.804930567741394 + ], + "logprob_reward": -0.17993058264255524, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17993058264255524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17556816339492798, + 0.12234163284301758, + 0.1046450212597847, + 0.04639841616153717, + 0.12322553992271423, + 0.08429605513811111, + 0.03814942017197609, + 0.0329502634704113, + 0.10103238373994827, + 0.17139309644699097 + ], + "credit_map": [ + 0.018415359780192375, + 0.02721073292195797, + 0.029601935297250748, + 0.019885452464222908, + 0.0180533304810524, + 0.018217984586954117, + 0.0256781205534935, + 0.018505262210965157, + 0.01684584841132164, + 0.01762811467051506, + 0.011312931776046753, + 0.011141456663608551, + 0.01574006862938404, + 0.0197310633957386, + 0.010917635634541512, + 0.01902451552450657, + 0.01273790467530489, + 0.016069745644927025, + 0.012872333638370037, + 0.01288506481796503, + 0.013459968380630016, + 0.0133764473721385, + 0.015235284343361855, + 0.008008268661797047, + 0.004621446598321199, + 0.006841352209448814, + 0.005553503055125475, + 0.004906481131911278, + 0.008327479474246502, + 0.004877567756921053, + 0.007856171578168869, + 0.003414412960410118, + 0.01358709018677473, + 0.015622303821146488, + 0.019362201914191246, + 0.017418567091226578, + 0.012464533559978008, + 0.018305493518710136, + 0.01469232514500618, + 0.011773028410971165, + 0.006489447318017483, + 0.004585553891956806, + 0.02739216387271881, + 0.006686792708933353, + 0.007265081163495779, + 0.022088119760155678, + 0.006506414618343115, + 0.0032824829686433077, + 0.0036765681579709053, + 0.0048208278603851795, + 0.0057756067253649235, + 0.004943286534398794, + 0.0030045360326766968, + 0.007519450504332781, + 0.0037151118740439415, + 0.004694030620157719, + 0.0044287992641329765, + 0.005330471321940422, + 0.0044938367791473866, + 0.0063532073982059956, + 0.0042928531765937805, + 0.003003645921126008, + 0.0027268575504422188, + 0.0023205920588225126, + 0.01771383360028267, + 0.01952381245791912, + 0.013396612368524075, + 0.00755980284884572, + 0.016001461073756218, + 0.008799402974545956, + 0.012462847866117954, + 0.0055746054276824, + 0.040833260864019394, + 0.039221201092004776, + 0.009766715578734875, + 0.013289923779666424, + 0.025272179394960403, + 0.009695874527096748, + 0.023844711482524872, + 0.009469220414757729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.17993058264255524 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978849083185196 + ], + "opposite_logprob_tokens": [ + -4.634788513183594 + ], + "logprob_reward": -0.00978849083185196, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978849083185196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20358115434646606, + 0.03159893676638603, + 0.0389467291533947, + 0.0395645909011364, + 0.05858512222766876, + 0.05936910957098007, + 0.09963847696781158, + 0.07327567785978317, + 0.206159308552742, + 0.1892809271812439 + ], + "credit_map": [ + 0.022293567657470703, + 0.038159046322107315, + 0.03800811991095543, + 0.04862166568636894, + 0.012257395312190056, + 0.011616063304245472, + 0.0193356741219759, + 0.013289622962474823, + 0.007414297666400671, + 0.004853027407079935, + 0.00260482938028872, + 0.003590974723920226, + 0.002967443550005555, + 0.003257310949265957, + 0.002552008256316185, + 0.004359041340649128, + 0.004044083412736654, + 0.00570646533742547, + 0.005884766578674316, + 0.0034386683255434036, + 0.004700776655226946, + 0.005585161969065666, + 0.0055206334218382835, + 0.004066175315529108, + 0.0046089268289506435, + 0.004684664309024811, + 0.005434281658381224, + 0.004441067110747099, + 0.006648026406764984, + 0.003675648709759116, + 0.006161949597299099, + 0.003910025581717491, + 0.004859571810811758, + 0.008439294062554836, + 0.006353625096380711, + 0.006262892857193947, + 0.00755001325160265, + 0.011207614094018936, + 0.007590828463435173, + 0.006321283057332039, + 0.00661676237359643, + 0.005441925022751093, + 0.014484954997897148, + 0.0049240305088460445, + 0.004955727607011795, + 0.012448412366211414, + 0.005605650134384632, + 0.0048916456289589405, + 0.009977270849049091, + 0.011620929464697838, + 0.014334060251712799, + 0.013810384087264538, + 0.009376567788422108, + 0.01559421420097351, + 0.011902923695743084, + 0.013022121973335743, + 0.01067835371941328, + 0.009088778868317604, + 0.009118559770286083, + 0.00870845839381218, + 0.009677606634795666, + 0.006971871946007013, + 0.009740571491420269, + 0.009291472844779491, + 0.04272707924246788, + 0.044260259717702866, + 0.02143743634223938, + 0.014684319496154785, + 0.02727684937417507, + 0.015420271083712578, + 0.027746059000492096, + 0.012607034295797348, + 0.03918805718421936, + 0.0356927327811718, + 0.010702344588935375, + 0.014158039353787899, + 0.031826891005039215, + 0.01285557821393013, + 0.034182991832494736, + 0.010674286633729935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.00978849083185196 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019393223337829113 + ], + "opposite_logprob_tokens": [ + -6.251939296722412 + ], + "logprob_reward": -0.0019393223337829113, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019393223337829113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2819099426269531, + 0.031171226873993874, + 0.04665583372116089, + 0.06347326934337616, + 0.07816099375486374, + 0.05268373340368271, + 0.11701172590255737, + 0.10008732229471207, + 0.15844199061393738, + 0.07040393352508545 + ], + "credit_map": [ + 0.03068782575428486, + 0.05002661049365997, + 0.06357945501804352, + 0.06317466497421265, + 0.018671875819563866, + 0.013586306944489479, + 0.024814115837216377, + 0.017369123175740242, + 0.005208536982536316, + 0.004302833694964647, + 0.002737109549343586, + 0.004096510354429483, + 0.004245745483785868, + 0.004651813767850399, + 0.0026308796368539333, + 0.0032977983355522156, + 0.007132972124963999, + 0.006551261525601149, + 0.0061295488849282265, + 0.005446149501949549, + 0.005966581404209137, + 0.005825147498399019, + 0.005882710684090853, + 0.003721464890986681, + 0.00864863209426403, + 0.006475070957094431, + 0.010611198842525482, + 0.0059709325432777405, + 0.010607106611132622, + 0.006394789554178715, + 0.008954033255577087, + 0.00581150408834219, + 0.008460947312414646, + 0.011639912612736225, + 0.009830520488321781, + 0.009649820625782013, + 0.009686043485999107, + 0.011158228851854801, + 0.01035663578659296, + 0.007378887850791216, + 0.00584336556494236, + 0.003725259331986308, + 0.01429535448551178, + 0.0039393724873661995, + 0.00389107852242887, + 0.013253552839159966, + 0.004451186396181583, + 0.0032845670357346535, + 0.010957958176732063, + 0.0079929418861866, + 0.020303204655647278, + 0.013254936784505844, + 0.009473748505115509, + 0.014502289704978466, + 0.017512930557131767, + 0.02301372028887272, + 0.013187115080654621, + 0.013088972307741642, + 0.012405071407556534, + 0.010919405147433281, + 0.01337627787142992, + 0.00948125496506691, + 0.014812500216066837, + 0.012816722504794598, + 0.02925630658864975, + 0.039531975984573364, + 0.018970446661114693, + 0.012560820206999779, + 0.01719425804913044, + 0.0123883206397295, + 0.016486262902617455, + 0.012053604237735271, + 0.012766744010150433, + 0.012041657231748104, + 0.005935864057391882, + 0.006847022101283073, + 0.009451444260776043, + 0.006014741491526365, + 0.013683045282959938, + 0.0036634167190641165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.0019393223337829113 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635440841317177 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.008635440841317177, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635440841317177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20617495477199554, + 0.0413309708237648, + 0.065520279109478, + 0.04769070819020271, + 0.05545022338628769, + 0.202895388007164, + 0.07650024443864822, + 0.05282345041632652, + 0.1431724578142166, + 0.10844124853610992 + ], + "credit_map": [ + 0.02818991430103779, + 0.039706695824861526, + 0.03244847059249878, + 0.046981751918792725, + 0.01306293997913599, + 0.016258930787444115, + 0.01639140024781227, + 0.013134857639670372, + 0.009856585413217545, + 0.004766132682561874, + 0.004834017716348171, + 0.004806176759302616, + 0.0035085207782685757, + 0.0046390206553041935, + 0.0030674822628498077, + 0.005853033624589443, + 0.006113818380981684, + 0.010277651250362396, + 0.00738977175205946, + 0.006285507697612047, + 0.009735433384776115, + 0.01173668634146452, + 0.007476441096514463, + 0.006504973862320185, + 0.008183385245501995, + 0.004170732572674751, + 0.005413321778178215, + 0.004087546840310097, + 0.007363026030361652, + 0.004713416565209627, + 0.00790927093476057, + 0.005850012879818678, + 0.007136262021958828, + 0.0069746901281178, + 0.0067068906500935555, + 0.006747403182089329, + 0.0070921992883086205, + 0.010258199647068977, + 0.0062393369153141975, + 0.0042952485382556915, + 0.01676027849316597, + 0.011858527548611164, + 0.07924579083919525, + 0.013525126501917839, + 0.009322259575128555, + 0.057302817702293396, + 0.009052099660038948, + 0.005828509572893381, + 0.005586825776845217, + 0.004816157277673483, + 0.012585125863552094, + 0.007618563715368509, + 0.005731245037168264, + 0.016504773870110512, + 0.010111055336892605, + 0.013546505942940712, + 0.007299468386918306, + 0.006826287135481834, + 0.007458209525793791, + 0.00632628146559, + 0.007437021937221289, + 0.005473771132528782, + 0.005462212022393942, + 0.006540201138705015, + 0.025842996314167976, + 0.03363349661231041, + 0.015834731981158257, + 0.012307479977607727, + 0.01761796325445175, + 0.012623869813978672, + 0.016139209270477295, + 0.009172718971967697, + 0.03326006606221199, + 0.030421581119298935, + 0.00480189872905612, + 0.006360654719173908, + 0.010581742972135544, + 0.00991594698280096, + 0.008375589735805988, + 0.004723773803561926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.008635440841317177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of a sports documentary, a low-angle, high-speed camera tracking shot, held close to the ground, captures in extreme slow motion the moment a professional motorcyclist conquers a precipitous mountain curve. Radiating the golden glow of the late afternoon sun, the athlete, clad in an aerodynamic black leather racing suit, its form-fitting leather embellished with striking red and white stripes, leans toward the apex at an improbable angle, his knees, covered in heavily worn sliders, almost parallel to the rough, dark gray asphalt. He controls a matte gray modern racing car, its lines sharp and its suspension compressed to the extreme by the immense centrifugal force. As the motorcycle leans to its limit, the motorcycle's metal footpegs and the athlete's knee sliders scrape against the road, instantly sending forth a shower of dazzling sparks. Under the extreme slow-motion particle effects, each spark is clearly visible. Like blazing miniature comets, they splash backwards with bright orange-yellow trails, briefly illuminating the sidewall of the rear tire that is spinning at high speed and slightly deformed by pressure, while stirring up a small amount of rubber and dust particles suspended in the warm air. The whole picture is full of the ultimate beauty of power, concentration and speed, creating a heart-pounding atmosphere of tension and awe.", + "seed": 19467, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_14/index_9256_seed_19467_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09375152802094817, + "vlm_logprob": -0.09375152802094817, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.09375152802094817, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28384068608283997 + ], + "opposite_logprob_tokens": [ + -1.4088406562805176 + ], + "logprob_reward": -0.28384068608283997, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28384068608283997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029598411172628403, + 0.004644407425075769, + 0.04536456987261772, + 0.014392374083399773, + 0.041072599589824677, + 0.015264172106981277, + 0.014214974828064442, + 0.05683497339487076, + 0.020240388810634613, + 0.7583730816841125 + ], + "credit_map": [ + 0.0033469051122665405, + 0.00492447055876255, + 0.0030982799362391233, + 0.005019808653742075, + 0.0032514980994164944, + 0.0027757720090448856, + 0.003712001023814082, + 0.003469674149528146, + 0.0006448558415286243, + 0.0006186682148836553, + 0.0007781547610647976, + 0.0005234336713328958, + 0.0003577081370167434, + 0.0005760378553532064, + 0.00037752490607090294, + 0.0007680241833440959, + 0.004596122540533543, + 0.010107608512043953, + 0.004169231280684471, + 0.006901201326400042, + 0.0071586016565561295, + 0.0043936860747635365, + 0.0030362482648342848, + 0.005001869052648544, + 0.0011415282497182488, + 0.0006402707076631486, + 0.0016094042221084237, + 0.0009777527302503586, + 0.0008201596210710704, + 0.0036117106210440397, + 0.000864008441567421, + 0.004727539140731096, + 0.004153560847043991, + 0.0027150509413331747, + 0.009318037889897823, + 0.0026429209392517805, + 0.0031112837605178356, + 0.00858911219984293, + 0.0027906070463359356, + 0.007752024568617344, + 0.0014028220903128386, + 0.0026368615217506886, + 0.0020982089918106794, + 0.001609237864613533, + 0.0011673270491883159, + 0.0020946869626641273, + 0.0025641038082540035, + 0.0016909238183870912, + 0.0013456816086545587, + 0.0014482017140835524, + 0.0020961950067430735, + 0.0012275924673303962, + 0.001854799105785787, + 0.0019220581743866205, + 0.0027992515824735165, + 0.0015211962163448334, + 0.006558472756296396, + 0.004337956663221121, + 0.010408974252641201, + 0.011692509055137634, + 0.006518464535474777, + 0.0069082207046449184, + 0.004910598509013653, + 0.0054997787810862064, + 0.0012010537320747972, + 0.0015914886025711894, + 0.001454801531508565, + 0.00238799792714417, + 0.0012213431764394045, + 0.0012854369124397635, + 0.0067448397167027, + 0.0043534268625080585, + 0.2716001570224762, + 0.02981097809970379, + 0.31318047642707825, + 0.039983443915843964, + 0.013558768667280674, + 0.018091566860675812, + 0.03360278159379959, + 0.038544911891222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.28384068608283997 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603996902704239 + ], + "opposite_logprob_tokens": [ + -1.9103996753692627 + ], + "logprob_reward": -0.1603996902704239, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603996902704239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04592561349272728, + 0.03516948223114014, + 0.028468674048781395, + 0.019397372379899025, + 0.02340320684015751, + 0.0684056431055069, + 0.020641963928937912, + 0.22059257328510284, + 0.054781507700681686, + 0.4832139015197754 + ], + "credit_map": [ + 0.005874432157725096, + 0.005396770779043436, + 0.005583552643656731, + 0.00684963958337903, + 0.0036524534225463867, + 0.005336083006113768, + 0.0033531386870890856, + 0.009879544377326965, + 0.004502824041992426, + 0.004656943958252668, + 0.004302294924855232, + 0.0041267056949436665, + 0.005504246335476637, + 0.0037318128161132336, + 0.003495412180200219, + 0.004849240183830261, + 0.003970074467360973, + 0.00461155641824007, + 0.0036002222914248705, + 0.004261441063135862, + 0.0032755897846072912, + 0.002761524636298418, + 0.0019443611381575465, + 0.004043904598802328, + 0.003175055142492056, + 0.001994180493056774, + 0.0026886111591011286, + 0.0020023270044475794, + 0.0016133686294779181, + 0.0031690269242972136, + 0.0013892316492274404, + 0.003365571377798915, + 0.0028255293145775795, + 0.002457693452015519, + 0.0024826861917972565, + 0.00273920432664454, + 0.0030791598837822676, + 0.0029268194921314716, + 0.0034558172337710857, + 0.003436295548453927, + 0.0057798121124506, + 0.0132368765771389, + 0.012521718628704548, + 0.006177187897264957, + 0.0065046981908380985, + 0.006958272773772478, + 0.009134059771895409, + 0.008093016222119331, + 0.0027676282916218042, + 0.0031656422652304173, + 0.002470677951350808, + 0.0027363761328160763, + 0.002539301523938775, + 0.002192129846662283, + 0.002120599616318941, + 0.0026496080681681633, + 0.03335980325937271, + 0.014643877744674683, + 0.055336978286504745, + 0.03355447202920914, + 0.020479198545217514, + 0.03234633803367615, + 0.014201113022863865, + 0.016670802608132362, + 0.004889624193310738, + 0.00644133472815156, + 0.005012562498450279, + 0.007059122901409864, + 0.005259297322481871, + 0.0076546333730220795, + 0.014328006654977798, + 0.004136923234909773, + 0.16637715697288513, + 0.019100915640592575, + 0.19269277155399323, + 0.027975184842944145, + 0.011676592752337456, + 0.013646057806909084, + 0.025133375078439713, + 0.02661183662712574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.1603996902704239 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544653378427029 + ], + "opposite_logprob_tokens": [ + -4.387544631958008 + ], + "logprob_reward": -0.012544653378427029, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544653378427029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03630759194493294, + 0.040290478616952896, + 0.030963292345404625, + 0.0359291136264801, + 0.07979130744934082, + 0.0928519070148468, + 0.060227829962968826, + 0.025574229657649994, + 0.14032195508480072, + 0.4577423334121704 + ], + "credit_map": [ + 0.005242487881332636, + 0.004474218934774399, + 0.004103417508304119, + 0.005411607213318348, + 0.004144834354519844, + 0.0036768517456948757, + 0.0038362646009773016, + 0.005417910870164633, + 0.004282460547983646, + 0.006060695741325617, + 0.0058068581856787205, + 0.007395205087959766, + 0.0044253915548324585, + 0.003607501508668065, + 0.0033874076325446367, + 0.005324959754943848, + 0.0038508123252540827, + 0.0056924945674836636, + 0.003259476739913225, + 0.005029399413615465, + 0.004772694315761328, + 0.002728862687945366, + 0.002032486256211996, + 0.0035970648750662804, + 0.0038599506951868534, + 0.0019644026178866625, + 0.006256104912608862, + 0.0033779803197830915, + 0.002243069000542164, + 0.007516067009419203, + 0.001957044703885913, + 0.008754493668675423, + 0.006113344803452492, + 0.0057601286098361015, + 0.013139021582901478, + 0.0054548801854252815, + 0.003767105285078287, + 0.006929105147719383, + 0.029030898585915565, + 0.009596815332770348, + 0.01921221613883972, + 0.011330920271575451, + 0.016535311937332153, + 0.010363472625613213, + 0.007133170962333679, + 0.009306523948907852, + 0.010301923379302025, + 0.008668366819620132, + 0.005834228824824095, + 0.004945030901581049, + 0.008232430554926395, + 0.005849186331033707, + 0.007210976909846067, + 0.008076013065874577, + 0.011043821461498737, + 0.009036143310368061, + 0.0028926145751029253, + 0.002278171246871352, + 0.0035342974588274956, + 0.004032401368021965, + 0.0037112832069396973, + 0.003471922595053911, + 0.002604994224384427, + 0.003048545913770795, + 0.007937935180962086, + 0.008223568089306355, + 0.008120168000459671, + 0.0167962945997715, + 0.008585409261286259, + 0.010114570148289204, + 0.050923701375722885, + 0.029620304703712463, + 0.1644836664199829, + 0.018124716356396675, + 0.1886962652206421, + 0.024548104032874107, + 0.009443862363696098, + 0.01091447938233614, + 0.019488859921693802, + 0.022042399272322655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.012544653378427029 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002195806708186865 + ], + "opposite_logprob_tokens": [ + -6.127195835113525 + ], + "logprob_reward": -0.002195806708186865, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002195806708186865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024976680055260658, + 0.008819131180644035, + 0.016166236251592636, + 0.011784779839217663, + 0.0257902592420578, + 0.037678614258766174, + 0.011473353020846844, + 0.026525797322392464, + 0.05490117147564888, + 0.781883955001831 + ], + "credit_map": [ + 0.004253311548382044, + 0.001885642996057868, + 0.0037508627865463495, + 0.0028628441505134106, + 0.0021323931869119406, + 0.0024945258628576994, + 0.002217171946540475, + 0.00537992874160409, + 0.0008968590991571546, + 0.0010818801820278168, + 0.000881330284755677, + 0.001858070376329124, + 0.0009261919185519218, + 0.0008911543409340084, + 0.0008835135959088802, + 0.0014001319650560617, + 0.0018600221956148744, + 0.0034040252212435007, + 0.0016493024304509163, + 0.002725288737565279, + 0.002367428969591856, + 0.0014744424261152744, + 0.0010894698789343238, + 0.001596257439814508, + 0.0013579399092122912, + 0.0005273845163173974, + 0.0017200206639245152, + 0.0009774185018613935, + 0.00046639691572636366, + 0.0024678159970790148, + 0.0005455738864839077, + 0.003722231136634946, + 0.0022445409558713436, + 0.0021532801911234856, + 0.005676666274666786, + 0.0021425862796604633, + 0.0017398822819814086, + 0.002170267514884472, + 0.008137204684317112, + 0.0015258300118148327, + 0.006640399340540171, + 0.004803638439625502, + 0.010011267848312855, + 0.003913562279194593, + 0.002942244755104184, + 0.003042287193238735, + 0.0033839521929621696, + 0.002941262209787965, + 0.001231424743309617, + 0.0015486974734812975, + 0.0014677203726023436, + 0.0012929945951327682, + 0.0013909603003412485, + 0.0017256560968235135, + 0.0012927219504490495, + 0.001523178187198937, + 0.003371774684637785, + 0.002054882701486349, + 0.004425916355103254, + 0.004524604883044958, + 0.003351568477228284, + 0.003418290987610817, + 0.0024948464706540108, + 0.0028839141596108675, + 0.004395367112010717, + 0.005949902348220348, + 0.005269037559628487, + 0.007177126593887806, + 0.005743223242461681, + 0.0038632231298834085, + 0.013726036995649338, + 0.00877725426107645, + 0.2799992561340332, + 0.03118581883609295, + 0.32299163937568665, + 0.042450275272130966, + 0.01337513793259859, + 0.018416928127408028, + 0.03429342806339264, + 0.039171554148197174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.002195806708186865 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977680366486311 + ], + "opposite_logprob_tokens": [ + -4.634776592254639 + ], + "logprob_reward": -0.00977680366486311, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977680366486311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03586014732718468, + 0.01252271980047226, + 0.06299220770597458, + 0.049336791038513184, + 0.016336429864168167, + 0.012681620195508003, + 0.017000878229737282, + 0.030972488224506378, + 0.09619324654340744, + 0.6661034226417542 + ], + "credit_map": [ + 0.003995048347860575, + 0.0036495605017989874, + 0.0040135071612894535, + 0.004746242426335812, + 0.004241279326379299, + 0.0034296968951821327, + 0.003947615623474121, + 0.007837200537323952, + 0.0011954345973208547, + 0.0014900334645062685, + 0.001881699077785015, + 0.0020121836569160223, + 0.001155390520580113, + 0.0012603085488080978, + 0.0009276755736209452, + 0.0025999960489571095, + 0.007430320605635643, + 0.015285874716937542, + 0.005945596378296614, + 0.008837704546749592, + 0.00901736319065094, + 0.005895530804991722, + 0.004060533829033375, + 0.0065192896872758865, + 0.003131068777292967, + 0.0022267955355346203, + 0.0051619429141283035, + 0.0033134282566607, + 0.0030340957455337048, + 0.013066884130239487, + 0.002225317293778062, + 0.017177265137434006, + 0.0018599851755425334, + 0.001242978498339653, + 0.0012330503668636084, + 0.000930732348933816, + 0.001068995101377368, + 0.0049012163653969765, + 0.0036368495784699917, + 0.0014626228949055076, + 0.0009764328715391457, + 0.001978892832994461, + 0.0028679685201495886, + 0.0014090035110712051, + 0.0014582774601876736, + 0.0014422689564526081, + 0.0013086905237287283, + 0.0012400869745761156, + 0.0018465404864400625, + 0.002154130721464753, + 0.002447141334414482, + 0.0021720807999372482, + 0.002443829318508506, + 0.0014917639782652259, + 0.0026911343447864056, + 0.0017542591085657477, + 0.0038760355673730373, + 0.002981380093842745, + 0.004414644092321396, + 0.0037280882243067026, + 0.005718099884688854, + 0.004491397645324469, + 0.0027019830886274576, + 0.0030608626548200846, + 0.0061714379116892815, + 0.009366260841488838, + 0.008477612398564816, + 0.011668854393064976, + 0.009067447856068611, + 0.006437927484512329, + 0.0269318837672472, + 0.018071826547384262, + 0.23997625708580017, + 0.026402220129966736, + 0.2758956253528595, + 0.03577413409948349, + 0.011378495953977108, + 0.015383325517177582, + 0.028869498521089554, + 0.03242393210530281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.00977680366486311 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of a sports documentary, a low-angle, high-speed camera tracking shot, held close to the ground, captures in extreme slow motion the moment a professional motorcyclist conquers a precipitous mountain curve. Radiating the golden glow of the late afternoon sun, the athlete, clad in an aerodynamic black leather racing suit, its form-fitting leather embellished with striking red and white stripes, leans toward the apex at an improbable angle, his knees, covered in heavily worn sliders, almost parallel to the rough, dark gray asphalt. He controls a matte gray modern racing car, its lines sharp and its suspension compressed to the extreme by the immense centrifugal force. As the motorcycle leans to its limit, the motorcycle's metal footpegs and the athlete's knee sliders scrape against the road, instantly sending forth a shower of dazzling sparks. Under the extreme slow-motion particle effects, each spark is clearly visible. Like blazing miniature comets, they splash backwards with bright orange-yellow trails, briefly illuminating the sidewall of the rear tire that is spinning at high speed and slightly deformed by pressure, while stirring up a small amount of rubber and dust particles suspended in the warm air. The whole picture is full of the ultimate beauty of power, concentration and speed, creating a heart-pounding atmosphere of tension and awe.", + "seed": 19468, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_14/index_9256_seed_19468_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14577048951759936, + "vlm_logprob": -0.14577048951759936, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.14577048951759936, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2543802559375763 + ], + "opposite_logprob_tokens": [ + -1.504380226135254 + ], + "logprob_reward": -0.2543802559375763, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2543802559375763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0614214763045311, + 0.05600482597947121, + 0.08123405277729034, + 0.06082088127732277, + 0.03731607273221016, + 0.04142477363348007, + 0.03459107503294945, + 0.029324239119887352, + 0.3861055076122284, + 0.21175703406333923 + ], + "credit_map": [ + 0.006889158394187689, + 0.006631949916481972, + 0.006819858681410551, + 0.008449801243841648, + 0.007533820811659098, + 0.008878407999873161, + 0.008703311905264854, + 0.007515176199376583, + 0.005127106327563524, + 0.005875538568943739, + 0.004091545473784208, + 0.0057749380357563496, + 0.003752651624381542, + 0.005977034103125334, + 0.012737437151372433, + 0.012668581679463387, + 0.009130013175308704, + 0.014687038958072662, + 0.004451833199709654, + 0.011652158573269844, + 0.004825914744287729, + 0.0045485179871320724, + 0.01674838922917843, + 0.015190188772976398, + 0.005115604493767023, + 0.009520127438008785, + 0.003582191187888384, + 0.011369348503649235, + 0.007129809819161892, + 0.006121301557868719, + 0.010696361772716045, + 0.007286140229552984, + 0.003056370420381427, + 0.004810447804629803, + 0.0035181117709726095, + 0.0063293008133769035, + 0.002950739348307252, + 0.004701875150203705, + 0.007748632691800594, + 0.004200599621981382, + 0.007689538411796093, + 0.0049048480577766895, + 0.006758346688002348, + 0.004356359597295523, + 0.0046297404915094376, + 0.005496094468981028, + 0.002745525212958455, + 0.004844324197620153, + 0.004101356957107782, + 0.005394577980041504, + 0.005392230115830898, + 0.0060984170995652676, + 0.004301848355680704, + 0.004333829041570425, + 0.0026418627239763737, + 0.00232695578597486, + 0.0034735535737127066, + 0.0024181832559406757, + 0.0032872639130800962, + 0.0038511231541633606, + 0.0034660573583096266, + 0.004765679128468037, + 0.004660493694245815, + 0.003401888068765402, + 0.03988367319107056, + 0.05010399594902992, + 0.0407402478158474, + 0.057865533977746964, + 0.04081860929727554, + 0.055522043257951736, + 0.04037487506866455, + 0.06079656258225441, + 0.06644056737422943, + 0.013299013487994671, + 0.09367828816175461, + 0.01776944287121296, + 0.0053760283626616, + 0.004855880979448557, + 0.0057625616900622845, + 0.004575274884700775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.2543802559375763 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42893946170806885 + ], + "opposite_logprob_tokens": [ + -1.0539394617080688 + ], + "logprob_reward": -0.42893946170806885, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42893946170806885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06283456832170486, + 0.017001619562506676, + 0.11062140017747879, + 0.06226731091737747, + 0.05025847256183624, + 0.06370028108358383, + 0.07606512308120728, + 0.06352528929710388, + 0.36215874552726746, + 0.13156722486019135 + ], + "credit_map": [ + 0.008005972020328045, + 0.005294345784932375, + 0.00752559769898653, + 0.007510220631957054, + 0.004722831770777702, + 0.008481212891638279, + 0.012811651453375816, + 0.008482733741402626, + 0.001692897523753345, + 0.0032430619467049837, + 0.001366844866424799, + 0.0018567906226962805, + 0.001537008211016655, + 0.002069362672045827, + 0.0025163721293210983, + 0.002719281008467078, + 0.013037626631557941, + 0.016510942950844765, + 0.009298746474087238, + 0.013424755074083805, + 0.006639353930950165, + 0.00799597054719925, + 0.023129772394895554, + 0.020584238693118095, + 0.004437747411429882, + 0.008220509625971317, + 0.004773576278239489, + 0.011045821011066437, + 0.008551840670406818, + 0.005905719939619303, + 0.012196054682135582, + 0.007136040832847357, + 0.004374985117465258, + 0.00881163403391838, + 0.0031833406537771225, + 0.007371074054390192, + 0.006554589606821537, + 0.004977790638804436, + 0.009766641072928905, + 0.005218420177698135, + 0.009081724099814892, + 0.006774360314011574, + 0.0062668416649103165, + 0.006924828048795462, + 0.008807656355202198, + 0.009872405789792538, + 0.006805439479649067, + 0.009167024865746498, + 0.0065086460672318935, + 0.008650656789541245, + 0.006580838933587074, + 0.008495161309838295, + 0.009398354217410088, + 0.012087381444871426, + 0.011176072061061859, + 0.013168004341423512, + 0.006847172509878874, + 0.007930244319140911, + 0.0067628840915858746, + 0.01023122202605009, + 0.007525850553065538, + 0.009549181908369064, + 0.007068426348268986, + 0.0076103112660348415, + 0.03654193878173828, + 0.04438335821032524, + 0.03893835470080376, + 0.05086769908666611, + 0.044847216457128525, + 0.05099796876311302, + 0.04050646722316742, + 0.05507573485374451, + 0.03928279131650925, + 0.008392786607146263, + 0.05504889786243439, + 0.010869858786463737, + 0.004515290725976229, + 0.004356041084975004, + 0.004849840886890888, + 0.004251719918102026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.42893946170806885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370697423815727 + ], + "opposite_logprob_tokens": [ + -3.4087069034576416 + ], + "logprob_reward": -0.03370697423815727, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370697423815727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19541579484939575, + 0.08494329452514648, + 0.12989361584186554, + 0.09625591337680817, + 0.1663946658372879, + 0.06943926215171814, + 0.04401165246963501, + 0.0499567985534668, + 0.07121443003416061, + 0.0924745723605156 + ], + "credit_map": [ + 0.02372051775455475, + 0.012133779004216194, + 0.018449408933520317, + 0.02066687121987343, + 0.018174054101109505, + 0.03371117636561394, + 0.03929762914776802, + 0.029262356460094452, + 0.00671021081507206, + 0.008194319903850555, + 0.012747242115437984, + 0.008298620581626892, + 0.008468396961688995, + 0.011855822056531906, + 0.016821928322315216, + 0.011846750974655151, + 0.006842669099569321, + 0.011416514404118061, + 0.0033967189956456423, + 0.0064875176176428795, + 0.005836309399455786, + 0.004753760993480682, + 0.04731019213795662, + 0.04384994134306908, + 0.00857190415263176, + 0.012558745220303535, + 0.006734792143106461, + 0.01605554297566414, + 0.011039790697395802, + 0.017320187762379646, + 0.013086365535855293, + 0.010888583026826382, + 0.010780468583106995, + 0.033702053129673004, + 0.011030913330614567, + 0.027742013335227966, + 0.011980247683823109, + 0.017619239166378975, + 0.038733240216970444, + 0.014806495979428291, + 0.0060860528610646725, + 0.0066726175136864185, + 0.00539701571688056, + 0.008827335201203823, + 0.010681540705263615, + 0.012310539372265339, + 0.005170492455363274, + 0.014293665997684002, + 0.004729133564978838, + 0.008376877754926682, + 0.007026731967926025, + 0.005653058178722858, + 0.0054840040393173695, + 0.005986143834888935, + 0.0029729772359132767, + 0.003782727289944887, + 0.003779226914048195, + 0.005746905226260424, + 0.004196155816316605, + 0.005625050514936447, + 0.008746427483856678, + 0.008000876754522324, + 0.0070013380609452724, + 0.006860812194645405, + 0.010082162916660309, + 0.007823009043931961, + 0.007102766539901495, + 0.010037827305495739, + 0.005981241352856159, + 0.009037000127136707, + 0.00821654312312603, + 0.012933874502778053, + 0.01756138913333416, + 0.007227166555821896, + 0.027138320729136467, + 0.01230417750775814, + 0.006005303934216499, + 0.00894888024777174, + 0.006240047514438629, + 0.00704928208142519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.03370697423815727 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319126108661294 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.00319126108661294, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319126108661294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09078676998615265, + 0.08471457660198212, + 0.13591735064983368, + 0.0679095983505249, + 0.06992486864328384, + 0.0294649638235569, + 0.038489047437906265, + 0.04161495342850685, + 0.04441063851118088, + 0.3967672288417816 + ], + "credit_map": [ + 0.00854637660086155, + 0.005465150345116854, + 0.010243992321193218, + 0.010101290419697762, + 0.0067612044513225555, + 0.01614765077829361, + 0.0206998772919178, + 0.012821227312088013, + 0.008062114007771015, + 0.009436274878680706, + 0.007367825135588646, + 0.009311390109360218, + 0.007081106770783663, + 0.00936552882194519, + 0.01672830432653427, + 0.01736203022301197, + 0.007275985088199377, + 0.01035866979509592, + 0.0030189715325832367, + 0.008050801232457161, + 0.005843175109475851, + 0.0040501742623746395, + 0.050897881388664246, + 0.04642169550061226, + 0.004939054138958454, + 0.009245285764336586, + 0.004133903421461582, + 0.013460339978337288, + 0.008411897346377373, + 0.007998581044375896, + 0.011091328226029873, + 0.008629213087260723, + 0.006837544962763786, + 0.011338558048009872, + 0.005691736936569214, + 0.011080583557486534, + 0.005753130652010441, + 0.010439792647957802, + 0.010243324562907219, + 0.008540199138224125, + 0.0032833192963153124, + 0.0034908156376332045, + 0.0033204026985913515, + 0.003328208113089204, + 0.003849887289106846, + 0.004767471458762884, + 0.002632314572110772, + 0.004792545456439257, + 0.003367356024682522, + 0.006496591027826071, + 0.004079440608620644, + 0.008257771842181683, + 0.0052071064710617065, + 0.005319688469171524, + 0.002719360636547208, + 0.003041731659322977, + 0.004251813981682062, + 0.005215111654251814, + 0.004056653007864952, + 0.005448432173579931, + 0.006712303962558508, + 0.005651154555380344, + 0.005372033454477787, + 0.004907452501356602, + 0.005849818699061871, + 0.004182864911854267, + 0.004048872273415327, + 0.007402433082461357, + 0.0036075825337320566, + 0.004372972995042801, + 0.005353772547096014, + 0.009592320770025253, + 0.1233975738286972, + 0.024916406720876694, + 0.17548508942127228, + 0.033682554960250854, + 0.010126544162631035, + 0.008688118308782578, + 0.010997881181538105, + 0.009473073296248913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.00319126108661294 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional motorcyclist wearing a black leather racing suit with red and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle matte gray with sharp lines and compressed suspension?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show sparks flying from the footpegs and knee sliders scraping against dark gray asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the sparks depicted as bright orange-yellow trails resembling blazing miniature comets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed camera tracking close to the ground in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634494617581367 + ], + "opposite_logprob_tokens": [ + -4.758634567260742 + ], + "logprob_reward": -0.008634494617581367, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634494617581367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14595341682434082, + 0.11427818238735199, + 0.16149291396141052, + 0.07561150938272476, + 0.10786343365907669, + 0.0709431990981102, + 0.04562585428357124, + 0.04505463317036629, + 0.07032159715890884, + 0.16285523772239685 + ], + "credit_map": [ + 0.013888556510210037, + 0.0094671119004488, + 0.01263192668557167, + 0.014332418330013752, + 0.011431361548602581, + 0.02740713767707348, + 0.033774662762880325, + 0.023020246997475624, + 0.01122137438505888, + 0.012375484220683575, + 0.008518115617334843, + 0.011540080420672894, + 0.008627261966466904, + 0.013012018986046314, + 0.02465139515697956, + 0.024332448840141296, + 0.00927591510117054, + 0.009836765006184578, + 0.005198957398533821, + 0.007947362028062344, + 0.006111491937190294, + 0.0053243027068674564, + 0.06174001097679138, + 0.05605810508131981, + 0.007468280848115683, + 0.011440916918218136, + 0.004720087628811598, + 0.013461817987263203, + 0.008680715225636959, + 0.006341344676911831, + 0.014428106136620045, + 0.009070240892469883, + 0.011343683116137981, + 0.014867694117128849, + 0.010915620252490044, + 0.01239666249603033, + 0.009378408081829548, + 0.014941303059458733, + 0.020020807161927223, + 0.013999252580106258, + 0.005756590981036425, + 0.0055608064867556095, + 0.006273038685321808, + 0.010720871388912201, + 0.01039169728755951, + 0.010655676014721394, + 0.007983393967151642, + 0.013601124286651611, + 0.005737468600273132, + 0.006054677534848452, + 0.004001033492386341, + 0.004712247289717197, + 0.007601576391607523, + 0.006836129352450371, + 0.00466399360448122, + 0.006018728017807007, + 0.005544849205762148, + 0.004597989376634359, + 0.00481657637283206, + 0.006566724739968777, + 0.005222974345088005, + 0.005790704861283302, + 0.006566887721419334, + 0.005947924219071865, + 0.007964221760630608, + 0.007340087555348873, + 0.006634427700191736, + 0.009640437550842762, + 0.009289151057600975, + 0.008481194265186787, + 0.007955940440297127, + 0.013016138225793839, + 0.03291046991944313, + 0.013141783885657787, + 0.044294826686382294, + 0.0157344788312912, + 0.014364346861839294, + 0.017811933532357216, + 0.012418913654983044, + 0.012178481556475163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.008634494617581367 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_15/index_tensor([2039, 2039])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_15/index_tensor([2039, 2039])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..901c90d7c4ffbb558ea1218f78ba6b719a04481c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_15/index_tensor([2039, 2039])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle, rotating shot, captured in a powerful, candid style, closely follows a cherry-red, family-use tricycle. Under the glaring afternoon sun, the tricycle hurtles toward the corner of a makeshift track constructed on the park's asphalt path at a speed belying its design. The metallic paintwork gleams with dazzling light, and the handlebars reveal the whitening knuckles of clenched fists. As it fiercely carves into the turn, the vehicle tilts perilously outward, nearly tipping over. In that instant, the inside rear wheel is completely lifted off the ground, its black rubber tire spinning futilely and wildly in mid-air, a stark contrast to the rough gray track surface below. The other two tires, gripped by the immense centrifugal force, kick up a cloud of fine dust and grit. In the background, blurred spectators and fluttering colorful flags flash by, their intense motion blur accentuating the tricycle's breakneck speed. The entire picture is filled with highly saturated colors and strong contrast between light and dark, perfectly capturing this comical and exciting competitive moment that pushes household items to their limits, creating an intense atmosphere of adrenaline rush.", + "seed": 5399, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_15/index_2039_seed_5399_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04105670116841793, + "vlm_logprob": -0.04105670116841793, + "vlm_yes_no_margin": 3.3999999523162843, + "vlm_token_credit": -0.04105670116841793, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05838760733604431 + ], + "opposite_logprob_tokens": [ + -3.308387517929077 + ], + "logprob_reward": -0.05838760733604431, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05838760733604431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1672395020723343, + 0.07360069453716278, + 0.13102152943611145, + 0.07122264802455902, + 0.05026684328913689, + 0.04269181564450264, + 0.0797731950879097, + 0.24611607193946838, + 0.053000956773757935, + 0.08506672084331512 + ], + "credit_map": [ + 0.021571103483438492, + 0.021269559860229492, + 0.016779765486717224, + 0.020249323919415474, + 0.016231130808591843, + 0.02211950160562992, + 0.023741822689771652, + 0.025277303531765938, + 0.009558018296957016, + 0.01794644631445408, + 0.010238302871584892, + 0.01198204793035984, + 0.006734516937285662, + 0.005670494865626097, + 0.005117711145430803, + 0.006353158503770828, + 0.012837163172662258, + 0.019199153408408165, + 0.014548525214195251, + 0.020465226843953133, + 0.018060024827718735, + 0.011899326927959919, + 0.019681204110383987, + 0.014330907724797726, + 0.004020631778985262, + 0.005820787511765957, + 0.006759990938007832, + 0.005746578797698021, + 0.006969715002924204, + 0.0075288936495780945, + 0.01787664368748665, + 0.01649940386414528, + 0.002911008894443512, + 0.005218344274908304, + 0.007613479159772396, + 0.005325266160070896, + 0.006616400089114904, + 0.008135315030813217, + 0.00669162767007947, + 0.007755402475595474, + 0.005317779257893562, + 0.005069013684988022, + 0.00483962707221508, + 0.005817226599901915, + 0.003928679972887039, + 0.007361397612839937, + 0.00572899030521512, + 0.004629103932529688, + 0.006086333654820919, + 0.006397719029337168, + 0.006681941915303469, + 0.011966768652200699, + 0.01550426334142685, + 0.012266675941646099, + 0.012291663326323032, + 0.00857782643288374, + 0.008222928270697594, + 0.0061080558225512505, + 0.0970316231250763, + 0.0923173576593399, + 0.011179494671523571, + 0.008642024360597134, + 0.01351857464760542, + 0.009096025489270687, + 0.00899268314242363, + 0.007217552047222853, + 0.007569785695523024, + 0.0060302820056676865, + 0.00592240784317255, + 0.005944184958934784, + 0.004850381053984165, + 0.006473680958151817, + 0.005542161408811808, + 0.008161376230418682, + 0.022628743201494217, + 0.01902376115322113, + 0.010175232775509357, + 0.004928213078528643, + 0.007711950223892927, + 0.006895282305777073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.05838760733604431 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002048194408417 + ], + "opposite_logprob_tokens": [ + -2.6950204372406006 + ], + "logprob_reward": -0.07002048194408417, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002048194408417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14917971193790436, + 0.07907506078481674, + 0.036850836127996445, + 0.026088079437613487, + 0.06656074523925781, + 0.10744904726743698, + 0.07179506868124008, + 0.3926089406013489, + 0.03749508783221245, + 0.0328974612057209 + ], + "credit_map": [ + 0.028017401695251465, + 0.03000299446284771, + 0.01699371449649334, + 0.022888988256454468, + 0.013103187084197998, + 0.013412011787295341, + 0.012296679429709911, + 0.012464742176234722, + 0.009193448349833488, + 0.012892909348011017, + 0.007670354563742876, + 0.01691070757806301, + 0.00851630698889494, + 0.00671692518517375, + 0.008922651410102844, + 0.008251756429672241, + 0.00409850524738431, + 0.005289407912641764, + 0.004049079492688179, + 0.005322608165442944, + 0.00585553701967001, + 0.0030795217026025057, + 0.00451378570869565, + 0.0046423920430243015, + 0.0034312536008656025, + 0.0027791913598775864, + 0.0025474154390394688, + 0.0029360181652009487, + 0.002224421827122569, + 0.003759861458092928, + 0.003953088074922562, + 0.004456829745322466, + 0.004495875909924507, + 0.005595861934125423, + 0.017938943579792976, + 0.011991201899945736, + 0.00475836219266057, + 0.0048401313833892345, + 0.007530194241553545, + 0.009410182014107704, + 0.012414820492267609, + 0.018968701362609863, + 0.01096335705369711, + 0.015374335460364819, + 0.010624988004565239, + 0.0150072630494833, + 0.013825152069330215, + 0.010270435363054276, + 0.008253870531916618, + 0.006575389765202999, + 0.007606381084769964, + 0.007812401745468378, + 0.013167550787329674, + 0.012008088640868664, + 0.008732781745493412, + 0.00763859786093235, + 0.012754534371197224, + 0.010044081136584282, + 0.16437044739723206, + 0.15973669290542603, + 0.01147050317376852, + 0.009325869381427765, + 0.01512978132814169, + 0.009777028113603592, + 0.003631167346611619, + 0.003964792471379042, + 0.004912262316793203, + 0.005158248823136091, + 0.004627796355634928, + 0.004510358441621065, + 0.005680050700902939, + 0.005010412074625492, + 0.004133328329771757, + 0.004779256880283356, + 0.004695795476436615, + 0.005725452210754156, + 0.005034859757870436, + 0.002653550822287798, + 0.002971102949231863, + 0.0029041108209639788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.07002048194408417 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811703622341156 + ], + "opposite_logprob_tokens": [ + -3.288116931915283 + ], + "logprob_reward": -0.03811703622341156, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811703622341156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07006024569272995, + 0.055585429072380066, + 0.0701923817396164, + 0.03684217855334282, + 0.027060585096478462, + 0.1508876085281372, + 0.026058794930577278, + 0.5038867592811584, + 0.02137330174446106, + 0.03805268183350563 + ], + "credit_map": [ + 0.008652802556753159, + 0.014096357859671116, + 0.006574329920113087, + 0.01034199446439743, + 0.009247408248484135, + 0.006872629281133413, + 0.006598372012376785, + 0.007676354609429836, + 0.00807590689510107, + 0.012638917192816734, + 0.006634915713220835, + 0.008800525218248367, + 0.00518576567992568, + 0.003507365006953478, + 0.004761498421430588, + 0.005980538669973612, + 0.009108739905059338, + 0.006549584213644266, + 0.007157018408179283, + 0.010272654704749584, + 0.00848469790071249, + 0.0060402522794902325, + 0.011729661375284195, + 0.010849777609109879, + 0.0035420353524386883, + 0.004538984037935734, + 0.004618129692971706, + 0.004774000030010939, + 0.003578877542167902, + 0.00448137242347002, + 0.006140675861388445, + 0.005168107803910971, + 0.001877638278529048, + 0.0024069829378277063, + 0.006202890537679195, + 0.002605013083666563, + 0.0025582730304449797, + 0.0026769405230879784, + 0.004471596330404282, + 0.004261252470314503, + 0.015265397727489471, + 0.026065973564982414, + 0.01659245975315571, + 0.021313948556780815, + 0.014665545895695686, + 0.021978672593832016, + 0.017270490527153015, + 0.017735138535499573, + 0.003173219272866845, + 0.002568522235378623, + 0.0025432598777115345, + 0.002096995012834668, + 0.005220321007072926, + 0.004971945192664862, + 0.0025253810454159975, + 0.0029591533821076155, + 0.01662828028202057, + 0.011949451640248299, + 0.22015711665153503, + 0.21742548048496246, + 0.0072579276748001575, + 0.006000788416713476, + 0.017129922285676003, + 0.007337779738008976, + 0.0025068880058825016, + 0.0025376526173204184, + 0.002208330435678363, + 0.0031182290986180305, + 0.002483044285327196, + 0.002584812929853797, + 0.003653957275673747, + 0.0022803894244134426, + 0.003722733585163951, + 0.004007870797067881, + 0.008154592476785183, + 0.008285212330520153, + 0.004530108533799648, + 0.00287066213786602, + 0.0038060788065195084, + 0.00267542852088809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.03811703622341156 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579738542437553 + ], + "opposite_logprob_tokens": [ + -3.8955798149108887 + ], + "logprob_reward": -0.020579738542437553, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579738542437553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10286278277635574, + 0.10392042994499207, + 0.12897683680057526, + 0.11808042973279953, + 0.03904343768954277, + 0.13294951617717743, + 0.039162810891866684, + 0.2790202796459198, + 0.017766760662198067, + 0.03821670264005661 + ], + "credit_map": [ + 0.012756108306348324, + 0.0158744677901268, + 0.010718738660216331, + 0.015145709738135338, + 0.010464474558830261, + 0.011385117657482624, + 0.01313464343547821, + 0.01338351983577013, + 0.013023789040744305, + 0.013797244057059288, + 0.011124798096716404, + 0.017510082572698593, + 0.01505100168287754, + 0.0076711131259799, + 0.014118074439466, + 0.011624321341514587, + 0.019406696781516075, + 0.011061807163059711, + 0.014098674058914185, + 0.014035386964678764, + 0.017838027328252792, + 0.017390789464116096, + 0.01781490445137024, + 0.01733054593205452, + 0.013460323214530945, + 0.008424142375588417, + 0.011585580185055733, + 0.012748817913234234, + 0.007515696343034506, + 0.008963547646999359, + 0.02984914369881153, + 0.02553318627178669, + 0.002617849735543132, + 0.0031168621499091387, + 0.007081410847604275, + 0.005136228632181883, + 0.004685147665441036, + 0.004582204855978489, + 0.006060219835489988, + 0.005763516761362553, + 0.014091442339122295, + 0.02104749158024788, + 0.012651779688894749, + 0.022030828520655632, + 0.01211160235106945, + 0.018020527437329292, + 0.017653310671448708, + 0.015342535451054573, + 0.00419339444488287, + 0.002951324451714754, + 0.004086507484316826, + 0.006493416614830494, + 0.006483475677669048, + 0.005472445394843817, + 0.005033552646636963, + 0.004448696039617062, + 0.010289843194186687, + 0.007344220299273729, + 0.11257291585206985, + 0.10850606858730316, + 0.011046496219933033, + 0.007700587157160044, + 0.01271551102399826, + 0.008844656869769096, + 0.003094066632911563, + 0.0032289260998368263, + 0.002033622469753027, + 0.0018754907650873065, + 0.0016084074741229415, + 0.001977324252948165, + 0.001774969045072794, + 0.0021739536896348, + 0.005593980196863413, + 0.005128066521137953, + 0.005885635502636433, + 0.005543284583836794, + 0.005285131279379129, + 0.003420023713260889, + 0.003559970064088702, + 0.0038006152026355267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.020579738542437553 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817864179611206 + ], + "opposite_logprob_tokens": [ + -4.018178462982178 + ], + "logprob_reward": -0.01817864179611206, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01817864179611206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10913468152284622, + 0.04215578734874725, + 0.030424611642956734, + 0.01908896677196026, + 0.020674588158726692, + 0.03340599685907364, + 0.07029197365045547, + 0.417467325925827, + 0.04510009288787842, + 0.21225591003894806 + ], + "credit_map": [ + 0.021472396329045296, + 0.020766282454133034, + 0.011370973661541939, + 0.01574157550930977, + 0.008022329770028591, + 0.009824891574680805, + 0.012052655220031738, + 0.00988358911126852, + 0.0056971050798892975, + 0.0072145829908549786, + 0.0064988043159246445, + 0.007944077253341675, + 0.004231075756251812, + 0.0022988663986325264, + 0.003692164784297347, + 0.004579117521643639, + 0.0036201714538037777, + 0.0020855844486504793, + 0.0031130642164498568, + 0.004155431408435106, + 0.004399721045047045, + 0.003306908765807748, + 0.00540301064029336, + 0.004340723156929016, + 0.0018866173923015594, + 0.0023942135740071535, + 0.002151522086933255, + 0.002422938821837306, + 0.0020342087373137474, + 0.002950704423710704, + 0.0023234826512634754, + 0.002925282809883356, + 0.0016921915812417865, + 0.0024123694747686386, + 0.004179669078439474, + 0.002260502427816391, + 0.002037400845438242, + 0.0030659388285130262, + 0.0024711075238883495, + 0.00255541130900383, + 0.0025487663224339485, + 0.003194272518157959, + 0.004445319529622793, + 0.0042664227075874805, + 0.0029738317243754864, + 0.0066511863842606544, + 0.00470332708209753, + 0.00462287338450551, + 0.005072947591543198, + 0.006587595213204622, + 0.009924628771841526, + 0.015273216180503368, + 0.007454333361238241, + 0.006774940062314272, + 0.011044262908399105, + 0.00816006027162075, + 0.013500750996172428, + 0.010763275437057018, + 0.18179625272750854, + 0.17983269691467285, + 0.007181850727647543, + 0.004427617881447077, + 0.014024348929524422, + 0.005940558388829231, + 0.006307702977210283, + 0.008156328462064266, + 0.006227407604455948, + 0.005232031922787428, + 0.004350653383880854, + 0.003973860759288073, + 0.006185273639857769, + 0.0046668401919305325, + 0.035319484770298004, + 0.028255341574549675, + 0.02866458147764206, + 0.03000352345407009, + 0.0330251008272171, + 0.01776399463415146, + 0.019395645707845688, + 0.01982826367020607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.01817864179611206 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, rotating shot, captured in a powerful, candid style, closely follows a cherry-red, family-use tricycle. Under the glaring afternoon sun, the tricycle hurtles toward the corner of a makeshift track constructed on the park's asphalt path at a speed belying its design. The metallic paintwork gleams with dazzling light, and the handlebars reveal the whitening knuckles of clenched fists. As it fiercely carves into the turn, the vehicle tilts perilously outward, nearly tipping over. In that instant, the inside rear wheel is completely lifted off the ground, its black rubber tire spinning futilely and wildly in mid-air, a stark contrast to the rough gray track surface below. The other two tires, gripped by the immense centrifugal force, kick up a cloud of fine dust and grit. In the background, blurred spectators and fluttering colorful flags flash by, their intense motion blur accentuating the tricycle's breakneck speed. The entire picture is filled with highly saturated colors and strong contrast between light and dark, perfectly capturing this comical and exciting competitive moment that pushes household items to their limits, creating an intense atmosphere of adrenaline rush.", + "seed": 5400, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_15/index_2039_seed_5400_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03603052068501711, + "vlm_logprob": -0.03603052068501711, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.03603052068501711, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0634877160191536 + ], + "opposite_logprob_tokens": [ + -2.9384877681732178 + ], + "logprob_reward": -0.0634877160191536, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0634877160191536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15843510627746582, + 0.12696179747581482, + 0.14628660678863525, + 0.200600266456604, + 0.09339337050914764, + 0.07079305499792099, + 0.0405103899538517, + 0.033368755131959915, + 0.039684660732746124, + 0.08996596187353134 + ], + "credit_map": [ + 0.01674586720764637, + 0.029339483007788658, + 0.019400451332330704, + 0.011268766596913338, + 0.024266589432954788, + 0.023840857669711113, + 0.016730966046452522, + 0.016842123121023178, + 0.014939648099243641, + 0.013629463501274586, + 0.01159762404859066, + 0.01842564158141613, + 0.012628563679754734, + 0.012781098484992981, + 0.015692295506596565, + 0.027267469093203545, + 0.027386553585529327, + 0.020011896267533302, + 0.018006667494773865, + 0.01646391674876213, + 0.014653201214969158, + 0.01467505656182766, + 0.020341863855719566, + 0.014747447334229946, + 0.02154015190899372, + 0.046194493770599365, + 0.01709580607712269, + 0.027166565880179405, + 0.02301262877881527, + 0.02308630384504795, + 0.02385343797504902, + 0.018650885671377182, + 0.0074074845761060715, + 0.012169959023594856, + 0.012159893289208412, + 0.012531688436865807, + 0.011561155319213867, + 0.011807755567133427, + 0.01301655825227499, + 0.012738877907395363, + 0.006489599123597145, + 0.006103409454226494, + 0.013996714726090431, + 0.021186525002121925, + 0.00678376667201519, + 0.004547822289168835, + 0.0054860603995621204, + 0.006199158262461424, + 0.006217937916517258, + 0.004744403064250946, + 0.005250029265880585, + 0.007093663793057203, + 0.0037128375843167305, + 0.005361987743526697, + 0.004026640672236681, + 0.004102890379726887, + 0.0031221245881170034, + 0.0037302940618246794, + 0.002466898411512375, + 0.004606545902788639, + 0.002618503523990512, + 0.009672615677118301, + 0.0031914489809423685, + 0.003960323520004749, + 0.0043784440495073795, + 0.005616020876914263, + 0.004401364829391241, + 0.006427847780287266, + 0.004492511507123709, + 0.006084249820560217, + 0.0037984352093189955, + 0.004485785961151123, + 0.006636610254645348, + 0.01001833751797676, + 0.009244592860341072, + 0.012247996404767036, + 0.007684667594730854, + 0.019980335608124733, + 0.012940817512571812, + 0.011212604120373726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.0634877160191536 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809959441423416 + ], + "opposite_logprob_tokens": [ + -3.288099527359009 + ], + "logprob_reward": -0.03809959441423416, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809959441423416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13821913301944733, + 0.0947614535689354, + 0.09746430069208145, + 0.05730363354086876, + 0.027212979272007942, + 0.02214302122592926, + 0.04655164107680321, + 0.02767316997051239, + 0.034929659217596054, + 0.45374104380607605 + ], + "credit_map": [ + 0.025659004226326942, + 0.029410691931843758, + 0.01668849214911461, + 0.011732256039977074, + 0.012543201446533203, + 0.011163095943629742, + 0.01057902816683054, + 0.020443355664610863, + 0.01288565807044506, + 0.017757873982191086, + 0.013399548828601837, + 0.012127360329031944, + 0.010601446032524109, + 0.006256338208913803, + 0.011599088087677956, + 0.010134132578969002, + 0.019336910918354988, + 0.01272793859243393, + 0.01126146037131548, + 0.01380570512264967, + 0.008611170575022697, + 0.01318934466689825, + 0.010331656783819199, + 0.008200114592909813, + 0.007528829388320446, + 0.006372557021677494, + 0.007173031568527222, + 0.010129520669579506, + 0.0039169322699308395, + 0.006522479932755232, + 0.007441775873303413, + 0.008218502625823021, + 0.003141270484775305, + 0.0033491498325020075, + 0.005274659488350153, + 0.004174110945314169, + 0.002542463829740882, + 0.0026161835994571447, + 0.0034923339262604713, + 0.0026228076312690973, + 0.0014398799976333976, + 0.0022196241188794374, + 0.004891015123575926, + 0.006062975153326988, + 0.0021091378293931484, + 0.00168799445964396, + 0.002049508970230818, + 0.001682885573245585, + 0.004138105548918247, + 0.0039214217104017735, + 0.005575636401772499, + 0.00520968995988369, + 0.004197467118501663, + 0.012535435147583485, + 0.003919400740414858, + 0.007054487708956003, + 0.002948623150587082, + 0.0028884524945169687, + 0.002978585660457611, + 0.0031745207961648703, + 0.0024230487179011106, + 0.007620410528033972, + 0.002417804906144738, + 0.0032217251136898994, + 0.003666206728667021, + 0.0028604576364159584, + 0.004079650156199932, + 0.0059637995436787605, + 0.0039675794541835785, + 0.006435683462768793, + 0.0033847580198198557, + 0.0045715272426605225, + 0.027841389179229736, + 0.03187161684036255, + 0.07758547365665436, + 0.06941819190979004, + 0.027666902169585228, + 0.07371360063552856, + 0.08055012673139572, + 0.06509372591972351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.03809959441423416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496309697628021 + ], + "opposite_logprob_tokens": [ + -2.9299631118774414 + ], + "logprob_reward": -0.05496309697628021, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496309697628021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15015240013599396, + 0.1434013694524765, + 0.13222768902778625, + 0.05943049117922783, + 0.04744040220975876, + 0.07345254719257355, + 0.04285208135843277, + 0.03956995904445648, + 0.05694989114999771, + 0.2545231878757477 + ], + "credit_map": [ + 0.02310618944466114, + 0.019564269110560417, + 0.017367886379361153, + 0.014249326661229134, + 0.018961643800139427, + 0.014836130663752556, + 0.013728710822761059, + 0.028338221833109856, + 0.015728790313005447, + 0.027350695803761482, + 0.01798071339726448, + 0.01844499632716179, + 0.017033763229846954, + 0.010991467162966728, + 0.01696081832051277, + 0.018910104408860207, + 0.020569657906889915, + 0.020701024681329727, + 0.01547818724066019, + 0.01857992820441723, + 0.012675563804805279, + 0.015183044597506523, + 0.013733378611505032, + 0.015306881628930569, + 0.006836203392595053, + 0.010094636119902134, + 0.008846868760883808, + 0.008133054710924625, + 0.0047554620541632175, + 0.007064734119921923, + 0.006708942353725433, + 0.006990585941821337, + 0.00380051601678133, + 0.004948684014379978, + 0.012088713236153126, + 0.010064729489386082, + 0.0034496102016419172, + 0.003915394656360149, + 0.005768002942204475, + 0.0034047469962388277, + 0.005347543396055698, + 0.005996237974613905, + 0.018723299726843834, + 0.01866704225540161, + 0.00873902440071106, + 0.0043685659766197205, + 0.0055108568631112576, + 0.006099968682974577, + 0.003922624047845602, + 0.003969348967075348, + 0.006535432767122984, + 0.007198271341621876, + 0.004078902769833803, + 0.006618835031986237, + 0.005386012606322765, + 0.00514264777302742, + 0.004655585624277592, + 0.0036851761396974325, + 0.004575932398438454, + 0.005183899309486151, + 0.0032226431649178267, + 0.010055352002382278, + 0.0035282440949231386, + 0.004663123283535242, + 0.004291352350264788, + 0.007231330964714289, + 0.004255176056176424, + 0.007967484183609486, + 0.010857843793928623, + 0.0074213831685483456, + 0.006697595585137606, + 0.008227720856666565, + 0.01633324846625328, + 0.0249796099960804, + 0.026032082736492157, + 0.03270453214645386, + 0.0220547653734684, + 0.05728476494550705, + 0.04100664332509041, + 0.03412749990820885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.05496309697628021 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253511756658554 + ], + "opposite_logprob_tokens": [ + -4.387535095214844 + ], + "logprob_reward": -0.01253511756658554, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253511756658554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11341589689254761, + 0.15547752380371094, + 0.11306682974100113, + 0.04948003590106964, + 0.1018151193857193, + 0.160108283162117, + 0.12221838533878326, + 0.04630294814705849, + 0.03639327734708786, + 0.10172174125909805 + ], + "credit_map": [ + 0.01870977133512497, + 0.013550007715821266, + 0.014219825156033039, + 0.010246211662888527, + 0.011165774427354336, + 0.01106717623770237, + 0.009281689301133156, + 0.025175437331199646, + 0.018785718828439713, + 0.029808515682816505, + 0.022514812648296356, + 0.01655544899404049, + 0.017073892056941986, + 0.010699332691729069, + 0.01741883158683777, + 0.022620972245931625, + 0.01094195619225502, + 0.021628187969326973, + 0.015210697427392006, + 0.01798868551850319, + 0.009840684942901134, + 0.010019601322710514, + 0.0142268231138587, + 0.01321018673479557, + 0.0073731630109250546, + 0.007242197170853615, + 0.005571485497057438, + 0.006463600788265467, + 0.00542364502325654, + 0.005969850812107325, + 0.005167280323803425, + 0.006268813740462065, + 0.010663543827831745, + 0.009542791172862053, + 0.024793261662125587, + 0.017476573586463928, + 0.008666837587952614, + 0.007863081991672516, + 0.0124511132016778, + 0.010357916355133057, + 0.016882291063666344, + 0.01223334763199091, + 0.029874490574002266, + 0.042645614594221115, + 0.015819448977708817, + 0.009864764288067818, + 0.015985701233148575, + 0.016802625730633736, + 0.010224594734609127, + 0.00963856466114521, + 0.016814185306429863, + 0.018411971628665924, + 0.009082893840968609, + 0.027435477823019028, + 0.012843536213040352, + 0.017767159268260002, + 0.005139322020113468, + 0.004036007449030876, + 0.006859458517283201, + 0.005758114159107208, + 0.0034383239690214396, + 0.009370539337396622, + 0.005890528671443462, + 0.005810652859508991, + 0.003305817488580942, + 0.002955714939162135, + 0.0031692571938037872, + 0.00743037648499012, + 0.0029818040784448385, + 0.006195484194904566, + 0.0039896052330732346, + 0.006365214008837938, + 0.006105284672230482, + 0.008697930723428726, + 0.02181928977370262, + 0.017529891803860664, + 0.006495595443993807, + 0.008765765465795994, + 0.01600104384124279, + 0.016306936740875244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.01253511756658554 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011067078448832035 + ], + "opposite_logprob_tokens": [ + -4.511066913604736 + ], + "logprob_reward": -0.011067078448832035, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011067078448832035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15982329845428467, + 0.14897191524505615, + 0.03506845980882645, + 0.03388885408639908, + 0.08810979872941971, + 0.04120226949453354, + 0.04650013521313667, + 0.017055151984095573, + 0.014738280326128006, + 0.4146418571472168 + ], + "credit_map": [ + 0.02113618515431881, + 0.03957388550043106, + 0.026028016582131386, + 0.014530849643051624, + 0.01395410392433405, + 0.015758797526359558, + 0.007876972667872906, + 0.020964469760656357, + 0.025802304968237877, + 0.020160825923085213, + 0.021857384592294693, + 0.016688724979758263, + 0.01416548527777195, + 0.01437464077025652, + 0.02072804979979992, + 0.015194485895335674, + 0.0064603714272379875, + 0.0035254245158284903, + 0.004153592977672815, + 0.004636457655578852, + 0.002785934368148446, + 0.003912979271262884, + 0.005141003988683224, + 0.004452690947800875, + 0.004903981927782297, + 0.004176790826022625, + 0.0034128064289689064, + 0.004350059665739536, + 0.0026456215418875217, + 0.003914955072104931, + 0.003960289526730776, + 0.006524343974888325, + 0.009167850017547607, + 0.0072004650719463825, + 0.02032315544784069, + 0.016320759430527687, + 0.00826408714056015, + 0.007226669229567051, + 0.009872670285403728, + 0.009734130464494228, + 0.005458751693367958, + 0.00414234446361661, + 0.007053834851831198, + 0.008834796026349068, + 0.004202495329082012, + 0.0028340909630060196, + 0.004133317619562149, + 0.004542637150734663, + 0.004225217737257481, + 0.0035773408599197865, + 0.004290168173611164, + 0.00520317954942584, + 0.004164037760347128, + 0.012214349582791328, + 0.005051260348409414, + 0.00777457607910037, + 0.001745654852129519, + 0.0022111646831035614, + 0.0025392272509634495, + 0.00151907280087471, + 0.0017693255795165896, + 0.004337869118899107, + 0.0013103409437462687, + 0.00162249687127769, + 0.0013349325163289905, + 0.0012750544119626284, + 0.002162253251299262, + 0.003010271117091179, + 0.001141150831244886, + 0.00219710567034781, + 0.00161815679166466, + 0.001999354688450694, + 0.026963602751493454, + 0.03990556672215462, + 0.04844997450709343, + 0.05822650343179703, + 0.03042101114988327, + 0.07935330271720886, + 0.07248268276453018, + 0.05883919820189476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.011067078448832035 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, rotating shot, captured in a powerful, candid style, closely follows a cherry-red, family-use tricycle. Under the glaring afternoon sun, the tricycle hurtles toward the corner of a makeshift track constructed on the park's asphalt path at a speed belying its design. The metallic paintwork gleams with dazzling light, and the handlebars reveal the whitening knuckles of clenched fists. As it fiercely carves into the turn, the vehicle tilts perilously outward, nearly tipping over. In that instant, the inside rear wheel is completely lifted off the ground, its black rubber tire spinning futilely and wildly in mid-air, a stark contrast to the rough gray track surface below. The other two tires, gripped by the immense centrifugal force, kick up a cloud of fine dust and grit. In the background, blurred spectators and fluttering colorful flags flash by, their intense motion blur accentuating the tricycle's breakneck speed. The entire picture is filled with highly saturated colors and strong contrast between light and dark, perfectly capturing this comical and exciting competitive moment that pushes household items to their limits, creating an intense atmosphere of adrenaline rush.", + "seed": 5401, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_15/index_2039_seed_5401_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0532607251778245, + "vlm_logprob": -0.0532607251778245, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.0532607251778245, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03735698387026787 + ], + "opposite_logprob_tokens": [ + -3.5373570919036865 + ], + "logprob_reward": -0.03735698387026787, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03735698387026787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16428041458129883, + 0.053478822112083435, + 0.02660425379872322, + 0.029481351375579834, + 0.056278400123119354, + 0.170575350522995, + 0.3422970473766327, + 0.08705301582813263, + 0.04137653112411499, + 0.0285748653113842 + ], + "credit_map": [ + 0.011465780436992645, + 0.01298537291586399, + 0.042721617966890335, + 0.043497100472450256, + 0.014200583100318909, + 0.011603502556681633, + 0.018241316080093384, + 0.009565139189362526, + 0.005368582438677549, + 0.004177192226052284, + 0.004568262025713921, + 0.004569187294691801, + 0.00608662236481905, + 0.005495106335729361, + 0.012338587082922459, + 0.010875282809138298, + 0.004556975327432156, + 0.003850628389045596, + 0.004818215034902096, + 0.00267019378952682, + 0.001975445309653878, + 0.002117832424119115, + 0.003596305148676038, + 0.003018657211214304, + 0.003158050123602152, + 0.0033196194563061, + 0.004277124535292387, + 0.0027196796145290136, + 0.003793757874518633, + 0.0040187593549489975, + 0.004251811653375626, + 0.003942549228668213, + 0.005377605091780424, + 0.0036223221104592085, + 0.005894177593290806, + 0.006625565700232983, + 0.006320127751678228, + 0.0070788078010082245, + 0.01299967896193266, + 0.008360113017261028, + 0.009958270937204361, + 0.0050714388489723206, + 0.06672026216983795, + 0.006106392480432987, + 0.004302613902837038, + 0.003893202403560281, + 0.06853356212377548, + 0.005989585537463427, + 0.011673478409647942, + 0.007467274088412523, + 0.14523017406463623, + 0.012757603079080582, + 0.006620919331908226, + 0.01126124057918787, + 0.036307625472545624, + 0.11097873747348785, + 0.006854903884232044, + 0.007900828495621681, + 0.010490135289728642, + 0.00926725473254919, + 0.008159070275723934, + 0.00892158318310976, + 0.02071942389011383, + 0.014739812351763248, + 0.004166180733591318, + 0.0039403242990374565, + 0.006092345342040062, + 0.0036829360760748386, + 0.006982977036386728, + 0.005693773273378611, + 0.005968268495053053, + 0.004849723074585199, + 0.0031247343868017197, + 0.0036327848210930824, + 0.004578104708343744, + 0.0036104307509958744, + 0.002705319318920374, + 0.002727217972278595, + 0.0038359160535037518, + 0.00436035729944706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.03735698387026787 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986387014389038 + ], + "opposite_logprob_tokens": [ + -1.8048639297485352 + ], + "logprob_reward": -0.17986387014389038, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986387014389038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12384927272796631, + 0.07525408267974854, + 0.04033100977540016, + 0.06118878349661827, + 0.07259605824947357, + 0.09200295060873032, + 0.26816803216934204, + 0.15688404440879822, + 0.06991544365882874, + 0.03981027752161026 + ], + "credit_map": [ + 0.013900885358452797, + 0.01592789590358734, + 0.013008964248001575, + 0.014842475764453411, + 0.018648995086550713, + 0.012435522861778736, + 0.02421431802213192, + 0.010870235040783882, + 0.008206123486161232, + 0.006921019870787859, + 0.012064383365213871, + 0.009382468648254871, + 0.006704599130898714, + 0.0070157418958842754, + 0.012944184243679047, + 0.012015574611723423, + 0.007771237287670374, + 0.006239274051040411, + 0.0053666699677705765, + 0.004567444324493408, + 0.005007047206163406, + 0.0033429651521146297, + 0.004297471605241299, + 0.0037389027420431376, + 0.012243269942700863, + 0.009345539845526218, + 0.00911590363830328, + 0.010235022753477097, + 0.00459142355248332, + 0.006148152053356171, + 0.003815904725342989, + 0.005693568382412195, + 0.008292715065181255, + 0.005089092534035444, + 0.011411797255277634, + 0.009440688416361809, + 0.0078028165735304356, + 0.008732649497687817, + 0.014184989035129547, + 0.007641316391527653, + 0.0068152062594890594, + 0.0033280192874372005, + 0.03047785349190235, + 0.0035270778462290764, + 0.0035713063552975655, + 0.003949364647269249, + 0.03528573364019394, + 0.005048402585089207, + 0.009672462940216064, + 0.007965297438204288, + 0.10091754049062729, + 0.011993272230029106, + 0.00724606029689312, + 0.010595514439046383, + 0.02328508347272873, + 0.09649281948804855, + 0.0167749784886837, + 0.01585560105741024, + 0.020925728604197502, + 0.01924898475408554, + 0.016365135088562965, + 0.01179230585694313, + 0.036481067538261414, + 0.019440259784460068, + 0.007996056228876114, + 0.007372849155217409, + 0.00897211953997612, + 0.0057336147874593735, + 0.00986239593476057, + 0.00964372418820858, + 0.010447131469845772, + 0.009887558408081532, + 0.004958413541316986, + 0.004321409855037928, + 0.006554632913321257, + 0.004722004756331444, + 0.004954885225743055, + 0.004334084689617157, + 0.004997145384550095, + 0.004967705812305212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.17986387014389038 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981428988277912 + ], + "opposite_logprob_tokens": [ + -3.5298142433166504 + ], + "logprob_reward": -0.02981428988277912, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981428988277912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12886187434196472, + 0.02687603048980236, + 0.022980425506830215, + 0.01809435524046421, + 0.015690194442868233, + 0.08554314076900482, + 0.6190882325172424, + 0.026408851146697998, + 0.03543054684996605, + 0.021026400849223137 + ], + "credit_map": [ + 0.010215576738119125, + 0.008787142112851143, + 0.024434521794319153, + 0.026280509307980537, + 0.012107755057513714, + 0.010113433003425598, + 0.025783419609069824, + 0.011139506474137306, + 0.002405945211648941, + 0.0025816024281084538, + 0.004266871139407158, + 0.002196439541876316, + 0.002842974616214633, + 0.0021930376533418894, + 0.004112274385988712, + 0.006276885978877544, + 0.00236670789308846, + 0.0048820748925209045, + 0.003974049352109432, + 0.0026522462721914053, + 0.0020063177216798067, + 0.00208727247081697, + 0.0028588094282895327, + 0.0021529491059482098, + 0.004042725544422865, + 0.002352724317461252, + 0.00260043703019619, + 0.002236807718873024, + 0.0015664873644709587, + 0.0023014128673821688, + 0.0013221047120168805, + 0.001671655336394906, + 0.001722368411719799, + 0.0012421852443367243, + 0.00277689634822309, + 0.0023117423988878727, + 0.001189206843264401, + 0.0015170705737546086, + 0.002777255140244961, + 0.002153468783944845, + 0.005290941335260868, + 0.0022796590346843004, + 0.033733926713466644, + 0.0028437271248549223, + 0.002165762474760413, + 0.0015377012314274907, + 0.034735504537820816, + 0.0029559137765318155, + 0.021491089835762978, + 0.013186652213335037, + 0.26198530197143555, + 0.02216460555791855, + 0.012219960801303387, + 0.02053758129477501, + 0.06584350019693375, + 0.20165956020355225, + 0.0032069215085357428, + 0.003487156005576253, + 0.004517692606896162, + 0.0029583582654595375, + 0.0017018868820741773, + 0.002252903999760747, + 0.005137370899319649, + 0.003146560862660408, + 0.005920025985687971, + 0.004854761995375156, + 0.004677414894104004, + 0.003706875490024686, + 0.004293861798942089, + 0.004023182671517134, + 0.004549885168671608, + 0.0034045411739498377, + 0.002348779933527112, + 0.0021612700074911118, + 0.004023784771561623, + 0.0028957936447113752, + 0.0023047495633363724, + 0.0018185219960287213, + 0.0029106743168085814, + 0.0025628271978348494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.02981428988277912 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673630740493536 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.00673630740493536, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673630740493536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033550821244716644, + 0.052620284259319305, + 0.031812019646167755, + 0.03210446983575821, + 0.06869339942932129, + 0.4570935070514679, + 0.22031864523887634, + 0.03633112087845802, + 0.04962565749883652, + 0.01785007305443287 + ], + "credit_map": [ + 0.0041228123009204865, + 0.0037587792612612247, + 0.003927348647266626, + 0.004322155844420195, + 0.00367718655616045, + 0.004161425866186619, + 0.005782388616353273, + 0.003798725549131632, + 0.010202698409557343, + 0.0045771850273013115, + 0.008630159310996532, + 0.005102402996271849, + 0.004382394719868898, + 0.0033887685276567936, + 0.008320765569806099, + 0.008015912026166916, + 0.006084983702749014, + 0.0041997553780674934, + 0.0051492261700332165, + 0.003838281612843275, + 0.003553698305040598, + 0.0018560788594186306, + 0.004074356984347105, + 0.003055636305361986, + 0.0044524879194796085, + 0.0042131259106099606, + 0.005297667812556028, + 0.005155351012945175, + 0.0028666581492871046, + 0.0033739455975592136, + 0.0033565270714461803, + 0.0033887075260281563, + 0.008589309640228748, + 0.00425209803506732, + 0.011404596269130707, + 0.010607616975903511, + 0.005179017782211304, + 0.006415632087737322, + 0.013333690352737904, + 0.0089114336296916, + 0.027375608682632446, + 0.012799650430679321, + 0.17583929002285004, + 0.01737193390727043, + 0.011835823766887188, + 0.008859585039317608, + 0.18552836775779724, + 0.017483271658420563, + 0.007826782763004303, + 0.004852405283600092, + 0.09463687241077423, + 0.007781887426972389, + 0.004870344884693623, + 0.0067528896033763885, + 0.028534023091197014, + 0.06506343930959702, + 0.0042558531276881695, + 0.004607430659234524, + 0.005193974357098341, + 0.005119512788951397, + 0.003830184228718281, + 0.002469488885253668, + 0.006496895104646683, + 0.0043577831238508224, + 0.008644245564937592, + 0.0059663476422429085, + 0.005798832047730684, + 0.004335033241659403, + 0.006084011867642403, + 0.0062966723926365376, + 0.004576003644615412, + 0.007924512960016727, + 0.0017183010932058096, + 0.0013611624017357826, + 0.0035663354210555553, + 0.004998806398361921, + 0.001225533545948565, + 0.001148846698924899, + 0.0014761174097657204, + 0.002354969969019294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.00673630740493536 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012532174587249756 + ], + "opposite_logprob_tokens": [ + -4.3875322341918945 + ], + "logprob_reward": -0.012532174587249756, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532174587249756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21214045584201813, + 0.06411846727132797, + 0.029659347608685493, + 0.06605706363916397, + 0.026821743696928024, + 0.048464611172676086, + 0.4458087682723999, + 0.021900760009884834, + 0.06393655389547348, + 0.02109219692647457 + ], + "credit_map": [ + 0.019794141873717308, + 0.016860000789165497, + 0.03324061259627342, + 0.03560575470328331, + 0.026326216757297516, + 0.023164020851254463, + 0.04112638160586357, + 0.016023334115743637, + 0.011646699160337448, + 0.00455007329583168, + 0.007842889055609703, + 0.006257314234972, + 0.005282410886138678, + 0.0032063033431768417, + 0.009153338149189949, + 0.016179440543055534, + 0.005929761566221714, + 0.0038164914585649967, + 0.006185382138937712, + 0.003340566297993064, + 0.002464352175593376, + 0.0022872556000947952, + 0.00307103106752038, + 0.0025645061396062374, + 0.014423537999391556, + 0.009672257117927074, + 0.010181250981986523, + 0.008928962051868439, + 0.005773191340267658, + 0.007502695079892874, + 0.004573718179017305, + 0.005001448094844818, + 0.0025881673209369183, + 0.0020225634798407555, + 0.0031232256442308426, + 0.004570337012410164, + 0.003485575085505843, + 0.003738939296454191, + 0.0032770512625575066, + 0.004015882965177298, + 0.003437717678025365, + 0.0016417969018220901, + 0.018977006897330284, + 0.0018351672915741801, + 0.0017191143706440926, + 0.0016017352463677526, + 0.016743015497922897, + 0.0025090565904974937, + 0.01612374559044838, + 0.009514791890978813, + 0.19088256359100342, + 0.017465176060795784, + 0.009028698317706585, + 0.013857226818799973, + 0.0555885024368763, + 0.13334810733795166, + 0.00260576861910522, + 0.0022177554201334715, + 0.0033819687087088823, + 0.003349797334522009, + 0.002147977240383625, + 0.0022998612839728594, + 0.00390655267983675, + 0.001991077559068799, + 0.012832232750952244, + 0.010294021107256413, + 0.007620485499501228, + 0.005651175510138273, + 0.007886695675551891, + 0.0064625246450304985, + 0.005820570979267359, + 0.007368848193436861, + 0.0019610519520938396, + 0.0023696718271821737, + 0.002625585300847888, + 0.004674858413636684, + 0.0017819487256929278, + 0.0021219835616648197, + 0.002765488810837269, + 0.002791607053950429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.012532174587249756 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, rotating shot, captured in a powerful, candid style, closely follows a cherry-red, family-use tricycle. Under the glaring afternoon sun, the tricycle hurtles toward the corner of a makeshift track constructed on the park's asphalt path at a speed belying its design. The metallic paintwork gleams with dazzling light, and the handlebars reveal the whitening knuckles of clenched fists. As it fiercely carves into the turn, the vehicle tilts perilously outward, nearly tipping over. In that instant, the inside rear wheel is completely lifted off the ground, its black rubber tire spinning futilely and wildly in mid-air, a stark contrast to the rough gray track surface below. The other two tires, gripped by the immense centrifugal force, kick up a cloud of fine dust and grit. In the background, blurred spectators and fluttering colorful flags flash by, their intense motion blur accentuating the tricycle's breakneck speed. The entire picture is filled with highly saturated colors and strong contrast between light and dark, perfectly capturing this comical and exciting competitive moment that pushes household items to their limits, creating an intense atmosphere of adrenaline rush.", + "seed": 5402, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_15/index_2039_seed_5402_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.040770523995161054, + "vlm_logprob": -0.040770523995161054, + "vlm_yes_no_margin": 3.874999976158142, + "vlm_token_credit": -0.040770523995161054, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028539536520838737 + ], + "opposite_logprob_tokens": [ + -4.153539657592773 + ], + "logprob_reward": -0.028539536520838737, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.028539536520838737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14759942889213562, + 0.10196122527122498, + 0.131802037358284, + 0.052497316151857376, + 0.05015096813440323, + 0.34727877378463745, + 0.02142971009016037, + 0.050645649433135986, + 0.056101467460393906, + 0.04053336754441261 + ], + "credit_map": [ + 0.019727902486920357, + 0.02471303753554821, + 0.017170527949929237, + 0.01618810184299946, + 0.015447551384568214, + 0.012907750904560089, + 0.01970142312347889, + 0.021743156015872955, + 0.009963608346879482, + 0.025824178010225296, + 0.012228420935571194, + 0.01270454004406929, + 0.013356562703847885, + 0.005318317096680403, + 0.01060576643794775, + 0.01195983961224556, + 0.009668576531112194, + 0.01235999260097742, + 0.018424851819872856, + 0.018316173925995827, + 0.01898539811372757, + 0.011052398011088371, + 0.02456619217991829, + 0.018428461626172066, + 0.0053092679008841515, + 0.004020709544420242, + 0.005225156433880329, + 0.005498028360307217, + 0.010183715261518955, + 0.01107972115278244, + 0.0044002048671245575, + 0.006780516356229782, + 0.006427520886063576, + 0.006101198494434357, + 0.008861741051077843, + 0.0068966480903327465, + 0.004433773458003998, + 0.0041058361530303955, + 0.006156479008495808, + 0.007167774718254805, + 0.06000099703669548, + 0.02869245782494545, + 0.1319928914308548, + 0.04496191069483757, + 0.0202373918145895, + 0.012362332083284855, + 0.027690259739756584, + 0.021340591832995415, + 0.0033889056649059057, + 0.0022576269693672657, + 0.004115120507776737, + 0.0031258726958185434, + 0.002590377116575837, + 0.0015463698655366898, + 0.003037110436707735, + 0.0013683289289474487, + 0.004059306345880032, + 0.0028363801538944244, + 0.010439952835440636, + 0.009262819774448872, + 0.0032375436276197433, + 0.007555875927209854, + 0.006060375366359949, + 0.007193404249846935, + 0.0031624813564121723, + 0.0031846824567764997, + 0.015870681032538414, + 0.004753315821290016, + 0.0038582170382142067, + 0.004375573713332415, + 0.017218446359038353, + 0.0036780773662030697, + 0.003746396629139781, + 0.005054304841905832, + 0.0035098146181553602, + 0.0053171939216554165, + 0.0063683814369142056, + 0.00633661774918437, + 0.006244509480893612, + 0.003956152126193047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.028539536520838737 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702569365501404 + ], + "opposite_logprob_tokens": [ + -2.127025604248047 + ], + "logprob_reward": -0.12702569365501404, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12702569365501404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1430201232433319, + 0.18477514386177063, + 0.07223113626241684, + 0.06994937360286713, + 0.052030231803655624, + 0.08710137754678726, + 0.028698373585939407, + 0.05744614452123642, + 0.13307058811187744, + 0.17167748510837555 + ], + "credit_map": [ + 0.014077432453632355, + 0.015731509774923325, + 0.021136386319994926, + 0.014512823894619942, + 0.02116374298930168, + 0.01749301888048649, + 0.01811828836798668, + 0.020786913111805916, + 0.01750333420932293, + 0.054895076900720596, + 0.022688796743750572, + 0.02366350218653679, + 0.01973763480782509, + 0.009367506951093674, + 0.01711351051926613, + 0.019805772230029106, + 0.007017976604402065, + 0.00883440300822258, + 0.011842601001262665, + 0.008729897439479828, + 0.012368150055408478, + 0.006086862646043301, + 0.008791163563728333, + 0.008560086600482464, + 0.00646360544487834, + 0.003979009576141834, + 0.005718023981899023, + 0.0056030601263046265, + 0.019761767238378525, + 0.01989070698618889, + 0.00362767418846488, + 0.004905525129288435, + 0.0060344566591084, + 0.008738039992749691, + 0.008085058070719242, + 0.008077725768089294, + 0.005722397938370705, + 0.004858301021158695, + 0.005464678164571524, + 0.005049576051533222, + 0.016791656613349915, + 0.007088265847414732, + 0.012383322231471539, + 0.007559661287814379, + 0.020635027438402176, + 0.004492100793868303, + 0.012808815576136112, + 0.005342532880604267, + 0.003410099307075143, + 0.0033030479680746794, + 0.0064484491012990475, + 0.003313877386972308, + 0.0026835231110453606, + 0.0022697974927723408, + 0.004331352189183235, + 0.002938227029517293, + 0.006163494661450386, + 0.007419831119477749, + 0.007242201361805201, + 0.008353123441338539, + 0.004690875299274921, + 0.009232909418642521, + 0.00788989569991827, + 0.00645381398499012, + 0.0071939085610210896, + 0.007970809005200863, + 0.029674548655748367, + 0.013171738013625145, + 0.00781671330332756, + 0.014616171829402447, + 0.04523072764277458, + 0.007395980879664421, + 0.015801332890987396, + 0.022400863468647003, + 0.016456758603453636, + 0.03803523629903793, + 0.01985609345138073, + 0.022107336670160294, + 0.0203613992780447, + 0.016658462584018707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.12702569365501404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370893374085426 + ], + "opposite_logprob_tokens": [ + -3.4087090492248535 + ], + "logprob_reward": -0.03370893374085426, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370893374085426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11907534301280975, + 0.11279454827308655, + 0.15708471834659576, + 0.07643456012010574, + 0.04941408336162567, + 0.06156421825289726, + 0.04931999742984772, + 0.18333324790000916, + 0.07206392288208008, + 0.11891540139913559 + ], + "credit_map": [ + 0.009771924465894699, + 0.01026628352701664, + 0.018455108627676964, + 0.012036752887070179, + 0.022716807201504707, + 0.009738368913531303, + 0.01638297364115715, + 0.019707120954990387, + 0.014595039188861847, + 0.020702270790934563, + 0.01377138588577509, + 0.012889240868389606, + 0.014922909438610077, + 0.007196307182312012, + 0.016231121495366096, + 0.012486274354159832, + 0.020128466188907623, + 0.018910083919763565, + 0.02255856618285179, + 0.032081540673971176, + 0.016441822052001953, + 0.010514847934246063, + 0.017427127808332443, + 0.019022256135940552, + 0.00570382084697485, + 0.0043270220048725605, + 0.006848631892353296, + 0.004752651788294315, + 0.02011881209909916, + 0.022014876827597618, + 0.0037082394119352102, + 0.008960503153502941, + 0.005162243265658617, + 0.006209524814039469, + 0.00809326022863388, + 0.009408343583345413, + 0.004434555768966675, + 0.0035086197312921286, + 0.007021085824817419, + 0.005576448515057564, + 0.004850372672080994, + 0.005531487055122852, + 0.008832979016005993, + 0.005710478872060776, + 0.01750498265028, + 0.004980877507477999, + 0.007428512442857027, + 0.006724529433995485, + 0.007221320178359747, + 0.005763784982264042, + 0.008600084111094475, + 0.008740177378058434, + 0.006072700954973698, + 0.003712462028488517, + 0.0053559462539851665, + 0.0038535187486559153, + 0.013044537976384163, + 0.014666750095784664, + 0.040159665048122406, + 0.03412483632564545, + 0.01698189787566662, + 0.022159701213240623, + 0.020689483731985092, + 0.021506378427147865, + 0.005340332631021738, + 0.00663298973813653, + 0.011010164394974709, + 0.010417553596198559, + 0.009139031171798706, + 0.010134217329323292, + 0.013465977273881435, + 0.0059236581437289715, + 0.010790998116135597, + 0.019086085259914398, + 0.018089313060045242, + 0.01785503327846527, + 0.015158639289438725, + 0.013929085806012154, + 0.013153930194675922, + 0.010852313600480556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.03370893374085426 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945852026343346 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.005945852026343346, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945852026343346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09815452247858047, + 0.10930798947811127, + 0.07556090503931046, + 0.1766611784696579, + 0.14192010462284088, + 0.1854424625635147, + 0.07421621680259705, + 0.07037247717380524, + 0.03655574098229408, + 0.03180832788348198 + ], + "credit_map": [ + 0.010208507999777794, + 0.012852160260081291, + 0.011972728185355663, + 0.012035032734274864, + 0.013167141005396843, + 0.006952918134629726, + 0.013769523240625858, + 0.01719650998711586, + 0.009604314342141151, + 0.026026776060461998, + 0.014870388433337212, + 0.012450188398361206, + 0.009624921716749668, + 0.008038617670536041, + 0.010793912224471569, + 0.01789887808263302, + 0.007716192863881588, + 0.007276829797774553, + 0.012200484052300453, + 0.015147927217185497, + 0.008714246563613415, + 0.005467340815812349, + 0.010706106200814247, + 0.008331780321896076, + 0.01122916117310524, + 0.010454698465764523, + 0.015355467796325684, + 0.015694953501224518, + 0.047948651015758514, + 0.049060337245464325, + 0.013126431033015251, + 0.013791469857096672, + 0.015735620632767677, + 0.021871117874979973, + 0.017751067876815796, + 0.03164353594183922, + 0.012623532675206661, + 0.009356443770229816, + 0.01973346807062626, + 0.013205313123762608, + 0.03730960562825203, + 0.014928129501640797, + 0.05722551420331001, + 0.020124148577451706, + 0.012762922793626785, + 0.009651991538703442, + 0.01907333917915821, + 0.01436680555343628, + 0.009481474757194519, + 0.010085348039865494, + 0.009174198843538761, + 0.010612858459353447, + 0.010885637253522873, + 0.006836197804659605, + 0.010451408103108406, + 0.006689099129289389, + 0.004302844405174255, + 0.004119215067476034, + 0.021713977679610252, + 0.018717123195528984, + 0.003779045771807432, + 0.006170080043375492, + 0.0058181025087833405, + 0.005752089433372021, + 0.0019317578990012407, + 0.0016894779400900006, + 0.009653562679886818, + 0.0036684179212898016, + 0.002176114823669195, + 0.0025809549260884523, + 0.012389893643558025, + 0.0024655635934323072, + 0.002270056866109371, + 0.00400685565546155, + 0.004000400193035603, + 0.005561721045523882, + 0.004542139824479818, + 0.004059105645865202, + 0.0036087380722165108, + 0.0037593094166368246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.005945852026343346 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cherry-red family-use tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle moving at high speed toward a corner of a makeshift track on an asphalt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the inside rear wheel completely lifted off the ground while spinning in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot captured from a low angle with a rotating camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image characterized by highly saturated colors and strong contrast between light and dark?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008632604032754898 + ], + "opposite_logprob_tokens": [ + -4.758632659912109 + ], + "logprob_reward": -0.008632604032754898, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008632604032754898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20080909132957458, + 0.09942874312400818, + 0.06378128379583359, + 0.08164387196302414, + 0.06577232480049133, + 0.20270690321922302, + 0.03644523397088051, + 0.08328953385353088, + 0.10006646066904068, + 0.06605657190084457 + ], + "credit_map": [ + 0.018836844712495804, + 0.022142088040709496, + 0.01631343364715576, + 0.031874824315309525, + 0.030245481058955193, + 0.014013457112014294, + 0.03777238726615906, + 0.02961054816842079, + 0.007348299026489258, + 0.02690889686346054, + 0.013335664756596088, + 0.013845567591488361, + 0.009146391414105892, + 0.007125884294509888, + 0.01051513385027647, + 0.011202896013855934, + 0.005535337142646313, + 0.008605863898992538, + 0.006568282842636108, + 0.00862935557961464, + 0.009363359771668911, + 0.004879858810454607, + 0.012785458005964756, + 0.007413762621581554, + 0.008355229161679745, + 0.007532388903200626, + 0.010437323711812496, + 0.010180044919252396, + 0.010550775565207005, + 0.01159244030714035, + 0.007934603840112686, + 0.01506105251610279, + 0.008038857951760292, + 0.009046236984431744, + 0.013296922668814659, + 0.007403986062854528, + 0.0059462678618729115, + 0.006492321845144033, + 0.008329986594617367, + 0.007217742968350649, + 0.018541360273957253, + 0.023151924833655357, + 0.05886254459619522, + 0.02784019336104393, + 0.025877466425299644, + 0.01606622338294983, + 0.016756432130932808, + 0.015610757283866405, + 0.00426497170701623, + 0.0043801902793347836, + 0.006717891898006201, + 0.004181649070233107, + 0.004588093608617783, + 0.003359622322022915, + 0.004819873720407486, + 0.004132937639951706, + 0.009029450826346874, + 0.008927474729716778, + 0.012097545899450779, + 0.012533527798950672, + 0.009112151339650154, + 0.010501486249268055, + 0.009979947470128536, + 0.011107943020761013, + 0.0060829585418105125, + 0.00806794036179781, + 0.020604245364665985, + 0.01302419975399971, + 0.010437721386551857, + 0.009136072359979153, + 0.025729605928063393, + 0.006983706261962652, + 0.0065155536867678165, + 0.009948233142495155, + 0.005624809768050909, + 0.010049721226096153, + 0.007438939996063709, + 0.012799643911421299, + 0.006954997777938843, + 0.0067246719263494015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.008632604032754898 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_2/index_tensor([5400, 5400])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_2/index_tensor([5400, 5400])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..aa15d266d74cd902b45625280e5b57990ce3127f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_2/index_tensor([5400, 5400])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, still photograph taken from a slightly elevated angle captures a tiny, iridescent black beetle perched on the richly textured, dark gray plastic dashboard of a moving car. Intense, direct sunlight filtering through the windshield casts harsh, bright spots on the beetle's smooth carapace and creates sharp, elongated micro-shadows beneath its body. The entire surface vibrates with the vehicle's subtle, high-frequency vibrations, causing the beetle to desperately search for its place. Its six spindly legs frantically wriggle, sliding across the smooth, sun-warmed plastic, its antennae twitching nervously. It periodically grasps the raised veins of the dashboard's texture, only to be jolted again by the constant movement, sliding a few millimeters with each jolt. In the background, a bright, overexposed landscape of sun-bleached road and passing trees, completely blurred through the out-of-focus windshield, emphasizes the car's speed and creates a tense, slightly dramatic atmosphere focused entirely on this tiny creature's Herculean efforts to survive.", + "seed": 24428, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_2/index_5400_seed_24428_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04079953320324421, + "vlm_logprob": -0.04079953320324421, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.04079953320324421, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027392376214265823 + ], + "opposite_logprob_tokens": [ + -3.7773923873901367 + ], + "logprob_reward": -0.027392376214265823, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027392376214265823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.286903440952301, + 0.05169187858700752, + 0.041112642735242844, + 0.0769752487540245, + 0.05629265680909157, + 0.08925622701644897, + 0.053417522460222244, + 0.06322280317544937, + 0.08679238706827164, + 0.19433514773845673 + ], + "credit_map": [ + 0.026723118498921394, + 0.013071930967271328, + 0.11094387620687485, + 0.092412069439888, + 0.007683506701141596, + 0.008363669738173485, + 0.015526756644248962, + 0.012178518809378147, + 0.006083603948354721, + 0.006165571045130491, + 0.009175215847790241, + 0.00984892062842846, + 0.004303559195250273, + 0.002969138091430068, + 0.007107535842806101, + 0.006038331426680088, + 0.005140876863151789, + 0.008930426090955734, + 0.005409738980233669, + 0.008040612563490868, + 0.0028981848154217005, + 0.001957810251042247, + 0.00448125135153532, + 0.004253744147717953, + 0.012265109457075596, + 0.01391840074211359, + 0.012137183919548988, + 0.012047801166772842, + 0.005428844131529331, + 0.0034335623495280743, + 0.008449683897197247, + 0.009294664487242699, + 0.006876525469124317, + 0.009078357368707657, + 0.0072568077594041824, + 0.007119473535567522, + 0.004890472628176212, + 0.010325910523533821, + 0.0036707092076539993, + 0.007074402179569006, + 0.009167129173874855, + 0.010074135847389698, + 0.01743869110941887, + 0.020634282380342484, + 0.008028714917600155, + 0.0064005241729319096, + 0.010525173507630825, + 0.006987580563873053, + 0.007855250500142574, + 0.008235446177423, + 0.00931063573807478, + 0.008052777498960495, + 0.0045660147443413734, + 0.0025318798143416643, + 0.009060794487595558, + 0.0038047221023589373, + 0.006328923627734184, + 0.005936064757406712, + 0.01194471213966608, + 0.010249634273350239, + 0.00438196025788784, + 0.005689567886292934, + 0.006770521402359009, + 0.011921419762074947, + 0.013385877944529057, + 0.019745487719774246, + 0.010716459713876247, + 0.012669840827584267, + 0.006441525649279356, + 0.004903391934931278, + 0.011558717116713524, + 0.007371085695922375, + 0.017727551981806755, + 0.03023110330104828, + 0.03455464541912079, + 0.029693931341171265, + 0.016655026003718376, + 0.013907932676374912, + 0.030137337744235992, + 0.021427616477012634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.027392376214265823 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865625500679016 + ], + "opposite_logprob_tokens": [ + -3.0486562252044678 + ], + "logprob_reward": -0.04865625500679016, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865625500679016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16394270956516266, + 0.09183220565319061, + 0.023738645017147064, + 0.12655024230480194, + 0.04060501977801323, + 0.08130326867103577, + 0.02691364847123623, + 0.11964137107133865, + 0.05099887400865555, + 0.27447399497032166 + ], + "credit_map": [ + 0.01360949408262968, + 0.013509641401469707, + 0.04840478673577309, + 0.044437527656555176, + 0.00882053654640913, + 0.009104427881538868, + 0.01570124551653862, + 0.010355041362345219, + 0.010163644328713417, + 0.013506009243428707, + 0.015101289376616478, + 0.01830917038023472, + 0.008023313246667385, + 0.005437629297375679, + 0.011166981421411037, + 0.010124167427420616, + 0.0035387848038226366, + 0.005135204643011093, + 0.0034569301642477512, + 0.004906323738396168, + 0.0013004499487578869, + 0.0009362181881442666, + 0.0022083644289523363, + 0.0022563692182302475, + 0.02101815491914749, + 0.02212616801261902, + 0.023218711838126183, + 0.02033846452832222, + 0.007852131500840187, + 0.006633195094764233, + 0.011990819126367569, + 0.013372594490647316, + 0.0069175586104393005, + 0.003948595374822617, + 0.006337014492601156, + 0.005345724523067474, + 0.003388582728803158, + 0.006574531551450491, + 0.003270848887041211, + 0.004822161048650742, + 0.00903544295579195, + 0.006626454647630453, + 0.022929223254323006, + 0.017040519043803215, + 0.005556722171604633, + 0.007671041879802942, + 0.006945073138922453, + 0.005498790647834539, + 0.0029381010681390762, + 0.0063870190642774105, + 0.004443246405571699, + 0.004072167910635471, + 0.002089708810672164, + 0.001559439697302878, + 0.00347554381005466, + 0.0019484235672280192, + 0.013660179451107979, + 0.017538970336318016, + 0.018899058923125267, + 0.017367210239171982, + 0.005882971454411745, + 0.010565413162112236, + 0.013060325756669044, + 0.02266724780201912, + 0.006629942916333675, + 0.00909806601703167, + 0.009408262558281422, + 0.009632688947021961, + 0.00315896887332201, + 0.0036739958450198174, + 0.0058346823789179325, + 0.0035622657742351294, + 0.029462531208992004, + 0.04245297983288765, + 0.04798455908894539, + 0.06774356961250305, + 0.017942186444997787, + 0.021234611049294472, + 0.026274263858795166, + 0.021379295736551285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.04865625500679016 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785539470613003 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.009785539470613003, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785539470613003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15786698460578918, + 0.030939245596528053, + 0.044650498777627945, + 0.0691908672451973, + 0.11725136637687683, + 0.12570305168628693, + 0.16173513233661652, + 0.04092930629849434, + 0.07101007550954819, + 0.18072345852851868 + ], + "credit_map": [ + 0.011344747617840767, + 0.018877491354942322, + 0.034744907170534134, + 0.04170725494623184, + 0.01268727146089077, + 0.007985362783074379, + 0.016472723335027695, + 0.014047244563698769, + 0.0030886612366884947, + 0.0033682226203382015, + 0.0046972972340881824, + 0.005161914974451065, + 0.003495245473459363, + 0.0024371801409870386, + 0.004837140440940857, + 0.0038535830099135637, + 0.004568345379084349, + 0.007187058217823505, + 0.006923085078597069, + 0.00777428550645709, + 0.004087958950549364, + 0.002049854723736644, + 0.004685852210968733, + 0.007374063599854708, + 0.009266608394682407, + 0.00913536362349987, + 0.014549153856933117, + 0.012837097980082035, + 0.005014178343117237, + 0.004506231285631657, + 0.007737680803984404, + 0.006144565064460039, + 0.012362495996057987, + 0.024255944415926933, + 0.015133097767829895, + 0.016113784164190292, + 0.009223248809576035, + 0.020503859966993332, + 0.0070764413103461266, + 0.012582509778439999, + 0.016367079690098763, + 0.021700909361243248, + 0.024446668103337288, + 0.022956280037760735, + 0.0054895589128136635, + 0.015012277290225029, + 0.009310105815529823, + 0.010420188307762146, + 0.022579113021492958, + 0.036073941737413406, + 0.02675771154463291, + 0.02907230518758297, + 0.010776682756841183, + 0.008484942838549614, + 0.01637803204357624, + 0.0116124227643013, + 0.005670585669577122, + 0.007932357490062714, + 0.006898257415741682, + 0.0062116156332194805, + 0.0023995109368115664, + 0.0024729089345782995, + 0.005261881276965141, + 0.004082190338522196, + 0.01008689310401678, + 0.015684546902775764, + 0.011344777420163155, + 0.013355345465242863, + 0.004441814962774515, + 0.005490794312208891, + 0.006079225800931454, + 0.004526677541434765, + 0.019732888787984848, + 0.027742689475417137, + 0.03128334879875183, + 0.04162400960922241, + 0.013137691654264927, + 0.01752135343849659, + 0.01591498963534832, + 0.013766488991677761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.009785539470613003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253083072602749 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253083072602749, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253083072602749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2581866681575775, + 0.10638076812028885, + 0.08018331974744797, + 0.14545677602291107, + 0.07754014432430267, + 0.12011729180812836, + 0.10038676857948303, + 0.06922619789838791, + 0.017867371439933777, + 0.024654673412442207 + ], + "credit_map": [ + 0.0215629730373621, + 0.023151075467467308, + 0.07611843198537827, + 0.06906432658433914, + 0.013710844330489635, + 0.013972771354019642, + 0.02416636236011982, + 0.016439884901046753, + 0.01142857875674963, + 0.0144351152703166, + 0.01575012318789959, + 0.020697323605418205, + 0.009735550731420517, + 0.006876066327095032, + 0.014236962422728539, + 0.013221056200563908, + 0.008186521008610725, + 0.014503298327326775, + 0.01527524832636118, + 0.015285998582839966, + 0.006404828745871782, + 0.0034105812665075064, + 0.007958377711474895, + 0.009158465079963207, + 0.025129131972789764, + 0.02534175105392933, + 0.022556820884346962, + 0.02318578027188778, + 0.010704020969569683, + 0.009637112729251385, + 0.015583454631268978, + 0.013318710960447788, + 0.0128439050167799, + 0.010521644726395607, + 0.013742583803832531, + 0.011977014131844044, + 0.005556314252316952, + 0.008975889533758163, + 0.005995616782456636, + 0.007927176542580128, + 0.015258637256920338, + 0.03122543916106224, + 0.01881723664700985, + 0.022015942260622978, + 0.0065907626412808895, + 0.009202352724969387, + 0.010431462898850441, + 0.00657545356079936, + 0.01438992191106081, + 0.02038702927529812, + 0.02046344429254532, + 0.019515499472618103, + 0.005558292847126722, + 0.005400646012276411, + 0.008760183118283749, + 0.00591175677254796, + 0.010878457687795162, + 0.016878150403499603, + 0.008181754499673843, + 0.00956698227673769, + 0.004546447657048702, + 0.005866311956197023, + 0.006350237876176834, + 0.006957858335226774, + 0.002403789199888706, + 0.0031896494328975677, + 0.003058799309656024, + 0.003355636028572917, + 0.001144913723692298, + 0.0012392554199323058, + 0.0019855028949677944, + 0.0014898257795721292, + 0.0024961577728390694, + 0.0038992834743112326, + 0.00506934430450201, + 0.004296230152249336, + 0.0017683319747447968, + 0.001955387881025672, + 0.002929608104750514, + 0.002240330446511507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.005253083072602749 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291041225194931 + ], + "opposite_logprob_tokens": [ + -2.237910509109497 + ], + "logprob_reward": -0.11291041225194931, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291041225194931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21994376182556152, + 0.037046320736408234, + 0.02361658215522766, + 0.15156909823417664, + 0.12990672886371613, + 0.10735217481851578, + 0.06448125094175339, + 0.08503556996583939, + 0.09161621332168579, + 0.08943237364292145 + ], + "credit_map": [ + 0.021518195047974586, + 0.01747717149555683, + 0.06820028275251389, + 0.05947496369481087, + 0.011579508893191814, + 0.0100662000477314, + 0.016419047489762306, + 0.01520836167037487, + 0.005161002278327942, + 0.004988561850041151, + 0.005745631642639637, + 0.00653864024206996, + 0.0034217145293951035, + 0.0026337222661823034, + 0.004780645947903395, + 0.003776397556066513, + 0.0024604785721749067, + 0.003378398949280381, + 0.004503682255744934, + 0.0049559324979782104, + 0.0020739594474434853, + 0.000994735979475081, + 0.0025880627799779177, + 0.0026613296940922737, + 0.02514347992837429, + 0.024094190448522568, + 0.02664978988468647, + 0.025173084810376167, + 0.010928161442279816, + 0.007936641573905945, + 0.01735585927963257, + 0.014287875033915043, + 0.017597466707229614, + 0.020696761086583138, + 0.017195353284478188, + 0.018856169655919075, + 0.010772238485515118, + 0.021784262731671333, + 0.007461014203727245, + 0.015543465502560139, + 0.01261639129370451, + 0.020294934511184692, + 0.019408676773309708, + 0.01867176964879036, + 0.006246332079172134, + 0.009435130283236504, + 0.013463654555380344, + 0.007215266115963459, + 0.007586481980979443, + 0.013903725892305374, + 0.012064339593052864, + 0.011951124295592308, + 0.003574167611077428, + 0.003286932362243533, + 0.007093010935932398, + 0.005021464545279741, + 0.013765377923846245, + 0.015971820801496506, + 0.012450040318071842, + 0.011633026413619518, + 0.0062017482705414295, + 0.004923181142657995, + 0.0080808624625206, + 0.01200950052589178, + 0.010711726732552052, + 0.014186660759150982, + 0.012816471979022026, + 0.014359921216964722, + 0.007537149824202061, + 0.007804649416357279, + 0.01355735119432211, + 0.010642279870808125, + 0.007430063094943762, + 0.00793017353862524, + 0.018575631082057953, + 0.01587083376944065, + 0.00844485405832529, + 0.005657133646309376, + 0.015415676869452, + 0.010107994079589844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.11291041225194931 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, still photograph taken from a slightly elevated angle captures a tiny, iridescent black beetle perched on the richly textured, dark gray plastic dashboard of a moving car. Intense, direct sunlight filtering through the windshield casts harsh, bright spots on the beetle's smooth carapace and creates sharp, elongated micro-shadows beneath its body. The entire surface vibrates with the vehicle's subtle, high-frequency vibrations, causing the beetle to desperately search for its place. Its six spindly legs frantically wriggle, sliding across the smooth, sun-warmed plastic, its antennae twitching nervously. It periodically grasps the raised veins of the dashboard's texture, only to be jolted again by the constant movement, sliding a few millimeters with each jolt. In the background, a bright, overexposed landscape of sun-bleached road and passing trees, completely blurred through the out-of-focus windshield, emphasizes the car's speed and creates a tense, slightly dramatic atmosphere focused entirely on this tiny creature's Herculean efforts to survive.", + "seed": 24429, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_2/index_5400_seed_24429_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13907738905400038, + "vlm_logprob": -0.13907738905400038, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.13907738905400038, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03308577090501785 + ], + "opposite_logprob_tokens": [ + -3.658085823059082 + ], + "logprob_reward": -0.03308577090501785, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03308577090501785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4588714838027954, + 0.12032954394817352, + 0.044467680156230927, + 0.1157945841550827, + 0.01699596457183361, + 0.05677901953458786, + 0.12866489589214325, + 0.017274577170610428, + 0.020881876349449158, + 0.01994040049612522 + ], + "credit_map": [ + 0.025598974898457527, + 0.16837406158447266, + 0.010608538053929806, + 0.029370049014687538, + 0.029835032299160957, + 0.0245786365121603, + 0.0366121307015419, + 0.13389405608177185, + 0.007189503405243158, + 0.03464072570204735, + 0.0030737603083252907, + 0.012753939256072044, + 0.007396341767162085, + 0.010302319191396236, + 0.012265237979590893, + 0.03270772099494934, + 0.003577345749363303, + 0.005283717066049576, + 0.003497048281133175, + 0.008611311204731464, + 0.0027243744116276503, + 0.005052446853369474, + 0.007873613387346268, + 0.007847821339964867, + 0.006686243694275618, + 0.042929258197546005, + 0.002983410144224763, + 0.0071865711361169815, + 0.00822141207754612, + 0.005893980152904987, + 0.00956524908542633, + 0.03232846036553383, + 0.002169419312849641, + 0.004069789312779903, + 0.0009381057461723685, + 0.0019207611912861466, + 0.0017341215861961246, + 0.0012486616615206003, + 0.0025643822737038136, + 0.0023507247678935528, + 0.0051430328749120235, + 0.009316088631749153, + 0.003652252722531557, + 0.008535702712833881, + 0.0030396480578929186, + 0.006031468976289034, + 0.011389916762709618, + 0.009670909494161606, + 0.013267356902360916, + 0.012815938331186771, + 0.008465968072414398, + 0.023771125823259354, + 0.014175568707287312, + 0.015692152082920074, + 0.02313079871237278, + 0.0173459742218256, + 0.0020460414234548807, + 0.0014218324795365334, + 0.0008874920313246548, + 0.002524162642657757, + 0.002029775409027934, + 0.0031812614761292934, + 0.0025037808809429407, + 0.002680230885744095, + 0.0020892603788524866, + 0.0021183984354138374, + 0.0013691080966964364, + 0.0038464905228465796, + 0.0017709910171106458, + 0.002861815271899104, + 0.0038232686929404736, + 0.003002544865012169, + 0.0027831527404487133, + 0.0015480731381103396, + 0.0017281160689890385, + 0.0032407387625426054, + 0.0013915150193497539, + 0.0025580991059541702, + 0.004066591151058674, + 0.0026241163723170757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.03308577090501785 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867521673440933 + ], + "opposite_logprob_tokens": [ + -3.048675298690796 + ], + "logprob_reward": -0.04867521673440933, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867521673440933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3702530562877655, + 0.12335988879203796, + 0.07313613593578339, + 0.10393837094306946, + 0.03399765491485596, + 0.061167363077402115, + 0.09829387068748474, + 0.08224425464868546, + 0.019813545048236847, + 0.033795759081840515 + ], + "credit_map": [ + 0.022658102214336395, + 0.12787912786006927, + 0.01203303411602974, + 0.025797627866268158, + 0.025455476716160774, + 0.02847227081656456, + 0.02952508069574833, + 0.09843237698078156, + 0.009039264172315598, + 0.027644891291856766, + 0.00460766302421689, + 0.013537168502807617, + 0.01360927615314722, + 0.019088514149188995, + 0.013670723885297775, + 0.022162403911352158, + 0.008729930967092514, + 0.008855300955474377, + 0.004694218281656504, + 0.013779062777757645, + 0.005009145941585302, + 0.006963546387851238, + 0.015948699787259102, + 0.009156248532235622, + 0.00602447846904397, + 0.03896554186940193, + 0.0025555668398737907, + 0.005013737827539444, + 0.0073740496300160885, + 0.004882193636149168, + 0.009413723833858967, + 0.029709096997976303, + 0.004081661347299814, + 0.007117975037544966, + 0.0017630602233111858, + 0.003317225491628051, + 0.002680862555280328, + 0.002913490403443575, + 0.005677743349224329, + 0.006445641629397869, + 0.004183115437626839, + 0.019511865451931953, + 0.002713794820010662, + 0.0044989935122430325, + 0.0037067539524286985, + 0.005519284401088953, + 0.006743015721440315, + 0.014290549792349339, + 0.010506086982786655, + 0.01009446568787098, + 0.00733160600066185, + 0.017269788309931755, + 0.01048116572201252, + 0.010937904007732868, + 0.017904268577694893, + 0.01376859936863184, + 0.010503017343580723, + 0.007150375284254551, + 0.006693722680211067, + 0.011071179993450642, + 0.009117476642131805, + 0.008716448210179806, + 0.0168764591217041, + 0.012115593999624252, + 0.0020989011973142624, + 0.0017823291709646583, + 0.0017060758545994759, + 0.002432398032397032, + 0.001401078305207193, + 0.003765427740290761, + 0.0037776061799377203, + 0.0028497318271547556, + 0.004871790297329426, + 0.0033270036801695824, + 0.0035641775466501713, + 0.0033289650455117226, + 0.002766065299510956, + 0.005228190682828426, + 0.006288975011557341, + 0.004420596174895763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.04867521673440933 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636326290667057 + ], + "opposite_logprob_tokens": [ + -3.6513633728027344 + ], + "logprob_reward": -0.02636326290667057, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636326290667057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08987436443567276, + 0.13447371125221252, + 0.08870091289281845, + 0.3561367094516754, + 0.0995275005698204, + 0.04167335480451584, + 0.09902933239936829, + 0.02552986331284046, + 0.028284160420298576, + 0.03677011653780937 + ], + "credit_map": [ + 0.013593188486993313, + 0.008812709711492062, + 0.006745415739715099, + 0.015758207067847252, + 0.00666357995942235, + 0.011209851130843163, + 0.012704827822744846, + 0.014386574737727642, + 0.01032327301800251, + 0.036093201488256454, + 0.005580898839980364, + 0.022616401314735413, + 0.009972083382308483, + 0.009626772254705429, + 0.016994373872876167, + 0.023266712203621864, + 0.007634438574314117, + 0.02487875334918499, + 0.004045779351145029, + 0.015398419462144375, + 0.005684919189661741, + 0.005806386470794678, + 0.009980401024222374, + 0.0152718061581254, + 0.02012861892580986, + 0.13623639941215515, + 0.010001939721405506, + 0.01662231609225273, + 0.024046849459409714, + 0.015866998583078384, + 0.026667069643735886, + 0.10656652599573135, + 0.008343414403498173, + 0.016691595315933228, + 0.006328604184091091, + 0.01182533148676157, + 0.006024934817105532, + 0.008177625946700573, + 0.018907669931650162, + 0.02322830818593502, + 0.0073869489133358, + 0.005375586450099945, + 0.002913770964369178, + 0.004030945245176554, + 0.002649228321388364, + 0.004818199202418327, + 0.007553246803581715, + 0.006945427972823381, + 0.010946289636194706, + 0.010123229585587978, + 0.008068908005952835, + 0.018033865839242935, + 0.01018412597477436, + 0.010094624012708664, + 0.017354100942611694, + 0.014224186539649963, + 0.002565207425504923, + 0.0026102750562131405, + 0.002190196653828025, + 0.003716956591233611, + 0.0023535932414233685, + 0.0028414083644747734, + 0.005583674181252718, + 0.003668547607958317, + 0.0027067377232015133, + 0.0020819930359721184, + 0.0017390198772773147, + 0.0030935758259147406, + 0.00308804027736187, + 0.007687933277338743, + 0.003680974943563342, + 0.004205884877592325, + 0.004544232506304979, + 0.002553805010393262, + 0.0034816598054021597, + 0.006378874182701111, + 0.0022639622911810875, + 0.004777208436280489, + 0.006242991425096989, + 0.006527382414788008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.02636326290667057 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086766608059406 + ], + "opposite_logprob_tokens": [ + -4.511086940765381 + ], + "logprob_reward": -0.011086766608059406, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086766608059406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19694915413856506, + 0.2500136196613312, + 0.031765811145305634, + 0.33726897835731506, + 0.02559337206184864, + 0.018235087394714355, + 0.10230541974306107, + 0.023836100473999977, + 0.008052871562540531, + 0.00597964134067297 + ], + "credit_map": [ + 0.011944087222218513, + 0.07494207471609116, + 0.004729495849460363, + 0.011234022676944733, + 0.013585743494331837, + 0.011045366525650024, + 0.013656800612807274, + 0.0558115616440773, + 0.015771295875310898, + 0.07462076097726822, + 0.007200252264738083, + 0.020465807989239693, + 0.026379847899079323, + 0.02822614274919033, + 0.021980326622724533, + 0.05536916106939316, + 0.003579629585146904, + 0.008127077482640743, + 0.001877833972685039, + 0.005251238122582436, + 0.0019754788372665644, + 0.0018750454764813185, + 0.0036636798176914454, + 0.00541582889854908, + 0.01883561909198761, + 0.1321088820695877, + 0.008908824995160103, + 0.012959528714418411, + 0.02372804470360279, + 0.014978060498833656, + 0.024699343368411064, + 0.10105067491531372, + 0.002055765362456441, + 0.007137706968933344, + 0.0009584298823028803, + 0.0021530610974878073, + 0.0016597595531493425, + 0.0023820996284484863, + 0.003597550792619586, + 0.005648997612297535, + 0.0017569835763424635, + 0.0032932308968156576, + 0.0012098693987354636, + 0.0029119672253727913, + 0.0012839195551350713, + 0.001832597772590816, + 0.003579623531550169, + 0.002366894157603383, + 0.0107954703271389, + 0.010229649022221565, + 0.00715372571721673, + 0.01786782220005989, + 0.010946601629257202, + 0.011547718197107315, + 0.019115474075078964, + 0.014648963697254658, + 0.003270745510235429, + 0.0019203810952603817, + 0.0019295171368867159, + 0.0028023505583405495, + 0.0027945127803832293, + 0.0036943540908396244, + 0.003987085539847612, + 0.003437155159190297, + 0.0008791617001406848, + 0.0005587836494669318, + 0.0005196384154260159, + 0.000979037256911397, + 0.0008274983265437186, + 0.0018851237837225199, + 0.001141777029260993, + 0.0012618514010682702, + 0.0012669319985434413, + 0.0005232703406363726, + 0.00047036525211296976, + 0.0006380731938406825, + 0.00048093812074512243, + 0.0009326618746854365, + 0.0009014763054437935, + 0.0007659238763153553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.011086766608059406 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761759281158447 + ], + "opposite_logprob_tokens": [ + -0.8261759281158447 + ], + "logprob_reward": -0.5761759281158447, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761759281158447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08668849617242813, + 0.09384593367576599, + 0.059260863810777664, + 0.5505045056343079, + 0.06461696326732635, + 0.026095645502209663, + 0.0549391433596611, + 0.010166386142373085, + 0.011506831273436546, + 0.04237521439790726 + ], + "credit_map": [ + 0.006688213907182217, + 0.01743672974407673, + 0.0040809535421431065, + 0.022693168371915817, + 0.004940547980368137, + 0.006948874332010746, + 0.010000240057706833, + 0.013899771496653557, + 0.007851658388972282, + 0.020245322957634926, + 0.0038342857733368874, + 0.01420583575963974, + 0.00630553113296628, + 0.010615267790853977, + 0.011771034449338913, + 0.01901700533926487, + 0.0049618324264883995, + 0.017657475546002388, + 0.0024757678620517254, + 0.007509779650717974, + 0.004094958771020174, + 0.005038559436798096, + 0.00563149806112051, + 0.011890996247529984, + 0.030380932614207268, + 0.2154572606086731, + 0.014478564262390137, + 0.02108493633568287, + 0.03851981833577156, + 0.024747442454099655, + 0.04071894660592079, + 0.16511662304401398, + 0.003692544996738434, + 0.022365489974617958, + 0.001777387224137783, + 0.003491105744615197, + 0.0032475260086357594, + 0.0042257653549313545, + 0.006917374208569527, + 0.018899772316217422, + 0.0037919054739177227, + 0.002008922165259719, + 0.0023862458765506744, + 0.004287194926291704, + 0.00174752960447222, + 0.0033444783184677362, + 0.005027315113693476, + 0.0035020567011088133, + 0.005708395037800074, + 0.005443437024950981, + 0.004217024426907301, + 0.009493377059698105, + 0.005949617363512516, + 0.0058885179460048676, + 0.010861742310225964, + 0.007377035915851593, + 0.0016079482156783342, + 0.0010551413288339972, + 0.0007861749618314207, + 0.001221736310981214, + 0.0009322386467829347, + 0.0011613158276304603, + 0.0019238743698224425, + 0.0014779566554352641, + 0.0017315307632088661, + 0.0010546408593654633, + 0.0008178515709005296, + 0.0019592628814280033, + 0.0008691188413649797, + 0.0012867114273831248, + 0.0022740401327610016, + 0.0015136756701394916, + 0.004332728683948517, + 0.0034604675602167845, + 0.004016665276139975, + 0.0079001784324646, + 0.0032261924352496862, + 0.004633110947906971, + 0.007762676104903221, + 0.007043200545012951 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.5761759281158447 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, still photograph taken from a slightly elevated angle captures a tiny, iridescent black beetle perched on the richly textured, dark gray plastic dashboard of a moving car. Intense, direct sunlight filtering through the windshield casts harsh, bright spots on the beetle's smooth carapace and creates sharp, elongated micro-shadows beneath its body. The entire surface vibrates with the vehicle's subtle, high-frequency vibrations, causing the beetle to desperately search for its place. Its six spindly legs frantically wriggle, sliding across the smooth, sun-warmed plastic, its antennae twitching nervously. It periodically grasps the raised veins of the dashboard's texture, only to be jolted again by the constant movement, sliding a few millimeters with each jolt. In the background, a bright, overexposed landscape of sun-bleached road and passing trees, completely blurred through the out-of-focus windshield, emphasizes the car's speed and creates a tense, slightly dramatic atmosphere focused entirely on this tiny creature's Herculean efforts to survive.", + "seed": 24430, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_2/index_5400_seed_24430_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.223550545796752, + "vlm_logprob": -0.223550545796752, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.223550545796752, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0365549772977829 + ], + "opposite_logprob_tokens": [ + -3.536555051803589 + ], + "logprob_reward": -0.0365549772977829, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0365549772977829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06647482514381409, + 0.031997568905353546, + 0.18522973358631134, + 0.4040450155735016, + 0.026379216462373734, + 0.06767044216394424, + 0.03385002538561821, + 0.0323752760887146, + 0.1090497076511383, + 0.04292820021510124 + ], + "credit_map": [ + 0.01241588220000267, + 0.014240533113479614, + 0.0032900406513363123, + 0.004523913376033306, + 0.006916244514286518, + 0.0059022908098995686, + 0.011884057894349098, + 0.007301865145564079, + 0.005220921244472265, + 0.004670695401728153, + 0.002083825645968318, + 0.0018691937439143658, + 0.005595048889517784, + 0.004460392519831657, + 0.004492031875997782, + 0.003605460748076439, + 0.057246070355176926, + 0.06410613656044006, + 0.006044487934559584, + 0.006258875597268343, + 0.013691863045096397, + 0.016125988215208054, + 0.012768598273396492, + 0.008987697772681713, + 0.12855280935764313, + 0.12590759992599487, + 0.012054752558469772, + 0.015220589004456997, + 0.03245870769023895, + 0.02174346148967743, + 0.04113427922129631, + 0.02697281911969185, + 0.005504906177520752, + 0.0059574078768491745, + 0.0015322727849707007, + 0.001441366970539093, + 0.00256309169344604, + 0.003993482328951359, + 0.00316215050406754, + 0.0022245391737669706, + 0.020652351900935173, + 0.02153754234313965, + 0.0025004788767546415, + 0.003960463684052229, + 0.0039627631194889545, + 0.0039253830909729, + 0.006176759023219347, + 0.004954703152179718, + 0.008713400922715664, + 0.00411932822316885, + 0.0024943333119153976, + 0.0020684930495917797, + 0.0027488789055496454, + 0.004724366590380669, + 0.005485783331096172, + 0.003495441749691963, + 0.00836904440075159, + 0.007857141084969044, + 0.0016898497706279159, + 0.0018296502530574799, + 0.0025017419829964638, + 0.0026436138432472944, + 0.003748305607587099, + 0.003735929261893034, + 0.035539522767066956, + 0.017438936978578568, + 0.0055568101815879345, + 0.0068658567033708096, + 0.006934172939509153, + 0.014775318093597889, + 0.009916789829730988, + 0.01202230155467987, + 0.009945600293576717, + 0.007387316320091486, + 0.002356843091547489, + 0.005171420052647591, + 0.0027661824133247137, + 0.00440675625577569, + 0.004811367020010948, + 0.006082712672650814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.0365549772977829 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038113825023174286 + ], + "opposite_logprob_tokens": [ + -3.288113832473755 + ], + "logprob_reward": -0.038113825023174286, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038113825023174286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052463896572589874, + 0.03592544421553612, + 0.2870400846004486, + 0.0938904732465744, + 0.0602479912340641, + 0.36654338240623474, + 0.04517342522740364, + 0.023978516459465027, + 0.015615608543157578, + 0.019121170043945312 + ], + "credit_map": [ + 0.013163913041353226, + 0.011161988601088524, + 0.0028469127137213945, + 0.002702345373108983, + 0.0049322727136313915, + 0.006797562353312969, + 0.006639260798692703, + 0.004219642840325832, + 0.008260495029389858, + 0.008565298281610012, + 0.0015772894257679582, + 0.0017834878526628017, + 0.004706191830337048, + 0.004902706015855074, + 0.0034117302857339382, + 0.002718247938901186, + 0.11420384049415588, + 0.12414266169071198, + 0.004330380354076624, + 0.005783720873296261, + 0.013006307184696198, + 0.012161589227616787, + 0.007473849691450596, + 0.0059377411380410194, + 0.033514056354761124, + 0.03409211337566376, + 0.00229645986109972, + 0.0037388924974948168, + 0.0065327975898981094, + 0.004823280498385429, + 0.005264948587864637, + 0.0036279181949794292, + 0.012345329858362675, + 0.014025002717971802, + 0.0026485321577638388, + 0.0026678668800741434, + 0.007022453937679529, + 0.00819206889718771, + 0.008617849089205265, + 0.004728888627141714, + 0.11354754120111465, + 0.12649914622306824, + 0.009579169563949108, + 0.017003031447529793, + 0.02359013631939888, + 0.020142612978816032, + 0.03234619274735451, + 0.02383553981781006, + 0.01168802659958601, + 0.005796627141535282, + 0.00302732246927917, + 0.002630306174978614, + 0.00509105296805501, + 0.006941249128431082, + 0.005140356719493866, + 0.004858485888689756, + 0.004676837008446455, + 0.005192642565816641, + 0.0015246145194396377, + 0.001593175227753818, + 0.0021448207553476095, + 0.0024582836776971817, + 0.002914383541792631, + 0.0034737586975097656, + 0.003782586893066764, + 0.0015706742415204644, + 0.0007659260299988091, + 0.0016583952819928527, + 0.0016112793236970901, + 0.001532542984932661, + 0.002627738518640399, + 0.002066465560346842, + 0.0032119315583258867, + 0.003459981642663479, + 0.0014112413628026843, + 0.002018081024289131, + 0.001793500268831849, + 0.0018721765372902155, + 0.002725274069234729, + 0.002628984395414591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.038113825023174286 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204738840460777 + ], + "opposite_logprob_tokens": [ + -2.8120474815368652 + ], + "logprob_reward": -0.06204738840460777, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204738840460777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14142271876335144, + 0.07124816626310349, + 0.3763642907142639, + 0.04822542145848274, + 0.08789882808923721, + 0.13731200993061066, + 0.08151239901781082, + 0.01661819778382778, + 0.013475622981786728, + 0.025922294706106186 + ], + "credit_map": [ + 0.03904154896736145, + 0.039831630885601044, + 0.0069846236146986485, + 0.006714141461998224, + 0.011248131282627583, + 0.01593419350683689, + 0.011470604687929153, + 0.010197849944233894, + 0.018246987834572792, + 0.017662007361650467, + 0.0040819174610078335, + 0.0036692586727440357, + 0.006330076139420271, + 0.007805803325027227, + 0.006692690774798393, + 0.006759425159543753, + 0.1272745132446289, + 0.14270134270191193, + 0.011307560838758945, + 0.010064719244837761, + 0.026272177696228027, + 0.026621662080287933, + 0.018418634310364723, + 0.013703681528568268, + 0.010271071456372738, + 0.011650527827441692, + 0.0027949754148721695, + 0.0021606911905109882, + 0.005475352052599192, + 0.005677568726241589, + 0.005826873239129782, + 0.004368361551314592, + 0.011944449506700039, + 0.013093914836645126, + 0.005891580134630203, + 0.005441301967948675, + 0.011829781346023083, + 0.0164753720164299, + 0.014256150461733341, + 0.008966277353465557, + 0.04160838946700096, + 0.046448688954114914, + 0.003986909985542297, + 0.0066802725195884705, + 0.008458727039396763, + 0.0078950896859169, + 0.013181735761463642, + 0.00905220303684473, + 0.020836329087615013, + 0.010514949448406696, + 0.006978943943977356, + 0.005138784646987915, + 0.008841213770210743, + 0.01312519982457161, + 0.008258120156824589, + 0.007818859070539474, + 0.00279436307027936, + 0.002316378289833665, + 0.001021030475385487, + 0.0011788286501541734, + 0.0015765974530950189, + 0.0020298820454627275, + 0.0032923517283052206, + 0.0024087654892355204, + 0.0032977901864796877, + 0.002673034556210041, + 0.0006174414302222431, + 0.0008494331850670278, + 0.0011294286232441664, + 0.0013125534169375896, + 0.0019254134967923164, + 0.0016705277375876904, + 0.00354630290530622, + 0.005102241411805153, + 0.00127036077901721, + 0.003368673613294959, + 0.0019387537613511086, + 0.0022762014996260405, + 0.004001350607722998, + 0.004418413620442152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.06204738840460777 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741398945450783 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741398945450783, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741398945450783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07389819622039795, + 0.05284503474831581, + 0.6107118129730225, + 0.03788300231099129, + 0.05120629072189331, + 0.027795951813459396, + 0.017961805686354637, + 0.036595895886421204, + 0.01140774879604578, + 0.07969433814287186 + ], + "credit_map": [ + 0.020347611978650093, + 0.020626293495297432, + 0.004566363524645567, + 0.0034900158643722534, + 0.007126973941922188, + 0.007755361497402191, + 0.00517240259796381, + 0.004813161212950945, + 0.010222881101071835, + 0.008546599186956882, + 0.0028780242428183556, + 0.0029651399236172438, + 0.00783002469688654, + 0.008712424896657467, + 0.006495162378996611, + 0.005194768309593201, + 0.23111344873905182, + 0.2538008689880371, + 0.01303874235600233, + 0.013139821588993073, + 0.031537432223558426, + 0.032418154180049896, + 0.019486790522933006, + 0.016176443547010422, + 0.006998192518949509, + 0.00981332827359438, + 0.0018217426259070635, + 0.002964199287816882, + 0.004200153052806854, + 0.003963578026741743, + 0.00424935994669795, + 0.003872442990541458, + 0.012639503926038742, + 0.016498487442731857, + 0.0017624861793592572, + 0.002121967962011695, + 0.004367412067949772, + 0.006261860020458698, + 0.004334470722824335, + 0.003220096928998828, + 0.006935172248631716, + 0.008862733840942383, + 0.0008414912736043334, + 0.0011564744636416435, + 0.0020753187127411366, + 0.002777606248855591, + 0.0029933410696685314, + 0.0021538096480071545, + 0.0045976173132658005, + 0.0026017760392278433, + 0.00149638787843287, + 0.001006494159810245, + 0.0020539446268230677, + 0.0020341931376606226, + 0.0023842391092330217, + 0.00178715109359473, + 0.009486645460128784, + 0.007081782445311546, + 0.0022262376733124256, + 0.0021496100816875696, + 0.0026882877573370934, + 0.004881128668785095, + 0.0035660825669765472, + 0.0045161135494709015, + 0.0025874690618366003, + 0.0017887685680761933, + 0.00046441034646704793, + 0.000965102284681052, + 0.0007809282396920025, + 0.0008867409778758883, + 0.0024033670779317617, + 0.0015309605514630675, + 0.011973060667514801, + 0.017680320888757706, + 0.004311985336244106, + 0.008740770630538464, + 0.006094784010201693, + 0.010245880112051964, + 0.010447093285620213, + 0.010200431570410728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.006741398945450783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742951393127441 + ], + "opposite_logprob_tokens": [ + -0.47429513931274414 + ], + "logprob_reward": -0.9742951393127441, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742951393127441, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06191849708557129, + 0.02553238905966282, + 0.2813471853733063, + 0.349307119846344, + 0.03267604112625122, + 0.10541358590126038, + 0.035312045365571976, + 0.03046446479856968, + 0.043303679674863815, + 0.03472498804330826 + ], + "credit_map": [ + 0.011048292741179466, + 0.009830532595515251, + 0.0032502240501344204, + 0.0031494067516177893, + 0.00976578053086996, + 0.008754154667258263, + 0.00953596830368042, + 0.006584141403436661, + 0.0034923588391393423, + 0.0039189402014017105, + 0.001169513096101582, + 0.0017629694193601608, + 0.004061496816575527, + 0.004433422815054655, + 0.004022008739411831, + 0.00267167785204947, + 0.09619710594415665, + 0.10741026699542999, + 0.0075786905363202095, + 0.008866089396178722, + 0.019946066662669182, + 0.01753821410238743, + 0.013922306708991528, + 0.009888449683785439, + 0.1120319589972496, + 0.11273937672376633, + 0.009145809337496758, + 0.012967261485755444, + 0.03359397128224373, + 0.01838397979736328, + 0.030214931815862656, + 0.020229868590831757, + 0.007746101822704077, + 0.008885152637958527, + 0.0013981860829517245, + 0.001443576067686081, + 0.0029327834490686655, + 0.004767200443893671, + 0.003315941197797656, + 0.0021871007047593594, + 0.03710009157657623, + 0.03728039562702179, + 0.00214815023355186, + 0.003470846451818943, + 0.005314004607498646, + 0.00525164557620883, + 0.008384320884943008, + 0.006464133504778147, + 0.012092598713934422, + 0.005425849463790655, + 0.0019248133758082986, + 0.001449413481168449, + 0.0035349028185009956, + 0.004733296576887369, + 0.0030843354761600494, + 0.003066837089136243, + 0.008755003102123737, + 0.004825997166335583, + 0.0019113613525405526, + 0.002035905374214053, + 0.0023101377300918102, + 0.004252488259226084, + 0.0035634287633001804, + 0.0028101452626287937, + 0.012808462604880333, + 0.006986933760344982, + 0.0021996984723955393, + 0.00290388660505414, + 0.0027564119081944227, + 0.005966499447822571, + 0.0044182357378304005, + 0.0052635520696640015, + 0.005162924062460661, + 0.006565967109054327, + 0.001736000063829124, + 0.004432235844433308, + 0.0022327520418912172, + 0.003927621524780989, + 0.005171542521566153, + 0.0054959459230303764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.9742951393127441 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extreme close-up, still photograph taken from a slightly elevated angle captures a tiny, iridescent black beetle perched on the richly textured, dark gray plastic dashboard of a moving car. Intense, direct sunlight filtering through the windshield casts harsh, bright spots on the beetle's smooth carapace and creates sharp, elongated micro-shadows beneath its body. The entire surface vibrates with the vehicle's subtle, high-frequency vibrations, causing the beetle to desperately search for its place. Its six spindly legs frantically wriggle, sliding across the smooth, sun-warmed plastic, its antennae twitching nervously. It periodically grasps the raised veins of the dashboard's texture, only to be jolted again by the constant movement, sliding a few millimeters with each jolt. In the background, a bright, overexposed landscape of sun-bleached road and passing trees, completely blurred through the out-of-focus windshield, emphasizes the car's speed and creates a tense, slightly dramatic atmosphere focused entirely on this tiny creature's Herculean efforts to survive.", + "seed": 24431, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_2/index_5400_seed_24431_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17706229379400612, + "vlm_logprob": -0.17706229379400612, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.17706229379400612, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06586895883083344 + ], + "opposite_logprob_tokens": [ + -2.940868854522705 + ], + "logprob_reward": -0.06586895883083344, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06586895883083344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1496163308620453, + 0.2263779193162918, + 0.2515040636062622, + 0.12408124655485153, + 0.03653177618980408, + 0.042160019278526306, + 0.01769939810037613, + 0.045624990016222, + 0.0843462124466896, + 0.022058075293898582 + ], + "credit_map": [ + 0.018620483577251434, + 0.03326716646552086, + 0.01313698384910822, + 0.010662224143743515, + 0.014660824090242386, + 0.028482835739850998, + 0.01947486773133278, + 0.011310950852930546, + 0.019239136949181557, + 0.06766027957201004, + 0.008961096405982971, + 0.007509410846978426, + 0.036838166415691376, + 0.05260015279054642, + 0.02322613075375557, + 0.01034353394061327, + 0.018240509554743767, + 0.04740581661462784, + 0.010242598131299019, + 0.014288672246038914, + 0.06385359168052673, + 0.05624166503548622, + 0.022971566766500473, + 0.018259640783071518, + 0.009274322539567947, + 0.01245320588350296, + 0.011798802763223648, + 0.01845529116690159, + 0.015310346148908138, + 0.01611732318997383, + 0.02682632766664028, + 0.013845626264810562, + 0.0029582923743873835, + 0.004168597515672445, + 0.003680138848721981, + 0.0027361195534467697, + 0.007219099439680576, + 0.005779932253062725, + 0.00643460126593709, + 0.003554995870217681, + 0.005267447791993618, + 0.005941503681242466, + 0.003847397631034255, + 0.0024920846335589886, + 0.004865394905209541, + 0.005129293538630009, + 0.008850602433085442, + 0.005766292568296194, + 0.0027240209747105837, + 0.002579756546765566, + 0.0015753648476675153, + 0.0011407617712393403, + 0.0019022218184545636, + 0.0020195709075778723, + 0.0037984049413353205, + 0.0019592957105487585, + 0.00533098541200161, + 0.006128733512014151, + 0.003232523798942566, + 0.0031459848396480083, + 0.009550167247653008, + 0.005941349547356367, + 0.007481937762349844, + 0.004813303705304861, + 0.009660044685006142, + 0.011832364834845066, + 0.007767928764224052, + 0.006418668199330568, + 0.00916769728064537, + 0.009064635261893272, + 0.017647098749876022, + 0.012787776999175549, + 0.002463496057316661, + 0.0021153271663933992, + 0.0025286769960075617, + 0.001347305136732757, + 0.002400577301159501, + 0.0018960057059302926, + 0.005119544453918934, + 0.004187141545116901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.06586895883083344 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018210716545581818 + ], + "opposite_logprob_tokens": [ + -4.0182108879089355 + ], + "logprob_reward": -0.018210716545581818, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018210716545581818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050322987139225006, + 0.14277446269989014, + 0.5043501853942871, + 0.03755933418869972, + 0.03726058453321457, + 0.034818440675735474, + 0.035916149616241455, + 0.028993552550673485, + 0.08495701104402542, + 0.04304732009768486 + ], + "credit_map": [ + 0.007313285022974014, + 0.007051569409668446, + 0.0058771101757884026, + 0.004169079475104809, + 0.006367906462401152, + 0.005619417410343885, + 0.006488921120762825, + 0.0074356901459395885, + 0.019664892926812172, + 0.017844093963503838, + 0.01049491111189127, + 0.009983843192458153, + 0.02947571687400341, + 0.02426873706281185, + 0.018280532211065292, + 0.012761722318828106, + 0.041958749294281006, + 0.09947890043258667, + 0.028172913938760757, + 0.026232756674289703, + 0.11381945759057999, + 0.11651027947664261, + 0.042406726628541946, + 0.03577038273215294, + 0.002637303899973631, + 0.005054567940533161, + 0.0037008728832006454, + 0.005155017599463463, + 0.006134286522865295, + 0.005947995465248823, + 0.005689007230103016, + 0.003240280319005251, + 0.0032876983750611544, + 0.005472241900861263, + 0.0027838849928230047, + 0.002966751344501972, + 0.00805758498609066, + 0.005145813338458538, + 0.006431526970118284, + 0.003115076804533601, + 0.004372322466224432, + 0.004502700641751289, + 0.0041301436722278595, + 0.003598729846999049, + 0.0041591450572013855, + 0.0037488292437046766, + 0.005233133211731911, + 0.005073432810604572, + 0.0036224275827407837, + 0.004216684028506279, + 0.0041915555484592915, + 0.003292255802080035, + 0.004131943453103304, + 0.003753225551918149, + 0.007465968374162912, + 0.005242085084319115, + 0.004572100937366486, + 0.005251572001725435, + 0.002495348919183016, + 0.0017929967725649476, + 0.0034111973363906145, + 0.003355009015649557, + 0.004285060800611973, + 0.0038302636239677668, + 0.009653333574533463, + 0.011847956106066704, + 0.009980540722608566, + 0.00700137484818697, + 0.011227278038859367, + 0.007460570894181728, + 0.012736433185636997, + 0.015049517154693604, + 0.00507982075214386, + 0.00526020722463727, + 0.004326849244534969, + 0.0028061422053724527, + 0.0047049312852323055, + 0.005532472860068083, + 0.009130376391112804, + 0.006206518970429897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.018210716545581818 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811921924352646 + ], + "opposite_logprob_tokens": [ + -3.288119316101074 + ], + "logprob_reward": -0.03811921924352646, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811921924352646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15041352808475494, + 0.09188788384199142, + 0.30429115891456604, + 0.0553889237344265, + 0.0601993203163147, + 0.05382267013192177, + 0.029681401327252388, + 0.08488469570875168, + 0.11265186220407486, + 0.05677855759859085 + ], + "credit_map": [ + 0.01891240105032921, + 0.023668812587857246, + 0.015019502490758896, + 0.012774779461324215, + 0.021775413304567337, + 0.021173011511564255, + 0.022854033857584, + 0.014235553331673145, + 0.007962045259773731, + 0.028930777683854103, + 0.005315569695085287, + 0.004937531426548958, + 0.0087197907269001, + 0.01970379613339901, + 0.01057239156216383, + 0.005745976697653532, + 0.02441692166030407, + 0.05639837309718132, + 0.01767968386411667, + 0.017519861459732056, + 0.07275243103504181, + 0.0701972246170044, + 0.024070359766483307, + 0.02125629596412182, + 0.005087682977318764, + 0.005267013795673847, + 0.004564619157463312, + 0.009677676483988762, + 0.007455827202647924, + 0.0067479549907147884, + 0.010062258690595627, + 0.006525889970362186, + 0.005751479882746935, + 0.00783501472324133, + 0.0032161481212824583, + 0.00298299640417099, + 0.01681344583630562, + 0.013879708014428616, + 0.004694031085819006, + 0.0050264932215213776, + 0.007134157698601484, + 0.007335917558521032, + 0.0048829661682248116, + 0.004366661421954632, + 0.006472662091255188, + 0.006347997579723597, + 0.008862649090588093, + 0.008419659920036793, + 0.002484010299667716, + 0.0029706056229770184, + 0.0038413105066865683, + 0.002485126256942749, + 0.0051646181382238865, + 0.003622042713686824, + 0.004614060278981924, + 0.004499628208577633, + 0.012350359000265598, + 0.015611708164215088, + 0.006473756395280361, + 0.007933699525892735, + 0.008639453910291195, + 0.00945530366152525, + 0.013330609537661076, + 0.011089804582297802, + 0.013681898824870586, + 0.011981220915913582, + 0.010719935409724712, + 0.012165121734142303, + 0.011483422480523586, + 0.01002274639904499, + 0.0245620459318161, + 0.01803547888994217, + 0.0073741646483540535, + 0.009020826779305935, + 0.0038934408221393824, + 0.0031346259638667107, + 0.009202436544001102, + 0.006776564754545689, + 0.010040294378995895, + 0.007336207199841738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.03811921924352646 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260079633444548 + ], + "opposite_logprob_tokens": [ + -5.255259990692139 + ], + "logprob_reward": -0.005260079633444548, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260079633444548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13751225173473358, + 0.09317833930253983, + 0.32330062985420227, + 0.05422895774245262, + 0.15295332670211792, + 0.03783193975687027, + 0.04272931441664696, + 0.02903670445084572, + 0.05642181262373924, + 0.07280674576759338 + ], + "credit_map": [ + 0.028101161122322083, + 0.026765674352645874, + 0.011044259183108807, + 0.008969899266958237, + 0.016495533287525177, + 0.020596347749233246, + 0.014294782653450966, + 0.011244598776102066, + 0.0086335064843297, + 0.01509883627295494, + 0.004782040603458881, + 0.004363850224763155, + 0.025195272639393806, + 0.015469247475266457, + 0.01056844461709261, + 0.009067140519618988, + 0.026670951396226883, + 0.06240638345479965, + 0.018556762486696243, + 0.01747916080057621, + 0.07641506940126419, + 0.07468798011541367, + 0.025630686432123184, + 0.021453646942973137, + 0.004280800465494394, + 0.007004172075539827, + 0.004663530737161636, + 0.005324020516127348, + 0.009926664642989635, + 0.010583043098449707, + 0.007361531723290682, + 0.005085194483399391, + 0.01778809353709221, + 0.025372842326760292, + 0.007763855159282684, + 0.007024004589766264, + 0.0385235920548439, + 0.0349002443253994, + 0.010834651067852974, + 0.010746038518846035, + 0.003860847558826208, + 0.005199905019253492, + 0.004112725146114826, + 0.003322838805615902, + 0.004420829471200705, + 0.004091127775609493, + 0.008126659318804741, + 0.004697006661444902, + 0.0037687858566641808, + 0.004865964408963919, + 0.004978077486157417, + 0.0037522409111261368, + 0.005897591356188059, + 0.004818929359316826, + 0.008410487323999405, + 0.006237236317247152, + 0.003936148248612881, + 0.005263213533908129, + 0.0022116980981081724, + 0.002077181125059724, + 0.0031949763651937246, + 0.0034272882621735334, + 0.00510740652680397, + 0.0038187899626791477, + 0.006589248310774565, + 0.008422723971307278, + 0.0052675218321383, + 0.003185404697433114, + 0.005879167467355728, + 0.005894591566175222, + 0.011850121431052685, + 0.009333031252026558, + 0.01154694426804781, + 0.01263775210827589, + 0.007186499889940023, + 0.0036061392165720463, + 0.007692396640777588, + 0.010397868230938911, + 0.00976179726421833, + 0.009977341629564762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.005260079633444548 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a tiny, iridescent black beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle perched on a dark gray plastic dashboard with raised veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does intense direct sunlight create harsh bright spots and sharp elongated micro-shadows on the beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a bright, overexposed landscape of a sun-bleached road and passing trees that is completely blurred?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the beetle exhibit frantic wriggling legs and antennae twitching while sliding across the vibrating surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578524947166443 + ], + "opposite_logprob_tokens": [ + -0.6328524947166443 + ], + "logprob_reward": -0.7578524947166443, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578524947166443, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15221306681632996, + 0.04600653797388077, + 0.5203019976615906, + 0.029917709529399872, + 0.019048139452934265, + 0.019604109227657318, + 0.009607192128896713, + 0.04163480922579765, + 0.14617270231246948, + 0.015493791550397873 + ], + "credit_map": [ + 0.026452992111444473, + 0.028877167031168938, + 0.012197877280414104, + 0.011334872804582119, + 0.018374091014266014, + 0.02067403309047222, + 0.020701520144939423, + 0.013600512407720089, + 0.004839090630412102, + 0.013862025924026966, + 0.0028962637297809124, + 0.0028244901914149523, + 0.005391843616962433, + 0.008483006618916988, + 0.004500051494687796, + 0.0032097615767270327, + 0.04241110756993294, + 0.09958165138959885, + 0.027691084891557693, + 0.027823811396956444, + 0.1207350641489029, + 0.11655396223068237, + 0.04920068755745888, + 0.03630456328392029, + 0.002204099204391241, + 0.0030772704631090164, + 0.0021852722857147455, + 0.006010665558278561, + 0.0038136886432766914, + 0.003381652059033513, + 0.005825044121593237, + 0.003420015564188361, + 0.0025829579681158066, + 0.0036939445417374372, + 0.0012596850283443928, + 0.0011634898837655783, + 0.003416795749217272, + 0.003507533809170127, + 0.0016967745032161474, + 0.0017269561067223549, + 0.0024045798927545547, + 0.0031588333658874035, + 0.0018578622257336974, + 0.001719829859212041, + 0.002513645915314555, + 0.002321353880688548, + 0.003082975745201111, + 0.0025450263638049364, + 0.0008742950740270317, + 0.0010083867236971855, + 0.0007628882303833961, + 0.0007994628394953907, + 0.0011236529098823667, + 0.0008756324532441795, + 0.002664232160896063, + 0.0014986402820795774, + 0.005527622532099485, + 0.003951602149754763, + 0.0034103828947991133, + 0.0028960041236132383, + 0.0054243747144937515, + 0.00458760978654027, + 0.010506276972591877, + 0.005330927670001984, + 0.018592681735754013, + 0.024760955944657326, + 0.01249600574374199, + 0.009004903957247734, + 0.02234463207423687, + 0.017978278920054436, + 0.024916095659136772, + 0.016079135239124298, + 0.002464542631059885, + 0.002836468629539013, + 0.0017210441874340177, + 0.001199877355247736, + 0.0018609716789796948, + 0.0022801996674388647, + 0.0016616644570603967, + 0.001469021663069725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.7578524947166443 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_3/index_tensor([6737, 6737])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_3/index_tensor([6737, 6737])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ff17cd992e09ec7c4327d49e13778f90fb307869 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_3/index_tensor([6737, 6737])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Immersive, frontal underwater filming captures the scene as the camera pans backward at a slow, steady pace, maintaining a constant distance from the subject. The setting is the deep, silent depths of the ocean, a rich palette of deep sapphires and cobalt blues. From afar, above the shimmering, rippled surface, bright, sharply defined golden sunlight penetrates the abyss, creating ethereal pillars of light that illuminate countless tiny suspended particles, making light itself appear tangible. A toned, athletic female figure is perfectly centered in the frame, gracefully swimming directly toward the viewer. She wears a simple, dark navy one-piece swimsuit that contrasts sharply with her skin, which is intermittently speckled by passing sunlight. Her long, dark brown hair flows freely behind her, undulating with the current like dark ribbons. She moves with calm purpose, executing a slow, rhythmic breaststroke; her arms deliberately reach forward, then sweep back in a powerful, fluid motion, effortlessly propelling her through the water. With each gentle exhalation, she releases a small cloud of silver bubbles that absorb the light and shimmer as they rise. Her facial expression, calm and focused, her eyes fixed forward, creates an atmosphere of stunning, almost spiritual beauty, evoking a profound sense of tranquility and freedom in the vast, silent underwater world.", + "seed": 18478, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_3/index_6737_seed_18478_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04153060410171747, + "vlm_logprob": -0.04153060410171747, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.04153060410171747, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007237762212753296 + ], + "opposite_logprob_tokens": [ + -5.632237911224365 + ], + "logprob_reward": -0.007237762212753296, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.007237762212753296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2482651174068451, + 0.044845100492239, + 0.03681177645921707, + 0.04556521400809288, + 0.09720588475465775, + 0.08250796049833298, + 0.1210811659693718, + 0.1909821629524231, + 0.09020335972309113, + 0.042532339692115784 + ], + "credit_map": [ + 0.03776077553629875, + 0.013298253528773785, + 0.06703242659568787, + 0.015177165158092976, + 0.02050935849547386, + 0.052139777690172195, + 0.011767338030040264, + 0.03058001957833767, + 0.007533021736890078, + 0.00458309194073081, + 0.006046953611075878, + 0.00482568284496665, + 0.00453927181661129, + 0.006367122754454613, + 0.0051391590386629105, + 0.005810795351862907, + 0.008043515495955944, + 0.0036547179333865643, + 0.005126990843564272, + 0.0039504701271653175, + 0.004260136280208826, + 0.004189435858279467, + 0.004686295986175537, + 0.0029002116061747074, + 0.007485855836421251, + 0.003612845903262496, + 0.005924278404563665, + 0.004794684704393148, + 0.0045058573596179485, + 0.007973674684762955, + 0.005440327804535627, + 0.005827692802995443, + 0.019696274772286415, + 0.008183027617633343, + 0.01609724946320057, + 0.005299297161400318, + 0.005362755618989468, + 0.024227293208241463, + 0.004639729391783476, + 0.013700257986783981, + 0.009302109479904175, + 0.00471611600369215, + 0.02298983372747898, + 0.005202310625463724, + 0.004061781335622072, + 0.011030664667487144, + 0.005046159029006958, + 0.02015899121761322, + 0.014419694431126118, + 0.0046996865421533585, + 0.038146089762449265, + 0.006535468157380819, + 0.0032814221922308207, + 0.012285982258617878, + 0.005172112490981817, + 0.036540694534778595, + 0.02536921761929989, + 0.01841903291642666, + 0.028290806338191032, + 0.023344220593571663, + 0.0131376963108778, + 0.034189458936452866, + 0.022004177793860435, + 0.0262275543063879, + 0.012519633397459984, + 0.009533457458019257, + 0.010020256973803043, + 0.013392893597483635, + 0.008507448248565197, + 0.00959972757846117, + 0.017647339031100273, + 0.008982609026134014, + 0.007829559966921806, + 0.003401807975023985, + 0.005786587949842215, + 0.005368059501051903, + 0.003809112822636962, + 0.005764056462794542, + 0.005462178494781256, + 0.005110976751893759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.007237762212753296 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953910294920206 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.005953910294920206, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953910294920206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32486921548843384, + 0.07144545763731003, + 0.040217529982328415, + 0.09081190824508667, + 0.07367270439863205, + 0.06643025577068329, + 0.11570889502763748, + 0.0529942624270916, + 0.07865312695503235, + 0.08519666641950607 + ], + "credit_map": [ + 0.03507040813565254, + 0.014113474637269974, + 0.07906955480575562, + 0.016293538734316826, + 0.038448166102170944, + 0.10525339096784592, + 0.015718193724751472, + 0.020902499556541443, + 0.010036845691502094, + 0.009989430196583271, + 0.008455215953290462, + 0.007578108459711075, + 0.009585083462297916, + 0.00856736022979021, + 0.00956607237458229, + 0.007667339872568846, + 0.006046914029866457, + 0.005401252303272486, + 0.005677275825291872, + 0.005002989899367094, + 0.004172130953520536, + 0.004881043918430805, + 0.00600639870390296, + 0.0030295224860310555, + 0.012094292789697647, + 0.010435197502374649, + 0.009893801063299179, + 0.009241730906069279, + 0.014751075766980648, + 0.007875174283981323, + 0.016495781019330025, + 0.010024847462773323, + 0.012189351953566074, + 0.0065907626412808895, + 0.012585981748998165, + 0.004148622043430805, + 0.007008121814578772, + 0.015634875744581223, + 0.004166843835264444, + 0.011348138563334942, + 0.011105135083198547, + 0.006248641759157181, + 0.012844516895711422, + 0.005755824036896229, + 0.004725574050098658, + 0.007187141105532646, + 0.006675029639154673, + 0.011888386681675911, + 0.014291106723248959, + 0.007203439250588417, + 0.035117875784635544, + 0.010911857709288597, + 0.0038303749170154333, + 0.009420315735042095, + 0.007359618786722422, + 0.027574297040700912, + 0.006727999541908503, + 0.005074677988886833, + 0.008526487275958061, + 0.0073944092728197575, + 0.0035637132823467255, + 0.010680868290364742, + 0.005435795057564974, + 0.005590311251580715, + 0.008561653085052967, + 0.008969074115157127, + 0.009753424674272537, + 0.01202582847326994, + 0.009231386706233025, + 0.008588278666138649, + 0.015686452388763428, + 0.0058370172046124935, + 0.015629079192876816, + 0.008449257351458073, + 0.010893366299569607, + 0.01079220324754715, + 0.009062912315130234, + 0.009913867339491844, + 0.010564868338406086, + 0.009891108609735966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.005953910294920206 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641467429697514 + ], + "opposite_logprob_tokens": [ + -4.758641242980957 + ], + "logprob_reward": -0.008641467429697514, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641467429697514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13984689116477966, + 0.05213916674256325, + 0.0654497742652893, + 0.05272530019283295, + 0.166897714138031, + 0.1582559049129486, + 0.039457064121961594, + 0.11813759058713913, + 0.06346537917852402, + 0.14362522959709167 + ], + "credit_map": [ + 0.016566075384616852, + 0.005986799020320177, + 0.0359543077647686, + 0.006544958800077438, + 0.015866223722696304, + 0.04428780823945999, + 0.0048576099798083305, + 0.009783108718693256, + 0.007437088061124086, + 0.005206183064728975, + 0.006339019164443016, + 0.005623635370284319, + 0.006621699780225754, + 0.007527243811637163, + 0.007551192305982113, + 0.005833104718476534, + 0.008335674181580544, + 0.007382262032479048, + 0.007911982014775276, + 0.008722304366528988, + 0.010165299288928509, + 0.009802366606891155, + 0.008630524389445782, + 0.004499361384660006, + 0.010912387631833553, + 0.005055304616689682, + 0.006534508429467678, + 0.004637823440134525, + 0.006794611923396587, + 0.0038757731672376394, + 0.009122977964580059, + 0.0057919141836464405, + 0.025828802958130836, + 0.013490823097527027, + 0.024653727188706398, + 0.010677295736968517, + 0.018357114866375923, + 0.03671581670641899, + 0.010070898570120335, + 0.02710323967039585, + 0.01466209627687931, + 0.009902560152113438, + 0.04414258524775505, + 0.015285981819033623, + 0.007309693843126297, + 0.015321815386414528, + 0.013314779847860336, + 0.038316383957862854, + 0.004669267684221268, + 0.00191264390014112, + 0.009112212806940079, + 0.0027896591927856207, + 0.0017098822863772511, + 0.008355846628546715, + 0.0021578751038759947, + 0.008749675936996937, + 0.02067265473306179, + 0.010258638300001621, + 0.01722106710076332, + 0.011909234337508678, + 0.007151995785534382, + 0.019764089956879616, + 0.012771104462444782, + 0.018388809636235237, + 0.007243224419653416, + 0.006362182088196278, + 0.007583306171000004, + 0.009263796731829643, + 0.006464656908065081, + 0.00733722560107708, + 0.013045977801084518, + 0.006165007594972849, + 0.0325043722987175, + 0.016168007627129555, + 0.02091592736542225, + 0.014065594412386417, + 0.012960858643054962, + 0.018939971923828125, + 0.015116694383323193, + 0.012953790836036205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.008641467429697514 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987331748008728 + ], + "opposite_logprob_tokens": [ + -1.8048733472824097 + ], + "logprob_reward": -0.17987331748008728, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987331748008728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13107886910438538, + 0.08573304116725922, + 0.04492534324526787, + 0.037492021918296814, + 0.11687131226062775, + 0.14151275157928467, + 0.13789817690849304, + 0.12888804078102112, + 0.07569687068462372, + 0.09990362077951431 + ], + "credit_map": [ + 0.01754274033010006, + 0.0073647103272378445, + 0.028988059610128403, + 0.009204075671732426, + 0.01418242696672678, + 0.034938011318445206, + 0.008535931818187237, + 0.010322915390133858, + 0.022933606058359146, + 0.008568847551941872, + 0.01211442332714796, + 0.007134397979825735, + 0.005750233773142099, + 0.008004212751984596, + 0.008046908304095268, + 0.01318040955811739, + 0.009009727276861668, + 0.0034666014835238457, + 0.006629139184951782, + 0.007040586322546005, + 0.0035878962371498346, + 0.0033018935937434435, + 0.007892927154898643, + 0.003996570594608784, + 0.005877625197172165, + 0.0037341006100177765, + 0.004672628827393055, + 0.002608813811093569, + 0.004538925364613533, + 0.005985486786812544, + 0.005454283207654953, + 0.0046201590448617935, + 0.023289458826184273, + 0.01117834635078907, + 0.019207851961255074, + 0.006694059818983078, + 0.010980761609971523, + 0.0240059532225132, + 0.006728471256792545, + 0.014786400832235813, + 0.014977826736867428, + 0.010413749143481255, + 0.03326375037431717, + 0.012473600916564465, + 0.00847367849200964, + 0.022084053605794907, + 0.01037993747740984, + 0.029446152970194817, + 0.013703216798603535, + 0.010100635699927807, + 0.037750083953142166, + 0.013910952024161816, + 0.00452700350433588, + 0.007922508753836155, + 0.012324369512498379, + 0.03765939176082611, + 0.023600351065397263, + 0.012814396992325783, + 0.01948121003806591, + 0.012639458291232586, + 0.007803007494658232, + 0.01817796379327774, + 0.012602504342794418, + 0.021769147366285324, + 0.008940276689827442, + 0.008150485344231129, + 0.008871368132531643, + 0.011542810127139091, + 0.008104187436401844, + 0.007074427790939808, + 0.016001595184206963, + 0.007011716719716787, + 0.022787347435951233, + 0.00891572143882513, + 0.014552034437656403, + 0.009808015078306198, + 0.008218028582632542, + 0.012998144142329693, + 0.010453103110194206, + 0.012171231210231781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.17987331748008728 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059465630911290646 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.0059465630911290646, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059465630911290646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13654403388500214, + 0.021798694506287575, + 0.07694891840219498, + 0.0650671049952507, + 0.08404669910669327, + 0.04869449511170387, + 0.31131818890571594, + 0.022184062749147415, + 0.021475201472640038, + 0.21192264556884766 + ], + "credit_map": [ + 0.016254236921668053, + 0.005089644808322191, + 0.03707714006304741, + 0.006109386682510376, + 0.015066424384713173, + 0.04311220347881317, + 0.002984229940921068, + 0.010850769467651844, + 0.0038951693568378687, + 0.0021033023949712515, + 0.0031716092489659786, + 0.002156091621145606, + 0.0024474661331623793, + 0.0025785763282328844, + 0.0028236976359039545, + 0.0026227817870676517, + 0.01567927561700344, + 0.008488355204463005, + 0.010646885260939598, + 0.007686267141252756, + 0.007574094459414482, + 0.011961517855525017, + 0.007469400763511658, + 0.0074431211687624454, + 0.009676938876509666, + 0.005432461854070425, + 0.00819349754601717, + 0.004843109752982855, + 0.00838443823158741, + 0.011725258082151413, + 0.0066048987209796906, + 0.010206502862274647, + 0.014888737350702286, + 0.007595823612064123, + 0.011445745825767517, + 0.0074309902265667915, + 0.008100672625005245, + 0.018038559705018997, + 0.0048150718212127686, + 0.011731098406016827, + 0.007823290303349495, + 0.0033968952484428883, + 0.008194436319172382, + 0.004370515700429678, + 0.003236631629988551, + 0.008407744579017162, + 0.004463973920792341, + 0.008801005780696869, + 0.02979922480881214, + 0.004911111202090979, + 0.12791955471038818, + 0.010031506419181824, + 0.0029633024241775274, + 0.01338673010468483, + 0.00563765037804842, + 0.11666911095380783, + 0.002670557936653495, + 0.00198400323279202, + 0.003515028627589345, + 0.0025718810502439737, + 0.0016454114811494946, + 0.005357707384973764, + 0.002395961433649063, + 0.002043511252850294, + 0.003406247589737177, + 0.00153304822742939, + 0.002704849001020193, + 0.0025702286511659622, + 0.0017734236316755414, + 0.0030848549213260412, + 0.0036499055568128824, + 0.0027526437770575285, + 0.040971849113702774, + 0.017535818740725517, + 0.029848676174879074, + 0.02354523167014122, + 0.01823640987277031, + 0.03126757591962814, + 0.01797807216644287, + 0.032539013773202896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.0059465630911290646 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Immersive, frontal underwater filming captures the scene as the camera pans backward at a slow, steady pace, maintaining a constant distance from the subject. The setting is the deep, silent depths of the ocean, a rich palette of deep sapphires and cobalt blues. From afar, above the shimmering, rippled surface, bright, sharply defined golden sunlight penetrates the abyss, creating ethereal pillars of light that illuminate countless tiny suspended particles, making light itself appear tangible. A toned, athletic female figure is perfectly centered in the frame, gracefully swimming directly toward the viewer. She wears a simple, dark navy one-piece swimsuit that contrasts sharply with her skin, which is intermittently speckled by passing sunlight. Her long, dark brown hair flows freely behind her, undulating with the current like dark ribbons. She moves with calm purpose, executing a slow, rhythmic breaststroke; her arms deliberately reach forward, then sweep back in a powerful, fluid motion, effortlessly propelling her through the water. With each gentle exhalation, she releases a small cloud of silver bubbles that absorb the light and shimmer as they rise. Her facial expression, calm and focused, her eyes fixed forward, creates an atmosphere of stunning, almost spiritual beauty, evoking a profound sense of tranquility and freedom in the vast, silent underwater world.", + "seed": 18479, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_3/index_6737_seed_18479_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.037242372240871194, + "vlm_logprob": -0.037242372240871194, + "vlm_yes_no_margin": 4.149999976158142, + "vlm_token_credit": -0.037242372240871194, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00365672679618001 + ], + "opposite_logprob_tokens": [ + -6.25365686416626 + ], + "logprob_reward": -0.00365672679618001, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.00365672679618001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42821574211120605, + 0.09199031442403793, + 0.08627510070800781, + 0.08840755373239517, + 0.12244965136051178, + 0.02514111064374447, + 0.02194293960928917, + 0.02573588117957115, + 0.08054856210947037, + 0.02929317019879818 + ], + "credit_map": [ + 0.026615384966135025, + 0.005842769518494606, + 0.17236925661563873, + 0.01189876813441515, + 0.008533092215657234, + 0.01903497613966465, + 0.006515320856124163, + 0.1774061769247055, + 0.008732317015528679, + 0.003035863395780325, + 0.03120799921452999, + 0.002811349229887128, + 0.0028459695167839527, + 0.011678485199809074, + 0.002345229499042034, + 0.029333099722862244, + 0.005753923207521439, + 0.001500026904977858, + 0.03194630518555641, + 0.0030371223110705614, + 0.0024635689333081245, + 0.006844759918749332, + 0.0018587721278890967, + 0.03287062048912048, + 0.006944295018911362, + 0.0018044967437162995, + 0.030918721109628677, + 0.003525784704834223, + 0.0023484486155211926, + 0.008573349565267563, + 0.0021810189355164766, + 0.032111432403326035, + 0.00936880148947239, + 0.0034746855963021517, + 0.04275814816355705, + 0.005272971000522375, + 0.002885672729462385, + 0.010885518044233322, + 0.0038289320655167103, + 0.043974921107292175, + 0.002365385415032506, + 0.0034451885148882866, + 0.003074090229347348, + 0.003350955666974187, + 0.0030236104503273964, + 0.003973494749516249, + 0.003009708831086755, + 0.0028986758552491665, + 0.006000186782330275, + 0.002269099233672023, + 0.004289258271455765, + 0.0018352088518440723, + 0.0018729991279542446, + 0.0018841420533135533, + 0.0020312683191150427, + 0.0017607762711122632, + 0.0046782707795500755, + 0.002207823796197772, + 0.003979921340942383, + 0.0027557574212551117, + 0.0028034362476319075, + 0.0038005479145795107, + 0.003002065932378173, + 0.0025080586783587933, + 0.01480433065444231, + 0.0047888425178825855, + 0.013185080140829086, + 0.0059193638153374195, + 0.006777801085263491, + 0.02261936664581299, + 0.006487603764981031, + 0.00596617441624403, + 0.00413847528398037, + 0.0019164825789630413, + 0.003912103362381458, + 0.0034610142465680838, + 0.0020447634160518646, + 0.006244474556297064, + 0.004225329961627722, + 0.003350526327267289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.00365672679618001 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196357689797878 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.014196357689797878, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196357689797878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1431828737258911, + 0.09312034398317337, + 0.10880128294229507, + 0.06028571352362633, + 0.1619095355272293, + 0.035658836364746094, + 0.08350913226604462, + 0.07718601077795029, + 0.16411279141902924, + 0.07223357260227203 + ], + "credit_map": [ + 0.01942886784672737, + 0.01312115229666233, + 0.029200736433267593, + 0.011386498808860779, + 0.01195567287504673, + 0.017870724201202393, + 0.010542609728872776, + 0.02967660315334797, + 0.014078867621719837, + 0.006059343460947275, + 0.015246988274157047, + 0.0048073832876980305, + 0.006299430038779974, + 0.027911020442843437, + 0.0063795046880841255, + 0.01233779639005661, + 0.008060027845203876, + 0.005794872529804707, + 0.031403589993715286, + 0.005304105579853058, + 0.005290542729198933, + 0.015627577900886536, + 0.004953210242092609, + 0.032367341220378876, + 0.005849991925060749, + 0.00417740922421217, + 0.01177368126809597, + 0.0075372084975242615, + 0.0036292695440351963, + 0.0097251757979393, + 0.004910584539175034, + 0.012682386673986912, + 0.012238097377121449, + 0.005532112438231707, + 0.05171286687254906, + 0.0059516411274671555, + 0.003977949731051922, + 0.02029278129339218, + 0.005954142194241285, + 0.05624992772936821, + 0.0031959149055182934, + 0.0038853392470628023, + 0.004542722832411528, + 0.003979082219302654, + 0.004500820767134428, + 0.006903266999870539, + 0.004801821894943714, + 0.003849861677736044, + 0.013715310022234917, + 0.008171044290065765, + 0.0125714847818017, + 0.008563980460166931, + 0.006646729074418545, + 0.010501348413527012, + 0.00819657277315855, + 0.015142656862735748, + 0.015138311311602592, + 0.007970539852976799, + 0.011028902605175972, + 0.008416056632995605, + 0.006429193541407585, + 0.012638645246624947, + 0.010018440894782543, + 0.005545914173126221, + 0.03155240789055824, + 0.008128332905471325, + 0.02850315347313881, + 0.009090178646147251, + 0.0114377960562706, + 0.05201477184891701, + 0.011278039775788784, + 0.012108087539672852, + 0.007975757122039795, + 0.009835146367549896, + 0.00971560925245285, + 0.008574414066970348, + 0.005749816540628672, + 0.012257940135896206, + 0.009282286278903484, + 0.0088425911962986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.014196357689797878 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033703286200761795 + ], + "opposite_logprob_tokens": [ + -3.408703327178955 + ], + "logprob_reward": -0.033703286200761795, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033703286200761795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5264015197753906, + 0.03003653511404991, + 0.06682384014129639, + 0.05528772622346878, + 0.03928852826356888, + 0.009970540180802345, + 0.019549032673239708, + 0.07263413071632385, + 0.1446465402841568, + 0.035361483693122864 + ], + "credit_map": [ + 0.05043147876858711, + 0.02096966654062271, + 0.18296021223068237, + 0.026196779683232307, + 0.01838815212249756, + 0.031236514449119568, + 0.017796745523810387, + 0.1784220188856125, + 0.0047738817520439625, + 0.002169432118535042, + 0.004872349556535482, + 0.0015162379713729024, + 0.002008943585678935, + 0.00895899161696434, + 0.0019080562051385641, + 0.0038286459166556597, + 0.004862786270678043, + 0.0014860979281365871, + 0.024593535810709, + 0.0021709841676056385, + 0.0013254898367449641, + 0.004698645789176226, + 0.001872456050477922, + 0.025813855230808258, + 0.004013029858469963, + 0.0012370561016723514, + 0.020032517611980438, + 0.0016937542241066694, + 0.001076922402717173, + 0.0047415271401405334, + 0.0010740720899775624, + 0.021418852731585503, + 0.0033772671595215797, + 0.0024821278639137745, + 0.011291157454252243, + 0.0021270676515996456, + 0.0027489729691296816, + 0.0040328288450837135, + 0.0022629841696470976, + 0.010966125875711441, + 0.0012806456070393324, + 0.0010707331821322441, + 0.0013576758792623878, + 0.0009320540120825171, + 0.0008089301409199834, + 0.002398049458861351, + 0.0011367413681000471, + 0.0009857108816504478, + 0.0042607844807207584, + 0.0027633255813270807, + 0.003261715639382601, + 0.0014916417421773076, + 0.002008603885769844, + 0.0026317904703319073, + 0.001545277307741344, + 0.0015858953120186925, + 0.01695961132645607, + 0.007706716191023588, + 0.012072067707777023, + 0.008060315623879433, + 0.00519771920517087, + 0.008581380359828472, + 0.009222795255482197, + 0.00483353529125452, + 0.026658207178115845, + 0.008071145042777061, + 0.02462947927415371, + 0.009989428333938122, + 0.011316986754536629, + 0.0410258024930954, + 0.012370316311717033, + 0.010585185140371323, + 0.007869669236242771, + 0.0019937369506806135, + 0.005816674791276455, + 0.0018120575696229935, + 0.0016699773259460926, + 0.010727660730481148, + 0.002265318762511015, + 0.0032063915859907866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.033703286200761795 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703020870685577 + ], + "opposite_logprob_tokens": [ + -2.12703013420105 + ], + "logprob_reward": -0.12703020870685577, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12703020870685577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18567553162574768, + 0.13267745077610016, + 0.07910532504320145, + 0.19190552830696106, + 0.2268853485584259, + 0.030170690268278122, + 0.02297261543571949, + 0.031043076887726784, + 0.04553161561489105, + 0.054032765328884125 + ], + "credit_map": [ + 0.028731901198625565, + 0.010297113098204136, + 0.04208841919898987, + 0.01641075871884823, + 0.010759126394987106, + 0.02058097906410694, + 0.01154588907957077, + 0.04526136443018913, + 0.01082517672330141, + 0.0038618468679487705, + 0.048643600195646286, + 0.004454826470464468, + 0.003223760286346078, + 0.010761508718132973, + 0.0032008104026317596, + 0.04770593345165253, + 0.00661134161055088, + 0.002750595100224018, + 0.024295702576637268, + 0.0036449667531996965, + 0.0042664711363613605, + 0.009918419644236565, + 0.0025598553474992514, + 0.025057978928089142, + 0.01609743945300579, + 0.006563989911228418, + 0.060364168137311935, + 0.009725160896778107, + 0.00496698310598731, + 0.02433178573846817, + 0.0071599544025957584, + 0.06269605457782745, + 0.017212076112627983, + 0.007913279347121716, + 0.07855614274740219, + 0.010759002529084682, + 0.007500529754906893, + 0.017253516241908073, + 0.00877417717128992, + 0.07891666144132614, + 0.0041983406990766525, + 0.004629100207239389, + 0.0038646243046969175, + 0.003946240060031414, + 0.0022131067235022783, + 0.005751220975071192, + 0.0027204991783946753, + 0.002847562776878476, + 0.00330449640750885, + 0.002626402536407113, + 0.003327962476760149, + 0.002490827115252614, + 0.002103405073285103, + 0.002472181571647525, + 0.0031016957946121693, + 0.0035456453915685415, + 0.007527728099375963, + 0.003201780840754509, + 0.005356050096452236, + 0.003829998429864645, + 0.002830364042893052, + 0.003199345897883177, + 0.0029635904356837273, + 0.002134220441803336, + 0.008009126409888268, + 0.0051412805914878845, + 0.006773847620934248, + 0.005662168376147747, + 0.0033567221835255623, + 0.00674508698284626, + 0.005688695702701807, + 0.00415468867868185, + 0.010720356367528439, + 0.005920677445828915, + 0.007460537366569042, + 0.004076699260622263, + 0.003794432384893298, + 0.012806059792637825, + 0.004648866597563028, + 0.004605142865329981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.12703020870685577 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625281810760498 + ], + "opposite_logprob_tokens": [ + -4.882625102996826 + ], + "logprob_reward": -0.007625281810760498, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625281810760498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6807814836502075, + 0.09045998752117157, + 0.057085007429122925, + 0.09039155393838882, + 0.03369844704866409, + 0.006742230150848627, + 0.005896823480725288, + 0.008282103575766087, + 0.014398772269487381, + 0.012263618409633636 + ], + "credit_map": [ + 0.04869234561920166, + 0.014491950161755085, + 0.2682287096977234, + 0.020685262978076935, + 0.017141306772828102, + 0.028544381260871887, + 0.012812783941626549, + 0.270184725522995, + 0.007273540366441011, + 0.002624466782435775, + 0.03331344574689865, + 0.0036149462684988976, + 0.0022598339710384607, + 0.005840643309056759, + 0.0033028677571564913, + 0.03223024681210518, + 0.0049401153810322285, + 0.00258362153545022, + 0.018828734755516052, + 0.0033305673860013485, + 0.0019054448930546641, + 0.004441455006599426, + 0.0022102207876741886, + 0.0188448466360569, + 0.015361922793090343, + 0.005321511533111334, + 0.015489811077713966, + 0.004315450321882963, + 0.00519703421741724, + 0.024870218709111214, + 0.004723717458546162, + 0.015111889690160751, + 0.0024921796284615993, + 0.0014615237014368176, + 0.009074353612959385, + 0.0022827768698334694, + 0.0017911617178469896, + 0.005298494361341, + 0.0022856700234115124, + 0.00901228841394186, + 0.0008957095560617745, + 0.0008092549978755414, + 0.0009077280992642045, + 0.0007841594633646309, + 0.000717445625923574, + 0.0012444929452612996, + 0.0006635015015490353, + 0.0007199382525868714, + 0.001608705148100853, + 0.00046494469279423356, + 0.0011596039403229952, + 0.000606838904786855, + 0.00044084840919822454, + 0.0005804542452096939, + 0.0005122609436511993, + 0.0005231670802459121, + 0.0014329177793115377, + 0.0009308019652962685, + 0.0011316428426653147, + 0.0013601729879155755, + 0.0008209461229853332, + 0.0009083783952519298, + 0.0010133568430319428, + 0.0006838866393081844, + 0.0024329235311597586, + 0.000929707835894078, + 0.00234580272808671, + 0.001128397649154067, + 0.00150627875700593, + 0.0036743427626788616, + 0.001373925944790244, + 0.0010073930025100708, + 0.001468698843382299, + 0.001197664998471737, + 0.001717026112601161, + 0.0013282445725053549, + 0.0008807213744148612, + 0.002896933350712061, + 0.0010087400441989303, + 0.001765589346177876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.007625281810760498 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Immersive, frontal underwater filming captures the scene as the camera pans backward at a slow, steady pace, maintaining a constant distance from the subject. The setting is the deep, silent depths of the ocean, a rich palette of deep sapphires and cobalt blues. From afar, above the shimmering, rippled surface, bright, sharply defined golden sunlight penetrates the abyss, creating ethereal pillars of light that illuminate countless tiny suspended particles, making light itself appear tangible. A toned, athletic female figure is perfectly centered in the frame, gracefully swimming directly toward the viewer. She wears a simple, dark navy one-piece swimsuit that contrasts sharply with her skin, which is intermittently speckled by passing sunlight. Her long, dark brown hair flows freely behind her, undulating with the current like dark ribbons. She moves with calm purpose, executing a slow, rhythmic breaststroke; her arms deliberately reach forward, then sweep back in a powerful, fluid motion, effortlessly propelling her through the water. With each gentle exhalation, she releases a small cloud of silver bubbles that absorb the light and shimmer as they rise. Her facial expression, calm and focused, her eyes fixed forward, creates an atmosphere of stunning, almost spiritual beauty, evoking a profound sense of tranquility and freedom in the vast, silent underwater world.", + "seed": 18480, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_3/index_6737_seed_18480_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04481030888855457, + "vlm_logprob": -0.04481030888855457, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.04481030888855457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0035355696454644203 + ], + "opposite_logprob_tokens": [ + -6.628535747528076 + ], + "logprob_reward": -0.0035355696454644203, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0035355696454644203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14812205731868744, + 0.33538854122161865, + 0.034389350563287735, + 0.03590478003025055, + 0.1684473156929016, + 0.0431985929608345, + 0.060812823474407196, + 0.0458059124648571, + 0.024219337850809097, + 0.1037113144993782 + ], + "credit_map": [ + 0.006633090320974588, + 0.0035832205321639776, + 0.06180647388100624, + 0.012988558039069176, + 0.0032107841689139605, + 0.037050675600767136, + 0.002910126931965351, + 0.0199391171336174, + 0.030948208644986153, + 0.006998129189014435, + 0.14417581260204315, + 0.015433240681886673, + 0.0046116653829813, + 0.016576826572418213, + 0.006553149316459894, + 0.11009145528078079, + 0.002609096933156252, + 0.002124731196090579, + 0.010485413484275341, + 0.0014873096952214837, + 0.0015773562481626868, + 0.003683092538267374, + 0.0013595069758594036, + 0.011062839068472385, + 0.004170040600001812, + 0.0032716267742216587, + 0.007870916277170181, + 0.0025061422493308783, + 0.0022193642798811197, + 0.006306246854364872, + 0.0021437224932014942, + 0.007416713051497936, + 0.021157996729016304, + 0.012501123361289501, + 0.036895591765642166, + 0.012754612602293491, + 0.01008953619748354, + 0.023685447871685028, + 0.010248029604554176, + 0.04111495614051819, + 0.0033205775544047356, + 0.003417161526158452, + 0.0063354698941111565, + 0.005096924491226673, + 0.003590333741158247, + 0.008077829144895077, + 0.005198090802878141, + 0.00816220324486494, + 0.009734831750392914, + 0.007571636699140072, + 0.00961112231016159, + 0.005726197734475136, + 0.00950859859585762, + 0.008395613171160221, + 0.005829238798469305, + 0.004435575567185879, + 0.010140749625861645, + 0.004121210891753435, + 0.008830097503960133, + 0.005371200852096081, + 0.004814612213522196, + 0.006291564088314772, + 0.0032281442545354366, + 0.003008326981216669, + 0.0027818698436021805, + 0.0036002665292471647, + 0.0025734773371368647, + 0.004061513114720583, + 0.001806104788556695, + 0.0038911791052669287, + 0.0021997911389917135, + 0.0033051345963031054, + 0.0073064174503088, + 0.013035919517278671, + 0.007584862411022186, + 0.02898278832435608, + 0.005237602163106203, + 0.009578446857631207, + 0.01763271726667881, + 0.014352544210851192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.0035355696454644203 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009788372553884983 + ], + "opposite_logprob_tokens": [ + -4.634788513183594 + ], + "logprob_reward": -0.009788372553884983, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009788372553884983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.077988401055336, + 0.28949832916259766, + 0.09151560813188553, + 0.08602409064769745, + 0.15131746232509613, + 0.08970843255519867, + 0.10451530665159225, + 0.041952021420001984, + 0.027126410976052284, + 0.040353961288928986 + ], + "credit_map": [ + 0.005296264309436083, + 0.006093621253967285, + 0.020872773602604866, + 0.00664111040532589, + 0.0077554816380143166, + 0.015018177218735218, + 0.00787344854325056, + 0.008437524549663067, + 0.02634614333510399, + 0.009511544369161129, + 0.1134277954697609, + 0.01583755388855934, + 0.010171148926019669, + 0.013997776433825493, + 0.010494659654796124, + 0.08971169590950012, + 0.007421153597533703, + 0.008696889504790306, + 0.025276880711317062, + 0.00593737605959177, + 0.005521803162992001, + 0.007220616098493338, + 0.006917414255440235, + 0.024523476138710976, + 0.00983083713799715, + 0.009032695554196835, + 0.015409543178975582, + 0.00733506353572011, + 0.006959677208214998, + 0.014101751148700714, + 0.0073056332767009735, + 0.016048889607191086, + 0.01978311687707901, + 0.011473028920590878, + 0.034602753818035126, + 0.010342312976717949, + 0.008396915160119534, + 0.019153865054249763, + 0.010622166097164154, + 0.03694329783320427, + 0.006847222335636616, + 0.008685038425028324, + 0.012058498337864876, + 0.011894246563315392, + 0.006941719446331263, + 0.019254783168435097, + 0.009188159368932247, + 0.01483876071870327, + 0.011686224490404129, + 0.01370061095803976, + 0.012555662542581558, + 0.013363144360482693, + 0.016392800956964493, + 0.011803878471255302, + 0.014895657077431679, + 0.010117330588400364, + 0.004722571466118097, + 0.006481490097939968, + 0.003587269689887762, + 0.004813650157302618, + 0.010586543008685112, + 0.003986541647464037, + 0.004616463091224432, + 0.003157492494210601, + 0.0027494749519973993, + 0.004187611397355795, + 0.0025958113837987185, + 0.004468252882361412, + 0.00285131111741066, + 0.00398813933134079, + 0.0025784061290323734, + 0.003707403317093849, + 0.004322327673435211, + 0.0046548605896532536, + 0.007200619671493769, + 0.008335514925420284, + 0.0025209328159689903, + 0.0034426955971866846, + 0.004126305691897869, + 0.005750706885010004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 6 + ], + "selected_reward": -0.009788372553884983 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029810702428221703 + ], + "opposite_logprob_tokens": [ + -3.529810667037964 + ], + "logprob_reward": -0.029810702428221703, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029810702428221703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1189219206571579, + 0.4911210238933563, + 0.01732969656586647, + 0.033673688769340515, + 0.20878958702087402, + 0.019350500777363777, + 0.021041372790932655, + 0.03396914526820183, + 0.023853972554206848, + 0.0319490022957325 + ], + "credit_map": [ + 0.005146924406290054, + 0.0025132233276963234, + 0.050412386655807495, + 0.011060389690101147, + 0.0030794511549174786, + 0.025203421711921692, + 0.002202241448685527, + 0.01930389367043972, + 0.04234432056546211, + 0.006549258250743151, + 0.21571972966194153, + 0.022687122225761414, + 0.005458739586174488, + 0.023262709379196167, + 0.008725515566766262, + 0.16637365520000458, + 0.0023861334193497896, + 0.001857544993981719, + 0.002107030712068081, + 0.0015909324865788221, + 0.0034032128751277924, + 0.0022164364345371723, + 0.0015104784397408366, + 0.0022579291835427284, + 0.003263193415477872, + 0.0029275882989168167, + 0.007500587496906519, + 0.0024697480257600546, + 0.0019192221807315946, + 0.005164225120097399, + 0.002502031158655882, + 0.007927101105451584, + 0.025792773813009262, + 0.015599796548485756, + 0.04603532329201698, + 0.01563224196434021, + 0.012735288590192795, + 0.029023775830864906, + 0.01264518965035677, + 0.05132522061467171, + 0.0013481280766427517, + 0.0018091755919158459, + 0.003248622640967369, + 0.002531616948544979, + 0.002042345004156232, + 0.0023756581358611584, + 0.002647185465320945, + 0.0033477714750915766, + 0.002060833852738142, + 0.003761903615668416, + 0.002470131730660796, + 0.002506990684196353, + 0.002940835664048791, + 0.0026708412915468216, + 0.002576680388301611, + 0.0020531590562313795, + 0.005129218101501465, + 0.004841328598558903, + 0.003453974612057209, + 0.0031583185773342848, + 0.008942648768424988, + 0.0036468019243329763, + 0.0026739933528006077, + 0.0021228657569736242, + 0.002209786791354418, + 0.0030561424791812897, + 0.002242429880425334, + 0.00521712563931942, + 0.002924610860645771, + 0.0023959309328347445, + 0.0035705766640603542, + 0.0022373702377080917, + 0.0031003481708467007, + 0.0034317204263061285, + 0.0033358989749103785, + 0.007624621503055096, + 0.001896551693789661, + 0.0032340760808438063, + 0.0032500773668289185, + 0.006075714249163866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.029810702428221703 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033844649791718 + ], + "opposite_logprob_tokens": [ + -1.9103384017944336 + ], + "logprob_reward": -0.16033844649791718, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033844649791718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13503219187259674, + 0.5023613572120667, + 0.011878467164933681, + 0.017607443034648895, + 0.05029195547103882, + 0.04304694011807442, + 0.030916312709450722, + 0.03636745736002922, + 0.01449944544583559, + 0.15799839794635773 + ], + "credit_map": [ + 0.0068548088893294334, + 0.003344758879393339, + 0.055675208568573, + 0.012941057793796062, + 0.0038328685332089663, + 0.0235118567943573, + 0.003814837196841836, + 0.02505679428577423, + 0.04241221025586128, + 0.0058396547101438046, + 0.22022847831249237, + 0.023202890530228615, + 0.005930569022893906, + 0.023658398538827896, + 0.009549248963594437, + 0.17153991758823395, + 0.0017597241094335914, + 0.001577134127728641, + 0.0019242329290136695, + 0.0008977768593467772, + 0.0011012793984264135, + 0.0019857038278132677, + 0.0007920898497104645, + 0.0018405268201604486, + 0.002099077682942152, + 0.0012784844730049372, + 0.004626757465302944, + 0.0010934413876384497, + 0.0009046882623806596, + 0.0027394010685384274, + 0.0008490695036016405, + 0.004016522783786058, + 0.006572036072611809, + 0.00407694885507226, + 0.011187827214598656, + 0.003498133271932602, + 0.0028727345634251833, + 0.006675117649137974, + 0.003586447099223733, + 0.011822709813714027, + 0.004206472542136908, + 0.00284750759601593, + 0.0071971118450164795, + 0.003941092640161514, + 0.0033620873000472784, + 0.00846568401902914, + 0.003913812339305878, + 0.009113171137869358, + 0.005464599933475256, + 0.0028295516967773438, + 0.006144456099718809, + 0.0026930500753223896, + 0.002982503268867731, + 0.006427414249628782, + 0.00219371123239398, + 0.0021810249891132116, + 0.005368893500417471, + 0.0041586859151721, + 0.005304479505866766, + 0.0045578014105558395, + 0.006239145528525114, + 0.003913120366632938, + 0.004012288525700569, + 0.002813041443005204, + 0.0015315765049308538, + 0.0019013406708836555, + 0.001494344905950129, + 0.0020763264037668705, + 0.0011937866220250726, + 0.0024417173117399216, + 0.001907271333038807, + 0.0019530815770849586, + 0.010632199235260487, + 0.02121547795832157, + 0.012049103155732155, + 0.054230667650699615, + 0.008082496002316475, + 0.009984592907130718, + 0.02137950249016285, + 0.02042436972260475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.16033844649791718 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020578453317284584 + ], + "opposite_logprob_tokens": [ + -3.895578384399414 + ], + "logprob_reward": -0.020578453317284584, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020578453317284584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07265870273113251, + 0.11663266271352768, + 0.0221407413482666, + 0.01937069743871689, + 0.623623788356781, + 0.027507254853844643, + 0.05646439641714096, + 0.015154091641306877, + 0.006642190273851156, + 0.03980544954538345 + ], + "credit_map": [ + 0.00372411054559052, + 0.0023163482546806335, + 0.030071746557950974, + 0.006547980476170778, + 0.0014655237318947911, + 0.014643418602645397, + 0.0016459973994642496, + 0.012243573553860188, + 0.010292382910847664, + 0.002230206271633506, + 0.050555795431137085, + 0.005302701611071825, + 0.0015035336837172508, + 0.0053394692949950695, + 0.002279906067997217, + 0.039128657430410385, + 0.002342411084100604, + 0.0031366634648293257, + 0.004312118049710989, + 0.0014778370968997478, + 0.0018314983462914824, + 0.0032823204528540373, + 0.0009362393757328391, + 0.004821652080863714, + 0.0028932865243405104, + 0.002700389828532934, + 0.003485724562779069, + 0.0010321598965674639, + 0.0015075805131345987, + 0.00376042234711349, + 0.0009366086451336741, + 0.0030545243062078953, + 0.07633049786090851, + 0.04651849344372749, + 0.13903531432151794, + 0.0464678555727005, + 0.03721287474036217, + 0.08393638581037521, + 0.03724599629640579, + 0.1568763554096222, + 0.003964617382735014, + 0.002323715714737773, + 0.004706792067736387, + 0.0017484674463048577, + 0.002407511230558157, + 0.0056737978011369705, + 0.002273229183629155, + 0.004409121815115213, + 0.012416021898388863, + 0.005209982395172119, + 0.013081015087664127, + 0.003386851167306304, + 0.0034329332411289215, + 0.012087645009160042, + 0.0033879668917506933, + 0.0034619804937392473, + 0.0035557979717850685, + 0.0011890723835676908, + 0.0031285651493817568, + 0.0016861081821843982, + 0.0012186586391180754, + 0.002128385240212083, + 0.0012272290186956525, + 0.0010202742414548993, + 0.0008121044375002384, + 0.0011747607495635748, + 0.0007983475807122886, + 0.001002509263344109, + 0.000752807769458741, + 0.0005723885260522366, + 0.0006939254817552865, + 0.0008353458251804113, + 0.002710800850763917, + 0.0035970951430499554, + 0.0029182713478803635, + 0.00797961838543415, + 0.003108440665528178, + 0.005515966564416885, + 0.002885427325963974, + 0.011089825071394444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.020578453317284584 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Immersive, frontal underwater filming captures the scene as the camera pans backward at a slow, steady pace, maintaining a constant distance from the subject. The setting is the deep, silent depths of the ocean, a rich palette of deep sapphires and cobalt blues. From afar, above the shimmering, rippled surface, bright, sharply defined golden sunlight penetrates the abyss, creating ethereal pillars of light that illuminate countless tiny suspended particles, making light itself appear tangible. A toned, athletic female figure is perfectly centered in the frame, gracefully swimming directly toward the viewer. She wears a simple, dark navy one-piece swimsuit that contrasts sharply with her skin, which is intermittently speckled by passing sunlight. Her long, dark brown hair flows freely behind her, undulating with the current like dark ribbons. She moves with calm purpose, executing a slow, rhythmic breaststroke; her arms deliberately reach forward, then sweep back in a powerful, fluid motion, effortlessly propelling her through the water. With each gentle exhalation, she releases a small cloud of silver bubbles that absorb the light and shimmer as they rise. Her facial expression, calm and focused, her eyes fixed forward, creates an atmosphere of stunning, almost spiritual beauty, evoking a profound sense of tranquility and freedom in the vast, silent underwater world.", + "seed": 18481, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_3/index_6737_seed_18481_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04663882572203874, + "vlm_logprob": -0.04663882572203874, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.04663882572203874, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016166433691978455 + ], + "opposite_logprob_tokens": [ + -4.3911662101745605 + ], + "logprob_reward": -0.016166433691978455, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.016166433691978455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06321489810943604, + 0.0380384735763073, + 0.06456007063388824, + 0.019555820152163506, + 0.024469919502735138, + 0.07347095757722855, + 0.1951175332069397, + 0.3955913186073303, + 0.03185080736875534, + 0.09413022547960281 + ], + "credit_map": [ + 0.006539412774145603, + 0.0071543431840837, + 0.00808063242584467, + 0.008860531263053417, + 0.008704107254743576, + 0.008862518705427647, + 0.008137336932122707, + 0.0068760160356760025, + 0.0035158072132617235, + 0.005441643763333559, + 0.0042407033033668995, + 0.0056146541610360146, + 0.005208037327975035, + 0.004164014011621475, + 0.005703656002879143, + 0.004149956628680229, + 0.009310206398367882, + 0.005518485326319933, + 0.007404322735965252, + 0.006185354199260473, + 0.004995109513401985, + 0.014830689877271652, + 0.006374216638505459, + 0.009941679425537586, + 0.00192743050865829, + 0.0016776890261098742, + 0.002187729114666581, + 0.002088014967739582, + 0.0020294152200222015, + 0.004714997950941324, + 0.001868874067440629, + 0.0030616691801697016, + 0.003213446820154786, + 0.001787426183000207, + 0.003755985526368022, + 0.0028981217183172703, + 0.00237973197363317, + 0.004611151292920113, + 0.0022855172865092754, + 0.0035385380033403635, + 0.010001329705119133, + 0.003653895342722535, + 0.009045364335179329, + 0.004781863186508417, + 0.008116944693028927, + 0.02233460359275341, + 0.0035371265839785337, + 0.011999832466244698, + 0.023787401616573334, + 0.014280786737799644, + 0.0287995133548975, + 0.026530729606747627, + 0.020246537402272224, + 0.03949645161628723, + 0.013221892528235912, + 0.028754225000739098, + 0.07089152932167053, + 0.0189101193100214, + 0.04751366376876831, + 0.03064150922000408, + 0.044153422117233276, + 0.09919081628322601, + 0.022885190322995186, + 0.06140506640076637, + 0.005208700429648161, + 0.0020444230176508427, + 0.002827719086781144, + 0.0033606605138629675, + 0.0024480819702148438, + 0.0076357717625796795, + 0.0027243783697485924, + 0.005601073615252972, + 0.008453505113720894, + 0.004598690662533045, + 0.008363689295947552, + 0.004363266285508871, + 0.005235430784523487, + 0.02846350148320198, + 0.005148325115442276, + 0.02950381115078926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.016166433691978455 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004104048945009708 + ], + "opposite_logprob_tokens": [ + -5.504104137420654 + ], + "logprob_reward": -0.004104048945009708, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104048945009708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11616778373718262, + 0.20498205721378326, + 0.07613236457109451, + 0.0818132534623146, + 0.071727454662323, + 0.03810249641537666, + 0.10683517903089523, + 0.17504854500293732, + 0.08853719383478165, + 0.040653668344020844 + ], + "credit_map": [ + 0.01086367852985859, + 0.013839609920978546, + 0.015135741792619228, + 0.015240142121911049, + 0.012434257194399834, + 0.013499203138053417, + 0.021306710317730904, + 0.0138484425842762, + 0.02241070754826069, + 0.02857167460024357, + 0.023935075849294662, + 0.03431286662817001, + 0.020136933773756027, + 0.018219193443655968, + 0.03366559371352196, + 0.023730026558041573, + 0.010159870609641075, + 0.006106805521994829, + 0.00893157534301281, + 0.01014566421508789, + 0.006627981085330248, + 0.016599921509623528, + 0.006705196108669043, + 0.010855347849428654, + 0.01034999918192625, + 0.008357211016118526, + 0.01046364102512598, + 0.009025651961565018, + 0.006438231561332941, + 0.016010135412216187, + 0.00914047472178936, + 0.012027915567159653, + 0.01351138111203909, + 0.00569614814594388, + 0.008090726099908352, + 0.005808057729154825, + 0.010596438311040401, + 0.012419231235980988, + 0.005364229902625084, + 0.010241241194307804, + 0.004264951683580875, + 0.0034978226758539677, + 0.004326946102082729, + 0.005045260768383741, + 0.005318156909197569, + 0.005576897878199816, + 0.004927122034132481, + 0.005145336966961622, + 0.0119453901425004, + 0.008968832902610302, + 0.014665384776890278, + 0.012963063083589077, + 0.011206516064703465, + 0.021953275427222252, + 0.01014520414173603, + 0.014987505972385406, + 0.026578284800052643, + 0.010154513642191887, + 0.02249377779662609, + 0.019069643691182137, + 0.02039938047528267, + 0.03804973512887955, + 0.01497215498238802, + 0.0233310516923666, + 0.009254449047148228, + 0.0066477213986217976, + 0.00962763000279665, + 0.016065599396824837, + 0.006242804694920778, + 0.012999536469578743, + 0.012858228757977486, + 0.01484122034162283, + 0.004474690183997154, + 0.004057562444359064, + 0.004947165492922068, + 0.004718956537544727, + 0.0033248027320951223, + 0.0077058239839971066, + 0.004229108802974224, + 0.007195556536316872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.004104048945009708 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741754245012999 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741754245012999, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741754245012999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07121633738279343, + 0.16030628979206085, + 0.06849582493305206, + 0.06712768971920013, + 0.16724395751953125, + 0.09848178923130035, + 0.040539734065532684, + 0.19584809243679047, + 0.06439041346311569, + 0.06634987145662308 + ], + "credit_map": [ + 0.00827133096754551, + 0.007916152477264404, + 0.007049218751490116, + 0.01022981945425272, + 0.007692030631005764, + 0.008101465180516243, + 0.014266126789152622, + 0.007690197788178921, + 0.014536953531205654, + 0.020467281341552734, + 0.015448703430593014, + 0.02541109174489975, + 0.018366646021604538, + 0.014879151247441769, + 0.0318213552236557, + 0.019375117495656013, + 0.00836382806301117, + 0.007674014195799828, + 0.007395550608634949, + 0.009020273573696613, + 0.005700405221432447, + 0.013209671713411808, + 0.0077920593321323395, + 0.009340016171336174, + 0.005943040829151869, + 0.007009096909314394, + 0.007432424463331699, + 0.01246077660471201, + 0.005408530589193106, + 0.009503427892923355, + 0.010731926187872887, + 0.008638458326458931, + 0.025384968146681786, + 0.021602768450975418, + 0.014868532307446003, + 0.02469639480113983, + 0.015109683386981487, + 0.02553439326584339, + 0.017506616190075874, + 0.022540608420968056, + 0.01182783953845501, + 0.01074374932795763, + 0.010316683910787106, + 0.012680421583354473, + 0.01245806273072958, + 0.019544580951333046, + 0.007720681838691235, + 0.013189772143959999, + 0.005497074220329523, + 0.003112850012257695, + 0.005496242083609104, + 0.0051154750399291515, + 0.0049407328478991985, + 0.008009751327335835, + 0.0024876121897250414, + 0.005879993084818125, + 0.03589637950062752, + 0.00936341192573309, + 0.02145257033407688, + 0.0140433544293046, + 0.02107604406774044, + 0.051094550639390945, + 0.011261656880378723, + 0.031660132110118866, + 0.010265739634633064, + 0.004747577011585236, + 0.006886275485157967, + 0.00532596604898572, + 0.005165504291653633, + 0.01567937433719635, + 0.006053123157471418, + 0.010266857221722603, + 0.005278274416923523, + 0.003932634834200144, + 0.005429815035313368, + 0.006412077229470015, + 0.006011704448610544, + 0.01710335537791252, + 0.005673795007169247, + 0.016508210450410843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.006741754245012999 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20154528319835663 + ], + "opposite_logprob_tokens": [ + -1.701545238494873 + ], + "logprob_reward": -0.20154528319835663, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20154528319835663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1605956256389618, + 0.16947847604751587, + 0.10821593552827835, + 0.10118772089481354, + 0.10051044076681137, + 0.07404496520757675, + 0.06569301337003708, + 0.08735331147909164, + 0.06426961719989777, + 0.06865086406469345 + ], + "credit_map": [ + 0.01194838061928749, + 0.017920205369591713, + 0.01918107457458973, + 0.020002372562885284, + 0.015031357295811176, + 0.017734285444021225, + 0.037021223455667496, + 0.021756736561655998, + 0.019305724650621414, + 0.019658299162983894, + 0.014732994139194489, + 0.02482319064438343, + 0.02075120061635971, + 0.016539478674530983, + 0.03269527852535248, + 0.020972302183508873, + 0.012719628401100636, + 0.00868020486086607, + 0.01181007269769907, + 0.010355870239436626, + 0.014917295426130295, + 0.021267404779791832, + 0.014197079464793205, + 0.014268376864492893, + 0.01403319463133812, + 0.007879799231886864, + 0.016025332733988762, + 0.010662494227290154, + 0.008659198880195618, + 0.018913481384515762, + 0.009369401261210442, + 0.015644822269678116, + 0.015878356993198395, + 0.009025403298437595, + 0.01002579741179943, + 0.007760239765048027, + 0.014460409060120583, + 0.020440440624952316, + 0.007474408019334078, + 0.015445388853549957, + 0.01052959356456995, + 0.004119541496038437, + 0.009021940641105175, + 0.005778565537184477, + 0.009658649563789368, + 0.019376231357455254, + 0.0036635517608374357, + 0.011896890588104725, + 0.009400473907589912, + 0.005292964167892933, + 0.009420790709555149, + 0.007936174049973488, + 0.005327496211975813, + 0.014164472930133343, + 0.003387208329513669, + 0.010763432830572128, + 0.016180578619241714, + 0.00511857308447361, + 0.01007892470806837, + 0.006544637028127909, + 0.008408129215240479, + 0.022581417113542557, + 0.005225175525993109, + 0.013215867802500725, + 0.009011965245008469, + 0.004621147643774748, + 0.0076720230281353, + 0.00726942578330636, + 0.005568011198192835, + 0.014083337970077991, + 0.007717728149145842, + 0.008325970731675625, + 0.0070504602044820786, + 0.006565222050994635, + 0.00655848253518343, + 0.012590263038873672, + 0.0047310721129179, + 0.011874641291797161, + 0.0073166340589523315, + 0.011964083649218082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.20154528319835663 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in the deep ocean with a blue color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a toned, athletic female figure swimming directly toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the swimmer wear a dark navy one-piece swimsuit and have long dark brown hair flowing behind her?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera panning backward at a slow, steady pace while maintaining a constant distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible pillars of golden sunlight penetrating the water and illuminating suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636608529835939 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636608529835939, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636608529835939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04394875839352608, + 0.10562995821237564, + 0.1061977744102478, + 0.022178180515766144, + 0.08064084500074387, + 0.24660006165504456, + 0.012136025354266167, + 0.3272698223590851, + 0.017101019620895386, + 0.038297563791275024 + ], + "credit_map": [ + 0.003728694748133421, + 0.004900727886706591, + 0.005765923764556646, + 0.005494719371199608, + 0.004428844433277845, + 0.004744518548250198, + 0.007876614108681679, + 0.007008716929703951, + 0.010926987044513226, + 0.012984039261937141, + 0.01002570427954197, + 0.01682615466415882, + 0.011658789590001106, + 0.010620479471981525, + 0.019994961097836494, + 0.012592854909598827, + 0.012456326745450497, + 0.0072286780923604965, + 0.013303233310580254, + 0.009774954058229923, + 0.010710076428949833, + 0.02496352046728134, + 0.01166610512882471, + 0.0160948745906353, + 0.003372682025656104, + 0.00139724463224411, + 0.0033530534710735083, + 0.0021449474152177572, + 0.0017902138642966747, + 0.004367270972579718, + 0.002357650315389037, + 0.00339511688798666, + 0.012698384933173656, + 0.005102729890495539, + 0.007577815558761358, + 0.0055551775731146336, + 0.011415722779929638, + 0.0203825905919075, + 0.00489220954477787, + 0.013016211800277233, + 0.036432817578315735, + 0.011575376614928246, + 0.029177017509937286, + 0.016613425686955452, + 0.029864320531487465, + 0.07415949553251266, + 0.009246746078133583, + 0.03953085467219353, + 0.001351086306385696, + 0.001249986351467669, + 0.0018171927658841014, + 0.0022416121792048216, + 0.001080203684978187, + 0.00195823865942657, + 0.0007874313159845769, + 0.0016502737998962402, + 0.059681713581085205, + 0.017940092831850052, + 0.03217344731092453, + 0.024530241265892982, + 0.032673660665750504, + 0.08762090653181076, + 0.020342111587524414, + 0.052307650446891785, + 0.002481121337041259, + 0.0012289519654586911, + 0.001686871750280261, + 0.001921983202919364, + 0.0014753135619685054, + 0.00376823334954679, + 0.0015887327026575804, + 0.0029498112853616476, + 0.0037385381292551756, + 0.0021295908372849226, + 0.003345224540680647, + 0.0023882626555860043, + 0.002790129743516445, + 0.0105552664026618, + 0.0026472851168364286, + 0.010703264735639095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.004636608529835939 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_4/index_tensor([5388, 5388])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_4/index_tensor([5388, 5388])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..17768ee4db64766711387f80e781cb02b79da8f6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_4/index_tensor([5388, 5388])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot begins, following a dusty crimson off-road jeep as it slowly but steadily rolls across the vast, jagged crest of the ancient Great Wall. The massive off-road tires rumble over the weathered gray masonry, their bodywork reflecting a dull luster in the setting sun. Then, the camera slowly and smoothly zooms out, backward and upward, the jeep gradually shrinking in the frame, transforming from a powerful, concrete figure into a solitary, moving red dot. As the camera continues to retreat, the Great Wall's majestic fullness is revealed: like a giant stone dragon, it winds endlessly along the ridges of rolling green mountains, its towering beacon towers looming in the distance. Soft golden light gilds the entire scene with a warm and tranquil hue, casting long, gentle shadows across the valley, and mist quietly rising between the mountain cols. Ultimately, the entire picture presents a breathtaking sense of solemnity and loneliness that transcends time and space, highlighting the insignificance and tenacity of human creation in the face of the majestic nature.", + "seed": 6124, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_4/index_5388_seed_6124_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01994168305536732, + "vlm_logprob": -0.01994168305536732, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.01994168305536732, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03960952162742615 + ], + "opposite_logprob_tokens": [ + -3.414609432220459 + ], + "logprob_reward": -0.03960952162742615, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03960952162742615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20778946578502655, + 0.09162933379411697, + 0.09781868010759354, + 0.01826826110482216, + 0.04467795789241791, + 0.05880296602845192, + 0.055885571986436844, + 0.06454702466726303, + 0.18094606697559357, + 0.17963458597660065 + ], + "credit_map": [ + 0.01648051291704178, + 0.028360700234770775, + 0.024668222293257713, + 0.026098039001226425, + 0.028496989980340004, + 0.02643447369337082, + 0.01880405843257904, + 0.03844646364450455, + 0.009487750008702278, + 0.00981771107763052, + 0.013859396800398827, + 0.015066983178257942, + 0.009180222637951374, + 0.008955455385148525, + 0.01437203399837017, + 0.01088978536427021, + 0.01773909293115139, + 0.009012192487716675, + 0.010959264822304249, + 0.009740478359162807, + 0.010315141640603542, + 0.019127430394291878, + 0.01016617938876152, + 0.010758906602859497, + 0.0023669942747801542, + 0.00166576006449759, + 0.002307255519554019, + 0.0032304003834724426, + 0.0019154626643285155, + 0.002603496890515089, + 0.0022263270802795887, + 0.001952566672116518, + 0.0032138123642653227, + 0.003075863467529416, + 0.003727481234818697, + 0.004606962203979492, + 0.00371177913621068, + 0.0032545558642596006, + 0.011501464992761612, + 0.01158604584634304, + 0.00808772724121809, + 0.008742301724851131, + 0.010507317259907722, + 0.004230209160596132, + 0.005872556008398533, + 0.009005245752632618, + 0.004980487283319235, + 0.007377125788480043, + 0.005088231060653925, + 0.004886831156909466, + 0.006522823125123978, + 0.006552237085998058, + 0.006557279732078314, + 0.01121919322758913, + 0.00931785348802805, + 0.005741129629313946, + 0.009414438158273697, + 0.008791908621788025, + 0.01077018678188324, + 0.006159808021038771, + 0.008063418790698051, + 0.01223761960864067, + 0.004468259867280722, + 0.004641394596546888, + 0.026079541072249413, + 0.012200613506138325, + 0.024480221793055534, + 0.010653992183506489, + 0.017113054171204567, + 0.0690821185708046, + 0.01284046284854412, + 0.008496078662574291, + 0.03563060611486435, + 0.0061770291067659855, + 0.013828509487211704, + 0.004036686383187771, + 0.025631733238697052, + 0.08313873410224915, + 0.0043305400758981705, + 0.006860767956823111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.03960952162742615 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009809857234358788 + ], + "opposite_logprob_tokens": [ + -4.634809970855713 + ], + "logprob_reward": -0.009809857234358788, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009809857234358788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09338241070508957, + 0.06500633805990219, + 0.0559975765645504, + 0.0519220307469368, + 0.15454719960689545, + 0.1114545613527298, + 0.21003393828868866, + 0.04994581639766693, + 0.0701824426651001, + 0.1375276893377304 + ], + "credit_map": [ + 0.006245837081223726, + 0.007751502096652985, + 0.010151475667953491, + 0.010671691037714481, + 0.007167516276240349, + 0.01753055676817894, + 0.007433618418872356, + 0.02643021196126938, + 0.008153082802891731, + 0.004226565361022949, + 0.009948990307748318, + 0.007193709723651409, + 0.005901015363633633, + 0.0064215948805212975, + 0.011686483398079872, + 0.011474895291030407, + 0.012772022746503353, + 0.003889575367793441, + 0.007924309000372887, + 0.007451794110238552, + 0.004180548246949911, + 0.007677546702325344, + 0.0057018231600522995, + 0.006399956997483969, + 0.005723446607589722, + 0.004870499484241009, + 0.0034373109228909016, + 0.006534113548696041, + 0.0033282735385000706, + 0.01007931586354971, + 0.006623424589633942, + 0.011325645260512829, + 0.016582682728767395, + 0.015050564892590046, + 0.0189497172832489, + 0.012584010139107704, + 0.008001556620001793, + 0.01493535004556179, + 0.03638036921620369, + 0.03206295147538185, + 0.013043615967035294, + 0.015820041298866272, + 0.016856612637639046, + 0.01072219293564558, + 0.016743438318371773, + 0.011486961506307125, + 0.00976687204092741, + 0.017014823853969574, + 0.013833051547408104, + 0.01535874791443348, + 0.025293292477726936, + 0.022843075916171074, + 0.02639642357826233, + 0.04314599186182022, + 0.03751343488693237, + 0.025649910792708397, + 0.00441857660189271, + 0.0067412774078547955, + 0.0072718895971775055, + 0.005963189527392387, + 0.004393397830426693, + 0.007860979996621609, + 0.008397773839533329, + 0.004898727871477604, + 0.008135728538036346, + 0.004482061602175236, + 0.012953621335327625, + 0.004377477802336216, + 0.006812467239797115, + 0.026238176971673965, + 0.003821724094450474, + 0.003361184149980545, + 0.021820148453116417, + 0.00605755066499114, + 0.016323266550898552, + 0.005439488682895899, + 0.017472688108682632, + 0.05835714936256409, + 0.006708687171339989, + 0.005348717328161001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -0.009809857234358788 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305563494563103 + ], + "opposite_logprob_tokens": [ + -3.168055534362793 + ], + "logprob_reward": -0.04305563494563103, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305563494563103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.065557561814785, + 0.10186850279569626, + 0.033880118280649185, + 0.05401630699634552, + 0.07958672940731049, + 0.09632960706949234, + 0.0755273848772049, + 0.06070352718234062, + 0.19877895712852478, + 0.2337513118982315 + ], + "credit_map": [ + 0.0075610182248055935, + 0.005426519550383091, + 0.008821289986371994, + 0.006447964813560247, + 0.005643183831125498, + 0.008673162199556828, + 0.0068548195995390415, + 0.01612960733473301, + 0.006415313575416803, + 0.0040767802856862545, + 0.035637374967336655, + 0.008529320359230042, + 0.004053889773786068, + 0.0066320765763521194, + 0.006523988675326109, + 0.029999755322933197, + 0.003888428211212158, + 0.0034289660397917032, + 0.005503751337528229, + 0.004416739102452993, + 0.0035327039659023285, + 0.004229641519486904, + 0.004918175283819437, + 0.003961714915931225, + 0.006774841342121363, + 0.006307896226644516, + 0.005111368838697672, + 0.006194696296006441, + 0.005944729316979647, + 0.009335577487945557, + 0.005731367506086826, + 0.008615830913186073, + 0.008811203762888908, + 0.007697324268519878, + 0.011386051774024963, + 0.007934645749628544, + 0.004703648388385773, + 0.008578882552683353, + 0.015696188434958458, + 0.014778784476220608, + 0.007969303987920284, + 0.007728223688900471, + 0.011244455352425575, + 0.010298053734004498, + 0.011982765048742294, + 0.012009570375084877, + 0.009702389128506184, + 0.025394847616553307, + 0.006841110996901989, + 0.008993974886834621, + 0.009399116970598698, + 0.007011794950813055, + 0.006514246575534344, + 0.010750389657914639, + 0.013367366045713425, + 0.012649383395910263, + 0.006064496468752623, + 0.0061609819531440735, + 0.007615158334374428, + 0.006237119436264038, + 0.0074347518384456635, + 0.01094688381999731, + 0.00980410072952509, + 0.006440037861466408, + 0.02899840474128723, + 0.008833321742713451, + 0.032514460384845734, + 0.008217875845730305, + 0.02006009966135025, + 0.08393220603466034, + 0.007827174849808216, + 0.008395402692258358, + 0.03824558109045029, + 0.008358185179531574, + 0.030028512701392174, + 0.006088185124099255, + 0.031006352975964546, + 0.10408255457878113, + 0.006831190548837185, + 0.009110752493143082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.04305563494563103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004963834653608501 + ], + "opposite_logprob_tokens": [ + -7.6254963874816895 + ], + "logprob_reward": -0.0004963834653608501, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0004963834653608501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06856158375740051, + 0.12023937702178955, + 0.033520471304655075, + 0.036955542862415314, + 0.12296026945114136, + 0.0716179683804512, + 0.11665353924036026, + 0.03802664205431938, + 0.11693248152732849, + 0.2745320796966553 + ], + "credit_map": [ + 0.006488355342298746, + 0.008378920145332813, + 0.008705233223736286, + 0.007019118871539831, + 0.007733662612736225, + 0.009472548961639404, + 0.0072227707132697105, + 0.01354097668081522, + 0.009388013742864132, + 0.004461720120161772, + 0.03669721260666847, + 0.01031775213778019, + 0.005996624007821083, + 0.008781569078564644, + 0.011159582994878292, + 0.03343689441680908, + 0.0073377350345253944, + 0.002944221720099449, + 0.005272949114441872, + 0.004123234190046787, + 0.0023370368871837854, + 0.0038141466211527586, + 0.0038150683976709843, + 0.0038760770112276077, + 0.004453373607248068, + 0.0037562812212854624, + 0.0029171935748308897, + 0.00423403037711978, + 0.0037081625778228045, + 0.006823374889791012, + 0.00447256350889802, + 0.006590561475604773, + 0.012770125642418861, + 0.010612393729388714, + 0.01680869236588478, + 0.011995966546237469, + 0.005534014664590359, + 0.013583256863057613, + 0.027150806039571762, + 0.02450500801205635, + 0.006185509730130434, + 0.007174540776759386, + 0.00644942419603467, + 0.0062520247884094715, + 0.01235838420689106, + 0.009616933763027191, + 0.00714189512655139, + 0.01643925905227661, + 0.009049573913216591, + 0.010750788263976574, + 0.014267497695982456, + 0.011845792643725872, + 0.01266603171825409, + 0.017444370314478874, + 0.022284140810370445, + 0.018345341086387634, + 0.0029144706204533577, + 0.0055105616338551044, + 0.004781445022672415, + 0.0035735045094043016, + 0.004880258347839117, + 0.006775300018489361, + 0.00498368963599205, + 0.004607414826750755, + 0.014169212430715561, + 0.007287267595529556, + 0.020048968493938446, + 0.007252562791109085, + 0.011266590096056461, + 0.044462475925683975, + 0.006709021516144276, + 0.005736387800425291, + 0.028903095051646233, + 0.011446601711213589, + 0.06436562538146973, + 0.01704091764986515, + 0.02521745301783085, + 0.09271419793367386, + 0.01383755449205637, + 0.021006634458899498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.0004963834653608501 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067370180040597916 + ], + "opposite_logprob_tokens": [ + -5.006737232208252 + ], + "logprob_reward": -0.0067370180040597916, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067370180040597916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08418433368206024, + 0.04996654391288757, + 0.031664587557315826, + 0.020461345091462135, + 0.0460357591509819, + 0.06154170632362366, + 0.08566837012767792, + 0.03936424106359482, + 0.044085003435611725, + 0.5370281934738159 + ], + "credit_map": [ + 0.0068603381514549255, + 0.006409703753888607, + 0.012391196563839912, + 0.008041776716709137, + 0.007540893740952015, + 0.015396798960864544, + 0.005675220396369696, + 0.021868396550416946, + 0.006215513218194246, + 0.0026573920622467995, + 0.008870545774698257, + 0.005310968495905399, + 0.0040769982151687145, + 0.005288862157613039, + 0.00772053049877286, + 0.00982572603970766, + 0.004196830093860626, + 0.0032724898774176836, + 0.004326577298343182, + 0.004266687668859959, + 0.0031380420550704002, + 0.005310352426022291, + 0.003385148011147976, + 0.0037684536073356867, + 0.0021207481622695923, + 0.0020898159127682447, + 0.0029804592486470938, + 0.002902163425460458, + 0.0026880966033786535, + 0.002406740328297019, + 0.002827445510774851, + 0.0024458710104227066, + 0.0071953050792217255, + 0.005489937029778957, + 0.009256046265363693, + 0.004582094959914684, + 0.002941457787528634, + 0.00670604407787323, + 0.00460016680881381, + 0.005264699924737215, + 0.009506230242550373, + 0.008039398118853569, + 0.010245631448924541, + 0.003475479781627655, + 0.0061483969911932945, + 0.011124152690172195, + 0.00397470872849226, + 0.009027699939906597, + 0.0061284564435482025, + 0.007301627192646265, + 0.011279594153165817, + 0.009426726028323174, + 0.00774325430393219, + 0.011686626821756363, + 0.017005516216158867, + 0.015096555463969707, + 0.005657270550727844, + 0.005360995419323444, + 0.004413405898958445, + 0.003864302998408675, + 0.005986446049064398, + 0.00568030122667551, + 0.005033513996750116, + 0.0033680000342428684, + 0.0044852327555418015, + 0.002986077917739749, + 0.01176508329808712, + 0.004029660020023584, + 0.0035967351868748665, + 0.009725382551550865, + 0.004123819060623646, + 0.003373008919879794, + 0.08383332192897797, + 0.01819349266588688, + 0.07846760004758835, + 0.017668412998318672, + 0.06738970428705215, + 0.2330254465341568, + 0.013772903010249138, + 0.024677203968167305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.0067370180040597916 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot begins, following a dusty crimson off-road jeep as it slowly but steadily rolls across the vast, jagged crest of the ancient Great Wall. The massive off-road tires rumble over the weathered gray masonry, their bodywork reflecting a dull luster in the setting sun. Then, the camera slowly and smoothly zooms out, backward and upward, the jeep gradually shrinking in the frame, transforming from a powerful, concrete figure into a solitary, moving red dot. As the camera continues to retreat, the Great Wall's majestic fullness is revealed: like a giant stone dragon, it winds endlessly along the ridges of rolling green mountains, its towering beacon towers looming in the distance. Soft golden light gilds the entire scene with a warm and tranquil hue, casting long, gentle shadows across the valley, and mist quietly rising between the mountain cols. Ultimately, the entire picture presents a breathtaking sense of solemnity and loneliness that transcends time and space, highlighting the insignificance and tenacity of human creation in the face of the majestic nature.", + "seed": 6125, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_4/index_5388_seed_6125_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05111590486485511, + "vlm_logprob": -0.05111590486485511, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.05111590486485511, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05957379564642906 + ], + "opposite_logprob_tokens": [ + -3.0595738887786865 + ], + "logprob_reward": -0.05957379564642906, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05957379564642906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14354750514030457, + 0.20859085023403168, + 0.07398362457752228, + 0.06467515230178833, + 0.1253185123205185, + 0.043229393661022186, + 0.050876762717962265, + 0.07221848517656326, + 0.12408329546451569, + 0.09347639977931976 + ], + "credit_map": [ + 0.01603536866605282, + 0.01638270728290081, + 0.021180281415581703, + 0.029475996270775795, + 0.009994436986744404, + 0.016260823234915733, + 0.015570685267448425, + 0.018647203221917152, + 0.018439069390296936, + 0.01154920645058155, + 0.07836294919252396, + 0.03666040673851967, + 0.00870082713663578, + 0.021285532042384148, + 0.018212009221315384, + 0.01538086123764515, + 0.010571474209427834, + 0.007125454023480415, + 0.006684072781354189, + 0.009790478274226189, + 0.012485185638070107, + 0.0132402703166008, + 0.007322455756366253, + 0.006764237303286791, + 0.013448415324091911, + 0.005824635270982981, + 0.009754408150911331, + 0.005948003381490707, + 0.005898891948163509, + 0.01250074990093708, + 0.0057070292532444, + 0.005593016743659973, + 0.012534228153526783, + 0.010472271591424942, + 0.017105832695961, + 0.025204071775078773, + 0.01801181770861149, + 0.012181803584098816, + 0.015583042986690998, + 0.014225445687770844, + 0.004565048031508923, + 0.004053137265145779, + 0.007836795412003994, + 0.006168272811919451, + 0.00486413948237896, + 0.005204495973885059, + 0.005752016790211201, + 0.00478548975661397, + 0.0057403575628995895, + 0.005146966781467199, + 0.0065111699514091015, + 0.008704802952706814, + 0.005283558741211891, + 0.008995972573757172, + 0.005457103252410889, + 0.005036828573793173, + 0.006318760570138693, + 0.006358758080750704, + 0.010694180615246296, + 0.01365136168897152, + 0.007017557043582201, + 0.008955833502113819, + 0.010735117830336094, + 0.008486917242407799, + 0.01250994298607111, + 0.013074195943772793, + 0.013144426047801971, + 0.01819412223994732, + 0.014715416356921196, + 0.023067517206072807, + 0.01305573619902134, + 0.016321932896971703, + 0.010090161114931107, + 0.006167097017168999, + 0.0169077068567276, + 0.025279397144913673, + 0.00610450329259038, + 0.015364116057753563, + 0.007837439887225628, + 0.005725983064621687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.05957379564642906 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006772421766072512 + ], + "opposite_logprob_tokens": [ + -5.006772518157959 + ], + "logprob_reward": -0.006772421766072512, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006772421766072512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.125209778547287, + 0.0595332495868206, + 0.05556808412075043, + 0.08422630280256271, + 0.04798813536763191, + 0.05080809444189072, + 0.0645986795425415, + 0.0670819953083992, + 0.18309378623962402, + 0.2618919014930725 + ], + "credit_map": [ + 0.014507378451526165, + 0.009234999306499958, + 0.011241803877055645, + 0.026939019560813904, + 0.007302863523364067, + 0.015177526511251926, + 0.0120549900457263, + 0.028751207515597343, + 0.005320544354617596, + 0.0035232973750680685, + 0.017821740359067917, + 0.009070388041436672, + 0.0036432165652513504, + 0.009532718919217587, + 0.005209125578403473, + 0.005412220489233732, + 0.005185007583349943, + 0.0055489796213805676, + 0.009570627473294735, + 0.0070168874226510525, + 0.008960981853306293, + 0.004902775399386883, + 0.007502127438783646, + 0.0068806977942585945, + 0.012064281851053238, + 0.008634456433355808, + 0.010531209409236908, + 0.012280277907848358, + 0.008415930904448032, + 0.009937522001564503, + 0.01133465301245451, + 0.011027979664504528, + 0.007206719368696213, + 0.004277125932276249, + 0.006953174713999033, + 0.011551686562597752, + 0.0038343043997883797, + 0.004591191653162241, + 0.005167671479284763, + 0.004406262189149857, + 0.005385356955230236, + 0.005628166254609823, + 0.009947242215275764, + 0.005443867761641741, + 0.005345870740711689, + 0.006207617931067944, + 0.006484189070761204, + 0.006365784909576178, + 0.006311839446425438, + 0.006572835147380829, + 0.010327314957976341, + 0.008448312059044838, + 0.007683190982788801, + 0.008838462643325329, + 0.011464901268482208, + 0.0049518197774887085, + 0.005613184534013271, + 0.005720399785786867, + 0.0089786471799016, + 0.006783793214708567, + 0.008786991238594055, + 0.008700486272573471, + 0.010931337252259254, + 0.011567160487174988, + 0.018619032576680183, + 0.02585051767528057, + 0.023372463881969452, + 0.026678623631596565, + 0.023940401151776314, + 0.03171014413237572, + 0.015929732471704483, + 0.01699286885559559, + 0.02522111125290394, + 0.016679149121046066, + 0.0665559396147728, + 0.08422215282917023, + 0.010225924663245678, + 0.0187295563519001, + 0.025399306789040565, + 0.014858749695122242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.006772421766072512 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987053096294403 + ], + "opposite_logprob_tokens": [ + -1.8048704862594604 + ], + "logprob_reward": -0.17987053096294403, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987053096294403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15409770607948303, + 0.08770637959241867, + 0.09224826842546463, + 0.13152289390563965, + 0.11053427308797836, + 0.08753374963998795, + 0.0878245159983635, + 0.08848007023334503, + 0.10302449017763138, + 0.05702764540910721 + ], + "credit_map": [ + 0.011477078311145306, + 0.016265874728560448, + 0.018104873597621918, + 0.037883758544921875, + 0.015597663819789886, + 0.016373543068766594, + 0.022059621289372444, + 0.016335299238562584, + 0.006724189966917038, + 0.007935992442071438, + 0.02025848627090454, + 0.01416496280580759, + 0.007869608700275421, + 0.007570122368633747, + 0.012621235102415085, + 0.010561784729361534, + 0.011903076432645321, + 0.015581713058054447, + 0.008547871373593807, + 0.011573229916393757, + 0.014748870395123959, + 0.010788343846797943, + 0.010249627754092216, + 0.00885554775595665, + 0.026042526587843895, + 0.011820425279438496, + 0.013937292620539665, + 0.014589081518352032, + 0.01095636934041977, + 0.029031870886683464, + 0.013361294753849506, + 0.011784027330577374, + 0.010762856341898441, + 0.008702150546014309, + 0.012080466374754906, + 0.01813657023012638, + 0.02011920139193535, + 0.012658064253628254, + 0.0162922665476799, + 0.011782691814005375, + 0.011590373702347279, + 0.008755448274314404, + 0.019581640139222145, + 0.009303583763539791, + 0.008689009584486485, + 0.009570339694619179, + 0.012018322013318539, + 0.00802503153681755, + 0.008985957130789757, + 0.008160344325006008, + 0.012110058218240738, + 0.014480466954410076, + 0.009403185918927193, + 0.01299747359007597, + 0.011570594273507595, + 0.010116436518728733, + 0.00763101177290082, + 0.007612486835569143, + 0.012904980219900608, + 0.019283700734376907, + 0.006754850968718529, + 0.007005779072642326, + 0.009663973934948444, + 0.017623290419578552, + 0.011373957619071007, + 0.01347373891621828, + 0.01257823221385479, + 0.01677238754928112, + 0.012475901283323765, + 0.013833098113536835, + 0.010914020240306854, + 0.011603161692619324, + 0.005079778376966715, + 0.007601217366755009, + 0.007870138622820377, + 0.00951739214360714, + 0.0061050341464579105, + 0.005543645936995745, + 0.006903238128870726, + 0.008407198823988438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.17987053096294403 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007205988513305783 + ], + "opposite_logprob_tokens": [ + -7.250720500946045 + ], + "logprob_reward": -0.0007205988513305783, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007205988513305783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13454850018024445, + 0.10762282460927963, + 0.051847923547029495, + 0.12413392961025238, + 0.0788143202662468, + 0.038762979209423065, + 0.07208613306283951, + 0.07484246045351028, + 0.12016545236110687, + 0.19717548787593842 + ], + "credit_map": [ + 0.014085442759096622, + 0.01037953607738018, + 0.011906063184142113, + 0.02686736173927784, + 0.01096847839653492, + 0.023545313626527786, + 0.013267830945551395, + 0.02352847345173359, + 0.008020791225135326, + 0.006862569600343704, + 0.04254186153411865, + 0.019690152257680893, + 0.004993319511413574, + 0.008562454953789711, + 0.008697660639882088, + 0.008254008367657661, + 0.008852844126522541, + 0.0054315063171088696, + 0.006482529919594526, + 0.006193654146045446, + 0.007639049552381039, + 0.006013359408825636, + 0.006640416570007801, + 0.004594563972204924, + 0.02159450389444828, + 0.013445092365145683, + 0.012911011464893818, + 0.014250424690544605, + 0.008977356366813183, + 0.02457396872341633, + 0.012966521084308624, + 0.015415052883327007, + 0.008521020412445068, + 0.005463750101625919, + 0.012452281080186367, + 0.015078735537827015, + 0.010575229302048683, + 0.006716148462146521, + 0.010351743549108505, + 0.00965540949255228, + 0.004561824724078178, + 0.004333162680268288, + 0.007545019034296274, + 0.004475818481296301, + 0.005047124810516834, + 0.004523993469774723, + 0.00468640960752964, + 0.003589628729969263, + 0.007232813164591789, + 0.0076580834574997425, + 0.006769936997443438, + 0.010779308155179024, + 0.009626632556319237, + 0.01125294715166092, + 0.009829547256231308, + 0.008936867117881775, + 0.006480678915977478, + 0.007862220518290997, + 0.010479365475475788, + 0.008008850738406181, + 0.009208439849317074, + 0.010388627648353577, + 0.010622833855450153, + 0.011791447177529335, + 0.012029125355184078, + 0.016598114743828773, + 0.017307182773947716, + 0.0187054593116045, + 0.014097176492214203, + 0.019534030929207802, + 0.010131207294762135, + 0.01176315825432539, + 0.019483184441924095, + 0.013161329552531242, + 0.04882654920220375, + 0.06425108015537262, + 0.009458499029278755, + 0.008642328903079033, + 0.022414639592170715, + 0.010937873274087906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.0007205988513305783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864217709749937 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.00864217709749937, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864217709749937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2470322996377945, + 0.15223056077957153, + 0.0573757030069828, + 0.055078089237213135, + 0.07217778265476227, + 0.05461021512746811, + 0.05522928759455681, + 0.05843742936849594, + 0.10019678622484207, + 0.1476319283246994 + ], + "credit_map": [ + 0.022257816046476364, + 0.021917810663580894, + 0.02914126217365265, + 0.05875253677368164, + 0.018757885321974754, + 0.04799612611532211, + 0.023455148562788963, + 0.024753721430897713, + 0.011532061733305454, + 0.009170952253043652, + 0.060446418821811676, + 0.02810869924724102, + 0.006094914861023426, + 0.01608813740313053, + 0.011825368739664555, + 0.008964000269770622, + 0.00857281032949686, + 0.0066702174954116344, + 0.008747220039367676, + 0.007777962367981672, + 0.006707027554512024, + 0.00787272397428751, + 0.004915306810289621, + 0.006112433038651943, + 0.008923649787902832, + 0.004884403198957443, + 0.008086943067610264, + 0.0075127966701984406, + 0.005161546636372805, + 0.008009730838239193, + 0.007329510524868965, + 0.005169507581740618, + 0.007579596247524023, + 0.007968840189278126, + 0.011958123184740543, + 0.012462242506444454, + 0.007269163150340319, + 0.0083913654088974, + 0.008065469563007355, + 0.008482977747917175, + 0.006031873635947704, + 0.006674936041235924, + 0.00925302691757679, + 0.00852782279253006, + 0.005961817689239979, + 0.0075003416277468204, + 0.005806630477309227, + 0.0048537664115428925, + 0.006284018978476524, + 0.005162177141755819, + 0.007247535511851311, + 0.012156650424003601, + 0.005412816535681486, + 0.007799364160746336, + 0.006378279533237219, + 0.00478844391182065, + 0.005642377305775881, + 0.004892537835985422, + 0.007881157100200653, + 0.012155000120401382, + 0.00642883125692606, + 0.005362104158848524, + 0.009804019704461098, + 0.006271402817219496, + 0.00992382038384676, + 0.014926204457879066, + 0.012831675820052624, + 0.012541363947093487, + 0.009391941130161285, + 0.015288950875401497, + 0.014554598368704319, + 0.01073822844773531, + 0.013539985753595829, + 0.012699314393103123, + 0.03074263036251068, + 0.03635887801647186, + 0.010804166086018085, + 0.01756991818547249, + 0.01657300814986229, + 0.009344019927084446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.00864217709749937 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot begins, following a dusty crimson off-road jeep as it slowly but steadily rolls across the vast, jagged crest of the ancient Great Wall. The massive off-road tires rumble over the weathered gray masonry, their bodywork reflecting a dull luster in the setting sun. Then, the camera slowly and smoothly zooms out, backward and upward, the jeep gradually shrinking in the frame, transforming from a powerful, concrete figure into a solitary, moving red dot. As the camera continues to retreat, the Great Wall's majestic fullness is revealed: like a giant stone dragon, it winds endlessly along the ridges of rolling green mountains, its towering beacon towers looming in the distance. Soft golden light gilds the entire scene with a warm and tranquil hue, casting long, gentle shadows across the valley, and mist quietly rising between the mountain cols. Ultimately, the entire picture presents a breathtaking sense of solemnity and loneliness that transcends time and space, highlighting the insignificance and tenacity of human creation in the face of the majestic nature.", + "seed": 6126, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_4/index_5388_seed_6126_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07033659457229077, + "vlm_logprob": -0.07033659457229077, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.07033659457229077, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.053989093750715256 + ], + "opposite_logprob_tokens": [ + -3.1789891719818115 + ], + "logprob_reward": -0.053989093750715256, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.053989093750715256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2619207799434662, + 0.11602954566478729, + 0.11892658472061157, + 0.18346406519412994, + 0.04593396559357643, + 0.022278685122728348, + 0.07345041632652283, + 0.04913291707634926, + 0.0695343092083931, + 0.05932873487472534 + ], + "credit_map": [ + 0.029762813821434975, + 0.03246847912669182, + 0.03587774187326431, + 0.027931444346904755, + 0.05148160085082054, + 0.02580244652926922, + 0.023267947137355804, + 0.03532828763127327, + 0.01039834599941969, + 0.008838810957968235, + 0.02046520635485649, + 0.018923982977867126, + 0.011034526862204075, + 0.012622429057955742, + 0.014258865267038345, + 0.019487356767058372, + 0.009427388198673725, + 0.010028183460235596, + 0.036734696477651596, + 0.024797534570097923, + 0.012434964999556541, + 0.0068341572768986225, + 0.006818823050707579, + 0.011850821785628796, + 0.01903769001364708, + 0.012177543714642525, + 0.02877981588244438, + 0.03403884172439575, + 0.031045973300933838, + 0.015013979747891426, + 0.020587943494319916, + 0.022782260552048683, + 0.005465279798954725, + 0.007522831670939922, + 0.006371191702783108, + 0.011206263676285744, + 0.004255247768014669, + 0.0038491995073854923, + 0.003562170546501875, + 0.003701774636283517, + 0.0021928870119154453, + 0.0036218450404703617, + 0.003509435337036848, + 0.0031861085444688797, + 0.002909224247559905, + 0.001948899356648326, + 0.0020850489381700754, + 0.002825235016644001, + 0.0067484332248568535, + 0.008285775780677795, + 0.013265666551887989, + 0.010866822674870491, + 0.0126753905788064, + 0.00514741986989975, + 0.00872876401990652, + 0.007732136640697718, + 0.007233988028019667, + 0.004860221408307552, + 0.013970396481454372, + 0.006111226975917816, + 0.0054549952037632465, + 0.004121946170926094, + 0.0033318737987428904, + 0.004048266913741827, + 0.00794217362999916, + 0.005391338374465704, + 0.011653685942292213, + 0.013923446647822857, + 0.01019336748868227, + 0.005412161350250244, + 0.007638064678758383, + 0.007380063645541668, + 0.007586086168885231, + 0.006976866628974676, + 0.008363834582269192, + 0.00639876164495945, + 0.013235975056886673, + 0.006483148317784071, + 0.004364796914160252, + 0.005919257178902626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.053989093750715256 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01111565064638853 + ], + "opposite_logprob_tokens": [ + -4.511115550994873 + ], + "logprob_reward": -0.01111565064638853, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01111565064638853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15334656834602356, + 0.10706577450037003, + 0.08081445842981339, + 0.10107147693634033, + 0.09802374988794327, + 0.04470027610659599, + 0.10636549443006516, + 0.11210384219884872, + 0.08951392769813538, + 0.1069943830370903 + ], + "credit_map": [ + 0.025429261848330498, + 0.026897896081209183, + 0.020181844010949135, + 0.021169915795326233, + 0.02182222716510296, + 0.012332023121416569, + 0.012028827331960201, + 0.013484586030244827, + 0.010739684104919434, + 0.009027096442878246, + 0.01857406087219715, + 0.021166188642382622, + 0.011164549738168716, + 0.009589274413883686, + 0.01243586651980877, + 0.014369047246873379, + 0.008647577837109566, + 0.010773264802992344, + 0.012312577106058598, + 0.013025890104472637, + 0.010958313941955566, + 0.00593404658138752, + 0.008127248845994473, + 0.011035545729100704, + 0.01015379000455141, + 0.015273786149919033, + 0.0177140012383461, + 0.012005984783172607, + 0.017813686281442642, + 0.010196113958954811, + 0.007390020415186882, + 0.010524094104766846, + 0.009699012152850628, + 0.013756857253611088, + 0.016807977110147476, + 0.017608657479286194, + 0.012746200896799564, + 0.010261181741952896, + 0.008971299976110458, + 0.008172562345862389, + 0.006109695881605148, + 0.00580387469381094, + 0.00784417986869812, + 0.005880582146346569, + 0.0034098776523023844, + 0.004249383229762316, + 0.004615721292793751, + 0.006786960642784834, + 0.013231621123850346, + 0.015795236453413963, + 0.01874733902513981, + 0.015441752970218658, + 0.01854742504656315, + 0.008371710777282715, + 0.008024253882467747, + 0.008206154219806194, + 0.014746845699846745, + 0.012451039627194405, + 0.014450609683990479, + 0.013649318367242813, + 0.015444111078977585, + 0.012449320405721664, + 0.01556952390819788, + 0.01334307435899973, + 0.009274194948375225, + 0.008720175363123417, + 0.011985558085143566, + 0.016574999317526817, + 0.009877604432404041, + 0.008118540048599243, + 0.015199283137917519, + 0.009763576090335846, + 0.01157286949455738, + 0.008797723799943924, + 0.013073793612420559, + 0.013793610036373138, + 0.023852137848734856, + 0.011564482934772968, + 0.013081828132271767, + 0.011257946491241455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.01111565064638853 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2812996804714203 + ], + "opposite_logprob_tokens": [ + -1.4062997102737427 + ], + "logprob_reward": -0.2812996804714203, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2812996804714203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13850809633731842, + 0.10221868753433228, + 0.1610352098941803, + 0.11408593505620956, + 0.08416104316711426, + 0.04465985298156738, + 0.17289945483207703, + 0.09287932515144348, + 0.036768022924661636, + 0.05278436467051506 + ], + "credit_map": [ + 0.01928708516061306, + 0.018427208065986633, + 0.017299264669418335, + 0.01289364229887724, + 0.03191601485013962, + 0.013582591898739338, + 0.012731286697089672, + 0.01237101387232542, + 0.011996393091976643, + 0.016643058508634567, + 0.01537577249109745, + 0.013348725624382496, + 0.019307317212224007, + 0.008676802739501, + 0.008230346255004406, + 0.008640282787382603, + 0.012880888767540455, + 0.014508260414004326, + 0.03941987827420235, + 0.0411815382540226, + 0.018403420224785805, + 0.015767866745591164, + 0.007464251015335321, + 0.011409119702875614, + 0.012391933239996433, + 0.01391317043453455, + 0.01729440875351429, + 0.02223876863718033, + 0.014104095287621021, + 0.010948012582957745, + 0.00956194382160902, + 0.013633610680699348, + 0.010325784794986248, + 0.0114233223721385, + 0.009907149709761143, + 0.011877881363034248, + 0.01253625750541687, + 0.01095748133957386, + 0.0064575811848044395, + 0.01067559327930212, + 0.005691145546734333, + 0.004772053565829992, + 0.008558143861591816, + 0.0049375384114682674, + 0.005956864450126886, + 0.005203074775636196, + 0.004301427863538265, + 0.005239612888544798, + 0.016902726143598557, + 0.01878530904650688, + 0.03153776749968529, + 0.01575339213013649, + 0.03214845433831215, + 0.012506996281445026, + 0.023295050486922264, + 0.02196977473795414, + 0.012414897792041302, + 0.011792821809649467, + 0.011755738407373428, + 0.008603333495557308, + 0.023442773148417473, + 0.009448119439184666, + 0.009045091457664967, + 0.006376560311764479, + 0.004075079225003719, + 0.002675484400242567, + 0.006404418498277664, + 0.005085780750960112, + 0.004178168717771769, + 0.0031339596025645733, + 0.006770768202841282, + 0.004444367717951536, + 0.00601954897865653, + 0.0072180163115262985, + 0.006638688500970602, + 0.004929214250296354, + 0.012737926095724106, + 0.005655098240822554, + 0.005098575260490179, + 0.004487303085625172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.2812996804714203 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011826909612864256 + ], + "opposite_logprob_tokens": [ + -6.751182556152344 + ], + "logprob_reward": -0.0011826909612864256, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011826909612864256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11422072350978851, + 0.053729817271232605, + 0.08299612253904343, + 0.13884806632995605, + 0.16118942201137543, + 0.05221128091216087, + 0.09085056930780411, + 0.13771894574165344, + 0.056463051587343216, + 0.11177201569080353 + ], + "credit_map": [ + 0.02145107463002205, + 0.01688370108604431, + 0.016363341361284256, + 0.017860515043139458, + 0.015170522034168243, + 0.007013134192675352, + 0.009803488850593567, + 0.00967494398355484, + 0.005854041315615177, + 0.007261219434440136, + 0.0072406562976539135, + 0.008085803128778934, + 0.009543835185468197, + 0.006290373392403126, + 0.004492178093641996, + 0.004961710888892412, + 0.007429827935993671, + 0.008619168773293495, + 0.02375062182545662, + 0.014276674948632717, + 0.010496741160750389, + 0.006693306379020214, + 0.004198635462671518, + 0.0075311437249183655, + 0.012099409475922585, + 0.01638174057006836, + 0.01720176823437214, + 0.01833321526646614, + 0.03145701065659523, + 0.012747999280691147, + 0.014984791167080402, + 0.015642134472727776, + 0.021369386464357376, + 0.019316622987389565, + 0.0195316169410944, + 0.02358923852443695, + 0.024655504152178764, + 0.01931632123887539, + 0.012903734110295773, + 0.02050699107348919, + 0.0052684820257127285, + 0.005070678424090147, + 0.006489926483482122, + 0.005370696075260639, + 0.01148603018373251, + 0.006692232098430395, + 0.006752897519618273, + 0.005080335307866335, + 0.010778848081827164, + 0.01942913606762886, + 0.011716794222593307, + 0.009492914192378521, + 0.012839200906455517, + 0.009443417191505432, + 0.008181905373930931, + 0.008968351408839226, + 0.01592201553285122, + 0.01560947299003601, + 0.03118237294256687, + 0.01441288460046053, + 0.026249205693602562, + 0.010095035657286644, + 0.009874124079942703, + 0.01437383983284235, + 0.006294446997344494, + 0.006521354429423809, + 0.008317194879055023, + 0.006804114673286676, + 0.005655522923916578, + 0.004667272791266441, + 0.010213232599198818, + 0.00798991322517395, + 0.010838882066309452, + 0.017458653077483177, + 0.011273255571722984, + 0.010974989272654057, + 0.028509344905614853, + 0.011856494471430779, + 0.012358739972114563, + 0.008501646108925343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.0011826909612864256 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095857031643391 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004095857031643391, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095857031643391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.370450884103775, + 0.08046842366456985, + 0.06313135474920273, + 0.07075700163841248, + 0.1305622011423111, + 0.025450922548770905, + 0.029427887871861458, + 0.09118134528398514, + 0.0578673779964447, + 0.08070264011621475 + ], + "credit_map": [ + 0.08044163882732391, + 0.05986100062727928, + 0.07069572806358337, + 0.04332267493009567, + 0.04341006278991699, + 0.021272677928209305, + 0.024538928642868996, + 0.0269081462174654, + 0.008601177483797073, + 0.011042517609894276, + 0.020482022315263748, + 0.014721576124429703, + 0.009497688151896, + 0.004564552567899227, + 0.005660582333803177, + 0.005898297764360905, + 0.007806912064552307, + 0.007864422164857388, + 0.00978982262313366, + 0.00971424113959074, + 0.011064398102462292, + 0.005844568368047476, + 0.004248981364071369, + 0.0067980047315359116, + 0.007357134483754635, + 0.006780758034437895, + 0.011294768191874027, + 0.013767918571829796, + 0.01008510123938322, + 0.005303380079567432, + 0.0072724181227386, + 0.00889552105218172, + 0.024654066190123558, + 0.017048358917236328, + 0.019146831706166267, + 0.015833687037229538, + 0.018331129103899002, + 0.013515239581465721, + 0.008888108655810356, + 0.013144763186573982, + 0.0027310361620038748, + 0.0027945232577621937, + 0.003933924250304699, + 0.0031911409460008144, + 0.0038635239470750093, + 0.002821484813466668, + 0.002828697208315134, + 0.00328659126535058, + 0.003566458821296692, + 0.002916156081482768, + 0.00392826646566391, + 0.003391850274056196, + 0.0048203859478235245, + 0.0034072939306497574, + 0.0037379085551947355, + 0.0036595661658793688, + 0.013679878786206245, + 0.008235502056777477, + 0.02536475658416748, + 0.009958227165043354, + 0.00982244685292244, + 0.010315349325537682, + 0.0059752753004431725, + 0.007829900830984116, + 0.0061394041404128075, + 0.007759850937873125, + 0.008628836832940578, + 0.007214243058115244, + 0.008782472461462021, + 0.005537519697099924, + 0.007459136191755533, + 0.006345909088850021, + 0.010018804110586643, + 0.012164623476564884, + 0.011026003398001194, + 0.008429660461843014, + 0.017381701618433, + 0.008371235802769661, + 0.007646344136446714, + 0.0056642573326826096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.004095857031643391 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot begins, following a dusty crimson off-road jeep as it slowly but steadily rolls across the vast, jagged crest of the ancient Great Wall. The massive off-road tires rumble over the weathered gray masonry, their bodywork reflecting a dull luster in the setting sun. Then, the camera slowly and smoothly zooms out, backward and upward, the jeep gradually shrinking in the frame, transforming from a powerful, concrete figure into a solitary, moving red dot. As the camera continues to retreat, the Great Wall's majestic fullness is revealed: like a giant stone dragon, it winds endlessly along the ridges of rolling green mountains, its towering beacon towers looming in the distance. Soft golden light gilds the entire scene with a warm and tranquil hue, casting long, gentle shadows across the valley, and mist quietly rising between the mountain cols. Ultimately, the entire picture presents a breathtaking sense of solemnity and loneliness that transcends time and space, highlighting the insignificance and tenacity of human creation in the face of the majestic nature.", + "seed": 6127, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_4/index_5388_seed_6127_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04736869843909517, + "vlm_logprob": -0.04736869843909517, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.04736869843909517, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10456191003322601 + ], + "opposite_logprob_tokens": [ + -2.4795618057250977 + ], + "logprob_reward": -0.10456191003322601, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10456191003322601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4850026071071625, + 0.103364497423172, + 0.05589751899242401, + 0.07890497893095016, + 0.03879120573401451, + 0.039595648646354675, + 0.06413383036851883, + 0.041993629187345505, + 0.03625018894672394, + 0.05606594309210777 + ], + "credit_map": [ + 0.04800054803490639, + 0.0576513335108757, + 0.06805216521024704, + 0.05434328690171242, + 0.05054929479956627, + 0.07910484820604324, + 0.06560464948415756, + 0.061696507036685944, + 0.010088379494845867, + 0.011016175150871277, + 0.01363184954971075, + 0.017326032742857933, + 0.010032812133431435, + 0.013490548357367516, + 0.012564538046717644, + 0.01521416287869215, + 0.006716834846884012, + 0.008143506944179535, + 0.004913696553558111, + 0.009996823035180569, + 0.008989142253994942, + 0.00634670490399003, + 0.006151337176561356, + 0.0046394700184464455, + 0.010834082029759884, + 0.012164140120148659, + 0.011826016008853912, + 0.014433285221457481, + 0.006868487223982811, + 0.006659358274191618, + 0.008028090000152588, + 0.008091521449387074, + 0.003890746971592307, + 0.0041281916201114655, + 0.006303780246526003, + 0.004211097955703735, + 0.0036501006688922644, + 0.003889994230121374, + 0.005797556135803461, + 0.0069197374396026134, + 0.004990518093109131, + 0.003443498630076647, + 0.00508984737098217, + 0.008331851102411747, + 0.0026287981308996677, + 0.003687882563099265, + 0.006724081002175808, + 0.0046991705894470215, + 0.008176755160093307, + 0.0049160681664943695, + 0.0077476720325648785, + 0.010350213386118412, + 0.005968480836600065, + 0.01198350079357624, + 0.005884748883545399, + 0.009106390178203583, + 0.004994899965822697, + 0.008289158344268799, + 0.006630865857005119, + 0.00622300710529089, + 0.0033913091756403446, + 0.0038686133921146393, + 0.004087063483893871, + 0.004508711863309145, + 0.0036752023734152317, + 0.0027465680614113808, + 0.003850442823022604, + 0.005658669862896204, + 0.003238135948777199, + 0.005498274229466915, + 0.00668004434555769, + 0.004902847111225128, + 0.007181304972618818, + 0.008689398877322674, + 0.01442271564155817, + 0.010420595295727253, + 0.00339163257740438, + 0.005092398729175329, + 0.0033552192617207766, + 0.003512678202241659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.10456191003322601 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076613640412688255 + ], + "opposite_logprob_tokens": [ + -4.88266134262085 + ], + "logprob_reward": -0.0076613640412688255, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076613640412688255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19149553775787354, + 0.1430434286594391, + 0.06736239790916443, + 0.06316573172807693, + 0.0626320168375969, + 0.06780519336462021, + 0.03183702379465103, + 0.28571823239326477, + 0.045318569988012314, + 0.04162183403968811 + ], + "credit_map": [ + 0.02775723487138748, + 0.021834438666701317, + 0.017076049000024796, + 0.019217919558286667, + 0.02105392888188362, + 0.029636211693286896, + 0.032730743288993835, + 0.02218901738524437, + 0.016648955643177032, + 0.011853337287902832, + 0.014072061516344547, + 0.017498891800642014, + 0.017146503552794456, + 0.023949380964040756, + 0.01836320571601391, + 0.023511093109846115, + 0.009409470483660698, + 0.007212681230157614, + 0.006466665770858526, + 0.009299242869019508, + 0.00893317349255085, + 0.006742521654814482, + 0.008281189948320389, + 0.011017451994121075, + 0.011110617779195309, + 0.007961252704262733, + 0.010157030075788498, + 0.010413781739771366, + 0.004596909508109093, + 0.0045688571408391, + 0.007608736399561167, + 0.00674854451790452, + 0.00720266904681921, + 0.006760602816939354, + 0.010386714711785316, + 0.0062896860763430595, + 0.005260533653199673, + 0.007932215929031372, + 0.00916347187012434, + 0.009636122733354568, + 0.005358410999178886, + 0.0046295891515910625, + 0.007326776161789894, + 0.012844019569456577, + 0.00613072793930769, + 0.005078466609120369, + 0.016340453177690506, + 0.010096753016114235, + 0.0027687991969287395, + 0.0028124931268393993, + 0.004381941631436348, + 0.006969033274799585, + 0.002473144792020321, + 0.00396673334762454, + 0.003932974301278591, + 0.00453190179541707, + 0.03075427934527397, + 0.04722387716174126, + 0.04817378148436546, + 0.07727405428886414, + 0.016371289268136024, + 0.020274324342608452, + 0.0242454893887043, + 0.02140115760266781, + 0.005476691760122776, + 0.0042996094562113285, + 0.008598990738391876, + 0.006158959586173296, + 0.004826548509299755, + 0.006439827382564545, + 0.005681097973138094, + 0.003836843417957425, + 0.0056586721912026405, + 0.0035891744773834944, + 0.0075107719749212265, + 0.006653684191405773, + 0.0023838954512029886, + 0.004848011769354343, + 0.004242736380547285, + 0.006734885741025209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.0076613640412688255 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290945112705231 + ], + "opposite_logprob_tokens": [ + -2.2379095554351807 + ], + "logprob_reward": -0.11290945112705231, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290945112705231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33945244550704956, + 0.1093827560544014, + 0.06039339303970337, + 0.10099831223487854, + 0.056814346462488174, + 0.13742154836654663, + 0.0466812402009964, + 0.06068456918001175, + 0.05716023966670036, + 0.03101121075451374 + ], + "credit_map": [ + 0.04748285934329033, + 0.037219636142253876, + 0.033070992678403854, + 0.030829856172204018, + 0.025423739105463028, + 0.05137066915631294, + 0.0751870647072792, + 0.038867611438035965, + 0.011489505879580975, + 0.009554388001561165, + 0.01566598191857338, + 0.01087099313735962, + 0.012585265561938286, + 0.017166538164019585, + 0.012769431807100773, + 0.019280659034848213, + 0.006010841578245163, + 0.007637017872184515, + 0.005650291219353676, + 0.008771839551627636, + 0.010416102595627308, + 0.0053069028072059155, + 0.008654465898871422, + 0.007945933379232883, + 0.011074481531977654, + 0.018535424023866653, + 0.01970897801220417, + 0.02068798430263996, + 0.007952520623803139, + 0.004713527858257294, + 0.010565604083240032, + 0.0077597880735993385, + 0.01146797277033329, + 0.004590366035699844, + 0.013035856187343597, + 0.006876020226627588, + 0.0037434955593198538, + 0.005693006329238415, + 0.005703244358301163, + 0.0057043819688260555, + 0.015704527497291565, + 0.013315063901245594, + 0.019240714609622955, + 0.02181091532111168, + 0.011024141684174538, + 0.011092184111475945, + 0.02758440375328064, + 0.01764960214495659, + 0.0065481895580887794, + 0.004489861894398928, + 0.005813864525407553, + 0.007610632572323084, + 0.0036690309643745422, + 0.005022342316806316, + 0.0070761325769126415, + 0.0064511834643781185, + 0.005450234282761812, + 0.004517544526606798, + 0.010649686679244041, + 0.01578592136502266, + 0.004820672795176506, + 0.005668679252266884, + 0.00644622603431344, + 0.007345602381974459, + 0.0040927487425506115, + 0.003148345509544015, + 0.009830773808062077, + 0.017035052180290222, + 0.003000169061124325, + 0.004408495035022497, + 0.010609855875372887, + 0.005034798756241798, + 0.002976377494633198, + 0.002220987109467387, + 0.007227027788758278, + 0.0093104038387537, + 0.001363781513646245, + 0.0024828724563121796, + 0.002656159456819296, + 0.0027736006304621696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.11290945112705231 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006367324967868626 + ], + "opposite_logprob_tokens": [ + -7.375636577606201 + ], + "logprob_reward": -0.0006367324967868626, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006367324967868626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1912839710712433, + 0.11311552673578262, + 0.08252382278442383, + 0.10161754488945007, + 0.06073008477687836, + 0.15569639205932617, + 0.07005112618207932, + 0.07273586094379425, + 0.0775788351893425, + 0.07466680556535721 + ], + "credit_map": [ + 0.019167693331837654, + 0.02498706430196762, + 0.022962788119912148, + 0.018010931089520454, + 0.020719850435853004, + 0.022743558511137962, + 0.03791802003979683, + 0.02477407641708851, + 0.013870333321392536, + 0.013491664081811905, + 0.013479324989020824, + 0.01520899124443531, + 0.01253530103713274, + 0.016497977077960968, + 0.013170375488698483, + 0.014861566945910454, + 0.008973024785518646, + 0.011325566098093987, + 0.008159282617270947, + 0.015867849811911583, + 0.01405137125402689, + 0.007590076420456171, + 0.009919148869812489, + 0.006637506652623415, + 0.014031820930540562, + 0.015722006559371948, + 0.011450707912445068, + 0.016952473670244217, + 0.00917696300894022, + 0.010286740027368069, + 0.011313127353787422, + 0.01268371194601059, + 0.012710841372609138, + 0.00717557268217206, + 0.010477789677679539, + 0.006701381411403418, + 0.0040063136257231236, + 0.005993513856083155, + 0.00693465443328023, + 0.006730022374540567, + 0.0144310612231493, + 0.009157444350421429, + 0.017022034153342247, + 0.031802885234355927, + 0.01321677677333355, + 0.017482789233326912, + 0.035346172749996185, + 0.017237229272723198, + 0.006868721917271614, + 0.006464908830821514, + 0.008165132254362106, + 0.015739308670163155, + 0.0050702583976089954, + 0.010077349841594696, + 0.008909592404961586, + 0.008755861781537533, + 0.008934713900089264, + 0.009295132011175156, + 0.01312883198261261, + 0.01830020733177662, + 0.0038901553489267826, + 0.0067225610837340355, + 0.006621632259339094, + 0.005842628423124552, + 0.009338281117379665, + 0.005063581746071577, + 0.011464111506938934, + 0.01874541863799095, + 0.005859578028321266, + 0.009431462734937668, + 0.009921373799443245, + 0.007755032740533352, + 0.00588661665096879, + 0.008609589189291, + 0.017517289146780968, + 0.02100394107401371, + 0.003552632872015238, + 0.005682831164449453, + 0.007388456724584103, + 0.005025456193834543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0006367324967868626 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dusty crimson off-road jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jeep rolling across the jagged crest of the Great Wall made of weathered gray masonry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a low-angle tracking shot following the jeep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly zoom out, backward, and upward to reveal the Great Wall winding along green mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden light with mist rising between the mountain cols?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074034497141838 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.011074034497141838, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074034497141838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31550976634025574, + 0.10080479830503464, + 0.04382007196545601, + 0.09652005136013031, + 0.10562267154455185, + 0.10926330834627151, + 0.04660337045788765, + 0.10107266157865524, + 0.022462179884314537, + 0.058321110904216766 + ], + "credit_map": [ + 0.04231500253081322, + 0.03034510836005211, + 0.022046852856874466, + 0.024886051192879677, + 0.028403962031006813, + 0.06926091015338898, + 0.05634276568889618, + 0.04190913215279579, + 0.012256994843482971, + 0.009968574158847332, + 0.010861109010875225, + 0.015174207277595997, + 0.013923094607889652, + 0.011740150861442089, + 0.011414826847612858, + 0.015465841628611088, + 0.005711347330361605, + 0.004400370176881552, + 0.0043342746794223785, + 0.009428774006664753, + 0.0037590174470096827, + 0.005451972130686045, + 0.005071849096566439, + 0.0056624701246619225, + 0.011604360304772854, + 0.015253406949341297, + 0.019031111150979996, + 0.018298190087080002, + 0.007695937063544989, + 0.006776674184948206, + 0.010118825361132622, + 0.007741548120975494, + 0.017353788018226624, + 0.018521467223763466, + 0.019989795982837677, + 0.010791448876261711, + 0.0064885965548455715, + 0.009355343878269196, + 0.011762247420847416, + 0.011359983123838902, + 0.011244155466556549, + 0.0132352439686656, + 0.014605369418859482, + 0.02102985605597496, + 0.007542725652456284, + 0.017149263992905617, + 0.010675391182303429, + 0.013781299814581871, + 0.006730850785970688, + 0.0038520877715200186, + 0.006222599186003208, + 0.0074470071122050285, + 0.003409380791708827, + 0.0070098889991641045, + 0.004811189137399197, + 0.007120361085981131, + 0.011018935590982437, + 0.01564444601535797, + 0.01715487241744995, + 0.025053923949599266, + 0.004758551251143217, + 0.008848639205098152, + 0.009517882019281387, + 0.009075412526726723, + 0.0033185146749019623, + 0.0031039374880492687, + 0.003281056648120284, + 0.0029520285315811634, + 0.0018437521066516638, + 0.0028661410324275494, + 0.0027329924050718546, + 0.002363756997510791, + 0.007585510611534119, + 0.006728961132466793, + 0.014109327457845211, + 0.013797649182379246, + 0.0025414442643523216, + 0.00416941475123167, + 0.0046860771253705025, + 0.0047027296386659145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.011074034497141838 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_5/index_tensor([6975, 6975])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_5/index_tensor([6975, 6975])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3a67d1ee340143b608e33a64fba087d6713d991a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_5/index_tensor([6975, 6975])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatically dynamic Snorri camera lens pinned a mysterious figure to the very center of the frame, creating a dizzying sensation of a stable subject amidst the world's turbulent motion. This figure, outlined in a form-fitting suit of dark armor gleaming with metallic and keratin, was enveloped in an ominous crimson glow of energy, like burning blood, a light that even stirred tiny arcs of electricity in the air around him. He tore through the air with thunderous force, slashing vertically from the sky. His body tensed under the immense acceleration of gravity, his grip on a massive blade composed of pure red light energy slicing a deadly arc through the air, poised to deliver a devastating blow. Around him, the world dissolved into a chaotic motion blur, gray clouds, sparse raindrops, and distant city lights stretched into countless vertical, rapidly streaking streaks of light and shadow, creating a stark visual contrast with the steady, incandescent glow of his body. This high-speed fall produces a strong cinematic light effect. The red halo creates glare and edge distortion on the lens, and the particles in the air are illuminated. The entire picture is filled with the destructive power of doomsday and a suffocating atmosphere of tension.", + "seed": 30329, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_5/index_6975_seed_30329_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.006396719580516219, + "vlm_logprob": -0.006396719580516219, + "vlm_yes_no_margin": 5.525, + "vlm_token_credit": -0.006396719580516219, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011640955694019794 + ], + "opposite_logprob_tokens": [ + -4.511641025543213 + ], + "logprob_reward": -0.011640955694019794, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011640955694019794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07937037944793701, + 0.03015955165028572, + 0.40245988965034485, + 0.014918800443410873, + 0.05324758589267731, + 0.09699350595474243, + 0.03652524948120117, + 0.05901392549276352, + 0.06786364316940308, + 0.15944746136665344 + ], + "credit_map": [ + 0.010171858593821526, + 0.008577484637498856, + 0.024016479030251503, + 0.006236013025045395, + 0.0064890459179878235, + 0.007392348255962133, + 0.0056478711776435375, + 0.010839274153113365, + 0.0046615516766905785, + 0.0023736392613500357, + 0.004319056868553162, + 0.003484070999547839, + 0.0031163112726062536, + 0.003734597470611334, + 0.003927232697606087, + 0.004543093498796225, + 0.07449306547641754, + 0.0151870958507061, + 0.11830326169729233, + 0.0193928275257349, + 0.015729524195194244, + 0.04482929781079292, + 0.024496501311659813, + 0.0900282934308052, + 0.0013979501090943813, + 0.0014822521479800344, + 0.002593406243249774, + 0.00227624224498868, + 0.0013582721585407853, + 0.001149920979514718, + 0.0025631363969296217, + 0.0020976189989596605, + 0.0046829660423099995, + 0.005655151326209307, + 0.00995731446892023, + 0.006003657355904579, + 0.004321670159697533, + 0.009654158726334572, + 0.005978034343570471, + 0.006994633004069328, + 0.008797001093626022, + 0.008891447447240353, + 0.014461763203144073, + 0.009901340119540691, + 0.010766024701297283, + 0.01099089253693819, + 0.01321075763553381, + 0.01997428573668003, + 0.004588371608406305, + 0.0036789686419069767, + 0.0048637716099619865, + 0.004463959485292435, + 0.0046171159483492374, + 0.00539219006896019, + 0.004218864720314741, + 0.004702002741396427, + 0.007097787223756313, + 0.004629067610949278, + 0.008440800942480564, + 0.005010118242353201, + 0.007322579622268677, + 0.011972120963037014, + 0.009296293370425701, + 0.005245158914476633, + 0.008735312148928642, + 0.0063294763676822186, + 0.01285262405872345, + 0.0066703734919428825, + 0.007854115217924118, + 0.00912129133939743, + 0.008113169111311436, + 0.008187280036509037, + 0.0467936135828495, + 0.009524082764983177, + 0.04160430654883385, + 0.006085927132517099, + 0.006827196571975946, + 0.018434982746839523, + 0.004773344378918409, + 0.02540399506688118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 5 + ], + "selected_reward": -0.011640955694019794 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000503770774230361 + ], + "opposite_logprob_tokens": [ + -7.6255035400390625 + ], + "logprob_reward": -0.000503770774230361, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.000503770774230361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06908848881721497, + 0.046220969408750534, + 0.5878407955169678, + 0.04095738008618355, + 0.042856018990278244, + 0.03511207923293114, + 0.07999426126480103, + 0.035923779010772705, + 0.04040495678782463, + 0.02160118706524372 + ], + "credit_map": [ + 0.010263764299452305, + 0.005355684086680412, + 0.011664103716611862, + 0.005432282108813524, + 0.0073340414091944695, + 0.015549679286777973, + 0.004890648182481527, + 0.008598288521170616, + 0.004991115070879459, + 0.004008946008980274, + 0.007098339963704348, + 0.00507381884381175, + 0.006619676016271114, + 0.007118092849850655, + 0.00618652394041419, + 0.005124461837112904, + 0.10780148953199387, + 0.02212919108569622, + 0.17211385071277618, + 0.026354406028985977, + 0.02190960757434368, + 0.06803971529006958, + 0.03696085512638092, + 0.1325317770242691, + 0.004529087338596582, + 0.0042396546341478825, + 0.00581055972725153, + 0.007757703773677349, + 0.004406031221151352, + 0.0034419496078044176, + 0.005438250955194235, + 0.005334142595529556, + 0.004051896743476391, + 0.003093593055382371, + 0.010025820694863796, + 0.003629946382716298, + 0.0033454285003244877, + 0.008368240669369698, + 0.004189318045973778, + 0.006151779089123011, + 0.0030404850840568542, + 0.004002787172794342, + 0.008280251175165176, + 0.0035801646299660206, + 0.0037273578345775604, + 0.004424971528351307, + 0.003257961478084326, + 0.004798105917870998, + 0.012838313356041908, + 0.006569686811417341, + 0.009966207668185234, + 0.008516235277056694, + 0.013460582122206688, + 0.012258317321538925, + 0.007262160070240498, + 0.009122762829065323, + 0.004842327907681465, + 0.004180594347417355, + 0.00402878550812602, + 0.0032697829883545637, + 0.004236634820699692, + 0.00747397867962718, + 0.0036993001122027636, + 0.004192375112324953, + 0.005078252870589495, + 0.0037447421345859766, + 0.005611661355942488, + 0.004783001262694597, + 0.005527595058083534, + 0.005558621138334274, + 0.005363405682146549, + 0.004737679846584797, + 0.0025206152349710464, + 0.001923294272273779, + 0.002614859724417329, + 0.0019150940934196115, + 0.004058203659951687, + 0.003241956699639559, + 0.0019220326794311404, + 0.0034051313996315002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.000503770774230361 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947748199105263 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947748199105263, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947748199105263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10721711069345474, + 0.1133759617805481, + 0.15042375028133392, + 0.06002730876207352, + 0.05173714831471443, + 0.038663581013679504, + 0.10250287503004074, + 0.14819134771823883, + 0.041305407881736755, + 0.18655547499656677 + ], + "credit_map": [ + 0.012479433789849281, + 0.01182585395872593, + 0.03025662526488304, + 0.008467716164886951, + 0.008710208348929882, + 0.0116244712844491, + 0.009652562439441681, + 0.014200249686837196, + 0.012732633389532566, + 0.010285348631441593, + 0.017494656145572662, + 0.016993552446365356, + 0.015585214830935001, + 0.017358409240841866, + 0.012230749242007732, + 0.010695410892367363, + 0.027053212746977806, + 0.005964036565274, + 0.04204484820365906, + 0.007808942347764969, + 0.0064089358784258366, + 0.016973767429590225, + 0.009851998649537563, + 0.034318022429943085, + 0.005292105488479137, + 0.005464402958750725, + 0.01008074264973402, + 0.008004066534340382, + 0.005534612573683262, + 0.00506582111120224, + 0.009983980096876621, + 0.010601584799587727, + 0.004156875889748335, + 0.005315736867487431, + 0.01041901670396328, + 0.006681038998067379, + 0.0039909714832901955, + 0.009114539250731468, + 0.005112117622047663, + 0.006946860812604427, + 0.004211350344121456, + 0.0045967744663357735, + 0.0065353866666555405, + 0.0033691045828163624, + 0.0048309070989489555, + 0.004674594383686781, + 0.005054939072579145, + 0.005390530452132225, + 0.016487635672092438, + 0.008446021005511284, + 0.01344026904553175, + 0.010790220461785793, + 0.01489927526563406, + 0.013657414354383945, + 0.009173063561320305, + 0.01560898870229721, + 0.024324612691998482, + 0.00528552895411849, + 0.013141625560820103, + 0.006936608348041773, + 0.023289838805794716, + 0.05108043923974037, + 0.008839769288897514, + 0.015292943455278873, + 0.004801622126251459, + 0.003025088692083955, + 0.006478765048086643, + 0.004067431669682264, + 0.005076911766082048, + 0.008034912869334221, + 0.004014837089926004, + 0.005805840715765953, + 0.043027814477682114, + 0.014313501305878162, + 0.028953323140740395, + 0.010870552621781826, + 0.01610570400953293, + 0.027353012934327126, + 0.012986201792955399, + 0.032945383340120316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.005947748199105263 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028139064088463783 + ], + "opposite_logprob_tokens": [ + -5.877813816070557 + ], + "logprob_reward": -0.0028139064088463783, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028139064088463783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16212645173072815, + 0.01762506365776062, + 0.5414544939994812, + 0.011170217767357826, + 0.02228420041501522, + 0.030280768871307373, + 0.04350104555487633, + 0.02784554474055767, + 0.05694492161273956, + 0.08676727861166 + ], + "credit_map": [ + 0.01975102350115776, + 0.019385499879717827, + 0.059534702450037, + 0.00870458222925663, + 0.005415910854935646, + 0.018932925537228584, + 0.007398514077067375, + 0.023003295063972473, + 0.001990032847970724, + 0.0013099099742248654, + 0.0021446216851472855, + 0.0021173760760575533, + 0.0022522183135151863, + 0.0021144344937056303, + 0.0022209594026207924, + 0.0034755105152726173, + 0.10028575360774994, + 0.019823821261525154, + 0.1582457721233368, + 0.023653877899050713, + 0.019993338733911514, + 0.0626847967505455, + 0.033453892916440964, + 0.12331322580575943, + 0.0010039587505161762, + 0.0008452530601061881, + 0.0018279229989275336, + 0.0014651530655100942, + 0.0009537727455608547, + 0.0011923251440748572, + 0.0021509320940822363, + 0.0017308983951807022, + 0.0020408006384968758, + 0.0016064989613369107, + 0.004482208285480738, + 0.0018069231882691383, + 0.0015588175738230348, + 0.00586562929674983, + 0.0018428447656333447, + 0.003080476773902774, + 0.0030382308177649975, + 0.002189593855291605, + 0.004377591423690319, + 0.002638391451910138, + 0.00291455932892859, + 0.003942225594073534, + 0.00377084338106215, + 0.007409331388771534, + 0.005113016813993454, + 0.0039834375493228436, + 0.005853069480508566, + 0.006643436383455992, + 0.004999849013984203, + 0.004719467367976904, + 0.006250225473195314, + 0.0059385462664067745, + 0.004111859016120434, + 0.0022486187517642975, + 0.0031860736198723316, + 0.0021872969809919596, + 0.003609853330999613, + 0.007484182715415955, + 0.0027698383200913668, + 0.002247819909825921, + 0.007708589546382427, + 0.004358931910246611, + 0.009932531043887138, + 0.004888817667961121, + 0.005140711087733507, + 0.010053594596683979, + 0.006214627996087074, + 0.008647119626402855, + 0.021912982687354088, + 0.00499145220965147, + 0.020124631002545357, + 0.003440532134845853, + 0.004601823631674051, + 0.012724699452519417, + 0.0029501630924642086, + 0.016020989045500755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.0028139064088463783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0110772168263793 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.0110772168263793, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0110772168263793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09153512865304947, + 0.03794237598776817, + 0.5169876217842102, + 0.02422521635890007, + 0.02082274854183197, + 0.027535049244761467, + 0.07605710625648499, + 0.03943585976958275, + 0.07931044697761536, + 0.08614847809076309 + ], + "credit_map": [ + 0.013436471112072468, + 0.009906072169542313, + 0.021133972331881523, + 0.008519737049937248, + 0.00941549427807331, + 0.013477561995387077, + 0.008547186851501465, + 0.007098633795976639, + 0.005851052235811949, + 0.0027030452620238066, + 0.005701269023120403, + 0.0032423879019916058, + 0.004440528340637684, + 0.006094465497881174, + 0.003813087707385421, + 0.006096539553254843, + 0.09571396559476852, + 0.018889764323830605, + 0.15270568430423737, + 0.023321082815527916, + 0.01934988610446453, + 0.058971185237169266, + 0.03152632713317871, + 0.11650972813367844, + 0.002742343582212925, + 0.0020718947052955627, + 0.004513776395469904, + 0.002500153612345457, + 0.00181469670496881, + 0.002336283214390278, + 0.00366358645260334, + 0.00458248145878315, + 0.0018640717025846243, + 0.0016038345638662577, + 0.004168774466961622, + 0.0025400810409337282, + 0.0014106554444879293, + 0.004321231506764889, + 0.0022939417976886034, + 0.002620156854391098, + 0.0023814337328076363, + 0.002742804354056716, + 0.003301669843494892, + 0.002478941110894084, + 0.0032889200374484062, + 0.003317355876788497, + 0.004101181868463755, + 0.0059227426536381245, + 0.010951690375804901, + 0.0076193176209926605, + 0.007358876522630453, + 0.010254253633320332, + 0.011334569193422794, + 0.010100571438670158, + 0.011126082390546799, + 0.007311752997338772, + 0.006377766840159893, + 0.0022512665018439293, + 0.0041364338248968124, + 0.003075538668781519, + 0.005155822262167931, + 0.010283275507390499, + 0.00332766305655241, + 0.004828094504773617, + 0.010968667455017567, + 0.009384234435856342, + 0.01039844285696745, + 0.006450263317674398, + 0.008214163593947887, + 0.013991991989314556, + 0.009816513396799564, + 0.010086164809763432, + 0.012965647503733635, + 0.004669246729463339, + 0.0150684155523777, + 0.0057768747210502625, + 0.007251545786857605, + 0.017246827483177185, + 0.005097289104014635, + 0.01807263493537903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.0110772168263793 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatically dynamic Snorri camera lens pinned a mysterious figure to the very center of the frame, creating a dizzying sensation of a stable subject amidst the world's turbulent motion. This figure, outlined in a form-fitting suit of dark armor gleaming with metallic and keratin, was enveloped in an ominous crimson glow of energy, like burning blood, a light that even stirred tiny arcs of electricity in the air around him. He tore through the air with thunderous force, slashing vertically from the sky. His body tensed under the immense acceleration of gravity, his grip on a massive blade composed of pure red light energy slicing a deadly arc through the air, poised to deliver a devastating blow. Around him, the world dissolved into a chaotic motion blur, gray clouds, sparse raindrops, and distant city lights stretched into countless vertical, rapidly streaking streaks of light and shadow, creating a stark visual contrast with the steady, incandescent glow of his body. This high-speed fall produces a strong cinematic light effect. The red halo creates glare and edge distortion on the lens, and the particles in the air are illuminated. The entire picture is filled with the destructive power of doomsday and a suffocating atmosphere of tension.", + "seed": 30330, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_5/index_6975_seed_30330_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026815745537169276, + "vlm_logprob": -0.026815745537169276, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.026815745537169276, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11363347619771957 + ], + "opposite_logprob_tokens": [ + -2.238633394241333 + ], + "logprob_reward": -0.11363347619771957, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11363347619771957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8184735178947449, + 0.01392550952732563, + 0.032454364001750946, + 0.01503781694918871, + 0.03291601687669754, + 0.057146452367305756, + 0.012836803682148457, + 0.006534361280500889, + 0.0047168731689453125, + 0.005958301480859518 + ], + "credit_map": [ + 0.022441774606704712, + 0.02170681580901146, + 0.010292183607816696, + 0.03436033055186272, + 0.026162272319197655, + 0.27203166484832764, + 0.052846070379018784, + 0.37863239645957947, + 0.0012034681858494878, + 0.0016545293619856238, + 0.001503891428001225, + 0.0014520110562443733, + 0.0036167013458907604, + 0.0013439743779599667, + 0.0013356052804738283, + 0.0018153280252590775, + 0.004706220235675573, + 0.004795868881046772, + 0.0027573229745030403, + 0.0016706457827240229, + 0.003559701144695282, + 0.008495338261127472, + 0.002236123662441969, + 0.004233142361044884, + 0.0013827817747369409, + 0.0020568862091749907, + 0.002179367933422327, + 0.0018519731238484383, + 0.001848178100772202, + 0.0016336325788870454, + 0.0011453864863142371, + 0.0029396105092018843, + 0.006380686070770025, + 0.005141605157405138, + 0.005664663389325142, + 0.0024332311004400253, + 0.003585433354601264, + 0.00275562540628016, + 0.0034763221628963947, + 0.003478452330455184, + 0.0068341814912855625, + 0.006236738059669733, + 0.011610383167862892, + 0.009573172777891159, + 0.00466876570135355, + 0.003582119010388851, + 0.0048468527384102345, + 0.009794236160814762, + 0.0017239508451893926, + 0.001338302856311202, + 0.001993532758206129, + 0.0016444422071799636, + 0.0013385304482653737, + 0.0015020107384771109, + 0.001905322540551424, + 0.001390711753629148, + 0.0007493682205677032, + 0.0009220910724252462, + 0.0007345137419179082, + 0.0006676768534816802, + 0.0008332313736900687, + 0.0007251054048538208, + 0.0008120561251416802, + 0.0010903185466304421, + 0.000682400306686759, + 0.0006183662335388362, + 0.000427344290073961, + 0.0004391283437144011, + 0.0004570779565256089, + 0.00045498795225284994, + 0.000596668163780123, + 0.0010409002425149083, + 0.0005397110362537205, + 0.0009795469231903553, + 0.000526614545378834, + 0.0007137211505323648, + 0.0006758869858458638, + 0.0009830208728089929, + 0.0008022667025215924, + 0.0007375336135737598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.11363347619771957 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011886443244293332 + ], + "opposite_logprob_tokens": [ + -6.7511887550354 + ], + "logprob_reward": -0.0011886443244293332, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011886443244293332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8179426789283752, + 0.012639360502362251, + 0.032334040850400925, + 0.02756349928677082, + 0.02131849154829979, + 0.03850748762488365, + 0.019325392320752144, + 0.013231400400400162, + 0.006856335327029228, + 0.010281429626047611 + ], + "credit_map": [ + 0.022858714684844017, + 0.021072398871183395, + 0.009285219945013523, + 0.03480791300535202, + 0.025755010545253754, + 0.2733832597732544, + 0.052455075085163116, + 0.3783249855041504, + 0.00127375905867666, + 0.0015848737675696611, + 0.001626941841095686, + 0.0013033959548920393, + 0.002579321386292577, + 0.0013444157084450126, + 0.0010564426193013787, + 0.0018702094675973058, + 0.005193094722926617, + 0.0046499441377818584, + 0.0023346063680946827, + 0.001824255334213376, + 0.002986437873914838, + 0.00939515046775341, + 0.0019463076023384929, + 0.00400423863902688, + 0.0020598452538251877, + 0.0037390319630503654, + 0.0033057834953069687, + 0.004120656289160252, + 0.0035124183632433414, + 0.0034998743794858456, + 0.002149086445569992, + 0.005176799837499857, + 0.0026847836561501026, + 0.002256373642012477, + 0.003474733093753457, + 0.001867204555310309, + 0.0022488622926175594, + 0.0030503151938319206, + 0.0027003162540495396, + 0.003035899717360735, + 0.004482390824705362, + 0.0038682548329234123, + 0.010411690920591354, + 0.004083946347236633, + 0.003395715029910207, + 0.0027081738226115704, + 0.003179890336468816, + 0.006377420853823423, + 0.0029064312111586332, + 0.0018191952258348465, + 0.0032529300078749657, + 0.00260065495967865, + 0.0016142838867381215, + 0.0024440577253699303, + 0.0028348788619041443, + 0.0018529590452089906, + 0.0024480291176587343, + 0.00144617585465312, + 0.0028778105042874813, + 0.001120182452723384, + 0.0011442848481237888, + 0.0011068694293498993, + 0.0012649487471207976, + 0.0018230986315757036, + 0.0008398597710765898, + 0.001238990924321115, + 0.000744908582419157, + 0.0008121237624436617, + 0.0006871857913210988, + 0.0006023694877512753, + 0.0008855180931277573, + 0.001045378390699625, + 0.0011215610429644585, + 0.0017659239238128066, + 0.001386980409733951, + 0.001632134197279811, + 0.0008503196877427399, + 0.0008956548408605158, + 0.001222696271724999, + 0.0014061590190976858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.0011886443244293332 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040932451374828815 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.0040932451374828815, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040932451374828815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8505843877792358, + 0.02157536894083023, + 0.022090844810009003, + 0.02619953267276287, + 0.016479264944791794, + 0.023293593898415565, + 0.011017798446118832, + 0.011480135843157768, + 0.011756722815334797, + 0.005522295366972685 + ], + "credit_map": [ + 0.02367136813700199, + 0.021971747279167175, + 0.01139668095856905, + 0.03622885420918465, + 0.026492713019251823, + 0.2832571268081665, + 0.0547577403485775, + 0.3928082585334778, + 0.0023375004529953003, + 0.0024443543516099453, + 0.002294577658176422, + 0.0023944717831909657, + 0.0049051521345973015, + 0.002076002536341548, + 0.0021967319771647453, + 0.0029265801422297955, + 0.0026521438267081976, + 0.002894906559959054, + 0.002332383533939719, + 0.002584056695923209, + 0.003555203089490533, + 0.0034068559762090445, + 0.002578712534159422, + 0.0020865844562649727, + 0.0023104827851057053, + 0.0038094674237072468, + 0.0030830095056444407, + 0.0042215497232973576, + 0.003197941230610013, + 0.003449653508141637, + 0.0024202263448387384, + 0.0037072033155709505, + 0.0029589931946247816, + 0.0026115241926163435, + 0.0030765049159526825, + 0.0010526871774345636, + 0.0018360797548666596, + 0.0012349358294159174, + 0.0017287310911342502, + 0.0019798101857304573, + 0.0027077514678239822, + 0.0022628940641880035, + 0.005686966236680746, + 0.002908681519329548, + 0.002163263037800789, + 0.0016408712835982442, + 0.0019228898454457521, + 0.004000277258455753, + 0.0020833141170442104, + 0.0011782575165852904, + 0.0019230805337429047, + 0.0016558457864448428, + 0.0009383756550960243, + 0.0010890060802921653, + 0.0009184657828882337, + 0.001231453730724752, + 0.001910538412630558, + 0.0009125808137468994, + 0.0023152080830186605, + 0.0013062661746516824, + 0.0010747385676950216, + 0.0011040964163839817, + 0.0013308088527992368, + 0.001525900443084538, + 0.001693147700279951, + 0.0017897904617711902, + 0.0012339908862486482, + 0.0011671327520161867, + 0.0009541819454170763, + 0.0011125531746074557, + 0.0015531315002590418, + 0.0022527959663420916, + 0.0006792424828745425, + 0.0007884753867983818, + 0.0007301049772650003, + 0.0007282066508196294, + 0.0004898566985502839, + 0.0006500704912468791, + 0.000608527974691242, + 0.0008478111121803522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0040932451374828815 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040893275290727615 + ], + "opposite_logprob_tokens": [ + -5.50408935546875 + ], + "logprob_reward": -0.0040893275290727615, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040893275290727615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7087045311927795, + 0.017766909673810005, + 0.03695584461092949, + 0.043783120810985565, + 0.09812267124652863, + 0.04308701306581497, + 0.019839845597743988, + 0.010090518742799759, + 0.01089994702488184, + 0.010749579407274723 + ], + "credit_map": [ + 0.019717436283826828, + 0.019033318385481834, + 0.008899176493287086, + 0.03015386126935482, + 0.022510627284646034, + 0.23606140911579132, + 0.04551965370774269, + 0.32680901885032654, + 0.0014950634213164449, + 0.0018179489998146892, + 0.0022526271641254425, + 0.002165304496884346, + 0.004217599518597126, + 0.0017615382093936205, + 0.0015385481528937817, + 0.0025182811077684164, + 0.006070266477763653, + 0.0048186699859797955, + 0.0032075203489512205, + 0.002542054047808051, + 0.0032260981388390064, + 0.010494913905858994, + 0.0024441685527563095, + 0.004152156878262758, + 0.005397575907409191, + 0.003964540082961321, + 0.007211368065327406, + 0.005186513997614384, + 0.0036565077025443316, + 0.006874397862702608, + 0.0029826813843101263, + 0.008509538136422634, + 0.022432509809732437, + 0.015532434917986393, + 0.01833098754286766, + 0.008832602761685848, + 0.00742202578112483, + 0.007173675578087568, + 0.007963730953633785, + 0.010434702970087528, + 0.0055411080829799175, + 0.004396109841763973, + 0.010982486419379711, + 0.007144138682633638, + 0.0036556802224367857, + 0.0027989305090159178, + 0.0029721204191446304, + 0.005596444010734558, + 0.0038746835198253393, + 0.002162148943170905, + 0.003913454711437225, + 0.0033724855165928602, + 0.0013457464519888163, + 0.0018425136804580688, + 0.0016294249799102545, + 0.0016993907047435641, + 0.0012124537024646997, + 0.0009323387057520449, + 0.0013632879126816988, + 0.0010418685851618648, + 0.0008999958517961204, + 0.0014568555634468794, + 0.0015268194256350398, + 0.0016569001600146294, + 0.0015317624201998115, + 0.002011120319366455, + 0.0010748780332505703, + 0.0012314869090914726, + 0.0009246387635357678, + 0.0007932573207654059, + 0.001284157857298851, + 0.002048645867034793, + 0.0012358305975794792, + 0.0021904511377215385, + 0.0010470558190718293, + 0.0014406819827854633, + 0.0010398422600701451, + 0.0012752782786265016, + 0.0012842180440202355, + 0.0012362223351374269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0040893275290727615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074034497141838 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.011074034497141838, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074034497141838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8245577216148376, + 0.01470749918371439, + 0.040123384445905685, + 0.01622135564684868, + 0.027154942974448204, + 0.027393778786063194, + 0.018845267593860626, + 0.006887112744152546, + 0.013455189764499664, + 0.010653767734766006 + ], + "credit_map": [ + 0.02334093116223812, + 0.02224162593483925, + 0.010073328390717506, + 0.03503942862153053, + 0.02629493735730648, + 0.274667888879776, + 0.052629634737968445, + 0.3802699148654938, + 0.0013070284621790051, + 0.0020257094874978065, + 0.001397209125570953, + 0.0013186887372285128, + 0.002419322030618787, + 0.0014679308515042067, + 0.0016701743006706238, + 0.0031014366541057825, + 0.006064190529286861, + 0.00591428205370903, + 0.0028643966652452946, + 0.0025748431216925383, + 0.0036283668596297503, + 0.011728650890290737, + 0.0026206683833152056, + 0.00472798990085721, + 0.0018459769198670983, + 0.002373042516410351, + 0.0020156968384981155, + 0.002632410265505314, + 0.0017495565116405487, + 0.0017776323948055506, + 0.0013757632113993168, + 0.002451278269290924, + 0.00588040379807353, + 0.003695750841870904, + 0.004602761473506689, + 0.0022682338021695614, + 0.0022554085589945316, + 0.002947969827800989, + 0.0023584766313433647, + 0.003145938040688634, + 0.0035300450399518013, + 0.0029246662743389606, + 0.007146851625293493, + 0.003152084769681096, + 0.0025234841741621494, + 0.001816281583160162, + 0.0026613182853907347, + 0.003639048431068659, + 0.004063046537339687, + 0.0019265500595793128, + 0.0032363904174417257, + 0.0024259397760033607, + 0.0012731663882732391, + 0.0016748224152252078, + 0.002293516881763935, + 0.0019518382614478469, + 0.0011566145112738013, + 0.0005530901835300028, + 0.0009425700409337878, + 0.0006861791480332613, + 0.000687959895003587, + 0.0006296731880865991, + 0.0010994046460837126, + 0.0011316211894154549, + 0.0015581488842144608, + 0.0019947541877627373, + 0.0020983614958822727, + 0.0014180002035573125, + 0.0011514712823554873, + 0.001005347352474928, + 0.001477943267673254, + 0.002751163672655821, + 0.0010201819241046906, + 0.0023123561404645443, + 0.0010115867480635643, + 0.0015792710473760962, + 0.0010737284319475293, + 0.0013576801866292953, + 0.0012128927046433091, + 0.00108607136644423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.011074034497141838 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatically dynamic Snorri camera lens pinned a mysterious figure to the very center of the frame, creating a dizzying sensation of a stable subject amidst the world's turbulent motion. This figure, outlined in a form-fitting suit of dark armor gleaming with metallic and keratin, was enveloped in an ominous crimson glow of energy, like burning blood, a light that even stirred tiny arcs of electricity in the air around him. He tore through the air with thunderous force, slashing vertically from the sky. His body tensed under the immense acceleration of gravity, his grip on a massive blade composed of pure red light energy slicing a deadly arc through the air, poised to deliver a devastating blow. Around him, the world dissolved into a chaotic motion blur, gray clouds, sparse raindrops, and distant city lights stretched into countless vertical, rapidly streaking streaks of light and shadow, creating a stark visual contrast with the steady, incandescent glow of his body. This high-speed fall produces a strong cinematic light effect. The red halo creates glare and edge distortion on the lens, and the particles in the air are illuminated. The entire picture is filled with the destructive power of doomsday and a suffocating atmosphere of tension.", + "seed": 30331, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_5/index_6975_seed_30331_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06146688172593713, + "vlm_logprob": -0.06146688172593713, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.06146688172593713, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.282480925321579 + ], + "opposite_logprob_tokens": [ + -1.4074809551239014 + ], + "logprob_reward": -0.282480925321579, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.282480925321579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19315220415592194, + 0.0748962014913559, + 0.11951795220375061, + 0.09216851741075516, + 0.10982032865285873, + 0.036511342972517014, + 0.11048848927021027, + 0.052858591079711914, + 0.03918534889817238, + 0.1714010387659073 + ], + "credit_map": [ + 0.024894388392567635, + 0.022085631266236305, + 0.028094103559851646, + 0.020763540640473366, + 0.020564578473567963, + 0.03424983471632004, + 0.01838444173336029, + 0.02411569096148014, + 0.010402614250779152, + 0.00840090587735176, + 0.007510116323828697, + 0.00893454346805811, + 0.008003630675375462, + 0.008082442916929722, + 0.008091539144515991, + 0.015470415353775024, + 0.0157940611243248, + 0.011563248001039028, + 0.0129396328702569, + 0.013255120255053043, + 0.010534040629863739, + 0.017634933814406395, + 0.010634109377861023, + 0.02716279961168766, + 0.008914535865187645, + 0.009402121417224407, + 0.010778102092444897, + 0.012294386513531208, + 0.009584462270140648, + 0.01571214571595192, + 0.016064614057540894, + 0.00941814761608839, + 0.013273543678224087, + 0.012068620882928371, + 0.013135090470314026, + 0.017516497522592545, + 0.011280905455350876, + 0.013180644251406193, + 0.01338567491620779, + 0.0159793458878994, + 0.003478247206658125, + 0.003592455293983221, + 0.00480624521151185, + 0.004671191330999136, + 0.0028444617055356503, + 0.0054545472376048565, + 0.0037037290167063475, + 0.007960463874042034, + 0.009725991636514664, + 0.010885991156101227, + 0.012011080980300903, + 0.018862351775169373, + 0.00832026544958353, + 0.01261398009955883, + 0.017318151891231537, + 0.020750675350427628, + 0.009451642632484436, + 0.00400265958160162, + 0.0049651917070150375, + 0.004881080240011215, + 0.004407181870192289, + 0.006791748572140932, + 0.008434546180069447, + 0.009924541227519512, + 0.005924041382968426, + 0.004379807505756617, + 0.005780236329883337, + 0.0038832398131489754, + 0.004568639677017927, + 0.0038184381555765867, + 0.005456774495542049, + 0.0053741722367703915, + 0.007765316404402256, + 0.003093836596235633, + 0.004790314007550478, + 0.00889484491199255, + 0.00305192731320858, + 0.07269075512886047, + 0.0032541584223508835, + 0.06785987317562103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.282480925321579 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00119304982945323 + ], + "opposite_logprob_tokens": [ + -6.751193046569824 + ], + "logprob_reward": -0.00119304982945323, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.00119304982945323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1243995651602745, + 0.06659550964832306, + 0.06669190526008606, + 0.07582323998212814, + 0.11416028439998627, + 0.03251674398779869, + 0.22857849299907684, + 0.041151367127895355, + 0.050001196563243866, + 0.20008166134357452 + ], + "credit_map": [ + 0.016091765835881233, + 0.01819767616689205, + 0.0160206388682127, + 0.011306867003440857, + 0.01180836372077465, + 0.02587861567735672, + 0.010720934718847275, + 0.01437469944357872, + 0.013121331110596657, + 0.009035926312208176, + 0.008337967097759247, + 0.005982876289635897, + 0.005510603077709675, + 0.008257022127509117, + 0.0053943549282848835, + 0.010955428704619408, + 0.006665891036391258, + 0.007098231930285692, + 0.0065614222548902035, + 0.007819383405148983, + 0.007105754688382149, + 0.010254085063934326, + 0.008109021931886673, + 0.013078116811811924, + 0.009096711874008179, + 0.007244702894240618, + 0.008387153968214989, + 0.010617942549288273, + 0.007156980689615011, + 0.01337744202464819, + 0.01241200789809227, + 0.007530297618359327, + 0.013728613033890724, + 0.013178004883229733, + 0.014993191696703434, + 0.01683216169476509, + 0.011029641143977642, + 0.013857826590538025, + 0.013737129047513008, + 0.016803719103336334, + 0.003374717431142926, + 0.0036751353181898594, + 0.00330067565664649, + 0.004006199073046446, + 0.0028247148729860783, + 0.004531611688435078, + 0.003497024066746235, + 0.0073066651821136475, + 0.01640566997230053, + 0.023637589067220688, + 0.027354393154382706, + 0.04743875190615654, + 0.018833665177226067, + 0.022784367203712463, + 0.04206099733710289, + 0.03006305918097496, + 0.0076344432309269905, + 0.003265416482463479, + 0.004242873750627041, + 0.0031571723520755768, + 0.003569863736629486, + 0.006086254958063364, + 0.004435668233782053, + 0.008759673684835434, + 0.007306015118956566, + 0.005316944792866707, + 0.006471836939454079, + 0.004943241830915213, + 0.006625436712056398, + 0.005085818003863096, + 0.006948073860257864, + 0.007303830236196518, + 0.009552528150379658, + 0.004316241480410099, + 0.005496331490576267, + 0.010486261919140816, + 0.0043113077990710735, + 0.08342108130455017, + 0.00474364310503006, + 0.07775426656007767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.00119304982945323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537590228021145 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537590228021145, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537590228021145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17172548174858093, + 0.0660485327243805, + 0.17257748544216156, + 0.09152311831712723, + 0.09038767218589783, + 0.020238639786839485, + 0.21241137385368347, + 0.05659734085202217, + 0.03685334324836731, + 0.08163704723119736 + ], + "credit_map": [ + 0.020348649471998215, + 0.021823972463607788, + 0.02356811985373497, + 0.02058015950024128, + 0.021476659923791885, + 0.02568332478404045, + 0.018248198553919792, + 0.019996386021375656, + 0.008719847537577152, + 0.007075731176882982, + 0.0077683283016085625, + 0.007276520598679781, + 0.006608051247894764, + 0.008794530294835567, + 0.007590143941342831, + 0.01221537683159113, + 0.021026788279414177, + 0.01761278323829174, + 0.016538023948669434, + 0.01799316518008709, + 0.020233536139130592, + 0.026052283123135567, + 0.018773742020130157, + 0.034347157925367355, + 0.012314263731241226, + 0.009343578480184078, + 0.00960925966501236, + 0.013302426785230637, + 0.01024892833083868, + 0.01148150209337473, + 0.015285968780517578, + 0.009937189519405365, + 0.01002575270831585, + 0.008367745205760002, + 0.01183701679110527, + 0.011496962048113346, + 0.013756419532001019, + 0.010755648836493492, + 0.010305558331310749, + 0.013842567801475525, + 0.002030561910942197, + 0.0022657865192741156, + 0.002704316284507513, + 0.002900433726608753, + 0.0020192249212414026, + 0.002922248560935259, + 0.0022860351018607616, + 0.0031100341584533453, + 0.013716774992644787, + 0.022875413298606873, + 0.025885503739118576, + 0.049546096473932266, + 0.017611101269721985, + 0.019208557903766632, + 0.04192456603050232, + 0.021643375977873802, + 0.009060073643922806, + 0.0045477780513465405, + 0.006244254298508167, + 0.005625746212899685, + 0.005255349911749363, + 0.007606589235365391, + 0.009282249957323074, + 0.008975301869213581, + 0.005190450698137283, + 0.0031790658831596375, + 0.005765060428529978, + 0.0042275856249034405, + 0.0049092695116996765, + 0.0038053281605243683, + 0.005027791950851679, + 0.004748791456222534, + 0.004595817998051643, + 0.0019389160443097353, + 0.0029684891924262047, + 0.004122848156839609, + 0.0017475479980930686, + 0.03332141414284706, + 0.002416351344436407, + 0.030525662004947662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.012537590228021145 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002483856864273548 + ], + "opposite_logprob_tokens": [ + -6.00248384475708 + ], + "logprob_reward": -0.002483856864273548, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002483856864273548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20502133667469025, + 0.06343915313482285, + 0.08220985531806946, + 0.058079320937395096, + 0.10328502207994461, + 0.02726755104959011, + 0.05019945278763771, + 0.028996605426073074, + 0.024414677172899246, + 0.3570869565010071 + ], + "credit_map": [ + 0.022051909938454628, + 0.03476203978061676, + 0.02406243234872818, + 0.021198179572820663, + 0.019533533602952957, + 0.03972272947430611, + 0.01782955415546894, + 0.025860978290438652, + 0.011186004616320133, + 0.007792187388986349, + 0.007932481355965137, + 0.005906302947551012, + 0.005494574084877968, + 0.006943541578948498, + 0.006399073172360659, + 0.011784999631345272, + 0.008916209451854229, + 0.0076097724959254265, + 0.008494271896779537, + 0.009382029995322227, + 0.007836183533072472, + 0.012984983623027802, + 0.009237311780452728, + 0.01774909906089306, + 0.0081713255494833, + 0.006864895112812519, + 0.005858836695551872, + 0.006783288903534412, + 0.005520238541066647, + 0.007729046046733856, + 0.00892214197665453, + 0.008229550905525684, + 0.014404145069420338, + 0.010124703869223595, + 0.012067338451743126, + 0.012257032096385956, + 0.011637009680271149, + 0.011767319403588772, + 0.01345270685851574, + 0.01757478341460228, + 0.002558676525950432, + 0.003604505443945527, + 0.003183698048815131, + 0.0036749744322150946, + 0.0029260534793138504, + 0.0037474767304956913, + 0.0031049782410264015, + 0.004467190708965063, + 0.005280258134007454, + 0.00402523810043931, + 0.005671713501214981, + 0.006232100538909435, + 0.0042930664494633675, + 0.007000866811722517, + 0.006457929499447346, + 0.011238285340368748, + 0.006346987560391426, + 0.002354685217142105, + 0.002755154622718692, + 0.0026787323877215385, + 0.002374502830207348, + 0.004527303855866194, + 0.0032837102189660072, + 0.004675531759858131, + 0.0037969020195305347, + 0.0021095070987939835, + 0.0039556436240673065, + 0.0026600316632539034, + 0.0024294378235936165, + 0.0034556598402559757, + 0.0034115517046302557, + 0.0025959450285881758, + 0.016771448776125908, + 0.006591582205146551, + 0.009839138947427273, + 0.018262110650539398, + 0.005846103653311729, + 0.15138208866119385, + 0.006915023550391197, + 0.1414794921875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.002483856864273548 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638986386358738 + ], + "opposite_logprob_tokens": [ + -4.758638858795166 + ], + "logprob_reward": -0.008638986386358738, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638986386358738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1040349155664444, + 0.07203949242830276, + 0.07254377752542496, + 0.04644107073545456, + 0.07610960304737091, + 0.04452740401029587, + 0.08747106045484543, + 0.07110073417425156, + 0.042449917644262314, + 0.38328200578689575 + ], + "credit_map": [ + 0.014841876924037933, + 0.012822523713111877, + 0.019810650497674942, + 0.008289076387882233, + 0.008460181765258312, + 0.018917370587587357, + 0.00747784785926342, + 0.013415384106338024, + 0.009511442855000496, + 0.006792434956878424, + 0.007782766129821539, + 0.009095586836338043, + 0.0058930059894919395, + 0.008501824922859669, + 0.008184387348592281, + 0.016278043389320374, + 0.005833800882101059, + 0.007029946427792311, + 0.010402741841971874, + 0.011071967892348766, + 0.009335736744105816, + 0.009634500369429588, + 0.007546713575720787, + 0.011688372120261192, + 0.007673872634768486, + 0.004517910536378622, + 0.00529932975769043, + 0.005874409805983305, + 0.003910491243004799, + 0.005439306143671274, + 0.006597386673092842, + 0.0071283611468970776, + 0.01036855485290289, + 0.007177452091127634, + 0.00910633523017168, + 0.011014563031494617, + 0.00831683911383152, + 0.008398698642849922, + 0.010334604419767857, + 0.011392554268240929, + 0.003743771929293871, + 0.004930779337882996, + 0.004416393116116524, + 0.005369465332478285, + 0.005439662374556065, + 0.006313755642622709, + 0.006077516358345747, + 0.008236059918999672, + 0.01008380576968193, + 0.007300334982573986, + 0.009192848578095436, + 0.009018736891448498, + 0.00632046302780509, + 0.013645004481077194, + 0.007574680261313915, + 0.02433517947793007, + 0.012505211867392063, + 0.005116209387779236, + 0.006758404430001974, + 0.005921688862144947, + 0.0055749439634382725, + 0.013229669071733952, + 0.005497483536601067, + 0.016497120261192322, + 0.006418288219720125, + 0.004570811055600643, + 0.006823429837822914, + 0.004465158563107252, + 0.003247086191549897, + 0.0060826316475868225, + 0.0038519466761499643, + 0.0069905659183859825, + 0.01811402477324009, + 0.007388757541775703, + 0.010573891922831535, + 0.01913156360387802, + 0.006119951140135527, + 0.1624675691127777, + 0.0073771304450929165, + 0.15210910141468048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.008638986386358738 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatically dynamic Snorri camera lens pinned a mysterious figure to the very center of the frame, creating a dizzying sensation of a stable subject amidst the world's turbulent motion. This figure, outlined in a form-fitting suit of dark armor gleaming with metallic and keratin, was enveloped in an ominous crimson glow of energy, like burning blood, a light that even stirred tiny arcs of electricity in the air around him. He tore through the air with thunderous force, slashing vertically from the sky. His body tensed under the immense acceleration of gravity, his grip on a massive blade composed of pure red light energy slicing a deadly arc through the air, poised to deliver a devastating blow. Around him, the world dissolved into a chaotic motion blur, gray clouds, sparse raindrops, and distant city lights stretched into countless vertical, rapidly streaking streaks of light and shadow, creating a stark visual contrast with the steady, incandescent glow of his body. This high-speed fall produces a strong cinematic light effect. The red halo creates glare and edge distortion on the lens, and the particles in the air are illuminated. The entire picture is filled with the destructive power of doomsday and a suffocating atmosphere of tension.", + "seed": 30332, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_5/index_6975_seed_30332_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.019418125064112245, + "vlm_logprob": -0.019418125064112245, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.019418125064112245, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07043764740228653 + ], + "opposite_logprob_tokens": [ + -2.6954376697540283 + ], + "logprob_reward": -0.07043764740228653, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07043764740228653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10166531801223755, + 0.04708648473024368, + 0.06232929602265358, + 0.03968716040253639, + 0.11392293125391006, + 0.2522624433040619, + 0.15428240597248077, + 0.08070439100265503, + 0.03172087296843529, + 0.11633867025375366 + ], + "credit_map": [ + 0.013801377266645432, + 0.01492980495095253, + 0.012674937956035137, + 0.016172613948583603, + 0.010464021004736423, + 0.008160867728292942, + 0.01235349103808403, + 0.013108192011713982, + 0.005263308994472027, + 0.006139859091490507, + 0.006190006621181965, + 0.006110018584877253, + 0.0065694586373865604, + 0.006322594825178385, + 0.004105481319129467, + 0.006385750137269497, + 0.007054350338876247, + 0.005068149883300066, + 0.0060791256837546825, + 0.006053218152374029, + 0.005886418279260397, + 0.01268081460148096, + 0.0051912120543420315, + 0.01431600097566843, + 0.004508941434323788, + 0.006267677526921034, + 0.005847495514899492, + 0.0031586773693561554, + 0.005022455006837845, + 0.004257191903889179, + 0.004493887536227703, + 0.0061308289878070354, + 0.014654165133833885, + 0.014782858081161976, + 0.017253652215003967, + 0.012990483082830906, + 0.010553929954767227, + 0.009660622105002403, + 0.014864040538668633, + 0.019163163378834724, + 0.015841078013181686, + 0.017488714307546616, + 0.041117459535598755, + 0.03720618039369583, + 0.027461450546979904, + 0.032784052193164825, + 0.0544709637761116, + 0.025892529636621475, + 0.02669585682451725, + 0.022356219589710236, + 0.027529790997505188, + 0.019312413409352303, + 0.011115605011582375, + 0.017955007031559944, + 0.01462821289896965, + 0.014689295552670956, + 0.01551473792642355, + 0.00806707888841629, + 0.008888023905456066, + 0.008187109604477882, + 0.007797840051352978, + 0.013329893350601196, + 0.009206641465425491, + 0.009713055565953255, + 0.00292580877430737, + 0.004456435330212116, + 0.004600940737873316, + 0.003333301981911063, + 0.0033857885282486677, + 0.0038552286569029093, + 0.004582885652780533, + 0.0045804791152477264, + 0.016165515407919884, + 0.022055519744753838, + 0.011545227840542793, + 0.015138085000216961, + 0.010321873240172863, + 0.013032387010753155, + 0.012743327766656876, + 0.01533672958612442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.07043764740228653 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015200738562271 + ], + "opposite_logprob_tokens": [ + -6.501520156860352 + ], + "logprob_reward": -0.0015200738562271, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015200738562271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15306688845157623, + 0.0735141709446907, + 0.05356905981898308, + 0.1063975915312767, + 0.10160854458808899, + 0.15723179280757904, + 0.0945642963051796, + 0.06942224502563477, + 0.1308780461549759, + 0.05974733829498291 + ], + "credit_map": [ + 0.018562141805887222, + 0.017956886440515518, + 0.017464498057961464, + 0.01938219740986824, + 0.016537673771381378, + 0.014221281744539738, + 0.017619024962186813, + 0.03132318705320358, + 0.0075346603989601135, + 0.00830203015357256, + 0.006090171635150909, + 0.009034006856381893, + 0.012994731776416302, + 0.01061951369047165, + 0.008428898639976978, + 0.010510163381695747, + 0.009648018516600132, + 0.005053190514445305, + 0.010031730867922306, + 0.005326429381966591, + 0.005070151295512915, + 0.0054516661912202835, + 0.006022993009537458, + 0.006964887026697397, + 0.011573996394872665, + 0.015143024735152721, + 0.013964461162686348, + 0.009113768115639687, + 0.011219456791877747, + 0.012968943454325199, + 0.011932475492358208, + 0.020481480285525322, + 0.012380479834973812, + 0.011376080103218555, + 0.01486921962350607, + 0.01190890558063984, + 0.00804163608700037, + 0.008003060705959797, + 0.013737942092120647, + 0.021291235461831093, + 0.012571780942380428, + 0.010078946128487587, + 0.03690061345696449, + 0.021746790036559105, + 0.012307308614253998, + 0.026340855285525322, + 0.023034773766994476, + 0.014250739477574825, + 0.014281041920185089, + 0.014256441034376621, + 0.01999971643090248, + 0.011819334700703621, + 0.0070568351075053215, + 0.009207997471094131, + 0.008603807538747787, + 0.009339124895632267, + 0.008966820314526558, + 0.010549901984632015, + 0.010156426578760147, + 0.007982230745255947, + 0.007043925113976002, + 0.008438640274107456, + 0.008590158075094223, + 0.007694150321185589, + 0.014116350561380386, + 0.01836300827562809, + 0.017668772488832474, + 0.011440393514931202, + 0.012962873093783855, + 0.029180284589529037, + 0.011902321130037308, + 0.015244047157466412, + 0.005298116244375706, + 0.007162721361964941, + 0.005982826463878155, + 0.010320104658603668, + 0.007581106852740049, + 0.007446624338626862, + 0.007369173225015402, + 0.008586669340729713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.0015200738562271 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538060545921326 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012538060545921326, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538060545921326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11620187759399414, + 0.060188375413417816, + 0.07756966352462769, + 0.049311257898807526, + 0.08696111291646957, + 0.1865328997373581, + 0.15517090260982513, + 0.05330180376768112, + 0.16963571310043335, + 0.045126352459192276 + ], + "credit_map": [ + 0.01231854036450386, + 0.015139492228627205, + 0.013559091836214066, + 0.015005873516201973, + 0.01262569334357977, + 0.010039342567324638, + 0.012283457443118095, + 0.025230396538972855, + 0.0073551912792027, + 0.0074177575297653675, + 0.00532509433105588, + 0.007045125123113394, + 0.00792788341641426, + 0.008307497948408127, + 0.006338447332382202, + 0.010471380315721035, + 0.01056757289916277, + 0.007922882214188576, + 0.00987790897488594, + 0.009018249809741974, + 0.007687253877520561, + 0.011094816960394382, + 0.008066484704613686, + 0.013334494084119797, + 0.00527590187266469, + 0.006361363455653191, + 0.00438782898709178, + 0.003522230312228203, + 0.005225644912570715, + 0.006614333018660545, + 0.007718415930867195, + 0.010205544531345367, + 0.011454460211098194, + 0.010770238004624844, + 0.014507243409752846, + 0.010217481292784214, + 0.007582369260489941, + 0.006859188433736563, + 0.010800434276461601, + 0.014769701287150383, + 0.014929167926311493, + 0.010877643711864948, + 0.040553558617830276, + 0.01643698289990425, + 0.016405267640948296, + 0.03882894665002823, + 0.017562301829457283, + 0.030939040705561638, + 0.04209564998745918, + 0.040134135633707047, + 0.02091236598789692, + 0.01246351096779108, + 0.008720390498638153, + 0.011751408688724041, + 0.009121039882302284, + 0.009972402825951576, + 0.006432915572077036, + 0.008396778255701065, + 0.008988973684608936, + 0.007002546451985836, + 0.005093952640891075, + 0.00743124121800065, + 0.005108967889100313, + 0.00484642805531621, + 0.021959519013762474, + 0.023717856034636497, + 0.02285972610116005, + 0.015374278649687767, + 0.014543280936777592, + 0.03771290183067322, + 0.013419346883893013, + 0.02004881761968136, + 0.006594178266823292, + 0.006498435512185097, + 0.006730835884809494, + 0.004423520993441343, + 0.007252211216837168, + 0.005395715590566397, + 0.0034062170889228582, + 0.00482524000108242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.012538060545921326 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002814144128933549 + ], + "opposite_logprob_tokens": [ + -5.877814292907715 + ], + "logprob_reward": -0.002814144128933549, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002814144128933549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1583537459373474, + 0.06685581058263779, + 0.04276871681213379, + 0.06601697206497192, + 0.034938257187604904, + 0.24576763808727264, + 0.19473165273666382, + 0.06585569679737091, + 0.08338303864002228, + 0.041328489780426025 + ], + "credit_map": [ + 0.0185081847012043, + 0.020928483456373215, + 0.016103800386190414, + 0.020772885531187057, + 0.01576514169573784, + 0.015213712118566036, + 0.015681607648730278, + 0.03537993133068085, + 0.006370053626596928, + 0.006868753582239151, + 0.006083390675485134, + 0.008846274577081203, + 0.011630034074187279, + 0.009239169768989086, + 0.007218563463538885, + 0.01059956569224596, + 0.005924785044044256, + 0.0032227174378931522, + 0.0047532119788229465, + 0.006210104562342167, + 0.0033578998409211636, + 0.007122994866222143, + 0.004167779814451933, + 0.008009224198758602, + 0.005876906681805849, + 0.01064665149897337, + 0.004845942836254835, + 0.004107298329472542, + 0.007920005358755589, + 0.010124261491000652, + 0.008342991583049297, + 0.01415291242301464, + 0.0060316454619169235, + 0.004115668125450611, + 0.006959867663681507, + 0.0032922723330557346, + 0.003133254824206233, + 0.0029825419187545776, + 0.004119568504393101, + 0.004303436726331711, + 0.022991055622696877, + 0.013517249375581741, + 0.07046236097812653, + 0.026021776720881462, + 0.017725447192788124, + 0.05593619495630264, + 0.02017621323466301, + 0.01893734745681286, + 0.055369842797517776, + 0.04280335083603859, + 0.03370559215545654, + 0.01425104308873415, + 0.009824373759329319, + 0.022187527269124985, + 0.008425585925579071, + 0.008164318278431892, + 0.008727961219847202, + 0.008403929881751537, + 0.013308534398674965, + 0.009597716853022575, + 0.004439440555870533, + 0.008921179920434952, + 0.005739417392760515, + 0.006717517971992493, + 0.009317023679614067, + 0.012690439820289612, + 0.013775059953331947, + 0.006660203915089369, + 0.007352028973400593, + 0.01733262836933136, + 0.007288292050361633, + 0.008967364206910133, + 0.005205562803894281, + 0.007185467053204775, + 0.005379834678024054, + 0.006148998159915209, + 0.0036262115463614464, + 0.004440451506525278, + 0.00360475922934711, + 0.005737203173339367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.002814144128933549 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a mysterious figure wearing a form-fitting dark armor suit visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure enveloped in an ominous crimson glow with visible arcs of electricity in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure holding a massive blade composed of pure red light energy and slashing vertically from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing a chaotic motion blur with gray clouds, sparse raindrops, and distant city lights stretched into vertical streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera maintain a stable subject while the world exhibits turbulent motion and vertical streaking effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780699387192726 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780699387192726, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780699387192726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1516275405883789, + 0.04481034725904465, + 0.04925663396716118, + 0.060538336634635925, + 0.1408362239599228, + 0.11966075748205185, + 0.10636887699365616, + 0.0816357433795929, + 0.136850044131279, + 0.10841544717550278 + ], + "credit_map": [ + 0.013823241926729679, + 0.01830008253455162, + 0.013165348209440708, + 0.02292937971651554, + 0.014514792710542679, + 0.024839619174599648, + 0.016121726483106613, + 0.02793336845934391, + 0.004593891557306051, + 0.004717712290585041, + 0.00407703872770071, + 0.006802630610764027, + 0.006916914600878954, + 0.0067411065101623535, + 0.00412657530978322, + 0.006834481842815876, + 0.007081786636263132, + 0.004941911902278662, + 0.008289067074656487, + 0.005342997610569, + 0.0046332585625350475, + 0.007338711526244879, + 0.004869419150054455, + 0.006759486626833677, + 0.007023579906672239, + 0.005392700433731079, + 0.010132577270269394, + 0.005074417218565941, + 0.005881577264517546, + 0.009856927208602428, + 0.005749213974922895, + 0.011427345685660839, + 0.017479995265603065, + 0.02082001231610775, + 0.0243830606341362, + 0.016473740339279175, + 0.015064375475049019, + 0.01219256967306137, + 0.020115310326218605, + 0.014307171106338501, + 0.020638059824705124, + 0.009454312734305859, + 0.026766793802380562, + 0.010868526063859463, + 0.012067783623933792, + 0.022937428206205368, + 0.008252890780568123, + 0.008674979209899902, + 0.02328767627477646, + 0.01687200553715229, + 0.016194602474570274, + 0.01306631974875927, + 0.007405181881040335, + 0.010407923720777035, + 0.009951701387763023, + 0.009183469228446484, + 0.020775670185685158, + 0.008517204783856869, + 0.008760122582316399, + 0.007131827529519796, + 0.007311403285712004, + 0.015256529673933983, + 0.007225463166832924, + 0.00665753148496151, + 0.018844125792384148, + 0.015068852342665195, + 0.0215358454734087, + 0.014402724802494049, + 0.011295432224869728, + 0.026239264756441116, + 0.01257412414997816, + 0.01688968762755394, + 0.013405323959887028, + 0.015108458697795868, + 0.01196248922497034, + 0.015014030039310455, + 0.014696596190333366, + 0.013095838017761707, + 0.012598590925335884, + 0.01253412663936615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.009780699387192726 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_6/index_tensor([2376, 2376])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_6/index_tensor([2376, 2376])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..76a83db69ebf3b09ac50004c66abf8eb058b7beb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_6/index_tensor([2376, 2376])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle macro shot, characterized by constant, subtle handheld shake, as if the camera were held by a trembling hand, focuses intensely on a classic, round, rainbow-swirled lollipop. The lollipop, with its vibrant, glossy swirls of cherry red, electric blue, and sunshine yellow, rests on a pristine, matte white surface under bright, clinical studio lighting. The action begins as the lollipop's crystal surface begins to sweat, forming tiny, shimmering beads of syrup that catch the overhead light. These beads quickly coalesce, and as the solid form dramatically liquefies, the disparate streams of color begin to chaotically interpenetrate. The lollipop collapses downward and outward, dissolving into a viscous, ever-expanding puddle. Once-separate colors now swirl together in a chaotic, hypnotic vortex, creating a miniature, turbulent galaxy of melted sugar. The relentless lens tremor, combined with the mesmerizing yet destructive shifts, creates a tense and slightly unsettling atmosphere, capturing the beautiful decay of a childhood icon.", + "seed": 23525, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_6/index_2376_seed_23525_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.32424762807786467, + "vlm_logprob": -0.32424762807786467, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.32424762807786467, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00845803041011095 + ], + "opposite_logprob_tokens": [ + -5.758458137512207 + ], + "logprob_reward": -0.00845803041011095, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00845803041011095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06243741139769554, + 0.09802764654159546, + 0.03275140002369881, + 0.1432049572467804, + 0.03577074036002159, + 0.16413365304470062, + 0.06519564986228943, + 0.22325265407562256, + 0.1344325691461563, + 0.040793322026729584 + ], + "credit_map": [ + 0.009032683447003365, + 0.005011339671909809, + 0.007892937399446964, + 0.010901370085775852, + 0.0065497239120304585, + 0.007453546859323978, + 0.008177518844604492, + 0.007418294902890921, + 0.013364840298891068, + 0.017287926748394966, + 0.008401971310377121, + 0.01368161290884018, + 0.015384242869913578, + 0.012659911997616291, + 0.008476936258375645, + 0.008770203217864037, + 0.005829880479723215, + 0.0035167280584573746, + 0.004558635409921408, + 0.00292295147664845, + 0.0038742476608604193, + 0.0037212660536170006, + 0.004791198298335075, + 0.0035364963114261627, + 0.022578904405236244, + 0.00736814783886075, + 0.024213341996073723, + 0.012968014925718307, + 0.008977113291621208, + 0.019208097830414772, + 0.0181728545576334, + 0.029718488454818726, + 0.005664368160068989, + 0.001965852687135339, + 0.004638289101421833, + 0.004130244255065918, + 0.00289347511716187, + 0.004413407761603594, + 0.0027343875262886286, + 0.009330717846751213, + 0.025371378287672997, + 0.014157415367662907, + 0.03146154433488846, + 0.020824352279305458, + 0.011578568257391453, + 0.025414736941456795, + 0.01810978353023529, + 0.017215868458151817, + 0.009067085571587086, + 0.008942783810198307, + 0.008687745779752731, + 0.007753428537398577, + 0.008596348576247692, + 0.007693022023886442, + 0.005405547562986612, + 0.009049695916473866, + 0.04715669900178909, + 0.012956652790307999, + 0.015142014250159264, + 0.011732823215425014, + 0.014958066865801811, + 0.017536504194140434, + 0.02767559140920639, + 0.07609429955482483, + 0.025913679972290993, + 0.015938187018036842, + 0.010278772562742233, + 0.01311376690864563, + 0.010526126250624657, + 0.016981810331344604, + 0.009412150830030441, + 0.03226806968450546, + 0.005619206000119448, + 0.006212771870195866, + 0.0062207006849348545, + 0.006168438121676445, + 0.004639781545847654, + 0.004861281253397465, + 0.004218251444399357, + 0.002852892503142357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.00845803041011095 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011096670292317867 + ], + "opposite_logprob_tokens": [ + -4.511096477508545 + ], + "logprob_reward": -0.011096670292317867, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011096670292317867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057317104190588, + 0.07463861256837845, + 0.11611012369394302, + 0.07251553237438202, + 0.09418375790119171, + 0.09302473068237305, + 0.05498012527823448, + 0.2770993709564209, + 0.09577301889657974, + 0.06435757875442505 + ], + "credit_map": [ + 0.016214264556765556, + 0.003284924663603306, + 0.006803908385336399, + 0.006047299597412348, + 0.0032923060934990644, + 0.00968991219997406, + 0.0036500273272395134, + 0.008334462530910969, + 0.011390200816094875, + 0.01281901728361845, + 0.0073930914513766766, + 0.008693304844200611, + 0.012461796402931213, + 0.009627220220863819, + 0.006743298377841711, + 0.005510692950338125, + 0.015679867938160896, + 0.012415788136422634, + 0.014483597129583359, + 0.012926655821502209, + 0.012741141952574253, + 0.015249907970428467, + 0.020170001313090324, + 0.012443170882761478, + 0.00909907091408968, + 0.004591790493577719, + 0.010409283451735973, + 0.008553323335945606, + 0.0054526058956980705, + 0.00907568633556366, + 0.01106859277933836, + 0.014265183359384537, + 0.015411147847771645, + 0.006183787249028683, + 0.012580962851643562, + 0.009868175722658634, + 0.007661757990717888, + 0.009387761354446411, + 0.0054405564442276955, + 0.027649616822600365, + 0.014993338845670223, + 0.0083446204662323, + 0.016631467267870903, + 0.012242020107805729, + 0.009202706627547741, + 0.010553642176091671, + 0.01178643200546503, + 0.009270509704947472, + 0.006657638121396303, + 0.007063017226755619, + 0.006931481882929802, + 0.006252100691199303, + 0.007129833102226257, + 0.00802332628518343, + 0.004815985914319754, + 0.008106749504804611, + 0.05450131744146347, + 0.01738319732248783, + 0.019280584529042244, + 0.018710387870669365, + 0.01903284713625908, + 0.022701825946569443, + 0.03448591008782387, + 0.09100330621004105, + 0.030497895553708076, + 0.0074465409852564335, + 0.008569987490773201, + 0.00818548258394003, + 0.007046179845929146, + 0.007165588438510895, + 0.004969065543264151, + 0.02189229428768158, + 0.007108220364898443, + 0.009687431156635284, + 0.007996531203389168, + 0.012730519287288189, + 0.0075595201924443245, + 0.006944980006664991, + 0.007337294053286314, + 0.004993080627173185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.011096670292317867 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900993525981903 + ], + "opposite_logprob_tokens": [ + -2.579010009765625 + ], + "logprob_reward": -0.07900993525981903, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900993525981903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023045014590024948, + 0.02199646085500717, + 0.026829788461327553, + 0.3037048280239105, + 0.02046162635087967, + 0.024285931140184402, + 0.018904419615864754, + 0.3085399270057678, + 0.04105618596076965, + 0.21117572486400604 + ], + "credit_map": [ + 0.007459804881364107, + 0.0017142691649496555, + 0.002375283744186163, + 0.0017693040426820517, + 0.0013847321970388293, + 0.0041886745020747185, + 0.001435765065252781, + 0.0027171852998435497, + 0.0035761792678385973, + 0.0016880340408533812, + 0.002054779324680567, + 0.004911594092845917, + 0.0016362069873139262, + 0.003452486591413617, + 0.0021467350888997316, + 0.0025304490700364113, + 0.004643267020583153, + 0.002042699372395873, + 0.004851241130381823, + 0.003206913359463215, + 0.0015217444160953164, + 0.003532648552209139, + 0.0032490729354321957, + 0.0037822043523192406, + 0.039241209626197815, + 0.02019241638481617, + 0.05006380379199982, + 0.03483818843960762, + 0.020897330716252327, + 0.038667500019073486, + 0.035336337983608246, + 0.06446806341409683, + 0.0031856398563832045, + 0.0012035418767482042, + 0.0027284794487059116, + 0.002696653362363577, + 0.001824383856728673, + 0.002181500429287553, + 0.0015681065851822495, + 0.005073323845863342, + 0.0035046699922531843, + 0.0021506662014871836, + 0.004017493221908808, + 0.004033450037240982, + 0.002789510879665613, + 0.0024886932224035263, + 0.0032691305968910456, + 0.002032321412116289, + 0.0022402030881494284, + 0.0019222149858251214, + 0.0031729561742395163, + 0.0033048016484826803, + 0.0020626189652830362, + 0.0023853753227740526, + 0.0015423361910507083, + 0.0022739158011972904, + 0.07089515030384064, + 0.01502588577568531, + 0.02021949738264084, + 0.011920874938368797, + 0.01985793747007847, + 0.022038135677576065, + 0.038223765790462494, + 0.11035872250795364, + 0.005337166134268045, + 0.0056913779117167, + 0.004136961419135332, + 0.0038317639846354723, + 0.003738819155842066, + 0.006983777973800898, + 0.0032143269199877977, + 0.008122000843286514, + 0.02691073901951313, + 0.03915239870548248, + 0.027593141421675682, + 0.028732310980558395, + 0.0224448312073946, + 0.023670509457588196, + 0.024944094941020012, + 0.017727725207805634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 9 + ], + "selected_reward": -0.07900993525981903 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592467859387398 + ], + "opposite_logprob_tokens": [ + -3.895592451095581 + ], + "logprob_reward": -0.020592467859387398, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592467859387398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02810530737042427, + 0.052176427096128464, + 0.04420699179172516, + 0.07371445745229721, + 0.04074535146355629, + 0.023701703175902367, + 0.048930574208498, + 0.42535459995269775, + 0.21311761438846588, + 0.04994706064462662 + ], + "credit_map": [ + 0.008613919839262962, + 0.001992351608350873, + 0.0026026626583188772, + 0.002652850467711687, + 0.0019436059519648552, + 0.004853147082030773, + 0.0019011033000424504, + 0.0035456661134958267, + 0.011412600986659527, + 0.006277819629758596, + 0.0055739376693964005, + 0.005144244991242886, + 0.006377390585839748, + 0.007290690205991268, + 0.0040535517036914825, + 0.006046189460903406, + 0.006113729905337095, + 0.006053450983017683, + 0.0047400034964084625, + 0.004889315459877253, + 0.005287675652652979, + 0.005942016839981079, + 0.006564890965819359, + 0.004615907557308674, + 0.009457233361899853, + 0.005213846452534199, + 0.010831746272742748, + 0.008293473161756992, + 0.0045660557225346565, + 0.0099728312343359, + 0.009463360533118248, + 0.01591591350734234, + 0.006500691641122103, + 0.002508326666429639, + 0.006713715847581625, + 0.004966170527040958, + 0.003836369840428233, + 0.005021488294005394, + 0.003305700607597828, + 0.007892887108027935, + 0.0036322670057415962, + 0.0018724027322605252, + 0.0024261611979454756, + 0.0027866752352565527, + 0.0034484858624637127, + 0.0030086978804320097, + 0.0037853620015084743, + 0.0027416504453867674, + 0.00735298078507185, + 0.005558468401432037, + 0.004946271888911724, + 0.010499684140086174, + 0.006071329116821289, + 0.005501287057995796, + 0.0038785389624536037, + 0.0051220133900642395, + 0.09752226620912552, + 0.02050783298909664, + 0.027677085250616074, + 0.016765369102358818, + 0.02749885432422161, + 0.03104506805539131, + 0.05180809274315834, + 0.1525300294160843, + 0.05297986790537834, + 0.022059187293052673, + 0.016153868287801743, + 0.0193936787545681, + 0.01692482829093933, + 0.023178761824965477, + 0.013893934898078442, + 0.04853349179029465, + 0.007105925120413303, + 0.007883247919380665, + 0.006472480949014425, + 0.008429239504039288, + 0.004890019074082375, + 0.005574495531618595, + 0.00569278234615922, + 0.003898868802934885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.020592467859387398 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.502081036567688 + ], + "opposite_logprob_tokens": [ + -0.2520810067653656 + ], + "logprob_reward": -1.502081036567688, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.502081036567688, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04796610027551651, + 0.07274540513753891, + 0.0571262389421463, + 0.2750852704048157, + 0.03955579176545143, + 0.027241243049502373, + 0.05548251047730446, + 0.09659571200609207, + 0.24878960847854614, + 0.07941210269927979 + ], + "credit_map": [ + 0.01342629175633192, + 0.0030913827940821648, + 0.004988369066268206, + 0.004703516140580177, + 0.003068484365940094, + 0.009654583409428596, + 0.0030858672689646482, + 0.005947601981461048, + 0.009835397824645042, + 0.009277181699872017, + 0.006721889600157738, + 0.016430040821433067, + 0.008290025405585766, + 0.010677632875740528, + 0.005400712136179209, + 0.006112533155828714, + 0.008637458086013794, + 0.007961916737258434, + 0.006678053643554449, + 0.005703139118850231, + 0.007223291322588921, + 0.0061990548856556416, + 0.008849241770803928, + 0.00587408198043704, + 0.03264973685145378, + 0.017973724752664566, + 0.04213748872280121, + 0.028876308351755142, + 0.015216591767966747, + 0.0375497043132782, + 0.039418723434209824, + 0.06126299872994423, + 0.007092433515936136, + 0.0022783430758863688, + 0.008232415653765202, + 0.005220724735409021, + 0.0026707355864346027, + 0.0036914951633661985, + 0.0031146276742219925, + 0.007255013566464186, + 0.0045506516471505165, + 0.0025999322533607483, + 0.004373863339424133, + 0.0037754143122583628, + 0.001971528632566333, + 0.0030301997903734446, + 0.003531278809532523, + 0.003408375894650817, + 0.007389930076897144, + 0.0061190384440124035, + 0.006582893896847963, + 0.011091932654380798, + 0.00638992665335536, + 0.00691866222769022, + 0.0048293378204107285, + 0.006160791497677565, + 0.02078437991440296, + 0.006232118234038353, + 0.006983257830142975, + 0.0071266707964241505, + 0.006623298395425081, + 0.007119329646229744, + 0.013593127019703388, + 0.028133530169725418, + 0.05750039592385292, + 0.028341418132185936, + 0.017765603959560394, + 0.023425979539752007, + 0.02002074010670185, + 0.028603114187717438, + 0.01574023813009262, + 0.05739210918545723, + 0.007313157897442579, + 0.011721092276275158, + 0.007525736466050148, + 0.015729380771517754, + 0.018166277557611465, + 0.007192597258836031, + 0.0072504086419939995, + 0.004513454623520374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -1.502081036567688 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "high-angle macro shot, characterized by constant, subtle handheld shake, as if the camera were held by a trembling hand, focuses intensely on a classic, round, rainbow-swirled lollipop. The lollipop, with its vibrant, glossy swirls of cherry red, electric blue, and sunshine yellow, rests on a pristine, matte white surface under bright, clinical studio lighting. The action begins as the lollipop's crystal surface begins to sweat, forming tiny, shimmering beads of syrup that catch the overhead light. These beads quickly coalesce, and as the solid form dramatically liquefies, the disparate streams of color begin to chaotically interpenetrate. The lollipop collapses downward and outward, dissolving into a viscous, ever-expanding puddle. Once-separate colors now swirl together in a chaotic, hypnotic vortex, creating a miniature, turbulent galaxy of melted sugar. The relentless lens tremor, combined with the mesmerizing yet destructive shifts, creates a tense and slightly unsettling atmosphere, capturing the beautiful decay of a childhood icon.", + "seed": 23526, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_6/index_2376_seed_23526_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.21992384055629371, + "vlm_logprob": -0.21992384055629371, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.21992384055629371, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017437132075428963 + ], + "opposite_logprob_tokens": [ + -4.892436981201172 + ], + "logprob_reward": -0.017437132075428963, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.017437132075428963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17175684869289398, + 0.1692822426557541, + 0.14675959944725037, + 0.20855726301670074, + 0.03100692667067051, + 0.08638104796409607, + 0.04882822185754776, + 0.0653972327709198, + 0.031354598701000214, + 0.040676068514585495 + ], + "credit_map": [ + 0.027274074032902718, + 0.01341587956994772, + 0.025927836075425148, + 0.012950715608894825, + 0.01435786858201027, + 0.03248370811343193, + 0.016755346208810806, + 0.028591424226760864, + 0.05658574029803276, + 0.011339868418872356, + 0.055270031094551086, + 0.0073601799085736275, + 0.00773387448862195, + 0.011258858256042004, + 0.00943413469940424, + 0.010299548506736755, + 0.027368199080228806, + 0.012188892811536789, + 0.03527729958295822, + 0.016479840502142906, + 0.011464474722743034, + 0.01719648391008377, + 0.013407093472778797, + 0.013377309776842594, + 0.029820483177900314, + 0.022348012775182724, + 0.029965698719024658, + 0.03757345303893089, + 0.01550658605992794, + 0.02111843414604664, + 0.03307623043656349, + 0.019148368388414383, + 0.0043624332174658775, + 0.0038493124302476645, + 0.00351872481405735, + 0.0034616219345480204, + 0.0032774177379906178, + 0.00429279962554574, + 0.005616662558168173, + 0.0026279515586793423, + 0.012215412221848965, + 0.007670280523598194, + 0.009225529618561268, + 0.007555885706096888, + 0.010093596763908863, + 0.016471069306135178, + 0.009297761134803295, + 0.01385150384157896, + 0.007602667436003685, + 0.005214689765125513, + 0.0047143264673650265, + 0.005660605616867542, + 0.00730129424482584, + 0.006868582684546709, + 0.005881721153855324, + 0.005584331229329109, + 0.008399905636906624, + 0.0076624006032943726, + 0.005730404052883387, + 0.007200428284704685, + 0.00667783385142684, + 0.009076387621462345, + 0.011838993057608604, + 0.008810870349407196, + 0.005429923068732023, + 0.0030410338658839464, + 0.002653746632859111, + 0.002782481489703059, + 0.0038355758879333735, + 0.004985503386706114, + 0.0026254986878484488, + 0.006000834051519632, + 0.004016305319964886, + 0.0027422348503023386, + 0.0026863724924623966, + 0.0028319789562374353, + 0.009324682876467705, + 0.011592145077884197, + 0.0030344079714268446, + 0.004447936546057463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.017437132075428963 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052719381637871265 + ], + "opposite_logprob_tokens": [ + -5.255271911621094 + ], + "logprob_reward": -0.0052719381637871265, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052719381637871265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15313462913036346, + 0.4032740294933319, + 0.06298346817493439, + 0.07981806248426437, + 0.09615711122751236, + 0.04029077664017677, + 0.01712164282798767, + 0.045514147728681564, + 0.08001887053251266, + 0.02168721705675125 + ], + "credit_map": [ + 0.024741819128394127, + 0.01816115714609623, + 0.019982794299721718, + 0.01465823408216238, + 0.016580350697040558, + 0.020200202241539955, + 0.015494422987103462, + 0.023315656930208206, + 0.15804095566272736, + 0.012747406028211117, + 0.15769526362419128, + 0.013236336410045624, + 0.01206028088927269, + 0.02130131609737873, + 0.009450465440750122, + 0.018742026761174202, + 0.007416385691612959, + 0.006635747849941254, + 0.00877573061734438, + 0.010295321233570576, + 0.005588649306446314, + 0.009746083989739418, + 0.008655736222863197, + 0.005869816057384014, + 0.010991834104061127, + 0.0072311703115701675, + 0.008021964691579342, + 0.010560503229498863, + 0.005208761431276798, + 0.01586972363293171, + 0.010387198999524117, + 0.011546903289854527, + 0.012575474567711353, + 0.009173120371997356, + 0.013253542594611645, + 0.00986434705555439, + 0.008447615429759026, + 0.013668612577021122, + 0.018883293494582176, + 0.010291105136275291, + 0.005402369424700737, + 0.0041814022697508335, + 0.00546974316239357, + 0.004186367150396109, + 0.00443678954616189, + 0.0058578974567353725, + 0.005199398845434189, + 0.005556807387620211, + 0.0027239955961704254, + 0.0017968319589272141, + 0.0017300229519605637, + 0.0020359221380203962, + 0.0017324346117675304, + 0.0031101438216865063, + 0.0018960874294862151, + 0.0020962052512913942, + 0.005611934699118137, + 0.004311758559197187, + 0.006511549465358257, + 0.005143808200955391, + 0.005070543847978115, + 0.007757002487778664, + 0.004825802985578775, + 0.006281748879700899, + 0.00816703774034977, + 0.007887199521064758, + 0.006476846057921648, + 0.005777538288384676, + 0.026019586250185966, + 0.010993321426212788, + 0.006189337000250816, + 0.008508004248142242, + 0.002808371791616082, + 0.002042367123067379, + 0.0027157755102962255, + 0.002052504103630781, + 0.003311608452349901, + 0.002943571424111724, + 0.0017825341783463955, + 0.004030483774840832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0052719381637871265 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039399802684784 + ], + "opposite_logprob_tokens": [ + -1.9103939533233643 + ], + "logprob_reward": -0.16039399802684784, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039399802684784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08811495453119278, + 0.48787233233451843, + 0.06292646378278732, + 0.18411235511302948, + 0.016623416915535927, + 0.027073636651039124, + 0.012966176494956017, + 0.03571918606758118, + 0.0637803003191948, + 0.02081110142171383 + ], + "credit_map": [ + 0.014329750090837479, + 0.0076398770324885845, + 0.01203505601733923, + 0.009018458425998688, + 0.008375563658773899, + 0.011810097843408585, + 0.009072897024452686, + 0.015833262354135513, + 0.13831308484077454, + 0.03768983855843544, + 0.12576773762702942, + 0.02708878181874752, + 0.029937149956822395, + 0.047261662781238556, + 0.027850687503814697, + 0.05396341532468796, + 0.008323898538947105, + 0.0066932933405041695, + 0.010350574739277363, + 0.008848357014358044, + 0.006031970959156752, + 0.00910999532788992, + 0.0060395305044949055, + 0.0075288512744009495, + 0.027096524834632874, + 0.018733400851488113, + 0.018802082166075706, + 0.029756125062704086, + 0.013918335549533367, + 0.02861851081252098, + 0.03046393394470215, + 0.016723476350307465, + 0.0019368643406778574, + 0.0015687898267060518, + 0.002529669785872102, + 0.0018993110861629248, + 0.002236337633803487, + 0.001977775478735566, + 0.002715960843488574, + 0.0017587102483958006, + 0.005486642010509968, + 0.001741899293847382, + 0.0024914517998695374, + 0.001571315573528409, + 0.0015703735407441854, + 0.0036434524226933718, + 0.002487767953425646, + 0.008080738596618176, + 0.0013536777114495635, + 0.0012635744642466307, + 0.0014594403328374028, + 0.001997987274080515, + 0.0011080042459070683, + 0.001580032054334879, + 0.001875721849501133, + 0.0023277406580746174, + 0.006529824808239937, + 0.004126535262912512, + 0.004636076744645834, + 0.003510689362883568, + 0.00482712360098958, + 0.004317735321819782, + 0.003415534505620599, + 0.0043556722812354565, + 0.006777418777346611, + 0.005136086605489254, + 0.005238597746938467, + 0.005044011399149895, + 0.011697421781718731, + 0.005920934025198221, + 0.016923170536756516, + 0.007042663637548685, + 0.0022476522717624903, + 0.0017582625150680542, + 0.002308906987309456, + 0.00243227556347847, + 0.004488139413297176, + 0.003194191260263324, + 0.0020303025376051664, + 0.002351372269913554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.16039399802684784 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819116808474064 + ], + "opposite_logprob_tokens": [ + -4.018191337585449 + ], + "logprob_reward": -0.01819116808474064, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819116808474064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0641278401017189, + 0.30458763241767883, + 0.1084696501493454, + 0.15168942511081696, + 0.02990729548037052, + 0.04646832123398781, + 0.021602973341941833, + 0.06721511483192444, + 0.14678090810775757, + 0.05915084853768349 + ], + "credit_map": [ + 0.010649505071341991, + 0.0048688855022192, + 0.010504753328859806, + 0.005562610924243927, + 0.005752401892095804, + 0.008051379583775997, + 0.007208732422441244, + 0.011529569514095783, + 0.09303422272205353, + 0.017574390396475792, + 0.08986982703208923, + 0.0144289406016469, + 0.01617596112191677, + 0.029073728248476982, + 0.01327514462172985, + 0.031155411154031754, + 0.01718842424452305, + 0.00869839545339346, + 0.022826673462986946, + 0.017197247594594955, + 0.007565953768789768, + 0.011332572437822819, + 0.012750040739774704, + 0.010910331271588802, + 0.02484937570989132, + 0.010918165557086468, + 0.021912338212132454, + 0.011976350098848343, + 0.010882062837481499, + 0.03855535760521889, + 0.012628335505723953, + 0.019967442378401756, + 0.003494337433949113, + 0.0034664517734199762, + 0.005432847421616316, + 0.0033966158516705036, + 0.0034805056639015675, + 0.0029940833337605, + 0.0052414522506296635, + 0.002401001052930951, + 0.006203705444931984, + 0.004609104245901108, + 0.006589507218450308, + 0.004697625525295734, + 0.003993985243141651, + 0.007354108616709709, + 0.004821644630283117, + 0.008198639377951622, + 0.0023299551103264093, + 0.002742585027590394, + 0.0019658494275063276, + 0.0023889965377748013, + 0.002116083400323987, + 0.004152688197791576, + 0.002649318892508745, + 0.003257496515288949, + 0.007918308489024639, + 0.008462808094918728, + 0.00722583569586277, + 0.0066808839328587055, + 0.006082233041524887, + 0.008462251164019108, + 0.011640941724181175, + 0.010741854086518288, + 0.017774729058146477, + 0.016006801277399063, + 0.01262615155428648, + 0.011962806805968285, + 0.03356979042291641, + 0.01998836174607277, + 0.016391746699810028, + 0.01846051961183548, + 0.0060618300922214985, + 0.005449850577861071, + 0.005581790115684271, + 0.005669846665114164, + 0.008478348143398762, + 0.013479026034474373, + 0.004401135724037886, + 0.010029017925262451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.01819116808474064 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983249664306641 + ], + "opposite_logprob_tokens": [ + -0.5233249664306641 + ], + "logprob_reward": -0.8983249664306641, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983249664306641, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07304774969816208, + 0.17064636945724487, + 0.057356253266334534, + 0.4905052185058594, + 0.017352133989334106, + 0.07715467363595963, + 0.01401611790060997, + 0.03919806703925133, + 0.044140443205833435, + 0.016582921147346497 + ], + "credit_map": [ + 0.013177908957004547, + 0.009129360318183899, + 0.009817111305892467, + 0.00876399502158165, + 0.007771138567477465, + 0.009507612325251102, + 0.006734557915478945, + 0.008146057836711407, + 0.053874772042036057, + 0.009141986258327961, + 0.05949557572603226, + 0.0073812915943562984, + 0.0066914684139192104, + 0.01528905238956213, + 0.008057807572185993, + 0.01071441825479269, + 0.006218580529093742, + 0.005680236965417862, + 0.006985013838857412, + 0.01074996218085289, + 0.006318966392427683, + 0.007722542155534029, + 0.00791832897812128, + 0.0057626268826425076, + 0.06973869353532791, + 0.04814799129962921, + 0.05717244744300842, + 0.07332446426153183, + 0.03202012926340103, + 0.07150030881166458, + 0.0701116994023323, + 0.06848946958780289, + 0.002923251362517476, + 0.0016549803549423814, + 0.0028410821687430143, + 0.0016127072740346193, + 0.002329331124201417, + 0.0022433691192418337, + 0.0017094658687710762, + 0.0020379475317895412, + 0.015250065363943577, + 0.0050762370228767395, + 0.009032187983393669, + 0.00686223478987813, + 0.00579090416431427, + 0.009480918757617474, + 0.007088594604283571, + 0.01857353001832962, + 0.002670939778909087, + 0.0013801191234961152, + 0.0013459826586768031, + 0.0012942379107698798, + 0.0016028512036427855, + 0.0022391786333173513, + 0.001717936247587204, + 0.0017648714128881693, + 0.005468726623803377, + 0.005111037753522396, + 0.003252575872465968, + 0.003697811160236597, + 0.005842373240739107, + 0.004553122445940971, + 0.0070380414836108685, + 0.004234380554407835, + 0.006285692565143108, + 0.004675380885601044, + 0.003309535328298807, + 0.003375972853973508, + 0.009820635430514812, + 0.006909701507538557, + 0.005025749560445547, + 0.0047377715818583965, + 0.0013196668587625027, + 0.0011975836241617799, + 0.0010454666335135698, + 0.0011433986946940422, + 0.003252060152590275, + 0.004037846811115742, + 0.0011804657988250256, + 0.0034064322244375944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.8983249664306641 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "high-angle macro shot, characterized by constant, subtle handheld shake, as if the camera were held by a trembling hand, focuses intensely on a classic, round, rainbow-swirled lollipop. The lollipop, with its vibrant, glossy swirls of cherry red, electric blue, and sunshine yellow, rests on a pristine, matte white surface under bright, clinical studio lighting. The action begins as the lollipop's crystal surface begins to sweat, forming tiny, shimmering beads of syrup that catch the overhead light. These beads quickly coalesce, and as the solid form dramatically liquefies, the disparate streams of color begin to chaotically interpenetrate. The lollipop collapses downward and outward, dissolving into a viscous, ever-expanding puddle. Once-separate colors now swirl together in a chaotic, hypnotic vortex, creating a miniature, turbulent galaxy of melted sugar. The relentless lens tremor, combined with the mesmerizing yet destructive shifts, creates a tense and slightly unsettling atmosphere, capturing the beautiful decay of a childhood icon.", + "seed": 23527, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_6/index_2376_seed_23527_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21964882146567105, + "vlm_logprob": -0.21964882146567105, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.21964882146567105, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013492346741259098 + ], + "opposite_logprob_tokens": [ + -5.013492584228516 + ], + "logprob_reward": -0.013492346741259098, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.013492346741259098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14412695169448853, + 0.03960622474551201, + 0.08964870125055313, + 0.04622543603181839, + 0.06415591388940811, + 0.041541289538145065, + 0.15716223418712616, + 0.15405242145061493, + 0.2294401228427887, + 0.034040726721286774 + ], + "credit_map": [ + 0.017478719353675842, + 0.01213507167994976, + 0.02637527883052826, + 0.013153529725968838, + 0.02395719289779663, + 0.019944356754422188, + 0.009563815779983997, + 0.021518968045711517, + 0.00519972899928689, + 0.004151954781264067, + 0.006941454019397497, + 0.004195252433419228, + 0.005953617859631777, + 0.004294844344258308, + 0.0037408447824418545, + 0.005128527991473675, + 0.010583936236798763, + 0.008390463888645172, + 0.01137575600296259, + 0.012418207712471485, + 0.012889984995126724, + 0.013691277243196964, + 0.009863004088401794, + 0.010436068288981915, + 0.006461137905716896, + 0.0043709720484912395, + 0.007937909103929996, + 0.006200626026839018, + 0.0052160234190523624, + 0.006145378574728966, + 0.004586078226566315, + 0.005307311192154884, + 0.008447287604212761, + 0.006198996677994728, + 0.014081260189414024, + 0.006112565286457539, + 0.00517083378508687, + 0.007831438444554806, + 0.0051931035704910755, + 0.011120429262518883, + 0.0056779878214001656, + 0.003981288056820631, + 0.007301982957869768, + 0.0031087009701877832, + 0.0034883127082139254, + 0.005543801002204418, + 0.004223322961479425, + 0.008215895853936672, + 0.05151207745075226, + 0.00884117279201746, + 0.006949837319552898, + 0.006707166787236929, + 0.007478949148207903, + 0.05966909974813461, + 0.005110323894768953, + 0.01089359913021326, + 0.018789788708090782, + 0.014736202545464039, + 0.026298759505152702, + 0.01902267336845398, + 0.019956381991505623, + 0.017960531637072563, + 0.016969677060842514, + 0.02031840570271015, + 0.02509765885770321, + 0.02199828065931797, + 0.041375745087862015, + 0.02849220670759678, + 0.021868467330932617, + 0.02231651358306408, + 0.03242644667625427, + 0.035864800214767456, + 0.0035703449975699186, + 0.0036889363545924425, + 0.004043997265398502, + 0.003983262460678816, + 0.0032957543153315783, + 0.0067348117008805275, + 0.003677157685160637, + 0.00504646310582757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.013492346741259098 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006754660978913307 + ], + "opposite_logprob_tokens": [ + -5.0067548751831055 + ], + "logprob_reward": -0.006754660978913307, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006754660978913307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13996386528015137, + 0.08580400794744492, + 0.0980873629450798, + 0.06080416589975357, + 0.07108833640813828, + 0.05124552175402641, + 0.10757706314325333, + 0.13456851243972778, + 0.16292887926101685, + 0.08793231844902039 + ], + "credit_map": [ + 0.020075557753443718, + 0.013487056829035282, + 0.021975353360176086, + 0.011624559760093689, + 0.013377857394516468, + 0.017666742205619812, + 0.018184449523687363, + 0.023572279140353203, + 0.010774660855531693, + 0.01107305008918047, + 0.011947189457714558, + 0.008040926419198513, + 0.018185695633292198, + 0.00960313156247139, + 0.0068917288444936275, + 0.009287628345191479, + 0.011926023289561272, + 0.00836251862347126, + 0.012973219156265259, + 0.01358259841799736, + 0.015100822784006596, + 0.01486942172050476, + 0.01016226876527071, + 0.011110491119325161, + 0.012348254211246967, + 0.005481318570673466, + 0.013217940926551819, + 0.005120913498103619, + 0.006393016315996647, + 0.006270550191402435, + 0.0043607959523797035, + 0.007611377164721489, + 0.012610947713255882, + 0.006605856120586395, + 0.010193000547587872, + 0.006511770188808441, + 0.005454739090055227, + 0.010164567269384861, + 0.006513431668281555, + 0.013034024275839329, + 0.00926483329385519, + 0.00534111587330699, + 0.00576427998021245, + 0.0035582403652369976, + 0.0044792406260967255, + 0.007767203729599714, + 0.00334644946269691, + 0.011724158190190792, + 0.022418225184082985, + 0.009072154760360718, + 0.011058456264436245, + 0.009912097826600075, + 0.00791506189852953, + 0.02574547752737999, + 0.009792879223823547, + 0.011662703938782215, + 0.017533008009195328, + 0.012005029246211052, + 0.022513609379529953, + 0.010743065737187862, + 0.013908807188272476, + 0.02707757242023945, + 0.010356204584240913, + 0.020431216806173325, + 0.022715723142027855, + 0.012081283144652843, + 0.03143399953842163, + 0.02144135907292366, + 0.01287960447371006, + 0.01675558090209961, + 0.020723458379507065, + 0.024897854775190353, + 0.013938181102275848, + 0.0077423760667443275, + 0.012800672091543674, + 0.011150548234581947, + 0.006362430285662413, + 0.013008073903620243, + 0.009038698859512806, + 0.013891337439417839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.006754660978913307 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033710893243551254 + ], + "opposite_logprob_tokens": [ + -3.4087109565734863 + ], + "logprob_reward": -0.033710893243551254, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033710893243551254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4034091532230377, + 0.026845403015613556, + 0.02278948947787285, + 0.019200501963496208, + 0.05024363845586777, + 0.02551967091858387, + 0.23976440727710724, + 0.1308879554271698, + 0.05389067903161049, + 0.0274491086602211 + ], + "credit_map": [ + 0.06634075939655304, + 0.03424111008644104, + 0.058272890746593475, + 0.03614483401179314, + 0.0731954574584961, + 0.041419100016355515, + 0.02879631146788597, + 0.06499869376420975, + 0.00322008621878922, + 0.00342945265583694, + 0.004197904374450445, + 0.0030024272855371237, + 0.004918909631669521, + 0.0021984484046697617, + 0.0024857171811163425, + 0.0033924588933587074, + 0.0028404018376022577, + 0.0017007086426019669, + 0.0028989166021347046, + 0.003308640792965889, + 0.0022313392255455256, + 0.002857072511687875, + 0.002878078492358327, + 0.004074333235621452, + 0.0022893394343554974, + 0.002830340526998043, + 0.0027220312040299177, + 0.0021960504818707705, + 0.0024692327715456486, + 0.0028619389049708843, + 0.0013895458541810513, + 0.002442025812342763, + 0.007632289547473192, + 0.004221880342811346, + 0.0068709952756762505, + 0.00577990198507905, + 0.0038572095800191164, + 0.008790766820311546, + 0.005018204916268587, + 0.00807239767163992, + 0.0027700632344931364, + 0.0018737063510343432, + 0.004572342149913311, + 0.001746680703945458, + 0.0022773281671106815, + 0.004008568357676268, + 0.0024354462511837482, + 0.005835536867380142, + 0.08329275250434875, + 0.011639954522252083, + 0.008244402706623077, + 0.004611383192241192, + 0.01020672358572483, + 0.10057210177183151, + 0.005078349262475967, + 0.016118738800287247, + 0.01920950785279274, + 0.010923195630311966, + 0.029677579179406166, + 0.011114099994301796, + 0.01347022783011198, + 0.02054489403963089, + 0.009632878005504608, + 0.016315585002303123, + 0.0059975143522024155, + 0.004263818729668856, + 0.010644153691828251, + 0.005853034555912018, + 0.004465767648071051, + 0.005205290392041206, + 0.008232245221734047, + 0.009228858165442944, + 0.0029421013314276934, + 0.0033053038641810417, + 0.0043839686550199986, + 0.004074768628925085, + 0.0027206435333937407, + 0.002990110544487834, + 0.003236538963392377, + 0.003795672906562686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.033710893243551254 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700267106294632 + ], + "opposite_logprob_tokens": [ + -2.6950266361236572 + ], + "logprob_reward": -0.0700267106294632, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700267106294632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16802436113357544, + 0.047691501677036285, + 0.05604344233870506, + 0.05121137574315071, + 0.0552530363202095, + 0.056574732065200806, + 0.14149093627929688, + 0.18857619166374207, + 0.12659916281700134, + 0.10853525996208191 + ], + "credit_map": [ + 0.028481487184762955, + 0.014549892395734787, + 0.02568899653851986, + 0.013408001512289047, + 0.023728329688310623, + 0.01638440228998661, + 0.016556309536099434, + 0.02922695502638817, + 0.0061822510324418545, + 0.0051663341000676155, + 0.006496192421764135, + 0.0064272466115653515, + 0.007334092631936073, + 0.00484224921092391, + 0.004536022432148457, + 0.006707110907882452, + 0.007388164289295673, + 0.007281098980456591, + 0.007916199043393135, + 0.006486634258180857, + 0.01156203355640173, + 0.005726556293666363, + 0.004535348154604435, + 0.005147406831383705, + 0.009878206998109818, + 0.004764769226312637, + 0.01130909938365221, + 0.00549697270616889, + 0.004998496267944574, + 0.004199296701699495, + 0.0045872158370912075, + 0.005977317690849304, + 0.0071992469020187855, + 0.00671023316681385, + 0.010451113805174828, + 0.005473968107253313, + 0.006334078963845968, + 0.0065094297751784325, + 0.005186678376048803, + 0.007388290949165821, + 0.011192438192665577, + 0.00909325759857893, + 0.00685798330232501, + 0.004369789268821478, + 0.004422138910740614, + 0.007722718175500631, + 0.0046448297798633575, + 0.008271578699350357, + 0.02409469522535801, + 0.013468851335346699, + 0.019567109644412994, + 0.016454895958304405, + 0.013142308220267296, + 0.019152207300066948, + 0.0163166131824255, + 0.01929425075650215, + 0.02303427830338478, + 0.01749914512038231, + 0.03142277151346207, + 0.015860823914408684, + 0.019177271053195, + 0.041160859167575836, + 0.014440891332924366, + 0.02598015032708645, + 0.018657637760043144, + 0.010079530999064445, + 0.023931194096803665, + 0.015918558463454247, + 0.00979625154286623, + 0.013127362355589867, + 0.016257472336292267, + 0.018831156194210052, + 0.01590781658887863, + 0.00930535513907671, + 0.017003871500492096, + 0.013882546685636044, + 0.007250375580042601, + 0.015658188611268997, + 0.012978757731616497, + 0.01654834672808647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.0700267106294632 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742594957351685 + ], + "opposite_logprob_tokens": [ + -0.47425949573516846 + ], + "logprob_reward": -0.9742594957351685, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742594957351685, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1433051973581314, + 0.09492628276348114, + 0.17010046541690826, + 0.048693228513002396, + 0.041103214025497437, + 0.03624769300222397, + 0.1802038550376892, + 0.0942988321185112, + 0.11607138812541962, + 0.07504988461732864 + ], + "credit_map": [ + 0.022318720817565918, + 0.01230646762996912, + 0.020892011001706123, + 0.01399984210729599, + 0.02312709577381611, + 0.017710048705339432, + 0.014075416140258312, + 0.01887558586895466, + 0.012533308006823063, + 0.010592508129775524, + 0.011545301415026188, + 0.01207135058939457, + 0.021090025082230568, + 0.008223132230341434, + 0.009560749866068363, + 0.009309899061918259, + 0.0225515253841877, + 0.018203623592853546, + 0.02337198331952095, + 0.020991357043385506, + 0.021557020023465157, + 0.02855405956506729, + 0.0163051076233387, + 0.018565787002444267, + 0.008423151448369026, + 0.004696284420788288, + 0.0075209373608231544, + 0.006567288190126419, + 0.005088553763926029, + 0.00597412372007966, + 0.004077553749084473, + 0.006345328874886036, + 0.005422031972557306, + 0.0035742574837058783, + 0.009364522993564606, + 0.003781105624511838, + 0.003434499492868781, + 0.0051735700108110905, + 0.003655540058389306, + 0.00669768126681447, + 0.007262025028467178, + 0.0029226848855614662, + 0.0045500569976866245, + 0.0031562703661620617, + 0.002841574838384986, + 0.0065659224055707455, + 0.0037356368266046047, + 0.005213515367358923, + 0.05730471387505531, + 0.0087810680270195, + 0.007521103601902723, + 0.005876168608665466, + 0.008371422067284584, + 0.07175274938344955, + 0.007552431896328926, + 0.013044181279838085, + 0.011445811949670315, + 0.010660717263817787, + 0.01696160063147545, + 0.010420483537018299, + 0.011213291436433792, + 0.011839771643280983, + 0.010197797790169716, + 0.01155934389680624, + 0.017055265605449677, + 0.00864699762314558, + 0.019503742456436157, + 0.016036562621593475, + 0.01015821099281311, + 0.011540250852704048, + 0.015217045322060585, + 0.017913294956088066, + 0.010670549236238003, + 0.00723605090752244, + 0.011642021127045155, + 0.010451923124492168, + 0.006705678999423981, + 0.00871842261403799, + 0.00808016862720251, + 0.011545064859092236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.9742594957351685 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "high-angle macro shot, characterized by constant, subtle handheld shake, as if the camera were held by a trembling hand, focuses intensely on a classic, round, rainbow-swirled lollipop. The lollipop, with its vibrant, glossy swirls of cherry red, electric blue, and sunshine yellow, rests on a pristine, matte white surface under bright, clinical studio lighting. The action begins as the lollipop's crystal surface begins to sweat, forming tiny, shimmering beads of syrup that catch the overhead light. These beads quickly coalesce, and as the solid form dramatically liquefies, the disparate streams of color begin to chaotically interpenetrate. The lollipop collapses downward and outward, dissolving into a viscous, ever-expanding puddle. Once-separate colors now swirl together in a chaotic, hypnotic vortex, creating a miniature, turbulent galaxy of melted sugar. The relentless lens tremor, combined with the mesmerizing yet destructive shifts, creates a tense and slightly unsettling atmosphere, capturing the beautiful decay of a childhood icon.", + "seed": 23528, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_6/index_2376_seed_23528_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3500372152775526, + "vlm_logprob": -0.3500372152775526, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.3500372152775526, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008779496885836124 + ], + "opposite_logprob_tokens": [ + -5.883779525756836 + ], + "logprob_reward": -0.008779496885836124, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.008779496885836124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07664331793785095, + 0.04289880394935608, + 0.07357866317033768, + 0.07634710520505905, + 0.030415289103984833, + 0.09480759501457214, + 0.34015563130378723, + 0.11198519915342331, + 0.11364647001028061, + 0.03952198475599289 + ], + "credit_map": [ + 0.009865052066743374, + 0.009454133920371532, + 0.012261460535228252, + 0.00489336671307683, + 0.011586860753595829, + 0.009987116791307926, + 0.0056408969685435295, + 0.012954429723322392, + 0.006438319571316242, + 0.0023365956731140614, + 0.006270187441259623, + 0.00656814593821764, + 0.0032794466242194176, + 0.005089218262583017, + 0.0050352755934000015, + 0.007881613448262215, + 0.008558829315006733, + 0.0050570485182106495, + 0.011492549441754818, + 0.00503938365727663, + 0.01708158478140831, + 0.009801366366446018, + 0.006580234970897436, + 0.00996765960007906, + 0.00949071254581213, + 0.006992264185100794, + 0.010885260067880154, + 0.010944410227239132, + 0.00999325979501009, + 0.013266468420624733, + 0.006491400301456451, + 0.008283325470983982, + 0.0030158539302647114, + 0.0032286958303302526, + 0.0025206231512129307, + 0.002938891062512994, + 0.003706923918798566, + 0.00791927706450224, + 0.002961160149425268, + 0.004123864229768515, + 0.009135385975241661, + 0.023683888837695122, + 0.009313044138252735, + 0.013444109819829464, + 0.013997935689985752, + 0.009954577311873436, + 0.007642542477697134, + 0.007636116351932287, + 0.02334507182240486, + 0.02870774082839489, + 0.011729002930223942, + 0.022470518946647644, + 0.011457711458206177, + 0.14502181112766266, + 0.015362842008471489, + 0.08206091076135635, + 0.006864441558718681, + 0.02231467515230179, + 0.006506394594907761, + 0.025535520166158676, + 0.016087915748357773, + 0.01068637240678072, + 0.011371820233762264, + 0.012618052773177624, + 0.013599974103271961, + 0.013086485676467419, + 0.006031274329870939, + 0.025010960176587105, + 0.007774911820888519, + 0.013404276221990585, + 0.010578463785350323, + 0.024160116910934448, + 0.0031348932534456253, + 0.01101759634912014, + 0.010754911229014397, + 0.004441684577614069, + 0.0021775488276034594, + 0.003062683856114745, + 0.0028660062234848738, + 0.0020666602067649364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.008779496885836124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006755489390343428 + ], + "opposite_logprob_tokens": [ + -5.006755352020264 + ], + "logprob_reward": -0.006755489390343428, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006755489390343428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060877081006765366, + 0.04371647164225578, + 0.0643259659409523, + 0.13372476398944855, + 0.03751009330153465, + 0.05871022865176201, + 0.319399356842041, + 0.041533563286066055, + 0.18393899500370026, + 0.056263476610183716 + ], + "credit_map": [ + 0.008185449987649918, + 0.00872625783085823, + 0.010105176828801632, + 0.004428568761795759, + 0.006661084480583668, + 0.007260921411216259, + 0.004436465445905924, + 0.011073155328631401, + 0.005602859891951084, + 0.002324468921869993, + 0.006127316039055586, + 0.005811803974211216, + 0.005358924623578787, + 0.00421688100323081, + 0.005270028952509165, + 0.009004189632833004, + 0.00831577368080616, + 0.006165655329823494, + 0.009663546457886696, + 0.007032481487840414, + 0.008643834851682186, + 0.009470557793974876, + 0.0047422004863619804, + 0.01029191818088293, + 0.016489209607243538, + 0.013559353537857533, + 0.023012587800621986, + 0.021535197272896767, + 0.01570175029337406, + 0.017690882086753845, + 0.01225221436470747, + 0.01348358765244484, + 0.003517481731250882, + 0.003428092924878001, + 0.003266340820118785, + 0.00440760375931859, + 0.005164756439626217, + 0.01015702448785305, + 0.003592820605263114, + 0.003975972067564726, + 0.005028525833040476, + 0.012745912186801434, + 0.005525232758373022, + 0.007986561395227909, + 0.009861648082733154, + 0.00572595652192831, + 0.0055085779167711735, + 0.006327813025563955, + 0.021941738203167915, + 0.027047887444496155, + 0.011470154859125614, + 0.019839679822325706, + 0.011565490625798702, + 0.1359693557024002, + 0.014258245937526226, + 0.07730679959058762, + 0.00303404382430017, + 0.008291816338896751, + 0.0027999221347272396, + 0.008160118944942951, + 0.0065952204167842865, + 0.004024113528430462, + 0.005073103122413158, + 0.00355522520840168, + 0.011322778649628162, + 0.03351220488548279, + 0.006967609282582998, + 0.06835630536079407, + 0.009686318226158619, + 0.011438095010817051, + 0.013735258020460606, + 0.02892041951417923, + 0.0038859921041876078, + 0.015111547894775867, + 0.01588529534637928, + 0.006145313382148743, + 0.0035171082708984613, + 0.0045995465479791164, + 0.004072641953825951, + 0.0030460311099886894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.006755489390343428 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026368023827672005 + ], + "opposite_logprob_tokens": [ + -3.6513681411743164 + ], + "logprob_reward": -0.026368023827672005, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026368023827672005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05874297767877579, + 0.1330977827310562, + 0.3308362364768982, + 0.04911896213889122, + 0.03901300206780434, + 0.07442357391119003, + 0.09221519529819489, + 0.07559774816036224, + 0.11904142796993256, + 0.027913151308894157 + ], + "credit_map": [ + 0.009503117762506008, + 0.004590168129652739, + 0.008727644570171833, + 0.00639134319499135, + 0.006121671758592129, + 0.0063959751278162, + 0.003284787992015481, + 0.013728261925280094, + 0.01993921957910061, + 0.009335003793239594, + 0.018628718331456184, + 0.017364203929901123, + 0.014165771193802357, + 0.011333738453686237, + 0.017976731061935425, + 0.02435438148677349, + 0.04636082798242569, + 0.027480779215693474, + 0.04302569106221199, + 0.022412214428186417, + 0.08231882005929947, + 0.035342827439308167, + 0.026407340541481972, + 0.04748772829771042, + 0.006585496477782726, + 0.004282291047275066, + 0.008896789513528347, + 0.006521675270050764, + 0.0050165606662631035, + 0.00816542748361826, + 0.004762664437294006, + 0.004888053983449936, + 0.00417395168915391, + 0.003682337701320648, + 0.0036005517467856407, + 0.003414659295231104, + 0.005015050061047077, + 0.011383747681975365, + 0.003388910088688135, + 0.004353789146989584, + 0.006227557547390461, + 0.01618000864982605, + 0.0067896246910095215, + 0.009680815041065216, + 0.011081493459641933, + 0.0078040435910224915, + 0.007025338709354401, + 0.009634692221879959, + 0.006163549143821001, + 0.00845406111329794, + 0.004301093984395266, + 0.005351128987967968, + 0.004230012185871601, + 0.036688610911369324, + 0.00510043790563941, + 0.02192629873752594, + 0.006946148816496134, + 0.008585440926253796, + 0.00675210589542985, + 0.024710139259696007, + 0.006783153396099806, + 0.0064485929906368256, + 0.007328822743147612, + 0.008043341338634491, + 0.009846477769315243, + 0.01632942073047161, + 0.0040876553393900394, + 0.04190884530544281, + 0.00561155378818512, + 0.008751102723181248, + 0.009687262587249279, + 0.022819088771939278, + 0.002204255433753133, + 0.008362391032278538, + 0.0074028572998940945, + 0.0027294158935546875, + 0.0017536545637995005, + 0.0016063604271039367, + 0.0023589052725583315, + 0.0014953064965084195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.026368023827672005 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067406888119876385 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.0067406888119876385, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067406888119876385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02145027555525303, + 0.0726611465215683, + 0.07150580734014511, + 0.06676787883043289, + 0.02740812487900257, + 0.029971066862344742, + 0.11449859291315079, + 0.06756958365440369, + 0.44494950771331787, + 0.08321801573038101 + ], + "credit_map": [ + 0.0035696907434612513, + 0.00158661138266325, + 0.003909602761268616, + 0.0023538083769381046, + 0.0018629598198458552, + 0.002808887278661132, + 0.0017661070451140404, + 0.003592605236917734, + 0.010106435976922512, + 0.003950282000005245, + 0.010850800201296806, + 0.013449357822537422, + 0.00703597255051136, + 0.005686196498572826, + 0.012084683403372765, + 0.009497418068349361, + 0.009446710348129272, + 0.00492442212998867, + 0.009994246065616608, + 0.005049348343163729, + 0.0171203650534153, + 0.008923611603677273, + 0.005751143209636211, + 0.010295957326889038, + 0.008402074687182903, + 0.006848233751952648, + 0.010652273893356323, + 0.010570742189884186, + 0.008017295971512794, + 0.008517376147210598, + 0.005656042601913214, + 0.008103832602500916, + 0.003180721076205373, + 0.003788054920732975, + 0.0027360175736248493, + 0.00396478408947587, + 0.002980752382427454, + 0.004854048602283001, + 0.002887206617742777, + 0.003016537055373192, + 0.0030385409481823444, + 0.006047964096069336, + 0.0033282414078712463, + 0.0025750435888767242, + 0.0035157245583832264, + 0.004568301606923342, + 0.0031200451776385307, + 0.0037772038485854864, + 0.008255582302808762, + 0.009617106057703495, + 0.003968540579080582, + 0.008708564564585686, + 0.00431954488158226, + 0.04837856441736221, + 0.0053772577084600925, + 0.025873415172100067, + 0.00449648592621088, + 0.01082772295922041, + 0.00537720276042819, + 0.018837081268429756, + 0.006743676960468292, + 0.00510941818356514, + 0.005986523814499378, + 0.010191461071372032, + 0.02823621779680252, + 0.07474805414676666, + 0.014879652298986912, + 0.16457132995128632, + 0.02096485160291195, + 0.028407560661435127, + 0.035163622349500656, + 0.0779782235622406, + 0.005582318641245365, + 0.021658463403582573, + 0.021588876843452454, + 0.010189787484705448, + 0.005406255833804607, + 0.007452608551830053, + 0.006353139877319336, + 0.004986558575183153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.0067406888119876385 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a high-angle macro view focusing on a classic round lollipop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lollipop display vibrant glossy swirls of cherry red, electric blue, and sunshine yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lollipop resting on a pristine matte white surface under bright clinical studio lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lollipop melt into a viscous puddle where the colors interpenetrate to form a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a constant subtle handheld shake in the camera movement throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7015423774719238 + ], + "opposite_logprob_tokens": [ + -0.20154236257076263 + ], + "logprob_reward": -1.7015423774719238, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7015423774719238, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.026679394766688347, + 0.04183262586593628, + 0.06724797189235687, + 0.033503707498311996, + 0.023560183122754097, + 0.04792838916182518, + 0.22825127840042114, + 0.14164137840270996, + 0.26747989654541016, + 0.12187515944242477 + ], + "credit_map": [ + 0.004916263744235039, + 0.002067475812509656, + 0.004214559216052294, + 0.002652703784406185, + 0.0027537730056792498, + 0.0035071615129709244, + 0.0022126403637230396, + 0.0043548173271119595, + 0.005956376902759075, + 0.002544617047533393, + 0.006130091845989227, + 0.006877816282212734, + 0.00515090674161911, + 0.0036289962008595467, + 0.007048900239169598, + 0.004494921304285526, + 0.008326704613864422, + 0.005221422296017408, + 0.009542512707412243, + 0.005117644555866718, + 0.01807042397558689, + 0.007235256023705006, + 0.006045992020517588, + 0.007688009180128574, + 0.005733824335038662, + 0.0026708149816840887, + 0.003612876171246171, + 0.005020991433411837, + 0.0038444248493760824, + 0.004328631330281496, + 0.003633100539445877, + 0.004659043159335852, + 0.0027959293220192194, + 0.0028865274507552385, + 0.0020692585967481136, + 0.002902924781665206, + 0.003038424998521805, + 0.00509566580876708, + 0.0023614238016307354, + 0.002410029759630561, + 0.0034441633615642786, + 0.009129934944212437, + 0.004072091076523066, + 0.007183628156781197, + 0.009126616641879082, + 0.005961537826806307, + 0.003988694399595261, + 0.0050217220559716225, + 0.015666499733924866, + 0.019356081262230873, + 0.008554305881261826, + 0.015582134947180748, + 0.008621332235634327, + 0.09721597284078598, + 0.010204780846834183, + 0.053050171583890915, + 0.012828946113586426, + 0.026581542566418648, + 0.011066123843193054, + 0.01714969426393509, + 0.02199176698923111, + 0.019461290910840034, + 0.019762462005019188, + 0.012799552641808987, + 0.020214319229125977, + 0.03719663247466087, + 0.008869899436831474, + 0.09599901735782623, + 0.01236687321215868, + 0.01915193349123001, + 0.022301509976387024, + 0.05137970671057701, + 0.006920953746885061, + 0.039018843322992325, + 0.030998004600405693, + 0.013296042568981647, + 0.008446186780929565, + 0.008205023594200611, + 0.009094052948057652, + 0.00589604489505291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -1.7015423774719238 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_7/index_tensor([5132, 5132])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_7/index_tensor([5132, 5132])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7ef28a84c934b92ed1cadafc7f6c92afedf84b2f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_7/index_tensor([5132, 5132])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a slightly overhead angle, the camera slowly and steadily tilts downward, focusing on a warm, light-grained wood tabletop. A slender, manicured hand elegantly grips a classic black lacquer fountain pen. The silver tip glides across the delicate, off-white hardcover journal page, the wrist steady and deliberate, leaving lines of elegant cursive handwriting. Next to the journal, a latte steams in a white ceramic cup, its frothed milk featuring an intricate latte art pattern. Wisps of delicate white steam rise, visible in the light. A ray of fresh morning light slants in from the window, casting a soft, elongated spot on the wooden table, giving the rising steam a textured quality. In the background, the rest of the cafe is softly blurred, with the faint silhouettes of other patrons and the warm interior lighting barely visible. The main colors of the whole picture are warm beige, white and wood color, creating a peaceful, focused and poetic atmosphere, perfectly capturing a leisurely and beautiful morning time.", + "seed": 6375, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_7/index_5132_seed_6375_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.024749944824725388, + "vlm_logprob": -0.024749944824725388, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.024749944824725388, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013906724750995636 + ], + "opposite_logprob_tokens": [ + -5.888906955718994 + ], + "logprob_reward": -0.013906724750995636, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.013906724750995636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2389768362045288, + 0.13743478059768677, + 0.05285871773958206, + 0.16123779118061066, + 0.025036098435521126, + 0.14451350271701813, + 0.06284864991903305, + 0.032814331352710724, + 0.09446559101343155, + 0.049813639372587204 + ], + "credit_map": [ + 0.0335959754884243, + 0.026394151151180267, + 0.023574918508529663, + 0.019278962165117264, + 0.03993067145347595, + 0.026153380051255226, + 0.04538207873702049, + 0.024666694924235344, + 0.012141041457653046, + 0.012663173489272594, + 0.021762119606137276, + 0.01810767687857151, + 0.016244594007730484, + 0.022122854366898537, + 0.013294672593474388, + 0.021098660305142403, + 0.005099146161228418, + 0.009088489226996899, + 0.008456061594188213, + 0.00700072618201375, + 0.006090542301535606, + 0.004634142853319645, + 0.006609367206692696, + 0.005880244076251984, + 0.01036916859447956, + 0.03208098188042641, + 0.0309566892683506, + 0.00780104985460639, + 0.007806482259184122, + 0.048659611493349075, + 0.008042803965508938, + 0.015521001070737839, + 0.003652534680441022, + 0.00283855851739645, + 0.0032165362499654293, + 0.00500484136864543, + 0.0026666095945984125, + 0.0017334906151518226, + 0.003576539922505617, + 0.0023469876032322645, + 0.010881714522838593, + 0.006227365229278803, + 0.04911382496356964, + 0.047205016016960144, + 0.005376595538109541, + 0.005716571118682623, + 0.008360362611711025, + 0.011632039211690426, + 0.004134494345635176, + 0.004549018107354641, + 0.02189517207443714, + 0.01666714809834957, + 0.004718647338449955, + 0.0026993337087333202, + 0.003731412347406149, + 0.004453425295650959, + 0.004671026952564716, + 0.003621498355641961, + 0.004973704926669598, + 0.0028558855410665274, + 0.005577161908149719, + 0.0030909734778106213, + 0.0032080872915685177, + 0.004815993830561638, + 0.014310302212834358, + 0.009872488677501678, + 0.012866409495472908, + 0.01756996475160122, + 0.01222633570432663, + 0.006851640529930592, + 0.010229156352579594, + 0.010539296083152294, + 0.006713658105581999, + 0.004702959209680557, + 0.007466915529221296, + 0.007382755167782307, + 0.005359086208045483, + 0.005832106806337833, + 0.003662095870822668, + 0.008694062940776348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.013906724750995636 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700269266963005 + ], + "opposite_logprob_tokens": [ + -2.6950268745422363 + ], + "logprob_reward": -0.0700269266963005, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700269266963005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11843957751989365, + 0.05510229244828224, + 0.04262683540582657, + 0.17790265381336212, + 0.12079129368066788, + 0.10959335416555405, + 0.030460679903626442, + 0.03726561367511749, + 0.0486966110765934, + 0.2591211497783661 + ], + "credit_map": [ + 0.018766338005661964, + 0.014301028102636337, + 0.01153561007231474, + 0.010892256163060665, + 0.02196034975349903, + 0.0157387163490057, + 0.014384203590452671, + 0.010861067101359367, + 0.006716335657984018, + 0.00566236674785614, + 0.008483024314045906, + 0.006090098526328802, + 0.008607834577560425, + 0.0065334271639585495, + 0.009393653832376003, + 0.0036155565176159143, + 0.003413262078538537, + 0.0037295676302164793, + 0.004493020940572023, + 0.004040301311761141, + 0.003411573125049472, + 0.007898042909801006, + 0.004753301851451397, + 0.010887766256928444, + 0.013119992800056934, + 0.018368493765592575, + 0.046460919082164764, + 0.010042749345302582, + 0.013511432334780693, + 0.04057407006621361, + 0.010684214532375336, + 0.025140784680843353, + 0.005866406951099634, + 0.007280953694134951, + 0.0452054999768734, + 0.04502241685986519, + 0.0029212518129497766, + 0.0029443264938890934, + 0.005834129173308611, + 0.005716312676668167, + 0.006770798936486244, + 0.003862692043185234, + 0.04323510453104973, + 0.04072294011712074, + 0.003260605735704303, + 0.0022992128506302834, + 0.0050155241042375565, + 0.004426481202244759, + 0.004363205749541521, + 0.0041090031154453754, + 0.005631194915622473, + 0.003970560152083635, + 0.0029881333466619253, + 0.0029731481336057186, + 0.003092145547270775, + 0.003333288012072444, + 0.011673396453261375, + 0.003577942494302988, + 0.0037391874939203262, + 0.0027632617857307196, + 0.004050358198583126, + 0.0038892559241503477, + 0.0029553945641964674, + 0.004616816993802786, + 0.006769358180463314, + 0.00545460032299161, + 0.005868085660040379, + 0.008146274834871292, + 0.004172308370471001, + 0.0069959452375769615, + 0.0062053631991147995, + 0.005084673874080181, + 0.040067147463560104, + 0.02342980168759823, + 0.04209060221910477, + 0.031856127083301544, + 0.021363940089941025, + 0.02443690225481987, + 0.028189068660140038, + 0.047687552869319916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.0700269266963005 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744833197444677 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.006744833197444677, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744833197444677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16429193317890167, + 0.07550006359815598, + 0.17905880510807037, + 0.059934671968221664, + 0.042016081511974335, + 0.033819496631622314, + 0.023568682372570038, + 0.0401223748922348, + 0.03682766482234001, + 0.3448602855205536 + ], + "credit_map": [ + 0.024364257231354713, + 0.018694138154387474, + 0.015181703492999077, + 0.013132822699844837, + 0.027085768058896065, + 0.019239313900470734, + 0.027933461591601372, + 0.01866045780479908, + 0.008170725777745247, + 0.008945989422500134, + 0.015367074869573116, + 0.008037775754928589, + 0.010086803697049618, + 0.009786983951926231, + 0.010435645468533039, + 0.00466906325891614, + 0.01282927580177784, + 0.018067313358187675, + 0.020253490656614304, + 0.020782023668289185, + 0.016627570614218712, + 0.028532028198242188, + 0.019927391782402992, + 0.04203971475362778, + 0.004970662295818329, + 0.019847821444272995, + 0.009496457874774933, + 0.0037687779404222965, + 0.0043390714563429356, + 0.005602775141596794, + 0.003870522603392601, + 0.00803858507424593, + 0.0025941524654626846, + 0.0025361112784594297, + 0.014911489561200142, + 0.007752949371933937, + 0.0026639066636562347, + 0.0028242389671504498, + 0.004264241550117731, + 0.004468991421163082, + 0.004303312394768, + 0.0030629343818873167, + 0.006673273164778948, + 0.005574759561568499, + 0.0031756197568029165, + 0.003188102738931775, + 0.004076279234141111, + 0.0037652140017598867, + 0.0031774453818798065, + 0.003841598518192768, + 0.0030834779608994722, + 0.0027023935690522194, + 0.0033113439567387104, + 0.0031424283515661955, + 0.002236056374385953, + 0.002073939424008131, + 0.007048961240798235, + 0.005917132366448641, + 0.0029058787040412426, + 0.002730385400354862, + 0.007401445414870977, + 0.003508572932332754, + 0.004969786386936903, + 0.005640210118144751, + 0.004917858634144068, + 0.003380216658115387, + 0.0058873677626252174, + 0.003853187896311283, + 0.004783149342983961, + 0.0031399279832839966, + 0.003760665189474821, + 0.007105293683707714, + 0.04851175844669342, + 0.03200995922088623, + 0.04937442019581795, + 0.05016498640179634, + 0.05047853663563728, + 0.033104658126831055, + 0.03236467018723488, + 0.04885128512978554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.006744833197444677 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340970769524574 + ], + "opposite_logprob_tokens": [ + -3.651340961456299 + ], + "logprob_reward": -0.026340970769524574, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340970769524574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10971168428659439, + 0.16187390685081482, + 0.14351926743984222, + 0.13185083866119385, + 0.1229502335190773, + 0.07864983379840851, + 0.03460662439465523, + 0.051683396100997925, + 0.03795503452420235, + 0.12719915807247162 + ], + "credit_map": [ + 0.011184257455170155, + 0.018414240330457687, + 0.015947693958878517, + 0.01233090553432703, + 0.014198740012943745, + 0.012754304334521294, + 0.012561640702188015, + 0.012319899164140224, + 0.01832456886768341, + 0.01804935559630394, + 0.024322794750332832, + 0.016133682802319527, + 0.02971610054373741, + 0.02028525061905384, + 0.02313932590186596, + 0.011902828700840473, + 0.013593749143183231, + 0.03433343395590782, + 0.024294069036841393, + 0.014050631783902645, + 0.010381832718849182, + 0.013484915718436241, + 0.010097989812493324, + 0.02328265644609928, + 0.010241658426821232, + 0.012007107026875019, + 0.037134621292352676, + 0.006954337004572153, + 0.009104849770665169, + 0.034656278789043427, + 0.006712744943797588, + 0.01503924559801817, + 0.00661914236843586, + 0.006205368787050247, + 0.04416157305240631, + 0.04488031193614006, + 0.004071657080203295, + 0.0041274274699389935, + 0.006335045676678419, + 0.006549702491611242, + 0.004782192409038544, + 0.0031342615839093924, + 0.030447550117969513, + 0.028745580464601517, + 0.0025234806817024946, + 0.0020102779380977154, + 0.003943655639886856, + 0.003062835428863764, + 0.0032703613396734, + 0.003685442265123129, + 0.006774534471333027, + 0.008397015742957592, + 0.00392004381865263, + 0.002465297235175967, + 0.0030402292031794786, + 0.0030536965932697058, + 0.015978476032614708, + 0.005314325913786888, + 0.003863085526973009, + 0.0044191875495016575, + 0.005922566168010235, + 0.003255729330703616, + 0.00516109075397253, + 0.0077689350582659245, + 0.0056242020800709724, + 0.003916103858500719, + 0.006245640106499195, + 0.005692336708307266, + 0.005009938031435013, + 0.0030192104168236256, + 0.0037722336128354073, + 0.004675370175391436, + 0.016466857865452766, + 0.01089551392942667, + 0.017617128789424896, + 0.02062772773206234, + 0.019018245860934258, + 0.012348953634500504, + 0.01217554323375225, + 0.018049191683530807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.026340970769524574 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730268709361553 + ], + "opposite_logprob_tokens": [ + -5.006730079650879 + ], + "logprob_reward": -0.006730268709361553, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730268709361553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1846960186958313, + 0.038473013788461685, + 0.10282181203365326, + 0.07619690895080566, + 0.30934351682662964, + 0.14289823174476624, + 0.04873446002602577, + 0.0267487820237875, + 0.03483354300260544, + 0.03525374084711075 + ], + "credit_map": [ + 0.022609097883105278, + 0.029600661247968674, + 0.02196863479912281, + 0.01849500648677349, + 0.02107810787856579, + 0.018969187512993813, + 0.03369227424263954, + 0.018283065408468246, + 0.00344602158293128, + 0.003186412388458848, + 0.0046425461769104, + 0.003934806678444147, + 0.003843998536467552, + 0.00809962023049593, + 0.004521884955465794, + 0.006797728128731251, + 0.006682258564978838, + 0.00926871970295906, + 0.013625110499560833, + 0.008025122806429863, + 0.007620525546371937, + 0.018076643347740173, + 0.006862748879939318, + 0.03266068175435066, + 0.004665068816393614, + 0.02961578592658043, + 0.0061186798848211765, + 0.003938890527933836, + 0.0037096315063536167, + 0.017991814762353897, + 0.0037243154365569353, + 0.0064327288419008255, + 0.012933766469359398, + 0.014913816004991531, + 0.12787652015686035, + 0.1155366450548172, + 0.007241981569677591, + 0.007185060065239668, + 0.015465029515326023, + 0.008190707303583622, + 0.015115432441234589, + 0.007369557861238718, + 0.04850655421614647, + 0.046793293207883835, + 0.004501401912420988, + 0.006742290686815977, + 0.0073774587363004684, + 0.006492260843515396, + 0.0034114879090338945, + 0.0033603699412196875, + 0.01522478461265564, + 0.016100909560918808, + 0.003384381765499711, + 0.0021307463757693768, + 0.0028267574962228537, + 0.002295027719810605, + 0.00740548362955451, + 0.0026972200721502304, + 0.0025308257900178432, + 0.003221239196136594, + 0.002870276104658842, + 0.0017025414854288101, + 0.002703801030293107, + 0.0036173982080072165, + 0.0054870606400072575, + 0.00408095633611083, + 0.004198528826236725, + 0.003937811590731144, + 0.0050718048587441444, + 0.0028113755397498608, + 0.003206908470019698, + 0.006039100233465433, + 0.005460270680487156, + 0.003463502274826169, + 0.004778490401804447, + 0.004953823052346706, + 0.004294710699468851, + 0.004146020393818617, + 0.0034837403800338507, + 0.0046731834299862385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.006730268709361553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, the camera slowly and steadily tilts downward, focusing on a warm, light-grained wood tabletop. A slender, manicured hand elegantly grips a classic black lacquer fountain pen. The silver tip glides across the delicate, off-white hardcover journal page, the wrist steady and deliberate, leaving lines of elegant cursive handwriting. Next to the journal, a latte steams in a white ceramic cup, its frothed milk featuring an intricate latte art pattern. Wisps of delicate white steam rise, visible in the light. A ray of fresh morning light slants in from the window, casting a soft, elongated spot on the wooden table, giving the rising steam a textured quality. In the background, the rest of the cafe is softly blurred, with the faint silhouettes of other patrons and the warm interior lighting barely visible. The main colors of the whole picture are warm beige, white and wood color, creating a peaceful, focused and poetic atmosphere, perfectly capturing a leisurely and beautiful morning time.", + "seed": 6376, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_7/index_5132_seed_6376_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02381632402539253, + "vlm_logprob": -0.02381632402539253, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.02381632402539253, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019223107025027275 + ], + "opposite_logprob_tokens": [ + -6.894223213195801 + ], + "logprob_reward": -0.019223107025027275, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.019223107025027275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05315458029508591, + 0.031330518424510956, + 0.0262849573045969, + 0.026549626141786575, + 0.10958573967218399, + 0.051945339888334274, + 0.22418873012065887, + 0.05161971226334572, + 0.4041309356689453, + 0.02120986022055149 + ], + "credit_map": [ + 0.006532180123031139, + 0.008584897965192795, + 0.005841481499373913, + 0.005566364154219627, + 0.004393038339912891, + 0.007553082425147295, + 0.006328750401735306, + 0.008354786783456802, + 0.0028509534895420074, + 0.004441250115633011, + 0.004688592627644539, + 0.0031954506412148476, + 0.0043329200707376, + 0.004052606876939535, + 0.005297791678458452, + 0.0024709568824619055, + 0.002534344093874097, + 0.005267264321446419, + 0.002729997271671891, + 0.002913188189268112, + 0.00237329606898129, + 0.003845739411190152, + 0.002797460649162531, + 0.0038236670661717653, + 0.002403007820248604, + 0.0029168121982365847, + 0.0026748771779239178, + 0.003374130465090275, + 0.0029091518372297287, + 0.0030923110898584127, + 0.003993283491581678, + 0.005186052527278662, + 0.026992537081241608, + 0.01256554014980793, + 0.02183612808585167, + 0.009010609239339828, + 0.007254679221659899, + 0.017855266109108925, + 0.005537362303584814, + 0.008533620275557041, + 0.008606010116636753, + 0.009315554983913898, + 0.006659389939159155, + 0.003873045789077878, + 0.008156451396644115, + 0.004677709192037582, + 0.004902116023004055, + 0.005755061283707619, + 0.026260973885655403, + 0.08489059656858444, + 0.022844793274998665, + 0.018179461359977722, + 0.012284536845982075, + 0.035147182643413544, + 0.007013162598013878, + 0.017568008974194527, + 0.007548842113465071, + 0.004773512948304415, + 0.007836912758648396, + 0.004888493102043867, + 0.009936396963894367, + 0.005080633331090212, + 0.0037142771761864424, + 0.007840640842914581, + 0.041510213166475296, + 0.017205961048603058, + 0.12888093292713165, + 0.020343197509646416, + 0.02155149355530739, + 0.13524053990840912, + 0.016419459134340286, + 0.022979138419032097, + 0.0028083620127290487, + 0.00220005726441741, + 0.0031230831518769264, + 0.0019692271016538143, + 0.004110601730644703, + 0.0025062481872737408, + 0.0017467793077230453, + 0.0027455026283860207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.019223107025027275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495237931609154 + ], + "opposite_logprob_tokens": [ + -2.929952383041382 + ], + "logprob_reward": -0.05495237931609154, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495237931609154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18784098327159882, + 0.06632011383771896, + 0.09620221704244614, + 0.08151309937238693, + 0.03687671944499016, + 0.08252160996198654, + 0.28387129306793213, + 0.08281257748603821, + 0.0670531839132309, + 0.014988157898187637 + ], + "credit_map": [ + 0.014570431783795357, + 0.046314988285303116, + 0.021903393790125847, + 0.015899570658802986, + 0.01888621784746647, + 0.032274845987558365, + 0.026250110939145088, + 0.011741426773369312, + 0.008824227377772331, + 0.00727026117965579, + 0.014552181586623192, + 0.006244413089007139, + 0.005421322304755449, + 0.00964379496872425, + 0.008294440805912018, + 0.006069468334317207, + 0.011035661213099957, + 0.018192077055573463, + 0.01798490434885025, + 0.00846512708812952, + 0.008103664964437485, + 0.014357632957398891, + 0.00896159466356039, + 0.00910155102610588, + 0.008836989291012287, + 0.008590705692768097, + 0.020119372755289078, + 0.007972569204866886, + 0.007583776954561472, + 0.010465891100466251, + 0.010259630158543587, + 0.007684160023927689, + 0.005544127430766821, + 0.005559225101023912, + 0.006445214617997408, + 0.0034214607439935207, + 0.0038886449765414, + 0.005324607249349356, + 0.003461462678387761, + 0.003231975482776761, + 0.010880311951041222, + 0.020182272419333458, + 0.01038286555558443, + 0.006997830234467983, + 0.011518481187522411, + 0.0068613337352871895, + 0.005828129593282938, + 0.009870383888483047, + 0.027191147208213806, + 0.1195889338850975, + 0.018268657848238945, + 0.012316136620938778, + 0.014999591745436192, + 0.05642027407884598, + 0.007420276291668415, + 0.02766629122197628, + 0.01017910148948431, + 0.00886563491076231, + 0.009645001962780952, + 0.004329954273998737, + 0.01773146726191044, + 0.010054554790258408, + 0.008933411911129951, + 0.013073445297777653, + 0.009144212119281292, + 0.0040591624565422535, + 0.01988021843135357, + 0.006965030450373888, + 0.004778435919433832, + 0.006950788665562868, + 0.004205599427223206, + 0.01106974482536316, + 0.0019670738838613033, + 0.0025002562906593084, + 0.0021797132212668657, + 0.001424960559234023, + 0.0018728497670963407, + 0.0017563867149874568, + 0.0014143767766654491, + 0.001872541499324143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.05495237931609154 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0125422989949584 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.0125422989949584, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0125422989949584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13661602139472961, + 0.030886614695191383, + 0.04150448366999626, + 0.031039733439683914, + 0.037681229412555695, + 0.06389806419610977, + 0.07801496982574463, + 0.2291206419467926, + 0.30739516019821167, + 0.043843068182468414 + ], + "credit_map": [ + 0.011195582337677479, + 0.026592306792736053, + 0.01512114703655243, + 0.012798616662621498, + 0.017411604523658752, + 0.02079244703054428, + 0.018745113164186478, + 0.01395920105278492, + 0.0039134398102760315, + 0.0030147223733365536, + 0.006190111860632896, + 0.002759333932772279, + 0.003517541568726301, + 0.0045969001948833466, + 0.004360392224043608, + 0.002534171100705862, + 0.004129994194954634, + 0.008315366692841053, + 0.0052300989627838135, + 0.0035391345154494047, + 0.004148868843913078, + 0.006180686876177788, + 0.00458534387871623, + 0.005374987609684467, + 0.0048182313330471516, + 0.005915778689086437, + 0.0036891812924295664, + 0.0030889385379850864, + 0.0027830214239656925, + 0.0041593038477003574, + 0.003412816673517227, + 0.003172458615154028, + 0.006796388421207666, + 0.004881494212895632, + 0.005968762561678886, + 0.002867565955966711, + 0.004631435964256525, + 0.0048862965777516365, + 0.002971478970721364, + 0.004677800461649895, + 0.012259447015821934, + 0.009088906459510326, + 0.006456831935793161, + 0.004631213378161192, + 0.01151344645768404, + 0.00570972915738821, + 0.005582859273999929, + 0.008655624464154243, + 0.008347827009856701, + 0.0295117050409317, + 0.005139926448464394, + 0.0031152796000242233, + 0.004642731510102749, + 0.012212530709803104, + 0.0031081296037882566, + 0.011936835944652557, + 0.018043622374534607, + 0.04600908234715462, + 0.021115435287356377, + 0.010502384044229984, + 0.045085061341524124, + 0.029988320544362068, + 0.02554658055305481, + 0.03283014893531799, + 0.02443774975836277, + 0.010808218270540237, + 0.11266529560089111, + 0.023634737357497215, + 0.013329336419701576, + 0.08381413668394089, + 0.017239462584257126, + 0.0214662104845047, + 0.00486085657030344, + 0.0066000972874462605, + 0.005712326616048813, + 0.003258044831454754, + 0.006652488838881254, + 0.006703529506921768, + 0.004156723618507385, + 0.005898996256291866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0125422989949584 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018178289756178856 + ], + "opposite_logprob_tokens": [ + -4.018178462982178 + ], + "logprob_reward": -0.018178289756178856, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018178289756178856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08027093857526779, + 0.06455590575933456, + 0.07260202616453171, + 0.05661853402853012, + 0.029849017038941383, + 0.0611221045255661, + 0.25973254442214966, + 0.09541794657707214, + 0.20765583217144012, + 0.07217516005039215 + ], + "credit_map": [ + 0.0076876250095665455, + 0.01831190101802349, + 0.010385550558567047, + 0.00899673905223608, + 0.006207482423633337, + 0.013254767283797264, + 0.007922312244772911, + 0.007504560984671116, + 0.005265910644084215, + 0.009077546186745167, + 0.006894040387123823, + 0.006119080353528261, + 0.011396676301956177, + 0.008635522797703743, + 0.012006542645394802, + 0.0051605841144919395, + 0.011435081250965595, + 0.010471353307366371, + 0.014459787867963314, + 0.008803803473711014, + 0.004359337966889143, + 0.008697946555912495, + 0.00708309281617403, + 0.007291626185178757, + 0.01247063186019659, + 0.007010937202721834, + 0.008946637623012066, + 0.005714959464967251, + 0.00400589918717742, + 0.006655847653746605, + 0.005560250021517277, + 0.006253372877836227, + 0.003363523632287979, + 0.0036864103749394417, + 0.004233808256685734, + 0.003223560517653823, + 0.004623555112630129, + 0.0028524177614599466, + 0.004163012374192476, + 0.003702727844938636, + 0.01021962333470583, + 0.006867765914648771, + 0.00759733933955431, + 0.004442143253982067, + 0.012321808375418186, + 0.005269553977996111, + 0.006910095456987619, + 0.0074937776662409306, + 0.02358037419617176, + 0.10947873443365097, + 0.01970553584396839, + 0.018804462626576424, + 0.01169399544596672, + 0.044624775648117065, + 0.006839466281235218, + 0.02500521019101143, + 0.020943019539117813, + 0.013489869423210621, + 0.017647478729486465, + 0.006168699823319912, + 0.0127640999853611, + 0.008536091074347496, + 0.0068007525987923145, + 0.00906793586909771, + 0.018861984834074974, + 0.00880963820964098, + 0.05826278403401375, + 0.01726091466844082, + 0.012312381528317928, + 0.07014411687850952, + 0.00940321572124958, + 0.012600792571902275, + 0.008717751130461693, + 0.009699287824332714, + 0.008289390243589878, + 0.004901405423879623, + 0.01474670134484768, + 0.009901824407279491, + 0.0060875918716192245, + 0.009831211529672146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.018178289756178856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014185545034706593 + ], + "opposite_logprob_tokens": [ + -4.264185428619385 + ], + "logprob_reward": -0.014185545034706593, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014185545034706593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22632034122943878, + 0.19086295366287231, + 0.04214473441243172, + 0.03252730146050453, + 0.09465084969997406, + 0.06733742356300354, + 0.1281254142522812, + 0.04288434982299805, + 0.13744550943374634, + 0.03770112991333008 + ], + "credit_map": [ + 0.0265050046145916, + 0.05118481069803238, + 0.03517291694879532, + 0.02526036463677883, + 0.01728161610662937, + 0.033063724637031555, + 0.023683033883571625, + 0.014168861322104931, + 0.022457754239439964, + 0.017259003594517708, + 0.05314606428146362, + 0.01864890195429325, + 0.01330641284584999, + 0.032623615115880966, + 0.02152980864048004, + 0.011891384609043598, + 0.00442154984921217, + 0.004246834199875593, + 0.0059250518679618835, + 0.005377608351409435, + 0.005483365152031183, + 0.00510752759873867, + 0.005169560667127371, + 0.006413233932107687, + 0.004304856061935425, + 0.003011132590472698, + 0.004938032012432814, + 0.004306023009121418, + 0.003323763143271208, + 0.004164424259215593, + 0.003979962784796953, + 0.004499106202274561, + 0.0244939886033535, + 0.011478830128908157, + 0.019024917855858803, + 0.008190728724002838, + 0.005234607495367527, + 0.014466368593275547, + 0.004515166860073805, + 0.007246232591569424, + 0.008226336911320686, + 0.014843473210930824, + 0.008620304986834526, + 0.005694974213838577, + 0.01049670111387968, + 0.0060209548100829124, + 0.00672488147392869, + 0.006709789391607046, + 0.011235128156840801, + 0.05421625077724457, + 0.008692597039043903, + 0.009119690395891666, + 0.00663309870287776, + 0.01990959793329239, + 0.003184311790391803, + 0.015134727582335472, + 0.01019140426069498, + 0.0067834104411304, + 0.007412930019199848, + 0.0021795190405100584, + 0.004881237167865038, + 0.004516624379903078, + 0.0021685112733393908, + 0.004750706255435944, + 0.018858280032873154, + 0.00793733261525631, + 0.04601411521434784, + 0.014482914470136166, + 0.014042952097952366, + 0.010728963650763035, + 0.01700947992503643, + 0.008371452800929546, + 0.004817023407667875, + 0.005075916647911072, + 0.004368292633444071, + 0.0027686955872923136, + 0.00776643818244338, + 0.004725178238004446, + 0.003302039811387658, + 0.004877541214227676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.014185545034706593 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, the camera slowly and steadily tilts downward, focusing on a warm, light-grained wood tabletop. A slender, manicured hand elegantly grips a classic black lacquer fountain pen. The silver tip glides across the delicate, off-white hardcover journal page, the wrist steady and deliberate, leaving lines of elegant cursive handwriting. Next to the journal, a latte steams in a white ceramic cup, its frothed milk featuring an intricate latte art pattern. Wisps of delicate white steam rise, visible in the light. A ray of fresh morning light slants in from the window, casting a soft, elongated spot on the wooden table, giving the rising steam a textured quality. In the background, the rest of the cafe is softly blurred, with the faint silhouettes of other patrons and the warm interior lighting barely visible. The main colors of the whole picture are warm beige, white and wood color, creating a peaceful, focused and poetic atmosphere, perfectly capturing a leisurely and beautiful morning time.", + "seed": 6377, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_7/index_5132_seed_6377_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03536772970110178, + "vlm_logprob": -0.03536772970110178, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.03536772970110178, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010505258105695248 + ], + "opposite_logprob_tokens": [ + -7.260505199432373 + ], + "logprob_reward": -0.010505258105695248, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.010505258105695248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08499152958393097, + 0.08778078854084015, + 0.01878325268626213, + 0.09124846011400223, + 0.08899600803852081, + 0.05124793201684952, + 0.03693639114499092, + 0.0472584068775177, + 0.12271346151828766, + 0.3700437843799591 + ], + "credit_map": [ + 0.008880091831088066, + 0.00894924160093069, + 0.007473359350115061, + 0.00868993904441595, + 0.015302790328860283, + 0.01111634075641632, + 0.014815337024629116, + 0.00976442638784647, + 0.009527723304927349, + 0.020225966349244118, + 0.006005128379911184, + 0.006941987667232752, + 0.012947088107466698, + 0.012487620115280151, + 0.012915873900055885, + 0.00672940444201231, + 0.002181079937145114, + 0.0035695116966962814, + 0.0016467173118144274, + 0.002098423196002841, + 0.0017820245120674372, + 0.0024306837003678083, + 0.002762788673862815, + 0.0023120231926441193, + 0.011393338441848755, + 0.01062384620308876, + 0.014046953991055489, + 0.01665453612804413, + 0.007807764690369368, + 0.010539562441408634, + 0.010931722819805145, + 0.009250734932720661, + 0.01142470445483923, + 0.009488790296018124, + 0.01655951328575611, + 0.011565018445253372, + 0.00779159227386117, + 0.012514314614236355, + 0.009306272491812706, + 0.010345806367695332, + 0.00495518185198307, + 0.009790351614356041, + 0.00388483889400959, + 0.006435053888708353, + 0.01092708669602871, + 0.006116574630141258, + 0.004325977526605129, + 0.004812866915017366, + 0.0041392589919269085, + 0.004138563759624958, + 0.004319966770708561, + 0.004795121029019356, + 0.003987469710409641, + 0.004090058151632547, + 0.007526341825723648, + 0.003939607180655003, + 0.00669431546702981, + 0.006349207367748022, + 0.004272363614290953, + 0.0066299536265432835, + 0.005332757253199816, + 0.006867514923214912, + 0.005496316123753786, + 0.005615977570414543, + 0.026240134611725807, + 0.012506931088864803, + 0.012342040427029133, + 0.010869812220335007, + 0.01373343262821436, + 0.00948743149638176, + 0.020999237895011902, + 0.016534442082047462, + 0.03889051452279091, + 0.056684352457523346, + 0.02406373806297779, + 0.043816614896059036, + 0.07257720828056335, + 0.04429173469543457, + 0.03232932463288307, + 0.05739029496908188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.010505258105695248 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291978508234024 + ], + "opposite_logprob_tokens": [ + -2.237919807434082 + ], + "logprob_reward": -0.11291978508234024, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291978508234024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0472923219203949, + 0.1289653480052948, + 0.03864135220646858, + 0.06890775263309479, + 0.11319741606712341, + 0.07075660675764084, + 0.17105212807655334, + 0.1107376292347908, + 0.06867887824773788, + 0.18177059292793274 + ], + "credit_map": [ + 0.00567486509680748, + 0.007065440062433481, + 0.005405684933066368, + 0.005095252301543951, + 0.006996511481702328, + 0.005529280751943588, + 0.006565485615283251, + 0.004959798417985439, + 0.018431242555379868, + 0.03177754580974579, + 0.009609677828848362, + 0.011004742234945297, + 0.013198383152484894, + 0.018067704513669014, + 0.01583915576338768, + 0.01103689894080162, + 0.004290054552257061, + 0.004019331652671099, + 0.007467187941074371, + 0.005560952704399824, + 0.0036646153312176466, + 0.004510902799665928, + 0.004365085624158382, + 0.00476321903988719, + 0.008313436061143875, + 0.007060016505420208, + 0.007599580567330122, + 0.00790121778845787, + 0.009579377248883247, + 0.008519001305103302, + 0.01137824822217226, + 0.00855687353760004, + 0.012743404135107994, + 0.010031049139797688, + 0.018087713047862053, + 0.011723801493644714, + 0.013550469651818275, + 0.016326213255524635, + 0.015413341112434864, + 0.01532141026109457, + 0.0070006330497562885, + 0.012013544328510761, + 0.006044596433639526, + 0.010153503157198429, + 0.014183173887431622, + 0.009008757770061493, + 0.005753637291491032, + 0.006598757579922676, + 0.02054072916507721, + 0.01716093346476555, + 0.015849599614739418, + 0.0186996478587389, + 0.01756005361676216, + 0.024561649188399315, + 0.03727218508720398, + 0.019407309591770172, + 0.012096096761524677, + 0.008911242708563805, + 0.00923339743167162, + 0.01583792269229889, + 0.01326893549412489, + 0.017777543514966965, + 0.010544316843152046, + 0.02306816354393959, + 0.017706463113427162, + 0.007075883448123932, + 0.0071511780843138695, + 0.006541081238538027, + 0.00764775276184082, + 0.006333355326205492, + 0.0069340611808001995, + 0.009289100766181946, + 0.022813713178038597, + 0.030905088409781456, + 0.011651486158370972, + 0.01866620033979416, + 0.03836827725172043, + 0.018574535846710205, + 0.017058510333299637, + 0.02373276837170124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.11291978508234024 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011082758195698261 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.011082758195698261, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082758195698261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09817102551460266, + 0.0738014206290245, + 0.03371378779411316, + 0.30859437584877014, + 0.15520063042640686, + 0.04189691320061684, + 0.09540122747421265, + 0.0801980271935463, + 0.045444555580616, + 0.06757802516222 + ], + "credit_map": [ + 0.008334740065038204, + 0.010804564692080021, + 0.009550768882036209, + 0.014412281103432178, + 0.01754111796617508, + 0.01311421301215887, + 0.015101274475455284, + 0.009312069043517113, + 0.008969029411673546, + 0.014057631604373455, + 0.007305940147489309, + 0.00628677848726511, + 0.007711971178650856, + 0.01084827445447445, + 0.010185067541897297, + 0.008436721749603748, + 0.004057332873344421, + 0.003973162267357111, + 0.00734986737370491, + 0.0031908212695270777, + 0.0029056689236313105, + 0.003828975837677717, + 0.0042982217855751514, + 0.00410973746329546, + 0.03209931403398514, + 0.041518405079841614, + 0.03667495399713516, + 0.05207464471459389, + 0.04229448735713959, + 0.0373423732817173, + 0.042414434254169464, + 0.02417575567960739, + 0.01625262387096882, + 0.02159462310373783, + 0.023449545726180077, + 0.02018093876540661, + 0.01437985897064209, + 0.02237558178603649, + 0.019500339403748512, + 0.017467116937041283, + 0.0042908950708806515, + 0.006333686411380768, + 0.004020905587822199, + 0.006739072035998106, + 0.00743305217474699, + 0.004625236615538597, + 0.004020393826067448, + 0.004433671478182077, + 0.014128603041172028, + 0.011558551341295242, + 0.010060344822704792, + 0.013064653612673283, + 0.010896611027419567, + 0.012431205250322819, + 0.011448945850133896, + 0.01181231252849102, + 0.011145472526550293, + 0.008215416222810745, + 0.006226195953786373, + 0.009271464310586452, + 0.00882344413548708, + 0.012608838267624378, + 0.007474175188690424, + 0.016433021053671837, + 0.010675881989300251, + 0.004137406591325998, + 0.004586073104292154, + 0.0031783576123416424, + 0.004110213369131088, + 0.004125717096030712, + 0.007507357746362686, + 0.007123549934476614, + 0.010140319354832172, + 0.008209221996366978, + 0.005854015238583088, + 0.013319117948412895, + 0.007646877784281969, + 0.008209168910980225, + 0.007510122377425432, + 0.006689176894724369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.011082758195698261 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033697061240673065 + ], + "opposite_logprob_tokens": [ + -3.4086971282958984 + ], + "logprob_reward": -0.033697061240673065, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033697061240673065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06512513756752014, + 0.20664048194885254, + 0.08478564769029617, + 0.09142223745584488, + 0.1668042093515396, + 0.043763235211372375, + 0.036798469722270966, + 0.046024076640605927, + 0.048159610480070114, + 0.2104768604040146 + ], + "credit_map": [ + 0.006107546854764223, + 0.007529455237090588, + 0.007591912057250738, + 0.009274006821215153, + 0.007322052493691444, + 0.00859159231185913, + 0.009596270509064198, + 0.009112302213907242, + 0.030519166961312294, + 0.05488021299242973, + 0.014989097602665424, + 0.015663621947169304, + 0.021083714440464973, + 0.028885265812277794, + 0.02122667245566845, + 0.01939273625612259, + 0.010300898924469948, + 0.01406541746109724, + 0.00848736334592104, + 0.00804253201931715, + 0.009143679402768612, + 0.010691738687455654, + 0.011597955599427223, + 0.012456060387194157, + 0.010979372076690197, + 0.0075746504589915276, + 0.015207521617412567, + 0.013275927864015102, + 0.010306364856660366, + 0.010231023654341698, + 0.015153877437114716, + 0.00869350042194128, + 0.016329333186149597, + 0.015704404562711716, + 0.022258734330534935, + 0.01943736895918846, + 0.018615612760186195, + 0.02449607476592064, + 0.027617143467068672, + 0.0223455261439085, + 0.003984346054494381, + 0.006975974887609482, + 0.0037510893307626247, + 0.0061368863098323345, + 0.006899614818394184, + 0.005204944871366024, + 0.005102021619677544, + 0.005708358250558376, + 0.005244199652224779, + 0.0045546418987214565, + 0.0038883264642208815, + 0.005591635126620531, + 0.004063664004206657, + 0.004473106004297733, + 0.004115015733987093, + 0.00486788060516119, + 0.005784962791949511, + 0.004037141799926758, + 0.0034457279834896326, + 0.006258096545934677, + 0.004900749307125807, + 0.007444726303219795, + 0.004190909676253796, + 0.009961762465536594, + 0.011117746122181416, + 0.005320088472217321, + 0.004913892596960068, + 0.003608797676861286, + 0.004609306808561087, + 0.0037762601859867573, + 0.008081070147454739, + 0.006732449401170015, + 0.019175803288817406, + 0.03157917782664299, + 0.013046849519014359, + 0.029071291908621788, + 0.04088369011878967, + 0.02555236779153347, + 0.019009632989764214, + 0.03215803951025009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.033697061240673065 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633785881102085 + ], + "opposite_logprob_tokens": [ + -4.758633613586426 + ], + "logprob_reward": -0.008633785881102085, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633785881102085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3307599723339081, + 0.2021840214729309, + 0.043171294033527374, + 0.12683722376823425, + 0.028117172420024872, + 0.04920268803834915, + 0.02853807434439659, + 0.032121654599905014, + 0.0513865202665329, + 0.10768143087625504 + ], + "credit_map": [ + 0.03708247095346451, + 0.034445032477378845, + 0.032167430967092514, + 0.03135661035776138, + 0.056826151907444, + 0.04221069812774658, + 0.052640400826931, + 0.04403117671608925, + 0.02456926368176937, + 0.04335953667759895, + 0.02154512144625187, + 0.015321226790547371, + 0.022264942526817322, + 0.02860165946185589, + 0.025805065408349037, + 0.020717209205031395, + 0.005181446671485901, + 0.006301350891590118, + 0.004802717827260494, + 0.004362838808447123, + 0.005115303210914135, + 0.005252778995782137, + 0.0053358483128249645, + 0.006819006055593491, + 0.013199307024478912, + 0.014501865953207016, + 0.014560279436409473, + 0.018295180052518845, + 0.026648283004760742, + 0.01433594711124897, + 0.01366081926971674, + 0.011635537259280682, + 0.0031315339729189873, + 0.0025219412054866552, + 0.004863185342401266, + 0.0037961730267852545, + 0.003156526479870081, + 0.003679995657876134, + 0.004063417203724384, + 0.002904401160776615, + 0.004396606236696243, + 0.009769046679139137, + 0.003995075356215239, + 0.006945539265871048, + 0.00913930218666792, + 0.00587147381156683, + 0.004658499266952276, + 0.0044271452352404594, + 0.0027918017003685236, + 0.003136457409709692, + 0.002828864846378565, + 0.0026382291689515114, + 0.003133102785795927, + 0.0035484973341226578, + 0.006868200376629829, + 0.003592920256778598, + 0.0034097193274646997, + 0.0045854090712964535, + 0.002628117334097624, + 0.004293086007237434, + 0.00462671834975481, + 0.004412079229950905, + 0.002685868414118886, + 0.005480656400322914, + 0.009642994031310081, + 0.004720475524663925, + 0.004873036406934261, + 0.004278501961380243, + 0.005899570882320404, + 0.006810153368860483, + 0.009723560884594917, + 0.005438227206468582, + 0.0094972038641572, + 0.01839453913271427, + 0.006069649010896683, + 0.013753307051956654, + 0.02420085296034813, + 0.01159017626196146, + 0.01030375063419342, + 0.013871951028704643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.008633785881102085 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, the camera slowly and steadily tilts downward, focusing on a warm, light-grained wood tabletop. A slender, manicured hand elegantly grips a classic black lacquer fountain pen. The silver tip glides across the delicate, off-white hardcover journal page, the wrist steady and deliberate, leaving lines of elegant cursive handwriting. Next to the journal, a latte steams in a white ceramic cup, its frothed milk featuring an intricate latte art pattern. Wisps of delicate white steam rise, visible in the light. A ray of fresh morning light slants in from the window, casting a soft, elongated spot on the wooden table, giving the rising steam a textured quality. In the background, the rest of the cafe is softly blurred, with the faint silhouettes of other patrons and the warm interior lighting barely visible. The main colors of the whole picture are warm beige, white and wood color, creating a peaceful, focused and poetic atmosphere, perfectly capturing a leisurely and beautiful morning time.", + "seed": 6378, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_7/index_5132_seed_6378_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02746367026120424, + "vlm_logprob": -0.02746367026120424, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.02746367026120424, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015113133005797863 + ], + "opposite_logprob_tokens": [ + -7.015113353729248 + ], + "logprob_reward": -0.015113133005797863, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.015113133005797863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04958318546414375, + 0.02313951961696148, + 0.3875453472137451, + 0.07589281350374222, + 0.05803445354104042, + 0.11979483813047409, + 0.043767210096120834, + 0.10001149028539658, + 0.07947182655334473, + 0.06275935471057892 + ], + "credit_map": [ + 0.006991180591285229, + 0.005678815767168999, + 0.005009678192436695, + 0.004979332443326712, + 0.005882516037672758, + 0.007393811363726854, + 0.00941492523998022, + 0.004232923500239849, + 0.004024213179945946, + 0.002237966749817133, + 0.0026754538994282484, + 0.0024617305025458336, + 0.00309883919544518, + 0.0030948431231081486, + 0.0032261062879115343, + 0.002320367144420743, + 0.02214682102203369, + 0.022635824978351593, + 0.11009345948696136, + 0.0743689090013504, + 0.017680149525403976, + 0.10874505341053009, + 0.017541611567139626, + 0.014333496801555157, + 0.006484484300017357, + 0.007870330475270748, + 0.004775976296514273, + 0.013425173237919807, + 0.0053160530515015125, + 0.02553119882941246, + 0.008416589349508286, + 0.004073009826242924, + 0.0044708410277962685, + 0.003848258638754487, + 0.004559056833386421, + 0.014674900099635124, + 0.00480735395103693, + 0.01671593263745308, + 0.0061364430002868176, + 0.0028216652572155, + 0.016533138230443, + 0.007924186997115612, + 0.009931118227541447, + 0.029332870617508888, + 0.011562878265976906, + 0.014576997607946396, + 0.023067550733685493, + 0.006866099778562784, + 0.003711123252287507, + 0.00506294472143054, + 0.005022888071835041, + 0.0072317770682275295, + 0.006712889298796654, + 0.0039940024726092815, + 0.006841345690190792, + 0.005190238356590271, + 0.01523134857416153, + 0.006068287417292595, + 0.009749808348715305, + 0.02115849032998085, + 0.006895513739436865, + 0.006554559804499149, + 0.023628544062376022, + 0.01072493102401495, + 0.02008664235472679, + 0.00853896327316761, + 0.008191131986677647, + 0.006690331734716892, + 0.015450706705451012, + 0.005632166285067797, + 0.006949256639927626, + 0.007932620123028755, + 0.01392423640936613, + 0.004785357974469662, + 0.01650208607316017, + 0.006242537405341864, + 0.006104903761297464, + 0.0037224802654236555, + 0.005177258979529142, + 0.006300494074821472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 7 + ], + "selected_reward": -0.015113133005797863 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890313908457756 + ], + "opposite_logprob_tokens": [ + -2.464031457901001 + ], + "logprob_reward": -0.0890313908457756, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890313908457756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047732945531606674, + 0.031233401969075203, + 0.30061304569244385, + 0.05171980336308479, + 0.14270929992198944, + 0.1423472762107849, + 0.02238314412534237, + 0.022918174043297768, + 0.1965695023536682, + 0.04177340492606163 + ], + "credit_map": [ + 0.006843680515885353, + 0.004379852209240198, + 0.003863350721076131, + 0.0038201967254281044, + 0.0054954662919044495, + 0.006569026969373226, + 0.010707496665418148, + 0.006053874734789133, + 0.0032167944591492414, + 0.003177241887897253, + 0.003042374039068818, + 0.0028495166916400194, + 0.0043434626422822475, + 0.004387843422591686, + 0.00647630775347352, + 0.0037398627027869225, + 0.017960064113140106, + 0.017360173165798187, + 0.08909469097852707, + 0.06264796853065491, + 0.013823609799146652, + 0.07791981846094131, + 0.012174688279628754, + 0.009632011875510216, + 0.009813806973397732, + 0.0046081868931651115, + 0.004245809279382229, + 0.0067893327213823795, + 0.0040914625860750675, + 0.00880988035351038, + 0.007476875092834234, + 0.005884452722966671, + 0.011777998879551888, + 0.008497553877532482, + 0.011890252120792866, + 0.042438335716724396, + 0.00942279677838087, + 0.037198979407548904, + 0.014363368041813374, + 0.00712002394720912, + 0.01953049749135971, + 0.01200876384973526, + 0.011808404698967934, + 0.03161510080099106, + 0.01565525121986866, + 0.022064026445150375, + 0.01831088215112686, + 0.0113543551415205, + 0.002104915212839842, + 0.0031992867588996887, + 0.003535463474690914, + 0.0026359071489423513, + 0.0033725362736731768, + 0.0018259974895045161, + 0.0030599588062614202, + 0.0026490790769457817, + 0.0024216799065470695, + 0.0017966286977753043, + 0.0022823242470622063, + 0.004828748758882284, + 0.0032860971987247467, + 0.0022563443053513765, + 0.0030886633321642876, + 0.0029576877132058144, + 0.02946213260293007, + 0.016475249081850052, + 0.013471368700265884, + 0.0174558162689209, + 0.030196206644177437, + 0.01396538969129324, + 0.02599063329398632, + 0.04955270513892174, + 0.010186107829213142, + 0.0020745510701090097, + 0.015123237855732441, + 0.004934331402182579, + 0.002526994328945875, + 0.002267788629978895, + 0.00266813556663692, + 0.0019922598730772734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.0890313908457756 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257352255284786 + ], + "opposite_logprob_tokens": [ + -5.2552571296691895 + ], + "logprob_reward": -0.005257352255284786, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257352255284786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07470466196537018, + 0.07515725493431091, + 0.14384539425373077, + 0.02035669796168804, + 0.10458143800497055, + 0.12179189920425415, + 0.05826426297426224, + 0.1690783053636551, + 0.17394466698169708, + 0.058275409042835236 + ], + "credit_map": [ + 0.012484250590205193, + 0.0071237641386687756, + 0.00796897616237402, + 0.006960380356758833, + 0.009707476943731308, + 0.011030798777937889, + 0.010128024034202099, + 0.00930099468678236, + 0.014266513288021088, + 0.008863916620612144, + 0.008576291613280773, + 0.0062870136462152, + 0.012245232239365578, + 0.009781086817383766, + 0.009598265402019024, + 0.0055389367043972015, + 0.008998551405966282, + 0.008503812365233898, + 0.04356330633163452, + 0.03136870637536049, + 0.007563830353319645, + 0.03351615369319916, + 0.006175418850034475, + 0.004155624657869339, + 0.0019150667358189821, + 0.0022492152638733387, + 0.0020865467377007008, + 0.0025535605382174253, + 0.002376517513766885, + 0.0038842889480292797, + 0.002766768215224147, + 0.0025247361045330763, + 0.008589728735387325, + 0.006469842046499252, + 0.008465683087706566, + 0.0306720994412899, + 0.00716894306242466, + 0.028154244646430016, + 0.010323419235646725, + 0.004737485200166702, + 0.02336684986948967, + 0.010482880286872387, + 0.012793350964784622, + 0.02045568823814392, + 0.0145187983289361, + 0.010743049904704094, + 0.021237555891275406, + 0.008193736895918846, + 0.005937909707427025, + 0.0064062983728945255, + 0.006933452561497688, + 0.007692388258874416, + 0.008651433512568474, + 0.006698147859424353, + 0.006638819817453623, + 0.009305820800364017, + 0.02890169620513916, + 0.012218952178955078, + 0.013269341550767422, + 0.039513252675533295, + 0.017820848152041435, + 0.017113063484430313, + 0.027316661551594734, + 0.012924508191645145, + 0.02565847709774971, + 0.01178241241723299, + 0.014036574400961399, + 0.01503363810479641, + 0.026352429762482643, + 0.01296914555132389, + 0.023578913882374763, + 0.04453309252858162, + 0.012664171867072582, + 0.003661038586869836, + 0.018789013847708702, + 0.00585966557264328, + 0.0036867668386548758, + 0.0038137149531394243, + 0.0035094814375042915, + 0.006291558034718037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.005257352255284786 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023282833397388458 + ], + "opposite_logprob_tokens": [ + -3.773282766342163 + ], + "logprob_reward": -0.023282833397388458, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023282833397388458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03553137928247452, + 0.008543812669813633, + 0.09931560605764389, + 0.017743917182087898, + 0.09013111144304276, + 0.09829967468976974, + 0.024328062310814857, + 0.032931920140981674, + 0.14945711195468903, + 0.44371742010116577 + ], + "credit_map": [ + 0.006948601454496384, + 0.003913967404514551, + 0.003671115031465888, + 0.003599681193009019, + 0.003024536184966564, + 0.003400282235816121, + 0.005897894501686096, + 0.0050752959214150906, + 0.0013106894912198186, + 0.0009322234545834363, + 0.000791615282651037, + 0.0008281056070700288, + 0.0013960824580863118, + 0.0010737631237134337, + 0.0014795900788158178, + 0.0007317432318814099, + 0.007496317848563194, + 0.006339114159345627, + 0.029108703136444092, + 0.021526897326111794, + 0.005432050675153732, + 0.020880628377199173, + 0.004942881874740124, + 0.003589014522731304, + 0.002400686964392662, + 0.002290551085025072, + 0.0010490500135347247, + 0.0021347887814044952, + 0.0023608391638845205, + 0.0030592221301048994, + 0.0024258249904960394, + 0.0020229551009833813, + 0.007643046323210001, + 0.005834782030433416, + 0.007339981850236654, + 0.023594480007886887, + 0.0063348799012601376, + 0.02535795047879219, + 0.009364414028823376, + 0.004661580082029104, + 0.017069388180971146, + 0.007471087388694286, + 0.010511757805943489, + 0.01952381432056427, + 0.010131326504051685, + 0.009447131305932999, + 0.017045725136995316, + 0.007099445443600416, + 0.0032479327637702227, + 0.002096722600981593, + 0.0030730890575796366, + 0.003751415526494384, + 0.0036352253518998623, + 0.0017452782485634089, + 0.003980758134275675, + 0.0027976385317742825, + 0.005372656974941492, + 0.0024957344867289066, + 0.0031681847758591175, + 0.007642294745892286, + 0.002804821589961648, + 0.0022310432977974415, + 0.006329615134745836, + 0.002887568436563015, + 0.01327264029532671, + 0.00976222101598978, + 0.013315643183887005, + 0.014940889552235603, + 0.018255086615681648, + 0.011240133084356785, + 0.02333652414381504, + 0.04533397778868675, + 0.10975372046232224, + 0.017844291403889656, + 0.14977888762950897, + 0.041328612715005875, + 0.02004578523337841, + 0.027951616793870926, + 0.024796107783913612, + 0.05221840739250183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.023282833397388458 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily tilt downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, manicured hand gripping a classic black lacquer fountain pen visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a latte with intricate latte art steaming in a white ceramic cup next to the journal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing faint silhouettes of other patrons and warm interior lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633641801774502 + ], + "opposite_logprob_tokens": [ + -5.37963342666626 + ], + "logprob_reward": -0.004633641801774502, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633641801774502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11936003714799881, + 0.04741482064127922, + 0.3706679344177246, + 0.047016214579343796, + 0.08286330103874207, + 0.09414191544055939, + 0.034832913428545, + 0.03661367669701576, + 0.07428359240293503, + 0.0928056389093399 + ], + "credit_map": [ + 0.01419094204902649, + 0.014906737022101879, + 0.011774204671382904, + 0.012867359444499016, + 0.016157638281583786, + 0.019984140992164612, + 0.016727745532989502, + 0.012751261703670025, + 0.006085983011871576, + 0.005051133222877979, + 0.006061015650629997, + 0.0048568956553936005, + 0.008085937239229679, + 0.005373166408389807, + 0.007840296253561974, + 0.004060391802340746, + 0.022267283871769905, + 0.022279413416981697, + 0.11379843205213547, + 0.08048979938030243, + 0.01628706604242325, + 0.0896589607000351, + 0.014823930338025093, + 0.011063015088438988, + 0.00783715769648552, + 0.006448367144912481, + 0.0033442385029047728, + 0.006193597335368395, + 0.00495815509930253, + 0.008765773847699165, + 0.005747615825384855, + 0.0037213079631328583, + 0.008481273427605629, + 0.0056949094869196415, + 0.006549437530338764, + 0.021718807518482208, + 0.00780606409534812, + 0.01908162422478199, + 0.009257828816771507, + 0.004273346625268459, + 0.011222793720662594, + 0.009907897561788559, + 0.008486098609864712, + 0.02226090617477894, + 0.00990025419741869, + 0.010984808206558228, + 0.016553372144699097, + 0.00482577970251441, + 0.004020991735160351, + 0.0028415240813046694, + 0.0057322122156620026, + 0.004197107162326574, + 0.004878485109657049, + 0.004596943501383066, + 0.0037358272820711136, + 0.004829817917197943, + 0.005060824099928141, + 0.002882531611248851, + 0.0035389531403779984, + 0.007240306586027145, + 0.00518212141469121, + 0.0038627926260232925, + 0.004887404851615429, + 0.003958738408982754, + 0.010729588568210602, + 0.005965943913906813, + 0.005658526439219713, + 0.0064134844578802586, + 0.010152079164981842, + 0.0051625752821564674, + 0.010848486796021461, + 0.019352907314896584, + 0.018786519765853882, + 0.006667973939329386, + 0.021141530945897102, + 0.008845419622957706, + 0.005643849726766348, + 0.007360144518315792, + 0.00808059610426426, + 0.016279594972729683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.004633641801774502 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_8/index_tensor([6819, 6819])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_8/index_tensor([6819, 6819])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c9e36e190f95fc799992ddf9b81e95b5e83c3d4c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_8/index_tensor([6819, 6819])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide-angle shot begins a slow, deliberate pan to the right, depicting the immense scale of a massive biomechanical tree against a surreal anime backdrop. The tree's trunk, a complex amalgam of gunmetal-gray metallic bark and interwoven fiber cables, dominates the frame, its surface etched with intricate patterns that emit a soft, internal cyan glow. At its base, vast, translucent root systems glow a vibrant emerald green, visibly channeling streams of light particles from the ground, creating a complex, mesmerizing representation of photosynthesis. As the camera glides, the background reveals a dense, neon-lit cityscape beneath a perpetual twilight, towering skyscrapers covered in holographic advertisements and flying vehicles leaving trails of light. Midway through the pan, a section of the trunk begins to bulge, its metal plates deforming before cracking and expanding outward with mechanical precision. From this newly formed hole, a thick, billowing cloud of golden, shimmering energy was forcefully expelled into the atmosphere, its countless glowing particles swirling and drifting upward, casting a warm, brilliant glow that momentarily overwhelmed the cool, electric tones of the surrounding metropolis, creating a breathtaking atmosphere of technological sublimity.", + "seed": 11933, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_8/index_6819_seed_11933_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01799671626649797, + "vlm_logprob": -0.01799671626649797, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.01799671626649797, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017792263999581337 + ], + "opposite_logprob_tokens": [ + -4.142792224884033 + ], + "logprob_reward": -0.017792263999581337, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.017792263999581337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2293124496936798, + 0.11719857901334763, + 0.04830143600702286, + 0.15385018289089203, + 0.1271689385175705, + 0.057502418756484985, + 0.05803686007857323, + 0.07376134395599365, + 0.09418439120054245, + 0.040683336555957794 + ], + "credit_map": [ + 0.060689814388751984, + 0.017915794625878334, + 0.017672130838036537, + 0.017955491319298744, + 0.06710381805896759, + 0.017012830823659897, + 0.014539924450218678, + 0.016422653570771217, + 0.012971298769116402, + 0.012759885750710964, + 0.01007926743477583, + 0.011855646967887878, + 0.0164516419172287, + 0.025975514203310013, + 0.013542838394641876, + 0.013562497682869434, + 0.004736961331218481, + 0.006709367036819458, + 0.005265916232019663, + 0.004735463764518499, + 0.008838191628456116, + 0.006883015390485525, + 0.004080880433320999, + 0.007051645778119564, + 0.01948065496981144, + 0.02421216107904911, + 0.015116481110453606, + 0.010548917576670647, + 0.015046371147036552, + 0.022105881944298744, + 0.014047985896468163, + 0.033291760832071304, + 0.01994374394416809, + 0.017004091292619705, + 0.013002675026655197, + 0.01380895171314478, + 0.010807910934090614, + 0.01782163791358471, + 0.01032473985105753, + 0.024455199018120766, + 0.008343863300979137, + 0.00578705407679081, + 0.009734994731843472, + 0.005461016669869423, + 0.005335130728781223, + 0.007280826102942228, + 0.004107865039259195, + 0.011451676487922668, + 0.010331209748983383, + 0.005041234660893679, + 0.009832419455051422, + 0.004903156775981188, + 0.007491766009479761, + 0.0069146412424743176, + 0.006871437653899193, + 0.006650999654084444, + 0.009661843068897724, + 0.008437251672148705, + 0.009775646030902863, + 0.006136239040642977, + 0.010290469042956829, + 0.01506852637976408, + 0.005280980374664068, + 0.009110393933951855, + 0.008654826320707798, + 0.01634758897125721, + 0.011919796466827393, + 0.013435626402497292, + 0.013416541740298271, + 0.013777810148894787, + 0.009280409663915634, + 0.0073517970740795135, + 0.0036935624666512012, + 0.004700215067714453, + 0.004168456885963678, + 0.004159674514085054, + 0.0066914744675159454, + 0.008311396464705467, + 0.004966162145137787, + 0.0039923968724906445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.017792263999581337 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005259249359369278 + ], + "opposite_logprob_tokens": [ + -5.255259037017822 + ], + "logprob_reward": -0.005259249359369278, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005259249359369278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3465518057346344, + 0.10407387465238571, + 0.09308988600969315, + 0.10983940958976746, + 0.08106215298175812, + 0.037715110927820206, + 0.050481166690588, + 0.0595555305480957, + 0.0759977251291275, + 0.041633374989032745 + ], + "credit_map": [ + 0.09810580313205719, + 0.0236109159886837, + 0.024913007393479347, + 0.016270872205495834, + 0.10874050855636597, + 0.032322317361831665, + 0.027060803025960922, + 0.015527544543147087, + 0.013356710784137249, + 0.014973610639572144, + 0.010798109695315361, + 0.009884266182780266, + 0.015007753856480122, + 0.016993505880236626, + 0.012589300982654095, + 0.010470624081790447, + 0.011361677199602127, + 0.016045667231082916, + 0.012001542374491692, + 0.009299065917730331, + 0.008807987906038761, + 0.014739954844117165, + 0.007320529315620661, + 0.013513457030057907, + 0.009739461354911327, + 0.018646633252501488, + 0.010890484787523746, + 0.011895191855728626, + 0.012673875316977501, + 0.016276899725198746, + 0.010787631385028362, + 0.01892922818660736, + 0.011235801503062248, + 0.008623099885880947, + 0.0074466075748205185, + 0.010934630408883095, + 0.009590083733201027, + 0.013527585193514824, + 0.008289195597171783, + 0.011415147222578526, + 0.004796931520104408, + 0.006329640746116638, + 0.005026108119636774, + 0.003958315588533878, + 0.004305408336222172, + 0.005010710563510656, + 0.002728744177147746, + 0.005559250246733427, + 0.011653084307909012, + 0.004488272592425346, + 0.009729892015457153, + 0.004031953401863575, + 0.0055381376296281815, + 0.004760687239468098, + 0.004286081530153751, + 0.005993058439344168, + 0.008737735450267792, + 0.007156104315072298, + 0.007712696213275194, + 0.004385090433061123, + 0.00892370380461216, + 0.010998030193150043, + 0.00420573353767395, + 0.007436439394950867, + 0.007944568991661072, + 0.011387424543499947, + 0.009981108829379082, + 0.012209255248308182, + 0.010696860030293465, + 0.010446094907820225, + 0.007281770464032888, + 0.006050639320164919, + 0.002777001354843378, + 0.0052192071452736855, + 0.003221258521080017, + 0.003503667889162898, + 0.010329843498766422, + 0.010316779837012291, + 0.0031582224182784557, + 0.003107395488768816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.005259249359369278 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350723579525948 + ], + "opposite_logprob_tokens": [ + -3.651350736618042 + ], + "logprob_reward": -0.026350723579525948, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350723579525948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26215440034866333, + 0.10672156512737274, + 0.05055272951722145, + 0.1479424387216568, + 0.1878948211669922, + 0.0392477922141552, + 0.0411200150847435, + 0.0861973688006401, + 0.028240298852324486, + 0.049928516149520874 + ], + "credit_map": [ + 0.045467644929885864, + 0.028755348175764084, + 0.02345465123653412, + 0.03001326508820057, + 0.043316226452589035, + 0.03572385013103485, + 0.02967863902449608, + 0.025744784623384476, + 0.00923123862594366, + 0.015957126393914223, + 0.008847981691360474, + 0.008662412874400616, + 0.016295138746500015, + 0.02111719734966755, + 0.011802262626588345, + 0.01480821892619133, + 0.005833531264215708, + 0.006640746723860502, + 0.007776186801493168, + 0.0055817426182329655, + 0.004140672739595175, + 0.007529082242399454, + 0.004086831584572792, + 0.0089639350771904, + 0.01295591238886118, + 0.017968468368053436, + 0.01554223895072937, + 0.013432460837066174, + 0.008478619158267975, + 0.023706601932644844, + 0.014477875083684921, + 0.04138026013970375, + 0.034623973071575165, + 0.02003471367061138, + 0.024255190044641495, + 0.01798192411661148, + 0.011598569341003895, + 0.02644866146147251, + 0.012363000772893429, + 0.04058878868818283, + 0.004733267705887556, + 0.005304676480591297, + 0.004773803520947695, + 0.003768044523894787, + 0.004445353057235479, + 0.005300921853631735, + 0.0027946988120675087, + 0.008127021603286266, + 0.010313617996871471, + 0.002992144785821438, + 0.007637436967343092, + 0.003185129491612315, + 0.004840688779950142, + 0.003644887125119567, + 0.003948308993130922, + 0.004557804670184851, + 0.010087508708238602, + 0.009045138955116272, + 0.011879154480993748, + 0.007678207475692034, + 0.011883608996868134, + 0.01707746647298336, + 0.006643399596214294, + 0.0119028864428401, + 0.0019789261277765036, + 0.0033302847295999527, + 0.002933342708274722, + 0.004007572308182716, + 0.005363760981708765, + 0.005266000982373953, + 0.0027307646814733744, + 0.0026296477299183607, + 0.00471163634210825, + 0.008135277777910233, + 0.005024802405387163, + 0.00467115081846714, + 0.007025253027677536, + 0.00866030901670456, + 0.005689662415534258, + 0.006010427605360746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.026350723579525948 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809213638305664 + ], + "opposite_logprob_tokens": [ + -3.2880921363830566 + ], + "logprob_reward": -0.03809213638305664, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809213638305664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1574685275554657, + 0.052791777998209, + 0.08900397270917892, + 0.09938076883554459, + 0.06244894117116928, + 0.1943337768316269, + 0.07430844753980637, + 0.11568553745746613, + 0.09717845916748047, + 0.057399824261665344 + ], + "credit_map": [ + 0.038824789226055145, + 0.011692122556269169, + 0.014780632220208645, + 0.0110291363671422, + 0.04354792833328247, + 0.013858295045793056, + 0.011940501630306244, + 0.011795133352279663, + 0.004205629229545593, + 0.004194290842860937, + 0.005754449404776096, + 0.005375548265874386, + 0.007996533066034317, + 0.011934061534702778, + 0.006014881655573845, + 0.007316382601857185, + 0.0082554891705513, + 0.010203494690358639, + 0.012186159379780293, + 0.01113495510071516, + 0.011258265003561974, + 0.013065535575151443, + 0.005886494647711515, + 0.017013579607009888, + 0.008905758149921894, + 0.016453353688120842, + 0.009635847993195057, + 0.013026004657149315, + 0.011214097030460835, + 0.013979172334074974, + 0.007225964684039354, + 0.01894056424498558, + 0.008917445316910744, + 0.008191809058189392, + 0.008992605842649937, + 0.008208412677049637, + 0.005494545679539442, + 0.009111813269555569, + 0.005519413389265537, + 0.008012894541025162, + 0.026317236945033073, + 0.029232356697320938, + 0.019768178462982178, + 0.01786954700946808, + 0.019272122532129288, + 0.029457466676831245, + 0.014744091778993607, + 0.037672776728868484, + 0.010515449568629265, + 0.010685805231332779, + 0.009454489685595036, + 0.004885196685791016, + 0.009950967505574226, + 0.009676169604063034, + 0.009203718043863773, + 0.00993665773421526, + 0.009246463887393475, + 0.01094158086925745, + 0.008693523705005646, + 0.005649982485920191, + 0.021246999502182007, + 0.029400287196040154, + 0.01091283280402422, + 0.019593864679336548, + 0.006869422271847725, + 0.011641288176178932, + 0.010186415165662766, + 0.01352016068994999, + 0.02096671238541603, + 0.020490027964115143, + 0.007663162890821695, + 0.005841268692165613, + 0.003747392911463976, + 0.004510337952524424, + 0.003554781898856163, + 0.0028408141806721687, + 0.01685258187353611, + 0.017365919426083565, + 0.004079891834408045, + 0.004448105115443468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.03809213638305664 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002489208010956645 + ], + "opposite_logprob_tokens": [ + -6.00248908996582 + ], + "logprob_reward": -0.002489208010956645, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002489208010956645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10068655759096146, + 0.042020730674266815, + 0.04155619069933891, + 0.07732784003019333, + 0.07164400070905685, + 0.0877545177936554, + 0.048703938722610474, + 0.23707376420497894, + 0.18610170483589172, + 0.10713080316781998 + ], + "credit_map": [ + 0.016705220565199852, + 0.013274379074573517, + 0.010792148299515247, + 0.013921048492193222, + 0.014065112918615341, + 0.011806359514594078, + 0.007035921793431044, + 0.013086370192468166, + 0.0038064057007431984, + 0.004586394876241684, + 0.005160967353731394, + 0.004020286723971367, + 0.0062621235847473145, + 0.007460406050086021, + 0.006190045271068811, + 0.004534102976322174, + 0.0046020494773983955, + 0.006659271661192179, + 0.004197555594146252, + 0.005866244435310364, + 0.006705897860229015, + 0.006088736467063427, + 0.0031974483281373978, + 0.004238988272845745, + 0.007627211976796389, + 0.010398447513580322, + 0.008201012387871742, + 0.009819678030908108, + 0.012403232045471668, + 0.010773680172860622, + 0.010399147868156433, + 0.0077054305002093315, + 0.013472994789481163, + 0.005900165531784296, + 0.009555897675454617, + 0.007531929295510054, + 0.006772024556994438, + 0.008837750181555748, + 0.0058734966441988945, + 0.013699747622013092, + 0.009787187911570072, + 0.012366772629320621, + 0.015003710053861141, + 0.009787067770957947, + 0.008039596490561962, + 0.012857293710112572, + 0.006933094467967749, + 0.012979794293642044, + 0.009069978259503841, + 0.0036021156702190638, + 0.009495432488620281, + 0.004865713883191347, + 0.005986584350466728, + 0.006833384279161692, + 0.0040836757980287075, + 0.004767056554555893, + 0.018629033118486404, + 0.016903989017009735, + 0.01711784116923809, + 0.010795243084430695, + 0.040884923189878464, + 0.06110081449151039, + 0.033081136643886566, + 0.03856077790260315, + 0.013891932554543018, + 0.01805366948246956, + 0.01773878000676632, + 0.020352520048618317, + 0.041888527572155, + 0.04673900455236435, + 0.015036369673907757, + 0.012400892563164234, + 0.007137497887015343, + 0.01154668815433979, + 0.005335466004908085, + 0.00602820236235857, + 0.030291182920336723, + 0.030794020742177963, + 0.00908514205366373, + 0.006912603043019772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.002489208010956645 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot begins a slow, deliberate pan to the right, depicting the immense scale of a massive biomechanical tree against a surreal anime backdrop. The tree's trunk, a complex amalgam of gunmetal-gray metallic bark and interwoven fiber cables, dominates the frame, its surface etched with intricate patterns that emit a soft, internal cyan glow. At its base, vast, translucent root systems glow a vibrant emerald green, visibly channeling streams of light particles from the ground, creating a complex, mesmerizing representation of photosynthesis. As the camera glides, the background reveals a dense, neon-lit cityscape beneath a perpetual twilight, towering skyscrapers covered in holographic advertisements and flying vehicles leaving trails of light. Midway through the pan, a section of the trunk begins to bulge, its metal plates deforming before cracking and expanding outward with mechanical precision. From this newly formed hole, a thick, billowing cloud of golden, shimmering energy was forcefully expelled into the atmosphere, its countless glowing particles swirling and drifting upward, casting a warm, brilliant glow that momentarily overwhelmed the cool, electric tones of the surrounding metropolis, creating a breathtaking atmosphere of technological sublimity.", + "seed": 11934, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_8/index_6819_seed_11934_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03691235282458365, + "vlm_logprob": -0.03691235282458365, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.03691235282458365, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0799991711974144 + ], + "opposite_logprob_tokens": [ + -2.5799992084503174 + ], + "logprob_reward": -0.0799991711974144, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0799991711974144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09658009558916092, + 0.024699384346604347, + 0.013684754259884357, + 0.026862476021051407, + 0.07069896161556244, + 0.42322176694869995, + 0.0481952503323555, + 0.16710643470287323, + 0.03659906983375549, + 0.09235178679227829 + ], + "credit_map": [ + 0.013966682367026806, + 0.01457453053444624, + 0.01347715500742197, + 0.013250867836177349, + 0.010103846900165081, + 0.008674453012645245, + 0.012000827118754387, + 0.010531743057072163, + 0.0033497631084173918, + 0.003954768646508455, + 0.003365485928952694, + 0.004714553244411945, + 0.0030264598317444324, + 0.001712318859063089, + 0.002254930790513754, + 0.0023211073130369186, + 0.001955342711880803, + 0.001746951718814671, + 0.0016189385205507278, + 0.0017064422136172652, + 0.0010947604896500707, + 0.0015840553678572178, + 0.002433929592370987, + 0.001544334925711155, + 0.0036020942497998476, + 0.0038783943746238947, + 0.004026179201900959, + 0.0029177095275372267, + 0.004006443079560995, + 0.0031186649575829506, + 0.002467527985572815, + 0.002845464274287224, + 0.009777809493243694, + 0.007915946654975414, + 0.009540915489196777, + 0.006754481699317694, + 0.009607750922441483, + 0.010180910117924213, + 0.005837552715092897, + 0.011083601042628288, + 0.021371541544795036, + 0.1163548156619072, + 0.025960173457860947, + 0.16552597284317017, + 0.01836061105132103, + 0.018440714105963707, + 0.02340574935078621, + 0.033802203834056854, + 0.004663417115807533, + 0.0051599168218672276, + 0.002954817144200206, + 0.004038932267576456, + 0.00738548906520009, + 0.00748027628287673, + 0.011427395977079868, + 0.005085010547190905, + 0.019042227417230606, + 0.024627257138490677, + 0.015133323147892952, + 0.020711181685328484, + 0.020259089767932892, + 0.021733839064836502, + 0.024092260748147964, + 0.021507274359464645, + 0.0017380336066707969, + 0.009659758768975735, + 0.0019362610764801502, + 0.005958189722150564, + 0.008623780682682991, + 0.002561222994700074, + 0.004310766700655222, + 0.00181105628143996, + 0.00706794997677207, + 0.008965977467596531, + 0.006828486453741789, + 0.009952706284821033, + 0.010801339522004128, + 0.025889858603477478, + 0.008352760225534439, + 0.01449271198362112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.0799991711974144 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028263882268220186 + ], + "opposite_logprob_tokens": [ + -5.87782621383667 + ], + "logprob_reward": -0.0028263882268220186, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028263882268220186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14444074034690857, + 0.053416188806295395, + 0.17236390709877014, + 0.03390706703066826, + 0.04613421484827995, + 0.22467041015625, + 0.059715382754802704, + 0.09722258150577545, + 0.032826196402311325, + 0.1353033035993576 + ], + "credit_map": [ + 0.016460241749882698, + 0.010790960863232613, + 0.024201616644859314, + 0.021164320409297943, + 0.016228828579187393, + 0.018902834504842758, + 0.02090466395020485, + 0.01578727923333645, + 0.005380480084568262, + 0.006284289062023163, + 0.008001004345715046, + 0.006605134345591068, + 0.007837001234292984, + 0.006552391219884157, + 0.006208100821822882, + 0.006547789089381695, + 0.016596242785453796, + 0.013040982186794281, + 0.023307960480451584, + 0.014031813479959965, + 0.016994502395391464, + 0.026781555265188217, + 0.03699352964758873, + 0.024617327377200127, + 0.005660783965140581, + 0.007443134672939777, + 0.003963489551097155, + 0.002365163527429104, + 0.004706711042672396, + 0.003871146822348237, + 0.003020750591531396, + 0.002875888254493475, + 0.006812054663896561, + 0.005582139361649752, + 0.0066566504538059235, + 0.004512146580964327, + 0.00658591790124774, + 0.005908044055104256, + 0.003617280162870884, + 0.006459981203079224, + 0.017912551760673523, + 0.025519173592329025, + 0.018869200721383095, + 0.0837857574224472, + 0.012294354848563671, + 0.017309635877609253, + 0.011814654804766178, + 0.037165071815252304, + 0.0053363461047410965, + 0.006742922123521566, + 0.0038161601405590773, + 0.005905375350266695, + 0.007607532199472189, + 0.014830764383077621, + 0.01053931936621666, + 0.004936965182423592, + 0.007982882671058178, + 0.017094811424613, + 0.007039796561002731, + 0.009845936670899391, + 0.011696944013237953, + 0.016077617183327675, + 0.011905171908438206, + 0.015579426661133766, + 0.0020288547966629267, + 0.007018732372671366, + 0.0020031307358294725, + 0.0057220254093408585, + 0.005679660476744175, + 0.0036518611013889313, + 0.0036626935470849276, + 0.0030592388939112425, + 0.00950209517031908, + 0.014495987445116043, + 0.008411368355154991, + 0.015447388403117657, + 0.016454143449664116, + 0.03287675976753235, + 0.01304356288164854, + 0.025071999058127403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.0028263882268220186 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585225895047188 + ], + "opposite_logprob_tokens": [ + -3.895585298538208 + ], + "logprob_reward": -0.020585225895047188, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585225895047188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1754855066537857, + 0.07074276357889175, + 0.07382820546627045, + 0.07119953632354736, + 0.07249362021684647, + 0.12967278063297272, + 0.1004483550786972, + 0.11449997872114182, + 0.044039640575647354, + 0.1475895643234253 + ], + "credit_map": [ + 0.03691484406590462, + 0.02619718573987484, + 0.02918778732419014, + 0.02282031625509262, + 0.024982141330838203, + 0.008148043416440487, + 0.016127031296491623, + 0.011108164675533772, + 0.013688884675502777, + 0.01348042767494917, + 0.008938894607126713, + 0.0074411416426301, + 0.008771506138145924, + 0.00511698704212904, + 0.006014337297528982, + 0.007290585897862911, + 0.011220860294997692, + 0.012357876636087894, + 0.008045816794037819, + 0.008042135275900364, + 0.004707677289843559, + 0.007055420894175768, + 0.015451081097126007, + 0.00694733951240778, + 0.012402081862092018, + 0.010897978208959103, + 0.008028230629861355, + 0.009665820747613907, + 0.006920796353369951, + 0.008725089952349663, + 0.0054480405524373055, + 0.009111502207815647, + 0.014480733312666416, + 0.011233467608690262, + 0.007993225008249283, + 0.005814566742628813, + 0.009952033869922161, + 0.009452913887798786, + 0.0043167793191969395, + 0.009249905124306679, + 0.006823762319982052, + 0.04140939190983772, + 0.006377606187015772, + 0.04470020532608032, + 0.007346489001065493, + 0.004806762561202049, + 0.007383378688246012, + 0.010825184173882008, + 0.010865269228816032, + 0.010694745928049088, + 0.007220807019621134, + 0.006788412109017372, + 0.015187248587608337, + 0.02297474816441536, + 0.016662832349538803, + 0.010054303333163261, + 0.009051674976944923, + 0.017539478838443756, + 0.005421396344900131, + 0.011929202824831009, + 0.02415873110294342, + 0.012346097268164158, + 0.023929044604301453, + 0.010124357417225838, + 0.002175798872485757, + 0.008962588384747505, + 0.002656068652868271, + 0.010736381635069847, + 0.007583041209727526, + 0.0031859595328569412, + 0.0063942535780370235, + 0.0023455508053302765, + 0.010805982165038586, + 0.012731039896607399, + 0.008846691809594631, + 0.015780210494995117, + 0.016461145132780075, + 0.044822029769420624, + 0.013184855692088604, + 0.02495761588215828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.020585225895047188 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305689036846161 + ], + "opposite_logprob_tokens": [ + -3.1680569648742676 + ], + "logprob_reward": -0.04305689036846161, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305689036846161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.126296728849411, + 0.0615512914955616, + 0.05796268209815025, + 0.0624871663749218, + 0.051871780306100845, + 0.20147928595542908, + 0.10272669047117233, + 0.12508848309516907, + 0.050164271146059036, + 0.16037170588970184 + ], + "credit_map": [ + 0.024007931351661682, + 0.01585162617266178, + 0.019071124494075775, + 0.016517970710992813, + 0.018849600106477737, + 0.007959767244756222, + 0.015406203456223011, + 0.008632499724626541, + 0.005825049243867397, + 0.004913386423140764, + 0.008812841959297657, + 0.01595407910645008, + 0.006606683135032654, + 0.005665592383593321, + 0.008439001627266407, + 0.005334661342203617, + 0.008270121179521084, + 0.005868867505341768, + 0.006058350671082735, + 0.004054654855281115, + 0.005147003103047609, + 0.007794650737196207, + 0.016761401668190956, + 0.004007629584521055, + 0.007307750638574362, + 0.008336490020155907, + 0.008511640131473541, + 0.006715505849570036, + 0.009671005420386791, + 0.007937852293252945, + 0.005766756366938353, + 0.008240162394940853, + 0.010083474218845367, + 0.0076783522963523865, + 0.00601236755028367, + 0.004239059519022703, + 0.0067541035823524, + 0.0061914739198982716, + 0.004688038025051355, + 0.006224912591278553, + 0.01531601045280695, + 0.03769683837890625, + 0.024402542039752007, + 0.06876104325056076, + 0.009094134904444218, + 0.012437033466994762, + 0.01175026036798954, + 0.02202141471207142, + 0.010523850098252296, + 0.011491035111248493, + 0.008205611258745193, + 0.008876234292984009, + 0.014706611633300781, + 0.02505652606487274, + 0.014346690848469734, + 0.009520130231976509, + 0.009695932269096375, + 0.015334944240748882, + 0.006784486584365368, + 0.012384557165205479, + 0.024065284058451653, + 0.013149394653737545, + 0.02873481623828411, + 0.014939060434699059, + 0.003769147442653775, + 0.007367465179413557, + 0.003410817589610815, + 0.008089677430689335, + 0.006582173984497786, + 0.007460056804120541, + 0.007956006564199924, + 0.005528926383703947, + 0.009733804501593113, + 0.025090184062719345, + 0.006768015678972006, + 0.016641106456518173, + 0.025444407016038895, + 0.038679324090480804, + 0.012518307194113731, + 0.025496551766991615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.04305689036846161 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038094088435173035 + ], + "opposite_logprob_tokens": [ + -3.2880940437316895 + ], + "logprob_reward": -0.038094088435173035, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038094088435173035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06076645106077194, + 0.04504233971238136, + 0.029663659632205963, + 0.04607699438929558, + 0.029929233714938164, + 0.3237414062023163, + 0.06075207516551018, + 0.19359655678272247, + 0.03547407314181328, + 0.17495721578598022 + ], + "credit_map": [ + 0.008811313658952713, + 0.010652878321707249, + 0.007633798755705357, + 0.006115060765296221, + 0.007361104246228933, + 0.004963185638189316, + 0.008345795795321465, + 0.0068833171389997005, + 0.004713121801614761, + 0.004499006550759077, + 0.006101301871240139, + 0.007197869010269642, + 0.006039787083864212, + 0.005969075486063957, + 0.0059657315723598, + 0.0045564486645162106, + 0.0046822018921375275, + 0.003522396320477128, + 0.0030698657501488924, + 0.002929145935922861, + 0.002336321398615837, + 0.003678356297314167, + 0.006774663459509611, + 0.0026707116048783064, + 0.007070410996675491, + 0.008593990467488766, + 0.005507712252438068, + 0.003998857922852039, + 0.005594079382717609, + 0.005290362052619457, + 0.004316539037972689, + 0.005705044139176607, + 0.005454402416944504, + 0.00419235322624445, + 0.003917815629392862, + 0.0025910669937729836, + 0.003929764963686466, + 0.004416171461343765, + 0.00244181789457798, + 0.002985844388604164, + 0.023326849564909935, + 0.04225234314799309, + 0.024710144847631454, + 0.11966293305158615, + 0.01953987404704094, + 0.02594747766852379, + 0.01736539416015148, + 0.05093641206622124, + 0.00456857169046998, + 0.0077442508190870285, + 0.0037920870818197727, + 0.004101678263396025, + 0.007162054069340229, + 0.007071655243635178, + 0.020465733483433723, + 0.005846046842634678, + 0.01448866818100214, + 0.02189168706536293, + 0.009482515044510365, + 0.009863171726465225, + 0.032496143132448196, + 0.03406910225749016, + 0.0408104732632637, + 0.030494797974824905, + 0.0022757796104997396, + 0.005119075998663902, + 0.0025725504383444786, + 0.006260504946112633, + 0.004694695584475994, + 0.005469165742397308, + 0.004490716848522425, + 0.004591582342982292, + 0.009976761415600777, + 0.02847946807742119, + 0.009324837476015091, + 0.018467936664819717, + 0.029313769191503525, + 0.029667656868696213, + 0.014787046238780022, + 0.03493974730372429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.038094088435173035 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot begins a slow, deliberate pan to the right, depicting the immense scale of a massive biomechanical tree against a surreal anime backdrop. The tree's trunk, a complex amalgam of gunmetal-gray metallic bark and interwoven fiber cables, dominates the frame, its surface etched with intricate patterns that emit a soft, internal cyan glow. At its base, vast, translucent root systems glow a vibrant emerald green, visibly channeling streams of light particles from the ground, creating a complex, mesmerizing representation of photosynthesis. As the camera glides, the background reveals a dense, neon-lit cityscape beneath a perpetual twilight, towering skyscrapers covered in holographic advertisements and flying vehicles leaving trails of light. Midway through the pan, a section of the trunk begins to bulge, its metal plates deforming before cracking and expanding outward with mechanical precision. From this newly formed hole, a thick, billowing cloud of golden, shimmering energy was forcefully expelled into the atmosphere, its countless glowing particles swirling and drifting upward, casting a warm, brilliant glow that momentarily overwhelmed the cool, electric tones of the surrounding metropolis, creating a breathtaking atmosphere of technological sublimity.", + "seed": 11935, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_8/index_6819_seed_11935_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03283794904127717, + "vlm_logprob": -0.03283794904127717, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.03283794904127717, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03942491486668587 + ], + "opposite_logprob_tokens": [ + -3.2894248962402344 + ], + "logprob_reward": -0.03942491486668587, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03942491486668587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14998087286949158, + 0.14881563186645508, + 0.040639497339725494, + 0.18303555250167847, + 0.06726717203855515, + 0.04525022208690643, + 0.2282577008008957, + 0.04943842440843582, + 0.030345896258950233, + 0.05696900933980942 + ], + "credit_map": [ + 0.017004111781716347, + 0.020253656432032585, + 0.01979255862534046, + 0.012541224248707294, + 0.01769436150789261, + 0.01915423572063446, + 0.018150074407458305, + 0.02539065293967724, + 0.020428931340575218, + 0.019413292407989502, + 0.024115677922964096, + 0.01744334027171135, + 0.014626126736402512, + 0.016812317073345184, + 0.0229311715811491, + 0.013044768944382668, + 0.009214979596436024, + 0.004369234666228294, + 0.004851415287703276, + 0.0025063150096684694, + 0.004115855786949396, + 0.005931657273322344, + 0.004041489213705063, + 0.005608551669865847, + 0.035918593406677246, + 0.008561733178794384, + 0.006011270917952061, + 0.007063637021929026, + 0.07301615178585052, + 0.026390599086880684, + 0.010996183380484581, + 0.015077377669513226, + 0.010302177630364895, + 0.004559312015771866, + 0.004199560731649399, + 0.005531310103833675, + 0.012233726680278778, + 0.013916942290961742, + 0.009927372448146343, + 0.006596773397177458, + 0.007265596184879541, + 0.005273993127048016, + 0.004546561744064093, + 0.0066024442203342915, + 0.004421125631779432, + 0.005070713348686695, + 0.007097126916050911, + 0.0049726637080311775, + 0.021958032622933388, + 0.022256478667259216, + 0.03166964650154114, + 0.017131127417087555, + 0.08599159866571426, + 0.01462873537093401, + 0.02360069938004017, + 0.01102138590067625, + 0.00547774787992239, + 0.004391987342387438, + 0.006100654602050781, + 0.004374894313514233, + 0.007830483838915825, + 0.005765336565673351, + 0.012437974102795124, + 0.0030593490228056908, + 0.002612423850223422, + 0.0029180082492530346, + 0.0016890591941773891, + 0.0027289772406220436, + 0.007176995277404785, + 0.005317118484526873, + 0.004486312624067068, + 0.0034170001745224, + 0.003537661163136363, + 0.009339329786598682, + 0.004953810945153236, + 0.00875687412917614, + 0.0072518750093877316, + 0.004197988193482161, + 0.0119114825502038, + 0.007019989658147097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.03942491486668587 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018201233819127083 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.018201233819127083, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018201233819127083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15890762209892273, + 0.029459919780492783, + 0.04487039893865585, + 0.3485369384288788, + 0.10145261883735657, + 0.04644080996513367, + 0.057591572403907776, + 0.10454930365085602, + 0.03828679770231247, + 0.0699039101600647 + ], + "credit_map": [ + 0.019207995384931564, + 0.019965503364801407, + 0.016464663669466972, + 0.021667122840881348, + 0.024050407111644745, + 0.021912045776844025, + 0.018777718767523766, + 0.016862189397215843, + 0.005419161636382341, + 0.0039800009690225124, + 0.002955697476863861, + 0.003052821382880211, + 0.0031529394909739494, + 0.004015917424112558, + 0.003750208066776395, + 0.003133179387077689, + 0.010344145819544792, + 0.004334229044616222, + 0.005861645098775625, + 0.0037304142024368048, + 0.004730521235615015, + 0.005647765472531319, + 0.005060703959316015, + 0.005160984117537737, + 0.07008396834135056, + 0.018271397799253464, + 0.011834044009447098, + 0.01132893655449152, + 0.143077090382576, + 0.05259914696216583, + 0.0215306356549263, + 0.01981177367269993, + 0.011880907230079174, + 0.012712731026113033, + 0.00943723414093256, + 0.012685030698776245, + 0.013785749673843384, + 0.013746439479291439, + 0.018223639577627182, + 0.008980906568467617, + 0.00600843271240592, + 0.01341680996119976, + 0.003347775200381875, + 0.004002631641924381, + 0.004707426764070988, + 0.006541589740663767, + 0.005486523266881704, + 0.0029296353459358215, + 0.00464193569496274, + 0.0052978526800870895, + 0.008296296931803226, + 0.004587309900671244, + 0.015588982030749321, + 0.008022299967706203, + 0.006563118193298578, + 0.004593786783516407, + 0.011784248985350132, + 0.009271081537008286, + 0.009077534079551697, + 0.008120689541101456, + 0.020654994994401932, + 0.022737009450793266, + 0.012781163677573204, + 0.010122609324753284, + 0.0035720316227525473, + 0.005469655618071556, + 0.0033044111914932728, + 0.004357805009931326, + 0.007145162206143141, + 0.0041219862177968025, + 0.00576476426795125, + 0.00455098832026124, + 0.009522732347249985, + 0.008766821585595608, + 0.007186930626630783, + 0.008593258447945118, + 0.010119324550032616, + 0.005537566728889942, + 0.01065949909389019, + 0.009517795406281948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.018201233819127083 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700334906578064 + ], + "opposite_logprob_tokens": [ + -2.695033550262451 + ], + "logprob_reward": -0.0700334906578064, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700334906578064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17958341538906097, + 0.16730618476867676, + 0.041602954268455505, + 0.2651987671852112, + 0.06228535994887352, + 0.024733245372772217, + 0.09032709896564484, + 0.033525727689266205, + 0.08625427633523941, + 0.04918298497796059 + ], + "credit_map": [ + 0.028621671721339226, + 0.04216460883617401, + 0.016829580068588257, + 0.01433323509991169, + 0.014495482668280602, + 0.032015036791563034, + 0.014574852772057056, + 0.016548937186598778, + 0.024222267791628838, + 0.020485343411564827, + 0.021253280341625214, + 0.020404357463121414, + 0.013952309265732765, + 0.01739788055419922, + 0.031202265992760658, + 0.018388478085398674, + 0.011876130476593971, + 0.005504193250089884, + 0.005999837536364794, + 0.002506502205505967, + 0.00378772197291255, + 0.005129106808453798, + 0.0034407542552798986, + 0.0033587110228836536, + 0.058466214686632156, + 0.011210856959223747, + 0.007621855475008488, + 0.008080720901489258, + 0.11050771176815033, + 0.04024466127157211, + 0.017602089792490005, + 0.011464661918580532, + 0.010399159975349903, + 0.004897479899227619, + 0.004601991269737482, + 0.005462882108986378, + 0.008553871884942055, + 0.01269533485174179, + 0.009414449334144592, + 0.006260187830775976, + 0.0021267880219966173, + 0.0022682081907987595, + 0.002882166998460889, + 0.0037483207415789366, + 0.0018784805433824658, + 0.0033773125614970922, + 0.0053796893917024136, + 0.0030722790397703648, + 0.009173773229122162, + 0.008639032952487469, + 0.01168488897383213, + 0.007534094620496035, + 0.032176386564970016, + 0.0071902028284966946, + 0.009000626392662525, + 0.004928096663206816, + 0.003956600558012724, + 0.002593323355540633, + 0.004602409899234772, + 0.002967328066006303, + 0.0051831346936523914, + 0.0048646386712789536, + 0.0070177302695810795, + 0.002340562641620636, + 0.009187205694615841, + 0.007091993000358343, + 0.008844071999192238, + 0.009025135077536106, + 0.012158973142504692, + 0.01958586648106575, + 0.011780095286667347, + 0.008580937050282955, + 0.0031536021269857883, + 0.007517373189330101, + 0.004614691715687513, + 0.008676157332956791, + 0.00627616373822093, + 0.004107882734388113, + 0.008743115700781345, + 0.006093999370932579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0700334906578064 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794272035360336 + ], + "opposite_logprob_tokens": [ + -3.529794216156006 + ], + "logprob_reward": -0.029794272035360336, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794272035360336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06942163407802582, + 0.019081827253103256, + 0.019813161343336105, + 0.32465800642967224, + 0.036045998334884644, + 0.03511790186166763, + 0.3428274989128113, + 0.030113719403743744, + 0.10495656728744507, + 0.01796369068324566 + ], + "credit_map": [ + 0.008498608134686947, + 0.010747676715254784, + 0.007265892345458269, + 0.005333833862096071, + 0.005300683900713921, + 0.007101279683411121, + 0.010646142065525055, + 0.014527513645589352, + 0.0033093122765421867, + 0.0027755540795624256, + 0.002220384543761611, + 0.0015183590585365891, + 0.0024457983672618866, + 0.0024759136140346527, + 0.0023574780207127333, + 0.0019790285732597113, + 0.003326138248667121, + 0.0025443113408982754, + 0.0022724284790456295, + 0.0018946562195196748, + 0.002927235094830394, + 0.0023465119302272797, + 0.002078254707157612, + 0.0024236261378973722, + 0.06901873648166656, + 0.013681088574230671, + 0.009052335284650326, + 0.010096934624016285, + 0.13612201809883118, + 0.048887692391872406, + 0.021422168239951134, + 0.016377035528421402, + 0.004702947102487087, + 0.004277953878045082, + 0.0025377182755619287, + 0.004701685160398483, + 0.00515403738245368, + 0.0050218598917126656, + 0.006415063049644232, + 0.0032347342930734158, + 0.005615789908915758, + 0.007344396784901619, + 0.0026706422213464975, + 0.0030541548039764166, + 0.0030907716136425734, + 0.004135075956583023, + 0.005725095514208078, + 0.0034819727297872305, + 0.03098299913108349, + 0.030373116955161095, + 0.05250122770667076, + 0.019475147128105164, + 0.13100479543209076, + 0.025374649092555046, + 0.03612205758690834, + 0.016993489116430283, + 0.003707712050527334, + 0.0024536033160984516, + 0.0021326555870473385, + 0.0022754124365746975, + 0.004913188982754946, + 0.009141580201685429, + 0.0032593298237770796, + 0.002230235608294606, + 0.009556440636515617, + 0.011422961950302124, + 0.004346299450844526, + 0.006832864601165056, + 0.027035046368837357, + 0.016303539276123047, + 0.019515354186296463, + 0.009944059886038303, + 0.0013524809619411826, + 0.0022175651974976063, + 0.0018207813845947385, + 0.0022374235559254885, + 0.002794076921418309, + 0.002424531616270542, + 0.0027739168144762516, + 0.002342914929613471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.029794272035360336 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735833827406168 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735833827406168, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735833827406168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06288572400808334, + 0.06306318193674088, + 0.015964794903993607, + 0.06473898142576218, + 0.025507140904664993, + 0.028402024880051613, + 0.23478904366493225, + 0.38597169518470764, + 0.0515604093670845, + 0.06711707264184952 + ], + "credit_map": [ + 0.008631682023406029, + 0.013623972423374653, + 0.006847563665360212, + 0.005723257083445787, + 0.006301649380475283, + 0.008211002685129642, + 0.006092017516493797, + 0.007454581093043089, + 0.008613994345068932, + 0.007357984781265259, + 0.006516813766211271, + 0.00792114157229662, + 0.005874634254723787, + 0.006838883738964796, + 0.012473768554627895, + 0.00746595673263073, + 0.0043236189521849155, + 0.0014897361397743225, + 0.0019941742066293955, + 0.0011372978333383799, + 0.0014272895641624928, + 0.0020916617941111326, + 0.001467980444431305, + 0.002033035270869732, + 0.010871924459934235, + 0.0029838986229151487, + 0.0019560896325856447, + 0.0023349199909716845, + 0.026110514998435974, + 0.009986771270632744, + 0.004543586168438196, + 0.005951272323727608, + 0.003383105155080557, + 0.0021151213441044092, + 0.002123145619407296, + 0.0028512163553386927, + 0.0038010585121810436, + 0.004367673769593239, + 0.002993914531543851, + 0.0038719044532626867, + 0.003981816116720438, + 0.004038266371935606, + 0.002221975941210985, + 0.0030273727606981993, + 0.003615014022216201, + 0.003411121666431427, + 0.005746175535023212, + 0.0023602836299687624, + 0.020603451877832413, + 0.020627181977033615, + 0.03522394597530365, + 0.01336066983640194, + 0.08981194347143173, + 0.017259808257222176, + 0.02618771232664585, + 0.01171433087438345, + 0.04092930257320404, + 0.034634221345186234, + 0.03655514493584633, + 0.03251959756016731, + 0.08068233728408813, + 0.08684323728084564, + 0.03789331391453743, + 0.03591454029083252, + 0.004675910808146, + 0.006609676405787468, + 0.0022447602823376656, + 0.0031416048295795918, + 0.013007275760173798, + 0.007924288511276245, + 0.010375299490988255, + 0.0035815960727632046, + 0.004092734307050705, + 0.010285626165568829, + 0.005610730033367872, + 0.006887896452099085, + 0.010554267093539238, + 0.007380831986665726, + 0.013714486733078957, + 0.008590496145188808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.006735833827406168 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot begins a slow, deliberate pan to the right, depicting the immense scale of a massive biomechanical tree against a surreal anime backdrop. The tree's trunk, a complex amalgam of gunmetal-gray metallic bark and interwoven fiber cables, dominates the frame, its surface etched with intricate patterns that emit a soft, internal cyan glow. At its base, vast, translucent root systems glow a vibrant emerald green, visibly channeling streams of light particles from the ground, creating a complex, mesmerizing representation of photosynthesis. As the camera glides, the background reveals a dense, neon-lit cityscape beneath a perpetual twilight, towering skyscrapers covered in holographic advertisements and flying vehicles leaving trails of light. Midway through the pan, a section of the trunk begins to bulge, its metal plates deforming before cracking and expanding outward with mechanical precision. From this newly formed hole, a thick, billowing cloud of golden, shimmering energy was forcefully expelled into the atmosphere, its countless glowing particles swirling and drifting upward, casting a warm, brilliant glow that momentarily overwhelmed the cool, electric tones of the surrounding metropolis, creating a breathtaking atmosphere of technological sublimity.", + "seed": 11936, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_8/index_6819_seed_11936_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.015445646457374096, + "vlm_logprob": -0.015445646457374096, + "vlm_yes_no_margin": 4.424999952316284, + "vlm_token_credit": -0.015445646457374096, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01741393655538559 + ], + "opposite_logprob_tokens": [ + -4.142414093017578 + ], + "logprob_reward": -0.01741393655538559, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01741393655538559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08854088932275772, + 0.05768951401114464, + 0.049877870827913284, + 0.03536558896303177, + 0.40084391832351685, + 0.052191268652677536, + 0.03547969087958336, + 0.12087831646203995, + 0.06411682814359665, + 0.09501609951257706 + ], + "credit_map": [ + 0.010254837572574615, + 0.009377515874803066, + 0.014285260811448097, + 0.007547096814960241, + 0.010032933205366135, + 0.01226403284817934, + 0.0115796634927392, + 0.013199555687606335, + 0.008588755503296852, + 0.007444686722010374, + 0.0077705164439976215, + 0.004497220274060965, + 0.010063150897622108, + 0.009054956026375294, + 0.004777469206601381, + 0.005492763128131628, + 0.008145992644131184, + 0.005286621395498514, + 0.008595112711191177, + 0.007969487458467484, + 0.0056737931445240974, + 0.005856221076101065, + 0.0037145009264349937, + 0.00463614147156477, + 0.0040060412138700485, + 0.00541615579277277, + 0.0035876904148608446, + 0.004618985578417778, + 0.004118716344237328, + 0.005930389743298292, + 0.0037347530014812946, + 0.003952860366553068, + 0.017270071431994438, + 0.0311447661370039, + 0.014453461393713951, + 0.019589664414525032, + 0.16043402254581451, + 0.030856717377901077, + 0.03712528571486473, + 0.0899699479341507, + 0.005034149158746004, + 0.0071234554052352905, + 0.004099631682038307, + 0.009472296573221684, + 0.0055376687087118626, + 0.0034114597365260124, + 0.006937611848115921, + 0.010574998334050179, + 0.004108229652047157, + 0.003580981632694602, + 0.005464162677526474, + 0.004502434749156237, + 0.0038594750221818686, + 0.006031285505741835, + 0.0036984621547162533, + 0.004234663676470518, + 0.021016767248511314, + 0.012123488821089268, + 0.02241813763976097, + 0.012423367239534855, + 0.012597001157701015, + 0.01796301268041134, + 0.0125197758898139, + 0.009816780686378479, + 0.005924198776483536, + 0.0069247158244252205, + 0.010778283700346947, + 0.007115052081644535, + 0.006887550000101328, + 0.008418734185397625, + 0.00977700762450695, + 0.00829128548502922, + 0.013569019734859467, + 0.012127581052482128, + 0.00887778215110302, + 0.011170399375259876, + 0.01422234158962965, + 0.013895507901906967, + 0.012531640939414501, + 0.008621836081147194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.01741393655538559 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003199579194188118 + ], + "opposite_logprob_tokens": [ + -5.753199577331543 + ], + "logprob_reward": -0.003199579194188118, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003199579194188118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09638518840074539, + 0.07762616872787476, + 0.09313614666461945, + 0.040770191699266434, + 0.1626289337873459, + 0.028553608804941177, + 0.026248134672641754, + 0.0677051767706871, + 0.051417119801044464, + 0.3555293381214142 + ], + "credit_map": [ + 0.013289513066411018, + 0.009861916303634644, + 0.012499812990427017, + 0.010873372666537762, + 0.012664100155234337, + 0.013235289603471756, + 0.011786463670432568, + 0.012174717150628567, + 0.008339060470461845, + 0.007810654118657112, + 0.005390186794102192, + 0.0053904494270682335, + 0.018404116854071617, + 0.01267792098224163, + 0.010551773943006992, + 0.009062002412974834, + 0.008305704221129417, + 0.006120757665485144, + 0.008531223982572556, + 0.007956522516906261, + 0.01584760844707489, + 0.01563621498644352, + 0.018007805570960045, + 0.012730302289128304, + 0.003683994757011533, + 0.004413369577378035, + 0.0035871933214366436, + 0.0036822757683694363, + 0.006213027518242598, + 0.007418885361403227, + 0.005496192257851362, + 0.00627525569871068, + 0.0067519317381083965, + 0.012630806304514408, + 0.0058638472110033035, + 0.007727018557488918, + 0.06377340853214264, + 0.011704323813319206, + 0.015555903315544128, + 0.0386216901242733, + 0.003617124632000923, + 0.00368875521235168, + 0.003840150311589241, + 0.004874378442764282, + 0.002817458240315318, + 0.0021412798669189215, + 0.003873401554301381, + 0.0037010612431913614, + 0.0031000031158328056, + 0.0028532862197607756, + 0.0035953186452388763, + 0.003161010565236211, + 0.003922964446246624, + 0.0040754578076303005, + 0.00324876606464386, + 0.002291328739374876, + 0.012648320756852627, + 0.006614783313125372, + 0.011327750980854034, + 0.003827061504125595, + 0.008476336486637592, + 0.009791135787963867, + 0.007131604477763176, + 0.007888185791671276, + 0.0050583877600729465, + 0.0061689685098826885, + 0.007942057214677334, + 0.0038592570926994085, + 0.0047276997938752174, + 0.006420713849365711, + 0.005797963123768568, + 0.011442073620855808, + 0.05744341015815735, + 0.057440582662820816, + 0.034471746534109116, + 0.038230739533901215, + 0.05222831293940544, + 0.05017618462443352, + 0.03981248289346695, + 0.025725863873958588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.003199579194188118 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638749830424786 + ], + "opposite_logprob_tokens": [ + -4.758638858795166 + ], + "logprob_reward": -0.008638749830424786, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638749830424786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24692532420158386, + 0.12927976250648499, + 0.043970413506031036, + 0.041461385786533356, + 0.31747639179229736, + 0.029975568875670433, + 0.05253913253545761, + 0.04579530656337738, + 0.03337199240922928, + 0.05920473858714104 + ], + "credit_map": [ + 0.05320191755890846, + 0.03948510065674782, + 0.023356979712843895, + 0.019090404734015465, + 0.03662920370697975, + 0.02939395233988762, + 0.02159656397998333, + 0.02417120710015297, + 0.01782592572271824, + 0.028979172930121422, + 0.015967044979333878, + 0.011179712601006031, + 0.01550563145428896, + 0.01741051860153675, + 0.013258922845125198, + 0.009152823127806187, + 0.0044378917664289474, + 0.0034659693483263254, + 0.005125050432980061, + 0.00582702225074172, + 0.005322537384927273, + 0.005848044995218515, + 0.0080724423751235, + 0.005871456582099199, + 0.00488862581551075, + 0.005819023121148348, + 0.003413175931200385, + 0.00505387969315052, + 0.0059866271913051605, + 0.00880509428679943, + 0.0038095449563115835, + 0.0036854129284620285, + 0.01386071927845478, + 0.024079564958810806, + 0.011744996532797813, + 0.013769143261015415, + 0.1263698786497116, + 0.023417171090841293, + 0.029130516573786736, + 0.07510440051555634, + 0.0036912879440933466, + 0.004843093920499086, + 0.0033081872388720512, + 0.004212615545839071, + 0.0030123733449727297, + 0.002058663871139288, + 0.004143328871577978, + 0.0047060162760317326, + 0.006362586747854948, + 0.004965539555996656, + 0.00925238337367773, + 0.005058341193944216, + 0.00498225586488843, + 0.01129485759884119, + 0.004190902225673199, + 0.006432263646274805, + 0.008414626121520996, + 0.00463496008887887, + 0.00815956573933363, + 0.004724693018943071, + 0.005374383181333542, + 0.004765352234244347, + 0.005235177930444479, + 0.004486543592065573, + 0.0035531322937458754, + 0.003014914458617568, + 0.0040201786905527115, + 0.0034584058448672295, + 0.0032675417605787516, + 0.00538939842954278, + 0.004316344857215881, + 0.006352076306939125, + 0.00718860886991024, + 0.00746546545997262, + 0.00517375348135829, + 0.0069172377698123455, + 0.008749944157898426, + 0.010175496339797974, + 0.007649407722055912, + 0.005884828977286816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.008638749830424786 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979346178472042 + ], + "opposite_logprob_tokens": [ + -3.5297935009002686 + ], + "logprob_reward": -0.02979346178472042, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979346178472042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12451645731925964, + 0.06471254676580429, + 0.05463714152574539, + 0.039225950837135315, + 0.435146301984787, + 0.09829097986221313, + 0.02880241721868515, + 0.04789268597960472, + 0.06149359419941902, + 0.04528196528553963 + ], + "credit_map": [ + 0.01582360453903675, + 0.012128937989473343, + 0.01687859371304512, + 0.010242377407848835, + 0.016411192715168, + 0.017595557495951653, + 0.015733376145362854, + 0.019702808931469917, + 0.006338301580399275, + 0.008432701230049133, + 0.007830199785530567, + 0.004370738286525011, + 0.010766828432679176, + 0.009893600828945637, + 0.010012797079980373, + 0.00706737395375967, + 0.009565089829266071, + 0.007892869412899017, + 0.004902044776827097, + 0.005351594183593988, + 0.007639255374670029, + 0.008683371357619762, + 0.0061668530106544495, + 0.004436058457940817, + 0.0035011630970984697, + 0.0030827680602669716, + 0.0039040714036673307, + 0.005107377655804157, + 0.006103478372097015, + 0.007647679187357426, + 0.005392459686845541, + 0.004486951045691967, + 0.02852984517812729, + 0.029279178008437157, + 0.01825016178190708, + 0.028002120554447174, + 0.15337195992469788, + 0.03406975418329239, + 0.04192490875720978, + 0.10171837359666824, + 0.009783529676496983, + 0.009195872582495213, + 0.010372958146035671, + 0.018328944221138954, + 0.012979446910321712, + 0.006383354775607586, + 0.01578196883201599, + 0.015464897267520428, + 0.004462924785912037, + 0.0029418894555419683, + 0.0039374493062496185, + 0.003849881934002042, + 0.0035005847457796335, + 0.003925339784473181, + 0.0032400009222328663, + 0.002944343723356724, + 0.006621782202273607, + 0.006536036729812622, + 0.0074877748265862465, + 0.0037720866966992617, + 0.007065911311656237, + 0.005203434731811285, + 0.004725223407149315, + 0.00648043118417263, + 0.005953551735728979, + 0.006430727895349264, + 0.010352769866585732, + 0.005095562431961298, + 0.005326071288436651, + 0.00801140908151865, + 0.007138447370380163, + 0.013185049407184124, + 0.005312497261911631, + 0.006309053860604763, + 0.003556830808520317, + 0.008002481423318386, + 0.006820174865424633, + 0.007567634340375662, + 0.003438039217144251, + 0.0042752488516271114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.02979346178472042 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive biomechanical tree with gunmetal-gray metallic bark and interwoven fiber cables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the tree's roots glow a vibrant emerald green and emit streams of light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dense, neon-lit cityscape with holographic advertisements and flying vehicles under perpetual twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a low-angle, wide-angle shot and perform a slow, deliberate pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a section of the trunk bulge, crack, and expel a thick cloud of golden, shimmering energy particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182504922151566 + ], + "opposite_logprob_tokens": [ + -4.018182277679443 + ], + "logprob_reward": -0.018182504922151566, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018182504922151566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04916563257575035, + 0.0327242910861969, + 0.03951188549399376, + 0.03533964976668358, + 0.12122324854135513, + 0.021913114935159683, + 0.02808382920920849, + 0.04559415578842163, + 0.030421392992138863, + 0.5960227251052856 + ], + "credit_map": [ + 0.006677618250250816, + 0.005990282632410526, + 0.008181310258805752, + 0.004495851695537567, + 0.006554222200065851, + 0.005971376318484545, + 0.005752534605562687, + 0.005542435683310032, + 0.003947146702557802, + 0.00419068057090044, + 0.0036035270895808935, + 0.002647567307576537, + 0.006283989641815424, + 0.005084095988422632, + 0.004694990813732147, + 0.0022722959984093904, + 0.005469858646392822, + 0.004019911866635084, + 0.006396750919520855, + 0.0059134396724402905, + 0.004886676091700792, + 0.004689992871135473, + 0.0052043115720152855, + 0.0029309457167983055, + 0.003639178117737174, + 0.004212541505694389, + 0.004296685568988323, + 0.004291090648621321, + 0.004369054920971394, + 0.005868326872587204, + 0.005549838300794363, + 0.0031129359267652035, + 0.008426089771091938, + 0.010028699412941933, + 0.00596247473731637, + 0.007255539298057556, + 0.043933380395174026, + 0.011848592199385166, + 0.011707091704010963, + 0.022061379626393318, + 0.003035892965272069, + 0.0029963506385684013, + 0.0036488075274974108, + 0.0024699883069843054, + 0.0020184882450848818, + 0.001896487083286047, + 0.002772318897768855, + 0.003074781736359, + 0.0037262060213834047, + 0.002913425210863352, + 0.0038106809370219707, + 0.002970379777252674, + 0.003097879234701395, + 0.004125386942178011, + 0.004128472879528999, + 0.003311397973448038, + 0.008105243556201458, + 0.004366810899227858, + 0.008416203781962395, + 0.004095915704965591, + 0.006025902461260557, + 0.005612339824438095, + 0.005263521336019039, + 0.0037082224152982235, + 0.003131585894152522, + 0.0037526744417846203, + 0.005137142725288868, + 0.004326197784394026, + 0.0032433457672595978, + 0.003056642133742571, + 0.004413981456309557, + 0.0033598218578845263, + 0.09149468690156937, + 0.09274261444807053, + 0.055589061230421066, + 0.07302224636077881, + 0.08833533525466919, + 0.08932147175073624, + 0.0652211382985115, + 0.040296223014593124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.018182504922151566 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_9/index_tensor([2530, 2530])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_9/index_tensor([2530, 2530])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8c90a3ccfd114aac725b2f307660c4f510fb97bd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_9/index_tensor([2530, 2530])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium shot, at eye level, captures the exuberant energy of Wadadou, a cheerful, anthropomorphic taiko drummer rendered in a vibrant pop art style defined by bold, flat colors and thick, decisive black outlines. His body is a perfect bright red cylinder, his face nestled against the surface of the drum, his large, expressive eyes closed in musical ecstasy, his mouth wide open as he belts out the song. He passionately grasps a classic silver microphone in one hand, gripping it tightly, his entire body seemingly vibrating with the power of his voice. The scene is set in a karaoke room, its backdrop a flat, electric yellow wall covered in a repeating pattern of large cyan dots. Multi-colored strobe lights rhythmically flicker and flash, casting rapidly changing patterns of fuchsia and green onto the character and the room, creating sharp, black-outlined shadows that dance and pulsate to the beat, amplifying the vibrant, joyful atmosphere of the throat performance.", + "seed": 30766, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_9/index_2530_seed_30766_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01723182140849531, + "vlm_logprob": -0.01723182140849531, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.01723182140849531, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010344106703996658 + ], + "opposite_logprob_tokens": [ + -4.6353440284729 + ], + "logprob_reward": -0.010344106703996658, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010344106703996658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1180284321308136, + 0.06419630348682404, + 0.13552692532539368, + 0.2995999753475189, + 0.06311020255088806, + 0.09046401083469391, + 0.047495003789663315, + 0.03672407194972038, + 0.048950858414173126, + 0.09590420871973038 + ], + "credit_map": [ + 0.008470026776194572, + 0.012741245329380035, + 0.008688108995556831, + 0.030808527022600174, + 0.019021889194846153, + 0.008369127288460732, + 0.018971579149365425, + 0.010957927443087101, + 0.012029494158923626, + 0.009207149036228657, + 0.006578547414392233, + 0.006894917692989111, + 0.007074282038956881, + 0.006815674249082804, + 0.007888276129961014, + 0.007707964163273573, + 0.012940453365445137, + 0.015141126699745655, + 0.010041298344731331, + 0.026320097967982292, + 0.018833743408322334, + 0.018692437559366226, + 0.019446931779384613, + 0.014110840857028961, + 0.01666191592812538, + 0.018514279276132584, + 0.014464356005191803, + 0.09947187453508377, + 0.021151866763830185, + 0.06333910673856735, + 0.02931937575340271, + 0.03667718917131424, + 0.0036754882894456387, + 0.00649390509352088, + 0.0030582298059016466, + 0.019705653190612793, + 0.007171499077230692, + 0.003066654782742262, + 0.00762880640104413, + 0.012309962883591652, + 0.006804735865443945, + 0.008896056562662125, + 0.004824954550713301, + 0.010315115563571453, + 0.010549464263021946, + 0.009210895746946335, + 0.02777363732457161, + 0.012089146301150322, + 0.005812502466142178, + 0.006842514034360647, + 0.005370826926082373, + 0.009122878313064575, + 0.005662035662680864, + 0.004904040601104498, + 0.005989454220980406, + 0.0037907520309090614, + 0.0026845973916351795, + 0.0024593593552708626, + 0.0035578387323766947, + 0.0048518795520067215, + 0.003785501467064023, + 0.004683695267885923, + 0.00875724945217371, + 0.005943950731307268, + 0.005443645641207695, + 0.0034361404832452536, + 0.005725749768316746, + 0.009891994297504425, + 0.004422983154654503, + 0.004022229928523302, + 0.005995156709104776, + 0.010012954473495483, + 0.005774934310466051, + 0.01067822054028511, + 0.00549627048894763, + 0.012687650509178638, + 0.02189445309340954, + 0.008284744806587696, + 0.01980440504848957, + 0.011283528059720993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.010344106703996658 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430544950067997 + ], + "opposite_logprob_tokens": [ + -3.1680545806884766 + ], + "logprob_reward": -0.0430544950067997, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430544950067997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1580478549003601, + 0.051880914717912674, + 0.07240843027830124, + 0.09655380994081497, + 0.05790944769978523, + 0.24340076744556427, + 0.04111317917704582, + 0.1214081272482872, + 0.06697633117437363, + 0.09030113369226456 + ], + "credit_map": [ + 0.011231396347284317, + 0.015457365661859512, + 0.011515182442963123, + 0.045842334628105164, + 0.02141980454325676, + 0.014398761093616486, + 0.027310341596603394, + 0.010872664861381054, + 0.007210710551589727, + 0.007184543646872044, + 0.004511836916208267, + 0.007630458567291498, + 0.006931908428668976, + 0.004437629599124193, + 0.006482081022113562, + 0.007491746451705694, + 0.006582814734429121, + 0.010964123532176018, + 0.007486618123948574, + 0.007641995791345835, + 0.011776702478528023, + 0.009334303438663483, + 0.010706596076488495, + 0.007915281690657139, + 0.007046673446893692, + 0.005734971258789301, + 0.007278404198586941, + 0.03046918287873268, + 0.008421639911830425, + 0.015296494588255882, + 0.011248710565268993, + 0.011057736352086067, + 0.003954231273382902, + 0.004857655614614487, + 0.0035541136749088764, + 0.009124581702053547, + 0.007729202974587679, + 0.0062211304903030396, + 0.007906774058938026, + 0.014561758376657963, + 0.016473768278956413, + 0.025725340470671654, + 0.012293530628085136, + 0.028218884021043777, + 0.03342346101999283, + 0.023053579032421112, + 0.07292138785123825, + 0.031290825456380844, + 0.0035086870193481445, + 0.006015198305249214, + 0.004350463859736919, + 0.006401594262570143, + 0.006541913840919733, + 0.0023996089585125446, + 0.005706327967345715, + 0.0061893826350569725, + 0.011421054601669312, + 0.010005760006606579, + 0.011285207234323025, + 0.016133742406964302, + 0.019450556486845016, + 0.012588911689817905, + 0.02204427309334278, + 0.018478617072105408, + 0.007726349867880344, + 0.005746059585362673, + 0.009065021760761738, + 0.01397391315549612, + 0.00815313309431076, + 0.0067602284252643585, + 0.0064943647012114525, + 0.009057261049747467, + 0.007805687841027975, + 0.007701206021010876, + 0.006786986719816923, + 0.016761599108576775, + 0.015036964789032936, + 0.006398496683686972, + 0.020844390615820885, + 0.008965798653662205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.0430544950067997 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026331797242164612 + ], + "opposite_logprob_tokens": [ + -3.651331901550293 + ], + "logprob_reward": -0.026331797242164612, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026331797242164612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11558370292186737, + 0.12669244408607483, + 0.07335299253463745, + 0.3335299491882324, + 0.052269864827394485, + 0.03635868430137634, + 0.03389924019575119, + 0.0877133458852768, + 0.03562526777386665, + 0.10497451573610306 + ], + "credit_map": [ + 0.018037524074316025, + 0.01087090466171503, + 0.017641007900238037, + 0.026036562398076057, + 0.009379029273986816, + 0.012281288392841816, + 0.010442408733069897, + 0.010894978418946266, + 0.02649487368762493, + 0.0116709740832448, + 0.02030370384454727, + 0.012451082468032837, + 0.009722603484988213, + 0.01670026034116745, + 0.012996324338018894, + 0.01635262928903103, + 0.0062528569251298904, + 0.013100354000926018, + 0.005860358476638794, + 0.012081385590136051, + 0.012534656561911106, + 0.006429668515920639, + 0.010462182573974133, + 0.006631528493016958, + 0.023730358108878136, + 0.018860766664147377, + 0.020278751850128174, + 0.13003799319267273, + 0.017799338325858116, + 0.07196445018053055, + 0.02457312121987343, + 0.026285188272595406, + 0.0065245842561125755, + 0.0055220406502485275, + 0.007161046378314495, + 0.00983879528939724, + 0.006451943889260292, + 0.005483289714902639, + 0.0059252651408314705, + 0.005362898577004671, + 0.00420139217749238, + 0.004128951113671064, + 0.0027367176953703165, + 0.004244827665388584, + 0.005696807522326708, + 0.0025738184340298176, + 0.008498563431203365, + 0.004277603700757027, + 0.004756537266075611, + 0.004493778105825186, + 0.003831692272797227, + 0.004252221435308456, + 0.0038975465577095747, + 0.004086563363671303, + 0.0035831236746162176, + 0.004997775889933109, + 0.007495297584682703, + 0.006110079120844603, + 0.006742341909557581, + 0.022820116952061653, + 0.007043026387691498, + 0.012093707919120789, + 0.012287165969610214, + 0.01312161237001419, + 0.0031595290638506413, + 0.0033140110317617655, + 0.0037185244727879763, + 0.007094970438629389, + 0.005136904306709766, + 0.0037185577675700188, + 0.005547251086682081, + 0.003935519140213728, + 0.01134142279624939, + 0.0075715514831244946, + 0.010495335794985294, + 0.011842098087072372, + 0.00963553600013256, + 0.011949962005019188, + 0.012156592682003975, + 0.02998201921582222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.026331797242164612 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028177103959023952 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028177103959023952, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028177103959023952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07596008479595184, + 0.05926063656806946, + 0.09374766796827316, + 0.21545197069644928, + 0.04927133768796921, + 0.114804707467556, + 0.11558178812265396, + 0.1542665809392929, + 0.08727837353944778, + 0.03437691554427147 + ], + "credit_map": [ + 0.010467085056006908, + 0.009052393026649952, + 0.007969412952661514, + 0.007412618491798639, + 0.008020964451134205, + 0.0065516154281795025, + 0.008673571981489658, + 0.017812427133321762, + 0.007671097759157419, + 0.008473237045109272, + 0.009898394346237183, + 0.006730851251631975, + 0.006740537937730551, + 0.005657331086695194, + 0.009343275800347328, + 0.004745906218886375, + 0.007719018030911684, + 0.012702425010502338, + 0.008341751992702484, + 0.015237276442348957, + 0.015686731785535812, + 0.009118727408349514, + 0.014343067072331905, + 0.010598665103316307, + 0.014141179621219635, + 0.011717552319169044, + 0.011548303999006748, + 0.08250097185373306, + 0.012383442372083664, + 0.049204692244529724, + 0.0157338734716177, + 0.018221935257315636, + 0.005969769321382046, + 0.004716763738542795, + 0.004549316130578518, + 0.006008628290146589, + 0.008621042594313622, + 0.004928172565996647, + 0.006562746595591307, + 0.007914896123111248, + 0.011933733709156513, + 0.01042160578072071, + 0.008561106398701668, + 0.010926165618002415, + 0.012052047066390514, + 0.020879874005913734, + 0.01545965950936079, + 0.024570511654019356, + 0.014629930257797241, + 0.017166711390018463, + 0.012382719665765762, + 0.02261176146566868, + 0.01323364581912756, + 0.01100105233490467, + 0.014200950041413307, + 0.01035500317811966, + 0.015665100887417793, + 0.009956480003893375, + 0.016639085486531258, + 0.043241024017333984, + 0.014795761555433273, + 0.015438530594110489, + 0.021461639553308487, + 0.01706894300878048, + 0.00839656125754118, + 0.010148514062166214, + 0.0070526376366615295, + 0.00686901668086648, + 0.020469333976507187, + 0.005579065065830946, + 0.018315505236387253, + 0.010447738692164421, + 0.0038342333864420652, + 0.0035407484974712133, + 0.0030033958610147238, + 0.004221497569233179, + 0.006842527072876692, + 0.00211612181738019, + 0.006180366035550833, + 0.0046380264684557915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.0028177103959023952 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036109976936131716 + ], + "opposite_logprob_tokens": [ + -5.628611087799072 + ], + "logprob_reward": -0.0036109976936131716, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036109976936131716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13776381313800812, + 0.04418869689106941, + 0.1225445494055748, + 0.09811979532241821, + 0.0912557914853096, + 0.050812866538763046, + 0.07383593171834946, + 0.20236988365650177, + 0.10959847271442413, + 0.06951019167900085 + ], + "credit_map": [ + 0.01440377440303564, + 0.016601255163550377, + 0.01703696884214878, + 0.02902407944202423, + 0.0180521123111248, + 0.013088568113744259, + 0.018255997449159622, + 0.011301050893962383, + 0.007225970737636089, + 0.0055944062769412994, + 0.006966093089431524, + 0.004548134282231331, + 0.003961833193898201, + 0.005749364383518696, + 0.005637885071337223, + 0.004505007527768612, + 0.013605146668851376, + 0.016574252396821976, + 0.02086254395544529, + 0.01839575543999672, + 0.012813798151910305, + 0.013787595555186272, + 0.015747105702757835, + 0.010758353397250175, + 0.00728195533156395, + 0.005635356064885855, + 0.006846231874078512, + 0.03734300285577774, + 0.005836342461407185, + 0.019842669367790222, + 0.007641337811946869, + 0.007692898623645306, + 0.007504013366997242, + 0.008199408650398254, + 0.0051105450838804245, + 0.030482694506645203, + 0.00891510397195816, + 0.005196308251470327, + 0.009668542072176933, + 0.016179176047444344, + 0.005285091698169708, + 0.006824237760156393, + 0.004269592463970184, + 0.006378351245075464, + 0.007545009255409241, + 0.0038968438748270273, + 0.01129313837736845, + 0.00532060069963336, + 0.007291185203939676, + 0.010352474637329578, + 0.006401219870895147, + 0.015458866022527218, + 0.008908659219741821, + 0.007536589168012142, + 0.011372605338692665, + 0.006514331791549921, + 0.02282252162694931, + 0.014209046959877014, + 0.024249743670225143, + 0.05394452437758446, + 0.017253227531909943, + 0.021345075219869614, + 0.025650564581155777, + 0.02289518341422081, + 0.0114232636988163, + 0.013221685774624348, + 0.02216508612036705, + 0.01426423154771328, + 0.012302251532673836, + 0.012361641973257065, + 0.012356282211840153, + 0.011504033580422401, + 0.0042289490811526775, + 0.00702096801251173, + 0.003465749090537429, + 0.008870978839695454, + 0.013225059024989605, + 0.007868489250540733, + 0.014020145870745182, + 0.010809848085045815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.0036109976936131716 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, at eye level, captures the exuberant energy of Wadadou, a cheerful, anthropomorphic taiko drummer rendered in a vibrant pop art style defined by bold, flat colors and thick, decisive black outlines. His body is a perfect bright red cylinder, his face nestled against the surface of the drum, his large, expressive eyes closed in musical ecstasy, his mouth wide open as he belts out the song. He passionately grasps a classic silver microphone in one hand, gripping it tightly, his entire body seemingly vibrating with the power of his voice. The scene is set in a karaoke room, its backdrop a flat, electric yellow wall covered in a repeating pattern of large cyan dots. Multi-colored strobe lights rhythmically flicker and flash, casting rapidly changing patterns of fuchsia and green onto the character and the room, creating sharp, black-outlined shadows that dance and pulsate to the beat, amplifying the vibrant, joyful atmosphere of the throat performance.", + "seed": 30767, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_9/index_2530_seed_30767_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.012136291153728962, + "vlm_logprob": -0.012136291153728962, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.012136291153728962, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004247334785759449 + ], + "opposite_logprob_tokens": [ + -5.629247188568115 + ], + "logprob_reward": -0.004247334785759449, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.004247334785759449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13750888407230377, + 0.08372492343187332, + 0.04041096940636635, + 0.05420069396495819, + 0.16218572854995728, + 0.05030330643057823, + 0.0635141134262085, + 0.2963445484638214, + 0.04596737399697304, + 0.06583940982818604 + ], + "credit_map": [ + 0.02394396811723709, + 0.012115526013076305, + 0.012070627883076668, + 0.03813547641038895, + 0.018169578164815903, + 0.012104335241019726, + 0.01091682817786932, + 0.010052545927464962, + 0.00842178426682949, + 0.013390650041401386, + 0.006803558673709631, + 0.012441383674740791, + 0.013571666553616524, + 0.009850704111158848, + 0.00961992796510458, + 0.009625251404941082, + 0.0063248006626963615, + 0.004661465995013714, + 0.005275399889796972, + 0.008156011812388897, + 0.004300260450690985, + 0.004843533504754305, + 0.0032186449971050024, + 0.0036308534909039736, + 0.008970976807177067, + 0.003954197745770216, + 0.005577590316534042, + 0.006832615472376347, + 0.006469986867159605, + 0.013536996208131313, + 0.0038056555204093456, + 0.005052677821367979, + 0.020488565787672997, + 0.016970181837677956, + 0.019372299313545227, + 0.053610909730196, + 0.010552938096225262, + 0.011135433800518513, + 0.01922002248466015, + 0.010835381224751472, + 0.005677611567080021, + 0.0047463965602219105, + 0.006011694669723511, + 0.009722936898469925, + 0.009614324197173119, + 0.0035252587404102087, + 0.00655702967196703, + 0.004448052495718002, + 0.004782564472407103, + 0.007742309011518955, + 0.010897149331867695, + 0.007156113162636757, + 0.008933371864259243, + 0.007076810114085674, + 0.008306893520057201, + 0.008618906140327454, + 0.0424078069627285, + 0.02413335070014, + 0.05072088912129402, + 0.05327809602022171, + 0.030149761587381363, + 0.0413154773414135, + 0.026609452441334724, + 0.027729712426662445, + 0.004393185488879681, + 0.0048264809884130955, + 0.004493071231991053, + 0.005490792449563742, + 0.009672602638602257, + 0.006537620443850756, + 0.005699970759451389, + 0.004853646270930767, + 0.006645113695412874, + 0.007682747673243284, + 0.010937601327896118, + 0.011026468127965927, + 0.007349623832851648, + 0.011014682240784168, + 0.006202723365277052, + 0.004980449564754963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.004247334785759449 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304821416735649 + ], + "opposite_logprob_tokens": [ + -3.168048143386841 + ], + "logprob_reward": -0.04304821416735649, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304821416735649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.125944122672081, + 0.05539831519126892, + 0.028127074241638184, + 0.02833898365497589, + 0.04133957624435425, + 0.03820588439702988, + 0.06328599154949188, + 0.5595643520355225, + 0.02032785303890705, + 0.03946780413389206 + ], + "credit_map": [ + 0.017306726425886154, + 0.013104499317705631, + 0.013400311581790447, + 0.0322866328060627, + 0.014831136912107468, + 0.012211873196065426, + 0.014027468860149384, + 0.00877547636628151, + 0.004675771575421095, + 0.008171988651156425, + 0.004728102590888739, + 0.008401045575737953, + 0.010674605146050453, + 0.007702211383730173, + 0.005473254714161158, + 0.005571335554122925, + 0.00442757410928607, + 0.003117763204500079, + 0.003089252160862088, + 0.0076555851846933365, + 0.0030771992169320583, + 0.002036848571151495, + 0.002465946599841118, + 0.0022569065913558006, + 0.005476333666592836, + 0.0018418903928250074, + 0.002219842281192541, + 0.004279200453311205, + 0.00486059719696641, + 0.004404111765325069, + 0.002651493763551116, + 0.0026055132038891315, + 0.00458404328674078, + 0.004451765678822994, + 0.006739270407706499, + 0.009256125427782536, + 0.004137563984841108, + 0.004128504544496536, + 0.004519876558333635, + 0.003522424725815654, + 0.003854164620861411, + 0.003799799829721451, + 0.0025557454209774733, + 0.0051318141631782055, + 0.008067951537668705, + 0.005778769962489605, + 0.0045243739150464535, + 0.004493265878409147, + 0.007756982464343309, + 0.006045490503311157, + 0.004253272898495197, + 0.01690622605383396, + 0.009193871170282364, + 0.004497727379202843, + 0.00882683601230383, + 0.005805585067719221, + 0.07740464061498642, + 0.045175861567258835, + 0.09720071405172348, + 0.09051842242479324, + 0.0684017539024353, + 0.07820405811071396, + 0.04825956001877785, + 0.05439935252070427, + 0.0023768714163452387, + 0.001555117778480053, + 0.0032402423676103354, + 0.005211833398789167, + 0.001965706003829837, + 0.0022548881825059652, + 0.0021610523108392954, + 0.0015621405327692628, + 0.0033983078319579363, + 0.004607148934155703, + 0.0057572657242417336, + 0.005109206307679415, + 0.0074491132982075214, + 0.0042387377470731735, + 0.005136711057275534, + 0.0037713139317929745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.04304821416735649 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005940400995314121 + ], + "opposite_logprob_tokens": [ + -5.1309404373168945 + ], + "logprob_reward": -0.005940400995314121, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005940400995314121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056303706020116806, + 0.12152501940727234, + 0.2049969881772995, + 0.11041636765003204, + 0.08290509879589081, + 0.05490799993276596, + 0.1208340972661972, + 0.1646905541419983, + 0.028672700747847557, + 0.054747432470321655 + ], + "credit_map": [ + 0.008508829399943352, + 0.005976254586130381, + 0.005731339566409588, + 0.013682384043931961, + 0.005044916179031134, + 0.006951621267944574, + 0.004410458263009787, + 0.005997906904667616, + 0.01267280988395214, + 0.02087116800248623, + 0.009311684407293797, + 0.01678047701716423, + 0.017234500497579575, + 0.011708451434969902, + 0.012135160155594349, + 0.020810775458812714, + 0.03364171087741852, + 0.019759638234972954, + 0.013818345032632351, + 0.06260768324136734, + 0.025581613183021545, + 0.015706397593021393, + 0.018378349021077156, + 0.01550326868891716, + 0.020448321476578712, + 0.009675920009613037, + 0.02573622576892376, + 0.020234134048223495, + 0.009180614724755287, + 0.009895989671349525, + 0.008068771101534367, + 0.007176402490586042, + 0.014577927999198437, + 0.009037191048264503, + 0.006754452828317881, + 0.02598756179213524, + 0.005189287941902876, + 0.006777416914701462, + 0.009418266825377941, + 0.005162990186363459, + 0.008919275365769863, + 0.0060163820162415504, + 0.0051560066640377045, + 0.013808580115437508, + 0.004310211166739464, + 0.005676095839589834, + 0.0058623929508030415, + 0.005159060005098581, + 0.011724943295121193, + 0.009923438541591167, + 0.01068964321166277, + 0.027811674401164055, + 0.01144763920456171, + 0.022463731467723846, + 0.017104892060160637, + 0.009668138809502125, + 0.02175959199666977, + 0.015350794419646263, + 0.0144082922488451, + 0.03592897951602936, + 0.01817052997648716, + 0.016907386481761932, + 0.022278418764472008, + 0.019886564463377, + 0.0029500150121748447, + 0.002791469683870673, + 0.0043711489997804165, + 0.007023553363978863, + 0.004000578075647354, + 0.002452509943395853, + 0.003457652172073722, + 0.001625774777494371, + 0.004849845543503761, + 0.008767001330852509, + 0.008918601088225842, + 0.006478590425103903, + 0.006874338258057833, + 0.0073752254247665405, + 0.006373564712703228, + 0.005110267549753189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.005940400995314121 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730505730956793 + ], + "opposite_logprob_tokens": [ + -5.006730556488037 + ], + "logprob_reward": -0.006730505730956793, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730505730956793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04219956323504448, + 0.044355958700180054, + 0.046134430915117264, + 0.07095684111118317, + 0.05408858135342598, + 0.016787471249699593, + 0.061363134533166885, + 0.40818285942077637, + 0.024956654757261276, + 0.23097455501556396 + ], + "credit_map": [ + 0.006994784809648991, + 0.004720805678516626, + 0.007736990228295326, + 0.006558088585734367, + 0.00455193966627121, + 0.005143572110682726, + 0.003455984639003873, + 0.003037400310859084, + 0.007623324636369944, + 0.0052987998351454735, + 0.004952054936438799, + 0.007573434617370367, + 0.0053440905176103115, + 0.004335142206400633, + 0.0052300915122032166, + 0.003999021369963884, + 0.006858767010271549, + 0.0050791604444384575, + 0.003481837222352624, + 0.014899629168212414, + 0.004653356969356537, + 0.0029776571318507195, + 0.004248781129717827, + 0.003935239277780056, + 0.014690163545310497, + 0.004804455209523439, + 0.005049028433859348, + 0.011080626398324966, + 0.006401561666280031, + 0.018002329394221306, + 0.004744610283523798, + 0.006184072233736515, + 0.006489143241196871, + 0.005051611922681332, + 0.01044544018805027, + 0.01763034798204899, + 0.002762887394055724, + 0.0030413575004786253, + 0.005537098273634911, + 0.0031306908931583166, + 0.0023706539068371058, + 0.002239273628219962, + 0.0023505985736846924, + 0.003115399507805705, + 0.001788610010407865, + 0.0015549923991784453, + 0.0016992561286315322, + 0.001668686280027032, + 0.007982385344803333, + 0.004824426956474781, + 0.00436342554166913, + 0.017543651163578033, + 0.006762548815459013, + 0.005746959708631039, + 0.009206206537783146, + 0.00493353046476841, + 0.05599117651581764, + 0.034134700894355774, + 0.08092517405748367, + 0.059059929102659225, + 0.04655245319008827, + 0.05914702266454697, + 0.033807914704084396, + 0.03856446221470833, + 0.0022273326758295298, + 0.003913336433470249, + 0.003262708894908428, + 0.0038251932710409164, + 0.004298593383282423, + 0.002567577874287963, + 0.002727454761043191, + 0.0021344569977372885, + 0.026878219097852707, + 0.03283345699310303, + 0.02907954715192318, + 0.043449871242046356, + 0.027949267998337746, + 0.028788575902581215, + 0.02203710749745369, + 0.019958525896072388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.006730505730956793 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007150000892579556 + ], + "opposite_logprob_tokens": [ + -7.2507147789001465 + ], + "logprob_reward": -0.0007150000892579556, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007150000892579556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037712302058935165, + 0.05011036619544029, + 0.10762622207403183, + 0.09566047787666321, + 0.04541386291384697, + 0.052234478294849396, + 0.07939742505550385, + 0.35589954257011414, + 0.036809783428907394, + 0.139135479927063 + ], + "credit_map": [ + 0.00470638507977128, + 0.004794263280928135, + 0.004242501221597195, + 0.004616874735802412, + 0.005517540033906698, + 0.005690006539225578, + 0.003770119044929743, + 0.004374610260128975, + 0.005440953187644482, + 0.006507030688226223, + 0.005423508118838072, + 0.006418925244361162, + 0.00674434844404459, + 0.006881339941173792, + 0.0060330587439239025, + 0.006661204621195793, + 0.01680447719991207, + 0.010512372478842735, + 0.007439802400767803, + 0.03478575125336647, + 0.012159322388470173, + 0.007110865321010351, + 0.01017390750348568, + 0.008639724925160408, + 0.016854973509907722, + 0.0071433838456869125, + 0.008662493899464607, + 0.015009259805083275, + 0.008105297572910786, + 0.025145400315523148, + 0.006452235858887434, + 0.008287428878247738, + 0.009567473083734512, + 0.005148046184331179, + 0.004376639146357775, + 0.009193981066346169, + 0.005471779033541679, + 0.0031523024663329124, + 0.005611898843199015, + 0.0028917461168020964, + 0.008608289994299412, + 0.005332729779183865, + 0.004940669983625412, + 0.01371291745454073, + 0.003960471600294113, + 0.005373015999794006, + 0.005787119269371033, + 0.004519265610724688, + 0.007689212914556265, + 0.008597545325756073, + 0.010765849612653255, + 0.013407440856099129, + 0.011480634100735188, + 0.008573313243687153, + 0.008650618605315685, + 0.01023280993103981, + 0.046436555683612823, + 0.035748668015003204, + 0.04257664829492569, + 0.06291002035140991, + 0.04074176773428917, + 0.04303620383143425, + 0.04631645604968071, + 0.03813318908214569, + 0.0039611691609025, + 0.0030808241572231054, + 0.007972333580255508, + 0.005907110869884491, + 0.005865038838237524, + 0.0031081910710781813, + 0.0038513028994202614, + 0.003063814714550972, + 0.014711875468492508, + 0.016250230371952057, + 0.029996847733855247, + 0.017457764595746994, + 0.013685801066458225, + 0.02465583197772503, + 0.009339251555502415, + 0.01303787436336279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.0007150000892579556 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, at eye level, captures the exuberant energy of Wadadou, a cheerful, anthropomorphic taiko drummer rendered in a vibrant pop art style defined by bold, flat colors and thick, decisive black outlines. His body is a perfect bright red cylinder, his face nestled against the surface of the drum, his large, expressive eyes closed in musical ecstasy, his mouth wide open as he belts out the song. He passionately grasps a classic silver microphone in one hand, gripping it tightly, his entire body seemingly vibrating with the power of his voice. The scene is set in a karaoke room, its backdrop a flat, electric yellow wall covered in a repeating pattern of large cyan dots. Multi-colored strobe lights rhythmically flicker and flash, casting rapidly changing patterns of fuchsia and green onto the character and the room, creating sharp, black-outlined shadows that dance and pulsate to the beat, amplifying the vibrant, joyful atmosphere of the throat performance.", + "seed": 30768, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_9/index_2530_seed_30768_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.014267453411594033, + "vlm_logprob": -0.014267453411594033, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.014267453411594033, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004909247159957886 + ], + "opposite_logprob_tokens": [ + -5.504909038543701 + ], + "logprob_reward": -0.004909247159957886, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004909247159957886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.312773197889328, + 0.18052130937576294, + 0.05705876275897026, + 0.15460671484470367, + 0.06514766812324524, + 0.06733284145593643, + 0.05180602893233299, + 0.032762449234724045, + 0.04317241534590721, + 0.034818653017282486 + ], + "credit_map": [ + 0.033497147262096405, + 0.033513881266117096, + 0.029664339497685432, + 0.06096041202545166, + 0.0503976009786129, + 0.024459073320031166, + 0.056521862745285034, + 0.023758837953209877, + 0.00999416783452034, + 0.01639416255056858, + 0.011147726327180862, + 0.06640908122062683, + 0.010631914250552654, + 0.007783063221722841, + 0.04753302037715912, + 0.010628165677189827, + 0.0049297683872282505, + 0.006737454794347286, + 0.006237096153199673, + 0.007967939600348473, + 0.007448168005794287, + 0.005114144645631313, + 0.009743898175656796, + 0.008880287408828735, + 0.013225422240793705, + 0.015544338151812553, + 0.012079281732439995, + 0.03334444388747215, + 0.01565009541809559, + 0.020853614434599876, + 0.02366582490503788, + 0.020243685692548752, + 0.005530898459255695, + 0.006785162724554539, + 0.005817829631268978, + 0.012954166159033775, + 0.007591069210320711, + 0.005671977531164885, + 0.013621509075164795, + 0.007175047416239977, + 0.0048685637302696705, + 0.00931993406265974, + 0.005871894769370556, + 0.011482851579785347, + 0.01123540848493576, + 0.004288989119231701, + 0.013677730225026608, + 0.006587456911802292, + 0.004227651283144951, + 0.009448852390050888, + 0.004991106688976288, + 0.007249579299241304, + 0.010909494012594223, + 0.0029478464275598526, + 0.007065713405609131, + 0.004965780768543482, + 0.0034711542539298534, + 0.005270039662718773, + 0.002869014162570238, + 0.0046480135060846806, + 0.0048512062057852745, + 0.0031979421619325876, + 0.00529104471206665, + 0.003164031310006976, + 0.004804928321391344, + 0.005050960928201675, + 0.0060659185983240604, + 0.006588810123503208, + 0.003829753724858165, + 0.0036950495559722185, + 0.008422978222370148, + 0.0047140116803348064, + 0.003740997053682804, + 0.005225851200520992, + 0.0035441722720861435, + 0.005333058536052704, + 0.005363659467548132, + 0.0023165536113083363, + 0.005270091351121664, + 0.0040242657996714115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.004909247159957886 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029790453612804413 + ], + "opposite_logprob_tokens": [ + -3.5297904014587402 + ], + "logprob_reward": -0.029790453612804413, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029790453612804413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1337808519601822, + 0.5544744729995728, + 0.03282175958156586, + 0.09015671163797379, + 0.04960978031158447, + 0.044392120093107224, + 0.02494559995830059, + 0.013307895511388779, + 0.028452524915337563, + 0.02805832028388977 + ], + "credit_map": [ + 0.013077761977910995, + 0.018385035917162895, + 0.011472981423139572, + 0.02317255176603794, + 0.02231030911207199, + 0.012927829287946224, + 0.02250986546278, + 0.009924523532390594, + 0.036669980734586716, + 0.06420250982046127, + 0.050818879157304764, + 0.16384239494800568, + 0.041682805866003036, + 0.048359062522649765, + 0.10652397572994232, + 0.04237482324242592, + 0.0024277193006128073, + 0.00494059594348073, + 0.003222829895094037, + 0.006204984150826931, + 0.0045735216699540615, + 0.0022866607178002596, + 0.006167319603264332, + 0.0029981264378875494, + 0.007838357239961624, + 0.009910114109516144, + 0.006939834449440241, + 0.018139366060495377, + 0.010189862921833992, + 0.011699720285832882, + 0.013879443518817425, + 0.011560008861124516, + 0.004589729011058807, + 0.005068602040410042, + 0.0037475719582289457, + 0.01074187271296978, + 0.004843646660447121, + 0.003777099074795842, + 0.011234213598072529, + 0.005607047583907843, + 0.0031386713963001966, + 0.006441899575293064, + 0.003239561803638935, + 0.01094916369765997, + 0.005307215731590986, + 0.0025266744196414948, + 0.009269163012504578, + 0.0035197704564779997, + 0.0013653516070917249, + 0.002635472221300006, + 0.0016372079262509942, + 0.005627624224871397, + 0.0035113715566694736, + 0.0015746736899018288, + 0.006757228169590235, + 0.0018366691656410694, + 0.001765322289429605, + 0.0016277842223644257, + 0.001350358477793634, + 0.0026425833348184824, + 0.001389827812090516, + 0.0015879186103120446, + 0.0019333750242367387, + 0.0010107258567586541, + 0.003345262259244919, + 0.004059311468154192, + 0.00284323887899518, + 0.0058365873992443085, + 0.003407492069527507, + 0.0021895926911383867, + 0.004172381944954395, + 0.0025986582040786743, + 0.0019708427134901285, + 0.002812817692756653, + 0.0017254031263291836, + 0.00899157952517271, + 0.002406519604846835, + 0.001196603523567319, + 0.006735143717378378, + 0.0022194127086549997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.029790453612804413 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02977564185857773 + ], + "opposite_logprob_tokens": [ + -3.529775619506836 + ], + "logprob_reward": -0.02977564185857773, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02977564185857773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2792380154132843, + 0.10368216037750244, + 0.11038247495889664, + 0.1202707514166832, + 0.14123763144016266, + 0.08668775111436844, + 0.02886587753891945, + 0.026739992201328278, + 0.050369758158922195, + 0.0525255911052227 + ], + "credit_map": [ + 0.0335492268204689, + 0.030278343707323074, + 0.026479411870241165, + 0.03986084461212158, + 0.0318191833794117, + 0.05375833436846733, + 0.03386596217751503, + 0.029626695439219475, + 0.009100774303078651, + 0.01729300059378147, + 0.01123067457228899, + 0.022427737712860107, + 0.008314504288136959, + 0.006979248486459255, + 0.015656646341085434, + 0.012679570354521275, + 0.010610274970531464, + 0.013752368278801441, + 0.013941905461251736, + 0.01845218613743782, + 0.01082511804997921, + 0.011212739162147045, + 0.0168131235986948, + 0.014774763956665993, + 0.011410636827349663, + 0.023894574493169785, + 0.01085655763745308, + 0.022042492404580116, + 0.015447104349732399, + 0.010073604062199593, + 0.014645174145698547, + 0.011900612153112888, + 0.01654479280114174, + 0.020211851224303246, + 0.021687710657715797, + 0.019434094429016113, + 0.010971681214869022, + 0.017532575875520706, + 0.016600003466010094, + 0.018254926428198814, + 0.012346671894192696, + 0.010929051786661148, + 0.01395296212285757, + 0.012486576102674007, + 0.008604181930422783, + 0.00865214690566063, + 0.009112786501646042, + 0.010603372007608414, + 0.0021438216790556908, + 0.0027145224157720804, + 0.002901465864852071, + 0.006087432149797678, + 0.002822056645527482, + 0.0025512236170470715, + 0.007303496822714806, + 0.0023418599739670753, + 0.003850878681987524, + 0.003454804653301835, + 0.003387083765119314, + 0.002591473516076803, + 0.003074972191825509, + 0.0045065078884363174, + 0.0028838247526437044, + 0.0029904458206146955, + 0.005816486664116383, + 0.007244796957820654, + 0.005832248833030462, + 0.011410491541028023, + 0.0046907151117920876, + 0.0035888003185391426, + 0.006374002434313297, + 0.005412218160927296, + 0.005377760622650385, + 0.007197702303528786, + 0.005576042924076319, + 0.007816799916327, + 0.007632801774889231, + 0.005816674325615168, + 0.006792111322283745, + 0.006315696518868208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.02977564185857773 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943719297647476 + ], + "opposite_logprob_tokens": [ + -5.130943775177002 + ], + "logprob_reward": -0.005943719297647476, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943719297647476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.240119069814682, + 0.2852940559387207, + 0.04678753390908241, + 0.10095597058534622, + 0.05941496789455414, + 0.052002836018800735, + 0.02592831663787365, + 0.012267868965864182, + 0.04019949212670326, + 0.13702987134456635 + ], + "credit_map": [ + 0.023454664275050163, + 0.03791113942861557, + 0.03719267621636391, + 0.032378118485212326, + 0.023295937106013298, + 0.03326503559947014, + 0.03218787536025047, + 0.020433641970157623, + 0.013896475546061993, + 0.03208981081843376, + 0.01907018944621086, + 0.11007913202047348, + 0.015453191474080086, + 0.007036448922008276, + 0.07389683276414871, + 0.013772006146609783, + 0.0040456922724843025, + 0.006525525823235512, + 0.006168108433485031, + 0.00731660658493638, + 0.0055318558588624, + 0.0044034854508936405, + 0.007721357513219118, + 0.005074909422546625, + 0.011501260101795197, + 0.016079362481832504, + 0.008698927238583565, + 0.014526817947626114, + 0.013018375262618065, + 0.012927457690238953, + 0.01152341440320015, + 0.012680359184741974, + 0.006768610328435898, + 0.006298957392573357, + 0.0077796620316803455, + 0.009021484293043613, + 0.006258015986531973, + 0.007869218476116657, + 0.008463895879685879, + 0.006955127697438002, + 0.004395672585815191, + 0.005761314649134874, + 0.005675082094967365, + 0.012265258468687534, + 0.005234096199274063, + 0.0043848007917404175, + 0.010206758975982666, + 0.0040798550471663475, + 0.002230510814115405, + 0.0035957396030426025, + 0.0026378061156719923, + 0.003295332193374634, + 0.004840061999857426, + 0.0023369607515633106, + 0.004033546894788742, + 0.002958358731120825, + 0.0013706700410693884, + 0.001700538326986134, + 0.001449520350433886, + 0.0016129622235894203, + 0.001929992693476379, + 0.0012321753893047571, + 0.001636172761209309, + 0.0013358386931940913, + 0.004431387409567833, + 0.005013103596866131, + 0.0033531785011291504, + 0.009743896313011646, + 0.005695654079318047, + 0.0028690986800938845, + 0.005777997430413961, + 0.0033151800744235516, + 0.007783791981637478, + 0.00984286516904831, + 0.010279584676027298, + 0.0503380186855793, + 0.008680779486894608, + 0.007165520451962948, + 0.035096220672130585, + 0.007843109779059887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.005943719297647476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009182051289826632 + ], + "opposite_logprob_tokens": [ + -7.000918388366699 + ], + "logprob_reward": -0.0009182051289826632, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009182051289826632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19905433058738708, + 0.19376437366008759, + 0.06639281660318375, + 0.08440002799034119, + 0.06179816648364067, + 0.07726185023784637, + 0.047989897429943085, + 0.04575087130069733, + 0.13458482921123505, + 0.08900286257266998 + ], + "credit_map": [ + 0.027450310066342354, + 0.026288336142897606, + 0.02842624858021736, + 0.031589996069669724, + 0.022001376375555992, + 0.01866849511861801, + 0.02605539560317993, + 0.018574154004454613, + 0.009373418055474758, + 0.015194403938949108, + 0.016280673444271088, + 0.07246315479278564, + 0.008759180083870888, + 0.006951236631721258, + 0.052059926092624664, + 0.012682384811341763, + 0.005839087534695864, + 0.009811951778829098, + 0.008743737824261189, + 0.008275533095002174, + 0.009755906648933887, + 0.0064950911328196526, + 0.008743740618228912, + 0.008727764710783958, + 0.006656791549175978, + 0.012602356262505054, + 0.005832744762301445, + 0.01242389902472496, + 0.01611059159040451, + 0.008807046338915825, + 0.01228718739002943, + 0.009679405018687248, + 0.005961790215224028, + 0.007118815556168556, + 0.008177254348993301, + 0.009770537726581097, + 0.007066573482006788, + 0.006043048109859228, + 0.011225731112062931, + 0.006434407085180283, + 0.004846824798732996, + 0.009982187300920486, + 0.007296459283679724, + 0.019861118867993355, + 0.009631435386836529, + 0.0045444113202393055, + 0.01579117216169834, + 0.005308234132826328, + 0.003539887722581625, + 0.004946901462972164, + 0.0024429208133369684, + 0.010700886137783527, + 0.005438971798866987, + 0.0025386824272572994, + 0.014617917127907276, + 0.0037637269124388695, + 0.0055799707770347595, + 0.006070410832762718, + 0.004316105507314205, + 0.00929232593625784, + 0.005778376944363117, + 0.003928271122276783, + 0.006696157157421112, + 0.004089249763637781, + 0.01991460472345352, + 0.013139737769961357, + 0.01833363249897957, + 0.021897777915000916, + 0.012819025665521622, + 0.013012206181883812, + 0.019592028111219406, + 0.015875810757279396, + 0.007907463237643242, + 0.007129854056984186, + 0.00755677092820406, + 0.02623961679637432, + 0.006602756679058075, + 0.00604118499904871, + 0.020026765763759613, + 0.007498445920646191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0009182051289826632 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, at eye level, captures the exuberant energy of Wadadou, a cheerful, anthropomorphic taiko drummer rendered in a vibrant pop art style defined by bold, flat colors and thick, decisive black outlines. His body is a perfect bright red cylinder, his face nestled against the surface of the drum, his large, expressive eyes closed in musical ecstasy, his mouth wide open as he belts out the song. He passionately grasps a classic silver microphone in one hand, gripping it tightly, his entire body seemingly vibrating with the power of his voice. The scene is set in a karaoke room, its backdrop a flat, electric yellow wall covered in a repeating pattern of large cyan dots. Multi-colored strobe lights rhythmically flicker and flash, casting rapidly changing patterns of fuchsia and green onto the character and the room, creating sharp, black-outlined shadows that dance and pulsate to the beat, amplifying the vibrant, joyful atmosphere of the throat performance.", + "seed": 30769, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000102/group_9/index_2530_seed_30769_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02966035179561004, + "vlm_logprob": -0.02966035179561004, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.02966035179561004, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017785940319299698 + ], + "opposite_logprob_tokens": [ + -4.142786026000977 + ], + "logprob_reward": -0.017785940319299698, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.017785940319299698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23066656291484833, + 0.10391835868358612, + 0.1370764970779419, + 0.05185159295797348, + 0.10935546457767487, + 0.10529390722513199, + 0.07878110557794571, + 0.050105128437280655, + 0.047780200839042664, + 0.085171177983284 + ], + "credit_map": [ + 0.03124019131064415, + 0.02200753055512905, + 0.02516784332692623, + 0.024973249062895775, + 0.023374944925308228, + 0.031166676431894302, + 0.019778583198785782, + 0.05295751616358757, + 0.01109872292727232, + 0.012702512554824352, + 0.009127078577876091, + 0.012660067528486252, + 0.017852691933512688, + 0.012987940572202206, + 0.015016136690974236, + 0.012473205104470253, + 0.016710376366972923, + 0.016275443136692047, + 0.011983592063188553, + 0.018311556428670883, + 0.02175956591963768, + 0.01835957169532776, + 0.014035817235708237, + 0.019640563055872917, + 0.006338631734251976, + 0.007514614146202803, + 0.006411384325474501, + 0.006276099011301994, + 0.006288942880928516, + 0.007693510502576828, + 0.004757679533213377, + 0.006570730358362198, + 0.0114236855879426, + 0.01488464418798685, + 0.010014249943196774, + 0.014015967026352882, + 0.015822751447558403, + 0.017810506746172905, + 0.011323755607008934, + 0.01405989658087492, + 0.01265977043658495, + 0.009914836846292019, + 0.009141124784946442, + 0.013063832186162472, + 0.01569340191781521, + 0.018347399309277534, + 0.015623211860656738, + 0.010850326158106327, + 0.010970246978104115, + 0.005826556123793125, + 0.005161968991160393, + 0.0052118850871920586, + 0.007227650843560696, + 0.02326910011470318, + 0.00627043517306447, + 0.014843258075416088, + 0.006152373738586903, + 0.007755750324577093, + 0.0036233982536941767, + 0.006212047301232815, + 0.01095999963581562, + 0.0051757898181676865, + 0.006216504145413637, + 0.004009264055639505, + 0.006306449882686138, + 0.006705311592668295, + 0.0027781531680375338, + 0.0045917765237390995, + 0.01214924268424511, + 0.0038745321799069643, + 0.006881265435367823, + 0.004493465647101402, + 0.01397500280290842, + 0.011661684140563011, + 0.005943938624113798, + 0.011745737865567207, + 0.013063594698905945, + 0.012699421495199203, + 0.00837151799350977, + 0.007710274308919907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.017785940319299698 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000847905874252 + ], + "opposite_logprob_tokens": [ + -2.6950085163116455 + ], + "logprob_reward": -0.07000847905874252, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000847905874252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11214008182287216, + 0.08767243474721909, + 0.15618999302387238, + 0.07520489394664764, + 0.09981321543455124, + 0.09787917137145996, + 0.2728196084499359, + 0.02689380943775177, + 0.04237053170800209, + 0.029016312211751938 + ], + "credit_map": [ + 0.01851540058851242, + 0.008899800479412079, + 0.011906852945685387, + 0.009120703674852848, + 0.008507764898240566, + 0.03633592277765274, + 0.006630034185945988, + 0.012223605066537857, + 0.009721734561026096, + 0.0114328907802701, + 0.007872085087001324, + 0.009267295710742474, + 0.014286176301538944, + 0.013287629932165146, + 0.012075016275048256, + 0.009729604236781597, + 0.018786510452628136, + 0.022006426006555557, + 0.01719536818563938, + 0.01957199163734913, + 0.024371162056922913, + 0.02038278430700302, + 0.015296900644898415, + 0.018578851595520973, + 0.00891065876930952, + 0.011552485637366772, + 0.007633906789124012, + 0.006254999898374081, + 0.006157303228974342, + 0.013381807133555412, + 0.005660030525177717, + 0.0156537052243948, + 0.00918591395020485, + 0.011876003816723824, + 0.010999520309269428, + 0.009258056990802288, + 0.01543069165199995, + 0.018898002803325653, + 0.010501731187105179, + 0.013663291931152344, + 0.011888806708157063, + 0.008422335609793663, + 0.009615994989871979, + 0.010357039980590343, + 0.01546011958271265, + 0.016619892790913582, + 0.014281830750405788, + 0.01123314630240202, + 0.03476988524198532, + 0.024958640336990356, + 0.02136858180165291, + 0.022221876308321953, + 0.04866299033164978, + 0.0535392165184021, + 0.031090548262000084, + 0.03620785474777222, + 0.0032737641595304012, + 0.003184559755027294, + 0.002328534610569477, + 0.0023568386677652597, + 0.003425775095820427, + 0.004868120886385441, + 0.003137819468975067, + 0.004318397492170334, + 0.004612405318766832, + 0.007081778720021248, + 0.0034079633187502623, + 0.005758956074714661, + 0.008088415488600731, + 0.004301009699702263, + 0.005004492588341236, + 0.00411551259458065, + 0.00412118062376976, + 0.003933224361389875, + 0.003180061699822545, + 0.003547798143699765, + 0.004546192474663258, + 0.004360112827271223, + 0.002313984790816903, + 0.0030137556605041027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.07000847905874252 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048625484108924866 + ], + "opposite_logprob_tokens": [ + -3.0486254692077637 + ], + "logprob_reward": -0.048625484108924866, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048625484108924866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09271862357854843, + 0.1576520800590515, + 0.041340943425893784, + 0.04870327189564705, + 0.040259383618831635, + 0.21180684864521027, + 0.13701829314231873, + 0.18121114373207092, + 0.025432972237467766, + 0.0638563483953476 + ], + "credit_map": [ + 0.013060773722827435, + 0.010193166323006153, + 0.006174794863909483, + 0.013383454643189907, + 0.008163150399923325, + 0.019160941243171692, + 0.009702037088572979, + 0.012880322523415089, + 0.01891404204070568, + 0.014540772885084152, + 0.023229064419865608, + 0.013226654380559921, + 0.01523874793201685, + 0.037720199674367905, + 0.014423709362745285, + 0.020358914509415627, + 0.0051413956098258495, + 0.0076871332712471485, + 0.003344783326610923, + 0.0051962342113256454, + 0.0052861412987113, + 0.00595677737146616, + 0.0037215950433164835, + 0.005006885621696711, + 0.007167380303144455, + 0.010633682832121849, + 0.006235006731003523, + 0.005767663940787315, + 0.005374190863221884, + 0.004567073192447424, + 0.004084727261215448, + 0.004873554687947035, + 0.004004223737865686, + 0.006591483019292355, + 0.002501873765140772, + 0.005269912071526051, + 0.008888446725904942, + 0.005539554636925459, + 0.004157444462180138, + 0.003306447993963957, + 0.02367616444826126, + 0.024431729689240456, + 0.016673069447278976, + 0.02658793143928051, + 0.038802120834589005, + 0.026229700073599815, + 0.033617518842220306, + 0.021788638085126877, + 0.015780603513121605, + 0.011742993257939816, + 0.008569454774260521, + 0.010769806802272797, + 0.023654066026210785, + 0.03390295058488846, + 0.011010904796421528, + 0.021587513387203217, + 0.01917334273457527, + 0.02297019399702549, + 0.014817778952419758, + 0.023698268458247185, + 0.03408064693212509, + 0.01789981685578823, + 0.030166177079081535, + 0.018404942005872726, + 0.0027735845651477575, + 0.003585129976272583, + 0.0026748126838356256, + 0.0020767145324498415, + 0.006160954013466835, + 0.0025943194050341845, + 0.0030707651749253273, + 0.0024966951459646225, + 0.009294400922954082, + 0.008399666287004948, + 0.0058616455644369125, + 0.008221465162932873, + 0.011858784593641758, + 0.009180343709886074, + 0.005578320473432541, + 0.0054617272689938545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.048625484108924866 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071322485804558 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.011071322485804558, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071322485804558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05260222777724266, + 0.11067754775285721, + 0.04022654891014099, + 0.04061303660273552, + 0.044410716742277145, + 0.2228420227766037, + 0.04552268981933594, + 0.05200892314314842, + 0.018784714862704277, + 0.3723115921020508 + ], + "credit_map": [ + 0.006843114271759987, + 0.004982631653547287, + 0.004918626043945551, + 0.007742396555840969, + 0.005068199243396521, + 0.011332903057336807, + 0.004448084160685539, + 0.007266269531100988, + 0.015079641714692116, + 0.01076591294258833, + 0.01493613701313734, + 0.009347301907837391, + 0.01323238667100668, + 0.020672300830483437, + 0.01073410827666521, + 0.01590975746512413, + 0.004770167637616396, + 0.006462331395596266, + 0.00440017506480217, + 0.005100790411233902, + 0.0050182463601231575, + 0.006143047474324703, + 0.004508753307163715, + 0.003823033766821027, + 0.0050885858945548534, + 0.008669624105095863, + 0.0027814251370728016, + 0.0037473002448678017, + 0.004805551376193762, + 0.006476125214248896, + 0.0031223620753735304, + 0.005922059528529644, + 0.004426275845617056, + 0.005556576885282993, + 0.006992637179791927, + 0.00378982862457633, + 0.005064792465418577, + 0.00865645706653595, + 0.004360326565802097, + 0.005563822574913502, + 0.029371488839387894, + 0.02391728013753891, + 0.017856495454907417, + 0.026507772505283356, + 0.040665898472070694, + 0.026829974725842476, + 0.035312578082084656, + 0.022380542010068893, + 0.006052614189684391, + 0.003510613227263093, + 0.0028926280792802572, + 0.00360558251850307, + 0.003712721634656191, + 0.015056292526423931, + 0.003071059240028262, + 0.0076211788691580296, + 0.0050160265527665615, + 0.006935830228030682, + 0.0034113277215510607, + 0.006237954832613468, + 0.010199910029768944, + 0.005878618452697992, + 0.0076868231408298016, + 0.006642433349043131, + 0.0023356785532087088, + 0.003334046807140112, + 0.0011510367039591074, + 0.0011688991216942668, + 0.005128733813762665, + 0.0016135742189362645, + 0.0028049491811543703, + 0.001247796812094748, + 0.051816873252391815, + 0.046278756111860275, + 0.03766166418790817, + 0.03981020674109459, + 0.07159718126058578, + 0.06289860606193542, + 0.028028899803757668, + 0.034219421446323395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 1 + ], + "selected_reward": -0.011071322485804558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character an anthropomorphic taiko drummer with a bright red cylindrical body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character holding a classic silver microphone in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a flat electric yellow wall covered in a repeating pattern of large cyan dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are multi-colored strobe lights casting rapidly changing patterns of fuchsia and green onto the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style defined by bold flat colors and thick decisive black outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008105330052785575 + ], + "opposite_logprob_tokens": [ + -7.125810623168945 + ], + "logprob_reward": -0.0008105330052785575, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008105330052785575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1052187979221344, + 0.10413399338722229, + 0.1305934637784958, + 0.09584333002567291, + 0.1005406305193901, + 0.06991560012102127, + 0.2223055362701416, + 0.02881951443850994, + 0.04217094928026199, + 0.10045818239450455 + ], + "credit_map": [ + 0.012579124420881271, + 0.010614188387989998, + 0.009237564168870449, + 0.015666136518120766, + 0.008318989537656307, + 0.01876056380569935, + 0.0077625843696296215, + 0.022279636934399605, + 0.011263562366366386, + 0.009502013213932514, + 0.011381183750927448, + 0.012170778587460518, + 0.014478490687906742, + 0.01848764903843403, + 0.011142878793179989, + 0.015707433223724365, + 0.016173169016838074, + 0.01986234448850155, + 0.011450848542153835, + 0.018140610307455063, + 0.014283890835940838, + 0.013132734224200249, + 0.016389166936278343, + 0.02116069197654724, + 0.013211657293140888, + 0.02215423993766308, + 0.010401418432593346, + 0.011544136330485344, + 0.009756477549672127, + 0.008779062889516354, + 0.009456970728933811, + 0.01053935568779707, + 0.010578261688351631, + 0.009746271185576916, + 0.01511839684098959, + 0.01093059778213501, + 0.01055724173784256, + 0.018900491297245026, + 0.009492553770542145, + 0.015216800384223461, + 0.011846862733364105, + 0.006548743229359388, + 0.007963424548506737, + 0.005874310620129108, + 0.010319173336029053, + 0.008809205144643784, + 0.0089338393881917, + 0.00962003879249096, + 0.029290584847331047, + 0.019016772508621216, + 0.00979197770357132, + 0.01630408875644207, + 0.04874780774116516, + 0.0440998338162899, + 0.020931681618094444, + 0.034122779965400696, + 0.00320274755358696, + 0.003916118294000626, + 0.0025422298349440098, + 0.003651545150205493, + 0.006049774587154388, + 0.0023344496730715036, + 0.004163967911154032, + 0.002958679338917136, + 0.004618586041033268, + 0.007161292713135481, + 0.001974673243239522, + 0.00531005859375, + 0.00998752936720848, + 0.0037306786980479956, + 0.006185116712003946, + 0.0032030099537223577, + 0.013762226328253746, + 0.012629678472876549, + 0.008928641676902771, + 0.013462628237903118, + 0.019077738747000694, + 0.01619012840092182, + 0.008968819864094257, + 0.007438307628035545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.0008105330052785575 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_0/index_tensor([3393, 3393])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_0/index_tensor([3393, 3393])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f458550706b75aa70f165d866418ca4216c9bba8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_0/index_tensor([3393, 3393])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long, newsreel-style shot, starting at eye level and slowly zooming in, focuses on an elderly eunuch, the chief eunuch of the Qing Dynasty, standing alone by a carved wooden window in a quiet study in the Forbidden City. Soft afternoon sunlight slanted through the thin paper covering the window pane, skewing the air with dancing dust and illuminating the delicate texture of his dark blue satin uniform, a patterned pattern. He was about sixty years old, with a shiny forehead and a head of graying hair that hung down to his waist. His pale, finely lined face was expressionless, only the dark green jade ring on his right thumb gleamed in the warm light. Carefully, his hands held a bright yellow memorial, his slender fingers moving steadily and deliberately along the neatly written calligraphy of the imperial court. As he reached a certain point, he gently lifted a page with his fingertips, his movements so gentle that they made almost no sound. At this moment, the camera has moved to a close-up of his face, clearly capturing a complex expression that flashed across his usually calm face: his brows frowned very slightly, his eyes withdrew from his concentrated reading, and briefly looked into the void, with a mixture of surprise and deep calculation in his pupils, and then the corners of his mouth sank imperceptibly, but all this lasted for less than a second, and his expression quickly returned to his usual, unfathomable calm. The whole process was completed in extreme silence, creating a tense atmosphere that contained a huge amount of information but was kept secret.", + "seed": 27727, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_0/index_3393_seed_27727_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04042419418692589, + "vlm_logprob": -0.04042419418692589, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.04042419418692589, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01653737761080265 + ], + "opposite_logprob_tokens": [ + -5.016537189483643 + ], + "logprob_reward": -0.01653737761080265, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.01653737761080265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2834921181201935, + 0.05286330357193947, + 0.03790689632296562, + 0.12952013313770294, + 0.07839127629995346, + 0.08227353543043137, + 0.08931563794612885, + 0.11030436307191849, + 0.07458771020174026, + 0.06134500727057457 + ], + "credit_map": [ + 0.07738269120454788, + 0.0463009811937809, + 0.06044624000787735, + 0.012224911712110043, + 0.01977313868701458, + 0.010625945404171944, + 0.017173966392874718, + 0.039564236998558044, + 0.0071975658647716045, + 0.009028596803545952, + 0.005970555357635021, + 0.004466955550014973, + 0.007820678874850273, + 0.005233969073742628, + 0.006715876515954733, + 0.006429104600101709, + 0.004351776093244553, + 0.004190970212221146, + 0.005037510301917791, + 0.004826551768928766, + 0.005802133120596409, + 0.002378420904278755, + 0.0068759857676923275, + 0.004443550482392311, + 0.01509963721036911, + 0.017987562343478203, + 0.017615072429180145, + 0.021161464974284172, + 0.011552920565009117, + 0.005536897107958794, + 0.02614976279437542, + 0.01441681757569313, + 0.008297651074826717, + 0.004950833041220903, + 0.009772684425115585, + 0.011923860758543015, + 0.012184635736048222, + 0.0034506951924413443, + 0.015032301656901836, + 0.012778613716363907, + 0.009706822223961353, + 0.005227622110396624, + 0.011173276230692863, + 0.007387413643300533, + 0.016431361436843872, + 0.00507825193926692, + 0.013528889045119286, + 0.013739894144237041, + 0.009214678779244423, + 0.004696437623351812, + 0.011175510473549366, + 0.008303211070597172, + 0.009640880860388279, + 0.0059057739563286304, + 0.016830815002322197, + 0.023548336699604988, + 0.01685917377471924, + 0.006116110365837812, + 0.019559212028980255, + 0.006241650320589542, + 0.025923311710357666, + 0.007628671824932098, + 0.01630709320306778, + 0.011669144965708256, + 0.012944575399160385, + 0.006298786960542202, + 0.010560538619756699, + 0.012598670087754726, + 0.008540602400898933, + 0.005792134907096624, + 0.00967756099998951, + 0.008174840360879898, + 0.006646226160228252, + 0.004825293552130461, + 0.008271533064544201, + 0.005174882709980011, + 0.009280446916818619, + 0.012447621673345566, + 0.007710856851190329, + 0.006988150998950005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.01653737761080265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207584589719772 + ], + "opposite_logprob_tokens": [ + -2.8120758533477783 + ], + "logprob_reward": -0.06207584589719772, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207584589719772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13168221712112427, + 0.10280349105596542, + 0.028897318989038467, + 0.04139655828475952, + 0.0719018429517746, + 0.14962010085582733, + 0.12204774469137192, + 0.07456658035516739, + 0.10662627220153809, + 0.17045795917510986 + ], + "credit_map": [ + 0.011898563243448734, + 0.01829422265291214, + 0.022931214421987534, + 0.013204354792833328, + 0.015619617886841297, + 0.01335869263857603, + 0.018218474462628365, + 0.01815706491470337, + 0.020257072523236275, + 0.015309135429561138, + 0.015996932983398438, + 0.006718698889017105, + 0.01160421408712864, + 0.0120663708075881, + 0.008770219050347805, + 0.012080837972462177, + 0.0036634986754506826, + 0.0033491041976958513, + 0.003318640636280179, + 0.004119171295315027, + 0.0031052727717906237, + 0.0015305246924981475, + 0.005765913985669613, + 0.004045193083584309, + 0.008581269532442093, + 0.004940339829772711, + 0.004266094882041216, + 0.005269899964332581, + 0.005237681325525045, + 0.0021325405687093735, + 0.006773361936211586, + 0.0041953688487410545, + 0.016011495143175125, + 0.005860028322786093, + 0.011067340150475502, + 0.008233924396336079, + 0.010020325891673565, + 0.004950486123561859, + 0.0070027257315814495, + 0.008755505084991455, + 0.021044788882136345, + 0.019603079184889793, + 0.02097972482442856, + 0.011906418949365616, + 0.02750563807785511, + 0.01373377162963152, + 0.01658843457698822, + 0.01825822703540325, + 0.018840918317437172, + 0.01688440889120102, + 0.016868026927113533, + 0.009288277477025986, + 0.025079846382141113, + 0.007282321806997061, + 0.01286072377115488, + 0.014943219721317291, + 0.008785955607891083, + 0.009050915017724037, + 0.010026732459664345, + 0.004696639254689217, + 0.019952276721596718, + 0.005971598904579878, + 0.008003287948668003, + 0.008079169318079948, + 0.019274504855275154, + 0.010431086644530296, + 0.01420161034911871, + 0.009381693787872791, + 0.014803717844188213, + 0.008934399113059044, + 0.018050549551844597, + 0.011548702605068684, + 0.025238998234272003, + 0.01984427310526371, + 0.02614680863916874, + 0.014226905070245266, + 0.023181777447462082, + 0.023944087326526642, + 0.023236550390720367, + 0.014638549648225307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.06207584589719772 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819690316915512 + ], + "opposite_logprob_tokens": [ + -4.018197059631348 + ], + "logprob_reward": -0.01819690316915512, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819690316915512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34560707211494446, + 0.08068692684173584, + 0.09973260015249252, + 0.0633491799235344, + 0.08101345598697662, + 0.04049261659383774, + 0.05785592645406723, + 0.08504380285739899, + 0.09645567834377289, + 0.04976273328065872 + ], + "credit_map": [ + 0.07571805268526077, + 0.06181412190198898, + 0.0582948662340641, + 0.026795243844389915, + 0.024180913344025612, + 0.014884213916957378, + 0.028732847422361374, + 0.055186815559864044, + 0.010518413968384266, + 0.008963093161582947, + 0.010532292537391186, + 0.010849347338080406, + 0.008520088158547878, + 0.00517041701823473, + 0.01510876975953579, + 0.01102449931204319, + 0.011767229065299034, + 0.010158790275454521, + 0.0086597865447402, + 0.014032579958438873, + 0.01739238016307354, + 0.006339031737297773, + 0.016190817579627037, + 0.015191987156867981, + 0.00845913216471672, + 0.008766480721533298, + 0.007924981415271759, + 0.0066215707920491695, + 0.007706882897764444, + 0.002786537166684866, + 0.012804578989744186, + 0.008279013447463512, + 0.010522693395614624, + 0.006873036734759808, + 0.012094543315470219, + 0.017210517078638077, + 0.011691614054143429, + 0.0032462691888213158, + 0.010064637288451195, + 0.009310143068432808, + 0.004549146164208651, + 0.0031233818735927343, + 0.004480336327105761, + 0.005311216227710247, + 0.009693733416497707, + 0.0022335487883538008, + 0.005897706374526024, + 0.005203550681471825, + 0.007339554373174906, + 0.004710806999355555, + 0.007337072864174843, + 0.0061577437445521355, + 0.009561186656355858, + 0.003650950500741601, + 0.008244155906140804, + 0.010854460299015045, + 0.010573047213256359, + 0.005772901698946953, + 0.013339336030185223, + 0.006541666109114885, + 0.021397382020950317, + 0.007303615100681782, + 0.01137478556483984, + 0.00874107051640749, + 0.018531177192926407, + 0.006397345568984747, + 0.015315033495426178, + 0.008468225598335266, + 0.01177988201379776, + 0.008386588655412197, + 0.014658239670097828, + 0.012919182889163494, + 0.005577758885920048, + 0.003142282832413912, + 0.0060704536736011505, + 0.0033013871870934963, + 0.008122939616441727, + 0.009239323437213898, + 0.008672048337757587, + 0.0056365421041846275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.01819690316915512 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.078971266746521 + ], + "opposite_logprob_tokens": [ + -2.5789713859558105 + ], + "logprob_reward": -0.078971266746521, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.078971266746521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10908900201320648, + 0.06690732389688492, + 0.05480318143963814, + 0.09143642336130142, + 0.16353276371955872, + 0.13551855087280273, + 0.159397155046463, + 0.04649660363793373, + 0.07341811805963516, + 0.09940094500780106 + ], + "credit_map": [ + 0.01836097612977028, + 0.015113325789570808, + 0.0152003588154912, + 0.010331211611628532, + 0.010075127705931664, + 0.006704359780997038, + 0.015410667285323143, + 0.017892973497509956, + 0.008332635276019573, + 0.008264406584203243, + 0.007490349933505058, + 0.007315647788345814, + 0.0091122817248106, + 0.006735396105796099, + 0.011965539306402206, + 0.007691072765737772, + 0.005646219942718744, + 0.0051064686849713326, + 0.007235675118863583, + 0.004420271143317223, + 0.012413255870342255, + 0.003595638321712613, + 0.00993705727159977, + 0.006448591593652964, + 0.009076754562556744, + 0.007732247933745384, + 0.012445596978068352, + 0.013020985759794712, + 0.009815212339162827, + 0.003766644047573209, + 0.02382049709558487, + 0.011758479289710522, + 0.02971666119992733, + 0.013048290275037289, + 0.02198784425854683, + 0.025912312790751457, + 0.021923301741480827, + 0.007360367104411125, + 0.023758405819535255, + 0.01982557587325573, + 0.01677352376282215, + 0.008921678178012371, + 0.01530887559056282, + 0.014353957027196884, + 0.02308233082294464, + 0.007241401821374893, + 0.028649117797613144, + 0.021187668666243553, + 0.017182204872369766, + 0.008198101073503494, + 0.01798863895237446, + 0.012773830443620682, + 0.03230471909046173, + 0.009408796206116676, + 0.032269541174173355, + 0.029271315783262253, + 0.006758729927241802, + 0.003753594122827053, + 0.007859627716243267, + 0.0034722068812698126, + 0.012406463734805584, + 0.002916284603998065, + 0.004609938710927963, + 0.004719758406281471, + 0.012074746191501617, + 0.005761676002293825, + 0.010038732551038265, + 0.007689945865422487, + 0.015308210626244545, + 0.006193740293383598, + 0.007162305060774088, + 0.009188761934638023, + 0.01374212559312582, + 0.004417745396494865, + 0.017680222168564796, + 0.005443960428237915, + 0.01285486575216055, + 0.011327957734465599, + 0.016369715332984924, + 0.017564354464411736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.078971266746521 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633957751095295 + ], + "opposite_logprob_tokens": [ + -3.651339530944824 + ], + "logprob_reward": -0.02633957751095295, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633957751095295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09593163430690765, + 0.04560592770576477, + 0.047329120337963104, + 0.04423113539814949, + 0.14562328159809113, + 0.0709337592124939, + 0.11603240668773651, + 0.10719921439886093, + 0.10983506590127945, + 0.21727851033210754 + ], + "credit_map": [ + 0.020916789770126343, + 0.014031194150447845, + 0.012603459879755974, + 0.009524751454591751, + 0.006307424511760473, + 0.006019524298608303, + 0.012445973232388496, + 0.014082510955631733, + 0.008154112845659256, + 0.006337560713291168, + 0.005186127964407206, + 0.00508839963003993, + 0.005459529813379049, + 0.003967531491070986, + 0.006136666983366013, + 0.0052759964019060135, + 0.005467982497066259, + 0.005801143124699593, + 0.005356325767934322, + 0.004708546679466963, + 0.007216086145490408, + 0.0036352677270770073, + 0.008785589598119259, + 0.006358179729431868, + 0.00502475630491972, + 0.003302005585283041, + 0.005981564987450838, + 0.005667879246175289, + 0.005967678036540747, + 0.0017800676869228482, + 0.010172472335398197, + 0.006334707140922546, + 0.024065999314188957, + 0.011059239506721497, + 0.019505925476551056, + 0.02570500783622265, + 0.02129332534968853, + 0.010036089457571507, + 0.017484067007899284, + 0.01647363044321537, + 0.013229132629930973, + 0.005399051588028669, + 0.010977543890476227, + 0.006591951008886099, + 0.012279878370463848, + 0.003512740833684802, + 0.010195361450314522, + 0.008748099207878113, + 0.01108340360224247, + 0.007767053786665201, + 0.013823740184307098, + 0.009400385431945324, + 0.02424810267984867, + 0.007692975923418999, + 0.01988244615495205, + 0.02213429845869541, + 0.01221416611224413, + 0.007705953903496265, + 0.01315563078969717, + 0.009244234301149845, + 0.030107468366622925, + 0.009285750798881054, + 0.012864316813647747, + 0.012621697038412094, + 0.01643224246799946, + 0.004515251610428095, + 0.01146206445991993, + 0.0069031077437102795, + 0.021219635382294655, + 0.008601965382695198, + 0.022852009162306786, + 0.01784878596663475, + 0.04052848368883133, + 0.011196146719157696, + 0.05412124842405319, + 0.013665742240846157, + 0.02478189580142498, + 0.023555520921945572, + 0.021412359550595284, + 0.028017112985253334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.02633957751095295 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, newsreel-style shot, starting at eye level and slowly zooming in, focuses on an elderly eunuch, the chief eunuch of the Qing Dynasty, standing alone by a carved wooden window in a quiet study in the Forbidden City. Soft afternoon sunlight slanted through the thin paper covering the window pane, skewing the air with dancing dust and illuminating the delicate texture of his dark blue satin uniform, a patterned pattern. He was about sixty years old, with a shiny forehead and a head of graying hair that hung down to his waist. His pale, finely lined face was expressionless, only the dark green jade ring on his right thumb gleamed in the warm light. Carefully, his hands held a bright yellow memorial, his slender fingers moving steadily and deliberately along the neatly written calligraphy of the imperial court. As he reached a certain point, he gently lifted a page with his fingertips, his movements so gentle that they made almost no sound. At this moment, the camera has moved to a close-up of his face, clearly capturing a complex expression that flashed across his usually calm face: his brows frowned very slightly, his eyes withdrew from his concentrated reading, and briefly looked into the void, with a mixture of surprise and deep calculation in his pupils, and then the corners of his mouth sank imperceptibly, but all this lasted for less than a second, and his expression quickly returned to his usual, unfathomable calm. The whole process was completed in extreme silence, creating a tense atmosphere that contained a huge amount of information but was kept secret.", + "seed": 27728, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_0/index_3393_seed_27728_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01631452590227127, + "vlm_logprob": -0.01631452590227127, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.01631452590227127, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009157324209809303 + ], + "opposite_logprob_tokens": [ + -6.509157180786133 + ], + "logprob_reward": -0.009157324209809303, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.009157324209809303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056307971477508545, + 0.0262092687189579, + 0.027677124366164207, + 0.07713101804256439, + 0.017006278038024902, + 0.05813341587781906, + 0.270413875579834, + 0.04880137741565704, + 0.08454708009958267, + 0.33377254009246826 + ], + "credit_map": [ + 0.009697921574115753, + 0.0035429487470537424, + 0.00796396192163229, + 0.010119729675352573, + 0.00721969036385417, + 0.003915785346180201, + 0.0060109482146799564, + 0.00783698633313179, + 0.002457832684740424, + 0.0014098018873482943, + 0.004503635689616203, + 0.003760319435968995, + 0.0035898920614272356, + 0.0018980327295139432, + 0.0035039125941693783, + 0.005085844546556473, + 0.004176930524408817, + 0.0016589302103966475, + 0.00475755101069808, + 0.00326096941716969, + 0.004380315076559782, + 0.0019878563471138477, + 0.003440849483013153, + 0.004013725556433201, + 0.0035462111700326204, + 0.0018320453818887472, + 0.021695591509342194, + 0.007660203613340855, + 0.003941117785871029, + 0.003475125879049301, + 0.006684429943561554, + 0.028296297416090965, + 0.0014666299102827907, + 0.001406910945661366, + 0.002415164839476347, + 0.001988665433600545, + 0.0024136702995747328, + 0.0010207948507741094, + 0.0029914129991084337, + 0.0033030291087925434, + 0.0022108862176537514, + 0.00148750189691782, + 0.021311337128281593, + 0.004435788840055466, + 0.0026902116369456053, + 0.002192259067669511, + 0.003016658592969179, + 0.02078877203166485, + 0.07959479838609695, + 0.019249051809310913, + 0.08642566949129105, + 0.019179528579115868, + 0.007572033442556858, + 0.013699059374630451, + 0.01320776715874672, + 0.03148600831627846, + 0.004992568399757147, + 0.004260671325027943, + 0.003797746729105711, + 0.005685419775545597, + 0.006533145438879728, + 0.009425661526620388, + 0.007729902397841215, + 0.006376262288540602, + 0.004166043363511562, + 0.012284774333238602, + 0.004587431903928518, + 0.010306961834430695, + 0.019225621595978737, + 0.005954890977591276, + 0.014835757203400135, + 0.013185604475438595, + 0.02120192162692547, + 0.11204766482114792, + 0.02126353047788143, + 0.02885800041258335, + 0.09218861162662506, + 0.025566961616277695, + 0.02196224220097065, + 0.010683630593121052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.009157324209809303 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608314923942089 + ], + "opposite_logprob_tokens": [ + -4.141083240509033 + ], + "logprob_reward": -0.01608314923942089, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608314923942089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18938982486724854, + 0.11332306265830994, + 0.04388739913702011, + 0.05856642127037048, + 0.124635249376297, + 0.07732897996902466, + 0.08586051315069199, + 0.06321679800748825, + 0.13536831736564636, + 0.10842350125312805 + ], + "credit_map": [ + 0.026715341955423355, + 0.013906425796449184, + 0.02251466177403927, + 0.03254769369959831, + 0.03131600469350815, + 0.010283608920872211, + 0.03248324990272522, + 0.019622810184955597, + 0.012040446512401104, + 0.0049875518307089806, + 0.023704728111624718, + 0.014135253615677357, + 0.010889289900660515, + 0.006911254487931728, + 0.013273747637867928, + 0.027380771934986115, + 0.00676333112642169, + 0.002306179841980338, + 0.006539218593388796, + 0.007411160506308079, + 0.0077401818707585335, + 0.002221036935225129, + 0.0054213134571909904, + 0.0054849665611982346, + 0.003530614310875535, + 0.002554412465542555, + 0.015334864147007465, + 0.006766154430806637, + 0.003516184166073799, + 0.0032802470959722996, + 0.004898513201624155, + 0.01868542470037937, + 0.012828318402171135, + 0.007406068965792656, + 0.026716526597738266, + 0.01732971891760826, + 0.009617576375603676, + 0.009599022567272186, + 0.018418822437524796, + 0.02271917276084423, + 0.003207453526556492, + 0.0018377798842266202, + 0.028637079522013664, + 0.00583100039511919, + 0.002422954188659787, + 0.0031896731816232204, + 0.0038900321815162897, + 0.028312992304563522, + 0.020234601572155952, + 0.006965447682887316, + 0.01972397044301033, + 0.010562256909906864, + 0.004941809456795454, + 0.006513425149023533, + 0.008646580390632153, + 0.00827240664511919, + 0.006431742571294308, + 0.006229592487215996, + 0.00544451829046011, + 0.006800508126616478, + 0.007350705098360777, + 0.010345932096242905, + 0.011042380705475807, + 0.009571406990289688, + 0.007266095839440823, + 0.01825619302690029, + 0.011526420712471008, + 0.01867818273603916, + 0.024486752226948738, + 0.011261132545769215, + 0.02262655273079872, + 0.02126695029437542, + 0.008541521616280079, + 0.03517404571175575, + 0.0049094632267951965, + 0.011035528965294361, + 0.024592608213424683, + 0.010216076858341694, + 0.007357473950833082, + 0.006596764549612999 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.01608314923942089 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005958057940006256 + ], + "opposite_logprob_tokens": [ + -5.130958080291748 + ], + "logprob_reward": -0.005958057940006256, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958057940006256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17742431163787842, + 0.06051788106560707, + 0.03075464628636837, + 0.19701623916625977, + 0.04621803015470505, + 0.04025638476014137, + 0.1098996102809906, + 0.0674993097782135, + 0.08037835359573364, + 0.19003531336784363 + ], + "credit_map": [ + 0.013968218117952347, + 0.008023709990084171, + 0.024533133953809738, + 0.029613159596920013, + 0.027390556409955025, + 0.008068662136793137, + 0.0389397032558918, + 0.026887141168117523, + 0.00506489397957921, + 0.0017975236987695098, + 0.014477661810815334, + 0.007671587634831667, + 0.006695414427667856, + 0.0024907980114221573, + 0.005764739587903023, + 0.01655525527894497, + 0.004413491580635309, + 0.0017279185121878982, + 0.004599807783961296, + 0.004442207049578428, + 0.0042777834460139275, + 0.0018766388529911637, + 0.005215559154748917, + 0.004201238043606281, + 0.009636227041482925, + 0.004862661473453045, + 0.05929001793265343, + 0.015489505603909492, + 0.006095543038100004, + 0.008618599735200405, + 0.019380079582333565, + 0.07364359498023987, + 0.0041733914986252785, + 0.004327518865466118, + 0.0064935642294585705, + 0.006842724047601223, + 0.004585193935781717, + 0.0021789520978927612, + 0.008758985437452793, + 0.008857696317136288, + 0.002596384845674038, + 0.00198775134049356, + 0.011400051414966583, + 0.004738032352179289, + 0.0031604680698364973, + 0.0017241219757124782, + 0.004962534178048372, + 0.009687035344541073, + 0.035699520260095596, + 0.009538265876471996, + 0.03004877269268036, + 0.008023499511182308, + 0.0036626504734158516, + 0.006351745221763849, + 0.006659096106886864, + 0.009916058741509914, + 0.0041747866198420525, + 0.005436246283352375, + 0.006361451465636492, + 0.010099350474774837, + 0.008818385191261768, + 0.0115983160212636, + 0.011889738962054253, + 0.009121035225689411, + 0.0051535931415855885, + 0.007328491657972336, + 0.0052100783213973045, + 0.009807721711695194, + 0.015634099021553993, + 0.005735027138143778, + 0.014049280434846878, + 0.01746004819869995, + 0.01114576868712902, + 0.05815773457288742, + 0.014032005332410336, + 0.01673327200114727, + 0.05227459967136383, + 0.01493105199187994, + 0.014241307973861694, + 0.008519557304680347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.005958057940006256 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029797280207276344 + ], + "opposite_logprob_tokens": [ + -3.529797315597534 + ], + "logprob_reward": -0.029797280207276344, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029797280207276344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035042062401771545, + 0.04486813396215439, + 0.04875851795077324, + 0.08570902049541473, + 0.05626458302140236, + 0.11449946463108063, + 0.23986363410949707, + 0.1262962818145752, + 0.13062261044979095, + 0.11807572096586227 + ], + "credit_map": [ + 0.003382590366527438, + 0.0015745841665193439, + 0.004355175886303186, + 0.006588836666196585, + 0.005711382254958153, + 0.0019086510874330997, + 0.0069447592832148075, + 0.004576077219098806, + 0.004734390415251255, + 0.002861802699044347, + 0.007012722548097372, + 0.007754925172775984, + 0.005467101000249386, + 0.003070288570597768, + 0.006684155203402042, + 0.007282746955752373, + 0.010903443209826946, + 0.003070971928536892, + 0.00864357128739357, + 0.00631594518199563, + 0.004145809914916754, + 0.0023889534641057253, + 0.005749127361923456, + 0.007540690712630749, + 0.004543992690742016, + 0.002793490421026945, + 0.02394327148795128, + 0.0071294656954705715, + 0.003478163154795766, + 0.0036291158758103848, + 0.010526428930461407, + 0.029665082693099976, + 0.00492115318775177, + 0.0040500289760529995, + 0.005936488974839449, + 0.012911257334053516, + 0.005236239172518253, + 0.0024181695189327, + 0.010575971566140652, + 0.010215267539024353, + 0.004952565301209688, + 0.0032324676867574453, + 0.040596432983875275, + 0.008936299942433834, + 0.0047250473871827126, + 0.004640909377485514, + 0.008807878009974957, + 0.038607850670814514, + 0.06328383088111877, + 0.014628010801970959, + 0.07522176206111908, + 0.019436847418546677, + 0.008368956856429577, + 0.01386935729533434, + 0.013346020132303238, + 0.03170883283019066, + 0.01930076628923416, + 0.010258950293064117, + 0.010054852813482285, + 0.012790937907993793, + 0.012840435840189457, + 0.022478457540273666, + 0.017991913482546806, + 0.020579952746629715, + 0.009523730725049973, + 0.01093886699527502, + 0.011394884437322617, + 0.014538062736392021, + 0.020295782014727592, + 0.009876085445284843, + 0.017902424558997154, + 0.03615276888012886, + 0.010779584757983685, + 0.039285510778427124, + 0.006117764860391617, + 0.011829869821667671, + 0.025702480226755142, + 0.007801600731909275, + 0.010648122988641262, + 0.005910785403102636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.029797280207276344 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057681791484356 + ], + "opposite_logprob_tokens": [ + -3.8955767154693604 + ], + "logprob_reward": -0.02057681791484356, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057681791484356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0316249243915081, + 0.054309435188770294, + 0.013842186890542507, + 0.10224231332540512, + 0.01609884947538376, + 0.044234566390514374, + 0.2799314558506012, + 0.08626428246498108, + 0.10586326569318771, + 0.26558876037597656 + ], + "credit_map": [ + 0.004276171326637268, + 0.0017481952672824264, + 0.0038487976416945457, + 0.00700695114210248, + 0.003496781690046191, + 0.0017011781455948949, + 0.00502440519630909, + 0.004522441886365414, + 0.004008352756500244, + 0.001627905759960413, + 0.014632103033363819, + 0.005162214860320091, + 0.004791141953319311, + 0.0024080700241029263, + 0.004847745411098003, + 0.0168319009244442, + 0.001160023151896894, + 0.0010618617525324225, + 0.0015201527858152986, + 0.002863823203369975, + 0.0031189427245408297, + 0.0006109815440140665, + 0.0014404545072466135, + 0.002065947512164712, + 0.004857596009969711, + 0.0026889259461313486, + 0.03004560060799122, + 0.009708949364721775, + 0.003719233674928546, + 0.004920854698866606, + 0.008181512355804443, + 0.038119640201330185, + 0.0021724929101765156, + 0.0016260611591860652, + 0.0018884757300838828, + 0.0037458627484738827, + 0.0012610654812306166, + 0.0007275839452631772, + 0.0020462372340261936, + 0.0026310698594897985, + 0.001903518452309072, + 0.001069971825927496, + 0.016540059819817543, + 0.00324011011980474, + 0.001293080858886242, + 0.0017909450689330697, + 0.002420037752017379, + 0.01597684621810913, + 0.08169709891080856, + 0.018634960055351257, + 0.08860277384519577, + 0.020497819408774376, + 0.008012747392058372, + 0.014638240449130535, + 0.015811948105692863, + 0.03203586861491203, + 0.015279639512300491, + 0.008461044169962406, + 0.006991560570895672, + 0.009300551377236843, + 0.010153991170227528, + 0.009399151429533958, + 0.013551321811974049, + 0.01312701590359211, + 0.006708624307066202, + 0.010679478757083416, + 0.015105843544006348, + 0.015662897378206253, + 0.019716434180736542, + 0.005990828387439251, + 0.016810482367873192, + 0.01518868189305067, + 0.017495345324277878, + 0.07100129872560501, + 0.02363341487944126, + 0.021249089390039444, + 0.07940899580717087, + 0.021880852058529854, + 0.020598677918314934, + 0.010321062058210373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.02057681791484356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, newsreel-style shot, starting at eye level and slowly zooming in, focuses on an elderly eunuch, the chief eunuch of the Qing Dynasty, standing alone by a carved wooden window in a quiet study in the Forbidden City. Soft afternoon sunlight slanted through the thin paper covering the window pane, skewing the air with dancing dust and illuminating the delicate texture of his dark blue satin uniform, a patterned pattern. He was about sixty years old, with a shiny forehead and a head of graying hair that hung down to his waist. His pale, finely lined face was expressionless, only the dark green jade ring on his right thumb gleamed in the warm light. Carefully, his hands held a bright yellow memorial, his slender fingers moving steadily and deliberately along the neatly written calligraphy of the imperial court. As he reached a certain point, he gently lifted a page with his fingertips, his movements so gentle that they made almost no sound. At this moment, the camera has moved to a close-up of his face, clearly capturing a complex expression that flashed across his usually calm face: his brows frowned very slightly, his eyes withdrew from his concentrated reading, and briefly looked into the void, with a mixture of surprise and deep calculation in his pupils, and then the corners of his mouth sank imperceptibly, but all this lasted for less than a second, and his expression quickly returned to his usual, unfathomable calm. The whole process was completed in extreme silence, creating a tense atmosphere that contained a huge amount of information but was kept secret.", + "seed": 27729, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_0/index_3393_seed_27729_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04256967194378376, + "vlm_logprob": -0.04256967194378376, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.04256967194378376, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016553910449147224 + ], + "opposite_logprob_tokens": [ + -5.01655387878418 + ], + "logprob_reward": -0.016553910449147224, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.016553910449147224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1273098587989807, + 0.07072906196117401, + 0.09131569415330887, + 0.12907972931861877, + 0.09659851342439651, + 0.03325268253684044, + 0.12071650475263596, + 0.07505245506763458, + 0.13499821722507477, + 0.12094727158546448 + ], + "credit_map": [ + 0.019066035747528076, + 0.017057053744792938, + 0.01821187697350979, + 0.016702162101864815, + 0.011514846235513687, + 0.014180940575897694, + 0.01564844325184822, + 0.01492849551141262, + 0.010169701650738716, + 0.01405580434948206, + 0.010835066437721252, + 0.005853826180100441, + 0.005770536605268717, + 0.007411182392388582, + 0.008587596006691456, + 0.008045349270105362, + 0.008339213207364082, + 0.008994875475764275, + 0.013272109441459179, + 0.017130712047219276, + 0.010921487584710121, + 0.007596684619784355, + 0.01031923945993185, + 0.014741378836333752, + 0.018818248063325882, + 0.01664356142282486, + 0.01868818886578083, + 0.01869644597172737, + 0.019340315833687782, + 0.013604581356048584, + 0.01613699086010456, + 0.007151395548135042, + 0.00962620135396719, + 0.015608007088303566, + 0.010799515061080456, + 0.007895262911915779, + 0.022887762635946274, + 0.010703070089221, + 0.008042942732572556, + 0.011035752482712269, + 0.0037092308048158884, + 0.0057937102392315865, + 0.005835136864334345, + 0.00374068901874125, + 0.0029064512345939875, + 0.004221587907522917, + 0.003766304813325405, + 0.003279573516920209, + 0.016257818788290024, + 0.020030124112963676, + 0.016746332868933678, + 0.019272956997156143, + 0.01064546499401331, + 0.010809230618178844, + 0.012068454176187515, + 0.014886117540299892, + 0.008041911758482456, + 0.013128804042935371, + 0.011150689795613289, + 0.012179332785308361, + 0.006410384550690651, + 0.00617119250819087, + 0.009529929608106613, + 0.00844020675867796, + 0.012689564377069473, + 0.0177493654191494, + 0.015954570844769478, + 0.018556224182248116, + 0.02248206175863743, + 0.018686432391405106, + 0.015161572955548763, + 0.0137184364721179, + 0.012813533656299114, + 0.02570153772830963, + 0.008378544822335243, + 0.018369754776358604, + 0.013651343993842602, + 0.01034967415034771, + 0.009550419636070728, + 0.02213246561586857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.016553910449147224 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295800656080246 + ], + "opposite_logprob_tokens": [ + -2.2379579544067383 + ], + "logprob_reward": -0.11295800656080246, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295800656080246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18142038583755493, + 0.031065965071320534, + 0.060686785727739334, + 0.057985708117485046, + 0.11113976687192917, + 0.1340557485818863, + 0.08743014931678772, + 0.138752743601799, + 0.09678853303194046, + 0.10067419707775116 + ], + "credit_map": [ + 0.014956340193748474, + 0.018620967864990234, + 0.02959945611655712, + 0.024757597595453262, + 0.017151249572634697, + 0.018359780311584473, + 0.032959841191768646, + 0.025015171617269516, + 0.005198769271373749, + 0.0038656576070934534, + 0.005302185658365488, + 0.003353826003149152, + 0.002522587077692151, + 0.002124036429449916, + 0.004799949005246162, + 0.003898957744240761, + 0.0065007139928638935, + 0.0053308820351958275, + 0.009537548758089542, + 0.009262355044484138, + 0.010251929983496666, + 0.005362493451684713, + 0.009080187417566776, + 0.005360680166631937, + 0.005223792977631092, + 0.00670338561758399, + 0.008345494046807289, + 0.008920826017856598, + 0.007042689248919487, + 0.006413192953914404, + 0.006811942905187607, + 0.008524385280907154, + 0.009999490343034267, + 0.00686135608702898, + 0.0165071040391922, + 0.008511023595929146, + 0.01752186007797718, + 0.038731057196855545, + 0.006779063958674669, + 0.006228816695511341, + 0.01827288046479225, + 0.015269216150045395, + 0.018651869148015976, + 0.014437234029173851, + 0.013381254859268665, + 0.022098086774349213, + 0.012255087494850159, + 0.0196901336312294, + 0.01382432784885168, + 0.010610569268465042, + 0.010568104684352875, + 0.007441259454935789, + 0.012091567739844322, + 0.011752947233617306, + 0.008216552436351776, + 0.012924829497933388, + 0.011804888024926186, + 0.018625492230057716, + 0.017465801909565926, + 0.017163561657071114, + 0.015790769830346107, + 0.0162812490016222, + 0.014047489501535892, + 0.02757350541651249, + 0.008182540535926819, + 0.01985946297645569, + 0.009238139726221561, + 0.013830814510583878, + 0.011777142994105816, + 0.013624012470245361, + 0.009996018372476101, + 0.010280407033860683, + 0.01287543959915638, + 0.012779117561876774, + 0.021837083622813225, + 0.01559515856206417, + 0.007447876036167145, + 0.00928920041769743, + 0.007848217152059078, + 0.013002109713852406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.11295800656080246 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033711355179548264 + ], + "opposite_logprob_tokens": [ + -3.4087114334106445 + ], + "logprob_reward": -0.033711355179548264, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033711355179548264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17701758444309235, + 0.1659127026796341, + 0.08627122640609741, + 0.03942280262708664, + 0.18338614702224731, + 0.04146314784884453, + 0.07534513622522354, + 0.04564354941248894, + 0.09201040118932724, + 0.09352727979421616 + ], + "credit_map": [ + 0.013665305450558662, + 0.020358042791485786, + 0.02676459774374962, + 0.019302116706967354, + 0.01707419753074646, + 0.021775688976049423, + 0.030513087287545204, + 0.02756454050540924, + 0.017645590007305145, + 0.020753001794219017, + 0.027673745527863503, + 0.020083408802747726, + 0.013414256274700165, + 0.013963088393211365, + 0.022964127361774445, + 0.02941548079252243, + 0.007122029550373554, + 0.00698372395709157, + 0.013558945618569851, + 0.01638134755194187, + 0.011249520815908909, + 0.007425337098538876, + 0.011687290854752064, + 0.011863026767969131, + 0.004572175443172455, + 0.003768157446756959, + 0.005123320966959, + 0.006313689053058624, + 0.005432683974504471, + 0.004152811132371426, + 0.004553997423499823, + 0.005505967885255814, + 0.020870396867394447, + 0.02222546376287937, + 0.025440528988838196, + 0.016364356502890587, + 0.022314833477139473, + 0.031541094183921814, + 0.016392027959227562, + 0.02823745645582676, + 0.004939781967550516, + 0.008409827016294003, + 0.006269550416618586, + 0.004784265533089638, + 0.004898860119283199, + 0.0043442449532449245, + 0.004456310998648405, + 0.0033603094052523375, + 0.006624705158174038, + 0.014626300893723965, + 0.0110669806599617, + 0.009464164264500141, + 0.007394915912300348, + 0.005693581886589527, + 0.010710165835916996, + 0.009764325805008411, + 0.004034148994833231, + 0.00888561550527811, + 0.007177503779530525, + 0.006393881514668465, + 0.003927024081349373, + 0.004103195387870073, + 0.004870747681707144, + 0.006251432001590729, + 0.009791804477572441, + 0.010872253216803074, + 0.009557791985571384, + 0.00927768088877201, + 0.010163189843297005, + 0.019383659586310387, + 0.007049831096082926, + 0.015914198011159897, + 0.009379944764077663, + 0.015706010162830353, + 0.01089007593691349, + 0.009235730394721031, + 0.00941407959908247, + 0.010872830636799335, + 0.007577033713459969, + 0.020451581105589867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.033711355179548264 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023285627365112305 + ], + "opposite_logprob_tokens": [ + -3.7732856273651123 + ], + "logprob_reward": -0.023285627365112305, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023285627365112305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1219092383980751, + 0.05907133221626282, + 0.03769156336784363, + 0.04411861672997475, + 0.23170970380306244, + 0.039116889238357544, + 0.08102025836706161, + 0.06907941401004791, + 0.1864091008901596, + 0.12987393140792847 + ], + "credit_map": [ + 0.015561167150735855, + 0.02018781006336212, + 0.02386900968849659, + 0.014150283299386501, + 0.010389653034508228, + 0.007073021028190851, + 0.01776624657213688, + 0.012912037782371044, + 0.006829971447587013, + 0.011505644768476486, + 0.008782078512012959, + 0.006706762593239546, + 0.0065130386501550674, + 0.0043732901103794575, + 0.00803933572024107, + 0.006321212276816368, + 0.0038885115645825863, + 0.0030299569480121136, + 0.00644410727545619, + 0.008406943641602993, + 0.002913683420047164, + 0.003080798778682947, + 0.004770312923938036, + 0.00515724765136838, + 0.006620475556701422, + 0.005773491691797972, + 0.005629390478134155, + 0.005971935577690601, + 0.005679931957274675, + 0.004435481503605843, + 0.006287642288953066, + 0.0037202658131718636, + 0.024719471111893654, + 0.018555957823991776, + 0.04129252955317497, + 0.019450174644589424, + 0.024848753586411476, + 0.0718151330947876, + 0.01655510440468788, + 0.01447256002575159, + 0.004860213957726955, + 0.004681324120610952, + 0.004879796877503395, + 0.004277150612324476, + 0.003782670246437192, + 0.005585229489952326, + 0.0058568320237100124, + 0.00519367353990674, + 0.01196579821407795, + 0.012368624098598957, + 0.010250238701701164, + 0.009102603420615196, + 0.007986360229551792, + 0.005467691458761692, + 0.0070248814299702644, + 0.016854064539074898, + 0.008055967278778553, + 0.015866633504629135, + 0.00943903811275959, + 0.008189888671040535, + 0.006707949563860893, + 0.006362224463373423, + 0.004709883127361536, + 0.009747826494276524, + 0.023163622245192528, + 0.03005063533782959, + 0.020281702280044556, + 0.02860279008746147, + 0.019443605095148087, + 0.02487361803650856, + 0.016629591584205627, + 0.023363543674349785, + 0.015539758838713169, + 0.02326771803200245, + 0.018686363473534584, + 0.02268320694565773, + 0.012888828292489052, + 0.010965784080326557, + 0.010571799241006374, + 0.015270470641553402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.023285627365112305 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026339460164308548 + ], + "opposite_logprob_tokens": [ + -3.651339530944824 + ], + "logprob_reward": -0.026339460164308548, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026339460164308548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1154472604393959, + 0.057168349623680115, + 0.12335449457168579, + 0.05095992982387543, + 0.14924590289592743, + 0.08083642274141312, + 0.05102790892124176, + 0.07301633805036545, + 0.12396480143070221, + 0.17497855424880981 + ], + "credit_map": [ + 0.009590922854840755, + 0.015588297508656979, + 0.01927516981959343, + 0.01612986996769905, + 0.011263295076787472, + 0.010130909271538258, + 0.019237563014030457, + 0.014231227338314056, + 0.006829777732491493, + 0.007027799729257822, + 0.009377931244671345, + 0.008303697220981121, + 0.004167965613305569, + 0.006035179365426302, + 0.007648729719221592, + 0.007777267135679722, + 0.012409997172653675, + 0.010657298378646374, + 0.018971256911754608, + 0.025324564427137375, + 0.01581612601876259, + 0.008062479086220264, + 0.016772495582699776, + 0.015340275131165981, + 0.005456339102238417, + 0.004623063839972019, + 0.007530247792601585, + 0.008334635756909847, + 0.006371627561748028, + 0.004653593525290489, + 0.006518894340842962, + 0.007471525110304356, + 0.014892159961163998, + 0.014119135215878487, + 0.022021101787686348, + 0.013694911263883114, + 0.018804974853992462, + 0.04206108674407005, + 0.012875040993094444, + 0.010777500458061695, + 0.010064766742289066, + 0.014283781871199608, + 0.012675989419221878, + 0.010118185542523861, + 0.00838651042431593, + 0.009057633578777313, + 0.01009648572653532, + 0.00615306943655014, + 0.006398519966751337, + 0.00865953229367733, + 0.007550898008048534, + 0.009001764468848705, + 0.005828803405165672, + 0.004313900135457516, + 0.0054881456308066845, + 0.003786345012485981, + 0.00883214920759201, + 0.015431275591254234, + 0.008998243138194084, + 0.01074996404349804, + 0.006945476401597261, + 0.006979470606893301, + 0.0059648375026881695, + 0.009114920161664486, + 0.009359443560242653, + 0.03469233959913254, + 0.014528507366776466, + 0.018469445407390594, + 0.01259307749569416, + 0.011225594207644463, + 0.011652658693492413, + 0.011443731375038624, + 0.01804434135556221, + 0.03318566828966141, + 0.01947163976728916, + 0.02377442643046379, + 0.019398430362343788, + 0.016822991892695427, + 0.014031147584319115, + 0.030249906703829765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.026339460164308548 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, newsreel-style shot, starting at eye level and slowly zooming in, focuses on an elderly eunuch, the chief eunuch of the Qing Dynasty, standing alone by a carved wooden window in a quiet study in the Forbidden City. Soft afternoon sunlight slanted through the thin paper covering the window pane, skewing the air with dancing dust and illuminating the delicate texture of his dark blue satin uniform, a patterned pattern. He was about sixty years old, with a shiny forehead and a head of graying hair that hung down to his waist. His pale, finely lined face was expressionless, only the dark green jade ring on his right thumb gleamed in the warm light. Carefully, his hands held a bright yellow memorial, his slender fingers moving steadily and deliberately along the neatly written calligraphy of the imperial court. As he reached a certain point, he gently lifted a page with his fingertips, his movements so gentle that they made almost no sound. At this moment, the camera has moved to a close-up of his face, clearly capturing a complex expression that flashed across his usually calm face: his brows frowned very slightly, his eyes withdrew from his concentrated reading, and briefly looked into the void, with a mixture of surprise and deep calculation in his pupils, and then the corners of his mouth sank imperceptibly, but all this lasted for less than a second, and his expression quickly returned to his usual, unfathomable calm. The whole process was completed in extreme silence, creating a tense atmosphere that contained a huge amount of information but was kept secret.", + "seed": 27730, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_0/index_3393_seed_27730_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03029817622154951, + "vlm_logprob": -0.03029817622154951, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.03029817622154951, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011757600121200085 + ], + "opposite_logprob_tokens": [ + -5.5117573738098145 + ], + "logprob_reward": -0.011757600121200085, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.011757600121200085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11943069100379944, + 0.0482197068631649, + 0.08712617307901382, + 0.1197623759508133, + 0.02590658888220787, + 0.06516300141811371, + 0.2275552600622177, + 0.11941301822662354, + 0.051750075072050095, + 0.1356731355190277 + ], + "credit_map": [ + 0.013285694643855095, + 0.024080026894807816, + 0.013215422630310059, + 0.017360009253025055, + 0.014146599918603897, + 0.009244604036211967, + 0.01843848079442978, + 0.009659857489168644, + 0.009622088633477688, + 0.011482386849820614, + 0.005464840214699507, + 0.007250683382153511, + 0.004785072989761829, + 0.0019867660012096167, + 0.005105649121105671, + 0.0025222175754606724, + 0.007612067274749279, + 0.008359583094716072, + 0.011528166010975838, + 0.008944802917540073, + 0.02163282036781311, + 0.006486393511295319, + 0.007237601559609175, + 0.01532473973929882, + 0.03154181316494942, + 0.014436897821724415, + 0.009577123448252678, + 0.010843374766409397, + 0.01835954748094082, + 0.015420770272612572, + 0.006539529655128717, + 0.013043321669101715, + 0.002529869321733713, + 0.003693970153108239, + 0.004421364516019821, + 0.0028939400799572468, + 0.003594831097871065, + 0.0031422656029462814, + 0.002427914412692189, + 0.0032024311367422342, + 0.006073308177292347, + 0.006080500315874815, + 0.02135302871465683, + 0.004917386919260025, + 0.005756980273872614, + 0.0032053536269813776, + 0.013592490926384926, + 0.004183951765298843, + 0.032029252499341965, + 0.017517264932394028, + 0.08106868714094162, + 0.013980508781969547, + 0.011023042723536491, + 0.016728103160858154, + 0.02826649881899357, + 0.026941897347569466, + 0.01759214513003826, + 0.011762858368456364, + 0.015025476925075054, + 0.011686408892273903, + 0.01819271594285965, + 0.018676722422242165, + 0.0125888641923666, + 0.013887823559343815, + 0.008904075250029564, + 0.005655086599290371, + 0.009098246693611145, + 0.004539589863270521, + 0.0069916220381855965, + 0.00517320865765214, + 0.006071810144931078, + 0.005316433496773243, + 0.023032769560813904, + 0.020890207961201668, + 0.020739663392305374, + 0.008843495510518551, + 0.024458540603518486, + 0.017664315178990364, + 0.010829525999724865, + 0.009214624762535095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 3 + ], + "selected_reward": -0.011757600121200085 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813115134835243 + ], + "opposite_logprob_tokens": [ + -3.2881312370300293 + ], + "logprob_reward": -0.03813115134835243, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813115134835243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47522881627082825, + 0.039426278322935104, + 0.08351978659629822, + 0.034784313291311264, + 0.037840697914361954, + 0.057669203728437424, + 0.06994210928678513, + 0.08050432801246643, + 0.035392407327890396, + 0.08569201081991196 + ], + "credit_map": [ + 0.05378158017992973, + 0.0619840994477272, + 0.050458233803510666, + 0.07579942792654037, + 0.060738950967788696, + 0.04490577057003975, + 0.08831018209457397, + 0.03925059735774994, + 0.005071895197033882, + 0.0072785927914083, + 0.005459306295961142, + 0.005398436915129423, + 0.0059110987931489944, + 0.0031493084970861673, + 0.004654289688915014, + 0.0025033499114215374, + 0.007685997057706118, + 0.006725708954036236, + 0.011840309016406536, + 0.00564919738098979, + 0.03037329763174057, + 0.011245009489357471, + 0.006391393020749092, + 0.0036088768392801285, + 0.004061510786414146, + 0.004614594858139753, + 0.0034245450515300035, + 0.00472564622759819, + 0.005398158449679613, + 0.004065093584358692, + 0.003937458153814077, + 0.0045573110692203045, + 0.004850201308727264, + 0.004762968514114618, + 0.0037476075813174248, + 0.004185602534562349, + 0.008163646794855595, + 0.004476346541196108, + 0.0041448031552135944, + 0.0035095203202217817, + 0.005957118701189756, + 0.006779595743864775, + 0.012485845945775509, + 0.005100643262267113, + 0.008352820761501789, + 0.005750713869929314, + 0.00866295862942934, + 0.004579506814479828, + 0.010996717028319836, + 0.005129518918693066, + 0.025394132360816002, + 0.0065244948491454124, + 0.004358314909040928, + 0.0045979940332472324, + 0.009294120594859123, + 0.003646818920969963, + 0.011046170257031918, + 0.0071174707263708115, + 0.012323345057666302, + 0.007289589382708073, + 0.009516658261418343, + 0.011484319344162941, + 0.010298790410161018, + 0.011427989229559898, + 0.007003831211477518, + 0.0035559963434934616, + 0.0033735265024006367, + 0.0028823385946452618, + 0.003448678180575371, + 0.007337830029428005, + 0.00431930972263217, + 0.0034708958119153976, + 0.01437170896679163, + 0.010914881713688374, + 0.008755182847380638, + 0.0058831083588302135, + 0.01766844652593136, + 0.009791157208383083, + 0.010453028604388237, + 0.00785449892282486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.03813115134835243 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371630981564522 + ], + "opposite_logprob_tokens": [ + -3.4087162017822266 + ], + "logprob_reward": -0.03371630981564522, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371630981564522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11546506732702255, + 0.28250521421432495, + 0.10055344551801682, + 0.18013012409210205, + 0.018821680918335915, + 0.048527464270591736, + 0.0425083227455616, + 0.05043371766805649, + 0.08868929743766785, + 0.07236570119857788 + ], + "credit_map": [ + 0.013264269568026066, + 0.020974185317754745, + 0.010963546112179756, + 0.013667132705450058, + 0.01810126006603241, + 0.011619118973612785, + 0.016504984349012375, + 0.010370561853051186, + 0.04883130267262459, + 0.0748961791396141, + 0.024878820404410362, + 0.04636305198073387, + 0.0287015400826931, + 0.010339606553316116, + 0.03444112837314606, + 0.014053607359528542, + 0.011279406026005745, + 0.007003930862993002, + 0.009311138652265072, + 0.017027296125888824, + 0.01607278734445572, + 0.0100019546225667, + 0.017774229869246483, + 0.012082703411579132, + 0.0369916632771492, + 0.029882987961173058, + 0.025292571634054184, + 0.02533848024904728, + 0.021766725927591324, + 0.0134379081428051, + 0.015406446531414986, + 0.012013338506221771, + 0.0014661778695881367, + 0.0032365715596824884, + 0.002484437543898821, + 0.0023409791756421328, + 0.0025868366938084364, + 0.002525561023503542, + 0.0015777730150148273, + 0.002603345550596714, + 0.005318203009665012, + 0.00510178878903389, + 0.014559068717062473, + 0.0038589686155319214, + 0.004604342859238386, + 0.0026046717539429665, + 0.00996134802699089, + 0.002519072499126196, + 0.004526748321950436, + 0.003986499737948179, + 0.00822411198168993, + 0.0058920192532241344, + 0.004129437264055014, + 0.0033700072672218084, + 0.004960382357239723, + 0.007419119123369455, + 0.007447891402989626, + 0.004756387323141098, + 0.0070497035048902035, + 0.004878050182014704, + 0.005791826173663139, + 0.008702510967850685, + 0.00692168902605772, + 0.00488565769046545, + 0.011229143477976322, + 0.008005620911717415, + 0.008654294535517693, + 0.011911775916814804, + 0.008017291314899921, + 0.01004557404667139, + 0.02156936749815941, + 0.009256227873265743, + 0.013050176203250885, + 0.010109989903867245, + 0.012025914154946804, + 0.0044447253458201885, + 0.009035632014274597, + 0.010018006898462772, + 0.005779414437711239, + 0.007901842705905437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.03371630981564522 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809305280447006 + ], + "opposite_logprob_tokens": [ + -3.288093090057373 + ], + "logprob_reward": -0.03809305280447006, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809305280447006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10258002579212189, + 0.060938648879528046, + 0.06925646960735321, + 0.06576909869909286, + 0.03365931659936905, + 0.04184038192033768, + 0.17266155779361725, + 0.14343507587909698, + 0.22203609347343445, + 0.08782339841127396 + ], + "credit_map": [ + 0.011117946356534958, + 0.018007919192314148, + 0.011602029204368591, + 0.017074033617973328, + 0.01309233345091343, + 0.008968416601419449, + 0.014359884895384312, + 0.008357455022633076, + 0.0073374188505113125, + 0.006029838230460882, + 0.00585173862054944, + 0.015054900199174881, + 0.008664804510772228, + 0.003216097131371498, + 0.010399994440376759, + 0.004383851308375597, + 0.007093712221831083, + 0.006602500565350056, + 0.009210560470819473, + 0.0030773684848099947, + 0.025826796889305115, + 0.009326795116066933, + 0.003673032158985734, + 0.004445692524313927, + 0.0077804275788366795, + 0.013903668150305748, + 0.009263388812541962, + 0.00895437691360712, + 0.009712888859212399, + 0.0050607407465577126, + 0.006745783146470785, + 0.004347821231931448, + 0.0033381273970007896, + 0.005335109308362007, + 0.0035794724244624376, + 0.004815004300326109, + 0.005318265873938799, + 0.004202515818178654, + 0.002849126700311899, + 0.004221691284328699, + 0.003984385635703802, + 0.00405512610450387, + 0.011494779959321022, + 0.0031140251085162163, + 0.005267102736979723, + 0.004084417130798101, + 0.006071080453693867, + 0.0037694626953452826, + 0.026867862790822983, + 0.012230794876813889, + 0.06171189993619919, + 0.016921717673540115, + 0.007406267337501049, + 0.011170127429068089, + 0.021779507398605347, + 0.014573362655937672, + 0.021727437153458595, + 0.01464191172271967, + 0.015216507948935032, + 0.01197348814457655, + 0.028319785371422768, + 0.02257358282804489, + 0.015129560604691505, + 0.013852794654667377, + 0.03099890425801277, + 0.018038434907794, + 0.018323039636015892, + 0.031033867970108986, + 0.017089998349547386, + 0.028781859204173088, + 0.05391822010278702, + 0.02385173738002777, + 0.014037467539310455, + 0.007642597891390324, + 0.0080511923879385, + 0.009399306029081345, + 0.011947101913392544, + 0.016214478760957718, + 0.01035621203482151, + 0.010175038129091263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.03809305280447006 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man wearing a dark blue satin uniform with a pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a quiet study featuring a carved wooden window with thin paper covering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a bright yellow memorial document and moving his fingers along the calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brief facial expression change including a slight frown, looking into the void, and a sinking mouth corner before returning to calm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at eye level, slowly zoom in, and end with a close-up of the character's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029792767018079758 + ], + "opposite_logprob_tokens": [ + -3.5297927856445312 + ], + "logprob_reward": -0.029792767018079758, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029792767018079758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09265990555286407, + 0.047142330557107925, + 0.051214173436164856, + 0.1821415275335312, + 0.08277639001607895, + 0.15100924670696259, + 0.07358162850141525, + 0.1149645745754242, + 0.10887519270181656, + 0.09563509374856949 + ], + "credit_map": [ + 0.010628549382090569, + 0.02294304035604, + 0.011239362880587578, + 0.01071268692612648, + 0.014672991819679737, + 0.005315094254910946, + 0.011275476776063442, + 0.005872698500752449, + 0.00784379430115223, + 0.009830493479967117, + 0.004525967873632908, + 0.0068743787705898285, + 0.007298940327018499, + 0.00329770939424634, + 0.004744077567011118, + 0.002726969076320529, + 0.004786591976881027, + 0.004836495500057936, + 0.006294673308730125, + 0.004650912247598171, + 0.013764917850494385, + 0.008032608777284622, + 0.004852960351854563, + 0.0039950148202478886, + 0.05642446503043175, + 0.02560477890074253, + 0.019712693989276886, + 0.01718246191740036, + 0.021562177687883377, + 0.018458401784300804, + 0.011642186902463436, + 0.011554357595741749, + 0.007649281527847052, + 0.01534966565668583, + 0.010356772691011429, + 0.00767854368314147, + 0.013125407509505749, + 0.007857639342546463, + 0.011183800175786018, + 0.009575283154845238, + 0.018354175612330437, + 0.013728705234825611, + 0.05090782418847084, + 0.010651699267327785, + 0.012735848315060139, + 0.0077295624651014805, + 0.02823008969426155, + 0.008671341463923454, + 0.006748305633664131, + 0.008467089384794235, + 0.013562079519033432, + 0.0119094867259264, + 0.00872271042317152, + 0.0077764177694916725, + 0.009061447344720364, + 0.007334093563258648, + 0.016195686534047127, + 0.01197861135005951, + 0.014956913888454437, + 0.007923229597508907, + 0.022174816578626633, + 0.017423273995518684, + 0.011177393607795238, + 0.013134648092091084, + 0.014458766207098961, + 0.008225208148360252, + 0.01121527235955, + 0.013286431320011616, + 0.0102235721424222, + 0.014559154398739338, + 0.025552894920110703, + 0.011353891342878342, + 0.01612049713730812, + 0.009232722222805023, + 0.01797664165496826, + 0.005646259058266878, + 0.012543766759335995, + 0.01569255441427231, + 0.009768929332494736, + 0.008653722703456879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.029792767018079758 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_1/index_tensor([5958, 5958])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_1/index_tensor([5958, 5958])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b2eae54e75c5b7753f93a28ae2a587898f93c1d7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_1/index_tensor([5958, 5958])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level close-up focuses on a stack of four clear, nested glass bowls on the dark, polished wood surface of a bedroom dresser. Realistic bedroom light, emitted by a warm, invisible lamp on the side, casts soft shadows and creates intricate, shimmering highlights, refracting through the layers of thick glass. A slender female hand, with immaculately manicured pale pink nails, enters the frame, clutching a collection of glittering, small jewels. With meticulous care, the hand gently places its contents into the top bowl: a delicate gold chain necklace, a silver ring set with tiny, sparkling diamonds, and a pair of sparkling pearl earrings. The jewels fall onto the glass with a soft, almost inaudible clink. The hand then smoothly withdraws, leaving the final composition: the three lower bowls remain perfectly stacked, while the upper bowl now supports the glittering jewels, their precious metals and gemstones catching the warm light, creating a tranquil, quiet, and personal atmosphere. A day-ending ritual.", + "seed": 1820, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_1/index_5958_seed_1820_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0197946404106915, + "vlm_logprob": -0.0197946404106915, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.0197946404106915, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.053249601274728775 + ], + "opposite_logprob_tokens": [ + -5.428249835968018 + ], + "logprob_reward": -0.053249601274728775, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.053249601274728775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17619624733924866, + 0.17666471004486084, + 0.07369744777679443, + 0.08355838805437088, + 0.1305675506591797, + 0.12988920509815216, + 0.06746063381433487, + 0.06282424181699753, + 0.07178248465061188, + 0.02735903114080429 + ], + "credit_map": [ + 0.045926012098789215, + 0.007232553791254759, + 0.04367402568459511, + 0.010657900013029575, + 0.008420892991125584, + 0.014901511371135712, + 0.010440465994179249, + 0.034942880272865295, + 0.0260136891156435, + 0.020839516073465347, + 0.028468728065490723, + 0.011511890217661858, + 0.0469098724424839, + 0.0171237550675869, + 0.013947715051472187, + 0.011849543079733849, + 0.012493249960243702, + 0.006202166900038719, + 0.01367778331041336, + 0.004979934077709913, + 0.004860704764723778, + 0.0049495697021484375, + 0.013010911643505096, + 0.013523127883672714, + 0.01220927108079195, + 0.007068160455673933, + 0.013892482034862041, + 0.012409226968884468, + 0.00702275475487113, + 0.0063911317847669125, + 0.013215686194598675, + 0.011349668726325035, + 0.030631456524133682, + 0.009698785841464996, + 0.026188183575868607, + 0.012188956141471863, + 0.009716874919831753, + 0.010134033858776093, + 0.02071334607899189, + 0.011295911855995655, + 0.021453041583299637, + 0.012620851397514343, + 0.018588965758681297, + 0.008755799382925034, + 0.01889028400182724, + 0.0282791368663311, + 0.01097168680280447, + 0.010329450480639935, + 0.013380175456404686, + 0.004779297858476639, + 0.01835077442228794, + 0.0048662712797522545, + 0.006478374358266592, + 0.004807847552001476, + 0.006504721939563751, + 0.008293170481920242, + 0.011385605670511723, + 0.004813548177480698, + 0.010833450593054295, + 0.007900455966591835, + 0.00599669199436903, + 0.007463316898792982, + 0.008732148446142673, + 0.005699023138731718, + 0.016027459874749184, + 0.003918341360986233, + 0.019916921854019165, + 0.00674496591091156, + 0.005303526762872934, + 0.005768917966634035, + 0.00713724410161376, + 0.006965105887502432, + 0.004441100172698498, + 0.0022876334842294455, + 0.005045999772846699, + 0.003563845995813608, + 0.002906921785324812, + 0.003328745486214757, + 0.003270149463787675, + 0.0025146338157355785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.053249601274728775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007653319742530584 + ], + "opposite_logprob_tokens": [ + -4.88265323638916 + ], + "logprob_reward": -0.007653319742530584, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007653319742530584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10394302010536194, + 0.13534843921661377, + 0.09385642409324646, + 0.042677830904722214, + 0.32698124647140503, + 0.08890745043754578, + 0.08118212223052979, + 0.0642862543463707, + 0.04219828173518181, + 0.020618924871087074 + ], + "credit_map": [ + 0.022953243926167488, + 0.005562201142311096, + 0.022048061713576317, + 0.0075048236176371574, + 0.0076806447468698025, + 0.014477386139333248, + 0.011488671414554119, + 0.012228002771735191, + 0.019086159765720367, + 0.015354327857494354, + 0.02089831605553627, + 0.01278381422162056, + 0.031355537474155426, + 0.01177451852709055, + 0.015477325767278671, + 0.008618448860943317, + 0.017930487170815468, + 0.009130963124334812, + 0.017770733684301376, + 0.008793415501713753, + 0.008892088197171688, + 0.009159781038761139, + 0.011717391200363636, + 0.01046157255768776, + 0.008164369501173496, + 0.003989082761108875, + 0.0076340870000422, + 0.0037756471429020166, + 0.004298851825296879, + 0.0038798684254288673, + 0.006702056620270014, + 0.004233872052282095, + 0.06256803125143051, + 0.025870800018310547, + 0.06264995038509369, + 0.03523390367627144, + 0.029038086533546448, + 0.022962870076298714, + 0.057382289320230484, + 0.031275346875190735, + 0.014194718562066555, + 0.011154556646943092, + 0.011755416169762611, + 0.006920007057487965, + 0.012396076694130898, + 0.017679929733276367, + 0.007363315671682358, + 0.0074434420093894005, + 0.02106969617307186, + 0.003745424561202526, + 0.02246321737766266, + 0.005220501217991114, + 0.004456083290278912, + 0.005180043634027243, + 0.00798012688755989, + 0.011067036539316177, + 0.009429202415049076, + 0.005336777772754431, + 0.010044687427580357, + 0.00827784650027752, + 0.006808080244809389, + 0.007563203107565641, + 0.010651457123458385, + 0.006175007205456495, + 0.0078542809933424, + 0.0028340735007077456, + 0.010050928220152855, + 0.0034816490951925516, + 0.004800914321094751, + 0.004953543655574322, + 0.004746806342154741, + 0.0034760914277285337, + 0.0031969996634870768, + 0.0016758551355451345, + 0.0032611764036118984, + 0.00218646670691669, + 0.0026780604384839535, + 0.002503809751942754, + 0.003077136818319559, + 0.0020394225139170885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.007653319742530584 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033734407275915146 + ], + "opposite_logprob_tokens": [ + -3.4087343215942383 + ], + "logprob_reward": -0.033734407275915146, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033734407275915146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0694056823849678, + 0.17103466391563416, + 0.04296567291021347, + 0.038483958691358566, + 0.4780856966972351, + 0.07453017681837082, + 0.049916259944438934, + 0.03226130083203316, + 0.026974637061357498, + 0.01634191907942295 + ], + "credit_map": [ + 0.014024347998201847, + 0.003664359450340271, + 0.01366201788187027, + 0.004188592545688152, + 0.005345637444406748, + 0.005685330368578434, + 0.0063742841593921185, + 0.01646110601723194, + 0.023414500057697296, + 0.018642496317625046, + 0.0315760038793087, + 0.010735159739851952, + 0.048782382160425186, + 0.016871124505996704, + 0.011986438184976578, + 0.009026563726365566, + 0.0072640120051801205, + 0.005651133134961128, + 0.007029229775071144, + 0.0048764958046376705, + 0.004637671168893576, + 0.0034493079874664545, + 0.005872232373803854, + 0.004185591824352741, + 0.0063998200930655, + 0.003916580695658922, + 0.005613964516669512, + 0.00370228779502213, + 0.004984676372259855, + 0.00441440986469388, + 0.005305687431246042, + 0.004146534018218517, + 0.09377404302358627, + 0.036914627999067307, + 0.08943437039852142, + 0.05285811424255371, + 0.040995679795742035, + 0.03067784197628498, + 0.08455710858106613, + 0.0488739050924778, + 0.013321706093847752, + 0.006167531479150057, + 0.012188136577606201, + 0.004183681681752205, + 0.011014948599040508, + 0.014799258671700954, + 0.0071908850222826, + 0.0056640347465872765, + 0.010702542029321194, + 0.003937611356377602, + 0.009046008810400963, + 0.005002544727176428, + 0.0046401480212807655, + 0.0038824090734124184, + 0.008657903410494328, + 0.004047093912959099, + 0.005749392788857222, + 0.0021419492550194263, + 0.0050423056818544865, + 0.003787675639614463, + 0.004063019994646311, + 0.004016410559415817, + 0.004299964290112257, + 0.003160581225529313, + 0.00504582142457366, + 0.0022722557187080383, + 0.005095728673040867, + 0.0033725358080118895, + 0.002218938432633877, + 0.0032118854578584433, + 0.0033160776365548372, + 0.0024413932114839554, + 0.0030139703303575516, + 0.0011767888208851218, + 0.00339271011762321, + 0.001867513870820403, + 0.0017188788624480367, + 0.001857033115811646, + 0.0017448042053729296, + 0.0015702196396887302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.033734407275915146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015135272406041622 + ], + "opposite_logprob_tokens": [ + -6.501513481140137 + ], + "logprob_reward": -0.0015135272406041622, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015135272406041622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14317259192466736, + 0.21061189472675323, + 0.09752144664525986, + 0.03902976214885712, + 0.08489759266376495, + 0.08914145082235336, + 0.18286564946174622, + 0.08138245344161987, + 0.042680028825998306, + 0.028697142377495766 + ], + "credit_map": [ + 0.02958625927567482, + 0.009027169086039066, + 0.025945385918021202, + 0.01136722695082426, + 0.00905678141862154, + 0.016472171992063522, + 0.014237597584724426, + 0.027480004355311394, + 0.03210641071200371, + 0.0234982967376709, + 0.03787161037325859, + 0.012047973461449146, + 0.05866048485040665, + 0.01923326589167118, + 0.01481085829436779, + 0.012382983230054379, + 0.020387357100844383, + 0.007786783389747143, + 0.021293096244335175, + 0.008385018445551395, + 0.008716646581888199, + 0.007714923471212387, + 0.013064704835414886, + 0.010172916576266289, + 0.008680584840476513, + 0.004458877258002758, + 0.0067823054268956184, + 0.004731275606900454, + 0.0030825661960989237, + 0.0036714000161737204, + 0.004133096430450678, + 0.0034896607976406813, + 0.014574292115867138, + 0.007793575059622526, + 0.014399325475096703, + 0.01008842047303915, + 0.009177505038678646, + 0.005491125397384167, + 0.015737297013401985, + 0.0076360455714166164, + 0.016846932470798492, + 0.009761873632669449, + 0.013327538967132568, + 0.005215850193053484, + 0.009434518404304981, + 0.01957004889845848, + 0.006929632741957903, + 0.008055057376623154, + 0.033383216708898544, + 0.015643754974007607, + 0.037315573543310165, + 0.017073122784495354, + 0.02269751764833927, + 0.012608127668499947, + 0.02490626834332943, + 0.019238069653511047, + 0.012591996230185032, + 0.007162679452449083, + 0.015029888600111008, + 0.009456776082515717, + 0.007404374424368143, + 0.010420965030789375, + 0.011376808397471905, + 0.00793896708637476, + 0.008618485182523727, + 0.003148695221170783, + 0.008797245100140572, + 0.004150770604610443, + 0.004220150411128998, + 0.004503404721617699, + 0.004667262081056833, + 0.004574013873934746, + 0.006047213450074196, + 0.002123243175446987, + 0.0061498116701841354, + 0.0028226596768945456, + 0.0023885800037533045, + 0.0030607632361352444, + 0.0033952645026147366, + 0.002709608292207122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.0015135272406041622 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002822346519678831 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.002822346519678831, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002822346519678831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.185635507106781, + 0.08438487350940704, + 0.06380443274974823, + 0.040226470679044724, + 0.08273738622665405, + 0.0609772726893425, + 0.1060953214764595, + 0.16950692236423492, + 0.14028902351856232, + 0.06634282320737839 + ], + "credit_map": [ + 0.05220329761505127, + 0.008845902979373932, + 0.040820490568876266, + 0.011248203925788403, + 0.012623416259884834, + 0.02096269465982914, + 0.014245770871639252, + 0.024685747921466827, + 0.012465668842196465, + 0.007264664862304926, + 0.012855677865445614, + 0.00787246972322464, + 0.0200357548892498, + 0.008464314974844456, + 0.008086643181741238, + 0.00733967637643218, + 0.012466545216739178, + 0.006400669924914837, + 0.013192806392908096, + 0.006808615289628506, + 0.006347028538584709, + 0.004147848580032587, + 0.007086159661412239, + 0.007354758679866791, + 0.006310456898063421, + 0.003981036599725485, + 0.007216739933937788, + 0.005368891172111034, + 0.003571993438526988, + 0.0030243757646530867, + 0.006286391988396645, + 0.004466585349291563, + 0.014674586243927479, + 0.006769970990717411, + 0.016151390969753265, + 0.009419364854693413, + 0.007922720164060593, + 0.005412644241005182, + 0.014331916347146034, + 0.008054790087044239, + 0.007772704586386681, + 0.008653451688587666, + 0.00807051919400692, + 0.007332830224186182, + 0.008525129407644272, + 0.006365117151290178, + 0.009342133067548275, + 0.004915387369692326, + 0.029994575306773186, + 0.006442924030125141, + 0.025635236874222755, + 0.007596103008836508, + 0.00824237521737814, + 0.007712230086326599, + 0.011029151268303394, + 0.009442723356187344, + 0.03617410361766815, + 0.011446698568761349, + 0.041900262236595154, + 0.015365423634648323, + 0.013859868980944157, + 0.017553189769387245, + 0.020010652020573616, + 0.013196725398302078, + 0.028269199654459953, + 0.008349127136170864, + 0.035337116569280624, + 0.014158235862851143, + 0.010833851993083954, + 0.013129586353898048, + 0.01829213835299015, + 0.011919763870537281, + 0.0130304591730237, + 0.00512135773897171, + 0.015058320015668869, + 0.006254661828279495, + 0.005810531787574291, + 0.007856981828808784, + 0.007681908085942268, + 0.005528603680431843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.002822346519678831 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up focuses on a stack of four clear, nested glass bowls on the dark, polished wood surface of a bedroom dresser. Realistic bedroom light, emitted by a warm, invisible lamp on the side, casts soft shadows and creates intricate, shimmering highlights, refracting through the layers of thick glass. A slender female hand, with immaculately manicured pale pink nails, enters the frame, clutching a collection of glittering, small jewels. With meticulous care, the hand gently places its contents into the top bowl: a delicate gold chain necklace, a silver ring set with tiny, sparkling diamonds, and a pair of sparkling pearl earrings. The jewels fall onto the glass with a soft, almost inaudible clink. The hand then smoothly withdraws, leaving the final composition: the three lower bowls remain perfectly stacked, while the upper bowl now supports the glittering jewels, their precious metals and gemstones catching the warm light, creating a tranquil, quiet, and personal atmosphere. A day-ending ritual.", + "seed": 1821, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_1/index_5958_seed_1821_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.018468968151137233, + "vlm_logprob": -0.018468968151137233, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.018468968151137233, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05406498908996582 + ], + "opposite_logprob_tokens": [ + -4.554064750671387 + ], + "logprob_reward": -0.05406498908996582, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.05406498908996582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1884169578552246, + 0.09574376046657562, + 0.09701504558324814, + 0.030832642689347267, + 0.07980753481388092, + 0.06435593962669373, + 0.16318772733211517, + 0.12625209987163544, + 0.07644835114479065, + 0.07794000208377838 + ], + "credit_map": [ + 0.02541785128414631, + 0.05557907372713089, + 0.02495729550719261, + 0.022701120004057884, + 0.020690375939011574, + 0.010987152345478535, + 0.013877658173441887, + 0.014206414110958576, + 0.014006058685481548, + 0.01761571131646633, + 0.013334597460925579, + 0.008045118302106857, + 0.01546974666416645, + 0.007304794155061245, + 0.010021100752055645, + 0.009946629405021667, + 0.0190231055021286, + 0.01010037399828434, + 0.016978314146399498, + 0.0073136575520038605, + 0.016831258311867714, + 0.007544339634478092, + 0.010598660446703434, + 0.008625323884189129, + 0.0055420976132154465, + 0.0035983012057840824, + 0.004547955933958292, + 0.0035344066563993692, + 0.003142348024994135, + 0.0030425691511482, + 0.004928327165544033, + 0.0024966346099972725, + 0.012967548333108425, + 0.015062774531543255, + 0.01467573270201683, + 0.009165524505078793, + 0.009120677597820759, + 0.005804637912660837, + 0.007234100718051195, + 0.005776530131697655, + 0.013409784063696861, + 0.005883920006453991, + 0.012401777319610119, + 0.005776901263743639, + 0.00903906300663948, + 0.006649088114500046, + 0.006816944573074579, + 0.0043784575536847115, + 0.021504418924450874, + 0.017731420695781708, + 0.02241319604218006, + 0.022491006180644035, + 0.024786584079265594, + 0.01931603066623211, + 0.02308376133441925, + 0.011861290782690048, + 0.01707012951374054, + 0.009372519329190254, + 0.017757318913936615, + 0.009831779636442661, + 0.02338375337421894, + 0.025400394573807716, + 0.01036029402166605, + 0.013075893744826317, + 0.008548601530492306, + 0.006480267737060785, + 0.010416967794299126, + 0.008664526976644993, + 0.022208943963050842, + 0.006678084377199411, + 0.008493712171912193, + 0.004957239609211683, + 0.011053092777729034, + 0.005050201900303364, + 0.00999639555811882, + 0.013405594043433666, + 0.01927284710109234, + 0.005891851615160704, + 0.008132202550768852, + 0.0051378183998167515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.05406498908996582 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014229852706193924 + ], + "opposite_logprob_tokens": [ + -4.264229774475098 + ], + "logprob_reward": -0.014229852706193924, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014229852706193924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3417966067790985, + 0.09153664112091064, + 0.0884176567196846, + 0.07894330471754074, + 0.15015868842601776, + 0.07282473891973495, + 0.06284739077091217, + 0.0577906034886837, + 0.03184101730585098, + 0.023843318223953247 + ], + "credit_map": [ + 0.0424439013004303, + 0.11309561878442764, + 0.04108695313334465, + 0.04234091937541962, + 0.043791964650154114, + 0.0209722388535738, + 0.017113525420427322, + 0.020951492711901665, + 0.025050750002264977, + 0.019218822941184044, + 0.012965050525963306, + 0.005832839757204056, + 0.008793010376393795, + 0.005411953199654818, + 0.006294002756476402, + 0.007970214821398258, + 0.01021929457783699, + 0.011936352588236332, + 0.0122382752597332, + 0.008487718179821968, + 0.020589279010891914, + 0.0072659156285226345, + 0.010478675365447998, + 0.007202166598290205, + 0.020193016156554222, + 0.01087369117885828, + 0.01580745168030262, + 0.007411644794046879, + 0.007052106317132711, + 0.0048071336932480335, + 0.006641277112066746, + 0.006156989373266697, + 0.030761295929551125, + 0.02334124781191349, + 0.0301046185195446, + 0.013784999959170818, + 0.018564267084002495, + 0.012290369719266891, + 0.010599607601761818, + 0.010712298564612865, + 0.01661226712167263, + 0.00932198017835617, + 0.01574532687664032, + 0.005268299952149391, + 0.009098094888031483, + 0.00579169113188982, + 0.006003272254019976, + 0.004983813967555761, + 0.010426710359752178, + 0.005313028581440449, + 0.009546511806547642, + 0.007492585573345423, + 0.012468854896724224, + 0.005765776615589857, + 0.007474874146282673, + 0.004359053913503885, + 0.008441473357379436, + 0.004671525210142136, + 0.008591133169829845, + 0.00362595496699214, + 0.012858790345489979, + 0.009291714988648891, + 0.00616419967263937, + 0.004145815037190914, + 0.004930389579385519, + 0.0028596692718565464, + 0.004350942559540272, + 0.0035372686106711626, + 0.0065535143949091434, + 0.0025955941528081894, + 0.004850736353546381, + 0.0021629049442708492, + 0.003399747656658292, + 0.0018259802600368857, + 0.003308890387415886, + 0.0035355661530047655, + 0.00536153232678771, + 0.0017324438085779548, + 0.0030397314112633467, + 0.0016394280828535557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.014229852706193924 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020617691799998283 + ], + "opposite_logprob_tokens": [ + -3.895617723464966 + ], + "logprob_reward": -0.020617691799998283, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020617691799998283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29413744807243347, + 0.07411996275186539, + 0.08708517998456955, + 0.07549642771482468, + 0.055679015815258026, + 0.030235256999731064, + 0.06612908840179443, + 0.10402656346559525, + 0.12241915613412857, + 0.09067192673683167 + ], + "credit_map": [ + 0.037576839327812195, + 0.09045831859111786, + 0.03623855486512184, + 0.036071132868528366, + 0.036920249462127686, + 0.019090812653303146, + 0.01499829813838005, + 0.022783230990171432, + 0.012753147631883621, + 0.013509121723473072, + 0.012153928168118, + 0.008274584077298641, + 0.009002503007650375, + 0.004244142211973667, + 0.006858127191662788, + 0.007324399426579475, + 0.017907412722706795, + 0.00951019674539566, + 0.015058758668601513, + 0.009113076142966747, + 0.011948791332542896, + 0.005849271547049284, + 0.009237445890903473, + 0.008460224606096745, + 0.02007467858493328, + 0.01081147138029337, + 0.014396675862371922, + 0.0062562827952206135, + 0.007793778087943792, + 0.0053202249109745026, + 0.00570600526407361, + 0.005137298256158829, + 0.01102891843765974, + 0.008211104199290276, + 0.01146505493670702, + 0.005578387528657913, + 0.005903471261262894, + 0.00423326063901186, + 0.004518435336649418, + 0.004740378353744745, + 0.0049489885568618774, + 0.005612920969724655, + 0.005948466248810291, + 0.002312374534085393, + 0.0040534986183047295, + 0.002716945018619299, + 0.0022619524970650673, + 0.0023801082279533148, + 0.008852977305650711, + 0.0070369550958275795, + 0.009060031734406948, + 0.008598451502621174, + 0.015104142017662525, + 0.00577120017260313, + 0.007273981813341379, + 0.004431342240422964, + 0.015443924814462662, + 0.008921793662011623, + 0.015131765976548195, + 0.007143034599721432, + 0.020656198263168335, + 0.018343502655625343, + 0.009783039800822735, + 0.008603290654718876, + 0.015665771439671516, + 0.01100441999733448, + 0.018198145553469658, + 0.012160424143075943, + 0.030843263491988182, + 0.011487838812172413, + 0.014323929324746132, + 0.008735345676541328, + 0.01065820176154375, + 0.007079221308231354, + 0.012520551681518555, + 0.00967187900096178, + 0.026299582794308662, + 0.0074888961389660835, + 0.011333364062011242, + 0.005620223470032215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.020617691799998283 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017151420470327139 + ], + "opposite_logprob_tokens": [ + -6.376715183258057 + ], + "logprob_reward": -0.0017151420470327139, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017151420470327139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2588684856891632, + 0.10897983610630035, + 0.0765795111656189, + 0.09406597912311554, + 0.0633118525147438, + 0.047522783279418945, + 0.13259200751781464, + 0.08602987229824066, + 0.07266505062580109, + 0.059384580701589584 + ], + "credit_map": [ + 0.034696660935878754, + 0.08139678090810776, + 0.03395315632224083, + 0.03123430535197258, + 0.032438892871141434, + 0.015489931218326092, + 0.015749294310808182, + 0.013909478671848774, + 0.02102123200893402, + 0.020802121609449387, + 0.014489473775029182, + 0.010410025715827942, + 0.01302132848650217, + 0.008274577558040619, + 0.011186940595507622, + 0.009774129837751389, + 0.012970334850251675, + 0.0093274200335145, + 0.011101212352514267, + 0.00888760481029749, + 0.013646448962390423, + 0.007215004414319992, + 0.008452952839434147, + 0.004978529643267393, + 0.011347224935889244, + 0.013997828587889671, + 0.013613306917250156, + 0.009780383668839931, + 0.02215786837041378, + 0.008733057416975498, + 0.009318476542830467, + 0.005117826163768768, + 0.006788271479308605, + 0.008453331887722015, + 0.010898185893893242, + 0.00576123408973217, + 0.014466977678239346, + 0.0058190044946968555, + 0.006921369582414627, + 0.004203481134027243, + 0.011457405053079128, + 0.00513166980817914, + 0.010570365004241467, + 0.0034218039363622665, + 0.005876141134649515, + 0.0039435457438230515, + 0.003613397479057312, + 0.0035084548871964216, + 0.02074798010289669, + 0.010914023965597153, + 0.021363014355301857, + 0.011754422448575497, + 0.029737228527665138, + 0.012409397400915623, + 0.016056403517723083, + 0.009609535336494446, + 0.013953812420368195, + 0.007256743032485247, + 0.0127479899674654, + 0.009351599961519241, + 0.01773849129676819, + 0.009207389317452908, + 0.009568716399371624, + 0.006205129437148571, + 0.0109651954844594, + 0.006249761208891869, + 0.010722248815000057, + 0.007155468687415123, + 0.015590770170092583, + 0.00579646322876215, + 0.010941123589873314, + 0.005244014319032431, + 0.007514936849474907, + 0.005114234983921051, + 0.008132029324769974, + 0.005818195175379515, + 0.016534196212887764, + 0.004956576973199844, + 0.007470241282135248, + 0.0038441719952970743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0017151420470327139 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017171651124954224 + ], + "opposite_logprob_tokens": [ + -6.3767170906066895 + ], + "logprob_reward": -0.0017171651124954224, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017171651124954224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46331921219825745, + 0.06180373951792717, + 0.07499007135629654, + 0.02748914621770382, + 0.03501119092106819, + 0.04680683836340904, + 0.13889180123806, + 0.07849542796611786, + 0.03116888925433159, + 0.042023759335279465 + ], + "credit_map": [ + 0.061464548110961914, + 0.15177620947360992, + 0.05410139262676239, + 0.06049935892224312, + 0.056966304779052734, + 0.02794879674911499, + 0.02069634012877941, + 0.029866231605410576, + 0.015491818077862263, + 0.011800569482147694, + 0.010453131049871445, + 0.005573551170527935, + 0.0055723837576806545, + 0.0037424727343022823, + 0.004849843215197325, + 0.004319965839385986, + 0.008990349248051643, + 0.006362685933709145, + 0.010090832598507404, + 0.007931286469101906, + 0.020038725808262825, + 0.0062345885671675205, + 0.009657453745603561, + 0.005684145260602236, + 0.0032083941623568535, + 0.003312314860522747, + 0.004774072673171759, + 0.003434370970353484, + 0.004788357298821211, + 0.0023491517640650272, + 0.0034354657400399446, + 0.002187016187235713, + 0.003977845422923565, + 0.0038355544675141573, + 0.005002611316740513, + 0.00518120126798749, + 0.0071130539290606976, + 0.003435717895627022, + 0.00417732447385788, + 0.0022878798190504313, + 0.010539044626057148, + 0.005177980288863182, + 0.007052725180983543, + 0.005129920784384012, + 0.006594640202820301, + 0.003774616401642561, + 0.004483265336602926, + 0.0040546441450715065, + 0.01720987819135189, + 0.013917460106313229, + 0.020020684227347374, + 0.018992634490132332, + 0.029349660500884056, + 0.011924326419830322, + 0.01680057682096958, + 0.010676572099328041, + 0.015482724644243717, + 0.007100608665496111, + 0.011399845592677593, + 0.009977343492209911, + 0.014068624004721642, + 0.008206713944673538, + 0.006439735181629658, + 0.005819831043481827, + 0.005804738495498896, + 0.002966652624309063, + 0.0048461789265275, + 0.0031822812743484974, + 0.004854065366089344, + 0.002511807018890977, + 0.004526629578322172, + 0.0024765327107161283, + 0.005508988630026579, + 0.003349246457219124, + 0.0055989534594118595, + 0.005330149549990892, + 0.010360934771597385, + 0.003899646457284689, + 0.004946761764585972, + 0.0030290738213807344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0017171651124954224 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up focuses on a stack of four clear, nested glass bowls on the dark, polished wood surface of a bedroom dresser. Realistic bedroom light, emitted by a warm, invisible lamp on the side, casts soft shadows and creates intricate, shimmering highlights, refracting through the layers of thick glass. A slender female hand, with immaculately manicured pale pink nails, enters the frame, clutching a collection of glittering, small jewels. With meticulous care, the hand gently places its contents into the top bowl: a delicate gold chain necklace, a silver ring set with tiny, sparkling diamonds, and a pair of sparkling pearl earrings. The jewels fall onto the glass with a soft, almost inaudible clink. The hand then smoothly withdraws, leaving the final composition: the three lower bowls remain perfectly stacked, while the upper bowl now supports the glittering jewels, their precious metals and gemstones catching the warm light, creating a tranquil, quiet, and personal atmosphere. A day-ending ritual.", + "seed": 1822, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_1/index_5958_seed_1822_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.020379633363336324, + "vlm_logprob": -0.020379633363336324, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.020379633363336324, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0584140308201313 + ], + "opposite_logprob_tokens": [ + -3.9334139823913574 + ], + "logprob_reward": -0.0584140308201313, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0584140308201313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21550239622592926, + 0.07386721670627594, + 0.16439774632453918, + 0.0915190801024437, + 0.09230273962020874, + 0.13691335916519165, + 0.10100477188825607, + 0.05606050416827202, + 0.03832234814763069, + 0.030109861865639687 + ], + "credit_map": [ + 0.04277363792061806, + 0.029032263904809952, + 0.03241399675607681, + 0.026769045740365982, + 0.020764218643307686, + 0.016644969582557678, + 0.0322040356695652, + 0.014900232665240765, + 0.01198976393789053, + 0.009106107987463474, + 0.010234163142740726, + 0.012235648930072784, + 0.013164707459509373, + 0.006313112564384937, + 0.006919521372765303, + 0.003904189681634307, + 0.019525699317455292, + 0.019143206998705864, + 0.02106260508298874, + 0.016643986105918884, + 0.044396523386240005, + 0.012563560158014297, + 0.01966044120490551, + 0.011401720345020294, + 0.01357530988752842, + 0.009516487829387188, + 0.012479807250201702, + 0.0180105771869421, + 0.01872965134680271, + 0.007171620149165392, + 0.008618033491075039, + 0.0034175969194620848, + 0.01123642735183239, + 0.014318197034299374, + 0.012001430615782738, + 0.015051295980811119, + 0.010814853943884373, + 0.012988380156457424, + 0.010523926466703415, + 0.005368225742131472, + 0.014905598945915699, + 0.019652778282761574, + 0.016068533062934875, + 0.012480913661420345, + 0.02806662768125534, + 0.017679085955023766, + 0.019944919273257256, + 0.008114896714687347, + 0.022607723250985146, + 0.011276612058281898, + 0.016985954716801643, + 0.011277089826762676, + 0.009320499375462532, + 0.009754819795489311, + 0.011541610583662987, + 0.008240456692874432, + 0.010858566500246525, + 0.0061705815605819225, + 0.008586188778281212, + 0.0064963470213115215, + 0.006455861963331699, + 0.006331317126750946, + 0.006741900462657213, + 0.004419742152094841, + 0.007192115765064955, + 0.004111332353204489, + 0.006261224392801523, + 0.0038782816845923662, + 0.004208515398204327, + 0.004423339385539293, + 0.005694564897567034, + 0.002552973572164774, + 0.009165188297629356, + 0.0034651453606784344, + 0.005871751345694065, + 0.002546016126871109, + 0.0018714385805651546, + 0.002155106980353594, + 0.0032187115866690874, + 0.0018165032379329205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.0584140308201313 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008669476956129074 + ], + "opposite_logprob_tokens": [ + -4.758669376373291 + ], + "logprob_reward": -0.008669476956129074, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008669476956129074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11179322749376297, + 0.10393166542053223, + 0.11789603531360626, + 0.22082965075969696, + 0.09242051839828491, + 0.1329873651266098, + 0.0746232196688652, + 0.04800404608249664, + 0.06898485124111176, + 0.028529442846775055 + ], + "credit_map": [ + 0.028404613956809044, + 0.01278079766780138, + 0.019512934610247612, + 0.008696884848177433, + 0.009401356801390648, + 0.008636037819087505, + 0.016946198418736458, + 0.007414404768496752, + 0.016038574278354645, + 0.011760731227695942, + 0.01311072800308466, + 0.009354114532470703, + 0.028645776212215424, + 0.008727885782718658, + 0.009674208238720894, + 0.006619656924158335, + 0.013536629267036915, + 0.011408409103751183, + 0.015315743163228035, + 0.010595371015369892, + 0.03528861701488495, + 0.011813501827418804, + 0.01328379288315773, + 0.006653958465903997, + 0.021072257310152054, + 0.025610554963350296, + 0.029896564781665802, + 0.02943970449268818, + 0.06303393840789795, + 0.018635384738445282, + 0.023076767101883888, + 0.010064461268484592, + 0.011478409171104431, + 0.012040411122143269, + 0.012258775532245636, + 0.011862856335937977, + 0.017291447147727013, + 0.011891026049852371, + 0.011247543618083, + 0.004350048024207354, + 0.013972403481602669, + 0.015238930471241474, + 0.01936447061598301, + 0.013675627298653126, + 0.026313191279768944, + 0.015606197528541088, + 0.02141047641634941, + 0.00740606477484107, + 0.011684498749673367, + 0.008318915031850338, + 0.011408680118620396, + 0.006973993498831987, + 0.01010427437722683, + 0.01033829152584076, + 0.009059973061084747, + 0.006734589580446482, + 0.012623720802366734, + 0.005967492703348398, + 0.007891533896327019, + 0.0031953665893524885, + 0.004621877335011959, + 0.00460025342181325, + 0.005808267276734114, + 0.0032955326605588198, + 0.018722640350461006, + 0.0065931593999266624, + 0.013381068594753742, + 0.005074659828096628, + 0.005385459400713444, + 0.006776208523660898, + 0.009130985476076603, + 0.003920670133084059, + 0.005137195810675621, + 0.003706237068399787, + 0.004201564472168684, + 0.003390606725588441, + 0.0030242830980569124, + 0.004042439628392458, + 0.002912752563133836, + 0.002114364178851247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.008669476956129074 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636999823153019 + ], + "opposite_logprob_tokens": [ + -3.651370048522949 + ], + "logprob_reward": -0.02636999823153019, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636999823153019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06137782707810402, + 0.13914178311824799, + 0.20553015172481537, + 0.18370883166790009, + 0.17260782420635223, + 0.07838521897792816, + 0.044952910393476486, + 0.02915041707456112, + 0.07282836735248566, + 0.012316559441387653 + ], + "credit_map": [ + 0.01225118339061737, + 0.0075963931158185005, + 0.008643693290650845, + 0.0076779830269515514, + 0.006061607506126165, + 0.007353744003921747, + 0.008023831062018871, + 0.003769393777474761, + 0.02335297502577305, + 0.015375286340713501, + 0.02536160498857498, + 0.016472086310386658, + 0.02863023616373539, + 0.011337467469274998, + 0.011120498180389404, + 0.007491655647754669, + 0.0212410856038332, + 0.026311717927455902, + 0.02451440691947937, + 0.019493816420435905, + 0.05841832980513573, + 0.01690545119345188, + 0.026804327964782715, + 0.01184104010462761, + 0.020613348111510277, + 0.020329680293798447, + 0.02614828757941723, + 0.01734306663274765, + 0.05386856943368912, + 0.01690836250782013, + 0.017223093658685684, + 0.011274444870650768, + 0.020379917696118355, + 0.019154058769345284, + 0.026779819279909134, + 0.017696227878332138, + 0.024294273927807808, + 0.021810956299304962, + 0.030804356560111046, + 0.011688233353197575, + 0.010908659547567368, + 0.008290349505841732, + 0.010392717085778713, + 0.009012443013489246, + 0.01698104664683342, + 0.007431172300130129, + 0.011147696524858475, + 0.004221141804009676, + 0.006811302155256271, + 0.005548623390495777, + 0.007104101590812206, + 0.004895678721368313, + 0.003897071583196521, + 0.0044659716077148914, + 0.00888613611459732, + 0.0033440289553254843, + 0.005266872700303793, + 0.003234978998079896, + 0.004000989254564047, + 0.003494338598102331, + 0.003313971683382988, + 0.004228397738188505, + 0.0033093071542680264, + 0.0023015651386231184, + 0.017969824373722076, + 0.007031054701656103, + 0.014568930491805077, + 0.00481152581050992, + 0.00560042355209589, + 0.0068765138275921345, + 0.011152900755405426, + 0.004817201755940914, + 0.0021604360081255436, + 0.001699752756394446, + 0.0017519639804959297, + 0.001173884724266827, + 0.001545832259580493, + 0.001383202732540667, + 0.0018040250288322568, + 0.000797464104834944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.02636999823153019 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024898024275898933 + ], + "opposite_logprob_tokens": [ + -6.0024895668029785 + ], + "logprob_reward": -0.0024898024275898933, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024898024275898933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11412358283996582, + 0.09447259455919266, + 0.12437743693590164, + 0.18204237520694733, + 0.13948000967502594, + 0.10547132790088654, + 0.139945849776268, + 0.029997749254107475, + 0.04877466335892677, + 0.021314386278390884 + ], + "credit_map": [ + 0.01768200844526291, + 0.015450729057192802, + 0.01830006018280983, + 0.012592989951372147, + 0.013989816419780254, + 0.010664119385182858, + 0.01679071970283985, + 0.008653138764202595, + 0.01544158160686493, + 0.010042108595371246, + 0.014614449813961983, + 0.010566111654043198, + 0.017067138105630875, + 0.00907449796795845, + 0.0103503642603755, + 0.007316343951970339, + 0.014872107654809952, + 0.010384432971477509, + 0.016191711649298668, + 0.010562838055193424, + 0.03432253748178482, + 0.01243705116212368, + 0.01850360631942749, + 0.007103145122528076, + 0.024102313444018364, + 0.019752494990825653, + 0.02334962785243988, + 0.020630991086363792, + 0.03328603878617287, + 0.024459468200802803, + 0.02361777238547802, + 0.012843663804233074, + 0.015353580005466938, + 0.01989043317735195, + 0.017940931022167206, + 0.02377539686858654, + 0.019876589998602867, + 0.01550325844436884, + 0.02015461027622223, + 0.006985214073210955, + 0.011208840645849705, + 0.012725519947707653, + 0.014139263890683651, + 0.01322206761687994, + 0.022158782929182053, + 0.00960627943277359, + 0.01473496574908495, + 0.007675607688724995, + 0.017622333019971848, + 0.02105366438627243, + 0.02072460949420929, + 0.014966334216296673, + 0.015173370018601418, + 0.020558388903737068, + 0.018050549551844597, + 0.011796598322689533, + 0.006217257119715214, + 0.0029970668256282806, + 0.005787948612123728, + 0.002863876987248659, + 0.003052053740248084, + 0.0027215485461056232, + 0.004133850336074829, + 0.0022241477854549885, + 0.011176226660609245, + 0.004830514546483755, + 0.00866229273378849, + 0.003545585321262479, + 0.005333544686436653, + 0.004687379114329815, + 0.007129009813070297, + 0.0034101095516234636, + 0.003653375431895256, + 0.00270812027156353, + 0.002975493436679244, + 0.0023477524518966675, + 0.0026981572154909372, + 0.0029419746715575457, + 0.002286897273734212, + 0.001702614943496883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.0024898024275898933 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954858381301165 + ], + "opposite_logprob_tokens": [ + -5.130954742431641 + ], + "logprob_reward": -0.005954858381301165, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954858381301165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18607693910598755, + 0.11317799240350723, + 0.27164867520332336, + 0.07752303779125214, + 0.06681471318006516, + 0.10000583529472351, + 0.07239587604999542, + 0.042433518916368484, + 0.04473086819052696, + 0.025192562490701675 + ], + "credit_map": [ + 0.05640941858291626, + 0.019313760101795197, + 0.0327899195253849, + 0.018044374883174896, + 0.012838157825171947, + 0.015012240037322044, + 0.019053006544709206, + 0.012616081163287163, + 0.018927961587905884, + 0.013214075937867165, + 0.017122434452176094, + 0.010194787755608559, + 0.02586616575717926, + 0.010542076081037521, + 0.01072571612894535, + 0.0065847719088196754, + 0.0375828817486763, + 0.030588600784540176, + 0.035315338522195816, + 0.02976769395172596, + 0.06942830979824066, + 0.03020140528678894, + 0.02279973216354847, + 0.01596471481025219, + 0.010356705635786057, + 0.009171362034976482, + 0.010967175476253033, + 0.014797499403357506, + 0.01157295797020197, + 0.0058320676907896996, + 0.01060845423489809, + 0.004216822329908609, + 0.008583120070397854, + 0.009007022716104984, + 0.009066459722816944, + 0.007635300047695637, + 0.009469103068113327, + 0.009064093232154846, + 0.009429696016013622, + 0.004559920635074377, + 0.0118784848600626, + 0.014323031529784203, + 0.011025124229490757, + 0.010442908853292465, + 0.01750447414815426, + 0.017092039808630943, + 0.011382541619241238, + 0.006357231177389622, + 0.02014808915555477, + 0.008102759718894958, + 0.01354962307959795, + 0.006645643152296543, + 0.005630254279822111, + 0.006011669989675283, + 0.007051797118037939, + 0.005256051197648048, + 0.012137936428189278, + 0.004614264238625765, + 0.006997561547905207, + 0.003575549926608801, + 0.00411499897018075, + 0.003477205755189061, + 0.005001177545636892, + 0.002514827996492386, + 0.011469602584838867, + 0.004572005942463875, + 0.008475994691252708, + 0.0033647408708930016, + 0.0035026504192501307, + 0.004295194987207651, + 0.006266912445425987, + 0.002783771138638258, + 0.005960153881460428, + 0.0030208807438611984, + 0.004080156330019236, + 0.002451877808198333, + 0.0028789544012397528, + 0.0024025309830904007, + 0.002593038836494088, + 0.0018049700884148479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.005954858381301165 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up focuses on a stack of four clear, nested glass bowls on the dark, polished wood surface of a bedroom dresser. Realistic bedroom light, emitted by a warm, invisible lamp on the side, casts soft shadows and creates intricate, shimmering highlights, refracting through the layers of thick glass. A slender female hand, with immaculately manicured pale pink nails, enters the frame, clutching a collection of glittering, small jewels. With meticulous care, the hand gently places its contents into the top bowl: a delicate gold chain necklace, a silver ring set with tiny, sparkling diamonds, and a pair of sparkling pearl earrings. The jewels fall onto the glass with a soft, almost inaudible clink. The hand then smoothly withdraws, leaving the final composition: the three lower bowls remain perfectly stacked, while the upper bowl now supports the glittering jewels, their precious metals and gemstones catching the warm light, creating a tranquil, quiet, and personal atmosphere. A day-ending ritual.", + "seed": 1823, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_1/index_5958_seed_1823_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.016056817257776856, + "vlm_logprob": -0.016056817257776856, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.016056817257776856, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03743242844939232 + ], + "opposite_logprob_tokens": [ + -4.9124321937561035 + ], + "logprob_reward": -0.03743242844939232, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.03743242844939232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08726039528846741, + 0.061918117105960846, + 0.13041438162326813, + 0.0798538401722908, + 0.055375587195158005, + 0.11637613922357559, + 0.1573098599910736, + 0.22785426676273346, + 0.05969643220305443, + 0.023940958082675934 + ], + "credit_map": [ + 0.015037220902740955, + 0.00980176031589508, + 0.016121739521622658, + 0.012527063488960266, + 0.007486676797270775, + 0.007601703517138958, + 0.008115317672491074, + 0.010568913072347641, + 0.007896463386714458, + 0.008781363256275654, + 0.008203537203371525, + 0.007357735652476549, + 0.012488747015595436, + 0.0064780390821397305, + 0.0052711437456309795, + 0.005441085901111364, + 0.0173324067145586, + 0.017656266689300537, + 0.016580460593104362, + 0.018598036840558052, + 0.025660358369350433, + 0.008986979722976685, + 0.01425202377140522, + 0.011347855441272259, + 0.009742818772792816, + 0.010443402454257011, + 0.011213159188628197, + 0.010722688399255276, + 0.012315639294683933, + 0.009142872877418995, + 0.006222685799002647, + 0.010050573386251926, + 0.007607261184602976, + 0.008150611072778702, + 0.00803682766854763, + 0.008379191160202026, + 0.00596188846975565, + 0.005931721068918705, + 0.004474150948226452, + 0.006833935156464577, + 0.023873422294855118, + 0.01189679279923439, + 0.02057725004851818, + 0.011031434871256351, + 0.006863818969577551, + 0.012942662462592125, + 0.010029226541519165, + 0.019161518663167953, + 0.02462308295071125, + 0.013240434229373932, + 0.027591926977038383, + 0.01943582482635975, + 0.01337355375289917, + 0.01003055926412344, + 0.02941574901342392, + 0.019598737359046936, + 0.03816244751214981, + 0.017520813271403313, + 0.040712226182222366, + 0.025308188050985336, + 0.016537129878997803, + 0.019552700221538544, + 0.038703352212905884, + 0.031357407569885254, + 0.011503600515425205, + 0.005042129196226597, + 0.008297676220536232, + 0.006145155522972345, + 0.003941818606108427, + 0.008408868685364723, + 0.005632130429148674, + 0.010725055821239948, + 0.0034384995233267546, + 0.002833707258105278, + 0.0034485654905438423, + 0.00305177248083055, + 0.0022838565055280924, + 0.002327470574527979, + 0.0038900491781532764, + 0.002667037071660161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.03743242844939232 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014224446378648281 + ], + "opposite_logprob_tokens": [ + -4.264224529266357 + ], + "logprob_reward": -0.014224446378648281, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014224446378648281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05933444947004318, + 0.1930149644613266, + 0.11779427528381348, + 0.07021404057741165, + 0.09104116261005402, + 0.1028653085231781, + 0.1716684252023697, + 0.1217280924320221, + 0.0627846047282219, + 0.009554727002978325 + ], + "credit_map": [ + 0.008751465007662773, + 0.007048333529382944, + 0.00924077071249485, + 0.0065735941752791405, + 0.005539247300475836, + 0.006699141580611467, + 0.006651445757597685, + 0.008830444887280464, + 0.021453963592648506, + 0.02329167164862156, + 0.025222817435860634, + 0.023032749071717262, + 0.05066025257110596, + 0.016257107257843018, + 0.01767778769135475, + 0.015418597497045994, + 0.016763530671596527, + 0.014673099853098392, + 0.015494279563426971, + 0.01779610477387905, + 0.021044500172138214, + 0.00892036035656929, + 0.013396457768976688, + 0.009705942124128342, + 0.00812366884201765, + 0.009707480669021606, + 0.008198846131563187, + 0.0069732218980789185, + 0.015941068530082703, + 0.006986864842474461, + 0.00625787815079093, + 0.008025010116398335, + 0.011281332932412624, + 0.013405014760792255, + 0.012238763272762299, + 0.013423982076346874, + 0.01480560377240181, + 0.0071955481544137, + 0.010044398717582226, + 0.008646517992019653, + 0.02044748328626156, + 0.010692446492612362, + 0.018010564148426056, + 0.009918120689690113, + 0.006054314319044352, + 0.011172082275152206, + 0.008789017796516418, + 0.017781278118491173, + 0.03129010275006294, + 0.013366389088332653, + 0.032209936529397964, + 0.01826445385813713, + 0.01176491566002369, + 0.014011071994900703, + 0.026301106438040733, + 0.0244604405015707, + 0.01863737963140011, + 0.01030191220343113, + 0.021287180483341217, + 0.01544464286416769, + 0.007090479601174593, + 0.012127364985644817, + 0.020461464300751686, + 0.01637767069041729, + 0.012921164743602276, + 0.0060992781072855, + 0.009579789824783802, + 0.008097139187157154, + 0.005146834068000317, + 0.0072553642094135284, + 0.008562025614082813, + 0.0051230089738965034, + 0.0019220913527533412, + 0.0014080285327509046, + 0.0012068268842995167, + 0.0011925135040655732, + 0.0007573380134999752, + 0.0009696642518974841, + 0.001030208426527679, + 0.001068055396899581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.014224446378648281 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331090159714222 + ], + "opposite_logprob_tokens": [ + -3.773310899734497 + ], + "logprob_reward": -0.02331090159714222, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331090159714222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11184737086296082, + 0.08732445538043976, + 0.07041836529970169, + 0.07017216831445694, + 0.10912121832370758, + 0.13826245069503784, + 0.2019188553094864, + 0.15064223110675812, + 0.04750768467783928, + 0.012785214930772781 + ], + "credit_map": [ + 0.015242571011185646, + 0.013193076476454735, + 0.017253953963518143, + 0.014093335717916489, + 0.011721701361238956, + 0.009200450032949448, + 0.014506598934531212, + 0.01663568615913391, + 0.009932074695825577, + 0.008649663999676704, + 0.011032483540475368, + 0.010953652672469616, + 0.02259633131325245, + 0.00617259694263339, + 0.007421513553708792, + 0.010566136799752712, + 0.009228764101862907, + 0.010066377930343151, + 0.009217062033712864, + 0.012427427805960178, + 0.010390523821115494, + 0.006234158296138048, + 0.0061631896533071995, + 0.006690859794616699, + 0.008581236936151981, + 0.007885400205850601, + 0.009121999144554138, + 0.00803333055227995, + 0.01452418603003025, + 0.006056275684386492, + 0.007487419527024031, + 0.008482316508889198, + 0.019393514841794968, + 0.009593318216502666, + 0.018100615590810776, + 0.010527326725423336, + 0.007945881225168705, + 0.017939109355211258, + 0.006091918330639601, + 0.01952952891588211, + 0.02738264761865139, + 0.015205977484583855, + 0.02366529032588005, + 0.013584769330918789, + 0.00863815937191248, + 0.015296930447220802, + 0.011096368543803692, + 0.02339230291545391, + 0.03382505103945732, + 0.016211383044719696, + 0.03635672479867935, + 0.02316303364932537, + 0.015301363542675972, + 0.014575476758182049, + 0.0349460132420063, + 0.027539817616343498, + 0.0255242008715868, + 0.01143650896847248, + 0.025757236406207085, + 0.016222398728132248, + 0.011245382949709892, + 0.014198441989719868, + 0.025025077164173126, + 0.02123299054801464, + 0.008943825960159302, + 0.004520629066973925, + 0.007294334936887026, + 0.005635562818497419, + 0.004740545060485601, + 0.004031250718981028, + 0.007864692248404026, + 0.004476846195757389, + 0.002496624831110239, + 0.0014497787924483418, + 0.0023006850387901068, + 0.0013066984247416258, + 0.0010904831578955054, + 0.001133193727582693, + 0.0016352589009329677, + 0.0013724926393479109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.02331090159714222 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024911104701459408 + ], + "opposite_logprob_tokens": [ + -6.002490997314453 + ], + "logprob_reward": -0.0024911104701459408, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024911104701459408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1368541419506073, + 0.07097535580396652, + 0.07598262280225754, + 0.07545429468154907, + 0.12408389896154404, + 0.1563214212656021, + 0.21137085556983948, + 0.0752762034535408, + 0.05041796714067459, + 0.023263277485966682 + ], + "credit_map": [ + 0.018191445618867874, + 0.01209200732409954, + 0.02344537153840065, + 0.017031041905283928, + 0.010547691956162453, + 0.009824548847973347, + 0.01764613948762417, + 0.028075894340872765, + 0.007596149574965239, + 0.008301383815705776, + 0.009190715849399567, + 0.007001904305070639, + 0.018086720257997513, + 0.005114799831062555, + 0.008483103476464748, + 0.0072005800902843475, + 0.009996108710765839, + 0.007143200375139713, + 0.011622142978012562, + 0.013488979078829288, + 0.010239707306027412, + 0.009604058228433132, + 0.005550554022192955, + 0.008337872102856636, + 0.008077447302639484, + 0.010974991135299206, + 0.01016885694116354, + 0.009701083414256573, + 0.014707610011100769, + 0.008955826982855797, + 0.006066215690225363, + 0.006802265532314777, + 0.01762537471950054, + 0.010728358291089535, + 0.016699377447366714, + 0.015826605260372162, + 0.007804344408214092, + 0.01586177386343479, + 0.010818332433700562, + 0.028719734400510788, + 0.030097510665655136, + 0.016650239005684853, + 0.024848418310284615, + 0.014540664851665497, + 0.009772115387022495, + 0.01906411722302437, + 0.012342128902673721, + 0.029006216675043106, + 0.03737456724047661, + 0.016616005450487137, + 0.038260605186223984, + 0.02219437062740326, + 0.01594090647995472, + 0.016529683023691177, + 0.034877609461545944, + 0.0295771062374115, + 0.013107120990753174, + 0.005899292416870594, + 0.011260929517447948, + 0.008074111305177212, + 0.00602673040702939, + 0.007537794299423695, + 0.012091421522200108, + 0.011278800666332245, + 0.01011150423437357, + 0.005031878128647804, + 0.007731918711215258, + 0.0049223448149859905, + 0.003720452543348074, + 0.00696519436314702, + 0.005440778564661741, + 0.006493892055004835, + 0.00422805966809392, + 0.0023942666593939066, + 0.002899196231737733, + 0.002805437659844756, + 0.0018401859560981393, + 0.0025293277576565742, + 0.0032888585701584816, + 0.0032779439352452755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.0024911104701459408 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, eye-level close-up of four clear, nested glass bowls on a dark, polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender female hand with immaculately manicured pale pink nails visible entering the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand place a gold chain necklace, a silver ring with diamonds, and pearl earrings into the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand smoothly withdraw after placing the items, leaving the jewels in the top bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting intricate highlights and shadows on the glass and jewels?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002825199393555522 + ], + "opposite_logprob_tokens": [ + -5.8778252601623535 + ], + "logprob_reward": -0.002825199393555522, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002825199393555522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19541878998279572, + 0.0769691914319992, + 0.062229909002780914, + 0.09596091508865356, + 0.09276825934648514, + 0.05366232618689537, + 0.06813212484121323, + 0.19135776162147522, + 0.1282053142786026, + 0.03529539331793785 + ], + "credit_map": [ + 0.04955232888460159, + 0.014573298394680023, + 0.03550420701503754, + 0.020315445959568024, + 0.012690294533967972, + 0.027226191014051437, + 0.011962306685745716, + 0.023594705387949944, + 0.007980785332620144, + 0.01003577746450901, + 0.010020587593317032, + 0.01134195551276207, + 0.015214336104691029, + 0.007456851191818714, + 0.007667835336178541, + 0.007251065224409103, + 0.010109166614711285, + 0.006012462545186281, + 0.011205211281776428, + 0.008356594480574131, + 0.01035173237323761, + 0.006166642066091299, + 0.005456092767417431, + 0.004572010599076748, + 0.012688352726399899, + 0.012854496017098427, + 0.015467782504856586, + 0.009253062307834625, + 0.019503798335790634, + 0.007731467951089144, + 0.00717325322329998, + 0.011288706213235855, + 0.017151465639472008, + 0.013564231805503368, + 0.013627725653350353, + 0.01103738322854042, + 0.012221244163811207, + 0.009876769036054611, + 0.005696702748537064, + 0.009592738002538681, + 0.011247048154473305, + 0.0049892994575202465, + 0.008820985443890095, + 0.00730338366702199, + 0.0032562273554503918, + 0.0051051536574959755, + 0.0061553362756967545, + 0.006784890778362751, + 0.009210542775690556, + 0.00871399324387312, + 0.006998023018240929, + 0.009664352051913738, + 0.007501392625272274, + 0.008758148178458214, + 0.011134474538266659, + 0.006151194684207439, + 0.03421078622341156, + 0.021770082414150238, + 0.025657352060079575, + 0.025531630963087082, + 0.01999174989759922, + 0.020553424954414368, + 0.02922693081200123, + 0.014415804296731949, + 0.024717826396226883, + 0.01305876113474369, + 0.02025250904262066, + 0.016392963007092476, + 0.01182519644498825, + 0.01407686248421669, + 0.018079690635204315, + 0.009801503270864487, + 0.005860296543687582, + 0.005295643117278814, + 0.004352199845016003, + 0.004247883800417185, + 0.0026883354876190424, + 0.003580401185899973, + 0.005272863432765007, + 0.003997770603746176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.002825199393555522 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_10/index_tensor([4180, 4180])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_10/index_tensor([4180, 4180])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3f82a2a47012254986e6a39c07821b8587257b55 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_10/index_tensor([4180, 4180])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a close-up, panning rightward, following the trajectory of a mahogany sword. An unseen hand swung the sword steadily and swiftly from left to right, shattering the gloomy silence. The blade, meticulously carved from deep reddish-brown mahogany, reveals its surface, embodying a patina of polished age, and revealing ancient runes and natural wood grain. As the blade sweeps across, a brilliant arc of golden light flashes through the air, leaving a fleeting afterimage that contrasts sharply with the surrounding darkness. The camera then quickly pulls back, revealing the full scene: a quaint Chinese-style hall, its sole source of light coming from a flickering oil lamp in the corner, casting long, undulating shadows across the wooden beams and the faintly visible bookshelves in the background. The entire picture is immersed in a retro tone like an old movie, with low color saturation, tending towards dark brown and yellowish warm colors. The grainy picture quality adds a sense of age, ultimately creating a solemn, mysterious and inner powerful visual atmosphere.", + "seed": 26252, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_10/index_4180_seed_26252_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.017216207180172204, + "vlm_logprob": -0.017216207180172204, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.017216207180172204, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021324990317225456 + ], + "opposite_logprob_tokens": [ + -4.14632511138916 + ], + "logprob_reward": -0.021324990317225456, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.021324990317225456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20884019136428833, + 0.1339731365442276, + 0.05223805457353592, + 0.047800250351428986, + 0.10921630263328552, + 0.12050098925828934, + 0.04614407569169998, + 0.1816740483045578, + 0.039172057062387466, + 0.06044090539216995 + ], + "credit_map": [ + 0.03007231093943119, + 0.02765575237572193, + 0.030513368546962738, + 0.022885536774992943, + 0.031352125108242035, + 0.02614501863718033, + 0.01713242568075657, + 0.02308366820216179, + 0.016398129984736443, + 0.012473197653889656, + 0.02141248993575573, + 0.017369814217090607, + 0.015066029503941536, + 0.01130708772689104, + 0.020238304510712624, + 0.019708098843693733, + 0.009333213791251183, + 0.006622752174735069, + 0.005233357660472393, + 0.005035857670009136, + 0.0038266752380877733, + 0.003593359375372529, + 0.011726759374141693, + 0.006866077892482281, + 0.00758233480155468, + 0.005734418518841267, + 0.0038180327974259853, + 0.002818132285028696, + 0.007136133033782244, + 0.007476747035980225, + 0.00607439037412405, + 0.007160059176385403, + 0.01224547065794468, + 0.012581522576510906, + 0.019071383401751518, + 0.010700078681111336, + 0.016291018575429916, + 0.018612660467624664, + 0.008706627413630486, + 0.011007540859282017, + 0.011980162933468819, + 0.02851732261478901, + 0.007784424815326929, + 0.012765522114932537, + 0.01204625889658928, + 0.017739038914442062, + 0.010133795440196991, + 0.019534457474946976, + 0.005143690854310989, + 0.005462956614792347, + 0.004905228037387133, + 0.004526703152805567, + 0.006691787391901016, + 0.006465802900493145, + 0.006720964331179857, + 0.0062269410118460655, + 0.008739056065678596, + 0.06706946343183517, + 0.005049319006502628, + 0.0322057269513607, + 0.019226623699069023, + 0.018761109560728073, + 0.019138433039188385, + 0.011484312824904919, + 0.004542181268334389, + 0.005890157073736191, + 0.0036694675218313932, + 0.005012288689613342, + 0.006274152547121048, + 0.0039254724979400635, + 0.0056359972804784775, + 0.00422233622521162, + 0.00579247809946537, + 0.008014987222850323, + 0.009436528198421001, + 0.008886045776307583, + 0.007926804013550282, + 0.004958756268024445, + 0.008349507115781307, + 0.007075797766447067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.021324990317225456 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369487449526787 + ], + "opposite_logprob_tokens": [ + -3.4086949825286865 + ], + "logprob_reward": -0.03369487449526787, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369487449526787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18324777483940125, + 0.1260315328836441, + 0.07100862264633179, + 0.017684288322925568, + 0.10329348593950272, + 0.13429506123065948, + 0.06777393817901611, + 0.22965992987155914, + 0.04031327739357948, + 0.026692116633057594 + ], + "credit_map": [ + 0.02928747422993183, + 0.026887163519859314, + 0.016524486243724823, + 0.013860795646905899, + 0.024061011150479317, + 0.026764480397105217, + 0.02276899293065071, + 0.02309337444603443, + 0.014937362633645535, + 0.01287374459207058, + 0.014875819906592369, + 0.015800897032022476, + 0.013651056215167046, + 0.00858058501034975, + 0.027461068704724312, + 0.017851000651717186, + 0.017829205840826035, + 0.007328586652874947, + 0.00600051274523139, + 0.005075996275991201, + 0.005444015841931105, + 0.0054092444479465485, + 0.01307607814669609, + 0.010844983160495758, + 0.002742983866482973, + 0.002715214155614376, + 0.0015165768563747406, + 0.0011955880327150226, + 0.002782084746286273, + 0.002524584997445345, + 0.0020809986162930727, + 0.002126257633790374, + 0.009605458937585354, + 0.011105425655841827, + 0.012978916987776756, + 0.011508702300488949, + 0.012147480621933937, + 0.015878358855843544, + 0.011470728553831577, + 0.018598418682813644, + 0.010521044954657555, + 0.01327377837151289, + 0.01204028632491827, + 0.011613326147198677, + 0.03391711786389351, + 0.023062681779265404, + 0.012274221517145634, + 0.017592601478099823, + 0.011573666706681252, + 0.006788791157305241, + 0.007944918237626553, + 0.006224792916327715, + 0.00887215230613947, + 0.009951514191925526, + 0.008852214552462101, + 0.007565893232822418, + 0.009741177782416344, + 0.08860070258378983, + 0.007493204437196255, + 0.052922025322914124, + 0.019449571147561073, + 0.01724458485841751, + 0.021287500858306885, + 0.012921158224344254, + 0.0039133639074862, + 0.0060817888006567955, + 0.004686946980655193, + 0.0050520882941782475, + 0.004493695683777332, + 0.004298336338251829, + 0.00706951878964901, + 0.00471753953024745, + 0.003329340135678649, + 0.004038657993078232, + 0.004317166283726692, + 0.003833855502307415, + 0.003908227197825909, + 0.0019002807093784213, + 0.003461164655163884, + 0.001903425669297576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.03369487449526787 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818636804819107 + ], + "opposite_logprob_tokens": [ + -4.018186569213867 + ], + "logprob_reward": -0.01818636804819107, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818636804819107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11226983368396759, + 0.062356434762477875, + 0.1809409111738205, + 0.03784025460481644, + 0.06723887473344803, + 0.06324208527803421, + 0.035408567637205124, + 0.3440038561820984, + 0.03995000198483467, + 0.05674915760755539 + ], + "credit_map": [ + 0.014868391677737236, + 0.02667040005326271, + 0.010719274170696735, + 0.009132321923971176, + 0.013535992242395878, + 0.011414796113967896, + 0.013290878385305405, + 0.012637773528695107, + 0.009521308355033398, + 0.007398994639515877, + 0.010473263449966908, + 0.007709248457103968, + 0.006911528296768665, + 0.004622180014848709, + 0.007356231566518545, + 0.00836368277668953, + 0.04077581316232681, + 0.018665624782443047, + 0.01645814999938011, + 0.014666915871202946, + 0.01204472966492176, + 0.013521447777748108, + 0.03894592821598053, + 0.025862300768494606, + 0.007373181637376547, + 0.005533790215849876, + 0.0018700410146266222, + 0.002560513559728861, + 0.006018789019435644, + 0.0039885626174509525, + 0.004951387643814087, + 0.00554399099200964, + 0.006894311401993036, + 0.006371087860316038, + 0.009369716979563236, + 0.007719092071056366, + 0.011797534301877022, + 0.00973043404519558, + 0.005297316238284111, + 0.010059377178549767, + 0.005268252920359373, + 0.01401775237172842, + 0.005893731024116278, + 0.006251723039895296, + 0.007280927151441574, + 0.009875485673546791, + 0.005698112305253744, + 0.008956102654337883, + 0.00372427050024271, + 0.004697943106293678, + 0.0034931099507957697, + 0.0037205314729362726, + 0.006301926448941231, + 0.004823393188416958, + 0.004571649711579084, + 0.004075746051967144, + 0.014931155368685722, + 0.1360110640525818, + 0.010301698930561543, + 0.09013862907886505, + 0.026593253016471863, + 0.0160012599080801, + 0.033912669867277145, + 0.016114117577672005, + 0.00463417824357748, + 0.005405720788985491, + 0.005154438782483339, + 0.0062358384020626545, + 0.005967531353235245, + 0.0030873571522533894, + 0.006275511812418699, + 0.0031894261483103037, + 0.006969675421714783, + 0.007570179179310799, + 0.007725453469902277, + 0.008458939380943775, + 0.006538181100040674, + 0.005550061352550983, + 0.00955283921211958, + 0.004383830353617668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.01818636804819107 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005247628316283226 + ], + "opposite_logprob_tokens": [ + -5.255247592926025 + ], + "logprob_reward": -0.005247628316283226, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005247628316283226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2381252497434616, + 0.053305696696043015, + 0.032739706337451935, + 0.027038894593715668, + 0.1736588180065155, + 0.06606723368167877, + 0.04073014482855797, + 0.1208283081650734, + 0.1034688875079155, + 0.14403700828552246 + ], + "credit_map": [ + 0.028077535331249237, + 0.06698434054851532, + 0.02046910859644413, + 0.018170421943068504, + 0.020343882963061333, + 0.017781924456357956, + 0.03712647780776024, + 0.029171563684940338, + 0.006855991203337908, + 0.0050957659259438515, + 0.007468329276889563, + 0.006040208041667938, + 0.004944986663758755, + 0.005525656510144472, + 0.011046504601836205, + 0.006328256335109472, + 0.007909928448498249, + 0.004070946015417576, + 0.003178780432790518, + 0.0025758487172424793, + 0.001561076263897121, + 0.0018822522833943367, + 0.006242028903216124, + 0.005318847019225359, + 0.00671871704980731, + 0.003434855258092284, + 0.0018154522404074669, + 0.001756387180648744, + 0.0027389617171138525, + 0.003695965511724353, + 0.0029449856374412775, + 0.003933571744710207, + 0.02339787967503071, + 0.029875226318836212, + 0.017150763422250748, + 0.016858480870723724, + 0.028190461918711662, + 0.01644887588918209, + 0.021869085729122162, + 0.01986805722117424, + 0.009506801143288612, + 0.0082588205114007, + 0.006019179243594408, + 0.005589040461927652, + 0.007831334136426449, + 0.009796153753995895, + 0.008318877778947353, + 0.01074703224003315, + 0.004917588550597429, + 0.004707304295152426, + 0.004621719941496849, + 0.007150007877498865, + 0.004579442087560892, + 0.005494425073266029, + 0.004966170527040958, + 0.004293491598218679, + 0.00511672580614686, + 0.0438811220228672, + 0.00439749239012599, + 0.030779484659433365, + 0.010718894191086292, + 0.0065772091038525105, + 0.012815757654607296, + 0.006541624199599028, + 0.012242366559803486, + 0.016310693696141243, + 0.010456796735525131, + 0.012772725895047188, + 0.012687155976891518, + 0.009138233959674835, + 0.01607903465628624, + 0.013781887479126453, + 0.017847517505288124, + 0.02150585874915123, + 0.019834965467453003, + 0.020030593499541283, + 0.015405582264065742, + 0.011137097142636776, + 0.02096531353890896, + 0.017310084775090218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.005247628316283226 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627174723893404 + ], + "opposite_logprob_tokens": [ + -4.882627010345459 + ], + "logprob_reward": -0.007627174723893404, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627174723893404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1440761834383011, + 0.0788833349943161, + 0.03385027498006821, + 0.03853515908122063, + 0.03286593779921532, + 0.06155776605010033, + 0.045449476689100266, + 0.3770185112953186, + 0.12028086185455322, + 0.06748256087303162 + ], + "credit_map": [ + 0.01760629005730152, + 0.023280370980501175, + 0.013187874108552933, + 0.019232187420129776, + 0.015162933617830276, + 0.02159751206636429, + 0.017202388495206833, + 0.01680661179125309, + 0.011146321892738342, + 0.01105821318924427, + 0.012262692674994469, + 0.008396284654736519, + 0.010442049242556095, + 0.006260095629841089, + 0.01016952283680439, + 0.009148144163191319, + 0.006281102076172829, + 0.005343012511730194, + 0.003307353937998414, + 0.0024763031397014856, + 0.0022182322572916746, + 0.0028023694176226854, + 0.006151046138256788, + 0.005270850844681263, + 0.009915941394865513, + 0.005191736854612827, + 0.001972217345610261, + 0.0023945330176502466, + 0.004120493307709694, + 0.0032806184608489275, + 0.003759186714887619, + 0.007900428026914597, + 0.003885146463289857, + 0.00393523508682847, + 0.005654125474393368, + 0.003127115312963724, + 0.0037787097971886396, + 0.004218050744384527, + 0.0036120598670095205, + 0.004655489698052406, + 0.0068565597757697105, + 0.010370437987148762, + 0.006185660604387522, + 0.00733543373644352, + 0.006015764083713293, + 0.009782567620277405, + 0.006075277924537659, + 0.008936057798564434, + 0.005020469427108765, + 0.006663868203759193, + 0.003709308337420225, + 0.004450580105185509, + 0.006431630812585354, + 0.0065652187913656235, + 0.0066156149841845036, + 0.0059927827678620815, + 0.016071433201432228, + 0.147700697183609, + 0.011004540137946606, + 0.09835118055343628, + 0.02998611144721508, + 0.01849021576344967, + 0.0367058664560318, + 0.018708426505327225, + 0.015230712480843067, + 0.01832064427435398, + 0.018789993599057198, + 0.015999823808670044, + 0.013968225568532944, + 0.01300987508147955, + 0.015629172325134277, + 0.009332398883998394, + 0.007020985707640648, + 0.010920734144747257, + 0.00788622535765171, + 0.011301540769636631, + 0.008876470848917961, + 0.004481824114918709, + 0.010438156314194202, + 0.006556617096066475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.007627174723893404 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a close-up, panning rightward, following the trajectory of a mahogany sword. An unseen hand swung the sword steadily and swiftly from left to right, shattering the gloomy silence. The blade, meticulously carved from deep reddish-brown mahogany, reveals its surface, embodying a patina of polished age, and revealing ancient runes and natural wood grain. As the blade sweeps across, a brilliant arc of golden light flashes through the air, leaving a fleeting afterimage that contrasts sharply with the surrounding darkness. The camera then quickly pulls back, revealing the full scene: a quaint Chinese-style hall, its sole source of light coming from a flickering oil lamp in the corner, casting long, undulating shadows across the wooden beams and the faintly visible bookshelves in the background. The entire picture is immersed in a retro tone like an old movie, with low color saturation, tending towards dark brown and yellowish warm colors. The grainy picture quality adds a sense of age, ultimately creating a solemn, mysterious and inner powerful visual atmosphere.", + "seed": 26253, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_10/index_4180_seed_26253_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02188096847385168, + "vlm_logprob": -0.02188096847385168, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.02188096847385168, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023602014407515526 + ], + "opposite_logprob_tokens": [ + -4.52360200881958 + ], + "logprob_reward": -0.023602014407515526, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.023602014407515526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13676753640174866, + 0.04735022783279419, + 0.06404794752597809, + 0.0527588352560997, + 0.0907549262046814, + 0.08710440248250961, + 0.05695931240916252, + 0.33692896366119385, + 0.06951572000980377, + 0.05781210958957672 + ], + "credit_map": [ + 0.017707983031868935, + 0.014568060636520386, + 0.0206412635743618, + 0.020331699401140213, + 0.017386527732014656, + 0.012410243973135948, + 0.014241119846701622, + 0.019480645656585693, + 0.006260542199015617, + 0.005814802832901478, + 0.01044907234609127, + 0.008322272449731827, + 0.0020838791970163584, + 0.0027560896705836058, + 0.006815942469984293, + 0.004847629461437464, + 0.008609216660261154, + 0.006420261226594448, + 0.014408563263714314, + 0.010222022421658039, + 0.002379738725721836, + 0.004910979885607958, + 0.007050974294543266, + 0.010046199895441532, + 0.0077421353198587894, + 0.0031787108164280653, + 0.006924337241798639, + 0.006288997828960419, + 0.0027651661075651646, + 0.004270903766155243, + 0.008388034999370575, + 0.013200554065406322, + 0.011548350565135479, + 0.007493075914680958, + 0.012393827550113201, + 0.009119155816733837, + 0.00802583247423172, + 0.015612073242664337, + 0.014783233404159546, + 0.011779380962252617, + 0.008813591673970222, + 0.011181870475411415, + 0.015262449160218239, + 0.012289621867239475, + 0.010595327243208885, + 0.014695351012051105, + 0.009021805599331856, + 0.005244391970336437, + 0.0062859863974153996, + 0.007025883067399263, + 0.008416389115154743, + 0.008517379872500896, + 0.01089109480381012, + 0.005079010967165232, + 0.005178377963602543, + 0.00556519627571106, + 0.029134053736925125, + 0.09956920892000198, + 0.10586077719926834, + 0.024458086118102074, + 0.0321197547018528, + 0.02132350578904152, + 0.01274797786027193, + 0.011715631000697613, + 0.007092601619660854, + 0.008802194148302078, + 0.008306891657412052, + 0.008651104755699635, + 0.013094479218125343, + 0.009051083587110043, + 0.006443930324167013, + 0.008073442615568638, + 0.010649978183209896, + 0.011365425772964954, + 0.008971589617431164, + 0.006137315183877945, + 0.0071562668308615685, + 0.004757866729050875, + 0.004014072008430958, + 0.004759600851684809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.023602014407515526 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809018433094025 + ], + "opposite_logprob_tokens": [ + -3.288090229034424 + ], + "logprob_reward": -0.03809018433094025, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809018433094025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21273502707481384, + 0.05044456943869591, + 0.04798030108213425, + 0.06172293797135353, + 0.1394660770893097, + 0.27020391821861267, + 0.0644301250576973, + 0.05776698887348175, + 0.07608382403850555, + 0.01916629634797573 + ], + "credit_map": [ + 0.03122221678495407, + 0.022344117984175682, + 0.023329636082053185, + 0.03224802389740944, + 0.029999783262610435, + 0.02154245600104332, + 0.021862652152776718, + 0.03018609806895256, + 0.00561052281409502, + 0.005734535399824381, + 0.006680914666503668, + 0.009864487685263157, + 0.0028275735676288605, + 0.0045735095627605915, + 0.006817301735281944, + 0.008335717022418976, + 0.006313157733529806, + 0.0058119031600654125, + 0.007317373529076576, + 0.00864733662456274, + 0.0021130982786417007, + 0.004043006803840399, + 0.006260971538722515, + 0.007473442703485489, + 0.007587452419102192, + 0.00408942298963666, + 0.011152353137731552, + 0.008470692671835423, + 0.0036032339558005333, + 0.006088626105338335, + 0.009449604898691177, + 0.011281542479991913, + 0.01927686482667923, + 0.011814036406576633, + 0.018706966191530228, + 0.019658954814076424, + 0.00986429862678051, + 0.017167065292596817, + 0.02279767021536827, + 0.02018020488321781, + 0.022728681564331055, + 0.01821511797606945, + 0.07283204793930054, + 0.021009953692555428, + 0.025999680161476135, + 0.07370999455451965, + 0.019480716437101364, + 0.016227679327130318, + 0.005720855668187141, + 0.012252395041286945, + 0.009499098174273968, + 0.007687387056648731, + 0.010284334421157837, + 0.007961241528391838, + 0.005608974955976009, + 0.005415827967226505, + 0.008085176348686218, + 0.012401261366903782, + 0.016575850546360016, + 0.005031077656894922, + 0.004810461774468422, + 0.0040937187150120735, + 0.002948564477264881, + 0.0038208705373108387, + 0.008395078592002392, + 0.015320627950131893, + 0.010054436512291431, + 0.009724093601107597, + 0.010433921590447426, + 0.006467341911047697, + 0.006279873661696911, + 0.009408439509570599, + 0.002110650995746255, + 0.002143740188330412, + 0.0030876186210662127, + 0.003058369504287839, + 0.0034335230011492968, + 0.0019953004084527493, + 0.0018169924151152372, + 0.0015200994675979018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.03809018433094025 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606355793774128 + ], + "opposite_logprob_tokens": [ + -4.141063690185547 + ], + "logprob_reward": -0.01606355793774128, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606355793774128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3336321711540222, + 0.05814887955784798, + 0.06756860762834549, + 0.0388038195669651, + 0.15002451837062836, + 0.09734227508306503, + 0.057994864881038666, + 0.03622140735387802, + 0.06368175894021988, + 0.09658167511224747 + ], + "credit_map": [ + 0.04608811065554619, + 0.04291297495365143, + 0.04948904737830162, + 0.03710075840353966, + 0.04514352232217789, + 0.04333280399441719, + 0.032494157552719116, + 0.037070807069540024, + 0.006963982712477446, + 0.007443615235388279, + 0.008110477589070797, + 0.009281224571168423, + 0.004305214621126652, + 0.006849369965493679, + 0.007282600738108158, + 0.00791239459067583, + 0.01219334825873375, + 0.005639605224132538, + 0.013043463230133057, + 0.009218696504831314, + 0.0030966580379754305, + 0.005165243986994028, + 0.00782808754593134, + 0.011383505538105965, + 0.003931890241801739, + 0.0033402990084141493, + 0.006044185254722834, + 0.004841012414544821, + 0.0034513582941144705, + 0.002098304918035865, + 0.009462322108447552, + 0.0056344494223594666, + 0.017511701211333275, + 0.01020466722548008, + 0.021518457680940628, + 0.017487330362200737, + 0.008582431823015213, + 0.015537421219050884, + 0.036093685775995255, + 0.023088818415999413, + 0.007109468337148428, + 0.008477621711790562, + 0.02090841345489025, + 0.013434242457151413, + 0.013354887254536152, + 0.013549667783081532, + 0.013185128569602966, + 0.007322842720896006, + 0.006189596373587847, + 0.007657355163246393, + 0.00883580558001995, + 0.006927394773811102, + 0.009839553385972977, + 0.005819797981530428, + 0.005609741434454918, + 0.007115617394447327, + 0.003474449273198843, + 0.007104178424924612, + 0.007600746117532253, + 0.003102819435298443, + 0.004128611646592617, + 0.004230996128171682, + 0.0025127434637397528, + 0.004066863097250462, + 0.008039449341595173, + 0.008223258890211582, + 0.009795425459742546, + 0.012220715172588825, + 0.009757019579410553, + 0.005910208448767662, + 0.004383747931569815, + 0.0053519271314144135, + 0.031431738287210464, + 0.03438722342252731, + 0.00788806565105915, + 0.005884876474738121, + 0.006670334376394749, + 0.004266665782779455, + 0.0025255067739635706, + 0.003527269931510091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.01606355793774128 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020576702430844307 + ], + "opposite_logprob_tokens": [ + -3.8955767154693604 + ], + "logprob_reward": -0.020576702430844307, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020576702430844307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0637722834944725, + 0.06042538955807686, + 0.02437032200396061, + 0.04819759353995323, + 0.0454389750957489, + 0.10200678557157516, + 0.09790525585412979, + 0.11866886913776398, + 0.21487481892108917, + 0.22433970868587494 + ], + "credit_map": [ + 0.008770350366830826, + 0.007361471187323332, + 0.0065523236989974976, + 0.010276050306856632, + 0.008929087780416012, + 0.004701424855738878, + 0.006666593719273806, + 0.010514984838664532, + 0.0065872869454324245, + 0.006924493238329887, + 0.011415312997996807, + 0.013527823612093925, + 0.003248788183555007, + 0.003707621479406953, + 0.0067481896840035915, + 0.008265869691967964, + 0.00382221769541502, + 0.0024832088965922594, + 0.005279116332530975, + 0.003900078358128667, + 0.0011185284238308668, + 0.0020768209360539913, + 0.002122033154591918, + 0.0035683189053088427, + 0.006340594496577978, + 0.0033012060448527336, + 0.010354388505220413, + 0.006576483603566885, + 0.0023062301333993673, + 0.004345666617155075, + 0.005596774630248547, + 0.009376252070069313, + 0.005631708540022373, + 0.0029104198329150677, + 0.005458749830722809, + 0.006440659519284964, + 0.003979246132075787, + 0.005786839406937361, + 0.008570870384573936, + 0.006660481449216604, + 0.00815823394805193, + 0.008202205412089825, + 0.018587153404951096, + 0.009887904860079288, + 0.014543650671839714, + 0.0216610599309206, + 0.012255704030394554, + 0.008710876107215881, + 0.013426200486719608, + 0.011470048688352108, + 0.016003627330064774, + 0.012959470972418785, + 0.013458900153636932, + 0.00989942904561758, + 0.00882445927709341, + 0.01186311524361372, + 0.019022539258003235, + 0.01522977277636528, + 0.018979022279381752, + 0.019881650805473328, + 0.014263267628848553, + 0.011397711001336575, + 0.00863285269588232, + 0.011262057349085808, + 0.020721208304166794, + 0.03309797868132591, + 0.029317421838641167, + 0.0285637229681015, + 0.039174921810626984, + 0.014816896989941597, + 0.019892485812306404, + 0.029290204867720604, + 0.06839108467102051, + 0.0751025602221489, + 0.01889759860932827, + 0.015247348695993423, + 0.017457541078329086, + 0.009938562288880348, + 0.009863000363111496, + 0.009441999718546867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.020576702430844307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072383262217045 + ], + "opposite_logprob_tokens": [ + -4.511072158813477 + ], + "logprob_reward": -0.011072383262217045, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072383262217045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11896290630102158, + 0.07928012311458588, + 0.037476081401109695, + 0.01872861385345459, + 0.06461230665445328, + 0.19379594922065735, + 0.1702033132314682, + 0.18670348823070526, + 0.07177349925041199, + 0.058463744819164276 + ], + "credit_map": [ + 0.0176398828625679, + 0.015045354142785072, + 0.01807641237974167, + 0.019089501351118088, + 0.01523249875754118, + 0.008982984349131584, + 0.014357446692883968, + 0.010538820177316666, + 0.00924734864383936, + 0.009704099968075752, + 0.015971407294273376, + 0.016616923734545708, + 0.0047222767025232315, + 0.005315401125699282, + 0.008382261730730534, + 0.00932040624320507, + 0.005631919950246811, + 0.004986180923879147, + 0.004264709539711475, + 0.005963299423456192, + 0.001551679102703929, + 0.002905801869928837, + 0.004270382225513458, + 0.007902108132839203, + 0.001983240246772766, + 0.001422533649019897, + 0.004106529522687197, + 0.0029182902071624994, + 0.0012297200737521052, + 0.001531261601485312, + 0.0030059663113206625, + 0.0025310730561614037, + 0.00666123628616333, + 0.004383096471428871, + 0.009730963967740536, + 0.009017415344715118, + 0.004701633937656879, + 0.010253972373902798, + 0.010513925924897194, + 0.009350065141916275, + 0.01499855238944292, + 0.013770601712167263, + 0.05321425199508667, + 0.012730223126709461, + 0.017304804176092148, + 0.0578564889729023, + 0.012338274158537388, + 0.011582736857235432, + 0.0207627322524786, + 0.03111126646399498, + 0.023746030405163765, + 0.019528057426214218, + 0.02813447080552578, + 0.01752449944615364, + 0.013582920655608177, + 0.015813345089554787, + 0.020974811166524887, + 0.04108068346977234, + 0.04138028621673584, + 0.025894952937960625, + 0.01893286034464836, + 0.017291273921728134, + 0.00996050052344799, + 0.011188105680048466, + 0.009373486042022705, + 0.010491791181266308, + 0.009576854296028614, + 0.008427402935922146, + 0.00750669976696372, + 0.009242787957191467, + 0.0075707514770329, + 0.009583724662661552, + 0.014208987355232239, + 0.016648905351758003, + 0.007135889492928982, + 0.006092417519539595, + 0.004544983617961407, + 0.003955658059567213, + 0.002734498120844364, + 0.0031424048356711864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.011072383262217045 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a close-up, panning rightward, following the trajectory of a mahogany sword. An unseen hand swung the sword steadily and swiftly from left to right, shattering the gloomy silence. The blade, meticulously carved from deep reddish-brown mahogany, reveals its surface, embodying a patina of polished age, and revealing ancient runes and natural wood grain. As the blade sweeps across, a brilliant arc of golden light flashes through the air, leaving a fleeting afterimage that contrasts sharply with the surrounding darkness. The camera then quickly pulls back, revealing the full scene: a quaint Chinese-style hall, its sole source of light coming from a flickering oil lamp in the corner, casting long, undulating shadows across the wooden beams and the faintly visible bookshelves in the background. The entire picture is immersed in a retro tone like an old movie, with low color saturation, tending towards dark brown and yellowish warm colors. The grainy picture quality adds a sense of age, ultimately creating a solemn, mysterious and inner powerful visual atmosphere.", + "seed": 26254, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_10/index_4180_seed_26254_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019778482615947723, + "vlm_logprob": -0.019778482615947723, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.019778482615947723, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029201893135905266 + ], + "opposite_logprob_tokens": [ + -3.9042019844055176 + ], + "logprob_reward": -0.029201893135905266, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.029201893135905266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2963680624961853, + 0.0474841333925724, + 0.16547845304012299, + 0.05560566112399101, + 0.05513787269592285, + 0.05460329353809357, + 0.04359599947929382, + 0.03790796548128128, + 0.09721948951482773, + 0.14659909904003143 + ], + "credit_map": [ + 0.02149609476327896, + 0.053023796528577805, + 0.04010359197854996, + 0.04327831417322159, + 0.039683934301137924, + 0.04775103181600571, + 0.025265086442232132, + 0.02576618082821369, + 0.004311977885663509, + 0.008685111999511719, + 0.0050435117445886135, + 0.006294107064604759, + 0.0060781994834542274, + 0.004723887424916029, + 0.006880840752273798, + 0.00546649657189846, + 0.006900649983435869, + 0.0044594318605959415, + 0.020195789635181427, + 0.010620958171784878, + 0.007349293678998947, + 0.05889361724257469, + 0.016041919589042664, + 0.04101678356528282, + 0.005375692620873451, + 0.007341759745031595, + 0.01392161101102829, + 0.005773204378783703, + 0.0036644008941948414, + 0.0037446473725140095, + 0.007867592386901379, + 0.00791674479842186, + 0.00702089536935091, + 0.00711441645398736, + 0.005823351908475161, + 0.009386450052261353, + 0.006223056931048632, + 0.005596050526946783, + 0.00731360400095582, + 0.006660043261945248, + 0.004646747838705778, + 0.004081501625478268, + 0.014949684031307697, + 0.006494683679193258, + 0.00648267287760973, + 0.005821002181619406, + 0.006816677749156952, + 0.005310323089361191, + 0.003286009654402733, + 0.005223631858825684, + 0.005543507635593414, + 0.005519102327525616, + 0.006144420709460974, + 0.0035888482816517353, + 0.010319268330931664, + 0.003971210215240717, + 0.003224174026399851, + 0.006023857742547989, + 0.004150493070483208, + 0.005055536050349474, + 0.0063402242958545685, + 0.0043677035719156265, + 0.004521102644503117, + 0.004224870819598436, + 0.008735505864024162, + 0.013326744548976421, + 0.020551374182105064, + 0.013396559283137321, + 0.01079491525888443, + 0.0041556209325790405, + 0.01659581996500492, + 0.009662937372922897, + 0.017705120146274567, + 0.05818372592329979, + 0.008083542808890343, + 0.0083564268425107, + 0.02375226467847824, + 0.0051704151555895805, + 0.009029567241668701, + 0.016318026930093765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.029201893135905266 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033695563673973083 + ], + "opposite_logprob_tokens": [ + -3.4086954593658447 + ], + "logprob_reward": -0.033695563673973083, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033695563673973083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0933610126376152, + 0.03413430228829384, + 0.20642536878585815, + 0.1744105964899063, + 0.07350347936153412, + 0.09780292212963104, + 0.029728949069976807, + 0.029483476653695107, + 0.18270668387413025, + 0.07844316214323044 + ], + "credit_map": [ + 0.005496957805007696, + 0.014274737797677517, + 0.009874498471617699, + 0.008410228416323662, + 0.018370267003774643, + 0.019229868426918983, + 0.010260899551212788, + 0.007443563546985388, + 0.0022946109529584646, + 0.004915895406156778, + 0.0038135689683258533, + 0.004288896918296814, + 0.0032616821117699146, + 0.0035073740873485804, + 0.008096426725387573, + 0.0039558494463562965, + 0.010997251607477665, + 0.005672064609825611, + 0.028809932991862297, + 0.009935205802321434, + 0.008038115687668324, + 0.08124256879091263, + 0.01618538610637188, + 0.045544855296611786, + 0.03319963067770004, + 0.019598742946982384, + 0.022495679557323456, + 0.026259027421474457, + 0.014315087348222733, + 0.01575435884296894, + 0.026814917102456093, + 0.015973176807165146, + 0.00905206985771656, + 0.009282341226935387, + 0.010058934800326824, + 0.013090185821056366, + 0.006943488493561745, + 0.008082639425992966, + 0.007879497483372688, + 0.009114332497119904, + 0.00830065831542015, + 0.007037915289402008, + 0.025977304205298424, + 0.011835072189569473, + 0.010353765450417995, + 0.010020189918577671, + 0.010546330362558365, + 0.013731691986322403, + 0.0026616882532835007, + 0.003581824479624629, + 0.004252014681696892, + 0.0031021905597299337, + 0.0039669219404459, + 0.0048204269260168076, + 0.004059736151248217, + 0.0032841505017131567, + 0.0027878766413778067, + 0.00290383817628026, + 0.0032743208575993776, + 0.004116901662200689, + 0.004183807875961065, + 0.004831233527511358, + 0.00320549588650465, + 0.004180006217211485, + 0.016474656760692596, + 0.012730986811220646, + 0.05142231658101082, + 0.021293627098202705, + 0.028708847239613533, + 0.009705726988613605, + 0.024399375542998314, + 0.017971165478229523, + 0.0075249504297971725, + 0.026216793805360794, + 0.006136205047369003, + 0.004543831571936607, + 0.011889132671058178, + 0.005095573607832193, + 0.006790900602936745, + 0.010245783254504204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.033695563673973083 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186016008257866 + ], + "opposite_logprob_tokens": [ + -4.018186092376709 + ], + "logprob_reward": -0.018186016008257866, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186016008257866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08821310102939606, + 0.030896561220288277, + 0.27481162548065186, + 0.26033321022987366, + 0.047208815813064575, + 0.08741065859794617, + 0.036473460495471954, + 0.024089625105261803, + 0.041930194944143295, + 0.10863273590803146 + ], + "credit_map": [ + 0.006011903285980225, + 0.012381537817418575, + 0.010356711223721504, + 0.009443667717278004, + 0.022375283762812614, + 0.01224592700600624, + 0.008114761672914028, + 0.007283309008926153, + 0.0021268026903271675, + 0.003709326731041074, + 0.004858480766415596, + 0.00482596643269062, + 0.0025005361530929804, + 0.002665624488145113, + 0.005674758460372686, + 0.004535062704235315, + 0.011196058243513107, + 0.009829806163907051, + 0.023215385153889656, + 0.01858038268983364, + 0.01317335944622755, + 0.08642522245645523, + 0.026266764849424362, + 0.08612462878227234, + 0.05077647417783737, + 0.040017884224653244, + 0.035073649138212204, + 0.028576184064149857, + 0.019953235983848572, + 0.024255817756056786, + 0.03569258004426956, + 0.02598738670349121, + 0.004735930357128382, + 0.006224227137863636, + 0.004597845952957869, + 0.010043914429843426, + 0.004817023407667875, + 0.005030968692153692, + 0.004997119773179293, + 0.006761786062270403, + 0.008194845169782639, + 0.00680927699431777, + 0.02585415355861187, + 0.01007218286395073, + 0.00806490983814001, + 0.008169792592525482, + 0.007841682061553001, + 0.012403815984725952, + 0.002987264422699809, + 0.005305710714310408, + 0.005424256902188063, + 0.0048796613700687885, + 0.0049810586497187614, + 0.0039599062874913216, + 0.0057670362293720245, + 0.0031685682479292154, + 0.0023310724645853043, + 0.0027138206642121077, + 0.0028996288310736418, + 0.0038468283601105213, + 0.0036151022650301456, + 0.002615223405882716, + 0.003018995514139533, + 0.0030489542987197638, + 0.003994294907897711, + 0.0066853854805231094, + 0.004833311773836613, + 0.005575099494308233, + 0.005937968380749226, + 0.001972809201106429, + 0.008111764676868916, + 0.004819562193006277, + 0.013181393034756184, + 0.043690554797649384, + 0.0070318421348929405, + 0.004881327971816063, + 0.017718518152832985, + 0.004488762002438307, + 0.006664818152785301, + 0.01097552664577961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.018186016008257866 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011069553904235363 + ], + "opposite_logprob_tokens": [ + -4.5110697746276855 + ], + "logprob_reward": -0.011069553904235363, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011069553904235363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11767031252384186, + 0.022205190733075142, + 0.26604196429252625, + 0.04848599061369896, + 0.07198978960514069, + 0.10106352716684341, + 0.057106856256723404, + 0.033519163727760315, + 0.17553569376468658, + 0.10638153553009033 + ], + "credit_map": [ + 0.006478703580796719, + 0.02208573743700981, + 0.009904066100716591, + 0.01595897041261196, + 0.02096616104245186, + 0.019054818898439407, + 0.013407744467258453, + 0.00981410313397646, + 0.0016872190171852708, + 0.003384316572919488, + 0.002898587379604578, + 0.0029859973583370447, + 0.0019474847940728068, + 0.0020363442599773407, + 0.00455906055867672, + 0.0027061793953180313, + 0.013730421662330627, + 0.007660098373889923, + 0.02353730984032154, + 0.012004864402115345, + 0.009598617441952229, + 0.10660772025585175, + 0.022392474114894867, + 0.07051046192646027, + 0.008064099587500095, + 0.004540954250842333, + 0.008924199268221855, + 0.009097317233681679, + 0.0027117012068629265, + 0.0036736996844410896, + 0.0065009137615561485, + 0.0049731056205928326, + 0.005303955636918545, + 0.00917523168027401, + 0.008407484740018845, + 0.013793469406664371, + 0.008853956125676632, + 0.007887185551226139, + 0.00810171477496624, + 0.010466794483363628, + 0.010184969753026962, + 0.013712803833186626, + 0.0207039937376976, + 0.011607400141656399, + 0.012155410833656788, + 0.009534870274364948, + 0.014165625907480717, + 0.008998449891805649, + 0.005001671612262726, + 0.005537459161132574, + 0.007190660573542118, + 0.007239907514303923, + 0.007542727515101433, + 0.008970526978373528, + 0.009225294925272465, + 0.006398604717105627, + 0.0030577420257031918, + 0.0033866027370095253, + 0.004476556554436684, + 0.00487788999453187, + 0.005748436786234379, + 0.0035044855903834105, + 0.004542727489024401, + 0.003924718592315912, + 0.016345130279660225, + 0.02228362485766411, + 0.036667756736278534, + 0.02501487359404564, + 0.018489709123969078, + 0.008993073366582394, + 0.02949756197631359, + 0.018243955448269844, + 0.010516592301428318, + 0.027998099103569984, + 0.013531730510294437, + 0.011576272547245026, + 0.012995582073926926, + 0.007767651695758104, + 0.009938477538526058, + 0.012057121843099594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.011069553904235363 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739386357367039 + ], + "opposite_logprob_tokens": [ + -5.006739616394043 + ], + "logprob_reward": -0.006739386357367039, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739386357367039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.124860018491745, + 0.03245972469449043, + 0.28331127762794495, + 0.06297294795513153, + 0.049581244587898254, + 0.08624149858951569, + 0.07505756616592407, + 0.06684724986553192, + 0.1502474993467331, + 0.06842099875211716 + ], + "credit_map": [ + 0.012510045431554317, + 0.014842557720839977, + 0.01783059909939766, + 0.01628865674138069, + 0.013632257468998432, + 0.020895879715681076, + 0.017713582143187523, + 0.011146431788802147, + 0.0020510985050350428, + 0.004341087769716978, + 0.003871422726660967, + 0.00412894319742918, + 0.004534510895609856, + 0.003328548977151513, + 0.006239792797714472, + 0.003964319359511137, + 0.01478593610227108, + 0.008173282258212566, + 0.024416126310825348, + 0.013261288404464722, + 0.010053995996713638, + 0.11152249574661255, + 0.02432147227227688, + 0.07677669823169708, + 0.009466392919421196, + 0.005978919565677643, + 0.010116532444953918, + 0.008126170374453068, + 0.005746207199990749, + 0.0065190414898097515, + 0.011025764048099518, + 0.005993923172354698, + 0.003623874392360449, + 0.007090302649885416, + 0.004844625946134329, + 0.008071087300777435, + 0.007103248964995146, + 0.005570889450609684, + 0.006023091729730368, + 0.007254128344357014, + 0.008059576153755188, + 0.009336535818874836, + 0.024455897510051727, + 0.011042924597859383, + 0.006793510168790817, + 0.007758719380944967, + 0.006250213831663132, + 0.012544117867946625, + 0.005744893103837967, + 0.007732849568128586, + 0.0070579322054982185, + 0.011540059931576252, + 0.010818691924214363, + 0.009170391596853733, + 0.015400945208966732, + 0.007591801229864359, + 0.005631817039102316, + 0.008729729801416397, + 0.007743483874946833, + 0.00999920628964901, + 0.010996310040354729, + 0.007132001221179962, + 0.009405161254107952, + 0.007209545001387596, + 0.013332915492355824, + 0.01120735239237547, + 0.04034162685275078, + 0.01506299152970314, + 0.027113396674394608, + 0.00924665480852127, + 0.017738675698637962, + 0.016203882172703743, + 0.007236703764647245, + 0.019880039617419243, + 0.00925917737185955, + 0.0043572853319346905, + 0.009067332372069359, + 0.005521163810044527, + 0.006013615522533655, + 0.007085682824254036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.006739386357367039 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a close-up, panning rightward, following the trajectory of a mahogany sword. An unseen hand swung the sword steadily and swiftly from left to right, shattering the gloomy silence. The blade, meticulously carved from deep reddish-brown mahogany, reveals its surface, embodying a patina of polished age, and revealing ancient runes and natural wood grain. As the blade sweeps across, a brilliant arc of golden light flashes through the air, leaving a fleeting afterimage that contrasts sharply with the surrounding darkness. The camera then quickly pulls back, revealing the full scene: a quaint Chinese-style hall, its sole source of light coming from a flickering oil lamp in the corner, casting long, undulating shadows across the wooden beams and the faintly visible bookshelves in the background. The entire picture is immersed in a retro tone like an old movie, with low color saturation, tending towards dark brown and yellowish warm colors. The grainy picture quality adds a sense of age, ultimately creating a solemn, mysterious and inner powerful visual atmosphere.", + "seed": 26255, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_10/index_4180_seed_26255_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.015360454190522432, + "vlm_logprob": -0.015360454190522432, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.015360454190522432, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03470426797866821 + ], + "opposite_logprob_tokens": [ + -3.9097042083740234 + ], + "logprob_reward": -0.03470426797866821, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03470426797866821, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03361115604639053, + 0.04198664054274559, + 0.05653082951903343, + 0.15480788052082062, + 0.47212547063827515, + 0.1063518300652504, + 0.036354515701532364, + 0.024464882910251617, + 0.03991985693573952, + 0.03384692966938019 + ], + "credit_map": [ + 0.006377924233675003, + 0.002419383730739355, + 0.007250436116009951, + 0.00293614505790174, + 0.002186025260016322, + 0.004402623046189547, + 0.002197750611230731, + 0.005840867757797241, + 0.006710388697683811, + 0.003448511241003871, + 0.00598708912730217, + 0.0038230058271437883, + 0.002492038533091545, + 0.004530881531536579, + 0.009646614082157612, + 0.005348110571503639, + 0.004677771124988794, + 0.003941952250897884, + 0.0049914345145225525, + 0.0036395436618477106, + 0.006205946672707796, + 0.013125344179570675, + 0.010537139140069485, + 0.00941169448196888, + 0.009009485132992268, + 0.008581342175602913, + 0.009700283408164978, + 0.008215252310037613, + 0.012007180601358414, + 0.012539882212877274, + 0.05116847902536392, + 0.04358597844839096, + 0.053619951009750366, + 0.020748840644955635, + 0.10114230215549469, + 0.01330711878836155, + 0.026392120867967606, + 0.20844769477844238, + 0.03141403570771217, + 0.017053408548235893, + 0.007967060431838036, + 0.012656218372285366, + 0.016699163243174553, + 0.019537469372153282, + 0.017926732078194618, + 0.01233817171305418, + 0.01289666723459959, + 0.006330346222966909, + 0.003180720377713442, + 0.004281308967620134, + 0.005704810377210379, + 0.0036083131562918425, + 0.005768979899585247, + 0.005232071503996849, + 0.004414890427142382, + 0.004163419362157583, + 0.002604707842692733, + 0.0037488979287445545, + 0.004864024929702282, + 0.0021777614019811153, + 0.0033923382870852947, + 0.002597637241706252, + 0.0016859965398907661, + 0.0033935196697711945, + 0.003561942372471094, + 0.006926350295543671, + 0.0065566725097596645, + 0.004230340011417866, + 0.005754288751631975, + 0.004847256932407618, + 0.004466845653951168, + 0.0035761611070483923, + 0.00341944326646626, + 0.006983271799981594, + 0.005077713634818792, + 0.003116068197414279, + 0.003213167190551758, + 0.006324616726487875, + 0.003729580668732524, + 0.0019830709788948298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.03470426797866821 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058148942887783 + ], + "opposite_logprob_tokens": [ + -3.8955814838409424 + ], + "logprob_reward": -0.02058148942887783, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058148942887783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0636819526553154, + 0.14659667015075684, + 0.06079241633415222, + 0.2864619195461273, + 0.2823816239833832, + 0.045903708785772324, + 0.029712960124015808, + 0.02412615716457367, + 0.032860636711120605, + 0.02748202346265316 + ], + "credit_map": [ + 0.011959969997406006, + 0.005573011469095945, + 0.008213404566049576, + 0.0047095282934606075, + 0.0037405621260404587, + 0.012957200407981873, + 0.0036965010222047567, + 0.012831768952310085, + 0.015095332637429237, + 0.008837209083139896, + 0.036057181656360626, + 0.009302730672061443, + 0.008651670999825, + 0.02512039616703987, + 0.016595052555203438, + 0.026937082409858704, + 0.005853051319718361, + 0.004384422674775124, + 0.008118271827697754, + 0.004252828191965818, + 0.00942739937454462, + 0.012385585345327854, + 0.008332810364663601, + 0.008038046769797802, + 0.013314886018633842, + 0.010996348224580288, + 0.013975041918456554, + 0.014445539563894272, + 0.015906600281596184, + 0.018298234790563583, + 0.11666619032621384, + 0.08285906165838242, + 0.05822677165269852, + 0.015050946734845638, + 0.10957817733287811, + 0.014374988153576851, + 0.014660721644759178, + 0.04692225158214569, + 0.013702514581382275, + 0.00986522901803255, + 0.004357106052339077, + 0.005067985504865646, + 0.005868077743798494, + 0.006939393002539873, + 0.00680241035297513, + 0.0058289142325520515, + 0.005879989359527826, + 0.005159827880561352, + 0.0021587226074188948, + 0.0032763571944087744, + 0.003640008857473731, + 0.004983463324606419, + 0.004606707952916622, + 0.0032132624182850122, + 0.004773372318595648, + 0.0030610612593591213, + 0.002430060412734747, + 0.0024001013953238726, + 0.003280005417764187, + 0.002932579256594181, + 0.0025208189617842436, + 0.0032669862266629934, + 0.0022786923218518496, + 0.005016911309212446, + 0.004126828629523516, + 0.006199688650667667, + 0.004808621946722269, + 0.00385479093529284, + 0.004526801872998476, + 0.004296622239053249, + 0.00287729618139565, + 0.002169982763007283, + 0.0032852974254637957, + 0.006466038059443235, + 0.004308649338781834, + 0.002696757670491934, + 0.0029597810935229063, + 0.0031885646749287844, + 0.0026524271816015244, + 0.00192450440954417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.02058148942887783 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638513274490833 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638513274490833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638513274490833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08034022897481918, + 0.06054674834012985, + 0.04724191501736641, + 0.0886433869600296, + 0.618186354637146, + 0.0284605473279953, + 0.018942369148135185, + 0.021802494302392006, + 0.022116312757134438, + 0.01371969934552908 + ], + "credit_map": [ + 0.01657661609351635, + 0.005456109996885061, + 0.008146771229803562, + 0.008474056608974934, + 0.004460169468075037, + 0.008525211364030838, + 0.006111057475209236, + 0.022590234875679016, + 0.006857719738036394, + 0.004432436544448137, + 0.010834637098014355, + 0.0034056103322654963, + 0.00471896817907691, + 0.01285803597420454, + 0.007223986089229584, + 0.01021535787731409, + 0.0054930634796619415, + 0.004483143333345652, + 0.004412793088704348, + 0.0031630713492631912, + 0.0072616045363247395, + 0.006975146010518074, + 0.0045689488761126995, + 0.010884144343435764, + 0.012180277146399021, + 0.005445344373583794, + 0.005929064005613327, + 0.0033847875893115997, + 0.004008545074611902, + 0.0059243482537567616, + 0.02207420952618122, + 0.029696816578507423, + 0.06040313094854355, + 0.025186672806739807, + 0.14849503338336945, + 0.016687743365764618, + 0.031776126474142075, + 0.2776600420475006, + 0.0365704707801342, + 0.021407105028629303, + 0.0027132686227560043, + 0.004416061565279961, + 0.003977959975600243, + 0.004499420989304781, + 0.0026419253554195166, + 0.0049600908532738686, + 0.0028041545301675797, + 0.002447666134685278, + 0.0015390466433018446, + 0.0025565410032868385, + 0.0023562866263091564, + 0.0029666144400835037, + 0.0036010255571454763, + 0.00209617312066257, + 0.0018201437778770924, + 0.0020065377466380596, + 0.0016612955369055271, + 0.0032816254533827305, + 0.0033865366131067276, + 0.00217448896728456, + 0.003779070219025016, + 0.0030408177990466356, + 0.001637160312384367, + 0.0028415003325790167, + 0.002171681495383382, + 0.004419451579451561, + 0.0055121504701673985, + 0.002301961649209261, + 0.002223061863332987, + 0.0015653386944904923, + 0.002112210961058736, + 0.0018104545306414366, + 0.0013657354284077883, + 0.0035409319680184126, + 0.002984195714816451, + 0.0013274064986035228, + 0.0010392515687271953, + 0.0015250525902956724, + 0.0012025011237710714, + 0.0007346243946813047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.008638513274490833 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248814355581999 + ], + "opposite_logprob_tokens": [ + -5.2552490234375 + ], + "logprob_reward": -0.005248814355581999, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248814355581999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02101556956768036, + 0.02549668587744236, + 0.016888335347175598, + 0.13379737734794617, + 0.3672546148300171, + 0.11576618254184723, + 0.06206493079662323, + 0.06965305656194687, + 0.09637654572725296, + 0.09168674051761627 + ], + "credit_map": [ + 0.004346698522567749, + 0.001501450315117836, + 0.0036748170386999846, + 0.0023945546709001064, + 0.0012084569316357374, + 0.0025597610510885715, + 0.0014878259971737862, + 0.003842006204649806, + 0.004348923917859793, + 0.0019380048615857959, + 0.004256919026374817, + 0.0021306925918906927, + 0.0015526830684393644, + 0.002040696097537875, + 0.004821436479687691, + 0.004407329484820366, + 0.0015058205462992191, + 0.0014083368005231023, + 0.0021087077911943197, + 0.0012249455321580172, + 0.0027981556486338377, + 0.0028344036545604467, + 0.002274378202855587, + 0.002733586821705103, + 0.007719779387116432, + 0.00525843258947134, + 0.009056986309587955, + 0.005147513933479786, + 0.007074674125760794, + 0.008316738530993462, + 0.04996504634618759, + 0.041258201003074646, + 0.03153161332011223, + 0.010829799808561802, + 0.13810370862483978, + 0.012340275570750237, + 0.013141789473593235, + 0.1420205533504486, + 0.010447019711136818, + 0.008839852176606655, + 0.008901010267436504, + 0.016300033777952194, + 0.01735875755548477, + 0.01456461101770401, + 0.018626052886247635, + 0.017891326919198036, + 0.013028147630393505, + 0.009096247144043446, + 0.004278136417269707, + 0.007862342521548271, + 0.007754122372716665, + 0.009367481805384159, + 0.011227065697312355, + 0.00867240596562624, + 0.005786026827991009, + 0.007117350585758686, + 0.007768901530653238, + 0.010122890584170818, + 0.014036325737833977, + 0.00426030158996582, + 0.009359333664178848, + 0.008588660508394241, + 0.005464361049234867, + 0.010052281431853771, + 0.00823462475091219, + 0.025685161352157593, + 0.019776342436671257, + 0.008038816042244434, + 0.011014437302947044, + 0.010955734178423882, + 0.006959987338632345, + 0.005711439065635204, + 0.00838388130068779, + 0.019589735195040703, + 0.016254976391792297, + 0.010489101521670818, + 0.008954950608313084, + 0.011619259603321552, + 0.00916504580527544, + 0.007229792885482311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.005248814355581999 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mahogany sword with deep reddish-brown color, visible wood grain, and ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a close-up shot of the sword being swung swiftly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a brilliant arc of golden light flashing through the air as the sword sweeps, leaving a fleeting afterimage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera quickly pull back to reveal a Chinese-style hall with a flickering oil lamp, wooden beams, and bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style retro with low color saturation, dark brown and yellowish warm tones, and a grainy picture quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629185914993286 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.007629185914993286, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629185914993286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02768576331436634, + 0.04112901911139488, + 0.035575736314058304, + 0.09276945143938065, + 0.42342129349708557, + 0.0612994022667408, + 0.046375297009944916, + 0.0871521532535553, + 0.15496136248111725, + 0.02963051199913025 + ], + "credit_map": [ + 0.004614087752997875, + 0.0020711454562842846, + 0.005386355333030224, + 0.003820443293079734, + 0.0017688260413706303, + 0.0036547682248055935, + 0.0023619025014340878, + 0.004008234012871981, + 0.0039206454530358315, + 0.0026694433763623238, + 0.009997673332691193, + 0.003307924373075366, + 0.002949964487925172, + 0.006490635685622692, + 0.004969608038663864, + 0.00682312436401844, + 0.0029358656611293554, + 0.002528571290895343, + 0.0033268872648477554, + 0.00284715392626822, + 0.004201098345220089, + 0.007147241849452257, + 0.0072076404467225075, + 0.005381274502724409, + 0.008835677057504654, + 0.004647486377507448, + 0.0046296073123812675, + 0.0034788702614605427, + 0.004478217102587223, + 0.00709853321313858, + 0.02706070989370346, + 0.032540351152420044, + 0.02371002919971943, + 0.014436026103794575, + 0.13680250942707062, + 0.01214044913649559, + 0.018775487318634987, + 0.18664585053920746, + 0.018596677109599113, + 0.012314272113144398, + 0.005051347892731428, + 0.007389974780380726, + 0.00791938416659832, + 0.00838488433510065, + 0.010856016539037228, + 0.008451537229120731, + 0.0078081232495605946, + 0.00543813593685627, + 0.004286312032490969, + 0.007354446686804295, + 0.005937040317803621, + 0.0065409294329583645, + 0.0077632125467062, + 0.004330407362431288, + 0.005709053482860327, + 0.004453894216567278, + 0.008068284951150417, + 0.015276586636900902, + 0.013909965753555298, + 0.01507079228758812, + 0.010724490508437157, + 0.009227392263710499, + 0.006590026430785656, + 0.008284623734652996, + 0.015773674473166466, + 0.04572406783699989, + 0.03528587892651558, + 0.017568202689290047, + 0.013151758350431919, + 0.009107680059969425, + 0.011554166674613953, + 0.006795941386371851, + 0.0024728509597480297, + 0.005773280281573534, + 0.006729112938046455, + 0.0026161239948123693, + 0.002303080400452018, + 0.004465204663574696, + 0.0027505881153047085, + 0.002520271809771657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.007629185914993286 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_11/index_tensor([5054, 5054])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_11/index_tensor([5054, 5054])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..454b3dbdada8028ddfb93400c6e65dd53a7e1768 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_11/index_tensor([5054, 5054])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tracking medium close-up shot, shot from a handheld camera and slightly tilted upward, locks on a thoroughly smudged clown, stumbling and scrambling down a dark, damp, cobblestoned back alley. His unkempt, colorful wig is tilted to one side, and his signature white greasepaint base, cracked and flaking from sweat, forms mottled patterns across his terrified face. The large, red smiley lipstick, which should have brought joy, has been roughly smeared, leaving a grotesque scar across his cheek. The ultra-high-definition detail reveals every glistening bead of sweat, seeping from his furrowed brow and temple, running down the furrows of greasepaint. His once vibrant polka-dot clown costume is now filthy and torn, and his heavy red clown shoes make a desperate, clattering sound on the slippery cobblestones. The only dim street lamp deep in the alley cast a bleak light from behind him, outlining a tense outline on his body. At the same time, his running figure was cast a huge and distorted shadow on the graffiti-covered brick wall in front of him. The whole picture was filled with an atmosphere of extreme tension, chaos and despair in the violent shaking.", + "seed": 18298, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_11/index_5054_seed_18298_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.056858055177144706, + "vlm_logprob": -0.056858055177144706, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.056858055177144706, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006126197054982185 + ], + "opposite_logprob_tokens": [ + -5.631126403808594 + ], + "logprob_reward": -0.006126197054982185, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.006126197054982185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1422790139913559, + 0.11518234014511108, + 0.07047078758478165, + 0.06808441877365112, + 0.06363477557897568, + 0.09270386397838593, + 0.04325709491968155, + 0.08348767459392548, + 0.2269803136587143, + 0.09391968697309494 + ], + "credit_map": [ + 0.01970468834042549, + 0.018552027642726898, + 0.015602405183017254, + 0.01999935321509838, + 0.018844369798898697, + 0.01865564100444317, + 0.01931360550224781, + 0.011606919579207897, + 0.014780694618821144, + 0.01619088277220726, + 0.016929803416132927, + 0.014150689356029034, + 0.013018239289522171, + 0.01072259433567524, + 0.018447119742631912, + 0.010942323133349419, + 0.009850222617387772, + 0.007250710856169462, + 0.011500263586640358, + 0.00781287346035242, + 0.008824985474348068, + 0.009941726922988892, + 0.006717880256474018, + 0.008572123013436794, + 0.007511636707931757, + 0.005067295860499144, + 0.013905171304941177, + 0.009940254501998425, + 0.006716567557305098, + 0.007307483348995447, + 0.005706647410988808, + 0.011929361149668694, + 0.005404678173363209, + 0.005759743042290211, + 0.010643928311765194, + 0.00624657841399312, + 0.009325758554041386, + 0.010691423900425434, + 0.008388230577111244, + 0.007174429018050432, + 0.013426204212009907, + 0.00757942721247673, + 0.018606068566441536, + 0.0051438817754387856, + 0.00878320075571537, + 0.009233321063220501, + 0.015001603402197361, + 0.014930158853530884, + 0.003961547743529081, + 0.004478538874536753, + 0.006001260131597519, + 0.0035469243302941322, + 0.005950161721557379, + 0.006034201476722956, + 0.007247720379382372, + 0.006036740727722645, + 0.008279153145849705, + 0.013248848728835583, + 0.008573820814490318, + 0.009464089758694172, + 0.008019980043172836, + 0.010009952820837498, + 0.013808374293148518, + 0.01208345778286457, + 0.024460675194859505, + 0.02956770732998848, + 0.012722518295049667, + 0.02244659885764122, + 0.020767241716384888, + 0.021190166473388672, + 0.05547364428639412, + 0.040351759642362595, + 0.008112505078315735, + 0.008187071420252323, + 0.005891079548746347, + 0.015804333612322807, + 0.012684129178524017, + 0.015212259255349636, + 0.016294872388243675, + 0.011733438819646835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.006126197054982185 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004149258136749 + ], + "opposite_logprob_tokens": [ + -2.6950414180755615 + ], + "logprob_reward": -0.07004149258136749, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004149258136749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16589763760566711, + 0.054494570940732956, + 0.08796600252389908, + 0.07637689262628555, + 0.05155433341860771, + 0.0996079295873642, + 0.04565127193927765, + 0.07610264420509338, + 0.2922585606575012, + 0.05009015277028084 + ], + "credit_map": [ + 0.02291649580001831, + 0.026125594973564148, + 0.02754954807460308, + 0.018068227916955948, + 0.024593893438577652, + 0.01879279688000679, + 0.015049717389047146, + 0.01280137337744236, + 0.0072583225555717945, + 0.007447737734764814, + 0.008109175600111485, + 0.005290063098073006, + 0.005771966185420752, + 0.0059452359564602375, + 0.008881613612174988, + 0.005790454335510731, + 0.010225233621895313, + 0.005350656807422638, + 0.015196295455098152, + 0.009391979314386845, + 0.015560495667159557, + 0.009802587330341339, + 0.009386027231812477, + 0.013052724301815033, + 0.007472015451639891, + 0.007086530793458223, + 0.012400311417877674, + 0.012175221927464008, + 0.009811868891119957, + 0.01000981591641903, + 0.008016654290258884, + 0.009404479525983334, + 0.006296340376138687, + 0.003937462344765663, + 0.009062484838068485, + 0.007095209322869778, + 0.00413990905508399, + 0.008460554294288158, + 0.005987763870507479, + 0.006574613507837057, + 0.011902778409421444, + 0.009768830612301826, + 0.0172678641974926, + 0.009075328707695007, + 0.014269043691456318, + 0.012449328787624836, + 0.012413420714437962, + 0.012461328878998756, + 0.003988457378000021, + 0.004160615149885416, + 0.005612276494503021, + 0.003279562573879957, + 0.005406776908785105, + 0.008508830331265926, + 0.006115227472037077, + 0.008579524233937263, + 0.007951987907290459, + 0.011208527721464634, + 0.0077107856050133705, + 0.007706877309828997, + 0.008784661069512367, + 0.007500913459807634, + 0.013458633795380592, + 0.011780265718698502, + 0.028406577184796333, + 0.03721073269844055, + 0.015475123189389706, + 0.030152516439557076, + 0.02753787860274315, + 0.025329651311039925, + 0.07411997765302658, + 0.05402609705924988, + 0.004220639821141958, + 0.005713587626814842, + 0.0038807708770036697, + 0.0077555617317557335, + 0.006110661663115025, + 0.008064218796789646, + 0.0073139420710504055, + 0.007030771113932133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.07004149258136749 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00134222989436239 + ], + "opposite_logprob_tokens": [ + -6.626342296600342 + ], + "logprob_reward": -0.00134222989436239, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.00134222989436239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14253826439380646, + 0.10015671700239182, + 0.07802640646696091, + 0.0531371608376503, + 0.13337045907974243, + 0.11613427102565765, + 0.04994657635688782, + 0.0866590216755867, + 0.17278200387954712, + 0.06724908202886581 + ], + "credit_map": [ + 0.02177211456000805, + 0.013770382851362228, + 0.01444154605269432, + 0.01610792987048626, + 0.020691314712166786, + 0.028338894248008728, + 0.013695447705686092, + 0.013720634393393993, + 0.01377660222351551, + 0.016335614025592804, + 0.014436733908951283, + 0.009777209721505642, + 0.008610543794929981, + 0.014288162812590599, + 0.014245404861867428, + 0.00868644192814827, + 0.010514674708247185, + 0.00734332948923111, + 0.012314924038946629, + 0.01149718090891838, + 0.007561807986348867, + 0.009240242652595043, + 0.010110155679285526, + 0.009444092400372028, + 0.006255307700484991, + 0.0046728309243917465, + 0.008131635375320911, + 0.008949216455221176, + 0.005234985146671534, + 0.008987722918391228, + 0.005384705029428005, + 0.005520760081708431, + 0.011820878833532333, + 0.008323571644723415, + 0.02769049070775509, + 0.018953371793031693, + 0.011903789825737476, + 0.01833687722682953, + 0.012470163404941559, + 0.023871321231126785, + 0.014836895279586315, + 0.008320193737745285, + 0.02247346192598343, + 0.009171905927360058, + 0.012173322960734367, + 0.014159257523715496, + 0.015572174452245235, + 0.019427062943577766, + 0.006484231911599636, + 0.004288067575544119, + 0.004464361816644669, + 0.003830250818282366, + 0.006995920091867447, + 0.006596000399440527, + 0.00914501678198576, + 0.008142723701894283, + 0.006937227211892605, + 0.009683790616691113, + 0.010078011080622673, + 0.008801081217825413, + 0.008061842061579227, + 0.012117613106966019, + 0.016480615362524986, + 0.014498834498226643, + 0.018205201253294945, + 0.024486161768436432, + 0.011383133009076118, + 0.02058856561779976, + 0.019302774220705032, + 0.013455137610435486, + 0.04196001961827278, + 0.023401016369462013, + 0.006690036039799452, + 0.007096346002072096, + 0.005465739872306585, + 0.01050869282335043, + 0.009299310855567455, + 0.01138714887201786, + 0.009086295031011105, + 0.007715513464063406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.00134222989436239 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052520157769322395 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.0052520157769322395, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052520157769322395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08251773566007614, + 0.0543663427233696, + 0.07969985902309418, + 0.09217897057533264, + 0.0920485258102417, + 0.10956092178821564, + 0.10008113086223602, + 0.16071243584156036, + 0.1511571854352951, + 0.07767687737941742 + ], + "credit_map": [ + 0.013989483006298542, + 0.008619433268904686, + 0.009108959697186947, + 0.00976521335542202, + 0.00861367117613554, + 0.012811629101634026, + 0.012538990005850792, + 0.007070350926369429, + 0.006299815606325865, + 0.004987095016986132, + 0.007773558609187603, + 0.007147533353418112, + 0.00806533545255661, + 0.006001442205160856, + 0.009454457089304924, + 0.004637105390429497, + 0.01173933781683445, + 0.008947446942329407, + 0.012982078827917576, + 0.01038384810090065, + 0.008837519213557243, + 0.008098702877759933, + 0.008410939946770668, + 0.010299988090991974, + 0.009359448216855526, + 0.009403769858181477, + 0.015455116517841816, + 0.012789996340870857, + 0.009348765015602112, + 0.015068565495312214, + 0.009429559111595154, + 0.011323750950396061, + 0.006518401205539703, + 0.005135562736541033, + 0.017595667392015457, + 0.014228666201233864, + 0.01100597158074379, + 0.012022909708321095, + 0.011024459265172482, + 0.014516884461045265, + 0.012532198801636696, + 0.009736905805766582, + 0.018716396763920784, + 0.008949286304414272, + 0.009129495359957218, + 0.013566680252552032, + 0.018961606547236443, + 0.017968345433473587, + 0.01605130545794964, + 0.008324946276843548, + 0.010435180738568306, + 0.0071733491495251656, + 0.010812328197062016, + 0.013836300000548363, + 0.017001405358314514, + 0.016446314752101898, + 0.013666892424225807, + 0.015851112082600594, + 0.025745555758476257, + 0.024009497836232185, + 0.012067769654095173, + 0.013333501294255257, + 0.029036549851298332, + 0.027001552283763885, + 0.017856189981102943, + 0.017605580389499664, + 0.007825169712305069, + 0.012102184817194939, + 0.011608476750552654, + 0.016005350276827812, + 0.03712931275367737, + 0.031024932861328125, + 0.010121265426278114, + 0.010008755140006542, + 0.007714372128248215, + 0.012654856778681278, + 0.010865981690585613, + 0.0066354707814753056, + 0.008979860693216324, + 0.010696322657167912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.0052520157769322395 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20152834057807922 + ], + "opposite_logprob_tokens": [ + -1.7015283107757568 + ], + "logprob_reward": -0.20152834057807922, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20152834057807922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11320917308330536, + 0.07456415891647339, + 0.10223107784986496, + 0.06345091760158539, + 0.05757194012403488, + 0.1252003312110901, + 0.13821955025196075, + 0.05726214498281479, + 0.22032317519187927, + 0.047967541962862015 + ], + "credit_map": [ + 0.016693342477083206, + 0.016885165125131607, + 0.015039963647723198, + 0.013816415332257748, + 0.014517235569655895, + 0.011164876632392406, + 0.013959129340946674, + 0.011133043095469475, + 0.009047946892678738, + 0.006926929112523794, + 0.01073781680315733, + 0.009184286929666996, + 0.006637738551944494, + 0.014254549518227577, + 0.010411986149847507, + 0.007362897042185068, + 0.010779529809951782, + 0.007349969819188118, + 0.011597700417041779, + 0.01129079982638359, + 0.02339046820998192, + 0.013953730463981628, + 0.01217729039490223, + 0.011691589839756489, + 0.008481694385409355, + 0.0069984858855605125, + 0.008593941107392311, + 0.008530685678124428, + 0.006701591890305281, + 0.010453899390995502, + 0.005975621286779642, + 0.007715000305324793, + 0.0051670679822564125, + 0.004933539312332869, + 0.007080704905092716, + 0.00908755138516426, + 0.009173717349767685, + 0.00901345256716013, + 0.006667658686637878, + 0.006448247004300356, + 0.011375389993190765, + 0.010782046243548393, + 0.018468104302883148, + 0.008034496568143368, + 0.018430382013320923, + 0.01860598661005497, + 0.020967325195670128, + 0.018536604940891266, + 0.014421761967241764, + 0.015748389065265656, + 0.016298901289701462, + 0.013958113268017769, + 0.02299671620130539, + 0.014751287177205086, + 0.018009407445788383, + 0.02203497104346752, + 0.004515216685831547, + 0.0051629748195409775, + 0.007161447312682867, + 0.008476896211504936, + 0.006740034092217684, + 0.0071090711280703545, + 0.009710444137454033, + 0.008386057801544666, + 0.02324533462524414, + 0.031124381348490715, + 0.013045592233538628, + 0.027830757200717926, + 0.02540675550699234, + 0.01724938303232193, + 0.05410771816968918, + 0.02831326052546501, + 0.0058977301232516766, + 0.00443796394392848, + 0.003467812202870846, + 0.007324101869016886, + 0.005541255231946707, + 0.007193263620138168, + 0.008109734393656254, + 0.005995681043714285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.20152834057807922 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tracking medium close-up shot, shot from a handheld camera and slightly tilted upward, locks on a thoroughly smudged clown, stumbling and scrambling down a dark, damp, cobblestoned back alley. His unkempt, colorful wig is tilted to one side, and his signature white greasepaint base, cracked and flaking from sweat, forms mottled patterns across his terrified face. The large, red smiley lipstick, which should have brought joy, has been roughly smeared, leaving a grotesque scar across his cheek. The ultra-high-definition detail reveals every glistening bead of sweat, seeping from his furrowed brow and temple, running down the furrows of greasepaint. His once vibrant polka-dot clown costume is now filthy and torn, and his heavy red clown shoes make a desperate, clattering sound on the slippery cobblestones. The only dim street lamp deep in the alley cast a bleak light from behind him, outlining a tense outline on his body. At the same time, his running figure was cast a huge and distorted shadow on the graffiti-covered brick wall in front of him. The whole picture was filled with an atmosphere of extreme tension, chaos and despair in the violent shaking.", + "seed": 18299, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_11/index_5054_seed_18299_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1394959533819929, + "vlm_logprob": -0.1394959533819929, + "vlm_yes_no_margin": 3.675000011920929, + "vlm_token_credit": -0.1394959533819929, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0044373623095452785 + ], + "opposite_logprob_tokens": [ + -6.254437446594238 + ], + "logprob_reward": -0.0044373623095452785, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0044373623095452785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11905136704444885, + 0.09495418518781662, + 0.059360604733228683, + 0.07165970653295517, + 0.1109306737780571, + 0.07128920406103134, + 0.12217867374420166, + 0.11558142304420471, + 0.10421064496040344, + 0.13078351318836212 + ], + "credit_map": [ + 0.018850643187761307, + 0.009947482496500015, + 0.018069235607981682, + 0.012535348534584045, + 0.011156751774251461, + 0.019087057560682297, + 0.01139169093221426, + 0.018013158813118935, + 0.015153031796216965, + 0.012795038521289825, + 0.014554355293512344, + 0.008176647126674652, + 0.009690380655229092, + 0.013265704736113548, + 0.013128773309290409, + 0.008190255612134933, + 0.010812394320964813, + 0.005898111499845982, + 0.009171605110168457, + 0.005730810575187206, + 0.005950890481472015, + 0.008772114291787148, + 0.007345954887568951, + 0.005678731016814709, + 0.00908285565674305, + 0.00636156415566802, + 0.01057863887399435, + 0.004776448011398315, + 0.010447042062878609, + 0.01330699771642685, + 0.007740974426269531, + 0.009365190751850605, + 0.012345905415713787, + 0.006545094307512045, + 0.012926141731441021, + 0.015240008942782879, + 0.01670253649353981, + 0.0062376828864216805, + 0.025027096271514893, + 0.015906214714050293, + 0.00766855850815773, + 0.007301942445337772, + 0.008717590942978859, + 0.0072017027996480465, + 0.014297161251306534, + 0.005455554462969303, + 0.01409709919244051, + 0.006549598649144173, + 0.01451217383146286, + 0.011160098016262054, + 0.018890351057052612, + 0.011702965945005417, + 0.014750372618436813, + 0.013575644232332706, + 0.025393854826688766, + 0.012193232774734497, + 0.015209059230983257, + 0.010133066214621067, + 0.017701243981719017, + 0.011612345464527607, + 0.009187810122966766, + 0.01574820652604103, + 0.01204113382846117, + 0.02394857257604599, + 0.018709426745772362, + 0.010820841416716576, + 0.017795400694012642, + 0.010632585734128952, + 0.009924734011292458, + 0.014059513807296753, + 0.014013736508786678, + 0.00825440976768732, + 0.014423416927456856, + 0.012499337084591389, + 0.011725199408829212, + 0.016222750768065453, + 0.011525745503604412, + 0.01937788911163807, + 0.02430632710456848, + 0.020702848210930824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.0044373623095452785 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34865134954452515 + ], + "opposite_logprob_tokens": [ + -1.22365140914917 + ], + "logprob_reward": -0.34865134954452515, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.34865134954452515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0621926449239254, + 0.1001000925898552, + 0.08672478049993515, + 0.1335039883852005, + 0.11986111104488373, + 0.1549384593963623, + 0.10235834866762161, + 0.09151552617549896, + 0.043094452470541, + 0.10571061819791794 + ], + "credit_map": [ + 0.008949954062700272, + 0.005279285833239555, + 0.008605209179222584, + 0.006541211158037186, + 0.008599507622420788, + 0.01134245004504919, + 0.006394628901034594, + 0.006480401381850243, + 0.016969960182905197, + 0.010775976814329624, + 0.009792418219149113, + 0.006595985032618046, + 0.017479244619607925, + 0.017918644472956657, + 0.010323576629161835, + 0.01024427730590105, + 0.013290602713823318, + 0.00875239446759224, + 0.01237998716533184, + 0.010603722184896469, + 0.008146361447870731, + 0.01152629591524601, + 0.01238282211124897, + 0.009642595425248146, + 0.022007033228874207, + 0.008780250325798988, + 0.02234756015241146, + 0.008153743110597134, + 0.012686049565672874, + 0.026121022179722786, + 0.013206562027335167, + 0.020201755687594414, + 0.013985540717840195, + 0.011766155250370502, + 0.018786951899528503, + 0.01815955713391304, + 0.02015501819550991, + 0.006801443640142679, + 0.019135747104883194, + 0.011070693843066692, + 0.013473977334797382, + 0.012526151724159718, + 0.019090399146080017, + 0.019126826897263527, + 0.021988188847899437, + 0.01616659201681614, + 0.02923973649740219, + 0.023326586931943893, + 0.011172987520694733, + 0.007290770765393972, + 0.01518833264708519, + 0.010618831031024456, + 0.01269421074539423, + 0.016799908131361008, + 0.020099515095353127, + 0.008493789471685886, + 0.011196334846317768, + 0.010890685953199863, + 0.012678587809205055, + 0.009296962060034275, + 0.0101854233071208, + 0.010681496001780033, + 0.010523839853703976, + 0.016062192618846893, + 0.00661542359739542, + 0.004644398111850023, + 0.005143395625054836, + 0.004287961404770613, + 0.005287863779813051, + 0.006812622770667076, + 0.005934383720159531, + 0.00436840346083045, + 0.009033946320414543, + 0.012871572747826576, + 0.011779666878283024, + 0.01631918177008629, + 0.008663997985422611, + 0.012747706845402718, + 0.02117978036403656, + 0.01311476156115532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.34865134954452515 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011819765204563737 + ], + "opposite_logprob_tokens": [ + -6.7511820793151855 + ], + "logprob_reward": -0.0011819765204563737, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011819765204563737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05376641824841499, + 0.0813252255320549, + 0.05633491277694702, + 0.11288686096668243, + 0.05976749211549759, + 0.12310141324996948, + 0.11688343435525894, + 0.06384246051311493, + 0.21066440641880035, + 0.12142741680145264 + ], + "credit_map": [ + 0.008873650804162025, + 0.005870239809155464, + 0.009334882721304893, + 0.006018135696649551, + 0.006637049838900566, + 0.006911792326718569, + 0.004488189239054918, + 0.005632475949823856, + 0.01417164970189333, + 0.009924707002937794, + 0.008781202137470245, + 0.006224518176168203, + 0.00759546272456646, + 0.01506553590297699, + 0.011399109847843647, + 0.008163035847246647, + 0.01046394556760788, + 0.006356206256896257, + 0.010324894450604916, + 0.005161111708730459, + 0.0042266687378287315, + 0.0067491899244487286, + 0.005943622440099716, + 0.007109273225069046, + 0.019497940316796303, + 0.006536087952554226, + 0.020797479897737503, + 0.006417061667889357, + 0.009746300987899303, + 0.021145347505807877, + 0.01146378181874752, + 0.017282862216234207, + 0.006239105947315693, + 0.004369776230305433, + 0.00712676253169775, + 0.008019676432013512, + 0.010170966386795044, + 0.0035530119203031063, + 0.009851809591054916, + 0.010436387732625008, + 0.011337203904986382, + 0.01148473471403122, + 0.01611373946070671, + 0.0111872348934412, + 0.02230742573738098, + 0.006690786685794592, + 0.03226570785045624, + 0.011714580468833447, + 0.012600780464708805, + 0.010682282038033009, + 0.017080068588256836, + 0.010860309936106205, + 0.015356539748609066, + 0.009589309804141521, + 0.028106428682804108, + 0.012607716023921967, + 0.007572363596409559, + 0.0062497700564563274, + 0.009261300787329674, + 0.004958270117640495, + 0.006223627366125584, + 0.009500164538621902, + 0.0061384886503219604, + 0.013938475400209427, + 0.06436295062303543, + 0.01728973351418972, + 0.02131018228828907, + 0.012056750245392323, + 0.019538432359695435, + 0.013795897364616394, + 0.03318796306848526, + 0.029122473672032356, + 0.013550695963203907, + 0.009628030471503735, + 0.008602752350270748, + 0.009325644932687283, + 0.016187703236937523, + 0.023029496893286705, + 0.01731172390282154, + 0.023791363462805748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.0011819765204563737 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980329655110836 + ], + "opposite_logprob_tokens": [ + -3.5298032760620117 + ], + "logprob_reward": -0.02980329655110836, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980329655110836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07133684307336807, + 0.054000306874513626, + 0.043750371783971786, + 0.1399238556623459, + 0.0417051687836647, + 0.04697348177433014, + 0.10421901196241379, + 0.09208665043115616, + 0.23608633875846863, + 0.16991795599460602 + ], + "credit_map": [ + 0.011030606925487518, + 0.005913263652473688, + 0.008972943760454655, + 0.009228593669831753, + 0.005979262292385101, + 0.01250644400715828, + 0.007624617777764797, + 0.010081115178763866, + 0.008745084516704082, + 0.0072295027785003185, + 0.007323442958295345, + 0.004986623302102089, + 0.005630779080092907, + 0.007006851490586996, + 0.0068650804460048676, + 0.006212945096194744, + 0.00852567981928587, + 0.006204533856362104, + 0.007505679503083229, + 0.004385245498269796, + 0.004631049931049347, + 0.004584642127156258, + 0.003973119892179966, + 0.003940420690923929, + 0.020472247153520584, + 0.009286514483392239, + 0.021837536245584488, + 0.008346989750862122, + 0.013812128454446793, + 0.02973603457212448, + 0.013650557957589626, + 0.022781843319535255, + 0.004679641220718622, + 0.0030362927354872227, + 0.005816632881760597, + 0.005863340571522713, + 0.0070516797713935375, + 0.0021919517312198877, + 0.006951950956135988, + 0.006113674491643906, + 0.005222988314926624, + 0.004078614991158247, + 0.006614845246076584, + 0.0051197391003370285, + 0.006531199440360069, + 0.003019454190507531, + 0.011327170766890049, + 0.005059466231614351, + 0.012282869778573513, + 0.008928793482482433, + 0.015213962644338608, + 0.009266304783523083, + 0.013827254064381123, + 0.010622058063745499, + 0.023417271673679352, + 0.010660501196980476, + 0.012086471542716026, + 0.007572099566459656, + 0.013495207764208317, + 0.008700449019670486, + 0.00969127006828785, + 0.013680826872587204, + 0.00910576805472374, + 0.017754556611180305, + 0.06680016219615936, + 0.021366938948631287, + 0.02354317530989647, + 0.015945348888635635, + 0.02445157989859581, + 0.017146633937954903, + 0.035102758556604385, + 0.03172975406050682, + 0.018347376957535744, + 0.015069256536662579, + 0.012772179208695889, + 0.014153174124658108, + 0.022793730720877647, + 0.03322356939315796, + 0.02298731356859207, + 0.030571363866329193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.02980329655110836 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134057819843292 + ], + "opposite_logprob_tokens": [ + -1.3134057521820068 + ], + "logprob_reward": -0.3134057819843292, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134057819843292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07668530195951462, + 0.0578615628182888, + 0.06074533611536026, + 0.1180504783987999, + 0.04977838695049286, + 0.10305634886026382, + 0.06221123784780502, + 0.11821509897708893, + 0.1965465396642685, + 0.1568496823310852 + ], + "credit_map": [ + 0.011968551203608513, + 0.007446094881743193, + 0.012739801779389381, + 0.007224082015454769, + 0.008570090867578983, + 0.013699186034500599, + 0.00615295534953475, + 0.008884548209607601, + 0.009966114535927773, + 0.005262152757495642, + 0.009475037455558777, + 0.005012362729758024, + 0.005618290044367313, + 0.00788835994899273, + 0.00777770671993494, + 0.006861544214189053, + 0.01071331836283207, + 0.007334697060286999, + 0.01170295849442482, + 0.0057580191642045975, + 0.004339409526437521, + 0.00757055776193738, + 0.007491725496947765, + 0.005834655370563269, + 0.019867371767759323, + 0.007953682914376259, + 0.02030542679131031, + 0.006423479411751032, + 0.01235068216919899, + 0.02449899911880493, + 0.009050332941114902, + 0.017600519582629204, + 0.0055507393553853035, + 0.0041084298864007, + 0.006906405091285706, + 0.006391176022589207, + 0.007349444553256035, + 0.005406087730079889, + 0.008132616989314556, + 0.005933492910116911, + 0.010287264361977577, + 0.006176401861011982, + 0.01515939086675644, + 0.012899012304842472, + 0.014547407627105713, + 0.005887198727577925, + 0.027934005483984947, + 0.010165677405893803, + 0.00999391172081232, + 0.005019373260438442, + 0.009874297305941582, + 0.005502557847648859, + 0.0073078349232673645, + 0.007147606927901506, + 0.009588348679244518, + 0.007777316030114889, + 0.013402708806097507, + 0.01408681832253933, + 0.016678759828209877, + 0.009404448792338371, + 0.010874680243432522, + 0.015509987249970436, + 0.013588926754891872, + 0.024668779224157333, + 0.04739450663328171, + 0.012094207108020782, + 0.016582569107413292, + 0.008100957609713078, + 0.01849154755473137, + 0.03654104098677635, + 0.02672736532986164, + 0.030614357441663742, + 0.017585717141628265, + 0.01214778982102871, + 0.01108654122799635, + 0.013839873485267162, + 0.021465204656124115, + 0.028941186144948006, + 0.02941090613603592, + 0.022372476756572723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.3134057819843292 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tracking medium close-up shot, shot from a handheld camera and slightly tilted upward, locks on a thoroughly smudged clown, stumbling and scrambling down a dark, damp, cobblestoned back alley. His unkempt, colorful wig is tilted to one side, and his signature white greasepaint base, cracked and flaking from sweat, forms mottled patterns across his terrified face. The large, red smiley lipstick, which should have brought joy, has been roughly smeared, leaving a grotesque scar across his cheek. The ultra-high-definition detail reveals every glistening bead of sweat, seeping from his furrowed brow and temple, running down the furrows of greasepaint. His once vibrant polka-dot clown costume is now filthy and torn, and his heavy red clown shoes make a desperate, clattering sound on the slippery cobblestones. The only dim street lamp deep in the alley cast a bleak light from behind him, outlining a tense outline on his body. At the same time, his running figure was cast a huge and distorted shadow on the graffiti-covered brick wall in front of him. The whole picture was filled with an atmosphere of extreme tension, chaos and despair in the violent shaking.", + "seed": 18300, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_11/index_5054_seed_18300_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11276341998018324, + "vlm_logprob": -0.11276341998018324, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.11276341998018324, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007854292169213295 + ], + "opposite_logprob_tokens": [ + -5.757854461669922 + ], + "logprob_reward": -0.007854292169213295, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007854292169213295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14996176958084106, + 0.0713777020573616, + 0.06670855730772018, + 0.06554397195577621, + 0.11237689852714539, + 0.1082175225019455, + 0.20381516218185425, + 0.03689657524228096, + 0.1159338653087616, + 0.06916792690753937 + ], + "credit_map": [ + 0.01940908469259739, + 0.015123806893825531, + 0.01923484541475773, + 0.020397204905748367, + 0.021668901666998863, + 0.020594274625182152, + 0.017611198127269745, + 0.01592245139181614, + 0.008174093440175056, + 0.008524540811777115, + 0.00898002926260233, + 0.007178218569606543, + 0.007498219143599272, + 0.01284999493509531, + 0.007217725273221731, + 0.010954881086945534, + 0.0056855701841413975, + 0.005565432365983725, + 0.009962750598788261, + 0.004556022118777037, + 0.011438361369073391, + 0.01277169119566679, + 0.010427895002067089, + 0.006300830282270908, + 0.004810761660337448, + 0.0059747169725596905, + 0.008758635260164738, + 0.006558435503393412, + 0.011784986592829227, + 0.012274914421141148, + 0.006832498591393232, + 0.008549020625650883, + 0.007451874669641256, + 0.008631488308310509, + 0.01190277747809887, + 0.01706547476351261, + 0.019680406898260117, + 0.014985717833042145, + 0.018773676827549934, + 0.01388548407703638, + 0.011320393532514572, + 0.009151670150458813, + 0.007969709113240242, + 0.008244304917752743, + 0.0202008243650198, + 0.019832806661725044, + 0.016217796131968498, + 0.015280014835298061, + 0.011375955305993557, + 0.029308082535862923, + 0.015401529148221016, + 0.02424432709813118, + 0.025967471301555634, + 0.027180708944797516, + 0.04316308721899986, + 0.027174001559615135, + 0.003985670395195484, + 0.003761610481888056, + 0.004829891491681337, + 0.0036967070773243904, + 0.005523476749658585, + 0.0038323604967445135, + 0.006291077937930822, + 0.004975779447704554, + 0.01095168199390173, + 0.0222676619887352, + 0.015272291377186775, + 0.016932882368564606, + 0.01454848051071167, + 0.009849132969975471, + 0.013637431897222996, + 0.012474307790398598, + 0.006125509273260832, + 0.011793019250035286, + 0.007416705600917339, + 0.006266733165830374, + 0.00757996691390872, + 0.008168669417500496, + 0.012028115801513195, + 0.009789209812879562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.007854292169213295 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038018465042114 + ], + "opposite_logprob_tokens": [ + -1.9103801250457764 + ], + "logprob_reward": -0.16038018465042114, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038018465042114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1952355057001114, + 0.09061408787965775, + 0.06753387302160263, + 0.1546933352947235, + 0.04993006959557533, + 0.21203374862670898, + 0.10133875161409378, + 0.035992901772260666, + 0.05177725479006767, + 0.04085046425461769 + ], + "credit_map": [ + 0.03511533513665199, + 0.017025377601385117, + 0.03267382085323334, + 0.019394712522625923, + 0.019345320761203766, + 0.03757382929325104, + 0.013899578712880611, + 0.020207546651363373, + 0.009130601771175861, + 0.0074081323109567165, + 0.013424803502857685, + 0.009791652671992779, + 0.007095037493854761, + 0.020991461351513863, + 0.008633658289909363, + 0.014138739556074142, + 0.004947681445628405, + 0.010373580269515514, + 0.009436496533453465, + 0.006246338132768869, + 0.008448508568108082, + 0.0123598612844944, + 0.008806530386209488, + 0.006914874073117971, + 0.008686457760632038, + 0.008819881826639175, + 0.017876479774713516, + 0.010310238227248192, + 0.033874887973070145, + 0.045072928071022034, + 0.012584532611072063, + 0.017467930912971497, + 0.004583335481584072, + 0.0034430879168212414, + 0.004875263199210167, + 0.005601613782346249, + 0.007984126918017864, + 0.006993206217885017, + 0.010526066645979881, + 0.005923372693359852, + 0.023252099752426147, + 0.016491388902068138, + 0.020226459950208664, + 0.02811095118522644, + 0.02533419243991375, + 0.03912412002682686, + 0.03000851534307003, + 0.029486017301678658, + 0.009374267421662807, + 0.013171632774174213, + 0.010128743015229702, + 0.008912518620491028, + 0.01022160891443491, + 0.014287213794887066, + 0.01970634050667286, + 0.015536422841250896, + 0.005052644293755293, + 0.0034072594717144966, + 0.00403728848323226, + 0.002968916902318597, + 0.005779577419161797, + 0.005335464142262936, + 0.004297935403883457, + 0.005113816820085049, + 0.006505880039185286, + 0.006778029724955559, + 0.005464417394250631, + 0.004784166347235441, + 0.0055626993998885155, + 0.010491438210010529, + 0.006344337947666645, + 0.005846289452165365, + 0.003415787825360894, + 0.005267437547445297, + 0.004446298815310001, + 0.004502738825976849, + 0.0047368197701871395, + 0.006206486374139786, + 0.006502665579319, + 0.005772230681031942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.16038018465042114 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009247555863112211 + ], + "opposite_logprob_tokens": [ + -7.000924587249756 + ], + "logprob_reward": -0.0009247555863112211, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009247555863112211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.136875718832016, + 0.05347064882516861, + 0.09764585644006729, + 0.1375090330839157, + 0.04810909181833267, + 0.17895309627056122, + 0.14677593111991882, + 0.045649655163288116, + 0.04281292110681534, + 0.11219799518585205 + ], + "credit_map": [ + 0.02544090710580349, + 0.013327322900295258, + 0.018173301592469215, + 0.016636215150356293, + 0.016378046944737434, + 0.01927895098924637, + 0.011817490682005882, + 0.0158234890550375, + 0.005323758348822594, + 0.0043950811959803104, + 0.007498272228986025, + 0.004096640273928642, + 0.003786214627325535, + 0.014521127566695213, + 0.0046131969429552555, + 0.009236357174813747, + 0.007910115644335747, + 0.0071179517544806, + 0.014442529529333115, + 0.005923612508922815, + 0.01218465156853199, + 0.026652202010154724, + 0.013119002804160118, + 0.010295797139406204, + 0.00748487701639533, + 0.008000604808330536, + 0.01889272965490818, + 0.010586339980363846, + 0.028412731364369392, + 0.03402555361390114, + 0.010879972949624062, + 0.019226225093007088, + 0.0037298998795449734, + 0.00431554950773716, + 0.005447409115731716, + 0.006898032501339912, + 0.007797747850418091, + 0.006197013892233372, + 0.007746738847345114, + 0.0059766992926597595, + 0.02146243490278721, + 0.011977532878518105, + 0.0181887187063694, + 0.019430385902523994, + 0.01738908514380455, + 0.04563124105334282, + 0.017854010686278343, + 0.0270196832716465, + 0.010363019071519375, + 0.02170047163963318, + 0.013927952386438847, + 0.012562272138893604, + 0.019841555505990982, + 0.019740844145417213, + 0.024663975462317467, + 0.02397584170103073, + 0.005643145646899939, + 0.004394878633320332, + 0.007378706242889166, + 0.004870374221354723, + 0.005684380419552326, + 0.0049108960665762424, + 0.005924660712480545, + 0.0068426174111664295, + 0.0056039500050246716, + 0.006808041594922543, + 0.00587418070062995, + 0.004205347038805485, + 0.005136159714311361, + 0.004008523654192686, + 0.006055988371372223, + 0.005120725370943546, + 0.00928101409226656, + 0.018749617040157318, + 0.011216003447771072, + 0.011166399344801903, + 0.012254500761628151, + 0.0133772362023592, + 0.01891898363828659, + 0.01723424158990383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.0009247555863112211 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763084227219224 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.00763084227219224, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763084227219224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11238143593072891, + 0.060064516961574554, + 0.1385205090045929, + 0.10782337188720703, + 0.05135791748762131, + 0.1404583752155304, + 0.11091796308755875, + 0.17228063941001892, + 0.05496012046933174, + 0.051235131919384 + ], + "credit_map": [ + 0.0146072618663311, + 0.01026797667145729, + 0.015545167960226536, + 0.0167876947671175, + 0.014331192709505558, + 0.01768314465880394, + 0.011384857818484306, + 0.011774136684834957, + 0.005868141073733568, + 0.005667244084179401, + 0.008352749049663544, + 0.006396578624844551, + 0.005995703395456076, + 0.012737469747662544, + 0.0066574034281075, + 0.00838923268020153, + 0.012503477744758129, + 0.015398599207401276, + 0.016327332705259323, + 0.01006886176764965, + 0.02086382731795311, + 0.028364745900034904, + 0.022294990718364716, + 0.012698673643171787, + 0.008145402185618877, + 0.008480113931000233, + 0.018581530079245567, + 0.012748954817652702, + 0.01277803722769022, + 0.020576953887939453, + 0.00988494697958231, + 0.016627434641122818, + 0.004852819722145796, + 0.004983813501894474, + 0.005248480476438999, + 0.0066446904093027115, + 0.006632868200540543, + 0.004680119454860687, + 0.012216485105454922, + 0.0060986424796283245, + 0.018259825184941292, + 0.009199864231050014, + 0.01349786389619112, + 0.012124619446694851, + 0.021567653864622116, + 0.022653035819530487, + 0.021250203251838684, + 0.02190531976521015, + 0.007843723520636559, + 0.008759201504290104, + 0.015409154817461967, + 0.01462297048419714, + 0.011530982330441475, + 0.020289096981287003, + 0.015132551081478596, + 0.01733027771115303, + 0.02130264602601528, + 0.014786065556108952, + 0.02918875403702259, + 0.01836438477039337, + 0.022333117201924324, + 0.017277050763368607, + 0.02396673522889614, + 0.025061890482902527, + 0.007733575999736786, + 0.010285592637956142, + 0.0059476071037352085, + 0.005288489628583193, + 0.005952616222202778, + 0.006664742715656757, + 0.006915634498000145, + 0.006171858869493008, + 0.004153431858867407, + 0.006891475059092045, + 0.005727527663111687, + 0.0040484461933374405, + 0.006346493028104305, + 0.006853602360934019, + 0.010850786231458187, + 0.006363370455801487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.00763084227219224 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870270252227783 + ], + "opposite_logprob_tokens": [ + -1.1370270252227783 + ], + "logprob_reward": -0.3870270252227783, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870270252227783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13380885124206543, + 0.18927520513534546, + 0.1693742722272873, + 0.0908445492386818, + 0.13971757888793945, + 0.0868331789970398, + 0.0726088210940361, + 0.058118052780628204, + 0.0305499155074358, + 0.028869541361927986 + ], + "credit_map": [ + 0.026862427592277527, + 0.012445775792002678, + 0.019148660823702812, + 0.015028414316475391, + 0.013991755433380604, + 0.018383627757430077, + 0.013045968487858772, + 0.014902234077453613, + 0.019034165889024734, + 0.012924021109938622, + 0.028432831168174744, + 0.01150075439363718, + 0.011000314727425575, + 0.058012597262859344, + 0.011877217330038548, + 0.036493320018053055, + 0.014611538499593735, + 0.015061551705002785, + 0.021815946325659752, + 0.011961722746491432, + 0.02663256973028183, + 0.02025071159005165, + 0.029738588258624077, + 0.029301650822162628, + 0.007547243032604456, + 0.005472400225698948, + 0.01770256832242012, + 0.01228129118680954, + 0.00898752547800541, + 0.01057092659175396, + 0.01003504078835249, + 0.018247567117214203, + 0.009936593472957611, + 0.009480796754360199, + 0.013625223189592361, + 0.02095211297273636, + 0.02529243938624859, + 0.016784770414233208, + 0.02675445005297661, + 0.016891194507479668, + 0.008307337760925293, + 0.006043443456292152, + 0.0073121520690619946, + 0.008884287439286709, + 0.017823223024606705, + 0.00910259410738945, + 0.017598090693354607, + 0.01176205649971962, + 0.005369700491428375, + 0.011825038120150566, + 0.0072048986330628395, + 0.008017046377062798, + 0.008863521739840508, + 0.01076368149369955, + 0.010474493727087975, + 0.010090441443026066, + 0.008245236240327358, + 0.00474851718172431, + 0.009725501760840416, + 0.005341022275388241, + 0.006512041669338942, + 0.0075218859128654, + 0.0063471486791968346, + 0.009676704183220863, + 0.003349670907482505, + 0.00506034167483449, + 0.0034992846194654703, + 0.003328308230265975, + 0.0028830606024712324, + 0.004549853038042784, + 0.003999804612249136, + 0.0038795964792370796, + 0.0022259815596044064, + 0.0030825159046798944, + 0.0041073220781981945, + 0.0034159186761826277, + 0.0030162972398102283, + 0.005025334190577269, + 0.004002607427537441, + 0.003993567079305649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.3870270252227783 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tracking medium close-up shot, shot from a handheld camera and slightly tilted upward, locks on a thoroughly smudged clown, stumbling and scrambling down a dark, damp, cobblestoned back alley. His unkempt, colorful wig is tilted to one side, and his signature white greasepaint base, cracked and flaking from sweat, forms mottled patterns across his terrified face. The large, red smiley lipstick, which should have brought joy, has been roughly smeared, leaving a grotesque scar across his cheek. The ultra-high-definition detail reveals every glistening bead of sweat, seeping from his furrowed brow and temple, running down the furrows of greasepaint. His once vibrant polka-dot clown costume is now filthy and torn, and his heavy red clown shoes make a desperate, clattering sound on the slippery cobblestones. The only dim street lamp deep in the alley cast a bleak light from behind him, outlining a tense outline on his body. At the same time, his running figure was cast a huge and distorted shadow on the graffiti-covered brick wall in front of him. The whole picture was filled with an atmosphere of extreme tension, chaos and despair in the violent shaking.", + "seed": 18301, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_11/index_5054_seed_18301_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.047519997344352305, + "vlm_logprob": -0.047519997344352305, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.047519997344352305, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007245099637657404 + ], + "opposite_logprob_tokens": [ + -5.632245063781738 + ], + "logprob_reward": -0.007245099637657404, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.007245099637657404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.185513436794281, + 0.11523871123790741, + 0.04940236359834671, + 0.2240961194038391, + 0.17234353721141815, + 0.07572748512029648, + 0.06548536568880081, + 0.03583996370434761, + 0.02550015039741993, + 0.05085284635424614 + ], + "credit_map": [ + 0.01454271748661995, + 0.02156244032084942, + 0.012566514313220978, + 0.03294149041175842, + 0.02034786157310009, + 0.03265026584267616, + 0.026644030585885048, + 0.024258112534880638, + 0.0063747684471309185, + 0.007295438554137945, + 0.005323401186615229, + 0.011329947039484978, + 0.014496286399662495, + 0.026199426501989365, + 0.018527928739786148, + 0.02569151483476162, + 0.0028831083327531815, + 0.003949239384382963, + 0.0016312823863700032, + 0.006710018031299114, + 0.007024758495390415, + 0.009094731882214546, + 0.010649469681084156, + 0.0074597569182515144, + 0.017050359398126602, + 0.028851279988884926, + 0.015944020822644234, + 0.02994348481297493, + 0.04288459196686745, + 0.03454761207103729, + 0.02448623813688755, + 0.030388526618480682, + 0.005689905025064945, + 0.015222764573991299, + 0.010666567832231522, + 0.016785085201263428, + 0.011552364565432072, + 0.05180976167321205, + 0.017973177134990692, + 0.042643919587135315, + 0.005992145277559757, + 0.010724475607275963, + 0.011470828205347061, + 0.009254734963178635, + 0.010316566564142704, + 0.009599740616977215, + 0.009325463324785233, + 0.009043531492352486, + 0.007644376251846552, + 0.007888206280767918, + 0.006447236053645611, + 0.00919791404157877, + 0.00994605291634798, + 0.00812690332531929, + 0.009291199035942554, + 0.0069434791803359985, + 0.0053886487148702145, + 0.004080092068761587, + 0.004913243465125561, + 0.004490569233894348, + 0.005924729164689779, + 0.004113554488867521, + 0.003972441423684359, + 0.0029566839803010225, + 0.003992011304944754, + 0.003145517548546195, + 0.004657191690057516, + 0.00395836029201746, + 0.0026935848873108625, + 0.0021077031269669533, + 0.00262798392213881, + 0.002317796228453517, + 0.01000664196908474, + 0.006177538074553013, + 0.01171929482370615, + 0.004082685336470604, + 0.007352454587817192, + 0.0027841839473694563, + 0.004223350435495377, + 0.004506698343902826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.007245099637657404 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294938623905182 + ], + "opposite_logprob_tokens": [ + -2.2379493713378906 + ], + "logprob_reward": -0.11294938623905182, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294938623905182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28155219554901123, + 0.31732428073883057, + 0.04580904170870781, + 0.035644568502902985, + 0.12194570899009705, + 0.08011295646429062, + 0.048138462007045746, + 0.02999962866306305, + 0.021670900285243988, + 0.01780225709080696 + ], + "credit_map": [ + 0.01348920725286007, + 0.03930509835481644, + 0.010821977630257607, + 0.025272028520703316, + 0.03808300569653511, + 0.0616375133395195, + 0.0445130355656147, + 0.04843030497431755, + 0.007186803501099348, + 0.008236194960772991, + 0.007974154315888882, + 0.012465016916394234, + 0.016983555629849434, + 0.12717081606388092, + 0.020477326586842537, + 0.11683040112257004, + 0.0036777714267373085, + 0.005417912732809782, + 0.002446298487484455, + 0.005450869910418987, + 0.007135788444429636, + 0.009387602098286152, + 0.006739797070622444, + 0.005552996415644884, + 0.0031044401694089174, + 0.004526861477643251, + 0.0030406061559915543, + 0.004206190817058086, + 0.00754921929910779, + 0.00456845061853528, + 0.004066053777933121, + 0.004582744091749191, + 0.005944283679127693, + 0.01883370243012905, + 0.010501569136977196, + 0.017444485798478127, + 0.010049867443740368, + 0.024649979546666145, + 0.013337763026356697, + 0.021184047684073448, + 0.007226245477795601, + 0.011672823689877987, + 0.00941409356892109, + 0.010341235436499119, + 0.010460958816111088, + 0.010537133552134037, + 0.012749363668262959, + 0.007711092941462994, + 0.006605200469493866, + 0.006416454445570707, + 0.006080695893615484, + 0.005172024946659803, + 0.004307396709918976, + 0.009394134394824505, + 0.00523121515288949, + 0.0049313330091536045, + 0.004635298624634743, + 0.004156487062573433, + 0.0028960099443793297, + 0.003735958132892847, + 0.00411213980987668, + 0.0034363160375505686, + 0.004287545569241047, + 0.0027398704551160336, + 0.005820184014737606, + 0.002135951304808259, + 0.0021403287537395954, + 0.002482707379385829, + 0.0021872695069760084, + 0.0014992435462772846, + 0.0017194612883031368, + 0.0036857526283711195, + 0.003541212994605303, + 0.0023322124034166336, + 0.0036820650566369295, + 0.0020164111629128456, + 0.0024058606941252947, + 0.0008000866509974003, + 0.0017210704972967505, + 0.0013033360010012984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.11294938623905182 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010457528987899423 + ], + "opposite_logprob_tokens": [ + -6.8760457038879395 + ], + "logprob_reward": -0.0010457528987899423, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010457528987899423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2172859013080597, + 0.08769424259662628, + 0.0547032430768013, + 0.2756008803844452, + 0.13763922452926636, + 0.06156007573008537, + 0.05035196617245674, + 0.05788800120353699, + 0.029189487919211388, + 0.028087018057703972 + ], + "credit_map": [ + 0.013684757985174656, + 0.02011575736105442, + 0.011394134722650051, + 0.030178960412740707, + 0.031308915466070175, + 0.04190874099731445, + 0.036149345338344574, + 0.032545264810323715, + 0.0030605588108301163, + 0.007011410780251026, + 0.0025344255845993757, + 0.009908891282975674, + 0.006959453225135803, + 0.026368828490376472, + 0.013020767830312252, + 0.018829908221960068, + 0.004124662838876247, + 0.004213578999042511, + 0.0028225111309438944, + 0.007723313756287098, + 0.007113352883607149, + 0.01118386723101139, + 0.010588128119707108, + 0.006933826021850109, + 0.021457519382238388, + 0.037762485444545746, + 0.01979709416627884, + 0.03440428897738457, + 0.05623043701052666, + 0.03961025923490524, + 0.02900872379541397, + 0.03733004629611969, + 0.004563306458294392, + 0.007432546000927687, + 0.005025456193834543, + 0.008137566968798637, + 0.006232020445168018, + 0.05345049127936363, + 0.00847369059920311, + 0.04432413727045059, + 0.005182916298508644, + 0.009601983241736889, + 0.005523141473531723, + 0.007714907173067331, + 0.008351925760507584, + 0.0078136520460248, + 0.010122306644916534, + 0.0072492389008402824, + 0.006642803084105253, + 0.00494631752371788, + 0.007342921569943428, + 0.008531401865184307, + 0.006688420195132494, + 0.004266645759344101, + 0.006294704042375088, + 0.005638753063976765, + 0.008649198338389397, + 0.009576023556292057, + 0.007088474929332733, + 0.006553384941071272, + 0.006388317793607712, + 0.00801691971719265, + 0.006464902311563492, + 0.00515077356249094, + 0.0040833791717886925, + 0.00323526025749743, + 0.005266750697046518, + 0.004409574903547764, + 0.003660692600533366, + 0.002388825174421072, + 0.0034076988231390715, + 0.002737304661422968, + 0.005852429196238518, + 0.0032879456412047148, + 0.005540060345083475, + 0.0027091808151453733, + 0.0037323152646422386, + 0.001583936158567667, + 0.0027580263558775187, + 0.002623119857162237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.0010457528987899423 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070127487182617 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.016070127487182617, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070127487182617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15624506771564484, + 0.13134655356407166, + 0.034509629011154175, + 0.4088434875011444, + 0.06077444553375244, + 0.04968024045228958, + 0.06088976562023163, + 0.045700713992118835, + 0.024623991921544075, + 0.027386100962758064 + ], + "credit_map": [ + 0.009575086645781994, + 0.01955503039062023, + 0.008223138749599457, + 0.02116156741976738, + 0.01640455052256584, + 0.033751122653484344, + 0.022781966254115105, + 0.024792609736323357, + 0.00368644786067307, + 0.0070321084931492805, + 0.004319057334214449, + 0.009637897834181786, + 0.007510250434279442, + 0.04694383218884468, + 0.012061216868460178, + 0.04015574976801872, + 0.0023232935927808285, + 0.002541028894484043, + 0.0019789983052760363, + 0.005013772286474705, + 0.004271264653652906, + 0.007439074572175741, + 0.006367678288370371, + 0.004574516788125038, + 0.030437540262937546, + 0.05835196003317833, + 0.03046587109565735, + 0.05106110870838165, + 0.08278382569551468, + 0.058280494064092636, + 0.04301344230771065, + 0.05444922670722008, + 0.002631731331348419, + 0.003471750533208251, + 0.002629190683364868, + 0.00505303917452693, + 0.003163969377055764, + 0.021199872717261314, + 0.00408507464453578, + 0.01853981241583824, + 0.005590913817286491, + 0.007039549760520458, + 0.004810363985598087, + 0.006173755973577499, + 0.005931973457336426, + 0.007629296276718378, + 0.008069987408816814, + 0.0044344025664031506, + 0.008305301889777184, + 0.005957228597253561, + 0.009294851683080196, + 0.010951897129416466, + 0.005845393519848585, + 0.009699145331978798, + 0.0048183961771428585, + 0.006017552688717842, + 0.0073064398020505905, + 0.0062193796038627625, + 0.0045600817538797855, + 0.004982271231710911, + 0.00526793859899044, + 0.00774531252682209, + 0.005201731808483601, + 0.00441755773499608, + 0.0052740806713700294, + 0.0024543581530451775, + 0.0034643046092242002, + 0.003250202862545848, + 0.0029981157276779413, + 0.0012596732703968883, + 0.002423158846795559, + 0.0035000985953956842, + 0.004707569722086191, + 0.0043301573023200035, + 0.006011696998029947, + 0.0028589582070708275, + 0.0030121193267405033, + 0.001271469984203577, + 0.0029321236070245504, + 0.002262005116790533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.016070127487182617 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a smudged clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera handheld, slightly tilted upward, and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a tilted, unkempt colorful wig with cracked white greasepaint and smeared red lipstick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown wearing a filthy, torn polka-dot costume and heavy red clown shoes while stumbling down a dark, damp cobblestoned alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dim street lamp behind the clown casting a huge, distorted shadow of his running figure onto a graffiti-covered brick wall in front?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028962045907974 + ], + "opposite_logprob_tokens": [ + -2.3502895832061768 + ], + "logprob_reward": -0.10028962045907974, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028962045907974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4253154695034027, + 0.11684348434209824, + 0.07685693353414536, + 0.09652213752269745, + 0.03836648911237717, + 0.11389465630054474, + 0.0628599002957344, + 0.021696927025914192, + 0.026415228843688965, + 0.021228762343525887 + ], + "credit_map": [ + 0.023922143504023552, + 0.053181011229753494, + 0.02925894968211651, + 0.05005364865064621, + 0.058921899646520615, + 0.0674431324005127, + 0.07745051383972168, + 0.06508421152830124, + 0.00630164472386241, + 0.013687483966350555, + 0.005575607065111399, + 0.02685258910059929, + 0.014294581487774849, + 0.016331030055880547, + 0.017920978367328644, + 0.015879571437835693, + 0.0047411187551915646, + 0.005589121021330357, + 0.004220372997224331, + 0.011600527912378311, + 0.01141057163476944, + 0.014173070900142193, + 0.01721113920211792, + 0.007911019958555698, + 0.01328358706086874, + 0.012151331640779972, + 0.01070321910083294, + 0.010035321116447449, + 0.01735910400748253, + 0.012378829531371593, + 0.010022149421274662, + 0.010588597506284714, + 0.0028374840039759874, + 0.005604126024991274, + 0.004135041031986475, + 0.004965145140886307, + 0.00420246971771121, + 0.006253721192479134, + 0.005127478390932083, + 0.005241023376584053, + 0.011571934446692467, + 0.022830016911029816, + 0.0172552652657032, + 0.014173194766044617, + 0.012340513058006763, + 0.013844071887433529, + 0.01093357615172863, + 0.010946092195808887, + 0.008905651047825813, + 0.006215634290128946, + 0.00929967500269413, + 0.006943429354578257, + 0.007443856447935104, + 0.010064459405839443, + 0.007037115283310413, + 0.006950080394744873, + 0.0030460094567388296, + 0.001964378636330366, + 0.0029102626722306013, + 0.002893742173910141, + 0.003999799955636263, + 0.0026024947874248028, + 0.0025709685869514942, + 0.001709271571598947, + 0.005169897805899382, + 0.003572296816855669, + 0.0027482362929731607, + 0.004687339998781681, + 0.002789345569908619, + 0.00150022201705724, + 0.0023628966882824898, + 0.0035849944688379765, + 0.0036422910634428263, + 0.0021749837324023247, + 0.0035610958002507687, + 0.002644446212798357, + 0.0034250780008733273, + 0.0016851993277668953, + 0.00204623956233263, + 0.002049428876489401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.10028962045907974 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_12/index_tensor([4467, 4467])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_12/index_tensor([4467, 4467])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2520456d00edf3ca4fa017213d1a6d26ccfb765a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_12/index_tensor([4467, 4467])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "perfectly static, extreme macro shot captures the moment a viscous drop of vibrant sapphire detergent pierces the pristine, crystal-clear surface of water within a glass container. For a fleeting moment, the drop maintains its spherical shape before beginning to dissolve. As the camera slowly and deliberately zooms out, a concentrated blue gracefully unfolds downward. It blooms into an ethereal, smoke-like plume that swirls and spirals hypnotically, creating a mesmerizing hydrodynamic ballet. Backlit by a clean, bright light source, the intricate tendrils of color become translucent, with luminous edges and deep, shadowy cores, showcasing a rich gradation from deep cobalt to pale, transparent azure. Set against a soft, out-of-focus neutral background, the entire scene evokes an atmosphere of serene, captivating, and almost meditative abstract beauty.", + "seed": 17951, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_12/index_4467_seed_17951_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.20397850424051284, + "vlm_logprob": -0.20397850424051284, + "vlm_yes_no_margin": 3.6250000059604646, + "vlm_token_credit": -0.20397850424051284, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011137224733829498 + ], + "opposite_logprob_tokens": [ + -4.636137008666992 + ], + "logprob_reward": -0.011137224733829498, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.011137224733829498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16118305921554565, + 0.10014253854751587, + 0.04503238573670387, + 0.07697034627199173, + 0.1705579310655594, + 0.2557339668273926, + 0.04978181794285774, + 0.04357403516769409, + 0.06776256114244461, + 0.02926134131848812 + ], + "credit_map": [ + 0.012240789830684662, + 0.013526310212910175, + 0.02031830884516239, + 0.018846137449145317, + 0.02580876834690571, + 0.019810952246189117, + 0.03083096072077751, + 0.019800826907157898, + 0.011468952521681786, + 0.016923323273658752, + 0.010704712942242622, + 0.018921229988336563, + 0.011953265406191349, + 0.008517810143530369, + 0.00988941453397274, + 0.011763833463191986, + 0.005620615091174841, + 0.007509178481996059, + 0.004948103800415993, + 0.005162248853594065, + 0.004070058465003967, + 0.004260975867509842, + 0.006029399111866951, + 0.007431805599480867, + 0.010629003867506981, + 0.008205389603972435, + 0.010860017500817776, + 0.007648041937500238, + 0.006891731638461351, + 0.00579428393393755, + 0.006702008657157421, + 0.02023986726999283, + 0.022581996396183968, + 0.01876208744943142, + 0.018110020086169243, + 0.017605559900403023, + 0.020702145993709564, + 0.01473291777074337, + 0.02612379938364029, + 0.031939391046762466, + 0.023027943447232246, + 0.025106890127062798, + 0.03508331626653671, + 0.05695401877164841, + 0.023365814238786697, + 0.021427327767014503, + 0.03762160614132881, + 0.033147044479846954, + 0.006445218343287706, + 0.006301954388618469, + 0.005755403079092503, + 0.00637496542185545, + 0.005855840165168047, + 0.006341164465993643, + 0.00595616502687335, + 0.006751111708581448, + 0.0060502029955387115, + 0.004293172154575586, + 0.006952532101422548, + 0.005330568179488182, + 0.006381471175700426, + 0.003122584894299507, + 0.007539159152656794, + 0.0039043452125042677, + 0.006689665839076042, + 0.008112267591059208, + 0.00775372888892889, + 0.009333230555057526, + 0.010927137918770313, + 0.0058203646913170815, + 0.009199203923344612, + 0.009926961734890938, + 0.0028431364335119724, + 0.0035530549939721823, + 0.003525987733155489, + 0.0035020452924072742, + 0.003254213836044073, + 0.0029822292272001505, + 0.0040153502486646175, + 0.005585324019193649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.011137224733829498 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789316914975643 + ], + "opposite_logprob_tokens": [ + -4.63478946685791 + ], + "logprob_reward": -0.009789316914975643, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789316914975643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27906760573387146, + 0.14120255410671234, + 0.09092050790786743, + 0.05509748309850693, + 0.08518385887145996, + 0.09797962009906769, + 0.05347466841340065, + 0.05678712949156761, + 0.0823812484741211, + 0.05790529400110245 + ], + "credit_map": [ + 0.022458944469690323, + 0.02413550764322281, + 0.0445927157998085, + 0.04428137093782425, + 0.033465441316366196, + 0.02544894441962242, + 0.05892479792237282, + 0.025759926065802574, + 0.018331212922930717, + 0.015867752954363823, + 0.01701870560646057, + 0.02696044184267521, + 0.021348021924495697, + 0.011330395936965942, + 0.015136062167584896, + 0.015209968201816082, + 0.008775071240961552, + 0.01795908436179161, + 0.01186674740165472, + 0.009539092890918255, + 0.010893288999795914, + 0.009019927121698856, + 0.01284983940422535, + 0.010017462074756622, + 0.006035853177309036, + 0.006637969985604286, + 0.006566756870597601, + 0.004545682109892368, + 0.00645394017919898, + 0.005620272364467382, + 0.007016986608505249, + 0.012220022268593311, + 0.012613079510629177, + 0.008072346448898315, + 0.010035257786512375, + 0.009180672466754913, + 0.007592989597469568, + 0.006774010602384806, + 0.010030926205217838, + 0.020884590223431587, + 0.009056388400495052, + 0.010431130416691303, + 0.012946855276823044, + 0.02092321589589119, + 0.00995786301791668, + 0.008619493804872036, + 0.01342911459505558, + 0.0126155661419034, + 0.009385205805301666, + 0.006126916967332363, + 0.0061082737520337105, + 0.00748016219586134, + 0.006593219470232725, + 0.00531380670145154, + 0.006636636331677437, + 0.005830450914800167, + 0.007665748707950115, + 0.006536913104355335, + 0.005282630678266287, + 0.008209343999624252, + 0.0069101732224226, + 0.005907618906348944, + 0.01012234017252922, + 0.006152370944619179, + 0.007929323241114616, + 0.00928564090281725, + 0.009521514177322388, + 0.011305440217256546, + 0.014780665747821331, + 0.006900543347001076, + 0.012183113023638725, + 0.010475017130374908, + 0.005059550516307354, + 0.006516694091260433, + 0.006819240748882294, + 0.006252241786569357, + 0.006472378503531218, + 0.008058669976890087, + 0.007495235186070204, + 0.01123129017651081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.009789316914975643 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742704629898071 + ], + "opposite_logprob_tokens": [ + -0.4742704927921295 + ], + "logprob_reward": -0.9742704629898071, + "yes_no_margin_reward": -0.4999999701976776, + "token_credit_reward": -0.9742704629898071, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24370963871479034, + 0.11626248806715012, + 0.14015407860279083, + 0.0858461782336235, + 0.16440263390541077, + 0.09292320162057877, + 0.02776191756129265, + 0.027990251779556274, + 0.06046430021524429, + 0.04048539325594902 + ], + "credit_map": [ + 0.019506098702549934, + 0.016924085095524788, + 0.06982730329036713, + 0.04157565161585808, + 0.017700431868433952, + 0.019921325147151947, + 0.018788790330290794, + 0.03946596011519432, + 0.012289047241210938, + 0.018899772316217422, + 0.01311794575303793, + 0.014507416635751724, + 0.020999347791075706, + 0.00964321754872799, + 0.012380844913423061, + 0.014424887485802174, + 0.014530337415635586, + 0.025836389511823654, + 0.020739587023854256, + 0.013964300975203514, + 0.016649043187499046, + 0.0090786162763834, + 0.021611792966723442, + 0.017744017764925957, + 0.011032302863895893, + 0.009325852617621422, + 0.011216389015316963, + 0.0075552514754235744, + 0.010853677056729794, + 0.007668898906558752, + 0.009601875208318233, + 0.018591927364468575, + 0.024527864530682564, + 0.013334659859538078, + 0.02182745561003685, + 0.01637645997107029, + 0.013990920037031174, + 0.012526866048574448, + 0.016215572133660316, + 0.04560283571481705, + 0.010694880969822407, + 0.008317893370985985, + 0.012037461623549461, + 0.018612533807754517, + 0.008564725518226624, + 0.00798578280955553, + 0.014620136469602585, + 0.012089786119759083, + 0.005345400422811508, + 0.0028496405575424433, + 0.0038250971119850874, + 0.004818696528673172, + 0.0028336476534605026, + 0.0023153377696871758, + 0.003364403499290347, + 0.0024096951819956303, + 0.003927316982299089, + 0.0026344649959355593, + 0.0033616831060498953, + 0.003961804788559675, + 0.003858933225274086, + 0.0019996995106339455, + 0.005216025747358799, + 0.0030303222592920065, + 0.006845811847597361, + 0.0052459160797297955, + 0.00891102198511362, + 0.005961432587355375, + 0.0076321447268128395, + 0.006893502548336983, + 0.009797760285437107, + 0.009176713414490223, + 0.00601231399923563, + 0.00456950394436717, + 0.004495940171182156, + 0.004513257648795843, + 0.004585933405905962, + 0.00434114458039403, + 0.00541822100058198, + 0.0065490794368088245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.9742704629898071 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020595503970980644 + ], + "opposite_logprob_tokens": [ + -3.8955955505371094 + ], + "logprob_reward": -0.020595503970980644, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020595503970980644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38121363520622253, + 0.07211431860923767, + 0.07930673658847809, + 0.06486642360687256, + 0.1993560791015625, + 0.03418690338730812, + 0.0408940464258194, + 0.04153015464544296, + 0.029724936932325363, + 0.056806780397892 + ], + "credit_map": [ + 0.027941644191741943, + 0.042323675006628036, + 0.07675144821405411, + 0.07261203974485397, + 0.03464711084961891, + 0.036712896078825, + 0.04912808537483215, + 0.0410967692732811, + 0.007373512256890535, + 0.010264446958899498, + 0.009286236949265003, + 0.011433201842010021, + 0.008186373859643936, + 0.0054050893522799015, + 0.007359658367931843, + 0.012805795297026634, + 0.011378432624042034, + 0.01424796599894762, + 0.0092378044500947, + 0.007935574278235435, + 0.005772837903350592, + 0.00722131785005331, + 0.00835760124027729, + 0.015155197121202946, + 0.01090838573873043, + 0.004342139698565006, + 0.01038278080523014, + 0.006268051918596029, + 0.0030524160247296095, + 0.0036352172028273344, + 0.004871247336268425, + 0.02140618860721588, + 0.026373624801635742, + 0.020627057179808617, + 0.026480821892619133, + 0.02818966656923294, + 0.01534744817763567, + 0.019163846969604492, + 0.026077933609485626, + 0.037095677107572556, + 0.00663339626044035, + 0.004467816092073917, + 0.004097985103726387, + 0.004060308448970318, + 0.004337333142757416, + 0.0020972294732928276, + 0.004409365821629763, + 0.004083470441401005, + 0.007281131111085415, + 0.004553619772195816, + 0.0048384289257228374, + 0.005279442761093378, + 0.004497680347412825, + 0.005123939365148544, + 0.005632222630083561, + 0.0036875817459076643, + 0.004618654027581215, + 0.003465189365670085, + 0.005995886400341988, + 0.007134026847779751, + 0.005433714482933283, + 0.002810204867273569, + 0.00847464520484209, + 0.00359783205203712, + 0.0032911566086113453, + 0.00337457237765193, + 0.0029431465081870556, + 0.00410705478861928, + 0.004503090400248766, + 0.0032550368923693895, + 0.0035417245235294104, + 0.004709157161414623, + 0.005640781484544277, + 0.005431115627288818, + 0.006841767579317093, + 0.006962592713534832, + 0.00567212887108326, + 0.009307519532740116, + 0.006821379531174898, + 0.010129495523869991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.020595503970980644 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100012592971325 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.004100012592971325, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100012592971325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2850230634212494, + 0.09757938235998154, + 0.08297190815210342, + 0.05587628856301308, + 0.08942580968141556, + 0.09401965141296387, + 0.06823081523180008, + 0.06450986862182617, + 0.04514336958527565, + 0.11721981316804886 + ], + "credit_map": [ + 0.017463630065321922, + 0.03025302290916443, + 0.05289243534207344, + 0.043863896280527115, + 0.02183867059648037, + 0.034701380878686905, + 0.03562718629837036, + 0.048382822424173355, + 0.009923866018652916, + 0.013570006936788559, + 0.011108286678791046, + 0.013578700833022594, + 0.015834316611289978, + 0.008985524997115135, + 0.011019322089850903, + 0.013559362851083279, + 0.008644603192806244, + 0.013376784510910511, + 0.010531431064009666, + 0.009770429693162441, + 0.007808756083250046, + 0.008282143622636795, + 0.012509576976299286, + 0.012048179283738136, + 0.005679261405020952, + 0.00757344625890255, + 0.0066102175042033195, + 0.0057403878308832645, + 0.008117974735796452, + 0.005107310600578785, + 0.007834993302822113, + 0.00921269878745079, + 0.012587674893438816, + 0.01006342750042677, + 0.01131259836256504, + 0.008349504321813583, + 0.00967300683259964, + 0.00711683789268136, + 0.010479064658284187, + 0.01984369568526745, + 0.007963686250150204, + 0.00866776891052723, + 0.012406336143612862, + 0.02103857696056366, + 0.007827560417354107, + 0.008345933631062508, + 0.016220245510339737, + 0.011549544520676136, + 0.012297207489609718, + 0.008616548031568527, + 0.008414756506681442, + 0.008736247196793556, + 0.00819921214133501, + 0.007293822709470987, + 0.007739579305052757, + 0.006933440919965506, + 0.006649782415479422, + 0.005333765409886837, + 0.009372430853545666, + 0.009460706263780594, + 0.009737358428537846, + 0.004233524203300476, + 0.013320773839950562, + 0.006401530466973782, + 0.006617099512368441, + 0.004209868609905243, + 0.005357192363590002, + 0.0064807371236383915, + 0.007297826465219259, + 0.004492141306400299, + 0.006086770445108414, + 0.004601730965077877, + 0.009356082417070866, + 0.013815830461680889, + 0.011496489867568016, + 0.0164504311978817, + 0.01314043439924717, + 0.015945253893733025, + 0.015945015475153923, + 0.02107028104364872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.004100012592971325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "perfectly static, extreme macro shot captures the moment a viscous drop of vibrant sapphire detergent pierces the pristine, crystal-clear surface of water within a glass container. For a fleeting moment, the drop maintains its spherical shape before beginning to dissolve. As the camera slowly and deliberately zooms out, a concentrated blue gracefully unfolds downward. It blooms into an ethereal, smoke-like plume that swirls and spirals hypnotically, creating a mesmerizing hydrodynamic ballet. Backlit by a clean, bright light source, the intricate tendrils of color become translucent, with luminous edges and deep, shadowy cores, showcasing a rich gradation from deep cobalt to pale, transparent azure. Set against a soft, out-of-focus neutral background, the entire scene evokes an atmosphere of serene, captivating, and almost meditative abstract beauty.", + "seed": 17952, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_12/index_4467_seed_17952_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.21309701800346376, + "vlm_logprob": -0.21309701800346376, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.21309701800346376, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007900300435721874 + ], + "opposite_logprob_tokens": [ + -5.132900238037109 + ], + "logprob_reward": -0.007900300435721874, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.007900300435721874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1609748899936676, + 0.07747264206409454, + 0.0896206647157669, + 0.07644018530845642, + 0.22157016396522522, + 0.0381176583468914, + 0.18149513006210327, + 0.07948245108127594, + 0.021549884229898453, + 0.053276244550943375 + ], + "credit_map": [ + 0.014528840780258179, + 0.010803347453474998, + 0.03613080829381943, + 0.022533223032951355, + 0.012989760376513004, + 0.013535923324525356, + 0.032289814203977585, + 0.018163185566663742, + 0.010770512744784355, + 0.008376861922442913, + 0.010850089602172375, + 0.00805498193949461, + 0.008645384572446346, + 0.014405913650989532, + 0.006369348149746656, + 0.009999556466937065, + 0.010738050565123558, + 0.008533347398042679, + 0.01134879793971777, + 0.011830596253275871, + 0.00773788383230567, + 0.014521045610308647, + 0.009975876659154892, + 0.014935075305402279, + 0.008732535876333714, + 0.00549329724162817, + 0.014905748888850212, + 0.008229751139879227, + 0.005447410978376865, + 0.01057400368154049, + 0.008044326677918434, + 0.015013113617897034, + 0.03240715712308884, + 0.013347571715712547, + 0.0459895133972168, + 0.017137590795755386, + 0.010542700998485088, + 0.03711138293147087, + 0.020666591823101044, + 0.044367678463459015, + 0.006104399915784597, + 0.0028230175375938416, + 0.005464077927172184, + 0.004353232681751251, + 0.0035291060339659452, + 0.0071031805127859116, + 0.004775582812726498, + 0.0039650616236031055, + 0.029311101883649826, + 0.015770450234413147, + 0.02046464942395687, + 0.01695784367620945, + 0.014385383576154709, + 0.03769394010305405, + 0.022034162655472755, + 0.024877609685063362, + 0.007696983870118856, + 0.004904419183731079, + 0.006416122429072857, + 0.01791076362133026, + 0.004570799879729748, + 0.011000353842973709, + 0.007581203244626522, + 0.01940181478857994, + 0.0026054605841636658, + 0.0028344157617539167, + 0.0030731388833373785, + 0.0027632699348032475, + 0.0026152676437050104, + 0.0028279253747314215, + 0.0024171778932213783, + 0.002413228154182434, + 0.0066072698682546616, + 0.005843190010637045, + 0.00647724187001586, + 0.00606412161141634, + 0.006450376473367214, + 0.00960167683660984, + 0.006389322690665722, + 0.005843047983944416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.007900300435721874 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645367808640003 + ], + "opposite_logprob_tokens": [ + -4.758645534515381 + ], + "logprob_reward": -0.008645367808640003, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645367808640003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17538198828697205, + 0.41521283984184265, + 0.09479587525129318, + 0.030785102397203445, + 0.06497479230165482, + 0.024810731410980225, + 0.046516600996255875, + 0.06832408159971237, + 0.030577173456549644, + 0.04862085357308388 + ], + "credit_map": [ + 0.015042773447930813, + 0.01972140744328499, + 0.027309177443385124, + 0.0393819734454155, + 0.015328442677855492, + 0.014992388896644115, + 0.023342881351709366, + 0.020262937992811203, + 0.05266207456588745, + 0.042654890567064285, + 0.056793469935655594, + 0.03782564029097557, + 0.02732878550887108, + 0.06287086009979248, + 0.04483712464570999, + 0.090239979326725, + 0.008687525987625122, + 0.00869693048298359, + 0.017012128606438637, + 0.013270694762468338, + 0.007093960884958506, + 0.010236087255179882, + 0.009735177271068096, + 0.020063364878296852, + 0.0031166367698460817, + 0.00303444336168468, + 0.0035021821968257427, + 0.005261780694127083, + 0.002891425509005785, + 0.005063215736299753, + 0.004183151293545961, + 0.003732265206053853, + 0.00867131631821394, + 0.003894153283908963, + 0.014746610075235367, + 0.005511689465492964, + 0.0046622189693152905, + 0.008245429955422878, + 0.006139249540865421, + 0.013104116544127464, + 0.0030452159699052572, + 0.0023600722197443247, + 0.002782403724268079, + 0.0038144246209412813, + 0.0028285644948482513, + 0.0038861443754285574, + 0.0034128318075090647, + 0.0026810711715370417, + 0.007642145734280348, + 0.0031157066114246845, + 0.004864524118602276, + 0.004270038567483425, + 0.0037783910520374775, + 0.0135231027379632, + 0.004412752576172352, + 0.00490993307903409, + 0.007628762628883123, + 0.00474920216947794, + 0.0058203344233334064, + 0.013217395171523094, + 0.0045482087880373, + 0.010129905305802822, + 0.006711084861308336, + 0.015519182197749615, + 0.003928633406758308, + 0.0028683410491794348, + 0.004359204787760973, + 0.0061263577081263065, + 0.0031346974428743124, + 0.0027090241201221943, + 0.0034858076833188534, + 0.0039651053957641125, + 0.005728579126298428, + 0.005225266329944134, + 0.006334260571748018, + 0.007143484428524971, + 0.004981237929314375, + 0.009025730192661285, + 0.006218396592885256, + 0.003963897004723549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.008645367808640003 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742507934570312 + ], + "opposite_logprob_tokens": [ + -0.47425076365470886 + ], + "logprob_reward": -0.9742507934570312, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742507934570312, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07402551174163818, + 0.0774395763874054, + 0.04582870379090309, + 0.06441649794578552, + 0.11121855676174164, + 0.03921051695942879, + 0.12026367336511612, + 0.34657201170921326, + 0.055466629564762115, + 0.06555835902690887 + ], + "credit_map": [ + 0.007084020879119635, + 0.008464071899652481, + 0.015187966637313366, + 0.010438544675707817, + 0.004615816753357649, + 0.007517565973103046, + 0.01341299619525671, + 0.007304519414901733, + 0.009778977371752262, + 0.005494674202054739, + 0.011577469296753407, + 0.005177422426640987, + 0.0065360539592802525, + 0.014167742803692818, + 0.007389736827462912, + 0.01731749065220356, + 0.00469117471948266, + 0.0026860053185373545, + 0.010142652317881584, + 0.003892931155860424, + 0.0022763062734156847, + 0.005339726805686951, + 0.003932426683604717, + 0.012867472134530544, + 0.008132664486765862, + 0.004079826176166534, + 0.012013720348477364, + 0.0073782396502792835, + 0.005289498716592789, + 0.007396419532597065, + 0.007204543799161911, + 0.01292157731950283, + 0.015432310290634632, + 0.008157504722476006, + 0.02223559282720089, + 0.009060928598046303, + 0.007620358839631081, + 0.0172549057751894, + 0.011048892512917519, + 0.02040805108845234, + 0.0060535818338394165, + 0.002345649991184473, + 0.006482764147222042, + 0.00502530112862587, + 0.0031965840607881546, + 0.0057175676338374615, + 0.005346629302948713, + 0.005042435601353645, + 0.018802763894200325, + 0.0107322558760643, + 0.01161041297018528, + 0.012186888605356216, + 0.009982515126466751, + 0.022176148369908333, + 0.01564394310116768, + 0.01912872865796089, + 0.018281767144799232, + 0.02009694278240204, + 0.013058396987617016, + 0.1047985777258873, + 0.012425858527421951, + 0.022055912762880325, + 0.028148841112852097, + 0.12770569324493408, + 0.005546276457607746, + 0.006828398443758488, + 0.006226381286978722, + 0.008905550464987755, + 0.006188301369547844, + 0.008941645734012127, + 0.006666877772659063, + 0.006163198500871658, + 0.006409729365259409, + 0.007847083732485771, + 0.006086159963160753, + 0.008597109466791153, + 0.0064299325458705425, + 0.013669904321432114, + 0.006495856214314699, + 0.010022581554949284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.9742507934570312 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004549354314804 + ], + "opposite_logprob_tokens": [ + -2.6950454711914062 + ], + "logprob_reward": -0.07004549354314804, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004549354314804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23939499258995056, + 0.06193054839968681, + 0.042718395590782166, + 0.05323838070034981, + 0.047062020748853683, + 0.03792929649353027, + 0.04040415212512016, + 0.20353932678699493, + 0.1459328681230545, + 0.1278500258922577 + ], + "credit_map": [ + 0.019662193953990936, + 0.02102268487215042, + 0.034665800631046295, + 0.03410518169403076, + 0.026691025123000145, + 0.03437292203307152, + 0.028853092342615128, + 0.0400221012532711, + 0.0063961707055568695, + 0.004933737218379974, + 0.008940308354794979, + 0.007743167225271463, + 0.004137569107115269, + 0.008778233081102371, + 0.008313850499689579, + 0.012687505222856998, + 0.005088089965283871, + 0.004042798653244972, + 0.0065393694676458836, + 0.005003655795007944, + 0.002705151680856943, + 0.006756929215043783, + 0.004561822395771742, + 0.00802057795226574, + 0.006090667098760605, + 0.0033573105465620756, + 0.008437485434114933, + 0.008193520829081535, + 0.0046585071831941605, + 0.00607747258618474, + 0.0074134888127446175, + 0.00900992751121521, + 0.005349384155124426, + 0.006477066315710545, + 0.0055909594520926476, + 0.0056182509288191795, + 0.006707419641315937, + 0.004683121107518673, + 0.00780629925429821, + 0.004829519893974066, + 0.005086648277938366, + 0.002713666995987296, + 0.004524766001850367, + 0.006341621279716492, + 0.004474910441786051, + 0.004750635474920273, + 0.005987345706671476, + 0.004049703944474459, + 0.0048964256420731544, + 0.003690610406920314, + 0.006819722708314657, + 0.005051294807344675, + 0.003323554527014494, + 0.006472418084740639, + 0.005492680706083775, + 0.004657446406781673, + 0.01616119220852852, + 0.01178743876516819, + 0.012550749816000462, + 0.05340276658535004, + 0.013687694445252419, + 0.017361167818307877, + 0.0147007517516613, + 0.06388755887746811, + 0.013871043920516968, + 0.015188475139439106, + 0.022631561383605003, + 0.02317913994193077, + 0.020137950778007507, + 0.01992892287671566, + 0.016456563025712967, + 0.014539211057126522, + 0.017275772988796234, + 0.012112473137676716, + 0.016301816329360008, + 0.013813354074954987, + 0.014621966518461704, + 0.025863589718937874, + 0.014234233647584915, + 0.01362682320177555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.07004549354314804 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643134772777557 + ], + "opposite_logprob_tokens": [ + -5.379642963409424 + ], + "logprob_reward": -0.004643134772777557, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643134772777557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14780078828334808, + 0.036316562443971634, + 0.043888289481401443, + 0.05859191343188286, + 0.1355735957622528, + 0.03730074688792229, + 0.0614565946161747, + 0.2876291573047638, + 0.056337542831897736, + 0.13510480523109436 + ], + "credit_map": [ + 0.008251057006418705, + 0.009926135651767254, + 0.03027474693953991, + 0.021673809736967087, + 0.011660751886665821, + 0.018085043877363205, + 0.0330914668738842, + 0.014837771654129028, + 0.0040544746443629265, + 0.003714166348800063, + 0.0053846570663154125, + 0.0037118582986295223, + 0.003586726263165474, + 0.005148778669536114, + 0.004724745638668537, + 0.005991156212985516, + 0.00544437812641263, + 0.003535077441483736, + 0.00659605348482728, + 0.004489224404096603, + 0.002624077722430229, + 0.007428659591823816, + 0.00515833031386137, + 0.008612489327788353, + 0.00732840783894062, + 0.0032799136824905872, + 0.013547064736485481, + 0.005242106970399618, + 0.0028620443772524595, + 0.008096665143966675, + 0.005208232440054417, + 0.013027483597397804, + 0.01976463571190834, + 0.007288011722266674, + 0.0277602206915617, + 0.008625400252640247, + 0.008548728190362453, + 0.02670839987695217, + 0.010872202925384045, + 0.026005996391177177, + 0.0057507846504449844, + 0.0027122150640934706, + 0.004993936978280544, + 0.005153561942279339, + 0.00409360509365797, + 0.004491151310503483, + 0.005625104997307062, + 0.004480382427573204, + 0.007551626302301884, + 0.004998465068638325, + 0.012561922892928123, + 0.007936570793390274, + 0.004517706111073494, + 0.010020529851317406, + 0.007569585461169481, + 0.00630018999800086, + 0.01245515514165163, + 0.016233492642641068, + 0.011737700551748276, + 0.09068650752305984, + 0.007741061970591545, + 0.020609121769666672, + 0.02334250509738922, + 0.10482361167669296, + 0.00548829510807991, + 0.00609869509935379, + 0.0076666162349283695, + 0.0068453215062618256, + 0.00828680582344532, + 0.007379893213510513, + 0.007464840542525053, + 0.007107075769454241, + 0.014982943423092365, + 0.01499943621456623, + 0.014629364013671875, + 0.01939011737704277, + 0.01511695422232151, + 0.030138451606035233, + 0.015383260324597359, + 0.010464292019605637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.004643134772777557 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "perfectly static, extreme macro shot captures the moment a viscous drop of vibrant sapphire detergent pierces the pristine, crystal-clear surface of water within a glass container. For a fleeting moment, the drop maintains its spherical shape before beginning to dissolve. As the camera slowly and deliberately zooms out, a concentrated blue gracefully unfolds downward. It blooms into an ethereal, smoke-like plume that swirls and spirals hypnotically, creating a mesmerizing hydrodynamic ballet. Backlit by a clean, bright light source, the intricate tendrils of color become translucent, with luminous edges and deep, shadowy cores, showcasing a rich gradation from deep cobalt to pale, transparent azure. Set against a soft, out-of-focus neutral background, the entire scene evokes an atmosphere of serene, captivating, and almost meditative abstract beauty.", + "seed": 17953, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_12/index_4467_seed_17953_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04087552381679416, + "vlm_logprob": -0.04087552381679416, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.04087552381679416, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00916382111608982 + ], + "opposite_logprob_tokens": [ + -5.134163856506348 + ], + "logprob_reward": -0.00916382111608982, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00916382111608982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11419083178043365, + 0.0950469821691513, + 0.1557535082101822, + 0.07076535373926163, + 0.06515098363161087, + 0.07621228694915771, + 0.038721028715372086, + 0.05957202613353729, + 0.21381475031375885, + 0.1107722669839859 + ], + "credit_map": [ + 0.011421754024922848, + 0.009469364769756794, + 0.03249223157763481, + 0.011651366017758846, + 0.007428609766066074, + 0.009323838166892529, + 0.015050419606268406, + 0.017353245988488197, + 0.008158457465469837, + 0.007736705709248781, + 0.013708396814763546, + 0.01644781231880188, + 0.007346277590841055, + 0.007561461068689823, + 0.018713653087615967, + 0.015374216251075268, + 0.018300756812095642, + 0.016286112368106842, + 0.025811055675148964, + 0.024471692740917206, + 0.01542961597442627, + 0.012970434501767159, + 0.023808643221855164, + 0.018675193190574646, + 0.004550842568278313, + 0.0063643730245530605, + 0.017769865691661835, + 0.006862655282020569, + 0.005269179120659828, + 0.006552564445883036, + 0.00721269566565752, + 0.016183186322450638, + 0.007456193212419748, + 0.005374038126319647, + 0.009783844463527203, + 0.0092075290158391, + 0.004992679692804813, + 0.011053766123950481, + 0.008128906600177288, + 0.009154022671282291, + 0.006915895733982325, + 0.008718634024262428, + 0.01168871484696865, + 0.010638752952218056, + 0.012060347013175488, + 0.0065802657045423985, + 0.008897170424461365, + 0.01071250345557928, + 0.004586311522871256, + 0.0030273953452706337, + 0.006541631184518337, + 0.0045889862813055515, + 0.003950483165681362, + 0.00364599353633821, + 0.004619075451046228, + 0.007761150598526001, + 0.005775175988674164, + 0.007760017178952694, + 0.007526912726461887, + 0.010408986359834671, + 0.009965099394321442, + 0.0035627889446914196, + 0.007454844657331705, + 0.007118201814591885, + 0.020244330167770386, + 0.0175380427390337, + 0.0320458710193634, + 0.041642602533102036, + 0.03245602920651436, + 0.014178681187331676, + 0.024599043652415276, + 0.031110160052776337, + 0.036276042461395264, + 0.004468944855034351, + 0.034018680453300476, + 0.007286312058568001, + 0.005463817622512579, + 0.00727850291877985, + 0.0061499448493123055, + 0.009830022230744362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.00916382111608982 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207052998244762 + ], + "opposite_logprob_tokens": [ + -4.264206886291504 + ], + "logprob_reward": -0.014207052998244762, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207052998244762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10048931837081909, + 0.17598572373390198, + 0.17934872210025787, + 0.06847824156284332, + 0.08187124878168106, + 0.07108671218156815, + 0.03549676761031151, + 0.03999493271112442, + 0.1398966908454895, + 0.10735155642032623 + ], + "credit_map": [ + 0.011918120086193085, + 0.012542926706373692, + 0.01346871443092823, + 0.017822757363319397, + 0.009943178854882717, + 0.009034374728798866, + 0.014325507916510105, + 0.011433752253651619, + 0.015891844406723976, + 0.01741424947977066, + 0.02282175049185753, + 0.04119647294282913, + 0.0201543141156435, + 0.011269455775618553, + 0.03001309745013714, + 0.017224552109837532, + 0.015753693878650665, + 0.016346272081136703, + 0.03981393203139305, + 0.02014850825071335, + 0.012751899659633636, + 0.01598879136145115, + 0.0226880069822073, + 0.035857632756233215, + 0.006760176736861467, + 0.011130906641483307, + 0.011809548363089561, + 0.008423506282269955, + 0.004774318542331457, + 0.007605074904859066, + 0.007681237068027258, + 0.01029348187148571, + 0.009295301511883736, + 0.007301947567611933, + 0.014438963495194912, + 0.006096712779253721, + 0.007560316007584333, + 0.013472079299390316, + 0.008107740432024002, + 0.015598200261592865, + 0.009839080274105072, + 0.005552482791244984, + 0.011789504438638687, + 0.012003965675830841, + 0.008437294512987137, + 0.004453015048056841, + 0.009804111905395985, + 0.009207267314195633, + 0.0038437298499047756, + 0.004020205698907375, + 0.004965456202626228, + 0.00463957479223609, + 0.006616111844778061, + 0.0027940908912569284, + 0.0044752792455255985, + 0.004142322111874819, + 0.003715706989169121, + 0.005617691203951836, + 0.006336141377687454, + 0.005280218552798033, + 0.005846608895808458, + 0.0026377278845757246, + 0.005583015736192465, + 0.004977828823029995, + 0.016432302072644234, + 0.01287537906318903, + 0.024785764515399933, + 0.019022414460778236, + 0.02040068805217743, + 0.00952551793307066, + 0.01624343730509281, + 0.02061120793223381, + 0.024468405172228813, + 0.006509009748697281, + 0.026520073413848877, + 0.011741865426301956, + 0.008256944827735424, + 0.009360045194625854, + 0.008694762364029884, + 0.01180045772343874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.014207052998244762 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603267788887024 + ], + "opposite_logprob_tokens": [ + -1.9103267192840576 + ], + "logprob_reward": -0.1603267788887024, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603267788887024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10752960294485092, + 0.04829641059041023, + 0.052837125957012177, + 0.05937642604112625, + 0.17087754607200623, + 0.11302829533815384, + 0.05776534974575043, + 0.03566012531518936, + 0.14897887408733368, + 0.205650195479393 + ], + "credit_map": [ + 0.005964240990579128, + 0.008303442969918251, + 0.03484687581658363, + 0.0135917654260993, + 0.006141699384897947, + 0.00508576724678278, + 0.016718167811632156, + 0.01687765121459961, + 0.0052501303143799305, + 0.0038770122919231653, + 0.008246817626059055, + 0.007526467088609934, + 0.003864492056891322, + 0.003960863687098026, + 0.007136355619877577, + 0.008434280753135681, + 0.005872748792171478, + 0.004902683664113283, + 0.00861255545169115, + 0.008061172440648079, + 0.004749837797135115, + 0.005131916608661413, + 0.009843860752880573, + 0.005662351381033659, + 0.005939624737948179, + 0.00638585677370429, + 0.010117623955011368, + 0.008455678820610046, + 0.005067497957497835, + 0.0053237685933709145, + 0.010441303253173828, + 0.007645078469067812, + 0.01816723868250847, + 0.015026109293103218, + 0.04117694869637489, + 0.013966811820864677, + 0.008883416652679443, + 0.019757386296987534, + 0.01594081148505211, + 0.03795883432030678, + 0.01925908401608467, + 0.006208924576640129, + 0.024384336546063423, + 0.012929772958159447, + 0.006979493424296379, + 0.013092291541397572, + 0.009826606139540672, + 0.020347798243165016, + 0.008432145230472088, + 0.004584543406963348, + 0.0075486404821276665, + 0.008122306317090988, + 0.006520756520330906, + 0.005182614084333181, + 0.008594544604420662, + 0.008779807016253471, + 0.0037389767821878195, + 0.0034267345909029245, + 0.004627024754881859, + 0.006200253963470459, + 0.004780645482242107, + 0.003034295281395316, + 0.004799730610102415, + 0.005052465479820967, + 0.014932617545127869, + 0.011565945111215115, + 0.020823722705245018, + 0.02612110786139965, + 0.015425123274326324, + 0.014167134650051594, + 0.01587350107729435, + 0.030069725587964058, + 0.028813466429710388, + 0.011031225323677063, + 0.029953138902783394, + 0.02908347174525261, + 0.02234295755624771, + 0.02560833841562271, + 0.017429612576961517, + 0.04138801619410515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.1603267788887024 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818695291876793 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.01818695291876793, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818695291876793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24950161576271057, + 0.06257694959640503, + 0.0898638442158699, + 0.03687824308872223, + 0.04269446060061455, + 0.047662850469350815, + 0.03421320393681526, + 0.22664864361286163, + 0.04573487117886543, + 0.16422533988952637 + ], + "credit_map": [ + 0.023549145087599754, + 0.020348841324448586, + 0.05614560842514038, + 0.027842838317155838, + 0.024527477100491524, + 0.032074663788080215, + 0.031689997762441635, + 0.03332303464412689, + 0.006118792109191418, + 0.004589523188769817, + 0.011691800318658352, + 0.009823531843721867, + 0.004989721346646547, + 0.005316834431141615, + 0.010044788010418415, + 0.010001953691244125, + 0.00683110486716032, + 0.007519678212702274, + 0.016292572021484375, + 0.016262395307421684, + 0.007359971757978201, + 0.006577352061867714, + 0.01680734194815159, + 0.012213419191539288, + 0.0043585775420069695, + 0.0036888676695525646, + 0.0063112354837358, + 0.004255342297255993, + 0.0031357721891254187, + 0.004735346417874098, + 0.005692011211067438, + 0.004701089579612017, + 0.005098918918520212, + 0.003142674220725894, + 0.008034639991819859, + 0.0037512609269469976, + 0.0036639964673668146, + 0.00916471891105175, + 0.0037299771793186665, + 0.006108269095420837, + 0.006515908986330032, + 0.0038010424468666315, + 0.010637873783707619, + 0.0036444473080337048, + 0.0047482275404036045, + 0.00502183148637414, + 0.004965356085449457, + 0.00832816120237112, + 0.005457797087728977, + 0.0022316796239465475, + 0.005741524510085583, + 0.004287024028599262, + 0.0027522295713424683, + 0.003416858147829771, + 0.003582942998036742, + 0.006743145175278187, + 0.04482588917016983, + 0.015067752450704575, + 0.04609015956521034, + 0.0208281259983778, + 0.015738915652036667, + 0.017838379368185997, + 0.019593290984630585, + 0.04666613042354584, + 0.006206910591572523, + 0.003685515373945236, + 0.010062769055366516, + 0.005092599894851446, + 0.004888561554253101, + 0.00302861537784338, + 0.005104290787130594, + 0.007665607612580061, + 0.05385779216885567, + 0.006277013104408979, + 0.059028852730989456, + 0.0079573979601264, + 0.006753339432179928, + 0.01074582152068615, + 0.005307480692863464, + 0.014297652058303356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.01818695291876793 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00249301316216588 + ], + "opposite_logprob_tokens": [ + -6.002492904663086 + ], + "logprob_reward": -0.00249301316216588, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00249301316216588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21527628600597382, + 0.06346923112869263, + 0.1012984961271286, + 0.08203217387199402, + 0.07116569578647614, + 0.07195860892534256, + 0.053188223391771317, + 0.0659158006310463, + 0.12182340770959854, + 0.1538720726966858 + ], + "credit_map": [ + 0.01833295077085495, + 0.010880881920456886, + 0.06679217517375946, + 0.025716228410601616, + 0.012603470124304295, + 0.0174239594489336, + 0.033450767397880554, + 0.03007584810256958, + 0.003943073097616434, + 0.00602369150146842, + 0.009344365447759628, + 0.011861156672239304, + 0.00619027903303504, + 0.0045332275331020355, + 0.012951593846082687, + 0.00862184539437294, + 0.008474564179778099, + 0.013050902634859085, + 0.014437491074204445, + 0.017243720591068268, + 0.008997700177133083, + 0.00948786549270153, + 0.01813671551644802, + 0.011469531804323196, + 0.0069928402081131935, + 0.008278275839984417, + 0.014700988307595253, + 0.0117970434948802, + 0.005611946806311607, + 0.009399707429111004, + 0.011600597761571407, + 0.013650776818394661, + 0.006781356874853373, + 0.004269958473742008, + 0.018937021493911743, + 0.004670248832553625, + 0.004831419326364994, + 0.006206597201526165, + 0.005267699249088764, + 0.020201392471790314, + 0.01035338919609785, + 0.00651271129027009, + 0.013484623283147812, + 0.007208364550024271, + 0.008673086762428284, + 0.00704346876591444, + 0.007993401028215885, + 0.010689571499824524, + 0.007061468902975321, + 0.002671726979315281, + 0.010969418101012707, + 0.005369048565626144, + 0.004410221707075834, + 0.004308258183300495, + 0.00661639217287302, + 0.011781693436205387, + 0.011190952733159065, + 0.008203540928661823, + 0.009842733852565289, + 0.00823603942990303, + 0.007050966378301382, + 0.004545469768345356, + 0.00568951852619648, + 0.011156579479575157, + 0.014224397018551826, + 0.009962841868400574, + 0.021365320309996605, + 0.019162479788064957, + 0.011990834958851337, + 0.00928806234151125, + 0.01602800376713276, + 0.01980147324502468, + 0.032060351222753525, + 0.01325273234397173, + 0.024098271504044533, + 0.019717786461114883, + 0.023712415248155594, + 0.013482331298291683, + 0.013150406070053577, + 0.014397775754332542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.00249301316216588 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "perfectly static, extreme macro shot captures the moment a viscous drop of vibrant sapphire detergent pierces the pristine, crystal-clear surface of water within a glass container. For a fleeting moment, the drop maintains its spherical shape before beginning to dissolve. As the camera slowly and deliberately zooms out, a concentrated blue gracefully unfolds downward. It blooms into an ethereal, smoke-like plume that swirls and spirals hypnotically, creating a mesmerizing hydrodynamic ballet. Backlit by a clean, bright light source, the intricate tendrils of color become translucent, with luminous edges and deep, shadowy cores, showcasing a rich gradation from deep cobalt to pale, transparent azure. Set against a soft, out-of-focus neutral background, the entire scene evokes an atmosphere of serene, captivating, and almost meditative abstract beauty.", + "seed": 17954, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_12/index_4467_seed_17954_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04105089106597006, + "vlm_logprob": -0.04105089106597006, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.04105089106597006, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01242762990295887 + ], + "opposite_logprob_tokens": [ + -4.512427806854248 + ], + "logprob_reward": -0.01242762990295887, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01242762990295887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2189929485321045, + 0.06425654143095016, + 0.03632470592856407, + 0.12343885749578476, + 0.07494883984327316, + 0.08870866149663925, + 0.16828086972236633, + 0.03361891955137253, + 0.048212070018053055, + 0.14321757853031158 + ], + "credit_map": [ + 0.029656685888767242, + 0.02859107404947281, + 0.033289287239313126, + 0.020732764154672623, + 0.025420255959033966, + 0.04088113456964493, + 0.017245225608348846, + 0.023176496848464012, + 0.007542576175183058, + 0.006084346678107977, + 0.011700066737830639, + 0.008006392978131771, + 0.0075068664737045765, + 0.007543139159679413, + 0.00703428266569972, + 0.008838866837322712, + 0.0054476140066981316, + 0.005926680285483599, + 0.005006417166441679, + 0.0036928115878254175, + 0.004872661549597979, + 0.0046007707715034485, + 0.002929418347775936, + 0.003848328022286296, + 0.012175963260233402, + 0.02602115459740162, + 0.014203231781721115, + 0.01746351271867752, + 0.01383026223629713, + 0.013069395907223225, + 0.010495968163013458, + 0.01617935672402382, + 0.007248517591506243, + 0.009488176554441452, + 0.00950008723884821, + 0.006037963554263115, + 0.008534743450582027, + 0.011052405461668968, + 0.007600023876875639, + 0.015486921183764935, + 0.008118759840726852, + 0.006821038201451302, + 0.01236577145755291, + 0.01491395290941, + 0.00896891113370657, + 0.017406119033694267, + 0.00823073461651802, + 0.011883368715643883, + 0.02226709946990013, + 0.0163123719394207, + 0.01531978975981474, + 0.013424642384052277, + 0.018902791664004326, + 0.0355355478823185, + 0.02851739712059498, + 0.018001211807131767, + 0.0038196775130927563, + 0.004180590622127056, + 0.004832922946661711, + 0.0034989449195563793, + 0.004267999902367592, + 0.00467459624633193, + 0.004666641354560852, + 0.0036775399930775166, + 0.005787916947156191, + 0.0038175906520336866, + 0.004513132851570845, + 0.004223014693707228, + 0.007989875040948391, + 0.009053290821611881, + 0.004701273515820503, + 0.008125970140099525, + 0.009406909346580505, + 0.014695169404149055, + 0.008548649027943611, + 0.011655731126666069, + 0.03861520066857338, + 0.016861746087670326, + 0.016191495582461357, + 0.02724265679717064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.01242762990295887 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205172657966614 + ], + "opposite_logprob_tokens": [ + -4.264204978942871 + ], + "logprob_reward": -0.014205172657966614, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205172657966614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26073017716407776, + 0.1443784534931183, + 0.08166118711233139, + 0.13133540749549866, + 0.04605214297771454, + 0.029590269550681114, + 0.03643099591135979, + 0.043899890035390854, + 0.02449367381632328, + 0.20142778754234314 + ], + "credit_map": [ + 0.03445831686258316, + 0.028996380046010017, + 0.03449089080095291, + 0.03245452791452408, + 0.028919540345668793, + 0.04601418599486351, + 0.021371886134147644, + 0.03402445465326309, + 0.016896283254027367, + 0.013298077508807182, + 0.03336808830499649, + 0.01919756457209587, + 0.013842692598700523, + 0.015566829591989517, + 0.015015758574008942, + 0.017193172127008438, + 0.008076213300228119, + 0.009968226775527, + 0.008721591904759407, + 0.009122698567807674, + 0.008230738341808319, + 0.014602404087781906, + 0.007207795046269894, + 0.01573152467608452, + 0.015327339991927147, + 0.02621685341000557, + 0.01644147001206875, + 0.018611717969179153, + 0.016008993610739708, + 0.013110242784023285, + 0.010951518081128597, + 0.014667275361716747, + 0.0036277752369642258, + 0.00791456364095211, + 0.0041930899024009705, + 0.007006652653217316, + 0.006456300616264343, + 0.005594149231910706, + 0.006405319087207317, + 0.004854298662394285, + 0.0036242229398339987, + 0.003190129529684782, + 0.004155945964157581, + 0.0023848742712289095, + 0.002908914815634489, + 0.005977770313620567, + 0.0033551359083503485, + 0.003993277903646231, + 0.004430997651070356, + 0.0038924976252019405, + 0.0036199011374264956, + 0.003170051146298647, + 0.004055955912917852, + 0.007187485229223967, + 0.0060578929260373116, + 0.004016216844320297, + 0.004798983223736286, + 0.005358306225389242, + 0.00555798877030611, + 0.004927261266857386, + 0.00580791337415576, + 0.006567355245351791, + 0.00615847110748291, + 0.004723614547401667, + 0.0022509892005473375, + 0.0019249286269769073, + 0.002621920546516776, + 0.002226862357929349, + 0.003928668797016144, + 0.004246352706104517, + 0.002526481170207262, + 0.004767472855746746, + 0.010621610097587109, + 0.02152586169540882, + 0.007770983036607504, + 0.014171783812344074, + 0.06412653625011444, + 0.019454287365078926, + 0.01444430835545063, + 0.04931243136525154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.014205172657966614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034728288650513 + ], + "opposite_logprob_tokens": [ + -1.9103472232818604 + ], + "logprob_reward": -0.16034728288650513, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034728288650513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25244954228401184, + 0.040076762437820435, + 0.04701025038957596, + 0.19996370375156403, + 0.09460052847862244, + 0.05992375314235687, + 0.21165621280670166, + 0.018643882125616074, + 0.033591900020837784, + 0.04208342730998993 + ], + "credit_map": [ + 0.03390051797032356, + 0.02591085433959961, + 0.040358491241931915, + 0.027256151661276817, + 0.025681916624307632, + 0.046606242656707764, + 0.020608389750123024, + 0.03212699666619301, + 0.004429821390658617, + 0.00371331418864429, + 0.007601766847074032, + 0.002775902394205332, + 0.004618213512003422, + 0.005919926799833775, + 0.002923550782725215, + 0.008094267919659615, + 0.004970741458237171, + 0.006031221244484186, + 0.005396811291575432, + 0.004610763397067785, + 0.005081535317003727, + 0.007813500240445137, + 0.004006002098321915, + 0.009099675342440605, + 0.02294217236340046, + 0.028466613963246346, + 0.022075040265917778, + 0.023518390953540802, + 0.040199629962444305, + 0.020982135087251663, + 0.021924355998635292, + 0.019855376332998276, + 0.007529138587415218, + 0.014137495309114456, + 0.008254965767264366, + 0.012755042873322964, + 0.015518741682171822, + 0.01258016750216484, + 0.012476172298192978, + 0.011348804458975792, + 0.00499057536944747, + 0.00533063942566514, + 0.006532873492687941, + 0.009558269754052162, + 0.007887122221291065, + 0.011346105486154556, + 0.006363208405673504, + 0.007914955727756023, + 0.029842276126146317, + 0.021030234172940254, + 0.019841337576508522, + 0.017556676641106606, + 0.023557467386126518, + 0.040832165628671646, + 0.03722282871603966, + 0.021773235872387886, + 0.0021215532906353474, + 0.0020961295813322067, + 0.0024827290326356888, + 0.0019702704157680273, + 0.002504007425159216, + 0.003143291687592864, + 0.002516223583370447, + 0.001809676643460989, + 0.005673902574926615, + 0.0034447740763425827, + 0.003452783450484276, + 0.0027915574610233307, + 0.0039518182165920734, + 0.0072629693895578384, + 0.0027666424866765738, + 0.004247453063726425, + 0.003523557912558317, + 0.004415643867105246, + 0.00342255225405097, + 0.003917632158845663, + 0.007166680879890919, + 0.008469377644360065, + 0.004750295076519251, + 0.006417684257030487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.16034728288650513 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070831567049026 + ], + "opposite_logprob_tokens": [ + -4.14107084274292 + ], + "logprob_reward": -0.016070831567049026, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070831567049026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40628671646118164, + 0.15867255628108978, + 0.056988541036844254, + 0.0649588406085968, + 0.05535126104950905, + 0.11050774902105331, + 0.028424615040421486, + 0.013953310437500477, + 0.02414385974407196, + 0.08071260899305344 + ], + "credit_map": [ + 0.05563894286751747, + 0.03977136313915253, + 0.06366947293281555, + 0.050154756754636765, + 0.04072507470846176, + 0.07198101282119751, + 0.03477966785430908, + 0.04956638440489769, + 0.021661899983882904, + 0.01030011847615242, + 0.04863044619560242, + 0.010909536853432655, + 0.008827057667076588, + 0.019131846725940704, + 0.015556496568024158, + 0.02365514449775219, + 0.005680157337337732, + 0.004704861436039209, + 0.008718352764844894, + 0.0041964794509112835, + 0.0045600286684930325, + 0.010330654680728912, + 0.0044435905292630196, + 0.014354410581290722, + 0.007021558936685324, + 0.008195936679840088, + 0.006377925630658865, + 0.008071629330515862, + 0.014389924705028534, + 0.0068557909689843655, + 0.005818475969135761, + 0.008227593265473843, + 0.004675578325986862, + 0.007451874669641256, + 0.004999546799808741, + 0.007015347946435213, + 0.009284996427595615, + 0.008551758714020252, + 0.005669168662279844, + 0.007702989503741264, + 0.010928560979664326, + 0.017726588994264603, + 0.010017015971243382, + 0.017836622893810272, + 0.014942655339837074, + 0.01601049117743969, + 0.011396539397537708, + 0.011649270541965961, + 0.003903436241671443, + 0.0023454988840967417, + 0.003640887327492237, + 0.0023948713205754757, + 0.003381071612238884, + 0.006012878380715847, + 0.0038595153018832207, + 0.002886455273255706, + 0.0015005262102931738, + 0.0011691980762407184, + 0.0015507303178310394, + 0.0013124218676239252, + 0.0016851827967911959, + 0.0031129131093621254, + 0.0017318192403763533, + 0.0018905176548287272, + 0.0031586564145982265, + 0.002283997368067503, + 0.0022769940551370382, + 0.0017792569706216455, + 0.0024539660662412643, + 0.006411985494196415, + 0.0033974621910601854, + 0.0023815406020730734, + 0.004225132986903191, + 0.008520851843059063, + 0.004545156378298998, + 0.008617265149950981, + 0.022686416283249855, + 0.006951523013412952, + 0.008913958445191383, + 0.016252301633358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.016070831567049026 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a viscous drop of vibrant sapphire detergent visible piercing the surface of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the drop maintain a spherical shape for a moment before beginning to dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and deliberately zoom out from the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background soft and out-of-focus with neutral colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene backlit by a clean, bright light source creating translucent tendrils with luminous edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002203538315370679 + ], + "opposite_logprob_tokens": [ + -6.127203464508057 + ], + "logprob_reward": -0.002203538315370679, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002203538315370679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22658133506774902, + 0.084238201379776, + 0.25210076570510864, + 0.0626487135887146, + 0.094392791390419, + 0.06442238390445709, + 0.07009999454021454, + 0.03737635910511017, + 0.03458492085337639, + 0.0735546126961708 + ], + "credit_map": [ + 0.03171248733997345, + 0.020328445360064507, + 0.033141542226076126, + 0.026960689574480057, + 0.025564225390553474, + 0.04294484108686447, + 0.017987165600061417, + 0.02794191800057888, + 0.011405380442738533, + 0.006456708535552025, + 0.023781899362802505, + 0.007577991578727961, + 0.006482291501015425, + 0.009017955511808395, + 0.008493559435009956, + 0.011022415943443775, + 0.024677058681845665, + 0.029738033190369606, + 0.02396201901137829, + 0.027114270254969597, + 0.032951630651950836, + 0.037602733820676804, + 0.03211313858628273, + 0.04394187033176422, + 0.007647528778761625, + 0.011197625659406185, + 0.011220665648579597, + 0.006026309449225664, + 0.008425293490290642, + 0.00666053919121623, + 0.0037473179399967194, + 0.007723433431237936, + 0.00767936697229743, + 0.015396294184029102, + 0.008123081177473068, + 0.011338305659592152, + 0.014544411562383175, + 0.013992447406053543, + 0.010236231610178947, + 0.013082657009363174, + 0.006413963157683611, + 0.011157047003507614, + 0.006028317846357822, + 0.009734722785651684, + 0.009909486398100853, + 0.008137685246765614, + 0.006347989663481712, + 0.0066931708715856075, + 0.009184375405311584, + 0.006781839765608311, + 0.0077499025501310825, + 0.0056629967875778675, + 0.007901959121227264, + 0.015373777598142624, + 0.010431360453367233, + 0.0070137823931872845, + 0.004253407008945942, + 0.0030623329803347588, + 0.004279161803424358, + 0.0035127426963299513, + 0.0038346198853105307, + 0.008533034473657608, + 0.005311605520546436, + 0.004589452873915434, + 0.0047376626171171665, + 0.0032982686534523964, + 0.004350751638412476, + 0.0029226962942630053, + 0.0036009373143315315, + 0.008905521593987942, + 0.003321368247270584, + 0.003447713330388069, + 0.004989475477486849, + 0.007706493604928255, + 0.005526362452656031, + 0.006230753380805254, + 0.011309944093227386, + 0.015545669943094254, + 0.007007652428001165, + 0.015238257125020027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.002203538315370679 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_13/index_tensor([5348, 5348])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_13/index_tensor([5348, 5348])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c485bc7e4682473ba2c4b7e1d87b3de2856054e6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_13/index_tensor([5348, 5348])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sleek, eye-level tracking shot maintains a tight close-up of a short, nervous Hispanic teenage boy as he walks across the spacious, polished wooden gymnasium floor. He has dark, wavy hair, neatly combed, but a hint of nervous dampness lingers on his cheeks. He wears a light blue cotton dress shirt, buttoned stiffly at the collar, its slightly oversized fit accentuating his slender frame. His jaw is set, his gaze fixed straight ahead, his gait stiff and unnatural, his arms tucked tightly to his sides. The high-gloss varnish of the floor catches the cool, even glare of the overhead fluorescent lights, creating a long, distorted reflection. He deliberately approaches a gray metal folding chair that stands alone in an empty space, mechanically turning and carefully positioning himself on the edge of the seat. His posture remains erect, and his hands immediately grasp his knees, creating an atmosphere of intense anxiety and self-conscious vulnerability.", + "seed": 17439, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_13/index_5348_seed_17439_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0316611297428608, + "vlm_logprob": -0.0316611297428608, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.0316611297428608, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03959393873810768 + ], + "opposite_logprob_tokens": [ + -3.4145939350128174 + ], + "logprob_reward": -0.03959393873810768, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03959393873810768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23133547604084015, + 0.03915271535515785, + 0.05847099423408508, + 0.025413965806365013, + 0.1058235615491867, + 0.20150023698806763, + 0.06587662547826767, + 0.04373961687088013, + 0.11261802166700363, + 0.1160687506198883 + ], + "credit_map": [ + 0.017406415194272995, + 0.029650205746293068, + 0.02397930808365345, + 0.046078626066446304, + 0.02094019018113613, + 0.0287469569593668, + 0.04292748495936394, + 0.021606288850307465, + 0.0032133818604052067, + 0.004320795647799969, + 0.005061057861894369, + 0.004955595824867487, + 0.004467466846108437, + 0.007297201082110405, + 0.004913764074444771, + 0.004923452157527208, + 0.005307512357831001, + 0.006810383405536413, + 0.006453996989876032, + 0.009251350536942482, + 0.009357219561934471, + 0.005825272295624018, + 0.011325203813612461, + 0.004140057135373354, + 0.002895427867770195, + 0.0030031956266611814, + 0.0026728438679128885, + 0.002520355861634016, + 0.003942190669476986, + 0.0031397289130836725, + 0.0038622261490672827, + 0.0033779959194362164, + 0.014031877741217613, + 0.013990696519613266, + 0.025132805109024048, + 0.010854233987629414, + 0.01089894212782383, + 0.008036969229578972, + 0.013603348284959793, + 0.00927469227463007, + 0.022987358272075653, + 0.017401302233338356, + 0.026888251304626465, + 0.02741430513560772, + 0.015850529074668884, + 0.022687477990984917, + 0.025795910507440567, + 0.04247508943080902, + 0.009136338718235493, + 0.006650871597230434, + 0.011216461658477783, + 0.008775342255830765, + 0.006018813233822584, + 0.007237995974719524, + 0.007834397256374359, + 0.00900640245527029, + 0.008788339793682098, + 0.005057205446064472, + 0.006310495547950268, + 0.00398257514461875, + 0.0037572034634649754, + 0.0037169293500483036, + 0.004805775824934244, + 0.007321094162762165, + 0.011273516342043877, + 0.01042815763503313, + 0.0154002346098423, + 0.017676757648587227, + 0.012757008895277977, + 0.016896450892090797, + 0.010736928321421146, + 0.017448963597416878, + 0.015636736527085304, + 0.01952766254544258, + 0.0168138574808836, + 0.0070584663189947605, + 0.019703971222043037, + 0.01812809705734253, + 0.011766789481043816, + 0.007433163467794657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.03959393873810768 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012563489377498627 + ], + "opposite_logprob_tokens": [ + -4.387563705444336 + ], + "logprob_reward": -0.012563489377498627, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012563489377498627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1258322149515152, + 0.05312449857592583, + 0.05993339419364929, + 0.03157956153154373, + 0.05181770771741867, + 0.07266837358474731, + 0.04799884557723999, + 0.05713118240237236, + 0.09652790427207947, + 0.4033862054347992 + ], + "credit_map": [ + 0.01122629176825285, + 0.014278453774750233, + 0.014539310708642006, + 0.02059338614344597, + 0.011337646283209324, + 0.01874394528567791, + 0.018273944035172462, + 0.01683925837278366, + 0.004352801013737917, + 0.007044302299618721, + 0.005512922536581755, + 0.008313735015690327, + 0.008763902820646763, + 0.00748857669532299, + 0.007890795357525349, + 0.0037574656307697296, + 0.005508935544639826, + 0.007371156942099333, + 0.006429459434002638, + 0.008439234457910061, + 0.008947540074586868, + 0.006524388212710619, + 0.012417142279446125, + 0.004295541439205408, + 0.004034585785120726, + 0.0031392553355544806, + 0.0037154213059693575, + 0.0032686134800314903, + 0.005365800578147173, + 0.004084295127540827, + 0.005005579441785812, + 0.002966011641547084, + 0.008352693170309067, + 0.007008193992078304, + 0.013286405242979527, + 0.004125190433114767, + 0.003179705236107111, + 0.004638137761503458, + 0.004176622722297907, + 0.007050763349980116, + 0.015172107145190239, + 0.00773801701143384, + 0.013685716316103935, + 0.006697813048958778, + 0.005441052373498678, + 0.0059329369105398655, + 0.008693968877196312, + 0.009306766092777252, + 0.007867010310292244, + 0.005737696308642626, + 0.007741711102426052, + 0.006075648590922356, + 0.005239867139607668, + 0.004072199109941721, + 0.005328197032213211, + 0.005936521105468273, + 0.012141459621489048, + 0.007719699293375015, + 0.009079080075025558, + 0.006019307766109705, + 0.006592720281332731, + 0.0044093746691942215, + 0.00566093111410737, + 0.00550860958173871, + 0.010204263962805271, + 0.009018916636705399, + 0.012927894480526447, + 0.013994372449815273, + 0.010399307124316692, + 0.014056273736059666, + 0.009346304461359978, + 0.016580572351813316, + 0.06424059718847275, + 0.08417075872421265, + 0.06868419796228409, + 0.02447211556136608, + 0.03205466642975807, + 0.08271385729312897, + 0.02286561205983162, + 0.024184439331293106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.012563489377498627 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786602109670639 + ], + "opposite_logprob_tokens": [ + -4.634786605834961 + ], + "logprob_reward": -0.009786602109670639, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786602109670639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22696936130523682, + 0.06377825140953064, + 0.03235291317105293, + 0.05576292425394058, + 0.06157776340842247, + 0.0827741026878357, + 0.03952506184577942, + 0.21232107281684875, + 0.1862497180700302, + 0.03868882358074188 + ], + "credit_map": [ + 0.0215133186429739, + 0.026032580062747, + 0.03270486742258072, + 0.018074549734592438, + 0.018960272893309593, + 0.05429437756538391, + 0.01883324421942234, + 0.03655613958835602, + 0.0062942407093942165, + 0.006484761834144592, + 0.009696447290480137, + 0.004224351141601801, + 0.004647550173103809, + 0.015213673003017902, + 0.005814597010612488, + 0.011402622796595097, + 0.0027014072984457016, + 0.003908847458660603, + 0.004459680058062077, + 0.0034777396358549595, + 0.007158374879509211, + 0.00305367773398757, + 0.005067519377917051, + 0.0025256641674786806, + 0.008672406896948814, + 0.009944697842001915, + 0.008806809782981873, + 0.0039418232627213, + 0.004986110143363476, + 0.007506916299462318, + 0.006442411802709103, + 0.0054617454297840595, + 0.009645040147006512, + 0.010080118663609028, + 0.017465859651565552, + 0.004972395487129688, + 0.004497216083109379, + 0.005075912456959486, + 0.004629631992429495, + 0.005211586132645607, + 0.022565504536032677, + 0.012924537062644958, + 0.01864933781325817, + 0.006243184674531221, + 0.005404407158493996, + 0.0074066235683858395, + 0.005008228588849306, + 0.004572269506752491, + 0.005535243544727564, + 0.006081591360270977, + 0.0057196537964046, + 0.004858097527176142, + 0.003858051961287856, + 0.003492455231025815, + 0.004727787803858519, + 0.005252175033092499, + 0.052438195794820786, + 0.031551018357276917, + 0.03589605167508125, + 0.019216308370232582, + 0.019317369908094406, + 0.017051802948117256, + 0.014125187881290913, + 0.022725122049450874, + 0.021851858124136925, + 0.03557029366493225, + 0.022227829322218895, + 0.02318040281534195, + 0.01717280223965645, + 0.02116650901734829, + 0.016263708472251892, + 0.028816306963562965, + 0.006952394265681505, + 0.00577829172834754, + 0.007053449284285307, + 0.0027246649842709303, + 0.002660078462213278, + 0.0066782631911337376, + 0.002839624648913741, + 0.004002055153250694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.009786602109670639 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635258063673973 + ], + "opposite_logprob_tokens": [ + -3.651352643966675 + ], + "logprob_reward": -0.02635258063673973, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635258063673973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10987969487905502, + 0.06970654428005219, + 0.03203346207737923, + 0.018326107412576675, + 0.08536727726459503, + 0.1989700198173523, + 0.0583888478577137, + 0.049034059047698975, + 0.15669776499271393, + 0.22159621119499207 + ], + "credit_map": [ + 0.009152896702289581, + 0.013479248620569706, + 0.012973722070455551, + 0.015717005357146263, + 0.010729248635470867, + 0.014485854655504227, + 0.017965205013751984, + 0.015376515686511993, + 0.005975780077278614, + 0.007305819075554609, + 0.007978693582117558, + 0.008231458254158497, + 0.010096976533532143, + 0.007732212543487549, + 0.012975912541151047, + 0.009409695863723755, + 0.0029679888393729925, + 0.003297478426247835, + 0.0037424147594720125, + 0.005197847727686167, + 0.00511200400069356, + 0.003308248473331332, + 0.005589888896793127, + 0.002817591419443488, + 0.0016817012801766396, + 0.0018604050856083632, + 0.001925092888996005, + 0.0025321159046143293, + 0.0022767267655581236, + 0.0020797457545995712, + 0.003231178503483534, + 0.002739140996709466, + 0.013428684324026108, + 0.007859299890697002, + 0.015999356284737587, + 0.007742611691355705, + 0.00927057396620512, + 0.0052616470493376255, + 0.011323593556880951, + 0.014481518417596817, + 0.016629144549369812, + 0.01746268756687641, + 0.02825900726020336, + 0.027215449139475822, + 0.014451284892857075, + 0.016854841262102127, + 0.03126278519630432, + 0.04683482274413109, + 0.008712238632142544, + 0.00699111120775342, + 0.009249776601791382, + 0.00746308034285903, + 0.0059688459150493145, + 0.0045421188697218895, + 0.007215969730168581, + 0.008245709352195263, + 0.007839947938919067, + 0.0045435610227286816, + 0.007299234624952078, + 0.005724808666855097, + 0.005130582023411989, + 0.004710026551038027, + 0.0060364301316440105, + 0.00774946715682745, + 0.017062218859791756, + 0.022776851430535316, + 0.020991157740354538, + 0.02348455786705017, + 0.016087381169199944, + 0.02102893404662609, + 0.012353186495602131, + 0.022913478314876556, + 0.037459518760442734, + 0.04725320637226105, + 0.038154229521751404, + 0.014494862407445908, + 0.013833267614245415, + 0.04753666743636131, + 0.008630792610347271, + 0.014233664609491825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.02635258063673973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000903785228729 + ], + "opposite_logprob_tokens": [ + -2.6950089931488037 + ], + "logprob_reward": -0.07000903785228729, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000903785228729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07750505954027176, + 0.02540895901620388, + 0.018978562206029892, + 0.017168991267681122, + 0.049630291759967804, + 0.08189324289560318, + 0.06352140754461288, + 0.17372991144657135, + 0.13531659543514252, + 0.35684695839881897 + ], + "credit_map": [ + 0.006307519972324371, + 0.006825858727097511, + 0.010094896890223026, + 0.010970509611070156, + 0.004920893348753452, + 0.015381732024252415, + 0.011232869699597359, + 0.01177077554166317, + 0.0022499829065054655, + 0.003126690862700343, + 0.0034983365330845118, + 0.0029444736428558826, + 0.002605245215818286, + 0.004505890421569347, + 0.003238365985453129, + 0.003239974146708846, + 0.0013686763122677803, + 0.0021517816931009293, + 0.0026134185027331114, + 0.0029835521709173918, + 0.0033813845366239548, + 0.002133652800694108, + 0.0029215198010206223, + 0.0014245767379179597, + 0.0025375878904014826, + 0.0023215310648083687, + 0.0031684942077845335, + 0.001016507507301867, + 0.002101818798109889, + 0.0020160130225121975, + 0.0020926741417497396, + 0.0019143649842590094, + 0.007109405938535929, + 0.007594716269522905, + 0.01205672137439251, + 0.004588352981954813, + 0.004894127603620291, + 0.00428757956251502, + 0.004893245175480843, + 0.004206136800348759, + 0.011618060991168022, + 0.010578692890703678, + 0.010867777280509472, + 0.00977801438421011, + 0.0072970036417245865, + 0.007749134209007025, + 0.009297511540353298, + 0.014707051217556, + 0.007900645956397057, + 0.009266705252230167, + 0.008917735889554024, + 0.008386855944991112, + 0.007405366748571396, + 0.005559931509196758, + 0.006607914809137583, + 0.0094762509688735, + 0.03695129230618477, + 0.02278522029519081, + 0.02711600624024868, + 0.019305812194943428, + 0.017955971881747246, + 0.015808148309588432, + 0.015179584734141827, + 0.01862788386642933, + 0.014235162176191807, + 0.02262365259230137, + 0.015614520758390427, + 0.0170766431838274, + 0.014091886579990387, + 0.014260229654610157, + 0.01586013473570347, + 0.021554358303546906, + 0.06112179905176163, + 0.07460270076990128, + 0.06210874766111374, + 0.021714486181735992, + 0.020769627764821053, + 0.07853282243013382, + 0.015576043166220188, + 0.022420713678002357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.07000903785228729 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sleek, eye-level tracking shot maintains a tight close-up of a short, nervous Hispanic teenage boy as he walks across the spacious, polished wooden gymnasium floor. He has dark, wavy hair, neatly combed, but a hint of nervous dampness lingers on his cheeks. He wears a light blue cotton dress shirt, buttoned stiffly at the collar, its slightly oversized fit accentuating his slender frame. His jaw is set, his gaze fixed straight ahead, his gait stiff and unnatural, his arms tucked tightly to his sides. The high-gloss varnish of the floor catches the cool, even glare of the overhead fluorescent lights, creating a long, distorted reflection. He deliberately approaches a gray metal folding chair that stands alone in an empty space, mechanically turning and carefully positioning himself on the edge of the seat. His posture remains erect, and his hands immediately grasp his knees, creating an atmosphere of intense anxiety and self-conscious vulnerability.", + "seed": 17440, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_13/index_5348_seed_17440_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04081710800528526, + "vlm_logprob": -0.04081710800528526, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.04081710800528526, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02132837474346161 + ], + "opposite_logprob_tokens": [ + -4.146328449249268 + ], + "logprob_reward": -0.02132837474346161, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02132837474346161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2046968787908554, + 0.0922575369477272, + 0.09681215137243271, + 0.07470790296792984, + 0.10796927660703659, + 0.07697810232639313, + 0.07797332108020782, + 0.09032284468412399, + 0.1372520625591278, + 0.04102996364235878 + ], + "credit_map": [ + 0.025064295157790184, + 0.01996598020195961, + 0.03879379481077194, + 0.019404355436563492, + 0.021251626312732697, + 0.02946246974170208, + 0.021296018734574318, + 0.029458321630954742, + 0.009806779213249683, + 0.019987430423498154, + 0.01334846019744873, + 0.010923389345407486, + 0.008433992974460125, + 0.012582281604409218, + 0.007959406822919846, + 0.00921578798443079, + 0.021703168749809265, + 0.013376731425523758, + 0.009676406159996986, + 0.005554792005568743, + 0.015170572325587273, + 0.01604718156158924, + 0.0055017657577991486, + 0.009781530126929283, + 0.011737565509974957, + 0.00912125501781702, + 0.01043267548084259, + 0.005764833651483059, + 0.006621594075113535, + 0.010936304926872253, + 0.006900979671627283, + 0.013192690908908844, + 0.01296159066259861, + 0.012688502669334412, + 0.012885917909443378, + 0.012179593555629253, + 0.014416491612792015, + 0.014915157109498978, + 0.012439055368304253, + 0.015482965856790543, + 0.018380407243967056, + 0.006886534858494997, + 0.010216571390628815, + 0.00533887417986989, + 0.008084564469754696, + 0.012703796848654747, + 0.004857499618083239, + 0.010509851388633251, + 0.012225783430039883, + 0.008217377588152885, + 0.014190442860126495, + 0.009769543074071407, + 0.004657685291022062, + 0.010799976997077465, + 0.007173519115895033, + 0.010938985273241997, + 0.012510222382843494, + 0.010416672565042973, + 0.013500778935849667, + 0.006288386415690184, + 0.015365133993327618, + 0.014636635780334473, + 0.007042217534035444, + 0.010562787763774395, + 0.02486131340265274, + 0.008593236096203327, + 0.023861879482865334, + 0.01130511611700058, + 0.011286670342087746, + 0.036205120384693146, + 0.007440011482685804, + 0.013698708266019821, + 0.00792553648352623, + 0.003387506352737546, + 0.007528382353484631, + 0.0029518064111471176, + 0.0035442488733679056, + 0.009220579639077187, + 0.0028158873319625854, + 0.0036560113076120615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.02132837474346161 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421281136572361 + ], + "opposite_logprob_tokens": [ + -4.264212608337402 + ], + "logprob_reward": -0.01421281136572361, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421281136572361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1321050524711609, + 0.07692395150661469, + 0.0919671505689621, + 0.08467607945203781, + 0.0981905609369278, + 0.0678824856877327, + 0.08885445445775986, + 0.20413269102573395, + 0.10335364937782288, + 0.05191393569111824 + ], + "credit_map": [ + 0.02086438611149788, + 0.01145906001329422, + 0.030080605298280716, + 0.010969499126076698, + 0.01068584993481636, + 0.017550257965922356, + 0.010881428606808186, + 0.01961396262049675, + 0.008391868323087692, + 0.01469546090811491, + 0.01110281515866518, + 0.007717772852629423, + 0.004906493704766035, + 0.010541001334786415, + 0.010576852597296238, + 0.008991684764623642, + 0.015384943224489689, + 0.00825658068060875, + 0.017767371609807014, + 0.007193170953541994, + 0.008137227036058903, + 0.01156875304877758, + 0.006347427144646645, + 0.017311671748757362, + 0.016994379460811615, + 0.008709313347935677, + 0.009919748641550541, + 0.007054693531244993, + 0.007860638201236725, + 0.01479886844754219, + 0.007806251756846905, + 0.011532188393175602, + 0.012257887050509453, + 0.010410087183117867, + 0.007016817107796669, + 0.006591149605810642, + 0.016539230942726135, + 0.02361706644296646, + 0.006885215174406767, + 0.014873096719384193, + 0.012982979416847229, + 0.010449185967445374, + 0.005261613056063652, + 0.005748503841459751, + 0.008599689230322838, + 0.011004890315234661, + 0.005851560737937689, + 0.007984056137502193, + 0.01408529281616211, + 0.00914721004664898, + 0.016305996105074883, + 0.008998604491353035, + 0.005829864181578159, + 0.014591681770980358, + 0.00974628422409296, + 0.010149515233933926, + 0.02801741287112236, + 0.020874226465821266, + 0.0378744937479496, + 0.015586374327540398, + 0.02185671031475067, + 0.042250897735357285, + 0.014978397637605667, + 0.022694159299135208, + 0.019444238394498825, + 0.008463085629045963, + 0.018229631707072258, + 0.005907765589654446, + 0.007254833355545998, + 0.02957640029489994, + 0.005252152215689421, + 0.00922553800046444, + 0.008260461501777172, + 0.006377766374498606, + 0.006981879472732544, + 0.0035302219912409782, + 0.006000080611556768, + 0.010239090770483017, + 0.004409046843647957, + 0.006115385331213474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.01421281136572361 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547360733151436 + ], + "opposite_logprob_tokens": [ + -4.387547492980957 + ], + "logprob_reward": -0.012547360733151436, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547360733151436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1808018982410431, + 0.034128908067941666, + 0.04512709006667137, + 0.022150808945298195, + 0.1412891000509262, + 0.04856746271252632, + 0.22705289721488953, + 0.1699957549571991, + 0.08739950507879257, + 0.04348656162619591 + ], + "credit_map": [ + 0.026135388761758804, + 0.021957023069262505, + 0.03670056164264679, + 0.01718733087182045, + 0.014306090772151947, + 0.023240884765982628, + 0.016026834025979042, + 0.025247791782021523, + 0.0034574856981635094, + 0.005009013693779707, + 0.005484733264893293, + 0.0035781152546405792, + 0.0025253882631659508, + 0.005272874608635902, + 0.003871978959068656, + 0.0049293190240859985, + 0.01028115302324295, + 0.004570738412439823, + 0.007106192409992218, + 0.0025193230248987675, + 0.004465478006750345, + 0.007061872165650129, + 0.002571940654888749, + 0.006550397723913193, + 0.0021443343721330166, + 0.0024748698342591524, + 0.003999701701104641, + 0.0017279558815062046, + 0.002586161717772484, + 0.0031256258953362703, + 0.002160358941182494, + 0.00393180176615715, + 0.01681685820221901, + 0.012558266520500183, + 0.010378433391451836, + 0.007269012276083231, + 0.027330361306667328, + 0.0352083258330822, + 0.008147967047989368, + 0.023579878732562065, + 0.011359915137290955, + 0.005396271590143442, + 0.006884891539812088, + 0.004357255529612303, + 0.004133084788918495, + 0.006561103276908398, + 0.0037070834077894688, + 0.006167855113744736, + 0.0594346821308136, + 0.03078008070588112, + 0.04555569961667061, + 0.0157411340624094, + 0.012598293833434582, + 0.034011490643024445, + 0.010461385361850262, + 0.01847013458609581, + 0.02181325852870941, + 0.01626996323466301, + 0.03330284357070923, + 0.012674547731876373, + 0.015875207260251045, + 0.039484694600105286, + 0.011555839329957962, + 0.01901940628886223, + 0.019432147964835167, + 0.0059576150961220264, + 0.015555552206933498, + 0.004948901943862438, + 0.006389649584889412, + 0.020430423319339752, + 0.00434014992788434, + 0.010345065966248512, + 0.008520863018929958, + 0.005085630342364311, + 0.006040314678102732, + 0.0027346559800207615, + 0.0037185747642070055, + 0.009393674321472645, + 0.003174956189468503, + 0.004817892797291279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.012547360733151436 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129482090473175 + ], + "opposite_logprob_tokens": [ + -2.237948179244995 + ], + "logprob_reward": -0.1129482090473175, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129482090473175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23183628916740417, + 0.044182658195495605, + 0.03209855407476425, + 0.0484846830368042, + 0.09977742284536362, + 0.0702693983912468, + 0.06971511244773865, + 0.19935911893844604, + 0.11321362107992172, + 0.09106317162513733 + ], + "credit_map": [ + 0.03723767772316933, + 0.029390182346105576, + 0.040112871676683426, + 0.017993619665503502, + 0.021103329956531525, + 0.04262042045593262, + 0.01644097827374935, + 0.026937203481793404, + 0.004341114312410355, + 0.008764163590967655, + 0.006100750528275967, + 0.0047905100509524345, + 0.0031764451414346695, + 0.006459327414631844, + 0.005153444595634937, + 0.005396904423832893, + 0.005016570910811424, + 0.003196087898686528, + 0.003365235636010766, + 0.00402462063357234, + 0.004079518839716911, + 0.0051778401248157024, + 0.0029707420617341995, + 0.004267937038093805, + 0.0058960379101336, + 0.007452342193573713, + 0.005456553306430578, + 0.0071404315531253815, + 0.005945651791989803, + 0.0058670188300311565, + 0.005710988771170378, + 0.005015656352043152, + 0.011121847666800022, + 0.00978055689483881, + 0.008272786624729633, + 0.006571162026375532, + 0.017686260864138603, + 0.02171815000474453, + 0.009397021494805813, + 0.015229628421366215, + 0.013502049259841442, + 0.009671851061284542, + 0.00704267667606473, + 0.006783348973840475, + 0.005361263174563646, + 0.013214665465056896, + 0.005573729053139687, + 0.009119812399148941, + 0.014112640172243118, + 0.007158858235925436, + 0.013945487327873707, + 0.006156178656965494, + 0.003776134690269828, + 0.01248222403228283, + 0.0038577388040721416, + 0.00822584517300129, + 0.02752630226314068, + 0.018341712653636932, + 0.037222567945718765, + 0.01475005317479372, + 0.019552985206246376, + 0.04338373243808746, + 0.015776753425598145, + 0.02280501089990139, + 0.021092427894473076, + 0.009861588478088379, + 0.017326749861240387, + 0.010946790687739849, + 0.008971970528364182, + 0.022144125774502754, + 0.008600101806223392, + 0.014269869774580002, + 0.020055098459124565, + 0.00893991719931364, + 0.01328058447688818, + 0.005575019866228104, + 0.005910683888942003, + 0.021284889429807663, + 0.006692752707749605, + 0.009324220940470695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.1129482090473175 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043048784136772156 + ], + "opposite_logprob_tokens": [ + -3.168048858642578 + ], + "logprob_reward": -0.043048784136772156, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043048784136772156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12242821604013443, + 0.03649504855275154, + 0.08857370913028717, + 0.05061456188559532, + 0.06941971182823181, + 0.0477219820022583, + 0.182521790266037, + 0.11103878915309906, + 0.17510312795639038, + 0.1160830557346344 + ], + "credit_map": [ + 0.01892976276576519, + 0.013490062206983566, + 0.018357660621404648, + 0.012029807083308697, + 0.01091075874865055, + 0.024090107530355453, + 0.011670618318021297, + 0.012949440628290176, + 0.0035843795631080866, + 0.007200232241302729, + 0.004317256156355143, + 0.003690554527565837, + 0.0027630648110061884, + 0.006198802962899208, + 0.0038966876454651356, + 0.0048440732061862946, + 0.016246071085333824, + 0.011297231540083885, + 0.008284658193588257, + 0.007076177280396223, + 0.013808295130729675, + 0.015762289986014366, + 0.005860030651092529, + 0.010238955728709698, + 0.007904178462922573, + 0.007149212062358856, + 0.006190809886902571, + 0.0050935023464262486, + 0.005620039999485016, + 0.006576457992196083, + 0.004499760922044516, + 0.007580604404211044, + 0.007365592755377293, + 0.007040840107947588, + 0.004913451615720987, + 0.007424888666719198, + 0.011687761172652245, + 0.013506042771041393, + 0.00737298047170043, + 0.010108167305588722, + 0.009919848293066025, + 0.006694141309708357, + 0.005948995240032673, + 0.005183883477002382, + 0.0036677487660199404, + 0.005424658767879009, + 0.004954051226377487, + 0.0059286621399223804, + 0.034563761204481125, + 0.022721366956830025, + 0.034516919404268265, + 0.016244271770119667, + 0.014535117894411087, + 0.034767214208841324, + 0.011850387789309025, + 0.013322754763066769, + 0.01303201075643301, + 0.01041486021131277, + 0.02007891610264778, + 0.007986178621649742, + 0.01640046015381813, + 0.022411102429032326, + 0.008453233167529106, + 0.01226203702390194, + 0.02851421758532524, + 0.013143126852810383, + 0.0266000647097826, + 0.01799388974905014, + 0.015061973594129086, + 0.03902627155184746, + 0.014465986751019955, + 0.020297599956393242, + 0.02115206979215145, + 0.012399557046592236, + 0.017559953033924103, + 0.007782577071338892, + 0.010804235003888607, + 0.02351911924779415, + 0.007764920592308044, + 0.015100633725523949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.043048784136772156 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sleek, eye-level tracking shot maintains a tight close-up of a short, nervous Hispanic teenage boy as he walks across the spacious, polished wooden gymnasium floor. He has dark, wavy hair, neatly combed, but a hint of nervous dampness lingers on his cheeks. He wears a light blue cotton dress shirt, buttoned stiffly at the collar, its slightly oversized fit accentuating his slender frame. His jaw is set, his gaze fixed straight ahead, his gait stiff and unnatural, his arms tucked tightly to his sides. The high-gloss varnish of the floor catches the cool, even glare of the overhead fluorescent lights, creating a long, distorted reflection. He deliberately approaches a gray metal folding chair that stands alone in an empty space, mechanically turning and carefully positioning himself on the edge of the seat. His posture remains erect, and his hands immediately grasp his knees, creating an atmosphere of intense anxiety and self-conscious vulnerability.", + "seed": 17441, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_13/index_5348_seed_17441_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04914089562371373, + "vlm_logprob": -0.04914089562371373, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.04914089562371373, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0527365505695343 + ], + "opposite_logprob_tokens": [ + -3.177736520767212 + ], + "logprob_reward": -0.0527365505695343, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0527365505695343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26864153146743774, + 0.06768173724412918, + 0.12930259108543396, + 0.05076586455106735, + 0.11652248352766037, + 0.11278792470693588, + 0.06792252510786057, + 0.0811462253332138, + 0.0607173852622509, + 0.04451172426342964 + ], + "credit_map": [ + 0.029584065079689026, + 0.04303916171193123, + 0.02857068181037903, + 0.042525772005319595, + 0.026479922235012054, + 0.030867924913764, + 0.03211386129260063, + 0.035460151731967926, + 0.0074238902889192104, + 0.006433126982301474, + 0.008691319264471531, + 0.010074969381093979, + 0.009804720059037209, + 0.010223799385130405, + 0.00846247561275959, + 0.006567446049302816, + 0.01598399132490158, + 0.012098020873963833, + 0.023851757869124413, + 0.012413564138114452, + 0.01563156396150589, + 0.016635194420814514, + 0.01618560589849949, + 0.01650290936231613, + 0.00547928549349308, + 0.008132290095090866, + 0.004105101805180311, + 0.004377793986350298, + 0.008106100372970104, + 0.005194647237658501, + 0.009029112756252289, + 0.006341536995023489, + 0.013420244678854942, + 0.014804476872086525, + 0.019790899008512497, + 0.012244924902915955, + 0.013760182075202465, + 0.02058027870953083, + 0.011702983640134335, + 0.01021850947290659, + 0.027295170351862907, + 0.009573737159371376, + 0.009030145592987537, + 0.009789619594812393, + 0.009515243582427502, + 0.029260024428367615, + 0.010631325654685497, + 0.007692674640566111, + 0.005849746055901051, + 0.005271376110613346, + 0.011850863695144653, + 0.007994826883077621, + 0.00581576069816947, + 0.012166500091552734, + 0.008657869882881641, + 0.010315585881471634, + 0.006808455567806959, + 0.005872424691915512, + 0.009872938506305218, + 0.006868897005915642, + 0.010612384416162968, + 0.012567059136927128, + 0.006945361383259296, + 0.02159871719777584, + 0.005904939025640488, + 0.006621262524276972, + 0.00934354867786169, + 0.006854563485831022, + 0.012155881151556969, + 0.00770528893917799, + 0.0056862980127334595, + 0.006445610895752907, + 0.004606458358466625, + 0.005899069365113974, + 0.007372376974672079, + 0.008410075679421425, + 0.00444032670930028, + 0.005893223453313112, + 0.0045328945852816105, + 0.0033573019318282604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0527365505695343 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018213994801044464 + ], + "opposite_logprob_tokens": [ + -4.018214225769043 + ], + "logprob_reward": -0.018213994801044464, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018213994801044464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2813803255558014, + 0.12691494822502136, + 0.13576465845108032, + 0.06209799647331238, + 0.05435299500823021, + 0.024504326283931732, + 0.17052043974399567, + 0.048611968755722046, + 0.04292465001344681, + 0.05292769521474838 + ], + "credit_map": [ + 0.026551295071840286, + 0.024569708853960037, + 0.05632452294230461, + 0.02674570307135582, + 0.025789879262447357, + 0.025880984961986542, + 0.03099912218749523, + 0.06451909989118576, + 0.017182163894176483, + 0.012565573677420616, + 0.015017297118902206, + 0.01661696285009384, + 0.019118232652544975, + 0.01627293974161148, + 0.011203330010175705, + 0.0189384575933218, + 0.022279653698205948, + 0.012671255506575108, + 0.017490344122052193, + 0.013089342042803764, + 0.018824659287929535, + 0.021092724055051804, + 0.014339223504066467, + 0.01597744971513748, + 0.0068627605214715, + 0.010865900665521622, + 0.004331097938120365, + 0.006928878370672464, + 0.008012101985514164, + 0.008374557830393314, + 0.010481176897883415, + 0.006241519935429096, + 0.008251383900642395, + 0.010565822012722492, + 0.00771335931494832, + 0.00534207746386528, + 0.006259303539991379, + 0.007370379753410816, + 0.0043797907419502735, + 0.004470879677683115, + 0.003960001282393932, + 0.002387403044849634, + 0.003467126516625285, + 0.003058947855606675, + 0.0035796344745904207, + 0.002365044318139553, + 0.0032201632857322693, + 0.0024660038761794567, + 0.009806329384446144, + 0.006072447169572115, + 0.05223654583096504, + 0.00925252865999937, + 0.013059102930128574, + 0.0545729398727417, + 0.007687089499086142, + 0.01783345267176628, + 0.0035306578502058983, + 0.002206655917689204, + 0.009539362974464893, + 0.004022129811346531, + 0.003673323430120945, + 0.009756986051797867, + 0.0027272533625364304, + 0.013155600987374783, + 0.0027579644229263067, + 0.0034251725301146507, + 0.006416836753487587, + 0.009507975541055202, + 0.003909118939191103, + 0.0062906788662076, + 0.004421971272677183, + 0.006194929592311382, + 0.0040996018797159195, + 0.0043238867074251175, + 0.006346708629280329, + 0.00707068620249629, + 0.008562413975596428, + 0.006330925039947033, + 0.006834174506366253, + 0.009359297342598438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.018213994801044464 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004644202534109354 + ], + "opposite_logprob_tokens": [ + -5.379644393920898 + ], + "logprob_reward": -0.004644202534109354, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004644202534109354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21988755464553833, + 0.045679982751607895, + 0.0372796505689621, + 0.08732647448778152, + 0.1822669357061386, + 0.17434336245059967, + 0.10364512354135513, + 0.030664294958114624, + 0.046582963317632675, + 0.07232367992401123 + ], + "credit_map": [ + 0.019885413348674774, + 0.01218374166637659, + 0.06768856197595596, + 0.012083903886377811, + 0.009228565730154514, + 0.014726666733622551, + 0.011113053187727928, + 0.07297764718532562, + 0.004679869394749403, + 0.004591894336044788, + 0.007606144528836012, + 0.004727994557470083, + 0.004855447448790073, + 0.008441107347607613, + 0.0043448470532894135, + 0.006432674825191498, + 0.0035141785629093647, + 0.003754701931029558, + 0.007033989764750004, + 0.0034626605920493603, + 0.006263046991080046, + 0.005186351481825113, + 0.003131932346150279, + 0.004932786803692579, + 0.011048872023820877, + 0.014213806949555874, + 0.01258105132728815, + 0.005996785592287779, + 0.010116956196725368, + 0.016280509531497955, + 0.008041175082325935, + 0.009047313593327999, + 0.026547981426119804, + 0.03499656170606613, + 0.032017018646001816, + 0.019562432542443275, + 0.015807611867785454, + 0.025831975042819977, + 0.010352452285587788, + 0.017150895670056343, + 0.05318869650363922, + 0.012170948088169098, + 0.012218380346894264, + 0.011470852419734001, + 0.012905485928058624, + 0.04992689937353134, + 0.010478880256414413, + 0.011983207426965237, + 0.012545966543257236, + 0.008921008557081223, + 0.026468954980373383, + 0.008075492456555367, + 0.006749467924237251, + 0.0236166063696146, + 0.0069802249781787395, + 0.010287384502589703, + 0.0043731145560741425, + 0.004872969351708889, + 0.004569276235997677, + 0.0029603689908981323, + 0.0038200037088245153, + 0.0045374599285423756, + 0.0020232328679412603, + 0.0035078665241599083, + 0.005283959209918976, + 0.008843264542520046, + 0.008359759114682674, + 0.005571957211941481, + 0.0038334564305841923, + 0.006784086115658283, + 0.0029519409872591496, + 0.004954536911100149, + 0.00755739863961935, + 0.010207253508269787, + 0.010410293936729431, + 0.015695763751864433, + 0.0063765146769583225, + 0.012102790176868439, + 0.005219589918851852, + 0.004754071123898029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.004644202534109354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307321831583977 + ], + "opposite_logprob_tokens": [ + -3.1680731773376465 + ], + "logprob_reward": -0.04307321831583977, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307321831583977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13213087618350983, + 0.0535748228430748, + 0.0846666693687439, + 0.05806119367480278, + 0.0885641947388649, + 0.21559877693653107, + 0.14431868493556976, + 0.08589645475149155, + 0.04690670222043991, + 0.09028168767690659 + ], + "credit_map": [ + 0.013184705749154091, + 0.01304907351732254, + 0.025032181292772293, + 0.012714151293039322, + 0.010710126720368862, + 0.019062504172325134, + 0.010261268354952335, + 0.02811683900654316, + 0.00654470082372427, + 0.005747755523771048, + 0.009976295754313469, + 0.004693862982094288, + 0.004829880781471729, + 0.008759576827287674, + 0.004119899123907089, + 0.008902844041585922, + 0.013195865787565708, + 0.009842243045568466, + 0.010390105657279491, + 0.006503691431134939, + 0.012869272381067276, + 0.013980566523969173, + 0.006826907396316528, + 0.011058004572987556, + 0.007716761436313391, + 0.008770816028118134, + 0.0025120677892118692, + 0.005679271183907986, + 0.008274202235043049, + 0.006985931191593409, + 0.011142488569021225, + 0.006979653146117926, + 0.018848372623324394, + 0.024103106930851936, + 0.013182789087295532, + 0.008118622936308384, + 0.006280309520661831, + 0.007114740088582039, + 0.0042816209606826305, + 0.006634625140577555, + 0.07013539224863052, + 0.011617896147072315, + 0.012033872306346893, + 0.010291311889886856, + 0.01727570965886116, + 0.06789533793926239, + 0.011340998113155365, + 0.015008234418928623, + 0.012343082576990128, + 0.010950203984975815, + 0.022184807807207108, + 0.0163352582603693, + 0.012262103147804737, + 0.029317500069737434, + 0.01867012493312359, + 0.022255586460232735, + 0.004818430170416832, + 0.005636346526443958, + 0.01499269064515829, + 0.010513831861317158, + 0.01787901483476162, + 0.014432156458497047, + 0.006770310923457146, + 0.010853666812181473, + 0.002689865417778492, + 0.005531421862542629, + 0.006776391062885523, + 0.008560438640415668, + 0.004662034101784229, + 0.00716622406616807, + 0.005007897969335318, + 0.006512423511594534, + 0.00497967517003417, + 0.005066188983619213, + 0.01182804349809885, + 0.015680363401770592, + 0.014446992427110672, + 0.010376031510531902, + 0.011761418543756008, + 0.016142960637807846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.04307321831583977 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703651189804077 + ], + "opposite_logprob_tokens": [ + -2.1270365715026855 + ], + "logprob_reward": -0.12703651189804077, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703651189804077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3196960389614105, + 0.04301541671156883, + 0.04152429476380348, + 0.049353502690792084, + 0.0805472731590271, + 0.11142175644636154, + 0.10409758239984512, + 0.07621343433856964, + 0.08480240404605865, + 0.08932831138372421 + ], + "credit_map": [ + 0.030247535556554794, + 0.016344614326953888, + 0.09981229156255722, + 0.016315871849656105, + 0.011675212532281876, + 0.020656004548072815, + 0.013045620173215866, + 0.11159887164831161, + 0.006086778827011585, + 0.004355183336883783, + 0.006671647075563669, + 0.004807989578694105, + 0.004988048225641251, + 0.007301585283130407, + 0.004411155823618174, + 0.004393028561025858, + 0.005869364831596613, + 0.00551966018974781, + 0.006884789559990168, + 0.004763831850141287, + 0.0048934719525277615, + 0.00357020553201437, + 0.005158623214811087, + 0.004864344373345375, + 0.0069968379102647305, + 0.008738028816878796, + 0.004306151531636715, + 0.005731948185712099, + 0.007518484722822905, + 0.004561058711260557, + 0.005130404606461525, + 0.006370586808770895, + 0.011175079271197319, + 0.008360283449292183, + 0.018715349957346916, + 0.011500726453959942, + 0.008738979697227478, + 0.009008171036839485, + 0.0053790900856256485, + 0.0076695880852639675, + 0.02623726986348629, + 0.007116515189409256, + 0.0072294920682907104, + 0.006920378655195236, + 0.009984207339584827, + 0.03206492215394974, + 0.013650038279592991, + 0.008218935690820217, + 0.007036683149635792, + 0.005716840736567974, + 0.023636963218450546, + 0.008535905741155148, + 0.010697016492486, + 0.023530814796686172, + 0.010894585400819778, + 0.014048767276108265, + 0.006011464633047581, + 0.005090252496302128, + 0.009427180513739586, + 0.00834941677749157, + 0.011885820887982845, + 0.011088134720921516, + 0.007233497221022844, + 0.017127661034464836, + 0.0072249420918524265, + 0.00802813470363617, + 0.014508054591715336, + 0.011338858865201473, + 0.010397697798907757, + 0.01202690415084362, + 0.007754702121019363, + 0.013523108325898647, + 0.007521582301706076, + 0.011286222375929356, + 0.013186631724238396, + 0.014484548009932041, + 0.0123689454048872, + 0.010505973361432552, + 0.008948293514549732, + 0.011026118882000446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.12703651189804077 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sleek, eye-level tracking shot maintains a tight close-up of a short, nervous Hispanic teenage boy as he walks across the spacious, polished wooden gymnasium floor. He has dark, wavy hair, neatly combed, but a hint of nervous dampness lingers on his cheeks. He wears a light blue cotton dress shirt, buttoned stiffly at the collar, its slightly oversized fit accentuating his slender frame. His jaw is set, his gaze fixed straight ahead, his gait stiff and unnatural, his arms tucked tightly to his sides. The high-gloss varnish of the floor catches the cool, even glare of the overhead fluorescent lights, creating a long, distorted reflection. He deliberately approaches a gray metal folding chair that stands alone in an empty space, mechanically turning and carefully positioning himself on the edge of the seat. His posture remains erect, and his hands immediately grasp his knees, creating an atmosphere of intense anxiety and self-conscious vulnerability.", + "seed": 17442, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_13/index_5348_seed_17442_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.023092705570161343, + "vlm_logprob": -0.023092705570161343, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.023092705570161343, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035048484802246094 + ], + "opposite_logprob_tokens": [ + -3.535048484802246 + ], + "logprob_reward": -0.035048484802246094, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.035048484802246094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8409300446510315, + 0.06689436733722687, + 0.007137666922062635, + 0.02421681024134159, + 0.0074115414172410965, + 0.0072240871377289295, + 0.017455844208598137, + 0.012600880116224289, + 0.006890504155308008, + 0.00923826638609171 + ], + "credit_map": [ + 0.057161640375852585, + 0.026088759303092957, + 0.346434086561203, + 0.03351447358727455, + 0.026285067200660706, + 0.2940639555454254, + 0.01891845464706421, + 0.038463592529296875, + 0.004698601551353931, + 0.0036376824136823416, + 0.023975545540452003, + 0.0041702971793711185, + 0.003029894782230258, + 0.019467981532216072, + 0.004329027608036995, + 0.003585333935916424, + 0.0008330506389029324, + 0.0007156279752962291, + 0.0009411554201506078, + 0.0010306719923391938, + 0.0008253539563156664, + 0.0008903499692678452, + 0.0010574505431577563, + 0.000844005960971117, + 0.0033641434274613857, + 0.0019492842257022858, + 0.005648259539157152, + 0.00245410343632102, + 0.001361858332529664, + 0.0039144959300756454, + 0.0022710724733769894, + 0.0032535921782255173, + 0.0009231138974428177, + 0.0008929927134886384, + 0.0012428259942680597, + 0.0010165873682126403, + 0.0005583190359175205, + 0.0006857942207716405, + 0.0012211407301947474, + 0.0008707677479833364, + 0.0011211051605641842, + 0.0008048617164604366, + 0.000903122709132731, + 0.0009346955339424312, + 0.0007827502558939159, + 0.000603015418164432, + 0.0009810772025957704, + 0.0010934590827673674, + 0.0028481928166002035, + 0.0017877033678814769, + 0.002355483127757907, + 0.002857705345377326, + 0.0011591330403462052, + 0.0029277000576257706, + 0.0016055390005931258, + 0.0019143864046782255, + 0.001022681943140924, + 0.0020527425222098827, + 0.0024585353676229715, + 0.0020752777345478535, + 0.0013472678838297725, + 0.0011082232231274247, + 0.0013682045973837376, + 0.0011679455637931824, + 0.0006481108721345663, + 0.0009953654371201992, + 0.0010311752557754517, + 0.0013305892935022712, + 0.0008326667011715472, + 0.0007526887929998338, + 0.0006220480427145958, + 0.0006778603419661522, + 0.0007023680373094976, + 0.002392062684521079, + 0.0016718513797968626, + 0.0013179739471524954, + 0.0008048720774240792, + 0.0011034889612346888, + 0.0006533956038765609, + 0.0005922538111917675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.035048484802246094 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012571494095027447 + ], + "opposite_logprob_tokens": [ + -4.387571334838867 + ], + "logprob_reward": -0.012571494095027447, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012571494095027447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7427355647087097, + 0.06765534728765488, + 0.015245988965034485, + 0.07328174263238907, + 0.007688465062528849, + 0.0322827585041523, + 0.024265572428703308, + 0.013195237144827843, + 0.011754459701478481, + 0.011894838884472847 + ], + "credit_map": [ + 0.05155227705836296, + 0.02063082344830036, + 0.31083112955093384, + 0.026627913117408752, + 0.02148866094648838, + 0.26699796319007874, + 0.012041336856782436, + 0.032565463334321976, + 0.00385475205257535, + 0.004834003746509552, + 0.021948402747511864, + 0.0057994406670331955, + 0.004445978906005621, + 0.01688896119594574, + 0.006423086393624544, + 0.00346072344109416, + 0.001576379407197237, + 0.00186557334382087, + 0.0016701904824003577, + 0.0020537502132356167, + 0.001637011650018394, + 0.002326769521459937, + 0.002011922188103199, + 0.002104392508044839, + 0.007970256730914116, + 0.004744180478155613, + 0.017250152304768562, + 0.007755965460091829, + 0.003423613728955388, + 0.010576724074780941, + 0.005931731779128313, + 0.0156291201710701, + 0.0008886443683877587, + 0.0008737352327443659, + 0.0010112228337675333, + 0.0011236798018217087, + 0.0005499761318787932, + 0.000807420292403549, + 0.0014395487960428, + 0.0009942371398210526, + 0.004816935397684574, + 0.0022740731947124004, + 0.0041955201886594296, + 0.002922662068158388, + 0.0026236779522150755, + 0.0026690210215747356, + 0.006780710536986589, + 0.006000159308314323, + 0.0040669203735888, + 0.0027695437893271446, + 0.0032176768872886896, + 0.004085539374500513, + 0.001742493943311274, + 0.0032766200602054596, + 0.002231636317446828, + 0.0028751399368047714, + 0.001266616745851934, + 0.0028671652544289827, + 0.0024962592869997025, + 0.001896731206215918, + 0.0010209508473053575, + 0.0010593398474156857, + 0.0010669281473383307, + 0.0015212458092719316, + 0.0013856621226295829, + 0.0014336653985083103, + 0.00173163041472435, + 0.0020953568164259195, + 0.0013403494376689196, + 0.0015078192809596658, + 0.001137673738412559, + 0.0011223024921491742, + 0.0011072780471295118, + 0.0012906886404380202, + 0.001508616958744824, + 0.0018862825818359852, + 0.002215215004980564, + 0.0012898786226287484, + 0.0013760823057964444, + 0.0012207970721647143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.012571494095027447 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820029877126217 + ], + "opposite_logprob_tokens": [ + -4.018200397491455 + ], + "logprob_reward": -0.01820029877126217, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820029877126217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6588493585586548, + 0.10230971872806549, + 0.016657788306474686, + 0.013570324517786503, + 0.01095307432115078, + 0.016744324937462807, + 0.04795597121119499, + 0.04265478253364563, + 0.06554215401411057, + 0.024762459099292755 + ], + "credit_map": [ + 0.046986065804958344, + 0.02201566845178604, + 0.2726452648639679, + 0.023243293166160583, + 0.020327795296907425, + 0.2334122359752655, + 0.00993574969470501, + 0.030283287167549133, + 0.008731186389923096, + 0.005990039557218552, + 0.031418900936841965, + 0.007091840263456106, + 0.006451688706874847, + 0.028667563572525978, + 0.007375647779554129, + 0.006582855712622404, + 0.00218599452637136, + 0.0025589619763195515, + 0.0026469703298062086, + 0.0018476912518963218, + 0.0018469432834535837, + 0.002236492931842804, + 0.0017671191599220037, + 0.001567615196108818, + 0.001016115420497954, + 0.0014908488374203444, + 0.0028146638069301844, + 0.0015996467554941773, + 0.0009712093160487711, + 0.003241336438804865, + 0.0009771068580448627, + 0.001459396560676396, + 0.0014378910418599844, + 0.0012110472889617085, + 0.0015456824330613017, + 0.001356229418888688, + 0.0008166144834831357, + 0.0010473349830135703, + 0.0013206592993810773, + 0.002217615256085992, + 0.0029699462465941906, + 0.0016155639896169305, + 0.0022673122584819794, + 0.0016273863147944212, + 0.0017235522391274571, + 0.0021446936298161745, + 0.00217692949809134, + 0.0022189412266016006, + 0.006833416409790516, + 0.003756466554477811, + 0.007416121196001768, + 0.004889968317002058, + 0.0036548003554344177, + 0.012156166136264801, + 0.0033741029910743237, + 0.005874929018318653, + 0.005375949665904045, + 0.007758236024528742, + 0.00542654050514102, + 0.008042865432798862, + 0.003479384584352374, + 0.0034429128281772137, + 0.0038366506341844797, + 0.0052922433242201805, + 0.00927183497697115, + 0.009366143494844437, + 0.013862863183021545, + 0.012097673490643501, + 0.006626510992646217, + 0.006117469165474176, + 0.004718142095953226, + 0.003481513587757945, + 0.003440680680796504, + 0.003789107548072934, + 0.004030460026115179, + 0.006082549225538969, + 0.0021949068177491426, + 0.001931745675392449, + 0.001995720202103257, + 0.0012972899712622166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.01820029877126217 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330123446881771 + ], + "opposite_logprob_tokens": [ + -3.773301124572754 + ], + "logprob_reward": -0.02330123446881771, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330123446881771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8525056838989258, + 0.03163354471325874, + 0.009250609204173088, + 0.01059896033257246, + 0.010911483317613602, + 0.017211472615599632, + 0.013151347637176514, + 0.029097743332386017, + 0.013591303490102291, + 0.012047812342643738 + ], + "credit_map": [ + 0.05903414636850357, + 0.02431461028754711, + 0.3577969968318939, + 0.030337508767843246, + 0.024597404524683952, + 0.30609139800071716, + 0.01228531077504158, + 0.03804831951856613, + 0.00187579239718616, + 0.0018964926712214947, + 0.01101883314549923, + 0.002239807741716504, + 0.0021539642475545406, + 0.008541540242731571, + 0.0021311643067747355, + 0.0017759520560503006, + 0.001183277927339077, + 0.001064371783286333, + 0.0011447991710156202, + 0.0010178260272368789, + 0.0011631646193563938, + 0.0012918751453980803, + 0.001022660406306386, + 0.0013626348227262497, + 0.0008933182107284665, + 0.0012046481715515256, + 0.0015662345103919506, + 0.0012488504871726036, + 0.0008949643815867603, + 0.0014789534034207463, + 0.0013566316338256001, + 0.001955360407009721, + 0.0014087478630244732, + 0.0014678968582302332, + 0.0009822060819715261, + 0.00148105644620955, + 0.0009152460261248052, + 0.0011424052063375711, + 0.001729808864183724, + 0.0017841164954006672, + 0.0024137659929692745, + 0.002099825069308281, + 0.0021420889534056187, + 0.002199748996645212, + 0.002255139872431755, + 0.0015148250386118889, + 0.0025645671412348747, + 0.002021513180807233, + 0.0016157657373696566, + 0.0010868953540921211, + 0.001975694438442588, + 0.0017440023366361856, + 0.0011938173556700349, + 0.0020523846615105867, + 0.0015285882400348783, + 0.0019542004447430372, + 0.002631682902574539, + 0.00386674958281219, + 0.0050063286907970905, + 0.004540866240859032, + 0.002760375617071986, + 0.003017183393239975, + 0.0037740827538073063, + 0.003500477410852909, + 0.0016717355465516448, + 0.001770476927049458, + 0.001988805830478668, + 0.0021906059700995684, + 0.0018143985653296113, + 0.001639596768654883, + 0.0012780632823705673, + 0.001237622112967074, + 0.0011492836056277156, + 0.0024471357464790344, + 0.0016656451625749469, + 0.001570481457747519, + 0.0012772122863680124, + 0.0013250896008685231, + 0.0013454088475555182, + 0.0012675567995756865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.02330123446881771 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a short Hispanic teenage boy with dark wavy hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a light blue cotton dress shirt that is buttoned stiffly at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place in a spacious gymnasium with a polished wooden floor reflecting overhead fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boy walk with a stiff, unnatural gait while keeping his arms tucked tightly to his sides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an eye-level tight close-up that tracks the boy as he approaches and sits on a gray metal folding chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026342015713453293 + ], + "opposite_logprob_tokens": [ + -3.6513419151306152 + ], + "logprob_reward": -0.026342015713453293, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026342015713453293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5568674206733704, + 0.07281584292650223, + 0.03834785893559456, + 0.04859718680381775, + 0.015210801735520363, + 0.07874373346567154, + 0.05336790159344673, + 0.043092623353004456, + 0.05800795927643776, + 0.03494869917631149 + ], + "credit_map": [ + 0.039161473512649536, + 0.015693342313170433, + 0.23283164203166962, + 0.020739559084177017, + 0.016619417816400528, + 0.19744837284088135, + 0.009433342143893242, + 0.02494027651846409, + 0.004759664181619883, + 0.004790253937244415, + 0.02656344696879387, + 0.0040862043388187885, + 0.0038339877501130104, + 0.020249594002962112, + 0.0036829994060099125, + 0.004849688615649939, + 0.006483947392553091, + 0.004883347079157829, + 0.008086313493549824, + 0.00355392973870039, + 0.003234204603359103, + 0.005656070541590452, + 0.0030329162254929543, + 0.0034171296283602715, + 0.006651564501225948, + 0.0030561573803424835, + 0.008570101112127304, + 0.0054798489436507225, + 0.002835774328559637, + 0.007894359529018402, + 0.004487037193030119, + 0.009622342884540558, + 0.001578702125698328, + 0.001588003127835691, + 0.0021554813720285892, + 0.001626635785214603, + 0.0012301933020353317, + 0.0016044750809669495, + 0.0033261175267398357, + 0.0021011929493397474, + 0.009389466606080532, + 0.005823513027280569, + 0.009662744589149952, + 0.009912151843309402, + 0.0071904477663338184, + 0.006646624766290188, + 0.014437003992497921, + 0.015681779012084007, + 0.00853817630559206, + 0.006484033074229956, + 0.007379284128546715, + 0.009333093650639057, + 0.004301039502024651, + 0.007094393949955702, + 0.004560333210974932, + 0.005677544511854649, + 0.004290651064366102, + 0.005210501607507467, + 0.008503585122525692, + 0.007553571835160255, + 0.004232517443597317, + 0.0037154166493564844, + 0.0052101886831223965, + 0.004376189783215523, + 0.005666317883878946, + 0.008503234945237637, + 0.008515791036188602, + 0.012379707768559456, + 0.007193886209279299, + 0.005613451823592186, + 0.005481363274157047, + 0.004654201678931713, + 0.004022499546408653, + 0.005491727031767368, + 0.005200878717005253, + 0.005147022195160389, + 0.005326804239302874, + 0.003840341465547681, + 0.0032002506777644157, + 0.0027191732078790665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.026342015713453293 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_14/index_tensor([7485, 7485])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_14/index_tensor([7485, 7485])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d73c6cba945551ae4b5e148b774a6d3b454490d7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_14/index_tensor([7485, 7485])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme macro-static shot delves into a microscopic world, focusing on a tiny snail nestled snugly within the iridescent interior of a large, empty shell. The snail's spiral shell is a polished gem of translucent amber and brown, delicately lined with growth lines that shimmer in soft, directed light. Its two slender eyestalks are fully extended, tentatively exploring its new surroundings as its pale, soft, muscular feet cling to the curved walls. With deliberate, mesmerizing slowness, the feet advance over the perfectly smooth surface, their movement creating a series of gentle, undulating ripples that propel them forward. The interior of the shell itself is a shimmering mother-of-pearl cavern, reflecting light into a captivating spectrum of pearlescent pinks, blues, and greens. The overall atmosphere is one of tranquil focus and quiet discovery, highlighting the beauty of natural, methodical exploration at the microscopic scale.", + "seed": 21313, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_14/index_7485_seed_21313_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1248988762497902, + "vlm_logprob": -0.1248988762497902, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.1248988762497902, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4238877594470978 + ], + "opposite_logprob_tokens": [ + -1.1738877296447754 + ], + "logprob_reward": -0.4238877594470978, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.4238877594470978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.283949077129364, + 0.11838696897029877, + 0.13395622372627258, + 0.04472988843917847, + 0.10319162905216217, + 0.131946861743927, + 0.021808471530675888, + 0.07569882273674011, + 0.051908060908317566, + 0.03442399576306343 + ], + "credit_map": [ + 0.009160533547401428, + 0.01220784243196249, + 0.010470576584339142, + 0.09226162731647491, + 0.01727394573390484, + 0.016132261604070663, + 0.11269275099039078, + 0.013749518431723118, + 0.013106529600918293, + 0.019679123535752296, + 0.014230935834348202, + 0.016160428524017334, + 0.020513013005256653, + 0.012835880741477013, + 0.013923952355980873, + 0.00793710257858038, + 0.005215054377913475, + 0.014553939923644066, + 0.014296882785856724, + 0.04067446291446686, + 0.01169517356902361, + 0.008545543067157269, + 0.015876613557338715, + 0.023098554462194443, + 0.0043679275549948215, + 0.005632862448692322, + 0.005482993088662624, + 0.00741330161690712, + 0.007206474430859089, + 0.0037143519148230553, + 0.007034088019281626, + 0.0038778900634497404, + 0.014721740037202835, + 0.0124490512534976, + 0.016616184264421463, + 0.01036764308810234, + 0.014989279210567474, + 0.011656160466372967, + 0.012259917333722115, + 0.010131649672985077, + 0.01010449044406414, + 0.015540695749223232, + 0.01247079111635685, + 0.01336574461311102, + 0.014365297742187977, + 0.02483612298965454, + 0.024856846779584885, + 0.016406865790486336, + 0.00308987102471292, + 0.0031087908428162336, + 0.002356678945943713, + 0.0028119718190282583, + 0.002028734190389514, + 0.0033942817244678736, + 0.002788814017549157, + 0.0022293294314295053, + 0.007437788415700197, + 0.008813578635454178, + 0.006430677138268948, + 0.01319561805576086, + 0.007744988426566124, + 0.007141925394535065, + 0.01789376698434353, + 0.007040486205369234, + 0.004042304120957851, + 0.004532393999397755, + 0.005434713792055845, + 0.013731050305068493, + 0.003979198168963194, + 0.004661670420318842, + 0.005621859338134527, + 0.009904869832098484, + 0.004405633080750704, + 0.004194028675556183, + 0.003751336131244898, + 0.005331816617399454, + 0.003847298910841346, + 0.004349092487245798, + 0.003475756384432316, + 0.005069030448794365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.4238877594470978 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020639877766370773 + ], + "opposite_logprob_tokens": [ + -3.8956398963928223 + ], + "logprob_reward": -0.020639877766370773, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020639877766370773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09717719256877899, + 0.07376892864704132, + 0.48324960470199585, + 0.03344029188156128, + 0.02117646299302578, + 0.034636471420526505, + 0.07496634125709534, + 0.1151934266090393, + 0.03460993617773056, + 0.03178133815526962 + ], + "credit_map": [ + 0.006291951052844524, + 0.008685714565217495, + 0.007954820990562439, + 0.029392434284090996, + 0.012308080680668354, + 0.008194985799491405, + 0.019641045480966568, + 0.004708159249275923, + 0.008194324560463428, + 0.010637541301548481, + 0.011071915738284588, + 0.008177576586604118, + 0.013115305453538895, + 0.010023149661719799, + 0.00653805723413825, + 0.00601105485111475, + 0.017328567802906036, + 0.05947098135948181, + 0.05666918307542801, + 0.17695988714694977, + 0.022007668390870094, + 0.024601077660918236, + 0.02598697692155838, + 0.10022526234388351, + 0.004406283609569073, + 0.003937583416700363, + 0.0043158880434930325, + 0.0035606056917458773, + 0.004061877727508545, + 0.003269168548285961, + 0.007234868593513966, + 0.0026540157850831747, + 0.0033847158774733543, + 0.0027232246939092875, + 0.003474810626357794, + 0.0017354940064251423, + 0.0026854651514440775, + 0.0027647139504551888, + 0.0021577489096671343, + 0.002250289311632514, + 0.003413580823689699, + 0.00441094720736146, + 0.004019528161734343, + 0.00316808489151299, + 0.004542381037026644, + 0.005978407338261604, + 0.004973103757947683, + 0.0041304356418550014, + 0.01728290505707264, + 0.010388705879449844, + 0.010496975854039192, + 0.0074403174221515656, + 0.007222929038107395, + 0.005965186282992363, + 0.010152538307011127, + 0.006016781087964773, + 0.012644024565815926, + 0.013802058063447475, + 0.011756018735468388, + 0.01911020837724209, + 0.011578037403523922, + 0.009962128475308418, + 0.02549581229686737, + 0.01084513496607542, + 0.002640858758240938, + 0.002611845964565873, + 0.003424059832468629, + 0.009115803055465221, + 0.0026147502940148115, + 0.0031342869624495506, + 0.0045645590871572495, + 0.006503770127892494, + 0.0027722741942852736, + 0.0025165353436022997, + 0.007857474498450756, + 0.004615365527570248, + 0.0018920995062217116, + 0.002815674990415573, + 0.0030825890135020018, + 0.006229322869330645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.020639877766370773 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009800177067518234 + ], + "opposite_logprob_tokens": [ + -4.634799957275391 + ], + "logprob_reward": -0.009800177067518234, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009800177067518234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3474601209163666, + 0.11645318567752838, + 0.06223462149500847, + 0.15502330660820007, + 0.037705738097429276, + 0.027288047596812248, + 0.07684523612260818, + 0.08400677889585495, + 0.05664721503853798, + 0.03633580729365349 + ], + "credit_map": [ + 0.01115432195365429, + 0.017309168353676796, + 0.014536257833242416, + 0.11107885837554932, + 0.019115569069981575, + 0.020714551210403442, + 0.13714873790740967, + 0.01640266738831997, + 0.011329755187034607, + 0.0184920821338892, + 0.015010880306363106, + 0.016760051250457764, + 0.021046366542577744, + 0.009826548397541046, + 0.015574416145682335, + 0.008413085713982582, + 0.004875523503869772, + 0.008577695116400719, + 0.007836601696908474, + 0.011965973302721977, + 0.006316662300378084, + 0.0060045006684958935, + 0.010009287856519222, + 0.006648378912359476, + 0.023466475307941437, + 0.02044794335961342, + 0.025157716125249863, + 0.027220508083701134, + 0.01649434305727482, + 0.009630805812776089, + 0.015780264511704445, + 0.016825241968035698, + 0.0040453034453094006, + 0.006084187421947718, + 0.0056190500035882, + 0.003411515848711133, + 0.005984033457934856, + 0.00394586892798543, + 0.004907125607132912, + 0.003708651289343834, + 0.002283358946442604, + 0.0023002151865512133, + 0.0030124997720122337, + 0.0030493703670799732, + 0.003754798322916031, + 0.0055837989784777164, + 0.0036676726303994656, + 0.0036363359540700912, + 0.015099522657692432, + 0.011478005908429623, + 0.009416390210390091, + 0.009171321988105774, + 0.007129852194339037, + 0.006921745371073484, + 0.010804316028952599, + 0.006824082229286432, + 0.008836735039949417, + 0.010517706163227558, + 0.008855769410729408, + 0.014275853522121906, + 0.007763801608234644, + 0.007755537051707506, + 0.0181166622787714, + 0.007884716615080833, + 0.004799290560185909, + 0.004014195874333382, + 0.005028368905186653, + 0.01536809653043747, + 0.004539679270237684, + 0.005101091228425503, + 0.00765365082770586, + 0.010142844170331955, + 0.003280027536675334, + 0.0036854424979537725, + 0.006758102215826511, + 0.0048555852845311165, + 0.003808211302384734, + 0.004307566210627556, + 0.0037732331547886133, + 0.005867642350494862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.009800177067518234 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097922682762146 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.0097922682762146, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097922682762146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30438631772994995, + 0.03951408341526985, + 0.26954615116119385, + 0.10982422530651093, + 0.018649304285645485, + 0.024598797783255577, + 0.06074915826320648, + 0.02125254087150097, + 0.12703338265419006, + 0.02444607950747013 + ], + "credit_map": [ + 0.007711836602538824, + 0.015324358828365803, + 0.01175231859087944, + 0.10695973038673401, + 0.020640388131141663, + 0.013760694302618504, + 0.11518459767103195, + 0.013052377849817276, + 0.0036735248286277056, + 0.00645549688488245, + 0.0046461839228868484, + 0.004472323227673769, + 0.006530031096190214, + 0.004945220425724983, + 0.0060342103242874146, + 0.0027570936363190413, + 0.008616442792117596, + 0.03360888361930847, + 0.03211349621415138, + 0.10182827711105347, + 0.010956761427223682, + 0.01362975500524044, + 0.012978502549231052, + 0.055814050137996674, + 0.00873112678527832, + 0.010791977867484093, + 0.008443983271718025, + 0.02865012362599373, + 0.00623574061319232, + 0.012927098199725151, + 0.009931887499988079, + 0.024112289771437645, + 0.0028253644704818726, + 0.00246679550036788, + 0.0028469518292695284, + 0.0019985635299235582, + 0.002578385639935732, + 0.0014200741425156593, + 0.002796136774122715, + 0.0017170325154438615, + 0.0022262129932641983, + 0.0022435258142650127, + 0.00313455518335104, + 0.0020463878754526377, + 0.003326677717268467, + 0.004590599797666073, + 0.0036307626869529486, + 0.003400074318051338, + 0.005827159155160189, + 0.008198127150535583, + 0.00627276161685586, + 0.008077663369476795, + 0.008268251083791256, + 0.007291203830391169, + 0.009873980656266212, + 0.0069400109350681305, + 0.002626302419230342, + 0.003158984240144491, + 0.0027719794306904078, + 0.0028720402624458075, + 0.002156771719455719, + 0.0018667211988940835, + 0.0026304498314857483, + 0.003169290954247117, + 0.009522083215415478, + 0.007435387931764126, + 0.011470177210867405, + 0.03926405683159828, + 0.007430471014231443, + 0.010793915018439293, + 0.018258051946759224, + 0.022859234362840652, + 0.0027939339634031057, + 0.0025721529964357615, + 0.005181311164051294, + 0.0033325855620205402, + 0.0022147030103951693, + 0.0027208535466343164, + 0.002613876247778535, + 0.0030166625510901213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0097922682762146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037429869174957 + ], + "opposite_logprob_tokens": [ + -1.9103742837905884 + ], + "logprob_reward": -0.16037429869174957, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037429869174957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23688331246376038, + 0.07724706083536148, + 0.31690600514411926, + 0.1391409933567047, + 0.02386903017759323, + 0.018119441345334053, + 0.03450636938214302, + 0.05366209149360657, + 0.07563287019729614, + 0.024032864719629288 + ], + "credit_map": [ + 0.008179725147783756, + 0.01068638265132904, + 0.009596780873835087, + 0.07094515860080719, + 0.012656843289732933, + 0.015150547958910465, + 0.0976172462105751, + 0.012050642631947994, + 0.007490613963454962, + 0.012360329739749432, + 0.00831575132906437, + 0.009344039484858513, + 0.017662135884165764, + 0.007571022491902113, + 0.008057835511863232, + 0.006445332430303097, + 0.012232519686222076, + 0.03908974304795265, + 0.03633987158536911, + 0.11188077926635742, + 0.014737092889845371, + 0.016329754143953323, + 0.025405455380678177, + 0.06089078262448311, + 0.009466013871133327, + 0.015649601817131042, + 0.012631259858608246, + 0.039399098604917526, + 0.007280789781361818, + 0.014947443269193172, + 0.00921894796192646, + 0.030547836795449257, + 0.002577590523287654, + 0.00327345822006464, + 0.0037727649323642254, + 0.0025728666223585606, + 0.0037272863555699587, + 0.0024426444433629513, + 0.0033886036835610867, + 0.0021138149313628674, + 0.001647468889132142, + 0.002427032683044672, + 0.0019509715493768454, + 0.0019822658505290747, + 0.002135102404281497, + 0.0031702853739261627, + 0.0025934225413948298, + 0.002212890889495611, + 0.0037593927700072527, + 0.002992140594869852, + 0.004289190284907818, + 0.004663551691919565, + 0.0046669538132846355, + 0.003433746052905917, + 0.007328046020120382, + 0.003373345360159874, + 0.005711990874260664, + 0.005949313752353191, + 0.0044868686236441135, + 0.009873917326331139, + 0.005314396228641272, + 0.004644578788429499, + 0.012411863543093204, + 0.005269158165901899, + 0.006116771139204502, + 0.0054696849547326565, + 0.007304495666176081, + 0.021178534254431725, + 0.005615552421659231, + 0.0066877263598144054, + 0.009395384229719639, + 0.013864722102880478, + 0.002458913717418909, + 0.0024849690962582827, + 0.004670730326324701, + 0.0031016641296446323, + 0.002084936946630478, + 0.002373953117057681, + 0.002445001620799303, + 0.004412694368511438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.16037429869174957 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro-static shot delves into a microscopic world, focusing on a tiny snail nestled snugly within the iridescent interior of a large, empty shell. The snail's spiral shell is a polished gem of translucent amber and brown, delicately lined with growth lines that shimmer in soft, directed light. Its two slender eyestalks are fully extended, tentatively exploring its new surroundings as its pale, soft, muscular feet cling to the curved walls. With deliberate, mesmerizing slowness, the feet advance over the perfectly smooth surface, their movement creating a series of gentle, undulating ripples that propel them forward. The interior of the shell itself is a shimmering mother-of-pearl cavern, reflecting light into a captivating spectrum of pearlescent pinks, blues, and greens. The overall atmosphere is one of tranquil focus and quiet discovery, highlighting the beauty of natural, methodical exploration at the microscopic scale.", + "seed": 21314, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_14/index_7485_seed_21314_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1914373952895403, + "vlm_logprob": -0.1914373952895403, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.1914373952895403, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8723835945129395 + ], + "opposite_logprob_tokens": [ + -0.6223835945129395 + ], + "logprob_reward": -0.8723835945129395, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8723835945129395, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.28221195936203003, + 0.09457764029502869, + 0.0252876877784729, + 0.03232653811573982, + 0.02742691896855831, + 0.05644826218485832, + 0.08102886378765106, + 0.19273871183395386, + 0.13233838975429535, + 0.07561498135328293 + ], + "credit_map": [ + 0.030251534655690193, + 0.023661937564611435, + 0.01905788481235504, + 0.023171938955783844, + 0.03151239454746246, + 0.03967258706688881, + 0.047754138708114624, + 0.06712955981492996, + 0.012889989651739597, + 0.011921684257686138, + 0.008007928729057312, + 0.008657734841108322, + 0.012395558878779411, + 0.010866598226130009, + 0.009623603895306587, + 0.020214540883898735, + 0.0031239758245646954, + 0.003635822096839547, + 0.0034013139083981514, + 0.0024919123388826847, + 0.0027212188579142094, + 0.0025299768894910812, + 0.0032926728017628193, + 0.004090794827789068, + 0.003598244395107031, + 0.0035383859649300575, + 0.004198145121335983, + 0.0030661013443022966, + 0.0035054010804742575, + 0.0038051356095820665, + 0.004269649274647236, + 0.006345476023852825, + 0.004079590085893869, + 0.005098311230540276, + 0.0038668697234243155, + 0.002776268171146512, + 0.003055101726204157, + 0.0025027599185705185, + 0.002432576147839427, + 0.003615441732108593, + 0.007076011039316654, + 0.008193961344659328, + 0.008585745468735695, + 0.0067378138191998005, + 0.0072036427445709705, + 0.004630488809198141, + 0.007318967487663031, + 0.006701629143208265, + 0.011844806373119354, + 0.011455077677965164, + 0.01149974949657917, + 0.01053500548005104, + 0.008560811169445515, + 0.006628134287893772, + 0.009679244831204414, + 0.010826033540070057, + 0.020645171403884888, + 0.023303618654608727, + 0.020610449835658073, + 0.016537461429834366, + 0.028370175510644913, + 0.022796927019953728, + 0.030151497572660446, + 0.030323414131999016, + 0.017383016645908356, + 0.019629182294011116, + 0.0211007222533226, + 0.014740814454853535, + 0.020206205546855927, + 0.010548635385930538, + 0.016411278396844864, + 0.012318523600697517, + 0.010830576531589031, + 0.008221722207963467, + 0.010863032191991806, + 0.012704581022262573, + 0.0052713993936777115, + 0.005652946885675192, + 0.016730286180973053, + 0.005340433679521084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.8723835945129395 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055004507303237915 + ], + "opposite_logprob_tokens": [ + -2.930004596710205 + ], + "logprob_reward": -0.055004507303237915, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055004507303237915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10253100842237473, + 0.11409348249435425, + 0.04056859016418457, + 0.06412098556756973, + 0.04958237707614899, + 0.04236724227666855, + 0.051034316420555115, + 0.20666980743408203, + 0.26381054520606995, + 0.06522167474031448 + ], + "credit_map": [ + 0.012355324812233448, + 0.011303648352622986, + 0.010118946433067322, + 0.01030676905065775, + 0.013840969651937485, + 0.012496519833803177, + 0.014345264062285423, + 0.017763560637831688, + 0.015425344929099083, + 0.0153803126886487, + 0.01367836631834507, + 0.01193889882415533, + 0.014619341120123863, + 0.011306527070701122, + 0.012244664132595062, + 0.019500022754073143, + 0.004192477557808161, + 0.0062431758269667625, + 0.0061208936385810375, + 0.004472081083804369, + 0.003977057058364153, + 0.004256632179021835, + 0.004725104663521051, + 0.006581168621778488, + 0.0069820997305214405, + 0.009691299870610237, + 0.008463137783110142, + 0.007115737069398165, + 0.005185465328395367, + 0.0075648450292646885, + 0.007147859316319227, + 0.011970548890531063, + 0.00530342198908329, + 0.006947265937924385, + 0.007097635883837938, + 0.006833081599324942, + 0.004854041617363691, + 0.004730905406177044, + 0.005877278279513121, + 0.007938748225569725, + 0.0051848022267222404, + 0.00593498395755887, + 0.0068046944215893745, + 0.005429669748991728, + 0.004720000084489584, + 0.003299063304439187, + 0.005524845328181982, + 0.005469182506203651, + 0.008472219109535217, + 0.006359429564327002, + 0.006750026717782021, + 0.0068833548575639725, + 0.004944156389683485, + 0.004137019161134958, + 0.006243078503757715, + 0.007245032116770744, + 0.023145947605371475, + 0.02936420775949955, + 0.03590220585465431, + 0.026079658418893814, + 0.03204188495874405, + 0.015245668590068817, + 0.02640414610505104, + 0.01848609559237957, + 0.050128817558288574, + 0.03505953401327133, + 0.0345149040222168, + 0.0328180156648159, + 0.030444370582699776, + 0.021817393600940704, + 0.02982630394399166, + 0.029201211407780647, + 0.009682726114988327, + 0.011037702672183514, + 0.009760864078998566, + 0.01037346851080656, + 0.005437776912003756, + 0.007739246357232332, + 0.005941188894212246, + 0.005248702596873045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.055004507303237915 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007642199285328388 + ], + "opposite_logprob_tokens": [ + -4.8826422691345215 + ], + "logprob_reward": -0.007642199285328388, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007642199285328388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25431060791015625, + 0.1928635835647583, + 0.04851935803890228, + 0.04620179533958435, + 0.07583463191986084, + 0.04686718434095383, + 0.07396762818098068, + 0.10467777401208878, + 0.06356901675462723, + 0.09318842738866806 + ], + "credit_map": [ + 0.030932361260056496, + 0.018938904628157616, + 0.018482325598597527, + 0.022305017337203026, + 0.024425415322184563, + 0.03549450635910034, + 0.04559587687253952, + 0.058136194944381714, + 0.022913401946425438, + 0.01810424216091633, + 0.020873211324214935, + 0.01891499198973179, + 0.028598085045814514, + 0.022323396056890488, + 0.024609319865703583, + 0.036526940762996674, + 0.005496834870427847, + 0.007700086571276188, + 0.007751299999654293, + 0.004379972815513611, + 0.004095308016985655, + 0.0050088344141840935, + 0.004734668880701065, + 0.009352351538836956, + 0.006446321029216051, + 0.005347706377506256, + 0.00940091721713543, + 0.00439350213855505, + 0.005219625774770975, + 0.004650338087230921, + 0.004759430885314941, + 0.005983950570225716, + 0.008703397586941719, + 0.008300328627228737, + 0.011218659579753876, + 0.010747051797807217, + 0.009485953487455845, + 0.008260378614068031, + 0.00906489510089159, + 0.010053969919681549, + 0.005885349586606026, + 0.007802226580679417, + 0.007265474647283554, + 0.005071784369647503, + 0.005490075331181288, + 0.0037445789203047752, + 0.006218796595931053, + 0.005388898309320211, + 0.008392044343054295, + 0.010505873709917068, + 0.012377725914120674, + 0.009365200996398926, + 0.01108369417488575, + 0.005423310212790966, + 0.008746554143726826, + 0.008073217235505581, + 0.010581126436591148, + 0.0135769909247756, + 0.014372910372912884, + 0.015075440518558025, + 0.016947610303759575, + 0.009011930786073208, + 0.011768932454288006, + 0.013342836871743202, + 0.008626243099570274, + 0.012745615094900131, + 0.008236026391386986, + 0.00959102064371109, + 0.008183344267308712, + 0.004929813090711832, + 0.00646616704761982, + 0.004790782928466797, + 0.015141641721129417, + 0.015947235748171806, + 0.01681448332965374, + 0.014423471875488758, + 0.006129116285592318, + 0.006083731539547443, + 0.011655899696052074, + 0.006992846727371216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.007642199285328388 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108264084905386 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.01108264084905386, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108264084905386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2342594861984253, + 0.1562846601009369, + 0.07872048020362854, + 0.09531515091657639, + 0.04871107637882233, + 0.0189262256026268, + 0.03984782099723816, + 0.14613015949726105, + 0.10972525179386139, + 0.07207963615655899 + ], + "credit_map": [ + 0.024463269859552383, + 0.021875161677598953, + 0.017060795798897743, + 0.016347529366612434, + 0.03756050020456314, + 0.034270431846380234, + 0.03211193159222603, + 0.05056985467672348, + 0.019126366823911667, + 0.014214891009032726, + 0.01398658100515604, + 0.013898834586143494, + 0.01830529235303402, + 0.02086634561419487, + 0.018208064138889313, + 0.03767829015851021, + 0.007369862869381905, + 0.007961968891322613, + 0.0072786868549883366, + 0.008196715265512466, + 0.005740811116993427, + 0.010973243974149227, + 0.009711923077702522, + 0.021487271413207054, + 0.01101509016007185, + 0.011949706822633743, + 0.015192856080830097, + 0.010497594252228737, + 0.00903788860887289, + 0.01093137264251709, + 0.010515152476727962, + 0.016175489872694016, + 0.00650865538045764, + 0.006168672349303961, + 0.007281655911356211, + 0.0065772137604653835, + 0.005530080292373896, + 0.004769227001816034, + 0.00544318649917841, + 0.006432380527257919, + 0.001980409724637866, + 0.002531129866838455, + 0.002289424417540431, + 0.0023180919233709574, + 0.0023780963383615017, + 0.0018356147920712829, + 0.0027195552829653025, + 0.0028739043045789003, + 0.0045967428013682365, + 0.007893204689025879, + 0.004833810962736607, + 0.00714245717972517, + 0.0036170892417430878, + 0.0032412807922810316, + 0.004546827170997858, + 0.0039764088578522205, + 0.019695386290550232, + 0.019929716363549232, + 0.025883402675390244, + 0.023753853514790535, + 0.015602550469338894, + 0.011204061098396778, + 0.013005678541958332, + 0.017055509611964226, + 0.021660462021827698, + 0.012019142508506775, + 0.019699525088071823, + 0.010225670412182808, + 0.014008205384016037, + 0.00979054719209671, + 0.012712485156953335, + 0.009609212167561054, + 0.012624545022845268, + 0.009101220406591892, + 0.010116389952600002, + 0.006161276716738939, + 0.0040587387047708035, + 0.009111146442592144, + 0.009879419580101967, + 0.011026897467672825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.01108264084905386 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074034497141838 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.011074034497141838, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074034497141838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3233640193939209, + 0.32995477318763733, + 0.03260648250579834, + 0.027555929496884346, + 0.0341520756483078, + 0.027537906542420387, + 0.014579539187252522, + 0.06225147098302841, + 0.08280631899833679, + 0.06519141048192978 + ], + "credit_map": [ + 0.028734082356095314, + 0.03191789984703064, + 0.02495487406849861, + 0.035357702523469925, + 0.04803205654025078, + 0.040490999817848206, + 0.04611847177147865, + 0.0677579715847969, + 0.03597240149974823, + 0.028790056705474854, + 0.03257754072546959, + 0.028117135167121887, + 0.05473949387669563, + 0.039308398962020874, + 0.03687415271997452, + 0.07357562333345413, + 0.00321041326969862, + 0.004577660467475653, + 0.003928437363356352, + 0.003113938495516777, + 0.0034145384561270475, + 0.0032017577905207872, + 0.003785806242376566, + 0.0073739318177104, + 0.003577845636755228, + 0.0032880904618650675, + 0.003950198646634817, + 0.002781165763735771, + 0.0027691496070474386, + 0.002812698483467102, + 0.0035381591878831387, + 0.004838623572140932, + 0.003494849195703864, + 0.0038058608770370483, + 0.004559765569865704, + 0.003520938567817211, + 0.003585649188607931, + 0.003886266378685832, + 0.005188058130443096, + 0.006110690534114838, + 0.003377236658707261, + 0.0034241564571857452, + 0.003753503318876028, + 0.002678541000932455, + 0.0030360177624970675, + 0.002436417853459716, + 0.0037071718834340572, + 0.005124864634126425, + 0.0017852092860266566, + 0.002027094829827547, + 0.0020746844820678234, + 0.0022852998226881027, + 0.0017364308005198836, + 0.0012856199173256755, + 0.0017677992582321167, + 0.0016174007905647159, + 0.006874398328363895, + 0.007829045876860619, + 0.011305514723062515, + 0.008242790587246418, + 0.009855289943516254, + 0.004732969682663679, + 0.00789451040327549, + 0.0055169593542814255, + 0.014316058717668056, + 0.013274638913571835, + 0.008801058866083622, + 0.009608108550310135, + 0.012327421456575394, + 0.009070741944015026, + 0.007607316132634878, + 0.0078009809367358685, + 0.009827882051467896, + 0.0062631769105792046, + 0.0050247409380972385, + 0.005590642336755991, + 0.004945381078869104, + 0.011791098862886429, + 0.013247709721326828, + 0.008500782772898674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.011074034497141838 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro-static shot delves into a microscopic world, focusing on a tiny snail nestled snugly within the iridescent interior of a large, empty shell. The snail's spiral shell is a polished gem of translucent amber and brown, delicately lined with growth lines that shimmer in soft, directed light. Its two slender eyestalks are fully extended, tentatively exploring its new surroundings as its pale, soft, muscular feet cling to the curved walls. With deliberate, mesmerizing slowness, the feet advance over the perfectly smooth surface, their movement creating a series of gentle, undulating ripples that propel them forward. The interior of the shell itself is a shimmering mother-of-pearl cavern, reflecting light into a captivating spectrum of pearlescent pinks, blues, and greens. The overall atmosphere is one of tranquil focus and quiet discovery, highlighting the beauty of natural, methodical exploration at the microscopic scale.", + "seed": 21315, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_14/index_7485_seed_21315_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.25045759696513414, + "vlm_logprob": -0.25045759696513414, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.25045759696513414, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1706691980361938 + ], + "opposite_logprob_tokens": [ + -0.42066922783851624 + ], + "logprob_reward": -1.1706691980361938, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1706691980361938, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10883796215057373, + 0.14070411026477814, + 0.03198058530688286, + 0.04576118290424347, + 0.04777488857507706, + 0.48567327857017517, + 0.028019003570079803, + 0.06746737658977509, + 0.019684210419654846, + 0.024097364395856857 + ], + "credit_map": [ + 0.010765757411718369, + 0.01566172018647194, + 0.011503147892653942, + 0.013438066467642784, + 0.013262415304780006, + 0.01311304047703743, + 0.01955857127904892, + 0.01153524685651064, + 0.012479541823267937, + 0.022130539640784264, + 0.010428055189549923, + 0.03323798254132271, + 0.013873451389372349, + 0.010254318825900555, + 0.019133349880576134, + 0.01916687749326229, + 0.003074487205594778, + 0.005593014415353537, + 0.0030760332010686398, + 0.006258721463382244, + 0.0032866939436644316, + 0.002632418880239129, + 0.005011730827391148, + 0.0030474874656647444, + 0.004884061403572559, + 0.0099342605099082, + 0.004087263718247414, + 0.007137201726436615, + 0.006027837749570608, + 0.0034071700647473335, + 0.0061568873934447765, + 0.0041264984756708145, + 0.004370731767266989, + 0.008589113131165504, + 0.003633680287748575, + 0.005981001071631908, + 0.006950316485017538, + 0.006562260910868645, + 0.007035890594124794, + 0.004651891533285379, + 0.0367649607360363, + 0.018451638519763947, + 0.010803199373185635, + 0.18511080741882324, + 0.016440950334072113, + 0.008862733840942383, + 0.18473957479000092, + 0.024499407038092613, + 0.0029864448588341475, + 0.0036180911120027304, + 0.0036065438762307167, + 0.004127909895032644, + 0.0025397189892828465, + 0.004313945304602385, + 0.004279473330825567, + 0.0025468734093010426, + 0.010913004167377949, + 0.005655115935951471, + 0.011481315828859806, + 0.010542033240199089, + 0.005828477442264557, + 0.0057699126191437244, + 0.009595409035682678, + 0.007682102266699076, + 0.0027166728395968676, + 0.002064520725980401, + 0.0033474969677627087, + 0.0035855197347700596, + 0.0018241259967908263, + 0.0018270727014169097, + 0.0022709507029503584, + 0.002047851914539933, + 0.003483504056930542, + 0.0029644013848155737, + 0.002933389274403453, + 0.003318868111819029, + 0.002952363807708025, + 0.0019459340255707502, + 0.003905645804479718, + 0.00259325816296041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -1.1706691980361938 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020646067336201668 + ], + "opposite_logprob_tokens": [ + -3.895646095275879 + ], + "logprob_reward": -0.020646067336201668, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020646067336201668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14126645028591156, + 0.28145018219947815, + 0.05186263844370842, + 0.08122710138559341, + 0.0915801003575325, + 0.13843241333961487, + 0.05872300639748573, + 0.08671675622463226, + 0.0477210097014904, + 0.0210203118622303 + ], + "credit_map": [ + 0.012658395804464817, + 0.021479977294802666, + 0.019033750519156456, + 0.017733046784996986, + 0.017930958420038223, + 0.01607438176870346, + 0.022610587999224663, + 0.013745357282459736, + 0.015730099752545357, + 0.041719794273376465, + 0.012024234980344772, + 0.07816645503044128, + 0.02464912459254265, + 0.029517367482185364, + 0.02205265499651432, + 0.05759045481681824, + 0.004629264585673809, + 0.011355281807482243, + 0.0036246133968234062, + 0.011469525285065174, + 0.0054981340654194355, + 0.005547472275793552, + 0.0061725955456495285, + 0.0035657479893416166, + 0.007508754730224609, + 0.012675049714744091, + 0.005769890267401934, + 0.011317728087306023, + 0.009092883206903934, + 0.007566788233816624, + 0.01566649228334427, + 0.011629519984126091, + 0.010125447995960712, + 0.014582541771233082, + 0.008263642899692059, + 0.016003679484128952, + 0.009411524049937725, + 0.0077522750943899155, + 0.01321549341082573, + 0.0122255003079772, + 0.010659155435860157, + 0.005820853635668755, + 0.003967590630054474, + 0.049977995455265045, + 0.0056494371965527534, + 0.003394006984308362, + 0.0505635030567646, + 0.008399873971939087, + 0.005599654279649258, + 0.005965420510619879, + 0.0037482408806681633, + 0.01291833259165287, + 0.005148538388311863, + 0.007599422708153725, + 0.008693583309650421, + 0.009049811400473118, + 0.011018162593245506, + 0.006894631311297417, + 0.009948654100298882, + 0.019430844113230705, + 0.007947401143610477, + 0.008998313918709755, + 0.013165315613150597, + 0.009313435293734074, + 0.0046753003261983395, + 0.00548196118324995, + 0.0047586774453520775, + 0.010824843309819698, + 0.0045334165915846825, + 0.0029189959168434143, + 0.011390218511223793, + 0.0031375987455248833, + 0.0028261973056942225, + 0.002701863879337907, + 0.0035492279566824436, + 0.0031771764624863863, + 0.0021605389192700386, + 0.002104853978380561, + 0.0023493622429668903, + 0.0021510920487344265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.020646067336201668 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008662267588078976 + ], + "opposite_logprob_tokens": [ + -4.758662223815918 + ], + "logprob_reward": -0.008662267588078976, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008662267588078976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06960143893957138, + 0.1943158507347107, + 0.05767717957496643, + 0.022742770612239838, + 0.02232864871621132, + 0.542731761932373, + 0.028990332037210464, + 0.03162478283047676, + 0.02103901281952858, + 0.008948155678808689 + ], + "credit_map": [ + 0.007370675448328257, + 0.008625487796962261, + 0.006075399927794933, + 0.008335397578775883, + 0.008170290850102901, + 0.009127606637775898, + 0.0137195885181427, + 0.008177001029253006, + 0.012176156044006348, + 0.028759362176060677, + 0.009569846093654633, + 0.04811514541506767, + 0.02124996855854988, + 0.020426275208592415, + 0.016571994870901108, + 0.037447117269039154, + 0.0038451505824923515, + 0.008091459050774574, + 0.0026996321976184845, + 0.020670156925916672, + 0.002132871188223362, + 0.0038945558480918407, + 0.007288717664778233, + 0.009054640308022499, + 0.002329080132767558, + 0.004413798917084932, + 0.002209755592048168, + 0.0035317877773195505, + 0.0033055120147764683, + 0.0020986751187592745, + 0.002907136222347617, + 0.0019470266997814178, + 0.0023943111300468445, + 0.003263666993007064, + 0.0018379311077296734, + 0.002882660599425435, + 0.0031380613800138235, + 0.002667295979335904, + 0.0026237131096422672, + 0.003521011210978031, + 0.04119522124528885, + 0.02071026712656021, + 0.010794048197567463, + 0.20673328638076782, + 0.018593918532133102, + 0.008998647332191467, + 0.2079540193080902, + 0.02775241807103157, + 0.003023377154022455, + 0.00455904845148325, + 0.002424323232844472, + 0.0037095763254910707, + 0.003864987986162305, + 0.004496697802096605, + 0.003338669892400503, + 0.0035736565478146076, + 0.003964007832109928, + 0.005296077113598585, + 0.003034143475815654, + 0.0053323740139603615, + 0.0034512921702116728, + 0.003056061454117298, + 0.004064030945301056, + 0.0034267976880073547, + 0.002752317814156413, + 0.003467516740784049, + 0.002457908121868968, + 0.0025778263807296753, + 0.0022207307629287243, + 0.002321472857147455, + 0.003605823963880539, + 0.0016354171093553305, + 0.0011549614137038589, + 0.0017332137795165181, + 0.000770921993535012, + 0.0015127536607906222, + 0.0011463051196187735, + 0.0006946224602870643, + 0.0013403272023424506, + 0.0005950502818450332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.008662267588078976 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207758009433746 + ], + "opposite_logprob_tokens": [ + -4.26420783996582 + ], + "logprob_reward": -0.014207758009433746, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207758009433746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06224190071225166, + 0.41482922434806824, + 0.08005504310131073, + 0.016116442158818245, + 0.029364947229623795, + 0.29377949237823486, + 0.023617375642061234, + 0.042764753103256226, + 0.02653796225786209, + 0.010692883282899857 + ], + "credit_map": [ + 0.006479070521891117, + 0.008202346041798592, + 0.005156003870069981, + 0.007591999135911465, + 0.006738467141985893, + 0.008352413773536682, + 0.012362519279122353, + 0.007359076291322708, + 0.023660628125071526, + 0.05680382251739502, + 0.014682522043585777, + 0.12512995302677155, + 0.02565739117562771, + 0.04326428100466728, + 0.032217685133218765, + 0.09341295808553696, + 0.005218002945184708, + 0.010492192581295967, + 0.0027382203843444586, + 0.03126630187034607, + 0.002130131935700774, + 0.005204669665545225, + 0.011618428863584995, + 0.011387093923985958, + 0.0015633412403985858, + 0.003805384738370776, + 0.0012352042831480503, + 0.0024215211160480976, + 0.0021880636923015118, + 0.0015161944320425391, + 0.002141020493581891, + 0.0012457119300961494, + 0.0042703477665781975, + 0.0044457754120230675, + 0.0028289365582168102, + 0.004935135133564472, + 0.002891968470066786, + 0.002458331175148487, + 0.0033150934614241123, + 0.004219359718263149, + 0.022272465750575066, + 0.011258840560913086, + 0.0057569947093725204, + 0.11125154048204422, + 0.01033853366971016, + 0.00495205307379365, + 0.11267971992492676, + 0.015269328840076923, + 0.002298858715221286, + 0.0023343332577496767, + 0.001821218989789486, + 0.007522621192038059, + 0.0020364592783153057, + 0.0018482410814613104, + 0.0028735592495650053, + 0.002882082015275955, + 0.005436321720480919, + 0.006681373808532953, + 0.0037700720131397247, + 0.00825762189924717, + 0.004312473814934492, + 0.0038812009152024984, + 0.005260169506072998, + 0.005165519658476114, + 0.0026291150134056807, + 0.0034272773191332817, + 0.00234088022261858, + 0.005511284340173006, + 0.0020406791009008884, + 0.001310160499997437, + 0.00800138060003519, + 0.0012771828332915902, + 0.0015578027814626694, + 0.0012634404702112079, + 0.0017150860512629151, + 0.0016748287016525865, + 0.0009476293926127255, + 0.0008837223867885768, + 0.0016342930030077696, + 0.0010160809615626931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.014207758009433746 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810269385576248 + ], + "opposite_logprob_tokens": [ + -3.288102626800537 + ], + "logprob_reward": -0.03810269385576248, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810269385576248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1076139584183693, + 0.2302219569683075, + 0.17158320546150208, + 0.013349602930247784, + 0.013763036578893661, + 0.39954930543899536, + 0.022631824016571045, + 0.029855692759156227, + 0.005963009316474199, + 0.005468416027724743 + ], + "credit_map": [ + 0.010515048168599606, + 0.014424581080675125, + 0.00980966817587614, + 0.012135829776525497, + 0.010249244049191475, + 0.01536373607814312, + 0.02206229418516159, + 0.013053552247583866, + 0.014500490389764309, + 0.03252401202917099, + 0.00675557367503643, + 0.07321857661008835, + 0.010987921617925167, + 0.022258883342146873, + 0.017389709129929543, + 0.05258680880069733, + 0.009769774042069912, + 0.022478055208921432, + 0.005913658533245325, + 0.06742643564939499, + 0.004161418881267309, + 0.011682309210300446, + 0.025099284946918488, + 0.025052284821867943, + 0.001332713640294969, + 0.0026275620330125093, + 0.0012448017951101065, + 0.0019991411827504635, + 0.0023548135068267584, + 0.0010369948577135801, + 0.001851496985182166, + 0.0009020803263410926, + 0.0014463653787970543, + 0.0024125976487994194, + 0.0014120476553216577, + 0.0019324758322909474, + 0.0015858197584748268, + 0.0011917714728042483, + 0.0016195623902603984, + 0.00216239714063704, + 0.030391640961170197, + 0.015464234165847301, + 0.00791161973029375, + 0.15211889147758484, + 0.013954547233879566, + 0.006385813467204571, + 0.15293274819850922, + 0.020389806479215622, + 0.002358741592615843, + 0.0017286513466387987, + 0.0022795372642576694, + 0.002781269606202841, + 0.0020559763070195913, + 0.005235509946942329, + 0.0025264688301831484, + 0.0036656681913882494, + 0.003910658415406942, + 0.004484184086322784, + 0.003842594102025032, + 0.004912425763905048, + 0.002322215586900711, + 0.0027950929943472147, + 0.00466262549161911, + 0.0029258967842906713, + 0.0009451061487197876, + 0.0007129998994059861, + 0.0007327236817218363, + 0.0008289889665320516, + 0.000585768953897059, + 0.0005760080530308187, + 0.000984250451438129, + 0.000597163219936192, + 0.0008257193840108812, + 0.0008999170968309045, + 0.0005680228350684047, + 0.0008633545367047191, + 0.0005642895121127367, + 0.0003723560075741261, + 0.0008917360100895166, + 0.00048302041250281036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.03810269385576248 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro-static shot delves into a microscopic world, focusing on a tiny snail nestled snugly within the iridescent interior of a large, empty shell. The snail's spiral shell is a polished gem of translucent amber and brown, delicately lined with growth lines that shimmer in soft, directed light. Its two slender eyestalks are fully extended, tentatively exploring its new surroundings as its pale, soft, muscular feet cling to the curved walls. With deliberate, mesmerizing slowness, the feet advance over the perfectly smooth surface, their movement creating a series of gentle, undulating ripples that propel them forward. The interior of the shell itself is a shimmering mother-of-pearl cavern, reflecting light into a captivating spectrum of pearlescent pinks, blues, and greens. The overall atmosphere is one of tranquil focus and quiet discovery, highlighting the beauty of natural, methodical exploration at the microscopic scale.", + "seed": 21316, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_14/index_7485_seed_21316_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14044785220175982, + "vlm_logprob": -0.14044785220175982, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.14044785220175982, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6094415783882141 + ], + "opposite_logprob_tokens": [ + -0.8594415783882141 + ], + "logprob_reward": -0.6094415783882141, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.6094415783882141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08316077291965485, + 0.05831645056605339, + 0.012612417340278625, + 0.021208886057138443, + 0.18054181337356567, + 0.45219314098358154, + 0.0338415689766407, + 0.02921706810593605, + 0.11180155724287033, + 0.017106322571635246 + ], + "credit_map": [ + 0.008153052069246769, + 0.011677525006234646, + 0.009025342762470245, + 0.007672806736081839, + 0.0156190674751997, + 0.010639253072440624, + 0.008702188730239868, + 0.011671536602079868, + 0.004752092529088259, + 0.011084041558206081, + 0.00858079269528389, + 0.008889471180737019, + 0.006436977069824934, + 0.008389411494135857, + 0.0054296511225402355, + 0.0047540077939629555, + 0.0011425415286794305, + 0.001926642144098878, + 0.0014793721493333578, + 0.003058265196159482, + 0.0013606998836621642, + 0.0012491736561059952, + 0.0012891942169517279, + 0.0011065286817029119, + 0.0019003041088581085, + 0.004531010054051876, + 0.0018320302478969097, + 0.005306168459355831, + 0.0018737586215138435, + 0.002312662312760949, + 0.002022991655394435, + 0.0014299607137218118, + 0.01589074917137623, + 0.024688810110092163, + 0.020193422213196754, + 0.04556412622332573, + 0.028607746586203575, + 0.012832855805754662, + 0.02415953204035759, + 0.008604573085904121, + 0.07163558900356293, + 0.04766025021672249, + 0.06456288695335388, + 0.049329712986946106, + 0.038771845400333405, + 0.046997614204883575, + 0.055837128311395645, + 0.07739809155464172, + 0.003972241189330816, + 0.002834434388205409, + 0.004328668117523193, + 0.0038699107244610786, + 0.00644304184243083, + 0.0039039093535393476, + 0.004591255448758602, + 0.0038981083780527115, + 0.0031724125146865845, + 0.002873215125873685, + 0.00264726928435266, + 0.0035234796814620495, + 0.0041747489012777805, + 0.0036735106259584427, + 0.00352188921533525, + 0.005630542989820242, + 0.012832444161176682, + 0.020667536184191704, + 0.01246012281626463, + 0.018855944275856018, + 0.012727734632790089, + 0.012440565042197704, + 0.013917231932282448, + 0.007899980060756207, + 0.0017029591836035252, + 0.002659579273313284, + 0.0013753465609624982, + 0.002057326026260853, + 0.003499222919344902, + 0.002708297222852707, + 0.0018698859494179487, + 0.0012337056687101722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.6094415783882141 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381472185254097 + ], + "opposite_logprob_tokens": [ + -3.288147211074829 + ], + "logprob_reward": -0.0381472185254097, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381472185254097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05280965566635132, + 0.04489472880959511, + 0.018486671149730682, + 0.010232733562588692, + 0.22150802612304688, + 0.45935964584350586, + 0.07259662449359894, + 0.0398237518966198, + 0.06131402403116226, + 0.018974116072058678 + ], + "credit_map": [ + 0.005541772581636906, + 0.0066636488772928715, + 0.0048834835179150105, + 0.005112943705171347, + 0.008763703517615795, + 0.01144054438918829, + 0.005168718285858631, + 0.005234838929027319, + 0.003562296275049448, + 0.007295099552720785, + 0.005748639348894358, + 0.007697692606598139, + 0.005100115668028593, + 0.005955853499472141, + 0.00446788314729929, + 0.005067149642854929, + 0.00244817603379488, + 0.003461708314716816, + 0.0018676938489079475, + 0.0033761647064238787, + 0.002364745596423745, + 0.0019270546035841107, + 0.0018325257115066051, + 0.0012086016358807683, + 0.0013305502943694592, + 0.001185609376989305, + 0.0015284843975678086, + 0.0013283725129440427, + 0.0009571545524522662, + 0.0011398788774386048, + 0.0013152705505490303, + 0.001447412185370922, + 0.019864708185195923, + 0.032181791961193085, + 0.024084074422717094, + 0.056190840899944305, + 0.032415252178907394, + 0.014439357444643974, + 0.030812926590442657, + 0.011519094929099083, + 0.07533083856105804, + 0.04649580642580986, + 0.06865181773900986, + 0.0490739606320858, + 0.037561774253845215, + 0.04752594605088234, + 0.05401768907904625, + 0.0807018056511879, + 0.00842821504920721, + 0.007465667091310024, + 0.012071889825165272, + 0.012708473019301891, + 0.007399840746074915, + 0.0074525619857013226, + 0.008427009917795658, + 0.008642969653010368, + 0.003853755071759224, + 0.005495419260114431, + 0.0053832679986953735, + 0.005260275211185217, + 0.006952605210244656, + 0.0042061107233166695, + 0.004391586408019066, + 0.004280730150640011, + 0.006647723726928234, + 0.009260204620659351, + 0.006029502488672733, + 0.00833950936794281, + 0.010147024877369404, + 0.008795656263828278, + 0.00559984939172864, + 0.006494551431387663, + 0.0020456404890865088, + 0.0029624069575220346, + 0.0014490209287032485, + 0.002076154574751854, + 0.0032765765208750963, + 0.002877306193113327, + 0.002531734062358737, + 0.0017552739009261131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.0381472185254097 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006750753149390221 + ], + "opposite_logprob_tokens": [ + -5.006750583648682 + ], + "logprob_reward": -0.006750753149390221, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006750753149390221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20638929307460785, + 0.05150536447763443, + 0.012426995672285557, + 0.03390664607286453, + 0.06770364195108414, + 0.4134727418422699, + 0.07688058912754059, + 0.04689081013202667, + 0.0347871296107769, + 0.05603684112429619 + ], + "credit_map": [ + 0.017321957275271416, + 0.02108900249004364, + 0.01832549460232258, + 0.01363853644579649, + 0.05046684294939041, + 0.058317579329013824, + 0.01556252408772707, + 0.011667345650494099, + 0.0040322584100067616, + 0.010329748503863811, + 0.006610159296542406, + 0.008489541709423065, + 0.006391589529812336, + 0.005918147508054972, + 0.0055757020600140095, + 0.004158212803304195, + 0.0019219524692744017, + 0.0019766082987189293, + 0.0012610004050657153, + 0.001670846831984818, + 0.001697902800515294, + 0.0013156767236068845, + 0.0012633978622034192, + 0.0013196090003475547, + 0.003589101368561387, + 0.005140353459864855, + 0.005444840993732214, + 0.006712150294333696, + 0.003375036409124732, + 0.003070281120017171, + 0.002829431090503931, + 0.0037454485427588224, + 0.005651276558637619, + 0.010378352366387844, + 0.006481793709099293, + 0.01551878359168768, + 0.010099835693836212, + 0.00516848498955369, + 0.009337364695966244, + 0.005067747551947832, + 0.06537849456071854, + 0.041897568851709366, + 0.06039731949567795, + 0.04685021936893463, + 0.0345970056951046, + 0.04242744296789169, + 0.04896164312958717, + 0.07296303659677505, + 0.008444675244390965, + 0.009333744645118713, + 0.013275817036628723, + 0.012284341268241405, + 0.009187710471451283, + 0.00716833770275116, + 0.010126115754246712, + 0.007059835363179445, + 0.005966885946691036, + 0.005257775541394949, + 0.0047517153434455395, + 0.0057815262116491795, + 0.004498948808759451, + 0.005162880755960941, + 0.00557112880051136, + 0.009899943135678768, + 0.003839008044451475, + 0.005126524716615677, + 0.0040398770943284035, + 0.0052071246318519115, + 0.0033389399759471416, + 0.004060182720422745, + 0.004707204177975655, + 0.0044682640582323074, + 0.005716316867619753, + 0.005491911433637142, + 0.005820746999233961, + 0.007726202253252268, + 0.006182210054248571, + 0.007905242033302784, + 0.009980684146285057, + 0.007213526405394077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.006750753149390221 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205760322511196 + ], + "opposite_logprob_tokens": [ + -4.2642059326171875 + ], + "logprob_reward": -0.014205760322511196, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205760322511196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29763633012771606, + 0.031261149793863297, + 0.019430462270975113, + 0.025851545855402946, + 0.05339735373854637, + 0.3588598370552063, + 0.06877642869949341, + 0.027063915506005287, + 0.0495687760412693, + 0.06815413385629654 + ], + "credit_map": [ + 0.0247476939111948, + 0.03160533681511879, + 0.029456745833158493, + 0.020109334960579872, + 0.07101244479417801, + 0.07985669374465942, + 0.0224860031157732, + 0.018362097442150116, + 0.0023779929615557194, + 0.004781766794621944, + 0.003702963935211301, + 0.00484114047139883, + 0.004385459702461958, + 0.004771574400365353, + 0.0036699434276670218, + 0.002730312291532755, + 0.0020373165607452393, + 0.003165513277053833, + 0.0020305055659264326, + 0.003667184617370367, + 0.0022888104431331158, + 0.002128885593265295, + 0.0022893669083714485, + 0.0018228826811537147, + 0.002793174935504794, + 0.0033530564978718758, + 0.0038633260410279036, + 0.004476095084100962, + 0.0028202426619827747, + 0.0031393584795296192, + 0.0022040382027626038, + 0.003202256979420781, + 0.004814166110008955, + 0.007570527959614992, + 0.004866947419941425, + 0.013168063946068287, + 0.007613871246576309, + 0.0036344544496387243, + 0.007862423546612263, + 0.0038669079076498747, + 0.05608987435698509, + 0.035611342638731, + 0.05346078798174858, + 0.037970591336488724, + 0.028816072270274162, + 0.035404011607170105, + 0.04649415239691734, + 0.06501302868127823, + 0.009011520072817802, + 0.008236165158450603, + 0.010424546897411346, + 0.009786291047930717, + 0.00838683731853962, + 0.006669893860816956, + 0.008504791185259819, + 0.00775638734921813, + 0.002875908510759473, + 0.0025990919675678015, + 0.0023711344692856073, + 0.002989639062434435, + 0.0038800060283392668, + 0.0032336448784917593, + 0.0041135395877063274, + 0.00500095309689641, + 0.0049742283299565315, + 0.008720651268959045, + 0.005278115160763264, + 0.007581169251352549, + 0.005472721066325903, + 0.006612007971853018, + 0.005514872260391712, + 0.005415013991296291, + 0.00699638994410634, + 0.0066271331161260605, + 0.006177059840410948, + 0.007538166828453541, + 0.011151732876896858, + 0.01097952015697956, + 0.01126804854720831, + 0.007416092790663242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.014205760322511196 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro-static view of a microscopic world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny snail visible nestled inside a large, empty shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail's shell translucent amber and brown with visible growth lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the snail's pale muscular foot advancing over a smooth surface creating gentle ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the interior of the shell a shimmering mother-of-pearl reflecting pinks, blues, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369395062327385 + ], + "opposite_logprob_tokens": [ + -3.40869402885437 + ], + "logprob_reward": -0.03369395062327385, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369395062327385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1422394961118698, + 0.03349412605166435, + 0.031043600291013718, + 0.041232816874980927, + 0.08299941569566727, + 0.36655542254447937, + 0.06870367377996445, + 0.032571613788604736, + 0.05245053395628929, + 0.14870937168598175 + ], + "credit_map": [ + 0.010440808720886707, + 0.015816230326890945, + 0.013681966811418533, + 0.009950629435479641, + 0.03154390677809715, + 0.034184373915195465, + 0.013244758360087872, + 0.013376833871006966, + 0.0027485184837132692, + 0.0039036739617586136, + 0.003793609794229269, + 0.0035384504590183496, + 0.00630216347053647, + 0.003854676615446806, + 0.004981434438377619, + 0.004371597431600094, + 0.004439912736415863, + 0.004125649109482765, + 0.0027826749719679356, + 0.0044154152274131775, + 0.004957542754709721, + 0.004200003109872341, + 0.003378037828952074, + 0.002744363620877266, + 0.004496192093938589, + 0.004984524101018906, + 0.005684574134647846, + 0.00625893147662282, + 0.005576469004154205, + 0.0045322696678340435, + 0.003422339679673314, + 0.006277515087276697, + 0.009048405103385448, + 0.011904015205800533, + 0.009371615946292877, + 0.019119614735245705, + 0.009171703830361366, + 0.0066328225657343864, + 0.011450477875769138, + 0.006300762761384249, + 0.057121966034173965, + 0.036974936723709106, + 0.05332707613706589, + 0.04182600602507591, + 0.030058719217777252, + 0.037559617310762405, + 0.0422172024846077, + 0.06746987998485565, + 0.00805481057614088, + 0.0077543193474411964, + 0.010446221567690372, + 0.009433488361537457, + 0.010909518226981163, + 0.006853817962110043, + 0.007634560577571392, + 0.0076169357635080814, + 0.0033882083371281624, + 0.004264844581484795, + 0.0033374461345374584, + 0.0036797255743294954, + 0.006246291566640139, + 0.003757079131901264, + 0.0037245762068778276, + 0.004173439461737871, + 0.0052549345418810844, + 0.008850551210343838, + 0.005337401293218136, + 0.006766657344996929, + 0.005648212041705847, + 0.006584843620657921, + 0.005460862070322037, + 0.008547075092792511, + 0.016778694465756416, + 0.017791306599974632, + 0.013375501148402691, + 0.015607742592692375, + 0.02371402643620968, + 0.022182252258062363, + 0.023576674982905388, + 0.015683168545365334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.03369395062327385 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_15/index_tensor([2345, 2345])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_15/index_tensor([2345, 2345])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c2dfac4e7d22545bf80595dd1d5609b637b4c808 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_15/index_tensor([2345, 2345])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "subsequent low-angle, high-speed tracking shot captures a nimble black panther sprinting at full speed through the dense tropical jungle. Its glossy black fur reflects a gorgeous, flowing silky sheen in the dappled sunlight filtering through the treetops and onto the forest floor. Every flexing muscle is clearly visible as it races. Its body is stretched to its limit, its powerful limbs alternately stamping the damp, leaf-strewn ground with astonishing frequency. Each thrust sends dirt and dead leaves flying backward, leaving a dynamic trail. The camera remains parallel to the panther as massive ferns, tangled vines, and moss-covered tree trunks in the background whiz by, creating a powerful sense of speed and visual impact. The black panther has sharp eyes, staring intently ahead, its mouth slightly open, and its long tail swinging gracefully behind it to maintain perfect balance. The whole picture is illuminated by the luster flowing from the black panther's body against a cold background of dark green and dark brown, full of primitive vitality, tension and an awe-inspiring wild beauty.", + "seed": 14116, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_15/index_2345_seed_14116_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.051434208638966084, + "vlm_logprob": -0.051434208638966084, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.051434208638966084, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21203327178955078 + ], + "opposite_logprob_tokens": [ + -1.7120332717895508 + ], + "logprob_reward": -0.21203327178955078, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21203327178955078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07343748956918716, + 0.04802127927541733, + 0.04170241579413414, + 0.01188039779663086, + 0.05439557507634163, + 0.020405571907758713, + 0.11633692681789398, + 0.043201994150877, + 0.5727436542510986, + 0.01787472702562809 + ], + "credit_map": [ + 0.004319377243518829, + 0.021761400625109673, + 0.004396930802613497, + 0.021444285288453102, + 0.007846822962164879, + 0.006467037368565798, + 0.00461904751136899, + 0.0025825807824730873, + 0.0059966472908854485, + 0.007991346530616283, + 0.005147078540176153, + 0.004964039660990238, + 0.005508179310709238, + 0.005604345817118883, + 0.005488510709255934, + 0.007321130018681288, + 0.003017036709934473, + 0.01266051921993494, + 0.0032499348744750023, + 0.0021023431327193975, + 0.012314299121499062, + 0.0019443286582827568, + 0.0038235557731240988, + 0.002590396674349904, + 0.001792442868463695, + 0.00173652870580554, + 0.0011206084163859487, + 0.0018110196106135845, + 0.0016318379202857614, + 0.000723703415133059, + 0.0017584174638614058, + 0.001305839279666543, + 0.005558322183787823, + 0.0055658635683357716, + 0.004246909637004137, + 0.014494732022285461, + 0.0039048362523317337, + 0.0034030878450721502, + 0.013146472163498402, + 0.0040753562934696674, + 0.0016130578005686402, + 0.0016472459537908435, + 0.003955441527068615, + 0.0032767574302852154, + 0.001646515796892345, + 0.002349655842408538, + 0.004167459439486265, + 0.0017494375351816416, + 0.012276525609195232, + 0.013184497132897377, + 0.013553882017731667, + 0.02305375412106514, + 0.013974243775010109, + 0.018771251663565636, + 0.011920166201889515, + 0.009602608159184456, + 0.003512780414894223, + 0.002065717475488782, + 0.0029356186278164387, + 0.0019641213584691286, + 0.016440117731690407, + 0.0029048246797174215, + 0.011660956777632236, + 0.0017178544076159596, + 0.031937871128320694, + 0.04072393476963043, + 0.04055151343345642, + 0.16048240661621094, + 0.1672145575284958, + 0.031159086152911186, + 0.05373954772949219, + 0.04693472385406494, + 0.0015452541410923004, + 0.002639058278873563, + 0.0012870560167357326, + 0.0022368512582033873, + 0.002850133227184415, + 0.001637685694731772, + 0.0028752218931913376, + 0.002803466049954295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.21203327178955078 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01822640188038349 + ], + "opposite_logprob_tokens": [ + -4.018226623535156 + ], + "logprob_reward": -0.01822640188038349, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01822640188038349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13537344336509705, + 0.06836659461259842, + 0.056967999786138535, + 0.04508185014128685, + 0.27166983485221863, + 0.029739992693066597, + 0.07547368109226227, + 0.1980774700641632, + 0.0862414687871933, + 0.03300771117210388 + ], + "credit_map": [ + 0.009190592914819717, + 0.03426555544137955, + 0.015918893739581108, + 0.01908189430832863, + 0.017100835219025612, + 0.014672987163066864, + 0.01673293113708496, + 0.008409750647842884, + 0.007856360636651516, + 0.014370343647897243, + 0.008402446284890175, + 0.008047852665185928, + 0.012660332024097443, + 0.006185727659612894, + 0.005403877701610327, + 0.005439655389636755, + 0.004373881500214338, + 0.01889282464981079, + 0.0034540060441941023, + 0.0028961552307009697, + 0.01860271580517292, + 0.0030463365837931633, + 0.0033603592310100794, + 0.0023417226038873196, + 0.006449960172176361, + 0.006683798041194677, + 0.0034603907261043787, + 0.006143586710095406, + 0.006159099750220776, + 0.003586752573028207, + 0.007282988168299198, + 0.005315272603183985, + 0.0845634937286377, + 0.018807390704751015, + 0.014930729754269123, + 0.07933606952428818, + 0.01255130860954523, + 0.015800699591636658, + 0.030944619327783585, + 0.014735526405274868, + 0.005135998595505953, + 0.0026808944530785084, + 0.004269215278327465, + 0.004034164361655712, + 0.002162654884159565, + 0.00326420902274549, + 0.00414242735132575, + 0.004050428047776222, + 0.007401018403470516, + 0.012676632963120937, + 0.01593947783112526, + 0.0077845812775194645, + 0.011324583552777767, + 0.008502930402755737, + 0.007090844679623842, + 0.004753617104142904, + 0.016751380637288094, + 0.009551966562867165, + 0.010237648151814938, + 0.0096024414524436, + 0.06886737048625946, + 0.014754112809896469, + 0.05833718925714493, + 0.009975360706448555, + 0.005936049856245518, + 0.008138682693243027, + 0.006914674304425716, + 0.025064505636692047, + 0.02256101742386818, + 0.004273095168173313, + 0.007470702286809683, + 0.005882746074348688, + 0.002919068792834878, + 0.005493441596627235, + 0.0025749164633452892, + 0.0044542658142745495, + 0.005639962386339903, + 0.0030489640776067972, + 0.00419942382723093, + 0.004677668679505587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.01822640188038349 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021998509764671326 + ], + "opposite_logprob_tokens": [ + -6.127199649810791 + ], + "logprob_reward": -0.0021998509764671326, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021998509764671326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2179238200187683, + 0.09418807178735733, + 0.09434207528829575, + 0.05190679803490639, + 0.2494383007287979, + 0.013946198858320713, + 0.07022415101528168, + 0.04901036620140076, + 0.1302517205476761, + 0.028768470510840416 + ], + "credit_map": [ + 0.010307969525456429, + 0.0542178750038147, + 0.01802610419690609, + 0.06263511627912521, + 0.026887943968176842, + 0.016659362241625786, + 0.021281659603118896, + 0.007907801307737827, + 0.013314381241798401, + 0.014701110310852528, + 0.011024216189980507, + 0.0071561033837497234, + 0.015124804340302944, + 0.011106365360319614, + 0.010947046801447868, + 0.010814047418534756, + 0.0063482047989964485, + 0.029869653284549713, + 0.005596263334155083, + 0.006134680937975645, + 0.02923046052455902, + 0.005155215505510569, + 0.007173357997089624, + 0.004834245424717665, + 0.010368592105805874, + 0.005956618115305901, + 0.003853423986583948, + 0.004681632854044437, + 0.007317098323255777, + 0.0035507038701325655, + 0.009058911353349686, + 0.007119821384549141, + 0.03844374790787697, + 0.011752966791391373, + 0.007061597891151905, + 0.10159113258123398, + 0.007596464361995459, + 0.00798534881323576, + 0.06718726456165314, + 0.00781979225575924, + 0.002108067972585559, + 0.0012489896034821868, + 0.002507532713934779, + 0.001707795774564147, + 0.0010810509556904435, + 0.0013309248024597764, + 0.002902151783928275, + 0.0010596868814900517, + 0.006869072560220957, + 0.008388656191527843, + 0.0074167572893202305, + 0.013966703787446022, + 0.008554932661354542, + 0.012408090755343437, + 0.007005229126662016, + 0.005614720284938812, + 0.005504780914634466, + 0.003082172479480505, + 0.0029238953720778227, + 0.0033762252423912287, + 0.011881526559591293, + 0.004900831263512373, + 0.014207610860466957, + 0.003133324906229973, + 0.007145195733755827, + 0.010331719182431698, + 0.011224542744457722, + 0.02657380700111389, + 0.040587037801742554, + 0.008380210027098656, + 0.011843706481158733, + 0.01416550949215889, + 0.0023359612096101046, + 0.003661427414044738, + 0.0018160882173106074, + 0.003161664819344878, + 0.006352642551064491, + 0.0028804242610931396, + 0.003853065660223365, + 0.004707196727395058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.0021998509764671326 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064614057540894 + ], + "opposite_logprob_tokens": [ + -4.141064643859863 + ], + "logprob_reward": -0.016064614057540894, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064614057540894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05817355215549469, + 0.07459253817796707, + 0.03607454523444176, + 0.03999105468392372, + 0.0890473797917366, + 0.04040162265300751, + 0.12206338346004486, + 0.03562932088971138, + 0.48258644342422485, + 0.021440202370285988 + ], + "credit_map": [ + 0.003950124606490135, + 0.010974045842885971, + 0.006058989092707634, + 0.013818316161632538, + 0.006206155754625797, + 0.005817746743559837, + 0.007355261594057083, + 0.003992906305938959, + 0.009692463092505932, + 0.012596468441188335, + 0.008312749676406384, + 0.006651915609836578, + 0.010077021084725857, + 0.00884195975959301, + 0.0067475601099431515, + 0.01167239435017109, + 0.004418134223669767, + 0.006364106200635433, + 0.004584387876093388, + 0.0035063850227743387, + 0.007511024829000235, + 0.0022347436752170324, + 0.004338372498750687, + 0.0031173902098089457, + 0.006755971349775791, + 0.004734621848911047, + 0.0027793957851827145, + 0.0058760144747793674, + 0.005297706928104162, + 0.0030590402893722057, + 0.006016981787979603, + 0.005471318494528532, + 0.02007414773106575, + 0.005641035735607147, + 0.004412198439240456, + 0.03412635251879692, + 0.0035652732476592064, + 0.002837872365489602, + 0.014452837407588959, + 0.0039376551285386086, + 0.0035788880195468664, + 0.003074029926210642, + 0.007395957130938768, + 0.006402985192835331, + 0.003396485699340701, + 0.004235723055899143, + 0.00875405129045248, + 0.00356349628418684, + 0.009774204343557358, + 0.018004367128014565, + 0.025433160364627838, + 0.025203749537467957, + 0.00978273618966341, + 0.011975591070950031, + 0.01283319666981697, + 0.00905636977404356, + 0.004367765970528126, + 0.0023095172364264727, + 0.0027376115322113037, + 0.0017937527736648917, + 0.013551879674196243, + 0.002743530785664916, + 0.006650135386735201, + 0.0014751265989616513, + 0.02509045973420143, + 0.035405196249485016, + 0.03375675901770592, + 0.1268017441034317, + 0.16146349906921387, + 0.026824548840522766, + 0.04117254912853241, + 0.032071638852357864, + 0.0019971062429249287, + 0.0028157266788184643, + 0.0013773577520623803, + 0.002026741858571768, + 0.003957451321184635, + 0.002172919688746333, + 0.003435635706409812, + 0.00365726207382977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.016064614057540894 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646904490888119 + ], + "opposite_logprob_tokens": [ + -4.7586469650268555 + ], + "logprob_reward": -0.008646904490888119, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646904490888119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13298603892326355, + 0.05712931603193283, + 0.14059565961360931, + 0.06220364198088646, + 0.11887364089488983, + 0.030507748946547508, + 0.07884231209754944, + 0.0691935271024704, + 0.28525981307029724, + 0.024408288300037384 + ], + "credit_map": [ + 0.008027433417737484, + 0.026872824877500534, + 0.013045496307313442, + 0.033122316002845764, + 0.016283802688121796, + 0.013771983794867992, + 0.015879929065704346, + 0.005982250906527042, + 0.007219824939966202, + 0.010775890201330185, + 0.006507678423076868, + 0.004928311333060265, + 0.009007852524518967, + 0.008141890168190002, + 0.005598875693976879, + 0.004948987625539303, + 0.008983464911580086, + 0.05211443454027176, + 0.006277516484260559, + 0.004038593266159296, + 0.050078049302101135, + 0.0064680143259465694, + 0.008007991127669811, + 0.004627602640539408, + 0.009842652827501297, + 0.009472501464188099, + 0.0033094221726059914, + 0.00758954556658864, + 0.008666597306728363, + 0.0052429912611842155, + 0.011829551309347153, + 0.006250376347452402, + 0.017656000331044197, + 0.009060535579919815, + 0.005534833297133446, + 0.03935509920120239, + 0.008565323427319527, + 0.007135265972465277, + 0.02372860535979271, + 0.007837963290512562, + 0.0032719802111387253, + 0.003101250622421503, + 0.005299889948219061, + 0.003453383222222328, + 0.003006281331181526, + 0.003315072041004896, + 0.005022715777158737, + 0.004037178121507168, + 0.00530726695433259, + 0.012039744295179844, + 0.020307481288909912, + 0.011301212944090366, + 0.005145392380654812, + 0.008503003977239132, + 0.00852375477552414, + 0.007714452687650919, + 0.009076287969946861, + 0.004414968192577362, + 0.006507853977382183, + 0.005365455057471991, + 0.01409422978758812, + 0.005481949541717768, + 0.01894398033618927, + 0.0053087989799678326, + 0.018611006438732147, + 0.023754877969622612, + 0.024257006123661995, + 0.07421624660491943, + 0.06445928663015366, + 0.019714100286364555, + 0.031268734484910965, + 0.028978558257222176, + 0.0034751328639686108, + 0.004257150925695896, + 0.0022147088311612606, + 0.002544337185099721, + 0.0040469118393957615, + 0.002299676416441798, + 0.002483112970367074, + 0.003087259130552411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.008646904490888119 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, high-speed tracking shot captures a nimble black panther sprinting at full speed through the dense tropical jungle. Its glossy black fur reflects a gorgeous, flowing silky sheen in the dappled sunlight filtering through the treetops and onto the forest floor. Every flexing muscle is clearly visible as it races. Its body is stretched to its limit, its powerful limbs alternately stamping the damp, leaf-strewn ground with astonishing frequency. Each thrust sends dirt and dead leaves flying backward, leaving a dynamic trail. The camera remains parallel to the panther as massive ferns, tangled vines, and moss-covered tree trunks in the background whiz by, creating a powerful sense of speed and visual impact. The black panther has sharp eyes, staring intently ahead, its mouth slightly open, and its long tail swinging gracefully behind it to maintain perfect balance. The whole picture is illuminated by the luster flowing from the black panther's body against a cold background of dark green and dark brown, full of primitive vitality, tension and an awe-inspiring wild beauty.", + "seed": 14117, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_15/index_2345_seed_14117_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1250076641794294, + "vlm_logprob": -0.1250076641794294, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.1250076641794294, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5852935314178467 + ], + "opposite_logprob_tokens": [ + -0.8352935314178467 + ], + "logprob_reward": -0.5852935314178467, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5852935314178467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23464041948318481, + 0.07271148264408112, + 0.2592690885066986, + 0.035871293395757675, + 0.032969776540994644, + 0.029901035130023956, + 0.030327534303069115, + 0.13238583505153656, + 0.1158844381570816, + 0.05603911727666855 + ], + "credit_map": [ + 0.019133439287543297, + 0.018849212676286697, + 0.029433611780405045, + 0.030883248895406723, + 0.03417659178376198, + 0.029184957966208458, + 0.05185991898179054, + 0.021119432523846626, + 0.0061814612708985806, + 0.005053714383393526, + 0.006078679580241442, + 0.009050354361534119, + 0.01498853974044323, + 0.010443096980452538, + 0.012842395342886448, + 0.008073240518569946, + 0.03308458253741264, + 0.012091772630810738, + 0.0706220418214798, + 0.03712071105837822, + 0.020199185237288475, + 0.03815754875540733, + 0.029401905834674835, + 0.01859131082892418, + 0.0041667018085718155, + 0.00450522406026721, + 0.004388205241411924, + 0.005406989250332117, + 0.002869244897738099, + 0.0050157480873167515, + 0.005901679862290621, + 0.0036175039131194353, + 0.0022048749960958958, + 0.004284068010747433, + 0.0032586397137492895, + 0.005407202523201704, + 0.003135226434096694, + 0.0027833518106490374, + 0.0070816874504089355, + 0.004814725834876299, + 0.0019630708266049623, + 0.00356583041138947, + 0.0025789474602788687, + 0.007390039972960949, + 0.003084616269916296, + 0.001843022764660418, + 0.0055128284730017185, + 0.003962676972150803, + 0.0015867267502471805, + 0.0032031277660280466, + 0.004047658760100603, + 0.005601617507636547, + 0.002707229694351554, + 0.0036084505263715982, + 0.005703027825802565, + 0.003869695821776986, + 0.006830328144133091, + 0.003795596770942211, + 0.007830306887626648, + 0.011790969409048557, + 0.007703911513090134, + 0.041046805679798126, + 0.042195938527584076, + 0.011191983707249165, + 0.01342869084328413, + 0.009508306160569191, + 0.02516101486980915, + 0.03766510263085365, + 0.004922151565551758, + 0.008630616590380669, + 0.007680932059884071, + 0.008887623436748981, + 0.0042574466206133366, + 0.005352603271603584, + 0.007215837016701698, + 0.007455318234860897, + 0.00831835437566042, + 0.006235496141016483, + 0.009626641869544983, + 0.0075774239376187325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.5852935314178467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012571611441671848 + ], + "opposite_logprob_tokens": [ + -4.387571811676025 + ], + "logprob_reward": -0.012571611441671848, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012571611441671848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17152291536331177, + 0.07766583561897278, + 0.12008777260780334, + 0.1363166719675064, + 0.09649086743593216, + 0.05405762791633606, + 0.04486726596951485, + 0.11721210181713104, + 0.09698338061571121, + 0.0847955271601677 + ], + "credit_map": [ + 0.010470520704984665, + 0.0202923770993948, + 0.02433633618056774, + 0.021848350763320923, + 0.02165958099067211, + 0.029861412942409515, + 0.0251191183924675, + 0.01793522946536541, + 0.008423208259046078, + 0.006391570903360844, + 0.008599127642810345, + 0.00852547399699688, + 0.014420468360185623, + 0.010754793882369995, + 0.01158900372684002, + 0.00896218791604042, + 0.015475505031645298, + 0.0075114406645298, + 0.029874371364712715, + 0.0188225656747818, + 0.008327900432050228, + 0.016734367236495018, + 0.013040312565863132, + 0.010301314294338226, + 0.0124545618891716, + 0.01563548855483532, + 0.014732440002262592, + 0.019394300878047943, + 0.019047651439905167, + 0.02267879992723465, + 0.01832861267030239, + 0.014044827781617641, + 0.0063087791204452515, + 0.010435855016112328, + 0.007656876929104328, + 0.0159681998193264, + 0.010256716050207615, + 0.018096869811415672, + 0.010771279223263264, + 0.01699630357325077, + 0.003104591742157936, + 0.007332632318139076, + 0.005641714669764042, + 0.010601622983813286, + 0.006527361460030079, + 0.003954362124204636, + 0.007674993947148323, + 0.009220353327691555, + 0.0027369691524654627, + 0.004320601001381874, + 0.006816401146352291, + 0.008579003624618053, + 0.0027747144922614098, + 0.0058212787844240665, + 0.005814024712890387, + 0.008004273287951946, + 0.00552883418276906, + 0.0037049513775855303, + 0.00844530388712883, + 0.009533059783279896, + 0.006657683756202459, + 0.03758871555328369, + 0.03742402791976929, + 0.008329535834491253, + 0.008613785728812218, + 0.008127031847834587, + 0.02591821551322937, + 0.03366191312670708, + 0.002976392162963748, + 0.005085359793156385, + 0.006052798125892878, + 0.006547893397510052, + 0.00513653177767992, + 0.008426270447671413, + 0.010874740779399872, + 0.012862041592597961, + 0.012529741041362286, + 0.010768015868961811, + 0.014578850008547306, + 0.009619341231882572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.012571611441671848 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003192092990502715 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.003192092990502715, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003192092990502715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1727832704782486, + 0.03830207511782646, + 0.2376859337091446, + 0.056814044713974, + 0.06551515311002731, + 0.11792664229869843, + 0.03770078346133232, + 0.051786188036203384, + 0.17977844178676605, + 0.04170747101306915 + ], + "credit_map": [ + 0.011684475466609001, + 0.021080251783132553, + 0.024272102862596512, + 0.0209575816988945, + 0.021838661283254623, + 0.025673706084489822, + 0.03020433522760868, + 0.0170721597969532, + 0.0037897920701652765, + 0.0033342589158564806, + 0.0033443307038396597, + 0.004531126003712416, + 0.007261466234922409, + 0.005177394952625036, + 0.006402114871889353, + 0.004461592063307762, + 0.028092963621020317, + 0.0163659006357193, + 0.04970651492476463, + 0.03894061967730522, + 0.021612292155623436, + 0.02699785865843296, + 0.03198695182800293, + 0.0239828173071146, + 0.0063527547754347324, + 0.00624379375949502, + 0.006563653703778982, + 0.009528536349534988, + 0.005656609311699867, + 0.009406938217580318, + 0.007433121092617512, + 0.005628636106848717, + 0.0038413405418395996, + 0.0100175142288208, + 0.004104770720005035, + 0.009393748827278614, + 0.005877583287656307, + 0.012687593698501587, + 0.009098856709897518, + 0.010493740439414978, + 0.007379805203527212, + 0.018833111971616745, + 0.012250829488039017, + 0.017901750281453133, + 0.02284124493598938, + 0.009756200946867466, + 0.017280766740441322, + 0.011682933196425438, + 0.0032189737539738417, + 0.0026731176767498255, + 0.004730958491563797, + 0.006951510906219482, + 0.0024276552721858025, + 0.0041247084736824036, + 0.004269667435437441, + 0.009304191917181015, + 0.005064805969595909, + 0.003901090705767274, + 0.006114631425589323, + 0.009882338345050812, + 0.005499731749296188, + 0.006285451352596283, + 0.008364582434296608, + 0.00667355814948678, + 0.015597025863826275, + 0.01327117346227169, + 0.047622211277484894, + 0.06392495334148407, + 0.007110308390110731, + 0.009584047831594944, + 0.011625384911894798, + 0.011043348349630833, + 0.0034935218282043934, + 0.004026408307254314, + 0.00529331574216485, + 0.007016444578766823, + 0.006491116713732481, + 0.005305523052811623, + 0.004702676087617874, + 0.005378464236855507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.003192092990502715 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193184673786163 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014193184673786163, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193184673786163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13471679389476776, + 0.05378153175115585, + 0.1352994740009308, + 0.05289802700281143, + 0.05414560064673424, + 0.04177644103765488, + 0.04365317150950432, + 0.32272911071777344, + 0.03467794135212898, + 0.12632189691066742 + ], + "credit_map": [ + 0.00987901072949171, + 0.010580409318208694, + 0.01904168911278248, + 0.016408147290349007, + 0.019099848344922066, + 0.01860092207789421, + 0.028727663680911064, + 0.012379092164337635, + 0.004202932119369507, + 0.004320262465626001, + 0.004467517137527466, + 0.005071834195405245, + 0.012821650132536888, + 0.008131353184580803, + 0.009184428490698338, + 0.005581554491072893, + 0.014795833267271519, + 0.00840822048485279, + 0.03156105801463127, + 0.021462803706526756, + 0.011570250615477562, + 0.018395768478512764, + 0.01893097348511219, + 0.01017457339912653, + 0.005801685154438019, + 0.005484541412442923, + 0.006607802584767342, + 0.007671650964766741, + 0.005813382565975189, + 0.007710497826337814, + 0.007343510165810585, + 0.0064649577252566814, + 0.004793023224920034, + 0.01121377944946289, + 0.005799249745905399, + 0.007740382570773363, + 0.004665903281420469, + 0.004267717245966196, + 0.008270434103906155, + 0.007395110093057156, + 0.003410111414268613, + 0.005984935909509659, + 0.0034555052407085896, + 0.00791921280324459, + 0.0062936139293015, + 0.0032318003941327333, + 0.005709602497518063, + 0.00577166024595499, + 0.003250609152019024, + 0.004254377447068691, + 0.006684557534754276, + 0.006887286435812712, + 0.0038154679350554943, + 0.008171889930963516, + 0.00563125591725111, + 0.004957725293934345, + 0.021741414442658424, + 0.007008603308349848, + 0.02571561001241207, + 0.025632238015532494, + 0.014945351518690586, + 0.09791442006826401, + 0.10426276177167892, + 0.025508707389235497, + 0.0030073446687310934, + 0.00567813403904438, + 0.005258643068373203, + 0.00629233755171299, + 0.0037398976273834705, + 0.003816677490249276, + 0.00374001101590693, + 0.0031448944937437773, + 0.008769313804805279, + 0.011180986650288105, + 0.015297622419893742, + 0.019850920885801315, + 0.016144147142767906, + 0.015202870592474937, + 0.024551643058657646, + 0.015324397943913937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.014193184673786163 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787900373339653 + ], + "opposite_logprob_tokens": [ + -4.6347880363464355 + ], + "logprob_reward": -0.009787900373339653, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787900373339653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.260053813457489, + 0.07400055974721909, + 0.16682302951812744, + 0.0533224418759346, + 0.06827987730503082, + 0.04390040412545204, + 0.025021439418196678, + 0.1477157175540924, + 0.0968022495508194, + 0.06408043950796127 + ], + "credit_map": [ + 0.016571305692195892, + 0.021467726677656174, + 0.03549473360180855, + 0.032697129994630814, + 0.04215149208903313, + 0.03474406152963638, + 0.051642145961523056, + 0.025285234674811363, + 0.007947081699967384, + 0.005436826031655073, + 0.007470000069588423, + 0.010042921639978886, + 0.014284214936196804, + 0.01039073709398508, + 0.00963663775473833, + 0.00879213958978653, + 0.019711561501026154, + 0.00954325683414936, + 0.04159412533044815, + 0.025918766856193542, + 0.013765761628746986, + 0.02386423572897911, + 0.019724510610103607, + 0.01270080916583538, + 0.00618929835036397, + 0.005409938748925924, + 0.006302943453192711, + 0.0075632729567587376, + 0.006662755738943815, + 0.010157729499042034, + 0.006272518076002598, + 0.004763985984027386, + 0.004172313958406448, + 0.008335362188518047, + 0.004538156557828188, + 0.009858676232397556, + 0.00754805002361536, + 0.014133610762655735, + 0.007288362365216017, + 0.01240534707903862, + 0.0027699978090822697, + 0.006247577257454395, + 0.004012278281152248, + 0.008473388850688934, + 0.007250814698636532, + 0.0030904337763786316, + 0.005256761331111193, + 0.006799147464334965, + 0.002298156265169382, + 0.0018788720481097698, + 0.004021934233605862, + 0.004683235660195351, + 0.002414531074464321, + 0.0032141096889972687, + 0.003491714596748352, + 0.0030188849195837975, + 0.0070044673047959805, + 0.004479157738387585, + 0.008479645475745201, + 0.013340644538402557, + 0.00825443770736456, + 0.047683026641607285, + 0.047895003110170364, + 0.010579343885183334, + 0.007676098495721817, + 0.010112719610333443, + 0.01838001050055027, + 0.017910687252879143, + 0.009498211555182934, + 0.013651307672262192, + 0.010221721604466438, + 0.009351490065455437, + 0.00467395456507802, + 0.005555026698857546, + 0.006883156485855579, + 0.009573183953762054, + 0.008464602753520012, + 0.007859105244278908, + 0.00954948365688324, + 0.011521927081048489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.009787900373339653 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, high-speed tracking shot captures a nimble black panther sprinting at full speed through the dense tropical jungle. Its glossy black fur reflects a gorgeous, flowing silky sheen in the dappled sunlight filtering through the treetops and onto the forest floor. Every flexing muscle is clearly visible as it races. Its body is stretched to its limit, its powerful limbs alternately stamping the damp, leaf-strewn ground with astonishing frequency. Each thrust sends dirt and dead leaves flying backward, leaving a dynamic trail. The camera remains parallel to the panther as massive ferns, tangled vines, and moss-covered tree trunks in the background whiz by, creating a powerful sense of speed and visual impact. The black panther has sharp eyes, staring intently ahead, its mouth slightly open, and its long tail swinging gracefully behind it to maintain perfect balance. The whole picture is illuminated by the luster flowing from the black panther's body against a cold background of dark green and dark brown, full of primitive vitality, tension and an awe-inspiring wild beauty.", + "seed": 14118, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_15/index_2345_seed_14118_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.15319374930113555, + "vlm_logprob": -0.15319374930113555, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.15319374930113555, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7042431235313416 + ], + "opposite_logprob_tokens": [ + -0.7042431235313416 + ], + "logprob_reward": -0.7042431235313416, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7042431235313416, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07900956273078918, + 0.1652599722146988, + 0.03311189264059067, + 0.2649204432964325, + 0.05376715585589409, + 0.14176301658153534, + 0.04573029652237892, + 0.07142489403486252, + 0.11013063788414001, + 0.0348820835351944 + ], + "credit_map": [ + 0.007223706226795912, + 0.0061818561516702175, + 0.009226717986166477, + 0.011483732610940933, + 0.00954387616366148, + 0.012640942819416523, + 0.0124399084597826, + 0.010268823243677616, + 0.012721261009573936, + 0.005660960916429758, + 0.011532938107848167, + 0.025347648188471794, + 0.06436270475387573, + 0.020724834874272346, + 0.01161124836653471, + 0.013298382051289082, + 0.002503135707229376, + 0.004118369892239571, + 0.003701373003423214, + 0.00423468416556716, + 0.0044722892343997955, + 0.00643069576472044, + 0.003585065482184291, + 0.0040662772953510284, + 0.03954366222023964, + 0.05567143112421036, + 0.026357844471931458, + 0.03625788167119026, + 0.03961026668548584, + 0.025119386613368988, + 0.022094134241342545, + 0.02026587724685669, + 0.005423577968031168, + 0.00822448544204235, + 0.008308122865855694, + 0.010652492754161358, + 0.005988040938973427, + 0.004129658918827772, + 0.006849462166428566, + 0.004191314801573753, + 0.024242861196398735, + 0.014985729940235615, + 0.019322095438838005, + 0.019398776814341545, + 0.01071864366531372, + 0.017309783026576042, + 0.017919747158885002, + 0.01786538027226925, + 0.004652359057217836, + 0.006585612427443266, + 0.005745690781623125, + 0.00563691183924675, + 0.007654187735170126, + 0.006213190034031868, + 0.004375064745545387, + 0.004867276176810265, + 0.006453999318182468, + 0.007740960922092199, + 0.011238777078688145, + 0.007185696624219418, + 0.010839778929948807, + 0.00885002315044403, + 0.014194825664162636, + 0.004920839332044125, + 0.007207168266177177, + 0.013075118884444237, + 0.012593631632626057, + 0.014638617634773254, + 0.016896143555641174, + 0.01388992927968502, + 0.020383769646286964, + 0.011446262709796429, + 0.0032464065589010715, + 0.004904140252619982, + 0.0030194539576768875, + 0.003988826647400856, + 0.003910840954631567, + 0.006630869582295418, + 0.006464485079050064, + 0.002717060735449195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.7042431235313416 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02062411420047283 + ], + "opposite_logprob_tokens": [ + -3.8956241607666016 + ], + "logprob_reward": -0.02062411420047283, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02062411420047283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06775343418121338, + 0.31191936135292053, + 0.01677742414176464, + 0.04964093863964081, + 0.021922104060649872, + 0.04716170206665993, + 0.09366854280233383, + 0.052007172256708145, + 0.11499208956956863, + 0.2241571694612503 + ], + "credit_map": [ + 0.00860053300857544, + 0.005728738382458687, + 0.0110629228875041, + 0.007381955627351999, + 0.007960793562233448, + 0.011876905336976051, + 0.0073121110908687115, + 0.007829482667148113, + 0.014786963351070881, + 0.010304437018930912, + 0.012042762711644173, + 0.07202951610088348, + 0.12392640858888626, + 0.04676445201039314, + 0.02046923153102398, + 0.011595599353313446, + 0.0020673491526395082, + 0.002406937535852194, + 0.0023677488788962364, + 0.0018295746995136142, + 0.001932778162881732, + 0.0025579556822776794, + 0.0018095561536028981, + 0.0018055237596854568, + 0.00503065949305892, + 0.01195459347218275, + 0.00660634133964777, + 0.00483805313706398, + 0.0071147712878882885, + 0.005510041024535894, + 0.005276104435324669, + 0.0033103798050433397, + 0.0041338615119457245, + 0.0027668853290379047, + 0.0038846435490995646, + 0.0034367465414106846, + 0.002151343272998929, + 0.0016048463294282556, + 0.0024416493251919746, + 0.001502130413427949, + 0.005445719696581364, + 0.005238393787294626, + 0.007877115160226822, + 0.008382716216146946, + 0.006213915999978781, + 0.0052838572300970554, + 0.0046690977178514, + 0.004050893243402243, + 0.008794807828962803, + 0.015125916339457035, + 0.010618560016155243, + 0.012830488383769989, + 0.014155891723930836, + 0.011402683332562447, + 0.011404699645936489, + 0.009335500188171864, + 0.0059058768674731255, + 0.007369017228484154, + 0.005491295829415321, + 0.005077274516224861, + 0.00725762452930212, + 0.008296255953609943, + 0.00634929770603776, + 0.006260528229176998, + 0.009259657002985477, + 0.017620714381337166, + 0.015042424201965332, + 0.009391489438712597, + 0.020833561196923256, + 0.012771285139024258, + 0.020883241668343544, + 0.009189719334244728, + 0.021544305607676506, + 0.014375755563378334, + 0.009129256010055542, + 0.016793139278888702, + 0.019239554181694984, + 0.09552143514156342, + 0.03481597825884819, + 0.012737756595015526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.02062411420047283 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036153923720121384 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.0036153923720121384, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036153923720121384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20578818023204803, + 0.29245057702064514, + 0.055769238620996475, + 0.09478966146707535, + 0.057209137827157974, + 0.05113264173269272, + 0.04587149992585182, + 0.05120764672756195, + 0.07879124581813812, + 0.06699013710021973 + ], + "credit_map": [ + 0.01995515450835228, + 0.015954671427607536, + 0.027111226692795753, + 0.022463630884885788, + 0.02977261133491993, + 0.04265069589018822, + 0.025298921391367912, + 0.022581253200769424, + 0.03741765394806862, + 0.013189476914703846, + 0.021455898880958557, + 0.03726897016167641, + 0.08769022673368454, + 0.025738626718521118, + 0.03702973201870918, + 0.032660000026226044, + 0.004148608073592186, + 0.006679844576865435, + 0.00811136607080698, + 0.006126514170318842, + 0.009289668872952461, + 0.007568119093775749, + 0.006912443321198225, + 0.006932673044502735, + 0.013275041244924068, + 0.01770368218421936, + 0.01099478267133236, + 0.014012711122632027, + 0.014159435406327248, + 0.007532098796218634, + 0.008177192881703377, + 0.008934713900089264, + 0.010007248260080814, + 0.01262879092246294, + 0.01209535263478756, + 0.00568078737705946, + 0.005407481919974089, + 0.003112661885097623, + 0.004044642206281424, + 0.004232172854244709, + 0.008637291379272938, + 0.006451614201068878, + 0.005536780226975679, + 0.006292048841714859, + 0.005728569347411394, + 0.006239541806280613, + 0.005598307121545076, + 0.00664849067106843, + 0.004818660672754049, + 0.007712587248533964, + 0.003691119374707341, + 0.005035607609897852, + 0.0073742950335145, + 0.005927213933318853, + 0.004556731786578894, + 0.006755282171070576, + 0.005604073405265808, + 0.005947391036897898, + 0.006841249298304319, + 0.005098965018987656, + 0.006760603282600641, + 0.00746197672560811, + 0.009042399935424328, + 0.004450992681086063, + 0.0064356946386396885, + 0.010367671959102154, + 0.009928998537361622, + 0.007874920964241028, + 0.010780202224850655, + 0.012209438718855381, + 0.010726233944296837, + 0.01046808622777462, + 0.005366943776607513, + 0.004297118633985519, + 0.0034327490720897913, + 0.005591672379523516, + 0.004246214404702187, + 0.018394511193037033, + 0.01834692992269993, + 0.007313998881727457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0036153923720121384 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283882066607475 + ], + "opposite_logprob_tokens": [ + -3.7732839584350586 + ], + "logprob_reward": -0.023283882066607475, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283882066607475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06981983780860901, + 0.11448366940021515, + 0.022581469267606735, + 0.10900896787643433, + 0.041042182594537735, + 0.10501685738563538, + 0.07864658534526825, + 0.0441754013299942, + 0.06659379601478577, + 0.3486311733722687 + ], + "credit_map": [ + 0.007109280675649643, + 0.006268253084272146, + 0.012216423638164997, + 0.007234370801597834, + 0.01083096768707037, + 0.010043291375041008, + 0.007856294512748718, + 0.00826095137745142, + 0.011166900396347046, + 0.005873473361134529, + 0.008466172963380814, + 0.038414161652326584, + 0.01544442493468523, + 0.007910637184977531, + 0.01620521768927574, + 0.011002684012055397, + 0.0027464653830975294, + 0.003556240117177367, + 0.0030431405175477266, + 0.002083188621327281, + 0.002705734223127365, + 0.0029206813778728247, + 0.0025958921760320663, + 0.002930128714069724, + 0.016765864565968513, + 0.02804568223655224, + 0.012985502369701862, + 0.011597069911658764, + 0.012984756380319595, + 0.008205394260585308, + 0.008518303744494915, + 0.009906405583024025, + 0.004822251386940479, + 0.007277676835656166, + 0.005356158595532179, + 0.006471874658018351, + 0.004599976819008589, + 0.0036675366573035717, + 0.005440018139779568, + 0.0034066890366375446, + 0.014923712238669395, + 0.014139835722744465, + 0.015097452327609062, + 0.016335299238562584, + 0.012120789848268032, + 0.012132905423641205, + 0.009785167872905731, + 0.010481695644557476, + 0.007937648333609104, + 0.014246491715312004, + 0.01073218323290348, + 0.009177685715258121, + 0.011335935443639755, + 0.00815033633261919, + 0.008499231189489365, + 0.008567075245082378, + 0.004247733391821384, + 0.005176224280148745, + 0.006387737113982439, + 0.0044202604331076145, + 0.0063450271263718605, + 0.006172499619424343, + 0.006798262242227793, + 0.004627656657248735, + 0.004368855617940426, + 0.008208232000470161, + 0.008763673715293407, + 0.0054457904770970345, + 0.009669088758528233, + 0.009350486099720001, + 0.014415724202990532, + 0.006371944677084684, + 0.036363884806632996, + 0.023050818592309952, + 0.014281358569860458, + 0.027279086410999298, + 0.031300049275159836, + 0.14894400537014008, + 0.04739394411444664, + 0.020018041133880615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.023283882066607475 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202234335243702 + ], + "opposite_logprob_tokens": [ + -4.264202117919922 + ], + "logprob_reward": -0.014202234335243702, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202234335243702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05783946067094803, + 0.21427541971206665, + 0.03723243251442909, + 0.09632158279418945, + 0.029039502143859863, + 0.09008090198040009, + 0.04364283010363579, + 0.04866711050271988, + 0.055177152156829834, + 0.3277236223220825 + ], + "credit_map": [ + 0.006077693775296211, + 0.004990554880350828, + 0.009546995162963867, + 0.007803566288203001, + 0.007028465624898672, + 0.008238734677433968, + 0.007211457937955856, + 0.006941991858184338, + 0.02930508926510811, + 0.01087696198374033, + 0.018063873052597046, + 0.04725591465830803, + 0.018922211602330208, + 0.03339652344584465, + 0.03076181933283806, + 0.025693031027913094, + 0.004600475076586008, + 0.0059267012402415276, + 0.005535203963518143, + 0.0034291520714759827, + 0.004145513754338026, + 0.007076564710587263, + 0.0037276851944625378, + 0.0027911358047276735, + 0.013134900480508804, + 0.022316178306937218, + 0.01343354769051075, + 0.009650121442973614, + 0.010941466316580772, + 0.007418646477162838, + 0.00757755758240819, + 0.011849164962768555, + 0.0037668903823941946, + 0.003504489315673709, + 0.005102687515318394, + 0.004940452985465527, + 0.0028241744730621576, + 0.0021113906987011433, + 0.004609449300915003, + 0.0021799698006361723, + 0.017002947628498077, + 0.00967550091445446, + 0.008382036350667477, + 0.010255862958729267, + 0.009350319392979145, + 0.010438277386128902, + 0.010358636267483234, + 0.014617327600717545, + 0.004089600872248411, + 0.006106619257479906, + 0.005546543747186661, + 0.00525181507691741, + 0.006447855848819017, + 0.005103663075715303, + 0.005040066782385111, + 0.006056663580238819, + 0.005452112294733524, + 0.0067353686317801476, + 0.004803883843123913, + 0.0042520505376160145, + 0.00758398650214076, + 0.008301204070448875, + 0.0057885837741196156, + 0.005749919451773167, + 0.005508817732334137, + 0.007283022627234459, + 0.007706289645284414, + 0.005445240531116724, + 0.01004142314195633, + 0.007183513138443232, + 0.007733709178864956, + 0.004275134298950434, + 0.02719130739569664, + 0.023708347231149673, + 0.013669582083821297, + 0.02690511755645275, + 0.027452031150460243, + 0.13622473180294037, + 0.05163181945681572, + 0.020940693095326424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.014202234335243702 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, high-speed tracking shot captures a nimble black panther sprinting at full speed through the dense tropical jungle. Its glossy black fur reflects a gorgeous, flowing silky sheen in the dappled sunlight filtering through the treetops and onto the forest floor. Every flexing muscle is clearly visible as it races. Its body is stretched to its limit, its powerful limbs alternately stamping the damp, leaf-strewn ground with astonishing frequency. Each thrust sends dirt and dead leaves flying backward, leaving a dynamic trail. The camera remains parallel to the panther as massive ferns, tangled vines, and moss-covered tree trunks in the background whiz by, creating a powerful sense of speed and visual impact. The black panther has sharp eyes, staring intently ahead, its mouth slightly open, and its long tail swinging gracefully behind it to maintain perfect balance. The whole picture is illuminated by the luster flowing from the black panther's body against a cold background of dark green and dark brown, full of primitive vitality, tension and an awe-inspiring wild beauty.", + "seed": 14119, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_15/index_2345_seed_14119_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2064557391218841, + "vlm_logprob": -0.2064557391218841, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.2064557391218841, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9796859622001648 + ], + "opposite_logprob_tokens": [ + -0.4796859622001648 + ], + "logprob_reward": -0.9796859622001648, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9796859622001648, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0850091353058815, + 0.06749933958053589, + 0.04582689329981804, + 0.13210168480873108, + 0.07025691121816635, + 0.030832968652248383, + 0.2619054913520813, + 0.09702659398317337, + 0.04837300628423691, + 0.1611679345369339 + ], + "credit_map": [ + 0.01211195532232523, + 0.010902785696089268, + 0.013203286565840244, + 0.013133661821484566, + 0.00730543676763773, + 0.011497315019369125, + 0.010747079737484455, + 0.00610762694850564, + 0.005477180704474449, + 0.008156736381351948, + 0.006914920173585415, + 0.011843073181807995, + 0.007398617919534445, + 0.013183556497097015, + 0.008950936608016491, + 0.00557433208450675, + 0.004208723548799753, + 0.0068732998333871365, + 0.00577878812327981, + 0.006932615302503109, + 0.006115974858403206, + 0.004640638362616301, + 0.00528344651684165, + 0.005993413273245096, + 0.011346434243023396, + 0.015253758057951927, + 0.01883929781615734, + 0.026782380416989326, + 0.014094742015004158, + 0.013635160401463509, + 0.020544586703181267, + 0.011605336330831051, + 0.007580543402582407, + 0.006871712394058704, + 0.004770294763147831, + 0.01196750532835722, + 0.010341187939047813, + 0.008094221353530884, + 0.010601281188428402, + 0.010030175559222698, + 0.0024290354922413826, + 0.003255183110013604, + 0.00297443475574255, + 0.005865932442247868, + 0.004667217377573252, + 0.0046911705285310745, + 0.004260258283466101, + 0.0026897399220615625, + 0.013139977119863033, + 0.012435922399163246, + 0.024436544626951218, + 0.07585979998111725, + 0.02536759153008461, + 0.03383192420005798, + 0.0392659567296505, + 0.037567801773548126, + 0.014890668913722038, + 0.009627161547541618, + 0.010334677062928677, + 0.015009526163339615, + 0.012480245903134346, + 0.010515854693949223, + 0.011437913402915001, + 0.012730560265481472, + 0.00818540807813406, + 0.006806553807109594, + 0.007853672839701176, + 0.0059160348027944565, + 0.005210634786635637, + 0.0030511885415762663, + 0.007470377255231142, + 0.003879143390804529, + 0.013117418624460697, + 0.059364113956689835, + 0.0446876585483551, + 0.008831087499856949, + 0.011656221002340317, + 0.010428829118609428, + 0.00625687837600708, + 0.006825749762356281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 3 + ], + "selected_reward": -0.9796859622001648 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01610778272151947 + ], + "opposite_logprob_tokens": [ + -4.141107559204102 + ], + "logprob_reward": -0.01610778272151947, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01610778272151947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13430529832839966, + 0.033382780849933624, + 0.05588095635175705, + 0.08009166270494461, + 0.054728321731090546, + 0.16450467705726624, + 0.20117199420928955, + 0.06134474650025368, + 0.04523526877164841, + 0.16935428977012634 + ], + "credit_map": [ + 0.013499862514436245, + 0.017664870247244835, + 0.020409341901540756, + 0.019411837682127953, + 0.01337486132979393, + 0.01919558271765709, + 0.01899048686027527, + 0.011758457869291306, + 0.0026644442696124315, + 0.0037723099812865257, + 0.0037804648745805025, + 0.006075756624341011, + 0.003031444735825062, + 0.005575115326792002, + 0.005044606979936361, + 0.0034386373590677977, + 0.005608800332993269, + 0.006196664646267891, + 0.009907030500471592, + 0.008069057017564774, + 0.006239402107894421, + 0.004191518295556307, + 0.007830825634300709, + 0.007837655022740364, + 0.008851361460983753, + 0.010508889332413673, + 0.009841748513281345, + 0.014006644487380981, + 0.011473284102976322, + 0.005941747687757015, + 0.011328089982271194, + 0.00813989993184805, + 0.006735676899552345, + 0.008076583966612816, + 0.005523188039660454, + 0.00528739532455802, + 0.0063993358053267, + 0.008079878985881805, + 0.00845450721681118, + 0.006171753164380789, + 0.011005074717104435, + 0.01545894518494606, + 0.018488267436623573, + 0.027673592790961266, + 0.03537101671099663, + 0.02162710204720497, + 0.020321188494563103, + 0.014559502713382244, + 0.010410594753921032, + 0.009977469220757484, + 0.017877161502838135, + 0.06088334321975708, + 0.017820212990045547, + 0.027439380064606667, + 0.028882930055260658, + 0.02788088470697403, + 0.006989754270762205, + 0.005270103458315134, + 0.007339293137192726, + 0.008418933488428593, + 0.007611304987221956, + 0.010514226742088795, + 0.006708559580147266, + 0.008492572233080864, + 0.00445238221436739, + 0.0071408553048968315, + 0.005167810712009668, + 0.009090124629437923, + 0.005006997846066952, + 0.003334677778184414, + 0.005455296486616135, + 0.0055871205404400826, + 0.012869592756032944, + 0.0689590573310852, + 0.0400741845369339, + 0.008057981729507446, + 0.014274569228291512, + 0.00832930114120245, + 0.011177286505699158, + 0.005612315610051155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.01610778272151947 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097638186067343 + ], + "opposite_logprob_tokens": [ + -5.5040974617004395 + ], + "logprob_reward": -0.004097638186067343, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097638186067343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3538002669811249, + 0.13930508494377136, + 0.10665088146924973, + 0.12007950246334076, + 0.03296970576047897, + 0.028411047533154488, + 0.06972532719373703, + 0.05220397561788559, + 0.043317459523677826, + 0.05353673920035362 + ], + "credit_map": [ + 0.025307223200798035, + 0.04518523067235947, + 0.03344981372356415, + 0.08938287198543549, + 0.03895699232816696, + 0.05874158442020416, + 0.036082156002521515, + 0.026694370433688164, + 0.013950846157968044, + 0.013317779637873173, + 0.011014499701559544, + 0.022777138277888298, + 0.014714706689119339, + 0.027007540687918663, + 0.025390924885869026, + 0.011131646111607552, + 0.011645718477666378, + 0.02352800779044628, + 0.01359203364700079, + 0.01281816978007555, + 0.014209587126970291, + 0.009984918870031834, + 0.009608845226466656, + 0.011263600550591946, + 0.010647028684616089, + 0.02005876414477825, + 0.012667320668697357, + 0.02188120037317276, + 0.014629616402089596, + 0.010501572862267494, + 0.016193265095353127, + 0.013500746339559555, + 0.0030603650957345963, + 0.005329362582415342, + 0.00281574041582644, + 0.005653302650898695, + 0.003229987807571888, + 0.005662790033966303, + 0.004008450545370579, + 0.0032097024377435446, + 0.0019189392914995551, + 0.003133301390334964, + 0.0030118131544440985, + 0.004893380217254162, + 0.005024005193263292, + 0.003469331655651331, + 0.0038058257196098566, + 0.003154451958835125, + 0.004540354944765568, + 0.006063100881874561, + 0.00584377022460103, + 0.01469347346574068, + 0.008683886379003525, + 0.010568628087639809, + 0.009208421222865582, + 0.010123697109520435, + 0.0118793286383152, + 0.004700853023678064, + 0.006641735788434744, + 0.0055220345966517925, + 0.005498989950865507, + 0.005759878549724817, + 0.00560612790286541, + 0.006595023907721043, + 0.00847463496029377, + 0.007914318703114986, + 0.008803720586001873, + 0.004089649301022291, + 0.0034511673729866743, + 0.0020061407703906298, + 0.00408053956925869, + 0.00449728686362505, + 0.004632368218153715, + 0.01662498526275158, + 0.00452246842905879, + 0.003499056678265333, + 0.005919067654758692, + 0.0018330771708860993, + 0.012412111274898052, + 0.004093605559319258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.004097638186067343 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818566583096981 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.01818566583096981, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818566583096981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1997251808643341, + 0.04469268023967743, + 0.08151070773601532, + 0.17250403761863708, + 0.06949138641357422, + 0.06626331061124802, + 0.13922660052776337, + 0.09436904639005661, + 0.021485580131411552, + 0.11073143035173416 + ], + "credit_map": [ + 0.01721694879233837, + 0.024907056242227554, + 0.02184557542204857, + 0.061657629907131195, + 0.014875872991979122, + 0.030317772179841995, + 0.01805044896900654, + 0.010853874497115612, + 0.004268729593604803, + 0.005516820587217808, + 0.00394684262573719, + 0.0065710702911019325, + 0.005041110794991255, + 0.008343233726918697, + 0.006494085770100355, + 0.004510787781327963, + 0.007932783104479313, + 0.012534143403172493, + 0.010040102526545525, + 0.011252295225858688, + 0.011163497343659401, + 0.011531676165759563, + 0.007397504057735205, + 0.009658710099756718, + 0.01835835911333561, + 0.020065994933247566, + 0.02083459310233593, + 0.029244450852274895, + 0.020387548953294754, + 0.015392223373055458, + 0.02542167901992798, + 0.022799206897616386, + 0.006850465666502714, + 0.00831430871039629, + 0.006957074627280235, + 0.009305915795266628, + 0.00724747171625495, + 0.012373032048344612, + 0.009604905731976032, + 0.008838212117552757, + 0.004139556549489498, + 0.005940197501331568, + 0.005243169143795967, + 0.010228553786873817, + 0.018602991476655006, + 0.009362529031932354, + 0.007838413119316101, + 0.0049078986048698425, + 0.007867703214287758, + 0.008369382470846176, + 0.011906378902494907, + 0.03771643340587616, + 0.017505550757050514, + 0.018835945054888725, + 0.0196218341588974, + 0.01740337163209915, + 0.010458843782544136, + 0.007446282543241978, + 0.01256465818732977, + 0.01449256855994463, + 0.01064236182719469, + 0.014330419711768627, + 0.011130399070680141, + 0.013303513638675213, + 0.004613587632775307, + 0.0026273857802152634, + 0.003845709376037121, + 0.002272143727168441, + 0.0019882540218532085, + 0.0018182138446718454, + 0.002151260618120432, + 0.002169023035094142, + 0.0092118289321661, + 0.008190974593162537, + 0.03301999345421791, + 0.006836749613285065, + 0.011039487086236477, + 0.005803967826068401, + 0.026938294991850853, + 0.009690137580037117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.01818566583096981 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black panther sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the panther's fur glossy and black with a silky sheen reflecting dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dense tropical jungle with massive ferns, tangled vines, and moss-covered tree trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle and tracking parallel to the panther's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the panther's mouth slightly open and its long tail swinging behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420164667069912 + ], + "opposite_logprob_tokens": [ + -4.264201641082764 + ], + "logprob_reward": -0.01420164667069912, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420164667069912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08379165083169937, + 0.029311666265130043, + 0.06379231065511703, + 0.11151177436113358, + 0.1144939661026001, + 0.1312456727027893, + 0.03493707254528999, + 0.06895222514867783, + 0.0854324996471405, + 0.2765311002731323 + ], + "credit_map": [ + 0.012956241145730019, + 0.010445088148117065, + 0.011563271284103394, + 0.017329739406704903, + 0.007730912417173386, + 0.010126389563083649, + 0.008352751843631268, + 0.0052872588858008385, + 0.002482058247551322, + 0.002969972090795636, + 0.004065552726387978, + 0.004598112311214209, + 0.0024622695054858923, + 0.005784515291452408, + 0.004564077593386173, + 0.0023851091973483562, + 0.005146734416484833, + 0.007199070416390896, + 0.010531012900173664, + 0.009564210660755634, + 0.005982925184071064, + 0.007830693386495113, + 0.010277618654072285, + 0.007260046433657408, + 0.010382752865552902, + 0.014262348413467407, + 0.012816742062568665, + 0.029113758355379105, + 0.010687683708965778, + 0.008097550831735134, + 0.015928180888295174, + 0.010222749784588814, + 0.014745708554983139, + 0.012108785100281239, + 0.008393339812755585, + 0.01830885000526905, + 0.015401816926896572, + 0.013217191211879253, + 0.016505563631653786, + 0.015812711790204048, + 0.011325131170451641, + 0.011636811308562756, + 0.00963528547435999, + 0.023590775206685066, + 0.03213828429579735, + 0.01786080002784729, + 0.015062663704156876, + 0.009995926171541214, + 0.002921375213190913, + 0.0026996885426342487, + 0.003731263568624854, + 0.007133367471396923, + 0.004545398056507111, + 0.005898038856685162, + 0.004206969868391752, + 0.0038009712006896734, + 0.011139974929392338, + 0.005984609015285969, + 0.008167595602571964, + 0.009992471896111965, + 0.00795102771371603, + 0.007970625534653664, + 0.008863137103617191, + 0.008882780559360981, + 0.011478034779429436, + 0.011906908825039864, + 0.013694926165044308, + 0.009531649760901928, + 0.00977354682981968, + 0.006821122486144304, + 0.013063758611679077, + 0.009162552654743195, + 0.02412383072078228, + 0.024613559246063232, + 0.07235889881849289, + 0.022975094616413116, + 0.02361837774515152, + 0.024191731587052345, + 0.06094155088067055, + 0.023708075284957886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.01420164667069912 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_2/index_tensor([4481, 4481])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_2/index_tensor([4481, 4481])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0c5874baf1e8378d87117dacfe9d0580b3126aca --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_2/index_tensor([4481, 4481])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle macro shot offers an intimate glimpse into a subterranean ant chamber, its walls composed of rough, dense, dark earth and fine sand. At the center of this dimly lit, oval space lies a majestic queen ant, her body a smooth, deep mahogany. Her abdomen, or stomach, is disproportionately large and swollen, its semi-translucent sections straining and shimmering under a soft, off-camera light source that simulates the careful lighting of a documentary filmmaker. With slow, rhythmic contractions, she deliberately squeezes a tiny, pearlescent white egg onto the chamber's floor. Moist and rectangular, the newly laid egg joins a nearby carefully arranged nest, tended by a pair of smaller worker ants who act with swift, instinctive precision. The scene, dominated by rich, earthy browns and blacks, contrasts sharply with the bright, creamy white cluster of eggs, creating an atmosphere of concentration and almost reverence, capturing the quiet miracle of life unfolding in a hidden world.", + "seed": 16987, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_2/index_4481_seed_16987_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.052495206787716596, + "vlm_logprob": -0.052495206787716596, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.052495206787716596, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021839246153831482 + ], + "opposite_logprob_tokens": [ + -4.271839141845703 + ], + "logprob_reward": -0.021839246153831482, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.021839246153831482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26297226548194885, + 0.11616551876068115, + 0.29246747493743896, + 0.06077869236469269, + 0.039910271763801575, + 0.05226489529013634, + 0.027720335870981216, + 0.04865049570798874, + 0.06498304009437561, + 0.03408708795905113 + ], + "credit_map": [ + 0.03466036170721054, + 0.03529190272092819, + 0.023679321631789207, + 0.02464974857866764, + 0.046786416321992874, + 0.03803407773375511, + 0.02999013476073742, + 0.029880274087190628, + 0.018386108800768852, + 0.012120949104428291, + 0.01325596310198307, + 0.008439376018941402, + 0.017910975962877274, + 0.01734357886016369, + 0.010162790305912495, + 0.018545769155025482, + 0.03358984738588333, + 0.03171001747250557, + 0.03275841847062111, + 0.03128020092844963, + 0.0404994823038578, + 0.04057770594954491, + 0.048437103629112244, + 0.03361464664340019, + 0.008383572101593018, + 0.007223920430988073, + 0.008553065359592438, + 0.005887571256607771, + 0.007396769244223833, + 0.008619487285614014, + 0.004687008447945118, + 0.010027292184531689, + 0.005270023830235004, + 0.004594953265041113, + 0.005464585497975349, + 0.003415364306420088, + 0.0037862632889300585, + 0.007319275755435228, + 0.0024170661345124245, + 0.0076427338644862175, + 0.005637002177536488, + 0.006036697421222925, + 0.003623761935159564, + 0.0027160083409398794, + 0.005782457534223795, + 0.011056512594223022, + 0.004720097873359919, + 0.012692353688180447, + 0.006781150121241808, + 0.003577999072149396, + 0.002574377693235874, + 0.002134145237505436, + 0.0028891649562865496, + 0.004116833209991455, + 0.001998760038986802, + 0.003647900652140379, + 0.004092489834874868, + 0.005268375389277935, + 0.005800072569400072, + 0.004662207327783108, + 0.006923769135028124, + 0.007678605150431395, + 0.004541458562016487, + 0.009683514013886452, + 0.010196859948337078, + 0.006107266992330551, + 0.00546047929674387, + 0.0055848718620836735, + 0.008179351687431335, + 0.009939546696841717, + 0.005655577871948481, + 0.013859082013368607, + 0.00503197405487299, + 0.0038884752430021763, + 0.002869583433493972, + 0.003332151100039482, + 0.0037069981917738914, + 0.005344814155250788, + 0.002252422273159027, + 0.007660667411983013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.021839246153831482 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026381840929389 + ], + "opposite_logprob_tokens": [ + -3.651381731033325 + ], + "logprob_reward": -0.026381840929389, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026381840929389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2152327299118042, + 0.16689296066761017, + 0.15747134387493134, + 0.0913720577955246, + 0.05132133886218071, + 0.03679906204342842, + 0.07414856553077698, + 0.05879776179790497, + 0.112394779920578, + 0.03556937724351883 + ], + "credit_map": [ + 0.03076826222240925, + 0.020939363166689873, + 0.022185875102877617, + 0.026647021993994713, + 0.03146395459771156, + 0.028340771794319153, + 0.022963762283325195, + 0.031923726201057434, + 0.01933286339044571, + 0.014101664535701275, + 0.025715384632349014, + 0.018348142504692078, + 0.018110480159521103, + 0.02554011158645153, + 0.01758204959332943, + 0.028162281960248947, + 0.020672336220741272, + 0.015149562619626522, + 0.017664697021245956, + 0.01848788745701313, + 0.020280644297599792, + 0.022474581375718117, + 0.02000236324965954, + 0.02273927442729473, + 0.010528898797929287, + 0.0111607126891613, + 0.009300878271460533, + 0.008967380039393902, + 0.016221845522522926, + 0.013980936259031296, + 0.00889170914888382, + 0.01231969054788351, + 0.009363370947539806, + 0.005873790942132473, + 0.006005707196891308, + 0.0032405040692538023, + 0.005420455243438482, + 0.008029558695852757, + 0.0037245405837893486, + 0.009663410484790802, + 0.005327227991074324, + 0.0044619254767894745, + 0.0033489521592855453, + 0.0028573658782988787, + 0.005226725246757269, + 0.0062674665823578835, + 0.003044032957404852, + 0.006265364587306976, + 0.012325691059231758, + 0.00949331559240818, + 0.005146628711372614, + 0.005524171981960535, + 0.010462786071002483, + 0.01149782631546259, + 0.0056031132116913795, + 0.014095029793679714, + 0.008690121583640575, + 0.0051797726191580296, + 0.004435868002474308, + 0.0047021652571856976, + 0.005922107957303524, + 0.011505387723445892, + 0.004546686541289091, + 0.013815651647746563, + 0.010121917352080345, + 0.01048679742962122, + 0.012857988476753235, + 0.010454247705638409, + 0.015298913232982159, + 0.020502343773841858, + 0.008315818384289742, + 0.02435675635933876, + 0.006509590428322554, + 0.004013989120721817, + 0.004786038305610418, + 0.004014166537672281, + 0.003809821791946888, + 0.005567018873989582, + 0.0031131987925618887, + 0.0037555519957095385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.026381840929389 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11299356818199158 + ], + "opposite_logprob_tokens": [ + -2.2379934787750244 + ], + "logprob_reward": -0.11299356818199158, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11299356818199158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1275796890258789, + 0.10058808326721191, + 0.16089856624603271, + 0.07514232397079468, + 0.09366055577993393, + 0.16952601075172424, + 0.05343158543109894, + 0.07730596512556076, + 0.09472896158695221, + 0.0471382662653923 + ], + "credit_map": [ + 0.01761399768292904, + 0.012313694693148136, + 0.013949701562523842, + 0.010434241034090519, + 0.017229953780770302, + 0.019446179270744324, + 0.012848805636167526, + 0.023743117228150368, + 0.010162636637687683, + 0.00943598523736, + 0.01386007759720087, + 0.01156652346253395, + 0.012495252303779125, + 0.015965670347213745, + 0.006264669355005026, + 0.02083727903664112, + 0.01914159208536148, + 0.016514746472239494, + 0.017009709030389786, + 0.01974995993077755, + 0.022572187706828117, + 0.021951980888843536, + 0.02548344060778618, + 0.01847495697438717, + 0.01030968502163887, + 0.009343422017991543, + 0.010136888362467289, + 0.0074425023049116135, + 0.011858335696160793, + 0.010728121735155582, + 0.005640782415866852, + 0.00968258548527956, + 0.011619588360190392, + 0.010976159013807774, + 0.010126393288373947, + 0.0063523962162435055, + 0.010903121903538704, + 0.01743961311876774, + 0.007049843203276396, + 0.019193438813090324, + 0.016740014776587486, + 0.017850469797849655, + 0.01255114283412695, + 0.00925417710095644, + 0.018779223784804344, + 0.03792000189423561, + 0.013758134096860886, + 0.04267283156514168, + 0.01025304850190878, + 0.00500982953235507, + 0.004817891400307417, + 0.004075017757713795, + 0.006608815863728523, + 0.0092490054666996, + 0.0049660298973321915, + 0.008451942354440689, + 0.00966425146907568, + 0.006902209483087063, + 0.006054505705833435, + 0.0035855737514793873, + 0.007287504617124796, + 0.016756806522607803, + 0.003622620366513729, + 0.023432496935129166, + 0.014372571371495724, + 0.007692386396229267, + 0.008638603612780571, + 0.006276882253587246, + 0.010746202431619167, + 0.017360147088766098, + 0.00713040865957737, + 0.02251175418496132, + 0.007883047685027122, + 0.004575875587761402, + 0.004690011963248253, + 0.005351309664547443, + 0.006330939009785652, + 0.007126967888325453, + 0.004747902974486351, + 0.006432215217500925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.11299356818199158 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033328831195831 + ], + "opposite_logprob_tokens": [ + -2.3503332138061523 + ], + "logprob_reward": -0.10033328831195831, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033328831195831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20344458520412445, + 0.1144976019859314, + 0.33683714270591736, + 0.04544232785701752, + 0.05517001822590828, + 0.08059968799352646, + 0.028673909604549408, + 0.05997858941555023, + 0.03660157322883606, + 0.038754578679800034 + ], + "credit_map": [ + 0.02396107092499733, + 0.021107004955410957, + 0.024490561336278915, + 0.019146114587783813, + 0.026691056787967682, + 0.029966257512569427, + 0.016087519004940987, + 0.04199500009417534, + 0.01254901010543108, + 0.01119177509099245, + 0.019655290991067886, + 0.009142099879682064, + 0.010831986553966999, + 0.018767476081848145, + 0.007740155328065157, + 0.024619804695248604, + 0.037593089044094086, + 0.03713533654808998, + 0.03754295036196709, + 0.0359664224088192, + 0.04674901068210602, + 0.047627419233322144, + 0.05032479017972946, + 0.04389812797307968, + 0.005287699401378632, + 0.004618500825017691, + 0.004962666425853968, + 0.0045452360063791275, + 0.006520369555801153, + 0.0077975718304514885, + 0.004095620941370726, + 0.00761466333642602, + 0.007288055028766394, + 0.007491768337786198, + 0.004686090629547834, + 0.007684174459427595, + 0.006249524652957916, + 0.008139419369399548, + 0.004744206555187702, + 0.008886775933206081, + 0.008754666894674301, + 0.008184321224689484, + 0.0050934432074427605, + 0.004527112003415823, + 0.010746064595878124, + 0.017490733414888382, + 0.007289332337677479, + 0.01851401850581169, + 0.004773491527885199, + 0.003067235229536891, + 0.002555568004027009, + 0.0018700598739087582, + 0.003528106724843383, + 0.004738139919936657, + 0.002438980620354414, + 0.005702325142920017, + 0.0096054058521986, + 0.0051965126767754555, + 0.004441533703356981, + 0.003037861781194806, + 0.005161916837096214, + 0.01206064410507679, + 0.0024174973368644714, + 0.018057215958833694, + 0.00605083117261529, + 0.003783403430134058, + 0.0036203907802700996, + 0.003627801313996315, + 0.005171562545001507, + 0.0049880086444318295, + 0.003836945164948702, + 0.005522631108760834, + 0.006233504973351955, + 0.0038080948870629072, + 0.004825163166970015, + 0.003326910315081477, + 0.004300483502447605, + 0.006352962926030159, + 0.003029886167496443, + 0.006877572275698185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.10033328831195831 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009280903614126146 + ], + "opposite_logprob_tokens": [ + -7.000927925109863 + ], + "logprob_reward": -0.0009280903614126146, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009280903614126146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24848884344100952, + 0.18192040920257568, + 0.07499951124191284, + 0.07862243056297302, + 0.09732676297426224, + 0.06884888559579849, + 0.05234140530228615, + 0.05162506923079491, + 0.0771963819861412, + 0.06863034516572952 + ], + "credit_map": [ + 0.030575338751077652, + 0.02942616492509842, + 0.026700779795646667, + 0.023653140291571617, + 0.03634824976325035, + 0.036556426435709, + 0.03773213177919388, + 0.027496609836816788, + 0.027098776772618294, + 0.0211888886988163, + 0.02123260870575905, + 0.015763921663165092, + 0.028783069923520088, + 0.023235904052853584, + 0.019105829298496246, + 0.025511395186185837, + 0.011791189201176167, + 0.008432206697762012, + 0.008856098167598248, + 0.008994437754154205, + 0.00875120609998703, + 0.009746911004185677, + 0.00954892486333847, + 0.008878536522388458, + 0.012751861475408077, + 0.010395399294793606, + 0.007147202733904123, + 0.004790788982063532, + 0.014561648480594158, + 0.011831839568912983, + 0.007470789831131697, + 0.009672899730503559, + 0.012660488486289978, + 0.010747257620096207, + 0.0099369827657938, + 0.009366339072585106, + 0.009485469199717045, + 0.017432216554880142, + 0.008838275447487831, + 0.018859729170799255, + 0.0082640266045928, + 0.009842108935117722, + 0.006010265089571476, + 0.005383648443967104, + 0.008005076088011265, + 0.011273565702140331, + 0.0043275728821754456, + 0.01574261672794819, + 0.010139066725969315, + 0.006867839023470879, + 0.004664536565542221, + 0.003464174224063754, + 0.006225652527064085, + 0.008833999745547771, + 0.0039154826663434505, + 0.008230648934841156, + 0.010302594862878323, + 0.00446174805983901, + 0.005582808516919613, + 0.0034036519937217236, + 0.004792691208422184, + 0.009109120815992355, + 0.0024166577495634556, + 0.01155579462647438, + 0.012718412093818188, + 0.007564162369817495, + 0.007546846754848957, + 0.006985390093177557, + 0.010423733852803707, + 0.01073408778756857, + 0.00850241631269455, + 0.012721323408186436, + 0.008653471246361732, + 0.006615314166992903, + 0.005050250329077244, + 0.0049341944977641106, + 0.006521941628307104, + 0.012750358320772648, + 0.0034255001228302717, + 0.020679308101534843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0009280903614126146 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle macro shot offers an intimate glimpse into a subterranean ant chamber, its walls composed of rough, dense, dark earth and fine sand. At the center of this dimly lit, oval space lies a majestic queen ant, her body a smooth, deep mahogany. Her abdomen, or stomach, is disproportionately large and swollen, its semi-translucent sections straining and shimmering under a soft, off-camera light source that simulates the careful lighting of a documentary filmmaker. With slow, rhythmic contractions, she deliberately squeezes a tiny, pearlescent white egg onto the chamber's floor. Moist and rectangular, the newly laid egg joins a nearby carefully arranged nest, tended by a pair of smaller worker ants who act with swift, instinctive precision. The scene, dominated by rich, earthy browns and blacks, contrasts sharply with the bright, creamy white cluster of eggs, creating an atmosphere of concentration and almost reverence, capturing the quiet miracle of life unfolding in a hidden world.", + "seed": 16988, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_2/index_4481_seed_16988_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.028426772000966594, + "vlm_logprob": -0.028426772000966594, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.028426772000966594, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006928111892193556 + ], + "opposite_logprob_tokens": [ + -5.881927967071533 + ], + "logprob_reward": -0.006928111892193556, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.006928111892193556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2844040095806122, + 0.12140704691410065, + 0.10513165593147278, + 0.08256640285253525, + 0.0750688686966896, + 0.095271036028862, + 0.07207470387220383, + 0.050001662224531174, + 0.05862520635128021, + 0.05544937029480934 + ], + "credit_map": [ + 0.030966278165578842, + 0.03218366205692291, + 0.03544046729803085, + 0.03161412477493286, + 0.037186287343502045, + 0.04832363501191139, + 0.03312142565846443, + 0.035568125545978546, + 0.01338806003332138, + 0.014542979188263416, + 0.012117148377001286, + 0.013929292559623718, + 0.022870387881994247, + 0.016075637191534042, + 0.01524467021226883, + 0.013238872401416302, + 0.01100916601717472, + 0.010714413598179817, + 0.012509468011558056, + 0.011850935406982899, + 0.01599598117172718, + 0.020382840186357498, + 0.009043623693287373, + 0.013625229708850384, + 0.009910558350384235, + 0.007675426080822945, + 0.010900610126554966, + 0.007918030954897404, + 0.010296358726918697, + 0.01451188512146473, + 0.007540817838162184, + 0.013812719844281673, + 0.005790119059383869, + 0.0043455311097204685, + 0.009637128561735153, + 0.004586210940033197, + 0.005540131591260433, + 0.015617499127984047, + 0.004938205704092979, + 0.02461404725909233, + 0.008969495072960854, + 0.009478054009377956, + 0.010155556723475456, + 0.009694932028651237, + 0.012219855561852455, + 0.01723090559244156, + 0.010838109068572521, + 0.016684122383594513, + 0.007403089199215174, + 0.007562675513327122, + 0.010049883276224136, + 0.008426295593380928, + 0.008369608782231808, + 0.010633585043251514, + 0.007311846595257521, + 0.012317719869315624, + 0.007228879723697901, + 0.0050542522221803665, + 0.007912786677479744, + 0.00587169174104929, + 0.004958202131092548, + 0.005863277241587639, + 0.00618315115571022, + 0.006929425057023764, + 0.008052464574575424, + 0.005753629375249147, + 0.007997274398803711, + 0.0070283315144479275, + 0.006831541657447815, + 0.008245185017585754, + 0.006625468377023935, + 0.008091311901807785, + 0.007949327118694782, + 0.0056528267450630665, + 0.008797179907560349, + 0.0062298947013914585, + 0.005827986169606447, + 0.00884147360920906, + 0.0054985759779810905, + 0.006652108859270811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.006928111892193556 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009806787595152855 + ], + "opposite_logprob_tokens": [ + -4.6348066329956055 + ], + "logprob_reward": -0.009806787595152855, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009806787595152855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18818999826908112, + 0.20485299825668335, + 0.25066256523132324, + 0.09957083314657211, + 0.05337394401431084, + 0.02918379195034504, + 0.02664545737206936, + 0.024776801466941833, + 0.04992993548512459, + 0.07281370460987091 + ], + "credit_map": [ + 0.019482413306832314, + 0.019800739362835884, + 0.025207161903381348, + 0.02490905672311783, + 0.023025620728731155, + 0.029152987524867058, + 0.024002650752663612, + 0.022609369829297066, + 0.017718594521284103, + 0.026700712740421295, + 0.024251017719507217, + 0.02456066571176052, + 0.029466962441802025, + 0.031378600746393204, + 0.023594899103045464, + 0.02718154340982437, + 0.023622168228030205, + 0.01813557557761669, + 0.026581183075904846, + 0.01586158573627472, + 0.048013776540756226, + 0.051461346447467804, + 0.01636669598519802, + 0.050620220601558685, + 0.01189462747424841, + 0.009224092587828636, + 0.015899233520030975, + 0.010571707971394062, + 0.011779067106544971, + 0.018302856013178825, + 0.007891619578003883, + 0.014007629826664925, + 0.0049382965080440044, + 0.005332385655492544, + 0.00568585516884923, + 0.005539545323699713, + 0.004265013616532087, + 0.011290700174868107, + 0.003974444232881069, + 0.012347700074315071, + 0.003131390083581209, + 0.0031638334039598703, + 0.003937062341719866, + 0.003399616340175271, + 0.002565867267549038, + 0.005515534430742264, + 0.0018672848818823695, + 0.0056032040156424046, + 0.003052254905924201, + 0.004177927505224943, + 0.003578629344701767, + 0.003169036004692316, + 0.003187797497957945, + 0.0033045930322259665, + 0.002614570315927267, + 0.0035606480669230223, + 0.003543718019500375, + 0.002424849197268486, + 0.0034316268283873796, + 0.0027048850897699594, + 0.0024676022585481405, + 0.004596831277012825, + 0.0022741947323083878, + 0.003333092899993062, + 0.0047618974931538105, + 0.004646722227334976, + 0.007302316837012768, + 0.005278625059872866, + 0.005346665158867836, + 0.010012133046984673, + 0.003951156046241522, + 0.008630417287349701, + 0.00975580420345068, + 0.006148539949208498, + 0.011619152501225471, + 0.008525071665644646, + 0.009647578932344913, + 0.011939296498894691, + 0.0069535584188997746, + 0.008224702440202236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.009806787595152855 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004882395267487 + ], + "opposite_logprob_tokens": [ + -2.6950488090515137 + ], + "logprob_reward": -0.07004882395267487, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004882395267487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12712693214416504, + 0.25572726130485535, + 0.1860073357820511, + 0.138079434633255, + 0.05494949221611023, + 0.03906065970659256, + 0.022722244262695312, + 0.03307725861668587, + 0.04354843124747276, + 0.09970088303089142 + ], + "credit_map": [ + 0.014347930438816547, + 0.0135347256436944, + 0.0204375721514225, + 0.013582129031419754, + 0.017082439735531807, + 0.02114819549024105, + 0.013789569027721882, + 0.013204379007220268, + 0.027329204604029655, + 0.024032866582274437, + 0.028724269941449165, + 0.027019290253520012, + 0.02476714551448822, + 0.043219536542892456, + 0.02979327365756035, + 0.05084166303277016, + 0.01887904293835163, + 0.015013040974736214, + 0.025260211899876595, + 0.014333778992295265, + 0.027786999940872192, + 0.031583067029714584, + 0.013271315023303032, + 0.03987988084554672, + 0.014690759591758251, + 0.009680744260549545, + 0.021081559360027313, + 0.010771672241389751, + 0.013038575649261475, + 0.025234952569007874, + 0.011802381835877895, + 0.03177879378199577, + 0.005734675098210573, + 0.007643736433237791, + 0.007056569214910269, + 0.007030653767287731, + 0.006069516763091087, + 0.007219322957098484, + 0.004673233721405268, + 0.0095217851921916, + 0.005540957674384117, + 0.003359378082677722, + 0.003645353950560093, + 0.004410406108945608, + 0.004101618193089962, + 0.007895790971815586, + 0.00359744462184608, + 0.006509712431579828, + 0.0028662197291851044, + 0.0027538451831787825, + 0.00241957139223814, + 0.003088108729571104, + 0.0030070957727730274, + 0.0031544696539640427, + 0.0026610882487148046, + 0.0027718443889170885, + 0.0057930066250264645, + 0.0025321277789771557, + 0.004699941724538803, + 0.0034452355466783047, + 0.004011410288512707, + 0.0048192348331213, + 0.0038801594637334347, + 0.0038961402606219053, + 0.005704501643776894, + 0.004645833745598793, + 0.005842296406626701, + 0.005320598836988211, + 0.0051357033662498, + 0.007355359848588705, + 0.004586430266499519, + 0.004957704339176416, + 0.01143727172166109, + 0.010019415989518166, + 0.013241918757557869, + 0.01241905614733696, + 0.01040312647819519, + 0.016261925920844078, + 0.007795874960720539, + 0.01812228560447693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.07004882395267487 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495598912239075 + ], + "opposite_logprob_tokens": [ + -2.9299559593200684 + ], + "logprob_reward": -0.05495598912239075, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495598912239075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1601806879043579, + 0.11473111808300018, + 0.14751233160495758, + 0.09101696312427521, + 0.32939624786376953, + 0.05406751483678818, + 0.01983514241874218, + 0.014778180047869682, + 0.019048983231186867, + 0.04943278059363365 + ], + "credit_map": [ + 0.021578963845968246, + 0.014548276551067829, + 0.02150927670300007, + 0.01585317961871624, + 0.022204240784049034, + 0.024572299793362617, + 0.0191350057721138, + 0.020779462531208992, + 0.01187288947403431, + 0.012661145068705082, + 0.01163917314261198, + 0.009060461074113846, + 0.018703637644648552, + 0.017409376800060272, + 0.016082344576716423, + 0.017302095890045166, + 0.013993913307785988, + 0.014569994993507862, + 0.01714356802403927, + 0.017038876190781593, + 0.026029985398054123, + 0.02242610789835453, + 0.01254009548574686, + 0.023769784718751907, + 0.007809935603290796, + 0.011322195641696453, + 0.012024949304759502, + 0.011444645002484322, + 0.011012461967766285, + 0.014444062486290932, + 0.008725518360733986, + 0.014233186841011047, + 0.023479070514440536, + 0.030842287465929985, + 0.03188258036971092, + 0.023710554465651512, + 0.01975095272064209, + 0.07404235750436783, + 0.01910635642707348, + 0.10658209770917892, + 0.007962089963257313, + 0.0071817017160356045, + 0.005649769678711891, + 0.006092604715377092, + 0.005794778000563383, + 0.007088383659720421, + 0.005226712208241224, + 0.009071475826203823, + 0.0021827791351825, + 0.002182111842557788, + 0.002169139450415969, + 0.0029513246845453978, + 0.0027803718112409115, + 0.002950097667053342, + 0.001949487254023552, + 0.00266983057372272, + 0.003519258461892605, + 0.0013467329554259777, + 0.0017015428747981787, + 0.001962499925866723, + 0.0012475716648623347, + 0.0022762527223676443, + 0.001288524130359292, + 0.0014357981272041798, + 0.002463515382260084, + 0.001840816461481154, + 0.002667501103132963, + 0.0023366783279925585, + 0.002515949308872223, + 0.0026266940403729677, + 0.0021426272578537464, + 0.0024552028626203537, + 0.0061592841520905495, + 0.004393319133669138, + 0.006098788231611252, + 0.005676506087183952, + 0.006018873769789934, + 0.007248393259942532, + 0.0035463417880237103, + 0.010291271843016148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.05495598912239075 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00039414744242094457 + ], + "opposite_logprob_tokens": [ + -7.875394344329834 + ], + "logprob_reward": -0.00039414744242094457, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.00039414744242094457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2702372074127197, + 0.23464268445968628, + 0.1406726837158203, + 0.08517110347747803, + 0.07738181948661804, + 0.014613326638936996, + 0.019282188266515732, + 0.03858373314142227, + 0.03443761542439461, + 0.0849776417016983 + ], + "credit_map": [ + 0.026597747579216957, + 0.02728475071489811, + 0.037186987698078156, + 0.03446141257882118, + 0.031027402728796005, + 0.04291475564241409, + 0.033408451825380325, + 0.037355709820985794, + 0.024555234238505363, + 0.021325230598449707, + 0.02674691006541252, + 0.022659001871943474, + 0.02929464727640152, + 0.03774859011173248, + 0.02951718680560589, + 0.042795877903699875, + 0.013607446104288101, + 0.011751367710530758, + 0.018483558669686317, + 0.012118755839765072, + 0.022548945620656013, + 0.025581708177924156, + 0.010936285369098186, + 0.02564462088048458, + 0.009902535006403923, + 0.007194621954113245, + 0.012807556428015232, + 0.007108096498996019, + 0.00905328243970871, + 0.014373869635164738, + 0.007923818193376064, + 0.016807325184345245, + 0.005907958373427391, + 0.008182198740541935, + 0.008193109184503555, + 0.006446246989071369, + 0.006072936113923788, + 0.015610802918672562, + 0.005393180530518293, + 0.021575389429926872, + 0.0020510193426162004, + 0.0015346823493018746, + 0.0015176705783233047, + 0.0018982519395649433, + 0.001998449442908168, + 0.0023760716430842876, + 0.0013615261996164918, + 0.0018756554927676916, + 0.002758558839559555, + 0.0018894568784162402, + 0.0026778429746627808, + 0.002256121253594756, + 0.0025322234723716974, + 0.0026445151306688786, + 0.002066149841994047, + 0.0024573204573243856, + 0.005804093088954687, + 0.003579421667382121, + 0.005257562734186649, + 0.005128807388246059, + 0.004628442227840424, + 0.004922126419842243, + 0.004728937055915594, + 0.0045343427918851376, + 0.0036187388468533754, + 0.0029473179019987583, + 0.004182251635938883, + 0.0034647039137780666, + 0.003674222156405449, + 0.007506144233047962, + 0.0025479369796812534, + 0.006496299058198929, + 0.012435336597263813, + 0.010142291896045208, + 0.00979945994913578, + 0.00831523071974516, + 0.008473588153719902, + 0.012086776085197926, + 0.006668753456324339, + 0.017056209966540337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.00039414744242094457 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle macro shot offers an intimate glimpse into a subterranean ant chamber, its walls composed of rough, dense, dark earth and fine sand. At the center of this dimly lit, oval space lies a majestic queen ant, her body a smooth, deep mahogany. Her abdomen, or stomach, is disproportionately large and swollen, its semi-translucent sections straining and shimmering under a soft, off-camera light source that simulates the careful lighting of a documentary filmmaker. With slow, rhythmic contractions, she deliberately squeezes a tiny, pearlescent white egg onto the chamber's floor. Moist and rectangular, the newly laid egg joins a nearby carefully arranged nest, tended by a pair of smaller worker ants who act with swift, instinctive precision. The scene, dominated by rich, earthy browns and blacks, contrasts sharply with the bright, creamy white cluster of eggs, creating an atmosphere of concentration and almost reverence, capturing the quiet miracle of life unfolding in a hidden world.", + "seed": 16989, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_2/index_4481_seed_16989_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04286802467540838, + "vlm_logprob": -0.04286802467540838, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.04286802467540838, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018436262384057045 + ], + "opposite_logprob_tokens": [ + -4.768436431884766 + ], + "logprob_reward": -0.018436262384057045, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.018436262384057045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1658230721950531, + 0.08452171832323074, + 0.08089133352041245, + 0.04542195796966553, + 0.08970985561609268, + 0.10012352466583252, + 0.21713510155677795, + 0.1378338783979416, + 0.04494558274745941, + 0.03359391540288925 + ], + "credit_map": [ + 0.023842591792345047, + 0.011866480112075806, + 0.017052439972758293, + 0.030268412083387375, + 0.014269361272454262, + 0.029392940923571587, + 0.02250269241631031, + 0.01662815734744072, + 0.014270471408963203, + 0.0068288277834653854, + 0.009843269363045692, + 0.011148217134177685, + 0.008532334119081497, + 0.013645514845848083, + 0.007675243075937033, + 0.012577839195728302, + 0.013490032404661179, + 0.0064854128286242485, + 0.008862716145813465, + 0.013465394265949726, + 0.0068378453142941, + 0.007876924239099026, + 0.009276860393583775, + 0.014596146531403065, + 0.007357813883572817, + 0.003518927376717329, + 0.005527188535779715, + 0.005283480975776911, + 0.003989522811025381, + 0.007478801999241114, + 0.004531916230916977, + 0.007734305690973997, + 0.01952582597732544, + 0.0070272586308419704, + 0.012269780039787292, + 0.008291498757898808, + 0.009942321106791496, + 0.014796067029237747, + 0.006391496397554874, + 0.011465608142316341, + 0.012501182034611702, + 0.008974539116024971, + 0.011891263537108898, + 0.010166171938180923, + 0.014567028731107712, + 0.018948841840028763, + 0.010528288781642914, + 0.012546214275062084, + 0.05158299580216408, + 0.018772751092910767, + 0.02913522720336914, + 0.019022779539227486, + 0.01884573884308338, + 0.028674839064478874, + 0.01547547709196806, + 0.03562530130147934, + 0.020693868398666382, + 0.013327248394489288, + 0.021397344768047333, + 0.012461697682738304, + 0.012950215488672256, + 0.02984100952744484, + 0.01235214900225401, + 0.014810348860919476, + 0.011712739244103432, + 0.002194549422711134, + 0.005894776899367571, + 0.002740546828135848, + 0.0023074294440448284, + 0.0071543497033417225, + 0.0037233743350952864, + 0.009217814542353153, + 0.00480630761012435, + 0.0023832025472074747, + 0.004518283065408468, + 0.004356806632131338, + 0.0024775145575404167, + 0.005247829481959343, + 0.005458097904920578, + 0.004345873836427927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.018436262384057045 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016099337488412857 + ], + "opposite_logprob_tokens": [ + -4.141099452972412 + ], + "logprob_reward": -0.016099337488412857, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016099337488412857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27480387687683105, + 0.09391319006681442, + 0.12540163099765778, + 0.04126423969864845, + 0.1227542832493782, + 0.06389052420854568, + 0.08456415683031082, + 0.09044481068849564, + 0.03868618607521057, + 0.06427714973688126 + ], + "credit_map": [ + 0.05461891368031502, + 0.020482392981648445, + 0.03187523037195206, + 0.028542039915919304, + 0.02634640596807003, + 0.03461811691522598, + 0.026771239936351776, + 0.0515495240688324, + 0.014304948039352894, + 0.009453405626118183, + 0.011072135530412197, + 0.01224080752581358, + 0.009720353409647942, + 0.014954882673919201, + 0.012477251701056957, + 0.009689400903880596, + 0.02487357333302498, + 0.009925846941769123, + 0.016156595200300217, + 0.01674112118780613, + 0.01023456733673811, + 0.012583077885210514, + 0.013306914828717709, + 0.021579932421445847, + 0.006075166631489992, + 0.004878293722867966, + 0.004834908526390791, + 0.004721013363450766, + 0.00592253590002656, + 0.004682187456637621, + 0.002971994923427701, + 0.0071781394071877, + 0.021254755556583405, + 0.011952404864132404, + 0.014458855614066124, + 0.015760358422994614, + 0.015448272228240967, + 0.02160612680017948, + 0.009634066373109818, + 0.012639438733458519, + 0.01031502801924944, + 0.0059374794363975525, + 0.007412218023091555, + 0.008472937159240246, + 0.007135797757655382, + 0.007896051742136478, + 0.00786960031837225, + 0.008851412683725357, + 0.017887191846966743, + 0.00914063211530447, + 0.010514259338378906, + 0.008261624723672867, + 0.008940763771533966, + 0.011239198967814445, + 0.006632757838815451, + 0.011947735212743282, + 0.013484184630215168, + 0.008748641237616539, + 0.01426334772258997, + 0.009553574025630951, + 0.010222377255558968, + 0.01723651960492134, + 0.008490068838000298, + 0.008446098305284977, + 0.009916857816278934, + 0.0019823117181658745, + 0.005629211664199829, + 0.0025914425496011972, + 0.0022561720106750727, + 0.006098920479416847, + 0.0024951014202088118, + 0.007716169115155935, + 0.007528703194111586, + 0.005364622455090284, + 0.007900025695562363, + 0.007716312538832426, + 0.005199509672820568, + 0.009521659463644028, + 0.013840524479746819, + 0.007205794099718332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.016099337488412857 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.100359708070755 + ], + "opposite_logprob_tokens": [ + -2.3503596782684326 + ], + "logprob_reward": -0.100359708070755, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.100359708070755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26373645663261414, + 0.11274627596139908, + 0.03267895430326462, + 0.0663711354136467, + 0.12811279296875, + 0.1414012461900711, + 0.04237727075815201, + 0.08695822209119797, + 0.05211177468299866, + 0.07350587099790573 + ], + "credit_map": [ + 0.044010698795318604, + 0.0206331554800272, + 0.030754514038562775, + 0.025686869397759438, + 0.0260306429117918, + 0.051460400223731995, + 0.03015724942088127, + 0.03500291705131531, + 0.018630007281899452, + 0.010696511715650558, + 0.012818804010748863, + 0.012933382764458656, + 0.012254754081368446, + 0.020236659795045853, + 0.011692523956298828, + 0.01348363608121872, + 0.006258925888687372, + 0.002681958256289363, + 0.004301498644053936, + 0.0036533435340970755, + 0.0025786678306758404, + 0.003327776910737157, + 0.0033706575632095337, + 0.006506125442683697, + 0.014144416898488998, + 0.006111376453191042, + 0.008595352992415428, + 0.007037285715341568, + 0.006255705840885639, + 0.010310602374374866, + 0.0054695080034434795, + 0.008446886204183102, + 0.020341698080301285, + 0.011231848038733006, + 0.015102251432836056, + 0.016264259815216064, + 0.015881258994340897, + 0.024852974340319633, + 0.010195261798799038, + 0.014243233017623425, + 0.023724276572465897, + 0.013601661659777164, + 0.017234470695257187, + 0.014327076263725758, + 0.022323135286569595, + 0.021257784217596054, + 0.013567378744482994, + 0.015365452505648136, + 0.006518341600894928, + 0.004340894985944033, + 0.006159247364848852, + 0.003865547012537718, + 0.003483362030237913, + 0.009211636148393154, + 0.0035577365197241306, + 0.0052405050955712795, + 0.015943020582199097, + 0.008087602443993092, + 0.012844246812164783, + 0.009545791894197464, + 0.009290765970945358, + 0.015358989126980305, + 0.0074206767603755, + 0.008467125706374645, + 0.010455001145601273, + 0.003266511019319296, + 0.007882136851549149, + 0.00453791581094265, + 0.004226531833410263, + 0.009122458286583424, + 0.004772253800183535, + 0.007848967798054218, + 0.014317804016172886, + 0.0059666200540959835, + 0.008325164206326008, + 0.006566356867551804, + 0.005601770710200071, + 0.012983066961169243, + 0.010944877751171589, + 0.0088002048432827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.100359708070755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899847626686096 + ], + "opposite_logprob_tokens": [ + -2.578998565673828 + ], + "logprob_reward": -0.07899847626686096, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899847626686096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.305177241563797, + 0.07165883481502533, + 0.06267877668142319, + 0.05078594759106636, + 0.07301084697246552, + 0.062160905450582504, + 0.1025431826710701, + 0.10109304636716843, + 0.041553981602191925, + 0.12933720648288727 + ], + "credit_map": [ + 0.04884069785475731, + 0.03077620081603527, + 0.033950094133615494, + 0.04510845988988876, + 0.03405587375164032, + 0.04572637379169464, + 0.03312714025378227, + 0.03359238803386688, + 0.013889682479202747, + 0.005750435404479504, + 0.007759415544569492, + 0.00814451277256012, + 0.0067297592759132385, + 0.011908487416803837, + 0.007051859516650438, + 0.010424674488604069, + 0.016017407178878784, + 0.0030931022483855486, + 0.010259765200316906, + 0.003994741011410952, + 0.0033336393535137177, + 0.005593451671302319, + 0.005044235847890377, + 0.01534243579953909, + 0.009137650020420551, + 0.004852185491472483, + 0.006258930079638958, + 0.0068711028434336185, + 0.0061315200291574, + 0.006437916774302721, + 0.005632651969790459, + 0.0054639894515275955, + 0.011605927720665932, + 0.007741228211671114, + 0.01041389163583517, + 0.012410604394972324, + 0.0073861232958734035, + 0.007111073471605778, + 0.0075519634410738945, + 0.008790036663413048, + 0.01236881036311388, + 0.005800654646009207, + 0.0083820391446352, + 0.007974964566528797, + 0.006759988609701395, + 0.007030182052403688, + 0.00688932603225112, + 0.006954938173294067, + 0.024490246549248695, + 0.00880433525890112, + 0.01339888758957386, + 0.008898862637579441, + 0.009228943847119808, + 0.013691278174519539, + 0.007578758057206869, + 0.016451874747872353, + 0.017424728721380234, + 0.011116677895188332, + 0.015797678381204605, + 0.009125183336436749, + 0.01010911539196968, + 0.0166347436606884, + 0.008914029225707054, + 0.011970890685915947, + 0.006194885820150375, + 0.0033043725416064262, + 0.004149067681282759, + 0.004873621743172407, + 0.0046007754281163216, + 0.007110161241143942, + 0.007805491331964731, + 0.0035156060475856066, + 0.018499033525586128, + 0.013149409554898739, + 0.01329636201262474, + 0.01163669303059578, + 0.011950027197599411, + 0.0280313640832901, + 0.014432279393076897, + 0.018342049792408943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.07899847626686096 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004463391669560224 + ], + "opposite_logprob_tokens": [ + -7.750446319580078 + ], + "logprob_reward": -0.0004463391669560224, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004463391669560224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17022965848445892, + 0.09384846687316895, + 0.12940235435962677, + 0.1026669293642044, + 0.09549777209758759, + 0.13255231082439423, + 0.1317577213048935, + 0.058892350643873215, + 0.03604285418987274, + 0.049109600484371185 + ], + "credit_map": [ + 0.029703589156270027, + 0.013092298060655594, + 0.021816110238432884, + 0.018744854256510735, + 0.014687628485262394, + 0.027262303978204727, + 0.020539985969662666, + 0.02438286691904068, + 0.01911110244691372, + 0.0071532391011714935, + 0.01158196572214365, + 0.010126507841050625, + 0.008425441570580006, + 0.01541579607874155, + 0.007547263987362385, + 0.014487147331237793, + 0.020170817151665688, + 0.011939975433051586, + 0.015492583625018597, + 0.022364994511008263, + 0.012623284943401814, + 0.014173552393913269, + 0.016471432521939278, + 0.016165703535079956, + 0.02851399965584278, + 0.005194489844143391, + 0.016081491485238075, + 0.007264701649546623, + 0.004921712912619114, + 0.01360415481030941, + 0.0057786875404417515, + 0.021307680755853653, + 0.018451746553182602, + 0.008956849575042725, + 0.011659768410027027, + 0.011757252737879753, + 0.012475362047553062, + 0.014354909770190716, + 0.0071607837453484535, + 0.0106810899451375, + 0.018678218126296997, + 0.013195326551795006, + 0.014916131272912025, + 0.014836888760328293, + 0.02145696058869362, + 0.01956956647336483, + 0.014121788553893566, + 0.015777423977851868, + 0.029223846271634102, + 0.012934587895870209, + 0.018006153404712677, + 0.011669185943901539, + 0.012829836457967758, + 0.017903927713632584, + 0.00926050916314125, + 0.019929666072130203, + 0.009687027893960476, + 0.005369705613702536, + 0.009349868632853031, + 0.005643763579428196, + 0.005084547214210033, + 0.011537684127688408, + 0.0057266224175691605, + 0.006493126507848501, + 0.008989730849862099, + 0.002403913764283061, + 0.004615018609911203, + 0.002603312721475959, + 0.0022797684650868177, + 0.005827728658914566, + 0.002573177218437195, + 0.006750200875103474, + 0.007464224006980658, + 0.004031945951282978, + 0.0074150776490569115, + 0.00529067637398839, + 0.0045205275528132915, + 0.008689003996551037, + 0.005073507782071829, + 0.006624636705964804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.0004463391669560224 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle macro shot offers an intimate glimpse into a subterranean ant chamber, its walls composed of rough, dense, dark earth and fine sand. At the center of this dimly lit, oval space lies a majestic queen ant, her body a smooth, deep mahogany. Her abdomen, or stomach, is disproportionately large and swollen, its semi-translucent sections straining and shimmering under a soft, off-camera light source that simulates the careful lighting of a documentary filmmaker. With slow, rhythmic contractions, she deliberately squeezes a tiny, pearlescent white egg onto the chamber's floor. Moist and rectangular, the newly laid egg joins a nearby carefully arranged nest, tended by a pair of smaller worker ants who act with swift, instinctive precision. The scene, dominated by rich, earthy browns and blacks, contrasts sharply with the bright, creamy white cluster of eggs, creating an atmosphere of concentration and almost reverence, capturing the quiet miracle of life unfolding in a hidden world.", + "seed": 16990, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_2/index_4481_seed_16990_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.036020043701864776, + "vlm_logprob": -0.036020043701864776, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.036020043701864776, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007463666144758463 + ], + "opposite_logprob_tokens": [ + -6.132463455200195 + ], + "logprob_reward": -0.007463666144758463, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.007463666144758463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23875916004180908, + 0.13976316154003143, + 0.06479841470718384, + 0.049010198563337326, + 0.06801822036504745, + 0.12404041737318039, + 0.0776793360710144, + 0.06233377754688263, + 0.10995543748140335, + 0.06564190238714218 + ], + "credit_map": [ + 0.025646653026342392, + 0.03704274073243141, + 0.029049912467598915, + 0.03012222796678543, + 0.039023492485284805, + 0.03642408922314644, + 0.022147513926029205, + 0.01930253393948078, + 0.012042190879583359, + 0.027538634836673737, + 0.018584294244647026, + 0.015183031558990479, + 0.025911914184689522, + 0.019747983664274216, + 0.00854243990033865, + 0.012212670408189297, + 0.006656579207628965, + 0.012144706211984158, + 0.006831603590399027, + 0.005361894145607948, + 0.00981731154024601, + 0.007787793874740601, + 0.009202714078128338, + 0.006995809264481068, + 0.004369825124740601, + 0.00728788785636425, + 0.004519475158303976, + 0.004886549897491932, + 0.008738239295780659, + 0.006040721666067839, + 0.00823406595736742, + 0.00493343360722065, + 0.005645516328513622, + 0.012401447631418705, + 0.00819899421185255, + 0.007394179701805115, + 0.011616873554885387, + 0.009285859763622284, + 0.0063997842371463776, + 0.007075564935803413, + 0.012049098499119282, + 0.023570813238620758, + 0.013662321493029594, + 0.012125848792493343, + 0.01986127905547619, + 0.01595880836248398, + 0.015449498780071735, + 0.011362751014530659, + 0.008445695973932743, + 0.01141010969877243, + 0.014233903959393501, + 0.007720374967902899, + 0.008572490885853767, + 0.010981447994709015, + 0.005981727037578821, + 0.010333583690226078, + 0.004561692010611296, + 0.012732834555208683, + 0.006201819982379675, + 0.006296013947576284, + 0.010402506217360497, + 0.009191092103719711, + 0.00483825383707881, + 0.008109565824270248, + 0.00857541523873806, + 0.013898010365664959, + 0.015039056539535522, + 0.01196036022156477, + 0.018794341012835503, + 0.016739744693040848, + 0.01470410916954279, + 0.010244391858577728, + 0.0062439837493002415, + 0.011102788150310516, + 0.007991745136678219, + 0.004398421850055456, + 0.01445576548576355, + 0.009936649352312088, + 0.005330910440534353, + 0.006181641016155481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.007463666144758463 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023307057097554207 + ], + "opposite_logprob_tokens": [ + -3.7733070850372314 + ], + "logprob_reward": -0.023307057097554207, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023307057097554207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27541062235832214, + 0.1505841612815857, + 0.037225622683763504, + 0.08901342004537582, + 0.07037421315908432, + 0.05483423173427582, + 0.09993015974760056, + 0.08462704718112946, + 0.07110141962766647, + 0.06689910590648651 + ], + "credit_map": [ + 0.03008505329489708, + 0.05592909827828407, + 0.032377976924180984, + 0.019284969195723534, + 0.05427509918808937, + 0.04348229989409447, + 0.017297329381108284, + 0.02267879992723465, + 0.013666938059031963, + 0.022363534197211266, + 0.018099592998623848, + 0.010088672861456871, + 0.026735050603747368, + 0.02017292007803917, + 0.022790171205997467, + 0.016667285934090614, + 0.004056974779814482, + 0.007335738744586706, + 0.005125814117491245, + 0.0032969005405902863, + 0.005084572359919548, + 0.0057960269041359425, + 0.0028897260781377554, + 0.0036398700904101133, + 0.008119851350784302, + 0.011781011708080769, + 0.01188539806753397, + 0.009705962613224983, + 0.012576060369610786, + 0.012303180061280727, + 0.01300741545855999, + 0.009634540416300297, + 0.009044786915183067, + 0.013443443924188614, + 0.010085050016641617, + 0.007450266275554895, + 0.008985691703855991, + 0.009662812575697899, + 0.005270561669021845, + 0.0064316000789403915, + 0.005763703025877476, + 0.008433391340076923, + 0.0074004121124744415, + 0.005360455252230167, + 0.010835413821041584, + 0.007487893104553223, + 0.004758051130920649, + 0.004794913809746504, + 0.010062206536531448, + 0.01438729278743267, + 0.014087626710534096, + 0.007775051519274712, + 0.021510493010282516, + 0.015802104026079178, + 0.006038281600922346, + 0.010267101228237152, + 0.006190072279423475, + 0.01447227317839861, + 0.011185603216290474, + 0.009238390251994133, + 0.011463564820587635, + 0.01105775497853756, + 0.010086090303957462, + 0.010933293960988522, + 0.0057178460992872715, + 0.012079359032213688, + 0.008771991357207298, + 0.0074612959288060665, + 0.012054631486535072, + 0.00830558966845274, + 0.009423385374248028, + 0.007287316955626011, + 0.0055680423974990845, + 0.011545251123607159, + 0.006689091213047504, + 0.006014971528202295, + 0.013255346566438675, + 0.008486518636345863, + 0.006862299516797066, + 0.008477580733597279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.023307057097554207 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003537565469742 + ], + "opposite_logprob_tokens": [ + -2.695035457611084 + ], + "logprob_reward": -0.07003537565469742, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003537565469742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14260810613632202, + 0.16070519387722015, + 0.05552952364087105, + 0.100471630692482, + 0.04069219529628754, + 0.04031192883849144, + 0.2208099067211151, + 0.12365154922008514, + 0.09921173751354218, + 0.01600833423435688 + ], + "credit_map": [ + 0.014424065127968788, + 0.02637016400694847, + 0.01622229442000389, + 0.014640421606600285, + 0.02093767747282982, + 0.019941044971346855, + 0.015585142187774181, + 0.01448726560920477, + 0.011968285776674747, + 0.02444208227097988, + 0.013155699707567692, + 0.014198532328009605, + 0.025500670075416565, + 0.02027805522084236, + 0.03075362555682659, + 0.020408224314451218, + 0.003854773472994566, + 0.008636604994535446, + 0.006371060386300087, + 0.006092840805649757, + 0.007466237526386976, + 0.007178342901170254, + 0.00905635766685009, + 0.006873296108096838, + 0.007058102637529373, + 0.013503734022378922, + 0.01321620773524046, + 0.009450425393879414, + 0.014486482366919518, + 0.014088908210396767, + 0.017720011994242668, + 0.010947750881314278, + 0.003473526332527399, + 0.0067162965424358845, + 0.004600743763148785, + 0.003948419354856014, + 0.004893932957202196, + 0.006097257602959871, + 0.006656607612967491, + 0.004305406007915735, + 0.0032535542268306017, + 0.005556232295930386, + 0.004944643937051296, + 0.004655587952584028, + 0.005931305233389139, + 0.004922355525195599, + 0.006377785932272673, + 0.0046704611741006374, + 0.01709391176700592, + 0.03737150505185127, + 0.02583482302725315, + 0.019295036792755127, + 0.038580991327762604, + 0.028088904917240143, + 0.02844032645225525, + 0.02610437572002411, + 0.010367458686232567, + 0.020093165338039398, + 0.015199179761111736, + 0.012537507340312004, + 0.017214970663189888, + 0.016026271507143974, + 0.01784531958401203, + 0.014367661438882351, + 0.007621017284691334, + 0.016732079908251762, + 0.011785958893597126, + 0.009845685213804245, + 0.015086717903614044, + 0.012078606523573399, + 0.015402992255985737, + 0.010658672079443932, + 0.0016050096601247787, + 0.001789107802323997, + 0.001798480749130249, + 0.0016875321744009852, + 0.0024355605710297823, + 0.002076613949611783, + 0.0027977561112493277, + 0.0018182711210101843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.07003537565469742 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789848193526268 + ], + "opposite_logprob_tokens": [ + -2.5789847373962402 + ], + "logprob_reward": -0.0789848193526268, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789848193526268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.216714009642601, + 0.18691854178905487, + 0.05770544707775116, + 0.06002677232027054, + 0.04457564651966095, + 0.07113153487443924, + 0.07689282298088074, + 0.07210614532232285, + 0.06961151957511902, + 0.14431752264499664 + ], + "credit_map": [ + 0.02222033590078354, + 0.03441159054636955, + 0.031795088201761246, + 0.02131287381052971, + 0.02843371219933033, + 0.0338544100522995, + 0.02322830632328987, + 0.021457679569721222, + 0.015764998272061348, + 0.025879770517349243, + 0.02107897214591503, + 0.014436484314501286, + 0.03255796432495117, + 0.024341559037566185, + 0.03281403332948685, + 0.02004474587738514, + 0.005641649477183819, + 0.01043420284986496, + 0.006060332525521517, + 0.0046523213386535645, + 0.007229558192193508, + 0.007625439204275608, + 0.00928441807627678, + 0.0067775254137814045, + 0.005026425234973431, + 0.008436326868832111, + 0.0068431394174695015, + 0.006836926098912954, + 0.008092401549220085, + 0.008707067929208279, + 0.009522865526378155, + 0.006561617366969585, + 0.005082468502223492, + 0.008004040457308292, + 0.0055518546141684055, + 0.004365202970802784, + 0.005743663292378187, + 0.006300326436758041, + 0.005262334831058979, + 0.0042657568119466305, + 0.00829883199185133, + 0.010059635154902935, + 0.0107285650447011, + 0.007804397493600845, + 0.00990080926567316, + 0.010746627114713192, + 0.007727486081421375, + 0.005865182261914015, + 0.007565297186374664, + 0.014041157439351082, + 0.010953090153634548, + 0.007090978790074587, + 0.01044598501175642, + 0.01046005543321371, + 0.006798560731112957, + 0.009537702426314354, + 0.006596804596483707, + 0.009039134718477726, + 0.009024000726640224, + 0.012641776353120804, + 0.006802242714911699, + 0.008526260033249855, + 0.008988299407064915, + 0.010487622581422329, + 0.00645768316462636, + 0.01361700426787138, + 0.010335870087146759, + 0.005845628213137388, + 0.008518384769558907, + 0.009253243915736675, + 0.008260463364422321, + 0.007323248311877251, + 0.01521468535065651, + 0.021257923915982246, + 0.014772937633097172, + 0.018195467069745064, + 0.025638587772846222, + 0.017481641843914986, + 0.020100800320506096, + 0.011655488051474094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0789848193526268 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, high-angle macro view of a subterranean ant chamber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a queen ant with a smooth, deep mahogany body and a disproportionately large, swollen abdomen visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the queen ant shown squeezing a tiny, pearlescent white egg onto the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two smaller worker ants visible tending to a cluster of eggs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by rich earthy browns and blacks contrasting with bright creamy white eggs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003093002596870065 + ], + "opposite_logprob_tokens": [ + -8.125308990478516 + ], + "logprob_reward": -0.0003093002596870065, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.0003093002596870065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30774080753326416, + 0.10873737931251526, + 0.06463650614023209, + 0.08159666508436203, + 0.03841736167669296, + 0.03812156990170479, + 0.11808031797409058, + 0.09463469684123993, + 0.05753026157617569, + 0.09050442278385162 + ], + "credit_map": [ + 0.029143525287508965, + 0.05625397711992264, + 0.040345609188079834, + 0.028055762872099876, + 0.04646541550755501, + 0.04485618695616722, + 0.030852321535348892, + 0.03176799789071083, + 0.008962237276136875, + 0.019778210669755936, + 0.011789423413574696, + 0.012719858437776566, + 0.015304925851523876, + 0.012827403843402863, + 0.01551698986440897, + 0.011838327161967754, + 0.005918810144066811, + 0.011552908457815647, + 0.005626919213682413, + 0.00535999471321702, + 0.014682958833873272, + 0.0082095293328166, + 0.007065438665449619, + 0.006219943985342979, + 0.006481255404651165, + 0.012598635628819466, + 0.009463598020374775, + 0.007747600320726633, + 0.013365506194531918, + 0.010613913647830486, + 0.01237888541072607, + 0.008947270922362804, + 0.0037728615570813417, + 0.007623101118952036, + 0.00406302697956562, + 0.004247892647981644, + 0.006109424866735935, + 0.004916648380458355, + 0.003998198080807924, + 0.0036862094420939684, + 0.004140314180403948, + 0.006126715801656246, + 0.004298339132219553, + 0.003809988731518388, + 0.005295285023748875, + 0.00482579693198204, + 0.00483521306887269, + 0.004789918661117554, + 0.0100905392318964, + 0.018645845353603363, + 0.01291594561189413, + 0.008637907914817333, + 0.02665826678276062, + 0.0179127249866724, + 0.009904368780553341, + 0.01331472396850586, + 0.007457914762198925, + 0.015814395621418953, + 0.011907227337360382, + 0.011007622815668583, + 0.0125050637871027, + 0.012506217695772648, + 0.011448495090007782, + 0.011987762525677681, + 0.0048586539924144745, + 0.009538402780890465, + 0.008721989579498768, + 0.004783843643963337, + 0.008499355986714363, + 0.007677186746150255, + 0.007216252386569977, + 0.006234575994312763, + 0.007859880104660988, + 0.014286777004599571, + 0.012357593514025211, + 0.009270219132304192, + 0.017501303926110268, + 0.012472761794924736, + 0.008650180883705616, + 0.008105706423521042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0003093002596870065 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_3/index_tensor([6371, 6371])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_3/index_tensor([6371, 6371])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..aaaa0dc10e5dcafd750eb40f9bfc30e285ed2cc4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_3/index_tensor([6371, 6371])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide shot pans right, closely following a young man as he zips along a rough asphalt surface on a skateboard. Shot with a Steve McCurry-esque documentary quality, the dazzling afternoon sun saturates the image's colors, setting against a vibrant backdrop of a vibrant urban streetscape—a vibrant canvas of colorful storefronts, intertwined power lines, and blurred pedestrians. The young man, wearing a black beanie, with wisps of sweat-soaked blond hair plastered to his forehead, wears a baggy charcoal gray T-shirt, frayed dark overalls, and dusty skateboard shoes. His gaze is fixed intently ahead. As the skateboard rushed towards an intersection, he leaned forward, and then with an extremely neat and smooth movement, he decisively stepped one foot on the ground, lifted the other foot, and jumped lightly from the skateboard that was still rushing forward at high speed, landing steadily on the sidewalk with his knees slightly bent, while the skateboard continued to slide forward and gradually disappeared from the screen. The whole process was full of unpretentious street energy and youthful dynamism under the sharp light and shadow.", + "seed": 15794, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_3/index_6371_seed_15794_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.026271213218569754, + "vlm_logprob": -0.026271213218569754, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.026271213218569754, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07160552591085434 + ], + "opposite_logprob_tokens": [ + -3.8216054439544678 + ], + "logprob_reward": -0.07160552591085434, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.07160552591085434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09466259926557541, + 0.05038655921816826, + 0.07183191925287247, + 0.03128954768180847, + 0.06282097101211548, + 0.0381191112101078, + 0.04437815025448799, + 0.10410993546247482, + 0.14281299710273743, + 0.35958823561668396 + ], + "credit_map": [ + 0.014514518901705742, + 0.012449147179722786, + 0.015036203898489475, + 0.010671763680875301, + 0.007789130322635174, + 0.010269809514284134, + 0.007483257446438074, + 0.016448771581053734, + 0.00877424143254757, + 0.008269153535366058, + 0.005972784478217363, + 0.005322540644556284, + 0.006732049863785505, + 0.005163211841136217, + 0.004497944377362728, + 0.005654637236148119, + 0.008035765960812569, + 0.008710619993507862, + 0.012212628498673439, + 0.006766530219465494, + 0.008126641623675823, + 0.010571695864200592, + 0.007072360720485449, + 0.010335675440728664, + 0.0026760732289403677, + 0.003221796592697501, + 0.0031637127976864576, + 0.004647157154977322, + 0.0041135684587061405, + 0.0029931622557342052, + 0.00633391784504056, + 0.004140160046517849, + 0.0061223143711686134, + 0.005046625155955553, + 0.00974955502897501, + 0.009372732602059841, + 0.007538652513176203, + 0.006492269225418568, + 0.00969620980322361, + 0.008802609518170357, + 0.003406410338357091, + 0.004916198551654816, + 0.006756586022675037, + 0.005253423936665058, + 0.003769837087020278, + 0.005348284728825092, + 0.003563324920833111, + 0.005105045158416033, + 0.0037892148829996586, + 0.007467180956155062, + 0.0063430024310946465, + 0.006473173853009939, + 0.004721201956272125, + 0.004230190999805927, + 0.005177870858460665, + 0.006176312454044819, + 0.006822885014116764, + 0.007119172252714634, + 0.02826373092830181, + 0.007657628506422043, + 0.005894831381738186, + 0.0095149464905262, + 0.008572394028306007, + 0.030264349654316902, + 0.009226351976394653, + 0.010529080405831337, + 0.021468505263328552, + 0.03104838728904724, + 0.010035510174930096, + 0.011037086136639118, + 0.03031095489859581, + 0.01915712095797062, + 0.02303597517311573, + 0.024362018331885338, + 0.0787658616900444, + 0.03094087354838848, + 0.01833261176943779, + 0.04580477252602577, + 0.03294413164258003, + 0.10540197789669037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.07160552591085434 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607622765004635 + ], + "opposite_logprob_tokens": [ + -4.14107608795166 + ], + "logprob_reward": -0.01607622765004635, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607622765004635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19224952161312103, + 0.09569195657968521, + 0.057126328349113464, + 0.044071588665246964, + 0.0680592879652977, + 0.04597076028585434, + 0.04403995722532272, + 0.09981875121593475, + 0.13837625086307526, + 0.21459555625915527 + ], + "credit_map": [ + 0.024109289050102234, + 0.02295583114027977, + 0.014756999909877777, + 0.02003447711467743, + 0.027639953419566154, + 0.021037153899669647, + 0.025056656450033188, + 0.03665916249155998, + 0.014764245599508286, + 0.014135722070932388, + 0.010355044156312943, + 0.012983487918972969, + 0.013728257268667221, + 0.006485540885478258, + 0.014296835288405418, + 0.008942825719714165, + 0.008798529393970966, + 0.007447168696671724, + 0.005918900016695261, + 0.008353397250175476, + 0.00707901269197464, + 0.005502961575984955, + 0.007249918766319752, + 0.006776437163352966, + 0.003569205990061164, + 0.004301630891859531, + 0.0066114808432757854, + 0.005898195318877697, + 0.005810654256492853, + 0.005402731709182262, + 0.00701921246945858, + 0.00545847974717617, + 0.006876344792544842, + 0.004978726152330637, + 0.009827975183725357, + 0.008300332352519035, + 0.008116776123642921, + 0.011654461733996868, + 0.008955619297921658, + 0.009349051862955093, + 0.0043066334910690784, + 0.005974805913865566, + 0.005545880179852247, + 0.005807841662317514, + 0.0054797022603452206, + 0.007434549741446972, + 0.00452904449775815, + 0.0068923053331673145, + 0.004448930732905865, + 0.006340263411402702, + 0.008815936744213104, + 0.006240602117031813, + 0.0039922152645885944, + 0.003673981409519911, + 0.0046930923126637936, + 0.005834933836013079, + 0.006919315084815025, + 0.00803952757269144, + 0.02733338251709938, + 0.00753512978553772, + 0.0072058942168951035, + 0.007497420068830252, + 0.012270009145140648, + 0.023018069565296173, + 0.009367728605866432, + 0.010135569609701633, + 0.017442554235458374, + 0.031002908945083618, + 0.008694067597389221, + 0.018875807523727417, + 0.023554624989628792, + 0.019302980974316597, + 0.014892956241965294, + 0.016593601554632187, + 0.02517295628786087, + 0.032575953751802444, + 0.020812874659895897, + 0.03387058898806572, + 0.03172912076115608, + 0.038947492837905884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.01607622765004635 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011069436557590961 + ], + "opposite_logprob_tokens": [ + -4.511069297790527 + ], + "logprob_reward": -0.011069436557590961, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011069436557590961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0755789577960968, + 0.08805155754089355, + 0.04938715696334839, + 0.058253563940525055, + 0.08082547783851624, + 0.07875674962997437, + 0.1354481428861618, + 0.13483895361423492, + 0.1546986997127533, + 0.14416071772575378 + ], + "credit_map": [ + 0.01150705385953188, + 0.010948157869279385, + 0.007006418891251087, + 0.008267734199762344, + 0.010770417749881744, + 0.009518525563180447, + 0.009082518517971039, + 0.008478134870529175, + 0.019965052604675293, + 0.00839821808040142, + 0.008285384625196457, + 0.012624635361135006, + 0.012501798570156097, + 0.0075698536820709705, + 0.009603368118405342, + 0.009103246964514256, + 0.006302602589130402, + 0.005893170367926359, + 0.007887654937803745, + 0.006316623650491238, + 0.007428256794810295, + 0.006322414148598909, + 0.004930196795612574, + 0.004306239075958729, + 0.006879087071865797, + 0.004827382508665323, + 0.009620307013392448, + 0.006057784892618656, + 0.006018097046762705, + 0.00733823049813509, + 0.009017576463520527, + 0.008495100773870945, + 0.006429224740713835, + 0.00506896385923028, + 0.01028035581111908, + 0.010685797780752182, + 0.008480553515255451, + 0.010431695729494095, + 0.01353428978472948, + 0.015914592891931534, + 0.0064347791485488415, + 0.009539095684885979, + 0.010641921311616898, + 0.011658595874905586, + 0.005094751249998808, + 0.009803906083106995, + 0.00876898318529129, + 0.016814718022942543, + 0.011409608647227287, + 0.02117227204144001, + 0.014387805014848709, + 0.024562230333685875, + 0.019188862293958664, + 0.00990142859518528, + 0.019233565777540207, + 0.015592364594340324, + 0.010135576128959656, + 0.009509105235338211, + 0.028739606961607933, + 0.011438756249845028, + 0.011805226095020771, + 0.01421371754258871, + 0.017833152785897255, + 0.031163809821009636, + 0.013893399387598038, + 0.010840127244591713, + 0.02284913696348667, + 0.024232923984527588, + 0.0121085774153471, + 0.027569537982344627, + 0.020550277084112167, + 0.022654719650745392, + 0.012778453528881073, + 0.011453492566943169, + 0.03277888894081116, + 0.015106855891644955, + 0.014685234054923058, + 0.0226689875125885, + 0.016855506226420403, + 0.01783330738544464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.011069436557590961 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002815927378833294 + ], + "opposite_logprob_tokens": [ + -5.8778157234191895 + ], + "logprob_reward": -0.002815927378833294, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002815927378833294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1323985457420349, + 0.036683011800050735, + 0.03128208965063095, + 0.03442467749118805, + 0.07671989500522614, + 0.04117990657687187, + 0.015613972209393978, + 0.17060039937496185, + 0.2062651813030243, + 0.25483232736587524 + ], + "credit_map": [ + 0.019862810149788857, + 0.03053704835474491, + 0.022464361041784286, + 0.011602818965911865, + 0.0075869266875088215, + 0.012038218788802624, + 0.0074285841546952724, + 0.020877769216895103, + 0.006897903513163328, + 0.006135793402791023, + 0.005632828455418348, + 0.0039051242638379335, + 0.003394021186977625, + 0.0036127560306340456, + 0.002559678629040718, + 0.004544906318187714, + 0.004159724339842796, + 0.004780901130288839, + 0.004725957754999399, + 0.003655951237305999, + 0.00358818331733346, + 0.003101459238678217, + 0.0029871375299990177, + 0.004282771609723568, + 0.005410414654761553, + 0.005400919821113348, + 0.004971568007022142, + 0.004288849420845509, + 0.0026012680027633905, + 0.003321070922538638, + 0.0038210758939385414, + 0.0046095093712210655, + 0.009750833734869957, + 0.009602833539247513, + 0.010713290423154831, + 0.010306892916560173, + 0.009104601107537746, + 0.012789096683263779, + 0.0062130531296133995, + 0.008239292539656162, + 0.004512436222285032, + 0.00646711140871048, + 0.005282930098474026, + 0.007762141525745392, + 0.004212989006191492, + 0.002510128077119589, + 0.004055473487824202, + 0.006376692559570074, + 0.0016276949318125844, + 0.0021532189566642046, + 0.002053426345810294, + 0.00230400706641376, + 0.00171365134883672, + 0.0015892644878476858, + 0.0017377034528180957, + 0.0024350048042833805, + 0.012462183833122253, + 0.01609213836491108, + 0.05893300473690033, + 0.012633227743208408, + 0.00674425857141614, + 0.011912642046809196, + 0.006507867015898228, + 0.045315083116292953, + 0.010567420162260532, + 0.014094549231231213, + 0.030366122722625732, + 0.05240439996123314, + 0.01074565015733242, + 0.014439474791288376, + 0.04492434859275818, + 0.028723223134875298, + 0.021167593076825142, + 0.014956878498196602, + 0.07123417407274246, + 0.01206066831946373, + 0.008588667958974838, + 0.03315533697605133, + 0.009104646742343903, + 0.08456436544656754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.002815927378833294 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029788948595523834 + ], + "opposite_logprob_tokens": [ + -3.5297889709472656 + ], + "logprob_reward": -0.029788948595523834, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029788948595523834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1869295835494995, + 0.05630062520503998, + 0.03943730890750885, + 0.0381554551422596, + 0.02924998290836811, + 0.03922522813081741, + 0.07960376143455505, + 0.13214464485645294, + 0.08073968440294266, + 0.31821367144584656 + ], + "credit_map": [ + 0.02939995750784874, + 0.02931688353419304, + 0.023806992918252945, + 0.021291885524988174, + 0.01567082293331623, + 0.02017672173678875, + 0.013672678731381893, + 0.03359364718198776, + 0.009465290233492851, + 0.006138224620372057, + 0.007782048545777798, + 0.00635621277615428, + 0.010172533802688122, + 0.003538684220984578, + 0.007446379400789738, + 0.005401254631578922, + 0.009019392542541027, + 0.003673339495435357, + 0.005462314002215862, + 0.003415177809074521, + 0.002810976468026638, + 0.003463021945208311, + 0.004124966915696859, + 0.007468121126294136, + 0.004551319871097803, + 0.005830733105540276, + 0.005428838077932596, + 0.006131398491561413, + 0.003198739606887102, + 0.0038550742901861668, + 0.0038442863151431084, + 0.005315067246556282, + 0.0023573492653667927, + 0.002126247389242053, + 0.004338410217314959, + 0.004479545634239912, + 0.003081723814830184, + 0.004613509867340326, + 0.0046426779590547085, + 0.00361051969230175, + 0.005324698518961668, + 0.0058250525034964085, + 0.005091375205665827, + 0.006939800921827555, + 0.003745370777323842, + 0.0035118006635457277, + 0.003734044497832656, + 0.005053084809333086, + 0.007524048909544945, + 0.01591106317937374, + 0.011754398234188557, + 0.00946857314556837, + 0.00834973156452179, + 0.006124611012637615, + 0.01177982334047556, + 0.008691512048244476, + 0.009791589342057705, + 0.010630592703819275, + 0.034095749258995056, + 0.013299895450472832, + 0.0074983094818890095, + 0.016740944236516953, + 0.014407164417207241, + 0.025680411607027054, + 0.006555120460689068, + 0.008419905789196491, + 0.013868726789951324, + 0.01475281547755003, + 0.00649746460840106, + 0.0066613745875656605, + 0.0123295234516263, + 0.011654751375317574, + 0.030399685725569725, + 0.02818395383656025, + 0.03956685587763786, + 0.028084484860301018, + 0.013715889304876328, + 0.07848668098449707, + 0.01847110129892826, + 0.08130502700805664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.029788948595523834 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide shot pans right, closely following a young man as he zips along a rough asphalt surface on a skateboard. Shot with a Steve McCurry-esque documentary quality, the dazzling afternoon sun saturates the image's colors, setting against a vibrant backdrop of a vibrant urban streetscape—a vibrant canvas of colorful storefronts, intertwined power lines, and blurred pedestrians. The young man, wearing a black beanie, with wisps of sweat-soaked blond hair plastered to his forehead, wears a baggy charcoal gray T-shirt, frayed dark overalls, and dusty skateboard shoes. His gaze is fixed intently ahead. As the skateboard rushed towards an intersection, he leaned forward, and then with an extremely neat and smooth movement, he decisively stepped one foot on the ground, lifted the other foot, and jumped lightly from the skateboard that was still rushing forward at high speed, landing steadily on the sidewalk with his knees slightly bent, while the skateboard continued to slide forward and gradually disappeared from the screen. The whole process was full of unpretentious street energy and youthful dynamism under the sharp light and shadow.", + "seed": 15795, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_3/index_6371_seed_15795_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04799729329533875, + "vlm_logprob": -0.04799729329533875, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.04799729329533875, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08212988078594208 + ], + "opposite_logprob_tokens": [ + -3.957129955291748 + ], + "logprob_reward": -0.08212988078594208, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.08212988078594208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1992454081773758, + 0.08428932726383209, + 0.06371039152145386, + 0.046491045504808426, + 0.09701237082481384, + 0.09007462859153748, + 0.05614475905895233, + 0.0998820960521698, + 0.14681124687194824, + 0.11633875221014023 + ], + "credit_map": [ + 0.02620220184326172, + 0.021764887496829033, + 0.030123243108391762, + 0.026204053312540054, + 0.0358121320605278, + 0.010702633298933506, + 0.03745002672076225, + 0.01098623126745224, + 0.01032013725489378, + 0.014263994060456753, + 0.00924451183527708, + 0.01005081832408905, + 0.009237207472324371, + 0.012911215424537659, + 0.007529738824814558, + 0.010731707327067852, + 0.009360204450786114, + 0.008330787532031536, + 0.01592249609529972, + 0.009206523187458515, + 0.003827974433079362, + 0.0050177304074168205, + 0.0053595746867358685, + 0.006685101427137852, + 0.004671217408031225, + 0.004034664016216993, + 0.005969843827188015, + 0.005417735315859318, + 0.007204936817288399, + 0.004405319690704346, + 0.007562886457890272, + 0.007224439177662134, + 0.0067584398202598095, + 0.007339260540902615, + 0.018139775842428207, + 0.014286522753536701, + 0.014727180823683739, + 0.01224198006093502, + 0.013493536040186882, + 0.010025675408542156, + 0.0077967047691345215, + 0.009751258417963982, + 0.013207059353590012, + 0.01708722487092018, + 0.008855406194925308, + 0.01018269918859005, + 0.011038755998015404, + 0.012155524455010891, + 0.006301324348896742, + 0.007766565307974815, + 0.00680337380617857, + 0.007567003834992647, + 0.005570198874920607, + 0.006534360349178314, + 0.007000419776886702, + 0.008601514622569084, + 0.012559633702039719, + 0.013017524965107441, + 0.015233202837407589, + 0.011905677616596222, + 0.009829230606555939, + 0.007179351523518562, + 0.020121736451983452, + 0.010035732761025429, + 0.01518864557147026, + 0.017803804948925972, + 0.021099243313074112, + 0.019130395725369453, + 0.014189540408551693, + 0.018507711589336395, + 0.02150825969874859, + 0.019383639097213745, + 0.015056110918521881, + 0.011204286478459835, + 0.018145548179745674, + 0.017246121540665627, + 0.012690954841673374, + 0.015520689077675343, + 0.01371253002434969, + 0.012762518599629402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.08212988078594208 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371031582355499 + ], + "opposite_logprob_tokens": [ + -3.408710241317749 + ], + "logprob_reward": -0.03371031582355499, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371031582355499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08961620926856995, + 0.06304369866847992, + 0.09992994368076324, + 0.1158510223031044, + 0.21006156504154205, + 0.11442844569683075, + 0.053775303065776825, + 0.07578910142183304, + 0.11984357982873917, + 0.057661160826683044 + ], + "credit_map": [ + 0.012734295800328255, + 0.015782276168465614, + 0.01154255960136652, + 0.009168682619929314, + 0.009747968055307865, + 0.006456580478698015, + 0.011424466036260128, + 0.012759375385940075, + 0.008721358142793179, + 0.005945904180407524, + 0.00736573338508606, + 0.015515970066189766, + 0.008605166338384151, + 0.004240457899868488, + 0.007766387891024351, + 0.0048827179707586765, + 0.008113102987408638, + 0.008015808649361134, + 0.022371407598257065, + 0.01436510682106018, + 0.010410478338599205, + 0.00760556198656559, + 0.017464952543377876, + 0.011583522893488407, + 0.006211892236024141, + 0.010745077393949032, + 0.014819187112152576, + 0.0122273750603199, + 0.019391192123293877, + 0.016719313338398933, + 0.01915603131055832, + 0.016580943018198013, + 0.014878532849252224, + 0.016261672601103783, + 0.03956467658281326, + 0.04786832258105278, + 0.025648191571235657, + 0.014477421529591084, + 0.03124009259045124, + 0.020122649148106575, + 0.0087185213342309, + 0.014321938157081604, + 0.017366649582982063, + 0.02226758375763893, + 0.015334160067141056, + 0.011968525126576424, + 0.014807279221713543, + 0.00964377447962761, + 0.007500407751649618, + 0.00875635165721178, + 0.0065153613686561584, + 0.005969803314656019, + 0.004177697002887726, + 0.004921335726976395, + 0.0062584844417870045, + 0.009675859473645687, + 0.006268814671784639, + 0.007039826363325119, + 0.014828629791736603, + 0.018430206924676895, + 0.007742980495095253, + 0.008692974224686623, + 0.0073579722084105015, + 0.005427692551165819, + 0.012821364216506481, + 0.014800634235143661, + 0.015550977550446987, + 0.014920838177204132, + 0.01307433657348156, + 0.016643265262246132, + 0.01712966337800026, + 0.014902488328516483, + 0.009468475356698036, + 0.0053602419793605804, + 0.00832875445485115, + 0.005673356354236603, + 0.005643732845783234, + 0.0069963582791388035, + 0.007543604355305433, + 0.008646629750728607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.03371031582355499 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624572142958641 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624572142958641, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624572142958641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13452830910682678, + 0.07308982312679291, + 0.04318947717547417, + 0.07038045674562454, + 0.16122500598430634, + 0.10672040283679962, + 0.05543685331940651, + 0.09926432371139526, + 0.17435461282730103, + 0.08181075006723404 + ], + "credit_map": [ + 0.017837215214967728, + 0.020933357998728752, + 0.015630554407835007, + 0.015355953015387058, + 0.01995556242763996, + 0.008200828917324543, + 0.02852802723646164, + 0.008086809888482094, + 0.011951552703976631, + 0.009824187494814396, + 0.011577034369111061, + 0.01685836911201477, + 0.00557120144367218, + 0.00539665762335062, + 0.0071783410385251045, + 0.004732477944344282, + 0.005730147007852793, + 0.005649417173117399, + 0.006835377775132656, + 0.006104436703026295, + 0.004013551399111748, + 0.004441662225872278, + 0.005222099367529154, + 0.005192783661186695, + 0.006739486940205097, + 0.006335228681564331, + 0.009798266924917698, + 0.008650371804833412, + 0.009686860255897045, + 0.006501934491097927, + 0.013225771486759186, + 0.009442532435059547, + 0.011199209839105606, + 0.01107206754386425, + 0.04151177778840065, + 0.021029245108366013, + 0.021747207269072533, + 0.01533158402889967, + 0.021460719406604767, + 0.017873192206025124, + 0.008174664340913296, + 0.010799656622111797, + 0.01709270291030407, + 0.024412695318460464, + 0.015429588034749031, + 0.011386106722056866, + 0.010642876848578453, + 0.008782107383012772, + 0.008014889433979988, + 0.006562969647347927, + 0.0065906718373298645, + 0.009213469922542572, + 0.0045155370607972145, + 0.005643431097269058, + 0.006326577626168728, + 0.008569308556616306, + 0.0064849830232560635, + 0.00778778363019228, + 0.025497134774923325, + 0.03136759623885155, + 0.0063939993269741535, + 0.006000040099024773, + 0.009605355560779572, + 0.006127440836280584, + 0.014517689123749733, + 0.01756255328655243, + 0.01917198859155178, + 0.024869007989764214, + 0.020160295069217682, + 0.025914208963513374, + 0.024941913783550262, + 0.027216970920562744, + 0.016001390293240547, + 0.006972188595682383, + 0.010729975998401642, + 0.01132030040025711, + 0.008196712471544743, + 0.010150475427508354, + 0.009480508044362068, + 0.00895919930189848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.007624572142958641 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036117101553827524 + ], + "opposite_logprob_tokens": [ + -5.6286115646362305 + ], + "logprob_reward": -0.0036117101553827524, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036117101553827524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15124347805976868, + 0.15695442259311676, + 0.038042839616537094, + 0.11776164919137955, + 0.09431717544794083, + 0.06380433589220047, + 0.05877454951405525, + 0.07069505751132965, + 0.10081195086240768, + 0.14759458601474762 + ], + "credit_map": [ + 0.0239560566842556, + 0.02273350954055786, + 0.021143198013305664, + 0.020875008776783943, + 0.020449021831154823, + 0.011309230700135231, + 0.016984200105071068, + 0.013793244026601315, + 0.03394382819533348, + 0.016581738367676735, + 0.02226886712014675, + 0.052107010036706924, + 0.00542491627857089, + 0.007693594321608543, + 0.010794986970722675, + 0.008139455690979958, + 0.005668295081704855, + 0.00389608065597713, + 0.004934101365506649, + 0.006442047655582428, + 0.0027236405294388533, + 0.003354409709572792, + 0.006527610123157501, + 0.004496648907661438, + 0.013283365406095982, + 0.01537216268479824, + 0.018289053812623024, + 0.022958196699619293, + 0.010306909680366516, + 0.012892781756818295, + 0.00946243666112423, + 0.015196731314063072, + 0.007939162664115429, + 0.008359507657587528, + 0.014137661084532738, + 0.012397372163832188, + 0.01387150026857853, + 0.011248797178268433, + 0.01623106747865677, + 0.010132099501788616, + 0.006119609344750643, + 0.006780865602195263, + 0.009197626262903214, + 0.009382955729961395, + 0.00848789606243372, + 0.007989255711436272, + 0.007309305481612682, + 0.00853681843727827, + 0.004314976744353771, + 0.006332450546324253, + 0.008629539981484413, + 0.006927995476871729, + 0.004307182971388102, + 0.00764424866065383, + 0.009271848015487194, + 0.011346302926540375, + 0.004723145626485348, + 0.00844593159854412, + 0.013418075628578663, + 0.01353913638740778, + 0.006105955224484205, + 0.00758986035361886, + 0.008415411226451397, + 0.008457536809146404, + 0.00798323005437851, + 0.01498729083687067, + 0.012430391274392605, + 0.009638747200369835, + 0.012162924744188786, + 0.010886632837355137, + 0.016337769106030464, + 0.01638496108353138, + 0.016666442155838013, + 0.01194851752370596, + 0.02296578511595726, + 0.012612889520823956, + 0.020028598606586456, + 0.024769295006990433, + 0.015995563939213753, + 0.022607484832406044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0036117101553827524 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290998756885529 + ], + "opposite_logprob_tokens": [ + -2.237910032272339 + ], + "logprob_reward": -0.11290998756885529, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290998756885529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27477821707725525, + 0.20644891262054443, + 0.1303582340478897, + 0.0738239511847496, + 0.05283686891198158, + 0.03892318531870842, + 0.028167473152279854, + 0.04718135669827461, + 0.06955137103796005, + 0.07793045043945312 + ], + "credit_map": [ + 0.039130233228206635, + 0.053817469626665115, + 0.027964357286691666, + 0.031962454319000244, + 0.030962804332375526, + 0.02673567272722721, + 0.03428298979997635, + 0.029922207817435265, + 0.0448249951004982, + 0.019357552751898766, + 0.02994612790644169, + 0.06964781135320663, + 0.008616585284471512, + 0.009930221363902092, + 0.014712799340486526, + 0.009412819519639015, + 0.024837128818035126, + 0.014504696242511272, + 0.01588810607790947, + 0.02052297815680504, + 0.011245851404964924, + 0.01735268533229828, + 0.011742047034204006, + 0.014264734461903572, + 0.008329643867909908, + 0.010205128230154514, + 0.00922261644154787, + 0.014172378927469254, + 0.006810750812292099, + 0.006372034549713135, + 0.0066857850179076195, + 0.012025616131722927, + 0.00839208997786045, + 0.008253148756921291, + 0.00784007553011179, + 0.007949061691761017, + 0.005817806348204613, + 0.0034148162230849266, + 0.005935207474976778, + 0.0052346657030284405, + 0.003009013133123517, + 0.003733416087925434, + 0.005576618015766144, + 0.0062972805462777615, + 0.004659595433622599, + 0.004044335335493088, + 0.004194822162389755, + 0.007408104371279478, + 0.0023425465915352106, + 0.0023219245485961437, + 0.003898061579093337, + 0.0031745273154228926, + 0.00254681590013206, + 0.003442215733230114, + 0.004507767502218485, + 0.005933613050729036, + 0.0054568140767514706, + 0.006092026364058256, + 0.007383489981293678, + 0.005175236612558365, + 0.005084079224616289, + 0.006086974870413542, + 0.0063125379383563995, + 0.0055901953019201756, + 0.007405674085021019, + 0.006722271908074617, + 0.00678649079054594, + 0.007540269289165735, + 0.00869623851031065, + 0.008046496659517288, + 0.011980319395661354, + 0.012373613193631172, + 0.011048962362110615, + 0.008793502114713192, + 0.009996582753956318, + 0.009784362278878689, + 0.010461485013365746, + 0.012298800982534885, + 0.007388389203697443, + 0.008158361539244652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.11290998756885529 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide shot pans right, closely following a young man as he zips along a rough asphalt surface on a skateboard. Shot with a Steve McCurry-esque documentary quality, the dazzling afternoon sun saturates the image's colors, setting against a vibrant backdrop of a vibrant urban streetscape—a vibrant canvas of colorful storefronts, intertwined power lines, and blurred pedestrians. The young man, wearing a black beanie, with wisps of sweat-soaked blond hair plastered to his forehead, wears a baggy charcoal gray T-shirt, frayed dark overalls, and dusty skateboard shoes. His gaze is fixed intently ahead. As the skateboard rushed towards an intersection, he leaned forward, and then with an extremely neat and smooth movement, he decisively stepped one foot on the ground, lifted the other foot, and jumped lightly from the skateboard that was still rushing forward at high speed, landing steadily on the sidewalk with his knees slightly bent, while the skateboard continued to slide forward and gradually disappeared from the screen. The whole process was full of unpretentious street energy and youthful dynamism under the sharp light and shadow.", + "seed": 15796, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_3/index_6371_seed_15796_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02702461234293878, + "vlm_logprob": -0.02702461234293878, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.02702461234293878, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08620189875364304 + ], + "opposite_logprob_tokens": [ + -3.3362019062042236 + ], + "logprob_reward": -0.08620189875364304, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.08620189875364304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13823294639587402, + 0.0559336394071579, + 0.03890243172645569, + 0.19001826643943787, + 0.0193931981921196, + 0.06451502442359924, + 0.3779011070728302, + 0.06599410623311996, + 0.018516601994633675, + 0.030592717230319977 + ], + "credit_map": [ + 0.018928689882159233, + 0.02155124396085739, + 0.01434379257261753, + 0.03600011020898819, + 0.013078240677714348, + 0.008781550452113152, + 0.014249473810195923, + 0.011299853213131428, + 0.006860681343823671, + 0.008386882022023201, + 0.00882477406412363, + 0.007901737466454506, + 0.005388898774981499, + 0.005954515188932419, + 0.00545301241800189, + 0.007163139525800943, + 0.004471967928111553, + 0.00471454206854105, + 0.0032928334549069405, + 0.004831012338399887, + 0.005124759394675493, + 0.005243552383035421, + 0.007144388277083635, + 0.004079372156411409, + 0.013894258998334408, + 0.0077578104101121426, + 0.04983939602971077, + 0.008448132313787937, + 0.006001469679176807, + 0.014474009163677692, + 0.018462948501110077, + 0.07114022970199585, + 0.0024832335766404867, + 0.0015319467056542635, + 0.002788197249174118, + 0.0015512540703639388, + 0.0021762442775070667, + 0.0033871473278850317, + 0.0018465904286131263, + 0.0036285831592977047, + 0.005628439597785473, + 0.0034064531791955233, + 0.0201448705047369, + 0.005956439301371574, + 0.005288895219564438, + 0.00588569138199091, + 0.00412901071831584, + 0.014075225219130516, + 0.025767860934138298, + 0.00839629303663969, + 0.11401236802339554, + 0.0394248329102993, + 0.027710752561688423, + 0.02000838704407215, + 0.021719954907894135, + 0.12086063623428345, + 0.0050825499929487705, + 0.00489283399656415, + 0.017357872799038887, + 0.011748737655580044, + 0.007532297167927027, + 0.005258035380393267, + 0.008919152431190014, + 0.005202631000429392, + 0.0019251538906246424, + 0.0019479993497952819, + 0.0018720810767263174, + 0.002086360240355134, + 0.002629691269248724, + 0.0034585134126245975, + 0.0021275593899190426, + 0.0024692434817552567, + 0.0044942512176930904, + 0.002179396338760853, + 0.0042161173187196255, + 0.004048697184771299, + 0.0028122097719460726, + 0.0030329525470733643, + 0.004887732211500406, + 0.004921360407024622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.08620189875364304 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023296808823943138 + ], + "opposite_logprob_tokens": [ + -3.77329683303833 + ], + "logprob_reward": -0.023296808823943138, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023296808823943138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0906202644109726, + 0.05143998563289642, + 0.07791987806558609, + 0.1900586187839508, + 0.044862810522317886, + 0.08008744567632675, + 0.34599384665489197, + 0.06584074348211288, + 0.02433868870139122, + 0.02883773110806942 + ], + "credit_map": [ + 0.00874757207930088, + 0.013265592977404594, + 0.00854488369077444, + 0.023324502632021904, + 0.00861765630543232, + 0.007997993379831314, + 0.01109230611473322, + 0.009029749780893326, + 0.006225834600627422, + 0.007577615790069103, + 0.006164259277284145, + 0.005086265504360199, + 0.009510618634521961, + 0.005425786133855581, + 0.006055579055100679, + 0.005394018720835447, + 0.007502307649701834, + 0.00932812038809061, + 0.007042353041470051, + 0.007161085028201342, + 0.013856587931513786, + 0.008137419819831848, + 0.01650560460984707, + 0.008386388421058655, + 0.017924753949046135, + 0.008220577612519264, + 0.034940093755722046, + 0.007225978188216686, + 0.008799627423286438, + 0.016433943063020706, + 0.024515608325600624, + 0.07199802249670029, + 0.010940480045974255, + 0.0041662524454295635, + 0.004454689100384712, + 0.0057189553044736385, + 0.004563912749290466, + 0.0052800956182181835, + 0.004494950640946627, + 0.005243467167019844, + 0.009169379249215126, + 0.005968777928501368, + 0.01885739527642727, + 0.007864793762564659, + 0.008651201613247395, + 0.00714093167334795, + 0.009467978030443192, + 0.012966987676918507, + 0.024065619334578514, + 0.009127418510615826, + 0.10277880728244781, + 0.03608613833785057, + 0.02526041492819786, + 0.01818361133337021, + 0.020162856206297874, + 0.11032898724079132, + 0.004432272166013718, + 0.005150855053216219, + 0.019535651430487633, + 0.020169902592897415, + 0.0037861070595681667, + 0.0029488522559404373, + 0.004778444766998291, + 0.0050386530347168446, + 0.002585559617727995, + 0.002263805363327265, + 0.0022349655628204346, + 0.0024487129412591457, + 0.003799684112891555, + 0.004208078607916832, + 0.002374138217419386, + 0.004423744510859251, + 0.003485914785414934, + 0.0024842671118676662, + 0.0038482288364320993, + 0.004265499301254749, + 0.0034442837350070477, + 0.0030925299506634474, + 0.003924322314560413, + 0.0042926836758852005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.023296808823943138 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251067224889994 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005251067224889994, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251067224889994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10101219266653061, + 0.04657145217061043, + 0.04824620485305786, + 0.20509220659732819, + 0.055761128664016724, + 0.157020702958107, + 0.21341396868228912, + 0.07994489371776581, + 0.02654377743601799, + 0.06639345735311508 + ], + "credit_map": [ + 0.012103564105927944, + 0.014336075633764267, + 0.012913990765810013, + 0.022702664136886597, + 0.009760943241417408, + 0.008349668234586716, + 0.012638000771403313, + 0.0082072913646698, + 0.005012271460145712, + 0.006675784941762686, + 0.0077241589315235615, + 0.006100942380726337, + 0.004929845687001944, + 0.006252804771065712, + 0.004220164846628904, + 0.005655477289110422, + 0.006738693919032812, + 0.007086620200425386, + 0.005621789023280144, + 0.008635121397674084, + 0.005230648443102837, + 0.004465359263122082, + 0.004466690588742495, + 0.006001284345984459, + 0.017506804317235947, + 0.007151561789214611, + 0.06524794548749924, + 0.009513241238892078, + 0.0060002803802490234, + 0.014550607651472092, + 0.01599755510687828, + 0.06912421435117722, + 0.01238071545958519, + 0.0038197257090359926, + 0.00841091014444828, + 0.004245573654770851, + 0.006160483229905367, + 0.007380460388958454, + 0.004444584250450134, + 0.008918674662709236, + 0.014033581130206585, + 0.011788050644099712, + 0.0433746799826622, + 0.01445178035646677, + 0.014738885685801506, + 0.013873777352273464, + 0.013116276822984219, + 0.031643666326999664, + 0.01593380980193615, + 0.005382961593568325, + 0.06393101811408997, + 0.022037643939256668, + 0.015608705580234528, + 0.01115355920046568, + 0.012256573885679245, + 0.06710968166589737, + 0.005911176558583975, + 0.006531835533678532, + 0.015469195321202278, + 0.01358792558312416, + 0.014330153353512287, + 0.006058526691049337, + 0.011384465731680393, + 0.00667161587625742, + 0.002175229135900736, + 0.002526103751733899, + 0.002570521552115679, + 0.003709740238264203, + 0.004131545312702656, + 0.005019948352128267, + 0.002944288309663534, + 0.003466402180492878, + 0.0063956002704799175, + 0.005494189448654652, + 0.00745073426514864, + 0.01032808143645525, + 0.007202211767435074, + 0.006955636199563742, + 0.01009200420230627, + 0.012475005351006985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.005251067224889994 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021964015904814005 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.0021964015904814005, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021964015904814005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1252085268497467, + 0.067215196788311, + 0.1067773625254631, + 0.12093424052000046, + 0.1020875945687294, + 0.09961345046758652, + 0.10134965181350708, + 0.06026948615908623, + 0.06114437058568001, + 0.15539999306201935 + ], + "credit_map": [ + 0.013428045436739922, + 0.01359329279512167, + 0.01703290082514286, + 0.03241125866770744, + 0.008729199878871441, + 0.014212317764759064, + 0.00759797403588891, + 0.018203558400273323, + 0.007812050171196461, + 0.007872497662901878, + 0.0105288689956069, + 0.009979295544326305, + 0.006946357432752848, + 0.006507020443677902, + 0.007525531109422445, + 0.01004359032958746, + 0.01467420719563961, + 0.01771409995853901, + 0.011420036666095257, + 0.017574165016412735, + 0.010778982192277908, + 0.010027523152530193, + 0.011594909243285656, + 0.012993456795811653, + 0.02228756621479988, + 0.005471483338624239, + 0.023417621850967407, + 0.005612659268081188, + 0.007768184877932072, + 0.01775282621383667, + 0.01684499904513359, + 0.02177891507744789, + 0.020501215010881424, + 0.006454753689467907, + 0.018481314182281494, + 0.006757275201380253, + 0.012898575514554977, + 0.013459458947181702, + 0.008537563495337963, + 0.014997459016740322, + 0.009239732287824154, + 0.008223461918532848, + 0.030101751908659935, + 0.009079087525606155, + 0.008391067385673523, + 0.00832333229482174, + 0.005137945059686899, + 0.021117080003023148, + 0.013775302097201347, + 0.00816346425563097, + 0.026876945048570633, + 0.008829435333609581, + 0.006539109628647566, + 0.005540824495255947, + 0.0056783296167850494, + 0.02594625949859619, + 0.004387349355965853, + 0.003217123681679368, + 0.017814451828598976, + 0.013590479269623756, + 0.006038649007678032, + 0.003809976624324918, + 0.007485632319003344, + 0.003925826400518417, + 0.005427282303571701, + 0.00643390417098999, + 0.005340247415006161, + 0.00559539208188653, + 0.006421846803277731, + 0.011810974217951298, + 0.006903637666255236, + 0.013211091049015522, + 0.02554711513221264, + 0.013778210617601871, + 0.023030102252960205, + 0.01816721446812153, + 0.013177018612623215, + 0.01664135977625847, + 0.02541070058941841, + 0.01964827999472618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.0021964015904814005 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018176885321736336 + ], + "opposite_logprob_tokens": [ + -4.018177032470703 + ], + "logprob_reward": -0.018176885321736336, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018176885321736336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42050880193710327, + 0.07590503245592117, + 0.06783458590507507, + 0.13473129272460938, + 0.013706173747777939, + 0.08358315378427505, + 0.05330102518200874, + 0.029253287240862846, + 0.07775187492370605, + 0.04342480003833771 + ], + "credit_map": [ + 0.05175551027059555, + 0.06536751240491867, + 0.04578498378396034, + 0.14581435918807983, + 0.023906603455543518, + 0.026291973888874054, + 0.027046505361795425, + 0.03454136475920677, + 0.008659600280225277, + 0.01014394499361515, + 0.01007517147809267, + 0.009797077625989914, + 0.011814618483185768, + 0.006231187842786312, + 0.010272097773849964, + 0.008911335840821266, + 0.009693666361272335, + 0.008889026008546352, + 0.008099056780338287, + 0.014613195322453976, + 0.00592466676607728, + 0.006448239553719759, + 0.006405196152627468, + 0.0077615417540073395, + 0.009923937730491161, + 0.005767108406871557, + 0.03662436828017235, + 0.005925836972892284, + 0.00458222022280097, + 0.00971710029989481, + 0.012863658368587494, + 0.0493270568549633, + 0.002463561249896884, + 0.0010467901593074203, + 0.001711002318188548, + 0.0010795838898047805, + 0.0017077544471248984, + 0.0023251569364219904, + 0.0015869553899392486, + 0.0017853688914328814, + 0.008052987046539783, + 0.004642676096409559, + 0.026047417894005775, + 0.008514925837516785, + 0.00679068174213171, + 0.006865368690341711, + 0.004781223367899656, + 0.017887873575091362, + 0.006173284258693457, + 0.0050580184906721115, + 0.01090964861214161, + 0.005784666631370783, + 0.004803352523595095, + 0.00438813678920269, + 0.006380265112966299, + 0.009803653694689274, + 0.002799204085022211, + 0.0026754611171782017, + 0.0032405825331807137, + 0.004322542808949947, + 0.004230356775224209, + 0.0034468919038772583, + 0.005403259303420782, + 0.003134987549856305, + 0.012121777981519699, + 0.010075710713863373, + 0.011646674945950508, + 0.011261343955993652, + 0.008046815171837807, + 0.008504500612616539, + 0.007909039035439491, + 0.008186014369130135, + 0.006857908796519041, + 0.003610416781157255, + 0.007848402485251427, + 0.00610687118023634, + 0.0032983694691210985, + 0.0034115593880414963, + 0.005587568506598473, + 0.0067037008702754974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.018176885321736336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide shot pans right, closely following a young man as he zips along a rough asphalt surface on a skateboard. Shot with a Steve McCurry-esque documentary quality, the dazzling afternoon sun saturates the image's colors, setting against a vibrant backdrop of a vibrant urban streetscape—a vibrant canvas of colorful storefronts, intertwined power lines, and blurred pedestrians. The young man, wearing a black beanie, with wisps of sweat-soaked blond hair plastered to his forehead, wears a baggy charcoal gray T-shirt, frayed dark overalls, and dusty skateboard shoes. His gaze is fixed intently ahead. As the skateboard rushed towards an intersection, he leaned forward, and then with an extremely neat and smooth movement, he decisively stepped one foot on the ground, lifted the other foot, and jumped lightly from the skateboard that was still rushing forward at high speed, landing steadily on the sidewalk with his knees slightly bent, while the skateboard continued to slide forward and gradually disappeared from the screen. The whole process was full of unpretentious street energy and youthful dynamism under the sharp light and shadow.", + "seed": 15797, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_3/index_6371_seed_15797_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.021760318148881196, + "vlm_logprob": -0.021760318148881196, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.021760318148881196, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05274571105837822 + ], + "opposite_logprob_tokens": [ + -4.677745819091797 + ], + "logprob_reward": -0.05274571105837822, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.05274571105837822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20356209576129913, + 0.04354378581047058, + 0.08115813881158829, + 0.04544474557042122, + 0.11155825853347778, + 0.31713488698005676, + 0.0958443284034729, + 0.03514664247632027, + 0.026304179802536964, + 0.04030295088887215 + ], + "credit_map": [ + 0.029928429052233696, + 0.035443264991045, + 0.026798931881785393, + 0.030308756977319717, + 0.01618940383195877, + 0.02453872747719288, + 0.01723637990653515, + 0.023118214681744576, + 0.0060595315881073475, + 0.0069815549068152905, + 0.004260707646608353, + 0.005406789015978575, + 0.005080288276076317, + 0.006984807550907135, + 0.00421576714143157, + 0.00455433689057827, + 0.02006744220852852, + 0.007975299842655659, + 0.02384619414806366, + 0.008761773817241192, + 0.003557485295459628, + 0.0056126173585653305, + 0.004987823776900768, + 0.006349498406052589, + 0.006900630425661802, + 0.007106503937393427, + 0.004215290304273367, + 0.004028359893709421, + 0.010315691120922565, + 0.004333777818828821, + 0.004759096074849367, + 0.003785394597798586, + 0.007140326779335737, + 0.036739833652973175, + 0.008376232348382473, + 0.039081692695617676, + 0.005876830779016018, + 0.002681934041902423, + 0.007621882949024439, + 0.004039520397782326, + 0.037958309054374695, + 0.09102948755025864, + 0.11786255240440369, + 0.019035324454307556, + 0.01595049351453781, + 0.0074793389067053795, + 0.014499185606837273, + 0.013320195488631725, + 0.010810286737978458, + 0.008993659168481827, + 0.009921789169311523, + 0.007029146421700716, + 0.02415965311229229, + 0.013382591307163239, + 0.014208624139428139, + 0.007338578347116709, + 0.004448010586202145, + 0.0038880459032952785, + 0.0025088617112487555, + 0.003200360108166933, + 0.008041221648454666, + 0.0036444268189370632, + 0.005376663524657488, + 0.004039050079882145, + 0.0029859805945307016, + 0.0030896065291017294, + 0.0023125961888581514, + 0.0034891851246356964, + 0.003400238463655114, + 0.00326576572842896, + 0.00399111257866025, + 0.003769694361835718, + 0.004274785052984953, + 0.004640292376279831, + 0.00355646968819201, + 0.003710969351232052, + 0.005780171602964401, + 0.007215477526187897, + 0.005504673812538385, + 0.0056201135739684105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.05274571105837822 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635246515274048 + ], + "opposite_logprob_tokens": [ + -3.6513524055480957 + ], + "logprob_reward": -0.02635246515274048, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635246515274048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10532442480325699, + 0.03306017816066742, + 0.08399612456560135, + 0.06209605932235718, + 0.21715696156024933, + 0.4016096293926239, + 0.02580069936811924, + 0.017100241035223007, + 0.03232560679316521, + 0.021530088037252426 + ], + "credit_map": [ + 0.013942074961960316, + 0.018788272514939308, + 0.011585320346057415, + 0.011953817680478096, + 0.016753118485212326, + 0.00865216925740242, + 0.012592755258083344, + 0.01105689350515604, + 0.004042241256684065, + 0.0033913911320269108, + 0.0033647259697318077, + 0.003861851990222931, + 0.005583996884524822, + 0.005385716911405325, + 0.0036798259243369102, + 0.0037504308857023716, + 0.016912197694182396, + 0.008596093393862247, + 0.014582342468202114, + 0.008834436535835266, + 0.007002621423453093, + 0.010736756958067417, + 0.0077202883549034595, + 0.009611384011805058, + 0.00727038225159049, + 0.010257801041007042, + 0.0051686824299395084, + 0.0061644879169762135, + 0.013408923521637917, + 0.005971903447061777, + 0.008437427692115307, + 0.00541645148769021, + 0.013423899188637733, + 0.07420241087675095, + 0.0164914783090353, + 0.0782332494854927, + 0.008440503850579262, + 0.004819595720618963, + 0.015024961903691292, + 0.006520857568830252, + 0.05348256230354309, + 0.108477883040905, + 0.15303483605384827, + 0.02293045073747635, + 0.021076004952192307, + 0.008508388884365559, + 0.01802712120115757, + 0.016072364524006844, + 0.0027560272719711065, + 0.0028349633794277906, + 0.0040293470956385136, + 0.003267566440626979, + 0.0030730890575796366, + 0.003117998130619526, + 0.0036958118434995413, + 0.0030258954502642155, + 0.0019751624204218388, + 0.0018112602410838008, + 0.0013624976854771376, + 0.001275748130865395, + 0.001913924003019929, + 0.0017267855582758784, + 0.0040111420676112175, + 0.0030237222090363503, + 0.004087273962795734, + 0.0028601104859262705, + 0.0024422076530754566, + 0.0038681374862790108, + 0.0038081512320786715, + 0.004062199965119362, + 0.004804104100912809, + 0.006393419113010168, + 0.002691320376470685, + 0.0034702469129115343, + 0.0036973331589251757, + 0.002708443906158209, + 0.0033318179193884134, + 0.0025248080492019653, + 0.0018697605701163411, + 0.0012363563291728497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.02635246515274048 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535824440419674 + ], + "opposite_logprob_tokens": [ + -4.38753604888916 + ], + "logprob_reward": -0.012535824440419674, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535824440419674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1663387268781662, + 0.06677760183811188, + 0.06763259321451187, + 0.06799737364053726, + 0.03280583769083023, + 0.33061110973358154, + 0.08844456821680069, + 0.052901044487953186, + 0.08016951382160187, + 0.046321604400873184 + ], + "credit_map": [ + 0.021776866167783737, + 0.03274291381239891, + 0.022236062213778496, + 0.016441768035292625, + 0.018884209915995598, + 0.017783241346478462, + 0.02196633070707321, + 0.014507331885397434, + 0.00930788367986679, + 0.007719660177826881, + 0.00992555357515812, + 0.0140501968562603, + 0.007805701345205307, + 0.007306802552193403, + 0.006318173836916685, + 0.004343634471297264, + 0.015633124858140945, + 0.0057647740468382835, + 0.018839320167899132, + 0.005976865068078041, + 0.0044935536570847034, + 0.006223028525710106, + 0.004233637358993292, + 0.006468288134783506, + 0.009857622906565666, + 0.009537887759506702, + 0.006081502418965101, + 0.008770696818828583, + 0.009104211814701557, + 0.0089454660192132, + 0.009009976871311665, + 0.006690013222396374, + 0.003657837864011526, + 0.004166289232671261, + 0.0030916209798306227, + 0.005120761226862669, + 0.004909707698971033, + 0.0021234045270830393, + 0.004230555612593889, + 0.005505660083144903, + 0.033590689301490784, + 0.09570468217134476, + 0.11629147827625275, + 0.01988818682730198, + 0.020349809899926186, + 0.011473369784653187, + 0.018680505454540253, + 0.01463242992758751, + 0.009459668770432472, + 0.00898959394544363, + 0.01060738880187273, + 0.009235742501914501, + 0.020349442958831787, + 0.009840971790254116, + 0.014022935181856155, + 0.005938823334872723, + 0.005731741897761822, + 0.00476409588009119, + 0.004096254240721464, + 0.003492196789011359, + 0.005454122554510832, + 0.004563225898891687, + 0.013980735093355179, + 0.010818676091730595, + 0.012702027335762978, + 0.007565371226519346, + 0.008849247358739376, + 0.010438768193125725, + 0.010746391490101814, + 0.009710718877613544, + 0.011152984574437141, + 0.009004011750221252, + 0.005362164694815874, + 0.004229156766086817, + 0.004595587495714426, + 0.0044809929095208645, + 0.007835503667593002, + 0.00895738136023283, + 0.004763249773532152, + 0.0060975695960223675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.012535824440419674 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634591285139322 + ], + "opposite_logprob_tokens": [ + -5.379634380340576 + ], + "logprob_reward": -0.004634591285139322, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634591285139322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10766071826219559, + 0.03133605420589447, + 0.22151730954647064, + 0.20114180445671082, + 0.12450098991394043, + 0.1749722957611084, + 0.052534252405166626, + 0.020405860617756844, + 0.03192204236984253, + 0.03400868922472 + ], + "credit_map": [ + 0.015468175522983074, + 0.01444346271455288, + 0.014921337366104126, + 0.012007422745227814, + 0.011982571333646774, + 0.012042243964970112, + 0.014765309169888496, + 0.012030191719532013, + 0.005261184647679329, + 0.007477069739252329, + 0.0023971134796738625, + 0.003546522930264473, + 0.004391776863485575, + 0.00350490165874362, + 0.0022542444057762623, + 0.0025032376870512962, + 0.06255429238080978, + 0.014520685188472271, + 0.07273232191801071, + 0.02654201164841652, + 0.010470597073435783, + 0.011116787791252136, + 0.009649597108364105, + 0.013931016437709332, + 0.03649013116955757, + 0.03933858871459961, + 0.021059731021523476, + 0.023228852078318596, + 0.024605417624115944, + 0.019907871261239052, + 0.020653005689382553, + 0.015858186408877373, + 0.009532093070447445, + 0.03921981155872345, + 0.009566090069711208, + 0.04198860749602318, + 0.0053860219195485115, + 0.0033351785968989134, + 0.008591713383793831, + 0.00688145961612463, + 0.017519524320960045, + 0.05620158091187477, + 0.06139491870999336, + 0.011289200745522976, + 0.009469582699239254, + 0.003855439368635416, + 0.007741075474768877, + 0.007500973995774984, + 0.00453856261447072, + 0.007693699561059475, + 0.009152298793196678, + 0.005581451579928398, + 0.009187719784677029, + 0.005669800564646721, + 0.006769082043319941, + 0.003941634204238653, + 0.0026292637921869755, + 0.001989529235288501, + 0.0019104633247479796, + 0.001585790072567761, + 0.0027416578959673643, + 0.0019299424020573497, + 0.004088965710252523, + 0.003530245739966631, + 0.0036391299217939377, + 0.003897103713825345, + 0.003888218430802226, + 0.0043023559264838696, + 0.00402133259922266, + 0.0036870248150080442, + 0.0039194799028337, + 0.004567396827042103, + 0.005392169114202261, + 0.00403050659224391, + 0.003552884329110384, + 0.003736309241503477, + 0.005648210644721985, + 0.005078996066004038, + 0.0034959588665515184, + 0.0030736499466001987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.004634591285139322 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle and does it pan right while following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man wearing a black beanie, baggy charcoal gray T-shirt, and frayed dark overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the young man jumping from a moving skateboard onto a sidewalk while the skateboard continues sliding forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant urban streetscape featuring colorful storefronts, intertwined power lines, and blurred pedestrians?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dazzling afternoon sun creating sharp light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253299880772829 + ], + "opposite_logprob_tokens": [ + -4.387533187866211 + ], + "logprob_reward": -0.01253299880772829, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253299880772829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15797452628612518, + 0.10620784759521484, + 0.06406264007091522, + 0.05813831090927124, + 0.12896068394184113, + 0.11981981247663498, + 0.034433647990226746, + 0.10117055475711823, + 0.045096609741449356, + 0.18413543701171875 + ], + "credit_map": [ + 0.020638050511479378, + 0.023332443088293076, + 0.028402645140886307, + 0.022580066695809364, + 0.017942072823643684, + 0.014462597668170929, + 0.014758002012968063, + 0.015858624130487442, + 0.017224641516804695, + 0.027311153709888458, + 0.008037424646317959, + 0.014087467454373837, + 0.011592816561460495, + 0.011671901680529118, + 0.00805865228176117, + 0.008223777636885643, + 0.012859589420258999, + 0.009450568817555904, + 0.015249987132847309, + 0.00871223397552967, + 0.00494482321664691, + 0.004176024813205004, + 0.00500965490937233, + 0.0036597561556845903, + 0.011538947932422161, + 0.011722040362656116, + 0.005567876622080803, + 0.005814431235194206, + 0.007583421654999256, + 0.0040811276994645596, + 0.006742726545780897, + 0.005087732803076506, + 0.0077156624756753445, + 0.04441063478589058, + 0.010153418406844139, + 0.04541603848338127, + 0.005102521274238825, + 0.0028966276440769434, + 0.009028159081935883, + 0.00423760712146759, + 0.017439302057027817, + 0.030174776911735535, + 0.04487255960702896, + 0.006863379385322332, + 0.006994433235377073, + 0.002641448751091957, + 0.005613543093204498, + 0.005220354069024324, + 0.0034645108971744776, + 0.005005123559385538, + 0.0057425303384661674, + 0.0060610356740653515, + 0.0038088213186711073, + 0.0035492684692144394, + 0.0037784615997225046, + 0.0030238921754062176, + 0.013776874169707298, + 0.007777519524097443, + 0.0060942810960114, + 0.007525146007537842, + 0.011829886585474014, + 0.0070911129005253315, + 0.02876594476401806, + 0.01830977387726307, + 0.004390705842524767, + 0.0031429575756192207, + 0.004002169240266085, + 0.003168724710121751, + 0.0033688684925436974, + 0.006833211984485388, + 0.009831435047090054, + 0.010358531959354877, + 0.015861282125115395, + 0.015705471858382225, + 0.017833495512604713, + 0.02240370772778988, + 0.029009155929088593, + 0.046658895909786224, + 0.017916467040777206, + 0.018746934831142426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.01253299880772829 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_4/index_tensor([5482, 5482])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_4/index_tensor([5482, 5482])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bbc336f70aa7fc2dc191763afa69b9bcc63ca956 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_4/index_tensor([5482, 5482])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The documentary-style photography begins with a horizontal medium shot that slowly pans rightward, panning across a row of honor guard soldiers standing erect, statuesquely erect. Their uniforms are clad in crisp, dark wool, with crisp collars and neatly arranged epaulettes. Shining brass buttons on their chests gleam faintly in the soft, slightly cool morning light. Each soldier's eyes are tightly closed, their jaws set in a resolute facial expression, seemingly shutting out all emotion. Yet, a single, close-up detail—their uncontrollable, slightly trembling lower lips—silently betrays the surging excitement and sense of honor beneath their iron discipline. With the vast, slightly out-of-focus square in the background, the entire image, captured in exceptional clarity, exudes an atmosphere of solemnity, intense focus, and inner tension, creating a striking and moving contrast between outward composure and inner turmoil.", + "seed": 24827, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_4/index_5482_seed_24827_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15901621282100678, + "vlm_logprob": -0.15901621282100678, + "vlm_yes_no_margin": 2.300000011920929, + "vlm_token_credit": -0.15901621282100678, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08234069496393204 + ], + "opposite_logprob_tokens": [ + -2.707340717315674 + ], + "logprob_reward": -0.08234069496393204, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08234069496393204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2754117548465729, + 0.06170806288719177, + 0.08995899558067322, + 0.14989374577999115, + 0.09658286720514297, + 0.09754297882318497, + 0.05928312614560127, + 0.114064060151577, + 0.03568759188055992, + 0.019866786897182465 + ], + "credit_map": [ + 0.023735500872135162, + 0.05960040166974068, + 0.020999932661652565, + 0.01586386188864708, + 0.021810853853821754, + 0.012058052234351635, + 0.07617826759815216, + 0.045164886862039566, + 0.005312560126185417, + 0.0073855724185705185, + 0.01180394645780325, + 0.006008865311741829, + 0.006880492903292179, + 0.00735860550776124, + 0.005716195795685053, + 0.011241826228797436, + 0.009178180247545242, + 0.009692917577922344, + 0.009395123459398746, + 0.015055335126817226, + 0.01391693577170372, + 0.011656486429274082, + 0.011576958000659943, + 0.009487071074545383, + 0.01708100363612175, + 0.020844532176852226, + 0.02075018733739853, + 0.02378624491393566, + 0.013460243120789528, + 0.015668760985136032, + 0.02025609277188778, + 0.018046682700514793, + 0.01158471591770649, + 0.012384899891912937, + 0.00886014848947525, + 0.012770997360348701, + 0.015150763094425201, + 0.01483320165425539, + 0.012555968947708607, + 0.008442167192697525, + 0.009346703067421913, + 0.01083785854279995, + 0.011882076039910316, + 0.011015122756361961, + 0.01221715472638607, + 0.011830195784568787, + 0.020555339753627777, + 0.009858525358140469, + 0.00466056726872921, + 0.005885138642042875, + 0.009853431954979897, + 0.004228997975587845, + 0.005305539816617966, + 0.007335476577281952, + 0.008638381026685238, + 0.013375593349337578, + 0.013835144229233265, + 0.013616877608001232, + 0.014743720181286335, + 0.0103209363296628, + 0.010775119066238403, + 0.02277246303856373, + 0.016629014164209366, + 0.011370781809091568, + 0.004147132392972708, + 0.005724555347114801, + 0.0037795475218445063, + 0.0037570649292320013, + 0.0038747519720345736, + 0.004849488381296396, + 0.0032328942324966192, + 0.006322159431874752, + 0.003244017018005252, + 0.003121323185041547, + 0.002608160488307476, + 0.002062475075945258, + 0.002137536648660898, + 0.0029453602619469166, + 0.0021357133518904448, + 0.0016122004017233849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.08234069496393204 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011104568839073181 + ], + "opposite_logprob_tokens": [ + -4.511104583740234 + ], + "logprob_reward": -0.011104568839073181, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011104568839073181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15022799372673035, + 0.07520692050457001, + 0.07844024896621704, + 0.06525704264640808, + 0.03926829993724823, + 0.06217391416430473, + 0.058693621307611465, + 0.284982293844223, + 0.14865319430828094, + 0.037096429616212845 + ], + "credit_map": [ + 0.00941450335085392, + 0.026316633448004723, + 0.01178074162453413, + 0.0109427934512496, + 0.012207760475575924, + 0.0071032228879630566, + 0.04256775975227356, + 0.029894579201936722, + 0.005239318124949932, + 0.00974026508629322, + 0.00948422309011221, + 0.00902547501027584, + 0.011233015917241573, + 0.012757450342178345, + 0.009689707309007645, + 0.008037472143769264, + 0.011950068175792694, + 0.010971879586577415, + 0.008371245115995407, + 0.011841668747365475, + 0.008782286196947098, + 0.010518603026866913, + 0.008786187507212162, + 0.007218305487185717, + 0.007178501226007938, + 0.006670328322798014, + 0.009026089683175087, + 0.011683453805744648, + 0.006008990574628115, + 0.00512162409722805, + 0.010313961654901505, + 0.009254086762666702, + 0.006003264803439379, + 0.0043764980509877205, + 0.004171514417976141, + 0.004217403009533882, + 0.007668249774724245, + 0.004061967134475708, + 0.004416877403855324, + 0.004352526273578405, + 0.0063636936247348785, + 0.01027581188827753, + 0.007237907033413649, + 0.007683205883949995, + 0.008214232511818409, + 0.006899677217006683, + 0.00972768198698759, + 0.00577170867472887, + 0.006713229697197676, + 0.006811340339481831, + 0.009021226316690445, + 0.005499651189893484, + 0.009167994372546673, + 0.004964086227118969, + 0.010687706060707569, + 0.005828389432281256, + 0.03397667035460472, + 0.03313909098505974, + 0.03798628970980644, + 0.024437228217720985, + 0.025191068649291992, + 0.05741880461573601, + 0.04567534476518631, + 0.02715780958533287, + 0.014483400620520115, + 0.023741085082292557, + 0.018887771293520927, + 0.021651513874530792, + 0.017840011045336723, + 0.019607597962021828, + 0.015502192080020905, + 0.01693962700664997, + 0.005881407763808966, + 0.005108810029923916, + 0.006588899064809084, + 0.004002433270215988, + 0.004199758637696505, + 0.004529590718448162, + 0.0035789129324257374, + 0.0032066158019006252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.011104568839073181 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20163747668266296 + ], + "opposite_logprob_tokens": [ + -1.7016375064849854 + ], + "logprob_reward": -0.20163747668266296, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20163747668266296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0818265900015831, + 0.08243657648563385, + 0.21469300985336304, + 0.18929073214530945, + 0.030977152287960052, + 0.04446369782090187, + 0.04514042288064957, + 0.21560853719711304, + 0.04295111447572708, + 0.052612125873565674 + ], + "credit_map": [ + 0.007648600731045008, + 0.006796945817768574, + 0.012874756008386612, + 0.015394522808492184, + 0.01178782805800438, + 0.008956996724009514, + 0.007587500847876072, + 0.01077944878488779, + 0.005351417697966099, + 0.01026355940848589, + 0.011778630316257477, + 0.007266846019774675, + 0.010562502779066563, + 0.01941593363881111, + 0.008262285962700844, + 0.009535406716167927, + 0.025242948904633522, + 0.028502942994236946, + 0.025081586092710495, + 0.03732183575630188, + 0.022541983053088188, + 0.030461281538009644, + 0.025313496589660645, + 0.020226934924721718, + 0.020306339487433434, + 0.021939635276794434, + 0.02699415758252144, + 0.03380194678902626, + 0.016656382009387016, + 0.01965199038386345, + 0.028074322268366814, + 0.021865960210561752, + 0.00392990093678236, + 0.004814661107957363, + 0.004835344385355711, + 0.003654223633930087, + 0.004514756146818399, + 0.0029598071705549955, + 0.003265058621764183, + 0.003003400517627597, + 0.0053343097679317, + 0.006464251317083836, + 0.006377656944096088, + 0.00578691391274333, + 0.004841149784624577, + 0.004510207567363977, + 0.007528301794081926, + 0.0036209120880812407, + 0.006213287822902203, + 0.0039404891431331635, + 0.006609691772609949, + 0.004465784411877394, + 0.006643760483711958, + 0.004482609685510397, + 0.0056338622234761715, + 0.007150944322347641, + 0.02596493624150753, + 0.023811722174286842, + 0.030309153720736504, + 0.019531430676579475, + 0.019085746258497238, + 0.042924556881189346, + 0.03453199565410614, + 0.019449027255177498, + 0.004312905017286539, + 0.006671045441180468, + 0.004519445821642876, + 0.004799286834895611, + 0.004195860121399164, + 0.007614381145685911, + 0.004592292942106724, + 0.006245898548513651, + 0.01095217652618885, + 0.006881270557641983, + 0.0072904969565570354, + 0.003644114127382636, + 0.005911334417760372, + 0.0055895010009408, + 0.004135829862207174, + 0.00820741057395935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.20163747668266296 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293607205152512 + ], + "opposite_logprob_tokens": [ + -2.237936019897461 + ], + "logprob_reward": -0.11293607205152512, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293607205152512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09909780323505402, + 0.14820496737957, + 0.113688625395298, + 0.0553838275372982, + 0.04215934872627258, + 0.07429154217243195, + 0.039305660873651505, + 0.33699727058410645, + 0.05023385211825371, + 0.04063707962632179 + ], + "credit_map": [ + 0.007519449573010206, + 0.007737181149423122, + 0.016731955111026764, + 0.02080116979777813, + 0.013523093424737453, + 0.009720217436552048, + 0.009638889692723751, + 0.013425841927528381, + 0.01041185762733221, + 0.015898145735263824, + 0.031716492027044296, + 0.017729684710502625, + 0.011451099067926407, + 0.017266998067498207, + 0.015063120983541012, + 0.028667563572525978, + 0.01277101319283247, + 0.014982308261096478, + 0.0109855392947793, + 0.02188759669661522, + 0.014018731191754341, + 0.01687639392912388, + 0.011553323827683926, + 0.01061371061950922, + 0.007381213363260031, + 0.0063071358017623425, + 0.008307560347020626, + 0.0074847182258963585, + 0.005177047569304705, + 0.006670932751148939, + 0.007025177590548992, + 0.00703004514798522, + 0.006427134852856398, + 0.005856242496520281, + 0.0046408218331635, + 0.004989513661712408, + 0.005995649844408035, + 0.004404073115438223, + 0.005542277358472347, + 0.004303636495023966, + 0.008785582147538662, + 0.012627432122826576, + 0.009817603975534439, + 0.009003154002130032, + 0.009694157168269157, + 0.008544034324586391, + 0.009458807297050953, + 0.006360769737511873, + 0.0048753940500319, + 0.004122605547308922, + 0.004836482461541891, + 0.0030001269187778234, + 0.004557818174362183, + 0.004893803503364325, + 0.0063080498948693275, + 0.006711380090564489, + 0.04157441481947899, + 0.040776364505290985, + 0.04097072407603264, + 0.028827616944909096, + 0.03169216215610504, + 0.06808063387870789, + 0.0502849817276001, + 0.03479037806391716, + 0.005895108450204134, + 0.006019876338541508, + 0.005539765115827322, + 0.005833786446601152, + 0.006036021746695042, + 0.0078115761280059814, + 0.006483840756118298, + 0.00661387899890542, + 0.006108802743256092, + 0.00444443617016077, + 0.0063278512097895145, + 0.005202082917094231, + 0.004300299100577831, + 0.005539008881896734, + 0.0036635855212807655, + 0.005051013082265854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 2 + ], + "selected_reward": -0.11293607205152512 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870622515678406 + ], + "opposite_logprob_tokens": [ + -1.1370623111724854 + ], + "logprob_reward": -0.3870622515678406, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3870622515678406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2514778673648834, + 0.08436478674411774, + 0.02236589603126049, + 0.021699951961636543, + 0.03135356307029724, + 0.020482588559389114, + 0.027546105906367302, + 0.48462581634521484, + 0.03917865455150604, + 0.016904791817069054 + ], + "credit_map": [ + 0.017620155587792397, + 0.05184538662433624, + 0.020059708505868912, + 0.020007506012916565, + 0.02042454108595848, + 0.010993131436407566, + 0.06830692291259766, + 0.04222051054239273, + 0.005479005631059408, + 0.008610193617641926, + 0.010392631404101849, + 0.005354276392608881, + 0.01768738590180874, + 0.020704517140984535, + 0.006780307274311781, + 0.009356468915939331, + 0.002496020868420601, + 0.0017456372734159231, + 0.0030326596461236477, + 0.0037193289026618004, + 0.00202841660939157, + 0.0025451339315623045, + 0.0034481261391192675, + 0.0033505719620734453, + 0.002868817886337638, + 0.0023115493822842836, + 0.0038204416632652283, + 0.002876326674595475, + 0.0017508757300674915, + 0.002240503905341029, + 0.002692085923627019, + 0.0031393517274409533, + 0.003923261072486639, + 0.0043147471733391285, + 0.003553478978574276, + 0.0033024917356669903, + 0.005193384364247322, + 0.004918183200061321, + 0.0031837644055485725, + 0.0029642547015100718, + 0.0017368837725371122, + 0.0017559699481353164, + 0.002690558321774006, + 0.002532700542360544, + 0.002952815033495426, + 0.0024349093437194824, + 0.0026116434019058943, + 0.0037671076133847237, + 0.0030105765908956528, + 0.003552195616066456, + 0.0033847817685455084, + 0.0021491735242307186, + 0.0036701096687465906, + 0.003589621977880597, + 0.004045231267809868, + 0.00414441479369998, + 0.06017868593335152, + 0.05773210525512695, + 0.05934075638651848, + 0.04211878031492233, + 0.046267319470644, + 0.09820803254842758, + 0.07174448668956757, + 0.04903562366962433, + 0.004029447212815285, + 0.006284826435148716, + 0.004404837265610695, + 0.005037476774305105, + 0.004681330174207687, + 0.00493244081735611, + 0.004627324175089598, + 0.005180974956601858, + 0.0021661468781530857, + 0.0018513351678848267, + 0.0022975432220846415, + 0.0022223920095711946, + 0.002106422558426857, + 0.001854304107837379, + 0.0016133280005306005, + 0.002793320454657078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.3870622515678406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The documentary-style photography begins with a horizontal medium shot that slowly pans rightward, panning across a row of honor guard soldiers standing erect, statuesquely erect. Their uniforms are clad in crisp, dark wool, with crisp collars and neatly arranged epaulettes. Shining brass buttons on their chests gleam faintly in the soft, slightly cool morning light. Each soldier's eyes are tightly closed, their jaws set in a resolute facial expression, seemingly shutting out all emotion. Yet, a single, close-up detail—their uncontrollable, slightly trembling lower lips—silently betrays the surging excitement and sense of honor beneath their iron discipline. With the vast, slightly out-of-focus square in the background, the entire image, captured in exceptional clarity, exudes an atmosphere of solemnity, intense focus, and inner tension, creating a striking and moving contrast between outward composure and inner turmoil.", + "seed": 24828, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_4/index_5482_seed_24828_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16369001977145672, + "vlm_logprob": -0.16369001977145672, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.16369001977145672, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.036632437258958817 + ], + "opposite_logprob_tokens": [ + -3.911632537841797 + ], + "logprob_reward": -0.036632437258958817, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.036632437258958817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12830325961112976, + 0.05801684781908989, + 0.041102148592472076, + 0.11174976080656052, + 0.03594258427619934, + 0.23452098667621613, + 0.12190083414316177, + 0.10450439155101776, + 0.14482589066028595, + 0.019133353605866432 + ], + "credit_map": [ + 0.010674151591956615, + 0.01546232495456934, + 0.01342769805341959, + 0.013331732712686062, + 0.0130305215716362, + 0.02091948129236698, + 0.01719709113240242, + 0.02426024153828621, + 0.005545215215533972, + 0.008168897591531277, + 0.00651930645108223, + 0.00798865593969822, + 0.005011378321796656, + 0.009647378697991371, + 0.008815241046249866, + 0.006320767104625702, + 0.0042315516620874405, + 0.004195962566882372, + 0.006939810235053301, + 0.00917884148657322, + 0.0031374222598969936, + 0.0031586477998644114, + 0.006205128971487284, + 0.004054776858538389, + 0.009020406752824783, + 0.024361852556467056, + 0.006318078376352787, + 0.012923060916364193, + 0.027137409895658493, + 0.009550001472234726, + 0.012113479897379875, + 0.010325462557375431, + 0.00496981106698513, + 0.005862944293767214, + 0.00296357786282897, + 0.005127262324094772, + 0.0032844915986061096, + 0.006119825411587954, + 0.003914795815944672, + 0.003699872875586152, + 0.029961327090859413, + 0.03965139761567116, + 0.02701219916343689, + 0.03106757625937462, + 0.04213691130280495, + 0.019921794533729553, + 0.02919633314013481, + 0.015573425218462944, + 0.01321764849126339, + 0.030185241252183914, + 0.009276475757360458, + 0.008912053890526295, + 0.021638255566358566, + 0.022201821208000183, + 0.009070578962564468, + 0.007398754358291626, + 0.01022915355861187, + 0.009462302550673485, + 0.015850912779569626, + 0.01356947049498558, + 0.011626278050243855, + 0.019403988495469093, + 0.01240356545895338, + 0.011958715505897999, + 0.019684474915266037, + 0.013815262354910374, + 0.02094634249806404, + 0.02323165535926819, + 0.02564547210931778, + 0.009031872265040874, + 0.02435644343495369, + 0.008114357478916645, + 0.0033371697645634413, + 0.002421621698886156, + 0.002167575992643833, + 0.001953201601281762, + 0.0029332665726542473, + 0.003401593305170536, + 0.0014648271026089787, + 0.0014540954725816846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.036632437258958817 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007655567489564419 + ], + "opposite_logprob_tokens": [ + -4.882655620574951 + ], + "logprob_reward": -0.007655567489564419, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007655567489564419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12385809421539307, + 0.149973064661026, + 0.09177849441766739, + 0.03813647851347923, + 0.13048556447029114, + 0.0722973495721817, + 0.08144727349281311, + 0.17146673798561096, + 0.08378659188747406, + 0.05677034333348274 + ], + "credit_map": [ + 0.008248139172792435, + 0.011464709416031837, + 0.017922526225447655, + 0.021033303812146187, + 0.013286484405398369, + 0.011197189800441265, + 0.01670779660344124, + 0.023997940123081207, + 0.012240526266396046, + 0.012420659884810448, + 0.02057843841612339, + 0.02171473577618599, + 0.013066940940916538, + 0.024934930726885796, + 0.024818802252411842, + 0.020198021084070206, + 0.006824356969445944, + 0.008726154454052448, + 0.02048218995332718, + 0.0058618634939193726, + 0.00910193007439375, + 0.0064230263233184814, + 0.027692491188645363, + 0.0066664800979197025, + 0.003071557031944394, + 0.0033957287669181824, + 0.003110104938969016, + 0.005912716034799814, + 0.005269038490951061, + 0.00632642675191164, + 0.00640037702396512, + 0.0046505313366651535, + 0.009972797706723213, + 0.016418740153312683, + 0.017723796889185905, + 0.021682661026716232, + 0.014175747521221638, + 0.01952703483402729, + 0.024655161425471306, + 0.006329610478132963, + 0.011365426704287529, + 0.013219993561506271, + 0.007007744628936052, + 0.009520480409264565, + 0.010137866251170635, + 0.006858783308416605, + 0.00907661858946085, + 0.005110437050461769, + 0.008055051788687706, + 0.020163865759968758, + 0.006678814999759197, + 0.006263154558837414, + 0.012966988608241081, + 0.014831168577075005, + 0.006063087843358517, + 0.00642513670027256, + 0.021130437031388283, + 0.02168881706893444, + 0.029654471203684807, + 0.022813547402620316, + 0.03284642472863197, + 0.018919065594673157, + 0.011939500458538532, + 0.012474467046558857, + 0.008919141255319118, + 0.009487537667155266, + 0.01171637512743473, + 0.012140809558331966, + 0.015545177273452282, + 0.008056706748902798, + 0.012241399846971035, + 0.00567944161593914, + 0.006858866196125746, + 0.009299956262111664, + 0.005776659585535526, + 0.008078347891569138, + 0.006939580664038658, + 0.009922885335981846, + 0.004303578287363052, + 0.005590468179434538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.007655567489564419 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6329205632209778 + ], + "opposite_logprob_tokens": [ + -0.7579205632209778 + ], + "logprob_reward": -0.6329205632209778, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6329205632209778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08994346857070923, + 0.10900648683309555, + 0.07520310580730438, + 0.10319118946790695, + 0.03751054033637047, + 0.06039290875196457, + 0.08006712049245834, + 0.08374238014221191, + 0.2764790952205658, + 0.0844637006521225 + ], + "credit_map": [ + 0.008545823395252228, + 0.008961232379078865, + 0.011461352929472923, + 0.01081343274563551, + 0.008929048664867878, + 0.015618543140590191, + 0.011022676713764668, + 0.014591355808079243, + 0.009639587253332138, + 0.011346641927957535, + 0.014796005561947823, + 0.012494846247136593, + 0.012143664993345737, + 0.021587645635008812, + 0.013722450472414494, + 0.013275650329887867, + 0.0074287657625973225, + 0.011093301698565483, + 0.011277555488049984, + 0.011409448459744453, + 0.006086974870413542, + 0.006955756805837154, + 0.013370199128985405, + 0.007581107784062624, + 0.008641846477985382, + 0.028741834685206413, + 0.010384414345026016, + 0.010832106694579124, + 0.018891725689172745, + 0.009112105704843998, + 0.011461131274700165, + 0.005126022268086672, + 0.005024333950132132, + 0.004338311497122049, + 0.00472007691860199, + 0.006391651928424835, + 0.004407361149787903, + 0.005549142602831125, + 0.005331138614565134, + 0.0017485242569819093, + 0.008289482444524765, + 0.009985691867768764, + 0.00722245778888464, + 0.007274148985743523, + 0.010237619280815125, + 0.007154935970902443, + 0.006540532689541578, + 0.003688037395477295, + 0.009355618618428707, + 0.016929717734456062, + 0.0067666443064808846, + 0.007894666865468025, + 0.010265086777508259, + 0.014434567652642727, + 0.005927930120378733, + 0.008492894470691681, + 0.011658585630357265, + 0.013366150669753551, + 0.010737836360931396, + 0.010657811537384987, + 0.011303791776299477, + 0.014787633903324604, + 0.0036694935988634825, + 0.007561083883047104, + 0.0392056368291378, + 0.030992304906249046, + 0.0399954691529274, + 0.04453559219837189, + 0.04616791382431984, + 0.016067329794168472, + 0.04536138474941254, + 0.014153456315398216, + 0.009559337049722672, + 0.009012401103973389, + 0.011873407289385796, + 0.012184146791696548, + 0.017216330394148827, + 0.01016957126557827, + 0.008807259611785412, + 0.005641252733767033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.6329205632209778 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704752385616302 + ], + "opposite_logprob_tokens": [ + -2.127047538757324 + ], + "logprob_reward": -0.12704752385616302, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704752385616302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08743024617433548, + 0.13813523948192596, + 0.10432924330234528, + 0.1242126002907753, + 0.03699849173426628, + 0.05217503383755684, + 0.08725287020206451, + 0.08355345577001572, + 0.1838941127061844, + 0.10201872885227203 + ], + "credit_map": [ + 0.006852529477328062, + 0.0065455022267997265, + 0.0145452581346035, + 0.011908200569450855, + 0.011204611510038376, + 0.008704961277544498, + 0.011793593876063824, + 0.01587558723986149, + 0.011794697493314743, + 0.01083139143884182, + 0.02381323277950287, + 0.01627369597554207, + 0.01303223054856062, + 0.021494554355740547, + 0.022827601060271263, + 0.018067840486764908, + 0.008101660758256912, + 0.012855974026024342, + 0.02007913403213024, + 0.01483793556690216, + 0.007585739251226187, + 0.007572739385068417, + 0.02533573843538761, + 0.007960319519042969, + 0.009622619487345219, + 0.03524358570575714, + 0.007502358872443438, + 0.00987028144299984, + 0.030502792447805405, + 0.011709777638316154, + 0.010762432590126991, + 0.008998746983706951, + 0.00506944116204977, + 0.006311076693236828, + 0.003965180367231369, + 0.004953421652317047, + 0.004877434112131596, + 0.004401303827762604, + 0.004444602411240339, + 0.0029760294128209352, + 0.006260312162339687, + 0.008809220045804977, + 0.006955632474273443, + 0.006625934969633818, + 0.0065590799786150455, + 0.00591800594702363, + 0.006322422530502081, + 0.004724425263702869, + 0.011047357693314552, + 0.021700050681829453, + 0.008307830430567265, + 0.0073471637442708015, + 0.013608471490442753, + 0.012140464037656784, + 0.005973330698907375, + 0.007128206081688404, + 0.010312969796359539, + 0.008242661133408546, + 0.013063556514680386, + 0.010316700674593449, + 0.0197274312376976, + 0.011772754602134228, + 0.004898614715784788, + 0.005218766164034605, + 0.023131240159273148, + 0.017836999148130417, + 0.026763910427689552, + 0.02836424484848976, + 0.037246208637952805, + 0.012349273078143597, + 0.028738198801875114, + 0.009464028291404247, + 0.010479027405381203, + 0.013689868152141571, + 0.010979345068335533, + 0.011179500259459019, + 0.01590527780354023, + 0.018375849351286888, + 0.011163418181240559, + 0.01024644635617733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.12704752385616302 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194007031619549 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014194007031619549, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194007031619549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2295609712600708, + 0.16427108645439148, + 0.11786999553442001, + 0.09149882942438126, + 0.15684649348258972, + 0.03375816345214844, + 0.044045887887477875, + 0.07863743603229523, + 0.04111332818865776, + 0.04239777848124504 + ], + "credit_map": [ + 0.01553453877568245, + 0.014803440310060978, + 0.041898008435964584, + 0.03718941658735275, + 0.02063646726310253, + 0.02691895328462124, + 0.03817442059516907, + 0.03440573439002037, + 0.016247713938355446, + 0.01588159240782261, + 0.032007623463869095, + 0.024269338697195053, + 0.014313419349491596, + 0.016421863809227943, + 0.02702704444527626, + 0.018102498725056648, + 0.011265476234257221, + 0.014583054929971695, + 0.012702124193310738, + 0.022137563675642014, + 0.011580313555896282, + 0.012556372210383415, + 0.01982365921139717, + 0.01322143618017435, + 0.0077338553965091705, + 0.017972514033317566, + 0.008650314062833786, + 0.007651916705071926, + 0.01933714561164379, + 0.010687777772545815, + 0.011217247694730759, + 0.008248057216405869, + 0.028814762830734253, + 0.034231819212436676, + 0.012724275700747967, + 0.0176410973072052, + 0.024914095178246498, + 0.01720765046775341, + 0.012987104244530201, + 0.008325709030032158, + 0.004886203445494175, + 0.005309319589287043, + 0.0035054779145866632, + 0.004336193203926086, + 0.005562254693359137, + 0.0030728254932910204, + 0.004358133766800165, + 0.00272775674238801, + 0.0037522667553275824, + 0.008466692641377449, + 0.0050687165930867195, + 0.0039070528000593185, + 0.006076928228139877, + 0.010863923467695713, + 0.0028530771378427744, + 0.003057228634133935, + 0.010058761574327946, + 0.009652775712311268, + 0.009886651299893856, + 0.010962741449475288, + 0.014345157891511917, + 0.011171648278832436, + 0.0067858281545341015, + 0.005773873999714851, + 0.00522245978936553, + 0.006112000439316034, + 0.00462466711178422, + 0.006382172927260399, + 0.007071914616972208, + 0.004044970963150263, + 0.0037720215041190386, + 0.0038831252604722977, + 0.006279174704104662, + 0.006257712375372648, + 0.004617674741894007, + 0.005590628366917372, + 0.004977511242032051, + 0.005742611829191446, + 0.004586566239595413, + 0.0043459003791213036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.014194007031619549 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The documentary-style photography begins with a horizontal medium shot that slowly pans rightward, panning across a row of honor guard soldiers standing erect, statuesquely erect. Their uniforms are clad in crisp, dark wool, with crisp collars and neatly arranged epaulettes. Shining brass buttons on their chests gleam faintly in the soft, slightly cool morning light. Each soldier's eyes are tightly closed, their jaws set in a resolute facial expression, seemingly shutting out all emotion. Yet, a single, close-up detail—their uncontrollable, slightly trembling lower lips—silently betrays the surging excitement and sense of honor beneath their iron discipline. With the vast, slightly out-of-focus square in the background, the entire image, captured in exceptional clarity, exudes an atmosphere of solemnity, intense focus, and inner tension, creating a striking and moving contrast between outward composure and inner turmoil.", + "seed": 24829, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_4/index_5482_seed_24829_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2380725484341383, + "vlm_logprob": -0.2380725484341383, + "vlm_yes_no_margin": 2.0, + "vlm_token_credit": -0.2380725484341383, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07620668411254883 + ], + "opposite_logprob_tokens": [ + -2.826206684112549 + ], + "logprob_reward": -0.07620668411254883, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07620668411254883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07021110504865646, + 0.06458433717489243, + 0.055727533996105194, + 0.05431922897696495, + 0.07447570562362671, + 0.06561267375946045, + 0.16182124614715576, + 0.1095641553401947, + 0.0712745189666748, + 0.27240949869155884 + ], + "credit_map": [ + 0.011708976700901985, + 0.011762965470552444, + 0.007505535613745451, + 0.00931047834455967, + 0.005802494939416647, + 0.0067122806794941425, + 0.006120787933468819, + 0.011287582106888294, + 0.006960680708289146, + 0.010368024930357933, + 0.007284342776983976, + 0.007278532721102238, + 0.01086272019892931, + 0.012262988835573196, + 0.005071415100246668, + 0.004495628643780947, + 0.004100745543837547, + 0.0070409164763987064, + 0.006105736363679171, + 0.009739457629621029, + 0.005405752919614315, + 0.005218260455876589, + 0.008879544213414192, + 0.009237125515937805, + 0.008266692981123924, + 0.006983570754528046, + 0.004115461837500334, + 0.003486745059490204, + 0.008632449433207512, + 0.013330986723303795, + 0.005459219682961702, + 0.004044104367494583, + 0.007844706997275352, + 0.010365594178438187, + 0.008881662972271442, + 0.009949198924005032, + 0.005368607584387064, + 0.008801714517176151, + 0.01637313701212406, + 0.006891082506626844, + 0.00861931499093771, + 0.009177514351904392, + 0.0065930685959756374, + 0.010153541341423988, + 0.008980699814856052, + 0.009168979711830616, + 0.007623881101608276, + 0.005295677576214075, + 0.018241271376609802, + 0.018770810216665268, + 0.011754787527024746, + 0.010710314847528934, + 0.04157107695937157, + 0.022520527243614197, + 0.01626730151474476, + 0.021985149011015892, + 0.01089931558817625, + 0.011612837202847004, + 0.011112751439213753, + 0.00998882856220007, + 0.02493441291153431, + 0.020154016092419624, + 0.01033137459307909, + 0.010530621744692326, + 0.007326565682888031, + 0.007960195653140545, + 0.007433919236063957, + 0.006655269302427769, + 0.017876654863357544, + 0.009378382936120033, + 0.00815360527485609, + 0.00648992694914341, + 0.045850109308958054, + 0.037664685398340225, + 0.02424737624824047, + 0.04037894681096077, + 0.02776161953806877, + 0.045679494738578796, + 0.030878227204084396, + 0.01994902454316616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.07620668411254883 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016101917251944542 + ], + "opposite_logprob_tokens": [ + -4.141101837158203 + ], + "logprob_reward": -0.016101917251944542, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016101917251944542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11609877645969391, + 0.09214848279953003, + 0.052428603172302246, + 0.13691364228725433, + 0.11975168436765671, + 0.10923956334590912, + 0.05461065471172333, + 0.09958184510469437, + 0.047092776745557785, + 0.17213395237922668 + ], + "credit_map": [ + 0.013805979862809181, + 0.019670519977808, + 0.010918389074504375, + 0.011707334779202938, + 0.015896327793598175, + 0.015251691453158855, + 0.012515862472355366, + 0.01633268967270851, + 0.00989923719316721, + 0.011085031554102898, + 0.013228964991867542, + 0.01043341401964426, + 0.009061063639819622, + 0.013036101125180721, + 0.01329522393643856, + 0.012109450995922089, + 0.005666764918714762, + 0.004861219320446253, + 0.006506447214633226, + 0.008647272363305092, + 0.005974231753498316, + 0.00552966957911849, + 0.007442752365022898, + 0.0078002517111599445, + 0.012427005916833878, + 0.010994790121912956, + 0.013246214017271996, + 0.015034975484013557, + 0.02346770092844963, + 0.026887770742177963, + 0.021832697093486786, + 0.01302250288426876, + 0.013431522063910961, + 0.017326360568404198, + 0.015111948363482952, + 0.013257346116006374, + 0.011443732306361198, + 0.011403418146073818, + 0.024649228900671005, + 0.013128144666552544, + 0.013134400360286236, + 0.014840659685432911, + 0.009727482683956623, + 0.01591017097234726, + 0.01786145009100437, + 0.014137199148535728, + 0.013928793370723724, + 0.009699421934783459, + 0.005107680801302195, + 0.00689500430598855, + 0.003861707169562578, + 0.0039536356925964355, + 0.017515450716018677, + 0.008116446435451508, + 0.0053776223212480545, + 0.0037831079680472612, + 0.015257307328283787, + 0.011465591378509998, + 0.014080313965678215, + 0.015347931534051895, + 0.01181697752326727, + 0.009683235548436642, + 0.009732088074088097, + 0.012198414653539658, + 0.004617700818926096, + 0.004091640468686819, + 0.004964496474713087, + 0.004374660551548004, + 0.008792940527200699, + 0.0071973539888858795, + 0.00602461863309145, + 0.007029368542134762, + 0.023185204714536667, + 0.019138876348733902, + 0.02039214037358761, + 0.03260035440325737, + 0.01666509360074997, + 0.023654755204916, + 0.0248888973146677, + 0.011608648113906384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.016101917251944542 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.474362313747406 + ], + "opposite_logprob_tokens": [ + -0.974362313747406 + ], + "logprob_reward": -0.474362313747406, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.474362313747406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15179993212223053, + 0.11864141374826431, + 0.09676139801740646, + 0.14399708807468414, + 0.11647991836071014, + 0.08762075006961823, + 0.110186368227005, + 0.05891667306423187, + 0.014135551638901234, + 0.10146089643239975 + ], + "credit_map": [ + 0.017696255818009377, + 0.029791420325636864, + 0.013176895678043365, + 0.015663795173168182, + 0.021016906946897507, + 0.01566101238131523, + 0.01411281619220972, + 0.024680817499756813, + 0.013938548043370247, + 0.015479886904358864, + 0.02059117890894413, + 0.01228250190615654, + 0.01210384164005518, + 0.015631631016731262, + 0.01561942882835865, + 0.012994402088224888, + 0.008675620891153812, + 0.010043674148619175, + 0.014846628531813622, + 0.015729060396552086, + 0.009702809154987335, + 0.014038999564945698, + 0.010577850043773651, + 0.013146757148206234, + 0.010607335716485977, + 0.014489044435322285, + 0.00995230209082365, + 0.008945868350565434, + 0.03469563648104668, + 0.04644032567739487, + 0.010437339544296265, + 0.008429251611232758, + 0.016009312123060226, + 0.01649036817252636, + 0.01497576106339693, + 0.013684170320630074, + 0.009875592775642872, + 0.01020197756588459, + 0.023686695843935013, + 0.011556046083569527, + 0.010709675028920174, + 0.01272306777536869, + 0.009820372797548771, + 0.01443352922797203, + 0.0082499198615551, + 0.010737930424511433, + 0.012310411781072617, + 0.008635842241346836, + 0.011298863217234612, + 0.011771030724048615, + 0.011882982216775417, + 0.008448824286460876, + 0.024295087903738022, + 0.011566133238375187, + 0.01894618384540081, + 0.011977256275713444, + 0.005580337718129158, + 0.005180768668651581, + 0.006690549198538065, + 0.005944511387497187, + 0.010050743818283081, + 0.008055687882006168, + 0.006967966910451651, + 0.010446109808981419, + 0.001993009587749839, + 0.0014253836125135422, + 0.0010568660218268633, + 0.0017999053234234452, + 0.0031005414202809334, + 0.0020098122768104076, + 0.0012844300363212824, + 0.0014656039420515299, + 0.015142865478992462, + 0.014970560558140278, + 0.009504483081400394, + 0.012073613703250885, + 0.010651705786585808, + 0.018401557579636574, + 0.01031374279409647, + 0.010402370244264603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.474362313747406 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032132267951965 + ], + "opposite_logprob_tokens": [ + -2.3503212928771973 + ], + "logprob_reward": -0.10032132267951965, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032132267951965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1654493510723114, + 0.05191131308674812, + 0.09796920418739319, + 0.04951849579811096, + 0.07813890278339386, + 0.0657520592212677, + 0.11565597355365753, + 0.08563075959682465, + 0.04489929974079132, + 0.24507468938827515 + ], + "credit_map": [ + 0.030336199328303337, + 0.01878940686583519, + 0.010106949135661125, + 0.016295915469527245, + 0.017311571165919304, + 0.011170689016580582, + 0.016129866242408752, + 0.04530871659517288, + 0.005077922251075506, + 0.008101041428744793, + 0.009169899858534336, + 0.007407318335026503, + 0.005864263512194157, + 0.005778952967375517, + 0.005688896868377924, + 0.004823008552193642, + 0.00753118097782135, + 0.007627955637872219, + 0.0114173898473382, + 0.025962788611650467, + 0.009109048172831535, + 0.010740634053945541, + 0.011436537839472294, + 0.01414366252720356, + 0.00459953723475337, + 0.00492757186293602, + 0.0037521945778280497, + 0.004118938464671373, + 0.009356273338198662, + 0.01414918527007103, + 0.005050895735621452, + 0.003563891863450408, + 0.0064677768386900425, + 0.00837006140500307, + 0.010732988826930523, + 0.011146028526127338, + 0.0052930130623281, + 0.008146798238158226, + 0.018621841445565224, + 0.00936038512736559, + 0.006727378349751234, + 0.008617701940238476, + 0.006811023224145174, + 0.010852230712771416, + 0.007306573912501335, + 0.008762185461819172, + 0.010532388463616371, + 0.006142564117908478, + 0.012529806233942509, + 0.013353174552321434, + 0.013341890648007393, + 0.008733379654586315, + 0.022312499582767487, + 0.011798637919127941, + 0.01911509782075882, + 0.014471466653048992, + 0.010405009612441063, + 0.007320437114685774, + 0.009648785926401615, + 0.010595357976853848, + 0.012962778098881245, + 0.010839119553565979, + 0.009779389016330242, + 0.01407987717539072, + 0.005649743136018515, + 0.00538892624899745, + 0.004671682603657246, + 0.0037554188165813684, + 0.007878106087446213, + 0.006884567439556122, + 0.006283273454755545, + 0.0043875789269804955, + 0.03975432366132736, + 0.033232033252716064, + 0.021828841418027878, + 0.03840353339910507, + 0.025746256113052368, + 0.03788406401872635, + 0.030343927443027496, + 0.017881685867905617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.10032132267951965 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233705043792725 + ], + "opposite_logprob_tokens": [ + -0.8983705043792725 + ], + "logprob_reward": -0.5233705043792725, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233705043792725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1453171819448471, + 0.4145899713039398, + 0.09183605760335922, + 0.025889834389090538, + 0.03357617184519768, + 0.031407780945301056, + 0.08800644427537918, + 0.03595557436347008, + 0.05774485319852829, + 0.07567612826824188 + ], + "credit_map": [ + 0.01705758646130562, + 0.031236810609698296, + 0.015968285501003265, + 0.010356650687754154, + 0.02296609804034233, + 0.01915154606103897, + 0.014135624282062054, + 0.014444577507674694, + 0.039447154849767685, + 0.05420157313346863, + 0.0774199590086937, + 0.059317708015441895, + 0.05432938039302826, + 0.04973512887954712, + 0.0395287461578846, + 0.040610335767269135, + 0.0077394479885697365, + 0.00851792935281992, + 0.009653514251112938, + 0.01306366827338934, + 0.01220645196735859, + 0.010700177401304245, + 0.018367120996117592, + 0.011587752029299736, + 0.0031023258343338966, + 0.0037491084076464176, + 0.003943920135498047, + 0.0025568108540028334, + 0.0033729635179042816, + 0.0030141063034534454, + 0.003714560763910413, + 0.002436037641018629, + 0.004933970049023628, + 0.005357656162232161, + 0.003507639979943633, + 0.0030147305224090815, + 0.0030191997066140175, + 0.0070225209929049015, + 0.0033061974681913853, + 0.0034142567310482264, + 0.0043383631855249405, + 0.004921569488942623, + 0.0029607671312987804, + 0.003988956566900015, + 0.0040028211660683155, + 0.005048599559813738, + 0.0033831081818789244, + 0.002763595199212432, + 0.008129410445690155, + 0.007910856045782566, + 0.008939351886510849, + 0.0072180964052677155, + 0.018140748143196106, + 0.008936962112784386, + 0.012517391704022884, + 0.01621362753212452, + 0.0039051095955073833, + 0.003989078104496002, + 0.0036741322837769985, + 0.004138159565627575, + 0.007951022125780582, + 0.004044991452246904, + 0.003791545517742634, + 0.004461534321308136, + 0.005972631741315126, + 0.0048292274586856365, + 0.007134472485631704, + 0.00603955565020442, + 0.009536894969642162, + 0.008858178742229939, + 0.007843178696930408, + 0.00753071578219533, + 0.009123505093157291, + 0.007313546724617481, + 0.009596601128578186, + 0.01433990802615881, + 0.006730611436069012, + 0.010949362069368362, + 0.011711944825947285, + 0.005910647101700306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.5233705043792725 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The documentary-style photography begins with a horizontal medium shot that slowly pans rightward, panning across a row of honor guard soldiers standing erect, statuesquely erect. Their uniforms are clad in crisp, dark wool, with crisp collars and neatly arranged epaulettes. Shining brass buttons on their chests gleam faintly in the soft, slightly cool morning light. Each soldier's eyes are tightly closed, their jaws set in a resolute facial expression, seemingly shutting out all emotion. Yet, a single, close-up detail—their uncontrollable, slightly trembling lower lips—silently betrays the surging excitement and sense of honor beneath their iron discipline. With the vast, slightly out-of-focus square in the background, the entire image, captured in exceptional clarity, exudes an atmosphere of solemnity, intense focus, and inner tension, creating a striking and moving contrast between outward composure and inner turmoil.", + "seed": 24830, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_4/index_5482_seed_24830_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14514669524505736, + "vlm_logprob": -0.14514669524505736, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.14514669524505736, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07673930376768112 + ], + "opposite_logprob_tokens": [ + -2.7017393112182617 + ], + "logprob_reward": -0.07673930376768112, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07673930376768112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12969358265399933, + 0.07791946828365326, + 0.06334829330444336, + 0.19788262248039246, + 0.050745394080877304, + 0.06277577579021454, + 0.040872395038604736, + 0.1866733878850937, + 0.10861995071172714, + 0.08146911859512329 + ], + "credit_map": [ + 0.02425265870988369, + 0.010554913431406021, + 0.02562832646071911, + 0.018322594463825226, + 0.006890386343002319, + 0.01015404798090458, + 0.01811889000236988, + 0.015771759673953056, + 0.012276856228709221, + 0.007405818440020084, + 0.00924727413803339, + 0.012698924168944359, + 0.006165559869259596, + 0.006547062192112207, + 0.014438947662711143, + 0.009139023721218109, + 0.01227858942002058, + 0.005931403953582048, + 0.00823182798922062, + 0.008605279959738255, + 0.008272015489637852, + 0.005052110180258751, + 0.007900928147137165, + 0.007076142355799675, + 0.02983691543340683, + 0.02911834977567196, + 0.026301313191652298, + 0.026199007406830788, + 0.01938987337052822, + 0.02168964594602585, + 0.021577589213848114, + 0.023769918829202652, + 0.00653274729847908, + 0.0040089343674480915, + 0.01040258351713419, + 0.005954194813966751, + 0.005966759752482176, + 0.006948051508516073, + 0.005406628362834454, + 0.005525494460016489, + 0.008327850140631199, + 0.007108347024768591, + 0.0073473406955599785, + 0.00995380524545908, + 0.008109137415885925, + 0.0062566534616053104, + 0.0071214777417480946, + 0.008551168255507946, + 0.006322480272501707, + 0.006403136067092419, + 0.005478253122419119, + 0.004050869960337877, + 0.00341372680850327, + 0.003527906257659197, + 0.0064375572837889194, + 0.005238468758761883, + 0.020273925736546516, + 0.025461679324507713, + 0.026478702202439308, + 0.02083507739007473, + 0.02178385481238365, + 0.020685896277427673, + 0.020763009786605835, + 0.03039124235510826, + 0.011887299828231335, + 0.015720898285508156, + 0.012815079651772976, + 0.012276103720068932, + 0.012353446334600449, + 0.010789143852889538, + 0.01300942711532116, + 0.019768550992012024, + 0.008505696430802345, + 0.012633946724236012, + 0.010365496389567852, + 0.010067244991660118, + 0.01050519198179245, + 0.008886829018592834, + 0.011350424960255623, + 0.009154286235570908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.07673930376768112 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006756792310625315 + ], + "opposite_logprob_tokens": [ + -5.006756782531738 + ], + "logprob_reward": -0.006756792310625315, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006756792310625315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18763107061386108, + 0.10362335294485092, + 0.07415284961462021, + 0.0757218673825264, + 0.06135974079370499, + 0.07440084964036942, + 0.07262717187404633, + 0.10907388478517532, + 0.06898540258407593, + 0.1724238395690918 + ], + "credit_map": [ + 0.03106386959552765, + 0.01936008781194687, + 0.027899328619241714, + 0.024268759414553642, + 0.014852246269583702, + 0.016841797158122063, + 0.032585471868515015, + 0.02075950987637043, + 0.012495631352066994, + 0.015193569473922253, + 0.011778402142226696, + 0.014444893226027489, + 0.014315331354737282, + 0.009085189551115036, + 0.016281330958008766, + 0.010029001161456108, + 0.007973810657858849, + 0.007392357103526592, + 0.008107678964734077, + 0.011858376674354076, + 0.009205104783177376, + 0.010433172807097435, + 0.011455115862190723, + 0.007727229036390781, + 0.010341987013816833, + 0.009607717394828796, + 0.010656180791556835, + 0.011690694838762283, + 0.00624458584934473, + 0.008869167417287827, + 0.010097041726112366, + 0.008214491419494152, + 0.008098024874925613, + 0.006735291797667742, + 0.006314938422292471, + 0.008231381885707378, + 0.006311600096523762, + 0.007298286072909832, + 0.010026239790022373, + 0.008343975059688091, + 0.01079552061855793, + 0.004727908875793219, + 0.014534994028508663, + 0.007561052683740854, + 0.005409238860011101, + 0.010218540206551552, + 0.012027109041810036, + 0.009126488119363785, + 0.008720085956156254, + 0.016702203080058098, + 0.007323074154555798, + 0.007353105116635561, + 0.009195365011692047, + 0.006858011707663536, + 0.008261792361736298, + 0.008213536813855171, + 0.024081138893961906, + 0.008278039284050465, + 0.017072828486561775, + 0.013637933880090714, + 0.007835779339075089, + 0.009915721602737904, + 0.015008723363280296, + 0.013243716210126877, + 0.007045296486467123, + 0.007962450385093689, + 0.012024535797536373, + 0.008253700099885464, + 0.005136559251695871, + 0.00584542378783226, + 0.009269513189792633, + 0.013447914272546768, + 0.019649207592010498, + 0.02232244424521923, + 0.030767057090997696, + 0.020203549414873123, + 0.01662198081612587, + 0.02164984494447708, + 0.01868782937526703, + 0.022521929815411568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.006756792310625315 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289594888687134 + ], + "opposite_logprob_tokens": [ + -1.0539594888687134 + ], + "logprob_reward": -0.4289594888687134, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289594888687134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17876999080181122, + 0.10137014836072922, + 0.04873194918036461, + 0.07913091778755188, + 0.10848256200551987, + 0.06550943851470947, + 0.18216460943222046, + 0.11943069845438004, + 0.07217168062925339, + 0.044237930327653885 + ], + "credit_map": [ + 0.026013322174549103, + 0.017738912254571915, + 0.029878677800297737, + 0.029347019270062447, + 0.015118688344955444, + 0.017744872719049454, + 0.023961851373314857, + 0.018966635689139366, + 0.014211397618055344, + 0.011596368625760078, + 0.014715000987052917, + 0.01412818394601345, + 0.009925076737999916, + 0.010739849880337715, + 0.01652713119983673, + 0.009527136571705341, + 0.006832957733422518, + 0.004872663412243128, + 0.00739132147282362, + 0.00588106969371438, + 0.006223401986062527, + 0.004926381632685661, + 0.006726238876581192, + 0.00587791483849287, + 0.009238244034349918, + 0.007359044160693884, + 0.023752113804221153, + 0.013571949675679207, + 0.0059632807970047, + 0.006805622950196266, + 0.006280628498643637, + 0.006160033401101828, + 0.014669905416667461, + 0.009441995993256569, + 0.009743303060531616, + 0.015976324677467346, + 0.01102970726788044, + 0.01106359250843525, + 0.01906401664018631, + 0.0174937155097723, + 0.009245474822819233, + 0.00463632820174098, + 0.01724063791334629, + 0.00896508526057005, + 0.005134819075465202, + 0.004733250476419926, + 0.008368665352463722, + 0.007185174152255058, + 0.022707754746079445, + 0.036073148250579834, + 0.027264941483736038, + 0.021274574100971222, + 0.015002825297415257, + 0.013101979158818722, + 0.024555973708629608, + 0.022183431312441826, + 0.018933327868580818, + 0.015198677778244019, + 0.01254742406308651, + 0.014525586739182472, + 0.013547663576900959, + 0.012511826120316982, + 0.01736004278063774, + 0.014806150458753109, + 0.010026120580732822, + 0.012052918784320354, + 0.007234142627567053, + 0.008364148437976837, + 0.008287847973406315, + 0.008557824417948723, + 0.00924783106893301, + 0.00840084906667471, + 0.004612745717167854, + 0.006059879437088966, + 0.004880974534898996, + 0.005739470478147268, + 0.005064763128757477, + 0.006126760970801115, + 0.005729817319661379, + 0.006023522932082415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.4289594888687134 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294395476579666 + ], + "opposite_logprob_tokens": [ + -2.2379438877105713 + ], + "logprob_reward": -0.11294395476579666, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294395476579666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08286259323358536, + 0.0482812374830246, + 0.03713347762823105, + 0.08898399025201797, + 0.024293607100844383, + 0.016824834048748016, + 0.06438331305980682, + 0.07599740475416183, + 0.227046936750412, + 0.33419257402420044 + ], + "credit_map": [ + 0.010411080904304981, + 0.009149069897830486, + 0.012035592459142208, + 0.013953047804534435, + 0.008284710347652435, + 0.0076599461026489735, + 0.012252289801836014, + 0.009116863831877708, + 0.005572233349084854, + 0.007958915084600449, + 0.006704388186335564, + 0.007105446420609951, + 0.006500760558992624, + 0.003687366144731641, + 0.006605332251638174, + 0.004146800376474857, + 0.005594965070486069, + 0.0030510544311255217, + 0.006705690640956163, + 0.00619586231186986, + 0.0035205106250941753, + 0.00340065429918468, + 0.005608746781945229, + 0.003055999055504799, + 0.011907641775906086, + 0.00557791069149971, + 0.03288601338863373, + 0.016262413933873177, + 0.0048021734692156315, + 0.004373176954686642, + 0.006618920248001814, + 0.006555745378136635, + 0.002734340960159898, + 0.0022697397507727146, + 0.005000253673642874, + 0.002762110671028495, + 0.00212550419382751, + 0.003445322159677744, + 0.0029637874104082584, + 0.0029925485141575336, + 0.003089064499363303, + 0.0027605402283370495, + 0.0026222437154501677, + 0.0022612959146499634, + 0.0013041493948549032, + 0.0011896687792614102, + 0.0018612927524372935, + 0.0017365803942084312, + 0.006784704979509115, + 0.016770122572779655, + 0.007006713654845953, + 0.00754543486982584, + 0.007539616897702217, + 0.004786428064107895, + 0.00794847123324871, + 0.006001824513077736, + 0.011377425864338875, + 0.013231913559138775, + 0.008442872203886509, + 0.008933337405323982, + 0.009118832647800446, + 0.005354574881494045, + 0.009992863051593304, + 0.009545582346618176, + 0.034240879118442535, + 0.025845473632216454, + 0.047911785542964935, + 0.02786913327872753, + 0.02158173732459545, + 0.020525172352790833, + 0.019886616617441177, + 0.029186144471168518, + 0.03853686898946762, + 0.036764200776815414, + 0.06414248049259186, + 0.040087874978780746, + 0.03297923877835274, + 0.045639898627996445, + 0.03390614688396454, + 0.04213588684797287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.11294395476579666 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a horizontal medium shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are honor guard soldiers standing erect in a row wearing dark wool uniforms with brass buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the soldiers have their eyes tightly closed with a resolute facial expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a close-up detail showing the soldiers' lower lips trembling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, slightly out-of-focus square?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033393651247025 + ], + "opposite_logprob_tokens": [ + -2.3503339290618896 + ], + "logprob_reward": -0.10033393651247025, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033393651247025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5506215691566467, + 0.07566562294960022, + 0.030905187129974365, + 0.09746845811605453, + 0.034611381590366364, + 0.02649037353694439, + 0.053054191172122955, + 0.04452197998762131, + 0.025044064968824387, + 0.0616171658039093 + ], + "credit_map": [ + 0.11088546365499496, + 0.047194309532642365, + 0.1193944439291954, + 0.06805961579084396, + 0.03799059987068176, + 0.03229716792702675, + 0.0725826546549797, + 0.06221729889512062, + 0.013594254851341248, + 0.008424903266131878, + 0.010919671505689621, + 0.009776059538125992, + 0.0055628614500164986, + 0.007606972940266132, + 0.012528937309980392, + 0.007251957897096872, + 0.004098793026059866, + 0.004750474821776152, + 0.0038159452378749847, + 0.0040700118988752365, + 0.0035537693183869123, + 0.0031788514461368322, + 0.004134142305701971, + 0.003303196979686618, + 0.012504789978265762, + 0.013541075401008129, + 0.021408725529909134, + 0.011381622403860092, + 0.009245866909623146, + 0.010420496575534344, + 0.009469722397625446, + 0.009496156126260757, + 0.005171110387891531, + 0.0037473919801414013, + 0.003868628293275833, + 0.004950079135596752, + 0.0040517239831388, + 0.004784198012202978, + 0.0038520062807947397, + 0.0041862428188323975, + 0.004093041177839041, + 0.0025155122857540846, + 0.0036480468697845936, + 0.00274142948910594, + 0.0027645297814160585, + 0.0032057249918580055, + 0.003766764188185334, + 0.0037553247530013323, + 0.0065530091524124146, + 0.012470252811908722, + 0.0068374499678611755, + 0.0071259005926549435, + 0.005617959890514612, + 0.003609017003327608, + 0.0055873217061161995, + 0.005253279581665993, + 0.006103802006691694, + 0.00396494846791029, + 0.007960319519042969, + 0.005192918237298727, + 0.003435614285990596, + 0.00268625863827765, + 0.005054411478340626, + 0.010123706422746181, + 0.0034175103064626455, + 0.0022349306382238865, + 0.0049164025112986565, + 0.0029887412674725056, + 0.002499894006177783, + 0.0028626935090869665, + 0.0034393847454339266, + 0.002684506354853511, + 0.0062832580879330635, + 0.0061533791013062, + 0.012949887663125992, + 0.007816149853169918, + 0.005943516734987497, + 0.009062381461262703, + 0.006120423786342144, + 0.007288170978426933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.10033393651247025 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_5/index_tensor([2782, 2782])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_5/index_tensor([2782, 2782])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5d006b49ac4861f7dd17cb7e929154ed619d1c29 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_5/index_tensor([2782, 2782])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "highly realistic, full-body static shot, captured from a slight overhead angle, captures a shirtless man lying on his side on dark sheets in a dimly lit room. His back is sharply defined, his muscles barely visible in the gloom, and a fine sheen of sweat covers his skin. The only warm light in the scene emanates from a slowly burning moxa stick in the center of his back. Its fiery red tip casts a small, concentrated, warm, orange-yellow glow, illuminating the surrounding skin with a translucent glow. As his chest and back rise and fall rhythmically with his steady, gentle breathing, wisps of soft white moxa smoke curl upward from the stick, swirling and dissipating in the still air, illuminated by the single, warm light, forming visible beams of light. This faint light creates a stark contrast with the surrounding darkness, profoundly sculpting the curves and shadows of his body, creating a paradoxical atmosphere of both healing tranquility and a subtle sense of endurance. The entire image is imbued with a dramatic interplay of light and shadow.", + "seed": 13131, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_5/index_2782_seed_13131_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.4126579684205353, + "vlm_logprob": -0.4126579684205353, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.4126579684205353, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028852412477135658 + ], + "opposite_logprob_tokens": [ + -3.6538524627685547 + ], + "logprob_reward": -0.028852412477135658, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.028852412477135658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20345132052898407, + 0.059556614607572556, + 0.035462722182273865, + 0.1375385969877243, + 0.1508571058511734, + 0.07569779455661774, + 0.04772520065307617, + 0.058506473898887634, + 0.18709613382816315, + 0.04410801827907562 + ], + "credit_map": [ + 0.019334174692630768, + 0.0244204830378294, + 0.023433171212673187, + 0.035690877586603165, + 0.03152330592274666, + 0.02071239799261093, + 0.02119462378323078, + 0.02714230865240097, + 0.0063309445977211, + 0.00868656113743782, + 0.006297655403614044, + 0.010786444880068302, + 0.008992553688585758, + 0.006158798933029175, + 0.006242966279387474, + 0.006060697603970766, + 0.004243536852300167, + 0.004361107479780912, + 0.004329288844019175, + 0.00494755432009697, + 0.004627931863069534, + 0.0037205384578555822, + 0.004937245510518551, + 0.004295520018786192, + 0.020760798826813698, + 0.021797703579068184, + 0.01786261424422264, + 0.026894643902778625, + 0.014772579073905945, + 0.009163030423223972, + 0.013236217200756073, + 0.013051020912826061, + 0.02760673686861992, + 0.02141251415014267, + 0.018421262502670288, + 0.026658453047275543, + 0.012297585606575012, + 0.012082703411579132, + 0.015244653448462486, + 0.017133209854364395, + 0.007718230132013559, + 0.010426882654428482, + 0.011931083165109158, + 0.013452638871967793, + 0.009358806535601616, + 0.006240874994546175, + 0.008794433437287807, + 0.007774845231324434, + 0.00575467711314559, + 0.005275202915072441, + 0.00617461372166872, + 0.011205045506358147, + 0.0046912552788853645, + 0.003670509671792388, + 0.0049735382199287415, + 0.005980361253023148, + 0.005779756233096123, + 0.007036997936666012, + 0.007213909644633532, + 0.010813718661665916, + 0.007706556469202042, + 0.0058119106106460094, + 0.006362517364323139, + 0.00778111070394516, + 0.018573330715298653, + 0.01871851459145546, + 0.022475579753518105, + 0.03215502202510834, + 0.044665418565273285, + 0.016056057065725327, + 0.015262259170413017, + 0.019189968705177307, + 0.0049591572023928165, + 0.003850560402497649, + 0.0076962788589298725, + 0.006632504519075155, + 0.0071090273559093475, + 0.003918422851711512, + 0.003967328928411007, + 0.005974743515253067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.028852412477135658 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005264111328870058 + ], + "opposite_logprob_tokens": [ + -5.2552642822265625 + ], + "logprob_reward": -0.005264111328870058, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005264111328870058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03596653416752815, + 0.027166252955794334, + 0.018838249146938324, + 0.08447284996509552, + 0.10707423835992813, + 0.0315791480243206, + 0.05603553354740143, + 0.04238666594028473, + 0.4924173355102539, + 0.10406322777271271 + ], + "credit_map": [ + 0.004327215254306793, + 0.0030253767035901546, + 0.004832473583519459, + 0.006482220254838467, + 0.005542906001210213, + 0.004045617766678333, + 0.003258249955251813, + 0.004452475346624851, + 0.005262601654976606, + 0.002856830134987831, + 0.003835967741906643, + 0.003973063547164202, + 0.0033363141119480133, + 0.002558276988565922, + 0.002153487876057625, + 0.0031897120643407106, + 0.0021878951229155064, + 0.0025379571598023176, + 0.002149593085050583, + 0.0034310210030525923, + 0.0021124156191945076, + 0.0017892803298309445, + 0.0024822941049933434, + 0.0021477940026670694, + 0.010056152008473873, + 0.011111554689705372, + 0.010629682801663876, + 0.020167309790849686, + 0.007197184022516012, + 0.005829459521919489, + 0.009254763834178448, + 0.01022674236446619, + 0.014927199110388756, + 0.016314763575792313, + 0.013659185729920864, + 0.02088811993598938, + 0.009487448260188103, + 0.009189287200570107, + 0.011424160562455654, + 0.01118407305330038, + 0.003326473757624626, + 0.004206878133118153, + 0.0037821896839886904, + 0.006125771906226873, + 0.0031631041783839464, + 0.002639890182763338, + 0.004691230598837137, + 0.003643611678853631, + 0.008000086061656475, + 0.0067561324685812, + 0.009667865931987762, + 0.008218443021178246, + 0.005982471629977226, + 0.004407523199915886, + 0.005715313833206892, + 0.007287696469575167, + 0.004718502517789602, + 0.005781283602118492, + 0.004949018359184265, + 0.007052881643176079, + 0.006683488842099905, + 0.004337273072451353, + 0.00448250537738204, + 0.004381714388728142, + 0.049153175204992294, + 0.05376564338803291, + 0.06578642874956131, + 0.08732365816831589, + 0.1078866571187973, + 0.03995063900947571, + 0.037145521491765976, + 0.05140560865402222, + 0.008985749445855618, + 0.010630139149725437, + 0.014034748077392578, + 0.011335890740156174, + 0.02608243003487587, + 0.01061113178730011, + 0.010871478356420994, + 0.011511663906276226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.005264111328870058 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948696285486221 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.005948696285486221, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948696285486221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16315710544586182, + 0.08401034772396088, + 0.0705142542719841, + 0.07273492962121964, + 0.13382327556610107, + 0.07368781417608261, + 0.1066683828830719, + 0.04839831963181496, + 0.12302495539188385, + 0.12398063391447067 + ], + "credit_map": [ + 0.018424170091748238, + 0.02173234522342682, + 0.018882915377616882, + 0.03613256290555, + 0.017283683642745018, + 0.015948139131069183, + 0.01747145876288414, + 0.01728183776140213, + 0.012430129572749138, + 0.013781809248030186, + 0.008149153552949429, + 0.012839851900935173, + 0.012834002263844013, + 0.005783777683973312, + 0.00838396605104208, + 0.009807657450437546, + 0.008364653214812279, + 0.009191615507006645, + 0.008801906369626522, + 0.008956901729106903, + 0.01324718352407217, + 0.005645879078656435, + 0.007861104793846607, + 0.008445008657872677, + 0.007633752189576626, + 0.015446089208126068, + 0.00868587102741003, + 0.013149017468094826, + 0.009834131225943565, + 0.004485254175961018, + 0.006485487800091505, + 0.007015333045274019, + 0.023267677053809166, + 0.0186298955231905, + 0.015401632525026798, + 0.026149701327085495, + 0.009734070859849453, + 0.012552105821669102, + 0.014446532353758812, + 0.013641653582453728, + 0.01090146042406559, + 0.010062159039080143, + 0.008326521143317223, + 0.01217679027467966, + 0.006852929480373859, + 0.006281836424022913, + 0.010355415754020214, + 0.008730697445571423, + 0.012125551700592041, + 0.014370203949511051, + 0.014668548479676247, + 0.018343130126595497, + 0.014167523942887783, + 0.008797838352620602, + 0.012245161458849907, + 0.011950426734983921, + 0.005387745797634125, + 0.007451810874044895, + 0.006316699553281069, + 0.00992017425596714, + 0.004238673020154238, + 0.004097919445484877, + 0.0036942255683243275, + 0.007291072513908148, + 0.01102455798536539, + 0.012050557881593704, + 0.013137363828718662, + 0.023130647838115692, + 0.02970043383538723, + 0.010067890398204327, + 0.011199818924069405, + 0.012713676318526268, + 0.009638242423534393, + 0.01576360873878002, + 0.012566257268190384, + 0.024935144931077957, + 0.02075646072626114, + 0.012523332610726357, + 0.01321615930646658, + 0.014581432566046715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.005948696285486221 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11289241909980774 + ], + "opposite_logprob_tokens": [ + -2.2378923892974854 + ], + "logprob_reward": -0.11289241909980774, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11289241909980774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18691982328891754, + 0.08272599428892136, + 0.08327481895685196, + 0.1306532472372055, + 0.11194425076246262, + 0.12911349534988403, + 0.06695592403411865, + 0.051383454352617264, + 0.08866962790489197, + 0.06835935264825821 + ], + "credit_map": [ + 0.018817156553268433, + 0.02749134786427021, + 0.022255340591073036, + 0.032993532717227936, + 0.02830781228840351, + 0.01698767952620983, + 0.019195493310689926, + 0.020871469751000404, + 0.009784399531781673, + 0.01115498412400484, + 0.008735693991184235, + 0.015476963482797146, + 0.014053618535399437, + 0.006927304435521364, + 0.007296624593436718, + 0.009296403266489506, + 0.012485957704484463, + 0.009682840667665005, + 0.010172906331717968, + 0.01410748902708292, + 0.01267467811703682, + 0.007234197109937668, + 0.008487430401146412, + 0.008429314009845257, + 0.020684558898210526, + 0.022723954170942307, + 0.013817491009831429, + 0.023959631100296974, + 0.013285340741276741, + 0.008039393462240696, + 0.01359940879046917, + 0.01454346813261509, + 0.01820005290210247, + 0.012564778327941895, + 0.015222926624119282, + 0.02080063335597515, + 0.011459852568805218, + 0.01000224333256483, + 0.012279011309146881, + 0.01141474861651659, + 0.015928668901324272, + 0.017913086339831352, + 0.016228003427386284, + 0.02660141885280609, + 0.010328610427677631, + 0.010235953144729137, + 0.017501549795269966, + 0.014376208186149597, + 0.0112905353307724, + 0.0074536665342748165, + 0.010839964263141155, + 0.011327916756272316, + 0.005964661482721567, + 0.004681936465203762, + 0.007358836941421032, + 0.008038406260311604, + 0.005406627897173166, + 0.007011245936155319, + 0.006399944424629211, + 0.011063850484788418, + 0.0053541758097708225, + 0.004352416843175888, + 0.005231332499533892, + 0.006563860923051834, + 0.0083873076364398, + 0.009693455882370472, + 0.01319003663957119, + 0.015251528471708298, + 0.018196526914834976, + 0.006609403993934393, + 0.00637813052162528, + 0.010963236913084984, + 0.005705866031348705, + 0.00709500489756465, + 0.008696157485246658, + 0.008464387618005276, + 0.016728349030017853, + 0.007443360518664122, + 0.006699668243527412, + 0.0075265588238835335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.11289241909980774 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.910332202911377 + ], + "opposite_logprob_tokens": [ + -0.16033226251602173 + ], + "logprob_reward": -1.910332202911377, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.910332202911377, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24112427234649658, + 0.12819717824459076, + 0.04920518770813942, + 0.07551173865795135, + 0.16403940320014954, + 0.14622999727725983, + 0.01905294880270958, + 0.09668653458356857, + 0.06235901266336441, + 0.017593760043382645 + ], + "credit_map": [ + 0.025111008435487747, + 0.031108230352401733, + 0.02768762595951557, + 0.048987023532390594, + 0.030915100127458572, + 0.021876465529203415, + 0.027424808591604233, + 0.028014007955789566, + 0.01275950763374567, + 0.012511732056736946, + 0.014629014767706394, + 0.015343844890594482, + 0.029946884140372276, + 0.012329215183854103, + 0.01697503589093685, + 0.013701955787837505, + 0.005817973520606756, + 0.006655706092715263, + 0.007118249777704477, + 0.008325593546032906, + 0.005643239710479975, + 0.0036264241207391024, + 0.006060285493731499, + 0.00595771474763751, + 0.011201450601220131, + 0.015318064950406551, + 0.007930626161396503, + 0.010843656957149506, + 0.00796240009367466, + 0.005643993616104126, + 0.007937567308545113, + 0.008673976175487041, + 0.023985324427485466, + 0.023936862125992775, + 0.019675157964229584, + 0.029634110629558563, + 0.016213754191994667, + 0.011219620704650879, + 0.019265195354819298, + 0.020109372213482857, + 0.019795862957835197, + 0.02004643715918064, + 0.020267346873879433, + 0.030366383492946625, + 0.01135973446071148, + 0.011001244187355042, + 0.017649715766310692, + 0.015743277966976166, + 0.002816593274474144, + 0.002993255155161023, + 0.00211655767634511, + 0.0037887990474700928, + 0.0015747006982564926, + 0.0014119991101324558, + 0.0021433585789054632, + 0.002207685261964798, + 0.009424356743693352, + 0.011879087425768375, + 0.012514350935816765, + 0.019040487706661224, + 0.012913602404296398, + 0.009835411794483662, + 0.009571429342031479, + 0.011507799848914146, + 0.006024159956723452, + 0.006298984866589308, + 0.007340623065829277, + 0.011475465260446072, + 0.014817425049841404, + 0.0052964654751122, + 0.0047896127216517925, + 0.006316277198493481, + 0.0017698509618639946, + 0.0019792986568063498, + 0.0022382873576134443, + 0.0037453423719853163, + 0.0032554829958826303, + 0.0014742034254595637, + 0.0014995215460658073, + 0.0016317734261974692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -1.910332202911377 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "highly realistic, full-body static shot, captured from a slight overhead angle, captures a shirtless man lying on his side on dark sheets in a dimly lit room. His back is sharply defined, his muscles barely visible in the gloom, and a fine sheen of sweat covers his skin. The only warm light in the scene emanates from a slowly burning moxa stick in the center of his back. Its fiery red tip casts a small, concentrated, warm, orange-yellow glow, illuminating the surrounding skin with a translucent glow. As his chest and back rise and fall rhythmically with his steady, gentle breathing, wisps of soft white moxa smoke curl upward from the stick, swirling and dissipating in the still air, illuminated by the single, warm light, forming visible beams of light. This faint light creates a stark contrast with the surrounding darkness, profoundly sculpting the curves and shadows of his body, creating a paradoxical atmosphere of both healing tranquility and a subtle sense of endurance. The entire image is imbued with a dramatic interplay of light and shadow.", + "seed": 13132, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_5/index_2782_seed_13132_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.48276100810617206, + "vlm_logprob": -0.48276100810617206, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.48276100810617206, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06400410830974579 + ], + "opposite_logprob_tokens": [ + -2.8140041828155518 + ], + "logprob_reward": -0.06400410830974579, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06400410830974579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2800978124141693, + 0.10380785167217255, + 0.08922404795885086, + 0.04561090096831322, + 0.1633341908454895, + 0.17254625260829926, + 0.05907735228538513, + 0.038508519530296326, + 0.02883378602564335, + 0.01895925961434841 + ], + "credit_map": [ + 0.033058252185583115, + 0.03511570766568184, + 0.0456569567322731, + 0.035233546048402786, + 0.019101636484265327, + 0.03434721753001213, + 0.03505164757370949, + 0.04253285005688667, + 0.013259832747280598, + 0.014236113987863064, + 0.01780099794268608, + 0.013978354632854462, + 0.006592203862965107, + 0.012506148777902126, + 0.011244454421103, + 0.014189746230840683, + 0.014458192512392998, + 0.008238286711275578, + 0.01656937412917614, + 0.012173098511993885, + 0.005942801479250193, + 0.010040984489023685, + 0.00876123458147049, + 0.013040079735219479, + 0.006287650670856237, + 0.0047262548469007015, + 0.007265982683748007, + 0.007374632637947798, + 0.0030546593479812145, + 0.004643505439162254, + 0.005619370844215155, + 0.00663884449750185, + 0.01964552141726017, + 0.019857926294207573, + 0.023624742403626442, + 0.03018227219581604, + 0.013495596125721931, + 0.01863722689449787, + 0.01458044070750475, + 0.02331046387553215, + 0.023691819980740547, + 0.017362704500555992, + 0.029355160892009735, + 0.02733444981276989, + 0.014702984131872654, + 0.01912379264831543, + 0.016876868903636932, + 0.024098467081785202, + 0.006527494639158249, + 0.005916472524404526, + 0.007989893667399883, + 0.011582452803850174, + 0.003686419455334544, + 0.006075954996049404, + 0.009223568253219128, + 0.008075092919170856, + 0.004529571160674095, + 0.003771260380744934, + 0.00513221463188529, + 0.008026148192584515, + 0.002776720793917775, + 0.004532369319349527, + 0.004578629042953253, + 0.005161605309695005, + 0.0037303855642676353, + 0.002718560630455613, + 0.004218250513076782, + 0.005058703478425741, + 0.0014774739975109696, + 0.0041880481876432896, + 0.0035885826218873262, + 0.003853783244267106, + 0.002707606879994273, + 0.0020883141551166773, + 0.0027661302592605352, + 0.002878876868635416, + 0.0016387563664466143, + 0.001601621275767684, + 0.002407924272119999, + 0.002870029304176569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.06400410830974579 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014218100346624851 + ], + "opposite_logprob_tokens": [ + -4.264218330383301 + ], + "logprob_reward": -0.014218100346624851, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014218100346624851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2556903660297394, + 0.13131952285766602, + 0.07220623642206192, + 0.10393035411834717, + 0.0689302533864975, + 0.17361390590667725, + 0.06670083850622177, + 0.0649234727025032, + 0.03212396800518036, + 0.030561130493879318 + ], + "credit_map": [ + 0.032516203820705414, + 0.02502531185746193, + 0.04403158649802208, + 0.03700783848762512, + 0.01794298365712166, + 0.031053174287080765, + 0.03157120943069458, + 0.03654204308986664, + 0.014706091023981571, + 0.02181337960064411, + 0.02160409279167652, + 0.017309460788965225, + 0.009052724577486515, + 0.015512409619987011, + 0.014366811141371727, + 0.01695455051958561, + 0.01014227420091629, + 0.006835611537098885, + 0.014990153722465038, + 0.010549675673246384, + 0.005346744786947966, + 0.007178476545959711, + 0.007782739587128162, + 0.009380554780364037, + 0.013905271887779236, + 0.00823323242366314, + 0.020858265459537506, + 0.014942406676709652, + 0.0062752775847911835, + 0.012439080514013767, + 0.010777028277516365, + 0.01649978570640087, + 0.008967830799520016, + 0.008464605547487736, + 0.009142326191067696, + 0.012672125361859798, + 0.005306575912982225, + 0.006449606269598007, + 0.008231150917708874, + 0.009696022607386112, + 0.01434391364455223, + 0.02049448899924755, + 0.028743663802742958, + 0.03415623679757118, + 0.012836236506700516, + 0.01784161478281021, + 0.018983811140060425, + 0.02621392533183098, + 0.00803866982460022, + 0.007033294998109341, + 0.009185283444821835, + 0.010280661284923553, + 0.005851839669048786, + 0.006868091877549887, + 0.010258989408612251, + 0.00918399728834629, + 0.00701551279053092, + 0.00668324762955308, + 0.008743787184357643, + 0.012970946729183197, + 0.005141918547451496, + 0.007325548212975264, + 0.007646565325558186, + 0.009395942091941833, + 0.005513187497854233, + 0.003579085925593972, + 0.004270700737833977, + 0.004490899853408337, + 0.0028197462670505047, + 0.0034602954983711243, + 0.003255670191720128, + 0.004734380636364222, + 0.0038912296295166016, + 0.003470401279628277, + 0.0037836788687855005, + 0.004315129015594721, + 0.003523536492139101, + 0.0037274740170687437, + 0.00389627437107265, + 0.00395340658724308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.014218100346624851 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640522137284279 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640522137284279, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640522137284279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23374183475971222, + 0.11624215543270111, + 0.05258950591087341, + 0.07849813252687454, + 0.09640476852655411, + 0.21455654501914978, + 0.10431282967329025, + 0.028955455869436264, + 0.03389885276556015, + 0.04079994186758995 + ], + "credit_map": [ + 0.025772439315915108, + 0.020314432680606842, + 0.037789639085531235, + 0.04314769059419632, + 0.016775041818618774, + 0.02618444710969925, + 0.028818754479289055, + 0.03493936359882355, + 0.011973288841545582, + 0.022692333906888962, + 0.016656238585710526, + 0.015851112082600594, + 0.010295913554728031, + 0.013847765512764454, + 0.010164458304643631, + 0.014761042781174183, + 0.010251014493405819, + 0.006873298902064562, + 0.006907568778842688, + 0.008191455155611038, + 0.0034002193715423346, + 0.005231776274740696, + 0.0044687045738101006, + 0.007265464868396521, + 0.008763259276747704, + 0.010931838303804398, + 0.015094864182174206, + 0.010560661554336548, + 0.005179471801966429, + 0.00866042822599411, + 0.0076349033042788506, + 0.011672694236040115, + 0.008492900058627129, + 0.01008699182420969, + 0.017164457589387894, + 0.019613761454820633, + 0.00647733686491847, + 0.00872368086129427, + 0.011088373139500618, + 0.014757250435650349, + 0.026794984936714172, + 0.018792107701301575, + 0.03603952005505562, + 0.03995629400014877, + 0.018212763592600822, + 0.02528209798038006, + 0.021852048113942146, + 0.027626723051071167, + 0.012942393310368061, + 0.010482171550393105, + 0.01866932027041912, + 0.01735648885369301, + 0.006849825382232666, + 0.011020008474588394, + 0.01305205374956131, + 0.013940555974841118, + 0.0027747100684791803, + 0.0025124293752014637, + 0.004831045866012573, + 0.0051500266417860985, + 0.002927417168393731, + 0.002965850755572319, + 0.0033400659449398518, + 0.004453904926776886, + 0.004536479711532593, + 0.00299207866191864, + 0.005810555070638657, + 0.005256111267954111, + 0.0027807727456092834, + 0.004214225336909294, + 0.0037704864516854286, + 0.004538143519312143, + 0.004619147162884474, + 0.004080817103385925, + 0.0065782079473137856, + 0.006913746241480112, + 0.00408853217959404, + 0.004548203200101852, + 0.004575850907713175, + 0.005395435728132725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.008640522137284279 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901678025722504 + ], + "opposite_logprob_tokens": [ + -2.4640166759490967 + ], + "logprob_reward": -0.08901678025722504, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901678025722504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08563683927059174, + 0.033975131809711456, + 0.04428552836179733, + 0.13395461440086365, + 0.3394302427768707, + 0.2195531725883484, + 0.09018970280885696, + 0.020424600690603256, + 0.01349143497645855, + 0.019058741629123688 + ], + "credit_map": [ + 0.01513037458062172, + 0.010869363322854042, + 0.013772341422736645, + 0.01081860065460205, + 0.005894645117223263, + 0.008388768881559372, + 0.008433866314589977, + 0.01232887338846922, + 0.0039319004863500595, + 0.0035906736738979816, + 0.005812152288854122, + 0.004951469600200653, + 0.0021356248762458563, + 0.004799740854650736, + 0.0038436432369053364, + 0.004909925162792206, + 0.007681563496589661, + 0.005323967430740595, + 0.007624539081007242, + 0.006353843025863171, + 0.0029376919846981764, + 0.0038715482223778963, + 0.004570625722408295, + 0.005921749398112297, + 0.022411108016967773, + 0.012785712257027626, + 0.01968909427523613, + 0.02179696597158909, + 0.007597814779728651, + 0.014554413966834545, + 0.015871936455368996, + 0.019247572869062424, + 0.038256123661994934, + 0.029565749689936638, + 0.07291469722986221, + 0.05410393700003624, + 0.021688256412744522, + 0.02850853092968464, + 0.039626140147447586, + 0.054766833782196045, + 0.022674262523651123, + 0.021218739449977875, + 0.040744706988334656, + 0.0384405218064785, + 0.016203418374061584, + 0.019260108470916748, + 0.027941932901740074, + 0.03306945785880089, + 0.010422474704682827, + 0.010035919025540352, + 0.014490587636828423, + 0.013410360552370548, + 0.006684460677206516, + 0.010408827103674412, + 0.012256326153874397, + 0.012480748817324638, + 0.001857819384895265, + 0.0019866516813635826, + 0.0033476895187050104, + 0.0036680835764855146, + 0.0018804647261276841, + 0.0020721855107694864, + 0.002535621402785182, + 0.0030760846566408873, + 0.0018749161390587687, + 0.0010639624670147896, + 0.0022213077172636986, + 0.0023259080480784178, + 0.0008597976411692798, + 0.0015967824729159474, + 0.0016478653997182846, + 0.0019008965464308858, + 0.0022507307585328817, + 0.0018231808207929134, + 0.0029195952229201794, + 0.003117692656815052, + 0.0016430882969871163, + 0.0022183407563716173, + 0.0024487730115652084, + 0.0026373399887233973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.08901678025722504 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.2379255294799805 + ], + "opposite_logprob_tokens": [ + -0.11292542517185211 + ], + "logprob_reward": -2.2379255294799805, + "yes_no_margin_reward": -2.125, + "token_credit_reward": -2.2379255294799805, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22951962053775787, + 0.12193302065134048, + 0.10398811101913452, + 0.037108223885297775, + 0.17505422234535217, + 0.1055612564086914, + 0.09829690307378769, + 0.08700114488601685, + 0.017903873696923256, + 0.02363366261124611 + ], + "credit_map": [ + 0.025868289172649384, + 0.021756451576948166, + 0.029553156346082687, + 0.04001066833734512, + 0.013725544326007366, + 0.03359442949295044, + 0.034322816878557205, + 0.030688272789120674, + 0.01641174405813217, + 0.013529001735150814, + 0.021219264715909958, + 0.016358966007828712, + 0.008563297800719738, + 0.015577079728245735, + 0.01385699026286602, + 0.01641668565571308, + 0.017957160249352455, + 0.01033724658191204, + 0.0167531855404377, + 0.017323479056358337, + 0.005323668476194143, + 0.010289349593222141, + 0.010880202054977417, + 0.015123819932341576, + 0.005295594688504934, + 0.0045537203550338745, + 0.005939392372965813, + 0.005853443406522274, + 0.002494597341865301, + 0.0030407090671360493, + 0.004117878153920174, + 0.005812888499349356, + 0.01809917762875557, + 0.01450053695589304, + 0.029175404459238052, + 0.034558702260255814, + 0.012276978231966496, + 0.01446376834064722, + 0.02446851134300232, + 0.027511129155755043, + 0.011247668415307999, + 0.009095950052142143, + 0.01883845217525959, + 0.01994178257882595, + 0.007280395831912756, + 0.009522724896669388, + 0.013865922577679157, + 0.01576835848391056, + 0.009125222451984882, + 0.009501543827354908, + 0.016052082180976868, + 0.01786614954471588, + 0.007252352312207222, + 0.009899678640067577, + 0.013971908017992973, + 0.014627961441874504, + 0.008125261403620243, + 0.008405489847064018, + 0.013709425926208496, + 0.01573030836880207, + 0.006676821503788233, + 0.010141425766050816, + 0.012209370732307434, + 0.012003039009869099, + 0.002073572250083089, + 0.0015939988661557436, + 0.003369611222296953, + 0.003147370647639036, + 0.0009367981110699475, + 0.002040949882939458, + 0.0018699908396229148, + 0.0028715813532471657, + 0.003584358375519514, + 0.0022619334049522877, + 0.004177414812147617, + 0.003861835924908519, + 0.0011996201938018203, + 0.0024406537413597107, + 0.00263350922614336, + 0.003474337048828602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -2.2379255294799805 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "highly realistic, full-body static shot, captured from a slight overhead angle, captures a shirtless man lying on his side on dark sheets in a dimly lit room. His back is sharply defined, his muscles barely visible in the gloom, and a fine sheen of sweat covers his skin. The only warm light in the scene emanates from a slowly burning moxa stick in the center of his back. Its fiery red tip casts a small, concentrated, warm, orange-yellow glow, illuminating the surrounding skin with a translucent glow. As his chest and back rise and fall rhythmically with his steady, gentle breathing, wisps of soft white moxa smoke curl upward from the stick, swirling and dissipating in the still air, illuminated by the single, warm light, forming visible beams of light. This faint light creates a stark contrast with the surrounding darkness, profoundly sculpting the curves and shadows of his body, creating a paradoxical atmosphere of both healing tranquility and a subtle sense of endurance. The entire image is imbued with a dramatic interplay of light and shadow.", + "seed": 13133, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_5/index_2782_seed_13133_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1553719975054264, + "vlm_logprob": -0.1553719975054264, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.1553719975054264, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01858324185013771 + ], + "opposite_logprob_tokens": [ + -4.143583297729492 + ], + "logprob_reward": -0.01858324185013771, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01858324185013771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.929121196269989, + 0.00941389799118042, + 0.005452387034893036, + 0.011869380250573158, + 0.005079479888081551, + 0.0073895822279155254, + 0.0038787091616541147, + 0.0024390078615397215, + 0.018245616927742958, + 0.007110722362995148 + ], + "credit_map": [ + 0.01520539727061987, + 0.01712937094271183, + 0.041265398263931274, + 0.0334196574985981, + 0.0207462627440691, + 0.042066629976034164, + 0.37630823254585266, + 0.38298019766807556, + 0.0007365396013483405, + 0.0009398245019838214, + 0.0010100112995132804, + 0.0009022812591865659, + 0.0005948719335719943, + 0.0005089336191304028, + 0.0023650869261473417, + 0.0023563485592603683, + 0.00040751739288680255, + 0.0007657127571292222, + 0.0005255544674582779, + 0.0008304683724418283, + 0.0009450698271393776, + 0.0006299970555119216, + 0.0006847383338026702, + 0.000663328857626766, + 0.0017104934668168426, + 0.0013317106058821082, + 0.0020532445050776005, + 0.001958305248990655, + 0.0008349631098099053, + 0.0011119323316961527, + 0.0013074028538540006, + 0.0015613286523148417, + 0.0006292493781074882, + 0.0005545590538531542, + 0.0006807415629737079, + 0.0007666907040402293, + 0.00047601229744032025, + 0.0005250244867056608, + 0.0005498614045791328, + 0.0008973408839665353, + 0.0006276500644162297, + 0.0006264978437684476, + 0.0007741129375062883, + 0.0012776406947523355, + 0.0007217502570711076, + 0.0007073741871863604, + 0.0010809595696628094, + 0.0015735970810055733, + 0.00019969332788605243, + 0.00015965373313520104, + 0.00023798314214218408, + 0.00018672352598514408, + 0.00015890369832050055, + 0.00022383018222171813, + 0.001448550377972424, + 0.0012633712030947208, + 0.0003178720362484455, + 0.00034618060453794897, + 0.00030677008908241987, + 0.00026766350492835045, + 0.00029038204229436815, + 0.0002891199546866119, + 0.0002969693159684539, + 0.0003240501682739705, + 0.0021000595297664404, + 0.0024205800145864487, + 0.0030021416023373604, + 0.0022814832627773285, + 0.0018769901944324374, + 0.001963872229680419, + 0.002384563907980919, + 0.0022159249056130648, + 0.00048167110071517527, + 0.0005197396385483444, + 0.0006170130800455809, + 0.0007116070482879877, + 0.0005924669676460326, + 0.0006568650715053082, + 0.0021157509181648493, + 0.0014156085671857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.01858324185013771 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0110928975045681 + ], + "opposite_logprob_tokens": [ + -4.511092662811279 + ], + "logprob_reward": -0.0110928975045681, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0110928975045681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5417884588241577, + 0.32269054651260376, + 0.0297729279845953, + 0.007262448780238628, + 0.006828621495515108, + 0.011148384772241116, + 0.012607035227119923, + 0.006443139165639877, + 0.02703792043030262, + 0.03442058712244034 + ], + "credit_map": [ + 0.010640904307365417, + 0.011905926279723644, + 0.025830699130892754, + 0.022351650521159172, + 0.01238708570599556, + 0.02392737567424774, + 0.2149612009525299, + 0.21978358924388885, + 0.005137468222528696, + 0.006025856826454401, + 0.01405156310647726, + 0.009698630310595036, + 0.007997473701834679, + 0.014738318510353565, + 0.13088472187519073, + 0.1341564953327179, + 0.0019206807482987642, + 0.0022812748793512583, + 0.0020153976511210203, + 0.0029923126567155123, + 0.0022476394660770893, + 0.001825303421355784, + 0.008112790994346142, + 0.008377525955438614, + 0.0009243916720151901, + 0.0009143778588622808, + 0.0012099031591787934, + 0.000998621922917664, + 0.0006647795671597123, + 0.0006284979754127562, + 0.0009827397298067808, + 0.0009391368948854506, + 0.000787924334872514, + 0.0007409550016745925, + 0.0008648686925880611, + 0.0006752461194992065, + 0.0005849751178175211, + 0.0006402501603588462, + 0.0011387644335627556, + 0.0013956372858956456, + 0.0009657572954893112, + 0.001115954015403986, + 0.0016735922545194626, + 0.0015664086677134037, + 0.0016539276111871004, + 0.0008872422040440142, + 0.0015280329389497638, + 0.001757470890879631, + 0.0010015740990638733, + 0.0010017218301072717, + 0.0011216525454074144, + 0.0009813524084165692, + 0.0007142076501622796, + 0.0009159998735412955, + 0.00386717077344656, + 0.0030033562798053026, + 0.0006214992026798427, + 0.000813354563433677, + 0.0008694338030181825, + 0.0007638594252057374, + 0.0006799811962991953, + 0.0007376885623671114, + 0.0010054147569462657, + 0.0009519074810668826, + 0.0029979085084050894, + 0.003763213288038969, + 0.0029396633617579937, + 0.002438232069835067, + 0.002830470446497202, + 0.00280013051815331, + 0.0044213528744876385, + 0.004846947733312845, + 0.002931599970906973, + 0.0030603345949202776, + 0.0035295335110276937, + 0.00460481783375144, + 0.0031564764212816954, + 0.0033471444621682167, + 0.00803433358669281, + 0.00575634278357029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0110928975045681 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253676325082779 + ], + "opposite_logprob_tokens": [ + -5.255253791809082 + ], + "logprob_reward": -0.005253676325082779, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253676325082779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9574286937713623, + 0.013262229040265083, + 0.004811575170606375, + 0.004258269909769297, + 0.002967286854982376, + 0.001987869618460536, + 0.0033441022969782352, + 0.001428465242497623, + 0.004786410368978977, + 0.005725019611418247 + ], + "credit_map": [ + 0.015024675987660885, + 0.01743408478796482, + 0.042326003313064575, + 0.03341102600097656, + 0.02083154395222664, + 0.04270600900053978, + 0.3893275260925293, + 0.3963679373264313, + 0.0007977609639056027, + 0.001036741305142641, + 0.000784409639891237, + 0.0008267408702522516, + 0.0005127649055793881, + 0.0008251869003288448, + 0.004345771390944719, + 0.004132853355258703, + 0.0004168007289990783, + 0.0007750768563710153, + 0.0005078426329419017, + 0.0007452211575582623, + 0.000535276485607028, + 0.00043171102879568934, + 0.0007013487629592419, + 0.0006982985069043934, + 0.0004169628955423832, + 0.00033349861041642725, + 0.0005804080865345895, + 0.00044094747863709927, + 0.00028502411441877484, + 0.00030133017571642995, + 0.0009429076453670859, + 0.0009571918053552508, + 0.0004105450352653861, + 0.0003473359683994204, + 0.0004866359813604504, + 0.0003679836227092892, + 0.00025431750691495836, + 0.0003232814196962863, + 0.0002611021918710321, + 0.0005160856526345015, + 0.00017804448725655675, + 0.0002388130669714883, + 0.00024207831302192062, + 0.0002478739188518375, + 0.0002745537494774908, + 0.00021915019897278398, + 0.0002131100627593696, + 0.0003742461558431387, + 0.0002953869989141822, + 0.00037826967309229076, + 0.0003218807396478951, + 0.00026859575882554054, + 0.00022642083058599383, + 0.00021269639546517283, + 0.0008705708314664662, + 0.0007702813018113375, + 0.0001820534816943109, + 0.00018759600061457604, + 0.00022901169722899795, + 0.00015493722457904369, + 0.0001518450299045071, + 0.0001448148541385308, + 0.00017169531201943755, + 0.0002065118751488626, + 0.0004896385944448411, + 0.0005591784720309079, + 0.000754701963160187, + 0.0006744916900061071, + 0.00044921666267327964, + 0.0004682753933593631, + 0.0007044100784696639, + 0.0006864983588457108, + 0.000623648171313107, + 0.0006970033864490688, + 0.0007512528682127595, + 0.0010123298270627856, + 0.0004440480552148074, + 0.0005787833943031728, + 0.000966887513641268, + 0.000651067472063005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.005253676325082779 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863445460796356 + ], + "opposite_logprob_tokens": [ + -3.0486345291137695 + ], + "logprob_reward": -0.04863445460796356, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863445460796356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9023857712745667, + 0.04341147840023041, + 0.009839226491749287, + 0.0036301943473517895, + 0.003060897346585989, + 0.002835166174918413, + 0.002187411766499281, + 0.0029237635899335146, + 0.011357451789081097, + 0.01836874708533287 + ], + "credit_map": [ + 0.013621463440358639, + 0.016441864892840385, + 0.040056321769952774, + 0.03157956153154373, + 0.020144881680607796, + 0.0401749387383461, + 0.3667256534099579, + 0.3736410439014435, + 0.0007469706470146775, + 0.0009303062688559294, + 0.0019121031509712338, + 0.0013453729916363955, + 0.0010735118994489312, + 0.001984280301257968, + 0.017501715570688248, + 0.017917215824127197, + 0.0003397527907509357, + 0.0004897608305327594, + 0.0005488815950229764, + 0.0006551950937137008, + 0.00041859192424453795, + 0.0005623014294542372, + 0.00346559751778841, + 0.0033591436222195625, + 0.00045937797403894365, + 0.0004321913584135473, + 0.0004654311924241483, + 0.00037236130447126925, + 0.0002709754044190049, + 0.0003172325377818197, + 0.0007131041493266821, + 0.0005995203391648829, + 0.0005346197285689414, + 0.00033831020118668675, + 0.00042746023973450065, + 0.0003203479282092303, + 0.0002549722557887435, + 0.0003499264130368829, + 0.00035964747075922787, + 0.00047561287647113204, + 0.00038919455255381763, + 0.0002410285669611767, + 0.0004198767419438809, + 0.0004025397065561265, + 0.00022455205908045173, + 0.00026270412490703166, + 0.00034633761970326304, + 0.0005489324685186148, + 0.00022509951668325812, + 0.00026614891248755157, + 0.00024498775019310415, + 0.00026588246691972017, + 0.0001314729597652331, + 0.00018450507195666432, + 0.0005004283739253879, + 0.00036888671456836164, + 0.0003034192486666143, + 0.0002995382237713784, + 0.0004075815959367901, + 0.00032760962494648993, + 0.0002561774745117873, + 0.0002354592870688066, + 0.0004939486389048398, + 0.0006000293651595712, + 0.0013276971876621246, + 0.001486675115302205, + 0.0018284375546500087, + 0.0017526608426123857, + 0.001189459697343409, + 0.0011430417653173208, + 0.0013588668080046773, + 0.0012706120032817125, + 0.0014396596234291792, + 0.0015556362923234701, + 0.0020366343669593334, + 0.0024320371448993683, + 0.0012045811163261533, + 0.0015628706896677613, + 0.004649418406188488, + 0.0034879082813858986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.04863445460796356 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6932957172393799 + ], + "opposite_logprob_tokens": [ + -0.6932957172393799 + ], + "logprob_reward": -0.6932957172393799, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6932957172393799, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.8536316752433777, + 0.04027579352259636, + 0.011511496268212795, + 0.007104115094989538, + 0.012659118510782719, + 0.007746867369860411, + 0.004014207981526852, + 0.006952066905796528, + 0.021179212257266045, + 0.03492550179362297 + ], + "credit_map": [ + 0.013741015456616879, + 0.01606079563498497, + 0.038268111646175385, + 0.031087245792150497, + 0.02017766237258911, + 0.039003968238830566, + 0.3450029492378235, + 0.3502899408340454, + 0.0008723788778297603, + 0.00100790208671242, + 0.0018692773301154375, + 0.0013326024636626244, + 0.0015920806908980012, + 0.0018441923893988132, + 0.015553980134427547, + 0.0162033773958683, + 0.0004455633752513677, + 0.000538928434252739, + 0.0007385795470327139, + 0.000815604638773948, + 0.0004716154362540692, + 0.0006269411533139646, + 0.004020733758807182, + 0.0038535294588655233, + 0.0007973863976076245, + 0.0011726075317710638, + 0.001044736010953784, + 0.0006053774268366396, + 0.00040643365355208516, + 0.0005466133006848395, + 0.0011752852005884051, + 0.0013556755147874355, + 0.0011226341594010592, + 0.001210627844557166, + 0.0015995396533980966, + 0.0019236619118601084, + 0.0014016730710864067, + 0.0013214648934081197, + 0.001671645906753838, + 0.0024078702554106712, + 0.0010373616823926568, + 0.0006886192713864148, + 0.0016916540917009115, + 0.0007561575039289892, + 0.0007357727736234665, + 0.0006577394669875503, + 0.0006827049655839801, + 0.0014968572650104761, + 0.0004108219873160124, + 0.0002775366883724928, + 0.0004482374060899019, + 0.0004835530708078295, + 0.00026845987304113805, + 0.00028723260038532317, + 0.0009257334168069065, + 0.0009126327931880951, + 0.0006147493259049952, + 0.0010188219603151083, + 0.0008370650466531515, + 0.0009196252794936299, + 0.0008848085417412221, + 0.0006061200401745737, + 0.0010268395999446511, + 0.0010440367041155696, + 0.0023703037295490503, + 0.002765645971521735, + 0.003255255054682493, + 0.003300176002085209, + 0.002362046856433153, + 0.0023155068047344685, + 0.002555529121309519, + 0.002254749182611704, + 0.0026095181237906218, + 0.0028634187765419483, + 0.0036339303478598595, + 0.004091473761945963, + 0.0020780512131750584, + 0.002405282575637102, + 0.0095788249745965, + 0.007665000390261412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.6932957172393799 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "highly realistic, full-body static shot, captured from a slight overhead angle, captures a shirtless man lying on his side on dark sheets in a dimly lit room. His back is sharply defined, his muscles barely visible in the gloom, and a fine sheen of sweat covers his skin. The only warm light in the scene emanates from a slowly burning moxa stick in the center of his back. Its fiery red tip casts a small, concentrated, warm, orange-yellow glow, illuminating the surrounding skin with a translucent glow. As his chest and back rise and fall rhythmically with his steady, gentle breathing, wisps of soft white moxa smoke curl upward from the stick, swirling and dissipating in the still air, illuminated by the single, warm light, forming visible beams of light. This faint light creates a stark contrast with the surrounding darkness, profoundly sculpting the curves and shadows of his body, creating a paradoxical atmosphere of both healing tranquility and a subtle sense of endurance. The entire image is imbued with a dramatic interplay of light and shadow.", + "seed": 13134, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_5/index_2782_seed_13134_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3961295035202056, + "vlm_logprob": -0.3961295035202056, + "vlm_yes_no_margin": 2.975000023841858, + "vlm_token_credit": -0.3961295035202056, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023090878501534462 + ], + "opposite_logprob_tokens": [ + -3.8980908393859863 + ], + "logprob_reward": -0.023090878501534462, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.023090878501534462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3046164810657501, + 0.08338293433189392, + 0.051309697329998016, + 0.07059921324253082, + 0.05306543782353401, + 0.13191303610801697, + 0.10916335135698318, + 0.07492092251777649, + 0.06311674416065216, + 0.0579121895134449 + ], + "credit_map": [ + 0.02639835514128208, + 0.03728102892637253, + 0.04378838464617729, + 0.04704449325799942, + 0.033846307545900345, + 0.02292943000793457, + 0.04162415489554405, + 0.05170430615544319, + 0.008307455107569695, + 0.007716445252299309, + 0.013916829600930214, + 0.013275379315018654, + 0.010560838505625725, + 0.005872465670108795, + 0.012245161458849907, + 0.011488350108265877, + 0.004586567170917988, + 0.0065859477035701275, + 0.006562971044331789, + 0.006790085230022669, + 0.008023967035114765, + 0.006644741166383028, + 0.0066181062720716, + 0.005497312173247337, + 0.004658751655369997, + 0.009725760668516159, + 0.009855563752353191, + 0.009059679694473743, + 0.006344950292259455, + 0.006903707515448332, + 0.010748002678155899, + 0.013302790001034737, + 0.006427744869142771, + 0.005255429539829493, + 0.006585162132978439, + 0.010120091028511524, + 0.004352468065917492, + 0.004375217016786337, + 0.007637250702828169, + 0.008312071673572063, + 0.019007408991456032, + 0.018676530569791794, + 0.01737031526863575, + 0.015164993703365326, + 0.016344891861081123, + 0.01577889733016491, + 0.017137549817562103, + 0.012432441115379333, + 0.016201715916395187, + 0.00990102719515562, + 0.015325537882745266, + 0.013789012096822262, + 0.012933836318552494, + 0.012447452172636986, + 0.015356083400547504, + 0.013208678923547268, + 0.006794745102524757, + 0.009196851402521133, + 0.01220779586583376, + 0.00969342514872551, + 0.007006481289863586, + 0.007228203117847443, + 0.011735639534890652, + 0.011057774536311626, + 0.011919981800019741, + 0.007030249107629061, + 0.007552185095846653, + 0.006433185655623674, + 0.007540506310760975, + 0.005032998509705067, + 0.00895413663238287, + 0.008653493598103523, + 0.008116334676742554, + 0.004434000235050917, + 0.011249811388552189, + 0.0076923007145524025, + 0.005375155247747898, + 0.004316335078328848, + 0.009462226182222366, + 0.007266018074005842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.023090878501534462 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00320504535920918 + ], + "opposite_logprob_tokens": [ + -5.753204822540283 + ], + "logprob_reward": -0.00320504535920918, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00320504535920918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20714490115642548, + 0.1909984052181244, + 0.07945311814546585, + 0.18096087872982025, + 0.0651073157787323, + 0.04847349599003792, + 0.049412988126277924, + 0.07121654599905014, + 0.041915539652109146, + 0.06531680375337601 + ], + "credit_map": [ + 0.0373651459813118, + 0.01704331673681736, + 0.025873925536870956, + 0.03396474942564964, + 0.015697719529271126, + 0.018304182216525078, + 0.026473572477698326, + 0.03242228925228119, + 0.031366392970085144, + 0.017426623031497, + 0.02187171019613743, + 0.027331937104463577, + 0.019088316708803177, + 0.021536897867918015, + 0.02385486289858818, + 0.028521666303277016, + 0.008413022384047508, + 0.009409240446984768, + 0.012600894086062908, + 0.011675681918859482, + 0.012696318328380585, + 0.0060820262879133224, + 0.011023804545402527, + 0.0075521320104599, + 0.018111106008291245, + 0.017917808145284653, + 0.02370961755514145, + 0.02124014124274254, + 0.021967386826872826, + 0.025318369269371033, + 0.023915929719805717, + 0.02878051996231079, + 0.0064468770287930965, + 0.00822160392999649, + 0.00803830660879612, + 0.008243480697274208, + 0.009742624126374722, + 0.007950276136398315, + 0.00904060434550047, + 0.007423541508615017, + 0.008420058526098728, + 0.005151643883436918, + 0.006344135385006666, + 0.00565289007499814, + 0.004717828240245581, + 0.004469596780836582, + 0.00633307034149766, + 0.00738427322357893, + 0.01075141504406929, + 0.006000433582812548, + 0.007369530852884054, + 0.0052812653593719006, + 0.0037436687853187323, + 0.004757555201649666, + 0.005865716841071844, + 0.00564340827986598, + 0.01033490989357233, + 0.007830338552594185, + 0.008402975276112556, + 0.006801454816013575, + 0.013620407320559025, + 0.007167877163738012, + 0.009369348175823689, + 0.007689237594604492, + 0.0068456437438726425, + 0.005483720451593399, + 0.005476390942931175, + 0.004640188999474049, + 0.007092582993209362, + 0.003967540804296732, + 0.005101747810840607, + 0.0033077255357056856, + 0.012864756397902966, + 0.006432740017771721, + 0.008210328407585621, + 0.0077059208415448666, + 0.00791867170482874, + 0.006951279938220978, + 0.00844394788146019, + 0.006789159029722214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.00320504535920918 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735478527843952 + ], + "opposite_logprob_tokens": [ + -5.006735324859619 + ], + "logprob_reward": -0.006735478527843952, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735478527843952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2665322721004486, + 0.10502508282661438, + 0.07909271866083145, + 0.2549678087234497, + 0.04361213743686676, + 0.04652735963463783, + 0.05109630152583122, + 0.04562980681657791, + 0.059457164257764816, + 0.04805934801697731 + ], + "credit_map": [ + 0.028034934774041176, + 0.04136461392045021, + 0.04684692993760109, + 0.02806444652378559, + 0.025074463337659836, + 0.022100578993558884, + 0.038983628153800964, + 0.03606267273426056, + 0.012918232940137386, + 0.011476312763988972, + 0.017762741073966026, + 0.013167162425816059, + 0.010930442251265049, + 0.008236393332481384, + 0.015048711560666561, + 0.015485083684325218, + 0.008089732378721237, + 0.008019162341952324, + 0.011126512661576271, + 0.01348235085606575, + 0.007209651172161102, + 0.007707201410084963, + 0.011425522156059742, + 0.012032588012516499, + 0.03511277213692665, + 0.05339842289686203, + 0.027030542492866516, + 0.026279034093022346, + 0.025748278945684433, + 0.019412094727158546, + 0.03098350763320923, + 0.03700314089655876, + 0.004745745565742254, + 0.004917565733194351, + 0.006016077473759651, + 0.006727383937686682, + 0.005489506758749485, + 0.003629491664469242, + 0.006096623372286558, + 0.0059897406026721, + 0.004883476532995701, + 0.008342381566762924, + 0.007210413459688425, + 0.005684728734195232, + 0.00449481001123786, + 0.002840023720636964, + 0.007147899363189936, + 0.005923631135374308, + 0.008355388417840004, + 0.0048253801651299, + 0.006600662134587765, + 0.005536985583603382, + 0.003549233777448535, + 0.005742251873016357, + 0.007325195707380772, + 0.009161202237010002, + 0.004749794490635395, + 0.005070917308330536, + 0.007548678200691938, + 0.007545220199972391, + 0.00396113283932209, + 0.0038416171446442604, + 0.007019686978310347, + 0.00589276198297739, + 0.010142681188881397, + 0.0068749370984733105, + 0.008593831211328506, + 0.006182653363794088, + 0.00617388729006052, + 0.004807730205357075, + 0.008990656584501266, + 0.00769078591838479, + 0.008166464976966381, + 0.002790325554087758, + 0.006026944145560265, + 0.0043191625736653805, + 0.004185694269835949, + 0.005704633891582489, + 0.007742370944470167, + 0.009123750030994415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.006735478527843952 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14275921881198883 + ], + "opposite_logprob_tokens": [ + -2.017759323120117 + ], + "logprob_reward": -0.14275921881198883, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14275921881198883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11755787581205368, + 0.08311222493648529, + 0.058064378798007965, + 0.08695466071367264, + 0.05431004613637924, + 0.031219372525811195, + 0.07461698353290558, + 0.18511070311069489, + 0.2910136878490448, + 0.018040021881461143 + ], + "credit_map": [ + 0.026654763147234917, + 0.009929770603775978, + 0.0167122520506382, + 0.014216400682926178, + 0.007124233525246382, + 0.00959154311567545, + 0.01586444117128849, + 0.017464470118284225, + 0.007821238599717617, + 0.009947235696017742, + 0.011554460041224957, + 0.01530926488339901, + 0.00896770041435957, + 0.008547479286789894, + 0.011690213344991207, + 0.009274632669985294, + 0.005512037314474583, + 0.0054382868111133575, + 0.007865017279982567, + 0.006295838858932257, + 0.007988878525793552, + 0.007059735246002674, + 0.007451139856129885, + 0.010453442111611366, + 0.0052358354441821575, + 0.009775341488420963, + 0.011834395118057728, + 0.01272729504853487, + 0.008688624948263168, + 0.009086843580007553, + 0.013409048318862915, + 0.01619727723300457, + 0.004259465262293816, + 0.00616683904081583, + 0.007130173500627279, + 0.008158953860402107, + 0.007105180993676186, + 0.005643569398671389, + 0.008974139578640461, + 0.006871727295219898, + 0.0031949500553309917, + 0.003875916823744774, + 0.004909294657409191, + 0.003993207588791847, + 0.0031458870507776737, + 0.002833752194419503, + 0.005059482995420694, + 0.004206882789731026, + 0.008040521293878555, + 0.015313807874917984, + 0.010072450153529644, + 0.0066747101955115795, + 0.01173386350274086, + 0.005324780009686947, + 0.008740645833313465, + 0.008716202341020107, + 0.015464349649846554, + 0.021484091877937317, + 0.02427057735621929, + 0.01982559822499752, + 0.021681219339370728, + 0.01499930676072836, + 0.028759144246578217, + 0.038626424968242645, + 0.05212385207414627, + 0.029491443186998367, + 0.03527902066707611, + 0.028876930475234985, + 0.02728782221674919, + 0.021612150594592094, + 0.04330844059586525, + 0.05303402245044708, + 0.0015611628768965602, + 0.0017763854702934623, + 0.0027258959598839283, + 0.002435623435303569, + 0.0017033610492944717, + 0.0021154822316020727, + 0.002901833038777113, + 0.0028202789835631847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.14275921881198883 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a shirtless man lying on his side on dark sheets visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moxa stick with a fiery red tip positioned on the man's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wisps of soft white moxa smoke curling upward from the stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's chest and back rise and fall rhythmically with steady, gentle breathing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle a slight overhead shot capturing a full-body view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.8048568964004517 + ], + "opposite_logprob_tokens": [ + -0.17985689640045166 + ], + "logprob_reward": -1.8048568964004517, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.8048568964004517, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3781103789806366, + 0.15816724300384521, + 0.05485033616423607, + 0.0397060327231884, + 0.0643218383193016, + 0.14340117573738098, + 0.07789965718984604, + 0.042257051914930344, + 0.022089146077632904, + 0.019197136163711548 + ], + "credit_map": [ + 0.031997907906770706, + 0.03262770548462868, + 0.06818249076604843, + 0.05774129554629326, + 0.03165234997868538, + 0.03534933924674988, + 0.061637263745069504, + 0.05892200395464897, + 0.013615605421364307, + 0.015650039538741112, + 0.03093341551721096, + 0.030975064262747765, + 0.018110204488039017, + 0.010253192856907845, + 0.022447848692536354, + 0.01618187688291073, + 0.004329999443143606, + 0.005444149021059275, + 0.011421812698245049, + 0.009265328757464886, + 0.0042757210321724415, + 0.004301025532186031, + 0.008113890886306763, + 0.007698410656303167, + 0.0026713937986642122, + 0.0033202318008989096, + 0.007179107517004013, + 0.0060219778679311275, + 0.004125438630580902, + 0.004238090477883816, + 0.005886717699468136, + 0.006263072602450848, + 0.005268865264952183, + 0.005926985293626785, + 0.010406960733234882, + 0.009264249354600906, + 0.007772868499159813, + 0.005623514764010906, + 0.010406199842691422, + 0.009652198292315006, + 0.013024215586483479, + 0.01646280474960804, + 0.02196350507438183, + 0.02104179933667183, + 0.018872177228331566, + 0.01687208004295826, + 0.018303316086530685, + 0.01686127297580242, + 0.009020315483212471, + 0.007743501104414463, + 0.01166749931871891, + 0.010607721284031868, + 0.008147569373250008, + 0.01086423173546791, + 0.010290421545505524, + 0.00955839641392231, + 0.004538616165518761, + 0.006391990464180708, + 0.006540379952639341, + 0.006338462233543396, + 0.002953913062810898, + 0.004659589845687151, + 0.005593717098236084, + 0.005240378901362419, + 0.0021800948306918144, + 0.00264555960893631, + 0.003102980088442564, + 0.003438633633777499, + 0.0023993186186999083, + 0.0022757300175726414, + 0.003183453343808651, + 0.002863375935703516, + 0.0020216028206050396, + 0.0019564987160265446, + 0.003527830820530653, + 0.0027637758757919073, + 0.0016974401660263538, + 0.0016772844828665257, + 0.0029179558623582125, + 0.0026347474195063114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -1.8048568964004517 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_6/index_tensor([8646, 8646])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_6/index_tensor([8646, 8646])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6f73f69ec168f619be6c06c4428f32489332a2bd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_6/index_tensor([8646, 8646])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a photorealistic style, a very slow, eye-level tracking shot follows a majestic Pacific green sea turtle as it feeds along the sun-drenched seafloor. Bright sunlight penetrates the crystal-clear turquoise water, creating a mesmerizing, ever-changing network of bright, shimmering caustics that dance across the turtle's massive, heart-shaped carapace—a mosaic of olive green and brown scales with a realistic texture. This ancient reptile, with its wrinkled skin and intelligent, dark eyes conveying a sense of timelessness, moves with ponderous grace. Lowering its head, it delicately and delicately uses its powerful, keratinous beak to carefully nibble at a patch of lush, vibrant seagrass rooted in the pristine white sand below, stirring up a small cloud of sand with each bite. The turtle is completely oblivious to the schools of colorful small fish, primarily yellow and blue sockeye, that frolic around its massive body. The entire scene is filled with a profoundly serene atmosphere and tranquil natural beauty, and the camera’s steady, unobtrusive movements make the viewer feel like a silent observer in this tranquil underwater world.", + "seed": 8039, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_6/index_8646_seed_8039_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19083626940846443, + "vlm_logprob": -0.19083626940846443, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.19083626940846443, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11640699952840805 + ], + "opposite_logprob_tokens": [ + -2.2414069175720215 + ], + "logprob_reward": -0.11640699952840805, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11640699952840805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12433018535375595, + 0.053586196154356, + 0.02983853593468666, + 0.14022815227508545, + 0.16574521362781525, + 0.042483989149332047, + 0.19648624956607819, + 0.09823069721460342, + 0.10161643475294113, + 0.04745436832308769 + ], + "credit_map": [ + 0.017892587929964066, + 0.00904702115803957, + 0.01867763325572014, + 0.01778663881123066, + 0.012137090787291527, + 0.014784452505409718, + 0.0107662258669734, + 0.0232385266572237, + 0.005783483386039734, + 0.004225385840982199, + 0.008467023260891438, + 0.009472628124058247, + 0.007203346583992243, + 0.006141511257737875, + 0.007688133977353573, + 0.004604677204042673, + 0.0036287230905145407, + 0.002573284786194563, + 0.0031394094694405794, + 0.004747169092297554, + 0.0038248857017606497, + 0.004912408534437418, + 0.00408140430226922, + 0.002931248629465699, + 0.01257727388292551, + 0.011418890208005905, + 0.01666105166077614, + 0.02363630384206772, + 0.014219890348613262, + 0.014738548547029495, + 0.0174657441675663, + 0.02951042912900448, + 0.021089661866426468, + 0.011500714346766472, + 0.017267495393753052, + 0.021585460752248764, + 0.01207992434501648, + 0.013418162241578102, + 0.012648196890950203, + 0.05615558847784996, + 0.004017518367618322, + 0.004586034920066595, + 0.004897065926343203, + 0.0075171818025410175, + 0.0056319572031497955, + 0.004080143291503191, + 0.00733331311494112, + 0.004420770797878504, + 0.023030385375022888, + 0.014816826209425926, + 0.0129069359973073, + 0.07352501899003983, + 0.008560823276638985, + 0.013592316769063473, + 0.030562542378902435, + 0.019491391256451607, + 0.014068360440433025, + 0.004917324520647526, + 0.004703054670244455, + 0.017665987834334373, + 0.006860222201794386, + 0.004210606217384338, + 0.03362656384706497, + 0.012178567238152027, + 0.012040132656693459, + 0.012161480262875557, + 0.02103191614151001, + 0.011712451465427876, + 0.009761303663253784, + 0.013321362435817719, + 0.012360982596874237, + 0.009226796217262745, + 0.006760431453585625, + 0.003779922379180789, + 0.011477258056402206, + 0.006703029852360487, + 0.0033132813405245543, + 0.006154263392090797, + 0.003918676637113094, + 0.005347500555217266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.11640699952840805 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868941009044647 + ], + "opposite_logprob_tokens": [ + -3.048689365386963 + ], + "logprob_reward": -0.04868941009044647, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868941009044647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23058174550533295, + 0.04931396245956421, + 0.03298286348581314, + 0.12218751013278961, + 0.25022366642951965, + 0.0320005789399147, + 0.1020314171910286, + 0.10955767333507538, + 0.02289598435163498, + 0.04822458326816559 + ], + "credit_map": [ + 0.0249327365309, + 0.02341480739414692, + 0.047329504042863846, + 0.04064542427659035, + 0.02314755693078041, + 0.016807084903120995, + 0.030891580507159233, + 0.02341306582093239, + 0.006599719636142254, + 0.004208325874060392, + 0.005576837342232466, + 0.006056701764464378, + 0.008786565624177456, + 0.005519543308764696, + 0.00807225052267313, + 0.004494021646678448, + 0.0032366362866014242, + 0.004076721146702766, + 0.0031097170431166887, + 0.0038502232637256384, + 0.006814764812588692, + 0.004164091777056456, + 0.0048617227002978325, + 0.002868984593078494, + 0.010308992117643356, + 0.014192146249115467, + 0.019844362512230873, + 0.020261868834495544, + 0.018850693479180336, + 0.00994119793176651, + 0.017013641074299812, + 0.011774605140089989, + 0.03357246518135071, + 0.01621721312403679, + 0.020207086578011513, + 0.03163404390215874, + 0.01927962340414524, + 0.021275797858834267, + 0.018958313390612602, + 0.0890791118144989, + 0.0032207295298576355, + 0.0033822141122072935, + 0.003547702683135867, + 0.004034410230815411, + 0.0034334762021899223, + 0.004027127753943205, + 0.0060394215397536755, + 0.0043154992163181305, + 0.008291797712445259, + 0.01152398157864809, + 0.008610855787992477, + 0.022757122293114662, + 0.010265267454087734, + 0.010145424865186214, + 0.011688893660902977, + 0.018748074769973755, + 0.01613045111298561, + 0.010792954824864864, + 0.010801661759614944, + 0.01169202383607626, + 0.009294279851019382, + 0.02626773528754711, + 0.01441243663430214, + 0.010166128166019917, + 0.001644724514335394, + 0.003117837244644761, + 0.002487647347152233, + 0.003910703118890524, + 0.002844934817403555, + 0.0020717994775623083, + 0.0036688712425529957, + 0.0031494670547544956, + 0.008001508191227913, + 0.003520271508023143, + 0.012251583859324455, + 0.0065505001693964005, + 0.0033330630976706743, + 0.006911871954798698, + 0.0029216231778264046, + 0.0047341627068817616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.04868941009044647 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204794719815254 + ], + "opposite_logprob_tokens": [ + -2.8120479583740234 + ], + "logprob_reward": -0.06204794719815254, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204794719815254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1982935220003128, + 0.052801843732595444, + 0.02457890287041664, + 0.05464908480644226, + 0.3538544476032257, + 0.01609368994832039, + 0.0868188887834549, + 0.047015104442834854, + 0.1394984871149063, + 0.02639603614807129 + ], + "credit_map": [ + 0.025263527408242226, + 0.015993744134902954, + 0.033445581793785095, + 0.023847321048378944, + 0.02203615941107273, + 0.022950921207666397, + 0.029031233862042427, + 0.02572501450777054, + 0.00503623578697443, + 0.005572508089244366, + 0.0057962373830378056, + 0.007949481718242168, + 0.010802512057125568, + 0.005944275297224522, + 0.007625992875546217, + 0.004074595402926207, + 0.00246720714494586, + 0.0023919017985463142, + 0.003198951715603471, + 0.0025373538956046104, + 0.004517592489719391, + 0.003998942207545042, + 0.003175537334755063, + 0.0022914146538823843, + 0.004749201238155365, + 0.008088680915534496, + 0.005282262340188026, + 0.00842010136693716, + 0.009834564290940762, + 0.00426911748945713, + 0.00966216903179884, + 0.004342981148511171, + 0.04626110568642616, + 0.03049985133111477, + 0.037746671587228775, + 0.04141540825366974, + 0.026915868744254112, + 0.028393596410751343, + 0.02858819253742695, + 0.11403373628854752, + 0.001827796339057386, + 0.001881152973510325, + 0.0017159637063741684, + 0.00362674449570477, + 0.002045336877927184, + 0.0012069379445165396, + 0.002386766020208597, + 0.0014029904268682003, + 0.007048314902931452, + 0.008819453418254852, + 0.00676396768540144, + 0.02082211524248123, + 0.007091332692652941, + 0.011195641942322254, + 0.014025581069290638, + 0.011052466928958893, + 0.004070126451551914, + 0.003974466118961573, + 0.004392748698592186, + 0.007892338559031487, + 0.0050813499838113785, + 0.0059027778916060925, + 0.011217479594051838, + 0.004483814351260662, + 0.020849090069532394, + 0.008094667457044125, + 0.03258248791098595, + 0.01097041741013527, + 0.008355850353837013, + 0.028359821066260338, + 0.012240729294717312, + 0.018045416101813316, + 0.0028518023900687695, + 0.0033326707780361176, + 0.0036821612156927586, + 0.0033736855257302523, + 0.002306625945493579, + 0.004086790140718222, + 0.0029513800982385874, + 0.003810918191447854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.06204794719815254 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.693346381187439 + ], + "opposite_logprob_tokens": [ + -0.693346381187439 + ], + "logprob_reward": -0.693346381187439, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.693346381187439, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3270033001899719, + 0.03174476698040962, + 0.016601746901869774, + 0.04558975249528885, + 0.05966070666909218, + 0.03693981096148491, + 0.11622454971075058, + 0.054764244705438614, + 0.22645533084869385, + 0.08501585572957993 + ], + "credit_map": [ + 0.03431451693177223, + 0.03251460939645767, + 0.08417528122663498, + 0.05626463145017624, + 0.02884211204946041, + 0.025389794260263443, + 0.03905871510505676, + 0.026443613693118095, + 0.004010406322777271, + 0.0025431725662201643, + 0.008520863950252533, + 0.003994937986135483, + 0.002973628928884864, + 0.0035370872355997562, + 0.003169954987242818, + 0.0029947138391435146, + 0.0023713232949376106, + 0.0010443286737427115, + 0.0016433072742074728, + 0.0016055942978709936, + 0.001329553546383977, + 0.002537619089707732, + 0.0015580697217956185, + 0.004511953331530094, + 0.0036137064453214407, + 0.005191698670387268, + 0.007449537981301546, + 0.006237837485969067, + 0.005495878402143717, + 0.0050232671201229095, + 0.008664465509355068, + 0.0039133611135184765, + 0.009110851213335991, + 0.004764935001730919, + 0.005189426708966494, + 0.006876190192997456, + 0.004066694062203169, + 0.0047344909980893135, + 0.004184782970696688, + 0.020733335986733437, + 0.004721899516880512, + 0.0036197593435645103, + 0.00497001688927412, + 0.009344861842691898, + 0.003054192056879401, + 0.002744669560343027, + 0.00555635429918766, + 0.0029280593153089285, + 0.014707519672811031, + 0.008411763235926628, + 0.008353890851140022, + 0.04215816408395767, + 0.006761801429092884, + 0.00844881683588028, + 0.01890118606388569, + 0.008481407538056374, + 0.005766740068793297, + 0.0030021625570952892, + 0.0038523476105183363, + 0.009941144846379757, + 0.0037963769864290953, + 0.0035797555465251207, + 0.018354270607233047, + 0.006471447646617889, + 0.03258966654539108, + 0.013148516416549683, + 0.04732841998338699, + 0.018802769482135773, + 0.01480091642588377, + 0.04627200961112976, + 0.02281327173113823, + 0.030699754133820534, + 0.009279835037887096, + 0.013520759530365467, + 0.010441676713526249, + 0.012599789537489414, + 0.009252944029867649, + 0.01125927921384573, + 0.010543296113610268, + 0.008118275552988052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.693346381187439 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369060903787613 + ], + "opposite_logprob_tokens": [ + -3.4086906909942627 + ], + "logprob_reward": -0.03369060903787613, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369060903787613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1462094783782959, + 0.042287591844797134, + 0.03827506676316261, + 0.07366564869880676, + 0.11843258142471313, + 0.06513963639736176, + 0.22816921770572662, + 0.04241359606385231, + 0.22199466824531555, + 0.023412493988871574 + ], + "credit_map": [ + 0.014268696308135986, + 0.01331161055713892, + 0.03468739986419678, + 0.02416412904858589, + 0.014292187988758087, + 0.013752994127571583, + 0.017306143417954445, + 0.014426317997276783, + 0.0058842250145971775, + 0.004796828143298626, + 0.007280171848833561, + 0.0058703310787677765, + 0.004588153678923845, + 0.006189813371747732, + 0.004489697515964508, + 0.0031883742194622755, + 0.004763214848935604, + 0.0031198938377201557, + 0.004769240040332079, + 0.0056884028017520905, + 0.004190295469015837, + 0.004995343275368214, + 0.0037177246995270252, + 0.007030951324850321, + 0.007486477494239807, + 0.007903083227574825, + 0.01185606699436903, + 0.008369691669940948, + 0.011892594397068024, + 0.008588156662881374, + 0.009736488573253155, + 0.007833086885511875, + 0.01783187873661518, + 0.008531219325959682, + 0.011190359480679035, + 0.013666464015841484, + 0.0084835821762681, + 0.010207846760749817, + 0.008582951501011848, + 0.03993827477097511, + 0.006263337098062038, + 0.006843638140708208, + 0.008722946047782898, + 0.010497133247554302, + 0.007365839090198278, + 0.006005126517266035, + 0.010915125720202923, + 0.008526496589183807, + 0.02510041929781437, + 0.013163258321583271, + 0.013206752017140388, + 0.09170451760292053, + 0.006800222676247358, + 0.01148940622806549, + 0.03636176884174347, + 0.030342863872647285, + 0.0037418866995722055, + 0.004983819555491209, + 0.007004491984844208, + 0.005225497763603926, + 0.0035636494867503643, + 0.005114531610161066, + 0.00893096812069416, + 0.00384874758310616, + 0.03148777782917023, + 0.014300424605607986, + 0.04171203449368477, + 0.019754819571971893, + 0.01582331955432892, + 0.04379268363118172, + 0.02393612079322338, + 0.031187478452920914, + 0.002971712965518236, + 0.0032374358270317316, + 0.0027000841218978167, + 0.003460000967606902, + 0.0025441250763833523, + 0.003471514442935586, + 0.0024466163013130426, + 0.002581004984676838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.03369060903787613 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "In a photorealistic style, a very slow, eye-level tracking shot follows a majestic Pacific green sea turtle as it feeds along the sun-drenched seafloor. Bright sunlight penetrates the crystal-clear turquoise water, creating a mesmerizing, ever-changing network of bright, shimmering caustics that dance across the turtle's massive, heart-shaped carapace—a mosaic of olive green and brown scales with a realistic texture. This ancient reptile, with its wrinkled skin and intelligent, dark eyes conveying a sense of timelessness, moves with ponderous grace. Lowering its head, it delicately and delicately uses its powerful, keratinous beak to carefully nibble at a patch of lush, vibrant seagrass rooted in the pristine white sand below, stirring up a small cloud of sand with each bite. The turtle is completely oblivious to the schools of colorful small fish, primarily yellow and blue sockeye, that frolic around its massive body. The entire scene is filled with a profoundly serene atmosphere and tranquil natural beauty, and the camera’s steady, unobtrusive movements make the viewer feel like a silent observer in this tranquil underwater world.", + "seed": 8040, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_6/index_8646_seed_8040_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.235935228317976, + "vlm_logprob": -0.235935228317976, + "vlm_yes_no_margin": 2.125, + "vlm_token_credit": -0.235935228317976, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13060130178928375 + ], + "opposite_logprob_tokens": [ + -2.130601406097412 + ], + "logprob_reward": -0.13060130178928375, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13060130178928375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09570710361003876, + 0.1123732402920723, + 0.12380404770374298, + 0.05433395132422447, + 0.08695957064628601, + 0.11463867127895355, + 0.14924931526184082, + 0.07508911937475204, + 0.12870104610919952, + 0.05914393067359924 + ], + "credit_map": [ + 0.011411747895181179, + 0.009475004859268665, + 0.01013206597417593, + 0.014423617161810398, + 0.007677505724132061, + 0.01792888343334198, + 0.01400255411863327, + 0.010655725374817848, + 0.010011282749474049, + 0.019938046112656593, + 0.008699197322130203, + 0.012083608657121658, + 0.015908006578683853, + 0.010822058655321598, + 0.01671881414949894, + 0.018192226067185402, + 0.0193466953933239, + 0.01457083411514759, + 0.015554358251392841, + 0.010387476533651352, + 0.015180140733718872, + 0.01392412930727005, + 0.019903365522623062, + 0.014937053434550762, + 0.006784531287848949, + 0.008197912946343422, + 0.00694305682554841, + 0.007164613343775272, + 0.005863870494067669, + 0.005846873391419649, + 0.007703922223299742, + 0.00582916708663106, + 0.011319546960294247, + 0.0070174154825508595, + 0.01267680712044239, + 0.010740229859948158, + 0.011404231190681458, + 0.008652527816593647, + 0.015426917001605034, + 0.00972189474850893, + 0.013826102018356323, + 0.012858202680945396, + 0.010807769373059273, + 0.013683132827281952, + 0.018258187919855118, + 0.010983951389789581, + 0.023600636050105095, + 0.010620693676173687, + 0.018493464216589928, + 0.018685244023799896, + 0.013895652256906033, + 0.018114134669303894, + 0.02354731597006321, + 0.018776142969727516, + 0.024062277749180794, + 0.013675075955688953, + 0.010151630267500877, + 0.011220209300518036, + 0.010582175105810165, + 0.008028091862797737, + 0.010070308111608028, + 0.009143118746578693, + 0.008612693287432194, + 0.0072808959521353245, + 0.017100857570767403, + 0.015756335109472275, + 0.021606208756566048, + 0.010932281613349915, + 0.01228009071201086, + 0.016489583998918533, + 0.012356453575193882, + 0.022179236635565758, + 0.009016036055982113, + 0.006492919754236937, + 0.010621501132845879, + 0.004851885139942169, + 0.007622752338647842, + 0.008458207361400127, + 0.004704264923930168, + 0.007376366760581732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.13060130178928375 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620885044336319 + ], + "opposite_logprob_tokens": [ + -2.8120884895324707 + ], + "logprob_reward": -0.0620885044336319, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620885044336319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08243690431118011, + 0.12329854816198349, + 0.04347069934010506, + 0.03338106721639633, + 0.050968147814273834, + 0.3027012348175049, + 0.03848093003034592, + 0.22797362506389618, + 0.05593438819050789, + 0.041354380548000336 + ], + "credit_map": [ + 0.011258386075496674, + 0.010394701734185219, + 0.009591796435415745, + 0.008239970542490482, + 0.008431894704699516, + 0.011267373338341713, + 0.010610725730657578, + 0.01264206226915121, + 0.009159868583083153, + 0.018621034920215607, + 0.013572834432125092, + 0.019957849755883217, + 0.012375151738524437, + 0.010651960037648678, + 0.020913610234856606, + 0.01804625801742077, + 0.00825880654156208, + 0.005253395065665245, + 0.00503972452133894, + 0.004473783541470766, + 0.004503324162214994, + 0.0050344872288405895, + 0.005061467178165913, + 0.0058457134291529655, + 0.00516136409714818, + 0.005298927426338196, + 0.004155466333031654, + 0.004033281933516264, + 0.0034398725256323814, + 0.003182732267305255, + 0.0032904469408094883, + 0.004818981047719717, + 0.005940567702054977, + 0.0068960334174335, + 0.006720120087265968, + 0.0073966169729828835, + 0.00656501017510891, + 0.004521999973803759, + 0.00720967585220933, + 0.0057181306183338165, + 0.08408667147159576, + 0.01267796941101551, + 0.016210416331887245, + 0.010717814788222313, + 0.015182976610958576, + 0.12835946679115295, + 0.02069087326526642, + 0.01477509643882513, + 0.004758213646709919, + 0.004729883279651403, + 0.0061738695949316025, + 0.0044431318528950214, + 0.0052294484339654446, + 0.004522313829511404, + 0.005155438557267189, + 0.0034686364233493805, + 0.029342280700802803, + 0.027872279286384583, + 0.02935905009508133, + 0.021312035620212555, + 0.03847122937440872, + 0.02628510445356369, + 0.02427752874791622, + 0.031054148450493813, + 0.007323428988456726, + 0.006774991750717163, + 0.008296133019030094, + 0.008608378469944, + 0.0060011050663888454, + 0.011302048340439796, + 0.0038391437847167253, + 0.003789167618378997, + 0.005015351343899965, + 0.0064451852813363075, + 0.007024993654340506, + 0.004542431328445673, + 0.005005559418350458, + 0.006173422560095787, + 0.0038070762529969215, + 0.003340366994962096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.0620885044336319 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370536118745804 + ], + "opposite_logprob_tokens": [ + -3.408705472946167 + ], + "logprob_reward": -0.03370536118745804, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370536118745804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08448242396116257, + 0.15191787481307983, + 0.07094664871692657, + 0.038857147097587585, + 0.06795917451381683, + 0.21143348515033722, + 0.02628837339580059, + 0.17174914479255676, + 0.08897131681442261, + 0.08739441633224487 + ], + "credit_map": [ + 0.0092713488265872, + 0.01700565218925476, + 0.010817577131092548, + 0.009694241918623447, + 0.006415301002562046, + 0.012830615974962711, + 0.008853158913552761, + 0.009594525210559368, + 0.016212565824389458, + 0.02091621793806553, + 0.01776040345430374, + 0.01916332356631756, + 0.014941534027457237, + 0.014571716077625751, + 0.01852509006857872, + 0.029827017337083817, + 0.011368046514689922, + 0.010238888673484325, + 0.007199935615062714, + 0.007776386570185423, + 0.0064698513597249985, + 0.00795940775424242, + 0.011171731166541576, + 0.00876239687204361, + 0.00880418810993433, + 0.004025967326015234, + 0.0056489137932658195, + 0.003352728672325611, + 0.0034854570403695107, + 0.00449187308549881, + 0.004561350215226412, + 0.004486666526645422, + 0.008179391734302044, + 0.011936772614717484, + 0.008421707898378372, + 0.010450842790305614, + 0.007273411378264427, + 0.006456712726503611, + 0.007941536605358124, + 0.007298791781067848, + 0.05024074390530586, + 0.010781928896903992, + 0.01216402929276228, + 0.01007081102579832, + 0.013729416765272617, + 0.07967614382505417, + 0.02193601429462433, + 0.012834415771067142, + 0.003981512971222401, + 0.0036289035342633724, + 0.0030505156610161066, + 0.003562529571354389, + 0.003596352878957987, + 0.0027961176820099354, + 0.0030054969247430563, + 0.002666943008080125, + 0.02336684800684452, + 0.022104227915406227, + 0.025920895859599113, + 0.018908897414803505, + 0.021628471091389656, + 0.019077232107520103, + 0.019989226013422012, + 0.02075336128473282, + 0.011695899069309235, + 0.011552855372428894, + 0.020599542185664177, + 0.0078108045272529125, + 0.009078971110284328, + 0.011172190308570862, + 0.0070663695223629475, + 0.0099946865811944, + 0.010028396733105183, + 0.011549064889550209, + 0.01452203281223774, + 0.010227122344076633, + 0.013543625362217426, + 0.012532415799796581, + 0.008334190584719181, + 0.006657571066170931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.03370536118745804 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.898329496383667 + ], + "opposite_logprob_tokens": [ + -0.523329496383667 + ], + "logprob_reward": -0.898329496383667, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.898329496383667, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11737105250358582, + 0.1866239756345749, + 0.09339851140975952, + 0.036523010581731796, + 0.12419161200523376, + 0.19248630106449127, + 0.060109592974185944, + 0.03896067291498184, + 0.09747882187366486, + 0.052856363356113434 + ], + "credit_map": [ + 0.015692435204982758, + 0.024917198345065117, + 0.010205714032053947, + 0.009986802004277706, + 0.00958928745239973, + 0.01645265892148018, + 0.012745127081871033, + 0.017781846225261688, + 0.023579301312565804, + 0.022888431325554848, + 0.01734866015613079, + 0.020324233919382095, + 0.020411459729075432, + 0.018889885395765305, + 0.024512846022844315, + 0.038669172674417496, + 0.023928876966238022, + 0.007792964112013578, + 0.011513536795973778, + 0.0060295360162854195, + 0.006470279768109322, + 0.013216654770076275, + 0.01399959810078144, + 0.010447079315781593, + 0.005691132508218288, + 0.006132279988378286, + 0.004992262925952673, + 0.003956545609980822, + 0.0036091499496251345, + 0.004002973437309265, + 0.004434486385434866, + 0.003704182105138898, + 0.025483259931206703, + 0.014410587958991528, + 0.018382498994469643, + 0.014753006398677826, + 0.01419601310044527, + 0.012223334982991219, + 0.014295690692961216, + 0.010447220876812935, + 0.057739127427339554, + 0.008138690143823624, + 0.009025037288665771, + 0.007013869006186724, + 0.007672121748328209, + 0.08463674783706665, + 0.009232506155967712, + 0.009028215892612934, + 0.009578726254403591, + 0.007320969365537167, + 0.008395016193389893, + 0.007780918385833502, + 0.009523838758468628, + 0.007086075376719236, + 0.006026298273354769, + 0.004397760145366192, + 0.0052698454819619656, + 0.005364654585719109, + 0.005064326338469982, + 0.003869342850521207, + 0.005096300970762968, + 0.005672293249517679, + 0.004432699643075466, + 0.004191212821751833, + 0.012080763466656208, + 0.009883742779493332, + 0.022461075335741043, + 0.009872935712337494, + 0.011762400157749653, + 0.011617778800427914, + 0.008994671516120434, + 0.010805472731590271, + 0.00602085841819644, + 0.006972273346036673, + 0.00943660270422697, + 0.005503010470420122, + 0.0089596351608634, + 0.006410754751414061, + 0.005049652419984341, + 0.004503584001213312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.898329496383667 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495147779583931 + ], + "opposite_logprob_tokens": [ + -2.9299514293670654 + ], + "logprob_reward": -0.05495147779583931, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495147779583931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11539997160434723, + 0.07584290206432343, + 0.028459936380386353, + 0.03749379143118858, + 0.038978271186351776, + 0.10693828016519547, + 0.04596889764070511, + 0.3781273066997528, + 0.13239799439907074, + 0.04039265587925911 + ], + "credit_map": [ + 0.009060096926987171, + 0.017827065661549568, + 0.01551287155598402, + 0.015253548510372639, + 0.01090832706540823, + 0.01686652936041355, + 0.015168152749538422, + 0.014803367666900158, + 0.010804014280438423, + 0.00900981668382883, + 0.006531220860779285, + 0.005407650489360094, + 0.011022237129509449, + 0.00861993059515953, + 0.012416733428835869, + 0.012031301856040955, + 0.00376585079357028, + 0.003936131950467825, + 0.004045537207275629, + 0.00311482697725296, + 0.003372966079041362, + 0.0027748732827603817, + 0.005124531686306, + 0.002325219800695777, + 0.004341685213148594, + 0.005672496277838945, + 0.006227209698408842, + 0.003956005442887545, + 0.004186993930488825, + 0.003191039664670825, + 0.0038568279705941677, + 0.0060615320689976215, + 0.0042852978222072124, + 0.005695318803191185, + 0.0051381587982177734, + 0.0054257214069366455, + 0.004857770167291164, + 0.0031018175650388002, + 0.006039241794496775, + 0.004434945061802864, + 0.03167734295129776, + 0.0054399617947638035, + 0.005049852188676596, + 0.005329395644366741, + 0.004636363126337528, + 0.04328616335988045, + 0.006208678241819143, + 0.0053105223923921585, + 0.005845040548592806, + 0.005333310924470425, + 0.005996838212013245, + 0.0062507521361112595, + 0.005634250119328499, + 0.005768040660768747, + 0.006748434156179428, + 0.004392227623611689, + 0.0529434010386467, + 0.03972537815570831, + 0.039812296628952026, + 0.044118113815784454, + 0.06367813050746918, + 0.04397924244403839, + 0.04586398974061012, + 0.04800673946738243, + 0.017710991203784943, + 0.020526094362139702, + 0.01818479597568512, + 0.01686910167336464, + 0.015320371836423874, + 0.023933034390211105, + 0.00926052313297987, + 0.01059306412935257, + 0.004698579199612141, + 0.005317535251379013, + 0.006700804922729731, + 0.004273617174476385, + 0.005117711611092091, + 0.006079057697206736, + 0.0046246228739619255, + 0.0035807283129543066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.05495147779583931 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "In a photorealistic style, a very slow, eye-level tracking shot follows a majestic Pacific green sea turtle as it feeds along the sun-drenched seafloor. Bright sunlight penetrates the crystal-clear turquoise water, creating a mesmerizing, ever-changing network of bright, shimmering caustics that dance across the turtle's massive, heart-shaped carapace—a mosaic of olive green and brown scales with a realistic texture. This ancient reptile, with its wrinkled skin and intelligent, dark eyes conveying a sense of timelessness, moves with ponderous grace. Lowering its head, it delicately and delicately uses its powerful, keratinous beak to carefully nibble at a patch of lush, vibrant seagrass rooted in the pristine white sand below, stirring up a small cloud of sand with each bite. The turtle is completely oblivious to the schools of colorful small fish, primarily yellow and blue sockeye, that frolic around its massive body. The entire scene is filled with a profoundly serene atmosphere and tranquil natural beauty, and the camera’s steady, unobtrusive movements make the viewer feel like a silent observer in this tranquil underwater world.", + "seed": 8041, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_6/index_8646_seed_8041_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05982592552900314, + "vlm_logprob": -0.05982592552900314, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.05982592552900314, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08140607923269272 + ], + "opposite_logprob_tokens": [ + -2.5814061164855957 + ], + "logprob_reward": -0.08140607923269272, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08140607923269272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3228917419910431, + 0.03827919811010361, + 0.02430511824786663, + 0.08606559783220291, + 0.03115610033273697, + 0.03521634265780449, + 0.07234899699687958, + 0.10886896401643753, + 0.26094850897789, + 0.019919512793421745 + ], + "credit_map": [ + 0.043093107640743256, + 0.008439033292233944, + 0.012923932634294033, + 0.008686654269695282, + 0.0115022799000144, + 0.13075749576091766, + 0.013790834695100784, + 0.09369838237762451, + 0.004517789930105209, + 0.004195673391222954, + 0.004093895200639963, + 0.0035430064890533686, + 0.003444983623921871, + 0.008551851846277714, + 0.006166690029203892, + 0.003765306668356061, + 0.002263749949634075, + 0.004249938298016787, + 0.003612278262153268, + 0.0030431232880800962, + 0.002660401863977313, + 0.0022730911150574684, + 0.004355153534561396, + 0.0018473835662007332, + 0.00868101418018341, + 0.013818279840052128, + 0.011197849176824093, + 0.010453296825289726, + 0.012530062347650528, + 0.008637558668851852, + 0.01391017809510231, + 0.006837352644652128, + 0.0031526703387498856, + 0.0035831641871482134, + 0.0051353308372199535, + 0.0027381128165870905, + 0.0034855660051107407, + 0.004301242996007204, + 0.004442366771399975, + 0.004317646846175194, + 0.0036796443164348602, + 0.003954899962991476, + 0.004624160937964916, + 0.0035244524478912354, + 0.00430888868868351, + 0.006063502747565508, + 0.00249455776065588, + 0.006566235795617104, + 0.00695570046082139, + 0.00995639432221651, + 0.00874682329595089, + 0.00744347320869565, + 0.012343780137598515, + 0.00828365609049797, + 0.011692504398524761, + 0.006926668342202902, + 0.01107916422188282, + 0.009884788654744625, + 0.02013595774769783, + 0.021028947085142136, + 0.01319968607276678, + 0.00822492316365242, + 0.01242087408900261, + 0.012894617393612862, + 0.03237685188651085, + 0.02977246791124344, + 0.03205709159374237, + 0.039969462901353836, + 0.024344630539417267, + 0.020114945247769356, + 0.05335254594683647, + 0.028960512951016426, + 0.002517856890335679, + 0.0019320629071444273, + 0.0016899846959859133, + 0.0028385119512677193, + 0.0015328016597777605, + 0.0018048577476292849, + 0.0036687052343040705, + 0.003934732638299465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.08140607923269272 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337178111076355 + ], + "opposite_logprob_tokens": [ + -3.4087178707122803 + ], + "logprob_reward": -0.0337178111076355, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337178111076355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3807450532913208, + 0.029486386105418205, + 0.02676103636622429, + 0.023646807298064232, + 0.028319915756583214, + 0.04049493372440338, + 0.07105035334825516, + 0.10282254219055176, + 0.2744238078594208, + 0.022249191999435425 + ], + "credit_map": [ + 0.05048532783985138, + 0.010305129922926426, + 0.015237303450703621, + 0.010253837332129478, + 0.013316274620592594, + 0.15446893870830536, + 0.015946518629789352, + 0.11073170602321625, + 0.0031924450304359198, + 0.003677935106679797, + 0.003648966085165739, + 0.003006487153470516, + 0.0024592371191829443, + 0.006549830082803965, + 0.00477713905274868, + 0.0021743474062532187, + 0.0026615920942276716, + 0.004660699982196093, + 0.004612252116203308, + 0.003606940619647503, + 0.002390519483014941, + 0.002384687541052699, + 0.004381274338811636, + 0.002063068328425288, + 0.0028805367182940245, + 0.004015298094600439, + 0.0032159187830984592, + 0.0025923424400389194, + 0.0031695961952209473, + 0.001827080617658794, + 0.0035186687018722296, + 0.0024273658636957407, + 0.0028360821306705475, + 0.0032353694550693035, + 0.004327076021581888, + 0.002526242285966873, + 0.003655088134109974, + 0.0034151310101151466, + 0.004550666082650423, + 0.0037742562126368284, + 0.003461134387180209, + 0.0036896616220474243, + 0.004351865965873003, + 0.003740007523447275, + 0.004118597134947777, + 0.00874378252774477, + 0.002764604054391384, + 0.009625277481973171, + 0.0062865098007023335, + 0.009024213999509811, + 0.008617733605206013, + 0.006741624791175127, + 0.01057329960167408, + 0.008618602529168129, + 0.01037110947072506, + 0.010817257687449455, + 0.010297084227204323, + 0.009345893748104572, + 0.018637778237462044, + 0.019617972895503044, + 0.012352689169347286, + 0.008069694973528385, + 0.011683478020131588, + 0.012817933224141598, + 0.033845070749521255, + 0.031162317842245102, + 0.03322143480181694, + 0.04182196408510208, + 0.02548959106206894, + 0.02131911739706993, + 0.05624166131019592, + 0.031322628259658813, + 0.002062987070530653, + 0.0029710358940064907, + 0.002185193356126547, + 0.0030972689855843782, + 0.0024164665956050158, + 0.002030685544013977, + 0.00427223602309823, + 0.003213314339518547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.0337178111076355 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328341454267502 + ], + "opposite_logprob_tokens": [ + -3.7732834815979004 + ], + "logprob_reward": -0.02328341454267502, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328341454267502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5447039604187012, + 0.02434333600103855, + 0.018656497821211815, + 0.03193763643503189, + 0.02310970053076744, + 0.026631981134414673, + 0.04832003638148308, + 0.07721156626939774, + 0.1921553760766983, + 0.01292989682406187 + ], + "credit_map": [ + 0.07299818098545074, + 0.01486084796488285, + 0.021544121205806732, + 0.014512899331748486, + 0.01918729767203331, + 0.22263433039188385, + 0.02225039340555668, + 0.15671588480472565, + 0.002412159461528063, + 0.003177548060193658, + 0.0025589060969650745, + 0.002047418151050806, + 0.0022922889329493046, + 0.004785101395100355, + 0.005110149271786213, + 0.001959763700142503, + 0.001892577507533133, + 0.0031202337704598904, + 0.00294698728248477, + 0.0024409517645835876, + 0.0017128591425716877, + 0.0020001460798084736, + 0.00308221229352057, + 0.0014605318428948522, + 0.0038044333923608065, + 0.004289731383323669, + 0.0046339030377566814, + 0.0038268945645540953, + 0.004026744980365038, + 0.0034560596104711294, + 0.005221094004809856, + 0.0026787747628986835, + 0.002004462294280529, + 0.002588886534795165, + 0.0029027583077549934, + 0.0020264664199203253, + 0.0030663260258734226, + 0.0031777690164744854, + 0.00419071689248085, + 0.003152313642203808, + 0.002780367387458682, + 0.002950561698526144, + 0.003485125955194235, + 0.002605590969324112, + 0.0032390241976827383, + 0.004586219787597656, + 0.002095688134431839, + 0.004889404401183128, + 0.004821341019123793, + 0.0070118666626513, + 0.006643218919634819, + 0.004922826308757067, + 0.008213515393435955, + 0.0045213857665658, + 0.008190601132810116, + 0.003995280712842941, + 0.00760295195505023, + 0.006946174893528223, + 0.014240109361708164, + 0.01483279187232256, + 0.00930763129144907, + 0.005878371652215719, + 0.008685988374054432, + 0.009717549197375774, + 0.024173909798264503, + 0.022209329530596733, + 0.024783862754702568, + 0.029730424284934998, + 0.018318431451916695, + 0.014700967818498611, + 0.036907754838466644, + 0.021330703049898148, + 0.0020494863856583834, + 0.0014460436068475246, + 0.001286710030399263, + 0.0014755656011402607, + 0.0014409177238121629, + 0.001160543761216104, + 0.0019781128503382206, + 0.0020925160497426987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.02328341454267502 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703828513622284 + ], + "opposite_logprob_tokens": [ + -2.1270382404327393 + ], + "logprob_reward": -0.12703828513622284, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703828513622284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4312793016433716, + 0.05055328831076622, + 0.018658673390746117, + 0.024364609271287918, + 0.023469584062695503, + 0.05781639739871025, + 0.04515296220779419, + 0.09139714390039444, + 0.19054116308689117, + 0.06676679849624634 + ], + "credit_map": [ + 0.05782129615545273, + 0.011206216178834438, + 0.017179345712065697, + 0.011321112513542175, + 0.015540990978479385, + 0.17639893293380737, + 0.017590342089533806, + 0.12422110140323639, + 0.005793678108602762, + 0.006445627193897963, + 0.005353984888643026, + 0.00472949119284749, + 0.004329301882535219, + 0.008760105818510056, + 0.011037040501832962, + 0.004104061052203178, + 0.0022088487166911364, + 0.002685746643692255, + 0.0025150738656520844, + 0.0020431724842637777, + 0.0020557905081659555, + 0.002241122769191861, + 0.0033489875495433807, + 0.0015599321341142058, + 0.003261374309659004, + 0.0034031912218779325, + 0.00306557584553957, + 0.0025973368901759386, + 0.0030015278607606888, + 0.002484760247170925, + 0.003956395201385021, + 0.002594448160380125, + 0.0021136763971298933, + 0.0018256920157000422, + 0.0027129852678626776, + 0.001838675350882113, + 0.002418429357931018, + 0.005073331296443939, + 0.0032301724422723055, + 0.004256624262779951, + 0.003916311077773571, + 0.0056424615904688835, + 0.005195999518036842, + 0.004615840036422014, + 0.0064196353778243065, + 0.013133169151842594, + 0.003998078405857086, + 0.014894907362759113, + 0.004010423552244902, + 0.005423140246421099, + 0.00533160287886858, + 0.003549761837348342, + 0.007304511032998562, + 0.007157821208238602, + 0.005056534428149462, + 0.007319170981645584, + 0.01027699839323759, + 0.009182407520711422, + 0.016616782173514366, + 0.014286448247730732, + 0.011024761945009232, + 0.009083489887416363, + 0.009134984575212002, + 0.011791280470788479, + 0.02399907074868679, + 0.021455027163028717, + 0.022704696282744408, + 0.028923027217388153, + 0.017662525177001953, + 0.014663215726613998, + 0.03952433168888092, + 0.021609285846352577, + 0.0070730941370129585, + 0.010903654620051384, + 0.008273636922240257, + 0.011006669141352177, + 0.007418771740049124, + 0.006605696864426136, + 0.008961517363786697, + 0.0065237656235694885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.12703828513622284 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368403762578964 + ], + "opposite_logprob_tokens": [ + -3.408684015274048 + ], + "logprob_reward": -0.03368403762578964, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368403762578964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.552620530128479, + 0.03599065542221069, + 0.02344769984483719, + 0.03045707568526268, + 0.027074608951807022, + 0.0332612581551075, + 0.05320126190781593, + 0.04978154972195625, + 0.17646226286888123, + 0.017703060060739517 + ], + "credit_map": [ + 0.07409334182739258, + 0.013951646164059639, + 0.02234760858118534, + 0.014359550550580025, + 0.01894705928862095, + 0.2246336191892624, + 0.023348622024059296, + 0.16093911230564117, + 0.004339941311627626, + 0.005013217683881521, + 0.0029414035379886627, + 0.003500181483104825, + 0.0030187745578587055, + 0.009483614936470985, + 0.004393897019326687, + 0.003299621632322669, + 0.0025761479046195745, + 0.003947515040636063, + 0.0035858030896633863, + 0.002800548216328025, + 0.002623344771564007, + 0.0019813179969787598, + 0.004223513882607222, + 0.0017095080111175776, + 0.003845500759780407, + 0.004622637294232845, + 0.004468477796763182, + 0.004335121717303991, + 0.0032407231628894806, + 0.003215269884094596, + 0.0036924092564731836, + 0.003036933019757271, + 0.002813128288835287, + 0.004240831825882196, + 0.004632171709090471, + 0.0025540669448673725, + 0.003995106089860201, + 0.002307058311998844, + 0.004094647243618965, + 0.002437599003314972, + 0.0030972748063504696, + 0.003479083301499486, + 0.003443149384111166, + 0.00301548489369452, + 0.0033211582340300083, + 0.007195682264864445, + 0.0018968109507113695, + 0.007812612690031528, + 0.005565913859754801, + 0.007440304849296808, + 0.006365019828081131, + 0.005150494631379843, + 0.008931757882237434, + 0.006576793733984232, + 0.007971740327775478, + 0.005199234466999769, + 0.004722872283309698, + 0.004310993477702141, + 0.009633931331336498, + 0.008901185356080532, + 0.005415826104581356, + 0.004326866008341312, + 0.00550202839076519, + 0.0069678477011621, + 0.021893421187996864, + 0.02002764865756035, + 0.021036366000771523, + 0.026402199640870094, + 0.016334718093276024, + 0.013809416443109512, + 0.03722384199500084, + 0.019734637811779976, + 0.002320455387234688, + 0.0023021430242806673, + 0.0019733780063688755, + 0.002867144299671054, + 0.0015888852067291737, + 0.0012169613037258387, + 0.0025678863748908043, + 0.002866205992177129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.03368403762578964 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a photorealistic style, a very slow, eye-level tracking shot follows a majestic Pacific green sea turtle as it feeds along the sun-drenched seafloor. Bright sunlight penetrates the crystal-clear turquoise water, creating a mesmerizing, ever-changing network of bright, shimmering caustics that dance across the turtle's massive, heart-shaped carapace—a mosaic of olive green and brown scales with a realistic texture. This ancient reptile, with its wrinkled skin and intelligent, dark eyes conveying a sense of timelessness, moves with ponderous grace. Lowering its head, it delicately and delicately uses its powerful, keratinous beak to carefully nibble at a patch of lush, vibrant seagrass rooted in the pristine white sand below, stirring up a small cloud of sand with each bite. The turtle is completely oblivious to the schools of colorful small fish, primarily yellow and blue sockeye, that frolic around its massive body. The entire scene is filled with a profoundly serene atmosphere and tranquil natural beauty, and the camera’s steady, unobtrusive movements make the viewer feel like a silent observer in this tranquil underwater world.", + "seed": 8042, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_6/index_8646_seed_8042_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05887755416333675, + "vlm_logprob": -0.05887755416333675, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.05887755416333675, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11640742421150208 + ], + "opposite_logprob_tokens": [ + -2.2414073944091797 + ], + "logprob_reward": -0.11640742421150208, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11640742421150208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1169271394610405, + 0.12922737002372742, + 0.061425305902957916, + 0.09852989763021469, + 0.05551249533891678, + 0.03909648582339287, + 0.20581836998462677, + 0.08083706349134445, + 0.05213974416255951, + 0.1604861617088318 + ], + "credit_map": [ + 0.018442276865243912, + 0.019265184178948402, + 0.013076551258563995, + 0.01313696801662445, + 0.012730482965707779, + 0.018326707184314728, + 0.011700167320668697, + 0.010248798877000809, + 0.012651191093027592, + 0.026772454380989075, + 0.027381999418139458, + 0.015171858482062817, + 0.011922457255423069, + 0.013447236269712448, + 0.010032259859144688, + 0.01184791885316372, + 0.007058281917124987, + 0.009577175602316856, + 0.007313030771911144, + 0.004825504496693611, + 0.00701952027156949, + 0.015283840708434582, + 0.00464589474722743, + 0.005702061578631401, + 0.011326083913445473, + 0.012067154981195927, + 0.013717259280383587, + 0.014442299492657185, + 0.01125510223209858, + 0.012271258048713207, + 0.013773432932794094, + 0.009677303954958916, + 0.005226080305874348, + 0.006894856225699186, + 0.004279311280697584, + 0.011298635974526405, + 0.005418697372078896, + 0.007505123503506184, + 0.010197943076491356, + 0.004691848531365395, + 0.00618067616596818, + 0.0049535115249454975, + 0.00364466174505651, + 0.006050830706954002, + 0.006503149401396513, + 0.004747654777020216, + 0.004275216255337, + 0.002740784315392375, + 0.04076562076807022, + 0.013926196843385696, + 0.017087068408727646, + 0.019484061747789383, + 0.023250693455338478, + 0.05470157414674759, + 0.025339029729366302, + 0.011264110915362835, + 0.011104815639555454, + 0.006699087098240852, + 0.010184560902416706, + 0.007803025655448437, + 0.009447088465094566, + 0.016897989436984062, + 0.009103401564061642, + 0.009597093798220158, + 0.007484029978513718, + 0.007762777153402567, + 0.006152932066470385, + 0.0037344577722251415, + 0.00723247230052948, + 0.00738914730027318, + 0.0063343411311507225, + 0.0060495841316878796, + 0.0399366170167923, + 0.013499900698661804, + 0.058864619582891464, + 0.010073610581457615, + 0.009787908755242825, + 0.013612993061542511, + 0.0074978182092309, + 0.007212694268673658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 1 + ], + "selected_reward": -0.11640742421150208 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308691993355751 + ], + "opposite_logprob_tokens": [ + -3.1680870056152344 + ], + "logprob_reward": -0.04308691993355751, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308691993355751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14202050864696503, + 0.18452633917331696, + 0.03300020098686218, + 0.15917783975601196, + 0.04740450158715248, + 0.029249833896756172, + 0.08864329755306244, + 0.035428013652563095, + 0.04023730009794235, + 0.24031217396259308 + ], + "credit_map": [ + 0.020228911191225052, + 0.025033649057149887, + 0.01975192315876484, + 0.016321735456585884, + 0.01933629810810089, + 0.015424979850649834, + 0.014084378257393837, + 0.011838635429739952, + 0.017322175204753876, + 0.04133612662553787, + 0.036769382655620575, + 0.01623084768652916, + 0.016646260395646095, + 0.029034731909632683, + 0.012763645499944687, + 0.014423156157135963, + 0.004081088118255138, + 0.0037827733904123306, + 0.005178077612072229, + 0.003217175602912903, + 0.004585315007716417, + 0.005376080516725779, + 0.0030892568174749613, + 0.003690434852614999, + 0.020018456503748894, + 0.023125166073441505, + 0.02071388065814972, + 0.021598193794488907, + 0.017792485654354095, + 0.013873035088181496, + 0.02571038529276848, + 0.016346244141459465, + 0.006413501221686602, + 0.004909590817987919, + 0.005490727256983519, + 0.00808799173682928, + 0.00381131237372756, + 0.008805334568023682, + 0.006651830859482288, + 0.0032342157792299986, + 0.005149513948708773, + 0.0038612589705735445, + 0.003216525772586465, + 0.0030754811596125364, + 0.00482667051255703, + 0.0031796270050108433, + 0.003419959917664528, + 0.0025207975413650274, + 0.01654955744743347, + 0.006566144991666079, + 0.009606653824448586, + 0.008079905062913895, + 0.010954262688755989, + 0.01872999593615532, + 0.010863447561860085, + 0.007293323054909706, + 0.004459490068256855, + 0.0048419213853776455, + 0.0055174268782138824, + 0.003547860076650977, + 0.004032042343169451, + 0.004437227267771959, + 0.0038840111810714006, + 0.004708035383373499, + 0.00496774073690176, + 0.0040401131846010685, + 0.005840434227138758, + 0.003082847688347101, + 0.0050165411084890366, + 0.008981049060821533, + 0.0037940815091133118, + 0.004514489322900772, + 0.06138487160205841, + 0.02614610642194748, + 0.05785764008760452, + 0.012913132086396217, + 0.010846425779163837, + 0.04036039859056473, + 0.018648821860551834, + 0.012154773809015751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.04308691993355751 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799247160553932 + ], + "opposite_logprob_tokens": [ + -3.529799222946167 + ], + "logprob_reward": -0.029799247160553932, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799247160553932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24096131324768066, + 0.11079032719135284, + 0.04543662816286087, + 0.15867193043231964, + 0.05202276259660721, + 0.038450662046670914, + 0.08586425334215164, + 0.08854641020298004, + 0.06252335757017136, + 0.11673229187726974 + ], + "credit_map": [ + 0.04112260416150093, + 0.04078854247927666, + 0.037881553173065186, + 0.02034536562860012, + 0.032654084265232086, + 0.023371100425720215, + 0.022857971489429474, + 0.021940119564533234, + 0.01537345815449953, + 0.017550978809595108, + 0.020472606644034386, + 0.01016330998390913, + 0.011296778917312622, + 0.018311623483896255, + 0.008419763296842575, + 0.009201818145811558, + 0.004771135281771421, + 0.005784356966614723, + 0.0054397620260715485, + 0.0045358589850366116, + 0.005864895414561033, + 0.009846043772995472, + 0.004436359740793705, + 0.004758220165967941, + 0.019927294924855232, + 0.019345417618751526, + 0.017670031636953354, + 0.022655945271253586, + 0.017253806814551353, + 0.0176682211458683, + 0.031994204968214035, + 0.012157008051872253, + 0.0070426128804683685, + 0.006577495019882917, + 0.003355358261615038, + 0.013784073293209076, + 0.0027389281895011663, + 0.006748562213033438, + 0.009329961612820625, + 0.0024457729887217283, + 0.004934308119118214, + 0.0034598330967128277, + 0.004825145937502384, + 0.0057856254279613495, + 0.00544598838314414, + 0.0036680058110505342, + 0.007533796597272158, + 0.002797960303723812, + 0.018457813188433647, + 0.004771062638610601, + 0.005291098728775978, + 0.012261535972356796, + 0.004906544461846352, + 0.027586156502366066, + 0.008702012710273266, + 0.003888034028932452, + 0.020470386371016502, + 0.005809399299323559, + 0.006620368920266628, + 0.004558905027806759, + 0.00528884818777442, + 0.02951309084892273, + 0.010502809658646584, + 0.005782600957900286, + 0.008800944313406944, + 0.006884456146508455, + 0.010813717730343342, + 0.004118170123547316, + 0.005651125218719244, + 0.013103226199746132, + 0.007282907143235207, + 0.005868818610906601, + 0.02530076541006565, + 0.006583019159734249, + 0.01892213150858879, + 0.004200262948870659, + 0.006150463130325079, + 0.045329079031944275, + 0.0052965981885790825, + 0.004949979484081268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.029799247160553932 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205007806420326 + ], + "opposite_logprob_tokens": [ + -2.8120501041412354 + ], + "logprob_reward": -0.06205007806420326, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205007806420326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06184830516576767, + 0.17342598736286163, + 0.12084753066301346, + 0.32848721742630005, + 0.053805284202098846, + 0.03232850506901741, + 0.06710150092840195, + 0.05523254722356796, + 0.015362962149083614, + 0.09156022220849991 + ], + "credit_map": [ + 0.011457244865596294, + 0.006238230038434267, + 0.007455575745552778, + 0.005412897560745478, + 0.006813952699303627, + 0.013074081391096115, + 0.004549546632915735, + 0.006846773903816938, + 0.016183355823159218, + 0.039143502712249756, + 0.034975647926330566, + 0.016199253499507904, + 0.01678282581269741, + 0.02467523142695427, + 0.011493409052491188, + 0.013972748070955276, + 0.01503706630319357, + 0.01948404312133789, + 0.009539957158267498, + 0.009204192087054253, + 0.012548776343464851, + 0.03291773423552513, + 0.0111665865406394, + 0.010949170216917992, + 0.04910522326827049, + 0.028776055201888084, + 0.03563184663653374, + 0.030745649710297585, + 0.027023136615753174, + 0.06964797526597977, + 0.045212600380182266, + 0.04234470799565315, + 0.0074304514564573765, + 0.006154194474220276, + 0.0029450715519487858, + 0.015484633855521679, + 0.0030979537405073643, + 0.007373534608632326, + 0.009233959019184113, + 0.00208548572845757, + 0.005231778137385845, + 0.002674936782568693, + 0.00436670146882534, + 0.004335527773946524, + 0.003357354551553726, + 0.0034729554317891598, + 0.006278101354837418, + 0.0026111493352800608, + 0.014609341509640217, + 0.005575152579694986, + 0.0059662386775016785, + 0.009258097968995571, + 0.006303390953689814, + 0.012606429867446423, + 0.007702438626438379, + 0.00508041400462389, + 0.013292405754327774, + 0.004465560428798199, + 0.004313269164413214, + 0.0026354051660746336, + 0.0037937697488814592, + 0.018412083387374878, + 0.005236165132373571, + 0.0030838886741548777, + 0.0016462071798741817, + 0.0016809949884191155, + 0.0022839659359306097, + 0.001152660115621984, + 0.001960384426638484, + 0.002356817713007331, + 0.002121842233464122, + 0.0021600897889584303, + 0.028462756425142288, + 0.006314726546406746, + 0.014598111622035503, + 0.0031832135282456875, + 0.003597535891458392, + 0.028863323852419853, + 0.0035695109982043505, + 0.002971039852127433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.06205007806420326 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Pacific green sea turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle's carapace olive green and brown with a heart shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the turtle actively nibbling on lush seagrass with its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are schools of yellow and blue sockeye fish visible around the turtle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot eye-level and tracking the turtle steadily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304410144686699 + ], + "opposite_logprob_tokens": [ + -3.168044090270996 + ], + "logprob_reward": -0.04304410144686699, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304410144686699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05643393099308014, + 0.13439424335956573, + 0.04852854460477829, + 0.12062802165746689, + 0.04679187014698982, + 0.05683518946170807, + 0.14351125061511993, + 0.14294196665287018, + 0.1965217888355255, + 0.05341311916708946 + ], + "credit_map": [ + 0.007935619913041592, + 0.00947397667914629, + 0.006560612935572863, + 0.0045634182170033455, + 0.007052511442452669, + 0.0098709836602211, + 0.004989702254533768, + 0.005987113807350397, + 0.011873333714902401, + 0.02952714078128338, + 0.027333127334713936, + 0.012077734805643559, + 0.010685482993721962, + 0.024024682119488716, + 0.009037675336003304, + 0.009835082106292248, + 0.005765081848949194, + 0.00666297459974885, + 0.006357891019433737, + 0.005445555783808231, + 0.006379921454936266, + 0.008090566843748093, + 0.004937228746712208, + 0.004889327101409435, + 0.018100347369909286, + 0.012625732459127903, + 0.011619248427450657, + 0.015114648267626762, + 0.01334857102483511, + 0.010920113883912563, + 0.029137849807739258, + 0.009761523455381393, + 0.003261307952925563, + 0.0055571505799889565, + 0.003024062840268016, + 0.016257816925644875, + 0.0020615451503545046, + 0.005980244372040033, + 0.008574411273002625, + 0.002075335243716836, + 0.006833103019744158, + 0.006454389076679945, + 0.006351264659315348, + 0.0079454705119133, + 0.007348247803747654, + 0.006982750259339809, + 0.010233910754323006, + 0.004686060361564159, + 0.030509045347571373, + 0.006364075001329184, + 0.007586199790239334, + 0.024192439392209053, + 0.006747312843799591, + 0.04613977670669556, + 0.01611069031059742, + 0.005861721467226744, + 0.03755519911646843, + 0.013602650724351406, + 0.010982686653733253, + 0.00601954897865653, + 0.010043720714747906, + 0.04734079912304878, + 0.009208114817738533, + 0.008189255371689796, + 0.03695765137672424, + 0.015983056277036667, + 0.026784949004650116, + 0.012733454816043377, + 0.017976518720388412, + 0.05792122706770897, + 0.0167183056473732, + 0.011446640826761723, + 0.005587546154856682, + 0.002131723565980792, + 0.018117474392056465, + 0.0038888384588062763, + 0.003041786141693592, + 0.014048067852854729, + 0.0044592609629035, + 0.0021384290885180235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.04304410144686699 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_7/index_tensor([2244, 2244])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_7/index_tensor([2244, 2244])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5bbda0b34075f513f4e01fab487775b6e755bbaf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_7/index_tensor([2244, 2244])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, worm-eye tracking shot, clinging to the dusty stone floor, follows a mysterious figure draped in a tattered black cloak. Due to the low angle, only the figure's feet, clad in scuffed dark leather boots, can be seen moving forward with a nearly silent, ghostly gait, the hem of the cloak trailing along the ground, occasionally stirring up dust. They are infiltrating the interior of a magnificent, ancient reclining Buddha temple, flanked by towering stone pillars carved with blurred reliefs of ancient deities. The only illumination comes from sporadic torches burning on the walls. The unsteady flickering flames cast long, grotesque, distorted shadows across the massive pillars and the surrounding, sinister guardian deities. These shadows seem to come alive, dancing and expanding with every flicker of the flames. The main colors of the entire picture are deep dark colors and warm but ominous orange-red firelight. The strong contrast between light and dark and the dynamic shadows together create a dark fantasy aesthetic that is a mixture of sacredness, repression and danger, full of disturbing suspense.", + "seed": 9315, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_7/index_2244_seed_9315_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05342923969728872, + "vlm_logprob": -0.05342923969728872, + "vlm_yes_no_margin": 4.249999976158142, + "vlm_token_credit": -0.05342923969728872, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1468750238418579 + ], + "opposite_logprob_tokens": [ + -2.1468749046325684 + ], + "logprob_reward": -0.1468750238418579, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1468750238418579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3258894383907318, + 0.05006489157676697, + 0.20331838726997375, + 0.11203154176473618, + 0.06919052451848984, + 0.1049218475818634, + 0.03622996434569359, + 0.037809357047080994, + 0.019004663452506065, + 0.04153940826654434 + ], + "credit_map": [ + 0.025374803692102432, + 0.049215350300073624, + 0.03107394278049469, + 0.02400503121316433, + 0.07598254084587097, + 0.039999399334192276, + 0.02565298043191433, + 0.05458540841937065, + 0.005389625672250986, + 0.009117293171584606, + 0.003965574782341719, + 0.006082897074520588, + 0.007315437775105238, + 0.00817745178937912, + 0.004557015839964151, + 0.005459594540297985, + 0.009745767340064049, + 0.08197402209043503, + 0.007278076838701963, + 0.004946916829794645, + 0.07736989110708237, + 0.013581042177975178, + 0.004363476764410734, + 0.004059188533574343, + 0.018783777952194214, + 0.02057209052145481, + 0.0150272436439991, + 0.008614697493612766, + 0.02128707431256771, + 0.014424245804548264, + 0.008555200882256031, + 0.004767206031829119, + 0.007631712593138218, + 0.012657389976084232, + 0.004618879407644272, + 0.008680940605700016, + 0.009298806078732014, + 0.008887460455298424, + 0.008721191436052322, + 0.00869414396584034, + 0.0218113511800766, + 0.016792211681604385, + 0.00828570406883955, + 0.00746151152998209, + 0.009894868358969688, + 0.012653707526624203, + 0.01263745129108429, + 0.015385041013360023, + 0.004182832315564156, + 0.005791994743049145, + 0.0031596622429788113, + 0.003684407565742731, + 0.006884012371301651, + 0.004668875131756067, + 0.003657920751720667, + 0.004200262483209372, + 0.004968036897480488, + 0.005279047880321741, + 0.004165509715676308, + 0.0027652608696371317, + 0.008249212987720966, + 0.004360772203654051, + 0.005024861544370651, + 0.0029966558795422316, + 0.002244349801912904, + 0.0021494945976883173, + 0.0023811564315110445, + 0.0020362338982522488, + 0.0027895085513591766, + 0.003386348718777299, + 0.001599404844455421, + 0.0024181657936424017, + 0.002971622860059142, + 0.004112966824322939, + 0.0015256127808243036, + 0.003013631096109748, + 0.0108030391857028, + 0.00802936777472496, + 0.0058212014846503735, + 0.00526196276769042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.1468750238418579 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907348662614822 + ], + "opposite_logprob_tokens": [ + -2.464073419570923 + ], + "logprob_reward": -0.08907348662614822, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907348662614822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23981553316116333, + 0.18380483984947205, + 0.05315040424466133, + 0.17803291976451874, + 0.0909113883972168, + 0.06904635578393936, + 0.07298261672258377, + 0.04112333059310913, + 0.023891838267445564, + 0.04724081978201866 + ], + "credit_map": [ + 0.020402420312166214, + 0.034130483865737915, + 0.018628772348165512, + 0.030277177691459656, + 0.05148167535662651, + 0.030303843319416046, + 0.03524535521864891, + 0.01934579201042652, + 0.027804819867014885, + 0.02024584263563156, + 0.0182109996676445, + 0.02698129042983055, + 0.025890734046697617, + 0.025456352159380913, + 0.016524959355592728, + 0.022689851000905037, + 0.009414899162948132, + 0.00862839724868536, + 0.006384813226759434, + 0.0050142384134233, + 0.007562556769698858, + 0.00963467825204134, + 0.002865599002689123, + 0.0036452235653996468, + 0.024392863735556602, + 0.03869768604636192, + 0.02315807342529297, + 0.014725309796631336, + 0.03732653334736824, + 0.02049863152205944, + 0.011771530844271183, + 0.007462272886186838, + 0.011488468386232853, + 0.012728486210107803, + 0.009387458674609661, + 0.011085537262260914, + 0.012504219077527523, + 0.010075828991830349, + 0.01327543519437313, + 0.01036594994366169, + 0.014230864122509956, + 0.012410346418619156, + 0.006423464044928551, + 0.004925284534692764, + 0.007232436910271645, + 0.00856014620512724, + 0.007265511900186539, + 0.007998299784958363, + 0.006602713838219643, + 0.010463688522577286, + 0.006183173973113298, + 0.006897728890180588, + 0.013470584526658058, + 0.010161599144339561, + 0.00794807355850935, + 0.011255059391260147, + 0.006740113720297813, + 0.0059264893643558025, + 0.0046646795235574245, + 0.004217787180095911, + 0.006147280335426331, + 0.00482653733342886, + 0.003975792322307825, + 0.004624649416655302, + 0.003762866836041212, + 0.003147693583741784, + 0.003685777308419347, + 0.0024676129687577486, + 0.002553521655499935, + 0.0030627339147031307, + 0.00278830504976213, + 0.0024233267176896334, + 0.004356127232313156, + 0.004676421172916889, + 0.002869499148800969, + 0.003791598603129387, + 0.011193418875336647, + 0.008338060230016708, + 0.006809831131249666, + 0.005205864552408457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.08907348662614822 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023293431848287582 + ], + "opposite_logprob_tokens": [ + -3.7732934951782227 + ], + "logprob_reward": -0.023293431848287582, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023293431848287582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08579390496015549, + 0.0391700454056263, + 0.09567955881357193, + 0.2575647830963135, + 0.12209353595972061, + 0.07377202063798904, + 0.09290052950382233, + 0.10079846531152725, + 0.06503776460886002, + 0.06718937307596207 + ], + "credit_map": [ + 0.005603358615189791, + 0.016647133976221085, + 0.007535128854215145, + 0.010893394239246845, + 0.015454334206879139, + 0.010407574474811554, + 0.011022438295185566, + 0.008230538107454777, + 0.0026667097117751837, + 0.006119240075349808, + 0.003464573062956333, + 0.005037272348999977, + 0.006342944223433733, + 0.008110405877232552, + 0.003406492294743657, + 0.004022403620183468, + 0.0032487623393535614, + 0.040285371243953705, + 0.0024188440293073654, + 0.0016169475857168436, + 0.038840197026729584, + 0.005702672526240349, + 0.0015563562046736479, + 0.002010397845879197, + 0.029345067217946053, + 0.050134263932704926, + 0.03292377293109894, + 0.023883866146206856, + 0.039744988083839417, + 0.049182191491127014, + 0.016187477856874466, + 0.01616315357387066, + 0.013651437126100063, + 0.026080375537276268, + 0.009622982703149319, + 0.01932164467871189, + 0.01797369308769703, + 0.01250515878200531, + 0.012605790980160236, + 0.010332440957427025, + 0.01182965375483036, + 0.012840142473578453, + 0.007027230225503445, + 0.006313508376479149, + 0.009422345086932182, + 0.00963878259062767, + 0.007594907656311989, + 0.009105443954467773, + 0.008029994554817677, + 0.013654526323080063, + 0.007528373505920172, + 0.011995739303529263, + 0.0120494170114398, + 0.012205900624394417, + 0.016428804025053978, + 0.011007757857441902, + 0.00998204480856657, + 0.011587433516979218, + 0.01185020711272955, + 0.008081271313130856, + 0.01900051161646843, + 0.010843813419342041, + 0.014860649593174458, + 0.01459252554923296, + 0.011843456886708736, + 0.008438215591013432, + 0.009998037479817867, + 0.0064583937637507915, + 0.007795624900609255, + 0.008347416296601295, + 0.0054388451389968395, + 0.0067177703604102135, + 0.005298452451825142, + 0.006879537831991911, + 0.004659177735447884, + 0.006966304965317249, + 0.012858849950134754, + 0.012878303416073322, + 0.009264270775020123, + 0.008384468033909798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.023293431848287582 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632853463292122 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.007632853463292122, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632853463292122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07329351454973221, + 0.03610541298985481, + 0.341781347990036, + 0.08312636613845825, + 0.07011407613754272, + 0.12096091359853745, + 0.09330079704523087, + 0.03440266102552414, + 0.05164011940360069, + 0.09527474641799927 + ], + "credit_map": [ + 0.004734789952635765, + 0.013808329589664936, + 0.007434725295752287, + 0.005998363718390465, + 0.012514984235167503, + 0.017271926626563072, + 0.005434362683445215, + 0.006096043158322573, + 0.005242317449301481, + 0.005243001505732536, + 0.003904830664396286, + 0.004089938011020422, + 0.005011738743633032, + 0.005421766545623541, + 0.003325394820421934, + 0.003866427345201373, + 0.0090551245957613, + 0.14757053554058075, + 0.00630703940987587, + 0.0056610009633004665, + 0.14694026112556458, + 0.01785369962453842, + 0.004176624119281769, + 0.004217086359858513, + 0.007975965738296509, + 0.015838079154491425, + 0.008075982332229614, + 0.008831441402435303, + 0.013243483379483223, + 0.01469645369797945, + 0.007830261252820492, + 0.006634707096964121, + 0.00958939827978611, + 0.013033163733780384, + 0.006099388934671879, + 0.009848281741142273, + 0.010223183780908585, + 0.00521194888278842, + 0.009218995459377766, + 0.006889717187732458, + 0.03186669200658798, + 0.02605154737830162, + 0.010696380399167538, + 0.010053593665361404, + 0.011579046957194805, + 0.014817940071225166, + 0.008489883504807949, + 0.00740584684535861, + 0.00891084037721157, + 0.012988924048841, + 0.006913620978593826, + 0.01225447840988636, + 0.0185711607336998, + 0.013322892598807812, + 0.008775611408054829, + 0.011563275009393692, + 0.0033698088955134153, + 0.0042165652848780155, + 0.004401199985295534, + 0.0035374476574361324, + 0.005204279441386461, + 0.0035960625391453505, + 0.006156728602945805, + 0.003920573275536299, + 0.007964675314724445, + 0.009360074065625668, + 0.005959098692983389, + 0.004843840841203928, + 0.006518094334751368, + 0.0057732411660254, + 0.005199461709707975, + 0.006021641194820404, + 0.014619997702538967, + 0.011749695986509323, + 0.007769501768052578, + 0.010302746668457985, + 0.014957182109355927, + 0.016359305009245872, + 0.00976182147860527, + 0.009754505008459091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 9 + ], + "selected_reward": -0.007632853463292122 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0002714027068577707 + ], + "opposite_logprob_tokens": [ + -8.250271797180176 + ], + "logprob_reward": -0.0002714027068577707, + "yes_no_margin_reward": 8.25, + "token_credit_reward": -0.0002714027068577707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06353951245546341, + 0.0358184389770031, + 0.298991322517395, + 0.17489251494407654, + 0.0855315625667572, + 0.07128006219863892, + 0.10897501558065414, + 0.026847388595342636, + 0.06261037290096283, + 0.0715138241648674 + ], + "credit_map": [ + 0.0068052224814891815, + 0.01264555100351572, + 0.005577178206294775, + 0.004481973592191935, + 0.010375848971307278, + 0.011849637143313885, + 0.005250314250588417, + 0.0065537854097783566, + 0.003029065439477563, + 0.005484733730554581, + 0.003662676550447941, + 0.004948719404637814, + 0.005148324649780989, + 0.005291692912578583, + 0.003436997765675187, + 0.00481622526422143, + 0.007630073465406895, + 0.12582919001579285, + 0.006756827235221863, + 0.005063608288764954, + 0.12702685594558716, + 0.017512673512101173, + 0.004342525731772184, + 0.004829541780054569, + 0.02205291949212551, + 0.03582564741373062, + 0.020714541897177696, + 0.01502184011042118, + 0.03085612878203392, + 0.031005702912807465, + 0.009322973899543285, + 0.010092755779623985, + 0.010163472034037113, + 0.017815643921494484, + 0.0068579502403736115, + 0.014488755725324154, + 0.012609152123332024, + 0.007753388490527868, + 0.008392098359763622, + 0.007451100740581751, + 0.01455716136842966, + 0.01171360444277525, + 0.006254288833588362, + 0.005444150418043137, + 0.006936864927411079, + 0.009163928218185902, + 0.00786948949098587, + 0.009340577758848667, + 0.010538539849221706, + 0.014721040613949299, + 0.009284761734306812, + 0.014266932383179665, + 0.017408637329936028, + 0.014110337942838669, + 0.012958395294845104, + 0.015686366707086563, + 0.0039233313873410225, + 0.004248382523655891, + 0.003279648721218109, + 0.0022423553746193647, + 0.00433996831998229, + 0.002846636576578021, + 0.0030192609410732985, + 0.002947803121060133, + 0.009401945397257805, + 0.006864333059638739, + 0.009959769435226917, + 0.006906377151608467, + 0.006281423848122358, + 0.009766721166670322, + 0.006500658579170704, + 0.006929145194590092, + 0.009172412566840649, + 0.007473453879356384, + 0.004574722610414028, + 0.006604403723031282, + 0.014711787924170494, + 0.01280808262526989, + 0.009540414437651634, + 0.006628554780036211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 6 + ], + "selected_reward": -0.0002714027068577707 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, worm-eye tracking shot, clinging to the dusty stone floor, follows a mysterious figure draped in a tattered black cloak. Due to the low angle, only the figure's feet, clad in scuffed dark leather boots, can be seen moving forward with a nearly silent, ghostly gait, the hem of the cloak trailing along the ground, occasionally stirring up dust. They are infiltrating the interior of a magnificent, ancient reclining Buddha temple, flanked by towering stone pillars carved with blurred reliefs of ancient deities. The only illumination comes from sporadic torches burning on the walls. The unsteady flickering flames cast long, grotesque, distorted shadows across the massive pillars and the surrounding, sinister guardian deities. These shadows seem to come alive, dancing and expanding with every flicker of the flames. The main colors of the entire picture are deep dark colors and warm but ominous orange-red firelight. The strong contrast between light and dark and the dynamic shadows together create a dark fantasy aesthetic that is a mixture of sacredness, repression and danger, full of disturbing suspense.", + "seed": 9316, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_7/index_2244_seed_9316_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07459720831247978, + "vlm_logprob": -0.07459720831247978, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.07459720831247978, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3016095757484436 + ], + "opposite_logprob_tokens": [ + -1.4266095161437988 + ], + "logprob_reward": -0.3016095757484436, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.3016095757484436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6194833517074585, + 0.07491832971572876, + 0.03771099075675011, + 0.054468657821416855, + 0.061473067849874496, + 0.03686239942908287, + 0.03348382189869881, + 0.03195791319012642, + 0.02499561943113804, + 0.024645822122693062 + ], + "credit_map": [ + 0.055131945759058, + 0.03921663016080856, + 0.013783692382276058, + 0.01502972561866045, + 0.2544270157814026, + 0.07854197174310684, + 0.028416775166988373, + 0.13493557274341583, + 0.009510471485555172, + 0.012056584469974041, + 0.006882665678858757, + 0.010575396940112114, + 0.008908646181225777, + 0.009364765137434006, + 0.009984307922422886, + 0.007635494694113731, + 0.00522657623514533, + 0.005883121397346258, + 0.004795548040419817, + 0.006038135848939419, + 0.004059917759150267, + 0.002390981186181307, + 0.006002840120345354, + 0.0033138704020529985, + 0.0036648602690547705, + 0.008709556423127651, + 0.004100358113646507, + 0.008845779113471508, + 0.005730305332690477, + 0.006015163846313953, + 0.011616981588304043, + 0.005785654298961163, + 0.009345324710011482, + 0.009953867644071579, + 0.008420730009675026, + 0.006059738807380199, + 0.005740920547395945, + 0.008462275378406048, + 0.00700407987460494, + 0.0064861285500228405, + 0.004468660335987806, + 0.0037347942125052214, + 0.004933160729706287, + 0.004311548080295324, + 0.004915240686386824, + 0.0034113340079784393, + 0.0063410853035748005, + 0.004746574442833662, + 0.004539376124739647, + 0.005095462780445814, + 0.004596266895532608, + 0.0055582039058208466, + 0.0032700770534574986, + 0.0025334726087749004, + 0.004033674020320177, + 0.0038572880439460278, + 0.0060529871843755245, + 0.00512374471873045, + 0.004363702144473791, + 0.005963521543890238, + 0.0033824294805526733, + 0.002013697288930416, + 0.0030206916853785515, + 0.002037142403423786, + 0.003172218566760421, + 0.005159570835530758, + 0.004490003455430269, + 0.00446737464517355, + 0.003524865722283721, + 0.0013625393621623516, + 0.001437768223695457, + 0.0013812801335006952, + 0.00312044401653111, + 0.0056236013770103455, + 0.0060309977270662785, + 0.003372613340616226, + 0.0025333170779049397, + 0.001271906541660428, + 0.0014075418002903461, + 0.0012854010565206409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.3016095757484436 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981880120933056 + ], + "opposite_logprob_tokens": [ + -3.5298187732696533 + ], + "logprob_reward": -0.02981880120933056, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981880120933056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43195661902427673, + 0.062402401119470596, + 0.035836219787597656, + 0.07432982325553894, + 0.03853804990649223, + 0.039926279336214066, + 0.11424022167921066, + 0.06409557908773422, + 0.09260091930627823, + 0.04607388749718666 + ], + "credit_map": [ + 0.037808869034051895, + 0.027010856196284294, + 0.014700772240757942, + 0.011177297681570053, + 0.17093825340270996, + 0.04952896013855934, + 0.021665658801794052, + 0.099125936627388, + 0.004500760231167078, + 0.010569272562861443, + 0.0038514204788953066, + 0.012088739313185215, + 0.007385743781924248, + 0.007377741392701864, + 0.008910352364182472, + 0.007718373090028763, + 0.004143114667385817, + 0.006114606279879808, + 0.0032641729339957237, + 0.006240935996174812, + 0.00297044450417161, + 0.0031481285113841295, + 0.006038754712790251, + 0.003916061483323574, + 0.005090018734335899, + 0.013298018835484982, + 0.007092893123626709, + 0.01685824804008007, + 0.006139611825346947, + 0.007909473031759262, + 0.011338848620653152, + 0.006602713372558355, + 0.004899514373391867, + 0.004203875549137592, + 0.0048551540821790695, + 0.0041829864494502544, + 0.002847483614459634, + 0.005823429208248854, + 0.004841732792556286, + 0.006883874535560608, + 0.004744274076074362, + 0.002614386845380068, + 0.009852772578597069, + 0.0017236137064173818, + 0.006653452292084694, + 0.005037487950176001, + 0.004303119610995054, + 0.0049971723929047585, + 0.014945605769753456, + 0.018673231825232506, + 0.01313110627233982, + 0.01758810505270958, + 0.01698131673038006, + 0.008668980561196804, + 0.013994208537042141, + 0.010257668793201447, + 0.0071462118066847324, + 0.01202619168907404, + 0.008687487803399563, + 0.009852453134953976, + 0.004915417172014713, + 0.0083236675709486, + 0.006100389175117016, + 0.007043762132525444, + 0.00841530691832304, + 0.013444938696920872, + 0.0117679787799716, + 0.009256869554519653, + 0.01481254305690527, + 0.005459111649543047, + 0.017780127003788948, + 0.011664037592709064, + 0.006912625394761562, + 0.006930590607225895, + 0.0068630678579211235, + 0.008232117630541325, + 0.007353525608778, + 0.00257131434045732, + 0.0037849850486963987, + 0.0034256591461598873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.02981880120933056 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809179365634918 + ], + "opposite_logprob_tokens": [ + -3.2880918979644775 + ], + "logprob_reward": -0.03809179365634918, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809179365634918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2573336660861969, + 0.08529788255691528, + 0.03709704801440239, + 0.25744199752807617, + 0.06476837396621704, + 0.07492271065711975, + 0.062493808567523956, + 0.06537912786006927, + 0.04440369829535484, + 0.0508616678416729 + ], + "credit_map": [ + 0.0228292066603899, + 0.016674386337399483, + 0.010866575874388218, + 0.01099485158920288, + 0.0992102324962616, + 0.03525463119149208, + 0.014183200895786285, + 0.04732058569788933, + 0.010431219823658466, + 0.013550490140914917, + 0.005532620940357447, + 0.012038799934089184, + 0.008374731056392193, + 0.016226286068558693, + 0.008578460663557053, + 0.010565274395048618, + 0.0041198753751814365, + 0.00618887972086668, + 0.0027365516871213913, + 0.005407954100519419, + 0.004753120243549347, + 0.005026696249842644, + 0.004898750223219395, + 0.003965219482779503, + 0.015145893208682537, + 0.037498846650123596, + 0.018154390156269073, + 0.03273239731788635, + 0.02765781059861183, + 0.042472850531339645, + 0.05029645934700966, + 0.0334833599627018, + 0.00899310689419508, + 0.006703554652631283, + 0.00968338456004858, + 0.006764502264559269, + 0.008228068239986897, + 0.007490915711969137, + 0.009631572291254997, + 0.007273268420249224, + 0.009206341579556465, + 0.00932129193097353, + 0.00885072909295559, + 0.007562557701021433, + 0.010312742553651333, + 0.008135601878166199, + 0.013282764703035355, + 0.008250677958130836, + 0.00756338145583868, + 0.011674322187900543, + 0.008177938871085644, + 0.012227191589772701, + 0.007396903354674578, + 0.0028704882133752108, + 0.007652340456843376, + 0.004931239876896143, + 0.011172108352184296, + 0.011409823782742023, + 0.009803983382880688, + 0.012631693854928017, + 0.005327776540070772, + 0.0043282536789774895, + 0.006421323399990797, + 0.004284167196601629, + 0.004588741809129715, + 0.006061446852982044, + 0.0062013184651732445, + 0.005460388958454132, + 0.0071401591412723064, + 0.003148287069052458, + 0.007314182817935944, + 0.004489175509661436, + 0.007623965851962566, + 0.008015255443751812, + 0.006437049247324467, + 0.007481354754418135, + 0.007389223203063011, + 0.003267427207902074, + 0.005621570628136396, + 0.005025819875299931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.03809179365634918 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031949449330568314 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031949449330568314, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031949449330568314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5645425319671631, + 0.0383809395134449, + 0.026279622688889503, + 0.046244580298662186, + 0.08766137808561325, + 0.05437086522579193, + 0.034789543598890305, + 0.10105124861001968, + 0.023103754967451096, + 0.023575514554977417 + ], + "credit_map": [ + 0.05121808871626854, + 0.0357695035636425, + 0.012409036979079247, + 0.015393177978694439, + 0.23420695960521698, + 0.0781334936618805, + 0.025296825915575027, + 0.11211545020341873, + 0.004544968716800213, + 0.00676957331597805, + 0.0030076364055275917, + 0.004796032328158617, + 0.004088103771209717, + 0.007579776458442211, + 0.0033287773840129375, + 0.004266068805009127, + 0.0025839845184236765, + 0.003902324941009283, + 0.004162963479757309, + 0.004355015233159065, + 0.0028668870218098164, + 0.0020289078820496798, + 0.004230402875691652, + 0.0021491365041583776, + 0.0021762591786682606, + 0.007355321664363146, + 0.0033061287831515074, + 0.0061312485486269, + 0.004333442077040672, + 0.005463056266307831, + 0.012760339304804802, + 0.004718787502497435, + 0.009638899005949497, + 0.011346084997057915, + 0.009475133381783962, + 0.007143392227590084, + 0.013834680430591106, + 0.011901568621397018, + 0.013375790789723396, + 0.010945827700197697, + 0.006575636565685272, + 0.002841013250872493, + 0.011999375186860561, + 0.003831323469057679, + 0.006741152610629797, + 0.010085332207381725, + 0.005354532040655613, + 0.0069424971006810665, + 0.006743338890373707, + 0.002942141145467758, + 0.0050707473419606686, + 0.002882360015064478, + 0.006037828512489796, + 0.0026614503003656864, + 0.004743495490401983, + 0.003708184463903308, + 0.015987778082489967, + 0.016557516530156136, + 0.01314419787377119, + 0.022430939599871635, + 0.008502279408276081, + 0.00778212072327733, + 0.01074977032840252, + 0.005896641407161951, + 0.002581661567091942, + 0.003994232043623924, + 0.003608791623264551, + 0.002379000186920166, + 0.004423591308295727, + 0.0018879111157730222, + 0.002185859251767397, + 0.0020427084527909756, + 0.002381476340815425, + 0.003686529118567705, + 0.005038742441684008, + 0.003984647337347269, + 0.0032777762971818447, + 0.0014354781014844775, + 0.0018443159060552716, + 0.0019265501759946346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.0031949449330568314 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00027092601521871984 + ], + "opposite_logprob_tokens": [ + -8.25027084350586 + ], + "logprob_reward": -0.00027092601521871984, + "yes_no_margin_reward": 8.25, + "token_credit_reward": -0.00027092601521871984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18715845048427582, + 0.08777101337909698, + 0.09925224632024765, + 0.09694548696279526, + 0.1427919864654541, + 0.06930005550384521, + 0.08405496925115585, + 0.10572017729282379, + 0.06041606515645981, + 0.0665896013379097 + ], + "credit_map": [ + 0.023214183747768402, + 0.01600106991827488, + 0.01537405326962471, + 0.02318805456161499, + 0.028165843337774277, + 0.022709986194968224, + 0.025845790281891823, + 0.03265944495797157, + 0.007184008602052927, + 0.014812862500548363, + 0.007839770056307316, + 0.014687687158584595, + 0.011159845627844334, + 0.009782271459698677, + 0.012585471384227276, + 0.009719086810946465, + 0.010677102953195572, + 0.017861632630228996, + 0.009098039008677006, + 0.01624559424817562, + 0.009563230909407139, + 0.007730964105576277, + 0.018210697919130325, + 0.009864975698292255, + 0.008317732252180576, + 0.015339324250817299, + 0.00647184532135725, + 0.015696126967668533, + 0.010518659837543964, + 0.013573979958891869, + 0.018357301130890846, + 0.008670511655509472, + 0.016106050461530685, + 0.01770489476621151, + 0.017069030553102493, + 0.015052182599902153, + 0.01857137121260166, + 0.016432499513030052, + 0.023193592205643654, + 0.018662353977560997, + 0.009194277226924896, + 0.005339029245078564, + 0.012810721062123775, + 0.0058387103490531445, + 0.008218806236982346, + 0.008688235655426979, + 0.009107518941164017, + 0.010102749802172184, + 0.012829671613872051, + 0.0096051674336195, + 0.011465050280094147, + 0.012896533124148846, + 0.012176513671875, + 0.0062127867713570595, + 0.01225703489035368, + 0.006612205877900124, + 0.017944155260920525, + 0.015370688401162624, + 0.01376332063227892, + 0.020638996735215187, + 0.010169235058128834, + 0.009565786458551884, + 0.011608265340328217, + 0.006659716833382845, + 0.006631277501583099, + 0.012070782482624054, + 0.006845927331596613, + 0.009415663778781891, + 0.009892696514725685, + 0.004181722644716501, + 0.0066508143208920956, + 0.004727174062281847, + 0.009013487957417965, + 0.008899567648768425, + 0.00909629836678505, + 0.010159051977097988, + 0.00914677046239376, + 0.004737933166325092, + 0.007256115321069956, + 0.00828036293387413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.00027092601521871984 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, worm-eye tracking shot, clinging to the dusty stone floor, follows a mysterious figure draped in a tattered black cloak. Due to the low angle, only the figure's feet, clad in scuffed dark leather boots, can be seen moving forward with a nearly silent, ghostly gait, the hem of the cloak trailing along the ground, occasionally stirring up dust. They are infiltrating the interior of a magnificent, ancient reclining Buddha temple, flanked by towering stone pillars carved with blurred reliefs of ancient deities. The only illumination comes from sporadic torches burning on the walls. The unsteady flickering flames cast long, grotesque, distorted shadows across the massive pillars and the surrounding, sinister guardian deities. These shadows seem to come alive, dancing and expanding with every flicker of the flames. The main colors of the entire picture are deep dark colors and warm but ominous orange-red firelight. The strong contrast between light and dark and the dynamic shadows together create a dark fantasy aesthetic that is a mixture of sacredness, repression and danger, full of disturbing suspense.", + "seed": 9317, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_7/index_2244_seed_9317_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.050564773543737826, + "vlm_logprob": -0.050564773543737826, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.050564773543737826, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08774339407682419 + ], + "opposite_logprob_tokens": [ + -2.7127432823181152 + ], + "logprob_reward": -0.08774339407682419, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08774339407682419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1078973188996315, + 0.24349206686019897, + 0.03176256641745567, + 0.07758399099111557, + 0.034483350813388824, + 0.05366657301783562, + 0.03913678973913193, + 0.07583342492580414, + 0.025145482271909714, + 0.3109983801841736 + ], + "credit_map": [ + 0.014523819088935852, + 0.01146650966256857, + 0.01007444504648447, + 0.006882981397211552, + 0.019147492945194244, + 0.020356984809041023, + 0.00960978027433157, + 0.01583530753850937, + 0.026921669021248817, + 0.03879018872976303, + 0.03669861704111099, + 0.03115064650774002, + 0.023739226162433624, + 0.02109290100634098, + 0.034397710114717484, + 0.030701104551553726, + 0.0032547018490731716, + 0.004502293653786182, + 0.004677992779761553, + 0.005283681210130453, + 0.0038271588273346424, + 0.003292757784947753, + 0.003792179049924016, + 0.003131801262497902, + 0.007764566224068403, + 0.011734329164028168, + 0.006842483766376972, + 0.009910933673381805, + 0.013342497870326042, + 0.008277074433863163, + 0.010417118668556213, + 0.009294981136918068, + 0.0031145873945206404, + 0.004289303440600634, + 0.0039368318393826485, + 0.0030681516509503126, + 0.005845607724040747, + 0.004513212013989687, + 0.007156675215810537, + 0.002558979904279113, + 0.007171279285103083, + 0.0063666850328445435, + 0.006761160213500261, + 0.00548112066462636, + 0.006656762678176165, + 0.007873833179473877, + 0.007161098998039961, + 0.006194634363055229, + 0.004214141983538866, + 0.005659609101712704, + 0.0052904472686350346, + 0.0041038827039301395, + 0.005272137466818094, + 0.005398574285209179, + 0.004729118663817644, + 0.004468875005841255, + 0.010754871182143688, + 0.012549479492008686, + 0.009106986224651337, + 0.008150873705744743, + 0.007415898144245148, + 0.006070921663194895, + 0.01095908135175705, + 0.010825314559042454, + 0.005599634721875191, + 0.0021508047357201576, + 0.004514728672802448, + 0.0020546612795442343, + 0.0018760629463940859, + 0.0012272379826754332, + 0.0023048350121825933, + 0.005417516455054283, + 0.058817874640226364, + 0.026415755972266197, + 0.07445897161960602, + 0.025310030207037926, + 0.03312626853585243, + 0.037693433463573456, + 0.026418523862957954, + 0.02875753678381443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.08774339407682419 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033480077981949 + ], + "opposite_logprob_tokens": [ + -2.350334882736206 + ], + "logprob_reward": -0.10033480077981949, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033480077981949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09246884286403656, + 0.0967603400349617, + 0.030150234699249268, + 0.17803099751472473, + 0.19461625814437866, + 0.18918396532535553, + 0.0790850967168808, + 0.04531630873680115, + 0.06357838213443756, + 0.030809583142399788 + ], + "credit_map": [ + 0.01173518318682909, + 0.011461692862212658, + 0.009778227657079697, + 0.006318687926977873, + 0.01471087895333767, + 0.01921849325299263, + 0.006698746234178543, + 0.012546923011541367, + 0.011046426370739937, + 0.019423549994826317, + 0.015553039498627186, + 0.011529114097356796, + 0.009834547527134418, + 0.009785853326320648, + 0.010856819339096546, + 0.008730989880859852, + 0.0033073441591113806, + 0.004691064357757568, + 0.0032168372999876738, + 0.0036377159412950277, + 0.0035878147464245558, + 0.0032361792400479317, + 0.0057015432976186275, + 0.002771733095869422, + 0.01533359196037054, + 0.033673614263534546, + 0.011801440268754959, + 0.01725226640701294, + 0.03905875235795975, + 0.017599837854504585, + 0.018066706135869026, + 0.02524477429687977, + 0.018515072762966156, + 0.033233728259801865, + 0.017003437504172325, + 0.01671464927494526, + 0.03314516693353653, + 0.020779602229595184, + 0.04019917547702789, + 0.015025407075881958, + 0.030012065544724464, + 0.02243007719516754, + 0.02194536291062832, + 0.013656691648066044, + 0.029077576473355293, + 0.026557808741927147, + 0.025687456130981445, + 0.019816916435956955, + 0.008457128889858723, + 0.011814005672931671, + 0.010980636812746525, + 0.006643504370003939, + 0.013387296348810196, + 0.01037663221359253, + 0.009344817139208317, + 0.008081072010099888, + 0.005996952764689922, + 0.0074009052477777, + 0.00481737544760108, + 0.004610637668520212, + 0.006393680348992348, + 0.003963307943195105, + 0.005440534558147192, + 0.006692913360893726, + 0.02217531017959118, + 0.004370693117380142, + 0.0048931376077234745, + 0.0026609166525304317, + 0.003066814737394452, + 0.00268155406229198, + 0.0028452796395868063, + 0.020884674042463303, + 0.0039626117795705795, + 0.0031912089325487614, + 0.0054531279020011425, + 0.003227437846362591, + 0.004001629073172808, + 0.004117831587791443, + 0.00301471259444952, + 0.003841019468382001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.10033480077981949 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062037866562604904 + ], + "opposite_logprob_tokens": [ + -2.812037944793701 + ], + "logprob_reward": -0.062037866562604904, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062037866562604904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19920390844345093, + 0.06574554741382599, + 0.11660598963499069, + 0.05159514769911766, + 0.052974898368120193, + 0.06578338891267776, + 0.09630850702524185, + 0.11466123163700104, + 0.158413365483284, + 0.07870800793170929 + ], + "credit_map": [ + 0.028308315202593803, + 0.011600128374993801, + 0.023366352543234825, + 0.011426067911088467, + 0.022812124341726303, + 0.05258908122777939, + 0.01640668697655201, + 0.03269514441490173, + 0.0074464790523052216, + 0.010581441223621368, + 0.010871394537389278, + 0.010884028859436512, + 0.009006465785205364, + 0.006942254025489092, + 0.005480992142111063, + 0.004532494116574526, + 0.010906928218901157, + 0.017062602564692497, + 0.01058491412550211, + 0.015031958930194378, + 0.017117910087108612, + 0.01855226792395115, + 0.013679912313818932, + 0.013669501058757305, + 0.005928622558712959, + 0.008040611632168293, + 0.005749151110649109, + 0.005232620518654585, + 0.010017773136496544, + 0.004628641065210104, + 0.0057582128793001175, + 0.006239515263587236, + 0.00402293773368001, + 0.007468838710337877, + 0.005301555618643761, + 0.004658874589949846, + 0.008991697803139687, + 0.007256677839905024, + 0.00946726556867361, + 0.00580704677850008, + 0.01115141250193119, + 0.007819593884050846, + 0.008477226831018925, + 0.004978892859071493, + 0.007133856415748596, + 0.008908523246645927, + 0.008105873130261898, + 0.009208010509610176, + 0.011768564581871033, + 0.02047731541097164, + 0.01245160959661007, + 0.009224548004567623, + 0.011200128123164177, + 0.01189491804689169, + 0.010419625788927078, + 0.008871791884303093, + 0.015577980317175388, + 0.02038704976439476, + 0.016201235353946686, + 0.011289198882877827, + 0.016329802572727203, + 0.008983658626675606, + 0.011510543525218964, + 0.014381763525307178, + 0.048624128103256226, + 0.012196092866361141, + 0.022109011188149452, + 0.009351114742457867, + 0.005437775049358606, + 0.00581680191680789, + 0.0066930269822478294, + 0.04818542301654816, + 0.010547991842031479, + 0.010683087632060051, + 0.008829833008348942, + 0.010805361904203892, + 0.0113570187240839, + 0.009198782034218311, + 0.009622016921639442, + 0.00766392145305872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.062037866562604904 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024925374891608953 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024925374891608953, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024925374891608953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08042757958173752, + 0.03300728276371956, + 0.08746933192014694, + 0.060682643204927444, + 0.06299841403961182, + 0.1708044558763504, + 0.1650753617286682, + 0.09582921862602234, + 0.028923850506544113, + 0.21478188037872314 + ], + "credit_map": [ + 0.011547797359526157, + 0.010820794850587845, + 0.0075879404321312904, + 0.0073307272978127, + 0.015747202560305595, + 0.015434314496815205, + 0.005256861913949251, + 0.0067019336856901646, + 0.0036322663072496653, + 0.005154459271579981, + 0.004092530347406864, + 0.005892745219171047, + 0.00454560574144125, + 0.0032273796387016773, + 0.0036542038433253765, + 0.0028080909978598356, + 0.007205684203654528, + 0.012678493745625019, + 0.009791851975023746, + 0.011986766010522842, + 0.014180712401866913, + 0.012600717134773731, + 0.008958691731095314, + 0.010066410526633263, + 0.007040319498628378, + 0.01066234428435564, + 0.007556079886853695, + 0.006876426283270121, + 0.011143948882818222, + 0.00657530315220356, + 0.005241456441581249, + 0.005586755927652121, + 0.006115215830504894, + 0.009260973893105984, + 0.0052508884109556675, + 0.0037890062667429447, + 0.010001225396990776, + 0.0075878300704061985, + 0.01491626724600792, + 0.006077003665268421, + 0.0275892224162817, + 0.030934542417526245, + 0.01916719228029251, + 0.015079528093338013, + 0.023037921637296677, + 0.02191932499408722, + 0.01851225271821022, + 0.014564472250640392, + 0.015145995654165745, + 0.02650460973381996, + 0.026034323498606682, + 0.032983116805553436, + 0.014138558879494667, + 0.020253175869584084, + 0.014311571605503559, + 0.015703992918133736, + 0.012413088232278824, + 0.01388783473521471, + 0.01096651516854763, + 0.010037600062787533, + 0.015130719169974327, + 0.009290988557040691, + 0.01030889991670847, + 0.013793566264212132, + 0.006284170318394899, + 0.004053598269820213, + 0.004255829844623804, + 0.0025036020670086145, + 0.0035355924628674984, + 0.0016161644598469138, + 0.0017850719159469008, + 0.004889816511422396, + 0.04067666828632355, + 0.015987496823072433, + 0.051536913961172104, + 0.017567088827490807, + 0.025215167552232742, + 0.025064725428819656, + 0.0185552965849638, + 0.020178500562906265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.0024925374891608953 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00021526881027966738 + ], + "opposite_logprob_tokens": [ + -8.500215530395508 + ], + "logprob_reward": -0.00021526881027966738, + "yes_no_margin_reward": 8.5, + "token_credit_reward": -0.00021526881027966738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13432204723358154, + 0.0650283470749855, + 0.09088509529829025, + 0.1826753467321396, + 0.1553218811750412, + 0.08934406191110611, + 0.06873149424791336, + 0.06238938868045807, + 0.028394058346748352, + 0.12290825694799423 + ], + "credit_map": [ + 0.017624519765377045, + 0.014767236076295376, + 0.01439112052321434, + 0.013228379189968109, + 0.02339782752096653, + 0.0257610771805048, + 0.00999719649553299, + 0.015154705382883549, + 0.007134362123906612, + 0.013414792716503143, + 0.007166096940636635, + 0.009565340355038643, + 0.007553172297775745, + 0.006801592651754618, + 0.0073019093833863735, + 0.00609108479693532, + 0.007901899516582489, + 0.013007642701268196, + 0.011155834421515465, + 0.01582786627113819, + 0.010161405429244041, + 0.0103379525244236, + 0.01331883855164051, + 0.009173667058348656, + 0.02787678875029087, + 0.03129112347960472, + 0.021820012480020523, + 0.0181924719363451, + 0.03388740122318268, + 0.016398770734667778, + 0.013709441758692265, + 0.019499333575367928, + 0.01693376898765564, + 0.02778674103319645, + 0.013006879016757011, + 0.010321004316210747, + 0.028093848377466202, + 0.016564618796110153, + 0.03352853283286095, + 0.009086485020816326, + 0.009899253956973553, + 0.016875537112355232, + 0.008985384367406368, + 0.009302155114710331, + 0.008987721055746078, + 0.010422817431390285, + 0.013185689225792885, + 0.011685503646731377, + 0.009341089986264706, + 0.01424392405897379, + 0.008353867568075657, + 0.006068350747227669, + 0.008362529799342155, + 0.008389980532228947, + 0.007421161979436874, + 0.0065505970269441605, + 0.008320450782775879, + 0.012285878881812096, + 0.008699888363480568, + 0.006347128190100193, + 0.0069678770378232, + 0.005035567097365856, + 0.0042795720510184765, + 0.010453032329678535, + 0.007164841052144766, + 0.002128191525116563, + 0.0037867086939513683, + 0.002341304440051317, + 0.0022333350498229265, + 0.0013787730131298304, + 0.0022686421871185303, + 0.0070922658778727055, + 0.020022664219141006, + 0.009477688930928707, + 0.027544744312763214, + 0.009912334382534027, + 0.010112373158335686, + 0.01787804625928402, + 0.010171028785407543, + 0.017789388075470924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.00021526881027966738 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, worm-eye tracking shot, clinging to the dusty stone floor, follows a mysterious figure draped in a tattered black cloak. Due to the low angle, only the figure's feet, clad in scuffed dark leather boots, can be seen moving forward with a nearly silent, ghostly gait, the hem of the cloak trailing along the ground, occasionally stirring up dust. They are infiltrating the interior of a magnificent, ancient reclining Buddha temple, flanked by towering stone pillars carved with blurred reliefs of ancient deities. The only illumination comes from sporadic torches burning on the walls. The unsteady flickering flames cast long, grotesque, distorted shadows across the massive pillars and the surrounding, sinister guardian deities. These shadows seem to come alive, dancing and expanding with every flicker of the flames. The main colors of the entire picture are deep dark colors and warm but ominous orange-red firelight. The strong contrast between light and dark and the dynamic shadows together create a dark fantasy aesthetic that is a mixture of sacredness, repression and danger, full of disturbing suspense.", + "seed": 9318, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_7/index_2244_seed_9318_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07641436270205304, + "vlm_logprob": -0.07641436270205304, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.07641436270205304, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22603414952754974 + ], + "opposite_logprob_tokens": [ + -1.726034164428711 + ], + "logprob_reward": -0.22603414952754974, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.22603414952754974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19956019520759583, + 0.06566688418388367, + 0.054881393909454346, + 0.1867741197347641, + 0.1615249365568161, + 0.16146817803382874, + 0.05626552924513817, + 0.039827294647693634, + 0.04266803339123726, + 0.03136341646313667 + ], + "credit_map": [ + 0.042299941182136536, + 0.031385958194732666, + 0.023992203176021576, + 0.019156688824295998, + 0.03313601016998291, + 0.015127002261579037, + 0.02512887306511402, + 0.009333515539765358, + 0.00715397484600544, + 0.014723379164934158, + 0.006814636755734682, + 0.0057018413208425045, + 0.01380765251815319, + 0.004968605935573578, + 0.008285846561193466, + 0.004210944287478924, + 0.004890107084065676, + 0.009337067604064941, + 0.005498729646205902, + 0.0064919195137917995, + 0.010502791032195091, + 0.004492831416428089, + 0.008059198968112469, + 0.005608749110251665, + 0.02024509757757187, + 0.018650494515895844, + 0.014025031588971615, + 0.02282477356493473, + 0.03842015191912651, + 0.037906747311353683, + 0.013781442306935787, + 0.020920392125844955, + 0.012396308593451977, + 0.012863450683653355, + 0.01125076413154602, + 0.014975357800722122, + 0.034065622836351395, + 0.04428137466311455, + 0.014761175028979778, + 0.016930880025029182, + 0.0033088538330048323, + 0.005271601025015116, + 0.002444928279146552, + 0.0032785627990961075, + 0.06908543407917023, + 0.06799860298633575, + 0.003435653168708086, + 0.006644542329013348, + 0.005008585285395384, + 0.0059725739993155, + 0.004069921560585499, + 0.0048973653465509415, + 0.012827214784920216, + 0.011158378794789314, + 0.0056741610169410706, + 0.0066573284566402435, + 0.004581442568451166, + 0.004372929688543081, + 0.002815737621858716, + 0.0035524351987987757, + 0.007117467001080513, + 0.007336876355111599, + 0.005429836921393871, + 0.004620567429810762, + 0.0029749921523034573, + 0.005193332210183144, + 0.0022293603979051113, + 0.0038642834406346083, + 0.003364365780726075, + 0.009020328521728516, + 0.004937056917697191, + 0.011084314435720444, + 0.0035978169180452824, + 0.003920162096619606, + 0.003019531024619937, + 0.002240039175376296, + 0.0056737978011369705, + 0.004715414252132177, + 0.004779692739248276, + 0.003416958497837186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.22603414952754974 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908002823591232 + ], + "opposite_logprob_tokens": [ + -2.4640800952911377 + ], + "logprob_reward": -0.08908002823591232, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908002823591232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19850920140743256, + 0.058813538402318954, + 0.06170876696705818, + 0.06851398944854736, + 0.08016380667686462, + 0.34955206513404846, + 0.022631769999861717, + 0.0452844612300396, + 0.07328306883573532, + 0.0415392704308033 + ], + "credit_map": [ + 0.029533125460147858, + 0.02319144643843174, + 0.02723821811378002, + 0.025557857006788254, + 0.027217041701078415, + 0.023319127038121223, + 0.021038834005594254, + 0.021413564682006836, + 0.00345567986369133, + 0.014730566181242466, + 0.004931593779474497, + 0.006381231360137463, + 0.014154257252812386, + 0.0025044709909707308, + 0.0103880250826478, + 0.0022677131928503513, + 0.004026622045785189, + 0.010615688748657703, + 0.004630328621715307, + 0.008009505458176136, + 0.012925253249704838, + 0.005776598118245602, + 0.008829481899738312, + 0.006895285565406084, + 0.007627920247614384, + 0.007897131145000458, + 0.005778266582638025, + 0.009152861312031746, + 0.01208720076829195, + 0.01317384373396635, + 0.005592514295130968, + 0.007204253226518631, + 0.008560475893318653, + 0.01060300599783659, + 0.007378246635198593, + 0.009559526108205318, + 0.012877861969172955, + 0.015356735326349735, + 0.0066032917238771915, + 0.009224665351212025, + 0.004482792690396309, + 0.009666313417255878, + 0.0029477053321897984, + 0.004819008521735668, + 0.15742303431034088, + 0.15603798627853394, + 0.007046491373330355, + 0.007128764409571886, + 0.0018219971098005772, + 0.0022621916141361, + 0.0024253963492810726, + 0.0015994381392374635, + 0.004996385890990496, + 0.003992775920778513, + 0.0032882201485335827, + 0.0022453658748418093, + 0.005870957858860493, + 0.005090522579848766, + 0.003103566123172641, + 0.004369765520095825, + 0.007517051883041859, + 0.008466949686408043, + 0.0050780740566551685, + 0.0057875728234648705, + 0.004869516938924789, + 0.009670610539615154, + 0.004869350232183933, + 0.015388372354209423, + 0.010550545528531075, + 0.010421080514788628, + 0.008591903373599052, + 0.00892169401049614, + 0.003399241017177701, + 0.005359152797609568, + 0.005084225442260504, + 0.0042571742087602615, + 0.008516073226928711, + 0.005865850485861301, + 0.004659211728721857, + 0.00439834501594305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.08908002823591232 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062041785567998886 + ], + "opposite_logprob_tokens": [ + -2.812041759490967 + ], + "logprob_reward": -0.062041785567998886, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062041785567998886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09838254004716873, + 0.08047281950712204, + 0.05543849617242813, + 0.11575049161911011, + 0.0753280371427536, + 0.3030690550804138, + 0.13126972317695618, + 0.06114833429455757, + 0.030895670875906944, + 0.048244912177324295 + ], + "credit_map": [ + 0.02661767229437828, + 0.0101816700771451, + 0.01916089467704296, + 0.008627092465758324, + 0.01564527489244938, + 0.005013079848140478, + 0.00767999654635787, + 0.005456854589283466, + 0.008828824386000633, + 0.019801635295152664, + 0.0072816116735339165, + 0.005079843103885651, + 0.01800229772925377, + 0.009772075340151787, + 0.006927324924618006, + 0.004779207985848188, + 0.006455910857766867, + 0.010225072503089905, + 0.005388081073760986, + 0.007239884696900845, + 0.008054307661950588, + 0.007859444245696068, + 0.006476249545812607, + 0.003739544888958335, + 0.01329509075731039, + 0.011676396243274212, + 0.009537719190120697, + 0.01308487355709076, + 0.022669970989227295, + 0.023134861141443253, + 0.008603483438491821, + 0.013748086988925934, + 0.007028283551335335, + 0.00936384592205286, + 0.006482290104031563, + 0.008615662343800068, + 0.0131881944835186, + 0.018025793135166168, + 0.006226783152669668, + 0.006397185381501913, + 0.0029683231841772795, + 0.005219906102865934, + 0.0020531087648123503, + 0.0035478719510138035, + 0.14090301096439362, + 0.1405627578496933, + 0.0031912182457745075, + 0.004622830078005791, + 0.014391260221600533, + 0.013644608668982983, + 0.01392181497067213, + 0.012044773437082767, + 0.028282467275857925, + 0.024819882586598396, + 0.013437430374324322, + 0.01072748564183712, + 0.006775187328457832, + 0.00799989141523838, + 0.0034174928441643715, + 0.00481270020827651, + 0.012476428411900997, + 0.011948889121413231, + 0.0070900870487093925, + 0.006627664901316166, + 0.003445602022111416, + 0.004003825131803751, + 0.0024312396999448538, + 0.00394843053072691, + 0.003097545588389039, + 0.005297357216477394, + 0.002936495700851083, + 0.005735177546739578, + 0.003827620530501008, + 0.004795958288013935, + 0.005400324240326881, + 0.004580322653055191, + 0.010472930036485195, + 0.007226540707051754, + 0.0071479142643511295, + 0.004793302156031132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.062041785567998886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046426597982645035 + ], + "opposite_logprob_tokens": [ + -5.379642486572266 + ], + "logprob_reward": -0.0046426597982645035, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046426597982645035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06989257037639618, + 0.05675091594457626, + 0.03403230011463165, + 0.357215940952301, + 0.05065812170505524, + 0.15796507894992828, + 0.15849769115447998, + 0.048753347247838974, + 0.020082524046301842, + 0.04615148529410362 + ], + "credit_map": [ + 0.017219461500644684, + 0.007325441110879183, + 0.01250782236456871, + 0.006033680867403746, + 0.01126511674374342, + 0.005944720469415188, + 0.005640134215354919, + 0.0039561958983540535, + 0.004903757944703102, + 0.012087357230484486, + 0.005794599652290344, + 0.007155661936849356, + 0.010443825274705887, + 0.004725589416921139, + 0.007706872187554836, + 0.003933255560696125, + 0.002799204783514142, + 0.006895130034536123, + 0.0025938558392226696, + 0.0036677184980362654, + 0.007209299132227898, + 0.0025883133057504892, + 0.003052191575989127, + 0.005226590670645237, + 0.03643917292356491, + 0.03914922848343849, + 0.026827773079276085, + 0.047064267098903656, + 0.07028516381978989, + 0.07191836088895798, + 0.027309924364089966, + 0.03822207823395729, + 0.004636358004063368, + 0.006360645871609449, + 0.004499964416027069, + 0.007173982914537191, + 0.008079543709754944, + 0.010783037170767784, + 0.004607079084962606, + 0.004517517983913422, + 0.002233287086710334, + 0.0053710839711129665, + 0.0014806975377723575, + 0.0026236167177557945, + 0.07049786299467087, + 0.0703275203704834, + 0.001998831983655691, + 0.003432192839682102, + 0.01696760766208172, + 0.016080185770988464, + 0.017268288880586624, + 0.01625845581293106, + 0.03353913500905037, + 0.029597105458378792, + 0.015873001888394356, + 0.012913936749100685, + 0.007000407669693232, + 0.006133256945759058, + 0.004057782236486673, + 0.004714156035333872, + 0.0072218687273561954, + 0.00873685535043478, + 0.005077152047306299, + 0.005811871029436588, + 0.0020649698562920094, + 0.0020563346333801746, + 0.0016296453541144729, + 0.0020757762249559164, + 0.002431317465379834, + 0.0036533919628709555, + 0.001789125264622271, + 0.004381964448839426, + 0.00415790406987071, + 0.0051824068650603294, + 0.005410617683082819, + 0.003927628509700298, + 0.008557910099625587, + 0.00669873645529151, + 0.005813382100313902, + 0.006402900442481041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.0046426597982645035 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, specifically a worm-eye view, looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a tracking shot that follows a mysterious figure draped in a tattered black cloak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the interior of an ancient reclining Buddha temple featuring towering stone pillars with carved reliefs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated primarily by sporadic torches on the walls, casting long, distorted shadows that appear to dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the dominant colors of the scene deep dark tones contrasted with warm orange-red firelight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0002731903805397451 + ], + "opposite_logprob_tokens": [ + -8.250272750854492 + ], + "logprob_reward": -0.0002731903805397451, + "yes_no_margin_reward": 8.25, + "token_credit_reward": -0.0002731903805397451, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13639475405216217, + 0.06546624004840851, + 0.04032839834690094, + 0.2229289710521698, + 0.08456788957118988, + 0.32539305090904236, + 0.031098606064915657, + 0.029025355353951454, + 0.023617472499608994, + 0.04117928445339203 + ], + "credit_map": [ + 0.03134404867887497, + 0.013319793157279491, + 0.021696817129850388, + 0.014525426551699638, + 0.021226132288575172, + 0.009482045657932758, + 0.013256511650979519, + 0.011543980799615383, + 0.006006781477481127, + 0.013265042565762997, + 0.008331065997481346, + 0.005617233924567699, + 0.013600582256913185, + 0.00572117930278182, + 0.007962189614772797, + 0.004962161183357239, + 0.00404170248657465, + 0.006685198750346899, + 0.003803775878623128, + 0.005466854199767113, + 0.008276082575321198, + 0.0035208193585276604, + 0.004796365275979042, + 0.003737600287422538, + 0.02393171563744545, + 0.024074137210845947, + 0.017124228179454803, + 0.02867749147117138, + 0.04503287002444267, + 0.044164471328258514, + 0.016191300004720688, + 0.023732760921120644, + 0.007998447865247726, + 0.009671036154031754, + 0.0065755704417824745, + 0.008735565468668938, + 0.015272894874215126, + 0.02120191790163517, + 0.007389006670564413, + 0.007723446469753981, + 0.004347886424511671, + 0.007092516403645277, + 0.0026840383652597666, + 0.0038237099070101976, + 0.15000854432582855, + 0.14971157908439636, + 0.003535924945026636, + 0.00418885936960578, + 0.0028111899737268686, + 0.0034938326571136713, + 0.0029801579657942057, + 0.0023334533907473087, + 0.006149384193122387, + 0.006074066739529371, + 0.0037841128651052713, + 0.0034724068827927113, + 0.0035618881229311228, + 0.0041127982549369335, + 0.0024673831649124622, + 0.002883634762838483, + 0.004079271573573351, + 0.005075262393802404, + 0.0035545045975595713, + 0.0032906129490584135, + 0.0018783960258588195, + 0.002284632297232747, + 0.0022343818563967943, + 0.0023584391456097364, + 0.002641581231728196, + 0.00462198443710804, + 0.002321976702660322, + 0.005276079755276442, + 0.002900121733546257, + 0.0037633602041751146, + 0.004389268811792135, + 0.0034013953991234303, + 0.007968430407345295, + 0.0066813211888074875, + 0.005615910515189171, + 0.006459477823227644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.0002731903805397451 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_8/index_tensor([4652, 4652])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_8/index_tensor([4652, 4652])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..800fc99a14b10e369fef6ea4dbc4ab1ab029d807 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_8/index_tensor([4652, 4652])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly advancing close-up shot first focuses on a vast mural depicting an ancient agricultural scene. Its mottled, earthy-yellow surface is riddled with tiny, spiderweb-like cracks. In the dim, serene surroundings, a mysterious blue light begins to cautiously seep from deep within these cracks, emitting a soft yet energetic glow that illuminates the surrounding rough texture of the wall. The camera then shifts focus to an ox plowing a field at the center of the mural. Its body, rendered in thick mineral pigment, boasts powerful yet archaic muscular lines. Yet, one of its eyes possesses an unusually deep depth, as if embedded in black obsidian. As a streak of blue light reaches this eye, a subtle ripple suddenly ripples through the still pupil, rapidly expanding to become transparent yet visible ripples in space and time, like ripples on water. These ripples, with the pupils of the plowing ox as their epicenter, spread steadily and calmly across the entire painting. Wherever they reach, the ancient pigments on the mural, the cracks in the wall, and even the air itself are slightly distorted and refracted, as if the physical laws of the entire space are being gently rewritten, creating a mysterious, solemn, and surreal atmosphere full of science fiction wonder.", + "seed": 29738, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_8/index_4652_seed_29738_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02585700377821922, + "vlm_logprob": -0.02585700377821922, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.02585700377821922, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07786467671394348 + ], + "opposite_logprob_tokens": [ + -2.827864646911621 + ], + "logprob_reward": -0.07786467671394348, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07786467671394348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11457520723342896, + 0.08948780596256256, + 0.038322120904922485, + 0.08509434759616852, + 0.07314994931221008, + 0.16634690761566162, + 0.24511025846004486, + 0.10775914788246155, + 0.040655575692653656, + 0.03949866443872452 + ], + "credit_map": [ + 0.010990304872393608, + 0.016213510185480118, + 0.01536374632269144, + 0.016748690977692604, + 0.015844719484448433, + 0.012263366021215916, + 0.017118122428655624, + 0.010032746009528637, + 0.00889996811747551, + 0.009039866738021374, + 0.014108095318078995, + 0.009916149079799652, + 0.016181645914912224, + 0.010392512194812298, + 0.010213036090135574, + 0.01073653157800436, + 0.0036013764329254627, + 0.003330729203298688, + 0.0064477501437067986, + 0.005477142985910177, + 0.004027336835861206, + 0.004972225986421108, + 0.005376964807510376, + 0.005088594742119312, + 0.010092311538755894, + 0.014429919421672821, + 0.007264562416821718, + 0.009097659960389137, + 0.005677313543856144, + 0.008690107613801956, + 0.00786032434552908, + 0.02198214828968048, + 0.006848374381661415, + 0.012737253680825233, + 0.00728312274441123, + 0.008312360383570194, + 0.01435905136168003, + 0.009308253414928913, + 0.005452405195683241, + 0.008849131874740124, + 0.022475630044937134, + 0.02225516550242901, + 0.019879980012774467, + 0.020796332508325577, + 0.019825533032417297, + 0.02397066541016102, + 0.013793259859085083, + 0.02335033193230629, + 0.02402704767882824, + 0.024986734613776207, + 0.031145986169576645, + 0.034948598593473434, + 0.03798094391822815, + 0.03934669867157936, + 0.02649819478392601, + 0.026176046580076218, + 0.01190959569066763, + 0.00823577493429184, + 0.012735504657030106, + 0.012971323914825916, + 0.02134341560304165, + 0.011365965008735657, + 0.012860305607318878, + 0.016337266191840172, + 0.00464866915717721, + 0.0029934823978692293, + 0.0038849669508635998, + 0.0035751061514019966, + 0.007280108518898487, + 0.005055869929492474, + 0.0061294082552194595, + 0.0070879640989005566, + 0.00408221734687686, + 0.0031322480645030737, + 0.0037722038105130196, + 0.004049714654684067, + 0.007245226763188839, + 0.005973032675683498, + 0.005987107288092375, + 0.005256915930658579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.07786467671394348 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006774553097784519 + ], + "opposite_logprob_tokens": [ + -5.006774425506592 + ], + "logprob_reward": -0.006774553097784519, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006774553097784519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12395599484443665, + 0.20765574276447296, + 0.05922919884324074, + 0.13786382973194122, + 0.07748138904571533, + 0.14130458235740662, + 0.07068870216608047, + 0.0628473311662674, + 0.04304521158337593, + 0.07592804729938507 + ], + "credit_map": [ + 0.014368556439876556, + 0.018728340044617653, + 0.016829052940011024, + 0.014569241553544998, + 0.016227545216679573, + 0.014475234784185886, + 0.015548963099718094, + 0.013209063559770584, + 0.014826974831521511, + 0.0318031869828701, + 0.013673611916601658, + 0.019882051274180412, + 0.05970543622970581, + 0.019835051149129868, + 0.022298285737633705, + 0.02563113160431385, + 0.0055514974519610405, + 0.004346996545791626, + 0.00912077073007822, + 0.00796492025256157, + 0.005574175622314215, + 0.006218323018401861, + 0.010561276227235794, + 0.009891241788864136, + 0.012817785143852234, + 0.0329013355076313, + 0.011888857930898666, + 0.010429180227220058, + 0.010769970715045929, + 0.010591999627649784, + 0.011080862022936344, + 0.03738383203744888, + 0.010120296850800514, + 0.012054428458213806, + 0.008144271560013294, + 0.008901946246623993, + 0.011269411072134972, + 0.01220131479203701, + 0.00592895457521081, + 0.00886076595634222, + 0.013358410447835922, + 0.013113939203321934, + 0.029550977051258087, + 0.010304264724254608, + 0.008604529313743114, + 0.011588064022362232, + 0.006466907449066639, + 0.04831749573349953, + 0.012053972110152245, + 0.009754710830748081, + 0.007799462415277958, + 0.011107517406344414, + 0.009705149568617344, + 0.006637778133153915, + 0.00740524148568511, + 0.006224870681762695, + 0.007195510435849428, + 0.004214595537632704, + 0.006465298123657703, + 0.007272656541317701, + 0.01473403163254261, + 0.004879050422459841, + 0.006970382295548916, + 0.011115807108581066, + 0.004733544774353504, + 0.0033081427682191133, + 0.004946536384522915, + 0.004273358732461929, + 0.009223474189639091, + 0.0038926885463297367, + 0.006704886443912983, + 0.005962577648460865, + 0.007237519137561321, + 0.006006873678416014, + 0.009279865771532059, + 0.013874552212655544, + 0.013784447684884071, + 0.008951453492045403, + 0.00950993224978447, + 0.007283396553248167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.006774553097784519 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016085729002952576 + ], + "opposite_logprob_tokens": [ + -4.141085624694824 + ], + "logprob_reward": -0.016085729002952576, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016085729002952576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09557828307151794, + 0.2711697220802307, + 0.07656539231538773, + 0.09073174744844437, + 0.06814666092395782, + 0.18546362221240997, + 0.05601213127374649, + 0.06517494469881058, + 0.04873332008719444, + 0.04242418706417084 + ], + "credit_map": [ + 0.009853960014879704, + 0.013571405783295631, + 0.011905043385922909, + 0.011753416620194912, + 0.017636312171816826, + 0.011438669636845589, + 0.01127084344625473, + 0.008148635737597942, + 0.02617747150361538, + 0.08585815876722336, + 0.018577223643660545, + 0.02400757372379303, + 0.028340786695480347, + 0.024962997063994408, + 0.016405601054430008, + 0.046839915215969086, + 0.008139477111399174, + 0.008185165002942085, + 0.009837022982537746, + 0.010302203707396984, + 0.011002397164702415, + 0.006885843817144632, + 0.011101522482931614, + 0.011111759580671787, + 0.011239214800298214, + 0.013097082264721394, + 0.008833510801196098, + 0.012254718691110611, + 0.010313970036804676, + 0.008066177368164062, + 0.009336373768746853, + 0.017590701580047607, + 0.007141320500522852, + 0.009973937645554543, + 0.007228522561490536, + 0.007152731530368328, + 0.009710759855806828, + 0.012680059298872948, + 0.005566434469074011, + 0.008692890405654907, + 0.0181803647428751, + 0.01672506146132946, + 0.03815508633852005, + 0.013463165611028671, + 0.01385197788476944, + 0.013248579576611519, + 0.010534815490245819, + 0.06130457669496536, + 0.009517048485577106, + 0.007717505097389221, + 0.005109814461320639, + 0.007668910548090935, + 0.007450730074197054, + 0.008093117736279964, + 0.005506553687155247, + 0.004948451183736324, + 0.007909586653113365, + 0.004299638792872429, + 0.008227690123021603, + 0.006392485462129116, + 0.01496698334813118, + 0.005968984682112932, + 0.007086986675858498, + 0.010322588495910168, + 0.004516403656452894, + 0.0030422259587794542, + 0.005101947113871574, + 0.005514212884008884, + 0.012730405665934086, + 0.003758999053388834, + 0.00690073985606432, + 0.007168381940573454, + 0.004176297690719366, + 0.003498333739116788, + 0.005266533698886633, + 0.00812320876866579, + 0.007134271785616875, + 0.0048490180633962154, + 0.004922468215227127, + 0.004454057198017836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.016085729002952576 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005261621437966824 + ], + "opposite_logprob_tokens": [ + -5.255261421203613 + ], + "logprob_reward": -0.005261621437966824, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005261621437966824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12296900153160095, + 0.09543997794389725, + 0.06698719412088394, + 0.0674472376704216, + 0.0665278285741806, + 0.10888748615980148, + 0.1478719413280487, + 0.14903846383094788, + 0.08129242807626724, + 0.09353845566511154 + ], + "credit_map": [ + 0.010866302065551281, + 0.016119765117764473, + 0.012684854678809643, + 0.019205551594495773, + 0.018703652545809746, + 0.013542050495743752, + 0.018839174881577492, + 0.013007644563913345, + 0.008994796313345432, + 0.014633175916969776, + 0.014856195077300072, + 0.013088797219097614, + 0.011961480602622032, + 0.010151330381631851, + 0.011339019984006882, + 0.010415171273052692, + 0.008171767927706242, + 0.005116589367389679, + 0.012342141009867191, + 0.005803046748042107, + 0.005068003665655851, + 0.006442304700613022, + 0.008482431061565876, + 0.015560903586447239, + 0.006957003381103277, + 0.01335659995675087, + 0.0065133837051689625, + 0.006502435076981783, + 0.006324946880340576, + 0.005951404571533203, + 0.007380405440926552, + 0.014461057260632515, + 0.007057036738842726, + 0.013466516509652138, + 0.0060281697660684586, + 0.006602198351174593, + 0.014113442040979862, + 0.0068565551191568375, + 0.004751323722302914, + 0.007652574684470892, + 0.013185132294893265, + 0.01836599037051201, + 0.012319723144173622, + 0.014516444876790047, + 0.016319165006279945, + 0.013354565016925335, + 0.00967674795538187, + 0.011149702593684196, + 0.012282587587833405, + 0.013131828978657722, + 0.02807704359292984, + 0.020298689603805542, + 0.018351443111896515, + 0.016950877383351326, + 0.014830542728304863, + 0.02394890785217285, + 0.015743490308523178, + 0.009450305253267288, + 0.015607617795467377, + 0.016649404540657997, + 0.037111129611730576, + 0.01184183731675148, + 0.018434440717101097, + 0.02420021966099739, + 0.00879399012774229, + 0.005039806477725506, + 0.007677457761019468, + 0.006515872199088335, + 0.019329944625496864, + 0.00853183027356863, + 0.012997642159461975, + 0.012405883520841599, + 0.008937428705394268, + 0.007586853113025427, + 0.012764581479132175, + 0.01586124300956726, + 0.012808545492589474, + 0.010374006815254688, + 0.013666238635778427, + 0.011539550498127937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.005261621437966824 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023298438638448715 + ], + "opposite_logprob_tokens": [ + -3.773298501968384 + ], + "logprob_reward": -0.023298438638448715, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023298438638448715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08199882507324219, + 0.0761256217956543, + 0.061062540858983994, + 0.07052452862262726, + 0.06679246574640274, + 0.14912262558937073, + 0.07325095683336258, + 0.1093270406126976, + 0.15857172012329102, + 0.15322370827198029 + ], + "credit_map": [ + 0.007275785785168409, + 0.009705588221549988, + 0.00991421565413475, + 0.01175917312502861, + 0.015972614288330078, + 0.00813321117311716, + 0.010719467885792255, + 0.008518760092556477, + 0.006975397001951933, + 0.016516923904418945, + 0.006526450626552105, + 0.007038109004497528, + 0.010007951408624649, + 0.009930653497576714, + 0.007345981430262327, + 0.01178414560854435, + 0.006591185461729765, + 0.005348449572920799, + 0.011623849160969257, + 0.007546081207692623, + 0.00455809012055397, + 0.007332432083785534, + 0.005814628675580025, + 0.012247818522155285, + 0.013142110779881477, + 0.011775770224630833, + 0.008126082830131054, + 0.010147490538656712, + 0.007985479198396206, + 0.006127543281763792, + 0.00780034065246582, + 0.005419708788394928, + 0.011128631420433521, + 0.010764756239950657, + 0.006731565110385418, + 0.0071230558678507805, + 0.007843470200896263, + 0.010985417291522026, + 0.005793389398604631, + 0.0064221699722111225, + 0.0190204456448555, + 0.02181951142847538, + 0.017549579963088036, + 0.01799059472978115, + 0.023715181276202202, + 0.018926534801721573, + 0.015482420101761818, + 0.014618339948356152, + 0.008621151559054852, + 0.006040987093001604, + 0.0161715317517519, + 0.009587650187313557, + 0.005307325627654791, + 0.006902334280312061, + 0.004622723441570997, + 0.01599724031984806, + 0.014247368089854717, + 0.009248342365026474, + 0.0115105165168643, + 0.013521624729037285, + 0.02143062837421894, + 0.009759204462170601, + 0.013403068296611309, + 0.016206270083785057, + 0.016309553757309914, + 0.007898321375250816, + 0.016297567635774612, + 0.01409225631505251, + 0.04172367602586746, + 0.011833896860480309, + 0.022160684689879417, + 0.028255758807063103, + 0.016172785311937332, + 0.01127958670258522, + 0.02072836086153984, + 0.021947801113128662, + 0.03158273175358772, + 0.020257916301488876, + 0.018064936622977257, + 0.013189573772251606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.023298438638448715 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing close-up shot first focuses on a vast mural depicting an ancient agricultural scene. Its mottled, earthy-yellow surface is riddled with tiny, spiderweb-like cracks. In the dim, serene surroundings, a mysterious blue light begins to cautiously seep from deep within these cracks, emitting a soft yet energetic glow that illuminates the surrounding rough texture of the wall. The camera then shifts focus to an ox plowing a field at the center of the mural. Its body, rendered in thick mineral pigment, boasts powerful yet archaic muscular lines. Yet, one of its eyes possesses an unusually deep depth, as if embedded in black obsidian. As a streak of blue light reaches this eye, a subtle ripple suddenly ripples through the still pupil, rapidly expanding to become transparent yet visible ripples in space and time, like ripples on water. These ripples, with the pupils of the plowing ox as their epicenter, spread steadily and calmly across the entire painting. Wherever they reach, the ancient pigments on the mural, the cracks in the wall, and even the air itself are slightly distorted and refracted, as if the physical laws of the entire space are being gently rewritten, creating a mysterious, solemn, and surreal atmosphere full of science fiction wonder.", + "seed": 29739, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_8/index_4652_seed_29739_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023145551048219203, + "vlm_logprob": -0.023145551048219203, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.023145551048219203, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07858652621507645 + ], + "opposite_logprob_tokens": [ + -2.7035865783691406 + ], + "logprob_reward": -0.07858652621507645, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07858652621507645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15980273485183716, + 0.1400957554578781, + 0.08249521255493164, + 0.12763383984565735, + 0.2704876959323883, + 0.06921035796403885, + 0.08440025895833969, + 0.03169841691851616, + 0.014366026036441326, + 0.01980971172451973 + ], + "credit_map": [ + 0.009103129617869854, + 0.02155350334942341, + 0.01839498244225979, + 0.026988105848431587, + 0.0351388156414032, + 0.009522109292447567, + 0.025560962036252022, + 0.01354113407433033, + 0.008786431513726711, + 0.019460562616586685, + 0.01908889040350914, + 0.024903470650315285, + 0.031877655535936356, + 0.007960998453199863, + 0.017108231782913208, + 0.010909509845077991, + 0.004731443710625172, + 0.013208452612161636, + 0.005042592529207468, + 0.006122063845396042, + 0.028224406763911247, + 0.009247292764484882, + 0.005560852587223053, + 0.01035810261964798, + 0.008863144554197788, + 0.022038228809833527, + 0.013359234668314457, + 0.01116457860916853, + 0.017006894573569298, + 0.014196043834090233, + 0.01152046024799347, + 0.029485249891877174, + 0.012932760640978813, + 0.011034848168492317, + 0.08895040303468704, + 0.014653347432613373, + 0.011946975253522396, + 0.1067565530538559, + 0.00880981981754303, + 0.015402988530695438, + 0.006348662078380585, + 0.013002208434045315, + 0.00596661027520895, + 0.008410453796386719, + 0.01044494193047285, + 0.009031667374074459, + 0.009686017408967018, + 0.006319799460470676, + 0.00646467600017786, + 0.008476493880152702, + 0.009555477648973465, + 0.009901267476379871, + 0.0062207793816924095, + 0.010868298821151257, + 0.011037558317184448, + 0.02187570370733738, + 0.002600910374894738, + 0.004430968780070543, + 0.0039983787573874, + 0.003674177685752511, + 0.0022078799083828926, + 0.003741869470104575, + 0.004730464890599251, + 0.006313768215477467, + 0.001359137473627925, + 0.0011571827344596386, + 0.002740102354437113, + 0.0011092418571934104, + 0.0007893560687080026, + 0.0027202763594686985, + 0.0014960495755076408, + 0.002994679380208254, + 0.0029275910928845406, + 0.002450239844620228, + 0.0030006568413227797, + 0.002182052470743656, + 0.0013631772017106414, + 0.0032049959991127253, + 0.002013917313888669, + 0.0026670817751437426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.07858652621507645 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036434237845242023 + ], + "opposite_logprob_tokens": [ + -5.62864351272583 + ], + "logprob_reward": -0.0036434237845242023, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036434237845242023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11786121875047684, + 0.19669410586357117, + 0.09990274906158447, + 0.0328187458217144, + 0.26784148812294006, + 0.03534974902868271, + 0.06245901435613632, + 0.06482039391994476, + 0.04704742133617401, + 0.07520511746406555 + ], + "credit_map": [ + 0.007435104809701443, + 0.013678515329957008, + 0.015238291583955288, + 0.022036336362361908, + 0.018549688160419464, + 0.006917961873114109, + 0.019886216148734093, + 0.014119095169007778, + 0.009052304551005363, + 0.03474492207169533, + 0.017464373260736465, + 0.024736464023590088, + 0.0645577535033226, + 0.013124119490385056, + 0.01968078501522541, + 0.013333369046449661, + 0.005769902840256691, + 0.013813893310725689, + 0.009788488037884235, + 0.01005885936319828, + 0.027967365458607674, + 0.009355224668979645, + 0.01119823008775711, + 0.01195078156888485, + 0.0038196456152945757, + 0.004800807218998671, + 0.003478900296613574, + 0.004004950635135174, + 0.004451641347259283, + 0.005048276390880346, + 0.003069210797548294, + 0.004145312123000622, + 0.012227256782352924, + 0.01669410616159439, + 0.07085596024990082, + 0.016860533505678177, + 0.015459416434168816, + 0.10245740413665771, + 0.012268857099115849, + 0.02101793885231018, + 0.0037046256475150585, + 0.005773662123829126, + 0.0035217164549976587, + 0.00404942873865366, + 0.004685120191425085, + 0.004425518214702606, + 0.005088866222649813, + 0.004100810270756483, + 0.005916466936469078, + 0.0043578362092375755, + 0.008192208595573902, + 0.005340839736163616, + 0.0036583433393388987, + 0.009426477365195751, + 0.006731558125466108, + 0.0188352819532156, + 0.005991465877741575, + 0.008040989749133587, + 0.009327475912868977, + 0.005695481318980455, + 0.004328266717493534, + 0.00787130557000637, + 0.009164098650217056, + 0.014401314780116081, + 0.005319288000464439, + 0.0048530809581279755, + 0.005796270444989204, + 0.006102194543927908, + 0.0060217599384486675, + 0.005928671453148127, + 0.005873850081115961, + 0.00715230405330658, + 0.015902796760201454, + 0.009600081481039524, + 0.01264519989490509, + 0.006485035642981529, + 0.00419603893533349, + 0.008755901828408241, + 0.006640957202762365, + 0.010979111306369305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.0036434237845242023 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210343360900879 + ], + "opposite_logprob_tokens": [ + -4.264210224151611 + ], + "logprob_reward": -0.014210343360900879, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210343360900879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1903093159198761, + 0.08758428692817688, + 0.08916274458169937, + 0.03286251798272133, + 0.46456149220466614, + 0.026666704565286636, + 0.03039320930838585, + 0.02759583108127117, + 0.022329680621623993, + 0.028534190729260445 + ], + "credit_map": [ + 0.012719584628939629, + 0.023433556780219078, + 0.023192470893263817, + 0.030804427340626717, + 0.036620836704969406, + 0.01218043826520443, + 0.03522080183029175, + 0.01613720692694187, + 0.004234544467180967, + 0.013570665381848812, + 0.008392265066504478, + 0.011675970628857613, + 0.02621516026556492, + 0.005791603587567806, + 0.008328983560204506, + 0.009375090710818768, + 0.004355458077043295, + 0.014681229367852211, + 0.009643277153372765, + 0.01174812950193882, + 0.02010134793817997, + 0.0075859068892896175, + 0.00964339543133974, + 0.011404002085328102, + 0.002889067865908146, + 0.00405564671382308, + 0.0025849691592156887, + 0.0038219443522393703, + 0.004914780147373676, + 0.0047271233052015305, + 0.003700383473187685, + 0.006168604362756014, + 0.02039559930562973, + 0.02226642705500126, + 0.13926614820957184, + 0.028238147497177124, + 0.021200230345129967, + 0.18479767441749573, + 0.0168861523270607, + 0.031511154025793076, + 0.0022117989137768745, + 0.005463320296257734, + 0.0022011767141520977, + 0.002958998316898942, + 0.004228262696415186, + 0.003794135758653283, + 0.0036050723865628242, + 0.002203940413892269, + 0.0024665528908371925, + 0.0025096964091062546, + 0.0037539759650826454, + 0.0030990284867584705, + 0.0024600853212177753, + 0.00422421982511878, + 0.003398922737687826, + 0.008480727672576904, + 0.0025330749340355396, + 0.0030531256925314665, + 0.003846667939797044, + 0.0026980817783623934, + 0.0020462041720747948, + 0.0028784209862351418, + 0.0037158504128456116, + 0.006824403069913387, + 0.0028511553537100554, + 0.0019608403090387583, + 0.0035199210979044437, + 0.002217256696894765, + 0.0012960447929799557, + 0.00335508119314909, + 0.0024682090152055025, + 0.004661170300096273, + 0.007160512264817953, + 0.003865509293973446, + 0.004889499396085739, + 0.0020100504625588655, + 0.001707072020508349, + 0.0029854844324290752, + 0.0021487995982170105, + 0.0037672617472708225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.014210343360900879 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740807089954615 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.006740807089954615, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740807089954615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0474533848464489, + 0.10665213316679001, + 0.06006237864494324, + 0.060748521238565445, + 0.4858260154724121, + 0.060399916023015976, + 0.057063475251197815, + 0.07694463431835175, + 0.017584022134542465, + 0.027265600860118866 + ], + "credit_map": [ + 0.0030771400779485703, + 0.004474218934774399, + 0.00825933925807476, + 0.008963889442384243, + 0.004797612316906452, + 0.0023990159388631582, + 0.010617785155773163, + 0.004864382557570934, + 0.006084234453737736, + 0.019190741702914238, + 0.008698493242263794, + 0.010821188800036907, + 0.035701729357242584, + 0.008021621033549309, + 0.009212037548422813, + 0.008922084234654903, + 0.003260401776060462, + 0.008935177698731422, + 0.004149299114942551, + 0.005366398487240076, + 0.02066330425441265, + 0.006358399987220764, + 0.004308560863137245, + 0.007020832505077124, + 0.0038432402070611715, + 0.008786707185208797, + 0.005425805225968361, + 0.005941865500062704, + 0.007934192195534706, + 0.008597703650593758, + 0.004940935876220465, + 0.01527806930243969, + 0.021722547709941864, + 0.020324064418673515, + 0.1517767459154129, + 0.02788127399981022, + 0.021158400923013687, + 0.19653205573558807, + 0.015424252487719059, + 0.03100663051009178, + 0.005882099736481905, + 0.008605275303125381, + 0.006335638929158449, + 0.010219977237284184, + 0.006708919070661068, + 0.006492698565125465, + 0.008400451391935349, + 0.007754854392260313, + 0.004552902653813362, + 0.004864053335040808, + 0.006579270586371422, + 0.006383506115525961, + 0.0043338024988770485, + 0.007753415498882532, + 0.0073187933303415775, + 0.015277732163667679, + 0.006155981682240963, + 0.01087157242000103, + 0.009112083353102207, + 0.006979017984122038, + 0.0058499849401414394, + 0.009972576051950455, + 0.009725099429488182, + 0.018278319388628006, + 0.0016883903881534934, + 0.0019340646686032414, + 0.0027592438273131847, + 0.0017194895772263408, + 0.0011059364769607782, + 0.002738610375672579, + 0.0020359084010124207, + 0.0036023780703544617, + 0.005101755261421204, + 0.003072672290727496, + 0.004508966580033302, + 0.003108156379312277, + 0.0014808602863922715, + 0.0030737542547285557, + 0.002476393710821867, + 0.004443042911589146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.006740807089954615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546654790639877 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.012546654790639877, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546654790639877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02909521758556366, + 0.063753142952919, + 0.01603216864168644, + 0.01576910726726055, + 0.7559367418289185, + 0.010005021467804909, + 0.011883498169481754, + 0.06032398343086243, + 0.024003440514206886, + 0.013197674416005611 + ], + "credit_map": [ + 0.001698151114396751, + 0.003130403347313404, + 0.003673722967505455, + 0.006390651687979698, + 0.004009231459349394, + 0.0016254093497991562, + 0.005410399287939072, + 0.0031572480220347643, + 0.0036273151636123657, + 0.008022653870284557, + 0.009550346061587334, + 0.01186518557369709, + 0.013019814155995846, + 0.004060060251504183, + 0.008368273265659809, + 0.005239494144916534, + 0.0010012600105255842, + 0.002146058948710561, + 0.0018012125510722399, + 0.002123925369232893, + 0.0038283171597868204, + 0.0013724116142839193, + 0.0021355266217142344, + 0.0016234570648521185, + 0.0014740179758518934, + 0.0016786137130111456, + 0.0013458302710205317, + 0.001526011386886239, + 0.0017078681848943233, + 0.002807511715218425, + 0.0015406954335048795, + 0.0036885575391352177, + 0.03319118544459343, + 0.035026904195547104, + 0.23181095719337463, + 0.045327380299568176, + 0.03512059897184372, + 0.3005506694316864, + 0.026336541399359703, + 0.04857248067855835, + 0.0009816510137170553, + 0.0013834623387083411, + 0.001130505814217031, + 0.0015365408034995198, + 0.0012631163699552417, + 0.0010837754234671593, + 0.001401051296852529, + 0.001224918756633997, + 0.0011109652696177363, + 0.0009459468419663608, + 0.0016610773745924234, + 0.0015134650748223066, + 0.0006827650358900428, + 0.0013077881885692477, + 0.0016000565374270082, + 0.00306143332272768, + 0.0051408265717327595, + 0.008757436648011208, + 0.007176906336098909, + 0.0059575107879936695, + 0.005313399713486433, + 0.008041777648031712, + 0.007462236564606428, + 0.012473885901272297, + 0.0020439513027668, + 0.0029665548354387283, + 0.003483040723949671, + 0.0025607242714613676, + 0.0015883174492046237, + 0.0026969141326844692, + 0.003453274955973029, + 0.0052106622606515884, + 0.002887950511649251, + 0.0014922404661774635, + 0.002199186012148857, + 0.0010915232123807073, + 0.0008070422918535769, + 0.0015654981834813952, + 0.00118371588177979, + 0.0019705165177583694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.012546654790639877 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing close-up shot first focuses on a vast mural depicting an ancient agricultural scene. Its mottled, earthy-yellow surface is riddled with tiny, spiderweb-like cracks. In the dim, serene surroundings, a mysterious blue light begins to cautiously seep from deep within these cracks, emitting a soft yet energetic glow that illuminates the surrounding rough texture of the wall. The camera then shifts focus to an ox plowing a field at the center of the mural. Its body, rendered in thick mineral pigment, boasts powerful yet archaic muscular lines. Yet, one of its eyes possesses an unusually deep depth, as if embedded in black obsidian. As a streak of blue light reaches this eye, a subtle ripple suddenly ripples through the still pupil, rapidly expanding to become transparent yet visible ripples in space and time, like ripples on water. These ripples, with the pupils of the plowing ox as their epicenter, spread steadily and calmly across the entire painting. Wherever they reach, the ancient pigments on the mural, the cracks in the wall, and even the air itself are slightly distorted and refracted, as if the physical laws of the entire space are being gently rewritten, creating a mysterious, solemn, and surreal atmosphere full of science fiction wonder.", + "seed": 29740, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_8/index_4652_seed_29740_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.016597982682287692, + "vlm_logprob": -0.016597982682287692, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.016597982682287692, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02856154926121235 + ], + "opposite_logprob_tokens": [ + -3.778561592102051 + ], + "logprob_reward": -0.02856154926121235, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02856154926121235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17823772132396698, + 0.07629440724849701, + 0.04045714810490608, + 0.0717938020825386, + 0.1267288327217102, + 0.06501126289367676, + 0.07228638976812363, + 0.1456727534532547, + 0.12514463067054749, + 0.09837300330400467 + ], + "credit_map": [ + 0.02380385808646679, + 0.0289456769824028, + 0.014544048346579075, + 0.039195626974105835, + 0.02155241183936596, + 0.012106732465326786, + 0.028213515877723694, + 0.009875857271254063, + 0.009165049530565739, + 0.014522627927362919, + 0.01643594540655613, + 0.011438896879553795, + 0.006465217098593712, + 0.005474649835377932, + 0.006897558458149433, + 0.00589445885270834, + 0.005503252148628235, + 0.0037838483694940805, + 0.006054633762687445, + 0.0034157263580709696, + 0.004655677825212479, + 0.006298401392996311, + 0.003132615704089403, + 0.007612992078065872, + 0.009914377704262733, + 0.01676289550960064, + 0.008681797422468662, + 0.006010857410728931, + 0.009372951462864876, + 0.010177523829042912, + 0.006353299133479595, + 0.004520096350461245, + 0.012063778936862946, + 0.027435924857854843, + 0.013566931709647179, + 0.009304895997047424, + 0.03135761618614197, + 0.015673154965043068, + 0.009018855169415474, + 0.00830767210572958, + 0.007163803558796644, + 0.005813059397041798, + 0.008682162500917912, + 0.0066044810228049755, + 0.01166666392236948, + 0.008317682892084122, + 0.009252149611711502, + 0.0075112637132406235, + 0.009120947681367397, + 0.014586851000785828, + 0.0086579080671072, + 0.005248566623777151, + 0.00616419967263937, + 0.006987179629504681, + 0.016425326466560364, + 0.005095405038446188, + 0.01807103492319584, + 0.016753574833273888, + 0.030921371653676033, + 0.02243819087743759, + 0.012330257333815098, + 0.01825944148004055, + 0.011810812167823315, + 0.015088057145476341, + 0.014055087231099606, + 0.0220932699739933, + 0.02741830050945282, + 0.02737819403409958, + 0.010878313332796097, + 0.007286236621439457, + 0.008571690879762173, + 0.007463539019227028, + 0.009997819550335407, + 0.006889909505844116, + 0.02265794388949871, + 0.030016576871275902, + 0.00559975765645504, + 0.005343906115740538, + 0.008696186356246471, + 0.0091708954423666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.02856154926121235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005274665541946888 + ], + "opposite_logprob_tokens": [ + -5.255274772644043 + ], + "logprob_reward": -0.005274665541946888, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005274665541946888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2459351122379303, + 0.07797708362340927, + 0.051237162202596664, + 0.024761835113167763, + 0.07974464446306229, + 0.03743772208690643, + 0.05207080394029617, + 0.24000924825668335, + 0.08028514683246613, + 0.11054132133722305 + ], + "credit_map": [ + 0.023116856813430786, + 0.029214367270469666, + 0.018926402553915977, + 0.05097106471657753, + 0.026157384738326073, + 0.0316612645983696, + 0.038407474756240845, + 0.027480315417051315, + 0.012021736241877079, + 0.011783301830291748, + 0.009135865606367588, + 0.009298888966441154, + 0.007966023869812489, + 0.009303414262831211, + 0.010118559934198856, + 0.008349289186298847, + 0.009766113013029099, + 0.004159101285040379, + 0.006547923199832439, + 0.003493515308946371, + 0.005696061998605728, + 0.010635902173817158, + 0.00408058101311326, + 0.0068579670041799545, + 0.0033699863124638796, + 0.002938841935247183, + 0.0026447977870702744, + 0.003206044202670455, + 0.00324713415466249, + 0.0033270532730966806, + 0.00357105047442019, + 0.002456926740705967, + 0.010256416164338589, + 0.014383948408067226, + 0.00870499387383461, + 0.008496807888150215, + 0.011211257427930832, + 0.012651997618377209, + 0.007952077314257622, + 0.006087150424718857, + 0.004104220774024725, + 0.0033480850979685783, + 0.005476216319948435, + 0.0034964303486049175, + 0.006985926069319248, + 0.004937090445309877, + 0.005336510483175516, + 0.0037532427813857794, + 0.005563879385590553, + 0.008701433427631855, + 0.0054548196494579315, + 0.00481509231030941, + 0.005526301916688681, + 0.006244177930057049, + 0.0103022875264287, + 0.005462812725454569, + 0.020471341907978058, + 0.0226803719997406, + 0.052545275539159775, + 0.03538452088832855, + 0.0285283662378788, + 0.029966609552502632, + 0.025560393929481506, + 0.024872368201613426, + 0.005365942604839802, + 0.005093051586300135, + 0.025013087317347527, + 0.02310372143983841, + 0.005469988100230694, + 0.0035579318646341562, + 0.0042707668617367744, + 0.008410654030740261, + 0.010975359939038754, + 0.009272956289350986, + 0.023934103548526764, + 0.0294791366904974, + 0.00856027938425541, + 0.005499106366187334, + 0.009922314435243607, + 0.012898058630526066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.005274665541946888 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016077283769845963 + ], + "opposite_logprob_tokens": [ + -4.141077518463135 + ], + "logprob_reward": -0.016077283769845963, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016077283769845963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18521562218666077, + 0.05683429166674614, + 0.13860748708248138, + 0.06748688966035843, + 0.083038330078125, + 0.10209058225154877, + 0.07541891932487488, + 0.1160636767745018, + 0.09688370674848557, + 0.07836055010557175 + ], + "credit_map": [ + 0.020801421254873276, + 0.029571928083896637, + 0.015339188277721405, + 0.02942481078207493, + 0.027995096519589424, + 0.01651936210691929, + 0.030422482639551163, + 0.015141342766582966, + 0.010530786588788033, + 0.00647463696077466, + 0.007645574864000082, + 0.006984330713748932, + 0.006691720802336931, + 0.006349639967083931, + 0.0063744522631168365, + 0.005783150438219309, + 0.0277390256524086, + 0.010196646675467491, + 0.011488653719425201, + 0.011036977171897888, + 0.01483497116714716, + 0.031088897958397865, + 0.012704476714134216, + 0.019517846405506134, + 0.00602914160117507, + 0.01151110790669918, + 0.006800684612244368, + 0.007329602260142565, + 0.010241623967885971, + 0.00999096967279911, + 0.008133742026984692, + 0.0074500152841210365, + 0.01205361820757389, + 0.01605512760579586, + 0.007683359552174807, + 0.008133894763886929, + 0.011665348894894123, + 0.014317464083433151, + 0.007024127058684826, + 0.0061053866520524025, + 0.010158720426261425, + 0.01149766519665718, + 0.014685817062854767, + 0.009296232834458351, + 0.02057749405503273, + 0.013139627873897552, + 0.013086715713143349, + 0.009648315608501434, + 0.0077756838873028755, + 0.012449022382497787, + 0.008089620620012283, + 0.00581565173342824, + 0.006467737257480621, + 0.008365596644580364, + 0.019220011308789253, + 0.007235599681735039, + 0.009005144238471985, + 0.0064759040251374245, + 0.026270920410752296, + 0.016778165474534035, + 0.015344819985330105, + 0.014980931766331196, + 0.01416828390210867, + 0.013039504177868366, + 0.008680611848831177, + 0.007828431203961372, + 0.02256832644343376, + 0.022304194048047066, + 0.010422355495393276, + 0.00794090423732996, + 0.008947483263909817, + 0.008191406726837158, + 0.0066814846359193325, + 0.007159726694226265, + 0.008872073143720627, + 0.013559888117015362, + 0.005725471768528223, + 0.00773479463532567, + 0.009260213002562523, + 0.019366897642612457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.016077283769845963 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786011651158333 + ], + "opposite_logprob_tokens": [ + -4.634786128997803 + ], + "logprob_reward": -0.009786011651158333, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786011651158333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08036431670188904, + 0.06853166967630386, + 0.03234830126166344, + 0.0467558428645134, + 0.12669815123081207, + 0.05284050479531288, + 0.03375359624624252, + 0.1024092361330986, + 0.3403618633747101, + 0.11593656986951828 + ], + "credit_map": [ + 0.00911215040832758, + 0.00948445126414299, + 0.006535005290061235, + 0.01764468476176262, + 0.007773433346301317, + 0.009292161092162132, + 0.013293884694576263, + 0.007228546775877476, + 0.00900841411203146, + 0.010639938525855541, + 0.014653103426098824, + 0.009880226105451584, + 0.005347900558263063, + 0.006170473527163267, + 0.006582082714885473, + 0.006249533500522375, + 0.007632965687662363, + 0.002657366916537285, + 0.003005585866048932, + 0.0029103721026331186, + 0.0040115294978022575, + 0.007700217887759209, + 0.002137452596798539, + 0.0022928137332201004, + 0.006891964469105005, + 0.007495590951293707, + 0.004358821548521519, + 0.0044071669690310955, + 0.007310510613024235, + 0.00625050812959671, + 0.005619599483907223, + 0.004421683494001627, + 0.015087592415511608, + 0.023988619446754456, + 0.016883356496691704, + 0.011587527580559254, + 0.022049609571695328, + 0.01669309474527836, + 0.010753925889730453, + 0.009654426015913486, + 0.005530188325792551, + 0.0049924785271286964, + 0.00656238803640008, + 0.005136969964951277, + 0.010383372195065022, + 0.006979155819863081, + 0.006322405766695738, + 0.006933542899787426, + 0.0036735034082084894, + 0.004842648282647133, + 0.006751579232513905, + 0.0035075421910732985, + 0.0036596355494111776, + 0.004848106764256954, + 0.0035080285742878914, + 0.002962552709504962, + 0.009985823184251785, + 0.010381310246884823, + 0.01579023338854313, + 0.012387407943606377, + 0.008760581724345684, + 0.013135564513504505, + 0.016013450920581818, + 0.015954861417412758, + 0.015590368770062923, + 0.020542722195386887, + 0.11961501091718674, + 0.11941730231046677, + 0.020587556064128876, + 0.013182013295590878, + 0.018556302413344383, + 0.012870585545897484, + 0.010568853467702866, + 0.00880417414009571, + 0.02472761645913124, + 0.03240154683589935, + 0.009416979737579823, + 0.006281607784330845, + 0.013208363205194473, + 0.010527427308261395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.009786011651158333 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023290403187274933 + ], + "opposite_logprob_tokens": [ + -3.7732903957366943 + ], + "logprob_reward": -0.023290403187274933, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290403187274933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13984207808971405, + 0.05673063173890114, + 0.0484844371676445, + 0.02265271171927452, + 0.08427108079195023, + 0.04140441119670868, + 0.0853317379951477, + 0.047004684805870056, + 0.3805224299430847, + 0.09375585615634918 + ], + "credit_map": [ + 0.015672801062464714, + 0.02449997141957283, + 0.011809547431766987, + 0.024729695171117783, + 0.019628893584012985, + 0.013141690753400326, + 0.02074086107313633, + 0.009618612937629223, + 0.006730119232088327, + 0.006758185103535652, + 0.007664116565138102, + 0.008054994978010654, + 0.006396552547812462, + 0.006118655204772949, + 0.008427195250988007, + 0.006580807268619537, + 0.006508121266961098, + 0.004818562418222427, + 0.006099720951169729, + 0.00387645629234612, + 0.006274521350860596, + 0.006462216377258301, + 0.004695574287325144, + 0.009749263525009155, + 0.0023443116806447506, + 0.0033818320371210575, + 0.0018121561734005809, + 0.0017699719173833728, + 0.003587330924347043, + 0.003946361597627401, + 0.002628675429150462, + 0.0031820714939385653, + 0.009896948002278805, + 0.019230956211686134, + 0.009336065500974655, + 0.007488722912967205, + 0.016268892213702202, + 0.009785929694771767, + 0.006499916315078735, + 0.005763641092926264, + 0.004177947994321585, + 0.004198330920189619, + 0.004970861598849297, + 0.004593883641064167, + 0.0073373871855437756, + 0.005085301585495472, + 0.005646375007927418, + 0.0053943186067044735, + 0.010899067856371403, + 0.01805095747113228, + 0.008697399869561195, + 0.005276291631162167, + 0.0041185952723026276, + 0.008657924830913544, + 0.021791327744722366, + 0.007840169593691826, + 0.006521889008581638, + 0.004595829639583826, + 0.006663728971034288, + 0.005273406393826008, + 0.006165627855807543, + 0.005968261044472456, + 0.005591383669525385, + 0.0062245577573776245, + 0.020130671560764313, + 0.02550131268799305, + 0.1353110671043396, + 0.13844631612300873, + 0.018150774762034416, + 0.010326928459107876, + 0.01612723246216774, + 0.016528088599443436, + 0.006859792862087488, + 0.007729455828666687, + 0.016373200342059135, + 0.02395673468708992, + 0.008079411461949348, + 0.005279602017253637, + 0.012698040343821049, + 0.012779615819454193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.023290403187274933 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing close-up shot first focuses on a vast mural depicting an ancient agricultural scene. Its mottled, earthy-yellow surface is riddled with tiny, spiderweb-like cracks. In the dim, serene surroundings, a mysterious blue light begins to cautiously seep from deep within these cracks, emitting a soft yet energetic glow that illuminates the surrounding rough texture of the wall. The camera then shifts focus to an ox plowing a field at the center of the mural. Its body, rendered in thick mineral pigment, boasts powerful yet archaic muscular lines. Yet, one of its eyes possesses an unusually deep depth, as if embedded in black obsidian. As a streak of blue light reaches this eye, a subtle ripple suddenly ripples through the still pupil, rapidly expanding to become transparent yet visible ripples in space and time, like ripples on water. These ripples, with the pupils of the plowing ox as their epicenter, spread steadily and calmly across the entire painting. Wherever they reach, the ancient pigments on the mural, the cracks in the wall, and even the air itself are slightly distorted and refracted, as if the physical laws of the entire space are being gently rewritten, creating a mysterious, solemn, and surreal atmosphere full of science fiction wonder.", + "seed": 29741, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_8/index_4652_seed_29741_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05412786789238453, + "vlm_logprob": -0.05412786789238453, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.05412786789238453, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19920073449611664 + ], + "opposite_logprob_tokens": [ + -1.8242007493972778 + ], + "logprob_reward": -0.19920073449611664, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.19920073449611664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05675044283270836, + 0.06564843654632568, + 0.12974482774734497, + 0.07983604818582535, + 0.11300303786993027, + 0.13646677136421204, + 0.21866904199123383, + 0.11442723125219345, + 0.050137266516685486, + 0.03531688451766968 + ], + "credit_map": [ + 0.004165226127952337, + 0.006769859232008457, + 0.00661309203132987, + 0.006306538358330727, + 0.007940076291561127, + 0.006388955749571323, + 0.007940198294818401, + 0.010626491159200668, + 0.004981104750186205, + 0.007557776290923357, + 0.0062371110543608665, + 0.006583712995052338, + 0.006964901462197304, + 0.009519134648144245, + 0.01332253310829401, + 0.010482162237167358, + 0.007078743539750576, + 0.006020416971296072, + 0.007500714622437954, + 0.007810673676431179, + 0.011584117077291012, + 0.04514811933040619, + 0.010696396231651306, + 0.033905643969774246, + 0.008371973410248756, + 0.008671118877828121, + 0.00948341004550457, + 0.006722620688378811, + 0.010769971646368504, + 0.011196792125701904, + 0.009046437218785286, + 0.015573716722428799, + 0.011647703126072884, + 0.013834900222718716, + 0.014419075101613998, + 0.012998358346521854, + 0.022527161985635757, + 0.011038406752049923, + 0.012054060585796833, + 0.014483359642326832, + 0.02090747281908989, + 0.015689248219132423, + 0.01173251960426569, + 0.010897437110543251, + 0.02131551317870617, + 0.020813308656215668, + 0.016537606716156006, + 0.01857364922761917, + 0.01754782348871231, + 0.01733964867889881, + 0.022860486060380936, + 0.020710214972496033, + 0.05938011780381203, + 0.022769371047616005, + 0.022609738633036613, + 0.035451631993055344, + 0.01258198544383049, + 0.010444987565279007, + 0.015776848420500755, + 0.012652477249503136, + 0.01800474338233471, + 0.012356924824416637, + 0.015951670706272125, + 0.016657590866088867, + 0.006646245252341032, + 0.005222403910011053, + 0.006073030177503824, + 0.007222839631140232, + 0.0059417798183858395, + 0.0064972518011927605, + 0.006778478156775236, + 0.005755236838012934, + 0.0026763000059872866, + 0.0021070437505841255, + 0.0019763903692364693, + 0.0026636598631739616, + 0.004598772153258324, + 0.010647051967680454, + 0.0026887457352131605, + 0.007958917878568172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.19920073449611664 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006785919889807701 + ], + "opposite_logprob_tokens": [ + -5.006785869598389 + ], + "logprob_reward": -0.006785919889807701, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006785919889807701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17330516874790192, + 0.0957123413681984, + 0.19958776235580444, + 0.11079662293195724, + 0.05949859693646431, + 0.045678142458200455, + 0.12528680264949799, + 0.10406819730997086, + 0.04634585604071617, + 0.03972051292657852 + ], + "credit_map": [ + 0.013625838793814182, + 0.02107871323823929, + 0.025847284123301506, + 0.020660903304815292, + 0.027422521263360977, + 0.015696987509727478, + 0.024285590276122093, + 0.024687329307198524, + 0.009623263962566853, + 0.012894443236291409, + 0.011386110447347164, + 0.010744111612439156, + 0.009147752076387405, + 0.017187193036079407, + 0.013322237879037857, + 0.011407231912016869, + 0.010113252326846123, + 0.017165277153253555, + 0.01155656948685646, + 0.01683216728270054, + 0.022312887012958527, + 0.05553366243839264, + 0.017981456592679024, + 0.04809247702360153, + 0.011864847503602505, + 0.008729248307645321, + 0.010340172797441483, + 0.016157643869519234, + 0.012128300033509731, + 0.013440918177366257, + 0.014725578017532825, + 0.023409904912114143, + 0.009953449480235577, + 0.006035920698195696, + 0.007605795748531818, + 0.006413938477635384, + 0.006975626107305288, + 0.006982825230807066, + 0.007418835535645485, + 0.008112210780382156, + 0.00679745152592659, + 0.005105327349156141, + 0.00403077807277441, + 0.0035071808379143476, + 0.005403041373938322, + 0.008394667878746986, + 0.005729059688746929, + 0.006710638292133808, + 0.008440460078418255, + 0.009230734780430794, + 0.01218574307858944, + 0.010733379051089287, + 0.03639562427997589, + 0.013300560414791107, + 0.013854254968464375, + 0.02114604413509369, + 0.009723308496177197, + 0.009003518149256706, + 0.011389569379389286, + 0.010037747211754322, + 0.019615909084677696, + 0.011273577809333801, + 0.01901007629930973, + 0.014014492742717266, + 0.005890249274671078, + 0.004754276480525732, + 0.004317941144108772, + 0.007606890983879566, + 0.00722515257075429, + 0.004700222983956337, + 0.006482964381575584, + 0.005368159618228674, + 0.004228200763463974, + 0.005357741378247738, + 0.0038515354972332716, + 0.004493761342018843, + 0.0066469525918364525, + 0.005258262623101473, + 0.005067650228738785, + 0.0048164078034460545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.006785919889807701 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371734917163849 + ], + "opposite_logprob_tokens": [ + -3.408717393875122 + ], + "logprob_reward": -0.03371734917163849, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371734917163849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16590799391269684, + 0.1071801409125328, + 0.19819194078445435, + 0.07684773206710815, + 0.05697875842452049, + 0.04996432736515999, + 0.08102839440107346, + 0.11963289976119995, + 0.09040828049182892, + 0.053859591484069824 + ], + "credit_map": [ + 0.014991319738328457, + 0.019260497763752937, + 0.03115083836019039, + 0.021729040890932083, + 0.023716188967227936, + 0.013373631983995438, + 0.022882435470819473, + 0.0188040379434824, + 0.014229540713131428, + 0.01218018401414156, + 0.013666256330907345, + 0.010993462055921555, + 0.010205598548054695, + 0.012643467634916306, + 0.01848236285150051, + 0.014779270626604557, + 0.009248694404959679, + 0.011719734407961369, + 0.013017057441174984, + 0.012376786209642887, + 0.016125617548823357, + 0.06302333623170853, + 0.01761866919696331, + 0.05506204068660736, + 0.01023211982101202, + 0.005913811735808849, + 0.007446801755577326, + 0.009833317250013351, + 0.010320867411792278, + 0.010124081745743752, + 0.009320518001914024, + 0.013656215742230415, + 0.007113580126315355, + 0.005406624637544155, + 0.0073669180274009705, + 0.005176166072487831, + 0.00786599051207304, + 0.0071751391515135765, + 0.006751658860594034, + 0.010122682899236679, + 0.006107926834374666, + 0.007091091945767403, + 0.005397371016442776, + 0.004468654748052359, + 0.007932550273835659, + 0.006487550213932991, + 0.006790006533265114, + 0.005689176265150309, + 0.00798311922699213, + 0.006608100607991219, + 0.012561730109155178, + 0.006674651522189379, + 0.01582314632833004, + 0.009010295383632183, + 0.0133378766477108, + 0.009029475972056389, + 0.012340768240392208, + 0.011213477700948715, + 0.01707742176949978, + 0.01254584826529026, + 0.022900938987731934, + 0.012889346107840538, + 0.017113665118813515, + 0.013551440089941025, + 0.007933506742119789, + 0.0069257961586117744, + 0.010599112138152122, + 0.010268999263644218, + 0.014541243202984333, + 0.009264860302209854, + 0.019758911803364754, + 0.01111584436148405, + 0.006720295641571283, + 0.006141255609691143, + 0.00636493694037199, + 0.005643341224640608, + 0.007624342571943998, + 0.007920323871076107, + 0.006857116706669331, + 0.0065879784524440765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.03371734917163849 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007638176903128624 + ], + "opposite_logprob_tokens": [ + -4.882637977600098 + ], + "logprob_reward": -0.007638176903128624, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007638176903128624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07831256091594696, + 0.053355272859334946, + 0.25395092368125916, + 0.07698700577020645, + 0.0807458758354187, + 0.07727295905351639, + 0.23488350212574005, + 0.05923454836010933, + 0.03786623850464821, + 0.047391168773174286 + ], + "credit_map": [ + 0.00678439112380147, + 0.009582874365150928, + 0.008990845642983913, + 0.00952956359833479, + 0.007825414650142193, + 0.011142264120280743, + 0.010889093391597271, + 0.01356810424476862, + 0.005888146348297596, + 0.007311223540455103, + 0.006983471103012562, + 0.005767172668129206, + 0.00507736811414361, + 0.008954633958637714, + 0.00671013630926609, + 0.006663115695118904, + 0.00993012823164463, + 0.01671917922794819, + 0.0114683723077178, + 0.016030222177505493, + 0.02374054491519928, + 0.08342935144901276, + 0.022336425259709358, + 0.07029667496681213, + 0.0070303031243383884, + 0.006388719193637371, + 0.007676970679312944, + 0.007986047305166721, + 0.00794907659292221, + 0.012724969536066055, + 0.007915272377431393, + 0.01931564137339592, + 0.008622533641755581, + 0.008636231534183025, + 0.01258394867181778, + 0.010154427029192448, + 0.013590377755463123, + 0.006670753471553326, + 0.009348886087536812, + 0.011138712987303734, + 0.011364556849002838, + 0.0084166768938303, + 0.0075361537747085094, + 0.006375148892402649, + 0.010794565081596375, + 0.012636924162507057, + 0.010074559599161148, + 0.010074364952743053, + 0.017682205885648727, + 0.016656965017318726, + 0.018468787893652916, + 0.02120649442076683, + 0.0728381797671318, + 0.02420138567686081, + 0.02569468878209591, + 0.038134776055812836, + 0.007750700227916241, + 0.005054452922195196, + 0.008659162558615208, + 0.005864752922207117, + 0.008028286509215832, + 0.005712062586098909, + 0.01155094988644123, + 0.006614176090806723, + 0.004628022667020559, + 0.003269221866503358, + 0.004123014397919178, + 0.004362759180366993, + 0.004016977269202471, + 0.004945134278386831, + 0.008281179703772068, + 0.004239922389388084, + 0.005055665038526058, + 0.004313236102461815, + 0.0036505903117358685, + 0.003885620506480336, + 0.005551714915782213, + 0.01216608565300703, + 0.004386946093291044, + 0.008381303399801254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 4 + ], + "selected_reward": -0.007638176903128624 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a vast mural depicting an ancient agricultural scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the mural surface appear mottled, earthy-yellow, and riddled with tiny spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blue light seep from the cracks and illuminate the rough texture of the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ox plowing a field visible at the center of the mural with thick mineral pigment and an eye resembling black obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do ripples expand from the ox's eye, distorting the pigments, cracks, and air as if rewriting physical laws?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023297159001231194 + ], + "opposite_logprob_tokens": [ + -3.773297071456909 + ], + "logprob_reward": -0.023297159001231194, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297159001231194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1382567137479782, + 0.061541322618722916, + 0.1205853521823883, + 0.10944639146327972, + 0.0740806832909584, + 0.08020193874835968, + 0.07822699099779129, + 0.11742045730352402, + 0.10125351697206497, + 0.11898666620254517 + ], + "credit_map": [ + 0.010164947248995304, + 0.015200001187622547, + 0.017650378867983818, + 0.015885518863797188, + 0.018010500818490982, + 0.018902193754911423, + 0.024335753172636032, + 0.01810741052031517, + 0.007113963831216097, + 0.006628262810409069, + 0.008109045214951038, + 0.006596811581403017, + 0.007194838020950556, + 0.009392248466610909, + 0.0069942781701684, + 0.009511872194707394, + 0.009887016378343105, + 0.017797313630580902, + 0.012933033518493176, + 0.014570903033018112, + 0.011666028760373592, + 0.02453661523759365, + 0.011091446503996849, + 0.018102990463376045, + 0.014101983979344368, + 0.011002752929925919, + 0.011689431965351105, + 0.019691508263349533, + 0.012480427511036396, + 0.010544417425990105, + 0.013533085584640503, + 0.01640278846025467, + 0.008500325493514538, + 0.007922917604446411, + 0.009198404848575592, + 0.008853564038872719, + 0.011938406154513359, + 0.00864789355546236, + 0.008206183090806007, + 0.010812994092702866, + 0.00753488065674901, + 0.00487121194601059, + 0.006750477943569422, + 0.003616360481828451, + 0.005935065448284149, + 0.02177533321082592, + 0.007011375389993191, + 0.02270723134279251, + 0.008922627195715904, + 0.00765051506459713, + 0.013568556867539883, + 0.007772083394229412, + 0.010139776393771172, + 0.008926105685532093, + 0.01224052906036377, + 0.009006790816783905, + 0.013541988097131252, + 0.01036798395216465, + 0.0182819627225399, + 0.011277545243501663, + 0.01720139943063259, + 0.014720698818564415, + 0.0184563547372818, + 0.01357252337038517, + 0.011771125718951225, + 0.006597512401640415, + 0.009151073172688484, + 0.011231397278606892, + 0.0228169783949852, + 0.010888163931667805, + 0.015218226239085197, + 0.013579034246504307, + 0.012206426821649075, + 0.01353455614298582, + 0.009171840734779835, + 0.010303332470357418, + 0.021771492436528206, + 0.02204013802111149, + 0.011809210292994976, + 0.018149666488170624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.023297159001231194 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_9/index_tensor([4772, 4772])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_9/index_tensor([4772, 4772])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6c83ac6fadbea0649ccc143b2be46f1fd7d4d648 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_9/index_tensor([4772, 4772])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a level angle, slowly pulls back to reveal a grand, dynamic scene. At the center, the historical figure Zhuge Liang stands calmly on a mountain plain shrouded in mist. The scene embodies the exquisite aesthetic of traditional Chinese ink landscape painting. He wears a flowing white silk Hanfu robe with wide sleeves, topped by a dark gray double-breasted jacket. The sleeves sway gently with his slow movements. His face is gaunt, with a neatly trimmed black beard. His eyes are half-closed, his expression focused and serene. His left hand calmly grasps a feather fan made of pure white feathers, its handle made of dark wood, a subtle sheen. He slowly practices Tai Chi, his movements flowing with the grace and fluidity of flowing water, transitioning gently from one pose to the next, his wide sleeves drawing soft arcs through the air as his body turns. The background depicts a classic ink painting: distant mountains, layered and serene, are outlined in varying shades of ink, shrouded in wisps of mist. Nearby, a few ancient pines, rendered with vigorous brushstrokes, can be seen. A subtle soft focus effect is employed throughout, blurring the edges and enhancing the scene's dreamlike and poetic quality. The lighting, rather than originating from a specific source, is a uniform, soft, diffused light that seems to permeate the rice paper itself. There are no distinct shadows, only the rich gradations of ink color in the folds of the figures' clothing and the mountains' contours, creating a philosophical atmosphere of transcendence, tranquility, and profound tranquility.", + "seed": 27723, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_9/index_4772_seed_27723_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04200586853548884, + "vlm_logprob": -0.04200586853548884, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.04200586853548884, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06881046295166016 + ], + "opposite_logprob_tokens": [ + -3.94381046295166 + ], + "logprob_reward": -0.06881046295166016, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.06881046295166016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34571564197540283, + 0.08494793623685837, + 0.13104532659053802, + 0.028145065531134605, + 0.0924304723739624, + 0.042984865605831146, + 0.04808861389756203, + 0.03351401537656784, + 0.04488633945584297, + 0.14824168384075165 + ], + "credit_map": [ + 0.042761605232954025, + 0.02693287841975689, + 0.05935944244265556, + 0.03729270026087761, + 0.02469191886484623, + 0.042469438165426254, + 0.04004530608654022, + 0.07216236740350723, + 0.011202510446310043, + 0.005809977184981108, + 0.010159811936318874, + 0.01124698854982853, + 0.006068943999707699, + 0.009829060174524784, + 0.016239013522863388, + 0.014391637407243252, + 0.013743323273956776, + 0.015672722831368446, + 0.01734405755996704, + 0.012706379406154156, + 0.015915673226118088, + 0.016099659726023674, + 0.01671559363603592, + 0.02284792810678482, + 0.003432060591876507, + 0.001762842177413404, + 0.003924194723367691, + 0.0035902969539165497, + 0.0020104357972741127, + 0.0038966182619333267, + 0.0029407544061541557, + 0.006587866693735123, + 0.012649992480874062, + 0.00922483392059803, + 0.00997687503695488, + 0.012818874791264534, + 0.011765610426664352, + 0.011573499999940395, + 0.010787955485284328, + 0.013632838614284992, + 0.004480807110667229, + 0.0047609275206923485, + 0.0049414304085075855, + 0.005880686454474926, + 0.007130817975848913, + 0.005358646158128977, + 0.005553344264626503, + 0.00487820990383625, + 0.006621272768825293, + 0.004426416475325823, + 0.0066429562866687775, + 0.004025809001177549, + 0.003665199736133218, + 0.007760406471788883, + 0.005021166522055864, + 0.00992539245635271, + 0.0036691816058009863, + 0.003445395966991782, + 0.004357951693236828, + 0.0038124790880829096, + 0.003782049287110567, + 0.004936004988849163, + 0.005705955438315868, + 0.0038050003349781036, + 0.004826960153877735, + 0.004866838455200195, + 0.00679858960211277, + 0.004164984915405512, + 0.005099190399050713, + 0.005763453897088766, + 0.008107714354991913, + 0.005258612334728241, + 0.014200078323483467, + 0.03513680770993233, + 0.016643501818180084, + 0.011483650654554367, + 0.007640152703970671, + 0.03492560610175133, + 0.008143037557601929, + 0.020068857818841934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.06881046295166016 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811738267540932 + ], + "opposite_logprob_tokens": [ + -3.2881174087524414 + ], + "logprob_reward": -0.03811738267540932, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811738267540932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12681713700294495, + 0.09010209143161774, + 0.19870539009571075, + 0.03366643190383911, + 0.26856473088264465, + 0.09040911495685577, + 0.05595659092068672, + 0.05732963979244232, + 0.028727160766720772, + 0.049721699208021164 + ], + "credit_map": [ + 0.014949453063309193, + 0.010081452317535877, + 0.029285524040460587, + 0.015145432204008102, + 0.009045949205756187, + 0.015547228045761585, + 0.01531229168176651, + 0.01744980923831463, + 0.011809810064733028, + 0.006811611820012331, + 0.009824976325035095, + 0.010160902515053749, + 0.006967833265662193, + 0.01137944683432579, + 0.010866054333746433, + 0.022281454876065254, + 0.022651351988315582, + 0.025576021522283554, + 0.02486983686685562, + 0.02263912558555603, + 0.022264692932367325, + 0.022842463105916977, + 0.022251654416322708, + 0.03561024367809296, + 0.003959125839173794, + 0.0031670755706727505, + 0.004975930787622929, + 0.0038802672643214464, + 0.0036134980618953705, + 0.004131785128265619, + 0.004978700540959835, + 0.0049600498750805855, + 0.03506594896316528, + 0.028298676013946533, + 0.043963514268398285, + 0.025522442534565926, + 0.0390242375433445, + 0.03011511266231537, + 0.027842095121741295, + 0.038732703775167465, + 0.00973483081907034, + 0.012538962997496128, + 0.009690378792583942, + 0.010413940995931625, + 0.01289707887917757, + 0.010934578254818916, + 0.012570170685648918, + 0.011629165150225163, + 0.0084114084020257, + 0.005963098723441362, + 0.006033582147210836, + 0.004229811951518059, + 0.0051778461784124374, + 0.009079644456505775, + 0.006710133980959654, + 0.010351069271564484, + 0.007625066675245762, + 0.007537395227700472, + 0.008085421286523342, + 0.003839566372334957, + 0.00600748835131526, + 0.013058222830295563, + 0.0066556790843605995, + 0.0045208013616502285, + 0.003722607158124447, + 0.002787591889500618, + 0.005066924728453159, + 0.00316157890483737, + 0.0023382161743938923, + 0.0035519336815923452, + 0.004661737009882927, + 0.003436570754274726, + 0.009566464461386204, + 0.016724316403269768, + 0.0032898872159421444, + 0.005095344502478838, + 0.003805317683145404, + 0.004820419941097498, + 0.0036145749036222696, + 0.0028053773567080498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.03811738267540932 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026357457041740417 + ], + "opposite_logprob_tokens": [ + -3.651357412338257 + ], + "logprob_reward": -0.026357457041740417, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026357457041740417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35530608892440796, + 0.054098568856716156, + 0.2713088393211365, + 0.02347283437848091, + 0.13527487218379974, + 0.026247071102261543, + 0.02112952433526516, + 0.03390318900346756, + 0.04238888993859291, + 0.03687018156051636 + ], + "credit_map": [ + 0.041089750826358795, + 0.02852615714073181, + 0.0666690319776535, + 0.042167145758867264, + 0.021532034501433372, + 0.04616565257310867, + 0.043851446360349655, + 0.06530486047267914, + 0.00592435197904706, + 0.0038526561111211777, + 0.012605120427906513, + 0.007748167961835861, + 0.00258831400424242, + 0.006223354488611221, + 0.006121213082224131, + 0.009035383351147175, + 0.029437387362122536, + 0.03592152148485184, + 0.03537222370505333, + 0.02696475386619568, + 0.031137194484472275, + 0.03171233832836151, + 0.02847359888255596, + 0.05228981003165245, + 0.0025753916706889868, + 0.0021355950739234686, + 0.0027581164613366127, + 0.003282216377556324, + 0.0025715609081089497, + 0.0028149220161139965, + 0.0038949986919760704, + 0.0034400303848087788, + 0.017685333266854286, + 0.014563986100256443, + 0.021967926993966103, + 0.013003063388168812, + 0.020241543650627136, + 0.014951296150684357, + 0.013943781144917011, + 0.018917925655841827, + 0.0030234293080866337, + 0.0030308302957564592, + 0.0037468154914677143, + 0.0032023689709603786, + 0.0035236976109445095, + 0.0031065901275724173, + 0.0026393672451376915, + 0.003973970655351877, + 0.0034995286259800196, + 0.0020790372509509325, + 0.002742371754720807, + 0.0016694579971954226, + 0.0019756879191845655, + 0.0025987476110458374, + 0.0029046223498880863, + 0.003660069312900305, + 0.0060108317993581295, + 0.0034908270463347435, + 0.0031412590760737658, + 0.0030123319011181593, + 0.0025379126891493797, + 0.007797956932336092, + 0.004357222933322191, + 0.003554843133315444, + 0.0042183175683021545, + 0.007071445230394602, + 0.006403742823749781, + 0.0032469802536070347, + 0.005160229280591011, + 0.006908469367772341, + 0.0059385523200035095, + 0.0034411523956805468, + 0.00840552244335413, + 0.004865646827965975, + 0.0030449917539954185, + 0.0025602744426578283, + 0.0024616720620542765, + 0.010249624028801918, + 0.002672333735972643, + 0.002610114635899663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.026357457041740417 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001081109046936 + ], + "opposite_logprob_tokens": [ + -2.6950109004974365 + ], + "logprob_reward": -0.07001081109046936, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001081109046936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1954660266637802, + 0.13835161924362183, + 0.18915604054927826, + 0.056116119027137756, + 0.18399588763713837, + 0.058960385620594025, + 0.06583771854639053, + 0.027191162109375, + 0.03266707807779312, + 0.052257947623729706 + ], + "credit_map": [ + 0.026036612689495087, + 0.0138472905382514, + 0.04800644889473915, + 0.02251623198390007, + 0.010759633034467697, + 0.02475094050168991, + 0.01825571618974209, + 0.031293150037527084, + 0.017523102462291718, + 0.008693667128682137, + 0.014069672673940659, + 0.014923418872058392, + 0.010634787380695343, + 0.01735522225499153, + 0.029551753774285316, + 0.025600001215934753, + 0.02009817212820053, + 0.024633124470710754, + 0.025371652096509933, + 0.019208338111639023, + 0.022114291787147522, + 0.022179508581757545, + 0.020020004361867905, + 0.0355309322476387, + 0.004739327821880579, + 0.003528305096551776, + 0.009404942393302917, + 0.00942745991051197, + 0.004641118925064802, + 0.006757868453860283, + 0.008034951984882355, + 0.009582143276929855, + 0.024240029975771904, + 0.019267991185188293, + 0.02636004611849785, + 0.018684064969420433, + 0.027560517191886902, + 0.02062002755701542, + 0.019911743700504303, + 0.027351470664143562, + 0.006493282038718462, + 0.006005318369716406, + 0.009887057356536388, + 0.006099727936089039, + 0.006200456991791725, + 0.007014109753072262, + 0.007293262053281069, + 0.009967165067791939, + 0.009161540307104588, + 0.007091217208653688, + 0.007370051462203264, + 0.005612954031676054, + 0.00541813625022769, + 0.007755665574222803, + 0.010396172292530537, + 0.013031980954110622, + 0.004236117471009493, + 0.0023513745982199907, + 0.0035204743035137653, + 0.002429127926006913, + 0.0022118284832686186, + 0.005340757314115763, + 0.0039249928668141365, + 0.0031764893792569637, + 0.004384532105177641, + 0.0030945874750614166, + 0.0038141224067658186, + 0.0033842355478554964, + 0.003358087269589305, + 0.004717547912150621, + 0.005991396028548479, + 0.003922570031136274, + 0.011283003725111485, + 0.004418194759637117, + 0.003289789194241166, + 0.003044353798031807, + 0.004072715528309345, + 0.01890304498374462, + 0.003870446002110839, + 0.0033764031250029802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.07001081109046936 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067332289181649685 + ], + "opposite_logprob_tokens": [ + -5.006733417510986 + ], + "logprob_reward": -0.0067332289181649685, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067332289181649685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16610412299633026, + 0.27893686294555664, + 0.038306355476379395, + 0.06839516758918762, + 0.18037888407707214, + 0.10865229368209839, + 0.023143846541643143, + 0.02854006551206112, + 0.02233060635626316, + 0.0852118581533432 + ], + "credit_map": [ + 0.020965704694390297, + 0.018424151465296745, + 0.029012955725193024, + 0.019329747185111046, + 0.010065963491797447, + 0.02208753675222397, + 0.014785178937017918, + 0.031432878226041794, + 0.037577349692583084, + 0.01647000014781952, + 0.0464327447116375, + 0.03097294270992279, + 0.017089925706386566, + 0.03266909345984459, + 0.042038097977638245, + 0.055686671286821365, + 0.004118411801755428, + 0.00458580581471324, + 0.00469192024320364, + 0.004782326519489288, + 0.00396383972838521, + 0.004675423260778189, + 0.004848363809287548, + 0.006640261970460415, + 0.007404664531350136, + 0.0038589336909353733, + 0.01034052949398756, + 0.009640629403293133, + 0.00505080446600914, + 0.00920073315501213, + 0.008425327949225903, + 0.014473535120487213, + 0.025298940017819405, + 0.018573513254523277, + 0.02785930410027504, + 0.019521303474903107, + 0.021233422681689262, + 0.02177826128900051, + 0.02122095599770546, + 0.02489316463470459, + 0.014586029574275017, + 0.011967233382165432, + 0.02136075124144554, + 0.011223581619560719, + 0.008711465634405613, + 0.012399873696267605, + 0.008408321999013424, + 0.019995035603642464, + 0.0036944495514035225, + 0.0025604618713259697, + 0.002174752065911889, + 0.003013049252331257, + 0.002419269410893321, + 0.0033918910194188356, + 0.003128345822915435, + 0.002761624753475189, + 0.005114749539643526, + 0.0026163870934396982, + 0.004235755652189255, + 0.0016816725255921483, + 0.002704466227442026, + 0.004789579659700394, + 0.003358755959197879, + 0.004038695711642504, + 0.0026613250374794006, + 0.002727943705394864, + 0.0036295808386057615, + 0.0021098298020660877, + 0.0024566035717725754, + 0.002691538305953145, + 0.0033703248482197523, + 0.002683460246771574, + 0.007164538372308016, + 0.022774014621973038, + 0.007858103141188622, + 0.005840573459863663, + 0.004179358948022127, + 0.024685196578502655, + 0.0035535874776542187, + 0.00915647391229868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.0067332289181649685 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly pulls back to reveal a grand, dynamic scene. At the center, the historical figure Zhuge Liang stands calmly on a mountain plain shrouded in mist. The scene embodies the exquisite aesthetic of traditional Chinese ink landscape painting. He wears a flowing white silk Hanfu robe with wide sleeves, topped by a dark gray double-breasted jacket. The sleeves sway gently with his slow movements. His face is gaunt, with a neatly trimmed black beard. His eyes are half-closed, his expression focused and serene. His left hand calmly grasps a feather fan made of pure white feathers, its handle made of dark wood, a subtle sheen. He slowly practices Tai Chi, his movements flowing with the grace and fluidity of flowing water, transitioning gently from one pose to the next, his wide sleeves drawing soft arcs through the air as his body turns. The background depicts a classic ink painting: distant mountains, layered and serene, are outlined in varying shades of ink, shrouded in wisps of mist. Nearby, a few ancient pines, rendered with vigorous brushstrokes, can be seen. A subtle soft focus effect is employed throughout, blurring the edges and enhancing the scene's dreamlike and poetic quality. The lighting, rather than originating from a specific source, is a uniform, soft, diffused light that seems to permeate the rice paper itself. There are no distinct shadows, only the rich gradations of ink color in the folds of the figures' clothing and the mountains' contours, creating a philosophical atmosphere of transcendence, tranquility, and profound tranquility.", + "seed": 27724, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_9/index_4772_seed_27724_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.038190783280879256, + "vlm_logprob": -0.038190783280879256, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.038190783280879256, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09887199848890305 + ], + "opposite_logprob_tokens": [ + -2.598871946334839 + ], + "logprob_reward": -0.09887199848890305, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09887199848890305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18337251245975494, + 0.07036533951759338, + 0.07664543390274048, + 0.12560462951660156, + 0.06514530628919601, + 0.10341111570596695, + 0.2173551321029663, + 0.07712459564208984, + 0.02678348682820797, + 0.05419246852397919 + ], + "credit_map": [ + 0.037107907235622406, + 0.019935352727770805, + 0.026664527133107185, + 0.01938267983496189, + 0.021942762657999992, + 0.026350680738687515, + 0.012936451472342014, + 0.01905214600265026, + 0.007876254618167877, + 0.007394136395305395, + 0.008675378747284412, + 0.008461296558380127, + 0.013536491431295872, + 0.00825659278780222, + 0.005637731868773699, + 0.010527458041906357, + 0.010357275605201721, + 0.008121328428387642, + 0.017419615760445595, + 0.006969335954636335, + 0.01171025913208723, + 0.008104969747364521, + 0.008402472361922264, + 0.005560178309679031, + 0.023962082341313362, + 0.012748681008815765, + 0.020945411175489426, + 0.012202917598187923, + 0.016373204067349434, + 0.019946185871958733, + 0.01137618813663721, + 0.008049948140978813, + 0.007424936629831791, + 0.007810840383172035, + 0.0056176260113716125, + 0.007795832119882107, + 0.011916626244783401, + 0.00596862006932497, + 0.010668955743312836, + 0.00794186256825924, + 0.013032475486397743, + 0.010507088154554367, + 0.007613287772983313, + 0.012470366433262825, + 0.015635348856449127, + 0.010447011329233646, + 0.01729460433125496, + 0.016410943120718002, + 0.023996837437152863, + 0.022367598488926888, + 0.019477281719446182, + 0.028616180643439293, + 0.038219694048166275, + 0.02399376779794693, + 0.04045351222157478, + 0.020230256021022797, + 0.0074322400614619255, + 0.008460769429802895, + 0.0061043486930429935, + 0.007492643315345049, + 0.01302762795239687, + 0.014755312353372574, + 0.010954169556498528, + 0.008897482417523861, + 0.0032764249481260777, + 0.0031342448201030493, + 0.0027227692771703005, + 0.0028046423103660345, + 0.004549414850771427, + 0.003866615006700158, + 0.00421449588611722, + 0.0022148792631924152, + 0.007428625598549843, + 0.012795236892998219, + 0.003718547523021698, + 0.006850419566035271, + 0.006991134025156498, + 0.005729590076953173, + 0.0078879464417696, + 0.002790966769680381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.09887199848890305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982470393180847 + ], + "opposite_logprob_tokens": [ + -3.529824733734131 + ], + "logprob_reward": -0.02982470393180847, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982470393180847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27427372336387634, + 0.14450478553771973, + 0.10477499663829803, + 0.05304905027151108, + 0.053583309054374695, + 0.058817632496356964, + 0.11746367812156677, + 0.048170022666454315, + 0.08134043961763382, + 0.06402238458395004 + ], + "credit_map": [ + 0.04067322611808777, + 0.030644351616501808, + 0.044097863137722015, + 0.02806803397834301, + 0.046002347022295, + 0.03640365228056908, + 0.02158711478114128, + 0.02679714560508728, + 0.025375045835971832, + 0.017259245738387108, + 0.021506350487470627, + 0.01492942962795496, + 0.021151993423700333, + 0.022272728383541107, + 0.012836412526667118, + 0.009173580445349216, + 0.013059922493994236, + 0.011779808439314365, + 0.013525440357625484, + 0.010509047657251358, + 0.017124779522418976, + 0.01089537050575018, + 0.01888992078602314, + 0.008990713395178318, + 0.0069432249292731285, + 0.0057197329588234425, + 0.0077706812880933285, + 0.005244139581918716, + 0.005588043015450239, + 0.007469086907804012, + 0.008733728900551796, + 0.00558041874319315, + 0.007902433164417744, + 0.005472402088344097, + 0.0056019495241343975, + 0.006177287083119154, + 0.00821853056550026, + 0.009215359576046467, + 0.006242974195629358, + 0.004752375651150942, + 0.007922369055449963, + 0.005904783960431814, + 0.008848818019032478, + 0.007292171940207481, + 0.008775283582508564, + 0.007937095127999783, + 0.0056600384414196014, + 0.006477077025920153, + 0.012236569076776505, + 0.011305452324450016, + 0.012529834173619747, + 0.01347018126398325, + 0.025260930880904198, + 0.01287255808711052, + 0.01949538290500641, + 0.010292786173522472, + 0.005294914357364178, + 0.005635291337966919, + 0.004278228618204594, + 0.007010634057223797, + 0.008182773366570473, + 0.005741368047893047, + 0.006438433658331633, + 0.0055883824825286865, + 0.008904707618057728, + 0.009297721087932587, + 0.00986428465694189, + 0.009601371362805367, + 0.01664462685585022, + 0.01016780361533165, + 0.011595042422413826, + 0.005264884326606989, + 0.01306151319295168, + 0.010893499478697777, + 0.006175957154482603, + 0.005623403936624527, + 0.0066580153070390224, + 0.009337114170193672, + 0.006495506968349218, + 0.005777378566563129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.02982470393180847 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811933100223541 + ], + "opposite_logprob_tokens": [ + -3.288119316101074 + ], + "logprob_reward": -0.03811933100223541, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811933100223541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24436227977275848, + 0.13158366084098816, + 0.11277744919061661, + 0.04907970130443573, + 0.07486121356487274, + 0.03940126672387123, + 0.11329755932092667, + 0.05081186071038246, + 0.10960457473993301, + 0.0742204561829567 + ], + "credit_map": [ + 0.03398647904396057, + 0.02812638133764267, + 0.05609867721796036, + 0.024841750040650368, + 0.029865598306059837, + 0.03364160656929016, + 0.01870688609778881, + 0.01909489557147026, + 0.01778029464185238, + 0.014781850390136242, + 0.01486702635884285, + 0.014422519132494926, + 0.021198401227593422, + 0.017131391912698746, + 0.016214631497859955, + 0.015187551267445087, + 0.015524514019489288, + 0.011829487048089504, + 0.022669771686196327, + 0.015430426225066185, + 0.017165295779705048, + 0.012124334461987019, + 0.009536691009998322, + 0.008496925234794617, + 0.005987864453345537, + 0.005116602405905724, + 0.005226115696132183, + 0.005653722211718559, + 0.009939049370586872, + 0.005186042282730341, + 0.007336723152548075, + 0.004633583128452301, + 0.0075066108256578445, + 0.008568568155169487, + 0.0062771267257630825, + 0.011379154399037361, + 0.019217558205127716, + 0.005818863864988089, + 0.009062203578650951, + 0.007031128741800785, + 0.0056522078812122345, + 0.004185395315289497, + 0.0029168815817683935, + 0.003668137127533555, + 0.005017328076064587, + 0.006054760422557592, + 0.006991666741669178, + 0.004914890043437481, + 0.012725351378321648, + 0.01186375506222248, + 0.011093529872596264, + 0.013282806612551212, + 0.023795804008841515, + 0.01326673198491335, + 0.01783629320561886, + 0.009433281607925892, + 0.006130908150225878, + 0.00627021212130785, + 0.004990092013031244, + 0.004837799351662397, + 0.007058765273541212, + 0.00950013566762209, + 0.006343066692352295, + 0.00568087724968791, + 0.01152978464961052, + 0.013681946322321892, + 0.01053494680672884, + 0.01240515522658825, + 0.030700424686074257, + 0.012653784826397896, + 0.010572112165391445, + 0.007526418659836054, + 0.010659480467438698, + 0.014336537569761276, + 0.006768820341676474, + 0.006885060574859381, + 0.010285091586411, + 0.011384344659745693, + 0.00811744574457407, + 0.005783672444522381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.03811933100223541 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818835735321045 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.01818835735321045, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818835735321045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13631027936935425, + 0.05074086785316467, + 0.06957625597715378, + 0.04593510553240776, + 0.04515649378299713, + 0.08402387797832489, + 0.17014266550540924, + 0.06752082705497742, + 0.26326796412467957, + 0.0673256516456604 + ], + "credit_map": [ + 0.014224497601389885, + 0.016574492678046227, + 0.0247725248336792, + 0.011719782836735249, + 0.01485135592520237, + 0.028329281136393547, + 0.011711125262081623, + 0.014127228409051895, + 0.007585159037262201, + 0.005876536946743727, + 0.0057428875006735325, + 0.006847165059298277, + 0.0074670445173978806, + 0.006031446624547243, + 0.005642687436193228, + 0.005547943525016308, + 0.008699457161128521, + 0.008394484408199787, + 0.013098557479679585, + 0.01100501511245966, + 0.006912337616086006, + 0.010564947500824928, + 0.005058143753558397, + 0.0058433241210877895, + 0.006606120616197586, + 0.0048258644528687, + 0.0057745566591620445, + 0.006746372673660517, + 0.005110261961817741, + 0.008714628405869007, + 0.004420457873493433, + 0.0037368438206613064, + 0.003977653570473194, + 0.005240637809038162, + 0.003467078320682049, + 0.0057229166850447655, + 0.009188271127641201, + 0.005774363409727812, + 0.0065182470716536045, + 0.005267326720058918, + 0.008259551599621773, + 0.008845594711601734, + 0.006072639022022486, + 0.006667234934866428, + 0.021585265174508095, + 0.008239543065428734, + 0.011616402305662632, + 0.012737644836306572, + 0.016378268599510193, + 0.023215342313051224, + 0.013814115896821022, + 0.02471625804901123, + 0.030872484669089317, + 0.019752778112888336, + 0.027286037802696228, + 0.014107372611761093, + 0.007136977277696133, + 0.007337613962590694, + 0.005461200140416622, + 0.006327885203063488, + 0.008175618015229702, + 0.010396632365882397, + 0.01599031686782837, + 0.006694577634334564, + 0.026887323707342148, + 0.03130481392145157, + 0.028743723407387733, + 0.031022805720567703, + 0.05712217837572098, + 0.0286391731351614, + 0.040567412972450256, + 0.018980564549565315, + 0.0064722285605967045, + 0.010213241912424564, + 0.005523603409528732, + 0.010685364715754986, + 0.008248130790889263, + 0.005312916357070208, + 0.015460792928934097, + 0.005409372039139271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.01818835735321045 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949525628238916 + ], + "opposite_logprob_tokens": [ + -5.1309494972229 + ], + "logprob_reward": -0.005949525628238916, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949525628238916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2097574919462204, + 0.06916486471891403, + 0.05935181677341461, + 0.19694115221500397, + 0.03789849951863289, + 0.05898871645331383, + 0.10526353120803833, + 0.05400528013706207, + 0.10212872177362442, + 0.10649994015693665 + ], + "credit_map": [ + 0.033728696405887604, + 0.024418089538812637, + 0.03350045531988144, + 0.016403764486312866, + 0.028458833694458008, + 0.03502817079424858, + 0.015344400890171528, + 0.022875070571899414, + 0.010812968015670776, + 0.00845392607152462, + 0.010639119893312454, + 0.007090249564498663, + 0.008906628005206585, + 0.013659953139722347, + 0.005060388706624508, + 0.004541629925370216, + 0.007055842783302069, + 0.007090333383530378, + 0.005807857029139996, + 0.009895737282931805, + 0.01122926827520132, + 0.005434696096926928, + 0.00806603766977787, + 0.004772046580910683, + 0.037725891917943954, + 0.019753778353333473, + 0.031789205968379974, + 0.020717665553092957, + 0.016762973740696907, + 0.043300189077854156, + 0.015009281225502491, + 0.011882155202329159, + 0.004455788526684046, + 0.004032627213746309, + 0.003976610489189625, + 0.003931938670575619, + 0.006335609592497349, + 0.006217753514647484, + 0.0048204329796135426, + 0.004127739928662777, + 0.009363867342472076, + 0.005815753247588873, + 0.008664031513035297, + 0.0057066213339567184, + 0.006017111241817474, + 0.008895268663764, + 0.007603353820741177, + 0.006922708824276924, + 0.017658017575740814, + 0.01251154113560915, + 0.00860962737351656, + 0.01382434368133545, + 0.016202080994844437, + 0.016895905137062073, + 0.011181032285094261, + 0.008380981162190437, + 0.0053456430323421955, + 0.004876644350588322, + 0.005707974079996347, + 0.006084763444960117, + 0.008983029052615166, + 0.00659120362251997, + 0.00967483501881361, + 0.006741182878613472, + 0.012508799321949482, + 0.011540771462023258, + 0.014845965430140495, + 0.010492060333490372, + 0.01915646158158779, + 0.012503363192081451, + 0.011232121847569942, + 0.00984918512403965, + 0.01691693440079689, + 0.02149442955851555, + 0.010230813175439835, + 0.009972603060305119, + 0.01679680123925209, + 0.010764281265437603, + 0.010552333667874336, + 0.009771748445928097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.005949525628238916 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly pulls back to reveal a grand, dynamic scene. At the center, the historical figure Zhuge Liang stands calmly on a mountain plain shrouded in mist. The scene embodies the exquisite aesthetic of traditional Chinese ink landscape painting. He wears a flowing white silk Hanfu robe with wide sleeves, topped by a dark gray double-breasted jacket. The sleeves sway gently with his slow movements. His face is gaunt, with a neatly trimmed black beard. His eyes are half-closed, his expression focused and serene. His left hand calmly grasps a feather fan made of pure white feathers, its handle made of dark wood, a subtle sheen. He slowly practices Tai Chi, his movements flowing with the grace and fluidity of flowing water, transitioning gently from one pose to the next, his wide sleeves drawing soft arcs through the air as his body turns. The background depicts a classic ink painting: distant mountains, layered and serene, are outlined in varying shades of ink, shrouded in wisps of mist. Nearby, a few ancient pines, rendered with vigorous brushstrokes, can be seen. A subtle soft focus effect is employed throughout, blurring the edges and enhancing the scene's dreamlike and poetic quality. The lighting, rather than originating from a specific source, is a uniform, soft, diffused light that seems to permeate the rice paper itself. There are no distinct shadows, only the rich gradations of ink color in the folds of the figures' clothing and the mountains' contours, creating a philosophical atmosphere of transcendence, tranquility, and profound tranquility.", + "seed": 27725, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_9/index_4772_seed_27725_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.033054542168974875, + "vlm_logprob": -0.033054542168974875, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.033054542168974875, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05839075520634651 + ], + "opposite_logprob_tokens": [ + -3.9333908557891846 + ], + "logprob_reward": -0.05839075520634651, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.05839075520634651, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10164020955562592, + 0.06089809164404869, + 0.07004688680171967, + 0.07732761651277542, + 0.15238678455352783, + 0.03989885747432709, + 0.09813238680362701, + 0.1588873565196991, + 0.1649678647518158, + 0.07581393420696259 + ], + "credit_map": [ + 0.018237750977277756, + 0.014852039515972137, + 0.010575400665402412, + 0.01705099083483219, + 0.012895691208541393, + 0.012328899465501308, + 0.009727894328534603, + 0.0059715392999351025, + 0.011412442661821842, + 0.012619223445653915, + 0.005508318077772856, + 0.004867481999099255, + 0.011690949089825153, + 0.006846561562269926, + 0.004957320634275675, + 0.0029957906808704138, + 0.01155067514628172, + 0.009828398004174232, + 0.004321357700973749, + 0.007065661251544952, + 0.013815479353070259, + 0.006634032353758812, + 0.013513023965060711, + 0.0033182636834681034, + 0.010192793793976307, + 0.010001072660088539, + 0.00693175895139575, + 0.010949869640171528, + 0.012968488968908787, + 0.005783188622444868, + 0.01550324261188507, + 0.0049972073175013065, + 0.0162662286311388, + 0.01951509155333042, + 0.012758961878716946, + 0.016140833497047424, + 0.030885914340615273, + 0.019448071718215942, + 0.024819454178214073, + 0.012552228756248951, + 0.006603377405554056, + 0.005741323810070753, + 0.0031230468302965164, + 0.004946490749716759, + 0.006315784063190222, + 0.004188711289316416, + 0.005207805894315243, + 0.003772316500544548, + 0.017000550404191017, + 0.011616201139986515, + 0.008873877115547657, + 0.008697117678821087, + 0.017104506492614746, + 0.015744352713227272, + 0.014501156285405159, + 0.00459462683647871, + 0.03612901270389557, + 0.027769295498728752, + 0.013119393028318882, + 0.014361674897372723, + 0.02322070486843586, + 0.026593072339892387, + 0.010401340201497078, + 0.0072928559966385365, + 0.03716704249382019, + 0.02440427616238594, + 0.0171987172216177, + 0.016124555841088295, + 0.020924022421240807, + 0.025083113461732864, + 0.013570110313594341, + 0.01049602497369051, + 0.00994002353399992, + 0.011063389480113983, + 0.00613489979878068, + 0.007720699533820152, + 0.015966977924108505, + 0.009143336676061153, + 0.008673431351780891, + 0.007171169389039278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.05839075520634651 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371976688504219 + ], + "opposite_logprob_tokens": [ + -3.408719778060913 + ], + "logprob_reward": -0.03371976688504219, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371976688504219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1294664442539215, + 0.05939297750592232, + 0.14851854741573334, + 0.0701233521103859, + 0.05581182613968849, + 0.059792470186948776, + 0.0634404793381691, + 0.09891477227210999, + 0.2728220820426941, + 0.04171702265739441 + ], + "credit_map": [ + 0.02319432981312275, + 0.02198649011552334, + 0.01100525725632906, + 0.023472309112548828, + 0.019149502739310265, + 0.011814328841865063, + 0.012102445587515831, + 0.006741799414157867, + 0.00796898640692234, + 0.007362141273915768, + 0.006911857984960079, + 0.008399803191423416, + 0.010067508555948734, + 0.009385177865624428, + 0.006051469594240189, + 0.0032460405491292477, + 0.020211609080433846, + 0.02039497159421444, + 0.01088478323072195, + 0.01650123856961727, + 0.024536222219467163, + 0.011831831187009811, + 0.03541339188814163, + 0.008744516409933567, + 0.010234632529318333, + 0.009854277595877647, + 0.007563605438917875, + 0.012324240058660507, + 0.007704064249992371, + 0.0046318587847054005, + 0.01377289742231369, + 0.0040377904660999775, + 0.0066041499376297, + 0.007371688727289438, + 0.0046140574850142, + 0.0065758200362324715, + 0.010224420577287674, + 0.007023596670478582, + 0.008749198168516159, + 0.004648903850466013, + 0.007776663172990084, + 0.008955636993050575, + 0.003462830325588584, + 0.006689315661787987, + 0.012449493631720543, + 0.005188539624214172, + 0.009400976821780205, + 0.005869023501873016, + 0.010397842153906822, + 0.008131296373903751, + 0.006574252620339394, + 0.010020321235060692, + 0.00814725924283266, + 0.0064595965668559074, + 0.010104880668222904, + 0.003605037461966276, + 0.016792140901088715, + 0.01542715821415186, + 0.0072873057797551155, + 0.011737623251974583, + 0.015543678775429726, + 0.01322366576641798, + 0.011257892474532127, + 0.007645314559340477, + 0.057098884135484695, + 0.038699325174093246, + 0.03339005634188652, + 0.022329887375235558, + 0.04192109405994415, + 0.041878487914800644, + 0.023267608135938644, + 0.014236753806471825, + 0.004124756902456284, + 0.005686165764927864, + 0.003407906275242567, + 0.0049715349450707436, + 0.009166362695395947, + 0.004170416388660669, + 0.0058940378949046135, + 0.00429584551602602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.03371976688504219 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038113366812467575 + ], + "opposite_logprob_tokens": [ + -3.2881133556365967 + ], + "logprob_reward": -0.038113366812467575, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038113366812467575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11899017542600632, + 0.09860938042402267, + 0.05348584055900574, + 0.0794774740934372, + 0.05754692852497101, + 0.08772288262844086, + 0.08933287858963013, + 0.10573375970125198, + 0.2404271364212036, + 0.0686735138297081 + ], + "credit_map": [ + 0.02272103540599346, + 0.01598854549229145, + 0.013554470613598824, + 0.026941785588860512, + 0.012797640636563301, + 0.011273890733718872, + 0.010193221271038055, + 0.005519594065845013, + 0.013141833245754242, + 0.013353374786674976, + 0.008121701888740063, + 0.016605358570814133, + 0.016149645671248436, + 0.011982275173068047, + 0.014749836176633835, + 0.00450536422431469, + 0.0068869865499436855, + 0.008167830295860767, + 0.004676107317209244, + 0.006312270648777485, + 0.008788046427071095, + 0.00499368691816926, + 0.010272323153913021, + 0.003388597397133708, + 0.019253600388765335, + 0.01301493775099516, + 0.0062639047391712666, + 0.00918667670339346, + 0.00767339114099741, + 0.008206598460674286, + 0.011313257738947868, + 0.004565107170492411, + 0.006536539178341627, + 0.007475127931684256, + 0.004651524126529694, + 0.006401033140718937, + 0.01251709833741188, + 0.005764715373516083, + 0.010104089044034481, + 0.004096807446330786, + 0.014561436139047146, + 0.012475878931581974, + 0.004942288622260094, + 0.009809090755879879, + 0.017291521653532982, + 0.008231081999838352, + 0.012508180923759937, + 0.007903406396508217, + 0.011480677872896194, + 0.011246678419411182, + 0.00890492182224989, + 0.0118208397179842, + 0.015141342766582966, + 0.009649598971009254, + 0.01430478971451521, + 0.006784042343497276, + 0.0166013166308403, + 0.013756470754742622, + 0.010820738971233368, + 0.015598716214299202, + 0.015269012190401554, + 0.014159414917230606, + 0.00988241471350193, + 0.009645686484873295, + 0.05052468925714493, + 0.033973585814237595, + 0.02967381663620472, + 0.019565759226679802, + 0.03626064583659172, + 0.03746967762708664, + 0.01915084756910801, + 0.013808139599859715, + 0.007694387808442116, + 0.010245608165860176, + 0.004628519527614117, + 0.008414492011070251, + 0.016820549964904785, + 0.007078233174979687, + 0.008025753311812878, + 0.005765975918620825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.038113366812467575 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029797280207276344 + ], + "opposite_logprob_tokens": [ + -3.529797315597534 + ], + "logprob_reward": -0.029797280207276344, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029797280207276344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08184779435396194, + 0.12085457146167755, + 0.04857353866100311, + 0.11498453468084335, + 0.045381054282188416, + 0.0872667133808136, + 0.09415882080793381, + 0.06891283392906189, + 0.2799931466579437, + 0.0580269955098629 + ], + "credit_map": [ + 0.01376426313072443, + 0.016385572031140327, + 0.008707373403012753, + 0.009973269887268543, + 0.009331637062132359, + 0.011741315014660358, + 0.007295270450413227, + 0.004649087321013212, + 0.015558130107820034, + 0.018278975039720535, + 0.015866467729210854, + 0.01645047776401043, + 0.019735705107450485, + 0.015310592949390411, + 0.013006295077502728, + 0.006647924426943064, + 0.005067143123596907, + 0.0059068170376122, + 0.008371050469577312, + 0.007256208918988705, + 0.009796073660254478, + 0.004493711981922388, + 0.004701376426964998, + 0.002981156576424837, + 0.018675893545150757, + 0.015030610375106335, + 0.0076786819845438, + 0.01227506436407566, + 0.0168309286236763, + 0.010645739734172821, + 0.02452899143099785, + 0.009318622760474682, + 0.007106528617441654, + 0.0059324000030756, + 0.0041658999398350716, + 0.006536722648888826, + 0.0068115959875285625, + 0.005541383754462004, + 0.005499419756233692, + 0.0037871019449084997, + 0.011725028976798058, + 0.011555243283510208, + 0.0058175246231257915, + 0.007831319235265255, + 0.019444987177848816, + 0.0073807211592793465, + 0.01275918260216713, + 0.010752703994512558, + 0.012772073037922382, + 0.011129245162010193, + 0.010229905135929585, + 0.013063895516097546, + 0.01702192984521389, + 0.008379001170396805, + 0.013775799423456192, + 0.007786968722939491, + 0.012126494199037552, + 0.011188550852239132, + 0.006506933830678463, + 0.008985484018921852, + 0.008826741017401218, + 0.009440743364393711, + 0.006511066574603319, + 0.00532681867480278, + 0.057828180491924286, + 0.038527995347976685, + 0.03476693853735924, + 0.026653921231627464, + 0.03879833593964577, + 0.043140310794115067, + 0.02342161536216736, + 0.016855832189321518, + 0.006535230204463005, + 0.007802702486515045, + 0.005916946567595005, + 0.008895195089280605, + 0.009840644896030426, + 0.00785109493881464, + 0.006764705758541822, + 0.0044204774312675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.029797280207276344 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525154173374176 + ], + "opposite_logprob_tokens": [ + -5.255251407623291 + ], + "logprob_reward": -0.00525154173374176, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525154173374176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11388807743787766, + 0.044591981917619705, + 0.02508593164384365, + 0.09011293947696686, + 0.029850410297513008, + 0.031971514225006104, + 0.12653326988220215, + 0.20826677978038788, + 0.13107559084892273, + 0.19862361252307892 + ], + "credit_map": [ + 0.02016112022101879, + 0.021013015881180763, + 0.014158902689814568, + 0.01325662899762392, + 0.016924453899264336, + 0.01360876765102148, + 0.008580799214541912, + 0.006184371653944254, + 0.007168206386268139, + 0.008175595663487911, + 0.005530328024178743, + 0.0048080082051455975, + 0.007614955306053162, + 0.005427776835858822, + 0.0034157298505306244, + 0.0024513774551451206, + 0.0041856965981423855, + 0.00360333570279181, + 0.002546139759942889, + 0.0027969535440206528, + 0.0034452371764928102, + 0.003925564233213663, + 0.00300190644338727, + 0.0015810965560376644, + 0.011437258683145046, + 0.010253866203129292, + 0.00809375662356615, + 0.014406094327569008, + 0.014737159013748169, + 0.007968232966959476, + 0.01747421734035015, + 0.005742346867918968, + 0.00554486270993948, + 0.004380427300930023, + 0.0024284024257212877, + 0.00401076627895236, + 0.003511992748826742, + 0.003879074938595295, + 0.0044351620599627495, + 0.0016597190406173468, + 0.006212092470377684, + 0.004461216740310192, + 0.004806367214769125, + 0.004367474932223558, + 0.0033789670560508966, + 0.0030291809234768152, + 0.0030950417276471853, + 0.0026211710646748543, + 0.0318189412355423, + 0.018253646790981293, + 0.01359530072659254, + 0.019564112648367882, + 0.011715926229953766, + 0.01396764162927866, + 0.010742703452706337, + 0.006874982267618179, + 0.03971187770366669, + 0.026602312922477722, + 0.019312266260385513, + 0.024652643129229546, + 0.028753528371453285, + 0.028658762574195862, + 0.02290349081158638, + 0.01767188496887684, + 0.029067885130643845, + 0.01776103489100933, + 0.01587267592549324, + 0.01201784610748291, + 0.018539125099778175, + 0.01863154023885727, + 0.010759334079921246, + 0.008426139131188393, + 0.037702776491642, + 0.029958240687847137, + 0.022420156747102737, + 0.02057119458913803, + 0.03164985775947571, + 0.02332182414829731, + 0.0174631979316473, + 0.01553633064031601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.00525154173374176 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly pulls back to reveal a grand, dynamic scene. At the center, the historical figure Zhuge Liang stands calmly on a mountain plain shrouded in mist. The scene embodies the exquisite aesthetic of traditional Chinese ink landscape painting. He wears a flowing white silk Hanfu robe with wide sleeves, topped by a dark gray double-breasted jacket. The sleeves sway gently with his slow movements. His face is gaunt, with a neatly trimmed black beard. His eyes are half-closed, his expression focused and serene. His left hand calmly grasps a feather fan made of pure white feathers, its handle made of dark wood, a subtle sheen. He slowly practices Tai Chi, his movements flowing with the grace and fluidity of flowing water, transitioning gently from one pose to the next, his wide sleeves drawing soft arcs through the air as his body turns. The background depicts a classic ink painting: distant mountains, layered and serene, are outlined in varying shades of ink, shrouded in wisps of mist. Nearby, a few ancient pines, rendered with vigorous brushstrokes, can be seen. A subtle soft focus effect is employed throughout, blurring the edges and enhancing the scene's dreamlike and poetic quality. The lighting, rather than originating from a specific source, is a uniform, soft, diffused light that seems to permeate the rice paper itself. There are no distinct shadows, only the rich gradations of ink color in the folds of the figures' clothing and the mountains' contours, creating a philosophical atmosphere of transcendence, tranquility, and profound tranquility.", + "seed": 27726, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000108/group_9/index_4772_seed_27726_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03118192180991173, + "vlm_logprob": -0.03118192180991173, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.03118192180991173, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05249553546309471 + ], + "opposite_logprob_tokens": [ + -3.6774954795837402 + ], + "logprob_reward": -0.05249553546309471, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.05249553546309471, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1529301553964615, + 0.12117801606655121, + 0.05735975131392479, + 0.0669327825307846, + 0.05508669838309288, + 0.19020019471645355, + 0.10573852062225342, + 0.10496840626001358, + 0.07581070810556412, + 0.06979480385780334 + ], + "credit_map": [ + 0.02498111128807068, + 0.019908184185624123, + 0.016849184408783913, + 0.020477792248129845, + 0.015196750871837139, + 0.03091822750866413, + 0.016481315717101097, + 0.008117586374282837, + 0.023229552432894707, + 0.019526055082678795, + 0.012589563615620136, + 0.015542082488536835, + 0.013403890654444695, + 0.0169968418776989, + 0.011724524199962616, + 0.008165501058101654, + 0.00594995217397809, + 0.007538404781371355, + 0.007004803512245417, + 0.006513276137411594, + 0.008698039688169956, + 0.012105218134820461, + 0.006010131910443306, + 0.0035399270709604025, + 0.00771918473765254, + 0.00943260919302702, + 0.0077018095180392265, + 0.0069033014588057995, + 0.010219894349575043, + 0.010370681993663311, + 0.00871297251433134, + 0.005872331093996763, + 0.005499567370861769, + 0.00769905187189579, + 0.006178408395498991, + 0.005758087150752544, + 0.007944614626467228, + 0.010083939880132675, + 0.008359701372683048, + 0.0035633263178169727, + 0.02779180370271206, + 0.031060658395290375, + 0.020873330533504486, + 0.024912208318710327, + 0.024040617048740387, + 0.026520641520619392, + 0.023718012496829033, + 0.011282900348305702, + 0.011991334147751331, + 0.016064487397670746, + 0.009507108479738235, + 0.025706326588988304, + 0.010986105538904667, + 0.010769493877887726, + 0.014931662008166313, + 0.005782001186162233, + 0.020690344274044037, + 0.013262585736811161, + 0.009207341820001602, + 0.011378435418009758, + 0.009707746095955372, + 0.020721344277262688, + 0.012353908270597458, + 0.007646697573363781, + 0.013713769614696503, + 0.009442687034606934, + 0.007485650014132261, + 0.007047964259982109, + 0.006368925794959068, + 0.017053868621587753, + 0.010537998750805855, + 0.00415984308347106, + 0.009695157408714294, + 0.010968533344566822, + 0.005628514569252729, + 0.0065752980299293995, + 0.0140227684751153, + 0.005746609531342983, + 0.012868613936007023, + 0.0042893048375844955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.05249553546309471 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811325132846832 + ], + "opposite_logprob_tokens": [ + -3.2881133556365967 + ], + "logprob_reward": -0.03811325132846832, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811325132846832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21777793765068054, + 0.11175782978534698, + 0.0429268516600132, + 0.055985692888498306, + 0.12063725292682648, + 0.10413738340139389, + 0.12051387131214142, + 0.04754230007529259, + 0.1322900801897049, + 0.0464307926595211 + ], + "credit_map": [ + 0.027821863070130348, + 0.03243796527385712, + 0.01663886569440365, + 0.047139693051576614, + 0.028392331674695015, + 0.025962254032492638, + 0.025821879506111145, + 0.013563078828155994, + 0.0229830089956522, + 0.021252579987049103, + 0.008428642526268959, + 0.01428907085210085, + 0.013645583763718605, + 0.012667924165725708, + 0.012610828503966331, + 0.005880191922187805, + 0.0036148845683783293, + 0.006039617117494345, + 0.003995464649051428, + 0.006775547284632921, + 0.005013621412217617, + 0.009938196279108524, + 0.005064773838967085, + 0.0024847486056387424, + 0.006941340398043394, + 0.007120945956557989, + 0.004895511548966169, + 0.00775046506896615, + 0.007426336407661438, + 0.009603296406567097, + 0.008136452175676823, + 0.004111346788704395, + 0.011847512796521187, + 0.018877791240811348, + 0.011699271388351917, + 0.014210770837962627, + 0.017523949965834618, + 0.020598484203219414, + 0.013722802512347698, + 0.012156671844422817, + 0.018340356647968292, + 0.02083367109298706, + 0.008270272985100746, + 0.01399881113320589, + 0.010156188160181046, + 0.012165099382400513, + 0.011956201866269112, + 0.008416780270636082, + 0.02612585574388504, + 0.018586626276373863, + 0.011531424708664417, + 0.012680971063673496, + 0.009901167824864388, + 0.02254781872034073, + 0.01214356068521738, + 0.006996450014412403, + 0.008494371548295021, + 0.005747782997786999, + 0.004758391994982958, + 0.007684988901019096, + 0.005459294654428959, + 0.005748485680669546, + 0.006023618392646313, + 0.0036253640428185463, + 0.027040354907512665, + 0.016919398680329323, + 0.009579673409461975, + 0.008548308163881302, + 0.012578792870044708, + 0.03807750344276428, + 0.012145120650529861, + 0.007400926668196917, + 0.005079903639853001, + 0.007897513918578625, + 0.003747593378648162, + 0.004327040631324053, + 0.009964553639292717, + 0.0035499264486134052, + 0.00947690848261118, + 0.002387350657954812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.03811325132846832 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370778262615204 + ], + "opposite_logprob_tokens": [ + -3.408707857131958 + ], + "logprob_reward": -0.03370778262615204, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370778262615204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19378355145454407, + 0.12246239185333252, + 0.11247923970222473, + 0.040544282644987106, + 0.07096157222986221, + 0.09484247863292694, + 0.08943953365087509, + 0.03930337354540825, + 0.1460869014263153, + 0.09009663015604019 + ], + "credit_map": [ + 0.026411239057779312, + 0.0250809695571661, + 0.019764207303524017, + 0.03980047628283501, + 0.02395549975335598, + 0.024199625477194786, + 0.021728822961449623, + 0.012842708267271519, + 0.027199149131774902, + 0.023424647748470306, + 0.008851609192788601, + 0.014888082630932331, + 0.014269698411226273, + 0.012693398632109165, + 0.013668006286025047, + 0.00746779004111886, + 0.011273698881268501, + 0.01659526862204075, + 0.011708545498549938, + 0.011398781090974808, + 0.019662557169795036, + 0.01904500089585781, + 0.017255732789635658, + 0.005539651494473219, + 0.004705523140728474, + 0.005647731013596058, + 0.0035911304876208305, + 0.00564698688685894, + 0.004334033466875553, + 0.007606371305882931, + 0.006281378213316202, + 0.0027311269659549, + 0.007040001451969147, + 0.009244932793080807, + 0.008382466621696949, + 0.010117179714143276, + 0.009732394479215145, + 0.009778185747563839, + 0.012220488861203194, + 0.004445914179086685, + 0.01691463217139244, + 0.01826239749789238, + 0.010140575468540192, + 0.010513780638575554, + 0.007673052605241537, + 0.013979067094624043, + 0.010298165492713451, + 0.007060807663947344, + 0.01433079969137907, + 0.012546178884804249, + 0.010759719647467136, + 0.013248097151517868, + 0.00997511763125658, + 0.015196878463029861, + 0.008278310298919678, + 0.005104427225887775, + 0.0065377941355109215, + 0.004856133833527565, + 0.0035747161600738764, + 0.0036993029061704874, + 0.0036222683265805244, + 0.008450745604932308, + 0.004732868168503046, + 0.0038295448757708073, + 0.03245428204536438, + 0.018309706822037697, + 0.011889412067830563, + 0.008853239938616753, + 0.011045310646295547, + 0.043369486927986145, + 0.011357800103724003, + 0.008807663805782795, + 0.010138805955648422, + 0.01577392779290676, + 0.0069676851853728294, + 0.008504391647875309, + 0.01984603889286518, + 0.006272722966969013, + 0.017223171889781952, + 0.005369887687265873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.03370778262615204 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634398825466633 + ], + "opposite_logprob_tokens": [ + -3.651344060897827 + ], + "logprob_reward": -0.02634398825466633, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634398825466633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14878472685813904, + 0.10898295044898987, + 0.06019861251115799, + 0.04839882627129555, + 0.06686125695705414, + 0.08526761084794998, + 0.07993929833173752, + 0.05489065870642662, + 0.12546274065971375, + 0.2212134152650833 + ], + "credit_map": [ + 0.01633795164525509, + 0.016958260908722878, + 0.01794494315981865, + 0.035303544253110886, + 0.012930847704410553, + 0.01974320225417614, + 0.017876191064715385, + 0.011689762584865093, + 0.016912147402763367, + 0.015834990888834, + 0.012503708712756634, + 0.013880389742553234, + 0.012910455465316772, + 0.012952263467013836, + 0.01855400763452053, + 0.005434975493699312, + 0.007094650994986296, + 0.008363060653209686, + 0.0050663212314248085, + 0.008303157985210419, + 0.010108301416039467, + 0.009238808415830135, + 0.008547808043658733, + 0.003476493526250124, + 0.006727909203618765, + 0.006636630743741989, + 0.004996865522116423, + 0.006325145717710257, + 0.006508445832878351, + 0.00818702857941389, + 0.005790113937109709, + 0.003226683707907796, + 0.005709122400730848, + 0.0073333322070539, + 0.008188427425920963, + 0.008388221263885498, + 0.00847556535154581, + 0.013889377936720848, + 0.010212854482233524, + 0.0046643479727208614, + 0.008801089599728584, + 0.013078144751489162, + 0.010098246857523918, + 0.011296774260699749, + 0.006991312839090824, + 0.01667344570159912, + 0.009888493455946445, + 0.008440092206001282, + 0.018589843064546585, + 0.012441041879355907, + 0.005985815543681383, + 0.007848355919122696, + 0.010471672751009464, + 0.011423089541494846, + 0.008754282258450985, + 0.004425191320478916, + 0.008873220533132553, + 0.007217985112220049, + 0.006222558673471212, + 0.008536390960216522, + 0.007257932331413031, + 0.005531375762075186, + 0.0075994920916855335, + 0.0036516983527690172, + 0.025178760290145874, + 0.015867309644818306, + 0.00990569218993187, + 0.0116842370480299, + 0.013839387334883213, + 0.027457693591713905, + 0.01520145870745182, + 0.00632818229496479, + 0.03204083815217018, + 0.036472633481025696, + 0.013637866824865341, + 0.018379071727395058, + 0.03775496780872345, + 0.018131472170352936, + 0.05455783009529114, + 0.010238703340291977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.02634398825466633 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure Zhuge Liang wearing a white silk Hanfu robe with a dark gray double-breasted jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character holding a feather fan with white feathers and a dark wood handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character perform slow Tai Chi movements with swaying sleeves in a misty mountain plain setting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a traditional Chinese ink landscape painting featuring soft focus and diffused lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052490513771772385 + ], + "opposite_logprob_tokens": [ + -5.2552490234375 + ], + "logprob_reward": -0.0052490513771772385, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052490513771772385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21506431698799133, + 0.0949864462018013, + 0.05445100739598274, + 0.032985467463731766, + 0.05341506004333496, + 0.12858419120311737, + 0.15240347385406494, + 0.05490103363990784, + 0.10461729764938354, + 0.10859176516532898 + ], + "credit_map": [ + 0.03242039680480957, + 0.022515922784805298, + 0.02367798052728176, + 0.025988154113292694, + 0.023766662925481796, + 0.04921935498714447, + 0.021685201674699783, + 0.015790633857250214, + 0.013316535390913486, + 0.014216354116797447, + 0.011506001465022564, + 0.009349876083433628, + 0.011745214462280273, + 0.01933138445019722, + 0.007457274943590164, + 0.008063802495598793, + 0.006307283416390419, + 0.007860833778977394, + 0.00498979352414608, + 0.007690527010709047, + 0.008254432119429111, + 0.010221400298178196, + 0.005834417883306742, + 0.003292321925982833, + 0.0037112433928996325, + 0.004342225845903158, + 0.0035348543897271156, + 0.0030852591153234243, + 0.0036515570245683193, + 0.009095555171370506, + 0.003363828407600522, + 0.0022009441163390875, + 0.005405678879469633, + 0.007048523984849453, + 0.008683864958584309, + 0.005535796284675598, + 0.005183760076761246, + 0.010422717779874802, + 0.007623718585819006, + 0.0035110018216073513, + 0.014554609544575214, + 0.017433367669582367, + 0.017184797674417496, + 0.02130017802119255, + 0.009928885847330093, + 0.02234577387571335, + 0.014712106436491013, + 0.011124465614557266, + 0.03437652811408043, + 0.023918693885207176, + 0.01472028624266386, + 0.01865677535533905, + 0.012185344472527504, + 0.02235097438097, + 0.01473894901573658, + 0.011455913074314594, + 0.0113315898925066, + 0.008800883777439594, + 0.004948554094880819, + 0.00537601625546813, + 0.005417710170149803, + 0.008455184288322926, + 0.0064438930712640285, + 0.004127208609133959, + 0.025873899459838867, + 0.01410109456628561, + 0.012926475144922733, + 0.011035352014005184, + 0.0071419705636799335, + 0.017552277073264122, + 0.010176792740821838, + 0.005809440277516842, + 0.021126877516508102, + 0.01297712977975607, + 0.013451655395328999, + 0.011429755948483944, + 0.011379709467291832, + 0.013875013217329979, + 0.015425289049744606, + 0.008926323615014553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0052490513771772385 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_0/index_tensor([8544, 8544])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_0/index_tensor([8544, 8544])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..463e1764cd3b2ca600d027cbbf897790f479f897 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_0/index_tensor([8544, 8544])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This low-angle, static, extreme close-up, captured with the cinematic clarity of a Sony FX3 paired with a master lens, focuses on a vibrant blue plastic bottle cap. The cap, with its delicately molded, serrated edge, spins rapidly and hypnotically on its flat top, becoming a circular blur against the gritty texture of the dusty concrete floor of the empty garage. A low golden hour light streaks across the scene from beneath the closed garage door, illuminating the fine dust particles on the ground and casting long, dynamic shadows from the spinning object. As the rotational energy dissipates, the cap's perfect rotation degenerates into an increasingly erratic and unstable oscillation, its edge beginning to rhythmically collide with the concrete surface. Finally, as its momentum completely depletes, the cap suddenly tilts sideways with a faint plastic click, slides a short distance, traces a tiny, clean arc through the dust layer, and then falls into absolute stillness, its brief, vibrant life ending in the quiet solitude of the garage. The shallow depth of field casts the crumbling concrete and shadows in the distance into a soft, blurred bokeh, creating an atmosphere of focus, contemplation, and strange poignancy in the final moments of movement of mundane objects.", + "seed": 7706, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_0/index_8544_seed_7706_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04370252760127187, + "vlm_logprob": -0.04370252760127187, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.04370252760127187, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028224823996424675 + ], + "opposite_logprob_tokens": [ + -3.9032247066497803 + ], + "logprob_reward": -0.028224823996424675, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.028224823996424675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2201451063156128, + 0.09143275767564774, + 0.06306197494268417, + 0.07832274585962296, + 0.2875564992427826, + 0.027002161368727684, + 0.07167026400566101, + 0.03042088821530342, + 0.06816983968019485, + 0.06221773102879524 + ], + "credit_map": [ + 0.02372688055038452, + 0.019075926393270493, + 0.020146457478404045, + 0.046450160443782806, + 0.01746438629925251, + 0.019986465573310852, + 0.043397191911935806, + 0.029897641390562057, + 0.010563910938799381, + 0.008438363671302795, + 0.012786125764250755, + 0.012317351065576077, + 0.009564302861690521, + 0.01632111147046089, + 0.007385880220681429, + 0.014055716805160046, + 0.008466175757348537, + 0.009229499846696854, + 0.00929486844688654, + 0.007950120605528355, + 0.0057735866867005825, + 0.006112386472523212, + 0.0069507877342402935, + 0.009284544736146927, + 0.007323793601244688, + 0.006193527020514011, + 0.01622338593006134, + 0.009775574319064617, + 0.007163583301007748, + 0.008489441126585007, + 0.00733638321980834, + 0.015817055478692055, + 0.039129965007305145, + 0.09847916662693024, + 0.0690445676445961, + 0.03281914070248604, + 0.01724693924188614, + 0.011624404229223728, + 0.006827476434409618, + 0.012384828180074692, + 0.003087329911068082, + 0.002774490974843502, + 0.0023972629569470882, + 0.004566961899399757, + 0.002210753969848156, + 0.002143457531929016, + 0.004687512293457985, + 0.0051343911327421665, + 0.009429359808564186, + 0.004013308323919773, + 0.010482092387974262, + 0.01117311604321003, + 0.006454658694565296, + 0.006772439461201429, + 0.0056182388216257095, + 0.017727041617035866, + 0.0038348438683897257, + 0.0019634864293038845, + 0.005841750651597977, + 0.006008615717291832, + 0.0015891892835497856, + 0.0028728146571666002, + 0.0030704308301210403, + 0.005239755846560001, + 0.006904576905071735, + 0.012244840152561665, + 0.009805301204323769, + 0.011983140371739864, + 0.00973505713045597, + 0.006457663141191006, + 0.004828990902751684, + 0.006210271269083023, + 0.0039916750974953175, + 0.0037122119683772326, + 0.007515404839068651, + 0.02461237460374832, + 0.002803632989525795, + 0.002429529093205929, + 0.0055105905048549175, + 0.01164231263101101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.028224823996424675 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707871198654175 + ], + "opposite_logprob_tokens": [ + -2.1270787715911865 + ], + "logprob_reward": -0.12707871198654175, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707871198654175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10400007665157318, + 0.10463768243789673, + 0.12542732059955597, + 0.11126793920993805, + 0.4645932614803314, + 0.010336167179048061, + 0.013432624749839306, + 0.006042541470378637, + 0.028729988262057304, + 0.031532373279333115 + ], + "credit_map": [ + 0.013190765865147114, + 0.011732138693332672, + 0.009289217181503773, + 0.019573930650949478, + 0.0077337054535746574, + 0.012040667235851288, + 0.0146465003490448, + 0.015793154016137123, + 0.01077247690409422, + 0.015304045751690865, + 0.01344391144812107, + 0.01007915660738945, + 0.010247173719108105, + 0.01774669997394085, + 0.005791158881038427, + 0.021253058686852455, + 0.016478808596730232, + 0.02241171896457672, + 0.012718583457171917, + 0.015616300515830517, + 0.013159722089767456, + 0.015414279885590076, + 0.013311558403074741, + 0.016316350549459457, + 0.010674149729311466, + 0.010338779538869858, + 0.02124202810227871, + 0.010675646364688873, + 0.0137835294008255, + 0.016219692304730415, + 0.010262398049235344, + 0.018071725964546204, + 0.03927520662546158, + 0.09598296135663986, + 0.1408138871192932, + 0.11560642719268799, + 0.019383180886507034, + 0.016381626948714256, + 0.01355914305895567, + 0.023590875789523125, + 0.0009189148549921811, + 0.0011623990722000599, + 0.0017233205726370215, + 0.0024197904858738184, + 0.0007744100294075906, + 0.0009778281673789024, + 0.0008382098749279976, + 0.0015212945872917771, + 0.001463017426431179, + 0.0009163357899524271, + 0.0018571752589195967, + 0.003350734943524003, + 0.0006981438491493464, + 0.0014009232399985194, + 0.0010388904483988881, + 0.0027074050158262253, + 0.0008891593315638602, + 0.0007810083334334195, + 0.0006942335166968405, + 0.0009634281741455197, + 0.0003598763723857701, + 0.0008552393992431462, + 0.0005662539624609053, + 0.0009333428461104631, + 0.0031278757378458977, + 0.00549464114010334, + 0.004412260837852955, + 0.0042893728241324425, + 0.0047573489136993885, + 0.0021374851930886507, + 0.002154815010726452, + 0.0023561902344226837, + 0.0020225965417921543, + 0.002126561477780342, + 0.004038010723888874, + 0.010636932216584682, + 0.001072200946509838, + 0.002114390954375267, + 0.0021657883189618587, + 0.007355893962085247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.12707871198654175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306533932685852 + ], + "opposite_logprob_tokens": [ + -3.168065309524536 + ], + "logprob_reward": -0.04306533932685852, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306533932685852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07063402980566025, + 0.09693784266710281, + 0.0738578662276268, + 0.04198816046118736, + 0.3295939266681671, + 0.04249245673418045, + 0.08226899057626724, + 0.014460197649896145, + 0.03762252628803253, + 0.2101440131664276 + ], + "credit_map": [ + 0.00964441616088152, + 0.00979128573089838, + 0.005732160992920399, + 0.011011593043804169, + 0.007197191473096609, + 0.010527350939810276, + 0.008452075533568859, + 0.008277954533696175, + 0.013227961026132107, + 0.014739957638084888, + 0.010425927117466927, + 0.008771128952503204, + 0.012377220205962658, + 0.021047335118055344, + 0.005169608164578676, + 0.011178704909980297, + 0.009727753698825836, + 0.00949739571660757, + 0.008896942250430584, + 0.008711457252502441, + 0.0068773613311350346, + 0.01054687611758709, + 0.00896310806274414, + 0.010636968538165092, + 0.004222513176500797, + 0.004306471440941095, + 0.006467481143772602, + 0.004404084756970406, + 0.006442546378821135, + 0.006559473462402821, + 0.004554835148155689, + 0.005030757747590542, + 0.04626579210162163, + 0.1067134216427803, + 0.07926397025585175, + 0.045184120535850525, + 0.0192654337733984, + 0.012082011438906193, + 0.008103218860924244, + 0.012715938501060009, + 0.005535355303436518, + 0.00527105713263154, + 0.007420049514621496, + 0.004560116678476334, + 0.004101628437638283, + 0.00733328377828002, + 0.0034834998659789562, + 0.0047874655574560165, + 0.010038300417363644, + 0.0086710499599576, + 0.011814774014055729, + 0.026590589433908463, + 0.005945057608187199, + 0.004338146187365055, + 0.008328987285494804, + 0.006542084738612175, + 0.0018424740992486477, + 0.0013903381768614054, + 0.0020637742709368467, + 0.004635729361325502, + 0.0007446578238159418, + 0.001098738401196897, + 0.0015349853783845901, + 0.0011495004873722792, + 0.003217775374650955, + 0.005678270943462849, + 0.008022960275411606, + 0.005870763212442398, + 0.004690372385084629, + 0.003127043368294835, + 0.003161399159580469, + 0.003853941336274147, + 0.014516565017402172, + 0.01625269651412964, + 0.02386852167546749, + 0.08850044757127762, + 0.009548879228532314, + 0.007672318257391453, + 0.019618501886725426, + 0.03016608953475952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.04306533932685852 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419494766741991 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.01419494766741991, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419494766741991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09782947599887848, + 0.0800117626786232, + 0.08300900459289551, + 0.08516637235879898, + 0.44858527183532715, + 0.025186657905578613, + 0.04747379198670387, + 0.043887458741664886, + 0.0504612922668457, + 0.0383889377117157 + ], + "credit_map": [ + 0.012527507729828358, + 0.011680486612021923, + 0.010115851648151875, + 0.016672691330313683, + 0.008776038885116577, + 0.01327341329306364, + 0.01304003968834877, + 0.011743438430130482, + 0.008107139728963375, + 0.01064489409327507, + 0.009389521554112434, + 0.009783846326172352, + 0.010230313055217266, + 0.013516352511942387, + 0.004668854176998138, + 0.013670834712684155, + 0.010783720761537552, + 0.014109711162745953, + 0.008975240401923656, + 0.012406953610479832, + 0.007936069741845131, + 0.010739810764789581, + 0.008247319608926773, + 0.009810172021389008, + 0.00880726519972086, + 0.007287928834557533, + 0.01616833731532097, + 0.010429071262478828, + 0.008370664902031422, + 0.011303069069981575, + 0.009928391315042973, + 0.0128716342151165, + 0.04935907945036888, + 0.10284840315580368, + 0.08714200556278229, + 0.14335685968399048, + 0.017677823081612587, + 0.012387300841510296, + 0.016846997663378716, + 0.01896676793694496, + 0.004352365620434284, + 0.002459673210978508, + 0.0019059856422245502, + 0.0020052583422511816, + 0.0025366477202624083, + 0.0031669330783188343, + 0.003494091099128127, + 0.005265701562166214, + 0.005389812868088484, + 0.003347198013216257, + 0.007534965872764587, + 0.015193426050245762, + 0.0022663353011012077, + 0.0035179874394088984, + 0.0042097619734704494, + 0.006014304235577583, + 0.003866827581077814, + 0.00375903001986444, + 0.006194327492266893, + 0.016407756134867668, + 0.0019872079137712717, + 0.0027333435136824846, + 0.00538458488881588, + 0.003554377006366849, + 0.006523469462990761, + 0.003666608827188611, + 0.007622436620295048, + 0.010844144970178604, + 0.0032063026446849108, + 0.00548448134213686, + 0.004653475247323513, + 0.00846036896109581, + 0.003310206113383174, + 0.0028208335861563683, + 0.005317783448845148, + 0.01114928349852562, + 0.001950470032170415, + 0.004402008838951588, + 0.0034444769844412804, + 0.005993872880935669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.01419494766741991 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948815029114485 + ], + "opposite_logprob_tokens": [ + -5.130949020385742 + ], + "logprob_reward": -0.005948815029114485, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948815029114485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0648370161652565, + 0.04295627400279045, + 0.033445488661527634, + 0.022756090387701988, + 0.5232152342796326, + 0.02575508877635002, + 0.06494908034801483, + 0.0374089851975441, + 0.03297968953847885, + 0.1516970694065094 + ], + "credit_map": [ + 0.008841672912240028, + 0.006750187370926142, + 0.006484222132712603, + 0.009986883029341698, + 0.005354173015803099, + 0.012100047431886196, + 0.00843973271548748, + 0.006880100350826979, + 0.006209464278072119, + 0.0074586449190974236, + 0.004093223251402378, + 0.0031589814461767673, + 0.007339306175708771, + 0.007135146297514439, + 0.0023307211231440306, + 0.005230786744505167, + 0.004325307905673981, + 0.0049042729660868645, + 0.0059069073759019375, + 0.0029592139180749655, + 0.0028316620737314224, + 0.004910183604806662, + 0.0019356978591531515, + 0.005672244820743799, + 0.0029506601858884096, + 0.003192161675542593, + 0.004990041255950928, + 0.002941350219771266, + 0.0020165499299764633, + 0.002357391407713294, + 0.0018361177062615752, + 0.0024718197528272867, + 0.07760395109653473, + 0.1725439578294754, + 0.05772694945335388, + 0.14600493013858795, + 0.02586844190955162, + 0.014839507639408112, + 0.011443957686424255, + 0.017183514311909676, + 0.0032085441052913666, + 0.0027490060310810804, + 0.003625191980972886, + 0.005093273241072893, + 0.0021918443962931633, + 0.0032595607917755842, + 0.0022645378485322, + 0.003363128285855055, + 0.0069252257235348225, + 0.009954028762876987, + 0.015822241082787514, + 0.012496342882514, + 0.005287164356559515, + 0.005078176036477089, + 0.004311599303036928, + 0.005074304062873125, + 0.00396909611299634, + 0.005276860669255257, + 0.007021358236670494, + 0.00670523289591074, + 0.0036535344552248716, + 0.0031728518661111593, + 0.0049502309411764145, + 0.0026598216500133276, + 0.003303976496681571, + 0.007323845289647579, + 0.003001383738592267, + 0.00518265925347805, + 0.005740257445722818, + 0.002930316375568509, + 0.002239049645140767, + 0.003258202690631151, + 0.009105286560952663, + 0.015186464414000511, + 0.01739722676575184, + 0.05956583842635155, + 0.007425142917782068, + 0.007571532856673002, + 0.014066165313124657, + 0.021379416808485985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.005948815029114485 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This low-angle, static, extreme close-up, captured with the cinematic clarity of a Sony FX3 paired with a master lens, focuses on a vibrant blue plastic bottle cap. The cap, with its delicately molded, serrated edge, spins rapidly and hypnotically on its flat top, becoming a circular blur against the gritty texture of the dusty concrete floor of the empty garage. A low golden hour light streaks across the scene from beneath the closed garage door, illuminating the fine dust particles on the ground and casting long, dynamic shadows from the spinning object. As the rotational energy dissipates, the cap's perfect rotation degenerates into an increasingly erratic and unstable oscillation, its edge beginning to rhythmically collide with the concrete surface. Finally, as its momentum completely depletes, the cap suddenly tilts sideways with a faint plastic click, slides a short distance, traces a tiny, clean arc through the dust layer, and then falls into absolute stillness, its brief, vibrant life ending in the quiet solitude of the garage. The shallow depth of field casts the crumbling concrete and shadows in the distance into a soft, blurred bokeh, creating an atmosphere of focus, contemplation, and strange poignancy in the final moments of movement of mundane objects.", + "seed": 7707, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_0/index_8544_seed_7707_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02515231054276228, + "vlm_logprob": -0.02515231054276228, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.02515231054276228, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02854440174996853 + ], + "opposite_logprob_tokens": [ + -3.7785444259643555 + ], + "logprob_reward": -0.02854440174996853, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02854440174996853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1535981297492981, + 0.06331641227006912, + 0.03312818706035614, + 0.20308789610862732, + 0.03249609097838402, + 0.3143242597579956, + 0.048244062811136246, + 0.06487546861171722, + 0.07023126631975174, + 0.01669822633266449 + ], + "credit_map": [ + 0.02557203359901905, + 0.026508932933211327, + 0.013905154541134834, + 0.011677112430334091, + 0.024910978972911835, + 0.019764136523008347, + 0.009062618017196655, + 0.022197159007191658, + 0.014852235093712807, + 0.008894355036318302, + 0.0043251654133200645, + 0.008041732013225555, + 0.008485393598675728, + 0.009501876309514046, + 0.00439397431910038, + 0.004821673966944218, + 0.006868244614452124, + 0.004192209802567959, + 0.002138958778232336, + 0.0020091738551855087, + 0.006607084535062313, + 0.004773264285176992, + 0.004315270110964775, + 0.0022239796817302704, + 0.031248481944203377, + 0.021144744008779526, + 0.012721100822091103, + 0.01739748753607273, + 0.0367262065410614, + 0.029019856825470924, + 0.036866478621959686, + 0.017963500693440437, + 0.004807009361684322, + 0.0046893334947526455, + 0.001947411336004734, + 0.003138398751616478, + 0.006500146817415953, + 0.005129448603838682, + 0.003375590778887272, + 0.002908749273046851, + 0.018474865704774857, + 0.1326267272233963, + 0.004064788576215506, + 0.002651615533977747, + 0.010892089456319809, + 0.1383044719696045, + 0.003066955367103219, + 0.004242731258273125, + 0.005333414301276207, + 0.00825707521289587, + 0.0014784774975851178, + 0.002639718586578965, + 0.01593039371073246, + 0.007508167997002602, + 0.004657318349927664, + 0.0024394900538027287, + 0.007658182643353939, + 0.00940418429672718, + 0.005641876254230738, + 0.007942281663417816, + 0.0108614731580019, + 0.007413350511342287, + 0.010287648998200893, + 0.0056664724834263325, + 0.0057484726421535015, + 0.010618448257446289, + 0.005327863153070211, + 0.00643118703737855, + 0.0144085343927145, + 0.009481527842581272, + 0.01228877529501915, + 0.005926452577114105, + 0.001715059275738895, + 0.002577628707513213, + 0.0014536986127495766, + 0.0029658556450158358, + 0.00271252216771245, + 0.0024855020456016064, + 0.0014587163459509611, + 0.001329240039922297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.02854440174996853 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029830489307641983 + ], + "opposite_logprob_tokens": [ + -3.5298304557800293 + ], + "logprob_reward": -0.029830489307641983, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029830489307641983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22279196977615356, + 0.052285484969615936, + 0.02395297959446907, + 0.0714796930551529, + 0.058774854987859726, + 0.2589366137981415, + 0.1211407482624054, + 0.09273114055395126, + 0.05270233377814293, + 0.04520423710346222 + ], + "credit_map": [ + 0.03776182979345322, + 0.041327446699142456, + 0.015912113711237907, + 0.02118874527513981, + 0.03203178197145462, + 0.03258265182375908, + 0.01350488979369402, + 0.028482502326369286, + 0.009794036857783794, + 0.01249044481664896, + 0.0032319845631718636, + 0.006030994467437267, + 0.005711412988603115, + 0.008178159594535828, + 0.0036768922582268715, + 0.0031715580262243748, + 0.005368824116885662, + 0.0035244005266577005, + 0.001565989456139505, + 0.0015547986840829253, + 0.004309778101742268, + 0.0033312842715531588, + 0.002716662595048547, + 0.0015812406782060862, + 0.017551405355334282, + 0.012018907815217972, + 0.004398491699248552, + 0.003914960660040379, + 0.009202725253999233, + 0.011083129793405533, + 0.007551541086286306, + 0.005758529994636774, + 0.010810957290232182, + 0.010593651793897152, + 0.003413453232496977, + 0.004780699498951435, + 0.008689412847161293, + 0.00994114764034748, + 0.0056760357692837715, + 0.004869499243795872, + 0.016043823212385178, + 0.10387372225522995, + 0.00394462002441287, + 0.0036822280380874872, + 0.01344748679548502, + 0.1076694056391716, + 0.004558850079774857, + 0.005716471001505852, + 0.018067436292767525, + 0.020594261586666107, + 0.007348743733018637, + 0.004401984624564648, + 0.029139330610632896, + 0.026578692719340324, + 0.0079804677516222, + 0.007029832806438208, + 0.009203126654028893, + 0.016711143776774406, + 0.007192778401076794, + 0.00904148630797863, + 0.019264180213212967, + 0.014652029611170292, + 0.010018556378781796, + 0.006647835019975901, + 0.005412184167653322, + 0.008984415791928768, + 0.004034920129925013, + 0.004468308296054602, + 0.014455992728471756, + 0.00799165666103363, + 0.004166131839156151, + 0.003188719507306814, + 0.004476360976696014, + 0.00731790391728282, + 0.0034184784162789583, + 0.005023836623877287, + 0.010900796391069889, + 0.0064950366504490376, + 0.003944861702620983, + 0.003626964520663023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.029830489307641983 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867635294795036 + ], + "opposite_logprob_tokens": [ + -3.0486762523651123 + ], + "logprob_reward": -0.04867635294795036, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867635294795036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09242677688598633, + 0.07548397034406662, + 0.047799233347177505, + 0.03815595060586929, + 0.054115455597639084, + 0.4988613426685333, + 0.09991022944450378, + 0.04479968175292015, + 0.02533009462058544, + 0.02311723120510578 + ], + "credit_map": [ + 0.020648635923862457, + 0.01788533478975296, + 0.006603699643164873, + 0.007454820908606052, + 0.011406244710087776, + 0.013829524628818035, + 0.005146591924130917, + 0.009451929479837418, + 0.01827780343592167, + 0.015758447349071503, + 0.006226852536201477, + 0.005912925582379103, + 0.00540900370106101, + 0.01064181700348854, + 0.004317102953791618, + 0.008940013125538826, + 0.013566821813583374, + 0.008425445295870304, + 0.0040415190160274506, + 0.002665654756128788, + 0.004208288621157408, + 0.008318924345076084, + 0.002188261365517974, + 0.004384319763630629, + 0.00901241134852171, + 0.005109714809805155, + 0.004368199035525322, + 0.0019002142362296581, + 0.004663476254791021, + 0.00547388382256031, + 0.002616979880258441, + 0.005011072847992182, + 0.009053043089807034, + 0.007872370071709156, + 0.0057349675334990025, + 0.0038097696378827095, + 0.007132230792194605, + 0.007469352800399065, + 0.0038592698983848095, + 0.009184451773762703, + 0.034021519124507904, + 0.1966111809015274, + 0.013678140006959438, + 0.007019436918199062, + 0.021445250138640404, + 0.2026176154613495, + 0.007977633737027645, + 0.015490575693547726, + 0.02039165608584881, + 0.01505269855260849, + 0.007223076187074184, + 0.004971904680132866, + 0.02043127454817295, + 0.015582996420562267, + 0.007276535965502262, + 0.008980096317827702, + 0.006006788462400436, + 0.007486552000045776, + 0.0033996952697634697, + 0.0030237098690122366, + 0.010564970783889294, + 0.007185113616287708, + 0.003208532929420471, + 0.003924326039850712, + 0.0024364860728383064, + 0.0046041556634008884, + 0.0019119055941700935, + 0.0021103806793689728, + 0.005882713478058577, + 0.004043948370963335, + 0.002174680121243, + 0.0021658255718648434, + 0.002621648134663701, + 0.003335690125823021, + 0.001780746504664421, + 0.002741435309872031, + 0.00514198886230588, + 0.0032408470287919044, + 0.0020026303827762604, + 0.0022522462531924248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.04867635294795036 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633563131093979 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633563131093979, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633563131093979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16838189959526062, + 0.061376895755529404, + 0.05694577470421791, + 0.06109345331788063, + 0.10036569833755493, + 0.1918383240699768, + 0.09331026673316956, + 0.1627165973186493, + 0.07164344191551208, + 0.032327715307474136 + ], + "credit_map": [ + 0.026881158351898193, + 0.03929630294442177, + 0.009131308645009995, + 0.017744064331054688, + 0.02816585823893547, + 0.024463575333356857, + 0.009381226263940334, + 0.013318399898707867, + 0.014852388761937618, + 0.00799574051052332, + 0.0037889450322836637, + 0.007454422768205404, + 0.00616078544408083, + 0.009469128213822842, + 0.0045867073349654675, + 0.007068767677992582, + 0.01321924477815628, + 0.007087899371981621, + 0.005589358042925596, + 0.004870643839240074, + 0.005307790823280811, + 0.009876958094537258, + 0.003968380857259035, + 0.007025493774563074, + 0.013293668627738953, + 0.009053466841578484, + 0.004646824207156897, + 0.004861076828092337, + 0.006469128653407097, + 0.009777799248695374, + 0.006785097997635603, + 0.006206383928656578, + 0.014167445711791515, + 0.013121317140758038, + 0.006229120772331953, + 0.008728161454200745, + 0.022059990093111992, + 0.012334410101175308, + 0.012035781517624855, + 0.011689465492963791, + 0.013138515874743462, + 0.06792435050010681, + 0.005916509311646223, + 0.005764464847743511, + 0.012157008983194828, + 0.07334808260202408, + 0.006087074521929026, + 0.007502302061766386, + 0.013201077468693256, + 0.016005104407668114, + 0.005967286415398121, + 0.004857494030147791, + 0.019231000915169716, + 0.021474409848451614, + 0.007654992863535881, + 0.0049188872799277306, + 0.021864714100956917, + 0.027694573625922203, + 0.010865820571780205, + 0.016902608796954155, + 0.027895310893654823, + 0.027023376896977425, + 0.021083863452076912, + 0.00938631221652031, + 0.007315554656088352, + 0.010929429903626442, + 0.0036639985628426075, + 0.007274335250258446, + 0.01714569702744484, + 0.012743277475237846, + 0.008951012045145035, + 0.003620131639763713, + 0.0032486182171851397, + 0.0041508604772388935, + 0.0024535050615668297, + 0.0054574040696024895, + 0.005124197341501713, + 0.004687782377004623, + 0.0035816372837871313, + 0.0036237079184502363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.007633563131093979 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076745577156544 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.011076745577156544, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076745577156544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08630514144897461, + 0.02417922392487526, + 0.009283454157412052, + 0.02402857318520546, + 0.026672864332795143, + 0.7008655667304993, + 0.03141364827752113, + 0.051739584654569626, + 0.021014025434851646, + 0.024497902020812035 + ], + "credit_map": [ + 0.01345909759402275, + 0.013705679215490818, + 0.0090944804251194, + 0.007644550409168005, + 0.011316444724798203, + 0.011682317592203617, + 0.0053789555095136166, + 0.014023622497916222, + 0.0057242815382778645, + 0.00464083906263113, + 0.0017845649272203445, + 0.0016864306526258588, + 0.002565806033089757, + 0.003544181352481246, + 0.0013956435723230243, + 0.002837477019056678, + 0.0020352511201053858, + 0.0011029926827177405, + 0.0009679470676928759, + 0.0005164826288819313, + 0.0013470550766214728, + 0.0013731864746659994, + 0.0009399342234246433, + 0.0010006052907556295, + 0.006606586277484894, + 0.003875696798786521, + 0.0015058157732710242, + 0.0014162466395646334, + 0.0024952481035143137, + 0.0038363183848559856, + 0.0026490751188248396, + 0.0016435873694717884, + 0.004549666307866573, + 0.0040663788095116615, + 0.0027218784671276808, + 0.0014314183499664068, + 0.0032571793999522924, + 0.004606463015079498, + 0.001974146580323577, + 0.004065734799951315, + 0.04339371621608734, + 0.29656368494033813, + 0.010413865558803082, + 0.004749665502458811, + 0.02228754200041294, + 0.30889230966567993, + 0.004858244210481644, + 0.009706555865705013, + 0.0037042368203401566, + 0.0047930097207427025, + 0.002565166214480996, + 0.0014172685332596302, + 0.006066437810659409, + 0.007605813443660736, + 0.0021248445846140385, + 0.0031368755735456944, + 0.0058233728632330894, + 0.0075976005755364895, + 0.0036620041355490685, + 0.004290603566914797, + 0.010761361569166183, + 0.008745714090764523, + 0.0042821564711630344, + 0.006576772779226303, + 0.001848442479968071, + 0.003319219686090946, + 0.0015809315955266356, + 0.002227189252153039, + 0.005572461523115635, + 0.002578237559646368, + 0.0021372989285737276, + 0.0017502467380836606, + 0.002390584209933877, + 0.003961564041674137, + 0.0021363080013543367, + 0.003139518667012453, + 0.0054101902060210705, + 0.002777011366561055, + 0.0024924965109676123, + 0.0021902313455939293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.011076745577156544 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This low-angle, static, extreme close-up, captured with the cinematic clarity of a Sony FX3 paired with a master lens, focuses on a vibrant blue plastic bottle cap. The cap, with its delicately molded, serrated edge, spins rapidly and hypnotically on its flat top, becoming a circular blur against the gritty texture of the dusty concrete floor of the empty garage. A low golden hour light streaks across the scene from beneath the closed garage door, illuminating the fine dust particles on the ground and casting long, dynamic shadows from the spinning object. As the rotational energy dissipates, the cap's perfect rotation degenerates into an increasingly erratic and unstable oscillation, its edge beginning to rhythmically collide with the concrete surface. Finally, as its momentum completely depletes, the cap suddenly tilts sideways with a faint plastic click, slides a short distance, traces a tiny, clean arc through the dust layer, and then falls into absolute stillness, its brief, vibrant life ending in the quiet solitude of the garage. The shallow depth of field casts the crumbling concrete and shadows in the distance into a soft, blurred bokeh, creating an atmosphere of focus, contemplation, and strange poignancy in the final moments of movement of mundane objects.", + "seed": 7708, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_0/index_8544_seed_7708_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0388007553294301, + "vlm_logprob": -0.0388007553294301, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.0388007553294301, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009939698502421379 + ], + "opposite_logprob_tokens": [ + -5.009939670562744 + ], + "logprob_reward": -0.009939698502421379, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009939698502421379, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2798152565956116, + 0.29695606231689453, + 0.0946313887834549, + 0.053221553564071655, + 0.05743467062711716, + 0.05642314627766609, + 0.05302094295620918, + 0.026058267802000046, + 0.054873302578926086, + 0.027565501630306244 + ], + "credit_map": [ + 0.06898913532495499, + 0.021463381126523018, + 0.02273615449666977, + 0.021028928458690643, + 0.02345178835093975, + 0.08978431671857834, + 0.017190489917993546, + 0.015171046368777752, + 0.08225075900554657, + 0.011642077937722206, + 0.08565206080675125, + 0.02883095107972622, + 0.010131514631211758, + 0.04270118102431297, + 0.011990824714303017, + 0.02375667169690132, + 0.023142050951719284, + 0.00899303425103426, + 0.007062328979372978, + 0.005600540898740292, + 0.008546240627765656, + 0.030428187921643257, + 0.0059861084446311, + 0.004872895311564207, + 0.006107286550104618, + 0.006418892648071051, + 0.004843215458095074, + 0.007169546093791723, + 0.01043644081801176, + 0.0067721824161708355, + 0.007146109361201525, + 0.004327873233705759, + 0.009606472216546535, + 0.0113123944029212, + 0.006560430862009525, + 0.005371800158172846, + 0.0070791649632155895, + 0.008706181310117245, + 0.003945454489439726, + 0.004852768499404192, + 0.006151835434138775, + 0.008434053510427475, + 0.005134236067533493, + 0.006893265526741743, + 0.010846657678484917, + 0.0078099691309034824, + 0.0070939622819423676, + 0.004059161059558392, + 0.007400796748697758, + 0.012164248153567314, + 0.004347509704530239, + 0.004649212583899498, + 0.009886863641440868, + 0.007123954594135284, + 0.003700156696140766, + 0.0037481975741684437, + 0.0034936019219458103, + 0.005426076706498861, + 0.0016057823086157441, + 0.0029987962916493416, + 0.005140370689332485, + 0.0039436607621610165, + 0.0018998200539499521, + 0.0015501572052016854, + 0.005462643224745989, + 0.009795255027711391, + 0.005428561009466648, + 0.0069591146893799305, + 0.010348253883421421, + 0.007085789926350117, + 0.00579578522592783, + 0.0039978958666324615, + 0.003432903904467821, + 0.0034206395503133535, + 0.0034127540420740843, + 0.003457524348050356, + 0.0038912715390324593, + 0.0045586624182760715, + 0.0030475701205432415, + 0.0023441738449037075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.009939698502421379 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381232351064682 + ], + "opposite_logprob_tokens": [ + -3.28812313079834 + ], + "logprob_reward": -0.0381232351064682, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381232351064682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20908069610595703, + 0.19202597439289093, + 0.09226242452859879, + 0.06743679195642471, + 0.08072202652692795, + 0.06052979454398155, + 0.15298958122730255, + 0.059216491878032684, + 0.03990793600678444, + 0.04582827538251877 + ], + "credit_map": [ + 0.03702143207192421, + 0.01473479438573122, + 0.044564589858055115, + 0.017761746421456337, + 0.019324734807014465, + 0.03818986564874649, + 0.017918894067406654, + 0.01956464722752571, + 0.04516987130045891, + 0.014979815110564232, + 0.05938496068120003, + 0.01588405855000019, + 0.010683265514671803, + 0.0244849044829607, + 0.008793010376393795, + 0.012646092101931572, + 0.022305959835648537, + 0.008196107111871243, + 0.009677238762378693, + 0.006819210946559906, + 0.006194944493472576, + 0.02782161347568035, + 0.004885685630142689, + 0.006361669395118952, + 0.013622986152768135, + 0.00666720001026988, + 0.008363167755305767, + 0.006227015983313322, + 0.011344482190907001, + 0.01080926600843668, + 0.006071341224014759, + 0.0043313284404575825, + 0.012042362242937088, + 0.01722172275185585, + 0.007607665378600359, + 0.007338216993957758, + 0.010121762752532959, + 0.014544988051056862, + 0.006608664058148861, + 0.005236642900854349, + 0.00803409330546856, + 0.010506534948945045, + 0.00666520930826664, + 0.006666379980742931, + 0.008144252933561802, + 0.00881950557231903, + 0.005384123884141445, + 0.006309689953923225, + 0.016500351950526237, + 0.03702418878674507, + 0.0104925986379385, + 0.01233084686100483, + 0.03509062901139259, + 0.022686151787638664, + 0.011326013132929802, + 0.007538797799497843, + 0.00703574949875474, + 0.014156699180603027, + 0.006159686483442783, + 0.005087228026241064, + 0.012191373854875565, + 0.0076919179409742355, + 0.003716732608154416, + 0.0031771028880029917, + 0.005191680509597063, + 0.005979812238365412, + 0.00584625406190753, + 0.004641937557607889, + 0.005751349963247776, + 0.00491706095635891, + 0.003538259770721197, + 0.004041582345962524, + 0.004997947718948126, + 0.011160115711390972, + 0.004993447568267584, + 0.003699934110045433, + 0.009302042424678802, + 0.005908413324505091, + 0.002567343646660447, + 0.0031990318093448877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0381232351064682 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295587569475174 + ], + "opposite_logprob_tokens": [ + -2.2379558086395264 + ], + "logprob_reward": -0.11295587569475174, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295587569475174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1551513373851776, + 0.11670204997062683, + 0.16320638358592987, + 0.07877881824970245, + 0.18422144651412964, + 0.1780746728181839, + 0.041060611605644226, + 0.027370145544409752, + 0.025941835716366768, + 0.02949274517595768 + ], + "credit_map": [ + 0.02985846810042858, + 0.013073612004518509, + 0.02049311064183712, + 0.016723765060305595, + 0.01630748249590397, + 0.02997250109910965, + 0.017777377739548683, + 0.010945008136332035, + 0.018233077600598335, + 0.009515085257589817, + 0.011837209574878216, + 0.020373670384287834, + 0.007661521900445223, + 0.01867544651031494, + 0.011395598761737347, + 0.019010422751307487, + 0.03888406231999397, + 0.015421665273606777, + 0.01877688243985176, + 0.009683659300208092, + 0.00857448298484087, + 0.054685186594724655, + 0.007303330581635237, + 0.009877104312181473, + 0.018336741253733635, + 0.008238198235630989, + 0.011679074726998806, + 0.007468970958143473, + 0.007412934675812721, + 0.012438968755304813, + 0.004987558349967003, + 0.008216354995965958, + 0.029726536944508553, + 0.05498427897691727, + 0.018661173060536385, + 0.008278957568109035, + 0.018555041402578354, + 0.03568815812468529, + 0.008690758608281612, + 0.009636539965867996, + 0.020254600793123245, + 0.0383710041642189, + 0.016678275540471077, + 0.013670907355844975, + 0.03899787366390228, + 0.02352323941886425, + 0.011863045394420624, + 0.014715708792209625, + 0.004897809587419033, + 0.008121254853904247, + 0.004001757595688105, + 0.003461815882474184, + 0.008522260934114456, + 0.006419915705919266, + 0.0028042932972311974, + 0.0028315000236034393, + 0.003145421389490366, + 0.005788534879684448, + 0.0018986089853569865, + 0.001985829556360841, + 0.006917711813002825, + 0.003965028561651707, + 0.001845251303166151, + 0.0018237559124827385, + 0.0028295908123254776, + 0.006547922734171152, + 0.002058088080957532, + 0.0018743212567642331, + 0.005711290519684553, + 0.003529717680066824, + 0.0017139887204393744, + 0.0016769147478044033, + 0.003305980935692787, + 0.006678292993456125, + 0.0026768234092742205, + 0.002934733871370554, + 0.0062775202095508575, + 0.00391450896859169, + 0.0018330836901441216, + 0.0018717993516474962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.11295587569475174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029793230816721916 + ], + "opposite_logprob_tokens": [ + -3.5297932624816895 + ], + "logprob_reward": -0.029793230816721916, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029793230816721916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12885601818561554, + 0.3707227110862732, + 0.057502735406160355, + 0.10188909620046616, + 0.06353834271430969, + 0.1210629791021347, + 0.08906690031290054, + 0.027409961447119713, + 0.028967173770070076, + 0.01098417118191719 + ], + "credit_map": [ + 0.027892177924513817, + 0.013247809372842312, + 0.013818355277180672, + 0.013451503589749336, + 0.011269844137132168, + 0.029527045786380768, + 0.010848887264728546, + 0.008800379931926727, + 0.10141992568969727, + 0.01540337223559618, + 0.10171128064393997, + 0.03903614357113838, + 0.010783687233924866, + 0.05347582697868347, + 0.014704497531056404, + 0.03418790549039841, + 0.008666162379086018, + 0.007029922679066658, + 0.006400838028639555, + 0.004420835990458727, + 0.005384793505072594, + 0.01594814658164978, + 0.005548682529479265, + 0.00410333601757884, + 0.014767879620194435, + 0.015086060389876366, + 0.008753623813390732, + 0.00985549297183752, + 0.01708763837814331, + 0.01891377381980419, + 0.008873255923390388, + 0.008551354520022869, + 0.007704323623329401, + 0.01124993059784174, + 0.008075169287621975, + 0.006676034536212683, + 0.008323585614562035, + 0.009352084249258041, + 0.007449535187333822, + 0.004707672633230686, + 0.011894960887730122, + 0.025551563128829002, + 0.010188765823841095, + 0.012355406768620014, + 0.021770529448986053, + 0.0169094055891037, + 0.012354646809399128, + 0.010037675499916077, + 0.011142285540699959, + 0.019274486228823662, + 0.006546516437083483, + 0.00738803343847394, + 0.014907483011484146, + 0.015856940299272537, + 0.006513645872473717, + 0.007437501102685928, + 0.003374427789822221, + 0.006936779711395502, + 0.0016677678795531392, + 0.0020284452475607395, + 0.005896870978176594, + 0.003955773543566465, + 0.0017984624719247222, + 0.0017514274222776294, + 0.003329848637804389, + 0.007075977511703968, + 0.0031061104964464903, + 0.002679774770513177, + 0.0057839020155370235, + 0.0035743850748986006, + 0.001807349850423634, + 0.0016098207561299205, + 0.0013758644927293062, + 0.0016805521445348859, + 0.0015926385531201959, + 0.0010294192470610142, + 0.0018835656810551882, + 0.001454332028515637, + 0.0008465800201520324, + 0.0011212168028578162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.029793230816721916 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003191736526787281 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.003191736526787281, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003191736526787281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08882385492324829, + 0.1550392508506775, + 0.2509210705757141, + 0.039690226316452026, + 0.1936269849538803, + 0.09013429284095764, + 0.030723104253411293, + 0.04434666037559509, + 0.05738808214664459, + 0.0493064820766449 + ], + "credit_map": [ + 0.01329243928194046, + 0.007961120456457138, + 0.013771693222224712, + 0.008085162378847599, + 0.010005551390349865, + 0.019029822200536728, + 0.008801737800240517, + 0.007876323536038399, + 0.030032820999622345, + 0.010669345036149025, + 0.02674134261906147, + 0.021211793646216393, + 0.008829126134514809, + 0.022014044225215912, + 0.012382391840219498, + 0.023158395662903786, + 0.05903860181570053, + 0.02553187869489193, + 0.02144935354590416, + 0.015474366024136543, + 0.014836536720395088, + 0.08923743665218353, + 0.01248797308653593, + 0.012864925898611546, + 0.0072548058815300465, + 0.004318558145314455, + 0.004324445500969887, + 0.004235157743096352, + 0.005315267480909824, + 0.005914630368351936, + 0.004017374012619257, + 0.004309988114982843, + 0.029913662001490593, + 0.04665669798851013, + 0.017531979829072952, + 0.012089143507182598, + 0.02314513549208641, + 0.03999389708042145, + 0.011314825154840946, + 0.012981650419533253, + 0.011658994480967522, + 0.013048372231423855, + 0.012019605375826359, + 0.007719327695667744, + 0.016963200643658638, + 0.012543074786663055, + 0.007588087115436792, + 0.00859362818300724, + 0.004818701185286045, + 0.004865127149969339, + 0.0033878437243402004, + 0.0028785134200006723, + 0.00465046102181077, + 0.005022426135838032, + 0.0021453206427395344, + 0.002954712137579918, + 0.0060302200727164745, + 0.007852384820580482, + 0.006934269797056913, + 0.003726224647834897, + 0.007996702566742897, + 0.005690149962902069, + 0.002721385797485709, + 0.003395321313291788, + 0.005413517821580172, + 0.01810610108077526, + 0.003304513171315193, + 0.004026305396109819, + 0.01464429497718811, + 0.007552820723503828, + 0.0019990834407508373, + 0.0023414441384375095, + 0.005350995343178511, + 0.010971330106258392, + 0.006399051286280155, + 0.004278603475540876, + 0.010034634731709957, + 0.0059363218024373055, + 0.0026826311368495226, + 0.0036529190838336945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.003191736526787281 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This low-angle, static, extreme close-up, captured with the cinematic clarity of a Sony FX3 paired with a master lens, focuses on a vibrant blue plastic bottle cap. The cap, with its delicately molded, serrated edge, spins rapidly and hypnotically on its flat top, becoming a circular blur against the gritty texture of the dusty concrete floor of the empty garage. A low golden hour light streaks across the scene from beneath the closed garage door, illuminating the fine dust particles on the ground and casting long, dynamic shadows from the spinning object. As the rotational energy dissipates, the cap's perfect rotation degenerates into an increasingly erratic and unstable oscillation, its edge beginning to rhythmically collide with the concrete surface. Finally, as its momentum completely depletes, the cap suddenly tilts sideways with a faint plastic click, slides a short distance, traces a tiny, clean arc through the dust layer, and then falls into absolute stillness, its brief, vibrant life ending in the quiet solitude of the garage. The shallow depth of field casts the crumbling concrete and shadows in the distance into a soft, blurred bokeh, creating an atmosphere of focus, contemplation, and strange poignancy in the final moments of movement of mundane objects.", + "seed": 7709, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_0/index_8544_seed_7709_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03552179012913257, + "vlm_logprob": -0.03552179012913257, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03552179012913257, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01971389912068844 + ], + "opposite_logprob_tokens": [ + -4.144713878631592 + ], + "logprob_reward": -0.01971389912068844, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01971389912068844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30405089259147644, + 0.09899482131004333, + 0.037185195833444595, + 0.033702678978443146, + 0.07135683298110962, + 0.06833784282207489, + 0.08729992806911469, + 0.14211341738700867, + 0.09620167315006256, + 0.06075683981180191 + ], + "credit_map": [ + 0.03194170445203781, + 0.0286936666816473, + 0.04078766331076622, + 0.05062020942568779, + 0.021316953003406525, + 0.06302618235349655, + 0.029018867760896683, + 0.03864561766386032, + 0.007974997162818909, + 0.010039777494966984, + 0.010392765514552593, + 0.014989416114985943, + 0.010491565801203251, + 0.012080607935786247, + 0.016828035935759544, + 0.01619764231145382, + 0.008687031455338001, + 0.0026085502468049526, + 0.003604200901463628, + 0.0034055281430482864, + 0.003094529267400503, + 0.006510360166430473, + 0.004416265524923801, + 0.004858724772930145, + 0.006209013983607292, + 0.0034683458507061005, + 0.003311607288196683, + 0.003584967227652669, + 0.004227464087307453, + 0.005134756676852703, + 0.004109174478799105, + 0.0036573470570147038, + 0.012789511121809483, + 0.00919121503829956, + 0.007048322353512049, + 0.006969102192670107, + 0.009094314649701118, + 0.011667211540043354, + 0.008408148773014545, + 0.006188997533172369, + 0.008242085576057434, + 0.006623796187341213, + 0.006306922994554043, + 0.008833571337163448, + 0.009052547626197338, + 0.012361479923129082, + 0.008327897638082504, + 0.008589530363678932, + 0.008491450920701027, + 0.010789208114147186, + 0.008611774072051048, + 0.010854759253561497, + 0.013014945201575756, + 0.012832609005272388, + 0.013386343605816364, + 0.009318829514086246, + 0.016213804483413696, + 0.014441738836467266, + 0.015669673681259155, + 0.02501751482486725, + 0.020355602726340294, + 0.020121751353144646, + 0.016265014186501503, + 0.014028285630047321, + 0.012570569291710854, + 0.009969285689294338, + 0.007439561188220978, + 0.014782470650970936, + 0.010627576150000095, + 0.011609969660639763, + 0.016715390607714653, + 0.012486832216382027, + 0.008230313658714294, + 0.005663603078573942, + 0.005965277552604675, + 0.00592726981267333, + 0.008778752759099007, + 0.009664761833846569, + 0.007470295764505863, + 0.009056556038558483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.01971389912068844 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048695314675569534 + ], + "opposite_logprob_tokens": [ + -3.0486953258514404 + ], + "logprob_reward": -0.048695314675569534, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048695314675569534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21739482879638672, + 0.10063432902097702, + 0.05057458579540253, + 0.04757861793041229, + 0.14802412688732147, + 0.11441904306411743, + 0.07857023924589157, + 0.10478310286998749, + 0.09193040430545807, + 0.04609077423810959 + ], + "credit_map": [ + 0.03296902775764465, + 0.02155306003987789, + 0.03086669184267521, + 0.03801070898771286, + 0.016285747289657593, + 0.04200029745697975, + 0.016645876690745354, + 0.019063401967287064, + 0.013952448964118958, + 0.009332003071904182, + 0.014843752607703209, + 0.011573524214327335, + 0.00761151360347867, + 0.013126901350915432, + 0.007729555945843458, + 0.02246461994946003, + 0.009447654709219933, + 0.0046657477505505085, + 0.006165672093629837, + 0.007061419077217579, + 0.0037517014425247908, + 0.006873032543808222, + 0.005615818779915571, + 0.0069935377687215805, + 0.012758440338075161, + 0.004638211335986853, + 0.00409192917868495, + 0.004186843987554312, + 0.004117944743484259, + 0.008498343639075756, + 0.0047658332623541355, + 0.004521064925938845, + 0.01367124728858471, + 0.027238771319389343, + 0.011924284510314465, + 0.01979924365878105, + 0.020485345274209976, + 0.02076089009642601, + 0.021383289247751236, + 0.012761056423187256, + 0.0203301552683115, + 0.010451365262269974, + 0.008921723812818527, + 0.013070284388959408, + 0.014851746149361134, + 0.02300083450973034, + 0.012581388466060162, + 0.011211542412638664, + 0.010907142423093319, + 0.009078004397451878, + 0.00772069999948144, + 0.009239123202860355, + 0.01149156503379345, + 0.009598278440535069, + 0.011063406243920326, + 0.009472017176449299, + 0.014109227806329727, + 0.008881858550012112, + 0.007456377614289522, + 0.011071987450122833, + 0.01579691655933857, + 0.02150314301252365, + 0.01426287367939949, + 0.011700709350407124, + 0.009881879203021526, + 0.00818647164851427, + 0.006712855771183968, + 0.011283909901976585, + 0.013187614269554615, + 0.015848208218812943, + 0.013736204244196415, + 0.013093252666294575, + 0.005933685228228569, + 0.004831974860280752, + 0.006585104390978813, + 0.005052295047789812, + 0.008817829191684723, + 0.0067885443568229675, + 0.0030668664257973433, + 0.00501446845009923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.048695314675569534 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906956017017365 + ], + "opposite_logprob_tokens": [ + -2.4640696048736572 + ], + "logprob_reward": -0.08906956017017365, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906956017017365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1424751877784729, + 0.07329794764518738, + 0.02823789417743683, + 0.0804065689444542, + 0.09760883450508118, + 0.029086196795105934, + 0.0769542008638382, + 0.04471556469798088, + 0.03253403306007385, + 0.3946835696697235 + ], + "credit_map": [ + 0.01642502285540104, + 0.024215780198574066, + 0.02276710234582424, + 0.018840091302990913, + 0.014080717228353024, + 0.023479264229536057, + 0.010711378417909145, + 0.01195583213120699, + 0.00951923057436943, + 0.009755641222000122, + 0.009917089715600014, + 0.010753536596894264, + 0.005514961667358875, + 0.011320804245769978, + 0.007028819061815739, + 0.009487860836088657, + 0.005043258424848318, + 0.002809553872793913, + 0.003984729293733835, + 0.003876051865518093, + 0.0026104310527443886, + 0.0038747512735426426, + 0.002629316644743085, + 0.003409800585359335, + 0.018898794427514076, + 0.008011545054614544, + 0.009121092967689037, + 0.0070559242740273476, + 0.005967980716377497, + 0.014438859187066555, + 0.005434634629637003, + 0.011477742344141006, + 0.012515970505774021, + 0.010350316762924194, + 0.015921561047434807, + 0.008057359606027603, + 0.011230039410293102, + 0.013782749883830547, + 0.011931370943784714, + 0.01381946261972189, + 0.005766375456005335, + 0.0028654756024479866, + 0.0029659802094101906, + 0.002268946962431073, + 0.0030137738212943077, + 0.004703094717115164, + 0.0024502878077328205, + 0.005052262917160988, + 0.010218728333711624, + 0.008610027842223644, + 0.007334826979786158, + 0.009796739555895329, + 0.007881350815296173, + 0.012105755507946014, + 0.01350473053753376, + 0.007502042688429356, + 0.006923126522451639, + 0.003162883687764406, + 0.0034557159524410963, + 0.004960446152836084, + 0.006020434200763702, + 0.008264702744781971, + 0.006028117146342993, + 0.005900137592107058, + 0.0029062279500067234, + 0.0036849272437393665, + 0.0029633240774273872, + 0.005183297209441662, + 0.004262400325387716, + 0.0038226665928959846, + 0.004436696879565716, + 0.005274494178593159, + 0.06673264503479004, + 0.037892282009124756, + 0.029731079936027527, + 0.03449278324842453, + 0.0718468725681305, + 0.06399063020944595, + 0.03380722925066948, + 0.056190039962530136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.08906956017017365 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818847469985485 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.01818847469985485, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818847469985485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08168061822652817, + 0.07785002887248993, + 0.05753621086478233, + 0.08166517317295074, + 0.12090356647968292, + 0.12312870472669601, + 0.09690534323453903, + 0.10430953651666641, + 0.09660764783620834, + 0.1594132035970688 + ], + "credit_map": [ + 0.01250183954834938, + 0.007155984174460173, + 0.011923630721867085, + 0.009288350120186806, + 0.008889098651707172, + 0.012199854478240013, + 0.0100851533934474, + 0.009636706672608852, + 0.007785934954881668, + 0.007018928416073322, + 0.00822930596768856, + 0.008881296031177044, + 0.012632648460566998, + 0.009387602098286152, + 0.009982486255466938, + 0.013931822963058949, + 0.010139293037354946, + 0.003987565636634827, + 0.0064778635278344154, + 0.009981286711990833, + 0.006253599189221859, + 0.007692247163504362, + 0.0077964975498616695, + 0.005207856651395559, + 0.015508527867496014, + 0.00873500294983387, + 0.007662898860871792, + 0.009187147952616215, + 0.008373172022402287, + 0.013398149982094765, + 0.009280461817979813, + 0.009519817307591438, + 0.01530587300658226, + 0.022201301530003548, + 0.011023767292499542, + 0.015673091635107994, + 0.015303920954465866, + 0.015883669257164, + 0.012540106661617756, + 0.012971839867532253, + 0.012731106020510197, + 0.013866736553609371, + 0.01112501323223114, + 0.017752841114997864, + 0.014737742021679878, + 0.020404063165187836, + 0.01651066727936268, + 0.016000540927052498, + 0.008289279416203499, + 0.00655796192586422, + 0.011567281559109688, + 0.00971406139433384, + 0.012302936054766178, + 0.011844472959637642, + 0.020816024392843246, + 0.015813324600458145, + 0.014023294672369957, + 0.008736548945307732, + 0.00564202107489109, + 0.011345270089805126, + 0.016443107277154922, + 0.02213362045586109, + 0.012669296003878117, + 0.0133163807913661, + 0.010925130918622017, + 0.008878196589648724, + 0.010145687498152256, + 0.01933910697698593, + 0.011010704562067986, + 0.012332894839346409, + 0.013275732286274433, + 0.01070018857717514, + 0.034762702882289886, + 0.014531721360981464, + 0.01698783040046692, + 0.015586313791573048, + 0.026240842416882515, + 0.026480719447135925, + 0.010951665230095387, + 0.013871394097805023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.01818847469985485 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant blue plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle a low-angle, static, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show a low golden hour light source coming from beneath a closed garage door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap transition from rapid spinning to erratic oscillation, then slide and fall into stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a gritty, dusty concrete floor with shallow depth of field creating blurred bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019417019793763757 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019417019793763757, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019417019793763757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2585361897945404, + 0.09015414863824844, + 0.052837055176496506, + 0.03771121799945831, + 0.08007613569498062, + 0.06708499789237976, + 0.08522441983222961, + 0.07026804983615875, + 0.05775255709886551, + 0.20035520195960999 + ], + "credit_map": [ + 0.031534790992736816, + 0.02352500520646572, + 0.03623993694782257, + 0.05223610997200012, + 0.016898896545171738, + 0.060332704335451126, + 0.01901833899319172, + 0.018750423565506935, + 0.010744315572082996, + 0.010925102047622204, + 0.011481723748147488, + 0.01264894288033247, + 0.007700771559029818, + 0.011557619087398052, + 0.008794896304607391, + 0.016300776973366737, + 0.010471292771399021, + 0.00418276758864522, + 0.005458586383610964, + 0.0047235991805791855, + 0.004929685965180397, + 0.008706016466021538, + 0.005286321975290775, + 0.009078784845769405, + 0.006028248928487301, + 0.00400103023275733, + 0.004699109122157097, + 0.0030327383428812027, + 0.003823351813480258, + 0.005842975340783596, + 0.0037882188335061073, + 0.00649554468691349, + 0.013059867545962334, + 0.005658210255205631, + 0.010297611355781555, + 0.00742768170312047, + 0.008668043650686741, + 0.011197434738278389, + 0.01080261543393135, + 0.012964674271643162, + 0.007796674035489559, + 0.007135561667382717, + 0.008483226411044598, + 0.006458453834056854, + 0.005904764402657747, + 0.00848833192139864, + 0.0067521193996071815, + 0.016065867617726326, + 0.011555245146155357, + 0.010507749393582344, + 0.008161850273609161, + 0.010857674293220043, + 0.008988677524030209, + 0.01335484255105257, + 0.014127363450825214, + 0.007671014871448278, + 0.009415862150490284, + 0.008602163754403591, + 0.005609304178506136, + 0.008621934801340103, + 0.009901808574795723, + 0.010912399739027023, + 0.009072237648069859, + 0.008132336661219597, + 0.006610752083361149, + 0.004405306652188301, + 0.0051216138526797295, + 0.00880627240985632, + 0.006606596522033215, + 0.006575769279152155, + 0.010375234298408031, + 0.00925101526081562, + 0.03775658831000328, + 0.01892751455307007, + 0.011988679878413677, + 0.0114118168130517, + 0.03805581107735634, + 0.03692413866519928, + 0.019024088978767395, + 0.026266558095812798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0019417019793763757 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_1/index_tensor([807, 807])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_1/index_tensor([807, 807])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b815d9c8cf9eb9fa5511eefeed40a4c4c89b7d5e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_1/index_tensor([807, 807])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, shot from a slightly overhead angle, slowly and almost imperceptibly zooms in, focusing on a cold, white metal bed. Lying on the bed is a patient, his face as pale as paper, bloodless. Fine beads of cold sweat, dripping from his forehead with agony, gleam faintly under the cold, white light of the ward's overhead lamp. His thinning, dark hair is sticky to his temples, clung to his eyes with sweat. His eyes are half-closed, his gaze unfocused, and his lips, dry and cracked, struggle to open. Each breath is a struggle, a short, hoarse sound. His chest, shrouded by a loose, washed-out, light-blue striped hospital gown, rises and falls with alarming force, his collarbone clearly visible with each rise. The pure white sheets are crumpled by his unconscious struggles. The entire scene is shrouded in a uniform and cold fluorescent light, and the main colors are white, light blue and sickly gray skin, creating a tense atmosphere of extreme depression, pain and the fragility of life.", + "seed": 31582, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_1/index_807_seed_31582_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.20061915861442686, + "vlm_logprob": -0.20061915861442686, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.20061915861442686, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0044373623095452785 + ], + "opposite_logprob_tokens": [ + -7.129437446594238 + ], + "logprob_reward": -0.0044373623095452785, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0044373623095452785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.705756425857544, + 0.06079151853919029, + 0.022971300408244133, + 0.03752632439136505, + 0.059428319334983826, + 0.04127732664346695, + 0.012657093815505505, + 0.025370946153998375, + 0.016674982383847237, + 0.017545800656080246 + ], + "credit_map": [ + 0.034582171589136124, + 0.033051278442144394, + 0.2835708558559418, + 0.03428288549184799, + 0.05452853813767433, + 0.03765692189335823, + 0.023582948371767998, + 0.20450077950954437, + 0.004035287071019411, + 0.0056514195166528225, + 0.005017337389290333, + 0.00694493530318141, + 0.012837537564337254, + 0.005517689976841211, + 0.006534690503031015, + 0.014252620749175549, + 0.0017174542881548405, + 0.0031781799625605345, + 0.0019029334653168917, + 0.002596202539280057, + 0.0026852604933083057, + 0.0035495583433657885, + 0.0029063818510621786, + 0.004435329232364893, + 0.006211448926478624, + 0.003768817288801074, + 0.0076186261139810085, + 0.0034081449266523123, + 0.004569315817207098, + 0.004829869605600834, + 0.003073456697165966, + 0.004046646878123283, + 0.006467847153544426, + 0.007775662466883659, + 0.0050001684576272964, + 0.008869247511029243, + 0.006095645017921925, + 0.009595714509487152, + 0.006255298387259245, + 0.009368736296892166, + 0.0036712440196424723, + 0.0069852229207754135, + 0.005227650515735149, + 0.006112896371632814, + 0.004280948080122471, + 0.0047319866716861725, + 0.006236974149942398, + 0.004030405543744564, + 0.0013533680466935039, + 0.0014485681895166636, + 0.0016865796642377973, + 0.0013344120234251022, + 0.0016259625554084778, + 0.0022646940778940916, + 0.0015149739338085055, + 0.0014285349752753973, + 0.002077010925859213, + 0.002270747907459736, + 0.003060251474380493, + 0.0035776991862803698, + 0.0030815533827990294, + 0.005038098432123661, + 0.0030687698163092136, + 0.0031968129333108664, + 0.00168539525475353, + 0.002489345381036401, + 0.001813013688661158, + 0.002923890482634306, + 0.0019410266540944576, + 0.002375175943598151, + 0.0017842311644926667, + 0.0016629042802378535, + 0.0024725133553147316, + 0.0022793610114604235, + 0.002147312741726637, + 0.0020543127320706844, + 0.0021463565062731504, + 0.0030041520949453115, + 0.001800072262994945, + 0.0016417214646935463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0044373623095452785 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983447551727295 + ], + "opposite_logprob_tokens": [ + -0.5233447551727295 + ], + "logprob_reward": -0.8983447551727295, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983447551727295, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.8524541258811951, + 0.054733987897634506, + 0.017801061272621155, + 0.010910655371844769, + 0.009653382003307343, + 0.021463695913553238, + 0.008407794870436192, + 0.00578411016613245, + 0.0037275892682373524, + 0.015063620172441006 + ], + "credit_map": [ + 0.03653911501169205, + 0.027815941721200943, + 0.35703185200691223, + 0.0439421720802784, + 0.05363987013697624, + 0.04497136175632477, + 0.027214163914322853, + 0.26129958033561707, + 0.003145362948998809, + 0.004861996043473482, + 0.004076583310961723, + 0.005061041563749313, + 0.009989526122808456, + 0.008031509816646576, + 0.0072125717997550964, + 0.012355390004813671, + 0.0018392637139186263, + 0.0017904669512063265, + 0.002159848576411605, + 0.00163172441534698, + 0.0019829522352665663, + 0.002875816309824586, + 0.0028377610724419355, + 0.002683226717635989, + 0.001069220481440425, + 0.0011914519127458334, + 0.0019776667468249798, + 0.0010958298807963729, + 0.0011404860997572541, + 0.0014382447116076946, + 0.0012733546318486333, + 0.0017244009068235755, + 0.0008185772458091378, + 0.0016893389401957393, + 0.0010734325041994452, + 0.0013304769527167082, + 0.000988350948318839, + 0.001440665335394442, + 0.0010603166883811355, + 0.0012522219913080335, + 0.002200598828494549, + 0.003116212785243988, + 0.0026001092046499252, + 0.002067773835733533, + 0.002324942732229829, + 0.003393030958250165, + 0.00286793801933527, + 0.00289308768697083, + 0.0007403715862892568, + 0.0012042926391586661, + 0.0009292413014918566, + 0.000891467381734401, + 0.000992465065792203, + 0.0016203998820856214, + 0.000954833987634629, + 0.0010747227352112532, + 0.0004532376187853515, + 0.0006226171972230077, + 0.0006366436718963087, + 0.0006202898803167045, + 0.0008884896524250507, + 0.0011996497632935643, + 0.0005075792432762682, + 0.0008556030807085335, + 0.0005132397054694593, + 0.0004907884867861867, + 0.0006014188984408975, + 0.0005896987277083099, + 0.0003874710528180003, + 0.0004463778168428689, + 0.00038841768400743604, + 0.00031017648871056736, + 0.001863409299403429, + 0.00214669038541615, + 0.0026784141082316637, + 0.0017776628956198692, + 0.0018384207505732775, + 0.001934191444888711, + 0.0015519896987825632, + 0.00127284019254148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.8983447551727295 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196240343153477 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.014196240343153477, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196240343153477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7085732221603394, + 0.13497741520404816, + 0.04357976093888283, + 0.02238881215453148, + 0.02056661993265152, + 0.017360107973217964, + 0.014154670760035515, + 0.007772673852741718, + 0.005299191921949387, + 0.025327447801828384 + ], + "credit_map": [ + 0.031379517167806625, + 0.041734661906957626, + 0.27477753162384033, + 0.045201271772384644, + 0.05319202318787575, + 0.03743359446525574, + 0.02572707086801529, + 0.19912762939929962, + 0.009385691955685616, + 0.009950920939445496, + 0.009918121621012688, + 0.015034244395792484, + 0.0267062708735466, + 0.017286472022533417, + 0.011055164970457554, + 0.03564053773880005, + 0.0033372847829014063, + 0.0070984032936394215, + 0.004632836673408747, + 0.003913009539246559, + 0.005130627192556858, + 0.007756530307233334, + 0.0033948912750929594, + 0.00831617508083582, + 0.0020071386825293303, + 0.0023339216131716967, + 0.0039814854972064495, + 0.0024561539757996798, + 0.0023749000392854214, + 0.002866238821297884, + 0.0022274970542639494, + 0.004141475539654493, + 0.0018804840510711074, + 0.003302142722532153, + 0.002656694734469056, + 0.002901543630287051, + 0.0021726482082158327, + 0.002626454457640648, + 0.0021805104333907366, + 0.002846141578629613, + 0.0017582324799150229, + 0.00236911466345191, + 0.0030975600238889456, + 0.0018765900749713182, + 0.0020645270124077797, + 0.002423225436359644, + 0.0019991814624518156, + 0.00177167565561831, + 0.001193941105157137, + 0.0018963515758514404, + 0.0018745331326499581, + 0.0015600607730448246, + 0.0016123810783028603, + 0.002387481741607189, + 0.0015855024103075266, + 0.002044418826699257, + 0.00048456809599883854, + 0.0009238088969141245, + 0.0008046207949519157, + 0.001018129289150238, + 0.0011022778926417232, + 0.00146026024594903, + 0.0008492973865941167, + 0.0011297118617221713, + 0.0005222285981290042, + 0.0007905278471298516, + 0.0006256572669371963, + 0.0008919431129470468, + 0.0005832512397319078, + 0.0006109408568590879, + 0.0008520634146407247, + 0.0004225797310937196, + 0.003081455361098051, + 0.0030423738062381744, + 0.004029103089123964, + 0.002918485552072525, + 0.0030353188049048185, + 0.003735200734809041, + 0.0029955452773720026, + 0.0024899656418710947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014196240343153477 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306123033165932 + ], + "opposite_logprob_tokens": [ + -3.1680612564086914 + ], + "logprob_reward": -0.04306123033165932, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306123033165932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6565169095993042, + 0.0773257240653038, + 0.03385643661022186, + 0.03035036474466324, + 0.02013547532260418, + 0.01651010848581791, + 0.016331713646650314, + 0.01784735918045044, + 0.011012192815542221, + 0.12011373788118362 + ], + "credit_map": [ + 0.03925500810146332, + 0.06153357774019241, + 0.20972803235054016, + 0.04239540174603462, + 0.08342629671096802, + 0.03578900918364525, + 0.029514791443943977, + 0.1548748016357422, + 0.005945791956037283, + 0.007277423981577158, + 0.007664604112505913, + 0.009974149987101555, + 0.0077280024997889996, + 0.0161654744297266, + 0.00974466186016798, + 0.012825618498027325, + 0.004227254074066877, + 0.005626022350043058, + 0.005638085305690765, + 0.0042285858653485775, + 0.003479011356830597, + 0.004743454046547413, + 0.002675227588042617, + 0.0032387941610068083, + 0.005586314480751753, + 0.005267019383609295, + 0.00375882675871253, + 0.0044603124260902405, + 0.002750992076471448, + 0.003167975228279829, + 0.0024334911722689867, + 0.002925433684140444, + 0.0018521700985729694, + 0.0030326987616717815, + 0.0028044262435287237, + 0.0028825222980231047, + 0.0019838165026158094, + 0.002732246881350875, + 0.001700852531939745, + 0.0031467422377318144, + 0.0014547163154929876, + 0.0026862649247050285, + 0.0025523770600557327, + 0.002716595074161887, + 0.001819161232560873, + 0.0015227930853143334, + 0.0025904371868819, + 0.001167762908153236, + 0.0015986994840204716, + 0.002850200282409787, + 0.001311269705183804, + 0.0025210718158632517, + 0.001772729097865522, + 0.0025057669263333082, + 0.0016753539675846696, + 0.002096623182296753, + 0.0017515704967081547, + 0.0020253092516213655, + 0.0017231237143278122, + 0.0027743023820221424, + 0.001975130056962371, + 0.0029395162127912045, + 0.0021045568864792585, + 0.0025538515765219927, + 0.0011119889095425606, + 0.001541979843750596, + 0.0015471286606043577, + 0.0015785235445946455, + 0.0013166619464755058, + 0.0013225993607193232, + 0.0013602501712739468, + 0.00123306002933532, + 0.023091040551662445, + 0.014295145869255066, + 0.025250624865293503, + 0.015590233728289604, + 0.013796468265354633, + 0.012841006740927696, + 0.008646104484796524, + 0.006603115238249302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.04306123033165932 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305620491504669 + ], + "opposite_logprob_tokens": [ + -3.1680562496185303 + ], + "logprob_reward": -0.04305620491504669, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305620491504669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.821937620639801, + 0.04959210753440857, + 0.02440514601767063, + 0.02030707523226738, + 0.01689210534095764, + 0.014793151058256626, + 0.013857327401638031, + 0.011250516399741173, + 0.00957114901393652, + 0.017393827438354492 + ], + "credit_map": [ + 0.034540604799985886, + 0.029941603541374207, + 0.34212324023246765, + 0.04375653713941574, + 0.05366818606853485, + 0.04202788695693016, + 0.026466719806194305, + 0.24941278994083405, + 0.004951184615492821, + 0.005927615333348513, + 0.007757696323096752, + 0.005401358008384705, + 0.005179954692721367, + 0.008892725221812725, + 0.005538060795515776, + 0.005943505559116602, + 0.0024931449443101883, + 0.0038269953802227974, + 0.002503564115613699, + 0.003273515962064266, + 0.0029870292637497187, + 0.003473574761301279, + 0.002234452171251178, + 0.0036128656938672066, + 0.002302406122907996, + 0.001815112424083054, + 0.004716913215816021, + 0.0023064061533659697, + 0.0021865894086658955, + 0.0027156490832567215, + 0.0014489179011434317, + 0.0028150787111371756, + 0.001349198748357594, + 0.002539532259106636, + 0.0019291604403406382, + 0.0021720908116549253, + 0.0019521195208653808, + 0.002617559628561139, + 0.0021125436760485172, + 0.0022198990918695927, + 0.0013201808324083686, + 0.0018647147808223963, + 0.0014628374483436346, + 0.001926634693518281, + 0.0016753684030845761, + 0.002427690662443638, + 0.0023281725589185953, + 0.0017875501653179526, + 0.0011704217176884413, + 0.0020725855138152838, + 0.0016839206218719482, + 0.0016765872715041041, + 0.0016003812197595835, + 0.0020606715697795153, + 0.0015521836467087269, + 0.002040575025603175, + 0.0007494041346944869, + 0.0012123080668970942, + 0.00108885217923671, + 0.0012819314142689109, + 0.001586827915161848, + 0.002599463565275073, + 0.0008358136983588338, + 0.0018959136214107275, + 0.0010166589636355639, + 0.0012483065947890282, + 0.0012511586537584662, + 0.0014161994913592935, + 0.0009624549420550466, + 0.001433970290236175, + 0.001098290435038507, + 0.001144108478911221, + 0.0026482506655156612, + 0.002281271619722247, + 0.0032024821266531944, + 0.0020865958649665117, + 0.0021486820187419653, + 0.002118391916155815, + 0.0015772371552884579, + 0.0013309132773429155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04305620491504669 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a slightly overhead angle, slowly and almost imperceptibly zooms in, focusing on a cold, white metal bed. Lying on the bed is a patient, his face as pale as paper, bloodless. Fine beads of cold sweat, dripping from his forehead with agony, gleam faintly under the cold, white light of the ward's overhead lamp. His thinning, dark hair is sticky to his temples, clung to his eyes with sweat. His eyes are half-closed, his gaze unfocused, and his lips, dry and cracked, struggle to open. Each breath is a struggle, a short, hoarse sound. His chest, shrouded by a loose, washed-out, light-blue striped hospital gown, rises and falls with alarming force, his collarbone clearly visible with each rise. The pure white sheets are crumpled by his unconscious struggles. The entire scene is shrouded in a uniform and cold fluorescent light, and the main colors are white, light blue and sickly gray skin, creating a tense atmosphere of extreme depression, pain and the fragility of life.", + "seed": 31583, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_1/index_807_seed_31583_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12134282309561968, + "vlm_logprob": -0.12134282309561968, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.12134282309561968, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014609739184379578 + ], + "opposite_logprob_tokens": [ + -4.7646098136901855 + ], + "logprob_reward": -0.014609739184379578, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.014609739184379578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07387593388557434, + 0.05221255496144295, + 0.0359504260122776, + 0.03699721023440361, + 0.026240888983011246, + 0.02051270380616188, + 0.40884485840797424, + 0.0761178508400917, + 0.14724573493003845, + 0.12200179696083069 + ], + "credit_map": [ + 0.004463029094040394, + 0.007029454689472914, + 0.009293404407799244, + 0.007586076855659485, + 0.009193792007863522, + 0.015138799324631691, + 0.010801601223647594, + 0.010369786061346531, + 0.004253279883414507, + 0.006233038380742073, + 0.00455043138936162, + 0.008863556198775768, + 0.0051113017834723, + 0.005289959255605936, + 0.00944831408560276, + 0.00846268143504858, + 0.0028978532645851374, + 0.004956208635121584, + 0.00496500451117754, + 0.004944341257214546, + 0.004674154799431562, + 0.003303373930975795, + 0.005141710862517357, + 0.005067782010883093, + 0.0034502940252423286, + 0.005941616836935282, + 0.005399643909186125, + 0.004731995519250631, + 0.004949283320456743, + 0.0037919909227639437, + 0.00391955254599452, + 0.004812837112694979, + 0.0018542984034866095, + 0.0046377419494092464, + 0.0023175817914307117, + 0.00531792314723134, + 0.003355027176439762, + 0.002187911421060562, + 0.003446379443630576, + 0.003124029142782092, + 0.0015990283573046327, + 0.004709349013864994, + 0.001246293424628675, + 0.0024817062076181173, + 0.0021041317377239466, + 0.0013537128688767552, + 0.0037603573873639107, + 0.003258127486333251, + 0.024536732584238052, + 0.046029482036828995, + 0.035553332418203354, + 0.05308329313993454, + 0.04791524261236191, + 0.08134278655052185, + 0.05945885553956032, + 0.0609251894056797, + 0.005075451917946339, + 0.011129800230264664, + 0.005362948402762413, + 0.008682486601173878, + 0.011854763142764568, + 0.011212551034986973, + 0.012345563620328903, + 0.010454289615154266, + 0.009617367759346962, + 0.008514097891747952, + 0.006788560654968023, + 0.007926413789391518, + 0.02998913638293743, + 0.042478278279304504, + 0.013126058503985405, + 0.028805840760469437, + 0.011485697701573372, + 0.010639162734150887, + 0.015142121352255344, + 0.013311557471752167, + 0.015764938667416573, + 0.015518536791205406, + 0.017447181046009064, + 0.022692611441016197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.014609739184379578 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234109163284302 + ], + "opposite_logprob_tokens": [ + -0.8984109163284302 + ], + "logprob_reward": -0.5234109163284302, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234109163284302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32776209712028503, + 0.16908693313598633, + 0.11042502522468567, + 0.041218869388103485, + 0.07697273790836334, + 0.08880596607923508, + 0.04574353247880936, + 0.04405519366264343, + 0.06251747906208038, + 0.033412087708711624 + ], + "credit_map": [ + 0.021200554445385933, + 0.04248570278286934, + 0.03397431597113609, + 0.04368637502193451, + 0.02868078462779522, + 0.041636448353528976, + 0.06707639247179031, + 0.04902154207229614, + 0.011865326203405857, + 0.017359476536512375, + 0.01197611540555954, + 0.030674438923597336, + 0.018131397664546967, + 0.014516920782625675, + 0.03676335886120796, + 0.027799896895885468, + 0.006957023870199919, + 0.011618363671004772, + 0.015188617631793022, + 0.015890130773186684, + 0.016397058963775635, + 0.007220159284770489, + 0.021527811884880066, + 0.015625864267349243, + 0.004487201105803251, + 0.006020293105393648, + 0.0035928389988839626, + 0.006463360972702503, + 0.0047492519952356815, + 0.0037774532102048397, + 0.006601863540709019, + 0.005526608321815729, + 0.005755807273089886, + 0.010874928906559944, + 0.007623289246112108, + 0.011263737455010414, + 0.008764436468482018, + 0.007817769423127174, + 0.015136349946260452, + 0.009736420586705208, + 0.008517883718013763, + 0.00963561236858368, + 0.006877432577311993, + 0.008221355266869068, + 0.01021555159240961, + 0.00792187824845314, + 0.02550450898706913, + 0.011911747045814991, + 0.0031269527971744537, + 0.007233082316815853, + 0.003988523967564106, + 0.006980076432228088, + 0.0038459477946162224, + 0.007048316765576601, + 0.006949481554329395, + 0.006571153178811073, + 0.0034507999662309885, + 0.005560973193496466, + 0.003773064585402608, + 0.0050336746498942375, + 0.00551008153706789, + 0.006895479280501604, + 0.007061370648443699, + 0.006769745144993067, + 0.0038891362491995096, + 0.006735788658261299, + 0.00410161167383194, + 0.005828614812344313, + 0.008741320110857487, + 0.013892004266381264, + 0.008287344127893448, + 0.011041665449738503, + 0.003838307224214077, + 0.00322775449603796, + 0.004253688734024763, + 0.003973776008933783, + 0.004190308041870594, + 0.0038697582203894854, + 0.004595525562763214, + 0.005462970118969679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.5234109163284302 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254394743591547 + ], + "opposite_logprob_tokens": [ + -4.38754415512085 + ], + "logprob_reward": -0.01254394743591547, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254394743591547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08379053324460983, + 0.1490744799375534, + 0.057933539152145386, + 0.06411252915859222, + 0.13622356951236725, + 0.18651607632637024, + 0.04602467641234398, + 0.11058540642261505, + 0.07347691804170609, + 0.09226230531930923 + ], + "credit_map": [ + 0.006408786866813898, + 0.008612055331468582, + 0.010535938665270805, + 0.01091625727713108, + 0.0074678827077150345, + 0.014438502490520477, + 0.014359943568706512, + 0.011051163077354431, + 0.009312045760452747, + 0.014036382548511028, + 0.01154231559485197, + 0.026555728167295456, + 0.01718413643538952, + 0.010968131013214588, + 0.034693919122219086, + 0.02478179708123207, + 0.0041867890395224094, + 0.007740663830190897, + 0.006216880399733782, + 0.007720067631453276, + 0.00747922295704484, + 0.004036601632833481, + 0.011529014445841312, + 0.009024295024573803, + 0.005178161431103945, + 0.008964561857283115, + 0.00840024184435606, + 0.009754653088748455, + 0.007337844930589199, + 0.007150484248995781, + 0.008943639695644379, + 0.00838293507695198, + 0.009448131546378136, + 0.015259388834238052, + 0.015393415465950966, + 0.03216129168868065, + 0.01742938905954361, + 0.013977138325572014, + 0.01863601803779602, + 0.013918794691562653, + 0.012837362475693226, + 0.018265152350068092, + 0.015527579933404922, + 0.017883425578475, + 0.01834922656416893, + 0.015367080457508564, + 0.05877799168229103, + 0.02950822003185749, + 0.003062500385567546, + 0.00795164704322815, + 0.0039212931878864765, + 0.007167241536080837, + 0.003980992361903191, + 0.004962657578289509, + 0.0077615766786038876, + 0.007216766010969877, + 0.008046461269259453, + 0.011954138055443764, + 0.008026049472391605, + 0.010396372526884079, + 0.017692260444164276, + 0.02054182067513466, + 0.016690397635102272, + 0.017237897962331772, + 0.005722332745790482, + 0.007339298725128174, + 0.00450350996106863, + 0.006600925698876381, + 0.01111635286360979, + 0.015228970907628536, + 0.009081742726266384, + 0.013883777894079685, + 0.008660931140184402, + 0.008157339878380299, + 0.010164394974708557, + 0.009104647673666477, + 0.012354857288300991, + 0.01291771698743105, + 0.01276825275272131, + 0.018134159967303276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.01254394743591547 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026351653039455414 + ], + "opposite_logprob_tokens": [ + -3.6513516902923584 + ], + "logprob_reward": -0.026351653039455414, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026351653039455414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05487256124615669, + 0.06379364430904388, + 0.18907056748867035, + 0.0707821473479271, + 0.08278238028287888, + 0.04773295298218727, + 0.2676985263824463, + 0.051545921713113785, + 0.0654405727982521, + 0.10628080368041992 + ], + "credit_map": [ + 0.00481936801224947, + 0.008373708464205265, + 0.007661949843168259, + 0.007157089654356241, + 0.005272974725812674, + 0.006324938032776117, + 0.0074437507428228855, + 0.007818781770765781, + 0.004136534407734871, + 0.006693204399198294, + 0.005736097693443298, + 0.010743166320025921, + 0.00794920977205038, + 0.004979001823812723, + 0.013720935210585594, + 0.009835497476160526, + 0.012055207043886185, + 0.019390899688005447, + 0.024061359465122223, + 0.029717007651925087, + 0.026555439457297325, + 0.011141450144350529, + 0.03966301679611206, + 0.026486186310648918, + 0.00958364736288786, + 0.009975218214094639, + 0.008627932518720627, + 0.007200153544545174, + 0.012278959155082703, + 0.007367212790995836, + 0.006724411156028509, + 0.00902460515499115, + 0.006570540834218264, + 0.00984978023916483, + 0.008411739952862263, + 0.021387288346886635, + 0.01017541904002428, + 0.009062468074262142, + 0.00850695837289095, + 0.008818181231617928, + 0.004700927529484034, + 0.0074506113305687904, + 0.005161015782505274, + 0.004428847227245569, + 0.00829022191464901, + 0.004801825154572725, + 0.006408426910638809, + 0.0064910766668617725, + 0.019380006939172745, + 0.04042376950383186, + 0.02287602797150612, + 0.04007392004132271, + 0.03494904935359955, + 0.03885296359658241, + 0.03412698581814766, + 0.03701578080654144, + 0.007657090667635202, + 0.006940656807273626, + 0.004405044950544834, + 0.007284718099981546, + 0.006664506625384092, + 0.006696747150272131, + 0.00518935639411211, + 0.00670779962092638, + 0.005638322327286005, + 0.008205666206777096, + 0.004468464758247137, + 0.006800477858632803, + 0.008655520156025887, + 0.01406450942158699, + 0.008956708014011383, + 0.008650901727378368, + 0.01229830365628004, + 0.01657785102725029, + 0.012183431535959244, + 0.013871612027287483, + 0.014800717122852802, + 0.011536535806953907, + 0.01143686007708311, + 0.013575485907495022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 9 + ], + "selected_reward": -0.026351653039455414 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029797859489917755 + ], + "opposite_logprob_tokens": [ + -3.5297977924346924 + ], + "logprob_reward": -0.029797859489917755, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029797859489917755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22488422691822052, + 0.17792999744415283, + 0.14838753640651703, + 0.025385510176420212, + 0.04110357165336609, + 0.05594949796795845, + 0.09734431654214859, + 0.0631108507514, + 0.07355271279811859, + 0.09235178679227829 + ], + "credit_map": [ + 0.01590474136173725, + 0.022737331688404083, + 0.026112565770745277, + 0.029536178335547447, + 0.020960761234164238, + 0.033933088183403015, + 0.042639557272195816, + 0.03306000679731369, + 0.011336767114698887, + 0.017523476853966713, + 0.013639001175761223, + 0.031867288053035736, + 0.020086774602532387, + 0.014140479266643524, + 0.0403001606464386, + 0.029036054387688637, + 0.008674940094351768, + 0.013258646242320538, + 0.019983181729912758, + 0.024556072428822517, + 0.023062212392687798, + 0.00753794377669692, + 0.030452165752649307, + 0.02086237259209156, + 0.0030631287954747677, + 0.0035392106510698795, + 0.0029099516104906797, + 0.0023303187917917967, + 0.003003497840836644, + 0.004273641388863325, + 0.003162954468280077, + 0.0031028063967823982, + 0.003181486390531063, + 0.004369331523776054, + 0.004696720279753208, + 0.007899516262114048, + 0.004583168774843216, + 0.0036878949031233788, + 0.008058489300310612, + 0.004626965615898371, + 0.0050002047792077065, + 0.007118951063603163, + 0.0037235519848763943, + 0.004577170126140118, + 0.006771352142095566, + 0.003747732611373067, + 0.01677834987640381, + 0.00823218747973442, + 0.0074330768547952175, + 0.011504238471388817, + 0.009254034608602524, + 0.0132287023589015, + 0.011441181413829327, + 0.019066795706748962, + 0.012667150236666203, + 0.012749134562909603, + 0.0058229463174939156, + 0.009637649171054363, + 0.005388688761740923, + 0.0053433761931955814, + 0.00856103003025055, + 0.01231924258172512, + 0.008248381316661835, + 0.007789537776261568, + 0.006552066188305616, + 0.008839543908834457, + 0.005584020633250475, + 0.008060352876782417, + 0.010311532765626907, + 0.012657631188631058, + 0.009216204285621643, + 0.01233136747032404, + 0.00798789318650961, + 0.009227316826581955, + 0.009248200803995132, + 0.008577467873692513, + 0.01280616968870163, + 0.015325812622904778, + 0.0113143902271986, + 0.017864538356661797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.029797859489917755 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a slightly overhead angle, slowly and almost imperceptibly zooms in, focusing on a cold, white metal bed. Lying on the bed is a patient, his face as pale as paper, bloodless. Fine beads of cold sweat, dripping from his forehead with agony, gleam faintly under the cold, white light of the ward's overhead lamp. His thinning, dark hair is sticky to his temples, clung to his eyes with sweat. His eyes are half-closed, his gaze unfocused, and his lips, dry and cracked, struggle to open. Each breath is a struggle, a short, hoarse sound. His chest, shrouded by a loose, washed-out, light-blue striped hospital gown, rises and falls with alarming force, his collarbone clearly visible with each rise. The pure white sheets are crumpled by his unconscious struggles. The entire scene is shrouded in a uniform and cold fluorescent light, and the main colors are white, light blue and sickly gray skin, creating a tense atmosphere of extreme depression, pain and the fragility of life.", + "seed": 31584, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_1/index_807_seed_31584_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04630132056772709, + "vlm_logprob": -0.04630132056772709, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.04630132056772709, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008195225149393082 + ], + "opposite_logprob_tokens": [ + -7.508195400238037 + ], + "logprob_reward": -0.008195225149393082, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.008195225149393082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15553678572177887, + 0.12288688123226166, + 0.09401454031467438, + 0.07694090902805328, + 0.05145258456468582, + 0.03778503090143204, + 0.0805964469909668, + 0.050263676792383194, + 0.05120992287993431, + 0.27931320667266846 + ], + "credit_map": [ + 0.013448542915284634, + 0.00883246399462223, + 0.04016440734267235, + 0.02513003535568714, + 0.016582826152443886, + 0.013043200597167015, + 0.0112618925049901, + 0.027073414996266365, + 0.015403826721012592, + 0.0069746882654726505, + 0.022411247715353966, + 0.016241081058979034, + 0.01518310233950615, + 0.010979571379721165, + 0.020198974758386612, + 0.015494395978748798, + 0.010424708016216755, + 0.008609146811068058, + 0.012170463800430298, + 0.011293902061879635, + 0.019451210275292397, + 0.010138092562556267, + 0.012557736597955227, + 0.00936928205192089, + 0.0055571445263922215, + 0.007159952539950609, + 0.009188057854771614, + 0.009540524333715439, + 0.01078738085925579, + 0.009025844745337963, + 0.01362977921962738, + 0.012052224017679691, + 0.006330207921564579, + 0.005774837452918291, + 0.009301122277975082, + 0.00531041668727994, + 0.007398051675409079, + 0.007727630436420441, + 0.004459674935787916, + 0.00515064038336277, + 0.0034319974947720766, + 0.003004784230142832, + 0.005277233198285103, + 0.004549458157271147, + 0.006771143060177565, + 0.005081137176603079, + 0.005733271595090628, + 0.0039360057562589645, + 0.007546175736933947, + 0.008133092895150185, + 0.012031536549329758, + 0.010808855295181274, + 0.008081726729869843, + 0.009639701806008816, + 0.01079324446618557, + 0.013562115840613842, + 0.007171907462179661, + 0.005866197403520346, + 0.01008276641368866, + 0.0066529386676847935, + 0.0038491394370794296, + 0.006445485632866621, + 0.005588584113866091, + 0.004606655798852444, + 0.003931399900466204, + 0.007767890579998493, + 0.006099522113800049, + 0.007728372700512409, + 0.0078789833933115, + 0.0067276060581207275, + 0.007396953646093607, + 0.0036791907623410225, + 0.01740434393286705, + 0.03790971264243126, + 0.033892206847667694, + 0.04374564811587334, + 0.03579938784241676, + 0.03445084020495415, + 0.05048040300607681, + 0.025630660355091095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.008195225149393082 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039623320102692 + ], + "opposite_logprob_tokens": [ + -1.9103962182998657 + ], + "logprob_reward": -0.16039623320102692, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039623320102692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18998661637306213, + 0.21194925904273987, + 0.14474888145923615, + 0.050475288182497025, + 0.03482016175985336, + 0.15027786791324615, + 0.07662257552146912, + 0.059958599507808685, + 0.02824668399989605, + 0.0529140867292881 + ], + "credit_map": [ + 0.021366775035858154, + 0.01580469124019146, + 0.052682776004076004, + 0.028371714055538177, + 0.015469730831682682, + 0.017548855394124985, + 0.014527679421007633, + 0.024214373901486397, + 0.041693150997161865, + 0.018695387989282608, + 0.0270194374024868, + 0.0312812402844429, + 0.022281458601355553, + 0.022541306912899017, + 0.01939629204571247, + 0.029040971770882607, + 0.014984546229243279, + 0.012700245715677738, + 0.019545497372746468, + 0.022984763607382774, + 0.021079272031784058, + 0.022925712168216705, + 0.015212616883218288, + 0.015316217206418514, + 0.008234755136072636, + 0.005401587579399347, + 0.008946076966822147, + 0.006148347165435553, + 0.005639741662889719, + 0.006896759383380413, + 0.005674478132277727, + 0.0035335372667759657, + 0.005401685368269682, + 0.00402862299233675, + 0.006088758818805218, + 0.0037300074473023415, + 0.003941136412322521, + 0.004809784237295389, + 0.0029025888070464134, + 0.003917575813829899, + 0.015269642695784569, + 0.01158476434648037, + 0.03530973941087723, + 0.018422024324536324, + 0.015339321456849575, + 0.02419712394475937, + 0.01467796042561531, + 0.015477268025279045, + 0.007212519180029631, + 0.006335671059787273, + 0.01960395835340023, + 0.009375753812491894, + 0.0060483915731310844, + 0.012760425917804241, + 0.007201884873211384, + 0.008083966560661793, + 0.007526608183979988, + 0.004766044672578573, + 0.01525614783167839, + 0.00529028195887804, + 0.00436051469296217, + 0.010680872946977615, + 0.008029424585402012, + 0.004048697650432587, + 0.0019667576998472214, + 0.0037863412871956825, + 0.0033603005576878786, + 0.004348364192992449, + 0.004017404280602932, + 0.003871259745210409, + 0.004057012498378754, + 0.0028392374515533447, + 0.006547179538756609, + 0.004153764806687832, + 0.011659177951514721, + 0.007448666263371706, + 0.004111846908926964, + 0.00754112284630537, + 0.005402978044003248, + 0.006049348972737789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.16039623320102692 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254265196621418 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.01254265196621418, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254265196621418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24225616455078125, + 0.13278725743293762, + 0.2026362419128418, + 0.09242097288370132, + 0.06756070256233215, + 0.0571041963994503, + 0.07374659180641174, + 0.05527546629309654, + 0.01713515631854534, + 0.05907725542783737 + ], + "credit_map": [ + 0.019606098532676697, + 0.014746589586138725, + 0.05361861363053322, + 0.025065377354621887, + 0.03616972640156746, + 0.020143236964941025, + 0.032080333679914474, + 0.040826182812452316, + 0.021764235571026802, + 0.013468068093061447, + 0.019292818382382393, + 0.015292969532310963, + 0.01916830986738205, + 0.016107259318232536, + 0.012442097999155521, + 0.015251495875418186, + 0.019046032801270485, + 0.016942353919148445, + 0.02313830703496933, + 0.030207686126232147, + 0.03636184334754944, + 0.02402954176068306, + 0.03186481446027756, + 0.02104567550122738, + 0.007485399022698402, + 0.008732287213206291, + 0.010684686712920666, + 0.009103827178478241, + 0.011326243169605732, + 0.011840282939374447, + 0.015605243854224682, + 0.017642999067902565, + 0.010682309046387672, + 0.008092395029962063, + 0.009871186688542366, + 0.0069306958466768265, + 0.00930524431169033, + 0.00982897263020277, + 0.007347298786044121, + 0.005502601619809866, + 0.004928659647703171, + 0.005107045639306307, + 0.015301985666155815, + 0.006799755152314901, + 0.005426976829767227, + 0.009767238050699234, + 0.004182467702776194, + 0.005590069107711315, + 0.007967621088027954, + 0.005197628866881132, + 0.020011067390441895, + 0.010263836942613125, + 0.0054977512918412685, + 0.011968899518251419, + 0.005464504472911358, + 0.007375274784862995, + 0.006705513224005699, + 0.004759388975799084, + 0.010975305922329426, + 0.00552277360111475, + 0.004057638347148895, + 0.008105205371975899, + 0.006302505731582642, + 0.008847133256494999, + 0.0014699018793180585, + 0.002776775509119034, + 0.002237060572952032, + 0.0022773032542318106, + 0.002529395744204521, + 0.00224021403118968, + 0.0022869776003062725, + 0.0013175284257158637, + 0.0064494977705180645, + 0.006569070741534233, + 0.009911530651152134, + 0.007601138204336166, + 0.007330282125622034, + 0.007706538774073124, + 0.008820089511573315, + 0.004689112305641174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.01254265196621418 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794156551361084 + ], + "opposite_logprob_tokens": [ + -3.529794216156006 + ], + "logprob_reward": -0.029794156551361084, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794156551361084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29230043292045593, + 0.07542215287685394, + 0.051092926412820816, + 0.0327749028801918, + 0.0450250506401062, + 0.02973182313144207, + 0.060739923268556595, + 0.030457697808742523, + 0.14407944679260254, + 0.23837563395500183 + ], + "credit_map": [ + 0.021575771272182465, + 0.007080166134983301, + 0.10665106028318405, + 0.028669504448771477, + 0.01277603954076767, + 0.015019203536212444, + 0.017200442031025887, + 0.0833282321691513, + 0.005609650630503893, + 0.005469959229230881, + 0.01489301584661007, + 0.012093143537640572, + 0.007566867861896753, + 0.004723174963146448, + 0.015069719403982162, + 0.00999662559479475, + 0.004642978310585022, + 0.0036467649042606354, + 0.005628201644867659, + 0.009561165235936642, + 0.006734473165124655, + 0.004894320387393236, + 0.009746463038027287, + 0.006238557863980532, + 0.004019689746201038, + 0.0036951573565602303, + 0.004128176718950272, + 0.003918477799743414, + 0.004072601906955242, + 0.004543324466794729, + 0.003713054582476616, + 0.0046844189055264, + 0.0060145738534629345, + 0.004561827518045902, + 0.008286834694445133, + 0.00435782503336668, + 0.007283973973244429, + 0.006889329291880131, + 0.004519876558333635, + 0.0031108062248677015, + 0.0033129449002444744, + 0.003144965274259448, + 0.004557301290333271, + 0.003491849172860384, + 0.004183601122349501, + 0.0035384194925427437, + 0.0035458628553897142, + 0.003956880420446396, + 0.004943664651364088, + 0.0075570158660411835, + 0.008664573542773724, + 0.007497703656554222, + 0.009103106334805489, + 0.00753479590639472, + 0.008180373348295689, + 0.007258689031004906, + 0.0034885434433817863, + 0.0026535133365541697, + 0.003847013460472226, + 0.0029294320847839117, + 0.0028892208356410265, + 0.003637946443632245, + 0.006292247679084539, + 0.004719780758023262, + 0.009946295991539955, + 0.02570788376033306, + 0.016606641933321953, + 0.01678765006363392, + 0.020152445882558823, + 0.019749131053686142, + 0.01944989711046219, + 0.015679500997066498, + 0.01943371444940567, + 0.029583550989627838, + 0.02844857983291149, + 0.034500036388635635, + 0.03452184423804283, + 0.027677956968545914, + 0.04404578357934952, + 0.02016417495906353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.029794156551361084 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020578335970640182 + ], + "opposite_logprob_tokens": [ + -3.895578384399414 + ], + "logprob_reward": -0.020578335970640182, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020578335970640182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3151090443134308, + 0.18155525624752045, + 0.047111548483371735, + 0.08846380561590195, + 0.06986688822507858, + 0.08713139593601227, + 0.0482526496052742, + 0.060726311057806015, + 0.03732113167643547, + 0.06446190923452377 + ], + "credit_map": [ + 0.02736945077776909, + 0.019967256113886833, + 0.08188394457101822, + 0.03420892730355263, + 0.025661759078502655, + 0.024737514555454254, + 0.017652899026870728, + 0.08362728357315063, + 0.027119331061840057, + 0.017834870144724846, + 0.022270360961556435, + 0.023688260465860367, + 0.028433427214622498, + 0.022465262562036514, + 0.016436368227005005, + 0.02330736629664898, + 0.00787384808063507, + 0.0029620237182825804, + 0.0062694307416677475, + 0.011038687080144882, + 0.00414191884920001, + 0.006854295264929533, + 0.004145734943449497, + 0.003825610503554344, + 0.00885610282421112, + 0.006109877955168486, + 0.01188631821423769, + 0.00914082583039999, + 0.009929961524903774, + 0.010804673656821251, + 0.015832366421818733, + 0.01590368151664734, + 0.009280923753976822, + 0.007240669801831245, + 0.01306837983429432, + 0.00722224498167634, + 0.009733901359140873, + 0.010314250364899635, + 0.0069693755358457565, + 0.0060371458530426025, + 0.00841501634567976, + 0.006382284685969353, + 0.021027972921729088, + 0.008995609357953072, + 0.00985636468976736, + 0.014575489796698093, + 0.008707357570528984, + 0.009171303361654282, + 0.005245980340987444, + 0.0030726706609129906, + 0.01110298652201891, + 0.004384552128612995, + 0.004534827545285225, + 0.007799929939210415, + 0.004616700578480959, + 0.007495000958442688, + 0.006467323284596205, + 0.006514967884868383, + 0.01133334543555975, + 0.006195627152919769, + 0.004809378180652857, + 0.008842699229717255, + 0.009999889880418777, + 0.006563078612089157, + 0.0025593419559299946, + 0.00795281957834959, + 0.003539257450029254, + 0.004486257676035166, + 0.0048456015065312386, + 0.00492494460195303, + 0.004422771744430065, + 0.004590136930346489, + 0.008739473298192024, + 0.006588908843696117, + 0.00967010110616684, + 0.010181907564401627, + 0.006371019408106804, + 0.008373642340302467, + 0.008035675622522831, + 0.006501184776425362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.020578335970640182 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a slightly overhead angle, slowly and almost imperceptibly zooms in, focusing on a cold, white metal bed. Lying on the bed is a patient, his face as pale as paper, bloodless. Fine beads of cold sweat, dripping from his forehead with agony, gleam faintly under the cold, white light of the ward's overhead lamp. His thinning, dark hair is sticky to his temples, clung to his eyes with sweat. His eyes are half-closed, his gaze unfocused, and his lips, dry and cracked, struggle to open. Each breath is a struggle, a short, hoarse sound. His chest, shrouded by a loose, washed-out, light-blue striped hospital gown, rises and falls with alarming force, his collarbone clearly visible with each rise. The pure white sheets are crumpled by his unconscious struggles. The entire scene is shrouded in a uniform and cold fluorescent light, and the main colors are white, light blue and sickly gray skin, creating a tense atmosphere of extreme depression, pain and the fragility of life.", + "seed": 31585, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_1/index_807_seed_31585_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.22020200891420244, + "vlm_logprob": -0.22020200891420244, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.22020200891420244, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059802173636853695 + ], + "opposite_logprob_tokens": [ + -7.255980014801025 + ], + "logprob_reward": -0.0059802173636853695, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0059802173636853695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17906993627548218, + 0.07563644647598267, + 0.22651471197605133, + 0.06274929642677307, + 0.09695160388946533, + 0.05248589813709259, + 0.07202577590942383, + 0.08483767509460449, + 0.08624379336833954, + 0.06348482519388199 + ], + "credit_map": [ + 0.013383644632995129, + 0.015848282724618912, + 0.016842462122440338, + 0.011087095364928246, + 0.030223267152905464, + 0.03658978268504143, + 0.016584521159529686, + 0.03851088508963585, + 0.0111861452460289, + 0.011640889570116997, + 0.010688915848731995, + 0.007570347283035517, + 0.010438243858516216, + 0.008673308417201042, + 0.006087821908295155, + 0.009350772015750408, + 0.014550539664924145, + 0.010942559689283371, + 0.012117099948227406, + 0.007626864593476057, + 0.091009221971035, + 0.03702891990542412, + 0.010080086067318916, + 0.04315941780805588, + 0.004953805357217789, + 0.01563943177461624, + 0.004212473519146442, + 0.006952883675694466, + 0.01231378223747015, + 0.008502856828272343, + 0.004015216138213873, + 0.006158851087093353, + 0.0042928969487547874, + 0.009783986955881119, + 0.0060770269483327866, + 0.005799145437777042, + 0.03043263591825962, + 0.022437362000346184, + 0.008791784755885601, + 0.00933676678687334, + 0.004984213970601559, + 0.007689411751925945, + 0.0057807741686701775, + 0.006783421617001295, + 0.006259515881538391, + 0.005876295268535614, + 0.005259275436401367, + 0.009852990508079529, + 0.00602344237267971, + 0.007394263055175543, + 0.008862902410328388, + 0.009000042453408241, + 0.006355604622513056, + 0.011097847484052181, + 0.008642039261758327, + 0.014649632386863232, + 0.005889662075787783, + 0.008369815535843372, + 0.00643179751932621, + 0.009328937157988548, + 0.011480312794446945, + 0.02022159844636917, + 0.01041081827133894, + 0.012704728171229362, + 0.009442462585866451, + 0.011502863839268684, + 0.01088763028383255, + 0.008705655112862587, + 0.010473079048097134, + 0.01628357172012329, + 0.009037408046424389, + 0.009911119937896729, + 0.00537973502650857, + 0.007545900531113148, + 0.005975337233394384, + 0.008923981338739395, + 0.007733771577477455, + 0.01063395943492651, + 0.007239286322146654, + 0.010052857920527458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.0059802173636853695 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742858409881592 + ], + "opposite_logprob_tokens": [ + -0.4742858409881592 + ], + "logprob_reward": -0.9742858409881592, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742858409881592, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10977595299482346, + 0.07923855632543564, + 0.4813159704208374, + 0.03397328034043312, + 0.07507891952991486, + 0.07786408066749573, + 0.028981922194361687, + 0.014476912096142769, + 0.029337136074900627, + 0.06995736807584763 + ], + "credit_map": [ + 0.007814041338860989, + 0.013026653788983822, + 0.008565811440348625, + 0.009503810666501522, + 0.023937473073601723, + 0.019597942009568214, + 0.011213771998882294, + 0.016116442158818245, + 0.004575587809085846, + 0.006126295309513807, + 0.004699668847024441, + 0.0043779280968010426, + 0.008781650103628635, + 0.01860802061855793, + 0.0051949964836239815, + 0.02687440626323223, + 0.029543356969952583, + 0.026250073686242104, + 0.028239261358976364, + 0.020517095923423767, + 0.17862053215503693, + 0.07568871974945068, + 0.04060598462820053, + 0.08185090869665146, + 0.002203770447522402, + 0.003751305164769292, + 0.001873001572676003, + 0.0021453003864735365, + 0.010403572581708431, + 0.0058836969546973705, + 0.0018796519143506885, + 0.005832979455590248, + 0.003982145339250565, + 0.007189604453742504, + 0.004550902638584375, + 0.0043936362490057945, + 0.012650509364902973, + 0.012803525663912296, + 0.008743389509618282, + 0.020765205845236778, + 0.005216570105403662, + 0.00825777929276228, + 0.007906782440841198, + 0.007055386435240507, + 0.009148365817964077, + 0.01276522409170866, + 0.009016413241624832, + 0.018497556447982788, + 0.00238039786927402, + 0.003094655927270651, + 0.0032652735244482756, + 0.0030986140482127666, + 0.003306441009044647, + 0.006227731239050627, + 0.002702629892155528, + 0.004906178452074528, + 0.0013496158644557, + 0.001996051287278533, + 0.0017284175846725702, + 0.0014819756615906954, + 0.001629544305615127, + 0.0021146503277122974, + 0.0014915608335286379, + 0.0026850965805351734, + 0.002871978562325239, + 0.0036862671840935946, + 0.0023221273440867662, + 0.003884687786921859, + 0.0038145482540130615, + 0.005050990264862776, + 0.0038259606808423996, + 0.0038805785588920116, + 0.00403454527258873, + 0.0048388466238975525, + 0.004467287566512823, + 0.005791190080344677, + 0.006672886200249195, + 0.019764386117458344, + 0.0047000618651509285, + 0.019688162952661514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.9742858409881592 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038112103939056396 + ], + "opposite_logprob_tokens": [ + -3.288112163543701 + ], + "logprob_reward": -0.038112103939056396, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038112103939056396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17451684176921844, + 0.11589442938566208, + 0.15742920339107513, + 0.07272486388683319, + 0.10717834532260895, + 0.20478522777557373, + 0.04848713055253029, + 0.021271945908665657, + 0.04178093001246452, + 0.05593108385801315 + ], + "credit_map": [ + 0.012566162273287773, + 0.014714053831994534, + 0.019249949604272842, + 0.010824430733919144, + 0.016472600400447845, + 0.030173027887940407, + 0.013165853917598724, + 0.057350751012563705, + 0.010301578789949417, + 0.015301579609513283, + 0.013100421987473965, + 0.011523871682584286, + 0.016056345775723457, + 0.015483643859624863, + 0.012826870195567608, + 0.021300116553902626, + 0.009883871302008629, + 0.011002909392118454, + 0.013356032781302929, + 0.009357038885354996, + 0.05292976647615433, + 0.02451936900615692, + 0.013816858641803265, + 0.022563356906175613, + 0.004247960634529591, + 0.010708118788897991, + 0.005868389271199703, + 0.005092634353786707, + 0.019584273919463158, + 0.01026405580341816, + 0.003248523687943816, + 0.013710909523069859, + 0.0053115603514015675, + 0.010887766256928444, + 0.005907793994992971, + 0.005819662008434534, + 0.03428473696112633, + 0.0242315586656332, + 0.01125367358326912, + 0.00948159210383892, + 0.016029279679059982, + 0.022151442244648933, + 0.01946219988167286, + 0.02629861980676651, + 0.029545968398451805, + 0.027409080415964127, + 0.02188999392092228, + 0.04199863597750664, + 0.005124646704643965, + 0.0077431718818843365, + 0.004182352684438229, + 0.006213841028511524, + 0.005498513113707304, + 0.008401434868574142, + 0.0036896420642733574, + 0.007633525412529707, + 0.0013852552510797977, + 0.0024672041181474924, + 0.0015566380461677909, + 0.0019462407799437642, + 0.003718285821378231, + 0.0043922532349824905, + 0.002514463849365711, + 0.003291604807600379, + 0.0032772859558463097, + 0.006760686170309782, + 0.003408947726711631, + 0.003848223714157939, + 0.005607579834759235, + 0.009406615048646927, + 0.0042233215644955635, + 0.005248273257166147, + 0.003379094647243619, + 0.00516996905207634, + 0.0035387794487178326, + 0.006198827177286148, + 0.005686035379767418, + 0.013638712465763092, + 0.0044310796074569225, + 0.013888584449887276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.038112103939056396 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583707839250565 + ], + "opposite_logprob_tokens": [ + -3.8955836296081543 + ], + "logprob_reward": -0.020583707839250565, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583707839250565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12707316875457764, + 0.048467475920915604, + 0.18634133040905, + 0.04032023996114731, + 0.09782454371452332, + 0.08252906799316406, + 0.11412745714187622, + 0.05335228517651558, + 0.12265317142009735, + 0.12731130421161652 + ], + "credit_map": [ + 0.011888675391674042, + 0.016324637457728386, + 0.013213710859417915, + 0.010475619696080685, + 0.025244886055588722, + 0.018528927117586136, + 0.012505742721259594, + 0.018890978768467903, + 0.004840788431465626, + 0.0049691954627633095, + 0.006158190313726664, + 0.004757173825055361, + 0.007177427411079407, + 0.006049134768545628, + 0.005310220643877983, + 0.00920534785836935, + 0.011901162564754486, + 0.011142573319375515, + 0.011942589655518532, + 0.007666759192943573, + 0.07083591818809509, + 0.0319892019033432, + 0.012959101237356663, + 0.027904024347662926, + 0.002941794926300645, + 0.00698002427816391, + 0.003532538888975978, + 0.003392921295017004, + 0.008993495255708694, + 0.006370900198817253, + 0.0026552623603492975, + 0.005453306250274181, + 0.006962033919990063, + 0.008456936106085777, + 0.00793008878827095, + 0.009381033480167389, + 0.016158035025000572, + 0.014877697452902794, + 0.013762924820184708, + 0.020295800641179085, + 0.006106224842369556, + 0.007941959425807, + 0.007392719388008118, + 0.0104363476857543, + 0.01014136616140604, + 0.011804001405835152, + 0.0096591142937541, + 0.01904732920229435, + 0.010839239694178104, + 0.016935963183641434, + 0.009508426301181316, + 0.014944263733923435, + 0.01213071495294571, + 0.017926281318068504, + 0.014970360323786736, + 0.016872212290763855, + 0.004937103018164635, + 0.007844575680792332, + 0.00339410244487226, + 0.005441171582788229, + 0.005836979951709509, + 0.009127265773713589, + 0.006693287752568722, + 0.01007780060172081, + 0.014480606652796268, + 0.027470817789435387, + 0.00945248082280159, + 0.015032860450446606, + 0.015121660195291042, + 0.01566411182284355, + 0.011416434310376644, + 0.014014196582138538, + 0.009641852229833603, + 0.017484955489635468, + 0.007621293887495995, + 0.01576278731226921, + 0.011960997246205807, + 0.024531276896595955, + 0.01108865812420845, + 0.029219474643468857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.020583707839250565 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a patient lying in a bed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bed made of cold white metal with pure white crumpled sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the patient wearing a loose, washed-out, light-blue striped hospital gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and almost imperceptibly zoom in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a uniform, cold fluorescent light creating a sickly gray skin tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204817444086075 + ], + "opposite_logprob_tokens": [ + -2.8120481967926025 + ], + "logprob_reward": -0.06204817444086075, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204817444086075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15453635156154633, + 0.04120342433452606, + 0.1131877601146698, + 0.06035034731030464, + 0.10981828719377518, + 0.06882303953170776, + 0.11638613045215607, + 0.16128982603549957, + 0.10745519399642944, + 0.06694965064525604 + ], + "credit_map": [ + 0.019770653918385506, + 0.015065299347043037, + 0.016372989863157272, + 0.010907766409218311, + 0.03327583149075508, + 0.025986332446336746, + 0.013864148408174515, + 0.01929333619773388, + 0.0036398633383214474, + 0.006084589287638664, + 0.0046706851571798325, + 0.005066220182925463, + 0.005746397189795971, + 0.005543543957173824, + 0.004150683525949717, + 0.006301439367234707, + 0.007196673192083836, + 0.006361587438732386, + 0.005682998336851597, + 0.005193586926907301, + 0.041707754135131836, + 0.01762998476624489, + 0.008227191865444183, + 0.02118798904120922, + 0.004208024125546217, + 0.0104709817096591, + 0.005394336301833391, + 0.005654646549373865, + 0.0122349439188838, + 0.008042492903769016, + 0.003724678885191679, + 0.010620244778692722, + 0.007781525142490864, + 0.009941915050148964, + 0.009039805270731449, + 0.010356918908655643, + 0.01984969526529312, + 0.01666080392897129, + 0.0155484015122056, + 0.020639220252633095, + 0.0038005909882485867, + 0.004971741233021021, + 0.006226289086043835, + 0.006495041772723198, + 0.006296207662671804, + 0.012194318696856499, + 0.007536747492849827, + 0.021302103996276855, + 0.009862546809017658, + 0.012482632882893085, + 0.009801909327507019, + 0.009905567392706871, + 0.015525903552770615, + 0.024544859305024147, + 0.010998008772730827, + 0.023264700546860695, + 0.011957556940615177, + 0.016136758029460907, + 0.0161444079130888, + 0.016137273982167244, + 0.021774442866444588, + 0.026081928983330727, + 0.02065817452967167, + 0.03239928558468819, + 0.009388391859829426, + 0.012549561448395252, + 0.014154428616166115, + 0.00954656582325697, + 0.01471692230552435, + 0.02358097769320011, + 0.010998652316629887, + 0.01251968089491129, + 0.004425101913511753, + 0.006374755408614874, + 0.007491309195756912, + 0.005499751772731543, + 0.007710727863013744, + 0.01385806966573, + 0.00672636553645134, + 0.014863568358123302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.06204817444086075 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_10/index_tensor([3632, 3632])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_10/index_tensor([3632, 3632])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2d5d3ffe8af2f1423bd0df1e23016e1435415a9c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_10/index_tensor([3632, 3632])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angle close-up captures a stack of worn magazines, resting on the dark, polished wood of a dusty table. The top magazine, a vintage publication, features a faded, saturated cover and yellowed, dog-eared pages, tilted slightly relative to the magazines below. A warm, off-camera light source casts a focused pool of illumination, highlighting the fine layer of dust on the pages and casting the surrounding room into deep, blurred shadows, creating an atmosphere of tension and mystery. The scene is initially tranquil, but then suddenly disturbed as the stack begins to tremble, a low-frequency vibration at first almost imperceptible throughout. Then, the thin, fragile edges of the pages begin to ripple and flutter erratically, not from scratch, but as if stirred by an inner, unseen force. The trembling intensifies, causing the entire stack to vibrate more violently, individual pages colliding with each other in a silent, chaotic dance, suggesting architectural, supernatural presences.", + "seed": 10964, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_10/index_3632_seed_10964_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.18513600192964078, + "vlm_logprob": -0.18513600192964078, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.18513600192964078, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019311392679810524 + ], + "opposite_logprob_tokens": [ + -4.394311428070068 + ], + "logprob_reward": -0.019311392679810524, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.019311392679810524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21678948402404785, + 0.03652773052453995, + 0.06676184386014938, + 0.034883543848991394, + 0.01843506470322609, + 0.08154682070016861, + 0.07828883826732635, + 0.08977939933538437, + 0.08348863571882248, + 0.2934986650943756 + ], + "credit_map": [ + 0.039890091866254807, + 0.051098957657814026, + 0.02370980754494667, + 0.01934818923473358, + 0.02444778010249138, + 0.02730165421962738, + 0.014268813654780388, + 0.016724197193980217, + 0.006506720092147589, + 0.0063407509587705135, + 0.005206855945289135, + 0.0032673468813300133, + 0.004753813613206148, + 0.005648581311106682, + 0.002221861155703664, + 0.002581800101324916, + 0.011318139731884003, + 0.02678551897406578, + 0.004162260796874762, + 0.002370631555095315, + 0.013668788596987724, + 0.005585961509495974, + 0.00134041509591043, + 0.0015301181701943278, + 0.004957177210599184, + 0.006027512717992067, + 0.0032396502792835236, + 0.003025814425200224, + 0.004277353640645742, + 0.0042404234409332275, + 0.0030004160944372416, + 0.006115193944424391, + 0.00336275901645422, + 0.0024872079957276583, + 0.0021174147259444, + 0.001546652289107442, + 0.0035704777110368013, + 0.0022483523935079575, + 0.0013069770066067576, + 0.0017952213529497385, + 0.011127964593470097, + 0.017877988517284393, + 0.010462009347975254, + 0.00822798814624548, + 0.009948967024683952, + 0.012229127809405327, + 0.0057088159956038, + 0.005963948555290699, + 0.012317494489252567, + 0.03178786113858223, + 0.004049778915941715, + 0.0032863037195056677, + 0.015614964999258518, + 0.0069908443838357925, + 0.0017310987459495664, + 0.002510488498955965, + 0.01393506582826376, + 0.02272743172943592, + 0.0070456028915941715, + 0.008275795727968216, + 0.011331330984830856, + 0.007803559303283691, + 0.007430163212120533, + 0.01123044639825821, + 0.013926439918577671, + 0.009590605273842812, + 0.01200298871845007, + 0.0075164358131587505, + 0.007071399129927158, + 0.012276372872292995, + 0.007423355244100094, + 0.013681030832231045, + 0.05045074597001076, + 0.021351441740989685, + 0.031119367107748985, + 0.017902709543704987, + 0.04146779328584671, + 0.061726611107587814, + 0.026079829782247543, + 0.04340014234185219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.019311392679810524 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029881512746214867 + ], + "opposite_logprob_tokens": [ + -3.529881477355957 + ], + "logprob_reward": -0.029881512746214867, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029881512746214867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19439883530139923, + 0.0783582478761673, + 0.0756065770983696, + 0.02822822332382202, + 0.021240567788481712, + 0.0618465282022953, + 0.10714233666658401, + 0.08743680268526077, + 0.07895588129758835, + 0.2667860686779022 + ], + "credit_map": [ + 0.03318146616220474, + 0.0329313650727272, + 0.022949175909161568, + 0.010479386895895004, + 0.021494148299098015, + 0.02515805885195732, + 0.017908209934830666, + 0.03029700182378292, + 0.009382291696965694, + 0.02464311011135578, + 0.006811483763158321, + 0.003551451489329338, + 0.014945119619369507, + 0.006676970049738884, + 0.004987902473658323, + 0.0073599074967205524, + 0.01303982175886631, + 0.026199840009212494, + 0.006673467345535755, + 0.003469476941972971, + 0.011711406521499157, + 0.007047660183161497, + 0.0027430960908532143, + 0.004721790086477995, + 0.0026372491847723722, + 0.007322031073272228, + 0.0027637453749775887, + 0.002237355336546898, + 0.005330261308699846, + 0.0037766487803310156, + 0.0015341717517003417, + 0.002626757137477398, + 0.0038380399346351624, + 0.0038575141225010157, + 0.0017954155337065458, + 0.0017590525094419718, + 0.0026261552702635527, + 0.0030135302804410458, + 0.0017005247063934803, + 0.0026503338012844324, + 0.007808416150510311, + 0.011754355393350124, + 0.00847510527819395, + 0.004978599492460489, + 0.007548320572823286, + 0.010887701995670795, + 0.004631540738046169, + 0.0057624829933047295, + 0.017361938953399658, + 0.04358730837702751, + 0.005117092281579971, + 0.005027095787227154, + 0.02097461000084877, + 0.00980434287339449, + 0.0023989835754036903, + 0.0028709566686302423, + 0.014953349716961384, + 0.021004585549235344, + 0.0077758291736245155, + 0.006392885930836201, + 0.010646376758813858, + 0.008403465151786804, + 0.006376856472343206, + 0.011883443221449852, + 0.013434087857604027, + 0.009360290132462978, + 0.012153249233961105, + 0.007935271598398685, + 0.007163344416767359, + 0.011212407611310482, + 0.006709124892950058, + 0.010988096706569195, + 0.04928045719861984, + 0.017591295763850212, + 0.02654261700809002, + 0.015820922330021858, + 0.04246249794960022, + 0.05290433391928673, + 0.022172024473547935, + 0.04001191630959511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.029881512746214867 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048818159848451614 + ], + "opposite_logprob_tokens": [ + -3.0488181114196777 + ], + "logprob_reward": -0.048818159848451614, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048818159848451614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10163408517837524, + 0.039034780114889145, + 0.07314807176589966, + 0.03440001979470253, + 0.06891369819641113, + 0.11268652975559235, + 0.1620636135339737, + 0.1570080816745758, + 0.07776651531457901, + 0.17334459722042084 + ], + "credit_map": [ + 0.020171349868178368, + 0.021952252835035324, + 0.009949751198291779, + 0.008428170345723629, + 0.012345447205007076, + 0.012864353135228157, + 0.00613691471517086, + 0.009785842150449753, + 0.006356758065521717, + 0.005229972768574953, + 0.004858921281993389, + 0.0038154784124344587, + 0.0071677290834486485, + 0.004493271466344595, + 0.0031772367656230927, + 0.003935409709811211, + 0.010980573482811451, + 0.027216404676437378, + 0.00449055852368474, + 0.0032717431895434856, + 0.014031452126801014, + 0.005847681313753128, + 0.0031169140711426735, + 0.004192742053419352, + 0.004919709637761116, + 0.012620832771062851, + 0.002421012846753001, + 0.001534205162897706, + 0.006444168277084827, + 0.003240276826545596, + 0.0012995332945138216, + 0.0019202814437448978, + 0.010464639402925968, + 0.021787531673908234, + 0.005377377383410931, + 0.004085982218384743, + 0.011258920654654503, + 0.005937452893704176, + 0.003998902160674334, + 0.006002879235893488, + 0.016651781275868416, + 0.024999963119626045, + 0.016855977475643158, + 0.008658443577587605, + 0.013502232730388641, + 0.01493703480809927, + 0.008066866546869278, + 0.009014229290187359, + 0.02346121147274971, + 0.0680125281214714, + 0.005880800075829029, + 0.006357147358357906, + 0.037137601524591446, + 0.013414835557341576, + 0.003284517675638199, + 0.0045149545185267925, + 0.030559878796339035, + 0.03792613744735718, + 0.015013903379440308, + 0.018237939104437828, + 0.017661385238170624, + 0.012820539064705372, + 0.012571567669510841, + 0.012216725386679173, + 0.011871113441884518, + 0.008712389506399632, + 0.010412982665002346, + 0.00776437483727932, + 0.009723643772304058, + 0.011880123056471348, + 0.006918628234416246, + 0.010483250953257084, + 0.028859784826636314, + 0.014736838638782501, + 0.015981091186404228, + 0.010919504798948765, + 0.023381300270557404, + 0.03861216455698013, + 0.0162018034607172, + 0.024652093648910522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.048818159848451614 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001518169417977333 + ], + "opposite_logprob_tokens": [ + -6.501518249511719 + ], + "logprob_reward": -0.001518169417977333, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001518169417977333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12439101934432983, + 0.03934621438384056, + 0.24110862612724304, + 0.03009742870926857, + 0.02918391115963459, + 0.0470077283680439, + 0.08317582309246063, + 0.09155725687742233, + 0.054653193801641464, + 0.2594788074493408 + ], + "credit_map": [ + 0.027222037315368652, + 0.013527662493288517, + 0.011284018866717815, + 0.007693794090300798, + 0.02437507174909115, + 0.017208650708198547, + 0.009229368530213833, + 0.013850419782102108, + 0.006621606647968292, + 0.01207682304084301, + 0.004145212471485138, + 0.0019814956467598677, + 0.006025930866599083, + 0.004798638168722391, + 0.0015657109906896949, + 0.002130795968696475, + 0.03889201208949089, + 0.10433464497327805, + 0.008835204876959324, + 0.005560384131968021, + 0.056123148649930954, + 0.018515123054385185, + 0.00338941253721714, + 0.005458708852529526, + 0.003577601397410035, + 0.005310457199811935, + 0.0030987299978733063, + 0.002425228012725711, + 0.0038692469242960215, + 0.003530621062964201, + 0.002633375348523259, + 0.005652167834341526, + 0.005421169102191925, + 0.007980380207300186, + 0.001956698950380087, + 0.0027911148499697447, + 0.0031547003891319036, + 0.004330582916736603, + 0.0013902129139751196, + 0.0021590518299490213, + 0.0056978072971105576, + 0.011919224634766579, + 0.004714588169008493, + 0.0037509140092879534, + 0.007084882352501154, + 0.005632915534079075, + 0.0035611141938716173, + 0.004646283108741045, + 0.012547525577247143, + 0.033851105719804764, + 0.00327565660700202, + 0.0032959170639514923, + 0.017945043742656708, + 0.007465181406587362, + 0.0017821387154981494, + 0.003013252280652523, + 0.014046761207282543, + 0.017917998135089874, + 0.00843683909624815, + 0.007667967583984137, + 0.009846351109445095, + 0.00955641083419323, + 0.008784962818026543, + 0.015299967490136623, + 0.009039780125021935, + 0.006110978778451681, + 0.008464803919196129, + 0.004931765142828226, + 0.005722024943679571, + 0.0064529855735599995, + 0.005167402792721987, + 0.0087634501978755, + 0.04240156337618828, + 0.019986294209957123, + 0.026894301176071167, + 0.01740426756441593, + 0.03164041414856911, + 0.05724094808101654, + 0.02463858760893345, + 0.039272453635931015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.001518169417977333 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261507749557495 + ], + "opposite_logprob_tokens": [ + -0.5761507749557495 + ], + "logprob_reward": -0.8261507749557495, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261507749557495, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.059712838381528854, + 0.05829085409641266, + 0.11814319342374802, + 0.0333939865231514, + 0.01984878070652485, + 0.038045484572649, + 0.05420791730284691, + 0.12275592982769012, + 0.1034851148724556, + 0.39211589097976685 + ], + "credit_map": [ + 0.012921925634145737, + 0.01594661734998226, + 0.006209977902472019, + 0.004008390940725803, + 0.005989740137010813, + 0.006363607943058014, + 0.0035359449684619904, + 0.004736633040010929, + 0.009890632703900337, + 0.020817385986447334, + 0.004589694552123547, + 0.002607539063319564, + 0.009371242485940456, + 0.004973160102963448, + 0.0024008890613913536, + 0.0036403127014636993, + 0.018958033993840218, + 0.04948314651846886, + 0.005419990047812462, + 0.00342409685254097, + 0.025511078536510468, + 0.00951044075191021, + 0.002263234229758382, + 0.003573172725737095, + 0.004577150102704763, + 0.010917107574641705, + 0.002877206075936556, + 0.001972378697246313, + 0.006647171452641487, + 0.003013498382642865, + 0.0012767183361575007, + 0.002112756948918104, + 0.002567390911281109, + 0.006048442330211401, + 0.0019335809629410505, + 0.0018219791818410158, + 0.003556253155693412, + 0.002209733473137021, + 0.0009066453785635531, + 0.0008047556621022522, + 0.004673982039093971, + 0.0071653202176094055, + 0.005058248061686754, + 0.004281689878553152, + 0.004714834038168192, + 0.005172219593077898, + 0.003026080084964633, + 0.00395311089232564, + 0.00975638348609209, + 0.01934407465159893, + 0.0041853939183056355, + 0.0030825315043330193, + 0.00833264272660017, + 0.005150109529495239, + 0.001955972518771887, + 0.0024008096661418676, + 0.016594940796494484, + 0.023208601400256157, + 0.010939949192106724, + 0.009791761636734009, + 0.014494606293737888, + 0.01233166828751564, + 0.012767664156854153, + 0.022626744583249092, + 0.016939587891101837, + 0.01190178282558918, + 0.014827257953584194, + 0.007812336552888155, + 0.008689945563673973, + 0.014979023486375809, + 0.009653370827436447, + 0.018681809306144714, + 0.0622008815407753, + 0.03390222042798996, + 0.037604764103889465, + 0.02414287067949772, + 0.05165134370326996, + 0.08521567285060883, + 0.03859526291489601, + 0.058802876621484756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.8261507749557495 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, slightly high-angle close-up captures a stack of worn magazines, resting on the dark, polished wood of a dusty table. The top magazine, a vintage publication, features a faded, saturated cover and yellowed, dog-eared pages, tilted slightly relative to the magazines below. A warm, off-camera light source casts a focused pool of illumination, highlighting the fine layer of dust on the pages and casting the surrounding room into deep, blurred shadows, creating an atmosphere of tension and mystery. The scene is initially tranquil, but then suddenly disturbed as the stack begins to tremble, a low-frequency vibration at first almost imperceptible throughout. Then, the thin, fragile edges of the pages begin to ripple and flutter erratically, not from scratch, but as if stirred by an inner, unseen force. The trembling intensifies, causing the entire stack to vibrate more violently, individual pages colliding with each other in a silent, chaotic dance, suggesting architectural, supernatural presences.", + "seed": 10965, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_10/index_3632_seed_10965_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.31303953342139723, + "vlm_logprob": -0.31303953342139723, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.31303953342139723, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008276332169771194 + ], + "opposite_logprob_tokens": [ + -5.633276462554932 + ], + "logprob_reward": -0.008276332169771194, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.008276332169771194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1353081315755844, + 0.35254499316215515, + 0.03187097981572151, + 0.08741112798452377, + 0.02240634709596634, + 0.030093353241682053, + 0.04230264574289322, + 0.10290337353944778, + 0.13730092346668243, + 0.057858116924762726 + ], + "credit_map": [ + 0.02007872797548771, + 0.021466074511408806, + 0.02312937192618847, + 0.014401347376406193, + 0.01665395498275757, + 0.013244378380477428, + 0.011232108809053898, + 0.015102173201739788, + 0.06162434443831444, + 0.023370927199721336, + 0.14597062766551971, + 0.022044332697987556, + 0.017706066370010376, + 0.014417784288525581, + 0.017757099121809006, + 0.04965383559465408, + 0.004578215070068836, + 0.0019672096241265535, + 0.010719859972596169, + 0.0044784811325371265, + 0.0021264338865876198, + 0.002015703357756138, + 0.0018672241130843759, + 0.004117849748581648, + 0.0110841179266572, + 0.006803017109632492, + 0.015629123896360397, + 0.019224395975470543, + 0.010558337904512882, + 0.005598454270511866, + 0.008143173530697823, + 0.010370505973696709, + 0.003618404734879732, + 0.001438937266357243, + 0.005403426941484213, + 0.003017361741513014, + 0.00175651500467211, + 0.001553983660414815, + 0.002148456871509552, + 0.003469261107966304, + 0.0037403167225420475, + 0.0026541410479694605, + 0.006623092573136091, + 0.006212174892425537, + 0.003522263141348958, + 0.001819175435230136, + 0.002938356949016452, + 0.002583833411335945, + 0.005949553567916155, + 0.006850265897810459, + 0.006226520519703627, + 0.004042947664856911, + 0.0064964452758431435, + 0.00394695159047842, + 0.0035332366824150085, + 0.005256724078208208, + 0.011598465964198112, + 0.01257331669330597, + 0.022387929260730743, + 0.01333731971681118, + 0.011609906330704689, + 0.007754557300359011, + 0.01058120559900999, + 0.013060671277344227, + 0.01318679004907608, + 0.010500133037567139, + 0.03806295245885849, + 0.023511391133069992, + 0.008946272544562817, + 0.007078087422996759, + 0.021534614264965057, + 0.01448067743331194, + 0.007034303620457649, + 0.005288116633892059, + 0.009317202493548393, + 0.009340587072074413, + 0.0057055228389799595, + 0.0052214618772268295, + 0.009217127226293087, + 0.006733794696629047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.008276332169771194 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023305194452404976 + ], + "opposite_logprob_tokens": [ + -3.7733051776885986 + ], + "logprob_reward": -0.023305194452404976, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023305194452404976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5465540289878845, + 0.30482468008995056, + 0.012324796989560127, + 0.023904256522655487, + 0.008608274161815643, + 0.007703764829784632, + 0.02276979573071003, + 0.051604125648736954, + 0.014114098623394966, + 0.007592240814119577 + ], + "credit_map": [ + 0.10414935648441315, + 0.028233900666236877, + 0.23930443823337555, + 0.032354895025491714, + 0.018209727481007576, + 0.020625468343496323, + 0.025444569066166878, + 0.07823161780834198, + 0.05724451318383217, + 0.016525933519005775, + 0.1316205859184265, + 0.01784532703459263, + 0.011825747787952423, + 0.011869278736412525, + 0.014471711590886116, + 0.04342157393693924, + 0.001910752383992076, + 0.0008676864090375602, + 0.0015007705660536885, + 0.0026902088429778814, + 0.0015381780685856938, + 0.0006882571615278721, + 0.0015314114280045033, + 0.0015975313726812601, + 0.0037369192577898502, + 0.0024611211847513914, + 0.002639164449647069, + 0.004272415302693844, + 0.003633751766756177, + 0.0020694336853921413, + 0.0021463274024426937, + 0.002945119282230735, + 0.0013570968294516206, + 0.0008039885433390737, + 0.002485760021954775, + 0.0006704064435325563, + 0.0009838060941547155, + 0.00043319971882738173, + 0.0005840043886564672, + 0.0012900116853415966, + 0.0012171472189947963, + 0.0009415073436684906, + 0.0008874384220689535, + 0.0010909440461546183, + 0.0012466087937355042, + 0.000516359752509743, + 0.0006362589192576706, + 0.001167499227449298, + 0.0024940252769738436, + 0.002052491996437311, + 0.006493073888123035, + 0.003950679674744606, + 0.002269865944981575, + 0.0014524472644552588, + 0.0016203761333599687, + 0.002436833456158638, + 0.006022456102073193, + 0.005341785494238138, + 0.013042696751654148, + 0.006272851023823023, + 0.004911885596811771, + 0.003516171360388398, + 0.005360335577279329, + 0.007135943043977022, + 0.0017308855894953012, + 0.0012965666828677058, + 0.002737120958045125, + 0.0016453368589282036, + 0.0016450645634904504, + 0.0013399505987763405, + 0.0014539607800543308, + 0.002265211660414934, + 0.0011826882837340236, + 0.0005927010206505656, + 0.0010528082493692636, + 0.0012975558638572693, + 0.0007589984452351928, + 0.000590645126067102, + 0.0011451237369328737, + 0.0009717194479890168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.023305194452404976 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983557991683483 + ], + "opposite_logprob_tokens": [ + -3.5298354625701904 + ], + "logprob_reward": -0.02983557991683483, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983557991683483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3010003864765167, + 0.5189535617828369, + 0.027995767071843147, + 0.037848100066185, + 0.022658301517367363, + 0.017855273559689522, + 0.009551133960485458, + 0.01986856386065483, + 0.03350742161273956, + 0.010761477053165436 + ], + "credit_map": [ + 0.05460312217473984, + 0.018053654581308365, + 0.12734858691692352, + 0.018975136801600456, + 0.013965670019388199, + 0.01208681333810091, + 0.014482391066849232, + 0.0414850078523159, + 0.09710415452718735, + 0.02843533270061016, + 0.22309066355228424, + 0.03069530427455902, + 0.02042875997722149, + 0.020351257175207138, + 0.02467072755098343, + 0.0741773396730423, + 0.005503895692527294, + 0.0013755260733887553, + 0.012003218755126, + 0.0021330274175852537, + 0.0008644409826956689, + 0.0010799096198752522, + 0.001169736497104168, + 0.0038660136051476, + 0.00448190001770854, + 0.00294058327563107, + 0.007752419915050268, + 0.007672383449971676, + 0.004780783783644438, + 0.0024933400563895702, + 0.003205440938472748, + 0.004521247930824757, + 0.0027437033131718636, + 0.0016902689822018147, + 0.0030941700097173452, + 0.0058418032713234425, + 0.0024872191715985537, + 0.0011289190733805299, + 0.002948138862848282, + 0.0027240782510489225, + 0.0023449792061001062, + 0.0008798572234809399, + 0.007125732488930225, + 0.002234319457784295, + 0.0007098122150637209, + 0.0006346841691993177, + 0.0009746321011334658, + 0.0029512587934732437, + 0.001211634837090969, + 0.0005895362119190395, + 0.003507049987092614, + 0.001448405790142715, + 0.0005462261033244431, + 0.0004253467486705631, + 0.0006077729049138725, + 0.0012151618720963597, + 0.002363263862207532, + 0.0025944551452994347, + 0.002948419190943241, + 0.002158457413315773, + 0.002904837019741535, + 0.001658627181313932, + 0.002332699950784445, + 0.002907806308940053, + 0.00311731337569654, + 0.002603839850053191, + 0.010166672989726067, + 0.0044447146356105804, + 0.002717586001381278, + 0.0018459425773471594, + 0.005429284647107124, + 0.003182066837325692, + 0.0013385729398578405, + 0.0007598736556246877, + 0.0014105705777183175, + 0.0019551231525838375, + 0.001432881224900484, + 0.0008705973741598427, + 0.001703211572021246, + 0.0012906469637528062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.02983557991683483 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017160940915346146 + ], + "opposite_logprob_tokens": [ + -6.376716136932373 + ], + "logprob_reward": -0.0017160940915346146, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017160940915346146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7013834714889526, + 0.055791713297367096, + 0.017973145470023155, + 0.031979940831661224, + 0.0108930179849267, + 0.01549861766397953, + 0.012001188471913338, + 0.047802213579416275, + 0.08421370387077332, + 0.022462932392954826 + ], + "credit_map": [ + 0.13560408353805542, + 0.03497212007641792, + 0.30870774388313293, + 0.04052138328552246, + 0.024044092744588852, + 0.026630103588104248, + 0.03217324987053871, + 0.09873075038194656, + 0.009845306165516376, + 0.003847323125228286, + 0.020187921822071075, + 0.003264510305598378, + 0.004318333696573973, + 0.00310027995146811, + 0.0031267742160707712, + 0.008101268671452999, + 0.002432611770927906, + 0.0013406574726104736, + 0.005278099328279495, + 0.0012611724669113755, + 0.0023638932034373283, + 0.0010649211471900344, + 0.0012452179798856378, + 0.0029865731485188007, + 0.0037886544596403837, + 0.002629334107041359, + 0.006004065275192261, + 0.0065655456855893135, + 0.004386201500892639, + 0.0022000845056027174, + 0.0027244838420301676, + 0.0036815751809626818, + 0.0012587410164996982, + 0.0009412824874743819, + 0.0011012089671567082, + 0.002391349757090211, + 0.0017638691933825612, + 0.0007311676163226366, + 0.0010075977770611644, + 0.0016978015191853046, + 0.001515501644462347, + 0.0010266462340950966, + 0.004826532676815987, + 0.0032941033132374287, + 0.0008510757470503449, + 0.0006699824007228017, + 0.0012082747416570783, + 0.002106502652168274, + 0.0013246689923107624, + 0.0009467875934205949, + 0.0026639304123818874, + 0.002548626158386469, + 0.000996965216472745, + 0.0007605609716847539, + 0.0011113079963251948, + 0.0016483422368764877, + 0.005973469465970993, + 0.0036067792680114508, + 0.01181626133620739, + 0.008675514720380306, + 0.005826044827699661, + 0.0031966252718120813, + 0.003939375746995211, + 0.004768148064613342, + 0.00849622767418623, + 0.006792624481022358, + 0.023085491731762886, + 0.012487372383475304, + 0.006769115570932627, + 0.0051943338476121426, + 0.012107624672353268, + 0.009280926547944546, + 0.0030141621828079224, + 0.0021270625293254852, + 0.003484823741018772, + 0.004072931595146656, + 0.001984151778742671, + 0.0018835472874343395, + 0.0036075860261917114, + 0.00228867051191628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0017160940915346146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5020644664764404 + ], + "opposite_logprob_tokens": [ + -0.2520644962787628 + ], + "logprob_reward": -1.5020644664764404, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5020644664764404, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2841137647628784, + 0.297476589679718, + 0.05701132491230965, + 0.053057897835969925, + 0.024462983012199402, + 0.023125246167182922, + 0.028892287984490395, + 0.0966203361749649, + 0.10878821462392807, + 0.02645137906074524 + ], + "credit_map": [ + 0.05559857562184334, + 0.013696088455617428, + 0.12557128071784973, + 0.016493046656250954, + 0.008511287160217762, + 0.010789819061756134, + 0.012794509530067444, + 0.040659159421920776, + 0.05532912537455559, + 0.016511185094714165, + 0.1275770664215088, + 0.017048832029104233, + 0.012040993198752403, + 0.011599252000451088, + 0.014362065121531487, + 0.043008070439100266, + 0.011957809329032898, + 0.0031031451653689146, + 0.02345220372080803, + 0.003008755389600992, + 0.002293178578838706, + 0.0028405410703271627, + 0.002625813474878669, + 0.007729882374405861, + 0.006128806620836258, + 0.0030284407548606396, + 0.017168791964650154, + 0.010064354166388512, + 0.004590112250298262, + 0.0023036820348352194, + 0.0036392156034708023, + 0.00613449839875102, + 0.0038288747891783714, + 0.001489305286668241, + 0.008431521244347095, + 0.00356367789208889, + 0.0017724087228998542, + 0.0011364301899448037, + 0.0016177432844415307, + 0.0026230216026306152, + 0.0038510113954544067, + 0.002285785274580121, + 0.005779137834906578, + 0.0025320600252598524, + 0.002549579832702875, + 0.001481919432990253, + 0.0019654768984764814, + 0.0026802769862115383, + 0.0037587580736726522, + 0.0035720732994377613, + 0.003241404891014099, + 0.006074910517781973, + 0.0034621579106897116, + 0.0018798039527609944, + 0.003342526266351342, + 0.0035606548190116882, + 0.012407684698700905, + 0.01210462860763073, + 0.017287680879235268, + 0.01217302680015564, + 0.01673915982246399, + 0.008384943008422852, + 0.007743386086076498, + 0.009779826737940311, + 0.010490705259144306, + 0.008151481859385967, + 0.032442644238471985, + 0.014668751507997513, + 0.007549321744590998, + 0.006359652150422335, + 0.01664845459163189, + 0.012477206066250801, + 0.002908285241574049, + 0.00247956276871264, + 0.0037621415685862303, + 0.004781005438417196, + 0.0031879637390375137, + 0.002902260748669505, + 0.00378434918820858, + 0.0026458108332008123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -1.5020644664764404 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, slightly high-angle close-up captures a stack of worn magazines, resting on the dark, polished wood of a dusty table. The top magazine, a vintage publication, features a faded, saturated cover and yellowed, dog-eared pages, tilted slightly relative to the magazines below. A warm, off-camera light source casts a focused pool of illumination, highlighting the fine layer of dust on the pages and casting the surrounding room into deep, blurred shadows, creating an atmosphere of tension and mystery. The scene is initially tranquil, but then suddenly disturbed as the stack begins to tremble, a low-frequency vibration at first almost imperceptible throughout. Then, the thin, fragile edges of the pages begin to ripple and flutter erratically, not from scratch, but as if stirred by an inner, unseen force. The trembling intensifies, causing the entire stack to vibrate more violently, individual pages colliding with each other in a silent, chaotic dance, suggesting architectural, supernatural presences.", + "seed": 10966, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_10/index_3632_seed_10966_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.39436296331696213, + "vlm_logprob": -0.39436296331696213, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.39436296331696213, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007844948209822178 + ], + "opposite_logprob_tokens": [ + -5.757844924926758 + ], + "logprob_reward": -0.007844948209822178, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007844948209822178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1374368816614151, + 0.07440896332263947, + 0.06506205350160599, + 0.13211287558078766, + 0.18609337508678436, + 0.08957710862159729, + 0.08267472684383392, + 0.04845234751701355, + 0.1203090101480484, + 0.06387260556221008 + ], + "credit_map": [ + 0.022640766575932503, + 0.013911789283156395, + 0.02928888611495495, + 0.011850034818053246, + 0.014330173842608929, + 0.016339940950274467, + 0.00915917195379734, + 0.019916130229830742, + 0.010314948856830597, + 0.006652826443314552, + 0.0136678796261549, + 0.007725478149950504, + 0.012674455530941486, + 0.006912783719599247, + 0.0064404853619635105, + 0.010020110756158829, + 0.009207838214933872, + 0.005548864137381315, + 0.007709259167313576, + 0.005497851874679327, + 0.01501511037349701, + 0.007156923413276672, + 0.006647472735494375, + 0.0082787424325943, + 0.009824569337069988, + 0.012426365166902542, + 0.011737710796296597, + 0.01821167767047882, + 0.0340522900223732, + 0.011602970771491528, + 0.017627384513616562, + 0.016629917547106743, + 0.019495226442813873, + 0.018708497285842896, + 0.01764906384050846, + 0.03451072424650192, + 0.029428955167531967, + 0.015762396156787872, + 0.027940068393945694, + 0.022598475217819214, + 0.006834540981799364, + 0.006964688189327717, + 0.006832588464021683, + 0.012545845471322536, + 0.012695611454546452, + 0.007506805472075939, + 0.02290646731853485, + 0.013290569186210632, + 0.010770324617624283, + 0.00767520209774375, + 0.009449909441173077, + 0.011836758814752102, + 0.01215345598757267, + 0.007476703263819218, + 0.011360418982803822, + 0.011951957829296589, + 0.004986668936908245, + 0.004206268582493067, + 0.005720558576285839, + 0.006822109222412109, + 0.00809620600193739, + 0.0060686031356453896, + 0.006693489849567413, + 0.00585844786837697, + 0.011832298710942268, + 0.014381390996277332, + 0.01180217880755663, + 0.012523854151368141, + 0.01872308738529682, + 0.020456187427043915, + 0.015311372466385365, + 0.01527866069227457, + 0.00747067853808403, + 0.004688786342740059, + 0.011150459758937359, + 0.008050546050071716, + 0.005561126861721277, + 0.006391371600329876, + 0.011213155463337898, + 0.009346485137939453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.007844948209822178 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821024715900421 + ], + "opposite_logprob_tokens": [ + -4.018210411071777 + ], + "logprob_reward": -0.01821024715900421, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01821024715900421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1810641586780548, + 0.09095898270606995, + 0.10037592053413391, + 0.12118265777826309, + 0.08187133073806763, + 0.054768119007349014, + 0.06516188383102417, + 0.06553231179714203, + 0.05743623897433281, + 0.181648388504982 + ], + "credit_map": [ + 0.025680579245090485, + 0.01734963431954384, + 0.04437054693698883, + 0.014841522090137005, + 0.022213390097022057, + 0.01899321936070919, + 0.013508972711861134, + 0.02410629577934742, + 0.014042913913726807, + 0.009846693836152554, + 0.013795227743685246, + 0.007780755870044231, + 0.01380383875221014, + 0.011364119127392769, + 0.007254041265696287, + 0.013071393594145775, + 0.012631882913410664, + 0.010516569949686527, + 0.017298882827162743, + 0.011417316272854805, + 0.01562810130417347, + 0.012287110090255737, + 0.007501778658479452, + 0.013094273395836353, + 0.012004561722278595, + 0.00924598053097725, + 0.013478056527674198, + 0.023601626977324486, + 0.014814765192568302, + 0.009515784680843353, + 0.02287832833826542, + 0.01564355380833149, + 0.007851808331906796, + 0.01047325599938631, + 0.0066741048358380795, + 0.010275062173604965, + 0.016523199155926704, + 0.008703820407390594, + 0.010094103403389454, + 0.011275982484221458, + 0.00410849042236805, + 0.006246353965252638, + 0.004385481122881174, + 0.006286167073994875, + 0.008533662185072899, + 0.0068318601697683334, + 0.011022942140698433, + 0.007353162858635187, + 0.005209358874708414, + 0.007280425634235144, + 0.007962167263031006, + 0.008678516373038292, + 0.007776197977364063, + 0.0071659698151052, + 0.012662303633987904, + 0.008426944725215435, + 0.008811862207949162, + 0.007159436121582985, + 0.007812035270035267, + 0.006647946313023567, + 0.008954053744673729, + 0.007453668396919966, + 0.00998294074088335, + 0.008710368536412716, + 0.0053773061372339725, + 0.007747076917439699, + 0.010462632402777672, + 0.005683927796781063, + 0.007334765512496233, + 0.006420993711799383, + 0.006914488039910793, + 0.007495050318539143, + 0.011572121642529964, + 0.010886565782129765, + 0.012874607928097248, + 0.022522391751408577, + 0.023076042532920837, + 0.018572887405753136, + 0.05334680154919624, + 0.0287969708442688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.01821024715900421 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033729564398527145 + ], + "opposite_logprob_tokens": [ + -3.4087295532226562 + ], + "logprob_reward": -0.033729564398527145, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033729564398527145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17088744044303894, + 0.07987677305936813, + 0.05552546679973602, + 0.14536741375923157, + 0.2667047083377838, + 0.034437961876392365, + 0.06966698914766312, + 0.04896767809987068, + 0.045993611216545105, + 0.08257199078798294 + ], + "credit_map": [ + 0.020887218415737152, + 0.016503551974892616, + 0.02094191685318947, + 0.013923855498433113, + 0.03949643671512604, + 0.018040485680103302, + 0.019540855661034584, + 0.021553119644522667, + 0.0072877840138971806, + 0.007520326878875494, + 0.009217350743710995, + 0.008036166429519653, + 0.02155313268303871, + 0.007730951998382807, + 0.008945991285145283, + 0.009585070423781872, + 0.00660480372607708, + 0.005501473788172007, + 0.009090980514883995, + 0.007657091598957777, + 0.007862298749387264, + 0.004964063409715891, + 0.006427196320146322, + 0.007417554035782814, + 0.010116808116436005, + 0.014608525671064854, + 0.011086832731962204, + 0.016228117048740387, + 0.042165420949459076, + 0.011552159674465656, + 0.01968722976744175, + 0.019922323524951935, + 0.025800105184316635, + 0.025250231847167015, + 0.027630966156721115, + 0.0535641573369503, + 0.03288524970412254, + 0.017194170504808426, + 0.05245961248874664, + 0.03192019835114479, + 0.003337165340781212, + 0.003799464786425233, + 0.0033174443524330854, + 0.0038613893557339907, + 0.007280050776898861, + 0.003294296795502305, + 0.004785394761711359, + 0.004762756172567606, + 0.00843728892505169, + 0.007850793190300465, + 0.007634023204445839, + 0.009602262638509274, + 0.008784844540059566, + 0.006030119024217129, + 0.011210709810256958, + 0.01011695060878992, + 0.006757368333637714, + 0.005512780975550413, + 0.006131850183010101, + 0.00537498202174902, + 0.007814827375113964, + 0.005060581024736166, + 0.005876654759049416, + 0.006438629701733589, + 0.005231660790741444, + 0.006330483593046665, + 0.005337402690201998, + 0.004682807717472315, + 0.008010619319975376, + 0.005263009108603001, + 0.005147866439074278, + 0.005989760626107454, + 0.004311730619519949, + 0.004471030551940203, + 0.005090861115604639, + 0.011060132645070553, + 0.010611051693558693, + 0.007915372960269451, + 0.025044318288564682, + 0.014067494310438633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.033729564398527145 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001716808183118701 + ], + "opposite_logprob_tokens": [ + -6.376716613769531 + ], + "logprob_reward": -0.001716808183118701, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001716808183118701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11071323603391647, + 0.03729942813515663, + 0.06135017052292824, + 0.08994735777378082, + 0.1331738531589508, + 0.07804962247610092, + 0.09211191534996033, + 0.19237686693668365, + 0.11904624849557877, + 0.08593124151229858 + ], + "credit_map": [ + 0.016574813053011894, + 0.010198080912232399, + 0.02606201358139515, + 0.009396940469741821, + 0.017194805666804314, + 0.010017783381044865, + 0.008412704803049564, + 0.012856095097959042, + 0.003987928852438927, + 0.003664815565571189, + 0.00592798413708806, + 0.004908877890557051, + 0.0063391816802322865, + 0.0042844791896641254, + 0.0036743462551385164, + 0.004511815495789051, + 0.006553951185196638, + 0.0056292228400707245, + 0.00915589276701212, + 0.007079202216118574, + 0.012267231941223145, + 0.006428959779441357, + 0.00570489140227437, + 0.008530816994607449, + 0.008746927604079247, + 0.008354801684617996, + 0.009925417602062225, + 0.014504163525998592, + 0.014905623160302639, + 0.009151214733719826, + 0.012632029131054878, + 0.011727187782526016, + 0.015233143232762814, + 0.015028721652925014, + 0.01450394093990326, + 0.020811278373003006, + 0.018260452896356583, + 0.012569407932460308, + 0.016934886574745178, + 0.019832035526633263, + 0.004955544602125883, + 0.007478907238692045, + 0.007178650703281164, + 0.009093006141483784, + 0.014319361187517643, + 0.007492183707654476, + 0.017094554379582405, + 0.010437420569360256, + 0.010236617177724838, + 0.010607919655740261, + 0.010354728437960148, + 0.01246601901948452, + 0.010513465851545334, + 0.009913589805364609, + 0.015469292178750038, + 0.012550285086035728, + 0.024305740371346474, + 0.020972242578864098, + 0.025043407455086708, + 0.020983919501304626, + 0.025347769260406494, + 0.023265596479177475, + 0.026192141696810722, + 0.026266047731041908, + 0.01867211051285267, + 0.013483692891895771, + 0.016426511108875275, + 0.011695193126797676, + 0.015240968205034733, + 0.011784125119447708, + 0.014996756799519062, + 0.016746893525123596, + 0.005555374082177877, + 0.00454495195299387, + 0.008688749745488167, + 0.010015054605901241, + 0.00803246721625328, + 0.008939629420638084, + 0.026200490072369576, + 0.013954522088170052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.001716808183118701 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9103132486343384 + ], + "opposite_logprob_tokens": [ + -0.16031326353549957 + ], + "logprob_reward": -1.9103132486343384, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9103132486343384, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14394812285900116, + 0.041952166706323624, + 0.056010253727436066, + 0.09310001134872437, + 0.08885501325130463, + 0.046268709003925323, + 0.08705349266529083, + 0.19264096021652222, + 0.19844496250152588, + 0.05172627046704292 + ], + "credit_map": [ + 0.019096340984106064, + 0.01333559025079012, + 0.01962083950638771, + 0.014800948090851307, + 0.028569012880325317, + 0.01715620793402195, + 0.01226795930415392, + 0.019101228564977646, + 0.0046235802583396435, + 0.0038219974376261234, + 0.005171090830117464, + 0.004308733157813549, + 0.009737526997923851, + 0.004094884730875492, + 0.004486048128455877, + 0.0057083056308329105, + 0.006402174476534128, + 0.005241819657385349, + 0.005975512787699699, + 0.005315213929861784, + 0.013640327379107475, + 0.005688869394361973, + 0.005509223323315382, + 0.008237110450863838, + 0.010443483479321003, + 0.009212492033839226, + 0.012140423990786076, + 0.012144060805439949, + 0.015848059207201004, + 0.008828643709421158, + 0.011518309824168682, + 0.012964535504579544, + 0.01211307942867279, + 0.008660669438540936, + 0.013979829847812653, + 0.009676555171608925, + 0.015212452970445156, + 0.007972799241542816, + 0.008539877831935883, + 0.012699754908680916, + 0.004363006446510553, + 0.004806261043995619, + 0.005260179750621319, + 0.0047788177616894245, + 0.008372635580599308, + 0.005839372053742409, + 0.006437062751501799, + 0.0064113750122487545, + 0.009613930247724056, + 0.008709322661161423, + 0.009697639383375645, + 0.011304634623229504, + 0.011769291944801807, + 0.010521982796490192, + 0.013525096699595451, + 0.011911598034203053, + 0.02464378997683525, + 0.022081783041357994, + 0.02548329532146454, + 0.02137904427945614, + 0.024073952808976173, + 0.021409841254353523, + 0.026527229696512222, + 0.027042031288146973, + 0.014797475188970566, + 0.022050216794013977, + 0.020486433058977127, + 0.020246220752596855, + 0.039643097668886185, + 0.03365963697433472, + 0.02268480882048607, + 0.024877075105905533, + 0.005370932165533304, + 0.00541815347969532, + 0.0072645219042897224, + 0.00530956219881773, + 0.0061121741309762, + 0.008482159115374088, + 0.007093838881701231, + 0.0066749332472682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -1.9103132486343384 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, slightly high-angle close-up captures a stack of worn magazines, resting on the dark, polished wood of a dusty table. The top magazine, a vintage publication, features a faded, saturated cover and yellowed, dog-eared pages, tilted slightly relative to the magazines below. A warm, off-camera light source casts a focused pool of illumination, highlighting the fine layer of dust on the pages and casting the surrounding room into deep, blurred shadows, creating an atmosphere of tension and mystery. The scene is initially tranquil, but then suddenly disturbed as the stack begins to tremble, a low-frequency vibration at first almost imperceptible throughout. Then, the thin, fragile edges of the pages begin to ripple and flutter erratically, not from scratch, but as if stirred by an inner, unseen force. The trembling intensifies, causing the entire stack to vibrate more violently, individual pages colliding with each other in a silent, chaotic dance, suggesting architectural, supernatural presences.", + "seed": 10967, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_10/index_3632_seed_10967_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.029783375025726855, + "vlm_logprob": -0.029783375025726855, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.029783375025726855, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006986474618315697 + ], + "opposite_logprob_tokens": [ + -6.381986618041992 + ], + "logprob_reward": -0.006986474618315697, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.006986474618315697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09257540106773376, + 0.033366307616233826, + 0.024906061589717865, + 0.036046937108039856, + 0.04371403157711029, + 0.037990812212228775, + 0.07364258915185928, + 0.10627631098031998, + 0.1607578843832016, + 0.3907236158847809 + ], + "credit_map": [ + 0.011880147270858288, + 0.007876163348555565, + 0.010541476309299469, + 0.014395548030734062, + 0.015900734812021255, + 0.01238554622977972, + 0.009795811958611012, + 0.009799971245229244, + 0.0048442985862493515, + 0.002387160900980234, + 0.005515023600310087, + 0.0031602822709828615, + 0.0039698719047009945, + 0.004862541798502207, + 0.003339450340718031, + 0.005287679377943277, + 0.0043656593188643456, + 0.0023408338893204927, + 0.004466939717531204, + 0.001879282295703888, + 0.003848097287118435, + 0.0033222297206521034, + 0.002091237111017108, + 0.0025917822495102882, + 0.005220730323344469, + 0.0037200001534074545, + 0.006194074638187885, + 0.0031333514489233494, + 0.0054601398296654224, + 0.004461667500436306, + 0.003243781393393874, + 0.004613193683326244, + 0.005971487145870924, + 0.004238148219883442, + 0.007069684565067291, + 0.003673231927677989, + 0.006262443959712982, + 0.005886612460017204, + 0.0050959596410393715, + 0.00551646389067173, + 0.00514061376452446, + 0.0030844262801110744, + 0.004874293226748705, + 0.00419611157849431, + 0.005963678006082773, + 0.005440902430564165, + 0.004105761647224426, + 0.005185021553188562, + 0.007670860271900892, + 0.006550501100718975, + 0.011194110848009586, + 0.007071002386510372, + 0.009777306579053402, + 0.011046722531318665, + 0.01014673337340355, + 0.010185351595282555, + 0.009285303764045238, + 0.021933216601610184, + 0.01193043403327465, + 0.0086968457326293, + 0.016433915123343468, + 0.016605909913778305, + 0.012283558025956154, + 0.009107125923037529, + 0.014831915497779846, + 0.026756323873996735, + 0.027296360582113266, + 0.01362274307757616, + 0.017394553869962692, + 0.02224501594901085, + 0.01991690881550312, + 0.018694061785936356, + 0.07259739190340042, + 0.0322134830057621, + 0.04464040696620941, + 0.06500429660081863, + 0.037789467722177505, + 0.056079816073179245, + 0.03800784796476364, + 0.04439091309905052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.006986474618315697 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014216925017535686 + ], + "opposite_logprob_tokens": [ + -4.264216899871826 + ], + "logprob_reward": -0.014216925017535686, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014216925017535686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054025910794734955, + 0.029892178252339363, + 0.036633457988500595, + 0.05643932893872261, + 0.06914305686950684, + 0.052868153899908066, + 0.43279850482940674, + 0.1329236775636673, + 0.07638008892536163, + 0.058895595371723175 + ], + "credit_map": [ + 0.006439533084630966, + 0.004666777327656746, + 0.009533785283565521, + 0.005462224595248699, + 0.007634282112121582, + 0.007062811404466629, + 0.006794452667236328, + 0.006432048976421356, + 0.0037081786431372166, + 0.0021915840916335583, + 0.004120413213968277, + 0.0025414919946342707, + 0.0050081610679626465, + 0.004037835635244846, + 0.003751761047169566, + 0.0045327539555728436, + 0.004230116959661245, + 0.0023104255087673664, + 0.0064789904281497, + 0.0036685019731521606, + 0.005059398245066404, + 0.004991680383682251, + 0.003680784022435546, + 0.006213563494384289, + 0.006653457880020142, + 0.00445113331079483, + 0.011171415448188782, + 0.004946739412844181, + 0.008088392205536366, + 0.007614747621119022, + 0.0039855013601481915, + 0.009527944028377533, + 0.008527758531272411, + 0.0046973880380392075, + 0.013906557112932205, + 0.007292230613529682, + 0.008503688499331474, + 0.0098151545971632, + 0.006300781387835741, + 0.01009950041770935, + 0.006903957575559616, + 0.004418105352669954, + 0.009657292626798153, + 0.004191297106444836, + 0.006961464416235685, + 0.008218961767852306, + 0.004695176612585783, + 0.007821901701390743, + 0.05094509571790695, + 0.026729991659522057, + 0.07496259361505508, + 0.023851215839385986, + 0.0466642901301384, + 0.07428994029760361, + 0.04637829214334488, + 0.08897709101438522, + 0.01204218715429306, + 0.021306835114955902, + 0.017268821597099304, + 0.008960690349340439, + 0.018255217000842094, + 0.016554560512304306, + 0.013276170939207077, + 0.02525920234620571, + 0.006069331429898739, + 0.010200749151408672, + 0.012855702079832554, + 0.008082060143351555, + 0.008391307666897774, + 0.011365204118192196, + 0.010595028288662434, + 0.008820708841085434, + 0.007455556187778711, + 0.007748703006654978, + 0.007352887652814388, + 0.009007817134261131, + 0.007509744260460138, + 0.009084908291697502, + 0.00507650850340724, + 0.005659470800310373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.014216925017535686 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637348137795925 + ], + "opposite_logprob_tokens": [ + -3.6513733863830566 + ], + "logprob_reward": -0.02637348137795925, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637348137795925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11304133385419846, + 0.06887675821781158, + 0.04351731762290001, + 0.028335409238934517, + 0.03927719593048096, + 0.027586109936237335, + 0.10983458161354065, + 0.08103601634502411, + 0.09663566201925278, + 0.39185962080955505 + ], + "credit_map": [ + 0.013914747163653374, + 0.00953865796327591, + 0.015301531180739403, + 0.014906201511621475, + 0.018496671691536903, + 0.013909639790654182, + 0.015349301509559155, + 0.011624583043158054, + 0.008391750045120716, + 0.007218678016215563, + 0.009577921591699123, + 0.009771640412509441, + 0.009855611249804497, + 0.008641079068183899, + 0.008425278589129448, + 0.006994795054197311, + 0.00531262531876564, + 0.004834676627069712, + 0.006310595665127039, + 0.004233088344335556, + 0.007454578764736652, + 0.0059221419505774975, + 0.004791280720382929, + 0.004658326972275972, + 0.0037518206518143415, + 0.00216893432661891, + 0.00542877335101366, + 0.0027183236088603735, + 0.0037632621824741364, + 0.0047138649970293045, + 0.0022899252362549305, + 0.003500506514683366, + 0.005519639700651169, + 0.0029734033159911633, + 0.004048025235533714, + 0.005750046111643314, + 0.004597661085426807, + 0.006202150601893663, + 0.006520531140267849, + 0.00366573897190392, + 0.003493883879855275, + 0.0031844794284552336, + 0.003667039331048727, + 0.003153150901198387, + 0.00400494085624814, + 0.0033700186759233475, + 0.004487957339733839, + 0.0022246420849114656, + 0.012771712616086006, + 0.006016389466822147, + 0.01936696656048298, + 0.006086363457143307, + 0.011220133863389492, + 0.018819278106093407, + 0.011367952451109886, + 0.02418578788638115, + 0.0077600376680493355, + 0.010177984833717346, + 0.010068063624203205, + 0.005939799826592207, + 0.009239772334694862, + 0.011828913353383541, + 0.010174370370805264, + 0.01584707200527191, + 0.008987064473330975, + 0.01626228727400303, + 0.015927869826555252, + 0.008282828144729137, + 0.010559851303696632, + 0.011857482604682446, + 0.012182404287159443, + 0.012575879693031311, + 0.07894793152809143, + 0.0364355742931366, + 0.046583015471696854, + 0.0595843531191349, + 0.03234434872865677, + 0.05591294914484024, + 0.03478432074189186, + 0.04726715013384819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.02637348137795925 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010444429935887456 + ], + "opposite_logprob_tokens": [ + -6.876044273376465 + ], + "logprob_reward": -0.0010444429935887456, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010444429935887456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0807456523180008, + 0.042825765907764435, + 0.030968645587563515, + 0.020164713263511658, + 0.03610323369503021, + 0.031353794038295746, + 0.12266392260789871, + 0.21597008407115936, + 0.13308405876159668, + 0.28612014651298523 + ], + "credit_map": [ + 0.00856329221278429, + 0.006414563395082951, + 0.009984930977225304, + 0.009791080839931965, + 0.015193202532827854, + 0.012010403908789158, + 0.009594038128852844, + 0.009194140322506428, + 0.005802772007882595, + 0.003230416914448142, + 0.004614363424479961, + 0.007910136133432388, + 0.006138375960290432, + 0.005350664258003235, + 0.005544002167880535, + 0.004235032945871353, + 0.0036705711390823126, + 0.0023167904000729322, + 0.0042825243435800076, + 0.0031796011608093977, + 0.005626442842185497, + 0.004332190845161676, + 0.0033146515488624573, + 0.004245872609317303, + 0.0023597420658916235, + 0.0017521384870633483, + 0.0033210262190550566, + 0.002971329726278782, + 0.0021919971331954002, + 0.0027124087791889906, + 0.00218774052336812, + 0.002668331377208233, + 0.004417729564011097, + 0.0023300210013985634, + 0.006232939660549164, + 0.002755132270976901, + 0.0055749686434865, + 0.005946366116404533, + 0.0035269358195364475, + 0.00531914085149765, + 0.003725763875991106, + 0.0022859135642647743, + 0.004458426497876644, + 0.0028512089047580957, + 0.004682968836277723, + 0.005069341044872999, + 0.003136777551844716, + 0.005143392831087112, + 0.01566675677895546, + 0.00895664468407631, + 0.022116634994745255, + 0.007011662237346172, + 0.012618424370884895, + 0.020352192223072052, + 0.011458530090749264, + 0.024483075365424156, + 0.018250280991196632, + 0.040725428611040115, + 0.02873711660504341, + 0.014769189059734344, + 0.025747422128915787, + 0.02852931059896946, + 0.01923208497464657, + 0.039979249238967896, + 0.009990639053285122, + 0.02693220041692257, + 0.015904219821095467, + 0.012164331041276455, + 0.020995348691940308, + 0.01789007894694805, + 0.016299203038215637, + 0.01290803961455822, + 0.05776194483041763, + 0.030205534771084785, + 0.03349636122584343, + 0.04251573979854584, + 0.027367113158106804, + 0.04037364199757576, + 0.02316240407526493, + 0.031237414106726646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.0010444429935887456 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle close-up of a stack of worn magazines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the magazines resting on a dark, polished wood table covered in a fine layer of dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the top magazine a vintage publication with a faded, saturated cover and yellowed, dog-eared pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm, off-camera light source creating a focused pool of illumination and deep, blurred shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stack of magazines begin to tremble and vibrate violently, causing pages to ripple and flutter erratically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1002955511212349 + ], + "opposite_logprob_tokens": [ + -2.3502955436706543 + ], + "logprob_reward": -0.1002955511212349, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1002955511212349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04588289558887482, + 0.03187824785709381, + 0.026158073917031288, + 0.02488940954208374, + 0.03640711307525635, + 0.024607233703136444, + 0.17270001769065857, + 0.21547213196754456, + 0.2440812885761261, + 0.17792350053787231 + ], + "credit_map": [ + 0.00522985914722085, + 0.004104042891412973, + 0.005763730965554714, + 0.007162913680076599, + 0.0066249556839466095, + 0.006656578741967678, + 0.00598505511879921, + 0.004355764947831631, + 0.004886343143880367, + 0.0032314348500221968, + 0.003873024368658662, + 0.0032401159405708313, + 0.0044621992856264114, + 0.005143717397004366, + 0.003966185264289379, + 0.003075231332331896, + 0.004374598152935505, + 0.002281505148857832, + 0.003413761965930462, + 0.003199944971129298, + 0.003328996244817972, + 0.004127752035856247, + 0.003163472982123494, + 0.0022680461406707764, + 0.0032694630790501833, + 0.001914649154059589, + 0.002719723619520664, + 0.003591971704736352, + 0.0037628221325576305, + 0.003612224478274584, + 0.002932619769126177, + 0.0030859378166496754, + 0.005480007268488407, + 0.0036893393844366074, + 0.00519998325034976, + 0.0042726341634988785, + 0.0049679819494485855, + 0.00627145916223526, + 0.003427237970754504, + 0.003098472021520138, + 0.0035284077748656273, + 0.002032478107139468, + 0.0028226759750396013, + 0.0029004213865846395, + 0.0026285923086106777, + 0.0040367902256548405, + 0.003051929408684373, + 0.0036059406120330095, + 0.01695198565721512, + 0.013571354560554028, + 0.028348704800009727, + 0.012844547629356384, + 0.01947413943707943, + 0.02673863247036934, + 0.022773059085011482, + 0.03199762478470802, + 0.01862550526857376, + 0.03314860165119171, + 0.025765353813767433, + 0.01786087267100811, + 0.029579132795333862, + 0.03003775142133236, + 0.020912131294608116, + 0.03954281657934189, + 0.018405331298708916, + 0.04695802181959152, + 0.03450644016265869, + 0.02074548602104187, + 0.03679155185818672, + 0.032145511358976364, + 0.028792601078748703, + 0.025736378505825996, + 0.030955392867326736, + 0.02561642788350582, + 0.019841104745864868, + 0.025081878527998924, + 0.021717451512813568, + 0.025460973381996155, + 0.01372187864035368, + 0.015528411604464054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.1002955511212349 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_11/index_tensor([7140, 7140])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_11/index_tensor([7140, 7140])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..740b51ee9412bc2c79095438f7e9a0f15bb43df4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_11/index_tensor([7140, 7140])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in a surreal style, this slow, orbiting medium shot smoothly orbits a woman going about her morning activities in a lush garden at dawn. She has long, dark brown hair, loosely tied in a high ponytail, with a few stray strands gently framing her face. Her eyes are gently closed, her expression serene and focused. She wears a fitted, long-sleeved heather-gray yoga top, the soft, breathable fabric subtly textured, paired with high-waisted, dark teal leggings. Her body gracefully stands in side bend; her right arm extends deliberately and fluidly upward and over her head, elongating her torso, while her left hand rests lightly on her hip. The entire pose maintains a steady, controlled grace. The camera's slow, circular path reveals the immersive environment: she stands barefoot on a manicured lawn, still damp with morning dew. The garden is filled with vibrant green ferns and blooming rose bushes with soft, pink and white petals. The first rays of warm, golden sunlight penetrate the light morning mist between the leaves of the tall, mature trees in the background, backlighting her body and capturing millions of tiny, shimmering dewdrops on the grass and leaves, creating a shimmering bokeh effect. The interplay of warm, low-angle light and cool, soft shadows creates a serene scene of profound calm and focus, and the camera movement enhances the contemplative and visually rich atmosphere.", + "seed": 22005, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_11/index_7140_seed_22005_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0463353767991066, + "vlm_logprob": -0.0463353767991066, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.0463353767991066, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013890853151679039 + ], + "opposite_logprob_tokens": [ + -4.638890743255615 + ], + "logprob_reward": -0.013890853151679039, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013890853151679039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22117438912391663, + 0.09632473438978195, + 0.15618282556533813, + 0.0515531450510025, + 0.20169280469417572, + 0.07695746421813965, + 0.07027655094861984, + 0.044841162860393524, + 0.04029088094830513, + 0.040705978870391846 + ], + "credit_map": [ + 0.021881520748138428, + 0.028974497690796852, + 0.05253703519701958, + 0.018289461731910706, + 0.022481946274638176, + 0.023305827751755714, + 0.023311834782361984, + 0.03039228729903698, + 0.005693477112799883, + 0.008091818541288376, + 0.029064374044537544, + 0.008159306831657887, + 0.007947959005832672, + 0.019059160724282265, + 0.006491483189165592, + 0.011817163787782192, + 0.011834226548671722, + 0.020885908976197243, + 0.015670720487833023, + 0.02047746628522873, + 0.03074750490486622, + 0.019271019846200943, + 0.021637074649333954, + 0.015658924356102943, + 0.0058665163815021515, + 0.008264951407909393, + 0.005841459613293409, + 0.007555332500487566, + 0.006383115891367197, + 0.004935207776725292, + 0.005680415313690901, + 0.00702615175396204, + 0.0511714369058609, + 0.013531163334846497, + 0.07233919203281403, + 0.008176741190254688, + 0.01195114478468895, + 0.02029116079211235, + 0.009124740026891232, + 0.015107247047126293, + 0.012181066907942295, + 0.0066320332698524, + 0.009134143590927124, + 0.01118402648717165, + 0.008640567772090435, + 0.008723185397684574, + 0.01172665972262621, + 0.008735784329473972, + 0.012848515063524246, + 0.007076793350279331, + 0.01670246198773384, + 0.007423893548548222, + 0.0062775867991149426, + 0.0064745391719043255, + 0.005592680536210537, + 0.007880088873207569, + 0.0046169087290763855, + 0.005095509812235832, + 0.0071746851317584515, + 0.008988099172711372, + 0.0052773617208004, + 0.003024595556780696, + 0.004612921271473169, + 0.00605108542367816, + 0.005108315497636795, + 0.004361825063824654, + 0.004783618729561567, + 0.004795484710484743, + 0.005040974356234074, + 0.003979256842285395, + 0.004531572572886944, + 0.007689833175390959, + 0.0062645552679896355, + 0.004762708675116301, + 0.0062972186133265495, + 0.006775350775569677, + 0.005321210715919733, + 0.002767883939668536, + 0.0038666613399982452, + 0.004650396760553122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.013890853151679039 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643134772777557 + ], + "opposite_logprob_tokens": [ + -5.379642963409424 + ], + "logprob_reward": -0.004643134772777557, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643134772777557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08692118525505066, + 0.08834615349769592, + 0.06575490534305573, + 0.07180509716272354, + 0.11568334698677063, + 0.07833471149206161, + 0.11624523997306824, + 0.2547227740287781, + 0.06816621869802475, + 0.054020386189222336 + ], + "credit_map": [ + 0.00932302325963974, + 0.009711192920804024, + 0.017262982204556465, + 0.010159400291740894, + 0.009746359661221504, + 0.009428746998310089, + 0.00960187055170536, + 0.01168760098516941, + 0.004863879643380642, + 0.008806604892015457, + 0.02349635399878025, + 0.009822209365665913, + 0.0071339900605380535, + 0.017966391518712044, + 0.0066468967124819756, + 0.00960981659591198, + 0.008721503429114819, + 0.007997442036867142, + 0.007765030954033136, + 0.009863457642495632, + 0.006289825774729252, + 0.00953573826700449, + 0.005569414235651493, + 0.010012480430305004, + 0.005958473309874535, + 0.011993056163191795, + 0.007685767952352762, + 0.009699815884232521, + 0.008452551439404488, + 0.008156056515872478, + 0.009844833984971046, + 0.010014533996582031, + 0.022543633356690407, + 0.008491219952702522, + 0.028499968349933624, + 0.006992497947067022, + 0.008437378332018852, + 0.019139770418405533, + 0.007134993094950914, + 0.014443875290453434, + 0.011104325763881207, + 0.007627441082149744, + 0.011706965044140816, + 0.010089503601193428, + 0.007969923317432404, + 0.010106116533279419, + 0.005766375921666622, + 0.013964053243398666, + 0.025658922269940376, + 0.012167314067482948, + 0.037070393562316895, + 0.009328062646090984, + 0.00604918273165822, + 0.007180918473750353, + 0.007306606508791447, + 0.011483822017908096, + 0.026951966807246208, + 0.035370953381061554, + 0.02999907173216343, + 0.04698570445179939, + 0.02715209312736988, + 0.020978691056370735, + 0.027111515402793884, + 0.0401727519929409, + 0.009962698444724083, + 0.012304302304983139, + 0.008733301423490047, + 0.008589640259742737, + 0.007827444933354855, + 0.004008735995739698, + 0.006578163243830204, + 0.010161924175918102, + 0.008887865580618382, + 0.009778241626918316, + 0.005427058320492506, + 0.00734013831242919, + 0.0071630338206887245, + 0.004526832140982151, + 0.005383135750889778, + 0.005514075979590416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.004643134772777557 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043049927800893784 + ], + "opposite_logprob_tokens": [ + -3.1680498123168945 + ], + "logprob_reward": -0.043049927800893784, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043049927800893784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08801168948411942, + 0.0828012153506279, + 0.06357640027999878, + 0.27025306224823, + 0.0955621525645256, + 0.11136748641729355, + 0.06142016872763634, + 0.03952237218618393, + 0.1011229008436203, + 0.08636254817247391 + ], + "credit_map": [ + 0.009184480644762516, + 0.012600469402968884, + 0.0128597691655159, + 0.010063881985843182, + 0.010018440894782543, + 0.011427023448050022, + 0.012790054082870483, + 0.009067569859325886, + 0.005617181770503521, + 0.006040120963007212, + 0.024866612628102303, + 0.008813481777906418, + 0.008370963856577873, + 0.012388044036924839, + 0.0065492684952914715, + 0.010155544616281986, + 0.004523636773228645, + 0.007784607354551554, + 0.0075139245018363, + 0.007708865683525801, + 0.010726387612521648, + 0.013082045130431652, + 0.006541355978697538, + 0.005695575848221779, + 0.026577461510896683, + 0.04932153970003128, + 0.035075049847364426, + 0.03645609691739082, + 0.035967450588941574, + 0.0265805684030056, + 0.02701619267463684, + 0.03325871005654335, + 0.018164897337555885, + 0.005422723013907671, + 0.02679005078971386, + 0.006847201380878687, + 0.008940436877310276, + 0.012062289752066135, + 0.005559925921261311, + 0.011774633079767227, + 0.021060096099972725, + 0.01674085669219494, + 0.013436596840620041, + 0.014512130059301853, + 0.01018628291785717, + 0.010743089951574802, + 0.01358985435217619, + 0.011098584160208702, + 0.010377274826169014, + 0.004690728150308132, + 0.01536103431135416, + 0.00731836911290884, + 0.006811824627220631, + 0.004317507613450289, + 0.005943170748651028, + 0.006600260268896818, + 0.004885743837803602, + 0.005156451836228371, + 0.0057761939242482185, + 0.005180847831070423, + 0.0039474815130233765, + 0.004478730261325836, + 0.0034346950706094503, + 0.006662228610366583, + 0.012236435897648335, + 0.017795536667108536, + 0.015522127971053123, + 0.014175744727253914, + 0.012057210318744183, + 0.004728630185127258, + 0.010631737299263477, + 0.013975479640066624, + 0.016666049137711525, + 0.009951181709766388, + 0.018450988456606865, + 0.0062168980948626995, + 0.010329474695026875, + 0.008190293796360493, + 0.006962407846003771, + 0.00959524605423212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.043049927800893784 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775741025805473 + ], + "opposite_logprob_tokens": [ + -4.634775638580322 + ], + "logprob_reward": -0.009775741025805473, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775741025805473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12071354687213898, + 0.044211309403181076, + 0.10597807168960571, + 0.0717153325676918, + 0.27251267433166504, + 0.23291851580142975, + 0.04788465425372124, + 0.023679740726947784, + 0.029899494722485542, + 0.05048667639493942 + ], + "credit_map": [ + 0.011670189909636974, + 0.012121325358748436, + 0.027297386899590492, + 0.012543187476694584, + 0.0097512723878026, + 0.012672627344727516, + 0.0165755283087492, + 0.0180820282548666, + 0.0029954598285257816, + 0.004567901603877544, + 0.00742323137819767, + 0.004682047758251429, + 0.005535914562642574, + 0.008380688726902008, + 0.003937095869332552, + 0.006688971538096666, + 0.009034928865730762, + 0.0057253725826740265, + 0.022149186581373215, + 0.009671119973063469, + 0.008666854351758957, + 0.0246575940400362, + 0.007415000814944506, + 0.018658017739653587, + 0.007498482707887888, + 0.007789460010826588, + 0.012662645429372787, + 0.010028840973973274, + 0.006782773416489363, + 0.008557463064789772, + 0.0067831710912287235, + 0.011612496338784695, + 0.06389017403125763, + 0.019141148775815964, + 0.09004812687635422, + 0.012297316454350948, + 0.015104865655303001, + 0.036568962037563324, + 0.009778325445950031, + 0.025683745741844177, + 0.03293674439191818, + 0.03256893903017044, + 0.04600425064563751, + 0.01983589679002762, + 0.014285171404480934, + 0.03461939096450806, + 0.013548750430345535, + 0.03911937400698662, + 0.01019469927996397, + 0.002814182545989752, + 0.014518027193844318, + 0.003344452241435647, + 0.003051342209801078, + 0.00495996093377471, + 0.0024059659335762262, + 0.006596022751182318, + 0.002973467344418168, + 0.005425125360488892, + 0.0030650594271719456, + 0.002827585441991687, + 0.003097803331911564, + 0.0015131679829210043, + 0.002336813835427165, + 0.002440716838464141, + 0.006076164543628693, + 0.0030787133146077394, + 0.00368468277156353, + 0.0030086622573435307, + 0.0028904308564960957, + 0.003304670099169016, + 0.0023943788837641478, + 0.005461791530251503, + 0.009822827763855457, + 0.005225055385380983, + 0.008552377112209797, + 0.004940965678542852, + 0.006017504259943962, + 0.0061722612008452415, + 0.003225581953302026, + 0.006530099548399448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.009775741025805473 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16031722724437714 + ], + "opposite_logprob_tokens": [ + -1.9103171825408936 + ], + "logprob_reward": -0.16031722724437714, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16031722724437714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13349930942058563, + 0.08293980360031128, + 0.0626048818230629, + 0.06698359549045563, + 0.2402888536453247, + 0.09337923675775528, + 0.13067659735679626, + 0.07601062208414078, + 0.03840966150164604, + 0.07520744204521179 + ], + "credit_map": [ + 0.012426347471773624, + 0.012053685262799263, + 0.035205189138650894, + 0.01319825742393732, + 0.011701717041432858, + 0.015546725131571293, + 0.013080141507089138, + 0.02028723433613777, + 0.006530627142637968, + 0.010654747486114502, + 0.01183078158646822, + 0.007248818874359131, + 0.011106795631349087, + 0.013513739220798016, + 0.009785229340195656, + 0.012269065715372562, + 0.012677418999373913, + 0.004950183443725109, + 0.009576295502483845, + 0.006371071562170982, + 0.005364287178963423, + 0.009112426079809666, + 0.005966015160083771, + 0.008587181568145752, + 0.00755081232637167, + 0.006430296692997217, + 0.008641600608825684, + 0.007107721175998449, + 0.005902407690882683, + 0.010601721704006195, + 0.007393638137727976, + 0.013355394825339317, + 0.06835641711950302, + 0.009113137610256672, + 0.10215219110250473, + 0.010848989710211754, + 0.010717405937612057, + 0.010379304178059101, + 0.007945227436721325, + 0.020776180550456047, + 0.014206048101186752, + 0.01515782717615366, + 0.01290923822671175, + 0.01082555204629898, + 0.009975006803870201, + 0.012852816842496395, + 0.00999019667506218, + 0.007462554611265659, + 0.035268981009721756, + 0.012962307780981064, + 0.04278269410133362, + 0.008612297475337982, + 0.00673610158264637, + 0.009591513313353062, + 0.005220787599682808, + 0.009501894935965538, + 0.009883958846330643, + 0.00948732066899538, + 0.007613060064613819, + 0.009925697930157185, + 0.009050870314240456, + 0.009302660822868347, + 0.010168313980102539, + 0.010578741319477558, + 0.007061993703246117, + 0.0039701503701508045, + 0.004169943276792765, + 0.0052725691348314285, + 0.005967465229332447, + 0.0032285978086292744, + 0.003739559091627598, + 0.004999381955713034, + 0.016048835590481758, + 0.006885833106935024, + 0.010598516091704369, + 0.007933379150927067, + 0.010148195549845695, + 0.008816519752144814, + 0.005476236809045076, + 0.009299930185079575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.16031722724437714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a surreal style, this slow, orbiting medium shot smoothly orbits a woman going about her morning activities in a lush garden at dawn. She has long, dark brown hair, loosely tied in a high ponytail, with a few stray strands gently framing her face. Her eyes are gently closed, her expression serene and focused. She wears a fitted, long-sleeved heather-gray yoga top, the soft, breathable fabric subtly textured, paired with high-waisted, dark teal leggings. Her body gracefully stands in side bend; her right arm extends deliberately and fluidly upward and over her head, elongating her torso, while her left hand rests lightly on her hip. The entire pose maintains a steady, controlled grace. The camera's slow, circular path reveals the immersive environment: she stands barefoot on a manicured lawn, still damp with morning dew. The garden is filled with vibrant green ferns and blooming rose bushes with soft, pink and white petals. The first rays of warm, golden sunlight penetrate the light morning mist between the leaves of the tall, mature trees in the background, backlighting her body and capturing millions of tiny, shimmering dewdrops on the grass and leaves, creating a shimmering bokeh effect. The interplay of warm, low-angle light and cool, soft shadows creates a serene scene of profound calm and focus, and the camera movement enhances the contemplative and visually rich atmosphere.", + "seed": 22006, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_11/index_7140_seed_22006_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03713804315775633, + "vlm_logprob": -0.03713804315775633, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.03713804315775633, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013279094360768795 + ], + "opposite_logprob_tokens": [ + -4.513278961181641 + ], + "logprob_reward": -0.013279094360768795, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013279094360768795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48652803897857666, + 0.050651732832193375, + 0.054413825273513794, + 0.04165190830826759, + 0.052227117121219635, + 0.04841213673353195, + 0.0919722244143486, + 0.08111079037189484, + 0.042033061385154724, + 0.05099916458129883 + ], + "credit_map": [ + 0.025468438863754272, + 0.023475386202335358, + 0.09110375493764877, + 0.024663805961608887, + 0.02832961082458496, + 0.10260476171970367, + 0.019846633076667786, + 0.17103566229343414, + 0.007327996660023928, + 0.00477051455527544, + 0.007545056752860546, + 0.005436680279672146, + 0.005951139144599438, + 0.007752332836389542, + 0.00544257927685976, + 0.006425434723496437, + 0.006633862387388945, + 0.004308880772441626, + 0.010267673060297966, + 0.006357002537697554, + 0.0034231524914503098, + 0.006667343899607658, + 0.003493895288556814, + 0.013262017630040646, + 0.0033686405513435602, + 0.0046296874061226845, + 0.004328943323343992, + 0.005342673510313034, + 0.0038102469407022, + 0.008614719845354557, + 0.003662017872557044, + 0.007894981652498245, + 0.007460393477231264, + 0.005702880211174488, + 0.005563782528042793, + 0.006262368056923151, + 0.009355621412396431, + 0.005979903507977724, + 0.0047712912783026695, + 0.007130878511816263, + 0.007027155719697475, + 0.006087655667215586, + 0.007382195442914963, + 0.005724817514419556, + 0.005938367452472448, + 0.005892490968108177, + 0.0034940678160637617, + 0.006865381728857756, + 0.00927679892629385, + 0.018037378787994385, + 0.011910807341337204, + 0.009007913060486317, + 0.01299022976309061, + 0.012449223548173904, + 0.006814784370362759, + 0.011485089547932148, + 0.011379744857549667, + 0.01669836975634098, + 0.011231872253119946, + 0.009235212579369545, + 0.008876769803464413, + 0.008647240698337555, + 0.006310486234724522, + 0.008731087669730186, + 0.006263699382543564, + 0.008037593215703964, + 0.004838373977690935, + 0.0053708492778241634, + 0.0055583156645298, + 0.003931747749447823, + 0.00436559971421957, + 0.0036668817047029734, + 0.006750232540071011, + 0.006757428869605064, + 0.00917489267885685, + 0.006696369498968124, + 0.008002606220543385, + 0.004757737275213003, + 0.005370950326323509, + 0.003488947870209813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.013279094360768795 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763675756752491 + ], + "opposite_logprob_tokens": [ + -4.882636547088623 + ], + "logprob_reward": -0.00763675756752491, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763675756752491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8226181268692017, + 0.014719977043569088, + 0.018183128908276558, + 0.020780250430107117, + 0.01911020837724209, + 0.020248079672455788, + 0.015698906034231186, + 0.016193833202123642, + 0.01333125215023756, + 0.03911623731255531 + ], + "credit_map": [ + 0.028531625866889954, + 0.01492259930819273, + 0.07998991012573242, + 0.014275937341153622, + 0.0407525934278965, + 0.3316062390804291, + 0.0294566061347723, + 0.28308266401290894, + 0.0016070352867245674, + 0.0018904033349826932, + 0.0019321662839502096, + 0.0019679006654769182, + 0.0020748337265104055, + 0.0018279360374435782, + 0.001686100964434445, + 0.0017336019081994891, + 0.002248351462185383, + 0.0030000649858266115, + 0.002068370347842574, + 0.003043264150619507, + 0.002030473668128252, + 0.0025822692550718784, + 0.0015961115714162588, + 0.0016142234671860933, + 0.0016938340850174427, + 0.0031669719610363245, + 0.002299216575920582, + 0.0029937352519482374, + 0.0020455727353692055, + 0.0038263495080173016, + 0.0018897374393418431, + 0.002864834852516651, + 0.0016545109683647752, + 0.002101206686347723, + 0.0015680789947509766, + 0.0021583735942840576, + 0.00335369142703712, + 0.002412680070847273, + 0.0032812319695949554, + 0.0025804375763982534, + 0.0021491372026503086, + 0.0028281810227781534, + 0.002610889496281743, + 0.0028495423030108213, + 0.0031001311726868153, + 0.002492170548066497, + 0.002053160686045885, + 0.0021648691035807133, + 0.0013301987200975418, + 0.002691094297915697, + 0.001949136727489531, + 0.0017007318092510104, + 0.0020230135414749384, + 0.002759573282673955, + 0.0012217621551826596, + 0.0020233970135450363, + 0.0019964135717600584, + 0.0027155179996043444, + 0.0018609932158142328, + 0.00219500670209527, + 0.002171273808926344, + 0.001759665785357356, + 0.001635700580663979, + 0.00185926363337785, + 0.003070933511480689, + 0.0016733923694118857, + 0.0020765250083059072, + 0.0016664594877511263, + 0.0013993916800245643, + 0.0014378898777067661, + 0.0010349391959607601, + 0.000971721950918436, + 0.005691629834473133, + 0.0031145503744482994, + 0.006335890851914883, + 0.0045124227181077, + 0.00603949511423707, + 0.004486816469579935, + 0.005776420701295137, + 0.003159015439450741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.00763675756752491 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369821608066559 + ], + "opposite_logprob_tokens": [ + -3.408698320388794 + ], + "logprob_reward": -0.03369821608066559, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369821608066559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.800177812576294, + 0.009756018407642841, + 0.01732426881790161, + 0.009654311463236809, + 0.02499750815331936, + 0.011184033006429672, + 0.009622025303542614, + 0.035206522792577744, + 0.05588733032345772, + 0.026190124452114105 + ], + "credit_map": [ + 0.03052132949233055, + 0.014807909727096558, + 0.07751377671957016, + 0.014427466318011284, + 0.03865811973810196, + 0.3200682997703552, + 0.0266721174120903, + 0.27750882506370544, + 0.0013250941410660744, + 0.001073329010978341, + 0.001449520350433886, + 0.0014827930135652423, + 0.0006409998168237507, + 0.0014661761233583093, + 0.0008660171297378838, + 0.0014520901022478938, + 0.0038316096179187298, + 0.0016316071851179004, + 0.0026359758339822292, + 0.002246260643005371, + 0.001842113328166306, + 0.0022728946059942245, + 0.0011078229872509837, + 0.001755985664203763, + 0.0010085571557283401, + 0.0014356916071847081, + 0.0010275421664118767, + 0.0011732341954484582, + 0.0016743266023695469, + 0.0012512746034190059, + 0.0013287708861753345, + 0.0007549154688604176, + 0.00240584509447217, + 0.00330604100599885, + 0.002365791704505682, + 0.0023368268739432096, + 0.004593168385326862, + 0.0027171692345291376, + 0.004036532249301672, + 0.0032361363992094994, + 0.0013911816058680415, + 0.0015893916133791208, + 0.0013940294738858938, + 0.0015064898179844022, + 0.001759168691933155, + 0.0012188103282824159, + 0.0010636800434440374, + 0.0012612827122211456, + 0.0007861033664084971, + 0.001287731109187007, + 0.0013312384253367782, + 0.0012499237200245261, + 0.0011745266383513808, + 0.0017231965903192759, + 0.0011072986526414752, + 0.0009620081982575357, + 0.004333529621362686, + 0.005694451741874218, + 0.004132551606744528, + 0.004106809385120869, + 0.0031487399246543646, + 0.005777252838015556, + 0.0027564994525164366, + 0.0052566914819180965, + 0.010234582237899303, + 0.01041810680180788, + 0.007799959741532803, + 0.006222942844033241, + 0.006643038708716631, + 0.005755624733865261, + 0.004268381744623184, + 0.004544700961560011, + 0.006442577578127384, + 0.004662416875362396, + 0.0050958083011209965, + 0.0015437380643561482, + 0.0017932243645191193, + 0.0027207538951188326, + 0.0015961650060489774, + 0.0023354438599199057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.03369821608066559 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818074844777584 + ], + "opposite_logprob_tokens": [ + -4.018180847167969 + ], + "logprob_reward": -0.01818074844777584, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818074844777584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7975714802742004, + 0.027704978361725807, + 0.022529156878590584, + 0.0098465196788311, + 0.034731920808553696, + 0.02175106108188629, + 0.01486285962164402, + 0.021147387102246284, + 0.013132919557392597, + 0.03672169893980026 + ], + "credit_map": [ + 0.03413790091872215, + 0.015836823731660843, + 0.09548633545637131, + 0.01965237595140934, + 0.04069359600543976, + 0.2699623703956604, + 0.023551030084490776, + 0.29825106263160706, + 0.003441969398409128, + 0.0025288851466029882, + 0.0065651582553982735, + 0.0022259920369833708, + 0.002461495576426387, + 0.004988266620784998, + 0.0021499237045645714, + 0.0033432887867093086, + 0.0037311618216335773, + 0.0017807960975915194, + 0.003543267725035548, + 0.0017594537930563092, + 0.0023364818189293146, + 0.005064066499471664, + 0.001901768147945404, + 0.002412161324173212, + 0.0011589565547183156, + 0.0009719395311549306, + 0.0018306850688531995, + 0.0010471363784745336, + 0.0009271184680983424, + 0.0017978486139327288, + 0.0007235189550556242, + 0.0013893160503357649, + 0.005226843990385532, + 0.0033660728950053453, + 0.004188026301562786, + 0.002888381015509367, + 0.003402585629373789, + 0.00414667883887887, + 0.004840928129851818, + 0.006672403309494257, + 0.004115939140319824, + 0.0023145403247326612, + 0.003409872530028224, + 0.001995102735236287, + 0.0025711790658533573, + 0.0024500046856701374, + 0.00126240449026227, + 0.0036320181097835302, + 0.0021896210964769125, + 0.0021730472799390554, + 0.0018085300689563155, + 0.001601683208718896, + 0.0016858706949278712, + 0.0018642973154783249, + 0.0014354713493958116, + 0.002104338491335511, + 0.0028114651795476675, + 0.004746909718960524, + 0.002867097267881036, + 0.001546342857182026, + 0.0015856012469157577, + 0.003028463339433074, + 0.001374014769680798, + 0.003187494119629264, + 0.0029675085097551346, + 0.0013321660226210952, + 0.0022113160230219364, + 0.0012490290682762861, + 0.0012017044937238097, + 0.0013439098838716745, + 0.0011966011952608824, + 0.0016306835459545255, + 0.010059753432869911, + 0.007045608479529619, + 0.006412556394934654, + 0.002475037705153227, + 0.0019201739924028516, + 0.003749368479475379, + 0.0017470057355239987, + 0.003312191925942898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.01818074844777584 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11289539933204651 + ], + "opposite_logprob_tokens": [ + -2.2378954887390137 + ], + "logprob_reward": -0.11289539933204651, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11289539933204651, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7858954071998596, + 0.026973964646458626, + 0.01420824695378542, + 0.013481521047651768, + 0.017543114721775055, + 0.014043185859918594, + 0.029192540794610977, + 0.04195495322346687, + 0.028098026290535927, + 0.028609048575162888 + ], + "credit_map": [ + 0.0340396873652935, + 0.017399443313479424, + 0.08734876662492752, + 0.015318751335144043, + 0.0338791199028492, + 0.31659024953842163, + 0.0258786603808403, + 0.25544071197509766, + 0.0030792513862252235, + 0.0019237145315855742, + 0.006208354141563177, + 0.002349957823753357, + 0.003007655031979084, + 0.00429535610601306, + 0.0024629642721265554, + 0.0036467122845351696, + 0.002529927995055914, + 0.0015590459806844592, + 0.0027066958136856556, + 0.0016040051123127341, + 0.0012270092265680432, + 0.002256967592984438, + 0.0007668078760616481, + 0.0015577867161482573, + 0.0017268317751586437, + 0.001097382395528257, + 0.0012877029366791248, + 0.0013465759111568332, + 0.0013350335648283362, + 0.0024496372789144516, + 0.0017739570466801524, + 0.002464401302859187, + 0.001994460355490446, + 0.0015311141032725573, + 0.0018007542239502072, + 0.0018474082462489605, + 0.0018453601514920592, + 0.001983349910005927, + 0.0037263890262693167, + 0.002814279170706868, + 0.0021367340814322233, + 0.00152755924500525, + 0.0017788003897294402, + 0.001284305821172893, + 0.002357455203309655, + 0.0016918272012844682, + 0.001109427073970437, + 0.0021570774260908365, + 0.004061393439769745, + 0.005318132694810629, + 0.0036972586531192064, + 0.002500980393961072, + 0.0031218165531754494, + 0.0034943362697958946, + 0.0025720270350575447, + 0.004426595289260149, + 0.005222303792834282, + 0.007531261537224054, + 0.005676142405718565, + 0.002957563381642103, + 0.003038412658497691, + 0.007869312539696693, + 0.001966241281479597, + 0.007693715393543243, + 0.003829488530755043, + 0.006403478793799877, + 0.0029028188437223434, + 0.0036779986694455147, + 0.003920247312635183, + 0.0027301267255097628, + 0.0029570329934358597, + 0.0016768347704783082, + 0.0063193179666996, + 0.005321708042174578, + 0.005283878184854984, + 0.0024834449868649244, + 0.001478035468608141, + 0.0033237338066101074, + 0.0017258677398785949, + 0.002673061564564705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.11289539933204651 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a surreal style, this slow, orbiting medium shot smoothly orbits a woman going about her morning activities in a lush garden at dawn. She has long, dark brown hair, loosely tied in a high ponytail, with a few stray strands gently framing her face. Her eyes are gently closed, her expression serene and focused. She wears a fitted, long-sleeved heather-gray yoga top, the soft, breathable fabric subtly textured, paired with high-waisted, dark teal leggings. Her body gracefully stands in side bend; her right arm extends deliberately and fluidly upward and over her head, elongating her torso, while her left hand rests lightly on her hip. The entire pose maintains a steady, controlled grace. The camera's slow, circular path reveals the immersive environment: she stands barefoot on a manicured lawn, still damp with morning dew. The garden is filled with vibrant green ferns and blooming rose bushes with soft, pink and white petals. The first rays of warm, golden sunlight penetrate the light morning mist between the leaves of the tall, mature trees in the background, backlighting her body and capturing millions of tiny, shimmering dewdrops on the grass and leaves, creating a shimmering bokeh effect. The interplay of warm, low-angle light and cool, soft shadows creates a serene scene of profound calm and focus, and the camera movement enhances the contemplative and visually rich atmosphere.", + "seed": 22007, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_11/index_7140_seed_22007_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04689784124493599, + "vlm_logprob": -0.04689784124493599, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.04689784124493599, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01779600977897644 + ], + "opposite_logprob_tokens": [ + -4.142796039581299 + ], + "logprob_reward": -0.01779600977897644, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01779600977897644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2220829874277115, + 0.24997837841510773, + 0.06715542823076248, + 0.046196889132261276, + 0.12774160504341125, + 0.06139426678419113, + 0.04615052044391632, + 0.07662682980298996, + 0.04622059315443039, + 0.0564524382352829 + ], + "credit_map": [ + 0.01964055746793747, + 0.02921740896999836, + 0.02958313561975956, + 0.02110452391207218, + 0.028734346851706505, + 0.03014199063181877, + 0.039183907210826874, + 0.024477142840623856, + 0.021720001474022865, + 0.035488903522491455, + 0.024991139769554138, + 0.020350445061922073, + 0.04655814915895462, + 0.030636219307780266, + 0.0466056652367115, + 0.023627880960702896, + 0.0056769344955682755, + 0.005525360349565744, + 0.007317947689443827, + 0.006134220864623785, + 0.009379471652209759, + 0.00734640983864665, + 0.011729391291737556, + 0.014045706018805504, + 0.0035675568506121635, + 0.005689632147550583, + 0.008055854588747025, + 0.00468292785808444, + 0.006269484292715788, + 0.004657704383134842, + 0.007764651905745268, + 0.0055090803653001785, + 0.013146860525012016, + 0.03227631747722626, + 0.015559148043394089, + 0.0119324317201972, + 0.015872029587626457, + 0.013355981558561325, + 0.012748696841299534, + 0.012850147671997547, + 0.006272705737501383, + 0.011027449741959572, + 0.008763247169554234, + 0.005810542497783899, + 0.005595141556113958, + 0.01016681082546711, + 0.005469829775393009, + 0.008288542740046978, + 0.005923957098275423, + 0.006974915042519569, + 0.005864556413143873, + 0.00447808438912034, + 0.007980069145560265, + 0.004039508756250143, + 0.006862206384539604, + 0.004027228802442551, + 0.022903647273778915, + 0.004597664810717106, + 0.011599143967032433, + 0.006480228155851364, + 0.01842622645199299, + 0.003927641082555056, + 0.005024883896112442, + 0.003667409997433424, + 0.004470439162105322, + 0.00683942623436451, + 0.007613683585077524, + 0.004958000499755144, + 0.005065108183771372, + 0.005825460888445377, + 0.00635445024818182, + 0.0050940304063260555, + 0.005432357080280781, + 0.004221681971102953, + 0.007464222144335508, + 0.005288118030875921, + 0.010495779104530811, + 0.006410975009202957, + 0.009860323742032051, + 0.0072789909318089485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.01779600977897644 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011082522571086884 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.011082522571086884, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082522571086884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09287431836128235, + 0.04669570550322533, + 0.05315127596259117, + 0.03329954296350479, + 0.466091513633728, + 0.06330225616693497, + 0.024026885628700256, + 0.06209719181060791, + 0.12241139262914658, + 0.0360499806702137 + ], + "credit_map": [ + 0.012668458744883537, + 0.010255767963826656, + 0.013128611259162426, + 0.00896704476326704, + 0.011102186515927315, + 0.011682439595460892, + 0.013869359157979488, + 0.011200451292097569, + 0.004587553441524506, + 0.006491710431873798, + 0.0038367838133126497, + 0.004386866930872202, + 0.007519994862377644, + 0.005821811035275459, + 0.008625722490251064, + 0.005425265524536371, + 0.007480120752006769, + 0.005678885150700808, + 0.00883467122912407, + 0.006268984638154507, + 0.006869471166282892, + 0.004022559151053429, + 0.009376024827361107, + 0.0046205585822463036, + 0.003128875745460391, + 0.002827657852321863, + 0.006022862158715725, + 0.003184717847034335, + 0.004179066512733698, + 0.003635556437075138, + 0.004776925314217806, + 0.005543880630284548, + 0.05225292965769768, + 0.06842762231826782, + 0.0577542819082737, + 0.03621514514088631, + 0.09919178485870361, + 0.055933479219675064, + 0.05072975903749466, + 0.04558649659156799, + 0.007986310869455338, + 0.009610083885490894, + 0.010498358868062496, + 0.006830442231148481, + 0.005824355874210596, + 0.009432202205061913, + 0.005551277194172144, + 0.007569222245365381, + 0.003890011692419648, + 0.0024389324244111776, + 0.003479650476947427, + 0.0022469398099929094, + 0.003868375439196825, + 0.00233464571647346, + 0.0030534942634403706, + 0.0027148353401571512, + 0.012034928426146507, + 0.003096122294664383, + 0.02089105360209942, + 0.006313636898994446, + 0.007120415102690458, + 0.00313583854585886, + 0.0036155576817691326, + 0.005889636464416981, + 0.012153888121247292, + 0.006543706636875868, + 0.03919032961130142, + 0.007271550130099058, + 0.00709547521546483, + 0.03039114736020565, + 0.008301666006445885, + 0.011463627219200134, + 0.002881283638998866, + 0.0031711356714367867, + 0.006251421291381121, + 0.004009270574897528, + 0.005161141511052847, + 0.004417443182319403, + 0.00579995708540082, + 0.00435832841321826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.011082522571086884 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864819347858429 + ], + "opposite_logprob_tokens": [ + -3.0486481189727783 + ], + "logprob_reward": -0.04864819347858429, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864819347858429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1224021464586258, + 0.07162107527256012, + 0.0470084585249424, + 0.1038399413228035, + 0.11635778099298477, + 0.07107748091220856, + 0.04142829403281212, + 0.27392783761024475, + 0.08021871000528336, + 0.07211831957101822 + ], + "credit_map": [ + 0.014181585051119328, + 0.016849976032972336, + 0.016288001090288162, + 0.01082717813551426, + 0.0166257806122303, + 0.014490244910120964, + 0.019809965044260025, + 0.013329396024346352, + 0.006155673880130053, + 0.010925252921879292, + 0.009162545204162598, + 0.006612503435462713, + 0.010620580054819584, + 0.008610483258962631, + 0.011623087339103222, + 0.007910937070846558, + 0.004268554970622063, + 0.005728084594011307, + 0.006134051363915205, + 0.006533748935908079, + 0.006098432000726461, + 0.005122694186866283, + 0.0043755946680903435, + 0.008747290819883347, + 0.009141040965914726, + 0.012507764622569084, + 0.017034610733389854, + 0.01037565153092146, + 0.008799104951322079, + 0.011209781281650066, + 0.013318161480128765, + 0.02145381271839142, + 0.012034879997372627, + 0.01669398695230484, + 0.012957398779690266, + 0.012312297709286213, + 0.027924036607146263, + 0.012241915799677372, + 0.012782908976078033, + 0.009410340338945389, + 0.008388488553464413, + 0.005798624362796545, + 0.012742341496050358, + 0.007379445713013411, + 0.005209656897932291, + 0.010466094128787518, + 0.005261904560029507, + 0.01583091914653778, + 0.005126211792230606, + 0.0037168546114116907, + 0.008229343220591545, + 0.004156762734055519, + 0.005957200191915035, + 0.004013385158032179, + 0.005018460098654032, + 0.005210070870816708, + 0.08140179514884949, + 0.013988635502755642, + 0.027741137892007828, + 0.01107141375541687, + 0.09217814356088638, + 0.018031371757388115, + 0.015838809311389923, + 0.013676503673195839, + 0.008968125097453594, + 0.012566981837153435, + 0.012060546316206455, + 0.00889294221997261, + 0.008758862502872944, + 0.007286186795681715, + 0.012591917999088764, + 0.009093130938708782, + 0.005506435409188271, + 0.008030091412365437, + 0.011575672775506973, + 0.007649092935025692, + 0.010000737383961678, + 0.00930943712592125, + 0.011528181843459606, + 0.00851866602897644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.04864819347858429 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191891998052597 + ], + "opposite_logprob_tokens": [ + -4.2641921043396 + ], + "logprob_reward": -0.014191891998052597, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191891998052597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20504136383533478, + 0.09143002331256866, + 0.0396287776529789, + 0.07683703303337097, + 0.22252057492733002, + 0.15310028195381165, + 0.04037082567811012, + 0.07472700625658035, + 0.041165415197610855, + 0.0551786869764328 + ], + "credit_map": [ + 0.02251080982387066, + 0.018223321065306664, + 0.03678525984287262, + 0.021808849647641182, + 0.017452247440814972, + 0.02907819300889969, + 0.02085069566965103, + 0.03833198919892311, + 0.008674086071550846, + 0.012676598504185677, + 0.011591844260692596, + 0.0074837468564510345, + 0.014650491997599602, + 0.011600549332797527, + 0.015971705317497253, + 0.008780996315181255, + 0.004097758326679468, + 0.0034167547710239887, + 0.005947559606283903, + 0.005922124721109867, + 0.003972383216023445, + 0.006520042661577463, + 0.0038888806011527777, + 0.00586327351629734, + 0.0067954589612782, + 0.012369414791464806, + 0.009096862748265266, + 0.006247665733098984, + 0.012193666771054268, + 0.01035069115459919, + 0.012240744195878506, + 0.007542532403022051, + 0.022708818316459656, + 0.03204819932579994, + 0.027831504121422768, + 0.018047751858830452, + 0.047843486070632935, + 0.028473583981394768, + 0.023136835545301437, + 0.022430377081036568, + 0.015561813488602638, + 0.008393803611397743, + 0.0314820371568203, + 0.017815418541431427, + 0.016786107793450356, + 0.027103343978524208, + 0.017246747389435768, + 0.01871100440621376, + 0.006624582689255476, + 0.0037724028807133436, + 0.005540038924664259, + 0.0026122250128537416, + 0.0044741397723555565, + 0.004021624103188515, + 0.0039888108149170876, + 0.009336999617516994, + 0.019809365272521973, + 0.0046281046234071255, + 0.019024593755602837, + 0.005019899923354387, + 0.01030284445732832, + 0.004820785019546747, + 0.005187527276575565, + 0.0059338822029531, + 0.004784794524312019, + 0.0064957826398313046, + 0.009295981377363205, + 0.0027906736359000206, + 0.0037560672499239445, + 0.005443323869258165, + 0.005282553844153881, + 0.0033162350300699472, + 0.0055379243567585945, + 0.004590856842696667, + 0.00941979605704546, + 0.005155443213880062, + 0.009470045566558838, + 0.007089802995324135, + 0.00651558255776763, + 0.007399235386401415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.014191891998052597 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14277058839797974 + ], + "opposite_logprob_tokens": [ + -2.017770528793335 + ], + "logprob_reward": -0.14277058839797974, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14277058839797974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18762964010238647, + 0.09380823373794556, + 0.12360569834709167, + 0.05871891230344772, + 0.08113835006952286, + 0.1696816384792328, + 0.07625387609004974, + 0.0863363966345787, + 0.0370599664747715, + 0.08576731383800507 + ], + "credit_map": [ + 0.029823336750268936, + 0.015098383650183678, + 0.027757931500673294, + 0.01727466657757759, + 0.019624441862106323, + 0.027278074994683266, + 0.021663455292582512, + 0.029109343886375427, + 0.007378785405308008, + 0.01559595949947834, + 0.007357615511864424, + 0.008034655824303627, + 0.01325161475688219, + 0.012500554323196411, + 0.02035262994468212, + 0.009336414746940136, + 0.01158570870757103, + 0.010302257724106312, + 0.010397657752037048, + 0.015914393588900566, + 0.022627973929047585, + 0.0116319190710783, + 0.022310057654976845, + 0.018835728988051414, + 0.005578687414526939, + 0.005958915688097477, + 0.010361126624047756, + 0.006220596842467785, + 0.007578580174595118, + 0.006787162274122238, + 0.007532777264714241, + 0.008701060898602009, + 0.00993370171636343, + 0.011088049970567226, + 0.008345672860741615, + 0.007105671335011721, + 0.014695183373987675, + 0.009878932498395443, + 0.01172766275703907, + 0.008363468572497368, + 0.02539120800793171, + 0.022241782397031784, + 0.022187476977705956, + 0.017865337431430817, + 0.015673546120524406, + 0.018536152318120003, + 0.01882825791835785, + 0.02895786054432392, + 0.014101475477218628, + 0.008108340203762054, + 0.011537419632077217, + 0.005616629496216774, + 0.009457295760512352, + 0.005789721384644508, + 0.007545456290245056, + 0.014097543433308601, + 0.007195429410785437, + 0.005589000415056944, + 0.028520602732896805, + 0.00445356871932745, + 0.023666558787226677, + 0.005357471760362387, + 0.005273149348795414, + 0.006280616857111454, + 0.004157629329711199, + 0.0033842248376458883, + 0.010752872563898563, + 0.0022024428471922874, + 0.002538163447752595, + 0.008281169459223747, + 0.0034609935246407986, + 0.00228246976621449, + 0.0070776804350316525, + 0.009204210713505745, + 0.010623065754771233, + 0.007439592853188515, + 0.02049892395734787, + 0.011255198158323765, + 0.011771675199270248, + 0.00789697002619505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.14277058839797974 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a surreal style, this slow, orbiting medium shot smoothly orbits a woman going about her morning activities in a lush garden at dawn. She has long, dark brown hair, loosely tied in a high ponytail, with a few stray strands gently framing her face. Her eyes are gently closed, her expression serene and focused. She wears a fitted, long-sleeved heather-gray yoga top, the soft, breathable fabric subtly textured, paired with high-waisted, dark teal leggings. Her body gracefully stands in side bend; her right arm extends deliberately and fluidly upward and over her head, elongating her torso, while her left hand rests lightly on her hip. The entire pose maintains a steady, controlled grace. The camera's slow, circular path reveals the immersive environment: she stands barefoot on a manicured lawn, still damp with morning dew. The garden is filled with vibrant green ferns and blooming rose bushes with soft, pink and white petals. The first rays of warm, golden sunlight penetrate the light morning mist between the leaves of the tall, mature trees in the background, backlighting her body and capturing millions of tiny, shimmering dewdrops on the grass and leaves, creating a shimmering bokeh effect. The interplay of warm, low-angle light and cool, soft shadows creates a serene scene of profound calm and focus, and the camera movement enhances the contemplative and visually rich atmosphere.", + "seed": 22008, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_11/index_7140_seed_22008_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04410820193588734, + "vlm_logprob": -0.04410820193588734, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.04410820193588734, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011728380806744099 + ], + "opposite_logprob_tokens": [ + -4.886728286743164 + ], + "logprob_reward": -0.011728380806744099, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.011728380806744099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13621146976947784, + 0.09915675967931747, + 0.05261920765042305, + 0.06945939362049103, + 0.13539183139801025, + 0.0714196264743805, + 0.08074852079153061, + 0.1582895666360855, + 0.11554712057113647, + 0.08115655928850174 + ], + "credit_map": [ + 0.01624881476163864, + 0.015895221382379532, + 0.021952683106064796, + 0.015086077153682709, + 0.015919065102934837, + 0.017840232700109482, + 0.017187077552080154, + 0.016082288697361946, + 0.016317469999194145, + 0.010293613187968731, + 0.0128964027389884, + 0.008443446829915047, + 0.010133985430002213, + 0.014191893860697746, + 0.013888217508792877, + 0.012991728261113167, + 0.0035796144511550665, + 0.006842547096312046, + 0.00790428463369608, + 0.0073850336484611034, + 0.006515411660075188, + 0.003970174118876457, + 0.006664463318884373, + 0.009757669642567635, + 0.007107082288712263, + 0.007526481989771128, + 0.01513255201280117, + 0.0066970414482057095, + 0.004707402549684048, + 0.00919631402939558, + 0.006140719633549452, + 0.012951796874403954, + 0.012301728129386902, + 0.01335824839770794, + 0.025924906134605408, + 0.014675232581794262, + 0.01009168941527605, + 0.018410762771964073, + 0.009530564770102501, + 0.031098689883947372, + 0.004415889270603657, + 0.00869147852063179, + 0.01004485972225666, + 0.008937629871070385, + 0.009563804604113102, + 0.006068704184144735, + 0.013155151158571243, + 0.010542101226747036, + 0.008268549107015133, + 0.015265189111232758, + 0.010556566528975964, + 0.009494182653725147, + 0.00852425117045641, + 0.00683174841105938, + 0.01050529908388853, + 0.011302725411951542, + 0.0105877835303545, + 0.00704538868740201, + 0.05025229975581169, + 0.01062808558344841, + 0.007977357134222984, + 0.010493704117834568, + 0.012116148136556149, + 0.0491887703537941, + 0.007870791479945183, + 0.008616064675152302, + 0.020440902560949326, + 0.028193863108754158, + 0.005699135363101959, + 0.0090982960537076, + 0.006637529004365206, + 0.028990529477596283, + 0.011466693133115768, + 0.012067792005836964, + 0.010986635461449623, + 0.0076685077510774136, + 0.006276989355683327, + 0.007816118188202381, + 0.01038870494812727, + 0.014485117048025131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.011728380806744099 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636994123458862 + ], + "opposite_logprob_tokens": [ + -4.882637023925781 + ], + "logprob_reward": -0.007636994123458862, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636994123458862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07938513159751892, + 0.09147535264492035, + 0.0324460007250309, + 0.025927333161234856, + 0.06709602475166321, + 0.06721427291631699, + 0.01945251040160656, + 0.06747956573963165, + 0.4854913055896759, + 0.06403254717588425 + ], + "credit_map": [ + 0.008222882635891438, + 0.007265585940331221, + 0.019173068925738335, + 0.008411657996475697, + 0.010978972539305687, + 0.009831258095800877, + 0.007522924803197384, + 0.00797878298908472, + 0.010527784936130047, + 0.006850990001112223, + 0.010223316960036755, + 0.006503356620669365, + 0.018594687804579735, + 0.01287870667874813, + 0.012414957396686077, + 0.013481561094522476, + 0.003632014151662588, + 0.0038036638870835304, + 0.005153217352926731, + 0.005526783876121044, + 0.0028105967212468386, + 0.003917650319635868, + 0.003159787505865097, + 0.004442289005964994, + 0.002146015176549554, + 0.0032626462634652853, + 0.004148016683757305, + 0.0036197106819599867, + 0.003470731433480978, + 0.0023733132984489202, + 0.004197458270937204, + 0.0027094415854662657, + 0.008557726629078388, + 0.006088268011808395, + 0.010154306888580322, + 0.007857310585677624, + 0.00665963813662529, + 0.00932897999882698, + 0.006553524639457464, + 0.011896270327270031, + 0.005206758622080088, + 0.004904540721327066, + 0.0118878073990345, + 0.005522295366972685, + 0.006574937142431736, + 0.008136145770549774, + 0.007238570600748062, + 0.017743218690156937, + 0.0019470305414870381, + 0.002558766631409526, + 0.0026689106598496437, + 0.002340769860893488, + 0.00221182475797832, + 0.0016493158182129264, + 0.0029610046185553074, + 0.0031148865818977356, + 0.006149497348815203, + 0.005980627611279488, + 0.014530320651829243, + 0.005710034165531397, + 0.006270027253776789, + 0.0058117457665503025, + 0.00831917766481638, + 0.014708137139678001, + 0.017631011083722115, + 0.024303019046783447, + 0.042327359318733215, + 0.16911394894123077, + 0.026591187343001366, + 0.020235219970345497, + 0.033378325402736664, + 0.15191122889518738, + 0.005965943448245525, + 0.005051444750279188, + 0.009362404234707355, + 0.011323501355946064, + 0.006637652404606342, + 0.004466056823730469, + 0.009178117848932743, + 0.012047422118484974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.007636994123458862 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864319786429405 + ], + "opposite_logprob_tokens": [ + -3.048643112182617 + ], + "logprob_reward": -0.04864319786429405, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864319786429405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10130516439676285, + 0.2221895307302475, + 0.06374919414520264, + 0.039536572992801666, + 0.1312353014945984, + 0.12530936300754547, + 0.051555074751377106, + 0.06137092411518097, + 0.13273555040359497, + 0.07101335376501083 + ], + "credit_map": [ + 0.009618782438337803, + 0.010248360224068165, + 0.012055598199367523, + 0.013514035381376743, + 0.015523095615208149, + 0.012885678559541702, + 0.013669081963598728, + 0.01379054132848978, + 0.03190930560231209, + 0.01887226291000843, + 0.023016659542918205, + 0.020409395918250084, + 0.0405295230448246, + 0.0263817198574543, + 0.03663070127367973, + 0.024439960718154907, + 0.011007320135831833, + 0.0073594944551587105, + 0.009110677987337112, + 0.004892352037131786, + 0.005319977179169655, + 0.012940122745931149, + 0.004415270406752825, + 0.008703986182808876, + 0.004231817089021206, + 0.004813977982848883, + 0.007478273473680019, + 0.0036190866958349943, + 0.0044416580349206924, + 0.004989157430827618, + 0.00444185733795166, + 0.005520747508853674, + 0.015588476322591305, + 0.012387102469801903, + 0.020195158198475838, + 0.015805086120963097, + 0.011565707623958588, + 0.018353836610913277, + 0.011852174997329712, + 0.025487756356596947, + 0.007180830929428339, + 0.010713846422731876, + 0.016811180859804153, + 0.0193231962621212, + 0.013718698173761368, + 0.014190460555255413, + 0.028544176369905472, + 0.014826973900198936, + 0.004553194157779217, + 0.007274566683918238, + 0.006134471390396357, + 0.007047105580568314, + 0.005145168397575617, + 0.005963539704680443, + 0.007826659828424454, + 0.007610369939357042, + 0.004683113191276789, + 0.007446729112416506, + 0.012739368714392185, + 0.006071927957236767, + 0.006976280827075243, + 0.005250109359622002, + 0.0072768619284033775, + 0.010926532559096813, + 0.012547752819955349, + 0.011285411193966866, + 0.014478784054517746, + 0.02417251653969288, + 0.011991949751973152, + 0.019173409789800644, + 0.010709167458117008, + 0.02837654948234558, + 0.006000074557960033, + 0.013415150344371796, + 0.010555350221693516, + 0.01759333536028862, + 0.005597803741693497, + 0.004032581113278866, + 0.006898808758705854, + 0.006920250598341227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 4 + ], + "selected_reward": -0.04864319786429405 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776213206350803 + ], + "opposite_logprob_tokens": [ + -4.6347761154174805 + ], + "logprob_reward": -0.009776213206350803, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776213206350803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25517934560775757, + 0.21225500106811523, + 0.04558410868048668, + 0.08017092198133469, + 0.06410277634859085, + 0.06370238959789276, + 0.06727223098278046, + 0.058417968451976776, + 0.10973889380693436, + 0.04357634112238884 + ], + "credit_map": [ + 0.041608911007642746, + 0.024772442877292633, + 0.04685847833752632, + 0.02042498253285885, + 0.030157171189785004, + 0.038352400064468384, + 0.01876426301896572, + 0.03424070030450821, + 0.03177483379840851, + 0.018782932311296463, + 0.02405204437673092, + 0.013106291182339191, + 0.047454316169023514, + 0.03264883905649185, + 0.021462244912981987, + 0.022973492741584778, + 0.0038358988240361214, + 0.003773442469537258, + 0.008763011544942856, + 0.005507591180503368, + 0.004633957054466009, + 0.004876771476119757, + 0.007944640703499317, + 0.006248796824365854, + 0.008907690644264221, + 0.005983029957860708, + 0.013516123406589031, + 0.0068576280027627945, + 0.010557596571743488, + 0.01301515195518732, + 0.008679156191647053, + 0.012654549442231655, + 0.007130415178835392, + 0.006470758467912674, + 0.00952131673693657, + 0.008690976537764072, + 0.00614267960190773, + 0.008017165586352348, + 0.008198036812245846, + 0.009931430220603943, + 0.0062726824544370174, + 0.006656994111835957, + 0.010617182590067387, + 0.005960522219538689, + 0.0055852849036455154, + 0.006422991398721933, + 0.006133047863841057, + 0.01605367660522461, + 0.007718005683273077, + 0.008321132510900497, + 0.0092396205291152, + 0.007043983321636915, + 0.005689780693501234, + 0.007387007586658001, + 0.013322672806680202, + 0.008550024591386318, + 0.005175346042960882, + 0.005969725549221039, + 0.013813457451760769, + 0.007013698574155569, + 0.003982933238148689, + 0.0044674803502857685, + 0.006172358989715576, + 0.011822969652712345, + 0.005428064148873091, + 0.006701268255710602, + 0.010663505643606186, + 0.036166731268167496, + 0.005516148637980223, + 0.005111144855618477, + 0.006774000357836485, + 0.033378031104803085, + 0.004640142899006605, + 0.008200154639780521, + 0.005266642197966576, + 0.003960947971791029, + 0.0045207468792796135, + 0.0041899955831468105, + 0.007268166169524193, + 0.005529545247554779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.009776213206350803 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with long, dark brown hair tied in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a heather-gray long-sleeved top and dark teal high-waisted leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing a side bend with her right arm extended upward and her left hand on her hip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush garden at dawn featuring a manicured lawn, ferns, and rose bushes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, circular orbiting movement around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14275622367858887 + ], + "opposite_logprob_tokens": [ + -2.017756223678589 + ], + "logprob_reward": -0.14275622367858887, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14275622367858887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12539902329444885, + 0.10881026089191437, + 0.027022389695048332, + 0.06260174512863159, + 0.15929146111011505, + 0.06482493132352829, + 0.042627692222595215, + 0.05696633830666542, + 0.2739875912666321, + 0.07846862822771072 + ], + "credit_map": [ + 0.011552744545042515, + 0.010943105444312096, + 0.020604688674211502, + 0.019592108204960823, + 0.01898694410920143, + 0.017141632735729218, + 0.014896903187036514, + 0.011680888943374157, + 0.014057333581149578, + 0.009709172882139683, + 0.012147204019129276, + 0.005391832906752825, + 0.025811534374952316, + 0.01440117135643959, + 0.013873999938368797, + 0.013418009504675865, + 0.00284323887899518, + 0.0033203877974301577, + 0.0038441086653620005, + 0.00436541298404336, + 0.002508654724806547, + 0.0032658767886459827, + 0.0031960606575012207, + 0.0036786484997719526, + 0.006122685968875885, + 0.005032112821936607, + 0.010072020813822746, + 0.007381144445389509, + 0.010723620653152466, + 0.008240521885454655, + 0.007965672761201859, + 0.007063967641443014, + 0.017242148518562317, + 0.015718622133135796, + 0.027513856068253517, + 0.02040204592049122, + 0.012907556258141994, + 0.024093307554721832, + 0.012579327449202538, + 0.028834611177444458, + 0.0037995202001184225, + 0.00656943628564477, + 0.0110777597874403, + 0.006394306663423777, + 0.007473526056855917, + 0.00472646951675415, + 0.008637315593659878, + 0.016146590933203697, + 0.0034373183734714985, + 0.005191318690776825, + 0.0059295170940458775, + 0.005150279495865107, + 0.003750000847503543, + 0.005256939679384232, + 0.005667218938469887, + 0.008245096541941166, + 0.004197551403194666, + 0.006396854761987925, + 0.012424197979271412, + 0.00549759017303586, + 0.006134204566478729, + 0.004129540175199509, + 0.007069385144859552, + 0.011117011308670044, + 0.011553307995200157, + 0.012087805196642876, + 0.027099354192614555, + 0.09855957329273224, + 0.009133568033576012, + 0.010906914249062538, + 0.014480480924248695, + 0.09016658365726471, + 0.005929257720708847, + 0.014054447412490845, + 0.010205172002315521, + 0.01791180856525898, + 0.00589903537184, + 0.00535836536437273, + 0.011617680080235004, + 0.007492862641811371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.14275622367858887 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_12/index_tensor([288, 288])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_12/index_tensor([288, 288])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6adf2b808213353ebd8ab80df741c9c1088d92f5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_12/index_tensor([288, 288])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, medium-length tracking shot smoothly follows a handsome young wizard as he effortlessly soars through the air. Rendered in vibrant and expressive Disney Pixar style, the wizard features windswept auburn hair and a joyful, confident expression. His opulent, magical costume is a masterpiece of detail: a deep sapphire velvet coat, its ornate high collar and cufflinks adorned with shimmering, iridescent peacock and phoenix feathers that catch the light. Delicate, luminous silver embroidery swirls across the fabric, pulsing with gentle energy. He sails through the air with balletic grace, his body in a gentle arc, one arm extended as if guiding his path, the feathers of his coat fluttering behind him like a silk banner. Below, a whimsical, rolling landscape unfolds, its hills of impossibly soft emerald green lawn dotted with oversized bioluminescent mushrooms and trees with spiraling trunks and cotton-candy pink leaves. The entire scene is bathed in the warm, soft light of the magical golden hour, creating long, soft shadows and making the dewdrops on the dreamy plants shimmer, evoking a profound sense of wonder, freedom, and pure animated magic.", + "seed": 17070, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_12/index_288_seed_17070_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.013581654150038957, + "vlm_logprob": -0.013581654150038957, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.013581654150038957, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02523045614361763 + ], + "opposite_logprob_tokens": [ + -3.7752304077148438 + ], + "logprob_reward": -0.02523045614361763, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02523045614361763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1826181858778, + 0.04630633443593979, + 0.0800456777215004, + 0.09640111774206161, + 0.018643490970134735, + 0.09212184697389603, + 0.042030952870845795, + 0.03872992843389511, + 0.08766317367553711, + 0.31543928384780884 + ], + "credit_map": [ + 0.016868524253368378, + 0.031465768814086914, + 0.044030800461769104, + 0.016590800136327744, + 0.012842291034758091, + 0.02236470952630043, + 0.016942866146564484, + 0.021512433886528015, + 0.004469726234674454, + 0.007798886392265558, + 0.008689528331160545, + 0.004197393544018269, + 0.004260243847966194, + 0.005334422457963228, + 0.0035767846275120974, + 0.007979349233210087, + 0.011435088701546192, + 0.009618331678211689, + 0.011817767284810543, + 0.00916011817753315, + 0.010755575262010098, + 0.010553487576544285, + 0.011022083461284637, + 0.00568322092294693, + 0.00893291924148798, + 0.008484253659844398, + 0.008648921735584736, + 0.010558541864156723, + 0.02333049476146698, + 0.013864222913980484, + 0.010473133064806461, + 0.01210862398147583, + 0.001967039657756686, + 0.0025582073722034693, + 0.0023614917881786823, + 0.0016984286485239863, + 0.002384098246693611, + 0.0038488805294036865, + 0.001961360452696681, + 0.001863983110524714, + 0.009098798967897892, + 0.03053290769457817, + 0.0030659837648272514, + 0.005126120522618294, + 0.005268827546387911, + 0.03179590404033661, + 0.0024890985805541277, + 0.004744201898574829, + 0.005940943956375122, + 0.007858434692025185, + 0.006807255558669567, + 0.0041416180320084095, + 0.004850984085351229, + 0.005202667787671089, + 0.0029773216228932142, + 0.0042517283000051975, + 0.005659950897097588, + 0.008138435892760754, + 0.0027821457479149103, + 0.003945619333535433, + 0.004210576880723238, + 0.005079489666968584, + 0.003157405648380518, + 0.005756305996328592, + 0.009985122829675674, + 0.020189959555864334, + 0.007326515857130289, + 0.007616635877639055, + 0.01579105854034424, + 0.0130836246535182, + 0.006207431200891733, + 0.007462827954441309, + 0.09807714819908142, + 0.015676258131861687, + 0.01489559281617403, + 0.012387631461024284, + 0.00945121981203556, + 0.035457100719213486, + 0.023055419325828552, + 0.10643891245126724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.02523045614361763 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004103098995983601 + ], + "opposite_logprob_tokens": [ + -5.504103183746338 + ], + "logprob_reward": -0.004103098995983601, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004103098995983601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15982399880886078, + 0.07294122129678726, + 0.14359451830387115, + 0.06219670549035072, + 0.024263300001621246, + 0.04538968950510025, + 0.03804757818579674, + 0.039682384580373764, + 0.04657395929098129, + 0.3674866557121277 + ], + "credit_map": [ + 0.018562478944659233, + 0.025694677606225014, + 0.03556157648563385, + 0.01318318210542202, + 0.014842171221971512, + 0.01806960254907608, + 0.0147545849904418, + 0.01915573701262474, + 0.008148297667503357, + 0.011228098534047604, + 0.013289080001413822, + 0.00513118039816618, + 0.009816399775445461, + 0.009844813495874405, + 0.0050784084014594555, + 0.010404937900602818, + 0.021779466420412064, + 0.014243570156395435, + 0.022868193686008453, + 0.013704560697078705, + 0.02001996338367462, + 0.022554902359843254, + 0.01908142678439617, + 0.009342443197965622, + 0.005511323921382427, + 0.006045219954103231, + 0.00827854499220848, + 0.004970448091626167, + 0.011467836797237396, + 0.009605351835489273, + 0.006748596206307411, + 0.009569378569722176, + 0.0023301055189222097, + 0.003341690404340625, + 0.0026012430898845196, + 0.0021655806340277195, + 0.003421003231778741, + 0.00562344491481781, + 0.0018711758311837912, + 0.002909055445343256, + 0.004807720426470041, + 0.008621414192020893, + 0.007730832789093256, + 0.00646051624789834, + 0.0033167048823088408, + 0.00942436046898365, + 0.002914660843089223, + 0.002113480819389224, + 0.004302740562707186, + 0.005895893555134535, + 0.006120037287473679, + 0.004047926515340805, + 0.004372613504528999, + 0.005419078748673201, + 0.0027376979123800993, + 0.005151592195034027, + 0.005912019871175289, + 0.005015912465751171, + 0.004259228240698576, + 0.004421166609972715, + 0.004006388131529093, + 0.007041805423796177, + 0.0035024769604206085, + 0.005523388739675283, + 0.005859692580997944, + 0.00705751683562994, + 0.005054744891822338, + 0.00926851388067007, + 0.004979977384209633, + 0.005242915824055672, + 0.005416726693511009, + 0.003693867940455675, + 0.10825308412313461, + 0.022028641775250435, + 0.017354482784867287, + 0.02023240365087986, + 0.014225881546735764, + 0.039960019290447235, + 0.0265582874417305, + 0.11887384206056595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.004103098995983601 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583592355251312 + ], + "opposite_logprob_tokens": [ + -3.8955836296081543 + ], + "logprob_reward": -0.020583592355251312, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583592355251312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0716894343495369, + 0.0429975725710392, + 0.10127122700214386, + 0.04313727095723152, + 0.02469005063176155, + 0.38227248191833496, + 0.017767099663615227, + 0.08047298341989517, + 0.040838588029146194, + 0.19486340880393982 + ], + "credit_map": [ + 0.0066830250434577465, + 0.011827506124973297, + 0.014072725549340248, + 0.006153923459351063, + 0.007658962160348892, + 0.010743814520537853, + 0.005256928503513336, + 0.009292537346482277, + 0.00632493058219552, + 0.0066434470936656, + 0.006781994830816984, + 0.0030278516933321953, + 0.0031471888069063425, + 0.0050215651281178, + 0.003775235265493393, + 0.008275352418422699, + 0.012653344310820103, + 0.009642661549150944, + 0.024973494932055473, + 0.011661119759082794, + 0.008524012751877308, + 0.014312585815787315, + 0.012076910585165024, + 0.007427088916301727, + 0.005183549597859383, + 0.0034394559916108847, + 0.005038965493440628, + 0.00351291848346591, + 0.007937482558190823, + 0.007369829807430506, + 0.00403928104788065, + 0.00661578169092536, + 0.0025137001648545265, + 0.0027534454129636288, + 0.003925071097910404, + 0.0023813056759536266, + 0.0030962983146309853, + 0.004597187973558903, + 0.002117518801242113, + 0.0033055199310183525, + 0.03833328187465668, + 0.12646956741809845, + 0.01126336120069027, + 0.020065996795892715, + 0.02274656482040882, + 0.13630777597427368, + 0.009891355410218239, + 0.01719450019299984, + 0.0015383362770080566, + 0.002984758699312806, + 0.0014055620413273573, + 0.0014291508123278618, + 0.004094727803021669, + 0.0029895019251853228, + 0.0013907625107094646, + 0.0019342980813235044, + 0.012148676440119743, + 0.014598672278225422, + 0.009363127872347832, + 0.007779313717037439, + 0.007938629016280174, + 0.010767465457320213, + 0.006399589590728283, + 0.011477498337626457, + 0.005435361992567778, + 0.005254427902400494, + 0.0045573515817523, + 0.009236838668584824, + 0.004473712760955095, + 0.0040170191787183285, + 0.004593660589307547, + 0.003270212560892105, + 0.05500167980790138, + 0.013010519556701183, + 0.011003991588950157, + 0.009694762527942657, + 0.010255045257508755, + 0.02262653037905693, + 0.013668199069797993, + 0.05960265174508095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.020583592355251312 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019374187104403973 + ], + "opposite_logprob_tokens": [ + -6.251937389373779 + ], + "logprob_reward": -0.0019374187104403973, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019374187104403973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12118351459503174, + 0.0493757463991642, + 0.039956096559762955, + 0.0574667826294899, + 0.046458903700113297, + 0.07754628360271454, + 0.016956662759184837, + 0.06528004258871078, + 0.06654702872037888, + 0.4592289328575134 + ], + "credit_map": [ + 0.018131913617253304, + 0.017433281987905502, + 0.02225285768508911, + 0.01198587752878666, + 0.00935409963130951, + 0.0147083830088377, + 0.01164123136550188, + 0.015675866976380348, + 0.006266008596867323, + 0.008192505687475204, + 0.010574949905276299, + 0.0035616476088762283, + 0.004990198649466038, + 0.005610604304820299, + 0.002636417979374528, + 0.007543409243226051, + 0.006390722002834082, + 0.004441678989678621, + 0.00795956701040268, + 0.003572817426174879, + 0.005235797259956598, + 0.005700952839106321, + 0.004175826907157898, + 0.0024787322618067265, + 0.009151638485491276, + 0.006100156810134649, + 0.006532710511237383, + 0.004404791630804539, + 0.008719702251255512, + 0.009975472465157509, + 0.004751087166368961, + 0.007831220515072346, + 0.005021121818572283, + 0.006013970356434584, + 0.005598800256848335, + 0.006450835149735212, + 0.005693026352673769, + 0.006553126964718103, + 0.0049804034642875195, + 0.0061476146802306175, + 0.006526842247694731, + 0.019886687397956848, + 0.007124930154532194, + 0.0064787911251187325, + 0.0043502203188836575, + 0.026070695370435715, + 0.003500767983496189, + 0.003607348073273897, + 0.0021708644926548004, + 0.00334448111243546, + 0.0017560971900820732, + 0.0013854491990059614, + 0.0020560501143336296, + 0.0032379860058426857, + 0.0015121599426493049, + 0.0014935744693502784, + 0.008431483060121536, + 0.008907278999686241, + 0.009151941165328026, + 0.007383085321635008, + 0.007756889797747135, + 0.01026932243257761, + 0.00602301862090826, + 0.007357021328061819, + 0.008500589057803154, + 0.008923085406422615, + 0.006011084653437138, + 0.01073151919990778, + 0.010564926080405712, + 0.009085878729820251, + 0.007355084642767906, + 0.0053748637437820435, + 0.11572373658418655, + 0.04782818257808685, + 0.024902258068323135, + 0.030324440449476242, + 0.03167897090315819, + 0.04761550948023796, + 0.03219067305326462, + 0.12896513938903809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.0019374187104403973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016053704544901848 + ], + "opposite_logprob_tokens": [ + -4.141053676605225 + ], + "logprob_reward": -0.016053704544901848, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016053704544901848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08779766410589218, + 0.1811298131942749, + 0.046368058770895004, + 0.08245664834976196, + 0.03377554938197136, + 0.3613077998161316, + 0.021119877696037292, + 0.034276507794857025, + 0.020269490778446198, + 0.13149864971637726 + ], + "credit_map": [ + 0.009376484900712967, + 0.014682063832879066, + 0.01613825000822544, + 0.005733598954975605, + 0.00903566088527441, + 0.01505904272198677, + 0.006269408389925957, + 0.011503150686621666, + 0.020263219252228737, + 0.028132809326052666, + 0.03203918784856796, + 0.010350987315177917, + 0.011412936262786388, + 0.017735010012984276, + 0.01473910640925169, + 0.046456534415483475, + 0.006779137533158064, + 0.006917217746376991, + 0.006429921369999647, + 0.005100547801703215, + 0.007790280506014824, + 0.0055588604882359505, + 0.004557936452329159, + 0.003234150819480419, + 0.010651645250618458, + 0.007716911844909191, + 0.009324890561401844, + 0.007062175776809454, + 0.015711527317762375, + 0.01720062457025051, + 0.005916252266615629, + 0.008872614242136478, + 0.0035745538771152496, + 0.004527613054960966, + 0.005277501884847879, + 0.0036776219494640827, + 0.004096884746104479, + 0.005705296993255615, + 0.0034721018746495247, + 0.003443970577791333, + 0.03355022519826889, + 0.1093999370932579, + 0.013967232778668404, + 0.018317000940442085, + 0.020081719383597374, + 0.13660112023353577, + 0.010883104056119919, + 0.018507475033402443, + 0.002471123356372118, + 0.003488140879198909, + 0.00248059444129467, + 0.001752681564539671, + 0.0033391958568245173, + 0.0028791159857064486, + 0.0016711545176804066, + 0.0030378696974366903, + 0.0036002853885293007, + 0.007258852943778038, + 0.003004910657182336, + 0.00328550161793828, + 0.005930793005973101, + 0.00455644354224205, + 0.003532244125381112, + 0.0031074753496795893, + 0.002433683257550001, + 0.003080173395574093, + 0.001407946110703051, + 0.0021359859965741634, + 0.004245230462402105, + 0.003573669120669365, + 0.001900131581351161, + 0.0014926689909771085, + 0.04149710014462471, + 0.006621865555644035, + 0.005947626195847988, + 0.004538908135145903, + 0.004084539134055376, + 0.01473255641758442, + 0.009561331011354923, + 0.04451470822095871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 9 + ], + "selected_reward": -0.016053704544901848 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, medium-length tracking shot smoothly follows a handsome young wizard as he effortlessly soars through the air. Rendered in vibrant and expressive Disney Pixar style, the wizard features windswept auburn hair and a joyful, confident expression. His opulent, magical costume is a masterpiece of detail: a deep sapphire velvet coat, its ornate high collar and cufflinks adorned with shimmering, iridescent peacock and phoenix feathers that catch the light. Delicate, luminous silver embroidery swirls across the fabric, pulsing with gentle energy. He sails through the air with balletic grace, his body in a gentle arc, one arm extended as if guiding his path, the feathers of his coat fluttering behind him like a silk banner. Below, a whimsical, rolling landscape unfolds, its hills of impossibly soft emerald green lawn dotted with oversized bioluminescent mushrooms and trees with spiraling trunks and cotton-candy pink leaves. The entire scene is bathed in the warm, soft light of the magical golden hour, creating long, soft shadows and making the dewdrops on the dreamy plants shimmer, evoking a profound sense of wonder, freedom, and pure animated magic.", + "seed": 17071, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_12/index_288_seed_17071_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0211546684615314, + "vlm_logprob": -0.0211546684615314, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.0211546684615314, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02210257016122341 + ], + "opposite_logprob_tokens": [ + -3.8971025943756104 + ], + "logprob_reward": -0.02210257016122341, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02210257016122341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16406312584877014, + 0.06992699950933456, + 0.03911777213215828, + 0.1498931497335434, + 0.10747501999139786, + 0.08677075803279877, + 0.03825562074780464, + 0.08522778004407883, + 0.16781006753444672, + 0.091459721326828 + ], + "credit_map": [ + 0.0232485830783844, + 0.019943859428167343, + 0.01950257271528244, + 0.024556143209338188, + 0.006638507358729839, + 0.015192757360637188, + 0.0255096685141325, + 0.029471026733517647, + 0.012472881004214287, + 0.007128767669200897, + 0.007813761942088604, + 0.007220966275781393, + 0.0058583724312484264, + 0.007225761655718088, + 0.011027216911315918, + 0.011179267428815365, + 0.005754303652793169, + 0.0035808037500828505, + 0.005272597540169954, + 0.0036582048051059246, + 0.005022762343287468, + 0.004498676396906376, + 0.003660461399704218, + 0.007669961079955101, + 0.011510385200381279, + 0.011686860583722591, + 0.01812579110264778, + 0.013370492495596409, + 0.015979668125510216, + 0.024939624592661858, + 0.035408299416303635, + 0.01887202449142933, + 0.02943655475974083, + 0.014463615603744984, + 0.018888723105192184, + 0.010945847257971764, + 0.00782232265919447, + 0.009448348544538021, + 0.009171726182103157, + 0.007297887001186609, + 0.010982651263475418, + 0.011806256137788296, + 0.013379327021539211, + 0.009450500831007957, + 0.007284708321094513, + 0.010855540633201599, + 0.011643345467746258, + 0.011368433013558388, + 0.00360616366378963, + 0.0052900598384439945, + 0.004551126156002283, + 0.004497838672250509, + 0.003826810047030449, + 0.0036275419406592846, + 0.007787571754306555, + 0.005068507976830006, + 0.008459365926682949, + 0.01164727658033371, + 0.012727702967822552, + 0.013355741277337074, + 0.011846223846077919, + 0.008754179812967777, + 0.010779881849884987, + 0.007657406851649284, + 0.026441557332873344, + 0.01771101914346218, + 0.04009396955370903, + 0.023173589259386063, + 0.014126456342637539, + 0.011264145374298096, + 0.018455661833286285, + 0.016543669626116753, + 0.012662799097597599, + 0.008919180370867252, + 0.021788250654935837, + 0.01625940203666687, + 0.0077866557985544205, + 0.009837248362600803, + 0.006577160209417343, + 0.007629013620316982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.02210257016122341 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639241870492697 + ], + "opposite_logprob_tokens": [ + -4.882639408111572 + ], + "logprob_reward": -0.007639241870492697, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639241870492697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11344614624977112, + 0.07732460647821426, + 0.08478907495737076, + 0.0974041074514389, + 0.13344383239746094, + 0.15784114599227905, + 0.07472868263721466, + 0.09117534011602402, + 0.11149340867996216, + 0.05835362523794174 + ], + "credit_map": [ + 0.015013457275927067, + 0.01339444238692522, + 0.012135105207562447, + 0.014504624530673027, + 0.0062468936666846275, + 0.012798776850104332, + 0.020168697461485863, + 0.019184142351150513, + 0.011952253058552742, + 0.011850843206048012, + 0.008855274878442287, + 0.007448463700711727, + 0.008020132780075073, + 0.007143786642700434, + 0.008422003127634525, + 0.013631846755743027, + 0.006521138362586498, + 0.007875769399106503, + 0.010079341940581799, + 0.007191464304924011, + 0.008096449077129364, + 0.014353806152939796, + 0.006570364348590374, + 0.02410074509680271, + 0.007589307148009539, + 0.006873295176774263, + 0.013731623999774456, + 0.01062548067420721, + 0.012407057918608189, + 0.012949489988386631, + 0.02410249598324299, + 0.009125354699790478, + 0.021863525733351707, + 0.012966875918209553, + 0.023105472326278687, + 0.017591848969459534, + 0.010774245485663414, + 0.012905608862638474, + 0.021354058757424355, + 0.012882190756499767, + 0.019656484946608543, + 0.033834703266620636, + 0.021184125915169716, + 0.01808485947549343, + 0.013956968672573566, + 0.018382830545306206, + 0.015050527639687061, + 0.01769064925611019, + 0.00889900978654623, + 0.011780955828726292, + 0.009421388618648052, + 0.009173255413770676, + 0.008063962683081627, + 0.00658020656555891, + 0.009734425693750381, + 0.011075477115809917, + 0.007870296016335487, + 0.009956875815987587, + 0.01377211045473814, + 0.01742769591510296, + 0.011047100648283958, + 0.009910966269671917, + 0.011087256483733654, + 0.01010303944349289, + 0.021619318053126335, + 0.010258050635457039, + 0.025091668590903282, + 0.016054049134254456, + 0.010706390254199505, + 0.008732539601624012, + 0.011844021268188953, + 0.0071873739361763, + 0.007690433878451586, + 0.005176529288291931, + 0.01223669108003378, + 0.00906150322407484, + 0.006195316091179848, + 0.007733519189059734, + 0.004738216754049063, + 0.0055214157328009605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.007639241870492697 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029812784865498543 + ], + "opposite_logprob_tokens": [ + -3.529812812805176 + ], + "logprob_reward": -0.029812784865498543, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029812784865498543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16090883314609528, + 0.09682721644639969, + 0.05066051334142685, + 0.06606218218803406, + 0.11897552013397217, + 0.09226278960704803, + 0.07991065084934235, + 0.15674547851085663, + 0.12227511405944824, + 0.05537169426679611 + ], + "credit_map": [ + 0.02817440591752529, + 0.01899722032248974, + 0.022368494421243668, + 0.023517005145549774, + 0.012033926323056221, + 0.015506323426961899, + 0.01893947646021843, + 0.02137199230492115, + 0.01118395570665598, + 0.009335444308817387, + 0.013361239805817604, + 0.011802667751908302, + 0.009661369025707245, + 0.011400205083191395, + 0.014403671957552433, + 0.01567867398262024, + 0.00533076049759984, + 0.0040381792932748795, + 0.006778282579034567, + 0.006936111953109503, + 0.009111863560974598, + 0.0056283012963831425, + 0.005698290653526783, + 0.007138724904507399, + 0.006154332775622606, + 0.006227627396583557, + 0.008953467942774296, + 0.007917871698737144, + 0.007374468259513378, + 0.011167260818183422, + 0.011227793991565704, + 0.00703936954960227, + 0.01712537184357643, + 0.01705477014183998, + 0.012676777318120003, + 0.014834604226052761, + 0.01400443259626627, + 0.011105729267001152, + 0.020844824612140656, + 0.011329024098813534, + 0.009823374450206757, + 0.0120529904961586, + 0.014648349955677986, + 0.010021485388278961, + 0.007872644811868668, + 0.009287738241255283, + 0.01384040154516697, + 0.014715810306370258, + 0.005410276353359222, + 0.008159142918884754, + 0.009658913128077984, + 0.008270747028291225, + 0.005988528020679951, + 0.008277502842247486, + 0.02051783911883831, + 0.013627707026898861, + 0.015634827315807343, + 0.016209008172154427, + 0.023437485098838806, + 0.02539222128689289, + 0.018343837931752205, + 0.01654898375272751, + 0.023648951202630997, + 0.017530176788568497, + 0.02725760079920292, + 0.01312930416315794, + 0.03190328925848007, + 0.015239009633660316, + 0.00829916913062334, + 0.00638606445863843, + 0.012660508044064045, + 0.007400183007121086, + 0.007700704038143158, + 0.0034804169554263353, + 0.011412921361625195, + 0.00939832255244255, + 0.006493112072348595, + 0.007729975041002035, + 0.005325653590261936, + 0.0038305933121591806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.029812784865498543 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537471950054169 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537471950054169, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537471950054169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20474177598953247, + 0.07961641252040863, + 0.12559141218662262, + 0.14695915579795837, + 0.03146303817629814, + 0.13026627898216248, + 0.05552256107330322, + 0.04879669472575188, + 0.10627599060535431, + 0.07076668739318848 + ], + "credit_map": [ + 0.035002753138542175, + 0.01934966631233692, + 0.024774521589279175, + 0.0296347476541996, + 0.019437426701188087, + 0.019441526383161545, + 0.023497553542256355, + 0.033603567630052567, + 0.010554427281022072, + 0.00707005150616169, + 0.010087917558848858, + 0.00850225705653429, + 0.00799567624926567, + 0.012742409482598305, + 0.009849931113421917, + 0.012813746929168701, + 0.008768984116613865, + 0.008411195129156113, + 0.021324433386325836, + 0.009116696193814278, + 0.0128978556022048, + 0.023428265005350113, + 0.009381239302456379, + 0.032262738794088364, + 0.009367357939481735, + 0.00882703810930252, + 0.02521773986518383, + 0.015932302922010422, + 0.017224058508872986, + 0.019388245418667793, + 0.036897558718919754, + 0.014104867354035378, + 0.005105046089738607, + 0.002515254309400916, + 0.004724610131233931, + 0.004612123128026724, + 0.002829133067280054, + 0.002433530054986477, + 0.0061631593853235245, + 0.00308018340729177, + 0.0143003910779953, + 0.01727858930826187, + 0.025764184072613716, + 0.013103890232741833, + 0.015546763315796852, + 0.014479989185929298, + 0.013876590877771378, + 0.01591588743031025, + 0.006034931633621454, + 0.005431005731225014, + 0.008075781166553497, + 0.008287389762699604, + 0.006110692396759987, + 0.006207518745213747, + 0.00860163290053606, + 0.006773609202355146, + 0.005309464875608683, + 0.0057429661974310875, + 0.007224857807159424, + 0.007804900407791138, + 0.0067535326816141605, + 0.005274554714560509, + 0.005770577583462, + 0.004915840458124876, + 0.021298706531524658, + 0.01171407662332058, + 0.021466108039021492, + 0.01503276638686657, + 0.010146453976631165, + 0.007587423548102379, + 0.01090303622186184, + 0.008127423003315926, + 0.008743960410356522, + 0.008293178863823414, + 0.014577989466488361, + 0.011582151986658573, + 0.00539244944229722, + 0.006323192268610001, + 0.005442665424197912, + 0.010411093942821026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.012537471950054169 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033681273460388184 + ], + "opposite_logprob_tokens": [ + -3.4086813926696777 + ], + "logprob_reward": -0.033681273460388184, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033681273460388184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2142261415719986, + 0.09163105487823486, + 0.07788988202810287, + 0.06813359260559082, + 0.09940019994974136, + 0.04971298575401306, + 0.05075744166970253, + 0.1361895203590393, + 0.135917067527771, + 0.0761420801281929 + ], + "credit_map": [ + 0.03096003085374832, + 0.027980146929621696, + 0.021288802847266197, + 0.03625127300620079, + 0.016889672726392746, + 0.021249359473586082, + 0.02988884225487709, + 0.029718054458498955, + 0.02067297138273716, + 0.009429052472114563, + 0.019461696967482567, + 0.008268252946436405, + 0.006283929571509361, + 0.008486788719892502, + 0.008216743357479572, + 0.01081162877380848, + 0.007729588076472282, + 0.005699981469660997, + 0.010440830141305923, + 0.007411370053887367, + 0.00863572582602501, + 0.013114823959767818, + 0.006751390174031258, + 0.018106183037161827, + 0.005690076854079962, + 0.006068205926567316, + 0.007541372440755367, + 0.00824078731238842, + 0.009724232368171215, + 0.009135449305176735, + 0.013090165331959724, + 0.008643309585750103, + 0.010721738450229168, + 0.012331167235970497, + 0.02047443576157093, + 0.012902025133371353, + 0.009503871202468872, + 0.008213922381401062, + 0.015947746112942696, + 0.009305305778980255, + 0.0044757043942809105, + 0.0062052481807768345, + 0.009376238100230694, + 0.004786280449479818, + 0.005410688929259777, + 0.0048042419366538525, + 0.0066465698182582855, + 0.00800801906734705, + 0.005584673024713993, + 0.006927503738552332, + 0.0064810276962816715, + 0.006514625623822212, + 0.006490250118076801, + 0.00502735422924161, + 0.006801912561058998, + 0.006930099800229073, + 0.018487069755792618, + 0.01887318678200245, + 0.019006479531526566, + 0.01997675560414791, + 0.022152887657284737, + 0.01275230199098587, + 0.013947584666311741, + 0.010993268340826035, + 0.03304456174373627, + 0.01005062460899353, + 0.037665095180273056, + 0.017438024282455444, + 0.008936312049627304, + 0.006723721977323294, + 0.014009395614266396, + 0.008049351163208485, + 0.009644589386880398, + 0.007768741808831692, + 0.01811089925467968, + 0.01226133480668068, + 0.007684527896344662, + 0.006467605475336313, + 0.006943032145500183, + 0.007261355873197317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.033681273460388184 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, medium-length tracking shot smoothly follows a handsome young wizard as he effortlessly soars through the air. Rendered in vibrant and expressive Disney Pixar style, the wizard features windswept auburn hair and a joyful, confident expression. His opulent, magical costume is a masterpiece of detail: a deep sapphire velvet coat, its ornate high collar and cufflinks adorned with shimmering, iridescent peacock and phoenix feathers that catch the light. Delicate, luminous silver embroidery swirls across the fabric, pulsing with gentle energy. He sails through the air with balletic grace, his body in a gentle arc, one arm extended as if guiding his path, the feathers of his coat fluttering behind him like a silk banner. Below, a whimsical, rolling landscape unfolds, its hills of impossibly soft emerald green lawn dotted with oversized bioluminescent mushrooms and trees with spiraling trunks and cotton-candy pink leaves. The entire scene is bathed in the warm, soft light of the magical golden hour, creating long, soft shadows and making the dewdrops on the dreamy plants shimmer, evoking a profound sense of wonder, freedom, and pure animated magic.", + "seed": 17072, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_12/index_288_seed_17072_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01749364463612437, + "vlm_logprob": -0.01749364463612437, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.01749364463612437, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02705676481127739 + ], + "opposite_logprob_tokens": [ + -3.6520566940307617 + ], + "logprob_reward": -0.02705676481127739, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02705676481127739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09397903829813004, + 0.027414532378315926, + 0.055435582995414734, + 0.20091542601585388, + 0.2284637689590454, + 0.023997986689209938, + 0.0539691224694252, + 0.22745808959007263, + 0.02935735508799553, + 0.0590091347694397 + ], + "credit_map": [ + 0.014596548862755299, + 0.01355914119631052, + 0.010426178574562073, + 0.00704118562862277, + 0.01053912565112114, + 0.013826430775225163, + 0.014705603010952473, + 0.009284812025725842, + 0.0033601021859794855, + 0.0031841169111430645, + 0.0041777449660003185, + 0.0025697448290884495, + 0.0033455989323556423, + 0.004698129836469889, + 0.0029950658790767193, + 0.0030840281397104263, + 0.006966054905205965, + 0.008698541671037674, + 0.01034132856875658, + 0.0044034025631845, + 0.006875266786664724, + 0.004313424229621887, + 0.007319412659853697, + 0.006518142763525248, + 0.029766865074634552, + 0.02147812210023403, + 0.012638903222978115, + 0.01106756180524826, + 0.04416189342737198, + 0.03367747738957405, + 0.02036301977932453, + 0.027761580422520638, + 0.05689116567373276, + 0.021917322650551796, + 0.03832797706127167, + 0.010278444737195969, + 0.006851706188172102, + 0.015454934909939766, + 0.007915359921753407, + 0.07082683593034744, + 0.0022283904254436493, + 0.002305843634530902, + 0.003807837376371026, + 0.0026536101941019297, + 0.003301439341157675, + 0.002082996303215623, + 0.0025153548922389746, + 0.005102512892335653, + 0.004502752795815468, + 0.0048539345152676105, + 0.006474534049630165, + 0.009747921489179134, + 0.007519400212913752, + 0.005247017834335566, + 0.009071042761206627, + 0.006552517879754305, + 0.01706843450665474, + 0.018246518447995186, + 0.04319240152835846, + 0.044608522206544876, + 0.02914537489414215, + 0.020991085097193718, + 0.024966269731521606, + 0.029239468276500702, + 0.004916675854474306, + 0.0027758891228586435, + 0.003614510875195265, + 0.004320899024605751, + 0.0029043925460428, + 0.002758335554972291, + 0.004917344078421593, + 0.003149305237457156, + 0.0071259899996221066, + 0.00843706913292408, + 0.008262342773377895, + 0.007669208571314812, + 0.00778138916939497, + 0.0042248317040503025, + 0.009573182091116905, + 0.005935114808380604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.02705676481127739 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595379201695323 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.00595379201695323, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595379201695323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22164252400398254, + 0.059507232159376144, + 0.0756429135799408, + 0.09120471775531769, + 0.27431440353393555, + 0.018701236695051193, + 0.033002521842718124, + 0.05128394812345505, + 0.03767592832446098, + 0.13702455163002014 + ], + "credit_map": [ + 0.030628323554992676, + 0.02791331149637699, + 0.024076515808701515, + 0.018196487799286842, + 0.031601160764694214, + 0.03771762549877167, + 0.031919777393341064, + 0.019589336588978767, + 0.0079964529722929, + 0.007094856817275286, + 0.00671607069671154, + 0.006180924363434315, + 0.007961694151163101, + 0.009268651716411114, + 0.008014057762920856, + 0.006274526007473469, + 0.009572996757924557, + 0.011916419491171837, + 0.010690249502658844, + 0.006070055533200502, + 0.011707743629813194, + 0.007819156162440777, + 0.008239194750785828, + 0.00962709542363882, + 0.013310077600181103, + 0.015308312140405178, + 0.0093332314863801, + 0.009147451259195805, + 0.012005752883851528, + 0.013037286698818207, + 0.007439714390784502, + 0.011622886173427105, + 0.06776580959558487, + 0.026899654418230057, + 0.04658184573054314, + 0.012828829698264599, + 0.007941055111587048, + 0.018680676817893982, + 0.009831830859184265, + 0.08378469944000244, + 0.0020357887260615826, + 0.0027382392436265945, + 0.0028045561630278826, + 0.0016356363194063306, + 0.002542386297136545, + 0.001555144670419395, + 0.0019499467452988029, + 0.003439538413658738, + 0.003189666196703911, + 0.0030731335282325745, + 0.003932338207960129, + 0.006208106409758329, + 0.00386505457572639, + 0.0035641135182231665, + 0.004732221830636263, + 0.004437884781509638, + 0.004380783997476101, + 0.005458589177578688, + 0.006841742899268866, + 0.007698971312493086, + 0.006765149068087339, + 0.009730692021548748, + 0.0047613573260605335, + 0.005646661389619112, + 0.006505847442895174, + 0.003647511126473546, + 0.005212015472352505, + 0.004573725163936615, + 0.0061783636920154095, + 0.003541567362844944, + 0.004233959596604109, + 0.003782939864322543, + 0.016647927463054657, + 0.018315250054001808, + 0.014559985138475895, + 0.018350958824157715, + 0.02170400880277157, + 0.013013500720262527, + 0.021114889532327652, + 0.013318033888936043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.00595379201695323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058872953057289 + ], + "opposite_logprob_tokens": [ + -3.8955886363983154 + ], + "logprob_reward": -0.02058872953057289, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058872953057289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11032289266586304, + 0.053414516150951385, + 0.07779451459646225, + 0.0707799419760704, + 0.44624024629592896, + 0.02269803173840046, + 0.025467848405241966, + 0.05480453372001648, + 0.045767784118652344, + 0.09270969033241272 + ], + "credit_map": [ + 0.01636020839214325, + 0.014221354387700558, + 0.018000679090619087, + 0.00805725995451212, + 0.015144369564950466, + 0.01404417585581541, + 0.012860896065831184, + 0.011633945629000664, + 0.006972291506826878, + 0.006041257176548243, + 0.007508903741836548, + 0.0051567102782428265, + 0.0058613987639546394, + 0.006533081643283367, + 0.009461142122745514, + 0.005879731848835945, + 0.00965051632374525, + 0.010743960738182068, + 0.013037823140621185, + 0.004967663437128067, + 0.010356161743402481, + 0.00796574354171753, + 0.011698289774358273, + 0.0093743447214365, + 0.008853234350681305, + 0.008609341457486153, + 0.007980247028172016, + 0.007292298134416342, + 0.01108287088572979, + 0.009940988384187222, + 0.007801628205925226, + 0.009219332598149776, + 0.11069245636463165, + 0.043960027396678925, + 0.07560763508081436, + 0.020577523857355118, + 0.0127046974375844, + 0.029578447341918945, + 0.016521060839295387, + 0.1365983486175537, + 0.002884410321712494, + 0.0031658848747611046, + 0.003774915123358369, + 0.0023755391594022512, + 0.002074239309877157, + 0.001534621580503881, + 0.0027967821806669235, + 0.004091637674719095, + 0.003911590203642845, + 0.003600876545533538, + 0.002739478135481477, + 0.0038519317749887705, + 0.0029855063185095787, + 0.0021907316986471415, + 0.003189093666151166, + 0.0029986368026584387, + 0.004259871318936348, + 0.005125065799802542, + 0.010073957964777946, + 0.007859594188630581, + 0.00699639180675149, + 0.0052371760830283165, + 0.006032746750861406, + 0.009219723753631115, + 0.006798821035772562, + 0.004220840521156788, + 0.005966839846223593, + 0.005501593928784132, + 0.006609117612242699, + 0.005426183342933655, + 0.004790712613612413, + 0.006453671958297491, + 0.011909900233149529, + 0.012197858653962612, + 0.012687056325376034, + 0.013170254416763783, + 0.011344457045197487, + 0.007493377197533846, + 0.014308993704617023, + 0.009597787633538246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.02058872953057289 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040906332433223724 + ], + "opposite_logprob_tokens": [ + -5.504090785980225 + ], + "logprob_reward": -0.0040906332433223724, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040906332433223724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2113976627588272, + 0.04057889059185982, + 0.03817136585712433, + 0.054962482303380966, + 0.395477294921875, + 0.026544135063886642, + 0.020882656797766685, + 0.05760195851325989, + 0.06980197131633759, + 0.08458162099123001 + ], + "credit_map": [ + 0.028166064992547035, + 0.026679877191781998, + 0.022011475637555122, + 0.02028663083910942, + 0.025185013189911842, + 0.032222162932157516, + 0.03490636870265007, + 0.0219400804489851, + 0.006696618627756834, + 0.004812040366232395, + 0.004162829834967852, + 0.0046974956057965755, + 0.006615652237087488, + 0.005084330681711435, + 0.00493814330548048, + 0.0035717820283025503, + 0.004826478660106659, + 0.006421833764761686, + 0.008217266760766506, + 0.0021667678374797106, + 0.0043198768980801105, + 0.0032034723553806543, + 0.003837690921500325, + 0.005177975632250309, + 0.009099125862121582, + 0.0065641324035823345, + 0.00739113288000226, + 0.004432057496160269, + 0.005087129771709442, + 0.007248982787132263, + 0.005818215664476156, + 0.009321705438196659, + 0.0982474833726883, + 0.037856608629226685, + 0.06730298697948456, + 0.018229901790618896, + 0.011262046173214912, + 0.02648356556892395, + 0.013896799646317959, + 0.12219791114330292, + 0.003279247088357806, + 0.0061735427007079124, + 0.003582937642931938, + 0.0024147185031324625, + 0.0025179858785122633, + 0.0020857383497059345, + 0.0016377028077840805, + 0.004852264653891325, + 0.003532163565978408, + 0.002228315221145749, + 0.002482484094798565, + 0.001979451160877943, + 0.0030066603794693947, + 0.0029083858244121075, + 0.0018778464291244745, + 0.002867351286113262, + 0.005388260819017887, + 0.007069888524711132, + 0.009250493720173836, + 0.00873840693384409, + 0.005375965964049101, + 0.0069550261832773685, + 0.005429259035736322, + 0.009394659660756588, + 0.016873706132173538, + 0.008248395286500454, + 0.00903417356312275, + 0.008532152511179447, + 0.006912978831678629, + 0.0055920761078596115, + 0.007081769872456789, + 0.007526714354753494, + 0.011005370877683163, + 0.013299897313117981, + 0.010590612888336182, + 0.010319323278963566, + 0.01091848686337471, + 0.008317587897181511, + 0.011276942677795887, + 0.008853397332131863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.0040906332433223724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02977830357849598 + ], + "opposite_logprob_tokens": [ + -3.529778242111206 + ], + "logprob_reward": -0.02977830357849598, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02977830357849598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12203145772218704, + 0.08398537337779999, + 0.1822780966758728, + 0.1027858778834343, + 0.33811843395233154, + 0.04940548539161682, + 0.012741218321025372, + 0.03536510840058327, + 0.044428396970033646, + 0.02886057086288929 + ], + "credit_map": [ + 0.017441637814044952, + 0.01623317040503025, + 0.012017002329230309, + 0.010147346183657646, + 0.012925485149025917, + 0.02032354660332203, + 0.018259283155202866, + 0.014683985151350498, + 0.014476398006081581, + 0.01192757673561573, + 0.009445232339203358, + 0.008957806043326855, + 0.010497941635549068, + 0.010221756063401699, + 0.010379350744187832, + 0.008079316467046738, + 0.02028142288327217, + 0.02852041646838188, + 0.03506331145763397, + 0.017501067370176315, + 0.02548522688448429, + 0.014336993917822838, + 0.02172403410077095, + 0.019365619868040085, + 0.017853988334536552, + 0.012259933166205883, + 0.012781420722603798, + 0.007236573379486799, + 0.012966888025403023, + 0.009962126612663269, + 0.01057726051658392, + 0.019147690385580063, + 0.08473396301269531, + 0.032725896686315536, + 0.057599905878305435, + 0.01520922314375639, + 0.009474188089370728, + 0.022648125886917114, + 0.010714258998632431, + 0.10501287132501602, + 0.006631838623434305, + 0.007344882003962994, + 0.009598693810403347, + 0.003687642514705658, + 0.003683021990582347, + 0.003579205134883523, + 0.003034661989659071, + 0.011845538392663002, + 0.0018110342789441347, + 0.0017371723661199212, + 0.0016717021353542805, + 0.0017389734275639057, + 0.0015426117461174726, + 0.00121123599819839, + 0.0009865566389635205, + 0.0020419321954250336, + 0.002920652972534299, + 0.003613799810409546, + 0.004326981958001852, + 0.006455018185079098, + 0.005078571382910013, + 0.0043603284284472466, + 0.004296008497476578, + 0.004313742276281118, + 0.005232470575720072, + 0.004202740732580423, + 0.005024177487939596, + 0.0059935469180345535, + 0.005857208278030157, + 0.005599671509116888, + 0.004942331463098526, + 0.007576249539852142, + 0.00419019628316164, + 0.00512021966278553, + 0.004742438904941082, + 0.003947370685636997, + 0.003011053428053856, + 0.0022481984924525023, + 0.0030707218684256077, + 0.002530371304601431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.02977830357849598 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, medium-length tracking shot smoothly follows a handsome young wizard as he effortlessly soars through the air. Rendered in vibrant and expressive Disney Pixar style, the wizard features windswept auburn hair and a joyful, confident expression. His opulent, magical costume is a masterpiece of detail: a deep sapphire velvet coat, its ornate high collar and cufflinks adorned with shimmering, iridescent peacock and phoenix feathers that catch the light. Delicate, luminous silver embroidery swirls across the fabric, pulsing with gentle energy. He sails through the air with balletic grace, his body in a gentle arc, one arm extended as if guiding his path, the feathers of his coat fluttering behind him like a silk banner. Below, a whimsical, rolling landscape unfolds, its hills of impossibly soft emerald green lawn dotted with oversized bioluminescent mushrooms and trees with spiraling trunks and cotton-candy pink leaves. The entire scene is bathed in the warm, soft light of the magical golden hour, creating long, soft shadows and making the dewdrops on the dreamy plants shimmer, evoking a profound sense of wonder, freedom, and pure animated magic.", + "seed": 17073, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_12/index_288_seed_17073_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07548682084307075, + "vlm_logprob": -0.07548682084307075, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.07548682084307075, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31648871302604675 + ], + "opposite_logprob_tokens": [ + -1.3164887428283691 + ], + "logprob_reward": -0.31648871302604675, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31648871302604675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15996327996253967, + 0.08761165291070938, + 0.04767969995737076, + 0.03665175288915634, + 0.03554961085319519, + 0.022815296426415443, + 0.06812060624361038, + 0.050960712134838104, + 0.01990608312189579, + 0.47074127197265625 + ], + "credit_map": [ + 0.01688673533499241, + 0.02310894802212715, + 0.01701219566166401, + 0.015537397004663944, + 0.021640421822667122, + 0.031601663678884506, + 0.016762111335992813, + 0.017413800582289696, + 0.010477544739842415, + 0.008297686465084553, + 0.00924219936132431, + 0.008630398660898209, + 0.019460823386907578, + 0.013557144440710545, + 0.008761265315115452, + 0.009184591472148895, + 0.005659032613039017, + 0.005911789368838072, + 0.005163155030459166, + 0.005785080138593912, + 0.005133194848895073, + 0.005653452128171921, + 0.009150379337370396, + 0.005223618820309639, + 0.003405071096494794, + 0.003938442561775446, + 0.00397745193913579, + 0.00446839677169919, + 0.004413713235408068, + 0.007097669877111912, + 0.004300795495510101, + 0.005050212610512972, + 0.0038757759612053633, + 0.005026727449148893, + 0.004317468497902155, + 0.004854564554989338, + 0.004956238903105259, + 0.003963159862905741, + 0.004618228413164616, + 0.003937446046620607, + 0.0022920819465070963, + 0.001974942162632942, + 0.0022376328706741333, + 0.0022201251704245806, + 0.0029559920076280832, + 0.0033366831485182047, + 0.0027856866363435984, + 0.005012152716517448, + 0.008076800033450127, + 0.006179285701364279, + 0.005521044135093689, + 0.005435678642243147, + 0.018200837075710297, + 0.008927574381232262, + 0.006669469177722931, + 0.009109919890761375, + 0.004444661550223827, + 0.004773567896336317, + 0.0037792357616126537, + 0.005350302904844284, + 0.012081312015652657, + 0.006967551540583372, + 0.005948673002421856, + 0.007615411654114723, + 0.0023625718895345926, + 0.001836851122789085, + 0.001968108583241701, + 0.0017116202507168055, + 0.002734735142439604, + 0.003535142634063959, + 0.0027224922087043524, + 0.003034562338143587, + 0.199154794216156, + 0.03146081045269966, + 0.012375127524137497, + 0.011401450261473656, + 0.014526003040373325, + 0.018401097506284714, + 0.013087240979075432, + 0.17033475637435913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.31648871302604675 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003626557532697916 + ], + "opposite_logprob_tokens": [ + -5.628626346588135 + ], + "logprob_reward": -0.003626557532697916, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003626557532697916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08977089077234268, + 0.09054995328187943, + 0.038978006690740585, + 0.04651496559381485, + 0.04586023464798927, + 0.0372946597635746, + 0.1656159907579422, + 0.07606597244739532, + 0.032667070627212524, + 0.3766823709011078 + ], + "credit_map": [ + 0.010134904645383358, + 0.011220637708902359, + 0.013953588902950287, + 0.009707985445857048, + 0.011522499844431877, + 0.012121354229748249, + 0.009915134869515896, + 0.011194773018360138, + 0.011527360416948795, + 0.01011725701391697, + 0.011654365807771683, + 0.01104014553129673, + 0.010812654159963131, + 0.014207205735147, + 0.011406798847019672, + 0.009784157387912273, + 0.004666164983063936, + 0.004775701090693474, + 0.004164929036051035, + 0.004916426260024309, + 0.004437551833689213, + 0.004031956195831299, + 0.007355422247201204, + 0.004629851318895817, + 0.0060726348310709, + 0.006554490886628628, + 0.005271477159112692, + 0.007520406972616911, + 0.005956048611551523, + 0.005024875048547983, + 0.00541397649794817, + 0.004701052792370319, + 0.005724854301661253, + 0.004518528934568167, + 0.00594990374520421, + 0.0044817132875323296, + 0.008367901667952538, + 0.006502087228000164, + 0.0049781533889472485, + 0.005337086971849203, + 0.003578809555619955, + 0.0032155984081327915, + 0.0026876747142523527, + 0.0028999533969908953, + 0.004954061470925808, + 0.006331653334200382, + 0.005566275212913752, + 0.008060627616941929, + 0.01975993625819683, + 0.01429288275539875, + 0.016873590648174286, + 0.017958249896764755, + 0.03368333354592323, + 0.016323795542120934, + 0.023754708468914032, + 0.022969471290707588, + 0.007174578495323658, + 0.007786867208778858, + 0.005259955767542124, + 0.008040855638682842, + 0.014512170106172562, + 0.011037602089345455, + 0.010748724453151226, + 0.01150520984083414, + 0.0040876604616642, + 0.005487320013344288, + 0.002370039466768503, + 0.0025991410948336124, + 0.00550741795450449, + 0.004684019833803177, + 0.0036437076050788164, + 0.004287759307771921, + 0.15851688385009766, + 0.023355843499302864, + 0.009533786214888096, + 0.009455324150621891, + 0.011528889648616314, + 0.014437823556363583, + 0.011962511576712132, + 0.1378912627696991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.003626557532697916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026356296613812447 + ], + "opposite_logprob_tokens": [ + -3.6513562202453613 + ], + "logprob_reward": -0.026356296613812447, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026356296613812447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16229724884033203, + 0.10615210235118866, + 0.06408608704805374, + 0.09162981063127518, + 0.10351870954036713, + 0.0667017325758934, + 0.05833038315176964, + 0.06791646033525467, + 0.04815080016851425, + 0.2312166541814804 + ], + "credit_map": [ + 0.020049618557095528, + 0.025127358734607697, + 0.02358085848391056, + 0.020957546308636665, + 0.016980566084384918, + 0.019941296428442, + 0.015977350994944572, + 0.019682645797729492, + 0.015658697113394737, + 0.010869207791984081, + 0.012954658828675747, + 0.011742710135877132, + 0.012740615755319595, + 0.017170598730444908, + 0.013605590909719467, + 0.011410025879740715, + 0.008898337371647358, + 0.007359514478594065, + 0.0069281370379030704, + 0.00735863670706749, + 0.006076162215322256, + 0.011167471297085285, + 0.008714256808161736, + 0.007583568803966045, + 0.01108547579497099, + 0.009210504591464996, + 0.010452013462781906, + 0.011715094558894634, + 0.01445109024643898, + 0.013054020702838898, + 0.008919596672058105, + 0.012742017395794392, + 0.014054695144295692, + 0.008853877894580364, + 0.011746540665626526, + 0.010842742398381233, + 0.01404793094843626, + 0.014671755023300648, + 0.016365142539143562, + 0.012936023995280266, + 0.005449317861348391, + 0.005275583826005459, + 0.004458259791135788, + 0.006151616107672453, + 0.008826534263789654, + 0.013711112551391125, + 0.007583826780319214, + 0.015245480462908745, + 0.011117285117506981, + 0.006148202810436487, + 0.0072895376943051815, + 0.005293100140988827, + 0.008764604106545448, + 0.0057115270756185055, + 0.006884370930492878, + 0.007121759466826916, + 0.00668539572507143, + 0.006706042215228081, + 0.00506583834066987, + 0.0074869138188660145, + 0.012331037782132626, + 0.009481862187385559, + 0.010086772963404655, + 0.010072598233819008, + 0.006284821778535843, + 0.00844998937100172, + 0.004259123466908932, + 0.003583472454920411, + 0.00699026370421052, + 0.006613665260374546, + 0.005467124283313751, + 0.006502341944724321, + 0.08190785348415375, + 0.012086560018360615, + 0.009413366205990314, + 0.009268276393413544, + 0.01749713160097599, + 0.011287646368145943, + 0.01317680161446333, + 0.0765790268778801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.026356296613812447 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633641801774502 + ], + "opposite_logprob_tokens": [ + -5.37963342666626 + ], + "logprob_reward": -0.004633641801774502, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633641801774502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20749707520008087, + 0.05549951270222664, + 0.04675179719924927, + 0.12270107865333557, + 0.07462706416845322, + 0.03184928745031357, + 0.02987351268529892, + 0.04179680347442627, + 0.030060630291700363, + 0.3593432903289795 + ], + "credit_map": [ + 0.022479014471173286, + 0.030508365482091904, + 0.03082519955933094, + 0.02454485557973385, + 0.023946618661284447, + 0.024819934740662575, + 0.024997703731060028, + 0.0253753624856472, + 0.005333328619599342, + 0.006039260886609554, + 0.008094550110399723, + 0.006132899317890406, + 0.00815078429877758, + 0.009612728841602802, + 0.006058720871806145, + 0.0060772341676056385, + 0.006781097501516342, + 0.006034297868609428, + 0.005336730275303125, + 0.004690682049840689, + 0.005390520207583904, + 0.00683550676330924, + 0.005904834251850843, + 0.005778120364993811, + 0.01908351294696331, + 0.013917366974055767, + 0.012236065231263638, + 0.012359467335045338, + 0.015567396767437458, + 0.018672313541173935, + 0.016403289511799812, + 0.014461654238402843, + 0.006922744680196047, + 0.009680885821580887, + 0.008696729317307472, + 0.009399817325174809, + 0.009963314980268478, + 0.009234697557985783, + 0.01200107578188181, + 0.008727786131203175, + 0.002742369659245014, + 0.0030004752334207296, + 0.0033508967608213425, + 0.0038282107561826706, + 0.0051973615773022175, + 0.004967740271240473, + 0.0033642223570495844, + 0.005398009438067675, + 0.003812334267422557, + 0.0032203595619648695, + 0.003242173232138157, + 0.0030832195188850164, + 0.005663824267685413, + 0.003428774420171976, + 0.0036465178709477186, + 0.0037763065192848444, + 0.003669107099995017, + 0.006863081827759743, + 0.0035096961073577404, + 0.003484124783426523, + 0.004907549824565649, + 0.004692916292697191, + 0.011142642237246037, + 0.003527680179104209, + 0.004087179899215698, + 0.0030993290711194277, + 0.0020776642486453056, + 0.002977406606078148, + 0.004192191641777754, + 0.005100200418382883, + 0.004629639443010092, + 0.003897015005350113, + 0.1476992517709732, + 0.018610665574669838, + 0.009273102506995201, + 0.00968926027417183, + 0.01406632550060749, + 0.013371494598686695, + 0.013821578584611416, + 0.13281159102916718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.004633641801774502 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young wizard soaring through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the wizard have windswept auburn hair and a joyful, confident expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the wizard wearing a deep sapphire velvet coat with ornate high collar and cufflinks adorned with iridescent peacock and phoenix feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a whimsical landscape with emerald green hills, oversized bioluminescent mushrooms, and trees with spiraling trunks and cotton-candy pink leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-length tracking shot that follows the wizard smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02632889524102211 + ], + "opposite_logprob_tokens": [ + -3.6513288021087646 + ], + "logprob_reward": -0.02632889524102211, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02632889524102211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1732146292924881, + 0.10281842201948166, + 0.1058632954955101, + 0.06056962162256241, + 0.12977920472621918, + 0.04353692755103111, + 0.05038983002305031, + 0.10554005205631256, + 0.12664096057415009, + 0.10164705663919449 + ], + "credit_map": [ + 0.016654569655656815, + 0.023626329377293587, + 0.026909520849585533, + 0.018103979527950287, + 0.015972351655364037, + 0.025895019993185997, + 0.023922869935631752, + 0.02212999016046524, + 0.015764787793159485, + 0.009872222319245338, + 0.011055264621973038, + 0.009898531250655651, + 0.013191286474466324, + 0.018422789871692657, + 0.012854610569775105, + 0.011758926324546337, + 0.017057739198207855, + 0.011881222017109394, + 0.011520205065608025, + 0.010900461114943027, + 0.011005943641066551, + 0.015185094438493252, + 0.015166795812547207, + 0.013145837001502514, + 0.006877693347632885, + 0.011122805997729301, + 0.00595872662961483, + 0.008256621658802032, + 0.005496206693351269, + 0.00892815925180912, + 0.006494089495390654, + 0.007435313425958157, + 0.012730156071484089, + 0.01555430144071579, + 0.016629528254270554, + 0.01761970855295658, + 0.01501760445535183, + 0.013399987481534481, + 0.02365432120859623, + 0.015173602849245071, + 0.004251779522746801, + 0.0050253295339643955, + 0.005921629257500172, + 0.005784416105598211, + 0.005103405099362135, + 0.005144566297531128, + 0.005259519908577204, + 0.0070462836883962154, + 0.006675309967249632, + 0.005241289734840393, + 0.00507262023165822, + 0.005052037071436644, + 0.008564244024455547, + 0.005232434719800949, + 0.008716822601854801, + 0.0058350712060928345, + 0.009236805140972137, + 0.007679615169763565, + 0.007829799316823483, + 0.009963699616491795, + 0.02310790866613388, + 0.012093737721443176, + 0.02494015172123909, + 0.010688336566090584, + 0.013914496637880802, + 0.026991451159119606, + 0.008460689336061478, + 0.012315653264522552, + 0.021605340763926506, + 0.012931929901242256, + 0.013677953742444515, + 0.016743451356887817, + 0.02189681865274906, + 0.007986507378518581, + 0.010425249114632607, + 0.009139315225183964, + 0.009557747282087803, + 0.011948194354772568, + 0.007628117688000202, + 0.02306511625647545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.02632889524102211 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_13/index_tensor([9788, 9788])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_13/index_tensor([9788, 9788])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fec71ec72caac0b11a7b0d09d3524cd373a96b67 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_13/index_tensor([9788, 9788])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, high-angle macro lens, positioned almost directly overhead, remains static, capturing mesmerizing detail as a thick, syrupy stream of raw, honey-gold motor oil pours from an unseen source onto a black, matte, textured metal surface coated with a fine, dark gray industrial grime. The viscous, translucent liquid strikes contact not with a splash but with a heavy, deliberate weight, immediately beginning to carve intricate, river-like channels through the powdery soot. The golden oil flows slowly and relentlessly outward, its leading edge pushing aside the grime to reveal the clean, dark surface beneath, creating a mesmerizing, branching pattern of glowing, flowing tendrils. A single, focused light source from the side casts a warm, dramatic glow, creating brilliant, lustrous highlights that glide across the oil's undulating surface and cast deep, micro-shadows, emphasizing the textures of the displaced grime. The scene, dominated by high-contrast gold and deep charcoal colors, creates a hypnotic and beautiful abstract atmosphere that finds elegance in industrial process.", + "seed": 17705, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_13/index_9788_seed_17705_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08922481313347816, + "vlm_logprob": -0.08922481313347816, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.08922481313347816, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13273055851459503 + ], + "opposite_logprob_tokens": [ + -2.257730484008789 + ], + "logprob_reward": -0.13273055851459503, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.13273055851459503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2128678262233734, + 0.08607702702283859, + 0.021325182169675827, + 0.09273400902748108, + 0.12423605471849442, + 0.03633595257997513, + 0.07453537732362747, + 0.07568086683750153, + 0.23283281922340393, + 0.043374836444854736 + ], + "credit_map": [ + 0.024245457723736763, + 0.015872836112976074, + 0.033306483179330826, + 0.0184938907623291, + 0.017523523420095444, + 0.03900324925780296, + 0.025218788534402847, + 0.03920360282063484, + 0.010407978668808937, + 0.006051268428564072, + 0.022269142791628838, + 0.005751931108534336, + 0.006431572139263153, + 0.01480176206678152, + 0.00949486531317234, + 0.010868505574762821, + 0.0028745264280587435, + 0.001551391207613051, + 0.002508780686184764, + 0.001540679601021111, + 0.0033267473336309195, + 0.0027440686244517565, + 0.003908843267709017, + 0.002870145719498396, + 0.007997359149158001, + 0.002438221825286746, + 0.027459971606731415, + 0.003593480447307229, + 0.004856992978602648, + 0.03179093450307846, + 0.004882830660790205, + 0.0097142169252038, + 0.013353748247027397, + 0.00743822893127799, + 0.027943478897213936, + 0.010309305973351002, + 0.007022069301456213, + 0.02360188029706478, + 0.007018520962446928, + 0.02754882723093033, + 0.004295155871659517, + 0.0049457852728664875, + 0.006626427173614502, + 0.0032153879292309284, + 0.003709063632413745, + 0.005209307186305523, + 0.00329389120452106, + 0.005040932446718216, + 0.0111930463463068, + 0.005334915593266487, + 0.01644265465438366, + 0.004757892806082964, + 0.005669332575052977, + 0.011877343989908695, + 0.0061795529909431934, + 0.013080640695989132, + 0.005164686124771833, + 0.004950577858835459, + 0.008787558414041996, + 0.0026850029826164246, + 0.004889858420938253, + 0.022670937702059746, + 0.0039679985493421555, + 0.022564247250556946, + 0.01496788952499628, + 0.005796255078166723, + 0.020026082172989845, + 0.006761334836483002, + 0.006568066775798798, + 0.08592674136161804, + 0.00940968468785286, + 0.08337678015232086, + 0.006117386277765036, + 0.002042800188064575, + 0.012597317807376385, + 0.0023856114130467176, + 0.0028825514018535614, + 0.007118937559425831, + 0.0036552546080201864, + 0.006574977654963732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.13273055851459503 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06212099269032478 + ], + "opposite_logprob_tokens": [ + -2.8121209144592285 + ], + "logprob_reward": -0.06212099269032478, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06212099269032478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22867894172668457, + 0.08730442821979523, + 0.020180227234959602, + 0.13668681681156158, + 0.1786855012178421, + 0.03192906826734543, + 0.03806911036372185, + 0.13123124837875366, + 0.10463935136795044, + 0.042595311999320984 + ], + "credit_map": [ + 0.015424852259457111, + 0.01226024329662323, + 0.029204223304986954, + 0.014155727811157703, + 0.015482948161661625, + 0.05720581114292145, + 0.024876100942492485, + 0.06006902456283569, + 0.008446062915027142, + 0.006216257344931364, + 0.016002735123038292, + 0.006455582566559315, + 0.006091869901865721, + 0.01673668809235096, + 0.009087646380066872, + 0.018267590552568436, + 0.002637169323861599, + 0.001541833276860416, + 0.004044937901198864, + 0.0016269569750875235, + 0.002283053006976843, + 0.0028924087528139353, + 0.002458245027810335, + 0.0026956216897815466, + 0.014370167627930641, + 0.00539669394493103, + 0.04574527591466904, + 0.008577235974371433, + 0.006967080757021904, + 0.0333651639521122, + 0.009623278863728046, + 0.0126419048756361, + 0.02069650962948799, + 0.006797274574637413, + 0.02053908258676529, + 0.0066207218915224075, + 0.00631738780066371, + 0.04199732467532158, + 0.011343939229846, + 0.06437326222658157, + 0.004291664808988571, + 0.0023690026719123125, + 0.008263357914984226, + 0.0019606738351285458, + 0.001741405576467514, + 0.0055450815707445145, + 0.002568232361227274, + 0.005189651623368263, + 0.00597476726397872, + 0.002844949020072818, + 0.009424582123756409, + 0.001961679430678487, + 0.0020337619353085756, + 0.00604445394128561, + 0.002613392425701022, + 0.007171522825956345, + 0.01237366534769535, + 0.009534044191241264, + 0.018836883828043938, + 0.00673321820795536, + 0.007349293678998947, + 0.03309161216020584, + 0.01185743696987629, + 0.03145509213209152, + 0.008380457758903503, + 0.004906502552330494, + 0.01920773647725582, + 0.004927352070808411, + 0.0059134019538760185, + 0.027134884148836136, + 0.006108799483627081, + 0.028060223907232285, + 0.005500989966094494, + 0.0026438392233103514, + 0.009379450231790543, + 0.0031657658983021975, + 0.0033308183774352074, + 0.007686471566557884, + 0.004877300933003426, + 0.006010675337165594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.06212099269032478 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016173005104065 + ], + "opposite_logprob_tokens": [ + -1.7016172409057617 + ], + "logprob_reward": -0.2016173005104065, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016173005104065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4179792106151581, + 0.07726848125457764, + 0.03028227761387825, + 0.049237143248319626, + 0.08592859655618668, + 0.0618705116212368, + 0.028163932263851166, + 0.08283849060535431, + 0.10980041325092316, + 0.05663096532225609 + ], + "credit_map": [ + 0.02740505523979664, + 0.018851425498723984, + 0.043050169944763184, + 0.01861034706234932, + 0.02479459159076214, + 0.12241428345441818, + 0.03903336450457573, + 0.1238199919462204, + 0.00898814108222723, + 0.00472780829295516, + 0.018471458926796913, + 0.005282122641801834, + 0.006463401019573212, + 0.012129313312470913, + 0.00783164706081152, + 0.01337458472698927, + 0.004423442296683788, + 0.001970309531316161, + 0.005860557314008474, + 0.0017204125178977847, + 0.003567403182387352, + 0.004042492713779211, + 0.0034188362769782543, + 0.0052788229659199715, + 0.006045951973646879, + 0.0036681671626865864, + 0.01127542182803154, + 0.0049521904438734055, + 0.004494309425354004, + 0.005042451433837414, + 0.005835138261318207, + 0.007923511788249016, + 0.008351074531674385, + 0.006277068052440882, + 0.01476229727268219, + 0.006211295258253813, + 0.004562544170767069, + 0.01577341929078102, + 0.0075378031469881535, + 0.022453095763921738, + 0.008257618173956871, + 0.0047280630096793175, + 0.013451511040329933, + 0.0038772437255829573, + 0.003702936228364706, + 0.011271861381828785, + 0.004189528524875641, + 0.01239174883812666, + 0.004786474164575338, + 0.002597876824438572, + 0.0029413311276584864, + 0.0016826404025778174, + 0.003773446660488844, + 0.004834970459342003, + 0.003056430025026202, + 0.00449076434597373, + 0.00812484696507454, + 0.002995272632688284, + 0.013581149280071259, + 0.0034041935577988625, + 0.0036507390905171633, + 0.02300255373120308, + 0.005349552258849144, + 0.022730182856321335, + 0.00845844205468893, + 0.003917421214282513, + 0.01054440625011921, + 0.004886440932750702, + 0.005460554733872414, + 0.03439288213849068, + 0.008093688637018204, + 0.034046582877635956, + 0.0056019010953605175, + 0.004056976642459631, + 0.01161906123161316, + 0.0041294400580227375, + 0.005347054451704025, + 0.010560923255980015, + 0.006773118861019611, + 0.008542487397789955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.2016173005104065 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302748799324036 + ], + "opposite_logprob_tokens": [ + -3.7733027935028076 + ], + "logprob_reward": -0.023302748799324036, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302748799324036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2471739947795868, + 0.04824429750442505, + 0.09533307701349258, + 0.08985819667577744, + 0.21184414625167847, + 0.07253605127334595, + 0.021030407398939133, + 0.1325560063123703, + 0.057627204805612564, + 0.02379656210541725 + ], + "credit_map": [ + 0.014330940321087837, + 0.008381973952054977, + 0.02594982087612152, + 0.008603568188846111, + 0.017274286597967148, + 0.07351803779602051, + 0.01960550993680954, + 0.07950989156961441, + 0.006085322704166174, + 0.0037173498421907425, + 0.008000155910849571, + 0.004392971284687519, + 0.00335005228407681, + 0.008508109487593174, + 0.005151746794581413, + 0.009038594551384449, + 0.015908751636743546, + 0.006189871113747358, + 0.01625119522213936, + 0.004216529428958893, + 0.017735695466399193, + 0.013542083092033863, + 0.008795211091637611, + 0.012693751603364944, + 0.007629708386957645, + 0.0025930197443813086, + 0.03202712908387184, + 0.004215783905237913, + 0.003998591564595699, + 0.02822665125131607, + 0.005189851392060518, + 0.005977463908493519, + 0.029519017785787582, + 0.008617567829787731, + 0.028285572305321693, + 0.007056325674057007, + 0.00799825694411993, + 0.04716737940907478, + 0.011257466860115528, + 0.07194257527589798, + 0.007546878419816494, + 0.006060849409550428, + 0.012941845692694187, + 0.004154778551310301, + 0.004613511264324188, + 0.014306356199085712, + 0.005092698149383068, + 0.01781914010643959, + 0.00338124786503613, + 0.00120002799667418, + 0.004516908433288336, + 0.0011721528135240078, + 0.0010880621848627925, + 0.00360189750790596, + 0.0015686005353927612, + 0.004501511808484793, + 0.01004659291356802, + 0.004376985132694244, + 0.016871480271220207, + 0.0036450910847634077, + 0.003839914221316576, + 0.043691687285900116, + 0.00546098779886961, + 0.044623278081417084, + 0.004896554164588451, + 0.002360433805733919, + 0.010109408758580685, + 0.002881178865209222, + 0.0028705489821732044, + 0.014284106902778149, + 0.0038533927872776985, + 0.0163715872913599, + 0.002474466571584344, + 0.0011064168065786362, + 0.00470344303175807, + 0.0011448387522250414, + 0.0012434294912964106, + 0.005598063580691814, + 0.002543790964409709, + 0.004982114303857088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.023302748799324036 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635246515274048 + ], + "opposite_logprob_tokens": [ + -3.6513524055480957 + ], + "logprob_reward": -0.02635246515274048, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635246515274048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22452357411384583, + 0.05003104731440544, + 0.025508154183626175, + 0.3200003504753113, + 0.13715019822120667, + 0.03888695687055588, + 0.03867406025528908, + 0.026364363729953766, + 0.09984873235225677, + 0.03901255503296852 + ], + "credit_map": [ + 0.01641661860048771, + 0.012730041518807411, + 0.027473030611872673, + 0.014916541986167431, + 0.01829424314200878, + 0.0536075085401535, + 0.026714138686656952, + 0.05437144637107849, + 0.006334802135825157, + 0.0025210431776940823, + 0.011271326802670956, + 0.003417622298002243, + 0.0038971705362200737, + 0.007730068638920784, + 0.005923926830291748, + 0.008935085497796535, + 0.0028725352603942156, + 0.001850827015005052, + 0.0052736192010343075, + 0.0015634879237040877, + 0.003888232633471489, + 0.00366563955321908, + 0.002986924722790718, + 0.0034068881068378687, + 0.031715892255306244, + 0.011766493320465088, + 0.10910096019506454, + 0.019506651908159256, + 0.015489086508750916, + 0.08069691807031631, + 0.023848146200180054, + 0.02787620574235916, + 0.01448857318609953, + 0.005666070152074099, + 0.017442159354686737, + 0.006036614999175072, + 0.004598574247211218, + 0.031530145555734634, + 0.00923954974859953, + 0.0481485016644001, + 0.004901302047073841, + 0.0035010199062526226, + 0.008143644779920578, + 0.003714534919708967, + 0.003931759390980005, + 0.005936608184129, + 0.003644349751994014, + 0.005113737192004919, + 0.004073098301887512, + 0.0022888563107699156, + 0.008858676068484783, + 0.0024803474079817533, + 0.00258290721103549, + 0.006531160324811935, + 0.004249592311680317, + 0.00760942418128252, + 0.004536750260740519, + 0.0015485429903492332, + 0.006702656392008066, + 0.0017999629490077496, + 0.0015117530710995197, + 0.004331849981099367, + 0.0023908335715532303, + 0.003542016027495265, + 0.00718081695958972, + 0.003422083565965295, + 0.01017134077847004, + 0.004381971433758736, + 0.004060492385178804, + 0.03259408846497536, + 0.006030027288943529, + 0.032007910311222076, + 0.003938550595194101, + 0.0031673472840338945, + 0.008921959437429905, + 0.0032553132623434067, + 0.0037957706954330206, + 0.005780043546110392, + 0.0044860453344881535, + 0.005667523015290499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.02635246515274048 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, high-angle macro lens, positioned almost directly overhead, remains static, capturing mesmerizing detail as a thick, syrupy stream of raw, honey-gold motor oil pours from an unseen source onto a black, matte, textured metal surface coated with a fine, dark gray industrial grime. The viscous, translucent liquid strikes contact not with a splash but with a heavy, deliberate weight, immediately beginning to carve intricate, river-like channels through the powdery soot. The golden oil flows slowly and relentlessly outward, its leading edge pushing aside the grime to reveal the clean, dark surface beneath, creating a mesmerizing, branching pattern of glowing, flowing tendrils. A single, focused light source from the side casts a warm, dramatic glow, creating brilliant, lustrous highlights that glide across the oil's undulating surface and cast deep, micro-shadows, emphasizing the textures of the displaced grime. The scene, dominated by high-contrast gold and deep charcoal colors, creates a hypnotic and beautiful abstract atmosphere that finds elegance in industrial process.", + "seed": 17706, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_13/index_9788_seed_17706_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07628434039652347, + "vlm_logprob": -0.07628434039652347, + "vlm_yes_no_margin": 2.9999999523162844, + "vlm_token_credit": -0.07628434039652347, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1676000952720642 + ], + "opposite_logprob_tokens": [ + -2.042600154876709 + ], + "logprob_reward": -0.1676000952720642, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1676000952720642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025712408125400543, + 0.0071800826117396355, + 0.022464150562882423, + 0.7670497298240662, + 0.01812325417995453, + 0.011195255443453789, + 0.016520382836461067, + 0.04805050790309906, + 0.07874732464551926, + 0.004956989083439112 + ], + "credit_map": [ + 0.0018691635923460126, + 0.0015675764298066497, + 0.0014023848343640566, + 0.001401186571456492, + 0.0010887127136811614, + 0.008320177905261517, + 0.001212799223139882, + 0.008850406855344772, + 0.000636293669231236, + 0.0003607604594435543, + 0.000620584876742214, + 0.0006492242100648582, + 0.000590914802160114, + 0.0018514897674322128, + 0.0005766437388956547, + 0.0018941700691357255, + 0.000897142686881125, + 0.0007950346334837377, + 0.0005739374901168048, + 0.0008133965893648565, + 0.0009150381083600223, + 0.008566115982830524, + 0.001005347352474928, + 0.008898136205971241, + 0.013399642892181873, + 0.02254568226635456, + 0.005707163829356432, + 0.009013506583869457, + 0.01671827770769596, + 0.3329300582408905, + 0.020265189930796623, + 0.3464701175689697, + 0.0023225913755595684, + 0.0014518583193421364, + 0.0022965334355831146, + 0.0012547664809972048, + 0.0026107279118150473, + 0.0028591935988515615, + 0.002859897445887327, + 0.0024676844477653503, + 0.001097955391742289, + 0.0012936678249388933, + 0.001105238450691104, + 0.0006678917561657727, + 0.0017419472569599748, + 0.0019034744473174214, + 0.001867034356109798, + 0.0015180449699983, + 0.0008557499386370182, + 0.000674462120514363, + 0.0005237903096713126, + 0.0005604902398772538, + 0.0012965637724846601, + 0.005728058982640505, + 0.001017836737446487, + 0.0058634295128285885, + 0.0015273885801434517, + 0.0013495837338268757, + 0.001146669965237379, + 0.0010408139787614346, + 0.002798984292894602, + 0.018464667722582817, + 0.002990437438711524, + 0.01873195730149746, + 0.0069129737094044685, + 0.003794347867369652, + 0.0031909600365906954, + 0.0029637294355779886, + 0.009298653341829777, + 0.01953701674938202, + 0.01254916563630104, + 0.02050047367811203, + 0.0005504441796801984, + 0.0005773447919636965, + 0.0004423634964041412, + 0.0002593692042864859, + 0.0009438342531211674, + 0.0007369577651843429, + 0.0006965329521335661, + 0.0007501419167965651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.1676000952720642 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026411911472678185 + ], + "opposite_logprob_tokens": [ + -3.651412010192871 + ], + "logprob_reward": -0.026411911472678185, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026411911472678185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01582040637731552, + 0.020346270874142647, + 0.04038511961698532, + 0.6044264435768127, + 0.07671673595905304, + 0.019398782402276993, + 0.07678616046905518, + 0.05812959745526314, + 0.08153645694255829, + 0.006454077549278736 + ], + "credit_map": [ + 0.0025736403185874224, + 0.0015849510673433542, + 0.0018878476694226265, + 0.002338357036933303, + 0.0019917578902095556, + 0.0019393416587263346, + 0.0013689558254554868, + 0.002135553630068898, + 0.0016024983488023281, + 0.001517545897513628, + 0.0016099201748147607, + 0.001076875370927155, + 0.001328128157183528, + 0.005732893943786621, + 0.0014263509074226022, + 0.006052055396139622, + 0.0017488289158791304, + 0.0019319382263347507, + 0.0011588274501264095, + 0.001316767418757081, + 0.001652766834013164, + 0.015236126258969307, + 0.001746089314110577, + 0.015593771822750568, + 0.011340572498738766, + 0.017792796716094017, + 0.004578371066600084, + 0.006967081222683191, + 0.013607230968773365, + 0.26166459918022156, + 0.016187580302357674, + 0.27228817343711853, + 0.0032229898497462273, + 0.003926384262740612, + 0.004723131190985441, + 0.0021532834507524967, + 0.005873033776879311, + 0.026083653792738914, + 0.005698232911527157, + 0.025036020204424858, + 0.0021635382436215878, + 0.0014037598157301545, + 0.001437654602341354, + 0.0008580491412431002, + 0.0033955315593630075, + 0.003263286082074046, + 0.0025277608074247837, + 0.004349199123680592, + 0.0017826466355472803, + 0.0022649543825536966, + 0.0011684319470077753, + 0.0008929857867769897, + 0.00315470015630126, + 0.03180348128080368, + 0.0033066892065107822, + 0.03241226449608803, + 0.0016925311647355556, + 0.00182330550160259, + 0.0010444032959640026, + 0.001109310076572001, + 0.0037933625280857086, + 0.022086359560489655, + 0.003776306752115488, + 0.022804012522101402, + 0.011092555709183216, + 0.005786731839179993, + 0.005312921945005655, + 0.00507431710138917, + 0.014118017628788948, + 0.013220195658504963, + 0.015054276213049889, + 0.011877433396875858, + 0.0006848560879006982, + 0.0006070021190680563, + 0.0005407893331721425, + 0.00042905748705379665, + 0.0010815508430823684, + 0.0008818025817163289, + 0.0013424577191472054, + 0.0008865605341270566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.026411911472678185 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285853505134583 + ], + "opposite_logprob_tokens": [ + -2.0178585052490234 + ], + "logprob_reward": -0.14285853505134583, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14285853505134583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08205326646566391, + 0.045682307332754135, + 0.06732945889234543, + 0.4725855886936188, + 0.12358494848012924, + 0.05927528813481331, + 0.06430111825466156, + 0.022210722789168358, + 0.04352725297212601, + 0.01945003680884838 + ], + "credit_map": [ + 0.006567422766238451, + 0.0072012669406831264, + 0.005950566381216049, + 0.007846102118492126, + 0.006358497776091099, + 0.02075744792819023, + 0.005808927118778229, + 0.021563027054071426, + 0.0060987393371760845, + 0.0040260073728859425, + 0.004983144346624613, + 0.004047025926411152, + 0.007153509650379419, + 0.006501196883618832, + 0.00599337974563241, + 0.006879298482090235, + 0.007344124373048544, + 0.005595772061496973, + 0.003833112306892872, + 0.0044530280865728855, + 0.005092258099466562, + 0.01713431626558304, + 0.005388759542256594, + 0.018488090485334396, + 0.008468912914395332, + 0.01422332227230072, + 0.004730530083179474, + 0.006448369938880205, + 0.010437315329909325, + 0.2032397836446762, + 0.013271255418658257, + 0.2117660790681839, + 0.006650408264249563, + 0.005060482770204544, + 0.003844105638563633, + 0.002272039419040084, + 0.006696862634271383, + 0.04347153380513191, + 0.010350874625146389, + 0.04523863643407822, + 0.004900040104985237, + 0.004742723423987627, + 0.0036046775057911873, + 0.0026365071535110474, + 0.013165310956537724, + 0.010755016468465328, + 0.009253508411347866, + 0.010217499919235706, + 0.0019268941832706332, + 0.0020348585676401854, + 0.0013525666436180472, + 0.0013906543608754873, + 0.0026604768354445696, + 0.025929225608706474, + 0.0024829143658280373, + 0.026523524895310402, + 0.001810819492675364, + 0.0018329162849113345, + 0.0014649401418864727, + 0.0009154640138149261, + 0.0024457916151732206, + 0.005328567232936621, + 0.00362600339576602, + 0.004786220379173756, + 0.005934989545494318, + 0.0036061061546206474, + 0.0032644036691635847, + 0.0027546731289476156, + 0.008338806219398975, + 0.007028036750853062, + 0.007586061023175716, + 0.005014174152165651, + 0.002435998758301139, + 0.0019049992552027106, + 0.0010087103582918644, + 0.0014028921723365784, + 0.0034749535843729973, + 0.002711878390982747, + 0.0034413281828165054, + 0.0030692745931446552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.14285853505134583 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018202990293502808 + ], + "opposite_logprob_tokens": [ + -4.018202781677246 + ], + "logprob_reward": -0.018202990293502808, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018202990293502808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04298238828778267, + 0.0671529620885849, + 0.3311721384525299, + 0.05614747107028961, + 0.08594995737075806, + 0.030984720215201378, + 0.10560598224401474, + 0.08598574250936508, + 0.10949723422527313, + 0.08452139794826508 + ], + "credit_map": [ + 0.003519403515383601, + 0.003334284294396639, + 0.0028936995659023523, + 0.0031736381351947784, + 0.002380908699706197, + 0.012668553739786148, + 0.0021993888076394796, + 0.012812514789402485, + 0.00397584680467844, + 0.0036357026547193527, + 0.004946164321154356, + 0.0030867080204188824, + 0.004831074737012386, + 0.019940968602895737, + 0.005043429788202047, + 0.021693067625164986, + 0.00962079782038927, + 0.011529451236128807, + 0.006699989549815655, + 0.00761222280561924, + 0.008541072718799114, + 0.1335221827030182, + 0.012358603067696095, + 0.14128784835338593, + 0.0026523314882069826, + 0.0026896370109170675, + 0.0038913763128221035, + 0.0022553084418177605, + 0.003801169339567423, + 0.018120912835001945, + 0.004176889546215534, + 0.018559850752353668, + 0.00880064070224762, + 0.0059215836226940155, + 0.011907919310033321, + 0.005480757914483547, + 0.014495620504021645, + 0.014613479375839233, + 0.010340495966374874, + 0.014389470219612122, + 0.004112115129828453, + 0.003233363851904869, + 0.002905394183471799, + 0.001667861477471888, + 0.0038256917614489794, + 0.005551558453589678, + 0.0034358249977231026, + 0.006252912804484367, + 0.002816810505464673, + 0.0032071347814053297, + 0.0013228723546490073, + 0.0014092089841142297, + 0.0023628338240087032, + 0.04475357383489609, + 0.004047276917845011, + 0.045686282217502594, + 0.0032005542889237404, + 0.0027892806101590395, + 0.0016213135095313191, + 0.001647497178055346, + 0.006244951859116554, + 0.0318799652159214, + 0.005759189836680889, + 0.03284299746155739, + 0.017856121063232422, + 0.010373908095061779, + 0.007426687050610781, + 0.00723639689385891, + 0.015200628899037838, + 0.020574243739247322, + 0.016013098880648613, + 0.014816160313785076, + 0.008996170945465565, + 0.0061644865199923515, + 0.007381747011095285, + 0.00827848445624113, + 0.012944146059453487, + 0.010615477338433266, + 0.018211914226412773, + 0.011928979307413101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.018202990293502808 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348169893026352 + ], + "opposite_logprob_tokens": [ + -3.651348114013672 + ], + "logprob_reward": -0.026348169893026352, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348169893026352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05161121487617493, + 0.029556667432188988, + 0.10177790373563766, + 0.3695034682750702, + 0.026122746989130974, + 0.036442942917346954, + 0.03426812216639519, + 0.16078098118305206, + 0.17109748721122742, + 0.018838535994291306 + ], + "credit_map": [ + 0.0038611688651144505, + 0.0036196736618876457, + 0.003019189229235053, + 0.004415425006300211, + 0.0029856509063392878, + 0.015055948868393898, + 0.003119206754490733, + 0.015534942969679832, + 0.002721495693549514, + 0.002260613953694701, + 0.0030666235834360123, + 0.002232542261481285, + 0.003479940816760063, + 0.006594549398869276, + 0.00263545336201787, + 0.0065654427744448185, + 0.002535901265218854, + 0.002984066028147936, + 0.001745868707075715, + 0.0019179024966433644, + 0.0021044996101409197, + 0.04246474802494049, + 0.003508320776745677, + 0.044516582041978836, + 0.007282195147126913, + 0.011436577886343002, + 0.00460490956902504, + 0.004840813111513853, + 0.008911259472370148, + 0.15803922712802887, + 0.010639606043696404, + 0.1637488156557083, + 0.0041884989477694035, + 0.0020326448138803244, + 0.002926464658230543, + 0.001648851903155446, + 0.0035361831542104483, + 0.0037179840728640556, + 0.00358329969458282, + 0.00448881508782506, + 0.0028620422817766666, + 0.0026620819699019194, + 0.002211780520156026, + 0.0013724557356908917, + 0.008475501090288162, + 0.006021537818014622, + 0.005261635407805443, + 0.007575905416160822, + 0.0029333983547985554, + 0.0018161619082093239, + 0.001388463075272739, + 0.0010394181590527296, + 0.004150391556322575, + 0.009905554354190826, + 0.002511057071387768, + 0.01052367314696312, + 0.006037060171365738, + 0.0047438922338187695, + 0.003594185458496213, + 0.0034718429669737816, + 0.010119178332388401, + 0.06045936793088913, + 0.010621779598295689, + 0.061733659356832504, + 0.019608434289693832, + 0.012698939070105553, + 0.01180982030928135, + 0.012088173069059849, + 0.026759110391139984, + 0.03777570649981499, + 0.02037353254854679, + 0.029983747750520706, + 0.0018241453217342496, + 0.001802021055482328, + 0.0013021143386140466, + 0.0009509547962807119, + 0.0030416285153478384, + 0.003603684715926647, + 0.003404319519177079, + 0.002909665461629629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.026348169893026352 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, high-angle macro lens, positioned almost directly overhead, remains static, capturing mesmerizing detail as a thick, syrupy stream of raw, honey-gold motor oil pours from an unseen source onto a black, matte, textured metal surface coated with a fine, dark gray industrial grime. The viscous, translucent liquid strikes contact not with a splash but with a heavy, deliberate weight, immediately beginning to carve intricate, river-like channels through the powdery soot. The golden oil flows slowly and relentlessly outward, its leading edge pushing aside the grime to reveal the clean, dark surface beneath, creating a mesmerizing, branching pattern of glowing, flowing tendrils. A single, focused light source from the side casts a warm, dramatic glow, creating brilliant, lustrous highlights that glide across the oil's undulating surface and cast deep, micro-shadows, emphasizing the textures of the displaced grime. The scene, dominated by high-contrast gold and deep charcoal colors, creates a hypnotic and beautiful abstract atmosphere that finds elegance in industrial process.", + "seed": 17707, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_13/index_9788_seed_17707_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07243389599025249, + "vlm_logprob": -0.07243389599025249, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.07243389599025249, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10168110579252243 + ], + "opposite_logprob_tokens": [ + -2.6016809940338135 + ], + "logprob_reward": -0.10168110579252243, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.10168110579252243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25644105672836304, + 0.08630608767271042, + 0.12940654158592224, + 0.05077227205038071, + 0.08610262721776962, + 0.12161935120820999, + 0.058374516665935516, + 0.06987593322992325, + 0.07770726084709167, + 0.06339435279369354 + ], + "credit_map": [ + 0.025065237656235695, + 0.016816509887576103, + 0.05656980723142624, + 0.02653353102505207, + 0.025284908711910248, + 0.0436696819961071, + 0.027997737750411034, + 0.034503646194934845, + 0.009920942597091198, + 0.005686354823410511, + 0.015123044140636921, + 0.010371212847530842, + 0.008557462133467197, + 0.014530308544635773, + 0.009753022342920303, + 0.012363731861114502, + 0.014270138926804066, + 0.011485730297863483, + 0.018069405108690262, + 0.013374281115829945, + 0.012709454633295536, + 0.01992090791463852, + 0.014478878118097782, + 0.025097733363509178, + 0.00560296606272459, + 0.006909323390573263, + 0.007483964785933495, + 0.0052153305150568485, + 0.004080347251147032, + 0.008064127527177334, + 0.0072843111120164394, + 0.006131898611783981, + 0.007021148223429918, + 0.010179854929447174, + 0.013359369710087776, + 0.011788385920226574, + 0.009458482265472412, + 0.01222038734704256, + 0.010656692087650299, + 0.01141830999404192, + 0.00802820548415184, + 0.01188149768859148, + 0.019966233521699905, + 0.0146632706746459, + 0.011547974310815334, + 0.016679294407367706, + 0.024368152022361755, + 0.014484718441963196, + 0.007057189010083675, + 0.006665946915745735, + 0.010737964883446693, + 0.006030498072504997, + 0.004869693424552679, + 0.008829236030578613, + 0.006066999863833189, + 0.008116986602544785, + 0.005203274078667164, + 0.006526396609842777, + 0.012148017063736916, + 0.008813954889774323, + 0.008368317037820816, + 0.011028600856661797, + 0.008902711793780327, + 0.008884665556252003, + 0.00417799549177289, + 0.004877476487308741, + 0.012200934812426567, + 0.012219667434692383, + 0.007775827776640654, + 0.012193876318633556, + 0.011594933457672596, + 0.012666548602283001, + 0.0048343162052333355, + 0.005759879015386105, + 0.009729775600135326, + 0.007768383715301752, + 0.008025841787457466, + 0.008540057577192783, + 0.011372344568371773, + 0.007363755255937576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.10168110579252243 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03816878795623779 + ], + "opposite_logprob_tokens": [ + -3.2881689071655273 + ], + "logprob_reward": -0.03816878795623779, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03816878795623779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13109691441059113, + 0.11480003595352173, + 0.14737346768379211, + 0.0827692523598671, + 0.09426965564489365, + 0.10439255833625793, + 0.09419311583042145, + 0.08404873311519623, + 0.06009034067392349, + 0.08696595579385757 + ], + "credit_map": [ + 0.01671895571053028, + 0.01206847932189703, + 0.023228630423545837, + 0.011586826294660568, + 0.011768492870032787, + 0.02315537817776203, + 0.011417249217629433, + 0.021152889356017113, + 0.009980732575058937, + 0.0073098898865282536, + 0.025378260761499405, + 0.01055386196821928, + 0.009960618801414967, + 0.020189320668578148, + 0.015439831651747227, + 0.015987517312169075, + 0.015559528023004532, + 0.011413558386266232, + 0.02495628595352173, + 0.015361906029284, + 0.011920944787561893, + 0.02279789187014103, + 0.02310149557888508, + 0.02226184494793415, + 0.006865864619612694, + 0.005717076361179352, + 0.011705980636179447, + 0.009773364290595055, + 0.006901796441525221, + 0.011895226314663887, + 0.019828028976917267, + 0.010081914253532887, + 0.008291526697576046, + 0.010693051852285862, + 0.019533328711986542, + 0.014067712239921093, + 0.005415197461843491, + 0.01551633421331644, + 0.008192065171897411, + 0.012560436502099037, + 0.008338702842593193, + 0.008931729011237621, + 0.012582716532051563, + 0.013933446258306503, + 0.008923604153096676, + 0.012778369709849358, + 0.023912781849503517, + 0.01499120146036148, + 0.006687107030302286, + 0.009928539395332336, + 0.0080189760774374, + 0.013161325827240944, + 0.014989124611020088, + 0.014994933269917965, + 0.014650154858827591, + 0.011762953363358974, + 0.0073417979292571545, + 0.008688089437782764, + 0.013797362335026264, + 0.013546633534133434, + 0.00877376925200224, + 0.011527993716299534, + 0.00968080759048462, + 0.010692278854548931, + 0.004531558603048325, + 0.004011636134237051, + 0.009351148270070553, + 0.006974761839956045, + 0.004725821781903505, + 0.009326555766165257, + 0.012311064638197422, + 0.00885778944939375, + 0.006245527416467667, + 0.008387708105146885, + 0.010868624784052372, + 0.012655263766646385, + 0.013404395431280136, + 0.012435942888259888, + 0.013819608837366104, + 0.00914888083934784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.03816878795623779 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17996154725551605 + ], + "opposite_logprob_tokens": [ + -1.8049615621566772 + ], + "logprob_reward": -0.17996154725551605, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17996154725551605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14077606797218323, + 0.11727001518011093, + 0.09179653972387314, + 0.10693693906068802, + 0.07542050629854202, + 0.057132914662361145, + 0.11653640121221542, + 0.15286193788051605, + 0.04674646630883217, + 0.09452225267887115 + ], + "credit_map": [ + 0.013743334449827671, + 0.015293796546757221, + 0.018253644928336143, + 0.015561445616185665, + 0.019440563395619392, + 0.019839882850646973, + 0.018511397764086723, + 0.02013201080262661, + 0.01122133620083332, + 0.007595536299049854, + 0.024591125547885895, + 0.01142383087426424, + 0.010694803670048714, + 0.01979045569896698, + 0.01522764191031456, + 0.016725270077586174, + 0.008664941415190697, + 0.007677173707634211, + 0.017092330381274223, + 0.009119964204728603, + 0.008072531782090664, + 0.015049785375595093, + 0.010130329988896847, + 0.015989484265446663, + 0.009295787662267685, + 0.011100616306066513, + 0.01576020009815693, + 0.015567894093692303, + 0.009755892679095268, + 0.01430286094546318, + 0.016429603099822998, + 0.014724082313477993, + 0.006361021660268307, + 0.009020990692079067, + 0.015319406054913998, + 0.01107922662049532, + 0.004766164347529411, + 0.01182568445801735, + 0.007155867759138346, + 0.009892141446471214, + 0.0050889537669718266, + 0.005356627516448498, + 0.007342486642301083, + 0.008558518253266811, + 0.005057892296463251, + 0.007180209271609783, + 0.010686598718166351, + 0.007861629128456116, + 0.009872449561953545, + 0.012222586199641228, + 0.009640493430197239, + 0.015422784723341465, + 0.01934133470058441, + 0.01771014928817749, + 0.018797585740685463, + 0.013529014773666859, + 0.014021980576217175, + 0.01765965111553669, + 0.022057661786675453, + 0.02522817626595497, + 0.0156876090914011, + 0.020672528073191643, + 0.020061608403921127, + 0.01747271418571472, + 0.0036600737366825342, + 0.0035841059871017933, + 0.007283471524715424, + 0.005785942077636719, + 0.004882595967501402, + 0.008121287450194359, + 0.006271181628108025, + 0.007157810032367706, + 0.007259988225996494, + 0.009282873943448067, + 0.012157035991549492, + 0.015134353190660477, + 0.01161605678498745, + 0.013078923337161541, + 0.013184704817831516, + 0.012808313593268394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.17996154725551605 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647376671433449 + ], + "opposite_logprob_tokens": [ + -4.758647441864014 + ], + "logprob_reward": -0.008647376671433449, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647376671433449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08093193173408508, + 0.08917949348688126, + 0.2799864709377289, + 0.1157832145690918, + 0.04242400825023651, + 0.14041857421398163, + 0.03630317002534866, + 0.14333751797676086, + 0.05328181013464928, + 0.018353814259171486 + ], + "credit_map": [ + 0.00921364314854145, + 0.0074373274110257626, + 0.01698385551571846, + 0.0079781049862504, + 0.007372476626187563, + 0.013549010269343853, + 0.006999833509325981, + 0.011397669091820717, + 0.007791428826749325, + 0.006104708649218082, + 0.01822243072092533, + 0.00973175186663866, + 0.007812775671482086, + 0.017092589288949966, + 0.009259022772312164, + 0.013164784759283066, + 0.03144359216094017, + 0.02029377408325672, + 0.05678803473711014, + 0.027006175369024277, + 0.022121770307421684, + 0.043648555874824524, + 0.033866338431835175, + 0.0448182187974453, + 0.010912316851317883, + 0.00729322200641036, + 0.02145400084555149, + 0.01157935056835413, + 0.008091229014098644, + 0.018138129264116287, + 0.022896112874150276, + 0.015418848022818565, + 0.0033170608803629875, + 0.004984947852790356, + 0.008231363259255886, + 0.005688453558832407, + 0.0030294766183942556, + 0.006710592191666365, + 0.0052520413883030415, + 0.005210073199123144, + 0.009354549460113049, + 0.008916476741433144, + 0.017021186649799347, + 0.01959924027323723, + 0.012005538679659367, + 0.016985779628157616, + 0.037695176899433136, + 0.018840616568922997, + 0.0035254661925137043, + 0.004307140596210957, + 0.005129951983690262, + 0.004734880290925503, + 0.0028082835488021374, + 0.005625752732157707, + 0.004578913561999798, + 0.00559277692809701, + 0.013947540894150734, + 0.011873641051352024, + 0.02226589247584343, + 0.018843816593289375, + 0.014209500513970852, + 0.017753303050994873, + 0.024647684767842293, + 0.019796133041381836, + 0.003630635794252157, + 0.003395721083506942, + 0.009435822255909443, + 0.007032788824290037, + 0.004712218418717384, + 0.009607069194316864, + 0.006881061941385269, + 0.008586491458117962, + 0.0014233705587685108, + 0.0015556724974885583, + 0.0024264040403068066, + 0.002134867710992694, + 0.0026122210547327995, + 0.002786133671179414, + 0.002583365421742201, + 0.0028317796532064676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 5 + ], + "selected_reward": -0.008647376671433449 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371066227555275 + ], + "opposite_logprob_tokens": [ + -3.4087107181549072 + ], + "logprob_reward": -0.03371066227555275, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371066227555275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14514313638210297, + 0.08828909695148468, + 0.11685458570718765, + 0.08249267935752869, + 0.059580832719802856, + 0.21086858212947845, + 0.12005504965782166, + 0.07417292892932892, + 0.06022021174430847, + 0.04232291877269745 + ], + "credit_map": [ + 0.01408267579972744, + 0.012528694234788418, + 0.021765900775790215, + 0.017865313217043877, + 0.015390826389193535, + 0.024786924943327904, + 0.014272606931626797, + 0.024450188502669334, + 0.005893170367926359, + 0.006594143342226744, + 0.012454602867364883, + 0.008982045575976372, + 0.013689639046788216, + 0.012244517914950848, + 0.0164265725761652, + 0.012004406191408634, + 0.012302449904382229, + 0.011970431543886662, + 0.01981719583272934, + 0.014654685743153095, + 0.009994341060519218, + 0.016606759279966354, + 0.01417060848325491, + 0.017338110134005547, + 0.005819616839289665, + 0.008091737516224384, + 0.011295006610453129, + 0.012076995335519314, + 0.009030750952661037, + 0.010771919041872025, + 0.01563699170947075, + 0.009769661352038383, + 0.005215917248278856, + 0.005486002191901207, + 0.010457434691488743, + 0.008179138414561749, + 0.00500738900154829, + 0.009255398996174335, + 0.008216848596930504, + 0.007762701716274023, + 0.013811101205646992, + 0.013878772966563702, + 0.025819582864642143, + 0.029129493981599808, + 0.018991781398653984, + 0.02500964142382145, + 0.05863640829920769, + 0.025591790676116943, + 0.006869915407150984, + 0.014260201714932919, + 0.012455291114747524, + 0.018122073262929916, + 0.01690969057381153, + 0.018485229462385178, + 0.01693093590438366, + 0.016021713614463806, + 0.008112743496894836, + 0.008574447594583035, + 0.010979421436786652, + 0.011867542751133442, + 0.007531359791755676, + 0.008596980012953281, + 0.0088318707421422, + 0.009678563103079796, + 0.0052542430348694324, + 0.00626452686265111, + 0.009112460538744926, + 0.007598554715514183, + 0.0060791620053350925, + 0.008500194177031517, + 0.008817782625555992, + 0.00859328918159008, + 0.0031585877295583487, + 0.004255236592143774, + 0.00527196703478694, + 0.006308069918304682, + 0.005027055740356445, + 0.006481315474957228, + 0.005877218209207058, + 0.005943470634520054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.03371066227555275 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, high-angle macro lens, positioned almost directly overhead, remains static, capturing mesmerizing detail as a thick, syrupy stream of raw, honey-gold motor oil pours from an unseen source onto a black, matte, textured metal surface coated with a fine, dark gray industrial grime. The viscous, translucent liquid strikes contact not with a splash but with a heavy, deliberate weight, immediately beginning to carve intricate, river-like channels through the powdery soot. The golden oil flows slowly and relentlessly outward, its leading edge pushing aside the grime to reveal the clean, dark surface beneath, creating a mesmerizing, branching pattern of glowing, flowing tendrils. A single, focused light source from the side casts a warm, dramatic glow, creating brilliant, lustrous highlights that glide across the oil's undulating surface and cast deep, micro-shadows, emphasizing the textures of the displaced grime. The scene, dominated by high-contrast gold and deep charcoal colors, creates a hypnotic and beautiful abstract atmosphere that finds elegance in industrial process.", + "seed": 17708, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_13/index_9788_seed_17708_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0853066073730588, + "vlm_logprob": -0.0853066073730588, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.0853066073730588, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11794472485780716 + ], + "opposite_logprob_tokens": [ + -2.4929447174072266 + ], + "logprob_reward": -0.11794472485780716, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.11794472485780716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36119523644447327, + 0.1229911595582962, + 0.16877590119838715, + 0.04995041713118553, + 0.04726748540997505, + 0.03313439339399338, + 0.022678328678011894, + 0.120221808552742, + 0.04278912767767906, + 0.03099609538912773 + ], + "credit_map": [ + 0.032810430973768234, + 0.025323817506432533, + 0.046188656240701675, + 0.036557797342538834, + 0.04854869470000267, + 0.05792246758937836, + 0.05732237175107002, + 0.05652102082967758, + 0.011362401768565178, + 0.009662190452218056, + 0.015659987926483154, + 0.011091984808444977, + 0.01740742288529873, + 0.022949080914258957, + 0.020982874557375908, + 0.013875230215489864, + 0.022025225684046745, + 0.00964327435940504, + 0.01784081943333149, + 0.018674788996577263, + 0.020346282050013542, + 0.03402302414178848, + 0.01740335300564766, + 0.02881915681064129, + 0.00787334144115448, + 0.0030848586466163397, + 0.005711423233151436, + 0.004844139330089092, + 0.0052512818947434425, + 0.008969863876700401, + 0.007117375731468201, + 0.00709814066067338, + 0.006045433226972818, + 0.003561011515557766, + 0.00439433054998517, + 0.004279003478586674, + 0.006510268431156874, + 0.008280591107904911, + 0.006784666795283556, + 0.00741218589246273, + 0.005102607887238264, + 0.0026618510019034147, + 0.003343694843351841, + 0.0021707562264055014, + 0.004844150505959988, + 0.006157232914119959, + 0.0035003700759261847, + 0.005353735294193029, + 0.002954750321805477, + 0.0016173989279195666, + 0.0029038169886916876, + 0.002500553848221898, + 0.0029714268166571856, + 0.0034462527837604284, + 0.0032546534202992916, + 0.003029480343684554, + 0.006558096967637539, + 0.0055564288049936295, + 0.007281350903213024, + 0.005855634342879057, + 0.023037677630782127, + 0.02167901210486889, + 0.03235023096203804, + 0.017903393134474754, + 0.005548973102122545, + 0.003229148918762803, + 0.004466430749744177, + 0.004284215159714222, + 0.005599476397037506, + 0.007410550490021706, + 0.005231487564742565, + 0.0070188529789447784, + 0.002568391617387533, + 0.002806085394695401, + 0.004162682686001062, + 0.0043550641275942326, + 0.004714942537248135, + 0.004553128499537706, + 0.004156507086008787, + 0.0036792990285903215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.11794472485780716 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02335038036108017 + ], + "opposite_logprob_tokens": [ + -3.773350477218628 + ], + "logprob_reward": -0.02335038036108017, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02335038036108017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2277996689081192, + 0.10935384035110474, + 0.19248764216899872, + 0.12346014380455017, + 0.05830085277557373, + 0.03602295368909836, + 0.023028602823615074, + 0.14753830432891846, + 0.04098772630095482, + 0.041020289063453674 + ], + "credit_map": [ + 0.022809213027358055, + 0.023942526429891586, + 0.022938121110200882, + 0.02049856446683407, + 0.02812349796295166, + 0.0404786616563797, + 0.027547873556613922, + 0.041461195796728134, + 0.008810799568891525, + 0.011497851461172104, + 0.014103860594332218, + 0.011409451253712177, + 0.012072228826582432, + 0.019259236752986908, + 0.015755433589220047, + 0.016444984823465347, + 0.024472787976264954, + 0.012069464661180973, + 0.020833130925893784, + 0.013040863908827305, + 0.028734268620610237, + 0.03518626093864441, + 0.031125841662287712, + 0.027025025337934494, + 0.017099015414714813, + 0.007587634027004242, + 0.01720181666314602, + 0.011723279021680355, + 0.010960361920297146, + 0.023564953356981277, + 0.01985376514494419, + 0.015469308942556381, + 0.0071983328089118, + 0.003589037572965026, + 0.007248422130942345, + 0.004603891633450985, + 0.00773172453045845, + 0.009990666061639786, + 0.010187845677137375, + 0.007750932592898607, + 0.005238765385001898, + 0.002182562369853258, + 0.0037018449511379004, + 0.0022583871614187956, + 0.005119356792420149, + 0.006868212018162012, + 0.004069108050316572, + 0.006584716495126486, + 0.003073332831263542, + 0.0017207623459398746, + 0.0026229543145745993, + 0.0019354200921952724, + 0.0035492125898599625, + 0.0038228339981287718, + 0.0037854863330721855, + 0.0025185991544276476, + 0.009025197476148605, + 0.008107599802315235, + 0.015426818281412125, + 0.012025866657495499, + 0.034738801419734955, + 0.028453204780817032, + 0.022237861528992653, + 0.01752294972538948, + 0.004493375774472952, + 0.003402557224035263, + 0.005503232125192881, + 0.0037106929812580347, + 0.005741656292229891, + 0.007286184001713991, + 0.004969389643520117, + 0.0058806417509913445, + 0.00430442625656724, + 0.0029521852266043425, + 0.006505890283733606, + 0.004556904546916485, + 0.004765292629599571, + 0.006988976150751114, + 0.004579281434416771, + 0.006367335096001625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.02335038036108017 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521558403968811 + ], + "opposite_logprob_tokens": [ + -1.5021557807922363 + ], + "logprob_reward": -0.2521558403968811, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521558403968811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2729799449443817, + 0.08579465001821518, + 0.20295850932598114, + 0.1083008423447609, + 0.09035491198301315, + 0.024752860888838768, + 0.02711424045264721, + 0.10339394211769104, + 0.03891504928469658, + 0.045435089617967606 + ], + "credit_map": [ + 0.028707057237625122, + 0.014688162133097649, + 0.020451903343200684, + 0.023122409358620644, + 0.041181400418281555, + 0.04960484430193901, + 0.0445873998105526, + 0.050636760890483856, + 0.00979290809482336, + 0.006967414636164904, + 0.009921779856085777, + 0.007521985098719597, + 0.01257163193076849, + 0.01385857816785574, + 0.012641704641282558, + 0.01251864992082119, + 0.02801421284675598, + 0.012060540728271008, + 0.02584805153310299, + 0.01661575213074684, + 0.025588495656847954, + 0.0345209464430809, + 0.033579617738723755, + 0.026730883866548538, + 0.015161008574068546, + 0.006822546012699604, + 0.012702756561338902, + 0.010711795650422573, + 0.010463111102581024, + 0.021728435531258583, + 0.01591699942946434, + 0.014794190414249897, + 0.00818210281431675, + 0.0060573238879442215, + 0.01203970331698656, + 0.009200077503919601, + 0.010733271948993206, + 0.015683839097619057, + 0.014620525762438774, + 0.013838060200214386, + 0.0029982614796608686, + 0.001781507395207882, + 0.0022061241324990988, + 0.002412754576653242, + 0.00337868626229465, + 0.0038173270877450705, + 0.004383391235023737, + 0.003774808021262288, + 0.003997670952230692, + 0.0022315089590847492, + 0.002670672722160816, + 0.002966665429994464, + 0.0041911001317203045, + 0.003952219150960445, + 0.0038286405615508556, + 0.003275761380791664, + 0.00963606871664524, + 0.0076537118293344975, + 0.01140652596950531, + 0.01288207434117794, + 0.013986521400511265, + 0.01590012013912201, + 0.01709619164466858, + 0.014832735992968082, + 0.0033030088525265455, + 0.0025502427015453577, + 0.004262956324964762, + 0.004283884074538946, + 0.006508713588118553, + 0.006606017705053091, + 0.005704193841665983, + 0.005696034524589777, + 0.004268256481736898, + 0.003676570486277342, + 0.0059954761527478695, + 0.00618871720507741, + 0.006475178990513086, + 0.007715955842286348, + 0.005223252810537815, + 0.005891676992177963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.2521558403968811 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792503900825977 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.009792503900825977, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792503900825977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26302996277809143, + 0.11274637281894684, + 0.14215488731861115, + 0.0633159652352333, + 0.06763563305139542, + 0.06918130815029144, + 0.06427910923957825, + 0.08161509037017822, + 0.03720112890005112, + 0.09884057939052582 + ], + "credit_map": [ + 0.0276188962161541, + 0.02466745488345623, + 0.025362471118569374, + 0.0227928776293993, + 0.028272205963730812, + 0.04950718954205513, + 0.029400434345006943, + 0.05540841072797775, + 0.010072526521980762, + 0.009865866042673588, + 0.009815153665840626, + 0.009804585948586464, + 0.011028285138309002, + 0.025132320821285248, + 0.014452899806201458, + 0.022574733942747116, + 0.01752413436770439, + 0.009984832257032394, + 0.017324503511190414, + 0.012960352934896946, + 0.019347649067640305, + 0.02441583201289177, + 0.019463781267404556, + 0.021133797243237495, + 0.009445181116461754, + 0.005719450768083334, + 0.006531076040118933, + 0.0050419531762599945, + 0.006141042336821556, + 0.012105206958949566, + 0.00868318136781454, + 0.009648867882788181, + 0.008817669935524464, + 0.006117452401667833, + 0.006469708867371082, + 0.004794761072844267, + 0.009221700020134449, + 0.01207678485661745, + 0.007430252153426409, + 0.012707301415503025, + 0.008173536509275436, + 0.0045937420800328255, + 0.0073536718264222145, + 0.004887823946774006, + 0.009209136478602886, + 0.013008288107812405, + 0.008774903602898121, + 0.013180206529796124, + 0.0086662657558918, + 0.004263324663043022, + 0.006807816214859486, + 0.006946737412363291, + 0.009172381833195686, + 0.009919212199747562, + 0.008586267940700054, + 0.009917102754116058, + 0.00763254938647151, + 0.006388330366462469, + 0.008939918130636215, + 0.008053467608988285, + 0.010923859663307667, + 0.012872390449047089, + 0.014285709708929062, + 0.012518870644271374, + 0.003294712631031871, + 0.002106209984049201, + 0.0034161475487053394, + 0.002027183538302779, + 0.008052500896155834, + 0.007042917422950268, + 0.005748913157731295, + 0.005512545816600323, + 0.013018403202295303, + 0.007915748283267021, + 0.008484621532261372, + 0.007605159655213356, + 0.015254548750817776, + 0.019025038927793503, + 0.009774246253073215, + 0.017762813717126846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.009792503900825977 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned high-angle, almost directly overhead, and remains static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick, syrupy stream of golden motor oil pouring onto a black, matte, textured metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the oil strike the surface with a heavy weight to carve intricate, river-like channels through dark gray industrial grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single, focused light source from the side casting warm highlights and deep micro-shadows on the oil's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil flow slowly and relentlessly outward, pushing aside the grime to reveal the clean surface beneath?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328958734869957 + ], + "opposite_logprob_tokens": [ + -3.773289680480957 + ], + "logprob_reward": -0.02328958734869957, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328958734869957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39439278841018677, + 0.10321883857250214, + 0.1304752677679062, + 0.08403664082288742, + 0.059744082391262054, + 0.030111221596598625, + 0.022109035402536392, + 0.09588100016117096, + 0.03240257129073143, + 0.04762856289744377 + ], + "credit_map": [ + 0.04284638538956642, + 0.0335649698972702, + 0.026442181318998337, + 0.03172043710947037, + 0.04604550823569298, + 0.08115795254707336, + 0.04691045731306076, + 0.08570487052202225, + 0.012509598396718502, + 0.009716826491057873, + 0.012601726688444614, + 0.009415887296199799, + 0.016298681497573853, + 0.01679287664592266, + 0.012681156396865845, + 0.013202070258557796, + 0.012775137089192867, + 0.009579747915267944, + 0.010614361613988876, + 0.013235064223408699, + 0.01540908869355917, + 0.026036646217107773, + 0.01788042113184929, + 0.024944795295596123, + 0.010784578509628773, + 0.005532427225261927, + 0.01210168469697237, + 0.007869813591241837, + 0.008129664696753025, + 0.01553848385810852, + 0.01145448163151741, + 0.012625506147742271, + 0.007443287875503302, + 0.0037151449359953403, + 0.0047706253826618195, + 0.006975550204515457, + 0.007174658123403788, + 0.010604091919958591, + 0.010314354673027992, + 0.008746366947889328, + 0.003267308697104454, + 0.0022260951809585094, + 0.0034314098302274942, + 0.0021512277889996767, + 0.004945265594869852, + 0.004413011483848095, + 0.0046195974573493, + 0.005057303234934807, + 0.003143085166811943, + 0.0017851572483778, + 0.002217995235696435, + 0.0022678303066641092, + 0.0029802245553582907, + 0.0034457824658602476, + 0.0031297942623496056, + 0.0031391633674502373, + 0.007787445560097694, + 0.006096730008721352, + 0.00744706392288208, + 0.0076047577895224094, + 0.018468743190169334, + 0.01401941291987896, + 0.022386183962225914, + 0.012070655822753906, + 0.003098420798778534, + 0.0029168077744543552, + 0.0036514257080852985, + 0.003984504379332066, + 0.003804061096161604, + 0.005705620627850294, + 0.004268433898687363, + 0.004973293282091618, + 0.004516199696809053, + 0.0036630050744861364, + 0.00689725810661912, + 0.005630830768495798, + 0.007137495093047619, + 0.007666626945137978, + 0.0055773877538740635, + 0.006539753172546625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.02328958734869957 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_14/index_tensor([2335, 2335])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_14/index_tensor([2335, 2335])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f3174c4dbd1ee80b2132927bde01ba6fa24a776c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_14/index_tensor([2335, 2335])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Inspired by the unique and elegant aesthetic of traditional Japanese scroll paintings, a slow, reverent crane shot begins at a low angle, focusing on the base of a translucent human figure standing against a richly textured, sepia-toned washi paper background. From the vaguely outlined feet of the ethereal figure, a complex system of delicate ink lines begins to glow, illuminating the circulatory system with a brilliant crimson and the nervous system with a brilliant, cool blue. These energy lines, with the delicate calligraphic quality of brushstrokes, flow gracefully upward in a steady, captivating ascent. As the camera steadily rises, it carefully follows the energetic path through the transparent legs, into the torso, and branching into the arms. The figure itself is defined not by a solid form but by a complex, luminous internal map of life force, contrasting starkly and beautifully against the minimalist ink background. The deliberate upward journey culminates when the camera reaches the head, where the red and blue lines converge into a dazzlingly complex network, illuminating the brain with a final, brilliant pulse of energy, creating a profoundly spiritual and meditative atmosphere, visualizing the ascending life force.", + "seed": 24568, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_14/index_2335_seed_24568_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04776507783681154, + "vlm_logprob": -0.04776507783681154, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.04776507783681154, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20220860838890076 + ], + "opposite_logprob_tokens": [ + -1.7022086381912231 + ], + "logprob_reward": -0.20220860838890076, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20220860838890076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32604119181632996, + 0.07598114758729935, + 0.10821308940649033, + 0.09805366396903992, + 0.1281103938817978, + 0.06344684213399887, + 0.052584175020456314, + 0.028951209038496017, + 0.09115903824567795, + 0.02745930105447769 + ], + "credit_map": [ + 0.04093964770436287, + 0.018984293565154076, + 0.1145525649189949, + 0.015525165013968945, + 0.014355480670928955, + 0.02815692499279976, + 0.013375375419855118, + 0.08015172183513641, + 0.010795729234814644, + 0.006226146128028631, + 0.012305967509746552, + 0.00501080509275198, + 0.004668107721954584, + 0.014021164737641811, + 0.008158938959240913, + 0.014794290997087955, + 0.020219692960381508, + 0.007288101129233837, + 0.014995724894106388, + 0.006310397759079933, + 0.015844807028770447, + 0.01818963699042797, + 0.009065664373338223, + 0.016299065202474594, + 0.010789643973112106, + 0.007680035661906004, + 0.01552205067127943, + 0.007268571760505438, + 0.013312296010553837, + 0.012014804407954216, + 0.014825418591499329, + 0.01664084382355213, + 0.01084154937416315, + 0.018780799582600594, + 0.0157370176166296, + 0.011015444062650204, + 0.02286945842206478, + 0.012962615117430687, + 0.012069642543792725, + 0.0238338615745306, + 0.007428646553307772, + 0.00882460456341505, + 0.005924586672335863, + 0.010256602428853512, + 0.008632398210465908, + 0.006454676389694214, + 0.008315003477036953, + 0.007610322441905737, + 0.007138561923056841, + 0.00603308854624629, + 0.007010682485997677, + 0.0071913315914571285, + 0.004727810621261597, + 0.009114369750022888, + 0.00557217700406909, + 0.005796152167022228, + 0.005064136348664761, + 0.003311383305117488, + 0.002245622221380472, + 0.006383856758475304, + 0.0026987360324710608, + 0.003044483484700322, + 0.004274646285921335, + 0.001928345882333815, + 0.025965234264731407, + 0.007094023283571005, + 0.006106237880885601, + 0.004319403320550919, + 0.0068635884672403336, + 0.030200336128473282, + 0.004845207091420889, + 0.0057650115340948105, + 0.002883328590542078, + 0.0040261875838041306, + 0.003124203300103545, + 0.00453043170273304, + 0.0031322448048740625, + 0.002673674374818802, + 0.003411959856748581, + 0.003677269909530878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.20220860838890076 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865092221647501 + ], + "opposite_logprob_tokens": [ + -4.758650779724121 + ], + "logprob_reward": -0.00865092221647501, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865092221647501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40145939588546753, + 0.17157863080501556, + 0.027309902012348175, + 0.037275101989507675, + 0.08591368049383163, + 0.025105569511651993, + 0.03178916126489639, + 0.05250087380409241, + 0.03518207371234894, + 0.1318855732679367 + ], + "credit_map": [ + 0.04184827208518982, + 0.02010047808289528, + 0.15222522616386414, + 0.019024096429347992, + 0.01583671011030674, + 0.02412387728691101, + 0.013865870423614979, + 0.11443488299846649, + 0.04512378200888634, + 0.013183624483644962, + 0.0193947646766901, + 0.013086559250950813, + 0.012586867436766624, + 0.0181545689702034, + 0.019002549350261688, + 0.031045909970998764, + 0.0039980122819542885, + 0.003078496316447854, + 0.004359659273177385, + 0.0024401426780968904, + 0.003695532213896513, + 0.004019758198410273, + 0.0022929024416953325, + 0.0034253988415002823, + 0.0036459744442254305, + 0.004081983584910631, + 0.005389282014220953, + 0.00396726792678237, + 0.004983815364539623, + 0.004704022314399481, + 0.0060471477918326855, + 0.004455605987459421, + 0.007026118692010641, + 0.01420626137405634, + 0.010995159856975079, + 0.009764396585524082, + 0.014735697768628597, + 0.009597198106348515, + 0.009347936138510704, + 0.010240916162729263, + 0.002633157651871443, + 0.004256763495504856, + 0.002533006016165018, + 0.003773132339119911, + 0.003478157566860318, + 0.0026656987611204386, + 0.002687897766008973, + 0.003077755682170391, + 0.004026906564831734, + 0.003226824104785919, + 0.004798208829015493, + 0.0046156421303749084, + 0.003923068754374981, + 0.0044757770374417305, + 0.0033274758607149124, + 0.0033952579833567142, + 0.01128874160349369, + 0.00358279375359416, + 0.0038409556727856398, + 0.012714086100459099, + 0.0024447825271636248, + 0.007143407594412565, + 0.007920836098492146, + 0.00356527091935277, + 0.008147920481860638, + 0.003898059483617544, + 0.00258029717952013, + 0.002198929898440838, + 0.0038002647925168276, + 0.010113821364939213, + 0.0025987974368035793, + 0.0018439851701259613, + 0.010546751320362091, + 0.021599439904093742, + 0.01770462654531002, + 0.02288629673421383, + 0.020619329065084457, + 0.015040770173072815, + 0.016172152012586594, + 0.007316213101148605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.00865092221647501 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021977098658680916 + ], + "opposite_logprob_tokens": [ + -6.127197742462158 + ], + "logprob_reward": -0.0021977098658680916, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021977098658680916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.404141902923584, + 0.10349990427494049, + 0.057686544954776764, + 0.03746756166219711, + 0.0371064618229866, + 0.025344694033265114, + 0.03848665580153465, + 0.05187005549669266, + 0.14973895251750946, + 0.09465727210044861 + ], + "credit_map": [ + 0.031735505908727646, + 0.0178185123950243, + 0.1537424772977829, + 0.015219036489725113, + 0.013972334563732147, + 0.025348598137497902, + 0.01521040964871645, + 0.1310950219631195, + 0.020706765353679657, + 0.0072831567376852036, + 0.014474117197096348, + 0.007290218491107225, + 0.006554626859724522, + 0.017184073105454445, + 0.010908456519246101, + 0.019098488613963127, + 0.00975880492478609, + 0.004823531489819288, + 0.006951835472136736, + 0.004146139603108168, + 0.00925798062235117, + 0.009309030137956142, + 0.005187389440834522, + 0.008251827210187912, + 0.004027460236102343, + 0.004120958503335714, + 0.005493251606822014, + 0.0032977282535284758, + 0.0049448637291789055, + 0.004342489875853062, + 0.005979570560157299, + 0.005261234473437071, + 0.003444012487307191, + 0.0044577475637197495, + 0.005259658209979534, + 0.005338186863809824, + 0.0049070571549236774, + 0.004143472760915756, + 0.003937522415071726, + 0.005618803668767214, + 0.002832331694662571, + 0.0034237669315189123, + 0.0023148527834564447, + 0.0036151011008769274, + 0.0036140214651823044, + 0.0028382912278175354, + 0.0036721909418702126, + 0.003034136025235057, + 0.005503532942384481, + 0.004355046432465315, + 0.004872365389019251, + 0.00743128452450037, + 0.0036787670105695724, + 0.004777302499860525, + 0.004056572448462248, + 0.003811781294643879, + 0.008162093348801136, + 0.0045306067913770676, + 0.005641224328428507, + 0.009808944538235664, + 0.004381697159260511, + 0.006943673826754093, + 0.007439969573169947, + 0.004961838945746422, + 0.041669704020023346, + 0.010707157664000988, + 0.012285818345844746, + 0.006348639260977507, + 0.010276226326823235, + 0.05079926177859306, + 0.007325123064219952, + 0.010327015072107315, + 0.013722251169383526, + 0.014010488986968994, + 0.01564698852598667, + 0.010442277416586876, + 0.014634187333285809, + 0.011084847152233124, + 0.0077967094257473946, + 0.007319511380046606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0021977098658680916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023281319066882133 + ], + "opposite_logprob_tokens": [ + -3.7732813358306885 + ], + "logprob_reward": -0.023281319066882133, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023281319066882133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23417817056179047, + 0.2034849226474762, + 0.04730166867375374, + 0.05038954317569733, + 0.03852139413356781, + 0.02679274044930935, + 0.09085480868816376, + 0.19043825566768646, + 0.052788812667131424, + 0.06524975597858429 + ], + "credit_map": [ + 0.02752993442118168, + 0.015127429738640785, + 0.06803501397371292, + 0.016404759138822556, + 0.011927736923098564, + 0.01932663843035698, + 0.012052811682224274, + 0.06377381831407547, + 0.04351692646741867, + 0.014295892789959908, + 0.026058519259095192, + 0.011261849664151669, + 0.012268844991922379, + 0.03282314911484718, + 0.020861007273197174, + 0.04239872470498085, + 0.008512166328728199, + 0.004183952696621418, + 0.006897382438182831, + 0.004150955472141504, + 0.006093710660934448, + 0.0060714129358530045, + 0.004399935249239206, + 0.00699215242639184, + 0.004610968753695488, + 0.006027190946042538, + 0.0076543656177818775, + 0.005294989328831434, + 0.007077344693243504, + 0.005874401889741421, + 0.007720204535871744, + 0.006130075082182884, + 0.004393374547362328, + 0.005383588839322329, + 0.00539436936378479, + 0.005171340890228748, + 0.0056375060230493546, + 0.0034106874372810125, + 0.003717973828315735, + 0.005412548314779997, + 0.0028921726625412703, + 0.004015759099274874, + 0.00297724362462759, + 0.004150294698774815, + 0.004040626343339682, + 0.003343381220474839, + 0.002950403606519103, + 0.002422856166958809, + 0.01307197380810976, + 0.009160540997982025, + 0.010340066626667976, + 0.021288976073265076, + 0.010816561989486217, + 0.00878165289759636, + 0.009838039986789227, + 0.007556987926363945, + 0.03807136416435242, + 0.014607414603233337, + 0.014392496086657047, + 0.045979440212249756, + 0.010657726787030697, + 0.027115285396575928, + 0.027120642364025116, + 0.012493857182562351, + 0.013717249035835266, + 0.0046245320700109005, + 0.004517737776041031, + 0.003698958782479167, + 0.004511604551225901, + 0.015143828466534615, + 0.002862976398319006, + 0.0037119206972420216, + 0.007877947762608528, + 0.009771578945219517, + 0.007838111370801926, + 0.011532283388078213, + 0.00975953508168459, + 0.006932586431503296, + 0.007439755368977785, + 0.004097951576113701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.023281319066882133 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024868296459317207 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.0024868296459317207, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024868296459317207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2745742201805115, + 0.057407017797231674, + 0.0981934517621994, + 0.047631360590457916, + 0.08084100484848022, + 0.06649310886859894, + 0.0582890585064888, + 0.1187717542052269, + 0.06053467094898224, + 0.1372644603252411 + ], + "credit_map": [ + 0.02407667599618435, + 0.007118951994925737, + 0.1009097620844841, + 0.010707749053835869, + 0.008692762814462185, + 0.018443696200847626, + 0.010560753755271435, + 0.09406380355358124, + 0.011030832305550575, + 0.005795578006654978, + 0.009019451215863228, + 0.0052673653699457645, + 0.005622256547212601, + 0.007171414792537689, + 0.005578316282480955, + 0.007921786978840828, + 0.018499502912163734, + 0.008451374247670174, + 0.012889036908745766, + 0.008461994118988514, + 0.010317363776266575, + 0.016018755733966827, + 0.007657104171812534, + 0.01589829847216606, + 0.006992519833147526, + 0.0038672948721796274, + 0.006356610916554928, + 0.0037414974067360163, + 0.003680639434605837, + 0.009046961553394794, + 0.005691310856491327, + 0.008254516869783401, + 0.010452710092067719, + 0.00770813412964344, + 0.011887069791555405, + 0.008343256078660488, + 0.009770305827260017, + 0.009235586039721966, + 0.012131894007325172, + 0.011312037706375122, + 0.00840678345412016, + 0.007325399201363325, + 0.00761830760166049, + 0.011414824984967709, + 0.00809699110686779, + 0.007749906275421381, + 0.008086131885647774, + 0.007794745732098818, + 0.007161343935877085, + 0.007043713703751564, + 0.008342253975570202, + 0.008823715150356293, + 0.0071242256090044975, + 0.007529116235673428, + 0.005613820161670446, + 0.006650859024375677, + 0.017322484403848648, + 0.018591368570923805, + 0.014986755326390266, + 0.012251953594386578, + 0.014683774672448635, + 0.014791928231716156, + 0.01417127437889576, + 0.011972181499004364, + 0.01178217027336359, + 0.008450380526483059, + 0.005976475775241852, + 0.005513666197657585, + 0.006264583207666874, + 0.014149521477520466, + 0.00534867774695158, + 0.0030491803772747517, + 0.011805578134953976, + 0.023592624813318253, + 0.019130216911435127, + 0.023151468485593796, + 0.0215041171759367, + 0.014982220716774464, + 0.012012938037514687, + 0.011085259728133678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.0024868296459317207 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Inspired by the unique and elegant aesthetic of traditional Japanese scroll paintings, a slow, reverent crane shot begins at a low angle, focusing on the base of a translucent human figure standing against a richly textured, sepia-toned washi paper background. From the vaguely outlined feet of the ethereal figure, a complex system of delicate ink lines begins to glow, illuminating the circulatory system with a brilliant crimson and the nervous system with a brilliant, cool blue. These energy lines, with the delicate calligraphic quality of brushstrokes, flow gracefully upward in a steady, captivating ascent. As the camera steadily rises, it carefully follows the energetic path through the transparent legs, into the torso, and branching into the arms. The figure itself is defined not by a solid form but by a complex, luminous internal map of life force, contrasting starkly and beautifully against the minimalist ink background. The deliberate upward journey culminates when the camera reaches the head, where the red and blue lines converge into a dazzlingly complex network, illuminating the brain with a final, brilliant pulse of energy, creating a profoundly spiritual and meditative atmosphere, visualizing the ascending life force.", + "seed": 24569, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_14/index_2335_seed_24569_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.042230316018685696, + "vlm_logprob": -0.042230316018685696, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.042230316018685696, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18069027364253998 + ], + "opposite_logprob_tokens": [ + -1.8056902885437012 + ], + "logprob_reward": -0.18069027364253998, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18069027364253998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2915875315666199, + 0.018991028890013695, + 0.09145757555961609, + 0.015031404793262482, + 0.024755170568823814, + 0.02216954343020916, + 0.025150442495942116, + 0.4733183681964874, + 0.02158482000231743, + 0.01595407724380493 + ], + "credit_map": [ + 0.03595674782991409, + 0.023307690396904945, + 0.045990951359272, + 0.023825209587812424, + 0.018945762887597084, + 0.06454816460609436, + 0.02035892941057682, + 0.058654073625802994, + 0.00349602778442204, + 0.0013549868017435074, + 0.003253029193729162, + 0.0025688272435218096, + 0.0016171354800462723, + 0.0022277289535850286, + 0.0018861335702240467, + 0.0025871603284031153, + 0.01728990487754345, + 0.004865056835114956, + 0.013513281010091305, + 0.0070430743508040905, + 0.004416102543473244, + 0.016594914719462395, + 0.006981625687330961, + 0.020753618329763412, + 0.0014147263718768954, + 0.0013035571901127696, + 0.002043170854449272, + 0.0027923474553972483, + 0.0010965998517349362, + 0.0020879439543932676, + 0.0019215617794543505, + 0.0023714981507509947, + 0.0026100389659404755, + 0.0027293795719742775, + 0.0024496009573340416, + 0.0023204199969768524, + 0.0028263276908546686, + 0.0036962181329727173, + 0.004151648376137018, + 0.003971536178141832, + 0.002767552388831973, + 0.002583489054813981, + 0.0017658042488619685, + 0.002798561006784439, + 0.0035082967951893806, + 0.0038449845742434263, + 0.0017291678814217448, + 0.0031716881785541773, + 0.003527262946590781, + 0.001869344967417419, + 0.004136546049267054, + 0.002230226993560791, + 0.0019273939542472363, + 0.0031981845386326313, + 0.0025220545940101147, + 0.005739428568631411, + 0.03177914023399353, + 0.02627825364470482, + 0.030519181862473488, + 0.024076469242572784, + 0.15176482498645782, + 0.02991418167948723, + 0.1576220542192459, + 0.021364297717809677, + 0.0039606294594705105, + 0.0016605101991444826, + 0.0023041721433401108, + 0.0014379994245246053, + 0.003126472467556596, + 0.003040747717022896, + 0.0025238748639822006, + 0.0035304129123687744, + 0.0016342895105481148, + 0.002042901935055852, + 0.002080893609672785, + 0.0016396917635574937, + 0.001417586812749505, + 0.0023864179383963346, + 0.0016212803311645985, + 0.003131016157567501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.18069027364253998 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005967182572931051 + ], + "opposite_logprob_tokens": [ + -5.130967140197754 + ], + "logprob_reward": -0.005967182572931051, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005967182572931051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23458625376224518, + 0.04080602899193764, + 0.05354692041873932, + 0.03589998930692673, + 0.1003735139966011, + 0.043313078582286835, + 0.019022716209292412, + 0.34162265062332153, + 0.10415475815534592, + 0.026674089953303337 + ], + "credit_map": [ + 0.021574147045612335, + 0.029902789741754532, + 0.021321719512343407, + 0.026898987591266632, + 0.020442422479391098, + 0.03972196951508522, + 0.030649330466985703, + 0.044074881821870804, + 0.006658573169261217, + 0.002259032567963004, + 0.008273337036371231, + 0.0033875664230436087, + 0.00277765397913754, + 0.004398271907120943, + 0.002685151295736432, + 0.010366443544626236, + 0.00982386153191328, + 0.0035357188899070024, + 0.007469024043530226, + 0.004104760941118002, + 0.003451868426054716, + 0.008785981684923172, + 0.004433570429682732, + 0.011942138895392418, + 0.005079230759292841, + 0.002912551863119006, + 0.004156981594860554, + 0.004495244473218918, + 0.002995534101501107, + 0.004970673006027937, + 0.00657902006059885, + 0.004710753448307514, + 0.01132923737168312, + 0.013691146858036518, + 0.010494266636669636, + 0.009685523808002472, + 0.013136284425854683, + 0.015593650750815868, + 0.01026589423418045, + 0.01617751084268093, + 0.006017394829541445, + 0.005454584024846554, + 0.0035541218239814043, + 0.006450514309108257, + 0.00418666610494256, + 0.008398308418691158, + 0.0035874778404831886, + 0.0056640105322003365, + 0.002373511204496026, + 0.002586894668638706, + 0.0020058657974004745, + 0.0025665615685284138, + 0.0020213532261550426, + 0.001950905891135335, + 0.0030766911804676056, + 0.0024409324396401644, + 0.02330912835896015, + 0.01870976947247982, + 0.022890647873282433, + 0.018239008262753487, + 0.1075320914387703, + 0.021849308162927628, + 0.11320428550243378, + 0.015888407826423645, + 0.01792137324810028, + 0.007287757471203804, + 0.015743069350719452, + 0.008514253422617912, + 0.013693020679056644, + 0.014684801921248436, + 0.01426358986645937, + 0.012046894989907742, + 0.0021839027758687735, + 0.004110290203243494, + 0.0034277273807674646, + 0.0023095454089343548, + 0.0019720010459423065, + 0.00398485129699111, + 0.003443059977144003, + 0.005242711398750544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.005967182572931051 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017162130679935217 + ], + "opposite_logprob_tokens": [ + -6.376716136932373 + ], + "logprob_reward": -0.0017162130679935217, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017162130679935217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14907784759998322, + 0.04270685464143753, + 0.031158506870269775, + 0.02151546999812126, + 0.02686188369989395, + 0.06094434857368469, + 0.038152970373630524, + 0.4657021462917328, + 0.10523572564125061, + 0.05864424258470535 + ], + "credit_map": [ + 0.01671609655022621, + 0.021176433190703392, + 0.022773226723074913, + 0.014440441504120827, + 0.017944123595952988, + 0.024722613394260406, + 0.013316107913851738, + 0.017988793551921844, + 0.008878655731678009, + 0.0016615245258435607, + 0.006159407552331686, + 0.0033534776885062456, + 0.0021071077790111303, + 0.0050857895985245705, + 0.0025827002245932817, + 0.01287818979471922, + 0.005133626516908407, + 0.0022506611421704292, + 0.004455129615962505, + 0.0035437268670648336, + 0.002440706128254533, + 0.005243975669145584, + 0.00254925899207592, + 0.005541421473026276, + 0.0023814004380255938, + 0.0019651581533253193, + 0.002652454422786832, + 0.0036761595401912928, + 0.0018520450685173273, + 0.0027519147843122482, + 0.0027934671379625797, + 0.0034428692888468504, + 0.0024208128452301025, + 0.003584692720323801, + 0.0031018094159662724, + 0.0022782390005886555, + 0.003906037425622344, + 0.0038144593127071857, + 0.0035852056462317705, + 0.004170626867562532, + 0.0060684410855174065, + 0.007609482388943434, + 0.006743519101291895, + 0.006298261694610119, + 0.008311041630804539, + 0.011422612704336643, + 0.0053581115789711475, + 0.009132876992225647, + 0.004807149060070515, + 0.004108654335141182, + 0.00584547221660614, + 0.0037263508420437574, + 0.004796984605491161, + 0.004013904370367527, + 0.004713225178420544, + 0.006141230463981628, + 0.0312265083193779, + 0.025477217510342598, + 0.03131788223981857, + 0.025029776617884636, + 0.14576099812984467, + 0.03019069693982601, + 0.15461955964565277, + 0.02207954227924347, + 0.016386937350034714, + 0.007718358654528856, + 0.01516364049166441, + 0.007332406472414732, + 0.01596193201839924, + 0.014915108680725098, + 0.015565089881420135, + 0.01219224650412798, + 0.00583169749006629, + 0.007057130802422762, + 0.007823104038834572, + 0.005288979504257441, + 0.006144221406430006, + 0.010693046264350414, + 0.007354659028351307, + 0.008451402187347412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.0017162130679935217 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020580438897013664 + ], + "opposite_logprob_tokens": [ + -3.895580530166626 + ], + "logprob_reward": -0.020580438897013664, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020580438897013664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3354181945323944, + 0.055345457047224045, + 0.06561215966939926, + 0.06396839022636414, + 0.06469885259866714, + 0.08340342342853546, + 0.06647492945194244, + 0.13799093663692474, + 0.0301753431558609, + 0.09691234678030014 + ], + "credit_map": [ + 0.03301024064421654, + 0.045525483787059784, + 0.030147302895784378, + 0.04228616878390312, + 0.033425260335206985, + 0.05302371829748154, + 0.043540023267269135, + 0.05446000024676323, + 0.010205202735960484, + 0.0031188500579446554, + 0.007805114611983299, + 0.004301364533603191, + 0.0035922250244766474, + 0.00593118742108345, + 0.0040419925935566425, + 0.016349518671631813, + 0.009802327491343021, + 0.007527749054133892, + 0.008659927174448967, + 0.007556622847914696, + 0.006252439226955175, + 0.009486021474003792, + 0.005003748927265406, + 0.011323319748044014, + 0.011860644444823265, + 0.007085486780852079, + 0.009146049618721008, + 0.008756636641919613, + 0.006270022597163916, + 0.008462795056402683, + 0.006066077388823032, + 0.0063206772319972515, + 0.006735091097652912, + 0.006717588286846876, + 0.008931506425142288, + 0.006321485619992018, + 0.0063267364166677, + 0.010876262560486794, + 0.00823680404573679, + 0.010553381405770779, + 0.007331076543778181, + 0.013049361295998096, + 0.006506515201181173, + 0.016131559386849403, + 0.011675729416310787, + 0.009485088288784027, + 0.007665696088224649, + 0.011558395810425282, + 0.006956405937671661, + 0.004857260268181562, + 0.015599580481648445, + 0.006469352170825005, + 0.005387095734477043, + 0.005671720951795578, + 0.006965878885239363, + 0.014567635022103786, + 0.009345191530883312, + 0.007944959215819836, + 0.009855842217803001, + 0.007848640903830528, + 0.0410299114882946, + 0.009650375694036484, + 0.045173123478889465, + 0.007142884191125631, + 0.004886218346655369, + 0.0037976314779371023, + 0.003944639582186937, + 0.0025389662478119135, + 0.0036449225153774023, + 0.0039083086885511875, + 0.003866213373839855, + 0.003588445018976927, + 0.01047599595040083, + 0.015635361894965172, + 0.010808499529957771, + 0.007488647475838661, + 0.010588303208351135, + 0.018538059666752815, + 0.009102287702262402, + 0.014275186695158482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.020580438897013664 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021974719129502773 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.0021974719129502773, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021974719129502773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19977132976055145, + 0.05005105212330818, + 0.0385141596198082, + 0.023281948640942574, + 0.032389085739851, + 0.1093888208270073, + 0.03273845091462135, + 0.3040038049221039, + 0.054667938500642776, + 0.15519337356090546 + ], + "credit_map": [ + 0.020060481503605843, + 0.03149466589093208, + 0.026414398103952408, + 0.01997447945177555, + 0.025746049359440804, + 0.02954842895269394, + 0.02013850025832653, + 0.02639433927834034, + 0.009481733664870262, + 0.0029589212499558926, + 0.00791015475988388, + 0.004299200605601072, + 0.0034696415532380342, + 0.00614979537203908, + 0.003173884004354477, + 0.012607716955244541, + 0.00660948920994997, + 0.0031017176806926727, + 0.005749638192355633, + 0.00430277967825532, + 0.003216089913621545, + 0.0067085945047438145, + 0.002604634501039982, + 0.006221218965947628, + 0.001994173740968108, + 0.0017875125631690025, + 0.003250120673328638, + 0.0031486214138567448, + 0.0016510991845279932, + 0.0033037038519978523, + 0.0027183396741747856, + 0.005428378004580736, + 0.0038820109330117702, + 0.0035235697869211435, + 0.004080707672983408, + 0.0033404724672436714, + 0.003944749012589455, + 0.005196229554712772, + 0.004388061352074146, + 0.004033284727483988, + 0.009232111275196075, + 0.01877930387854576, + 0.012087097391486168, + 0.017400434240698814, + 0.013802746310830116, + 0.015851736068725586, + 0.008219127543270588, + 0.01401625107973814, + 0.006093703210353851, + 0.004870790522545576, + 0.005112338345497847, + 0.0027988916262984276, + 0.003381238551810384, + 0.002976544899865985, + 0.0030417381785809994, + 0.004463206976652145, + 0.02124829590320587, + 0.016852600499987602, + 0.020382624119520187, + 0.015901701524853706, + 0.0966581255197525, + 0.019719701260328293, + 0.09891204535961151, + 0.014328693971037865, + 0.006687719840556383, + 0.005206779111176729, + 0.008056939579546452, + 0.0053085931576788425, + 0.009588642045855522, + 0.00659424951300025, + 0.007610377389937639, + 0.005614642519503832, + 0.018408624455332756, + 0.02345072105526924, + 0.015589703805744648, + 0.011233126744627953, + 0.015614395961165428, + 0.03165879100561142, + 0.014694145880639553, + 0.024543873965740204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.0021974719129502773 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Inspired by the unique and elegant aesthetic of traditional Japanese scroll paintings, a slow, reverent crane shot begins at a low angle, focusing on the base of a translucent human figure standing against a richly textured, sepia-toned washi paper background. From the vaguely outlined feet of the ethereal figure, a complex system of delicate ink lines begins to glow, illuminating the circulatory system with a brilliant crimson and the nervous system with a brilliant, cool blue. These energy lines, with the delicate calligraphic quality of brushstrokes, flow gracefully upward in a steady, captivating ascent. As the camera steadily rises, it carefully follows the energetic path through the transparent legs, into the torso, and branching into the arms. The figure itself is defined not by a solid form but by a complex, luminous internal map of life force, contrasting starkly and beautifully against the minimalist ink background. The deliberate upward journey culminates when the camera reaches the head, where the red and blue lines converge into a dazzlingly complex network, illuminating the brain with a final, brilliant pulse of energy, creating a profoundly spiritual and meditative atmosphere, visualizing the ascending life force.", + "seed": 24570, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_14/index_2335_seed_24570_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03479292960837484, + "vlm_logprob": -0.03479292960837484, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.03479292960837484, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1437244564294815 + ], + "opposite_logprob_tokens": [ + -2.0187244415283203 + ], + "logprob_reward": -0.1437244564294815, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1437244564294815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8882613182067871, + 0.0143766850233078, + 0.04517767205834389, + 0.006179800257086754, + 0.006088991649448872, + 0.00503613892942667, + 0.007619843352586031, + 0.012851957231760025, + 0.004791715648025274, + 0.009615851566195488 + ], + "credit_map": [ + 0.08568213135004044, + 0.012441110797226429, + 0.29534879326820374, + 0.04627405107021332, + 0.011543567292392254, + 0.0932018905878067, + 0.03133068233728409, + 0.31243911385536194, + 0.0014984302688390017, + 0.0009111575782299042, + 0.003548185108229518, + 0.0006562480702996254, + 0.0010156907374039292, + 0.0017654206603765488, + 0.0008883925038389862, + 0.004093159921467304, + 0.006156809162348509, + 0.0037565818056464195, + 0.009199715219438076, + 0.004354892298579216, + 0.0038698683492839336, + 0.006825596559792757, + 0.005221218336373568, + 0.005792986135929823, + 0.0007607985171489418, + 0.0006950735114514828, + 0.0009979273891076446, + 0.0007275897078216076, + 0.0006631250726059079, + 0.0006126585067249835, + 0.0007947466219775379, + 0.0009278802899643779, + 0.0005375220207497478, + 0.0009761921828612685, + 0.000701704528182745, + 0.0005855554481968284, + 0.0006061085732653737, + 0.0012786859879270196, + 0.0005623952602036297, + 0.0008408271824009717, + 0.0006992056150920689, + 0.0005159291904419661, + 0.0004072319425176829, + 0.00036470266059041023, + 0.0006211148574948311, + 0.0013796788407489657, + 0.000334531650878489, + 0.0007137436186894774, + 0.0007604751153849065, + 0.001438054721802473, + 0.0004687750188168138, + 0.0008224351331591606, + 0.0017345844535157084, + 0.0010959100909531116, + 0.0007529784343205392, + 0.0005466299480758607, + 0.0014383506495505571, + 0.0013466209638863802, + 0.0015222398797050118, + 0.0022194350603967905, + 0.0015048935310915112, + 0.0012780383694916964, + 0.0019214675994589925, + 0.0016209103632718325, + 0.00077183882240206, + 0.0006434226525016129, + 0.0006817473331466317, + 0.00046945540816523135, + 0.0004908915143460035, + 0.0006463977042585611, + 0.0005769205163232982, + 0.0005110418424010277, + 0.0015132854459807277, + 0.0011739577166736126, + 0.0013628251617774367, + 0.0008321827044710517, + 0.0010703189764171839, + 0.001516148098744452, + 0.0009790544863790274, + 0.0011680787429213524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.1437244564294815 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012566078454256058 + ], + "opposite_logprob_tokens": [ + -4.387566089630127 + ], + "logprob_reward": -0.012566078454256058, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012566078454256058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43596869707107544, + 0.09069320559501648, + 0.07865245640277863, + 0.0959777981042862, + 0.030343012884259224, + 0.06631188839673996, + 0.07160428911447525, + 0.05591398477554321, + 0.03082682378590107, + 0.043707773089408875 + ], + "credit_map": [ + 0.04146817699074745, + 0.016179492697119713, + 0.14100921154022217, + 0.02544073387980461, + 0.012016418389976025, + 0.03898816555738449, + 0.018986225128173828, + 0.14188025891780853, + 0.008873983286321163, + 0.006780034862458706, + 0.019046781584620476, + 0.005124974064528942, + 0.0062125129625201225, + 0.010259835049510002, + 0.008191263303160667, + 0.026203813031315804, + 0.009281693957746029, + 0.008205367252230644, + 0.011760056018829346, + 0.008346478454768658, + 0.009257351979613304, + 0.01116607803851366, + 0.012640612199902534, + 0.0079948203638196, + 0.006897014565765858, + 0.009624646976590157, + 0.014640102162957191, + 0.016950853168964386, + 0.008332549594342709, + 0.008311360143125057, + 0.015748118981719017, + 0.015473144128918648, + 0.0031213939655572176, + 0.00408159289509058, + 0.004334843717515469, + 0.0031113692093640566, + 0.002368458779528737, + 0.005371930543333292, + 0.0026545876171439886, + 0.005298836622387171, + 0.009738421998918056, + 0.009591387584805489, + 0.0036418281961232424, + 0.0030740341171622276, + 0.0058287628926336765, + 0.026633724570274353, + 0.0025121679063886404, + 0.005291558802127838, + 0.006597442552447319, + 0.017331067472696304, + 0.005586328916251659, + 0.004919543396681547, + 0.01576758362352848, + 0.00999480951577425, + 0.005047304090112448, + 0.006360211409628391, + 0.008109326474368572, + 0.008157055824995041, + 0.006385582499206066, + 0.0073255011811852455, + 0.007230041548609734, + 0.006064722314476967, + 0.007606058847159147, + 0.005035692360252142, + 0.006931246258318424, + 0.0033493652008473873, + 0.003740319749340415, + 0.0020986986346542835, + 0.0024068320635706186, + 0.004598420113325119, + 0.003050015075132251, + 0.004651925526559353, + 0.007643909193575382, + 0.004963286686688662, + 0.007078280672430992, + 0.004200606141239405, + 0.004271928686648607, + 0.005805710796266794, + 0.004393359646201134, + 0.0053506894037127495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.012566078454256058 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031969649717211723 + ], + "opposite_logprob_tokens": [ + -5.753197193145752 + ], + "logprob_reward": -0.0031969649717211723, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031969649717211723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6247332692146301, + 0.01976003125309944, + 0.019680196419358253, + 0.015808837488293648, + 0.01126505434513092, + 0.081370048224926, + 0.0308139119297266, + 0.05175396427512169, + 0.07185694575309753, + 0.07295779883861542 + ], + "credit_map": [ + 0.060876939445734024, + 0.008321826346218586, + 0.20503836870193481, + 0.03194183483719826, + 0.007909918203949928, + 0.06913638859987259, + 0.02229353040456772, + 0.21921434998512268, + 0.0020486346911638975, + 0.0015765014104545116, + 0.003752374555915594, + 0.0011055973591282964, + 0.0012006046017631888, + 0.0026400426868349314, + 0.0018622491043061018, + 0.005574024748057127, + 0.002509617479518056, + 0.0015390372136607766, + 0.004345760215073824, + 0.0018163148779422045, + 0.0016400936292484403, + 0.0029046267736703157, + 0.002291229786351323, + 0.0026335157454013824, + 0.0020270859822630882, + 0.001448378898203373, + 0.003337973030284047, + 0.0017049078596755862, + 0.0012979978928342462, + 0.0015571161638945341, + 0.0019238502718508244, + 0.0025115255266427994, + 0.0008249279926531017, + 0.0015778765082359314, + 0.00192673085257411, + 0.001067116390913725, + 0.0013423183700069785, + 0.0019388125510886312, + 0.0008372933370992541, + 0.001749977353028953, + 0.010088879615068436, + 0.015328265726566315, + 0.0022292016074061394, + 0.0033181069884449244, + 0.008977694436907768, + 0.03218529000878334, + 0.0031358185224235058, + 0.0061067864298820496, + 0.0038122960831969976, + 0.004323897883296013, + 0.0034545506350696087, + 0.003693149657920003, + 0.005740324500948191, + 0.003398091997951269, + 0.003516125027090311, + 0.0028754728846251965, + 0.006752731744199991, + 0.006803040858358145, + 0.005718220490962267, + 0.006963654886931181, + 0.008978406898677349, + 0.004519358742982149, + 0.006178730633109808, + 0.00583981815725565, + 0.010628441348671913, + 0.009463413618505001, + 0.006843598559498787, + 0.009086739271879196, + 0.005754717160016298, + 0.008205811493098736, + 0.012813868001103401, + 0.009060352109372616, + 0.00988455954939127, + 0.006644304841756821, + 0.014345268718898296, + 0.0070922947488725185, + 0.007323312573134899, + 0.009777223691344261, + 0.006468676496297121, + 0.01142214797437191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.0031969649717211723 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537825852632523 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012537825852632523, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537825852632523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4618387520313263, + 0.03330259770154953, + 0.12600916624069214, + 0.07993268221616745, + 0.055134788155555725, + 0.11908581107854843, + 0.03938952833414078, + 0.052147313952445984, + 0.013805811293423176, + 0.019353587180376053 + ], + "credit_map": [ + 0.04306688532233238, + 0.00906812958419323, + 0.1449122577905655, + 0.024999387562274933, + 0.008164871484041214, + 0.051963966339826584, + 0.018231987953186035, + 0.16143128275871277, + 0.0038770537357777357, + 0.0032394968438893557, + 0.00554521381855011, + 0.003153929254040122, + 0.0024939191062003374, + 0.00561004551127553, + 0.0036414351779967546, + 0.005741504020988941, + 0.01563437655568123, + 0.01145673543214798, + 0.01905130222439766, + 0.015850355848670006, + 0.01064146775752306, + 0.014663588255643845, + 0.02634887397289276, + 0.0123624661937356, + 0.006720826029777527, + 0.007943982258439064, + 0.01313774660229683, + 0.012315161526203156, + 0.006772615015506744, + 0.006738697178661823, + 0.012986443936824799, + 0.013317215256392956, + 0.00457829050719738, + 0.0035877323243767023, + 0.011552534066140652, + 0.0070123812183737755, + 0.004970825742930174, + 0.008982985280454159, + 0.003972599282860756, + 0.010477438569068909, + 0.021463902667164803, + 0.012964385561645031, + 0.007340195123106241, + 0.006504325196146965, + 0.014489371329545975, + 0.03714457154273987, + 0.0061790733598172665, + 0.01299999002367258, + 0.005481217987835407, + 0.005571222864091396, + 0.004732451401650906, + 0.004455805290490389, + 0.008929054252803326, + 0.0035009710118174553, + 0.0034909266978502274, + 0.003227877663448453, + 0.0068498956970870495, + 0.007464839145541191, + 0.005853588227182627, + 0.00889728032052517, + 0.0061661009676754475, + 0.005032198503613472, + 0.0068768467754125595, + 0.005006565246731043, + 0.002216304885223508, + 0.001434852252714336, + 0.001601780648343265, + 0.001235014176927507, + 0.0014325708616524935, + 0.0018683711532503366, + 0.0018007004400715232, + 0.0022162171080708504, + 0.0030735258478671312, + 0.0026071358006447554, + 0.0022756445687264204, + 0.0017045839922502637, + 0.002265143906697631, + 0.0026710506062954664, + 0.0017765350639820099, + 0.0029799677431583405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.012537825852632523 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019393223337829113 + ], + "opposite_logprob_tokens": [ + -6.251939296722412 + ], + "logprob_reward": -0.0019393223337829113, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019393223337829113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5385861992835999, + 0.03312774375081062, + 0.042779918760061264, + 0.02867811918258667, + 0.028152190148830414, + 0.08439251780509949, + 0.04532923176884651, + 0.06132880225777626, + 0.04065365716814995, + 0.09697160124778748 + ], + "credit_map": [ + 0.05081450566649437, + 0.009462233632802963, + 0.17831794917583466, + 0.028395842760801315, + 0.008726993575692177, + 0.05622798949480057, + 0.018834171816706657, + 0.18780647218227386, + 0.0034242619294673204, + 0.0019013246055692434, + 0.007671068888157606, + 0.0020365926902741194, + 0.0018768274458125234, + 0.004087101202458143, + 0.002267299685627222, + 0.009863263927400112, + 0.005685324780642986, + 0.0038718453142791986, + 0.00877153780311346, + 0.003690871177241206, + 0.004094922915101051, + 0.0063353292644023895, + 0.00534254452213645, + 0.004987536463886499, + 0.0028500868938863277, + 0.0027451557107269764, + 0.004637703765183687, + 0.0032633638475090265, + 0.0029590025078505278, + 0.004159078001976013, + 0.0036565163172781467, + 0.004407206084579229, + 0.001947447657585144, + 0.002108615590259433, + 0.0062787183560431, + 0.004085651598870754, + 0.003597223199903965, + 0.0027134581468999386, + 0.002600798150524497, + 0.004820275120437145, + 0.013590758666396141, + 0.013557147234678268, + 0.002368537476286292, + 0.0035869234707206488, + 0.008262368850409985, + 0.03482695296406746, + 0.00303017464466393, + 0.005169653799384832, + 0.004254302475601435, + 0.009407393634319305, + 0.0038945390842854977, + 0.005374472588300705, + 0.007390289567410946, + 0.0060045914724469185, + 0.004901522304862738, + 0.004102116450667381, + 0.008004753850400448, + 0.00763269979506731, + 0.007127434015274048, + 0.009811146184802055, + 0.007373526692390442, + 0.005935202818363905, + 0.006290446035563946, + 0.00915359053760767, + 0.007032881490886211, + 0.007156643550843, + 0.0037194450851529837, + 0.0034487787634134293, + 0.006597052328288555, + 0.005255075637251139, + 0.0035358185414224863, + 0.0039079575799405575, + 0.01550145074725151, + 0.01604841835796833, + 0.011737492866814137, + 0.008797840215265751, + 0.012652344070374966, + 0.013163396157324314, + 0.009989612735807896, + 0.009081034921109676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0019393223337829113 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Inspired by the unique and elegant aesthetic of traditional Japanese scroll paintings, a slow, reverent crane shot begins at a low angle, focusing on the base of a translucent human figure standing against a richly textured, sepia-toned washi paper background. From the vaguely outlined feet of the ethereal figure, a complex system of delicate ink lines begins to glow, illuminating the circulatory system with a brilliant crimson and the nervous system with a brilliant, cool blue. These energy lines, with the delicate calligraphic quality of brushstrokes, flow gracefully upward in a steady, captivating ascent. As the camera steadily rises, it carefully follows the energetic path through the transparent legs, into the torso, and branching into the arms. The figure itself is defined not by a solid form but by a complex, luminous internal map of life force, contrasting starkly and beautifully against the minimalist ink background. The deliberate upward journey culminates when the camera reaches the head, where the red and blue lines converge into a dazzlingly complex network, illuminating the brain with a final, brilliant pulse of energy, creating a profoundly spiritual and meditative atmosphere, visualizing the ascending life force.", + "seed": 24571, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_14/index_2335_seed_24571_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03051130778621882, + "vlm_logprob": -0.03051130778621882, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.03051130778621882, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1284947693347931 + ], + "opposite_logprob_tokens": [ + -2.1284947395324707 + ], + "logprob_reward": -0.1284947693347931, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1284947693347931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4050554633140564, + 0.21588268876075745, + 0.055185891687870026, + 0.09073891490697861, + 0.030239084735512733, + 0.03323005884885788, + 0.03441399708390236, + 0.02797006443142891, + 0.07670151442289352, + 0.030582260340452194 + ], + "credit_map": [ + 0.04224846512079239, + 0.020278966054320335, + 0.11883899569511414, + 0.053821586072444916, + 0.01552528329193592, + 0.026548460125923157, + 0.02992476150393486, + 0.09786896407604218, + 0.015538000501692295, + 0.0034793298691511154, + 0.0821724534034729, + 0.00936434231698513, + 0.004394992254674435, + 0.013498242013156414, + 0.005427639931440353, + 0.08200770616531372, + 0.008123882114887238, + 0.0038243825547397137, + 0.012193349190056324, + 0.004142201505601406, + 0.004515547771006823, + 0.008304651826620102, + 0.0048124054446816444, + 0.009269476868212223, + 0.01905381679534912, + 0.010513954795897007, + 0.00913690309971571, + 0.0065120491199195385, + 0.012023797258734703, + 0.01962384767830372, + 0.005020631477236748, + 0.008853928185999393, + 0.005033312365412712, + 0.003002712270244956, + 0.004458180628716946, + 0.0032381247729063034, + 0.002635495737195015, + 0.004119378048926592, + 0.003399680135771632, + 0.004352204035967588, + 0.0072233714163303375, + 0.006794297602027655, + 0.0039613451808691025, + 0.002568986965343356, + 0.003311012638732791, + 0.003483339911326766, + 0.0028439711313694715, + 0.003043738193809986, + 0.003644486889243126, + 0.005184549372643232, + 0.004705770406872034, + 0.004900758154690266, + 0.004330489784479141, + 0.004140871576964855, + 0.0037298656534403563, + 0.0037772059440612793, + 0.0032077766954898834, + 0.002588308649137616, + 0.0031139091588556767, + 0.003878312185406685, + 0.0030454162042587996, + 0.0033705481328070164, + 0.00435875728726387, + 0.004407040774822235, + 0.026091933250427246, + 0.005021417047828436, + 0.01772121526300907, + 0.002787039615213871, + 0.0026949357707053423, + 0.013093598186969757, + 0.0032650986686348915, + 0.006026286166161299, + 0.00452075432986021, + 0.002829022938385606, + 0.006082408130168915, + 0.0030414164066314697, + 0.0028952420689165592, + 0.0055215987376868725, + 0.002417194191366434, + 0.003274626797065139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.1284947693347931 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007644920144230127 + ], + "opposite_logprob_tokens": [ + -4.882645130157471 + ], + "logprob_reward": -0.007644920144230127, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007644920144230127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2516719698905945, + 0.02824946492910385, + 0.02943245694041252, + 0.13536614179611206, + 0.04698289558291435, + 0.034064602106809616, + 0.32427841424942017, + 0.0646364763379097, + 0.038125719875097275, + 0.04719182476401329 + ], + "credit_map": [ + 0.019951321184635162, + 0.010241871699690819, + 0.0732269287109375, + 0.02351577952504158, + 0.007536947261542082, + 0.019408579915761948, + 0.021281998604536057, + 0.0765085443854332, + 0.0033985497429966927, + 0.0022413781844079494, + 0.005210991017520428, + 0.00402797618880868, + 0.0021367408335208893, + 0.0035450195427984, + 0.003270128509029746, + 0.004418679513037205, + 0.004425837658345699, + 0.00261226505972445, + 0.0035140919499099255, + 0.003162203822284937, + 0.005042726639658213, + 0.004437238909304142, + 0.0028065326623618603, + 0.0034315609373152256, + 0.026110785081982613, + 0.016466494649648666, + 0.015420673415064812, + 0.00991182029247284, + 0.014726282097399235, + 0.0201350636780262, + 0.009007204324007034, + 0.023587819188833237, + 0.004944517742842436, + 0.003513317322358489, + 0.00885822344571352, + 0.00431860750541091, + 0.005616704002022743, + 0.006309471093118191, + 0.004611695185303688, + 0.008810357190668583, + 0.008402066305279732, + 0.011517568491399288, + 0.0032204855233430862, + 0.0016930076526477933, + 0.0030723782256245613, + 0.0019292819779366255, + 0.001801881822757423, + 0.0024279351346194744, + 0.03500652313232422, + 0.04117526486515999, + 0.03727782890200615, + 0.03679974004626274, + 0.05962822958827019, + 0.05150757357478142, + 0.02994569018483162, + 0.032937582582235336, + 0.007626431528478861, + 0.00592321902513504, + 0.006467905361205339, + 0.009110841900110245, + 0.007473152130842209, + 0.008475832641124725, + 0.009768242947757244, + 0.009790853597223759, + 0.005891248118132353, + 0.002481865230947733, + 0.008787626400589943, + 0.0026975416112691164, + 0.002607629867270589, + 0.007811630144715309, + 0.004797838162630796, + 0.003050339175388217, + 0.006141879595816135, + 0.005530338268727064, + 0.00917759072035551, + 0.004791451618075371, + 0.004627236630767584, + 0.006554648745805025, + 0.004637205507606268, + 0.005731473211199045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.007644920144230127 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001942891743965447 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.001942891743965447, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001942891743965447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10563741624355316, + 0.03612597659230232, + 0.030647579580545425, + 0.41596683859825134, + 0.02246083691716194, + 0.019502563402056694, + 0.0898432508111, + 0.022201556712388992, + 0.20485277473926544, + 0.0527612529695034 + ], + "credit_map": [ + 0.006539870984852314, + 0.005433525424450636, + 0.024697475135326385, + 0.012194603681564331, + 0.005203119944781065, + 0.00959221925586462, + 0.01271645724773407, + 0.029260143637657166, + 0.0022685853764414787, + 0.0029450596775859594, + 0.009087915532290936, + 0.0038767654914408922, + 0.0026564989238977432, + 0.0024550745729357004, + 0.003038941416889429, + 0.009797138161957264, + 0.004044932313263416, + 0.003257475094869733, + 0.005050111096352339, + 0.002638603560626507, + 0.0048061260022223, + 0.003933269996196032, + 0.003460401203483343, + 0.0034566582180559635, + 0.10320528596639633, + 0.041665658354759216, + 0.042377226054668427, + 0.027282465249300003, + 0.046715378761291504, + 0.07466484606266022, + 0.022839508950710297, + 0.05721646174788475, + 0.002677470911294222, + 0.0019640817772597075, + 0.0030729256104677916, + 0.002631737384945154, + 0.002263524103909731, + 0.0027928792405873537, + 0.003253211500123143, + 0.0038050052244216204, + 0.002259753178805113, + 0.004045740235596895, + 0.0026743707712739706, + 0.0017969845794141293, + 0.002411650726571679, + 0.0020951975602656603, + 0.002132358029484749, + 0.0020865085534751415, + 0.011065511964261532, + 0.011861102655529976, + 0.01408405601978302, + 0.009972370229661465, + 0.015510671772062778, + 0.00951800774782896, + 0.009690409526228905, + 0.008141126483678818, + 0.0024095589760690928, + 0.003087840275838971, + 0.0022515368182212114, + 0.003219103440642357, + 0.004525824915617704, + 0.0017253401456400752, + 0.002471992978826165, + 0.002510359510779381, + 0.05007088556885719, + 0.011442653834819794, + 0.021120397374033928, + 0.01397792249917984, + 0.009822987020015717, + 0.061896272003650665, + 0.0177264753729105, + 0.01879516802728176, + 0.007902558892965317, + 0.005510724149644375, + 0.008333135396242142, + 0.004033019300550222, + 0.005234984681010246, + 0.008617728017270565, + 0.004149328451603651, + 0.008979777805507183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.001942891743965447 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535706162452698 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535706162452698, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535706162452698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10961852967739105, + 0.10580174624919891, + 0.03533834218978882, + 0.36520543694496155, + 0.03948172181844711, + 0.07689981907606125, + 0.10092480480670929, + 0.059241849929094315, + 0.08036895096302032, + 0.027118774130940437 + ], + "credit_map": [ + 0.008672604337334633, + 0.00910257175564766, + 0.024529991671442986, + 0.019336776807904243, + 0.008442492224276066, + 0.007611248642206192, + 0.01078268326818943, + 0.02114015817642212, + 0.0075133247300982475, + 0.00482534896582365, + 0.03244040906429291, + 0.01032464299350977, + 0.004196998197585344, + 0.007271660957485437, + 0.006021025124937296, + 0.03320832923054695, + 0.0036675273440778255, + 0.0035201236605644226, + 0.006054329685866833, + 0.004167879931628704, + 0.006982624530792236, + 0.003994917497038841, + 0.0034943250939249992, + 0.003456615610048175, + 0.05214424431324005, + 0.05162140354514122, + 0.04764334857463837, + 0.03131350129842758, + 0.05633304640650749, + 0.0602991059422493, + 0.027697861194610596, + 0.03815295174717903, + 0.004051665309816599, + 0.0028905069921165705, + 0.0074953604489564896, + 0.004142030142247677, + 0.005090534687042236, + 0.0039038825780153275, + 0.003538916353136301, + 0.008368827402591705, + 0.01881161332130432, + 0.02404763735830784, + 0.007250712253153324, + 0.0038226793985813856, + 0.007425060961395502, + 0.0047269645147025585, + 0.004433762282133102, + 0.0063813915476202965, + 0.012236198410391808, + 0.014388710260391235, + 0.014035687781870365, + 0.011034020222723484, + 0.016900692135095596, + 0.011850670911371708, + 0.009084649384021759, + 0.011394178494811058, + 0.006734869908541441, + 0.006888832431286573, + 0.006183020770549774, + 0.00762291019782424, + 0.007651729043573141, + 0.007082320749759674, + 0.008816336281597614, + 0.008261832408607006, + 0.020629286766052246, + 0.004937382880598307, + 0.008305654861032963, + 0.003692585276439786, + 0.004789524246007204, + 0.0260172039270401, + 0.0044618346728384495, + 0.007535479962825775, + 0.004154134076088667, + 0.003051181323826313, + 0.004363471642136574, + 0.0026144732255488634, + 0.0032702218741178513, + 0.0045118690468370914, + 0.002499764086678624, + 0.0026536581572145224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.012535706162452698 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a richly textured, sepia-toned washi paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the human figure translucent with vaguely outlined feet and no solid form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do glowing ink lines illuminate the circulatory system in crimson and the nervous system in cool blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin at a low angle and steadily rise upward following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the red and blue lines converge into a complex network illuminating the brain at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019382515456527472 + ], + "opposite_logprob_tokens": [ + -6.251938343048096 + ], + "logprob_reward": -0.0019382515456527472, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019382515456527472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06110252067446709, + 0.038970865309238434, + 0.027613012120127678, + 0.2753915786743164, + 0.04860778898000717, + 0.1597704291343689, + 0.08693893998861313, + 0.056729260832071304, + 0.11512506753206253, + 0.12975049018859863 + ], + "credit_map": [ + 0.007723255082964897, + 0.004896435420960188, + 0.011913366615772247, + 0.008372659794986248, + 0.006388308480381966, + 0.007101996336132288, + 0.006555435713380575, + 0.008151073008775711, + 0.0051338388584554195, + 0.004275121726095676, + 0.0049752299673855305, + 0.0063333939760923386, + 0.004532619379460812, + 0.004295498598366976, + 0.005249864421784878, + 0.004175302106887102, + 0.0032490859739482403, + 0.0026684445329010487, + 0.005606404971331358, + 0.0029110824689269066, + 0.004141944460570812, + 0.003288757288828492, + 0.0024885244201868773, + 0.0032587689347565174, + 0.0559510663151741, + 0.02931782603263855, + 0.03545556217432022, + 0.015828389674425125, + 0.029050685465335846, + 0.035186152905225754, + 0.018212363123893738, + 0.05638954043388367, + 0.004748302046209574, + 0.0035564671270549297, + 0.008470412343740463, + 0.00511432159692049, + 0.00583859346807003, + 0.0048044840805232525, + 0.006327633745968342, + 0.009747582487761974, + 0.03666941449046135, + 0.04388329014182091, + 0.013775089755654335, + 0.009534340351819992, + 0.024088386446237564, + 0.010197962634265423, + 0.012015503831207752, + 0.00960646290332079, + 0.010034807957708836, + 0.009320875629782677, + 0.011045727878808975, + 0.007228604517877102, + 0.010758155025541782, + 0.019497588276863098, + 0.0057799057103693485, + 0.013273282907903194, + 0.0065688504837453365, + 0.0068839555606245995, + 0.007174473721534014, + 0.006536264903843403, + 0.006848557852208614, + 0.007243210915476084, + 0.006505733821541071, + 0.008968215435743332, + 0.025659188628196716, + 0.007164358161389828, + 0.012597057968378067, + 0.010043816640973091, + 0.007193406578153372, + 0.019250279292464256, + 0.01617320440709591, + 0.01704375259578228, + 0.02077270671725273, + 0.01430632546544075, + 0.01970876380801201, + 0.011180135421454906, + 0.0157786812633276, + 0.020885009318590164, + 0.012916389852762222, + 0.014202491380274296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 9 + ], + "selected_reward": -0.0019382515456527472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_15/index_tensor([9958, 9958])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_15/index_tensor([9958, 9958])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b3d00759bf21cb775c4b0da736e647c067817831 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_15/index_tensor([9958, 9958])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme macro shot focuses on the slightly open human mouth, creating a remarkably intimate and clinical image. The lighting is bright and directional, casting sharp, moist highlights across the textured landscape of the tongue's surface and revealing each individual papilla in surreal detail. A small, round, uncoated white pill, its chalky, porous texture clearly visible, is deliberately placed on the center of the tongue with a fingertip that has just been removed from the frame. As the pill touches the moist surface, a subtle, involuntary grimace of disgust ripples through the visible features; the corners of the soft pink lips almost imperceptibly draw downward, and the skin around the mouth tightens in a fleeting expression of pain. The stark, clinical white of the pill contrasts sharply with the deep, organic pinks and reds of the tongue and inner mouth, creating a visually appealing and slightly uncomfortable atmosphere of concentrated sensory discomfort.", + "seed": 5142, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_15/index_9958_seed_5142_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08760576687054708, + "vlm_logprob": -0.08760576687054708, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.08760576687054708, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013925652019679546 + ], + "opposite_logprob_tokens": [ + -5.888925552368164 + ], + "logprob_reward": -0.013925652019679546, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.013925652019679546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08255337923765182, + 0.0914098396897316, + 0.37205249071121216, + 0.08309946954250336, + 0.043447498232126236, + 0.049058184027671814, + 0.06301923096179962, + 0.05198703333735466, + 0.09534092992544174, + 0.06803189963102341 + ], + "credit_map": [ + 0.008927511982619762, + 0.007542303763329983, + 0.008716609328985214, + 0.014279304072260857, + 0.010555700398981571, + 0.009584877640008926, + 0.010887641459703445, + 0.01205943338572979, + 0.007665713783353567, + 0.005738010164350271, + 0.015689166262745857, + 0.030068138614296913, + 0.007002627477049828, + 0.0083925174549222, + 0.007276694756001234, + 0.009576966054737568, + 0.030249791219830513, + 0.017221814021468163, + 0.058664463460445404, + 0.11892259120941162, + 0.027765506878495216, + 0.03130798414349556, + 0.05272350832819939, + 0.03519681841135025, + 0.010744494386017323, + 0.006628209259361029, + 0.009727941825985909, + 0.01308802142739296, + 0.010720925405621529, + 0.008349314332008362, + 0.01157641876488924, + 0.012264142744243145, + 0.004958799108862877, + 0.004329378716647625, + 0.0038652068469673395, + 0.007586629595607519, + 0.007621627300977707, + 0.0052397591061890125, + 0.004225051961839199, + 0.005621044896543026, + 0.007956797257065773, + 0.0046591199934482574, + 0.004490567371249199, + 0.0053611090406775475, + 0.006951407995074987, + 0.006973059847950935, + 0.005409469828009605, + 0.007256651762872934, + 0.006741188000887632, + 0.009207794442772865, + 0.005218627862632275, + 0.007850958034396172, + 0.009501907974481583, + 0.009375646710395813, + 0.00869412999600172, + 0.006428978405892849, + 0.005585918668657541, + 0.007808854803442955, + 0.005429861601442099, + 0.0055876970291137695, + 0.008074428886175156, + 0.006178659852594137, + 0.006599942222237587, + 0.00672167306765914, + 0.01326675247400999, + 0.015668457373976707, + 0.013912556692957878, + 0.00893297791481018, + 0.011577613651752472, + 0.008416492491960526, + 0.01339090708643198, + 0.010175172239542007, + 0.011104357428848743, + 0.010489780455827713, + 0.010729840956628323, + 0.009756244719028473, + 0.005931831430643797, + 0.005461518187075853, + 0.007467364892363548, + 0.007090963888913393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 1 + ], + "selected_reward": -0.013925652019679546 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026373015716671944 + ], + "opposite_logprob_tokens": [ + -3.6513729095458984 + ], + "logprob_reward": -0.026373015716671944, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026373015716671944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06005096808075905, + 0.05251026153564453, + 0.3585842549800873, + 0.3052344024181366, + 0.045001763850450516, + 0.033000148832798004, + 0.021926026791334152, + 0.025353308767080307, + 0.03788624331355095, + 0.06045264005661011 + ], + "credit_map": [ + 0.006359195336699486, + 0.007260913960635662, + 0.005638863891363144, + 0.00712992949411273, + 0.011056645773351192, + 0.008581072092056274, + 0.005030472297221422, + 0.008993875235319138, + 0.004470215644687414, + 0.0027989656664431095, + 0.008087633177638054, + 0.016224589198827744, + 0.004584026522934437, + 0.005431501194834709, + 0.005299203563481569, + 0.0056141274981200695, + 0.029720744118094444, + 0.01696833036839962, + 0.054954469203948975, + 0.10866105556488037, + 0.028746118769049644, + 0.03223629295825958, + 0.05208389833569527, + 0.03521335870027542, + 0.04529563710093498, + 0.02547849342226982, + 0.03219562768936157, + 0.04190291091799736, + 0.037209004163742065, + 0.0404704324901104, + 0.035744283348321915, + 0.04693802818655968, + 0.006496873684227467, + 0.004054917488247156, + 0.003790169022977352, + 0.004173282068222761, + 0.008739284239709377, + 0.006435036193579435, + 0.00417223758995533, + 0.007139964494854212, + 0.0028984840027987957, + 0.004034929443150759, + 0.0026436506304889917, + 0.004135704133659601, + 0.00697366101667285, + 0.005559687037020922, + 0.0038161948323249817, + 0.0029378358740359545, + 0.002428588690236211, + 0.002919100923463702, + 0.0027033588849008083, + 0.0027252223808318377, + 0.0031376953702419996, + 0.002522652503103018, + 0.002586476970463991, + 0.00290293013677001, + 0.002635628217831254, + 0.004006527829915285, + 0.002639155136421323, + 0.003082666313275695, + 0.0032656921539455652, + 0.0034043414052575827, + 0.003542402060702443, + 0.0027768949512392282, + 0.0038071973249316216, + 0.007870914414525032, + 0.004266330040991306, + 0.0038390501867979765, + 0.004979902412742376, + 0.004057709127664566, + 0.0054758344776928425, + 0.0035893095191568136, + 0.00793301872909069, + 0.010338684543967247, + 0.008240715600550175, + 0.006663513835519552, + 0.009396877139806747, + 0.00429026922211051, + 0.007561543490737677, + 0.006028014700859785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.026373015716671944 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794747456908226 + ], + "opposite_logprob_tokens": [ + -4.63479471206665 + ], + "logprob_reward": -0.009794747456908226, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794747456908226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15837806463241577, + 0.05811479315161705, + 0.2808351516723633, + 0.0955217182636261, + 0.04949111491441727, + 0.059232234954833984, + 0.03453387692570686, + 0.041939329355955124, + 0.07432867586612701, + 0.14762510359287262 + ], + "credit_map": [ + 0.01370343565940857, + 0.019133558496832848, + 0.013130624778568745, + 0.01964450068771839, + 0.02846956066787243, + 0.018291397020220757, + 0.020571939647197723, + 0.025433041155338287, + 0.005314566660672426, + 0.004285247530788183, + 0.007905837148427963, + 0.01536076981574297, + 0.005360763520002365, + 0.006320815533399582, + 0.0070024817250669, + 0.006564313545823097, + 0.024189874529838562, + 0.01413088571280241, + 0.04067347198724747, + 0.08020158112049103, + 0.02496662363409996, + 0.026721926406025887, + 0.042026586830616, + 0.027924207970499992, + 0.009917552582919598, + 0.009679346345365047, + 0.008569171652197838, + 0.01186784915626049, + 0.014883332885801792, + 0.014034493826329708, + 0.015522236935794353, + 0.011047731153666973, + 0.007202635053545237, + 0.005542967934161425, + 0.0034213464241474867, + 0.004947203677147627, + 0.009151552803814411, + 0.005989783443510532, + 0.006440829485654831, + 0.00679479306563735, + 0.006216250825673342, + 0.004999155178666115, + 0.004809477366507053, + 0.0076075769029557705, + 0.011298423632979393, + 0.011761419475078583, + 0.006087655201554298, + 0.006452277768403292, + 0.0038977826479822397, + 0.004696028772741556, + 0.0033876432571560144, + 0.004320322535932064, + 0.004280631896108389, + 0.005265033338218927, + 0.0036389832384884357, + 0.005047449842095375, + 0.005246732383966446, + 0.006526268552988768, + 0.004750135820358992, + 0.005529969930648804, + 0.0051128072664141655, + 0.00568007118999958, + 0.005135313142091036, + 0.0039580343291163445, + 0.011186812072992325, + 0.012079889886081219, + 0.010627761483192444, + 0.007842197082936764, + 0.010743949562311172, + 0.005726325325667858, + 0.008249478414654732, + 0.007872259244322777, + 0.0192323736846447, + 0.02485569193959236, + 0.028766943141818047, + 0.01721826381981373, + 0.017826320603489876, + 0.011653348803520203, + 0.016588136553764343, + 0.01148401852697134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.009794747456908226 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38711273670196533 + ], + "opposite_logprob_tokens": [ + -1.1371127367019653 + ], + "logprob_reward": -0.38711273670196533, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38711273670196533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06759630143642426, + 0.1110568642616272, + 0.12347237020730972, + 0.12265711277723312, + 0.05491602048277855, + 0.14151330292224884, + 0.03618907928466797, + 0.07298111170530319, + 0.07323525846004486, + 0.1963825672864914 + ], + "credit_map": [ + 0.008996092714369297, + 0.00716758705675602, + 0.006778467912226915, + 0.007732364349067211, + 0.009733866900205612, + 0.009861506521701813, + 0.006884224712848663, + 0.010442190803587437, + 0.007593852933496237, + 0.005966172553598881, + 0.019270343706011772, + 0.042156368494033813, + 0.0062823835760355, + 0.009916911832988262, + 0.010089153423905373, + 0.009781680069863796, + 0.009317776188254356, + 0.007703086361289024, + 0.01570577174425125, + 0.024690845981240273, + 0.021364690735936165, + 0.014018623158335686, + 0.018709426745772362, + 0.011962149292230606, + 0.021013041958212852, + 0.008051330223679543, + 0.013528920710086823, + 0.015883278101682663, + 0.012681220658123493, + 0.0118423942476511, + 0.01467872690409422, + 0.02497820183634758, + 0.008765767328441143, + 0.0049506304785609245, + 0.005255029071122408, + 0.006753781344741583, + 0.00938440952450037, + 0.00706113874912262, + 0.00617353618144989, + 0.0065717254765331745, + 0.012412612326443195, + 0.01343587227165699, + 0.010221929289400578, + 0.02006221003830433, + 0.023972010239958763, + 0.030973266810178757, + 0.014469861052930355, + 0.015965551137924194, + 0.004140338860452175, + 0.004826890770345926, + 0.0035396385937929153, + 0.003608187660574913, + 0.006086824461817741, + 0.004676180891692638, + 0.004484393633902073, + 0.004826626740396023, + 0.007490086834877729, + 0.011279016733169556, + 0.009125572629272938, + 0.009831661358475685, + 0.009731465019285679, + 0.008026906289160252, + 0.01002028863877058, + 0.007476120255887508, + 0.008260644972324371, + 0.009668477810919285, + 0.010220707394182682, + 0.008432388305664062, + 0.012638605199754238, + 0.0067252907902002335, + 0.008395043201744556, + 0.00889410451054573, + 0.022185901179909706, + 0.03410525619983673, + 0.031994130462408066, + 0.024187538772821426, + 0.019930774345993996, + 0.012223334051668644, + 0.037268903106451035, + 0.014486725442111492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.38711273670196533 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008226824575103819 + ], + "opposite_logprob_tokens": [ + -7.1258225440979 + ], + "logprob_reward": -0.0008226824575103819, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008226824575103819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07788082212209702, + 0.09018878638744354, + 0.34519514441490173, + 0.1719878613948822, + 0.04731616750359535, + 0.03618781268596649, + 0.03896639868617058, + 0.05100702494382858, + 0.056273430585861206, + 0.08499657362699509 + ], + "credit_map": [ + 0.012027079239487648, + 0.004943130537867546, + 0.00793746393173933, + 0.01117435097694397, + 0.009838533587753773, + 0.010419503785669804, + 0.006422864738851786, + 0.015117894858121872, + 0.009299451485276222, + 0.005494231823831797, + 0.014477483928203583, + 0.02912592701613903, + 0.00529443146660924, + 0.008191595785319805, + 0.007044471334666014, + 0.011261185631155968, + 0.02911672741174698, + 0.01654510758817196, + 0.053446605801582336, + 0.10559042543172836, + 0.026449766010046005, + 0.03074651025235653, + 0.04909204691648483, + 0.03420792520046234, + 0.026346459984779358, + 0.012628174386918545, + 0.019766459241509438, + 0.026297686621546745, + 0.01962141878902912, + 0.019538408145308495, + 0.019345281645655632, + 0.028443973511457443, + 0.006719235330820084, + 0.004322227090597153, + 0.004930471070110798, + 0.009040490724146366, + 0.0058012790977954865, + 0.00561500946059823, + 0.005034590605646372, + 0.005852863658219576, + 0.003874578746035695, + 0.004417398478835821, + 0.0030854500364512205, + 0.004511300008744001, + 0.0052245305851101875, + 0.006188130006194115, + 0.00477246381342411, + 0.0041139558888971806, + 0.00420341594144702, + 0.006072048097848892, + 0.004172696731984615, + 0.004378878977149725, + 0.005472666118294001, + 0.004928603768348694, + 0.0037864840123802423, + 0.005951603874564171, + 0.006842745002359152, + 0.009396657347679138, + 0.004864165559411049, + 0.005691677797585726, + 0.008338423445820808, + 0.0056157042272388935, + 0.005359578877687454, + 0.004898069426417351, + 0.007224063389003277, + 0.007710568606853485, + 0.005775402300059795, + 0.006849758327007294, + 0.010126507841050625, + 0.005616249982267618, + 0.0061185965314507484, + 0.006852282676845789, + 0.010781915858387947, + 0.013535205274820328, + 0.011587010696530342, + 0.008976908400654793, + 0.011613944545388222, + 0.006881156470626593, + 0.012830852530896664, + 0.008789568208158016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.0008226824575103819 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot focuses on the slightly open human mouth, creating a remarkably intimate and clinical image. The lighting is bright and directional, casting sharp, moist highlights across the textured landscape of the tongue's surface and revealing each individual papilla in surreal detail. A small, round, uncoated white pill, its chalky, porous texture clearly visible, is deliberately placed on the center of the tongue with a fingertip that has just been removed from the frame. As the pill touches the moist surface, a subtle, involuntary grimace of disgust ripples through the visible features; the corners of the soft pink lips almost imperceptibly draw downward, and the skin around the mouth tightens in a fleeting expression of pain. The stark, clinical white of the pill contrasts sharply with the deep, organic pinks and reds of the tongue and inner mouth, creating a visually appealing and slightly uncomfortable atmosphere of concentrated sensory discomfort.", + "seed": 5143, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_15/index_9958_seed_5143_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05694766560336575, + "vlm_logprob": -0.05694766560336575, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.05694766560336575, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009946072474122047 + ], + "opposite_logprob_tokens": [ + -5.759945869445801 + ], + "logprob_reward": -0.009946072474122047, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.009946072474122047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1468626856803894, + 0.047528594732284546, + 0.05253036320209503, + 0.07974445819854736, + 0.06969816982746124, + 0.06532005220651627, + 0.09267360717058182, + 0.13276584446430206, + 0.11550307273864746, + 0.19737325608730316 + ], + "credit_map": [ + 0.020066410303115845, + 0.01686626486480236, + 0.018514618277549744, + 0.0245322547852993, + 0.01851651445031166, + 0.016860971227288246, + 0.01582919992506504, + 0.015676429495215416, + 0.004837205167859793, + 0.004474136978387833, + 0.007002843543887138, + 0.01067439466714859, + 0.005846246611326933, + 0.004459274932742119, + 0.005536054261028767, + 0.004698434378951788, + 0.00511713745072484, + 0.006533160340040922, + 0.004895675927400589, + 0.006364301312714815, + 0.011858499608933926, + 0.006692847702652216, + 0.005779573228210211, + 0.005289160180836916, + 0.012312143109738827, + 0.009944208897650242, + 0.007738964632153511, + 0.009120549075305462, + 0.017109498381614685, + 0.009539003483951092, + 0.006546025164425373, + 0.007434062194079161, + 0.012302854098379612, + 0.011615484952926636, + 0.008136370219290257, + 0.010655902326107025, + 0.008610893040895462, + 0.0069967033341526985, + 0.005202084314078093, + 0.006177871022373438, + 0.005829998757690191, + 0.011917476542294025, + 0.006495968904346228, + 0.009763291105628014, + 0.012309796176850796, + 0.0073303841054439545, + 0.005660793278366327, + 0.006012337282299995, + 0.009507673792541027, + 0.009632476605474949, + 0.011381939984858036, + 0.015589039772748947, + 0.013285933062434196, + 0.00932877603918314, + 0.010769489221274853, + 0.01317827682942152, + 0.015818968415260315, + 0.018260661512613297, + 0.01765579730272293, + 0.02252410538494587, + 0.020769603550434113, + 0.015064452774822712, + 0.010198856703937054, + 0.01247336994856596, + 0.010951114818453789, + 0.01073004212230444, + 0.014053111895918846, + 0.015378543175756931, + 0.011329117231070995, + 0.01999092474579811, + 0.011882521212100983, + 0.021187689155340195, + 0.026777658611536026, + 0.028104597702622414, + 0.028859324753284454, + 0.02959642931818962, + 0.017423415556550026, + 0.02058563381433487, + 0.027012143284082413, + 0.0190140288323164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.009946072474122047 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608525961637497 + ], + "opposite_logprob_tokens": [ + -4.141085147857666 + ], + "logprob_reward": -0.01608525961637497, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608525961637497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1169188991189003, + 0.0654803216457367, + 0.19490490853786469, + 0.06285817921161652, + 0.04586898908019066, + 0.07385964691638947, + 0.05407246574759483, + 0.073670893907547, + 0.11633718013763428, + 0.19602850079536438 + ], + "credit_map": [ + 0.012909895740449429, + 0.012053648009896278, + 0.01393693033605814, + 0.015077869407832623, + 0.023660726845264435, + 0.01383434422314167, + 0.011914033442735672, + 0.013531448319554329, + 0.007216886151582003, + 0.006703323684632778, + 0.009834006428718567, + 0.012697316706180573, + 0.007596702314913273, + 0.006858300883322954, + 0.008347755298018456, + 0.0062260315753519535, + 0.01950974576175213, + 0.01800973154604435, + 0.020927198231220245, + 0.019399648532271385, + 0.05230318009853363, + 0.025349095463752747, + 0.02294825203716755, + 0.01645806059241295, + 0.006459204945713282, + 0.005549210589379072, + 0.004994827322661877, + 0.006956383585929871, + 0.020302798599004745, + 0.008070866577327251, + 0.005118669010698795, + 0.005406217183917761, + 0.004791653249412775, + 0.004122089594602585, + 0.004461437463760376, + 0.006817436311393976, + 0.007712486200034618, + 0.005582277663052082, + 0.004351470153778791, + 0.008030139841139317, + 0.006577317137271166, + 0.010189029388129711, + 0.006821685936301947, + 0.008379139937460423, + 0.017681540921330452, + 0.009109375067055225, + 0.006731998175382614, + 0.008369562216103077, + 0.0050280774012207985, + 0.005726515781134367, + 0.006842862349003553, + 0.00858909823000431, + 0.006634593941271305, + 0.005101681686937809, + 0.008220762014389038, + 0.007928877137601376, + 0.0066774701699614525, + 0.008634604513645172, + 0.008833501487970352, + 0.01137031801044941, + 0.014309386722743511, + 0.009016825817525387, + 0.006507841404527426, + 0.008320946246385574, + 0.014776811003684998, + 0.011719875037670135, + 0.013348473235964775, + 0.013518239371478558, + 0.01150444708764553, + 0.017051219940185547, + 0.016974790021777153, + 0.017443325370550156, + 0.015175674110651016, + 0.029022183269262314, + 0.03003215417265892, + 0.036461587995290756, + 0.015854772180318832, + 0.022708920761942863, + 0.018835626542568207, + 0.027937566861510277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.01608525961637497 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005958176217973232 + ], + "opposite_logprob_tokens": [ + -5.130958080291748 + ], + "logprob_reward": -0.005958176217973232, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958176217973232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1021609902381897, + 0.08112647384405136, + 0.12164095044136047, + 0.05467211827635765, + 0.0668751522898674, + 0.041808225214481354, + 0.1729365438222885, + 0.12490738928318024, + 0.13465560972690582, + 0.09921656548976898 + ], + "credit_map": [ + 0.009865648113191128, + 0.010910422541201115, + 0.01184541080147028, + 0.014271094463765621, + 0.014529863372445107, + 0.015161920338869095, + 0.010168095119297504, + 0.015408532693982124, + 0.011699964292347431, + 0.009467785246670246, + 0.01014718972146511, + 0.010693054646253586, + 0.008140957914292812, + 0.010159499943256378, + 0.007695559877902269, + 0.013122454285621643, + 0.0180220864713192, + 0.014278064481914043, + 0.013045142404735088, + 0.0134327607229352, + 0.021953098475933075, + 0.015520738437771797, + 0.011800846084952354, + 0.0135881956666708, + 0.009099663235247135, + 0.006330345757305622, + 0.005993044003844261, + 0.008163763210177422, + 0.007776797283440828, + 0.006015710532665253, + 0.005734944716095924, + 0.005557843018323183, + 0.009160451591014862, + 0.008797846734523773, + 0.005892850924283266, + 0.007515239529311657, + 0.008687654510140419, + 0.006975849159061909, + 0.0054999347776174545, + 0.014345319010317326, + 0.0038474288303405046, + 0.007857922464609146, + 0.003602560143917799, + 0.00456651346758008, + 0.00866700243204832, + 0.005319270305335522, + 0.0033391232136636972, + 0.0046083987690508366, + 0.013861050829291344, + 0.01892310194671154, + 0.021962327882647514, + 0.029898004606366158, + 0.022071726620197296, + 0.018822479993104935, + 0.02350069023668766, + 0.02389715611934662, + 0.01640075258910656, + 0.014003412798047066, + 0.019108237698674202, + 0.024016547948122025, + 0.012443074025213718, + 0.01064224448055029, + 0.015581602230668068, + 0.01271151378750801, + 0.014850667677819729, + 0.014226377941668034, + 0.014115488156676292, + 0.015521620400249958, + 0.015680260956287384, + 0.022401554509997368, + 0.015816908329725266, + 0.02204272150993347, + 0.010088111273944378, + 0.011503963731229305, + 0.01333522703498602, + 0.016585933044552803, + 0.00852750800549984, + 0.013150000013411045, + 0.011965082958340645, + 0.014060735702514648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.005958176217973232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25210684537887573 + ], + "opposite_logprob_tokens": [ + -1.5021069049835205 + ], + "logprob_reward": -0.25210684537887573, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25210684537887573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06399859488010406, + 0.06012691557407379, + 0.053070638328790665, + 0.0931147113442421, + 0.07383965700864792, + 0.07031253725290298, + 0.05327983945608139, + 0.08409387618303299, + 0.1639842987060547, + 0.2841789424419403 + ], + "credit_map": [ + 0.007835565134882927, + 0.007261991035193205, + 0.007929726503789425, + 0.008945710025727749, + 0.010096203535795212, + 0.00676962872967124, + 0.00868061650544405, + 0.006479151546955109, + 0.00688478397205472, + 0.0072583709843456745, + 0.007789751049131155, + 0.010207919403910637, + 0.00649558799341321, + 0.005906245671212673, + 0.008228668011724949, + 0.007355589885264635, + 0.007581694051623344, + 0.005430885124951601, + 0.0059082466177642345, + 0.006181840319186449, + 0.011518693529069424, + 0.006130666937679052, + 0.004500230774283409, + 0.005818380042910576, + 0.006055580452084541, + 0.006951564457267523, + 0.006892435718327761, + 0.008173942565917969, + 0.03068200871348381, + 0.013155057094991207, + 0.008425885811448097, + 0.012778245843946934, + 0.010646963492035866, + 0.010453823953866959, + 0.0069552455097436905, + 0.007758833933621645, + 0.009782067500054836, + 0.007521090097725391, + 0.007060646079480648, + 0.013660984113812447, + 0.006768363527953625, + 0.011660376563668251, + 0.008262847550213337, + 0.01103868056088686, + 0.010102624073624611, + 0.00863795354962349, + 0.006069608498364687, + 0.007772079668939114, + 0.006780311465263367, + 0.0058301156386733055, + 0.006090393755584955, + 0.006960504222661257, + 0.007481180131435394, + 0.007986188866198063, + 0.0042650289833545685, + 0.00788611825555563, + 0.009243703447282314, + 0.012964176945388317, + 0.010077867656946182, + 0.013932725414633751, + 0.011180821806192398, + 0.009287778288125992, + 0.0090448884293437, + 0.008361908607184887, + 0.014926417730748653, + 0.015251594595611095, + 0.016434842720627785, + 0.01675764098763466, + 0.021539101377129555, + 0.029767576605081558, + 0.01701619289815426, + 0.03229093179106712, + 0.02268279902637005, + 0.03906605765223503, + 0.04380479082465172, + 0.07030697911977768, + 0.01968519762158394, + 0.02161063440144062, + 0.0439215786755085, + 0.023100877180695534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.25210684537887573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000641974329482764 + ], + "opposite_logprob_tokens": [ + -7.375641822814941 + ], + "logprob_reward": -0.000641974329482764, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.000641974329482764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1212119534611702, + 0.048848025500774384, + 0.11395633220672607, + 0.06716135144233704, + 0.04240453615784645, + 0.03204454481601715, + 0.11665704846382141, + 0.08196795731782913, + 0.0973944142460823, + 0.2783537805080414 + ], + "credit_map": [ + 0.01847209967672825, + 0.010271966457366943, + 0.014212926849722862, + 0.012667972594499588, + 0.02012215554714203, + 0.014844221994280815, + 0.010485869832336903, + 0.02013475075364113, + 0.008239390328526497, + 0.0049989065155386925, + 0.006821033079177141, + 0.007796004880219698, + 0.003685122122988105, + 0.005407834891229868, + 0.004337604157626629, + 0.007562132086604834, + 0.012163267470896244, + 0.010208703577518463, + 0.010980562306940556, + 0.01181851513683796, + 0.035609398037195206, + 0.011827216483652592, + 0.00953688658773899, + 0.011811800301074982, + 0.008867405354976654, + 0.006364037282764912, + 0.005831024143844843, + 0.00636921264231205, + 0.01718619279563427, + 0.01092519424855709, + 0.005458621773868799, + 0.006159669253975153, + 0.005266345571726561, + 0.006043276283890009, + 0.004495606757700443, + 0.006377116311341524, + 0.003625225042924285, + 0.00315720122307539, + 0.004425242077559233, + 0.009014527313411236, + 0.003162394743412733, + 0.004238042049109936, + 0.0026971122715622187, + 0.003244479186832905, + 0.008974442258477211, + 0.004506136756390333, + 0.0024019074626266956, + 0.002820033812895417, + 0.010153078474104404, + 0.013337072916328907, + 0.01496442686766386, + 0.02099885232746601, + 0.01319327112287283, + 0.012471511960029602, + 0.015819890424609184, + 0.01571894809603691, + 0.008091721683740616, + 0.008950122632086277, + 0.012074540369212627, + 0.016301855444908142, + 0.009020056575536728, + 0.008788705803453922, + 0.008587980642914772, + 0.01015298068523407, + 0.008197343908250332, + 0.015764186158776283, + 0.011255956254899502, + 0.011562949977815151, + 0.016943814232945442, + 0.01359524205327034, + 0.007890069857239723, + 0.01218485925346613, + 0.02057250402867794, + 0.039490632712841034, + 0.04600193724036217, + 0.06680206954479218, + 0.01987306959927082, + 0.022874606773257256, + 0.03804776072502136, + 0.024691203609108925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.000641974329482764 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot focuses on the slightly open human mouth, creating a remarkably intimate and clinical image. The lighting is bright and directional, casting sharp, moist highlights across the textured landscape of the tongue's surface and revealing each individual papilla in surreal detail. A small, round, uncoated white pill, its chalky, porous texture clearly visible, is deliberately placed on the center of the tongue with a fingertip that has just been removed from the frame. As the pill touches the moist surface, a subtle, involuntary grimace of disgust ripples through the visible features; the corners of the soft pink lips almost imperceptibly draw downward, and the skin around the mouth tightens in a fleeting expression of pain. The stark, clinical white of the pill contrasts sharply with the deep, organic pinks and reds of the tongue and inner mouth, creating a visually appealing and slightly uncomfortable atmosphere of concentrated sensory discomfort.", + "seed": 5144, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_15/index_9958_seed_5144_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06961648085853085, + "vlm_logprob": -0.06961648085853085, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.06961648085853085, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01428602822124958 + ], + "opposite_logprob_tokens": [ + -6.389286041259766 + ], + "logprob_reward": -0.01428602822124958, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.01428602822124958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05710623413324356, + 0.03672416880726814, + 0.04940393939614296, + 0.027127645909786224, + 0.025193316861987114, + 0.029823729768395424, + 0.035807859152555466, + 0.3511539697647095, + 0.05369863286614418, + 0.33396047353744507 + ], + "credit_map": [ + 0.005857095122337341, + 0.009766323491930962, + 0.007506181485950947, + 0.008530781604349613, + 0.004023325629532337, + 0.006786473095417023, + 0.006236708257347345, + 0.008399344980716705, + 0.004906128626316786, + 0.0036199772730469704, + 0.004266539588570595, + 0.00459698960185051, + 0.0026025581173598766, + 0.00437880540266633, + 0.0045738983899354935, + 0.007779269013553858, + 0.005341929849237204, + 0.005540359299629927, + 0.006054237950593233, + 0.007273631636053324, + 0.003599210176616907, + 0.005530168768018484, + 0.006260153371840715, + 0.009804245084524155, + 0.002486253622919321, + 0.0031615907792001963, + 0.0030660873744636774, + 0.0035256571136415005, + 0.0037807340268045664, + 0.003214774653315544, + 0.0032776203006505966, + 0.004614928271621466, + 0.002148501807823777, + 0.0029384505469352007, + 0.002576541854068637, + 0.002734786830842495, + 0.0035050027072429657, + 0.0029708861839026213, + 0.0044549559243023396, + 0.0038641912396997213, + 0.0030095702968537807, + 0.00399085832759738, + 0.0032936620991677046, + 0.0041373190470039845, + 0.003899705596268177, + 0.0033417034428566694, + 0.004674906376749277, + 0.0034760064445436, + 0.003907018341124058, + 0.004127936903387308, + 0.0048441700637340546, + 0.0034172795712947845, + 0.007760249078273773, + 0.004023512359708548, + 0.003717346116900444, + 0.004010348580777645, + 0.030151762068271637, + 0.019263451918959618, + 0.0689241886138916, + 0.033300094306468964, + 0.03613719344139099, + 0.037825990468263626, + 0.03569987416267395, + 0.08985141664743423, + 0.009716239757835865, + 0.006294589024037123, + 0.0058126808144152164, + 0.006625004578381777, + 0.004995966795831919, + 0.005231589078903198, + 0.008308066055178642, + 0.0067145004868507385, + 0.028891079127788544, + 0.034224096685647964, + 0.016872037202119827, + 0.10553940385580063, + 0.016281427815556526, + 0.024804163724184036, + 0.02586057223379612, + 0.08148769289255142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.01428602822124958 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011092661879956722 + ], + "opposite_logprob_tokens": [ + -4.511092662811279 + ], + "logprob_reward": -0.011092661879956722, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011092661879956722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07636085897684097, + 0.07485329359769821, + 0.27916455268859863, + 0.05408990755677223, + 0.15924440324306488, + 0.034485429525375366, + 0.04915948584675789, + 0.1393367350101471, + 0.0379340723156929, + 0.0953712984919548 + ], + "credit_map": [ + 0.006730619817972183, + 0.013995225541293621, + 0.007363544777035713, + 0.00956927239894867, + 0.006922160740941763, + 0.010080941952764988, + 0.011091905646026134, + 0.010607191361486912, + 0.008969434536993504, + 0.009371460415422916, + 0.010662663727998734, + 0.00782688707113266, + 0.011165863834321499, + 0.009060594253242016, + 0.009560058824717999, + 0.00823633186519146, + 0.02623201347887516, + 0.03219069540500641, + 0.036531154066324234, + 0.0250041913241148, + 0.04271002113819122, + 0.03677742928266525, + 0.030598865821957588, + 0.04912015423178673, + 0.005255246534943581, + 0.006425748113542795, + 0.006100426893681288, + 0.007895813323557377, + 0.004373521078377962, + 0.005813250783830881, + 0.0068231974728405476, + 0.011402701959013939, + 0.010093332268297672, + 0.01722079887986183, + 0.02126547507941723, + 0.01574135199189186, + 0.023106873035430908, + 0.02266802452504635, + 0.023628458380699158, + 0.025520093739032745, + 0.002491294639185071, + 0.004588457755744457, + 0.0033794355113059282, + 0.00404086010530591, + 0.005605394020676613, + 0.0041405451484024525, + 0.0047198813408613205, + 0.005519562866538763, + 0.005114525090903044, + 0.007403333205729723, + 0.004612436052411795, + 0.004696754273027182, + 0.009815662167966366, + 0.006281067617237568, + 0.006849135272204876, + 0.004386572167277336, + 0.01259638275951147, + 0.014859642833471298, + 0.011275098659098148, + 0.0171157568693161, + 0.025644449517130852, + 0.019944917410612106, + 0.017515582963824272, + 0.020384911447763443, + 0.004172191489487886, + 0.006068387534469366, + 0.0033978354185819626, + 0.005268266424536705, + 0.00508060771971941, + 0.003508985973894596, + 0.00524574238806963, + 0.0051920549012720585, + 0.00821276381611824, + 0.013041679747402668, + 0.004968469962477684, + 0.027113240212202072, + 0.0072079915553331375, + 0.006680717691779137, + 0.007155023980885744, + 0.020991405472159386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.011092661879956722 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008652695454657078 + ], + "opposite_logprob_tokens": [ + -4.758652687072754 + ], + "logprob_reward": -0.008652695454657078, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008652695454657078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0662509873509407, + 0.04587503522634506, + 0.11624853312969208, + 0.08976740390062332, + 0.03332824632525444, + 0.03961309418082237, + 0.03218211606144905, + 0.3322771191596985, + 0.050181880593299866, + 0.1942756474018097 + ], + "credit_map": [ + 0.008314303122460842, + 0.0092545785009861, + 0.00885920412838459, + 0.007051583379507065, + 0.004625945817679167, + 0.007654581218957901, + 0.008703532628715038, + 0.011787247844040394, + 0.005741619039326906, + 0.006376165430992842, + 0.0061128814704716206, + 0.004701609723269939, + 0.0036011184565722942, + 0.005967901088297367, + 0.005106925033032894, + 0.0082668112590909, + 0.017433946952223778, + 0.012785960920155048, + 0.015097844414412975, + 0.011143644340336323, + 0.01038092840462923, + 0.014386479742825031, + 0.010422240942716599, + 0.024597469717264175, + 0.00892714411020279, + 0.012821096926927567, + 0.012083872221410275, + 0.014995316043496132, + 0.007132928352802992, + 0.009629606269299984, + 0.011047117412090302, + 0.013130311854183674, + 0.0031793001107871532, + 0.004160227254033089, + 0.0036746456753462553, + 0.004154745489358902, + 0.00509081594645977, + 0.003609113395214081, + 0.0058593060821294785, + 0.003600089345127344, + 0.002876827958971262, + 0.004027115181088448, + 0.0037970468401908875, + 0.006759224459528923, + 0.004922875203192234, + 0.0057469382882118225, + 0.005609422922134399, + 0.005873636808246374, + 0.0037438631989061832, + 0.0044264947064220905, + 0.0029924814589321613, + 0.0029277207795530558, + 0.006216741632670164, + 0.004258494824171066, + 0.00399906188249588, + 0.0036172529216855764, + 0.03058391623198986, + 0.01701604388654232, + 0.06640124320983887, + 0.034109219908714294, + 0.033748749643564224, + 0.028245016932487488, + 0.0442710779607296, + 0.07790181785821915, + 0.006892939563840628, + 0.007230044808238745, + 0.004471321124583483, + 0.0050721680745482445, + 0.00561766792088747, + 0.007455557119101286, + 0.006600464228540659, + 0.006841710302978754, + 0.017912812530994415, + 0.02946178987622261, + 0.011537252925336361, + 0.050177305936813354, + 0.016314756125211716, + 0.012584359385073185, + 0.016122348606586456, + 0.0401650033891201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.008652695454657078 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31348124146461487 + ], + "opposite_logprob_tokens": [ + -1.3134812116622925 + ], + "logprob_reward": -0.31348124146461487, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31348124146461487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03417245298624039, + 0.020592736080288887, + 0.06160248443484306, + 0.026289666071534157, + 0.048193398863077164, + 0.05420912802219391, + 0.06314331293106079, + 0.060177020728588104, + 0.04811611771583557, + 0.5835036635398865 + ], + "credit_map": [ + 0.003191334428265691, + 0.005110026802867651, + 0.0034358096309006214, + 0.004437974654138088, + 0.0034218640066683292, + 0.005088801961392164, + 0.0037323280703276396, + 0.005754312500357628, + 0.0025762736331671476, + 0.002559660468250513, + 0.00248922873288393, + 0.002285749651491642, + 0.0018189253751188517, + 0.0034459568560123444, + 0.0022977222688496113, + 0.0031192193273454905, + 0.007076161447912455, + 0.005874151363968849, + 0.00792090967297554, + 0.007084499578922987, + 0.005284187383949757, + 0.008113259449601173, + 0.006902390159666538, + 0.01334692258387804, + 0.002427781233564019, + 0.004090338014066219, + 0.0026215589605271816, + 0.003831359324976802, + 0.0021976870484650135, + 0.0026545762084424496, + 0.003734846832230687, + 0.004731518682092428, + 0.004619631916284561, + 0.0059949173592031, + 0.005342195276170969, + 0.007292281370609999, + 0.0037686359137296677, + 0.005248623434454203, + 0.009258059784770012, + 0.006669055204838514, + 0.003941051196306944, + 0.005657150875777006, + 0.005874615162611008, + 0.008705547079443932, + 0.005984473042190075, + 0.009692590683698654, + 0.006963638123124838, + 0.007390064187347889, + 0.0067223492078483105, + 0.009413938969373703, + 0.007140763569623232, + 0.00682029128074646, + 0.011945502832531929, + 0.005930277053266764, + 0.006583754904568195, + 0.008586427196860313, + 0.005002876278012991, + 0.006470934487879276, + 0.006533799692988396, + 0.007986328564584255, + 0.007874800823628902, + 0.0070159705355763435, + 0.01008179783821106, + 0.009210511110723019, + 0.005431837402284145, + 0.005958395544439554, + 0.005480116233229637, + 0.0050220550037920475, + 0.006767421029508114, + 0.006779166869819164, + 0.006422868464142084, + 0.006254257168620825, + 0.04696159437298775, + 0.07261542975902557, + 0.028682898730039597, + 0.17486539483070374, + 0.0357041209936142, + 0.04037034511566162, + 0.04934389516711235, + 0.13495996594429016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.31348124146461487 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005697772721759975 + ], + "opposite_logprob_tokens": [ + -7.500569820404053 + ], + "logprob_reward": -0.0005697772721759975, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005697772721759975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13992130756378174, + 0.048067279160022736, + 0.15272650122642517, + 0.04549941420555115, + 0.05979614704847336, + 0.04342382401227951, + 0.06490118056535721, + 0.17090365290641785, + 0.06945120543241501, + 0.2053094208240509 + ], + "credit_map": [ + 0.010538185015320778, + 0.022590529173612595, + 0.015695806592702866, + 0.018783366307616234, + 0.015285689383745193, + 0.016176017001271248, + 0.01920895092189312, + 0.02164277620613575, + 0.004402074031531811, + 0.006557504180818796, + 0.006194374989718199, + 0.006398608908057213, + 0.004566417075693607, + 0.00604680459946394, + 0.005609533749520779, + 0.008291961625218391, + 0.01790934056043625, + 0.013064748607575893, + 0.020634042099118233, + 0.016574354842305183, + 0.01269015297293663, + 0.020137935876846313, + 0.016672860831022263, + 0.03504307568073273, + 0.004572940990328789, + 0.005256724543869495, + 0.005521201528608799, + 0.00715238694101572, + 0.003972032573074102, + 0.004676301963627338, + 0.006513022352010012, + 0.00783480703830719, + 0.004243005532771349, + 0.007069180719554424, + 0.007130635902285576, + 0.007065108045935631, + 0.008280610665678978, + 0.00737985223531723, + 0.00972356740385294, + 0.008904194459319115, + 0.0036244718357920647, + 0.005346190650016069, + 0.004389594309031963, + 0.007295764982700348, + 0.00602262606844306, + 0.005045836325734854, + 0.006785614416003227, + 0.004913724958896637, + 0.006596930790692568, + 0.00802427064627409, + 0.0060928682796657085, + 0.0068320948630571365, + 0.012808924540877342, + 0.009988841600716114, + 0.007923888973891735, + 0.006633361801505089, + 0.01442309282720089, + 0.016147414222359657, + 0.02059653215110302, + 0.016184693202376366, + 0.02506597712635994, + 0.017805790528655052, + 0.027038564905524254, + 0.03364160656929016, + 0.007844245992600918, + 0.01192677766084671, + 0.005630241706967354, + 0.008287324570119381, + 0.008943047374486923, + 0.009061899036169052, + 0.00914494413882494, + 0.00861272867769003, + 0.01843922957777977, + 0.034454114735126495, + 0.011680335737764835, + 0.049400586634874344, + 0.021000437438488007, + 0.013060826808214188, + 0.016112569719552994, + 0.041161321103572845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.0005697772721759975 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot focuses on the slightly open human mouth, creating a remarkably intimate and clinical image. The lighting is bright and directional, casting sharp, moist highlights across the textured landscape of the tongue's surface and revealing each individual papilla in surreal detail. A small, round, uncoated white pill, its chalky, porous texture clearly visible, is deliberately placed on the center of the tongue with a fingertip that has just been removed from the frame. As the pill touches the moist surface, a subtle, involuntary grimace of disgust ripples through the visible features; the corners of the soft pink lips almost imperceptibly draw downward, and the skin around the mouth tightens in a fleeting expression of pain. The stark, clinical white of the pill contrasts sharply with the deep, organic pinks and reds of the tongue and inner mouth, creating a visually appealing and slightly uncomfortable atmosphere of concentrated sensory discomfort.", + "seed": 5145, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_15/index_9958_seed_5145_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0964028435992077, + "vlm_logprob": -0.0964028435992077, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.0964028435992077, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01575111411511898 + ], + "opposite_logprob_tokens": [ + -5.140750885009766 + ], + "logprob_reward": -0.01575111411511898, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01575111411511898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2113897055387497, + 0.038274332880973816, + 0.07017819583415985, + 0.1623872071504593, + 0.12877139449119568, + 0.09509290754795074, + 0.07670049369335175, + 0.09232953190803528, + 0.07893967628479004, + 0.045936618000268936 + ], + "credit_map": [ + 0.03168733790516853, + 0.02205568365752697, + 0.022429324686527252, + 0.03017592243850231, + 0.024386249482631683, + 0.03248589485883713, + 0.020148877054452896, + 0.02802041545510292, + 0.005398687440901995, + 0.003078364534303546, + 0.004208799917250872, + 0.004198497626930475, + 0.004351085051894188, + 0.006513455882668495, + 0.005349399987608194, + 0.0051760426722466946, + 0.012646381743252277, + 0.006033642217516899, + 0.008611028082668781, + 0.007746923249214888, + 0.00867573544383049, + 0.009732616133987904, + 0.007882808335125446, + 0.008849059231579304, + 0.02067628689110279, + 0.011765182949602604, + 0.02543564885854721, + 0.03218581900000572, + 0.016048120334744453, + 0.010679059661924839, + 0.02416754513978958, + 0.02142953872680664, + 0.014320739544928074, + 0.014069734141230583, + 0.020613567903637886, + 0.021828122437000275, + 0.01269130315631628, + 0.010366647504270077, + 0.016664763912558556, + 0.01821652241051197, + 0.015421787276864052, + 0.009123130701482296, + 0.016222937032580376, + 0.008871596306562424, + 0.007650002371519804, + 0.008557160384953022, + 0.008786763064563274, + 0.020459527149796486, + 0.007997512817382812, + 0.00748757878318429, + 0.013089499436318874, + 0.010549024678766727, + 0.008870312012732029, + 0.005384940188378096, + 0.007963184267282486, + 0.01535844337195158, + 0.007822138257324696, + 0.005345896352082491, + 0.014320354908704758, + 0.009631279855966568, + 0.006991447415202856, + 0.005304953549057245, + 0.025859728455543518, + 0.017053736373782158, + 0.007117385044693947, + 0.00958603248000145, + 0.0113399438560009, + 0.010806087404489517, + 0.00939302146434784, + 0.006027220282703638, + 0.011789069510996342, + 0.012880916707217693, + 0.0040338896214962006, + 0.00408642552793026, + 0.0067302570678293705, + 0.00806382019072771, + 0.004503021016716957, + 0.003253613132983446, + 0.006523745134472847, + 0.00874184537678957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.01575111411511898 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609042100608349 + ], + "opposite_logprob_tokens": [ + -4.141090393066406 + ], + "logprob_reward": -0.01609042100608349, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609042100608349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15518295764923096, + 0.08957904577255249, + 0.1352989375591278, + 0.22460715472698212, + 0.052094560116529465, + 0.08207472413778305, + 0.06286243349313736, + 0.03716389089822769, + 0.11950331181287766, + 0.041632991284132004 + ], + "credit_map": [ + 0.021430673077702522, + 0.01635267771780491, + 0.016072915866971016, + 0.015949690714478493, + 0.024167699739336967, + 0.021209795027971268, + 0.014357000589370728, + 0.02564249373972416, + 0.013387310318648815, + 0.008278082124888897, + 0.010464427061378956, + 0.012385695241391659, + 0.013381683267652988, + 0.00755487522110343, + 0.01121679786592722, + 0.012910178862512112, + 0.018677912652492523, + 0.006072915159165859, + 0.02181393653154373, + 0.010834968648850918, + 0.013880434446036816, + 0.010627387091517448, + 0.012582043185830116, + 0.040809329599142075, + 0.029096122831106186, + 0.01465741079300642, + 0.03618175908923149, + 0.04371614009141922, + 0.023693451657891273, + 0.014539241790771484, + 0.03200382739305496, + 0.030719196423888206, + 0.00736062228679657, + 0.005134128965437412, + 0.007524602115154266, + 0.007571992464363575, + 0.004909326322376728, + 0.0041488343849778175, + 0.006040386855602264, + 0.009404663927853107, + 0.010490195825695992, + 0.006186373997479677, + 0.01605028472840786, + 0.010569220408797264, + 0.00531639251857996, + 0.005159914027899504, + 0.009511345066130161, + 0.018790997564792633, + 0.007604789454489946, + 0.005968281999230385, + 0.009255524724721909, + 0.009595386683940887, + 0.007248186971992254, + 0.004621356725692749, + 0.006131620611995459, + 0.012437280267477036, + 0.002692205598577857, + 0.0028420051094144583, + 0.005531103815883398, + 0.004073777701705694, + 0.004392153117805719, + 0.0026986522134393454, + 0.009002097882330418, + 0.005931892432272434, + 0.010542670264840126, + 0.017553580924868584, + 0.015861907973885536, + 0.01489832904189825, + 0.016697432845830917, + 0.01320086233317852, + 0.014273413456976414, + 0.01647510938346386, + 0.0038271539378911257, + 0.004375366494059563, + 0.006522580981254578, + 0.006029663607478142, + 0.004207686521112919, + 0.003054742468520999, + 0.004995478317141533, + 0.008620317094027996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.01609042100608349 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020599473267793655 + ], + "opposite_logprob_tokens": [ + -3.895599365234375 + ], + "logprob_reward": -0.020599473267793655, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020599473267793655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21685396134853363, + 0.06639760732650757, + 0.10936292260885239, + 0.12958210706710815, + 0.05721700191497803, + 0.13082173466682434, + 0.09154002368450165, + 0.07198450714349747, + 0.0856364369392395, + 0.04060373827815056 + ], + "credit_map": [ + 0.04296185448765755, + 0.019558686763048172, + 0.02613450400531292, + 0.02211630530655384, + 0.02258015051484108, + 0.03363306447863579, + 0.020113099366426468, + 0.029756294563412666, + 0.008776891976594925, + 0.005964124575257301, + 0.006595739163458347, + 0.010547096841037273, + 0.008334871381521225, + 0.008131809532642365, + 0.010777153074741364, + 0.007269921712577343, + 0.014566872268915176, + 0.00681340554729104, + 0.017506495118141174, + 0.007233723998069763, + 0.012523511424660683, + 0.010313097387552261, + 0.009105095639824867, + 0.03130071610212326, + 0.011570747010409832, + 0.010787254199385643, + 0.020934028550982475, + 0.025283394381403923, + 0.012834270484745502, + 0.011107627302408218, + 0.02046356163918972, + 0.016601217910647392, + 0.0075659556314349174, + 0.007428399752825499, + 0.008290190249681473, + 0.008177101612091064, + 0.006217004265636206, + 0.004815877415239811, + 0.006982075050473213, + 0.007740399334579706, + 0.019839081913232803, + 0.008697851561009884, + 0.023604772984981537, + 0.012167246080935001, + 0.011248989962041378, + 0.009498299099504948, + 0.014543084427714348, + 0.03122241050004959, + 0.005860539618879557, + 0.008786329999566078, + 0.012666304595768452, + 0.010103551670908928, + 0.021729957312345505, + 0.009332007728517056, + 0.012272228486835957, + 0.010789104737341404, + 0.005203180015087128, + 0.006731496658176184, + 0.010511711239814758, + 0.007845399901270866, + 0.008601830340921879, + 0.0052748145535588264, + 0.017189305275678635, + 0.010626774281263351, + 0.006075276993215084, + 0.012261948548257351, + 0.011189164593815804, + 0.01163795217871666, + 0.01222901325672865, + 0.00827556848526001, + 0.011715850792825222, + 0.012251663021743298, + 0.0036170766688883305, + 0.0038951158057898283, + 0.005573117174208164, + 0.006378404330462217, + 0.0053968350403010845, + 0.003677833592519164, + 0.006040818989276886, + 0.006024536676704884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.020599473267793655 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289296865463257 + ], + "opposite_logprob_tokens": [ + -1.0539296865463257 + ], + "logprob_reward": -0.4289296865463257, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289296865463257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09887859970331192, + 0.030937690287828445, + 0.07054467499256134, + 0.05522187799215317, + 0.0627482533454895, + 0.07238362729549408, + 0.1507277637720108, + 0.08298774063587189, + 0.28752371668815613, + 0.0880461037158966 + ], + "credit_map": [ + 0.013227631337940693, + 0.007466843351721764, + 0.01098522450774908, + 0.013792821206152439, + 0.01584828644990921, + 0.014209141954779625, + 0.009106849320232868, + 0.014241799712181091, + 0.00411982461810112, + 0.0023726546205580235, + 0.0035192915238440037, + 0.003297016490250826, + 0.003523987252265215, + 0.003613624954596162, + 0.0049051325768232346, + 0.005586159881204367, + 0.010977081023156643, + 0.005765287671238184, + 0.00900462456047535, + 0.009831327944993973, + 0.01215384528040886, + 0.006809710059314966, + 0.007455725222826004, + 0.008547072298824787, + 0.008861983194947243, + 0.004702033940702677, + 0.008116361685097218, + 0.006868612952530384, + 0.0055532921105623245, + 0.005389745347201824, + 0.007011577486991882, + 0.008718269877135754, + 0.007822968997061253, + 0.007291568908840418, + 0.009749329648911953, + 0.008494771085679531, + 0.006108142901211977, + 0.0067379167303442955, + 0.007959495298564434, + 0.008584057912230492, + 0.008419190533459187, + 0.011428716592490673, + 0.011607681401073933, + 0.009490152820944786, + 0.006365118082612753, + 0.0066691734828054905, + 0.007476836442947388, + 0.010926755145192146, + 0.009307023137807846, + 0.014734388329088688, + 0.02012006752192974, + 0.016122447326779366, + 0.034019362181425095, + 0.01674066111445427, + 0.02275761216878891, + 0.016926202923059464, + 0.011066563427448273, + 0.011493614874780178, + 0.012621761299669743, + 0.010192492976784706, + 0.0077742040157318115, + 0.0072207972407341, + 0.009688079357147217, + 0.01293022558093071, + 0.02246035262942314, + 0.03998248651623726, + 0.038323670625686646, + 0.04197743535041809, + 0.03829287737607956, + 0.0265160221606493, + 0.04011928290128708, + 0.039851609617471695, + 0.00763305788859725, + 0.009476393461227417, + 0.012955456972122192, + 0.016216358169913292, + 0.00985701009631157, + 0.007713937200605869, + 0.00931401178240776, + 0.014879883266985416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.4289296865463257 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, extreme macro view focused on a slightly open human mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small, round, uncoated white pill with a chalky, porous texture visible on the center of the tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting cast sharp, moist highlights revealing individual papillae on the tongue's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the corners of the soft pink lips draw downward in a subtle grimace of disgust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a visible contrast between the stark white pill and the deep organic pinks and reds of the mouth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006435230607166886 + ], + "opposite_logprob_tokens": [ + -7.375643730163574 + ], + "logprob_reward": -0.0006435230607166886, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006435230607166886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18329326808452606, + 0.08505798131227493, + 0.15914274752140045, + 0.11979953199625015, + 0.12324568629264832, + 0.033612173050642014, + 0.060346443206071854, + 0.06535105407238007, + 0.11619312316179276, + 0.05395801365375519 + ], + "credit_map": [ + 0.03443670645356178, + 0.015277615748345852, + 0.022394103929400444, + 0.019929662346839905, + 0.024258144199848175, + 0.027636684477329254, + 0.017465148121118546, + 0.021895188838243484, + 0.008048002608120441, + 0.009055586531758308, + 0.007613780442625284, + 0.01616295427083969, + 0.014522525481879711, + 0.008271767757833004, + 0.012730763293802738, + 0.008652599528431892, + 0.022632068023085594, + 0.008010241203010082, + 0.024910615757107735, + 0.016803322359919548, + 0.013223538175225258, + 0.01249645370990038, + 0.0153930289670825, + 0.04567348584532738, + 0.017661741003394127, + 0.007542258128523827, + 0.01940815895795822, + 0.021792497485876083, + 0.012378537096083164, + 0.008194890804588795, + 0.01671513356268406, + 0.016106318682432175, + 0.02114538662135601, + 0.010208898223936558, + 0.01822821982204914, + 0.015574801713228226, + 0.013793686404824257, + 0.008897051215171814, + 0.01286710798740387, + 0.022530538961291313, + 0.004148891661316156, + 0.005067124031484127, + 0.005491337738931179, + 0.003383383620530367, + 0.003410049481317401, + 0.0031569343991577625, + 0.0040323566645383835, + 0.004922093823552132, + 0.004616871476173401, + 0.006470137275755405, + 0.009241122752428055, + 0.008279940113425255, + 0.011048718355596066, + 0.00501764053478837, + 0.007190237287431955, + 0.008481774479150772, + 0.004254186991602182, + 0.004901582840830088, + 0.009919523261487484, + 0.0058858441188931465, + 0.006404022686183453, + 0.0038665987085551023, + 0.017624931409955025, + 0.012494365684688091, + 0.008981987833976746, + 0.01657073199748993, + 0.0158846452832222, + 0.017238032072782516, + 0.015086395666003227, + 0.009494496509432793, + 0.01634891889989376, + 0.01658792234957218, + 0.005484175402671099, + 0.006552936974912882, + 0.007992137223482132, + 0.007344658020883799, + 0.006550442427396774, + 0.003978800028562546, + 0.005943397991359234, + 0.010111462324857712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0006435230607166886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_2/index_tensor([2340, 2340])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_2/index_tensor([2340, 2340])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ef5742121f3a5cd43901791f57a28171ce42d928 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_2/index_tensor([2340, 2340])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly zooming drone shot zooms in, focusing on the very center of the frame. Across a cracked, black volcanic landscape shrouded in darkness, molten lava flows with an unstoppable heaviness. The viscous, orange-red torrent sometimes solidifies into a dark, hard crust, then fractures under internal pressure, revealing a more dazzling, liquid core beneath, emitting waves of searing heat. At the center of this sea of inferno, a lone, two-story wooden house, its weathered white walls pale in the flames, is relentlessly devoured by this destructive force. Lava has already overflowed the house's foundations, instantly igniting the wooden planks that touch the walls, reducing them to charred charcoal. The structure groans and tilts under the intense heat. Thick black smoke and sparking ash erupt from shattered windows and cracks in the roof, rising upwards to merge with the night sky, stained a dark red by the volcanic ash. The only lighting in the entire scene comes from the lava itself, its dancing light casting long, dancing shadows on the remaining structure of the house, creating an apocalyptic atmosphere that is both spectacular and terrifying. As the camera continuously and steadily pulls back, the dying house appears increasingly small and helpless in the boundless sea of fire, ultimately revealing a grand and tragic picture of destruction.", + "seed": 11952, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_2/index_2340_seed_11952_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12736086249351503, + "vlm_logprob": -0.12736086249351503, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.12736086249351503, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02202456071972847 + ], + "opposite_logprob_tokens": [ + -5.147024631500244 + ], + "logprob_reward": -0.02202456071972847, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.02202456071972847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.300963431596756, + 0.047631531953811646, + 0.02926539070904255, + 0.04828943684697151, + 0.11314602196216583, + 0.07608812302350998, + 0.1015387624502182, + 0.1793186366558075, + 0.033312879502773285, + 0.07044576108455658 + ], + "credit_map": [ + 0.035365160554647446, + 0.05017527565360069, + 0.0332767479121685, + 0.022652065381407738, + 0.044365279376506805, + 0.041482385247945786, + 0.03386107459664345, + 0.03978542983531952, + 0.005692728329449892, + 0.005691679194569588, + 0.003694620681926608, + 0.0041218409314751625, + 0.007522225379943848, + 0.005699718836694956, + 0.005100496578961611, + 0.010108222253620625, + 0.0046156467869877815, + 0.003581833094358444, + 0.002964332466945052, + 0.003441851120442152, + 0.004269298631697893, + 0.004207942634820938, + 0.0031094844453036785, + 0.0030750033911317587, + 0.008113366551697254, + 0.00486329011619091, + 0.0047762151807546616, + 0.0036927354522049427, + 0.007804316934198141, + 0.0071127102710306644, + 0.00488697923719883, + 0.007039823569357395, + 0.020721565932035446, + 0.016252703964710236, + 0.010929889976978302, + 0.008367533795535564, + 0.012975837104022503, + 0.019017374143004417, + 0.011852019466459751, + 0.013029090128839016, + 0.010237808339297771, + 0.008307603187859058, + 0.007292713969945908, + 0.0058555398136377335, + 0.011488622985780239, + 0.011835471726953983, + 0.012235336005687714, + 0.00883503258228302, + 0.017202289775013924, + 0.013616738840937614, + 0.009387016296386719, + 0.009904451668262482, + 0.014768749475479126, + 0.015892501920461655, + 0.01145491749048233, + 0.009312087669968605, + 0.020259248092770576, + 0.02090454287827015, + 0.020463230088353157, + 0.015345309861004353, + 0.03531017526984215, + 0.02563265711069107, + 0.02364257164299488, + 0.017760906368494034, + 0.0030896649695932865, + 0.0037814269308000803, + 0.0030483065638691187, + 0.004980079829692841, + 0.004753655754029751, + 0.0028023142367601395, + 0.003479637671262026, + 0.007377793081104755, + 0.010944359004497528, + 0.008962278254330158, + 0.004072684794664383, + 0.005822035018354654, + 0.008284385316073895, + 0.018385309725999832, + 0.00522216223180294, + 0.008752545341849327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.02202456071972847 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020604494959115982 + ], + "opposite_logprob_tokens": [ + -3.8956046104431152 + ], + "logprob_reward": -0.020604494959115982, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020604494959115982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23054054379463196, + 0.05955272912979126, + 0.05464712157845497, + 0.10914319008588791, + 0.071937195956707, + 0.09216547757387161, + 0.14355507493019104, + 0.14698193967342377, + 0.03900220990180969, + 0.052474409341812134 + ], + "credit_map": [ + 0.03795187547802925, + 0.021261001005768776, + 0.027580736204981804, + 0.018170885741710663, + 0.03933052718639374, + 0.028398603200912476, + 0.024615973234176636, + 0.03323093801736832, + 0.010351143777370453, + 0.00891825184226036, + 0.0044710757210850716, + 0.005231622606515884, + 0.009478208608925343, + 0.007309828419238329, + 0.006543143652379513, + 0.007249452639371157, + 0.008639274165034294, + 0.00733985286206007, + 0.005618408787995577, + 0.00367585476487875, + 0.008474240079522133, + 0.007670377381145954, + 0.006548543926328421, + 0.006680572871118784, + 0.017407063394784927, + 0.011631499975919724, + 0.010622370056807995, + 0.012583509087562561, + 0.014344933442771435, + 0.014739949256181717, + 0.010331579484045506, + 0.01748228818178177, + 0.009767020121216774, + 0.006975920870900154, + 0.008684715256094933, + 0.005317090544849634, + 0.00810910202562809, + 0.010664340108633041, + 0.007568178232759237, + 0.014850829727947712, + 0.011879886500537395, + 0.010147065855562687, + 0.009856725111603737, + 0.008283027447760105, + 0.008438404649496078, + 0.017615096643567085, + 0.01591477543115616, + 0.010030507110059261, + 0.01883821375668049, + 0.013285443186759949, + 0.01711924560368061, + 0.01279564667493105, + 0.021176699548959732, + 0.023989150300621986, + 0.01844535581767559, + 0.017905330285429955, + 0.01440038625150919, + 0.01642468571662903, + 0.020562104880809784, + 0.01610834337770939, + 0.018986456096172333, + 0.02116721123456955, + 0.021227778866887093, + 0.01810496300458908, + 0.003480377607047558, + 0.005731943529099226, + 0.004326160531491041, + 0.004333566874265671, + 0.006676113698631525, + 0.0035112001933157444, + 0.0040445891208946705, + 0.006898257415741682, + 0.008319049142301083, + 0.0070442627184093, + 0.003864731639623642, + 0.003567451611161232, + 0.008001130074262619, + 0.0070166634395718575, + 0.0040784357115626335, + 0.01058268453925848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.020604494959115982 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07904980331659317 + ], + "opposite_logprob_tokens": [ + -2.579049825668335 + ], + "logprob_reward": -0.07904980331659317, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07904980331659317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3894338309764862, + 0.06205935776233673, + 0.0684845969080925, + 0.15235066413879395, + 0.0898512750864029, + 0.059531792998313904, + 0.0578349269926548, + 0.05046778917312622, + 0.02814088948071003, + 0.04184488579630852 + ], + "credit_map": [ + 0.04190484434366226, + 0.06672558933496475, + 0.04369879141449928, + 0.030199306085705757, + 0.06089870631694794, + 0.053188979625701904, + 0.043992094695568085, + 0.048825543373823166, + 0.008040009997785091, + 0.007259354926645756, + 0.005875455681234598, + 0.005803517531603575, + 0.0099723469465971, + 0.008588683791458607, + 0.006013514939695597, + 0.01050647720694542, + 0.010629404336214066, + 0.008805899880826473, + 0.00468303170055151, + 0.0074588945135474205, + 0.015607769601047039, + 0.008808224461972713, + 0.00800417922437191, + 0.004487198777496815, + 0.02522406354546547, + 0.017904886975884438, + 0.012981553561985493, + 0.01554409135133028, + 0.02350601740181446, + 0.02015715092420578, + 0.013480335474014282, + 0.023552563041448593, + 0.011286209337413311, + 0.00949240755289793, + 0.009830623865127563, + 0.009242641739547253, + 0.014062406495213509, + 0.013907946646213531, + 0.011479232460260391, + 0.01054981630295515, + 0.00927552580833435, + 0.0077504320070147514, + 0.007146570831537247, + 0.005643828306347132, + 0.0051756491884589195, + 0.008532924577593803, + 0.008296660147607327, + 0.007710207253694534, + 0.008627885021269321, + 0.006132558453828096, + 0.006203402765095234, + 0.004672281444072723, + 0.007143080700188875, + 0.010055203922092915, + 0.007630814798176289, + 0.007369700353592634, + 0.0060120318084955215, + 0.0057632713578641415, + 0.006401128601282835, + 0.005174350459128618, + 0.006799047812819481, + 0.008344994857907295, + 0.006702032405883074, + 0.005270936992019415, + 0.0026614749804139137, + 0.004685236141085625, + 0.0032959680538624525, + 0.003656344721093774, + 0.0046285889111459255, + 0.002307827351614833, + 0.0029373676516115665, + 0.003968085162341595, + 0.005571827758103609, + 0.005622373893857002, + 0.0035708181094378233, + 0.002873245859518647, + 0.005620656535029411, + 0.00862833485007286, + 0.004572021774947643, + 0.005385606549680233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.07904980331659317 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31357821822166443 + ], + "opposite_logprob_tokens": [ + -1.3135782480239868 + ], + "logprob_reward": -0.31357821822166443, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31357821822166443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14919669926166534, + 0.11693558096885681, + 0.06748805940151215, + 0.304884672164917, + 0.05003723129630089, + 0.0641750618815422, + 0.05294951796531677, + 0.11499430984258652, + 0.037904638797044754, + 0.04143417626619339 + ], + "credit_map": [ + 0.02110755443572998, + 0.0156578216701746, + 0.02068488486111164, + 0.009606806561350822, + 0.026589006185531616, + 0.01933938078582287, + 0.01221606694161892, + 0.023995181545615196, + 0.017383964732289314, + 0.01611238345503807, + 0.008042379282414913, + 0.009016476571559906, + 0.024926409125328064, + 0.015665367245674133, + 0.009208638221025467, + 0.016579966992139816, + 0.008520970121026039, + 0.008273894898593426, + 0.008095541968941689, + 0.005910604726523161, + 0.012953274883329868, + 0.009886723943054676, + 0.007191086187958717, + 0.006655965466052294, + 0.04453613609075546, + 0.038959842175245285, + 0.025632958859205246, + 0.033383507281541824, + 0.04691919684410095, + 0.03937030956149101, + 0.02947894297540188, + 0.046603813767433167, + 0.008254699409008026, + 0.006130910944193602, + 0.004189648199826479, + 0.003416866296902299, + 0.00956930685788393, + 0.005915860179811716, + 0.004547521937638521, + 0.008012423291802406, + 0.009247875772416592, + 0.009760275483131409, + 0.006601018365472555, + 0.004363168962299824, + 0.008879683911800385, + 0.00783238373696804, + 0.007974967360496521, + 0.009515694342553616, + 0.007423216477036476, + 0.004872083198279142, + 0.006821162533015013, + 0.004984635394066572, + 0.0063501219265162945, + 0.007742315996438265, + 0.005963101051747799, + 0.008792880922555923, + 0.015413441695272923, + 0.011026461608707905, + 0.011823456734418869, + 0.007813292555510998, + 0.027671335265040398, + 0.016651349142193794, + 0.012823851779103279, + 0.011771121993660927, + 0.0037635292392224073, + 0.005271984729915857, + 0.005492150783538818, + 0.006178154144436121, + 0.004038182087242603, + 0.003600724507123232, + 0.004860169254243374, + 0.00469974847510457, + 0.006506791338324547, + 0.006189079489558935, + 0.00270751747302711, + 0.002848045201972127, + 0.005102999974042177, + 0.010702438652515411, + 0.0026046670973300934, + 0.0047726379707455635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.31357821822166443 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20154723525047302 + ], + "opposite_logprob_tokens": [ + -1.7015472650527954 + ], + "logprob_reward": -0.20154723525047302, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20154723525047302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18917231261730194, + 0.05476658418774605, + 0.02346888743340969, + 0.049099914729595184, + 0.053350649774074554, + 0.2150312215089798, + 0.2391519397497177, + 0.13208472728729248, + 0.016616366803646088, + 0.02725744992494583 + ], + "credit_map": [ + 0.02254670485854149, + 0.032468561083078384, + 0.021406054496765137, + 0.014618177898228168, + 0.029734868556261063, + 0.026891477406024933, + 0.02126842550933361, + 0.02023804932832718, + 0.007073211017996073, + 0.0077959634363651276, + 0.005925292614847422, + 0.005182687658816576, + 0.010886424221098423, + 0.0069275121204555035, + 0.0055942474864423275, + 0.0053812419064342976, + 0.0041004689410328865, + 0.0032035040203481913, + 0.0014221016317605972, + 0.0025790189392864704, + 0.003780471393838525, + 0.003566150786355138, + 0.00240401946939528, + 0.0024131538812071085, + 0.0054116882383823395, + 0.00550862867385149, + 0.005374155472964048, + 0.005411411169916391, + 0.006686744745820761, + 0.006113838870078325, + 0.0052155074663460255, + 0.009377941489219666, + 0.008876008912920952, + 0.005839453544467688, + 0.008802562952041626, + 0.004663717932999134, + 0.005807075649499893, + 0.008508840575814247, + 0.006008240859955549, + 0.004844747483730316, + 0.02676469273865223, + 0.02235199511051178, + 0.012749115005135536, + 0.01526712067425251, + 0.02686075121164322, + 0.05477171763777733, + 0.03505047410726547, + 0.021215364336967468, + 0.030420904979109764, + 0.025261858478188515, + 0.03233816474676132, + 0.021677173674106598, + 0.03453550115227699, + 0.03382585197687149, + 0.028967518359422684, + 0.03212495893239975, + 0.01624307408928871, + 0.014793908223509789, + 0.015715762972831726, + 0.014558837749063969, + 0.02047591097652912, + 0.01586310751736164, + 0.016931544989347458, + 0.01750257983803749, + 0.0017069110181182623, + 0.0022604158148169518, + 0.0019792714156210423, + 0.00220458023250103, + 0.0021101252641528845, + 0.001743200933560729, + 0.002283004578202963, + 0.0023288570810109377, + 0.003258072305470705, + 0.004116641357541084, + 0.0016289182240143418, + 0.001852712593972683, + 0.0035101566463708878, + 0.00712598254904151, + 0.0018128978554159403, + 0.003952068742364645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.20154723525047302 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly zooming drone shot zooms in, focusing on the very center of the frame. Across a cracked, black volcanic landscape shrouded in darkness, molten lava flows with an unstoppable heaviness. The viscous, orange-red torrent sometimes solidifies into a dark, hard crust, then fractures under internal pressure, revealing a more dazzling, liquid core beneath, emitting waves of searing heat. At the center of this sea of inferno, a lone, two-story wooden house, its weathered white walls pale in the flames, is relentlessly devoured by this destructive force. Lava has already overflowed the house's foundations, instantly igniting the wooden planks that touch the walls, reducing them to charred charcoal. The structure groans and tilts under the intense heat. Thick black smoke and sparking ash erupt from shattered windows and cracks in the roof, rising upwards to merge with the night sky, stained a dark red by the volcanic ash. The only lighting in the entire scene comes from the lava itself, its dancing light casting long, dancing shadows on the remaining structure of the house, creating an apocalyptic atmosphere that is both spectacular and terrifying. As the camera continuously and steadily pulls back, the dying house appears increasingly small and helpless in the boundless sea of fire, ultimately revealing a grand and tragic picture of destruction.", + "seed": 11953, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_2/index_2340_seed_11953_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.28434925973415376, + "vlm_logprob": -0.28434925973415376, + "vlm_yes_no_margin": 1.7, + "vlm_token_credit": -0.28434925973415376, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05255785956978798 + ], + "opposite_logprob_tokens": [ + -3.677557945251465 + ], + "logprob_reward": -0.05255785956978798, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.05255785956978798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15056179463863373, + 0.047887347638607025, + 0.06389091163873672, + 0.16757476329803467, + 0.07117800414562225, + 0.049253158271312714, + 0.11048998683691025, + 0.12400981783866882, + 0.18075570464134216, + 0.034398507326841354 + ], + "credit_map": [ + 0.020168378949165344, + 0.017341015860438347, + 0.016931459307670593, + 0.012921083718538284, + 0.02372864820063114, + 0.015871573239564896, + 0.01426650956273079, + 0.02933313138782978, + 0.008898423984646797, + 0.00697185005992651, + 0.005710380617529154, + 0.004529704339802265, + 0.006437964737415314, + 0.005424906965345144, + 0.004058263264596462, + 0.005855855531990528, + 0.01026410423219204, + 0.014762600883841515, + 0.0070987846702337265, + 0.005525426939129829, + 0.008714355528354645, + 0.006660622078925371, + 0.005518600810319185, + 0.005346419755369425, + 0.02793644182384014, + 0.05222271755337715, + 0.011883999221026897, + 0.012534205801784992, + 0.023421846330165863, + 0.021069340407848358, + 0.007306970655918121, + 0.011199229396879673, + 0.010212846100330353, + 0.011825060471892357, + 0.008692996576428413, + 0.011393196880817413, + 0.008607247844338417, + 0.006934563163667917, + 0.006639809347689152, + 0.006872277706861496, + 0.008895850740373135, + 0.00812959112226963, + 0.004478172864764929, + 0.005255929194390774, + 0.005611145403236151, + 0.00523568969219923, + 0.004278738982975483, + 0.007368039805442095, + 0.025537140667438507, + 0.009265602566301823, + 0.013376959599554539, + 0.008117794059216976, + 0.009663485921919346, + 0.028360141441226006, + 0.006698961835354567, + 0.009469893760979176, + 0.021970154717564583, + 0.017686378210783005, + 0.014292125590145588, + 0.011189297772943974, + 0.013249222189188004, + 0.019190212711691856, + 0.01182631030678749, + 0.014606120996177197, + 0.03061460144817829, + 0.03008480742573738, + 0.022753911092877388, + 0.018774546682834625, + 0.022638777270913124, + 0.017827657982707024, + 0.013659908436238766, + 0.024401497095823288, + 0.004104420077055693, + 0.0072844140231609344, + 0.0028837977442890406, + 0.0028945880476385355, + 0.005845263134688139, + 0.0034384296741336584, + 0.003329745726659894, + 0.00461784889921546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.05255785956978798 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500585958361626 + ], + "opposite_logprob_tokens": [ + -2.9300057888031006 + ], + "logprob_reward": -0.05500585958361626, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500585958361626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2011767476797104, + 0.0755421593785286, + 0.07859617471694946, + 0.1481558084487915, + 0.07979433238506317, + 0.05809587612748146, + 0.14652779698371887, + 0.1073310375213623, + 0.07148855924606323, + 0.0332915373146534 + ], + "credit_map": [ + 0.026135174557566643, + 0.03632007911801338, + 0.022120408713817596, + 0.018530219793319702, + 0.02924356609582901, + 0.02593725547194481, + 0.021118972450494766, + 0.021771054714918137, + 0.011573569849133492, + 0.009832131676375866, + 0.01612965390086174, + 0.008710276335477829, + 0.007894046604633331, + 0.006495453417301178, + 0.0059923334047198296, + 0.008914696052670479, + 0.011588071472942829, + 0.017634116113185883, + 0.009484181180596352, + 0.005958698224276304, + 0.013989515602588654, + 0.00853696558624506, + 0.005272441077977419, + 0.006132187321782112, + 0.02311057038605213, + 0.034158580005168915, + 0.017081022262573242, + 0.015220372937619686, + 0.021529780700802803, + 0.01651156321167946, + 0.009803183376789093, + 0.010740729980170727, + 0.011601222679018974, + 0.014979559928178787, + 0.009257561527192593, + 0.006872209254652262, + 0.011716557666659355, + 0.00928763672709465, + 0.006876562722027302, + 0.009203021414577961, + 0.007973361760377884, + 0.012404941022396088, + 0.006264522206038237, + 0.004956449847668409, + 0.009667162783443928, + 0.007614755537360907, + 0.0050883060321211815, + 0.004126375541090965, + 0.05387398973107338, + 0.008789371699094772, + 0.006148240994662046, + 0.0023660100996494293, + 0.007562866900116205, + 0.057600751519203186, + 0.004114343319088221, + 0.006072223652154207, + 0.017125898972153664, + 0.01921560801565647, + 0.01473002228885889, + 0.008972552604973316, + 0.01168164610862732, + 0.013394379056990147, + 0.010659012012183666, + 0.011551915667951107, + 0.011486921459436417, + 0.010765481740236282, + 0.008502720855176449, + 0.0057836067862808704, + 0.01076204888522625, + 0.0075432718731462955, + 0.007298932410776615, + 0.009345573373138905, + 0.0038122606929391623, + 0.004354179371148348, + 0.003031004685908556, + 0.003282292280346155, + 0.008224361576139927, + 0.003022165969014168, + 0.0038179117254912853, + 0.0037473624106496572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.05500585958361626 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25227460265159607 + ], + "opposite_logprob_tokens": [ + -1.5022746324539185 + ], + "logprob_reward": -0.25227460265159607, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25227460265159607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14035539329051971, + 0.20570974051952362, + 0.13673633337020874, + 0.10364367812871933, + 0.07157690823078156, + 0.07853351533412933, + 0.08163206279277802, + 0.05203633010387421, + 0.11467593908309937, + 0.015100067481398582 + ], + "credit_map": [ + 0.01836315169930458, + 0.018772834911942482, + 0.020109079778194427, + 0.010234619490802288, + 0.02109770104289055, + 0.017911702394485474, + 0.0113567765802145, + 0.022509530186653137, + 0.045975152403116226, + 0.02956453710794449, + 0.01788116991519928, + 0.012215492315590382, + 0.03700828179717064, + 0.021132731810212135, + 0.010817992500960827, + 0.031114380806684494, + 0.018233522772789, + 0.03151833638548851, + 0.012831348925828934, + 0.011172586120665073, + 0.03451618552207947, + 0.013831683434545994, + 0.007649484556168318, + 0.006983195897191763, + 0.01458369567990303, + 0.020831529051065445, + 0.008701573126018047, + 0.01004820317029953, + 0.021559344604611397, + 0.010847849771380424, + 0.008064791560173035, + 0.009006691165268421, + 0.010261510498821735, + 0.007889955304563046, + 0.007795725017786026, + 0.008044572547078133, + 0.015792913734912872, + 0.00713317422196269, + 0.008144870400428772, + 0.006514189299196005, + 0.012769411318004131, + 0.02271183207631111, + 0.006054597906768322, + 0.0051469625905156136, + 0.010341624729335308, + 0.012431864626705647, + 0.004025234840810299, + 0.0050519853830337524, + 0.025981828570365906, + 0.006245281081646681, + 0.0037029380910098553, + 0.0030393146444112062, + 0.007809388916939497, + 0.027876948937773705, + 0.002892639022320509, + 0.004083716776221991, + 0.007304224651306868, + 0.010818849317729473, + 0.00602344935759902, + 0.0050619542598724365, + 0.006000664085149765, + 0.006891457363963127, + 0.00476178340613842, + 0.005173944868147373, + 0.01893015392124653, + 0.025550147518515587, + 0.012540725991129875, + 0.010744847357273102, + 0.012432782910764217, + 0.012300338596105576, + 0.007080358453094959, + 0.015096589922904968, + 0.0020752258133143187, + 0.0020349062979221344, + 0.0011983372969552875, + 0.001343454234302044, + 0.0032362849451601505, + 0.001869473373517394, + 0.0015853452496230602, + 0.0017570405034348369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.25227460265159607 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4290846288204193 + ], + "opposite_logprob_tokens": [ + -1.0540846586227417 + ], + "logprob_reward": -0.4290846288204193, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4290846288204193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11796458810567856, + 0.15441930294036865, + 0.07295916974544525, + 0.08821620047092438, + 0.1101846992969513, + 0.10207629203796387, + 0.08197904378175735, + 0.06706324219703674, + 0.15394170582294464, + 0.05119578167796135 + ], + "credit_map": [ + 0.016572581604123116, + 0.014504577964544296, + 0.01802798919379711, + 0.011637083254754543, + 0.01710103079676628, + 0.011615242809057236, + 0.011285279877483845, + 0.01722080446779728, + 0.03414555639028549, + 0.028293218463659286, + 0.011097865179181099, + 0.009330873377621174, + 0.025209402665495872, + 0.014938784763216972, + 0.009531054645776749, + 0.021872537210583687, + 0.010067070834338665, + 0.01676003634929657, + 0.007193597033619881, + 0.006633388344198465, + 0.014888632111251354, + 0.005976853892207146, + 0.006448979489505291, + 0.004990613088011742, + 0.011924958787858486, + 0.021759461611509323, + 0.0075813704170286655, + 0.008764076046645641, + 0.011477258056402206, + 0.012254852801561356, + 0.005311537068337202, + 0.009142681024968624, + 0.017076905816793442, + 0.01631242223083973, + 0.014154604636132717, + 0.009035550989210606, + 0.01947326585650444, + 0.012919635511934757, + 0.008090675808489323, + 0.013121634721755981, + 0.016021236777305603, + 0.022611891850829124, + 0.010321271605789661, + 0.009438316337764263, + 0.013969031162559986, + 0.011690898798406124, + 0.008809627965092659, + 0.009214022196829319, + 0.023882249370217323, + 0.009918585419654846, + 0.0051249428652226925, + 0.003893979825079441, + 0.0064804465509951115, + 0.024922586977481842, + 0.0031369251664727926, + 0.0046193283051252365, + 0.01210961863398552, + 0.009515703655779362, + 0.005824447609484196, + 0.0062245712615549564, + 0.007645745761692524, + 0.012143164873123169, + 0.006290903780609369, + 0.007309082895517349, + 0.025867216289043427, + 0.03355026617646217, + 0.01638307236135006, + 0.012693694792687893, + 0.017458273097872734, + 0.016325682401657104, + 0.011385481804609299, + 0.020278025418519974, + 0.007212208118289709, + 0.009249523282051086, + 0.0035272741224616766, + 0.004133511334657669, + 0.007403794210404158, + 0.007936811074614525, + 0.005956659559160471, + 0.0057759955525398254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.4290846288204193 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328233480453491 + ], + "opposite_logprob_tokens": [ + -0.7578233480453491 + ], + "logprob_reward": -0.6328233480453491, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328233480453491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19974836707115173, + 0.12217545509338379, + 0.059994570910930634, + 0.1457747370004654, + 0.06286145001649857, + 0.05150236934423447, + 0.10194086283445358, + 0.06364463269710541, + 0.12735408544540405, + 0.06500351428985596 + ], + "credit_map": [ + 0.021964408457279205, + 0.037473414093256, + 0.02231576293706894, + 0.016251029446721077, + 0.04644877091050148, + 0.022238699719309807, + 0.0149751091375947, + 0.018081165850162506, + 0.024165449663996696, + 0.020164210349321365, + 0.01097926776856184, + 0.006854631006717682, + 0.028106030076742172, + 0.012013806030154228, + 0.006290772929787636, + 0.013601276092231274, + 0.008112158626317978, + 0.01125064305961132, + 0.004771850537508726, + 0.003777910489588976, + 0.019361840561032295, + 0.0048589324578642845, + 0.003625513520091772, + 0.004235721193253994, + 0.01964404247701168, + 0.03535156697034836, + 0.014117999002337456, + 0.012215359136462212, + 0.020990746095776558, + 0.0193153265863657, + 0.008740400895476341, + 0.01539929024875164, + 0.008475800976157188, + 0.011431305669248104, + 0.0069231754168868065, + 0.006546975113451481, + 0.008419028483331203, + 0.008419759571552277, + 0.005316868890076876, + 0.007328537292778492, + 0.007569178473204374, + 0.01128784567117691, + 0.005070465616881847, + 0.00380568765103817, + 0.006755697540938854, + 0.009171659126877785, + 0.003990175202488899, + 0.0038516647182404995, + 0.01938704214990139, + 0.018074506893754005, + 0.011032531037926674, + 0.006662089843302965, + 0.011615044437348843, + 0.02139444462954998, + 0.006358055863529444, + 0.007417145650833845, + 0.009330520406365395, + 0.009462404996156693, + 0.010251701809465885, + 0.005371754989027977, + 0.007587563246488571, + 0.008182439021766186, + 0.005796578712761402, + 0.007661672309041023, + 0.020126644521951675, + 0.022075634449720383, + 0.0162021666765213, + 0.011256248690187931, + 0.016591781750321388, + 0.012199828401207924, + 0.012042839080095291, + 0.016858939081430435, + 0.008139198645949364, + 0.009417071007192135, + 0.008180994540452957, + 0.007490741088986397, + 0.009979129768908024, + 0.005890172906219959, + 0.007925501093268394, + 0.007980708964169025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.6328233480453491 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly zooming drone shot zooms in, focusing on the very center of the frame. Across a cracked, black volcanic landscape shrouded in darkness, molten lava flows with an unstoppable heaviness. The viscous, orange-red torrent sometimes solidifies into a dark, hard crust, then fractures under internal pressure, revealing a more dazzling, liquid core beneath, emitting waves of searing heat. At the center of this sea of inferno, a lone, two-story wooden house, its weathered white walls pale in the flames, is relentlessly devoured by this destructive force. Lava has already overflowed the house's foundations, instantly igniting the wooden planks that touch the walls, reducing them to charred charcoal. The structure groans and tilts under the intense heat. Thick black smoke and sparking ash erupt from shattered windows and cracks in the roof, rising upwards to merge with the night sky, stained a dark red by the volcanic ash. The only lighting in the entire scene comes from the lava itself, its dancing light casting long, dancing shadows on the remaining structure of the house, creating an apocalyptic atmosphere that is both spectacular and terrifying. As the camera continuously and steadily pulls back, the dying house appears increasingly small and helpless in the boundless sea of fire, ultimately revealing a grand and tragic picture of destruction.", + "seed": 11954, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_2/index_2340_seed_11954_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2518379457294941, + "vlm_logprob": -0.2518379457294941, + "vlm_yes_no_margin": 2.025, + "vlm_token_credit": -0.2518379457294941, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023663487285375595 + ], + "opposite_logprob_tokens": [ + -4.523663520812988 + ], + "logprob_reward": -0.023663487285375595, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.023663487285375595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23616017401218414, + 0.17988263070583344, + 0.09735145419836044, + 0.08205460011959076, + 0.06862224638462067, + 0.060799386352300644, + 0.09267177432775497, + 0.07054460048675537, + 0.07182459533214569, + 0.04008852317929268 + ], + "credit_map": [ + 0.031221823766827583, + 0.022587241604924202, + 0.03430863097310066, + 0.025590816512703896, + 0.02329246699810028, + 0.02663389965891838, + 0.02528650127351284, + 0.047238789498806, + 0.030915630981326103, + 0.010864131152629852, + 0.0693112388253212, + 0.01149804424494505, + 0.012659545987844467, + 0.011392001062631607, + 0.00804225541651249, + 0.025199776515364647, + 0.008592426776885986, + 0.014668505638837814, + 0.01143241673707962, + 0.014861511066555977, + 0.018605263903737068, + 0.009500323794782162, + 0.007829491049051285, + 0.011861521750688553, + 0.016745086759328842, + 0.008135033771395683, + 0.007380455732345581, + 0.006341019179672003, + 0.008187839761376381, + 0.020915118977427483, + 0.007207468152046204, + 0.0071425847709178925, + 0.009250881150364876, + 0.009007466956973076, + 0.012040061876177788, + 0.007898975163698196, + 0.0058440109714865685, + 0.008432525210082531, + 0.009498492814600468, + 0.0066498336382210255, + 0.008515268564224243, + 0.007033718284219503, + 0.00841619074344635, + 0.006335246376693249, + 0.009022517129778862, + 0.0057798330672085285, + 0.007605551276355982, + 0.008091062307357788, + 0.011871044524013996, + 0.013402442447841167, + 0.010838931426405907, + 0.009925156831741333, + 0.011346322484314442, + 0.0092412568628788, + 0.013149851001799107, + 0.012896766886115074, + 0.011839761398732662, + 0.00697307288646698, + 0.006758799310773611, + 0.007761909160763025, + 0.01090866606682539, + 0.010909153148531914, + 0.008813714608550072, + 0.00657952344045043, + 0.008857285603880882, + 0.008244248107075691, + 0.00830073468387127, + 0.008150971494615078, + 0.00967427622526884, + 0.008668445982038975, + 0.013297860510647297, + 0.006630768533796072, + 0.006120017264038324, + 0.003920715302228928, + 0.005479259416460991, + 0.004626926500350237, + 0.0052275084890425205, + 0.006656750105321407, + 0.003679879242554307, + 0.004377467092126608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.023663487285375595 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867555946111679 + ], + "opposite_logprob_tokens": [ + -3.048675537109375 + ], + "logprob_reward": -0.04867555946111679, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867555946111679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11197195947170258, + 0.1359189748764038, + 0.1477021723985672, + 0.07069596648216248, + 0.11198253184556961, + 0.09708710014820099, + 0.0731433779001236, + 0.09040273725986481, + 0.06638125330209732, + 0.09471388906240463 + ], + "credit_map": [ + 0.01678697019815445, + 0.01666009984910488, + 0.01306756678968668, + 0.012456079013645649, + 0.012879635207355022, + 0.013270940631628036, + 0.011396028101444244, + 0.015454648993909359, + 0.015012141317129135, + 0.017945043742656708, + 0.026745671406388283, + 0.008586048148572445, + 0.029498044401407242, + 0.017091497778892517, + 0.009625254198908806, + 0.011415287852287292, + 0.011651664040982723, + 0.02627253718674183, + 0.009909864515066147, + 0.020222606137394905, + 0.03665183484554291, + 0.010739968158304691, + 0.007130193058401346, + 0.025123512372374535, + 0.008067387156188488, + 0.011749579571187496, + 0.007765637245029211, + 0.007819565013051033, + 0.012612348422408104, + 0.008891178295016289, + 0.00746844382956624, + 0.006321831606328487, + 0.01329067349433899, + 0.01601393334567547, + 0.011912892572581768, + 0.021001409739255905, + 0.012184510938823223, + 0.014325909316539764, + 0.013125869445502758, + 0.01012734230607748, + 0.012007148936390877, + 0.011310772970318794, + 0.013015576638281345, + 0.014371027238667011, + 0.01308143325150013, + 0.00976272951811552, + 0.010849099606275558, + 0.012689315713942051, + 0.007974175736308098, + 0.009021002799272537, + 0.009671706706285477, + 0.009924055077135563, + 0.008471415378153324, + 0.008323578163981438, + 0.009457845240831375, + 0.010299604386091232, + 0.016071593388915062, + 0.009168596006929874, + 0.008238323032855988, + 0.013128729537129402, + 0.009249736554920673, + 0.014436113648116589, + 0.011857310310006142, + 0.008252344094216824, + 0.011883502826094627, + 0.007363462820649147, + 0.009658042341470718, + 0.007011737208813429, + 0.00872934702783823, + 0.009839450009167194, + 0.005924808327108622, + 0.005970904603600502, + 0.0126457205042243, + 0.015028201043605804, + 0.009073077701032162, + 0.010108129121363163, + 0.014997454360127449, + 0.014543754979968071, + 0.010509723797440529, + 0.007807838264852762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.04867555946111679 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799524873495102 + ], + "opposite_logprob_tokens": [ + -1.8049525022506714 + ], + "logprob_reward": -0.1799524873495102, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799524873495102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0944710373878479, + 0.0997469574213028, + 0.05839165672659874, + 0.35492101311683655, + 0.05599137023091316, + 0.15795783698558807, + 0.06097731366753578, + 0.0683494359254837, + 0.023264534771442413, + 0.025928834453225136 + ], + "credit_map": [ + 0.012020159512758255, + 0.009010055102407932, + 0.014455463737249374, + 0.009808845818042755, + 0.010095426812767982, + 0.010828562080860138, + 0.00979047641158104, + 0.018462050706148148, + 0.013906539417803288, + 0.006152474787086248, + 0.03819621726870537, + 0.00793391838669777, + 0.0071562957018613815, + 0.006350787356495857, + 0.004215216264128685, + 0.015835510566830635, + 0.005218997597694397, + 0.008367368020117283, + 0.005166889168322086, + 0.007029244210571051, + 0.012251622043550014, + 0.0053715333342552185, + 0.004313515964895487, + 0.010672487318515778, + 0.11521966010332108, + 0.018620608374476433, + 0.01339398417621851, + 0.006522946059703827, + 0.020678715780377388, + 0.13366122543811798, + 0.008647255599498749, + 0.03817658871412277, + 0.007794706150889397, + 0.00803903304040432, + 0.006463699508458376, + 0.006759761832654476, + 0.005797922611236572, + 0.0073900483548641205, + 0.008914663456380367, + 0.004831536673009396, + 0.019836658611893654, + 0.020487461239099503, + 0.01914217509329319, + 0.02481800504028797, + 0.01931157521903515, + 0.015540024265646935, + 0.020164398476481438, + 0.018657535314559937, + 0.007046585436910391, + 0.006368221715092659, + 0.008340009488165379, + 0.00812986958771944, + 0.00629887729883194, + 0.006780577357858419, + 0.009075291454792023, + 0.00893788319081068, + 0.01422099769115448, + 0.006474254187196493, + 0.006353167816996574, + 0.006073821801692247, + 0.008634766563773155, + 0.01220341119915247, + 0.009049884043633938, + 0.00533913541585207, + 0.0035144747234880924, + 0.003144301241263747, + 0.002851658733561635, + 0.003635384375229478, + 0.002676768461242318, + 0.003143581096082926, + 0.002415113616734743, + 0.0018832518253475428, + 0.002978941425681114, + 0.00321006216108799, + 0.0025070125702768564, + 0.0039848871529102325, + 0.003782358719035983, + 0.003585719270631671, + 0.003732675453647971, + 0.002147177467122674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.1799524873495102 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31353166699409485 + ], + "opposite_logprob_tokens": [ + -1.3135316371917725 + ], + "logprob_reward": -0.31353166699409485, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31353166699409485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10141430795192719, + 0.2019166499376297, + 0.08355286717414856, + 0.18769977986812592, + 0.04038689285516739, + 0.17628724873065948, + 0.06375254690647125, + 0.07029170542955399, + 0.03674280270934105, + 0.03795524314045906 + ], + "credit_map": [ + 0.014165694825351238, + 0.00945703312754631, + 0.015501421876251698, + 0.01147452648729086, + 0.009313393384218216, + 0.012870382517576218, + 0.009523414075374603, + 0.01910843886435032, + 0.0300790686160326, + 0.010275918990373611, + 0.08212138712406158, + 0.014699230901896954, + 0.012767646461725235, + 0.012595131993293762, + 0.007800252176821232, + 0.03157800808548927, + 0.00897146761417389, + 0.011792339384555817, + 0.013598301447927952, + 0.011346861720085144, + 0.013420983217656612, + 0.007641446311026812, + 0.008881386369466782, + 0.007900078780949116, + 0.049259137362241745, + 0.017190122976899147, + 0.012185030616819859, + 0.011133338324725628, + 0.016621358692646027, + 0.05559168756008148, + 0.010195043869316578, + 0.015524061396718025, + 0.004450434353202581, + 0.005320199765264988, + 0.003934609238058329, + 0.006454190239310265, + 0.005807948298752308, + 0.004627523943781853, + 0.005877879448235035, + 0.003914107568562031, + 0.025095034390687943, + 0.02387167513370514, + 0.018034640699625015, + 0.02378392592072487, + 0.02503899671137333, + 0.016430767253041267, + 0.02418064884841442, + 0.019851556047797203, + 0.00665568420663476, + 0.00916738249361515, + 0.005786423105746508, + 0.013107317499816418, + 0.010018308646976948, + 0.007201012223958969, + 0.005575464107096195, + 0.006240956485271454, + 0.01322193443775177, + 0.0074612898752093315, + 0.00608708756044507, + 0.0089337807148695, + 0.010653764940798283, + 0.011218172498047352, + 0.007476720027625561, + 0.005238957237452269, + 0.006148125045001507, + 0.0042546275071799755, + 0.004726932849735022, + 0.0043429844081401825, + 0.004566045943647623, + 0.005589823704212904, + 0.003993581049144268, + 0.0031206796411424875, + 0.006599476095288992, + 0.004490198567509651, + 0.003672543913125992, + 0.003659015754237771, + 0.0063360463827848434, + 0.005922478623688221, + 0.00421901885420084, + 0.0030564656481146812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.31353166699409485 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.693366527557373 + ], + "opposite_logprob_tokens": [ + -0.693366527557373 + ], + "logprob_reward": -0.693366527557373, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.693366527557373, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07425285130739212, + 0.27351275086402893, + 0.0554557666182518, + 0.292082279920578, + 0.03460988402366638, + 0.05925677716732025, + 0.08170688152313232, + 0.02910423092544079, + 0.05518839880824089, + 0.044830072671175 + ], + "credit_map": [ + 0.009341119788587093, + 0.009847963228821754, + 0.007652497384697199, + 0.007631841581314802, + 0.00706642447039485, + 0.012327894568443298, + 0.00733850197866559, + 0.013046623207628727, + 0.052993543446063995, + 0.01946737989783287, + 0.09825726598501205, + 0.01566421426832676, + 0.01849033683538437, + 0.01507987454533577, + 0.011493210680782795, + 0.04206697270274162, + 0.006080470513552427, + 0.010786062106490135, + 0.00416834419593215, + 0.006531808525323868, + 0.01489465031772852, + 0.004003304056823254, + 0.003471251344308257, + 0.005519886035472155, + 0.09028153121471405, + 0.018966887146234512, + 0.01908094622194767, + 0.00911931972950697, + 0.01774098351597786, + 0.09781632572412491, + 0.012289231643080711, + 0.026787113398313522, + 0.0032152526546269655, + 0.005569663364440203, + 0.0035724996123462915, + 0.005151315592229366, + 0.005215614568442106, + 0.004357069730758667, + 0.004360614810138941, + 0.00316785741597414, + 0.007643170189112425, + 0.005198119208216667, + 0.005794014316052198, + 0.006874153856188059, + 0.007085457444190979, + 0.005464975722134113, + 0.013216516003012657, + 0.007980383932590485, + 0.006777207832783461, + 0.01279392372816801, + 0.007620708551257849, + 0.019491681829094887, + 0.009908434003591537, + 0.009688595309853554, + 0.00861095916479826, + 0.0068153925240039825, + 0.003530474379658699, + 0.005200606305152178, + 0.002269058721140027, + 0.005591702647507191, + 0.0034219971857964993, + 0.0037533107679337263, + 0.0035013826563954353, + 0.0018357032677158713, + 0.01013909187167883, + 0.007521785795688629, + 0.00557293463498354, + 0.005111624486744404, + 0.005791736301034689, + 0.007670767139643431, + 0.008261959068477154, + 0.005118511151522398, + 0.006024578120559454, + 0.007092244457453489, + 0.004859499633312225, + 0.007356908172369003, + 0.005425920244306326, + 0.006356134545058012, + 0.004148536361753941, + 0.003566259518265724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 6 + ], + "selected_reward": -0.693366527557373 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slowly zooming drone shot zooms in, focusing on the very center of the frame. Across a cracked, black volcanic landscape shrouded in darkness, molten lava flows with an unstoppable heaviness. The viscous, orange-red torrent sometimes solidifies into a dark, hard crust, then fractures under internal pressure, revealing a more dazzling, liquid core beneath, emitting waves of searing heat. At the center of this sea of inferno, a lone, two-story wooden house, its weathered white walls pale in the flames, is relentlessly devoured by this destructive force. Lava has already overflowed the house's foundations, instantly igniting the wooden planks that touch the walls, reducing them to charred charcoal. The structure groans and tilts under the intense heat. Thick black smoke and sparking ash erupt from shattered windows and cracks in the roof, rising upwards to merge with the night sky, stained a dark red by the volcanic ash. The only lighting in the entire scene comes from the lava itself, its dancing light casting long, dancing shadows on the remaining structure of the house, creating an apocalyptic atmosphere that is both spectacular and terrifying. As the camera continuously and steadily pulls back, the dying house appears increasingly small and helpless in the boundless sea of fire, ultimately revealing a grand and tragic picture of destruction.", + "seed": 11955, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_2/index_2340_seed_11955_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.25658756867051125, + "vlm_logprob": -0.25658756867051125, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.25658756867051125, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023998143151402473 + ], + "opposite_logprob_tokens": [ + -4.648998260498047 + ], + "logprob_reward": -0.023998143151402473, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.023998143151402473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09434429556131363, + 0.07781185954809189, + 0.10887023061513901, + 0.09834637492895126, + 0.35264652967453003, + 0.049625497311353683, + 0.06777828931808472, + 0.04292585328221321, + 0.048923902213573456, + 0.058727141469717026 + ], + "credit_map": [ + 0.018413450568914413, + 0.014138204976916313, + 0.010326559655368328, + 0.008570699952542782, + 0.010952065698802471, + 0.014117274433374405, + 0.0082306619733572, + 0.00959538109600544, + 0.011522354558110237, + 0.015484461560845375, + 0.012123696506023407, + 0.006150832399725914, + 0.006875538732856512, + 0.009908116422593594, + 0.0065053487196564674, + 0.009241510182619095, + 0.03110996074974537, + 0.006590775679796934, + 0.013014400377869606, + 0.004792771302163601, + 0.007121978793293238, + 0.030126310884952545, + 0.005535576492547989, + 0.010578448884189129, + 0.013576575554907322, + 0.03213925659656525, + 0.006570940371602774, + 0.00514508131891489, + 0.007194619160145521, + 0.021143175661563873, + 0.005064947530627251, + 0.007511777803301811, + 0.0840030387043953, + 0.022049354389309883, + 0.02646612375974655, + 0.008561281487345695, + 0.017274195328354836, + 0.1462833136320114, + 0.0107500609010458, + 0.03725915402173996, + 0.00976557657122612, + 0.004857409745454788, + 0.0066184671595692635, + 0.006703345105051994, + 0.004698832985013723, + 0.007932734675705433, + 0.00345601886510849, + 0.005593113601207733, + 0.017452502623200417, + 0.006329858675599098, + 0.00435210345312953, + 0.005269220564514399, + 0.005493467673659325, + 0.01755673624575138, + 0.005601958837360144, + 0.005722443573176861, + 0.007224432658404112, + 0.005798073951154947, + 0.003912335727363825, + 0.005073108244687319, + 0.005559154786169529, + 0.005308942403644323, + 0.005184684880077839, + 0.004865123890340328, + 0.005480725318193436, + 0.008607795462012291, + 0.006884059403091669, + 0.006267203949391842, + 0.005349087994545698, + 0.006947454996407032, + 0.00488606933504343, + 0.004501505754888058, + 0.006817438639700413, + 0.009172804653644562, + 0.0050933170132339, + 0.008621365763247013, + 0.00811576098203659, + 0.009613774716854095, + 0.006326048634946346, + 0.00496663386002183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.023998143151402473 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026375802233815193 + ], + "opposite_logprob_tokens": [ + -3.6513757705688477 + ], + "logprob_reward": -0.026375802233815193, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026375802233815193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12294748425483704, + 0.0714256763458252, + 0.04986600950360298, + 0.04388011246919632, + 0.4918457567691803, + 0.025583788752555847, + 0.09366157650947571, + 0.030801916494965553, + 0.036782413721084595, + 0.033205192536115646 + ], + "credit_map": [ + 0.03502355143427849, + 0.010526934638619423, + 0.011096046306192875, + 0.006707597058266401, + 0.007979130372405052, + 0.008511674590408802, + 0.008313622325658798, + 0.034788940101861954, + 0.013183852657675743, + 0.014421754516661167, + 0.007011928129941225, + 0.004827475640922785, + 0.007713582832366228, + 0.007644556928426027, + 0.0061074309051036835, + 0.010515102185308933, + 0.009742779657244682, + 0.0034061260521411896, + 0.0044960565865039825, + 0.002696024486795068, + 0.004375146701931953, + 0.014814524911344051, + 0.003241174388676882, + 0.007094183005392551, + 0.00457519106566906, + 0.011786822229623795, + 0.0034908936358988285, + 0.003393968567252159, + 0.005411633290350437, + 0.007503540255129337, + 0.0030877182725816965, + 0.004630349110811949, + 0.12652859091758728, + 0.030391938984394073, + 0.03377203270792961, + 0.00943120289593935, + 0.024022871628403664, + 0.20466898381710052, + 0.013627242296934128, + 0.049402933567762375, + 0.0048715416342020035, + 0.0034274705685675144, + 0.002556389197707176, + 0.002295192563906312, + 0.004406046122312546, + 0.0030695670284330845, + 0.0024429820477962494, + 0.002514600520953536, + 0.024593500420451164, + 0.00732619222253561, + 0.005857693497091532, + 0.00492433924227953, + 0.013387157581746578, + 0.024378851056098938, + 0.008024091832339764, + 0.005169753450900316, + 0.005129603203386068, + 0.004511735402047634, + 0.0028878694865852594, + 0.0027505659963935614, + 0.004450123757123947, + 0.0044757346622645855, + 0.0037926146760582924, + 0.0028036723379045725, + 0.004778405651450157, + 0.004821533802896738, + 0.005515268538147211, + 0.005169075448065996, + 0.003644788870587945, + 0.004705434199422598, + 0.003695783903822303, + 0.004452127031981945, + 0.006307232193648815, + 0.0037645152769982815, + 0.002435625996440649, + 0.003482737811282277, + 0.002660129452124238, + 0.007806610781699419, + 0.0029461276717483997, + 0.003802213119342923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.026375802233815193 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023329302668571472 + ], + "opposite_logprob_tokens": [ + -3.773329257965088 + ], + "logprob_reward": -0.023329302668571472, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023329302668571472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10656020045280457, + 0.09487837553024292, + 0.13171808421611786, + 0.027647148817777634, + 0.37475350499153137, + 0.030638456344604492, + 0.125125452876091, + 0.04649295285344124, + 0.031424492597579956, + 0.030761301517486572 + ], + "credit_map": [ + 0.02696670964360237, + 0.01012884359806776, + 0.010459034703671932, + 0.008364038541913033, + 0.006633414886891842, + 0.009616248309612274, + 0.008427112363278866, + 0.02596481330692768, + 0.016668148338794708, + 0.019089149311184883, + 0.010415012016892433, + 0.00734406104311347, + 0.010422414168715477, + 0.01007130928337574, + 0.008291461504995823, + 0.012576824985444546, + 0.03160072863101959, + 0.007945680059492588, + 0.01361373532563448, + 0.005604257341474295, + 0.009937670081853867, + 0.03803005814552307, + 0.007824511267244816, + 0.017161451280117035, + 0.003765573026612401, + 0.00409302581101656, + 0.0021901775617152452, + 0.002350650494918227, + 0.004079476464539766, + 0.002725085010752082, + 0.002932571340352297, + 0.005510593298822641, + 0.09040232747793198, + 0.021822843700647354, + 0.028074176982045174, + 0.010468919761478901, + 0.018819836899638176, + 0.1533854603767395, + 0.012459342367947102, + 0.039320606738328934, + 0.006594143342226744, + 0.003594358917325735, + 0.002954855328425765, + 0.002417603973299265, + 0.0033003201242536306, + 0.004430969245731831, + 0.002865718910470605, + 0.004480487201362848, + 0.02179623395204544, + 0.020200299099087715, + 0.01674550399184227, + 0.012798360548913479, + 0.013006411492824554, + 0.01895371824502945, + 0.012003418989479542, + 0.009621514938771725, + 0.0068551525473594666, + 0.005710109136998653, + 0.004074508789926767, + 0.003751924494281411, + 0.007601557765156031, + 0.007357510272413492, + 0.006944654975086451, + 0.004197539296001196, + 0.004372339230030775, + 0.0036925729364156723, + 0.004582889378070831, + 0.0042842780239880085, + 0.0032781255431473255, + 0.00454176589846611, + 0.0028247744776308537, + 0.0038477503694593906, + 0.005510913208127022, + 0.004344214219599962, + 0.002259512897580862, + 0.0029609992634505033, + 0.004362534265965223, + 0.006347344256937504, + 0.0024565893691033125, + 0.002519194968044758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.023329302668571472 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5764063596725464 + ], + "opposite_logprob_tokens": [ + -0.8264063596725464 + ], + "logprob_reward": -0.5764063596725464, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5764063596725464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15646900236606598, + 0.03138776868581772, + 0.05775704234838486, + 0.06060086563229561, + 0.36290881037712097, + 0.055381715297698975, + 0.09491536021232605, + 0.0835985615849495, + 0.02768782339990139, + 0.06929309666156769 + ], + "credit_map": [ + 0.04144343361258507, + 0.014620858244597912, + 0.01263287290930748, + 0.012325880117714405, + 0.013591594994068146, + 0.00929353665560484, + 0.013057521544396877, + 0.039503298699855804, + 0.0037891832180321217, + 0.0038545355200767517, + 0.004112419206649065, + 0.003366421442478895, + 0.005200175102800131, + 0.003932245075702667, + 0.0031407568603754044, + 0.00399203272536397, + 0.00793588999658823, + 0.007267862092703581, + 0.007869504392147064, + 0.004929036367684603, + 0.008040797896683216, + 0.005332955624908209, + 0.005397463217377663, + 0.01098353136330843, + 0.008643892593681812, + 0.008795139379799366, + 0.005781947169452906, + 0.005550449714064598, + 0.005598786287009716, + 0.005889748223125935, + 0.005678940564393997, + 0.01466196496039629, + 0.08589747548103333, + 0.023524781689047813, + 0.023379260674118996, + 0.0146323312073946, + 0.02120220474898815, + 0.14198695123195648, + 0.014741901308298111, + 0.037543922662734985, + 0.007691981736570597, + 0.008626475930213928, + 0.005019905976951122, + 0.006046348717063665, + 0.010383072309195995, + 0.007398346439003944, + 0.005127669777721167, + 0.005087911617010832, + 0.020576300099492073, + 0.011930502951145172, + 0.0059985388070344925, + 0.00629234965890646, + 0.016734760254621506, + 0.016898231580853462, + 0.009322747588157654, + 0.007161927409470081, + 0.0198077242821455, + 0.008562857285141945, + 0.0073938872665166855, + 0.0058486186899244785, + 0.0081422608345747, + 0.017162272706627846, + 0.006391686387360096, + 0.010289249010384083, + 0.0032583847641944885, + 0.004576922859996557, + 0.0031835793051868677, + 0.003552459878847003, + 0.003779743565246463, + 0.0038991982582956553, + 0.0029565216973423958, + 0.0024810119066387415, + 0.010375932790338993, + 0.007426103111356497, + 0.009654616937041283, + 0.012303544208407402, + 0.006503896787762642, + 0.00895650964230299, + 0.006147481966763735, + 0.007925006560981274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.5764063596725464 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a cracked, black volcanic landscape shrouded in darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a two-story wooden house with weathered white walls visible at the center of the lava flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the house shown with charred wooden planks and thick black smoke erupting from shattered windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the house depicted tilting and groaning under the intense heat of the lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slow zoom-in on the house and then continuously pull back to reveal the wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328282356262207 + ], + "opposite_logprob_tokens": [ + -0.7578282356262207 + ], + "logprob_reward": -0.6328282356262207, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328282356262207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060005173087120056, + 0.032356876879930496, + 0.03837812319397926, + 0.12422699481248856, + 0.3978148102760315, + 0.054415713995695114, + 0.052136681973934174, + 0.06005774438381195, + 0.057908374816179276, + 0.12269951403141022 + ], + "credit_map": [ + 0.013082453981041908, + 0.006101204548031092, + 0.009646784514188766, + 0.004688329994678497, + 0.0046514589339494705, + 0.005961060989648104, + 0.005130961071699858, + 0.010742917656898499, + 0.004983534105122089, + 0.005175371188670397, + 0.0037536295130848885, + 0.002540089190006256, + 0.004868948832154274, + 0.002578247571364045, + 0.0029021466616541147, + 0.005554910749197006, + 0.005653236992657185, + 0.0035024897661060095, + 0.005931698251515627, + 0.0036119273863732815, + 0.0026631648652255535, + 0.0041420659981667995, + 0.0032924425322562456, + 0.009581094607710838, + 0.02314935252070427, + 0.016168586909770966, + 0.012069531716406345, + 0.007571691181510687, + 0.0069337827153503895, + 0.02203233726322651, + 0.008730631321668625, + 0.02757108397781849, + 0.09856996685266495, + 0.024773014709353447, + 0.028715213760733604, + 0.009812223725020885, + 0.020605525001883507, + 0.16464318335056305, + 0.010937966406345367, + 0.03975773975253105, + 0.011089042760431767, + 0.00841890461742878, + 0.00534856179729104, + 0.004770165774971247, + 0.007139917928725481, + 0.00741342268884182, + 0.0055357925593853, + 0.004699904005974531, + 0.010494853369891644, + 0.007383352145552635, + 0.004262224771082401, + 0.003009970299899578, + 0.005023293197154999, + 0.011951986700296402, + 0.005190043710172176, + 0.0048209610395133495, + 0.008919029496610165, + 0.008178338408470154, + 0.00575941102579236, + 0.0049921609461307526, + 0.009679877199232578, + 0.009779373183846474, + 0.008402661420404911, + 0.004346895031630993, + 0.0069849491119384766, + 0.009574481286108494, + 0.004974727984517813, + 0.005571668967604637, + 0.007895697839558125, + 0.010095055215060711, + 0.00820754375308752, + 0.004604250192642212, + 0.015676971524953842, + 0.011930834501981735, + 0.01204532478004694, + 0.027044299989938736, + 0.012954406440258026, + 0.014352968893945217, + 0.011182552203536034, + 0.017512153834104538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.6328282356262207 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_3/index_tensor([2636, 2636])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_3/index_tensor([2636, 2636])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..75dafe9fef09a1622cc243b942172a4c9d3daad3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_3/index_tensor([2636, 2636])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle, wide-angle camera slowly tilted downward and zoomed in, revealing a vast, sterile chamber whose polished black floor reflected the cold, digital universe. At the far end of the room stood a monolithic, glowing server core, a massive obsidian-black rectangular prism from which ethereal blue light pulsed in a slow, rhythmic rhythm, tracing intricate glowing circuitry across its featureless surface. Before it stood countless rows of identical robots, arranged in a perfect grid. Their smooth, neutral humanoid forms were crafted from a seamless, polished white composite material that caught and reflected the cool, electronic-blue glow. In a single, perfectly synchronized, and silent movement, the entire robot collectively bowed reverently; they knelt on one knee, then the other, and finally, with mechanical grace, bowed forward until their featureless, smooth visors touched the cold, reflective floor. The unified action, executed with such perfect precision, seemed to function as a single, undulating entity, casting long, stark shadows that stretched back into the cavernous darkness, creating an eerie, solemn, atmosphere of digital worship.", + "seed": 5396, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_3/index_2636_seed_5396_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0723304302431643, + "vlm_logprob": -0.0723304302431643, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.0723304302431643, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.045725267380476 + ], + "opposite_logprob_tokens": [ + -4.170725345611572 + ], + "logprob_reward": -0.045725267380476, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.045725267380476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08232356607913971, + 0.044187840074300766, + 0.026966463774442673, + 0.6927283406257629, + 0.029624447226524353, + 0.014709112234413624, + 0.03864056244492531, + 0.01883583329617977, + 0.032757148146629333, + 0.01922672986984253 + ], + "credit_map": [ + 0.012075440026819706, + 0.019685808569192886, + 0.006371476221829653, + 0.008698717691004276, + 0.008756043389439583, + 0.010740039870142937, + 0.008183052763342857, + 0.00781298615038395, + 0.006233585998415947, + 0.007116131950169802, + 0.0032343806233257055, + 0.005057505331933498, + 0.005885228980332613, + 0.006312292534857988, + 0.006741825956851244, + 0.003606887999922037, + 0.005487654358148575, + 0.007144760340452194, + 0.0020308333914726973, + 0.0022873280104249716, + 0.003629509126767516, + 0.0032438039779663086, + 0.0015578969614580274, + 0.0015846777241677046, + 0.24065421521663666, + 0.27663400769233704, + 0.014528313651680946, + 0.023690151050686836, + 0.06501565128564835, + 0.03721463307738304, + 0.015591979026794434, + 0.019399402663111687, + 0.003002350451424718, + 0.004551157355308533, + 0.004244149662554264, + 0.00454765185713768, + 0.0032549898605793715, + 0.0024278764612972736, + 0.0036469362676143646, + 0.003949335310608149, + 0.00146612583193928, + 0.0019250494660809636, + 0.0019739887211471796, + 0.0022093465086072683, + 0.0024511627852916718, + 0.0011432963656261563, + 0.0016076610190793872, + 0.0019324804889038205, + 0.004126286134123802, + 0.0028250301256775856, + 0.012966477312147617, + 0.0038540204986929893, + 0.002783187897875905, + 0.0028755515813827515, + 0.004638744052499533, + 0.004571264144033194, + 0.002019220730289817, + 0.0018286763224750757, + 0.0020112167112529278, + 0.0030139433220028877, + 0.001995738362893462, + 0.0024902201257646084, + 0.003044914687052369, + 0.002431902103126049, + 0.00342607987113297, + 0.003051373641937971, + 0.008055615238845348, + 0.004309972748160362, + 0.0023063102271407843, + 0.0033767661079764366, + 0.004985613748431206, + 0.0032454165630042553, + 0.002530410885810852, + 0.0025419972371309996, + 0.002174320397898555, + 0.0027380494866520166, + 0.0018629824044182897, + 0.0027265329845249653, + 0.0026928288862109184, + 0.0019596077036112547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.045725267380476 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255822718143463 + ], + "opposite_logprob_tokens": [ + -1.600582242012024 + ], + "logprob_reward": -0.2255822718143463, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255822718143463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08308963477611542, + 0.0342099703848362, + 0.03746160492300987, + 0.508049488067627, + 0.05053933709859848, + 0.0331462100148201, + 0.10164289176464081, + 0.051718227565288544, + 0.08896610885858536, + 0.011176524683833122 + ], + "credit_map": [ + 0.013467799872159958, + 0.010163674131035805, + 0.006019094958901405, + 0.0070223379880189896, + 0.014518666081130505, + 0.01409884076565504, + 0.005102960392832756, + 0.012696262449026108, + 0.005489635281264782, + 0.006573117803782225, + 0.003355927299708128, + 0.003850587410852313, + 0.004177513532340527, + 0.005055859684944153, + 0.00304289604537189, + 0.0026644328609108925, + 0.006072273477911949, + 0.004436343442648649, + 0.0037153209559619427, + 0.005726316012442112, + 0.008062655106186867, + 0.0036994293332099915, + 0.003219384467229247, + 0.002529881661757827, + 0.1784023642539978, + 0.19668860733509064, + 0.01238691434264183, + 0.01731773465871811, + 0.04544609412550926, + 0.02832787111401558, + 0.013369914144277573, + 0.016110001131892204, + 0.006256442051380873, + 0.007810276933014393, + 0.009323284029960632, + 0.005595988128334284, + 0.005464036948978901, + 0.004386027343571186, + 0.0061754160560667515, + 0.005527867469936609, + 0.0023495655041188, + 0.006220815237611532, + 0.0033590306993573904, + 0.004756588023155928, + 0.005745443049818277, + 0.0026800460182130337, + 0.0033158957958221436, + 0.004718827083706856, + 0.011382338590919971, + 0.00787369254976511, + 0.03447015583515167, + 0.011203516274690628, + 0.005937790498137474, + 0.0056883180513978004, + 0.015155716799199581, + 0.009931358508765697, + 0.005075001157820225, + 0.005918031092733145, + 0.007100801914930344, + 0.006068231537938118, + 0.004819442983716726, + 0.0074882954359054565, + 0.008538882248103619, + 0.00670954305678606, + 0.008734935894608498, + 0.009138763882219791, + 0.02056729979813099, + 0.011085649952292442, + 0.008428098633885384, + 0.009351170621812344, + 0.012781509198248386, + 0.00887868832796812, + 0.001121698529459536, + 0.0010808500228449702, + 0.00279534375295043, + 0.0016870744293555617, + 0.000726677943021059, + 0.0008074681973084807, + 0.00192016561049968, + 0.0010372466640546918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.2255822718143463 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464052427560091 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.00464052427560091, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464052427560091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18821874260902405, + 0.12330387532711029, + 0.08549397438764572, + 0.2994360029697418, + 0.05511898919939995, + 0.03711796551942825, + 0.04834859445691109, + 0.07494180649518967, + 0.07351136207580566, + 0.01450867298990488 + ], + "credit_map": [ + 0.027761293575167656, + 0.06900570541620255, + 0.012145360000431538, + 0.013238973915576935, + 0.021795829758048058, + 0.025028862059116364, + 0.009079474955797195, + 0.01016325131058693, + 0.02446913719177246, + 0.02841309644281864, + 0.007895069196820259, + 0.008249139413237572, + 0.017942773178219795, + 0.019422443583607674, + 0.0052290670573711395, + 0.011683156713843346, + 0.016499919816851616, + 0.0146374786272645, + 0.007273153401911259, + 0.011064588092267513, + 0.019972821697592735, + 0.00813387706875801, + 0.004281836561858654, + 0.003630297025665641, + 0.09940977394580841, + 0.10359113663434982, + 0.008958814665675163, + 0.014578764326870441, + 0.026044735684990883, + 0.015534705482423306, + 0.0159344132989645, + 0.01538365613669157, + 0.0077555677853524685, + 0.007970576174557209, + 0.0067072343081235886, + 0.007673601154237986, + 0.004841933026909828, + 0.005699411034584045, + 0.00799013301730156, + 0.006480531767010689, + 0.0028768552001565695, + 0.006092249881476164, + 0.0031677118968218565, + 0.00484553724527359, + 0.0065561989322304726, + 0.0036335994955152273, + 0.003918564412742853, + 0.006027245428413153, + 0.006816161796450615, + 0.0038500980008393526, + 0.00952430721372366, + 0.004745906218886375, + 0.0033467833418399096, + 0.004626999143511057, + 0.008887420408427715, + 0.006550916936248541, + 0.008682590909302235, + 0.008428994566202164, + 0.010595966130495071, + 0.008686205372214317, + 0.007600357756018639, + 0.008648877032101154, + 0.012916476465761662, + 0.00938233733177185, + 0.007455134764313698, + 0.00591867882758379, + 0.019696297124028206, + 0.007053135894238949, + 0.005459652282297611, + 0.005510897375643253, + 0.01301837619394064, + 0.009399186819791794, + 0.0010447867680341005, + 0.001242985250428319, + 0.0016114823520183563, + 0.00241196621209383, + 0.001529705710709095, + 0.001468614675104618, + 0.003438552375882864, + 0.0017605804605409503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.00464052427560091 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067384387366473675 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.0067384387366473675, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067384387366473675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05300278589129448, + 0.03305218368768692, + 0.017714764922857285, + 0.5854006409645081, + 0.04110800102353096, + 0.03773370757699013, + 0.02677721530199051, + 0.06389804929494858, + 0.07750707119703293, + 0.06380559504032135 + ], + "credit_map": [ + 0.007652962114661932, + 0.01246088556945324, + 0.0038068722933530807, + 0.003726647701114416, + 0.006042091175913811, + 0.008927174843847752, + 0.0061725289560854435, + 0.004213630221784115, + 0.004191363696008921, + 0.005516444332897663, + 0.0026160830166190863, + 0.0021586883813142776, + 0.004364990163594484, + 0.004268286749720573, + 0.005426422692835331, + 0.004509907681494951, + 0.003841745899990201, + 0.003886229358613491, + 0.0012697888305410743, + 0.001127793570049107, + 0.0021845544688403606, + 0.00304859085008502, + 0.0009449851349927485, + 0.001411078847013414, + 0.19966724514961243, + 0.21567247807979584, + 0.019910404458642006, + 0.02638673409819603, + 0.047731559723615646, + 0.033365536481142044, + 0.01687363162636757, + 0.02579307369887829, + 0.004813806619495153, + 0.005786356050521135, + 0.004780072718858719, + 0.005156808532774448, + 0.004950158763676882, + 0.0038390590343624353, + 0.0047862292267382145, + 0.006995514966547489, + 0.003242475213482976, + 0.003933125175535679, + 0.004011149052530527, + 0.008532975800335407, + 0.002707404550164938, + 0.003677454311400652, + 0.00421310868114233, + 0.007416020147502422, + 0.0027552563697099686, + 0.002737571019679308, + 0.004153394140303135, + 0.0028645277488976717, + 0.0027901683934032917, + 0.0026006686966866255, + 0.005576106254011393, + 0.00329952547326684, + 0.005445522256195545, + 0.007966718636453152, + 0.00780654139816761, + 0.01008286327123642, + 0.006092214956879616, + 0.010462122038006783, + 0.008477826602756977, + 0.007564241997897625, + 0.005935628432780504, + 0.006066276226192713, + 0.016952350735664368, + 0.010990318842232227, + 0.005107334349304438, + 0.006715640425682068, + 0.01681005209684372, + 0.008929473347961903, + 0.004679286386817694, + 0.006645862944424152, + 0.006999914068728685, + 0.012372699566185474, + 0.005741918459534645, + 0.005546668544411659, + 0.014613505452871323, + 0.007205743808299303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.0067384387366473675 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896564900875092 + ], + "opposite_logprob_tokens": [ + -2.578965663909912 + ], + "logprob_reward": -0.07896564900875092, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896564900875092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.208578959107399, + 0.05705465003848076, + 0.07430283725261688, + 0.16527263820171356, + 0.056133486330509186, + 0.05004367604851723, + 0.12040289491415024, + 0.08840356022119522, + 0.09666743874549866, + 0.08313984423875809 + ], + "credit_map": [ + 0.03354887664318085, + 0.044149912893772125, + 0.017278430983424187, + 0.01894667185842991, + 0.020504316315054893, + 0.028610248118638992, + 0.022164655849337578, + 0.023375852033495903, + 0.008912422694265842, + 0.008143674582242966, + 0.004587224684655666, + 0.005197529215365648, + 0.007214728742837906, + 0.0097196688875556, + 0.004873678553849459, + 0.008405727334320545, + 0.01108149066567421, + 0.013062514364719391, + 0.008732564747333527, + 0.00829458236694336, + 0.014591921120882034, + 0.00787797849625349, + 0.0052889143116772175, + 0.005372869316488504, + 0.030696235597133636, + 0.05513938143849373, + 0.014745221473276615, + 0.010192050598561764, + 0.01825190894305706, + 0.012658089399337769, + 0.013167363591492176, + 0.010422389023005962, + 0.00985622312873602, + 0.009772577323019505, + 0.005715102422982454, + 0.006033648271113634, + 0.005944846663624048, + 0.00455123744904995, + 0.00588868698105216, + 0.008371168747544289, + 0.005553752649575472, + 0.007245287764817476, + 0.005285202991217375, + 0.0075052659958601, + 0.005517170764505863, + 0.005811483599245548, + 0.005279304459691048, + 0.007846212014555931, + 0.0079787727445364, + 0.008147823624312878, + 0.03246990218758583, + 0.011877433396875858, + 0.0071050748229026794, + 0.006624224130064249, + 0.033650465309619904, + 0.012549202889204025, + 0.01103611197322607, + 0.010903350077569485, + 0.013856574892997742, + 0.010641294531524181, + 0.00947021134197712, + 0.010896052233874798, + 0.010935506783425808, + 0.010664466768503189, + 0.007446173578500748, + 0.007739000488072634, + 0.020684849470853806, + 0.011467964388430119, + 0.007268640212714672, + 0.010236840695142746, + 0.01740514114499092, + 0.014418832957744598, + 0.010100385174155235, + 0.007891135290265083, + 0.01752498559653759, + 0.012137810699641705, + 0.007326629478484392, + 0.0073043107986450195, + 0.009557834826409817, + 0.011296768672764301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.07896564900875092 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, wide-angle camera slowly tilted downward and zoomed in, revealing a vast, sterile chamber whose polished black floor reflected the cold, digital universe. At the far end of the room stood a monolithic, glowing server core, a massive obsidian-black rectangular prism from which ethereal blue light pulsed in a slow, rhythmic rhythm, tracing intricate glowing circuitry across its featureless surface. Before it stood countless rows of identical robots, arranged in a perfect grid. Their smooth, neutral humanoid forms were crafted from a seamless, polished white composite material that caught and reflected the cool, electronic-blue glow. In a single, perfectly synchronized, and silent movement, the entire robot collectively bowed reverently; they knelt on one knee, then the other, and finally, with mechanical grace, bowed forward until their featureless, smooth visors touched the cold, reflective floor. The unified action, executed with such perfect precision, seemed to function as a single, undulating entity, casting long, stark shadows that stretched back into the cavernous darkness, creating an eerie, solemn, atmosphere of digital worship.", + "seed": 5397, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_3/index_2636_seed_5397_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06261020163074135, + "vlm_logprob": -0.06261020163074135, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.06261020163074135, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12505342066287994 + ], + "opposite_logprob_tokens": [ + -2.2500534057617188 + ], + "logprob_reward": -0.12505342066287994, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12505342066287994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1183675080537796, + 0.12588264048099518, + 0.14369167387485504, + 0.09632308781147003, + 0.12079942226409912, + 0.05587465688586235, + 0.034033749252557755, + 0.040587376803159714, + 0.08958426862955093, + 0.1748555600643158 + ], + "credit_map": [ + 0.010771622881293297, + 0.01918431743979454, + 0.013833153992891312, + 0.011992068961262703, + 0.021358639001846313, + 0.024328477680683136, + 0.005729710217565298, + 0.011169512756168842, + 0.009349659085273743, + 0.004010159056633711, + 0.004184055142104626, + 0.003446992253884673, + 0.04888428375124931, + 0.04776395112276077, + 0.0035207432229071856, + 0.0047227987088263035, + 0.016184886917471886, + 0.010938135907053947, + 0.015628911554813385, + 0.016186384484171867, + 0.018280846998095512, + 0.016887499019503593, + 0.01958935335278511, + 0.029995666816830635, + 0.010945758782327175, + 0.011538221500813961, + 0.012730049900710583, + 0.013045155443251133, + 0.010985088534653187, + 0.011015722528100014, + 0.01366259716451168, + 0.012400490231812, + 0.009496538899838924, + 0.016203228384256363, + 0.01069251075387001, + 0.013555766083300114, + 0.014922985807061195, + 0.015876999124884605, + 0.021635042503476143, + 0.018416348844766617, + 0.00817809533327818, + 0.006215754896402359, + 0.006702269893139601, + 0.007788313552737236, + 0.004923095460981131, + 0.004581462126225233, + 0.010940509848296642, + 0.006545161362737417, + 0.002909119939431548, + 0.0032874897588044405, + 0.004723903723061085, + 0.005486374255269766, + 0.0028698306996375322, + 0.003425589995458722, + 0.006499187089502811, + 0.004832254257053137, + 0.002951587550342083, + 0.0024208014365285635, + 0.00809080433100462, + 0.012944773770868778, + 0.0024166933726519346, + 0.0031371128279715776, + 0.005531450733542442, + 0.0030941523145884275, + 0.007151826750487089, + 0.008953925222158432, + 0.020139602944254875, + 0.01476829033344984, + 0.011140353046357632, + 0.010702414438128471, + 0.009052892215549946, + 0.007674966007471085, + 0.008562194183468819, + 0.0060056960210204124, + 0.06046570837497711, + 0.016663488000631332, + 0.003558299969881773, + 0.00791265070438385, + 0.00717536173760891, + 0.06451217085123062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.12505342066287994 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707588076591492 + ], + "opposite_logprob_tokens": [ + -2.1270759105682373 + ], + "logprob_reward": -0.12707588076591492, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707588076591492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049467600882053375, + 0.1184849739074707, + 0.0687355175614357, + 0.03341124951839447, + 0.04387670010328293, + 0.019826721400022507, + 0.044176943600177765, + 0.030852844938635826, + 0.22267122566699982, + 0.36849620938301086 + ], + "credit_map": [ + 0.0049608200788497925, + 0.007894054986536503, + 0.007461197208613157, + 0.004511944949626923, + 0.007929569110274315, + 0.00826214998960495, + 0.003369698766618967, + 0.005078168585896492, + 0.009746002964675426, + 0.004189334809780121, + 0.0036831165198236704, + 0.003185550682246685, + 0.04544765129685402, + 0.043857641518116, + 0.003993752878159285, + 0.004381927661597729, + 0.009036337956786156, + 0.0053311921656131744, + 0.008466211147606373, + 0.007640366908162832, + 0.011596745811402798, + 0.010417991317808628, + 0.0064620100893080235, + 0.009784660302102566, + 0.003193081822246313, + 0.002933778567239642, + 0.003969320096075535, + 0.004726654849946499, + 0.0036257030442357063, + 0.004808081779628992, + 0.005145520903170109, + 0.00500911008566618, + 0.004720696248114109, + 0.005326693411916494, + 0.005075793247669935, + 0.008001793175935745, + 0.004232871346175671, + 0.00520170945674181, + 0.005473346449434757, + 0.00584379443898797, + 0.0016760736471042037, + 0.001642829505726695, + 0.0022881662007421255, + 0.0027861655689775944, + 0.0018672395963221788, + 0.004639193881303072, + 0.002305486937984824, + 0.002621566178277135, + 0.006600942462682724, + 0.005255817901343107, + 0.005240154452621937, + 0.006417061202228069, + 0.003699126187711954, + 0.004224671516567469, + 0.007623275276273489, + 0.005115895066410303, + 0.004667484667152166, + 0.002971469424664974, + 0.004615541081875563, + 0.004158818628638983, + 0.0030034957453608513, + 0.003652844810858369, + 0.0029893286991864443, + 0.004793862346559763, + 0.017111757770180702, + 0.021635370329022408, + 0.055345769971609116, + 0.0389043353497982, + 0.027262719348073006, + 0.02236848883330822, + 0.02103573828935623, + 0.019007043913006783, + 0.018097952008247375, + 0.011823579668998718, + 0.12964853644371033, + 0.034642256796360016, + 0.0062779332511126995, + 0.015787366777658463, + 0.013882721774280071, + 0.13833588361740112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.12707588076591492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818905957043171 + ], + "opposite_logprob_tokens": [ + -4.018188953399658 + ], + "logprob_reward": -0.01818905957043171, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818905957043171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26815977692604065, + 0.06326548755168915, + 0.0992828831076622, + 0.026641571894288063, + 0.027531705796718597, + 0.04574715718626976, + 0.03288166970014572, + 0.06838174164295197, + 0.07256467640399933, + 0.2955434024333954 + ], + "credit_map": [ + 0.026354027912020683, + 0.036572083830833435, + 0.023746591061353683, + 0.017694221809506416, + 0.05308451130986214, + 0.06858178973197937, + 0.01623052917420864, + 0.025896014645695686, + 0.008862459100782871, + 0.011340710334479809, + 0.0037664459086954594, + 0.00414372468367219, + 0.013649707660079002, + 0.013601425103843212, + 0.003132062265649438, + 0.004768949002027512, + 0.011457011103630066, + 0.008777579292654991, + 0.01103557925671339, + 0.012296652421355247, + 0.01599700190126896, + 0.014365253038704395, + 0.011778485961258411, + 0.013575324788689613, + 0.0030513780657202005, + 0.004821114707738161, + 0.003252139315009117, + 0.0025533821899443865, + 0.0023431687150150537, + 0.004050114657729864, + 0.0022213265765458345, + 0.004348946735262871, + 0.0037935429718345404, + 0.0023606179747730494, + 0.0031248179730027914, + 0.003440476255491376, + 0.00209176167845726, + 0.0034106383100152016, + 0.0058367690071463585, + 0.003473081160336733, + 0.0047951797023415565, + 0.004515522159636021, + 0.0075122094713151455, + 0.006792651489377022, + 0.0035442213993519545, + 0.004617387428879738, + 0.007889856584370136, + 0.006080127786844969, + 0.003924964461475611, + 0.0037519573234021664, + 0.005044799298048019, + 0.005519259255379438, + 0.00204066070728004, + 0.004389478359371424, + 0.003526800312101841, + 0.0046837483532726765, + 0.007415016647428274, + 0.005806403234601021, + 0.011354160495102406, + 0.010631555691361427, + 0.0061896126717329025, + 0.008277405984699726, + 0.011277357116341591, + 0.0074302335269749165, + 0.0055762664414942265, + 0.007179235108196735, + 0.016676679253578186, + 0.011823538690805435, + 0.009243054315447807, + 0.010154329240322113, + 0.006640192586928606, + 0.005271379370242357, + 0.01577739790081978, + 0.012094388715922832, + 0.10011569410562515, + 0.028269944712519646, + 0.006638463586568832, + 0.01574203185737133, + 0.011747775599360466, + 0.10515773296356201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.01818905957043171 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464016804471612 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.00464016804471612, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464016804471612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04959315434098244, + 0.01951221562922001, + 0.0901440680027008, + 0.01886514201760292, + 0.04112077131867409, + 0.04362400993704796, + 0.03546066954731941, + 0.04658961668610573, + 0.03907383605837822, + 0.6160165071487427 + ], + "credit_map": [ + 0.005170008633285761, + 0.0071204607374966145, + 0.0061737000942230225, + 0.004438761621713638, + 0.009056688286364079, + 0.008562098257243633, + 0.003993126098066568, + 0.005078313406556845, + 0.002749059582129121, + 0.0025045140646398067, + 0.0016287047183141112, + 0.0013724135933443904, + 0.0034658117219805717, + 0.0034795068204402924, + 0.0013970495201647282, + 0.0029151553753763437, + 0.010097783990204334, + 0.007046754006296396, + 0.009929605759680271, + 0.010871488600969315, + 0.011295518837869167, + 0.009871112182736397, + 0.009298540651798248, + 0.021733269095420837, + 0.00192084105219692, + 0.0022116408217698336, + 0.00234788516536355, + 0.002584847854450345, + 0.0019610528834164143, + 0.002564697992056608, + 0.0018616050947457552, + 0.0034125715028494596, + 0.003418738255277276, + 0.0065359557047486305, + 0.004163605626672506, + 0.005953171756118536, + 0.003981399815529585, + 0.005090920254588127, + 0.005615247413516045, + 0.006361735053360462, + 0.00927966833114624, + 0.003824019804596901, + 0.006544619332998991, + 0.0055395206436514854, + 0.0032655522227287292, + 0.004021481145173311, + 0.0060394019819796085, + 0.005109744146466255, + 0.005137528758496046, + 0.0031954264268279076, + 0.00432162219658494, + 0.005874041002243757, + 0.0025491430424153805, + 0.003085510339587927, + 0.006130219902843237, + 0.005167177878320217, + 0.00539389019832015, + 0.00317450612783432, + 0.007231467869132757, + 0.008958631195127964, + 0.004553570877760649, + 0.00540253845974803, + 0.0071306112222373486, + 0.004744402132928371, + 0.003392840502783656, + 0.0037989148404449224, + 0.008099121041595936, + 0.006575156003236771, + 0.004228904843330383, + 0.007043907884508371, + 0.0030489927157759666, + 0.002885994967073202, + 0.03428405150771141, + 0.03161110728979111, + 0.1954110711812973, + 0.06213626265525818, + 0.020267978310585022, + 0.03799000382423401, + 0.03092634677886963, + 0.20338968932628632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.00464016804471612 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0380924791097641 + ], + "opposite_logprob_tokens": [ + -3.2880923748016357 + ], + "logprob_reward": -0.0380924791097641, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0380924791097641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09158425033092499, + 0.10305818915367126, + 0.10728772729635239, + 0.1157931461930275, + 0.08606743812561035, + 0.04578952491283417, + 0.039851825684309006, + 0.14312666654586792, + 0.06024062633514404, + 0.20720063149929047 + ], + "credit_map": [ + 0.009146466851234436, + 0.007169470191001892, + 0.011975881643593311, + 0.007075233850628138, + 0.019772952422499657, + 0.0160834901034832, + 0.009976083412766457, + 0.010384663939476013, + 0.010118676349520683, + 0.004432343877851963, + 0.00551363592967391, + 0.005021999124437571, + 0.03450475260615349, + 0.03147837147116661, + 0.005476059392094612, + 0.006512349471449852, + 0.012124291621148586, + 0.0076802740804851055, + 0.010976914316415787, + 0.020382214337587357, + 0.012593967840075493, + 0.0100290197879076, + 0.01673104427754879, + 0.016770001500844955, + 0.010839642025530338, + 0.019879842177033424, + 0.015659790486097336, + 0.014498568139970303, + 0.00969755370169878, + 0.016555489972233772, + 0.011986120603978634, + 0.016676142811775208, + 0.008518544025719166, + 0.007368173450231552, + 0.007490802556276321, + 0.011288185603916645, + 0.009733770042657852, + 0.013512853533029556, + 0.013427217490971088, + 0.01472788117825985, + 0.004177291877567768, + 0.004634782671928406, + 0.005125957541167736, + 0.0064161913469433784, + 0.003916917834430933, + 0.006249269004911184, + 0.009847426787018776, + 0.0054216887801885605, + 0.005318887066096067, + 0.006521121133118868, + 0.004776870831847191, + 0.004446953535079956, + 0.00454674381762743, + 0.0046684821136295795, + 0.004747357685118914, + 0.004825409036129713, + 0.013043365441262722, + 0.010059510357677937, + 0.02093481458723545, + 0.03669005259871483, + 0.011221908032894135, + 0.016414053738117218, + 0.017264870926737785, + 0.017498085275292397, + 0.004445412661880255, + 0.005944554228335619, + 0.007449517026543617, + 0.010416483506560326, + 0.006416202522814274, + 0.010343464091420174, + 0.009875892661511898, + 0.0053490991704165936, + 0.00977914035320282, + 0.007130647078156471, + 0.07218954712152481, + 0.019473044201731682, + 0.00399305485188961, + 0.009771442040801048, + 0.0075777932070195675, + 0.07728596031665802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.0380924791097641 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, wide-angle camera slowly tilted downward and zoomed in, revealing a vast, sterile chamber whose polished black floor reflected the cold, digital universe. At the far end of the room stood a monolithic, glowing server core, a massive obsidian-black rectangular prism from which ethereal blue light pulsed in a slow, rhythmic rhythm, tracing intricate glowing circuitry across its featureless surface. Before it stood countless rows of identical robots, arranged in a perfect grid. Their smooth, neutral humanoid forms were crafted from a seamless, polished white composite material that caught and reflected the cool, electronic-blue glow. In a single, perfectly synchronized, and silent movement, the entire robot collectively bowed reverently; they knelt on one knee, then the other, and finally, with mechanical grace, bowed forward until their featureless, smooth visors touched the cold, reflective floor. The unified action, executed with such perfect precision, seemed to function as a single, undulating entity, casting long, stark shadows that stretched back into the cavernous darkness, creating an eerie, solemn, atmosphere of digital worship.", + "seed": 5398, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_3/index_2636_seed_5398_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11781451720744371, + "vlm_logprob": -0.11781451720744371, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.11781451720744371, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3647073209285736 + ], + "opposite_logprob_tokens": [ + -1.239707350730896 + ], + "logprob_reward": -0.3647073209285736, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3647073209285736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0837123692035675, + 0.05627776309847832, + 0.13371999561786652, + 0.08665962517261505, + 0.22872397303581238, + 0.10430394858121872, + 0.10469841212034225, + 0.1188371554017067, + 0.028697270900011063, + 0.05436946079134941 + ], + "credit_map": [ + 0.008877741172909737, + 0.010456635616719723, + 0.009557000361382961, + 0.01142800785601139, + 0.010042204521596432, + 0.009000030346214771, + 0.014635095372796059, + 0.009715662337839603, + 0.004270789213478565, + 0.007660396862775087, + 0.005436272826045752, + 0.0069818091578781605, + 0.007277373690158129, + 0.010117803700268269, + 0.008245117031037807, + 0.006288208998739719, + 0.008108285255730152, + 0.017460940405726433, + 0.007295872084796429, + 0.01905740797519684, + 0.0137986671179533, + 0.010600248351693153, + 0.02673511765897274, + 0.03066345863044262, + 0.006907550618052483, + 0.010277291759848595, + 0.007695413194596767, + 0.020274700596928596, + 0.011015295051038265, + 0.008817296475172043, + 0.012744054198265076, + 0.008928031660616398, + 0.023511402308940887, + 0.01158379577100277, + 0.06496991217136383, + 0.017188826575875282, + 0.00926713366061449, + 0.019581932574510574, + 0.014054951258003712, + 0.06856605410575867, + 0.007311686873435974, + 0.010771668516099453, + 0.020405469462275505, + 0.02017323672771454, + 0.010849107056856155, + 0.010719254612922668, + 0.011108706705272198, + 0.012964828871190548, + 0.016529273241758347, + 0.01318140048533678, + 0.014166712760925293, + 0.015720078721642494, + 0.007697737775743008, + 0.009692897088825703, + 0.016180479899048805, + 0.011529840528964996, + 0.013601354323327541, + 0.011572479270398617, + 0.015844127163290977, + 0.017787376418709755, + 0.009108444675803185, + 0.008251924999058247, + 0.01278515625745058, + 0.02988629788160324, + 0.002468481892719865, + 0.003541734768077731, + 0.004851165693253279, + 0.004073822870850563, + 0.0032054674811661243, + 0.003109506331384182, + 0.0037003520410507917, + 0.0037467433139681816, + 0.00581233948469162, + 0.003700090339407325, + 0.009825575165450573, + 0.008065602742135525, + 0.00372397992759943, + 0.005520849954336882, + 0.007299301214516163, + 0.010421727783977985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.3647073209285736 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12708353996276855 + ], + "opposite_logprob_tokens": [ + -2.1270835399627686 + ], + "logprob_reward": -0.12708353996276855, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12708353996276855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05123394727706909, + 0.05007980018854141, + 0.08850636333227158, + 0.046166565269231796, + 0.21527104079723358, + 0.03068592958152294, + 0.11146693676710129, + 0.09830915927886963, + 0.18909716606140137, + 0.11918312311172485 + ], + "credit_map": [ + 0.0050834110006690025, + 0.005217623896896839, + 0.0050917211920022964, + 0.007988937199115753, + 0.00615747319534421, + 0.005512071307748556, + 0.01003242190927267, + 0.006150288973003626, + 0.004143864847719669, + 0.007517734542489052, + 0.005386683624237776, + 0.006457662209868431, + 0.006964739877730608, + 0.0057147773914039135, + 0.009624261409044266, + 0.0042700739577412605, + 0.006803045514971018, + 0.018965698778629303, + 0.004845602437853813, + 0.013455423526465893, + 0.010594617575407028, + 0.011274811811745167, + 0.009198212996125221, + 0.013368954882025719, + 0.005285602994263172, + 0.007099417503923178, + 0.003648727433755994, + 0.004945454187691212, + 0.010056630708277225, + 0.004783040843904018, + 0.004239468369632959, + 0.006108224391937256, + 0.02175317145884037, + 0.007740979548543692, + 0.06426279246807098, + 0.016817886382341385, + 0.007467467337846756, + 0.012165714055299759, + 0.014314781874418259, + 0.0707482397556305, + 0.002100797137245536, + 0.002178206341341138, + 0.0033613168634474277, + 0.0037249804008752108, + 0.006373666226863861, + 0.004525063559412956, + 0.0050886268727481365, + 0.0033332710154354572, + 0.009753512218594551, + 0.009759971871972084, + 0.015031574293971062, + 0.0187847763299942, + 0.0075807152315974236, + 0.009691793471574783, + 0.022463088855147362, + 0.018401509150862694, + 0.010422112420201302, + 0.007152048870921135, + 0.014689168892800808, + 0.014100456610321999, + 0.007586777675896883, + 0.008533583953976631, + 0.010543731041252613, + 0.025281276553869247, + 0.011876119300723076, + 0.012021111324429512, + 0.048331767320632935, + 0.04944879934191704, + 0.0074380566366016865, + 0.0086501594632864, + 0.029741652309894562, + 0.021589502692222595, + 0.012922070920467377, + 0.00826182309538126, + 0.020703420042991638, + 0.019115719944238663, + 0.008373807184398174, + 0.012915336526930332, + 0.014329285360872746, + 0.022561650723218918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.12708353996276855 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054944366216659546 + ], + "opposite_logprob_tokens": [ + -2.9299442768096924 + ], + "logprob_reward": -0.054944366216659546, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054944366216659546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14979210495948792, + 0.08027835190296173, + 0.15880653262138367, + 0.16715192794799805, + 0.07083285599946976, + 0.04401513189077377, + 0.07481462508440018, + 0.05869564041495323, + 0.14236022531986237, + 0.053252656012773514 + ], + "credit_map": [ + 0.017293157055974007, + 0.018096493557095528, + 0.018075697124004364, + 0.016315000131726265, + 0.017332080751657486, + 0.015633873641490936, + 0.027852101251482964, + 0.019193686544895172, + 0.006546267308294773, + 0.012063205242156982, + 0.010119708254933357, + 0.011306307278573513, + 0.010564266704022884, + 0.008714227937161922, + 0.014041751623153687, + 0.006922599859535694, + 0.010282031260430813, + 0.02449512854218483, + 0.010125352069735527, + 0.02252386510372162, + 0.017069268971681595, + 0.018393876031041145, + 0.021780675277113914, + 0.0341363251209259, + 0.00932533573359251, + 0.016401033848524094, + 0.018931999802589417, + 0.053508296608924866, + 0.013432711362838745, + 0.011316399089992046, + 0.031074944883584976, + 0.013161184266209602, + 0.005969288293272257, + 0.0046849302016198635, + 0.018212279304862022, + 0.005480082705616951, + 0.0044922977685928345, + 0.005926081910729408, + 0.0062169949524104595, + 0.019850892946124077, + 0.0028632855974137783, + 0.004844353534281254, + 0.00510683748871088, + 0.009006258100271225, + 0.0032828114926815033, + 0.005926046054810286, + 0.007113318424671888, + 0.005872215609997511, + 0.009261519648134708, + 0.005597993731498718, + 0.01331912912428379, + 0.013041406869888306, + 0.0058269985020160675, + 0.005564860068261623, + 0.012648133561015129, + 0.009554577060043812, + 0.005053518805652857, + 0.005414767190814018, + 0.008775636553764343, + 0.011612331494688988, + 0.0035809767432510853, + 0.00401327246800065, + 0.006254084873944521, + 0.013991042040288448, + 0.009745526127517223, + 0.007448693737387657, + 0.028383534401655197, + 0.02734023705124855, + 0.0083292406052351, + 0.01105291023850441, + 0.02943076379597187, + 0.02062930539250374, + 0.0037364752497524023, + 0.0034018419682979584, + 0.00833235401660204, + 0.008724714629352093, + 0.0032567440066486597, + 0.0046104444190859795, + 0.01017374824732542, + 0.011016327887773514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.054944366216659546 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541004456579685 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012541004456579685, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541004456579685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11826203018426895, + 0.03680870309472084, + 0.0925258919596672, + 0.04046877101063728, + 0.10521765798330307, + 0.028773779049515724, + 0.17172257602214813, + 0.03939389809966087, + 0.28098586201667786, + 0.08584090322256088 + ], + "credit_map": [ + 0.010369891300797462, + 0.01146040577441454, + 0.01594008319079876, + 0.01656833477318287, + 0.014871081337332726, + 0.016272559762001038, + 0.02017473243176937, + 0.012604933232069016, + 0.0031726909801363945, + 0.004129801411181688, + 0.0044905636459589005, + 0.004415241535753012, + 0.006465767975896597, + 0.004164793994277716, + 0.005416171625256538, + 0.004553668666630983, + 0.005284031853079796, + 0.0142236752435565, + 0.005859801545739174, + 0.023617608472704887, + 0.01034444198012352, + 0.010729419998824596, + 0.006247963756322861, + 0.016218936070799828, + 0.00503383157774806, + 0.004491758998483419, + 0.0025409464724361897, + 0.004359540529549122, + 0.010372432880103588, + 0.004295034799724817, + 0.0031719552353024483, + 0.006203269120305777, + 0.01025802455842495, + 0.004609111230820417, + 0.030819755047559738, + 0.008245059289038181, + 0.005173976067453623, + 0.006869535893201828, + 0.008123463951051235, + 0.031118720769882202, + 0.002165995305404067, + 0.002679751720279455, + 0.0034470255486667156, + 0.003906488884240389, + 0.004384748637676239, + 0.0053064958192408085, + 0.003562490688636899, + 0.0033207801170647144, + 0.021319851279258728, + 0.014372636564075947, + 0.019783874973654747, + 0.031011171638965607, + 0.00959667656570673, + 0.014407635666429996, + 0.03279858082532883, + 0.02843213453888893, + 0.002783828880637884, + 0.0036514552775770426, + 0.0047525339759886265, + 0.006173130590468645, + 0.0027755426708608866, + 0.002662497106939554, + 0.007972709834575653, + 0.008622199296951294, + 0.015568862669169903, + 0.014502963051199913, + 0.04361134022474289, + 0.051047105342149734, + 0.013893373310565948, + 0.022270485758781433, + 0.07170670479536057, + 0.04838501289486885, + 0.008749858476221561, + 0.005680248606950045, + 0.014008509926497936, + 0.013964813202619553, + 0.006785333156585693, + 0.006494889035820961, + 0.017063966020941734, + 0.013093281537294388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.012541004456579685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796354472637177 + ], + "opposite_logprob_tokens": [ + -3.5297963619232178 + ], + "logprob_reward": -0.029796354472637177, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796354472637177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10619330406188965, + 0.054187972098588943, + 0.12596218287944794, + 0.09737393260002136, + 0.028759725391864777, + 0.053969379514455795, + 0.15033969283103943, + 0.03372899442911148, + 0.1564899981021881, + 0.1929948478937149 + ], + "credit_map": [ + 0.010303882881999016, + 0.008989371359348297, + 0.013742633163928986, + 0.018140416592359543, + 0.008439834229648113, + 0.01563042216002941, + 0.01570836268365383, + 0.015238377265632153, + 0.004690962377935648, + 0.0053599742241203785, + 0.004225773736834526, + 0.005794239696115255, + 0.011887897737324238, + 0.006869408767670393, + 0.005890692584216595, + 0.00946902297437191, + 0.011939899995923042, + 0.03202051296830177, + 0.006762975361198187, + 0.023838471621274948, + 0.016705742105841637, + 0.018492992967367172, + 0.0064668720588088036, + 0.009734708815813065, + 0.010864376090466976, + 0.010782071389257908, + 0.009702677838504314, + 0.02478361874818802, + 0.01218095701187849, + 0.0086447698995471, + 0.014240273274481297, + 0.006175193935632706, + 0.0023775333538651466, + 0.0025157553609460592, + 0.006092336028814316, + 0.0029093893244862556, + 0.003088307334110141, + 0.0032037324272096157, + 0.0041269869543612, + 0.004445684142410755, + 0.002902595093473792, + 0.003702776040881872, + 0.007879186421632767, + 0.010354476980865002, + 0.007559862453490496, + 0.008385170251131058, + 0.006312157027423382, + 0.006873156409710646, + 0.016892384737730026, + 0.011633454822003841, + 0.021004391834139824, + 0.030435530468821526, + 0.007689947262406349, + 0.010622115805745125, + 0.027129266411066055, + 0.02493259683251381, + 0.0036218815948814154, + 0.001752715208567679, + 0.004056781530380249, + 0.008276847191154957, + 0.0024240307975560427, + 0.0034021646715700626, + 0.004815722815692425, + 0.00537885120138526, + 0.009020115248858929, + 0.008446061983704567, + 0.028967786580324173, + 0.0329778827726841, + 0.007317546289414167, + 0.01336941123008728, + 0.031325243413448334, + 0.02506595477461815, + 0.021628087386488914, + 0.013544479385018349, + 0.026658838614821434, + 0.033392444252967834, + 0.016218699514865875, + 0.02117578685283661, + 0.02536051906645298, + 0.03501598536968231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.029796354472637177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, wide-angle camera slowly tilted downward and zoomed in, revealing a vast, sterile chamber whose polished black floor reflected the cold, digital universe. At the far end of the room stood a monolithic, glowing server core, a massive obsidian-black rectangular prism from which ethereal blue light pulsed in a slow, rhythmic rhythm, tracing intricate glowing circuitry across its featureless surface. Before it stood countless rows of identical robots, arranged in a perfect grid. Their smooth, neutral humanoid forms were crafted from a seamless, polished white composite material that caught and reflected the cool, electronic-blue glow. In a single, perfectly synchronized, and silent movement, the entire robot collectively bowed reverently; they knelt on one knee, then the other, and finally, with mechanical grace, bowed forward until their featureless, smooth visors touched the cold, reflective floor. The unified action, executed with such perfect precision, seemed to function as a single, undulating entity, casting long, stark shadows that stretched back into the cavernous darkness, creating an eerie, solemn, atmosphere of digital worship.", + "seed": 5399, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_3/index_2636_seed_5399_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14335803110152484, + "vlm_logprob": -0.14335803110152484, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.14335803110152484, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.33955684304237366 + ], + "opposite_logprob_tokens": [ + -1.3395568132400513 + ], + "logprob_reward": -0.33955684304237366, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.33955684304237366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2580382525920868, + 0.05850723013281822, + 0.07880061119794846, + 0.15825022757053375, + 0.0620909258723259, + 0.03738187998533249, + 0.08862628042697906, + 0.07484730333089828, + 0.1171758845448494, + 0.06628139317035675 + ], + "credit_map": [ + 0.03417748957872391, + 0.04104932025074959, + 0.024311479181051254, + 0.018173594027757645, + 0.03449726849794388, + 0.05822589248418808, + 0.01937159337103367, + 0.02823162078857422, + 0.005778547376394272, + 0.00914599746465683, + 0.005546325817704201, + 0.004237847402691841, + 0.009432589635252953, + 0.008199289441108704, + 0.00831823330372572, + 0.007848396897315979, + 0.011015775613486767, + 0.009191994555294514, + 0.01086045429110527, + 0.010226135142147541, + 0.007496647536754608, + 0.009227607399225235, + 0.010671393014490604, + 0.010110597126185894, + 0.015173434279859066, + 0.028201047331094742, + 0.013597981072962284, + 0.015475839376449585, + 0.025956502184271812, + 0.023758621886372566, + 0.017741555348038673, + 0.018345234915614128, + 0.005717861000448465, + 0.004381077829748392, + 0.016249937936663628, + 0.013781746849417686, + 0.003448879113420844, + 0.006972082424908876, + 0.006261906120926142, + 0.005277434829622507, + 0.004450927954167128, + 0.00504829827696085, + 0.004556313157081604, + 0.006544096861034632, + 0.003465042682364583, + 0.004178616218268871, + 0.004651639610528946, + 0.00448694359511137, + 0.012618870474398136, + 0.00862372387200594, + 0.01780225895345211, + 0.012419986538589, + 0.005040162708610296, + 0.00881220679730177, + 0.010813971050083637, + 0.012495097704231739, + 0.006207033526152372, + 0.004298156127333641, + 0.006914537400007248, + 0.012902596034109592, + 0.010927010327577591, + 0.01538511086255312, + 0.006101781036704779, + 0.012111078016459942, + 0.008038457483053207, + 0.016477419063448906, + 0.013613688759505749, + 0.02058139629662037, + 0.009316404350101948, + 0.020707329735159874, + 0.011677870526909828, + 0.016763318330049515, + 0.005931391846388578, + 0.006163321435451508, + 0.011049500666558743, + 0.010104917921125889, + 0.007874990813434124, + 0.006073727272450924, + 0.008586039766669273, + 0.0104975001886487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.33955684304237366 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813628315925598 + ], + "opposite_logprob_tokens": [ + -1.406362771987915 + ], + "logprob_reward": -0.2813628315925598, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813628315925598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18365433812141418, + 0.10631405562162399, + 0.09239222854375839, + 0.1018749251961708, + 0.05970793589949608, + 0.025890298187732697, + 0.015058853663504124, + 0.13654395937919617, + 0.1004994660615921, + 0.17806391417980194 + ], + "credit_map": [ + 0.0214676596224308, + 0.022563662379980087, + 0.01923263631761074, + 0.017298217862844467, + 0.028411686420440674, + 0.03532714024186134, + 0.01746206171810627, + 0.021891264244914055, + 0.010101920925080776, + 0.016769327223300934, + 0.010813883505761623, + 0.008855819702148438, + 0.013971632346510887, + 0.013749595731496811, + 0.012930165976285934, + 0.01912170648574829, + 0.010429862886667252, + 0.012417496182024479, + 0.012791525572538376, + 0.008413692936301231, + 0.009936876595020294, + 0.011525468900799751, + 0.011997982859611511, + 0.014879326336085796, + 0.012954515404999256, + 0.018305545672774315, + 0.015163276344537735, + 0.012050069868564606, + 0.008386574685573578, + 0.009696545079350471, + 0.010448194108903408, + 0.014870205894112587, + 0.00825183279812336, + 0.008878578431904316, + 0.007091901730746031, + 0.007007654290646315, + 0.005215740296989679, + 0.006681358441710472, + 0.008934623561799526, + 0.007646247278898954, + 0.003079600166529417, + 0.003579085925593972, + 0.003123464062809944, + 0.004469909705221653, + 0.0022639939561486244, + 0.0031235297210514545, + 0.003484843298792839, + 0.0027658711187541485, + 0.001826941268518567, + 0.0013652554480358958, + 0.0027745324186980724, + 0.0019342921441420913, + 0.0012402880238369107, + 0.0014826044207438827, + 0.002438629511743784, + 0.0019963099621236324, + 0.010402867570519447, + 0.009868558496236801, + 0.016641099005937576, + 0.027239536866545677, + 0.015506693162024021, + 0.02539031393826008, + 0.010229995474219322, + 0.02126488834619522, + 0.007305237930268049, + 0.01764998957514763, + 0.008102018386125565, + 0.012178591452538967, + 0.007561936974525452, + 0.02967315912246704, + 0.009069196879863739, + 0.008959333412349224, + 0.015736574307084084, + 0.02064613252878189, + 0.025592481717467308, + 0.027971943840384483, + 0.02382529526948929, + 0.012846113182604313, + 0.01904352568089962, + 0.03240184485912323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.2813628315925598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075406223535538 + ], + "opposite_logprob_tokens": [ + -4.141075611114502 + ], + "logprob_reward": -0.016075406223535538, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075406223535538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23847822844982147, + 0.24515286087989807, + 0.09465296566486359, + 0.04356938228011131, + 0.02193496935069561, + 0.023828327655792236, + 0.02247147634625435, + 0.21635273098945618, + 0.07978267967700958, + 0.013776401057839394 + ], + "credit_map": [ + 0.031931087374687195, + 0.0468418151140213, + 0.021686876192688942, + 0.021014196798205376, + 0.031095970422029495, + 0.042132697999477386, + 0.018849456682801247, + 0.024926135316491127, + 0.036395229399204254, + 0.06832980364561081, + 0.011965114623308182, + 0.013853979296982288, + 0.049861736595630646, + 0.03336348012089729, + 0.013267203234136105, + 0.018116319552063942, + 0.013688945211470127, + 0.013265001587569714, + 0.011503659188747406, + 0.010208574123680592, + 0.011087278835475445, + 0.014490997418761253, + 0.009579656645655632, + 0.010828852653503418, + 0.006381665356457233, + 0.006710059009492397, + 0.004838157445192337, + 0.0040562995709478855, + 0.00553289195522666, + 0.0061258659698069096, + 0.004398729186505079, + 0.005525712855160236, + 0.0020325016230344772, + 0.002212303224951029, + 0.002837278414517641, + 0.0034917122684419155, + 0.001979266991838813, + 0.002410477725788951, + 0.00456204591318965, + 0.002409382024779916, + 0.002932917093858123, + 0.002661433769389987, + 0.0040572104044258595, + 0.0034413975663483143, + 0.0018844747683033347, + 0.0030294156167656183, + 0.0032361543271690607, + 0.002585325390100479, + 0.0025892520789057016, + 0.0018679928034543991, + 0.006211866158992052, + 0.0038030333817005157, + 0.0010125809349119663, + 0.0017516892403364182, + 0.0019075417658314109, + 0.0033275207970291376, + 0.02079762890934944, + 0.014572889544069767, + 0.03131053224205971, + 0.055378902703523636, + 0.013940352946519852, + 0.031791821122169495, + 0.01593855954706669, + 0.032622043043375015, + 0.0061862412840127945, + 0.009336422197520733, + 0.010374851524829865, + 0.01206654030829668, + 0.006358025129884481, + 0.016375143080949783, + 0.00911925733089447, + 0.009966198354959488, + 0.0010736004915088415, + 0.001734961406327784, + 0.002234307350590825, + 0.0021834424696862698, + 0.0017062737606465816, + 0.001312767155468464, + 0.0013490993296727538, + 0.0021819493267685175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.016075406223535538 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786483831703663 + ], + "opposite_logprob_tokens": [ + -4.634786605834961 + ], + "logprob_reward": -0.009786483831703663, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786483831703663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20053893327713013, + 0.06077507510781288, + 0.07394417375326157, + 0.06210174784064293, + 0.07664065808057785, + 0.04576088488101959, + 0.050301384180784225, + 0.24016690254211426, + 0.10521616786718369, + 0.08455411344766617 + ], + "credit_map": [ + 0.0297324787825346, + 0.024037685245275497, + 0.022738473489880562, + 0.015456901863217354, + 0.03229090943932533, + 0.04032827913761139, + 0.0163053460419178, + 0.019648851826786995, + 0.006869670934975147, + 0.009736809879541397, + 0.0052381716668605804, + 0.004587613046169281, + 0.009652337990701199, + 0.009750835597515106, + 0.006349033676087856, + 0.00859060138463974, + 0.008509119041264057, + 0.011344701051712036, + 0.009681591764092445, + 0.006846248637884855, + 0.007453871425241232, + 0.008559197187423706, + 0.010920042172074318, + 0.010629399679601192, + 0.007459814194589853, + 0.011511672288179398, + 0.005061883945018053, + 0.006628448609262705, + 0.0075866966508328915, + 0.009928110055625439, + 0.006249970756471157, + 0.00767514668405056, + 0.006796946283429861, + 0.006016090977936983, + 0.00913894921541214, + 0.010432882234454155, + 0.008218519389629364, + 0.00973503664135933, + 0.017117924988269806, + 0.009184303693473339, + 0.006081614177674055, + 0.004969703499227762, + 0.006529799196869135, + 0.007413607090711594, + 0.003659334732219577, + 0.005368866957724094, + 0.007004569750279188, + 0.004733389243483543, + 0.006502059288322926, + 0.005257848184555769, + 0.008965294808149338, + 0.00619354285299778, + 0.0036710109561681747, + 0.00409062672406435, + 0.009144666604697704, + 0.006476333364844322, + 0.021102698519825935, + 0.014535480178892612, + 0.028979357331991196, + 0.05102331191301346, + 0.026493532583117485, + 0.04239543154835701, + 0.017175253480672836, + 0.0384618379175663, + 0.010852150619029999, + 0.010987760499119759, + 0.013609255664050579, + 0.01660008355975151, + 0.009626338258385658, + 0.01571759395301342, + 0.01425738912075758, + 0.013565589673817158, + 0.007362581789493561, + 0.013229820877313614, + 0.012518336065113544, + 0.009953943081200123, + 0.011868841014802456, + 0.007586849853396416, + 0.007757497485727072, + 0.014276242814958096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.009786483831703663 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle looking down into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly tilting downward and zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a monolithic, glowing server core visible at the far end of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless rows of identical robots with smooth, white humanoid forms stand in a perfect grid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots perform a synchronized movement where they kneel and bow forward until their visors touch the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000859081745148 + ], + "opposite_logprob_tokens": [ + -2.6950085163116455 + ], + "logprob_reward": -0.07000859081745148, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000859081745148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29349833726882935, + 0.0910939946770668, + 0.0991552323102951, + 0.11870723962783813, + 0.024610212072730064, + 0.07275698333978653, + 0.03253583610057831, + 0.1324816793203354, + 0.07568150758743286, + 0.059479009360075 + ], + "credit_map": [ + 0.0426124632358551, + 0.02984754927456379, + 0.02860209345817566, + 0.022275244817137718, + 0.05135778337717056, + 0.06332461535930634, + 0.029141580685973167, + 0.02633698284626007, + 0.008923841640353203, + 0.019103150814771652, + 0.0077166506089270115, + 0.006635061930865049, + 0.016331426799297333, + 0.014018583111464977, + 0.008590052835643291, + 0.00977521575987339, + 0.013107005506753922, + 0.013021334074437618, + 0.01703927479684353, + 0.014736619777977467, + 0.008122299797832966, + 0.010076123289763927, + 0.010200520977377892, + 0.012852051295340061, + 0.014882349409162998, + 0.019111623987555504, + 0.0159380491822958, + 0.012316059321165085, + 0.014288266189396381, + 0.01636713743209839, + 0.011085099540650845, + 0.01471865177154541, + 0.002520449925214052, + 0.002569986041635275, + 0.002656688215211034, + 0.00362158240750432, + 0.0026111057959496975, + 0.003500797087326646, + 0.004223327152431011, + 0.0029062728863209486, + 0.008943697437644005, + 0.006454877555370331, + 0.010752991773188114, + 0.011485748924314976, + 0.005541812162846327, + 0.007789180148392916, + 0.014087536372244358, + 0.0077011361718177795, + 0.004044775385409594, + 0.0025052756536751986, + 0.006097130011767149, + 0.006014333106577396, + 0.0018970322562381625, + 0.0027686897665262222, + 0.0037148436531424522, + 0.005493753589689732, + 0.010767132975161076, + 0.008195054717361927, + 0.01547974906861782, + 0.027668138965964317, + 0.015339367091655731, + 0.025254501029849052, + 0.009263374842703342, + 0.020514363422989845, + 0.006309294141829014, + 0.0077450890094041824, + 0.01129358634352684, + 0.01149671245366335, + 0.007492884993553162, + 0.012219695374369621, + 0.007368588354438543, + 0.01175564806908369, + 0.006329916883260012, + 0.007395354565232992, + 0.009701638482511044, + 0.008947276510298252, + 0.005748791620135307, + 0.004759923554956913, + 0.007946056313812733, + 0.008650046773254871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.07000859081745148 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_4/index_tensor([1190, 1190])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_4/index_tensor([1190, 1190])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d010ee39257997d919d8acf3c5e7f234f4f1eb70 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_4/index_tensor([1190, 1190])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, full-body, eye-level photograph captures a mime performing at the center of an immersive projection theater, rendered in a surreal and clean style. The seamless walls and floor become a continuous canvas, showcasing a swirling, lifelike galaxy in high-fidelity detail; vast clouds of deep indigo and vibrant fuchsia drift slowly among clusters of diamond-sharp stars. The performer's face is painted with flawless white makeup, with stark black accents around his eyes and a single painted teardrop. He wears a fitted matte black turtleneck and pants, and pristine white gloves. His entire body is dramatically illuminated by the shifting, ethereal light of the cosmic projection, casting soft, colorful shadows that move with him. With silent, exaggerated precision, he carefully traces the boundaries of an invisible box, his gloved palm flat against an invisible barrier before him, his fingers splayed in mock despair. Then, he slides his hands sideways, his body tensing, pantomiming the hard surface, his facial expression a concentrated blend of concentration and theatrical panic, creating a visually stunning and emotionally resonant atmosphere of isolated art within the infinite expanse of space.", + "seed": 16716, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_4/index_1190_seed_16716_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03309412850067019, + "vlm_logprob": -0.03309412850067019, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.03309412850067019, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004112359136343002 + ], + "opposite_logprob_tokens": [ + -5.629112243652344 + ], + "logprob_reward": -0.004112359136343002, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.004112359136343002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4024788737297058, + 0.05281880870461464, + 0.05368555709719658, + 0.03526340425014496, + 0.021233011037111282, + 0.12223415821790695, + 0.05986197665333748, + 0.04361364617943764, + 0.11711730062961578, + 0.09169328957796097 + ], + "credit_map": [ + 0.07931303232908249, + 0.011009196750819683, + 0.1640668362379074, + 0.018542438745498657, + 0.013630790635943413, + 0.0839712843298912, + 0.007650025654584169, + 0.024295253679156303, + 0.0045423428528010845, + 0.005088976118713617, + 0.007446566130965948, + 0.005456062965095043, + 0.013013802468776703, + 0.007908652536571026, + 0.003943468444049358, + 0.005418932996690273, + 0.00852519366890192, + 0.005629877094179392, + 0.005425684154033661, + 0.0037990256678313017, + 0.008001592010259628, + 0.013412680476903915, + 0.0028714467771351337, + 0.0060200560837984085, + 0.00407701451331377, + 0.0031375105027109385, + 0.006732909008860588, + 0.0038883774541318417, + 0.0037052109837532043, + 0.006471951026469469, + 0.0022879817988723516, + 0.0049624452367424965, + 0.00279830489307642, + 0.0023941053077578545, + 0.003548311535269022, + 0.0022290307097136974, + 0.0031578221824020147, + 0.0031960599590092897, + 0.0018079959554597735, + 0.002101378981024027, + 0.018274297937750816, + 0.011058568954467773, + 0.024481581524014473, + 0.011388920247554779, + 0.013608066365122795, + 0.024554457515478134, + 0.005169975105673075, + 0.013698283582925797, + 0.006233665160834789, + 0.0069366819225251675, + 0.007129327859729528, + 0.006979287136346102, + 0.014697587117552757, + 0.007704040501266718, + 0.005193626042455435, + 0.004987757187336683, + 0.006259250454604626, + 0.00695981178432703, + 0.011089611798524857, + 0.004156791605055332, + 0.005961373448371887, + 0.0029865854885429144, + 0.003112773410975933, + 0.003087445627897978, + 0.012106657028198242, + 0.013501555658876896, + 0.014223041944205761, + 0.011167111806571484, + 0.028928902000188828, + 0.01687183789908886, + 0.008631209842860699, + 0.011686982586979866, + 0.007491345051676035, + 0.021203381940722466, + 0.016829971224069595, + 0.007044968660920858, + 0.017002316191792488, + 0.008901294320821762, + 0.004673770163208246, + 0.008546236902475357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.004112359136343002 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785421192646027 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.009785421192646027, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785421192646027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6212096214294434, + 0.04371068626642227, + 0.06844187527894974, + 0.02401142753660679, + 0.028111154213547707, + 0.050418488681316376, + 0.055643368512392044, + 0.0171780027449131, + 0.04958578571677208, + 0.041689638048410416 + ], + "credit_map": [ + 0.14616522192955017, + 0.014761079102754593, + 0.2579575181007385, + 0.027568379417061806, + 0.017288517206907272, + 0.10662253946065903, + 0.011576730757951736, + 0.03926955163478851, + 0.0030269301496446133, + 0.00513229938223958, + 0.00930796843022108, + 0.002740264870226383, + 0.004477447364479303, + 0.0129811717197299, + 0.0020712476689368486, + 0.0039733536541461945, + 0.015153458341956139, + 0.005312467459589243, + 0.0061717829667031765, + 0.003603754099458456, + 0.004955760203301907, + 0.02451421320438385, + 0.002379051176831126, + 0.006351380608975887, + 0.0025722479913383722, + 0.003889500629156828, + 0.003807587083429098, + 0.0032445676624774933, + 0.002099431585520506, + 0.002966569736599922, + 0.0018239812925457954, + 0.003607537830248475, + 0.004703141283243895, + 0.003619074122980237, + 0.006746251136064529, + 0.003081419738009572, + 0.00201270985417068, + 0.004392798990011215, + 0.001276734983548522, + 0.002279021544381976, + 0.006778968498110771, + 0.0066234818659722805, + 0.0081150708720088, + 0.0037641001399606466, + 0.005129852332174778, + 0.01238250732421875, + 0.0020940841641277075, + 0.005530413240194321, + 0.00649673817679286, + 0.005243360064923763, + 0.008944910019636154, + 0.00791968498378992, + 0.005917245987802744, + 0.009760481305420399, + 0.0040031843818724155, + 0.007357760798186064, + 0.0025742065627127886, + 0.002911805873736739, + 0.0040335338562726974, + 0.0012829839251935482, + 0.0014479252276942134, + 0.0020896336063742638, + 0.0011957758106291294, + 0.0016421355539932847, + 0.00665639340877533, + 0.007881959900259972, + 0.008192526176571846, + 0.004160871729254723, + 0.007038473151624203, + 0.009082908742129803, + 0.002650239272043109, + 0.003922408912330866, + 0.004457816015928984, + 0.013051029294729233, + 0.007714743725955486, + 0.0023850256111472845, + 0.005117054097354412, + 0.0045128511264920235, + 0.0018418304389342666, + 0.002609283896163106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.009785421192646027 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630487438291311 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630487438291311, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630487438291311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5357037782669067, + 0.1068234071135521, + 0.08873982727527618, + 0.03188798204064369, + 0.024551067501306534, + 0.06814353913068771, + 0.023877190425992012, + 0.02043384499847889, + 0.028571533039212227, + 0.07126782834529877 + ], + "credit_map": [ + 0.1352233588695526, + 0.014989998191595078, + 0.2205272912979126, + 0.024737147614359856, + 0.014232982881367207, + 0.08227802813053131, + 0.010053343139588833, + 0.033661678433418274, + 0.010275129228830338, + 0.010088394396007061, + 0.009784571826457977, + 0.00982972327619791, + 0.030770568177103996, + 0.014390826225280762, + 0.010180962271988392, + 0.011503227055072784, + 0.021967792883515358, + 0.005459120962768793, + 0.007633472327142954, + 0.0035403259098529816, + 0.0052878400310873985, + 0.034529682248830795, + 0.0027360892854630947, + 0.00758550176396966, + 0.004402569029480219, + 0.0029428433626890182, + 0.005584349390119314, + 0.003303205594420433, + 0.0034701430704444647, + 0.006228638347238302, + 0.0019702427089214325, + 0.003985994029790163, + 0.0030874726362526417, + 0.0022841894533485174, + 0.0033852553460747004, + 0.0019099856726825237, + 0.003259404795244336, + 0.00445987842977047, + 0.002140694996342063, + 0.0040241871029138565, + 0.009873351082205772, + 0.006268346216529608, + 0.013400326482951641, + 0.00703794090077281, + 0.007260418031364679, + 0.014170380309224129, + 0.0031770302448421717, + 0.006955746561288834, + 0.0029801269993185997, + 0.0019383840262889862, + 0.003896343056112528, + 0.002855327446013689, + 0.004383077844977379, + 0.0034978340845555067, + 0.0020744013600051403, + 0.0022516967728734016, + 0.0022909962572157383, + 0.002090285997837782, + 0.003473664168268442, + 0.002127718413248658, + 0.0035166332963854074, + 0.0025559368077665567, + 0.0022647585719823837, + 0.0021138505544513464, + 0.003232799004763365, + 0.004608388990163803, + 0.003877257229760289, + 0.002541160210967064, + 0.005584867671132088, + 0.003534030169248581, + 0.0028128130361437798, + 0.002380215097218752, + 0.00912829115986824, + 0.016549382358789444, + 0.02300387993454933, + 0.00410314928740263, + 0.008409696631133556, + 0.004116747062653303, + 0.002057908335700631, + 0.003898776136338711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.007630487438291311 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054940756410360336 + ], + "opposite_logprob_tokens": [ + -2.929940700531006 + ], + "logprob_reward": -0.054940756410360336, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054940756410360336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6845643520355225, + 0.03018260933458805, + 0.03783625736832619, + 0.036977656185626984, + 0.01574370451271534, + 0.05615109205245972, + 0.01899663917720318, + 0.034212443977594376, + 0.05049292370676994, + 0.03484240174293518 + ], + "credit_map": [ + 0.1640392392873764, + 0.018054071813821793, + 0.283476859331131, + 0.029658975079655647, + 0.019272610545158386, + 0.11336637288331985, + 0.013116301968693733, + 0.04357990622520447, + 0.0023831750731915236, + 0.003065364668145776, + 0.0045038675889372826, + 0.0025042016059160233, + 0.0042833443731069565, + 0.007171141915023327, + 0.0021050660870969296, + 0.004166449885815382, + 0.007519679609686136, + 0.00375424325466156, + 0.0042020902037620544, + 0.0023556682281196117, + 0.004174857400357723, + 0.010797670111060143, + 0.0017241504974663258, + 0.0033078964333981276, + 0.005693935323506594, + 0.002974816830828786, + 0.007899848744273186, + 0.0034468527883291245, + 0.004210553131997585, + 0.006151425186544657, + 0.002140549011528492, + 0.004459675867110491, + 0.002724356483668089, + 0.0021428344771265984, + 0.0031379249412566423, + 0.0017614495009183884, + 0.0017103003337979317, + 0.0018387414747849107, + 0.0012029451318085194, + 0.0012251522857695818, + 0.008186683058738708, + 0.0042793783359229565, + 0.010985029861330986, + 0.00506165437400341, + 0.004282334353774786, + 0.00830758735537529, + 0.003451306140050292, + 0.01159711740911007, + 0.0026905618142336607, + 0.0018417617538943887, + 0.003846081206575036, + 0.002682898659259081, + 0.0019150375155732036, + 0.0024808188900351524, + 0.0015553347766399384, + 0.0019841445609927177, + 0.004319836851209402, + 0.005559021141380072, + 0.006598292384296656, + 0.0027337446808815002, + 0.005183642264455557, + 0.00423151720315218, + 0.002315378515049815, + 0.003271013032644987, + 0.006133191753178835, + 0.007207533810287714, + 0.006625817157328129, + 0.0047583929263055325, + 0.011452889069914818, + 0.006430064328014851, + 0.004718694370239973, + 0.0031663384288549423, + 0.0054794857278466225, + 0.006228678859770298, + 0.010904114693403244, + 0.0021506422199308872, + 0.004129753913730383, + 0.002547626383602619, + 0.0013670678017660975, + 0.0020350308623164892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.054940756410360336 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890016183257103 + ], + "opposite_logprob_tokens": [ + -2.4640016555786133 + ], + "logprob_reward": -0.0890016183257103, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890016183257103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13637425005435944, + 0.03906083479523659, + 0.21191145479679108, + 0.031090525910258293, + 0.0259074866771698, + 0.20440757274627686, + 0.21660484373569489, + 0.041761331260204315, + 0.04366268217563629, + 0.04921901226043701 + ], + "credit_map": [ + 0.031336694955825806, + 0.0065074204467237, + 0.05393468961119652, + 0.006435940973460674, + 0.004796668887138367, + 0.021289680153131485, + 0.0032494820188730955, + 0.008823677897453308, + 0.003822221653535962, + 0.0035309316590428352, + 0.004698407370597124, + 0.004339625593274832, + 0.007091776467859745, + 0.0067537324503064156, + 0.0029665117617696524, + 0.00585762457922101, + 0.05560186132788658, + 0.011965667828917503, + 0.016711892560124397, + 0.006948649417608976, + 0.011733106337487698, + 0.08587769418954849, + 0.00528743676841259, + 0.017785126343369484, + 0.0028600033838301897, + 0.0024230375420302153, + 0.006417900323867798, + 0.004601968452334404, + 0.0026912870816886425, + 0.0056330254301428795, + 0.0019492530263960361, + 0.0045140511356294155, + 0.00404656957834959, + 0.0025485544465482235, + 0.00467350659891963, + 0.0022565305698662996, + 0.003129075514152646, + 0.0029549202881753445, + 0.0016975132748484612, + 0.00460081547498703, + 0.026754407212138176, + 0.02126743458211422, + 0.03771017864346504, + 0.014927548356354237, + 0.02035415917634964, + 0.048318587243556976, + 0.008251039311289787, + 0.026824217289686203, + 0.027841342613101006, + 0.017117422074079514, + 0.03276357799768448, + 0.03884251415729523, + 0.02212381921708584, + 0.0289327222853899, + 0.018550930544734, + 0.03043251670897007, + 0.0051600029692053795, + 0.005448224954307079, + 0.007209237664937973, + 0.004736826755106449, + 0.0053021712228655815, + 0.004640047438442707, + 0.0033747402485460043, + 0.005890081636607647, + 0.004515400156378746, + 0.0060503301210701466, + 0.004934169817715883, + 0.004906540736556053, + 0.010762779973447323, + 0.005939376540482044, + 0.0030065509490668774, + 0.0035475343465805054, + 0.008784249424934387, + 0.010993817821145058, + 0.00759248249232769, + 0.0030118413269519806, + 0.0072616771794855595, + 0.006343905348330736, + 0.0017419197829440236, + 0.0034891190007328987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.0890016183257103 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, full-body, eye-level photograph captures a mime performing at the center of an immersive projection theater, rendered in a surreal and clean style. The seamless walls and floor become a continuous canvas, showcasing a swirling, lifelike galaxy in high-fidelity detail; vast clouds of deep indigo and vibrant fuchsia drift slowly among clusters of diamond-sharp stars. The performer's face is painted with flawless white makeup, with stark black accents around his eyes and a single painted teardrop. He wears a fitted matte black turtleneck and pants, and pristine white gloves. His entire body is dramatically illuminated by the shifting, ethereal light of the cosmic projection, casting soft, colorful shadows that move with him. With silent, exaggerated precision, he carefully traces the boundaries of an invisible box, his gloved palm flat against an invisible barrier before him, his fingers splayed in mock despair. Then, he slides his hands sideways, his body tensing, pantomiming the hard surface, his facial expression a concentrated blend of concentration and theatrical panic, creating a visually stunning and emotionally resonant atmosphere of isolated art within the infinite expanse of space.", + "seed": 16717, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_4/index_1190_seed_16717_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2672733074054122, + "vlm_logprob": -0.2672733074054122, + "vlm_yes_no_margin": 2.8999999880790712, + "vlm_token_credit": -0.2672733074054122, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01242162473499775 + ], + "opposite_logprob_tokens": [ + -4.512421607971191 + ], + "logprob_reward": -0.01242162473499775, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01242162473499775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15633277595043182, + 0.0615382082760334, + 0.15554004907608032, + 0.1385529339313507, + 0.08644046634435654, + 0.12349092215299606, + 0.10421392321586609, + 0.08368256688117981, + 0.05482618138194084, + 0.0353819839656353 + ], + "credit_map": [ + 0.016649363562464714, + 0.026430565863847733, + 0.020075548440217972, + 0.019279733300209045, + 0.032580096274614334, + 0.011453898623585701, + 0.01530611701309681, + 0.014557450078427792, + 0.007211713586002588, + 0.010253194719552994, + 0.010439012199640274, + 0.008028630167245865, + 0.008898930624127388, + 0.0056470418348908424, + 0.004474349785596132, + 0.006585330236703157, + 0.018650533631443977, + 0.016574624925851822, + 0.022520840167999268, + 0.015440640971064568, + 0.013765016570687294, + 0.0194342490285635, + 0.01253675390034914, + 0.03661739081144333, + 0.016210272908210754, + 0.01550669502466917, + 0.029297590255737305, + 0.010239864699542522, + 0.020814597606658936, + 0.027027860283851624, + 0.007199026178568602, + 0.012257027439773083, + 0.008922552689909935, + 0.012967877089977264, + 0.009637855924665928, + 0.0077431355603039265, + 0.00933846179395914, + 0.011493989266455173, + 0.006588307674974203, + 0.01974828541278839, + 0.013430959545075893, + 0.01595814898610115, + 0.025951799005270004, + 0.006572750397026539, + 0.02655637077987194, + 0.00911559909582138, + 0.008426502346992493, + 0.01747879385948181, + 0.01426155399531126, + 0.017039090394973755, + 0.014194690622389317, + 0.005949602462351322, + 0.012687805108726025, + 0.011578737758100033, + 0.011275036260485649, + 0.01722741685807705, + 0.012377115897834301, + 0.022550635039806366, + 0.006840875372290611, + 0.004767565056681633, + 0.011934448964893818, + 0.008784908801317215, + 0.005718777887523174, + 0.01070824172347784, + 0.00827974546700716, + 0.008779622614383698, + 0.00730260880663991, + 0.003911423031240702, + 0.005989375989884138, + 0.005396628752350807, + 0.009528253227472305, + 0.005638522561639547, + 0.0038247168995440006, + 0.004076170269399881, + 0.0054176729172468185, + 0.0021853805519640446, + 0.004618105012923479, + 0.004244420677423477, + 0.003189662005752325, + 0.007825857028365135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.01242162473499775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048659320920705795 + ], + "opposite_logprob_tokens": [ + -3.048659324645996 + ], + "logprob_reward": -0.048659320920705795, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048659320920705795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10424021631479263, + 0.06996023654937744, + 0.10391939431428909, + 0.2600245773792267, + 0.06297322362661362, + 0.11686448752880096, + 0.17319120466709137, + 0.03982516750693321, + 0.02083667740225792, + 0.048164837062358856 + ], + "credit_map": [ + 0.02686334028840065, + 0.018925292417407036, + 0.013839379884302616, + 0.009498448111116886, + 0.00909465178847313, + 0.01221400499343872, + 0.005082556512206793, + 0.00872254092246294, + 0.008989858441054821, + 0.012887386605143547, + 0.010070553049445152, + 0.008817697875201702, + 0.009859022684395313, + 0.007883524522185326, + 0.004470741376280785, + 0.006981456186622381, + 0.011858162470161915, + 0.011813884600996971, + 0.01517687551677227, + 0.011025814339518547, + 0.008604830130934715, + 0.011948397383093834, + 0.008551171980798244, + 0.024940256029367447, + 0.03557156026363373, + 0.01757209002971649, + 0.07492782175540924, + 0.014884816482663155, + 0.017555663362145424, + 0.07405661046504974, + 0.010099230334162712, + 0.015356777235865593, + 0.008165841922163963, + 0.01211507897824049, + 0.006523299030959606, + 0.004016231279820204, + 0.007978848181664944, + 0.007462295237928629, + 0.0056531247682869434, + 0.011058511212468147, + 0.011587339453399181, + 0.014002368785440922, + 0.024914048612117767, + 0.0066190315410494804, + 0.02799203060567379, + 0.008946582674980164, + 0.008248835802078247, + 0.014554250985383987, + 0.01015509758144617, + 0.009979129768908024, + 0.0048522730357944965, + 0.005032450892031193, + 0.009706195443868637, + 0.060864102095365524, + 0.00467346515506506, + 0.0679284855723381, + 0.003947611898183823, + 0.00541525986045599, + 0.003978578373789787, + 0.0022186331916600466, + 0.006638261489570141, + 0.0047559430822730064, + 0.002639159094542265, + 0.01023172028362751, + 0.0030361388344317675, + 0.002927483292296529, + 0.0032124826684594154, + 0.0011605891631916165, + 0.0030126364436000586, + 0.0025888613890856504, + 0.00208752928301692, + 0.002810956910252571, + 0.005773675162345171, + 0.007489549461752176, + 0.007192870136350393, + 0.003406799864023924, + 0.0072990842163562775, + 0.005221704952418804, + 0.004903973545879126, + 0.006877180188894272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.048659320920705795 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644540794193745 + ], + "opposite_logprob_tokens": [ + -4.7586445808410645 + ], + "logprob_reward": -0.008644540794193745, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644540794193745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12105293571949005, + 0.032622940838336945, + 0.03554121032357216, + 0.10438501834869385, + 0.0319196917116642, + 0.06452345103025436, + 0.5172544717788696, + 0.04565396159887314, + 0.021365802735090256, + 0.025680528953671455 + ], + "credit_map": [ + 0.015836188569664955, + 0.017425324767827988, + 0.015161675401031971, + 0.0158779788762331, + 0.02202102355659008, + 0.009217019192874432, + 0.01062407623976469, + 0.014889643527567387, + 0.0034663458354771137, + 0.0051236022263765335, + 0.004271544981747866, + 0.0038895762991160154, + 0.00623352313414216, + 0.0026866616681218147, + 0.0023897262290120125, + 0.004561958834528923, + 0.004764472134411335, + 0.0029631892684847116, + 0.005648328457027674, + 0.002551519311964512, + 0.003837971482425928, + 0.004822866525501013, + 0.0027949074283242226, + 0.008157954551279545, + 0.011348378844559193, + 0.015435533598065376, + 0.019968099892139435, + 0.008886521682143211, + 0.020034542307257652, + 0.014816565439105034, + 0.0056882090866565704, + 0.008207169361412525, + 0.003004551399499178, + 0.006081747356802225, + 0.0043008034117519855, + 0.0033807444851845503, + 0.0036495495587587357, + 0.00361340818926692, + 0.0032179616391658783, + 0.004670925438404083, + 0.006856225430965424, + 0.00806722417473793, + 0.012735649012029171, + 0.0033478543628007174, + 0.013366030529141426, + 0.00560426339507103, + 0.00433749333024025, + 0.010208713822066784, + 0.023743731901049614, + 0.017341315746307373, + 0.013110416010022163, + 0.013059738092124462, + 0.023378480225801468, + 0.19966073334217072, + 0.010384069755673409, + 0.21657593548297882, + 0.005651491228491068, + 0.013264684937894344, + 0.003669225610792637, + 0.0025276278611272573, + 0.006233103107661009, + 0.004637670703232288, + 0.0024786971043795347, + 0.007191463373601437, + 0.00324591645039618, + 0.003364135744050145, + 0.002566927345469594, + 0.0011629282962530851, + 0.0024028595071285963, + 0.0023694587871432304, + 0.002263373229652643, + 0.00399020267650485, + 0.002579974476248026, + 0.004692013841122389, + 0.003431204706430435, + 0.0016957102343440056, + 0.0035483227111399174, + 0.002840582048520446, + 0.0018696037586778402, + 0.0050231157802045345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.008644540794193745 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305186867713928 + ], + "opposite_logprob_tokens": [ + -3.1680519580841064 + ], + "logprob_reward": -0.04305186867713928, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305186867713928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1392723023891449, + 0.049938008189201355, + 0.10395049303770065, + 0.14243751764297485, + 0.10292597860097885, + 0.20131312310695648, + 0.14214275777339935, + 0.02612464316189289, + 0.023394666612148285, + 0.06850049644708633 + ], + "credit_map": [ + 0.017048224806785583, + 0.02133183740079403, + 0.01774376630783081, + 0.017305636778473854, + 0.026746662333607674, + 0.01036603283137083, + 0.011484921909868717, + 0.017245225608348846, + 0.00686128344386816, + 0.008864058181643486, + 0.007794106379151344, + 0.005500750616192818, + 0.006938878446817398, + 0.006792164873331785, + 0.0023873590398579836, + 0.004799411166459322, + 0.009730163030326366, + 0.01105555985122919, + 0.015965502709150314, + 0.009054767899215221, + 0.012692931108176708, + 0.014696413651108742, + 0.008280553855001926, + 0.022474611178040504, + 0.017070572823286057, + 0.011767109856009483, + 0.03622463718056679, + 0.008936263620853424, + 0.011914590373635292, + 0.038188543170690536, + 0.0062536802142858505, + 0.012082124128937721, + 0.012073073536157608, + 0.022383078932762146, + 0.013135191053152084, + 0.008003170602023602, + 0.011064096353948116, + 0.012059429660439491, + 0.009904036298394203, + 0.014303901232779026, + 0.02475530467927456, + 0.030317893251776695, + 0.035151343792676926, + 0.010280926711857319, + 0.03462643548846245, + 0.015379180200397968, + 0.0122140571475029, + 0.038587991148233414, + 0.010257722809910774, + 0.007327302824705839, + 0.004647566005587578, + 0.0038095987401902676, + 0.007848871871829033, + 0.048955343663692474, + 0.00619408767670393, + 0.053102266043424606, + 0.0031967454124242067, + 0.005184781271964312, + 0.0029235067777335644, + 0.0012309382436797023, + 0.00278440467081964, + 0.0032632164657115936, + 0.0020192544907331467, + 0.005521797575056553, + 0.003910102881491184, + 0.004433606751263142, + 0.00292630004696548, + 0.001533093280158937, + 0.0026373101864010096, + 0.002172060776501894, + 0.0025063129141926765, + 0.0032758808229118586, + 0.007940655574202538, + 0.008172172121703625, + 0.008033528923988342, + 0.0037363911978900433, + 0.008556239306926727, + 0.009568760171532631, + 0.005222525913268328, + 0.017270224168896675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.04305186867713928 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2235891819000244 + ], + "opposite_logprob_tokens": [ + -0.34858912229537964 + ], + "logprob_reward": -1.2235891819000244, + "yes_no_margin_reward": -0.8750000596046448, + "token_credit_reward": -1.2235891819000244, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15499892830848694, + 0.08342709392309189, + 0.03355848789215088, + 0.20554271340370178, + 0.06821254640817642, + 0.04960742965340614, + 0.2572441101074219, + 0.06479115039110184, + 0.056054335087537766, + 0.026563281193375587 + ], + "credit_map": [ + 0.022627126425504684, + 0.024630529806017876, + 0.019241612404584885, + 0.019619295373558998, + 0.02884155698120594, + 0.01136624626815319, + 0.01343715563416481, + 0.01523540262132883, + 0.007376568857580423, + 0.011413921602070332, + 0.01138822641223669, + 0.009521850384771824, + 0.018154390156269073, + 0.006868091877549887, + 0.007345389574766159, + 0.011358655989170074, + 0.004766650963574648, + 0.00423726812005043, + 0.004836126696318388, + 0.004501239396631718, + 0.003818552242591977, + 0.0037220690865069628, + 0.00293261231854558, + 0.004743968602269888, + 0.02539290115237236, + 0.01365906186401844, + 0.05728015676140785, + 0.010605418123304844, + 0.015952525660395622, + 0.06097795441746712, + 0.007425975985825062, + 0.014248711988329887, + 0.006367666181176901, + 0.013475368730723858, + 0.010398182086646557, + 0.005743440240621567, + 0.008516105823218822, + 0.008124505169689655, + 0.0074895876459777355, + 0.008097697049379349, + 0.006977670360356569, + 0.0078643299639225, + 0.007524792570620775, + 0.0031595572363585234, + 0.005621976684778929, + 0.005391603801399469, + 0.003276087576523423, + 0.009791410528123379, + 0.015789953991770744, + 0.014581006951630116, + 0.00990820862352848, + 0.007422504015266895, + 0.01139278057962656, + 0.09075362980365753, + 0.010524539276957512, + 0.09687148779630661, + 0.008961930871009827, + 0.018194027245044708, + 0.005999340210109949, + 0.0038946480490267277, + 0.006734905298799276, + 0.007639479823410511, + 0.00500614894554019, + 0.008360663428902626, + 0.007988532073795795, + 0.011295875534415245, + 0.006718112155795097, + 0.0036924167070537806, + 0.006732885260134935, + 0.0056272391229867935, + 0.008582865819334984, + 0.005416411906480789, + 0.0033829931635409594, + 0.006394778378307819, + 0.0024628820829093456, + 0.0017711545806378126, + 0.003522941144183278, + 0.0034569408744573593, + 0.0015218064654618502, + 0.004049782641232014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -1.2235891819000244 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, full-body, eye-level photograph captures a mime performing at the center of an immersive projection theater, rendered in a surreal and clean style. The seamless walls and floor become a continuous canvas, showcasing a swirling, lifelike galaxy in high-fidelity detail; vast clouds of deep indigo and vibrant fuchsia drift slowly among clusters of diamond-sharp stars. The performer's face is painted with flawless white makeup, with stark black accents around his eyes and a single painted teardrop. He wears a fitted matte black turtleneck and pants, and pristine white gloves. His entire body is dramatically illuminated by the shifting, ethereal light of the cosmic projection, casting soft, colorful shadows that move with him. With silent, exaggerated precision, he carefully traces the boundaries of an invisible box, his gloved palm flat against an invisible barrier before him, his fingers splayed in mock despair. Then, he slides his hands sideways, his body tensing, pantomiming the hard surface, his facial expression a concentrated blend of concentration and theatrical panic, creating a visually stunning and emotionally resonant atmosphere of isolated art within the infinite expanse of space.", + "seed": 16718, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_4/index_1190_seed_16718_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.029918006528168915, + "vlm_logprob": -0.029918006528168915, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.029918006528168915, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010603162460029125 + ], + "opposite_logprob_tokens": [ + -4.635602951049805 + ], + "logprob_reward": -0.010603162460029125, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010603162460029125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24463605880737305, + 0.030102306976914406, + 0.019179614260792732, + 0.07147880643606186, + 0.07772789150476456, + 0.049981553107500076, + 0.08026765286922455, + 0.04941004887223244, + 0.3429354429244995, + 0.03428053483366966 + ], + "credit_map": [ + 0.02141267992556095, + 0.022652341052889824, + 0.03191544860601425, + 0.018445156514644623, + 0.03418553248047829, + 0.06424115598201752, + 0.018670592457056046, + 0.03311316668987274, + 0.0021203213836997747, + 0.002633092226460576, + 0.006512508727610111, + 0.002296608407050371, + 0.0034259327221661806, + 0.0034117642790079117, + 0.002667247084900737, + 0.00703483447432518, + 0.0024685997050255537, + 0.0027104706969112158, + 0.0023136588279157877, + 0.0016857680166140199, + 0.002407717751339078, + 0.002104449551552534, + 0.0020474628545343876, + 0.0034414895344525576, + 0.010607041418552399, + 0.0057344282977283, + 0.010751013644039631, + 0.008210163563489914, + 0.0062551903538405895, + 0.010992474853992462, + 0.009472720324993134, + 0.009455781430006027, + 0.007795069832354784, + 0.009710531681776047, + 0.009708134457468987, + 0.007734464015811682, + 0.012006321921944618, + 0.013052009977400303, + 0.008852250874042511, + 0.008869113400578499, + 0.007631796877831221, + 0.005157529842108488, + 0.006372401956468821, + 0.003995767328888178, + 0.007166224531829357, + 0.00880640558898449, + 0.004164169076830149, + 0.006687262561172247, + 0.01191499549895525, + 0.004533614031970501, + 0.009962312877178192, + 0.005334206856787205, + 0.007031500339508057, + 0.01223736722022295, + 0.009735917672514915, + 0.019517745822668076, + 0.009043202735483646, + 0.0050125098787248135, + 0.008826018311083317, + 0.004277861677110195, + 0.003196708159521222, + 0.005818508565425873, + 0.002872442128136754, + 0.010362803004682064, + 0.01997494511306286, + 0.006581185385584831, + 0.037536926567554474, + 0.009213477373123169, + 0.02516940049827099, + 0.14559084177017212, + 0.011762888170778751, + 0.08710581064224243, + 0.0028997743502259254, + 0.0010721838334575295, + 0.011148156598210335, + 0.0011638762662187219, + 0.0018139288295060396, + 0.01156072411686182, + 0.0023730802349746227, + 0.002248812932521105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.010603162460029125 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059547401033341885 + ], + "opposite_logprob_tokens": [ + -5.130954742431641 + ], + "logprob_reward": -0.0059547401033341885, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059547401033341885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.170915424823761, + 0.1018279418349266, + 0.032108109444379807, + 0.19017472863197327, + 0.042331621050834656, + 0.05832710489630699, + 0.03574151545763016, + 0.1255459189414978, + 0.2004753053188324, + 0.04255234822630882 + ], + "credit_map": [ + 0.016041351482272148, + 0.02059921808540821, + 0.02497284486889839, + 0.01516256108880043, + 0.02478577382862568, + 0.03126174211502075, + 0.01618114858865738, + 0.021910786628723145, + 0.008134475909173489, + 0.008844826370477676, + 0.019060533493757248, + 0.00903148204088211, + 0.0147322379052639, + 0.010952368378639221, + 0.014402264729142189, + 0.016669753938913345, + 0.0036251924466341734, + 0.004110138863325119, + 0.004160016309469938, + 0.004881745669990778, + 0.004425440914928913, + 0.003708835458382964, + 0.003356063738465309, + 0.0038406741805374622, + 0.0226415004581213, + 0.022563040256500244, + 0.03774545341730118, + 0.01652809977531433, + 0.014415099285542965, + 0.025338586419820786, + 0.016655636951327324, + 0.03428731858730316, + 0.006049595773220062, + 0.004598293919116259, + 0.00847445148974657, + 0.003014977788552642, + 0.004112664144486189, + 0.009442544542253017, + 0.0033716147299855947, + 0.0032674758695065975, + 0.010466801933944225, + 0.004755814094096422, + 0.011200321838259697, + 0.004929115995764732, + 0.00638396805152297, + 0.009669864550232887, + 0.004968173336237669, + 0.005953045561909676, + 0.005006273742765188, + 0.002995599526911974, + 0.005484527442604303, + 0.002623742213472724, + 0.002988020423799753, + 0.006277673412114382, + 0.004978856071829796, + 0.005386820528656244, + 0.012013896368443966, + 0.005563414189964533, + 0.043053045868873596, + 0.00686499522998929, + 0.008133175782859325, + 0.011878014542162418, + 0.005210000090301037, + 0.032829370349645615, + 0.013207227922976017, + 0.005515936762094498, + 0.023270271718502045, + 0.009583818726241589, + 0.016116321086883545, + 0.0777847170829773, + 0.013661547563970089, + 0.041335444897413254, + 0.005203003529459238, + 0.0031012746039777994, + 0.006210200022906065, + 0.0025734156370162964, + 0.006895461585372686, + 0.006332899443805218, + 0.005169524811208248, + 0.007066572085022926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.0059547401033341885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607118360698223 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.01607118360698223, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607118360698223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09232912957668304, + 0.03689662739634514, + 0.012341151013970375, + 0.055372562259435654, + 0.034716371446847916, + 0.04430210217833519, + 0.11009233444929123, + 0.0712900459766388, + 0.45321398973464966, + 0.08944568783044815 + ], + "credit_map": [ + 0.009964312426745892, + 0.009385518729686737, + 0.014513436704874039, + 0.006746834609657526, + 0.013846936635673046, + 0.014701957814395428, + 0.010689187794923782, + 0.01248093694448471, + 0.0022529070265591145, + 0.002483939751982689, + 0.006928844377398491, + 0.0034563750959932804, + 0.003804245498031378, + 0.004489255137741566, + 0.0043425848707556725, + 0.009138474240899086, + 0.0010689260670915246, + 0.0015621291240677238, + 0.0015776043292135, + 0.0015161623014137149, + 0.0015329568414017558, + 0.0014055391075089574, + 0.0018794769421219826, + 0.001798354904167354, + 0.006773774046450853, + 0.004968318156898022, + 0.007607827894389629, + 0.007748466916382313, + 0.007062478922307491, + 0.008442499674856663, + 0.00741642527282238, + 0.005352768115699291, + 0.004934699274599552, + 0.0037452992983162403, + 0.005144785158336163, + 0.003489201655611396, + 0.004197696223855019, + 0.00425142515450716, + 0.004075037315487862, + 0.004878226667642593, + 0.004653021227568388, + 0.0049082934856414795, + 0.005840603727847338, + 0.0046721044927835464, + 0.008084663189947605, + 0.007072057109326124, + 0.005319911986589432, + 0.0037514439318329096, + 0.019696002826094627, + 0.0073189521208405495, + 0.015631895512342453, + 0.005033970810472965, + 0.007793329656124115, + 0.01810690388083458, + 0.008148781023919582, + 0.028362495824694633, + 0.014390337280929089, + 0.006924731656908989, + 0.011271428316831589, + 0.00614144979044795, + 0.00511270621791482, + 0.009779654443264008, + 0.004317037295550108, + 0.013352684676647186, + 0.026301022619009018, + 0.009913908317685127, + 0.05034807324409485, + 0.015004577115178108, + 0.03337159752845764, + 0.1879502683877945, + 0.02048611082136631, + 0.10983841866254807, + 0.008581419475376606, + 0.0028130626305937767, + 0.029886487871408463, + 0.003010358428582549, + 0.0044146133586764336, + 0.03132880479097366, + 0.004389018751680851, + 0.005021916702389717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.01607118360698223 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494481697678566 + ], + "opposite_logprob_tokens": [ + -2.9299447536468506 + ], + "logprob_reward": -0.05494481697678566, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494481697678566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08427957445383072, + 0.01069481298327446, + 0.009891054593026638, + 0.01332370936870575, + 0.013341783545911312, + 0.015386080369353294, + 0.1039169505238533, + 0.026308687403798103, + 0.7050561904907227, + 0.01780112087726593 + ], + "credit_map": [ + 0.007469136267900467, + 0.007477925159037113, + 0.010017814114689827, + 0.006476639769971371, + 0.011140627786517143, + 0.022976357489824295, + 0.005906177684664726, + 0.012814903631806374, + 0.0012183386133983731, + 0.0014014816842973232, + 0.0019492156570777297, + 0.000915739219635725, + 0.0012011367361992598, + 0.0011811770964413881, + 0.0009999347385019064, + 0.0018277897033840418, + 0.001072251470759511, + 0.0009039681171998382, + 0.0012754686176776886, + 0.0014321297639980912, + 0.0009413203224539757, + 0.00104809133335948, + 0.0015428037149831653, + 0.0016750229988247156, + 0.0018146858783438802, + 0.001348966034129262, + 0.001614346751011908, + 0.001963114831596613, + 0.0015920085133984685, + 0.0012462727027013898, + 0.002514124847948551, + 0.0012301899259909987, + 0.0014752557035535574, + 0.0013656516093760729, + 0.002119379583746195, + 0.0013411255786195397, + 0.001458004000596702, + 0.0020648720674216747, + 0.001785398693755269, + 0.0017320974729955196, + 0.002561053028330207, + 0.0012658591149374843, + 0.002493876963853836, + 0.0010016747983172536, + 0.0015381589764729142, + 0.0034331907518208027, + 0.0010261471616104245, + 0.002066120272502303, + 0.02003895305097103, + 0.006313772406429052, + 0.014424017630517483, + 0.005130572244524956, + 0.00726621737703681, + 0.01674230769276619, + 0.008217411115765572, + 0.025783710181713104, + 0.0049285548739135265, + 0.002529148943722248, + 0.006879962049424648, + 0.0018744046101346612, + 0.0018899996066465974, + 0.0039342897944152355, + 0.001455680001527071, + 0.0028166496194899082, + 0.0406024344265461, + 0.01401998195797205, + 0.07980389893054962, + 0.018398957327008247, + 0.051631662994623184, + 0.3013571798801422, + 0.02224813774228096, + 0.17699399590492249, + 0.0022557880729436874, + 0.000575295474845916, + 0.0051529803313314915, + 0.0006759970565326512, + 0.0013733288506045938, + 0.00510274525731802, + 0.00103692093398422, + 0.0016280669951811433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.05494481697678566 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06201612949371338 + ], + "opposite_logprob_tokens": [ + -2.812016010284424 + ], + "logprob_reward": -0.06201612949371338, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06201612949371338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05941972881555557, + 0.0321643128991127, + 0.03289914131164551, + 0.042255949229002, + 0.058829959481954575, + 0.04194140434265137, + 0.1144287958741188, + 0.062234699726104736, + 0.5430496335029602, + 0.012776370160281658 + ], + "credit_map": [ + 0.004334756173193455, + 0.004273544996976852, + 0.006936610676348209, + 0.004568569362163544, + 0.007994726300239563, + 0.014034982770681381, + 0.008025935851037502, + 0.009250604547560215, + 0.0032598725520074368, + 0.0029009401332587004, + 0.004962429869920015, + 0.0025337999686598778, + 0.0035681144800037146, + 0.004066241905093193, + 0.004497184883803129, + 0.0063757277093827724, + 0.003315741429105401, + 0.0032600874546915293, + 0.00449084397405386, + 0.004358097445219755, + 0.004291045479476452, + 0.0035231916699558496, + 0.005310882348567247, + 0.004349254071712494, + 0.005741427652537823, + 0.003269955050200224, + 0.008360338397324085, + 0.004660027101635933, + 0.004578073509037495, + 0.00565434480085969, + 0.003980068489909172, + 0.006011712830513716, + 0.007045526523143053, + 0.006092581897974014, + 0.00828772783279419, + 0.005420598201453686, + 0.008625362999737263, + 0.009242948144674301, + 0.00665312958881259, + 0.007462086621671915, + 0.005286922212690115, + 0.0041051991283893585, + 0.005910736508667469, + 0.0033584972843527794, + 0.005708640906959772, + 0.007944482378661633, + 0.005162072833627462, + 0.004464851226657629, + 0.021386297419667244, + 0.007874658331274986, + 0.014615008607506752, + 0.006580744404345751, + 0.008205716498196125, + 0.02010398916900158, + 0.010129240341484547, + 0.02553313970565796, + 0.005229366943240166, + 0.003718277672305703, + 0.019410759210586548, + 0.004746222868561745, + 0.003848128719255328, + 0.006834155879914761, + 0.004668496549129486, + 0.013779292814433575, + 0.03099558874964714, + 0.012592639774084091, + 0.058830197900533676, + 0.01748628541827202, + 0.04019879549741745, + 0.22695428133010864, + 0.02190234139561653, + 0.13408949971199036, + 0.0014495516661554575, + 0.0007036394672468305, + 0.0035537725780159235, + 0.0005962944705970585, + 0.0015648280968889594, + 0.002750552026554942, + 0.0008528304169885814, + 0.0013049016706645489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.06201612949371338 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, full-body, eye-level photograph captures a mime performing at the center of an immersive projection theater, rendered in a surreal and clean style. The seamless walls and floor become a continuous canvas, showcasing a swirling, lifelike galaxy in high-fidelity detail; vast clouds of deep indigo and vibrant fuchsia drift slowly among clusters of diamond-sharp stars. The performer's face is painted with flawless white makeup, with stark black accents around his eyes and a single painted teardrop. He wears a fitted matte black turtleneck and pants, and pristine white gloves. His entire body is dramatically illuminated by the shifting, ethereal light of the cosmic projection, casting soft, colorful shadows that move with him. With silent, exaggerated precision, he carefully traces the boundaries of an invisible box, his gloved palm flat against an invisible barrier before him, his fingers splayed in mock despair. Then, he slides his hands sideways, his body tensing, pantomiming the hard surface, his facial expression a concentrated blend of concentration and theatrical panic, creating a visually stunning and emotionally resonant atmosphere of isolated art within the infinite expanse of space.", + "seed": 16719, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_4/index_1190_seed_16719_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03335111178457737, + "vlm_logprob": -0.03335111178457737, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.03335111178457737, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01226101815700531 + ], + "opposite_logprob_tokens": [ + -4.512260913848877 + ], + "logprob_reward": -0.01226101815700531, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01226101815700531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08707576990127563, + 0.08467794209718704, + 0.039676204323768616, + 0.056995879858732224, + 0.0386807918548584, + 0.5299749970436096, + 0.043412014842033386, + 0.025405656546354294, + 0.04381518438458443, + 0.05028560385107994 + ], + "credit_map": [ + 0.011378259398043156, + 0.008222795091569424, + 0.011073834262788296, + 0.006597189698368311, + 0.006050313822925091, + 0.01522147934883833, + 0.009341341443359852, + 0.019190555438399315, + 0.008709793910384178, + 0.008423913270235062, + 0.011716953478753567, + 0.009332964196801186, + 0.011358534917235374, + 0.014483093284070492, + 0.010374180041253567, + 0.010278514586389065, + 0.0051351445727050304, + 0.0029044020920991898, + 0.00730217294767499, + 0.0032765462528914213, + 0.0034452390391379595, + 0.00862071942538023, + 0.003627294907346368, + 0.005364681594073772, + 0.006158108357340097, + 0.0033240565098822117, + 0.01920216903090477, + 0.006795382127165794, + 0.004251249600201845, + 0.008603744208812714, + 0.00466893520206213, + 0.003992235753685236, + 0.003966654650866985, + 0.00309580541215837, + 0.005025286227464676, + 0.0022180441301316023, + 0.0077550276182591915, + 0.005483198910951614, + 0.004560774192214012, + 0.006576002575457096, + 0.1590876281261444, + 0.034777648746967316, + 0.13518370687961578, + 0.036681078374385834, + 0.027590690180659294, + 0.07515586912631989, + 0.029123658314347267, + 0.03237469866871834, + 0.0048934523947536945, + 0.0030918808188289404, + 0.008081793785095215, + 0.004871769808232784, + 0.007623820565640926, + 0.00542496656998992, + 0.004930722992867231, + 0.004493610933423042, + 0.0027232319116592407, + 0.0018492713570594788, + 0.0036036199890077114, + 0.003886148799210787, + 0.0043313950300216675, + 0.0037863999605178833, + 0.0026201356668025255, + 0.002605453599244356, + 0.00649177236482501, + 0.003981634043157101, + 0.00390864722430706, + 0.004230451304465532, + 0.0050438144244253635, + 0.009650958701968193, + 0.005771939177066088, + 0.004735969938337803, + 0.0074401977472007275, + 0.003768871771171689, + 0.0030252370052039623, + 0.002689539920538664, + 0.006252841092646122, + 0.007946066558361053, + 0.004531417042016983, + 0.01463143341243267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.01226101815700531 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007635574322193861 + ], + "opposite_logprob_tokens": [ + -4.882635593414307 + ], + "logprob_reward": -0.007635574322193861, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007635574322193861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11896374821662903, + 0.05610366538167, + 0.0326954610645771, + 0.03660616651177406, + 0.04628610610961914, + 0.4984293580055237, + 0.06196656450629234, + 0.09113974124193192, + 0.024894529953598976, + 0.03291459009051323 + ], + "credit_map": [ + 0.013601887971162796, + 0.010166008956730366, + 0.01802303083240986, + 0.005340614356100559, + 0.019934946671128273, + 0.01818513125181198, + 0.011309402994811535, + 0.022402724251151085, + 0.0068164546974003315, + 0.00650459760800004, + 0.007524040061980486, + 0.0040894667617976665, + 0.008243669755756855, + 0.0120884133502841, + 0.004556294996291399, + 0.006280731409788132, + 0.004025283735245466, + 0.004080510698258877, + 0.005351088475435972, + 0.003084517316892743, + 0.0037659630179405212, + 0.006179077550768852, + 0.002340126084163785, + 0.0038688983768224716, + 0.004883747082203627, + 0.002489418024197221, + 0.01074234303086996, + 0.0030451614875346422, + 0.0027084704488515854, + 0.00695844367146492, + 0.0024337777867913246, + 0.0033448082394897938, + 0.004707551095634699, + 0.004181866068392992, + 0.006811394356191158, + 0.0026527310255914927, + 0.007763328030705452, + 0.00835386011749506, + 0.004671583417803049, + 0.007143794558942318, + 0.15447407960891724, + 0.03277616575360298, + 0.12742330133914948, + 0.03484337404370308, + 0.01916695572435856, + 0.07172097265720367, + 0.029999790713191032, + 0.028024744242429733, + 0.009514985606074333, + 0.004186342470347881, + 0.012341848574578762, + 0.008290855213999748, + 0.0069205970503389835, + 0.0065427981317043304, + 0.00784525740891695, + 0.006323882378637791, + 0.014152271673083305, + 0.005625638645142317, + 0.007247475441545248, + 0.013206216506659985, + 0.004819853231310844, + 0.031016454100608826, + 0.0051633273251354694, + 0.009908515028655529, + 0.005528850946575403, + 0.0014358178013935685, + 0.0016522438963875175, + 0.001720230095088482, + 0.002209681086242199, + 0.006496895104646683, + 0.0022860001772642136, + 0.0035648136399686337, + 0.005247105844318867, + 0.0034728895407170057, + 0.002448535058647394, + 0.0019463787321001291, + 0.0024873160291463137, + 0.004950670525431633, + 0.00347507419064641, + 0.00888662226498127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.007635574322193861 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631551939994097 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631551939994097, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631551939994097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13899771869182587, + 0.06049467995762825, + 0.021286914125084877, + 0.042074184864759445, + 0.09229099005460739, + 0.22577659785747528, + 0.0334017239511013, + 0.14952301979064941, + 0.10378306359052658, + 0.13237111270427704 + ], + "credit_map": [ + 0.021738851442933083, + 0.011759454384446144, + 0.01159161888062954, + 0.009254466742277145, + 0.016977112740278244, + 0.01991371251642704, + 0.013638444244861603, + 0.03412405401468277, + 0.006560076028108597, + 0.005716639570891857, + 0.00640102056786418, + 0.006114807445555925, + 0.01076258160173893, + 0.012767275795340538, + 0.0058508035726845264, + 0.00632148003205657, + 0.002610725350677967, + 0.0018865318270400167, + 0.0032384409569203854, + 0.0023528626188635826, + 0.0031996024772524834, + 0.003927211742848158, + 0.001916410750709474, + 0.0021551272366195917, + 0.007692029699683189, + 0.0031454835552722216, + 0.014103449881076813, + 0.002385117346420884, + 0.003529585665091872, + 0.005147411487996578, + 0.002519038738682866, + 0.003552067559212446, + 0.012863202951848507, + 0.008489424362778664, + 0.013613211922347546, + 0.0056006149388849735, + 0.016343288123607635, + 0.011407419107854366, + 0.010872479528188705, + 0.013101345859467983, + 0.06596224755048752, + 0.015671784058213234, + 0.05359429866075516, + 0.01567358709871769, + 0.014575732871890068, + 0.031133869662880898, + 0.01504992600530386, + 0.01411514077335596, + 0.004126833286136389, + 0.0024685002863407135, + 0.006837352178990841, + 0.003311469219624996, + 0.005942466668784618, + 0.004787224344909191, + 0.0023584861773997545, + 0.003569389460608363, + 0.022757554426789284, + 0.007925629615783691, + 0.017529230564832687, + 0.017079437151551247, + 0.008388311602175236, + 0.05138210952281952, + 0.00965438038110733, + 0.01480636652559042, + 0.011303124018013477, + 0.010266666300594807, + 0.009106689132750034, + 0.011617131531238556, + 0.015062037855386734, + 0.019165731966495514, + 0.015568749979138374, + 0.011692937463521957, + 0.022536054253578186, + 0.010433757677674294, + 0.007536640390753746, + 0.005216383375227451, + 0.014725523069500923, + 0.01961269974708557, + 0.011114189401268959, + 0.04119586944580078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.007631551939994097 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026337021961808205 + ], + "opposite_logprob_tokens": [ + -3.651336908340454 + ], + "logprob_reward": -0.026337021961808205, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026337021961808205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07158361375331879, + 0.14858753979206085, + 0.07217205315828323, + 0.0755411684513092, + 0.06181394308805466, + 0.04255234822630882, + 0.03622829169034958, + 0.3904610276222229, + 0.05950850620865822, + 0.041551489382982254 + ], + "credit_map": [ + 0.00899807084351778, + 0.0051492005586624146, + 0.013360974378883839, + 0.008678109385073185, + 0.00697896396741271, + 0.010033573023974895, + 0.008723446168005466, + 0.009661278687417507, + 0.01517224870622158, + 0.014905265532433987, + 0.02024771273136139, + 0.018649766221642494, + 0.01678280718624592, + 0.02605605125427246, + 0.01884675957262516, + 0.017926936969161034, + 0.008716286160051823, + 0.006312988232821226, + 0.016727373003959656, + 0.00762467086315155, + 0.00620043370872736, + 0.008469420485198498, + 0.008993702940642834, + 0.009127175435423851, + 0.007921222597360611, + 0.004799525719136, + 0.02008868008852005, + 0.008333754725754261, + 0.006667478010058403, + 0.010212922468781471, + 0.00829322636127472, + 0.009224357083439827, + 0.006156732328236103, + 0.0072077554650604725, + 0.00725586386397481, + 0.004740648902952671, + 0.007794762495905161, + 0.011740794405341148, + 0.008486070670187473, + 0.008431317284703255, + 0.0068542975932359695, + 0.005196064710617065, + 0.006127294152975082, + 0.004020994063466787, + 0.005532326642423868, + 0.005555830895900726, + 0.005074650049209595, + 0.004190891515463591, + 0.0033577056601643562, + 0.0028749005869030952, + 0.00348376901820302, + 0.004794502630829811, + 0.007829230278730392, + 0.004679444245994091, + 0.0047317976132035255, + 0.004476939793676138, + 0.05459214001893997, + 0.025692246854305267, + 0.022616025060415268, + 0.06964892148971558, + 0.018894366919994354, + 0.13605286180973053, + 0.022016528993844986, + 0.040947940200567245, + 0.007706501055508852, + 0.004816580098122358, + 0.00562557578086853, + 0.006537260953336954, + 0.010931557044386864, + 0.005719399079680443, + 0.009656718000769615, + 0.00851491466164589, + 0.006858683656901121, + 0.0033934128005057573, + 0.0029251493979245424, + 0.002148325787857175, + 0.00430699298158288, + 0.005940601695328951, + 0.004287518095225096, + 0.011690804734826088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.026337021961808205 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mime performing at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a seamless projection theater with walls and floor displaying a swirling galaxy of indigo clouds and fuchsia stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mime wearing a matte black turtleneck, pants, and pristine white gloves with white face makeup and black accents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mime performing the specific action of tracing an invisible box with splayed fingers and sliding hands sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level capturing a static, full-body shot of the performer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11289039254188538 + ], + "opposite_logprob_tokens": [ + -2.2378904819488525 + ], + "logprob_reward": -0.11289039254188538, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11289039254188538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04427246376872063, + 0.025458944961428642, + 0.014158613979816437, + 0.02773989364504814, + 0.039300791919231415, + 0.22645653784275055, + 0.030413668602705002, + 0.5514261722564697, + 0.014595680870115757, + 0.026177290827035904 + ], + "credit_map": [ + 0.005844186991453171, + 0.0037783216685056686, + 0.0075934696942567825, + 0.0026564470026642084, + 0.00602745870128274, + 0.006401960272341967, + 0.004232696257531643, + 0.007737922947853804, + 0.003453147830441594, + 0.0030882793944329023, + 0.0031665312126278877, + 0.0015838221879675984, + 0.004858193453401327, + 0.004119043238461018, + 0.0025882262270897627, + 0.002601698972284794, + 0.001273627276532352, + 0.0016861784970387816, + 0.0018997450824826956, + 0.0011263960041105747, + 0.0018501993035897613, + 0.0029948223382234573, + 0.0016260092379525304, + 0.0017016367055475712, + 0.005452538840472698, + 0.0014125137822702527, + 0.009610943496227264, + 0.001336709945462644, + 0.002277975669130683, + 0.0032720102462917566, + 0.0014073107158765197, + 0.002969890832901001, + 0.003911490086466074, + 0.0034984457306563854, + 0.004345477558672428, + 0.002203674055635929, + 0.006302498746663332, + 0.005706968251615763, + 0.0044381460174918175, + 0.008894090540707111, + 0.06698210537433624, + 0.01424372661858797, + 0.058441609144210815, + 0.015801062807440758, + 0.01180043164640665, + 0.032262664288282394, + 0.012565388344228268, + 0.01435956172645092, + 0.004645817447453737, + 0.002377716125920415, + 0.004977148026227951, + 0.0032264713663607836, + 0.004945479333400726, + 0.003237761789932847, + 0.002845496404916048, + 0.004157778806984425, + 0.09000583738088608, + 0.04278223589062691, + 0.03524380177259445, + 0.1506306380033493, + 0.03637164086103439, + 0.10329746454954147, + 0.05046231672167778, + 0.04263220727443695, + 0.0023750755935907364, + 0.001068169018253684, + 0.0018069557845592499, + 0.0013119624927639961, + 0.0020357624161988497, + 0.003221284132450819, + 0.0015286877751350403, + 0.0012477828422561288, + 0.003842258593067527, + 0.0021345652639865875, + 0.0019119399366900325, + 0.0012578864116221666, + 0.0033048344776034355, + 0.0034407179336994886, + 0.0023595693055540323, + 0.007925518788397312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.11289039254188538 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_5/index_tensor([9753, 9753])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_5/index_tensor([9753, 9753])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c3e4d899ad3387a727cb0ff5c5b1f8ba6d5187ef --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_5/index_tensor([9753, 9753])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up, slow-motion camera zooms in on the silhouette of a mother seated by a window. The cool moonlight filters through the old-fashioned wooden lattice, precisely outlining her slightly hunched figure, her hair tied in a bun. The entire image takes on a rich, ink-wash quality, the transitions of light and shadow flowing naturally like ink on rice paper. Clad in a dark-colored cotton shirt, she stands serenely, intently mending a slightly worn quilt on her lap. Her hands, marked by the marks of time, hold the needle steady and composed. As the camera inched ever closer, focus finally settled on the fine silver needle weaving through the thick fabric, its tip gleaming faintly in the cold moonlight. Each piercing and pulling of the thread was remarkably clear and rhythmic, like a silent ritual. A tranquil, slightly bitter atmosphere permeated the air, as if each stitch and thread wove memories of time and unspoken love.", + "seed": 14660, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_5/index_9753_seed_14660_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1188839789479971, + "vlm_logprob": -0.1188839789479971, + "vlm_yes_no_margin": 2.6499999880790712, + "vlm_token_credit": -0.1188839789479971, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1484825611114502 + ], + "opposite_logprob_tokens": [ + -2.02348256111145 + ], + "logprob_reward": -0.1484825611114502, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1484825611114502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3701222240924835, + 0.06586838513612747, + 0.04615554213523865, + 0.04635895416140556, + 0.04362408071756363, + 0.03075592778623104, + 0.16423505544662476, + 0.05200756713747978, + 0.1262950599193573, + 0.05457727238535881 + ], + "credit_map": [ + 0.024505650624632835, + 0.02098858915269375, + 0.06297590583562851, + 0.05417477712035179, + 0.04783675819635391, + 0.018839098513126373, + 0.06127398833632469, + 0.07952743023633957, + 0.006208876613527536, + 0.007197306491434574, + 0.010384793393313885, + 0.005643898155540228, + 0.012274129316210747, + 0.0048034596256911755, + 0.00949146132916212, + 0.009864455088973045, + 0.003553238231688738, + 0.0033757449127733707, + 0.0076035261154174805, + 0.007694311905652285, + 0.006210341118276119, + 0.0032001545187085867, + 0.007058694027364254, + 0.0074595240876078606, + 0.0063333469443023205, + 0.004994194023311138, + 0.00803404487669468, + 0.0033245752565562725, + 0.005592736881226301, + 0.0049188463017344475, + 0.004424545913934708, + 0.008736658841371536, + 0.00197041523642838, + 0.0016960033681243658, + 0.007338040042668581, + 0.00704751955345273, + 0.006548112723976374, + 0.0036491048522293568, + 0.008781226351857185, + 0.006593653466552496, + 0.0029200606513768435, + 0.002894169883802533, + 0.005120953544974327, + 0.004236449487507343, + 0.004485337529331446, + 0.0022713711950927973, + 0.0025409962981939316, + 0.006286583840847015, + 0.010118825361132622, + 0.010334886610507965, + 0.03037899173796177, + 0.021527674049139023, + 0.02392871305346489, + 0.022173073142766953, + 0.016685213893651962, + 0.029087664559483528, + 0.005641994997859001, + 0.004860189743340015, + 0.00820627436041832, + 0.005604906473308802, + 0.008969800546765327, + 0.007239209953695536, + 0.006177304312586784, + 0.0053078788332641125, + 0.03817540407180786, + 0.018665969371795654, + 0.020118605345487595, + 0.006488974671810865, + 0.010999318212270737, + 0.015324050560593605, + 0.007548848632723093, + 0.008973883464932442, + 0.004899947438389063, + 0.002795490203425288, + 0.008545657619833946, + 0.010173921473324299, + 0.007519925478845835, + 0.00953146442770958, + 0.006364439148455858, + 0.004746418911963701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.1484825611114502 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0431363619863987 + ], + "opposite_logprob_tokens": [ + -3.1681363582611084 + ], + "logprob_reward": -0.0431363619863987, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0431363619863987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19649451971054077, + 0.08066830039024353, + 0.06797993928194046, + 0.055952660739421844, + 0.06961964815855026, + 0.12492774426937103, + 0.061752788722515106, + 0.06253174692392349, + 0.20673434436321259, + 0.07333824783563614 + ], + "credit_map": [ + 0.01347521971911192, + 0.007477901875972748, + 0.03540993481874466, + 0.03290493041276932, + 0.02429870143532753, + 0.010933920741081238, + 0.03181885555386543, + 0.04017506167292595, + 0.005350891500711441, + 0.003631157334893942, + 0.01495584100484848, + 0.011403665877878666, + 0.010518682189285755, + 0.005302806850522757, + 0.012971816584467888, + 0.016533445566892624, + 0.005793515592813492, + 0.0039845784194767475, + 0.01175927184522152, + 0.012439719401299953, + 0.007526454981416464, + 0.003704552538692951, + 0.011589293368160725, + 0.011182554066181183, + 0.005872346926480532, + 0.0036253388971090317, + 0.010516208596527576, + 0.008089404553174973, + 0.006262154318392277, + 0.0038827930111438036, + 0.007632913533598185, + 0.010071506723761559, + 0.0058943298645317554, + 0.003020875621587038, + 0.013003185391426086, + 0.006334033794701099, + 0.011903279460966587, + 0.0050425478257238865, + 0.008060586638748646, + 0.016360817477107048, + 0.009477511048316956, + 0.009481041692197323, + 0.020316731184720993, + 0.017588796094059944, + 0.017265373840928078, + 0.011472715064883232, + 0.012153017334640026, + 0.02717255987226963, + 0.0056169284507632256, + 0.006226920057088137, + 0.009933616034686565, + 0.008236321620643139, + 0.009478751569986343, + 0.007650425657629967, + 0.0067219724878668785, + 0.007887857966125011, + 0.008554687723517418, + 0.0050549884326756, + 0.009890029206871986, + 0.0071197091601789, + 0.009109610691666603, + 0.006931767798960209, + 0.006883954629302025, + 0.00898700300604105, + 0.04956045374274254, + 0.020547224208712578, + 0.02864467166364193, + 0.019095689058303833, + 0.01836872659623623, + 0.028437716886401176, + 0.0219523087143898, + 0.020127572119235992, + 0.006355257239192724, + 0.0037454653065651655, + 0.011942824348807335, + 0.0061458395794034, + 0.00980732124298811, + 0.020612608641386032, + 0.007947051897644997, + 0.0067818909883499146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.0431363619863987 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812266141176224 + ], + "opposite_logprob_tokens": [ + -3.2881226539611816 + ], + "logprob_reward": -0.03812266141176224, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812266141176224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2556847035884857, + 0.09306523948907852, + 0.10314817726612091, + 0.04639912024140358, + 0.07433505356311798, + 0.019292393699288368, + 0.06770762801170349, + 0.08576823770999908, + 0.1591336578130722, + 0.0954657793045044 + ], + "credit_map": [ + 0.02651059441268444, + 0.01909748651087284, + 0.0405002124607563, + 0.03526553511619568, + 0.024586059153079987, + 0.03172755241394043, + 0.04130181670188904, + 0.036695446819067, + 0.010435577481985092, + 0.005504284054040909, + 0.015018569305539131, + 0.008513425476849079, + 0.009905694983899593, + 0.012415492907166481, + 0.013210771605372429, + 0.018061425536870956, + 0.0073694935999810696, + 0.0047265635803341866, + 0.017197974026203156, + 0.021440107375383377, + 0.009261402301490307, + 0.0047899531200528145, + 0.02214415930211544, + 0.016218531876802444, + 0.006183568388223648, + 0.0057090348564088345, + 0.007399882189929485, + 0.0037592852022498846, + 0.006189222913235426, + 0.006507658865302801, + 0.004692792892456055, + 0.005957676563411951, + 0.006686907261610031, + 0.004008883144706488, + 0.012931525707244873, + 0.006299392785876989, + 0.011016258969902992, + 0.0053091938607394695, + 0.012354723177850246, + 0.015728171914815903, + 0.0023044105619192123, + 0.0016629339661449194, + 0.0027281963266432285, + 0.0024151704274117947, + 0.002629881026223302, + 0.002624833257868886, + 0.002520092995837331, + 0.0024068739730864763, + 0.004945776890963316, + 0.007543564308434725, + 0.011723869480192661, + 0.012237190268933773, + 0.008887714706361294, + 0.00776753481477499, + 0.0073081874288618565, + 0.007293784059584141, + 0.008409758098423481, + 0.008717191405594349, + 0.015242327004671097, + 0.008138513192534447, + 0.010063384659588337, + 0.016527492552995682, + 0.008911318145692348, + 0.00975826010107994, + 0.05117505416274071, + 0.023154214024543762, + 0.02504528872668743, + 0.010712842456996441, + 0.01064490806311369, + 0.019269390031695366, + 0.008204692043364048, + 0.01092726830393076, + 0.011586903594434261, + 0.008060907945036888, + 0.012416374869644642, + 0.010632759891450405, + 0.01148991473019123, + 0.023668069392442703, + 0.00951126404106617, + 0.008099588565528393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.03812266141176224 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486073613166809 + ], + "opposite_logprob_tokens": [ + -1.2236073017120361 + ], + "logprob_reward": -0.3486073613166809, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3486073613166809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07707862555980682, + 0.03656468167901039, + 0.023477906361222267, + 0.049436986446380615, + 0.0598866268992424, + 0.047873079776763916, + 0.1218041330575943, + 0.0662773922085762, + 0.26501771807670593, + 0.2525828182697296 + ], + "credit_map": [ + 0.008506995625793934, + 0.005467137787491083, + 0.01303160097450018, + 0.00925840437412262, + 0.010261540301144123, + 0.008569616824388504, + 0.006791058927774429, + 0.015192270278930664, + 0.0037451512180268764, + 0.0023590498603880405, + 0.006832341197878122, + 0.004153369460254908, + 0.005114174447953701, + 0.00259478110820055, + 0.004523577634245157, + 0.007242238614708185, + 0.0019920000340789557, + 0.0014691426185891032, + 0.0037560230121016502, + 0.0031201241072267294, + 0.003464627079665661, + 0.0017036627978086472, + 0.0031888207886368036, + 0.004783505108207464, + 0.004263728391379118, + 0.004200879484415054, + 0.00799619685858488, + 0.0035748411901295185, + 0.005600373260676861, + 0.006779135204851627, + 0.007201376371085644, + 0.009820455685257912, + 0.00565374968573451, + 0.005610354244709015, + 0.010032801888883114, + 0.0069645545445382595, + 0.0106204180046916, + 0.005733828525990248, + 0.007774165365844965, + 0.007496750447899103, + 0.00433580856770277, + 0.0031247443985193968, + 0.00811309739947319, + 0.005327474791556597, + 0.005866659805178642, + 0.0044346158392727375, + 0.008180581033229828, + 0.008490100502967834, + 0.011870368383824825, + 0.02128692716360092, + 0.020450226962566376, + 0.018052224069833755, + 0.014609994366765022, + 0.011141384951770306, + 0.013233073055744171, + 0.011159933172166348, + 0.009991948492825031, + 0.00550922192633152, + 0.011003421619534492, + 0.004974475596100092, + 0.011185717768967152, + 0.007141751237213612, + 0.0059232632629573345, + 0.010547592304646969, + 0.09453590959310532, + 0.023759586736559868, + 0.04407771676778793, + 0.017661139369010925, + 0.017758384346961975, + 0.030279936268925667, + 0.019039828330278397, + 0.01790522038936615, + 0.019987119361758232, + 0.014522031880915165, + 0.031368229538202286, + 0.021178975701332092, + 0.0370294488966465, + 0.06690336763858795, + 0.029985880479216576, + 0.03160776197910309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.3486073613166809 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070948913693428 + ], + "opposite_logprob_tokens": [ + -4.14107084274292 + ], + "logprob_reward": -0.016070948913693428, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070948913693428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1503240019083023, + 0.07512559741735458, + 0.05076010897755623, + 0.024891367182135582, + 0.06137147545814514, + 0.02862563170492649, + 0.0843685194849968, + 0.10602293163537979, + 0.34939414262771606, + 0.06911622732877731 + ], + "credit_map": [ + 0.015204035677015781, + 0.008985915221273899, + 0.025687433779239655, + 0.015095931477844715, + 0.021966824308037758, + 0.01005135104060173, + 0.019543970003724098, + 0.03378854691982269, + 0.005283381324261427, + 0.0042015984654426575, + 0.013033545576035976, + 0.00794798880815506, + 0.012916618958115578, + 0.006981497630476952, + 0.01125903520733118, + 0.013501929119229317, + 0.003188452450558543, + 0.0022336659021675587, + 0.009223445318639278, + 0.008366990834474564, + 0.006819804199039936, + 0.003870537271723151, + 0.008290126919746399, + 0.008767086081206799, + 0.0025663767009973526, + 0.0021121848840266466, + 0.004178812727332115, + 0.00217379885725677, + 0.002609746064990759, + 0.002227027900516987, + 0.0041267708875238895, + 0.0048966496251523495, + 0.007854720577597618, + 0.004781483672559261, + 0.009927603416144848, + 0.00781730655580759, + 0.007279621437191963, + 0.005378715228289366, + 0.008802594617009163, + 0.009529427625238895, + 0.0034870230592787266, + 0.00232733809389174, + 0.004155087284743786, + 0.0030626256484538317, + 0.003925655968487263, + 0.0027122925966978073, + 0.005238555371761322, + 0.0037170532159507275, + 0.006653970573097467, + 0.009381900541484356, + 0.013227642513811588, + 0.011968141421675682, + 0.011602476239204407, + 0.012861853465437889, + 0.007203878369182348, + 0.011468661017715931, + 0.01587141491472721, + 0.010875298641622066, + 0.018022600561380386, + 0.011757386848330498, + 0.014452979899942875, + 0.012363775633275509, + 0.008855137974023819, + 0.013824338093400002, + 0.11454758793115616, + 0.02885730005800724, + 0.057941898703575134, + 0.019517075270414352, + 0.030062047764658928, + 0.044645681977272034, + 0.026951314881443977, + 0.026871241629123688, + 0.006674144882708788, + 0.005688029807060957, + 0.010513318702578545, + 0.008677341043949127, + 0.00794387236237526, + 0.014091872610151768, + 0.00694869551807642, + 0.008578957058489323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.016070948913693428 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up, slow-motion camera zooms in on the silhouette of a mother seated by a window. The cool moonlight filters through the old-fashioned wooden lattice, precisely outlining her slightly hunched figure, her hair tied in a bun. The entire image takes on a rich, ink-wash quality, the transitions of light and shadow flowing naturally like ink on rice paper. Clad in a dark-colored cotton shirt, she stands serenely, intently mending a slightly worn quilt on her lap. Her hands, marked by the marks of time, hold the needle steady and composed. As the camera inched ever closer, focus finally settled on the fine silver needle weaving through the thick fabric, its tip gleaming faintly in the cold moonlight. Each piercing and pulling of the thread was remarkably clear and rhythmic, like a silent ritual. A tranquil, slightly bitter atmosphere permeated the air, as if each stitch and thread wove memories of time and unspoken love.", + "seed": 14661, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_5/index_9753_seed_14661_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05024067163467407, + "vlm_logprob": -0.05024067163467407, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.05024067163467407, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05541805550456047 + ], + "opposite_logprob_tokens": [ + -3.055418014526367 + ], + "logprob_reward": -0.05541805550456047, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05541805550456047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18011029064655304, + 0.06349009275436401, + 0.07828925549983978, + 0.04550349712371826, + 0.1334334909915924, + 0.12345890700817108, + 0.042650774121284485, + 0.1407942771911621, + 0.12061016261577606, + 0.07165923714637756 + ], + "credit_map": [ + 0.026384081691503525, + 0.03760397434234619, + 0.028854798525571823, + 0.028685281053185463, + 0.01616719737648964, + 0.00804367195814848, + 0.017594024538993835, + 0.016777273267507553, + 0.008716730400919914, + 0.01541146356612444, + 0.016161222010850906, + 0.007214790675789118, + 0.004486472345888615, + 0.002349516609683633, + 0.0040240040980279446, + 0.005125900264829397, + 0.012158594094216824, + 0.015345959924161434, + 0.014306269586086273, + 0.010795127600431442, + 0.009969811886548996, + 0.0037843824829906225, + 0.006653306074440479, + 0.005275812931358814, + 0.005605635233223438, + 0.010359677486121655, + 0.007943391799926758, + 0.005264181178063154, + 0.004929801914840937, + 0.002367530483752489, + 0.004441075958311558, + 0.004592211451381445, + 0.03843504562973976, + 0.018431514501571655, + 0.034749194979667664, + 0.006465781945735216, + 0.008957397192716599, + 0.005077044013887644, + 0.011485842056572437, + 0.009831666946411133, + 0.026233751326799393, + 0.01914461888372898, + 0.024995125830173492, + 0.014503619633615017, + 0.0068479967303574085, + 0.004122718703001738, + 0.015254541300237179, + 0.012356556951999664, + 0.006468809675425291, + 0.007498035207390785, + 0.007371319457888603, + 0.006786870304495096, + 0.003965180367231369, + 0.0017572266515344381, + 0.004540850408375263, + 0.00426248786970973, + 0.02604822814464569, + 0.026467064395546913, + 0.0209963358938694, + 0.016062607988715172, + 0.016712067648768425, + 0.006744393613189459, + 0.014717602171003819, + 0.013046005740761757, + 0.025204692035913467, + 0.01861037127673626, + 0.015431071631610394, + 0.009757643565535545, + 0.010056935250759125, + 0.01361106801778078, + 0.01570550538599491, + 0.012232891283929348, + 0.009599816054105759, + 0.01084663812071085, + 0.009563020430505276, + 0.009914644993841648, + 0.007878389209508896, + 0.006394754163920879, + 0.009101762436330318, + 0.008360227569937706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.05541805550456047 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.089200958609581 + ], + "opposite_logprob_tokens": [ + -2.464200973510742 + ], + "logprob_reward": -0.089200958609581, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.089200958609581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25828978419303894, + 0.04242204502224922, + 0.05745222046971321, + 0.06874971091747284, + 0.0755753293633461, + 0.169942706823349, + 0.05219261720776558, + 0.14654487371444702, + 0.07669385522603989, + 0.05213683843612671 + ], + "credit_map": [ + 0.034753795713186264, + 0.05806123837828636, + 0.04610290378332138, + 0.0494113489985466, + 0.014566853642463684, + 0.009453810751438141, + 0.022935492917895317, + 0.023004325106739998, + 0.006343154702335596, + 0.008753704838454723, + 0.006168203428387642, + 0.00727890245616436, + 0.002704489743337035, + 0.0027817334048449993, + 0.0031928797252476215, + 0.005198977887630463, + 0.010661281645298004, + 0.011116022244095802, + 0.009461872279644012, + 0.005482386332005262, + 0.007027792278677225, + 0.0026776567101478577, + 0.007527704816311598, + 0.0034975057933479548, + 0.009745808318257332, + 0.015476655215024948, + 0.011407334357500076, + 0.005941057577729225, + 0.008116496726870537, + 0.00404283357784152, + 0.008036048151552677, + 0.005983477924019098, + 0.016580386087298393, + 0.012870684266090393, + 0.016264161095023155, + 0.009029986336827278, + 0.00493462011218071, + 0.0034780369605869055, + 0.007355438079684973, + 0.00506202457472682, + 0.02916804701089859, + 0.02845628187060356, + 0.03596503287553787, + 0.016478853300213814, + 0.018656602129340172, + 0.0063011799938976765, + 0.017567167058587074, + 0.017349541187286377, + 0.009145745076239109, + 0.008731560781598091, + 0.010735067538917065, + 0.006726921070367098, + 0.00430179201066494, + 0.001913981861434877, + 0.004824408330023289, + 0.005813136231154203, + 0.02147611230611801, + 0.02859865128993988, + 0.02401861920952797, + 0.02194635570049286, + 0.019329959526658058, + 0.005547564011067152, + 0.01418975181877613, + 0.011437864042818546, + 0.0209217332303524, + 0.007443454582244158, + 0.013500350527465343, + 0.009740377776324749, + 0.004144236445426941, + 0.006364195607602596, + 0.005059490446001291, + 0.00952001940459013, + 0.007914806716144085, + 0.007888898253440857, + 0.006220752373337746, + 0.009035653434693813, + 0.006479928735643625, + 0.0038289669901132584, + 0.0056344568729400635, + 0.005133376456797123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.089200958609581 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033721037209033966 + ], + "opposite_logprob_tokens": [ + -3.4087209701538086 + ], + "logprob_reward": -0.033721037209033966, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033721037209033966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11075623333454132, + 0.06093771010637283, + 0.059362296015024185, + 0.0642242580652237, + 0.22165925800800323, + 0.2040800303220749, + 0.05828721076250076, + 0.07730410248041153, + 0.04906082525849342, + 0.09432811290025711 + ], + "credit_map": [ + 0.022357823327183723, + 0.022449960932135582, + 0.018608922138810158, + 0.016921907663345337, + 0.007995260879397392, + 0.005165854003280401, + 0.008379467763006687, + 0.008877034299075603, + 0.009805604815483093, + 0.010622259229421616, + 0.0103291692212224, + 0.007160909939557314, + 0.00800394918769598, + 0.0032146696466952562, + 0.005872681271284819, + 0.005928471684455872, + 0.01175034698098898, + 0.01089423056691885, + 0.011077545583248138, + 0.007363737560808659, + 0.005005732644349337, + 0.00246853893622756, + 0.006655104923993349, + 0.004147056024521589, + 0.009478664956986904, + 0.011956931091845036, + 0.010711845010519028, + 0.008402238599956036, + 0.006357366684824228, + 0.0037276616785675287, + 0.007213110104203224, + 0.006376437842845917, + 0.06034740060567856, + 0.027693066745996475, + 0.05826807767152786, + 0.015198254957795143, + 0.015529136173427105, + 0.00822965707629919, + 0.020959816873073578, + 0.015433842316269875, + 0.03605860844254494, + 0.03195321932435036, + 0.04187683016061783, + 0.021347424015402794, + 0.023469382897019386, + 0.00874283630400896, + 0.0235757976770401, + 0.017055917531251907, + 0.009744352661073208, + 0.010889319702982903, + 0.010354544967412949, + 0.00821409747004509, + 0.004628836177289486, + 0.002003628760576248, + 0.0076036169193685055, + 0.004848811775445938, + 0.013658826239407063, + 0.01434541866183281, + 0.009155947715044022, + 0.00910441018640995, + 0.008879040367901325, + 0.003202789695933461, + 0.011422617360949516, + 0.007535052951425314, + 0.008004346862435341, + 0.006016990169882774, + 0.008804173208773136, + 0.00959254801273346, + 0.003391602775081992, + 0.0041535296477377415, + 0.0054773264564573765, + 0.003620309755206108, + 0.013456467539072037, + 0.013025261461734772, + 0.01351948082447052, + 0.014352370984852314, + 0.01154464203864336, + 0.00580209493637085, + 0.011635088361799717, + 0.010992711409926414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.033721037209033966 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305746406316757 + ], + "opposite_logprob_tokens": [ + -3.168057441711426 + ], + "logprob_reward": -0.04305746406316757, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305746406316757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08925531804561615, + 0.09834647178649902, + 0.08910509943962097, + 0.10133296996355057, + 0.09282314032316208, + 0.09212026745080948, + 0.06439833343029022, + 0.08568084985017776, + 0.2416725903749466, + 0.045264896005392075 + ], + "credit_map": [ + 0.017499836161732674, + 0.02208959124982357, + 0.013154243119060993, + 0.012892888858914375, + 0.005911226850003004, + 0.004979638382792473, + 0.006125548388808966, + 0.006602343171834946, + 0.013001294806599617, + 0.026071542873978615, + 0.02826189436018467, + 0.01083772350102663, + 0.005205454770475626, + 0.003310883417725563, + 0.006826732773333788, + 0.004830945748835802, + 0.01770860329270363, + 0.014617427252233028, + 0.01277106162160635, + 0.0110098235309124, + 0.012516058050096035, + 0.0044103884138166904, + 0.009903926402330399, + 0.006167811341583729, + 0.012057431973516941, + 0.01753333956003189, + 0.01448542159050703, + 0.019123191013932228, + 0.010512938722968102, + 0.00460538174957037, + 0.012851362116634846, + 0.01016390509903431, + 0.026729444041848183, + 0.010290577076375484, + 0.025278452783823013, + 0.006638461258262396, + 0.007209213450551033, + 0.003044228535145521, + 0.008407844230532646, + 0.005224910564720631, + 0.015336341224610806, + 0.010963723994791508, + 0.02146599069237709, + 0.01012369990348816, + 0.007635521702468395, + 0.005263667553663254, + 0.012930653989315033, + 0.008400672115385532, + 0.009737789630889893, + 0.013456144370138645, + 0.012287170626223087, + 0.005869917571544647, + 0.00669204443693161, + 0.003309786319732666, + 0.00568398367613554, + 0.007361497264355421, + 0.01381151843816042, + 0.020924801006913185, + 0.011134584434330463, + 0.011155813932418823, + 0.006776215974241495, + 0.003746133064851165, + 0.01161355059593916, + 0.006518233567476273, + 0.03903477266430855, + 0.03351011499762535, + 0.025250714272260666, + 0.04118393361568451, + 0.021303584799170494, + 0.02722465805709362, + 0.029172291979193687, + 0.024992508813738823, + 0.007649035193026066, + 0.005988378543406725, + 0.004407923202961683, + 0.005638050381094217, + 0.0053454008884727955, + 0.003359987633302808, + 0.006357038393616676, + 0.0065190838649868965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 1 + ], + "selected_reward": -0.04305746406316757 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980584278702736 + ], + "opposite_logprob_tokens": [ + -3.529805898666382 + ], + "logprob_reward": -0.02980584278702736, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980584278702736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26555320620536804, + 0.032277610152959824, + 0.14720036089420319, + 0.04840158298611641, + 0.13921420276165009, + 0.10338552296161652, + 0.05629577115178108, + 0.14376690983772278, + 0.033169157803058624, + 0.030735723674297333 + ], + "credit_map": [ + 0.055936966091394424, + 0.05786935240030289, + 0.046075012534856796, + 0.034892160445451736, + 0.016673997044563293, + 0.011496370658278465, + 0.018726445734500885, + 0.02388289012014866, + 0.006882428657263517, + 0.005640631075948477, + 0.005347886122763157, + 0.004673824645578861, + 0.0022075276356190443, + 0.0015870225615799427, + 0.0028895705472677946, + 0.003048716112971306, + 0.02343904972076416, + 0.02741999737918377, + 0.024802222847938538, + 0.014162478037178516, + 0.01682315766811371, + 0.006600586231797934, + 0.023950055241584778, + 0.010002789087593555, + 0.011058026924729347, + 0.009054681286215782, + 0.007575348950922489, + 0.004221128299832344, + 0.0042207869701087475, + 0.0028491432312875986, + 0.004868857562541962, + 0.004553604405373335, + 0.044356171041727066, + 0.015025457367300987, + 0.04141966998577118, + 0.0075546675361692905, + 0.00776447681710124, + 0.004287851043045521, + 0.011137954890727997, + 0.007667939644306898, + 0.01709655486047268, + 0.014806563965976238, + 0.021770382300019264, + 0.009781130589544773, + 0.008870111778378487, + 0.004702501930296421, + 0.016007956117391586, + 0.010350314900279045, + 0.009600433520972729, + 0.008832035586237907, + 0.009356259368360043, + 0.008946369402110577, + 0.004302744287997484, + 0.0023788241669535637, + 0.0075255283154547215, + 0.005353571847081184, + 0.02386936917901039, + 0.025901591405272484, + 0.025351060554385185, + 0.01600014977157116, + 0.01926587149500847, + 0.008643453009426594, + 0.011364823207259178, + 0.013370580971240997, + 0.005677828565239906, + 0.006028265692293644, + 0.0049308598972857, + 0.003581360913813114, + 0.0029235626570880413, + 0.002346331486478448, + 0.004792008548974991, + 0.0028889363165944815, + 0.004552717320621014, + 0.003958316519856453, + 0.0038274305406957865, + 0.002823740476742387, + 0.0047880480997264385, + 0.003092969534918666, + 0.0034930938854813576, + 0.004199406132102013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.02980584278702736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up, slow-motion camera zooms in on the silhouette of a mother seated by a window. The cool moonlight filters through the old-fashioned wooden lattice, precisely outlining her slightly hunched figure, her hair tied in a bun. The entire image takes on a rich, ink-wash quality, the transitions of light and shadow flowing naturally like ink on rice paper. Clad in a dark-colored cotton shirt, she stands serenely, intently mending a slightly worn quilt on her lap. Her hands, marked by the marks of time, hold the needle steady and composed. As the camera inched ever closer, focus finally settled on the fine silver needle weaving through the thick fabric, its tip gleaming faintly in the cold moonlight. Each piercing and pulling of the thread was remarkably clear and rhythmic, like a silent ritual. A tranquil, slightly bitter atmosphere permeated the air, as if each stitch and thread wove memories of time and unspoken love.", + "seed": 14662, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_5/index_9753_seed_14662_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03880801759660244, + "vlm_logprob": -0.03880801759660244, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.03880801759660244, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034758154302835464 + ], + "opposite_logprob_tokens": [ + -3.4097580909729004 + ], + "logprob_reward": -0.034758154302835464, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.034758154302835464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34241074323654175, + 0.031633250415325165, + 0.039078231900930405, + 0.14799344539642334, + 0.07820291817188263, + 0.042549580335617065, + 0.06128138303756714, + 0.08494362980127335, + 0.053605418652296066, + 0.11830141395330429 + ], + "credit_map": [ + 0.016549047082662582, + 0.012484326958656311, + 0.04522278532385826, + 0.08358102291822433, + 0.03788535296916962, + 0.008689380250871181, + 0.08446051180362701, + 0.05353833734989166, + 0.0027410576585680246, + 0.0018134901765733957, + 0.0046314094215631485, + 0.005266720894724131, + 0.004067385569214821, + 0.001882844022475183, + 0.006270998157560825, + 0.004959344398230314, + 0.002694888273254037, + 0.002660903846845031, + 0.006456557661294937, + 0.00791962817311287, + 0.0055406061001122, + 0.0022116294130682945, + 0.006401095073670149, + 0.005192925687879324, + 0.021715698763728142, + 0.011811246164143085, + 0.01922960765659809, + 0.0178407970815897, + 0.02427312359213829, + 0.013131005689501762, + 0.017656562849879265, + 0.02233540080487728, + 0.0031380360014736652, + 0.005267653614282608, + 0.015219805762171745, + 0.012264842167496681, + 0.014364202506840229, + 0.004963292274624109, + 0.012737639248371124, + 0.010247446596622467, + 0.0032634269446134567, + 0.002736226422712207, + 0.00555563485249877, + 0.0047398945316672325, + 0.009047912433743477, + 0.0026781216729432344, + 0.007507511880248785, + 0.00702085392549634, + 0.006124131381511688, + 0.0055819242261350155, + 0.006933548953384161, + 0.0058385333977639675, + 0.01285400427877903, + 0.007874633185565472, + 0.006428332068026066, + 0.009646283462643623, + 0.007452988065779209, + 0.009283930994570255, + 0.007950830273330212, + 0.006886437069624662, + 0.019822027534246445, + 0.009034382179379463, + 0.007907640188932419, + 0.016605399549007416, + 0.006264748517423868, + 0.006414500530809164, + 0.006296624895185232, + 0.005700233392417431, + 0.008436993695795536, + 0.007442502770572901, + 0.006346690468490124, + 0.006703127641230822, + 0.008978348225355148, + 0.00963820144534111, + 0.015110249631106853, + 0.013357092626392841, + 0.027385825291275978, + 0.014892923645675182, + 0.012821600772440434, + 0.01611717790365219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.034758154302835464 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07908307015895844 + ], + "opposite_logprob_tokens": [ + -2.57908296585083 + ], + "logprob_reward": -0.07908307015895844, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07908307015895844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1772320568561554, + 0.056965362280607224, + 0.048047516494989395, + 0.12338932603597641, + 0.15951253473758698, + 0.02824888750910759, + 0.07401146739721298, + 0.07827302068471909, + 0.1033981442451477, + 0.15092170238494873 + ], + "credit_map": [ + 0.008111759088933468, + 0.009562250226736069, + 0.024091515690088272, + 0.03624650835990906, + 0.01971123367547989, + 0.007002143654972315, + 0.04075084626674652, + 0.031755805015563965, + 0.002547056647017598, + 0.002151224995031953, + 0.010432248935103416, + 0.011078842915594578, + 0.005479529500007629, + 0.0028936718590557575, + 0.011332837864756584, + 0.011049949564039707, + 0.003938918933272362, + 0.002036870224401355, + 0.009373468346893787, + 0.010054275393486023, + 0.005767839960753918, + 0.0020505967549979687, + 0.007197528146207333, + 0.0076280198991298676, + 0.01538001000881195, + 0.007941051386296749, + 0.016299381852149963, + 0.014032886363565922, + 0.020699014887213707, + 0.010866557247936726, + 0.01708155684173107, + 0.021088872104883194, + 0.007639538496732712, + 0.010654188692569733, + 0.02690003253519535, + 0.018733760342001915, + 0.030065705999732018, + 0.01760636642575264, + 0.022376850247383118, + 0.025536110624670982, + 0.0016244786093011498, + 0.001696246094070375, + 0.004170068074017763, + 0.002722454257309437, + 0.006723752710968256, + 0.002707233652472496, + 0.004180573858320713, + 0.004424080718308687, + 0.006506168749183416, + 0.0060803769156336784, + 0.0102748554199934, + 0.007514492142945528, + 0.015980402007699013, + 0.00645511643961072, + 0.009330101311206818, + 0.011869953013956547, + 0.010694033466279507, + 0.007831962779164314, + 0.005953970365226269, + 0.007245182991027832, + 0.01902439258992672, + 0.008139346726238728, + 0.007094904780387878, + 0.012289229780435562, + 0.008925791829824448, + 0.00945322960615158, + 0.014947188086807728, + 0.00995936244726181, + 0.021997850388288498, + 0.011484040878713131, + 0.011938244104385376, + 0.01469243410974741, + 0.013960649259388447, + 0.013538820669054985, + 0.02127738855779171, + 0.013536971062421799, + 0.03726142644882202, + 0.021380893886089325, + 0.010690242983400822, + 0.019275307655334473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.07908307015895844 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029819726943969727 + ], + "opposite_logprob_tokens": [ + -3.5298197269439697 + ], + "logprob_reward": -0.029819726943969727, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029819726943969727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11727989464998245, + 0.027850190177559853, + 0.05003434792160988, + 0.45139971375465393, + 0.0409359447658062, + 0.03966688737273216, + 0.04572727903723717, + 0.05168062448501587, + 0.05717386677861214, + 0.11825131624937057 + ], + "credit_map": [ + 0.006671186536550522, + 0.0060339574702084064, + 0.016105981543660164, + 0.024628886952996254, + 0.013394434936344624, + 0.0046441880986094475, + 0.027332542464137077, + 0.018468696624040604, + 0.002466160338371992, + 0.0018795443465933204, + 0.0047487919218838215, + 0.0036631810944527388, + 0.0030207450035959482, + 0.0021263905800879, + 0.0048989104107022285, + 0.00504646310582757, + 0.005594990216195583, + 0.0026504213456064463, + 0.008386536501348019, + 0.009761362336575985, + 0.0053149983286857605, + 0.002620850456878543, + 0.008848972618579865, + 0.0068562086671590805, + 0.09312471002340317, + 0.038990892469882965, + 0.05815386772155762, + 0.041666075587272644, + 0.056382834911346436, + 0.05042380094528198, + 0.047013476490974426, + 0.06564400345087051, + 0.0019516611937433481, + 0.0016609260346740484, + 0.0073324162513017654, + 0.0038414113223552704, + 0.009305424056947231, + 0.0036915966775268316, + 0.0064583998173475266, + 0.006694103591144085, + 0.0025704526342451572, + 0.0027537106070667505, + 0.005000336095690727, + 0.003395410953089595, + 0.009483558125793934, + 0.0038892945740371943, + 0.005368693266063929, + 0.007205425761640072, + 0.004108864348381758, + 0.0037363676819950342, + 0.004893095698207617, + 0.0034655313938856125, + 0.009920851327478886, + 0.0038123601116240025, + 0.007950904779136181, + 0.007839296013116837, + 0.003943600691854954, + 0.005268608685582876, + 0.006992881651967764, + 0.005650810431689024, + 0.006106148473918438, + 0.004652633797377348, + 0.010902820155024529, + 0.00816311128437519, + 0.00355414766818285, + 0.004153124056756496, + 0.008640436455607414, + 0.007271727081388235, + 0.010228833183646202, + 0.00630394509062171, + 0.008721735328435898, + 0.008299903944134712, + 0.011335625313222408, + 0.014269694685935974, + 0.02118268609046936, + 0.012727780267596245, + 0.015739791095256805, + 0.014858031645417213, + 0.010830911807715893, + 0.017306780442595482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.029819726943969727 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979554422199726 + ], + "opposite_logprob_tokens": [ + -3.5297956466674805 + ], + "logprob_reward": -0.02979554422199726, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979554422199726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08971522003412247, + 0.0493016354739666, + 0.05236111581325531, + 0.06432565301656723, + 0.10228539258241653, + 0.05121948570013046, + 0.09750135242938995, + 0.2675001323223114, + 0.10221448540687561, + 0.12357550114393234 + ], + "credit_map": [ + 0.00537933548912406, + 0.011299611069262028, + 0.011901813559234142, + 0.017231041565537453, + 0.006168851628899574, + 0.009456584230065346, + 0.013525989837944508, + 0.014751987531781197, + 0.002864360809326172, + 0.002346224617213011, + 0.007558733690530062, + 0.010235922411084175, + 0.004064595326781273, + 0.003961569629609585, + 0.009970353916287422, + 0.0082998750731349, + 0.004088213667273521, + 0.0036152710672467947, + 0.009283139370381832, + 0.009382853284478188, + 0.007099037524312735, + 0.0027831795159727335, + 0.00735120614990592, + 0.008758217096328735, + 0.007030232343822718, + 0.00316436355933547, + 0.009946865029633045, + 0.009879045188426971, + 0.011004374362528324, + 0.0035775741562247276, + 0.009520354680716991, + 0.010202839970588684, + 0.008483897894620895, + 0.012494324706494808, + 0.011632879264652729, + 0.006021575070917606, + 0.013786702416837215, + 0.02480602264404297, + 0.009920267388224602, + 0.015139730647206306, + 0.005434541963040829, + 0.004163206554949284, + 0.008561333641409874, + 0.0054153576493263245, + 0.009180111810564995, + 0.002997600706294179, + 0.007555564399808645, + 0.007911771535873413, + 0.008570469915866852, + 0.008452433161437511, + 0.010403424501419067, + 0.008207014761865139, + 0.025857936590909958, + 0.008818991482257843, + 0.012134494259953499, + 0.01505658682435751, + 0.021352635696530342, + 0.029523909091949463, + 0.021916231140494347, + 0.022430097684264183, + 0.06233939528465271, + 0.034053150564432144, + 0.026937995105981827, + 0.04894672706723213, + 0.011661752127110958, + 0.01491603534668684, + 0.01127241738140583, + 0.010224024765193462, + 0.01737275905907154, + 0.012400729581713676, + 0.010929943062365055, + 0.013436831533908844, + 0.008914527483284473, + 0.012825448997318745, + 0.013828094117343426, + 0.014235136099159718, + 0.026185857132077217, + 0.018218617886304855, + 0.013738167472183704, + 0.01562964916229248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.02979554422199726 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583592355251312 + ], + "opposite_logprob_tokens": [ + -3.8955836296081543 + ], + "logprob_reward": -0.020583592355251312, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583592355251312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1192750558257103, + 0.07660143822431564, + 0.07468552887439728, + 0.06651727110147476, + 0.047221556305885315, + 0.03212621435523033, + 0.1591821312904358, + 0.14994318783283234, + 0.22463929653167725, + 0.049808330833911896 + ], + "credit_map": [ + 0.007859227247536182, + 0.013567326590418816, + 0.013154012151062489, + 0.02427029237151146, + 0.012016274966299534, + 0.010770826600492, + 0.021838195621967316, + 0.01579890586435795, + 0.005849204491823912, + 0.004802506882697344, + 0.012234382331371307, + 0.0132552869617939, + 0.007709185592830181, + 0.005043602082878351, + 0.01242316234856844, + 0.015284106135368347, + 0.006124765146523714, + 0.0033776555210351944, + 0.010033529251813889, + 0.014487380161881447, + 0.012526090256869793, + 0.0029912730678915977, + 0.01394021138548851, + 0.011204627342522144, + 0.013992341235280037, + 0.005999183747917414, + 0.008108261972665787, + 0.006314651574939489, + 0.008227129466831684, + 0.006656209006905556, + 0.007492975797504187, + 0.009726510383188725, + 0.004479766823351383, + 0.00304823974147439, + 0.007718071807175875, + 0.0041782185435295105, + 0.0066801998764276505, + 0.008132168091833591, + 0.006168328691273928, + 0.006816566456109285, + 0.004079320468008518, + 0.0036972647067159414, + 0.003911439795047045, + 0.0038257029373198748, + 0.005334188695997, + 0.0030455265659838915, + 0.004049499519169331, + 0.004183270502835512, + 0.009706893935799599, + 0.00982188805937767, + 0.019518109038472176, + 0.016003744676709175, + 0.036888621747493744, + 0.012032750062644482, + 0.026752468198537827, + 0.028457647189497948, + 0.017197532579302788, + 0.013594625517725945, + 0.017576923593878746, + 0.01750396564602852, + 0.02572031132876873, + 0.015450367704033852, + 0.020175816491246223, + 0.022723648697137833, + 0.024273645132780075, + 0.02596244402229786, + 0.033092837780714035, + 0.01787399873137474, + 0.03833959251642227, + 0.025447478517889977, + 0.03035242296755314, + 0.029296888038516045, + 0.003976019099354744, + 0.005713839083909988, + 0.007828665897250175, + 0.006535097025334835, + 0.007608190178871155, + 0.005669825244694948, + 0.004646544344723225, + 0.007830149494111538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.020583592355251312 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up, slow-motion camera zooms in on the silhouette of a mother seated by a window. The cool moonlight filters through the old-fashioned wooden lattice, precisely outlining her slightly hunched figure, her hair tied in a bun. The entire image takes on a rich, ink-wash quality, the transitions of light and shadow flowing naturally like ink on rice paper. Clad in a dark-colored cotton shirt, she stands serenely, intently mending a slightly worn quilt on her lap. Her hands, marked by the marks of time, hold the needle steady and composed. As the camera inched ever closer, focus finally settled on the fine silver needle weaving through the thick fabric, its tip gleaming faintly in the cold moonlight. Each piercing and pulling of the thread was remarkably clear and rhythmic, like a silent ritual. A tranquil, slightly bitter atmosphere permeated the air, as if each stitch and thread wove memories of time and unspoken love.", + "seed": 14663, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_5/index_9753_seed_14663_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16352249458432197, + "vlm_logprob": -0.16352249458432197, + "vlm_yes_no_margin": 2.1, + "vlm_token_credit": -0.16352249458432197, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.32779574394226074 + ], + "opposite_logprob_tokens": [ + -1.3277957439422607 + ], + "logprob_reward": -0.32779574394226074, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32779574394226074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20687571167945862, + 0.05733012035489082, + 0.028509056195616722, + 0.1494433581829071, + 0.04655315354466438, + 0.05176227167248726, + 0.05624757334589958, + 0.05113499239087105, + 0.13082869350910187, + 0.22131508588790894 + ], + "credit_map": [ + 0.013021998107433319, + 0.020111408084630966, + 0.020024806261062622, + 0.028488287702202797, + 0.019134653732180595, + 0.04266834631562233, + 0.028991464525461197, + 0.034434761852025986, + 0.004122454673051834, + 0.008859141729772091, + 0.007151021156460047, + 0.007061721757054329, + 0.004385124891996384, + 0.00838112086057663, + 0.008755038492381573, + 0.0086144944652915, + 0.0019021566258743405, + 0.0024877723772078753, + 0.0019101796206086874, + 0.003376908367499709, + 0.0019606712739914656, + 0.006817972287535667, + 0.003442953573539853, + 0.006610443349927664, + 0.003563359845429659, + 0.0034685602877289057, + 0.00446129497140646, + 0.007030199747532606, + 0.0033846753649413586, + 0.059241630136966705, + 0.010917131789028645, + 0.05737649276852608, + 0.003984366077929735, + 0.0032798550091683865, + 0.0031977584585547447, + 0.005931174382567406, + 0.005183804780244827, + 0.008629417046904564, + 0.007638912182301283, + 0.008707866072654724, + 0.004686394240707159, + 0.0030206290539354086, + 0.004194080363959074, + 0.004340407904237509, + 0.004401100799441338, + 0.010165149345993996, + 0.010524926707148552, + 0.010429584421217442, + 0.0038641849532723427, + 0.0034125028178095818, + 0.004047432914376259, + 0.004516439978033304, + 0.007515124510973692, + 0.010934198275208473, + 0.011025220155715942, + 0.010932469740509987, + 0.003088075667619705, + 0.002761616138741374, + 0.0040542446076869965, + 0.01076145377010107, + 0.004903309978544712, + 0.008979029953479767, + 0.008608954027295113, + 0.00797831080853939, + 0.008122874423861504, + 0.007153557613492012, + 0.021490730345249176, + 0.010087309405207634, + 0.014432224445044994, + 0.023574886843562126, + 0.02364024519920349, + 0.02232685685157776, + 0.015173778869211674, + 0.0119862649589777, + 0.02238388918340206, + 0.0340251587331295, + 0.04041997343301773, + 0.03567655012011528, + 0.03606986626982689, + 0.02557961829006672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.32779574394226074 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12732937932014465 + ], + "opposite_logprob_tokens": [ + -2.1273293495178223 + ], + "logprob_reward": -0.12732937932014465, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12732937932014465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16412867605686188, + 0.10154764354228973, + 0.03516000881791115, + 0.2980530261993408, + 0.10804399847984314, + 0.05941959470510483, + 0.05078303441405296, + 0.04145846143364906, + 0.06048307567834854, + 0.08092241734266281 + ], + "credit_map": [ + 0.010589420795440674, + 0.02285991981625557, + 0.016948021948337555, + 0.023468151688575745, + 0.014745092950761318, + 0.025540627539157867, + 0.028142334893345833, + 0.02183510549366474, + 0.006559262052178383, + 0.013847221620380878, + 0.010893440805375576, + 0.013758523389697075, + 0.006920613348484039, + 0.01774762198328972, + 0.015007946640253067, + 0.016813015565276146, + 0.0017499469686299562, + 0.002582745859399438, + 0.0021057880949229, + 0.003282597055658698, + 0.0016596578061580658, + 0.009809247218072414, + 0.00437649991363287, + 0.00959352869540453, + 0.005613165441900492, + 0.006398984231054783, + 0.005826232023537159, + 0.0061595323495566845, + 0.005218165460973978, + 0.1310255378484726, + 0.009637109935283661, + 0.1281743347644806, + 0.006763729732483625, + 0.006505783647298813, + 0.008420652709901333, + 0.010189034044742584, + 0.00685835350304842, + 0.02689618431031704, + 0.01588226482272148, + 0.02652798779308796, + 0.0046863798052072525, + 0.0040838392451405525, + 0.004172143526375294, + 0.007492728065699339, + 0.004315536469221115, + 0.011380721814930439, + 0.01386220846325159, + 0.009426038712263107, + 0.004079452715814114, + 0.003241858445107937, + 0.0021549889352172613, + 0.0036061410792171955, + 0.008750268258154392, + 0.009162564761936665, + 0.01210716925561428, + 0.007680593058466911, + 0.003420325228944421, + 0.0028246541041880846, + 0.003754973877221346, + 0.008736586198210716, + 0.004157599061727524, + 0.005628661718219519, + 0.007513735443353653, + 0.0054219258017838, + 0.005423959344625473, + 0.0030486490577459335, + 0.007934700697660446, + 0.005780862644314766, + 0.0075645046308636665, + 0.011357397772371769, + 0.01141836866736412, + 0.007954628206789494, + 0.006136738229542971, + 0.006322085857391357, + 0.009776114486157894, + 0.010377843864262104, + 0.013038472272455692, + 0.012937096878886223, + 0.01256241463124752, + 0.009771657176315784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.12732937932014465 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812679275870323 + ], + "opposite_logprob_tokens": [ + -3.2881267070770264 + ], + "logprob_reward": -0.03812679275870323, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812679275870323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08352506905794144, + 0.07291003316640854, + 0.05620693042874336, + 0.1727243810892105, + 0.0670132115483284, + 0.09282735735177994, + 0.04660627990961075, + 0.0370941124856472, + 0.1535569727420807, + 0.21753565967082977 + ], + "credit_map": [ + 0.006629555486142635, + 0.01012254785746336, + 0.009088918566703796, + 0.01183802168816328, + 0.007824094966053963, + 0.013129678554832935, + 0.01260475255548954, + 0.012287499383091927, + 0.008606165647506714, + 0.010259062983095646, + 0.006148034706711769, + 0.008960051462054253, + 0.008755507878959179, + 0.009047899395227432, + 0.010350022464990616, + 0.01078328862786293, + 0.005709334276616573, + 0.0068373531103134155, + 0.008084814995527267, + 0.0052898661233484745, + 0.004522415809333324, + 0.009248326532542706, + 0.008238562382757664, + 0.008276257663965225, + 0.005783165339380503, + 0.006955661810934544, + 0.006477633956819773, + 0.008362329564988613, + 0.006536789704114199, + 0.06436637043952942, + 0.010116285644471645, + 0.06412613391876221, + 0.005929282866418362, + 0.0057287635281682014, + 0.006182513199746609, + 0.008830787613987923, + 0.006325529422610998, + 0.01220347173511982, + 0.009538067504763603, + 0.012274794280529022, + 0.009465264156460762, + 0.008338001556694508, + 0.00761151360347867, + 0.012985198758542538, + 0.008208435028791428, + 0.014668396674096584, + 0.01718558557331562, + 0.014364962466061115, + 0.005532062146812677, + 0.003477318910881877, + 0.003756341291591525, + 0.0045155189000070095, + 0.005880876909941435, + 0.007597994524985552, + 0.009534221142530441, + 0.006311948411166668, + 0.0028563577216118574, + 0.002694891532883048, + 0.002889149123802781, + 0.0065593719482421875, + 0.003681531408801675, + 0.005875536240637302, + 0.006371323484927416, + 0.0061659542843699455, + 0.009235937148332596, + 0.0077272988855838776, + 0.026992766186594963, + 0.013423582538962364, + 0.018495408818125725, + 0.027479402720928192, + 0.026754116639494896, + 0.023448459804058075, + 0.015086139552295208, + 0.012189238332211971, + 0.018502777442336082, + 0.03402001038193703, + 0.04147106036543846, + 0.035430897027254105, + 0.03539744392037392, + 0.025438103824853897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.03812679275870323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2812967002391815 + ], + "opposite_logprob_tokens": [ + -1.406296730041504 + ], + "logprob_reward": -0.2812967002391815, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2812967002391815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0604027658700943, + 0.03146091103553772, + 0.029927726835012436, + 0.17629554867744446, + 0.028868436813354492, + 0.206888809800148, + 0.02561439387500286, + 0.04003085568547249, + 0.06175624579191208, + 0.3387543559074402 + ], + "credit_map": [ + 0.004352310206741095, + 0.004477973096072674, + 0.004969698376953602, + 0.007704607676714659, + 0.0052792406640946865, + 0.01166796125471592, + 0.012683343142271042, + 0.009267626330256462, + 0.001603287528268993, + 0.0037127111572772264, + 0.0021948334760963917, + 0.005784182343631983, + 0.0018007948528975248, + 0.005734334234148264, + 0.0054428353905677795, + 0.005187931936234236, + 0.001884876168332994, + 0.0019427434308454394, + 0.0022869480308145285, + 0.0031732094939798117, + 0.001576082082465291, + 0.007177193183451891, + 0.004726141691207886, + 0.00716053182259202, + 0.002933803014457226, + 0.0029614693485200405, + 0.0018168991664424539, + 0.003117023268714547, + 0.002399103483185172, + 0.07908977568149567, + 0.006271487567573786, + 0.0777059942483902, + 0.0022399078588932753, + 0.0021111699752509594, + 0.0015775691717863083, + 0.003977428190410137, + 0.002912362804636359, + 0.005229742266237736, + 0.006319971289485693, + 0.004500283394008875, + 0.012843253090977669, + 0.011159010231494904, + 0.014383549802005291, + 0.021104734390974045, + 0.01605171151459217, + 0.04255428537726402, + 0.05909916013479233, + 0.02969309501349926, + 0.001880777534097433, + 0.0018995145801454782, + 0.001588091254234314, + 0.002709572436287999, + 0.003570026019588113, + 0.004010927863419056, + 0.0063861352391541, + 0.0035693482495844364, + 0.002931029535830021, + 0.002531968289986253, + 0.0039016755763441324, + 0.004413184244185686, + 0.004319810774177313, + 0.007898823358118534, + 0.007221118081361055, + 0.006813241634517908, + 0.005530562251806259, + 0.004068824928253889, + 0.009617703035473824, + 0.005225979257375002, + 0.008168975822627544, + 0.009811733849346638, + 0.011360686272382736, + 0.007971778512001038, + 0.02287735417485237, + 0.018508154898881912, + 0.031096452847123146, + 0.04952656105160713, + 0.06333290040493011, + 0.05562078580260277, + 0.05746468901634216, + 0.04032744839787483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 3 + ], + "selected_reward": -0.2812967002391815 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother seated by a window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark-colored cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman mending a slightly worn quilt on her lap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow-motion zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a rich ink-wash visual quality with light and shadow transitions resembling ink on rice paper?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306385666131973 + ], + "opposite_logprob_tokens": [ + -3.1680638790130615 + ], + "logprob_reward": -0.04306385666131973, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306385666131973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06346884369850159, + 0.0686921551823616, + 0.05561519414186478, + 0.2987453043460846, + 0.06329617649316788, + 0.12365608662366867, + 0.039671845734119415, + 0.05134204775094986, + 0.09755686670541763, + 0.13795539736747742 + ], + "credit_map": [ + 0.005943947471678257, + 0.00652366504073143, + 0.007474950049072504, + 0.0076631964184343815, + 0.00613979110494256, + 0.010093810968101025, + 0.01100021693855524, + 0.008629269897937775, + 0.006659078877419233, + 0.006845072377473116, + 0.0054587931372225285, + 0.008760686963796616, + 0.006377501878887415, + 0.012738393619656563, + 0.008653194643557072, + 0.013199436478316784, + 0.0037078389432281256, + 0.00528277363628149, + 0.005626472644507885, + 0.005245973821729422, + 0.0032534345518797636, + 0.013054316863417625, + 0.007014116272330284, + 0.012430270202457905, + 0.005784137640148401, + 0.006646053399890661, + 0.004458154551684856, + 0.005686887539923191, + 0.004118351265788078, + 0.13202734291553497, + 0.0091911181807518, + 0.13083326816558838, + 0.0035359461326152086, + 0.0033680074848234653, + 0.0047104861587285995, + 0.008011379279196262, + 0.002818773500621319, + 0.015530316159129143, + 0.011330575682222843, + 0.013990693725645542, + 0.01119430921971798, + 0.009239775128662586, + 0.010072893463075161, + 0.018137969076633453, + 0.012558365240693092, + 0.01909494958817959, + 0.025140663608908653, + 0.018217166885733604, + 0.0039052022621035576, + 0.0033799814991652966, + 0.002958297496661544, + 0.0034918442834168673, + 0.006081429775804281, + 0.006461357697844505, + 0.008488726802170277, + 0.004905004985630512, + 0.0051170773804187775, + 0.0032161171548068523, + 0.004970741458237171, + 0.0043957591988146305, + 0.0057215020060539246, + 0.009576761163771152, + 0.009605183266103268, + 0.008738908916711807, + 0.006196707487106323, + 0.005754312500357628, + 0.01070562656968832, + 0.007322403602302074, + 0.012028982862830162, + 0.017864707857370377, + 0.02110603079199791, + 0.016578108072280884, + 0.009546403773128986, + 0.007630605250597, + 0.012706111185252666, + 0.01977367512881756, + 0.02609534002840519, + 0.02223127707839012, + 0.023100975900888443, + 0.016871018335223198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 5 + ], + "selected_reward": -0.04306385666131973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_6/index_tensor([488, 488])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_6/index_tensor([488, 488])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9fb1858f5fe445188f58c29a82c9245b67c6c560 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_6/index_tensor([488, 488])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long shot, beginning from a slightly upward angle and slowly advancing forward, transports us into a world shrouded in a milky morning mist. A weathered arched stone bridge spans the center of the frame, its slippery slabs dotted with moss, its railings etched with the marks of age. In the center of the bridge, an elderly man, silver hair pulled back into a neat bun, dressed in a loose-fitting white silk training suit, is immersed in a Tai Chi practice. His face is gaunt, wrinkled with wisdom, his eyes slightly closed, his expression serene and focused. With his slow, fluid movements, his arms trace gentle arcs like flowing clouds and water, his weight shifting steadily and effortlessly between his legs, his wide sleeves fluttering softly. The first soft rays of morning light penetrated the mist, casting a faint golden halo on his outline and silver hair, forming a subtle contrast with the cool, gray-green background tones around him. The whole picture was filled with an otherworldly sense of tranquility and concentration, as if time had frozen at this moment.", + "seed": 12200, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_6/index_488_seed_12200_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.015673020202666522, + "vlm_logprob": -0.015673020202666522, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.015673020202666522, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04143963381648064 + ], + "opposite_logprob_tokens": [ + -3.7914395332336426 + ], + "logprob_reward": -0.04143963381648064, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.04143963381648064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13920606672763824, + 0.05238981544971466, + 0.040648408234119415, + 0.029070794582366943, + 0.07498937100172043, + 0.057738203555345535, + 0.026631560176610947, + 0.018439538776874542, + 0.08716839551925659, + 0.4737177789211273 + ], + "credit_map": [ + 0.01505199633538723, + 0.02212570421397686, + 0.016833452507853508, + 0.01731867529451847, + 0.028803158551454544, + 0.011797389015555382, + 0.018053093925118446, + 0.00922261830419302, + 0.0047164480201900005, + 0.010122918523848057, + 0.005629569757729769, + 0.006043008994311094, + 0.009141691960394382, + 0.007971676997840405, + 0.00530671002343297, + 0.0034577990882098675, + 0.0037973688449710608, + 0.010723297484219074, + 0.0032473031897097826, + 0.007491034455597401, + 0.006057309452444315, + 0.003359819995239377, + 0.0036778785288333893, + 0.0022944023367017508, + 0.0049720569513738155, + 0.004237870220094919, + 0.0033737539779394865, + 0.002757325768470764, + 0.003892053384333849, + 0.004173737484961748, + 0.0024527781642973423, + 0.0032112235203385353, + 0.007544749416410923, + 0.009485839866101742, + 0.013780400156974792, + 0.011923862621188164, + 0.008577036671340466, + 0.008151770569384098, + 0.008567577227950096, + 0.006958143785595894, + 0.004839407280087471, + 0.0063781095668673515, + 0.004642283543944359, + 0.009019175544381142, + 0.00913532916456461, + 0.00950045045465231, + 0.009083434008061886, + 0.005140019115060568, + 0.002724750665947795, + 0.003027748316526413, + 0.0032734340056777, + 0.0046652574092149734, + 0.00374682922847569, + 0.003677109954878688, + 0.0027509902138262987, + 0.002765444340184331, + 0.002417060313746333, + 0.0026509275194257498, + 0.0021365568973124027, + 0.0030539212748408318, + 0.0021288800053298473, + 0.001995290396735072, + 0.002437470480799675, + 0.0016194344498217106, + 0.011073715053498745, + 0.009141619317233562, + 0.007149084005504847, + 0.011097877286374569, + 0.015133754350244999, + 0.00924466922879219, + 0.016711650416254997, + 0.007616035640239716, + 0.17282268404960632, + 0.20153221487998962, + 0.006350876297801733, + 0.00942941289395094, + 0.016471147537231445, + 0.03724270313978195, + 0.022891348227858543, + 0.006977436598390341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.04143963381648064 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009798052720725536 + ], + "opposite_logprob_tokens": [ + -4.634798049926758 + ], + "logprob_reward": -0.009798052720725536, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009798052720725536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05860096216201782, + 0.022502847015857697, + 0.021807879209518433, + 0.03439755365252495, + 0.028164608404040337, + 0.030299408361315727, + 0.013128198683261871, + 0.0187517162412405, + 0.01618538610637188, + 0.7561615109443665 + ], + "credit_map": [ + 0.005439344793558121, + 0.0072537255473434925, + 0.006850420963019133, + 0.006675083655864, + 0.014914950355887413, + 0.005719705019146204, + 0.007523889187723398, + 0.0042238435707986355, + 0.0021719238720834255, + 0.003335915971547365, + 0.0030824225395917892, + 0.005850551649928093, + 0.0027053100056946278, + 0.0018500377191230655, + 0.0019069276750087738, + 0.0015997581649571657, + 0.002502845600247383, + 0.0023133549839258194, + 0.0026920821983367205, + 0.004445548169314861, + 0.0036967936903238297, + 0.0018169002141803503, + 0.0027554547414183617, + 0.0015848985640332103, + 0.0034720716066658497, + 0.005104715935885906, + 0.0032208897173404694, + 0.008140597492456436, + 0.004298713989555836, + 0.0027036441024392843, + 0.004338799510151148, + 0.0031181222293525934, + 0.0022015899885445833, + 0.003675698535516858, + 0.004135047551244497, + 0.0037419397849589586, + 0.00464441254734993, + 0.0030037332326173782, + 0.003378983587026596, + 0.0033832036424428225, + 0.0022826488129794598, + 0.004226310644298792, + 0.001646010554395616, + 0.002511376515030861, + 0.008288335986435413, + 0.0032017016783356667, + 0.005976156797260046, + 0.0021668653935194016, + 0.0013169161975383759, + 0.0017767867539077997, + 0.0011930663604289293, + 0.0019306031754240394, + 0.002585430396720767, + 0.0015863763401284814, + 0.0018279536161571741, + 0.0009110659011639655, + 0.0024017200339585543, + 0.002401506993919611, + 0.001480430131778121, + 0.002111717825755477, + 0.0037322507705539465, + 0.0021929594222456217, + 0.0029146787710487843, + 0.0015164513606578112, + 0.0015274580800905824, + 0.0020272897090762854, + 0.0014060249086469412, + 0.0018920187139883637, + 0.003542103338986635, + 0.0017877215286716819, + 0.0027037118561565876, + 0.0012990572722628713, + 0.2771501839160919, + 0.31503576040267944, + 0.012264643795788288, + 0.014115195721387863, + 0.028149466961622238, + 0.05571182072162628, + 0.03870474919676781, + 0.01502963900566101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.009798052720725536 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643477223813534 + ], + "opposite_logprob_tokens": [ + -4.758643627166748 + ], + "logprob_reward": -0.008643477223813534, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643477223813534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11524083465337753, + 0.05280977115035057, + 0.04816466197371483, + 0.05153852328658104, + 0.04589018598198891, + 0.04848325625061989, + 0.022121960297226906, + 0.046215225011110306, + 0.014445800334215164, + 0.5550897717475891 + ], + "credit_map": [ + 0.015381169505417347, + 0.01831856369972229, + 0.012568006291985512, + 0.013188245706260204, + 0.01797979325056076, + 0.01576431281864643, + 0.01091725379228592, + 0.011123490519821644, + 0.004436167888343334, + 0.006724285427480936, + 0.007284205872565508, + 0.008906707167625427, + 0.010497272945940495, + 0.0066246031783521175, + 0.004775179084390402, + 0.0035613521467894316, + 0.003901757299900055, + 0.009698042646050453, + 0.0045366971753537655, + 0.006443770602345467, + 0.007598872296512127, + 0.004438008647412062, + 0.007944012992084026, + 0.003603496588766575, + 0.004618421196937561, + 0.0077325101010501385, + 0.003886864520609379, + 0.00713609391823411, + 0.014219184406101704, + 0.005423147697001696, + 0.005491677206009626, + 0.003030624007806182, + 0.002944855485111475, + 0.006655644625425339, + 0.0038142804987728596, + 0.004400214180350304, + 0.01144917868077755, + 0.003703343216329813, + 0.007275068666785955, + 0.005647606216371059, + 0.003543947823345661, + 0.006762519013136625, + 0.004669524729251862, + 0.006552937440574169, + 0.011229956522583961, + 0.006527923513203859, + 0.005142980720847845, + 0.0040534608997404575, + 0.002026786096394062, + 0.0030199033208191395, + 0.002119189128279686, + 0.0028697792440652847, + 0.003508594585582614, + 0.0035788530949503183, + 0.0032675873953849077, + 0.0017312669660896063, + 0.006144975312054157, + 0.005677333567291498, + 0.003418081207200885, + 0.005374296568334103, + 0.005824394524097443, + 0.005585792940109968, + 0.009937187656760216, + 0.004253167659044266, + 0.0017921264516189694, + 0.002190882107242942, + 0.0014568041078746319, + 0.00205779867246747, + 0.0022254555951803923, + 0.0019464213401079178, + 0.001861155848018825, + 0.0009151560952886939, + 0.20186063647270203, + 0.22922712564468384, + 0.008817602880299091, + 0.01081901416182518, + 0.021000301465392113, + 0.040330708026885986, + 0.03148866072297096, + 0.011545713059604168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.008643477223813534 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949525628238916 + ], + "opposite_logprob_tokens": [ + -5.1309494972229 + ], + "logprob_reward": -0.005949525628238916, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949525628238916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10844459384679794, + 0.06171433627605438, + 0.07137902826070786, + 0.09539239853620529, + 0.05892086774110794, + 0.04811680689454079, + 0.057404566556215286, + 0.05510660633444786, + 0.074885293841362, + 0.3686354458332062 + ], + "credit_map": [ + 0.010866742581129074, + 0.015924787148833275, + 0.011848239228129387, + 0.015776317566633224, + 0.017888247966766357, + 0.009437942877411842, + 0.016205057501792908, + 0.010497266426682472, + 0.004587842617183924, + 0.010692938230931759, + 0.00812843069434166, + 0.013576866127550602, + 0.010388068854808807, + 0.004213958978652954, + 0.005598064046353102, + 0.0045281704515218735, + 0.006341299507766962, + 0.009353594854474068, + 0.007508974988013506, + 0.010109025053679943, + 0.01438057329505682, + 0.007242020219564438, + 0.011394607834517956, + 0.00504893297329545, + 0.007550223730504513, + 0.012170053087174892, + 0.014690585434436798, + 0.015509774908423424, + 0.016220049932599068, + 0.009966181591153145, + 0.010891883634030819, + 0.00839365366846323, + 0.0044203754514455795, + 0.00941548403352499, + 0.004459307063370943, + 0.006001971662044525, + 0.01362635474652052, + 0.004571408033370972, + 0.009965082630515099, + 0.006460888776928186, + 0.004293834790587425, + 0.007328083273023367, + 0.004128748085349798, + 0.005574342329055071, + 0.007235358469188213, + 0.007545880973339081, + 0.007327876053750515, + 0.00468268571421504, + 0.006656449753791094, + 0.0061202035285532475, + 0.004000839777290821, + 0.00404495932161808, + 0.010766362771391869, + 0.009658471681177616, + 0.011392829939723015, + 0.004764455836266279, + 0.006011981051415205, + 0.005827167071402073, + 0.004839291330426931, + 0.009561062790453434, + 0.009674112312495708, + 0.0084524629637599, + 0.006490844301879406, + 0.004249688237905502, + 0.009149051271378994, + 0.011320429854094982, + 0.006686102598905563, + 0.009962865151464939, + 0.010758884251117706, + 0.009064760990440845, + 0.013862838968634605, + 0.004080364480614662, + 0.12902706861495972, + 0.14780844748020172, + 0.006944154854863882, + 0.01009331550449133, + 0.016348421573638916, + 0.0267082117497921, + 0.019678952172398567, + 0.012026886455714703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.005949525628238916 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534411624073982 + ], + "opposite_logprob_tokens": [ + -4.3875346183776855 + ], + "logprob_reward": -0.012534411624073982, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534411624073982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19500641524791718, + 0.04736890643835068, + 0.06333906948566437, + 0.10814008116722107, + 0.08528371155261993, + 0.03148104250431061, + 0.020570214837789536, + 0.05281703546643257, + 0.024676380679011345, + 0.3713170886039734 + ], + "credit_map": [ + 0.026202581822872162, + 0.026232847943902016, + 0.019481051713228226, + 0.02727343700826168, + 0.040051113814115524, + 0.018169961869716644, + 0.02535761334002018, + 0.01223780494183302, + 0.004794389009475708, + 0.006836877670139074, + 0.005108766257762909, + 0.012119559571146965, + 0.006467794068157673, + 0.003935311455279589, + 0.004719204735010862, + 0.003387002507224679, + 0.009098613634705544, + 0.008544890210032463, + 0.006619745399802923, + 0.008320407010614872, + 0.008963363245129585, + 0.007257931400090456, + 0.009530586190521717, + 0.005003530997782946, + 0.013197182677686214, + 0.013530214317142963, + 0.01685073971748352, + 0.012554597109556198, + 0.01299669686704874, + 0.013222034089267254, + 0.010531781241297722, + 0.015256837010383606, + 0.008304707705974579, + 0.010652466677129269, + 0.015741879120469093, + 0.012101094238460064, + 0.010800361633300781, + 0.009185467846691608, + 0.009240476414561272, + 0.009257257916033268, + 0.0034702748525887728, + 0.003497394500300288, + 0.002843675669282675, + 0.004238871391862631, + 0.0067674871534109116, + 0.003325141966342926, + 0.005168739706277847, + 0.002169456798583269, + 0.0023321788758039474, + 0.0026486944407224655, + 0.0018831283086910844, + 0.0025501262862235308, + 0.0031349665950983763, + 0.002844986505806446, + 0.003459138097241521, + 0.0017169957282021642, + 0.007534690201282501, + 0.006789566017687321, + 0.010818217881023884, + 0.009045467711985111, + 0.003780683036893606, + 0.0057834056206047535, + 0.003683973802253604, + 0.005381035152822733, + 0.003298540133982897, + 0.0035378667525947094, + 0.003615105990320444, + 0.0031733703799545765, + 0.003326213452965021, + 0.0034436716232448816, + 0.002393618691712618, + 0.00188799190800637, + 0.13654732704162598, + 0.1510939598083496, + 0.00780125055462122, + 0.009642519056797028, + 0.01478751190006733, + 0.02901208959519863, + 0.017711719498038292, + 0.004720731172710657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.012534411624073982 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, beginning from a slightly upward angle and slowly advancing forward, transports us into a world shrouded in a milky morning mist. A weathered arched stone bridge spans the center of the frame, its slippery slabs dotted with moss, its railings etched with the marks of age. In the center of the bridge, an elderly man, silver hair pulled back into a neat bun, dressed in a loose-fitting white silk training suit, is immersed in a Tai Chi practice. His face is gaunt, wrinkled with wisdom, his eyes slightly closed, his expression serene and focused. With his slow, fluid movements, his arms trace gentle arcs like flowing clouds and water, his weight shifting steadily and effortlessly between his legs, his wide sleeves fluttering softly. The first soft rays of morning light penetrated the mist, casting a faint golden halo on his outline and silver hair, forming a subtle contrast with the cool, gray-green background tones around him. The whole picture was filled with an otherworldly sense of tranquility and concentration, as if time had frozen at this moment.", + "seed": 12201, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_6/index_488_seed_12201_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02929575266316533, + "vlm_logprob": -0.02929575266316533, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.02929575266316533, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11797703802585602 + ], + "opposite_logprob_tokens": [ + -2.3679771423339844 + ], + "logprob_reward": -0.11797703802585602, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11797703802585602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28049442172050476, + 0.05559202656149864, + 0.09337770193815231, + 0.10649366676807404, + 0.049882423132658005, + 0.10609103739261627, + 0.08844044059515, + 0.044879838824272156, + 0.05593348294496536, + 0.11881493031978607 + ], + "credit_map": [ + 0.04110105708241463, + 0.028875455260276794, + 0.050440844148397446, + 0.027236929163336754, + 0.02071181684732437, + 0.042676880955696106, + 0.030578024685382843, + 0.03887341171503067, + 0.006928440183401108, + 0.00676974980160594, + 0.007885332219302654, + 0.006274686194956303, + 0.010261625051498413, + 0.007283850573003292, + 0.0048906877636909485, + 0.0052976543083786964, + 0.010271770879626274, + 0.02685638889670372, + 0.012180203571915627, + 0.008169395849108696, + 0.005389449652284384, + 0.014047971926629543, + 0.0059639643877744675, + 0.010498550720512867, + 0.015754535794258118, + 0.009410034865140915, + 0.010403322987258434, + 0.00959016289561987, + 0.02195855602622032, + 0.014293069019913673, + 0.017116965726017952, + 0.007967019453644753, + 0.006537188775837421, + 0.006316813640296459, + 0.004853447433561087, + 0.004203945863991976, + 0.007699580397456884, + 0.008761457167565823, + 0.005151866003870964, + 0.0063581219874322414, + 0.013099024072289467, + 0.008793042972683907, + 0.00878442358225584, + 0.03456259146332741, + 0.007350921630859375, + 0.010148142464458942, + 0.014198284596204758, + 0.00915460754185915, + 0.014673516154289246, + 0.009130815044045448, + 0.00965285673737526, + 0.008318685926496983, + 0.007767345756292343, + 0.010144069790840149, + 0.01114309299737215, + 0.017610058188438416, + 0.005103345960378647, + 0.00660724239423871, + 0.0053480868227779865, + 0.004865058697760105, + 0.00512060709297657, + 0.007145806215703487, + 0.005745781119912863, + 0.004943912383168936, + 0.007283905055373907, + 0.010977011173963547, + 0.003917966969311237, + 0.005778878927230835, + 0.007696318440139294, + 0.008332544937729836, + 0.006423634476959705, + 0.005523226223886013, + 0.03145788609981537, + 0.007244383916258812, + 0.006459243595600128, + 0.007409542333334684, + 0.010084724985063076, + 0.0175479743629694, + 0.02939736470580101, + 0.009213821031153202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.11797703802585602 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002833995968103409 + ], + "opposite_logprob_tokens": [ + -5.877833843231201 + ], + "logprob_reward": -0.002833995968103409, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002833995968103409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04693244397640228, + 0.09893681108951569, + 0.062396060675382614, + 0.019556747749447823, + 0.03673169016838074, + 0.43112847208976746, + 0.029368065297603607, + 0.036895256489515305, + 0.04270344600081444, + 0.19535093009471893 + ], + "credit_map": [ + 0.00489816116169095, + 0.006246800534427166, + 0.0050659882836043835, + 0.005814622156322002, + 0.007303988561034203, + 0.006369558162987232, + 0.006357951555401087, + 0.004875378683209419, + 0.012549305334687233, + 0.010937019251286983, + 0.010069498792290688, + 0.008834353648126125, + 0.026226188987493515, + 0.012028748169541359, + 0.009119951166212559, + 0.00917175691574812, + 0.009892135858535767, + 0.022975793108344078, + 0.0038780425675213337, + 0.00461545679718256, + 0.005992259364575148, + 0.008351104333996773, + 0.0025581999216228724, + 0.004133080132305622, + 0.0026207594200968742, + 0.002079085214063525, + 0.0017133854562416673, + 0.0015911831287667155, + 0.005281830206513405, + 0.0022885724902153015, + 0.002323469379916787, + 0.0016584655968472362, + 0.004749559797346592, + 0.003630767809227109, + 0.002262560185045004, + 0.003608803963288665, + 0.009116199798882008, + 0.003626122372224927, + 0.0051385643891990185, + 0.004599117208272219, + 0.03531210869550705, + 0.020749112591147423, + 0.01828494295477867, + 0.15098491311073303, + 0.04115894064307213, + 0.03332046419382095, + 0.09987162053585052, + 0.03144639730453491, + 0.005187139380723238, + 0.00438826996833086, + 0.0029982931446284056, + 0.002692247275263071, + 0.003919126931577921, + 0.0035794361028820276, + 0.0019220829708501697, + 0.0046814740635454655, + 0.008046582341194153, + 0.008817044086754322, + 0.0026207873597741127, + 0.002371137961745262, + 0.003398710861802101, + 0.004022578708827496, + 0.0023619940038770437, + 0.005256426054984331, + 0.008183971978724003, + 0.007264078129082918, + 0.0032772596459835768, + 0.007013258058577776, + 0.005201054736971855, + 0.0038772146217525005, + 0.0039111836813390255, + 0.003975432366132736, + 0.058242328464984894, + 0.010522802360355854, + 0.007008437067270279, + 0.016200408339500427, + 0.01638481579720974, + 0.015041339211165905, + 0.05983436107635498, + 0.012116474099457264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.002833995968103409 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011082522571086884 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.011082522571086884, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082522571086884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20928052067756653, + 0.11659316718578339, + 0.06217328831553459, + 0.10241992026567459, + 0.06463422626256943, + 0.18424616754055023, + 0.060174982994794846, + 0.06516042351722717, + 0.0450800284743309, + 0.09023728966712952 + ], + "credit_map": [ + 0.02327018417418003, + 0.02023155614733696, + 0.02140877954661846, + 0.03176464885473251, + 0.025020474568009377, + 0.034856587648391724, + 0.024183500558137894, + 0.02854478918015957, + 0.013836496509611607, + 0.01167637575417757, + 0.012858334928750992, + 0.012769270688295364, + 0.025001324713230133, + 0.016348116099834442, + 0.012610778212547302, + 0.011492469348013401, + 0.009225099347531796, + 0.022090217098593712, + 0.0025675883516669273, + 0.0045318603515625, + 0.007385636679828167, + 0.007216067053377628, + 0.0037117854226380587, + 0.005445033311843872, + 0.016240548342466354, + 0.011095481924712658, + 0.00765871349722147, + 0.006373709533363581, + 0.027987612411379814, + 0.01477561704814434, + 0.010988913476467133, + 0.0072993263602256775, + 0.007751955185085535, + 0.006381526123732328, + 0.005210367031395435, + 0.008616341277956963, + 0.011123613454401493, + 0.0064342571422457695, + 0.009104865603148937, + 0.01001130323857069, + 0.026997782289981842, + 0.011129442602396011, + 0.00932011753320694, + 0.07193531841039658, + 0.009939795359969139, + 0.013691750355064869, + 0.02143567055463791, + 0.019796276465058327, + 0.010611971840262413, + 0.010064708068966866, + 0.0073762149550020695, + 0.0053094662725925446, + 0.007620117627084255, + 0.007637659087777138, + 0.003599885618314147, + 0.007954959757626057, + 0.011702983640134335, + 0.012887255288660526, + 0.006551067810505629, + 0.0053994543850421906, + 0.0058045536279678345, + 0.007603996433317661, + 0.005035328678786755, + 0.010175785981118679, + 0.006139263045042753, + 0.008166728541254997, + 0.003321624593809247, + 0.005074750632047653, + 0.0072994958609342575, + 0.005295082926750183, + 0.0048846472054719925, + 0.004898433107882738, + 0.015507224947214127, + 0.00972388219088316, + 0.0074879019521176815, + 0.0083932401612401, + 0.012165323831140995, + 0.011101527139544487, + 0.01613112911581993, + 0.009727058932185173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.011082522571086884 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950829479843378 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950829479843378, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950829479843378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09455486387014389, + 0.07115508615970612, + 0.06028778478503227, + 0.17085422575473785, + 0.05087512731552124, + 0.20027092099189758, + 0.13478870689868927, + 0.05448341369628906, + 0.06727717816829681, + 0.09545265138149261 + ], + "credit_map": [ + 0.010546060279011726, + 0.009861168451607227, + 0.0074635944329202175, + 0.01016747485846281, + 0.014079215005040169, + 0.016954772174358368, + 0.012244300916790962, + 0.013238280080258846, + 0.010841070674359798, + 0.0040842583402991295, + 0.01735565811395645, + 0.006504377815872431, + 0.011848925612866879, + 0.008099525235593319, + 0.005348716862499714, + 0.007072557229548693, + 0.00939768087118864, + 0.019444787874817848, + 0.005848485045135021, + 0.005614870693534613, + 0.0036071259528398514, + 0.008633513003587723, + 0.004479562398046255, + 0.0032617568504065275, + 0.01926441676914692, + 0.021465793251991272, + 0.016269972547888756, + 0.014164775609970093, + 0.04705876484513283, + 0.017155155539512634, + 0.021193401888012886, + 0.01428194623440504, + 0.007539838552474976, + 0.0054155816324055195, + 0.004399138502776623, + 0.005610981490463018, + 0.0068479860201478004, + 0.006101933773607016, + 0.006083655171096325, + 0.008876010775566101, + 0.02532033994793892, + 0.013048469088971615, + 0.009524990804493427, + 0.07663384079933167, + 0.012928517535328865, + 0.015953008085489273, + 0.028651658445596695, + 0.01821010746061802, + 0.0211234949529171, + 0.03063010983169079, + 0.01515133772045374, + 0.016007110476493835, + 0.017021194100379944, + 0.014002452604472637, + 0.00989879947155714, + 0.010954204946756363, + 0.007331199012696743, + 0.008555673994123936, + 0.012556292116641998, + 0.00611032173037529, + 0.004390215966850519, + 0.005099515896290541, + 0.005018618889153004, + 0.005421576090157032, + 0.016409797593951225, + 0.008190602995455265, + 0.003486855421215296, + 0.016690630465745926, + 0.0056903911754488945, + 0.006957006175071001, + 0.005458956118673086, + 0.00439293822273612, + 0.013022075407207012, + 0.011040867306292057, + 0.007376275956630707, + 0.010335690341889858, + 0.01651710271835327, + 0.012739467434585094, + 0.014255648478865623, + 0.010165522806346416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.005950829479843378 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634377270936966 + ], + "opposite_logprob_tokens": [ + -4.758634567260742 + ], + "logprob_reward": -0.008634377270936966, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634377270936966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23373548686504364, + 0.16840684413909912, + 0.1551835685968399, + 0.05839615315198898, + 0.03630351647734642, + 0.12628763914108276, + 0.07990463078022003, + 0.07408186793327332, + 0.027720022946596146, + 0.039980217814445496 + ], + "credit_map": [ + 0.021950744092464447, + 0.026106063276529312, + 0.030748849734663963, + 0.031076207756996155, + 0.018221842125058174, + 0.04300586134195328, + 0.028642579913139343, + 0.03398334980010986, + 0.027427902445197105, + 0.010146575048565865, + 0.05218608304858208, + 0.016594072803854942, + 0.00780186802148819, + 0.02427176758646965, + 0.009110978804528713, + 0.020867574959993362, + 0.020414205268025398, + 0.054259736090898514, + 0.011414364911615849, + 0.01657089963555336, + 0.009352640248835087, + 0.01816309429705143, + 0.008394363336265087, + 0.0166142750531435, + 0.00950416550040245, + 0.006982902996242046, + 0.006513781379908323, + 0.005046900361776352, + 0.00927676074206829, + 0.011315946467220783, + 0.005524265579879284, + 0.004231433849781752, + 0.003703953232616186, + 0.003827661508694291, + 0.0037492874544113874, + 0.006851860787719488, + 0.005437315441668034, + 0.004073585383594036, + 0.0031751426868140697, + 0.005484711844474077, + 0.011223429813981056, + 0.006622954271733761, + 0.010159237310290337, + 0.03337832912802696, + 0.010344274342060089, + 0.013162952847778797, + 0.025203153491020203, + 0.01619330793619156, + 0.01679043658077717, + 0.006093979813158512, + 0.006206455174833536, + 0.006014649290591478, + 0.006812226492911577, + 0.015388148836791515, + 0.004840198438614607, + 0.017758535221219063, + 0.019949940964579582, + 0.008002222515642643, + 0.004586593247950077, + 0.003658743342384696, + 0.004375819582492113, + 0.010873492807149887, + 0.004241912625730038, + 0.018393145874142647, + 0.00635036313906312, + 0.002073288196697831, + 0.0016792803071439266, + 0.006428701803088188, + 0.002115911804139614, + 0.004256493877619505, + 0.0024001221172511578, + 0.0024158628657460213, + 0.006913186516612768, + 0.003922407981008291, + 0.0028773739468306303, + 0.0039278315380215645, + 0.005582764744758606, + 0.005715512670576572, + 0.006962575484067202, + 0.004078563302755356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.008634377270936966 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, beginning from a slightly upward angle and slowly advancing forward, transports us into a world shrouded in a milky morning mist. A weathered arched stone bridge spans the center of the frame, its slippery slabs dotted with moss, its railings etched with the marks of age. In the center of the bridge, an elderly man, silver hair pulled back into a neat bun, dressed in a loose-fitting white silk training suit, is immersed in a Tai Chi practice. His face is gaunt, wrinkled with wisdom, his eyes slightly closed, his expression serene and focused. With his slow, fluid movements, his arms trace gentle arcs like flowing clouds and water, his weight shifting steadily and effortlessly between his legs, his wide sleeves fluttering softly. The first soft rays of morning light penetrated the mist, casting a faint golden halo on his outline and silver hair, forming a subtle contrast with the cool, gray-green background tones around him. The whole picture was filled with an otherworldly sense of tranquility and concentration, as if time had frozen at this moment.", + "seed": 12202, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_6/index_488_seed_12202_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03394501470029354, + "vlm_logprob": -0.03394501470029354, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.03394501470029354, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1258944272994995 + ], + "opposite_logprob_tokens": [ + -2.375894546508789 + ], + "logprob_reward": -0.1258944272994995, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1258944272994995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22433142364025116, + 0.06412249058485031, + 0.1050625592470169, + 0.07344594597816467, + 0.04785685986280441, + 0.060645028948783875, + 0.192948117852211, + 0.08499909937381744, + 0.09459559619426727, + 0.05199287086725235 + ], + "credit_map": [ + 0.02657156251370907, + 0.04086583852767944, + 0.02499687485396862, + 0.048585474491119385, + 0.025232551619410515, + 0.02830188162624836, + 0.01799154281616211, + 0.011785697191953659, + 0.010865403339266777, + 0.010370076633989811, + 0.004388085100799799, + 0.004425594117492437, + 0.011549352668225765, + 0.009748884476721287, + 0.00792074203491211, + 0.004854346625506878, + 0.01586553081870079, + 0.022185999900102615, + 0.008120029233396053, + 0.007169507909566164, + 0.019728442654013634, + 0.016632486134767532, + 0.008949212729930878, + 0.006411352660506964, + 0.00931879598647356, + 0.01392761617898941, + 0.007714768871665001, + 0.0073952036909759045, + 0.0051319957710802555, + 0.013766216114163399, + 0.0051994104869663715, + 0.010991940274834633, + 0.007502197753638029, + 0.008039241656661034, + 0.00458961958065629, + 0.005538007244467735, + 0.0062920148484408855, + 0.005398477893322706, + 0.00743166683241725, + 0.0030656331218779087, + 0.007404728326946497, + 0.006932068150490522, + 0.011011069640517235, + 0.0066118063405156136, + 0.004320940934121609, + 0.008455502800643444, + 0.006599892862141132, + 0.009309021756052971, + 0.029006795957684517, + 0.03276795893907547, + 0.030622590333223343, + 0.02357044443488121, + 0.02091989479959011, + 0.01929430104792118, + 0.0196352768689394, + 0.017130855470895767, + 0.01049316581338644, + 0.021068433299660683, + 0.01164688728749752, + 0.007597421295940876, + 0.007811700459569693, + 0.012674887664616108, + 0.008212082087993622, + 0.005494526121765375, + 0.011239532381296158, + 0.00884302519261837, + 0.012369601987302303, + 0.014423949643969536, + 0.01220726314932108, + 0.007640409283339977, + 0.02048739790916443, + 0.0073844147846102715, + 0.0070028407499194145, + 0.0048874663189053535, + 0.0046964120119810104, + 0.004451286047697067, + 0.013414197601377964, + 0.0055304174311459064, + 0.006975407712161541, + 0.005034842994064093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.1258944272994995 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005267906002700329 + ], + "opposite_logprob_tokens": [ + -5.255268096923828 + ], + "logprob_reward": -0.005267906002700329, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005267906002700329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09405602514743805, + 0.13140401244163513, + 0.04574648290872574, + 0.062102142721414566, + 0.022736717015504837, + 0.034514494240283966, + 0.1479150801897049, + 0.06293393671512604, + 0.07635828107595444, + 0.32223284244537354 + ], + "credit_map": [ + 0.013932235538959503, + 0.015870947390794754, + 0.007434861734509468, + 0.013682455755770206, + 0.01325117889791727, + 0.012127828784286976, + 0.01217608992010355, + 0.005580425262451172, + 0.022487083449959755, + 0.022174816578626633, + 0.008262058719992638, + 0.008065925911068916, + 0.03179553151130676, + 0.016096701845526695, + 0.016695359721779823, + 0.005826525390148163, + 0.006537756882607937, + 0.010552293621003628, + 0.004941219929605722, + 0.0046278806403279305, + 0.005864547099918127, + 0.005769653245806694, + 0.003719709115102887, + 0.003733423538506031, + 0.008647711016237736, + 0.01273256167769432, + 0.006427802611142397, + 0.005793506745249033, + 0.004527380224317312, + 0.011625469662249088, + 0.0045347921550273895, + 0.007812917232513428, + 0.00336074479855597, + 0.0035439569037407637, + 0.002691644476726651, + 0.002273299265652895, + 0.0032399189658463, + 0.0024620583280920982, + 0.0035435024183243513, + 0.0016215941868722439, + 0.005111773032695055, + 0.005339345429092646, + 0.005075753666460514, + 0.002756331814453006, + 0.002798873232677579, + 0.005982994567602873, + 0.0031619591172784567, + 0.00428746035322547, + 0.024946890771389008, + 0.01792350970208645, + 0.023673275485634804, + 0.016740571707487106, + 0.01711405999958515, + 0.016578972339630127, + 0.01625392958521843, + 0.014683874323964119, + 0.007645988371223211, + 0.010008175857365131, + 0.0042526815086603165, + 0.005361352115869522, + 0.009670121595263481, + 0.00965888611972332, + 0.009558812715113163, + 0.006777919828891754, + 0.007858333177864552, + 0.008138014934957027, + 0.013573170639574528, + 0.012458177283406258, + 0.007882749661803246, + 0.0071668787859380245, + 0.01260882057249546, + 0.0066721318289637566, + 0.0349532850086689, + 0.04312777519226074, + 0.027216032147407532, + 0.04503965377807617, + 0.06749550998210907, + 0.03237074986100197, + 0.054302021861076355, + 0.017727823927998543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.005267906002700329 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954266060143709 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.005954266060143709, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954266060143709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1561255156993866, + 0.14770576357841492, + 0.08411960303783417, + 0.10985426604747772, + 0.10364089906215668, + 0.04537351056933403, + 0.03014366328716278, + 0.07718372344970703, + 0.14577314257621765, + 0.10007994621992111 + ], + "credit_map": [ + 0.02824866585433483, + 0.02228318527340889, + 0.011923268437385559, + 0.0175701305270195, + 0.018679527565836906, + 0.019769078120589256, + 0.027398426085710526, + 0.010253236629068851, + 0.028819680213928223, + 0.021301422268152237, + 0.01046011596918106, + 0.010165663436055183, + 0.035966191440820694, + 0.01753229834139347, + 0.01462294440716505, + 0.00883745402097702, + 0.011762376874685287, + 0.01854228600859642, + 0.009376544505357742, + 0.008176014758646488, + 0.008481031283736229, + 0.011543766595423222, + 0.008265555836260319, + 0.00797202531248331, + 0.017233923077583313, + 0.022328654304146767, + 0.012714370153844357, + 0.008665703237056732, + 0.0075207361951470375, + 0.020525755360722542, + 0.007783182431012392, + 0.013081934303045273, + 0.014413433149456978, + 0.016420111060142517, + 0.013836167752742767, + 0.008695977739989758, + 0.010045735165476799, + 0.01636822149157524, + 0.011619847267866135, + 0.012241406366229057, + 0.006030390504747629, + 0.00634061498567462, + 0.006803053896874189, + 0.004492172505706549, + 0.004345178138464689, + 0.006724745966494083, + 0.004777503199875355, + 0.005859849508851767, + 0.0046068052761256695, + 0.006084760185331106, + 0.0029858178459107876, + 0.0025790182407945395, + 0.002858464838936925, + 0.005194393917918205, + 0.002909769769757986, + 0.002924633212387562, + 0.009650415740907192, + 0.009324637241661549, + 0.013179702684283257, + 0.009080278687179089, + 0.009434527717530727, + 0.009013151749968529, + 0.013149864040315151, + 0.0043511479161679745, + 0.016026530414819717, + 0.014662506990134716, + 0.026502154767513275, + 0.02753937430679798, + 0.013006315566599369, + 0.013187442906200886, + 0.021063853055238724, + 0.013784985058009624, + 0.011721590533852577, + 0.013659476302564144, + 0.011042927391827106, + 0.011648212559521198, + 0.01443119440227747, + 0.012607630342245102, + 0.015656784176826477, + 0.009312130510807037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.005954266060143709 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028206822462379932 + ], + "opposite_logprob_tokens": [ + -5.8778204917907715 + ], + "logprob_reward": -0.0028206822462379932, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028206822462379932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1138516440987587, + 0.08394943177700043, + 0.058825183659791946, + 0.05333225429058075, + 0.036440659314394, + 0.041548311710357666, + 0.14268381893634796, + 0.18326497077941895, + 0.19491684436798096, + 0.09118698537349701 + ], + "credit_map": [ + 0.014791862107813358, + 0.017560278996825218, + 0.014232750050723553, + 0.024676615372300148, + 0.010112089104950428, + 0.013896001502871513, + 0.011779691092669964, + 0.006802348420023918, + 0.012334422208368778, + 0.017966577783226967, + 0.006624181754887104, + 0.006917368154972792, + 0.013989144004881382, + 0.01446984987705946, + 0.006220605224370956, + 0.005427277646958828, + 0.008683663792908192, + 0.01579257659614086, + 0.00489802798256278, + 0.004240073263645172, + 0.006768074817955494, + 0.008131015114486217, + 0.0057595837861299515, + 0.004552166443318129, + 0.00931560155004263, + 0.010397241450846195, + 0.007357853930443525, + 0.0041678668931126595, + 0.0037694377824664116, + 0.008088703267276287, + 0.004198049195110798, + 0.006037500686943531, + 0.005652968771755695, + 0.005907428450882435, + 0.0032063086982816458, + 0.0036998088471591473, + 0.0050531597808003426, + 0.004223066847771406, + 0.006469589192420244, + 0.0022283282596617937, + 0.005817027762532234, + 0.009263096377253532, + 0.003880333388224244, + 0.00301960832439363, + 0.0029001543298363686, + 0.009985769167542458, + 0.0030082196462899446, + 0.003674101084470749, + 0.022763023152947426, + 0.02825314737856388, + 0.019295578822493553, + 0.010913594625890255, + 0.012287857942283154, + 0.022280525416135788, + 0.012537210248410702, + 0.014352882280945778, + 0.02288271114230156, + 0.026591263711452484, + 0.03422387316823006, + 0.01920846849679947, + 0.01879648119211197, + 0.026303676888346672, + 0.024955829605460167, + 0.010302656330168247, + 0.021975690498948097, + 0.0215724129229784, + 0.03481157496571541, + 0.03150126338005066, + 0.020421018823981285, + 0.01991342008113861, + 0.027068762108683586, + 0.01765270344913006, + 0.01218121126294136, + 0.0101671377196908, + 0.011527438648045063, + 0.01631837897002697, + 0.009505619294941425, + 0.007015656679868698, + 0.013563624583184719, + 0.010907911695539951 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.0028206822462379932 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029787791892886162 + ], + "opposite_logprob_tokens": [ + -3.52978777885437 + ], + "logprob_reward": -0.029787791892886162, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029787791892886162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.127240389585495, + 0.1036236509680748, + 0.2695355713367462, + 0.12005628645420074, + 0.05745980516076088, + 0.032826103270053864, + 0.06730453670024872, + 0.04168189689517021, + 0.09859161078929901, + 0.08168020099401474 + ], + "credit_map": [ + 0.016244778409600258, + 0.01881597191095352, + 0.013808767311275005, + 0.015873389318585396, + 0.012292244471609592, + 0.022610871121287346, + 0.014685339294373989, + 0.012909024022519588, + 0.01494878064841032, + 0.016021504998207092, + 0.013596860691905022, + 0.01642726920545101, + 0.009882169775664806, + 0.010582303628325462, + 0.010719056241214275, + 0.011445707641541958, + 0.04097443446516991, + 0.07602612674236298, + 0.028140980750322342, + 0.02526230178773403, + 0.017309393733739853, + 0.03974128141999245, + 0.01870356686413288, + 0.02337747812271118, + 0.01941903494298458, + 0.02443491294980049, + 0.016940288245677948, + 0.011335100047290325, + 0.007186776492744684, + 0.02024471014738083, + 0.008352180942893028, + 0.012143278494477272, + 0.008477490395307541, + 0.01207644771784544, + 0.009740099310874939, + 0.00454246299341321, + 0.004008605144917965, + 0.00846797414124012, + 0.003855954622849822, + 0.006290771532803774, + 0.004007487557828426, + 0.0035145345609635115, + 0.006451663561165333, + 0.003468422219157219, + 0.0022888644598424435, + 0.004386106040328741, + 0.0032025109976530075, + 0.0055065155029296875, + 0.011540367268025875, + 0.01341952383518219, + 0.007578934542834759, + 0.005302491132169962, + 0.005987685639411211, + 0.009554827585816383, + 0.006079006940126419, + 0.007841701619327068, + 0.005328807048499584, + 0.010378173552453518, + 0.0067085823975503445, + 0.0025771160144358873, + 0.0031762144062668085, + 0.007086408324539661, + 0.0037453824188560247, + 0.0026812145952135324, + 0.01192642655223608, + 0.013299830257892609, + 0.0081834327429533, + 0.010391403920948505, + 0.011510323733091354, + 0.013348150067031384, + 0.020203586667776108, + 0.0097284484654665, + 0.010126947425305843, + 0.007127922959625721, + 0.006109490990638733, + 0.013266089372336864, + 0.01047753170132637, + 0.007886589504778385, + 0.017218835651874542, + 0.00946679525077343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.029787791892886162 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, beginning from a slightly upward angle and slowly advancing forward, transports us into a world shrouded in a milky morning mist. A weathered arched stone bridge spans the center of the frame, its slippery slabs dotted with moss, its railings etched with the marks of age. In the center of the bridge, an elderly man, silver hair pulled back into a neat bun, dressed in a loose-fitting white silk training suit, is immersed in a Tai Chi practice. His face is gaunt, wrinkled with wisdom, his eyes slightly closed, his expression serene and focused. With his slow, fluid movements, his arms trace gentle arcs like flowing clouds and water, his weight shifting steadily and effortlessly between his legs, his wide sleeves fluttering softly. The first soft rays of morning light penetrated the mist, casting a faint golden halo on his outline and silver hair, forming a subtle contrast with the cool, gray-green background tones around him. The whole picture was filled with an otherworldly sense of tranquility and concentration, as if time had frozen at this moment.", + "seed": 12203, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_6/index_488_seed_12203_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02577099474146962, + "vlm_logprob": -0.02577099474146962, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.02577099474146962, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07293464243412018 + ], + "opposite_logprob_tokens": [ + -2.947934627532959 + ], + "logprob_reward": -0.07293464243412018, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07293464243412018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1146305575966835, + 0.07250562310218811, + 0.02303747646510601, + 0.06329284608364105, + 0.10518957674503326, + 0.17270386219024658, + 0.20843806862831116, + 0.019137777388095856, + 0.07884368300437927, + 0.14222058653831482 + ], + "credit_map": [ + 0.015810763463377953, + 0.026075005531311035, + 0.0072910236194729805, + 0.01033149752765894, + 0.019124463200569153, + 0.017558716237545013, + 0.010231989435851574, + 0.008207100443542004, + 0.010009564459323883, + 0.010835904628038406, + 0.006534514017403126, + 0.006503625772893429, + 0.013960979878902435, + 0.008840486407279968, + 0.00740817328915, + 0.00841237511485815, + 0.00422359723597765, + 0.0035124158021062613, + 0.0020573933143168688, + 0.0015482584713026881, + 0.0029181530699133873, + 0.003948500845581293, + 0.0024035372771322727, + 0.0024256196338683367, + 0.008410073816776276, + 0.009476379491388798, + 0.005020135547965765, + 0.006964075844734907, + 0.012686803936958313, + 0.006987440399825573, + 0.009573686867952347, + 0.004174249246716499, + 0.036940593272447586, + 0.009612508118152618, + 0.002476533642038703, + 0.002930769231170416, + 0.007296726573258638, + 0.035927675664424896, + 0.00515394052490592, + 0.00485082296654582, + 0.05371159687638283, + 0.07071121782064438, + 0.004607102368026972, + 0.0038012396544218063, + 0.009563987143337727, + 0.020005570724606514, + 0.004583029542118311, + 0.005720124579966068, + 0.05377944931387901, + 0.07985758781433105, + 0.006439908407628536, + 0.005887146107852459, + 0.009900557808578014, + 0.04050777852535248, + 0.00623276736587286, + 0.005832839757204056, + 0.0027530305087566376, + 0.0030853634234517813, + 0.002017743419855833, + 0.001584332319907844, + 0.0020456817001104355, + 0.0033596076536923647, + 0.002252807840704918, + 0.002039210870862007, + 0.011760981753468513, + 0.01117662526667118, + 0.00646376796066761, + 0.010833868756890297, + 0.012230520136654377, + 0.005008226726204157, + 0.012974642217159271, + 0.008395051583647728, + 0.015772823244333267, + 0.019952617585659027, + 0.015293688513338566, + 0.01570720411837101, + 0.02543327398598194, + 0.00917631946504116, + 0.01990598998963833, + 0.020978674292564392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.07293464243412018 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018201233819127083 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.018201233819127083, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018201233819127083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2119998037815094, + 0.22360236942768097, + 0.05339812859892845, + 0.08034753054380417, + 0.15704049170017242, + 0.032116055488586426, + 0.06280708312988281, + 0.028404472395777702, + 0.03158646821975708, + 0.11869760602712631 + ], + "credit_map": [ + 0.03309266269207001, + 0.04011267423629761, + 0.01873943954706192, + 0.015446300618350506, + 0.0459137037396431, + 0.03085947036743164, + 0.016822893172502518, + 0.01101266872137785, + 0.03425393998622894, + 0.02708383835852146, + 0.019470373168587685, + 0.019092703238129616, + 0.03940390795469284, + 0.027532078325748444, + 0.029870333150029182, + 0.026895200833678246, + 0.007042902987450361, + 0.006406531203538179, + 0.005544243846088648, + 0.005430587567389011, + 0.007468924392014742, + 0.007725812494754791, + 0.006966962944716215, + 0.006812160834670067, + 0.020517749711871147, + 0.010257232934236526, + 0.0034153368324041367, + 0.0051646362990140915, + 0.011897772550582886, + 0.017994415014982224, + 0.00659363716840744, + 0.004506747703999281, + 0.051899150013923645, + 0.013939755968749523, + 0.0050191027112305164, + 0.005855887196958065, + 0.01162982452660799, + 0.051618777215480804, + 0.009586452506482601, + 0.007491535507142544, + 0.0036913263611495495, + 0.006620577536523342, + 0.0021653512958437204, + 0.0032541172113269567, + 0.005706233438104391, + 0.004979935474693775, + 0.004155323375016451, + 0.0015431924257427454, + 0.01693463884294033, + 0.016254430636763573, + 0.0024432281497865915, + 0.004774151835590601, + 0.0074149309657514095, + 0.004754208028316498, + 0.0070412252098321915, + 0.0031902724876999855, + 0.003613643813878298, + 0.004809264559298754, + 0.0020871986635029316, + 0.002771255560219288, + 0.004713778849691153, + 0.003084469586610794, + 0.005064778495579958, + 0.0022600830998271704, + 0.0036770778242498636, + 0.005302556324750185, + 0.003107303287833929, + 0.003601332660764456, + 0.004527816083282232, + 0.002558167092502117, + 0.004493555054068565, + 0.00431866105645895, + 0.017802445217967033, + 0.02210388518869877, + 0.009601068682968616, + 0.010424229316413403, + 0.02218548208475113, + 0.01259525865316391, + 0.010648687370121479, + 0.013336543925106525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.018201233819127083 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742938421666622 + ], + "opposite_logprob_tokens": [ + -5.00674295425415 + ], + "logprob_reward": -0.006742938421666622, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742938421666622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06226961687207222, + 0.06383066624403, + 0.035519473254680634, + 0.11694759875535965, + 0.20652265846729279, + 0.2314429134130478, + 0.05581650882959366, + 0.08376647531986237, + 0.0369427427649498, + 0.10694140940904617 + ], + "credit_map": [ + 0.009435229934751987, + 0.009025376290082932, + 0.007427959702908993, + 0.006741873454302549, + 0.009770005010068417, + 0.007897447794675827, + 0.005712708458304405, + 0.006259017623960972, + 0.007620139513164759, + 0.00865904986858368, + 0.004710495937615633, + 0.005180466454476118, + 0.016169380396604538, + 0.007793071214109659, + 0.008983985520899296, + 0.004714073613286018, + 0.004495191387832165, + 0.004590952303260565, + 0.0017529056640341878, + 0.0029363331850618124, + 0.009799379855394363, + 0.0049819680862128735, + 0.0044114068150520325, + 0.002551338169723749, + 0.029294798150658607, + 0.02134898118674755, + 0.004903751891106367, + 0.006517964880913496, + 0.011570870876312256, + 0.028816377744078636, + 0.008750188164412975, + 0.005744667258113623, + 0.06445637345314026, + 0.020065462216734886, + 0.006083752028644085, + 0.010136055760085583, + 0.01793275773525238, + 0.06485045701265335, + 0.013328871689736843, + 0.009668929502367973, + 0.061601705849170685, + 0.0855620801448822, + 0.00477879261597991, + 0.01042473316192627, + 0.01754477433860302, + 0.03330318257212639, + 0.011084123514592648, + 0.0071435109712183475, + 0.014056566171348095, + 0.020419200882315636, + 0.0014164070598781109, + 0.0022292183712124825, + 0.0030220667831599712, + 0.010332376696169376, + 0.002935592085123062, + 0.0014050810132175684, + 0.019933369010686874, + 0.02525152638554573, + 0.0033145237248390913, + 0.004956312943249941, + 0.009756110608577728, + 0.009086953476071358, + 0.007628432475030422, + 0.0038392378482967615, + 0.005548068787902594, + 0.005374379921704531, + 0.0029879191424697638, + 0.0048763323575258255, + 0.005417614709585905, + 0.0026253724936395884, + 0.005998642649501562, + 0.004114415030926466, + 0.014199567958712578, + 0.017934996634721756, + 0.008299766108393669, + 0.009700274094939232, + 0.021958302706480026, + 0.010613920167088509, + 0.011379042640328407, + 0.012855539098381996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.006742938421666622 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464076129719615 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.00464076129719615, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464076129719615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09816696494817734, + 0.09965148568153381, + 0.031512852758169174, + 0.17881283164024353, + 0.047401610761880875, + 0.0888364389538765, + 0.05251482501626015, + 0.07196538895368576, + 0.11641211062669754, + 0.214725524187088 + ], + "credit_map": [ + 0.013405141420662403, + 0.020649705082178116, + 0.008233346976339817, + 0.009190605953335762, + 0.013299161568284035, + 0.01993652433156967, + 0.006307858508080244, + 0.007144629023969173, + 0.012777771800756454, + 0.015000026673078537, + 0.008415628224611282, + 0.009798062033951283, + 0.018083248287439346, + 0.011778448708355427, + 0.013822091743350029, + 0.009976215660572052, + 0.004384710919111967, + 0.0038587970193475485, + 0.0027042042929679155, + 0.003390568308532238, + 0.005495980381965637, + 0.00470309890806675, + 0.003863319056108594, + 0.0031121757347136736, + 0.054504550993442535, + 0.03087497688829899, + 0.005632671993225813, + 0.0073673175647854805, + 0.011676090769469738, + 0.05452905595302582, + 0.006530766375362873, + 0.007697393652051687, + 0.011056908406317234, + 0.006426386069506407, + 0.0018648896366357803, + 0.0029460119549185038, + 0.0061120460741221905, + 0.011520625092089176, + 0.004641866777092218, + 0.0028328760527074337, + 0.020892122760415077, + 0.03547578305006027, + 0.001809046952985227, + 0.002451918786391616, + 0.00643707811832428, + 0.01612447202205658, + 0.003250839887186885, + 0.002395173069089651, + 0.010250462219119072, + 0.007094184402376413, + 0.0026484851259738207, + 0.004265390336513519, + 0.006745757535099983, + 0.011328679509460926, + 0.007468254771083593, + 0.0027136087883263826, + 0.017672264948487282, + 0.024043088778853416, + 0.0026331341359764338, + 0.0033907312899827957, + 0.007323864847421646, + 0.00887890625745058, + 0.005024482496082783, + 0.0029989106114953756, + 0.018680693581700325, + 0.01720574125647545, + 0.012138744816184044, + 0.013026154600083828, + 0.014734460972249508, + 0.008138051256537437, + 0.016627417877316475, + 0.01586083322763443, + 0.02664673514664173, + 0.0339493528008461, + 0.012824528850615025, + 0.03210067003965378, + 0.047392964363098145, + 0.02024233713746071, + 0.026531178504228592, + 0.015037749893963337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.00464076129719615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a long shot that slowly advances forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver hair in a neat bun visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose-fitting white silk training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man performing slow, fluid Tai Chi movements with wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a weathered arched stone bridge with mossy slabs in a milky morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026335397735238075 + ], + "opposite_logprob_tokens": [ + -3.6513354778289795 + ], + "logprob_reward": -0.026335397735238075, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026335397735238075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13263356685638428, + 0.11475523561239243, + 0.016333453357219696, + 0.034789085388183594, + 0.058768659830093384, + 0.10457991808652878, + 0.4405510723590851, + 0.02298247255384922, + 0.029487375169992447, + 0.04511919617652893 + ], + "credit_map": [ + 0.01714218594133854, + 0.027786502614617348, + 0.013777989894151688, + 0.014098073355853558, + 0.014455706812441349, + 0.022347912192344666, + 0.011286823078989983, + 0.011738372966647148, + 0.019131481647491455, + 0.014675210230052471, + 0.015186703763902187, + 0.013431502506136894, + 0.014596914872527122, + 0.012091845273971558, + 0.015591291710734367, + 0.010050279088318348, + 0.002307398011907935, + 0.0022725064773112535, + 0.002219217596575618, + 0.001608343212865293, + 0.0017112436471506953, + 0.002441967837512493, + 0.0018587998347356915, + 0.0019139769719913602, + 0.005933698266744614, + 0.006628640927374363, + 0.0017790324054658413, + 0.00319095142185688, + 0.0050940243527293205, + 0.005787841510027647, + 0.004740254953503609, + 0.001634641783311963, + 0.012287246063351631, + 0.009331471286714077, + 0.0037187207490205765, + 0.004085969179868698, + 0.006807214580476284, + 0.014079917222261429, + 0.004929560236632824, + 0.0035285602789372206, + 0.03209729120135307, + 0.02576327696442604, + 0.007838084362447262, + 0.004449952393770218, + 0.007903963327407837, + 0.012361799366772175, + 0.004033708944916725, + 0.0101318359375, + 0.11289074271917343, + 0.17290014028549194, + 0.009934843517839909, + 0.011318176984786987, + 0.020002633333206177, + 0.09148313850164413, + 0.011903115548193455, + 0.01011828426271677, + 0.005720080342143774, + 0.007126667071133852, + 0.0012668472481891513, + 0.0012706532143056393, + 0.0019657837692648172, + 0.0027867844328284264, + 0.001445877831429243, + 0.0013997782953083515, + 0.004711532033979893, + 0.004467991646379232, + 0.003667276119813323, + 0.0034683551639318466, + 0.003034684807062149, + 0.0022910083644092083, + 0.0034895106218755245, + 0.004357016645371914, + 0.006500902585685253, + 0.0070548043586313725, + 0.004400641191750765, + 0.002971170237287879, + 0.00709085026755929, + 0.004961862228810787, + 0.0038863851223140955, + 0.0082525834441185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.026335397735238075 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_7/index_tensor([8555, 8555])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_7/index_tensor([8555, 8555])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0a0d3e0dc7cf09bdff1dd599fc390b3024d71a25 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_7/index_tensor([8555, 8555])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level wide-angle shot captures a stunning desert oasis, a living jewel amidst a vast, arid expanse. In the foreground, a pristine, glassy pool of water lies completely still, its surface a perfect mirror, reflecting the deep blue of the sky and the vibrant green of the surrounding palm trees. A cluster of tall, lush date palms, their fibrous brown trunks rising proudly, frame the water's edge, their dense, emerald-green fronds fanning out like intricate lace. Beyond this verdant haven, vast, rolling dunes of fine, golden-orange sand stretch endlessly to the distant, razor-sharp horizon, where the earth meets the brilliant, completely cloudless azure sky. The entire scene is bathed in the brilliant, unfiltered sunlight of midday, casting sharp, dark shadows beneath the palm trees. The only perceptible movement is the slow, gentle swaying of the palm fronds, which drift lazily in the breeze. This delicate dance contrasts with the absolute stillness of the water, creating a profound atmosphere of tranquility and serenity.", + "seed": 33389, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_7/index_8555_seed_33389_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.040523398166988044, + "vlm_logprob": -0.040523398166988044, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.040523398166988044, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004002180881798267 + ], + "opposite_logprob_tokens": [ + -6.504002094268799 + ], + "logprob_reward": -0.004002180881798267, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.004002180881798267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22390154004096985, + 0.09856432676315308, + 0.07878239452838898, + 0.12671078741550446, + 0.06512713432312012, + 0.07712212949991226, + 0.1161816269159317, + 0.14533677697181702, + 0.03363008797168732, + 0.03464319556951523 + ], + "credit_map": [ + 0.034311313182115555, + 0.02928801067173481, + 0.0241486057639122, + 0.019798001274466515, + 0.02510075643658638, + 0.030279291793704033, + 0.02120724320411682, + 0.039768319576978683, + 0.01083105243742466, + 0.01833846978843212, + 0.008664078079164028, + 0.009094052016735077, + 0.009912197478115559, + 0.0131897097453475, + 0.011582602746784687, + 0.016952168196439743, + 0.008144532330334187, + 0.013927760533988476, + 0.007360019721090794, + 0.006454683840274811, + 0.006919564679265022, + 0.010621240362524986, + 0.007807031273841858, + 0.01754756085574627, + 0.014719905331730843, + 0.019316285848617554, + 0.013566941022872925, + 0.012967231683433056, + 0.018358036875724792, + 0.014329738914966583, + 0.011684222146868706, + 0.02176842838525772, + 0.006407753564417362, + 0.011622671969234943, + 0.007984543219208717, + 0.006670665927231312, + 0.005559498909860849, + 0.007408888079226017, + 0.006540745031088591, + 0.012932371348142624, + 0.008142405189573765, + 0.010918550193309784, + 0.008108899928629398, + 0.007397190202027559, + 0.007195618934929371, + 0.008269356563687325, + 0.013348923996090889, + 0.013741178438067436, + 0.01514551043510437, + 0.018006736412644386, + 0.015113373287022114, + 0.011742595583200455, + 0.013391304761171341, + 0.0154029605910182, + 0.014358093030750751, + 0.013021060265600681, + 0.017914457246661186, + 0.017094960436224937, + 0.015554020181298256, + 0.011548659764230251, + 0.015749627724289894, + 0.018642647191882133, + 0.021485624834895134, + 0.0273467805236578, + 0.003508967347443104, + 0.006126613821834326, + 0.0033201260957866907, + 0.0026768939569592476, + 0.0055580176413059235, + 0.00440974673256278, + 0.002864319831132889, + 0.005165402777493, + 0.0042761266231536865, + 0.00688349362462759, + 0.004061602056026459, + 0.004151613917201757, + 0.0034938727039843798, + 0.0038099680095911026, + 0.003918349742889404, + 0.004048169124871492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.004002180881798267 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008229206432588398 + ], + "opposite_logprob_tokens": [ + -7.125823020935059 + ], + "logprob_reward": -0.0008229206432588398, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008229206432588398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20647916197776794, + 0.1410144418478012, + 0.08095230907201767, + 0.09593567997217178, + 0.0756167620420456, + 0.09252430498600006, + 0.06583751738071442, + 0.08731509745121002, + 0.10061727464199066, + 0.053707461804151535 + ], + "credit_map": [ + 0.02510502003133297, + 0.027982160449028015, + 0.03275549039244652, + 0.020848315209150314, + 0.024711214005947113, + 0.022707365453243256, + 0.020990557968616486, + 0.03137902542948723, + 0.014501411467790604, + 0.022460831329226494, + 0.011479560285806656, + 0.013742092065513134, + 0.012713543139398098, + 0.019931336864829063, + 0.016517259180545807, + 0.029668418690562248, + 0.011509397998452187, + 0.009160046465694904, + 0.007963085547089577, + 0.00846842210739851, + 0.011410565115511417, + 0.008651498705148697, + 0.010684730485081673, + 0.013104569166898727, + 0.010207029059529305, + 0.010411221534013748, + 0.008565808646380901, + 0.010856729932129383, + 0.00859435647726059, + 0.011845583096146584, + 0.009813821874558926, + 0.02564113587141037, + 0.008327222429215908, + 0.010443653911352158, + 0.008043399080634117, + 0.007650915998965502, + 0.010438883677124977, + 0.009408196434378624, + 0.008600981906056404, + 0.01270350906997919, + 0.007916383445262909, + 0.012976277619600296, + 0.010284362360835075, + 0.014962373301386833, + 0.006563117261976004, + 0.009672969579696655, + 0.010166570544242859, + 0.019982242956757545, + 0.008799132890999317, + 0.012000549584627151, + 0.00806585792452097, + 0.006629341281950474, + 0.007305601611733437, + 0.008817538619041443, + 0.006451533641666174, + 0.007767967414110899, + 0.008635773323476315, + 0.010175741277635098, + 0.008996185846626759, + 0.008122414350509644, + 0.00859161652624607, + 0.00971220526844263, + 0.017460724338889122, + 0.015620436519384384, + 0.009767446666955948, + 0.013621651567518711, + 0.009595837444067001, + 0.006055998150259256, + 0.009839171543717384, + 0.014397088438272476, + 0.011391259729862213, + 0.025948816910386086, + 0.007305492181330919, + 0.006617520935833454, + 0.00590392155572772, + 0.004378302022814751, + 0.005661384202539921, + 0.006956939585506916, + 0.005159132182598114, + 0.011724771000444889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0008229206432588398 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011066842824220657 + ], + "opposite_logprob_tokens": [ + -4.511066913604736 + ], + "logprob_reward": -0.011066842824220657, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011066842824220657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3128166198730469, + 0.12007375061511993, + 0.16181007027626038, + 0.14937688410282135, + 0.04775112867355347, + 0.05802096053957939, + 0.04203776270151138, + 0.0345771498978138, + 0.04036998003721237, + 0.03316568583250046 + ], + "credit_map": [ + 0.03955042362213135, + 0.07461702078580856, + 0.0292363278567791, + 0.025000860914587975, + 0.034054458141326904, + 0.04595063254237175, + 0.02313317358493805, + 0.04127371311187744, + 0.02025039866566658, + 0.01723076030611992, + 0.012456188909709454, + 0.012375466525554657, + 0.013985862024128437, + 0.015826698392629623, + 0.013138202019035816, + 0.014810166321694851, + 0.01967158541083336, + 0.017082978039979935, + 0.0113325584679842, + 0.010558253154158592, + 0.014423600398004055, + 0.02401488460600376, + 0.01641746051609516, + 0.048308756202459335, + 0.01806284859776497, + 0.010973531752824783, + 0.010736297816038132, + 0.010950133204460144, + 0.01176274474710226, + 0.02117512747645378, + 0.015153258107602596, + 0.05056294798851013, + 0.004480320960283279, + 0.004102996550500393, + 0.0053347148932516575, + 0.00370279373601079, + 0.004322092514485121, + 0.005691756960004568, + 0.008098090998828411, + 0.012018360197544098, + 0.0047967033460736275, + 0.006511217448860407, + 0.006792047061026096, + 0.007399830035865307, + 0.004874120000749826, + 0.006479967851191759, + 0.007155743893235922, + 0.01401133369654417, + 0.00501578813418746, + 0.00628979317843914, + 0.0053685917519032955, + 0.004722364712506533, + 0.004864907823503017, + 0.00563359260559082, + 0.004962901584804058, + 0.0051798224449157715, + 0.0040264190174639225, + 0.005025168880820274, + 0.00452730618417263, + 0.0027542184107005596, + 0.004473861772567034, + 0.004351076669991016, + 0.004457300063222647, + 0.004961799364537001, + 0.003539319382980466, + 0.003407447598874569, + 0.004182462580502033, + 0.002378537319600582, + 0.0038127789739519358, + 0.005939935799688101, + 0.004566027782857418, + 0.012543471530079842, + 0.004164442420005798, + 0.006927611771970987, + 0.004154632333666086, + 0.002916947240009904, + 0.0028333282098174095, + 0.004078736994415522, + 0.0030280801001936197, + 0.005061904899775982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.011066842824220657 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263565294444561 + ], + "opposite_logprob_tokens": [ + -3.6513564586639404 + ], + "logprob_reward": -0.0263565294444561, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263565294444561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25712841749191284, + 0.13642992079257965, + 0.15877243876457214, + 0.10098107904195786, + 0.0832076221704483, + 0.04333726316690445, + 0.0651962012052536, + 0.06900005787611008, + 0.05312766134738922, + 0.03281928971409798 + ], + "credit_map": [ + 0.03351357579231262, + 0.06576158851385117, + 0.022904416546225548, + 0.019410695880651474, + 0.02990514039993286, + 0.038036614656448364, + 0.016278840601444244, + 0.03131754323840141, + 0.012730265967547894, + 0.03404804319143295, + 0.014022325165569782, + 0.013172974810004234, + 0.013689115643501282, + 0.01961277797818184, + 0.01623954251408577, + 0.012914883904159069, + 0.018968548625707626, + 0.012998818419873714, + 0.013996771536767483, + 0.017597638070583344, + 0.01286868192255497, + 0.02087223529815674, + 0.018447335809469223, + 0.043022435158491135, + 0.011388235725462437, + 0.008266144432127476, + 0.007889213040471077, + 0.009833498857915401, + 0.009244034998118877, + 0.013116398826241493, + 0.011656675487756729, + 0.02958688884973526, + 0.009018500335514545, + 0.010818392038345337, + 0.010276558808982372, + 0.008784686215221882, + 0.007098446134477854, + 0.008882420137524605, + 0.013284968212246895, + 0.01504365261644125, + 0.004354021977633238, + 0.004157466348260641, + 0.005773958750069141, + 0.004532746039330959, + 0.0032425473909825087, + 0.0050313882529735565, + 0.0048430077731609344, + 0.011402131989598274, + 0.006481442134827375, + 0.01461769174784422, + 0.006945205852389336, + 0.007824537344276905, + 0.007937893271446228, + 0.008318128995597363, + 0.0068579139187932014, + 0.0062133874744176865, + 0.007596627809107304, + 0.007608918938785791, + 0.00776782538741827, + 0.006658061873167753, + 0.0080740200355649, + 0.00799382384866476, + 0.010889960452914238, + 0.01241081953048706, + 0.006180599797517061, + 0.006557714659720659, + 0.0049854363314807415, + 0.0041778539307415485, + 0.007909923791885376, + 0.006904824171215296, + 0.006166471168398857, + 0.010244838893413544, + 0.004555790219455957, + 0.007731454446911812, + 0.003399553010240197, + 0.002012828132137656, + 0.00289990846067667, + 0.004289972595870495, + 0.0030954640824347734, + 0.004834318999201059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0263565294444561 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036851704120636 + ], + "opposite_logprob_tokens": [ + -1.91036856174469 + ], + "logprob_reward": -0.16036851704120636, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036851704120636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22055262327194214, + 0.16391701996326447, + 0.08555072546005249, + 0.1416209489107132, + 0.07486535608768463, + 0.04973540082573891, + 0.056620873510837555, + 0.058683015406131744, + 0.08332837373018265, + 0.06512568891048431 + ], + "credit_map": [ + 0.02019071951508522, + 0.03755801543593407, + 0.024639887735247612, + 0.02762761525809765, + 0.01703486032783985, + 0.027906108647584915, + 0.020791027694940567, + 0.0448044016957283, + 0.01589024066925049, + 0.03204871714115143, + 0.013710235245525837, + 0.016176670789718628, + 0.01526451576501131, + 0.023068420588970184, + 0.01585252210497856, + 0.03190568834543228, + 0.008914435282349586, + 0.019025269895792007, + 0.007390462793409824, + 0.008409036323428154, + 0.007408709730952978, + 0.011917788535356522, + 0.010681239888072014, + 0.011803782545030117, + 0.016104204580187798, + 0.012432740069925785, + 0.01359604112803936, + 0.014145158231258392, + 0.01742781512439251, + 0.017679059877991676, + 0.014220179989933968, + 0.036015741527080536, + 0.00954620260745287, + 0.011022988706827164, + 0.010712482035160065, + 0.00633035134524107, + 0.0070767174474895, + 0.008418213576078415, + 0.007197825238108635, + 0.014560575596988201, + 0.005248252768069506, + 0.008016502484679222, + 0.006020150613039732, + 0.006851166486740112, + 0.004355221521109343, + 0.0055062612518668175, + 0.0049938964657485485, + 0.008743946440517902, + 0.0058748298324644566, + 0.007521315012127161, + 0.005634615197777748, + 0.008309366181492805, + 0.004431352484971285, + 0.006516339723020792, + 0.010687866248190403, + 0.0076451851055026054, + 0.004875333048403263, + 0.008595221675932407, + 0.005970322526991367, + 0.0061779809184372425, + 0.004534050356596708, + 0.007124943658709526, + 0.01154111884534359, + 0.009864043444395065, + 0.009416059590876102, + 0.008404115214943886, + 0.0084372004494071, + 0.009285924956202507, + 0.009084751829504967, + 0.010515567846596241, + 0.010832599364221096, + 0.01735215075314045, + 0.008274571970105171, + 0.012900693342089653, + 0.00890919379889965, + 0.006638803519308567, + 0.005825691856443882, + 0.008066212758421898, + 0.006556154694408178, + 0.007954363711178303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.16036851704120636 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level wide-angle shot captures a stunning desert oasis, a living jewel amidst a vast, arid expanse. In the foreground, a pristine, glassy pool of water lies completely still, its surface a perfect mirror, reflecting the deep blue of the sky and the vibrant green of the surrounding palm trees. A cluster of tall, lush date palms, their fibrous brown trunks rising proudly, frame the water's edge, their dense, emerald-green fronds fanning out like intricate lace. Beyond this verdant haven, vast, rolling dunes of fine, golden-orange sand stretch endlessly to the distant, razor-sharp horizon, where the earth meets the brilliant, completely cloudless azure sky. The entire scene is bathed in the brilliant, unfiltered sunlight of midday, casting sharp, dark shadows beneath the palm trees. The only perceptible movement is the slow, gentle swaying of the palm fronds, which drift lazily in the breeze. This delicate dance contrasts with the absolute stillness of the water, creating a profound atmosphere of tranquility and serenity.", + "seed": 33390, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_7/index_8555_seed_33390_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0214300240855664, + "vlm_logprob": -0.0214300240855664, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.0214300240855664, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01046349760144949 + ], + "opposite_logprob_tokens": [ + -4.885463714599609 + ], + "logprob_reward": -0.01046349760144949, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01046349760144949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3347311019897461, + 0.14897452294826508, + 0.05361533164978027, + 0.0627286285161972, + 0.0509527213871479, + 0.04593528062105179, + 0.05189115181565285, + 0.09727798402309418, + 0.10197655111551285, + 0.05191672220826149 + ], + "credit_map": [ + 0.0276749599725008, + 0.03759140893816948, + 0.02699168026447296, + 0.02386491186916828, + 0.040505632758140564, + 0.050418511033058167, + 0.03619538992643356, + 0.09148859232664108, + 0.016192691400647163, + 0.0183491799980402, + 0.012767069973051548, + 0.016338076442480087, + 0.016215745359659195, + 0.019360994920134544, + 0.016384687274694443, + 0.033366069197654724, + 0.007092313840985298, + 0.008316949009895325, + 0.0031351977959275246, + 0.004047012887895107, + 0.005688318982720375, + 0.006877187639474869, + 0.007478324696421623, + 0.010980024002492428, + 0.005908978637307882, + 0.00558696873486042, + 0.003993376158177853, + 0.0036912451032549143, + 0.006043901201337576, + 0.012157841585576534, + 0.0072897328063845634, + 0.018056588247418404, + 0.00623156875371933, + 0.006105052772909403, + 0.004664757288992405, + 0.005666963756084442, + 0.007218468002974987, + 0.0063696554861962795, + 0.006210941821336746, + 0.008485307916998863, + 0.006765465252101421, + 0.0029077844228595495, + 0.0032068269792944193, + 0.004175347741693258, + 0.003708453383296728, + 0.006636857520788908, + 0.005863120313733816, + 0.012671424075961113, + 0.010502802208065987, + 0.007244364358484745, + 0.0041619292460381985, + 0.004121501464396715, + 0.006822295021265745, + 0.005669916048645973, + 0.007158266846090555, + 0.006210080813616514, + 0.02462996169924736, + 0.009852245450019836, + 0.010887516662478447, + 0.006560411304235458, + 0.01053793728351593, + 0.013119105249643326, + 0.012143095023930073, + 0.009547710418701172, + 0.015330003574490547, + 0.012477966956794262, + 0.010033235885202885, + 0.008647708222270012, + 0.013946836814284325, + 0.014470400288701057, + 0.014397921971976757, + 0.012672479264438152, + 0.005566922947764397, + 0.005005300510674715, + 0.004370748531073332, + 0.0035223988816142082, + 0.00711920065805316, + 0.0069129373878240585, + 0.005692038685083389, + 0.013727170415222645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.01046349760144949 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008247073274105787 + ], + "opposite_logprob_tokens": [ + -7.125824928283691 + ], + "logprob_reward": -0.0008247073274105787, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008247073274105787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14463308453559875, + 0.07066753506660461, + 0.10921698063611984, + 0.10272721201181412, + 0.08821356296539307, + 0.05417534336447716, + 0.06992509961128235, + 0.020777234807610512, + 0.18476514518260956, + 0.15489880740642548 + ], + "credit_map": [ + 0.016971006989479065, + 0.016578737646341324, + 0.012355221435427666, + 0.013489920645952225, + 0.014253532513976097, + 0.02348526567220688, + 0.014609378762543201, + 0.03289002552628517, + 0.009118209592998028, + 0.008852388709783554, + 0.0063947224989533424, + 0.005946298595517874, + 0.009049344807863235, + 0.007955901324748993, + 0.007329996675252914, + 0.016020674258470535, + 0.01552581787109375, + 0.019593246281147003, + 0.007717452477663755, + 0.00903151836246252, + 0.012134288437664509, + 0.01761821284890175, + 0.013439641334116459, + 0.014156799763441086, + 0.012678108178079128, + 0.008013698272407055, + 0.0062146312557160854, + 0.006671641953289509, + 0.009928560815751553, + 0.01655702479183674, + 0.012558777816593647, + 0.03010476566851139, + 0.010350115597248077, + 0.009928019717335701, + 0.011309344321489334, + 0.00873253121972084, + 0.012975485995411873, + 0.012545600533485413, + 0.01069699227809906, + 0.011675474233925343, + 0.00924649741500616, + 0.004929003305733204, + 0.00562388775870204, + 0.005542099475860596, + 0.005178690422326326, + 0.00756625272333622, + 0.006274087354540825, + 0.009814823046326637, + 0.013194885104894638, + 0.009391152299940586, + 0.006980961188673973, + 0.006231484469026327, + 0.007558675482869148, + 0.007587347645312548, + 0.009655101224780083, + 0.009325491264462471, + 0.00418930035084486, + 0.0022454974241554737, + 0.00227765878662467, + 0.001627858611755073, + 0.0020863679237663746, + 0.0025879424065351486, + 0.0020017516799271107, + 0.0037608565762639046, + 0.022868243977427483, + 0.011129516176879406, + 0.015380450524389744, + 0.01177940797060728, + 0.021173415705561638, + 0.028930822387337685, + 0.020789392292499542, + 0.052713893353939056, + 0.01677069067955017, + 0.010950412601232529, + 0.010505614802241325, + 0.008942571468651295, + 0.019197799265384674, + 0.023471860215067863, + 0.017394904047250748, + 0.047664955258369446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0008247073274105787 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011070143431425095 + ], + "opposite_logprob_tokens": [ + -4.511070251464844 + ], + "logprob_reward": -0.011070143431425095, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011070143431425095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10840737819671631, + 0.0835452526807785, + 0.06153231859207153, + 0.14801153540611267, + 0.09194935858249664, + 0.04576142504811287, + 0.03714890033006668, + 0.022901903837919235, + 0.20557433366775513, + 0.19516761600971222 + ], + "credit_map": [ + 0.010172148235142231, + 0.009277337230741978, + 0.007806790992617607, + 0.007733131758868694, + 0.008727057836949825, + 0.019380085170269012, + 0.011130712926387787, + 0.034180112183094025, + 0.009447108022868633, + 0.009427065961062908, + 0.007320302538573742, + 0.007069924846291542, + 0.00989482644945383, + 0.010235688649117947, + 0.008006184361875057, + 0.02214413695037365, + 0.005442436784505844, + 0.008493542671203613, + 0.00396802369505167, + 0.003924498334527016, + 0.007313033565878868, + 0.010393200442194939, + 0.0065037584863603115, + 0.015493816696107388, + 0.013881382532417774, + 0.010978526435792446, + 0.008629487827420235, + 0.008407233282923698, + 0.01315336488187313, + 0.026214037090539932, + 0.017359087243676186, + 0.04938840866088867, + 0.011374865658581257, + 0.00797487422823906, + 0.008901561610400677, + 0.006491236854344606, + 0.009644011966884136, + 0.01534630823880434, + 0.01198575645685196, + 0.020230738446116447, + 0.004874886479228735, + 0.003542625345289707, + 0.004098684061318636, + 0.0036646586377173662, + 0.0047719781287014484, + 0.008933932520449162, + 0.005903774406760931, + 0.009970884770154953, + 0.004044739995151758, + 0.00313725508749485, + 0.0025390316732227802, + 0.0026828646659851074, + 0.003813779214397073, + 0.005891844164580107, + 0.004877990111708641, + 0.010161392390727997, + 0.004567344207316637, + 0.0027330259326845407, + 0.0028299412224441767, + 0.001432204619050026, + 0.0024473515804857016, + 0.0028435117565095425, + 0.002647706773132086, + 0.0034008140210062265, + 0.023191990330815315, + 0.012248429469764233, + 0.016907280310988426, + 0.014038624241948128, + 0.02407759241759777, + 0.03184942156076431, + 0.02360738068819046, + 0.05965357646346092, + 0.02056494913995266, + 0.013079840689897537, + 0.012473562732338905, + 0.011981330811977386, + 0.022877080366015434, + 0.029144294559955597, + 0.023659689351916313, + 0.06138685718178749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.011070143431425095 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498442426323891 + ], + "opposite_logprob_tokens": [ + -2.9299843311309814 + ], + "logprob_reward": -0.05498442426323891, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498442426323891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11196304857730865, + 0.07568462938070297, + 0.03309765085577965, + 0.13596715033054352, + 0.021933984011411667, + 0.049297530204057693, + 0.05728884041309357, + 0.02996964380145073, + 0.17534148693084717, + 0.3094560205936432 + ], + "credit_map": [ + 0.010554616339504719, + 0.008554699830710888, + 0.007933625020086765, + 0.008000734262168407, + 0.01001653727144003, + 0.02119426615536213, + 0.011597019620239735, + 0.03411155194044113, + 0.008510688319802284, + 0.009715089574456215, + 0.005641474388539791, + 0.005868696607649326, + 0.007323086261749268, + 0.009444516152143478, + 0.00754792895168066, + 0.021633150056004524, + 0.0031311737839132547, + 0.004151484463363886, + 0.002467390615493059, + 0.0024410237092524767, + 0.003037493210285902, + 0.006173964589834213, + 0.003464414505288005, + 0.008230708539485931, + 0.013060532510280609, + 0.01094946637749672, + 0.008161082863807678, + 0.008481768891215324, + 0.011272217147052288, + 0.022703630849719048, + 0.016923831775784492, + 0.044414620846509933, + 0.002801369410008192, + 0.0015837694518268108, + 0.0016888901591300964, + 0.001134512829594314, + 0.0018840443808585405, + 0.0034767198376357555, + 0.0027466497849673033, + 0.006618030834943056, + 0.005106596741825342, + 0.004340946674346924, + 0.0042986138723790646, + 0.003919500391930342, + 0.004597831517457962, + 0.007398488000035286, + 0.006489318795502186, + 0.01314623560756445, + 0.009186762385070324, + 0.005472885444760323, + 0.0055459593422710896, + 0.0047660451382398605, + 0.005843660794198513, + 0.0077734701335430145, + 0.005811571143567562, + 0.012888485565781593, + 0.005241253413259983, + 0.0035992786288261414, + 0.0025533775333315134, + 0.00196596491150558, + 0.0027187191881239414, + 0.005167922005057335, + 0.0026728154625743628, + 0.006050314754247665, + 0.01791643723845482, + 0.010007093660533428, + 0.012512607499957085, + 0.011622462421655655, + 0.019681677222251892, + 0.02975665032863617, + 0.01698911003768444, + 0.05685547739267349, + 0.03314224258065224, + 0.019426964223384857, + 0.018339725211262703, + 0.0186370387673378, + 0.037586260586977005, + 0.05119799077510834, + 0.0363524854183197, + 0.0947733148932457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.05498442426323891 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029807347804307938 + ], + "opposite_logprob_tokens": [ + -3.5298073291778564 + ], + "logprob_reward": -0.029807347804307938, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807347804307938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15896964073181152, + 0.09897169470787048, + 0.18372943997383118, + 0.09267014265060425, + 0.06443312019109726, + 0.046470414847135544, + 0.03990482911467552, + 0.03368237614631653, + 0.1661854237318039, + 0.11498290300369263 + ], + "credit_map": [ + 0.014457288198173046, + 0.019692551344633102, + 0.012860485352575779, + 0.015468008816242218, + 0.021991576999425888, + 0.021145300939679146, + 0.018054163083434105, + 0.035300273448228836, + 0.01200932078063488, + 0.013173251412808895, + 0.012028985656797886, + 0.009437336586415768, + 0.013326844200491905, + 0.012208810076117516, + 0.01257668249309063, + 0.01421046257019043, + 0.025610268115997314, + 0.02924429439008236, + 0.016828786581754684, + 0.01662060059607029, + 0.02148362621665001, + 0.029484141618013382, + 0.019534796476364136, + 0.02492292784154415, + 0.00869877077639103, + 0.007423533126711845, + 0.007873790338635445, + 0.005917510949075222, + 0.007735228631645441, + 0.01662541925907135, + 0.010883679613471031, + 0.027512211352586746, + 0.008202608674764633, + 0.0056276628747582436, + 0.005486858543008566, + 0.004995056428015232, + 0.008519968949258327, + 0.00827836338430643, + 0.0084491902962327, + 0.014873410575091839, + 0.010962224565446377, + 0.004873672500252724, + 0.0042480467818677425, + 0.003031842177733779, + 0.006414731498807669, + 0.005087088327854872, + 0.005629976280033588, + 0.006222832482308149, + 0.00494548212736845, + 0.004250045400112867, + 0.00330758118070662, + 0.0029730387032032013, + 0.004750972148030996, + 0.0056333341635763645, + 0.004046052228659391, + 0.009998323395848274, + 0.006117905490100384, + 0.0055236658081412315, + 0.0026855890173465014, + 0.002863154746592045, + 0.003050070023164153, + 0.0046483115293085575, + 0.00332861440256238, + 0.0054650683887302876, + 0.017992334440350533, + 0.009761654771864414, + 0.012941191904246807, + 0.01070212759077549, + 0.019720153883099556, + 0.02692633680999279, + 0.019423527643084526, + 0.048718106001615524, + 0.01247446145862341, + 0.008120601996779442, + 0.00830706488341093, + 0.00775372376665473, + 0.01313637476414442, + 0.018552595749497414, + 0.014679531566798687, + 0.031958550214767456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.029807347804307938 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level wide-angle shot captures a stunning desert oasis, a living jewel amidst a vast, arid expanse. In the foreground, a pristine, glassy pool of water lies completely still, its surface a perfect mirror, reflecting the deep blue of the sky and the vibrant green of the surrounding palm trees. A cluster of tall, lush date palms, their fibrous brown trunks rising proudly, frame the water's edge, their dense, emerald-green fronds fanning out like intricate lace. Beyond this verdant haven, vast, rolling dunes of fine, golden-orange sand stretch endlessly to the distant, razor-sharp horizon, where the earth meets the brilliant, completely cloudless azure sky. The entire scene is bathed in the brilliant, unfiltered sunlight of midday, casting sharp, dark shadows beneath the palm trees. The only perceptible movement is the slow, gentle swaying of the palm fronds, which drift lazily in the breeze. This delicate dance contrasts with the absolute stillness of the water, creating a profound atmosphere of tranquility and serenity.", + "seed": 33391, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_7/index_8555_seed_33391_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02409396630246192, + "vlm_logprob": -0.02409396630246192, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.02409396630246192, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004143225960433483 + ], + "opposite_logprob_tokens": [ + -6.129143238067627 + ], + "logprob_reward": -0.004143225960433483, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.004143225960433483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2871303856372833, + 0.13019654154777527, + 0.20112541317939758, + 0.06477916985750198, + 0.05934237316250801, + 0.032390449196100235, + 0.03680481016635895, + 0.06225040182471275, + 0.08847155421972275, + 0.037508971989154816 + ], + "credit_map": [ + 0.027440018951892853, + 0.020335447043180466, + 0.011988221667706966, + 0.013772807084023952, + 0.080317422747612, + 0.024125853553414345, + 0.027307577431201935, + 0.08184301853179932, + 0.012579282745718956, + 0.00957877654582262, + 0.008842773735523224, + 0.014483589679002762, + 0.03639701381325722, + 0.012399916537106037, + 0.023388192057609558, + 0.012526988983154297, + 0.01315500121563673, + 0.019885869696736336, + 0.009741558693349361, + 0.0373644195497036, + 0.045080337673425674, + 0.014691772870719433, + 0.04445192962884903, + 0.016754519194364548, + 0.0059857419691979885, + 0.008355529978871346, + 0.004597743973135948, + 0.0099993497133255, + 0.004686512518674135, + 0.004960792604833841, + 0.010439092293381691, + 0.01575441099703312, + 0.006012252531945705, + 0.015720224007964134, + 0.004557786975055933, + 0.004668788518756628, + 0.0073814173229038715, + 0.00799389835447073, + 0.007385545875877142, + 0.005622455850243568, + 0.0036790366284549236, + 0.006481146439909935, + 0.003069524886086583, + 0.003683659480884671, + 0.004222905728965998, + 0.00434653926640749, + 0.00310905440710485, + 0.003798585617914796, + 0.005827075801789761, + 0.003462978871539235, + 0.0051737562753260136, + 0.0043274154886603355, + 0.0032088947482407093, + 0.0036463818978518248, + 0.0039783925749361515, + 0.007179914508014917, + 0.006158469244837761, + 0.021476084366440773, + 0.0038112755864858627, + 0.005924558266997337, + 0.0061334543861448765, + 0.008837048895657063, + 0.005777077749371529, + 0.004132434725761414, + 0.009769158437848091, + 0.00918222963809967, + 0.006556573323905468, + 0.015075783245265484, + 0.010794127359986305, + 0.008660957217216492, + 0.019260365515947342, + 0.009172358550131321, + 0.006928892340511084, + 0.003976351115852594, + 0.004349106457084417, + 0.0024916213005781174, + 0.004823949653655291, + 0.0063687399961054325, + 0.004175292328000069, + 0.004395016934722662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.004143225960433483 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005690624238923192 + ], + "opposite_logprob_tokens": [ + -7.500568866729736 + ], + "logprob_reward": -0.0005690624238923192, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005690624238923192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3809691369533539, + 0.07294194400310516, + 0.0985075831413269, + 0.11323799937963486, + 0.05114920437335968, + 0.04587584361433983, + 0.04860293120145798, + 0.048881903290748596, + 0.052890386432409286, + 0.08694308996200562 + ], + "credit_map": [ + 0.044162727892398834, + 0.02023400366306305, + 0.014337120577692986, + 0.01457394752651453, + 0.14187400043010712, + 0.0314059853553772, + 0.028024420142173767, + 0.08635691553354263, + 0.00786326453089714, + 0.005266266874969006, + 0.005845468025654554, + 0.007713301572948694, + 0.01316310279071331, + 0.007373801898211241, + 0.012189838103950024, + 0.013526900671422482, + 0.010375197045505047, + 0.010646142065525055, + 0.004905536770820618, + 0.012217247858643532, + 0.017817454412579536, + 0.010761850513517857, + 0.02287306636571884, + 0.008911091834306717, + 0.008418007753789425, + 0.03086145967245102, + 0.006018721032887697, + 0.007047888357192278, + 0.013954156078398228, + 0.008411111310124397, + 0.009226026944816113, + 0.029300620779395103, + 0.005207080394029617, + 0.013756075873970985, + 0.004420223645865917, + 0.007072371430695057, + 0.005903040058910847, + 0.0054513923823833466, + 0.004545967560261488, + 0.004793053027242422, + 0.0039002373814582825, + 0.014578485861420631, + 0.0025602623354643583, + 0.00610755942761898, + 0.0036631692200899124, + 0.006996288429945707, + 0.0037950892001390457, + 0.00427475618198514, + 0.008476356975734234, + 0.007156849838793278, + 0.007256988435983658, + 0.004036556929349899, + 0.003685110015794635, + 0.005012317560613155, + 0.005405991803854704, + 0.007572758011519909, + 0.00546853244304657, + 0.014843467622995377, + 0.0032265980262309313, + 0.003954077139496803, + 0.00582701014354825, + 0.006931476294994354, + 0.004889241885393858, + 0.0037414992693811655, + 0.006254186853766441, + 0.005622513592243195, + 0.004052100703120232, + 0.00881523359566927, + 0.006780213676393032, + 0.005916794762015343, + 0.009978377260267735, + 0.0054709711112082005, + 0.013468930497765541, + 0.012938372790813446, + 0.008340849541127682, + 0.004734773654490709, + 0.012127207592129707, + 0.01629711128771305, + 0.008321528322994709, + 0.010714314877986908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0005690624238923192 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622915785759687 + ], + "opposite_logprob_tokens": [ + -4.882622718811035 + ], + "logprob_reward": -0.007622915785759687, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622915785759687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22135484218597412, + 0.3085898458957672, + 0.15845616161823273, + 0.1331540048122406, + 0.03242287039756775, + 0.020872879773378372, + 0.022857002913951874, + 0.0250480268150568, + 0.05889316275715828, + 0.018351150676608086 + ], + "credit_map": [ + 0.031099168583750725, + 0.03329470008611679, + 0.010573575273156166, + 0.014175406657159328, + 0.05442893132567406, + 0.01803840883076191, + 0.015192771330475807, + 0.04455189034342766, + 0.032248467206954956, + 0.02261294797062874, + 0.026697751134634018, + 0.03700676187872887, + 0.05188268423080444, + 0.03168359771370888, + 0.07360292226076126, + 0.032854724675416946, + 0.013736759312450886, + 0.02356462925672531, + 0.005310636945068836, + 0.006843040231615305, + 0.03837541863322258, + 0.017867039889097214, + 0.021170545369386673, + 0.031588103622198105, + 0.011424553580582142, + 0.03017904795706272, + 0.008696731179952621, + 0.007881876081228256, + 0.02979620359838009, + 0.010727212764322758, + 0.0101712541654706, + 0.024277128279209137, + 0.0031497045420110226, + 0.00849184300750494, + 0.003142878646031022, + 0.002808589255437255, + 0.004957037977874279, + 0.0038381945341825485, + 0.002780134789645672, + 0.003254491137340665, + 0.0020612541120499372, + 0.005259896162897348, + 0.0014770436100661755, + 0.00283720507286489, + 0.0017893135081976652, + 0.00300081935711205, + 0.002590084681287408, + 0.0018572642002254725, + 0.004777610767632723, + 0.002790408907458186, + 0.003245861269533634, + 0.0021743089891970158, + 0.001860265270806849, + 0.0023967460729181767, + 0.0022657655645161867, + 0.003346039215102792, + 0.0028552848380059004, + 0.006873789709061384, + 0.001558374147862196, + 0.004117905627936125, + 0.0019338884158059955, + 0.00330702168866992, + 0.002754217479377985, + 0.0016475478187203407, + 0.004671297036111355, + 0.007929553277790546, + 0.004110788460820913, + 0.0115859629586339, + 0.005879176314920187, + 0.006193134002387524, + 0.01273582223802805, + 0.005787428468465805, + 0.0031350445933640003, + 0.0027273332234472036, + 0.0018163560889661312, + 0.001595130073837936, + 0.001646892400458455, + 0.0019906149245798588, + 0.002241690643131733, + 0.003198090009391308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.007622915785759687 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811703622341156 + ], + "opposite_logprob_tokens": [ + -3.288116931915283 + ], + "logprob_reward": -0.03811703622341156, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811703622341156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23593564331531525, + 0.1418571174144745, + 0.2319304198026657, + 0.22780625522136688, + 0.03812573105096817, + 0.02122712880373001, + 0.024593878537416458, + 0.028505399823188782, + 0.03844650462269783, + 0.011571934446692467 + ], + "credit_map": [ + 0.02810407429933548, + 0.026230864226818085, + 0.008028892800211906, + 0.013614311814308167, + 0.05714065581560135, + 0.018402740359306335, + 0.021246200427412987, + 0.06316789984703064, + 0.01335048582404852, + 0.010731381364166737, + 0.012794706970453262, + 0.016914427280426025, + 0.019542813301086426, + 0.012616807594895363, + 0.03849431127309799, + 0.01741219125688076, + 0.019451923668384552, + 0.03836733475327492, + 0.012690544128417969, + 0.03850157931447029, + 0.05227933079004288, + 0.01719442568719387, + 0.034818559885025024, + 0.01862671598792076, + 0.019275406375527382, + 0.04114025831222534, + 0.00876249186694622, + 0.009455169551074505, + 0.08129408210515976, + 0.024539612233638763, + 0.027799217030405998, + 0.015540030784904957, + 0.0030555014964193106, + 0.0028079324401915073, + 0.0030257347971200943, + 0.003643974894657731, + 0.006817266810685396, + 0.0044561391696333885, + 0.009120950475335121, + 0.005198231898248196, + 0.0020896971691399813, + 0.0027661891654133797, + 0.002238085027784109, + 0.0031648457515984774, + 0.002481949981302023, + 0.0024221439380198717, + 0.0036059641279280186, + 0.002458255272358656, + 0.004226480610668659, + 0.0020084658171981573, + 0.003610437037423253, + 0.002479809569194913, + 0.002569075906649232, + 0.002384347841143608, + 0.0035124842543154955, + 0.0038027772679924965, + 0.002780413720756769, + 0.008878505788743496, + 0.0019610212184488773, + 0.0034349397756159306, + 0.002818210981786251, + 0.00432168785482645, + 0.002288373187184334, + 0.0020222479943186045, + 0.0047080437652766705, + 0.0052626165561378, + 0.0029763428028672934, + 0.006510969717055559, + 0.004889579489827156, + 0.004101139027625322, + 0.006535346154123545, + 0.0034624654799699783, + 0.0011709686368703842, + 0.0013992628082633018, + 0.0010038074105978012, + 0.001342995441518724, + 0.001464134082198143, + 0.0016603150870651007, + 0.0016576320631429553, + 0.001872819266282022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.03811703622341156 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001759111881256 + ], + "opposite_logprob_tokens": [ + -2.6950175762176514 + ], + "logprob_reward": -0.07001759111881256, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001759111881256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12639985978603363, + 0.08695952594280243, + 0.13263045251369476, + 0.09396833181381226, + 0.10687142610549927, + 0.14162616431713104, + 0.13347522914409637, + 0.06826303899288177, + 0.0750235840678215, + 0.03478235751390457 + ], + "credit_map": [ + 0.012233751825988293, + 0.010647847317159176, + 0.008509574458003044, + 0.008730857633054256, + 0.015903957188129425, + 0.014945176430046558, + 0.01759372651576996, + 0.037834975868463516, + 0.008058453910052776, + 0.005761167034506798, + 0.005371308885514736, + 0.00868882518261671, + 0.02683323249220848, + 0.007499490398913622, + 0.012360055930912495, + 0.012386998161673546, + 0.017251620069146156, + 0.015437759459018707, + 0.005647299811244011, + 0.00914101954549551, + 0.03949557989835739, + 0.017231786623597145, + 0.012516353279352188, + 0.015909049659967422, + 0.008445113897323608, + 0.009421849623322487, + 0.00400112010538578, + 0.006124426145106554, + 0.02376772090792656, + 0.01015570666640997, + 0.01886448636651039, + 0.013187921606004238, + 0.012772299349308014, + 0.007823986001312733, + 0.009775115177035332, + 0.007871837355196476, + 0.010629967786371708, + 0.011846126988530159, + 0.02125583216547966, + 0.024896269664168358, + 0.014934200793504715, + 0.038511183112859726, + 0.011730695143342018, + 0.012916292063891888, + 0.016248593106865883, + 0.019916806370019913, + 0.011705063283443451, + 0.015663329511880875, + 0.019170423969626427, + 0.016718508675694466, + 0.020535344257950783, + 0.015217374078929424, + 0.011887168511748314, + 0.013804050162434578, + 0.01853237673640251, + 0.017609978094697, + 0.006606631446629763, + 0.023680491372942924, + 0.004103184677660465, + 0.008416986092925072, + 0.005970856640487909, + 0.010688775219023228, + 0.004591684322804213, + 0.004204432480037212, + 0.007876802235841751, + 0.01467097271233797, + 0.005132228601723909, + 0.011509040370583534, + 0.00840412825345993, + 0.0113816624507308, + 0.01049635000526905, + 0.005552396643906832, + 0.006000479217618704, + 0.005058199167251587, + 0.003486005589365959, + 0.002420709701254964, + 0.004050055053085089, + 0.005653079133480787, + 0.003550740424543619, + 0.004563091788440943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.07001759111881256 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level wide-angle shot captures a stunning desert oasis, a living jewel amidst a vast, arid expanse. In the foreground, a pristine, glassy pool of water lies completely still, its surface a perfect mirror, reflecting the deep blue of the sky and the vibrant green of the surrounding palm trees. A cluster of tall, lush date palms, their fibrous brown trunks rising proudly, frame the water's edge, their dense, emerald-green fronds fanning out like intricate lace. Beyond this verdant haven, vast, rolling dunes of fine, golden-orange sand stretch endlessly to the distant, razor-sharp horizon, where the earth meets the brilliant, completely cloudless azure sky. The entire scene is bathed in the brilliant, unfiltered sunlight of midday, casting sharp, dark shadows beneath the palm trees. The only perceptible movement is the slow, gentle swaying of the palm fronds, which drift lazily in the breeze. This delicate dance contrasts with the absolute stillness of the water, creating a profound atmosphere of tranquility and serenity.", + "seed": 33392, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_7/index_8555_seed_33392_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.034988274762872605, + "vlm_logprob": -0.034988274762872605, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.034988274762872605, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01328603457659483 + ], + "opposite_logprob_tokens": [ + -4.763286113739014 + ], + "logprob_reward": -0.01328603457659483, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01328603457659483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3743187487125397, + 0.12777692079544067, + 0.14022326469421387, + 0.09586641192436218, + 0.02711998112499714, + 0.04282104969024658, + 0.0456664115190506, + 0.059995830059051514, + 0.049447204917669296, + 0.036764152348041534 + ], + "credit_map": [ + 0.047065384685993195, + 0.03388554975390434, + 0.03384257107973099, + 0.04730190709233284, + 0.04269114136695862, + 0.04017550125718117, + 0.06449668854475021, + 0.064860038459301, + 0.026483338326215744, + 0.014113377779722214, + 0.009730345569550991, + 0.012511654756963253, + 0.01631683111190796, + 0.013503961265087128, + 0.019709864631295204, + 0.015407565981149673, + 0.02054557017982006, + 0.015247388742864132, + 0.013126143254339695, + 0.012128029018640518, + 0.019970932975411415, + 0.014739176258444786, + 0.024448543787002563, + 0.020017500966787338, + 0.01649417169392109, + 0.010844704695045948, + 0.008920864202082157, + 0.008377796970307827, + 0.0120278000831604, + 0.010323172435164452, + 0.016440019011497498, + 0.01243788842111826, + 0.005109922494739294, + 0.002754479879513383, + 0.0025262441486120224, + 0.001996929757297039, + 0.003413824364542961, + 0.0032216745894402266, + 0.004782166797667742, + 0.0033147414214909077, + 0.005424711387604475, + 0.004096880089491606, + 0.004282417707145214, + 0.005837452132254839, + 0.005025938618928194, + 0.004341679625213146, + 0.006713124457746744, + 0.007098850794136524, + 0.011005867272615433, + 0.0060333809815347195, + 0.0057579344138503075, + 0.0048370882868766785, + 0.004370366223156452, + 0.0050821988843381405, + 0.00399568909779191, + 0.004583888221532106, + 0.010311990045011044, + 0.0063103376887738705, + 0.005645364057272673, + 0.005792947951704264, + 0.006085313390940428, + 0.007088970392942429, + 0.008077342994511127, + 0.010683568194508553, + 0.013299640268087387, + 0.007755576632916927, + 0.005351378582417965, + 0.003269523149356246, + 0.004845498129725456, + 0.005396247375756502, + 0.0046654934994876385, + 0.004863853100687265, + 0.008091334253549576, + 0.005793314427137375, + 0.004301270470023155, + 0.0030631544068455696, + 0.00464276410639286, + 0.004341274965554476, + 0.003178107552230358, + 0.0033529356587678194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.01328603457659483 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008235162240453064 + ], + "opposite_logprob_tokens": [ + -7.125823497772217 + ], + "logprob_reward": -0.0008235162240453064, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008235162240453064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20514562726020813, + 0.1446855515241623, + 0.18897782266139984, + 0.08574267476797104, + 0.1090870350599289, + 0.062074899673461914, + 0.03623202443122864, + 0.0640055388212204, + 0.042387235909700394, + 0.061661649495363235 + ], + "credit_map": [ + 0.026241162791848183, + 0.026708750054240227, + 0.01954224891960621, + 0.017211023718118668, + 0.02971956692636013, + 0.02443793974816799, + 0.031745798885822296, + 0.029539141803979874, + 0.03257685899734497, + 0.016717305406928062, + 0.01053787674754858, + 0.007758197840303183, + 0.01929025538265705, + 0.01623816415667534, + 0.020555837079882622, + 0.021011054515838623, + 0.025659412145614624, + 0.020731858909130096, + 0.017918290570378304, + 0.012004995718598366, + 0.030296333134174347, + 0.020080039277672768, + 0.03519546613097191, + 0.02709142677485943, + 0.013607004657387733, + 0.008981131948530674, + 0.009438272565603256, + 0.00944362673908472, + 0.009664920158684254, + 0.009285520762205124, + 0.01610153168439865, + 0.009220660664141178, + 0.014007755555212498, + 0.013133062981069088, + 0.01227929349988699, + 0.011401982977986336, + 0.015874672681093216, + 0.012115157209336758, + 0.016875745728611946, + 0.013399358838796616, + 0.008810040540993214, + 0.006952532101422548, + 0.00623686145991087, + 0.004951382055878639, + 0.007433837279677391, + 0.006070492789149284, + 0.011358758434653282, + 0.01026099268347025, + 0.0040932949632406235, + 0.003980977460741997, + 0.003983248956501484, + 0.0022046524100005627, + 0.005956271197646856, + 0.004387021996080875, + 0.005995339713990688, + 0.0056312186643481255, + 0.008430433459579945, + 0.007584509905427694, + 0.005554493982344866, + 0.005767871160060167, + 0.008034766651690006, + 0.00798239279538393, + 0.00938631035387516, + 0.011264757253229618, + 0.004166026134043932, + 0.005658825859427452, + 0.0039377398788928986, + 0.003177775302901864, + 0.008062286302447319, + 0.005144114606082439, + 0.00772364204749465, + 0.004516828805208206, + 0.009157311171293259, + 0.009301911108195782, + 0.005978958681225777, + 0.004865862429141998, + 0.01060230378061533, + 0.007957013323903084, + 0.007791477255523205, + 0.0060068112798035145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0008235162240453064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418777834624052 + ], + "opposite_logprob_tokens": [ + -4.264187812805176 + ], + "logprob_reward": -0.01418777834624052, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418777834624052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2551805078983307, + 0.0899886041879654, + 0.11623839288949966, + 0.07105864584445953, + 0.1515359729528427, + 0.06964123249053955, + 0.04360774904489517, + 0.15510404109954834, + 0.026269931346178055, + 0.021374888718128204 + ], + "credit_map": [ + 0.04269128292798996, + 0.034892771393060684, + 0.02021329291164875, + 0.020628387108445168, + 0.02878536470234394, + 0.027205415070056915, + 0.0399412177503109, + 0.04082274064421654, + 0.01296555157750845, + 0.010025531984865665, + 0.009183701127767563, + 0.007911575958132744, + 0.012879476882517338, + 0.008880081586539745, + 0.014860274270176888, + 0.013282411731779575, + 0.01272030733525753, + 0.013040597550570965, + 0.011308956891298294, + 0.008526893332600594, + 0.017134539783000946, + 0.012590601108968258, + 0.022631941363215446, + 0.01828456111252308, + 0.008236559107899666, + 0.008098498918116093, + 0.007204726338386536, + 0.005244679283350706, + 0.011263984255492687, + 0.007847323082387447, + 0.013169826939702034, + 0.009993040934205055, + 0.016325000673532486, + 0.01663462445139885, + 0.017546547576785088, + 0.012169920839369297, + 0.022678790614008904, + 0.01575705036520958, + 0.029221396893262863, + 0.02120264060795307, + 0.008156309835612774, + 0.00739253917708993, + 0.008486632257699966, + 0.006233140826225281, + 0.00979604385793209, + 0.007043248042464256, + 0.013483097776770592, + 0.009050223045051098, + 0.005434830207377672, + 0.0038036855403333902, + 0.004926628898829222, + 0.004789961967617273, + 0.005667201243340969, + 0.004776076413691044, + 0.007428433746099472, + 0.006780928932130337, + 0.013816410675644875, + 0.015190321020781994, + 0.016156096011400223, + 0.013147933408617973, + 0.019923215731978416, + 0.015755798667669296, + 0.03344089537858963, + 0.027673369273543358, + 0.00576845882460475, + 0.003621362615376711, + 0.0018754114862531424, + 0.001457005157135427, + 0.0036593673285096884, + 0.003275313414633274, + 0.0032520717941224575, + 0.0033609403762966394, + 0.0032515062484890223, + 0.002770795486867428, + 0.0013722797157242894, + 0.0015088273212313652, + 0.003948565572500229, + 0.0029125798027962446, + 0.0027229206170886755, + 0.002887415001168847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.01418777834624052 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371308371424675 + ], + "opposite_logprob_tokens": [ + -3.4087131023406982 + ], + "logprob_reward": -0.03371308371424675, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371308371424675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1867523491382599, + 0.0969436764717102, + 0.11986836045980453, + 0.09944713860750198, + 0.06460250914096832, + 0.11079854518175125, + 0.09137248992919922, + 0.1634034514427185, + 0.03563305363059044, + 0.03117838129401207 + ], + "credit_map": [ + 0.023411253467202187, + 0.023140419274568558, + 0.01617610827088356, + 0.01631401851773262, + 0.028720147907733917, + 0.022185856476426125, + 0.03229660913348198, + 0.024507932364940643, + 0.025451192632317543, + 0.014157180674374104, + 0.006093889009207487, + 0.0060053919441998005, + 0.011587806046009064, + 0.01151735894382, + 0.011788406409323215, + 0.010342451743781567, + 0.012376816011965275, + 0.012051914818584919, + 0.012016938999295235, + 0.00966691318899393, + 0.01908077485859394, + 0.012884839437901974, + 0.024700572714209557, + 0.017089588567614555, + 0.012027936987578869, + 0.010671268217265606, + 0.009130901657044888, + 0.008887175470590591, + 0.013544457964599133, + 0.01044947374612093, + 0.020394664257764816, + 0.014341256581246853, + 0.008297236636281013, + 0.006620810832828283, + 0.006925626657903194, + 0.008734800852835178, + 0.007817237637937069, + 0.006213835906237364, + 0.009730513207614422, + 0.01026244554668665, + 0.01365923136472702, + 0.01133109163492918, + 0.010207738727331161, + 0.01078358106315136, + 0.014074605889618397, + 0.011109218001365662, + 0.02179192565381527, + 0.01784115470945835, + 0.020289693027734756, + 0.010107027366757393, + 0.009174982085824013, + 0.008872889913618565, + 0.01009388454258442, + 0.009858318604528904, + 0.012250758707523346, + 0.010724936611950397, + 0.015126781538128853, + 0.018750453367829323, + 0.015541227534413338, + 0.013522871769964695, + 0.021479260176420212, + 0.02152932807803154, + 0.02539513260126114, + 0.03205839544534683, + 0.00920198205858469, + 0.005105368793010712, + 0.00291622057557106, + 0.001892021275125444, + 0.004144761245697737, + 0.004537145607173443, + 0.004389349836856127, + 0.003446203889325261, + 0.005734783597290516, + 0.0038713642861694098, + 0.002654359908774495, + 0.0022030966356396675, + 0.005240553058683872, + 0.004008388612419367, + 0.003509551053866744, + 0.003956284373998642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.03371308371424675 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a desert oasis featuring a glassy pool of water in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall date palms with fibrous brown trunks and emerald-green fronds visible framing the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the background show vast, rolling dunes of golden-orange sand stretching to a distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky completely cloudless and azure, with sharp, dark shadows cast by the palm trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only visible motion the slow, gentle swaying of the palm fronds while the water remains perfectly still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293096095323563 + ], + "opposite_logprob_tokens": [ + -2.2379310131073 + ], + "logprob_reward": -0.11293096095323563, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293096095323563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17497926950454712, + 0.13764619827270508, + 0.09421216696500778, + 0.18509036302566528, + 0.12935395538806915, + 0.1004902571439743, + 0.0678727924823761, + 0.039322301745414734, + 0.044465914368629456, + 0.026566727086901665 + ], + "credit_map": [ + 0.026873912662267685, + 0.022223154082894325, + 0.01578873209655285, + 0.011955643072724342, + 0.020095016807317734, + 0.018923526629805565, + 0.03302581608295441, + 0.026093482971191406, + 0.03686663135886192, + 0.01964683085680008, + 0.012853182852268219, + 0.008936379104852676, + 0.013956787995994091, + 0.016105707734823227, + 0.01485403347760439, + 0.014426643960177898, + 0.010117280296981335, + 0.011198767460882664, + 0.00948235671967268, + 0.007511634845286608, + 0.014455393888056278, + 0.010751883499324322, + 0.017786076292395592, + 0.012908783741295338, + 0.02252569980919361, + 0.01894989237189293, + 0.018578000366687775, + 0.019956817850470543, + 0.022823821753263474, + 0.01907547190785408, + 0.03671964257955551, + 0.026461025699973106, + 0.017072206363081932, + 0.011840153485536575, + 0.011779190972447395, + 0.0184173583984375, + 0.015492863953113556, + 0.013421577401459217, + 0.01959526538848877, + 0.02173534221947193, + 0.00962920393794775, + 0.012605179101228714, + 0.009826463647186756, + 0.00969360675662756, + 0.015362112782895565, + 0.010970708914101124, + 0.015717603266239166, + 0.016685383394360542, + 0.011580170132219791, + 0.00760938273742795, + 0.007326570805162191, + 0.005448284093290567, + 0.008109992370009422, + 0.0070838285610079765, + 0.012860006652772427, + 0.007854563184082508, + 0.006726894993335009, + 0.004662553779780865, + 0.0033507728949189186, + 0.0031602366361767054, + 0.004571341909468174, + 0.0043432158417999744, + 0.006854247767478228, + 0.005653040949255228, + 0.00399765046313405, + 0.004379764664918184, + 0.005060219205915928, + 0.004088596440851688, + 0.006809013895690441, + 0.004628702066838741, + 0.00904067512601614, + 0.0064612943679094315, + 0.004917027894407511, + 0.0024212405551224947, + 0.002243038499727845, + 0.002341953571885824, + 0.0037882993929088116, + 0.0027408068999648094, + 0.004145944025367498, + 0.003968419507145882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.11293096095323563 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_8/index_tensor([5294, 5294])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_8/index_tensor([5294, 5294])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ff081bd872403aedec8cb3a2c8a74ea969728cfb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_8/index_tensor([5294, 5294])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Through a single, static, eye-level medium shot, rendered in Studio Ghibli's signature hand-drawn style, this time-lapse captures the miraculous growth of several potted plants on a warm, sun-drenched windowsill. The camera remains stationary, observing as tiny, vibrant green sprouts delicately pierce the surface of the rich, dark soil within weathered terracotta pots. As time accelerates, these sprouts gracefully unfold, transforming into lush, verdant leaves, their stems reaching upward toward the light. Golden sunlight filters through the clean glass of the white wooden windows, casting dynamic shadows and illuminating the magical dance of dust particles in the air. As the plants mature, glistening water droplets, remnants of recent watering, glitter like tiny gems on the leaves. In the final moments of the shot, a tiny, bright red ladybug with a distinct black spot cautiously crawls across a newly opened, broad green leaf, its antennae twitching curiously, adding a touch of gentle life to the serene and wondrous spectacle of nature's quiet power.", + "seed": 32400, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_8/index_5294_seed_32400_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.16427248967811464, + "vlm_logprob": -0.16427248967811464, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.16427248967811464, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6361713409423828 + ], + "opposite_logprob_tokens": [ + -0.7611713409423828 + ], + "logprob_reward": -0.6361713409423828, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6361713409423828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37632009387016296, + 0.12263426929712296, + 0.05435846373438835, + 0.06875208020210266, + 0.2359284609556198, + 0.025511326268315315, + 0.012818505987524986, + 0.047190748155117035, + 0.030479945242404938, + 0.026006169617176056 + ], + "credit_map": [ + 0.060877230018377304, + 0.03836604580283165, + 0.022433066740632057, + 0.03537049889564514, + 0.07102285325527191, + 0.08970034867525101, + 0.027090895920991898, + 0.03145912289619446, + 0.01812426745891571, + 0.01743934117257595, + 0.004493623971939087, + 0.01372128538787365, + 0.04102630913257599, + 0.016751375049352646, + 0.0058709681034088135, + 0.0052070920355618, + 0.008298017084598541, + 0.006734820548444986, + 0.003445978043600917, + 0.006169603206217289, + 0.007247308734804392, + 0.013302463106811047, + 0.0036032365169376135, + 0.005557031370699406, + 0.013102083466947079, + 0.006045920308679342, + 0.003796013304963708, + 0.009336595423519611, + 0.01167477760463953, + 0.013937047682702541, + 0.0062490166164934635, + 0.004610614851117134, + 0.031546950340270996, + 0.05165385454893112, + 0.007663677912205458, + 0.012923534959554672, + 0.08053812384605408, + 0.02829894982278347, + 0.01245369203388691, + 0.010849657468497753, + 0.002257926855236292, + 0.007053642999380827, + 0.0019396244315430522, + 0.003628468606621027, + 0.0035921253729611635, + 0.003613043809309602, + 0.00190505082719028, + 0.0015214422019198537, + 0.0013597332872450352, + 0.0025326337199658155, + 0.0008899766835384071, + 0.0014828740386292338, + 0.00141771393828094, + 0.002905344357714057, + 0.0009660520008765161, + 0.0012641777284443378, + 0.003964552190154791, + 0.010971184819936752, + 0.0022440943866968155, + 0.0066428473219275475, + 0.012252709828317165, + 0.00541791133582592, + 0.0035292906686663628, + 0.0021681510843336582, + 0.006910886149853468, + 0.003922595176845789, + 0.002125588245689869, + 0.004500017501413822, + 0.004129047971218824, + 0.003577328519895673, + 0.0028830908704549074, + 0.0024313866160809994, + 0.004180802032351494, + 0.003578110132366419, + 0.0021914164535701275, + 0.00369873084127903, + 0.0032168885227292776, + 0.003223368199542165, + 0.0032107445877045393, + 0.0027061065193265676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.6361713409423828 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020617574453353882 + ], + "opposite_logprob_tokens": [ + -3.8956174850463867 + ], + "logprob_reward": -0.020617574453353882, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020617574453353882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3709750473499298, + 0.12178341299295425, + 0.0683872178196907, + 0.0793146938085556, + 0.10232491046190262, + 0.04584527388215065, + 0.029063379392027855, + 0.0908420979976654, + 0.055649854242801666, + 0.03581413999199867 + ], + "credit_map": [ + 0.06682256609201431, + 0.04234535992145538, + 0.01954469084739685, + 0.034258052706718445, + 0.0862397849559784, + 0.0654529258608818, + 0.023875847458839417, + 0.03243577852845192, + 0.011853734962642193, + 0.013645919039845467, + 0.009341197088360786, + 0.016853174194693565, + 0.031584035605192184, + 0.012473063543438911, + 0.011136378161609173, + 0.01489589735865593, + 0.008543981239199638, + 0.009341276250779629, + 0.003538998309522867, + 0.010072801262140274, + 0.01340407319366932, + 0.012845159508287907, + 0.004643120802938938, + 0.005997799802571535, + 0.008386331610381603, + 0.010582450777292252, + 0.007122487295418978, + 0.01486955862492323, + 0.01131867989897728, + 0.011031636036932468, + 0.008667533285915852, + 0.0073360055685043335, + 0.008952430449426174, + 0.024921823292970657, + 0.004216266330331564, + 0.011433113366365433, + 0.03100087121129036, + 0.01139170303940773, + 0.004838160704821348, + 0.005570536945015192, + 0.004223705269396305, + 0.009267031215131283, + 0.0036079506389796734, + 0.005623239558190107, + 0.005643434356898069, + 0.009956005029380322, + 0.003555324859917164, + 0.003968578763306141, + 0.0035570450127124786, + 0.005488983355462551, + 0.0022630225867033005, + 0.003760588588193059, + 0.004540815483778715, + 0.00418873643502593, + 0.002655870746821165, + 0.0026083155535161495, + 0.009849990718066692, + 0.00908251665532589, + 0.009066787548363209, + 0.012674124911427498, + 0.012156922370195389, + 0.015710635110735893, + 0.011122972704470158, + 0.011178137734532356, + 0.010232596658170223, + 0.009239128790795803, + 0.00503444904461503, + 0.0056693581864237785, + 0.007281270809471607, + 0.005489996634423733, + 0.006091554649174213, + 0.006611492019146681, + 0.005794309079647064, + 0.004033316858112812, + 0.002670514164492488, + 0.005695014260709286, + 0.006284643895924091, + 0.0035361885093152523, + 0.004275206476449966, + 0.003524942323565483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.020617574453353882 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022077015601098537 + ], + "opposite_logprob_tokens": [ + -6.1272077560424805 + ], + "logprob_reward": -0.0022077015601098537, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022077015601098537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4095959961414337, + 0.11607273668050766, + 0.09892117977142334, + 0.06746906042098999, + 0.0946505069732666, + 0.06275391578674316, + 0.015349975787103176, + 0.073848195374012, + 0.029874125495553017, + 0.03146432340145111 + ], + "credit_map": [ + 0.06841132044792175, + 0.047501638531684875, + 0.01646740548312664, + 0.026946397498250008, + 0.1034293919801712, + 0.09924409538507462, + 0.021403713151812553, + 0.026192011311650276, + 0.015482169575989246, + 0.018766405060887337, + 0.007315456867218018, + 0.01523028127849102, + 0.030236108228564262, + 0.01436117384582758, + 0.007120615337044001, + 0.00756050692871213, + 0.009966673329472542, + 0.01669788546860218, + 0.00821295753121376, + 0.01452108845114708, + 0.019821474328637123, + 0.013564497232437134, + 0.00768347829580307, + 0.008453111164271832, + 0.012367858551442623, + 0.011000028811395168, + 0.0044202134013175964, + 0.009723073802888393, + 0.009217162616550922, + 0.010745425708591938, + 0.00619426928460598, + 0.0038010214921087027, + 0.014118624851107597, + 0.017983365803956985, + 0.005182580556720495, + 0.01245473325252533, + 0.018019970506429672, + 0.012993237935006618, + 0.007916375994682312, + 0.005981610622256994, + 0.006749213207513094, + 0.014525983482599258, + 0.003755171550437808, + 0.009794342331588268, + 0.008264299482107162, + 0.012177103199064732, + 0.004012275952845812, + 0.0034755240194499493, + 0.0017986965831369162, + 0.003763139247894287, + 0.0010545124532654881, + 0.0024473299272358418, + 0.0021564695052802563, + 0.0015361255500465631, + 0.001266227220185101, + 0.001327473553828895, + 0.007863890379667282, + 0.013468658551573753, + 0.0057835569605231285, + 0.009608727879822254, + 0.010144870728254318, + 0.013320148922502995, + 0.006631532683968544, + 0.007026795297861099, + 0.004736911505460739, + 0.006214874796569347, + 0.0021507253404706717, + 0.004773584194481373, + 0.0033026845194399357, + 0.004648517817258835, + 0.0023241303861141205, + 0.0017226955387741327, + 0.0074048833921551704, + 0.0037522087804973125, + 0.002283918671309948, + 0.004949748981744051, + 0.0039021624252200127, + 0.004043145105242729, + 0.0027554163243621588, + 0.0023728369269520044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0022077015601098537 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031463205814362 + ], + "opposite_logprob_tokens": [ + -2.3503146171569824 + ], + "logprob_reward": -0.10031463205814362, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031463205814362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2662471532821655, + 0.039706919342279434, + 0.1285083293914795, + 0.06135745346546173, + 0.07549172639846802, + 0.032620079815387726, + 0.04494649916887283, + 0.1791822761297226, + 0.06388559192419052, + 0.1080540120601654 + ], + "credit_map": [ + 0.052241191267967224, + 0.024276556447148323, + 0.012951833195984364, + 0.021607190370559692, + 0.052312176674604416, + 0.069798544049263, + 0.01464637741446495, + 0.01841326802968979, + 0.005323613062500954, + 0.0037934936117380857, + 0.0030018130782991648, + 0.006597206927835941, + 0.006854770239442587, + 0.005650092847645283, + 0.003442217130213976, + 0.005043707322329283, + 0.010669504292309284, + 0.022483017295598984, + 0.008918010629713535, + 0.023672500625252724, + 0.023759592324495316, + 0.015691310167312622, + 0.011848758906126022, + 0.01146563421934843, + 0.009817828424274921, + 0.005806895904242992, + 0.004281706176698208, + 0.008137202821671963, + 0.01150533277541399, + 0.011760074645280838, + 0.004590933211147785, + 0.005457477644085884, + 0.005299903452396393, + 0.020933950319886208, + 0.003804340958595276, + 0.006286334246397018, + 0.021901946514844894, + 0.007589614950120449, + 0.00488266721367836, + 0.004792958032339811, + 0.0035895940382033587, + 0.008831499144434929, + 0.001841099001467228, + 0.005119082052260637, + 0.004044524393975735, + 0.00432779872789979, + 0.002326702931895852, + 0.0025397799909114838, + 0.004796413704752922, + 0.011592401191592216, + 0.002800764050334692, + 0.0061968909576535225, + 0.007670608814805746, + 0.005174826364964247, + 0.0038274123799055815, + 0.0028871793765574694, + 0.015003494918346405, + 0.03805755823850632, + 0.01238387543708086, + 0.030061781406402588, + 0.03314858302474022, + 0.0233243890106678, + 0.018032826483249664, + 0.009169748984277248, + 0.008120100013911724, + 0.01126865204423666, + 0.004901690874248743, + 0.010752067901194096, + 0.007644448429346085, + 0.007208410184830427, + 0.006598061881959438, + 0.007392155006527901, + 0.021625585854053497, + 0.019837575033307076, + 0.008714210242033005, + 0.01626584492623806, + 0.012637055478990078, + 0.011250205338001251, + 0.008624709211289883, + 0.009098817594349384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.10031463205814362 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620511993765831 + ], + "opposite_logprob_tokens": [ + -2.812051296234131 + ], + "logprob_reward": -0.0620511993765831, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620511993765831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15409883856773376, + 0.24751269817352295, + 0.10351231694221497, + 0.16198953986167908, + 0.07983745634555817, + 0.02733125165104866, + 0.024386612698435783, + 0.11524905264377594, + 0.03851659595966339, + 0.04756563529372215 + ], + "credit_map": [ + 0.02328077331185341, + 0.019606731832027435, + 0.009816950187087059, + 0.02070782706141472, + 0.026601402088999748, + 0.03151315078139305, + 0.009972365573048592, + 0.012599634937942028, + 0.03827250376343727, + 0.03412714973092079, + 0.01038403995335102, + 0.031348731368780136, + 0.0737345740199089, + 0.035686854273080826, + 0.01226203516125679, + 0.01169680617749691, + 0.01551146525889635, + 0.015031917952001095, + 0.00655336631461978, + 0.009481561370193958, + 0.014333619736135006, + 0.026139678433537483, + 0.005403552670031786, + 0.011057154275476933, + 0.032634198665618896, + 0.0341997966170311, + 0.0073388200253248215, + 0.010889002121984959, + 0.0181354321539402, + 0.03661927208304405, + 0.01045356784015894, + 0.011719453148543835, + 0.0101031344383955, + 0.018070843070745468, + 0.0026607364416122437, + 0.005793623626232147, + 0.027107587084174156, + 0.008327290415763855, + 0.004514198284596205, + 0.0032600397244095802, + 0.0026849647983908653, + 0.004191091749817133, + 0.0017478294903412461, + 0.0034138287883251905, + 0.004216802306473255, + 0.00701924366876483, + 0.0018259231001138687, + 0.00223156763240695, + 0.002631702460348606, + 0.005397176370024681, + 0.0017020937521010637, + 0.003279599826782942, + 0.003646459896117449, + 0.0025510385166853666, + 0.0023620608262717724, + 0.0028164798859506845, + 0.008969223126769066, + 0.015165499411523342, + 0.008711915463209152, + 0.015241187065839767, + 0.02028658054769039, + 0.026223527267575264, + 0.010468997992575169, + 0.010182123631238937, + 0.007338834926486015, + 0.006430471781641245, + 0.0028627682477235794, + 0.005766893271356821, + 0.0038169105537235737, + 0.0042732455767691135, + 0.0029276441782712936, + 0.005099826958030462, + 0.0057441601529717445, + 0.004901002626866102, + 0.004244210664182901, + 0.008373633027076721, + 0.005937485955655575, + 0.006382135208696127, + 0.0055372645147144794, + 0.006445741280913353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.0620511993765831 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Through a single, static, eye-level medium shot, rendered in Studio Ghibli's signature hand-drawn style, this time-lapse captures the miraculous growth of several potted plants on a warm, sun-drenched windowsill. The camera remains stationary, observing as tiny, vibrant green sprouts delicately pierce the surface of the rich, dark soil within weathered terracotta pots. As time accelerates, these sprouts gracefully unfold, transforming into lush, verdant leaves, their stems reaching upward toward the light. Golden sunlight filters through the clean glass of the white wooden windows, casting dynamic shadows and illuminating the magical dance of dust particles in the air. As the plants mature, glistening water droplets, remnants of recent watering, glitter like tiny gems on the leaves. In the final moments of the shot, a tiny, bright red ladybug with a distinct black spot cautiously crawls across a newly opened, broad green leaf, its antennae twitching curiously, adding a touch of gentle life to the serene and wondrous spectacle of nature's quiet power.", + "seed": 32401, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_8/index_5294_seed_32401_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.26236665872856973, + "vlm_logprob": -0.26236665872856973, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.26236665872856973, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1413586139678955 + ], + "opposite_logprob_tokens": [ + -0.3913585841655731 + ], + "logprob_reward": -1.1413586139678955, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1413586139678955, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2553630471229553, + 0.061671722680330276, + 0.013050679117441177, + 0.02108803205192089, + 0.01413502637296915, + 0.047644227743148804, + 0.06605470180511475, + 0.08635076135396957, + 0.3303011655807495, + 0.10434062033891678 + ], + "credit_map": [ + 0.02073596604168415, + 0.08693353831768036, + 0.011956597678363323, + 0.08272578567266464, + 0.013742570765316486, + 0.010140781290829182, + 0.015025763772428036, + 0.0141020892187953, + 0.0070160445757210255, + 0.023720357567071915, + 0.002085997024551034, + 0.0062612262554466724, + 0.009741615504026413, + 0.003289566608145833, + 0.0044594574719667435, + 0.005097466986626387, + 0.0015975177520886064, + 0.00208286102861166, + 0.000996218528598547, + 0.0017756196903064847, + 0.002721434226259589, + 0.001131760305725038, + 0.0012770897010341287, + 0.0014681797474622726, + 0.002812824910506606, + 0.0027031663339585066, + 0.001815934432670474, + 0.003656357992440462, + 0.002649847185239196, + 0.0013242372078821063, + 0.004021842963993549, + 0.002103823469951749, + 0.00141250086016953, + 0.0029379127081483603, + 0.0014136697864159942, + 0.00154009903781116, + 0.0026329061947762966, + 0.0011614960385486484, + 0.0015019734855741262, + 0.0015344693092629313, + 0.0033761137165129185, + 0.008749512024223804, + 0.0033281075302511454, + 0.014737666584551334, + 0.005845853127539158, + 0.0030376617796719074, + 0.004146696534007788, + 0.004422618541866541, + 0.00620924262329936, + 0.017711080610752106, + 0.004205486737191677, + 0.01197455171495676, + 0.007789545692503452, + 0.005916004069149494, + 0.004408434499055147, + 0.007840363308787346, + 0.0101563585922122, + 0.022240398451685905, + 0.006707221735268831, + 0.010237245820462704, + 0.01315640565007925, + 0.008482638746500015, + 0.006274664308875799, + 0.009095828980207443, + 0.02546446956694126, + 0.03455817699432373, + 0.01413660403341055, + 0.09283839166164398, + 0.09545505791902542, + 0.0239412784576416, + 0.02411835640668869, + 0.01978885941207409, + 0.014433603733778, + 0.015120442025363445, + 0.00912987720221281, + 0.014806890860199928, + 0.0156653244048357, + 0.011040857993066311, + 0.01148203294724226, + 0.012661603279411793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -1.1413586139678955 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04312163218855858 + ], + "opposite_logprob_tokens": [ + -3.168121576309204 + ], + "logprob_reward": -0.04312163218855858, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04312163218855858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12145649641752243, + 0.322354257106781, + 0.0352010615170002, + 0.03408990427851677, + 0.01703689619898796, + 0.025062205269932747, + 0.03770548850297928, + 0.060512762516736984, + 0.2023886740207672, + 0.14419229328632355 + ], + "credit_map": [ + 0.007205752655863762, + 0.01960074156522751, + 0.008336134254932404, + 0.047416992485523224, + 0.011072400026023388, + 0.006947489455342293, + 0.010348406620323658, + 0.01052857469767332, + 0.023223552852869034, + 0.04186014458537102, + 0.02171771600842476, + 0.13015629351139069, + 0.03211262449622154, + 0.019475366920232773, + 0.029722873121500015, + 0.02408568188548088, + 0.0035124055575579405, + 0.005253275390714407, + 0.002844745060428977, + 0.006102518178522587, + 0.007491393014788628, + 0.0032591279596090317, + 0.0032390274573117495, + 0.0034985686652362347, + 0.0038509685546159744, + 0.0029903848189860582, + 0.002283466048538685, + 0.00400205934420228, + 0.006805474404245615, + 0.003373296931385994, + 0.006927961483597755, + 0.003856290364637971, + 0.0014959329273551702, + 0.00378088909201324, + 0.001747174421325326, + 0.0016552780289202929, + 0.0028192445170134306, + 0.0014331556158140302, + 0.001865981612354517, + 0.0022392398677766323, + 0.00292784976772964, + 0.005364617332816124, + 0.0018995353020727634, + 0.0035655167885124683, + 0.004786712117493153, + 0.0020270238164812326, + 0.002231015358120203, + 0.0022599336225539446, + 0.003556725336238742, + 0.007009352557361126, + 0.0029144601430743933, + 0.009358826093375683, + 0.004212262574583292, + 0.0032132957130670547, + 0.003029460785910487, + 0.004411101341247559, + 0.007759917061775923, + 0.008486658334732056, + 0.005271161440759897, + 0.00802689976990223, + 0.010591275058686733, + 0.006902854423969984, + 0.006637484300881624, + 0.006836507469415665, + 0.016093270853161812, + 0.019998973235487938, + 0.009776582010090351, + 0.05463477224111557, + 0.05813772603869438, + 0.014497273601591587, + 0.014865816570818424, + 0.01438424363732338, + 0.016998611390590668, + 0.02710474282503128, + 0.01016231533139944, + 0.02183002233505249, + 0.019681710749864578, + 0.017071111127734184, + 0.012249038554728031, + 0.019094714894890785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.04312163218855858 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052745467983186245 + ], + "opposite_logprob_tokens": [ + -5.255274772644043 + ], + "logprob_reward": -0.0052745467983186245, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052745467983186245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07466059923171997, + 0.08587928861379623, + 0.03974480554461479, + 0.03932061791419983, + 0.026397960260510445, + 0.032530028373003006, + 0.07928085327148438, + 0.055579666048288345, + 0.4417600929737091, + 0.12484610825777054 + ], + "credit_map": [ + 0.006211108062416315, + 0.012647312134504318, + 0.004654471762478352, + 0.010633217170834541, + 0.006604523863643408, + 0.014265409670770168, + 0.010237347334623337, + 0.009407203644514084, + 0.007356276735663414, + 0.030691873282194138, + 0.0032145751174539328, + 0.023609595373272896, + 0.006353212520480156, + 0.004558023530989885, + 0.0049868556670844555, + 0.00510887848213315, + 0.003826332278549671, + 0.0056054252199828625, + 0.002499945927411318, + 0.0040258304215967655, + 0.009736389853060246, + 0.004997398238629103, + 0.004124644678086042, + 0.004928841255605221, + 0.00425329152494669, + 0.008371593430638313, + 0.0020454037003219128, + 0.005254640709608793, + 0.008138729259371758, + 0.0037076971493661404, + 0.0037235587369650602, + 0.0038257045671343803, + 0.002441979479044676, + 0.0048941620625555515, + 0.0017880011582747102, + 0.003840017132461071, + 0.005329806357622147, + 0.0022212350741028786, + 0.002901917789131403, + 0.002980840625241399, + 0.003391527570784092, + 0.006115806754678488, + 0.002230654703453183, + 0.007381142117083073, + 0.004940543789416552, + 0.002700090641155839, + 0.0028137722983956337, + 0.0029564888682216406, + 0.008592244237661362, + 0.01583860255777836, + 0.006619942374527454, + 0.0171984676271677, + 0.007385760545730591, + 0.008978808298707008, + 0.0062852720730006695, + 0.00838176067918539, + 0.005481432657688856, + 0.013451156206429005, + 0.005112344864755869, + 0.006671931128948927, + 0.009106851182878017, + 0.006136405281722546, + 0.004652021452784538, + 0.004967518616467714, + 0.03243155777454376, + 0.04224158078432083, + 0.01862393319606781, + 0.1309138387441635, + 0.13030071556568146, + 0.031939394772052765, + 0.032324522733688354, + 0.02298457734286785, + 0.01716388575732708, + 0.02027541771531105, + 0.010517229326069355, + 0.01829945296049118, + 0.01776459440588951, + 0.012716402299702168, + 0.01324036531150341, + 0.014868761412799358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.0052745467983186245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899352163076401 + ], + "opposite_logprob_tokens": [ + -2.578993558883667 + ], + "logprob_reward": -0.07899352163076401, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899352163076401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11837401986122131, + 0.11091656237840652, + 0.032849859446287155, + 0.07839342951774597, + 0.016220713034272194, + 0.0621720589697361, + 0.05077564716339111, + 0.07252802699804306, + 0.11257141083478928, + 0.34519827365875244 + ], + "credit_map": [ + 0.020974814891815186, + 0.01865282468497753, + 0.006613518111407757, + 0.02341839112341404, + 0.013741578906774521, + 0.009809144772589207, + 0.013330241665244102, + 0.01183351594954729, + 0.01117816660553217, + 0.0168241448700428, + 0.006573912687599659, + 0.027778366580605507, + 0.016573801636695862, + 0.012982925400137901, + 0.009985039941966534, + 0.009020199067890644, + 0.0032556019723415375, + 0.006145009305328131, + 0.002442811382934451, + 0.007241160608828068, + 0.005547448992729187, + 0.002793704392388463, + 0.0026392359286546707, + 0.0027848861645907164, + 0.005894447676837444, + 0.0056389993987977505, + 0.002469449769705534, + 0.022289320826530457, + 0.005981664638966322, + 0.005363943986594677, + 0.024986080825328827, + 0.00576952425763011, + 0.0013396169524639845, + 0.0035767070949077606, + 0.0012640421045944095, + 0.0013822225155308843, + 0.003772503463551402, + 0.0013047584798187017, + 0.0017574455123394728, + 0.0018234173767268658, + 0.0034389118663966656, + 0.017538322135806084, + 0.0029807486571371555, + 0.017273280769586563, + 0.008593342266976833, + 0.0033052980434149504, + 0.004598085302859545, + 0.004444068297743797, + 0.004534691572189331, + 0.012360679917037487, + 0.0037248260341584682, + 0.0058264899998903275, + 0.009638091549277306, + 0.005725076887756586, + 0.004016418009996414, + 0.004949373193085194, + 0.008883995935320854, + 0.020735088735818863, + 0.005527855828404427, + 0.009735235944390297, + 0.009474638849496841, + 0.006176792550832033, + 0.005874927621334791, + 0.006119485013186932, + 0.008764628320932388, + 0.013392181135714054, + 0.006920492276549339, + 0.02754082903265953, + 0.031518809497356415, + 0.010331783443689346, + 0.006723674014210701, + 0.007379014510661364, + 0.03682325407862663, + 0.07292214781045914, + 0.022396625950932503, + 0.04581080749630928, + 0.04761025309562683, + 0.056751035153865814, + 0.029248058795928955, + 0.03363610804080963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.07899352163076401 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308497905731201 + ], + "opposite_logprob_tokens": [ + -3.1680850982666016 + ], + "logprob_reward": -0.04308497905731201, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308497905731201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1215428039431572, + 0.14093148708343506, + 0.05670201778411865, + 0.0939338356256485, + 0.08716033399105072, + 0.06555073708295822, + 0.05672147125005722, + 0.08643365651369095, + 0.08612740784883499, + 0.20489619672298431 + ], + "credit_map": [ + 0.015951208770275116, + 0.013894946314394474, + 0.009930332191288471, + 0.014215007424354553, + 0.019829118624329567, + 0.011582406237721443, + 0.016318362206220627, + 0.01982143148779869, + 0.01274084486067295, + 0.03159726411104202, + 0.006628331262618303, + 0.015359466895461082, + 0.03464578837156296, + 0.012225953862071037, + 0.013100961223244667, + 0.014632896520197392, + 0.005265009589493275, + 0.006870171055197716, + 0.004660585429519415, + 0.015124457888305187, + 0.007449612021446228, + 0.006169630214571953, + 0.005110993515700102, + 0.0060515692457556725, + 0.013350889086723328, + 0.008998343721032143, + 0.006049342919141054, + 0.019197145476937294, + 0.015196701511740685, + 0.007454212289303541, + 0.01210404559969902, + 0.011583161540329456, + 0.008507311344146729, + 0.013907617889344692, + 0.007425257004797459, + 0.012269340455532074, + 0.016043368726968765, + 0.009283801540732384, + 0.01177175436168909, + 0.007951893843710423, + 0.005942592862993479, + 0.015430067665874958, + 0.004972635302692652, + 0.012089412659406662, + 0.008671055547893047, + 0.0077864909544587135, + 0.005037239752709866, + 0.005621252115815878, + 0.006783211603760719, + 0.012698417529463768, + 0.004402010701596737, + 0.008066775277256966, + 0.006628385744988918, + 0.008148078806698322, + 0.004069915506988764, + 0.005924678407609463, + 0.011351252906024456, + 0.0140071427449584, + 0.009131993167102337, + 0.012003520503640175, + 0.011340625584125519, + 0.007363468408584595, + 0.009017924778163433, + 0.012217734940350056, + 0.009833645075559616, + 0.015737785026431084, + 0.007161368150264025, + 0.013425175100564957, + 0.011288771405816078, + 0.008674952201545238, + 0.007767726667225361, + 0.012237993068993092, + 0.0316033735871315, + 0.03549548238515854, + 0.014916525222361088, + 0.02883378602564335, + 0.02683841995894909, + 0.024258287623524666, + 0.017563672736287117, + 0.02538667619228363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.04308497905731201 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Through a single, static, eye-level medium shot, rendered in Studio Ghibli's signature hand-drawn style, this time-lapse captures the miraculous growth of several potted plants on a warm, sun-drenched windowsill. The camera remains stationary, observing as tiny, vibrant green sprouts delicately pierce the surface of the rich, dark soil within weathered terracotta pots. As time accelerates, these sprouts gracefully unfold, transforming into lush, verdant leaves, their stems reaching upward toward the light. Golden sunlight filters through the clean glass of the white wooden windows, casting dynamic shadows and illuminating the magical dance of dust particles in the air. As the plants mature, glistening water droplets, remnants of recent watering, glitter like tiny gems on the leaves. In the final moments of the shot, a tiny, bright red ladybug with a distinct black spot cautiously crawls across a newly opened, broad green leaf, its antennae twitching curiously, adding a touch of gentle life to the serene and wondrous spectacle of nature's quiet power.", + "seed": 32402, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_8/index_5294_seed_32402_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.24491455773822962, + "vlm_logprob": -0.24491455773822962, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.24491455773822962, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9777864813804626 + ], + "opposite_logprob_tokens": [ + -0.47778648138046265 + ], + "logprob_reward": -0.9777864813804626, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9777864813804626, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21492904424667358, + 0.12312344461679459, + 0.03286996856331825, + 0.07687177509069443, + 0.044658295810222626, + 0.03737049549818039, + 0.13079865276813507, + 0.07830505818128586, + 0.1318458616733551, + 0.12922737002372742 + ], + "credit_map": [ + 0.041308969259262085, + 0.05469319596886635, + 0.02035362459719181, + 0.01813425123691559, + 0.020328069105744362, + 0.02246156521141529, + 0.018626084551215172, + 0.019023261964321136, + 0.016988184303045273, + 0.013542468659579754, + 0.007887876592576504, + 0.02878074161708355, + 0.007785026915371418, + 0.008799776434898376, + 0.028847569599747658, + 0.010491794906556606, + 0.004441097844392061, + 0.005493143107742071, + 0.0036935838870704174, + 0.0064278957433998585, + 0.0027209518011659384, + 0.0025639294181019068, + 0.004392401780933142, + 0.003136963117867708, + 0.010391131043434143, + 0.0067016431130468845, + 0.011449807323515415, + 0.010739929042756557, + 0.006843763869255781, + 0.011562315747141838, + 0.010111338458955288, + 0.00907184649258852, + 0.007877347059547901, + 0.0036677359603345394, + 0.0036134521942585707, + 0.004669227171689272, + 0.004477406851947308, + 0.011055768467485905, + 0.0055030654184520245, + 0.003794291988015175, + 0.004345830995589495, + 0.004640560131520033, + 0.003510871436446905, + 0.005323893390595913, + 0.004997239913791418, + 0.006364448461681604, + 0.003743827808648348, + 0.004443823359906673, + 0.015941403806209564, + 0.014947823248803616, + 0.014677059836685658, + 0.016103599220514297, + 0.0190220195800066, + 0.012995883822441101, + 0.01870805397629738, + 0.018402807414531708, + 0.01080346666276455, + 0.016031943261623383, + 0.007279799319803715, + 0.012659846805036068, + 0.007002534344792366, + 0.010696658864617348, + 0.0073630171827971935, + 0.006467786151915789, + 0.01833595708012581, + 0.016322685405611992, + 0.023112500086426735, + 0.013838885352015495, + 0.0147571861743927, + 0.01874782145023346, + 0.010569289326667786, + 0.01616153120994568, + 0.017177697271108627, + 0.019576601684093475, + 0.022027073428034782, + 0.011395817622542381, + 0.018545685335993767, + 0.01753273978829384, + 0.012618658132851124, + 0.010353096760809422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.9777864813804626 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028608606662601233 + ], + "opposite_logprob_tokens": [ + -5.877861022949219 + ], + "logprob_reward": -0.0028608606662601233, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028608606662601233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6855944991111755, + 0.04117449000477791, + 0.030407920479774475, + 0.029712865129113197, + 0.01841586083173752, + 0.018425535410642624, + 0.03104369528591633, + 0.02733944170176983, + 0.03586191684007645, + 0.08202376961708069 + ], + "credit_map": [ + 0.20685243606567383, + 0.23740778863430023, + 0.046557798981666565, + 0.031657624989748, + 0.07634381949901581, + 0.03185161575675011, + 0.03335994482040405, + 0.021563464775681496, + 0.006164475344121456, + 0.004503112751990557, + 0.0031182996463030577, + 0.00955069437623024, + 0.004446486476808786, + 0.0032196161337196827, + 0.006253316067159176, + 0.003918488975614309, + 0.0036739197093993425, + 0.003486328525468707, + 0.0019034150755032897, + 0.007459179498255253, + 0.002804104471579194, + 0.0022713958751410246, + 0.006163457408547401, + 0.0026461186353117228, + 0.0040209125727415085, + 0.0024692362640053034, + 0.0027106781490147114, + 0.005253827665001154, + 0.002463639946654439, + 0.005083403084427118, + 0.004775557667016983, + 0.002935609081760049, + 0.003872121451422572, + 0.001215794705785811, + 0.0010804595658555627, + 0.0016398605657741427, + 0.0015527858631685376, + 0.0059356363490223885, + 0.0020130868069827557, + 0.001106114941649139, + 0.002078988589346409, + 0.002599441446363926, + 0.0017768117832019925, + 0.002171499654650688, + 0.0029053939506411552, + 0.0028858256991952658, + 0.0022561559453606606, + 0.0017514178762212396, + 0.0038080441299825907, + 0.003582711098715663, + 0.0030343837570399046, + 0.005479564890265465, + 0.00329594686627388, + 0.0029951666947454214, + 0.0040510534308850765, + 0.004796824883669615, + 0.004307070281356573, + 0.005446204915642738, + 0.0019159221556037664, + 0.004460119642317295, + 0.0024429750628769398, + 0.0034746446181088686, + 0.002733314409852028, + 0.0025591892190277576, + 0.006091570481657982, + 0.006532483268529177, + 0.004042129963636398, + 0.0038462618831545115, + 0.004545256961137056, + 0.004240487702190876, + 0.0028227930888533592, + 0.0037409323267638683, + 0.011344022117555141, + 0.010422359220683575, + 0.00828672107309103, + 0.007344511337578297, + 0.013916048221290112, + 0.013983300887048244, + 0.008006945252418518, + 0.008719855919480324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0028608606662601233 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012560192495584488 + ], + "opposite_logprob_tokens": [ + -4.3875603675842285 + ], + "logprob_reward": -0.012560192495584488, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012560192495584488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4076210856437683, + 0.07957138866186142, + 0.06640715152025223, + 0.03821387141942978, + 0.12166345119476318, + 0.037770532071590424, + 0.026745542883872986, + 0.06023075804114342, + 0.09478060156106949, + 0.06699559092521667 + ], + "credit_map": [ + 0.1135583221912384, + 0.13141736388206482, + 0.030751243233680725, + 0.01994927041232586, + 0.048944782465696335, + 0.024536022916436195, + 0.023633666336536407, + 0.014830442145466805, + 0.012049905955791473, + 0.009438945911824703, + 0.005837373901158571, + 0.016955258324742317, + 0.008555869571864605, + 0.007220769766718149, + 0.012498084455728531, + 0.007015187293291092, + 0.010330352932214737, + 0.009760677814483643, + 0.004398247692734003, + 0.01263410970568657, + 0.007355427369475365, + 0.005049118306487799, + 0.009978042915463448, + 0.0069011785089969635, + 0.005996217019855976, + 0.005722973495721817, + 0.004090708214789629, + 0.005882893223315477, + 0.002925524488091469, + 0.0040965755470097065, + 0.006216900423169136, + 0.0032820836640894413, + 0.028780583292245865, + 0.007350328378379345, + 0.00602192897349596, + 0.009137594141066074, + 0.008384719491004944, + 0.04347250238060951, + 0.011358166113495827, + 0.007157641928642988, + 0.003736813087016344, + 0.00444052554666996, + 0.003348542610183358, + 0.004733957350254059, + 0.005003081168979406, + 0.0065963417291641235, + 0.00452694995328784, + 0.005384327843785286, + 0.003269332693889737, + 0.004325751680880785, + 0.002104984363541007, + 0.004553081933408976, + 0.003861908335238695, + 0.0025628022849559784, + 0.0030540607403963804, + 0.0030136227142065763, + 0.007856961339712143, + 0.011091119609773159, + 0.006541647017002106, + 0.008205368183553219, + 0.008692274801433086, + 0.007877462543547153, + 0.005588024388998747, + 0.004377908539026976, + 0.01682683639228344, + 0.014624150469899178, + 0.010302659124135971, + 0.009631093591451645, + 0.014593766070902348, + 0.010879657231271267, + 0.00826107058674097, + 0.009661376476287842, + 0.009229371324181557, + 0.010263973847031593, + 0.00848663505166769, + 0.005140727385878563, + 0.009070695377886295, + 0.008856487460434437, + 0.0070871347561478615, + 0.008860579691827297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.012560192495584488 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156438648700714 + ], + "opposite_logprob_tokens": [ + -1.7015644311904907 + ], + "logprob_reward": -0.20156438648700714, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156438648700714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5791498422622681, + 0.11170979589223862, + 0.023810381069779396, + 0.04374929144978523, + 0.035851284861564636, + 0.033764977008104324, + 0.03078693337738514, + 0.013711066916584969, + 0.06593415886163712, + 0.06153227388858795 + ], + "credit_map": [ + 0.17308270931243896, + 0.1991182118654251, + 0.04267540201544762, + 0.023500969633460045, + 0.06822329759597778, + 0.02752627618610859, + 0.0277037862688303, + 0.01731915771961212, + 0.01589151844382286, + 0.016780145466327667, + 0.010350670665502548, + 0.018515560775995255, + 0.01038600504398346, + 0.00881140772253275, + 0.020643770694732666, + 0.010330716148018837, + 0.0033505139872431755, + 0.004091199021786451, + 0.001456407830119133, + 0.003757316619157791, + 0.002476451685652137, + 0.002617757534608245, + 0.003353533800691366, + 0.002707201521843672, + 0.005871263332664967, + 0.004438140895217657, + 0.0039867255836725235, + 0.008265706710517406, + 0.003745411057025194, + 0.0061784242279827595, + 0.0068443771451711655, + 0.004419244360178709, + 0.00509137287735939, + 0.004326629918068647, + 0.003016616916283965, + 0.0052747116424143314, + 0.002678110497072339, + 0.005065021105110645, + 0.00573571166023612, + 0.004663108382374048, + 0.003594518406316638, + 0.006308612879365683, + 0.003599198069423437, + 0.004720896016806364, + 0.003699185326695442, + 0.003866649931296706, + 0.004068195819854736, + 0.003907719627022743, + 0.0037845636252313852, + 0.003956094849854708, + 0.004317380487918854, + 0.00397754181176424, + 0.004455784801393747, + 0.003262483049184084, + 0.0035653384402394295, + 0.0034677472431212664, + 0.0019678298849612474, + 0.0026023683603852987, + 0.0016769750509411097, + 0.002016422338783741, + 0.0013618792872875929, + 0.0012340362882241607, + 0.0013559162616729736, + 0.001495639095082879, + 0.00981738232076168, + 0.009828724898397923, + 0.011183347553014755, + 0.007548270747065544, + 0.007447643671184778, + 0.008394891396164894, + 0.0048381672240793705, + 0.006875728722661734, + 0.011362383142113686, + 0.00797035451978445, + 0.01043024007230997, + 0.006543882191181183, + 0.006210328545421362, + 0.0076479725539684296, + 0.006205159705132246, + 0.005161958280950785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.20156438648700714 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029800867661833763 + ], + "opposite_logprob_tokens": [ + -3.5298008918762207 + ], + "logprob_reward": -0.029800867661833763, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029800867661833763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2678821384906769, + 0.07247991114854813, + 0.13173247873783112, + 0.08686221390962601, + 0.128270223736763, + 0.027761317789554596, + 0.04450857266783714, + 0.059720076620578766, + 0.11035842448472977, + 0.07042470574378967 + ], + "credit_map": [ + 0.07641569525003433, + 0.08703599125146866, + 0.01648491621017456, + 0.013340191915631294, + 0.028599480167031288, + 0.016342952847480774, + 0.017568474635481834, + 0.012094439007341862, + 0.009265192784368992, + 0.010826749727129936, + 0.005202832631766796, + 0.01531622651964426, + 0.007925627753138542, + 0.005641251802444458, + 0.012120283208787441, + 0.006181745324283838, + 0.018610848113894463, + 0.015290354378521442, + 0.008110316470265388, + 0.032372549176216125, + 0.00899757444858551, + 0.008942456915974617, + 0.02709304355084896, + 0.01231533195823431, + 0.012458277866244316, + 0.006990955676883459, + 0.013726144097745419, + 0.008413758128881454, + 0.010588706471025944, + 0.013310592621564865, + 0.010136950761079788, + 0.011236829683184624, + 0.030663490295410156, + 0.009065011516213417, + 0.005099241156131029, + 0.008602666668593884, + 0.008881763555109501, + 0.04634715989232063, + 0.012591910548508167, + 0.00701897731050849, + 0.0032244452740997076, + 0.003476860234513879, + 0.002654769690707326, + 0.004390740767121315, + 0.0026907455176115036, + 0.004306448623538017, + 0.004118762910366058, + 0.002898547099903226, + 0.005716957617551088, + 0.005266239400953054, + 0.004536802880465984, + 0.006308072712272406, + 0.005708401557058096, + 0.004061196930706501, + 0.005444933660328388, + 0.007465969305485487, + 0.008510464802384377, + 0.011984451673924923, + 0.006191960070282221, + 0.008317488245666027, + 0.005204183515161276, + 0.008321421220898628, + 0.0055029550567269325, + 0.005687152501195669, + 0.01893709786236286, + 0.01861833781003952, + 0.014728077687323093, + 0.008749132975935936, + 0.013513266108930111, + 0.015633083879947662, + 0.009981898590922356, + 0.010197524912655354, + 0.011212107725441456, + 0.01015771459788084, + 0.009077147580683231, + 0.007571694441139698, + 0.00911516509950161, + 0.01013627927750349, + 0.006304322741925716, + 0.00685027614235878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.029800867661833763 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Through a single, static, eye-level medium shot, rendered in Studio Ghibli's signature hand-drawn style, this time-lapse captures the miraculous growth of several potted plants on a warm, sun-drenched windowsill. The camera remains stationary, observing as tiny, vibrant green sprouts delicately pierce the surface of the rich, dark soil within weathered terracotta pots. As time accelerates, these sprouts gracefully unfold, transforming into lush, verdant leaves, their stems reaching upward toward the light. Golden sunlight filters through the clean glass of the white wooden windows, casting dynamic shadows and illuminating the magical dance of dust particles in the air. As the plants mature, glistening water droplets, remnants of recent watering, glitter like tiny gems on the leaves. In the final moments of the shot, a tiny, bright red ladybug with a distinct black spot cautiously crawls across a newly opened, broad green leaf, its antennae twitching curiously, adding a touch of gentle life to the serene and wondrous spectacle of nature's quiet power.", + "seed": 32403, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_8/index_5294_seed_32403_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2234612605534494, + "vlm_logprob": -0.2234612605534494, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.2234612605534494, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8312841653823853 + ], + "opposite_logprob_tokens": [ + -0.5812841653823853 + ], + "logprob_reward": -0.8312841653823853, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8312841653823853, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.053036246448755264, + 0.1744336038827896, + 0.035039544105529785, + 0.23203209042549133, + 0.09079297631978989, + 0.0964025929570198, + 0.17939017713069916, + 0.09724405407905579, + 0.022326787933707237, + 0.01930188573896885 + ], + "credit_map": [ + 0.0061427634209394455, + 0.005717565305531025, + 0.00416808295994997, + 0.011614887975156307, + 0.008214742876589298, + 0.004739715252071619, + 0.008169424720108509, + 0.004269066266715527, + 0.053716130554676056, + 0.014989802613854408, + 0.005076853558421135, + 0.008791454136371613, + 0.0693296417593956, + 0.007637364324182272, + 0.009263991378247738, + 0.00562837952747941, + 0.007200801745057106, + 0.003983958624303341, + 0.0025932553689926863, + 0.00537665281444788, + 0.005422178190201521, + 0.002374091651290655, + 0.004718729294836521, + 0.003369879676029086, + 0.05943502485752106, + 0.02804298885166645, + 0.0064923446625471115, + 0.00722104124724865, + 0.09891147911548615, + 0.010216865688562393, + 0.010832577012479305, + 0.010879795998334885, + 0.01650615595281124, + 0.010444048792123795, + 0.0046773930080235004, + 0.008392971940338612, + 0.0323876217007637, + 0.0050945221446454525, + 0.00893297977745533, + 0.004357292316854, + 0.009955331683158875, + 0.016151640564203262, + 0.008975238539278507, + 0.028464948758482933, + 0.012946750968694687, + 0.0066670519299805164, + 0.008200345560908318, + 0.005041291005909443, + 0.02064843289554119, + 0.0237477608025074, + 0.015525844879448414, + 0.028666436672210693, + 0.034943170845508575, + 0.013574165292084217, + 0.024388378486037254, + 0.017895996570587158, + 0.010832194238901138, + 0.016191979870200157, + 0.008515521883964539, + 0.012392447330057621, + 0.013241564854979515, + 0.01210565585643053, + 0.013234060257673264, + 0.01073063351213932, + 0.003187907626852393, + 0.003910775762051344, + 0.0017183327581733465, + 0.0024019472766667604, + 0.004117179196327925, + 0.0019308202899992466, + 0.0024214310105890036, + 0.002638396341353655, + 0.002275877632200718, + 0.003382055088877678, + 0.0013045758241787553, + 0.0026596342213451862, + 0.00354817695915699, + 0.002061477629467845, + 0.001997709972783923, + 0.0020723792258650064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.8312841653823853 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04310736060142517 + ], + "opposite_logprob_tokens": [ + -3.168107271194458 + ], + "logprob_reward": -0.04310736060142517, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04310736060142517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0884503573179245, + 0.41749340295791626, + 0.026860646903514862, + 0.16424499452114105, + 0.08892101794481277, + 0.039326246827840805, + 0.036288078874349594, + 0.03717777505517006, + 0.05023651570081711, + 0.05100097134709358 + ], + "credit_map": [ + 0.018598465248942375, + 0.010955157689750195, + 0.005632766056805849, + 0.014898264780640602, + 0.012235574424266815, + 0.009113229811191559, + 0.00938696600496769, + 0.007629940286278725, + 0.11260986328125, + 0.03236629441380501, + 0.016877133399248123, + 0.037960994988679886, + 0.14367851614952087, + 0.02091607078909874, + 0.02842213399708271, + 0.02466239221394062, + 0.004052972421050072, + 0.003011907683685422, + 0.002204196760430932, + 0.004678066819906235, + 0.004729500971734524, + 0.0021587428636848927, + 0.0030900314450263977, + 0.0029352274723351, + 0.04227227345108986, + 0.018838083371520042, + 0.005011405795812607, + 0.005270888097584248, + 0.06930242478847504, + 0.00774625176563859, + 0.007977421395480633, + 0.007826232351362705, + 0.01689688116312027, + 0.009206682443618774, + 0.004964957945048809, + 0.00671741459518671, + 0.028173387050628662, + 0.005332890897989273, + 0.006331000942736864, + 0.011297808960080147, + 0.00464236456900835, + 0.005234353244304657, + 0.0036017633974552155, + 0.010558760724961758, + 0.00524767953902483, + 0.00276001519523561, + 0.003895918605849147, + 0.003385391319170594, + 0.004276736173778772, + 0.0041765752248466015, + 0.00302688917145133, + 0.005385877098888159, + 0.006234598346054554, + 0.0024158006999641657, + 0.006320539861917496, + 0.004451061598956585, + 0.00471937470138073, + 0.006395464763045311, + 0.0026322808116674423, + 0.005890564527362585, + 0.005592600908130407, + 0.0033162108156830072, + 0.004515929613262415, + 0.004115347284823656, + 0.00514819985255599, + 0.00843814667314291, + 0.0035483010578900576, + 0.009544489905238152, + 0.009325819090008736, + 0.0036032560747116804, + 0.006628150120377541, + 0.0040001519955694675, + 0.006008564960211515, + 0.008183509111404419, + 0.004640189930796623, + 0.006455177906900644, + 0.008965167216956615, + 0.004902506247162819, + 0.005808629561215639, + 0.006037228275090456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.04310736060142517 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007649652194231749 + ], + "opposite_logprob_tokens": [ + -4.8826494216918945 + ], + "logprob_reward": -0.007649652194231749, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007649652194231749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11047059297561646, + 0.21658888459205627, + 0.06294836848974228, + 0.23660025000572205, + 0.0464884452521801, + 0.04357108101248741, + 0.1354467123746872, + 0.07852902263402939, + 0.046270888298749924, + 0.02308576926589012 + ], + "credit_map": [ + 0.011523179709911346, + 0.00917914416640997, + 0.009106315672397614, + 0.025794701650738716, + 0.01754741743206978, + 0.007721079979091883, + 0.02285647578537464, + 0.006742272060364485, + 0.049439936876297, + 0.014678188599646091, + 0.012732511386275291, + 0.0369606576859951, + 0.054107796400785446, + 0.013771535828709602, + 0.022693181410431862, + 0.012205062434077263, + 0.008260712958872318, + 0.006894038990139961, + 0.0044517251662909985, + 0.010710074566304684, + 0.011667696759104729, + 0.00585751561447978, + 0.00882582925260067, + 0.006280775647610426, + 0.05916723236441612, + 0.027948686853051186, + 0.007030140142887831, + 0.007596976589411497, + 0.10075869411230087, + 0.01078890822827816, + 0.011809363029897213, + 0.011500228196382523, + 0.005002522375434637, + 0.007658533751964569, + 0.004312634468078613, + 0.006627944298088551, + 0.005333447363227606, + 0.0031826605554670095, + 0.007511896081268787, + 0.006858800072222948, + 0.006232365500181913, + 0.003721039043739438, + 0.003645900869742036, + 0.0049782101996243, + 0.005596731789410114, + 0.004517380613833666, + 0.004628299735486507, + 0.010251150466501713, + 0.015664253383874893, + 0.016509439796209335, + 0.013568544760346413, + 0.023695021867752075, + 0.02817131020128727, + 0.009020666591823101, + 0.017598874866962433, + 0.011218588799238205, + 0.009893521666526794, + 0.013635468669235706, + 0.006638761609792709, + 0.01058235950767994, + 0.011171319521963596, + 0.008300240151584148, + 0.011184412986040115, + 0.007122935261577368, + 0.005948949139565229, + 0.006674972362816334, + 0.003531249240040779, + 0.004562624264508486, + 0.01124308630824089, + 0.00463296752423048, + 0.004886749666184187, + 0.004790285136550665, + 0.0022169628646224737, + 0.002684216247871518, + 0.001339946873486042, + 0.004064737819135189, + 0.0052048563957214355, + 0.0025695401709526777, + 0.00313186994753778, + 0.0018736377824097872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 6 + ], + "selected_reward": -0.007649652194231749 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156264305114746 + ], + "opposite_logprob_tokens": [ + -1.7015626430511475 + ], + "logprob_reward": -0.20156264305114746, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156264305114746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18015335500240326, + 0.07935091108083725, + 0.05190471187233925, + 0.12627598643302917, + 0.0736137256026268, + 0.1036924496293068, + 0.15226125717163086, + 0.1036250963807106, + 0.03227926790714264, + 0.09684327989816666 + ], + "credit_map": [ + 0.01417427510023117, + 0.01544838584959507, + 0.014473030343651772, + 0.041484035551548004, + 0.02833208627998829, + 0.013433770276606083, + 0.03872131556272507, + 0.014086461625993252, + 0.018498927354812622, + 0.007039916235953569, + 0.0037609331775456667, + 0.011149321682751179, + 0.020396435633301735, + 0.003750462317839265, + 0.008870118297636509, + 0.005884801968932152, + 0.007467356976121664, + 0.0051511493511497974, + 0.004329436458647251, + 0.00889802910387516, + 0.009861896745860577, + 0.0037164133973419666, + 0.006103818770498037, + 0.006376613862812519, + 0.03248682990670204, + 0.014713958837091923, + 0.003915808629244566, + 0.004034705925732851, + 0.053361281752586365, + 0.005951095372438431, + 0.005898298695683479, + 0.005914009176194668, + 0.01116932462900877, + 0.008284068666398525, + 0.0048720501363277435, + 0.006410080473870039, + 0.02389838732779026, + 0.0045320759527385235, + 0.008942115120589733, + 0.005505619570612907, + 0.012603182345628738, + 0.010647992603480816, + 0.010476251132786274, + 0.025337420403957367, + 0.014909517019987106, + 0.007500096224248409, + 0.015032310038805008, + 0.007185671478509903, + 0.017483772709965706, + 0.020536024123430252, + 0.01316699106246233, + 0.02656654641032219, + 0.02006366103887558, + 0.011736886575818062, + 0.02837696485221386, + 0.014330401085317135, + 0.019865384325385094, + 0.01539683062583208, + 0.007785726338624954, + 0.01185249350965023, + 0.013425685465335846, + 0.012456494383513927, + 0.013213977217674255, + 0.00962850358337164, + 0.003965216223150492, + 0.00512299919500947, + 0.0025384423788636923, + 0.005404005292803049, + 0.004394486080855131, + 0.0029948102310299873, + 0.004009779077023268, + 0.003849528729915619, + 0.008358174934983253, + 0.012287978082895279, + 0.00806141272187233, + 0.018063630908727646, + 0.01582595892250538, + 0.009633098728954792, + 0.01618819683790207, + 0.008424829691648483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.20156264305114746 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Studio Ghibli hand-drawn animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene show potted plants growing on a sun-drenched windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the plants grow from tiny green sprouts into lush leaves with glistening water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tiny bright red ladybug with a black spot crawling on a green leaf in the final moments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, eye-level, and medium throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370248153805733 + ], + "opposite_logprob_tokens": [ + -3.4087023735046387 + ], + "logprob_reward": -0.03370248153805733, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370248153805733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10317006707191467, + 0.10518831759691238, + 0.04025163874030113, + 0.16518811881542206, + 0.1102706715464592, + 0.11804947257041931, + 0.14689897000789642, + 0.12529684603214264, + 0.0524408295750618, + 0.0332450233399868 + ], + "credit_map": [ + 0.012747258879244328, + 0.012112876400351524, + 0.007608507759869099, + 0.023855436593294144, + 0.01309393160045147, + 0.006640130653977394, + 0.018079442903399467, + 0.0090324766933918, + 0.009878071025013924, + 0.009921005927026272, + 0.00800326094031334, + 0.024586990475654602, + 0.016329754143953323, + 0.007675275206565857, + 0.015628257766366005, + 0.013165699318051338, + 0.0055530923418700695, + 0.004456022288650274, + 0.003223544219508767, + 0.006458032876253128, + 0.0056455135345458984, + 0.005363797768950462, + 0.0037326416932046413, + 0.005818995181471109, + 0.043914541602134705, + 0.017194082960486412, + 0.00544336624443531, + 0.008544792421162128, + 0.06624495983123779, + 0.007519194390624762, + 0.007774463389068842, + 0.008552725426852703, + 0.021021967753767967, + 0.01033395528793335, + 0.004323760513216257, + 0.007071508560329676, + 0.04368601366877556, + 0.0060752625577151775, + 0.009624434635043144, + 0.008133777417242527, + 0.013617154210805893, + 0.015735357999801636, + 0.00959533080458641, + 0.024451585486531258, + 0.01484694518148899, + 0.009674601256847382, + 0.010930865071713924, + 0.019197629764676094, + 0.014463376253843307, + 0.017006585374474525, + 0.014957808889448643, + 0.018924234434962273, + 0.03616061806678772, + 0.01046566478908062, + 0.018908701837062836, + 0.016011983156204224, + 0.017976969480514526, + 0.01777486689388752, + 0.009656264446675777, + 0.015936367213726044, + 0.01939358375966549, + 0.012793242931365967, + 0.019720526412129402, + 0.012045026756823063, + 0.005011469125747681, + 0.008543194271624088, + 0.0034016803838312626, + 0.011201990768313408, + 0.009648837149143219, + 0.003334545297548175, + 0.0074540493078529835, + 0.0038450625725090504, + 0.004320594482123852, + 0.007323449477553368, + 0.002131807152181864, + 0.004293103236705065, + 0.0065121958032250404, + 0.0027593630366027355, + 0.002857727464288473, + 0.0030467843171209097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.03370248153805733 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_9/index_tensor([6540, 6540])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_9/index_tensor([6540, 6540])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..29cc2cfb19016825d4796384514e1199688185cc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_9/index_tensor([6540, 6540])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "lingering, static wide shot plunges the viewer into the heart of a chaotic shipyard, ripped apart by a fierce storm. The vast skeletal hulls of unfinished vessels, their raw steel frames a web of rust-orange and deep shadows, twisted into jagged, anxious shapes clawing at a wounded, roiling sky. The entire scene is rendered with expressionistic, emotive brushstrokes and distorted perspective, with thick, diagonal flakes of rain streaking across the frame like frantic scratches on canvas. Erratic, blinding lightning momentarily paints the world a stark, electric white, casting long, twisting shadows that wriggle and contort with a life of their own. The ground is a roiling sea of mud and dark, reflective puddles, reflecting the sickly greens and deep indigos of the clouds above. Every line is sharp and unsettling, every shape feels unstable, creating a powerful, claustrophobic atmosphere, as if the landscape itself is screaming in silent terror.", + "seed": 13908, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_9/index_6540_seed_13908_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04766296213492751, + "vlm_logprob": -0.04766296213492751, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.04766296213492751, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16294334828853607 + ], + "opposite_logprob_tokens": [ + -1.9129433631896973 + ], + "logprob_reward": -0.16294334828853607, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16294334828853607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08458292484283447, + 0.041773948818445206, + 0.027193883433938026, + 0.2769750654697418, + 0.30531802773475647, + 0.01614314317703247, + 0.015415778383612633, + 0.09579148143529892, + 0.04632321745157242, + 0.09048260003328323 + ], + "credit_map": [ + 0.006351180374622345, + 0.005412714555859566, + 0.016584526747465134, + 0.015159988775849342, + 0.006960253696888685, + 0.008871938101947308, + 0.013442931696772575, + 0.01179939042776823, + 0.006614670157432556, + 0.003344952827319503, + 0.0062631540931761265, + 0.004622566048055887, + 0.0038489908911287785, + 0.003592811059206724, + 0.007954585365951061, + 0.0055322167463600636, + 0.0026068626902997494, + 0.0026125977747142315, + 0.0026861289516091347, + 0.0041067590937018394, + 0.0038700022269040346, + 0.003127735573798418, + 0.004506363067775965, + 0.0036774345207959414, + 0.014163554646074772, + 0.08260582387447357, + 0.018115032464265823, + 0.02859719842672348, + 0.02723272517323494, + 0.014785351231694221, + 0.02989201247692108, + 0.06158336624503136, + 0.016549767926335335, + 0.044393591582775116, + 0.1147063672542572, + 0.016517549753189087, + 0.013275020755827427, + 0.012025509960949421, + 0.019776713103055954, + 0.06807353347539902, + 0.0016518537886440754, + 0.002053075935691595, + 0.0019697898533195257, + 0.0019367113709449768, + 0.0024568578228354454, + 0.0023479689843952656, + 0.00172014394775033, + 0.0020067417062819004, + 0.0017814033199101686, + 0.002688518026843667, + 0.0018809433095157146, + 0.0018731398740783334, + 0.0022238087840378284, + 0.0020499101374298334, + 0.001602521282620728, + 0.0013155334163457155, + 0.006823028903454542, + 0.0168902687728405, + 0.013923131860792637, + 0.008287954144179821, + 0.025140173733234406, + 0.009946634992957115, + 0.008673246949911118, + 0.006107036489993334, + 0.004873640835285187, + 0.0063788434490561485, + 0.0070974561385810375, + 0.006520553957670927, + 0.006887344177812338, + 0.005291914101690054, + 0.003870126325637102, + 0.00540333753451705, + 0.0050813439302146435, + 0.031005268916487694, + 0.023326365277171135, + 0.005833594594150782, + 0.010447365231812, + 0.0071136122569441795, + 0.0025169854052364826, + 0.005158068612217903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.16294334828853607 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018203342333436012 + ], + "opposite_logprob_tokens": [ + -4.018203258514404 + ], + "logprob_reward": -0.018203342333436012, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018203342333436012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10111183673143387, + 0.14010606706142426, + 0.05257290601730347, + 0.080454021692276, + 0.10576039552688599, + 0.015162522904574871, + 0.03867356479167938, + 0.0462980791926384, + 0.10191389173269272, + 0.31794673204421997 + ], + "credit_map": [ + 0.008748060092329979, + 0.011159772984683514, + 0.010688194073736668, + 0.013529697433114052, + 0.017491625621914864, + 0.008023371919989586, + 0.018557842820882797, + 0.012913260608911514, + 0.01769259199500084, + 0.015563498251140118, + 0.019808366894721985, + 0.01590646244585514, + 0.016669725999236107, + 0.009759464301168919, + 0.02179633639752865, + 0.022909602150321007, + 0.005293943453580141, + 0.003939472138881683, + 0.006454738322645426, + 0.009689821861684322, + 0.005597739480435848, + 0.00365380197763443, + 0.009964266791939735, + 0.007979116402566433, + 0.008029029704630375, + 0.009513584896922112, + 0.012162133120000362, + 0.014531317166984081, + 0.008393575437366962, + 0.006122916471213102, + 0.01245070993900299, + 0.009250749833881855, + 0.006829578895121813, + 0.009942250326275826, + 0.03629468381404877, + 0.0069069950841367245, + 0.007582553196698427, + 0.005575845949351788, + 0.0066586388275027275, + 0.025969842448830605, + 0.0014590114587917924, + 0.0017206505872309208, + 0.002067787339910865, + 0.002531250938773155, + 0.001838961965404451, + 0.0018874414963647723, + 0.0020304799545556307, + 0.0016269388142973185, + 0.0035858789924532175, + 0.005944087635725737, + 0.004930643830448389, + 0.005702740978449583, + 0.004634459502995014, + 0.006583711132407188, + 0.0037757845129817724, + 0.003516254248097539, + 0.0039012404158711433, + 0.004928032401949167, + 0.005991337820887566, + 0.009007206186652184, + 0.006995197851210833, + 0.006832699757069349, + 0.0053849671967327595, + 0.0032573931384831667, + 0.009954353794455528, + 0.008961358107626438, + 0.018437176942825317, + 0.021051466464996338, + 0.011502886191010475, + 0.010888826102018356, + 0.008454159833490849, + 0.012663657777011395, + 0.040262266993522644, + 0.06559155136346817, + 0.08329315483570099, + 0.04550246521830559, + 0.017705485224723816, + 0.027661969885230064, + 0.017632663249969482, + 0.02029714547097683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.018203342333436012 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096213262528181 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004096213262528181, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096213262528181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13699784874916077, + 0.1020621508359909, + 0.03897232189774513, + 0.1259223222732544, + 0.12173688411712646, + 0.020292166620492935, + 0.033835988491773605, + 0.1381460428237915, + 0.1671820878982544, + 0.1148521900177002 + ], + "credit_map": [ + 0.010925274342298508, + 0.012180785648524761, + 0.014807203784584999, + 0.017004920169711113, + 0.020046453922986984, + 0.014388572424650192, + 0.02933458238840103, + 0.018310047686100006, + 0.0101653803139925, + 0.01225016638636589, + 0.011191447265446186, + 0.012576193548738956, + 0.0129478070884943, + 0.008808701299130917, + 0.021510997787117958, + 0.012611465528607368, + 0.003378911642357707, + 0.005018694791942835, + 0.004398019053041935, + 0.004131662659347057, + 0.0049486649222671986, + 0.0034309565089643, + 0.006433486007153988, + 0.007231926079839468, + 0.006461756303906441, + 0.04188338294625282, + 0.0064798337407410145, + 0.007998459041118622, + 0.014010540209710598, + 0.007344784680753946, + 0.01233800407499075, + 0.029405564069747925, + 0.0069968923926353455, + 0.01224532537162304, + 0.04502062872052193, + 0.006716035772114992, + 0.004906211979687214, + 0.005784314591437578, + 0.007989596575498581, + 0.032077889889478683, + 0.002414002316072583, + 0.0030850241892039776, + 0.002636232879012823, + 0.0030544097535312176, + 0.002356128301471472, + 0.0024235190358012915, + 0.0020214137621223927, + 0.002301436150446534, + 0.003911267034709454, + 0.0043433429673314095, + 0.004546452779322863, + 0.004754925146698952, + 0.00491816783323884, + 0.003538393648341298, + 0.004200619645416737, + 0.003622820833697915, + 0.0069162193685770035, + 0.03621014952659607, + 0.031048890203237534, + 0.010456264950335026, + 0.028501925989985466, + 0.010350930504500866, + 0.008807864040136337, + 0.0058537935838103294, + 0.01790555752813816, + 0.03010578453540802, + 0.02384701929986477, + 0.01666068471968174, + 0.03465979918837547, + 0.014584401622414589, + 0.018624041229486465, + 0.01079480815678835, + 0.008208952844142914, + 0.02922871895134449, + 0.03470294177532196, + 0.010497529059648514, + 0.009091212414205074, + 0.010450179688632488, + 0.0047691077925264835, + 0.007903552614152431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.004096213262528181 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978767640888691 + ], + "opposite_logprob_tokens": [ + -3.52978777885437 + ], + "logprob_reward": -0.02978767640888691, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978767640888691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14696630835533142, + 0.0450780875980854, + 0.05244661495089531, + 0.15324290096759796, + 0.12028421461582184, + 0.04072119668126106, + 0.03462382033467293, + 0.13234713673591614, + 0.2196083962917328, + 0.05468136444687843 + ], + "credit_map": [ + 0.011046375147998333, + 0.01303968857973814, + 0.01615975983440876, + 0.019456885755062103, + 0.01891624927520752, + 0.02077868953347206, + 0.028719812631607056, + 0.01884884014725685, + 0.004161427728831768, + 0.0034192961174994707, + 0.007197943050414324, + 0.005792978219687939, + 0.004049441311508417, + 0.008310914039611816, + 0.005906149744987488, + 0.006239930633455515, + 0.00435686856508255, + 0.006843782961368561, + 0.004673202987760305, + 0.005880683194845915, + 0.008606422692537308, + 0.00524910306558013, + 0.005346499849110842, + 0.011490045115351677, + 0.007850301451981068, + 0.04865378141403198, + 0.007472071796655655, + 0.010913086123764515, + 0.016538796946406364, + 0.008430326357483864, + 0.01653432846069336, + 0.0368502102792263, + 0.006975291762501001, + 0.016712287440896034, + 0.0421517938375473, + 0.008010308258235455, + 0.006725077982991934, + 0.005877544172108173, + 0.008522889576852322, + 0.025309011340141296, + 0.0029379185289144516, + 0.008555702865123749, + 0.004425725433975458, + 0.002951876726001501, + 0.004655440337955952, + 0.0048693763092160225, + 0.005495255347341299, + 0.0068299006670713425, + 0.0032306823413819075, + 0.004403185565024614, + 0.004438475240021944, + 0.004054829943925142, + 0.006853354629129171, + 0.0052978661842644215, + 0.0031413049437105656, + 0.0032041179947555065, + 0.007691968698054552, + 0.0281671192497015, + 0.026403123512864113, + 0.013422084972262383, + 0.022587789222598076, + 0.011743802577257156, + 0.010149598121643066, + 0.012181639671325684, + 0.02138807624578476, + 0.052177343517541885, + 0.025641800835728645, + 0.02417411468923092, + 0.03900105506181717, + 0.021930983290076256, + 0.01913703791797161, + 0.01615796983242035, + 0.005982513539493084, + 0.012061611749231815, + 0.014433035627007484, + 0.006904827430844307, + 0.0038249630015343428, + 0.004907118156552315, + 0.0026961518451571465, + 0.003871142864227295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.02978767640888691 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328423038125038 + ], + "opposite_logprob_tokens": [ + -3.7732841968536377 + ], + "logprob_reward": -0.02328423038125038, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328423038125038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08512751013040543, + 0.04403373599052429, + 0.03160538896918297, + 0.03901955112814903, + 0.08317688852548599, + 0.04420953989028931, + 0.04229366406798363, + 0.4610193371772766, + 0.06326919794082642, + 0.10624521225690842 + ], + "credit_map": [ + 0.005734306760132313, + 0.00680269580334425, + 0.013624841347336769, + 0.015637708827853203, + 0.009481367655098438, + 0.00852756667882204, + 0.013569733127951622, + 0.01174929365515709, + 0.00654076412320137, + 0.0034654990304261446, + 0.006380494683980942, + 0.005091182421892881, + 0.0035562128759920597, + 0.004633570555597544, + 0.008917653933167458, + 0.005448359064757824, + 0.0025737553369253874, + 0.003887321101501584, + 0.0027972261887043715, + 0.0036033696960657835, + 0.005654309410601854, + 0.0030509154312312603, + 0.0055908882059156895, + 0.004447604063898325, + 0.002880842424929142, + 0.009731084108352661, + 0.003829436143860221, + 0.003954881802201271, + 0.0044311899691820145, + 0.003006596118211746, + 0.003893130924552679, + 0.00729239359498024, + 0.006438099313527346, + 0.016326135024428368, + 0.024659983813762665, + 0.005797004327178001, + 0.009026535786688328, + 0.004703134763985872, + 0.006654513534158468, + 0.009571495465934277, + 0.00316369510255754, + 0.00789558608084917, + 0.0052312216721475124, + 0.00348661164753139, + 0.008430419489741325, + 0.006724564358592033, + 0.004917355719953775, + 0.004360090475529432, + 0.003678803099319339, + 0.007185760419815779, + 0.005223450250923634, + 0.005551307462155819, + 0.007407354656606913, + 0.006916802376508713, + 0.003384083043783903, + 0.0029461050871759653, + 0.025446521118283272, + 0.11038965731859207, + 0.09476064145565033, + 0.029971621930599213, + 0.10960833728313446, + 0.037618789821863174, + 0.032918140292167664, + 0.020305639132857323, + 0.0037652950268238783, + 0.01554903294891119, + 0.01342759095132351, + 0.0055363355204463005, + 0.009232836775481701, + 0.005146014969795942, + 0.005244700238108635, + 0.005367395468056202, + 0.005342202726751566, + 0.03860719874501228, + 0.02096865512430668, + 0.0055569554679095745, + 0.017448412254452705, + 0.007614624686539173, + 0.004587376024574041, + 0.00611979840323329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.02328423038125038 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "lingering, static wide shot plunges the viewer into the heart of a chaotic shipyard, ripped apart by a fierce storm. The vast skeletal hulls of unfinished vessels, their raw steel frames a web of rust-orange and deep shadows, twisted into jagged, anxious shapes clawing at a wounded, roiling sky. The entire scene is rendered with expressionistic, emotive brushstrokes and distorted perspective, with thick, diagonal flakes of rain streaking across the frame like frantic scratches on canvas. Erratic, blinding lightning momentarily paints the world a stark, electric white, casting long, twisting shadows that wriggle and contort with a life of their own. The ground is a roiling sea of mud and dark, reflective puddles, reflecting the sickly greens and deep indigos of the clouds above. Every line is sharp and unsettling, every shape feels unstable, creating a powerful, claustrophobic atmosphere, as if the landscape itself is screaming in silent terror.", + "seed": 13909, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_9/index_6540_seed_13909_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.027602043421939015, + "vlm_logprob": -0.027602043421939015, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.027602043421939015, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03586193546652794 + ], + "opposite_logprob_tokens": [ + -3.4108619689941406 + ], + "logprob_reward": -0.03586193546652794, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03586193546652794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13382597267627716, + 0.03571973741054535, + 0.03459392115473747, + 0.0295160673558712, + 0.07931909710168839, + 0.20944936573505402, + 0.10479572415351868, + 0.10607200115919113, + 0.2232525646686554, + 0.043455496430397034 + ], + "credit_map": [ + 0.014864682219922543, + 0.027267470955848694, + 0.01751432567834854, + 0.009904555045068264, + 0.018520191311836243, + 0.01195838674902916, + 0.013042368926107883, + 0.02075398713350296, + 0.005318587180227041, + 0.003690164303407073, + 0.00452315853908658, + 0.0031149019487202168, + 0.0035173462238162756, + 0.004660175647586584, + 0.005424785427749157, + 0.005470620933920145, + 0.0050240387208759785, + 0.005311113782227039, + 0.0045877015218138695, + 0.0025397262070327997, + 0.005625036079436541, + 0.0047409795224666595, + 0.0027331507299095392, + 0.004032174125313759, + 0.0031780663412064314, + 0.0031434122938662767, + 0.005118194967508316, + 0.004675732459872961, + 0.0031782444566488266, + 0.004345799796283245, + 0.0023353705182671547, + 0.003541249781847, + 0.006861768197268248, + 0.012318886816501617, + 0.00539760896936059, + 0.005190432537347078, + 0.0227204617112875, + 0.010436120443046093, + 0.004831066355109215, + 0.011562759056687355, + 0.013386421836912632, + 0.023567207157611847, + 0.02628088742494583, + 0.022698882967233658, + 0.022231586277484894, + 0.02995982952415943, + 0.040267083793878555, + 0.031057490035891533, + 0.02216058410704136, + 0.01228174939751625, + 0.013008982874453068, + 0.010827234946191311, + 0.011569340713322163, + 0.006568953860551119, + 0.013959529809653759, + 0.014419353567063808, + 0.015729838982224464, + 0.009902894496917725, + 0.016395380720496178, + 0.008454876951873302, + 0.013841082341969013, + 0.01572989672422409, + 0.01289554126560688, + 0.013122495263814926, + 0.019802840426564217, + 0.06115449592471123, + 0.014971117489039898, + 0.026061417534947395, + 0.03316410258412361, + 0.01769907958805561, + 0.017738856375217438, + 0.03266065940260887, + 0.004198824055492878, + 0.0048079355619847775, + 0.005013497546315193, + 0.007963617332279682, + 0.006234728731215, + 0.004019981250166893, + 0.005299602635204792, + 0.005917315371334553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.03586193546652794 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016080450266599655 + ], + "opposite_logprob_tokens": [ + -4.141080379486084 + ], + "logprob_reward": -0.016080450266599655, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016080450266599655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15193843841552734, + 0.26641690731048584, + 0.03148959204554558, + 0.025088360533118248, + 0.059556473046541214, + 0.1220937967300415, + 0.0592537485063076, + 0.05918627977371216, + 0.15473672747612, + 0.07023964077234268 + ], + "credit_map": [ + 0.033351823687553406, + 0.012701379135251045, + 0.026948973536491394, + 0.019328352063894272, + 0.012902368791401386, + 0.012841131538152695, + 0.0161164328455925, + 0.017747977748513222, + 0.040277495980262756, + 0.05658680945634842, + 0.03143022954463959, + 0.02002762071788311, + 0.0372256338596344, + 0.02467500790953636, + 0.025204215198755264, + 0.03098989650607109, + 0.004164815414696932, + 0.005073827225714922, + 0.004253231454640627, + 0.002641854342073202, + 0.003952547442167997, + 0.0031306135933846235, + 0.0032373894937336445, + 0.005035311449319124, + 0.0031250216998159885, + 0.0034605080727487803, + 0.0028226932045072317, + 0.004115221556276083, + 0.002775664906948805, + 0.0016871828120201826, + 0.003929563332349062, + 0.0031725040171295404, + 0.004454553592950106, + 0.00530181685462594, + 0.00674035819247365, + 0.00532350642606616, + 0.008721861056983471, + 0.008425082080066204, + 0.011463365517556667, + 0.00912593211978674, + 0.009817814454436302, + 0.018470093607902527, + 0.021376628428697586, + 0.012531202286481857, + 0.011183838360011578, + 0.01432598102837801, + 0.016802119091153145, + 0.017586130648851395, + 0.011982099153101444, + 0.006169147323817015, + 0.0074530430138111115, + 0.0056348321959376335, + 0.00931224413216114, + 0.0046774945221841335, + 0.005942763760685921, + 0.008082118816673756, + 0.007731403689831495, + 0.005029810592532158, + 0.007101289927959442, + 0.004269982688128948, + 0.005329765845090151, + 0.012774458155035973, + 0.009183386340737343, + 0.00776618393138051, + 0.013751856982707977, + 0.041671525686979294, + 0.009825787506997585, + 0.0181757602840662, + 0.022196734324097633, + 0.012464539147913456, + 0.013748027384281158, + 0.02290249615907669, + 0.007817212492227554, + 0.006934668868780136, + 0.005003813188523054, + 0.0066102719865739346, + 0.018834615126252174, + 0.00752475019544363, + 0.007589784916490316, + 0.009924530051648617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.016080450266599655 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013367536012083292 + ], + "opposite_logprob_tokens": [ + -6.626336574554443 + ], + "logprob_reward": -0.0013367536012083292, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013367536012083292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08180064707994461, + 0.07619994133710861, + 0.043848130851984024, + 0.12305724620819092, + 0.051587700843811035, + 0.18113815784454346, + 0.06389426440000534, + 0.11450719833374023, + 0.22486811876296997, + 0.03909863904118538 + ], + "credit_map": [ + 0.012128612026572227, + 0.010384714230895042, + 0.014126214198768139, + 0.006771374959498644, + 0.009441708214581013, + 0.009938500821590424, + 0.006826790515333414, + 0.012182729318737984, + 0.00765127083286643, + 0.011629029177129269, + 0.007797589525580406, + 0.00537111284211278, + 0.009135612286627293, + 0.013247278518974781, + 0.009818396531045437, + 0.011549648828804493, + 0.005434037651866674, + 0.00659747002646327, + 0.005457908380776644, + 0.0035580534022301435, + 0.006299681030213833, + 0.005196596495807171, + 0.00472565833479166, + 0.006578722968697548, + 0.015749795362353325, + 0.01978559046983719, + 0.014677092432975769, + 0.016066448763012886, + 0.013610731810331345, + 0.015824757516384125, + 0.011895398609340191, + 0.015447425656020641, + 0.004051005467772484, + 0.006688035558909178, + 0.002761624753475189, + 0.00366219156421721, + 0.011698821559548378, + 0.010390715673565865, + 0.004428060259670019, + 0.007907241582870483, + 0.013611492700874805, + 0.023766107857227325, + 0.025019077584147453, + 0.02037639170885086, + 0.014484668150544167, + 0.026961402967572212, + 0.03022194653749466, + 0.026697048917412758, + 0.008880223147571087, + 0.005042430479079485, + 0.008087915368378162, + 0.00533136585727334, + 0.012149007059633732, + 0.009344160556793213, + 0.006998573895543814, + 0.008060581050813198, + 0.01770695112645626, + 0.01121680811047554, + 0.017589803785085678, + 0.009493044577538967, + 0.015937399119138718, + 0.014732858166098595, + 0.013523049652576447, + 0.014307277277112007, + 0.019999202340841293, + 0.05971141904592514, + 0.014252296648919582, + 0.02580106444656849, + 0.033614981919527054, + 0.018932826817035675, + 0.018433529883623123, + 0.03412279486656189, + 0.0034719740506261587, + 0.006378547754138708, + 0.003793456358835101, + 0.006152831017971039, + 0.0044662198051810265, + 0.004961986560374498, + 0.004233787767589092, + 0.005639833863824606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.0013367536012083292 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978975884616375 + ], + "opposite_logprob_tokens": [ + -3.529789686203003 + ], + "logprob_reward": -0.02978975884616375, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978975884616375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11609144508838654, + 0.09199976921081543, + 0.08901941776275635, + 0.1684199720621109, + 0.052695922553539276, + 0.1251291185617447, + 0.06904598325490952, + 0.05781247839331627, + 0.17114868760108948, + 0.05863719433546066 + ], + "credit_map": [ + 0.01664232462644577, + 0.013778090476989746, + 0.018019383773207664, + 0.010265108197927475, + 0.014475241303443909, + 0.015788108110427856, + 0.012886017560958862, + 0.01423716265708208, + 0.01098098885267973, + 0.018637649714946747, + 0.011446639895439148, + 0.006922098807990551, + 0.011165397241711617, + 0.008063585497438908, + 0.012726400047540665, + 0.012057009153068066, + 0.010453632101416588, + 0.016948124393820763, + 0.013628182001411915, + 0.00676943501457572, + 0.017157841473817825, + 0.007700519170612097, + 0.0075224400497972965, + 0.008839234709739685, + 0.015268340706825256, + 0.019124934449791908, + 0.024391261860728264, + 0.025940945371985435, + 0.02124776318669319, + 0.014477641321718693, + 0.024678021669387817, + 0.023291047662496567, + 0.005573006346821785, + 0.0074968705885112286, + 0.005538639146834612, + 0.006373962387442589, + 0.008277947083115578, + 0.007057173643261194, + 0.0063171545043587685, + 0.006061169318854809, + 0.008916610851883888, + 0.014762918464839458, + 0.014521326869726181, + 0.012997269630432129, + 0.016676297411322594, + 0.022640570998191833, + 0.014825255610048771, + 0.01978885754942894, + 0.008433233946561813, + 0.005773174110800028, + 0.007453264202922583, + 0.006291586440056562, + 0.009280932135879993, + 0.015841541811823845, + 0.007213211618363857, + 0.008759032003581524, + 0.006198427174240351, + 0.004976909141987562, + 0.006366358604282141, + 0.005349680781364441, + 0.00819645170122385, + 0.008397904224693775, + 0.010700788348913193, + 0.007625950500369072, + 0.014947833493351936, + 0.046380311250686646, + 0.011101911775767803, + 0.019689081236720085, + 0.0248210821300745, + 0.014801343902945518, + 0.013771777041256428, + 0.02563532069325447, + 0.006525313947349787, + 0.007645720615983009, + 0.006275166757404804, + 0.006501153577119112, + 0.009074249304831028, + 0.008021468296647072, + 0.005996995139867067, + 0.008597118780016899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -0.02978975884616375 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054941318929195404 + ], + "opposite_logprob_tokens": [ + -2.929941415786743 + ], + "logprob_reward": -0.054941318929195404, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054941318929195404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12692002952098846, + 0.07179596275091171, + 0.12247206270694733, + 0.1462179720401764, + 0.05504614859819412, + 0.1548488438129425, + 0.04099356010556221, + 0.08116307109594345, + 0.14872696995735168, + 0.05181533470749855 + ], + "credit_map": [ + 0.010766575112938881, + 0.030421888455748558, + 0.013091767206788063, + 0.008484091609716415, + 0.019991694018244743, + 0.013068284839391708, + 0.011482736095786095, + 0.01961299404501915, + 0.006887648720294237, + 0.0078823771327734, + 0.006885630078613758, + 0.006646409165114164, + 0.012654162012040615, + 0.010838469490408897, + 0.012300971895456314, + 0.007700297981500626, + 0.012968792580068111, + 0.01800072379410267, + 0.014766253530979156, + 0.008257429115474224, + 0.022823378443717957, + 0.021628480404615402, + 0.011181016452610493, + 0.012845991179347038, + 0.012022764421999454, + 0.015500265173614025, + 0.018907155841588974, + 0.021332815289497375, + 0.022635644301772118, + 0.01053106039762497, + 0.02564818784594536, + 0.01964009553194046, + 0.005732893478125334, + 0.007259360980242491, + 0.005630147643387318, + 0.006733627989888191, + 0.008329883217811584, + 0.006415353622287512, + 0.006754089146852493, + 0.008190792985260487, + 0.013513674028217793, + 0.020164193585515022, + 0.02550983615219593, + 0.015115351416170597, + 0.020721515640616417, + 0.0166337750852108, + 0.019593287259340286, + 0.02359720505774021, + 0.004222211427986622, + 0.004232907667756081, + 0.0044205463491380215, + 0.003804665757343173, + 0.007189767435193062, + 0.006034889258444309, + 0.005739156622439623, + 0.005349416751414537, + 0.010048125870525837, + 0.007836854085326195, + 0.011309860274195671, + 0.006212894804775715, + 0.012373505160212517, + 0.013557303696870804, + 0.009994791820645332, + 0.009829742833971977, + 0.013234821148216724, + 0.03726481646299362, + 0.010471316054463387, + 0.017514055594801903, + 0.021003857254981995, + 0.013644455000758171, + 0.01279521081596613, + 0.022798441350460052, + 0.0045646182261407375, + 0.004950973205268383, + 0.0036877004895359278, + 0.006258510053157806, + 0.011565339751541615, + 0.00714375963434577, + 0.006087759975343943, + 0.007556670345366001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.054941318929195404 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "lingering, static wide shot plunges the viewer into the heart of a chaotic shipyard, ripped apart by a fierce storm. The vast skeletal hulls of unfinished vessels, their raw steel frames a web of rust-orange and deep shadows, twisted into jagged, anxious shapes clawing at a wounded, roiling sky. The entire scene is rendered with expressionistic, emotive brushstrokes and distorted perspective, with thick, diagonal flakes of rain streaking across the frame like frantic scratches on canvas. Erratic, blinding lightning momentarily paints the world a stark, electric white, casting long, twisting shadows that wriggle and contort with a life of their own. The ground is a roiling sea of mud and dark, reflective puddles, reflecting the sickly greens and deep indigos of the clouds above. Every line is sharp and unsettling, every shape feels unstable, creating a powerful, claustrophobic atmosphere, as if the landscape itself is screaming in silent terror.", + "seed": 13910, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_9/index_6540_seed_13910_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1882107098121196, + "vlm_logprob": -0.1882107098121196, + "vlm_yes_no_margin": 3.699999952316284, + "vlm_token_credit": -0.1882107098121196, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8996831178665161 + ], + "opposite_logprob_tokens": [ + -0.5246831178665161 + ], + "logprob_reward": -0.8996831178665161, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8996831178665161, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1805606484413147, + 0.07215346395969391, + 0.08279719948768616, + 0.07754101604223251, + 0.1182241290807724, + 0.10204271972179413, + 0.07835325598716736, + 0.03327463939785957, + 0.2309456616640091, + 0.024107323959469795 + ], + "credit_map": [ + 0.01289865281432867, + 0.018221378326416016, + 0.021727224811911583, + 0.023384882137179375, + 0.021711496636271477, + 0.0174398235976696, + 0.04905520752072334, + 0.016121959313750267, + 0.007314061280339956, + 0.006084095221012831, + 0.005688309669494629, + 0.01166507974267006, + 0.011608710512518883, + 0.006972507573664188, + 0.013541517779231071, + 0.009279171004891396, + 0.006938538514077663, + 0.009259107522666454, + 0.0069242860190570354, + 0.012836247682571411, + 0.016832539811730385, + 0.011812697164714336, + 0.01022629626095295, + 0.007967479526996613, + 0.008955036289989948, + 0.008711904287338257, + 0.006692597176879644, + 0.005460190586745739, + 0.011860242113471031, + 0.013276941142976284, + 0.012735674157738686, + 0.009848417714238167, + 0.011793156154453754, + 0.01131289079785347, + 0.012941843830049038, + 0.011249180883169174, + 0.011712377890944481, + 0.021567028015851974, + 0.018901608884334564, + 0.018746038898825645, + 0.008690771646797657, + 0.014753932133316994, + 0.00825173407793045, + 0.009958385489881039, + 0.013555560261011124, + 0.014425108209252357, + 0.016205059364438057, + 0.016202161088585854, + 0.007801175117492676, + 0.007955855689942837, + 0.008208257146179676, + 0.013429826125502586, + 0.009293509647250175, + 0.008551020175218582, + 0.011335194110870361, + 0.011778416112065315, + 0.0023522418923676014, + 0.0026435067411512136, + 0.0034217694774270058, + 0.00511257303878665, + 0.0051162042655050755, + 0.004369688220322132, + 0.005466165486723185, + 0.004792488645762205, + 0.01562112383544445, + 0.01732657477259636, + 0.07425378262996674, + 0.0163606945425272, + 0.01854587160050869, + 0.01502174325287342, + 0.016952799633145332, + 0.056863050907850266, + 0.0018856943352147937, + 0.001993779093027115, + 0.007108940277248621, + 0.001748439623042941, + 0.001577294897288084, + 0.0027453529182821512, + 0.0016844738274812698, + 0.005363346543163061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.8996831178665161 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009798170067369938 + ], + "opposite_logprob_tokens": [ + -4.634798049926758 + ], + "logprob_reward": -0.009798170067369938, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009798170067369938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12080785632133484, + 0.15385091304779053, + 0.11780904233455658, + 0.06933067739009857, + 0.0831800326704979, + 0.10680573433637619, + 0.09101563692092896, + 0.025159263983368874, + 0.219662144780159, + 0.012378694489598274 + ], + "credit_map": [ + 0.008148640394210815, + 0.01182244997471571, + 0.015158620662987232, + 0.01632172241806984, + 0.0160768274217844, + 0.010659277439117432, + 0.03256314620375633, + 0.010057179257273674, + 0.011670488864183426, + 0.014215150848031044, + 0.017607038840651512, + 0.03125201165676117, + 0.024085616692900658, + 0.01346966065466404, + 0.025043759495019913, + 0.016507206484675407, + 0.010798200033605099, + 0.013170219026505947, + 0.01052090898156166, + 0.014352795667946339, + 0.02374955266714096, + 0.018195226788520813, + 0.014858505688607693, + 0.01216362789273262, + 0.008494578301906586, + 0.008654887787997723, + 0.006951803807169199, + 0.00527372257784009, + 0.010636557824909687, + 0.009804892353713512, + 0.009987210854887962, + 0.009527028538286686, + 0.009947698563337326, + 0.008939388208091259, + 0.00912198144942522, + 0.006820543669164181, + 0.008902381174266338, + 0.015648748725652695, + 0.010867143981158733, + 0.012932147830724716, + 0.009882253594696522, + 0.01477089710533619, + 0.008429312147200108, + 0.010598787106573582, + 0.014342611655592918, + 0.014400710351765156, + 0.017514316365122795, + 0.016866840422153473, + 0.009116658009588718, + 0.009561309590935707, + 0.009074965491890907, + 0.015524987131357193, + 0.011598283424973488, + 0.0108290771022439, + 0.012443275190889835, + 0.012867084704339504, + 0.0018368064193055034, + 0.001904974109493196, + 0.0025166175328195095, + 0.0035245518665760756, + 0.0038749822415411472, + 0.0041895764879882336, + 0.003950555808842182, + 0.0033612006809562445, + 0.015257774852216244, + 0.016354257240891457, + 0.07268229126930237, + 0.01390247792005539, + 0.014650016091763973, + 0.013774443417787552, + 0.01693478599190712, + 0.05610610172152519, + 0.0010166049469262362, + 0.0019904777873307467, + 0.0017134486697614193, + 0.001527984393760562, + 0.0017768549732863903, + 0.00204230984672904, + 0.001081600901670754, + 0.0012294132029637694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.009798170067369938 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036140859592705965 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036140859592705965, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036140859592705965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10505972057580948, + 0.09013833850622177, + 0.09591042250394821, + 0.06533630192279816, + 0.0881568193435669, + 0.08433734625577927, + 0.0680384412407875, + 0.04293800890445709, + 0.3125404119491577, + 0.04754415899515152 + ], + "credit_map": [ + 0.007665397133678198, + 0.008911906741559505, + 0.013383641839027405, + 0.01581050083041191, + 0.011685606092214584, + 0.010469453409314156, + 0.027086880058050156, + 0.010046334937214851, + 0.010102035477757454, + 0.007383482065051794, + 0.007092111743986607, + 0.013711639679968357, + 0.014112362638115883, + 0.007923656143248081, + 0.01771407015621662, + 0.012098983861505985, + 0.008832884952425957, + 0.010712157934904099, + 0.008059537038207054, + 0.009684456512331963, + 0.018308909609913826, + 0.015313200652599335, + 0.01307042594999075, + 0.011928858235478401, + 0.009063966572284698, + 0.008170721121132374, + 0.007672208361327648, + 0.005077228881418705, + 0.007788418792188168, + 0.009190893732011318, + 0.009420640766620636, + 0.008952229283750057, + 0.009067147970199585, + 0.008999892510473728, + 0.009839601814746857, + 0.007375551387667656, + 0.008781607262790203, + 0.016397058963775635, + 0.013346361927688122, + 0.014349602162837982, + 0.008034907281398773, + 0.011469242163002491, + 0.0063284398056566715, + 0.007927807047963142, + 0.011659110896289349, + 0.012207115069031715, + 0.013625195249915123, + 0.013085536658763885, + 0.006949933245778084, + 0.0071896458975970745, + 0.007112914230674505, + 0.011309604160487652, + 0.008554583415389061, + 0.007707424461841583, + 0.00934554636478424, + 0.009868795052170753, + 0.0038827320095151663, + 0.005110090132802725, + 0.003249793779104948, + 0.0067464206367731094, + 0.007593454793095589, + 0.004405444022268057, + 0.0056106396950781345, + 0.006339436396956444, + 0.021532300859689713, + 0.02807917632162571, + 0.10509423911571503, + 0.020642444491386414, + 0.019895296543836594, + 0.018431231379508972, + 0.02223360165953636, + 0.07663212716579437, + 0.0032070702873170376, + 0.009090925566852093, + 0.012652775272727013, + 0.003080141730606556, + 0.003994436468929052, + 0.0028624290134757757, + 0.0024436763487756252, + 0.010212704539299011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.0036140859592705965 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018176767975091934 + ], + "opposite_logprob_tokens": [ + -4.018176555633545 + ], + "logprob_reward": -0.018176767975091934, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018176767975091934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26850759983062744, + 0.04822809621691704, + 0.05794564634561539, + 0.04613322764635086, + 0.08165167272090912, + 0.08119671791791916, + 0.06190939247608185, + 0.028585145249962807, + 0.29522159695625305, + 0.030620895326137543 + ], + "credit_map": [ + 0.016787540167570114, + 0.03793371841311455, + 0.03698080778121948, + 0.03569994494318962, + 0.028228625655174255, + 0.024921976029872894, + 0.06500736624002457, + 0.022947607561945915, + 0.004592613782733679, + 0.005072017665952444, + 0.005414485465735197, + 0.0065210312604904175, + 0.008522662334144115, + 0.004254504106938839, + 0.008188200183212757, + 0.005662580952048302, + 0.005508871749043465, + 0.004831536673009396, + 0.005741225555539131, + 0.005402495618909597, + 0.009948783554136753, + 0.009520887397229671, + 0.008101066574454308, + 0.008890786208212376, + 0.007329981308430433, + 0.0060548847541213036, + 0.006527482531964779, + 0.003909151069819927, + 0.005261266604065895, + 0.005217676982283592, + 0.005525781773030758, + 0.006307005882263184, + 0.009039592929184437, + 0.008127916604280472, + 0.010362730361521244, + 0.007736897561699152, + 0.008091485127806664, + 0.014694767072796822, + 0.011498761363327503, + 0.012099526822566986, + 0.007493453565984964, + 0.01131441816687584, + 0.006452263332903385, + 0.005962809547781944, + 0.0100236302241683, + 0.013328798115253448, + 0.013495502062141895, + 0.013125843368470669, + 0.006238861475139856, + 0.0067254118621349335, + 0.006483714561909437, + 0.01011541299521923, + 0.008002618327736855, + 0.007721879053860903, + 0.007961945608258247, + 0.008659550920128822, + 0.002223358489573002, + 0.003338068723678589, + 0.002469318686053157, + 0.0038591232150793076, + 0.006013147067278624, + 0.002770665567368269, + 0.0038331381510943174, + 0.004078328609466553, + 0.02025468833744526, + 0.027373410761356354, + 0.09929971396923065, + 0.019630806520581245, + 0.01758742332458496, + 0.017186516895890236, + 0.02160843461751938, + 0.0722806379199028, + 0.003395423525944352, + 0.005033720284700394, + 0.006120703183114529, + 0.0028596508782356977, + 0.00401698611676693, + 0.003482907311990857, + 0.002297409577295184, + 0.003414096776396036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.018176767975091934 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781407192349434 + ], + "opposite_logprob_tokens": [ + -4.634781360626221 + ], + "logprob_reward": -0.009781407192349434, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781407192349434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10293048620223999, + 0.04852159321308136, + 0.04935101419687271, + 0.05277344956994057, + 0.0798126831650734, + 0.05501178652048111, + 0.028498897328972816, + 0.04301466420292854, + 0.4173409044742584, + 0.12274452298879623 + ], + "credit_map": [ + 0.00815487653017044, + 0.020648067817091942, + 0.011105233803391457, + 0.01123079564422369, + 0.014910041354596615, + 0.010124185122549534, + 0.016081469133496284, + 0.010675815865397453, + 0.004682360216975212, + 0.004672061651945114, + 0.003976645413786173, + 0.008735665120184422, + 0.007806326728314161, + 0.004904727451503277, + 0.007175033912062645, + 0.006568775046616793, + 0.005875670351088047, + 0.004446974955499172, + 0.004410258494317532, + 0.008262399584054947, + 0.008148848079144955, + 0.006460051983594894, + 0.006989344023168087, + 0.004757465794682503, + 0.009268207475543022, + 0.006644234526902437, + 0.005370843689888716, + 0.003086416982114315, + 0.008531918749213219, + 0.007276229560375214, + 0.006047220900654793, + 0.006548374891281128, + 0.006511747371405363, + 0.007112966850399971, + 0.009986098855733871, + 0.007612510118633509, + 0.007584459614008665, + 0.014772693626582623, + 0.013034133240580559, + 0.013198076747357845, + 0.005664162803441286, + 0.007582071702927351, + 0.003934529609978199, + 0.005329441279172897, + 0.007283458020538092, + 0.0079777417704463, + 0.008337078616023064, + 0.00890329945832491, + 0.002786008408293128, + 0.003284214297309518, + 0.0023497501388192177, + 0.004934606608003378, + 0.00280965818092227, + 0.003513069823384285, + 0.004445246886461973, + 0.004376343917101622, + 0.004914578050374985, + 0.005056418012827635, + 0.004335358273237944, + 0.005628514569252729, + 0.006402960512787104, + 0.004255547653883696, + 0.004918838385492563, + 0.007502450607717037, + 0.029409628361463547, + 0.0668337270617485, + 0.14538300037384033, + 0.02693268470466137, + 0.019043320789933205, + 0.017884884029626846, + 0.0253909844905138, + 0.08646269142627716, + 0.010114478878676891, + 0.021214505657553673, + 0.032826218754053116, + 0.008931991644203663, + 0.008995553478598595, + 0.007860437966883183, + 0.008421013131737709, + 0.024380328133702278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.009781407192349434 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "lingering, static wide shot plunges the viewer into the heart of a chaotic shipyard, ripped apart by a fierce storm. The vast skeletal hulls of unfinished vessels, their raw steel frames a web of rust-orange and deep shadows, twisted into jagged, anxious shapes clawing at a wounded, roiling sky. The entire scene is rendered with expressionistic, emotive brushstrokes and distorted perspective, with thick, diagonal flakes of rain streaking across the frame like frantic scratches on canvas. Erratic, blinding lightning momentarily paints the world a stark, electric white, casting long, twisting shadows that wriggle and contort with a life of their own. The ground is a roiling sea of mud and dark, reflective puddles, reflecting the sickly greens and deep indigos of the clouds above. Every line is sharp and unsettling, every shape feels unstable, creating a powerful, claustrophobic atmosphere, as if the landscape itself is screaming in silent terror.", + "seed": 13911, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000123/group_9/index_6540_seed_13911_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12916282499209047, + "vlm_logprob": -0.12916282499209047, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.12916282499209047, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5268784761428833 + ], + "opposite_logprob_tokens": [ + -0.9018784761428833 + ], + "logprob_reward": -0.5268784761428833, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5268784761428833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16002601385116577, + 0.12890030443668365, + 0.07182849943637848, + 0.14687487483024597, + 0.06137481704354286, + 0.03461335599422455, + 0.058593060821294785, + 0.1383209377527237, + 0.02910720184445381, + 0.17036092281341553 + ], + "credit_map": [ + 0.023497968912124634, + 0.01627686806023121, + 0.025262076407670975, + 0.01413729228079319, + 0.021219108253717422, + 0.023536376655101776, + 0.013611529022455215, + 0.022484781220555305, + 0.01283970195800066, + 0.012170991860330105, + 0.018922962248325348, + 0.011871350929141045, + 0.022026926279067993, + 0.018448559567332268, + 0.021797485649585724, + 0.010822330601513386, + 0.006629385519772768, + 0.007164057809859514, + 0.009538012556731701, + 0.007642390206456184, + 0.01024074386805296, + 0.01019600871950388, + 0.013072951696813107, + 0.007344952318817377, + 0.007791321258991957, + 0.009495431557297707, + 0.0122104836627841, + 0.009993471205234528, + 0.016502194106578827, + 0.0362851619720459, + 0.020563557744026184, + 0.034033261239528656, + 0.004071113653481007, + 0.007095314562320709, + 0.005083663389086723, + 0.013799143023788929, + 0.007001298945397139, + 0.00857513677328825, + 0.006065607536584139, + 0.00968354195356369, + 0.0024993487168103456, + 0.0037994973827153444, + 0.0038619637489318848, + 0.005142203997820616, + 0.005027976352721453, + 0.004990037064999342, + 0.0036698307376354933, + 0.005622497294098139, + 0.005068779457360506, + 0.005153066013008356, + 0.004439990036189556, + 0.00649708928540349, + 0.00871145911514759, + 0.01055698748677969, + 0.005392834078520536, + 0.012772858142852783, + 0.015689775347709656, + 0.02821870520710945, + 0.015286996029317379, + 0.013768747448921204, + 0.02027440257370472, + 0.01905055344104767, + 0.015096647664904594, + 0.010935110971331596, + 0.0021705010440200567, + 0.0031159124337136745, + 0.002581871347501874, + 0.0032833407167345285, + 0.005972905084490776, + 0.004257197491824627, + 0.003689458826556802, + 0.004036015830934048, + 0.01971340924501419, + 0.021501198410987854, + 0.027524279430508614, + 0.020083805546164513, + 0.016246402636170387, + 0.01827002875506878, + 0.021540945395827293, + 0.025480851531028748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.5268784761428833 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007645274978131056 + ], + "opposite_logprob_tokens": [ + -4.882645130157471 + ], + "logprob_reward": -0.007645274978131056, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007645274978131056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09153823554515839, + 0.12300915271043777, + 0.08475351333618164, + 0.135177344083786, + 0.03925047069787979, + 0.03920648619532585, + 0.0467887781560421, + 0.11427903175354004, + 0.033050745725631714, + 0.2929462492465973 + ], + "credit_map": [ + 0.0094412537291646, + 0.010164202190935612, + 0.0125671261921525, + 0.006912853568792343, + 0.021985827013850212, + 0.015269693918526173, + 0.007914458401501179, + 0.00728282006457448, + 0.014682919718325138, + 0.011871703900396824, + 0.015644563362002373, + 0.012850693427026272, + 0.020694848150014877, + 0.01582004502415657, + 0.01817990653216839, + 0.013264474458992481, + 0.007865638472139835, + 0.012433835305273533, + 0.007378851994872093, + 0.008608837611973286, + 0.012729059904813766, + 0.01486461702734232, + 0.00892600417137146, + 0.011946667917072773, + 0.008307747542858124, + 0.008748985826969147, + 0.016922229900956154, + 0.007584599312394857, + 0.021231302991509438, + 0.029389135539531708, + 0.02275865152478218, + 0.020234692841768265, + 0.002646669978275895, + 0.0036192655097693205, + 0.0026807172689586878, + 0.007947754114866257, + 0.005500192288309336, + 0.005630152300000191, + 0.005640552379190922, + 0.005585165694355965, + 0.003793546697124839, + 0.0030806278809905052, + 0.003812705399468541, + 0.004618159029632807, + 0.0059246402233839035, + 0.005882947705686092, + 0.005027299746870995, + 0.00706655764952302, + 0.005948625970631838, + 0.0046880110166966915, + 0.0064536454156041145, + 0.0034169384744018316, + 0.0036641855258494616, + 0.0074382079765200615, + 0.005750374868512154, + 0.009428789839148521, + 0.02094070240855217, + 0.015514632686972618, + 0.018566928803920746, + 0.012423079460859299, + 0.015509937889873981, + 0.015444530174136162, + 0.009261466562747955, + 0.006617754232138395, + 0.0030134913977235556, + 0.003616838250309229, + 0.0039022911805659533, + 0.0038837692700326443, + 0.006153514143079519, + 0.004858057014644146, + 0.0032953177578747272, + 0.004327466711401939, + 0.036636464297771454, + 0.037355899810791016, + 0.046607039868831635, + 0.03259040787816048, + 0.037080541253089905, + 0.033503055572509766, + 0.03354618698358536, + 0.03562667593359947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.007645274978131056 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040968069806694984 + ], + "opposite_logprob_tokens": [ + -5.504096984863281 + ], + "logprob_reward": -0.0040968069806694984, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040968069806694984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3578203320503235, + 0.10567565262317657, + 0.0438421294093132, + 0.1878138929605484, + 0.0419401191174984, + 0.03200308233499527, + 0.030422599986195564, + 0.03878048434853554, + 0.06218191236257553, + 0.09951983392238617 + ], + "credit_map": [ + 0.034843023866415024, + 0.029136354103684425, + 0.06012020632624626, + 0.018899139016866684, + 0.07897541671991348, + 0.082736536860466, + 0.025528034195303917, + 0.027581622824072838, + 0.00937863253057003, + 0.010172118432819843, + 0.013449895195662975, + 0.010861171409487724, + 0.018558889627456665, + 0.013694400899112225, + 0.017035746946930885, + 0.0125248022377491, + 0.0030436438973993063, + 0.00274329143576324, + 0.0037222725804895163, + 0.003230705624446273, + 0.00484322477132082, + 0.00990793202072382, + 0.007492578588426113, + 0.008858481422066689, + 0.009604955092072487, + 0.010776796378195286, + 0.02374911680817604, + 0.010966201312839985, + 0.019108524546027184, + 0.05051277577877045, + 0.019330481067299843, + 0.04376504197716713, + 0.002605356741696596, + 0.0039849793538451195, + 0.003613098058849573, + 0.010822939686477184, + 0.003568108892068267, + 0.006151476409286261, + 0.005234815645962954, + 0.005959345027804375, + 0.003122547175735235, + 0.004433813970535994, + 0.00256592920050025, + 0.004577597137540579, + 0.004773087799549103, + 0.004937073215842247, + 0.0025892751291394234, + 0.005003754515200853, + 0.002877614926546812, + 0.0029140058904886246, + 0.004126385319977999, + 0.004494847264140844, + 0.005579773802310228, + 0.004220993258059025, + 0.0026209307834506035, + 0.0035880496725440025, + 0.0029252739623188972, + 0.005557726603001356, + 0.0036710898857563734, + 0.0035300981253385544, + 0.007597377989441156, + 0.005686814431101084, + 0.005452667362987995, + 0.004359435755759478, + 0.004625474102795124, + 0.006261523813009262, + 0.004414772614836693, + 0.005163709167391062, + 0.012987693771719933, + 0.011747429147362709, + 0.006757140159606934, + 0.010224169120192528, + 0.016071300953626633, + 0.009464237838983536, + 0.010982626117765903, + 0.009599057957530022, + 0.015024947933852673, + 0.017167555168271065, + 0.007870267145335674, + 0.013339840807020664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0040968069806694984 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018177002668380737 + ], + "opposite_logprob_tokens": [ + -4.018177032470703 + ], + "logprob_reward": -0.018177002668380737, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018177002668380737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.347766250371933, + 0.059724994003772736, + 0.033919405192136765, + 0.4342614710330963, + 0.01288829930126667, + 0.016850557178258896, + 0.017887206748127937, + 0.018286295235157013, + 0.025558285415172577, + 0.0328572653234005 + ], + "credit_map": [ + 0.03079730086028576, + 0.027610892429947853, + 0.0630284771323204, + 0.0162828229367733, + 0.0772518441081047, + 0.08458966761827469, + 0.02295321226119995, + 0.025252025574445724, + 0.005799110047519207, + 0.00429862504824996, + 0.013741693459451199, + 0.005928944330662489, + 0.008053737692534924, + 0.00853844452649355, + 0.007509069982916117, + 0.005855364724993706, + 0.002679621335119009, + 0.002582334214821458, + 0.004047711845487356, + 0.003619334427639842, + 0.005794460419565439, + 0.0041886200197041035, + 0.006659944076091051, + 0.0043473755940794945, + 0.038929425179958344, + 0.021165555343031883, + 0.06228237599134445, + 0.021860983222723007, + 0.05490405485033989, + 0.11562050133943558, + 0.03963655233383179, + 0.07986200600862503, + 0.0010072975419461727, + 0.0013718106783926487, + 0.0012670792639255524, + 0.002041769213974476, + 0.0014646529452875257, + 0.0015788995660841465, + 0.0021523404866456985, + 0.00200444832444191, + 0.0013714939123019576, + 0.0015711908927187324, + 0.0019395392155274749, + 0.003164921421557665, + 0.0020503888372331858, + 0.0027016622480005026, + 0.0017196083208546042, + 0.0023317504674196243, + 0.0013300994178280234, + 0.0013656046940013766, + 0.0021111031528562307, + 0.0015158408787101507, + 0.003063018899410963, + 0.0026714736595749855, + 0.002363217994570732, + 0.003466846654191613, + 0.0016633577179163694, + 0.0025016001891344786, + 0.0019932365976274014, + 0.001810575951822102, + 0.003031195141375065, + 0.0027192989364266396, + 0.0023880652152001858, + 0.0021789642050862312, + 0.002287521492689848, + 0.0020253900438547134, + 0.001898789661936462, + 0.0021248862612992525, + 0.004926898516714573, + 0.005086020566523075, + 0.0027527555357664824, + 0.004456021822988987, + 0.004638381768018007, + 0.003758992999792099, + 0.004132230766117573, + 0.003228327725082636, + 0.00527528952807188, + 0.005317362491041422, + 0.002755537396296859, + 0.003751143114641309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.018177002668380737 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a chaotic shipyard featuring skeletal hulls of unfinished vessels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steel frames of the vessels depicted in rust-orange and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with expressionistic, emotive brushstrokes and distorted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick, diagonal flakes of rain streaking across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video include erratic, blinding lightning that momentarily paints the world stark white?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901656419038773 + ], + "opposite_logprob_tokens": [ + -2.4640166759490967 + ], + "logprob_reward": -0.08901656419038773, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901656419038773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7172240614891052, + 0.01804581843316555, + 0.022762419655919075, + 0.11521387100219727, + 0.017767926678061485, + 0.011941139586269855, + 0.01175807323306799, + 0.019675977528095245, + 0.015514973551034927, + 0.05009562149643898 + ], + "credit_map": [ + 0.0668368935585022, + 0.058088235557079315, + 0.12302622944116592, + 0.036651916801929474, + 0.162127286195755, + 0.17075477540493011, + 0.047586604952812195, + 0.05215217545628548, + 0.0017105634324252605, + 0.0016629945021122694, + 0.0026023928076028824, + 0.001950694597326219, + 0.0025765837635844946, + 0.0025436612777411938, + 0.0028445504140108824, + 0.002154378453269601, + 0.0015902715967968106, + 0.0017125187441706657, + 0.0021630360279232264, + 0.0021836410742253065, + 0.002705285092815757, + 0.00460708886384964, + 0.003422386012971401, + 0.004378193989396095, + 0.010578413493931293, + 0.006379397120326757, + 0.015994571149349213, + 0.006157295778393745, + 0.014009551145136356, + 0.027576610445976257, + 0.012193231843411922, + 0.022324811667203903, + 0.0009220119100064039, + 0.0016161733074113727, + 0.0013224652502685785, + 0.004261869937181473, + 0.0018939449219033122, + 0.002504464704543352, + 0.002529362216591835, + 0.0027176355943083763, + 0.0013664155267179012, + 0.0014991892967373133, + 0.0016495280433446169, + 0.001785824541002512, + 0.001499753212556243, + 0.0018419428961351514, + 0.0008886129944585264, + 0.0014098742976784706, + 0.0011598854325711727, + 0.0007990518352016807, + 0.0008861033711582422, + 0.0007607092848047614, + 0.001768050598911941, + 0.0020501501858234406, + 0.0014356463216245174, + 0.0028984760865569115, + 0.0015620592748746276, + 0.002688131295144558, + 0.002539075678214431, + 0.002291433745995164, + 0.0026098936796188354, + 0.0027695917524397373, + 0.002746043261140585, + 0.002469752449542284, + 0.001553810783661902, + 0.0017719848547130823, + 0.001407140982337296, + 0.0015727141872048378, + 0.00227368320338428, + 0.002442418597638607, + 0.0018034707754850388, + 0.002689751097932458, + 0.007050961256027222, + 0.005809818394482136, + 0.006947650574147701, + 0.004000168293714523, + 0.006682634819298983, + 0.007416663225740194, + 0.005116243846714497, + 0.007071484811604023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.08901656419038773 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_0/index_tensor([9202, 9202])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_0/index_tensor([9202, 9202])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..412b04e5b9d11b63ac46a08b6f01905b1ff3df3c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_0/index_tensor([9202, 9202])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "striking close-up, captured from a slightly upward angle, depicts the camera in a slow, steady clockwise motion, circling the majestic face of a massive Buddha statue within the Yungang Grottoes in Datong. In stark black and white, light slants in from an unseen direction, carving deep shadows across the statue's features. Its compassionate and serene expression appears even more enduring, a testament to the weathering of millennia. Countless tiny grains of sandstone, whipped up by the strong winds outside the caves, form a series of gray, dynamic curtains that relentlessly impact the statue's face. As the camera pans, viewers can clearly see a small flake of sandstone, buffeted by the wind, lift slightly from the statue's weathered cheek, only to break away from the subject in the next instant, swept away by the whirling dust. This flaking process is captured with pinpoint precision. The statue's once smooth stone surface is now covered in fine cracks and pits. The lines of the eyelids and nose, while still discernible, have been worn away by the edges. The entire picture is full of Robert Capa's documentary sense and tension, condensing the grand theme of time's ruthless erosion of immortal things into this frame of visual narrative full of conflict between dynamic (flying sand) and static (Buddha statue), conveying a sense of historical heaviness and a heartbreaking fragility.", + "seed": 27401, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_0/index_9202_seed_27401_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17155494256876408, + "vlm_logprob": -0.17155494256876408, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.17155494256876408, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.322447270154953 + ], + "opposite_logprob_tokens": [ + -1.3224472999572754 + ], + "logprob_reward": -0.322447270154953, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.322447270154953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21856746077537537, + 0.12797695398330688, + 0.09832599014043808, + 0.029413212090730667, + 0.06479644775390625, + 0.027546973899006844, + 0.09237688034772873, + 0.08892982453107834, + 0.1225396990776062, + 0.12952665984630585 + ], + "credit_map": [ + 0.02843942865729332, + 0.012912931852042675, + 0.02383190579712391, + 0.022682657465338707, + 0.02565876580774784, + 0.03312910348176956, + 0.0326928049325943, + 0.03921984136104584, + 0.010773833841085434, + 0.005509400740265846, + 0.01205685269087553, + 0.0069958604872226715, + 0.009673872962594032, + 0.028221724554896355, + 0.009750011377036572, + 0.04499538615345955, + 0.010762304067611694, + 0.005915515590459108, + 0.0075718797743320465, + 0.010783903300762177, + 0.01184384897351265, + 0.019356783479452133, + 0.014505339786410332, + 0.01758640818297863, + 0.0036841286346316338, + 0.0020554261282086372, + 0.0031723028514534235, + 0.003593705128878355, + 0.002401085337623954, + 0.006184425670653582, + 0.0030824714340269566, + 0.005239667370915413, + 0.008530131541192532, + 0.0028497683815658092, + 0.013324886560440063, + 0.005599464755505323, + 0.005473164841532707, + 0.011809400282800198, + 0.0060022794641554356, + 0.011207339353859425, + 0.004135136026889086, + 0.0016759398858994246, + 0.004594285041093826, + 0.0021750200539827347, + 0.004535053391009569, + 0.003513416275382042, + 0.0027166237123310566, + 0.004201498813927174, + 0.014370224438607693, + 0.004002145957201719, + 0.011813323944807053, + 0.007777773309499025, + 0.01568448171019554, + 0.011692102998495102, + 0.014293493703007698, + 0.012743322178721428, + 0.018471797928214073, + 0.004864641465246677, + 0.015238417312502861, + 0.00789645966142416, + 0.00801724474877119, + 0.010410195216536522, + 0.008234263397753239, + 0.01579679362475872, + 0.015866409987211227, + 0.012934531085193157, + 0.009409356862306595, + 0.018799394369125366, + 0.01058774534612894, + 0.015267138369381428, + 0.02143190987408161, + 0.018243202939629555, + 0.01200892124325037, + 0.012223443947732449, + 0.012431041337549686, + 0.029889186844229698, + 0.009296183474361897, + 0.011872228235006332, + 0.022262757644057274, + 0.019542884081602097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.322447270154953 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004131235647946596 + ], + "opposite_logprob_tokens": [ + -5.504131317138672 + ], + "logprob_reward": -0.004131235647946596, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004131235647946596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12765981256961823, + 0.12118591368198395, + 0.04942883923649788, + 0.020946897566318512, + 0.032810382544994354, + 0.03309056535363197, + 0.22913698852062225, + 0.2161567360162735, + 0.10746389627456665, + 0.062119949609041214 + ], + "credit_map": [ + 0.017023395746946335, + 0.005468712188303471, + 0.012992780655622482, + 0.007979978807270527, + 0.013560435734689236, + 0.026875777170062065, + 0.012752645649015903, + 0.03100607916712761, + 0.012839367613196373, + 0.005969977471977472, + 0.012908237054944038, + 0.008922720327973366, + 0.017428312450647354, + 0.021635500714182854, + 0.012705453671514988, + 0.028776349499821663, + 0.004190297331660986, + 0.003558067139238119, + 0.003345358418300748, + 0.0066267563961446285, + 0.0068752397783100605, + 0.008968151174485683, + 0.009348724968731403, + 0.006516244262456894, + 0.0036769285798072815, + 0.0011707592057064176, + 0.0023282046895474195, + 0.0029196874238550663, + 0.0024052958469837904, + 0.0029580763075500727, + 0.0024954804684966803, + 0.0029924644622951746, + 0.004483228083699942, + 0.001603413256816566, + 0.006973925046622753, + 0.004105743486434221, + 0.002839131746441126, + 0.0055550467222929, + 0.0028217460494488478, + 0.004428147803992033, + 0.005381180439144373, + 0.0015228837728500366, + 0.006448423955589533, + 0.002587605267763138, + 0.003150597447529435, + 0.005085834302008152, + 0.002664207713678479, + 0.006249833386391401, + 0.047410786151885986, + 0.016042141243815422, + 0.028072575107216835, + 0.022937551140785217, + 0.022802473977208138, + 0.03259299695491791, + 0.02512148581445217, + 0.03415699303150177, + 0.06123523414134979, + 0.009045612066984177, + 0.04354670271277428, + 0.014677817933261395, + 0.012495487928390503, + 0.020564991980791092, + 0.015367926098406315, + 0.039222970604896545, + 0.020050060003995895, + 0.008075636811554432, + 0.015580440871417522, + 0.011194354854524136, + 0.006928302813321352, + 0.014256150461733341, + 0.012204629369080067, + 0.01917431689798832, + 0.008378631435334682, + 0.003924455028027296, + 0.006795123219490051, + 0.007891600951552391, + 0.004129085689783096, + 0.008644521236419678, + 0.011731661856174469, + 0.010624874383211136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.004131235647946596 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278960227966309 + ], + "opposite_logprob_tokens": [ + -2.017789602279663 + ], + "logprob_reward": -0.14278960227966309, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278960227966309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15998299419879913, + 0.04681515321135521, + 0.1072390004992485, + 0.10796630382537842, + 0.13949580490589142, + 0.13557671010494232, + 0.07378023862838745, + 0.07280376553535461, + 0.0988653376698494, + 0.05747472122311592 + ], + "credit_map": [ + 0.01957220770418644, + 0.006688884925097227, + 0.020537352189421654, + 0.009799588471651077, + 0.0132325179874897, + 0.03589692339301109, + 0.01272466778755188, + 0.04153083264827728, + 0.004600220359861851, + 0.0022600339725613594, + 0.003355656284838915, + 0.003083041636273265, + 0.0091111334040761, + 0.008077130652964115, + 0.004607758019119501, + 0.011720171198248863, + 0.010492353700101376, + 0.005736225750297308, + 0.007355880457907915, + 0.010542938485741615, + 0.013685929588973522, + 0.024036675691604614, + 0.014624659903347492, + 0.02076433226466179, + 0.012937493622303009, + 0.006071283482015133, + 0.010064471513032913, + 0.015206051990389824, + 0.010208077728748322, + 0.01905960403382778, + 0.01571379043161869, + 0.018705520778894424, + 0.030318139120936394, + 0.009513664990663528, + 0.028132092207670212, + 0.015881678089499474, + 0.01312638632953167, + 0.016656868159770966, + 0.00964454934000969, + 0.01622241921722889, + 0.01542539894580841, + 0.008002715185284615, + 0.018471762537956238, + 0.010294059291481972, + 0.011114494875073433, + 0.02346612885594368, + 0.010579509660601616, + 0.03822262957692146, + 0.012274740263819695, + 0.00454334681853652, + 0.010769647546112537, + 0.005788999143987894, + 0.010969873517751694, + 0.008736194111406803, + 0.011988059617578983, + 0.008709363639354706, + 0.014703873544931412, + 0.003948200028389692, + 0.010944297537207603, + 0.006413372233510017, + 0.005234424024820328, + 0.009054606780409813, + 0.012115544639527798, + 0.010389444418251514, + 0.015404201112687588, + 0.009293667040765285, + 0.01150247361510992, + 0.01298150047659874, + 0.009032417088747025, + 0.007982321083545685, + 0.015088123269379139, + 0.01758061908185482, + 0.005632358603179455, + 0.004851886071264744, + 0.005551931448280811, + 0.008956093341112137, + 0.005168381612747908, + 0.007278561592102051, + 0.009604212827980518, + 0.01043128501623869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.14278960227966309 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38706985116004944 + ], + "opposite_logprob_tokens": [ + -1.137069821357727 + ], + "logprob_reward": -0.38706985116004944, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38706985116004944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1370886117219925, + 0.09249713271856308, + 0.14820429682731628, + 0.14647842943668365, + 0.08774212747812271, + 0.04644564911723137, + 0.0645168274641037, + 0.11557948589324951, + 0.09928581118583679, + 0.06216160207986832 + ], + "credit_map": [ + 0.016733484342694283, + 0.007382608484476805, + 0.014351755380630493, + 0.0090349605306983, + 0.019511083140969276, + 0.021935319527983665, + 0.015239653177559376, + 0.032899729907512665, + 0.007250238209962845, + 0.005681236274540424, + 0.008163874968886375, + 0.008033178746700287, + 0.01379940565675497, + 0.01784662716090679, + 0.008314020000398159, + 0.02340855821967125, + 0.012032260186970234, + 0.010624881833791733, + 0.009893324226140976, + 0.017745887860655785, + 0.015343887731432915, + 0.0334225594997406, + 0.020892759785056114, + 0.028248731046915054, + 0.02211260423064232, + 0.008548152633011341, + 0.019822053611278534, + 0.023162933066487312, + 0.014547825790941715, + 0.018897468224167824, + 0.018796199932694435, + 0.02059120126068592, + 0.021241897717118263, + 0.00449052918702364, + 0.013427572324872017, + 0.007918902672827244, + 0.006169503554701805, + 0.01638221926987171, + 0.007190223317593336, + 0.01092128548771143, + 0.006946447305381298, + 0.002182117197662592, + 0.008100834675133228, + 0.00369984726421535, + 0.005911160726100206, + 0.006782212760299444, + 0.004174091387540102, + 0.008648940362036228, + 0.009913481771945953, + 0.004009870812296867, + 0.011362719349563122, + 0.006366077810525894, + 0.00912301242351532, + 0.008327501825988293, + 0.007287024054676294, + 0.008127138018608093, + 0.024329932406544685, + 0.007774379104375839, + 0.018312055617570877, + 0.011035947129130363, + 0.008752414025366306, + 0.014061174355447292, + 0.0149192800745368, + 0.016394304111599922, + 0.017397476360201836, + 0.007903257384896278, + 0.010836020112037659, + 0.014941412024199963, + 0.010244578123092651, + 0.008575455285608768, + 0.012252509593963623, + 0.017135104164481163, + 0.007028166204690933, + 0.005442810244858265, + 0.00494777737185359, + 0.008442315272986889, + 0.004514873493462801, + 0.007399435620754957, + 0.013043520972132683, + 0.011342701502144337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.38706985116004944 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013367536012083292 + ], + "opposite_logprob_tokens": [ + -6.626336574554443 + ], + "logprob_reward": -0.0013367536012083292, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013367536012083292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17978721857070923, + 0.15098035335540771, + 0.22097864747047424, + 0.0724925622344017, + 0.028597144410014153, + 0.02868475951254368, + 0.05473349988460541, + 0.057395223528146744, + 0.12769176065921783, + 0.0786588191986084 + ], + "credit_map": [ + 0.025133240967988968, + 0.00926738791167736, + 0.019468214362859726, + 0.014569993130862713, + 0.02100532501935959, + 0.034462835639715195, + 0.021063994616270065, + 0.03481622040271759, + 0.015026724897325039, + 0.005661058705300093, + 0.01639689691364765, + 0.007075759116560221, + 0.010043236427009106, + 0.03313669189810753, + 0.010873767547309399, + 0.05276620760560036, + 0.02226523496210575, + 0.010545125231146812, + 0.01564798876643181, + 0.017483944073319435, + 0.02461189404129982, + 0.0542198047041893, + 0.03172900527715683, + 0.04447563737630844, + 0.012516600079834461, + 0.003753283526748419, + 0.007669432554394007, + 0.007079349830746651, + 0.005145985633134842, + 0.014369385316967964, + 0.0098632937297225, + 0.012095233425498009, + 0.004587708506733179, + 0.0011097899405285716, + 0.005761992651969194, + 0.00230844016186893, + 0.0021653787698596716, + 0.004749258980154991, + 0.0024069061037153006, + 0.005507668945938349, + 0.004468539729714394, + 0.002031040843576193, + 0.003757759463042021, + 0.0034822362940758467, + 0.002821307396516204, + 0.003907059319317341, + 0.002871862379834056, + 0.00534495385363698, + 0.010442380793392658, + 0.004623075481504202, + 0.007832770235836506, + 0.005180422216653824, + 0.00658347737044096, + 0.006915392354130745, + 0.005630603060126305, + 0.007525378372520208, + 0.007806507404893637, + 0.004123489838093519, + 0.006596080027520657, + 0.007559970952570438, + 0.006447736639529467, + 0.007255207747220993, + 0.008892487734556198, + 0.008713747374713421, + 0.018910199403762817, + 0.009234013967216015, + 0.01309137511998415, + 0.018332459032535553, + 0.012085676193237305, + 0.011580578051507473, + 0.021361645311117172, + 0.023095816373825073, + 0.007679014466702938, + 0.00707494281232357, + 0.005400441586971283, + 0.014163264073431492, + 0.008250411599874496, + 0.007493786513805389, + 0.015081114135682583, + 0.013515844941139221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0013367536012083292 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking close-up, captured from a slightly upward angle, depicts the camera in a slow, steady clockwise motion, circling the majestic face of a massive Buddha statue within the Yungang Grottoes in Datong. In stark black and white, light slants in from an unseen direction, carving deep shadows across the statue's features. Its compassionate and serene expression appears even more enduring, a testament to the weathering of millennia. Countless tiny grains of sandstone, whipped up by the strong winds outside the caves, form a series of gray, dynamic curtains that relentlessly impact the statue's face. As the camera pans, viewers can clearly see a small flake of sandstone, buffeted by the wind, lift slightly from the statue's weathered cheek, only to break away from the subject in the next instant, swept away by the whirling dust. This flaking process is captured with pinpoint precision. The statue's once smooth stone surface is now covered in fine cracks and pits. The lines of the eyelids and nose, while still discernible, have been worn away by the edges. The entire picture is full of Robert Capa's documentary sense and tension, condensing the grand theme of time's ruthless erosion of immortal things into this frame of visual narrative full of conflict between dynamic (flying sand) and static (Buddha statue), conveying a sense of historical heaviness and a heartbreaking fragility.", + "seed": 27402, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_0/index_9202_seed_27402_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1048823214136064, + "vlm_logprob": -0.1048823214136064, + "vlm_yes_no_margin": 3.500000023841858, + "vlm_token_credit": -0.1048823214136064, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.263895183801651 + ], + "opposite_logprob_tokens": [ + -1.5138951539993286 + ], + "logprob_reward": -0.263895183801651, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.263895183801651, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07428339868783951, + 0.14165851473808289, + 0.028212347999215126, + 0.054192785173654556, + 0.04563729465007782, + 0.4041121006011963, + 0.0857119932770729, + 0.04307899624109268, + 0.07799764722585678, + 0.045114900916814804 + ], + "credit_map": [ + 0.009286168031394482, + 0.009948055259883404, + 0.015539182350039482, + 0.004176280461251736, + 0.002777757588773966, + 0.011865622363984585, + 0.004989521577954292, + 0.015700813382864, + 0.014243867248296738, + 0.019732913002371788, + 0.028783541172742844, + 0.011618651449680328, + 0.007564933970570564, + 0.024222001433372498, + 0.007021165918558836, + 0.02847144566476345, + 0.003879675641655922, + 0.0035240966826677322, + 0.005615635775029659, + 0.0018405219307169318, + 0.00107770215254277, + 0.004874295089393854, + 0.0016557812923565507, + 0.005744638852775097, + 0.004701866768300533, + 0.008743065409362316, + 0.007980257272720337, + 0.005300984252244234, + 0.005018559284508228, + 0.009056327864527702, + 0.00450182007625699, + 0.00888990517705679, + 0.005461819004267454, + 0.006917799822986126, + 0.00583879416808486, + 0.009331300854682922, + 0.0024933703243732452, + 0.0058459085412323475, + 0.00325017049908638, + 0.006498131435364485, + 0.06189197301864624, + 0.018804002553224564, + 0.0815645307302475, + 0.0227938462048769, + 0.007177555467933416, + 0.09090101718902588, + 0.00918823853135109, + 0.11179094016551971, + 0.006943000014871359, + 0.011443627066910267, + 0.009786264970898628, + 0.01219925656914711, + 0.010828825645148754, + 0.0124500198289752, + 0.0076087461784482, + 0.014452255330979824, + 0.007142361719161272, + 0.004852815996855497, + 0.009180945344269276, + 0.005437425337731838, + 0.002227732678875327, + 0.004824720788747072, + 0.0031909095123410225, + 0.006222084630280733, + 0.009929009713232517, + 0.00853604357689619, + 0.00986788421869278, + 0.012244445271790028, + 0.010107658803462982, + 0.00978034920990467, + 0.010018006898462772, + 0.007514252793043852, + 0.005287252366542816, + 0.005973014980554581, + 0.006748789921402931, + 0.008799091912806034, + 0.002580791711807251, + 0.004270288627594709, + 0.004760426003485918, + 0.006695246789604425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.263895183801651 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032406931277364492 + ], + "opposite_logprob_tokens": [ + -5.753240585327148 + ], + "logprob_reward": -0.0032406931277364492, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032406931277364492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12963806092739105, + 0.0648505911231041, + 0.03643748536705971, + 0.15977135300636292, + 0.11527157574892044, + 0.19914603233337402, + 0.04887879639863968, + 0.06661354005336761, + 0.04847130551934242, + 0.1309211701154709 + ], + "credit_map": [ + 0.015887951478362083, + 0.014885214157402515, + 0.024374945089221, + 0.014725765213370323, + 0.011522174812853336, + 0.019429681822657585, + 0.01394590362906456, + 0.01486644335091114, + 0.0074730319902300835, + 0.009007747285068035, + 0.010059350170195103, + 0.006060752552002668, + 0.003979905042797327, + 0.010553847067058086, + 0.006263097748160362, + 0.011452869512140751, + 0.005106270778924227, + 0.00492196436971426, + 0.006334566976875067, + 0.0030101353768259287, + 0.0022119057830423117, + 0.006187427323311567, + 0.00305056176148355, + 0.005614657420665026, + 0.008598508313298225, + 0.009808046743273735, + 0.009818167425692081, + 0.007957490161061287, + 0.005106666125357151, + 0.057798873633146286, + 0.004789890255779028, + 0.05589373782277107, + 0.013124912045896053, + 0.017287183552980423, + 0.015491641126573086, + 0.015945618972182274, + 0.006931745912879705, + 0.017651209607720375, + 0.00836020614951849, + 0.020479070022702217, + 0.03198365494608879, + 0.012374404817819595, + 0.03980132192373276, + 0.015817057341337204, + 0.005116203799843788, + 0.0396980345249176, + 0.00736518669873476, + 0.046990178525447845, + 0.004039001651108265, + 0.005037848372012377, + 0.00610366091132164, + 0.0067754569463431835, + 0.005629242397844791, + 0.0073944032192230225, + 0.007432740181684494, + 0.006466444581747055, + 0.010777795687317848, + 0.006669783033430576, + 0.012695835903286934, + 0.00865418091416359, + 0.005321305710822344, + 0.008159775286912918, + 0.005274896509945393, + 0.009059980511665344, + 0.005336542148143053, + 0.007337603252381086, + 0.004486189689487219, + 0.010948221199214458, + 0.0038228356279432774, + 0.006213428918272257, + 0.004673526156693697, + 0.005652963183820248, + 0.01687825471162796, + 0.0202725101262331, + 0.017915481701493263, + 0.026850882917642593, + 0.009327605366706848, + 0.011377846822142601, + 0.013936608098447323, + 0.014362000860273838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 9 + ], + "selected_reward": -0.0032406931277364492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280717074871063 + ], + "opposite_logprob_tokens": [ + -2.0178072452545166 + ], + "logprob_reward": -0.14280717074871063, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14280717074871063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025149552151560783, + 0.03828379884362221, + 0.0157132800668478, + 0.06633619964122772, + 0.024396663531661034, + 0.683752179145813, + 0.042220454663038254, + 0.05263625457882881, + 0.018910426646471024, + 0.03260117769241333 + ], + "credit_map": [ + 0.003357540350407362, + 0.0029083117842674255, + 0.004377264529466629, + 0.0024600953329354525, + 0.0028670779429376125, + 0.003442496759817004, + 0.0026072142645716667, + 0.0031295542139559984, + 0.003263056045398116, + 0.0041554756462574005, + 0.005133742466568947, + 0.003006135579198599, + 0.0031012718100100756, + 0.00834360159933567, + 0.002576860599219799, + 0.008703657425940037, + 0.0020710837561637163, + 0.002227708464488387, + 0.0027687049005180597, + 0.001379384659230709, + 0.0008666463545523584, + 0.0021903214510530233, + 0.0019430677639320493, + 0.0022663648705929518, + 0.004349126480519772, + 0.005570987239480019, + 0.0056711602956056595, + 0.004281362984329462, + 0.0026868502609431744, + 0.020746689289808273, + 0.00247750012204051, + 0.02055252715945244, + 0.00216938485391438, + 0.0030738224741071463, + 0.0023789391852915287, + 0.0025970942806452513, + 0.0035930874291807413, + 0.0035327959340065718, + 0.002700717421248555, + 0.004350823815912008, + 0.09561282396316528, + 0.029532644897699356, + 0.1260504573583603, + 0.037617962807416916, + 0.013535372912883759, + 0.16525687277317047, + 0.017771463841199875, + 0.19837455451488495, + 0.004327049478888512, + 0.0042931572534143925, + 0.006666339468210936, + 0.005391398910433054, + 0.0030289252754300833, + 0.006781769450753927, + 0.00499718775972724, + 0.006734633352607489, + 0.008865824900567532, + 0.005627832375466824, + 0.01081122551113367, + 0.006403572391718626, + 0.003949476405978203, + 0.006410576403141022, + 0.003480185056105256, + 0.007087562698870897, + 0.002821710193529725, + 0.0022786303889006376, + 0.0023938578087836504, + 0.0029542322736233473, + 0.0016939706401899457, + 0.0024665421806275845, + 0.0019036972662433982, + 0.002397787058725953, + 0.003926321864128113, + 0.0035345647484064102, + 0.0039426833391189575, + 0.004554294049739838, + 0.0028308872133493423, + 0.004657287150621414, + 0.004005288705229759, + 0.005149851553142071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 7 + ], + "selected_reward": -0.14280717074871063 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295396089553833 + ], + "opposite_logprob_tokens": [ + -2.2379539012908936 + ], + "logprob_reward": -0.11295396089553833, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295396089553833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17191770672798157, + 0.06667915731668472, + 0.03454875573515892, + 0.11825776845216751, + 0.0571414977312088, + 0.1840844452381134, + 0.15352049469947815, + 0.07635533809661865, + 0.07522661983966827, + 0.06226819381117821 + ], + "credit_map": [ + 0.013178754597902298, + 0.01488184928894043, + 0.028846636414527893, + 0.010097559541463852, + 0.00895632617175579, + 0.035787273198366165, + 0.011526811867952347, + 0.04864248260855675, + 0.007661743555217981, + 0.008365306071937084, + 0.013676795177161694, + 0.007366368547081947, + 0.004475933965295553, + 0.009052739478647709, + 0.005239998456090689, + 0.010840279050171375, + 0.005529370624572039, + 0.005545326042920351, + 0.0061760288663208485, + 0.0029872339218854904, + 0.002358390251174569, + 0.005163432564586401, + 0.0023376874160021544, + 0.004451286513358355, + 0.015755604952573776, + 0.023504018783569336, + 0.025310995057225227, + 0.013519598171114922, + 0.0066950502805411816, + 0.01614406518638134, + 0.005058553069829941, + 0.0122698824852705, + 0.005952801089733839, + 0.0088937534019351, + 0.007641208358108997, + 0.010062778368592262, + 0.0043981969356536865, + 0.008482031524181366, + 0.0036385052371770144, + 0.008072220720350742, + 0.04147212952375412, + 0.015502529218792915, + 0.044795505702495575, + 0.017662538215517998, + 0.0167428869754076, + 0.015454738400876522, + 0.015089262276887894, + 0.01736484467983246, + 0.013763226568698883, + 0.018179049715399742, + 0.019775312393903732, + 0.02254904806613922, + 0.01654866524040699, + 0.020574452355504036, + 0.016730407252907753, + 0.02540033496916294, + 0.009464702568948269, + 0.007649111561477184, + 0.008947313763201237, + 0.008540966548025608, + 0.006519390270113945, + 0.010617215186357498, + 0.011727318167686462, + 0.012889323756098747, + 0.00835302285850048, + 0.012522122822701931, + 0.007713978178799152, + 0.016866935417056084, + 0.0050318543799221516, + 0.009551937691867352, + 0.0070015545934438705, + 0.008185218088328838, + 0.006424007937312126, + 0.008378366939723492, + 0.00848359614610672, + 0.012651955708861351, + 0.004026629030704498, + 0.006218055263161659, + 0.008292247541248798, + 0.007793338503688574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.11295396089553833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015145984943956137 + ], + "opposite_logprob_tokens": [ + -6.501514434814453 + ], + "logprob_reward": -0.0015145984943956137, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015145984943956137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16978789865970612, + 0.08540874719619751, + 0.06666384637355804, + 0.18660742044448853, + 0.04288550838828087, + 0.16187404096126556, + 0.10170908272266388, + 0.04094943776726723, + 0.04682443290948868, + 0.0972895398736 + ], + "credit_map": [ + 0.013737061992287636, + 0.014217407442629337, + 0.029281482100486755, + 0.00972135178744793, + 0.007375240791589022, + 0.03556244820356369, + 0.011894352734088898, + 0.047998566180467606, + 0.010164815001189709, + 0.009815467521548271, + 0.01772632822394371, + 0.006552302744239569, + 0.004799559712409973, + 0.015089068561792374, + 0.004929947201162577, + 0.016331268474459648, + 0.008966326713562012, + 0.010797017253935337, + 0.015406562946736813, + 0.004483893048018217, + 0.002847511786967516, + 0.008875273168087006, + 0.004908938426524401, + 0.010378332808613777, + 0.013157805427908897, + 0.016398100182414055, + 0.01696872152388096, + 0.011565249413251877, + 0.006388833746314049, + 0.0590633749961853, + 0.005785398185253143, + 0.05727994814515114, + 0.006551980972290039, + 0.006922445725649595, + 0.006613398436456919, + 0.004871555604040623, + 0.0029004011303186417, + 0.006646231282502413, + 0.0028506587259471416, + 0.005528837908059359, + 0.012804236263036728, + 0.020384421572089195, + 0.018030842766165733, + 0.017696000635623932, + 0.012683609500527382, + 0.034894056618213654, + 0.011379809118807316, + 0.034001074731349945, + 0.010516406036913395, + 0.010096929967403412, + 0.0113758509978652, + 0.01992076449096203, + 0.0054636853747069836, + 0.01808272860944271, + 0.00931203830987215, + 0.016940690577030182, + 0.0047698961570858955, + 0.0038368995301425457, + 0.006096034776419401, + 0.0055693588219583035, + 0.002653594361618161, + 0.005558460019528866, + 0.00522747915238142, + 0.007237717974931002, + 0.005232074297964573, + 0.00545509671792388, + 0.005448911339044571, + 0.005852652247995138, + 0.006724248174577951, + 0.006170964799821377, + 0.006447461899369955, + 0.005493029020726681, + 0.012303917668759823, + 0.014832546934485435, + 0.013925809413194656, + 0.018107078969478607, + 0.006107966415584087, + 0.009448884986341, + 0.01131336111575365, + 0.011249978095293045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.0015145984943956137 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking close-up, captured from a slightly upward angle, depicts the camera in a slow, steady clockwise motion, circling the majestic face of a massive Buddha statue within the Yungang Grottoes in Datong. In stark black and white, light slants in from an unseen direction, carving deep shadows across the statue's features. Its compassionate and serene expression appears even more enduring, a testament to the weathering of millennia. Countless tiny grains of sandstone, whipped up by the strong winds outside the caves, form a series of gray, dynamic curtains that relentlessly impact the statue's face. As the camera pans, viewers can clearly see a small flake of sandstone, buffeted by the wind, lift slightly from the statue's weathered cheek, only to break away from the subject in the next instant, swept away by the whirling dust. This flaking process is captured with pinpoint precision. The statue's once smooth stone surface is now covered in fine cracks and pits. The lines of the eyelids and nose, while still discernible, have been worn away by the edges. The entire picture is full of Robert Capa's documentary sense and tension, condensing the grand theme of time's ruthless erosion of immortal things into this frame of visual narrative full of conflict between dynamic (flying sand) and static (Buddha statue), conveying a sense of historical heaviness and a heartbreaking fragility.", + "seed": 27403, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_0/index_9202_seed_27403_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.18179221237078308, + "vlm_logprob": -0.18179221237078308, + "vlm_yes_no_margin": 2.499999976158142, + "vlm_token_credit": -0.18179221237078308, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28841203451156616 + ], + "opposite_logprob_tokens": [ + -1.413412094116211 + ], + "logprob_reward": -0.28841203451156616, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28841203451156616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03530409559607506, + 0.026406878605484962, + 0.050276026129722595, + 0.013980403542518616, + 0.0495799295604229, + 0.10065535455942154, + 0.05021948739886284, + 0.3799690902233124, + 0.26496878266334534, + 0.02863992564380169 + ], + "credit_map": [ + 0.0022658638190478086, + 0.001955946208909154, + 0.00480923755094409, + 0.0025222976692020893, + 0.005501354578882456, + 0.005854785442352295, + 0.006901037413626909, + 0.005493572913110256, + 0.0024267619010061026, + 0.0015842897119000554, + 0.003513869596645236, + 0.0020612450316548347, + 0.004080583807080984, + 0.004852197598665953, + 0.0038936769124120474, + 0.003994254861027002, + 0.0028048832900822163, + 0.001389126991853118, + 0.003911824431270361, + 0.002102271653711796, + 0.015085907652974129, + 0.004744506441056728, + 0.003526997286826372, + 0.01671050861477852, + 0.001295361784286797, + 0.0007739982102066278, + 0.0020695216953754425, + 0.0014269606908783317, + 0.0026870681904256344, + 0.0017342547653242946, + 0.0018935189582407475, + 0.0020997184328734875, + 0.0038282033056020737, + 0.0036078710108995438, + 0.009038258343935013, + 0.008565862663090229, + 0.005889165215194225, + 0.0046738265082240105, + 0.006704671308398247, + 0.007272071670740843, + 0.008335924707353115, + 0.00855174195021391, + 0.016338177025318146, + 0.015591291710734367, + 0.012957753613591194, + 0.011102893389761448, + 0.0129130519926548, + 0.01486452016979456, + 0.006001753732562065, + 0.003968618344515562, + 0.006767401471734047, + 0.007466456852853298, + 0.004974351264536381, + 0.0050033507868647575, + 0.008568480610847473, + 0.00746907340362668, + 0.04233383387327194, + 0.03237323462963104, + 0.05798141285777092, + 0.05291062593460083, + 0.0491667166352272, + 0.04222503677010536, + 0.06333945691585541, + 0.039638787508010864, + 0.03103695623576641, + 0.018442988395690918, + 0.04253099486231804, + 0.01752130128443241, + 0.024568891152739525, + 0.036894865334033966, + 0.02484028786420822, + 0.0691324919462204, + 0.003969789948314428, + 0.002693417016416788, + 0.004581992514431477, + 0.002649603644385934, + 0.002646220149472356, + 0.003605996025726199, + 0.004772291984409094, + 0.003720614593476057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.28841203451156616 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12710726261138916 + ], + "opposite_logprob_tokens": [ + -2.1271071434020996 + ], + "logprob_reward": -0.12710726261138916, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12710726261138916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06742652505636215, + 0.03452355042099953, + 0.1900523453950882, + 0.0157193373888731, + 0.03780582547187805, + 0.0910077765583992, + 0.03117876872420311, + 0.2478574961423874, + 0.19885848462581635, + 0.08556989580392838 + ], + "credit_map": [ + 0.004574783146381378, + 0.005023784004151821, + 0.008295753970742226, + 0.005191908683627844, + 0.011076703667640686, + 0.011734981089830399, + 0.012084349058568478, + 0.009444259107112885, + 0.002725864527747035, + 0.0021136559080332518, + 0.004043676424771547, + 0.0029141679406166077, + 0.005096209701150656, + 0.0065485150553286076, + 0.005510716233402491, + 0.005570744164288044, + 0.008738888427615166, + 0.0034129119012504816, + 0.013033347204327583, + 0.004217594861984253, + 0.07491026818752289, + 0.023004621267318726, + 0.00951241236180067, + 0.05322229489684105, + 0.0012373782228678465, + 0.0007014669827185571, + 0.0017005656845867634, + 0.001917832181788981, + 0.0030578277073800564, + 0.0016703552100807428, + 0.003268314292654395, + 0.0021655980963259935, + 0.00399361364543438, + 0.0028599752113223076, + 0.007795953191816807, + 0.005772854667156935, + 0.004348952788859606, + 0.003213875461369753, + 0.004331517964601517, + 0.005489081609994173, + 0.006666782312095165, + 0.004822959192097187, + 0.009802771732211113, + 0.023383602499961853, + 0.007628440856933594, + 0.00887765921652317, + 0.007805566303431988, + 0.022019989788532257, + 0.004570723976939917, + 0.002117072930559516, + 0.004186752252280712, + 0.005627663340419531, + 0.002780684968456626, + 0.00299430126324296, + 0.0047225672751665115, + 0.004179004114121199, + 0.02638760209083557, + 0.020317258313298225, + 0.036093130707740784, + 0.03447534888982773, + 0.03244183212518692, + 0.029152417555451393, + 0.04222244396805763, + 0.026767456904053688, + 0.025937911123037338, + 0.012677808292210102, + 0.03446071594953537, + 0.011654503643512726, + 0.015874700620770454, + 0.02735547348856926, + 0.017404116690158844, + 0.053493257611989975, + 0.016603179275989532, + 0.005534601863473654, + 0.02749749831855297, + 0.006423012353479862, + 0.005134124308824539, + 0.006973378825932741, + 0.006046007853001356, + 0.011358089745044708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.12710726261138916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204346567392349 + ], + "opposite_logprob_tokens": [ + -2.8120434284210205 + ], + "logprob_reward": -0.06204346567392349, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204346567392349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040621317923069, + 0.04832028970122337, + 0.05456554517149925, + 0.031099749729037285, + 0.10793211311101913, + 0.12201160192489624, + 0.08240726590156555, + 0.2900911271572113, + 0.2016190141439438, + 0.02133198454976082 + ], + "credit_map": [ + 0.0035389603581279516, + 0.0041509997099637985, + 0.005581858567893505, + 0.00506530562415719, + 0.005530704278498888, + 0.005022295750677586, + 0.008149413391947746, + 0.00358177674934268, + 0.00377201521769166, + 0.0030140571761876345, + 0.005277421325445175, + 0.004506338853389025, + 0.007406358607113361, + 0.007953573018312454, + 0.009819919243454933, + 0.00657060369849205, + 0.003343405434861779, + 0.0018175605218857527, + 0.007086490746587515, + 0.003409575903788209, + 0.018806181848049164, + 0.00780627503991127, + 0.003425488481298089, + 0.008870563469827175, + 0.0027544863987714052, + 0.001287677325308323, + 0.0050191315822303295, + 0.003909478895366192, + 0.004896284081041813, + 0.003765435190871358, + 0.004862319678068161, + 0.00460493378341198, + 0.008505010046064854, + 0.009747995063662529, + 0.014195912517607212, + 0.01807665079832077, + 0.01539013534784317, + 0.0124183539301157, + 0.016486981883645058, + 0.01311106700450182, + 0.009770276956260204, + 0.008155991323292255, + 0.016312409192323685, + 0.026106201112270355, + 0.012865670025348663, + 0.012277139350771904, + 0.012815142050385475, + 0.023708760738372803, + 0.01142085436731577, + 0.00504737813025713, + 0.015428917482495308, + 0.00917031615972519, + 0.010051521472632885, + 0.008770625106990337, + 0.012278533540666103, + 0.01023910939693451, + 0.03277687728404999, + 0.023864051327109337, + 0.04577019438147545, + 0.0407065786421299, + 0.03673289716243744, + 0.03213152661919594, + 0.04728196561336517, + 0.03082704357802868, + 0.022830259054899216, + 0.014244962483644485, + 0.03137717396020889, + 0.013430071994662285, + 0.018917694687843323, + 0.028697699308395386, + 0.01977820135653019, + 0.05234294384717941, + 0.003785994602367282, + 0.0013751588994637132, + 0.005714020226150751, + 0.0017574919620528817, + 0.001360291731543839, + 0.002119483659043908, + 0.00258993380703032, + 0.0026296081487089396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.06204346567392349 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42890825867652893 + ], + "opposite_logprob_tokens": [ + -1.0539082288742065 + ], + "logprob_reward": -0.42890825867652893, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42890825867652893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12546305358409882, + 0.04785836115479469, + 0.10033918917179108, + 0.047276537865400314, + 0.1077793538570404, + 0.11400150507688522, + 0.07327853888273239, + 0.18370714783668518, + 0.10321783274412155, + 0.09707845002412796 + ], + "credit_map": [ + 0.0068406760692596436, + 0.007691885810345411, + 0.014551138505339622, + 0.013187899254262447, + 0.020107390359044075, + 0.02289215289056301, + 0.020776620134711266, + 0.019415292888879776, + 0.005492015741765499, + 0.0034280098043382168, + 0.007436763495206833, + 0.004490434192121029, + 0.006581438239663839, + 0.006810304708778858, + 0.007420738693326712, + 0.006198652554303408, + 0.006770063191652298, + 0.002274752827361226, + 0.013857242651283741, + 0.0055259703658521175, + 0.024259842932224274, + 0.00795530341565609, + 0.005555958487093449, + 0.03414006158709526, + 0.003044976620003581, + 0.0021292937453836203, + 0.005889044608920813, + 0.003971376456320286, + 0.008921843022108078, + 0.005852434318512678, + 0.009880528785288334, + 0.007587041240185499, + 0.009262041188776493, + 0.010010897181928158, + 0.016500763595104218, + 0.016888834536075592, + 0.012442554347217083, + 0.011911415494978428, + 0.016360757872462273, + 0.014402085915207863, + 0.009126701392233372, + 0.007253201212733984, + 0.01607142947614193, + 0.026629118248820305, + 0.009006810374557972, + 0.010599061846733093, + 0.01134510524570942, + 0.023970073089003563, + 0.011107244528830051, + 0.005419458262622356, + 0.010902592912316322, + 0.009065797552466393, + 0.008460601791739464, + 0.008688820526003838, + 0.008614393882453442, + 0.011019628494977951, + 0.017942998558282852, + 0.0151450764387846, + 0.031963709741830826, + 0.025077424943447113, + 0.022736111655831337, + 0.020983649417757988, + 0.030292686074972153, + 0.019565491005778313, + 0.013323999010026455, + 0.006385045126080513, + 0.017719751223921776, + 0.0060217175632715225, + 0.008124099113047123, + 0.014498291537165642, + 0.010723683051764965, + 0.02642124891281128, + 0.01323549635708332, + 0.0086336899548769, + 0.013646163046360016, + 0.009945572353899479, + 0.009464605711400509, + 0.01292632706463337, + 0.016802599653601646, + 0.012423993088304996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.42890825867652893 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024900403805077076 + ], + "opposite_logprob_tokens": [ + -6.002490043640137 + ], + "logprob_reward": -0.0024900403805077076, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024900403805077076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1002090573310852, + 0.06481572985649109, + 0.03117041289806366, + 0.016546666622161865, + 0.02631406858563423, + 0.10495174676179886, + 0.03696130961179733, + 0.32154858112335205, + 0.23951762914657593, + 0.057964831590652466 + ], + "credit_map": [ + 0.010162802413105965, + 0.007504607085138559, + 0.012779694981873035, + 0.01869179494678974, + 0.015392609871923923, + 0.010314210318028927, + 0.015989253297448158, + 0.009374071843922138, + 0.004898576531559229, + 0.003747576614841819, + 0.006875211372971535, + 0.0050549376755952835, + 0.010626588948071003, + 0.01330264937132597, + 0.011739451438188553, + 0.008570739068090916, + 0.003042243653908372, + 0.0015686029801145196, + 0.0063393497839570045, + 0.003004303202033043, + 0.002884390763938427, + 0.0035789120011031628, + 0.0023957614321261644, + 0.008356847800314426, + 0.001734329154714942, + 0.0012779822573065758, + 0.0021170242689549923, + 0.0024243169464170933, + 0.0019202448893338442, + 0.002179674804210663, + 0.0024281993974000216, + 0.002464892575517297, + 0.002510028425604105, + 0.0022143879905343056, + 0.005158359184861183, + 0.003573718713596463, + 0.0034563399385660887, + 0.0025792797096073627, + 0.0029665911570191383, + 0.0038553625345230103, + 0.006990585941821337, + 0.00803125835955143, + 0.013139246962964535, + 0.02321542054414749, + 0.010313333943486214, + 0.010333914309740067, + 0.011349765583872795, + 0.021578220650553703, + 0.00535156624391675, + 0.0032179555855691433, + 0.006107802037149668, + 0.004762142430990934, + 0.0035212768707424402, + 0.0036835793871432543, + 0.005275072529911995, + 0.005041909869760275, + 0.037785645574331284, + 0.025666672736406326, + 0.052236203104257584, + 0.04480067268013954, + 0.03873676806688309, + 0.03707832843065262, + 0.0499686673283577, + 0.03527557849884033, + 0.02783939056098461, + 0.016595901921391487, + 0.04039113596081734, + 0.016153182834386826, + 0.02237711288034916, + 0.03276420384645462, + 0.021880270913243294, + 0.0615164153277874, + 0.008063865825533867, + 0.0052437023259699345, + 0.008471596986055374, + 0.00642413692548871, + 0.0063912211917340755, + 0.006885239388793707, + 0.009510677307844162, + 0.0069743916392326355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.0024900403805077076 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking close-up, captured from a slightly upward angle, depicts the camera in a slow, steady clockwise motion, circling the majestic face of a massive Buddha statue within the Yungang Grottoes in Datong. In stark black and white, light slants in from an unseen direction, carving deep shadows across the statue's features. Its compassionate and serene expression appears even more enduring, a testament to the weathering of millennia. Countless tiny grains of sandstone, whipped up by the strong winds outside the caves, form a series of gray, dynamic curtains that relentlessly impact the statue's face. As the camera pans, viewers can clearly see a small flake of sandstone, buffeted by the wind, lift slightly from the statue's weathered cheek, only to break away from the subject in the next instant, swept away by the whirling dust. This flaking process is captured with pinpoint precision. The statue's once smooth stone surface is now covered in fine cracks and pits. The lines of the eyelids and nose, while still discernible, have been worn away by the edges. The entire picture is full of Robert Capa's documentary sense and tension, condensing the grand theme of time's ruthless erosion of immortal things into this frame of visual narrative full of conflict between dynamic (flying sand) and static (Buddha statue), conveying a sense of historical heaviness and a heartbreaking fragility.", + "seed": 27404, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_0/index_9202_seed_27404_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17138029967900364, + "vlm_logprob": -0.17138029967900364, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.17138029967900364, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2943209111690521 + ], + "opposite_logprob_tokens": [ + -1.4193209409713745 + ], + "logprob_reward": -0.2943209111690521, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2943209111690521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12087848037481308, + 0.04142973572015762, + 0.05464464798569679, + 0.03773248940706253, + 0.025549788028001785, + 0.18136480450630188, + 0.1592807024717331, + 0.16271495819091797, + 0.06799465417861938, + 0.14840969443321228 + ], + "credit_map": [ + 0.012252368032932281, + 0.01020719762891531, + 0.012895996682345867, + 0.021451599895954132, + 0.009914785623550415, + 0.020816203206777573, + 0.013147068209946156, + 0.02019326388835907, + 0.005473581608384848, + 0.00311972270719707, + 0.00762674305588007, + 0.0025248487945646048, + 0.0027800763491541147, + 0.007906540296971798, + 0.0030547657515853643, + 0.008943459019064903, + 0.007284577935934067, + 0.004120876546949148, + 0.006390112917870283, + 0.009126514196395874, + 0.00502996938303113, + 0.00828337762504816, + 0.005682454910129309, + 0.008726765401661396, + 0.006189853418618441, + 0.0025203388649970293, + 0.010197745636105537, + 0.001985260983929038, + 0.002158600138500333, + 0.0066173868253827095, + 0.001687792711891234, + 0.006375513039529324, + 0.003095506224781275, + 0.0016470561968162656, + 0.004017003811895847, + 0.0016260079573839903, + 0.0020608408376574516, + 0.00536241102963686, + 0.0023854293394833803, + 0.0053555346094071865, + 0.009360969066619873, + 0.004485455807298422, + 0.01340891420841217, + 0.006536795292049646, + 0.00811824295669794, + 0.06334119290113449, + 0.011603781953454018, + 0.06450946629047394, + 0.007670090068131685, + 0.0040677497163414955, + 0.010785150341689587, + 0.004657875280827284, + 0.006862734444439411, + 0.059220943599939346, + 0.009621351025998592, + 0.056394826620817184, + 0.02262577973306179, + 0.02120848186314106, + 0.022383790463209152, + 0.014515258371829987, + 0.019248116761446, + 0.02350534126162529, + 0.014109566807746887, + 0.025118624791502953, + 0.009496481157839298, + 0.007690673228353262, + 0.011060895398259163, + 0.00758862029761076, + 0.00496656633913517, + 0.011727890931069851, + 0.0048492224887013435, + 0.01061430387198925, + 0.02089601755142212, + 0.019903816282749176, + 0.022222021594643593, + 0.013171250000596046, + 0.018319811671972275, + 0.01988067477941513, + 0.012358319014310837, + 0.021657781675457954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.2943209111690521 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004690003115683794 + ], + "opposite_logprob_tokens": [ + -5.379690170288086 + ], + "logprob_reward": -0.004690003115683794, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004690003115683794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17121148109436035, + 0.07245777547359467, + 0.18500526249408722, + 0.14297707378864288, + 0.03026917763054371, + 0.09712335467338562, + 0.10239461809396744, + 0.07495353370904922, + 0.07328486442565918, + 0.050322938710451126 + ], + "credit_map": [ + 0.025934429839253426, + 0.015486335381865501, + 0.021781357005238533, + 0.020388174802064896, + 0.011114276014268398, + 0.032758306711912155, + 0.013781954534351826, + 0.029966630041599274, + 0.013620559126138687, + 0.005305691622197628, + 0.01617598719894886, + 0.004346342291682959, + 0.004278899170458317, + 0.01255561038851738, + 0.005048890132457018, + 0.011125780642032623, + 0.03397882357239723, + 0.012546318583190441, + 0.03521442413330078, + 0.012527319602668285, + 0.010142838582396507, + 0.03383122384548187, + 0.009625880047678947, + 0.03713841736316681, + 0.024706289172172546, + 0.010784310288727283, + 0.04638312757015228, + 0.005711016245186329, + 0.007052076049149036, + 0.024014299735426903, + 0.0048775989562273026, + 0.019448349252343178, + 0.005728370975703001, + 0.0018381497357040644, + 0.006693804636597633, + 0.0014673108235001564, + 0.0018459060229361057, + 0.005544522777199745, + 0.0020756705198436975, + 0.005075439345091581, + 0.005654304753988981, + 0.0024041959550231695, + 0.0053821103647351265, + 0.003151559503749013, + 0.00280004576779902, + 0.03555700182914734, + 0.004136130679398775, + 0.0380379892885685, + 0.017800670117139816, + 0.006978880148380995, + 0.020135635510087013, + 0.005877593997865915, + 0.005995933897793293, + 0.02013120800256729, + 0.005677421577274799, + 0.01979726552963257, + 0.012836641632020473, + 0.008454900234937668, + 0.012406373396515846, + 0.0055758035741746426, + 0.007284329738467932, + 0.011076269671320915, + 0.005631619598716497, + 0.011687583290040493, + 0.010753877460956573, + 0.008594309911131859, + 0.010112577117979527, + 0.009066595695912838, + 0.005769597366452217, + 0.012118329294025898, + 0.006062052678316832, + 0.010807517915964127, + 0.007195568177849054, + 0.007641281001269817, + 0.008159655146300793, + 0.004572998266667128, + 0.006026464980095625, + 0.0067531089298427105, + 0.003785109845921397, + 0.0061887456104159355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.004690003115683794 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704312801361084 + ], + "opposite_logprob_tokens": [ + -2.1270432472229004 + ], + "logprob_reward": -0.12704312801361084, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704312801361084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07120309770107269, + 0.03999650478363037, + 0.03308563679456711, + 0.036149438470602036, + 0.0541292242705822, + 0.24763794243335724, + 0.1866922825574875, + 0.1984785944223404, + 0.06836728006601334, + 0.06426002830266953 + ], + "credit_map": [ + 0.007344831246882677, + 0.006425164174288511, + 0.008715365082025528, + 0.006737867370247841, + 0.005573386792093515, + 0.014410258270800114, + 0.006746930535882711, + 0.01524929516017437, + 0.005205804016441107, + 0.002611681818962097, + 0.008694818243384361, + 0.0032618108671158552, + 0.00302365655079484, + 0.0058946553617715836, + 0.00498723890632391, + 0.006316840182989836, + 0.0031241110991686583, + 0.002844694070518017, + 0.00381006789393723, + 0.0068160551600158215, + 0.0031701060943305492, + 0.004345988854765892, + 0.004586012102663517, + 0.0043886019848287106, + 0.0056350152008235455, + 0.00396180571988225, + 0.008111556991934776, + 0.0024237330071628094, + 0.0029737346339970827, + 0.006289559416472912, + 0.0018740348750725389, + 0.004879997577518225, + 0.008469616994261742, + 0.00513390963897109, + 0.00782304909080267, + 0.003521176753565669, + 0.007280111778527498, + 0.008912644349038601, + 0.0038441051729023457, + 0.009144609794020653, + 0.007277799770236015, + 0.0029591333586722612, + 0.007832609117031097, + 0.0030799845699220896, + 0.004530827049165964, + 0.10621471703052521, + 0.009774588979780674, + 0.10596828162670135, + 0.012160764075815678, + 0.014520109631121159, + 0.011464860290288925, + 0.011882412247359753, + 0.012756304815411568, + 0.0597776435315609, + 0.011111365631222725, + 0.053018808364868164, + 0.026949642226099968, + 0.024440258741378784, + 0.02521171234548092, + 0.0193069726228714, + 0.024351341649889946, + 0.02902834303677082, + 0.01944749988615513, + 0.02974281646311283, + 0.011104228906333447, + 0.00923745334148407, + 0.00996297039091587, + 0.007940092124044895, + 0.005268157925456762, + 0.009634272195398808, + 0.00453439075499773, + 0.010685714893043041, + 0.008980385027825832, + 0.007195927202701569, + 0.010555801913142204, + 0.005329235922545195, + 0.008092742413282394, + 0.009489276446402073, + 0.005590247455984354, + 0.009026412852108479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.12704312801361084 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289051592350006 + ], + "opposite_logprob_tokens": [ + -1.0539051294326782 + ], + "logprob_reward": -0.4289051592350006, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289051592350006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12554167211055756, + 0.10381956398487091, + 0.046112287789583206, + 0.05924653634428978, + 0.07828734070062637, + 0.148359015583992, + 0.10971929877996445, + 0.13044457137584686, + 0.05442383512854576, + 0.144045889377594 + ], + "credit_map": [ + 0.016174910590052605, + 0.011738664470613003, + 0.018366362899541855, + 0.012081338092684746, + 0.01022140122950077, + 0.020761657506227493, + 0.014389390125870705, + 0.021807938814163208, + 0.01389683410525322, + 0.006992730312049389, + 0.023219844326376915, + 0.006738970521837473, + 0.0059257205575704575, + 0.019773490726947784, + 0.006954892538487911, + 0.0203170794993639, + 0.005503264721482992, + 0.0036231728736311197, + 0.00755275646224618, + 0.005845516454428434, + 0.004127861466258764, + 0.007466578856110573, + 0.004591846838593483, + 0.007401293143630028, + 0.0076864794827997684, + 0.0060132225044071674, + 0.008882139809429646, + 0.0063240681774914265, + 0.005951086524873972, + 0.009165836498141289, + 0.005992815364152193, + 0.00923088751733303, + 0.00887120608240366, + 0.004195391200482845, + 0.009630377404391766, + 0.004302689805626869, + 0.006786658428609371, + 0.02194254845380783, + 0.006369810085743666, + 0.016188660636544228, + 0.010397347621619701, + 0.0038879266940057278, + 0.02230953611433506, + 0.004591815173625946, + 0.007918418385088444, + 0.04134224355220795, + 0.009237653575837612, + 0.04867406561970711, + 0.00661788834258914, + 0.0023142839781939983, + 0.009783625602722168, + 0.00300223333761096, + 0.0030386336147785187, + 0.041202377527952194, + 0.00444794213399291, + 0.03931231051683426, + 0.01925406977534294, + 0.014031085185706615, + 0.015840038657188416, + 0.00910203717648983, + 0.014354453422129154, + 0.022222256287932396, + 0.011058849282562733, + 0.02458178997039795, + 0.00835261307656765, + 0.004131598863750696, + 0.012509054504334927, + 0.0034370915964245796, + 0.0027523855678737164, + 0.009845462627708912, + 0.003497383091598749, + 0.009898246265947819, + 0.020806118845939636, + 0.017426878213882446, + 0.02446100302040577, + 0.01573563925921917, + 0.016487419605255127, + 0.01784270629286766, + 0.012972951866686344, + 0.01831318624317646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.4289051592350006 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Buddha statue located in the Yungang Grottoes in Datong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with deep shadows carved by slanting light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow, steady clockwise motion circling the statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small flake of sandstone visible lifting from the statue's cheek and breaking away due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the statue's surface covered in fine cracks and pits with worn-away eyelid and nose lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019422968616709113 + ], + "opposite_logprob_tokens": [ + -6.251942157745361 + ], + "logprob_reward": -0.0019422968616709113, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019422968616709113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08651474863290787, + 0.06858514994382858, + 0.030053630471229553, + 0.046653445810079575, + 0.02985510230064392, + 0.19540327787399292, + 0.09417286515235901, + 0.12041784077882767, + 0.15551699697971344, + 0.17282697558403015 + ], + "credit_map": [ + 0.00957681704312563, + 0.007141183130443096, + 0.011725801974534988, + 0.012531646527349949, + 0.005832009483128786, + 0.015067816711962223, + 0.009077800437808037, + 0.015561672858893871, + 0.011083485558629036, + 0.00474610086530447, + 0.018701724708080292, + 0.004191739484667778, + 0.003947558347135782, + 0.011689814738929272, + 0.0038301905151456594, + 0.01039453037083149, + 0.004543578252196312, + 0.0024988041259348392, + 0.005196605809032917, + 0.002439419738948345, + 0.002029295079410076, + 0.005138978362083435, + 0.0018535575363785028, + 0.006353389471769333, + 0.005633723456412554, + 0.0049613225273787975, + 0.006523767951875925, + 0.005288703367114067, + 0.004544537514448166, + 0.007059294730424881, + 0.004979681223630905, + 0.0076624187640845776, + 0.00536341080442071, + 0.0018840280827134848, + 0.006273521576076746, + 0.00127638247795403, + 0.0019065685337409377, + 0.006294082850217819, + 0.0017710963729768991, + 0.005086010787636042, + 0.013012314215302467, + 0.009739233180880547, + 0.02079075388610363, + 0.007204216904938221, + 0.013661602512001991, + 0.05331876873970032, + 0.016022246330976486, + 0.06165414676070213, + 0.004905834794044495, + 0.0017728721722960472, + 0.006132187787443399, + 0.002018195344135165, + 0.0015193689614534378, + 0.03821609169244766, + 0.0028546929825097322, + 0.03675362095236778, + 0.014644465409219265, + 0.013265417888760567, + 0.019056642428040504, + 0.00935125071555376, + 0.014157888479530811, + 0.020268764346837997, + 0.009393970482051373, + 0.020279444754123688, + 0.027553226798772812, + 0.016547225415706635, + 0.030155597254633904, + 0.013619212433695793, + 0.009184295311570168, + 0.026581978425383568, + 0.00840880535542965, + 0.023466652259230614, + 0.02796999365091324, + 0.012093611992895603, + 0.02998383902013302, + 0.012122630141675472, + 0.013131518848240376, + 0.029791636392474174, + 0.010891520418226719, + 0.03684224933385849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.0019422968616709113 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_1/index_tensor([1776, 1776])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_1/index_tensor([1776, 1776])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bc8de30a9529ba416ca01536ae0a7d78fedbaeba --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_1/index_tensor([1776, 1776])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a dramatic low-angle shot, a swift and aggressive advance of a powerful emperor, his entire form and world rendered in the stark, dynamic style of a black-and-white ink calligraphy painting. The emperor stands in a commanding pose, his form defined by bold, decisive ink strokes. He wears a voluminous, layered imperial robe, its heavy folds suggested by varying shades of gray ink wash painting, and intricate dragon motifs artfully suggested by flowing calligraphic lines. His face, chiseled and stern, is a mask of cold rage, piercing eyes set beneath an intricate, ornate crown painted with sharp lines. As the camera draws closer, he decisively extends his right arm forward, extending his hand. From the center of his open palm, a burst of pure energy erupts—not light, but a dazzling white void that fiercely repels the ink. This explosion causes the surrounding ink to splatter outward from the epicenter in a chaotic yet beautiful splatter, with tendrils and droplets of black ink scattering in all directions like a liquid explosion, creating a powerful and artistic spectacle against a textured off-white background reminiscent of rice paper.", + "seed": 1890, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_1/index_1776_seed_1890_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.016323746275156738, + "vlm_logprob": -0.016323746275156738, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.016323746275156738, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00827384926378727 + ], + "opposite_logprob_tokens": [ + -7.383274078369141 + ], + "logprob_reward": -0.00827384926378727, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.00827384926378727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14926770329475403, + 0.06253424286842346, + 0.07772481441497803, + 0.04351242631673813, + 0.1037387028336525, + 0.04852021858096123, + 0.06637214869260788, + 0.08457746356725693, + 0.1756347417831421, + 0.18811750411987305 + ], + "credit_map": [ + 0.0229896679520607, + 0.010631725192070007, + 0.025120161473751068, + 0.012092133983969688, + 0.01469385251402855, + 0.02333446964621544, + 0.020375512540340424, + 0.0200301893055439, + 0.014779864810407162, + 0.005538382567465305, + 0.008902882225811481, + 0.003891324158757925, + 0.005487029906362295, + 0.007812595926225185, + 0.008041346445679665, + 0.008080814965069294, + 0.010762427933514118, + 0.00884095299988985, + 0.009819306433200836, + 0.006293167360126972, + 0.009592634625732899, + 0.009997392073273659, + 0.012775968760251999, + 0.009642967022955418, + 0.006366346497088671, + 0.004086633212864399, + 0.006403082516044378, + 0.0039055137895047665, + 0.005627644248306751, + 0.006337611936032772, + 0.007191809359937906, + 0.0035937915090471506, + 0.012640878558158875, + 0.009890400804579258, + 0.018828360363841057, + 0.015210127457976341, + 0.01574484072625637, + 0.014735188335180283, + 0.00970651488751173, + 0.006982401479035616, + 0.006208270322531462, + 0.008058998733758926, + 0.008232034742832184, + 0.006126148160547018, + 0.005015839356929064, + 0.00567198172211647, + 0.004765007179230452, + 0.0044419411569833755, + 0.0093306303024292, + 0.009442565031349659, + 0.009713415987789631, + 0.012269312515854836, + 0.0075049977749586105, + 0.006916123908013105, + 0.0049950433894991875, + 0.006200066767632961, + 0.01075634267181158, + 0.010944189503788948, + 0.011430254206061363, + 0.025488249957561493, + 0.005564255639910698, + 0.007101130206137896, + 0.007600948680192232, + 0.00569209735840559, + 0.026155633851885796, + 0.023155663162469864, + 0.023919828236103058, + 0.035690560936927795, + 0.015172290615737438, + 0.015267026610672474, + 0.02422979474067688, + 0.012043953873217106, + 0.03053640015423298, + 0.015404599718749523, + 0.014231722801923752, + 0.021994348615407944, + 0.014777947217226028, + 0.02000603638589382, + 0.020549291744828224, + 0.05061716213822365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.00827384926378727 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008652695454657078 + ], + "opposite_logprob_tokens": [ + -4.758652687072754 + ], + "logprob_reward": -0.008652695454657078, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008652695454657078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3483494818210602, + 0.040794916450977325, + 0.11956045031547546, + 0.11033201217651367, + 0.06799618899822235, + 0.08830465376377106, + 0.033211275935173035, + 0.060360681265592575, + 0.05004431679844856, + 0.08104606717824936 + ], + "credit_map": [ + 0.04057883098721504, + 0.017460057511925697, + 0.08342043310403824, + 0.02026261016726494, + 0.02658676542341709, + 0.10225266963243484, + 0.02678300067782402, + 0.031005090102553368, + 0.010022738017141819, + 0.00403190590441227, + 0.0066558546386659145, + 0.0021423164289444685, + 0.0033524318132549524, + 0.0054299901239573956, + 0.0038584547583013773, + 0.005301220342516899, + 0.018504548817873, + 0.014245841652154922, + 0.018688948825001717, + 0.009229066781699657, + 0.017350204288959503, + 0.016365038231015205, + 0.009568968787789345, + 0.015607824549078941, + 0.015392299741506577, + 0.01259058341383934, + 0.018581457436084747, + 0.011653888039290905, + 0.016455497592687607, + 0.014588470570743084, + 0.012700186111032963, + 0.008369616232812405, + 0.008296984247863293, + 0.0064888219349086285, + 0.010365556925535202, + 0.010748784989118576, + 0.00959569402039051, + 0.009830078110098839, + 0.008544180542230606, + 0.0041260854341089725, + 0.012651794590055943, + 0.013151884078979492, + 0.02532108500599861, + 0.009317600168287754, + 0.008491755463182926, + 0.0068900627084076405, + 0.008330811746418476, + 0.0041496530175209045, + 0.004364044405519962, + 0.005526794120669365, + 0.004804491531103849, + 0.004933786578476429, + 0.004653468262404203, + 0.0035806065425276756, + 0.002886057598516345, + 0.002462024800479412, + 0.005373798310756683, + 0.004260578192770481, + 0.018373414874076843, + 0.0184590145945549, + 0.002898834878578782, + 0.0038609106559306383, + 0.0037343825679272413, + 0.0033997430000454187, + 0.006854168139398098, + 0.0072262948378920555, + 0.007461973931640387, + 0.008319503627717495, + 0.00382784241810441, + 0.004339087288826704, + 0.00608676066622138, + 0.00592868123203516, + 0.011343700811266899, + 0.0068552070297300816, + 0.006020914763212204, + 0.008609970100224018, + 0.00831262581050396, + 0.00850842334330082, + 0.009856010787189007, + 0.02153920754790306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.008652695454657078 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263621024787426 + ], + "opposite_logprob_tokens": [ + -3.651362180709839 + ], + "logprob_reward": -0.0263621024787426, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263621024787426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16568423807621002, + 0.05249634385108948, + 0.12082965672016144, + 0.12878459692001343, + 0.10325150936841965, + 0.13088183104991913, + 0.0662904754281044, + 0.09263112396001816, + 0.07028317451477051, + 0.0688670352101326 + ], + "credit_map": [ + 0.024091703817248344, + 0.007597310934215784, + 0.03903894126415253, + 0.009108442813158035, + 0.009327526204288006, + 0.04602082073688507, + 0.012523330748081207, + 0.017976170405745506, + 0.0100714648142457, + 0.004800593946129084, + 0.007582901045680046, + 0.00339370034635067, + 0.007184745278209448, + 0.0075172195211052895, + 0.006062053609639406, + 0.005883669015020132, + 0.017555654048919678, + 0.015494630672037601, + 0.020043572410941124, + 0.007729537785053253, + 0.01876622438430786, + 0.012964434921741486, + 0.014827389270067215, + 0.013448222540318966, + 0.018577562645077705, + 0.0179812703281641, + 0.018053360283374786, + 0.009859787300229073, + 0.021418534219264984, + 0.022670354694128036, + 0.01185720693320036, + 0.008366523310542107, + 0.012482727877795696, + 0.011934987269341946, + 0.014048712328076363, + 0.016803078353405, + 0.013274029828608036, + 0.018334919586777687, + 0.010119766928255558, + 0.0062532974407076836, + 0.01771872118115425, + 0.01914372108876705, + 0.037331245839595795, + 0.014089374803006649, + 0.014204666018486023, + 0.00994722731411457, + 0.012313748709857464, + 0.006133134011179209, + 0.01139003038406372, + 0.01042494922876358, + 0.011637859046459198, + 0.01164107583463192, + 0.007990372367203236, + 0.005723319016396999, + 0.004108990076929331, + 0.0033738838974386454, + 0.0064631132408976555, + 0.009450319223105907, + 0.0276260357350111, + 0.03214561939239502, + 0.004511241801083088, + 0.005274321883916855, + 0.0039952900260686874, + 0.0031651894096285105, + 0.009959192015230656, + 0.006694680079817772, + 0.012539424002170563, + 0.014008153229951859, + 0.006318786647170782, + 0.006568145006895065, + 0.008694939315319061, + 0.005499856546521187, + 0.009568675421178341, + 0.008900230750441551, + 0.004979588557034731, + 0.009183568879961967, + 0.008338377811014652, + 0.006245272699743509, + 0.008286078460514545, + 0.013365252874791622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0263621024787426 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033689916133880615 + ], + "opposite_logprob_tokens": [ + -3.4086899757385254 + ], + "logprob_reward": -0.033689916133880615, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033689916133880615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13806167244911194, + 0.07666697353124619, + 0.12032578885555267, + 0.13521546125411987, + 0.11493868380784988, + 0.09555017948150635, + 0.0603090263903141, + 0.1385912448167801, + 0.07941631972789764, + 0.04092458635568619 + ], + "credit_map": [ + 0.030552765354514122, + 0.010126778855919838, + 0.03538689389824867, + 0.007496583741158247, + 0.010274630039930344, + 0.023769019171595573, + 0.008999060839414597, + 0.011455953121185303, + 0.01461031474173069, + 0.005681714043021202, + 0.010206341743469238, + 0.006255298852920532, + 0.008983350358903408, + 0.009109938517212868, + 0.009961630217730999, + 0.011858389712870121, + 0.018785426393151283, + 0.012745378538966179, + 0.013624976389110088, + 0.009427055716514587, + 0.013212655670940876, + 0.026897165924310684, + 0.011897115968167782, + 0.013736016117036343, + 0.0164177343249321, + 0.012707791291177273, + 0.015382174402475357, + 0.013184759765863419, + 0.016854947432875633, + 0.020642194896936417, + 0.02517644502222538, + 0.01484941691160202, + 0.013467393815517426, + 0.010287960059940815, + 0.026540126651525497, + 0.01390921138226986, + 0.015129034407436848, + 0.01743234135210514, + 0.011593780480325222, + 0.006578835658729076, + 0.011815561912953854, + 0.015338258817791939, + 0.02039262093603611, + 0.013266095891594887, + 0.010310775600373745, + 0.01095336489379406, + 0.008363411761820316, + 0.005110087804496288, + 0.0102365892380476, + 0.008860073983669281, + 0.0070329755544662476, + 0.014029138721525669, + 0.005839082878082991, + 0.006126145366579294, + 0.004493975546211004, + 0.00369104347191751, + 0.009994206950068474, + 0.012717613950371742, + 0.03667789697647095, + 0.049814559519290924, + 0.006271042861044407, + 0.007865766994655132, + 0.010224903002381325, + 0.005025265738368034, + 0.01049654558300972, + 0.008290935307741165, + 0.010711976327002048, + 0.013360981829464436, + 0.008282781578600407, + 0.007245857268571854, + 0.012306473217904568, + 0.00872077513486147, + 0.004923999309539795, + 0.004614047706127167, + 0.0032062670215964317, + 0.0054389433935284615, + 0.004305141977965832, + 0.003390044905245304, + 0.004431437700986862, + 0.010614700615406036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.033689916133880615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464016804471612 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.00464016804471612, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464016804471612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12624922394752502, + 0.02455078810453415, + 0.03305567800998688, + 0.03205954283475876, + 0.09372445195913315, + 0.053083084523677826, + 0.13823308050632477, + 0.14195041358470917, + 0.22651170194149017, + 0.13058209419250488 + ], + "credit_map": [ + 0.02212219312787056, + 0.010695725679397583, + 0.034729793667793274, + 0.006246018689125776, + 0.009506519883871078, + 0.02624610625207424, + 0.008203942328691483, + 0.00849891547113657, + 0.006012995261698961, + 0.0019513852894306183, + 0.005086899269372225, + 0.0014821906806901097, + 0.0017270926618948579, + 0.0034388420172035694, + 0.0021490224171429873, + 0.0027023619040846825, + 0.005378290079534054, + 0.003063797252252698, + 0.004615599289536476, + 0.0023617991246283054, + 0.0046399743296206, + 0.007106202654540539, + 0.0027079633437097073, + 0.0031820505391806364, + 0.0043074022978544235, + 0.004619000945240259, + 0.0028042448684573174, + 0.0031823799945414066, + 0.00443944102153182, + 0.004575528670102358, + 0.006110755261033773, + 0.002020788611844182, + 0.011107201687991619, + 0.008318714797496796, + 0.017378466203808784, + 0.012115651741623878, + 0.013931834138929844, + 0.01317007839679718, + 0.011414583772420883, + 0.00628791656345129, + 0.005811640992760658, + 0.0072548771277070045, + 0.009173225611448288, + 0.007165732327848673, + 0.006672822870314121, + 0.006129254121333361, + 0.007694297004491091, + 0.003181236330419779, + 0.026333408430218697, + 0.017082897946238518, + 0.012686260975897312, + 0.044105105102062225, + 0.011972957290709019, + 0.014523321762681007, + 0.00771544361487031, + 0.0038136891089379787, + 0.017254246398806572, + 0.012368520721793175, + 0.040643807500600815, + 0.03307076543569565, + 0.006783056538552046, + 0.013345573097467422, + 0.01081536803394556, + 0.007669069338589907, + 0.03703729063272476, + 0.02758650667965412, + 0.03616618737578392, + 0.05370611697435379, + 0.01497621275484562, + 0.020528381690382957, + 0.023163726553320885, + 0.013347268104553223, + 0.016594072803854942, + 0.013278024271130562, + 0.011124971322715282, + 0.02868626080453396, + 0.010364784859120846, + 0.012210747227072716, + 0.013048242777585983, + 0.025274986401200294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.00464016804471612 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a dramatic low-angle shot, a swift and aggressive advance of a powerful emperor, his entire form and world rendered in the stark, dynamic style of a black-and-white ink calligraphy painting. The emperor stands in a commanding pose, his form defined by bold, decisive ink strokes. He wears a voluminous, layered imperial robe, its heavy folds suggested by varying shades of gray ink wash painting, and intricate dragon motifs artfully suggested by flowing calligraphic lines. His face, chiseled and stern, is a mask of cold rage, piercing eyes set beneath an intricate, ornate crown painted with sharp lines. As the camera draws closer, he decisively extends his right arm forward, extending his hand. From the center of his open palm, a burst of pure energy erupts—not light, but a dazzling white void that fiercely repels the ink. This explosion causes the surrounding ink to splatter outward from the epicenter in a chaotic yet beautiful splatter, with tendrils and droplets of black ink scattering in all directions like a liquid explosion, creating a powerful and artistic spectacle against a textured off-white background reminiscent of rice paper.", + "seed": 1891, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_1/index_1776_seed_1891_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016588253527879716, + "vlm_logprob": -0.016588253527879716, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.016588253527879716, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006075248587876558 + ], + "opposite_logprob_tokens": [ + -7.131075382232666 + ], + "logprob_reward": -0.006075248587876558, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.006075248587876558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1712789684534073, + 0.09793629497289658, + 0.04414278268814087, + 0.07090172916650772, + 0.07322473078966141, + 0.06281103938817978, + 0.3238786458969116, + 0.04650678113102913, + 0.04897542670369148, + 0.06034364178776741 + ], + "credit_map": [ + 0.025605402886867523, + 0.0212883111089468, + 0.03720736876130104, + 0.013158450834453106, + 0.014405323192477226, + 0.020467441529035568, + 0.016371432691812515, + 0.022775214165449142, + 0.015005994588136673, + 0.015352009795606136, + 0.02062571980059147, + 0.009476042352616787, + 0.00996086560189724, + 0.012512268498539925, + 0.00732832495123148, + 0.0076750596053898335, + 0.007384962867945433, + 0.0037508413661271334, + 0.008369662798941135, + 0.00519937090575695, + 0.004723364487290382, + 0.0058953287079930305, + 0.004218137823045254, + 0.004601112101227045, + 0.011475932784378529, + 0.004506664350628853, + 0.013376033864915371, + 0.004016426391899586, + 0.004523534793406725, + 0.02250271663069725, + 0.004300112836062908, + 0.006200301926583052, + 0.013064638711512089, + 0.007986688055098057, + 0.01506273914128542, + 0.007260187529027462, + 0.007417870219796896, + 0.006293703801929951, + 0.008463875390589237, + 0.007675022352486849, + 0.013842136599123478, + 0.007614825386554003, + 0.012739107944071293, + 0.009351246058940887, + 0.004417912103235722, + 0.005827088840305805, + 0.005179920699447393, + 0.0038387898821383715, + 0.07551174610853195, + 0.04884488880634308, + 0.039570778608322144, + 0.03813566640019417, + 0.023210974410176277, + 0.043376002460718155, + 0.028377912938594818, + 0.02685065194964409, + 0.00951336044818163, + 0.004764670971781015, + 0.00833259429782629, + 0.003366759279742837, + 0.00519774341955781, + 0.00565967196598649, + 0.0038462337106466293, + 0.005825742147862911, + 0.010441956110298634, + 0.004810838028788567, + 0.00685952790081501, + 0.005410597659647465, + 0.005368947517126799, + 0.005818123463541269, + 0.004020595923066139, + 0.006244834512472153, + 0.01103418692946434, + 0.013406029902398586, + 0.011060746386647224, + 0.004329145886003971, + 0.006841293536126614, + 0.005786133464425802, + 0.003498987527564168, + 0.004387114197015762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.006075248587876558 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645982291549444 + ], + "opposite_logprob_tokens": [ + -5.379645824432373 + ], + "logprob_reward": -0.004645982291549444, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645982291549444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11181685328483582, + 0.051514655351638794, + 0.0499890111386776, + 0.041763823479413986, + 0.03765270486474037, + 0.016994081437587738, + 0.529586672782898, + 0.07492183893918991, + 0.023231206461787224, + 0.06252916157245636 + ], + "credit_map": [ + 0.017305176705121994, + 0.017247717827558517, + 0.03553589805960655, + 0.0064417896792292595, + 0.005836622789502144, + 0.018035363405942917, + 0.005697818007320166, + 0.005716461222618818, + 0.008413579314947128, + 0.007486541755497456, + 0.0076878871768713, + 0.005114400759339333, + 0.005916799884289503, + 0.006355670280754566, + 0.004874084610491991, + 0.005665686912834644, + 0.007612336426973343, + 0.007658974267542362, + 0.00786144845187664, + 0.006682239472866058, + 0.0045253438875079155, + 0.007264676038175821, + 0.004301904700696468, + 0.004082088358700275, + 0.006667953450232744, + 0.005109047517180443, + 0.006234671920537949, + 0.004072645679116249, + 0.0039417012594640255, + 0.006940681021660566, + 0.0036727851256728172, + 0.00512433797121048, + 0.006279587745666504, + 0.005141827743500471, + 0.010197381488978863, + 0.0045481836423277855, + 0.0022283049765974283, + 0.003120161360129714, + 0.0028615386690944433, + 0.003275718539953232, + 0.002310666022822261, + 0.002207651035860181, + 0.002493388019502163, + 0.002269771881401539, + 0.002260434441268444, + 0.001793865580111742, + 0.0020306548103690147, + 0.001627650810405612, + 0.13181811571121216, + 0.07729648053646088, + 0.06890764832496643, + 0.056105777621269226, + 0.03500758856534958, + 0.07031387090682983, + 0.04612646996974945, + 0.04401072487235069, + 0.02037729136645794, + 0.009795659221708775, + 0.011859974823892117, + 0.003525660140439868, + 0.006448531989008188, + 0.008762748911976814, + 0.004122538026422262, + 0.01002943329513073, + 0.003803364234045148, + 0.0031097624450922012, + 0.0030116585548967123, + 0.002365944441407919, + 0.0035861192736774683, + 0.002713496331125498, + 0.0022948661353439093, + 0.0023459945805370808, + 0.011419309303164482, + 0.007838385179638863, + 0.00965844839811325, + 0.004500563256442547, + 0.009178879670798779, + 0.007006446365267038, + 0.0068903579376637936, + 0.006036781240254641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.004645982291549444 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642767556011677 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.008642767556011677, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642767556011677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12054786831140518, + 0.035145316272974014, + 0.04152657836675644, + 0.04574132710695267, + 0.05199827253818512, + 0.04038742557168007, + 0.49541178345680237, + 0.07841160148382187, + 0.027457326650619507, + 0.06337250024080276 + ], + "credit_map": [ + 0.018070362508296967, + 0.017257066443562508, + 0.035968367010354996, + 0.0072111631743609905, + 0.008130326867103577, + 0.01707634888589382, + 0.00845701526850462, + 0.008377224206924438, + 0.006879217457026243, + 0.004614613484591246, + 0.007850932888686657, + 0.00370771624147892, + 0.0023673176765441895, + 0.004996982403099537, + 0.0024127569049596786, + 0.002315778750926256, + 0.006570098455995321, + 0.004290242213755846, + 0.005011099856346846, + 0.004135434050112963, + 0.006340974010527134, + 0.006276998668909073, + 0.0034924380015581846, + 0.005409290548413992, + 0.00720691354945302, + 0.0031578857451677322, + 0.008191319182515144, + 0.003880303120240569, + 0.0034831208176910877, + 0.008453425951302052, + 0.0036070444621145725, + 0.007761313579976559, + 0.007676985580474138, + 0.005698158871382475, + 0.008714229799807072, + 0.007722045294940472, + 0.008739032782614231, + 0.00425085611641407, + 0.0051322211511433125, + 0.0040647410787642, + 0.005453488323837519, + 0.010116071440279484, + 0.005353196524083614, + 0.005544299725443125, + 0.004128691274672747, + 0.004444472957402468, + 0.00328142405487597, + 0.0020657784771174192, + 0.12376949936151505, + 0.07330232858657837, + 0.06443984061479568, + 0.053201865404844284, + 0.032607629895210266, + 0.06394042819738388, + 0.043948546051979065, + 0.040201667696237564, + 0.01975644752383232, + 0.009376851841807365, + 0.01200927421450615, + 0.004601052030920982, + 0.008517655543982983, + 0.00954519771039486, + 0.006282585207372904, + 0.00832253135740757, + 0.00594563502818346, + 0.0041666640900075436, + 0.004117709118872881, + 0.0018492004601284862, + 0.00327504426240921, + 0.0034890936221927404, + 0.0020058229565620422, + 0.002608155831694603, + 0.01093686930835247, + 0.008573785424232483, + 0.011335092596709728, + 0.005039475858211517, + 0.008449745364487171, + 0.008308298885822296, + 0.005161657929420471, + 0.005567572545260191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.008642767556011677 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494052916765213 + ], + "opposite_logprob_tokens": [ + -2.9299404621124268 + ], + "logprob_reward": -0.05494052916765213, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494052916765213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11683982610702515, + 0.049916625022888184, + 0.04041709005832672, + 0.08154429495334625, + 0.07085129618644714, + 0.02462853491306305, + 0.49915608763694763, + 0.08573233336210251, + 0.01334923505783081, + 0.017564721405506134 + ], + "credit_map": [ + 0.013630392029881477, + 0.01489793136715889, + 0.02622111141681671, + 0.009760953485965729, + 0.011225500144064426, + 0.012710093520581722, + 0.012172137387096882, + 0.01622169278562069, + 0.008047187700867653, + 0.006240025628358126, + 0.011431626975536346, + 0.004517148248851299, + 0.0036885777954012156, + 0.005674323532730341, + 0.0049743810668587685, + 0.00534334871917963, + 0.006101603619754314, + 0.004116694908589125, + 0.007093746680766344, + 0.0036770300939679146, + 0.0027833671774715185, + 0.007913177832961082, + 0.0034572617150843143, + 0.005274202674627304, + 0.019366856664419174, + 0.007245600689202547, + 0.012752148322761059, + 0.005456879734992981, + 0.00758003443479538, + 0.014546920545399189, + 0.0051494622603058815, + 0.009446392767131329, + 0.0151780154556036, + 0.00897513423115015, + 0.02237783558666706, + 0.006095198914408684, + 0.004619987215846777, + 0.005124291870743036, + 0.004356015007942915, + 0.004124813713133335, + 0.004801448434591293, + 0.0030652384739369154, + 0.004539296962320805, + 0.0025607445277273655, + 0.0028595596086233854, + 0.002624282380566001, + 0.0023187657352536917, + 0.001859195763245225, + 0.12256019562482834, + 0.07397429645061493, + 0.0641571655869484, + 0.053525760769844055, + 0.033376291394233704, + 0.06628264486789703, + 0.04194766655564308, + 0.043332070112228394, + 0.022274591028690338, + 0.01387090515345335, + 0.013234544545412064, + 0.0048141819424927235, + 0.008246442303061485, + 0.011752798222005367, + 0.004434268921613693, + 0.007104591466486454, + 0.003415633924305439, + 0.001198279787786305, + 0.0025942365173250437, + 0.0009164128568954766, + 0.001535336603410542, + 0.0014377875486388803, + 0.0011031933827325702, + 0.001148353097960353, + 0.0028667012229561806, + 0.0024223830550909042, + 0.0029202031437307596, + 0.0014674188569188118, + 0.0022848795633763075, + 0.0018098077271133661, + 0.002072500064969063, + 0.0017208266071975231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.05494052916765213 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636740036308765 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636740036308765, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636740036308765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08941938728094101, + 0.04578137397766113, + 0.047718968242406845, + 0.036214057356119156, + 0.037483397871255875, + 0.05247437208890915, + 0.36993882060050964, + 0.2517068684101105, + 0.02351413480937481, + 0.045748647302389145 + ], + "credit_map": [ + 0.01246761716902256, + 0.013616153970360756, + 0.02841973304748535, + 0.005679610185325146, + 0.004723358899354935, + 0.013304825872182846, + 0.004817354027181864, + 0.006390733644366264, + 0.006740341428667307, + 0.0067480700090527534, + 0.009417201392352581, + 0.004668355453759432, + 0.003188020084053278, + 0.005807948764413595, + 0.0047032698057591915, + 0.0045081693679094315, + 0.007415682543069124, + 0.005812132265418768, + 0.008673605509102345, + 0.006296275649219751, + 0.004109190311282873, + 0.00745990127325058, + 0.0038203070871531963, + 0.004131875466555357, + 0.0066511486656963825, + 0.0025406735949218273, + 0.005350242834538221, + 0.0019920868799090385, + 0.003034055233001709, + 0.008734773844480515, + 0.0028147948905825615, + 0.0050962804816663265, + 0.008093427866697311, + 0.005333422217518091, + 0.011910942383110523, + 0.002800823887810111, + 0.0024137538857758045, + 0.0028950294945389032, + 0.002309231786057353, + 0.0017267639050260186, + 0.00961071252822876, + 0.0065243602730333805, + 0.009585272520780563, + 0.009284298866987228, + 0.005036898422986269, + 0.00491481414064765, + 0.0042037758976221085, + 0.0033142438624054193, + 0.09407896548509598, + 0.05447642132639885, + 0.04773664101958275, + 0.03941449522972107, + 0.02386663667857647, + 0.04660090059041977, + 0.03248902037739754, + 0.03127577528357506, + 0.05638711154460907, + 0.04051220044493675, + 0.037795841693878174, + 0.01787322759628296, + 0.025458430871367455, + 0.03896113485097885, + 0.013091230764985085, + 0.021627679467201233, + 0.005406295880675316, + 0.00259809079580009, + 0.004121196456253529, + 0.001815910218283534, + 0.0022636447101831436, + 0.0031666294671595097, + 0.0018544255290180445, + 0.002287941752001643, + 0.007637430913746357, + 0.012372901663184166, + 0.009383520111441612, + 0.0029603398870676756, + 0.00465195020660758, + 0.0026983588468283415, + 0.0031377184204757214, + 0.0029064237605780363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.008636740036308765 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a dramatic low-angle shot, a swift and aggressive advance of a powerful emperor, his entire form and world rendered in the stark, dynamic style of a black-and-white ink calligraphy painting. The emperor stands in a commanding pose, his form defined by bold, decisive ink strokes. He wears a voluminous, layered imperial robe, its heavy folds suggested by varying shades of gray ink wash painting, and intricate dragon motifs artfully suggested by flowing calligraphic lines. His face, chiseled and stern, is a mask of cold rage, piercing eyes set beneath an intricate, ornate crown painted with sharp lines. As the camera draws closer, he decisively extends his right arm forward, extending his hand. From the center of his open palm, a burst of pure energy erupts—not light, but a dazzling white void that fiercely repels the ink. This explosion causes the surrounding ink to splatter outward from the epicenter in a chaotic yet beautiful splatter, with tendrils and droplets of black ink scattering in all directions like a liquid explosion, creating a powerful and artistic spectacle against a textured off-white background reminiscent of rice paper.", + "seed": 1892, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_1/index_1776_seed_1892_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01197740128263831, + "vlm_logprob": -0.01197740128263831, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.01197740128263831, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009457310661673546 + ], + "opposite_logprob_tokens": [ + -7.134457111358643 + ], + "logprob_reward": -0.009457310661673546, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.009457310661673546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16374081373214722, + 0.0648220032453537, + 0.06656473875045776, + 0.08670662343502045, + 0.0970604419708252, + 0.10654336959123611, + 0.09944899380207062, + 0.057540543377399445, + 0.05530380830168724, + 0.20226868987083435 + ], + "credit_map": [ + 0.03907756507396698, + 0.01226208545267582, + 0.02063828893005848, + 0.022678466513752937, + 0.014342989772558212, + 0.020308321341872215, + 0.013894704170525074, + 0.02053837664425373, + 0.013989447616040707, + 0.005103532690554857, + 0.009438718669116497, + 0.006437883246690035, + 0.007477245293557644, + 0.009786226786673069, + 0.004995359107851982, + 0.007593591697514057, + 0.011731237173080444, + 0.005643295589834452, + 0.014097858220338821, + 0.005039115436375141, + 0.005661905277520418, + 0.011449200101196766, + 0.005024379119277, + 0.007917741313576698, + 0.01725645363330841, + 0.006431412883102894, + 0.021982770413160324, + 0.0063849245198071, + 0.006348073482513428, + 0.012788979336619377, + 0.0076371668837964535, + 0.007876839488744736, + 0.019709328189492226, + 0.008512768894433975, + 0.020658712834119797, + 0.006590978242456913, + 0.008876393549144268, + 0.01217679027467966, + 0.011198140680789948, + 0.009337327443063259, + 0.01286640577018261, + 0.016839567571878433, + 0.014956320635974407, + 0.01164854597300291, + 0.011076530441641808, + 0.010682997293770313, + 0.02118200622498989, + 0.007290998473763466, + 0.012135827913880348, + 0.013581164181232452, + 0.01355291623622179, + 0.011421243660151958, + 0.012179981917142868, + 0.0136424470692873, + 0.016096103936433792, + 0.006839308887720108, + 0.00661683501675725, + 0.007329357787966728, + 0.004951069131493568, + 0.01112858671694994, + 0.007372603751718998, + 0.00923373643308878, + 0.006783265620470047, + 0.004125090315937996, + 0.008725271560251713, + 0.0074602775275707245, + 0.006737800315022469, + 0.005359445698559284, + 0.007216614205390215, + 0.005677533335983753, + 0.00794206652790308, + 0.006184800062328577, + 0.027998875826597214, + 0.023387420922517776, + 0.028258802369236946, + 0.016359811648726463, + 0.04670606553554535, + 0.016049211844801903, + 0.026820579543709755, + 0.0166879091411829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.009457310661673546 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648913353681564 + ], + "opposite_logprob_tokens": [ + -4.758648872375488 + ], + "logprob_reward": -0.008648913353681564, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648913353681564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3082644045352936, + 0.11650952696800232, + 0.10577042400836945, + 0.029009243473410606, + 0.12132006138563156, + 0.07491067051887512, + 0.04621507227420807, + 0.05468003824353218, + 0.05910295248031616, + 0.08421759307384491 + ], + "credit_map": [ + 0.04105919599533081, + 0.030490271747112274, + 0.04327875003218651, + 0.036187153309583664, + 0.04071792960166931, + 0.04204948991537094, + 0.03405135124921799, + 0.04043028503656387, + 0.018539777025580406, + 0.00925090629607439, + 0.014076050370931625, + 0.013642470352351665, + 0.014894715510308743, + 0.01619355008006096, + 0.013489097356796265, + 0.01642296463251114, + 0.011399684473872185, + 0.009098583832383156, + 0.01615048199892044, + 0.012676719576120377, + 0.016084250062704086, + 0.01457861065864563, + 0.014286011457443237, + 0.011496087536215782, + 0.0054740263149142265, + 0.0026586486492305994, + 0.004328215029090643, + 0.002014943165704608, + 0.002814497333019972, + 0.0044713872484862804, + 0.00251022563315928, + 0.00473730219528079, + 0.025361420586705208, + 0.011212369427084923, + 0.02622160315513611, + 0.007986203767359257, + 0.012966113165020943, + 0.015691151842474937, + 0.011959224008023739, + 0.009921968914568424, + 0.009245941415429115, + 0.01675346866250038, + 0.013805384747684002, + 0.005622823256999254, + 0.009243195876479149, + 0.007734067738056183, + 0.007590167690068483, + 0.004915622528642416, + 0.00805457029491663, + 0.006045266054570675, + 0.005338484421372414, + 0.0070513044483959675, + 0.005286952015012503, + 0.005683331284672022, + 0.005674847401678562, + 0.0030803184490650892, + 0.006853050552308559, + 0.007481704466044903, + 0.006286499090492725, + 0.00721442885696888, + 0.008158069103956223, + 0.007851538248360157, + 0.006532623432576656, + 0.004302123095840216, + 0.008124292828142643, + 0.007042057812213898, + 0.006515525281429291, + 0.006657266523689032, + 0.010903709568083286, + 0.004193980246782303, + 0.009134391322731972, + 0.00653172517195344, + 0.01597602292895317, + 0.007955036126077175, + 0.013115397654473782, + 0.005113963037729263, + 0.013020597398281097, + 0.007262601982802153, + 0.01271107979118824, + 0.009062889963388443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.008648913353681564 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012552070431411266 + ], + "opposite_logprob_tokens": [ + -4.387552261352539 + ], + "logprob_reward": -0.012552070431411266, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012552070431411266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19444310665130615, + 0.10336105525493622, + 0.14401714503765106, + 0.025047926232218742, + 0.17659203708171844, + 0.06861972063779831, + 0.13953611254692078, + 0.03856261074542999, + 0.051032986491918564, + 0.05878734216094017 + ], + "credit_map": [ + 0.025135435163974762, + 0.01894586905837059, + 0.027156364172697067, + 0.025388479232788086, + 0.023247769102454185, + 0.026079261675477028, + 0.02403852343559265, + 0.024451415985822678, + 0.017250608652830124, + 0.007460791151970625, + 0.012372966855764389, + 0.010653303004801273, + 0.011859619058668613, + 0.015013065189123154, + 0.013526271097362041, + 0.015224429778754711, + 0.013561681844294071, + 0.0120222307741642, + 0.019461119547486305, + 0.016682656481862068, + 0.02090374380350113, + 0.020783133804798126, + 0.02231227606534958, + 0.018290311098098755, + 0.0038191629573702812, + 0.0029240348376333714, + 0.0033252094872295856, + 0.0022345450706779957, + 0.003809735644608736, + 0.003321235766634345, + 0.0023383507505059242, + 0.003275652416050434, + 0.03569137677550316, + 0.017212722450494766, + 0.035292014479637146, + 0.012363648973405361, + 0.019495472311973572, + 0.021904630586504936, + 0.020556876435875893, + 0.014075294137001038, + 0.009365508332848549, + 0.014816689305007458, + 0.00873681902885437, + 0.006209813989698887, + 0.00880809873342514, + 0.008278725668787956, + 0.007767565548419952, + 0.004636493511497974, + 0.01697968877851963, + 0.017504818737506866, + 0.013828184455633163, + 0.03127572312951088, + 0.021188082173466682, + 0.015536690130829811, + 0.01466827280819416, + 0.008554652333259583, + 0.003946816548705101, + 0.005058052483946085, + 0.0031661821994930506, + 0.007246027234941721, + 0.005999402143061161, + 0.0043737757951021194, + 0.005669623613357544, + 0.0031027323566377163, + 0.007847930304706097, + 0.005981120280921459, + 0.0058611915446817875, + 0.005759791471064091, + 0.008243261836469173, + 0.0041809710673987865, + 0.007716076914221048, + 0.005442637484520674, + 0.010395274497568607, + 0.006007254123687744, + 0.007054754067212343, + 0.005238696932792664, + 0.011201147921383381, + 0.003848334075883031, + 0.00775721063837409, + 0.007284670602530241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.012552070431411266 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023281319066882133 + ], + "opposite_logprob_tokens": [ + -3.7732813358306885 + ], + "logprob_reward": -0.023281319066882133, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023281319066882133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09047648310661316, + 0.07349492609500885, + 0.05034145340323448, + 0.07682954519987106, + 0.16282068192958832, + 0.11374679952859879, + 0.3131001591682434, + 0.030571753159165382, + 0.0240593533962965, + 0.06455889344215393 + ], + "credit_map": [ + 0.015485009178519249, + 0.005749179981648922, + 0.011284427717328072, + 0.01411350630223751, + 0.008059514686465263, + 0.010580691508948803, + 0.010722023434937, + 0.014482132159173489, + 0.016800956800580025, + 0.0057347603142261505, + 0.010136986151337624, + 0.007944596000015736, + 0.007856212556362152, + 0.009661379270255566, + 0.006961315404623747, + 0.008398712612688541, + 0.0065643396228551865, + 0.004659730941057205, + 0.007740773260593414, + 0.0046303123235702515, + 0.005653184838593006, + 0.00811796449124813, + 0.00574580067768693, + 0.007229343988001347, + 0.011748646385967731, + 0.008421399630606174, + 0.017512964084744453, + 0.007116473279893398, + 0.005656798370182514, + 0.011029359884560108, + 0.006301890593022108, + 0.009042012505233288, + 0.03517034649848938, + 0.016668323427438736, + 0.03299863636493683, + 0.012139590457081795, + 0.015915460884571075, + 0.02045251987874508, + 0.016674712300300598, + 0.012801106087863445, + 0.016745036467909813, + 0.013070200569927692, + 0.0164501816034317, + 0.016945401206612587, + 0.012023068964481354, + 0.011480329558253288, + 0.014346524141728878, + 0.012686061672866344, + 0.0456516407430172, + 0.040834102779626846, + 0.032591547816991806, + 0.06864410638809204, + 0.04058265686035156, + 0.03030105121433735, + 0.031266987323760986, + 0.023228073492646217, + 0.003688905853778124, + 0.004663008730858564, + 0.0034336233511567116, + 0.004706108942627907, + 0.0040239556692540646, + 0.0036781129892915487, + 0.004007730633020401, + 0.002370305825024843, + 0.003669719910249114, + 0.004128590226173401, + 0.003175697522237897, + 0.002847552765160799, + 0.0025948372203856707, + 0.002249682554975152, + 0.0034019146114587784, + 0.001991358818486333, + 0.011660867370665073, + 0.005633233580738306, + 0.009701837785542011, + 0.004122097510844469, + 0.010805079713463783, + 0.005991286598145962, + 0.009150338359177113, + 0.007494151592254639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.023281319066882133 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947392899543047 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947392899543047, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947392899543047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1616756021976471, + 0.0601196214556694, + 0.05235682427883148, + 0.054033078253269196, + 0.0997670516371727, + 0.21498897671699524, + 0.1336614340543747, + 0.11571715027093887, + 0.05595226213335991, + 0.051727984100580215 + ], + "credit_map": [ + 0.031695760786533356, + 0.010695490054786205, + 0.02094462513923645, + 0.01724577695131302, + 0.018449600785970688, + 0.01639048382639885, + 0.018499532714486122, + 0.02775432914495468, + 0.01100454106926918, + 0.004265936091542244, + 0.008888044394552708, + 0.006663896609097719, + 0.005900857504457235, + 0.008210857398808002, + 0.006484003737568855, + 0.00870148278772831, + 0.006440762896090746, + 0.0047923182137310505, + 0.00830734334886074, + 0.006448853760957718, + 0.005843381863087416, + 0.007441740483045578, + 0.007255233824253082, + 0.005827190820127726, + 0.00869556050747633, + 0.006379888858646154, + 0.010219545103609562, + 0.006008348893374205, + 0.003981690388172865, + 0.007354832720011473, + 0.0045744264498353004, + 0.00681878300383687, + 0.018801746889948845, + 0.01011890172958374, + 0.019412659108638763, + 0.007393138017505407, + 0.00956003274768591, + 0.012238282710313797, + 0.01390816830098629, + 0.008334120735526085, + 0.022876445204019547, + 0.04324851557612419, + 0.03241200000047684, + 0.025529813021421432, + 0.023607805371284485, + 0.02098485641181469, + 0.03216388449072838, + 0.014165668748319149, + 0.015691787004470825, + 0.018390163779258728, + 0.015493923798203468, + 0.029740968719124794, + 0.01856403239071369, + 0.012089123018085957, + 0.014090071432292461, + 0.009601357392966747, + 0.012637226842343807, + 0.01683378778398037, + 0.01025365013629198, + 0.018378300592303276, + 0.015557621605694294, + 0.015045673586428165, + 0.01856840029358864, + 0.00844249501824379, + 0.008575942367315292, + 0.007265003398060799, + 0.008835647255182266, + 0.007056743372231722, + 0.008149722591042519, + 0.0053687854669988155, + 0.005721771623939276, + 0.004978641867637634, + 0.008846850134432316, + 0.00480890367180109, + 0.006271198391914368, + 0.004810845013707876, + 0.009519774466753006, + 0.00501563074067235, + 0.0072927167639136314, + 0.005162063054740429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.005947392899543047 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a dramatic low-angle shot, a swift and aggressive advance of a powerful emperor, his entire form and world rendered in the stark, dynamic style of a black-and-white ink calligraphy painting. The emperor stands in a commanding pose, his form defined by bold, decisive ink strokes. He wears a voluminous, layered imperial robe, its heavy folds suggested by varying shades of gray ink wash painting, and intricate dragon motifs artfully suggested by flowing calligraphic lines. His face, chiseled and stern, is a mask of cold rage, piercing eyes set beneath an intricate, ornate crown painted with sharp lines. As the camera draws closer, he decisively extends his right arm forward, extending his hand. From the center of his open palm, a burst of pure energy erupts—not light, but a dazzling white void that fiercely repels the ink. This explosion causes the surrounding ink to splatter outward from the epicenter in a chaotic yet beautiful splatter, with tendrils and droplets of black ink scattering in all directions like a liquid explosion, creating a powerful and artistic spectacle against a textured off-white background reminiscent of rice paper.", + "seed": 1893, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_1/index_1776_seed_1893_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01690060468390584, + "vlm_logprob": -0.01690060468390584, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.01690060468390584, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013355790637433529 + ], + "opposite_logprob_tokens": [ + -7.138355731964111 + ], + "logprob_reward": -0.013355790637433529, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.013355790637433529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10287018865346909, + 0.11174037307500839, + 0.060725681483745575, + 0.07671679556369781, + 0.06753958761692047, + 0.23516276478767395, + 0.18839417397975922, + 0.024765130132436752, + 0.045584820210933685, + 0.08650054037570953 + ], + "credit_map": [ + 0.0190875306725502, + 0.01243620179593563, + 0.01844569481909275, + 0.011838716454803944, + 0.009108849801123142, + 0.012473689392209053, + 0.008587096817791462, + 0.010892395861446857, + 0.01736300066113472, + 0.011001096107065678, + 0.025432555004954338, + 0.012882656417787075, + 0.010797474533319473, + 0.01712234690785408, + 0.008935121819376945, + 0.008206119760870934, + 0.00937225203961134, + 0.006527554709464312, + 0.013062220998108387, + 0.00661997776478529, + 0.0063945879228413105, + 0.0067811147309839725, + 0.005578826181590557, + 0.006389141082763672, + 0.013074182905256748, + 0.008938536047935486, + 0.013243090361356735, + 0.010747930034995079, + 0.007135557010769844, + 0.009292022325098515, + 0.007038448005914688, + 0.0072470298036932945, + 0.00897219404578209, + 0.008937191218137741, + 0.006320254411548376, + 0.0058457376435399055, + 0.010334869846701622, + 0.010386106558144093, + 0.007886387407779694, + 0.008856834843754768, + 0.03131701797246933, + 0.03121836669743061, + 0.03783351555466652, + 0.026960458606481552, + 0.03145328536629677, + 0.02824237011373043, + 0.02290996164083481, + 0.02522776648402214, + 0.029837489128112793, + 0.016893746331334114, + 0.029190413653850555, + 0.020731419324874878, + 0.016888141632080078, + 0.02248879335820675, + 0.021494871005415916, + 0.03086928278207779, + 0.004383191931992769, + 0.0024532503448426723, + 0.00408515240997076, + 0.001950700068846345, + 0.0017849362920969725, + 0.003977760206907988, + 0.0018158701714128256, + 0.004314265679568052, + 0.007028762251138687, + 0.006415351293981075, + 0.010334241203963757, + 0.0031436930876225233, + 0.004819305147975683, + 0.004678326193243265, + 0.00310160918161273, + 0.006063527427613735, + 0.012501812539994717, + 0.012152004055678844, + 0.014878036454319954, + 0.011125997640192509, + 0.006489895284175873, + 0.010788524523377419, + 0.009140846319496632, + 0.00942341610789299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.013355790637433529 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260791163891554 + ], + "opposite_logprob_tokens": [ + -5.255260944366455 + ], + "logprob_reward": -0.005260791163891554, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260791163891554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4661877155303955, + 0.0656876266002655, + 0.10969667136669159, + 0.062058109790086746, + 0.02745072729885578, + 0.034793421626091, + 0.08134005218744278, + 0.025997724384069443, + 0.07301010936498642, + 0.05377788096666336 + ], + "credit_map": [ + 0.07627888768911362, + 0.07598838955163956, + 0.06859879940748215, + 0.037779442965984344, + 0.06431996077299118, + 0.07392558455467224, + 0.03190715238451958, + 0.03738949075341225, + 0.011693194508552551, + 0.005948038771748543, + 0.008658629842102528, + 0.00447545200586319, + 0.005994376726448536, + 0.015253105200827122, + 0.0047743734903633595, + 0.00889044813811779, + 0.013706132769584656, + 0.013107677921652794, + 0.012531704269349575, + 0.010010994039475918, + 0.009485374204814434, + 0.022659970447421074, + 0.010503152385354042, + 0.017691660672426224, + 0.007538218051195145, + 0.0060301111079752445, + 0.011875173076987267, + 0.011191688477993011, + 0.004506430123001337, + 0.00641494058072567, + 0.008799288421869278, + 0.005702253896743059, + 0.004646996036171913, + 0.0024992183316498995, + 0.006584138609468937, + 0.002636154182255268, + 0.0022644889540970325, + 0.0022724734153598547, + 0.0025542660150676966, + 0.0039929901249706745, + 0.006470383610576391, + 0.003566802479326725, + 0.006894036661833525, + 0.0032025224063545465, + 0.0050217099487781525, + 0.0031984017696231604, + 0.0033988216891884804, + 0.00304074352607131, + 0.015905428677797318, + 0.005538115743547678, + 0.017780931666493416, + 0.005105046089738607, + 0.005029899068176746, + 0.006022325251251459, + 0.005543824750930071, + 0.020414479076862335, + 0.004552380181849003, + 0.002559974789619446, + 0.004271743819117546, + 0.002391314832493663, + 0.0020000177901238203, + 0.004440242424607277, + 0.0016515053575858474, + 0.004130542743951082, + 0.012325290590524673, + 0.013730794191360474, + 0.01043054647743702, + 0.004933805670589209, + 0.012997454963624477, + 0.005415778607130051, + 0.0037603983655571938, + 0.009416037239134312, + 0.0062182205729186535, + 0.008598760701715946, + 0.009814523160457611, + 0.005984670016914606, + 0.004528877791017294, + 0.006693484727293253, + 0.005683425348252058, + 0.006255915388464928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.005260791163891554 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644777350127697 + ], + "opposite_logprob_tokens": [ + -4.7586445808410645 + ], + "logprob_reward": -0.008644777350127697, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644777350127697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3695110082626343, + 0.04719100892543793, + 0.061326321214437485, + 0.10683553665876389, + 0.03228023275732994, + 0.10649775713682175, + 0.0802239403128624, + 0.047219764441251755, + 0.08616430312395096, + 0.06275004148483276 + ], + "credit_map": [ + 0.05716818943619728, + 0.06580370664596558, + 0.03656775876879692, + 0.024165375158190727, + 0.060055095702409744, + 0.060854721814394, + 0.02822595089673996, + 0.036670226603746414, + 0.007223740220069885, + 0.00624834792688489, + 0.008297884836792946, + 0.0037257540971040726, + 0.004900510422885418, + 0.009085290133953094, + 0.0030823713168501854, + 0.004627110902220011, + 0.009356117807328701, + 0.005487745627760887, + 0.012505535036325455, + 0.006032855715602636, + 0.005920022260397673, + 0.008503076620399952, + 0.004303519614040852, + 0.009217449463903904, + 0.013215111568570137, + 0.01281374879181385, + 0.019742479547858238, + 0.017534486949443817, + 0.009516128338873386, + 0.012276334688067436, + 0.012604737654328346, + 0.009132510051131248, + 0.005607918370515108, + 0.0033949336502701044, + 0.005654101725667715, + 0.002433685353025794, + 0.0029001710936427116, + 0.0035954290069639683, + 0.00315061560831964, + 0.005543379113078117, + 0.01796172931790352, + 0.008029998280107975, + 0.0076356930658221245, + 0.01152208261191845, + 0.029989417642354965, + 0.015408908948302269, + 0.006987325847148895, + 0.008962606079876423, + 0.013009010814130306, + 0.007502872962504625, + 0.014356762170791626, + 0.006087785121053457, + 0.00866682082414627, + 0.007465459872037172, + 0.006887731142342091, + 0.016247499734163284, + 0.007756592705845833, + 0.006667429581284523, + 0.005200992338359356, + 0.007777998223900795, + 0.004849367309361696, + 0.00518472446128726, + 0.003624474396929145, + 0.0061581870540976524, + 0.012107562273740768, + 0.01889760047197342, + 0.010249902494251728, + 0.00542259169742465, + 0.016706567257642746, + 0.006488335318863392, + 0.004836042411625385, + 0.011455698870122433, + 0.008307796902954578, + 0.01074234303086996, + 0.011094609275460243, + 0.007615480571985245, + 0.005541875958442688, + 0.005750696640461683, + 0.0056250328198075294, + 0.008072205819189548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.008644777350127697 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043051183223724365 + ], + "opposite_logprob_tokens": [ + -3.168051242828369 + ], + "logprob_reward": -0.043051183223724365, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043051183223724365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16687288880348206, + 0.056935880333185196, + 0.06280112266540527, + 0.09076786041259766, + 0.07727141678333282, + 0.2611517012119293, + 0.17787061631679535, + 0.017825588583946228, + 0.039193592965602875, + 0.049309391528367996 + ], + "credit_map": [ + 0.027076197788119316, + 0.025651050731539726, + 0.03211940824985504, + 0.014695516787469387, + 0.017429308965802193, + 0.023780642077326775, + 0.012268555350601673, + 0.013852188363671303, + 0.008795281872153282, + 0.010059279389679432, + 0.013439304195344448, + 0.004432598128914833, + 0.005366229452192783, + 0.0071744294837117195, + 0.0034498809836804867, + 0.004218869376927614, + 0.009020254947245121, + 0.007786208763718605, + 0.007310355547815561, + 0.005280110985040665, + 0.00574071891605854, + 0.012988165020942688, + 0.005496565252542496, + 0.009178736247122288, + 0.00995353702455759, + 0.01101876050233841, + 0.015468861907720566, + 0.015187868848443031, + 0.00870630145072937, + 0.009701637551188469, + 0.010966627858579159, + 0.009764265269041061, + 0.010180396027863026, + 0.010923492722213268, + 0.014777332544326782, + 0.005921980831772089, + 0.00936309527605772, + 0.009496549144387245, + 0.00871257483959198, + 0.00789598934352398, + 0.036804571747779846, + 0.03284597024321556, + 0.0385873056948185, + 0.026299212127923965, + 0.05276099219918251, + 0.03129967674612999, + 0.019728146493434906, + 0.022825825959444046, + 0.02908029593527317, + 0.014496127143502235, + 0.03149706497788429, + 0.019022604450583458, + 0.014167475514113903, + 0.020837049931287766, + 0.019321199506521225, + 0.029448777437210083, + 0.0033749465364962816, + 0.002032372634857893, + 0.0033286442048847675, + 0.00204993924126029, + 0.0009122469928115606, + 0.0019117859192192554, + 0.0011583395535126328, + 0.003057309193536639, + 0.005304241087287664, + 0.00735529325902462, + 0.003923672251403332, + 0.0036230175755918026, + 0.006972135975956917, + 0.005182568449527025, + 0.002231988124549389, + 0.004600674379616976, + 0.006096562370657921, + 0.007782204542309046, + 0.0053614056669175625, + 0.005937201902270317, + 0.008623052388429642, + 0.004418645054101944, + 0.005693934857845306, + 0.0053963796235620975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.043051183223724365 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a black-and-white ink calligraphy painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful emperor wearing a voluminous, layered imperial robe with dragon motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the emperor have a stern face with piercing eyes and an ornate crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and advance swiftly toward the emperor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the emperor extend his right arm forward to release a burst of white energy that causes black ink to splatter outward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190481044352055 + ], + "opposite_logprob_tokens": [ + -4.264190673828125 + ], + "logprob_reward": -0.014190481044352055, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190481044352055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13691726326942444, + 0.04111287742853165, + 0.09738905727863312, + 0.13186682760715485, + 0.13139592111110687, + 0.10381794720888138, + 0.07263734936714172, + 0.03115946054458618, + 0.15624076128005981, + 0.09746251255273819 + ], + "credit_map": [ + 0.02871861681342125, + 0.016681049019098282, + 0.03365880250930786, + 0.01117897778749466, + 0.009725203737616539, + 0.014302249066531658, + 0.010691113770008087, + 0.011961258016526699, + 0.006823377683758736, + 0.004628724418580532, + 0.00519845774397254, + 0.004288955125957727, + 0.004227424506098032, + 0.0060729156248271465, + 0.0033003955613821745, + 0.006572627928107977, + 0.01415465958416462, + 0.009868250228464603, + 0.017287474125623703, + 0.010438896715641022, + 0.008579286746680737, + 0.01507189217954874, + 0.008517072536051273, + 0.013471524231135845, + 0.017552070319652557, + 0.014113394543528557, + 0.025439543649554253, + 0.022306503728032112, + 0.011671836487948895, + 0.013877651654183865, + 0.014983617700636387, + 0.011922216974198818, + 0.019913144409656525, + 0.011829983443021774, + 0.015531733632087708, + 0.010615313425660133, + 0.01849183440208435, + 0.017580589279532433, + 0.01649654284119606, + 0.02093677967786789, + 0.020394423976540565, + 0.012386584654450417, + 0.018533682450652122, + 0.008142007514834404, + 0.00893763080239296, + 0.009897570125758648, + 0.008797915652394295, + 0.016728132963180542, + 0.012281767092645168, + 0.009545830078423023, + 0.011350437998771667, + 0.006250514183193445, + 0.005757721606642008, + 0.009904906153678894, + 0.00638387817889452, + 0.011162295937538147, + 0.005642611067742109, + 0.002707580104470253, + 0.005172728560864925, + 0.0033434501383453608, + 0.0016781255835667253, + 0.0037716254591941833, + 0.0023923150729388, + 0.006451025605201721, + 0.023209085687994957, + 0.02633940801024437, + 0.021336229518055916, + 0.012833657674491405, + 0.027230296283960342, + 0.015514284372329712, + 0.010970455594360828, + 0.018807346001267433, + 0.012559561990201473, + 0.016671383753418922, + 0.014756008982658386, + 0.012608966790139675, + 0.007407379802316427, + 0.007651867810636759, + 0.012557946145534515, + 0.013249398209154606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.014190481044352055 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_10/index_tensor([2350, 2350])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_10/index_tensor([2350, 2350])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3a95fc76abca4f683e0aa376110c418e33dd909b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_10/index_tensor([2350, 2350])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot moves parallel to a majestic old steam locomotive as it powerfully churns along tracks laid across the rolling countryside. The locomotive is pitch black, its surface stained by age and soot to a heavy metallic finish. Its massive drive wheels and complex linkages rotate with a steady, relentless rhythm. Atop the locomotive's chimney, thick, white steam billows backward, slowly dissipating in the soft evening air. Following closely behind are several passenger cars painted dark green, their brass-framed windows shimmering with warm light, hinting at the comfort within. As the wheels glide over the joints of the rails, each car sways slightly, rhythmically, side to side, a cadence unmistakable. The entire scene is bathed in the warm light of the golden hour, the setting sun gilding the steam clouds and casting long, flowing shadows across the train's sleek sides, creating an atmosphere of nostalgia, tranquility, and power.", + "seed": 27544, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_10/index_2350_seed_27544_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07409915756434202, + "vlm_logprob": -0.07409915756434202, + "vlm_yes_no_margin": 2.874999976158142, + "vlm_token_credit": -0.07409915756434202, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12943637371063232 + ], + "opposite_logprob_tokens": [ + -2.129436492919922 + ], + "logprob_reward": -0.12943637371063232, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12943637371063232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.278326153755188, + 0.09466568380594254, + 0.042649947106838226, + 0.05032564699649811, + 0.23835702240467072, + 0.05274803936481476, + 0.060298167169094086, + 0.03603816404938698, + 0.07736719399690628, + 0.06922409683465958 + ], + "credit_map": [ + 0.032560478895902634, + 0.0629352480173111, + 0.029497448354959488, + 0.03761116415262222, + 0.039702143520116806, + 0.022371212020516396, + 0.03219318017363548, + 0.021455250680446625, + 0.012061163783073425, + 0.01659589447081089, + 0.010260671377182007, + 0.009229145012795925, + 0.010948349721729755, + 0.011592351831495762, + 0.008550603874027729, + 0.01542749721556902, + 0.004887805320322514, + 0.0043569388799369335, + 0.003490124363452196, + 0.004512069746851921, + 0.007197823375463486, + 0.006332860793918371, + 0.004634635057300329, + 0.007237686775624752, + 0.005958300083875656, + 0.008396167308092117, + 0.006405944004654884, + 0.005215250421315432, + 0.009896932169795036, + 0.005838270299136639, + 0.003937833476811647, + 0.004676944576203823, + 0.02090744487941265, + 0.053514737635850906, + 0.029987022280693054, + 0.030086584389209747, + 0.022818537428975105, + 0.03235161677002907, + 0.025561833754181862, + 0.023129206150770187, + 0.0048525892198085785, + 0.007252905517816544, + 0.007659217808395624, + 0.00716192414984107, + 0.008035384118556976, + 0.005850065965205431, + 0.006213468499481678, + 0.005722483154386282, + 0.005475599318742752, + 0.007203879300504923, + 0.008997086435556412, + 0.008768546395003796, + 0.007010587025433779, + 0.007472730241715908, + 0.006899499800056219, + 0.00847023818641901, + 0.003634650958701968, + 0.0053540244698524475, + 0.006132315844297409, + 0.0050680432468652725, + 0.004525988828390837, + 0.0033345462288707495, + 0.0041295853443443775, + 0.003859007963910699, + 0.016546742990612984, + 0.008745054714381695, + 0.009612871333956718, + 0.007051042281091213, + 0.004979592747986317, + 0.0059984782710671425, + 0.008069857023656368, + 0.01636354625225067, + 0.009225385263562202, + 0.007351789623498917, + 0.005831506568938494, + 0.0069295247085392475, + 0.012622487731277943, + 0.010428999550640583, + 0.009103650227189064, + 0.007730746641755104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.12943637371063232 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209403656423092 + ], + "opposite_logprob_tokens": [ + -4.264209270477295 + ], + "logprob_reward": -0.014209403656423092, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209403656423092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048202499747276306, + 0.12404394894838333, + 0.03194967657327652, + 0.04114135727286339, + 0.14829157292842865, + 0.038232650607824326, + 0.11396816372871399, + 0.03182404860854149, + 0.22001485526561737, + 0.20233125984668732 + ], + "credit_map": [ + 0.005214024800807238, + 0.006124787963926792, + 0.006861845962703228, + 0.007992199622094631, + 0.005885242484509945, + 0.0042822351679205894, + 0.007602962199598551, + 0.00423920014873147, + 0.013166755437850952, + 0.017176926136016846, + 0.014449063688516617, + 0.0129480492323637, + 0.0176123958081007, + 0.01906285062432289, + 0.014127382077276707, + 0.015500519424676895, + 0.002961954101920128, + 0.002947953762486577, + 0.0028530091512948275, + 0.004631405230611563, + 0.004890941083431244, + 0.004491564817726612, + 0.0045393118634819984, + 0.00463353656232357, + 0.004808268044143915, + 0.0045906244777143, + 0.005030779633671045, + 0.006061296444386244, + 0.0066759102046489716, + 0.004902131389826536, + 0.003973802085965872, + 0.005098545923829079, + 0.013615651056170464, + 0.0291345976293087, + 0.023783179000020027, + 0.020891951397061348, + 0.01682012528181076, + 0.016253294423222542, + 0.013474415056407452, + 0.014318362809717655, + 0.0031923488713800907, + 0.0077118887566030025, + 0.00615334790199995, + 0.005419776309281588, + 0.004749588668346405, + 0.0030657367315143347, + 0.004174865782260895, + 0.003765095490962267, + 0.011338062584400177, + 0.017331117764115334, + 0.020744221284985542, + 0.01646735891699791, + 0.010919920168817043, + 0.010104386135935783, + 0.012487558647990227, + 0.014575524255633354, + 0.0035211127251386642, + 0.00432940199971199, + 0.004124788101762533, + 0.00444285711273551, + 0.0033112461678683758, + 0.003577491734176874, + 0.004253051243722439, + 0.004264097660779953, + 0.04263954237103462, + 0.02201981656253338, + 0.04303744435310364, + 0.0221241544932127, + 0.014927215874195099, + 0.0156928151845932, + 0.014651653356850147, + 0.044922199100255966, + 0.032885100692510605, + 0.04214019328355789, + 0.015659678727388382, + 0.017176436260342598, + 0.03587895631790161, + 0.02419751137495041, + 0.018719380721449852, + 0.015674004331231117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.014209403656423092 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001525908708572 + ], + "opposite_logprob_tokens": [ + -2.6950151920318604 + ], + "logprob_reward": -0.07001525908708572, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001525908708572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16481506824493408, + 0.1865922212600708, + 0.031872402876615524, + 0.05039087310433388, + 0.16093334555625916, + 0.056787628680467606, + 0.05298596993088722, + 0.054702311754226685, + 0.1487242430448532, + 0.09219592064619064 + ], + "credit_map": [ + 0.025308066979050636, + 0.01844114437699318, + 0.01932213269174099, + 0.025310853496193886, + 0.022675011307001114, + 0.015930084511637688, + 0.0247245691716671, + 0.013103206641972065, + 0.023567069321870804, + 0.025700777769088745, + 0.03081699274480343, + 0.020015934482216835, + 0.02449394389986992, + 0.023085998371243477, + 0.018491121008992195, + 0.020420391112565994, + 0.0038782660849392414, + 0.004102373495697975, + 0.003695826278999448, + 0.003392864717170596, + 0.005020883399993181, + 0.003382576396688819, + 0.0037451437674462795, + 0.004654469434171915, + 0.007025677245110273, + 0.006845494266599417, + 0.006227321922779083, + 0.006895580794662237, + 0.006773290224373341, + 0.005356966983526945, + 0.005734650883823633, + 0.005531893577426672, + 0.01594938337802887, + 0.04396599158644676, + 0.015727804973721504, + 0.024730512872338295, + 0.013175688683986664, + 0.010947629809379578, + 0.018798505887389183, + 0.017637820914387703, + 0.0060536921955645084, + 0.013871402479708195, + 0.009966766461730003, + 0.0070496792905032635, + 0.005690738093107939, + 0.0036704314406961203, + 0.005940889474004507, + 0.0045440285466611385, + 0.005895315203815699, + 0.013197346590459347, + 0.005431717727333307, + 0.00466095469892025, + 0.005233187694102526, + 0.006420150864869356, + 0.005332081578671932, + 0.006815216038376093, + 0.006182425189763308, + 0.00788118876516819, + 0.01013344619423151, + 0.008493587374687195, + 0.005127281416207552, + 0.006098050624132156, + 0.004822366405278444, + 0.005963967181742191, + 0.040437955409288406, + 0.01786091923713684, + 0.011635562404990196, + 0.010347390547394753, + 0.0112638920545578, + 0.01161457970738411, + 0.009576607495546341, + 0.03598733991384506, + 0.014404087327420712, + 0.017645591869950294, + 0.010740933939814568, + 0.010235331952571869, + 0.012553293257951736, + 0.011782907880842686, + 0.007200161460787058, + 0.007633612025529146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.07001525908708572 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029790684580802917 + ], + "opposite_logprob_tokens": [ + -3.5297906398773193 + ], + "logprob_reward": -0.029790684580802917, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029790684580802917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09880906343460083, + 0.15638436377048492, + 0.04838003218173981, + 0.028881195932626724, + 0.1364775151014328, + 0.09797300398349762, + 0.07841815054416656, + 0.08205404877662659, + 0.151864156126976, + 0.12075848132371902 + ], + "credit_map": [ + 0.013633202761411667, + 0.015550645999610424, + 0.013874361291527748, + 0.012292865663766861, + 0.012259037233889103, + 0.008729669265449047, + 0.011196659877896309, + 0.01127261109650135, + 0.018927354365587234, + 0.023369615897536278, + 0.023976512253284454, + 0.01896166242659092, + 0.016765698790550232, + 0.019210519269108772, + 0.01677156426012516, + 0.018401436507701874, + 0.006291121710091829, + 0.004796988796442747, + 0.005413152743130922, + 0.005548608023673296, + 0.006331960670650005, + 0.006655117496848106, + 0.0053403242491185665, + 0.008002761751413345, + 0.0030646505765616894, + 0.0039503080770373344, + 0.004687992390245199, + 0.004283358342945576, + 0.004271580837666988, + 0.003581025404855609, + 0.002394292503595352, + 0.0026479905936867, + 0.013005241751670837, + 0.041536469012498856, + 0.013579800724983215, + 0.018486471846699715, + 0.010122437961399555, + 0.014557607471942902, + 0.015455184504389763, + 0.009734299033880234, + 0.00963530782610178, + 0.021127576008439064, + 0.014472434297204018, + 0.01109450962394476, + 0.010562731884419918, + 0.008631748147308826, + 0.012723194435238838, + 0.009725511074066162, + 0.007182490546256304, + 0.00800962746143341, + 0.015910236164927483, + 0.009883243590593338, + 0.006188283208757639, + 0.008636527694761753, + 0.009718244895339012, + 0.012889508157968521, + 0.008955317549407482, + 0.013782047666609287, + 0.014327501878142357, + 0.012658528052270412, + 0.006924562621861696, + 0.008497456088662148, + 0.008024558424949646, + 0.008884084410965443, + 0.03398913890123367, + 0.019539356231689453, + 0.0213181059807539, + 0.013661334291100502, + 0.0075985402800142765, + 0.014953074976801872, + 0.008379848673939705, + 0.03242475911974907, + 0.018027497455477715, + 0.018489304929971695, + 0.015262591652572155, + 0.016390429809689522, + 0.014831701293587685, + 0.01529016438871622, + 0.010493622161448002, + 0.011973166838288307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 4 + ], + "selected_reward": -0.029790684580802917 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704406678676605 + ], + "opposite_logprob_tokens": [ + -2.1270439624786377 + ], + "logprob_reward": -0.12704406678676605, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12704406678676605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21547691524028778, + 0.09174228459596634, + 0.057215120643377304, + 0.062337033450603485, + 0.2200605720281601, + 0.048641324043273926, + 0.04901890829205513, + 0.1290445327758789, + 0.03194931894540787, + 0.09451394528150558 + ], + "credit_map": [ + 0.03635359928011894, + 0.03625904768705368, + 0.01668260619044304, + 0.022386543452739716, + 0.018351225182414055, + 0.0233040489256382, + 0.042615536600351334, + 0.019524313509464264, + 0.011688443832099438, + 0.014308740384876728, + 0.015999944880604744, + 0.010222868993878365, + 0.00947446096688509, + 0.007206854410469532, + 0.010620285756886005, + 0.012220690958201885, + 0.007814917713403702, + 0.006012772675603628, + 0.006219915580004454, + 0.00493838544934988, + 0.005853625014424324, + 0.007976680994033813, + 0.006346375215798616, + 0.012052450329065323, + 0.007856215350329876, + 0.008105509914457798, + 0.01018583495169878, + 0.009735419414937496, + 0.00883660651743412, + 0.005632568150758743, + 0.005620972253382206, + 0.006363906431943178, + 0.021821971982717514, + 0.04882141202688217, + 0.04112868010997772, + 0.038213133811950684, + 0.01638212986290455, + 0.025466272607445717, + 0.0150851896032691, + 0.013141771778464317, + 0.005182167049497366, + 0.01202799379825592, + 0.006473115645349026, + 0.006108021829277277, + 0.004261340014636517, + 0.003594023874029517, + 0.0051393150351941586, + 0.005855342373251915, + 0.0058124433271586895, + 0.010431824252009392, + 0.004175310488790274, + 0.0059561035595834255, + 0.006611381657421589, + 0.0050719380378723145, + 0.004991051275283098, + 0.005968853365629911, + 0.014146455563604832, + 0.015793826431035995, + 0.028330013155937195, + 0.020385103300213814, + 0.014007178135216236, + 0.011128869839012623, + 0.012561163865029812, + 0.012691912241280079, + 0.003289037849754095, + 0.0038688990753144026, + 0.005103475879877806, + 0.004118281416594982, + 0.0032132526393979788, + 0.004307441413402557, + 0.004146745894104242, + 0.0039021852426230907, + 0.014813021756708622, + 0.01140210498124361, + 0.00803537666797638, + 0.01210787519812584, + 0.016020681709051132, + 0.010420435108244419, + 0.012340291403234005, + 0.009374160319566727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.12704406678676605 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot moves parallel to a majestic old steam locomotive as it powerfully churns along tracks laid across the rolling countryside. The locomotive is pitch black, its surface stained by age and soot to a heavy metallic finish. Its massive drive wheels and complex linkages rotate with a steady, relentless rhythm. Atop the locomotive's chimney, thick, white steam billows backward, slowly dissipating in the soft evening air. Following closely behind are several passenger cars painted dark green, their brass-framed windows shimmering with warm light, hinting at the comfort within. As the wheels glide over the joints of the rails, each car sways slightly, rhythmically, side to side, a cadence unmistakable. The entire scene is bathed in the warm light of the golden hour, the setting sun gilding the steam clouds and casting long, flowing shadows across the train's sleek sides, creating an atmosphere of nostalgia, tranquility, and power.", + "seed": 27545, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_10/index_2350_seed_27545_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1292693704366684, + "vlm_logprob": -0.1292693704366684, + "vlm_yes_no_margin": 2.3, + "vlm_token_credit": -0.1292693704366684, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3182055652141571 + ], + "opposite_logprob_tokens": [ + -1.3182055950164795 + ], + "logprob_reward": -0.3182055652141571, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3182055652141571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2535473704338074, + 0.03272543475031853, + 0.13465261459350586, + 0.06554824858903885, + 0.08075917512178421, + 0.03948688134551048, + 0.09637569636106491, + 0.07285290211439133, + 0.16218262910842896, + 0.061869144439697266 + ], + "credit_map": [ + 0.039520662277936935, + 0.020301593467593193, + 0.04407750815153122, + 0.030811185017228127, + 0.022427471354603767, + 0.02862999029457569, + 0.02876790054142475, + 0.03901103138923645, + 0.003803838277235627, + 0.0036668803077191114, + 0.0041701048612594604, + 0.0044423737563192844, + 0.0029022477101534605, + 0.0038854952435940504, + 0.0035849306732416153, + 0.006269562989473343, + 0.01954520121216774, + 0.011761985719203949, + 0.01354169100522995, + 0.015124433673918247, + 0.020693078637123108, + 0.010887717828154564, + 0.01398863922804594, + 0.029109850525856018, + 0.010681664571166039, + 0.005863358732312918, + 0.008989176712930202, + 0.010273683816194534, + 0.0045687537640333176, + 0.0054905954748392105, + 0.01070625800639391, + 0.008974750526249409, + 0.006609149277210236, + 0.005135207902640104, + 0.011782047338783741, + 0.013521603308618069, + 0.006161856930702925, + 0.008262485265731812, + 0.012603494338691235, + 0.016683323308825493, + 0.0031794363167136908, + 0.004428364336490631, + 0.004537417087703943, + 0.00690269423648715, + 0.0034087770618498325, + 0.003274354850873351, + 0.007272738963365555, + 0.006483092438429594, + 0.010278056375682354, + 0.015065708197653294, + 0.009118102490901947, + 0.015146144665777683, + 0.007337802555412054, + 0.007733161095529795, + 0.017611518502235413, + 0.014085193164646626, + 0.00755372503772378, + 0.0062460084445774555, + 0.015418590977787971, + 0.009439739398658276, + 0.008400368504226208, + 0.006186258513480425, + 0.00980065856128931, + 0.009807540103793144, + 0.0211799293756485, + 0.02733134850859642, + 0.023789793252944946, + 0.018281487748026848, + 0.015830228105187416, + 0.01747058890759945, + 0.02508912980556488, + 0.013210097327828407, + 0.006617615465074778, + 0.0068151443265378475, + 0.008295197039842606, + 0.01159780751913786, + 0.0072156586684286594, + 0.00576344970613718, + 0.008120022714138031, + 0.007444237824529409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.3182055652141571 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023309968411922455 + ], + "opposite_logprob_tokens": [ + -3.7733099460601807 + ], + "logprob_reward": -0.023309968411922455, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023309968411922455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.103789322078228, + 0.026942746713757515, + 0.16429997980594635, + 0.16142000257968903, + 0.10592960566282272, + 0.10919562727212906, + 0.09940756112337112, + 0.12175263464450836, + 0.050011392682790756, + 0.05725118890404701 + ], + "credit_map": [ + 0.008977181278169155, + 0.009045255370438099, + 0.015176868997514248, + 0.015841186046600342, + 0.013795898295938969, + 0.009645490907132626, + 0.015774408355355263, + 0.0155330253764987, + 0.0022523202933371067, + 0.0020461571402847767, + 0.0032102956902235746, + 0.003752014832571149, + 0.003554668975993991, + 0.0030913271475583315, + 0.003802910214290023, + 0.0052330526523292065, + 0.021435394883155823, + 0.01233986672013998, + 0.020759355276823044, + 0.022909026592969894, + 0.023220401257276535, + 0.01590116322040558, + 0.01801620051264763, + 0.029718557372689247, + 0.025111928582191467, + 0.01214456744492054, + 0.026450881734490395, + 0.018965573981404305, + 0.011515443213284016, + 0.017371393740177155, + 0.027536163106560707, + 0.022324033081531525, + 0.007935757748782635, + 0.008655574172735214, + 0.017681041732430458, + 0.01819680444896221, + 0.010419500060379505, + 0.012181248515844345, + 0.016025464981794357, + 0.014834214001893997, + 0.009453004226088524, + 0.013437471352517605, + 0.013090063817799091, + 0.016317568719387054, + 0.008672202005982399, + 0.013048596680164337, + 0.014366740360856056, + 0.02080998197197914, + 0.010165377520024776, + 0.008213896304368973, + 0.013354046270251274, + 0.020618677139282227, + 0.0052632614970207214, + 0.006912491284310818, + 0.016745010390877724, + 0.018134798854589462, + 0.012060167267918587, + 0.011791246011853218, + 0.02056678757071495, + 0.018994873389601707, + 0.008530305698513985, + 0.013449911028146744, + 0.022989094257354736, + 0.013370252214372158, + 0.006779154762625694, + 0.011529766023159027, + 0.006103230640292168, + 0.00698039261624217, + 0.0050298357382416725, + 0.005714230239391327, + 0.004183205775916576, + 0.0036915745586156845, + 0.007057759910821915, + 0.0071137272752821445, + 0.008680431172251701, + 0.007390533573925495, + 0.005024414509534836, + 0.00884417723864317, + 0.006935952231287956, + 0.006204189732670784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.023309968411922455 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294416338205338 + ], + "opposite_logprob_tokens": [ + -2.2379441261291504 + ], + "logprob_reward": -0.11294416338205338, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294416338205338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24420417845249176, + 0.06330211460590363, + 0.06800523400306702, + 0.05577865242958069, + 0.06043248251080513, + 0.04493140056729317, + 0.04328838735818863, + 0.0638316348195076, + 0.30797845125198364, + 0.04824751615524292 + ], + "credit_map": [ + 0.01969737932085991, + 0.019891813397407532, + 0.02792687714099884, + 0.03874858096241951, + 0.021704772487282753, + 0.02425258420407772, + 0.040360935032367706, + 0.051621194928884506, + 0.00523109408095479, + 0.0065973298624157906, + 0.007886799052357674, + 0.010957498103380203, + 0.004720208700746298, + 0.005626836325973272, + 0.00928710401058197, + 0.012995238415896893, + 0.009255754761397839, + 0.005760509520769119, + 0.009027927182614803, + 0.006775465793907642, + 0.011289314366877079, + 0.0067005455493927, + 0.0063130841590464115, + 0.012882629409432411, + 0.0050451988354325294, + 0.004225455224514008, + 0.007867011241614819, + 0.011251569725573063, + 0.004334155935794115, + 0.004559989087283611, + 0.011541401036083698, + 0.006953864358365536, + 0.005389063619077206, + 0.004938414320349693, + 0.0090798893943429, + 0.009670067578554153, + 0.004639871418476105, + 0.006385812535881996, + 0.008408060297369957, + 0.011921296827495098, + 0.00398928951472044, + 0.004467722959816456, + 0.005273575894534588, + 0.00871625728905201, + 0.0034933944698423147, + 0.003756704507395625, + 0.008309520781040192, + 0.006924932356923819, + 0.004022559151053429, + 0.005732614081352949, + 0.005054878070950508, + 0.007684264797717333, + 0.0037249329034239054, + 0.0038665973115712404, + 0.006838884670287371, + 0.0063636512495577335, + 0.005941446404904127, + 0.006853094790130854, + 0.012221027165651321, + 0.008799592964351177, + 0.005673782434314489, + 0.004605497233569622, + 0.009935841895639896, + 0.009801344014704227, + 0.04474812000989914, + 0.0744568407535553, + 0.038616664707660675, + 0.03515703231096268, + 0.02704884111881256, + 0.020098157227039337, + 0.04213313013315201, + 0.02571963332593441, + 0.00623825378715992, + 0.008241899311542511, + 0.005783586297184229, + 0.00842259917408228, + 0.003679516725242138, + 0.003316165180876851, + 0.007126100827008486, + 0.00543938810005784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.11294416338205338 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896300405263901 + ], + "opposite_logprob_tokens": [ + -2.578963041305542 + ], + "logprob_reward": -0.07896300405263901, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896300405263901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09271159768104553, + 0.02204572968184948, + 0.3428308665752411, + 0.15805691480636597, + 0.048729319125413895, + 0.12301066517829895, + 0.027691205963492393, + 0.03565125912427902, + 0.1292736530303955, + 0.019998760893940926 + ], + "credit_map": [ + 0.011349746957421303, + 0.008164213970303535, + 0.012296877801418304, + 0.012116599828004837, + 0.008234778419137001, + 0.009100859053432941, + 0.01500679925084114, + 0.016441715881228447, + 0.002054905751720071, + 0.003389692399650812, + 0.002957054413855076, + 0.003887447062879801, + 0.002013458637520671, + 0.0022682901471853256, + 0.0026503349654376507, + 0.0028245446737855673, + 0.04454559087753296, + 0.034339938312768936, + 0.03337930142879486, + 0.039531808346509933, + 0.06526806950569153, + 0.030271122232079506, + 0.036955587565898895, + 0.05853944644331932, + 0.023613683879375458, + 0.015290537849068642, + 0.021247852593660355, + 0.027068613097071648, + 0.010422859340906143, + 0.015346427448093891, + 0.024440551176667213, + 0.020626390352845192, + 0.0036805197596549988, + 0.003373545128852129, + 0.006963832303881645, + 0.008519886061549187, + 0.004295816645026207, + 0.0047225747257471085, + 0.008625068701803684, + 0.008548079058527946, + 0.008755498565733433, + 0.014873262494802475, + 0.018259206786751747, + 0.02044183947145939, + 0.013980941846966743, + 0.012026580981910229, + 0.019174009561538696, + 0.015499324537813663, + 0.002164847683161497, + 0.003442836692556739, + 0.0034417102579027414, + 0.0045835645869374275, + 0.0022393029648810625, + 0.0027785857673734426, + 0.005755517166107893, + 0.0032848401460796595, + 0.004404651466757059, + 0.004015607759356499, + 0.007984270341694355, + 0.004253937862813473, + 0.0025264557916671038, + 0.0038537527434527874, + 0.0045901997946202755, + 0.004022381268441677, + 0.018250776454806328, + 0.026301948353648186, + 0.017009349539875984, + 0.01235897932201624, + 0.013073640875518322, + 0.011144903488457203, + 0.020163794979453087, + 0.010970261879265308, + 0.002155153313651681, + 0.002249176846817136, + 0.003054865635931492, + 0.0037855091504752636, + 0.0019227775046601892, + 0.0014063791604712605, + 0.0034229166340082884, + 0.002001982880756259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.07896300405263901 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292415112257004 + ], + "opposite_logprob_tokens": [ + -2.237924098968506 + ], + "logprob_reward": -0.11292415112257004, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292415112257004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29567989706993103, + 0.0603170320391655, + 0.14102013409137726, + 0.07928849756717682, + 0.2127191573381424, + 0.04743138700723648, + 0.05185309424996376, + 0.03430606424808502, + 0.03732838109135628, + 0.04005640372633934 + ], + "credit_map": [ + 0.032308075577020645, + 0.0375472791492939, + 0.034825313836336136, + 0.051234353333711624, + 0.030764365568757057, + 0.024328529834747314, + 0.04243502393364906, + 0.04223693534731865, + 0.007922722958028316, + 0.006605411414057016, + 0.007796929683536291, + 0.008344478905200958, + 0.003524373285472393, + 0.003705780953168869, + 0.006714164279401302, + 0.015703167766332626, + 0.017662813887000084, + 0.011578901670873165, + 0.015274343080818653, + 0.018026597797870636, + 0.019225571304559708, + 0.01440356019884348, + 0.017502879723906517, + 0.02734546549618244, + 0.006727392785251141, + 0.006103442516177893, + 0.010453423485159874, + 0.017230994999408722, + 0.005253853276371956, + 0.005187814123928547, + 0.017053132876753807, + 0.011278437450528145, + 0.015190141275525093, + 0.016349807381629944, + 0.02677404135465622, + 0.03585801273584366, + 0.02340961992740631, + 0.027682457119226456, + 0.03349015861749649, + 0.03396492451429367, + 0.005214831326156855, + 0.0037782234139740467, + 0.0059311059303581715, + 0.006557045970112085, + 0.003557338612154126, + 0.00499899685382843, + 0.006287891883403063, + 0.011105953715741634, + 0.004031576681882143, + 0.004253910388797522, + 0.004219725262373686, + 0.01362115889787674, + 0.002874641679227352, + 0.003859535325318575, + 0.010364186018705368, + 0.008628356270492077, + 0.004646160639822483, + 0.005205501802265644, + 0.006701023783534765, + 0.004041103180497885, + 0.003026986960321665, + 0.001959610730409622, + 0.003954125102609396, + 0.0047715529799461365, + 0.0052658431231975555, + 0.008008092641830444, + 0.0061242287047207355, + 0.004469504579901695, + 0.0030366487335413694, + 0.0035628208424896, + 0.0027152991387993097, + 0.004145939834415913, + 0.005365822464227676, + 0.004051432479172945, + 0.007983976975083351, + 0.005234560463577509, + 0.003080901922658086, + 0.0050106169655919075, + 0.004949342459440231, + 0.004379752092063427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.11292415112257004 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot moves parallel to a majestic old steam locomotive as it powerfully churns along tracks laid across the rolling countryside. The locomotive is pitch black, its surface stained by age and soot to a heavy metallic finish. Its massive drive wheels and complex linkages rotate with a steady, relentless rhythm. Atop the locomotive's chimney, thick, white steam billows backward, slowly dissipating in the soft evening air. Following closely behind are several passenger cars painted dark green, their brass-framed windows shimmering with warm light, hinting at the comfort within. As the wheels glide over the joints of the rails, each car sways slightly, rhythmically, side to side, a cadence unmistakable. The entire scene is bathed in the warm light of the golden hour, the setting sun gilding the steam clouds and casting long, flowing shadows across the train's sleek sides, creating an atmosphere of nostalgia, tranquility, and power.", + "seed": 27546, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_10/index_2350_seed_27546_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13519484549760818, + "vlm_logprob": -0.13519484549760818, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.13519484549760818, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31713026762008667 + ], + "opposite_logprob_tokens": [ + -1.3171303272247314 + ], + "logprob_reward": -0.31713026762008667, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31713026762008667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10616294294595718, + 0.10007142275571823, + 0.051787178963422775, + 0.06971272081136703, + 0.21435992419719696, + 0.21174509823322296, + 0.08846411108970642, + 0.06906116008758545, + 0.0545889250934124, + 0.03404654189944267 + ], + "credit_map": [ + 0.01839165762066841, + 0.014407663606107235, + 0.017061619088053703, + 0.012973693199455738, + 0.009582778438925743, + 0.011943507008254528, + 0.007607198785990477, + 0.014194827526807785, + 0.007433153223246336, + 0.008077587001025677, + 0.01737576350569725, + 0.012404795736074448, + 0.010398141108453274, + 0.01433979906141758, + 0.01483929343521595, + 0.015202891081571579, + 0.004454890266060829, + 0.005285440478473902, + 0.007829312235116959, + 0.007932969368994236, + 0.005245011765509844, + 0.006495046894997358, + 0.007931958884000778, + 0.00661255419254303, + 0.006945289205759764, + 0.008952375501394272, + 0.010660680010914803, + 0.006920245476067066, + 0.010743963532149792, + 0.006402779836207628, + 0.009536916390061378, + 0.009550472721457481, + 0.023271413519978523, + 0.02692261151969433, + 0.025493985041975975, + 0.02897528186440468, + 0.02948176860809326, + 0.023627890273928642, + 0.033211637288331985, + 0.02337533049285412, + 0.01945248432457447, + 0.022206895053386688, + 0.03217103332281113, + 0.031455863267183304, + 0.024820489808917046, + 0.029507674276828766, + 0.02271636761724949, + 0.029414284974336624, + 0.013875744305551052, + 0.006062975153326988, + 0.008355734869837761, + 0.008561481721699238, + 0.010888764634728432, + 0.013481320813298225, + 0.012930014170706272, + 0.014308072626590729, + 0.006723576691001654, + 0.006530898157507181, + 0.00843901839107275, + 0.008888479322195053, + 0.00767596485093236, + 0.00876125693321228, + 0.009340799413621426, + 0.01270117238163948, + 0.006539894267916679, + 0.005553087219595909, + 0.00532692251726985, + 0.003802954452112317, + 0.005915939342230558, + 0.010820120573043823, + 0.007069172337651253, + 0.009560832753777504, + 0.003725508926436305, + 0.004477409180253744, + 0.00532798795029521, + 0.004305379930883646, + 0.00339585286565125, + 0.00407069455832243, + 0.004598498810082674, + 0.004145211540162563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.31713026762008667 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016187608242035 + ], + "opposite_logprob_tokens": [ + -1.7016187906265259 + ], + "logprob_reward": -0.2016187608242035, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016187608242035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05630353465676308, + 0.08473025262355804, + 0.043247975409030914, + 0.058973528444767, + 0.1598222553730011, + 0.06917751580476761, + 0.08509556204080582, + 0.14234502613544464, + 0.18480660021305084, + 0.11549770832061768 + ], + "credit_map": [ + 0.008463587611913681, + 0.0049341460689902306, + 0.007634427398443222, + 0.005576407071202993, + 0.005301777273416519, + 0.0068506780080497265, + 0.006875097751617432, + 0.010667412541806698, + 0.005880304146558046, + 0.007620776072144508, + 0.01025389600545168, + 0.008509617298841476, + 0.010659176856279373, + 0.01578020304441452, + 0.012585793621838093, + 0.01344048511236906, + 0.003668806981295347, + 0.004191110376268625, + 0.007072585169225931, + 0.005103142000734806, + 0.003606067504733801, + 0.00517380703240633, + 0.006763683632016182, + 0.007668772712349892, + 0.007448232267051935, + 0.007103581912815571, + 0.006164956372231245, + 0.0054408893920481205, + 0.0061890240758657455, + 0.005894499830901623, + 0.008934292010962963, + 0.011798054911196232, + 0.01522747427225113, + 0.02065698616206646, + 0.01770772784948349, + 0.018951447680592537, + 0.022543055936694145, + 0.01927725411951542, + 0.02701564133167267, + 0.0184426661580801, + 0.007549856323748827, + 0.008211400359869003, + 0.011513886041939259, + 0.00982851255685091, + 0.007959932088851929, + 0.008756093680858612, + 0.007115489337593317, + 0.008242345415055752, + 0.019110839813947678, + 0.008053310215473175, + 0.008906086906790733, + 0.005333161447197199, + 0.006407993845641613, + 0.010213314555585384, + 0.0087039889767766, + 0.018366867676377296, + 0.015608330257236958, + 0.01453445479273796, + 0.027788003906607628, + 0.02424320951104164, + 0.012851431965827942, + 0.016213679686188698, + 0.017229123041033745, + 0.013876787386834621, + 0.020457642152905464, + 0.017073772847652435, + 0.03542131930589676, + 0.016850939020514488, + 0.020047856494784355, + 0.01769263856112957, + 0.03219982609152794, + 0.02506260573863983, + 0.013982128351926804, + 0.016506487503647804, + 0.022521011531352997, + 0.015160728245973587, + 0.008793116547167301, + 0.011918655596673489, + 0.013543365523219109, + 0.013072214089334011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.2016187608242035 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898955047130585 + ], + "opposite_logprob_tokens": [ + -2.5789895057678223 + ], + "logprob_reward": -0.07898955047130585, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898955047130585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0934264063835144, + 0.08302217721939087, + 0.23336048424243927, + 0.040593527257442474, + 0.11625459045171738, + 0.11535415798425674, + 0.06303242594003677, + 0.0986570492386818, + 0.10891355574131012, + 0.04738568887114525 + ], + "credit_map": [ + 0.011043530888855457, + 0.010284943506121635, + 0.01269170455634594, + 0.011037258431315422, + 0.011568963527679443, + 0.012660758569836617, + 0.013261656276881695, + 0.010877588763833046, + 0.008150123991072178, + 0.010191580280661583, + 0.00819526705890894, + 0.010431443341076374, + 0.013539623469114304, + 0.011689146049320698, + 0.010177542455494404, + 0.01064745057374239, + 0.021339673548936844, + 0.018473975360393524, + 0.055264946073293686, + 0.0318523570895195, + 0.019256791099905968, + 0.021914377808570862, + 0.029208006337285042, + 0.03605034947395325, + 0.003129713935777545, + 0.004583948291838169, + 0.0059559098444879055, + 0.0038375803269445896, + 0.007209240458905697, + 0.004079922568053007, + 0.006207235157489777, + 0.005589976441115141, + 0.009423949755728245, + 0.011346978135406971, + 0.01560027152299881, + 0.013978186063468456, + 0.017266282811760902, + 0.015981236472725868, + 0.015276383608579636, + 0.017381301149725914, + 0.011355367489159107, + 0.01844322495162487, + 0.010470474138855934, + 0.011038970202207565, + 0.021477580070495605, + 0.017245745286345482, + 0.010388387367129326, + 0.014934411272406578, + 0.011283330619335175, + 0.005654459353536367, + 0.006814927328377962, + 0.006747584789991379, + 0.007231543306261301, + 0.005741351284086704, + 0.009239069186151028, + 0.010320160537958145, + 0.010264371521770954, + 0.008366061374545097, + 0.009389162063598633, + 0.015688668936491013, + 0.011089960113167763, + 0.01664850115776062, + 0.012865662574768066, + 0.014344658702611923, + 0.010815621353685856, + 0.009079898707568645, + 0.01447684783488512, + 0.010245484299957752, + 0.011571395210921764, + 0.02295069582760334, + 0.016304852440953255, + 0.013468761928379536, + 0.007169452961534262, + 0.0077958665788173676, + 0.006537824869155884, + 0.0052348957397043705, + 0.004565911367535591, + 0.005312490742653608, + 0.005085152108222246, + 0.005684094503521919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.07898955047130585 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328597754240036 + ], + "opposite_logprob_tokens": [ + -3.7732858657836914 + ], + "logprob_reward": -0.02328597754240036, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328597754240036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04229402914643288, + 0.10858451575040817, + 0.13505592942237854, + 0.05725879594683647, + 0.07589257508516312, + 0.10069401562213898, + 0.11016688495874405, + 0.1236892119050026, + 0.1998337209224701, + 0.04653025418519974 + ], + "credit_map": [ + 0.007047398015856743, + 0.004314957186579704, + 0.005881073419004679, + 0.005044437944889069, + 0.004082230385392904, + 0.0060239676386117935, + 0.004656771197915077, + 0.005243196152150631, + 0.00857486017048359, + 0.008444932289421558, + 0.020183973014354706, + 0.010051893070340157, + 0.012874914333224297, + 0.020360535010695457, + 0.011784051544964314, + 0.016309361904859543, + 0.008156283758580685, + 0.011507535353302956, + 0.03365309163928032, + 0.01925034075975418, + 0.01018354669213295, + 0.00978542398661375, + 0.016810985282063484, + 0.02570871077477932, + 0.006184905301779509, + 0.005368281155824661, + 0.008554221130907536, + 0.007300392258912325, + 0.005934954155236483, + 0.007165773306041956, + 0.008261434733867645, + 0.008488830178976059, + 0.007353888358920813, + 0.008019229397177696, + 0.011197208426892757, + 0.011816395446658134, + 0.007450243923813105, + 0.006584119517356157, + 0.013124769553542137, + 0.010346716269850731, + 0.008703946135938168, + 0.016030168160796165, + 0.011690656654536724, + 0.013672804459929466, + 0.01649446412920952, + 0.013048659078776836, + 0.00930400937795639, + 0.011749312281608582, + 0.02674701064825058, + 0.009433818981051445, + 0.0068464926443994045, + 0.009120231494307518, + 0.008953982964158058, + 0.010476534254848957, + 0.013286986388266087, + 0.02530183270573616, + 0.013754360377788544, + 0.012769744731485844, + 0.02581183984875679, + 0.021195203065872192, + 0.010808673687279224, + 0.00972022395581007, + 0.01498386263847351, + 0.014645298011600971, + 0.019815441220998764, + 0.014446276240050793, + 0.01944143883883953, + 0.014223002828657627, + 0.020546693354845047, + 0.04802624508738518, + 0.036012426018714905, + 0.027322186157107353, + 0.005718863569200039, + 0.006654995493590832, + 0.005488365422934294, + 0.006008314900100231, + 0.004704651422798634, + 0.004695625510066748, + 0.00644200574606657, + 0.00681743398308754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.02328597754240036 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054949671030044556 + ], + "opposite_logprob_tokens": [ + -2.9299497604370117 + ], + "logprob_reward": -0.054949671030044556, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054949671030044556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09501487761735916, + 0.07330478727817535, + 0.05993200093507767, + 0.04613250866532326, + 0.07384688407182693, + 0.09870360046625137, + 0.041085246950387955, + 0.18578602373600006, + 0.1026221364736557, + 0.22357194125652313 + ], + "credit_map": [ + 0.013300850056111813, + 0.00868489034473896, + 0.01406051404774189, + 0.009764430113136768, + 0.011061020195484161, + 0.014833830296993256, + 0.011799715459346771, + 0.01150962058454752, + 0.006040226202458143, + 0.008137994445860386, + 0.009024024941027164, + 0.009144501760601997, + 0.010064241476356983, + 0.010681727901101112, + 0.009534298442304134, + 0.010677763260900974, + 0.005055650137364864, + 0.004572233650833368, + 0.01904403045773506, + 0.00593661330640316, + 0.0037481081672012806, + 0.004493196029216051, + 0.004942712839692831, + 0.012139455415308475, + 0.0042067114263772964, + 0.0049366517923772335, + 0.008007572032511234, + 0.005040440242737532, + 0.00742038618773222, + 0.004354827105998993, + 0.005436139181256294, + 0.0067297774367034435, + 0.008839086629450321, + 0.00808747112751007, + 0.010664173401892185, + 0.007552540395408869, + 0.010599922388792038, + 0.007443748414516449, + 0.010810112580657005, + 0.009849825873970985, + 0.012418130412697792, + 0.013284699991345406, + 0.011427829973399639, + 0.013392188586294651, + 0.012047245167195797, + 0.009625394828617573, + 0.011750571429729462, + 0.01475752517580986, + 0.004977004136890173, + 0.004046981688588858, + 0.0065558734349906445, + 0.0032019298523664474, + 0.0052058217115700245, + 0.006315123289823532, + 0.005437618587166071, + 0.0053448923863470554, + 0.01988396607339382, + 0.017639270052313805, + 0.02782873623073101, + 0.03272900730371475, + 0.018922023475170135, + 0.017809730023145676, + 0.023762892931699753, + 0.027210375294089317, + 0.012484103441238403, + 0.006499325856566429, + 0.008601375855505466, + 0.007154328748583794, + 0.01011411752551794, + 0.024146977812051773, + 0.01685170643031597, + 0.01677018776535988, + 0.03196224942803383, + 0.03048228845000267, + 0.03228827938437462, + 0.04230864718556404, + 0.014086313545703888, + 0.022189030423760414, + 0.02368020825088024, + 0.026574913412332535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.054949671030044556 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot moves parallel to a majestic old steam locomotive as it powerfully churns along tracks laid across the rolling countryside. The locomotive is pitch black, its surface stained by age and soot to a heavy metallic finish. Its massive drive wheels and complex linkages rotate with a steady, relentless rhythm. Atop the locomotive's chimney, thick, white steam billows backward, slowly dissipating in the soft evening air. Following closely behind are several passenger cars painted dark green, their brass-framed windows shimmering with warm light, hinting at the comfort within. As the wheels glide over the joints of the rails, each car sways slightly, rhythmically, side to side, a cadence unmistakable. The entire scene is bathed in the warm light of the golden hour, the setting sun gilding the steam clouds and casting long, flowing shadows across the train's sleek sides, creating an atmosphere of nostalgia, tranquility, and power.", + "seed": 27547, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_10/index_2350_seed_27547_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10880114436149597, + "vlm_logprob": -0.10880114436149597, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.10880114436149597, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25587862730026245 + ], + "opposite_logprob_tokens": [ + -1.5058786869049072 + ], + "logprob_reward": -0.25587862730026245, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25587862730026245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07843734323978424, + 0.01723582297563553, + 0.03261982277035713, + 0.04168481379747391, + 0.4557282328605652, + 0.03723195567727089, + 0.034248873591423035, + 0.04869227483868599, + 0.12354262173175812, + 0.13057823479175568 + ], + "credit_map": [ + 0.008147207088768482, + 0.012602419592440128, + 0.010294027626514435, + 0.010526097379624844, + 0.01331799291074276, + 0.0058950926177203655, + 0.010129834525287151, + 0.0075246659107506275, + 0.0025106342509388924, + 0.0038448746781796217, + 0.001968268770724535, + 0.0012804976431652904, + 0.002826468087732792, + 0.002105901949107647, + 0.0014699142193421721, + 0.0012292632600292563, + 0.004657910671085119, + 0.005871395114809275, + 0.002995674032717943, + 0.0029897508211433887, + 0.0037366223987191916, + 0.004569782875478268, + 0.003943185321986675, + 0.0038554994389414787, + 0.004261590074747801, + 0.004343067295849323, + 0.0032456982880830765, + 0.0036768270656466484, + 0.007529159542173147, + 0.008270802907645702, + 0.00552809052169323, + 0.004829575773328543, + 0.0507158599793911, + 0.06149972975254059, + 0.030076447874307632, + 0.03012658655643463, + 0.046942874789237976, + 0.11283553391695023, + 0.0612030066549778, + 0.06232814863324165, + 0.004043142311275005, + 0.0034845885820686817, + 0.0036844233982264996, + 0.0038367249071598053, + 0.002994827227666974, + 0.004273748490959406, + 0.008865661919116974, + 0.006048835813999176, + 0.003935865126550198, + 0.00476795295253396, + 0.003307248465716839, + 0.005339549854397774, + 0.0037339343689382076, + 0.002858328865841031, + 0.00457643810659647, + 0.005729555618017912, + 0.005412629339843988, + 0.00612251553684473, + 0.00298704719170928, + 0.005056062713265419, + 0.0034006377682089806, + 0.005677063949406147, + 0.009191052056849003, + 0.010845259763300419, + 0.01171601377427578, + 0.009798312559723854, + 0.017117999494075775, + 0.01624004729092121, + 0.014938459731638432, + 0.02644285559654236, + 0.016890471801161766, + 0.010398454032838345, + 0.018724197521805763, + 0.009079735726118088, + 0.011844885535538197, + 0.016169840469956398, + 0.0141445966437459, + 0.025006724521517754, + 0.02011590264737606, + 0.015492336824536324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.25587862730026245 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11299815028905869 + ], + "opposite_logprob_tokens": [ + -2.2379982471466064 + ], + "logprob_reward": -0.11299815028905869, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11299815028905869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06349897384643555, + 0.026647385209798813, + 0.01712278090417385, + 0.03616608679294586, + 0.2567538917064667, + 0.049108877778053284, + 0.0644998699426651, + 0.14720547199249268, + 0.11179907619953156, + 0.2271975427865982 + ], + "credit_map": [ + 0.01010038610547781, + 0.00681162578985095, + 0.011480744928121567, + 0.011252196505665779, + 0.005489550065249205, + 0.005543275736272335, + 0.005576460622251034, + 0.007244730833917856, + 0.0030082743614912033, + 0.004227190278470516, + 0.0037127374671399593, + 0.002945826854556799, + 0.00379596883431077, + 0.003646287601441145, + 0.0028180780354887247, + 0.002493023406714201, + 0.002227151533588767, + 0.002503028605133295, + 0.0016834636917337775, + 0.0017914492636919022, + 0.0025213281624019146, + 0.0021962877362966537, + 0.0025933249853551388, + 0.0016067489050328732, + 0.004349426366388798, + 0.0049798982217907906, + 0.0031367980409413576, + 0.0024628699757158756, + 0.005084979813545942, + 0.00437454367056489, + 0.007041838485747576, + 0.0047357333824038506, + 0.029770974069833755, + 0.04147862643003464, + 0.02053968980908394, + 0.016475128009915352, + 0.027098244056105614, + 0.058679644018411636, + 0.030010735616087914, + 0.03270087018609047, + 0.0070476592518389225, + 0.004732784815132618, + 0.004463902674615383, + 0.004306421149522066, + 0.0033167870715260506, + 0.0046257455833256245, + 0.009607424028217793, + 0.011008160188794136, + 0.009172786958515644, + 0.013675713911652565, + 0.006698342971503735, + 0.006787028629332781, + 0.006170667242258787, + 0.0055303494445979595, + 0.0074523817747831345, + 0.009012607857584953, + 0.017209207639098167, + 0.02202233485877514, + 0.023067329078912735, + 0.018445834517478943, + 0.01737293042242527, + 0.01814619079232216, + 0.016668232157826424, + 0.014273421838879585, + 0.013422285206615925, + 0.010476638562977314, + 0.015584450215101242, + 0.010427781380712986, + 0.011125745251774788, + 0.021698063239455223, + 0.013711451552808285, + 0.015352667309343815, + 0.03377983719110489, + 0.013999988324940205, + 0.03656071051955223, + 0.02802325412631035, + 0.02103779837489128, + 0.041202984750270844, + 0.032953936606645584, + 0.019639045000076294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.11299815028905869 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043067965656518936 + ], + "opposite_logprob_tokens": [ + -3.1680679321289062 + ], + "logprob_reward": -0.043067965656518936, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043067965656518936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37328118085861206, + 0.17292694747447968, + 0.024646611884236336, + 0.04571685567498207, + 0.019427521154284477, + 0.11462026089429855, + 0.10300496965646744, + 0.08148849755525589, + 0.03271784633398056, + 0.03216927498579025 + ], + "credit_map": [ + 0.04790481925010681, + 0.07821591943502426, + 0.043814897537231445, + 0.04751843214035034, + 0.06287913769483566, + 0.023265987634658813, + 0.04353995993733406, + 0.026142070069909096, + 0.024267245084047318, + 0.04322630912065506, + 0.02599937468767166, + 0.015012104995548725, + 0.02345454879105091, + 0.012739405035972595, + 0.016436953097581863, + 0.01179101224988699, + 0.0034346231259405613, + 0.003397855209186673, + 0.002270811004564166, + 0.0014985122252255678, + 0.003246881300583482, + 0.004779324401170015, + 0.0022886130027472973, + 0.0037299946416169405, + 0.006021363195031881, + 0.007665529847145081, + 0.005682189948856831, + 0.0037503137718886137, + 0.00433886656537652, + 0.006731449160724878, + 0.005260201171040535, + 0.006266946438699961, + 0.00285935215651989, + 0.0036163318436592817, + 0.002811080077663064, + 0.0016548257553949952, + 0.001636325498111546, + 0.003183170221745968, + 0.0015070136869326234, + 0.002159423427656293, + 0.01949354074895382, + 0.0174249317497015, + 0.00811714306473732, + 0.008192664943635464, + 0.0077519891783595085, + 0.008790085092186928, + 0.01694747805595398, + 0.027902448549866676, + 0.013317640870809555, + 0.01581471413373947, + 0.006238449364900589, + 0.009997624903917313, + 0.010848920792341232, + 0.011128085665404797, + 0.01692097820341587, + 0.01873856596648693, + 0.009623604826629162, + 0.010815273970365524, + 0.00816779863089323, + 0.0064223892986774445, + 0.010511535219848156, + 0.013474434614181519, + 0.011264113709330559, + 0.011209356598556042, + 0.0026092459447681904, + 0.002427593106403947, + 0.0031312766950577497, + 0.003989061806350946, + 0.0035623745061457157, + 0.005658440291881561, + 0.006765694357454777, + 0.004574162419885397, + 0.002784323412925005, + 0.003369459416717291, + 0.0037549107801169157, + 0.0047616297379136086, + 0.005126628093421459, + 0.002567603951320052, + 0.005720747169107199, + 0.004083976615220308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.043067965656518936 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001392543315887 + ], + "opposite_logprob_tokens": [ + -2.695013999938965 + ], + "logprob_reward": -0.07001392543315887, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001392543315887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09640318900346756, + 0.05510881915688515, + 0.017781579867005348, + 0.06244650110602379, + 0.1330079287290573, + 0.05642382428050041, + 0.12320500612258911, + 0.3573046922683716, + 0.055925920605659485, + 0.04239258915185928 + ], + "credit_map": [ + 0.01211522612720728, + 0.011396977119147778, + 0.01185604091733694, + 0.012675946578383446, + 0.014329230412840843, + 0.007457956206053495, + 0.016539450734853745, + 0.010032357648015022, + 0.007545731961727142, + 0.007958571426570415, + 0.006688724737614393, + 0.005764999892562628, + 0.007872912101447582, + 0.006543370429426432, + 0.0072799380868673325, + 0.00545456912368536, + 0.0024921679869294167, + 0.0022111088037490845, + 0.0019313213415443897, + 0.0020911151077598333, + 0.00244215433485806, + 0.0027502188459038734, + 0.0018556735012680292, + 0.0020078199449926615, + 0.005677530076354742, + 0.0048659625463187695, + 0.004960017278790474, + 0.005668756552040577, + 0.0073455339297652245, + 0.012182366102933884, + 0.013499917462468147, + 0.008246418088674545, + 0.015674525871872902, + 0.021405788138508797, + 0.00999319925904274, + 0.008459965698421001, + 0.013645553030073643, + 0.030743323266506195, + 0.014628484845161438, + 0.01845707930624485, + 0.008957899175584316, + 0.006692386232316494, + 0.006033026147633791, + 0.004569471348077059, + 0.004420074634253979, + 0.007287138607352972, + 0.007904766127467155, + 0.010559061542153358, + 0.014429666101932526, + 0.011670875363051891, + 0.01252434216439724, + 0.01235674973577261, + 0.011760038323700428, + 0.015404683537781239, + 0.025386974215507507, + 0.019671672955155373, + 0.043067943304777145, + 0.031389638781547546, + 0.030966749414801598, + 0.033851224929094315, + 0.027233216911554337, + 0.04505900666117668, + 0.06417401880025864, + 0.08156286925077438, + 0.006207410246133804, + 0.007331300061196089, + 0.006574716884642839, + 0.00840665027499199, + 0.004155205097049475, + 0.007366040255874395, + 0.007663763128221035, + 0.008220837451517582, + 0.005811281502246857, + 0.0024206945672631264, + 0.003839947050437331, + 0.004579868167638779, + 0.004315313417464495, + 0.008735264651477337, + 0.006584105081856251, + 0.006106116343289614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.07001392543315887 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pitch black steam locomotive with a heavy metallic finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are passenger cars painted dark green with brass-framed windows visible behind the locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is thick white steam billowing backward from the locomotive's chimney?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a low-angle tracking shot moving parallel to the train?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by the warm light of the golden hour with long, flowing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204705312848091 + ], + "opposite_logprob_tokens": [ + -2.812047004699707 + ], + "logprob_reward": -0.06204705312848091, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204705312848091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22069454193115234, + 0.042546890676021576, + 0.10636565089225769, + 0.08121276646852493, + 0.09857477992773056, + 0.0483732745051384, + 0.15560758113861084, + 0.03727371245622635, + 0.1723565012216568, + 0.03699425980448723 + ], + "credit_map": [ + 0.02675729990005493, + 0.04052140191197395, + 0.02732018008828163, + 0.029138974845409393, + 0.03772926330566406, + 0.017230018973350525, + 0.022325841709971428, + 0.01967158168554306, + 0.005713548511266708, + 0.0064560482278466225, + 0.00772643368691206, + 0.00501265237107873, + 0.005986964330077171, + 0.003952018450945616, + 0.004797513131052256, + 0.0029017203487455845, + 0.015326043590903282, + 0.023893114179372787, + 0.010956323705613613, + 0.008314570412039757, + 0.012996497564017773, + 0.010662700980901718, + 0.011265803128480911, + 0.012950610369443893, + 0.010814602486789227, + 0.009119955822825432, + 0.007448118180036545, + 0.006777381524443626, + 0.009292230941355228, + 0.010371429845690727, + 0.0127480523660779, + 0.014640997163951397, + 0.013814513571560383, + 0.022273346781730652, + 0.012828783132135868, + 0.010050589218735695, + 0.008454924449324608, + 0.009808152914047241, + 0.011019548401236534, + 0.010324923321604729, + 0.0070616682060062885, + 0.009060626849532127, + 0.0050440202467143536, + 0.004044040106236935, + 0.004356598015874624, + 0.004363554995507002, + 0.004985434468835592, + 0.00945733580738306, + 0.01797056756913662, + 0.024778975173830986, + 0.01688091643154621, + 0.020408477634191513, + 0.017323609441518784, + 0.018190449103713036, + 0.018933046609163284, + 0.021121559664607048, + 0.0035299118608236313, + 0.005046338774263859, + 0.004124023485928774, + 0.004730623681098223, + 0.0030749128200113773, + 0.0034080655314028263, + 0.006524443626403809, + 0.006835395935922861, + 0.021325742825865746, + 0.030597910284996033, + 0.02361462078988552, + 0.019775932654738426, + 0.01663188822567463, + 0.01678118295967579, + 0.021869465708732605, + 0.021759772673249245, + 0.003495896002277732, + 0.0038314422126859426, + 0.005598008632659912, + 0.005874447524547577, + 0.003957726061344147, + 0.0035569928586483, + 0.006801679730415344, + 0.003878070507198572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.06204705312848091 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_11/index_tensor([6698, 6698])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_11/index_tensor([6698, 6698])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..28484237462247158bef574c7b7b6b05d7b32a44 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_11/index_tensor([6698, 6698])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly descending, ritualistic overhead shot unfolds, opening from a solid wood tabletop with delicate dark wood grain and slowly approaching a handleless, thin-walled, transparent glass cup placed in its center. A soft ray of afternoon sunlight slants through the window, gently shimmering on the cup's walls and filtering through the clear, light amber tea within. As the camera continues to descend, we clearly see several previously dry, shrunken golden chrysanthemums. As if awakened from a deep slumber, the hot water gently and delicately unfolds their petals, layer by layer. This blossoming process is captured with exquisite precision, each petal gracefully stretching outward, ultimately fully blossoming within the tea, forming plump, delicate golden corollas. Their golden hue appears remarkably translucent and luminous in the light, with a soft halo emanating from the edges. The entire scene is bathed in a warm, soft beige hue, creating an atmosphere of profound tranquility, focus, and healing Zen.", + "seed": 13593, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_11/index_6698_seed_13593_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15482995435595512, + "vlm_logprob": -0.15482995435595512, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.15482995435595512, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5821131467819214 + ], + "opposite_logprob_tokens": [ + -0.8321131467819214 + ], + "logprob_reward": -0.5821131467819214, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5821131467819214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14480721950531006, + 0.025509437546133995, + 0.047015827149152756, + 0.6565427184104919, + 0.04137471318244934, + 0.019536564126610756, + 0.014429840259253979, + 0.009597407653927803, + 0.031126966699957848, + 0.010059280321002007 + ], + "credit_map": [ + 0.014554374851286411, + 0.05162665992975235, + 0.005339813884347677, + 0.006931583397090435, + 0.04631764069199562, + 0.007584712002426386, + 0.007186524104326963, + 0.005265895277261734, + 0.0023984722793102264, + 0.003899696981534362, + 0.0029512501787394285, + 0.003413273487240076, + 0.0019007910741493106, + 0.00254882057197392, + 0.0034270600881427526, + 0.00497007230296731, + 0.003889994928613305, + 0.0035326494835317135, + 0.009993016719818115, + 0.007437701337039471, + 0.003802908817306161, + 0.004861443769186735, + 0.00816321186721325, + 0.005334899760782719, + 0.10997045785188675, + 0.023103006184101105, + 0.09743645787239075, + 0.04391559213399887, + 0.018911423161625862, + 0.04307417571544647, + 0.043132808059453964, + 0.2769987881183624, + 0.0032406251411885023, + 0.004811885766685009, + 0.005722224246710539, + 0.006217328831553459, + 0.003548925742506981, + 0.0037506071384996176, + 0.005379064008593559, + 0.008704050444066525, + 0.0013847033260390162, + 0.0012417559046298265, + 0.001859538722783327, + 0.0032559337560087442, + 0.0018249788554385304, + 0.0013706701574847102, + 0.005899770185351372, + 0.0026992126367986202, + 0.0012345494469627738, + 0.0020308885723352432, + 0.001551992492750287, + 0.0024808072485029697, + 0.001278229639865458, + 0.0012093880213797092, + 0.0029683734755963087, + 0.0016756108961999416, + 0.000878566992469132, + 0.001724794739857316, + 0.0013707475736737251, + 0.0011674200650304556, + 0.0012410818599164486, + 0.0007639631512574852, + 0.001633280306123197, + 0.0008175530820153654, + 0.004737334791570902, + 0.0074303229339420795, + 0.0031513089779764414, + 0.0033903964795172215, + 0.0032017945777624846, + 0.002265404909849167, + 0.004151757340878248, + 0.0027986473869532347, + 0.0010186629369854927, + 0.001895892433822155, + 0.0008799908100627363, + 0.0012287256540730596, + 0.0015293735777959228, + 0.0010035866871476173, + 0.0017124811420217156, + 0.0007905670208856463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.5821131467819214 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14286400377750397 + ], + "opposite_logprob_tokens": [ + -2.0178639888763428 + ], + "logprob_reward": -0.14286400377750397, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14286400377750397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18373751640319824, + 0.06410963833332062, + 0.05342933535575867, + 0.29554489254951477, + 0.10997086018323898, + 0.03829625993967056, + 0.044205013662576675, + 0.10951771587133408, + 0.0907837375998497, + 0.010405045934021473 + ], + "credit_map": [ + 0.01129150576889515, + 0.07029961794614792, + 0.0062348973006010056, + 0.007340109907090664, + 0.06401119381189346, + 0.009875811636447906, + 0.007936269976198673, + 0.0067481109872460365, + 0.009090126492083073, + 0.021110596135258675, + 0.00351261836476624, + 0.0066978042013943195, + 0.0058687287382781506, + 0.004883408080786467, + 0.005296455696225166, + 0.007649898063391447, + 0.0036970791406929493, + 0.0043618702329695225, + 0.00549273518845439, + 0.009570464491844177, + 0.009077818132936954, + 0.004989124834537506, + 0.009120791219174862, + 0.007119449786841869, + 0.04221539944410324, + 0.011381617747247219, + 0.05246023088693619, + 0.020992286503314972, + 0.008159385062754154, + 0.01594654470682144, + 0.0184466689825058, + 0.12594275176525116, + 0.0077188629657030106, + 0.011890956200659275, + 0.010507243685424328, + 0.0234854593873024, + 0.006809244863688946, + 0.006343619432300329, + 0.028571832925081253, + 0.014643646776676178, + 0.003193474840372801, + 0.0025576509069651365, + 0.003318070201203227, + 0.0077127087861299515, + 0.003069705795496702, + 0.0027405184227973223, + 0.011347061023116112, + 0.004357069730758667, + 0.0035764402709901333, + 0.005214512813836336, + 0.004204573109745979, + 0.007286107633262873, + 0.0032133255153894424, + 0.004853239748626947, + 0.008647498674690723, + 0.007209316361695528, + 0.01631643995642662, + 0.022842315956950188, + 0.014719311147928238, + 0.009134089574217796, + 0.01508688647300005, + 0.01159617304801941, + 0.010028094984591007, + 0.009794404730200768, + 0.013523218221962452, + 0.01837925985455513, + 0.008761377073824406, + 0.01239596027880907, + 0.010024570859968662, + 0.006883293390274048, + 0.012039046734571457, + 0.00877701211720705, + 0.0015627286629751325, + 0.002167477272450924, + 0.0009494263213127851, + 0.0011283213971182704, + 0.00140131707303226, + 0.001435152138583362, + 0.000913659343495965, + 0.0008469639578834176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.14286400377750397 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330007031559944 + ], + "opposite_logprob_tokens": [ + -3.7733001708984375 + ], + "logprob_reward": -0.02330007031559944, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330007031559944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15898513793945312, + 0.08788369596004486, + 0.0759533941745758, + 0.2527374029159546, + 0.06823954731225967, + 0.1812799870967865, + 0.03086957335472107, + 0.021687166765332222, + 0.028351882472634315, + 0.09401220083236694 + ], + "credit_map": [ + 0.027990039438009262, + 0.06583783775568008, + 0.005811095703393221, + 0.006680207327008247, + 0.03272183984518051, + 0.007934987545013428, + 0.006681844592094421, + 0.005327280610799789, + 0.010240291245281696, + 0.02914336323738098, + 0.006031657103449106, + 0.009218908846378326, + 0.007757196668535471, + 0.008235315792262554, + 0.0075476816855371, + 0.009709282778203487, + 0.0051459986716508865, + 0.005914948880672455, + 0.009606531821191311, + 0.019650235772132874, + 0.010327109135687351, + 0.004801237024366856, + 0.011862064711749554, + 0.008645260706543922, + 0.04972933977842331, + 0.01143439207226038, + 0.02938249334692955, + 0.019594797864556313, + 0.00788832176476717, + 0.01812904141843319, + 0.015404805541038513, + 0.10117421299219131, + 0.005341079086065292, + 0.01236117072403431, + 0.009275401942431927, + 0.010963836684823036, + 0.003876484464854002, + 0.006062588654458523, + 0.011014813557267189, + 0.00934416614472866, + 0.013430536724627018, + 0.013649974949657917, + 0.018631407991051674, + 0.028038883581757545, + 0.019952593371272087, + 0.014795766212046146, + 0.0486905574798584, + 0.024090278893709183, + 0.0024728570133447647, + 0.004295188467949629, + 0.00334010343067348, + 0.005978386849164963, + 0.002527390606701374, + 0.0024150938261300325, + 0.006825092248618603, + 0.0030154609121382236, + 0.0029773118440061808, + 0.003969402052462101, + 0.002626169938594103, + 0.0024767573922872543, + 0.0025449029635638, + 0.0020960902329534292, + 0.002808252116665244, + 0.0021882778964936733, + 0.0015077829593792558, + 0.003115271683782339, + 0.0026496367063373327, + 0.0034763820003718138, + 0.002727913437411189, + 0.00316215748898685, + 0.003724886802956462, + 0.007987851276993752, + 0.011327833868563175, + 0.02290426380932331, + 0.008186445571482182, + 0.010660435073077679, + 0.014989947900176048, + 0.009300184436142445, + 0.009496797807514668, + 0.007146290037781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.02330007031559944 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608842797577381 + ], + "opposite_logprob_tokens": [ + -4.141088485717773 + ], + "logprob_reward": -0.01608842797577381, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608842797577381, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10185252875089645, + 0.04252365231513977, + 0.03657027706503868, + 0.4972224831581116, + 0.032635219395160675, + 0.04845406115055084, + 0.05721893906593323, + 0.07610242813825607, + 0.08066073805093765, + 0.026759738102555275 + ], + "credit_map": [ + 0.025650912895798683, + 0.03880394250154495, + 0.00459075253456831, + 0.006260338705033064, + 0.008922506123781204, + 0.006595810875296593, + 0.005983336828649044, + 0.005044921301305294, + 0.00583589868620038, + 0.01403926033526659, + 0.002830910263583064, + 0.003465882735326886, + 0.003896104171872139, + 0.003335027489811182, + 0.003978662192821503, + 0.005141905043274164, + 0.002826507668942213, + 0.002695387229323387, + 0.004220437258481979, + 0.009435930289328098, + 0.00379098835401237, + 0.0027467275504022837, + 0.006132922600954771, + 0.004721371456980705, + 0.08257363736629486, + 0.017179075628519058, + 0.07644001394510269, + 0.032765425741672516, + 0.013629663735628128, + 0.03163183480501175, + 0.031676121056079865, + 0.2113266885280609, + 0.0023539485409855843, + 0.00510254455730319, + 0.003974886145442724, + 0.006611342541873455, + 0.0023588526528328657, + 0.0017272067489102483, + 0.006239172071218491, + 0.004267266485840082, + 0.0037270127795636654, + 0.004041390959173441, + 0.006685574073344469, + 0.007230130955576897, + 0.006498390808701515, + 0.0029990447219461203, + 0.011902819387614727, + 0.005369692575186491, + 0.004393325187265873, + 0.008699803613126278, + 0.0061827776953577995, + 0.00991518422961235, + 0.004796245601028204, + 0.004803383257240057, + 0.011533848010003567, + 0.006894364021718502, + 0.008058225736021996, + 0.01478826068341732, + 0.008288472890853882, + 0.009248113259673119, + 0.010646459646522999, + 0.0069518918171525, + 0.011285499669611454, + 0.006835497450083494, + 0.00873665139079094, + 0.012095607817173004, + 0.006742119323462248, + 0.00987175852060318, + 0.00977468304336071, + 0.008185791783034801, + 0.014146269299089909, + 0.011107849888503551, + 0.0020324939396232367, + 0.003853433532640338, + 0.002245764946565032, + 0.003877260023728013, + 0.005379564128816128, + 0.0025254266802221537, + 0.003212776966392994, + 0.0036330162547528744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.01608842797577381 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784122928977013 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784122928977013, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784122928977013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18240030109882355, + 0.06652326881885529, + 0.06048691272735596, + 0.4412713646888733, + 0.027364080771803856, + 0.02039734087884426, + 0.028705716133117676, + 0.061082810163497925, + 0.038599349558353424, + 0.07316882163286209 + ], + "credit_map": [ + 0.0337110199034214, + 0.027411719784140587, + 0.02031293697655201, + 0.015065828338265419, + 0.03396204113960266, + 0.016732102259993553, + 0.01976723037660122, + 0.015437429770827293, + 0.007851377129554749, + 0.02054702118039131, + 0.00319335306994617, + 0.006169966422021389, + 0.005717939231544733, + 0.005042303819209337, + 0.006033641751855612, + 0.011967670172452927, + 0.0044069066643714905, + 0.005859323777258396, + 0.007449242286384106, + 0.009976692497730255, + 0.00999815296381712, + 0.006712146569043398, + 0.009438577108085155, + 0.006645869463682175, + 0.07902143895626068, + 0.01662517711520195, + 0.05975829064846039, + 0.030635571107268333, + 0.014444740489125252, + 0.03176817670464516, + 0.02934320829808712, + 0.17967472970485687, + 0.0019007642986252904, + 0.0032864732202142477, + 0.0028777853585779667, + 0.005371365696191788, + 0.002571161836385727, + 0.0017426955746486783, + 0.006092457100749016, + 0.0035213795490562916, + 0.0017014205222949386, + 0.0020195443648844957, + 0.0026412608567625284, + 0.003433363977819681, + 0.0019356473349034786, + 0.0015780802350491285, + 0.0047238548286259174, + 0.00236417050473392, + 0.002191001083701849, + 0.004027585498988628, + 0.003263564081862569, + 0.005645942408591509, + 0.0029200089629739523, + 0.0021278008352965117, + 0.005902481731027365, + 0.002627329668030143, + 0.005177770741283894, + 0.007211573421955109, + 0.0066663408651947975, + 0.007323870435357094, + 0.013051298446953297, + 0.006180710159242153, + 0.009594676084816456, + 0.005876570474356413, + 0.0053433822467923164, + 0.007769536226987839, + 0.0031575686298310757, + 0.004961680620908737, + 0.003460196778178215, + 0.003041744465008378, + 0.005897509399801493, + 0.004967734217643738, + 0.008271120488643646, + 0.01748909428715706, + 0.006302989553660154, + 0.008384390734136105, + 0.01141798309981823, + 0.00719995703548193, + 0.008324439637362957, + 0.005778845865279436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.009784122928977013 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly descending, ritualistic overhead shot unfolds, opening from a solid wood tabletop with delicate dark wood grain and slowly approaching a handleless, thin-walled, transparent glass cup placed in its center. A soft ray of afternoon sunlight slants through the window, gently shimmering on the cup's walls and filtering through the clear, light amber tea within. As the camera continues to descend, we clearly see several previously dry, shrunken golden chrysanthemums. As if awakened from a deep slumber, the hot water gently and delicately unfolds their petals, layer by layer. This blossoming process is captured with exquisite precision, each petal gracefully stretching outward, ultimately fully blossoming within the tea, forming plump, delicate golden corollas. Their golden hue appears remarkably translucent and luminous in the light, with a soft halo emanating from the edges. The entire scene is bathed in a warm, soft beige hue, creating an atmosphere of profound tranquility, focus, and healing Zen.", + "seed": 13594, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_11/index_6698_seed_13594_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1752826080657542, + "vlm_logprob": -0.1752826080657542, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.1752826080657542, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7606732845306396 + ], + "opposite_logprob_tokens": [ + -0.6356732845306396 + ], + "logprob_reward": -0.7606732845306396, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7606732845306396, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.25622308254241943, + 0.0875692367553711, + 0.15874330699443817, + 0.03535992652177811, + 0.07324845343828201, + 0.05700380355119705, + 0.09979317337274551, + 0.17619380354881287, + 0.027629517018795013, + 0.028235696256160736 + ], + "credit_map": [ + 0.015776846557855606, + 0.07486867159605026, + 0.07095400989055634, + 0.02768889255821705, + 0.016962431371212006, + 0.01563451811671257, + 0.02079010009765625, + 0.013547602109611034, + 0.007501542568206787, + 0.01184407901018858, + 0.00853432435542345, + 0.016805199906229973, + 0.007246822584420443, + 0.010985872708261013, + 0.00966054480522871, + 0.01499085407704115, + 0.011458228342235088, + 0.016628948971629143, + 0.024061009287834167, + 0.016521455720067024, + 0.015176791697740555, + 0.02225710265338421, + 0.0248727910220623, + 0.02776697650551796, + 0.002161678159609437, + 0.004939483478665352, + 0.004598330706357956, + 0.004291832447052002, + 0.0037659797817468643, + 0.004790616221725941, + 0.005677812732756138, + 0.005134193692356348, + 0.0047652204521000385, + 0.008544474840164185, + 0.007683428004384041, + 0.007716909982264042, + 0.006830002646893263, + 0.010489258915185928, + 0.01916709914803505, + 0.008052059449255466, + 0.004164533689618111, + 0.009657018817961216, + 0.0080652991309762, + 0.0056131016463041306, + 0.0067547946237027645, + 0.00962416548281908, + 0.006105000618845224, + 0.007019883953034878, + 0.011522507295012474, + 0.02259702980518341, + 0.01540499459952116, + 0.008850901387631893, + 0.0087449811398983, + 0.015336870215833187, + 0.008827063255012035, + 0.008508829399943352, + 0.03304368257522583, + 0.01820891536772251, + 0.016468804329633713, + 0.020060112699866295, + 0.012459940277040005, + 0.01802491769194603, + 0.014942526817321777, + 0.04298489913344383, + 0.0023857669439166784, + 0.0028092393185943365, + 0.0030889029148966074, + 0.004887084010988474, + 0.002935584867373109, + 0.005083474330604076, + 0.0033175009302794933, + 0.0031219632364809513, + 0.003164076479151845, + 0.003078433917835355, + 0.0036653042770922184, + 0.005614141933619976, + 0.0036321240477263927, + 0.0028447594959288836, + 0.0037917285226285458, + 0.002445128047838807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.7606732845306396 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499209463596344 + ], + "opposite_logprob_tokens": [ + -2.929992198944092 + ], + "logprob_reward": -0.05499209463596344, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499209463596344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11917182058095932, + 0.0719994381070137, + 0.1784645915031433, + 0.10241669416427612, + 0.07206495851278305, + 0.24750939011573792, + 0.04272564500570297, + 0.035980142652988434, + 0.020756477490067482, + 0.10891079902648926 + ], + "credit_map": [ + 0.008941815234720707, + 0.02046150527894497, + 0.015791811048984528, + 0.017344480380415916, + 0.013636651448905468, + 0.013021468184888363, + 0.01850876584649086, + 0.011465328745543957, + 0.004759203642606735, + 0.006038550287485123, + 0.005696858745068312, + 0.013647722080349922, + 0.009469188749790192, + 0.007052056957036257, + 0.013054870069026947, + 0.012280987575650215, + 0.011991330422461033, + 0.011860215105116367, + 0.017471587285399437, + 0.021243540570139885, + 0.021678339689970016, + 0.03133111074566841, + 0.02472708560526371, + 0.03816137835383415, + 0.008256027474999428, + 0.028960801661014557, + 0.013920441269874573, + 0.008047332987189293, + 0.00917551014572382, + 0.014274181798100471, + 0.00956275500357151, + 0.010219641029834747, + 0.004872772376984358, + 0.009405598975718021, + 0.009635788388550282, + 0.007377445697784424, + 0.007607869803905487, + 0.010803679935634136, + 0.01492873951792717, + 0.007433067075908184, + 0.019377553835511208, + 0.03737109899520874, + 0.040655311197042465, + 0.033409640192985535, + 0.023926017805933952, + 0.04181934520602226, + 0.021204806864261627, + 0.029745621606707573, + 0.003845423460006714, + 0.00691190455108881, + 0.004623353946954012, + 0.005645017605274916, + 0.004847755655646324, + 0.006015683989971876, + 0.006620700471103191, + 0.004215807653963566, + 0.004039444960653782, + 0.005146936047822237, + 0.004078997299075127, + 0.006446871906518936, + 0.0032408724073320627, + 0.004050521180033684, + 0.005376716610044241, + 0.0035997836384922266, + 0.002024070592597127, + 0.001808301080018282, + 0.00288861314766109, + 0.003044706303626299, + 0.002982196630910039, + 0.002210961189121008, + 0.0036441530101001263, + 0.002153475070372224, + 0.009106937795877457, + 0.00972188264131546, + 0.015088618732988834, + 0.014003492891788483, + 0.01318973395973444, + 0.02270975150167942, + 0.014590424485504627, + 0.010499958880245686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.05499209463596344 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029811281710863113 + ], + "opposite_logprob_tokens": [ + -3.529811382293701 + ], + "logprob_reward": -0.029811281710863113, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029811281710863113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1332383006811142, + 0.06913679838180542, + 0.4023233950138092, + 0.09092855453491211, + 0.0495280958712101, + 0.046521611511707306, + 0.0624469630420208, + 0.05203507840633392, + 0.040965497493743896, + 0.052875712513923645 + ], + "credit_map": [ + 0.008392558433115482, + 0.0306093692779541, + 0.02822147309780121, + 0.020563475787639618, + 0.009488734416663647, + 0.011456598527729511, + 0.014972958713769913, + 0.009533128701150417, + 0.0045226155780255795, + 0.007533187512308359, + 0.00793926790356636, + 0.015483623370528221, + 0.007342672906816006, + 0.008517993614077568, + 0.008190332911908627, + 0.009607103653252125, + 0.031579624861478806, + 0.035021670162677765, + 0.04303526133298874, + 0.04389435425400734, + 0.04915979504585266, + 0.056560128927230835, + 0.06835179775953293, + 0.07472075521945953, + 0.005626278463751078, + 0.010104681365191936, + 0.014973286539316177, + 0.012536564841866493, + 0.00915856845676899, + 0.012372702360153198, + 0.014492056332528591, + 0.011664415709674358, + 0.0035226449836045504, + 0.007713445462286472, + 0.008439640514552593, + 0.004981781356036663, + 0.005099333357065916, + 0.007831915281713009, + 0.007397508714348078, + 0.004541830159723759, + 0.003511660499498248, + 0.00874221883714199, + 0.005753834266215563, + 0.003353408770635724, + 0.008226423524320126, + 0.008193825371563435, + 0.005004273261874914, + 0.0037359660491347313, + 0.004154114052653313, + 0.007856999523937702, + 0.010927819646894932, + 0.007659571710973978, + 0.009174318984150887, + 0.009200074709951878, + 0.007441401481628418, + 0.006032660603523254, + 0.00727705005556345, + 0.004680802579969168, + 0.00794867891818285, + 0.007373488042503595, + 0.0039292327128350735, + 0.005693563260138035, + 0.005727487150579691, + 0.009404774755239487, + 0.004380908794701099, + 0.004341987892985344, + 0.007263378705829382, + 0.005189826246351004, + 0.0035352904815226793, + 0.006816925946623087, + 0.005432962905615568, + 0.00400421442463994, + 0.004076159093528986, + 0.0047837006859481335, + 0.007572639267891645, + 0.005353901069611311, + 0.008494448848068714, + 0.009128383360803127, + 0.008049137890338898, + 0.00541734742000699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.029811281710863113 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330601029098034 + ], + "opposite_logprob_tokens": [ + -3.773305892944336 + ], + "logprob_reward": -0.02330601029098034, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330601029098034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09358486533164978, + 0.03737812116742134, + 0.2330893725156784, + 0.15613867342472076, + 0.1794256865978241, + 0.08889095485210419, + 0.0525439977645874, + 0.10437503457069397, + 0.03169723600149155, + 0.022876037284731865 + ], + "credit_map": [ + 0.0062536816112697124, + 0.03231244534254074, + 0.030878329649567604, + 0.006574911065399647, + 0.004444279242306948, + 0.004198060370981693, + 0.004700610879808664, + 0.004222552757710218, + 0.002482255455106497, + 0.003758784616366029, + 0.0027468865737318993, + 0.006875672377645969, + 0.00476445909589529, + 0.004091739654541016, + 0.005276157986372709, + 0.0073821633122861385, + 0.017993884161114693, + 0.02305588126182556, + 0.032204609364271164, + 0.024792145937681198, + 0.022602293640375137, + 0.0319201797246933, + 0.037996482104063034, + 0.04252389073371887, + 0.007632852531969547, + 0.023157965391874313, + 0.020173175260424614, + 0.01927507482469082, + 0.02043129876255989, + 0.019868575036525726, + 0.025311831384897232, + 0.02028789557516575, + 0.012831627391278744, + 0.019963601604104042, + 0.023683572188019753, + 0.018754778429865837, + 0.018796483054757118, + 0.02732311561703682, + 0.037715137004852295, + 0.02035737968981266, + 0.007516189012676477, + 0.011371943168342113, + 0.015509058721363544, + 0.008812177926301956, + 0.01104578748345375, + 0.01280501764267683, + 0.011504300870001316, + 0.010326479561626911, + 0.004057369194924831, + 0.005944134201854467, + 0.005524341948330402, + 0.007027823477983475, + 0.007843855768442154, + 0.00718805892392993, + 0.008753632195293903, + 0.0062047806568443775, + 0.015869220718741417, + 0.008470465429127216, + 0.009981263428926468, + 0.016476906836032867, + 0.00992782972753048, + 0.009444013237953186, + 0.017396531999111176, + 0.016808802261948586, + 0.0030491764191538095, + 0.002882107626646757, + 0.005874283611774445, + 0.00347784161567688, + 0.0035607500467449427, + 0.004916646052151918, + 0.004938293248414993, + 0.002998136915266514, + 0.0025508874095976353, + 0.0027555779088288546, + 0.0024703098461031914, + 0.003048650221899152, + 0.00312410993501544, + 0.002538184868171811, + 0.003352176398038864, + 0.0030361423268914223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.02330601029098034 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630369160324335 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630369160324335, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630369160324335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5557416677474976, + 0.05980926379561424, + 0.04360077902674675, + 0.10952039808034897, + 0.04792070761322975, + 0.051807500422000885, + 0.031905438750982285, + 0.032108861953020096, + 0.025457464158535004, + 0.042127978056669235 + ], + "credit_map": [ + 0.03985363245010376, + 0.1881202608346939, + 0.1850004941225052, + 0.041343800723552704, + 0.027332082390785217, + 0.0229264535009861, + 0.02883468009531498, + 0.022330258041620255, + 0.004816037602722645, + 0.009132307954132557, + 0.004555147141218185, + 0.010485888458788395, + 0.005489781033247709, + 0.006126458290964365, + 0.009370461106300354, + 0.009833182208240032, + 0.0023722629994153976, + 0.004436874762177467, + 0.003493264550343156, + 0.006152979098260403, + 0.006686551496386528, + 0.0065803793258965015, + 0.007805848494172096, + 0.006072616670280695, + 0.006016591563820839, + 0.0218410212546587, + 0.013187580741941929, + 0.011412004008889198, + 0.01582644321024418, + 0.01426897943019867, + 0.014014476910233498, + 0.012953300960361958, + 0.003599283518269658, + 0.007915139198303223, + 0.005825065542012453, + 0.0044802092015743256, + 0.007430892437696457, + 0.007806881330907345, + 0.005692045670002699, + 0.0051711914129555225, + 0.005194396246224642, + 0.0070129395462572575, + 0.0060523818247020245, + 0.005729818716645241, + 0.006122208666056395, + 0.007417955901473761, + 0.008230959065258503, + 0.0060468376614153385, + 0.004110173787921667, + 0.00455485237762332, + 0.005003603640943766, + 0.003395230509340763, + 0.002867432776838541, + 0.005051637068390846, + 0.0035507935099303722, + 0.003371716244146228, + 0.003914820495992899, + 0.004965534899383783, + 0.00475308857858181, + 0.004564404487609863, + 0.0034251997712999582, + 0.0035699624568223953, + 0.0036191032268106937, + 0.0032967468723654747, + 0.002377359662204981, + 0.0021106204949319363, + 0.0036905999295413494, + 0.003407648764550686, + 0.0033432214986532927, + 0.002862551948055625, + 0.005306753795593977, + 0.0023587082978338003, + 0.004390215501189232, + 0.003726532217115164, + 0.004801948554813862, + 0.004288629163056612, + 0.007368321996182203, + 0.006567332893610001, + 0.0064744665287435055, + 0.0045105330646038055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.007630369160324335 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly descending, ritualistic overhead shot unfolds, opening from a solid wood tabletop with delicate dark wood grain and slowly approaching a handleless, thin-walled, transparent glass cup placed in its center. A soft ray of afternoon sunlight slants through the window, gently shimmering on the cup's walls and filtering through the clear, light amber tea within. As the camera continues to descend, we clearly see several previously dry, shrunken golden chrysanthemums. As if awakened from a deep slumber, the hot water gently and delicately unfolds their petals, layer by layer. This blossoming process is captured with exquisite precision, each petal gracefully stretching outward, ultimately fully blossoming within the tea, forming plump, delicate golden corollas. Their golden hue appears remarkably translucent and luminous in the light, with a soft halo emanating from the edges. The entire scene is bathed in a warm, soft beige hue, creating an atmosphere of profound tranquility, focus, and healing Zen.", + "seed": 13595, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_11/index_6698_seed_13595_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.27445386704057456, + "vlm_logprob": -0.27445386704057456, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.27445386704057456, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2257397174835205 + ], + "opposite_logprob_tokens": [ + -0.3507396876811981 + ], + "logprob_reward": -1.2257397174835205, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2257397174835205, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2944606840610504, + 0.09449200332164764, + 0.11654524505138397, + 0.05508652701973915, + 0.13409793376922607, + 0.09107791632413864, + 0.061096735298633575, + 0.08474867045879364, + 0.04388083145022392, + 0.024513499811291695 + ], + "credit_map": [ + 0.028331737965345383, + 0.034707553684711456, + 0.05044872313737869, + 0.03479462489485741, + 0.024936847388744354, + 0.03598382696509361, + 0.027217522263526917, + 0.058039817959070206, + 0.010606386698782444, + 0.015530581586062908, + 0.010134495794773102, + 0.011671542190015316, + 0.014103726483881474, + 0.012621021829545498, + 0.012219580821692944, + 0.007604654878377914, + 0.00940803810954094, + 0.012864619493484497, + 0.010697959922254086, + 0.008885560557246208, + 0.01720874384045601, + 0.04251088574528694, + 0.008089173585176468, + 0.006880251690745354, + 0.004886504262685776, + 0.006483681034296751, + 0.0044779591262340546, + 0.0038217201363295317, + 0.01448803674429655, + 0.010890771634876728, + 0.005412720143795013, + 0.00462512718513608, + 0.01592104136943817, + 0.017145320773124695, + 0.016025705263018608, + 0.014724298380315304, + 0.017119791358709335, + 0.019510537385940552, + 0.019840531051158905, + 0.01381069328635931, + 0.010383836925029755, + 0.01045634038746357, + 0.016098784282803535, + 0.011726231314241886, + 0.01101023517549038, + 0.014682943001389503, + 0.008758272975683212, + 0.00796126201748848, + 0.0044273980893194675, + 0.002963860984891653, + 0.0045410566963255405, + 0.010838647373020649, + 0.003181569743901491, + 0.014026298187673092, + 0.00985836610198021, + 0.011259532533586025, + 0.008293996565043926, + 0.007983681745827198, + 0.007576117757707834, + 0.009542490355670452, + 0.008566685020923615, + 0.01657419092953205, + 0.011157347820699215, + 0.015054150484502316, + 0.005128310061991215, + 0.005247772205621004, + 0.0067049977369606495, + 0.005078939720988274, + 0.0065250517800450325, + 0.003590199165046215, + 0.00641283905133605, + 0.0051927161403000355, + 0.0016351662343367934, + 0.004228550475090742, + 0.001948452671058476, + 0.003135927254334092, + 0.005050856154412031, + 0.002619560342282057, + 0.0029898276552557945, + 0.002905158093199134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -1.2257397174835205 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08910848945379257 + ], + "opposite_logprob_tokens": [ + -2.464108467102051 + ], + "logprob_reward": -0.08910848945379257, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08910848945379257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09503523260354996, + 0.06460535526275635, + 0.1062091588973999, + 0.024652749300003052, + 0.0433594286441803, + 0.14306119084358215, + 0.39916518330574036, + 0.03411141037940979, + 0.061343565583229065, + 0.02845664694905281 + ], + "credit_map": [ + 0.00855280738323927, + 0.011833340860903263, + 0.01181742548942566, + 0.010142057202756405, + 0.016302799805998802, + 0.011312403716146946, + 0.011218184605240822, + 0.013856221921741962, + 0.005258962046355009, + 0.008083381690084934, + 0.004958108998835087, + 0.009032174944877625, + 0.009443884715437889, + 0.013114127330482006, + 0.009668889455497265, + 0.005045829340815544, + 0.014334489591419697, + 0.029984641820192337, + 0.0072394805029034615, + 0.007119602523744106, + 0.007267438806593418, + 0.028603358194231987, + 0.006905023939907551, + 0.004755127709358931, + 0.0018043025629594922, + 0.002925676992163062, + 0.0028655303176492453, + 0.003103072987869382, + 0.004429381340742111, + 0.0035673396196216345, + 0.003815938252955675, + 0.002141509670764208, + 0.005694030784070492, + 0.005060598719865084, + 0.003962401766330004, + 0.004704730119556189, + 0.006073876284062862, + 0.005609809421002865, + 0.007526027038693428, + 0.004727956838905811, + 0.01694110967218876, + 0.015499159693717957, + 0.03207603469491005, + 0.023647258058190346, + 0.014652774669229984, + 0.014367425814270973, + 0.01214611902832985, + 0.013731329701840878, + 0.029369894415140152, + 0.016754508018493652, + 0.02395717054605484, + 0.029557915404438972, + 0.02093949355185032, + 0.1446714699268341, + 0.04124459996819496, + 0.09267017245292664, + 0.004489136394113302, + 0.004332548938691616, + 0.004111416172236204, + 0.003123672679066658, + 0.005071289837360382, + 0.005159890279173851, + 0.003628683974966407, + 0.0041947741992771626, + 0.009244708344340324, + 0.007920055650174618, + 0.007649356033653021, + 0.00697175832465291, + 0.0055694133043289185, + 0.00687512056902051, + 0.009494921192526817, + 0.007618238218128681, + 0.00399635499343276, + 0.004047737456858158, + 0.00319734588265419, + 0.0036631824914366007, + 0.0031624252442270517, + 0.004378152079880238, + 0.0034295793157070875, + 0.002581871347501874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.08910848945379257 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255018636584282 + ], + "opposite_logprob_tokens": [ + -4.387550354003906 + ], + "logprob_reward": -0.01255018636584282, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255018636584282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1484900861978531, + 0.08165495842695236, + 0.08282209932804108, + 0.0688481330871582, + 0.136493518948555, + 0.09201651066541672, + 0.09857930988073349, + 0.19077320396900177, + 0.08279973268508911, + 0.017522532492876053 + ], + "credit_map": [ + 0.014521163888275623, + 0.017044302076101303, + 0.019873108714818954, + 0.016937604174017906, + 0.020054779946804047, + 0.016729312017560005, + 0.016553211957216263, + 0.026776598766446114, + 0.012332231737673283, + 0.015702424570918083, + 0.00759555259719491, + 0.00896269641816616, + 0.010460104793310165, + 0.01335850078612566, + 0.007025812286883593, + 0.006217633839696646, + 0.008137394674122334, + 0.009496724233031273, + 0.013375949114561081, + 0.0073485421016812325, + 0.009436093270778656, + 0.020213063806295395, + 0.006364923436194658, + 0.008449399843811989, + 0.0070473686791956425, + 0.008431768044829369, + 0.0075613330118358135, + 0.008151788264513016, + 0.012282099574804306, + 0.009616577997803688, + 0.0089110117405653, + 0.006846183445304632, + 0.016589337959885597, + 0.01566511020064354, + 0.017770230770111084, + 0.017964165657758713, + 0.01780918799340725, + 0.016440222039818764, + 0.01977628469467163, + 0.014478979632258415, + 0.009450222365558147, + 0.007427658420056105, + 0.011764973402023315, + 0.013126152567565441, + 0.009400888346135616, + 0.01332107838243246, + 0.021235521882772446, + 0.0062900111079216, + 0.009708148427307606, + 0.007028623484075069, + 0.007862555794417858, + 0.010103357955813408, + 0.0056915911845862865, + 0.029092911630868912, + 0.019856486469507217, + 0.009235631674528122, + 0.02150566689670086, + 0.018994148820638657, + 0.018431534990668297, + 0.021388884633779526, + 0.017858508974313736, + 0.03605557605624199, + 0.021734613925218582, + 0.034804269671440125, + 0.01107985619455576, + 0.010379567742347717, + 0.011144875548779964, + 0.0077763511799275875, + 0.010690269060432911, + 0.00739994365721941, + 0.013329673558473587, + 0.010999191552400589, + 0.0015025283209979534, + 0.001431104727089405, + 0.003955800551921129, + 0.0011782851070165634, + 0.0015204873634502292, + 0.0014423980610445142, + 0.0012550167739391327, + 0.005236910656094551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.01255018636584282 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812966123223305 + ], + "opposite_logprob_tokens": [ + -3.2881295680999756 + ], + "logprob_reward": -0.03812966123223305, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812966123223305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1121850460767746, + 0.07145890593528748, + 0.04209510609507561, + 0.06015831232070923, + 0.07432039082050323, + 0.15322671830654144, + 0.2820540964603424, + 0.058147359639406204, + 0.0505221001803875, + 0.095831960439682 + ], + "credit_map": [ + 0.009395945817232132, + 0.012788227759301662, + 0.02196594886481762, + 0.015860842540860176, + 0.010512161999940872, + 0.011584065854549408, + 0.012949888594448566, + 0.01712797023355961, + 0.01053041685372591, + 0.014007930643856525, + 0.00517615769058466, + 0.007551606744527817, + 0.008586776442825794, + 0.011838700622320175, + 0.008155368268489838, + 0.005611952859908342, + 0.003713998943567276, + 0.006092095281928778, + 0.0061669559217989445, + 0.005770813673734665, + 0.0034921166952699423, + 0.00530609954148531, + 0.007496257312595844, + 0.004056769423186779, + 0.0045815808698534966, + 0.005192529875785112, + 0.009810969233512878, + 0.008787061087787151, + 0.010546145029366016, + 0.006250108126550913, + 0.007854120805859566, + 0.00713579636067152, + 0.008437460288405418, + 0.00830436684191227, + 0.009796369820833206, + 0.008889605291187763, + 0.010019269771873951, + 0.009609603323042393, + 0.01206386648118496, + 0.007199851330369711, + 0.016549929976463318, + 0.020493507385253906, + 0.016663430258631706, + 0.027929263189435005, + 0.016867848113179207, + 0.014391601085662842, + 0.024047711864113808, + 0.01628340780735016, + 0.02481762506067753, + 0.013539442792534828, + 0.01203643623739481, + 0.02364516630768776, + 0.015777017921209335, + 0.10736241191625595, + 0.04790631681680679, + 0.036969687789678574, + 0.007516880985349417, + 0.007426266558468342, + 0.00548262195661664, + 0.0063882796093821526, + 0.004342233296483755, + 0.011447055265307426, + 0.0061926268972456455, + 0.009351394139230251, + 0.006468411535024643, + 0.006165563128888607, + 0.007718282286077738, + 0.004608257673680782, + 0.0063578044064342976, + 0.004558068700134754, + 0.009759909473359585, + 0.004885801114141941, + 0.007918383926153183, + 0.008365161716938019, + 0.01919124647974968, + 0.010051972232758999, + 0.009215282276272774, + 0.007719516754150391, + 0.009101595729589462, + 0.024268805980682373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.03812966123223305 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741280667483807 + ], + "opposite_logprob_tokens": [ + -5.006741046905518 + ], + "logprob_reward": -0.006741280667483807, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741280667483807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2694849669933319, + 0.15435753762722015, + 0.050912611186504364, + 0.08660176396369934, + 0.11124727129936218, + 0.06196577101945877, + 0.08267509192228317, + 0.048981256783008575, + 0.02543029934167862, + 0.10834337025880814 + ], + "credit_map": [ + 0.02342209406197071, + 0.031710900366306305, + 0.04842734709382057, + 0.02646428719162941, + 0.026224488392472267, + 0.03759501129388809, + 0.026730455458164215, + 0.04891040921211243, + 0.024894921109080315, + 0.03164328634738922, + 0.01374563667923212, + 0.012461699545383453, + 0.02247754856944084, + 0.022595930844545364, + 0.01419536117464304, + 0.012343162670731544, + 0.005151467863470316, + 0.006451580207794905, + 0.006905199494212866, + 0.006982604041695595, + 0.00728306919336319, + 0.007424124050885439, + 0.006166541948914528, + 0.0045480248518288136, + 0.00658330786973238, + 0.00961947813630104, + 0.009191554971039295, + 0.006150853354483843, + 0.023154333233833313, + 0.014828910119831562, + 0.007398865185678005, + 0.009674468077719212, + 0.012893887236714363, + 0.013876710087060928, + 0.018598491325974464, + 0.010500707663595676, + 0.016302715986967087, + 0.016473276540637016, + 0.011183236725628376, + 0.011418257839977741, + 0.008419811725616455, + 0.005721628665924072, + 0.006739373318850994, + 0.007265815045684576, + 0.006667222362011671, + 0.009894270449876785, + 0.012091557495296001, + 0.005166096147149801, + 0.007586532738059759, + 0.004188684280961752, + 0.006950707640498877, + 0.007839953526854515, + 0.004556612111628056, + 0.022184817120432854, + 0.007846704684197903, + 0.021521084010601044, + 0.006725624203681946, + 0.005801905412226915, + 0.005261383950710297, + 0.005076166242361069, + 0.006254165433347225, + 0.0053600831888616085, + 0.00597677007317543, + 0.008525160141289234, + 0.0031754036899656057, + 0.003513480070978403, + 0.003640082897618413, + 0.0023863576352596283, + 0.003519263118505478, + 0.0021256909240037203, + 0.003492089919745922, + 0.0035779348108917475, + 0.007184580899775028, + 0.013578531332314014, + 0.019155582413077354, + 0.010852755047380924, + 0.014850827865302563, + 0.007870894856750965, + 0.010820094496011734, + 0.024030111730098724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.006741280667483807 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly descending, ritualistic overhead shot unfolds, opening from a solid wood tabletop with delicate dark wood grain and slowly approaching a handleless, thin-walled, transparent glass cup placed in its center. A soft ray of afternoon sunlight slants through the window, gently shimmering on the cup's walls and filtering through the clear, light amber tea within. As the camera continues to descend, we clearly see several previously dry, shrunken golden chrysanthemums. As if awakened from a deep slumber, the hot water gently and delicately unfolds their petals, layer by layer. This blossoming process is captured with exquisite precision, each petal gracefully stretching outward, ultimately fully blossoming within the tea, forming plump, delicate golden corollas. Their golden hue appears remarkably translucent and luminous in the light, with a soft halo emanating from the edges. The entire scene is bathed in a warm, soft beige hue, creating an atmosphere of profound tranquility, focus, and healing Zen.", + "seed": 13596, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_11/index_6698_seed_13596_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05211514532566071, + "vlm_logprob": -0.05211514532566071, + "vlm_yes_no_margin": 3.2249999761581423, + "vlm_token_credit": -0.05211514532566071, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1336575448513031 + ], + "opposite_logprob_tokens": [ + -2.133657455444336 + ], + "logprob_reward": -0.1336575448513031, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1336575448513031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06822127103805542, + 0.35096922516822815, + 0.07034259289503098, + 0.025854837149381638, + 0.12437572330236435, + 0.11191937327384949, + 0.11625663191080093, + 0.10154516249895096, + 0.012403067201375961, + 0.018112173303961754 + ], + "credit_map": [ + 0.005255628377199173, + 0.002850712276995182, + 0.0071006095968186855, + 0.020157093182206154, + 0.003969792742282152, + 0.0032707408536225557, + 0.016351550817489624, + 0.009265132248401642, + 0.012133212760090828, + 0.012997215613722801, + 0.07337121665477753, + 0.04751277342438698, + 0.014022648334503174, + 0.01426725834608078, + 0.12507426738739014, + 0.05159061402082443, + 0.002747158519923687, + 0.009993578307330608, + 0.0020503876730799675, + 0.00543550681322813, + 0.022059941664338112, + 0.010358426719903946, + 0.0074580335058271885, + 0.010239550843834877, + 0.0017150569474324584, + 0.0030036806128919125, + 0.0022904807701706886, + 0.005365653894841671, + 0.0023645779583603144, + 0.0027772027533501387, + 0.004860162269324064, + 0.003478019032627344, + 0.00849542673677206, + 0.013843885622918606, + 0.017821647226810455, + 0.016573337838053703, + 0.013112332671880722, + 0.013149956241250038, + 0.016261190176010132, + 0.02511792629957199, + 0.01056432630866766, + 0.022594375535845757, + 0.023294059559702873, + 0.00966368243098259, + 0.009659131988883018, + 0.012375467456877232, + 0.0070962305180728436, + 0.016672084107995033, + 0.01242914516478777, + 0.015764674171805382, + 0.013559846207499504, + 0.015192858874797821, + 0.01158630009740591, + 0.013576840050518513, + 0.01843339577317238, + 0.015713566914200783, + 0.009569436311721802, + 0.012481164187192917, + 0.015440229326486588, + 0.011959768831729889, + 0.011679637245833874, + 0.012920942157506943, + 0.01101596374064684, + 0.01647801324725151, + 0.001065125223249197, + 0.0016329374630004168, + 0.0011265779612585902, + 0.0023421617224812508, + 0.0017950035398826003, + 0.0014086543815210462, + 0.0015793811762705445, + 0.0014532239874824882, + 0.002309326082468033, + 0.0017324588261544704, + 0.00206632143817842, + 0.0026466120034456253, + 0.001636783592402935, + 0.0029767509549856186, + 0.001517373719252646, + 0.003226544475182891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 6 + ], + "selected_reward": -0.1336575448513031 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03814021870493889 + ], + "opposite_logprob_tokens": [ + -3.288140296936035 + ], + "logprob_reward": -0.03814021870493889, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03814021870493889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08108898997306824, + 0.16155864298343658, + 0.04366978630423546, + 0.04820431396365166, + 0.30693772435188293, + 0.08596247434616089, + 0.08957253396511078, + 0.06940668821334839, + 0.07682457566261292, + 0.03677427023649216 + ], + "credit_map": [ + 0.008299760520458221, + 0.004608238581568003, + 0.008491537533700466, + 0.019724609330296516, + 0.0072089675813913345, + 0.00544089637696743, + 0.016561172902584076, + 0.010753803886473179, + 0.0064144316129386425, + 0.005742969457060099, + 0.04348663240671158, + 0.024484829977154732, + 0.009928332641720772, + 0.010202975012362003, + 0.04507145658135414, + 0.016227005049586296, + 0.0025880211032927036, + 0.011329981498420238, + 0.0025262681301683187, + 0.005750857759267092, + 0.006144281011074781, + 0.004680221900343895, + 0.005649303086102009, + 0.005000852048397064, + 0.0026827086694538593, + 0.0046617123298347, + 0.0047553787007927895, + 0.009470614604651928, + 0.004222457762807608, + 0.003928852267563343, + 0.008947067894041538, + 0.009535519406199455, + 0.01861335150897503, + 0.03934623301029205, + 0.04989276081323624, + 0.03410516306757927, + 0.03123023733496666, + 0.030393145978450775, + 0.033569496124982834, + 0.06978733092546463, + 0.007391549181193113, + 0.018928024917840958, + 0.013050071895122528, + 0.00790080800652504, + 0.01148669421672821, + 0.010936111211776733, + 0.0065217954106628895, + 0.009747418574988842, + 0.006618996150791645, + 0.00980665534734726, + 0.012163585983216763, + 0.01717200316488743, + 0.008741574361920357, + 0.009442473761737347, + 0.01559898629784584, + 0.010028256103396416, + 0.009084506891667843, + 0.009423831477761269, + 0.009378968738019466, + 0.006940317340195179, + 0.007937232963740826, + 0.011456762440502644, + 0.006217063870280981, + 0.008968006819486618, + 0.007151366677135229, + 0.011355179361999035, + 0.008056347258388996, + 0.009148013778030872, + 0.00914773903787136, + 0.009375513531267643, + 0.00870515312999487, + 0.013885261490941048, + 0.004341185092926025, + 0.005369469523429871, + 0.005575929302722216, + 0.005008138250559568, + 0.00447609880939126, + 0.00452182674780488, + 0.002981157973408699, + 0.00450046407058835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.03814021870493889 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810935094952583 + ], + "opposite_logprob_tokens": [ + -3.288109302520752 + ], + "logprob_reward": -0.03810935094952583, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810935094952583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12251441925764084, + 0.24658572673797607, + 0.33116164803504944, + 0.06226452812552452, + 0.059522271156311035, + 0.04980067163705826, + 0.05927712842822075, + 0.0417509451508522, + 0.01259290985763073, + 0.014529747888445854 + ], + "credit_map": [ + 0.007475072983652353, + 0.005838352721184492, + 0.010445141233503819, + 0.028405645862221718, + 0.008861258625984192, + 0.006267095450311899, + 0.031458258628845215, + 0.023763610050082207, + 0.006817820481956005, + 0.011639723554253578, + 0.04990655928850174, + 0.016589751467108727, + 0.01687886379659176, + 0.009402196854352951, + 0.09447628259658813, + 0.04087453708052635, + 0.012219659052789211, + 0.07847340404987335, + 0.009326561354100704, + 0.01587061956524849, + 0.12909063696861267, + 0.047128938138484955, + 0.02082999236881733, + 0.018221866339445114, + 0.0028427881188690662, + 0.005895101930946112, + 0.004980116151273251, + 0.014766071923077106, + 0.005562221631407738, + 0.004884584806859493, + 0.013033761642873287, + 0.01029988657683134, + 0.004970172420144081, + 0.006086209323257208, + 0.0068437689915299416, + 0.009633238427340984, + 0.005776974838227034, + 0.006347158458083868, + 0.007040835916996002, + 0.0128239206969738, + 0.005543519277125597, + 0.005256996024399996, + 0.008873076178133488, + 0.00548808416351676, + 0.005398663692176342, + 0.0055082207545638084, + 0.0051427483558654785, + 0.008589368313550949, + 0.005407698452472687, + 0.008700817823410034, + 0.005961865186691284, + 0.008712368085980415, + 0.0061228894628584385, + 0.007757515646517277, + 0.008213909342885017, + 0.008400068618357182, + 0.004420469049364328, + 0.004099717829376459, + 0.00506580388173461, + 0.004733707755804062, + 0.00513824075460434, + 0.005632356740534306, + 0.0031482044141739607, + 0.009512444026768208, + 0.0013494340237230062, + 0.0013530434807762504, + 0.0013227162417024374, + 0.001698363572359085, + 0.001999237574636936, + 0.0016193011542782187, + 0.0011796329636126757, + 0.00207118084654212, + 0.001696492196060717, + 0.002166438614949584, + 0.0016936727333813906, + 0.0023371397983282804, + 0.002250580582767725, + 0.0014347027754411101, + 0.0014442682731896639, + 0.0015064531471580267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.03810935094952583 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038122545927762985 + ], + "opposite_logprob_tokens": [ + -3.2881226539611816 + ], + "logprob_reward": -0.038122545927762985, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038122545927762985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0766829326748848, + 0.0907629132270813, + 0.11298424750566483, + 0.08780314773321152, + 0.13375991582870483, + 0.14749807119369507, + 0.16744664311408997, + 0.09272096306085587, + 0.04246765375137329, + 0.047873564064502716 + ], + "credit_map": [ + 0.008543726056814194, + 0.005316621158272028, + 0.006432860158383846, + 0.016265833750367165, + 0.006635013967752457, + 0.007099592126905918, + 0.013720673508942127, + 0.012668611481785774, + 0.006360304541885853, + 0.0034765813034027815, + 0.025712639093399048, + 0.0131677919998765, + 0.004554962273687124, + 0.004303425550460815, + 0.020928507670760155, + 0.012258701026439667, + 0.005975872278213501, + 0.0391571968793869, + 0.004305251874029636, + 0.010313073173165321, + 0.029224824160337448, + 0.008065495640039444, + 0.00997614860534668, + 0.005966384895145893, + 0.005791506730020046, + 0.008020345121622086, + 0.0093266936019063, + 0.01925080455839634, + 0.008110717870295048, + 0.008360067382454872, + 0.016308708116412163, + 0.012634309940040112, + 0.008899006992578506, + 0.03665221482515335, + 0.007294693496078253, + 0.015621587634086609, + 0.016734112054109573, + 0.018755227327346802, + 0.01049086358398199, + 0.01931220106780529, + 0.009804144501686096, + 0.012195575051009655, + 0.02009170688688755, + 0.02651030384004116, + 0.024519087746739388, + 0.01847933605313301, + 0.018250999972224236, + 0.017646916210651398, + 0.022903865203261375, + 0.020997004583477974, + 0.019708743318915367, + 0.020178699865937233, + 0.01561366580426693, + 0.020117565989494324, + 0.025975262746214867, + 0.02195184864103794, + 0.01017940603196621, + 0.01092013530433178, + 0.010598861612379551, + 0.008947192691266537, + 0.01105812843888998, + 0.015057990327477455, + 0.009830530732870102, + 0.016128715127706528, + 0.0033387818839401007, + 0.005962601862847805, + 0.004347886890172958, + 0.005481847561895847, + 0.005980728659778833, + 0.00506315752863884, + 0.004973717965185642, + 0.007318931166082621, + 0.004292990081012249, + 0.005922539159655571, + 0.006000934634357691, + 0.007352548651397228, + 0.0067787314765155315, + 0.006491683889180422, + 0.004065337590873241, + 0.006968798115849495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.038122545927762985 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video start with a view of a solid wood tabletop with dark wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a handleless, thin-walled, transparent glass cup in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow descending overhead shot approaching the cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do golden chrysanthemums unfold their petals in the hot tea, forming translucent golden corollas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a warm, soft beige hue with afternoon sunlight shimmering on the cup?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254606619477272 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.01254606619477272, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254606619477272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.137479767203331, + 0.31700870394706726, + 0.15352606773376465, + 0.03655141219496727, + 0.1458110362291336, + 0.09567663818597794, + 0.015252282842993736, + 0.028902526944875717, + 0.024651046842336655, + 0.04514046758413315 + ], + "credit_map": [ + 0.009682318195700645, + 0.006518900860100985, + 0.009163636714220047, + 0.0168902687728405, + 0.013237289153039455, + 0.0075395312160253525, + 0.04684579372406006, + 0.027602020651102066, + 0.01004709955304861, + 0.008741294965147972, + 0.09038373082876205, + 0.030737966299057007, + 0.012415031902492046, + 0.011299366131424904, + 0.11846999824047089, + 0.03491424024105072, + 0.008493487723171711, + 0.05422576889395714, + 0.006086125038564205, + 0.009589903056621552, + 0.031356845051050186, + 0.01778753474354744, + 0.011985658667981625, + 0.014000746421515942, + 0.003320480929687619, + 0.0043095992878079414, + 0.0027347183786332607, + 0.0063070147298276424, + 0.003820385318249464, + 0.004046339076012373, + 0.006997942924499512, + 0.005014931317418814, + 0.009647050872445107, + 0.019039908424019814, + 0.02515777200460434, + 0.013612085022032261, + 0.011251197196543217, + 0.01550951786339283, + 0.012144604697823524, + 0.03944890573620796, + 0.007490175310522318, + 0.024731045588850975, + 0.010339866392314434, + 0.009268038906157017, + 0.013293064199388027, + 0.012709256261587143, + 0.007203706074506044, + 0.010641483590006828, + 0.0017143783625215292, + 0.0016866223886609077, + 0.002900954568758607, + 0.0016760872676968575, + 0.001526572392322123, + 0.002161725191399455, + 0.001896056579425931, + 0.001689885975793004, + 0.004004594404250383, + 0.00348974228836596, + 0.005615557078272104, + 0.003317847615107894, + 0.002687314758077264, + 0.003788307774811983, + 0.0028174554463475943, + 0.0031817061826586723, + 0.0025758985430002213, + 0.0033046561293303967, + 0.0021800044924020767, + 0.003650082740932703, + 0.00392620824277401, + 0.0026078480295836926, + 0.00290170731022954, + 0.003504641354084015, + 0.004375510849058628, + 0.0056483023799955845, + 0.004770951345562935, + 0.006482399068772793, + 0.005826265085488558, + 0.005630943924188614, + 0.0037917199078947306, + 0.008614378049969673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.01254606619477272 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_12/index_tensor([5639, 5639])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_12/index_tensor([5639, 5639])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..23d5627ba8790323bc3e87ffdf2cf34cb61b4b9a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_12/index_tensor([5639, 5639])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "drone scan begins at a high angle, slowly pulling back and panning right, revealing a majestic section of the Great Wall, winding through rolling, mist-shrouded green hills at dawn. This ancient fortification is rendered in surreal detail, its dappled gray stones revealing centuries of weathering through cracks, uneven surfaces, and deep green moss. From deep crevices between these stones, tiny luminous buds begin to miraculously emerge, emitting a soft inner glow. In fluid, time-lapse motion, these buds gracefully unfold into delicate, translucent blossoms, their petals radiating a vibrant spectrum of pink, white, and ethereal blue. This magical flowering begins as a small cluster, then spreads like a swift and silent wave of vibrant color, a cascading river of flowers that flows organically along the ancient fortification's zigzag spine and its watchtowers. The warm golden rays of the rising sun catch the dew on the petals, making them shimmer, while their own soft glow contrasts stunningly with the rugged, shadowed texture of the centuries-old stone, creating an atmosphere of breathtaking, serene magic as nature reclaims this massive structure.", + "seed": 4505, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_12/index_5639_seed_4505_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0995248556137085, + "vlm_logprob": -0.0995248556137085, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.0995248556137085, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09110389649868011 + ], + "opposite_logprob_tokens": [ + -2.5911037921905518 + ], + "logprob_reward": -0.09110389649868011, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09110389649868011, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08681240677833557, + 0.04107801243662834, + 0.035301193594932556, + 0.07178424298763275, + 0.17754045128822327, + 0.06320871412754059, + 0.39679670333862305, + 0.04545298591256142, + 0.05054293945431709, + 0.031482305377721786 + ], + "credit_map": [ + 0.005862704012542963, + 0.00931241363286972, + 0.014949781820178032, + 0.01363068725913763, + 0.01358423288911581, + 0.012306800112128258, + 0.009440785273909569, + 0.0077250138856470585, + 0.0021325305569916964, + 0.002370179630815983, + 0.00638890266418457, + 0.004731304943561554, + 0.006008915603160858, + 0.008591449819505215, + 0.007058371789753437, + 0.003796364413574338, + 0.0018377173691987991, + 0.0032087871804833412, + 0.00556443864479661, + 0.005605175159871578, + 0.006045971531420946, + 0.006593162659555674, + 0.003494021948426962, + 0.00295192189514637, + 0.0024101349990814924, + 0.003246681299060583, + 0.007746104616671801, + 0.0041241697035729885, + 0.02182563580572605, + 0.009329084306955338, + 0.004136035684496164, + 0.018966399133205414, + 0.006947935558855534, + 0.009443399496376514, + 0.023473918437957764, + 0.007982442155480385, + 0.07535356283187866, + 0.040144115686416626, + 0.005227935966104269, + 0.008967159315943718, + 0.0040502604097127914, + 0.005652945023030043, + 0.007151201833039522, + 0.007320373319089413, + 0.015004162676632404, + 0.012342266738414764, + 0.007416985929012299, + 0.004270524717867374, + 0.03875909745693207, + 0.0307135209441185, + 0.07612654566764832, + 0.044627513736486435, + 0.05303458869457245, + 0.0778069719672203, + 0.03306572884321213, + 0.042662810534238815, + 0.0037159540224820375, + 0.0033237324096262455, + 0.0077888029627501965, + 0.004794848617166281, + 0.008930372074246407, + 0.009184790775179863, + 0.003347010351717472, + 0.0043674795888364315, + 0.0033789444714784622, + 0.004397667478770018, + 0.009260612539947033, + 0.00601888308301568, + 0.009677085094153881, + 0.00853333342820406, + 0.0041190255433321, + 0.005157394800335169, + 0.0031950250267982483, + 0.0028816063422709703, + 0.004848369397222996, + 0.005252873990684748, + 0.002807921264320612, + 0.0035570019390434027, + 0.003448597388342023, + 0.005490911193192005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.09110389649868011 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008667940273880959 + ], + "opposite_logprob_tokens": [ + -4.758667945861816 + ], + "logprob_reward": -0.008667940273880959, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008667940273880959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05537375435233116, + 0.029004482552409172, + 0.0174544844776392, + 0.05207002907991409, + 0.15020005404949188, + 0.16196922957897186, + 0.30540943145751953, + 0.10630019009113312, + 0.08204220980405807, + 0.040176086127758026 + ], + "credit_map": [ + 0.003798802150413394, + 0.004352365620434284, + 0.014099139720201492, + 0.007065332494676113, + 0.006863945163786411, + 0.0074521019123494625, + 0.006205379031598568, + 0.005536685697734356, + 0.002018324099481106, + 0.0030510846991091967, + 0.004423106554895639, + 0.005439234897494316, + 0.00290549430064857, + 0.00347566488198936, + 0.005091760773211718, + 0.002599813975393772, + 0.0012651063734665513, + 0.0015255151083692908, + 0.002505985088646412, + 0.002861154731363058, + 0.0031563586089760065, + 0.0020417054183781147, + 0.001803424907848239, + 0.0022952354047447443, + 0.0015707715647295117, + 0.0018656618194654584, + 0.005206860136240721, + 0.002489553764462471, + 0.01746004819869995, + 0.007274061441421509, + 0.00241758581250906, + 0.013785486109554768, + 0.005716284736990929, + 0.010175527073442936, + 0.014836897142231464, + 0.009836799465119839, + 0.06029890105128288, + 0.03233768418431282, + 0.007202222011983395, + 0.009795738384127617, + 0.006290634628385305, + 0.011443398892879486, + 0.013478733599185944, + 0.010014607571065426, + 0.04901054501533508, + 0.05556563287973404, + 0.010232499800622463, + 0.005933176260441542, + 0.03022577054798603, + 0.029954703524708748, + 0.050878170877695084, + 0.04133540764451027, + 0.03859995678067207, + 0.0493927039206028, + 0.031704869121313095, + 0.033317871391773224, + 0.007741919718682766, + 0.00737585686147213, + 0.02297298237681389, + 0.009834365919232368, + 0.024696622043848038, + 0.01653219945728779, + 0.007211317308247089, + 0.009934925474226475, + 0.0065718223340809345, + 0.007060390897095203, + 0.017945220693945885, + 0.013534724712371826, + 0.009703682735562325, + 0.006337763275951147, + 0.010348263196647167, + 0.010540337301790714, + 0.0033112922683358192, + 0.002804854651913047, + 0.007177454419434071, + 0.0070211929269135, + 0.0031338315457105637, + 0.0026465023402124643, + 0.0060370550490915775, + 0.00804390199482441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.008667940273880959 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031996979378163815 + ], + "opposite_logprob_tokens": [ + -5.753199577331543 + ], + "logprob_reward": -0.0031996979378163815, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031996979378163815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03862239420413971, + 0.02537442557513714, + 0.02477412298321724, + 0.05987633019685745, + 0.05323532223701477, + 0.2856188118457794, + 0.3467405140399933, + 0.09341161698102951, + 0.05465679243206978, + 0.017689676955342293 + ], + "credit_map": [ + 0.002241067821159959, + 0.00320397294126451, + 0.008715550415217876, + 0.005746033973991871, + 0.004975609015673399, + 0.005626734346151352, + 0.005040542688220739, + 0.0030728804413229227, + 0.002327963011339307, + 0.003699282882735133, + 0.0031245779246091843, + 0.004988497123122215, + 0.002411887049674988, + 0.0027635223232209682, + 0.004205347970128059, + 0.0018533460097387433, + 0.001782119506970048, + 0.002780330367386341, + 0.0040740338154137135, + 0.0038448974955826998, + 0.0041368198581039906, + 0.003182873362675309, + 0.0023639407008886337, + 0.00260910764336586, + 0.0030906591564416885, + 0.0038165946025401354, + 0.007697926834225655, + 0.004546740557998419, + 0.01439649797976017, + 0.005745626054704189, + 0.0037418317515403032, + 0.01684045046567917, + 0.0024116800632327795, + 0.003239769022911787, + 0.0075288088992238045, + 0.0026253550313413143, + 0.02181081846356392, + 0.009583541192114353, + 0.002492342609912157, + 0.0035430039279162884, + 0.011406871490180492, + 0.015090364962816238, + 0.026499353349208832, + 0.03202312812209129, + 0.0716664120554924, + 0.08056843280792236, + 0.037116311490535736, + 0.011247904971241951, + 0.03434247151017189, + 0.035474348813295364, + 0.058069586753845215, + 0.04569508880376816, + 0.04387916997075081, + 0.0567181296646595, + 0.034564655274152756, + 0.037997033447027206, + 0.006444774568080902, + 0.005181405693292618, + 0.030548706650733948, + 0.010727820917963982, + 0.009538578800857067, + 0.013702942058444023, + 0.007430605124682188, + 0.009836772456765175, + 0.0040111057460308075, + 0.004603395238518715, + 0.01187643501907587, + 0.009171615354716778, + 0.008157112635672092, + 0.004817027132958174, + 0.006140733137726784, + 0.005879364442080259, + 0.0012884221505373716, + 0.0019158120267093182, + 0.00255411840043962, + 0.003071125829592347, + 0.0015593888238072395, + 0.0018531463574618101, + 0.002344548236578703, + 0.0031031144317239523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.0031996979378163815 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626701612025499 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626701612025499, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626701612025499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062006574124097824, + 0.03304685279726982, + 0.03722578287124634, + 0.016659462824463844, + 0.0714384913444519, + 0.14880602061748505, + 0.3496343791484833, + 0.06092013046145439, + 0.13802485167980194, + 0.08223745971918106 + ], + "credit_map": [ + 0.0041905296966433525, + 0.005526058375835419, + 0.009947160258889198, + 0.011941378004848957, + 0.008447340689599514, + 0.007600483950227499, + 0.008366286754608154, + 0.005987335462123156, + 0.0024111210368573666, + 0.004374141339212656, + 0.004473940934985876, + 0.006686036475002766, + 0.003996437881141901, + 0.003652597311884165, + 0.0049165282398462296, + 0.002536049112677574, + 0.0031249341554939747, + 0.0047788298688828945, + 0.004999860189855099, + 0.0064245169050991535, + 0.005015081260353327, + 0.004510022699832916, + 0.004820247180759907, + 0.0035522913094609976, + 0.0012324450071901083, + 0.0010462106438353658, + 0.0016532501904293895, + 0.0016826424980536103, + 0.00420518359169364, + 0.0030141519382596016, + 0.0012305452255532146, + 0.002595033496618271, + 0.004311568569391966, + 0.0086744399741292, + 0.00918737892061472, + 0.00940675474703312, + 0.013808822259306908, + 0.012699902057647705, + 0.005886885337531567, + 0.007462732028216124, + 0.011030000634491444, + 0.017480088397860527, + 0.01805242709815502, + 0.0190298892557621, + 0.030202645808458328, + 0.023096557706594467, + 0.01913948729634285, + 0.010774923488497734, + 0.034686584025621414, + 0.03198952600359917, + 0.06112852320075035, + 0.04345560073852539, + 0.047275152057409286, + 0.05853709205985069, + 0.03365888446569443, + 0.03890300542116165, + 0.00513393385335803, + 0.004136939067393541, + 0.01498380396515131, + 0.009029855020344257, + 0.0065824054181575775, + 0.010499934665858746, + 0.005394996143877506, + 0.0051582627929747105, + 0.01000719703733921, + 0.009518200531601906, + 0.034051645547151566, + 0.021388553082942963, + 0.022138971835374832, + 0.015142031013965607, + 0.013478055596351624, + 0.012300191447138786, + 0.00622358825057745, + 0.004669242072850466, + 0.01755092665553093, + 0.008387880399823189, + 0.006827548611909151, + 0.009475575760006905, + 0.018258189782500267, + 0.010844513773918152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.007626701612025499 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870260417461395 + ], + "opposite_logprob_tokens": [ + -1.137026071548462 + ], + "logprob_reward": -0.3870260417461395, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870260417461395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05441915988922119, + 0.03895376995205879, + 0.024986324831843376, + 0.08663525432348251, + 0.07317432016134262, + 0.09627000242471695, + 0.3374626636505127, + 0.0710904598236084, + 0.13786137104034424, + 0.07914665341377258 + ], + "credit_map": [ + 0.0033023071009665728, + 0.005553258117288351, + 0.006828270386904478, + 0.013650917448103428, + 0.006494893226772547, + 0.006463767495006323, + 0.008479905314743519, + 0.003645840100944042, + 0.002334322314709425, + 0.003923133946955204, + 0.005528191104531288, + 0.009055575355887413, + 0.004853168968111277, + 0.0038055451586842537, + 0.006342296488583088, + 0.0031115363817662, + 0.0018735523335635662, + 0.0016632411861792207, + 0.006307559087872505, + 0.0035998292732983828, + 0.00348457507789135, + 0.003276224248111248, + 0.0027232079301029444, + 0.0020581362769007683, + 0.003088618628680706, + 0.0040101357735693455, + 0.008051647804677486, + 0.004291709512472153, + 0.030491767451167107, + 0.01335559505969286, + 0.0042568170465528965, + 0.019088957458734512, + 0.002970667090266943, + 0.004011692013591528, + 0.008540130220353603, + 0.004008695483207703, + 0.027417728677392006, + 0.016324415802955627, + 0.005692662205547094, + 0.004208330996334553, + 0.008214175701141357, + 0.01114229578524828, + 0.01197579875588417, + 0.013362533412873745, + 0.015798049047589302, + 0.014056064188480377, + 0.01430516503751278, + 0.007415924221277237, + 0.033230673521757126, + 0.0275842547416687, + 0.06319139152765274, + 0.03777208551764488, + 0.045940544456243515, + 0.06229935586452484, + 0.029966922476887703, + 0.03747742995619774, + 0.0069928355515003204, + 0.005399249028414488, + 0.00878414697945118, + 0.010095669887959957, + 0.011582822538912296, + 0.014733487740159035, + 0.006720033008605242, + 0.00678221182897687, + 0.009009145200252533, + 0.007735772058367729, + 0.03746519982814789, + 0.01792083866894245, + 0.029310692101716995, + 0.015563140623271465, + 0.011138375848531723, + 0.009718202985823154, + 0.005905918776988983, + 0.004792156163603067, + 0.016909953206777573, + 0.00824062991887331, + 0.0058914124965667725, + 0.00757810240611434, + 0.021777404472231865, + 0.008051074109971523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.3870260417461395 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "drone scan begins at a high angle, slowly pulling back and panning right, revealing a majestic section of the Great Wall, winding through rolling, mist-shrouded green hills at dawn. This ancient fortification is rendered in surreal detail, its dappled gray stones revealing centuries of weathering through cracks, uneven surfaces, and deep green moss. From deep crevices between these stones, tiny luminous buds begin to miraculously emerge, emitting a soft inner glow. In fluid, time-lapse motion, these buds gracefully unfold into delicate, translucent blossoms, their petals radiating a vibrant spectrum of pink, white, and ethereal blue. This magical flowering begins as a small cluster, then spreads like a swift and silent wave of vibrant color, a cascading river of flowers that flows organically along the ancient fortification's zigzag spine and its watchtowers. The warm golden rays of the rising sun catch the dew on the petals, making them shimmer, while their own soft glow contrasts stunningly with the rugged, shadowed texture of the centuries-old stone, creating an atmosphere of breathtaking, serene magic as nature reclaims this massive structure.", + "seed": 4506, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_12/index_5639_seed_4506_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09209160930477082, + "vlm_logprob": -0.09209160930477082, + "vlm_yes_no_margin": 4.025000011920929, + "vlm_token_credit": -0.09209160930477082, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06084786728024483 + ], + "opposite_logprob_tokens": [ + -2.935847759246826 + ], + "logprob_reward": -0.06084786728024483, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06084786728024483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2539530396461487, + 0.18462209403514862, + 0.0681399255990982, + 0.06374664604663849, + 0.07199524343013763, + 0.0721789225935936, + 0.06723629683256149, + 0.06795282661914825, + 0.08249617367982864, + 0.06767883896827698 + ], + "credit_map": [ + 0.03397025540471077, + 0.04474693164229393, + 0.03401196748018265, + 0.036320094019174576, + 0.02608978934586048, + 0.022579554468393326, + 0.03236737847328186, + 0.02386706881225109, + 0.016784792765975, + 0.024419276043772697, + 0.02640318125486374, + 0.017543785274028778, + 0.03902995586395264, + 0.027252748608589172, + 0.017752565443515778, + 0.015435784123837948, + 0.006103011313825846, + 0.013030207715928555, + 0.009143457747995853, + 0.006781957112252712, + 0.008318929001688957, + 0.006329562049359083, + 0.008981081657111645, + 0.009451717138290405, + 0.006464140024036169, + 0.013700153678655624, + 0.009682520292699337, + 0.00551218818873167, + 0.009045838378369808, + 0.004741788841784, + 0.006999921519309282, + 0.00760009465739131, + 0.007170386612415314, + 0.008525513112545013, + 0.01032924372702837, + 0.008512463420629501, + 0.008371907286345959, + 0.010013909079134464, + 0.00854646135121584, + 0.010525356978178024, + 0.010364916175603867, + 0.0106618357822299, + 0.00994894653558731, + 0.0069471136666834354, + 0.008828051388263702, + 0.00825589057058096, + 0.0074848998337984085, + 0.009687271900475025, + 0.007781427353620529, + 0.010598489083349705, + 0.010432112962007523, + 0.0077720750123262405, + 0.009312603622674942, + 0.006379737984389067, + 0.004796669352799654, + 0.010163177736103535, + 0.005786137655377388, + 0.00988142192363739, + 0.012975594960153103, + 0.008015022613108158, + 0.0076148174703121185, + 0.006501404568552971, + 0.007655042689293623, + 0.009523382410407066, + 0.009992331266403198, + 0.009107762947678566, + 0.009787212125957012, + 0.009311688132584095, + 0.007509170565754175, + 0.015881380066275597, + 0.00890607200562954, + 0.012000557035207748, + 0.013033337891101837, + 0.00902246218174696, + 0.012295268476009369, + 0.006914646364748478, + 0.004904601257294416, + 0.0064147282391786575, + 0.005519721657037735, + 0.009574074298143387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.06084786728024483 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052717006765306 + ], + "opposite_logprob_tokens": [ + -5.255271911621094 + ], + "logprob_reward": -0.0052717006765306, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052717006765306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15178918838500977, + 0.06674009561538696, + 0.061911724507808685, + 0.022539086639881134, + 0.05635729432106018, + 0.07916367799043655, + 0.03880412131547928, + 0.09203050285577774, + 0.24522890150547028, + 0.18543535470962524 + ], + "credit_map": [ + 0.017619118094444275, + 0.023142701014876366, + 0.021691439673304558, + 0.022703558206558228, + 0.013873017393052578, + 0.014168309979140759, + 0.021237241104245186, + 0.01735379919409752, + 0.00532658863812685, + 0.009256946854293346, + 0.011045425198972225, + 0.00575398001819849, + 0.012430793605744839, + 0.010900945402681828, + 0.006090107839554548, + 0.005935311783105135, + 0.006347208749502897, + 0.012186401523649693, + 0.008625444956123829, + 0.0061564575880765915, + 0.007783922832459211, + 0.004495447501540184, + 0.00894345436245203, + 0.007373385597020388, + 0.0029358533211052418, + 0.0034566721878945827, + 0.003388047218322754, + 0.0024584324564784765, + 0.0025212832260876894, + 0.001821015845052898, + 0.0028970809653401375, + 0.0030607013031840324, + 0.005978315137326717, + 0.007856959477066994, + 0.008174804970622063, + 0.006178311072289944, + 0.007254201918840408, + 0.006686790380626917, + 0.006250015459954739, + 0.007977896369993687, + 0.011336436495184898, + 0.013065428473055363, + 0.01268193032592535, + 0.006950853858143091, + 0.010861774906516075, + 0.007211696822196245, + 0.00813191570341587, + 0.008923640474677086, + 0.004170251544564962, + 0.003634880529716611, + 0.005670823156833649, + 0.005543826147913933, + 0.005088422447443008, + 0.004056049045175314, + 0.004136691801249981, + 0.006503176409751177, + 0.009511428885161877, + 0.01620192639529705, + 0.014669996686279774, + 0.01064220629632473, + 0.01341316383332014, + 0.006910419557243586, + 0.011135119944810867, + 0.009546247310936451, + 0.032954782247543335, + 0.0343259796500206, + 0.03690450266003609, + 0.0311806071549654, + 0.0166416447609663, + 0.028409598395228386, + 0.03288332372903824, + 0.03192846477031708, + 0.021368300542235374, + 0.018886320292949677, + 0.025921059772372246, + 0.024066394194960594, + 0.01574895903468132, + 0.024131344631314278, + 0.02939484268426895, + 0.025918133556842804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0052717006765306 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004106185864657164 + ], + "opposite_logprob_tokens": [ + -5.504106044769287 + ], + "logprob_reward": -0.004106185864657164, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004106185864657164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08095227926969528, + 0.05317176878452301, + 0.05364282429218292, + 0.03280164301395416, + 0.025468112900853157, + 0.02643577754497528, + 0.04428322613239288, + 0.09212574362754822, + 0.1472303718328476, + 0.44388824701309204 + ], + "credit_map": [ + 0.012110361829400063, + 0.011615648865699768, + 0.013841159641742706, + 0.010781505145132542, + 0.005744314752519131, + 0.010111436247825623, + 0.008942600339651108, + 0.007805251982063055, + 0.004784324672073126, + 0.006498556584119797, + 0.008734225295484066, + 0.007908911444246769, + 0.007905597798526287, + 0.005281758960336447, + 0.006215461529791355, + 0.005842935293912888, + 0.004719976335763931, + 0.007476386614143848, + 0.008712498471140862, + 0.0052574859000742435, + 0.007149007637053728, + 0.0043930783867836, + 0.006495323497802019, + 0.009439065121114254, + 0.0034482539631426334, + 0.004879916086792946, + 0.004381750710308552, + 0.003481087973341346, + 0.003618523245677352, + 0.00293272128328681, + 0.005001783836632967, + 0.005057607311755419, + 0.0024378469679504633, + 0.0026036144699901342, + 0.0037559643387794495, + 0.003013653913512826, + 0.0030878905672580004, + 0.0035495779011398554, + 0.002855734433978796, + 0.004163830541074276, + 0.0027434458024799824, + 0.0029614195227622986, + 0.005111448001116514, + 0.002756912726908922, + 0.0036586951464414597, + 0.0026259783189743757, + 0.0032660861033946276, + 0.003311791457235813, + 0.006023467052727938, + 0.008283182978630066, + 0.0055234478786587715, + 0.004632709547877312, + 0.006435468327254057, + 0.0036028828471899033, + 0.0032333398703485727, + 0.006548726465553045, + 0.00884031131863594, + 0.011564822867512703, + 0.021267646923661232, + 0.008613118901848793, + 0.009580652229487896, + 0.009452523663640022, + 0.008230460807681084, + 0.014576206915080547, + 0.015069914981722832, + 0.01361057534813881, + 0.030008649453520775, + 0.021614614874124527, + 0.010497142560780048, + 0.014632133767008781, + 0.02286641299724579, + 0.01893092505633831, + 0.05860999971628189, + 0.04955045133829117, + 0.06948084384202957, + 0.055085934698581696, + 0.0366300493478775, + 0.05068635940551758, + 0.06137029454112053, + 0.06247428432106972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.004106185864657164 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031914988067001104 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.0031914988067001104, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031914988067001104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20839351415634155, + 0.12433432787656784, + 0.0701414942741394, + 0.056306056678295135, + 0.09569088369607925, + 0.06007549911737442, + 0.050774555653333664, + 0.08353932201862335, + 0.1847558319568634, + 0.06598851084709167 + ], + "credit_map": [ + 0.02455422468483448, + 0.03006223402917385, + 0.031203975901007652, + 0.030827123671770096, + 0.020567137748003006, + 0.01985945925116539, + 0.027211206033825874, + 0.024108154699206352, + 0.010024003684520721, + 0.01774490810930729, + 0.018984567373991013, + 0.01581750623881817, + 0.022023599594831467, + 0.01597423478960991, + 0.013286257162690163, + 0.010479261167347431, + 0.006518030073493719, + 0.01193223986774683, + 0.010100526735186577, + 0.009196368046104908, + 0.006045564543455839, + 0.004929517861455679, + 0.009147257544100285, + 0.012271991930902004, + 0.005552364978939295, + 0.006357936654239893, + 0.007007719483226538, + 0.006768791005015373, + 0.004402726888656616, + 0.006125248037278652, + 0.011098732240498066, + 0.008992533199489117, + 0.012286892160773277, + 0.016925569623708725, + 0.015107396990060806, + 0.009417052380740643, + 0.01050082128494978, + 0.010458401404321194, + 0.009278449229896069, + 0.011716301552951336, + 0.008100357837975025, + 0.008286179974675179, + 0.008251029998064041, + 0.00481790816411376, + 0.010864817537367344, + 0.005745128262788057, + 0.0053014918230473995, + 0.0087085897102952, + 0.006455425638705492, + 0.007785826455801725, + 0.008382751606404781, + 0.006106727756559849, + 0.008538654074072838, + 0.004329661838710308, + 0.0032448023557662964, + 0.0059307049959897995, + 0.007811881136149168, + 0.010165248066186905, + 0.012502502650022507, + 0.008512238040566444, + 0.01129845716059208, + 0.009627679362893105, + 0.00911704357713461, + 0.014504271559417248, + 0.018782567232847214, + 0.02087634988129139, + 0.030061163008213043, + 0.0263226181268692, + 0.01117040030658245, + 0.019942931830883026, + 0.03220604732632637, + 0.02539374865591526, + 0.007848088629543781, + 0.00697091780602932, + 0.010432691313326359, + 0.008155474439263344, + 0.005798080004751682, + 0.006491838488727808, + 0.010500753298401833, + 0.00979067012667656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0031914988067001104 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38704079389572144 + ], + "opposite_logprob_tokens": [ + -1.1370408535003662 + ], + "logprob_reward": -0.38704079389572144, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.38704079389572144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1127915307879448, + 0.11634761840105057, + 0.052302271127700806, + 0.06704410910606384, + 0.10888271778821945, + 0.12941499054431915, + 0.11905238777399063, + 0.19658434391021729, + 0.06011206656694412, + 0.037467990070581436 + ], + "credit_map": [ + 0.010587061755359173, + 0.014660158194601536, + 0.0161488838493824, + 0.016770513728260994, + 0.014290811493992805, + 0.011959107592701912, + 0.01625390537083149, + 0.012121091596782207, + 0.009376662783324718, + 0.01852378435432911, + 0.018139086663722992, + 0.014384124428033829, + 0.020039351657032967, + 0.015068919397890568, + 0.010414259508252144, + 0.010401432402431965, + 0.005839725956320763, + 0.00815889798104763, + 0.007217040751129389, + 0.005527586676180363, + 0.005893292836844921, + 0.0050065484829247, + 0.006125628016889095, + 0.008533550426363945, + 0.008388295769691467, + 0.015185846947133541, + 0.00929059274494648, + 0.006038766819983721, + 0.008902435190975666, + 0.005454323720186949, + 0.007252577226608992, + 0.006531271152198315, + 0.01422568503767252, + 0.020350992679595947, + 0.016630064696073532, + 0.0098673515021801, + 0.011943256482481956, + 0.008084708824753761, + 0.017428983002901077, + 0.010351686738431454, + 0.01646474562585354, + 0.01755332387983799, + 0.018826698884367943, + 0.011754918843507767, + 0.015000492334365845, + 0.01871904544532299, + 0.015606316737830639, + 0.015489441342651844, + 0.015435542911291122, + 0.02100740559399128, + 0.015211676247417927, + 0.01403866894543171, + 0.015039262361824512, + 0.0102608697488904, + 0.009965415112674236, + 0.018093546852469444, + 0.018512429669499397, + 0.02650093473494053, + 0.027704082429409027, + 0.028395215049386024, + 0.024328667670488358, + 0.01749270036816597, + 0.02950696088373661, + 0.024143341928720474, + 0.008363224565982819, + 0.00988225918263197, + 0.011006752960383892, + 0.004095801617950201, + 0.0042680371552705765, + 0.005102202296257019, + 0.0061124470084905624, + 0.011281341314315796, + 0.004540449008345604, + 0.0052055977284908295, + 0.004325319547206163, + 0.003713856218382716, + 0.0037347329780459404, + 0.006815593224018812, + 0.004875360522419214, + 0.0042570787481963634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.38704079389572144 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "drone scan begins at a high angle, slowly pulling back and panning right, revealing a majestic section of the Great Wall, winding through rolling, mist-shrouded green hills at dawn. This ancient fortification is rendered in surreal detail, its dappled gray stones revealing centuries of weathering through cracks, uneven surfaces, and deep green moss. From deep crevices between these stones, tiny luminous buds begin to miraculously emerge, emitting a soft inner glow. In fluid, time-lapse motion, these buds gracefully unfold into delicate, translucent blossoms, their petals radiating a vibrant spectrum of pink, white, and ethereal blue. This magical flowering begins as a small cluster, then spreads like a swift and silent wave of vibrant color, a cascading river of flowers that flows organically along the ancient fortification's zigzag spine and its watchtowers. The warm golden rays of the rising sun catch the dew on the petals, making them shimmer, while their own soft glow contrasts stunningly with the rugged, shadowed texture of the centuries-old stone, creating an atmosphere of breathtaking, serene magic as nature reclaims this massive structure.", + "seed": 4507, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_12/index_5639_seed_4507_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.101539048505947, + "vlm_logprob": -0.101539048505947, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.101539048505947, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02278807945549488 + ], + "opposite_logprob_tokens": [ + -4.147788047790527 + ], + "logprob_reward": -0.02278807945549488, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02278807945549488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22776159644126892, + 0.07021000981330872, + 0.2878713309764862, + 0.04039791598916054, + 0.12405341118574142, + 0.01939285174012184, + 0.09051819890737534, + 0.03093133307993412, + 0.04831959679722786, + 0.060543764382600784 + ], + "credit_map": [ + 0.026672255247831345, + 0.030634338036179543, + 0.037443388253450394, + 0.025789963081479073, + 0.03763975203037262, + 0.024775272235274315, + 0.017602428793907166, + 0.02720419317483902, + 0.007072791922837496, + 0.0055985040962696075, + 0.006721626501530409, + 0.017219001427292824, + 0.005970284342765808, + 0.006333968136459589, + 0.0089196115732193, + 0.012374218553304672, + 0.017105335369706154, + 0.011313050985336304, + 0.02172115631401539, + 0.0908869057893753, + 0.01744738221168518, + 0.025121290236711502, + 0.030632639303803444, + 0.07364356517791748, + 0.005034177098423243, + 0.0030663381330668926, + 0.005006975959986448, + 0.004438049625605345, + 0.006696032825857401, + 0.005552274640649557, + 0.004829742480069399, + 0.00577432569116354, + 0.013726607896387577, + 0.012638548389077187, + 0.015892067924141884, + 0.019824298098683357, + 0.016705790534615517, + 0.016866711899638176, + 0.01589067280292511, + 0.012508709914982319, + 0.0017732396954670548, + 0.002536121057346463, + 0.002320735715329647, + 0.002549203112721443, + 0.003078279783949256, + 0.002026225673034787, + 0.0029615985695272684, + 0.0021474480163306, + 0.007613716181367636, + 0.006780585274100304, + 0.00891347136348486, + 0.022065354511141777, + 0.006787913851439953, + 0.00791161134839058, + 0.020997319370508194, + 0.00944822933524847, + 0.0054826391860842705, + 0.0042661321349442005, + 0.004228544421494007, + 0.0030108618084341288, + 0.0031270263716578484, + 0.003167749149724841, + 0.0038877148181200027, + 0.003760664025321603, + 0.004679139703512192, + 0.003791408147662878, + 0.008625362068414688, + 0.0050422074273228645, + 0.004340601619333029, + 0.005308805499225855, + 0.007588828448206186, + 0.008943245746195316, + 0.00470438739284873, + 0.006006996147334576, + 0.011443011462688446, + 0.006903353147208691, + 0.00807628221809864, + 0.0054167103953659534, + 0.011132020503282547, + 0.0068610068410634995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.02278807945549488 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002220190828666091 + ], + "opposite_logprob_tokens": [ + -6.127220153808594 + ], + "logprob_reward": -0.002220190828666091, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002220190828666091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08412621915340424, + 0.06699599325656891, + 0.1367584764957428, + 0.13318473100662231, + 0.05429850518703461, + 0.0874144583940506, + 0.07564719766378403, + 0.08606592565774918, + 0.15997529029846191, + 0.11553321778774261 + ], + "credit_map": [ + 0.0072365873493254185, + 0.00827279407531023, + 0.01596202701330185, + 0.009961352683603764, + 0.013750697486102581, + 0.010858272202312946, + 0.008692005649209023, + 0.009392489679157734, + 0.006600601598620415, + 0.004850954748690128, + 0.005635909736156464, + 0.018086697906255722, + 0.004721869248896837, + 0.005597215611487627, + 0.00767407426610589, + 0.013828671537339687, + 0.00807508360594511, + 0.005534439347684383, + 0.009539096616208553, + 0.042045723646879196, + 0.008035515435039997, + 0.012676510959863663, + 0.014117352664470673, + 0.03673475235700607, + 0.008702977560460567, + 0.009187651798129082, + 0.017513681203126907, + 0.014559110626578331, + 0.01911752298474312, + 0.02521524578332901, + 0.01693839021027088, + 0.021950142458081245, + 0.004787227138876915, + 0.004649946931749582, + 0.00538474740460515, + 0.005561120808124542, + 0.00896306149661541, + 0.010294007137417793, + 0.008570547215640545, + 0.006087843328714371, + 0.007968653924763203, + 0.00698321545496583, + 0.012009486556053162, + 0.01726205088198185, + 0.012005750089883804, + 0.009605839848518372, + 0.010169205255806446, + 0.01141025684773922, + 0.005878895055502653, + 0.006014620885252953, + 0.006008617579936981, + 0.018946507945656776, + 0.007083237171173096, + 0.0075919777154922485, + 0.016855228692293167, + 0.007268110755831003, + 0.011316025629639626, + 0.009433919563889503, + 0.00845132302492857, + 0.009674514643847942, + 0.009650489315390587, + 0.008607427589595318, + 0.018959609791636467, + 0.009972622618079185, + 0.019594581797719002, + 0.01323939673602581, + 0.021952876821160316, + 0.015836387872695923, + 0.013775525614619255, + 0.01769387349486351, + 0.02559617906808853, + 0.03228646144270897, + 0.012190742418169975, + 0.01317526027560234, + 0.016970869153738022, + 0.011202272027730942, + 0.016312429681420326, + 0.009589675813913345, + 0.020002124831080437, + 0.016089845448732376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.002220190828666091 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017201402224600315 + ], + "opposite_logprob_tokens": [ + -6.376719951629639 + ], + "logprob_reward": -0.0017201402224600315, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017201402224600315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1332278549671173, + 0.04397323355078697, + 0.04654663801193237, + 0.060688432306051254, + 0.03137779235839844, + 0.06711606681346893, + 0.0687689483165741, + 0.07625743746757507, + 0.14186926186084747, + 0.3301743268966675 + ], + "credit_map": [ + 0.011569162830710411, + 0.016738539561629295, + 0.02362992614507675, + 0.01584763266146183, + 0.02367982268333435, + 0.01635063998401165, + 0.00959409773349762, + 0.01581803523004055, + 0.005406196229159832, + 0.0034499589819461107, + 0.004415447358042002, + 0.009211676195263863, + 0.004619470331817865, + 0.004727497696876526, + 0.007518390659242868, + 0.004624599125236273, + 0.0030484781600534916, + 0.0026208963245153427, + 0.0056579397059977055, + 0.013451218605041504, + 0.004382073879241943, + 0.004336316138505936, + 0.010352184996008873, + 0.0026975288055837154, + 0.006030954420566559, + 0.00584616931155324, + 0.00733719440177083, + 0.006965556647628546, + 0.00931429024785757, + 0.008279108442366123, + 0.007271031383424997, + 0.009644127450883389, + 0.0027772702742367983, + 0.002612612210214138, + 0.003333861008286476, + 0.003947682678699493, + 0.0063356379978358746, + 0.003698350628837943, + 0.004624382127076387, + 0.004047994967550039, + 0.006268329918384552, + 0.006560466252267361, + 0.010042806155979633, + 0.01055236253887415, + 0.010344387963414192, + 0.007111168000847101, + 0.007434885483235121, + 0.008801658637821674, + 0.006932355929166079, + 0.005959848873317242, + 0.00847476627677679, + 0.0102164875715971, + 0.008022946305572987, + 0.006989338900893927, + 0.013494540937244892, + 0.00867866724729538, + 0.014936108142137527, + 0.011203628033399582, + 0.009696641005575657, + 0.008675000630319118, + 0.006975650787353516, + 0.006577459629625082, + 0.012248069979250431, + 0.005944880191236734, + 0.01741654984652996, + 0.013448023237287998, + 0.027307337149977684, + 0.01720792055130005, + 0.014088542200624943, + 0.01531901117414236, + 0.016697270795702934, + 0.020384592935442924, + 0.030176810920238495, + 0.038412246853113174, + 0.05201571434736252, + 0.04282616823911667, + 0.04251036420464516, + 0.03311210870742798, + 0.05519441142678261, + 0.03592650219798088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0017201402224600315 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732400041073561 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.006732400041073561, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732400041073561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1456000953912735, + 0.038901057094335556, + 0.38019564747810364, + 0.044824566692113876, + 0.0360836461186409, + 0.06783388555049896, + 0.08053373545408249, + 0.06849794089794159, + 0.05135451629757881, + 0.0861748531460762 + ], + "credit_map": [ + 0.011135836131870747, + 0.01715795136988163, + 0.019506284967064857, + 0.018770458176732063, + 0.02624320797622204, + 0.021676355972886086, + 0.013297278434038162, + 0.017812730744481087, + 0.004399956203997135, + 0.00327153317630291, + 0.005031906068325043, + 0.0046254713088274, + 0.006471648346632719, + 0.003891980042681098, + 0.007103337440639734, + 0.004105222877115011, + 0.022220689803361893, + 0.0146086560562253, + 0.033547792583703995, + 0.12235350906848907, + 0.021376438438892365, + 0.03303898125886917, + 0.02615687996149063, + 0.10689270496368408, + 0.004258075729012489, + 0.003891912056133151, + 0.006846684031188488, + 0.0061638508923351765, + 0.007550427690148354, + 0.00511776190251112, + 0.006027800031006336, + 0.004968053195625544, + 0.0028824862092733383, + 0.0034263087436556816, + 0.004196271765977144, + 0.0032821306958794594, + 0.004991806577891111, + 0.005716073326766491, + 0.005539935082197189, + 0.006048630923032761, + 0.006501189898699522, + 0.007195471785962582, + 0.008206559345126152, + 0.009337713941931725, + 0.012635238468647003, + 0.00940446276217699, + 0.008311686106026173, + 0.006241559516638517, + 0.009584619663655758, + 0.010311681777238846, + 0.010720457881689072, + 0.010006571188569069, + 0.010777558200061321, + 0.0077661373652517796, + 0.012598445639014244, + 0.008768255822360516, + 0.009654725901782513, + 0.007227130699902773, + 0.009719633497297764, + 0.008333159610629082, + 0.007843013852834702, + 0.007307223044335842, + 0.010832632891833782, + 0.007580428384244442, + 0.00697817699983716, + 0.005253925919532776, + 0.0059385839849710464, + 0.005645164754241705, + 0.0064725312404334545, + 0.007142491638660431, + 0.007468436844646931, + 0.006455203052610159, + 0.007928211241960526, + 0.011540773324668407, + 0.013256172649562359, + 0.010650699958205223, + 0.009280350990593433, + 0.006935464218258858, + 0.016147224232554436, + 0.010435955598950386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.006732400041073561 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742344319820404 + ], + "opposite_logprob_tokens": [ + -0.9742344617843628 + ], + "logprob_reward": -0.4742344319820404, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742344319820404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11055149883031845, + 0.05852305144071579, + 0.40575823187828064, + 0.03473614156246185, + 0.05128980055451393, + 0.07371464371681213, + 0.15982814133167267, + 0.017704572528600693, + 0.04026256501674652, + 0.04763130098581314 + ], + "credit_map": [ + 0.00916559062898159, + 0.013632390648126602, + 0.014443673193454742, + 0.011926109902560711, + 0.017064491286873817, + 0.01481026504188776, + 0.012220842763781548, + 0.01728813908994198, + 0.004589144606143236, + 0.004410784691572189, + 0.007024663500487804, + 0.009799695573747158, + 0.007485216949135065, + 0.005164336413145065, + 0.009627105668187141, + 0.01042210217565298, + 0.023992598056793213, + 0.015483209863305092, + 0.037521108984947205, + 0.12987811863422394, + 0.023641126230359077, + 0.03486798331141472, + 0.028580915182828903, + 0.11179319024085999, + 0.00443572411313653, + 0.002406186191365123, + 0.004657464567571878, + 0.006130414083600044, + 0.0071827108040452, + 0.003349319100379944, + 0.003880321979522705, + 0.0026940007228404284, + 0.006378555204719305, + 0.006005728617310524, + 0.004757749382406473, + 0.006803383585065603, + 0.007948416285216808, + 0.007005177438259125, + 0.006829300429672003, + 0.005561484955251217, + 0.007275667041540146, + 0.005785355810075998, + 0.009216120466589928, + 0.015060130506753922, + 0.008934756740927696, + 0.00915450882166624, + 0.010689038783311844, + 0.00759906554594636, + 0.013045864179730415, + 0.011241784319281578, + 0.01605086214840412, + 0.03649062290787697, + 0.013393583707511425, + 0.014114332385361195, + 0.03788209334015846, + 0.017608996480703354, + 0.0019651453476399183, + 0.002046918496489525, + 0.0023180716671049595, + 0.0017518907552585006, + 0.0022432177793234587, + 0.0017913217889145017, + 0.0038465384859591722, + 0.0017414689064025879, + 0.008780671283602715, + 0.004827840253710747, + 0.0058695063926279545, + 0.004420311655849218, + 0.003104287898167968, + 0.003893927438184619, + 0.004864333663135767, + 0.004501683637499809, + 0.005000143311917782, + 0.005696112290024757, + 0.006800384726375341, + 0.007164254784584045, + 0.00559220090508461, + 0.004082840401679277, + 0.007036727387458086, + 0.006258637644350529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.4742344319820404 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "drone scan begins at a high angle, slowly pulling back and panning right, revealing a majestic section of the Great Wall, winding through rolling, mist-shrouded green hills at dawn. This ancient fortification is rendered in surreal detail, its dappled gray stones revealing centuries of weathering through cracks, uneven surfaces, and deep green moss. From deep crevices between these stones, tiny luminous buds begin to miraculously emerge, emitting a soft inner glow. In fluid, time-lapse motion, these buds gracefully unfold into delicate, translucent blossoms, their petals radiating a vibrant spectrum of pink, white, and ethereal blue. This magical flowering begins as a small cluster, then spreads like a swift and silent wave of vibrant color, a cascading river of flowers that flows organically along the ancient fortification's zigzag spine and its watchtowers. The warm golden rays of the rising sun catch the dew on the petals, making them shimmer, while their own soft glow contrasts stunningly with the rugged, shadowed texture of the centuries-old stone, creating an atmosphere of breathtaking, serene magic as nature reclaims this massive structure.", + "seed": 4508, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_12/index_5639_seed_4508_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10159280886873603, + "vlm_logprob": -0.10159280886873603, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.10159280886873603, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05714833363890648 + ], + "opposite_logprob_tokens": [ + -3.0571482181549072 + ], + "logprob_reward": -0.05714833363890648, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05714833363890648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1924629956483841, + 0.11563540995121002, + 0.09549865871667862, + 0.03304605185985565, + 0.06524674594402313, + 0.2599758803844452, + 0.045576442033052444, + 0.01828858070075512, + 0.019212033599615097, + 0.15505719184875488 + ], + "credit_map": [ + 0.023656649515032768, + 0.03249542787671089, + 0.02748761884868145, + 0.022546809166669846, + 0.021581824868917465, + 0.02912503480911255, + 0.02382056787610054, + 0.011749082244932652, + 0.014453815296292305, + 0.01717405579984188, + 0.011785081587731838, + 0.013902511447668076, + 0.025898268446326256, + 0.015022744424641132, + 0.00957624614238739, + 0.007822692394256592, + 0.013251684606075287, + 0.013802308589220047, + 0.010204505175352097, + 0.01170152798295021, + 0.016369475051760674, + 0.013110428117215633, + 0.009908132255077362, + 0.007150604855269194, + 0.0031621241942048073, + 0.0033226381056010723, + 0.003678226377815008, + 0.0034245033748447895, + 0.005216069053858519, + 0.005794398952275515, + 0.004992673173546791, + 0.003455422818660736, + 0.005542555823922157, + 0.008123583160340786, + 0.006046577822417021, + 0.008871374651789665, + 0.011478077620267868, + 0.009645657613873482, + 0.006486332044005394, + 0.009052589535713196, + 0.010352338664233685, + 0.010294520296156406, + 0.010338989086449146, + 0.09978018701076508, + 0.009612015448510647, + 0.0823468267917633, + 0.02754691243171692, + 0.00970409344881773, + 0.004222552292048931, + 0.005297484342008829, + 0.00481785275042057, + 0.006653954274952412, + 0.004784740041941404, + 0.005425573792308569, + 0.0075460029765963554, + 0.006828285753726959, + 0.0018318617949262261, + 0.0015531072858721018, + 0.002238500164821744, + 0.001766762463375926, + 0.0016405549831688404, + 0.0025024053175002337, + 0.002363210776820779, + 0.0043921805918216705, + 0.0016032602870836854, + 0.001506809494458139, + 0.001973637379705906, + 0.00190206419210881, + 0.0023665300104767084, + 0.0019456733716651797, + 0.0043205083347857, + 0.003593550994992256, + 0.008038097061216831, + 0.023050013929605484, + 0.010044890455901623, + 0.014935126528143883, + 0.017343904823064804, + 0.019065357744693756, + 0.030319467186927795, + 0.032260339707136154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.05714833363890648 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004673866089433432 + ], + "opposite_logprob_tokens": [ + -5.379673957824707 + ], + "logprob_reward": -0.004673866089433432, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004673866089433432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11648423969745636, + 0.06184600293636322, + 0.14832010865211487, + 0.06981385499238968, + 0.04714782163500786, + 0.06589330732822418, + 0.15714038908481598, + 0.09545355290174484, + 0.06474444270133972, + 0.17315633594989777 + ], + "credit_map": [ + 0.01237562671303749, + 0.01235940121114254, + 0.01746261678636074, + 0.014705682173371315, + 0.01438014954328537, + 0.01778944954276085, + 0.01722612977027893, + 0.010185186751186848, + 0.007868975400924683, + 0.007508676499128342, + 0.01016157679259777, + 0.00807911530137062, + 0.010452637448906898, + 0.00838976725935936, + 0.005640325136482716, + 0.0037449311930686235, + 0.015725914388895035, + 0.018515681847929955, + 0.02129317820072174, + 0.02653534896671772, + 0.018655797466635704, + 0.017679277807474136, + 0.015814580023288727, + 0.014100329019129276, + 0.006808130536228418, + 0.00785616971552372, + 0.0070462762378156185, + 0.0068252976052463055, + 0.012012826278805733, + 0.010275784879922867, + 0.011719832196831703, + 0.007269532419741154, + 0.003814432770013809, + 0.004067497793585062, + 0.003549301065504551, + 0.005218860227614641, + 0.005425968673080206, + 0.010051362216472626, + 0.007191138342022896, + 0.00782926194369793, + 0.004210060928016901, + 0.007010679692029953, + 0.004368934780359268, + 0.01971207931637764, + 0.006872439291328192, + 0.011052205227315426, + 0.0074210693128407, + 0.00524583924561739, + 0.010338502936065197, + 0.011136017739772797, + 0.011466857977211475, + 0.023889979347586632, + 0.016407037153840065, + 0.019492117688059807, + 0.025320131331682205, + 0.03908974677324295, + 0.007226626854389906, + 0.009105134755373001, + 0.008993502706289291, + 0.007047609426081181, + 0.016991768032312393, + 0.018988629803061485, + 0.013007540255784988, + 0.014092744328081608, + 0.005349252372980118, + 0.01020997203886509, + 0.006669973488897085, + 0.007237851619720459, + 0.007660962641239166, + 0.008042762987315655, + 0.007570739835500717, + 0.012002924457192421, + 0.011680880561470985, + 0.019760772585868835, + 0.01860141195356846, + 0.01693948730826378, + 0.026482388377189636, + 0.024664200842380524, + 0.023450350388884544, + 0.03157683089375496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.004673866089433432 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646431379020214 + ], + "opposite_logprob_tokens": [ + -4.758646488189697 + ], + "logprob_reward": -0.008646431379020214, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646431379020214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14177444577217102, + 0.04705478250980377, + 0.08982165157794952, + 0.08701440691947937, + 0.07365349680185318, + 0.13704751431941986, + 0.08657269924879074, + 0.10609091818332672, + 0.06808056682348251, + 0.1628894805908203 + ], + "credit_map": [ + 0.016877632588148117, + 0.015241963788866997, + 0.021409645676612854, + 0.020328108221292496, + 0.013716860674321651, + 0.0199714545160532, + 0.02317115105688572, + 0.01105762179940939, + 0.004246744327247143, + 0.005225569475442171, + 0.008070133626461029, + 0.006700789090245962, + 0.005946806166321039, + 0.006253641098737717, + 0.0063765873201191425, + 0.004234509542584419, + 0.010841074399650097, + 0.009123923256993294, + 0.011990534141659737, + 0.010873214341700077, + 0.009980959817767143, + 0.013661750592291355, + 0.012274544686079025, + 0.011075649410486221, + 0.008184341713786125, + 0.010187864303588867, + 0.008568531833589077, + 0.010375211015343666, + 0.011418981477618217, + 0.011907109059393406, + 0.015264065936207771, + 0.011108305305242538, + 0.006283602211624384, + 0.007594519294798374, + 0.006908628158271313, + 0.00968869123607874, + 0.013376626186072826, + 0.007932676933705807, + 0.010634475387632847, + 0.011234280653297901, + 0.009064234793186188, + 0.005125543102622032, + 0.0048075285740196705, + 0.04169467091560364, + 0.006835054140537977, + 0.05403467267751694, + 0.010393960401415825, + 0.005091859493404627, + 0.009258292615413666, + 0.009374525398015976, + 0.009719781577587128, + 0.01066930964589119, + 0.013055169023573399, + 0.008899484761059284, + 0.011328562162816525, + 0.014267570339143276, + 0.008325109258294106, + 0.010269967839121819, + 0.009115072898566723, + 0.008611811324954033, + 0.012759865261614323, + 0.01954323798418045, + 0.015891674906015396, + 0.02157418057322502, + 0.006892483215779066, + 0.010631999000906944, + 0.006384931970387697, + 0.006015283986926079, + 0.00827092956751585, + 0.006843077950179577, + 0.014866597950458527, + 0.008175259456038475, + 0.013143233954906464, + 0.02710464596748352, + 0.013056465424597263, + 0.02005475014448166, + 0.025022324174642563, + 0.023210769519209862, + 0.023613231256604195, + 0.017684059217572212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.008646431379020214 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863614957779646 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.00863614957779646, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863614957779646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1031501516699791, + 0.042876821011304855, + 0.04804724454879761, + 0.061910029500722885, + 0.05368206277489662, + 0.21299728751182556, + 0.1966530829668045, + 0.07348867505788803, + 0.07058564573526382, + 0.1366090178489685 + ], + "credit_map": [ + 0.012399780564010143, + 0.00955219380557537, + 0.014046345837414265, + 0.014857910573482513, + 0.011291497386991978, + 0.015239197760820389, + 0.01609935611486435, + 0.009663864970207214, + 0.004261220805346966, + 0.006504829507321119, + 0.005868387408554554, + 0.0061724078841507435, + 0.006841527298092842, + 0.006078905425965786, + 0.004410796333104372, + 0.002738748211413622, + 0.006156700663268566, + 0.008282773196697235, + 0.006987457629293203, + 0.006578444503247738, + 0.004687075037509203, + 0.006599485408514738, + 0.004582767840474844, + 0.004172544460743666, + 0.006198897957801819, + 0.008513519540429115, + 0.006102643441408873, + 0.006697838194668293, + 0.008583244867622852, + 0.008774757385253906, + 0.010702692903578281, + 0.006336436606943607, + 0.005014397203922272, + 0.00461135059595108, + 0.006092987954616547, + 0.008730636909604073, + 0.006814173422753811, + 0.009007878601551056, + 0.005983260925859213, + 0.007427374366670847, + 0.010791406966745853, + 0.010324773378670216, + 0.006696473341435194, + 0.07227173447608948, + 0.010978744365274906, + 0.07528665661811829, + 0.015897514298558235, + 0.010749991051852703, + 0.015205449424684048, + 0.014568246901035309, + 0.02319885604083538, + 0.024420008063316345, + 0.02181187830865383, + 0.022995762526988983, + 0.02688632905483246, + 0.047566574066877365, + 0.007848496548831463, + 0.009356691502034664, + 0.007837114855647087, + 0.0059955245815217495, + 0.010202779434621334, + 0.011914522387087345, + 0.009157632477581501, + 0.011175913736224174, + 0.007988093420863152, + 0.009436883963644505, + 0.008098931983113289, + 0.008843977935612202, + 0.007304583676159382, + 0.007979759946465492, + 0.011825618334114552, + 0.009107791818678379, + 0.010834718123078346, + 0.01709623821079731, + 0.010170225985348225, + 0.016100604087114334, + 0.017953496426343918, + 0.02025735378265381, + 0.022113492712378502, + 0.02208288013935089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.00863614957779646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the Great Wall winding through mist-shrouded green hills at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny luminous buds emerge from deep crevices in the gray stones and unfold into translucent blossoms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the blossoms display a vibrant spectrum of pink, white, and ethereal blue colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flowering spread rapidly along the fortification's zigzag spine and watchtowers in a fluid time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a high angle and slowly pull back while panning right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42885926365852356 + ], + "opposite_logprob_tokens": [ + -1.0538592338562012 + ], + "logprob_reward": -0.42885926365852356, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42885926365852356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10637329518795013, + 0.07803241163492203, + 0.06390826404094696, + 0.07121843844652176, + 0.03398972749710083, + 0.2791585624217987, + 0.1218099519610405, + 0.04340493679046631, + 0.04767116531729698, + 0.15443319082260132 + ], + "credit_map": [ + 0.012917713262140751, + 0.00971534475684166, + 0.015036175958812237, + 0.014960738830268383, + 0.01518572773784399, + 0.014893251471221447, + 0.013653300702571869, + 0.010011042468249798, + 0.008845712058246136, + 0.012380729429423809, + 0.008912528865039349, + 0.010199734941124916, + 0.012869510799646378, + 0.011849869973957539, + 0.007426103111356497, + 0.00554822338744998, + 0.008359665982425213, + 0.009444636292755604, + 0.007298405282199383, + 0.008785417303442955, + 0.0075718373991549015, + 0.010885881260037422, + 0.006779169663786888, + 0.004783254116773605, + 0.007111969869583845, + 0.010451928712427616, + 0.006507741287350655, + 0.008244075812399387, + 0.008372131735086441, + 0.009629609063267708, + 0.012824959121644497, + 0.008076027035713196, + 0.0032169977203011513, + 0.0048390282317996025, + 0.0031216414645314217, + 0.004206505604088306, + 0.007010309025645256, + 0.004302490968257189, + 0.0032551665790379047, + 0.004037589766085148, + 0.01136577595025301, + 0.01138308271765709, + 0.011736923828721046, + 0.10723764449357986, + 0.008985313586890697, + 0.09095001965761185, + 0.02658388391137123, + 0.010915939696133137, + 0.009992456994950771, + 0.009619020856916904, + 0.014087595976889133, + 0.015701662749052048, + 0.015313813462853432, + 0.0134019386023283, + 0.018938260152935982, + 0.0247552078217268, + 0.003389746882021427, + 0.004308739677071571, + 0.003425862640142441, + 0.004643817897886038, + 0.004932966083288193, + 0.00556212617084384, + 0.006260789465159178, + 0.010880887508392334, + 0.003612214932218194, + 0.004562971647828817, + 0.003647630335763097, + 0.005741623695939779, + 0.0083254873752594, + 0.006669738329946995, + 0.009771530516445637, + 0.005339967552572489, + 0.013410651125013828, + 0.016720276325941086, + 0.016010893508791924, + 0.020133432000875473, + 0.015892084687948227, + 0.022563667967915535, + 0.02314717508852482, + 0.026555022224783897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.42885926365852356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_13/index_tensor([6141, 6141])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_13/index_tensor([6141, 6141])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..867f8dd183a8b113604316516dc1a4d637701928 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_13/index_tensor([6141, 6141])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "strikingly close-up shot, shot from a sideways, frontal angle, captures a young man fast asleep against the window of a moving car. He has short, shaggy dark brown hair, a few strands tangled across his forehead from his sleeping position. His eyes are closed, his face serene as he rests his head gently against the cool car window. He wears a soft, off-white cotton hooded sweatshirt. Warm afternoon sunlight filters through swaying leaves on the roadside, casting a shifting, soft, playful pattern across his smooth cheeks and high nose. Light and shadow flow slowly across his face. As the car jolted gently and rhythmically along the road, his long, thick eyelashes fluttered ever-so-slightly, as if capturing a peaceful dream. Through the car window behind him, the green fields outside, rapidly receding into a blur, are visible. The entire picture is shrouded in a warm golden halo, with the main colors being soft beige and gold, creating an extremely quiet, serene and dreamy atmosphere, perfectly capturing a moment of tranquility during the journey.", + "seed": 28740, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_13/index_6141_seed_28740_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.044253043457865716, + "vlm_logprob": -0.044253043457865716, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.044253043457865716, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10139656811952591 + ], + "opposite_logprob_tokens": [ + -2.4763965606689453 + ], + "logprob_reward": -0.10139656811952591, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10139656811952591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28557470440864563, + 0.04967804625630379, + 0.027389610186219215, + 0.08489413559436798, + 0.05285028740763664, + 0.12661834061145782, + 0.25799286365509033, + 0.030752738937735558, + 0.029365424066781998, + 0.05488383769989014 + ], + "credit_map": [ + 0.05766545981168747, + 0.029680170118808746, + 0.033158667385578156, + 0.027066530659794807, + 0.029613569378852844, + 0.04229585826396942, + 0.038511551916599274, + 0.027582883834838867, + 0.007471080403774977, + 0.003478964092209935, + 0.004152536392211914, + 0.005108687095344067, + 0.004648021887987852, + 0.0074479300528764725, + 0.011328228749334812, + 0.006042597349733114, + 0.0037546036764979362, + 0.002021064283326268, + 0.004267552401870489, + 0.0025211831089109182, + 0.0029461304657161236, + 0.003851201618090272, + 0.004886412527412176, + 0.0031414623372256756, + 0.023459095507860184, + 0.0037730266340076923, + 0.006004026159644127, + 0.004557365085929632, + 0.004100610502064228, + 0.006819499656558037, + 0.009211156517267227, + 0.02696935646235943, + 0.008350334130227566, + 0.005515660624951124, + 0.006488584913313389, + 0.005377169232815504, + 0.005404581781476736, + 0.006664809305220842, + 0.008400876075029373, + 0.006648271810263395, + 0.015120680443942547, + 0.011361955665051937, + 0.015343517996370792, + 0.01222379133105278, + 0.015330872498452663, + 0.0230488833039999, + 0.016946889460086823, + 0.01724175363779068, + 0.029607966542243958, + 0.03680931776762009, + 0.031126564368605614, + 0.028459345921874046, + 0.03312091901898384, + 0.030396131798624992, + 0.029388295486569405, + 0.03908431529998779, + 0.0047095115296542645, + 0.0027046767063438892, + 0.003006501356139779, + 0.004098204895853996, + 0.00399389723315835, + 0.006493933964520693, + 0.003012817120179534, + 0.0027331975288689137, + 0.002963713137432933, + 0.0030428848695009947, + 0.0029557819943875074, + 0.003985298331826925, + 0.004252539947628975, + 0.003821000223979354, + 0.003208112670108676, + 0.005136091262102127, + 0.004389163572341204, + 0.004397890530526638, + 0.005512200761586428, + 0.005768138449639082, + 0.005732094403356314, + 0.007331244181841612, + 0.015366248786449432, + 0.006386859342455864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.10139656811952591 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07905663549900055 + ], + "opposite_logprob_tokens": [ + -2.579056739807129 + ], + "logprob_reward": -0.07905663549900055, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07905663549900055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14442788064479828, + 0.06323929876089096, + 0.2428690493106842, + 0.1325133591890335, + 0.04471798613667488, + 0.064902663230896, + 0.12314900010824203, + 0.04990019649267197, + 0.10998983681201935, + 0.024290721863508224 + ], + "credit_map": [ + 0.018354900181293488, + 0.02059333026409149, + 0.016084223985671997, + 0.016438236460089684, + 0.016520902514457703, + 0.024601252749562263, + 0.017858155071735382, + 0.013976871967315674, + 0.00878644734621048, + 0.004185365047305822, + 0.006681431550532579, + 0.006089971400797367, + 0.007142978720366955, + 0.006456395611166954, + 0.015186537988483906, + 0.008710174821317196, + 0.047471269965171814, + 0.014288585633039474, + 0.04204649478197098, + 0.023984363302588463, + 0.017416443675756454, + 0.03134724497795105, + 0.027186265215277672, + 0.0391283854842186, + 0.01958031766116619, + 0.008986560627818108, + 0.021848425269126892, + 0.014246740378439426, + 0.012058776803314686, + 0.013484349474310875, + 0.026471566408872604, + 0.015836622565984726, + 0.008435272611677647, + 0.0037503105122596025, + 0.0052505675703287125, + 0.004107991699129343, + 0.005047248676419258, + 0.008032536134123802, + 0.004687987267971039, + 0.005406069569289684, + 0.009511586278676987, + 0.00768430158495903, + 0.0056785824708640575, + 0.006076867692172527, + 0.008380145765841007, + 0.012053263373672962, + 0.006815540604293346, + 0.008702375926077366, + 0.012140250764787197, + 0.019009418785572052, + 0.01401588600128889, + 0.0142316659912467, + 0.016633886843919754, + 0.01428762637078762, + 0.014457507990300655, + 0.018372762948274612, + 0.006149398162961006, + 0.005350854247808456, + 0.00497068976983428, + 0.00786155927926302, + 0.007345298305153847, + 0.00745115103200078, + 0.007010423578321934, + 0.003760820720344782, + 0.010922994464635849, + 0.011826328933238983, + 0.015354948118329048, + 0.01592193730175495, + 0.014031417667865753, + 0.014405996538698673, + 0.01480154786258936, + 0.012724676169455051, + 0.0037397202104330063, + 0.0029724345076829195, + 0.002055080607533455, + 0.0030467708129435778, + 0.003409056691452861, + 0.0037714058998972178, + 0.003123778384178877, + 0.0021724761463701725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.07905663549900055 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609945483505726 + ], + "opposite_logprob_tokens": [ + -4.141099452972412 + ], + "logprob_reward": -0.01609945483505726, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609945483505726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28814220428466797, + 0.07087641209363937, + 0.12399537116289139, + 0.09532303363084793, + 0.024055829271674156, + 0.054214634001255035, + 0.12386179715394974, + 0.1285260021686554, + 0.056938089430332184, + 0.03406664729118347 + ], + "credit_map": [ + 0.0768219456076622, + 0.023229237645864487, + 0.030997667461633682, + 0.015768826007843018, + 0.029314212501049042, + 0.049909695982933044, + 0.02578783594071865, + 0.036312781274318695, + 0.014709718525409698, + 0.006325294263660908, + 0.006272606551647186, + 0.0067201037891209126, + 0.008256465196609497, + 0.011541290208697319, + 0.007951101288199425, + 0.009099827148020267, + 0.018020030111074448, + 0.007361200172454119, + 0.012797843664884567, + 0.013025892898440361, + 0.009513778612017632, + 0.01494514849036932, + 0.034912437200546265, + 0.013419031165540218, + 0.02484944835305214, + 0.00491984561085701, + 0.0062048970721662045, + 0.005372087936848402, + 0.005308256950229406, + 0.007605609949678183, + 0.013289939612150192, + 0.02777293510735035, + 0.0023785156663507223, + 0.0028939375188201666, + 0.0029167395550757647, + 0.0032844615634530783, + 0.003058352740481496, + 0.0037593990564346313, + 0.0035960718523710966, + 0.002168351085856557, + 0.010627401061356068, + 0.006730320863425732, + 0.005441342480480671, + 0.005223765503615141, + 0.005380102898925543, + 0.00939101167023182, + 0.005179013125598431, + 0.006241668481379747, + 0.020163236185908318, + 0.013081612065434456, + 0.016293872147798538, + 0.011946427635848522, + 0.015162601135671139, + 0.01703709363937378, + 0.014462091960012913, + 0.015714852139353752, + 0.018574636429548264, + 0.010477437637746334, + 0.012359635904431343, + 0.015978001058101654, + 0.01658784970641136, + 0.028158532455563545, + 0.01430509053170681, + 0.012084818445146084, + 0.007647609803825617, + 0.00382537255063653, + 0.003917773719877005, + 0.0037955816369503736, + 0.005731618497520685, + 0.014162519946694374, + 0.005056504625827074, + 0.012801099568605423, + 0.006459106225520372, + 0.00355109223164618, + 0.004568387288600206, + 0.0038576945662498474, + 0.004243203904479742, + 0.004490981809794903, + 0.0037500897888094187, + 0.0031460910104215145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.01609945483505726 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067663207650185 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.016067663207650185, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067663207650185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07958772033452988, + 0.13470229506492615, + 0.08099235594272614, + 0.05784541741013527, + 0.06398604065179825, + 0.1383451223373413, + 0.15433858335018158, + 0.050480931997299194, + 0.0855196863412857, + 0.1542019098997116 + ], + "credit_map": [ + 0.009655540809035301, + 0.007736917119473219, + 0.013860155828297138, + 0.00948004238307476, + 0.007981553673744202, + 0.009207970462739468, + 0.013697220012545586, + 0.007968321442604065, + 0.035963717848062515, + 0.013492110185325146, + 0.017042003571987152, + 0.011873005889356136, + 0.010172098875045776, + 0.01732523739337921, + 0.010782621800899506, + 0.01805148832499981, + 0.01113448478281498, + 0.004432868678122759, + 0.008402986451983452, + 0.007569999434053898, + 0.006380060687661171, + 0.010515689849853516, + 0.023103969171643257, + 0.009452302940189838, + 0.011878467164933681, + 0.003631728468462825, + 0.003514444688335061, + 0.004488568753004074, + 0.004661517217755318, + 0.004318816587328911, + 0.012037956155836582, + 0.013313918374478817, + 0.008708082139492035, + 0.005580282770097256, + 0.006039922125637531, + 0.007520147133618593, + 0.008611040189862251, + 0.011614993214607239, + 0.00916951522231102, + 0.0067420583218336105, + 0.02590409480035305, + 0.013532467186450958, + 0.012679553590714931, + 0.011830712668597698, + 0.013438831083476543, + 0.03296347334980965, + 0.012178861536085606, + 0.015817122533917427, + 0.021165234968066216, + 0.022423379123210907, + 0.014964590780436993, + 0.015065238811075687, + 0.0194130577147007, + 0.02719738520681858, + 0.016715088859200478, + 0.01739461161196232, + 0.007101610768586397, + 0.0038777205627411604, + 0.004972402006387711, + 0.005878690630197525, + 0.006143261678516865, + 0.011838133446872234, + 0.005006089340895414, + 0.005663022864609957, + 0.009791558608412743, + 0.004868768621236086, + 0.0045205638743937016, + 0.004542273934930563, + 0.007088189013302326, + 0.025222009047865868, + 0.004213682841509581, + 0.02527264505624771, + 0.030816007405519485, + 0.020896874368190765, + 0.02469274029135704, + 0.012581724673509598, + 0.017917616292834282, + 0.0188869871199131, + 0.011701083742082119, + 0.01670888438820839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.016067663207650185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644895628094673 + ], + "opposite_logprob_tokens": [ + -4.758645057678223 + ], + "logprob_reward": -0.008644895628094673, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644895628094673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3334601819515228, + 0.05310782790184021, + 0.04838468134403229, + 0.037533827126026154, + 0.07814403623342514, + 0.0578785240650177, + 0.12904778122901917, + 0.13564446568489075, + 0.0719173401594162, + 0.05488131567835808 + ], + "credit_map": [ + 0.06684067100286484, + 0.036143895238637924, + 0.04718925431370735, + 0.031120270490646362, + 0.036846619099378586, + 0.04264684021472931, + 0.03760652616620064, + 0.03506610170006752, + 0.01261067483574152, + 0.0041616833768785, + 0.005550331436097622, + 0.005824132822453976, + 0.004300807602703571, + 0.009192947298288345, + 0.005244921892881393, + 0.00622232910245657, + 0.00668269582092762, + 0.0037679283414036036, + 0.006022052373737097, + 0.004965845495462418, + 0.003937124740332365, + 0.007037636358290911, + 0.010097046382725239, + 0.005874352995306253, + 0.009512640535831451, + 0.001726632472127676, + 0.003254887880757451, + 0.0023639400023967028, + 0.002402025507763028, + 0.002598466817289591, + 0.0046423026360571384, + 0.011032931506633759, + 0.015760233625769615, + 0.006805612239986658, + 0.00902724638581276, + 0.006958356127142906, + 0.007848082110285759, + 0.013958350755274296, + 0.008184823207557201, + 0.009601333178579807, + 0.011590722016990185, + 0.003119627246633172, + 0.005536545533686876, + 0.004876927472651005, + 0.005110646598041058, + 0.012044699862599373, + 0.006859539542347193, + 0.008739810436964035, + 0.012874158099293709, + 0.0192249845713377, + 0.01584586128592491, + 0.016412809491157532, + 0.016591684892773628, + 0.014471029862761497, + 0.016065387055277824, + 0.017561862245202065, + 0.024176672101020813, + 0.01225563045591116, + 0.012647881172597408, + 0.009368359111249447, + 0.015204977244138718, + 0.03485628217458725, + 0.009623798541724682, + 0.017510859295725822, + 0.012798302806913853, + 0.005137164611369371, + 0.005042761564254761, + 0.005389815662056208, + 0.0059920623898506165, + 0.017523201182484627, + 0.0035737736616283655, + 0.01646025851368904, + 0.012083111330866814, + 0.007824970409274101, + 0.007122039329260588, + 0.0040596602484583855, + 0.007183671463280916, + 0.007291189394891262, + 0.004696011543273926, + 0.004620661493390799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.008644895628094673 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly close-up shot, shot from a sideways, frontal angle, captures a young man fast asleep against the window of a moving car. He has short, shaggy dark brown hair, a few strands tangled across his forehead from his sleeping position. His eyes are closed, his face serene as he rests his head gently against the cool car window. He wears a soft, off-white cotton hooded sweatshirt. Warm afternoon sunlight filters through swaying leaves on the roadside, casting a shifting, soft, playful pattern across his smooth cheeks and high nose. Light and shadow flow slowly across his face. As the car jolted gently and rhythmically along the road, his long, thick eyelashes fluttered ever-so-slightly, as if capturing a peaceful dream. Through the car window behind him, the green fields outside, rapidly receding into a blur, are visible. The entire picture is shrouded in a warm golden halo, with the main colors being soft beige and gold, creating an extremely quiet, serene and dreamy atmosphere, perfectly capturing a moment of tranquility during the journey.", + "seed": 28741, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_13/index_6141_seed_28741_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12014584802091122, + "vlm_logprob": -0.12014584802091122, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.12014584802091122, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.32221177220344543 + ], + "opposite_logprob_tokens": [ + -1.322211742401123 + ], + "logprob_reward": -0.32221177220344543, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32221177220344543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5811985731124878, + 0.06334668397903442, + 0.03292129188776016, + 0.06109876558184624, + 0.08484767377376556, + 0.048797156661748886, + 0.0436863973736763, + 0.03142659366130829, + 0.03735750913619995, + 0.015319324098527431 + ], + "credit_map": [ + 0.02237415313720703, + 0.020015038549900055, + 0.01489945687353611, + 0.01577049307525158, + 0.019611656665802002, + 0.22515524923801422, + 0.01919139176607132, + 0.24418117105960846, + 0.005824601277709007, + 0.008635547943413258, + 0.004228249657899141, + 0.005351966712623835, + 0.00945565290749073, + 0.008150065317749977, + 0.009888878092169762, + 0.011811723932623863, + 0.0022307175677269697, + 0.003371462458744645, + 0.002949940739199519, + 0.0033012330532073975, + 0.003412490477785468, + 0.007659620605409145, + 0.005720236804336309, + 0.004275593440979719, + 0.00912327878177166, + 0.005686561111360788, + 0.00711841369047761, + 0.0034561248030513525, + 0.005134819075465202, + 0.015788421034812927, + 0.006419352721422911, + 0.008371799252927303, + 0.003403910668566823, + 0.00426657497882843, + 0.004901384003460407, + 0.0052503542974591255, + 0.0058207386173307896, + 0.01959792710840702, + 0.011955899186432362, + 0.029650887474417686, + 0.0031790805514901876, + 0.005842917133122683, + 0.005511890631169081, + 0.004144166596233845, + 0.006289611104875803, + 0.009584701620042324, + 0.0068397982977330685, + 0.007404990494251251, + 0.0038101153913885355, + 0.004337948746979237, + 0.004721481818705797, + 0.004305030219256878, + 0.006430621258914471, + 0.0063569447956979275, + 0.00712470430880785, + 0.006599548738449812, + 0.003272994188591838, + 0.0033216241281479597, + 0.0028299165423959494, + 0.002259837696328759, + 0.0040429383516311646, + 0.006820401176810265, + 0.003559916513040662, + 0.005318965297192335, + 0.003024858655408025, + 0.0037834884133189917, + 0.0037791964132338762, + 0.0036303699016571045, + 0.00605438556522131, + 0.0058702342212200165, + 0.006248541176319122, + 0.004966433625668287, + 0.0011498858220875263, + 0.001597863738425076, + 0.0012403882574290037, + 0.0011675453279167414, + 0.0023302468471229076, + 0.0023941281251609325, + 0.003363978350535035, + 0.0020752872806042433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.32221177220344543 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03376979008316994 + ], + "opposite_logprob_tokens": [ + -3.4087698459625244 + ], + "logprob_reward": -0.03376979008316994, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03376979008316994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3596400022506714, + 0.06324602663516998, + 0.08269316703081131, + 0.07320473343133926, + 0.14620670676231384, + 0.029755178838968277, + 0.06145130842924118, + 0.10670457780361176, + 0.036294884979724884, + 0.04080341011285782 + ], + "credit_map": [ + 0.014214050024747849, + 0.013948814943432808, + 0.015647195279598236, + 0.012443379499018192, + 0.013890724629163742, + 0.13360589742660522, + 0.016070958226919174, + 0.13981899619102478, + 0.00611525634303689, + 0.006778476759791374, + 0.0063655260019004345, + 0.0063209449872374535, + 0.007565068546682596, + 0.011956211179494858, + 0.006044983398169279, + 0.012099554762244225, + 0.007307040970772505, + 0.009835859760642052, + 0.011840692721307278, + 0.01079114805907011, + 0.009204264730215073, + 0.011491885408759117, + 0.01162540353834629, + 0.010596874170005322, + 0.006343698129057884, + 0.007650317158550024, + 0.01444415096193552, + 0.00747430557385087, + 0.008232268504798412, + 0.011910767294466496, + 0.007397392299026251, + 0.00975183304399252, + 0.0040550814010202885, + 0.003397585591301322, + 0.005375460255891085, + 0.00648881820961833, + 0.005946632008999586, + 0.04334023967385292, + 0.01426722016185522, + 0.06333566457033157, + 0.0024497052654623985, + 0.0033543086610734463, + 0.003769099013879895, + 0.003361113602295518, + 0.005200967658311129, + 0.00309489737264812, + 0.004170604050159454, + 0.004354480654001236, + 0.0047981999814510345, + 0.007367850746959448, + 0.0119266826659441, + 0.007465870585292578, + 0.008140078745782375, + 0.004036951344460249, + 0.009199324063956738, + 0.008516351692378521, + 0.011026142165064812, + 0.012954698875546455, + 0.018843019381165504, + 0.011973189190030098, + 0.0135268559679389, + 0.007584510836750269, + 0.01298472099006176, + 0.01781143620610237, + 0.003120016073808074, + 0.004958995617926121, + 0.0049251774325966835, + 0.0037810325156897306, + 0.0044114720076322556, + 0.004602555651217699, + 0.0057898350059986115, + 0.004705800209194422, + 0.0042274934239685535, + 0.006562733557075262, + 0.004819619003683329, + 0.003870819229632616, + 0.0060297781601548195, + 0.00515170581638813, + 0.004574760794639587, + 0.005566500127315521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.03376979008316994 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799958050251007 + ], + "opposite_logprob_tokens": [ + -1.8049957752227783 + ], + "logprob_reward": -0.1799958050251007, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799958050251007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08598931133747101, + 0.08059787005186081, + 0.09054781496524811, + 0.10185682773590088, + 0.15945351123809814, + 0.2948152720928192, + 0.026461588218808174, + 0.016231779009103775, + 0.13232313096523285, + 0.011722845956683159 + ], + "credit_map": [ + 0.006617299746721983, + 0.007328186649829149, + 0.01001287717372179, + 0.006319060921669006, + 0.014696810394525528, + 0.013652903027832508, + 0.011561784893274307, + 0.015800394117832184, + 0.004963414277881384, + 0.006271470803767443, + 0.004653522279113531, + 0.005742844194173813, + 0.010503974743187428, + 0.016310472041368484, + 0.009492505341768265, + 0.02265966683626175, + 0.00811227411031723, + 0.01011157501488924, + 0.010386589914560318, + 0.009074337780475616, + 0.010170940309762955, + 0.018179597333073616, + 0.008934005163609982, + 0.015578493475914001, + 0.008223486132919788, + 0.007730845827609301, + 0.012263537384569645, + 0.008656012825667858, + 0.010969727300107479, + 0.02724076807498932, + 0.014071457087993622, + 0.01270099263638258, + 0.007166162598878145, + 0.005582768470048904, + 0.00728091923519969, + 0.008610332384705544, + 0.009372128173708916, + 0.045117106288671494, + 0.0170877818018198, + 0.0592363178730011, + 0.021775612607598305, + 0.02925906702876091, + 0.03428136929869652, + 0.025484193116426468, + 0.0407399944961071, + 0.05327746644616127, + 0.046895481646060944, + 0.04310211166739464, + 0.001837344840168953, + 0.0024659931659698486, + 0.0029717280995100737, + 0.0021757164504379034, + 0.0033831805922091007, + 0.005908056627959013, + 0.0038391367997974157, + 0.0038804314099252224, + 0.00181415188126266, + 0.002697047544643283, + 0.0018152327975258231, + 0.0012701336527243257, + 0.002153457375243306, + 0.0018761528190225363, + 0.002366109751164913, + 0.0022394927218556404, + 0.011115252040326595, + 0.011360067874193192, + 0.011824154295027256, + 0.008984945714473724, + 0.02055242471396923, + 0.02592984400689602, + 0.022004373371601105, + 0.02055206336081028, + 0.001204484375193715, + 0.001348418416455388, + 0.0013948754640296102, + 0.0010068153496831656, + 0.0017581870779395103, + 0.0017343983054161072, + 0.0017302290070801973, + 0.0015454377280548215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.1799958050251007 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496208369731903 + ], + "opposite_logprob_tokens": [ + -2.929962158203125 + ], + "logprob_reward": -0.05496208369731903, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496208369731903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08578351885080338, + 0.23902706801891327, + 0.028092218562960625, + 0.0704990029335022, + 0.14210635423660278, + 0.26167455315589905, + 0.052390944212675095, + 0.04618063569068909, + 0.016390424221754074, + 0.05785531550645828 + ], + "credit_map": [ + 0.006347151007503271, + 0.008602338843047619, + 0.008030657656490803, + 0.006747623439878225, + 0.00813790038228035, + 0.017740264534950256, + 0.008389566093683243, + 0.021788012236356735, + 0.019566986709833145, + 0.02336130291223526, + 0.014240715652704239, + 0.020351223647594452, + 0.029504448175430298, + 0.06633424013853073, + 0.037504080682992935, + 0.02816406823694706, + 0.002381241647526622, + 0.003843628102913499, + 0.002762899035587907, + 0.0026086256839334965, + 0.002880807500332594, + 0.005150184500962496, + 0.0036721869837492704, + 0.004792643710970879, + 0.007039830554276705, + 0.006874598562717438, + 0.011071433313190937, + 0.004299158696085215, + 0.006973030511289835, + 0.0174483023583889, + 0.00778645183891058, + 0.009006195701658726, + 0.003520856611430645, + 0.0031293421052396297, + 0.005049742758274078, + 0.006070657167583704, + 0.005193253513425589, + 0.04386340081691742, + 0.013405435718595982, + 0.061873652040958405, + 0.02261911705136299, + 0.028872808441519737, + 0.039223939180374146, + 0.019069140776991844, + 0.035162489861249924, + 0.049676522612571716, + 0.03180999308824539, + 0.035240527242422104, + 0.005600820295512676, + 0.005866057705134153, + 0.007705397438257933, + 0.004832181613892317, + 0.00564928213134408, + 0.007331583183258772, + 0.00797354057431221, + 0.007432072423398495, + 0.004582432098686695, + 0.0052444892935454845, + 0.006176225375384092, + 0.0033161845058202744, + 0.004420601297169924, + 0.008715279400348663, + 0.007707048673182726, + 0.00601837458088994, + 0.0015307731227949262, + 0.001740539213642478, + 0.0017639342695474625, + 0.0016525405226275325, + 0.002422143705189228, + 0.0024203374050557613, + 0.0026226185727864504, + 0.0022375357802957296, + 0.005647747311741114, + 0.0049974434077739716, + 0.005484510213136673, + 0.0029059103690087795, + 0.007304378319531679, + 0.012194336391985416, + 0.010276565328240395, + 0.009044423699378967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.05496208369731903 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789789095520973 + ], + "opposite_logprob_tokens": [ + -4.634789943695068 + ], + "logprob_reward": -0.009789789095520973, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789789095520973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5067330598831177, + 0.11067806929349899, + 0.0360659584403038, + 0.03677603602409363, + 0.15416574478149414, + 0.04670663923025131, + 0.04263171926140785, + 0.03061503730714321, + 0.019669797271490097, + 0.015957899391651154 + ], + "credit_map": [ + 0.023566989228129387, + 0.020458098500967026, + 0.012912245467305183, + 0.01210623700171709, + 0.026929618790745735, + 0.1896202564239502, + 0.02001504972577095, + 0.20112457871437073, + 0.009518799372017384, + 0.01235402561724186, + 0.008464673534035683, + 0.0087918471544981, + 0.015353797934949398, + 0.0213994849473238, + 0.015381831675767899, + 0.01941361092031002, + 0.002653396688401699, + 0.003917317371815443, + 0.005577755160629749, + 0.0037744638975709677, + 0.003712807549163699, + 0.006432229187339544, + 0.004065151792019606, + 0.005932837259024382, + 0.0031147415284067392, + 0.0037671790923923254, + 0.004521278664469719, + 0.0035841099452227354, + 0.005574717652052641, + 0.00587082002311945, + 0.005341467447578907, + 0.005001723300665617, + 0.012790161184966564, + 0.010158998891711235, + 0.01095450110733509, + 0.012662517838180065, + 0.01754974201321602, + 0.0326555073261261, + 0.03598882257938385, + 0.021405480802059174, + 0.004065472632646561, + 0.005025258753448725, + 0.0069804722443223, + 0.003938557114452124, + 0.0046555595472455025, + 0.007174380123615265, + 0.004794715438038111, + 0.010072223842144012, + 0.004580010659992695, + 0.006098099052906036, + 0.00494174100458622, + 0.004920262843370438, + 0.006284724920988083, + 0.004890753421932459, + 0.005261141341179609, + 0.005654983688145876, + 0.0042109182104468346, + 0.0029572497587651014, + 0.004449884872883558, + 0.002262712689116597, + 0.003571654437109828, + 0.004776504822075367, + 0.003969094250351191, + 0.0044170161709189415, + 0.002367936074733734, + 0.002488476224243641, + 0.0033655737061053514, + 0.0017676876159384847, + 0.002167491475120187, + 0.0023750036489218473, + 0.002860418288037181, + 0.0022772119846194983, + 0.0011259563034400344, + 0.0019659469835460186, + 0.003084236290305853, + 0.0013980595394968987, + 0.0016603197436779737, + 0.002437995048239827, + 0.0019403811311349273, + 0.002345004817470908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.009789789095520973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly close-up shot, shot from a sideways, frontal angle, captures a young man fast asleep against the window of a moving car. He has short, shaggy dark brown hair, a few strands tangled across his forehead from his sleeping position. His eyes are closed, his face serene as he rests his head gently against the cool car window. He wears a soft, off-white cotton hooded sweatshirt. Warm afternoon sunlight filters through swaying leaves on the roadside, casting a shifting, soft, playful pattern across his smooth cheeks and high nose. Light and shadow flow slowly across his face. As the car jolted gently and rhythmically along the road, his long, thick eyelashes fluttered ever-so-slightly, as if capturing a peaceful dream. Through the car window behind him, the green fields outside, rapidly receding into a blur, are visible. The entire picture is shrouded in a warm golden halo, with the main colors being soft beige and gold, creating an extremely quiet, serene and dreamy atmosphere, perfectly capturing a moment of tranquility during the journey.", + "seed": 28742, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_13/index_6141_seed_28742_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14303847551345825, + "vlm_logprob": -0.14303847551345825, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.14303847551345825, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35731372237205505 + ], + "opposite_logprob_tokens": [ + -1.2323137521743774 + ], + "logprob_reward": -0.35731372237205505, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35731372237205505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06881656497716904, + 0.04829486086964607, + 0.05760016664862633, + 0.023222921416163445, + 0.025126874446868896, + 0.7183281183242798, + 0.01049947738647461, + 0.02065362222492695, + 0.017492374405264854, + 0.009965002536773682 + ], + "credit_map": [ + 0.009225658141076565, + 0.0058218915946781635, + 0.0055653126910328865, + 0.010329862125217915, + 0.007569907233119011, + 0.007436686661094427, + 0.011840807273983955, + 0.011026439256966114, + 0.0034877685829997063, + 0.004605850204825401, + 0.002456245245411992, + 0.006155639886856079, + 0.004072955343872309, + 0.005806309171020985, + 0.011654232628643513, + 0.010055861435830593, + 0.005030861124396324, + 0.005057049449533224, + 0.0038575197104364634, + 0.008713064715266228, + 0.006347093731164932, + 0.006234398577362299, + 0.015231067314743996, + 0.007129111792892218, + 0.0019261096604168415, + 0.0031203965190798044, + 0.003080552676692605, + 0.0035029228311032057, + 0.0041869934648275375, + 0.0017888270085677505, + 0.0030773687176406384, + 0.0025397506542503834, + 0.002307112794369459, + 0.0034411856904625893, + 0.001480961567722261, + 0.00286357244476676, + 0.0035652979277074337, + 0.003297653514891863, + 0.003431168617680669, + 0.0047399201430380344, + 0.05467834323644638, + 0.042896561324596405, + 0.012692037038505077, + 0.014755915850400925, + 0.3021053969860077, + 0.06026782467961311, + 0.14588579535484314, + 0.08504629135131836, + 0.001305350917391479, + 0.001440110499970615, + 0.0008306492818519473, + 0.0015950872329995036, + 0.0012277070200070739, + 0.0012442879378795624, + 0.001579040545038879, + 0.0012772443005815148, + 0.002141556702554226, + 0.004452909342944622, + 0.00183581851888448, + 0.003246084088459611, + 0.00225370517000556, + 0.0015561714535579085, + 0.0027195895090699196, + 0.0024477867409586906, + 0.0010415867436677217, + 0.0025115038733929396, + 0.0008896689978428185, + 0.0014986174646764994, + 0.0019549669232219458, + 0.004128516651690006, + 0.0029355117585510015, + 0.0025320027489215136, + 0.0011571442009881139, + 0.0012517796130850911, + 0.0009130479302257299, + 0.0010677868267521262, + 0.0016123612876981497, + 0.001241931808181107, + 0.0014114778023213148, + 0.0013094733003526926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.35731372237205505 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0891714096069336 + ], + "opposite_logprob_tokens": [ + -2.4641714096069336 + ], + "logprob_reward": -0.0891714096069336, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0891714096069336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08300846070051193, + 0.04623570665717125, + 0.09432017803192139, + 0.02647402323782444, + 0.0477723628282547, + 0.3030012845993042, + 0.043540842831134796, + 0.23241475224494934, + 0.040413565933704376, + 0.0828188955783844 + ], + "credit_map": [ + 0.012060320004820824, + 0.009499463252723217, + 0.008220615796744823, + 0.010905667208135128, + 0.010075114667415619, + 0.00983610562980175, + 0.01045161485671997, + 0.011959563940763474, + 0.004930209834128618, + 0.006395229138433933, + 0.005443619564175606, + 0.004598494619131088, + 0.005076150875538588, + 0.007165578659623861, + 0.005600443109869957, + 0.007025980390608311, + 0.006751819979399443, + 0.007242743857204914, + 0.007436216343194246, + 0.013077311217784882, + 0.008535237051546574, + 0.011444728821516037, + 0.027423206716775894, + 0.012408913113176823, + 0.003416567575186491, + 0.00415221368893981, + 0.0035893928725272417, + 0.0032371231354773045, + 0.0030289061833173037, + 0.0028913328424096107, + 0.002740222029387951, + 0.003418263513594866, + 0.004891304299235344, + 0.005269201472401619, + 0.00734924478456378, + 0.006225381046533585, + 0.004982879851013422, + 0.007823196239769459, + 0.004987177439033985, + 0.006243978627026081, + 0.024309180676937103, + 0.01908860355615616, + 0.005836566910147667, + 0.006539391353726387, + 0.12668552994728088, + 0.027172580361366272, + 0.038571059703826904, + 0.05479835718870163, + 0.003253784729167819, + 0.004619324579834938, + 0.0036629342939704657, + 0.00620269775390625, + 0.003562541212886572, + 0.005508851259946823, + 0.010582242161035538, + 0.006148467771708965, + 0.03425682708621025, + 0.0322599895298481, + 0.031331926584243774, + 0.02554510161280632, + 0.021111691370606422, + 0.026378342881798744, + 0.031111912801861763, + 0.030418960377573967, + 0.0035375873558223248, + 0.004888621624559164, + 0.0032125147990882397, + 0.004790783394128084, + 0.005591262131929398, + 0.006960367318242788, + 0.006301082670688629, + 0.005131347104907036, + 0.008421394042670727, + 0.011889929883182049, + 0.01225731335580349, + 0.008370070718228817, + 0.011863140389323235, + 0.008607047609984875, + 0.00932234711945057, + 0.012087653391063213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.0891714096069336 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05503440648317337 + ], + "opposite_logprob_tokens": [ + -2.9300343990325928 + ], + "logprob_reward": -0.05503440648317337, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05503440648317337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11293194442987442, + 0.11010970920324326, + 0.10068406164646149, + 0.04413244128227234, + 0.06378606706857681, + 0.20471256971359253, + 0.08075807988643646, + 0.14908529818058014, + 0.08846615999937057, + 0.04533368721604347 + ], + "credit_map": [ + 0.010951930657029152, + 0.013119027018547058, + 0.008787300437688828, + 0.01668580435216427, + 0.012280820868909359, + 0.017701108008623123, + 0.016185177490115166, + 0.01722078025341034, + 0.010937185026705265, + 0.010619728825986385, + 0.008300467394292355, + 0.010179582983255386, + 0.01198527216911316, + 0.021665802225470543, + 0.021560102701187134, + 0.01486156415194273, + 0.007821092382073402, + 0.009402982890605927, + 0.007502841297537088, + 0.015532911755144596, + 0.012086538597941399, + 0.013882610015571117, + 0.02119518257677555, + 0.013259898871183395, + 0.004017787519842386, + 0.005483423825353384, + 0.003370646620169282, + 0.00615641800686717, + 0.006227852776646614, + 0.004248799756169319, + 0.008400620892643929, + 0.006226891186088324, + 0.00937612820416689, + 0.006137153133749962, + 0.005147268064320087, + 0.0068172430619597435, + 0.007117412053048611, + 0.012641232460737228, + 0.008490432053804398, + 0.008059200830757618, + 0.016151852905750275, + 0.01632232591509819, + 0.01032811775803566, + 0.012217400595545769, + 0.024798184633255005, + 0.016619514673948288, + 0.06746461242437363, + 0.04081057384610176, + 0.006281939335167408, + 0.012776646763086319, + 0.0056845699436962605, + 0.013787737116217613, + 0.006624770350754261, + 0.009402982890605927, + 0.015713540837168694, + 0.010485894046723843, + 0.019717412069439888, + 0.027648836374282837, + 0.012737154960632324, + 0.016383735463023186, + 0.017417944967746735, + 0.017938563600182533, + 0.019565394148230553, + 0.01767624355852604, + 0.004782931879162788, + 0.009610779583454132, + 0.00503126997500658, + 0.0086814034730196, + 0.00846134964376688, + 0.020346080884337425, + 0.019089944660663605, + 0.01246239896863699, + 0.004218146670609713, + 0.004566568415611982, + 0.0031709279865026474, + 0.004996895790100098, + 0.005250346846878529, + 0.006339990999549627, + 0.010474737733602524, + 0.006316072773188353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.05503440648317337 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799509972333908 + ], + "opposite_logprob_tokens": [ + -1.8049509525299072 + ], + "logprob_reward": -0.1799509972333908, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799509972333908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13946126401424408, + 0.041311170905828476, + 0.20072303712368011, + 0.03102477826178074, + 0.027378534898161888, + 0.4454379379749298, + 0.01528999488800764, + 0.0425436869263649, + 0.031385745853185654, + 0.025443818420171738 + ], + "credit_map": [ + 0.013270513154566288, + 0.01015632413327694, + 0.008791955187916756, + 0.019645612686872482, + 0.01888744905591011, + 0.0300043486058712, + 0.018793275579810143, + 0.019911780953407288, + 0.003957527223974466, + 0.004529146011918783, + 0.0036211658734828234, + 0.004007128998637199, + 0.004687372129410505, + 0.00739015219733119, + 0.006942771375179291, + 0.0061759063974022865, + 0.03554461896419525, + 0.012096228078007698, + 0.006256361026316881, + 0.019945748150348663, + 0.02026938647031784, + 0.030402030795812607, + 0.03996884450316429, + 0.03623982518911362, + 0.0026724343188107014, + 0.0033033243380486965, + 0.0024609260726720095, + 0.003977074753493071, + 0.0028167192358523607, + 0.003735265228897333, + 0.007199063431471586, + 0.004859971348196268, + 0.003152408404275775, + 0.0031912541016936302, + 0.0019446664955466986, + 0.0026921613607555628, + 0.00282076932489872, + 0.005403007380664349, + 0.004750964697450399, + 0.00342330290004611, + 0.02729499153792858, + 0.022071637213230133, + 0.009658905677497387, + 0.01576712168753147, + 0.0730481818318367, + 0.04191264510154724, + 0.11358770728111267, + 0.14209671318531036, + 0.001138770137913525, + 0.002776856767013669, + 0.0009408573969267309, + 0.00193163868971169, + 0.0015770131722092628, + 0.0023185231257230043, + 0.0028425275813788176, + 0.001763808191753924, + 0.0038755221758037806, + 0.004834997467696667, + 0.002581140026450157, + 0.004734130576252937, + 0.004254981409758329, + 0.008989573456346989, + 0.008033615536987782, + 0.005239726044237614, + 0.0033341688103973866, + 0.004589039366692305, + 0.0027732504531741142, + 0.0027808574959635735, + 0.0034199957735836506, + 0.0049260519444942474, + 0.005702263209968805, + 0.0038601194974035025, + 0.003323810640722513, + 0.0033845931757241488, + 0.0026035711634904146, + 0.0024040613789111376, + 0.0029817833565175533, + 0.0036252401769161224, + 0.0032066074199974537, + 0.003914151340723038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.1799509972333908 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033721841871738434 + ], + "opposite_logprob_tokens": [ + -3.408721923828125 + ], + "logprob_reward": -0.033721841871738434, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033721841871738434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08747826516628265, + 0.051417604088783264, + 0.0629429817199707, + 0.030902620404958725, + 0.043501581996679306, + 0.587294340133667, + 0.02944495528936386, + 0.040469612926244736, + 0.03381160646677017, + 0.03273650258779526 + ], + "credit_map": [ + 0.00798413623124361, + 0.010255297645926476, + 0.007206846494227648, + 0.010922091081738472, + 0.009156808257102966, + 0.016413908451795578, + 0.01400458812713623, + 0.011534583754837513, + 0.004038162063807249, + 0.005941662937402725, + 0.003191563067957759, + 0.005522164981812239, + 0.0050296541303396225, + 0.008232875727117062, + 0.007855311967432499, + 0.011606205254793167, + 0.00641519483178854, + 0.006614101585000753, + 0.0039048055186867714, + 0.0076113510876894, + 0.007235928438603878, + 0.007208523340523243, + 0.014740205369889736, + 0.009212863631546497, + 0.0027444688603281975, + 0.0036128880456089973, + 0.002383225364610553, + 0.00417599081993103, + 0.0028682327829301357, + 0.0038246249314397573, + 0.0069622136652469635, + 0.004330974072217941, + 0.0049234419129788876, + 0.004838146269321442, + 0.004676994867622852, + 0.004519742913544178, + 0.0042833550833165646, + 0.007738073356449604, + 0.005359861068427563, + 0.007161965128034353, + 0.04503132402896881, + 0.036795370280742645, + 0.009750813245773315, + 0.012361307628452778, + 0.24720953404903412, + 0.051960766315460205, + 0.0931057408452034, + 0.09107940644025803, + 0.0029368398245424032, + 0.0031676546204835176, + 0.0019381444435566664, + 0.00310658011585474, + 0.002636396326124668, + 0.0060513862408697605, + 0.006041104905307293, + 0.0035668464843183756, + 0.0038091111928224564, + 0.0062233624048531055, + 0.004977921023964882, + 0.005224491935223341, + 0.003791485447436571, + 0.005274041555821896, + 0.006031414959579706, + 0.00513777881860733, + 0.0019404791528359056, + 0.004160943906754255, + 0.0019096878822892904, + 0.0037964764051139355, + 0.003248050110414624, + 0.009058661758899689, + 0.00532089127227664, + 0.004376410041004419, + 0.0023266514763236046, + 0.003107934957370162, + 0.0021701930090785027, + 0.003346298821270466, + 0.004148843232542276, + 0.005942091345787048, + 0.0063131931237876415, + 0.005381294991821051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.033721841871738434 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly close-up shot, shot from a sideways, frontal angle, captures a young man fast asleep against the window of a moving car. He has short, shaggy dark brown hair, a few strands tangled across his forehead from his sleeping position. His eyes are closed, his face serene as he rests his head gently against the cool car window. He wears a soft, off-white cotton hooded sweatshirt. Warm afternoon sunlight filters through swaying leaves on the roadside, casting a shifting, soft, playful pattern across his smooth cheeks and high nose. Light and shadow flow slowly across his face. As the car jolted gently and rhythmically along the road, his long, thick eyelashes fluttered ever-so-slightly, as if capturing a peaceful dream. Through the car window behind him, the green fields outside, rapidly receding into a blur, are visible. The entire picture is shrouded in a warm golden halo, with the main colors being soft beige and gold, creating an extremely quiet, serene and dreamy atmosphere, perfectly capturing a moment of tranquility during the journey.", + "seed": 28743, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_13/index_6141_seed_28743_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07345859576016664, + "vlm_logprob": -0.07345859576016664, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.07345859576016664, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.046817928552627563 + ], + "opposite_logprob_tokens": [ + -3.2968180179595947 + ], + "logprob_reward": -0.046817928552627563, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.046817928552627563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19342301785945892, + 0.07297243177890778, + 0.11979760974645615, + 0.04024855047464371, + 0.11676192283630371, + 0.06225031614303589, + 0.05403280258178711, + 0.12059301137924194, + 0.09776553511619568, + 0.12215475738048553 + ], + "credit_map": [ + 0.012620099820196629, + 0.018712231889367104, + 0.03552315756678581, + 0.02641924098134041, + 0.025695107877254486, + 0.018847964704036713, + 0.024606050923466682, + 0.030999159440398216, + 0.00955040380358696, + 0.0073561337776482105, + 0.009849190711975098, + 0.006212558131664991, + 0.009901825338602066, + 0.007844183593988419, + 0.012820704840123653, + 0.009437432512640953, + 0.009426149539649487, + 0.010044352151453495, + 0.014737744815647602, + 0.012104306370019913, + 0.013827618211507797, + 0.01846211776137352, + 0.024215763434767723, + 0.01697956770658493, + 0.006204358302056789, + 0.004050293006002903, + 0.005115986801683903, + 0.0038684834726154804, + 0.005333202891051769, + 0.005843946244567633, + 0.005326339043676853, + 0.004505939316004515, + 0.01970500685274601, + 0.013224423862993717, + 0.017241382971405983, + 0.01274533849209547, + 0.015917090699076653, + 0.008657577447593212, + 0.013787645846605301, + 0.015483456663787365, + 0.004890356678515673, + 0.007284366525709629, + 0.008229739964008331, + 0.011997903697192669, + 0.0077806576155126095, + 0.009140204638242722, + 0.005903496406972408, + 0.0070235878229141235, + 0.00445298757404089, + 0.005080260802060366, + 0.004984488245099783, + 0.004312598146498203, + 0.007797630503773689, + 0.014104124158620834, + 0.004826633725315332, + 0.008474080823361874, + 0.011830918490886688, + 0.012816217727959156, + 0.013966956175863743, + 0.012864289805293083, + 0.017364678904414177, + 0.029016979038715363, + 0.009141594171524048, + 0.013591372407972813, + 0.01719539985060692, + 0.0092830965295434, + 0.008716332726180553, + 0.008115820586681366, + 0.01549910195171833, + 0.020299920812249184, + 0.008371184580028057, + 0.010284681804478168, + 0.0058545260690152645, + 0.017955008894205093, + 0.014108937233686447, + 0.014587286859750748, + 0.0253020990639925, + 0.017100311815738678, + 0.01561399083584547, + 0.01163260918110609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.046817928552627563 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04872574284672737 + ], + "opposite_logprob_tokens": [ + -3.0487258434295654 + ], + "logprob_reward": -0.04872574284672737, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04872574284672737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0730130672454834, + 0.04875485599040985, + 0.04646633565425873, + 0.045584000647068024, + 0.10256842523813248, + 0.04116996377706528, + 0.06215229257941246, + 0.5365144610404968, + 0.014361919835209846, + 0.029414653778076172 + ], + "credit_map": [ + 0.007449384778738022, + 0.005410437472164631, + 0.010261527262628078, + 0.008803724311292171, + 0.010389568284153938, + 0.007968654856085777, + 0.012372794561088085, + 0.010356971062719822, + 0.012129989452660084, + 0.004225735552608967, + 0.005855186376720667, + 0.0033795058261603117, + 0.005937399808317423, + 0.005470968782901764, + 0.004842665512114763, + 0.006913401652127504, + 0.00871409010142088, + 0.003971271216869354, + 0.005619966425001621, + 0.004066289868205786, + 0.004839134868234396, + 0.006593575235456228, + 0.005654953885823488, + 0.0070070549845695496, + 0.011416252702474594, + 0.005973090883344412, + 0.005388968158513308, + 0.0033722082152962685, + 0.003957059700042009, + 0.006804178468883038, + 0.0037020219024270773, + 0.004970216657966375, + 0.005576267372816801, + 0.007795118726789951, + 0.030405478551983833, + 0.026084616780281067, + 0.008087163791060448, + 0.004329677205532789, + 0.007033633533865213, + 0.013256472535431385, + 0.005362711846828461, + 0.004124533385038376, + 0.00824083760380745, + 0.0072556184604763985, + 0.00417333422228694, + 0.0041802022606134415, + 0.0036407005973160267, + 0.004192025400698185, + 0.006004705093801022, + 0.0051200673915445805, + 0.005545531865209341, + 0.00557396374642849, + 0.008136616088449955, + 0.015974385663866997, + 0.007697794120758772, + 0.008099229075014591, + 0.03231952339410782, + 0.057336777448654175, + 0.10057494789361954, + 0.08255785703659058, + 0.06777460873126984, + 0.08485978841781616, + 0.056480057537555695, + 0.05461091920733452, + 0.0016392591642215848, + 0.0015359719982370734, + 0.00202164170332253, + 0.0018468209309503436, + 0.0018019524868577719, + 0.0026028365828096867, + 0.0016301164869219065, + 0.0012833198998123407, + 0.003507428802549839, + 0.0028319749981164932, + 0.003341993549838662, + 0.0024319670628756285, + 0.0034509601537138224, + 0.007871825248003006, + 0.0016647264128550887, + 0.004313777666538954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.04872574284672737 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009813044220209122 + ], + "opposite_logprob_tokens": [ + -4.634812831878662 + ], + "logprob_reward": -0.009813044220209122, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009813044220209122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09539587795734406, + 0.05925509333610535, + 0.08960597962141037, + 0.10167549550533295, + 0.07923371344804764, + 0.102069191634655, + 0.12490707635879517, + 0.1309996247291565, + 0.0711071565747261, + 0.1457507610321045 + ], + "credit_map": [ + 0.011293093673884869, + 0.007865102961659431, + 0.021022198721766472, + 0.012083085253834724, + 0.008578210137784481, + 0.007494116667658091, + 0.012306399643421173, + 0.014753676019608974, + 0.008959833532571793, + 0.0048841917887330055, + 0.0058096544817090034, + 0.006307217758148909, + 0.008324910886585712, + 0.009284150786697865, + 0.007587844971567392, + 0.008097290992736816, + 0.017441073432564735, + 0.006309614051133394, + 0.011872095987200737, + 0.008345570415258408, + 0.00939879845827818, + 0.014162495732307434, + 0.008876361884176731, + 0.013199970126152039, + 0.010883127339184284, + 0.010917948558926582, + 0.010945071466267109, + 0.007711195386946201, + 0.015302627347409725, + 0.02104726806282997, + 0.012832432985305786, + 0.012035822495818138, + 0.008518978953361511, + 0.00849863886833191, + 0.011144253425300121, + 0.015490425750613213, + 0.010988712310791016, + 0.005889626685529947, + 0.008887222968041897, + 0.009815849363803864, + 0.017422756180167198, + 0.009884544648230076, + 0.017924901098012924, + 0.012497332878410816, + 0.01203662995249033, + 0.012800161726772785, + 0.009130913764238358, + 0.010371951386332512, + 0.008733098395168781, + 0.008600296452641487, + 0.01184918638318777, + 0.012726383283734322, + 0.014937857165932655, + 0.034366898238658905, + 0.012627705931663513, + 0.021065643057227135, + 0.008980737999081612, + 0.011744694784283638, + 0.014133226126432419, + 0.01350861694663763, + 0.016560668125748634, + 0.037028536200523376, + 0.011085117235779762, + 0.017958030104637146, + 0.006854640319943428, + 0.006598300766199827, + 0.008600757457315922, + 0.003957942593842745, + 0.009399458765983582, + 0.019994378089904785, + 0.006774326786398888, + 0.008927352726459503, + 0.02613963931798935, + 0.013851210474967957, + 0.017482975497841835, + 0.012607846409082413, + 0.013633930124342442, + 0.03465401008725166, + 0.007834089919924736, + 0.01954706385731697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.009813044220209122 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521456480026245 + ], + "opposite_logprob_tokens": [ + -1.5021456480026245 + ], + "logprob_reward": -0.2521456480026245, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521456480026245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07166437804698944, + 0.04290448874235153, + 0.046628061681985855, + 0.041167087852954865, + 0.05848062410950661, + 0.04669426381587982, + 0.31591445207595825, + 0.22041194140911102, + 0.1139654666185379, + 0.04216926917433739 + ], + "credit_map": [ + 0.007002986501902342, + 0.005905298516154289, + 0.00973924994468689, + 0.009993135929107666, + 0.008987825363874435, + 0.009012246504426003, + 0.010413427837193012, + 0.010610203258693218, + 0.004864979535341263, + 0.003379808273166418, + 0.004555803723633289, + 0.004682473838329315, + 0.005529702641069889, + 0.007472172845155001, + 0.006211773958057165, + 0.006207770202308893, + 0.0027886689640581608, + 0.004846989177167416, + 0.006466492544859648, + 0.005297898314893246, + 0.007977859117090702, + 0.004728211089968681, + 0.008345468901097775, + 0.006176468450576067, + 0.006459731608629227, + 0.00353153795003891, + 0.004912888165563345, + 0.0026572870556265116, + 0.004623464774340391, + 0.007873092778027058, + 0.00647750636562705, + 0.004631574731320143, + 0.005766432266682386, + 0.005321220960468054, + 0.011661898344755173, + 0.012568838894367218, + 0.00573579641059041, + 0.004125920124351978, + 0.005671062972396612, + 0.00762944994494319, + 0.007463290821760893, + 0.004568641539663076, + 0.005245243664830923, + 0.004836099222302437, + 0.008055873215198517, + 0.005216089077293873, + 0.00591228948906064, + 0.005396737717092037, + 0.02701316773891449, + 0.028670264407992363, + 0.023308895528316498, + 0.02594318427145481, + 0.03531302511692047, + 0.09943247586488724, + 0.02744133770465851, + 0.048792097717523575, + 0.016224397346377373, + 0.020152773708105087, + 0.01709330454468727, + 0.017963990569114685, + 0.030573904514312744, + 0.07234911620616913, + 0.013390107080340385, + 0.03266433626413345, + 0.01861191727221012, + 0.010720937512814999, + 0.009549538604915142, + 0.007401104550808668, + 0.015571183525025845, + 0.030348103493452072, + 0.007635185495018959, + 0.014127488248050213, + 0.004850474651902914, + 0.0038706138730049133, + 0.004822935443371534, + 0.0029613827355206013, + 0.005068338941782713, + 0.011503378860652447, + 0.0025972984731197357, + 0.006494843866676092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.2521456480026245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a strikingly close-up view of a young man sleeping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, off-white cotton hooded sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is warm afternoon sunlight filtering through swaying leaves to cast a shifting, soft pattern on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are green fields visible through the car window behind him, appearing as a rapidly receding blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image shrouded in a warm golden halo with dominant soft beige and gold colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790615178644657 + ], + "opposite_logprob_tokens": [ + -4.634790420532227 + ], + "logprob_reward": -0.009790615178644657, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790615178644657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10719393193721771, + 0.08039160072803497, + 0.054838359355926514, + 0.08491713553667068, + 0.09245315939188004, + 0.04952409490942955, + 0.19189177453517914, + 0.18932685256004333, + 0.1086684837937355, + 0.04079468548297882 + ], + "credit_map": [ + 0.01261045504361391, + 0.011344017460942268, + 0.018641842529177666, + 0.012601848691701889, + 0.01422489620745182, + 0.0072693065740168095, + 0.014655841514468193, + 0.015845727175474167, + 0.010533440858125687, + 0.006695899181067944, + 0.007883058860898018, + 0.012398303486406803, + 0.014376292936503887, + 0.007599883247166872, + 0.01107914000749588, + 0.009825579822063446, + 0.009352600201964378, + 0.004867244977504015, + 0.007105314638465643, + 0.005084426142275333, + 0.005593659356236458, + 0.006510572042316198, + 0.008557568304240704, + 0.0077669695019721985, + 0.009177712723612785, + 0.00841795839369297, + 0.011086457408964634, + 0.00635420111939311, + 0.009212896227836609, + 0.016706479713320732, + 0.013718933798372746, + 0.010242490097880363, + 0.010351504199206829, + 0.011180524714291096, + 0.01581605337560177, + 0.014065800234675407, + 0.011669926345348358, + 0.007090156432241201, + 0.008606759831309319, + 0.013672430999577045, + 0.007858144119381905, + 0.0062021962366998196, + 0.006225929129868746, + 0.007548539433628321, + 0.004793539177626371, + 0.006666300352662802, + 0.004269598051905632, + 0.005959846545010805, + 0.024411218240857124, + 0.017720624804496765, + 0.01433697622269392, + 0.013449442572891712, + 0.021657191216945648, + 0.055545005947351456, + 0.01710665412247181, + 0.027664661407470703, + 0.014694753102958202, + 0.018119635060429573, + 0.01499196607619524, + 0.013451737351715565, + 0.02653580904006958, + 0.06138882413506508, + 0.010877178981900215, + 0.029266949743032455, + 0.013562082313001156, + 0.009644373320043087, + 0.007325869519263506, + 0.00641090190038085, + 0.015945585444569588, + 0.03324778378009796, + 0.007709655445069075, + 0.014822232536971569, + 0.004773562774062157, + 0.004490627907216549, + 0.006256149150431156, + 0.004109199158847332, + 0.0048486338928341866, + 0.007912755012512207, + 0.0035384302027523518, + 0.004865326453000307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.009790615178644657 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_14/index_tensor([8764, 8764])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_14/index_tensor([8764, 8764])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..95b0d84327fdeaa9a5bcd0c51244c5811e6a736b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_14/index_tensor([8764, 8764])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up shot from a level angle centered a mechanical tentacle gleaming with a cold metallic sheen. Its segmented, gunmetal-gray shell was covered in meticulously carved lines and faint signs of wear. At its tip, a sharp, multi-faceted probe pierced the black glass surface of a modern, borderless monitor with cold, precise precision. The moment of contact, the blue data stream flowing across the screen collapsed, exploding from the point of penetration. Countless sharp streaks of cyan, magenta, and yellow digital glitches flashed wildly, blending with a pixelated mosaic to form a visual digital storm. Particles of shattered glass flew outward, accompanied by sparks of arcing electricity. Then, the camera began to steadily and smoothly pull back, revealing that the mechanical tentacle was only part of a huge, complex mechanical body lurking in the shadows, and the destroyed monitor was embedded in a row of server cabinets. The entire dim control room environment was illuminated by the only, chaotic and stroboscopic light emitted by the dying screen, creating a tense and destructive atmosphere of a high-tech system being violently invaded.", + "seed": 17704, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_14/index_8764_seed_17704_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04068217799067497, + "vlm_logprob": -0.04068217799067497, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.04068217799067497, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05741693079471588 + ], + "opposite_logprob_tokens": [ + -2.9324169158935547 + ], + "logprob_reward": -0.05741693079471588, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05741693079471588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2174997478723526, + 0.0981411412358284, + 0.1218351498246193, + 0.10648205131292343, + 0.14335831999778748, + 0.049506548792123795, + 0.09944578260183334, + 0.023213986307382584, + 0.08832208812236786, + 0.052195169031620026 + ], + "credit_map": [ + 0.03843303024768829, + 0.019734712317585945, + 0.03349277377128601, + 0.024886159226298332, + 0.024435317143797874, + 0.03282667323946953, + 0.021371684968471527, + 0.02231939323246479, + 0.01360535342246294, + 0.008510900661349297, + 0.014380625449120998, + 0.008203490637242794, + 0.01957497000694275, + 0.016698556020855904, + 0.009299105033278465, + 0.007868140935897827, + 0.05173517018556595, + 0.004326730500906706, + 0.046225644648075104, + 0.0043386733159422874, + 0.002680066507309675, + 0.006287512369453907, + 0.0028347382321953773, + 0.0034066198859363794, + 0.013366232626140118, + 0.007637667935341597, + 0.014787948690354824, + 0.02256315015256405, + 0.010687456466257572, + 0.014343086630105972, + 0.012162701226770878, + 0.010933805257081985, + 0.013042490929365158, + 0.017582420259714127, + 0.013341890648007393, + 0.011678261682391167, + 0.03331566974520683, + 0.022755904123187065, + 0.015756702050566673, + 0.015884986147284508, + 0.006226013880223036, + 0.0069808512926101685, + 0.00755195040255785, + 0.005506314802914858, + 0.009331529960036278, + 0.004184699151664972, + 0.005127157550305128, + 0.00459803082048893, + 0.012860077433288097, + 0.00838993676006794, + 0.008320135995745659, + 0.00883813388645649, + 0.015170914120972157, + 0.01648798957467079, + 0.013606473803520203, + 0.015772130340337753, + 0.0036293109878897667, + 0.0019821992609649897, + 0.003579238895326853, + 0.002705878810957074, + 0.002856081584468484, + 0.0020959859248250723, + 0.0030921115539968014, + 0.003273179056122899, + 0.018930738791823387, + 0.01117336843162775, + 0.012169972993433475, + 0.012232468463480473, + 0.008403806947171688, + 0.007796090096235275, + 0.009318425320088863, + 0.008297213353216648, + 0.007789038121700287, + 0.006215113215148449, + 0.00639738142490387, + 0.00784500502049923, + 0.008053062483668327, + 0.005436541512608528, + 0.007172503042966127, + 0.0032865239772945642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.05741693079471588 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033713892102241516 + ], + "opposite_logprob_tokens": [ + -3.4087138175964355 + ], + "logprob_reward": -0.033713892102241516, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033713892102241516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11097094416618347, + 0.043703529983758926, + 0.137918159365654, + 0.09589507430791855, + 0.09534820914268494, + 0.061420805752277374, + 0.11395065486431122, + 0.052123744040727615, + 0.13797670602798462, + 0.1506921499967575 + ], + "credit_map": [ + 0.01949838548898697, + 0.008364085108041763, + 0.015853777527809143, + 0.013447179459035397, + 0.012695491313934326, + 0.01980908028781414, + 0.009996514767408371, + 0.011306431144475937, + 0.006524373311549425, + 0.0043863048776984215, + 0.006563407834619284, + 0.00462864525616169, + 0.005695653613656759, + 0.007135104387998581, + 0.003600313328206539, + 0.005169726442545652, + 0.05242365971207619, + 0.0066298493184149265, + 0.04865435138344765, + 0.005769878625869751, + 0.006142702419310808, + 0.00874082650989294, + 0.005193993449211121, + 0.004362890031188726, + 0.011488229036331177, + 0.009038135409355164, + 0.011303285136818886, + 0.012655026279389858, + 0.007476076018065214, + 0.011575741693377495, + 0.020593907684087753, + 0.011764680966734886, + 0.009630396030843258, + 0.011755670420825481, + 0.00926552526652813, + 0.010333098471164703, + 0.021542122587561607, + 0.01076404470950365, + 0.010708685033023357, + 0.01134866289794445, + 0.0072541795670986176, + 0.008298797532916069, + 0.007016741670668125, + 0.006653680466115475, + 0.011973867192864418, + 0.006170290056616068, + 0.007588901091367006, + 0.006464347708970308, + 0.01201446820050478, + 0.010462522506713867, + 0.014070989564061165, + 0.011664551682770252, + 0.01618681289255619, + 0.014689343981444836, + 0.017543241381645203, + 0.017318719998002052, + 0.004927290603518486, + 0.005884560756385326, + 0.006494087632745504, + 0.007029250264167786, + 0.005351909901946783, + 0.007276791147887707, + 0.007224894128739834, + 0.007934962399303913, + 0.03453318402171135, + 0.014221685007214546, + 0.021981913596391678, + 0.01877020299434662, + 0.012455666437745094, + 0.011811701580882072, + 0.01084036473184824, + 0.013362000696361065, + 0.02923016995191574, + 0.01714441552758217, + 0.02394505962729454, + 0.024691738188266754, + 0.017906006425619125, + 0.013236564584076405, + 0.0148407481610775, + 0.009697441011667252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.033713892102241516 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589664578437805 + ], + "opposite_logprob_tokens": [ + -3.895589590072632 + ], + "logprob_reward": -0.020589664578437805, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589664578437805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08670027554035187, + 0.14991430938243866, + 0.24879398941993713, + 0.05543612688779831, + 0.08703344315290451, + 0.04203498736023903, + 0.10751689225435257, + 0.09517646580934525, + 0.07564578950405121, + 0.0517476424574852 + ], + "credit_map": [ + 0.01137295737862587, + 0.012156778946518898, + 0.008955910801887512, + 0.009645170532166958, + 0.01093839667737484, + 0.01295049861073494, + 0.008615029975771904, + 0.01206553541123867, + 0.023110685870051384, + 0.01594618707895279, + 0.0235061626881361, + 0.009834214113652706, + 0.03346945345401764, + 0.02083647809922695, + 0.011886555701494217, + 0.011324587278068066, + 0.10162847489118576, + 0.010154318995773792, + 0.09284225106239319, + 0.008166328072547913, + 0.008404710330069065, + 0.01358311902731657, + 0.007155995350331068, + 0.006858782842755318, + 0.007964104413986206, + 0.0049045030027627945, + 0.010627636685967445, + 0.006155749782919884, + 0.004859509412199259, + 0.00504404166713357, + 0.009451182559132576, + 0.00642940029501915, + 0.008820203132927418, + 0.009022794663906097, + 0.016292816027998924, + 0.01030817348510027, + 0.01107283029705286, + 0.010390236042439938, + 0.010458361357450485, + 0.010668027214705944, + 0.00447525130584836, + 0.004607217852026224, + 0.005430653225630522, + 0.004724083002656698, + 0.007114619482308626, + 0.0052219415083527565, + 0.005119914188981056, + 0.005341305863112211, + 0.011828196235001087, + 0.00991066638380289, + 0.0169075895100832, + 0.012001111172139645, + 0.016732558608055115, + 0.015292767435312271, + 0.01130293682217598, + 0.01354107353836298, + 0.010150667279958725, + 0.009980336762964725, + 0.012492025271058083, + 0.013396617025136948, + 0.013554306700825691, + 0.009738919325172901, + 0.012825937010347843, + 0.013037660159170628, + 0.015271710231900215, + 0.006819311995059252, + 0.006174038629978895, + 0.011421157978475094, + 0.006312808021903038, + 0.00977533869445324, + 0.0060768136754632, + 0.013794613070786, + 0.0071079544723033905, + 0.007135479710996151, + 0.00539006432518363, + 0.008449582383036613, + 0.008637867867946625, + 0.004594790283590555, + 0.0069303992204368114, + 0.003501501865684986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.020589664578437805 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863990470767021 + ], + "opposite_logprob_tokens": [ + -3.048640012741089 + ], + "logprob_reward": -0.04863990470767021, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863990470767021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10194309055805206, + 0.1332257091999054, + 0.09170086681842804, + 0.05993598327040672, + 0.10654760897159576, + 0.057345595210790634, + 0.07669529318809509, + 0.10630040615797043, + 0.06877195835113525, + 0.197533518075943 + ], + "credit_map": [ + 0.015592498704791069, + 0.011416575871407986, + 0.015676608309149742, + 0.011210198514163494, + 0.01392056979238987, + 0.015235486440360546, + 0.010183288715779781, + 0.008707863278687, + 0.026273582130670547, + 0.011992854997515678, + 0.02790236473083496, + 0.009390272200107574, + 0.012649829499423504, + 0.029079245403409004, + 0.007992666214704514, + 0.007944890297949314, + 0.02023586817085743, + 0.008556408807635307, + 0.01708766631782055, + 0.006950901821255684, + 0.009005067870020866, + 0.016262149438261986, + 0.008451906964182854, + 0.0051508997566998005, + 0.006762286648154259, + 0.005101905670017004, + 0.008420410566031933, + 0.008813024498522282, + 0.0070270695723593235, + 0.007147636730223894, + 0.008390557020902634, + 0.008273089304566383, + 0.00985637679696083, + 0.01177250500768423, + 0.012193792499601841, + 0.012291928753256798, + 0.014703557826578617, + 0.0194594357162714, + 0.009797911159694195, + 0.016472110524773598, + 0.006905816029757261, + 0.005285768769681454, + 0.00934350211173296, + 0.005022954661399126, + 0.012156219221651554, + 0.007505094166845083, + 0.0050441971980035305, + 0.006082037463784218, + 0.008949163369834423, + 0.007518360391259193, + 0.009911797009408474, + 0.00931097473949194, + 0.010668637230992317, + 0.01046131644397974, + 0.010630294680595398, + 0.009244748391211033, + 0.0202105063945055, + 0.012068134732544422, + 0.013109317049384117, + 0.012650479562580585, + 0.013717846013605595, + 0.009076082147657871, + 0.01210566982626915, + 0.013362374156713486, + 0.015486441552639008, + 0.006751071661710739, + 0.008037089370191097, + 0.008537962101399899, + 0.008104179985821247, + 0.00840846635401249, + 0.006249132566154003, + 0.007197610102593899, + 0.021463721990585327, + 0.02610114961862564, + 0.01373008731752634, + 0.028184136375784874, + 0.03258756175637245, + 0.02121082879602909, + 0.03993948549032211, + 0.01431653369218111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.04863990470767021 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305049777030945 + ], + "opposite_logprob_tokens": [ + -3.168050527572632 + ], + "logprob_reward": -0.04305049777030945, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305049777030945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21654069423675537, + 0.1489911824464798, + 0.08122620731592178, + 0.1005011722445488, + 0.07818230986595154, + 0.043093472719192505, + 0.06767577677965164, + 0.04805808514356613, + 0.11005616933107376, + 0.10567492246627808 + ], + "credit_map": [ + 0.03623234108090401, + 0.021671434864401817, + 0.03524445742368698, + 0.02558896876871586, + 0.023002447560429573, + 0.03688697889447212, + 0.017662370577454567, + 0.020251696929335594, + 0.02198680117726326, + 0.01860632933676243, + 0.01993964985013008, + 0.012997147627174854, + 0.02563641406595707, + 0.021959582343697548, + 0.011546565219759941, + 0.016318682581186295, + 0.023381123319268227, + 0.005446641705930233, + 0.021535387262701988, + 0.005979950539767742, + 0.005533214658498764, + 0.008375541307032108, + 0.005158901680260897, + 0.005815453361719847, + 0.010706583969295025, + 0.00811704434454441, + 0.014234366826713085, + 0.022431712597608566, + 0.008787685073912144, + 0.011191075667738914, + 0.012554875575006008, + 0.012477832846343517, + 0.008546954020857811, + 0.008428565226495266, + 0.01262044906616211, + 0.008773627690970898, + 0.012426990084350109, + 0.013331578113138676, + 0.007133739069104195, + 0.006920404266566038, + 0.005169203504920006, + 0.005712292157113552, + 0.006621969398111105, + 0.005125164054334164, + 0.004797964822500944, + 0.005624669604003429, + 0.004511271137744188, + 0.0055309380404651165, + 0.005424662958830595, + 0.0058787246234714985, + 0.01046146359294653, + 0.007002468686550856, + 0.00761752063408494, + 0.009556560777127743, + 0.01258905790746212, + 0.00914531946182251, + 0.007932332344353199, + 0.004031837452203035, + 0.0065976218320429325, + 0.005397434812039137, + 0.005340944975614548, + 0.005637068301439285, + 0.006492034066468477, + 0.006628811359405518, + 0.028620358556509018, + 0.008240261115133762, + 0.012486315332353115, + 0.01625068299472332, + 0.010301640257239342, + 0.012458117678761482, + 0.006952790077775717, + 0.014746006578207016, + 0.01718115247786045, + 0.010533150285482407, + 0.009702067822217941, + 0.01955278031527996, + 0.013429485261440277, + 0.010557375848293304, + 0.014131543226540089, + 0.010587358847260475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.04305049777030945 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot from a level angle centered a mechanical tentacle gleaming with a cold metallic sheen. Its segmented, gunmetal-gray shell was covered in meticulously carved lines and faint signs of wear. At its tip, a sharp, multi-faceted probe pierced the black glass surface of a modern, borderless monitor with cold, precise precision. The moment of contact, the blue data stream flowing across the screen collapsed, exploding from the point of penetration. Countless sharp streaks of cyan, magenta, and yellow digital glitches flashed wildly, blending with a pixelated mosaic to form a visual digital storm. Particles of shattered glass flew outward, accompanied by sparks of arcing electricity. Then, the camera began to steadily and smoothly pull back, revealing that the mechanical tentacle was only part of a huge, complex mechanical body lurking in the shadows, and the destroyed monitor was embedded in a row of server cabinets. The entire dim control room environment was illuminated by the only, chaotic and stroboscopic light emitted by the dying screen, creating a tense and destructive atmosphere of a high-tech system being violently invaded.", + "seed": 17705, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_14/index_8764_seed_17705_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.032473881356418134, + "vlm_logprob": -0.032473881356418134, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.032473881356418134, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04947017505764961 + ], + "opposite_logprob_tokens": [ + -3.0494701862335205 + ], + "logprob_reward": -0.04947017505764961, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04947017505764961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31209269165992737, + 0.20802636444568634, + 0.07625722140073776, + 0.08598366379737854, + 0.02660108357667923, + 0.04635941982269287, + 0.11086896061897278, + 0.051708705723285675, + 0.028523210436105728, + 0.053578782826662064 + ], + "credit_map": [ + 0.03803471848368645, + 0.04189252853393555, + 0.044759318232536316, + 0.02884729765355587, + 0.026093291118741035, + 0.02251327969133854, + 0.06400775909423828, + 0.04594448208808899, + 0.030262580141425133, + 0.027720237150788307, + 0.02956967242062092, + 0.020155098289251328, + 0.02172860875725746, + 0.009808375500142574, + 0.032266341149806976, + 0.03651542216539383, + 0.007722442038357258, + 0.004124784376472235, + 0.009746881201863289, + 0.004815820604562759, + 0.0054075527004897594, + 0.01355479471385479, + 0.0071374657563865185, + 0.02374747209250927, + 0.006931492127478123, + 0.010118382051587105, + 0.004607940558344126, + 0.012146055698394775, + 0.009266857989132404, + 0.017180034890770912, + 0.005000550299882889, + 0.020732343196868896, + 0.001766644068993628, + 0.004453390371054411, + 0.004098993726074696, + 0.0026460839435458183, + 0.002424120670184493, + 0.004358004312962294, + 0.0024279507342725992, + 0.004425891675055027, + 0.0035433312878012657, + 0.005438472609966993, + 0.005979630164802074, + 0.007834048010408878, + 0.004327480215579271, + 0.006711248308420181, + 0.0060875918716192245, + 0.006437608040869236, + 0.00930738728493452, + 0.015254049561917782, + 0.013728350400924683, + 0.020859021693468094, + 0.016496438533067703, + 0.0059165311977267265, + 0.01770414412021637, + 0.01160302385687828, + 0.005122633185237646, + 0.003700096160173416, + 0.009832640178501606, + 0.005283509846776724, + 0.0038328636437654495, + 0.007898504845798016, + 0.00476989196613431, + 0.011268561705946922, + 0.0046753715723752975, + 0.0030311099253594875, + 0.004629423841834068, + 0.00251292297616601, + 0.0034998892806470394, + 0.0046165818348526955, + 0.002972780028358102, + 0.0025851286482065916, + 0.0047695329412817955, + 0.00418724725022912, + 0.01297434326261282, + 0.0036645138170570135, + 0.00507460068911314, + 0.012195578776299953, + 0.004935642704367638, + 0.005777318961918354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.04947017505764961 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026361405849456787 + ], + "opposite_logprob_tokens": [ + -3.6513614654541016 + ], + "logprob_reward": -0.026361405849456787, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026361405849456787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0535045862197876, + 0.0508304163813591, + 0.11379639804363251, + 0.32924723625183105, + 0.04370534047484398, + 0.04124622046947479, + 0.06712096184492111, + 0.06573127210140228, + 0.08909010887145996, + 0.14572742581367493 + ], + "credit_map": [ + 0.006005970761179924, + 0.0076018706895411015, + 0.007257079239934683, + 0.0072709848172962666, + 0.004517395049333572, + 0.003524115774780512, + 0.0088691096752882, + 0.00845805648714304, + 0.0082113491371274, + 0.007890033535659313, + 0.010621941648423672, + 0.0034107144456356764, + 0.00487325806170702, + 0.002658585086464882, + 0.008168362081050873, + 0.004996176343411207, + 0.015560416504740715, + 0.005407034885138273, + 0.007561448961496353, + 0.007058904971927404, + 0.006666642148047686, + 0.021428167819976807, + 0.012213031761348248, + 0.037900760769844055, + 0.010477833449840546, + 0.009591130539774895, + 0.007778243627399206, + 0.025208717212080956, + 0.008846493437886238, + 0.12665097415447235, + 0.009727470576763153, + 0.13096636533737183, + 0.003370571881532669, + 0.0057261185720562935, + 0.0041415756568312645, + 0.0032858627382665873, + 0.003151985350996256, + 0.009391695261001587, + 0.004258289933204651, + 0.010379242710769176, + 0.003512005088850856, + 0.003894207766279578, + 0.005439353175461292, + 0.007000685669481754, + 0.004706081934273243, + 0.005859550554305315, + 0.004645831882953644, + 0.00618850439786911, + 0.008821939118206501, + 0.008856069296598434, + 0.011929765343666077, + 0.005958086811006069, + 0.00756312720477581, + 0.00662889052182436, + 0.006449525710195303, + 0.010913549922406673, + 0.006821350660175085, + 0.00785151869058609, + 0.008279727771878242, + 0.011086264625191689, + 0.005384388379752636, + 0.008379719220101833, + 0.007652752101421356, + 0.010275544598698616, + 0.013676081784069538, + 0.007560616359114647, + 0.012759150937199593, + 0.007380282040685415, + 0.009244360961019993, + 0.01758495159447193, + 0.009158685803413391, + 0.011725983582437038, + 0.016487808898091316, + 0.013782410882413387, + 0.026422204449772835, + 0.013383209705352783, + 0.01336770597845316, + 0.03100333735346794, + 0.017673872411251068, + 0.01360688079148531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 2 + ], + "selected_reward": -0.026361405849456787 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783650748431683 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783650748431683, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783650748431683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11009350419044495, + 0.0778217762708664, + 0.14637458324432373, + 0.15232272446155548, + 0.057083770632743835, + 0.15249386429786682, + 0.052178721874952316, + 0.06033266335725784, + 0.0753725916147232, + 0.11592575907707214 + ], + "credit_map": [ + 0.011545178480446339, + 0.013757230713963509, + 0.01516713947057724, + 0.009037171490490437, + 0.011422846466302872, + 0.00886219646781683, + 0.023350274190306664, + 0.016951467841863632, + 0.009435545653104782, + 0.009306383319199085, + 0.010743779130280018, + 0.006720654666423798, + 0.008077959530055523, + 0.004353003576397896, + 0.010396946221590042, + 0.018787506967782974, + 0.0167225431650877, + 0.010772498324513435, + 0.015606030821800232, + 0.013930308632552624, + 0.011694769375026226, + 0.023259496316313744, + 0.015280653722584248, + 0.039108291268348694, + 0.021167879924178123, + 0.012024657800793648, + 0.015046870335936546, + 0.01947343349456787, + 0.009820220060646534, + 0.03158218413591385, + 0.02002662420272827, + 0.023180857300758362, + 0.005983315408229828, + 0.008966128341853619, + 0.0076772915199398994, + 0.005447608884423971, + 0.0035007658880203962, + 0.00964971724897623, + 0.003941642120480537, + 0.011917301453649998, + 0.010350684635341167, + 0.015149434097111225, + 0.021606823429465294, + 0.030454805120825768, + 0.01552360225468874, + 0.01318224985152483, + 0.026259783655405045, + 0.01996648870408535, + 0.0057479385286569595, + 0.007929613813757896, + 0.00669638579711318, + 0.0065805730409920216, + 0.005959003232419491, + 0.004263365641236305, + 0.007597575429826975, + 0.007404271513223648, + 0.005941960960626602, + 0.007265291642397642, + 0.010571952909231186, + 0.006816336885094643, + 0.005662271287292242, + 0.008462400175631046, + 0.006494737230241299, + 0.00911771785467863, + 0.012481489218771458, + 0.006248328369110823, + 0.012700412422418594, + 0.006332086864858866, + 0.008729947730898857, + 0.012017449364066124, + 0.008543850854039192, + 0.008319028653204441, + 0.01908779889345169, + 0.010312920436263084, + 0.025812814012169838, + 0.010384947061538696, + 0.008442488498985767, + 0.020592043176293373, + 0.010285397991538048, + 0.011007350869476795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.009783650748431683 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043057918548583984 + ], + "opposite_logprob_tokens": [ + -3.168057918548584 + ], + "logprob_reward": -0.043057918548583984, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043057918548583984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0983855128288269, + 0.08489672094583511, + 0.14242762327194214, + 0.0721498429775238, + 0.13856106996536255, + 0.05021769180893898, + 0.058475226163864136, + 0.13021717965602875, + 0.18482251465320587, + 0.03984659165143967 + ], + "credit_map": [ + 0.009156670421361923, + 0.014114717952907085, + 0.01320730708539486, + 0.012556242756545544, + 0.0087578808888793, + 0.00873563066124916, + 0.020350592210888863, + 0.011506467126309872, + 0.01562696136534214, + 0.009650379419326782, + 0.017392192035913467, + 0.005564971826970577, + 0.010610491968691349, + 0.005726502742618322, + 0.011771867983043194, + 0.008553355932235718, + 0.010940568521618843, + 0.01139073632657528, + 0.023862240836024284, + 0.006690796464681625, + 0.010263603180646896, + 0.025879835709929466, + 0.011116777546703815, + 0.042283061891794205, + 0.006760876160115004, + 0.006860068067908287, + 0.004389800596982241, + 0.01155199483036995, + 0.007450481411069632, + 0.01317561510950327, + 0.005688121076673269, + 0.016272883862257004, + 0.007967024110257626, + 0.014957506209611893, + 0.014221125282347202, + 0.008663764223456383, + 0.008221614174544811, + 0.03661021962761879, + 0.009115145541727543, + 0.03880468010902405, + 0.004408974200487137, + 0.004872680641710758, + 0.007372198160737753, + 0.005907576065510511, + 0.004531835205852985, + 0.009066554717719555, + 0.005508015863597393, + 0.008549857884645462, + 0.005959285888820887, + 0.009165270254015923, + 0.00827384926378727, + 0.006119645666331053, + 0.008162807673215866, + 0.006296582520008087, + 0.007900902070105076, + 0.00659688375890255, + 0.020036730915308, + 0.01358113344758749, + 0.029076961800456047, + 0.01027706265449524, + 0.008058563806116581, + 0.01076972484588623, + 0.009020102210342884, + 0.029396891593933105, + 0.025395724922418594, + 0.015079140663146973, + 0.02767845429480076, + 0.01170244812965393, + 0.017230987548828125, + 0.04726061969995499, + 0.01584022119641304, + 0.024634933099150658, + 0.00553950946778059, + 0.006011399440467358, + 0.0053034378215670586, + 0.0033404428977519274, + 0.0036814226768910885, + 0.006677333731204271, + 0.004267442971467972, + 0.005025602411478758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.043057918548583984 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336962565779686 + ], + "opposite_logprob_tokens": [ + -3.408696174621582 + ], + "logprob_reward": -0.0336962565779686, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336962565779686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13682952523231506, + 0.06801772117614746, + 0.10695992410182953, + 0.19512830674648285, + 0.0752536877989769, + 0.05195893347263336, + 0.1015111431479454, + 0.08575523644685745, + 0.12206153571605682, + 0.05652396008372307 + ], + "credit_map": [ + 0.016293175518512726, + 0.01772727258503437, + 0.017275918275117874, + 0.012881321832537651, + 0.011921648867428303, + 0.011598522774875164, + 0.020274721086025238, + 0.028856946155428886, + 0.007591160945594311, + 0.012334061786532402, + 0.009477258659899235, + 0.004830086603760719, + 0.009706171229481697, + 0.005206243600696325, + 0.009675418958067894, + 0.009197317995131016, + 0.01679990626871586, + 0.008229461498558521, + 0.015443110838532448, + 0.009169715456664562, + 0.008249437436461449, + 0.013108554296195507, + 0.012199535965919495, + 0.023760192096233368, + 0.010013509541749954, + 0.007233848329633474, + 0.0071874381974339485, + 0.01736711710691452, + 0.006963219027966261, + 0.06898508220911026, + 0.005997380241751671, + 0.07138071209192276, + 0.004548523109406233, + 0.005683224182575941, + 0.006381244398653507, + 0.004597758874297142, + 0.004341687075793743, + 0.02185080572962761, + 0.003962620627135038, + 0.02388782426714897, + 0.004391264170408249, + 0.005655432119965553, + 0.007495793513953686, + 0.00940384715795517, + 0.004721058998256922, + 0.004789451602846384, + 0.007780489977449179, + 0.007721597328782082, + 0.013383558951318264, + 0.01466897502541542, + 0.02043396607041359, + 0.007633368484675884, + 0.010863329283893108, + 0.007017289288341999, + 0.013155781663954258, + 0.014354866929352283, + 0.008447997272014618, + 0.006904780864715576, + 0.010981649160385132, + 0.007555103395134211, + 0.008657482452690601, + 0.018246835097670555, + 0.009081816300749779, + 0.015879575163125992, + 0.021607285365462303, + 0.010025247000157833, + 0.01895102486014366, + 0.010569984093308449, + 0.010694155469536781, + 0.02184944599866867, + 0.011060619726777077, + 0.017303774133324623, + 0.008275305852293968, + 0.0070455316454172134, + 0.012422742322087288, + 0.003787266556173563, + 0.0038477692287415266, + 0.00944431684911251, + 0.0050278836861252785, + 0.0066731455735862255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.0336962565779686 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot from a level angle centered a mechanical tentacle gleaming with a cold metallic sheen. Its segmented, gunmetal-gray shell was covered in meticulously carved lines and faint signs of wear. At its tip, a sharp, multi-faceted probe pierced the black glass surface of a modern, borderless monitor with cold, precise precision. The moment of contact, the blue data stream flowing across the screen collapsed, exploding from the point of penetration. Countless sharp streaks of cyan, magenta, and yellow digital glitches flashed wildly, blending with a pixelated mosaic to form a visual digital storm. Particles of shattered glass flew outward, accompanied by sparks of arcing electricity. Then, the camera began to steadily and smoothly pull back, revealing that the mechanical tentacle was only part of a huge, complex mechanical body lurking in the shadows, and the destroyed monitor was embedded in a row of server cabinets. The entire dim control room environment was illuminated by the only, chaotic and stroboscopic light emitted by the dying screen, creating a tense and destructive atmosphere of a high-tech system being violently invaded.", + "seed": 17706, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_14/index_8764_seed_17706_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06552562117576599, + "vlm_logprob": -0.06552562117576599, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.06552562117576599, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05456561595201492 + ], + "opposite_logprob_tokens": [ + -3.054565668106079 + ], + "logprob_reward": -0.05456561595201492, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05456561595201492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35959455370903015, + 0.19790709018707275, + 0.06924767047166824, + 0.07659653574228287, + 0.07217846065759659, + 0.02619493566453457, + 0.07458361983299255, + 0.03195760026574135, + 0.049274325370788574, + 0.042465176433324814 + ], + "credit_map": [ + 0.023563746362924576, + 0.020999467000365257, + 0.023998955264687538, + 0.06758669763803482, + 0.04265156388282776, + 0.0644255205988884, + 0.07161781191825867, + 0.04475083202123642, + 0.01437843032181263, + 0.015029788948595524, + 0.018062325194478035, + 0.017710600048303604, + 0.02317032963037491, + 0.039697080850601196, + 0.04494589567184448, + 0.024912621825933456, + 0.016150640323758125, + 0.006177178118377924, + 0.005669102072715759, + 0.005052617285400629, + 0.007583737839013338, + 0.012597363442182541, + 0.011579139158129692, + 0.004437888041138649, + 0.00780283659696579, + 0.013987616635859013, + 0.014801591634750366, + 0.0116578359156847, + 0.00971693079918623, + 0.005304987542331219, + 0.006157531403005123, + 0.007167206611484289, + 0.0064480421133339405, + 0.013629455119371414, + 0.021428605541586876, + 0.004978744778782129, + 0.0066319615580141544, + 0.008261581882834435, + 0.0054639047011733055, + 0.005336170084774494, + 0.0034200078807771206, + 0.002879953011870384, + 0.0029263522010296583, + 0.0022705323062837124, + 0.003357816021889448, + 0.004305111709982157, + 0.0033148627262562513, + 0.0037202988751232624, + 0.008257037028670311, + 0.006581156048923731, + 0.005588898900896311, + 0.010839821770787239, + 0.008479914627969265, + 0.009610466659069061, + 0.016983548179268837, + 0.008242773823440075, + 0.004819976631551981, + 0.003009677864611149, + 0.003875012742355466, + 0.003550530644133687, + 0.0027509292121976614, + 0.00384620763361454, + 0.005686026532202959, + 0.004419241100549698, + 0.007548250257968903, + 0.004810390993952751, + 0.004572713281959295, + 0.005638625007122755, + 0.00515331793576479, + 0.0074833896942436695, + 0.00552422134205699, + 0.00854342058300972, + 0.0075656878761947155, + 0.003315901616588235, + 0.005767819471657276, + 0.0036666549276560545, + 0.007299248594790697, + 0.007014458067715168, + 0.003755215322598815, + 0.004080190788954496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.05456561595201492 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048674650490283966 + ], + "opposite_logprob_tokens": [ + -3.0486745834350586 + ], + "logprob_reward": -0.048674650490283966, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048674650490283966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20449361205101013, + 0.07029330730438232, + 0.10137660056352615, + 0.06866481900215149, + 0.2515028417110443, + 0.049768026918172836, + 0.03572101145982742, + 0.07641731947660446, + 0.10255406051874161, + 0.039208486676216125 + ], + "credit_map": [ + 0.016921652480959892, + 0.011522631160914898, + 0.010619807988405228, + 0.03661825880408287, + 0.021297810599207878, + 0.03514275327324867, + 0.044186610728502274, + 0.0281840730458498, + 0.006073968019336462, + 0.0057092830538749695, + 0.00619062315672636, + 0.00814842153340578, + 0.006839357782155275, + 0.008701087906956673, + 0.01278023049235344, + 0.015850331634283066, + 0.022176964208483696, + 0.01305735856294632, + 0.014332665130496025, + 0.007812241092324257, + 0.008097887970507145, + 0.011639435775578022, + 0.01912255957722664, + 0.0051374915055930614, + 0.008146323263645172, + 0.013743488118052483, + 0.013862059451639652, + 0.007549090310931206, + 0.006517602596431971, + 0.00591174466535449, + 0.006033154670149088, + 0.00690135033801198, + 0.012316853739321232, + 0.08420377224683762, + 0.09832576662302017, + 0.013486268930137157, + 0.012184031307697296, + 0.013003242202103138, + 0.009544258937239647, + 0.008438645862042904, + 0.005878729745745659, + 0.005917245987802744, + 0.005419143009930849, + 0.006915941368788481, + 0.007035060320049524, + 0.005790372844785452, + 0.005974725820124149, + 0.006836803629994392, + 0.004524799529463053, + 0.004447480197995901, + 0.002923225983977318, + 0.005197013728320599, + 0.005424058064818382, + 0.005203744862228632, + 0.003946234937757254, + 0.00405445322394371, + 0.009185157716274261, + 0.012888564728200436, + 0.007635209709405899, + 0.010392574593424797, + 0.007324825972318649, + 0.007971239276230335, + 0.011121690273284912, + 0.009898047894239426, + 0.01890232227742672, + 0.007352230604737997, + 0.01091703586280346, + 0.011524672619998455, + 0.01000580657273531, + 0.01720736175775528, + 0.011383580975234509, + 0.015261048451066017, + 0.008403252810239792, + 0.0024741620291024446, + 0.004212032537907362, + 0.003138334956020117, + 0.005224436987191439, + 0.006356500089168549, + 0.004891588352620602, + 0.00450817309319973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.048674650490283966 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645722642540932 + ], + "opposite_logprob_tokens": [ + -4.758645534515381 + ], + "logprob_reward": -0.008645722642540932, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645722642540932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11000917106866837, + 0.06683743745088577, + 0.0982743427157402, + 0.09274598956108093, + 0.2903348207473755, + 0.06566537916660309, + 0.047531142830848694, + 0.0835876390337944, + 0.11617480218410492, + 0.028839319944381714 + ], + "credit_map": [ + 0.011222965084016323, + 0.014208468608558178, + 0.007537199184298515, + 0.016583401709794998, + 0.009408781304955482, + 0.01738840714097023, + 0.021457649767398834, + 0.012202288024127483, + 0.006785167381167412, + 0.0065322513692080975, + 0.008527103811502457, + 0.006879856344312429, + 0.009138551540672779, + 0.009955145418643951, + 0.011007876135408878, + 0.00801147986203432, + 0.01496204361319542, + 0.016638437286019325, + 0.00984292570501566, + 0.01796015165746212, + 0.009388035163283348, + 0.013254053890705109, + 0.008459755219519138, + 0.0077689350582659245, + 0.012109084986150265, + 0.012195464223623276, + 0.015617857687175274, + 0.015608387999236584, + 0.011276972480118275, + 0.00926741398870945, + 0.006713110953569412, + 0.009957684203982353, + 0.014302445575594902, + 0.10444317013025284, + 0.10923483967781067, + 0.014383330941200256, + 0.015097452327609062, + 0.012866485863924026, + 0.011113477870821953, + 0.008893598802387714, + 0.007167526055127382, + 0.010835365392267704, + 0.006884338334202766, + 0.006440117955207825, + 0.006096328608691692, + 0.015179762616753578, + 0.0058756680227816105, + 0.007186271250247955, + 0.005046124570071697, + 0.005104436539113522, + 0.003828952554613352, + 0.0061076669953763485, + 0.00788742396980524, + 0.00599920516833663, + 0.007300843019038439, + 0.006256486754864454, + 0.00983025785535574, + 0.010737872682511806, + 0.006931687705218792, + 0.014072239398956299, + 0.00849733967334032, + 0.01126191858202219, + 0.01279448438435793, + 0.009461823850870132, + 0.01536431722342968, + 0.011417845264077187, + 0.010730018839240074, + 0.019511260092258453, + 0.010472381487488747, + 0.01751149818301201, + 0.01633080653846264, + 0.014836667105555534, + 0.004581386223435402, + 0.002608698792755604, + 0.0025734209921211004, + 0.003011468332260847, + 0.005985800176858902, + 0.004170581232756376, + 0.0030582279432564974, + 0.002849734853953123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.008645722642540932 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20154987275600433 + ], + "opposite_logprob_tokens": [ + -1.7015498876571655 + ], + "logprob_reward": -0.20154987275600433, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20154987275600433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04285624623298645, + 0.02909552864730358, + 0.07777299731969833, + 0.04341968521475792, + 0.5569683909416199, + 0.06724503636360168, + 0.08224021643400192, + 0.016003409400582314, + 0.04999413341283798, + 0.034404411911964417 + ], + "credit_map": [ + 0.003801185404881835, + 0.0027659002225846052, + 0.003466953756287694, + 0.0057848552241921425, + 0.004102648235857487, + 0.007518044672906399, + 0.008905196562409401, + 0.006511462386697531, + 0.001894077518954873, + 0.0021553784608840942, + 0.0033916477113962173, + 0.0025269798934459686, + 0.003002050332725048, + 0.006993800401687622, + 0.005401259753853083, + 0.0037303329445421696, + 0.02020486816763878, + 0.005440223962068558, + 0.004706578329205513, + 0.00455884076654911, + 0.009359344840049744, + 0.01609615609049797, + 0.013133483938872814, + 0.00427350215613842, + 0.005292592570185661, + 0.007216436322778463, + 0.006670695263892412, + 0.006696569267660379, + 0.004897597711533308, + 0.005398237146437168, + 0.0030558735597878695, + 0.004191682208329439, + 0.025615859776735306, + 0.20401085913181305, + 0.21418580412864685, + 0.027273885905742645, + 0.028269002214074135, + 0.024370474740862846, + 0.019642513245344162, + 0.013599952682852745, + 0.006438312120735645, + 0.008729927241802216, + 0.008337362669408321, + 0.007810493465512991, + 0.007424910552799702, + 0.011768941767513752, + 0.009442436508834362, + 0.00729265995323658, + 0.009590198285877705, + 0.008300847373902798, + 0.0062148370780050755, + 0.011334759183228016, + 0.011902349069714546, + 0.010980932973325253, + 0.014848058111965656, + 0.009068235754966736, + 0.002610007766634226, + 0.002654476324096322, + 0.0013822244945913553, + 0.0015971757238730788, + 0.00167036650236696, + 0.0018198871985077858, + 0.002756717847660184, + 0.0015125529607757926, + 0.008487922139465809, + 0.005797340534627438, + 0.004634658340364695, + 0.0051666018553078175, + 0.005672858562320471, + 0.007395146880298853, + 0.005196757148951292, + 0.007642850745469332, + 0.005918019451200962, + 0.002502607414498925, + 0.004856803920120001, + 0.003127342788502574, + 0.00630785059183836, + 0.005239414516836405, + 0.003351195016875863, + 0.003101176582276821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.20154987275600433 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192244037985802 + ], + "opposite_logprob_tokens": [ + -4.2641921043396 + ], + "logprob_reward": -0.014192244037985802, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192244037985802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24380426108837128, + 0.09380261600017548, + 0.12024754285812378, + 0.04569842666387558, + 0.19149410724639893, + 0.027016643434762955, + 0.054279908537864685, + 0.037893861532211304, + 0.11215262115001678, + 0.07360999286174774 + ], + "credit_map": [ + 0.03128809854388237, + 0.031944505870342255, + 0.017268138006329536, + 0.03763324394822121, + 0.018855702131986618, + 0.03876788541674614, + 0.04397425800561905, + 0.02407243847846985, + 0.009138287045061588, + 0.009109913371503353, + 0.011913834139704704, + 0.009709053672850132, + 0.012047262862324715, + 0.015282458625733852, + 0.015878019854426384, + 0.010723789222538471, + 0.027409417554736137, + 0.01622251234948635, + 0.01198651734739542, + 0.014151979237794876, + 0.013045698404312134, + 0.014681357890367508, + 0.017824476584792137, + 0.004925583489239216, + 0.006894615478813648, + 0.005386088974773884, + 0.00944737158715725, + 0.007271047681570053, + 0.004272737540304661, + 0.0032274872064590454, + 0.004938987549394369, + 0.004260090179741383, + 0.008737616240978241, + 0.06825347244739532, + 0.07388415932655334, + 0.009887684136629105, + 0.009245854802429676, + 0.009652618318796158, + 0.007140762638300657, + 0.004691938869655132, + 0.0035440411884337664, + 0.00354268541559577, + 0.003221663413569331, + 0.002194348257035017, + 0.002383152488619089, + 0.005796722136437893, + 0.0037267315201461315, + 0.002607299480587244, + 0.006838852073997259, + 0.004921630956232548, + 0.003381231799721718, + 0.007063052151352167, + 0.0059633394703269005, + 0.008288666605949402, + 0.010446231812238693, + 0.007376904599368572, + 0.0053388141095638275, + 0.005458284635096788, + 0.0030117533169686794, + 0.004855351522564888, + 0.004152277484536171, + 0.005387643352150917, + 0.005388737190514803, + 0.004300997592508793, + 0.01686697266995907, + 0.010868458077311516, + 0.007985821925103664, + 0.012150904163718224, + 0.00919316802173853, + 0.01818709634244442, + 0.01394230592995882, + 0.022957898676395416, + 0.018643677234649658, + 0.007708196993917227, + 0.009516952559351921, + 0.005727963987737894, + 0.007778209168463945, + 0.012835904024541378, + 0.005901393014937639, + 0.005497692618519068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.014192244037985802 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot from a level angle centered a mechanical tentacle gleaming with a cold metallic sheen. Its segmented, gunmetal-gray shell was covered in meticulously carved lines and faint signs of wear. At its tip, a sharp, multi-faceted probe pierced the black glass surface of a modern, borderless monitor with cold, precise precision. The moment of contact, the blue data stream flowing across the screen collapsed, exploding from the point of penetration. Countless sharp streaks of cyan, magenta, and yellow digital glitches flashed wildly, blending with a pixelated mosaic to form a visual digital storm. Particles of shattered glass flew outward, accompanied by sparks of arcing electricity. Then, the camera began to steadily and smoothly pull back, revealing that the mechanical tentacle was only part of a huge, complex mechanical body lurking in the shadows, and the destroyed monitor was embedded in a row of server cabinets. The entire dim control room environment was illuminated by the only, chaotic and stroboscopic light emitted by the dying screen, creating a tense and destructive atmosphere of a high-tech system being violently invaded.", + "seed": 17707, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_14/index_8764_seed_17707_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.016657229512929916, + "vlm_logprob": -0.016657229512929916, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.016657229512929916, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009838659316301346 + ], + "opposite_logprob_tokens": [ + -4.884838581085205 + ], + "logprob_reward": -0.009838659316301346, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.009838659316301346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20582754909992218, + 0.1725623607635498, + 0.05395158752799034, + 0.06823942810297012, + 0.06390415877103806, + 0.23887310922145844, + 0.060911890119314194, + 0.02824315056204796, + 0.04575832560658455, + 0.061728425323963165 + ], + "credit_map": [ + 0.02145564928650856, + 0.03027835302054882, + 0.025875896215438843, + 0.022980568930506706, + 0.020127013325691223, + 0.02272997982800007, + 0.029843434691429138, + 0.032536666840314865, + 0.02150627039372921, + 0.014720393344759941, + 0.03795890510082245, + 0.020704180002212524, + 0.009262146428227425, + 0.024408036842942238, + 0.015810830518603325, + 0.02819158509373665, + 0.007809741888195276, + 0.004510987084358931, + 0.010947652161121368, + 0.005559505894780159, + 0.003176124067977071, + 0.011949365958571434, + 0.00436506187543273, + 0.005633150227367878, + 0.008562243543565273, + 0.004288081079721451, + 0.010579998604953289, + 0.009523160755634308, + 0.0063780066557228565, + 0.008639223873615265, + 0.012583381496369839, + 0.00768533069640398, + 0.006812551524490118, + 0.0051284790970385075, + 0.011930981650948524, + 0.006791912950575352, + 0.006523442454636097, + 0.010061579756438732, + 0.007215382996946573, + 0.009439832530915737, + 0.021553872153162956, + 0.0629439651966095, + 0.03309614956378937, + 0.03116089105606079, + 0.03226032480597496, + 0.029424820095300674, + 0.013597147539258003, + 0.014835933223366737, + 0.005031488835811615, + 0.005992223042994738, + 0.009195777587592602, + 0.00713625131174922, + 0.008005580864846706, + 0.008943749591708183, + 0.007445998955518007, + 0.009160821326076984, + 0.004826754797250032, + 0.0028938413597643375, + 0.0023478271905332804, + 0.0026354528963565826, + 0.004801569506525993, + 0.004497855436056852, + 0.0021814603824168444, + 0.004058388527482748, + 0.0047363718040287495, + 0.0030871834605932236, + 0.004489891696721315, + 0.004809993784874678, + 0.009255457669496536, + 0.010355855338275433, + 0.004966349340975285, + 0.0040572211146354675, + 0.006802555173635483, + 0.0048071593046188354, + 0.006547245662659407, + 0.010207628831267357, + 0.007379382383078337, + 0.010393425822257996, + 0.008043277077376842, + 0.007547751534730196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.009838659316301346 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350490748882294 + ], + "opposite_logprob_tokens": [ + -3.651350498199463 + ], + "logprob_reward": -0.026350490748882294, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350490748882294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12196695804595947, + 0.09762846678495407, + 0.04611504077911377, + 0.030295081436634064, + 0.0786435678601265, + 0.3570462465286255, + 0.040003702044487, + 0.08669144660234451, + 0.09093831479549408, + 0.05067114531993866 + ], + "credit_map": [ + 0.013923019170761108, + 0.019945943728089333, + 0.015497277490794659, + 0.015018399804830551, + 0.008984234184026718, + 0.014189597219228745, + 0.017187975347042084, + 0.017220495268702507, + 0.011129037477076054, + 0.0062625654973089695, + 0.021802326664328575, + 0.010700675658881664, + 0.005381936673074961, + 0.015024635009467602, + 0.008897325024008751, + 0.018429959192872047, + 0.005511526484042406, + 0.0027872470673173666, + 0.00852432195097208, + 0.005574327893555164, + 0.003480554325506091, + 0.011110581457614899, + 0.004013535566627979, + 0.005112943705171347, + 0.004868191201239824, + 0.0013567100977525115, + 0.005758018232882023, + 0.0031669768504798412, + 0.0019446653313934803, + 0.005681382026523352, + 0.0032085038255900145, + 0.004310629330575466, + 0.00959204975515604, + 0.004642324987798929, + 0.01755221001803875, + 0.008499163202941418, + 0.005975899752229452, + 0.013219772838056087, + 0.006559092551469803, + 0.012603041715919971, + 0.031852878630161285, + 0.09938020259141922, + 0.041276536881923676, + 0.04733879864215851, + 0.05564158782362938, + 0.04253090173006058, + 0.018522094935178757, + 0.020503230392932892, + 0.004114649258553982, + 0.005506493616849184, + 0.005887907929718494, + 0.003908696584403515, + 0.005767103750258684, + 0.005326454993337393, + 0.00448629492893815, + 0.005006097257137299, + 0.012079714797437191, + 0.012633119709789753, + 0.008533567190170288, + 0.01062571071088314, + 0.012715577147901058, + 0.012191134504973888, + 0.00886286236345768, + 0.009049756452441216, + 0.008506196551024914, + 0.005977313499897718, + 0.007059349212795496, + 0.010586349293589592, + 0.02021077834069729, + 0.02208094857633114, + 0.007929081097245216, + 0.008588295429944992, + 0.0066606937907636166, + 0.00442859111353755, + 0.005013772752135992, + 0.00706856744363904, + 0.006196820177137852, + 0.008326335810124874, + 0.0072633144445717335, + 0.005713046994060278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.026350490748882294 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951184779405594 + ], + "opposite_logprob_tokens": [ + -5.130951404571533 + ], + "logprob_reward": -0.005951184779405594, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951184779405594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07799822092056274, + 0.221054807305336, + 0.03575017675757408, + 0.034963443875312805, + 0.11723147332668304, + 0.07567130029201508, + 0.036480821669101715, + 0.11854718625545502, + 0.21014948189258575, + 0.07215310633182526 + ], + "credit_map": [ + 0.01126356702297926, + 0.01533551700413227, + 0.00959515105932951, + 0.007972966879606247, + 0.006267374847084284, + 0.006423081737011671, + 0.008994756266474724, + 0.012145807035267353, + 0.028477191925048828, + 0.012912762351334095, + 0.05179775133728981, + 0.023476094007492065, + 0.011880178935825825, + 0.021343227475881577, + 0.01879732869565487, + 0.05237027257680893, + 0.004855554085224867, + 0.002331859664991498, + 0.007781605236232281, + 0.003979158122092485, + 0.0018946323543787003, + 0.008017778396606445, + 0.0032123704440891743, + 0.003677218686789274, + 0.0053405468352139, + 0.00204296107403934, + 0.005509017966687679, + 0.004552984610199928, + 0.0026296726427972317, + 0.006362777203321457, + 0.0028946271631866693, + 0.005630859639495611, + 0.01286406721919775, + 0.007227480877190828, + 0.027544597163796425, + 0.010970904491841793, + 0.011746479198336601, + 0.023123040795326233, + 0.006476570386439562, + 0.01727832853794098, + 0.008107535541057587, + 0.018839897587895393, + 0.00876858551055193, + 0.009556552395224571, + 0.011082830838859081, + 0.008988647721707821, + 0.005343392491340637, + 0.004983860068023205, + 0.003794717136770487, + 0.006139026489108801, + 0.0041627730242908, + 0.004197028931230307, + 0.005178631749004126, + 0.004739189520478249, + 0.0036567228380590677, + 0.004612732212990522, + 0.019010841846466064, + 0.012200785800814629, + 0.01144938264042139, + 0.015459112823009491, + 0.011142074130475521, + 0.018614137545228004, + 0.01747138611972332, + 0.013199463486671448, + 0.028351852670311928, + 0.015680504962801933, + 0.022793902084231377, + 0.028696604073047638, + 0.034766215831041336, + 0.035398729145526886, + 0.0291658416390419, + 0.015295821242034435, + 0.006796801928430796, + 0.0066479905508458614, + 0.007890133187174797, + 0.011668838560581207, + 0.008972639217972755, + 0.010141829960048199, + 0.011430760845541954, + 0.008604113012552261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.005951184779405594 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057168073952198 + ], + "opposite_logprob_tokens": [ + -3.895571708679199 + ], + "logprob_reward": -0.02057168073952198, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057168073952198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051056645810604095, + 0.3010949194431305, + 0.04189048334956169, + 0.03214172273874283, + 0.046140387654304504, + 0.20340117812156677, + 0.0782167837023735, + 0.09134311228990555, + 0.06695327907800674, + 0.08776143938302994 + ], + "credit_map": [ + 0.006837991997599602, + 0.006661531049758196, + 0.007116199471056461, + 0.007043962366878986, + 0.005366295110434294, + 0.006191272754222155, + 0.0058962213806807995, + 0.005943171679973602, + 0.038520947098731995, + 0.013059752061963081, + 0.07775206863880157, + 0.03174664452672005, + 0.015277734957635403, + 0.028211822733283043, + 0.025617478415369987, + 0.07090846449136734, + 0.006865049712359905, + 0.0024083296302706003, + 0.0092429518699646, + 0.003796202829107642, + 0.002426179125905037, + 0.01112862303853035, + 0.003004544647410512, + 0.0030186036601662636, + 0.0053069498389959335, + 0.0018031998770311475, + 0.0057357423938810825, + 0.004653484094887972, + 0.001998626859858632, + 0.005133609753102064, + 0.002804666757583618, + 0.004705445840954781, + 0.006194699555635452, + 0.002898125909268856, + 0.010205330327153206, + 0.0042296820320189, + 0.004746921826153994, + 0.007905961945652962, + 0.003273668698966503, + 0.00668599596247077, + 0.01768709346652031, + 0.05608349293470383, + 0.02129463478922844, + 0.028079453855752945, + 0.029246093705296516, + 0.02389923296868801, + 0.017598208039999008, + 0.009512976743280888, + 0.006435714662075043, + 0.00648475019261241, + 0.010650400072336197, + 0.00968281552195549, + 0.011983497999608517, + 0.010935744270682335, + 0.010890588164329529, + 0.01115327700972557, + 0.0073082479648292065, + 0.009456857107579708, + 0.009313132613897324, + 0.009491529315710068, + 0.009183499962091446, + 0.017292020842432976, + 0.010473479516804218, + 0.018824344500899315, + 0.006329150404781103, + 0.004339974839240313, + 0.004633814096450806, + 0.006655239965766668, + 0.013598136603832245, + 0.01395664643496275, + 0.008716425858438015, + 0.008723887614905834, + 0.009130710735917091, + 0.007140338886529207, + 0.007509775925427675, + 0.013347726315259933, + 0.011559235863387585, + 0.01695489138364792, + 0.012140550650656223, + 0.009978216141462326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.02057168073952198 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical tentacle with a gunmetal-gray shell and carved lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tentacle pierce a black glass monitor with a sharp, multi-faceted probe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the screen display a blue data stream that collapses and explodes into cyan, magenta, and yellow digital glitches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as an extreme close-up and then steadily pull back to reveal a huge mechanical body and server cabinets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by chaotic, stroboscopic light from the dying screen in a dim control room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020574131980538368 + ], + "opposite_logprob_tokens": [ + -3.8955740928649902 + ], + "logprob_reward": -0.020574131980538368, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020574131980538368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0905090942978859, + 0.23037438094615936, + 0.05108090117573738, + 0.03157666325569153, + 0.10618098080158234, + 0.1709330976009369, + 0.04448191449046135, + 0.06185409054160118, + 0.11308892071247101, + 0.09991998225450516 + ], + "credit_map": [ + 0.013450739905238152, + 0.015937156975269318, + 0.012465608306229115, + 0.011357085779309273, + 0.006609147880226374, + 0.008481637574732304, + 0.009222328662872314, + 0.012985387817025185, + 0.02984541840851307, + 0.015138479880988598, + 0.056722722947597504, + 0.026021448895335197, + 0.012178773991763592, + 0.019769104197621346, + 0.01888830028474331, + 0.051810119301080704, + 0.008195750415325165, + 0.0034308384638279676, + 0.011434429325163364, + 0.005201196298003197, + 0.0029641552828252316, + 0.01139903999865055, + 0.004093475639820099, + 0.004362016916275024, + 0.004341660067439079, + 0.0018031696090474725, + 0.005192007403820753, + 0.0049675507470965385, + 0.002528758253902197, + 0.005135202314704657, + 0.002681971527636051, + 0.004926344379782677, + 0.014193814247846603, + 0.005065010394901037, + 0.027639595791697502, + 0.00976544339209795, + 0.005632433574646711, + 0.01974775269627571, + 0.00830326322466135, + 0.015833670273423195, + 0.014871199615299702, + 0.04509030282497406, + 0.020555200055241585, + 0.024985987693071365, + 0.026306945830583572, + 0.019850945100188255, + 0.009819402359426022, + 0.009453115053474903, + 0.006054705008864403, + 0.006510986480861902, + 0.005949110724031925, + 0.003518559504300356, + 0.007160546723753214, + 0.006462725345045328, + 0.003155685728415847, + 0.005669597070664167, + 0.007674186956137419, + 0.005393608473241329, + 0.0071448166854679585, + 0.006885726470500231, + 0.007992702536284924, + 0.00911275576800108, + 0.005689583253115416, + 0.011960712261497974, + 0.010730747133493423, + 0.009138593450188637, + 0.013018961064517498, + 0.015993375331163406, + 0.012152374722063541, + 0.016887348145246506, + 0.01827315241098404, + 0.016894372180104256, + 0.008935155346989632, + 0.007278081960976124, + 0.008164439350366592, + 0.009812899865210056, + 0.016946598887443542, + 0.02345685288310051, + 0.010365852154791355, + 0.014960099942982197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 8 + ], + "selected_reward": -0.020574131980538368 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_15/index_tensor([1172, 1172])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_15/index_tensor([1172, 1172])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..553fa30d410da95d642b4f6553225f02c9dbca02 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_15/index_tensor([1172, 1172])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a gentle picture-book style, a close-up, static shot, shot from a slightly upward angle, focuses on a wondrous creature composed entirely of living vines and green leaves. Its flexible, emerald green vines intertwine to form its body and arms, punctuated by heart-shaped, distinctly veined leaves. Its face is a blooming white flower, similar to a sunflower but with a softer hue. Its eyes are gleaming with innocent black stamens at its center. In its hands, formed from delicate vines and tender leaves, it carefully holds and shapes a tiny, glowing seed. Its vine-like fingertips, with a loving, focused touch, slowly and steadily stroke the seed's shell. The background is a watercolor-blurred, softly illuminated forest. Suddenly, as a tiny green sprout gently peeked out from the top of the seed, a warm and magical golden halo instantly pulsated and emanated from it. This light not only made the seed itself brilliant, but also softly illuminated its slightly forward, curious flower face upwards, casting warm light and shadows on the petals, and letting its black stamens reflect little starlight. The whole picture was full of tranquility, miracle and the hope of the new life.", + "seed": 32714, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_15/index_1172_seed_32714_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03463755352422595, + "vlm_logprob": -0.03463755352422595, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.03463755352422595, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11573626101016998 + ], + "opposite_logprob_tokens": [ + -2.240736246109009 + ], + "logprob_reward": -0.11573626101016998, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11573626101016998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1774318516254425, + 0.07520419359207153, + 0.37630099058151245, + 0.10586286336183548, + 0.11193272471427917, + 0.018922390416264534, + 0.026108315214514732, + 0.04678615182638168, + 0.019950317218899727, + 0.04150021821260452 + ], + "credit_map": [ + 0.013922709040343761, + 0.01457719225436449, + 0.017028801143169403, + 0.012745260260999203, + 0.015130612067878246, + 0.04738839715719223, + 0.01868179254233837, + 0.037957094609737396, + 0.010476930998265743, + 0.00825341884046793, + 0.005810077767819166, + 0.007413485553115606, + 0.008351461961865425, + 0.014291536994278431, + 0.009657750837504864, + 0.010949524119496346, + 0.011739308014512062, + 0.00875459797680378, + 0.013768513686954975, + 0.021143916994333267, + 0.01578499935567379, + 0.15139728784561157, + 0.0124945268034935, + 0.14121785759925842, + 0.011103805154561996, + 0.01151364017277956, + 0.009623649530112743, + 0.010154199786484241, + 0.014487243257462978, + 0.021619854494929314, + 0.00925417710095644, + 0.01810629293322563, + 0.010700618848204613, + 0.011215436272323132, + 0.018356511369347572, + 0.014278658665716648, + 0.007062055170536041, + 0.019245123490691185, + 0.011791563592851162, + 0.0192827507853508, + 0.003124522976577282, + 0.0019806288182735443, + 0.0037434198893606663, + 0.002065073698759079, + 0.0014888111036270857, + 0.0022408582735806704, + 0.0017067876178771257, + 0.0025722873397171497, + 0.0023423749953508377, + 0.002618566155433655, + 0.0035835877060890198, + 0.0035202419385313988, + 0.003654863452538848, + 0.004396427888423204, + 0.0027418320532888174, + 0.003250422654673457, + 0.005300373770296574, + 0.005435672122985125, + 0.006880949251353741, + 0.004590289667248726, + 0.005410651210695505, + 0.00784825999289751, + 0.0037511407863348722, + 0.007568814791738987, + 0.004866487812250853, + 0.00277070514857769, + 0.003010798478499055, + 0.001860150252468884, + 0.0023412914015352726, + 0.002308361465111375, + 0.0013597524957731366, + 0.0014327699318528175, + 0.0067321681417524815, + 0.006502205040305853, + 0.005486570298671722, + 0.006187393795698881, + 0.0053596762008965015, + 0.00436395313590765, + 0.0038083670660853386, + 0.0030598831363022327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.11573626101016998 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060566283762455 + ], + "opposite_logprob_tokens": [ + -3.8956055641174316 + ], + "logprob_reward": -0.02060566283762455, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060566283762455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12291257083415985, + 0.05002908781170845, + 0.5066289901733398, + 0.11883009225130081, + 0.06564070284366608, + 0.028817448765039444, + 0.026742924004793167, + 0.05415710434317589, + 0.011338028125464916, + 0.014903039671480656 + ], + "credit_map": [ + 0.010745844803750515, + 0.007476538419723511, + 0.009489699266850948, + 0.006159927695989609, + 0.00852191261947155, + 0.037998754531145096, + 0.008786801248788834, + 0.033733099699020386, + 0.005192626267671585, + 0.005841020029038191, + 0.0031196593772619963, + 0.005655079148709774, + 0.005346068646758795, + 0.010695292614400387, + 0.004621019586920738, + 0.009558327496051788, + 0.01371448952704668, + 0.009889375418424606, + 0.017586352303624153, + 0.02330455370247364, + 0.011464811861515045, + 0.21574810147285461, + 0.00999275129288435, + 0.20492854714393616, + 0.01023164950311184, + 0.014854460954666138, + 0.011727850884199142, + 0.007150850258767605, + 0.027326328679919243, + 0.01962609775364399, + 0.01231067068874836, + 0.015602190047502518, + 0.006459721364080906, + 0.006002819165587425, + 0.01116899959743023, + 0.005721765570342541, + 0.011962021701037884, + 0.010165264829993248, + 0.006302200257778168, + 0.007857914082705975, + 0.003000871744006872, + 0.003257839707657695, + 0.004368339665234089, + 0.002948223613202572, + 0.003961784765124321, + 0.003649674588814378, + 0.0038343574851751328, + 0.0037963595241308212, + 0.0018737070495262742, + 0.0021657878533005714, + 0.003447978524491191, + 0.0025685566943138838, + 0.0033017140813171864, + 0.006500851828604937, + 0.003047329606488347, + 0.0038370033726096153, + 0.006534669082611799, + 0.004827054217457771, + 0.004727712832391262, + 0.004868110176175833, + 0.005365658085793257, + 0.011073248460888863, + 0.004370620474219322, + 0.012390032410621643, + 0.0015167421661317348, + 0.0011591778602451086, + 0.0012554002460092306, + 0.0012704605469480157, + 0.002060621976852417, + 0.0015803161077201366, + 0.001333216903731227, + 0.001162092201411724, + 0.0022527386900037527, + 0.0017845433903858066, + 0.002289317548274994, + 0.0015306383138522506, + 0.0026382827199995518, + 0.0016424821224063635, + 0.0016164195258170366, + 0.001148618757724762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.02060566283762455 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058405987918377 + ], + "opposite_logprob_tokens": [ + -3.8955841064453125 + ], + "logprob_reward": -0.02058405987918377, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058405987918377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19739282131195068, + 0.11080381274223328, + 0.10648234933614731, + 0.10943480581045151, + 0.15047962963581085, + 0.06337343156337738, + 0.07266504317522049, + 0.10297680646181107, + 0.04076329618692398, + 0.04562794044613838 + ], + "credit_map": [ + 0.014221199788153172, + 0.012096806429326534, + 0.014327515847980976, + 0.012256369926035404, + 0.013300369493663311, + 0.05715014785528183, + 0.024386784061789513, + 0.04965364187955856, + 0.011150489561259747, + 0.017299329861998558, + 0.009128410369157791, + 0.007989166304469109, + 0.016703946515917778, + 0.02181619592010975, + 0.010210911743342876, + 0.016505368053913116, + 0.00952374842017889, + 0.011988944374024868, + 0.008678491227328777, + 0.00751488609239459, + 0.012474537827074528, + 0.027736539021134377, + 0.011407463811337948, + 0.017157740890979767, + 0.009126733988523483, + 0.008383628912270069, + 0.009438734501600266, + 0.006795808207243681, + 0.01801326870918274, + 0.025111431255936623, + 0.008426248095929623, + 0.024138951674103737, + 0.016958478838205338, + 0.010854879394173622, + 0.02802657149732113, + 0.016127880662679672, + 0.018125057220458984, + 0.026679502800107002, + 0.012680144980549812, + 0.02102711983025074, + 0.007242110557854176, + 0.006093578413128853, + 0.01110615860670805, + 0.005712059326469898, + 0.008496475405991077, + 0.009079854004085064, + 0.00748163927346468, + 0.008161553181707859, + 0.007610004395246506, + 0.00696511659771204, + 0.007561877835541964, + 0.007188970223069191, + 0.015523807145655155, + 0.013336142525076866, + 0.006548932287842035, + 0.007930190302431583, + 0.01482469867914915, + 0.013795051723718643, + 0.01836714707314968, + 0.010514157824218273, + 0.009684385731816292, + 0.012887264601886272, + 0.008385131135582924, + 0.01451896596699953, + 0.005713633727282286, + 0.0037648698780685663, + 0.004573074635118246, + 0.0036389436572790146, + 0.008096146397292614, + 0.005592697765678167, + 0.005328124389052391, + 0.004055806435644627, + 0.007831133902072906, + 0.006046591326594353, + 0.006059179548174143, + 0.005336371716111898, + 0.007113021798431873, + 0.004494186956435442, + 0.004890205804258585, + 0.0038572463672608137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.02058405987918377 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638513274490833 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638513274490833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638513274490833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14809168875217438, + 0.051396407186985016, + 0.2542259991168976, + 0.12451310455799103, + 0.08416609466075897, + 0.06399059295654297, + 0.050151657313108444, + 0.16340228915214539, + 0.038402725011110306, + 0.021659482270479202 + ], + "credit_map": [ + 0.010824397206306458, + 0.010718340054154396, + 0.012098558247089386, + 0.010304716415703297, + 0.011592126451432705, + 0.036708250641822815, + 0.022526351734995842, + 0.03331894427537918, + 0.006236819084733725, + 0.00804105307906866, + 0.004236867185682058, + 0.004582803696393967, + 0.006895274855196476, + 0.008114243857562542, + 0.006977381184697151, + 0.006311957258731127, + 0.009106230922043324, + 0.008538633584976196, + 0.00894330907613039, + 0.013933725655078888, + 0.013408594764769077, + 0.09966609627008438, + 0.008419379591941833, + 0.09221000224351883, + 0.007661453448235989, + 0.00892767496407032, + 0.010236679576337337, + 0.00782779324799776, + 0.016308721154928207, + 0.03188829496502876, + 0.010085667483508587, + 0.03157681226730347, + 0.0074093337170779705, + 0.008865493349730968, + 0.011584148742258549, + 0.009637588635087013, + 0.010987632907927036, + 0.016640515998005867, + 0.00731929624453187, + 0.0117220813408494, + 0.006290431134402752, + 0.004989807028323412, + 0.013623501174151897, + 0.006658039055764675, + 0.009605352766811848, + 0.0076062423177063465, + 0.0065728649497032166, + 0.008644349873065948, + 0.00484974030405283, + 0.006267131306231022, + 0.008166026324033737, + 0.00522485701367259, + 0.006894068792462349, + 0.008105922490358353, + 0.005475449841469526, + 0.005168458912521601, + 0.028989171609282494, + 0.019110379740595818, + 0.02794700860977173, + 0.01526195090264082, + 0.014349455945193768, + 0.019539987668395042, + 0.016593104228377342, + 0.021611230447888374, + 0.007567315362393856, + 0.0025464415084570646, + 0.008140197955071926, + 0.0024784149136394262, + 0.005147367250174284, + 0.004989726934581995, + 0.0031531285494565964, + 0.0043801297433674335, + 0.0034100196789950132, + 0.003328772261738777, + 0.0027982015162706375, + 0.00209847092628479, + 0.0031306298915296793, + 0.002438148483633995, + 0.0023945490829646587, + 0.0020606901962310076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.008638513274490833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623270619660616 + ], + "opposite_logprob_tokens": [ + -4.882623195648193 + ], + "logprob_reward": -0.007623270619660616, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623270619660616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6317439675331116, + 0.042472030967473984, + 0.0341709703207016, + 0.03807127848267555, + 0.05371439829468727, + 0.0790441632270813, + 0.02324691228568554, + 0.033006452023983, + 0.034936822950839996, + 0.029593057930469513 + ], + "credit_map": [ + 0.026063216850161552, + 0.027712352573871613, + 0.04680941626429558, + 0.0239223912358284, + 0.03753781318664551, + 0.23565393686294556, + 0.044968150556087494, + 0.18907663226127625, + 0.005914932582527399, + 0.005833627190440893, + 0.0039035012014210224, + 0.0036350951995700598, + 0.004091027192771435, + 0.007666941732168198, + 0.004838484805077314, + 0.00658841710537672, + 0.004128367640078068, + 0.0034115032758563757, + 0.0036755951587110758, + 0.002778960857540369, + 0.0037726876325905323, + 0.006850377190858126, + 0.0034385232720524073, + 0.006114957854151726, + 0.003943612799048424, + 0.005067325197160244, + 0.004586625378578901, + 0.003486247966066003, + 0.005875035189092159, + 0.006435164716094732, + 0.0027386939618736506, + 0.005938570946455002, + 0.004963401705026627, + 0.004656330682337284, + 0.009136499837040901, + 0.005485872272402048, + 0.00475767906755209, + 0.012038789689540863, + 0.004994265269488096, + 0.007681559771299362, + 0.013864146545529366, + 0.0066688284277915955, + 0.015616782009601593, + 0.008505118079483509, + 0.006263516377657652, + 0.009173650294542313, + 0.008479870855808258, + 0.010472245514392853, + 0.0018665145616978407, + 0.002436460927128792, + 0.0038539182860404253, + 0.002988268854096532, + 0.002998371608555317, + 0.003798489924520254, + 0.002560882829129696, + 0.002744004363194108, + 0.00590761611238122, + 0.003642193740233779, + 0.0056130606681108475, + 0.003546958090737462, + 0.002980879507958889, + 0.003939607180655003, + 0.003379952162504196, + 0.0039961859583854675, + 0.006442087236791849, + 0.0028262250125408173, + 0.007771054748445749, + 0.003594309790059924, + 0.002642790088430047, + 0.004601988475769758, + 0.003079519374296069, + 0.003978848922997713, + 0.005354486405849457, + 0.005108693148940802, + 0.0041738650761544704, + 0.003043233882635832, + 0.0032312136609107256, + 0.0033365015406161547, + 0.002537025371566415, + 0.0028080404736101627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.007623270619660616 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a gentle picture-book style, a close-up, static shot, shot from a slightly upward angle, focuses on a wondrous creature composed entirely of living vines and green leaves. Its flexible, emerald green vines intertwine to form its body and arms, punctuated by heart-shaped, distinctly veined leaves. Its face is a blooming white flower, similar to a sunflower but with a softer hue. Its eyes are gleaming with innocent black stamens at its center. In its hands, formed from delicate vines and tender leaves, it carefully holds and shapes a tiny, glowing seed. Its vine-like fingertips, with a loving, focused touch, slowly and steadily stroke the seed's shell. The background is a watercolor-blurred, softly illuminated forest. Suddenly, as a tiny green sprout gently peeked out from the top of the seed, a warm and magical golden halo instantly pulsated and emanated from it. This light not only made the seed itself brilliant, but also softly illuminated its slightly forward, curious flower face upwards, casting warm light and shadows on the petals, and letting its black stamens reflect little starlight. The whole picture was full of tranquility, miracle and the hope of the new life.", + "seed": 32715, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_15/index_1172_seed_32715_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.028766934480518102, + "vlm_logprob": -0.028766934480518102, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.028766934480518102, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10182183980941772 + ], + "opposite_logprob_tokens": [ + -2.3518218994140625 + ], + "logprob_reward": -0.10182183980941772, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10182183980941772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2256423830986023, + 0.07056956738233566, + 0.04838693514466286, + 0.11507347971200943, + 0.0506589412689209, + 0.0838758572936058, + 0.1930726021528244, + 0.036911629140377045, + 0.134779691696167, + 0.04102897644042969 + ], + "credit_map": [ + 0.03377237170934677, + 0.02125079371035099, + 0.03260704502463341, + 0.0340641587972641, + 0.021094582974910736, + 0.022656211629509926, + 0.028919322416186333, + 0.031277868896722794, + 0.009431237354874611, + 0.008836964145302773, + 0.00746495695784688, + 0.0077063050121068954, + 0.011360475793480873, + 0.010069197975099087, + 0.008042735978960991, + 0.0076576899737119675, + 0.005637882277369499, + 0.00536405760794878, + 0.007516049314290285, + 0.0037648051511496305, + 0.0068418774753808975, + 0.004407179541885853, + 0.004367996007204056, + 0.010487085208296776, + 0.01713525503873825, + 0.009928839281201363, + 0.020613692700862885, + 0.014798649586737156, + 0.009898592717945576, + 0.010328912176191807, + 0.012636924162507057, + 0.019732603803277016, + 0.00542967114597559, + 0.0036386323627084494, + 0.012693786062300205, + 0.007786770351231098, + 0.004176411312073469, + 0.004681859631091356, + 0.004570298828184605, + 0.007681508082896471, + 0.009278473444283009, + 0.009448809549212456, + 0.010911958292126656, + 0.013472843915224075, + 0.011413182131946087, + 0.00876373890787363, + 0.010942958295345306, + 0.009643886238336563, + 0.017658714205026627, + 0.009722081013023853, + 0.06692111492156982, + 0.010340220294892788, + 0.03312186896800995, + 0.02114122360944748, + 0.0076715294271707535, + 0.026495851576328278, + 0.005150068085640669, + 0.0033877617679536343, + 0.007093511521816254, + 0.003237325930967927, + 0.002263220027089119, + 0.0035684723407030106, + 0.004526506178081036, + 0.007684761192649603, + 0.013008205220103264, + 0.010845905169844627, + 0.014103707857429981, + 0.03944087028503418, + 0.02141755260527134, + 0.013260210864245892, + 0.012969783507287502, + 0.009733438491821289, + 0.005172691773623228, + 0.003758701728656888, + 0.008407254703342915, + 0.0036197470035403967, + 0.005386874545365572, + 0.005856786388903856, + 0.00442073168233037, + 0.004406183026731014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.10182183980941772 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016090303659439087 + ], + "opposite_logprob_tokens": [ + -4.141090393066406 + ], + "logprob_reward": -0.016090303659439087, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016090303659439087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08193179219961166, + 0.23320049047470093, + 0.08282312750816345, + 0.058435071259737015, + 0.04246767610311508, + 0.08055199682712555, + 0.20614413917064667, + 0.1548730880022049, + 0.032138049602508545, + 0.027434581890702248 + ], + "credit_map": [ + 0.008809841237962246, + 0.009248631075024605, + 0.011988935992121696, + 0.009897836484014988, + 0.0068209352903068066, + 0.012039512395858765, + 0.011649740859866142, + 0.011476351879537106, + 0.02142352983355522, + 0.03668372705578804, + 0.025007067248225212, + 0.02660978212952614, + 0.045935627073049545, + 0.03339119255542755, + 0.021408285945653915, + 0.022741246968507767, + 0.011421995237469673, + 0.011792977340519428, + 0.010270663537085056, + 0.005265816580504179, + 0.018010003492236137, + 0.011879192665219307, + 0.0065413303673267365, + 0.007641139440238476, + 0.006140169687569141, + 0.00671357661485672, + 0.007354999892413616, + 0.0081515247002244, + 0.011686335317790508, + 0.0035372276324778795, + 0.0060071758925914764, + 0.008844058960676193, + 0.003984312992542982, + 0.0048128897324204445, + 0.006227539386600256, + 0.004790762439370155, + 0.006458544172346592, + 0.004019229207187891, + 0.0055977050215005875, + 0.006576689891517162, + 0.007773219607770443, + 0.007173639722168446, + 0.015133248642086983, + 0.007835354655981064, + 0.007148837670683861, + 0.013325992040336132, + 0.007720074150711298, + 0.01444162055850029, + 0.018864400684833527, + 0.010038885287940502, + 0.07447700202465057, + 0.012254675850272179, + 0.041861798614263535, + 0.012323863804340363, + 0.006988415494561195, + 0.02933509461581707, + 0.023482948541641235, + 0.01778809167444706, + 0.017700711265206337, + 0.010409179143607616, + 0.03528542071580887, + 0.01764788292348385, + 0.01715625450015068, + 0.015402600169181824, + 0.004616055637598038, + 0.003596321679651737, + 0.0052376920357346535, + 0.004339441657066345, + 0.0028714630752801895, + 0.004815936088562012, + 0.0035715089179575443, + 0.0030896274838596582, + 0.003250967711210251, + 0.0020342518109828234, + 0.008900105021893978, + 0.0018751416355371475, + 0.003564844373613596, + 0.003256400814279914, + 0.002180295530706644, + 0.0023725724313408136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.016090303659439087 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420364435762167 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.01420364435762167, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420364435762167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1560419350862503, + 0.05306236818432808, + 0.07289589941501617, + 0.15083816647529602, + 0.058498531579971313, + 0.2104012370109558, + 0.16452725231647491, + 0.056478843092918396, + 0.052001576870679855, + 0.02525411546230316 + ], + "credit_map": [ + 0.023523425683379173, + 0.012963450513780117, + 0.02487231232225895, + 0.01820771023631096, + 0.013472898863255978, + 0.022579729557037354, + 0.020115699619054794, + 0.02030670829117298, + 0.007231493014842272, + 0.008081055246293545, + 0.006120601668953896, + 0.0066533880308270454, + 0.007799001410603523, + 0.007477264851331711, + 0.004698093980550766, + 0.005001469515264034, + 0.006235789507627487, + 0.00776576716452837, + 0.013804985210299492, + 0.007694570813328028, + 0.01020418293774128, + 0.007024374324828386, + 0.007252716924995184, + 0.012913511134684086, + 0.017094163224101067, + 0.008799929171800613, + 0.024035373702645302, + 0.03677653521299362, + 0.009208525530993938, + 0.0082175862044096, + 0.021089397370815277, + 0.025616656988859177, + 0.0059201037511229515, + 0.004759584553539753, + 0.013198547065258026, + 0.008350061252713203, + 0.008247055113315582, + 0.0046213548630476, + 0.005852293223142624, + 0.007549526169896126, + 0.020329579710960388, + 0.024179581552743912, + 0.03460876643657684, + 0.020239530131220818, + 0.02780311554670334, + 0.030394621193408966, + 0.024148546159267426, + 0.028697501868009567, + 0.014485800638794899, + 0.009344005957245827, + 0.05831320583820343, + 0.011929240077733994, + 0.03272051736712456, + 0.01005446445196867, + 0.0065007684752345085, + 0.02117924764752388, + 0.008357098326086998, + 0.0069046891294419765, + 0.0082187345251441, + 0.004237933550029993, + 0.007671063765883446, + 0.007096449378877878, + 0.007137379143387079, + 0.006855495274066925, + 0.004844474606215954, + 0.0032937845680862665, + 0.017365949228405952, + 0.012496359646320343, + 0.0020177187398076057, + 0.003479711012914777, + 0.0027908894699066877, + 0.005712689831852913, + 0.0039040232077240944, + 0.00255299499258399, + 0.0059546069242060184, + 0.002018173225224018, + 0.0017429727595299482, + 0.003650742582976818, + 0.0021206520032137632, + 0.0033099509309977293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.01420364435762167 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093957599252462 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004093957599252462, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093957599252462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1527319699525833, + 0.05935480445623398, + 0.04808427765965462, + 0.05662509799003601, + 0.02277989313006401, + 0.36920836567878723, + 0.10033048689365387, + 0.1201971024274826, + 0.04671722650527954, + 0.023970773443579674 + ], + "credit_map": [ + 0.01629498600959778, + 0.011891859583556652, + 0.02438175119459629, + 0.016962602734565735, + 0.014656439423561096, + 0.022209258750081062, + 0.026105187833309174, + 0.02022988349199295, + 0.006748594343662262, + 0.00977702159434557, + 0.007527695968747139, + 0.004801524803042412, + 0.011426456272602081, + 0.009062561206519604, + 0.0043232678435742855, + 0.005687682889401913, + 0.008829265832901001, + 0.0033836783841252327, + 0.0075435335747897625, + 0.002752169966697693, + 0.005254135467112064, + 0.003945212345570326, + 0.0035163869615644217, + 0.012859894894063473, + 0.008217515423893929, + 0.004571220371872187, + 0.008475321345031261, + 0.010857555083930492, + 0.004221417475491762, + 0.004460511263459921, + 0.0068398392759263515, + 0.008981718681752682, + 0.0022318519186228514, + 0.0023359302431344986, + 0.004846919793635607, + 0.0017910711467266083, + 0.004935644567012787, + 0.00245985621586442, + 0.0016216468065977097, + 0.0025569729041308165, + 0.03719378262758255, + 0.04588402435183525, + 0.06094459444284439, + 0.03219074755907059, + 0.05410746857523918, + 0.05002225562930107, + 0.040869638323783875, + 0.04799587279558182, + 0.008770111948251724, + 0.0062578157521784306, + 0.033332426100969315, + 0.0067670634016394615, + 0.02122054249048233, + 0.007107892073690891, + 0.004043271765112877, + 0.012831362895667553, + 0.015424210578203201, + 0.013821891508996487, + 0.017337338998913765, + 0.008656947873532772, + 0.015824101865291595, + 0.014030465856194496, + 0.013832071796059608, + 0.021270068362355232, + 0.00560740428045392, + 0.0032140498515218496, + 0.014964461326599121, + 0.003960455767810345, + 0.0034548116382211447, + 0.0043610562570393085, + 0.003154908074066043, + 0.008000076748430729, + 0.003603150136768818, + 0.0023411435540765524, + 0.007343820296227932, + 0.001564630540087819, + 0.001645560609176755, + 0.003322659293189645, + 0.0016352699603885412, + 0.0025145383551716805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.004093957599252462 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076249269768595695 + ], + "opposite_logprob_tokens": [ + -4.882625102996826 + ], + "logprob_reward": -0.0076249269768595695, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076249269768595695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18976692855358124, + 0.11547936499118805, + 0.09148840606212616, + 0.0485757440328598, + 0.07173528522253036, + 0.0903417319059372, + 0.17024360597133636, + 0.0454590767621994, + 0.11705584079027176, + 0.059854015707969666 + ], + "credit_map": [ + 0.024625055491924286, + 0.013499966822564602, + 0.02809940092265606, + 0.020281337201595306, + 0.021454188972711563, + 0.038474392145872116, + 0.01856861636042595, + 0.024763990193605423, + 0.015256719663739204, + 0.013374792411923409, + 0.012996900826692581, + 0.01445610262453556, + 0.016116714105010033, + 0.016269341111183167, + 0.014486508443951607, + 0.012522285804152489, + 0.01638001948595047, + 0.005901026539504528, + 0.017459748312830925, + 0.007911872118711472, + 0.0070776441134512424, + 0.014500112272799015, + 0.008607814088463783, + 0.013650165870785713, + 0.00572796817868948, + 0.00509636988863349, + 0.009420277550816536, + 0.007298056036233902, + 0.00465445825830102, + 0.004447305575013161, + 0.006255434826016426, + 0.00567587511613965, + 0.008140536025166512, + 0.007312797009944916, + 0.013407478109002113, + 0.009422271512448788, + 0.008002757094800472, + 0.008520947769284248, + 0.0074175577610731125, + 0.009510939009487629, + 0.00868808850646019, + 0.0072539071552455425, + 0.013464945368468761, + 0.013884851709008217, + 0.008839728310704231, + 0.012396601028740406, + 0.013393484987318516, + 0.012420124374330044, + 0.016407068818807602, + 0.007636859081685543, + 0.05841915309429169, + 0.009699616581201553, + 0.02690805308520794, + 0.019809026271104813, + 0.0056578051298856735, + 0.02570601925253868, + 0.006374003365635872, + 0.004234994761645794, + 0.006683200132101774, + 0.0030022403225302696, + 0.009028041735291481, + 0.006211458705365658, + 0.005197893362492323, + 0.004727245774120092, + 0.01643846370279789, + 0.013434631749987602, + 0.018388068303465843, + 0.01984982192516327, + 0.009780723601579666, + 0.013831141404807568, + 0.01621088758111, + 0.009122099727392197, + 0.0077683827839791775, + 0.004982353653758764, + 0.014750506728887558, + 0.004307112656533718, + 0.007970189675688744, + 0.007915765047073364, + 0.006757007446140051, + 0.005402699112892151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.0076249269768595695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a gentle picture-book style, a close-up, static shot, shot from a slightly upward angle, focuses on a wondrous creature composed entirely of living vines and green leaves. Its flexible, emerald green vines intertwine to form its body and arms, punctuated by heart-shaped, distinctly veined leaves. Its face is a blooming white flower, similar to a sunflower but with a softer hue. Its eyes are gleaming with innocent black stamens at its center. In its hands, formed from delicate vines and tender leaves, it carefully holds and shapes a tiny, glowing seed. Its vine-like fingertips, with a loving, focused touch, slowly and steadily stroke the seed's shell. The background is a watercolor-blurred, softly illuminated forest. Suddenly, as a tiny green sprout gently peeked out from the top of the seed, a warm and magical golden halo instantly pulsated and emanated from it. This light not only made the seed itself brilliant, but also softly illuminated its slightly forward, curious flower face upwards, casting warm light and shadows on the petals, and letting its black stamens reflect little starlight. The whole picture was full of tranquility, miracle and the hope of the new life.", + "seed": 32716, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_15/index_1172_seed_32716_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.021261506713926792, + "vlm_logprob": -0.021261506713926792, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.021261506713926792, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.050611112266778946 + ], + "opposite_logprob_tokens": [ + -3.0506110191345215 + ], + "logprob_reward": -0.050611112266778946, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.050611112266778946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20184937119483948, + 0.03450765833258629, + 0.03711463138461113, + 0.04877432435750961, + 0.2829744517803192, + 0.033468738198280334, + 0.0505879670381546, + 0.14372852444648743, + 0.03499233350157738, + 0.13200199604034424 + ], + "credit_map": [ + 0.0287002045661211, + 0.019752521067857742, + 0.025889405980706215, + 0.02083936147391796, + 0.032576851546764374, + 0.036534614861011505, + 0.01857568882405758, + 0.01898069679737091, + 0.0037383795715868473, + 0.003914469853043556, + 0.004232315346598625, + 0.003579795593395829, + 0.005978437606245279, + 0.005551538430154324, + 0.003680373542010784, + 0.0038323451299220324, + 0.006551527883857489, + 0.0027276556938886642, + 0.00463606882840395, + 0.00432391045615077, + 0.00393152330070734, + 0.005219749640673399, + 0.005612576380372047, + 0.0041116150096058846, + 0.004926077090203762, + 0.0037880444433540106, + 0.008926267735660076, + 0.0031950867269188166, + 0.003703590016812086, + 0.008455399423837662, + 0.0031223976984620094, + 0.012657455168664455, + 0.0941464975476265, + 0.013314216397702694, + 0.014595266431570053, + 0.010125821456313133, + 0.011437970213592052, + 0.09816216677427292, + 0.009739240631461143, + 0.031453266739845276, + 0.00685231015086174, + 0.0026827258989214897, + 0.006615324877202511, + 0.003081271657720208, + 0.003029405837878585, + 0.005210237577557564, + 0.002103248843923211, + 0.0038942061364650726, + 0.016087891533970833, + 0.003744589164853096, + 0.006982036866247654, + 0.0036516543477773666, + 0.0028716903179883957, + 0.008064818568527699, + 0.0022688163444399834, + 0.006916465703397989, + 0.018385112285614014, + 0.028094308450818062, + 0.013297619298100471, + 0.007851220667362213, + 0.015266446396708488, + 0.041943419724702835, + 0.005944293458014727, + 0.012946093454957008, + 0.004457483999431133, + 0.004457646980881691, + 0.0050466638058424, + 0.003373570740222931, + 0.004507084842771292, + 0.004945328459143639, + 0.0044950335286557674, + 0.00370951765216887, + 0.017337078228592873, + 0.031840939074754715, + 0.007997692562639713, + 0.00544147239997983, + 0.012143638916313648, + 0.041557349264621735, + 0.006348481867462397, + 0.009335329756140709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.050611112266778946 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016087019816040993 + ], + "opposite_logprob_tokens": [ + -4.141087055206299 + ], + "logprob_reward": -0.016087019816040993, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016087019816040993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12598149478435516, + 0.06991151720285416, + 0.07517456263303757, + 0.09973572194576263, + 0.35035791993141174, + 0.0622236505150795, + 0.032785139977931976, + 0.03625163808465004, + 0.03567145764827728, + 0.11190692335367203 + ], + "credit_map": [ + 0.012774148024618626, + 0.014932272955775261, + 0.014887661673128605, + 0.012943661771714687, + 0.027871420606970787, + 0.015843166038393974, + 0.014186417683959007, + 0.012542727403342724, + 0.008395268581807613, + 0.010131572373211384, + 0.006404933985322714, + 0.008082178421318531, + 0.013505185954272747, + 0.007719559129327536, + 0.00913719367235899, + 0.006535619962960482, + 0.01046751718968153, + 0.011018321849405766, + 0.005483587272465229, + 0.007132281549274921, + 0.014163614250719547, + 0.009924610145390034, + 0.008602536283433437, + 0.008382085710763931, + 0.00883560348302126, + 0.013214142061769962, + 0.01852279342710972, + 0.006470974069088697, + 0.01628495752811432, + 0.009449743665754795, + 0.005746078677475452, + 0.021211422979831696, + 0.109960176050663, + 0.01804502122104168, + 0.014618556946516037, + 0.011514616198837757, + 0.01222573034465313, + 0.12700828909873962, + 0.00934680737555027, + 0.047638699412345886, + 0.010088438168168068, + 0.005784255918115377, + 0.010294433683156967, + 0.006045342888683081, + 0.008206913247704506, + 0.008046053349971771, + 0.006190813146531582, + 0.0075673991814255714, + 0.007408203557133675, + 0.0032639235723763704, + 0.0045143552124500275, + 0.0032230110373347998, + 0.004371623508632183, + 0.003054831875488162, + 0.0025063459761440754, + 0.00444284500554204, + 0.004117751959711313, + 0.007095845881849527, + 0.0035568540915846825, + 0.0016689246986061335, + 0.004255283158272505, + 0.01027266588062048, + 0.0016821840545162559, + 0.003602120792493224, + 0.004114313516765833, + 0.004811993334442377, + 0.0037305860314518213, + 0.0038594226352870464, + 0.005276392214000225, + 0.004876631312072277, + 0.005449519958347082, + 0.003552597714588046, + 0.013823338784277439, + 0.025141602382063866, + 0.006948310881853104, + 0.004424870014190674, + 0.013945449143648148, + 0.034236013889312744, + 0.005302935838699341, + 0.008084386587142944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.016087019816040993 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107910368591547 + ], + "opposite_logprob_tokens": [ + -4.51107931137085 + ], + "logprob_reward": -0.01107910368591547, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107910368591547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018235309049487114, + 0.009418483823537827, + 0.009035334922373295, + 0.016503194347023964, + 0.907055139541626, + 0.005174100864678621, + 0.004015958867967129, + 0.007925841957330704, + 0.006117194890975952, + 0.016519442200660706 + ], + "credit_map": [ + 0.0023531068582087755, + 0.0013625809224322438, + 0.0024946669582277536, + 0.002465699100866914, + 0.0019107875414192677, + 0.0030558877624571323, + 0.0024536417331546545, + 0.002138938056305051, + 0.0010067288530990481, + 0.0011764033697545528, + 0.0012193896109238267, + 0.0009867111220955849, + 0.001666621770709753, + 0.0013740966096520424, + 0.0009415271342732012, + 0.0010470055276528, + 0.001164899324066937, + 0.0009414658416062593, + 0.0015495042316615582, + 0.0010635539656504989, + 0.0011369743151590228, + 0.0011550355702638626, + 0.0010589496232569218, + 0.0009649523417465389, + 0.0014111216878518462, + 0.0023564454168081284, + 0.0025854967534542084, + 0.0013684681616723537, + 0.002863209694623947, + 0.00172996218316257, + 0.0009622460347600281, + 0.003226244356483221, + 0.31718769669532776, + 0.0357031486928463, + 0.03501545637845993, + 0.02298988774418831, + 0.02597668394446373, + 0.339254766702652, + 0.023312900215387344, + 0.10761459171772003, + 0.0009876972762867808, + 0.00036353827454149723, + 0.0011572421062737703, + 0.000436067784903571, + 0.0003878084826283157, + 0.000773690000642091, + 0.0003468806098680943, + 0.000721176213119179, + 0.000687737949192524, + 0.00034862689790315926, + 0.0007383950287476182, + 0.0004207580059301108, + 0.0003643712552729994, + 0.0005508429603651166, + 0.00036768431891687214, + 0.0005375424516387284, + 0.000990251312032342, + 0.0012186700478196144, + 0.0010230221087113023, + 0.0006292849429883063, + 0.0008741238852962852, + 0.001867766142822802, + 0.0005266212974675, + 0.0007961025112308562, + 0.0007434230064973235, + 0.0007608869345858693, + 0.0008046464063227177, + 0.000493067258503288, + 0.000949796405620873, + 0.0009707590215839446, + 0.0005786276306025684, + 0.0008159883436746895, + 0.002354641444981098, + 0.003530031768605113, + 0.0012028093915432692, + 0.0006718421936966479, + 0.0017502430127933621, + 0.004872728604823351, + 0.0006847511976957321, + 0.0014523927820846438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.01107910368591547 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525154173374176 + ], + "opposite_logprob_tokens": [ + -5.255251407623291 + ], + "logprob_reward": -0.00525154173374176, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525154173374176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038494620472192764, + 0.020111078396439552, + 0.015813060104846954, + 0.033305972814559937, + 0.7195110321044922, + 0.013373066671192646, + 0.014880194328725338, + 0.016162412241101265, + 0.01871892437338829, + 0.10962970554828644 + ], + "credit_map": [ + 0.0040711648762226105, + 0.004482917487621307, + 0.004198858048766851, + 0.004022366367280483, + 0.006875297520309687, + 0.005290383007377386, + 0.004953878931701183, + 0.004599751904606819, + 0.00200342177413404, + 0.0027255674358457327, + 0.002416344126686454, + 0.003670643549412489, + 0.002004806650802493, + 0.00234368909150362, + 0.002817995846271515, + 0.0021286068949848413, + 0.002217549132183194, + 0.0014092623023316264, + 0.002121588448062539, + 0.0023665388580411673, + 0.002196941524744034, + 0.0020087433513253927, + 0.0020473692566156387, + 0.0014450644375756383, + 0.0037463712505996227, + 0.005026287864893675, + 0.004447513725608587, + 0.002875884994864464, + 0.0056279003620147705, + 0.003752402262762189, + 0.0021391368936747313, + 0.005690472200512886, + 0.24046951532363892, + 0.03120911680161953, + 0.02791748195886612, + 0.018798816949129105, + 0.023290541023015976, + 0.2669978737831116, + 0.019108884036540985, + 0.09171874076128006, + 0.0028789422940462828, + 0.0011663534678518772, + 0.0019704806618392467, + 0.0012961223255842924, + 0.001347842626273632, + 0.0018389078322798014, + 0.0009027824853546917, + 0.0019716324750334024, + 0.004388143308460712, + 0.001193943084217608, + 0.0022522860672324896, + 0.0009520006715320051, + 0.0009277340723201632, + 0.00202503171749413, + 0.000610193470492959, + 0.0025308621115982533, + 0.0021759613882750273, + 0.0016967372503131628, + 0.0031661801040172577, + 0.0015231029829010367, + 0.002198020229116082, + 0.0018440736457705498, + 0.0013567740097641945, + 0.0022015615832060575, + 0.002684691920876503, + 0.0021946835331618786, + 0.002612051786854863, + 0.0016126304399222136, + 0.0023203608579933643, + 0.002784839365631342, + 0.0020744414068758488, + 0.0024352234322577715, + 0.013747994787991047, + 0.026716360822319984, + 0.006831307429820299, + 0.004343480337411165, + 0.01116236299276352, + 0.03470796346664429, + 0.0041719465516507626, + 0.007948275655508041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.00525154173374176 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327875606715679 + ], + "opposite_logprob_tokens": [ + -3.7732787132263184 + ], + "logprob_reward": -0.02327875606715679, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327875606715679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07980510592460632, + 0.042264390736818314, + 0.01426113024353981, + 0.03307317942380905, + 0.3949643075466156, + 0.024304626509547234, + 0.05961082503199577, + 0.06079902499914169, + 0.054920487105846405, + 0.23599688708782196 + ], + "credit_map": [ + 0.014731667004525661, + 0.006103286519646645, + 0.012856570072472095, + 0.008914744481444359, + 0.009954958222806454, + 0.010599290020763874, + 0.008122613653540611, + 0.008521975949406624, + 0.006449424661695957, + 0.003365821437910199, + 0.006744038313627243, + 0.004868927877396345, + 0.0045578316785395145, + 0.0062613170593976974, + 0.0057236431166529655, + 0.0042933872900903225, + 0.0019079837948083878, + 0.001283339224755764, + 0.0018772798357531428, + 0.00198082672432065, + 0.0018365263240411878, + 0.0019086090615019202, + 0.001966618001461029, + 0.0014999466948211193, + 0.0045089093036949635, + 0.0023878957144916058, + 0.005076196976006031, + 0.002567353192716837, + 0.002427033381536603, + 0.0064752851612865925, + 0.002298208652064204, + 0.007332297042012215, + 0.14931906759738922, + 0.01208505965769291, + 0.015377691946923733, + 0.00961392093449831, + 0.007973290048539639, + 0.14999616146087646, + 0.010025199502706528, + 0.04057390242815018, + 0.005649057682603598, + 0.0018868917832151055, + 0.004153937101364136, + 0.0022481053601950407, + 0.0019295931560918689, + 0.0036256148014217615, + 0.0016856928123161197, + 0.003125734394416213, + 0.012990006245672703, + 0.004367504734545946, + 0.010721439495682716, + 0.005209969822317362, + 0.005069590173661709, + 0.006962071638554335, + 0.004258399363607168, + 0.010031845420598984, + 0.008775215595960617, + 0.010057690553367138, + 0.005503565538674593, + 0.0032387380488216877, + 0.006627731956541538, + 0.017678119242191315, + 0.0033253959845751524, + 0.0055925664491951466, + 0.008317837491631508, + 0.006356109399348497, + 0.008140967227518559, + 0.004604593850672245, + 0.007183436769992113, + 0.007604946382343769, + 0.004638893995434046, + 0.008073707111179829, + 0.029111512005329132, + 0.05872957408428192, + 0.014407340437173843, + 0.00831629429012537, + 0.021717874333262444, + 0.07634100317955017, + 0.00934274960309267, + 0.018030546605587006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.02327875606715679 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a gentle picture-book style, a close-up, static shot, shot from a slightly upward angle, focuses on a wondrous creature composed entirely of living vines and green leaves. Its flexible, emerald green vines intertwine to form its body and arms, punctuated by heart-shaped, distinctly veined leaves. Its face is a blooming white flower, similar to a sunflower but with a softer hue. Its eyes are gleaming with innocent black stamens at its center. In its hands, formed from delicate vines and tender leaves, it carefully holds and shapes a tiny, glowing seed. Its vine-like fingertips, with a loving, focused touch, slowly and steadily stroke the seed's shell. The background is a watercolor-blurred, softly illuminated forest. Suddenly, as a tiny green sprout gently peeked out from the top of the seed, a warm and magical golden halo instantly pulsated and emanated from it. This light not only made the seed itself brilliant, but also softly illuminated its slightly forward, curious flower face upwards, casting warm light and shadows on the petals, and letting its black stamens reflect little starlight. The whole picture was full of tranquility, miracle and the hope of the new life.", + "seed": 32717, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_15/index_1172_seed_32717_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02446899195201695, + "vlm_logprob": -0.02446899195201695, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02446899195201695, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07222555577754974 + ], + "opposite_logprob_tokens": [ + -2.697225570678711 + ], + "logprob_reward": -0.07222555577754974, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07222555577754974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0732814222574234, + 0.029685257002711296, + 0.3843887150287628, + 0.07119236141443253, + 0.10781701654195786, + 0.09156876057386398, + 0.11849561333656311, + 0.0902320146560669, + 0.024833787232637405, + 0.008505096659064293 + ], + "credit_map": [ + 0.00898815132677555, + 0.011605631560087204, + 0.006648161448538303, + 0.009435002692043781, + 0.012331042438745499, + 0.009000623598694801, + 0.008381071500480175, + 0.006891737226396799, + 0.0035648515913635492, + 0.002871386706829071, + 0.0033185638021677732, + 0.0035254908725619316, + 0.0027331681922078133, + 0.00672279205173254, + 0.0029743651393800974, + 0.003974635619670153, + 0.048982005566358566, + 0.008320827968418598, + 0.1554446518421173, + 0.0213759895414114, + 0.011206812225282192, + 0.01755595952272415, + 0.02285507321357727, + 0.09864740073680878, + 0.005831161513924599, + 0.0037099053151905537, + 0.007931827567517757, + 0.003547488246113062, + 0.002858042949810624, + 0.02431557886302471, + 0.004311330616474152, + 0.01868702657520771, + 0.0099976547062397, + 0.01493972260504961, + 0.014001774601638317, + 0.0066047473810613155, + 0.02624659053981304, + 0.011997857131063938, + 0.010440124198794365, + 0.013588545843958855, + 0.010161357000470161, + 0.009754667989909649, + 0.016386641189455986, + 0.00645547965541482, + 0.011175454594194889, + 0.016229689121246338, + 0.00921714585274458, + 0.012188319116830826, + 0.011784104630351067, + 0.016839830204844475, + 0.01565529778599739, + 0.014128264971077442, + 0.019107729196548462, + 0.01390235498547554, + 0.011677379719913006, + 0.015400647185742855, + 0.014022953808307648, + 0.010204065591096878, + 0.014886523596942425, + 0.006515636574476957, + 0.012820051051676273, + 0.012561851181089878, + 0.00623664353042841, + 0.01298428326845169, + 0.003519220743328333, + 0.003671813290566206, + 0.003683550748974085, + 0.0029629888013005257, + 0.0033582448959350586, + 0.0029755604919046164, + 0.0023869210854172707, + 0.0022754862438887358, + 0.0012861138675361872, + 0.0008402653620578349, + 0.0016286203172057867, + 0.0008510556654073298, + 0.0010338654974475503, + 0.00116198998875916, + 0.0006655155448243022, + 0.0010376697173342109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 4 + ], + "selected_reward": -0.07222555577754974 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023314394056797028 + ], + "opposite_logprob_tokens": [ + -3.7733144760131836 + ], + "logprob_reward": -0.023314394056797028, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023314394056797028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14222100377082825, + 0.050961289554834366, + 0.07332611083984375, + 0.07824081927537918, + 0.04197558015584946, + 0.045787278562784195, + 0.09506233781576157, + 0.42159196734428406, + 0.031685639172792435, + 0.01914798654615879 + ], + "credit_map": [ + 0.017209051176905632, + 0.024140024557709694, + 0.011410909704864025, + 0.01563129760324955, + 0.02881200984120369, + 0.014606538228690624, + 0.017799226567149162, + 0.012611938640475273, + 0.003612218890339136, + 0.004409609362483025, + 0.0037323024589568377, + 0.005915607791393995, + 0.003527124645188451, + 0.013866440393030643, + 0.004610199015587568, + 0.011287786066532135, + 0.005733651574701071, + 0.0031967300456017256, + 0.024059562012553215, + 0.004903585650026798, + 0.0024309318978339434, + 0.005697496235370636, + 0.005447154399007559, + 0.02185700833797455, + 0.010194875299930573, + 0.009205324575304985, + 0.009548811241984367, + 0.008510316722095013, + 0.006119143683463335, + 0.01446863729506731, + 0.0100551201030612, + 0.010138600133359432, + 0.0055982028134167194, + 0.004448780789971352, + 0.005897473078221083, + 0.004149163607507944, + 0.0056066191755235195, + 0.0060232169926166534, + 0.00414570327848196, + 0.006106420885771513, + 0.004834179766476154, + 0.0054991100914776325, + 0.006878408137708902, + 0.0034709470346570015, + 0.004722115118056536, + 0.007074577268213034, + 0.005762691143900156, + 0.007545250002294779, + 0.009118307381868362, + 0.013479199260473251, + 0.011738481931388378, + 0.011083772405982018, + 0.017431970685720444, + 0.010393361561000347, + 0.008815605193376541, + 0.013001636601984501, + 0.06863683462142944, + 0.049719974398612976, + 0.0703265517950058, + 0.03129040449857712, + 0.05719872564077377, + 0.06060225889086723, + 0.030266035348176956, + 0.05355118587613106, + 0.004079191479831934, + 0.004397442564368248, + 0.005819747690111399, + 0.0034014680422842503, + 0.0038116967771202326, + 0.0038540796376764774, + 0.003025219077244401, + 0.0032967906445264816, + 0.00416635861620307, + 0.0022394899278879166, + 0.003257188480347395, + 0.0015217303298413754, + 0.0021692339796572924, + 0.002651070011779666, + 0.0015397348906844854, + 0.0016031801933422685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.023314394056797028 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541004456579685 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012541004456579685, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541004456579685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26691117882728577, + 0.03510338068008423, + 0.30918988585472107, + 0.07819865643978119, + 0.04734300822019577, + 0.03007815219461918, + 0.019356559962034225, + 0.1828443706035614, + 0.016467895358800888, + 0.014506842941045761 + ], + "credit_map": [ + 0.02778630703687668, + 0.04294285178184509, + 0.019727690145373344, + 0.030313733965158463, + 0.051526907831430435, + 0.026468953117728233, + 0.04348680004477501, + 0.024657946079969406, + 0.00377708300948143, + 0.004789771046489477, + 0.004032748751342297, + 0.004568065982311964, + 0.005113695748150349, + 0.004654351156204939, + 0.0030460068956017494, + 0.005121663678437471, + 0.0798213854432106, + 0.01614469103515148, + 0.0269420575350523, + 0.018826976418495178, + 0.011476090177893639, + 0.026417136192321777, + 0.025812651962041855, + 0.10374893248081207, + 0.0070990887470543385, + 0.00819981750100851, + 0.013661591336131096, + 0.009729841724038124, + 0.005467692390084267, + 0.006479361094534397, + 0.01624181866645813, + 0.011319455690681934, + 0.004591824021190405, + 0.0074007101356983185, + 0.007008506450802088, + 0.003564900951460004, + 0.00873600784689188, + 0.005830894224345684, + 0.004772008862346411, + 0.0054381596855819225, + 0.0035961205139756203, + 0.0031900692265480757, + 0.005476371385157108, + 0.002304541412740946, + 0.002893079537898302, + 0.004403162747621536, + 0.0041072131134569645, + 0.004107600077986717, + 0.0024942925665527582, + 0.0016275107627734542, + 0.003624429227784276, + 0.0019836549181491137, + 0.0016744217136874795, + 0.003307296894490719, + 0.0017028669826686382, + 0.002942088758572936, + 0.02570095844566822, + 0.021663954481482506, + 0.029467295855283737, + 0.014240344986319542, + 0.02727423794567585, + 0.02462940663099289, + 0.018282001838088036, + 0.021586183458566666, + 0.002190958010032773, + 0.0019542051013559103, + 0.001738267601467669, + 0.0016447702655568719, + 0.0016953571466729045, + 0.002959804842248559, + 0.0020158078987151384, + 0.002268725773319602, + 0.001838002703152597, + 0.0017329893307760358, + 0.002899079117923975, + 0.0008872377220541239, + 0.0015673504676669836, + 0.0026058494113385677, + 0.0010429082904011011, + 0.0019334278767928481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.012541004456579685 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031922117341309786 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.0031922117341309786, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031922117341309786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12365257740020752, + 0.09801290929317474, + 0.5378714203834534, + 0.06939439475536346, + 0.042274974286556244, + 0.028926076367497444, + 0.014036586508154869, + 0.031317088752985, + 0.025858990848064423, + 0.028654998168349266 + ], + "credit_map": [ + 0.012429202906787395, + 0.017857631668448448, + 0.009641403332352638, + 0.015922171995043755, + 0.019846579059958458, + 0.014919379726052284, + 0.020291371271014214, + 0.012744830921292305, + 0.008417651057243347, + 0.00455179950222373, + 0.005616321228444576, + 0.005210474599152803, + 0.004473828710615635, + 0.034961141645908356, + 0.005675709806382656, + 0.02910597063601017, + 0.039112016558647156, + 0.012663602828979492, + 0.1930284947156906, + 0.03294625133275986, + 0.013976133428514004, + 0.03382010757923126, + 0.03357846662402153, + 0.17874635756015778, + 0.006077060475945473, + 0.004420516546815634, + 0.009856927208602428, + 0.007472900673747063, + 0.0028035405557602644, + 0.014186562970280647, + 0.009973229840397835, + 0.014603657647967339, + 0.0042052846401929855, + 0.006173755973577499, + 0.006240317597985268, + 0.003748239018023014, + 0.006369438488036394, + 0.006323465146124363, + 0.004046021495014429, + 0.005168451927602291, + 0.0026992368511855602, + 0.0038240484427660704, + 0.006058389320969582, + 0.0021277323830872774, + 0.0035164367873221636, + 0.0033993737306445837, + 0.0026328254025429487, + 0.004668033216148615, + 0.001459578052163124, + 0.001702278619632125, + 0.0025623016990721226, + 0.0011655838461592793, + 0.001769354217685759, + 0.002197620691731572, + 0.0010916761821135879, + 0.0020881928503513336, + 0.0034571001306176186, + 0.003353861626237631, + 0.005722925532609224, + 0.002397014992311597, + 0.004806989803910255, + 0.00402871472761035, + 0.0026328829117119312, + 0.00491759879514575, + 0.0030138790607452393, + 0.0018215232994407415, + 0.005707358475774527, + 0.0025002029724419117, + 0.002288549905642867, + 0.004767670761793852, + 0.0021304097026586533, + 0.003629396203905344, + 0.004228705540299416, + 0.003156292950734496, + 0.006172969006001949, + 0.0030217175371944904, + 0.00286182202398777, + 0.004060181323438883, + 0.0018334960332140326, + 0.003319814568385482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0031922117341309786 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the creature composed entirely of living vines and green leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a face that is a blooming white flower with black stamens for eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature holding and stroking a tiny glowing seed with its vine-like hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a tiny green sprout emerge from the seed, causing a warm golden halo to pulse and illuminate the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a watercolor-blurred, softly illuminated forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071793735027313 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.011071793735027313, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071793735027313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08728896081447601, + 0.19099517166614532, + 0.12145127356052399, + 0.0893205925822258, + 0.11165472865104675, + 0.11908850073814392, + 0.07888375967741013, + 0.05782105028629303, + 0.10366538912057877, + 0.039830539375543594 + ], + "credit_map": [ + 0.010010835714638233, + 0.012187103740870953, + 0.008429713547229767, + 0.009163515642285347, + 0.013989956118166447, + 0.01197119802236557, + 0.009809466078877449, + 0.011727175675332546, + 0.01573665253818035, + 0.009120726026594639, + 0.012111506424844265, + 0.011565622873604298, + 0.009061584249138832, + 0.06800135970115662, + 0.011980011127889156, + 0.05341772362589836, + 0.03675123304128647, + 0.005673946347087622, + 0.011575968004763126, + 0.008579655550420284, + 0.003702267073094845, + 0.01108215469866991, + 0.009599343873560429, + 0.03448670729994774, + 0.005871863104403019, + 0.0032762924674898386, + 0.00798859167844057, + 0.004662723280489445, + 0.0029478929936885834, + 0.02996494062244892, + 0.00748284999281168, + 0.02712542936205864, + 0.010669820941984653, + 0.014666955918073654, + 0.014794893562793732, + 0.007600345648825169, + 0.02561151050031185, + 0.012469541281461716, + 0.010992072522640228, + 0.014849597588181496, + 0.01764821633696556, + 0.00757310027256608, + 0.025878742337226868, + 0.008345567621290684, + 0.010782362893223763, + 0.018665237352252007, + 0.010190579108893871, + 0.020004702731966972, + 0.008121566846966743, + 0.010100539773702621, + 0.01493754144757986, + 0.007953505031764507, + 0.010287128388881683, + 0.0098715890198946, + 0.007480982691049576, + 0.010130907408893108, + 0.007949257269501686, + 0.006364299450069666, + 0.009036701172590256, + 0.004015778191387653, + 0.007066104095429182, + 0.00805989932268858, + 0.004225731827318668, + 0.01110327709466219, + 0.012024756520986557, + 0.01285333652049303, + 0.01546945795416832, + 0.010941429063677788, + 0.013244499452412128, + 0.01861223392188549, + 0.009913133457303047, + 0.010606537573039532, + 0.007195982616394758, + 0.0036200429312884808, + 0.009212575852870941, + 0.00337447552010417, + 0.004329221788793802, + 0.004883508197963238, + 0.002408856526017189, + 0.004805875476449728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.011071793735027313 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_2/index_tensor([4531, 4531])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_2/index_tensor([4531, 4531])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c946b7b0f7ace62146944c6be1e0ee5f1300952a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_2/index_tensor([4531, 4531])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, at eye level, begins focusing on a woman with long, wavy auburn hair, calmly walking forward along a stone path. The camera immediately begins tracking backward in a smooth, steady motion, maintaining a consistent distance. She wears a flowing cream linen sundress that flutters gently around her knees, her bare feet pressing lightly against the cool stones. A gentle breeze rustles through the surrounding leaves, ruffling strands of her hair and enhancing the sense of tranquil movement. As the camera continues to pull back, the frame widens, revealing a vast, lush botanical garden bathed in the warm, golden light of the afternoon. The stunning landscape is showcased, including manicured green lawns, vibrant flower pots blooming with pink roses and purple lavender, and ancient, mossy trees whose leaves create dappled patterns of light and shadow on the ground. Her graceful, unhurried gait and contented expression, combined with the idyllic surroundings, create an atmosphere of profound calm and quiet contemplation.", + "seed": 32812, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_2/index_4531_seed_32812_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.011059877229854465, + "vlm_logprob": -0.011059877229854465, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.011059877229854465, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012281861156225204 + ], + "opposite_logprob_tokens": [ + -4.762281894683838 + ], + "logprob_reward": -0.012281861156225204, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.012281861156225204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45442166924476624, + 0.05798077955842018, + 0.07081776112318039, + 0.0905345007777214, + 0.08765942603349686, + 0.05862586572766304, + 0.03796123340725899, + 0.06536855548620224, + 0.02857162430882454, + 0.04805860295891762 + ], + "credit_map": [ + 0.0851740911602974, + 0.02823052927851677, + 0.025668520480394363, + 0.020577751100063324, + 0.05659780651330948, + 0.03401095047593117, + 0.15531326830387115, + 0.04884876310825348, + 0.007782017346471548, + 0.00432172603905201, + 0.00454969285055995, + 0.005194999743252993, + 0.007044143043458462, + 0.010700199753046036, + 0.010160187259316444, + 0.008227820508182049, + 0.008890604600310326, + 0.010403330437839031, + 0.0065610273741185665, + 0.006049822550266981, + 0.011054521426558495, + 0.009175535291433334, + 0.010144508443772793, + 0.008538422174751759, + 0.004662845749408007, + 0.010302701033651829, + 0.005489589646458626, + 0.009581616148352623, + 0.021221961826086044, + 0.01566704735159874, + 0.012936396524310112, + 0.010672357864677906, + 0.005147857125848532, + 0.007730172481387854, + 0.006288496311753988, + 0.021402815356850624, + 0.013089860789477825, + 0.012957938946783543, + 0.011078925803303719, + 0.009963368996977806, + 0.004591174889355898, + 0.007793989963829517, + 0.0033193849958479404, + 0.008212807588279247, + 0.012240433134138584, + 0.008729612454771996, + 0.008898520842194557, + 0.0048399497754871845, + 0.003166759619489312, + 0.009074902161955833, + 0.002317892387509346, + 0.003683500224724412, + 0.006665004417300224, + 0.005295678973197937, + 0.0038022922817617655, + 0.003955209627747536, + 0.004667564295232296, + 0.011906963773071766, + 0.004834433551877737, + 0.007658574730157852, + 0.008974375203251839, + 0.008788703940808773, + 0.009683086536824703, + 0.008854865096509457, + 0.0015675927279517055, + 0.002779459347948432, + 0.002249234588816762, + 0.0036677331663668156, + 0.002814434003084898, + 0.0034358371049165726, + 0.007671127561479807, + 0.004386208951473236, + 0.0021698533091694117, + 0.004332454409450293, + 0.0030017420649528503, + 0.0053531876765191555, + 0.005906912963837385, + 0.00824861042201519, + 0.010861571878194809, + 0.008184278383851051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.012281861156225204 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009795336984097958 + ], + "opposite_logprob_tokens": [ + -4.634795188903809 + ], + "logprob_reward": -0.009795336984097958, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009795336984097958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3234175443649292, + 0.05396873131394386, + 0.09684542566537857, + 0.05894196406006813, + 0.11290398985147476, + 0.05470866337418556, + 0.06056058034300804, + 0.06101477891206741, + 0.05498084798455238, + 0.1226574182510376 + ], + "credit_map": [ + 0.04834115877747536, + 0.018358293920755386, + 0.018569275736808777, + 0.013591417111456394, + 0.03955855965614319, + 0.02085903100669384, + 0.10956306755542755, + 0.05457674339413643, + 0.006527315825223923, + 0.0031531122513115406, + 0.002778272144496441, + 0.005045516882091761, + 0.011661588214337826, + 0.0052309040911495686, + 0.012508709914982319, + 0.007063316181302071, + 0.010872733779251575, + 0.009567198343575, + 0.009076046757400036, + 0.014341897331178188, + 0.017557738348841667, + 0.013487489894032478, + 0.010093897581100464, + 0.011848432011902332, + 0.007897983305156231, + 0.0060377768240869045, + 0.00706604216247797, + 0.0065913996659219265, + 0.009748171083629131, + 0.009094251319766045, + 0.00707089900970459, + 0.0054354495368897915, + 0.008000416681170464, + 0.007840084843337536, + 0.010429954156279564, + 0.034695882350206375, + 0.010364219546318054, + 0.017293095588684082, + 0.014705343171954155, + 0.00957499910145998, + 0.0032620569691061974, + 0.0060179270803928375, + 0.0025374351534992456, + 0.007160987704992294, + 0.014770095236599445, + 0.008793172426521778, + 0.00799767579883337, + 0.00416931277140975, + 0.0049210842698812485, + 0.014807187020778656, + 0.0031396187841892242, + 0.007875831797719002, + 0.009668732061982155, + 0.008213473483920097, + 0.005488149356096983, + 0.006446508225053549, + 0.004166312515735626, + 0.009780933149158955, + 0.004226122982800007, + 0.00716156791895628, + 0.01061322446912527, + 0.008671279065310955, + 0.008739756420254707, + 0.007655585650354624, + 0.002582845976576209, + 0.00503306882455945, + 0.004242241848260164, + 0.007447402458637953, + 0.006029828917235136, + 0.007239809259772301, + 0.014188653789460659, + 0.008217000402510166, + 0.008066785521805286, + 0.011364183388650417, + 0.013185934163630009, + 0.013527250848710537, + 0.015277081169188023, + 0.020666679367423058, + 0.023428652435541153, + 0.017140861600637436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.009795336984097958 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538296170532703 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538296170532703, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538296170532703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7652525305747986, + 0.013842524029314518, + 0.021285442635416985, + 0.02099701017141342, + 0.030535034835338593, + 0.02817966416478157, + 0.01728767715394497, + 0.02935667708516121, + 0.03919781371951103, + 0.034065697342157364 + ], + "credit_map": [ + 0.12278950959444046, + 0.040113821625709534, + 0.05191197246313095, + 0.029156966134905815, + 0.09270203113555908, + 0.05121472105383873, + 0.257988840341568, + 0.1193746030330658, + 0.0015865083551034331, + 0.001397420885041356, + 0.0014817978953942657, + 0.0014652209356427193, + 0.002281541470438242, + 0.002504530129954219, + 0.0013165127020329237, + 0.0018089916557073593, + 0.0026933513581752777, + 0.004344608169049025, + 0.002200518501922488, + 0.002109331078827381, + 0.0024053631350398064, + 0.003085485426709056, + 0.002352624898776412, + 0.0020941596012562513, + 0.002821393543854356, + 0.0030997367575764656, + 0.0014825587859377265, + 0.0018688320415094495, + 0.004204542841762304, + 0.003316390560939908, + 0.0026673562824726105, + 0.0015361985424533486, + 0.001938902190886438, + 0.0038061898667365313, + 0.0018150899559259415, + 0.005506407003849745, + 0.0048301019705832005, + 0.004873748403042555, + 0.004438592586666346, + 0.0033260034397244453, + 0.0024152647238224745, + 0.004617528989911079, + 0.0016418925952166319, + 0.004175756126642227, + 0.004847383592277765, + 0.00417314562946558, + 0.003771307412534952, + 0.002537384396418929, + 0.0016296907560899854, + 0.0034923842176795006, + 0.0011165231699123979, + 0.002227253746241331, + 0.0025810967199504375, + 0.002211841056123376, + 0.0021037659607827663, + 0.0019251216435804963, + 0.0019695351365953684, + 0.0041009820997715, + 0.0026893741451203823, + 0.0034836484119296074, + 0.0031576643232256174, + 0.003904207842424512, + 0.0054941922426223755, + 0.004557071253657341, + 0.0020151869393885136, + 0.003534705610945821, + 0.0027555322740226984, + 0.005629367660731077, + 0.004547462332993746, + 0.0045278132893145084, + 0.011044113896787167, + 0.00514362845569849, + 0.001752421841956675, + 0.0027816195506602526, + 0.0029348721727728844, + 0.004139442928135395, + 0.004452376160770655, + 0.005768992472440004, + 0.007371395360678434, + 0.004864574875682592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.012538296170532703 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186835572123528 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.018186835572123528, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186835572123528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6010904312133789, + 0.04639286547899246, + 0.05726529285311699, + 0.018151257187128067, + 0.09956761449575424, + 0.03489718958735466, + 0.05075593292713165, + 0.018207812681794167, + 0.02092185989022255, + 0.052749715745449066 + ], + "credit_map": [ + 0.08155466616153717, + 0.03055243007838726, + 0.05105520784854889, + 0.022368518635630608, + 0.07236064970493317, + 0.03472087159752846, + 0.20782820880413055, + 0.1006498634815216, + 0.005553199909627438, + 0.00361474696546793, + 0.003488146001473069, + 0.0042413221672177315, + 0.008899817243218422, + 0.0073297154158353806, + 0.00721674133092165, + 0.006049177143722773, + 0.006243299227207899, + 0.010282771661877632, + 0.006209770683199167, + 0.006554583087563515, + 0.008391049690544605, + 0.007742258720099926, + 0.006126357242465019, + 0.005715206265449524, + 0.0026239037979394197, + 0.0021595654543489218, + 0.002092509064823389, + 0.0016566847916692495, + 0.002872147597372532, + 0.002918998943641782, + 0.002130360109731555, + 0.001697086961939931, + 0.010340580716729164, + 0.007423557806760073, + 0.009660881012678146, + 0.02915092371404171, + 0.0075178565457463264, + 0.013661328703165054, + 0.012605390511453152, + 0.00920709315687418, + 0.0034353353548794985, + 0.004787234123796225, + 0.0018515404080972075, + 0.004392912611365318, + 0.008375364355742931, + 0.00558848911896348, + 0.004631887190043926, + 0.0018344277050346136, + 0.00503423111513257, + 0.012997402809560299, + 0.002878271508961916, + 0.005233359057456255, + 0.008505178615450859, + 0.006782823242247105, + 0.0038705889601260424, + 0.005454080644994974, + 0.0012223967351019382, + 0.002887298818677664, + 0.0012932743411511183, + 0.002015868667513132, + 0.0029653285164386034, + 0.0026171270292252302, + 0.0027273334562778473, + 0.0024791862815618515, + 0.002278724452480674, + 0.003025737823918462, + 0.0014283821219578385, + 0.0022823100443929434, + 0.0021103585604578257, + 0.002554577775299549, + 0.0037661930546164513, + 0.003475575940683484, + 0.0026589506305754185, + 0.004290941171348095, + 0.003992867656052113, + 0.0071114650927484035, + 0.006644679699093103, + 0.008555983193218708, + 0.010984552092850208, + 0.008510282263159752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.018186835572123528 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024970562662929296 + ], + "opposite_logprob_tokens": [ + -6.00249719619751 + ], + "logprob_reward": -0.0024970562662929296, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024970562662929296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32146432995796204, + 0.06207786872982979, + 0.1419612318277359, + 0.06713777780532837, + 0.040619101375341415, + 0.03673815354704857, + 0.08745874464511871, + 0.09265512973070145, + 0.046266648918390274, + 0.10362108051776886 + ], + "credit_map": [ + 0.05265761911869049, + 0.017989005893468857, + 0.05414114519953728, + 0.010636520572006702, + 0.027457328513264656, + 0.02686186134815216, + 0.07247531414031982, + 0.059245530515909195, + 0.007699658628553152, + 0.0043967426754534245, + 0.003823918988928199, + 0.004939442966133356, + 0.012592865154147148, + 0.010667149908840656, + 0.009227105416357517, + 0.008730985224246979, + 0.01880543678998947, + 0.01840132847428322, + 0.015038634650409222, + 0.016467293724417686, + 0.013041965663433075, + 0.01971200853586197, + 0.014844940043985844, + 0.025649631395936012, + 0.005404382012784481, + 0.005563771817833185, + 0.007837487384676933, + 0.008869913406670094, + 0.012502034194767475, + 0.011013317853212357, + 0.007713676430284977, + 0.008233193308115005, + 0.004868174437433481, + 0.00425341259688139, + 0.003196747973561287, + 0.004779397509992123, + 0.006799327675253153, + 0.005952323786914349, + 0.0049768975004553795, + 0.005792818032205105, + 0.0020832535810768604, + 0.003504999913275242, + 0.001980397617444396, + 0.005497465841472149, + 0.009424785152077675, + 0.00576991681009531, + 0.005054876673966646, + 0.0034224572591483593, + 0.008287422358989716, + 0.02198173478245735, + 0.0050625004805624485, + 0.01030699722468853, + 0.01585247553884983, + 0.01134547684341669, + 0.0072314683347940445, + 0.007390669081360102, + 0.006651438772678375, + 0.017672022804617882, + 0.005244389176368713, + 0.009799351915717125, + 0.013609392568469048, + 0.012737617827951908, + 0.015201876871287823, + 0.011739037930965424, + 0.0030721949879080057, + 0.004534891806542873, + 0.003767341375350952, + 0.004380599595606327, + 0.005784593056887388, + 0.006053134333342314, + 0.012626501731574535, + 0.006047392264008522, + 0.005232964642345905, + 0.008424459956586361, + 0.007984761148691177, + 0.011604520492255688, + 0.015014193952083588, + 0.01647687517106533, + 0.022737417370080948, + 0.01614588499069214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.0024970562662929296 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, at eye level, begins focusing on a woman with long, wavy auburn hair, calmly walking forward along a stone path. The camera immediately begins tracking backward in a smooth, steady motion, maintaining a consistent distance. She wears a flowing cream linen sundress that flutters gently around her knees, her bare feet pressing lightly against the cool stones. A gentle breeze rustles through the surrounding leaves, ruffling strands of her hair and enhancing the sense of tranquil movement. As the camera continues to pull back, the frame widens, revealing a vast, lush botanical garden bathed in the warm, golden light of the afternoon. The stunning landscape is showcased, including manicured green lawns, vibrant flower pots blooming with pink roses and purple lavender, and ancient, mossy trees whose leaves create dappled patterns of light and shadow on the ground. Her graceful, unhurried gait and contented expression, combined with the idyllic surroundings, create an atmosphere of profound calm and quiet contemplation.", + "seed": 32813, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_2/index_4531_seed_32813_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.013834981946274638, + "vlm_logprob": -0.013834981946274638, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.013834981946274638, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017792614176869392 + ], + "opposite_logprob_tokens": [ + -4.142792701721191 + ], + "logprob_reward": -0.017792614176869392, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.017792614176869392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43748030066490173, + 0.06607015430927277, + 0.04498656466603279, + 0.033428825438022614, + 0.04550759121775627, + 0.0608895979821682, + 0.14179475605487823, + 0.12630367279052734, + 0.027704041451215744, + 0.015834461897611618 + ], + "credit_map": [ + 0.025328896939754486, + 0.08236651122570038, + 0.04959741607308388, + 0.06118228659033775, + 0.07900024205446243, + 0.026296989992260933, + 0.0765073299407959, + 0.03720065578818321, + 0.006816832348704338, + 0.005890451371669769, + 0.007517349440604448, + 0.009818077087402344, + 0.010307464748620987, + 0.005777123849838972, + 0.012464473024010658, + 0.007478390820324421, + 0.005700615234673023, + 0.003388345241546631, + 0.004777682013809681, + 0.005453427322208881, + 0.005222452338784933, + 0.006325719878077507, + 0.007015618961304426, + 0.007102705538272858, + 0.0035658928100019693, + 0.0037577238399535418, + 0.005485836416482925, + 0.003953508101403713, + 0.00578019954264164, + 0.004348684102296829, + 0.0026175223756581545, + 0.003919463139027357, + 0.004052978940308094, + 0.004927688278257847, + 0.0037841983139514923, + 0.003260394325479865, + 0.006017766427248716, + 0.012106255628168583, + 0.007521737366914749, + 0.003836575895547867, + 0.004565765615552664, + 0.01827850006520748, + 0.003261350095272064, + 0.008700777776539326, + 0.007363440934568644, + 0.008227979764342308, + 0.005487882532179356, + 0.0050039105117321014, + 0.011141405440866947, + 0.017973413690924644, + 0.007715594954788685, + 0.009486465714871883, + 0.04042595997452736, + 0.021078070625662804, + 0.020832354202866554, + 0.01314152218401432, + 0.015492918901145458, + 0.02250763587653637, + 0.00988868810236454, + 0.01353220734745264, + 0.02328668348491192, + 0.01801551692187786, + 0.013280252926051617, + 0.010299776680767536, + 0.002786797471344471, + 0.002726333448663354, + 0.0034636035561561584, + 0.0022049450781196356, + 0.005811089184135199, + 0.0036983962636440992, + 0.0032671319786459208, + 0.003745745401829481, + 0.0014343177899718285, + 0.0012144725769758224, + 0.002107918495312333, + 0.0014334280276671052, + 0.0026956915389746428, + 0.0015259404899552464, + 0.0025769483763724566, + 0.0028457462321966887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.017792614176869392 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792976081371307 + ], + "opposite_logprob_tokens": [ + -4.634792804718018 + ], + "logprob_reward": -0.009792976081371307, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792976081371307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7117197513580322, + 0.02306556887924671, + 0.029626358300447464, + 0.022758206352591515, + 0.022693542763590813, + 0.06641290336847305, + 0.04116037115454674, + 0.05675680935382843, + 0.015615703538060188, + 0.010190734639763832 + ], + "credit_map": [ + 0.06670375168323517, + 0.1535091996192932, + 0.10387661308050156, + 0.08450835198163986, + 0.1254916787147522, + 0.054218001663684845, + 0.07352705299854279, + 0.04988512024283409, + 0.0024816556833684444, + 0.002025187946856022, + 0.002673317678272724, + 0.0034145682584494352, + 0.0028803416062146425, + 0.0026054754853248596, + 0.003979078959673643, + 0.003005941864103079, + 0.0028441271279007196, + 0.0029032258316874504, + 0.002562764333561063, + 0.004254127852618694, + 0.004905289970338345, + 0.0029039664659649134, + 0.005487656220793724, + 0.003765200264751911, + 0.0020087817683815956, + 0.0027633283752948046, + 0.003025277052074671, + 0.0028708719182759523, + 0.00449663819745183, + 0.0030740888323634863, + 0.0021590329706668854, + 0.0023601865395903587, + 0.0022148939315229654, + 0.00247498182579875, + 0.0022972626611590385, + 0.003103003604337573, + 0.004040603991597891, + 0.002674136310815811, + 0.0036693913862109184, + 0.0022192690521478653, + 0.0056368084624409676, + 0.024073557928204536, + 0.0037461952306330204, + 0.011112038046121597, + 0.004296223632991314, + 0.009747831150889397, + 0.003277380019426346, + 0.004522871691733599, + 0.0034039069432765245, + 0.005692839622497559, + 0.006630662828683853, + 0.00498435040935874, + 0.004843766335397959, + 0.006993940565735102, + 0.003829255932942033, + 0.004781648982316256, + 0.009404192678630352, + 0.009454422630369663, + 0.006778184790164232, + 0.009105902165174484, + 0.004991320427507162, + 0.008310440927743912, + 0.003077112138271332, + 0.005635238252580166, + 0.0014372940640896559, + 0.0012824517907574773, + 0.0018129926174879074, + 0.0015689694555476308, + 0.0024419683031737804, + 0.0024159655440598726, + 0.0025237929075956345, + 0.002132267924025655, + 0.0012057626154273748, + 0.0009033109527081251, + 0.0013352194800972939, + 0.0011888878652825952, + 0.0014817878836765885, + 0.0009342565899714828, + 0.0016049743862822652, + 0.0015365348663181067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.009792976081371307 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183674663305283 + ], + "opposite_logprob_tokens": [ + -4.018183708190918 + ], + "logprob_reward": -0.018183674663305283, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183674663305283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31011784076690674, + 0.09628085047006607, + 0.06463617831468582, + 0.10102029144763947, + 0.07033367455005646, + 0.07958125323057175, + 0.10339067876338959, + 0.04988836869597435, + 0.08573783189058304, + 0.03901299834251404 + ], + "credit_map": [ + 0.027589531615376472, + 0.06780609488487244, + 0.045077331364154816, + 0.03732765465974808, + 0.050807952880859375, + 0.025432759895920753, + 0.03410443663597107, + 0.02197207137942314, + 0.00883859395980835, + 0.006654171738773584, + 0.008890204131603241, + 0.01130190771073103, + 0.0105538135394454, + 0.010598459281027317, + 0.019842976704239845, + 0.019600728526711464, + 0.007571924477815628, + 0.004737896844744682, + 0.007182370405644178, + 0.010196348652243614, + 0.008608285337686539, + 0.00809579249471426, + 0.010325216688215733, + 0.007918342016637325, + 0.010507695376873016, + 0.012837076559662819, + 0.015131546184420586, + 0.009225714020431042, + 0.02108887955546379, + 0.012556998990476131, + 0.009030500426888466, + 0.01064187940210104, + 0.008330799639225006, + 0.0074464427307248116, + 0.00631799316033721, + 0.004407028201967478, + 0.010031517595052719, + 0.016427848488092422, + 0.00971132516860962, + 0.007660719100385904, + 0.005998310633003712, + 0.023060392588377, + 0.004642914980649948, + 0.01281795185059309, + 0.007425986230373383, + 0.009990084916353226, + 0.006112618371844292, + 0.009532991796731949, + 0.009489819407463074, + 0.012447049841284752, + 0.019421998411417007, + 0.011611145921051502, + 0.010377871803939342, + 0.02161647006869316, + 0.007731880061328411, + 0.010694446042180061, + 0.005443800240755081, + 0.006967552006244659, + 0.007949008606374264, + 0.0030970906373113394, + 0.009097505360841751, + 0.00671770516782999, + 0.006031924858689308, + 0.004583780653774738, + 0.00724096829071641, + 0.007095875684171915, + 0.01473203208297491, + 0.006350185722112656, + 0.01269895676523447, + 0.014849095605313778, + 0.010070141404867172, + 0.012700577266514301, + 0.004340154118835926, + 0.003838673932477832, + 0.005654349457472563, + 0.004419345408678055, + 0.005879587028175592, + 0.003875848138704896, + 0.0062250676564872265, + 0.004779970273375511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.018183674663305283 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058008871972561 + ], + "opposite_logprob_tokens": [ + -3.8955800533294678 + ], + "logprob_reward": -0.02058008871972561, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058008871972561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28195151686668396, + 0.06776607781648636, + 0.028823247179389, + 0.14393974840641022, + 0.0447859913110733, + 0.11653704941272736, + 0.11138477176427841, + 0.15216514468193054, + 0.03313868120312691, + 0.019507789984345436 + ], + "credit_map": [ + 0.036030497401952744, + 0.05690331384539604, + 0.042705509811639786, + 0.03107336349785328, + 0.04279835522174835, + 0.026324613019824028, + 0.022485757246613503, + 0.023630093783140182, + 0.006334993522614241, + 0.006949246395379305, + 0.00824191328138113, + 0.008639280684292316, + 0.011681994423270226, + 0.005920014344155788, + 0.009681752882897854, + 0.010316881351172924, + 0.0029387311078608036, + 0.002294686622917652, + 0.003755912883207202, + 0.0036875796504318714, + 0.004285437986254692, + 0.003226948669180274, + 0.00458851782605052, + 0.0040454319678246975, + 0.009009577333927155, + 0.01970476284623146, + 0.012186448089778423, + 0.027346404269337654, + 0.03242170065641403, + 0.017337176948785782, + 0.014966567978262901, + 0.010967111214995384, + 0.007776118814945221, + 0.006228444632142782, + 0.0063818031921982765, + 0.005044992547482252, + 0.0038535662461072206, + 0.0065622832626104355, + 0.0033343485556542873, + 0.005604431964457035, + 0.009823846630752087, + 0.040640346705913544, + 0.007992315106093884, + 0.021888693794608116, + 0.00645851856097579, + 0.014837246388196945, + 0.005394080188125372, + 0.009502002038061619, + 0.008858106099069118, + 0.0162625964730978, + 0.018492449074983597, + 0.016674047335982323, + 0.009730177000164986, + 0.019774459302425385, + 0.00802804995328188, + 0.013564893044531345, + 0.024525469169020653, + 0.031734101474285126, + 0.015482866205275059, + 0.026148881763219833, + 0.00760194007307291, + 0.02535107173025608, + 0.006454463116824627, + 0.014866353012621403, + 0.0035454086028039455, + 0.0038672559894621372, + 0.005136239342391491, + 0.0034899208694696426, + 0.0036382635589689016, + 0.005633270833641291, + 0.0031611330341547728, + 0.004667188040912151, + 0.002592426957562566, + 0.0021332765463739634, + 0.0025050684344023466, + 0.0031082737259566784, + 0.00186181312892586, + 0.0021702584344893694, + 0.0019816418644040823, + 0.003155030542984605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.02058008871972561 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028255560901015997 + ], + "opposite_logprob_tokens": [ + -5.877825736999512 + ], + "logprob_reward": -0.0028255560901015997, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028255560901015997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33651241660118103, + 0.08216319978237152, + 0.05474076420068741, + 0.11991804093122482, + 0.1020304337143898, + 0.04036937654018402, + 0.07723566889762878, + 0.08513044565916061, + 0.06527843326330185, + 0.03662124276161194 + ], + "credit_map": [ + 0.047553662210702896, + 0.05450594052672386, + 0.04059012606739998, + 0.044487982988357544, + 0.03838624060153961, + 0.03543051332235336, + 0.03846648335456848, + 0.03709147125482559, + 0.008701134473085403, + 0.006526091136038303, + 0.008737974800169468, + 0.009743047878146172, + 0.00938344281166792, + 0.007949572056531906, + 0.01743663288652897, + 0.013685307465493679, + 0.005866602528840303, + 0.004896495491266251, + 0.005126229487359524, + 0.006307411007583141, + 0.005544999614357948, + 0.008855954743921757, + 0.007342274766415358, + 0.010800795629620552, + 0.011915093287825584, + 0.01604493521153927, + 0.018013738095760345, + 0.013000616803765297, + 0.023924175649881363, + 0.015106800943613052, + 0.0119545953348279, + 0.009958087466657162, + 0.010710581205785275, + 0.016037916764616966, + 0.013335599564015865, + 0.015186459757387638, + 0.008509323000907898, + 0.01807432621717453, + 0.009757230058312416, + 0.01041900273412466, + 0.003737102262675762, + 0.006414863746613264, + 0.0049518742598593235, + 0.004632547963410616, + 0.007326030638068914, + 0.004122503101825714, + 0.004835603758692741, + 0.0043488494120538235, + 0.007470385637134314, + 0.009219993837177753, + 0.009676672518253326, + 0.00885208323597908, + 0.0075719053857028484, + 0.019052840769290924, + 0.006754610221832991, + 0.008637174032628536, + 0.011942841112613678, + 0.014259648509323597, + 0.011298955418169498, + 0.012112117372453213, + 0.008668000809848309, + 0.013578616082668304, + 0.006250694859772921, + 0.007019577547907829, + 0.006233097054064274, + 0.007578207645565271, + 0.010261328890919685, + 0.006596214137971401, + 0.0075604296289384365, + 0.011665395461022854, + 0.0074949804693460464, + 0.007888780906796455, + 0.0038403344806283712, + 0.0035047896672040224, + 0.005408271681517363, + 0.005138637963682413, + 0.004791693761944771, + 0.0038573311176151037, + 0.004943199455738068, + 0.005136983934789896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.0028255560901015997 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, at eye level, begins focusing on a woman with long, wavy auburn hair, calmly walking forward along a stone path. The camera immediately begins tracking backward in a smooth, steady motion, maintaining a consistent distance. She wears a flowing cream linen sundress that flutters gently around her knees, her bare feet pressing lightly against the cool stones. A gentle breeze rustles through the surrounding leaves, ruffling strands of her hair and enhancing the sense of tranquil movement. As the camera continues to pull back, the frame widens, revealing a vast, lush botanical garden bathed in the warm, golden light of the afternoon. The stunning landscape is showcased, including manicured green lawns, vibrant flower pots blooming with pink roses and purple lavender, and ancient, mossy trees whose leaves create dappled patterns of light and shadow on the ground. Her graceful, unhurried gait and contented expression, combined with the idyllic surroundings, create an atmosphere of profound calm and quiet contemplation.", + "seed": 32814, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_2/index_4531_seed_32814_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.00701035694219172, + "vlm_logprob": -0.00701035694219172, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.00701035694219172, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008087624795734882 + ], + "opposite_logprob_tokens": [ + -5.008087635040283 + ], + "logprob_reward": -0.008087624795734882, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008087624795734882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054235316812992096, + 0.026757564395666122, + 0.04452671483159065, + 0.6339790225028992, + 0.03169816732406616, + 0.047651972621679306, + 0.06464405357837677, + 0.027087071910500526, + 0.02209462597966194, + 0.0473254956305027 + ], + "credit_map": [ + 0.0046798596158623695, + 0.004306684248149395, + 0.007994575425982475, + 0.005087883211672306, + 0.010970665141940117, + 0.0044464427046477795, + 0.007896993309259415, + 0.008852214552462101, + 0.002305645728483796, + 0.0034011893440037966, + 0.0030199799221009016, + 0.0037298777606338263, + 0.0036926318425685167, + 0.001984764588996768, + 0.0062375906854867935, + 0.002385883592069149, + 0.0028645892161875963, + 0.005254612769931555, + 0.00426889443770051, + 0.009729981422424316, + 0.00814027339220047, + 0.003282504854723811, + 0.007342214696109295, + 0.0036436475347727537, + 0.03518315404653549, + 0.07068981975317001, + 0.03924327343702316, + 0.255922794342041, + 0.03437350317835808, + 0.020184257999062538, + 0.1519424319267273, + 0.026439806446433067, + 0.004646761808544397, + 0.002676549134775996, + 0.0037559117190539837, + 0.00311528448946774, + 0.0037881010212004185, + 0.00579018983989954, + 0.0031748423352837563, + 0.004750525113195181, + 0.007809740491211414, + 0.006113107781857252, + 0.0052047655917704105, + 0.00805701594799757, + 0.0056160748936235905, + 0.004063780885189772, + 0.0073909149505198, + 0.0033965737093240023, + 0.005672135390341282, + 0.006285790354013443, + 0.010204973630607128, + 0.005456200335174799, + 0.015499528497457504, + 0.011086598969995975, + 0.005148650147020817, + 0.005290173459798098, + 0.0032229081261903048, + 0.0027943034656345844, + 0.005750881973654032, + 0.0024633724242448807, + 0.0036825588904321194, + 0.004743647761642933, + 0.0023330459371209145, + 0.002096354030072689, + 0.0032996931113302708, + 0.0015152349369600415, + 0.0043591842986643314, + 0.002147423569113016, + 0.002600436331704259, + 0.0030684496741741896, + 0.002464256016537547, + 0.0026399472262710333, + 0.004931380040943623, + 0.010948672890663147, + 0.004563501570373774, + 0.004643876105546951, + 0.003609317820519209, + 0.004313770681619644, + 0.00467927660793066, + 0.009635700844228268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.008087624795734882 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046449145302176476 + ], + "opposite_logprob_tokens": [ + -5.379644870758057 + ], + "logprob_reward": -0.0046449145302176476, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046449145302176476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1601042002439499, + 0.2477264702320099, + 0.08824312686920166, + 0.15436901152133942, + 0.040073003619909286, + 0.06227108836174011, + 0.048128075897693634, + 0.10690878331661224, + 0.04304886236786842, + 0.049127429723739624 + ], + "credit_map": [ + 0.013973813503980637, + 0.016039587557315826, + 0.02400854043662548, + 0.01699022203683853, + 0.03100951761007309, + 0.02041931077837944, + 0.014881250448524952, + 0.02278195507824421, + 0.014290301129221916, + 0.03122456930577755, + 0.018374985083937645, + 0.043401703238487244, + 0.042356546968221664, + 0.014633861370384693, + 0.06396064907312393, + 0.01948384940624237, + 0.008682419545948505, + 0.008526097051799297, + 0.009772857651114464, + 0.017813818529248238, + 0.011825958266854286, + 0.012019305489957333, + 0.011495723389089108, + 0.008106939494609833, + 0.009943485260009766, + 0.018002666532993317, + 0.015584106557071209, + 0.056828003376722336, + 0.009788693860173225, + 0.007639244198799133, + 0.029582396149635315, + 0.007000409997999668, + 0.005124953109771013, + 0.004622786305844784, + 0.00421008188277483, + 0.006200565956532955, + 0.0049071768298745155, + 0.0052696107886731625, + 0.005913253407925367, + 0.003824575338512659, + 0.009256896562874317, + 0.00961351953446865, + 0.010038292035460472, + 0.005558648146688938, + 0.007307569496333599, + 0.006465468090027571, + 0.007453604135662317, + 0.006577091757208109, + 0.0060728187672793865, + 0.003524468280375004, + 0.010410667397081852, + 0.0030843615531921387, + 0.00507004838436842, + 0.008968459442257881, + 0.0026861908845603466, + 0.008311057463288307, + 0.014836483635008335, + 0.014094507321715355, + 0.013280682265758514, + 0.013664055615663528, + 0.012467785738408566, + 0.01569966971874237, + 0.014624384231865406, + 0.008241213858127594, + 0.005726453848183155, + 0.0032015005126595497, + 0.01029653288424015, + 0.0031852128449827433, + 0.0044087814167141914, + 0.006708015222102404, + 0.003991643898189068, + 0.005530722439289093, + 0.00567104946821928, + 0.011139566078782082, + 0.0055676973424851894, + 0.003312841057777405, + 0.0044555035419762135, + 0.004583321046084166, + 0.01053833682090044, + 0.0038591132033616304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0046449145302176476 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626583334058523 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626583334058523, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626583334058523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13173744082450867, + 0.07952205091714859, + 0.1275414526462555, + 0.0964444950222969, + 0.038406968116760254, + 0.05561518296599388, + 0.05305885523557663, + 0.035051994025707245, + 0.04644669592380524, + 0.3361749053001404 + ], + "credit_map": [ + 0.010529626160860062, + 0.01537462417036295, + 0.021873753517866135, + 0.01291150413453579, + 0.023199396207928658, + 0.014477557502686977, + 0.015312647446990013, + 0.018058331683278084, + 0.005074960179626942, + 0.009413142688572407, + 0.006848912686109543, + 0.014240968972444534, + 0.010861730203032494, + 0.004747605882585049, + 0.021665992215275764, + 0.006668725982308388, + 0.013716303743422031, + 0.01191672682762146, + 0.015312177129089832, + 0.027333879843354225, + 0.012895805761218071, + 0.01760677434504032, + 0.015516344457864761, + 0.013243434950709343, + 0.006488494109362364, + 0.015331635251641273, + 0.01264665275812149, + 0.02594452165067196, + 0.011065352708101273, + 0.007622449658811092, + 0.012272810563445091, + 0.005072572734206915, + 0.005199097562581301, + 0.0034456748981028795, + 0.004748234990984201, + 0.004519654903560877, + 0.003255253192037344, + 0.004909294657409191, + 0.007388795260339975, + 0.004940958693623543, + 0.008129787631332874, + 0.006452889647334814, + 0.009967050515115261, + 0.005838419776409864, + 0.005387388169765472, + 0.005301706492900848, + 0.005657648202031851, + 0.008880283683538437, + 0.004864908289164305, + 0.00555306114256382, + 0.010329753160476685, + 0.004684069659560919, + 0.00848795473575592, + 0.008498293347656727, + 0.004499811679124832, + 0.006141000427305698, + 0.0036337878555059433, + 0.0029406866524368525, + 0.005241428967565298, + 0.00327519653365016, + 0.005222939886152744, + 0.004830323159694672, + 0.004294347949326038, + 0.005613283254206181, + 0.006990163587033749, + 0.00369427353143692, + 0.009205573238432407, + 0.0038403202779591084, + 0.004009145777672529, + 0.006126394961029291, + 0.0031404669862240553, + 0.009440351277589798, + 0.038644030690193176, + 0.075168676674366, + 0.030137917026877403, + 0.012423647567629814, + 0.024430792778730392, + 0.026816660538315773, + 0.09432119131088257, + 0.03423195704817772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.007626583334058523 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107026170939207 + ], + "opposite_logprob_tokens": [ + -4.511070251464844 + ], + "logprob_reward": -0.01107026170939207, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107026170939207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1430777907371521, + 0.054055240005254745, + 0.06215917691588402, + 0.4310143291950226, + 0.02601582370698452, + 0.03194389119744301, + 0.04587696120142937, + 0.07423056662082672, + 0.05598146468400955, + 0.07564471662044525 + ], + "credit_map": [ + 0.011686776764690876, + 0.014764325693249702, + 0.020606886595487595, + 0.013798831962049007, + 0.028594492003321648, + 0.01719764620065689, + 0.014963708817958832, + 0.021465132012963295, + 0.003939639311283827, + 0.005955772008746862, + 0.005318681709468365, + 0.008829683065414429, + 0.009028560481965542, + 0.0035581854172050953, + 0.012997172772884369, + 0.004427552688866854, + 0.006175063084810972, + 0.00554449949413538, + 0.01158959697932005, + 0.006617930717766285, + 0.00794396735727787, + 0.00903924461454153, + 0.00687420042231679, + 0.008374679833650589, + 0.02541397325694561, + 0.050418391823768616, + 0.037479840219020844, + 0.16673584282398224, + 0.025614894926548004, + 0.018030107021331787, + 0.09053649008274078, + 0.016784846782684326, + 0.004070727154612541, + 0.00265314569696784, + 0.0034394022077322006, + 0.0023267127107828856, + 0.0019223588751628995, + 0.006102692801505327, + 0.0018335810163989663, + 0.0036672079004347324, + 0.004483677446842194, + 0.004995367489755154, + 0.004333097022026777, + 0.0034834900870919228, + 0.0042303199879825115, + 0.002772762905806303, + 0.004178395960479975, + 0.0034667851869016886, + 0.004962042439728975, + 0.0034025695640593767, + 0.009920242242515087, + 0.003172211814671755, + 0.004859941080212593, + 0.008633364923298359, + 0.0029015911277383566, + 0.008025006391108036, + 0.009579350240528584, + 0.005612313747406006, + 0.0199714545160532, + 0.004753962624818087, + 0.007102832198143005, + 0.013715908862650394, + 0.005774263292551041, + 0.007720493711531162, + 0.010072030127048492, + 0.003438022453337908, + 0.011828888207674026, + 0.0036656982265412807, + 0.003244055900722742, + 0.007237492594867945, + 0.0026224744506180286, + 0.013872811570763588, + 0.008915378712117672, + 0.018140224739909172, + 0.011941773816943169, + 0.003918249160051346, + 0.004830700810998678, + 0.006212692707777023, + 0.016051992774009705, + 0.005633714143186808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.01107026170939207 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003622400341555476 + ], + "opposite_logprob_tokens": [ + -5.628622531890869 + ], + "logprob_reward": -0.003622400341555476, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003622400341555476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1505584567785263, + 0.03746194764971733, + 0.041004959493875504, + 0.033357180655002594, + 0.041682541370391846, + 0.04962761327624321, + 0.06719046086072922, + 0.03614257648587227, + 0.18905776739120483, + 0.3539165258407593 + ], + "credit_map": [ + 0.013342060148715973, + 0.01339433342218399, + 0.019229929894208908, + 0.016472192481160164, + 0.03223614767193794, + 0.015303936786949635, + 0.019430583342909813, + 0.021149270236492157, + 0.00277361529879272, + 0.004356912337243557, + 0.004425732418894768, + 0.0050821928307414055, + 0.005259139463305473, + 0.002665511332452297, + 0.00926826149225235, + 0.003630578750744462, + 0.003751414129510522, + 0.005097199697047472, + 0.004979124758392572, + 0.006613283883780241, + 0.006450142711400986, + 0.005166530143469572, + 0.0055091604590415955, + 0.0034381034784018993, + 0.002722719917073846, + 0.005335886497050524, + 0.0029396144673228264, + 0.003984036389738321, + 0.006117194425314665, + 0.0045904843136668205, + 0.004036356694996357, + 0.003630885621532798, + 0.005651447456330061, + 0.004276539199054241, + 0.007988166995346546, + 0.004218161106109619, + 0.003078269073739648, + 0.006258262321352959, + 0.0049276091158390045, + 0.005284085404127836, + 0.007208344992250204, + 0.008418765850365162, + 0.005092963110655546, + 0.0065808966755867004, + 0.00638375012204051, + 0.00398494116961956, + 0.007386046927422285, + 0.004571903962641954, + 0.009358126670122147, + 0.004790293984115124, + 0.014409849420189857, + 0.005391557700932026, + 0.006628782022744417, + 0.0127895288169384, + 0.0039548687636852264, + 0.00986744649708271, + 0.0058709303848445415, + 0.003241840749979019, + 0.007865352556109428, + 0.00246067950502038, + 0.002846162300556898, + 0.006476325914263725, + 0.0021133534610271454, + 0.005267929285764694, + 0.03239106386899948, + 0.012770925648510456, + 0.0397237204015255, + 0.01244033221155405, + 0.01207246258854866, + 0.024826733395457268, + 0.009719721041619778, + 0.04511280730366707, + 0.04059365391731262, + 0.08730281889438629, + 0.031847406178712845, + 0.013366823084652424, + 0.02512279897928238, + 0.029132621362805367, + 0.09681915491819382, + 0.029731228947639465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.003622400341555476 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, at eye level, begins focusing on a woman with long, wavy auburn hair, calmly walking forward along a stone path. The camera immediately begins tracking backward in a smooth, steady motion, maintaining a consistent distance. She wears a flowing cream linen sundress that flutters gently around her knees, her bare feet pressing lightly against the cool stones. A gentle breeze rustles through the surrounding leaves, ruffling strands of her hair and enhancing the sense of tranquil movement. As the camera continues to pull back, the frame widens, revealing a vast, lush botanical garden bathed in the warm, golden light of the afternoon. The stunning landscape is showcased, including manicured green lawns, vibrant flower pots blooming with pink roses and purple lavender, and ancient, mossy trees whose leaves create dappled patterns of light and shadow on the ground. Her graceful, unhurried gait and contented expression, combined with the idyllic surroundings, create an atmosphere of profound calm and quiet contemplation.", + "seed": 32815, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_2/index_4531_seed_32815_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.008617420401424169, + "vlm_logprob": -0.008617420401424169, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.008617420401424169, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010716980323195457 + ], + "opposite_logprob_tokens": [ + -4.635716915130615 + ], + "logprob_reward": -0.010716980323195457, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010716980323195457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14233344793319702, + 0.11951754987239838, + 0.1157870963215828, + 0.0806170329451561, + 0.19125039875507355, + 0.14733648300170898, + 0.07467816025018692, + 0.07304064929485321, + 0.024502107873558998, + 0.030937012284994125 + ], + "credit_map": [ + 0.0117436358705163, + 0.014063095673918724, + 0.012190394103527069, + 0.014889517799019814, + 0.025827545672655106, + 0.012719344347715378, + 0.029760245233774185, + 0.021139681339263916, + 0.009205116890370846, + 0.011293049901723862, + 0.008530947379767895, + 0.01339261420071125, + 0.016148457303643227, + 0.013893119059503078, + 0.02931704930961132, + 0.017737217247486115, + 0.008563224226236343, + 0.014080381952226162, + 0.009935680776834488, + 0.018073230981826782, + 0.023537680506706238, + 0.010645364411175251, + 0.017829226329922676, + 0.013122310861945152, + 0.005112035199999809, + 0.005674163345247507, + 0.007930498570203781, + 0.011488527990877628, + 0.019878651946783066, + 0.00502180028706789, + 0.011711258441209793, + 0.013800106011331081, + 0.020475490018725395, + 0.016714198514819145, + 0.026348210871219635, + 0.031152736395597458, + 0.036037713289260864, + 0.01616266556084156, + 0.02148834429681301, + 0.022871065884828568, + 0.04120427742600441, + 0.012843482196331024, + 0.017759375274181366, + 0.012942010536789894, + 0.008655108511447906, + 0.015513334423303604, + 0.010224559344351292, + 0.028194356709718704, + 0.006094497162848711, + 0.011107136495411396, + 0.006340622436255217, + 0.010465094819664955, + 0.017545759677886963, + 0.008487872779369354, + 0.006081266328692436, + 0.00855591893196106, + 0.006873755715787411, + 0.008727099746465683, + 0.006657538469880819, + 0.009098147042095661, + 0.015131647698581219, + 0.009132789447903633, + 0.011259107850492, + 0.006160566583275795, + 0.00295383483171463, + 0.004082729574292898, + 0.003910913597792387, + 0.001958988606929779, + 0.003688076976686716, + 0.0030601152684539557, + 0.0026977057568728924, + 0.00214974582195282, + 0.003137691877782345, + 0.003939451649785042, + 0.003906875383108854, + 0.005387462675571442, + 0.0042020599357783794, + 0.0037231240421533585, + 0.003321475815027952, + 0.003318873466923833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.010716980323195457 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595793966203928 + ], + "opposite_logprob_tokens": [ + -5.130958080291748 + ], + "logprob_reward": -0.00595793966203928, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595793966203928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15065646171569824, + 0.3128267228603363, + 0.058680787682533264, + 0.07529820501804352, + 0.13469579815864563, + 0.10946181416511536, + 0.046604230999946594, + 0.052834030240774155, + 0.030054738745093346, + 0.028887230902910233 + ], + "credit_map": [ + 0.00815964862704277, + 0.013337473385035992, + 0.009632737375795841, + 0.019788645207881927, + 0.027183841913938522, + 0.010670127347111702, + 0.03334139660000801, + 0.028542593121528625, + 0.02311483584344387, + 0.02462604083120823, + 0.02255384996533394, + 0.04952393099665642, + 0.035978544503450394, + 0.02771032229065895, + 0.06482160836458206, + 0.06449756771326065, + 0.005117930006235838, + 0.006966293323785067, + 0.004944699816405773, + 0.0074314093217253685, + 0.011142716743052006, + 0.004151988308876753, + 0.010582208633422852, + 0.008343541994690895, + 0.00461371848359704, + 0.008655920624732971, + 0.005196488928049803, + 0.010538407601416111, + 0.019191782921552658, + 0.004890149459242821, + 0.011481672525405884, + 0.01073006633669138, + 0.008974909782409668, + 0.013284391723573208, + 0.01073465496301651, + 0.02237226627767086, + 0.03876098245382309, + 0.009477589279413223, + 0.02136017195880413, + 0.009730842895805836, + 0.033769652247428894, + 0.00990957673639059, + 0.013859841972589493, + 0.007107714656740427, + 0.007575638592243195, + 0.011482789181172848, + 0.007192493416368961, + 0.01856410689651966, + 0.006872400175780058, + 0.004866438452154398, + 0.00675598019734025, + 0.0037782026920467615, + 0.00808025524020195, + 0.005989220924675465, + 0.004017109051346779, + 0.006244628690183163, + 0.005779766943305731, + 0.0063439467921853065, + 0.0071007502265274525, + 0.006379949860274792, + 0.009018859826028347, + 0.007443548180162907, + 0.006319598760455847, + 0.00444761011749506, + 0.003417228814214468, + 0.004055889323353767, + 0.005884234327822924, + 0.0033872390631586313, + 0.0032119802199304104, + 0.0038894934114068747, + 0.003263298887759447, + 0.0029453751631081104, + 0.0035081706009805202, + 0.003765682689845562, + 0.003980332985520363, + 0.0031906315125524998, + 0.0038677561096847057, + 0.0032997927628457546, + 0.0035917023196816444, + 0.003683162620291114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.00595793966203928 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777040220797062 + ], + "opposite_logprob_tokens": [ + -4.634777069091797 + ], + "logprob_reward": -0.009777040220797062, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777040220797062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07601223140954971, + 0.17050442099571228, + 0.07699999958276749, + 0.10073486715555191, + 0.23021449148654938, + 0.11430244892835617, + 0.08751675486564636, + 0.03951486945152283, + 0.07705098390579224, + 0.027148935943841934 + ], + "credit_map": [ + 0.0062222531996667385, + 0.0052210926078259945, + 0.008563991636037827, + 0.01230594702064991, + 0.008662035688757896, + 0.006499808747321367, + 0.013598809018731117, + 0.014938301406800747, + 0.016694992780685425, + 0.018402045592665672, + 0.015856541693210602, + 0.026475723832845688, + 0.01643485017120838, + 0.02342807874083519, + 0.028111008927226067, + 0.025101173669099808, + 0.005294511094689369, + 0.006930377334356308, + 0.00593741238117218, + 0.00915847159922123, + 0.014691430144011974, + 0.005693400278687477, + 0.017899038270115852, + 0.011395360343158245, + 0.005646717734634876, + 0.011062627658247948, + 0.008976010605692863, + 0.010377523489296436, + 0.02698766626417637, + 0.006793975830078125, + 0.015094137750566006, + 0.015796206891536713, + 0.033279139548540115, + 0.015854865312576294, + 0.05902448669075966, + 0.023728622123599052, + 0.030237464234232903, + 0.024265063926577568, + 0.020450839772820473, + 0.02337399497628212, + 0.03131252899765968, + 0.010293742641806602, + 0.01279483549296856, + 0.008288874290883541, + 0.013193625025451183, + 0.011111099272966385, + 0.008006386458873749, + 0.01930135302245617, + 0.012230639345943928, + 0.009271723218262196, + 0.01133659016340971, + 0.008623247966170311, + 0.01050200778990984, + 0.012369873933494091, + 0.008990230970084667, + 0.014192437753081322, + 0.003870817134156823, + 0.005008204374462366, + 0.005139679182320833, + 0.004104724619537592, + 0.006378914695233107, + 0.00516835693269968, + 0.005064604338258505, + 0.004779570270329714, + 0.006038969848304987, + 0.009885541163384914, + 0.017796959728002548, + 0.00839858129620552, + 0.009424423798918724, + 0.010796137154102325, + 0.006353979464620352, + 0.008356393314898014, + 0.003332708729431033, + 0.003906069789081812, + 0.0033990079537034035, + 0.00239776074886322, + 0.004022389184683561, + 0.0031927404925227165, + 0.003935415297746658, + 0.0029628449119627476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.009777040220797062 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533704750239849 + ], + "opposite_logprob_tokens": [ + -4.387533664703369 + ], + "logprob_reward": -0.012533704750239849, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533704750239849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10437440872192383, + 0.08600598573684692, + 0.10681383311748505, + 0.08260967582464218, + 0.26843270659446716, + 0.07071923464536667, + 0.0521165207028389, + 0.12801827490329742, + 0.06224147602915764, + 0.03866789862513542 + ], + "credit_map": [ + 0.010310165584087372, + 0.008123768493533134, + 0.012884334661066532, + 0.016745293512940407, + 0.013032081536948681, + 0.010214314796030521, + 0.018865831196308136, + 0.01419861987233162, + 0.008969804272055626, + 0.00889632198959589, + 0.010058173909783363, + 0.016041338443756104, + 0.006639814004302025, + 0.013247093185782433, + 0.011223207227885723, + 0.010930231772363186, + 0.009892175905406475, + 0.01158967800438404, + 0.008163618855178356, + 0.010713687166571617, + 0.018592799082398415, + 0.008729715831577778, + 0.023200765252113342, + 0.0159313902258873, + 0.0088481605052948, + 0.008360320702195168, + 0.009798883460462093, + 0.011358530260622501, + 0.01243255753070116, + 0.006774661596864462, + 0.010042231529951096, + 0.014994323253631592, + 0.036903902888298035, + 0.015698233619332314, + 0.06827457249164581, + 0.02917327545583248, + 0.03523280471563339, + 0.030491339042782784, + 0.02474464848637581, + 0.02791394479572773, + 0.007405237760394812, + 0.008145326748490334, + 0.005455091129988432, + 0.01213344931602478, + 0.011718777939677238, + 0.008009695447981358, + 0.007282117381691933, + 0.010569537989795208, + 0.007004257757216692, + 0.005990944802761078, + 0.006929432507604361, + 0.005710914731025696, + 0.00933840125799179, + 0.005979530978947878, + 0.0037221938837319613, + 0.007440844550728798, + 0.013708945363759995, + 0.016680536791682243, + 0.019414152950048447, + 0.011815755628049374, + 0.01835673488676548, + 0.0156269371509552, + 0.01051397155970335, + 0.021901240572333336, + 0.005362705793231726, + 0.0076087722554802895, + 0.014589759521186352, + 0.006992534268647432, + 0.005705259740352631, + 0.00945424847304821, + 0.005406791809946299, + 0.0071214064955711365, + 0.004276246763765812, + 0.005714711267501116, + 0.005602008663117886, + 0.0057452889159321785, + 0.004818044137209654, + 0.004903056658804417, + 0.0039781229570508, + 0.003630418563261628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 2 + ], + "selected_reward": -0.012533704750239849 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot at eye level featuring a woman with long, wavy auburn hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a flowing cream linen sundress and walking barefoot on a stone path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera track backward in a smooth, steady motion while maintaining a consistent distance from the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast, lush botanical garden with manicured green lawns, pink roses, purple lavender, and ancient mossy trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, golden afternoon light with a gentle breeze rustling the leaves and the woman's hair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101437050849199 + ], + "opposite_logprob_tokens": [ + -5.504101276397705 + ], + "logprob_reward": -0.004101437050849199, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101437050849199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1291465312242508, + 0.13295073807239532, + 0.06407488137483597, + 0.06506171822547913, + 0.15184751152992249, + 0.0399237722158432, + 0.10813799500465393, + 0.10414976626634598, + 0.1169981136918068, + 0.08770893514156342 + ], + "credit_map": [ + 0.010734851472079754, + 0.010871750302612782, + 0.008937006816267967, + 0.025302033871412277, + 0.016676004976034164, + 0.009399371221661568, + 0.02225571498274803, + 0.024969790130853653, + 0.006925182417035103, + 0.010349501855671406, + 0.011701836250722408, + 0.020476432517170906, + 0.018040012568235397, + 0.011868898756802082, + 0.02581564523279667, + 0.027773238718509674, + 0.004955369513481855, + 0.007023381069302559, + 0.005857577081769705, + 0.011237683705985546, + 0.010120133869349957, + 0.00437864800915122, + 0.009077461436390877, + 0.011424622498452663, + 0.005058330949395895, + 0.0059748743660748005, + 0.006373927928507328, + 0.006118178367614746, + 0.015179669484496117, + 0.00421179411932826, + 0.010434780269861221, + 0.011710161343216896, + 0.02233608067035675, + 0.00836279708892107, + 0.030673837289214134, + 0.01525831874459982, + 0.027735857293009758, + 0.015516775660216808, + 0.015836363658308983, + 0.01612747274339199, + 0.003782758256420493, + 0.0037990808486938477, + 0.0037887676153331995, + 0.00391022302210331, + 0.010313273407518864, + 0.003697232576087117, + 0.004273820202797651, + 0.006358616054058075, + 0.015755845233798027, + 0.009090057574212551, + 0.013551066629588604, + 0.009294570423662663, + 0.02764878235757351, + 0.010169350542128086, + 0.008573961444199085, + 0.014054358936846256, + 0.015988951548933983, + 0.01275764312595129, + 0.01829145848751068, + 0.009671601466834545, + 0.011495789512991905, + 0.01636611856520176, + 0.007399940397590399, + 0.012178263626992702, + 0.012860167771577835, + 0.014581531286239624, + 0.02439104951918125, + 0.012168141081929207, + 0.014873446896672249, + 0.014651201665401459, + 0.012239508330821991, + 0.011233069002628326, + 0.008270974270999432, + 0.011428212746977806, + 0.011959352530539036, + 0.015013040974736214, + 0.009533513337373734, + 0.010620522312819958, + 0.009083003737032413, + 0.011800317093729973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.004101437050849199 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_3/index_tensor([337, 337])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_3/index_tensor([337, 337])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b32f27f8e5de7edc7f7040ddfca0e61747012512 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_3/index_tensor([337, 337])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle shot captures a towering skeleton clad in tattered black robes, standing motionless in the center of a vast, eerie, abandoned hall. A cold, moonlit beam streams through a ripped hole in the dome, illuminating its ancient, yellowed bones. A faint, dying crimson glow flickers in its hollow eye sockets. The process of dissolution begins at its feet. As if consumed by an invisible flame, the solid bone structure steadily and silently disintegrates into a dense, inky black vapor. This black vapor spirals upward with an ominous energy, and the skeleton's heavy, tattered black robes dissolve into smoke, billowing in the updraft. As the decomposition process spread upward, its entire body gradually became transparent. Finally, when the last wisp of black steam rose from its slightly tilted head and completely dissipated in the air above the dim hall, the entire scene was left with only the lonely column of light, creating a solemn, sad and final magical atmosphere.", + "seed": 28232, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_3/index_337_seed_28232_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06311018597334624, + "vlm_logprob": -0.06311018597334624, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.06311018597334624, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012902350164949894 + ], + "opposite_logprob_tokens": [ + -4.88790225982666 + ], + "logprob_reward": -0.012902350164949894, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012902350164949894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3463662266731262, + 0.11599026620388031, + 0.026975996792316437, + 0.1292957067489624, + 0.03435462340712547, + 0.11077608168125153, + 0.04320219159126282, + 0.0940709114074707, + 0.05376574769616127, + 0.04520227387547493 + ], + "credit_map": [ + 0.03545044735074043, + 0.02913336455821991, + 0.12414001673460007, + 0.02133851684629917, + 0.009838003665208817, + 0.019272105768322945, + 0.01090810727328062, + 0.09628567844629288, + 0.006893159355968237, + 0.007570283487439156, + 0.03440125659108162, + 0.004990424029529095, + 0.006845173425972462, + 0.009257047437131405, + 0.005029373336583376, + 0.041003547608852386, + 0.0023762511555105448, + 0.005119228735566139, + 0.0034759410191327333, + 0.0024547020439058542, + 0.005432085134088993, + 0.002146573504433036, + 0.0025609189178794622, + 0.0034102965146303177, + 0.007908366620540619, + 0.004507032688707113, + 0.04213058203458786, + 0.007060337346047163, + 0.00652404734864831, + 0.010174700990319252, + 0.008030438795685768, + 0.04296019300818443, + 0.003914283588528633, + 0.003576349699869752, + 0.0047186897136271, + 0.0034642363898456097, + 0.0036815444473177195, + 0.003875160589814186, + 0.007542315870523453, + 0.0035820454359054565, + 0.006417759694159031, + 0.0076022217981517315, + 0.018361425027251244, + 0.00713646924123168, + 0.005671483464539051, + 0.03366601839661598, + 0.008586471900343895, + 0.023334238678216934, + 0.0031354245729744434, + 0.007129254285246134, + 0.0032784477807581425, + 0.005683987867087126, + 0.008703868836164474, + 0.004195090848952532, + 0.007459205575287342, + 0.0036169099621474743, + 0.011992476880550385, + 0.00965961255133152, + 0.008215942420065403, + 0.015391038730740547, + 0.01650926284492016, + 0.010323993861675262, + 0.011955670081079006, + 0.010022909380495548, + 0.006484979297965765, + 0.006481140851974487, + 0.00429613096639514, + 0.007699429988861084, + 0.008251509629189968, + 0.005570933222770691, + 0.009466412477195263, + 0.005515215918421745, + 0.005058335606008768, + 0.0035546866711229086, + 0.006337275728583336, + 0.005394547712057829, + 0.003979802597314119, + 0.005495897494256496, + 0.005940582137554884, + 0.009441143833100796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.012902350164949894 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309685528278351 + ], + "opposite_logprob_tokens": [ + -3.1680967807769775 + ], + "logprob_reward": -0.04309685528278351, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309685528278351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.324657142162323, + 0.1192021295428276, + 0.037452634423971176, + 0.1580711007118225, + 0.08968929201364517, + 0.09347884356975555, + 0.045989740639925, + 0.04339940473437309, + 0.040829092264175415, + 0.04723065346479416 + ], + "credit_map": [ + 0.03544478118419647, + 0.030088404193520546, + 0.10819626599550247, + 0.026021765545010567, + 0.01247806753963232, + 0.01798328012228012, + 0.013425813056528568, + 0.08101873099803925, + 0.006649964489042759, + 0.01629532501101494, + 0.019549624994397163, + 0.011455828323960304, + 0.01585404947400093, + 0.00945843756198883, + 0.015069452114403248, + 0.024869440123438835, + 0.004874365869909525, + 0.005789651069790125, + 0.004845907445997, + 0.003704247996211052, + 0.006220471579581499, + 0.003627862548455596, + 0.00348330638371408, + 0.004906817805022001, + 0.008874224498867989, + 0.008262085728347301, + 0.0483844131231308, + 0.009709717705845833, + 0.012385495007038116, + 0.015460703521966934, + 0.01121349073946476, + 0.04378095269203186, + 0.008420814760029316, + 0.009597823955118656, + 0.012309592217206955, + 0.008540447801351547, + 0.010558032430708408, + 0.011196793057024479, + 0.018737878650426865, + 0.010327897034585476, + 0.007860816083848476, + 0.005125370807945728, + 0.015626901760697365, + 0.007474695332348347, + 0.005494480486959219, + 0.02549479529261589, + 0.007723924703896046, + 0.018677853047847748, + 0.0037550712004303932, + 0.006639084778726101, + 0.004953188356012106, + 0.005423358175903559, + 0.007618382107466459, + 0.005205539520829916, + 0.008246894925832748, + 0.004148215986788273, + 0.005929472390562296, + 0.005472181364893913, + 0.0034192854072898626, + 0.0083580007776618, + 0.006894085556268692, + 0.004352965857833624, + 0.006022442132234573, + 0.0029509635642170906, + 0.005359348375350237, + 0.004883965011686087, + 0.005646092351526022, + 0.00536889536306262, + 0.005119844805449247, + 0.003962055779993534, + 0.0058320993557572365, + 0.004656786099076271, + 0.005971243605017662, + 0.004388320725411177, + 0.0022169824223965406, + 0.005444038659334183, + 0.005607614293694496, + 0.006508891005069017, + 0.00744399381801486, + 0.009649564512073994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.04309685528278351 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763675756752491 + ], + "opposite_logprob_tokens": [ + -4.882636547088623 + ], + "logprob_reward": -0.00763675756752491, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763675756752491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5510480999946594, + 0.08884308487176895, + 0.04530705139040947, + 0.07881194353103638, + 0.04699167236685753, + 0.04964670538902283, + 0.033456187695264816, + 0.03582127019762993, + 0.022549087181687355, + 0.04752486199140549 + ], + "credit_map": [ + 0.04622858390212059, + 0.04580468311905861, + 0.20307199656963348, + 0.02993624471127987, + 0.014683264307677746, + 0.031972989439964294, + 0.015536320395767689, + 0.1638140231370926, + 0.014285262674093246, + 0.006255446933209896, + 0.019521087408065796, + 0.005694916937500238, + 0.006507986690849066, + 0.007839024998247623, + 0.005204808432608843, + 0.023534560576081276, + 0.0031923402566462755, + 0.009351101703941822, + 0.0062189106829464436, + 0.00518313841894269, + 0.007779485080391169, + 0.0035726085770875216, + 0.005672542378306389, + 0.004336930811405182, + 0.0070206583477556705, + 0.0073412018828094006, + 0.019741712138056755, + 0.005220623686909676, + 0.006362424697726965, + 0.007585045415908098, + 0.005490530747920275, + 0.02004975639283657, + 0.006747102830559015, + 0.006763409357517958, + 0.006287943106144667, + 0.005175069905817509, + 0.004277597647160292, + 0.004824657924473286, + 0.008131889626383781, + 0.00478400569409132, + 0.0061888182535767555, + 0.005025618243962526, + 0.009965370409190655, + 0.00553007610142231, + 0.004284546710550785, + 0.009105381555855274, + 0.00519441207870841, + 0.0043524811044335365, + 0.0032524666748940945, + 0.004281040281057358, + 0.0043139373883605, + 0.004116796422749758, + 0.006134100258350372, + 0.0031609726138412952, + 0.005326904822140932, + 0.0028699710965156555, + 0.004504567477852106, + 0.005160186905413866, + 0.004419761709868908, + 0.003728459822013974, + 0.006512912455946207, + 0.0035438830964267254, + 0.004324904642999172, + 0.003626595251262188, + 0.002852684585377574, + 0.0031498712487518787, + 0.003079651854932308, + 0.002256285399198532, + 0.0018751000752672553, + 0.00254578934982419, + 0.0030095952097326517, + 0.0037801109720021486, + 0.005658985581248999, + 0.004625312052667141, + 0.0045572505332529545, + 0.005261489190161228, + 0.004001135006546974, + 0.00624395115301013, + 0.005588519852608442, + 0.011588220484554768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.00763675756752491 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348400861024857 + ], + "opposite_logprob_tokens": [ + -3.651348352432251 + ], + "logprob_reward": -0.026348400861024857, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348400861024857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4052373766899109, + 0.14714394509792328, + 0.03398441895842552, + 0.07368916273117065, + 0.043033115565776825, + 0.04930835962295532, + 0.04937691614031792, + 0.07695958763360977, + 0.08710968494415283, + 0.03415754809975624 + ], + "credit_map": [ + 0.0358651801943779, + 0.03471888229250908, + 0.13972289860248566, + 0.02546756900846958, + 0.02071279287338257, + 0.025340886786580086, + 0.016488878056406975, + 0.1069202870130539, + 0.009744076989591122, + 0.010424379259347916, + 0.040706224739551544, + 0.008484473451972008, + 0.008165576495230198, + 0.011986961588263512, + 0.00868945848196745, + 0.04894278571009636, + 0.0041846539825201035, + 0.005485523492097855, + 0.0042224060744047165, + 0.0032323456835001707, + 0.005690116435289383, + 0.0032255996484309435, + 0.004238292574882507, + 0.0037054806016385555, + 0.005407415330410004, + 0.00698784040287137, + 0.013143356889486313, + 0.006578370463103056, + 0.010826158337295055, + 0.008428853005170822, + 0.008817620575428009, + 0.01349954679608345, + 0.0063413577154278755, + 0.0049604130908846855, + 0.005522930528968573, + 0.004702298901975155, + 0.0049166386015713215, + 0.005232039839029312, + 0.005913074593991041, + 0.00544436601921916, + 0.005943658761680126, + 0.008335375227034092, + 0.006014293525367975, + 0.004830129910260439, + 0.004471327643841505, + 0.007115853484719992, + 0.0043947696685791016, + 0.008202950470149517, + 0.004906988702714443, + 0.005047356244176626, + 0.00933307595551014, + 0.0066210306249558926, + 0.008176974952220917, + 0.005532421637326479, + 0.006242196075618267, + 0.0035168705508112907, + 0.012072991579771042, + 0.008940726518630981, + 0.009311364032328129, + 0.010170492343604565, + 0.007932614535093307, + 0.010554035194218159, + 0.010759078897535801, + 0.007218286395072937, + 0.012949026189744473, + 0.007069814950227737, + 0.017656825482845306, + 0.00879652053117752, + 0.005696731619536877, + 0.010013051331043243, + 0.01015031524002552, + 0.014777394942939281, + 0.004358465317636728, + 0.0036499500274658203, + 0.004452364053577185, + 0.004589165095239878, + 0.003323026467114687, + 0.004190407227724791, + 0.0038778341840952635, + 0.005716334562748671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.026348400861024857 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.225566565990448 + ], + "opposite_logprob_tokens": [ + -1.6005666255950928 + ], + "logprob_reward": -0.225566565990448, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.225566565990448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3052161931991577, + 0.1770000010728836, + 0.03130091354250908, + 0.04707510769367218, + 0.03296399861574173, + 0.05841226130723953, + 0.09080418944358826, + 0.0729934573173523, + 0.07927524298429489, + 0.1049586832523346 + ], + "credit_map": [ + 0.027741745114326477, + 0.030386582016944885, + 0.10169775038957596, + 0.018900005146861076, + 0.015208481810986996, + 0.016004571691155434, + 0.0137880127876997, + 0.08148901909589767, + 0.011522116139531136, + 0.014410172589123249, + 0.05008576810359955, + 0.00892284233123064, + 0.009903592988848686, + 0.014730836264789104, + 0.00757985794916749, + 0.05984481796622276, + 0.0027655726298689842, + 0.006121574901044369, + 0.0033246371895074844, + 0.002933247247710824, + 0.005258152727037668, + 0.0032772300764918327, + 0.0033453102223575115, + 0.0042751869186758995, + 0.003458902006968856, + 0.00450114905834198, + 0.011669539846479893, + 0.004441745113581419, + 0.0045714364387094975, + 0.004041178617626429, + 0.004346603527665138, + 0.010044552385807037, + 0.0032709294464439154, + 0.0029649492353200912, + 0.007471158169209957, + 0.0037766871973872185, + 0.0034741591662168503, + 0.0023837087210267782, + 0.0037491454277187586, + 0.0058732605539262295, + 0.007602239493280649, + 0.005494589917361736, + 0.011204807087779045, + 0.0056025180965662, + 0.006142124067991972, + 0.006771814078092575, + 0.004798728972673416, + 0.010795439593493938, + 0.007742930203676224, + 0.009908461011946201, + 0.00819061417132616, + 0.012246204540133476, + 0.019531184807419777, + 0.010121646337211132, + 0.016309572383761406, + 0.006753579247742891, + 0.009610053151845932, + 0.008641594089567661, + 0.008626539260149002, + 0.0074655297212302685, + 0.011448309756815434, + 0.009896572679281235, + 0.009175203740596771, + 0.008129655383527279, + 0.009413119405508041, + 0.011316361837089062, + 0.008253064006567001, + 0.009917239658534527, + 0.011182151734828949, + 0.008018149994313717, + 0.011955360881984234, + 0.00921979546546936, + 0.01182516384869814, + 0.011344576254487038, + 0.008717804215848446, + 0.016213463619351387, + 0.016065437346696854, + 0.010872859507799149, + 0.01756782829761505, + 0.012351544573903084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.225566565990448 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle shot captures a towering skeleton clad in tattered black robes, standing motionless in the center of a vast, eerie, abandoned hall. A cold, moonlit beam streams through a ripped hole in the dome, illuminating its ancient, yellowed bones. A faint, dying crimson glow flickers in its hollow eye sockets. The process of dissolution begins at its feet. As if consumed by an invisible flame, the solid bone structure steadily and silently disintegrates into a dense, inky black vapor. This black vapor spirals upward with an ominous energy, and the skeleton's heavy, tattered black robes dissolve into smoke, billowing in the updraft. As the decomposition process spread upward, its entire body gradually became transparent. Finally, when the last wisp of black steam rose from its slightly tilted head and completely dissipated in the air above the dim hall, the entire scene was left with only the lonely column of light, creating a solemn, sad and final magical atmosphere.", + "seed": 28233, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_3/index_337_seed_28233_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05178005173802376, + "vlm_logprob": -0.05178005173802376, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.05178005173802376, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012291045859456062 + ], + "opposite_logprob_tokens": [ + -4.887290954589844 + ], + "logprob_reward": -0.012291045859456062, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012291045859456062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03984319791197777, + 0.025771673768758774, + 0.06107199937105179, + 0.2607438266277313, + 0.05103115364909172, + 0.03323308750987053, + 0.37380245327949524, + 0.04756250977516174, + 0.02162824757397175, + 0.08531176298856735 + ], + "credit_map": [ + 0.005733573343604803, + 0.005359915550798178, + 0.0026459465734660625, + 0.004508862271904945, + 0.00371127319522202, + 0.006951472256332636, + 0.004554103594273329, + 0.006378057412803173, + 0.004007503855973482, + 0.002823303220793605, + 0.0025433918926864862, + 0.0034641544334590435, + 0.0021729699801653624, + 0.004219201859086752, + 0.0032739201560616493, + 0.003267229301854968, + 0.01389843039214611, + 0.006449779029935598, + 0.004987244959920645, + 0.006359357386827469, + 0.006738410331308842, + 0.011034785769879818, + 0.0070692696608603, + 0.004534730222076178, + 0.09299194067716599, + 0.008937624283134937, + 0.1108323261141777, + 0.011051676236093044, + 0.006826852913945913, + 0.007543439511209726, + 0.005694596562534571, + 0.01686539500951767, + 0.012762323021888733, + 0.0022211018949747086, + 0.017385605722665787, + 0.00502183148637414, + 0.003916638437658548, + 0.0021815760992467403, + 0.00399566488340497, + 0.003546417923644185, + 0.0030002344865351915, + 0.0018767219735309482, + 0.0046494803391397, + 0.0044293212704360485, + 0.0025708891917020082, + 0.0015497150598093867, + 0.00640152720734477, + 0.008755205199122429, + 0.01811918430030346, + 0.008243098855018616, + 0.13078102469444275, + 0.020845428109169006, + 0.019273392856121063, + 0.011094154790043831, + 0.1476639211177826, + 0.01778229884803295, + 0.0040561743080616, + 0.004161521792411804, + 0.007436036132276058, + 0.009274372830986977, + 0.005466152913868427, + 0.003614910179749131, + 0.007094547152519226, + 0.00645880401134491, + 0.0024638681206852198, + 0.002449726453050971, + 0.0027052240911871195, + 0.0036278164479881525, + 0.0025839535519480705, + 0.0017074510687962174, + 0.004280698020011187, + 0.0018095126142725348, + 0.005100132431834936, + 0.004843340255320072, + 0.02691175602376461, + 0.028364650905132294, + 0.0039785923436284065, + 0.004054203163832426, + 0.008523276075720787, + 0.0035358252935111523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -0.012291045859456062 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207931786775589 + ], + "opposite_logprob_tokens": [ + -2.812079429626465 + ], + "logprob_reward": -0.06207931786775589, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207931786775589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048283159732818604, + 0.030277565121650696, + 0.07020130753517151, + 0.26655691862106323, + 0.041133858263492584, + 0.07575526088476181, + 0.29364046454429626, + 0.031167984008789062, + 0.08070426434278488, + 0.062279198318719864 + ], + "credit_map": [ + 0.007472609169781208, + 0.003943422343581915, + 0.00435628229752183, + 0.002874505240470171, + 0.003531038062646985, + 0.011201148852705956, + 0.004679051227867603, + 0.010225106962025166, + 0.005920663475990295, + 0.0025759160052984953, + 0.0028162943199276924, + 0.003048853250220418, + 0.0019762166775763035, + 0.006390222813934088, + 0.003161587053909898, + 0.004387813154608011, + 0.008487985469400883, + 0.0035648406483232975, + 0.003542443970218301, + 0.004057277459651232, + 0.003557623364031315, + 0.021166933700442314, + 0.004014897160232067, + 0.021809305995702744, + 0.09527686983346939, + 0.009247557260096073, + 0.11408720165491104, + 0.011651158332824707, + 0.007344647776335478, + 0.006932508200407028, + 0.005395755171775818, + 0.01662120223045349, + 0.007932274602353573, + 0.0036269230768084526, + 0.010345363058149815, + 0.004998546093702316, + 0.0034387458581477404, + 0.0022269615437835455, + 0.0041600377298891544, + 0.004405003972351551, + 0.009779758751392365, + 0.003140363609418273, + 0.008755411952733994, + 0.004614207427948713, + 0.0035256461706012487, + 0.003551609581336379, + 0.011598318815231323, + 0.030789941549301147, + 0.0161686223000288, + 0.006220056209713221, + 0.1065627932548523, + 0.016865115612745285, + 0.012793050147593021, + 0.0077465251088142395, + 0.1150604784488678, + 0.01222381554543972, + 0.00394837511703372, + 0.003086959244683385, + 0.003824743442237377, + 0.004827850963920355, + 0.004903510678559542, + 0.0027812558691948652, + 0.004570893943309784, + 0.0032243935856968164, + 0.007946331053972244, + 0.008727260865271091, + 0.013302017003297806, + 0.012437613680958748, + 0.0077823130413889885, + 0.008682172745466232, + 0.013169304467737675, + 0.008657251484692097, + 0.004409142769873142, + 0.004329270217567682, + 0.010702068917453289, + 0.01617477834224701, + 0.00521528534591198, + 0.004822980612516403, + 0.011924449354410172, + 0.004701222758740187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.06207931786775589 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005960546433925629 + ], + "opposite_logprob_tokens": [ + -5.130960464477539 + ], + "logprob_reward": -0.005960546433925629, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005960546433925629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06614238768815994, + 0.029029160737991333, + 0.10226466506719589, + 0.2039405256509781, + 0.08936188369989395, + 0.043923269957304, + 0.3026212155818939, + 0.02958499640226364, + 0.03032275289297104, + 0.1028091236948967 + ], + "credit_map": [ + 0.010817809030413628, + 0.004402184393256903, + 0.007894058711826801, + 0.006208668928593397, + 0.004326889291405678, + 0.014911253936588764, + 0.0065646045841276646, + 0.011016922071576118, + 0.004383622203022242, + 0.003850847715511918, + 0.00327998585999012, + 0.004224032629281282, + 0.0023398795165121555, + 0.004131903871893883, + 0.002928945468738675, + 0.0038899441715329885, + 0.028628384694457054, + 0.0062807281501591206, + 0.006658606696873903, + 0.008959249593317509, + 0.006004051771014929, + 0.01021644938737154, + 0.0071317278780043125, + 0.028385460376739502, + 0.07914333045482635, + 0.007142704911530018, + 0.08458296209573746, + 0.008577041327953339, + 0.005463819485157728, + 0.00612602848559618, + 0.004662912338972092, + 0.00824171956628561, + 0.02227795124053955, + 0.0037517501041293144, + 0.030808428302407265, + 0.008646742440760136, + 0.005183606408536434, + 0.004012680146843195, + 0.007298070937395096, + 0.00738265085965395, + 0.0044380491599440575, + 0.003545340383425355, + 0.006741965189576149, + 0.005154678598046303, + 0.0039772577583789825, + 0.003878367133438587, + 0.006760220509022474, + 0.009427391923964024, + 0.014673253521323204, + 0.0062093413434922695, + 0.108678437769413, + 0.017456240952014923, + 0.014653520658612251, + 0.008207416161894798, + 0.12039510160684586, + 0.012347938492894173, + 0.003574329661205411, + 0.003042650641873479, + 0.003969094250351191, + 0.004093426279723644, + 0.003563772654160857, + 0.002381514059379697, + 0.004666522145271301, + 0.004293688107281923, + 0.003364086151123047, + 0.003283863188698888, + 0.004173988942056894, + 0.004785431083291769, + 0.0036762633826583624, + 0.0028188019059598446, + 0.00546399736776948, + 0.002756321569904685, + 0.008260384202003479, + 0.008555464446544647, + 0.011812026612460613, + 0.020281346514821053, + 0.010679973289370537, + 0.008083793334662914, + 0.025700006633996964, + 0.009436129592359066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -0.005960546433925629 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818905957043171 + ], + "opposite_logprob_tokens": [ + -4.018188953399658 + ], + "logprob_reward": -0.01818905957043171, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818905957043171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05323284491896629, + 0.04338844493031502, + 0.06824066489934921, + 0.13628464937210083, + 0.2262619584798813, + 0.1766178458929062, + 0.13247641921043396, + 0.1084320917725563, + 0.028272178024053574, + 0.02679290995001793 + ], + "credit_map": [ + 0.008898417465388775, + 0.004793051164597273, + 0.004403123632073402, + 0.005445685237646103, + 0.002600406762212515, + 0.014444563537836075, + 0.0045635635033249855, + 0.008084029890596867, + 0.005710342433303595, + 0.0029318437445908785, + 0.0045190793462097645, + 0.004258754663169384, + 0.0024705762043595314, + 0.009662392549216747, + 0.004183010198175907, + 0.009652444161474705, + 0.0074490876868367195, + 0.009668751619756222, + 0.0054557607509195805, + 0.011154735460877419, + 0.009440556168556213, + 0.01097920723259449, + 0.006137307733297348, + 0.007955258712172508, + 0.044482383877038956, + 0.004966138396412134, + 0.05616959556937218, + 0.007471871562302113, + 0.003873890731483698, + 0.004338942933827639, + 0.0035049733705818653, + 0.01147684920579195, + 0.05897838994860649, + 0.009000447578728199, + 0.0830598995089531, + 0.022143632173538208, + 0.01161789521574974, + 0.005016854498535395, + 0.019054478034377098, + 0.017390340566635132, + 0.021207617595791817, + 0.007051821798086166, + 0.02205345407128334, + 0.009621987119317055, + 0.007307937368750572, + 0.008564993739128113, + 0.026245497167110443, + 0.07456453889608383, + 0.007910201326012611, + 0.006486867088824511, + 0.035009872168302536, + 0.012422525323927402, + 0.008544474840164185, + 0.005271423142403364, + 0.04647456482052803, + 0.01035649050027132, + 0.010834249667823315, + 0.010677574202418327, + 0.013908631168305874, + 0.015005853027105331, + 0.012790026143193245, + 0.010937260463833809, + 0.018787646666169167, + 0.015490849502384663, + 0.003614612855017185, + 0.0023656566627323627, + 0.003521321341395378, + 0.005804803688079119, + 0.003062994685024023, + 0.002605533692985773, + 0.005110979545861483, + 0.0021862746216356754, + 0.0019503135699778795, + 0.002533938502892852, + 0.003205053275451064, + 0.004735862370580435, + 0.002583142137154937, + 0.0024036995600908995, + 0.006795167922973633, + 0.002585733076557517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.01818905957043171 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603802889585495 + ], + "opposite_logprob_tokens": [ + -1.910380244255066 + ], + "logprob_reward": -0.1603802889585495, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603802889585495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.019341962411999702, + 0.011414213106036186, + 0.2249506711959839, + 0.12470658123493195, + 0.1329134702682495, + 0.022381357848644257, + 0.3772668242454529, + 0.017120545729994774, + 0.03064919263124466, + 0.03925520181655884 + ], + "credit_map": [ + 0.003125758608803153, + 0.0019824097398668528, + 0.0022661348339170218, + 0.0018979122396558523, + 0.0014351261779665947, + 0.0035130796022713184, + 0.0019411883549764752, + 0.0031803541351109743, + 0.001906957128085196, + 0.0013485393719747663, + 0.0008969844784587622, + 0.0021938420832157135, + 0.0007450521225109696, + 0.00166996568441391, + 0.0013446648372337222, + 0.0013082075165584683, + 0.04047893360257149, + 0.008128316141664982, + 0.010676585137844086, + 0.013731637969613075, + 0.0067476448602974415, + 0.04988718405365944, + 0.01491056103259325, + 0.08038982003927231, + 0.03760526701807976, + 0.007472256198525429, + 0.04384524002671242, + 0.010679957456886768, + 0.007412091828882694, + 0.004738515708595514, + 0.005763753782957792, + 0.007189491298049688, + 0.026508212089538574, + 0.00782908033579588, + 0.034291692078113556, + 0.01951448991894722, + 0.011378399096429348, + 0.005720668472349644, + 0.01567026972770691, + 0.012000657618045807, + 0.0019302752334624529, + 0.002250608755275607, + 0.002526234369724989, + 0.0028658758383244276, + 0.002564823953434825, + 0.0010016957530751824, + 0.0036230790428817272, + 0.005618764087557793, + 0.018294448032975197, + 0.00756927952170372, + 0.13428497314453125, + 0.021135825663805008, + 0.01801379956305027, + 0.01043672300875187, + 0.15136410295963287, + 0.01616767793893814, + 0.002300656633451581, + 0.002285226248204708, + 0.002362367231398821, + 0.002686992986127734, + 0.0025428819935768843, + 0.0009881012374535203, + 0.002536119893193245, + 0.001418199040926993, + 0.0029399984050542116, + 0.003427618881687522, + 0.00632034707814455, + 0.004039842169731855, + 0.002906159032136202, + 0.0027539946604520082, + 0.005518157035112381, + 0.002743075368925929, + 0.002091234317049384, + 0.0020133471116423607, + 0.01305330265313387, + 0.012830879539251328, + 0.0018021668074652553, + 0.001798540586605668, + 0.00400380278006196, + 0.0016619289526715875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.1603802889585495 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle shot captures a towering skeleton clad in tattered black robes, standing motionless in the center of a vast, eerie, abandoned hall. A cold, moonlit beam streams through a ripped hole in the dome, illuminating its ancient, yellowed bones. A faint, dying crimson glow flickers in its hollow eye sockets. The process of dissolution begins at its feet. As if consumed by an invisible flame, the solid bone structure steadily and silently disintegrates into a dense, inky black vapor. This black vapor spirals upward with an ominous energy, and the skeleton's heavy, tattered black robes dissolve into smoke, billowing in the updraft. As the decomposition process spread upward, its entire body gradually became transparent. Finally, when the last wisp of black steam rose from its slightly tilted head and completely dissipated in the air above the dim hall, the entire scene was left with only the lonely column of light, creating a solemn, sad and final magical atmosphere.", + "seed": 28234, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_3/index_337_seed_28234_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03950344333425164, + "vlm_logprob": -0.03950344333425164, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.03950344333425164, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01719776913523674 + ], + "opposite_logprob_tokens": [ + -5.392197608947754 + ], + "logprob_reward": -0.01719776913523674, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.01719776913523674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06062345206737518, + 0.020992152392864227, + 0.048419829457998276, + 0.04075291380286217, + 0.0991932600736618, + 0.061678532510995865, + 0.10801898688077927, + 0.14574497938156128, + 0.3586721122264862, + 0.05590376257896423 + ], + "credit_map": [ + 0.009380239062011242, + 0.007835290394723415, + 0.010585330426692963, + 0.006746447179466486, + 0.005286861211061478, + 0.005984595511108637, + 0.008376086130738258, + 0.006428598426282406, + 0.0033985539339482784, + 0.0025243149138987064, + 0.002573444740846753, + 0.002841320587322116, + 0.0019299613777548075, + 0.00308589031919837, + 0.002277856692671776, + 0.002360810060054064, + 0.006893131881952286, + 0.006478835362941027, + 0.004069103859364986, + 0.007899072021245956, + 0.0047641717828810215, + 0.007556534372270107, + 0.004522525705397129, + 0.00623645493760705, + 0.006222423631697893, + 0.005175126250833273, + 0.003087913617491722, + 0.0043258629739284515, + 0.007619449403136969, + 0.006576935760676861, + 0.005517900921404362, + 0.002227301709353924, + 0.014146403409540653, + 0.008749006316065788, + 0.014092047698795795, + 0.01181972585618496, + 0.0064595770090818405, + 0.02597752772271633, + 0.008510266430675983, + 0.00943869911134243, + 0.007351262029260397, + 0.0062128109857439995, + 0.006922829896211624, + 0.009569887071847916, + 0.007312651723623276, + 0.00941748172044754, + 0.0065209218300879, + 0.008370688185095787, + 0.010640740394592285, + 0.015741238370537758, + 0.0132064837962389, + 0.013334454037249088, + 0.013750186190009117, + 0.014651786535978317, + 0.011678018607199192, + 0.015016073361039162, + 0.014077452011406422, + 0.020829742774367332, + 0.01437489315867424, + 0.02033046819269657, + 0.02328905649483204, + 0.019377095624804497, + 0.021650010719895363, + 0.011816254816949368, + 0.031331367790699005, + 0.009324912913143635, + 0.13102255761623383, + 0.023682890459895134, + 0.007912046276032925, + 0.019623959437012672, + 0.021373076364398003, + 0.11440130323171616, + 0.0044221412390470505, + 0.0035657959524542093, + 0.009758698754012585, + 0.00815808679908514, + 0.005843743681907654, + 0.005519551690667868, + 0.010335062630474567, + 0.00830068439245224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.01719776913523674 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023318586871027946 + ], + "opposite_logprob_tokens": [ + -3.7733185291290283 + ], + "logprob_reward": -0.023318586871027946, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023318586871027946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03852028399705887, + 0.027996519580483437, + 0.01997309923171997, + 0.01886654645204544, + 0.1244288757443428, + 0.07723207026720047, + 0.11129635572433472, + 0.09200222045183182, + 0.42327409982681274, + 0.06640993803739548 + ], + "credit_map": [ + 0.0051447488367557526, + 0.0048213438130915165, + 0.0072537316009402275, + 0.005117460619658232, + 0.003091071965172887, + 0.0036188694648444653, + 0.005590649321675301, + 0.0038824090734124184, + 0.0038521550595760345, + 0.004072572570294142, + 0.003512658877298236, + 0.004522689152508974, + 0.0025814364198595285, + 0.003519320860505104, + 0.0033312884625047445, + 0.0026043972466140985, + 0.0032536955550312996, + 0.0031670222524553537, + 0.001837854739278555, + 0.0023797573521733284, + 0.0025490745902061462, + 0.002939054276794195, + 0.0015153740532696247, + 0.002331267111003399, + 0.00390700064599514, + 0.0025703045539557934, + 0.0011335774324834347, + 0.0015042184386402369, + 0.0016482369974255562, + 0.004327422007918358, + 0.0023137368261814117, + 0.0014620490837842226, + 0.012082708068192005, + 0.004163633566349745, + 0.040701981633901596, + 0.006703067570924759, + 0.004883599001914263, + 0.04385652765631676, + 0.004981467965990305, + 0.007055898196995258, + 0.009089830331504345, + 0.012569810263812542, + 0.008652493357658386, + 0.008381295949220657, + 0.009479710832238197, + 0.008886310271918774, + 0.006696425843983889, + 0.013476192951202393, + 0.01389153953641653, + 0.015099923126399517, + 0.01325184665620327, + 0.01166678499430418, + 0.017311302945017815, + 0.01615658961236477, + 0.011625420302152634, + 0.012292949482798576, + 0.009247795678675175, + 0.009538823738694191, + 0.01375366561114788, + 0.013577534817159176, + 0.01181564200669527, + 0.011931190267205238, + 0.013627101667225361, + 0.008510468527674675, + 0.036564841866493225, + 0.010335363447666168, + 0.16097187995910645, + 0.02457917109131813, + 0.00981929711997509, + 0.020238220691680908, + 0.018072092905640602, + 0.14269323647022247, + 0.005637981463223696, + 0.003812528681010008, + 0.01094684936106205, + 0.0110352523624897, + 0.006195071153342724, + 0.004135643132030964, + 0.013665024191141129, + 0.01098159234970808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.023318586871027946 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005960783455520868 + ], + "opposite_logprob_tokens": [ + -5.130960941314697 + ], + "logprob_reward": -0.005960783455520868, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005960783455520868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04438948631286621, + 0.03236207365989685, + 0.022585783153772354, + 0.055444639176130295, + 0.05999908596277237, + 0.050225112587213516, + 0.07991882413625717, + 0.20425920188426971, + 0.3016464114189148, + 0.1491694152355194 + ], + "credit_map": [ + 0.005117849446833134, + 0.006385405547916889, + 0.0066988589242100716, + 0.005828015971928835, + 0.004615163896232843, + 0.006013545207679272, + 0.005065407138317823, + 0.004665236454457045, + 0.005125652067363262, + 0.0036019948311150074, + 0.0034630917944014072, + 0.005538708530366421, + 0.0028432586695998907, + 0.005287091713398695, + 0.004350232891738415, + 0.002152044093236327, + 0.004217163659632206, + 0.003525423351675272, + 0.0019394556293264031, + 0.0027524263132363558, + 0.0024156984873116016, + 0.0037512928247451782, + 0.0020136197563260794, + 0.0019707034807652235, + 0.011494262143969536, + 0.006803597789257765, + 0.0039392015896737576, + 0.003247331827878952, + 0.00649342080578208, + 0.014266258105635643, + 0.005795895121991634, + 0.003404672723263502, + 0.007673280313611031, + 0.0036552438978105783, + 0.020103562623262405, + 0.004814078565686941, + 0.0037590432912111282, + 0.013740183785557747, + 0.003498835489153862, + 0.002754863118752837, + 0.005458621308207512, + 0.007127937860786915, + 0.006437070667743683, + 0.0054052770137786865, + 0.007198186591267586, + 0.006081876344978809, + 0.006737856660038233, + 0.0057782880030572414, + 0.008709684945642948, + 0.0121697336435318, + 0.011489124037325382, + 0.008806772530078888, + 0.010427655652165413, + 0.009950847364962101, + 0.008299540728330612, + 0.010065467096865177, + 0.01872936077415943, + 0.026898309588432312, + 0.025600142776966095, + 0.02654418908059597, + 0.041107743978500366, + 0.027659159153699875, + 0.02265343628823757, + 0.015066866762936115, + 0.027938351035118103, + 0.016483383253216743, + 0.09429637342691422, + 0.025630464777350426, + 0.015767009928822517, + 0.01778256520628929, + 0.024422260001301765, + 0.07932596653699875, + 0.013023009523749352, + 0.011759746819734573, + 0.031219730153679848, + 0.015924755483865738, + 0.010848747566342354, + 0.009305724874138832, + 0.02139122039079666, + 0.03569647669792175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.005960783455520868 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809867799282074 + ], + "opposite_logprob_tokens": [ + -3.2880985736846924 + ], + "logprob_reward": -0.03809867799282074, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809867799282074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05025176331400871, + 0.03111252747476101, + 0.05379636958241463, + 0.06759348511695862, + 0.1420184224843979, + 0.07412847876548767, + 0.10121413320302963, + 0.11302418261766434, + 0.292300283908844, + 0.07456045597791672 + ], + "credit_map": [ + 0.0075561110861599445, + 0.004852617625147104, + 0.011871210299432278, + 0.005680099595338106, + 0.0030333115719258785, + 0.004419337026774883, + 0.0059847403317689896, + 0.006854332517832518, + 0.004225564654916525, + 0.0028873630799353123, + 0.004445262253284454, + 0.0046095591969788074, + 0.0026081178802996874, + 0.004753145854920149, + 0.004595675505697727, + 0.002987839514389634, + 0.008178123272955418, + 0.006901022512465715, + 0.005684767384082079, + 0.006260867230594158, + 0.005734928883612156, + 0.00841530878096819, + 0.005438269581645727, + 0.007183084264397621, + 0.010507664643228054, + 0.009091629646718502, + 0.005639106035232544, + 0.005869037471711636, + 0.010055125690996647, + 0.008323157206177711, + 0.012727657333016396, + 0.0053801038302481174, + 0.01693657785654068, + 0.009502258151769638, + 0.043878886848688126, + 0.010212313383817673, + 0.007887679152190685, + 0.03741912171244621, + 0.008654549717903137, + 0.0075270384550094604, + 0.00942248571664095, + 0.011407146230340004, + 0.00786946713924408, + 0.011196229606866837, + 0.00972425751388073, + 0.008517008274793625, + 0.006469158921390772, + 0.00952272117137909, + 0.009663357399404049, + 0.016607161611318588, + 0.013646402396261692, + 0.013309715315699577, + 0.01367698423564434, + 0.010513555258512497, + 0.010561018250882626, + 0.013235935010015965, + 0.012184497900307178, + 0.013909609988331795, + 0.017006296664476395, + 0.015581033192574978, + 0.012305539101362228, + 0.016893837600946426, + 0.013240558095276356, + 0.01190281379967928, + 0.028642302379012108, + 0.005669932346791029, + 0.11086700111627579, + 0.018311547115445137, + 0.008257614448666573, + 0.014851679094135761, + 0.013251245021820068, + 0.09244895726442337, + 0.004813108593225479, + 0.004704681690782309, + 0.010755171068012714, + 0.011832522228360176, + 0.0067559643648564816, + 0.005083227530121803, + 0.019806865602731705, + 0.010808915831148624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.03809867799282074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294139921665192 + ], + "opposite_logprob_tokens": [ + -2.2379415035247803 + ], + "logprob_reward": -0.11294139921665192, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294139921665192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023430366069078445, + 0.03252396360039711, + 0.03638686612248421, + 0.03417650982737541, + 0.2009221911430359, + 0.10049556940793991, + 0.1262427419424057, + 0.18389147520065308, + 0.20660467445850372, + 0.05532563850283623 + ], + "credit_map": [ + 0.003275469643995166, + 0.002385648200288415, + 0.003980916924774647, + 0.0026551613118499517, + 0.0024667109828442335, + 0.0032834894955158234, + 0.0028901314362883568, + 0.0024928366765379906, + 0.004535312298685312, + 0.0046036336570978165, + 0.004583895672112703, + 0.004476322792470455, + 0.0031146127730607986, + 0.004859032575041056, + 0.0025494371075183153, + 0.003801715560257435, + 0.007247465196996927, + 0.005350620485842228, + 0.0023280035238713026, + 0.004596395883709192, + 0.004419949371367693, + 0.006969383917748928, + 0.0021375759970396757, + 0.0033374689519405365, + 0.005576439201831818, + 0.00559043325483799, + 0.00298767420463264, + 0.0033273568842560053, + 0.005633763503283262, + 0.00430793734267354, + 0.003908502869307995, + 0.0028444011695683002, + 0.018709592521190643, + 0.010650726966559887, + 0.05822627991437912, + 0.014656989835202694, + 0.011482249945402145, + 0.06018783152103424, + 0.011630941182374954, + 0.015377574600279331, + 0.011884399689733982, + 0.01598431169986725, + 0.013053652830421925, + 0.010468781925737858, + 0.014105333015322685, + 0.01053586509078741, + 0.009871477261185646, + 0.014591740444302559, + 0.01267371978610754, + 0.019635500386357307, + 0.012972800061106682, + 0.017827432602643967, + 0.021608933806419373, + 0.016005396842956543, + 0.013015412725508213, + 0.012503551319241524, + 0.020596185699105263, + 0.02869129180908203, + 0.02092142216861248, + 0.023218955844640732, + 0.02754203788936138, + 0.026723526418209076, + 0.022807223722338676, + 0.013390838168561459, + 0.014210177585482597, + 0.007334024645388126, + 0.07028723508119583, + 0.014386764727532864, + 0.011166227981448174, + 0.011045712046325207, + 0.012584208510816097, + 0.06559032201766968, + 0.00370587850920856, + 0.003866729326546192, + 0.008389362134039402, + 0.009136402048170567, + 0.0048256441950798035, + 0.002743968041613698, + 0.013770561665296555, + 0.008887090720236301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.11294139921665192 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle shot captures a towering skeleton clad in tattered black robes, standing motionless in the center of a vast, eerie, abandoned hall. A cold, moonlit beam streams through a ripped hole in the dome, illuminating its ancient, yellowed bones. A faint, dying crimson glow flickers in its hollow eye sockets. The process of dissolution begins at its feet. As if consumed by an invisible flame, the solid bone structure steadily and silently disintegrates into a dense, inky black vapor. This black vapor spirals upward with an ominous energy, and the skeleton's heavy, tattered black robes dissolve into smoke, billowing in the updraft. As the decomposition process spread upward, its entire body gradually became transparent. Finally, when the last wisp of black steam rose from its slightly tilted head and completely dissipated in the air above the dim hall, the entire scene was left with only the lonely column of light, creating a solemn, sad and final magical atmosphere.", + "seed": 28235, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_3/index_337_seed_28235_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05492917913943529, + "vlm_logprob": -0.05492917913943529, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.05492917913943529, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015370487235486507 + ], + "opposite_logprob_tokens": [ + -5.0153703689575195 + ], + "logprob_reward": -0.015370487235486507, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.015370487235486507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14837156236171722, + 0.18605446815490723, + 0.08684559911489487, + 0.15910932421684265, + 0.06970538198947906, + 0.05833035707473755, + 0.08997545391321182, + 0.075768381357193, + 0.0798291265964508, + 0.04601031169295311 + ], + "credit_map": [ + 0.016870779916644096, + 0.013203958980739117, + 0.022070541977882385, + 0.019973397254943848, + 0.009575748816132545, + 0.024964770302176476, + 0.013798519037663937, + 0.027913857251405716, + 0.0194650050252676, + 0.013370731845498085, + 0.02793561853468418, + 0.02519289217889309, + 0.01716652698814869, + 0.03130029886960983, + 0.015209191478788853, + 0.03641420975327492, + 0.010133140720427036, + 0.007935697212815285, + 0.017213033512234688, + 0.015356446616351604, + 0.009993918240070343, + 0.010684472508728504, + 0.005682986229658127, + 0.009845901280641556, + 0.028233487159013748, + 0.006499321665614843, + 0.03472818061709404, + 0.014722028747200966, + 0.010041407309472561, + 0.021767396479845047, + 0.01731693372130394, + 0.025800568982958794, + 0.008355142548680305, + 0.006213086191564798, + 0.01755555532872677, + 0.005148949567228556, + 0.0071633863262832165, + 0.005611248314380646, + 0.005211756564676762, + 0.014446256682276726, + 0.006402877625077963, + 0.009151019155979156, + 0.005401879549026489, + 0.01100828591734171, + 0.007417155895382166, + 0.007658443879336119, + 0.006573802325874567, + 0.004716894123703241, + 0.01058134064078331, + 0.01131333876401186, + 0.009269000962376595, + 0.013833479955792427, + 0.006995042320340872, + 0.008160021156072617, + 0.007236212491989136, + 0.022587014362215996, + 0.009738198481500149, + 0.0064164879731833935, + 0.00784100592136383, + 0.008279168047010899, + 0.012765050865709782, + 0.010304314084351063, + 0.010417181998491287, + 0.010006971657276154, + 0.00818696804344654, + 0.011908693239092827, + 0.011336944997310638, + 0.00792798213660717, + 0.010038241744041443, + 0.007763173431158066, + 0.009856845252215862, + 0.012810281477868557, + 0.007933284156024456, + 0.00937847513705492, + 0.004647118039429188, + 0.00376010499894619, + 0.006632384844124317, + 0.0064547741785645485, + 0.003560007084161043, + 0.0036441637203097343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.015370487235486507 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905582129955292 + ], + "opposite_logprob_tokens": [ + -2.4640557765960693 + ], + "logprob_reward": -0.08905582129955292, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905582129955292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03795214369893074, + 0.07899683713912964, + 0.030085846781730652, + 0.07509820908308029, + 0.023176785558462143, + 0.03800203278660774, + 0.3125171959400177, + 0.05342315509915352, + 0.12816835939884186, + 0.2225794494152069 + ], + "credit_map": [ + 0.006532709114253521, + 0.0025371662341058254, + 0.004180144052952528, + 0.005026767496019602, + 0.004700429271906614, + 0.005347713362425566, + 0.0035750102251768112, + 0.006052199751138687, + 0.007603430189192295, + 0.005150339566171169, + 0.008395342156291008, + 0.014528254978358746, + 0.005126597825437784, + 0.01277356967329979, + 0.006688101217150688, + 0.018731193616986275, + 0.0040283724665641785, + 0.00364521611481905, + 0.004545418545603752, + 0.0051140631549060345, + 0.0037974873557686806, + 0.003768518101423979, + 0.0019597632344812155, + 0.003227004548534751, + 0.012180310674011707, + 0.0028789727948606014, + 0.016985082998871803, + 0.005338679067790508, + 0.005580282770097256, + 0.00977679155766964, + 0.008447632193565369, + 0.013910455629229546, + 0.0019324987661093473, + 0.0016712271608412266, + 0.004698798526078463, + 0.0016918496694415808, + 0.0026735446881502867, + 0.0017201894661411643, + 0.003083028830587864, + 0.005705645773559809, + 0.00428418256342411, + 0.005847673397511244, + 0.003485636319965124, + 0.006043809000402689, + 0.005454752594232559, + 0.00595081178471446, + 0.00375025044195354, + 0.003184914356097579, + 0.046031489968299866, + 0.019736846908926964, + 0.01991710253059864, + 0.05610695481300354, + 0.010994400829076767, + 0.025234686210751534, + 0.021338846534490585, + 0.11315684765577316, + 0.005287394858896732, + 0.005550734233111143, + 0.006290277931839228, + 0.005715271923691034, + 0.009983886033296585, + 0.007791836746037006, + 0.006960474420338869, + 0.0058432756923139095, + 0.014925537630915642, + 0.03370874747633934, + 0.009983102791011333, + 0.010485751554369926, + 0.015530452132225037, + 0.01700868457555771, + 0.012829354032874107, + 0.013696723617613316, + 0.03612585738301277, + 0.04236755892634392, + 0.03661360964179039, + 0.021189847961068153, + 0.023395253345370293, + 0.027188679203391075, + 0.01847102865576744, + 0.01722760871052742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.08905582129955292 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645249530673027 + ], + "opposite_logprob_tokens": [ + -4.758645057678223 + ], + "logprob_reward": -0.008645249530673027, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645249530673027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09125664085149765, + 0.10414601862430573, + 0.060414738953113556, + 0.12750260531902313, + 0.07953176647424698, + 0.05090886354446411, + 0.11431682109832764, + 0.0349988117814064, + 0.1574084311723709, + 0.1795153170824051 + ], + "credit_map": [ + 0.0122189586982131, + 0.0068546743132174015, + 0.010797924362123013, + 0.008539478294551373, + 0.011989898048341274, + 0.014341086149215698, + 0.01126829907298088, + 0.015246322378516197, + 0.011433915235102177, + 0.007179488893598318, + 0.01788209192454815, + 0.011615284718573093, + 0.00990015733987093, + 0.015917791053652763, + 0.010409491136670113, + 0.01980779506266117, + 0.008572759106755257, + 0.0038994522765278816, + 0.012629993259906769, + 0.007491846103221178, + 0.00648861937224865, + 0.008671838790178299, + 0.005316784605383873, + 0.0073434473015367985, + 0.03100704587996006, + 0.004550138954073191, + 0.026057424023747444, + 0.006968329194933176, + 0.009960873052477837, + 0.02222379855811596, + 0.00959794782102108, + 0.01713704504072666, + 0.00869796983897686, + 0.004624244756996632, + 0.023989781737327576, + 0.004832811653614044, + 0.005937353242188692, + 0.0048787700943648815, + 0.004879300016909838, + 0.02169153280556202, + 0.004809478763490915, + 0.008102457039058208, + 0.006034988909959793, + 0.008636901155114174, + 0.006439573597162962, + 0.004675367381423712, + 0.007992465980350971, + 0.004217627923935652, + 0.01584811322391033, + 0.00828468706458807, + 0.007000827230513096, + 0.019435163587331772, + 0.006364427041262388, + 0.0090045016258955, + 0.008627941831946373, + 0.03975115716457367, + 0.0046412707306444645, + 0.0024961940944194794, + 0.004219702910631895, + 0.0040904381312429905, + 0.004971735179424286, + 0.0037524786312133074, + 0.006787997670471668, + 0.004038992337882519, + 0.017583834007382393, + 0.03637564182281494, + 0.024585340172052383, + 0.013995928689837456, + 0.016695113852620125, + 0.020624905824661255, + 0.012048845179378986, + 0.0154988132417202, + 0.03198005259037018, + 0.0415710024535656, + 0.024892941117286682, + 0.015907976776361465, + 0.017959831282496452, + 0.02248384989798069, + 0.011457555927336216, + 0.013262116350233555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.008645249530673027 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048649441450834274 + ], + "opposite_logprob_tokens": [ + -3.048649549484253 + ], + "logprob_reward": -0.048649441450834274, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048649441450834274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13244518637657166, + 0.08437386900186539, + 0.1228746697306633, + 0.0679955706000328, + 0.10234174132347107, + 0.09589030593633652, + 0.19822759926319122, + 0.05843989923596382, + 0.07683218270540237, + 0.06057897210121155 + ], + "credit_map": [ + 0.01583523117005825, + 0.012394813820719719, + 0.01762271858751774, + 0.01461231242865324, + 0.009622030891478062, + 0.02192710153758526, + 0.014927810989320278, + 0.02550315670669079, + 0.009385129436850548, + 0.007312598172575235, + 0.012530993670225143, + 0.013531051576137543, + 0.00826657097786665, + 0.013102976605296135, + 0.007481606677174568, + 0.012762943282723427, + 0.01470686960965395, + 0.01060513686388731, + 0.023951075971126556, + 0.013673706911504269, + 0.015415086410939693, + 0.014828147366642952, + 0.018624361604452133, + 0.011070288717746735, + 0.01047611329704523, + 0.003949265461415052, + 0.012158744037151337, + 0.004592237062752247, + 0.005800120998173952, + 0.010843700729310513, + 0.007241389248520136, + 0.01293400302529335, + 0.009991960600018501, + 0.007140490226447582, + 0.029718736186623573, + 0.007928500883281231, + 0.0089305704459548, + 0.007290627807378769, + 0.00747006805613637, + 0.02387077920138836, + 0.00830758921802044, + 0.010257607325911522, + 0.019386503845453262, + 0.013477427884936333, + 0.00899998564273119, + 0.012492338195443153, + 0.00897887907922268, + 0.013989968225359917, + 0.02790135145187378, + 0.017658760771155357, + 0.013345773331820965, + 0.033139538019895554, + 0.0105325011536479, + 0.016881413757801056, + 0.01622030884027481, + 0.0625479593873024, + 0.0053842258639633656, + 0.0058250827714800835, + 0.00740505987778306, + 0.008366633206605911, + 0.009957964532077312, + 0.006050966680049896, + 0.007870840840041637, + 0.007579123601317406, + 0.007959925569593906, + 0.011511716991662979, + 0.007832909934222698, + 0.00915092509239912, + 0.012575061991810799, + 0.008428051136434078, + 0.01185978576540947, + 0.007513805292546749, + 0.01032213494181633, + 0.01093615498393774, + 0.01133993361145258, + 0.005480582360178232, + 0.00544749153777957, + 0.007458163425326347, + 0.00398121727630496, + 0.005613294430077076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.048649441450834274 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle showing a towering skeleton in tattered black robes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skeleton standing motionless in the center of a vast, eerie, abandoned hall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cold, moonlit beam stream through a ripped hole in the dome to illuminate the skeleton's yellowed bones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the skeleton disintegrate into black vapor starting from its feet and spreading upward until only the light beam remains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the skeleton's head slightly tilted during the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292489618062973 + ], + "opposite_logprob_tokens": [ + -2.237924814224243 + ], + "logprob_reward": -0.11292489618062973, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292489618062973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10129106789827347, + 0.08368537575006485, + 0.05026833713054657, + 0.09775020182132721, + 0.035352278500795364, + 0.03186928480863571, + 0.330349862575531, + 0.12084563076496124, + 0.06692866235971451, + 0.08165931701660156 + ], + "credit_map": [ + 0.01294528879225254, + 0.008286821655929089, + 0.009492583572864532, + 0.010087798349559307, + 0.01095639169216156, + 0.014942820183932781, + 0.013963405974209309, + 0.020615950226783752, + 0.007838329300284386, + 0.004217496141791344, + 0.011611889116466045, + 0.012047851458191872, + 0.006067041307687759, + 0.013177113607525826, + 0.007500796113163233, + 0.0212248545140028, + 0.00477024307474494, + 0.004284264054149389, + 0.009509383700788021, + 0.007162067107856274, + 0.004764461424201727, + 0.006480153650045395, + 0.008513963781297207, + 0.00478379474952817, + 0.02005564235150814, + 0.0030059670098125935, + 0.01790081523358822, + 0.005668384954333305, + 0.007260477636009455, + 0.019521519541740417, + 0.010607222095131874, + 0.013730157166719437, + 0.003737300867214799, + 0.0032880736980587244, + 0.007986214943230152, + 0.002837956650182605, + 0.004000517539680004, + 0.003059946233406663, + 0.004149244166910648, + 0.006293018814176321, + 0.0037438031286001205, + 0.004583785310387611, + 0.0029313534032553434, + 0.004854791332036257, + 0.006547677796334028, + 0.003111116820946336, + 0.003539940109476447, + 0.0025568122509866953, + 0.04146812856197357, + 0.02134978212416172, + 0.0186904426664114, + 0.057853423058986664, + 0.021355783566832542, + 0.02787582203745842, + 0.028692996129393578, + 0.11306346952915192, + 0.012928670272231102, + 0.008147557266056538, + 0.008654973469674587, + 0.013567794114351273, + 0.01715122163295746, + 0.011536656878888607, + 0.02236679010093212, + 0.026491951197385788, + 0.00783283170312643, + 0.009049956686794758, + 0.0064675211906433105, + 0.007428546901792288, + 0.011045461520552635, + 0.007238651625812054, + 0.012190516106784344, + 0.005675165913999081, + 0.010740394704043865, + 0.011441153474152088, + 0.011577252298593521, + 0.008332811295986176, + 0.011366808786988258, + 0.013717839494347572, + 0.006733729969710112, + 0.007749308831989765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.11292489618062973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_4/index_tensor([2949, 2949])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_4/index_tensor([2949, 2949])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5c13920f7ffdac3adb437d94b0a648fcc88e89bf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_4/index_tensor([2949, 2949])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera, held perfectly still in a high-angle macro shot, captures a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface that fills the frame. Under bright, unfocused studio lighting, the liquid, with its deep, oily sheen, begins to slowly and relentlessly seep across the entire surface, its leading edge forming a perfect, glossy curve against the stark white. Without warning, the smooth surface of the liquid trembles before suddenly erupting upward into a dense forest of sharp, needle-like spikes. These structures are not static; they undulate and sway mesmerizingly in synchronized, wave-like motion, their shimmering tips dancing and reorganizing as if responding to the currents of a powerful, ever-shifting magnetic force unseen beyond the frame. The stark, high-contrast visuals and the liquid's controlled yet chaotic movement create a hypnotic and slightly unsettling atmosphere, underscoring the alien, organic quality of its beautiful yet sinister transformations.", + "seed": 12461, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_4/index_2949_seed_12461_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.174718544119969, + "vlm_logprob": -0.174718544119969, + "vlm_yes_no_margin": 2.5250000238418577, + "vlm_token_credit": -0.174718544119969, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2683805525302887 + ], + "opposite_logprob_tokens": [ + -1.5183805227279663 + ], + "logprob_reward": -0.2683805525302887, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2683805525302887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07991261780261993, + 0.04091149941086769, + 0.11339181661605835, + 0.13163645565509796, + 0.1934480369091034, + 0.06053026765584946, + 0.0829005241394043, + 0.10732614994049072, + 0.14058741927146912, + 0.04935522377490997 + ], + "credit_map": [ + 0.0079153748229146, + 0.010133066214621067, + 0.00993375200778246, + 0.012839754112064838, + 0.007192780263721943, + 0.011610014364123344, + 0.00956002902239561, + 0.010727851651608944, + 0.0043800906278193, + 0.004180994350463152, + 0.005770108196884394, + 0.008061225526034832, + 0.0030843045096844435, + 0.004913203418254852, + 0.0048289368860423565, + 0.005692632403224707, + 0.011774803511798382, + 0.020501744002103806, + 0.010170765221118927, + 0.00977136567234993, + 0.010889044962823391, + 0.010595922358334064, + 0.016229070723056793, + 0.02345910295844078, + 0.014645383693277836, + 0.01928851753473282, + 0.016454555094242096, + 0.015332085080444813, + 0.021908527240157127, + 0.013019324280321598, + 0.01726817525923252, + 0.013719876296818256, + 0.022334514185786247, + 0.023978285491466522, + 0.02578798308968544, + 0.0331047959625721, + 0.024986693635582924, + 0.023146428167819977, + 0.019646700471639633, + 0.020462628453969955, + 0.005521989893168211, + 0.00693303719162941, + 0.005388072691857815, + 0.0065860869362950325, + 0.006846441421657801, + 0.009095700457692146, + 0.01146398764103651, + 0.008694947697222233, + 0.012065225280821323, + 0.014897509478032589, + 0.0050166696310043335, + 0.007232472766190767, + 0.011546083725988865, + 0.008765208534896374, + 0.010376028716564178, + 0.01300132367759943, + 0.02061348408460617, + 0.01599775440990925, + 0.011632349342107773, + 0.008661928586661816, + 0.012596804648637772, + 0.009704722091555595, + 0.012599076144397259, + 0.015520020388066769, + 0.01457530539482832, + 0.03187816962599754, + 0.010527758859097958, + 0.021661238744854927, + 0.011647761799395084, + 0.022391024976968765, + 0.012507462874054909, + 0.015398700721561909, + 0.009970230981707573, + 0.007244687993079424, + 0.0062794154509902, + 0.0062027787789702415, + 0.005234909243881702, + 0.003537662560120225, + 0.00409688288345933, + 0.006788650061935186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.2683805525302887 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813335955142975 + ], + "opposite_logprob_tokens": [ + -1.406333565711975 + ], + "logprob_reward": -0.2813335955142975, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813335955142975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12128246575593948, + 0.04200199246406555, + 0.131304532289505, + 0.13540144264698029, + 0.22833257913589478, + 0.1036374494433403, + 0.12596771121025085, + 0.06404563784599304, + 0.026649516075849533, + 0.021376686170697212 + ], + "credit_map": [ + 0.012109600938856602, + 0.016456888988614082, + 0.013005714863538742, + 0.01447705551981926, + 0.015989040955901146, + 0.019191766157746315, + 0.013200617395341396, + 0.016851790249347687, + 0.0054503390565514565, + 0.004744944628328085, + 0.004217728041112423, + 0.008028642274439335, + 0.002988160587847233, + 0.006192388478666544, + 0.004272219724953175, + 0.006107574328780174, + 0.014171946793794632, + 0.03413470461964607, + 0.005230807699263096, + 0.00639681238681078, + 0.011117554269731045, + 0.012861660681664944, + 0.00640333816409111, + 0.04098772257566452, + 0.015096821822226048, + 0.024982307106256485, + 0.014212354086339474, + 0.013878299854695797, + 0.017361152917146683, + 0.016959434375166893, + 0.014792346395552158, + 0.01811874285340309, + 0.029581116512417793, + 0.03733238950371742, + 0.027926621958613396, + 0.034887682646512985, + 0.018637461587786674, + 0.024999279528856277, + 0.02254859358072281, + 0.03241945058107376, + 0.008373663760721684, + 0.016957230865955353, + 0.008222753182053566, + 0.010073569603264332, + 0.012189486064016819, + 0.017503120005130768, + 0.013928518630564213, + 0.016389118507504463, + 0.01674865186214447, + 0.03415447100996971, + 0.007340217474848032, + 0.008769465610384941, + 0.014841960743069649, + 0.015202482230961323, + 0.01724436692893505, + 0.011666103266179562, + 0.009760563261806965, + 0.009981602430343628, + 0.006152636371552944, + 0.006807804107666016, + 0.007418092340230942, + 0.004972006194293499, + 0.009380989708006382, + 0.00957194808870554, + 0.002761933719739318, + 0.0026991567574441433, + 0.0026828597765415907, + 0.003505207598209381, + 0.003183349035680294, + 0.0042503527365624905, + 0.003886341815814376, + 0.003680317895486951, + 0.0021860452834516764, + 0.003228318877518177, + 0.0030172010883688927, + 0.003247426124289632, + 0.003050270490348339, + 0.002302475506439805, + 0.002265850082039833, + 0.0020791003480553627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.2813335955142975 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279642701148987 + ], + "opposite_logprob_tokens": [ + -2.017796516418457 + ], + "logprob_reward": -0.14279642701148987, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14279642701148987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16419392824172974, + 0.062326788902282715, + 0.12331664562225342, + 0.10422085225582123, + 0.07148751616477966, + 0.07139241695404053, + 0.17858268320560455, + 0.076381154358387, + 0.1047857329249382, + 0.04331236332654953 + ], + "credit_map": [ + 0.01364874467253685, + 0.0209573432803154, + 0.01484702154994011, + 0.023821841925382614, + 0.020990760996937752, + 0.03031645156443119, + 0.01693490892648697, + 0.02267683856189251, + 0.006498835049569607, + 0.007097345311194658, + 0.006201989948749542, + 0.008195516653358936, + 0.006150455214083195, + 0.010530442930758, + 0.006930261384695768, + 0.01072194054722786, + 0.012123635970056057, + 0.024034462869167328, + 0.015774564817547798, + 0.012926357798278332, + 0.011071300134062767, + 0.00884885061532259, + 0.014582554809749126, + 0.023954909294843674, + 0.011226130649447441, + 0.021225659176707268, + 0.010145553387701511, + 0.01290228869765997, + 0.01607632450759411, + 0.01287426520138979, + 0.008932908065617085, + 0.010837709531188011, + 0.00843001902103424, + 0.01042137574404478, + 0.008385957218706608, + 0.011896561831235886, + 0.008506589569151402, + 0.007631505839526653, + 0.007891770452260971, + 0.008323729038238525, + 0.00861276127398014, + 0.011001868173480034, + 0.012959658168256283, + 0.01110039558261633, + 0.007956752553582191, + 0.005053731147199869, + 0.009129964746534824, + 0.005577270872890949, + 0.02685321681201458, + 0.04094674065709114, + 0.011340775527060032, + 0.015616619028151035, + 0.02435234934091568, + 0.0177558995783329, + 0.020679766312241554, + 0.021037286147475243, + 0.015601515769958496, + 0.010348140262067318, + 0.009938814677298069, + 0.008053997531533241, + 0.008246988989412785, + 0.005961187183856964, + 0.008479496464133263, + 0.009751005098223686, + 0.011246507056057453, + 0.023766882717609406, + 0.010567541234195232, + 0.010834846645593643, + 0.008961845189332962, + 0.01826987788081169, + 0.008601440116763115, + 0.012536784633994102, + 0.006903391797095537, + 0.006015833467245102, + 0.011057236231863499, + 0.0036853866185992956, + 0.004570486024022102, + 0.003529135137796402, + 0.0030162360053509474, + 0.00453465199097991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.14279642701148987 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011823337990790606 + ], + "opposite_logprob_tokens": [ + -6.751182556152344 + ], + "logprob_reward": -0.0011823337990790606, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011823337990790606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14393390715122223, + 0.04688326269388199, + 0.1150323897600174, + 0.09259048849344254, + 0.12291514873504639, + 0.08274412155151367, + 0.22362032532691956, + 0.08270857483148575, + 0.04701503366231918, + 0.04255678877234459 + ], + "credit_map": [ + 0.01451956294476986, + 0.019197914749383926, + 0.01816757209599018, + 0.017429405823349953, + 0.0201265811920166, + 0.020530512556433678, + 0.014369885437190533, + 0.01959247700870037, + 0.0031332054641097784, + 0.0027690951246768236, + 0.00879734568297863, + 0.012352030724287033, + 0.0025884988717734814, + 0.004089874681085348, + 0.006463098339736462, + 0.006690117530524731, + 0.014722298830747604, + 0.018528811633586884, + 0.013737152330577374, + 0.011577600613236427, + 0.014534622430801392, + 0.011020191013813019, + 0.014203226193785667, + 0.016708478331565857, + 0.01094258762896061, + 0.01769300363957882, + 0.009195851162075996, + 0.010182340629398823, + 0.01170006487518549, + 0.010589632205665112, + 0.010578596964478493, + 0.011708414182066917, + 0.016855932772159576, + 0.01811998151242733, + 0.015299041755497456, + 0.017301207408308983, + 0.012540188618004322, + 0.013519549742341042, + 0.011906747706234455, + 0.017372502014040947, + 0.010114955715835094, + 0.010145680978894234, + 0.012590471655130386, + 0.0090620256960392, + 0.0071343849413096905, + 0.005087218247354031, + 0.021816782653331757, + 0.0067926025949418545, + 0.027945710346102715, + 0.04666776955127716, + 0.013218395411968231, + 0.020020371302962303, + 0.028238337486982346, + 0.019510488957166672, + 0.0341988168656826, + 0.03382043167948723, + 0.016383636742830276, + 0.01277890708297491, + 0.007850383408367634, + 0.006750981789082289, + 0.011652886867523193, + 0.007318397518247366, + 0.010611114092171192, + 0.009362270124256611, + 0.004825742915272713, + 0.009274976328015327, + 0.00474834768101573, + 0.006438847631216049, + 0.004763433709740639, + 0.007515565492212772, + 0.0038269739598035812, + 0.005621142219752073, + 0.0064448099583387375, + 0.00812240969389677, + 0.005741678643971682, + 0.004667502827942371, + 0.005060205236077309, + 0.0037049814127385616, + 0.0030536821577697992, + 0.005761518143117428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.0011823337990790606 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989981174468994 + ], + "opposite_logprob_tokens": [ + -1.80489981174469 + ], + "logprob_reward": -0.17989981174468994, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989981174468994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08206024765968323, + 0.06484778225421906, + 0.06708300858736038, + 0.12514449656009674, + 0.3163456916809082, + 0.09079957008361816, + 0.13957472145557404, + 0.050318025052547455, + 0.03923968970775604, + 0.024586766958236694 + ], + "credit_map": [ + 0.008198867551982403, + 0.008854999206960201, + 0.009349197149276733, + 0.009640979580581188, + 0.008468625135719776, + 0.01483895257115364, + 0.010644564405083656, + 0.012064055539667606, + 0.006248610559850931, + 0.007324543781578541, + 0.006286749616265297, + 0.013257162645459175, + 0.0047300392761826515, + 0.0072718216106295586, + 0.012406702153384686, + 0.007322146091610193, + 0.010508768260478973, + 0.01281135343015194, + 0.00404618913307786, + 0.00634019635617733, + 0.008103513158857822, + 0.007009769789874554, + 0.009026662446558475, + 0.009236550889909267, + 0.013271531090140343, + 0.02098286710679531, + 0.01430718693882227, + 0.012167725712060928, + 0.019839806482195854, + 0.017596976831555367, + 0.014603577554225922, + 0.012374808080494404, + 0.0346740260720253, + 0.06416443735361099, + 0.02580266073346138, + 0.03936666250228882, + 0.029915018007159233, + 0.03612256795167923, + 0.03287293016910553, + 0.05342736467719078, + 0.009641490876674652, + 0.009177323430776596, + 0.011763904243707657, + 0.008986388333141804, + 0.007060792297124863, + 0.011567682027816772, + 0.0211507398635149, + 0.011451235972344875, + 0.01684550754725933, + 0.03535321354866028, + 0.011469496414065361, + 0.01218412071466446, + 0.01625099778175354, + 0.016713986173272133, + 0.02011726051568985, + 0.010640131309628487, + 0.007739562541246414, + 0.008049349300563335, + 0.0086302375420928, + 0.005136632826179266, + 0.005807339213788509, + 0.0033390712924301624, + 0.006248388439416885, + 0.005367437843233347, + 0.004070184659212828, + 0.0035611584316939116, + 0.004740857984870672, + 0.006168559193611145, + 0.003680552588775754, + 0.006081237457692623, + 0.005055999383330345, + 0.005881136748939753, + 0.004349933937191963, + 0.004908416420221329, + 0.0031868903897702694, + 0.002670014975592494, + 0.002162704011425376, + 0.002113721799105406, + 0.001597938477061689, + 0.0035971461329609156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.17989981174468994 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera, held perfectly still in a high-angle macro shot, captures a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface that fills the frame. Under bright, unfocused studio lighting, the liquid, with its deep, oily sheen, begins to slowly and relentlessly seep across the entire surface, its leading edge forming a perfect, glossy curve against the stark white. Without warning, the smooth surface of the liquid trembles before suddenly erupting upward into a dense forest of sharp, needle-like spikes. These structures are not static; they undulate and sway mesmerizingly in synchronized, wave-like motion, their shimmering tips dancing and reorganizing as if responding to the currents of a powerful, ever-shifting magnetic force unseen beyond the frame. The stark, high-contrast visuals and the liquid's controlled yet chaotic movement create a hypnotic and slightly unsettling atmosphere, underscoring the alien, organic quality of its beautiful yet sinister transformations.", + "seed": 12462, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_4/index_2949_seed_12462_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09240001819562167, + "vlm_logprob": -0.09240001819562167, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.09240001819562167, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1265602707862854 + ], + "opposite_logprob_tokens": [ + -2.2515602111816406 + ], + "logprob_reward": -0.1265602707862854, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1265602707862854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045636359602212906, + 0.026190614327788353, + 0.03316666558384895, + 0.03710620477795601, + 0.0706564337015152, + 0.06698090583086014, + 0.019958315417170525, + 0.11610853672027588, + 0.03320366144180298, + 0.5509922504425049 + ], + "credit_map": [ + 0.0030924840830266476, + 0.009284166619181633, + 0.0034531198907643557, + 0.009562588296830654, + 0.007475682534277439, + 0.0034936692100018263, + 0.0061124954372644424, + 0.003162155859172344, + 0.0021452894434332848, + 0.004664784763008356, + 0.002655200194567442, + 0.006990478374063969, + 0.003178223269060254, + 0.001519284094683826, + 0.003572150133550167, + 0.0014652055688202381, + 0.0031278065871447325, + 0.0040829358622431755, + 0.0032990891486406326, + 0.007119050249457359, + 0.005711840931326151, + 0.002013595076277852, + 0.00553803239017725, + 0.0022743153385818005, + 0.0040677450597286224, + 0.006150401197373867, + 0.003656171727925539, + 0.008802015334367752, + 0.0035437080077826977, + 0.002426442224532366, + 0.005457966588437557, + 0.003001757664605975, + 0.004904052708297968, + 0.012488150969147682, + 0.005539408884942532, + 0.007732496131211519, + 0.024282105267047882, + 0.004762487020343542, + 0.007243584841489792, + 0.0037041502073407173, + 0.006247056648135185, + 0.016984310001134872, + 0.00673010153695941, + 0.013257257640361786, + 0.007359087001532316, + 0.0066825030371546745, + 0.005600952077656984, + 0.0041196392849087715, + 0.00234786095097661, + 0.002727954648435116, + 0.0016712682554498315, + 0.003332399530336261, + 0.003368380945175886, + 0.0014211043016985059, + 0.0033748613204807043, + 0.0017144882585853338, + 0.008391136303544044, + 0.03576035425066948, + 0.006460691802203655, + 0.005029253661632538, + 0.04025585949420929, + 0.007277798373252153, + 0.00846841186285019, + 0.004465045407414436, + 0.002623521024361253, + 0.005153419449925423, + 0.0029840823262929916, + 0.003816890297457576, + 0.005579597316682339, + 0.0021265314426273108, + 0.007822396233677864, + 0.003097228240221739, + 0.02951844595372677, + 0.09592423588037491, + 0.043892208486795425, + 0.13976161181926727, + 0.1086650937795639, + 0.016760773956775665, + 0.10103072971105576, + 0.015439234673976898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.1265602707862854 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033742014318704605 + ], + "opposite_logprob_tokens": [ + -3.4087419509887695 + ], + "logprob_reward": -0.033742014318704605, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033742014318704605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06007035821676254, + 0.043238669633865356, + 0.07062073796987534, + 0.15123574435710907, + 0.313494473695755, + 0.10104777663946152, + 0.024217555299401283, + 0.13168340921401978, + 0.019968770444393158, + 0.0844225063920021 + ], + "credit_map": [ + 0.0049453359097242355, + 0.011140871793031693, + 0.004994550719857216, + 0.011705520562827587, + 0.009057859890162945, + 0.005108376033604145, + 0.005819268524646759, + 0.007298572920262814, + 0.0033237726893275976, + 0.007689637131989002, + 0.004346594680100679, + 0.010511456988751888, + 0.00624167500063777, + 0.0027721296064555645, + 0.005758993793278933, + 0.0025944069493561983, + 0.008022993803024292, + 0.0108365248888731, + 0.006749160122126341, + 0.010957845486700535, + 0.01018484216183424, + 0.005992847494781017, + 0.011717524379491806, + 0.006158990319818258, + 0.013841148465871811, + 0.027000293135643005, + 0.013114255852997303, + 0.0459502674639225, + 0.01256124209612608, + 0.011623989790678024, + 0.01771276257932186, + 0.009431783109903336, + 0.019517527893185616, + 0.04135287180542946, + 0.013708694837987423, + 0.06384850293397903, + 0.11195597797632217, + 0.021944470703601837, + 0.03105773963034153, + 0.01010867953300476, + 0.009047973901033401, + 0.030183233320713043, + 0.010065953247249126, + 0.020468762144446373, + 0.009553633630275726, + 0.008557870052754879, + 0.0073336344212293625, + 0.005836710799485445, + 0.00248228688724339, + 0.006710073910653591, + 0.0016175201162695885, + 0.0030405663419514894, + 0.0034039760939776897, + 0.0017313167918473482, + 0.0033872281201183796, + 0.0018445837777107954, + 0.007132645230740309, + 0.04324888810515404, + 0.003749655093997717, + 0.0036600150633603334, + 0.050364214926958084, + 0.010203625075519085, + 0.008711837232112885, + 0.00461251987144351, + 0.0020419051870703697, + 0.003352477215230465, + 0.001548627857118845, + 0.002458353294059634, + 0.0037102701608091593, + 0.0013375323032960296, + 0.004254534374922514, + 0.001265067607164383, + 0.0040429639630019665, + 0.01864059641957283, + 0.005771494936197996, + 0.016746886074543, + 0.021470965817570686, + 0.0028061745688319206, + 0.012319372966885567, + 0.0026240497827529907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.033742014318704605 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521020174026489 + ], + "opposite_logprob_tokens": [ + -1.502102017402649 + ], + "logprob_reward": -0.2521020174026489, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521020174026489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07551540434360504, + 0.09196554124355316, + 0.13264471292495728, + 0.0651194229722023, + 0.04216068983078003, + 0.10272380709648132, + 0.06719408929347992, + 0.04559243470430374, + 0.08369099348783493, + 0.2933928966522217 + ], + "credit_map": [ + 0.005928501021116972, + 0.01600019633769989, + 0.00624534348025918, + 0.013670708052814007, + 0.01177437137812376, + 0.006820306181907654, + 0.007475920952856541, + 0.007600057870149612, + 0.007228301838040352, + 0.009261318482458591, + 0.008480190299451351, + 0.028336381539702415, + 0.006970737595111132, + 0.008289217017591, + 0.00938024278730154, + 0.014019145630300045, + 0.00992418359965086, + 0.02131650783121586, + 0.0143687529489398, + 0.023747200146317482, + 0.016380025073885918, + 0.007694237865507603, + 0.031148914247751236, + 0.008064894936978817, + 0.006136667914688587, + 0.012387702241539955, + 0.006957608740776777, + 0.017145846039056778, + 0.005298508796840906, + 0.0034944643266499043, + 0.010564029216766357, + 0.0031346017494797707, + 0.0030628545209765434, + 0.0077345590107142925, + 0.0028304033912718296, + 0.006604787427932024, + 0.013481657020747662, + 0.0021629873663187027, + 0.004184572491794825, + 0.0020988688338547945, + 0.011672858148813248, + 0.020810024812817574, + 0.007557579781860113, + 0.018564296886324883, + 0.014521456323564053, + 0.00932954903692007, + 0.012749063782393932, + 0.007518978323787451, + 0.007664606440812349, + 0.015120119787752628, + 0.006527316756546497, + 0.006965877488255501, + 0.008879336528480053, + 0.004871245939284563, + 0.008481276221573353, + 0.008684307336807251, + 0.003043357515707612, + 0.012375560589134693, + 0.0030442867428064346, + 0.0034941211342811584, + 0.013740003108978271, + 0.0031882920302450657, + 0.004478291608393192, + 0.0022285189479589462, + 0.006464012432843447, + 0.013644564896821976, + 0.005998351611196995, + 0.012652544304728508, + 0.01147988997399807, + 0.005551124457269907, + 0.022544564679265022, + 0.005355948116630316, + 0.014879653230309486, + 0.055568207055330276, + 0.02291453443467617, + 0.06959708780050278, + 0.0637204572558403, + 0.009036511182785034, + 0.04932121932506561, + 0.008355224505066872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 5 + ], + "selected_reward": -0.2521020174026489 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000925112864933908 + ], + "opposite_logprob_tokens": [ + -7.000925064086914 + ], + "logprob_reward": -0.000925112864933908, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.000925112864933908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040634140372276306, + 0.018654808402061462, + 0.03086927719414234, + 0.030741453170776367, + 0.03175950050354004, + 0.2062971442937851, + 0.025338858366012573, + 0.0796767920255661, + 0.04058327525854111, + 0.49544474482536316 + ], + "credit_map": [ + 0.003425684291869402, + 0.00963544100522995, + 0.0035918427165597677, + 0.008869485929608345, + 0.0058771357871592045, + 0.0026142876595258713, + 0.003975964151322842, + 0.0026442953385412693, + 0.001545157516375184, + 0.0022936228197067976, + 0.001987532014027238, + 0.0031644091941416264, + 0.0024952194653451443, + 0.002033528406172991, + 0.002892816672101617, + 0.0022425211500376463, + 0.003171256510540843, + 0.004804839845746756, + 0.0034102287609130144, + 0.004681534133851528, + 0.005319346208125353, + 0.0018351557664573193, + 0.005517353303730488, + 0.0021295619662851095, + 0.0025084293447434902, + 0.009486380033195019, + 0.002567985327914357, + 0.004788115154951811, + 0.004129445645958185, + 0.0015944818733260036, + 0.004165216349065304, + 0.001501397811807692, + 0.0023641111329197884, + 0.004558922722935677, + 0.001997391227632761, + 0.005833889357745647, + 0.01091831922531128, + 0.0020824309904128313, + 0.0028110668063163757, + 0.0011933650821447372, + 0.01919632963836193, + 0.06565006077289581, + 0.0208832286298275, + 0.0402127243578434, + 0.01934327371418476, + 0.018336936831474304, + 0.011502502486109734, + 0.01117206085473299, + 0.003176594153046608, + 0.005627840291708708, + 0.002341965911909938, + 0.003093570936471224, + 0.003541341284289956, + 0.0019349928479641676, + 0.003802679246291518, + 0.0018198703182861209, + 0.004983263090252876, + 0.02293049730360508, + 0.0036658653989434242, + 0.00409784447401762, + 0.025034533813595772, + 0.006554081570357084, + 0.005382592789828777, + 0.007028107065707445, + 0.0034998436458408833, + 0.006169436499476433, + 0.002611585659906268, + 0.006292352918535471, + 0.005529175512492657, + 0.002406455809250474, + 0.010646332055330276, + 0.0034280880354344845, + 0.022979410365223885, + 0.11719218641519547, + 0.028983531519770622, + 0.0930740237236023, + 0.13674655556678772, + 0.0172076765447855, + 0.06684494763612747, + 0.012416386045515537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.000925112864933908 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867067560553551 + ], + "opposite_logprob_tokens": [ + -3.048670768737793 + ], + "logprob_reward": -0.04867067560553551, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867067560553551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021153178066015244, + 0.035148900002241135, + 0.02700318768620491, + 0.03631613403558731, + 0.1040082648396492, + 0.08637873083353043, + 0.02430584654211998, + 0.08517610281705856, + 0.02771059423685074, + 0.5527990460395813 + ], + "credit_map": [ + 0.002170670311897993, + 0.004508272744715214, + 0.0018228659173473716, + 0.003645701799541712, + 0.003071699757128954, + 0.0016555598704144359, + 0.0023394771851599216, + 0.00193893164396286, + 0.002738700481131673, + 0.003034301334992051, + 0.0034834719263017178, + 0.012581335380673409, + 0.0020854317117482424, + 0.0024101831950247288, + 0.0037614142056554556, + 0.005054060369729996, + 0.0028249353636056185, + 0.004610604140907526, + 0.002191211562603712, + 0.004249075427651405, + 0.0037675427738577127, + 0.0017747109523043036, + 0.005659021437168121, + 0.0019260863773524761, + 0.002980421297252178, + 0.004647633992135525, + 0.003940674010664225, + 0.011501026339828968, + 0.003697900101542473, + 0.0026528961025178432, + 0.004932297859340906, + 0.0019632866606116295, + 0.0060204025357961655, + 0.014284276403486729, + 0.00435036513954401, + 0.01964843086898327, + 0.03991111367940903, + 0.007303041405975819, + 0.008941390551626682, + 0.003549239132553339, + 0.00810086727142334, + 0.026129096746444702, + 0.00881228968501091, + 0.01740802451968193, + 0.008386209607124329, + 0.0075422171503305435, + 0.0053795441053807735, + 0.004620482213795185, + 0.0027851855847984552, + 0.004222496412694454, + 0.0023674259427934885, + 0.003319305367767811, + 0.003935104236006737, + 0.0016706314636394382, + 0.004044457338750362, + 0.001961241476237774, + 0.0044678375124931335, + 0.02741275727748871, + 0.0026802276261150837, + 0.0023679290898144245, + 0.03254086896777153, + 0.006824500858783722, + 0.0058000823482871056, + 0.003081894712522626, + 0.0027530051302164793, + 0.004628384951502085, + 0.0031016780994832516, + 0.003739756066352129, + 0.004375234711915255, + 0.002075767843052745, + 0.0048847077414393425, + 0.0021520587615668774, + 0.02619410865008831, + 0.11767597496509552, + 0.03648200258612633, + 0.11852110922336578, + 0.1365075707435608, + 0.017970295622944832, + 0.08551803231239319, + 0.013929953798651695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.04867067560553551 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera, held perfectly still in a high-angle macro shot, captures a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface that fills the frame. Under bright, unfocused studio lighting, the liquid, with its deep, oily sheen, begins to slowly and relentlessly seep across the entire surface, its leading edge forming a perfect, glossy curve against the stark white. Without warning, the smooth surface of the liquid trembles before suddenly erupting upward into a dense forest of sharp, needle-like spikes. These structures are not static; they undulate and sway mesmerizingly in synchronized, wave-like motion, their shimmering tips dancing and reorganizing as if responding to the currents of a powerful, ever-shifting magnetic force unseen beyond the frame. The stark, high-contrast visuals and the liquid's controlled yet chaotic movement create a hypnotic and slightly unsettling atmosphere, underscoring the alien, organic quality of its beautiful yet sinister transformations.", + "seed": 12463, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_4/index_2949_seed_12463_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12802339429035783, + "vlm_logprob": -0.12802339429035783, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.12802339429035783, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1445896327495575 + ], + "opposite_logprob_tokens": [ + -2.14458966255188 + ], + "logprob_reward": -0.1445896327495575, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1445896327495575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15879622101783752, + 0.08788175880908966, + 0.1582300215959549, + 0.07043260335922241, + 0.025774570181965828, + 0.0608571320772171, + 0.02684028260409832, + 0.11588194966316223, + 0.2632294297218323, + 0.03207610547542572 + ], + "credit_map": [ + 0.0066866339184343815, + 0.007463413756340742, + 0.05411281809210777, + 0.007055351976305246, + 0.00820916797965765, + 0.009179096668958664, + 0.00669356994330883, + 0.05939615145325661, + 0.006423509679734707, + 0.012133043259382248, + 0.007622177246958017, + 0.008519881404936314, + 0.02409384958446026, + 0.011044356040656567, + 0.007827307097613811, + 0.010217638686299324, + 0.005671995226293802, + 0.018166840076446533, + 0.0066544474102556705, + 0.003920721355825663, + 0.060394953936338425, + 0.022598328068852425, + 0.008807873353362083, + 0.0320148691534996, + 0.005642308853566647, + 0.01171080581843853, + 0.0034586151596158743, + 0.006826804485172033, + 0.014745575375854969, + 0.013822407461702824, + 0.008981618098914623, + 0.005244466941803694, + 0.002712095621973276, + 0.005106869153678417, + 0.0017023984109982848, + 0.0021401920821517706, + 0.004552893340587616, + 0.004306253511458635, + 0.0026904242113232613, + 0.002563443500548601, + 0.00530529348179698, + 0.012531603686511517, + 0.003192446194589138, + 0.005258623976260424, + 0.008437981829047203, + 0.013210533186793327, + 0.006216987036168575, + 0.006703661289066076, + 0.003630302380770445, + 0.004041081760078669, + 0.0021924928296357393, + 0.0016730932984501123, + 0.004785535391420126, + 0.0052794539369642735, + 0.002042973879724741, + 0.0031953472644090652, + 0.00663414504379034, + 0.011645232327282429, + 0.004091375973075628, + 0.004014170728623867, + 0.046278826892375946, + 0.02956724725663662, + 0.005413029342889786, + 0.008237922564148903, + 0.022991299629211426, + 0.03875698894262314, + 0.011871017515659332, + 0.021408498287200928, + 0.07314807921648026, + 0.05984140932559967, + 0.016621273010969162, + 0.018590843304991722, + 0.003488947870209813, + 0.0067252810113132, + 0.002163143130019307, + 0.0026324631180614233, + 0.005404037423431873, + 0.007593125570565462, + 0.002162518911063671, + 0.0019065880915150046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.1445896327495575 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296876519918442 + ], + "opposite_logprob_tokens": [ + -2.237968683242798 + ], + "logprob_reward": -0.11296876519918442, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296876519918442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14902284741401672, + 0.09343385696411133, + 0.29760652780532837, + 0.05664973333477974, + 0.03452523425221443, + 0.03136684373021126, + 0.10698229819536209, + 0.1353585422039032, + 0.0703924372792244, + 0.02466168813407421 + ], + "credit_map": [ + 0.00702167022973299, + 0.00941325444728136, + 0.04417715594172478, + 0.010367875918745995, + 0.010720783844590187, + 0.007678277790546417, + 0.007958211936056614, + 0.05168561264872551, + 0.010057567618787289, + 0.010427935980260372, + 0.010948286391794682, + 0.004909273236989975, + 0.02559611387550831, + 0.010659800842404366, + 0.0073507800698280334, + 0.01348409615457058, + 0.010846921242773533, + 0.03384353965520859, + 0.012691771611571312, + 0.008830496110022068, + 0.11290882527828217, + 0.04148942232131958, + 0.01821226254105568, + 0.05878327041864395, + 0.004546977113932371, + 0.01036912202835083, + 0.0032502557151019573, + 0.005819696933031082, + 0.010263812728226185, + 0.00904793944209814, + 0.005412286147475243, + 0.007939640432596207, + 0.003231541719287634, + 0.006626179441809654, + 0.0019692981150001287, + 0.003945476375520229, + 0.007555786054581404, + 0.004002539906650782, + 0.0037225927226245403, + 0.0034718161914497614, + 0.003036020789295435, + 0.005519591737538576, + 0.0017051873728632927, + 0.0029848814010620117, + 0.005882965866476297, + 0.006865214090794325, + 0.0028275882359594107, + 0.0025453914422541857, + 0.010128108784556389, + 0.019158992916345596, + 0.006025037262588739, + 0.008179957047104836, + 0.02457479014992714, + 0.021825864911079407, + 0.007731035817414522, + 0.009358497336506844, + 0.009514957666397095, + 0.015998706221580505, + 0.007062119897454977, + 0.007926437072455883, + 0.04330141842365265, + 0.035595133900642395, + 0.00707263732329011, + 0.008887127973139286, + 0.007332755718380213, + 0.013013366609811783, + 0.00429747998714447, + 0.007428626995533705, + 0.014204664155840874, + 0.012705634348094463, + 0.005423282273113728, + 0.005986622069031, + 0.002920635510236025, + 0.004748594015836716, + 0.0018699063220992684, + 0.0017701656324788928, + 0.0042746402323246, + 0.0055810557678341866, + 0.0017189013306051493, + 0.0017777862958610058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.11296876519918442 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029806768521666527 + ], + "opposite_logprob_tokens": [ + -3.5298068523406982 + ], + "logprob_reward": -0.029806768521666527, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029806768521666527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13004064559936523, + 0.034365154802799225, + 0.0640546903014183, + 0.05656499043107033, + 0.014642849564552307, + 0.02128644473850727, + 0.020776595920324326, + 0.1344068944454193, + 0.509064257144928, + 0.014797468669712543 + ], + "credit_map": [ + 0.0062691885977983475, + 0.006929625757038593, + 0.04140986502170563, + 0.007979683578014374, + 0.008673814125359058, + 0.006987351458519697, + 0.005932826083153486, + 0.045858290046453476, + 0.0024308173451572657, + 0.005223896820098162, + 0.003453668672591448, + 0.0017270715907216072, + 0.010554653592407703, + 0.004343785345554352, + 0.0025738603435456753, + 0.004057401325553656, + 0.003441121894866228, + 0.008489992469549179, + 0.0026420000940561295, + 0.003518168581649661, + 0.018615305423736572, + 0.010766676627099514, + 0.0072984215803444386, + 0.009282998740673065, + 0.005745438858866692, + 0.009566448628902435, + 0.004594568163156509, + 0.005177913699299097, + 0.009450926445424557, + 0.012075110338628292, + 0.005369366146624088, + 0.004585216753184795, + 0.0015231394208967686, + 0.0031014783307909966, + 0.0010931749129667878, + 0.0014981548301875591, + 0.0023325805086642504, + 0.0024070926010608673, + 0.0013167372671887279, + 0.0013704919256269932, + 0.0023270954843610525, + 0.003593465546146035, + 0.001524246297776699, + 0.0020136672537773848, + 0.003318753791972995, + 0.0039541469886898994, + 0.002360865008085966, + 0.002194205764681101, + 0.001977111678570509, + 0.003212653798982501, + 0.001856148475781083, + 0.0013002714840695262, + 0.003480165032669902, + 0.005356937181204557, + 0.0015275218756869435, + 0.0020657845307141542, + 0.007829539477825165, + 0.013955285772681236, + 0.0033654356375336647, + 0.00458300719037652, + 0.055652108043432236, + 0.03574739024043083, + 0.005610681604593992, + 0.007663445081561804, + 0.02225079946219921, + 0.036289624869823456, + 0.012553864158689976, + 0.015277017839252949, + 0.20858901739120483, + 0.17943303287029266, + 0.01655537448823452, + 0.018115520477294922, + 0.0012193849543109536, + 0.0024868641048669815, + 0.0010173497721552849, + 0.0010209525935351849, + 0.0030407316517084837, + 0.0038458823692053556, + 0.0008564405143260956, + 0.0013098620111122727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.029806768521666527 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100962076336145 + ], + "opposite_logprob_tokens": [ + -5.504100799560547 + ], + "logprob_reward": -0.004100962076336145, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100962076336145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05414300784468651, + 0.07467503845691681, + 0.14545823633670807, + 0.10569179058074951, + 0.046216484159231186, + 0.09794096648693085, + 0.03268121927976608, + 0.08691881597042084, + 0.2989044487476349, + 0.05736995488405228 + ], + "credit_map": [ + 0.004051294643431902, + 0.008568175137043, + 0.010096292942762375, + 0.004667083732783794, + 0.006009995937347412, + 0.005741550121456385, + 0.0035290997475385666, + 0.011479522101581097, + 0.00511052506044507, + 0.011014637537300587, + 0.006279690191149712, + 0.0077290949411690235, + 0.021036330610513687, + 0.007987098768353462, + 0.007596748415380716, + 0.007920919917523861, + 0.005583213642239571, + 0.016125882044434547, + 0.005882992409169674, + 0.004328375682234764, + 0.055304739624261856, + 0.021205898374319077, + 0.00852506048977375, + 0.028502080589532852, + 0.00891568511724472, + 0.01757601648569107, + 0.0059617673978209496, + 0.011175554245710373, + 0.01739480160176754, + 0.0216089878231287, + 0.015449007973074913, + 0.007609978783875704, + 0.003997961524873972, + 0.010132952593266964, + 0.0034743514843285084, + 0.0043428936041891575, + 0.0062037063762545586, + 0.00708110723644495, + 0.005394581705331802, + 0.005588932428508997, + 0.012199737131595612, + 0.016256915405392647, + 0.0072822486981749535, + 0.009180370718240738, + 0.01679769530892372, + 0.017238952219486237, + 0.010490614920854568, + 0.008494438603520393, + 0.003481942228972912, + 0.005287530366331339, + 0.002772423205897212, + 0.0027606203220784664, + 0.005786009132862091, + 0.006977644748985767, + 0.0023438837379217148, + 0.003271167166531086, + 0.006131441332399845, + 0.0102352574467659, + 0.004608606453984976, + 0.0061219739727675915, + 0.02616528421640396, + 0.022312890738248825, + 0.004584856331348419, + 0.006758515723049641, + 0.012832503765821457, + 0.0205816850066185, + 0.006900697015225887, + 0.007985339500010014, + 0.1241026446223259, + 0.10599683225154877, + 0.009693586267530918, + 0.010811175219714642, + 0.006044880487024784, + 0.015088032931089401, + 0.0031615037005394697, + 0.004663105122745037, + 0.010754900984466076, + 0.010098465718328953, + 0.0038220658898353577, + 0.0037370019126683474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.004100962076336145 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34865084290504456 + ], + "opposite_logprob_tokens": [ + -1.2236508131027222 + ], + "logprob_reward": -0.34865084290504456, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34865084290504456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11094966530799866, + 0.061309073120355606, + 0.23275819420814514, + 0.05756770819425583, + 0.03216219320893288, + 0.038782037794589996, + 0.03669263795018196, + 0.1081610843539238, + 0.28902384638786316, + 0.03259354084730148 + ], + "credit_map": [ + 0.005305960774421692, + 0.010722498409450054, + 0.02715889923274517, + 0.009763968177139759, + 0.009612377732992172, + 0.006988218054175377, + 0.00787926372140646, + 0.03351848945021629, + 0.006700828671455383, + 0.010860014706850052, + 0.006753683555871248, + 0.004646581131964922, + 0.0089512187987566, + 0.007993733510375023, + 0.009338323026895523, + 0.006064689252525568, + 0.007565946783870459, + 0.025792617350816727, + 0.010238842107355595, + 0.008326835930347443, + 0.08656203001737595, + 0.03256143629550934, + 0.016020797193050385, + 0.04568970203399658, + 0.005127732176333666, + 0.011261703446507454, + 0.004015421494841576, + 0.006885741371661425, + 0.009808572940528393, + 0.010435860604047775, + 0.005790536291897297, + 0.0042421454563736916, + 0.0028587107080966234, + 0.006063822191208601, + 0.0021942337043583393, + 0.003969493322074413, + 0.005540512967854738, + 0.004518243949860334, + 0.003402132773771882, + 0.003615048248320818, + 0.003602623473852873, + 0.007090856321156025, + 0.0021253987215459347, + 0.003315675538033247, + 0.006362051237374544, + 0.008280233480036259, + 0.004409450571984053, + 0.0035957512445747852, + 0.003118795808404684, + 0.005916777532547712, + 0.002364913234487176, + 0.0031252887565642595, + 0.007715640123933554, + 0.008074031211435795, + 0.003093432169407606, + 0.0032837586477398872, + 0.006351337302476168, + 0.010699131526052952, + 0.0036824482958763838, + 0.0047197709791362286, + 0.042221251875162125, + 0.029963476583361626, + 0.005524049513041973, + 0.004999621771275997, + 0.014894066378474236, + 0.02601446397602558, + 0.008133661933243275, + 0.01335732452571392, + 0.10945390164852142, + 0.09219612181186676, + 0.011659229174256325, + 0.01331509742885828, + 0.004265028517693281, + 0.006768241059035063, + 0.0021658854093402624, + 0.0023864663671702147, + 0.005404610186815262, + 0.006741065066307783, + 0.002550263423472643, + 0.0023119819816201925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.34865084290504456 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera, held perfectly still in a high-angle macro shot, captures a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface that fills the frame. Under bright, unfocused studio lighting, the liquid, with its deep, oily sheen, begins to slowly and relentlessly seep across the entire surface, its leading edge forming a perfect, glossy curve against the stark white. Without warning, the smooth surface of the liquid trembles before suddenly erupting upward into a dense forest of sharp, needle-like spikes. These structures are not static; they undulate and sway mesmerizingly in synchronized, wave-like motion, their shimmering tips dancing and reorganizing as if responding to the currents of a powerful, ever-shifting magnetic force unseen beyond the frame. The stark, high-contrast visuals and the liquid's controlled yet chaotic movement create a hypnotic and slightly unsettling atmosphere, underscoring the alien, organic quality of its beautiful yet sinister transformations.", + "seed": 12464, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_4/index_2949_seed_12464_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1373543043853715, + "vlm_logprob": -0.1373543043853715, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.1373543043853715, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.36281833052635193 + ], + "opposite_logprob_tokens": [ + -1.2378183603286743 + ], + "logprob_reward": -0.36281833052635193, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.36281833052635193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1736396849155426, + 0.2017544060945511, + 0.08872254192829132, + 0.06160422042012215, + 0.06304249167442322, + 0.0414612703025341, + 0.056228380650281906, + 0.07057276368141174, + 0.12545335292816162, + 0.11752092838287354 + ], + "credit_map": [ + 0.03493133559823036, + 0.025726327672600746, + 0.03258461877703667, + 0.013424308970570564, + 0.01832502894103527, + 0.01918427087366581, + 0.014379104599356651, + 0.015084697864949703, + 0.04328741878271103, + 0.07726075500249863, + 0.017860110849142075, + 0.015328758396208286, + 0.012799659743905067, + 0.011731992475688457, + 0.015551140531897545, + 0.007934569381177425, + 0.014796026982367039, + 0.017105309292674065, + 0.010671772994101048, + 0.007883467711508274, + 0.006702135317027569, + 0.008134241215884686, + 0.009419010952115059, + 0.014010576531291008, + 0.007096842862665653, + 0.010986640118062496, + 0.005745399743318558, + 0.006044508423656225, + 0.006385886110365391, + 0.009532763622701168, + 0.007432866841554642, + 0.00837931502610445, + 0.00967747438699007, + 0.009163479320704937, + 0.004344888497143984, + 0.009365844540297985, + 0.00651793647557497, + 0.009211542084813118, + 0.007862756960093975, + 0.006898568943142891, + 0.005258188582956791, + 0.010044241324067116, + 0.0026373162399977446, + 0.0052810292690992355, + 0.004292033612728119, + 0.004559187218546867, + 0.006677238270640373, + 0.0027120367158204317, + 0.006743764970451593, + 0.010810481384396553, + 0.005799721926450729, + 0.007230572402477264, + 0.006468720268458128, + 0.0074995034374296665, + 0.0069622197188436985, + 0.00471339700743556, + 0.008890126831829548, + 0.01232825592160225, + 0.008547216653823853, + 0.010757342912256718, + 0.008692240342497826, + 0.007858986034989357, + 0.006778671406209469, + 0.00671992264688015, + 0.026216138154268265, + 0.009107095189392567, + 0.0062735616229474545, + 0.010647010058164597, + 0.05126497521996498, + 0.009902505204081535, + 0.00737000023946166, + 0.004672064911574125, + 0.025072503834962845, + 0.017334217205643654, + 0.009477773681282997, + 0.014517881907522678, + 0.023365387693047523, + 0.00921460147947073, + 0.00747430557385087, + 0.011064252816140652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.36281833052635193 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005693763494492 + ], + "opposite_logprob_tokens": [ + -2.695056915283203 + ], + "logprob_reward": -0.07005693763494492, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005693763494492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08650325983762741, + 0.303760290145874, + 0.06405365467071533, + 0.04067494720220566, + 0.02754540555179119, + 0.05214014649391174, + 0.04179792106151581, + 0.027246570214629173, + 0.0709790289402008, + 0.28529873490333557 + ], + "credit_map": [ + 0.011231022886931896, + 0.014382371678948402, + 0.015404456295073032, + 0.008274764753878117, + 0.010351250879466534, + 0.011176570318639278, + 0.008094098418951035, + 0.007588720880448818, + 0.029849253594875336, + 0.12705403566360474, + 0.0831715315580368, + 0.015602678060531616, + 0.012958516366779804, + 0.012473976239562035, + 0.016459070146083832, + 0.006191227585077286, + 0.006857273634523153, + 0.012944458983838558, + 0.008049729280173779, + 0.006383270490914583, + 0.007899084128439426, + 0.007305077742785215, + 0.009782117791473866, + 0.004832638427615166, + 0.004500644281506538, + 0.008155951276421547, + 0.003382912604138255, + 0.00424534734338522, + 0.00410350039601326, + 0.004643023479729891, + 0.006668221205472946, + 0.004975346848368645, + 0.002827486488968134, + 0.0031215210910886526, + 0.0018037802074104548, + 0.0035412651486694813, + 0.0033860800322145224, + 0.005861580837517977, + 0.003677617758512497, + 0.003326073521748185, + 0.004707059357315302, + 0.008132404647767544, + 0.002180074341595173, + 0.007000314071774483, + 0.004008730407804251, + 0.005053319502621889, + 0.017952613532543182, + 0.0031056301668286324, + 0.005270184949040413, + 0.004751415923237801, + 0.0029501470271497965, + 0.007425654213875532, + 0.0032697715796530247, + 0.0030032929498702288, + 0.012260527350008488, + 0.002866927534341812, + 0.0036762263625860214, + 0.00484547670930624, + 0.0026521196123212576, + 0.0039061768911778927, + 0.003254728624597192, + 0.002957204356789589, + 0.0036808678414672613, + 0.002273769350722432, + 0.01395613793283701, + 0.0063532632775604725, + 0.0036223442293703556, + 0.006444873288273811, + 0.028322890400886536, + 0.00523611344397068, + 0.004040200263261795, + 0.00300321145914495, + 0.05605630949139595, + 0.056795526295900345, + 0.018953446298837662, + 0.03201288357377052, + 0.05788184329867363, + 0.02157049998641014, + 0.016483493149280548, + 0.025544751435518265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.07005693763494492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026357922703027725 + ], + "opposite_logprob_tokens": [ + -3.651357889175415 + ], + "logprob_reward": -0.026357922703027725, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026357922703027725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1223183199763298, + 0.3580143451690674, + 0.058291271328926086, + 0.052146054804325104, + 0.055056095123291016, + 0.06122925505042076, + 0.05229319632053375, + 0.10423112660646439, + 0.06979632377624512, + 0.06662400811910629 + ], + "credit_map": [ + 0.01694229058921337, + 0.020785296335816383, + 0.015178771689534187, + 0.01111417543143034, + 0.015873383730649948, + 0.01574978418648243, + 0.013650898821651936, + 0.01302371732890606, + 0.03394652530550957, + 0.14831621944904327, + 0.09861892461776733, + 0.01884058304131031, + 0.016654061153531075, + 0.014205380342900753, + 0.019974125549197197, + 0.007458521518856287, + 0.008999978192150593, + 0.014203977771103382, + 0.0067886728793382645, + 0.005306751932948828, + 0.004748173523694277, + 0.006200137082487345, + 0.006613105535507202, + 0.005430472083389759, + 0.005228058900684118, + 0.00545381847769022, + 0.004547757562249899, + 0.007623481098562479, + 0.00541295949369669, + 0.008359553292393684, + 0.006721685640513897, + 0.00879873987287283, + 0.012248157523572445, + 0.011440356262028217, + 0.00470714271068573, + 0.005557978991419077, + 0.004129245411604643, + 0.007325596176087856, + 0.0043304054997861385, + 0.00531721580773592, + 0.009778177365660667, + 0.006082243286073208, + 0.003916472662240267, + 0.007095795590430498, + 0.0068948958069086075, + 0.00971614196896553, + 0.01364065520465374, + 0.004104868974536657, + 0.006408315617591143, + 0.005693203303962946, + 0.00556497648358345, + 0.007366999052464962, + 0.006419911049306393, + 0.007227014284580946, + 0.009339950978755951, + 0.004272826481610537, + 0.01566925086081028, + 0.02055167965590954, + 0.009209422394633293, + 0.012405266053974628, + 0.01800657995045185, + 0.011081814765930176, + 0.009478429332375526, + 0.007828686386346817, + 0.014032197184860706, + 0.004886726848781109, + 0.003852068679407239, + 0.005694856401532888, + 0.028733398765325546, + 0.00570355961099267, + 0.0039437864907085896, + 0.002949727699160576, + 0.007133117876946926, + 0.012684517540037632, + 0.005391956306993961, + 0.006796677131205797, + 0.009363699704408646, + 0.011803272180259228, + 0.0064800577238202095, + 0.006970701739192009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.026357922703027725 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019455092260614038 + ], + "opposite_logprob_tokens": [ + -6.251945495605469 + ], + "logprob_reward": -0.0019455092260614038, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019455092260614038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060614507645368576, + 0.13907845318317413, + 0.015626655891537666, + 0.04003239795565605, + 0.018641529604792595, + 0.04717598110437393, + 0.39311736822128296, + 0.10696633905172348, + 0.14676237106323242, + 0.03198442980647087 + ], + "credit_map": [ + 0.009733133018016815, + 0.010107466019690037, + 0.008341450244188309, + 0.0050759646110236645, + 0.005903610959649086, + 0.007813097909092903, + 0.006554442457854748, + 0.007085349410772324, + 0.023882601410150528, + 0.03360261768102646, + 0.05584852024912834, + 0.007125616539269686, + 0.004258935339748859, + 0.004203184507787228, + 0.006869340781122446, + 0.0032876471523195505, + 0.0019740930292755365, + 0.0035332327242940664, + 0.0015372654888778925, + 0.0018217549659311771, + 0.0014637800632044673, + 0.0019409176893532276, + 0.0021265631075948477, + 0.0012290515005588531, + 0.004207592457532883, + 0.007884564809501171, + 0.0035720914602279663, + 0.0038748332299292088, + 0.003972203005105257, + 0.0051677413284778595, + 0.0051679727621376514, + 0.0061854030936956406, + 0.001947580138221383, + 0.002345958724617958, + 0.00160212314222008, + 0.002171702217310667, + 0.0021835521329194307, + 0.003107464173808694, + 0.0025159826036542654, + 0.002767168218269944, + 0.004150029271841049, + 0.0059938011690974236, + 0.0022186546120792627, + 0.0062601324170827866, + 0.0045344945974648, + 0.004930592142045498, + 0.01588885486125946, + 0.0031994283199310303, + 0.06523261219263077, + 0.01667855866253376, + 0.017231035977602005, + 0.06930790096521378, + 0.024811822921037674, + 0.02518651634454727, + 0.1478578746318817, + 0.02681106887757778, + 0.015004211105406284, + 0.014489737339317799, + 0.01418626680970192, + 0.015536824241280556, + 0.017272811383008957, + 0.013568633235991001, + 0.010484945960342884, + 0.006422922480851412, + 0.030696984380483627, + 0.009499792009592056, + 0.007317210081964731, + 0.012366256676614285, + 0.06190025806427002, + 0.011152142658829689, + 0.008373969234526157, + 0.00545577984303236, + 0.005385307595133781, + 0.003671941114589572, + 0.0028768901247531176, + 0.00471383286640048, + 0.006647814996540546, + 0.0034078729804605246, + 0.0023425037506967783, + 0.002938270103186369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.0019455092260614038 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera held perfectly still in a high-angle macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a puddle of viscous, inky black ferrofluid on a pristine, non-reflective matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ferrofluid slowly seep across the surface, forming a perfect glossy curve at its leading edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid suddenly erupt upward into a dense forest of sharp, needle-like spikes that undulate and sway in synchronized wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright and unfocused, creating a stark, high-contrast visual style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559282183647156 + ], + "opposite_logprob_tokens": [ + -1.600592851638794 + ], + "logprob_reward": -0.22559282183647156, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559282183647156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15814678370952606, + 0.08522463589906693, + 0.06616921722888947, + 0.05148342251777649, + 0.03155110776424408, + 0.04099554196000099, + 0.2120230793952942, + 0.1322048157453537, + 0.08358258754014969, + 0.13861878216266632 + ], + "credit_map": [ + 0.021395182237029076, + 0.03182525187730789, + 0.02678167074918747, + 0.013896891847252846, + 0.016311436891555786, + 0.02114640548825264, + 0.012760224752128124, + 0.014029702171683311, + 0.024518828839063644, + 0.006224901881068945, + 0.028950640931725502, + 0.007176708430051804, + 0.005037092603743076, + 0.004208768717944622, + 0.005394112318754196, + 0.0037135882303118706, + 0.005872508976608515, + 0.020287996158003807, + 0.007917994633316994, + 0.004937654826790094, + 0.0058151246048510075, + 0.005878169555217028, + 0.010346007533371449, + 0.005113756749778986, + 0.008754021488130093, + 0.011326820589601994, + 0.0037753183860331774, + 0.0043985904194414616, + 0.0042539313435554504, + 0.0076169297099113464, + 0.005158527288585901, + 0.006199284456670284, + 0.004550025798380375, + 0.00406845984980464, + 0.0027756758499890566, + 0.0054954299703240395, + 0.0027778290677815676, + 0.0048225196078419685, + 0.004122988786548376, + 0.002938179997727275, + 0.00569957448169589, + 0.011660436168313026, + 0.0022021704353392124, + 0.005035605281591415, + 0.0036034754011780024, + 0.005084593314677477, + 0.005935021210461855, + 0.0017746653174981475, + 0.030241744592785835, + 0.015226820483803749, + 0.013513428159058094, + 0.03601711243391037, + 0.018581710755825043, + 0.019550366327166557, + 0.06247669830918312, + 0.01641521416604519, + 0.01300518587231636, + 0.026390675455331802, + 0.013978361152112484, + 0.0202945526689291, + 0.01822967268526554, + 0.015066970139741898, + 0.01647273637354374, + 0.008766673505306244, + 0.01798442378640175, + 0.008423170074820518, + 0.004376930184662342, + 0.00737414788454771, + 0.03006795607507229, + 0.005985751282423735, + 0.004616240970790386, + 0.004753962624818087, + 0.022623613476753235, + 0.028337281197309494, + 0.008804388344287872, + 0.017255527898669243, + 0.019442200660705566, + 0.015632720664143562, + 0.011661791242659092, + 0.014861264266073704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.22559282183647156 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_5/index_tensor([6194, 6194])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_5/index_tensor([6194, 6194])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..525d68a45656f5069ad59d69ece45ead28d74e31 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_5/index_tensor([6194, 6194])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme macro shot captures the tip of a weathered brass spout, its surface a tapestry of aged tapestry, a dull metallic sheen punctuated by green patina of oxidized copper. A cool, diffuse light illuminates the scene, creating soft, crisp highlights on the metal and a single, shimmering droplet of water clinging to the spout. Pulled by the unstoppable force of gravity, the droplet slowly and deliberately elongates, stretching downward into a teardrop shape, its surface tension visibly strained. With one final, clean break, it separates from the brass and quickly falls from the bottom of the frame, leaving the spout momentarily exposed. A new, smaller droplet immediately begins to expand and form, continuing a quiet, rhythmic cycle amidst a serene and focused atmosphere.", + "seed": 19023, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_5/index_6194_seed_19023_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.056971911597065626, + "vlm_logprob": -0.056971911597065626, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.056971911597065626, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008439235389232635 + ], + "opposite_logprob_tokens": [ + -7.133439064025879 + ], + "logprob_reward": -0.008439235389232635, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.008439235389232635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06658393889665604, + 0.22736580669879913, + 0.08996214717626572, + 0.0674654096364975, + 0.013162558898329735, + 0.16968679428100586, + 0.11127543449401855, + 0.07594972103834152, + 0.14893649518489838, + 0.029611650854349136 + ], + "credit_map": [ + 0.016358988359570503, + 0.006904360838234425, + 0.006421905010938644, + 0.003257657401263714, + 0.005193722900003195, + 0.006269107572734356, + 0.004018001724034548, + 0.018160194158554077, + 0.07753970474004745, + 0.020184319466352463, + 0.015774523839354515, + 0.010074971243739128, + 0.010671127587556839, + 0.01962832361459732, + 0.00946240033954382, + 0.06403046101331711, + 0.024477213621139526, + 0.006885872222483158, + 0.017870739102363586, + 0.007887977175414562, + 0.006563800852745771, + 0.006077907979488373, + 0.012019907124340534, + 0.008178738877177238, + 0.012247411534190178, + 0.004538446664810181, + 0.008380677551031113, + 0.0036136989947408438, + 0.004324506036937237, + 0.0070163230411708355, + 0.006584567949175835, + 0.020759787410497665, + 0.002120526973158121, + 0.0009133261628448963, + 0.001060476410202682, + 0.0013160684611648321, + 0.001317918417043984, + 0.0012930788798257709, + 0.0020686492789536715, + 0.003072514431551099, + 0.05691252276301384, + 0.012382376939058304, + 0.013415215536952019, + 0.007226111367344856, + 0.0043984148651361465, + 0.014326614327728748, + 0.009057722985744476, + 0.051967840641736984, + 0.03055676631629467, + 0.003381217597052455, + 0.0441398099064827, + 0.00497070699930191, + 0.0029837111942470074, + 0.006782961077988148, + 0.0039020248223096132, + 0.014558246359229088, + 0.014421256259083748, + 0.005618794355541468, + 0.010463087819516659, + 0.00883669126778841, + 0.007367413491010666, + 0.007585511077195406, + 0.013271844945847988, + 0.00838513020426035, + 0.041729796677827835, + 0.0044059958308935165, + 0.0577850304543972, + 0.005917666479945183, + 0.00483422027900815, + 0.009704825468361378, + 0.0061065321788191795, + 0.018452439457178116, + 0.008450375869870186, + 0.0021475786343216896, + 0.004921786021441221, + 0.0027226556558161974, + 0.002452307613566518, + 0.0029260204173624516, + 0.0031882058829069138, + 0.002802722156047821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 8 + ], + "selected_reward": -0.008439235389232635 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370229199528694 + ], + "opposite_logprob_tokens": [ + -3.6513702869415283 + ], + "logprob_reward": -0.026370229199528694, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370229199528694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24791830778121948, + 0.03947025537490845, + 0.025953784584999084, + 0.04097210615873337, + 0.0435267835855484, + 0.021533384919166565, + 0.0770929604768753, + 0.3264974057674408, + 0.11572889238595963, + 0.06130620837211609 + ], + "credit_map": [ + 0.08264421671628952, + 0.022186581045389175, + 0.018377264961600304, + 0.012443347834050655, + 0.013320897705852985, + 0.02085990086197853, + 0.013584003783762455, + 0.0645020455121994, + 0.005529025103896856, + 0.00515888724476099, + 0.005626150406897068, + 0.0024448991753160954, + 0.004471090156584978, + 0.004310503602027893, + 0.002764798002317548, + 0.009164891205728054, + 0.004955132491886616, + 0.002436723094433546, + 0.0023299409076571465, + 0.002272670855745673, + 0.002174589317291975, + 0.002924263011664152, + 0.0031384271569550037, + 0.005722031928598881, + 0.009033799171447754, + 0.003434626618400216, + 0.0041578407399356365, + 0.002480514580383897, + 0.002757161622866988, + 0.0038028156850486994, + 0.005128348711878061, + 0.010176992043852806, + 0.003855837509036064, + 0.003102485090494156, + 0.0068823606707155704, + 0.004324824083596468, + 0.00422700447961688, + 0.004630111623555422, + 0.00668902974575758, + 0.009815121069550514, + 0.003241585101932287, + 0.0011602832237258554, + 0.00291646015830338, + 0.001816264120861888, + 0.0020759382750838995, + 0.0023797324392944574, + 0.005116830579936504, + 0.0028262862470000982, + 0.020382190123200417, + 0.0019806353375315666, + 0.031116684898734093, + 0.0030519745778292418, + 0.0020045689307153225, + 0.004632008261978626, + 0.0027147596701979637, + 0.011210129596292973, + 0.07651741802692413, + 0.014466168358922005, + 0.09791773557662964, + 0.026317941024899483, + 0.020357808098196983, + 0.02233484759926796, + 0.041565604507923126, + 0.02701980248093605, + 0.009303465485572815, + 0.007529034279286861, + 0.0358104445040226, + 0.006593978498131037, + 0.005620364565402269, + 0.009357931092381477, + 0.007911542430520058, + 0.033602118492126465, + 0.0207901019603014, + 0.0037740853149443865, + 0.012316678650677204, + 0.0035600168630480766, + 0.0027713903691619635, + 0.004541066940873861, + 0.005769349168986082, + 0.007783509790897369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.026370229199528694 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006410212954506278 + ], + "opposite_logprob_tokens": [ + -7.375640869140625 + ], + "logprob_reward": -0.0006410212954506278, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006410212954506278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0982884094119072, + 0.03760659322142601, + 0.06752389669418335, + 0.05761196091771126, + 0.034655749797821045, + 0.14713361859321594, + 0.06466494500637054, + 0.32406365871429443, + 0.14002777636051178, + 0.028423335403203964 + ], + "credit_map": [ + 0.034522946923971176, + 0.007975985296070576, + 0.009361443109810352, + 0.006174927111715078, + 0.004232735838741064, + 0.007803240325301886, + 0.004654506221413612, + 0.023562638089060783, + 0.010274963453412056, + 0.0035663852468132973, + 0.00390581670217216, + 0.0028902445919811726, + 0.003143309149891138, + 0.003899139352142811, + 0.0026132911443710327, + 0.007313442416489124, + 0.008132560178637505, + 0.006090880371630192, + 0.009745083749294281, + 0.006513070780783892, + 0.00634040916338563, + 0.007072541397064924, + 0.0058061215095222, + 0.017823228612542152, + 0.014169142581522465, + 0.005256233736872673, + 0.004937784280627966, + 0.0040453658439219, + 0.004856865853071213, + 0.005192737560719252, + 0.007427133619785309, + 0.011726702563464642, + 0.007493925280869007, + 0.0021297300700098276, + 0.0032909971196204424, + 0.002476528286933899, + 0.002348549198359251, + 0.0032217775005847216, + 0.004075287841260433, + 0.00961895752698183, + 0.040957625955343246, + 0.010302652604877949, + 0.017524704337120056, + 0.006381239742040634, + 0.00413784384727478, + 0.011749427765607834, + 0.009447740390896797, + 0.04663239046931267, + 0.020911335945129395, + 0.003567925887182355, + 0.019840717315673828, + 0.004359425976872444, + 0.003866913029924035, + 0.005192240700125694, + 0.00319121265783906, + 0.0037351830396801233, + 0.05171623080968857, + 0.01570018380880356, + 0.07900816947221756, + 0.03429009020328522, + 0.023717425763607025, + 0.02293267846107483, + 0.0651639923453331, + 0.03153491020202637, + 0.020523475483059883, + 0.00851338729262352, + 0.04360704496502876, + 0.00826256349682808, + 0.010801905766129494, + 0.011718989349901676, + 0.008464915677905083, + 0.02813551016151905, + 0.003584484104067087, + 0.002334624994546175, + 0.006632487289607525, + 0.002180827083066106, + 0.003095826832577586, + 0.003114615334197879, + 0.002541930414736271, + 0.004938540514558554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.0006410212954506278 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890391394495964 + ], + "opposite_logprob_tokens": [ + -2.4640390872955322 + ], + "logprob_reward": -0.0890391394495964, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890391394495964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04536781087517738, + 0.1089891567826271, + 0.07902560383081436, + 0.046779654920101166, + 0.088812455534935, + 0.3113573491573334, + 0.042803481221199036, + 0.17375774681568146, + 0.087466299533844, + 0.01564040593802929 + ], + "credit_map": [ + 0.0118989497423172, + 0.0041031260043382645, + 0.004060923121869564, + 0.002775523578748107, + 0.003103792667388916, + 0.004263847600668669, + 0.0036514773964881897, + 0.011510174721479416, + 0.0329802930355072, + 0.00945685338228941, + 0.009584619663655758, + 0.004733183421194553, + 0.0049918838776648045, + 0.00965137965977192, + 0.004638045560568571, + 0.03295291215181351, + 0.027476003393530846, + 0.006179444491863251, + 0.011160710826516151, + 0.004752339329570532, + 0.003927416633814573, + 0.00537368468940258, + 0.00820216629654169, + 0.011953848414123058, + 0.012993681244552135, + 0.004584182985126972, + 0.004119459073990583, + 0.0034779291599988937, + 0.0033988631330430508, + 0.0038287288043648005, + 0.007784085348248482, + 0.006592726334929466, + 0.033099789172410965, + 0.006813715677708387, + 0.008932244963943958, + 0.004291492979973555, + 0.002960417652502656, + 0.007331437431275845, + 0.006087261252105236, + 0.019296107813715935, + 0.08130297809839249, + 0.02071574702858925, + 0.03648994117975235, + 0.014843156561255455, + 0.011617900803685188, + 0.024885766208171844, + 0.026880884543061256, + 0.09462103247642517, + 0.007406739983707666, + 0.002656857715919614, + 0.01132882572710514, + 0.004029368981719017, + 0.00391680933535099, + 0.0037401372101157904, + 0.005535620264708996, + 0.004189126659184694, + 0.031664688140153885, + 0.009695300832390785, + 0.042703885585069656, + 0.01743280328810215, + 0.014398450031876564, + 0.012711625546216965, + 0.030554672703146935, + 0.014596333727240562, + 0.009643349796533585, + 0.00835406593978405, + 0.015190934762358665, + 0.007317120674997568, + 0.010264808312058449, + 0.009843415580689907, + 0.00739694619551301, + 0.019455665722489357, + 0.003613476874306798, + 0.0013628624146804214, + 0.004312917590141296, + 0.00137804402038455, + 0.0013760467991232872, + 0.0012696677586063743, + 0.0013161477399989963, + 0.0010112443706020713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.0890391394495964 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036993265151978 + ], + "opposite_logprob_tokens": [ + -1.910369873046875 + ], + "logprob_reward": -0.16036993265151978, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036993265151978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1506490558385849, + 0.038271814584732056, + 0.10812031477689743, + 0.12970107793807983, + 0.05788996443152428, + 0.09702508896589279, + 0.07948143035173416, + 0.21655361354351044, + 0.07107745856046677, + 0.05123016983270645 + ], + "credit_map": [ + 0.05527978762984276, + 0.011765125207602978, + 0.016574572771787643, + 0.007101447321474552, + 0.006796345580369234, + 0.012395567260682583, + 0.006719868630170822, + 0.03401634469628334, + 0.009701070375740528, + 0.004193602595478296, + 0.002842186251655221, + 0.0031937304884195328, + 0.004636966623365879, + 0.0036153828259557486, + 0.0025450559332966805, + 0.007543819956481457, + 0.04010463133454323, + 0.008389228023588657, + 0.014654159545898438, + 0.004842006601393223, + 0.005973134655505419, + 0.008226750418543816, + 0.007236409466713667, + 0.01869400590658188, + 0.041384123265743256, + 0.010237160138785839, + 0.017239397391676903, + 0.01023429911583662, + 0.006881108041852713, + 0.00990673154592514, + 0.021796124055981636, + 0.012022142298519611, + 0.020924340933561325, + 0.004436003975570202, + 0.0055230287835001945, + 0.0032239709980785847, + 0.002016513142734766, + 0.0046936566941440105, + 0.004368695430457592, + 0.01270375493913889, + 0.023506060242652893, + 0.0069974083453416824, + 0.00720186997205019, + 0.006202077027410269, + 0.006856870371848345, + 0.008658504113554955, + 0.015009520575404167, + 0.022592786699533463, + 0.0251421257853508, + 0.0043298183009028435, + 0.024843543767929077, + 0.004469594918191433, + 0.003986124414950609, + 0.00489162327721715, + 0.006622988265007734, + 0.005195614416152239, + 0.03850826993584633, + 0.00951377023011446, + 0.06700775027275085, + 0.0186019204556942, + 0.013451947830617428, + 0.015569877810776234, + 0.024899182841181755, + 0.02900090254843235, + 0.008968398906290531, + 0.005876623559743166, + 0.016238076612353325, + 0.005262420512735844, + 0.008039007894694805, + 0.00877521000802517, + 0.006365372333675623, + 0.011552347801625729, + 0.014517199248075485, + 0.004469584207981825, + 0.005676050204783678, + 0.0032407159451395273, + 0.0040762340649962425, + 0.005725665017962456, + 0.0031733782961964607, + 0.010351346805691719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.16036993265151978 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot captures the tip of a weathered brass spout, its surface a tapestry of aged tapestry, a dull metallic sheen punctuated by green patina of oxidized copper. A cool, diffuse light illuminates the scene, creating soft, crisp highlights on the metal and a single, shimmering droplet of water clinging to the spout. Pulled by the unstoppable force of gravity, the droplet slowly and deliberately elongates, stretching downward into a teardrop shape, its surface tension visibly strained. With one final, clean break, it separates from the brass and quickly falls from the bottom of the frame, leaving the spout momentarily exposed. A new, smaller droplet immediately begins to expand and form, continuing a quiet, rhythmic cycle amidst a serene and focused atmosphere.", + "seed": 19024, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_5/index_6194_seed_19024_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12973131723701953, + "vlm_logprob": -0.12973131723701953, + "vlm_yes_no_margin": 2.924999988079071, + "vlm_token_credit": -0.12973131723701953, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06096968054771423 + ], + "opposite_logprob_tokens": [ + -3.185969591140747 + ], + "logprob_reward": -0.06096968054771423, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.06096968054771423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16450677812099457, + 0.14549648761749268, + 0.08520371466875076, + 0.030421148985624313, + 0.0471113882958889, + 0.07011915743350983, + 0.18592752516269684, + 0.04534413665533066, + 0.16640077531337738, + 0.0594688355922699 + ], + "credit_map": [ + 0.02375299483537674, + 0.01668364740908146, + 0.024783415719866753, + 0.026538366451859474, + 0.011944285593926907, + 0.02248050831258297, + 0.01402713917195797, + 0.024296415969729424, + 0.010858582332730293, + 0.01495942659676075, + 0.014649104326963425, + 0.018368348479270935, + 0.025457801297307014, + 0.02012539841234684, + 0.01999131590127945, + 0.021086523309350014, + 0.006402355618774891, + 0.005914442706853151, + 0.008426523767411709, + 0.011067626997828484, + 0.004179973155260086, + 0.02083149552345276, + 0.007414846681058407, + 0.02096644975244999, + 0.0032440524082630873, + 0.003100388450548053, + 0.004296452272683382, + 0.0039394996128976345, + 0.003845762461423874, + 0.003768066642805934, + 0.004899566527456045, + 0.0033273608423769474, + 0.003201159182935953, + 0.0033000819385051727, + 0.005616846494376659, + 0.007448725402355194, + 0.003729937830939889, + 0.009017203003168106, + 0.005067028570920229, + 0.009730411693453789, + 0.003391844453290105, + 0.003532308153808117, + 0.0038378285244107246, + 0.004413134418427944, + 0.003034953959286213, + 0.02409505471587181, + 0.0035634655505418777, + 0.0242505744099617, + 0.013388355262577534, + 0.013460211455821991, + 0.014888748526573181, + 0.018939407542347908, + 0.015901127830147743, + 0.040147650986909866, + 0.027202680706977844, + 0.041999347507953644, + 0.004139385186135769, + 0.00391909247264266, + 0.0045232018455863, + 0.0048859864473342896, + 0.0054102493450045586, + 0.008190648630261421, + 0.006500367075204849, + 0.007775206118822098, + 0.015639465302228928, + 0.014812970533967018, + 0.018349794670939445, + 0.02440664730966091, + 0.01574822887778282, + 0.027813320979475975, + 0.013399498537182808, + 0.03623085841536522, + 0.004181969910860062, + 0.0040334006771445274, + 0.006487340200692415, + 0.009471999481320381, + 0.003975112922489643, + 0.012318176217377186, + 0.005408322438597679, + 0.01359251793473959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.06096968054771423 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02638009935617447 + ], + "opposite_logprob_tokens": [ + -3.6513800621032715 + ], + "logprob_reward": -0.02638009935617447, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638009935617447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14034855365753174, + 0.15251271426677704, + 0.11637751013040543, + 0.0733625739812851, + 0.08229280263185501, + 0.07668757438659668, + 0.1698022037744522, + 0.07903783023357391, + 0.0738690122961998, + 0.0357092060148716 + ], + "credit_map": [ + 0.021388161927461624, + 0.01325027085840702, + 0.022274507209658623, + 0.02413269318640232, + 0.009357936680316925, + 0.017364736646413803, + 0.013278653845191002, + 0.01930159516632557, + 0.012107516638934612, + 0.01765536144375801, + 0.015540136024355888, + 0.01861937902867794, + 0.030958034098148346, + 0.017916375771164894, + 0.02213309518992901, + 0.017582807689905167, + 0.010803740471601486, + 0.010209127329289913, + 0.011539313942193985, + 0.014410953037440777, + 0.008602446876466274, + 0.024214938282966614, + 0.011995689012110233, + 0.024601299315690994, + 0.007275964133441448, + 0.00488695316016674, + 0.009980703704059124, + 0.00832778587937355, + 0.006336651276797056, + 0.014734885655343533, + 0.007517330348491669, + 0.01430229377001524, + 0.01000602263957262, + 0.008941650390625, + 0.009942581877112389, + 0.010956848040223122, + 0.006819053553044796, + 0.01259687077254057, + 0.008772374130785465, + 0.014257395640015602, + 0.005728164687752724, + 0.0044844732619822025, + 0.005021002609282732, + 0.00587781798094511, + 0.0035325183998793364, + 0.02362283691763878, + 0.004742046352475882, + 0.023678716272115707, + 0.01069553941488266, + 0.011577156372368336, + 0.013688972219824791, + 0.01808156631886959, + 0.013777242973446846, + 0.035400133579969406, + 0.029246557503938675, + 0.03733503445982933, + 0.0036048097535967827, + 0.003775793593376875, + 0.004853859543800354, + 0.005407068412750959, + 0.005384143907576799, + 0.023182356730103493, + 0.008869124576449394, + 0.02396067976951599, + 0.009215706959366798, + 0.008833101950585842, + 0.009538223035633564, + 0.010787060484290123, + 0.007777174469083548, + 0.009250599890947342, + 0.006790038663893938, + 0.011677101254463196, + 0.003095427295193076, + 0.0037320717237889767, + 0.004046596586704254, + 0.004608278162777424, + 0.00356730492785573, + 0.005113318562507629, + 0.0054764654487371445, + 0.0060697393491864204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.02638009935617447 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005265297368168831 + ], + "opposite_logprob_tokens": [ + -5.255265235900879 + ], + "logprob_reward": -0.005265297368168831, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005265297368168831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3617249131202698, + 0.07767169177532196, + 0.05665237084031105, + 0.07462674379348755, + 0.04184921458363533, + 0.07625219225883484, + 0.07054443657398224, + 0.08732565492391586, + 0.06693623960018158, + 0.08641652762889862 + ], + "credit_map": [ + 0.02611318603157997, + 0.0341327004134655, + 0.041438229382038116, + 0.06635943800210953, + 0.037551697343587875, + 0.04999808967113495, + 0.04558582976460457, + 0.060545746237039566, + 0.0056655206717550755, + 0.007958578877151012, + 0.009321490302681923, + 0.010017535649240017, + 0.015279613435268402, + 0.009841643273830414, + 0.011351557448506355, + 0.008235747925937176, + 0.005168667063117027, + 0.00408639432862401, + 0.007861054502427578, + 0.00906333513557911, + 0.0052687437273561954, + 0.0081699900329113, + 0.008634146302938461, + 0.008400038816034794, + 0.004789500497281551, + 0.005091981031000614, + 0.007518880069255829, + 0.009727292694151402, + 0.0054261405020952225, + 0.01572035439312458, + 0.008347253315150738, + 0.01800534687936306, + 0.0016988218994811177, + 0.0022994372993707657, + 0.0024777345824986696, + 0.002707314444705844, + 0.002359134843572974, + 0.013533608987927437, + 0.003296556184068322, + 0.013476603664457798, + 0.0029366803355515003, + 0.003668423742055893, + 0.00795870739966631, + 0.010792933404445648, + 0.0104985935613513, + 0.011711226776242256, + 0.014311271719634533, + 0.014374357648193836, + 0.002985042752698064, + 0.0025595040060579777, + 0.0058736978098750114, + 0.00777216162532568, + 0.006475729867815971, + 0.014382991008460522, + 0.015404297970235348, + 0.015091012232005596, + 0.0031089261174201965, + 0.003479254199191928, + 0.0047773076221346855, + 0.006493463646620512, + 0.004788978490978479, + 0.029358983039855957, + 0.005583265796303749, + 0.029735472053289413, + 0.0047177961096167564, + 0.004364748485386372, + 0.007245767395943403, + 0.009991162456572056, + 0.006831629667431116, + 0.011603397317230701, + 0.005859823897480965, + 0.01632191427052021, + 0.00861132238060236, + 0.0085889408364892, + 0.009448024444282055, + 0.013476422987878323, + 0.007197331637144089, + 0.015864327549934387, + 0.007085585035383701, + 0.016144566237926483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.005265297368168831 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707051634788513 + ], + "opposite_logprob_tokens": [ + -2.127070426940918 + ], + "logprob_reward": -0.12707051634788513, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12707051634788513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0960140973329544, + 0.07747814059257507, + 0.05026337131857872, + 0.12552784383296967, + 0.07765953987836838, + 0.18813258409500122, + 0.10245393961668015, + 0.12054698914289474, + 0.08697034418582916, + 0.0749531090259552 + ], + "credit_map": [ + 0.008229169994592667, + 0.009538345038890839, + 0.013572760857641697, + 0.018063601106405258, + 0.006881774868816137, + 0.013970513828098774, + 0.0099661098793149, + 0.01579183340072632, + 0.004396946169435978, + 0.00499966973438859, + 0.009537087753415108, + 0.013453694991767406, + 0.0076961759477853775, + 0.014573859050869942, + 0.008183637633919716, + 0.014637075364589691, + 0.004644347354769707, + 0.004387926775962114, + 0.007028082385659218, + 0.008560544811189175, + 0.004104964435100555, + 0.007064081262797117, + 0.006641998887062073, + 0.007831432856619358, + 0.006492646411061287, + 0.007900684140622616, + 0.01023799553513527, + 0.014476196840405464, + 0.010724383406341076, + 0.0330294594168663, + 0.0083195511251688, + 0.03434694558382034, + 0.004232927691191435, + 0.00538993626832962, + 0.008017715066671371, + 0.010189072228968143, + 0.006038131192326546, + 0.017265360802412033, + 0.008830070495605469, + 0.017696332186460495, + 0.008107894100248814, + 0.010054923593997955, + 0.018273277208209038, + 0.026104988530278206, + 0.025108205154538155, + 0.029124386608600616, + 0.037212468683719635, + 0.03414645045995712, + 0.005483634769916534, + 0.005751262418925762, + 0.008314630016684532, + 0.011396045796573162, + 0.008784947916865349, + 0.023004982620477676, + 0.015844345092773438, + 0.023874104022979736, + 0.007616602815687656, + 0.006885140668600798, + 0.01012029591947794, + 0.013765660114586353, + 0.006913044489920139, + 0.03342188894748688, + 0.007314564660191536, + 0.0345098078250885, + 0.007849293760955334, + 0.008288423530757427, + 0.009509490802884102, + 0.012240728363394737, + 0.009747426956892014, + 0.01206472609192133, + 0.009885355830192566, + 0.017384905368089676, + 0.008177011273801327, + 0.008777244947850704, + 0.007841618731617928, + 0.011122514493763447, + 0.008650721050798893, + 0.011310863308608532, + 0.007544511929154396, + 0.01152863446623087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 7 + ], + "selected_reward": -0.12707051634788513 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42897099256515503 + ], + "opposite_logprob_tokens": [ + -1.0539710521697998 + ], + "logprob_reward": -0.42897099256515503, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.42897099256515503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1065012663602829, + 0.06392757594585419, + 0.0608244426548481, + 0.0993647575378418, + 0.06573443114757538, + 0.08763234317302704, + 0.09490659832954407, + 0.17695608735084534, + 0.09953727573156357, + 0.14461518824100494 + ], + "credit_map": [ + 0.008699529804289341, + 0.009773263707756996, + 0.012341619469225407, + 0.017046773806214333, + 0.011253868229687214, + 0.018407313153147697, + 0.009382582269608974, + 0.019596315920352936, + 0.003730283584445715, + 0.005546342581510544, + 0.00778660224750638, + 0.011186127550899982, + 0.006652272772043943, + 0.010515598580241203, + 0.007812689989805222, + 0.01069766003638506, + 0.00444800266996026, + 0.004431955050677061, + 0.008524660021066666, + 0.00896576140075922, + 0.004362230654805899, + 0.010601775720715523, + 0.007950360886752605, + 0.011539705097675323, + 0.003768160007894039, + 0.003145421389490366, + 0.007646753918379545, + 0.010137457400560379, + 0.007068187464028597, + 0.030611660331487656, + 0.005860032048076391, + 0.031127087771892548, + 0.0026347499806433916, + 0.002274158177897334, + 0.003855682210996747, + 0.005154036916792393, + 0.002993028610944748, + 0.021892594173550606, + 0.004585570655763149, + 0.02234460785984993, + 0.007001892663538456, + 0.005605625454336405, + 0.010353663004934788, + 0.01460410188883543, + 0.012126367539167404, + 0.010583830066025257, + 0.013727396726608276, + 0.013629480265080929, + 0.005616714712232351, + 0.005620642099529505, + 0.007395949214696884, + 0.009841938503086567, + 0.0075798737816512585, + 0.022150855511426926, + 0.013808506540954113, + 0.022892121225595474, + 0.006747541017830372, + 0.006752369925379753, + 0.01077235583215952, + 0.014565792866051197, + 0.006950043141841888, + 0.05881910398602486, + 0.01221842598170042, + 0.06013047695159912, + 0.009148787707090378, + 0.009365399368107319, + 0.011192920617759228, + 0.012448328547179699, + 0.0100112808868289, + 0.01567695662379265, + 0.015671413391828537, + 0.016022199764847755, + 0.012994088232517242, + 0.01431373879313469, + 0.013774403370916843, + 0.01975654438138008, + 0.01293737068772316, + 0.028574902564287186, + 0.012556538917124271, + 0.029707621783018112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.42897099256515503 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot captures the tip of a weathered brass spout, its surface a tapestry of aged tapestry, a dull metallic sheen punctuated by green patina of oxidized copper. A cool, diffuse light illuminates the scene, creating soft, crisp highlights on the metal and a single, shimmering droplet of water clinging to the spout. Pulled by the unstoppable force of gravity, the droplet slowly and deliberately elongates, stretching downward into a teardrop shape, its surface tension visibly strained. With one final, clean break, it separates from the brass and quickly falls from the bottom of the frame, leaving the spout momentarily exposed. A new, smaller droplet immediately begins to expand and form, continuing a quiet, rhythmic cycle amidst a serene and focused atmosphere.", + "seed": 19025, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_5/index_6194_seed_19025_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.026922141201794147, + "vlm_logprob": -0.026922141201794147, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.026922141201794147, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008438407443463802 + ], + "opposite_logprob_tokens": [ + -6.008438587188721 + ], + "logprob_reward": -0.008438407443463802, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.008438407443463802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1903589963912964, + 0.10585055500268936, + 0.05682807043194771, + 0.04565364494919777, + 0.05458569899201393, + 0.0717371478676796, + 0.07487417757511139, + 0.12749440968036652, + 0.1809815615415573, + 0.09163583815097809 + ], + "credit_map": [ + 0.01878347434103489, + 0.011170081794261932, + 0.03468678519129753, + 0.04670269787311554, + 0.019359061494469643, + 0.01790856570005417, + 0.024814141914248466, + 0.016934162005782127, + 0.013658705167472363, + 0.006873135920614004, + 0.018027862533926964, + 0.027585485950112343, + 0.0069169653579592705, + 0.005714879836887121, + 0.017704177647829056, + 0.009369328618049622, + 0.008029312826693058, + 0.004838166292756796, + 0.007185841910541058, + 0.009783743880689144, + 0.0061650993302464485, + 0.00714635057374835, + 0.005345896817743778, + 0.008333656005561352, + 0.004800351336598396, + 0.003918670583516359, + 0.0059005445800721645, + 0.007077462505549192, + 0.008187638595700264, + 0.005398656241595745, + 0.005282656755298376, + 0.005087660159915686, + 0.007489086594432592, + 0.004573214333504438, + 0.007801440078765154, + 0.009694294072687626, + 0.007022164296358824, + 0.0037591555155813694, + 0.00900771003216505, + 0.005238630343228579, + 0.010555833578109741, + 0.006983333267271519, + 0.01133965328335762, + 0.015828732401132584, + 0.006361799314618111, + 0.005617818329483271, + 0.008745159022510052, + 0.006304810754954815, + 0.006700127851217985, + 0.006998027674853802, + 0.010234047658741474, + 0.013581151142716408, + 0.008799416013062, + 0.007731258869171143, + 0.010736104100942612, + 0.010094036348164082, + 0.014562675729393959, + 0.01477128453552723, + 0.027735361829400063, + 0.025607768446207047, + 0.014474394731223583, + 0.007904269732534885, + 0.012237455695867538, + 0.010201195254921913, + 0.025241557508707047, + 0.01208515278995037, + 0.027549346908926964, + 0.03022368811070919, + 0.011702951975166798, + 0.010307827964425087, + 0.02949490211904049, + 0.0343761146068573, + 0.01187377143651247, + 0.005134928040206432, + 0.02397722192108631, + 0.020377954468131065, + 0.004444790072739124, + 0.0072507490403950214, + 0.012091735377907753, + 0.006484682206064463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.008438407443463802 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208698645234108 + ], + "opposite_logprob_tokens": [ + -4.264208793640137 + ], + "logprob_reward": -0.014208698645234108, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208698645234108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19406548142433167, + 0.10674885660409927, + 0.08014023303985596, + 0.04414758086204529, + 0.11332198232412338, + 0.09238500893115997, + 0.04391317442059517, + 0.07498271763324738, + 0.10733117163181305, + 0.14296382665634155 + ], + "credit_map": [ + 0.029467055574059486, + 0.013668599538505077, + 0.02700679376721382, + 0.03668892756104469, + 0.023898178711533546, + 0.015341570600867271, + 0.03138601779937744, + 0.01660832390189171, + 0.018254833295941353, + 0.006292989011853933, + 0.019173514097929, + 0.027483846992254257, + 0.007238695863634348, + 0.008867388591170311, + 0.010511116124689579, + 0.008926470763981342, + 0.014166324399411678, + 0.006480831187218428, + 0.013085893355309963, + 0.01798567734658718, + 0.007133643142879009, + 0.006604365073144436, + 0.0075958166271448135, + 0.007087684702128172, + 0.005107170436531305, + 0.004569421522319317, + 0.00568670779466629, + 0.0069106752052903175, + 0.004431739915162325, + 0.005141715053468943, + 0.0069986144080758095, + 0.005301534663885832, + 0.01599496230483055, + 0.009095759131014347, + 0.018388664349913597, + 0.030401691794395447, + 0.008526762016117573, + 0.007738008629530668, + 0.011330249719321728, + 0.011845884844660759, + 0.010788201354444027, + 0.008145522326231003, + 0.014563933946192265, + 0.016389599069952965, + 0.012113374657928944, + 0.007978628389537334, + 0.011242031119763851, + 0.011163725517690182, + 0.004692241083830595, + 0.004562505055218935, + 0.0053649721667170525, + 0.006367904134094715, + 0.005884280428290367, + 0.00482086930423975, + 0.0064139110036194324, + 0.005806492641568184, + 0.00822847243398428, + 0.007389399688690901, + 0.010567102581262589, + 0.016416653990745544, + 0.0075671495869755745, + 0.006126568652689457, + 0.010424157604575157, + 0.008263216353952885, + 0.011388592422008514, + 0.007598017808049917, + 0.016633225604891777, + 0.02318352647125721, + 0.01194628607481718, + 0.007842833176255226, + 0.015722671523690224, + 0.013016019947826862, + 0.020236747339367867, + 0.009103753603994846, + 0.0278227087110281, + 0.025829099118709564, + 0.013709107413887978, + 0.014895438216626644, + 0.01932189054787159, + 0.012045091018080711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.014208698645234108 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00056429672986269 + ], + "opposite_logprob_tokens": [ + -7.500564098358154 + ], + "logprob_reward": -0.00056429672986269, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.00056429672986269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12101072818040848, + 0.052302587777376175, + 0.04490400478243828, + 0.028829574584960938, + 0.23942242562770844, + 0.06879134476184845, + 0.028401680290699005, + 0.05739127844572067, + 0.239144966006279, + 0.11980145424604416 + ], + "credit_map": [ + 0.014496274292469025, + 0.006468549836426973, + 0.022110724821686745, + 0.0321507453918457, + 0.009599377401173115, + 0.009148845449090004, + 0.0169247854501009, + 0.010111413896083832, + 0.004337238613516092, + 0.004551670514047146, + 0.00647012097761035, + 0.00807773508131504, + 0.005086355376988649, + 0.006034146063029766, + 0.013091512024402618, + 0.004653805401176214, + 0.006514016538858414, + 0.0036683815997093916, + 0.00712224654853344, + 0.008923225104808807, + 0.00447801174595952, + 0.004923027474433184, + 0.004235451575368643, + 0.00503963977098465, + 0.004209807142615318, + 0.0021901417057961226, + 0.003964999224990606, + 0.004357229918241501, + 0.0029662277083843946, + 0.003866632003337145, + 0.003909136634320021, + 0.003365397686138749, + 0.03349589556455612, + 0.019246850162744522, + 0.040051233023405075, + 0.06244610249996185, + 0.019769007340073586, + 0.01297956146299839, + 0.02763277105987072, + 0.023800978437066078, + 0.009924880228936672, + 0.0065657529048621655, + 0.0117751844227314, + 0.016521288082003593, + 0.004506435710936785, + 0.005276965908706188, + 0.0064925868064165115, + 0.007728249300271273, + 0.0037991884164512157, + 0.004002007190138102, + 0.00362015631981194, + 0.004241371527314186, + 0.0028789713978767395, + 0.0021339606028050184, + 0.004309183917939663, + 0.003416839288547635, + 0.0062627969309687614, + 0.004478401970118284, + 0.008980103768408298, + 0.01162499189376831, + 0.00895615667104721, + 0.005654935725033283, + 0.005209612660109997, + 0.006224275100976229, + 0.030163297429680824, + 0.01795457862317562, + 0.03466160222887993, + 0.040789101272821426, + 0.020660560578107834, + 0.012930824421346188, + 0.044383566826581955, + 0.03760140761733055, + 0.015376747585833073, + 0.007673643529415131, + 0.027639757841825485, + 0.02411978878080845, + 0.009663493372499943, + 0.009609435684978962, + 0.016955656930804253, + 0.008762913756072521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.00056429672986269 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073562316596508 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073562316596508, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073562316596508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05253876745700836, + 0.07057413458824158, + 0.045586198568344116, + 0.03167865797877312, + 0.4519413709640503, + 0.08844631165266037, + 0.029444070532917976, + 0.030211832374334335, + 0.12407296150922775, + 0.07550577819347382 + ], + "credit_map": [ + 0.00451930146664381, + 0.003221750259399414, + 0.009717484936118126, + 0.01035326812416315, + 0.006497695576399565, + 0.00573364645242691, + 0.006343399174511433, + 0.006152212619781494, + 0.0056592971086502075, + 0.005668346304446459, + 0.011008206754922867, + 0.013758731074631214, + 0.006067509762942791, + 0.006039734929800034, + 0.016123687848448753, + 0.006248604506254196, + 0.007780927699059248, + 0.003931354731321335, + 0.007834662683308125, + 0.008776708506047726, + 0.0043655261397361755, + 0.004124908242374659, + 0.004811055026948452, + 0.0039610471576452255, + 0.0033078701235353947, + 0.0032132985070347786, + 0.003926475066691637, + 0.00445743789896369, + 0.004904956556856632, + 0.002903444692492485, + 0.005444186273962259, + 0.003520984435454011, + 0.07105018198490143, + 0.029911182820796967, + 0.081564761698246, + 0.12983085215091705, + 0.026350999251008034, + 0.02137082815170288, + 0.0478387326002121, + 0.044023752212524414, + 0.011198252439498901, + 0.006859087385237217, + 0.013676059432327747, + 0.01590093784034252, + 0.011689601466059685, + 0.005411503836512566, + 0.013209543190896511, + 0.010501308366656303, + 0.002685127779841423, + 0.004711845424026251, + 0.003997955471277237, + 0.0042285253293812275, + 0.0034468539524823427, + 0.0030368748120963573, + 0.003688737517222762, + 0.0036481444258242846, + 0.003053050022572279, + 0.004275024402886629, + 0.005489378701895475, + 0.005850383546203375, + 0.003186766989529133, + 0.0024273258168250322, + 0.0025902702473104, + 0.0033396268263459206, + 0.012079968117177486, + 0.009288523346185684, + 0.016497058793902397, + 0.02127852849662304, + 0.009874587878584862, + 0.006991992238909006, + 0.02529989928007126, + 0.022762387990951538, + 0.007162644062191248, + 0.005460480693727732, + 0.013610794208943844, + 0.012184998951852322, + 0.008274377323687077, + 0.011427661404013634, + 0.010466932319104671, + 0.006917873397469521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.011073562316596508 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032574087381363 + ], + "opposite_logprob_tokens": [ + -2.3503258228302 + ], + "logprob_reward": -0.10032574087381363, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032574087381363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06304151564836502, + 0.026921093463897705, + 0.03599339723587036, + 0.024238020181655884, + 0.276839941740036, + 0.20490729808807373, + 0.04658467695116997, + 0.07192263007164001, + 0.1425543874502182, + 0.1069970354437828 + ], + "credit_map": [ + 0.007206586189568043, + 0.004424443002790213, + 0.013882503844797611, + 0.012932137586176395, + 0.00659519387409091, + 0.004711478482931852, + 0.00790735986083746, + 0.005381810013204813, + 0.002661402802914381, + 0.0019396735588088632, + 0.002953421324491501, + 0.004589793737977743, + 0.002645309781655669, + 0.0021158347371965647, + 0.007111785933375359, + 0.0029038728680461645, + 0.006384260021150112, + 0.0024028047919273376, + 0.005628990940749645, + 0.005902079399675131, + 0.003203650936484337, + 0.004031939432024956, + 0.003769379109144211, + 0.004670292604714632, + 0.003943399991840124, + 0.0017024835105985403, + 0.003505936125293374, + 0.003702443325892091, + 0.003828587243333459, + 0.0021147322840988636, + 0.0023595085367560387, + 0.003080929396674037, + 0.04313366115093231, + 0.017849920317530632, + 0.05107314512133598, + 0.08118674159049988, + 0.01690761372447014, + 0.013119744136929512, + 0.026521330699324608, + 0.02704780362546444, + 0.024792902171611786, + 0.01494667399674654, + 0.03565715625882149, + 0.048834413290023804, + 0.018863722681999207, + 0.01126003172248602, + 0.026036230847239494, + 0.024516161531209946, + 0.005061862990260124, + 0.00536700151860714, + 0.0074541387148201466, + 0.008805060759186745, + 0.005701400805264711, + 0.004111079499125481, + 0.005207929760217667, + 0.004876206628978252, + 0.007517901249229908, + 0.009173242375254631, + 0.012719602324068546, + 0.011407511308789253, + 0.009775116108357906, + 0.005176928360015154, + 0.008090618066489697, + 0.008061708882451057, + 0.01638042740523815, + 0.010218411684036255, + 0.020710250362753868, + 0.02565060183405876, + 0.014200557954609394, + 0.008246769197285175, + 0.02919483743607998, + 0.01795251853764057, + 0.011505547910928726, + 0.008354117162525654, + 0.02357523702085018, + 0.015036672353744507, + 0.014089779928326607, + 0.011430297046899796, + 0.013626201078295708, + 0.009379183873534203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.10032574087381363 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot captures the tip of a weathered brass spout, its surface a tapestry of aged tapestry, a dull metallic sheen punctuated by green patina of oxidized copper. A cool, diffuse light illuminates the scene, creating soft, crisp highlights on the metal and a single, shimmering droplet of water clinging to the spout. Pulled by the unstoppable force of gravity, the droplet slowly and deliberately elongates, stretching downward into a teardrop shape, its surface tension visibly strained. With one final, clean break, it separates from the brass and quickly falls from the bottom of the frame, leaving the spout momentarily exposed. A new, smaller droplet immediately begins to expand and form, continuing a quiet, rhythmic cycle amidst a serene and focused atmosphere.", + "seed": 19026, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_5/index_6194_seed_19026_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08833887833170592, + "vlm_logprob": -0.08833887833170592, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.08833887833170592, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006769698113203049 + ], + "opposite_logprob_tokens": [ + -6.50676965713501 + ], + "logprob_reward": -0.006769698113203049, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.006769698113203049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09588455408811569, + 0.19348427653312683, + 0.017820250242948532, + 0.05941447243094444, + 0.2834884226322174, + 0.040509115904569626, + 0.08554553240537643, + 0.0893755629658699, + 0.10522881150245667, + 0.02924901805818081 + ], + "credit_map": [ + 0.007372355088591576, + 0.009068540297448635, + 0.013418413698673248, + 0.0064436038956046104, + 0.005088658072054386, + 0.02759406901896, + 0.009934037923812866, + 0.016964873299002647, + 0.008157826028764248, + 0.00590835465118289, + 0.019621919840574265, + 0.008775308728218079, + 0.009077110327780247, + 0.06686124950647354, + 0.01696535013616085, + 0.05811714753508568, + 0.00176770961843431, + 0.0018491008086130023, + 0.0021204776130616665, + 0.002107510808855295, + 0.0013894265284761786, + 0.002445048652589321, + 0.0025625950656831264, + 0.0035783809144049883, + 0.002488688100129366, + 0.00203500222414732, + 0.005583473946899176, + 0.0024272070731967688, + 0.0024195825681090355, + 0.02057747170329094, + 0.004279453307390213, + 0.01960359327495098, + 0.009423429146409035, + 0.0045752148143947124, + 0.02351764217019081, + 0.009406591765582561, + 0.008687070570886135, + 0.11094369739294052, + 0.009922266006469727, + 0.10701248794794083, + 0.0014997117687016726, + 0.0018221988575533032, + 0.0031152782030403614, + 0.0022041143383830786, + 0.0020300974138081074, + 0.013460434041917324, + 0.0033840476535260677, + 0.012993233278393745, + 0.0087844617664814, + 0.007934077642858028, + 0.010504291392862797, + 0.010944019071757793, + 0.005597023293375969, + 0.011065671220421791, + 0.019227109849452972, + 0.011488874442875385, + 0.0035258878488093615, + 0.002025546971708536, + 0.009023175574839115, + 0.005288056097924709, + 0.0030553096439689398, + 0.025499410927295685, + 0.011934044770896435, + 0.02902413159608841, + 0.0021185139194130898, + 0.0017471658065915108, + 0.00856294296681881, + 0.0027748544234782457, + 0.0021698065102100372, + 0.04276207834482193, + 0.005516791716217995, + 0.039576657116413116, + 0.0015495793195441365, + 0.0017361914506182075, + 0.00336232571862638, + 0.0024576117284595966, + 0.0012369423639029264, + 0.0071349553763866425, + 0.004273754078894854, + 0.007497655693441629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 8 + ], + "selected_reward": -0.006769698113203049 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007646339945495129 + ], + "opposite_logprob_tokens": [ + -4.882646560668945 + ], + "logprob_reward": -0.007646339945495129, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007646339945495129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15038518607616425, + 0.1635265350341797, + 0.1348789632320404, + 0.0822964683175087, + 0.12358450144529343, + 0.04925376921892166, + 0.12853790819644928, + 0.022986292839050293, + 0.09554023295640945, + 0.04901013895869255 + ], + "credit_map": [ + 0.012184545397758484, + 0.016466690227389336, + 0.016525806859135628, + 0.011610089801251888, + 0.009900405071675777, + 0.02835073322057724, + 0.017381377518177032, + 0.037965528666973114, + 0.011251023970544338, + 0.015405124984681606, + 0.01721654087305069, + 0.007871517911553383, + 0.008806833066046238, + 0.04518139734864235, + 0.013892839662730694, + 0.04390125349164009, + 0.005452250130474567, + 0.0065377745777368546, + 0.012505260296165943, + 0.006386231165379286, + 0.0044321115128695965, + 0.04813820496201515, + 0.008605726063251495, + 0.04282141476869583, + 0.005434422288089991, + 0.004674394149333239, + 0.009111479856073856, + 0.005198960658162832, + 0.0044263508170843124, + 0.02561139687895775, + 0.008655449375510216, + 0.019184013828635216, + 0.005538594909012318, + 0.005805603228509426, + 0.01045489776879549, + 0.005250745452940464, + 0.0064459629356861115, + 0.04290448874235153, + 0.007496565114706755, + 0.039687640964984894, + 0.0035673838574439287, + 0.0037889042869210243, + 0.004034500103443861, + 0.0030344713013619184, + 0.004230811260640621, + 0.012868350371718407, + 0.004902430344372988, + 0.012826921418309212, + 0.008246056735515594, + 0.006471752189099789, + 0.015462027862668037, + 0.014578159898519516, + 0.005670909304171801, + 0.027196012437343597, + 0.024480199441313744, + 0.026432787999510765, + 0.0013329023495316505, + 0.0016929196426644921, + 0.0034340417478233576, + 0.002567172748968005, + 0.0012631636345759034, + 0.0035870240535587072, + 0.005471576936542988, + 0.003637491725385189, + 0.0025555763859301805, + 0.0024421196430921555, + 0.008285054005682468, + 0.002377243246883154, + 0.002074936870485544, + 0.03859167546033859, + 0.0060705626383423805, + 0.0331430621445179, + 0.0019938419573009014, + 0.0019539997447282076, + 0.004339815583080053, + 0.0028918355237692595, + 0.0022553554736077785, + 0.015542539767920971, + 0.0044286856427788734, + 0.015604064799845219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.007646339945495129 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008201810996979475 + ], + "opposite_logprob_tokens": [ + -7.125820159912109 + ], + "logprob_reward": -0.0008201810996979475, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008201810996979475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07452523708343506, + 0.049260277301073074, + 0.09273593872785568, + 0.02867957018315792, + 0.2412087470293045, + 0.11785522103309631, + 0.11415258049964905, + 0.20023325085639954, + 0.04941316694021225, + 0.031936030834913254 + ], + "credit_map": [ + 0.0067222099751234055, + 0.006452732719480991, + 0.009493284858763218, + 0.00887179747223854, + 0.010572795756161213, + 0.011942473240196705, + 0.011513498611748219, + 0.008956441655755043, + 0.0030492895748466253, + 0.002464490244165063, + 0.005475118290632963, + 0.002323512453585863, + 0.003247986314818263, + 0.014835162088274956, + 0.005156381521373987, + 0.012708337977528572, + 0.002900252118706703, + 0.0019028435926884413, + 0.008687916211783886, + 0.0032436756882816553, + 0.0029312409460544586, + 0.03619418665766716, + 0.005575635936111212, + 0.03130018338561058, + 0.0026454224716871977, + 0.001946987584233284, + 0.0037574388552457094, + 0.0037317173555493355, + 0.0017560700653120875, + 0.0037109986878931522, + 0.006556474603712559, + 0.004574459977447987, + 0.006458139978349209, + 0.005487654823809862, + 0.02041253075003624, + 0.013970287516713142, + 0.006691655609756708, + 0.08847130835056305, + 0.013540099374949932, + 0.08617707341909409, + 0.005442282650619745, + 0.003671616315841675, + 0.010207860730588436, + 0.0069098020903766155, + 0.00719938799738884, + 0.03515315055847168, + 0.013274709694087505, + 0.035996418446302414, + 0.008901815861463547, + 0.007736396975815296, + 0.01639745570719242, + 0.014733298681676388, + 0.006725303363054991, + 0.015169832855463028, + 0.03011825494468212, + 0.014370213262736797, + 0.007754737976938486, + 0.004508272744715214, + 0.01814158260822296, + 0.011312323622405529, + 0.006431219633668661, + 0.06205739453434944, + 0.02324659191071987, + 0.06678114086389542, + 0.00240321340970695, + 0.002633659401908517, + 0.004179404582828283, + 0.0026911990717053413, + 0.002009103773161769, + 0.015223264694213867, + 0.0052840267308056355, + 0.014989296905696392, + 0.0015205530216917396, + 0.0019053897121921182, + 0.002948771696537733, + 0.0021787865553051233, + 0.0016330615617334843, + 0.009998324327170849, + 0.0027862528804689646, + 0.008964892476797104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.0008201810996979475 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294128745794296 + ], + "opposite_logprob_tokens": [ + -2.237941265106201 + ], + "logprob_reward": -0.11294128745794296, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294128745794296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05906108021736145, + 0.14854739606380463, + 0.0499153807759285, + 0.03395039215683937, + 0.1143040880560875, + 0.2917625606060028, + 0.15338486433029175, + 0.05395740270614624, + 0.07480709999799728, + 0.020309794694185257 + ], + "credit_map": [ + 0.005550494883209467, + 0.006453350652009249, + 0.007349834311753511, + 0.00519828125834465, + 0.005583367310464382, + 0.010015604086220264, + 0.007428300566971302, + 0.011481834575533867, + 0.007056402042508125, + 0.005054543260484934, + 0.015448837541043758, + 0.008458808995783329, + 0.00712103396654129, + 0.04960721731185913, + 0.014375953003764153, + 0.04142456129193306, + 0.0024245998356491327, + 0.0019173778127878904, + 0.0043737743981182575, + 0.0016654786886647344, + 0.0019260670524090528, + 0.018205996602773666, + 0.003233300754800439, + 0.016168775036931038, + 0.002258880063891411, + 0.002125436905771494, + 0.004023847170174122, + 0.003972321283072233, + 0.0021293654572218657, + 0.007418501190841198, + 0.0055819302797317505, + 0.006440106313675642, + 0.007182396948337555, + 0.005071129649877548, + 0.013371352106332779, + 0.01041855476796627, + 0.005965008400380611, + 0.03232153132557869, + 0.013631870970129967, + 0.026342235505580902, + 0.012043073773384094, + 0.007105773780494928, + 0.025132564827799797, + 0.015843557193875313, + 0.015399180352687836, + 0.09101730585098267, + 0.030854662880301476, + 0.0943664163351059, + 0.0087925735861063, + 0.007353353779762983, + 0.014510387554764748, + 0.01233590766787529, + 0.007067128550261259, + 0.03937967121601105, + 0.02442503720521927, + 0.039520785212516785, + 0.004487572703510523, + 0.004357369616627693, + 0.008961883373558521, + 0.00571675319224596, + 0.0028138242196291685, + 0.005921403411775827, + 0.013499284163117409, + 0.008199305273592472, + 0.002747452585026622, + 0.0027431182097643614, + 0.0057297335006296635, + 0.0028591537848114967, + 0.0020929959136992693, + 0.027068834751844406, + 0.005826130975037813, + 0.025739675387740135, + 0.0017048848094418645, + 0.0019434954738244414, + 0.002671665046364069, + 0.0023356794845312834, + 0.001750965602695942, + 0.003434064332395792, + 0.002996204188093543, + 0.0034728317987173796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.11294128745794296 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the tip of a weathered brass spout?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the brass spout surface show a dull metallic sheen with green patina of oxidized copper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single water droplet visible clinging to the spout and elongating into a teardrop shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet separate from the spout and fall out of the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a new, smaller droplet immediately begin to form on the spout after the first one falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31351688504219055 + ], + "opposite_logprob_tokens": [ + -1.3135168552398682 + ], + "logprob_reward": -0.31351688504219055, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31351688504219055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07597161829471588, + 0.05010657012462616, + 0.045521657913923264, + 0.09983186423778534, + 0.11085914075374603, + 0.16451717913150787, + 0.1508195847272873, + 0.1008865237236023, + 0.15723435580730438, + 0.04425155371427536 + ], + "credit_map": [ + 0.0058783963322639465, + 0.004645376931875944, + 0.008467991836369038, + 0.006982023362070322, + 0.008223130367696285, + 0.018776286393404007, + 0.008908729068934917, + 0.014089680276811123, + 0.0031248244922608137, + 0.0036538466811180115, + 0.0051857600919902325, + 0.00379841192625463, + 0.002850144635885954, + 0.012776251882314682, + 0.005523218773305416, + 0.013194108381867409, + 0.002196002984419465, + 0.00170711160171777, + 0.004161587450653315, + 0.002144928788766265, + 0.00233917823061347, + 0.014712493866682053, + 0.003707402152940631, + 0.014552954584360123, + 0.009075259789824486, + 0.006206604186445475, + 0.013087822124361992, + 0.010688100010156631, + 0.006843882147222757, + 0.015026949346065521, + 0.022554872557520866, + 0.016348378732800484, + 0.007821951992809772, + 0.004581325687468052, + 0.012959555722773075, + 0.012108900584280491, + 0.008024244569242, + 0.02082880586385727, + 0.02346552163362503, + 0.021068837493658066, + 0.008472115732729435, + 0.004989274777472019, + 0.01416748482733965, + 0.009091420099139214, + 0.009320802986621857, + 0.05139594152569771, + 0.015456587076187134, + 0.05162355676293373, + 0.008245679549872875, + 0.008803299628198147, + 0.013486228883266449, + 0.009035295806825161, + 0.006468010600656271, + 0.044692933559417725, + 0.016316665336489677, + 0.043771468102931976, + 0.004257924389094114, + 0.00455475552007556, + 0.009648989886045456, + 0.005910002160817385, + 0.0036446894519031048, + 0.02715413086116314, + 0.013645677827298641, + 0.032070353627204895, + 0.005368393380194902, + 0.005560808815062046, + 0.012006024830043316, + 0.005062980577349663, + 0.004349668510258198, + 0.05997181683778763, + 0.009285642765462399, + 0.05562901869416237, + 0.004179887939244509, + 0.004411112982779741, + 0.005537799093872309, + 0.004771056119352579, + 0.003761114552617073, + 0.007277670782059431, + 0.005985038820654154, + 0.00832787062972784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.31351688504219055 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_6/index_tensor([326, 326])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_6/index_tensor([326, 326])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b40872b0de87d73fc9fb8223b654e7260ef50032 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_6/index_tensor([326, 326])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide-angle shot establishes a spacious yoga studio bathed in the golden glow of afternoon sunlight filtering through large arched windows, then the camera begins a slow and graceful pan to the right, capturing the scene. Warm light illuminates countless motes of dust dancing in the visible sunlight that spreads across the highly polished, honey-toned wooden floor, while tall potted ferns with lush, vibrant green foliage add a touch of nature to the serene space. At the front of the room, a tall, graceful giraffe instructor, dressed in a simple lavender tank top and yoga pants, gracefully demonstrates a pose, her long neck curved in a gentle arc, a serene smile on her face. A variety of anthropomorphic animals—including a slender fox in a warrior pose and a plump rabbit in a child's pose—are spread across colorful yoga mats, quietly holding their poses. As the camera continues, a furry little brown bear cub peeks hesitantly from the doorway, clutching a rolled-up blue yoga mat. His large, expressive eyes scan the room, and upon finding an empty spot, he cautiously and adorably clumsily tips in. Unfolding his mat, his expression determined but unsteady, he attempts to join the group, striking a precarious downward dog pose. The entire scene is rendered in the vibrant and captivating Pixar style, characterized by soft, saturated colors, detailed textures in the animals' fur, and an atmosphere that blends gentle humor with focused tranquility.", + "seed": 21187, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_6/index_326_seed_21187_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.008825672045350074, + "vlm_logprob": -0.008825672045350074, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.008825672045350074, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006989315617829561 + ], + "opposite_logprob_tokens": [ + -6.881989479064941 + ], + "logprob_reward": -0.006989315617829561, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.006989315617829561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1941005438566208, + 0.1619640588760376, + 0.0669240653514862, + 0.06822487711906433, + 0.040126003324985504, + 0.06682302057743073, + 0.14527806639671326, + 0.09466484189033508, + 0.06477760523557663, + 0.09711690992116928 + ], + "credit_map": [ + 0.019312886521220207, + 0.016230948269367218, + 0.028329508379101753, + 0.020561659708619118, + 0.023947661742568016, + 0.03893698751926422, + 0.015422612428665161, + 0.03135829046368599, + 0.013711914420127869, + 0.043004222214221954, + 0.011562258936464787, + 0.014624432660639286, + 0.013938832096755505, + 0.018537307158112526, + 0.012750338762998581, + 0.03383474051952362, + 0.006125917192548513, + 0.012694737873971462, + 0.006700115278363228, + 0.011870997957885265, + 0.005022154655307531, + 0.00663405004888773, + 0.009029323235154152, + 0.0088467663154006, + 0.008485493250191212, + 0.006886395160108805, + 0.010352294892072678, + 0.011243505403399467, + 0.008443409577012062, + 0.0094298105686903, + 0.0074381958693265915, + 0.005945777054876089, + 0.004367728717625141, + 0.004394188988953829, + 0.003980838693678379, + 0.00501006655395031, + 0.009466210380196571, + 0.0037833398673683405, + 0.00467115780338645, + 0.004452474880963564, + 0.008246153593063354, + 0.008571760728955269, + 0.011387414298951626, + 0.008162029087543488, + 0.005253504030406475, + 0.0075711593963205814, + 0.00835736095905304, + 0.009273632429540157, + 0.029859507456421852, + 0.018782932311296463, + 0.026323864236474037, + 0.01853470876812935, + 0.012120306491851807, + 0.014215357601642609, + 0.010355108417570591, + 0.015086277388036251, + 0.023021122440695763, + 0.008498918265104294, + 0.010354413650929928, + 0.011843560263514519, + 0.011283403262495995, + 0.013655639253556728, + 0.00600188784301281, + 0.010005896911025047, + 0.011150181293487549, + 0.005949843674898148, + 0.010423867031931877, + 0.008339866064488888, + 0.008292949758470058, + 0.006219023838639259, + 0.008166583254933357, + 0.0062352875247597694, + 0.009152550250291824, + 0.015892216935753822, + 0.011067958548665047, + 0.008608442731201649, + 0.00974282342940569, + 0.02069883979856968, + 0.010048050433397293, + 0.011906026862561703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.006989315617829561 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734057795256376 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.006734057795256376, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734057795256376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1982974261045456, + 0.1852739304304123, + 0.10494573414325714, + 0.11870754510164261, + 0.03249668329954147, + 0.06714021414518356, + 0.0861978754401207, + 0.02296726033091545, + 0.06940139085054398, + 0.11457191407680511 + ], + "credit_map": [ + 0.020351586863398552, + 0.015821276232600212, + 0.026037132367491722, + 0.02449730411171913, + 0.03104390762746334, + 0.03952634334564209, + 0.01641194522380829, + 0.024607954546809196, + 0.012725383043289185, + 0.06037220358848572, + 0.009241723455488682, + 0.015762489289045334, + 0.052456751465797424, + 0.012804935686290264, + 0.010720616206526756, + 0.01118983794003725, + 0.011674800887703896, + 0.02043759636580944, + 0.008524691686034203, + 0.017284514382481575, + 0.009452901780605316, + 0.01157945953309536, + 0.012100331485271454, + 0.01389144454151392, + 0.011872224509716034, + 0.028317758813500404, + 0.010784391313791275, + 0.020779116079211235, + 0.010031829588115215, + 0.009625683538615704, + 0.013635444454848766, + 0.013661104254424572, + 0.0033917040564119816, + 0.00395748158916831, + 0.003496425459161401, + 0.0024771238677203655, + 0.008146204054355621, + 0.003783139633014798, + 0.0038894282188266516, + 0.0033551796805113554, + 0.007876683957874775, + 0.007520487066358328, + 0.012174161151051521, + 0.006884758826345205, + 0.008070939220488071, + 0.007103718817234039, + 0.007422447670251131, + 0.010087022557854652, + 0.008134203962981701, + 0.009228314273059368, + 0.0072218081913888454, + 0.014282877556979656, + 0.015136847272515297, + 0.014983835630118847, + 0.0082942433655262, + 0.00891574565321207, + 0.002675386378541589, + 0.003701169975101948, + 0.00258967699483037, + 0.0024368194863200188, + 0.0035041042137891054, + 0.002798155415803194, + 0.0024459667038172483, + 0.0028159827925264835, + 0.007384845521301031, + 0.006557315122336149, + 0.010287423618137836, + 0.009525160305202007, + 0.009031346999108791, + 0.005515124648809433, + 0.014176568947732449, + 0.006923605687916279, + 0.01027142908424139, + 0.01807516999542713, + 0.015015517361462116, + 0.010026533156633377, + 0.010704295709729195, + 0.02254047989845276, + 0.009357299655675888, + 0.018581198528409004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.006734057795256376 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605464331805706 + ], + "opposite_logprob_tokens": [ + -4.141054630279541 + ], + "logprob_reward": -0.01605464331805706, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605464331805706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13596785068511963, + 0.16225548088550568, + 0.03995368629693985, + 0.04213868826627731, + 0.03815499693155289, + 0.05195965617895126, + 0.10116313397884369, + 0.05790995806455612, + 0.05347391963005066, + 0.3170226514339447 + ], + "credit_map": [ + 0.024372711777687073, + 0.015965910628437996, + 0.015997005626559258, + 0.016141192987561226, + 0.024376628920435905, + 0.016700323671102524, + 0.009924674406647682, + 0.012489397078752518, + 0.009994869120419025, + 0.052947815507650375, + 0.007971919141709805, + 0.010698896832764149, + 0.04868931695818901, + 0.01123995240777731, + 0.011517515406012535, + 0.009195196442306042, + 0.003781980136409402, + 0.005320328287780285, + 0.003929599188268185, + 0.0043334076181054115, + 0.004585515707731247, + 0.0049299513921141624, + 0.004996392410248518, + 0.00807651225477457, + 0.00495869992300868, + 0.00621302193030715, + 0.004668900743126869, + 0.0068154847249388695, + 0.004570335149765015, + 0.004206142388284206, + 0.005356109235435724, + 0.005349994637072086, + 0.0041188085451722145, + 0.005142133217304945, + 0.003462911583483219, + 0.003184031695127487, + 0.006653440650552511, + 0.005892285145819187, + 0.004211596213281155, + 0.005489793140441179, + 0.005357434507459402, + 0.004460989031940699, + 0.006635449361056089, + 0.007384962402284145, + 0.007393566891551018, + 0.010939919389784336, + 0.004839158151298761, + 0.0049481745809316635, + 0.015623167157173157, + 0.009891245514154434, + 0.01079401932656765, + 0.011607383377850056, + 0.016912726685404778, + 0.017915744334459305, + 0.007130063138902187, + 0.011288785375654697, + 0.010906378738582134, + 0.00832291878759861, + 0.006869389209896326, + 0.00587713997811079, + 0.006597656290978193, + 0.008094439283013344, + 0.005608630366623402, + 0.005633406341075897, + 0.007292891386896372, + 0.0073220087215304375, + 0.009292061440646648, + 0.007994581945240498, + 0.005477363709360361, + 0.005537957418709993, + 0.005251517053693533, + 0.00530553562566638, + 0.029025869444012642, + 0.06560046970844269, + 0.036872945725917816, + 0.028991185128688812, + 0.03553792089223862, + 0.033412277698516846, + 0.04665566235780716, + 0.040926333516836166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.01605464331805706 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006728373933583498 + ], + "opposite_logprob_tokens": [ + -5.006728172302246 + ], + "logprob_reward": -0.006728373933583498, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006728373933583498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09426628053188324, + 0.1608014702796936, + 0.04188307747244835, + 0.0510360449552536, + 0.03185756877064705, + 0.08017294108867645, + 0.11057614535093307, + 0.04729584604501724, + 0.1394013613462448, + 0.24270926415920258 + ], + "credit_map": [ + 0.014740398153662682, + 0.010115724056959152, + 0.01051670778542757, + 0.011522757820785046, + 0.013021118938922882, + 0.018096383661031723, + 0.008766134269535542, + 0.0074870530515909195, + 0.012357158586382866, + 0.042696110904216766, + 0.010513623245060444, + 0.01639813371002674, + 0.02824493870139122, + 0.01764579489827156, + 0.016145266592502594, + 0.016800444573163986, + 0.004314021673053503, + 0.005861472804099321, + 0.007027012761682272, + 0.0048574223183095455, + 0.0032657261472195387, + 0.005527340807020664, + 0.004214039538055658, + 0.006816042121499777, + 0.004152468405663967, + 0.013641996309161186, + 0.004319263156503439, + 0.00627449806779623, + 0.004430757835507393, + 0.0057450770400464535, + 0.006164808291941881, + 0.0063071767799556255, + 0.005663249175995588, + 0.0038890859577804804, + 0.0036959084682166576, + 0.003405724884942174, + 0.004197285044938326, + 0.0027465438470244408, + 0.005286990664899349, + 0.002972779795527458, + 0.010922410525381565, + 0.010199627839028835, + 0.012640123255550861, + 0.01256716251373291, + 0.0067257387563586235, + 0.00856585893779993, + 0.010661156848073006, + 0.007890860550105572, + 0.00854854192584753, + 0.016661280766129494, + 0.009028183296322823, + 0.01652262918651104, + 0.014622475020587444, + 0.021806852892041206, + 0.010259204544126987, + 0.013126973994076252, + 0.009552855044603348, + 0.005163051653653383, + 0.0066054570488631725, + 0.004762152675539255, + 0.004316483158618212, + 0.00545160798355937, + 0.006815050262957811, + 0.004629188682883978, + 0.014080272987484932, + 0.015011017210781574, + 0.018402643501758575, + 0.019478002563118935, + 0.01289242785423994, + 0.01227065734565258, + 0.024928105995059013, + 0.02233823388814926, + 0.020637469366192818, + 0.05351562425494194, + 0.029386157169938087, + 0.02541518025100231, + 0.02458495832979679, + 0.021399687975645065, + 0.036288030445575714, + 0.03148215636610985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.006728373933583498 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007621969562023878 + ], + "opposite_logprob_tokens": [ + -4.882621765136719 + ], + "logprob_reward": -0.007621969562023878, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007621969562023878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12983208894729614, + 0.16496491432189941, + 0.09992994368076324, + 0.0487954244017601, + 0.05025865510106087, + 0.0295411329716444, + 0.09668537974357605, + 0.1194654330611229, + 0.07116609811782837, + 0.18936093151569366 + ], + "credit_map": [ + 0.021180100739002228, + 0.012254605069756508, + 0.019533803686499596, + 0.015178682282567024, + 0.015252386219799519, + 0.022972550243139267, + 0.01029196660965681, + 0.01316799782216549, + 0.014208491891622543, + 0.043988537043333054, + 0.008513095788657665, + 0.016118545085191727, + 0.0242594126611948, + 0.022319283336400986, + 0.01115233451128006, + 0.024405213072896004, + 0.00939661730080843, + 0.01417427696287632, + 0.012754566967487335, + 0.012278755195438862, + 0.00878575537353754, + 0.013691600412130356, + 0.011522900313138962, + 0.017325472086668015, + 0.004057248122990131, + 0.010708209127187729, + 0.007038146257400513, + 0.006781126372516155, + 0.005321382079273462, + 0.00348618789575994, + 0.006048052106052637, + 0.005355071742087603, + 0.005994305945932865, + 0.008611741475760937, + 0.004951484967023134, + 0.007000948768109083, + 0.005374980624765158, + 0.004642820451408625, + 0.007594309747219086, + 0.006088060326874256, + 0.004396140109747648, + 0.00472270930185914, + 0.0034785931929945946, + 0.0022857561707496643, + 0.003578495467081666, + 0.003247288055717945, + 0.0038884927053004503, + 0.003943657968193293, + 0.010342041961848736, + 0.012863604351878166, + 0.008206554688513279, + 0.012992903590202332, + 0.015895450487732887, + 0.021752161905169487, + 0.005961058661341667, + 0.008671607822179794, + 0.018075915053486824, + 0.01686549186706543, + 0.015296092256903648, + 0.011672829277813435, + 0.025119449943304062, + 0.014289855025708675, + 0.007195560727268457, + 0.010950238443911076, + 0.013034471310675144, + 0.008457905612885952, + 0.01143441442400217, + 0.009679228998720646, + 0.005157198756933212, + 0.005343536380678415, + 0.008629554882645607, + 0.009429792873561382, + 0.020738644525408745, + 0.028229989111423492, + 0.019847914576530457, + 0.015721481293439865, + 0.017143994569778442, + 0.04368859529495239, + 0.015109017491340637, + 0.028881289064884186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.007621969562023878 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot establishes a spacious yoga studio bathed in the golden glow of afternoon sunlight filtering through large arched windows, then the camera begins a slow and graceful pan to the right, capturing the scene. Warm light illuminates countless motes of dust dancing in the visible sunlight that spreads across the highly polished, honey-toned wooden floor, while tall potted ferns with lush, vibrant green foliage add a touch of nature to the serene space. At the front of the room, a tall, graceful giraffe instructor, dressed in a simple lavender tank top and yoga pants, gracefully demonstrates a pose, her long neck curved in a gentle arc, a serene smile on her face. A variety of anthropomorphic animals—including a slender fox in a warrior pose and a plump rabbit in a child's pose—are spread across colorful yoga mats, quietly holding their poses. As the camera continues, a furry little brown bear cub peeks hesitantly from the doorway, clutching a rolled-up blue yoga mat. His large, expressive eyes scan the room, and upon finding an empty spot, he cautiously and adorably clumsily tips in. Unfolding his mat, his expression determined but unsteady, he attempts to join the group, striking a precarious downward dog pose. The entire scene is rendered in the vibrant and captivating Pixar style, characterized by soft, saturated colors, detailed textures in the animals' fur, and an atmosphere that blends gentle humor with focused tranquility.", + "seed": 21188, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_6/index_326_seed_21188_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.008373954333364964, + "vlm_logprob": -0.008373954333364964, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.008373954333364964, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008447864092886448 + ], + "opposite_logprob_tokens": [ + -5.258447647094727 + ], + "logprob_reward": -0.008447864092886448, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008447864092886448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20309659838676453, + 0.060368627309799194, + 0.15747804939746857, + 0.05551480874419212, + 0.1878318041563034, + 0.14014163613319397, + 0.14380276203155518, + 0.02415481209754944, + 0.013597558252513409, + 0.01401333138346672 + ], + "credit_map": [ + 0.022475840523838997, + 0.02465687319636345, + 0.02925569750368595, + 0.02451392449438572, + 0.019702693447470665, + 0.025361109524965286, + 0.026156587526202202, + 0.030973870307207108, + 0.005020217038691044, + 0.004027357324957848, + 0.013876216486096382, + 0.012588533572852612, + 0.002390325302258134, + 0.006381346844136715, + 0.009593515656888485, + 0.006491117179393768, + 0.012349274009466171, + 0.01741466484963894, + 0.025091737508773804, + 0.016706738620996475, + 0.029957842081785202, + 0.022429464384913445, + 0.014875280670821667, + 0.01865304633975029, + 0.00701161427423358, + 0.005566891748458147, + 0.0067601618357002735, + 0.0035117375664412975, + 0.00601579574868083, + 0.009719357825815678, + 0.009460212662816048, + 0.007469034753739834, + 0.024530703201889992, + 0.01623748615384102, + 0.015803979709744453, + 0.026460664346814156, + 0.01671064831316471, + 0.0199465099722147, + 0.04349898919463158, + 0.024642815813422203, + 0.007570615503937006, + 0.011069433763623238, + 0.021642714738845825, + 0.030607646331191063, + 0.012755533680319786, + 0.011015081778168678, + 0.020084138959646225, + 0.025396468117833138, + 0.01693607121706009, + 0.021452734246850014, + 0.016030751168727875, + 0.0162543635815382, + 0.024339823052287102, + 0.01612149178981781, + 0.020158106461167336, + 0.012509427033364773, + 0.002624993911013007, + 0.0020739964675158262, + 0.0032263684552162886, + 0.0032579940743744373, + 0.004328667651861906, + 0.002188366372138262, + 0.0023190355859696865, + 0.004135387483984232, + 0.0015559662133455276, + 0.0018365943105891347, + 0.002061209175735712, + 0.0018219489138573408, + 0.001875067362561822, + 0.0015279290964826941, + 0.0013930713757872581, + 0.0015257721533998847, + 0.0016820163000375032, + 0.002041702624410391, + 0.002255701459944248, + 0.0011600512079894543, + 0.002248060889542103, + 0.0015316720819100738, + 0.0015545287169516087, + 0.0015395982190966606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.008447864092886448 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535824440419674 + ], + "opposite_logprob_tokens": [ + -4.38753604888916 + ], + "logprob_reward": -0.012535824440419674, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535824440419674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08009102195501328, + 0.0896894708275795, + 0.0646292194724083, + 0.03413766250014305, + 0.11231875419616699, + 0.312732070684433, + 0.052272893488407135, + 0.03963632881641388, + 0.18608087301254272, + 0.028411615639925003 + ], + "credit_map": [ + 0.008529415354132652, + 0.005967957433313131, + 0.012350295670330524, + 0.009323382750153542, + 0.008275099098682404, + 0.011999513022601604, + 0.011845553293824196, + 0.01179980393499136, + 0.008259676396846771, + 0.00824551098048687, + 0.018813451752066612, + 0.018805410712957382, + 0.005106884054839611, + 0.007455075159668922, + 0.012720401398837566, + 0.01028305385261774, + 0.004564139526337385, + 0.006339057814329863, + 0.014507362619042397, + 0.013250733725726604, + 0.007088220212608576, + 0.0055983844213187695, + 0.007064132951200008, + 0.006217189133167267, + 0.005406194366514683, + 0.0035944629926234484, + 0.0032687161583453417, + 0.0024971042294055223, + 0.004483310505747795, + 0.005912573542445898, + 0.004572233185172081, + 0.0044030677527189255, + 0.013168933801352978, + 0.012168077751994133, + 0.009270097129046917, + 0.016003474593162537, + 0.014044209383428097, + 0.010634079575538635, + 0.0222071073949337, + 0.014822776429355145, + 0.024970227852463722, + 0.02929370291531086, + 0.04430581256747246, + 0.05482737720012665, + 0.029112571850419044, + 0.027191180735826492, + 0.044366899877786636, + 0.058664340525865555, + 0.005194677039980888, + 0.007768653798848391, + 0.005798120517283678, + 0.005401662550866604, + 0.008639521896839142, + 0.00697133457288146, + 0.008489582687616348, + 0.0040093413554131985, + 0.010831248015165329, + 0.0025481723714619875, + 0.010219825431704521, + 0.0035628946498036385, + 0.003218452213332057, + 0.00225820392370224, + 0.00274569820612669, + 0.004251832142472267, + 0.02014804072678089, + 0.033652011305093765, + 0.03015984408557415, + 0.018133267760276794, + 0.03234143182635307, + 0.013310710899531841, + 0.01886013150215149, + 0.01947544515132904, + 0.003317964496091008, + 0.004285257309675217, + 0.0022820611484348774, + 0.0029545535799115896, + 0.006133070681244135, + 0.0029834324959665537, + 0.003624449484050274, + 0.0028308271430432796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.012535824440419674 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008631067350506783 + ], + "opposite_logprob_tokens": [ + -4.758631229400635 + ], + "logprob_reward": -0.008631067350506783, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008631067350506783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15459100902080536, + 0.04904688894748688, + 0.1505526900291443, + 0.019835615530610085, + 0.04540308564901352, + 0.20398369431495667, + 0.15070809423923492, + 0.07228735089302063, + 0.11923476308584213, + 0.034356795251369476 + ], + "credit_map": [ + 0.01889929734170437, + 0.014314912259578705, + 0.02511083148419857, + 0.01901325397193432, + 0.01386425830423832, + 0.014620915986597538, + 0.027026211842894554, + 0.02174132503569126, + 0.004836622159928083, + 0.003903463017195463, + 0.009491645731031895, + 0.008503719232976437, + 0.0028724917210638523, + 0.00351309753023088, + 0.007191188167780638, + 0.008734658360481262, + 0.007572179194539785, + 0.012390858493745327, + 0.05022076517343521, + 0.039594635367393494, + 0.009963934309780598, + 0.01137247309088707, + 0.007957229390740395, + 0.011480613611638546, + 0.0022898174356669188, + 0.002333038253709674, + 0.002930873539298773, + 0.0016389485681429505, + 0.0022715458180755377, + 0.0023072741460055113, + 0.0032873055897653103, + 0.002776810899376869, + 0.005413484293967485, + 0.007288306951522827, + 0.00271207164041698, + 0.005123375449329615, + 0.005212857853621244, + 0.007081091403961182, + 0.006919223349541426, + 0.005652676336467266, + 0.01494335662573576, + 0.016803273931145668, + 0.028942156583070755, + 0.03509971499443054, + 0.019483253359794617, + 0.017498645931482315, + 0.0312739722430706, + 0.03993932157754898, + 0.016502955928444862, + 0.02065402828156948, + 0.016522232443094254, + 0.015070776455104351, + 0.027729960158467293, + 0.01914534717798233, + 0.024594498798251152, + 0.010488301515579224, + 0.020812949165701866, + 0.005856796633452177, + 0.020653201267123222, + 0.004960660822689533, + 0.006582679692655802, + 0.0045085507445037365, + 0.0034123272635042667, + 0.00550018111243844, + 0.014748060144484043, + 0.02204439975321293, + 0.016060857102274895, + 0.012411639094352722, + 0.022651836276054382, + 0.008776881732046604, + 0.011472084559500217, + 0.01106900442391634, + 0.003722773166373372, + 0.004582745488733053, + 0.004185795783996582, + 0.005018530413508415, + 0.004930414725095034, + 0.0038326671347022057, + 0.004981862846761942, + 0.0031020059250295162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.008631067350506783 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004632336553186178 + ], + "opposite_logprob_tokens": [ + -5.379632472991943 + ], + "logprob_reward": -0.004632336553186178, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004632336553186178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043979719281196594, + 0.025867057964205742, + 0.04143037647008896, + 0.005527321249246597, + 0.05456268787384033, + 0.043780792504549026, + 0.17710132896900177, + 0.022063245996832848, + 0.5537464618682861, + 0.03194108605384827 + ], + "credit_map": [ + 0.0050984937697649, + 0.004584369249641895, + 0.008283407427370548, + 0.005927987862378359, + 0.0034301055129617453, + 0.004159309435635805, + 0.006018560379743576, + 0.006477478425949812, + 0.002420484321191907, + 0.0020823595114052296, + 0.005325150676071644, + 0.004853906109929085, + 0.001722085988149047, + 0.002851291326805949, + 0.0035671463701874018, + 0.003044630866497755, + 0.0025474964641034603, + 0.004093270283192396, + 0.010565647855401039, + 0.007414159364998341, + 0.005020711105316877, + 0.003991666249930859, + 0.0034963737707585096, + 0.004301047418266535, + 0.0009496897691860795, + 0.0006999648758210242, + 0.0006064253393560648, + 0.00045842889812774956, + 0.0007606153376400471, + 0.0007351686945185065, + 0.0006988442037254572, + 0.0006181836943142116, + 0.006348761264234781, + 0.005107139237225056, + 0.004043140448629856, + 0.007333163637667894, + 0.004497794900089502, + 0.007332887500524521, + 0.01212279312312603, + 0.007777001708745956, + 0.004529309459030628, + 0.0030128771904855967, + 0.006800676696002483, + 0.012536324560642242, + 0.002834727056324482, + 0.0026965236756950617, + 0.00593103002756834, + 0.00543931731954217, + 0.017210692167282104, + 0.029213130474090576, + 0.019760137423872948, + 0.018379133194684982, + 0.03032601997256279, + 0.020644772797822952, + 0.028915047645568848, + 0.012652375735342503, + 0.004325259942561388, + 0.0023062711115926504, + 0.00433883722871542, + 0.0026097772642970085, + 0.0027231317944824696, + 0.0015006187604740262, + 0.0017295817378908396, + 0.002529766410589218, + 0.05465557426214218, + 0.0992003008723259, + 0.09786342084407806, + 0.05240101367235184, + 0.09320253878831863, + 0.04366123303771019, + 0.056101467460393906, + 0.05666084587574005, + 0.0036993655376136303, + 0.0029503912664949894, + 0.005137432832270861, + 0.0030029709450900555, + 0.006333861034363508, + 0.0028294818475842476, + 0.004710803739726543, + 0.0032767769880592823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.004632336553186178 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622679229825735 + ], + "opposite_logprob_tokens": [ + -4.882622718811035 + ], + "logprob_reward": -0.007622679229825735, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622679229825735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13565272092819214, + 0.07739780098199844, + 0.04418064281344414, + 0.031581711024045944, + 0.044695645570755005, + 0.12227297574281693, + 0.09916277974843979, + 0.028451746329665184, + 0.37315768003463745, + 0.04344635084271431 + ], + "credit_map": [ + 0.01838451623916626, + 0.014179767109453678, + 0.012712686322629452, + 0.016467055305838585, + 0.015566304326057434, + 0.016346991062164307, + 0.016974616795778275, + 0.02502075955271721, + 0.006447670049965382, + 0.005310262553393841, + 0.015593687072396278, + 0.013978683389723301, + 0.0039131916128098965, + 0.010833094827830791, + 0.01129230111837387, + 0.010028903372585773, + 0.0030217119492590427, + 0.004245314747095108, + 0.006998980883508921, + 0.005661664064973593, + 0.007543391548097134, + 0.007009581662714481, + 0.004640696570277214, + 0.005059299990534782, + 0.003634697524830699, + 0.0028025771025568247, + 0.0035247826017439365, + 0.0016153142787516117, + 0.004157313611358404, + 0.006277392152696848, + 0.005258480086922646, + 0.004311152268201113, + 0.004763100296258926, + 0.0049582188948988914, + 0.003196004079654813, + 0.005862079560756683, + 0.0039838463999331, + 0.005326633807271719, + 0.00856710784137249, + 0.008038651198148727, + 0.014496072195470333, + 0.01426622923463583, + 0.016221866011619568, + 0.022004730999469757, + 0.009603257291018963, + 0.008225399069488049, + 0.017150141298770905, + 0.020305262878537178, + 0.013931016437709332, + 0.011845256201922894, + 0.012055735103785992, + 0.011304114945232868, + 0.013785257004201412, + 0.010726354084908962, + 0.012439695186913013, + 0.013075342401862144, + 0.0031039968598634005, + 0.003976779989898205, + 0.00454770028591156, + 0.0028832440730184317, + 0.003962681163102388, + 0.0036823221016675234, + 0.0020722791086882353, + 0.004222739953547716, + 0.038875363767147064, + 0.06730443239212036, + 0.06460107117891312, + 0.03623633831739426, + 0.06222028285264969, + 0.028656363487243652, + 0.03749991953372955, + 0.03776385635137558, + 0.005214231088757515, + 0.005231557879596949, + 0.004938946105539799, + 0.003939488902688026, + 0.008467694744467735, + 0.003993592225015163, + 0.006822064984589815, + 0.004838770255446434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.007622679229825735 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot establishes a spacious yoga studio bathed in the golden glow of afternoon sunlight filtering through large arched windows, then the camera begins a slow and graceful pan to the right, capturing the scene. Warm light illuminates countless motes of dust dancing in the visible sunlight that spreads across the highly polished, honey-toned wooden floor, while tall potted ferns with lush, vibrant green foliage add a touch of nature to the serene space. At the front of the room, a tall, graceful giraffe instructor, dressed in a simple lavender tank top and yoga pants, gracefully demonstrates a pose, her long neck curved in a gentle arc, a serene smile on her face. A variety of anthropomorphic animals—including a slender fox in a warrior pose and a plump rabbit in a child's pose—are spread across colorful yoga mats, quietly holding their poses. As the camera continues, a furry little brown bear cub peeks hesitantly from the doorway, clutching a rolled-up blue yoga mat. His large, expressive eyes scan the room, and upon finding an empty spot, he cautiously and adorably clumsily tips in. Unfolding his mat, his expression determined but unsteady, he attempts to join the group, striking a precarious downward dog pose. The entire scene is rendered in the vibrant and captivating Pixar style, characterized by soft, saturated colors, detailed textures in the animals' fur, and an atmosphere that blends gentle humor with focused tranquility.", + "seed": 21189, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_6/index_326_seed_21189_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010136602981947362, + "vlm_logprob": -0.010136602981947362, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.010136602981947362, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006578100845217705 + ], + "opposite_logprob_tokens": [ + -6.256577968597412 + ], + "logprob_reward": -0.006578100845217705, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.006578100845217705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3005935549736023, + 0.07077716290950775, + 0.1264064460992813, + 0.039295587688684464, + 0.07236915826797485, + 0.03490908071398735, + 0.10799791663885117, + 0.1078595221042633, + 0.06313677877187729, + 0.076654814183712 + ], + "credit_map": [ + 0.05145947262644768, + 0.026394663378596306, + 0.04565722495317459, + 0.038769595324993134, + 0.03979918733239174, + 0.03404470905661583, + 0.03185808286070824, + 0.03261063992977142, + 0.005875932052731514, + 0.008308476768434048, + 0.007779902312904596, + 0.008334482088685036, + 0.0071046133525669575, + 0.010215513408184052, + 0.01326568890362978, + 0.009892554022371769, + 0.014985836111009121, + 0.013189088553190231, + 0.020055824890732765, + 0.017751621082425117, + 0.019605929031968117, + 0.01396366860717535, + 0.014214764349162579, + 0.012639705091714859, + 0.004702642560005188, + 0.004098216071724892, + 0.004359493963420391, + 0.0053117950446903706, + 0.0037511151749640703, + 0.004109119530767202, + 0.007452190387994051, + 0.0055110156536102295, + 0.006320988293737173, + 0.008450663648545742, + 0.006677937228232622, + 0.01137184351682663, + 0.013791990466415882, + 0.007991007529199123, + 0.010468832217156887, + 0.007295896764844656, + 0.0038427284453064203, + 0.00394363421946764, + 0.0054478151723742485, + 0.0049024708569049835, + 0.0037056419532746077, + 0.004481710493564606, + 0.004058771766722202, + 0.004526310600340366, + 0.013809511438012123, + 0.013426476158201694, + 0.011242109350860119, + 0.018113281577825546, + 0.013677528128027916, + 0.013016976416110992, + 0.015564095228910446, + 0.009147941134870052, + 0.015918653458356857, + 0.014210394583642483, + 0.016217142343521118, + 0.008961433544754982, + 0.01972026936709881, + 0.008658882230520248, + 0.008804875425994396, + 0.015367867425084114, + 0.006315100938081741, + 0.008557585999369621, + 0.0055977110750973225, + 0.006271182559430599, + 0.0125676104798913, + 0.012152113951742649, + 0.006519582588225603, + 0.0051558902487158775, + 0.007353255059570074, + 0.007859415374696255, + 0.010240412317216396, + 0.007791128940880299, + 0.010160526260733604, + 0.01348993368446827, + 0.010835764929652214, + 0.008924373425543308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.006578100845217705 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636622689664364 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636622689664364, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636622689664364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21006594598293304, + 0.03922388702630997, + 0.28878509998321533, + 0.04553470015525818, + 0.07750637829303741, + 0.04927525296807289, + 0.09429673850536346, + 0.1141090989112854, + 0.04870407283306122, + 0.03249878063797951 + ], + "credit_map": [ + 0.048670779913663864, + 0.01930435188114643, + 0.02672627381980419, + 0.02007872797548771, + 0.033931341022253036, + 0.020105326548218727, + 0.021219773218035698, + 0.020029393956065178, + 0.0036376784555613995, + 0.0046881441958248615, + 0.004260940942913294, + 0.006405295804142952, + 0.003804122796282172, + 0.004578837193548679, + 0.00582599313929677, + 0.00602287333458662, + 0.030338583514094353, + 0.03138742595911026, + 0.03811752051115036, + 0.07513943314552307, + 0.022500377148389816, + 0.03599105775356293, + 0.0319899320602417, + 0.023320786654949188, + 0.006708038970828056, + 0.004952683579176664, + 0.006888805888593197, + 0.006734316237270832, + 0.004358591511845589, + 0.004981855396181345, + 0.005662429146468639, + 0.005247981753200293, + 0.007403945550322533, + 0.012683740817010403, + 0.006546262186020613, + 0.008102510124444962, + 0.015044170431792736, + 0.009326295927166939, + 0.008810172788798809, + 0.009589281864464283, + 0.004663950763642788, + 0.006423264741897583, + 0.005634574685245752, + 0.004617478232830763, + 0.0071152979508042336, + 0.010860566049814224, + 0.005117037333548069, + 0.004843087866902351, + 0.01268857903778553, + 0.014493444934487343, + 0.010467763990163803, + 0.01630186289548874, + 0.011537435464560986, + 0.01131114736199379, + 0.011294309981167316, + 0.006202209275215864, + 0.013358589261770248, + 0.012978331185877323, + 0.013768564909696579, + 0.012065790593624115, + 0.02155883051455021, + 0.010676617734134197, + 0.018843915313482285, + 0.010858467780053616, + 0.0044279699213802814, + 0.005261326674371958, + 0.004613925237208605, + 0.006083790212869644, + 0.0076559074223041534, + 0.0057469317689538, + 0.010109503753483295, + 0.004804722033441067, + 0.002460370073094964, + 0.003287763334810734, + 0.003532672068104148, + 0.004991332069039345, + 0.003908336628228426, + 0.003767345566302538, + 0.0059416769072413445, + 0.004609288182109594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.008636622689664364 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026334119960665703 + ], + "opposite_logprob_tokens": [ + -3.651334047317505 + ], + "logprob_reward": -0.026334119960665703, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026334119960665703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2755826413631439, + 0.03278800845146179, + 0.14219558238983154, + 0.03218049183487892, + 0.10700836032629013, + 0.05500692129135132, + 0.04986940696835518, + 0.10647785663604736, + 0.0693308636546135, + 0.12955985963344574 + ], + "credit_map": [ + 0.05288602411746979, + 0.024874228984117508, + 0.043891943991184235, + 0.027015672996640205, + 0.037223391234874725, + 0.02447649836540222, + 0.03291924670338631, + 0.03229561448097229, + 0.004229535814374685, + 0.004353173542767763, + 0.0040468377992510796, + 0.0046454984694719315, + 0.0034246882423758507, + 0.003430838929489255, + 0.003977617248892784, + 0.004679820500314236, + 0.014595040120184422, + 0.013920658268034458, + 0.021960383281111717, + 0.03248078376054764, + 0.014672322198748589, + 0.01755213737487793, + 0.01374838501214981, + 0.013265866786241531, + 0.00381304114125669, + 0.003189877839758992, + 0.0038405517116189003, + 0.004530687816441059, + 0.003269046079367399, + 0.0035468493588268757, + 0.006152702029794455, + 0.003837736789137125, + 0.011974465101957321, + 0.013954661786556244, + 0.013493546284735203, + 0.013491753488779068, + 0.019793640822172165, + 0.011557698249816895, + 0.013739594258368015, + 0.009002997539937496, + 0.007003041449934244, + 0.007375798653811216, + 0.009412419982254505, + 0.005065863020718098, + 0.00707088690251112, + 0.006844755262136459, + 0.006443572696298361, + 0.005790582858026028, + 0.004467653576284647, + 0.009095603600144386, + 0.005440995097160339, + 0.006286717019975185, + 0.006886716466397047, + 0.005255395546555519, + 0.0070030870847404, + 0.005433238111436367, + 0.01227966882288456, + 0.015905756503343582, + 0.012776849791407585, + 0.010537875816226006, + 0.019160738214850426, + 0.012350395321846008, + 0.015005010180175304, + 0.00846156571060419, + 0.006881499197334051, + 0.009351382963359356, + 0.006144762970507145, + 0.007682533469051123, + 0.012845581397414207, + 0.012833643704652786, + 0.008256061933934689, + 0.005335400812327862, + 0.010631195269525051, + 0.01110425591468811, + 0.012531929649412632, + 0.021328791975975037, + 0.01916191726922989, + 0.012046282179653645, + 0.029579469934105873, + 0.013176023960113525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.026334119960665703 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015095992712303996 + ], + "opposite_logprob_tokens": [ + -6.501509666442871 + ], + "logprob_reward": -0.0015095992712303996, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015095992712303996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06755674630403519, + 0.06693581491708755, + 0.04845735803246498, + 0.04696497693657875, + 0.10807434469461441, + 0.08752632141113281, + 0.17049650847911835, + 0.16135752201080322, + 0.1131768599152565, + 0.12945355474948883 + ], + "credit_map": [ + 0.00928542111068964, + 0.007736832834780216, + 0.009694360196590424, + 0.006111424416303635, + 0.007957620546221733, + 0.007990505546331406, + 0.009445653297007084, + 0.009334921836853027, + 0.006199384108185768, + 0.0069265044294297695, + 0.003971709869801998, + 0.008007924072444439, + 0.0050320397131145, + 0.015597863122820854, + 0.014088313095271587, + 0.0071120792999863625, + 0.005836748983711004, + 0.005202645435929298, + 0.007002684753388166, + 0.00688980845734477, + 0.006401641294360161, + 0.005957625340670347, + 0.0054705021902918816, + 0.005695699714124203, + 0.007204681169241667, + 0.004884518217295408, + 0.0067254286259412766, + 0.0056083048693835735, + 0.006818625144660473, + 0.0053964657709002495, + 0.004334903322160244, + 0.005992050748318434, + 0.014751894399523735, + 0.012308337725698948, + 0.014742031693458557, + 0.014306015335023403, + 0.01766817830502987, + 0.010771281085908413, + 0.014550250954926014, + 0.008976359851658344, + 0.009678659029304981, + 0.01100090891122818, + 0.013455946929752827, + 0.009020308032631874, + 0.010690297931432724, + 0.009726404212415218, + 0.010220307856798172, + 0.013733487576246262, + 0.026054363697767258, + 0.023697860538959503, + 0.019761715084314346, + 0.03005523607134819, + 0.020117051899433136, + 0.016849424690008163, + 0.015346554107964039, + 0.018614303320646286, + 0.011869525536894798, + 0.015573528595268726, + 0.01838805340230465, + 0.02140824683010578, + 0.03126700967550278, + 0.01451532356441021, + 0.03062567487359047, + 0.017710156738758087, + 0.010515706613659859, + 0.011270240880548954, + 0.013765737414360046, + 0.012586393393576145, + 0.021386338397860527, + 0.01059796568006277, + 0.019843336194753647, + 0.013211140409111977, + 0.011761601082980633, + 0.010315785184502602, + 0.015623665414750576, + 0.016178777441382408, + 0.027446391060948372, + 0.018139535561203957, + 0.017023760825395584, + 0.012964046560227871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.0015095992712303996 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624572142958641 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624572142958641, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624572142958641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24018694460391998, + 0.07976236939430237, + 0.04690784215927124, + 0.045447882264852524, + 0.05395886301994324, + 0.04139767214655876, + 0.15487191081047058, + 0.13151724636554718, + 0.10144522041082382, + 0.10450409352779388 + ], + "credit_map": [ + 0.04669734835624695, + 0.020481280982494354, + 0.03353637829422951, + 0.027450693771243095, + 0.03013111837208271, + 0.021848976612091064, + 0.03369476646184921, + 0.026346370577812195, + 0.006608237512409687, + 0.005710999947041273, + 0.0049770730547606945, + 0.008242265321314335, + 0.006391645409166813, + 0.01992625929415226, + 0.02160578779876232, + 0.006300103385001421, + 0.0052063302136957645, + 0.006883248221129179, + 0.006724958308041096, + 0.006793193984776735, + 0.006947693880647421, + 0.004762211348861456, + 0.004869753960520029, + 0.004720455501228571, + 0.0054833414033055305, + 0.004661323968321085, + 0.004944559186697006, + 0.00560480821877718, + 0.006249354686588049, + 0.00412777392193675, + 0.009907281026244164, + 0.004469445440918207, + 0.007458553649485111, + 0.010536478832364082, + 0.006826152093708515, + 0.004927372559905052, + 0.008343592286109924, + 0.004772419575601816, + 0.005943056661635637, + 0.005151235964149237, + 0.00432269461452961, + 0.004487318452447653, + 0.004808970261365175, + 0.004859029781073332, + 0.006470272317528725, + 0.008714402094483376, + 0.003690310288220644, + 0.0040446738712489605, + 0.02473158948123455, + 0.021308546885848045, + 0.016532542183995247, + 0.02418118715286255, + 0.018524665385484695, + 0.01612078957259655, + 0.021083425730466843, + 0.012389165349304676, + 0.017608707770705223, + 0.021516242995858192, + 0.019261743873357773, + 0.012562304735183716, + 0.01603490114212036, + 0.009411169216036797, + 0.011485707014799118, + 0.023636458441615105, + 0.009178422391414642, + 0.010741247795522213, + 0.00964946486055851, + 0.011970226652920246, + 0.017147386446595192, + 0.011126836761832237, + 0.01917647011578083, + 0.012455168180167675, + 0.010893087834119797, + 0.008581562899053097, + 0.010082362219691277, + 0.013804659247398376, + 0.014659952372312546, + 0.020132815465331078, + 0.016722917556762695, + 0.009626739658415318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.007624572142958641 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot establishes a spacious yoga studio bathed in the golden glow of afternoon sunlight filtering through large arched windows, then the camera begins a slow and graceful pan to the right, capturing the scene. Warm light illuminates countless motes of dust dancing in the visible sunlight that spreads across the highly polished, honey-toned wooden floor, while tall potted ferns with lush, vibrant green foliage add a touch of nature to the serene space. At the front of the room, a tall, graceful giraffe instructor, dressed in a simple lavender tank top and yoga pants, gracefully demonstrates a pose, her long neck curved in a gentle arc, a serene smile on her face. A variety of anthropomorphic animals—including a slender fox in a warrior pose and a plump rabbit in a child's pose—are spread across colorful yoga mats, quietly holding their poses. As the camera continues, a furry little brown bear cub peeks hesitantly from the doorway, clutching a rolled-up blue yoga mat. His large, expressive eyes scan the room, and upon finding an empty spot, he cautiously and adorably clumsily tips in. Unfolding his mat, his expression determined but unsteady, he attempts to join the group, striking a precarious downward dog pose. The entire scene is rendered in the vibrant and captivating Pixar style, characterized by soft, saturated colors, detailed textures in the animals' fur, and an atmosphere that blends gentle humor with focused tranquility.", + "seed": 21190, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_6/index_326_seed_21190_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.010950675886124372, + "vlm_logprob": -0.010950675886124372, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.010950675886124372, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012255837209522724 + ], + "opposite_logprob_tokens": [ + -4.762255668640137 + ], + "logprob_reward": -0.012255837209522724, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.012255837209522724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25304803252220154, + 0.08746068924665451, + 0.1551293283700943, + 0.04663110896945, + 0.14761506021022797, + 0.06531432271003723, + 0.07570259273052216, + 0.07752576470375061, + 0.07266681641340256, + 0.018906285986304283 + ], + "credit_map": [ + 0.026169173419475555, + 0.020823664963245392, + 0.0475548580288887, + 0.0230241809040308, + 0.039099857211112976, + 0.05182189121842384, + 0.02300788089632988, + 0.021546537056565285, + 0.0125661576166749, + 0.013553145341575146, + 0.008012780919671059, + 0.010875964537262917, + 0.01617547869682312, + 0.008154702372848988, + 0.01035257987678051, + 0.007769881282001734, + 0.01548695657402277, + 0.018994027748703957, + 0.01684003882110119, + 0.028339749202132225, + 0.022636979818344116, + 0.014384680427610874, + 0.017846308648586273, + 0.020600590854883194, + 0.004696492571383715, + 0.0068994187749922276, + 0.005940714385360479, + 0.007201012689620256, + 0.0065408856607973576, + 0.005352362059056759, + 0.0057471441105008125, + 0.004253077786415815, + 0.018578344956040382, + 0.01533287763595581, + 0.02270987071096897, + 0.016406744718551636, + 0.018410874530673027, + 0.01611366868019104, + 0.019723104313015938, + 0.020339565351605415, + 0.011690713465213776, + 0.008451825007796288, + 0.007136601954698563, + 0.007444554939866066, + 0.009250812232494354, + 0.008140559308230877, + 0.006262348499149084, + 0.006936902180314064, + 0.012253519147634506, + 0.011846909299492836, + 0.008312786929309368, + 0.009173029102385044, + 0.010686042718589306, + 0.009640238247811794, + 0.007228008471429348, + 0.006562056485563517, + 0.010164063423871994, + 0.013397540897130966, + 0.008252869360148907, + 0.008985213935375214, + 0.00931481085717678, + 0.011126653291285038, + 0.00713858800008893, + 0.009146026335656643, + 0.011419854126870632, + 0.006055880803614855, + 0.004479643888771534, + 0.003823093604296446, + 0.00799847487360239, + 0.02577509544789791, + 0.0063045877031981945, + 0.006810187827795744, + 0.002562898676842451, + 0.0015508518554270267, + 0.001758668338879943, + 0.0034171370789408684, + 0.0020132672507315874, + 0.003285468090325594, + 0.002088101813569665, + 0.002229893347248435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.012255837209522724 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637921884655952 + ], + "opposite_logprob_tokens": [ + -4.75863790512085 + ], + "logprob_reward": -0.008637921884655952, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637921884655952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33811551332473755, + 0.10283049196004868, + 0.08697012066841125, + 0.1241157203912735, + 0.07217545062303543, + 0.1316198706626892, + 0.029402872547507286, + 0.05667487904429436, + 0.04264184087514877, + 0.015453213825821877 + ], + "credit_map": [ + 0.05367720499634743, + 0.04152983799576759, + 0.0458671972155571, + 0.047586068511009216, + 0.0331885851919651, + 0.06217992305755615, + 0.028305577114224434, + 0.025781111791729927, + 0.013732746243476868, + 0.00961253046989441, + 0.010637213476002216, + 0.012985671870410442, + 0.019561687484383583, + 0.014897771179676056, + 0.01190020702779293, + 0.009502653032541275, + 0.008814280852675438, + 0.009606540203094482, + 0.00939657911658287, + 0.0177838783711195, + 0.006703365594148636, + 0.008555945940315723, + 0.011155283078551292, + 0.014954241923987865, + 0.013625959865748882, + 0.012064923532307148, + 0.020042022690176964, + 0.026896843686699867, + 0.011520200408995152, + 0.012701986357569695, + 0.013812488876283169, + 0.013451279141008854, + 0.008693559095263481, + 0.007935653440654278, + 0.007090989500284195, + 0.008959597907960415, + 0.011387588456273079, + 0.007740700617432594, + 0.011231821030378342, + 0.009135534055531025, + 0.024469414725899696, + 0.022943949326872826, + 0.013715622946619987, + 0.012869377620518208, + 0.01943007856607437, + 0.017060479149222374, + 0.009697102941572666, + 0.01143383514136076, + 0.003553529502823949, + 0.0035744537599384785, + 0.0024284895043820143, + 0.002477596513926983, + 0.0064935022965073586, + 0.003993112128227949, + 0.0033020435366779566, + 0.003580139484256506, + 0.0065621440298855305, + 0.005502532701939344, + 0.011004596948623657, + 0.009394030086696148, + 0.00635929498821497, + 0.0075904605910182, + 0.003708712989464402, + 0.006553099490702152, + 0.007891255430877209, + 0.003168590599671006, + 0.0035157273523509502, + 0.002493740525096655, + 0.004182014614343643, + 0.014057636260986328, + 0.0034728064201772213, + 0.0038600624538958073, + 0.0015451739309355617, + 0.0014246219070628285, + 0.0017679312732070684, + 0.002338123507797718, + 0.00244511547498405, + 0.002534926636144519, + 0.0017068670131266117, + 0.0016904536169022322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.008637921884655952 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006729913409799337 + ], + "opposite_logprob_tokens": [ + -5.006730079650879 + ], + "logprob_reward": -0.006729913409799337, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006729913409799337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4941302537918091, + 0.11202883720397949, + 0.07904483377933502, + 0.04984496161341667, + 0.04066812992095947, + 0.04167143628001213, + 0.033893369138240814, + 0.057880934327840805, + 0.0683591440320015, + 0.022478127852082253 + ], + "credit_map": [ + 0.05505886301398277, + 0.06220662593841553, + 0.07484328001737595, + 0.08343242853879929, + 0.04501757025718689, + 0.09848276525735855, + 0.041122693568468094, + 0.033966004848480225, + 0.016165170818567276, + 0.022066975012421608, + 0.012739601545035839, + 0.014394163154065609, + 0.018414122983813286, + 0.009843355976045132, + 0.010527103208005428, + 0.007878336124122143, + 0.007581906393170357, + 0.010087646543979645, + 0.00878433883190155, + 0.01067850086838007, + 0.008671529591083527, + 0.007661592680960894, + 0.011650554835796356, + 0.013928763568401337, + 0.005950587801635265, + 0.007735799066722393, + 0.00776294432580471, + 0.008078843355178833, + 0.005724785383790731, + 0.0049082012847065926, + 0.004737896844744682, + 0.004945902619510889, + 0.005628403276205063, + 0.004165468271821737, + 0.005936960689723492, + 0.004857297986745834, + 0.004520116839557886, + 0.0037636347115039825, + 0.0055036889389157295, + 0.00629255548119545, + 0.006184617057442665, + 0.003761798143386841, + 0.005117271561175585, + 0.005627465434372425, + 0.0062894620932638645, + 0.005205227993428707, + 0.004526556935161352, + 0.004959034267812967, + 0.003950587473809719, + 0.003530702320858836, + 0.0030361360404640436, + 0.002570405136793852, + 0.008122351951897144, + 0.005361031740903854, + 0.003818526165559888, + 0.003503624117001891, + 0.006292681209743023, + 0.010227864608168602, + 0.007127242162823677, + 0.0075966413132846355, + 0.006299284286797047, + 0.00827076192945242, + 0.0061676171608269215, + 0.005898838397115469, + 0.011145412921905518, + 0.005279728677123785, + 0.004643773194402456, + 0.004057550337165594, + 0.006848597899079323, + 0.022595055401325226, + 0.006412746850401163, + 0.007376275025308132, + 0.003417812054976821, + 0.0019118508789688349, + 0.0032395494636148214, + 0.0024765250273048878, + 0.0030299355275928974, + 0.0034679328091442585, + 0.002169158775359392, + 0.0027653633151203394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.006729913409799337 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016060978174209595 + ], + "opposite_logprob_tokens": [ + -4.141060829162598 + ], + "logprob_reward": -0.016060978174209595, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016060978174209595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26355013251304626, + 0.06719046086072922, + 0.09186088293790817, + 0.07627898454666138, + 0.059632398188114166, + 0.09039919078350067, + 0.07567895203828812, + 0.037916507571935654, + 0.16735191643238068, + 0.07014060020446777 + ], + "credit_map": [ + 0.06198790296912193, + 0.02119707129895687, + 0.023231040686368942, + 0.01685248129069805, + 0.030952174216508865, + 0.07944171130657196, + 0.016567053273320198, + 0.013320697471499443, + 0.007498153019696474, + 0.011861446313560009, + 0.006207285448908806, + 0.011873307637870312, + 0.008617119863629341, + 0.00696027185767889, + 0.007425257004797459, + 0.006747621111571789, + 0.01135027501732111, + 0.010216468945145607, + 0.012710129842162132, + 0.011610658839344978, + 0.010405872017145157, + 0.007088606245815754, + 0.011249754577875137, + 0.017229123041033745, + 0.008912009187042713, + 0.008995755575597286, + 0.01162203773856163, + 0.015935931354761124, + 0.008465965278446674, + 0.0067763724364340305, + 0.008482662960886955, + 0.007088252808898687, + 0.009930960834026337, + 0.006647135596722364, + 0.008176939561963081, + 0.007435779552906752, + 0.005891846492886543, + 0.005995945073664188, + 0.0067083220928907394, + 0.008845467120409012, + 0.018405184149742126, + 0.010622759349644184, + 0.01145730260759592, + 0.008746827952563763, + 0.010050643235445023, + 0.009595869109034538, + 0.009650890715420246, + 0.011869718320667744, + 0.01024443656206131, + 0.009444856084883213, + 0.006533339153975248, + 0.006033648271113634, + 0.016661450266838074, + 0.01244006771594286, + 0.0070610628463327885, + 0.007260088808834553, + 0.0047086142003536224, + 0.0032897309865802526, + 0.007154220249503851, + 0.005900322459638119, + 0.003939392045140266, + 0.00461189029738307, + 0.004375164862722158, + 0.003937173169106245, + 0.033383581787347794, + 0.011203055270016193, + 0.011013549752533436, + 0.009022092446684837, + 0.0167890302836895, + 0.059711650013923645, + 0.011516266502439976, + 0.01471269503235817, + 0.014726080931723118, + 0.004564513918012381, + 0.008214552886784077, + 0.0070861754938960075, + 0.006844100076705217, + 0.017614630982279778, + 0.004243150353431702, + 0.006847389508038759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.016060978174209595 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a spacious yoga studio with large arched windows and a honey-toned wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow and graceful pan to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tall giraffe instructor wearing a lavender tank top and yoga pants demonstrating a pose at the front of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a brown bear cub peeking from the doorway and clumsily attempting a downward dog pose on a blue yoga mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a Pixar style with soft, saturated colors and detailed fur textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011068728752434254 + ], + "opposite_logprob_tokens": [ + -4.511068820953369 + ], + "logprob_reward": -0.011068728752434254, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011068728752434254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4405015707015991, + 0.0561068020761013, + 0.04488186165690422, + 0.0399775430560112, + 0.04163958132266998, + 0.11108610033988953, + 0.07460486143827438, + 0.06653975695371628, + 0.09934476017951965, + 0.02531718648970127 + ], + "credit_map": [ + 0.12206605076789856, + 0.027525652199983597, + 0.022183436900377274, + 0.024036459624767303, + 0.038871023803949356, + 0.15346166491508484, + 0.02061551623046398, + 0.031741730868816376, + 0.010007643140852451, + 0.006583203561604023, + 0.005782833322882652, + 0.006505647208541632, + 0.0076538678258657455, + 0.006566296797245741, + 0.007788129150867462, + 0.005219180602580309, + 0.00569024495780468, + 0.005645822733640671, + 0.005996438208967447, + 0.006012899335473776, + 0.005352813750505447, + 0.0035891723819077015, + 0.005798823200166225, + 0.0067956410348415375, + 0.006291261874139309, + 0.006135319359600544, + 0.005742180626839399, + 0.00453639542683959, + 0.005208118353039026, + 0.004633305128663778, + 0.003613448701798916, + 0.0038175093941390514, + 0.006732854060828686, + 0.004829511512070894, + 0.005668997298926115, + 0.004381336271762848, + 0.004796787165105343, + 0.004337462130934, + 0.005605150479823351, + 0.00528748007491231, + 0.021720537915825844, + 0.014130905270576477, + 0.013872455805540085, + 0.012127125635743141, + 0.012108642607927322, + 0.011318660341203213, + 0.011525211855769157, + 0.014282552525401115, + 0.009511061944067478, + 0.009084583260118961, + 0.006845911033451557, + 0.005404825322329998, + 0.018026866018772125, + 0.012582852505147457, + 0.006637820042669773, + 0.0065109338611364365, + 0.00817019958049059, + 0.010704613290727139, + 0.011384560726583004, + 0.012015367858111858, + 0.005982798524200916, + 0.007597292773425579, + 0.005517602898180485, + 0.005167309660464525, + 0.019847717136144638, + 0.0070851631462574005, + 0.00786419678479433, + 0.005764294881373644, + 0.010044167749583721, + 0.03203223645687103, + 0.006788602564483881, + 0.009918374009430408, + 0.004510074853897095, + 0.002148283878341317, + 0.00318903592415154, + 0.0032893524039536715, + 0.002797985216602683, + 0.005627754610031843, + 0.0015367926098406315, + 0.00221790443174541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.011068728752434254 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_7/index_tensor([3016, 3016])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_7/index_tensor([3016, 3016])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..999649481c6d6c72ba3a9c232ceb31f6f728a32e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_7/index_tensor([3016, 3016])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow-motion shot, shot from a low, upward angle, brims with a raw sense of power and a violent aesthetic. A fighter, incredibly muscular, his torso bare, sweat-soaked skin reflecting an oily sheen under the single, harsh spotlight overhead. He wears a pair of ragged black combat shorts, his hands bandaged in frayed white bandages. His face is one of intense focus and ferocity, his teeth clenched as he lifts his opponent, clad in red shorts, high above his head. The opponent's body arcs helplessly in the air, his face etched with terror and agony. Using one foot as the pivot, the fighter launches into a swift yet steady spin across the rough, cracked, dusty concrete. The centrifugal force sends his opponent's limbs flailing helplessly. When the rotation reached its peak, he suddenly exerted force and slammed his opponent to the ground. The huge impact force instantly made the concrete ground under the fighter's feet overwhelmed. Spider-web-like cracks quickly spread with the impact point as the center, followed by large pieces of concrete fragments and dust bursting upward. The flying dust was illuminated in the light beam, forming a visible particle trail. The entire ground collapsed violently under his feet, forming an expanding, bottomless pit. The whole scene, with the cold gray and dark black main colors, created an unstoppable and suffocating atmosphere of destruction.", + "seed": 11054, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_7/index_3016_seed_11054_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.026416387036442758, + "vlm_logprob": -0.026416387036442758, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.026416387036442758, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06890816241502762 + ], + "opposite_logprob_tokens": [ + -3.193908214569092 + ], + "logprob_reward": -0.06890816241502762, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.06890816241502762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12145765125751495, + 0.093268483877182, + 0.18071329593658447, + 0.1691235601902008, + 0.13034209609031677, + 0.03310755267739296, + 0.09053568542003632, + 0.045038528740406036, + 0.08948974311351776, + 0.046923473477363586 + ], + "credit_map": [ + 0.01291063241660595, + 0.011111362837255001, + 0.01844938099384308, + 0.014101916924118996, + 0.016040274873375893, + 0.0172852985560894, + 0.021220937371253967, + 0.010337849147617817, + 0.014700161293148994, + 0.00997982732951641, + 0.012215521186590195, + 0.012961463071405888, + 0.010304373688995838, + 0.011345668695867062, + 0.012526215054094791, + 0.009235256351530552, + 0.03424587473273277, + 0.014337562024593353, + 0.037604834884405136, + 0.01690729148685932, + 0.008598518557846546, + 0.03546600416302681, + 0.01786462590098381, + 0.015688573941588402, + 0.018296878784894943, + 0.02623429149389267, + 0.036085858941078186, + 0.017265895381569862, + 0.013987529091536999, + 0.01659262925386429, + 0.022997738793492317, + 0.01766274683177471, + 0.01436107512563467, + 0.017192265018820763, + 0.013295283541083336, + 0.015244215726852417, + 0.019942285493016243, + 0.008397688157856464, + 0.018603026866912842, + 0.02330624684691429, + 0.0024228275287896395, + 0.004118757322430611, + 0.004986889194697142, + 0.004492939915508032, + 0.00465509993955493, + 0.003914989996701479, + 0.004586944822221994, + 0.0039291041903197765, + 0.0074953436851501465, + 0.01235397718846798, + 0.008938565850257874, + 0.011262821964919567, + 0.013653356581926346, + 0.014428133144974709, + 0.00977250561118126, + 0.012630981393158436, + 0.005685231648385525, + 0.004437717143446207, + 0.003235473297536373, + 0.0035632995422929525, + 0.008581084199249744, + 0.005275044124573469, + 0.004905697889626026, + 0.00935498345643282, + 0.004415522795170546, + 0.003399791195988655, + 0.0035662290174514055, + 0.009160121902823448, + 0.0074434783309698105, + 0.030768152326345444, + 0.004523762967437506, + 0.02621268294751644, + 0.0037015217822045088, + 0.0028651042375713587, + 0.006094132550060749, + 0.003831121139228344, + 0.003906899597495794, + 0.00926929246634245, + 0.008119184523820877, + 0.009136218577623367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.06890816241502762 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007658288348466158 + ], + "opposite_logprob_tokens": [ + -4.8826584815979 + ], + "logprob_reward": -0.007658288348466158, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007658288348466158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2074662446975708, + 0.11290842294692993, + 0.14261843264102936, + 0.050710856914520264, + 0.08639796823263168, + 0.03981929272413254, + 0.04727378115057945, + 0.12630026042461395, + 0.136366069316864, + 0.050138700753450394 + ], + "credit_map": [ + 0.019205739721655846, + 0.020365137606859207, + 0.023380525410175323, + 0.02563258819282055, + 0.03248140588402748, + 0.03022925741970539, + 0.039747897535562515, + 0.01642369106411934, + 0.014445526525378227, + 0.01669829525053501, + 0.01608959585428238, + 0.017238738015294075, + 0.010633088648319244, + 0.012512332759797573, + 0.013438737951219082, + 0.011852103285491467, + 0.025508945807814598, + 0.012569809332489967, + 0.02023002691566944, + 0.015448127873241901, + 0.010029839351773262, + 0.02196602337062359, + 0.019085155799984932, + 0.017780495807528496, + 0.006745391525328159, + 0.007021167315542698, + 0.010183160193264484, + 0.004343357402831316, + 0.0044992282055318356, + 0.004457662347704172, + 0.007097989786416292, + 0.0063628992065787315, + 0.011426595039665699, + 0.010773122310638428, + 0.010154088959097862, + 0.009861106052994728, + 0.01107548363506794, + 0.006851817946881056, + 0.012847469188272953, + 0.01340828649699688, + 0.0029595927335321903, + 0.005025445483624935, + 0.0064201862551271915, + 0.005520382430404425, + 0.0048642815090715885, + 0.004270704463124275, + 0.005464477930217981, + 0.0052942195907235146, + 0.0038584803696721792, + 0.0064551145769655704, + 0.0045169140212237835, + 0.007202950306236744, + 0.005390281323343515, + 0.0061728013679385185, + 0.0068210419267416, + 0.0068561965599656105, + 0.017093421891331673, + 0.020978597924113274, + 0.014260626398026943, + 0.008310580626130104, + 0.023176265880465508, + 0.012630067765712738, + 0.013221614994108677, + 0.016629068180918694, + 0.0039761969819664955, + 0.0036777781788259745, + 0.0030662824865430593, + 0.015507507137954235, + 0.006614371202886105, + 0.051967523992061615, + 0.005403134040534496, + 0.04615326598286629, + 0.004061004612594843, + 0.006899713538587093, + 0.003781119594350457, + 0.004201368894428015, + 0.004421137738972902, + 0.010583559051156044, + 0.00597906531766057, + 0.010211726650595665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.007658288348466158 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419494766741991 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.01419494766741991, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419494766741991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11401841044425964, + 0.17157317698001862, + 0.0869843065738678, + 0.03539533168077469, + 0.0680117979645729, + 0.04076734557747841, + 0.06727775186300278, + 0.18649260699748993, + 0.130457803606987, + 0.09902147948741913 + ], + "credit_map": [ + 0.013600694946944714, + 0.013997096568346024, + 0.013830597512423992, + 0.015991078689694405, + 0.013955190777778625, + 0.01502539124339819, + 0.017651809379458427, + 0.009966561570763588, + 0.02174903266131878, + 0.019848328083753586, + 0.02487906813621521, + 0.027322661131620407, + 0.017252612859010696, + 0.017928682267665863, + 0.02160988189280033, + 0.020982906222343445, + 0.017085935920476913, + 0.007067400496453047, + 0.014478079974651337, + 0.009219381958246231, + 0.00547888595610857, + 0.01155299786478281, + 0.011281734332442284, + 0.010819890536367893, + 0.004441647324711084, + 0.004851118661463261, + 0.00822279043495655, + 0.0023149338085204363, + 0.0029702268075197935, + 0.003744506975635886, + 0.00423433305695653, + 0.004615774843841791, + 0.008735954761505127, + 0.008871973492205143, + 0.006597577594220638, + 0.007873153313994408, + 0.009729291312396526, + 0.005461385007947683, + 0.009789529256522655, + 0.010952935554087162, + 0.0030437372624874115, + 0.0049116904847323895, + 0.005424952134490013, + 0.006534653250128031, + 0.006950591690838337, + 0.004295700695365667, + 0.004683943465352058, + 0.004922077991068363, + 0.005941040813922882, + 0.008542824536561966, + 0.00900335144251585, + 0.009494883939623833, + 0.007152621168643236, + 0.00838544499129057, + 0.009366382844746113, + 0.009391196072101593, + 0.026307782158255577, + 0.03082170896232128, + 0.020074987784028053, + 0.011264361441135406, + 0.031593870371580124, + 0.01793748512864113, + 0.02029808610677719, + 0.02819431945681572, + 0.003986489027738571, + 0.004571986850351095, + 0.0030158173758536577, + 0.014179251156747341, + 0.0066325184889137745, + 0.04851849004626274, + 0.00472737243399024, + 0.04482588171958923, + 0.0061144451610744, + 0.005247943568974733, + 0.008637804538011551, + 0.008668585680425167, + 0.008718718774616718, + 0.017063317820429802, + 0.019436277449131012, + 0.025134393945336342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.01419494766741991 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626701612025499 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626701612025499, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626701612025499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09967786818742752, + 0.08970609307289124, + 0.11468297243118286, + 0.058669716119766235, + 0.11008533835411072, + 0.09197980910539627, + 0.07763586193323135, + 0.2049006223678589, + 0.11955874413251877, + 0.033102940768003464 + ], + "credit_map": [ + 0.010827058926224709, + 0.007573341950774193, + 0.011189818382263184, + 0.018346108496189117, + 0.015202498063445091, + 0.011574976146221161, + 0.016428913921117783, + 0.008535156026482582, + 0.013688988983631134, + 0.01646088995039463, + 0.012253155000507832, + 0.009254171513020992, + 0.007634826470166445, + 0.012453600764274597, + 0.008629095740616322, + 0.009331367909908295, + 0.02056204341351986, + 0.01443879771977663, + 0.0243174247443676, + 0.00682061305269599, + 0.006956267170608044, + 0.023255815729498863, + 0.007844282314181328, + 0.010487723164260387, + 0.007384740747511387, + 0.007387886755168438, + 0.011770850047469139, + 0.004480649251490831, + 0.00785856507718563, + 0.00590180791914463, + 0.006444195285439491, + 0.007441025692969561, + 0.011809847317636013, + 0.013254206627607346, + 0.012729262001812458, + 0.014868893660604954, + 0.017606636509299278, + 0.008528517559170723, + 0.016171319410204887, + 0.015116650611162186, + 0.007237017620354891, + 0.010040261782705784, + 0.012132743373513222, + 0.015251188538968563, + 0.01464760396629572, + 0.010900917463004589, + 0.010526391677558422, + 0.011243687942624092, + 0.0060303653590381145, + 0.009572030045092106, + 0.01073571015149355, + 0.010035374201834202, + 0.00961377564817667, + 0.009469609707593918, + 0.01050180196762085, + 0.011677194386720657, + 0.016937008127570152, + 0.025084923952817917, + 0.013806933537125587, + 0.017745744436979294, + 0.058346863836050034, + 0.026809683069586754, + 0.022189529612660408, + 0.023979930207133293, + 0.003950517624616623, + 0.004219833295792341, + 0.0036664525978267193, + 0.01217848714441061, + 0.008661546744406223, + 0.04203321784734726, + 0.0048079295083880424, + 0.0400407612323761, + 0.0023938431404531, + 0.00352094485424459, + 0.0036699026823043823, + 0.0029277505818754435, + 0.0031229115556925535, + 0.005563949700444937, + 0.0052355616353452206, + 0.006668076850473881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.007626701612025499 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336938351392746 + ], + "opposite_logprob_tokens": [ + -3.408693790435791 + ], + "logprob_reward": -0.0336938351392746, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336938351392746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04997725039720535, + 0.05988293141126633, + 0.0746266320347786, + 0.13193649053573608, + 0.12450354546308517, + 0.2337496280670166, + 0.15399949252605438, + 0.07435644418001175, + 0.06374479830265045, + 0.03322276473045349 + ], + "credit_map": [ + 0.005727824755012989, + 0.005666324403136969, + 0.00688248360529542, + 0.006798311602324247, + 0.005691494792699814, + 0.006420116871595383, + 0.009097537957131863, + 0.0036931538488715887, + 0.007428593002259731, + 0.006178740411996841, + 0.008221443742513657, + 0.010608118027448654, + 0.006707662716507912, + 0.006032440811395645, + 0.006103164050728083, + 0.008602762594819069, + 0.01112472265958786, + 0.007573344744741917, + 0.01649816520512104, + 0.007786874659359455, + 0.004905402194708586, + 0.011412408202886581, + 0.0071161407977342606, + 0.008209572173655033, + 0.017334595322608948, + 0.019548358395695686, + 0.03296647593379021, + 0.011189606040716171, + 0.011661392636597157, + 0.015558794140815735, + 0.011440083384513855, + 0.012237191200256348, + 0.02020534873008728, + 0.015863116830587387, + 0.014372024685144424, + 0.0157702025026083, + 0.013003687374293804, + 0.013151239603757858, + 0.01399514265358448, + 0.018142785876989365, + 0.01413717120885849, + 0.027408279478549957, + 0.03368452936410904, + 0.036224380135536194, + 0.0284104086458683, + 0.024677742272615433, + 0.03646005317568779, + 0.03274707868695259, + 0.011938507668673992, + 0.017858583480119705, + 0.01959165558218956, + 0.02178555354475975, + 0.023081323131918907, + 0.025362281128764153, + 0.01623672991991043, + 0.018144862726330757, + 0.010706355795264244, + 0.007843432016670704, + 0.005339444614946842, + 0.004601380322128534, + 0.017855241894721985, + 0.008541257120668888, + 0.007614034693688154, + 0.011855293065309525, + 0.003843998070806265, + 0.0033806986175477505, + 0.0027947856578975916, + 0.0061050280928611755, + 0.005524452310055494, + 0.020396186038851738, + 0.003155904822051525, + 0.01854374259710312, + 0.0037369669880717993, + 0.005149850156158209, + 0.0031362471636384726, + 0.0024912338703870773, + 0.0033650074619799852, + 0.006301858928054571, + 0.0037066913209855556, + 0.005334910005331039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.0336938351392746 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion shot, shot from a low, upward angle, brims with a raw sense of power and a violent aesthetic. A fighter, incredibly muscular, his torso bare, sweat-soaked skin reflecting an oily sheen under the single, harsh spotlight overhead. He wears a pair of ragged black combat shorts, his hands bandaged in frayed white bandages. His face is one of intense focus and ferocity, his teeth clenched as he lifts his opponent, clad in red shorts, high above his head. The opponent's body arcs helplessly in the air, his face etched with terror and agony. Using one foot as the pivot, the fighter launches into a swift yet steady spin across the rough, cracked, dusty concrete. The centrifugal force sends his opponent's limbs flailing helplessly. When the rotation reached its peak, he suddenly exerted force and slammed his opponent to the ground. The huge impact force instantly made the concrete ground under the fighter's feet overwhelmed. Spider-web-like cracks quickly spread with the impact point as the center, followed by large pieces of concrete fragments and dust bursting upward. The flying dust was illuminated in the light beam, forming a visible particle trail. The entire ground collapsed violently under his feet, forming an expanding, bottomless pit. The whole scene, with the cold gray and dark black main colors, created an unstoppable and suffocating atmosphere of destruction.", + "seed": 11055, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_7/index_3016_seed_11055_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.013431319687515497, + "vlm_logprob": -0.013431319687515497, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.013431319687515497, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03159714117646217 + ], + "opposite_logprob_tokens": [ + -4.531597137451172 + ], + "logprob_reward": -0.03159714117646217, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.03159714117646217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11167691648006439, + 0.10704852640628815, + 0.17353592813014984, + 0.06770075857639313, + 0.17294472455978394, + 0.03895603492856026, + 0.06562035530805588, + 0.0731520801782608, + 0.07236405462026596, + 0.11700063198804855 + ], + "credit_map": [ + 0.01521963719278574, + 0.007811363320797682, + 0.015919741243124008, + 0.021298548206686974, + 0.008818859234452248, + 0.011752636171877384, + 0.01905411295592785, + 0.011802013032138348, + 0.015963073819875717, + 0.007399725262075663, + 0.019949674606323242, + 0.005746159702539444, + 0.008883066475391388, + 0.02272222377359867, + 0.0108876321464777, + 0.015496970154345036, + 0.03988141193985939, + 0.012191837653517723, + 0.03946581482887268, + 0.016250427812337875, + 0.013026976026594639, + 0.012267961166799068, + 0.018822766840457916, + 0.021628733724355698, + 0.009370485320687294, + 0.012865078635513783, + 0.007554459385573864, + 0.005844655446708202, + 0.007491993252187967, + 0.008997496217489243, + 0.007313176523894072, + 0.008263417519629002, + 0.016561850905418396, + 0.02837686985731125, + 0.02090044692158699, + 0.010513370856642723, + 0.02232215367257595, + 0.04538177698850632, + 0.015227209776639938, + 0.013661050237715244, + 0.007109031546860933, + 0.004609609954059124, + 0.006087027955800295, + 0.0028354290407150984, + 0.00439848005771637, + 0.007148709148168564, + 0.0028562559746205807, + 0.003911490552127361, + 0.0073990048840641975, + 0.0049564428627491, + 0.011193829588592052, + 0.010022217407822609, + 0.006677531637251377, + 0.013367985375225544, + 0.005686875898391008, + 0.006316464859992266, + 0.006396610289812088, + 0.006011650897562504, + 0.015976062044501305, + 0.010574974119663239, + 0.005578064825385809, + 0.01194907259196043, + 0.009473851881921291, + 0.007191789802163839, + 0.009766632691025734, + 0.005354140419512987, + 0.006311443634331226, + 0.007966370321810246, + 0.007112011779099703, + 0.012137176468968391, + 0.013546058908104897, + 0.010170218534767628, + 0.011789271607995033, + 0.008498916402459145, + 0.03858114778995514, + 0.013472424820065498, + 0.011916451156139374, + 0.00813286378979683, + 0.011853696778416634, + 0.012755856849253178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 9 + ], + "selected_reward": -0.03159714117646217 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007650361862033606 + ], + "opposite_logprob_tokens": [ + -4.882650375366211 + ], + "logprob_reward": -0.007650361862033606, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007650361862033606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09398166835308075, + 0.3444276750087738, + 0.08207631856203079, + 0.04738324135541916, + 0.053666580468416214, + 0.035344742238521576, + 0.10202031582593918, + 0.050118185579776764, + 0.12595424056053162, + 0.06502711772918701 + ], + "credit_map": [ + 0.014566194266080856, + 0.007257648278027773, + 0.013566127978265285, + 0.014730933122336864, + 0.0076910750940442085, + 0.008873801678419113, + 0.013637380674481392, + 0.013658498413860798, + 0.05439502000808716, + 0.01838834211230278, + 0.07174749672412872, + 0.02533668652176857, + 0.027992764487862587, + 0.05873187258839607, + 0.048731882125139236, + 0.03910362347960472, + 0.017464715987443924, + 0.006188719533383846, + 0.017074506729841232, + 0.007757838815450668, + 0.006913567427545786, + 0.007856122218072414, + 0.010121962055563927, + 0.008698881603777409, + 0.006605086848139763, + 0.005421774461865425, + 0.0072900839149951935, + 0.003807665314525366, + 0.004812088795006275, + 0.006632853765040636, + 0.004909097217023373, + 0.007904591970145702, + 0.00455124955624342, + 0.010832630097866058, + 0.006011016201227903, + 0.0034121146891266108, + 0.006748022045940161, + 0.012288284488022327, + 0.006126981228590012, + 0.003696284256875515, + 0.005897045135498047, + 0.003477141261100769, + 0.005568459164351225, + 0.0029748694505542517, + 0.003752367105334997, + 0.004747090861201286, + 0.003858504118397832, + 0.005069264210760593, + 0.009133460931479931, + 0.006220283452421427, + 0.01627565175294876, + 0.01084520947188139, + 0.010521809570491314, + 0.028820350766181946, + 0.011079381220042706, + 0.009124169126152992, + 0.00400139857083559, + 0.00410939147695899, + 0.010435436852276325, + 0.006530712824314833, + 0.005257190205156803, + 0.006218270864337683, + 0.006488146726042032, + 0.007077633868902922, + 0.036782458424568176, + 0.009930716827511787, + 0.012454296462237835, + 0.009229375049471855, + 0.014538927003741264, + 0.018091991543769836, + 0.012224509380757809, + 0.012701965868473053, + 0.007444095332175493, + 0.0050778742879629135, + 0.017915494740009308, + 0.006848120596259832, + 0.007277282886207104, + 0.006249863654375076, + 0.006671846844255924, + 0.0075425393879413605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -0.007650361862033606 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253841444849968 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.01253841444849968, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253841444849968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10547615587711334, + 0.060114871710538864, + 0.10169569402933121, + 0.04242532700300217, + 0.1614406555891037, + 0.037138503044843674, + 0.19954003393650055, + 0.06486880034208298, + 0.17648327350616455, + 0.050816603004932404 + ], + "credit_map": [ + 0.011905724182724953, + 0.008883880451321602, + 0.015429473482072353, + 0.017052920535206795, + 0.011874591000378132, + 0.009459949098527431, + 0.01872309111058712, + 0.012146527878940105, + 0.007906116545200348, + 0.007480486761778593, + 0.00994323194026947, + 0.006074362900108099, + 0.005163043737411499, + 0.0073284669779241085, + 0.008925434201955795, + 0.0072937291115522385, + 0.023315176367759705, + 0.006734265945851803, + 0.024349018931388855, + 0.010695435106754303, + 0.007178506813943386, + 0.0074815088883042336, + 0.01045920792967081, + 0.01148257963359356, + 0.0063437195494771, + 0.009480317123234272, + 0.003733891062438488, + 0.0033714724704623222, + 0.004274341277778149, + 0.00657244585454464, + 0.0035908871795982122, + 0.005058251786977053, + 0.015048043802380562, + 0.02896478772163391, + 0.017190570011734962, + 0.007995352149009705, + 0.022950682789087296, + 0.04509211704134941, + 0.012846137396991253, + 0.011352967470884323, + 0.005246981978416443, + 0.005268320441246033, + 0.0061981710605323315, + 0.0035295526031404734, + 0.004754256922751665, + 0.004449719563126564, + 0.0036193132400512695, + 0.004072186537086964, + 0.015251094475388527, + 0.010025490075349808, + 0.0318266786634922, + 0.0243403110653162, + 0.01944774203002453, + 0.04906654357910156, + 0.0286662969738245, + 0.020915893837809563, + 0.005202845670282841, + 0.005110105965286493, + 0.013697185553610325, + 0.008453622460365295, + 0.005702291149646044, + 0.010000946931540966, + 0.008942999877035618, + 0.007758804596960545, + 0.03948875144124031, + 0.014355792663991451, + 0.01630093902349472, + 0.012555965222418308, + 0.017026469111442566, + 0.037368156015872955, + 0.020742684602737427, + 0.018644515424966812, + 0.005244359839707613, + 0.007686482276767492, + 0.009390079416334629, + 0.005019717384129763, + 0.009572189301252365, + 0.003614390967413783, + 0.0052167256362736225, + 0.005072655621916056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.01253841444849968 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863614957779646 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.00863614957779646, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863614957779646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07094433158636093, + 0.10302338004112244, + 0.11050115525722504, + 0.07297788560390472, + 0.06735256314277649, + 0.06159929931163788, + 0.10504905879497528, + 0.08886834233999252, + 0.17552721500396729, + 0.14415685832500458 + ], + "credit_map": [ + 0.010614205151796341, + 0.004979881923645735, + 0.010729227215051651, + 0.012877710163593292, + 0.00609799288213253, + 0.006833571009337902, + 0.010152488015592098, + 0.008659248240292072, + 0.015395338647067547, + 0.006769835948944092, + 0.02104034274816513, + 0.006617344915866852, + 0.00852307677268982, + 0.019490106031298637, + 0.014154297299683094, + 0.011033032089471817, + 0.021028129383921623, + 0.0061654034070670605, + 0.023375900462269783, + 0.011653481982648373, + 0.009362535551190376, + 0.007389667443931103, + 0.014197524636983871, + 0.017328493297100067, + 0.010135885328054428, + 0.015958840027451515, + 0.008415156044065952, + 0.0052681127563118935, + 0.008380240760743618, + 0.010455991141498089, + 0.006853668484836817, + 0.007509979885071516, + 0.006133323535323143, + 0.014321730472147465, + 0.0071376352570950985, + 0.004074959550052881, + 0.009855193085968494, + 0.012894608080387115, + 0.008090903982520103, + 0.004844203591346741, + 0.00936280470341444, + 0.007590467110276222, + 0.009630423039197922, + 0.004695270210504532, + 0.006868868134915829, + 0.009667660109698772, + 0.006347292568534613, + 0.007436507847160101, + 0.010071908123791218, + 0.009299376979470253, + 0.016652608290314674, + 0.010706990957260132, + 0.009938106872141361, + 0.024207329377532005, + 0.012602321803569794, + 0.011570402421057224, + 0.01039732713252306, + 0.007721382658928633, + 0.02047070302069187, + 0.01203307043761015, + 0.0066294195130467415, + 0.011996215209364891, + 0.010957876220345497, + 0.008662340231239796, + 0.034700024873018265, + 0.021111231297254562, + 0.016226470470428467, + 0.01402775663882494, + 0.01601555570960045, + 0.02696601115167141, + 0.023089678958058357, + 0.02339046448469162, + 0.015212258324027061, + 0.010755922645330429, + 0.04513370245695114, + 0.015151671133935452, + 0.015175746753811836, + 0.012113500386476517, + 0.014297056943178177, + 0.016316987574100494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.00863614957779646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734531372785568 + ], + "opposite_logprob_tokens": [ + -5.006734371185303 + ], + "logprob_reward": -0.006734531372785568, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734531372785568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17539501190185547, + 0.2568484842777252, + 0.08162223547697067, + 0.05828242376446724, + 0.07726824283599854, + 0.07850275188684464, + 0.020951613783836365, + 0.06171935796737671, + 0.04615982249379158, + 0.14325004816055298 + ], + "credit_map": [ + 0.02320881001651287, + 0.01187404990196228, + 0.0255593229085207, + 0.024752097204327583, + 0.019141685217618942, + 0.020401500165462494, + 0.03027491830289364, + 0.020182626321911812, + 0.04773251712322235, + 0.018627149984240532, + 0.061893682926893234, + 0.010920092463493347, + 0.017071716487407684, + 0.05620676651597023, + 0.013580475933849812, + 0.030816087499260902, + 0.016136353835463524, + 0.003870990825816989, + 0.018603209406137466, + 0.007707796525210142, + 0.008367357775568962, + 0.007391975261271, + 0.009697015397250652, + 0.00984753668308258, + 0.007392174564301968, + 0.008816405199468136, + 0.008071946911513805, + 0.0038424257654696703, + 0.006041177082806826, + 0.010527454316616058, + 0.005163102876394987, + 0.008427735418081284, + 0.007242134772241116, + 0.014422047883272171, + 0.010041043162345886, + 0.004243294708430767, + 0.013459090143442154, + 0.015356751158833504, + 0.007088475860655308, + 0.005415409803390503, + 0.015226785093545914, + 0.008950174786150455, + 0.010136793367564678, + 0.00480165658518672, + 0.008373302407562733, + 0.019805260002613068, + 0.005198328755795956, + 0.006010459735989571, + 0.001662924769334495, + 0.0023515885695815086, + 0.003525447566062212, + 0.002180139534175396, + 0.0029140335973352194, + 0.0032297028228640556, + 0.0026417383924126625, + 0.002446036320179701, + 0.007864486426115036, + 0.00960541795939207, + 0.008278853259980679, + 0.006169441621750593, + 0.006093787960708141, + 0.007188105024397373, + 0.007781028747558594, + 0.00873823557049036, + 0.005597343202680349, + 0.005733747035264969, + 0.004909921903163195, + 0.004158843774348497, + 0.005668171681463718, + 0.006882573943585157, + 0.006429814733564854, + 0.006779410410672426, + 0.01674162596464157, + 0.01506232563406229, + 0.03580432012677193, + 0.018901402130723, + 0.012671452015638351, + 0.01245170272886753, + 0.01568845473229885, + 0.01592877507209778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.006734531372785568 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion shot, shot from a low, upward angle, brims with a raw sense of power and a violent aesthetic. A fighter, incredibly muscular, his torso bare, sweat-soaked skin reflecting an oily sheen under the single, harsh spotlight overhead. He wears a pair of ragged black combat shorts, his hands bandaged in frayed white bandages. His face is one of intense focus and ferocity, his teeth clenched as he lifts his opponent, clad in red shorts, high above his head. The opponent's body arcs helplessly in the air, his face etched with terror and agony. Using one foot as the pivot, the fighter launches into a swift yet steady spin across the rough, cracked, dusty concrete. The centrifugal force sends his opponent's limbs flailing helplessly. When the rotation reached its peak, he suddenly exerted force and slammed his opponent to the ground. The huge impact force instantly made the concrete ground under the fighter's feet overwhelmed. Spider-web-like cracks quickly spread with the impact point as the center, followed by large pieces of concrete fragments and dust bursting upward. The flying dust was illuminated in the light beam, forming a visible particle trail. The entire ground collapsed violently under his feet, forming an expanding, bottomless pit. The whole scene, with the cold gray and dark black main colors, created an unstoppable and suffocating atmosphere of destruction.", + "seed": 11056, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_7/index_3016_seed_11056_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.012392700091004372, + "vlm_logprob": -0.012392700091004372, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.012392700091004372, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032037798315286636 + ], + "opposite_logprob_tokens": [ + -4.157037734985352 + ], + "logprob_reward": -0.032037798315286636, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.032037798315286636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2308279126882553, + 0.11540395766496658, + 0.05349090322852135, + 0.06429246813058853, + 0.11006617546081543, + 0.1597096174955368, + 0.05127963796257973, + 0.05833180993795395, + 0.07744698226451874, + 0.07915056496858597 + ], + "credit_map": [ + 0.030742961913347244, + 0.016103971749544144, + 0.04919581487774849, + 0.015326322056353092, + 0.014610089361667633, + 0.054060570895671844, + 0.013531296513974667, + 0.037256870418787, + 0.02090282365679741, + 0.009265996515750885, + 0.02078874781727791, + 0.00906901340931654, + 0.007345254998654127, + 0.02795240841805935, + 0.008859382010996342, + 0.011220332235097885, + 0.01034131832420826, + 0.006428574211895466, + 0.005636366084218025, + 0.00503160897642374, + 0.008832848630845547, + 0.007168086711317301, + 0.0052918787114322186, + 0.004760220646858215, + 0.00608567800372839, + 0.005717304069548845, + 0.006678834557533264, + 0.007487620692700148, + 0.011491712182760239, + 0.011670690029859543, + 0.006455396767705679, + 0.008705227635800838, + 0.00845201313495636, + 0.012539820745587349, + 0.00997029710561037, + 0.007202764041721821, + 0.016646364703774452, + 0.022883174940943718, + 0.017128925770521164, + 0.015242824330925941, + 0.011160423047840595, + 0.015632381662726402, + 0.02189989574253559, + 0.014673814177513123, + 0.026637298986315727, + 0.017637060955166817, + 0.031120674684643745, + 0.020948071032762527, + 0.0038772164843976498, + 0.005535754840821028, + 0.004561045207083225, + 0.0046840752474963665, + 0.008073361590504646, + 0.011870254762470722, + 0.00522094639018178, + 0.0074569834396243095, + 0.006177108734846115, + 0.006443114951252937, + 0.013155058026313782, + 0.008305598981678486, + 0.006860620807856321, + 0.00494436826556921, + 0.004762649070471525, + 0.007683290634304285, + 0.009934396483004093, + 0.011935998685657978, + 0.011543071828782558, + 0.010507742874324322, + 0.00740561168640852, + 0.009896820411086082, + 0.006421633530408144, + 0.009801704436540604, + 0.006664242595434189, + 0.018623346462845802, + 0.010695327073335648, + 0.007279952988028526, + 0.010315529070794582, + 0.010024436749517918, + 0.007749014534056187, + 0.0077987127006053925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.032037798315286636 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008663804270327091 + ], + "opposite_logprob_tokens": [ + -4.758663654327393 + ], + "logprob_reward": -0.008663804270327091, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008663804270327091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30594176054000854, + 0.2402265965938568, + 0.030862005427479744, + 0.050500981509685516, + 0.05813566967844963, + 0.07894489914178848, + 0.038531020283699036, + 0.08343840390443802, + 0.03912653401494026, + 0.07429216802120209 + ], + "credit_map": [ + 0.027742261067032814, + 0.022018438205122948, + 0.07144547998905182, + 0.024382920935750008, + 0.023479731753468513, + 0.07813559472560883, + 0.021008307114243507, + 0.037729017436504364, + 0.02863936871290207, + 0.014440770260989666, + 0.056829530745744705, + 0.01327416580170393, + 0.016010964289307594, + 0.07478458434343338, + 0.013415074907243252, + 0.022832155227661133, + 0.0043767946772277355, + 0.0037606169935315847, + 0.003417195985093713, + 0.0024997121654450893, + 0.006639169063419104, + 0.0038457890041172504, + 0.00379940471611917, + 0.002523321658372879, + 0.00451506394892931, + 0.0074641830287873745, + 0.008664291352033615, + 0.005642769392579794, + 0.007293327711522579, + 0.0065878937020897865, + 0.005493066273629665, + 0.004840387497097254, + 0.004996510688215494, + 0.005354450549930334, + 0.007069397252053022, + 0.008173543959856033, + 0.006877681706100702, + 0.008795585483312607, + 0.009465960785746574, + 0.007402543444186449, + 0.006746454164385796, + 0.011575854383409023, + 0.008017775602638721, + 0.00968258548527956, + 0.011849765665829182, + 0.010404177941381931, + 0.01180370431393385, + 0.008864585310220718, + 0.0034585969988256693, + 0.003567330539226532, + 0.00418931944295764, + 0.002577511826530099, + 0.006294189020991325, + 0.010148651897907257, + 0.002765911165624857, + 0.005529510788619518, + 0.007936403155326843, + 0.014363427646458149, + 0.012164939194917679, + 0.00715874508023262, + 0.00845498126000166, + 0.011768664233386517, + 0.010483012534677982, + 0.01110823079943657, + 0.005458453670144081, + 0.005963893141597509, + 0.005344299133867025, + 0.004428151994943619, + 0.0038714588154107332, + 0.006297851912677288, + 0.002783764386549592, + 0.004978659562766552, + 0.006353742443025112, + 0.02632945030927658, + 0.011376578360795975, + 0.006060415878891945, + 0.006906446069478989, + 0.006575142964720726, + 0.00412318529561162, + 0.006567205302417278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.008663804270327091 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533587403595448 + ], + "opposite_logprob_tokens": [ + -4.387533664703369 + ], + "logprob_reward": -0.012533587403595448, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533587403595448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16543953120708466, + 0.12418924272060394, + 0.03257337957620621, + 0.07363002002239227, + 0.13186855614185333, + 0.11843129992485046, + 0.04970363527536392, + 0.13147097826004028, + 0.06217832863330841, + 0.1105150654911995 + ], + "credit_map": [ + 0.023896854370832443, + 0.009610189124941826, + 0.04004732519388199, + 0.011316699907183647, + 0.007714623119682074, + 0.04291719198226929, + 0.01124500297009945, + 0.01869162917137146, + 0.012675981037318707, + 0.008791855536401272, + 0.028019655495882034, + 0.008397916331887245, + 0.010123380459845066, + 0.03415752202272415, + 0.012618465349078178, + 0.009404475800693035, + 0.005867708008736372, + 0.005280427169054747, + 0.003039507195353508, + 0.003710305318236351, + 0.003300826530903578, + 0.003784374799579382, + 0.0045476737432181835, + 0.003042558440938592, + 0.005663435906171799, + 0.007391194347292185, + 0.010087602771818638, + 0.00907152146100998, + 0.01226859726011753, + 0.011289327405393124, + 0.009219009429216385, + 0.008639330044388771, + 0.008318590000271797, + 0.014046706259250641, + 0.012913621962070465, + 0.012486238963901997, + 0.017380958423018456, + 0.020635752007365227, + 0.022758351638913155, + 0.02332833968102932, + 0.008186020888388157, + 0.016665296629071236, + 0.01099332980811596, + 0.01518833264708519, + 0.021230755373835564, + 0.013795225881040096, + 0.019197681918740273, + 0.013174649327993393, + 0.0038772900588810444, + 0.00474139116704464, + 0.0049885171465575695, + 0.003367834025993943, + 0.009881813079118729, + 0.012329598888754845, + 0.003643796546384692, + 0.006873393431305885, + 0.00867655873298645, + 0.0197631623595953, + 0.022566411644220352, + 0.01702720671892166, + 0.018059538677334785, + 0.013240806758403778, + 0.014731273986399174, + 0.017406025901436806, + 0.00783502496778965, + 0.008230263367295265, + 0.009536348283290863, + 0.00725799985229969, + 0.005246961954981089, + 0.010408764705061913, + 0.005376711487770081, + 0.008286255411803722, + 0.009334376081824303, + 0.03257868066430092, + 0.017202815040946007, + 0.010199612937867641, + 0.010073723271489143, + 0.011484441347420216, + 0.008368270471692085, + 0.01127314381301403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.012533587403595448 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092651419341564 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092651419341564, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092651419341564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1411619633436203, + 0.13275179266929626, + 0.03357239440083504, + 0.054303333163261414, + 0.07908523082733154, + 0.2468084990978241, + 0.04664808511734009, + 0.10219814628362656, + 0.0570131354033947, + 0.10645744949579239 + ], + "credit_map": [ + 0.022019650787115097, + 0.009068584069609642, + 0.027805214747786522, + 0.009706181474030018, + 0.00842322688549757, + 0.029060209169983864, + 0.011920491233468056, + 0.023158401250839233, + 0.016952209174633026, + 0.011498348787426949, + 0.023459970951080322, + 0.011462073773145676, + 0.010331030003726482, + 0.029275156557559967, + 0.01609817147254944, + 0.013674827292561531, + 0.004045937675982714, + 0.005630123429000378, + 0.004390537738800049, + 0.004334069788455963, + 0.0037732915952801704, + 0.0036329617723822594, + 0.0041300952434539795, + 0.00363537878729403, + 0.004203401040285826, + 0.006471441127359867, + 0.008556726388633251, + 0.006656264420598745, + 0.007631572894752026, + 0.006108378525823355, + 0.008122031576931477, + 0.006553517654538155, + 0.0058348579332232475, + 0.008651032112538815, + 0.007621423806995153, + 0.008879508823156357, + 0.012663720175623894, + 0.018382802605628967, + 0.009501531720161438, + 0.00755035737529397, + 0.015138092450797558, + 0.02629590779542923, + 0.030640892684459686, + 0.04064033925533295, + 0.032201509922742844, + 0.028334027156233788, + 0.04134410247206688, + 0.03221360221505165, + 0.004571368917822838, + 0.0032263589091598988, + 0.006496064830571413, + 0.0030568954534828663, + 0.006678002420812845, + 0.01324042771011591, + 0.0036049673799425364, + 0.0057739983312785625, + 0.00935908779501915, + 0.015305270440876484, + 0.01708926633000374, + 0.010125395841896534, + 0.009611624293029308, + 0.011931701563298702, + 0.014598711393773556, + 0.014177089557051659, + 0.007476886734366417, + 0.007364402059465647, + 0.00843583419919014, + 0.006129233632236719, + 0.005682322662323713, + 0.008862320333719254, + 0.004842530936002731, + 0.008219605311751366, + 0.009044929407536983, + 0.034160010516643524, + 0.016755972057580948, + 0.009280938655138016, + 0.008981337770819664, + 0.011591038666665554, + 0.0064870696514844894, + 0.010156149975955486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.004092651419341564 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635659046471119 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004635659046471119, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635659046471119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16957131028175354, + 0.06270790100097656, + 0.08878444135189056, + 0.0614621639251709, + 0.18417908251285553, + 0.21759051084518433, + 0.02042529359459877, + 0.054215554147958755, + 0.039359718561172485, + 0.10170400142669678 + ], + "credit_map": [ + 0.01910248212516308, + 0.008216776885092258, + 0.04332476854324341, + 0.008227543905377388, + 0.006823433097451925, + 0.04235263913869858, + 0.009492991492152214, + 0.032030679285526276, + 0.009582583792507648, + 0.008461774326860905, + 0.009854952804744244, + 0.00470703886821866, + 0.004880765452980995, + 0.008581655099987984, + 0.009808662347495556, + 0.006830465979874134, + 0.026659436523914337, + 0.00890277698636055, + 0.010061031207442284, + 0.0054187895730137825, + 0.011476089246571064, + 0.012891322374343872, + 0.006715868599712849, + 0.006659137085080147, + 0.005607385654002428, + 0.007315588183701038, + 0.008693725802004337, + 0.005743597634136677, + 0.00845405925065279, + 0.008985524997115135, + 0.0073684039525687695, + 0.009293882176280022, + 0.014927892945706844, + 0.03270525857806206, + 0.01448623463511467, + 0.015766451135277748, + 0.030835013836622238, + 0.026894761249423027, + 0.027648741379380226, + 0.02091471664607525, + 0.01645936816930771, + 0.03171782195568085, + 0.024510078132152557, + 0.023536667227745056, + 0.03591572865843773, + 0.024246281012892723, + 0.03619036450982094, + 0.02501419000327587, + 0.0018382577691227198, + 0.0024251092690974474, + 0.0018865481251850724, + 0.0027028657495975494, + 0.0027554230764508247, + 0.002973525784909725, + 0.003448925679549575, + 0.002394639188423753, + 0.005653647240251303, + 0.0068407440558075905, + 0.009518197737634182, + 0.00549030676484108, + 0.006267152260988951, + 0.006711774971336126, + 0.005892049055546522, + 0.007841681130230427, + 0.0033586854115128517, + 0.004807638935744762, + 0.006273164413869381, + 0.0040444438345730305, + 0.0035591786727309227, + 0.005119048058986664, + 0.0030699139460921288, + 0.009127646684646606, + 0.008763136342167854, + 0.030704878270626068, + 0.01596308872103691, + 0.009367186576128006, + 0.008347226306796074, + 0.010588767006993294, + 0.006947179324924946, + 0.011022532358765602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.004635659046471119 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion shot, shot from a low, upward angle, brims with a raw sense of power and a violent aesthetic. A fighter, incredibly muscular, his torso bare, sweat-soaked skin reflecting an oily sheen under the single, harsh spotlight overhead. He wears a pair of ragged black combat shorts, his hands bandaged in frayed white bandages. His face is one of intense focus and ferocity, his teeth clenched as he lifts his opponent, clad in red shorts, high above his head. The opponent's body arcs helplessly in the air, his face etched with terror and agony. Using one foot as the pivot, the fighter launches into a swift yet steady spin across the rough, cracked, dusty concrete. The centrifugal force sends his opponent's limbs flailing helplessly. When the rotation reached its peak, he suddenly exerted force and slammed his opponent to the ground. The huge impact force instantly made the concrete ground under the fighter's feet overwhelmed. Spider-web-like cracks quickly spread with the impact point as the center, followed by large pieces of concrete fragments and dust bursting upward. The flying dust was illuminated in the light beam, forming a visible particle trail. The entire ground collapsed violently under his feet, forming an expanding, bottomless pit. The whole scene, with the cold gray and dark black main colors, created an unstoppable and suffocating atmosphere of destruction.", + "seed": 11057, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_7/index_3016_seed_11057_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0231715053319931, + "vlm_logprob": -0.0231715053319931, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.0231715053319931, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05670655518770218 + ], + "opposite_logprob_tokens": [ + -3.431706666946411 + ], + "logprob_reward": -0.05670655518770218, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.05670655518770218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11409103125333786, + 0.17111890017986298, + 0.05496063455939293, + 0.059757381677627563, + 0.04442353919148445, + 0.3406223952770233, + 0.09450028836727142, + 0.054626017808914185, + 0.02567490004003048, + 0.04022478312253952 + ], + "credit_map": [ + 0.018757391721010208, + 0.010248757898807526, + 0.013383188284933567, + 0.01584458164870739, + 0.013043071143329144, + 0.015135238878428936, + 0.012440350838005543, + 0.015238465741276741, + 0.029227055609226227, + 0.013689402490854263, + 0.02013988047838211, + 0.01873452216386795, + 0.01783336140215397, + 0.022194012999534607, + 0.01942000724375248, + 0.029880687594413757, + 0.006683358922600746, + 0.00688996072858572, + 0.0047514610923826694, + 0.005859445780515671, + 0.006855615880340338, + 0.008525160141289234, + 0.008199549280107021, + 0.0071960873901844025, + 0.005927796009927988, + 0.009468824602663517, + 0.005562596023082733, + 0.007306887302547693, + 0.007042154669761658, + 0.009901461191475391, + 0.0059743281453847885, + 0.008573339320719242, + 0.004430932924151421, + 0.0038790120743215084, + 0.005087753292173147, + 0.005572166293859482, + 0.0036958942655473948, + 0.007030485197901726, + 0.005340181291103363, + 0.009387115947902203, + 0.010029819793999195, + 0.022508813068270683, + 0.01038223598152399, + 0.008750232867896557, + 0.1417297124862671, + 0.12757529318332672, + 0.008378659375011921, + 0.011267707683146, + 0.009200021624565125, + 0.010112352669239044, + 0.012329421006143093, + 0.012436099350452423, + 0.009511083364486694, + 0.011843687854707241, + 0.010828289203345776, + 0.0182393416762352, + 0.0060911583714187145, + 0.009434389881789684, + 0.006007292307913303, + 0.006220868322998285, + 0.005347945261746645, + 0.005192663986235857, + 0.010037681087851524, + 0.006294027902185917, + 0.003528292290866375, + 0.0029683110769838095, + 0.003036360489204526, + 0.0028776226099580526, + 0.0034036594443023205, + 0.0024337873328477144, + 0.004658591002225876, + 0.0027682804502546787, + 0.0038923572283238173, + 0.005341311451047659, + 0.0037698368541896343, + 0.005432242061942816, + 0.009177680127322674, + 0.0027687191031873226, + 0.00582121592015028, + 0.00402142433449626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.05670655518770218 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018221251666545868 + ], + "opposite_logprob_tokens": [ + -4.018221378326416 + ], + "logprob_reward": -0.018221251666545868, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018221251666545868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23629523813724518, + 0.06571254134178162, + 0.06131097301840782, + 0.10889185220003128, + 0.10472144186496735, + 0.11912958323955536, + 0.1051071360707283, + 0.056748948991298676, + 0.07423768937587738, + 0.06784460693597794 + ], + "credit_map": [ + 0.0408831425011158, + 0.022052908316254616, + 0.04738010838627815, + 0.03714403882622719, + 0.022745467722415924, + 0.01617344096302986, + 0.02806234173476696, + 0.02185376174747944, + 0.008153989911079407, + 0.005207268055528402, + 0.007316581904888153, + 0.011713621206581593, + 0.009717194363474846, + 0.006374537944793701, + 0.0073070223443210125, + 0.009922320954501629, + 0.00902758352458477, + 0.008671450428664684, + 0.008036233484745026, + 0.0063309017568826675, + 0.00561229232698679, + 0.00622372655197978, + 0.008351373486220837, + 0.009057406336069107, + 0.011895598843693733, + 0.010626069270074368, + 0.014674695208668709, + 0.0193049143999815, + 0.009604017250239849, + 0.019304586574435234, + 0.007948528975248337, + 0.015533426776528358, + 0.009624596685171127, + 0.015676647424697876, + 0.013839715160429478, + 0.014112122356891632, + 0.010981165803968906, + 0.013861515559256077, + 0.013900660909712315, + 0.01272500492632389, + 0.006059245206415653, + 0.028096772730350494, + 0.007855475880205631, + 0.00673258863389492, + 0.0147702069953084, + 0.040249790996313095, + 0.006297951098531485, + 0.009067535400390625, + 0.011376040056347847, + 0.010452421382069588, + 0.012754426337778568, + 0.01556103304028511, + 0.009892692789435387, + 0.012588429264724255, + 0.013770805671811104, + 0.0187112744897604, + 0.0064750174060463905, + 0.006354900076985359, + 0.005984937772154808, + 0.006973394192755222, + 0.00775539455935359, + 0.004220698960125446, + 0.009486309252679348, + 0.00949829164892435, + 0.012020387686789036, + 0.009694593958556652, + 0.008582225069403648, + 0.00955166108906269, + 0.010967416688799858, + 0.006069696508347988, + 0.008462006226181984, + 0.008889690041542053, + 0.006056691985577345, + 0.01079196110367775, + 0.006468131206929684, + 0.010356158949434757, + 0.01044708862900734, + 0.006080275867134333, + 0.010693724267184734, + 0.006950573995709419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.018221251666545868 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419647503644228 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.01419647503644228, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419647503644228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1159185916185379, + 0.08536262065172195, + 0.05224737524986267, + 0.08547817915678024, + 0.06547120958566666, + 0.38934746384620667, + 0.0523272342979908, + 0.045592572540044785, + 0.0632084384560585, + 0.045046381652355194 + ], + "credit_map": [ + 0.0196235328912735, + 0.007895705290138721, + 0.015912989154458046, + 0.019274413585662842, + 0.012862221337854862, + 0.009814269840717316, + 0.014280620962381363, + 0.016254831105470657, + 0.011061318218708038, + 0.006977919954806566, + 0.00839553214609623, + 0.021012743934988976, + 0.010031462647020817, + 0.0063837142661213875, + 0.009965715929865837, + 0.011534207500517368, + 0.006344007793813944, + 0.007085664197802544, + 0.005249605514109135, + 0.006801861338317394, + 0.00677573261782527, + 0.0061420476995408535, + 0.007735131774097681, + 0.006113323848694563, + 0.005703252740204334, + 0.008659372106194496, + 0.011465378105640411, + 0.02239762619137764, + 0.006293070036917925, + 0.012789984233677387, + 0.009121016599237919, + 0.0090484619140625, + 0.006292731501162052, + 0.005268569104373455, + 0.009888019412755966, + 0.007396947592496872, + 0.005913126282393932, + 0.010049628093838692, + 0.00874157715588808, + 0.011920608580112457, + 0.010645176284015179, + 0.053822923451662064, + 0.008856195956468582, + 0.01881137490272522, + 0.16293825209140778, + 0.10977630317211151, + 0.012594712898135185, + 0.01190251111984253, + 0.004722058307379484, + 0.005437071900814772, + 0.006406176369637251, + 0.008767731487751007, + 0.005678102374076843, + 0.0076394760981202126, + 0.006543561350554228, + 0.007133051287382841, + 0.005688243545591831, + 0.007831254042685032, + 0.005130535922944546, + 0.006100930273532867, + 0.004429864697158337, + 0.003209386020898819, + 0.008270833641290665, + 0.00493151880800724, + 0.007738148793578148, + 0.010588202625513077, + 0.008092249743640423, + 0.006936031859368086, + 0.007945122197270393, + 0.0049004810862243176, + 0.009164807386696339, + 0.00784339290112257, + 0.003973718732595444, + 0.0058783129788935184, + 0.004766353406012058, + 0.005999343004077673, + 0.009723705239593983, + 0.002947563072666526, + 0.006724016275256872, + 0.005033362191170454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.01419647503644228 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194360002875328 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194360002875328, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194360002875328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0737805962562561, + 0.04021771624684334, + 0.04222017899155617, + 0.017314013093709946, + 0.05099979788064957, + 0.439441055059433, + 0.1480298936367035, + 0.0714978575706482, + 0.07186231762170792, + 0.0446365587413311 + ], + "credit_map": [ + 0.012875095941126347, + 0.00554845342412591, + 0.010990842245519161, + 0.009650489315390587, + 0.007826690562069416, + 0.007999580353498459, + 0.007012594025582075, + 0.01187685877084732, + 0.006392540410161018, + 0.0031036315485835075, + 0.006114926654845476, + 0.007164500653743744, + 0.003141308669000864, + 0.003446447430178523, + 0.0035789827816188335, + 0.007275382522493601, + 0.0052490499801933765, + 0.006096545606851578, + 0.005126831587404013, + 0.004732843488454819, + 0.004886038135737181, + 0.004937805235385895, + 0.00646436819806695, + 0.004726701881736517, + 0.0017831031000241637, + 0.0024705096147954464, + 0.0016962462104856968, + 0.002813400700688362, + 0.0016237342497333884, + 0.0028292853385210037, + 0.0017872534226626158, + 0.002310482319444418, + 0.0049993060529232025, + 0.0057377382181584835, + 0.005994708742946386, + 0.00881450530141592, + 0.004072173498570919, + 0.006249070633202791, + 0.0077692368067801, + 0.007363065145909786, + 0.009296460077166557, + 0.07123669981956482, + 0.0077786860056221485, + 0.009034495800733566, + 0.19774466753005981, + 0.12621401250362396, + 0.008775182999670506, + 0.009360884316265583, + 0.011867191642522812, + 0.01990235224366188, + 0.018582798540592194, + 0.023537594825029373, + 0.015212973579764366, + 0.015835963189601898, + 0.015576845034956932, + 0.027514180168509483, + 0.009434269741177559, + 0.010275944136083126, + 0.008505426347255707, + 0.008039108477532864, + 0.007266748696565628, + 0.005860861856490374, + 0.013171940110623837, + 0.008943559601902962, + 0.010747453197836876, + 0.012555141001939774, + 0.009321462363004684, + 0.0060770404525101185, + 0.009922495111823082, + 0.007068741600960493, + 0.008876299485564232, + 0.00729369604960084, + 0.004750579595565796, + 0.005910544656217098, + 0.005021474324166775, + 0.006346974056214094, + 0.005813714116811752, + 0.004687412641942501, + 0.006461362354457378, + 0.005644499324262142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.014194360002875328 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main fighter muscular, bare-chested, wearing ragged black combat shorts, and has bandaged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fighter perform a swift spin on rough, cracked, dusty concrete while lifting an opponent in red shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact cause spider-web-like cracks to spread from the center, with concrete fragments and dust bursting upward into a visible particle trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by cold gray and dark black colors with a single harsh spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253888476639986 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.01253888476639986, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253888476639986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17900899052619934, + 0.16512541472911835, + 0.0968586727976799, + 0.09261039644479752, + 0.04801282659173012, + 0.20962312817573547, + 0.08228032290935516, + 0.03583642840385437, + 0.03975146636366844, + 0.05089230090379715 + ], + "credit_map": [ + 0.028445128351449966, + 0.014305329881608486, + 0.024678125977516174, + 0.023983249440789223, + 0.01928943209350109, + 0.025574084371328354, + 0.020259952172636986, + 0.02247370406985283, + 0.02788619138300419, + 0.013140449300408363, + 0.018362145870923996, + 0.015493705868721008, + 0.019917810335755348, + 0.024286771193146706, + 0.014345662668347359, + 0.03169269114732742, + 0.012224451638758183, + 0.01248281728476286, + 0.008756005205214024, + 0.010588649660348892, + 0.012268872000277042, + 0.015705300495028496, + 0.01224539615213871, + 0.012587180361151695, + 0.008975466713309288, + 0.007462562993168831, + 0.008672691881656647, + 0.014409283176064491, + 0.006384194828569889, + 0.019163457676768303, + 0.011308577843010426, + 0.016234170645475388, + 0.005560161545872688, + 0.0048892986960709095, + 0.005409001372754574, + 0.00543625745922327, + 0.0036960402503609657, + 0.006613133009523153, + 0.006968855392187834, + 0.009440083056688309, + 0.00688787829130888, + 0.046922579407691956, + 0.005461276043206453, + 0.009992743842303753, + 0.08428547531366348, + 0.037673406302928925, + 0.009316575713455677, + 0.00908320490270853, + 0.006635414902120829, + 0.011178537271916866, + 0.010247254744172096, + 0.011967272497713566, + 0.007975822314620018, + 0.009412478655576706, + 0.009345157071948051, + 0.015518390573561192, + 0.006241423077881336, + 0.004249257501214743, + 0.004595195408910513, + 0.0041032941080629826, + 0.004135872237384319, + 0.0029755544383078814, + 0.004608386196196079, + 0.004927450325340033, + 0.005443372298032045, + 0.0071103558875620365, + 0.0046922252513468266, + 0.004741381388157606, + 0.00570383807644248, + 0.0030183657072484493, + 0.004654808901250362, + 0.0043871235102415085, + 0.005072108935564756, + 0.00700393645092845, + 0.006832568906247616, + 0.006309931632131338, + 0.006882590241730213, + 0.0037902919575572014, + 0.007081009913235903, + 0.00791986659169197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.01253888476639986 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_8/index_tensor([2991, 2991])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_8/index_tensor([2991, 2991])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..964892e5206fc3c00442483c6876b19ed0882250 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_8/index_tensor([2991, 2991])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium-length, static shot places us in a dimly lit traditional kendo arena, where two kendo practitioners, clad in full traditional protective gear (bogu), are locked in a tense standoff. A harsh overhead light cuts fiercely into the frame from above and sideways, casting deep, sharply defined shadows on their bodies and the polished wooden floor. The entire scene is rendered in Daido Moriyama's signature grainy, high-contrast black-and-white style, with virtually no gray between the stark black and white. The light creates sharp, bright edges on the lacquered metal visors of their breastplates (do) and helmets (men), contrasting starkly with the sheer darkness of their dark kendo uniforms (keikogi) and hakama culottes (skirt pants). Suddenly, the swordsman in the foreground plunges forward, then lunges forward with explosive force, his right foot stomping decisively into the floor, stirring up a fine dust cloud. He raises his bamboo sword (shinai) overhead in a sharp arc, its immense speed leaving a visible, blurry white afterimage as it hurtles towards his opponent, who, poised to block the devastating blow, holds his balance. The entire scene is imbued with extreme dynamic tension and the aesthetics of a frozen moment, conveying a raw, intense, and intensely focused atmosphere of the duel.", + "seed": 25467, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_8/index_2991_seed_25467_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.040976897813379765, + "vlm_logprob": -0.040976897813379765, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.040976897813379765, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06092313304543495 + ], + "opposite_logprob_tokens": [ + -2.9359230995178223 + ], + "logprob_reward": -0.06092313304543495, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06092313304543495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18399441242218018, + 0.12716516852378845, + 0.06926180422306061, + 0.05416379123926163, + 0.020434198901057243, + 0.06921788305044174, + 0.09658444672822952, + 0.07456588745117188, + 0.1511096954345703, + 0.1535027176141739 + ], + "credit_map": [ + 0.031834378838539124, + 0.022199362516403198, + 0.02013378031551838, + 0.026923932135105133, + 0.022038858383893967, + 0.01071811094880104, + 0.027279214933514595, + 0.022866755723953247, + 0.020779764279723167, + 0.014864259399473667, + 0.010046526789665222, + 0.021709328517317772, + 0.015177062712609768, + 0.010221101343631744, + 0.01804284378886223, + 0.016324274241924286, + 0.009999813511967659, + 0.010709311813116074, + 0.004331266041845083, + 0.009103893302381039, + 0.008728204295039177, + 0.00824033748358488, + 0.0073854667134583, + 0.010763508267700672, + 0.007930454798042774, + 0.01246356125921011, + 0.0035686346236616373, + 0.007159910164773464, + 0.0050300112925469875, + 0.00492500513792038, + 0.00601561414077878, + 0.007070593070238829, + 0.0027008408214896917, + 0.003107944503426552, + 0.002002351451665163, + 0.002605391200631857, + 0.002618570113554597, + 0.0018496228149160743, + 0.0027002282440662384, + 0.0028492470737546682, + 0.011637275107204914, + 0.007248874753713608, + 0.004315752536058426, + 0.006912310607731342, + 0.0057752314023673534, + 0.005114327650517225, + 0.00909202080219984, + 0.019122084602713585, + 0.015316533856093884, + 0.015107271261513233, + 0.005958341062068939, + 0.0093615697696805, + 0.008694429881870747, + 0.01137508824467659, + 0.015503893606364727, + 0.01526730414479971, + 0.013419159688055515, + 0.008635930716991425, + 0.004998486023396254, + 0.007364839315414429, + 0.005939064547419548, + 0.012338058091700077, + 0.007822978310286999, + 0.014047360979020596, + 0.021192070096731186, + 0.024759801104664803, + 0.010381476953625679, + 0.0291645098477602, + 0.011586369946599007, + 0.007290471810847521, + 0.02145148627460003, + 0.025283508002758026, + 0.022305183112621307, + 0.023856692016124725, + 0.01572057232260704, + 0.030829958617687225, + 0.01833951473236084, + 0.011674943380057812, + 0.01687074825167656, + 0.013905084691941738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.06092313304543495 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020633455365896225 + ], + "opposite_logprob_tokens": [ + -3.8956334590911865 + ], + "logprob_reward": -0.020633455365896225, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020633455365896225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18460534512996674, + 0.17295250296592712, + 0.0729721188545227, + 0.07981926947832108, + 0.03475167229771614, + 0.04174480587244034, + 0.19931504130363464, + 0.09743159264326096, + 0.06869921088218689, + 0.04770850017666817 + ], + "credit_map": [ + 0.026443595066666603, + 0.029230967164039612, + 0.017286133021116257, + 0.0247334074229002, + 0.025295313447713852, + 0.013210919685661793, + 0.026577789336442947, + 0.021827228367328644, + 0.026441974565386772, + 0.03041083738207817, + 0.01274188607931137, + 0.02024206891655922, + 0.018140796571969986, + 0.014753533527255058, + 0.029652750119566917, + 0.020568659529089928, + 0.011150055564939976, + 0.013837413862347603, + 0.00588473305106163, + 0.010067008435726166, + 0.008671125397086143, + 0.0068522305227816105, + 0.008369805291295052, + 0.008139746263623238, + 0.011537599377334118, + 0.016954118385910988, + 0.003840382443740964, + 0.010129962116479874, + 0.007603755220770836, + 0.006368946749716997, + 0.007496863603591919, + 0.015887636691331863, + 0.005703547038137913, + 0.005235984921455383, + 0.0019271712517365813, + 0.0039054714143276215, + 0.00394909642636776, + 0.002683706348761916, + 0.004396637436002493, + 0.006950057111680508, + 0.006247838493436575, + 0.005238644313067198, + 0.0029786599334329367, + 0.003908670041710138, + 0.004112448543310165, + 0.0031955097801983356, + 0.010137229226529598, + 0.005925805773586035, + 0.031653258949518204, + 0.01422697864472866, + 0.008921634405851364, + 0.016239289194345474, + 0.012347903102636337, + 0.029276471585035324, + 0.030814489349722862, + 0.055834993720054626, + 0.016841497272253036, + 0.012580379843711853, + 0.00473498459905386, + 0.0063486783765256405, + 0.010020894929766655, + 0.016306547448039055, + 0.012377719394862652, + 0.01822088658809662, + 0.010199212469160557, + 0.011591845192015171, + 0.0038380881305783987, + 0.007563105784356594, + 0.00811169296503067, + 0.004167396575212479, + 0.012222090736031532, + 0.011005774140357971, + 0.006734101101756096, + 0.0060847303830087185, + 0.004077110439538956, + 0.007729582954198122, + 0.005935992579907179, + 0.007005529012531042, + 0.004429732449352741, + 0.005711722187697887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.020633455365896225 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004648474045097828 + ], + "opposite_logprob_tokens": [ + -5.379648685455322 + ], + "logprob_reward": -0.004648474045097828, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004648474045097828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40957680344581604, + 0.07674214243888855, + 0.0901249349117279, + 0.044052548706531525, + 0.06119180843234062, + 0.05364029482007027, + 0.056684572249650955, + 0.05368068814277649, + 0.08808349072933197, + 0.06622274219989777 + ], + "credit_map": [ + 0.081576406955719, + 0.041380029171705246, + 0.037873316556215286, + 0.0389186255633831, + 0.05118048936128616, + 0.030503232032060623, + 0.06081049144268036, + 0.06733421236276627, + 0.012849071994423866, + 0.013115180656313896, + 0.007529975846409798, + 0.010718880221247673, + 0.008508112281560898, + 0.0057659330777823925, + 0.011572330258786678, + 0.006682658568024635, + 0.012872054241597652, + 0.010503149591386318, + 0.005255575757473707, + 0.014172819443047047, + 0.012109928764402866, + 0.005483547225594521, + 0.018671350553631783, + 0.011056501418352127, + 0.006309653632342815, + 0.007425037212669849, + 0.0026876134797930717, + 0.005927287507802248, + 0.0053317006677389145, + 0.0028652590699493885, + 0.009543294087052345, + 0.003962698858231306, + 0.009402783587574959, + 0.010846319608390331, + 0.005187875125557184, + 0.00922919437289238, + 0.006023964378982782, + 0.0038253620732575655, + 0.007851705886423588, + 0.008824600838124752, + 0.00717698922380805, + 0.008604468777775764, + 0.0033034475054591894, + 0.0046684350818395615, + 0.005412462633103132, + 0.00480950903147459, + 0.008828862570226192, + 0.010836118832230568, + 0.010052885860204697, + 0.007962431758642197, + 0.004069025628268719, + 0.004923479165881872, + 0.0036660428158938885, + 0.005038991570472717, + 0.009119786322116852, + 0.011851930059492588, + 0.009404884651303291, + 0.006631303578615189, + 0.003613373963162303, + 0.004534444771707058, + 0.004450658801943064, + 0.008234689943492413, + 0.006908653769642115, + 0.009902680292725563, + 0.013379903510212898, + 0.01155027374625206, + 0.004173711873590946, + 0.00959815550595522, + 0.009397291578352451, + 0.004859709180891514, + 0.013896554708480835, + 0.021227886900305748, + 0.008110524155199528, + 0.007785520516335964, + 0.004681548103690147, + 0.013581718318164349, + 0.008241506293416023, + 0.007961825467646122, + 0.006941468454897404, + 0.00891862902790308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.004648474045097828 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486467182636261 + ], + "opposite_logprob_tokens": [ + -3.0486466884613037 + ], + "logprob_reward": -0.0486467182636261, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486467182636261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13407067954540253, + 0.07542046159505844, + 0.09952732920646667, + 0.08247458934783936, + 0.06066827476024628, + 0.08081861585378647, + 0.228220134973526, + 0.04931499436497688, + 0.13054707646369934, + 0.05893780663609505 + ], + "credit_map": [ + 0.02170104905962944, + 0.021535467356443405, + 0.015197689644992352, + 0.034193530678749084, + 0.011867697350680828, + 0.006668661721050739, + 0.01239858753979206, + 0.010508004575967789, + 0.011443049646914005, + 0.012330139987170696, + 0.005274502094835043, + 0.00882758479565382, + 0.009538805112242699, + 0.006077696569263935, + 0.0131114162504673, + 0.008817268535494804, + 0.013514085672795773, + 0.009141125716269016, + 0.006018532440066338, + 0.011716003529727459, + 0.01445846725255251, + 0.008002026006579399, + 0.02201836369931698, + 0.014658727683126926, + 0.011773361824452877, + 0.016653865575790405, + 0.0036000702530145645, + 0.010678938589990139, + 0.006243739742785692, + 0.0055209132842719555, + 0.014152168296277523, + 0.01385154016315937, + 0.008623939007520676, + 0.00823048036545515, + 0.0041755023412406445, + 0.006321482360363007, + 0.00892453920096159, + 0.005805301479995251, + 0.009755409322679043, + 0.008831620216369629, + 0.01226876676082611, + 0.009719614870846272, + 0.004257457796484232, + 0.006804297678172588, + 0.008907640352845192, + 0.006480667274445295, + 0.012168608605861664, + 0.020211568102240562, + 0.038699228316545486, + 0.017365146428346634, + 0.009929648600518703, + 0.014503013342618942, + 0.012817499227821827, + 0.026969054713845253, + 0.039043281227350235, + 0.06889329105615616, + 0.0072677298448979855, + 0.007013401482254267, + 0.0024413366336375475, + 0.004227567929774523, + 0.004829810466617346, + 0.006990876980125904, + 0.008031808771193027, + 0.00851246528327465, + 0.018995271995663643, + 0.019590556621551514, + 0.007878205738961697, + 0.014382002875208855, + 0.016313496977090836, + 0.007856673561036587, + 0.022328399121761322, + 0.023202477023005486, + 0.008561944589018822, + 0.009024779312312603, + 0.00621465127915144, + 0.008813465014100075, + 0.008187036029994488, + 0.0063658528961241245, + 0.005619299132376909, + 0.006150783505290747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.0486467182636261 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003270834684372 + ], + "opposite_logprob_tokens": [ + -2.6950325965881348 + ], + "logprob_reward": -0.07003270834684372, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003270834684372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3727964758872986, + 0.057610902935266495, + 0.07049478590488434, + 0.10270281136035919, + 0.03777424991130829, + 0.04740521311759949, + 0.05979295074939728, + 0.036021556705236435, + 0.11924765259027481, + 0.0961533933877945 + ], + "credit_map": [ + 0.06620743870735168, + 0.04012904316186905, + 0.04010625556111336, + 0.037846289575099945, + 0.04664609208703041, + 0.031647853553295135, + 0.049905575811862946, + 0.06030794605612755, + 0.009516901336610317, + 0.007500451058149338, + 0.005464581307023764, + 0.009255655109882355, + 0.00594544131308794, + 0.003728718962520361, + 0.008113794028759003, + 0.008085362613201141, + 0.011950880289077759, + 0.011329823173582554, + 0.004166059661656618, + 0.0076067703776061535, + 0.008314735256135464, + 0.009157711640000343, + 0.0071384496986866, + 0.010830355808138847, + 0.014709101058542728, + 0.021486740559339523, + 0.004986208397895098, + 0.014906818978488445, + 0.008006622083485126, + 0.006221282761543989, + 0.013206059113144875, + 0.01917998306453228, + 0.006114872172474861, + 0.007347323931753635, + 0.002942399587482214, + 0.005140360910445452, + 0.003749032272025943, + 0.0025374949909746647, + 0.005026547703891993, + 0.004916221369057894, + 0.007725344970822334, + 0.005013358313590288, + 0.0020390735007822514, + 0.0038248347118496895, + 0.008116758428514004, + 0.003984496463090181, + 0.006656421814113855, + 0.010044928640127182, + 0.010644516907632351, + 0.00799639243632555, + 0.003940626513212919, + 0.008663475513458252, + 0.004125622101128101, + 0.004945802036672831, + 0.006991143338382244, + 0.012485374696552753, + 0.005774273071438074, + 0.0050082397647202015, + 0.0019612398464232683, + 0.0031789152417331934, + 0.0034477480221539736, + 0.004639317747205496, + 0.005229723174124956, + 0.006782099138945341, + 0.01846279576420784, + 0.022149749100208282, + 0.008024566806852818, + 0.011986473575234413, + 0.018844883888959885, + 0.007274816744029522, + 0.015772400423884392, + 0.016731970012187958, + 0.013553984463214874, + 0.013181603513658047, + 0.007132653146982193, + 0.012371753342449665, + 0.013338463380932808, + 0.01307657826691866, + 0.011673073284327984, + 0.011825291439890862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.07003270834684372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-length, static shot places us in a dimly lit traditional kendo arena, where two kendo practitioners, clad in full traditional protective gear (bogu), are locked in a tense standoff. A harsh overhead light cuts fiercely into the frame from above and sideways, casting deep, sharply defined shadows on their bodies and the polished wooden floor. The entire scene is rendered in Daido Moriyama's signature grainy, high-contrast black-and-white style, with virtually no gray between the stark black and white. The light creates sharp, bright edges on the lacquered metal visors of their breastplates (do) and helmets (men), contrasting starkly with the sheer darkness of their dark kendo uniforms (keikogi) and hakama culottes (skirt pants). Suddenly, the swordsman in the foreground plunges forward, then lunges forward with explosive force, his right foot stomping decisively into the floor, stirring up a fine dust cloud. He raises his bamboo sword (shinai) overhead in a sharp arc, its immense speed leaving a visible, blurry white afterimage as it hurtles towards his opponent, who, poised to block the devastating blow, holds his balance. The entire scene is imbued with extreme dynamic tension and the aesthetics of a frozen moment, conveying a raw, intense, and intensely focused atmosphere of the duel.", + "seed": 25468, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_8/index_2991_seed_25468_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13251264886930586, + "vlm_logprob": -0.13251264886930586, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.13251264886930586, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5858272910118103 + ], + "opposite_logprob_tokens": [ + -0.8358272910118103 + ], + "logprob_reward": -0.5858272910118103, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5858272910118103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23216202855110168, + 0.09304092824459076, + 0.049832895398139954, + 0.11588042229413986, + 0.08411398530006409, + 0.07294882833957672, + 0.15617333352565765, + 0.031210828572511673, + 0.11965357512235641, + 0.04498317092657089 + ], + "credit_map": [ + 0.04008602350950241, + 0.04245327413082123, + 0.03335091844201088, + 0.021805981174111366, + 0.029450593516230583, + 0.025532159954309464, + 0.016185514628887177, + 0.023297559469938278, + 0.014593550004065037, + 0.01070913951843977, + 0.013195044361054897, + 0.008412377908825874, + 0.009266186505556107, + 0.015887368470430374, + 0.007313176523894072, + 0.01366408634930849, + 0.007374709937721491, + 0.004120992962270975, + 0.007564593106508255, + 0.004023863468319178, + 0.003774260636419058, + 0.009610421024262905, + 0.003645913442596793, + 0.00971813965588808, + 0.01597411558032036, + 0.008219782263040543, + 0.014831200242042542, + 0.009958531707525253, + 0.014566454105079174, + 0.021804282441735268, + 0.014650014229118824, + 0.015876032412052155, + 0.012721620500087738, + 0.01435215026140213, + 0.010275699198246002, + 0.005422092508524656, + 0.010105294175446033, + 0.013342098332941532, + 0.0068466393277049065, + 0.011048391461372375, + 0.009276866912841797, + 0.011251821182668209, + 0.010230466723442078, + 0.004282223526388407, + 0.008823361247777939, + 0.010807315818965435, + 0.007117464207112789, + 0.011159304529428482, + 0.0178487878292799, + 0.015970503911376, + 0.017995035275816917, + 0.014743747189640999, + 0.02632143720984459, + 0.023738078773021698, + 0.022590219974517822, + 0.01696552336215973, + 0.005723468028008938, + 0.00366609008051455, + 0.003573303809389472, + 0.0024189201649278402, + 0.00534192705526948, + 0.004072611220180988, + 0.003691332880407572, + 0.00272317579947412, + 0.013800528831779957, + 0.011326196603477001, + 0.01721089705824852, + 0.009949189610779285, + 0.012886001728475094, + 0.022304072976112366, + 0.01691688783466816, + 0.015259798616170883, + 0.005943930707871914, + 0.004325828514993191, + 0.006144877057522535, + 0.004204344004392624, + 0.005123134236782789, + 0.006449637468904257, + 0.006151043344289064, + 0.006640381645411253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.5858272910118103 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028477851301431656 + ], + "opposite_logprob_tokens": [ + -5.877847671508789 + ], + "logprob_reward": -0.0028477851301431656, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028477851301431656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21461504697799683, + 0.07280989736318588, + 0.048001814633607864, + 0.2610229253768921, + 0.05953631550073624, + 0.0431206151843071, + 0.14115847647190094, + 0.045489661395549774, + 0.057953573763370514, + 0.056291721761226654 + ], + "credit_map": [ + 0.028401102870702744, + 0.0420122928917408, + 0.026821352541446686, + 0.026386167854070663, + 0.026883743703365326, + 0.023538773879408836, + 0.019865823909640312, + 0.02070576697587967, + 0.00925231259316206, + 0.007416096515953541, + 0.011276868171989918, + 0.008068029768764973, + 0.0075434185564517975, + 0.011208437383174896, + 0.007647112011909485, + 0.010397618636488914, + 0.0063876984640955925, + 0.0036738449707627296, + 0.0074522001668810844, + 0.004389294423162937, + 0.004584928974509239, + 0.00947619043290615, + 0.003830078523606062, + 0.008207570761442184, + 0.0254481453448534, + 0.011316818185150623, + 0.0739893764257431, + 0.012582071125507355, + 0.015749063342809677, + 0.09535817056894302, + 0.014418764971196651, + 0.01216050609946251, + 0.009437765926122665, + 0.011228960007429123, + 0.0077682118862867355, + 0.0034101358614861965, + 0.007415690924972296, + 0.007797213736921549, + 0.004898506682366133, + 0.0075798239558935165, + 0.004479824565351009, + 0.004855082370340824, + 0.005702789407223463, + 0.00330949854105711, + 0.006702860817313194, + 0.005985387600958347, + 0.0070488532073795795, + 0.005036314483731985, + 0.014103824272751808, + 0.014308713376522064, + 0.016335126012563705, + 0.014418467879295349, + 0.021453384310007095, + 0.02306566759943962, + 0.022173108533024788, + 0.015300184488296509, + 0.006096530240029097, + 0.005928960628807545, + 0.004046973306685686, + 0.0038354145362973213, + 0.01004115678369999, + 0.005514026153832674, + 0.0056257834658026695, + 0.0044008102267980576, + 0.007253625895828009, + 0.005383261013776064, + 0.007396083325147629, + 0.00389788462780416, + 0.0063170455396175385, + 0.011182468384504318, + 0.008416895754635334, + 0.008106301538646221, + 0.005906597711145878, + 0.005418932065367699, + 0.008144636638462543, + 0.004325374495238066, + 0.007606217637658119, + 0.008841452188789845, + 0.006809971760958433, + 0.009238533675670624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.0028477851301431656 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004650965798646212 + ], + "opposite_logprob_tokens": [ + -5.379651069641113 + ], + "logprob_reward": -0.004650965798646212, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004650965798646212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2382378727197647, + 0.07532694935798645, + 0.07100562751293182, + 0.28258851170539856, + 0.03984494134783745, + 0.0353064090013504, + 0.07326395064592361, + 0.11750569194555283, + 0.038354575634002686, + 0.028565507382154465 + ], + "credit_map": [ + 0.033395878970623016, + 0.04752443730831146, + 0.030321883037686348, + 0.024039436131715775, + 0.03583924099802971, + 0.024894876405596733, + 0.018149159848690033, + 0.024072960019111633, + 0.010282762348651886, + 0.010263199917972088, + 0.009238158352673054, + 0.00739271380007267, + 0.008993766270577908, + 0.013140292838215828, + 0.006111341994255781, + 0.009904714301228523, + 0.010868433862924576, + 0.0059186965227127075, + 0.014244034886360168, + 0.007084527052938938, + 0.006539517547935247, + 0.01140590850263834, + 0.005453102290630341, + 0.009491405449807644, + 0.020457126200199127, + 0.008053604513406754, + 0.09516821056604385, + 0.009701204486191273, + 0.01115842629224062, + 0.10732093453407288, + 0.015655452385544777, + 0.015073554590344429, + 0.004704460501670837, + 0.00481448695063591, + 0.005654327105730772, + 0.0025581682566553354, + 0.006606698501855135, + 0.006602538749575615, + 0.003928407095372677, + 0.004975850693881512, + 0.0052515240386128426, + 0.004755072761327028, + 0.004045999143272638, + 0.003334901062771678, + 0.005352971144020557, + 0.004121517296880484, + 0.004671134985983372, + 0.003773289266973734, + 0.007854192517697811, + 0.00888881180435419, + 0.009393525309860706, + 0.006586868315935135, + 0.011353458277881145, + 0.01209911610931158, + 0.009532086551189423, + 0.007555889431387186, + 0.018369009718298912, + 0.015417642891407013, + 0.011220415122807026, + 0.007430807687342167, + 0.025785868987441063, + 0.016263343393802643, + 0.012467646040022373, + 0.01055095810443163, + 0.004201331175863743, + 0.003771711140871048, + 0.004559026565402746, + 0.0032126575242727995, + 0.006545747630298138, + 0.005515587516129017, + 0.0054363105446100235, + 0.0051122019067406654, + 0.002861947286874056, + 0.0027317453641444445, + 0.004283437971025705, + 0.002382541773840785, + 0.003580959513783455, + 0.004154152702540159, + 0.004142272286117077, + 0.004428450483828783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.004650965798646212 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020578335970640182 + ], + "opposite_logprob_tokens": [ + -3.895578384399414 + ], + "logprob_reward": -0.020578335970640182, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020578335970640182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09357240796089172, + 0.04111681878566742, + 0.1271921545267105, + 0.25163912773132324, + 0.06825985759496689, + 0.0977277010679245, + 0.055346548557281494, + 0.12959878146648407, + 0.07342306524515152, + 0.062123529613018036 + ], + "credit_map": [ + 0.014386185444891453, + 0.014229868538677692, + 0.011830405332148075, + 0.01161721721291542, + 0.011315261013805866, + 0.010683435946702957, + 0.008040998131036758, + 0.011469040997326374, + 0.006530824117362499, + 0.003888180013746023, + 0.0059224991127848625, + 0.004392520524561405, + 0.004245951771736145, + 0.006863415241241455, + 0.003405520226806402, + 0.005867909640073776, + 0.02064715512096882, + 0.010109703056514263, + 0.01895499788224697, + 0.008808408863842487, + 0.009672820568084717, + 0.018157806247472763, + 0.009948916733264923, + 0.030892331153154373, + 0.043036993592977524, + 0.022780122235417366, + 0.057452112436294556, + 0.01709265448153019, + 0.018092181533575058, + 0.037584371864795685, + 0.018874026834964752, + 0.036726657301187515, + 0.01068382989615202, + 0.00834503024816513, + 0.007877279072999954, + 0.0040588672272861, + 0.009639596566557884, + 0.011753504164516926, + 0.007892921566963196, + 0.008008833974599838, + 0.013276916928589344, + 0.011531271040439606, + 0.010358951985836029, + 0.007331131026148796, + 0.017577672377228737, + 0.014131865464150906, + 0.011821866035461426, + 0.011698027141392231, + 0.006207671016454697, + 0.005624087061733007, + 0.006237606052309275, + 0.005141503177583218, + 0.00945420190691948, + 0.007132906001061201, + 0.00858316384255886, + 0.006965408101677895, + 0.014121877029538155, + 0.015004079788923264, + 0.012875862419605255, + 0.011161844246089458, + 0.03285375237464905, + 0.015790676698088646, + 0.015594176948070526, + 0.01219650823622942, + 0.007457412779331207, + 0.007068336009979248, + 0.008100497536361217, + 0.0060735284350812435, + 0.009361507371068, + 0.012897041626274586, + 0.01409609243273735, + 0.008368650451302528, + 0.006436392664909363, + 0.0063113863579928875, + 0.008899589069187641, + 0.007636264897882938, + 0.009577615186572075, + 0.007268196437507868, + 0.008961153216660023, + 0.007032930850982666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 2 + ], + "selected_reward": -0.020578335970640182 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865886643528938 + ], + "opposite_logprob_tokens": [ + -3.048658847808838 + ], + "logprob_reward": -0.04865886643528938, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865886643528938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0763428658246994, + 0.05273260921239853, + 0.08868918567895889, + 0.5092064738273621, + 0.03104582615196705, + 0.04155638813972473, + 0.03889878839254379, + 0.03981037065386772, + 0.06701989471912384, + 0.05469764769077301 + ], + "credit_map": [ + 0.008533102460205555, + 0.011706468649208546, + 0.011353451758623123, + 0.008894868195056915, + 0.0089202169328928, + 0.00917807687073946, + 0.008774722926318645, + 0.008981948718428612, + 0.00707298843190074, + 0.005181134678423405, + 0.007456368301063776, + 0.0049964687786996365, + 0.004074313677847385, + 0.010027975775301456, + 0.004278885666280985, + 0.009644470177590847, + 0.014069092459976673, + 0.0063943322747945786, + 0.016402676701545715, + 0.006051019765436649, + 0.007670159451663494, + 0.012931986711919308, + 0.005446014925837517, + 0.01972389593720436, + 0.043598152697086334, + 0.02002343162894249, + 0.1743623912334442, + 0.01613806001842022, + 0.015157213434576988, + 0.193220317363739, + 0.020878583192825317, + 0.02582833543419838, + 0.004300661850720644, + 0.002889452502131462, + 0.004885418340563774, + 0.002908351831138134, + 0.003205913119018078, + 0.005240779835730791, + 0.003663009265437722, + 0.003952237777411938, + 0.004421209450811148, + 0.004528198856860399, + 0.0035389899276196957, + 0.003210993018001318, + 0.00925084762275219, + 0.0058640786446630955, + 0.006279237102717161, + 0.004462831187993288, + 0.0041740573942661285, + 0.0036173255648463964, + 0.004346273839473724, + 0.0034306286834180355, + 0.0062580471858382225, + 0.006286791060119867, + 0.006847770418971777, + 0.003937890287488699, + 0.004829672165215015, + 0.004252366255968809, + 0.00448024133220315, + 0.0031942464411258698, + 0.00913149956613779, + 0.005133265163749456, + 0.00517621822655201, + 0.0036128596402704716, + 0.007057428825646639, + 0.007800288964062929, + 0.00801083818078041, + 0.006089197937399149, + 0.00861207116395235, + 0.010440578684210777, + 0.008827241137623787, + 0.010182245634496212, + 0.006204416509717703, + 0.005169422831386328, + 0.007912806235253811, + 0.0049761212430894375, + 0.006154970731586218, + 0.007145491428673267, + 0.008781163021922112, + 0.00835325289517641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.04865886643528938 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-length, static shot places us in a dimly lit traditional kendo arena, where two kendo practitioners, clad in full traditional protective gear (bogu), are locked in a tense standoff. A harsh overhead light cuts fiercely into the frame from above and sideways, casting deep, sharply defined shadows on their bodies and the polished wooden floor. The entire scene is rendered in Daido Moriyama's signature grainy, high-contrast black-and-white style, with virtually no gray between the stark black and white. The light creates sharp, bright edges on the lacquered metal visors of their breastplates (do) and helmets (men), contrasting starkly with the sheer darkness of their dark kendo uniforms (keikogi) and hakama culottes (skirt pants). Suddenly, the swordsman in the foreground plunges forward, then lunges forward with explosive force, his right foot stomping decisively into the floor, stirring up a fine dust cloud. He raises his bamboo sword (shinai) overhead in a sharp arc, its immense speed leaving a visible, blurry white afterimage as it hurtles towards his opponent, who, poised to block the devastating blow, holds his balance. The entire scene is imbued with extreme dynamic tension and the aesthetics of a frozen moment, conveying a raw, intense, and intensely focused atmosphere of the duel.", + "seed": 25469, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_8/index_2991_seed_25469_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0314234871417284, + "vlm_logprob": -0.0314234871417284, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.0314234871417284, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07528357207775116 + ], + "opposite_logprob_tokens": [ + -2.7002835273742676 + ], + "logprob_reward": -0.07528357207775116, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07528357207775116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34118548035621643, + 0.03410867601633072, + 0.032866042107343674, + 0.032037653028964996, + 0.025612331926822662, + 0.04762856662273407, + 0.03125395625829697, + 0.03963238000869751, + 0.057025499641895294, + 0.35864943265914917 + ], + "credit_map": [ + 0.04941311478614807, + 0.06040113791823387, + 0.050720904022455215, + 0.04751502349972725, + 0.05089531093835831, + 0.026894723996520042, + 0.029476093128323555, + 0.025869157165288925, + 0.004998825024813414, + 0.0067110927775502205, + 0.0036721238866448402, + 0.0035821266938000917, + 0.005549892783164978, + 0.0033715746831148863, + 0.0027550661470741034, + 0.0034679763484746218, + 0.004573503974825144, + 0.005121372174471617, + 0.0027266666293144226, + 0.0042559914290905, + 0.005548472981899977, + 0.0029145395383238792, + 0.004141140263527632, + 0.0035843534860759974, + 0.0048539466224610806, + 0.004219597205519676, + 0.0034607069101184607, + 0.002689009765163064, + 0.004970321897417307, + 0.003779300721362233, + 0.005271882750093937, + 0.0027928869239985943, + 0.003978042863309383, + 0.0037800646387040615, + 0.0033668323885649443, + 0.0023356880992650986, + 0.0033391485922038555, + 0.0018684687092900276, + 0.0032852806616574526, + 0.0036588062066584826, + 0.007173218764364719, + 0.007250004913657904, + 0.006252564489841461, + 0.0043289815075695515, + 0.007332773879170418, + 0.004958793520927429, + 0.004694086499512196, + 0.0056381430476903915, + 0.005027238745242357, + 0.003771288553252816, + 0.004105324391275644, + 0.001839703880250454, + 0.004442030563950539, + 0.0033898288384079933, + 0.0033591166138648987, + 0.0053194244392216206, + 0.005893666297197342, + 0.006929328665137291, + 0.0035380010958760977, + 0.002659743884578347, + 0.0083081079646945, + 0.0035002015065401793, + 0.004768097773194313, + 0.004035235848277807, + 0.0077041564509272575, + 0.012925621122121811, + 0.008543003350496292, + 0.005101439543068409, + 0.00743196252733469, + 0.0055354442447423935, + 0.005212276708334684, + 0.004571592900902033, + 0.05281483381986618, + 0.038303546607494354, + 0.061580460518598557, + 0.04899205267429352, + 0.046986039727926254, + 0.03467213734984398, + 0.03667445853352547, + 0.038625892251729965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.07528357207775116 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009817293845117092 + ], + "opposite_logprob_tokens": [ + -4.634817123413086 + ], + "logprob_reward": -0.009817293845117092, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009817293845117092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20316144824028015, + 0.08357557654380798, + 0.13153491914272308, + 0.08546672016382217, + 0.07141221314668655, + 0.06856709718704224, + 0.06641823053359985, + 0.12270669639110565, + 0.12480820715427399, + 0.042348962277173996 + ], + "credit_map": [ + 0.026551293209195137, + 0.029812179505825043, + 0.023525912314653397, + 0.04889904335141182, + 0.026108309626579285, + 0.011511106975376606, + 0.0209676343947649, + 0.015785951167345047, + 0.013598094694316387, + 0.01340706367045641, + 0.006044032983481884, + 0.00973585620522499, + 0.021366402506828308, + 0.006054719444364309, + 0.007278879638761282, + 0.0060905166901648045, + 0.0185526255518198, + 0.016216235235333443, + 0.017184903845191002, + 0.022163668647408485, + 0.023575380444526672, + 0.010851312428712845, + 0.012715300545096397, + 0.010275479406118393, + 0.011987539008259773, + 0.013591221533715725, + 0.007531636860221624, + 0.006985723040997982, + 0.01815006136894226, + 0.0050649214535951614, + 0.01437152735888958, + 0.007784082088619471, + 0.010921084322035313, + 0.009480930864810944, + 0.008828421123325825, + 0.007800402119755745, + 0.010899481363594532, + 0.004885527305305004, + 0.011732000857591629, + 0.006864361464977264, + 0.010088318027555943, + 0.01068027876317501, + 0.0075743417255580425, + 0.004382443148642778, + 0.017506705597043037, + 0.006046256981790066, + 0.005674675107002258, + 0.00661407271400094, + 0.010667400434613228, + 0.01007122453302145, + 0.005586711224168539, + 0.0036824450362473726, + 0.011992614716291428, + 0.00904409121721983, + 0.008568406105041504, + 0.006805334705859423, + 0.0158867035061121, + 0.021048834547400475, + 0.012703432701528072, + 0.007474200334399939, + 0.01930469460785389, + 0.014607361517846584, + 0.01588071696460247, + 0.01580074243247509, + 0.016037672758102417, + 0.024356527253985405, + 0.008408932946622372, + 0.009484636597335339, + 0.02519647590816021, + 0.011438990011811256, + 0.018454082310199738, + 0.011430880054831505, + 0.006180628202855587, + 0.004306131042540073, + 0.003139178967103362, + 0.0022856646683067083, + 0.008428281173110008, + 0.0043399883434176445, + 0.0067962221801280975, + 0.006872861180454493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.009817293845117092 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004105354659259319 + ], + "opposite_logprob_tokens": [ + -5.504105567932129 + ], + "logprob_reward": -0.004105354659259319, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004105354659259319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3022642731666565, + 0.11597080528736115, + 0.2696544826030731, + 0.037250228226184845, + 0.049518782645463943, + 0.030599866062402725, + 0.06521698832511902, + 0.05534540116786957, + 0.029917512089014053, + 0.0442616306245327 + ], + "credit_map": [ + 0.045088041573762894, + 0.046787671744823456, + 0.036288898438215256, + 0.03132098540663719, + 0.07106947898864746, + 0.026380201801657677, + 0.029093287885189056, + 0.016235727816820145, + 0.016763580963015556, + 0.017396608367562294, + 0.009317160584032536, + 0.01672440953552723, + 0.028201933950185776, + 0.00854488369077444, + 0.011455767787992954, + 0.007566469721496105, + 0.04023074358701706, + 0.03460255637764931, + 0.03366163745522499, + 0.041242074221372604, + 0.041461627930402756, + 0.024698106572031975, + 0.025608697906136513, + 0.028149034827947617, + 0.004922799300402403, + 0.004892353899776936, + 0.0039042532444000244, + 0.0034091859124600887, + 0.007778834085911512, + 0.0033207919914275408, + 0.0064742048271000385, + 0.0025478051975369453, + 0.0071443659253418446, + 0.010112208314239979, + 0.00618118979036808, + 0.005672161001712084, + 0.0063310484401881695, + 0.00429797638207674, + 0.00453296210616827, + 0.00524686835706234, + 0.005227961111813784, + 0.0057212114334106445, + 0.004062636289745569, + 0.0019811587408185005, + 0.0034456816501915455, + 0.003200779901817441, + 0.002827201271429658, + 0.004133237991482019, + 0.010095059871673584, + 0.010954751633107662, + 0.012917893007397652, + 0.005067978985607624, + 0.007564673200249672, + 0.007910555228590965, + 0.004851773846894503, + 0.005854298360645771, + 0.007840219885110855, + 0.008760657161474228, + 0.005597829353064299, + 0.0029886355623602867, + 0.010751825757324696, + 0.006988417357206345, + 0.006367741618305445, + 0.006050067953765392, + 0.004165248479694128, + 0.005704717710614204, + 0.003459234954789281, + 0.0023119712714105844, + 0.004834484308958054, + 0.002502788556739688, + 0.0036905582528561354, + 0.003248506458476186, + 0.0061492775566875935, + 0.005819873884320259, + 0.0059568569995462894, + 0.0032289247028529644, + 0.006751351989805698, + 0.004474357236176729, + 0.005766859278082848, + 0.006114129908382893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.004105354659259319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796700924634933 + ], + "opposite_logprob_tokens": [ + -3.529796600341797 + ], + "logprob_reward": -0.029796700924634933, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796700924634933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1872633844614029, + 0.06505262106657028, + 0.10599944740533829, + 0.17850305140018463, + 0.08900080621242523, + 0.0659596398472786, + 0.07618329674005508, + 0.07672001421451569, + 0.1078554093837738, + 0.04746236652135849 + ], + "credit_map": [ + 0.030948713421821594, + 0.02929433062672615, + 0.020052021369338036, + 0.02397262677550316, + 0.04014420881867409, + 0.011534336023032665, + 0.016667071729898453, + 0.014650066383183002, + 0.00940671656280756, + 0.009750796481966972, + 0.00554717518389225, + 0.007750598713755608, + 0.016937825828790665, + 0.005211748648434877, + 0.004892451222985983, + 0.0055553060956299305, + 0.0172569677233696, + 0.014551571570336819, + 0.011674173176288605, + 0.013595685362815857, + 0.015387787483632565, + 0.009975435212254524, + 0.010287392884492874, + 0.013270438648760319, + 0.025850903242826462, + 0.02527735009789467, + 0.019790638238191605, + 0.024036001414060593, + 0.032369472086429596, + 0.009993714280426502, + 0.029872287064790726, + 0.011312678456306458, + 0.013438527472317219, + 0.012032129801809788, + 0.012063227593898773, + 0.008206001482903957, + 0.015112548135221004, + 0.007593045011162758, + 0.010221992619335651, + 0.01033332571387291, + 0.009270359762012959, + 0.00781412422657013, + 0.011058192700147629, + 0.004870886914432049, + 0.011898157186806202, + 0.007409898564219475, + 0.006259816233068705, + 0.007378201000392437, + 0.011758247390389442, + 0.014775145798921585, + 0.0070095849223434925, + 0.006796128116548061, + 0.010678359307348728, + 0.0074431817047297955, + 0.011414083652198315, + 0.0063085611909627914, + 0.011000641621649265, + 0.010984296910464764, + 0.007221680134534836, + 0.004309612326323986, + 0.019739367067813873, + 0.0066072833724319935, + 0.010505909100174904, + 0.006351222284138203, + 0.01732056401669979, + 0.01840941049158573, + 0.008550714701414108, + 0.006379387341439724, + 0.018036669120192528, + 0.008644099347293377, + 0.014447394758462906, + 0.016067175194621086, + 0.007424552459269762, + 0.005266280844807625, + 0.005086992867290974, + 0.003457627259194851, + 0.006887839641422033, + 0.003907995764166117, + 0.008966871537268162, + 0.006464208941906691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.029796700924634933 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381145142018795 + ], + "opposite_logprob_tokens": [ + -3.288114547729492 + ], + "logprob_reward": -0.0381145142018795, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381145142018795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10678602010011673, + 0.052860014140605927, + 0.10842304676771164, + 0.06575792282819748, + 0.11552859097719193, + 0.10752632468938828, + 0.04568399116396904, + 0.2138904631137848, + 0.059044551104307175, + 0.12449906021356583 + ], + "credit_map": [ + 0.014227109961211681, + 0.015828318893909454, + 0.015654467046260834, + 0.016846656799316406, + 0.016620613634586334, + 0.006503115873783827, + 0.011621682904660702, + 0.009484052658081055, + 0.007398572284728289, + 0.010125947184860706, + 0.003997497726231813, + 0.00674663158133626, + 0.011412044055759907, + 0.004226084798574448, + 0.004449123982340097, + 0.004504113923758268, + 0.01890905573964119, + 0.019941486418247223, + 0.008602899499237537, + 0.01234529074281454, + 0.013599301688373089, + 0.008236749097704887, + 0.012294521555304527, + 0.014493752270936966, + 0.008771068416535854, + 0.007300186436623335, + 0.005111261736601591, + 0.006355937570333481, + 0.011461202055215836, + 0.005241671577095985, + 0.015477176755666733, + 0.006039419211447239, + 0.01657439023256302, + 0.019703010097146034, + 0.01947384513914585, + 0.014677055180072784, + 0.01390034705400467, + 0.008789327926933765, + 0.011577479541301727, + 0.010833140462636948, + 0.015622105449438095, + 0.020557858049869537, + 0.010628310963511467, + 0.007530177943408489, + 0.022051172330975533, + 0.008999759331345558, + 0.01124055776745081, + 0.010896380990743637, + 0.0064777503721416, + 0.007962415926158428, + 0.004334752913564444, + 0.00250225979834795, + 0.006530939135700464, + 0.006385695654898882, + 0.005236721131950617, + 0.006253460422158241, + 0.030654223635792732, + 0.02949335426092148, + 0.02823980152606964, + 0.010787086561322212, + 0.04349885508418083, + 0.02567484602332115, + 0.026998411864042282, + 0.018543889746069908, + 0.008275647647678852, + 0.008435492403805256, + 0.007668091915547848, + 0.00577439134940505, + 0.008272122591733932, + 0.0055822646245360374, + 0.007330169435590506, + 0.007706371136009693, + 0.016759607940912247, + 0.01637968420982361, + 0.012220298871397972, + 0.013121378608047962, + 0.023054849356412888, + 0.012487745843827724, + 0.019220741465687752, + 0.011254750192165375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.0381145142018795 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium-length, static shot places us in a dimly lit traditional kendo arena, where two kendo practitioners, clad in full traditional protective gear (bogu), are locked in a tense standoff. A harsh overhead light cuts fiercely into the frame from above and sideways, casting deep, sharply defined shadows on their bodies and the polished wooden floor. The entire scene is rendered in Daido Moriyama's signature grainy, high-contrast black-and-white style, with virtually no gray between the stark black and white. The light creates sharp, bright edges on the lacquered metal visors of their breastplates (do) and helmets (men), contrasting starkly with the sheer darkness of their dark kendo uniforms (keikogi) and hakama culottes (skirt pants). Suddenly, the swordsman in the foreground plunges forward, then lunges forward with explosive force, his right foot stomping decisively into the floor, stirring up a fine dust cloud. He raises his bamboo sword (shinai) overhead in a sharp arc, its immense speed leaving a visible, blurry white afterimage as it hurtles towards his opponent, who, poised to block the devastating blow, holds his balance. The entire scene is imbued with extreme dynamic tension and the aesthetics of a frozen moment, conveying a raw, intense, and intensely focused atmosphere of the duel.", + "seed": 25470, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000129/group_8/index_2991_seed_25470_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02835881458595395, + "vlm_logprob": -0.02835881458595395, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.02835881458595395, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04227578639984131 + ], + "opposite_logprob_tokens": [ + -3.417275905609131 + ], + "logprob_reward": -0.04227578639984131, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04227578639984131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16698777675628662, + 0.06794311851263046, + 0.05102870985865593, + 0.030431250110268593, + 0.08842578530311584, + 0.16289173066616058, + 0.21243053674697876, + 0.08902229368686676, + 0.08847421407699585, + 0.042364589869976044 + ], + "credit_map": [ + 0.012922844849526882, + 0.020785793662071228, + 0.019027385860681534, + 0.026376349851489067, + 0.02032248117029667, + 0.02508675679564476, + 0.011464583687484264, + 0.031001582741737366, + 0.004551347345113754, + 0.013192329555749893, + 0.0040550222620368, + 0.011813112534582615, + 0.006494268774986267, + 0.009572165086865425, + 0.007398947142064571, + 0.010865925811231136, + 0.004358474165201187, + 0.006161123048514128, + 0.004651073832064867, + 0.009822643361985683, + 0.005857352167367935, + 0.008873797953128815, + 0.0043880934827029705, + 0.00691614905372262, + 0.00284945173189044, + 0.003565341467037797, + 0.0027842943090945482, + 0.005462720058858395, + 0.0037748506292700768, + 0.004607677925378084, + 0.0031365705654025078, + 0.004250342026352882, + 0.005052570253610611, + 0.0075590224005281925, + 0.005735518876463175, + 0.011579669080674648, + 0.020592454820871353, + 0.014391722157597542, + 0.013683531433343887, + 0.009831291623413563, + 0.010035627521574497, + 0.015594047494232655, + 0.007919404655694962, + 0.027456380426883698, + 0.0378325879573822, + 0.027890928089618683, + 0.02035565860569477, + 0.015807099640369415, + 0.011544182896614075, + 0.023256709799170494, + 0.011780143715441227, + 0.0413120836019516, + 0.024039486423134804, + 0.04461060091853142, + 0.015423115342855453, + 0.04046422615647316, + 0.005148329772055149, + 0.006197579670697451, + 0.003549752989783883, + 0.007598449010401964, + 0.00668258685618639, + 0.03236263990402222, + 0.005780145525932312, + 0.02170281484723091, + 0.0052130138501524925, + 0.008693937212228775, + 0.0043838368728756905, + 0.012088432908058167, + 0.019198287278413773, + 0.015510477125644684, + 0.009983403608202934, + 0.013402823358774185, + 0.0026016668416559696, + 0.0036181702744215727, + 0.002791395876556635, + 0.005920421332120895, + 0.002891508862376213, + 0.009036066010594368, + 0.0059086657129228115, + 0.009596694260835648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.04227578639984131 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00866274069994688 + ], + "opposite_logprob_tokens": [ + -4.758662700653076 + ], + "logprob_reward": -0.00866274069994688, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00866274069994688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22975972294807434, + 0.11471999436616898, + 0.041076015681028366, + 0.06071312725543976, + 0.03411320969462395, + 0.07129079848527908, + 0.0755847692489624, + 0.14301671087741852, + 0.19103923439979553, + 0.03868643939495087 + ], + "credit_map": [ + 0.009615852497518063, + 0.03082170896232128, + 0.021161731332540512, + 0.050363972783088684, + 0.027190877124667168, + 0.029467394575476646, + 0.018433623015880585, + 0.04270455613732338, + 0.006565706804394722, + 0.02460576966404915, + 0.006756102666258812, + 0.02710186503827572, + 0.009581946767866611, + 0.015464645810425282, + 0.009439067915081978, + 0.015204883180558681, + 0.0024618434254080057, + 0.00589235732331872, + 0.003011486493051052, + 0.006733478978276253, + 0.004627786111086607, + 0.009887101128697395, + 0.0029166850727051497, + 0.005545276682823896, + 0.003756124060600996, + 0.006270524114370346, + 0.002606211695820093, + 0.01898684725165367, + 0.005999429617077112, + 0.008722626604139805, + 0.004792932886630297, + 0.009578431025147438, + 0.0017001398373395205, + 0.004013489466160536, + 0.0013745754258707166, + 0.005312009248882532, + 0.008340255357325077, + 0.005075049586594105, + 0.005422378424555063, + 0.00287531316280365, + 0.003921304363757372, + 0.009953342378139496, + 0.0032966274302452803, + 0.014591292478144169, + 0.012387503869831562, + 0.010880985297262669, + 0.008220815099775791, + 0.008038928732275963, + 0.004243914503604174, + 0.009587403386831284, + 0.003647724399343133, + 0.013980600982904434, + 0.009713280014693737, + 0.014275225810706615, + 0.0064513664692640305, + 0.013685254380106926, + 0.008548158220946789, + 0.010916456580162048, + 0.005627119913697243, + 0.009051881730556488, + 0.008203686214983463, + 0.05402227118611336, + 0.006069682538509369, + 0.04057746008038521, + 0.011208304204046726, + 0.02177664265036583, + 0.010719822719693184, + 0.025744987651705742, + 0.023400986567139626, + 0.04005539417266846, + 0.01655440218746662, + 0.04157869517803192, + 0.0022195603232830763, + 0.00600054394453764, + 0.002722497098147869, + 0.00777829997241497, + 0.0033351897727698088, + 0.0051707010716199875, + 0.006247326731681824, + 0.005212321411818266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.00866274069994688 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004106185864657164 + ], + "opposite_logprob_tokens": [ + -5.504106044769287 + ], + "logprob_reward": -0.004106185864657164, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004106185864657164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15600837767124176, + 0.04596756398677826, + 0.026528682559728622, + 0.018660182133316994, + 0.04884721338748932, + 0.037173476070165634, + 0.28345948457717896, + 0.12124741077423096, + 0.2164457142353058, + 0.04566185176372528 + ], + "credit_map": [ + 0.011103814467787743, + 0.029502088204026222, + 0.00875230971723795, + 0.020136522129178047, + 0.027056844905018806, + 0.021800758317112923, + 0.014584343880414963, + 0.02307169698178768, + 0.0027079100254923105, + 0.006542489398270845, + 0.0024990886449813843, + 0.00935183372348547, + 0.0040426477789878845, + 0.009238009341061115, + 0.004096870310604572, + 0.007488717790693045, + 0.002310380572453141, + 0.0033840348478406668, + 0.002910793060436845, + 0.004961393307894468, + 0.002784609328955412, + 0.004677186254411936, + 0.00259271077811718, + 0.00290757417678833, + 0.0015948484651744366, + 0.0022135975304991007, + 0.001260224380530417, + 0.0034715114161372185, + 0.0014375413302332163, + 0.003981459885835648, + 0.002041397150605917, + 0.002659602090716362, + 0.0038241720758378506, + 0.006089200731366873, + 0.0030741530936211348, + 0.006344265304505825, + 0.007094434462487698, + 0.01031806692481041, + 0.006765155587345362, + 0.005337765906006098, + 0.002175914356485009, + 0.004764472600072622, + 0.002492584055289626, + 0.007610332686454058, + 0.006398613099008799, + 0.005867651663720608, + 0.004098610952496529, + 0.0037652961909770966, + 0.015454715117812157, + 0.03925151377916336, + 0.008801538497209549, + 0.03435105085372925, + 0.06350930780172348, + 0.04897437244653702, + 0.0327686071395874, + 0.04034839943051338, + 0.006004860159009695, + 0.013327966444194317, + 0.004253724589943886, + 0.01163202803581953, + 0.007504853419959545, + 0.041971176862716675, + 0.004766505677253008, + 0.03178630769252777, + 0.009573335759341717, + 0.022740710526704788, + 0.012341299094259739, + 0.025540314614772797, + 0.024024074897170067, + 0.05939994379878044, + 0.015348469838500023, + 0.04747757315635681, + 0.002778697991743684, + 0.0054282271303236485, + 0.0029600244015455246, + 0.009311599656939507, + 0.003218429395928979, + 0.007202793378382921, + 0.008022270165383816, + 0.006739807780832052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.004106185864657164 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808720037341118 + ], + "opposite_logprob_tokens": [ + -3.2880871295928955 + ], + "logprob_reward": -0.03808720037341118, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808720037341118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13962745666503906, + 0.08687040954828262, + 0.038136281073093414, + 0.0775999128818512, + 0.057788435369729996, + 0.20683054625988007, + 0.1265692263841629, + 0.06785853952169418, + 0.16583342850208282, + 0.032885827124118805 + ], + "credit_map": [ + 0.00798050593584776, + 0.022223301231861115, + 0.009061448276042938, + 0.03283844143152237, + 0.015684671700000763, + 0.01888100802898407, + 0.007954888045787811, + 0.025003191083669662, + 0.006391332019120455, + 0.01305861584842205, + 0.00661972351372242, + 0.01892964541912079, + 0.0073241861537098885, + 0.013616305775940418, + 0.006535031367093325, + 0.014395571313798428, + 0.0026128666941076517, + 0.003190276911482215, + 0.00289138057269156, + 0.0073982542380690575, + 0.003779478371143341, + 0.009209360927343369, + 0.0029978088568896055, + 0.006056854035705328, + 0.003698816290125251, + 0.009506648406386375, + 0.0030559184961020947, + 0.024040836840867996, + 0.007594838738441467, + 0.012204967439174652, + 0.005014337133616209, + 0.012483551166951656, + 0.0034497848246246576, + 0.006353508681058884, + 0.0025557850021868944, + 0.005942827556282282, + 0.011673679575324059, + 0.007683704607188702, + 0.014734012074768543, + 0.005395133513957262, + 0.012321523390710354, + 0.0237454641610384, + 0.007679404225200415, + 0.028569625690579414, + 0.046646248549222946, + 0.03991758078336716, + 0.020636266097426414, + 0.02731442078948021, + 0.006608359050005674, + 0.01240013912320137, + 0.003785429522395134, + 0.014543836936354637, + 0.018786806613206863, + 0.03199835121631622, + 0.011396211571991444, + 0.027050096541643143, + 0.0034518232569098473, + 0.009640871547162533, + 0.0037609965074807405, + 0.011154471896588802, + 0.008942320942878723, + 0.013427630066871643, + 0.005938150454312563, + 0.011542271822690964, + 0.011771506629884243, + 0.017602937296032906, + 0.008323697373270988, + 0.02611033245921135, + 0.02799658477306366, + 0.036067623645067215, + 0.01788371242582798, + 0.020077038556337357, + 0.0018650353886187077, + 0.004646644461899996, + 0.0019738194532692432, + 0.007548262365162373, + 0.002184330252930522, + 0.0051360419020056725, + 0.004644421860575676, + 0.004887271206825972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.03808720037341118 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit traditional kendo arena with a polished wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two kendo practitioners visible wearing full traditional protective gear (bogu)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a high-contrast black-and-white style with sharp, bright edges on the visors and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foreground swordsman lunge forward with his right foot stomping and raising a bamboo sword (shinai) overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot from a low-angle perspective with a static camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866215959191322 + ], + "opposite_logprob_tokens": [ + -3.0486621856689453 + ], + "logprob_reward": -0.04866215959191322, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866215959191322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16673897206783295, + 0.07221881300210953, + 0.05857465788722038, + 0.04269322007894516, + 0.03463510796427727, + 0.11961348354816437, + 0.1211952194571495, + 0.15263715386390686, + 0.15697738528251648, + 0.0747160017490387 + ], + "credit_map": [ + 0.0111719761043787, + 0.01761999912559986, + 0.00957871787250042, + 0.03913116827607155, + 0.021203849464654922, + 0.02421012707054615, + 0.01334102638065815, + 0.030482130125164986, + 0.00424948800355196, + 0.013015606440603733, + 0.004302800167351961, + 0.016189085319638252, + 0.007518005557358265, + 0.011261139065027237, + 0.005594329442828894, + 0.010088365525007248, + 0.0035752563271671534, + 0.006556808948516846, + 0.0039510843344032764, + 0.010519826784729958, + 0.007602148689329624, + 0.014630116522312164, + 0.0030407358426600695, + 0.008698681369423866, + 0.0025119376368820667, + 0.005026872269809246, + 0.0031146255787461996, + 0.007719496730715036, + 0.005615205504000187, + 0.007433381397277117, + 0.004642674699425697, + 0.006629028357565403, + 0.0030864907894283533, + 0.003058105241507292, + 0.002635680139064789, + 0.004625212866812944, + 0.00654074689373374, + 0.00710675586014986, + 0.004061078652739525, + 0.003521039616316557, + 0.007335560861974955, + 0.016594886779785156, + 0.005537723191082478, + 0.02175520360469818, + 0.030062520876526833, + 0.015609439462423325, + 0.009848043322563171, + 0.012870111502707005, + 0.007786424830555916, + 0.011273854412138462, + 0.005768513306975365, + 0.017238400876522064, + 0.016975848004221916, + 0.030305542051792145, + 0.009155248291790485, + 0.0226913932710886, + 0.00800341833382845, + 0.010932757519185543, + 0.005821348167955875, + 0.010809509083628654, + 0.012981251813471317, + 0.054879456758499146, + 0.008130290545523167, + 0.04107912629842758, + 0.01143313106149435, + 0.016508042812347412, + 0.007698327302932739, + 0.012788302265107632, + 0.01487768068909645, + 0.04785667732357979, + 0.012036599218845367, + 0.033778637647628784, + 0.006156112998723984, + 0.010077391751110554, + 0.006145513616502285, + 0.015333002433180809, + 0.005856504198163748, + 0.010590048506855965, + 0.011054543778300285, + 0.009502886794507504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.04866215959191322 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_0/index_tensor([8496, 8496])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_0/index_tensor([8496, 8496])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..61f7259409127e67733c5526ebe227f0bcdaa2d1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_0/index_tensor([8496, 8496])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The shot begins with a low-angle shot, the camera slowly and steadily panning from left to right, gliding over the floor of a remarkably tranquil room. The camera's focus is on the intricate golden-brown woven straws of a traditional tatami mat, their texture captured in surreal detail. In a surreal and beautiful display, tiny, colorful wildflowers delicately yet relentlessly push through the tightly woven mats. Rather than being randomly scattered, these blooms unfold in a gentle, sprawling pattern, meandering across the floor like a living river of color. We see tiny bell-shaped flowers of deep purple, multi-petaled heads of soft pink, and sunny yellow buds, their slender green stems and delicate leaves slightly replacing the straw fibers as they sprout. The scene is illuminated by soft, warm light filtered through an off-screen shoji paper screen, creating a gentle glow that accentuates the translucent quality of the petals and casts long, soft shadows. The background remains minimalist and out of focus, suggesting a simple, uncluttered space, heightening the focus on these unlikely blossoms. The entire vision creates a powerful atmosphere of quiet magic, tranquility, and the gentle, enduring power of nature within a space of human order.", + "seed": 150, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_0/index_8496_seed_150_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05971591020934284, + "vlm_logprob": -0.05971591020934284, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.05971591020934284, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1593058705329895 + ], + "opposite_logprob_tokens": [ + -2.4093058109283447 + ], + "logprob_reward": -0.1593058705329895, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1593058705329895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2364899069070816, + 0.04424393177032471, + 0.03181423619389534, + 0.036667585372924805, + 0.04984330013394356, + 0.025952594354748726, + 0.11585652083158493, + 0.27752578258514404, + 0.16146613657474518, + 0.020140010863542557 + ], + "credit_map": [ + 0.027228951454162598, + 0.014771082438528538, + 0.03588177263736725, + 0.04283428564667702, + 0.020375220105051994, + 0.02344253659248352, + 0.041422829031944275, + 0.03053322806954384, + 0.0033255896996706724, + 0.0031571390572935343, + 0.00813962146639824, + 0.008031248115003109, + 0.005553825758397579, + 0.004350378643721342, + 0.0076162791810929775, + 0.004069848917424679, + 0.00385899911634624, + 0.00283731147646904, + 0.0031981507781893015, + 0.005135064013302326, + 0.00511968694627285, + 0.003564594080671668, + 0.004494920372962952, + 0.0036055066157132387, + 0.002332604257389903, + 0.002470480278134346, + 0.005547310691326857, + 0.004787092097103596, + 0.005811385344713926, + 0.004459409508854151, + 0.004986066371202469, + 0.006273234263062477, + 0.0037860730662941933, + 0.004183961544185877, + 0.005465186201035976, + 0.007303155958652496, + 0.00479121832177043, + 0.00797939207404852, + 0.009498256258666515, + 0.006836055312305689, + 0.0018422116991132498, + 0.001769607188180089, + 0.0043244254775345325, + 0.004411208909004927, + 0.0029580523259937763, + 0.00323809077963233, + 0.004393232520669699, + 0.00301576335914433, + 0.015473662875592709, + 0.012214647606015205, + 0.01550885010510683, + 0.012566447257995605, + 0.016799157485365868, + 0.01108592189848423, + 0.01833902671933174, + 0.013868793845176697, + 0.009742933325469494, + 0.004386114422231913, + 0.008744842372834682, + 0.026494160294532776, + 0.009148198179900646, + 0.00786623079329729, + 0.12204450368881226, + 0.08909879624843597, + 0.006812415551394224, + 0.009098339825868607, + 0.027200937271118164, + 0.024255525320768356, + 0.007772508542984724, + 0.007617414463311434, + 0.05249164626002312, + 0.02621733397245407, + 0.0016486137174069881, + 0.0027414713986217976, + 0.00391281908378005, + 0.0031604061368852854, + 0.0021593787241727114, + 0.0014550529886037111, + 0.00293918140232563, + 0.0021230867132544518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.1593058705329895 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032746940851212 + ], + "opposite_logprob_tokens": [ + -2.350327491760254 + ], + "logprob_reward": -0.10032746940851212, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032746940851212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1179267168045044, + 0.030865170061588287, + 0.035260412842035294, + 0.026085931807756424, + 0.05504368245601654, + 0.040936313569545746, + 0.1987282931804657, + 0.23180943727493286, + 0.16556532680988312, + 0.09777870774269104 + ], + "credit_map": [ + 0.008743098005652428, + 0.00598051818087697, + 0.017346229404211044, + 0.02782982587814331, + 0.00811530090868473, + 0.008940252475440502, + 0.02430724911391735, + 0.016664253547787666, + 0.0024329759180545807, + 0.0020270429085940123, + 0.0061736819334328175, + 0.005129472352564335, + 0.002754009095951915, + 0.002903193235397339, + 0.006031001452356577, + 0.0034137945622205734, + 0.00313188461586833, + 0.0031972501892596483, + 0.00415166188031435, + 0.0072869546711444855, + 0.0027444742154330015, + 0.003390400903299451, + 0.004911686759442091, + 0.006446104031056166, + 0.001967577962204814, + 0.0020958874374628067, + 0.002883573528379202, + 0.0049425954930484295, + 0.001968629425391555, + 0.003183357883244753, + 0.004542260896414518, + 0.00450205197557807, + 0.0024887260515242815, + 0.003424812573939562, + 0.009078899398446083, + 0.00968024879693985, + 0.0065435790456831455, + 0.004537587054073811, + 0.010736390016973019, + 0.00855344720184803, + 0.0026329488027840853, + 0.0028632262255996466, + 0.008477495983242989, + 0.004903037566691637, + 0.003968068398535252, + 0.0030674126464873552, + 0.008339036256074905, + 0.00668509304523468, + 0.029709920287132263, + 0.023049484938383102, + 0.023224962875247, + 0.017988821491599083, + 0.03393080085515976, + 0.021716106683015823, + 0.02967541292309761, + 0.019432784989476204, + 0.012839188799262047, + 0.004146191757172346, + 0.008948288857936859, + 0.0217538271099329, + 0.007545174565166235, + 0.006480262149125338, + 0.10125990957021713, + 0.06883659958839417, + 0.01259851735085249, + 0.01927756331861019, + 0.022320788353681564, + 0.02686266601085663, + 0.013348960317671299, + 0.015896594151854515, + 0.028784308582544327, + 0.026475947350263596, + 0.006430324167013168, + 0.00891051348298788, + 0.024932539090514183, + 0.016618138179183006, + 0.0049929507076740265, + 0.0065334332175552845, + 0.014278668910264969, + 0.015082146041095257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.10032746940851212 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002826506970450282 + ], + "opposite_logprob_tokens": [ + -5.877826690673828 + ], + "logprob_reward": -0.002826506970450282, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002826506970450282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10670014470815659, + 0.08651906251907349, + 0.05069035664200783, + 0.04298024624586105, + 0.04317846894264221, + 0.03982778638601303, + 0.02054479718208313, + 0.47595471143722534, + 0.09967032074928284, + 0.0339340977370739 + ], + "credit_map": [ + 0.011189513839781284, + 0.008199105970561504, + 0.014378590509295464, + 0.01498821284621954, + 0.011156783439218998, + 0.01571819931268692, + 0.018949953839182854, + 0.012119783088564873, + 0.00615983922034502, + 0.005627189297229052, + 0.012607516720890999, + 0.0199112668633461, + 0.010474803857505322, + 0.009016801603138447, + 0.01475832425057888, + 0.007963319309055805, + 0.0033498320262879133, + 0.004112728405743837, + 0.006624794565141201, + 0.012013168074190617, + 0.004218154586851597, + 0.0047312527894973755, + 0.006674605887383223, + 0.008965818211436272, + 0.0036860830150544643, + 0.002012009499594569, + 0.004704232327640057, + 0.00984240509569645, + 0.0037275832146406174, + 0.0043953522108495235, + 0.007919039577245712, + 0.006693539209663868, + 0.0029267750214785337, + 0.00297357514500618, + 0.005230371840298176, + 0.007251592818647623, + 0.004627270624041557, + 0.004622230306267738, + 0.009413138963282108, + 0.006133515387773514, + 0.00219349330291152, + 0.001993684098124504, + 0.008420437574386597, + 0.008914177305996418, + 0.003278181655332446, + 0.002806128468364477, + 0.007536048069596291, + 0.004685633350163698, + 0.0022461300250142813, + 0.001391286146827042, + 0.0024363251868635416, + 0.0021786275319755077, + 0.0037531410343945026, + 0.002728378167375922, + 0.003753592260181904, + 0.0020573153160512447, + 0.03479839488863945, + 0.010069728828966618, + 0.014250346459448338, + 0.05325651168823242, + 0.014964022673666477, + 0.01059869583696127, + 0.210064098238945, + 0.12795288860797882, + 0.003694879822432995, + 0.003533331211656332, + 0.026315931230783463, + 0.01370159164071083, + 0.0041965008713305, + 0.003590039676055312, + 0.032063715159893036, + 0.012574324384331703, + 0.0025260562542825937, + 0.0037938093300908804, + 0.004913610871881247, + 0.005804151762276888, + 0.003929886035621166, + 0.004354897886514664, + 0.004845057614147663, + 0.0037666300777345896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.002826506970450282 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781171567738056 + ], + "opposite_logprob_tokens": [ + -4.634781360626221 + ], + "logprob_reward": -0.009781171567738056, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781171567738056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21820496022701263, + 0.06263697147369385, + 0.051384929567575455, + 0.07662197947502136, + 0.05103596672415733, + 0.062479034066200256, + 0.030596934258937836, + 0.24983352422714233, + 0.0991377905011177, + 0.09806789457798004 + ], + "credit_map": [ + 0.015167921781539917, + 0.012352542020380497, + 0.029579367488622665, + 0.04940146207809448, + 0.0135365454480052, + 0.01683196611702442, + 0.04851394519209862, + 0.03282121196389198, + 0.005386109463870525, + 0.003842879319563508, + 0.009957930073142052, + 0.015238402411341667, + 0.004770101513713598, + 0.006477898918092251, + 0.010096478275954723, + 0.006867165677249432, + 0.0033783982507884502, + 0.0039359936490654945, + 0.006594408303499222, + 0.013914988376200199, + 0.004248645156621933, + 0.0043046441860497, + 0.005944773089140654, + 0.00906307715922594, + 0.006649272982031107, + 0.004283545073121786, + 0.007896039634943008, + 0.016109265387058258, + 0.00570052582770586, + 0.009115521796047688, + 0.014026034623384476, + 0.012841765768826008, + 0.004006735514849424, + 0.0034157338086515665, + 0.007731466554105282, + 0.006474272347986698, + 0.003641190007328987, + 0.009929431602358818, + 0.007851595059037209, + 0.007985543459653854, + 0.004166254308074713, + 0.0030776553321629763, + 0.006857625674456358, + 0.007303166668862104, + 0.010144863277673721, + 0.005380179267376661, + 0.00971805490553379, + 0.01583123579621315, + 0.003979172091931105, + 0.002968370448797941, + 0.003589808940887451, + 0.002704695798456669, + 0.005935820285230875, + 0.003248807741329074, + 0.004606174770742655, + 0.003564083483070135, + 0.023571571335196495, + 0.006778827868402004, + 0.00822385586798191, + 0.029890410602092743, + 0.009983149357140064, + 0.005709081422537565, + 0.10749291628599167, + 0.058183711022138596, + 0.005788043141365051, + 0.009759870357811451, + 0.0240778885781765, + 0.013147988356649876, + 0.006120903883129358, + 0.0056642708368599415, + 0.021310780197381973, + 0.01326804794371128, + 0.007136810105293989, + 0.009556017816066742, + 0.01350085623562336, + 0.018127435818314552, + 0.0062544215470552444, + 0.008820596151053905, + 0.020685797557234764, + 0.01398596353828907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.009781171567738056 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633853256702423 + ], + "opposite_logprob_tokens": [ + -3.651338577270508 + ], + "logprob_reward": -0.02633853256702423, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633853256702423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2813304662704468, + 0.03596676513552666, + 0.06079910323023796, + 0.10320213437080383, + 0.05681395158171654, + 0.029831841588020325, + 0.0658903568983078, + 0.16898326575756073, + 0.07377997040748596, + 0.12340211123228073 + ], + "credit_map": [ + 0.013158493675291538, + 0.017302624881267548, + 0.04151814058423042, + 0.058865513652563095, + 0.022503899410367012, + 0.018165867775678635, + 0.07018453627824783, + 0.039631400257349014, + 0.0031082474160939455, + 0.0029079325031489134, + 0.0047038099728524685, + 0.006784038618206978, + 0.00342753273434937, + 0.00506036588922143, + 0.0049550579860806465, + 0.00501978350803256, + 0.004520485643297434, + 0.00506497360765934, + 0.008618107996881008, + 0.013044376857578754, + 0.0072883423417806625, + 0.005913248751312494, + 0.006288816221058369, + 0.010060752741992474, + 0.007705517578870058, + 0.00683864951133728, + 0.013229108415544033, + 0.013832847587764263, + 0.01660793274641037, + 0.015473308973014355, + 0.014074787497520447, + 0.015439990907907486, + 0.004072902258485556, + 0.004204833880066872, + 0.004872738383710384, + 0.004995109047740698, + 0.008166524581611156, + 0.01326259970664978, + 0.007448834832757711, + 0.009790414944291115, + 0.0026826104149222374, + 0.001964029623195529, + 0.0027585942298173904, + 0.0034450748935341835, + 0.004987107124179602, + 0.004253566265106201, + 0.00414178567007184, + 0.005599075462669134, + 0.00904273334890604, + 0.006727295462042093, + 0.007078779861330986, + 0.005780272651463747, + 0.013145704753696918, + 0.008654101751744747, + 0.009144874289631844, + 0.0063166008330881596, + 0.006898059044033289, + 0.004940504673868418, + 0.004664637614041567, + 0.013858666643500328, + 0.010121896862983704, + 0.005671690683811903, + 0.06727217882871628, + 0.05555564910173416, + 0.003562329337000847, + 0.005509309004992247, + 0.012804572470486164, + 0.00645662285387516, + 0.003745605470612645, + 0.003426317125558853, + 0.024670729413628578, + 0.013604491017758846, + 0.007482930552214384, + 0.008562441915273666, + 0.0117730051279068, + 0.029012173414230347, + 0.00875519122928381, + 0.010289023630321026, + 0.03410687297582626, + 0.013420487754046917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.02633853256702423 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with a low-angle shot, the camera slowly and steadily panning from left to right, gliding over the floor of a remarkably tranquil room. The camera's focus is on the intricate golden-brown woven straws of a traditional tatami mat, their texture captured in surreal detail. In a surreal and beautiful display, tiny, colorful wildflowers delicately yet relentlessly push through the tightly woven mats. Rather than being randomly scattered, these blooms unfold in a gentle, sprawling pattern, meandering across the floor like a living river of color. We see tiny bell-shaped flowers of deep purple, multi-petaled heads of soft pink, and sunny yellow buds, their slender green stems and delicate leaves slightly replacing the straw fibers as they sprout. The scene is illuminated by soft, warm light filtered through an off-screen shoji paper screen, creating a gentle glow that accentuates the translucent quality of the petals and casts long, soft shadows. The background remains minimalist and out of focus, suggesting a simple, uncluttered space, heightening the focus on these unlikely blossoms. The entire vision creates a powerful atmosphere of quiet magic, tranquility, and the gentle, enduring power of nature within a space of human order.", + "seed": 151, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_0/index_8496_seed_151_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08485485576093196, + "vlm_logprob": -0.08485485576093196, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.08485485576093196, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17235928773880005 + ], + "opposite_logprob_tokens": [ + -2.1723592281341553 + ], + "logprob_reward": -0.17235928773880005, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.17235928773880005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24954093992710114, + 0.09507628530263901, + 0.03361697122454643, + 0.06679485738277435, + 0.0314541719853878, + 0.058850232511758804, + 0.025956375524401665, + 0.0725773274898529, + 0.3131754696369171, + 0.05295734852552414 + ], + "credit_map": [ + 0.014954917132854462, + 0.01569742150604725, + 0.05606218799948692, + 0.04453015699982643, + 0.01573771797120571, + 0.01622026413679123, + 0.04677790775895119, + 0.039560385048389435, + 0.008934061042964458, + 0.020568620413541794, + 0.013873355463147163, + 0.014804196543991566, + 0.006161229219287634, + 0.011710134334862232, + 0.009901652112603188, + 0.00912304688245058, + 0.002682038815692067, + 0.0021109087392687798, + 0.006746607832610607, + 0.007300719153136015, + 0.0018848023610189557, + 0.002878340194001794, + 0.004914748482406139, + 0.00509880855679512, + 0.00446267519146204, + 0.0038655181415379047, + 0.014023366384208202, + 0.012091496959328651, + 0.006478658877313137, + 0.0032062504906207323, + 0.012173007242381573, + 0.010493886657059193, + 0.0024097186978906393, + 0.0017994926311075687, + 0.0059685371816158295, + 0.004847906995564699, + 0.002905899193137884, + 0.001756800338625908, + 0.006913732271641493, + 0.004852086305618286, + 0.002995455637574196, + 0.0023125973530113697, + 0.012552834115922451, + 0.011346218176186085, + 0.0037815796677023172, + 0.0024595342110842466, + 0.011449071578681469, + 0.011952943168580532, + 0.0021329750306904316, + 0.0017910742899402976, + 0.003903194796293974, + 0.003579766722396016, + 0.0017815076280385256, + 0.0036775104235857725, + 0.004497605375945568, + 0.004592742305248976, + 0.008117527700960636, + 0.004444134887307882, + 0.011232667602598667, + 0.015388712286949158, + 0.006246177479624748, + 0.0038638897240161896, + 0.010890641249716282, + 0.012393581680953503, + 0.02409842424094677, + 0.012034333311021328, + 0.04603250324726105, + 0.06831365078687668, + 0.020733918994665146, + 0.01671670936048031, + 0.06075343117117882, + 0.06449250876903534, + 0.003131929785013199, + 0.007015562616288662, + 0.007477517705410719, + 0.010648935101926327, + 0.004639068618416786, + 0.004446185193955898, + 0.00803919043391943, + 0.007558960467576981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.17235928773880005 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159050822257996 + ], + "opposite_logprob_tokens": [ + -1.7015905380249023 + ], + "logprob_reward": -0.20159050822257996, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159050822257996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12484003603458405, + 0.06142880395054817, + 0.1281629353761673, + 0.20809613168239594, + 0.07631681859493256, + 0.045675136148929596, + 0.07128551602363586, + 0.061187125742435455, + 0.1581445187330246, + 0.06486297398805618 + ], + "credit_map": [ + 0.0111039774492383, + 0.011426601558923721, + 0.021772829815745354, + 0.018737027421593666, + 0.012412559241056442, + 0.008992183953523636, + 0.0225252453237772, + 0.017869621515274048, + 0.005320060066878796, + 0.007359358016401529, + 0.012811235152184963, + 0.008906588889658451, + 0.006125129759311676, + 0.005992593243718147, + 0.007789337541908026, + 0.007124497555196285, + 0.005206122063100338, + 0.006260802503675222, + 0.03455733135342598, + 0.014572663232684135, + 0.005307050887495279, + 0.005586408078670502, + 0.044629719108343124, + 0.012042833492159843, + 0.013639949262142181, + 0.008918564766645432, + 0.03864605724811554, + 0.036475520581007004, + 0.01338057778775692, + 0.010361000895500183, + 0.04721985384821892, + 0.03945459425449371, + 0.005741292145103216, + 0.003385293297469616, + 0.008288659155368805, + 0.0201732087880373, + 0.004292071796953678, + 0.004192939028143883, + 0.00898020900785923, + 0.02126314304769039, + 0.005913539789617062, + 0.0023805410601198673, + 0.007246998604387045, + 0.007852120324969292, + 0.004199064336717129, + 0.0028641102835536003, + 0.009190472774207592, + 0.006028290838003159, + 0.005627569742500782, + 0.0033872155472636223, + 0.01012794952839613, + 0.011890840716660023, + 0.0056885965168476105, + 0.006155238952487707, + 0.014154392294585705, + 0.014253716915845871, + 0.004101429134607315, + 0.0030976678244769573, + 0.009482596069574356, + 0.011777197942137718, + 0.004585068207234144, + 0.004269770346581936, + 0.012413251213729382, + 0.011460140347480774, + 0.009840653277933598, + 0.007592394016683102, + 0.022966094315052032, + 0.0469033382833004, + 0.009220967069268227, + 0.011081648990511894, + 0.02353244088590145, + 0.027006976306438446, + 0.004084063693881035, + 0.00801385659724474, + 0.006436228286474943, + 0.01246759481728077, + 0.005343568045645952, + 0.0046362015418708324, + 0.010963858105242252, + 0.012917605228722095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.20159050822257996 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635258063673973 + ], + "opposite_logprob_tokens": [ + -3.651352643966675 + ], + "logprob_reward": -0.02635258063673973, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635258063673973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2656295597553253, + 0.07069656252861023, + 0.1102932021021843, + 0.14251276850700378, + 0.08140335977077484, + 0.01919739320874214, + 0.05935075879096985, + 0.05347822234034538, + 0.15060371160507202, + 0.046834446489810944 + ], + "credit_map": [ + 0.0162538830190897, + 0.013623582199215889, + 0.03819306939840317, + 0.04792620986700058, + 0.029743468388915062, + 0.02417439967393875, + 0.03487459942698479, + 0.060840342193841934, + 0.0054380870424211025, + 0.007510636001825333, + 0.014011026360094547, + 0.013138576410710812, + 0.0036258017644286156, + 0.003976917825639248, + 0.010181593708693981, + 0.012813922017812729, + 0.006465384270995855, + 0.009791489690542221, + 0.019846495240926743, + 0.01763540506362915, + 0.009524165652692318, + 0.005649586208164692, + 0.02147475816309452, + 0.019905921071767807, + 0.008238260634243488, + 0.006150382570922375, + 0.0261622816324234, + 0.021079296246170998, + 0.01110917516052723, + 0.006642487831413746, + 0.03720391169190407, + 0.02592698484659195, + 0.00473807891830802, + 0.005082902032881975, + 0.01384932640939951, + 0.014534931629896164, + 0.009321038611233234, + 0.005916336551308632, + 0.014046892523765564, + 0.013913854956626892, + 0.0012156900484114885, + 0.0009006425971165299, + 0.0035478617064654827, + 0.00361058721318841, + 0.001986369024962187, + 0.001125060603953898, + 0.003909438848495483, + 0.0029017431661486626, + 0.003889766288921237, + 0.0044980659149587154, + 0.01300759706646204, + 0.010282924398779869, + 0.0035447273403406143, + 0.00671191094443202, + 0.007825353182852268, + 0.009590410627424717, + 0.004303112626075745, + 0.006239764392375946, + 0.006438677664846182, + 0.009744217619299889, + 0.005953919608145952, + 0.004534413106739521, + 0.008368326351046562, + 0.007895788177847862, + 0.008204434998333454, + 0.020380575209856033, + 0.015276112593710423, + 0.0475282296538353, + 0.00848342478275299, + 0.007891615852713585, + 0.021644417196512222, + 0.021194903180003166, + 0.0034129272680729628, + 0.007918966002762318, + 0.0055536190047860146, + 0.007396957837045193, + 0.00339144398458302, + 0.0027011402416974306, + 0.008013812825083733, + 0.008445581421256065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.02635258063673973 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784241206943989 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784241206943989, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784241206943989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2589173913002014, + 0.07847528159618378, + 0.06015726923942566, + 0.05730145797133446, + 0.020737236365675926, + 0.038177620619535446, + 0.04895869642496109, + 0.13341033458709717, + 0.15462824702262878, + 0.14923644065856934 + ], + "credit_map": [ + 0.02095017023384571, + 0.02242172695696354, + 0.04223654419183731, + 0.04747210443019867, + 0.02460760436952114, + 0.019566388800740242, + 0.03621197119355202, + 0.04545092210173607, + 0.007685832213610411, + 0.018202442675828934, + 0.010126210749149323, + 0.01364767923951149, + 0.0071008652448654175, + 0.00722121074795723, + 0.00819330383092165, + 0.0062977406196296215, + 0.0034516232553869486, + 0.003758140141144395, + 0.01763933151960373, + 0.009583843871951103, + 0.004146439954638481, + 0.003460750449448824, + 0.011176347732543945, + 0.006940794177353382, + 0.0039269146509468555, + 0.0030176560394465923, + 0.01372743584215641, + 0.01024298369884491, + 0.004468035884201527, + 0.003239632584154606, + 0.010242249816656113, + 0.008436550386250019, + 0.0025609461590647697, + 0.0015179445035755634, + 0.0025893915444612503, + 0.002730904147028923, + 0.0015016898978501558, + 0.0026586418971419334, + 0.003235495649278164, + 0.003942225128412247, + 0.002968507120385766, + 0.0024236927274614573, + 0.005889554042369127, + 0.006398722529411316, + 0.0032450782600790262, + 0.0022109418641775846, + 0.008432477712631226, + 0.006608649156987667, + 0.002678715391084552, + 0.005354013293981552, + 0.007066978607326746, + 0.006961446721106768, + 0.0029058828949928284, + 0.004810463637113571, + 0.011138590052723885, + 0.008042609319090843, + 0.011285340413451195, + 0.01123780570924282, + 0.018865332007408142, + 0.027769865468144417, + 0.011152727529406548, + 0.010128691792488098, + 0.017671717330813408, + 0.025298863649368286, + 0.008901761844754219, + 0.010456124320626259, + 0.01988108642399311, + 0.04114425182342529, + 0.01225812453776598, + 0.010814153589308262, + 0.025802655145525932, + 0.025370098650455475, + 0.01078141387552023, + 0.011633225716650486, + 0.021525755524635315, + 0.0177758801728487, + 0.031001172959804535, + 0.014006041921675205, + 0.021420126780867577, + 0.02109282836318016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.009784241206943989 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014187660999596119 + ], + "opposite_logprob_tokens": [ + -4.264187812805176 + ], + "logprob_reward": -0.014187660999596119, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014187660999596119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2219482958316803, + 0.06564754247665405, + 0.10665112733840942, + 0.05174964666366577, + 0.044040970504283905, + 0.05181049555540085, + 0.03748985752463341, + 0.1723160445690155, + 0.20436665415763855, + 0.043979354202747345 + ], + "credit_map": [ + 0.022798065096139908, + 0.02649773098528385, + 0.031298402696847916, + 0.029653461650013924, + 0.03522130846977234, + 0.02332964725792408, + 0.029255444183945656, + 0.02389424480497837, + 0.006852960214018822, + 0.017265716567635536, + 0.007864025421440601, + 0.009178552776575089, + 0.004901102744042873, + 0.006353321019560099, + 0.006996185518801212, + 0.0062356856651604176, + 0.005343096796423197, + 0.008658499456942081, + 0.01801711693406105, + 0.022022919729351997, + 0.01428836677223444, + 0.007907182909548283, + 0.01515079103410244, + 0.015263166278600693, + 0.0036757830530405045, + 0.003227364271879196, + 0.00760224275290966, + 0.014325332827866077, + 0.0038314128760248423, + 0.0026612491346895695, + 0.009752659127116203, + 0.006673604249954224, + 0.0037439882289618254, + 0.0034231171011924744, + 0.004549821373075247, + 0.010352900251746178, + 0.0028233276680111885, + 0.0025346442125737667, + 0.005675841588526964, + 0.01093733124434948, + 0.009130851365625858, + 0.003220068523660302, + 0.009527086280286312, + 0.0064864009618759155, + 0.00363427447155118, + 0.003631263505667448, + 0.009637221693992615, + 0.006543334107846022, + 0.003269354347139597, + 0.006747494451701641, + 0.0040143933147192, + 0.004501917399466038, + 0.0032617768738418818, + 0.00622218195348978, + 0.004098271951079369, + 0.005374467466026545, + 0.016238393262028694, + 0.007732408121228218, + 0.027174554765224457, + 0.03840698301792145, + 0.010531102307140827, + 0.007854022085666656, + 0.028025681152939796, + 0.036352913826704025, + 0.010925858281552792, + 0.022365650162100792, + 0.027565306052565575, + 0.05988851562142372, + 0.008711795322597027, + 0.011154942214488983, + 0.030903073027729988, + 0.032851532101631165, + 0.0019801335874944925, + 0.0024967750068753958, + 0.005824695341289043, + 0.009199786931276321, + 0.004291648045182228, + 0.004126986023038626, + 0.007348064798861742, + 0.008711264468729496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.014187660999596119 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with a low-angle shot, the camera slowly and steadily panning from left to right, gliding over the floor of a remarkably tranquil room. The camera's focus is on the intricate golden-brown woven straws of a traditional tatami mat, their texture captured in surreal detail. In a surreal and beautiful display, tiny, colorful wildflowers delicately yet relentlessly push through the tightly woven mats. Rather than being randomly scattered, these blooms unfold in a gentle, sprawling pattern, meandering across the floor like a living river of color. We see tiny bell-shaped flowers of deep purple, multi-petaled heads of soft pink, and sunny yellow buds, their slender green stems and delicate leaves slightly replacing the straw fibers as they sprout. The scene is illuminated by soft, warm light filtered through an off-screen shoji paper screen, creating a gentle glow that accentuates the translucent quality of the petals and casts long, soft shadows. The background remains minimalist and out of focus, suggesting a simple, uncluttered space, heightening the focus on these unlikely blossoms. The entire vision creates a powerful atmosphere of quiet magic, tranquility, and the gentle, enduring power of nature within a space of human order.", + "seed": 152, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_0/index_8496_seed_152_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14355265833437442, + "vlm_logprob": -0.14355265833437442, + "vlm_yes_no_margin": 2.4499999642372132, + "vlm_token_credit": -0.14355265833437442, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3668997883796692 + ], + "opposite_logprob_tokens": [ + -1.3668997287750244 + ], + "logprob_reward": -0.3668997883796692, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.3668997883796692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17949187755584717, + 0.06253714859485626, + 0.06630486994981766, + 0.08565791696310043, + 0.014711626805365086, + 0.037580959498882294, + 0.03089875541627407, + 0.09601224958896637, + 0.1636144369840622, + 0.2631900906562805 + ], + "credit_map": [ + 0.011249945499002934, + 0.014888866804540157, + 0.030424924567341805, + 0.025154603645205498, + 0.015160366892814636, + 0.01821143552660942, + 0.03357913717627525, + 0.030822616070508957, + 0.0032496091444045305, + 0.00667102076113224, + 0.008637335151433945, + 0.009568304754793644, + 0.004905520472675562, + 0.006638381630182266, + 0.011485785245895386, + 0.011381193995475769, + 0.0039399489760398865, + 0.004902931861579418, + 0.011056934483349323, + 0.011834867298603058, + 0.003117242595180869, + 0.005540627054870129, + 0.012522906064987183, + 0.013389415107667446, + 0.005820447579026222, + 0.0077796513214707375, + 0.012371230870485306, + 0.01114506833255291, + 0.010777745395898819, + 0.01229818630963564, + 0.009523794054985046, + 0.0159417986869812, + 0.0013288151239976287, + 0.0012794096255674958, + 0.002469151047989726, + 0.002374384319409728, + 0.0013740778667852283, + 0.0015368727035820484, + 0.0021722165402024984, + 0.0021767013240605593, + 0.0030935942195355892, + 0.0037156096659600735, + 0.0049556707963347435, + 0.004408321343362331, + 0.00396200455725193, + 0.004652991425246, + 0.006958417594432831, + 0.005834350362420082, + 0.0022711153142154217, + 0.00185545498970896, + 0.006145006977021694, + 0.004259628243744373, + 0.0030079176649451256, + 0.0030426608864217997, + 0.005610555876046419, + 0.004706420004367828, + 0.007884562015533447, + 0.003695543622598052, + 0.012191294692456722, + 0.017324984073638916, + 0.005460800137370825, + 0.006123446859419346, + 0.011726472526788712, + 0.031605158001184464, + 0.005592780653387308, + 0.007597554940730333, + 0.05811549723148346, + 0.05900293216109276, + 0.005108620971441269, + 0.004867049399763346, + 0.01116744801402092, + 0.012162575498223305, + 0.010459249839186668, + 0.01715158298611641, + 0.030160456895828247, + 0.059862565249204636, + 0.010172256268560886, + 0.015402758494019508, + 0.043166689574718475, + 0.07681456953287125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.3668997883796692 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17993396520614624 + ], + "opposite_logprob_tokens": [ + -1.804934024810791 + ], + "logprob_reward": -0.17993396520614624, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17993396520614624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046287816017866135, + 0.09902025759220123, + 0.02797117456793785, + 0.05975912883877754, + 0.022076833993196487, + 0.0172134879976511, + 0.054800745099782944, + 0.026669859886169434, + 0.17654363811016083, + 0.4696570932865143 + ], + "credit_map": [ + 0.005469692870974541, + 0.003653160994872451, + 0.007465943694114685, + 0.005389743484556675, + 0.004993945825845003, + 0.004300788976252079, + 0.008076624944806099, + 0.006937918718904257, + 0.004937155172228813, + 0.010089851915836334, + 0.014419169165194035, + 0.015797654166817665, + 0.006475408561527729, + 0.011745205149054527, + 0.01726267673075199, + 0.018293138593435287, + 0.0014976144302636385, + 0.003540931036695838, + 0.004338917322456837, + 0.003831519279628992, + 0.002059605438262224, + 0.002649095142260194, + 0.004718760494142771, + 0.0053347316570580006, + 0.002901997184380889, + 0.0040930043905973434, + 0.010517152026295662, + 0.009479663334786892, + 0.004653036594390869, + 0.0063882628455758095, + 0.008425134234130383, + 0.0133008798584342, + 0.0022544676903635263, + 0.0020924988202750683, + 0.0031869783997535706, + 0.0036013226490467787, + 0.002580813132226467, + 0.0023808516561985016, + 0.003497803583741188, + 0.0024820975959300995, + 0.0017076224321499467, + 0.0013226018054410815, + 0.002374173142015934, + 0.002755299210548401, + 0.0014376291073858738, + 0.0018064038595184684, + 0.003047092817723751, + 0.00276266410946846, + 0.0022787160705775023, + 0.0022582218516618013, + 0.008706250227987766, + 0.007012049201875925, + 0.003906401805579662, + 0.003302208846434951, + 0.012252869084477425, + 0.015084030106663704, + 0.0023562503047287464, + 0.0016910925041884184, + 0.004875142592936754, + 0.003565147053450346, + 0.001963619142770767, + 0.002789850579574704, + 0.002873319434002042, + 0.006555438973009586, + 0.007303133141249418, + 0.011432966217398643, + 0.06453727930784225, + 0.06219519302248955, + 0.003938144072890282, + 0.004219694063067436, + 0.01275577861815691, + 0.010161439888179302, + 0.015889829024672508, + 0.028542106971144676, + 0.05360176041722298, + 0.12174780666828156, + 0.019820237532258034, + 0.028286950662732124, + 0.07086477428674698, + 0.13090363144874573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.17993396520614624 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705929577350616 + ], + "opposite_logprob_tokens": [ + -2.1270592212677 + ], + "logprob_reward": -0.12705929577350616, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12705929577350616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11283780634403229, + 0.1060575619339943, + 0.04710202291607857, + 0.031027836725115776, + 0.025183796882629395, + 0.039384953677654266, + 0.04255048185586929, + 0.23802721500396729, + 0.18461638689041138, + 0.17321190237998962 + ], + "credit_map": [ + 0.008765855804085732, + 0.009181699715554714, + 0.02117174305021763, + 0.019117364659905434, + 0.008931008167564869, + 0.007587597239762545, + 0.020492374897003174, + 0.01759016513824463, + 0.006003528367727995, + 0.009559785015881062, + 0.016314538195729256, + 0.024596234783530235, + 0.00703792879357934, + 0.008210730738937855, + 0.019385671243071556, + 0.014949153177440166, + 0.0037459763698279858, + 0.004764602053910494, + 0.007964922115206718, + 0.0066686468198895454, + 0.002738233422860503, + 0.004094691015779972, + 0.009537465870380402, + 0.007587489206343889, + 0.0021984013728797436, + 0.0033272618893533945, + 0.0042284713126719, + 0.0043882206082344055, + 0.004315739031881094, + 0.003294829512014985, + 0.003572771791368723, + 0.0057021440006792545, + 0.001310095191001892, + 0.0015803694259375334, + 0.004110666923224926, + 0.003923011012375355, + 0.0024313817266374826, + 0.0017078982200473547, + 0.004851571749895811, + 0.005268803332000971, + 0.00550731411203742, + 0.0037033045664429665, + 0.004890031646937132, + 0.005273963790386915, + 0.0038099277298897505, + 0.0035590617917478085, + 0.005342758726328611, + 0.007298596668988466, + 0.002771573606878519, + 0.003746742382645607, + 0.007200729567557573, + 0.005345011595636606, + 0.004898637533187866, + 0.0030703661032021046, + 0.007180436514317989, + 0.008336987346410751, + 0.020072167739272118, + 0.011889437213540077, + 0.024382002651691437, + 0.037518780678510666, + 0.014356137253344059, + 0.013721809722483158, + 0.027739757671952248, + 0.08834713697433472, + 0.008502411656081676, + 0.02016351744532585, + 0.038183264434337616, + 0.04565955325961113, + 0.007055656053125858, + 0.006745781283825636, + 0.01651042141020298, + 0.041795793920755386, + 0.007174385711550713, + 0.014531743712723255, + 0.022132139652967453, + 0.039164889603853226, + 0.009333721362054348, + 0.00898693222552538, + 0.022818880155682564, + 0.04906922206282616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.12705929577350616 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329576015472412 + ], + "opposite_logprob_tokens": [ + -3.7732958793640137 + ], + "logprob_reward": -0.02329576015472412, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329576015472412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10241927951574326, + 0.09539271146059036, + 0.05718306824564934, + 0.1987362653017044, + 0.015429949387907982, + 0.0374305285513401, + 0.11565130203962326, + 0.06456780433654785, + 0.08380724489688873, + 0.22938185930252075 + ], + "credit_map": [ + 0.007475038059055805, + 0.007723500486463308, + 0.0175620187073946, + 0.013242614455521107, + 0.010075097903609276, + 0.011730410158634186, + 0.01712031476199627, + 0.017490282654762268, + 0.0046752202324569225, + 0.011346816085278988, + 0.014952459372580051, + 0.026693707332015038, + 0.004450797103345394, + 0.006414420437067747, + 0.01238949503749609, + 0.014469795860350132, + 0.0031394283287227154, + 0.003922818694263697, + 0.010091658681631088, + 0.010659375227987766, + 0.002800960326567292, + 0.005033576861023903, + 0.010888174176216125, + 0.010647077113389969, + 0.012434511445462704, + 0.012970171868801117, + 0.03134569898247719, + 0.026679782196879387, + 0.017369380220770836, + 0.0260673388838768, + 0.023649411275982857, + 0.048219963908195496, + 0.0013486557872965932, + 0.001062445342540741, + 0.002520211972296238, + 0.0021292136516422033, + 0.001282924204133451, + 0.0010362094035372138, + 0.0025920779444277287, + 0.00345821026712656, + 0.0044152759946882725, + 0.00442682858556509, + 0.005610190331935883, + 0.00616380013525486, + 0.003505944972857833, + 0.0036820718087255955, + 0.005651083774864674, + 0.003975332714617252, + 0.006015420891344547, + 0.005902784410864115, + 0.01573244109749794, + 0.01632026955485344, + 0.006313340738415718, + 0.007115013897418976, + 0.022361483424901962, + 0.035890549421310425, + 0.004282331094145775, + 0.0024527304340153933, + 0.007950501516461372, + 0.011787989176809788, + 0.0034087386447936296, + 0.00453942222520709, + 0.008403167128562927, + 0.02174292877316475, + 0.004753713961690664, + 0.010990317910909653, + 0.018337629735469818, + 0.02056116610765457, + 0.004439637530595064, + 0.003975884523242712, + 0.007957980036735535, + 0.012790912762284279, + 0.009260415099561214, + 0.01577441394329071, + 0.025501804426312447, + 0.049609024077653885, + 0.01015782356262207, + 0.014663951471447945, + 0.03682229667901993, + 0.06759212166070938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.02329576015472412 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020574482157826424 + ], + "opposite_logprob_tokens": [ + -3.8955745697021484 + ], + "logprob_reward": -0.020574482157826424, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020574482157826424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15958936512470245, + 0.08372440189123154, + 0.06624508649110794, + 0.08220925182104111, + 0.014674014411866665, + 0.15446096658706665, + 0.04298148304224014, + 0.101382315158844, + 0.15115360915660858, + 0.14357952773571014 + ], + "credit_map": [ + 0.0173338633030653, + 0.014337114989757538, + 0.022737011313438416, + 0.01923244260251522, + 0.016826512292027473, + 0.018766820430755615, + 0.023831887170672417, + 0.026523703709244728, + 0.0049485620111227036, + 0.011064580641686916, + 0.013281576335430145, + 0.019342944025993347, + 0.004381848499178886, + 0.008396042510867119, + 0.010964819230139256, + 0.01134402584284544, + 0.00414218008518219, + 0.006077584810554981, + 0.010700195096433163, + 0.015470923855900764, + 0.0040911715477705, + 0.0048614623956382275, + 0.008216166868805885, + 0.01268540509045124, + 0.007463810034096241, + 0.006371228024363518, + 0.012941939756274223, + 0.010573028586804867, + 0.006172075401991606, + 0.00878995843231678, + 0.013040352612733841, + 0.016856851056218147, + 0.0009403005824424326, + 0.0010453116847202182, + 0.0024111901875585318, + 0.002229108242318034, + 0.0017256455030292273, + 0.001362943323329091, + 0.0024157718289643526, + 0.0025437436997890472, + 0.01881483756005764, + 0.008329348638653755, + 0.021253911778330803, + 0.025328805670142174, + 0.012899973429739475, + 0.015373144298791885, + 0.023739594966173172, + 0.028721362352371216, + 0.0029427949339151382, + 0.0034173233434557915, + 0.007625063415616751, + 0.005494629964232445, + 0.004244577139616013, + 0.003111194586381316, + 0.008179878816008568, + 0.007966022938489914, + 0.007052286062389612, + 0.004287959076464176, + 0.014456866309046745, + 0.018470073118805885, + 0.0037611350417137146, + 0.007387853227555752, + 0.011814702302217484, + 0.03415144234895706, + 0.005565676838159561, + 0.010993130505084991, + 0.041870035231113434, + 0.04436632990837097, + 0.006531362887471914, + 0.00732404226437211, + 0.012737205252051353, + 0.021765828132629395, + 0.006587660871446133, + 0.012847552075982094, + 0.01703447476029396, + 0.02880459651350975, + 0.00591528695076704, + 0.009395592845976353, + 0.022853203117847443, + 0.04014115408062935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.020574482157826424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with a low-angle shot, the camera slowly and steadily panning from left to right, gliding over the floor of a remarkably tranquil room. The camera's focus is on the intricate golden-brown woven straws of a traditional tatami mat, their texture captured in surreal detail. In a surreal and beautiful display, tiny, colorful wildflowers delicately yet relentlessly push through the tightly woven mats. Rather than being randomly scattered, these blooms unfold in a gentle, sprawling pattern, meandering across the floor like a living river of color. We see tiny bell-shaped flowers of deep purple, multi-petaled heads of soft pink, and sunny yellow buds, their slender green stems and delicate leaves slightly replacing the straw fibers as they sprout. The scene is illuminated by soft, warm light filtered through an off-screen shoji paper screen, creating a gentle glow that accentuates the translucent quality of the petals and casts long, soft shadows. The background remains minimalist and out of focus, suggesting a simple, uncluttered space, heightening the focus on these unlikely blossoms. The entire vision creates a powerful atmosphere of quiet magic, tranquility, and the gentle, enduring power of nature within a space of human order.", + "seed": 153, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_0/index_8496_seed_153_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06884462106972933, + "vlm_logprob": -0.06884462106972933, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.06884462106972933, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17155466973781586 + ], + "opposite_logprob_tokens": [ + -2.2965545654296875 + ], + "logprob_reward": -0.17155466973781586, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.17155466973781586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24348165094852448, + 0.02247515693306923, + 0.0628294050693512, + 0.16261741518974304, + 0.07660994678735733, + 0.07620933651924133, + 0.16447880864143372, + 0.10299064964056015, + 0.04044923558831215, + 0.04785839095711708 + ], + "credit_map": [ + 0.025905489921569824, + 0.014716566540300846, + 0.03897484391927719, + 0.03895833343267441, + 0.017086733132600784, + 0.01494322344660759, + 0.05373159050941467, + 0.039164889603853226, + 0.0019478548783808947, + 0.001800214871764183, + 0.003076929831877351, + 0.0033609310630708933, + 0.002977220807224512, + 0.0016942047514021397, + 0.005377532914280891, + 0.0022402717731893063, + 0.006400415673851967, + 0.0035936504136770964, + 0.009520243853330612, + 0.012758278287947178, + 0.007117478642612696, + 0.0055006202310323715, + 0.007519001141190529, + 0.010419721715152264, + 0.009220524691045284, + 0.010076438076794147, + 0.0214675385504961, + 0.028999637812376022, + 0.027816176414489746, + 0.01889137364923954, + 0.024040713906288147, + 0.02210502326488495, + 0.0035471345763653517, + 0.006167541723698378, + 0.012441582977771759, + 0.020543290302157402, + 0.008190871216356754, + 0.004423083737492561, + 0.011454504914581776, + 0.00984194502234459, + 0.007455229293555021, + 0.009857765398919582, + 0.009902313351631165, + 0.010434286668896675, + 0.009194137528538704, + 0.0075692986138165, + 0.01034893374890089, + 0.011447376571595669, + 0.012060721404850483, + 0.014583316631615162, + 0.03300320729613304, + 0.037226561456918716, + 0.010285702534019947, + 0.014807797968387604, + 0.026419205591082573, + 0.01609230786561966, + 0.009105956181883812, + 0.012308930978178978, + 0.020888833329081535, + 0.01390967145562172, + 0.009284107945859432, + 0.010126052424311638, + 0.01580231450498104, + 0.011564790271222591, + 0.004854358267039061, + 0.004219320602715015, + 0.005760363303124905, + 0.004088183399289846, + 0.008778220042586327, + 0.004744403995573521, + 0.004953557159751654, + 0.0030508304480463266, + 0.005062587093561888, + 0.004552865866571665, + 0.006640955805778503, + 0.00629206420853734, + 0.006806078366935253, + 0.006731298752129078, + 0.0070102582685649395, + 0.00476228678599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.17155466973781586 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706327438354492 + ], + "opposite_logprob_tokens": [ + -2.127063274383545 + ], + "logprob_reward": -0.12706327438354492, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706327438354492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08653675764799118, + 0.048125848174095154, + 0.052996471524238586, + 0.1588410586118698, + 0.09371256828308105, + 0.09392208606004715, + 0.12644246220588684, + 0.13651280105113983, + 0.14455506205558777, + 0.058354973793029785 + ], + "credit_map": [ + 0.005569391418248415, + 0.006733777467161417, + 0.015019803307950497, + 0.01649007946252823, + 0.012994659133255482, + 0.0061432416550815105, + 0.013496953062713146, + 0.010088849812746048, + 0.0038039982318878174, + 0.004274161532521248, + 0.00576039357110858, + 0.0071288649924099445, + 0.005134078674018383, + 0.007956705056130886, + 0.007357749156653881, + 0.00670989602804184, + 0.003947822842746973, + 0.0035775592550635338, + 0.00666682468727231, + 0.009308242239058018, + 0.009194813668727875, + 0.00315518188290298, + 0.006745875813066959, + 0.010400149039924145, + 0.008826334960758686, + 0.010169423185288906, + 0.02238539606332779, + 0.030605383217334747, + 0.027053700760006905, + 0.013996679335832596, + 0.024585919454693794, + 0.021218225359916687, + 0.005588355474174023, + 0.006712427828460932, + 0.013921027071774006, + 0.020116858184337616, + 0.009049679152667522, + 0.011732903309166431, + 0.01374818291515112, + 0.012843135744333267, + 0.006015750579535961, + 0.012511271983385086, + 0.0117545947432518, + 0.018661808222532272, + 0.014011155813932419, + 0.011341609060764313, + 0.011062057688832283, + 0.008563842624425888, + 0.00637404527515173, + 0.019370300695300102, + 0.019153131172060966, + 0.02117600291967392, + 0.011426585726439953, + 0.015053984709084034, + 0.025916142389178276, + 0.00797225534915924, + 0.00937711913138628, + 0.01654059812426567, + 0.022836964577436447, + 0.017732243984937668, + 0.010031135752797127, + 0.015001310966908932, + 0.02587062679231167, + 0.01912280172109604, + 0.0180914718657732, + 0.0178424920886755, + 0.016213946044445038, + 0.01399311888962984, + 0.023919852450489998, + 0.018345165997743607, + 0.02135409787297249, + 0.014794915914535522, + 0.0057913693599402905, + 0.00619119219481945, + 0.010482745245099068, + 0.012596013024449348, + 0.007232521194964647, + 0.005361420568078756, + 0.00584569713100791, + 0.004854016937315464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.12706327438354492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545948848128319 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.012545948848128319, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545948848128319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11198467016220093, + 0.056466516107320786, + 0.14051738381385803, + 0.22243331372737885, + 0.0374605692923069, + 0.1785491406917572, + 0.12103667855262756, + 0.04392554610967636, + 0.06643275171518326, + 0.021193459630012512 + ], + "credit_map": [ + 0.007075718604028225, + 0.007981977425515652, + 0.019375041127204895, + 0.022964738309383392, + 0.012424299493432045, + 0.0070102340541779995, + 0.022721242159605026, + 0.012431420385837555, + 0.00564690912142396, + 0.0033961478620767593, + 0.009428524412214756, + 0.009931277483701706, + 0.006518998183310032, + 0.002228465862572193, + 0.013444000855088234, + 0.005872196517884731, + 0.009249143302440643, + 0.008387437090277672, + 0.022662652656435966, + 0.02689911425113678, + 0.021935435011982918, + 0.009258667938411236, + 0.016388216987252235, + 0.025736724957823753, + 0.011846409179270267, + 0.013448181562125683, + 0.032260045409202576, + 0.04927867278456688, + 0.02797195129096508, + 0.021801084280014038, + 0.03420707583427429, + 0.03161991387605667, + 0.002179803093895316, + 0.002938843797892332, + 0.006983157247304916, + 0.006978692952543497, + 0.002841911045834422, + 0.0018296547932550311, + 0.005475519225001335, + 0.008232983760535717, + 0.019251927733421326, + 0.018807053565979004, + 0.020574122667312622, + 0.03463461250066757, + 0.016025254502892494, + 0.025805054232478142, + 0.020740127190947533, + 0.022710980847477913, + 0.007019126787781715, + 0.0354093462228775, + 0.01566827856004238, + 0.013156765140593052, + 0.009915580973029137, + 0.013215693645179272, + 0.018613366410136223, + 0.008038532920181751, + 0.0035609849728643894, + 0.005375354550778866, + 0.007577627431601286, + 0.005238104611635208, + 0.003454401856288314, + 0.003996535670012236, + 0.007681633811444044, + 0.007040904834866524, + 0.006496782880276442, + 0.007416841574013233, + 0.01113917212933302, + 0.009044318459928036, + 0.009002956561744213, + 0.00817484688013792, + 0.007426454219967127, + 0.007731383666396141, + 0.0022213070187717676, + 0.0026450452860444784, + 0.002506872173398733, + 0.002730586566030979, + 0.0032359070610255003, + 0.002746140118688345, + 0.002966427942737937, + 0.002141176024451852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.012545948848128319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779990650713444 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779990650713444, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779990650713444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22926977276802063, + 0.07719632983207703, + 0.05662071332335472, + 0.10785121470689774, + 0.09666861593723297, + 0.05552370473742485, + 0.19230204820632935, + 0.10485357791185379, + 0.03928500413894653, + 0.04042905941605568 + ], + "credit_map": [ + 0.01859923265874386, + 0.018031230196356773, + 0.03778485208749771, + 0.046052150428295135, + 0.01561260037124157, + 0.01852361485362053, + 0.041994426399469376, + 0.03267162665724754, + 0.007527529262006283, + 0.004300244152545929, + 0.009880417957901955, + 0.014275746420025826, + 0.0058984411880373955, + 0.009379073977470398, + 0.015139848925173283, + 0.01079502236098051, + 0.0041315616108477116, + 0.00356283038854599, + 0.007623703218996525, + 0.010742053389549255, + 0.009582038037478924, + 0.004759336356073618, + 0.00741993635892868, + 0.008799250237643719, + 0.005182134918868542, + 0.006905944086611271, + 0.014987433329224586, + 0.018146280199289322, + 0.01248614676296711, + 0.01596415974199772, + 0.01581568829715252, + 0.018363410606980324, + 0.005998983047902584, + 0.008388212881982327, + 0.018512222915887833, + 0.022862663492560387, + 0.007788053713738918, + 0.006975135300308466, + 0.013543111272156239, + 0.012600229121744633, + 0.0037083972711116076, + 0.010469562374055386, + 0.006439316552132368, + 0.008219525218009949, + 0.008952939882874489, + 0.005946436896920204, + 0.007217881735414267, + 0.004569637589156628, + 0.010230009444057941, + 0.02160433493554592, + 0.027724646031856537, + 0.049926578998565674, + 0.015613543801009655, + 0.018070731312036514, + 0.03768179193139076, + 0.011450395919382572, + 0.007152773439884186, + 0.014342249371111393, + 0.013371472246944904, + 0.01664385013282299, + 0.010873398743569851, + 0.0133215906098485, + 0.01687603071331978, + 0.012272198684513569, + 0.0058086165226995945, + 0.005511879455298185, + 0.0061800796538591385, + 0.004676199052482843, + 0.0050474596209824085, + 0.0043808333575725555, + 0.004379464779049158, + 0.0033004675060510635, + 0.003252446185797453, + 0.0036536529660224915, + 0.0058420319110155106, + 0.010795568116009235, + 0.005363114643841982, + 0.0034093644935637712, + 0.004003468900918961, + 0.004109408240765333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.009779990650713444 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny colorful wildflowers (purple, pink, and yellow) visible growing through a traditional tatami mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm light creating long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232792217284441 + ], + "opposite_logprob_tokens": [ + -3.7732791900634766 + ], + "logprob_reward": -0.0232792217284441, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232792217284441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23491765558719635, + 0.08181381970643997, + 0.07075275480747223, + 0.15247967839241028, + 0.08321162313222885, + 0.02494838275015354, + 0.08808080106973648, + 0.19471947848796844, + 0.02585434727370739, + 0.04322146624326706 + ], + "credit_map": [ + 0.021579725667834282, + 0.01918848790228367, + 0.036863066256046295, + 0.04352284595370293, + 0.02720949612557888, + 0.01859893649816513, + 0.03774208202958107, + 0.030213017016649246, + 0.004933884367346764, + 0.004040167201310396, + 0.012787994928658009, + 0.019901851192116737, + 0.006291369907557964, + 0.003820448648184538, + 0.01981043629348278, + 0.010227665305137634, + 0.0061875986866652966, + 0.004329460673034191, + 0.011056462302803993, + 0.014131725765764713, + 0.008154544048011303, + 0.005160188768059015, + 0.008422726765275002, + 0.01331004686653614, + 0.008478314615786076, + 0.010499412193894386, + 0.02015325054526329, + 0.02645883336663246, + 0.027884740382432938, + 0.018166620284318924, + 0.02179385907948017, + 0.019044646993279457, + 0.0039466144517064095, + 0.009719623252749443, + 0.015063125640153885, + 0.02141614817082882, + 0.0077115935273468494, + 0.0066824075765907764, + 0.010265781544148922, + 0.008406326174736023, + 0.003007136285305023, + 0.002509730402380228, + 0.002778854686766863, + 0.003200066741555929, + 0.002277118619531393, + 0.003992785234004259, + 0.004540050867944956, + 0.002642639447003603, + 0.005353245418518782, + 0.02218310721218586, + 0.01472857128828764, + 0.010578019544482231, + 0.007227523718029261, + 0.006332498509436846, + 0.01513394434005022, + 0.006543896161019802, + 0.016644982621073723, + 0.02898118644952774, + 0.04008551687002182, + 0.02773449383676052, + 0.016407540068030357, + 0.017167257145047188, + 0.027246830984950066, + 0.02045166678726673, + 0.003604237223044038, + 0.004268219694495201, + 0.00346370623447001, + 0.0025696726515889168, + 0.002697490621358156, + 0.0031859884038567543, + 0.002951111877337098, + 0.0031139219645410776, + 0.005415748339146376, + 0.004529127385467291, + 0.007794099394232035, + 0.007458905223757029, + 0.00436924584209919, + 0.004313447512686253, + 0.004457139875739813, + 0.004883751273155212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.0232792217284441 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_1/index_tensor([7500, 7500])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_1/index_tensor([7500, 7500])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..648620f754ce65f93c79becf393d9c0894f62cd0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_1/index_tensor([7500, 7500])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, upward, medium close-up shot focuses on a grizzled general clad in heavy, battle-worn steel armor. His armor is etched with battle scars and dents. The once intricate carvings on his shoulder plate and gorget are now stained with soot and dried mud, reflecting a dull metallic sheen in the fiery light of the battlefield. His face is matted with sweat and grime, his tangled, sweat-soaked gray hair plastered to his forehead. His eyes, sunken in their sockets, blaze with fury. He stretches his mouth to its limit, releasing a silent but visually arresting roar of fury. The tense jaw and bulging veins in his neck betray the concentration of all his strength. In the background, a chaotic, smoke-filled battlefield unfolds. Thick, gray-black smoke billows, nearly engulfing the blurred silhouettes of soldiers in the distance. Scarlet sparks and embers hang in the air. The entire scene is illuminated by an ominous, flickering light from a distant fire, which carves deep contours on his face and armor, creating an atmosphere of tension, despair, and primitive, savage power.", + "seed": 4083, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_1/index_7500_seed_4083_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.004622384230606258, + "vlm_logprob": -0.004622384230606258, + "vlm_yes_no_margin": 5.575, + "vlm_token_credit": -0.004622384230606258, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00816507451236248 + ], + "opposite_logprob_tokens": [ + -5.133164882659912 + ], + "logprob_reward": -0.00816507451236248, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00816507451236248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6247540712356567, + 0.05121708661317825, + 0.03954833373427391, + 0.026719922199845314, + 0.04246501624584198, + 0.03431030362844467, + 0.07100926339626312, + 0.030038097873330116, + 0.026065994054079056, + 0.05387186259031296 + ], + "credit_map": [ + 0.23383326828479767, + 0.28087618947029114, + 0.011779991909861565, + 0.016194796189665794, + 0.01775846630334854, + 0.043823644518852234, + 0.010718170553445816, + 0.0097695617005229, + 0.0036862515844404697, + 0.011496609076857567, + 0.003575345268473029, + 0.005342374090105295, + 0.0065704998560249805, + 0.014390142634510994, + 0.003485050518065691, + 0.0026708145160228014, + 0.003672564635053277, + 0.003737947205081582, + 0.00442348001524806, + 0.005984411109238863, + 0.007799769286066294, + 0.0053992061875760555, + 0.004458536393940449, + 0.004072417039424181, + 0.002883540466427803, + 0.0028841085731983185, + 0.003359686117619276, + 0.0035573048517107964, + 0.0034055046271532774, + 0.00230360752902925, + 0.005068213678896427, + 0.003257955191656947, + 0.00497589772567153, + 0.003959338180720806, + 0.004696395248174667, + 0.007135801017284393, + 0.008336735889315605, + 0.004833802580833435, + 0.004528451710939407, + 0.003998595289885998, + 0.009721151553094387, + 0.002624089829623699, + 0.0019668617751449347, + 0.0027283786330372095, + 0.004137141164392233, + 0.008307830430567265, + 0.0025716449599713087, + 0.0022532043512910604, + 0.011959383264183998, + 0.026367386803030968, + 0.002510549034923315, + 0.005088526755571365, + 0.0038187075406312943, + 0.016272561624646187, + 0.002864053240045905, + 0.0021280923392623663, + 0.0027904408052563667, + 0.004030388314276934, + 0.003434211714193225, + 0.006492888554930687, + 0.005589339416474104, + 0.002821277594193816, + 0.002884529298171401, + 0.0019950210116803646, + 0.003232988528907299, + 0.002765062265098095, + 0.0027638047467917204, + 0.004714037757366896, + 0.0026729628443717957, + 0.004325972869992256, + 0.003531903028488159, + 0.002059263177216053, + 0.00779950013384223, + 0.00553981214761734, + 0.0033037997782230377, + 0.012682214379310608, + 0.008757025003433228, + 0.0048004803247749805, + 0.006583623122423887, + 0.004405407700687647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.00816507451236248 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002504547592252493 + ], + "opposite_logprob_tokens": [ + -6.002504348754883 + ], + "logprob_reward": -0.002504547592252493, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002504547592252493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6600862145423889, + 0.06570345908403397, + 0.018158797174692154, + 0.028227834030985832, + 0.025783758610486984, + 0.06536348164081573, + 0.018115855753421783, + 0.030731618404388428, + 0.05758963152766228, + 0.03023936040699482 + ], + "credit_map": [ + 0.24237951636314392, + 0.2939029932022095, + 0.015284479595720768, + 0.01888110861182213, + 0.01921015791594982, + 0.04794944450259209, + 0.011862264946103096, + 0.010616232641041279, + 0.008174053393304348, + 0.010917405597865582, + 0.013392294757068157, + 0.0077193910256028175, + 0.004897752311080694, + 0.01402205415070057, + 0.003843697253614664, + 0.0027368124574422836, + 0.0024633246939629316, + 0.001946206670254469, + 0.003136732615530491, + 0.0030357406940311193, + 0.002370120957493782, + 0.0022997634951025248, + 0.0018446274334564805, + 0.001062279217876494, + 0.003417534986510873, + 0.0032749264501035213, + 0.004127637483179569, + 0.0037395439576357603, + 0.00332877179607749, + 0.0026828970294445753, + 0.004768202546983957, + 0.0028883209452033043, + 0.003074409905821085, + 0.0027168970555067062, + 0.0027418259996920824, + 0.0043780868873000145, + 0.005200807936489582, + 0.002569273579865694, + 0.002668282948434353, + 0.0024341752287000418, + 0.014725510962307453, + 0.006011966150254011, + 0.005902300588786602, + 0.004231098107993603, + 0.00626890454441309, + 0.01978369429707527, + 0.004026002716273069, + 0.004414000548422337, + 0.0035090851597487926, + 0.0027652913704514503, + 0.0019622533582150936, + 0.003349660662934184, + 0.0018736940110102296, + 0.0018404553411528468, + 0.0014977871906012297, + 0.0013176295906305313, + 0.003918714355677366, + 0.005942184012383223, + 0.004330549854785204, + 0.004098696168512106, + 0.004886186681687832, + 0.0025147630367428064, + 0.002602455671876669, + 0.0024380700197070837, + 0.014885364100337029, + 0.003447593655437231, + 0.005257845390588045, + 0.0038237476255744696, + 0.0034102899953722954, + 0.019128313288092613, + 0.0043874941766262054, + 0.003248983295634389, + 0.005359196104109287, + 0.0032062926329672337, + 0.002398091135546565, + 0.005730157718062401, + 0.004440475720912218, + 0.0027739638462662697, + 0.0034760155249387026, + 0.00285517075099051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.002504547592252493 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001943010720424354 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.001943010720424354, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001943010720424354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7909933924674988, + 0.028947308659553528, + 0.017812516540288925, + 0.015132555738091469, + 0.013495393097400665, + 0.04451022669672966, + 0.017530914396047592, + 0.007723212242126465, + 0.040633674710989, + 0.0232208501547575 + ], + "credit_map": [ + 0.3003484606742859, + 0.3565208613872528, + 0.014283394441008568, + 0.019087042659521103, + 0.02205497771501541, + 0.05383447930216789, + 0.013410481624305248, + 0.011453627608716488, + 0.0023094401694834232, + 0.006636098492890596, + 0.002512468956410885, + 0.0025825544726103544, + 0.004319042898714542, + 0.007015451323240995, + 0.002168715000152588, + 0.0014035332715138793, + 0.0017984185833483934, + 0.0023077030200511217, + 0.0022649194579571486, + 0.0033602535258978605, + 0.0025313892401754856, + 0.0026537631638348103, + 0.0013939283089712262, + 0.0015021382132545114, + 0.0014904642011970282, + 0.0018764669075608253, + 0.0018395145889371634, + 0.0020854936446994543, + 0.002730197971686721, + 0.0014338736655190587, + 0.0014541229465976357, + 0.0022224201820790768, + 0.0017472143517807126, + 0.0012543410994112492, + 0.002150418236851692, + 0.0018756849458441138, + 0.0022265897132456303, + 0.001982248155400157, + 0.001150500145740807, + 0.001108393887989223, + 0.016143834218382835, + 0.005164667498320341, + 0.0019786665216088295, + 0.0021603982895612717, + 0.0036221498157829046, + 0.010334130376577377, + 0.0017513257917016745, + 0.0033550492953509092, + 0.0029017634224146605, + 0.0047935801558196545, + 0.0013971272855997086, + 0.0024054066743701696, + 0.0013138310750946403, + 0.002748443279415369, + 0.0011518363608047366, + 0.0008189225918613374, + 0.0006621995125897229, + 0.0010928587289527059, + 0.0008500762633047998, + 0.0009689557482488453, + 0.0019034820143133402, + 0.0006972315022721887, + 0.0008509302861057222, + 0.000697477487847209, + 0.00898884516209364, + 0.0031803615856915712, + 0.0022919001057744026, + 0.0048431153409183025, + 0.002512359293177724, + 0.013147715479135513, + 0.0033166585490107536, + 0.002352713607251644, + 0.002599794650450349, + 0.0029766547959297895, + 0.0013908790424466133, + 0.0038199659902602434, + 0.0037089833058416843, + 0.002964559942483902, + 0.0037450760137289762, + 0.002014933153986931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.001943010720424354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052507114596664906 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.0052507114596664906, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052507114596664906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6128007173538208, + 0.03189335763454437, + 0.059832196682691574, + 0.06397109478712082, + 0.02213325724005699, + 0.02751549892127514, + 0.10514476150274277, + 0.02986261248588562, + 0.027009207755327225, + 0.019837278872728348 + ], + "credit_map": [ + 0.22296589612960815, + 0.2642715573310852, + 0.012803967110812664, + 0.016605500131845474, + 0.028602007776498795, + 0.042706042528152466, + 0.013034705072641373, + 0.011811056174337864, + 0.004131767898797989, + 0.0024801313411444426, + 0.006390572991222143, + 0.004618678241968155, + 0.00592093076556921, + 0.004669941961765289, + 0.002099413424730301, + 0.0015819226391613483, + 0.007219724357128143, + 0.0071098338812589645, + 0.005289019085466862, + 0.008955098688602448, + 0.012460296973586082, + 0.008405473083257675, + 0.006266971118748188, + 0.004125779494643211, + 0.00685182074084878, + 0.007870162837207317, + 0.00838053785264492, + 0.009047471918165684, + 0.009020970202982426, + 0.0058420090936124325, + 0.008695408701896667, + 0.008262715302407742, + 0.002932255621999502, + 0.0023054475896060467, + 0.0027695326134562492, + 0.0032065780833363533, + 0.003566078608855605, + 0.0022987041156738997, + 0.002304167952388525, + 0.0027504952158778906, + 0.006259374786168337, + 0.0029965578578412533, + 0.003430198412388563, + 0.0019038242753595114, + 0.002501252107322216, + 0.007096693851053715, + 0.0014340545749291778, + 0.0018935444531962276, + 0.014387466944754124, + 0.039295896887779236, + 0.0037601422518491745, + 0.006908994633704424, + 0.0068460325710475445, + 0.028052447363734245, + 0.0030532123055309057, + 0.0028405652847141027, + 0.004091076087206602, + 0.00539585342630744, + 0.0040235454216599464, + 0.00413935724645853, + 0.004783748649060726, + 0.002639733487740159, + 0.0027603397611528635, + 0.0020289593376219273, + 0.005413270555436611, + 0.0021870394703000784, + 0.0033917834516614676, + 0.004904341883957386, + 0.0016285940073430538, + 0.005790636409074068, + 0.0018883892334997654, + 0.0018051518127322197, + 0.002315738471224904, + 0.0046662259846925735, + 0.0021838960237801075, + 0.0017972528003156185, + 0.0022919930052012205, + 0.003020215779542923, + 0.001833354472182691, + 0.0017286029178649187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.0052507114596664906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248576868325472 + ], + "opposite_logprob_tokens": [ + -5.255248546600342 + ], + "logprob_reward": -0.005248576868325472, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248576868325472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5160759091377258, + 0.07669949531555176, + 0.030483271926641464, + 0.020348045974969864, + 0.024052325636148453, + 0.017708253115415573, + 0.08026646822690964, + 0.025840330868959427, + 0.1483144462108612, + 0.060211487114429474 + ], + "credit_map": [ + 0.1885516196489334, + 0.22319184243679047, + 0.016172535717487335, + 0.01944727636873722, + 0.01435895636677742, + 0.03744642436504364, + 0.008917177096009254, + 0.007990053854882717, + 0.008559788577258587, + 0.01742989383637905, + 0.01218520849943161, + 0.00872641708701849, + 0.005901702679693699, + 0.017514223232865334, + 0.0037690841127187014, + 0.0026131735648959875, + 0.0036049236077815294, + 0.004468224477022886, + 0.0022950400598347187, + 0.006098196376115084, + 0.003567630657926202, + 0.003970759455114603, + 0.003009001724421978, + 0.0034694939386099577, + 0.0025906458031386137, + 0.002367530483752489, + 0.0023083030246198177, + 0.003262524027377367, + 0.002187717705965042, + 0.0019351816736161709, + 0.0031218654476106167, + 0.0025742785073816776, + 0.003809808287769556, + 0.002319875406101346, + 0.00319037027657032, + 0.003069656202569604, + 0.0040493798442184925, + 0.0027781527023762465, + 0.002518757013604045, + 0.002316325670108199, + 0.002064540283754468, + 0.002400106517598033, + 0.0028739883564412594, + 0.002111878013238311, + 0.0022566947154700756, + 0.0021154729183763266, + 0.002006016205996275, + 0.0018795568030327559, + 0.010023615323007107, + 0.024697987362742424, + 0.004757770337164402, + 0.01025400310754776, + 0.006098743993788958, + 0.017649445682764053, + 0.004110285546630621, + 0.0026746168732643127, + 0.003509131958708167, + 0.00562673807144165, + 0.003958285786211491, + 0.0039606718346476555, + 0.0025195088237524033, + 0.0019996098708361387, + 0.002342690248042345, + 0.001923695090226829, + 0.04254772141575813, + 0.008796915411949158, + 0.012249486520886421, + 0.005401895381510258, + 0.006005230825394392, + 0.05743587762117386, + 0.006250668317079544, + 0.009626666083931923, + 0.011184357106685638, + 0.011059578508138657, + 0.0034092222340404987, + 0.018870247527956963, + 0.003869520965963602, + 0.0036449821200221777, + 0.0033940228167921305, + 0.00477956049144268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.005248576868325472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward, medium close-up shot focuses on a grizzled general clad in heavy, battle-worn steel armor. His armor is etched with battle scars and dents. The once intricate carvings on his shoulder plate and gorget are now stained with soot and dried mud, reflecting a dull metallic sheen in the fiery light of the battlefield. His face is matted with sweat and grime, his tangled, sweat-soaked gray hair plastered to his forehead. His eyes, sunken in their sockets, blaze with fury. He stretches his mouth to its limit, releasing a silent but visually arresting roar of fury. The tense jaw and bulging veins in his neck betray the concentration of all his strength. In the background, a chaotic, smoke-filled battlefield unfolds. Thick, gray-black smoke billows, nearly engulfing the blurred silhouettes of soldiers in the distance. Scarlet sparks and embers hang in the air. The entire scene is illuminated by an ominous, flickering light from a distant fire, which carves deep contours on his face and armor, creating an atmosphere of tension, despair, and primitive, savage power.", + "seed": 4084, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_1/index_7500_seed_4084_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.004541045194491744, + "vlm_logprob": -0.004541045194491744, + "vlm_yes_no_margin": 5.625, + "vlm_token_credit": -0.004541045194491744, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005445409100502729 + ], + "opposite_logprob_tokens": [ + -5.50544548034668 + ], + "logprob_reward": -0.005445409100502729, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005445409100502729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08030184358358383, + 0.029083723202347755, + 0.03317180275917053, + 0.17503392696380615, + 0.05188365280628204, + 0.1303841769695282, + 0.20170985162258148, + 0.16761218011379242, + 0.06892374157905579, + 0.061895109713077545 + ], + "credit_map": [ + 0.00967444758862257, + 0.008098611608147621, + 0.01374843716621399, + 0.009399691596627235, + 0.007478024810552597, + 0.013557963073253632, + 0.005875733215361834, + 0.012468940578401089, + 0.0030309902504086494, + 0.002643014071509242, + 0.004120298195630312, + 0.004436733666807413, + 0.002724862191826105, + 0.0035575623624026775, + 0.004892494063824415, + 0.0036777681671082973, + 0.004113445989787579, + 0.004827100783586502, + 0.0033588793594390154, + 0.0028274073265492916, + 0.004731131251901388, + 0.004002845846116543, + 0.0052638910710811615, + 0.004047102760523558, + 0.02166389487683773, + 0.014457196928560734, + 0.013967855833470821, + 0.01573231816291809, + 0.03584175184369087, + 0.032419733703136444, + 0.0266492310911417, + 0.014301944524049759, + 0.008074909448623657, + 0.011540171690285206, + 0.004893920384347439, + 0.004198233596980572, + 0.004586631897836924, + 0.008763767778873444, + 0.004853145219385624, + 0.004972871858626604, + 0.022123202681541443, + 0.011268826201558113, + 0.008180445991456509, + 0.013626638799905777, + 0.012523960322141647, + 0.014590401202440262, + 0.015071102418005466, + 0.03299959376454353, + 0.023004556074738503, + 0.02331528440117836, + 0.019729623571038246, + 0.02714083157479763, + 0.032066550105810165, + 0.02134394459426403, + 0.032170090824365616, + 0.022938983514904976, + 0.037930406630039215, + 0.008660739287734032, + 0.0106589887291193, + 0.012697647325694561, + 0.018261568620800972, + 0.047335900366306305, + 0.015212849713861942, + 0.016854064539074898, + 0.02073149010539055, + 0.01964644342660904, + 0.005940892733633518, + 0.00396032864227891, + 0.004728391300886869, + 0.004449916537851095, + 0.005557419266551733, + 0.003908860497176647, + 0.00940766092389822, + 0.0077033755369484425, + 0.005033170338720083, + 0.010380083695054054, + 0.007625741418451071, + 0.007239623926579952, + 0.008459211327135563, + 0.006046243477612734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.005445409100502729 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032071841415017843 + ], + "opposite_logprob_tokens": [ + -5.753207206726074 + ], + "logprob_reward": -0.0032071841415017843, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032071841415017843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1532231867313385, + 0.06370196491479874, + 0.04650137573480606, + 0.06896250694990158, + 0.07310807704925537, + 0.08875404298305511, + 0.11553403735160828, + 0.12670332193374634, + 0.1562337428331375, + 0.1072777509689331 + ], + "credit_map": [ + 0.02405407838523388, + 0.01974225603044033, + 0.01944045163691044, + 0.01893533393740654, + 0.01662421226501465, + 0.02277257852256298, + 0.012362947687506676, + 0.019291335716843605, + 0.007600318640470505, + 0.004644367843866348, + 0.010201909579336643, + 0.010806947015225887, + 0.008502170443534851, + 0.0065819500014185905, + 0.007866141386330128, + 0.007498160004615784, + 0.005278330296278, + 0.00475776894018054, + 0.005352068692445755, + 0.004722972866147757, + 0.008884035982191563, + 0.007029938045889139, + 0.005406858399510384, + 0.00506940484046936, + 0.010452555492520332, + 0.006560575217008591, + 0.0058758375234901905, + 0.006113140843808651, + 0.012165986001491547, + 0.015241380780935287, + 0.007868414744734764, + 0.004684614483267069, + 0.009019528515636921, + 0.007034734357148409, + 0.00556073896586895, + 0.00766646396368742, + 0.011795535683631897, + 0.012029128149151802, + 0.013070957735180855, + 0.006930988281965256, + 0.01104450598359108, + 0.011328574270009995, + 0.012203565798699856, + 0.008500132709741592, + 0.010228044353425503, + 0.012069674208760262, + 0.012247422710061073, + 0.011132124811410904, + 0.01711134798824787, + 0.013073070906102657, + 0.007369325961917639, + 0.012314952909946442, + 0.02230011112987995, + 0.01355697587132454, + 0.014263421297073364, + 0.015544828027486801, + 0.029747797176241875, + 0.007383636198937893, + 0.008434412069618702, + 0.009090318344533443, + 0.012268111109733582, + 0.03411860391497612, + 0.011957147158682346, + 0.013703293167054653, + 0.034355055540800095, + 0.026379510760307312, + 0.01960711181163788, + 0.016203369945287704, + 0.016330605372786522, + 0.010916542261838913, + 0.020243573933839798, + 0.012197976931929588, + 0.018553204834461212, + 0.01837768405675888, + 0.011031095869839191, + 0.022169863805174828, + 0.009306550957262516, + 0.009480969980359077, + 0.010128084570169449, + 0.008230287581682205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.0032071841415017843 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011809049174189568 + ], + "opposite_logprob_tokens": [ + -6.751181125640869 + ], + "logprob_reward": -0.0011809049174189568, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011809049174189568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08703447133302689, + 0.05377225577831268, + 0.04655150696635246, + 0.14766308665275574, + 0.08981925249099731, + 0.19655366241931915, + 0.06488589942455292, + 0.171708881855011, + 0.06843803822994232, + 0.07357292622327805 + ], + "credit_map": [ + 0.013221930712461472, + 0.008084860630333424, + 0.014638371765613556, + 0.0075874668546020985, + 0.008884761482477188, + 0.012944059446454048, + 0.006419350393116474, + 0.015253669582307339, + 0.005779317580163479, + 0.003689134493470192, + 0.00905431155115366, + 0.009374285116791725, + 0.005417820066213608, + 0.005994725506752729, + 0.007381422445178032, + 0.007081239018589258, + 0.005707135424017906, + 0.004983632359653711, + 0.006742025259882212, + 0.005017902702093124, + 0.006069207563996315, + 0.0062734996899962425, + 0.005116465501487255, + 0.006641637533903122, + 0.01670898124575615, + 0.01239839382469654, + 0.011991042643785477, + 0.012359595857560635, + 0.033436018973588943, + 0.0270441472530365, + 0.022494228556752205, + 0.01123068481683731, + 0.014668531715869904, + 0.02419060282409191, + 0.012232727371156216, + 0.009768436662852764, + 0.005865470971912146, + 0.008980527520179749, + 0.006321001332253218, + 0.007791962940245867, + 0.03697013109922409, + 0.018677419051527977, + 0.016706285998225212, + 0.014033923856914043, + 0.015241263434290886, + 0.020024806261062622, + 0.020380590111017227, + 0.05451923981308937, + 0.008501676842570305, + 0.005082039162516594, + 0.010345387272536755, + 0.008265928365290165, + 0.00807320699095726, + 0.00829447153955698, + 0.008704440668225288, + 0.007618749514222145, + 0.04317498207092285, + 0.008550428785383701, + 0.011098810471594334, + 0.013603057712316513, + 0.014574752189218998, + 0.04823223128914833, + 0.013810815289616585, + 0.0186638031154871, + 0.01588769629597664, + 0.018654899671673775, + 0.003414368024095893, + 0.004781246650964022, + 0.008555498905479908, + 0.005125861149281263, + 0.008005081675946712, + 0.004013386555016041, + 0.012190110981464386, + 0.014580579474568367, + 0.005090131424367428, + 0.008281554095447063, + 0.009013253264129162, + 0.007567774970084429, + 0.011404748074710369, + 0.00544477254152298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 3 + ], + "selected_reward": -0.0011809049174189568 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762398075312376 + ], + "opposite_logprob_tokens": [ + -4.88262414932251 + ], + "logprob_reward": -0.00762398075312376, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762398075312376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14696446061134338, + 0.04120970889925957, + 0.04483220726251602, + 0.05148925632238388, + 0.04008256644010544, + 0.22585538029670715, + 0.06917235255241394, + 0.18985435366630554, + 0.08356794714927673, + 0.10697180032730103 + ], + "credit_map": [ + 0.02606946974992752, + 0.013501938432455063, + 0.021096758544445038, + 0.014986258931457996, + 0.013543998822569847, + 0.01867051050066948, + 0.012140856124460697, + 0.026954663917422295, + 0.0032045196276158094, + 0.002696175826713443, + 0.004471269436180592, + 0.0057361130602657795, + 0.005990896373987198, + 0.0061357333324849606, + 0.0076677571050822735, + 0.005307246930897236, + 0.00488682696595788, + 0.005486653186380863, + 0.005155968014150858, + 0.0032322730403393507, + 0.007626205682754517, + 0.007246678229421377, + 0.004669356159865856, + 0.006528245750814676, + 0.006871745456010103, + 0.005659061949700117, + 0.0052560968324542046, + 0.004875962156802416, + 0.008505248464643955, + 0.007732139900326729, + 0.006861994508653879, + 0.005727005656808615, + 0.006428876426070929, + 0.007950649596750736, + 0.004372789058834314, + 0.004367989953607321, + 0.0036720060743391514, + 0.004596721846610308, + 0.004831328056752682, + 0.003862204495817423, + 0.04980608448386192, + 0.01563280262053013, + 0.014145496301352978, + 0.01641770265996456, + 0.013564873486757278, + 0.021777493879199028, + 0.0243462435901165, + 0.07016469538211823, + 0.008751967921853065, + 0.007195963989943266, + 0.006382108200341463, + 0.007038671988993883, + 0.011070827953517437, + 0.008417886681854725, + 0.011508218012750149, + 0.008806713856756687, + 0.06098305433988571, + 0.007084937766194344, + 0.006420606281608343, + 0.008066880516707897, + 0.0077745928429067135, + 0.06975854188203812, + 0.011746692471206188, + 0.018019039183855057, + 0.018520286306738853, + 0.015278236009180546, + 0.005377037450671196, + 0.007336348760873079, + 0.010428598150610924, + 0.009337048046290874, + 0.01005037222057581, + 0.007240019738674164, + 0.011453897692263126, + 0.013390179723501205, + 0.009043770842254162, + 0.01575305312871933, + 0.016492100432515144, + 0.011534427292644978, + 0.019736414775252342, + 0.009567955508828163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.00762398075312376 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052477470599114895 + ], + "opposite_logprob_tokens": [ + -5.255247592926025 + ], + "logprob_reward": -0.0052477470599114895, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052477470599114895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19671325385570526, + 0.07650526612997055, + 0.0387699268758297, + 0.10786083340644836, + 0.08538807183504105, + 0.05333029478788376, + 0.06159088388085365, + 0.08256226778030396, + 0.24684010446071625, + 0.05043908581137657 + ], + "credit_map": [ + 0.03524142503738403, + 0.013872766867280006, + 0.02664903551340103, + 0.016605842858552933, + 0.01564055308699608, + 0.03138837590813637, + 0.01570258103311062, + 0.04161267727613449, + 0.008645439520478249, + 0.006562296766787767, + 0.012292152270674706, + 0.01668146438896656, + 0.008400478400290012, + 0.005840047728270292, + 0.008616351522505283, + 0.009467030875384808, + 0.004280107095837593, + 0.00353518920019269, + 0.0028086998499929905, + 0.0031495322473347187, + 0.008685806766152382, + 0.008115771226584911, + 0.0048886616714298725, + 0.0033061576541513205, + 0.021209612488746643, + 0.007675260305404663, + 0.009131735190749168, + 0.0070646158419549465, + 0.017859207466244698, + 0.026736581698060036, + 0.01142941415309906, + 0.006754414178431034, + 0.014361408539116383, + 0.018284181132912636, + 0.007700024172663689, + 0.008948770351707935, + 0.008486349135637283, + 0.01231587678194046, + 0.007314932066947222, + 0.007976535707712173, + 0.010080115869641304, + 0.004124543629586697, + 0.004200029652565718, + 0.006380017846822739, + 0.005368561949580908, + 0.006452224217355251, + 0.006419675890356302, + 0.010305123403668404, + 0.008254542015492916, + 0.0053563970141112804, + 0.010038772597908974, + 0.006193187553435564, + 0.007604627404361963, + 0.010599905624985695, + 0.006864402908831835, + 0.006679048761725426, + 0.013256208039820194, + 0.005490934941917658, + 0.009629851207137108, + 0.009508317336440086, + 0.010132290422916412, + 0.016574162989854813, + 0.009045298211276531, + 0.008925204165279865, + 0.07294612377882004, + 0.0778450146317482, + 0.01919768750667572, + 0.01764528825879097, + 0.017200563102960587, + 0.014177828095853329, + 0.017082473263144493, + 0.01074514165520668, + 0.008531083352863789, + 0.008556879125535488, + 0.004722871817648411, + 0.007540583144873381, + 0.005106513388454914, + 0.005170310381799936, + 0.006695459131151438, + 0.004115384537726641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.0052477470599114895 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward, medium close-up shot focuses on a grizzled general clad in heavy, battle-worn steel armor. His armor is etched with battle scars and dents. The once intricate carvings on his shoulder plate and gorget are now stained with soot and dried mud, reflecting a dull metallic sheen in the fiery light of the battlefield. His face is matted with sweat and grime, his tangled, sweat-soaked gray hair plastered to his forehead. His eyes, sunken in their sockets, blaze with fury. He stretches his mouth to its limit, releasing a silent but visually arresting roar of fury. The tense jaw and bulging veins in his neck betray the concentration of all his strength. In the background, a chaotic, smoke-filled battlefield unfolds. Thick, gray-black smoke billows, nearly engulfing the blurred silhouettes of soldiers in the distance. Scarlet sparks and embers hang in the air. The entire scene is illuminated by an ominous, flickering light from a distant fire, which carves deep contours on his face and armor, creating an atmosphere of tension, despair, and primitive, savage power.", + "seed": 4085, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_1/index_7500_seed_4085_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.007942613121122122, + "vlm_logprob": -0.007942613121122122, + "vlm_yes_no_margin": 5.35, + "vlm_token_credit": -0.007942613121122122, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004155097529292107 + ], + "opposite_logprob_tokens": [ + -6.129155158996582 + ], + "logprob_reward": -0.004155097529292107, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.004155097529292107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12356971949338913, + 0.07924757152795792, + 0.13629120588302612, + 0.05409158021211624, + 0.05467512086033821, + 0.261740118265152, + 0.08558936417102814, + 0.027793381363153458, + 0.09605739265680313, + 0.08094450831413269 + ], + "credit_map": [ + 0.015589391812682152, + 0.017774896696209908, + 0.01496587973088026, + 0.016429787501692772, + 0.014393256045877934, + 0.01623951643705368, + 0.013044636696577072, + 0.01513235829770565, + 0.006402675993740559, + 0.010767624713480473, + 0.011359712108969688, + 0.012089909054338932, + 0.011375188827514648, + 0.006463285535573959, + 0.01334104873239994, + 0.007448130287230015, + 0.036308277398347855, + 0.011826179921627045, + 0.011012321338057518, + 0.011981855146586895, + 0.022448981180787086, + 0.018365884199738503, + 0.014902365393936634, + 0.00944534596055746, + 0.006499019451439381, + 0.006762437988072634, + 0.007780622690916061, + 0.0057186163030564785, + 0.008607965894043446, + 0.005399039480835199, + 0.007171675097197294, + 0.006152199115604162, + 0.007826527580618858, + 0.004743831232190132, + 0.008594297803938389, + 0.005547532346099615, + 0.008397508412599564, + 0.005876997951418161, + 0.0077626691199839115, + 0.0059257554821670055, + 0.053466103971004486, + 0.02094597928225994, + 0.0380411371588707, + 0.013404528610408306, + 0.017253335565328598, + 0.08282977342605591, + 0.01739438623189926, + 0.018404869362711906, + 0.012464831583201885, + 0.013415733352303505, + 0.009275808930397034, + 0.013664737343788147, + 0.012274087406694889, + 0.009349077008664608, + 0.008689961396157742, + 0.0064551266841590405, + 0.0036200364120304585, + 0.0031645020935684443, + 0.0019254301441833377, + 0.0029474732000380754, + 0.0032137050293385983, + 0.0033279319759458303, + 0.005546643864363432, + 0.004047658760100603, + 0.013731974177062511, + 0.009336894378066063, + 0.008628427982330322, + 0.00954305101186037, + 0.015938151627779007, + 0.01359586138278246, + 0.015502273105084896, + 0.009780759923160076, + 0.012389780022203922, + 0.007410643622279167, + 0.005339252296835184, + 0.01130285207182169, + 0.010527883656322956, + 0.013197925873100758, + 0.012706938199698925, + 0.008069236762821674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.004155097529292107 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004116395954042673 + ], + "opposite_logprob_tokens": [ + -5.504116535186768 + ], + "logprob_reward": -0.004116395954042673, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004116395954042673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19842392206192017, + 0.06943801045417786, + 0.12272588908672333, + 0.0775819644331932, + 0.30046603083610535, + 0.06681910902261734, + 0.05442064628005028, + 0.02445858344435692, + 0.046780895441770554, + 0.03888492286205292 + ], + "credit_map": [ + 0.04542166367173195, + 0.02171466313302517, + 0.03999185189604759, + 0.016677645966410637, + 0.013758592307567596, + 0.020784590393304825, + 0.01626436784863472, + 0.023810558021068573, + 0.005679579451680183, + 0.006774351932108402, + 0.011827399954199791, + 0.010004492476582527, + 0.009157918393611908, + 0.0072927954606711864, + 0.011728343553841114, + 0.0069731236435472965, + 0.028963472694158554, + 0.009429527446627617, + 0.006270694080740213, + 0.010439135134220123, + 0.018042031675577164, + 0.01809321530163288, + 0.02530715800821781, + 0.0061806621961295605, + 0.015015188604593277, + 0.008937089703977108, + 0.009114358574151993, + 0.006335534155368805, + 0.0073918914422392845, + 0.006975633092224598, + 0.014511806890368462, + 0.00930046383291483, + 0.03857037425041199, + 0.021464582532644272, + 0.0577593669295311, + 0.03696824237704277, + 0.04226868972182274, + 0.040634624660015106, + 0.030889514833688736, + 0.03191061690449715, + 0.010168896056711674, + 0.006539716385304928, + 0.008566617965698242, + 0.006754951551556587, + 0.009538860991597176, + 0.009854407981038094, + 0.008381773717701435, + 0.0070138853043317795, + 0.007134395185858011, + 0.006143143400549889, + 0.009168694727122784, + 0.006267441436648369, + 0.00939869973808527, + 0.005378768779337406, + 0.005462530069053173, + 0.005466978996992111, + 0.003498999634757638, + 0.003083078656345606, + 0.0016277480171993375, + 0.0019917746540158987, + 0.0036471583880484104, + 0.0036011734046041965, + 0.0033400515094399452, + 0.003668599994853139, + 0.007273546885699034, + 0.0048223501071333885, + 0.00369728310033679, + 0.005601116921752691, + 0.007143283262848854, + 0.006192540284246206, + 0.008410979062318802, + 0.0036397979129105806, + 0.005014265887439251, + 0.0039221555925905704, + 0.004873414523899555, + 0.005086980294436216, + 0.005050596781075001, + 0.005480452906340361, + 0.004889545496553183, + 0.004567509517073631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.004116395954042673 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011848341673612595 + ], + "opposite_logprob_tokens": [ + -6.751184940338135 + ], + "logprob_reward": -0.0011848341673612595, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011848341673612595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13184933364391327, + 0.07751154154539108, + 0.1125999242067337, + 0.05846010893583298, + 0.13475272059440613, + 0.09382354468107224, + 0.04361614212393761, + 0.031116275116801262, + 0.22685424983501434, + 0.08941611647605896 + ], + "credit_map": [ + 0.0168977789580822, + 0.015085294842720032, + 0.01968594826757908, + 0.018244819715619087, + 0.013933629728853703, + 0.014710713177919388, + 0.017718983814120293, + 0.015572180040180683, + 0.007414430379867554, + 0.010918579995632172, + 0.010467449203133583, + 0.008146604523062706, + 0.009747643955051899, + 0.006616682279855013, + 0.016372662037611008, + 0.007827485911548138, + 0.030923480167984962, + 0.007154709659516811, + 0.006175625137984753, + 0.008621755987405777, + 0.011937136761844158, + 0.015952952206134796, + 0.025852857157588005, + 0.005981399677693844, + 0.006869797129184008, + 0.006003261543810368, + 0.009290745481848717, + 0.006839779671281576, + 0.0061128041706979275, + 0.004895254038274288, + 0.013101388700306416, + 0.005347079131752253, + 0.01844790391623974, + 0.009708906523883343, + 0.024142295122146606, + 0.016098013147711754, + 0.017230836674571037, + 0.01999446377158165, + 0.015430975705385208, + 0.013699330389499664, + 0.01557048037648201, + 0.010093733668327332, + 0.01823638565838337, + 0.008016047067940235, + 0.009960753843188286, + 0.017017798498272896, + 0.007225631270557642, + 0.007702713832259178, + 0.005869904067367315, + 0.005511871539056301, + 0.005629758816212416, + 0.004788827616721392, + 0.007552117109298706, + 0.004617081955075264, + 0.004563325550407171, + 0.0050832550041377544, + 0.004083661828190088, + 0.003687457414343953, + 0.0033278614282608032, + 0.0037954794242978096, + 0.003981659654527903, + 0.0034973465371876955, + 0.005376699846237898, + 0.0033661096822470427, + 0.030859779566526413, + 0.02164660207927227, + 0.029016686603426933, + 0.024705642834305763, + 0.0371762253344059, + 0.02275451458990574, + 0.0310433991253376, + 0.02965139038860798, + 0.014062163420021534, + 0.00984796229749918, + 0.01113191433250904, + 0.010212165303528309, + 0.014380170032382011, + 0.010811446234583855, + 0.010873064398765564, + 0.00809723325073719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.0011848341673612595 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066256910562515 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.016066256910562515, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066256910562515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09697794914245605, + 0.06729964911937714, + 0.1279759407043457, + 0.11357579380273819, + 0.052241403609514236, + 0.12736567854881287, + 0.1386319398880005, + 0.03534986078739166, + 0.11952099949121475, + 0.12106078863143921 + ], + "credit_map": [ + 0.013506423681974411, + 0.013725299388170242, + 0.013572050258517265, + 0.012169820256531239, + 0.011419915594160557, + 0.011349908076226711, + 0.010088258422911167, + 0.011146276257932186, + 0.006990219000726938, + 0.007963170297443867, + 0.005179302766919136, + 0.00794088002294302, + 0.013607984408736229, + 0.010570542886853218, + 0.007183500099927187, + 0.007864044979214668, + 0.02030445635318756, + 0.012747352942824364, + 0.010174675844609737, + 0.00993373990058899, + 0.02588699571788311, + 0.024041498079895973, + 0.0167243592441082, + 0.008162862621247768, + 0.019598709419369698, + 0.011430277489125729, + 0.009412200190126896, + 0.0074156057089567184, + 0.019956625998020172, + 0.016824888065457344, + 0.01301288791000843, + 0.0159246064722538, + 0.009920801967382431, + 0.00471225380897522, + 0.0035210817586630583, + 0.003973428625613451, + 0.004930121824145317, + 0.014518016017973423, + 0.006108468864113092, + 0.004557234235107899, + 0.029262728989124298, + 0.010992849245667458, + 0.010229724459350109, + 0.00733755761757493, + 0.008563682436943054, + 0.04323628544807434, + 0.006537029519677162, + 0.011205838061869144, + 0.021748678758740425, + 0.01357552781701088, + 0.009922653436660767, + 0.013713136315345764, + 0.03371145576238632, + 0.016317669302225113, + 0.015598513185977936, + 0.014044307172298431, + 0.005032951012253761, + 0.005881246645003557, + 0.0036819111555814743, + 0.0032312532421201468, + 0.006683495361357927, + 0.003983191680163145, + 0.0031390173826366663, + 0.003716791747137904, + 0.014875827357172966, + 0.011920442804694176, + 0.008762706071138382, + 0.01072031818330288, + 0.026414301246404648, + 0.014259363524615765, + 0.01876986399292946, + 0.013798176310956478, + 0.014204692095518112, + 0.011775015853345394, + 0.010137570090591908, + 0.013055982068181038, + 0.026108838617801666, + 0.014666194096207619, + 0.01895308867096901, + 0.012159407138824463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.016066256910562515 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190481044352055 + ], + "opposite_logprob_tokens": [ + -4.264190673828125 + ], + "logprob_reward": -0.014190481044352055, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190481044352055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4796600937843323, + 0.05146665871143341, + 0.0773836001753807, + 0.04766746610403061, + 0.12193991988897324, + 0.09489452838897705, + 0.028933417052030563, + 0.038527023047208786, + 0.030412966385483742, + 0.029114285483956337 + ], + "credit_map": [ + 0.0850830003619194, + 0.053159285336732864, + 0.0820237547159195, + 0.04303713142871857, + 0.04834892600774765, + 0.05756481736898422, + 0.05219903960824013, + 0.05824416130781174, + 0.005275746341794729, + 0.009416846558451653, + 0.004860579036176205, + 0.006306082475930452, + 0.00732204457744956, + 0.006671625655144453, + 0.006105881184339523, + 0.0055078561417758465, + 0.023419873788952827, + 0.006627193186432123, + 0.006227067671716213, + 0.0051338099874556065, + 0.008145779371261597, + 0.015288354828953743, + 0.008491475135087967, + 0.004050054121762514, + 0.00526496022939682, + 0.004908102564513683, + 0.00822741910815239, + 0.005316622089594603, + 0.005435979459434748, + 0.003673422383144498, + 0.009698687121272087, + 0.005142278037965298, + 0.024179210886359215, + 0.012020296417176723, + 0.01245624665170908, + 0.010874121449887753, + 0.014766602776944637, + 0.0251457542181015, + 0.010121823288500309, + 0.01237587071955204, + 0.0207426268607378, + 0.008430778980255127, + 0.012013300321996212, + 0.004493151791393757, + 0.0054520610719919205, + 0.03166402131319046, + 0.0052515724673867226, + 0.006847022566944361, + 0.004249590914696455, + 0.0024679899215698242, + 0.003747363807633519, + 0.004685295280069113, + 0.003010965185239911, + 0.002290158299729228, + 0.005979849491268396, + 0.0025022050831466913, + 0.006453903391957283, + 0.00301984790712595, + 0.0021790422033518553, + 0.0024633356370031834, + 0.0047140843234956264, + 0.0060499683022499084, + 0.006918687839061022, + 0.006728156004101038, + 0.00442545348778367, + 0.00318037043325603, + 0.002736635273322463, + 0.0032038437202572823, + 0.005299933720380068, + 0.004017062485218048, + 0.005165221635252237, + 0.002384448191151023, + 0.003573628840968013, + 0.0035614054650068283, + 0.003980571404099464, + 0.003387556178495288, + 0.004388326779007912, + 0.0025715103838592768, + 0.003977329935878515, + 0.0036739595234394073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.014190481044352055 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward, medium close-up shot focuses on a grizzled general clad in heavy, battle-worn steel armor. His armor is etched with battle scars and dents. The once intricate carvings on his shoulder plate and gorget are now stained with soot and dried mud, reflecting a dull metallic sheen in the fiery light of the battlefield. His face is matted with sweat and grime, his tangled, sweat-soaked gray hair plastered to his forehead. His eyes, sunken in their sockets, blaze with fury. He stretches his mouth to its limit, releasing a silent but visually arresting roar of fury. The tense jaw and bulging veins in his neck betray the concentration of all his strength. In the background, a chaotic, smoke-filled battlefield unfolds. Thick, gray-black smoke billows, nearly engulfing the blurred silhouettes of soldiers in the distance. Scarlet sparks and embers hang in the air. The entire scene is illuminated by an ominous, flickering light from a distant fire, which carves deep contours on his face and armor, creating an atmosphere of tension, despair, and primitive, savage power.", + "seed": 4086, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_1/index_7500_seed_4086_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.008844583621248603, + "vlm_logprob": -0.008844583621248603, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.008844583621248603, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004545119125396013 + ], + "opposite_logprob_tokens": [ + -5.754545211791992 + ], + "logprob_reward": -0.004545119125396013, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.004545119125396013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031170543283224106, + 0.021451840177178383, + 0.02500133402645588, + 0.08230010420084, + 0.020083842799067497, + 0.031244592741131783, + 0.017132528126239777, + 0.03191480413079262, + 0.034324850887060165, + 0.7053755521774292 + ], + "credit_map": [ + 0.0034766802564263344, + 0.005629465915262699, + 0.0027097982820123434, + 0.0036711671855300665, + 0.003526408690959215, + 0.004329217132180929, + 0.004430863074958324, + 0.0033969401847571135, + 0.002737452043220401, + 0.0036653489805758, + 0.001674640690907836, + 0.0033615774009376764, + 0.003396127140149474, + 0.0023815170861780643, + 0.002679102122783661, + 0.0015560733154416084, + 0.002387936692684889, + 0.0025398610159754753, + 0.002132831374183297, + 0.0038867774419486523, + 0.006180354859679937, + 0.002831194084137678, + 0.003011079505085945, + 0.0020312978886067867, + 0.01218811608850956, + 0.005758051294833422, + 0.012653494253754616, + 0.007443311158567667, + 0.006528620142489672, + 0.020325927063822746, + 0.00906453188508749, + 0.00833805464208126, + 0.0025844001211225986, + 0.0026057595387101173, + 0.0015262061497196555, + 0.002211944665759802, + 0.0028915046714246273, + 0.004192463122308254, + 0.0026597296819090843, + 0.001411835546605289, + 0.004161341581493616, + 0.0035796831361949444, + 0.003949108999222517, + 0.004853175953030586, + 0.004824878182262182, + 0.003979030530899763, + 0.0033656328450888395, + 0.002531742211431265, + 0.0024668448604643345, + 0.002308956813067198, + 0.0025464394129812717, + 0.0016301816795021296, + 0.002158376621082425, + 0.0027415137737989426, + 0.001461259089410305, + 0.0018189552938565612, + 0.004209529608488083, + 0.004507833626121283, + 0.004893742501735687, + 0.0029714773409068584, + 0.004512881860136986, + 0.0057120067067444324, + 0.0024856841191649437, + 0.0026216479018330574, + 0.008791603147983551, + 0.0037889350205659866, + 0.0032403592485934496, + 0.003118680790066719, + 0.002044111490249634, + 0.009109697304666042, + 0.0021144752390682697, + 0.0021169891115278006, + 0.09440964460372925, + 0.09290221333503723, + 0.023573970422148705, + 0.02069907635450363, + 0.016599083319306374, + 0.2893347442150116, + 0.018417194485664368, + 0.14943961799144745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.004545119125396013 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003212888026610017 + ], + "opposite_logprob_tokens": [ + -5.753212928771973 + ], + "logprob_reward": -0.003212888026610017, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003212888026610017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08254143595695496, + 0.03475397825241089, + 0.03282656893134117, + 0.05093171447515488, + 0.051188770681619644, + 0.05776066333055496, + 0.04500655084848404, + 0.14057469367980957, + 0.04316037893295288, + 0.46125519275665283 + ], + "credit_map": [ + 0.008949005976319313, + 0.012710317969322205, + 0.011353996582329273, + 0.007592404261231422, + 0.009730268269777298, + 0.009906315244734287, + 0.011341110803186893, + 0.010958013124763966, + 0.0032155599910765886, + 0.005209433380514383, + 0.003216571407392621, + 0.004219045862555504, + 0.006125808693468571, + 0.0038962995167821646, + 0.005007435102015734, + 0.0038638231344521046, + 0.003388620214536786, + 0.0032252983655780554, + 0.0033862206619232893, + 0.00495001208037138, + 0.006456824950873852, + 0.004905098117887974, + 0.003958169370889664, + 0.0025563249364495277, + 0.005628182552754879, + 0.004106439650058746, + 0.004599921405315399, + 0.0038666194304823875, + 0.005457452032715082, + 0.01566717028617859, + 0.008588074706494808, + 0.0030178565066307783, + 0.007187255192548037, + 0.0069267903454601765, + 0.0041785091161727905, + 0.004784677177667618, + 0.007468890864402056, + 0.010573280975222588, + 0.00707871001213789, + 0.0029906579293310642, + 0.010902299545705318, + 0.010444315150380135, + 0.0059335012920200825, + 0.005621221382170916, + 0.006822382565587759, + 0.008452894166111946, + 0.004749252460896969, + 0.004834798630326986, + 0.008344138041138649, + 0.010876183398067951, + 0.00511567760258913, + 0.0032058265060186386, + 0.004580655600875616, + 0.004803686868399382, + 0.0034361882135272026, + 0.004644196480512619, + 0.02168172039091587, + 0.022429635748267174, + 0.026145292446017265, + 0.013075527735054493, + 0.011816275306046009, + 0.022074248641729355, + 0.00872740801423788, + 0.014624593779444695, + 0.007099334616214037, + 0.004960933700203896, + 0.005222694482654333, + 0.006463134661316872, + 0.0046892003156244755, + 0.006078910082578659, + 0.0043919626623392105, + 0.004254212137311697, + 0.06710456311702728, + 0.07367027550935745, + 0.015277212485671043, + 0.012319229543209076, + 0.01138325221836567, + 0.1923019140958786, + 0.011252452619373798, + 0.07794628292322159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.003212888026610017 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024952725507318974 + ], + "opposite_logprob_tokens": [ + -6.002495288848877 + ], + "logprob_reward": -0.0024952725507318974, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024952725507318974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0720270499587059, + 0.02034805156290531, + 0.02914857678115368, + 0.05058080330491066, + 0.033665306866168976, + 0.02279093861579895, + 0.025669097900390625, + 0.07327834516763687, + 0.02766183577477932, + 0.6448299884796143 + ], + "credit_map": [ + 0.006928788963705301, + 0.01009669341146946, + 0.008696986362338066, + 0.010093530640006065, + 0.008007660508155823, + 0.008665643632411957, + 0.009449051693081856, + 0.010088693350553513, + 0.0018954275874421, + 0.0032918795477598906, + 0.0019416772993281484, + 0.0023180153220891953, + 0.003905304241925478, + 0.0024700576905161142, + 0.0028618222568184137, + 0.0016638668021187186, + 0.0026307462248951197, + 0.0029684898909181356, + 0.004253813531249762, + 0.0038779983296990395, + 0.007046540733426809, + 0.0025072034914046526, + 0.003174296347424388, + 0.0026894868351519108, + 0.011878052726387978, + 0.0029873645398765802, + 0.002947137923911214, + 0.0023606312461197376, + 0.002782148774713278, + 0.01964571140706539, + 0.005886849947273731, + 0.0020929027814418077, + 0.004561637062579393, + 0.004443359095603228, + 0.0034659348893910646, + 0.003475927049294114, + 0.003967622760683298, + 0.005048280116170645, + 0.004950664006173611, + 0.0037518825847655535, + 0.0035764973144978285, + 0.0030254197772592306, + 0.002614946337416768, + 0.003096470609307289, + 0.004025137983262539, + 0.002463334472849965, + 0.0019413691479712725, + 0.002047765301540494, + 0.004560509696602821, + 0.004246571101248264, + 0.003341453382745385, + 0.0024154777638614178, + 0.003150178352370858, + 0.003364270320162177, + 0.0023555904626846313, + 0.002235047286376357, + 0.010055704973638058, + 0.00937788374722004, + 0.00801172386854887, + 0.004741691052913666, + 0.010614690370857716, + 0.015404979698359966, + 0.006154637783765793, + 0.008917036466300488, + 0.00449382234364748, + 0.0051534781232476234, + 0.003184292232617736, + 0.004323835019022226, + 0.0025015685241669416, + 0.003940952941775322, + 0.002450430067256093, + 0.00161345477681607, + 0.08719225227832794, + 0.08885013312101364, + 0.021791022270917892, + 0.018454881384968758, + 0.015615526586771011, + 0.26513221859931946, + 0.01701413281261921, + 0.13077984750270844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.0024952725507318974 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343291625380516 + ], + "opposite_logprob_tokens": [ + -3.65134334564209 + ], + "logprob_reward": -0.026343291625380516, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343291625380516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12854629755020142, + 0.051530495285987854, + 0.08216962963342667, + 0.13795801997184753, + 0.06140860915184021, + 0.061776455491781235, + 0.05971377342939377, + 0.13303129374980927, + 0.12891198694705963, + 0.15495340526103973 + ], + "credit_map": [ + 0.012696216814219952, + 0.013987457379698753, + 0.01538541354238987, + 0.018659885972738266, + 0.020997650921344757, + 0.020060371607542038, + 0.01560812909156084, + 0.011151176877319813, + 0.004770977422595024, + 0.007908616214990616, + 0.005124722141772509, + 0.009320058859884739, + 0.007364665623754263, + 0.006003322545439005, + 0.007384121883660555, + 0.0036540161818265915, + 0.009787038899958134, + 0.009936866350471973, + 0.0069040232338011265, + 0.010925824753940105, + 0.01995566301047802, + 0.009137443266808987, + 0.009488765150308609, + 0.006034017074853182, + 0.037238381803035736, + 0.01022893376648426, + 0.007455144543200731, + 0.007762649562209845, + 0.008458920754492283, + 0.04682758077979088, + 0.013705635443329811, + 0.00628078356385231, + 0.011764757335186005, + 0.008793224580585957, + 0.005336006172001362, + 0.006661850493401289, + 0.010120906867086887, + 0.007153185550123453, + 0.007251892238855362, + 0.004326793365180492, + 0.009067809209227562, + 0.0071030487306416035, + 0.007077056914567947, + 0.005774879362434149, + 0.01082714181393385, + 0.009369389154016972, + 0.005553517024964094, + 0.007003614213317633, + 0.0072948564775288105, + 0.007919039577245712, + 0.006953636649996042, + 0.006197781767696142, + 0.009142131544649601, + 0.00804720725864172, + 0.008281237445771694, + 0.005877885967493057, + 0.01936991885304451, + 0.014969482086598873, + 0.012776748277246952, + 0.007253079209476709, + 0.024561380967497826, + 0.028002455830574036, + 0.01171838492155075, + 0.014379845932126045, + 0.026936901733279228, + 0.018216831609606743, + 0.011541643179953098, + 0.016013128682971, + 0.009820627048611641, + 0.030684657394886017, + 0.008809325285255909, + 0.0068888855166733265, + 0.017956262454390526, + 0.027231603860855103, + 0.010567308403551579, + 0.0113137261942029, + 0.015692010521888733, + 0.040587496012449265, + 0.017359061166644096, + 0.014245934784412384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 7 + ], + "selected_reward": -0.026343291625380516 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grizzled general wearing heavy, battle-worn steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly etched with battle scars, dents, soot, and dried mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face matted with sweat and grime, with tangled gray hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general shown stretching his mouth wide in a silent, visually arresting roar with a tense jaw and bulging neck veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a chaotic, smoke-filled battlefield with thick gray-black smoke, blurred soldier silhouettes, and scarlet sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626346778124571 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626346778124571, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626346778124571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17721614241600037, + 0.05497090891003609, + 0.08601266890764236, + 0.06921020150184631, + 0.09379050880670547, + 0.07655443251132965, + 0.05380947142839432, + 0.10191964358091354, + 0.12400883436203003, + 0.16250720620155334 + ], + "credit_map": [ + 0.023462029173970222, + 0.027722295373678207, + 0.019792355597019196, + 0.016881346702575684, + 0.023641271516680717, + 0.019654016941785812, + 0.02212376892566681, + 0.023939065635204315, + 0.007721260655671358, + 0.009055971167981625, + 0.004902784712612629, + 0.00494140200316906, + 0.00902187917381525, + 0.006139653734862804, + 0.00823342613875866, + 0.004954530857503414, + 0.01179733220487833, + 0.010170748457312584, + 0.014242680743336678, + 0.012427511624991894, + 0.012562340125441551, + 0.008161651901900768, + 0.009838075377047062, + 0.0068123359233140945, + 0.015259534120559692, + 0.00454592565074563, + 0.005161434877663851, + 0.004165431950241327, + 0.0049383374862372875, + 0.023321790620684624, + 0.007860192097723484, + 0.003957556560635567, + 0.011734707280993462, + 0.010890820994973183, + 0.007623431272804737, + 0.008405404165387154, + 0.012701191939413548, + 0.023686278611421585, + 0.011950762011110783, + 0.00679791159927845, + 0.014451760798692703, + 0.012516987510025501, + 0.008962654508650303, + 0.005374415311962366, + 0.010879989713430405, + 0.012065622955560684, + 0.0060860514640808105, + 0.006216962356120348, + 0.010117017664015293, + 0.008316016755998135, + 0.008838600479066372, + 0.004325490910559893, + 0.005669410340487957, + 0.006068493705242872, + 0.00522996811196208, + 0.0052444771863520145, + 0.01602640375494957, + 0.013108097948133945, + 0.01311689056456089, + 0.009169746190309525, + 0.01167172472923994, + 0.021261019632220268, + 0.006412974093109369, + 0.011152790859341621, + 0.03454297408461571, + 0.01358417235314846, + 0.007776971440762281, + 0.00961901806294918, + 0.007821732200682163, + 0.03601918742060661, + 0.007332319859415293, + 0.007312461268156767, + 0.018080931156873703, + 0.032567400485277176, + 0.010671858675777912, + 0.011660052463412285, + 0.008829536847770214, + 0.016710488125681877, + 0.00842952262610197, + 0.05555741861462593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.007626346778124571 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_10/index_tensor([2413, 2413])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_10/index_tensor([2413, 2413])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6586de65f5559b974c249324edc38f86e3a6a3c8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_10/index_tensor([2413, 2413])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a tight close-up shot, the camera pans steadily to the right, following Sun Wukong, crowned by an intricately carved golden headband, as he soars at high speed atop a swirling cloud. A hint of defiant confidence graces his majestic monkey face, the corners of his mouth curling upwards slightly, while his signature fiery eyes gleam with a sharp, fiery golden light, piercing through all illusions. The wind whipping up from his high-speed flight ruffles the golden fur along his cheeks and sends the red tassels dangling from his headband fluttering violently backward. Below, the edge of his signature golden chain mail armor catches a dazzling glint in the sunlight. Behind him, a vast expanse of azure sky and rolling, indigo-hued mountains, like an ink painting, shrouded in mist and clouds, create an epic and ethereal feeling. The entire picture is bathed in the warm golden sunlight of the afternoon. The light outlines every detail of his silhouette, emphasizing the texture of the metal and the softness of his hair. The overall tone is warm and powerful, perfectly capturing the moment of the Monkey King's fearlessness and high spirits.", + "seed": 14733, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_10/index_2413_seed_14733_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.040526227094233036, + "vlm_logprob": -0.040526227094233036, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.040526227094233036, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019757144153118134 + ], + "opposite_logprob_tokens": [ + -4.519757270812988 + ], + "logprob_reward": -0.019757144153118134, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.019757144153118134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10744684189558029, + 0.09971282631158829, + 0.11878301948308945, + 0.06217046082019806, + 0.1544884443283081, + 0.07107478380203247, + 0.05956569314002991, + 0.11843414604663849, + 0.10667774081230164, + 0.1016460731625557 + ], + "credit_map": [ + 0.008417836390435696, + 0.011178996413946152, + 0.02261102944612503, + 0.02470305562019348, + 0.013233265839517117, + 0.011091100051999092, + 0.008370448835194111, + 0.007841106504201889, + 0.017571181058883667, + 0.023106107488274574, + 0.014513020403683186, + 0.013094942085444927, + 0.01125834695994854, + 0.005885886028409004, + 0.008627931587398052, + 0.00565541023388505, + 0.02090311050415039, + 0.010202916339039803, + 0.04087916761636734, + 0.02071918547153473, + 0.007176592946052551, + 0.005841770209372044, + 0.007465683855116367, + 0.005594593472778797, + 0.010890373960137367, + 0.009983262047171593, + 0.00893313903361559, + 0.005374810192734003, + 0.007481744978576899, + 0.005039420444518328, + 0.010054869577288628, + 0.004412839189171791, + 0.012204206548631191, + 0.018530353903770447, + 0.018213845789432526, + 0.018124965950846672, + 0.024394161999225616, + 0.02076711878180504, + 0.023596415296196938, + 0.018657365813851357, + 0.008612611331045628, + 0.013502327725291252, + 0.007340509444475174, + 0.007718736305832863, + 0.010363656096160412, + 0.006227463949471712, + 0.0077671161852777, + 0.009542363695800304, + 0.006797098554670811, + 0.007877097465097904, + 0.006428905297070742, + 0.008609763346612453, + 0.00800279900431633, + 0.00769420899450779, + 0.006057581398636103, + 0.008098234422504902, + 0.015417748130857944, + 0.01994730718433857, + 0.022746318951249123, + 0.013104327023029327, + 0.010663987137377262, + 0.013476107269525528, + 0.012884107418358326, + 0.010194230824708939, + 0.011123846285045147, + 0.0195892583578825, + 0.009465407580137253, + 0.014568120241165161, + 0.013639109209179878, + 0.013910951092839241, + 0.012989684008061886, + 0.011391359381377697, + 0.009035812690854073, + 0.020749960094690323, + 0.009682831354439259, + 0.011814610101282597, + 0.0075651598162949085, + 0.009730366058647633, + 0.02035989612340927, + 0.012707438319921494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.019757144153118134 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16041196882724762 + ], + "opposite_logprob_tokens": [ + -1.9104119539260864 + ], + "logprob_reward": -0.16041196882724762, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16041196882724762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10583694279193878, + 0.05884122848510742, + 0.11664776504039764, + 0.0518362857401371, + 0.08212262392044067, + 0.10498209297657013, + 0.10099296271800995, + 0.21693764626979828, + 0.13589458167552948, + 0.025907907634973526 + ], + "credit_map": [ + 0.020345384255051613, + 0.01354172732681036, + 0.01787170022726059, + 0.016596613451838493, + 0.008697877638041973, + 0.011097708716988564, + 0.00923354271799326, + 0.008452391251921654, + 0.009395700879395008, + 0.007962582632899284, + 0.009270211681723595, + 0.006245636846870184, + 0.007451262325048447, + 0.006942506413906813, + 0.0075712925754487514, + 0.004002033267170191, + 0.014368070289492607, + 0.020891809836030006, + 0.021986404433846474, + 0.019399594515562057, + 0.01211497187614441, + 0.008182761259377003, + 0.013683679513633251, + 0.006020471453666687, + 0.008491195738315582, + 0.01070513017475605, + 0.006587684154510498, + 0.004008367192000151, + 0.006130005232989788, + 0.00406178692355752, + 0.008414458483457565, + 0.003437658539041877, + 0.008688228204846382, + 0.0059684826992452145, + 0.017965849488973618, + 0.006355594377964735, + 0.007677098270505667, + 0.011159233748912811, + 0.017621897161006927, + 0.006686241831630468, + 0.008557623252272606, + 0.017914941534399986, + 0.011257849633693695, + 0.010980469174683094, + 0.016118643805384636, + 0.009990986436605453, + 0.015738537535071373, + 0.014423051849007607, + 0.011572821997106075, + 0.015598542056977749, + 0.011555622331798077, + 0.01781672053039074, + 0.013777109794318676, + 0.009667211212217808, + 0.00892624445259571, + 0.01207868754863739, + 0.03142249956727028, + 0.027806686237454414, + 0.030479876324534416, + 0.020149946212768555, + 0.01867128163576126, + 0.03172401711344719, + 0.028432974591851234, + 0.02825036086142063, + 0.010957412421703339, + 0.017936434596776962, + 0.019314967095851898, + 0.013097669929265976, + 0.02372417226433754, + 0.02287406474351883, + 0.014007623307406902, + 0.013982237316668034, + 0.002658550161868334, + 0.0045812190510332584, + 0.002186393365263939, + 0.002820083172991872, + 0.001957674277946353, + 0.0030207326635718346, + 0.004850981757044792, + 0.003832275280728936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.16041196882724762 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004650728777050972 + ], + "opposite_logprob_tokens": [ + -5.379650592803955 + ], + "logprob_reward": -0.004650728777050972, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004650728777050972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056785520166158676, + 0.04752247408032417, + 0.22856025397777557, + 0.050097886472940445, + 0.03364560380578041, + 0.09472653269767761, + 0.04460020735859871, + 0.17834877967834473, + 0.2197222262620926, + 0.045990537852048874 + ], + "credit_map": [ + 0.00859958678483963, + 0.007354883477091789, + 0.010333863086998463, + 0.01107014063745737, + 0.0049613420851528645, + 0.0050708758644759655, + 0.0049330187030136585, + 0.0044618044048547745, + 0.006509553641080856, + 0.007318382617086172, + 0.006465552374720573, + 0.004831777885556221, + 0.006464631296694279, + 0.004648621194064617, + 0.006339548621326685, + 0.0049444022588431835, + 0.03663172572851181, + 0.020825428888201714, + 0.07844654470682144, + 0.05148596689105034, + 0.016246814280748367, + 0.008243905380368233, + 0.010301959700882435, + 0.006377886515110731, + 0.006720897741615772, + 0.008836932480335236, + 0.0061046648770570755, + 0.004988280590623617, + 0.007635326590389013, + 0.006092107854783535, + 0.00536621967330575, + 0.00435345433652401, + 0.005064365919679403, + 0.0043823011219501495, + 0.00382052781060338, + 0.006129460409283638, + 0.003140117973089218, + 0.0028181758243590593, + 0.004932428244501352, + 0.0033582253381609917, + 0.00818970613181591, + 0.01448665652424097, + 0.009527123533189297, + 0.010559272021055222, + 0.014069482684135437, + 0.00851774588227272, + 0.013229185715317726, + 0.01614736020565033, + 0.006882159039378166, + 0.005315718241035938, + 0.008727512322366238, + 0.005938822403550148, + 0.004873110447078943, + 0.0053272549994289875, + 0.00378149701282382, + 0.003754132194444537, + 0.023079993203282356, + 0.018688183277845383, + 0.036571621894836426, + 0.01841903105378151, + 0.016955792903900146, + 0.019373055547475815, + 0.018397076055407524, + 0.026864001527428627, + 0.019583512097597122, + 0.024466555565595627, + 0.03659071400761604, + 0.020704127848148346, + 0.04387916624546051, + 0.031059078872203827, + 0.02209741249680519, + 0.021341657266020775, + 0.003599165938794613, + 0.008952035568654537, + 0.005287120118737221, + 0.005352575331926346, + 0.004368616733700037, + 0.004573584999889135, + 0.006461538840085268, + 0.007395896594971418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.004650728777050972 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003616699017584324 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.003616699017584324, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616699017584324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11695397645235062, + 0.05134427174925804, + 0.4112697243690491, + 0.05369152873754501, + 0.07726159691810608, + 0.07473039627075195, + 0.04695321246981621, + 0.07090532779693604, + 0.062255579978227615, + 0.034634318202733994 + ], + "credit_map": [ + 0.016233127564191818, + 0.013490281067788601, + 0.02218586392700672, + 0.02269984968006611, + 0.011886047199368477, + 0.01082264818251133, + 0.010793965309858322, + 0.008842205628752708, + 0.008749785833060741, + 0.01132022961974144, + 0.00803919043391943, + 0.007556774653494358, + 0.005598270334303379, + 0.0031116604804992676, + 0.0040642558597028255, + 0.002904105931520462, + 0.08757422119379044, + 0.038017965853214264, + 0.15338942408561707, + 0.07160396128892899, + 0.024272466078400612, + 0.011861221864819527, + 0.012635672464966774, + 0.01191481202840805, + 0.008029566146433353, + 0.011904459446668625, + 0.005962171591818333, + 0.004194248467683792, + 0.007169801741838455, + 0.0041848705150187016, + 0.008935099467635155, + 0.003311315318569541, + 0.01385833602398634, + 0.00705223623663187, + 0.011347098276019096, + 0.013281542807817459, + 0.008745209313929081, + 0.00850065890699625, + 0.007958358153700829, + 0.006518165115267038, + 0.010884693823754787, + 0.008382605388760567, + 0.012217218987643719, + 0.008385393768548965, + 0.007037909235805273, + 0.01154826395213604, + 0.010039872489869595, + 0.006234447006136179, + 0.008668213151395321, + 0.005403163842856884, + 0.009998736903071404, + 0.00518482830375433, + 0.00542058888822794, + 0.005259099882096052, + 0.0034781231079250574, + 0.0035404653754085302, + 0.007627997547388077, + 0.007256245706230402, + 0.010960384272038937, + 0.009396365843713284, + 0.009120889008045197, + 0.012250757776200771, + 0.007014872040599585, + 0.007277820724993944, + 0.006578308995813131, + 0.009477849118411541, + 0.007150732912123203, + 0.00592963770031929, + 0.011739984154701233, + 0.00912492722272873, + 0.00667774211615324, + 0.005576406139880419, + 0.0034207813441753387, + 0.006286292336881161, + 0.003832746995612979, + 0.0042639318853616714, + 0.00354205840267241, + 0.004196716472506523, + 0.0037096792366355658, + 0.005382112227380276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.003616699017584324 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194594696164131 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194594696164131, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194594696164131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13775493204593658, + 0.07728967815637589, + 0.2152741402387619, + 0.059993550181388855, + 0.03109630197286606, + 0.06219048053026199, + 0.08827730268239975, + 0.11065804958343506, + 0.05650026723742485, + 0.16096529364585876 + ], + "credit_map": [ + 0.014121904037892818, + 0.014733952470123768, + 0.03292088210582733, + 0.04375964775681496, + 0.010742573998868465, + 0.007610465865582228, + 0.007800472434610128, + 0.006065035704523325, + 0.012484160251915455, + 0.017748994752764702, + 0.012341829016804695, + 0.011408007703721523, + 0.008214599452912807, + 0.00426926976069808, + 0.007072731852531433, + 0.0037500839680433273, + 0.041721027344465256, + 0.031420160084962845, + 0.07213514298200607, + 0.030290471389889717, + 0.012617964297533035, + 0.007846536114811897, + 0.011980217881500721, + 0.007262609899044037, + 0.009215776808559895, + 0.01275993138551712, + 0.008663765154778957, + 0.005929897073656321, + 0.00818155612796545, + 0.00472429720684886, + 0.006998924072831869, + 0.003519403049722314, + 0.0032751192338764668, + 0.0031435240525752306, + 0.005070537328720093, + 0.003765571629628539, + 0.0050029512494802475, + 0.0024141212925314903, + 0.00576403271406889, + 0.002660444239154458, + 0.010708415880799294, + 0.007147468160837889, + 0.007461757864803076, + 0.00799896102398634, + 0.0066644782200455666, + 0.01007675938308239, + 0.006539584137499332, + 0.005593057721853256, + 0.011692697182297707, + 0.013700399547815323, + 0.017162634059786797, + 0.00929854717105627, + 0.012812941335141659, + 0.007945931516587734, + 0.007915669120848179, + 0.007748478092253208, + 0.011951116845011711, + 0.0227934792637825, + 0.011510084383189678, + 0.01743261143565178, + 0.01276436448097229, + 0.00880496110767126, + 0.013311095535755157, + 0.012090341188013554, + 0.005136240739375353, + 0.007394755259156227, + 0.009112321771681309, + 0.0055427406914532185, + 0.011198434978723526, + 0.004786927253007889, + 0.005785565823316574, + 0.007543276995420456, + 0.011872075498104095, + 0.033072445541620255, + 0.017904313281178474, + 0.017797181382775307, + 0.011715707369148731, + 0.01564115472137928, + 0.029432345181703568, + 0.023530080914497375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.014194594696164131 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a tight close-up shot, the camera pans steadily to the right, following Sun Wukong, crowned by an intricately carved golden headband, as he soars at high speed atop a swirling cloud. A hint of defiant confidence graces his majestic monkey face, the corners of his mouth curling upwards slightly, while his signature fiery eyes gleam with a sharp, fiery golden light, piercing through all illusions. The wind whipping up from his high-speed flight ruffles the golden fur along his cheeks and sends the red tassels dangling from his headband fluttering violently backward. Below, the edge of his signature golden chain mail armor catches a dazzling glint in the sunlight. Behind him, a vast expanse of azure sky and rolling, indigo-hued mountains, like an ink painting, shrouded in mist and clouds, create an epic and ethereal feeling. The entire picture is bathed in the warm golden sunlight of the afternoon. The light outlines every detail of his silhouette, emphasizing the texture of the metal and the softness of his hair. The overall tone is warm and powerful, perfectly capturing the moment of the Monkey King's fearlessness and high spirits.", + "seed": 14734, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_10/index_2413_seed_14734_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09208563375286757, + "vlm_logprob": -0.09208563375286757, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.09208563375286757, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012905645184218884 + ], + "opposite_logprob_tokens": [ + -5.262905597686768 + ], + "logprob_reward": -0.012905645184218884, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.012905645184218884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13079895079135895, + 0.08902978152036667, + 0.05641256645321846, + 0.10849770158529282, + 0.15345004200935364, + 0.08398199826478958, + 0.07906949520111084, + 0.05472182109951973, + 0.12472545355558395, + 0.11931223422288895 + ], + "credit_map": [ + 0.0212604608386755, + 0.016287943348288536, + 0.01767413318157196, + 0.014174575917422771, + 0.01580214872956276, + 0.013259254395961761, + 0.019253388047218323, + 0.013087046332657337, + 0.012163190171122551, + 0.010506579652428627, + 0.010391920804977417, + 0.007151713129132986, + 0.01243071723729372, + 0.009472857229411602, + 0.01846316084265709, + 0.008449630811810493, + 0.007541719824075699, + 0.008708848617970943, + 0.017235033214092255, + 0.004662890452891588, + 0.004730572924017906, + 0.003631612053140998, + 0.003885034006088972, + 0.006016853731125593, + 0.010227220132946968, + 0.011518942192196846, + 0.024264948442578316, + 0.010250409133732319, + 0.01737137883901596, + 0.010659289546310902, + 0.01102732215076685, + 0.013178182765841484, + 0.013638043776154518, + 0.019085271283984184, + 0.014467276632785797, + 0.021548066288232803, + 0.02290971949696541, + 0.01771136373281479, + 0.01827957108616829, + 0.025810718536376953, + 0.0063486904837191105, + 0.009699678048491478, + 0.006010178476572037, + 0.03193036466836929, + 0.004210538696497679, + 0.0058930376544594765, + 0.011939845979213715, + 0.00794965960085392, + 0.015447513200342655, + 0.0044051846489310265, + 0.025401469320058823, + 0.006031087599694729, + 0.004195617511868477, + 0.005223081912845373, + 0.011087960563600063, + 0.00727757066488266, + 0.00999361090362072, + 0.005454475525766611, + 0.00516773946583271, + 0.006085682660341263, + 0.00445721996948123, + 0.005983690731227398, + 0.006190712098032236, + 0.0113886883482337, + 0.017945609986782074, + 0.011434350162744522, + 0.012619259767234325, + 0.03984549269080162, + 0.013022377155721188, + 0.005797330755740404, + 0.014193113893270493, + 0.009867911227047443, + 0.00898769125342369, + 0.014672142453491688, + 0.015460261143743992, + 0.03329528495669365, + 0.020871227607131004, + 0.006024166941642761, + 0.010893809609115124, + 0.009107636287808418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.012905645184218884 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42898619174957275 + ], + "opposite_logprob_tokens": [ + -1.0539861917495728 + ], + "logprob_reward": -0.42898619174957275, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42898619174957275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12480248510837555, + 0.11102870106697083, + 0.07442829757928848, + 0.0397987961769104, + 0.039858151227235794, + 0.2149888277053833, + 0.17742331326007843, + 0.05976912006735802, + 0.10727248340845108, + 0.05062982067465782 + ], + "credit_map": [ + 0.017136409878730774, + 0.011409993283450603, + 0.016777928918600082, + 0.011277838610112667, + 0.01460955012589693, + 0.011058507487177849, + 0.03132225200533867, + 0.011210007593035698, + 0.01635308749973774, + 0.010133435018360615, + 0.012155719101428986, + 0.008985313586890697, + 0.011429142206907272, + 0.011416408233344555, + 0.027290020138025284, + 0.013265583664178848, + 0.007503250148147345, + 0.00753282755613327, + 0.015434985049068928, + 0.007416185922920704, + 0.007377469912171364, + 0.007847647182643414, + 0.012126758694648743, + 0.009189171716570854, + 0.00407621543854475, + 0.003864976577460766, + 0.007279526442289352, + 0.004707339685410261, + 0.005040719639509916, + 0.003434708807617426, + 0.006537471432238817, + 0.004857835825532675, + 0.0038237774278968573, + 0.005574097391217947, + 0.006169191561639309, + 0.005871399771422148, + 0.003148815594613552, + 0.004943932872265577, + 0.004764634184539318, + 0.005562303587794304, + 0.016736343502998352, + 0.026759706437587738, + 0.010286088101565838, + 0.083420529961586, + 0.009765980765223503, + 0.011838343925774097, + 0.03646925464272499, + 0.019712582230567932, + 0.03849007189273834, + 0.008447467349469662, + 0.06598816812038422, + 0.010979401879012585, + 0.008491259068250656, + 0.006728942971676588, + 0.028029516339302063, + 0.010268484242260456, + 0.005092149134725332, + 0.008325576782226562, + 0.006410602480173111, + 0.0074825286865234375, + 0.0059113879688084126, + 0.0070322067476809025, + 0.005578112322837114, + 0.013936555944383144, + 0.009435662068426609, + 0.018484920263290405, + 0.008594974875450134, + 0.013914629817008972, + 0.017530769109725952, + 0.010510128922760487, + 0.011924050748348236, + 0.016877343878149986, + 0.005042646080255508, + 0.006866232957690954, + 0.006819119676947594, + 0.009582029655575752, + 0.008261794224381447, + 0.003446451388299465, + 0.005783603992313147, + 0.0048279426991939545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.42898619174957275 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028272203635424376 + ], + "opposite_logprob_tokens": [ + -5.877827167510986 + ], + "logprob_reward": -0.0028272203635424376, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028272203635424376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07392758876085281, + 0.06520528346300125, + 0.08338667452335358, + 0.06435994058847427, + 0.12268900126218796, + 0.23532846570014954, + 0.23875202238559723, + 0.06295837461948395, + 0.023475725203752518, + 0.0299169160425663 + ], + "credit_map": [ + 0.010605642572045326, + 0.00787835568189621, + 0.010359363630414009, + 0.009680393151938915, + 0.00789360050112009, + 0.0060165682807564735, + 0.012429728172719479, + 0.009063933975994587, + 0.007246030028909445, + 0.005731791723519564, + 0.01109133567661047, + 0.006060891319066286, + 0.008850940503180027, + 0.007648078259080648, + 0.010829049162566662, + 0.0077471681870520115, + 0.008776526898145676, + 0.012420808896422386, + 0.014089399948716164, + 0.008062812499701977, + 0.009123608469963074, + 0.00980375800281763, + 0.013093383982777596, + 0.008016371168196201, + 0.0076616001315414906, + 0.008750523440539837, + 0.014314203523099422, + 0.00651873042806983, + 0.0064649423584342, + 0.004447109531611204, + 0.007591766770929098, + 0.008611060678958893, + 0.010974097065627575, + 0.01551548670977354, + 0.014467135071754456, + 0.01764303259551525, + 0.01378648355603218, + 0.018342314288020134, + 0.012545148842036724, + 0.01941531151533127, + 0.018029114231467247, + 0.024218907579779625, + 0.014147822745144367, + 0.08634001761674881, + 0.010886118747293949, + 0.022296685725450516, + 0.03785085678100586, + 0.021558938547968864, + 0.04510273039340973, + 0.013221974484622478, + 0.07821270078420639, + 0.015186306089162827, + 0.014387738890945911, + 0.015169182792305946, + 0.04244114086031914, + 0.015030239708721638, + 0.004750032443553209, + 0.004642596933990717, + 0.004719663877040148, + 0.021941272541880608, + 0.0037155586760491133, + 0.0034408217761665583, + 0.007715384475886822, + 0.012033043429255486, + 0.0022231866605579853, + 0.003850152250379324, + 0.0017162678996101022, + 0.0031150924041867256, + 0.0033408552408218384, + 0.0027936992701143026, + 0.0026894041802734137, + 0.003747067181393504, + 0.0029996808152645826, + 0.0030798076186329126, + 0.004306848160922527, + 0.006596899125725031, + 0.003785210195928812, + 0.0023304889909923077, + 0.0036065189633518457, + 0.0032114640343934298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.0028272203635424376 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636845551431179 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636845551431179, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636845551431179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15339143574237823, + 0.13339902460575104, + 0.26089736819267273, + 0.04144005849957466, + 0.06975385546684265, + 0.10231198370456696, + 0.05816788598895073, + 0.09083949774503708, + 0.052043937146663666, + 0.03775491937994957 + ], + "credit_map": [ + 0.026668688282370567, + 0.025769991800189018, + 0.027566123753786087, + 0.020294826477766037, + 0.014622943475842476, + 0.01101592369377613, + 0.013605543412268162, + 0.01384738925844431, + 0.019225969910621643, + 0.020364979282021523, + 0.021573269739747047, + 0.010978790000081062, + 0.013506283983588219, + 0.01210884191095829, + 0.023423556238412857, + 0.012217334471642971, + 0.03765864297747612, + 0.051441170275211334, + 0.07007914036512375, + 0.021301813423633575, + 0.023371705785393715, + 0.015652213245630264, + 0.017383787781000137, + 0.024008896201848984, + 0.0043997070752084255, + 0.005375855602324009, + 0.0094298729673028, + 0.00410728994756937, + 0.004222288727760315, + 0.0032378067262470722, + 0.004928493872284889, + 0.005738743115216494, + 0.00703446427360177, + 0.008906664326786995, + 0.01639292761683464, + 0.010537667199969292, + 0.008285957388579845, + 0.004738687537610531, + 0.0069233737885952, + 0.006934110075235367, + 0.007903940975666046, + 0.01465523336082697, + 0.0077892751432955265, + 0.025177501142024994, + 0.008091187104582787, + 0.01059237401932478, + 0.017780674621462822, + 0.010321795009076595, + 0.012957824394106865, + 0.0025660647079348564, + 0.020596839487552643, + 0.0034482458140701056, + 0.00331807229667902, + 0.003107881173491478, + 0.008538832888007164, + 0.003634123830124736, + 0.0176794845610857, + 0.006915819365531206, + 0.010486829094588757, + 0.015307953581213951, + 0.006816173903644085, + 0.006739148870110512, + 0.011654488742351532, + 0.015239597298204899, + 0.009122475050389767, + 0.006204333156347275, + 0.005031163338571787, + 0.015105806291103363, + 0.005935291759669781, + 0.0023278528824448586, + 0.004956129938364029, + 0.0033608858939260244, + 0.003919358365237713, + 0.004658127203583717, + 0.003618457820266485, + 0.007706339005380869, + 0.005508200265467167, + 0.0022544509265571833, + 0.005847081076353788, + 0.004242904018610716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.004636845551431179 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072265915572643 + ], + "opposite_logprob_tokens": [ + -4.511072158813477 + ], + "logprob_reward": -0.011072265915572643, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072265915572643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16671662032604218, + 0.0973106324672699, + 0.07100546360015869, + 0.06777475774288177, + 0.07209160178899765, + 0.11244937032461166, + 0.07605675607919693, + 0.13077320158481598, + 0.047796186059713364, + 0.15802541375160217 + ], + "credit_map": [ + 0.02286270633339882, + 0.023849140852689743, + 0.027243372052907944, + 0.02414625510573387, + 0.016292719170451164, + 0.015220873057842255, + 0.018257085233926773, + 0.018844474107027054, + 0.010007956065237522, + 0.009517154656350613, + 0.00985668320208788, + 0.012565117329359055, + 0.011902141384780407, + 0.014630370773375034, + 0.01514917891472578, + 0.01368202455341816, + 0.009398712776601315, + 0.006965144071727991, + 0.018836522474884987, + 0.005847932770848274, + 0.0067112925462424755, + 0.005539562087506056, + 0.009407348930835724, + 0.008298944681882858, + 0.007215368561446667, + 0.007488054223358631, + 0.015541259199380875, + 0.007409683894366026, + 0.006479841191321611, + 0.003838589880615473, + 0.010994709096848965, + 0.008807250298559666, + 0.006449396256357431, + 0.011865070089697838, + 0.007051215972751379, + 0.012803468853235245, + 0.006862280890345573, + 0.00810573436319828, + 0.006711198948323727, + 0.012243234552443027, + 0.008434310555458069, + 0.014468967914581299, + 0.006066807080060244, + 0.035201169550418854, + 0.006565957795828581, + 0.009364543482661247, + 0.020816504955291748, + 0.011531113646924496, + 0.017376093193888664, + 0.004350143019109964, + 0.024889178574085236, + 0.005515930708497763, + 0.004787960555404425, + 0.007123611867427826, + 0.005910827312618494, + 0.006103008985519409, + 0.011870736256241798, + 0.010319128632545471, + 0.007106750272214413, + 0.048521626740694046, + 0.005655964370816946, + 0.005785339046269655, + 0.015463451854884624, + 0.026050202548503876, + 0.005696077831089497, + 0.007459499407559633, + 0.0053132809698581696, + 0.0063963113352656364, + 0.007870389148592949, + 0.004467104095965624, + 0.0046212924644351006, + 0.005972232669591904, + 0.0125908637419343, + 0.01380163710564375, + 0.02343754470348358, + 0.038971591740846634, + 0.018100852146744728, + 0.01630633883178234, + 0.014324798248708248, + 0.02049177698791027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.011072265915572643 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a tight close-up shot, the camera pans steadily to the right, following Sun Wukong, crowned by an intricately carved golden headband, as he soars at high speed atop a swirling cloud. A hint of defiant confidence graces his majestic monkey face, the corners of his mouth curling upwards slightly, while his signature fiery eyes gleam with a sharp, fiery golden light, piercing through all illusions. The wind whipping up from his high-speed flight ruffles the golden fur along his cheeks and sends the red tassels dangling from his headband fluttering violently backward. Below, the edge of his signature golden chain mail armor catches a dazzling glint in the sunlight. Behind him, a vast expanse of azure sky and rolling, indigo-hued mountains, like an ink painting, shrouded in mist and clouds, create an epic and ethereal feeling. The entire picture is bathed in the warm golden sunlight of the afternoon. The light outlines every detail of his silhouette, emphasizing the texture of the metal and the softness of his hair. The overall tone is warm and powerful, perfectly capturing the moment of the Monkey King's fearlessness and high spirits.", + "seed": 14735, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_10/index_2413_seed_14735_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03937111208215356, + "vlm_logprob": -0.03937111208215356, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.03937111208215356, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01619364693760872 + ], + "opposite_logprob_tokens": [ + -4.391193866729736 + ], + "logprob_reward": -0.01619364693760872, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01619364693760872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10155824571847916, + 0.052527036517858505, + 0.05843851715326309, + 0.034411221742630005, + 0.11159868538379669, + 0.05637931078672409, + 0.045886170119047165, + 0.14401528239250183, + 0.17102673649787903, + 0.2241588681936264 + ], + "credit_map": [ + 0.010303640738129616, + 0.01144885178655386, + 0.010469126515090466, + 0.006770734209567308, + 0.015882259234786034, + 0.012659840285778046, + 0.01853914186358452, + 0.015484641306102276, + 0.005768805276602507, + 0.005121547728776932, + 0.006737555842846632, + 0.008011025376617908, + 0.006576444022357464, + 0.0065868208184838295, + 0.006979287136346102, + 0.006745546590536833, + 0.006598672363907099, + 0.006528018042445183, + 0.007783768232911825, + 0.01166579220443964, + 0.009957414120435715, + 0.005420070141553879, + 0.0064806039445102215, + 0.004004173446446657, + 0.003736418206244707, + 0.004351435229182243, + 0.0031863688491284847, + 0.0037704117130488157, + 0.0056846048682928085, + 0.003828079905360937, + 0.007292966824024916, + 0.002560935216024518, + 0.015237532556056976, + 0.019012395292520523, + 0.00646202452480793, + 0.011614333838224411, + 0.02044404298067093, + 0.01353849284350872, + 0.01740402542054653, + 0.007885828614234924, + 0.006849611643701792, + 0.008166346698999405, + 0.008790523745119572, + 0.006735837087035179, + 0.006839404348284006, + 0.008255311287939548, + 0.006520899012684822, + 0.004221371375024319, + 0.006084890570491552, + 0.008563236333429813, + 0.004943210165947676, + 0.0048830402083694935, + 0.005901633761823177, + 0.005857426207512617, + 0.004248323384672403, + 0.00540440808981657, + 0.022966325283050537, + 0.027355585247278214, + 0.016239460557699203, + 0.017147723585367203, + 0.01747824065387249, + 0.015242625027894974, + 0.009911197237670422, + 0.017674116417765617, + 0.021348627284169197, + 0.031422991305589676, + 0.018062802031636238, + 0.025856206193566322, + 0.017557833343744278, + 0.012998647056519985, + 0.01769060641527176, + 0.02608899027109146, + 0.020911550149321556, + 0.03939331695437431, + 0.024291228502988815, + 0.021904677152633667, + 0.022070206701755524, + 0.021203158423304558, + 0.022839466109871864, + 0.051545239984989166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.01619364693760872 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604192852973938 + ], + "opposite_logprob_tokens": [ + -1.910419225692749 + ], + "logprob_reward": -0.1604192852973938, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604192852973938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15531136095523834, + 0.0454988069832325, + 0.1429748684167862, + 0.07505053281784058, + 0.09722627699375153, + 0.08996381610631943, + 0.06279963999986649, + 0.09317194670438766, + 0.14314621686935425, + 0.09485657513141632 + ], + "credit_map": [ + 0.012236173264682293, + 0.023796187713742256, + 0.013773244805634022, + 0.011890031397342682, + 0.02963382937014103, + 0.023711584508419037, + 0.019428489729762077, + 0.020841818302869797, + 0.005430612713098526, + 0.005726109724491835, + 0.004347577691078186, + 0.005895925220102072, + 0.005711883306503296, + 0.008233671076595783, + 0.005530238151550293, + 0.004622787702828646, + 0.026118073612451553, + 0.01787833869457245, + 0.02082109823822975, + 0.020488325506448746, + 0.01337791234254837, + 0.012878043577075005, + 0.017466552555561066, + 0.013946518301963806, + 0.008726219646632671, + 0.008236602880060673, + 0.008733890950679779, + 0.01181381568312645, + 0.007439109962433577, + 0.0075217559933662415, + 0.01610444486141205, + 0.006474689580500126, + 0.01559279765933752, + 0.020789196714758873, + 0.0072906711138784885, + 0.014604700729250908, + 0.013544115237891674, + 0.008102028630673885, + 0.009789438918232918, + 0.007513329386711121, + 0.012160932645201683, + 0.017903344705700874, + 0.011709408834576607, + 0.010511309839785099, + 0.0101209357380867, + 0.008554945699870586, + 0.01181397307664156, + 0.007188962772488594, + 0.011026035994291306, + 0.015081322751939297, + 0.006557947490364313, + 0.00626115407794714, + 0.005956506822258234, + 0.006996068172156811, + 0.004732012748718262, + 0.0061885942704975605, + 0.012712376192212105, + 0.013656854629516602, + 0.011714774183928967, + 0.011376641690731049, + 0.009608989581465721, + 0.016352472826838493, + 0.006330153439193964, + 0.011419685557484627, + 0.013698498718440533, + 0.023374896496534348, + 0.017942791804671288, + 0.019100632518529892, + 0.01593388430774212, + 0.014005930162966251, + 0.013853300362825394, + 0.025236284360289574, + 0.010240796022117138, + 0.01843651384115219, + 0.014651619829237461, + 0.00822215061634779, + 0.009177612140774727, + 0.006374355871230364, + 0.013111956417560577, + 0.014641574583947659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.1604192852973938 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007641489617526531 + ], + "opposite_logprob_tokens": [ + -4.882641315460205 + ], + "logprob_reward": -0.007641489617526531, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007641489617526531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11316569149494171, + 0.0698661059141159, + 0.06799795478582382, + 0.1233019232749939, + 0.1581718474626541, + 0.10577072948217392, + 0.07503703981637955, + 0.05291811749339104, + 0.07968416810035706, + 0.15408648550510406 + ], + "credit_map": [ + 0.011850330047309399, + 0.01635085418820381, + 0.013588037341833115, + 0.007972780615091324, + 0.018114415928721428, + 0.012915930710732937, + 0.01860208250582218, + 0.013771245256066322, + 0.0061496845446527, + 0.006844099145382643, + 0.006088487338274717, + 0.007388857658952475, + 0.008680076338350773, + 0.014036580920219421, + 0.007671562489122152, + 0.013006742112338543, + 0.0059510646387934685, + 0.006630093324929476, + 0.004803987219929695, + 0.007390649989247322, + 0.013127753511071205, + 0.010431181639432907, + 0.01003681868314743, + 0.009626401588320732, + 0.008761941455304623, + 0.010441424325108528, + 0.006111170630902052, + 0.012971147894859314, + 0.018790869042277336, + 0.014517546631395817, + 0.04146215692162514, + 0.010245651006698608, + 0.025603631511330605, + 0.02726169489324093, + 0.0073039042763412, + 0.016479337587952614, + 0.023839356377720833, + 0.01887083798646927, + 0.02718195877969265, + 0.011631111614406109, + 0.01481139101088047, + 0.01815129444003105, + 0.014891412109136581, + 0.013794878497719765, + 0.011190085671842098, + 0.01129197794944048, + 0.013503043912351131, + 0.008136636577546597, + 0.010041266679763794, + 0.012255150824785233, + 0.010025386698544025, + 0.006625522393733263, + 0.008925319649279118, + 0.009699052199721336, + 0.006711518857628107, + 0.010753806680440903, + 0.007165560033172369, + 0.008914451114833355, + 0.005805228836834431, + 0.0042136129923164845, + 0.005753465928137302, + 0.010945403017103672, + 0.003888953709974885, + 0.006231436040252447, + 0.006999981589615345, + 0.010645528323948383, + 0.009732628241181374, + 0.008733163587749004, + 0.007971327751874924, + 0.007988957688212395, + 0.00796421803534031, + 0.01964835822582245, + 0.017265567556023598, + 0.031878817826509476, + 0.02636900544166565, + 0.013775000348687172, + 0.014191779308021069, + 0.011170883662998676, + 0.013338192366063595, + 0.02609722875058651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 3 + ], + "selected_reward": -0.007641489617526531 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002819731365889311 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.002819731365889311, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002819731365889311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2958506643772125, + 0.08694373071193695, + 0.10434595495462418, + 0.0319870188832283, + 0.0561225488781929, + 0.025095390155911446, + 0.09169149398803711, + 0.17350956797599792, + 0.09400556981563568, + 0.04044802486896515 + ], + "credit_map": [ + 0.02983758971095085, + 0.03406144306063652, + 0.03455238416790962, + 0.025158466771245003, + 0.04721904173493385, + 0.03930119797587395, + 0.04700176417827606, + 0.038718800991773605, + 0.00819143746048212, + 0.010345187969505787, + 0.009958158247172832, + 0.014423120766878128, + 0.01268361508846283, + 0.011963726952672005, + 0.009608287364244461, + 0.00977020338177681, + 0.009142155759036541, + 0.011629081331193447, + 0.006401333026587963, + 0.010424227453768253, + 0.02432899735867977, + 0.019322469830513, + 0.013206114992499352, + 0.009891582652926445, + 0.003618486924096942, + 0.003689521923661232, + 0.0035554508212953806, + 0.004853853490203619, + 0.006933312863111496, + 0.0034202139358967543, + 0.003949675243347883, + 0.0019665067084133625, + 0.00670338049530983, + 0.009364243596792221, + 0.003835561452433467, + 0.008818753063678741, + 0.010572017170488834, + 0.004218089394271374, + 0.00808537658303976, + 0.004525130148977041, + 0.003730705240741372, + 0.006377764977514744, + 0.0029283035546541214, + 0.002568024443462491, + 0.002786321332678199, + 0.0020799990743398666, + 0.002953068818897009, + 0.0016712063224986196, + 0.014185072854161263, + 0.013575547374784946, + 0.00991765595972538, + 0.00825507566332817, + 0.007695370353758335, + 0.0116852056235075, + 0.010006800293922424, + 0.016370762139558792, + 0.025813069194555283, + 0.032586246728897095, + 0.016927070915699005, + 0.01555183157324791, + 0.025719938799738884, + 0.028136370703577995, + 0.010400424711406231, + 0.018374642357230186, + 0.00912256445735693, + 0.014964662492275238, + 0.011028645560145378, + 0.012252639047801495, + 0.010655304417014122, + 0.007697668857872486, + 0.009677504189312458, + 0.018606580793857574, + 0.003898107446730137, + 0.00724753038957715, + 0.003955823834985495, + 0.006078723352402449, + 0.004428934771567583, + 0.004429575055837631, + 0.003695880528539419, + 0.006713453214615583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.002819731365889311 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781407192349434 + ], + "opposite_logprob_tokens": [ + -4.634781360626221 + ], + "logprob_reward": -0.009781407192349434, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781407192349434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14298738539218903, + 0.15396249294281006, + 0.032073162496089935, + 0.0480668880045414, + 0.14647594094276428, + 0.047016799449920654, + 0.032183777540922165, + 0.12873876094818115, + 0.09639692306518555, + 0.1720978170633316 + ], + "credit_map": [ + 0.01462024636566639, + 0.0288582444190979, + 0.015330027788877487, + 0.008410644717514515, + 0.024688594043254852, + 0.013240580447018147, + 0.018226295709609985, + 0.019612766802310944, + 0.012805548496544361, + 0.01789398491382599, + 0.00990307703614235, + 0.012547791004180908, + 0.022809667512774467, + 0.030363652855157852, + 0.02148001827299595, + 0.02615876868367195, + 0.003718293271958828, + 0.003862607991322875, + 0.003211118746548891, + 0.0037833601236343384, + 0.0052107470110058784, + 0.0028318455442786217, + 0.006520354654639959, + 0.0029348377138376236, + 0.004439245443791151, + 0.004470487125217915, + 0.003301772754639387, + 0.004104357212781906, + 0.006844867020845413, + 0.005664893891662359, + 0.01523599773645401, + 0.004005272872745991, + 0.01883445307612419, + 0.026750430464744568, + 0.01028866320848465, + 0.023691225796937943, + 0.0203119944781065, + 0.015946164727211, + 0.0222992692142725, + 0.00835375115275383, + 0.00604252191260457, + 0.008450639434158802, + 0.008127029053866863, + 0.00669734925031662, + 0.005498487036675215, + 0.004271033685654402, + 0.004528486635535955, + 0.0034012545365840197, + 0.004550023470073938, + 0.004931529052555561, + 0.00366044114343822, + 0.0033448419999331236, + 0.0035772956907749176, + 0.004000252578407526, + 0.003402784001082182, + 0.004716613329946995, + 0.018249809741973877, + 0.026683276519179344, + 0.011306654661893845, + 0.013341228477656841, + 0.019865570589900017, + 0.014493870548903942, + 0.007365906611084938, + 0.01743246056139469, + 0.009739851579070091, + 0.015818113461136818, + 0.011456462554633617, + 0.013041660189628601, + 0.010739892721176147, + 0.0074890679679811, + 0.009714122861623764, + 0.018397757783532143, + 0.019677408039569855, + 0.03369991481304169, + 0.030336806550621986, + 0.017026349902153015, + 0.01618969812989235, + 0.012594727799296379, + 0.01519359927624464, + 0.02737932838499546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.009781407192349434 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a tight close-up shot, the camera pans steadily to the right, following Sun Wukong, crowned by an intricately carved golden headband, as he soars at high speed atop a swirling cloud. A hint of defiant confidence graces his majestic monkey face, the corners of his mouth curling upwards slightly, while his signature fiery eyes gleam with a sharp, fiery golden light, piercing through all illusions. The wind whipping up from his high-speed flight ruffles the golden fur along his cheeks and sends the red tassels dangling from his headband fluttering violently backward. Below, the edge of his signature golden chain mail armor catches a dazzling glint in the sunlight. Behind him, a vast expanse of azure sky and rolling, indigo-hued mountains, like an ink painting, shrouded in mist and clouds, create an epic and ethereal feeling. The entire picture is bathed in the warm golden sunlight of the afternoon. The light outlines every detail of his silhouette, emphasizing the texture of the metal and the softness of his hair. The overall tone is warm and powerful, perfectly capturing the moment of the Monkey King's fearlessness and high spirits.", + "seed": 14736, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_10/index_2413_seed_14736_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07635382181033493, + "vlm_logprob": -0.07635382181033493, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.07635382181033493, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013917069882154465 + ], + "opposite_logprob_tokens": [ + -4.513916969299316 + ], + "logprob_reward": -0.013917069882154465, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013917069882154465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12846598029136658, + 0.050260890275239944, + 0.13748478889465332, + 0.2856829762458801, + 0.05567270517349243, + 0.07815904915332794, + 0.08109430223703384, + 0.038748644292354584, + 0.11826024949550629, + 0.026170456781983376 + ], + "credit_map": [ + 0.016801873221993446, + 0.010078102350234985, + 0.02334713563323021, + 0.014580260030925274, + 0.01534777320921421, + 0.020463701337575912, + 0.01457897573709488, + 0.013268154114484787, + 0.010991862043738365, + 0.0032675471156835556, + 0.01522583793848753, + 0.0046882638707757, + 0.005229102913290262, + 0.0034680606331676245, + 0.004353221505880356, + 0.003036993555724621, + 0.039825502783060074, + 0.010318788699805737, + 0.042366839945316315, + 0.00915785413235426, + 0.01068798080086708, + 0.008579177781939507, + 0.009102045558393002, + 0.007446602452546358, + 0.07529135793447495, + 0.022239675745368004, + 0.08682485669851303, + 0.018957167863845825, + 0.03228802978992462, + 0.017690138891339302, + 0.017809506505727768, + 0.014582260511815548, + 0.006200793664902449, + 0.0077536748722195625, + 0.0047987764701247215, + 0.006725630257278681, + 0.011446231044828892, + 0.003913977649062872, + 0.009115892462432384, + 0.005717727821320295, + 0.019010508432984352, + 0.00976420845836401, + 0.005082858260720968, + 0.009057825431227684, + 0.014622137881815434, + 0.005811488255858421, + 0.0070854960940778255, + 0.007724526338279247, + 0.01945267990231514, + 0.010042061097919941, + 0.007871133275330067, + 0.0064314925111830235, + 0.006932531483471394, + 0.01722782477736473, + 0.006344641093164682, + 0.006791937164962292, + 0.005623117089271545, + 0.005966802593320608, + 0.003457867307588458, + 0.005496335215866566, + 0.004539184272289276, + 0.0052830553613603115, + 0.004620371386408806, + 0.0037619101349264383, + 0.008109613321721554, + 0.01086815819144249, + 0.006001799367368221, + 0.011098120361566544, + 0.013930462300777435, + 0.016767330467700958, + 0.04429343342781067, + 0.007191330660134554, + 0.0019622300751507282, + 0.0029378151521086693, + 0.003554629860445857, + 0.0036881633568555117, + 0.001985449343919754, + 0.0017617265693843365, + 0.007639888674020767, + 0.0026405546814203262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.013917069882154465 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.313480019569397 + ], + "opposite_logprob_tokens": [ + -1.313480019569397 + ], + "logprob_reward": -0.313480019569397, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.313480019569397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22403639554977417, + 0.1580529808998108, + 0.0708950087428093, + 0.11408219486474991, + 0.05544188991189003, + 0.03443078324198723, + 0.02215326577425003, + 0.04773997887969017, + 0.18768973648548126, + 0.08547782897949219 + ], + "credit_map": [ + 0.04828479886054993, + 0.017659097909927368, + 0.05805077776312828, + 0.020835502073168755, + 0.024571409448981285, + 0.018273571506142616, + 0.018448349088430405, + 0.017912857234477997, + 0.023632599040865898, + 0.01652645319700241, + 0.021593885496258736, + 0.020751213654875755, + 0.023916058242321014, + 0.01729498989880085, + 0.019138362258672714, + 0.015199403278529644, + 0.009242502972483635, + 0.010065445676445961, + 0.006538217421621084, + 0.007706788368523121, + 0.009389245882630348, + 0.011332862079143524, + 0.009207716211676598, + 0.007412225939333439, + 0.018717467784881592, + 0.013255580328404903, + 0.015000345185399055, + 0.012732777744531631, + 0.021851740777492523, + 0.00909368135035038, + 0.012527888640761375, + 0.010902701877057552, + 0.00734579935669899, + 0.0087768929079175, + 0.005603491794317961, + 0.00536802364513278, + 0.009450395591557026, + 0.0047624362632632256, + 0.006345255766063929, + 0.007789591792970896, + 0.0027475093957036734, + 0.005900667514652014, + 0.003726919647306204, + 0.004575182683765888, + 0.004889708943665028, + 0.0032977904193103313, + 0.004581636283546686, + 0.0047113653272390366, + 0.00284446869045496, + 0.003952005412429571, + 0.0021142272744327784, + 0.0029584658332169056, + 0.0025318500120192766, + 0.0023562770802527666, + 0.002070659538730979, + 0.0033253091387450695, + 0.005032294895499945, + 0.00578460656106472, + 0.003273583482950926, + 0.006289146840572357, + 0.007941148243844509, + 0.007554898038506508, + 0.006330091506242752, + 0.005534205120056868, + 0.012451095506548882, + 0.0164673812687397, + 0.01065017655491829, + 0.019458536058664322, + 0.018214548006653786, + 0.028475338593125343, + 0.06659943610429764, + 0.01537320390343666, + 0.0047488100826740265, + 0.0068807341158390045, + 0.007957102730870247, + 0.009400229901075363, + 0.00499515887349844, + 0.01236387062817812, + 0.033993199467659, + 0.005138717591762543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.313480019569397 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007640306372195482 + ], + "opposite_logprob_tokens": [ + -4.882640361785889 + ], + "logprob_reward": -0.007640306372195482, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007640306372195482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13710716366767883, + 0.11877968907356262, + 0.17695890367031097, + 0.13425591588020325, + 0.02502591721713543, + 0.04530435428023338, + 0.033819667994976044, + 0.052254319190979004, + 0.13923953473567963, + 0.1372545063495636 + ], + "credit_map": [ + 0.02351023070514202, + 0.010200412012636662, + 0.02802436240017414, + 0.022883232682943344, + 0.016913319006562233, + 0.013009779155254364, + 0.012108508497476578, + 0.01045731920748949, + 0.029315287247300148, + 0.009118420071899891, + 0.03852280601859093, + 0.01748647168278694, + 0.007375354412943125, + 0.0051232329569756985, + 0.006631284486502409, + 0.005206828471273184, + 0.059637658298015594, + 0.011793063022196293, + 0.06540515273809433, + 0.008660949766635895, + 0.00902674999088049, + 0.005216393154114485, + 0.0076600308530032635, + 0.009558903984725475, + 0.0290292389690876, + 0.010710537433624268, + 0.035355981439352036, + 0.012081893160939217, + 0.018532220274209976, + 0.009160579182207584, + 0.010711940936744213, + 0.008673518896102905, + 0.002898991806432605, + 0.003110929625108838, + 0.002195703564211726, + 0.003198381746187806, + 0.004763047676533461, + 0.0019110357388854027, + 0.0036372824106365442, + 0.0033105448819696903, + 0.004482528660446405, + 0.007068227510899305, + 0.0029523768462240696, + 0.00498373294249177, + 0.005988643504679203, + 0.004052022006362677, + 0.007967488840222359, + 0.007809337694197893, + 0.004672115668654442, + 0.00572229502722621, + 0.003700380213558674, + 0.0037034268025308847, + 0.0048033976927399635, + 0.003873501205816865, + 0.003541294252499938, + 0.003803257131949067, + 0.0067502353340387344, + 0.005973232444375753, + 0.005478199105709791, + 0.005708550568670034, + 0.00542645575478673, + 0.007873616181313992, + 0.008251882158219814, + 0.006792146246880293, + 0.009118651039898396, + 0.010081250220537186, + 0.005638007074594498, + 0.012514442205429077, + 0.021203864365816116, + 0.01691232994198799, + 0.05494696646928787, + 0.008824020624160767, + 0.0067823403514921665, + 0.012682628817856312, + 0.012524905614554882, + 0.014786242507398129, + 0.008549408987164497, + 0.019757797941565514, + 0.05327194929122925, + 0.008899232372641563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.007640306372195482 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635676465928555 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.008635676465928555, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635676465928555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20220907032489777, + 0.5657816529273987, + 0.037365954369306564, + 0.034749675542116165, + 0.010455392301082611, + 0.02477949485182762, + 0.06498394906520844, + 0.0228972677141428, + 0.018359141424298286, + 0.018418392166495323 + ], + "credit_map": [ + 0.02995162084698677, + 0.018881486728787422, + 0.034890759736299515, + 0.04043792560696602, + 0.02502446621656418, + 0.01678847335278988, + 0.022776171565055847, + 0.01345815695822239, + 0.16803067922592163, + 0.03597090393304825, + 0.2065625637769699, + 0.06333917379379272, + 0.022894954308867455, + 0.01790587231516838, + 0.0231640487909317, + 0.02791348658502102, + 0.010796680115163326, + 0.0036810908932238817, + 0.010170851834118366, + 0.0023432259913533926, + 0.0027319297660142183, + 0.0023573231883347034, + 0.0023927632719278336, + 0.0028920916374772787, + 0.006329454015940428, + 0.0042500365525484085, + 0.007805021479725838, + 0.0024468216579407454, + 0.005064391065388918, + 0.002436346374452114, + 0.004415819421410561, + 0.002001783112064004, + 0.0013414775021374226, + 0.0016805721679702401, + 0.0014645564369857311, + 0.0013383063487708569, + 0.0010701611172407866, + 0.0009077017894014716, + 0.0013625241117551923, + 0.0012900930596515536, + 0.005235432647168636, + 0.0030895948875695467, + 0.0017297632293775678, + 0.0025321664288640022, + 0.005198162980377674, + 0.0019384343177080154, + 0.002326906193047762, + 0.002729033352807164, + 0.011066708713769913, + 0.008220380172133446, + 0.007420248817652464, + 0.00630919961258769, + 0.0064420197159051895, + 0.011826308444142342, + 0.006226816214621067, + 0.007472271099686623, + 0.004520961549133062, + 0.0027090110816061497, + 0.002328001894056797, + 0.002253910293802619, + 0.002816850319504738, + 0.0033129665534943342, + 0.003025046084076166, + 0.0019305192399770021, + 0.0015926604392006993, + 0.0020112202037125826, + 0.00171513925306499, + 0.002152280882000923, + 0.0019000004976987839, + 0.0032998123206198215, + 0.004404222592711449, + 0.001283804653212428, + 0.0010464335791766644, + 0.0024734686594456434, + 0.001954497769474983, + 0.0020435042679309845, + 0.001836618990637362, + 0.002387239597737789, + 0.004688916727900505, + 0.0019877131562680006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.008635676465928555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right following Sun Wukong?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Sun Wukong wearing an intricately carved golden headband with red tassels that flutter backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of an azure sky and rolling indigo-hued mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight outlining the silhouette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809603676199913 + ], + "opposite_logprob_tokens": [ + -3.2880959510803223 + ], + "logprob_reward": -0.03809603676199913, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809603676199913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2838361859321594, + 0.08424828946590424, + 0.07865110784769058, + 0.09373606741428375, + 0.060366738587617874, + 0.07717438787221909, + 0.1798105686903, + 0.03965369239449501, + 0.020773224532604218, + 0.08174967020750046 + ], + "credit_map": [ + 0.04672062769532204, + 0.021555006504058838, + 0.06068585813045502, + 0.04031377285718918, + 0.03249792382121086, + 0.030582245439291, + 0.025710027664899826, + 0.02577069401741028, + 0.014204116538167, + 0.010273509658873081, + 0.011175796389579773, + 0.016808563843369484, + 0.00879964791238308, + 0.00823911651968956, + 0.008322779089212418, + 0.006424758583307266, + 0.01820291206240654, + 0.00781718734651804, + 0.01702905260026455, + 0.005504667293280363, + 0.0072482586838305, + 0.00568860350176692, + 0.009691505692899227, + 0.007468922063708305, + 0.02043326385319233, + 0.008857255801558495, + 0.020728740841150284, + 0.00973597913980484, + 0.008824479766190052, + 0.007346183527261019, + 0.011860056780278683, + 0.005950110033154488, + 0.01141880452632904, + 0.005930121522396803, + 0.0047378018498420715, + 0.00520290294662118, + 0.016868051141500473, + 0.0034007804933935404, + 0.005881395656615496, + 0.006926882546395063, + 0.014683591201901436, + 0.009793070144951344, + 0.004655842669308186, + 0.007509369868785143, + 0.015648500993847847, + 0.00602729944512248, + 0.00763482041656971, + 0.011221903376281261, + 0.02871628664433956, + 0.016321642324328423, + 0.022361258044838905, + 0.0189621914178133, + 0.019037935882806778, + 0.04188910499215126, + 0.01631678268313408, + 0.01620537042617798, + 0.005536921322345734, + 0.005166046787053347, + 0.004885275848209858, + 0.004594097379595041, + 0.005031173583120108, + 0.004757674410939217, + 0.0050791483372449875, + 0.004603354725986719, + 0.002276361221447587, + 0.0027755270712077618, + 0.0023653556127101183, + 0.002141991164535284, + 0.003815452568233013, + 0.0016880352050065994, + 0.0030100937001407146, + 0.002700407523661852, + 0.004320558160543442, + 0.0066372789442539215, + 0.007513346616178751, + 0.00953032448887825, + 0.00642016064375639, + 0.011272412724792957, + 0.028773020952939987, + 0.007282567676156759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.03809603676199913 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_11/index_tensor([5683, 5683])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_11/index_tensor([5683, 5683])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..af5f4d47215de68076abc80f1b743056d3390ddf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_11/index_tensor([5683, 5683])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static telephoto lens, shot from a slightly upward angle, focuses on a young girl in a dark blue sailor-collar school uniform, her hair tied in a neat black ponytail, a few strands falling against her cheeks from the effort. Her brow furrowed, her lips pursed, her eyes etched with unyielding determination as she struggled to scale a crumbling concrete wall. Leaning forward, her hands gripped the edge of a cracked concrete block, her knuckles white from the strain, as she cautiously searched for a firm foothold. A hint of dust clung to the hem of her uniform. She was amidst a vast expanse of ruin, a backdrop of layers of shattered walls and twisted rebar blurred by the telephoto lens's compression. The soft afternoon sunlight penetrates the golden dust in the air, forming clearly visible beams of light. One of the warm spots of light happens to be projected onto her focused profile, illuminating the fine hairs on her cheek and her determined expression. The whole picture forms a sharp contrast between the cold gray tone of the ruins and the warm golden light and shadow, rendering a moment full of challenges and struggles with a tranquil and healing sense of power.", + "seed": 24093, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_11/index_5683_seed_24093_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04294864125549793, + "vlm_logprob": -0.04294864125549793, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.04294864125549793, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018900301307439804 + ], + "opposite_logprob_tokens": [ + -4.018900394439697 + ], + "logprob_reward": -0.018900301307439804, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018900301307439804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25869330763816833, + 0.09732162207365036, + 0.025950459763407707, + 0.07109364867210388, + 0.06828752905130386, + 0.27023404836654663, + 0.11626443266868591, + 0.034507837146520615, + 0.041113898158073425, + 0.01653316617012024 + ], + "credit_map": [ + 0.019635288044810295, + 0.05309019237756729, + 0.03519458323717117, + 0.032190751284360886, + 0.053565092384815216, + 0.023411313071846962, + 0.025842418894171715, + 0.0157636646181345, + 0.017180781811475754, + 0.020947204902768135, + 0.01751398853957653, + 0.00994479563087225, + 0.01359871868044138, + 0.00531685771420598, + 0.006192987784743309, + 0.006626290734857321, + 0.0035068748984485865, + 0.00419606314972043, + 0.004362691193819046, + 0.0024122348986566067, + 0.005363358184695244, + 0.0021797791123390198, + 0.002064615488052368, + 0.0018648420227691531, + 0.006035865284502506, + 0.007671693805605173, + 0.007919895462691784, + 0.011769348755478859, + 0.012985678389668465, + 0.00982117559760809, + 0.00897852424532175, + 0.005911469925194979, + 0.004095397423952818, + 0.017320100218057632, + 0.01718721352517605, + 0.005323322024196386, + 0.009788753464818, + 0.006288876291364431, + 0.004013922996819019, + 0.004269944969564676, + 0.013202734291553497, + 0.009097488597035408, + 0.11055583506822586, + 0.012084009125828743, + 0.024712683632969856, + 0.0849381610751152, + 0.006697752512991428, + 0.008945397101342678, + 0.018745986744761467, + 0.01355708297342062, + 0.011662628501653671, + 0.01772981323301792, + 0.015706732869148254, + 0.014671845361590385, + 0.015436570160090923, + 0.008753778412938118, + 0.005087331868708134, + 0.004555150400847197, + 0.003743905806913972, + 0.0034941521007567644, + 0.004575020167976618, + 0.006006840616464615, + 0.003722978523001075, + 0.003322457429021597, + 0.008315956220030785, + 0.006130364257842302, + 0.0052025518380105495, + 0.00407631928101182, + 0.004366707522422075, + 0.0059035774320364, + 0.0028300669509917498, + 0.004288353957235813, + 0.002609307412058115, + 0.0019345846958458424, + 0.001956164836883545, + 0.00235308357514441, + 0.0021160708274692297, + 0.0013133807806298137, + 0.002200834918767214, + 0.002049739472568035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.018900301307439804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187304958701134 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.018187304958701134, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018187304958701134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23355549573898315, + 0.13082486391067505, + 0.037613917142152786, + 0.100316621363163, + 0.08537262678146362, + 0.18624848127365112, + 0.12810631096363068, + 0.03925696387887001, + 0.03990081325173378, + 0.01880393922328949 + ], + "credit_map": [ + 0.018298408016562462, + 0.04448263719677925, + 0.03219031170010567, + 0.029713625088334084, + 0.04822501540184021, + 0.019612722098827362, + 0.026583300903439522, + 0.014449464157223701, + 0.01995866186916828, + 0.024944467470049858, + 0.03073621354997158, + 0.012188538908958435, + 0.027362573891878128, + 0.004826055373996496, + 0.006309922784566879, + 0.004498419817537069, + 0.0055524748750030994, + 0.0059194075874984264, + 0.006434492766857147, + 0.0033536262344568968, + 0.006226460449397564, + 0.003548630280420184, + 0.0033762564416974783, + 0.0032025654800236225, + 0.007314692251384258, + 0.008204005658626556, + 0.011194713413715363, + 0.013893730007112026, + 0.019824568182229996, + 0.016760239377617836, + 0.01374775916337967, + 0.00937691144645214, + 0.005150984972715378, + 0.03036392852663994, + 0.029278336092829704, + 0.004399801138788462, + 0.006812193430960178, + 0.0042418744415044785, + 0.0028111720457673073, + 0.0023143368307501078, + 0.010743704624474049, + 0.008838066831231117, + 0.07651685178279877, + 0.009069664403796196, + 0.01704361103475094, + 0.053263869136571884, + 0.005340798757970333, + 0.005431911442428827, + 0.018715547397732735, + 0.015138665214180946, + 0.012589599005877972, + 0.01996253989636898, + 0.01735961250960827, + 0.01624038629233837, + 0.017925720661878586, + 0.010174252092838287, + 0.007027280516922474, + 0.005578954704105854, + 0.004537084139883518, + 0.004643913358449936, + 0.004641386680305004, + 0.0069504352286458015, + 0.0034267986193299294, + 0.002451106673106551, + 0.009286847896873951, + 0.005227349698543549, + 0.005326389335095882, + 0.0038928487338125706, + 0.003660193644464016, + 0.005488581955432892, + 0.0026737248990684748, + 0.004344874061644077, + 0.003318707225844264, + 0.002180896233767271, + 0.0023062487598508596, + 0.0028044250793755054, + 0.002505409996956587, + 0.0014762419741600752, + 0.0019791140221059322, + 0.00223289686255157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.018187304958701134 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903422951698303 + ], + "opposite_logprob_tokens": [ + -2.46403431892395 + ], + "logprob_reward": -0.08903422951698303, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903422951698303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21107898652553558, + 0.09498656541109085, + 0.024655774235725403, + 0.052527304738759995, + 0.15981104969978333, + 0.19727663695812225, + 0.15850546956062317, + 0.023256603628396988, + 0.05895128846168518, + 0.01895037852227688 + ], + "credit_map": [ + 0.014941608533263206, + 0.03794080764055252, + 0.02955055981874466, + 0.02512414939701557, + 0.0453004464507103, + 0.01916251704096794, + 0.02565869316458702, + 0.013400169089436531, + 0.015946893021464348, + 0.019178928807377815, + 0.020007632672786713, + 0.009522461332380772, + 0.016890032216906548, + 0.004087620414793491, + 0.005460015032440424, + 0.0038929665461182594, + 0.0032056928612291813, + 0.003322524717077613, + 0.0035367023665457964, + 0.002373647876083851, + 0.004703310783952475, + 0.0022597911302000284, + 0.00221233069896698, + 0.0030417696107178926, + 0.005059884395450354, + 0.0066378978081047535, + 0.0062527284026145935, + 0.011823396198451519, + 0.0073922076262533665, + 0.0056068748235702515, + 0.006660848390311003, + 0.0030934573151171207, + 0.009057208895683289, + 0.06346457451581955, + 0.05817437171936035, + 0.0064102536998689175, + 0.011456661857664585, + 0.004754221066832542, + 0.004021020606160164, + 0.00247270823456347, + 0.010550577193498611, + 0.008725689724087715, + 0.08057913929224014, + 0.01026034913957119, + 0.01803506724536419, + 0.05809610337018967, + 0.005624380894005299, + 0.005405315198004246, + 0.02641691081225872, + 0.01883038319647312, + 0.01603413000702858, + 0.023153165355324745, + 0.020711423829197884, + 0.020127801224589348, + 0.02127305045723915, + 0.011958597227931023, + 0.0038904917892068624, + 0.0033813186455518007, + 0.002458751667290926, + 0.0026983602438122034, + 0.0026756839361041784, + 0.0030000514816492796, + 0.002541798632591963, + 0.002610144205391407, + 0.012304349802434444, + 0.008522105403244495, + 0.007592155132442713, + 0.005643244832754135, + 0.006402455270290375, + 0.008016064763069153, + 0.0043522086925804615, + 0.006118700373917818, + 0.003521692007780075, + 0.001972140045836568, + 0.0025039981119334698, + 0.002788300858810544, + 0.0023269131779670715, + 0.0014894501073285937, + 0.002082507824525237, + 0.0022653755731880665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.08903422951698303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368738293647766 + ], + "opposite_logprob_tokens": [ + -3.4086873531341553 + ], + "logprob_reward": -0.03368738293647766, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368738293647766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24936896562576294, + 0.03791943937540054, + 0.06731679290533066, + 0.037740666419267654, + 0.09412503242492676, + 0.3637697696685791, + 0.06123950704932213, + 0.04669852554798126, + 0.026722492650151253, + 0.01509877759963274 + ], + "credit_map": [ + 0.018428321927785873, + 0.052946314215660095, + 0.03374467045068741, + 0.02840588614344597, + 0.056835681200027466, + 0.018126068636775017, + 0.0273066908121109, + 0.013575322926044464, + 0.006119462661445141, + 0.005880596116185188, + 0.00892767682671547, + 0.00413048779591918, + 0.006991819478571415, + 0.0018569036619737744, + 0.00224706600420177, + 0.0017654249677434564, + 0.009607606567442417, + 0.01519747357815504, + 0.01183905266225338, + 0.006273822393268347, + 0.009890842251479626, + 0.004160726442933083, + 0.005724478047341108, + 0.004622796084731817, + 0.003817999269813299, + 0.003920821473002434, + 0.0041501144878566265, + 0.006990132387727499, + 0.006174289621412754, + 0.005174647551029921, + 0.004433584865182638, + 0.0030790765304118395, + 0.005232871510088444, + 0.03679906204342842, + 0.03375338762998581, + 0.0037568784318864346, + 0.006875656545162201, + 0.003459857776761055, + 0.002381941769272089, + 0.0018653871957212687, + 0.016726981848478317, + 0.01058940589427948, + 0.15012498199939728, + 0.014906075783073902, + 0.033442143350839615, + 0.12058238685131073, + 0.00737141165882349, + 0.010026379488408566, + 0.007801890838891268, + 0.007978399284183979, + 0.005922190845012665, + 0.010144716128706932, + 0.008331812918186188, + 0.008020028471946716, + 0.008851367048919201, + 0.004189105238765478, + 0.010349446907639503, + 0.006800008937716484, + 0.005458362400531769, + 0.005843719933182001, + 0.004473359789699316, + 0.006914185825735331, + 0.0038462388329207897, + 0.0030132047832012177, + 0.0032371575944125652, + 0.003901658346876502, + 0.0033640742767602205, + 0.0029138901736587286, + 0.0030832956545054913, + 0.005353718064725399, + 0.0018673304002732038, + 0.0030013700015842915, + 0.002717530122026801, + 0.001674037892371416, + 0.0016990011790767312, + 0.0017752144485712051, + 0.0020288778468966484, + 0.0012840271228924394, + 0.002088013803586364, + 0.0018320742528885603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.03368738293647766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493398755788803 + ], + "opposite_logprob_tokens": [ + -2.929934024810791 + ], + "logprob_reward": -0.05493398755788803, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493398755788803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30091023445129395, + 0.05770299211144447, + 0.07345570623874664, + 0.02554032765328884, + 0.07173148542642593, + 0.2972072660923004, + 0.07107566297054291, + 0.05295335873961449, + 0.03322261944413185, + 0.01620035618543625 + ], + "credit_map": [ + 0.02517004869878292, + 0.0681772232055664, + 0.03962774574756622, + 0.03789462149143219, + 0.07124137133359909, + 0.018696583807468414, + 0.027126023545861244, + 0.012976634316146374, + 0.00471092713996768, + 0.004574890248477459, + 0.017991939559578896, + 0.004544155206531286, + 0.018190938979387283, + 0.002925032516941428, + 0.002408075612038374, + 0.0023570340126752853, + 0.009910892695188522, + 0.01694042608141899, + 0.013027850538492203, + 0.00725182332098484, + 0.010733036324381828, + 0.004491347819566727, + 0.0057870796881616116, + 0.005313252564519644, + 0.0022429018281400204, + 0.0027412844356149435, + 0.0035322806797921658, + 0.0029745507054030895, + 0.004448290448635817, + 0.0036861200351268053, + 0.003669968107715249, + 0.0022449325770139694, + 0.004210225772112608, + 0.0271239522844553, + 0.02470872364938259, + 0.003136647865176201, + 0.005884262733161449, + 0.002671928843483329, + 0.0020584072917699814, + 0.0019373331451788545, + 0.013515160419046879, + 0.008469685912132263, + 0.12281584739685059, + 0.011682897806167603, + 0.027450503781437874, + 0.09917369484901428, + 0.005885396618396044, + 0.008214081637561321, + 0.006709837820380926, + 0.00833871215581894, + 0.007447804324328899, + 0.010471438989043236, + 0.010686323046684265, + 0.010855711996555328, + 0.011844544671475887, + 0.004721284843981266, + 0.011850782670080662, + 0.007858005352318287, + 0.006256486754864454, + 0.005640806630253792, + 0.005964200012385845, + 0.009895969182252884, + 0.0030267653055489063, + 0.002460342599079013, + 0.005084598902612925, + 0.004250277299433947, + 0.003487485693767667, + 0.005050057545304298, + 0.003654747037217021, + 0.005187093745917082, + 0.003114779479801655, + 0.0033935795072466135, + 0.003111122874543071, + 0.001858197501860559, + 0.00193480693269521, + 0.002170959487557411, + 0.002107893116772175, + 0.0010655614314600825, + 0.002195243490859866, + 0.0017565707676112652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.05493398755788803 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static telephoto lens, shot from a slightly upward angle, focuses on a young girl in a dark blue sailor-collar school uniform, her hair tied in a neat black ponytail, a few strands falling against her cheeks from the effort. Her brow furrowed, her lips pursed, her eyes etched with unyielding determination as she struggled to scale a crumbling concrete wall. Leaning forward, her hands gripped the edge of a cracked concrete block, her knuckles white from the strain, as she cautiously searched for a firm foothold. A hint of dust clung to the hem of her uniform. She was amidst a vast expanse of ruin, a backdrop of layers of shattered walls and twisted rebar blurred by the telephoto lens's compression. The soft afternoon sunlight penetrates the golden dust in the air, forming clearly visible beams of light. One of the warm spots of light happens to be projected onto her focused profile, illuminating the fine hairs on her cheek and her determined expression. The whole picture forms a sharp contrast between the cold gray tone of the ruins and the warm golden light and shadow, rendering a moment full of challenges and struggles with a tranquil and healing sense of power.", + "seed": 24094, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_11/index_5683_seed_24094_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01472411472350359, + "vlm_logprob": -0.01472411472350359, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.01472411472350359, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01388415228575468 + ], + "opposite_logprob_tokens": [ + -4.3888840675354 + ], + "logprob_reward": -0.01388415228575468, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01388415228575468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1937934011220932, + 0.08089820295572281, + 0.14351661503314972, + 0.1684362143278122, + 0.07796633243560791, + 0.04444418102502823, + 0.06409997493028641, + 0.06027518957853317, + 0.12407421320676804, + 0.042495667934417725 + ], + "credit_map": [ + 0.03347504511475563, + 0.01995416358113289, + 0.020210053771734238, + 0.027774613350629807, + 0.01352695282548666, + 0.027648992836475372, + 0.015179741196334362, + 0.036023832857608795, + 0.008809756487607956, + 0.010657006874680519, + 0.011379134841263294, + 0.009576205164194107, + 0.010107982903718948, + 0.010529165156185627, + 0.009968257509171963, + 0.009870697744190693, + 0.013440988957881927, + 0.015550581738352776, + 0.01104319654405117, + 0.016787759959697723, + 0.015148235484957695, + 0.01988697610795498, + 0.036041781306266785, + 0.015617094933986664, + 0.011317613534629345, + 0.016777904704213142, + 0.01496898289769888, + 0.03401004150509834, + 0.030854543671011925, + 0.015477869659662247, + 0.018258893862366676, + 0.02677038125693798, + 0.0064911688677966595, + 0.009075414389371872, + 0.006337899249047041, + 0.01312341634184122, + 0.010194201953709126, + 0.012921535409986973, + 0.008505563251674175, + 0.011317132972180843, + 0.003369981423020363, + 0.0055180806666612625, + 0.004014401230961084, + 0.008737778291106224, + 0.0050367265939712524, + 0.004818172659724951, + 0.004909581039100885, + 0.008039457723498344, + 0.0057515245862305164, + 0.013431447558104992, + 0.0031127152033150196, + 0.009133606217801571, + 0.006704386323690414, + 0.007677983958274126, + 0.008481089025735855, + 0.009807221591472626, + 0.004496586974710226, + 0.009350658394396305, + 0.004523710813373327, + 0.009430864825844765, + 0.0036007200833410025, + 0.005935138091444969, + 0.010408715344965458, + 0.012528793886303902, + 0.01262673083692789, + 0.014913982711732388, + 0.014944166876375675, + 0.029981527477502823, + 0.010480840690433979, + 0.013006926514208317, + 0.012285555712878704, + 0.01583448238670826, + 0.004559722729027271, + 0.0032818056643009186, + 0.003547647036612034, + 0.003973071929067373, + 0.006399884354323149, + 0.00792697723954916, + 0.006453329231590033, + 0.006353229749947786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.01388415228575468 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786483831703663 + ], + "opposite_logprob_tokens": [ + -4.634786605834961 + ], + "logprob_reward": -0.009786483831703663, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786483831703663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1563941091299057, + 0.09977924823760986, + 0.06472016870975494, + 0.26352086663246155, + 0.09509854018688202, + 0.06477504223585129, + 0.11068868637084961, + 0.02939329668879509, + 0.045268770307302475, + 0.07036127895116806 + ], + "credit_map": [ + 0.025540854781866074, + 0.015311136841773987, + 0.019580049440264702, + 0.01861731708049774, + 0.009970617480576038, + 0.02383245714008808, + 0.012880253605544567, + 0.030661417171359062, + 0.008592689409852028, + 0.022353462874889374, + 0.007624116726219654, + 0.010448959656059742, + 0.00974767841398716, + 0.02312668040394783, + 0.00610932195559144, + 0.011776340194046497, + 0.005543140694499016, + 0.008056078106164932, + 0.006503385957330465, + 0.004691296722739935, + 0.007960019633173943, + 0.014402913860976696, + 0.00707619171589613, + 0.010487145744264126, + 0.018059147521853447, + 0.03539898246526718, + 0.018877293914556503, + 0.041310109198093414, + 0.043870341032743454, + 0.03126676753163338, + 0.030576026067137718, + 0.04416217282414436, + 0.007956095971167088, + 0.008768738247454166, + 0.005792774260044098, + 0.015967220067977905, + 0.014118976891040802, + 0.027470184490084648, + 0.006887506693601608, + 0.008137045428156853, + 0.005652140825986862, + 0.007816559635102749, + 0.006615334190428257, + 0.010900240391492844, + 0.01092776283621788, + 0.006891347002238035, + 0.0072341482155025005, + 0.00873751100152731, + 0.006899876054376364, + 0.017401253804564476, + 0.0085974195972085, + 0.02025642991065979, + 0.012124932371079922, + 0.014207389205694199, + 0.01641922816634178, + 0.014782147482037544, + 0.00336838373914361, + 0.003989201970398426, + 0.0027718376368284225, + 0.0037981318309903145, + 0.0026382815558463335, + 0.003979946952313185, + 0.004425177816301584, + 0.004422334488481283, + 0.0047962600365281105, + 0.006479793228209019, + 0.005160058382898569, + 0.0073259323835372925, + 0.0049538458697497845, + 0.004882353823632002, + 0.005926281679421663, + 0.005744244437664747, + 0.007457490544766188, + 0.006810975726693869, + 0.004971790127456188, + 0.008234143257141113, + 0.010417967103421688, + 0.010310974903404713, + 0.008858677931129932, + 0.013299251906573772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.009786483831703663 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979704923927784 + ], + "opposite_logprob_tokens": [ + -3.529797077178955 + ], + "logprob_reward": -0.02979704923927784, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979704923927784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19325508177280426, + 0.1354413777589798, + 0.07855428010225296, + 0.1365039348602295, + 0.1034553125500679, + 0.04541170224547386, + 0.07914341241121292, + 0.06748514622449875, + 0.06834018230438232, + 0.09240952879190445 + ], + "credit_map": [ + 0.024779047816991806, + 0.01635599508881569, + 0.020971568301320076, + 0.020787445828318596, + 0.022157447412610054, + 0.02974332496523857, + 0.015259698033332825, + 0.04320056736469269, + 0.013067400082945824, + 0.018439549952745438, + 0.013964815996587276, + 0.012297892943024635, + 0.021686626598238945, + 0.01713108830153942, + 0.012500619515776634, + 0.026353392750024796, + 0.0063701774924993515, + 0.00861749891191721, + 0.009949963539838791, + 0.008751503191888332, + 0.01435372605919838, + 0.013066804967820644, + 0.009049473330378532, + 0.00839513260871172, + 0.012384387664496899, + 0.019848283380270004, + 0.012826512567698956, + 0.02021877095103264, + 0.019421184435486794, + 0.016098668798804283, + 0.021543331444263458, + 0.014162793755531311, + 0.009158030152320862, + 0.0140719348564744, + 0.006115797441452742, + 0.014689454808831215, + 0.01842028833925724, + 0.0171321090310812, + 0.013695619069039822, + 0.010172083042562008, + 0.0037257785443216562, + 0.005490608513355255, + 0.0034926149528473616, + 0.007255735341459513, + 0.008775044232606888, + 0.0055539715103805065, + 0.004862556699663401, + 0.006255393847823143, + 0.005522837862372398, + 0.01274273730814457, + 0.006032404489815235, + 0.016953574493527412, + 0.007578080985695124, + 0.009686709381639957, + 0.010310214012861252, + 0.010316847823560238, + 0.0048342375084757805, + 0.009134205058217049, + 0.003919071052223444, + 0.008822848089039326, + 0.00450379541143775, + 0.012634220533072948, + 0.012223106808960438, + 0.011413661763072014, + 0.006602782756090164, + 0.009337634779512882, + 0.007504671812057495, + 0.013749774545431137, + 0.006975856609642506, + 0.00447830930352211, + 0.010918817482888699, + 0.008772334083914757, + 0.009645390324294567, + 0.00815563090145588, + 0.007724585942924023, + 0.008213113062083721, + 0.020116765052080154, + 0.011023850180208683, + 0.012760168872773647, + 0.014770030044019222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.02979704923927784 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040895650163292885 + ], + "opposite_logprob_tokens": [ + -5.50408935546875 + ], + "logprob_reward": -0.0040895650163292885, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040895650163292885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3339795470237732, + 0.3041592836380005, + 0.06512913107872009, + 0.04548818990588188, + 0.047992732375860214, + 0.0361347571015358, + 0.044862862676382065, + 0.03412822261452675, + 0.024869674816727638, + 0.0632556676864624 + ], + "credit_map": [ + 0.05721152946352959, + 0.02908838354051113, + 0.028242897242307663, + 0.02763952873647213, + 0.029679078608751297, + 0.05779673531651497, + 0.02254115417599678, + 0.08178017288446426, + 0.03332490473985672, + 0.028461309149861336, + 0.029302340000867844, + 0.01887846365571022, + 0.05794085934758186, + 0.0641336664557457, + 0.023453956469893456, + 0.04866376146674156, + 0.005762264598160982, + 0.007890417240560055, + 0.006354086101055145, + 0.004518604837357998, + 0.006737167481333017, + 0.013909853994846344, + 0.011387570761144161, + 0.00856915581971407, + 0.004289281088858843, + 0.004573616199195385, + 0.0045850384049117565, + 0.006357768550515175, + 0.0056001353077590466, + 0.008594833314418793, + 0.005455680191516876, + 0.006031831726431847, + 0.0037451479583978653, + 0.004833453334867954, + 0.0027388071175664663, + 0.006315488368272781, + 0.008745251223444939, + 0.011831602081656456, + 0.004689320921897888, + 0.005093655548989773, + 0.002853943267837167, + 0.003168074879795313, + 0.003156858729198575, + 0.006616563536226749, + 0.007024490274488926, + 0.004187107551842928, + 0.004928837530314922, + 0.004198876209557056, + 0.0026401772629469633, + 0.0076792556792497635, + 0.003870365209877491, + 0.011156778782606125, + 0.0056293499656021595, + 0.004090580623596907, + 0.004427356645464897, + 0.005368989892303944, + 0.003353675827383995, + 0.005500867962837219, + 0.0022409672383219004, + 0.00550720002502203, + 0.0034605897963047028, + 0.002736997790634632, + 0.005035734735429287, + 0.00629218528047204, + 0.003239830257371068, + 0.003180979983881116, + 0.002635352546349168, + 0.003596062073484063, + 0.002388512948527932, + 0.0029857181943953037, + 0.0027704176027327776, + 0.004072797950357199, + 0.007351181469857693, + 0.0053288741037249565, + 0.004855307284742594, + 0.006599800195544958, + 0.012428377754986286, + 0.009628013707697392, + 0.0075909909792244434, + 0.009473111480474472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0040895650163292885 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016063323244452477 + ], + "opposite_logprob_tokens": [ + -4.141063213348389 + ], + "logprob_reward": -0.016063323244452477, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016063323244452477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11233494430780411, + 0.08582273870706558, + 0.06315543502569199, + 0.04802991449832916, + 0.0697789266705513, + 0.04065969958901405, + 0.07756996154785156, + 0.0677880123257637, + 0.1644715815782547, + 0.2703888714313507 + ], + "credit_map": [ + 0.013524304144084454, + 0.009928856045007706, + 0.011754200793802738, + 0.01204329077154398, + 0.012148804031312466, + 0.017674723640084267, + 0.009006919339299202, + 0.0262538380920887, + 0.007933993823826313, + 0.011137099005281925, + 0.008524607867002487, + 0.005906140431761742, + 0.01254194788634777, + 0.02043890580534935, + 0.006602840963751078, + 0.012737204320728779, + 0.006352676544338465, + 0.005445337854325771, + 0.005661455914378166, + 0.008661719039082527, + 0.009671817533671856, + 0.009648236446082592, + 0.008766244165599346, + 0.008947945199906826, + 0.0035986625589430332, + 0.005271119996905327, + 0.004355006851255894, + 0.00909755751490593, + 0.007649194449186325, + 0.006232967600226402, + 0.005567612592130899, + 0.006257792469114065, + 0.005958959925919771, + 0.008112186565995216, + 0.004748941399157047, + 0.009953655302524567, + 0.011159746907651424, + 0.015047864057123661, + 0.006267757620662451, + 0.00852980837225914, + 0.0031283495482057333, + 0.00398772768676281, + 0.0033754876349121332, + 0.005457917228341103, + 0.006205968093127012, + 0.005121081136167049, + 0.00825536623597145, + 0.005127798765897751, + 0.005494299344718456, + 0.01389286294579506, + 0.006507008336484432, + 0.017558224499225616, + 0.009215247817337513, + 0.006699579302221537, + 0.007266126107424498, + 0.0109366150572896, + 0.005843068473041058, + 0.008734557777643204, + 0.004283646587282419, + 0.009395393542945385, + 0.006384946871548891, + 0.011255539953708649, + 0.009980262257158756, + 0.011910595931112766, + 0.012435351498425007, + 0.023746183142066002, + 0.015506594441831112, + 0.019887903705239296, + 0.026351770386099815, + 0.027766749262809753, + 0.020893769338726997, + 0.017883259803056717, + 0.033075153827667236, + 0.01842219941318035, + 0.021419119089841843, + 0.024151010438799858, + 0.050474971532821655, + 0.041406504809856415, + 0.03931274265050888, + 0.042127180844545364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.016063323244452477 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static telephoto lens, shot from a slightly upward angle, focuses on a young girl in a dark blue sailor-collar school uniform, her hair tied in a neat black ponytail, a few strands falling against her cheeks from the effort. Her brow furrowed, her lips pursed, her eyes etched with unyielding determination as she struggled to scale a crumbling concrete wall. Leaning forward, her hands gripped the edge of a cracked concrete block, her knuckles white from the strain, as she cautiously searched for a firm foothold. A hint of dust clung to the hem of her uniform. She was amidst a vast expanse of ruin, a backdrop of layers of shattered walls and twisted rebar blurred by the telephoto lens's compression. The soft afternoon sunlight penetrates the golden dust in the air, forming clearly visible beams of light. One of the warm spots of light happens to be projected onto her focused profile, illuminating the fine hairs on her cheek and her determined expression. The whole picture forms a sharp contrast between the cold gray tone of the ruins and the warm golden light and shadow, rendering a moment full of challenges and struggles with a tranquil and healing sense of power.", + "seed": 24095, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_11/index_5683_seed_24095_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02316059162840247, + "vlm_logprob": -0.02316059162840247, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.02316059162840247, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006513318512588739 + ], + "opposite_logprob_tokens": [ + -5.1315131187438965 + ], + "logprob_reward": -0.006513318512588739, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.006513318512588739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1884237676858902, + 0.118683822453022, + 0.10142535716295242, + 0.08493892848491669, + 0.07816542685031891, + 0.12818460166454315, + 0.07773113995790482, + 0.10333500057458878, + 0.06809868663549423, + 0.051013194024562836 + ], + "credit_map": [ + 0.014908800832927227, + 0.02921256422996521, + 0.03249416872859001, + 0.020261049270629883, + 0.015683773905038834, + 0.031391438096761703, + 0.015259461477398872, + 0.029212534427642822, + 0.011820164509117603, + 0.01498535368591547, + 0.015083126723766327, + 0.013836352154612541, + 0.01245980802923441, + 0.017632612958550453, + 0.02289636805653572, + 0.009970048442482948, + 0.011407301761209965, + 0.014110597781836987, + 0.008679024875164032, + 0.012014828622341156, + 0.014187008142471313, + 0.01764814928174019, + 0.010843292810022831, + 0.012535170651972294, + 0.009669196791946888, + 0.011348722502589226, + 0.008146391250193119, + 0.00845863576978445, + 0.012609659694135189, + 0.01283853780478239, + 0.010734218172729015, + 0.011133578605949879, + 0.0070386179722845554, + 0.011563279666006565, + 0.004889272153377533, + 0.006426606792956591, + 0.008359678089618683, + 0.015954244881868362, + 0.007183560635894537, + 0.016750173643231392, + 0.014829396270215511, + 0.01616763137280941, + 0.013492200523614883, + 0.0160590298473835, + 0.01571742631494999, + 0.015955273061990738, + 0.015674540773034096, + 0.020289115607738495, + 0.006280573550611734, + 0.011942719109356403, + 0.009979961439967155, + 0.013545363210141659, + 0.007836204953491688, + 0.011416977271437645, + 0.0075125666335225105, + 0.009216779842972755, + 0.0065579405054450035, + 0.011315491981804371, + 0.009467890486121178, + 0.010938222520053387, + 0.0058943904004991055, + 0.028394300490617752, + 0.009664405137300491, + 0.02110237069427967, + 0.0051693362183868885, + 0.007190641947090626, + 0.010798284783959389, + 0.009303178638219833, + 0.008545850403606892, + 0.0109416414052248, + 0.006829220801591873, + 0.009320538491010666, + 0.008062917739152908, + 0.007590165827423334, + 0.005353657063096762, + 0.006377455312758684, + 0.005623084958642721, + 0.0063322316855192184, + 0.004545754287391901, + 0.007127927150577307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.006513318512588739 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200235716998577 + ], + "opposite_logprob_tokens": [ + -4.264200210571289 + ], + "logprob_reward": -0.014200235716998577, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200235716998577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12675702571868896, + 0.09361135959625244, + 0.0718088448047638, + 0.09877699613571167, + 0.07266714423894882, + 0.23086562752723694, + 0.05647142603993416, + 0.10215399414300919, + 0.040252700448036194, + 0.10663490742444992 + ], + "credit_map": [ + 0.010785439983010292, + 0.01474007684737444, + 0.0241110697388649, + 0.013588983565568924, + 0.013001913204789162, + 0.022295869886875153, + 0.010639241896569729, + 0.017594438046216965, + 0.010624523274600506, + 0.010576615110039711, + 0.013629324734210968, + 0.01046948041766882, + 0.009492198936641216, + 0.014434192329645157, + 0.015399792231619358, + 0.00898523349314928, + 0.007504187989979982, + 0.008746171370148659, + 0.006388969719409943, + 0.008573988452553749, + 0.00910076592117548, + 0.015032888390123844, + 0.0072902194224298, + 0.009171652607619762, + 0.010171483270823956, + 0.014593313448131084, + 0.007652440574020147, + 0.01004907488822937, + 0.014525561593472958, + 0.016138486564159393, + 0.01167620811611414, + 0.013970425352454185, + 0.006829829886555672, + 0.01167292334139347, + 0.005442840978503227, + 0.00568152591586113, + 0.00899068545550108, + 0.013351818546652794, + 0.0051806094124913216, + 0.015516906045377254, + 0.03336063399910927, + 0.025101661682128906, + 0.021679092198610306, + 0.022028949111700058, + 0.03665715456008911, + 0.031763531267642975, + 0.0230539683252573, + 0.03722062706947327, + 0.005513446871191263, + 0.007017964031547308, + 0.008037908002734184, + 0.007775686215609312, + 0.005680030677467585, + 0.009196353144943714, + 0.0052324640564620495, + 0.008017569780349731, + 0.00909128412604332, + 0.01197754219174385, + 0.011128689162433147, + 0.012254844419658184, + 0.018180476501584053, + 0.019084015861153603, + 0.010995030403137207, + 0.009442109614610672, + 0.004211470950394869, + 0.008001306094229221, + 0.005097297020256519, + 0.004984352272003889, + 0.003952017053961754, + 0.004614158067852259, + 0.0028614881448447704, + 0.006530609913170338, + 0.013623370788991451, + 0.014897005632519722, + 0.011877297423779964, + 0.011637534946203232, + 0.01318990346044302, + 0.017648963257670403, + 0.009705021046102047, + 0.01405581459403038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.014200235716998577 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781407192349434 + ], + "opposite_logprob_tokens": [ + -4.634781360626221 + ], + "logprob_reward": -0.009781407192349434, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781407192349434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1133093312382698, + 0.08263560384511948, + 0.09913861006498337, + 0.16448113322257996, + 0.07840435951948166, + 0.12743332982063293, + 0.1330397129058838, + 0.052815575152635574, + 0.10275519639253616, + 0.04598712921142578 + ], + "credit_map": [ + 0.00911461841315031, + 0.017422424629330635, + 0.021150166168808937, + 0.012485094368457794, + 0.00809397827833891, + 0.021246980875730515, + 0.008246336132287979, + 0.015549725852906704, + 0.006613952107727528, + 0.009905675426125526, + 0.013258075341582298, + 0.008917841129004955, + 0.008982737548649311, + 0.01595614105463028, + 0.010837841778993607, + 0.008163338527083397, + 0.01046255137771368, + 0.014760675840079784, + 0.006755471229553223, + 0.011342117562890053, + 0.013940954580903053, + 0.016020480543375015, + 0.008806008845567703, + 0.017050359398126602, + 0.017500249668955803, + 0.01934535801410675, + 0.011016243137419224, + 0.014482430182397366, + 0.03582146763801575, + 0.02415015920996666, + 0.022137803956866264, + 0.020027421414852142, + 0.008272175677120686, + 0.010487656109035015, + 0.0054842084646224976, + 0.006599301937967539, + 0.011725683696568012, + 0.013521605171263218, + 0.011445273645222187, + 0.010868453420698643, + 0.013377495110034943, + 0.012758152559399605, + 0.00921647623181343, + 0.010461145080626011, + 0.024674609303474426, + 0.018048347905278206, + 0.0168901477009058, + 0.022006964311003685, + 0.014452088624238968, + 0.01299143023788929, + 0.00936069618910551, + 0.01149174477905035, + 0.02282201126217842, + 0.024480482563376427, + 0.016201667487621307, + 0.021239595487713814, + 0.003198557533323765, + 0.005605412181466818, + 0.003562712576240301, + 0.004120966419577599, + 0.003422291949391365, + 0.013435115106403828, + 0.0036601515021175146, + 0.015810372307896614, + 0.014041312970221043, + 0.011529745534062386, + 0.014137497171759605, + 0.008331098593771458, + 0.009724140167236328, + 0.015811167657375336, + 0.009492204524576664, + 0.019688019528985023, + 0.003048586193472147, + 0.005156210623681545, + 0.007722021080553532, + 0.006227323319762945, + 0.003971568308770657, + 0.008190851658582687, + 0.006472324021160603, + 0.005198243539780378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.009781407192349434 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023275844752788544 + ], + "opposite_logprob_tokens": [ + -3.773275852203369 + ], + "logprob_reward": -0.023275844752788544, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023275844752788544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08750640600919724, + 0.27068349719047546, + 0.1282586306333542, + 0.04269395396113396, + 0.047073084861040115, + 0.13629916310310364, + 0.11473988741636276, + 0.03785107284784317, + 0.05307386443018913, + 0.08182041347026825 + ], + "credit_map": [ + 0.0065802838653326035, + 0.016252001747488976, + 0.009167619980871677, + 0.011685057543218136, + 0.010534659959375858, + 0.013646801933646202, + 0.00841254461556673, + 0.011227437295019627, + 0.021227015182375908, + 0.021099764853715897, + 0.04992448911070824, + 0.022464780136942863, + 0.02303439937531948, + 0.06563273072242737, + 0.036123257130384445, + 0.03117707557976246, + 0.010535974986851215, + 0.012986559420824051, + 0.007469472475349903, + 0.014588664285838604, + 0.015871426090598106, + 0.033511459827423096, + 0.012237770482897758, + 0.021057303994894028, + 0.004823622759431601, + 0.005062280222773552, + 0.0032176142558455467, + 0.004227780736982822, + 0.006239001639187336, + 0.007761075627058744, + 0.0047075324691832066, + 0.006655051838606596, + 0.004158100578933954, + 0.00917578674852848, + 0.0035428982228040695, + 0.004453914240002632, + 0.006732509471476078, + 0.008444256149232388, + 0.003691712161526084, + 0.0068739112466573715, + 0.0171374399214983, + 0.00983469095081091, + 0.010334276594221592, + 0.010710429400205612, + 0.0259130597114563, + 0.03168078884482384, + 0.012296569533646107, + 0.018391920253634453, + 0.014164873398840427, + 0.010819941759109497, + 0.011014859192073345, + 0.011764916591346264, + 0.014551992528140545, + 0.02109706960618496, + 0.009678942151367664, + 0.02164730615913868, + 0.0025995962787419558, + 0.0034411922097206116, + 0.0031119449995458126, + 0.002673323266208172, + 0.002383548766374588, + 0.012269913218915462, + 0.0041785091161727905, + 0.007193045690655708, + 0.004149359185248613, + 0.005353904794901609, + 0.00846769381314516, + 0.004545522853732109, + 0.006512034684419632, + 0.010300969704985619, + 0.00614923657849431, + 0.007595143746584654, + 0.009783896617591381, + 0.011757250875234604, + 0.011417245492339134, + 0.011306583881378174, + 0.009988265112042427, + 0.009521333500742912, + 0.00939633697271347, + 0.008649507537484169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.023275844752788544 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062032151967287064 + ], + "opposite_logprob_tokens": [ + -2.8120322227478027 + ], + "logprob_reward": -0.062032151967287064, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062032151967287064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12079861015081406, + 0.13488420844078064, + 0.09558268636465073, + 0.04203232005238533, + 0.061649080365896225, + 0.07403048127889633, + 0.07266250252723694, + 0.09496507793664932, + 0.053258053958415985, + 0.25013697147369385 + ], + "credit_map": [ + 0.007846047170460224, + 0.013711906969547272, + 0.01963474228978157, + 0.012248571962118149, + 0.014753546565771103, + 0.02186037041246891, + 0.012973529286682606, + 0.01776989921927452, + 0.012658176012337208, + 0.014002499170601368, + 0.019333366304636, + 0.013842860236763954, + 0.014866722747683525, + 0.02706136181950569, + 0.01921098306775093, + 0.013908229768276215, + 0.007620199583470821, + 0.0066939592361450195, + 0.005708544515073299, + 0.011940598487854004, + 0.010067038238048553, + 0.028348539024591446, + 0.009503204375505447, + 0.015700597316026688, + 0.0075378865003585815, + 0.006572892423719168, + 0.0043158624321222305, + 0.0045820120722055435, + 0.004279060754925013, + 0.004873406141996384, + 0.005218399688601494, + 0.004652800969779491, + 0.005348195321857929, + 0.012374969199299812, + 0.00554624292999506, + 0.00678190728649497, + 0.00948792602866888, + 0.010216168127954006, + 0.0051698023453354836, + 0.006723865866661072, + 0.014195145107805729, + 0.005780077539384365, + 0.007185664027929306, + 0.006340672727674246, + 0.01129145547747612, + 0.014369156211614609, + 0.004820205736905336, + 0.010048103518784046, + 0.008793357759714127, + 0.008263308554887772, + 0.005903179291635752, + 0.004371493589133024, + 0.01112576387822628, + 0.01639818400144577, + 0.006282677873969078, + 0.011524534784257412, + 0.005741003435105085, + 0.008147121407091618, + 0.006271746475249529, + 0.008809336461126804, + 0.005784037057310343, + 0.026269497349858284, + 0.010586574673652649, + 0.023355763405561447, + 0.004825375974178314, + 0.006422510836273432, + 0.007154250051826239, + 0.00648073386400938, + 0.0048961821012198925, + 0.009058509953320026, + 0.003737651975825429, + 0.010682841762900352, + 0.020288832485675812, + 0.0353645496070385, + 0.02836555242538452, + 0.043260686099529266, + 0.028710579499602318, + 0.02770325168967247, + 0.03295045718550682, + 0.033493079245090485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.062032151967287064 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static telephoto lens, shot from a slightly upward angle, focuses on a young girl in a dark blue sailor-collar school uniform, her hair tied in a neat black ponytail, a few strands falling against her cheeks from the effort. Her brow furrowed, her lips pursed, her eyes etched with unyielding determination as she struggled to scale a crumbling concrete wall. Leaning forward, her hands gripped the edge of a cracked concrete block, her knuckles white from the strain, as she cautiously searched for a firm foothold. A hint of dust clung to the hem of her uniform. She was amidst a vast expanse of ruin, a backdrop of layers of shattered walls and twisted rebar blurred by the telephoto lens's compression. The soft afternoon sunlight penetrates the golden dust in the air, forming clearly visible beams of light. One of the warm spots of light happens to be projected onto her focused profile, illuminating the fine hairs on her cheek and her determined expression. The whole picture forms a sharp contrast between the cold gray tone of the ruins and the warm golden light and shadow, rendering a moment full of challenges and struggles with a tranquil and healing sense of power.", + "seed": 24096, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_11/index_5683_seed_24096_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.022119599767029284, + "vlm_logprob": -0.022119599767029284, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.022119599767029284, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008447864092886448 + ], + "opposite_logprob_tokens": [ + -4.883447647094727 + ], + "logprob_reward": -0.008447864092886448, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008447864092886448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06024681776762009, + 0.03210948780179024, + 0.2496153712272644, + 0.32780778408050537, + 0.03295058384537697, + 0.0812012180685997, + 0.12115912139415741, + 0.025867462158203125, + 0.03250569850206375, + 0.03653644025325775 + ], + "credit_map": [ + 0.008459939621388912, + 0.008168334141373634, + 0.007096440065652132, + 0.006057812832295895, + 0.011656206101179123, + 0.00778784416615963, + 0.005423435941338539, + 0.005596804898232222, + 0.0033104713074862957, + 0.003643834963440895, + 0.004073115065693855, + 0.004487309139221907, + 0.0037394394166767597, + 0.004174032714217901, + 0.0036268183030188084, + 0.005054471082985401, + 0.02175636775791645, + 0.029967427253723145, + 0.01367174182087183, + 0.07230620086193085, + 0.010926675982773304, + 0.0278482623398304, + 0.014079015702009201, + 0.05905969440937042, + 0.028605522587895393, + 0.0938958153128624, + 0.020032111555337906, + 0.10730433464050293, + 0.01023197267204523, + 0.014280348084867, + 0.04346790909767151, + 0.009989776648581028, + 0.005964995827525854, + 0.004804011434316635, + 0.007527600973844528, + 0.0028225944843143225, + 0.003992903046309948, + 0.002885807305574417, + 0.002348410664126277, + 0.00260426034219563, + 0.011606592684984207, + 0.006976023782044649, + 0.009676349349319935, + 0.010688415728509426, + 0.005745135713368654, + 0.007535621989518404, + 0.006590383127331734, + 0.022382697090506554, + 0.014821444638073444, + 0.013067071326076984, + 0.018140269443392754, + 0.012298954650759697, + 0.01109836995601654, + 0.024964500218629837, + 0.009012253023684025, + 0.0177562665194273, + 0.0033859170507639647, + 0.004320226144045591, + 0.004333777353167534, + 0.0032635186798870564, + 0.004565708339214325, + 0.002287867944687605, + 0.001773168914951384, + 0.0019372819224372506, + 0.0035457031335681677, + 0.0033362829126417637, + 0.0030550716910511255, + 0.001928813522681594, + 0.003647790290415287, + 0.008725649677217007, + 0.002239835448563099, + 0.006026556249707937, + 0.0038781866896897554, + 0.004561660345643759, + 0.006019568536430597, + 0.004093318246304989, + 0.0052149030379951, + 0.004703735467046499, + 0.0045730904676020145, + 0.003491979790851474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 6 + ], + "selected_reward": -0.008447864092886448 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789552539587021 + ], + "opposite_logprob_tokens": [ + -4.63478946685791 + ], + "logprob_reward": -0.009789552539587021, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789552539587021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08305994421243668, + 0.05650503933429718, + 0.07823427021503448, + 0.435157835483551, + 0.030692944303154945, + 0.07973861694335938, + 0.13023263216018677, + 0.04921969771385193, + 0.027379952371120453, + 0.029779091477394104 + ], + "credit_map": [ + 0.009909119457006454, + 0.012650926597416401, + 0.007678153924643993, + 0.007389516104012728, + 0.014824814163148403, + 0.012345409952104092, + 0.00812887866050005, + 0.010133124887943268, + 0.005454116966575384, + 0.005470930598676205, + 0.008159288205206394, + 0.006374964024871588, + 0.006794884335249662, + 0.010337688028812408, + 0.0056755514815449715, + 0.00823761336505413, + 0.0074322521686553955, + 0.011262944899499416, + 0.006417062599211931, + 0.020043684169650078, + 0.003763309447094798, + 0.008207133039832115, + 0.004452960100024939, + 0.016654925420880318, + 0.051776863634586334, + 0.0943574383854866, + 0.028956331312656403, + 0.14319080114364624, + 0.0137721486389637, + 0.01799343340098858, + 0.0690528005361557, + 0.01605796255171299, + 0.004001989960670471, + 0.0034662149846553802, + 0.0053736246190965176, + 0.003032444277778268, + 0.004165331367403269, + 0.004353752359747887, + 0.002897372469305992, + 0.003402212867513299, + 0.012014206498861313, + 0.007668560836464167, + 0.011314970441162586, + 0.006734452210366726, + 0.005320203024893999, + 0.01315066684037447, + 0.006109399721026421, + 0.017426159232854843, + 0.014248479157686234, + 0.016090355813503265, + 0.023981040343642235, + 0.01849331147968769, + 0.013552404940128326, + 0.014416909776628017, + 0.016076872125267982, + 0.01337326318025589, + 0.0054877218790352345, + 0.006580869201570749, + 0.006774569395929575, + 0.006234026048332453, + 0.008184068836271763, + 0.0037195514887571335, + 0.004372075200080872, + 0.007866817526519299, + 0.003100506030023098, + 0.0036155818961560726, + 0.0016216239891946316, + 0.001876931288279593, + 0.0034450667444616556, + 0.007374031934887171, + 0.001981595763936639, + 0.004364615771919489, + 0.0036555465776473284, + 0.0039810859598219395, + 0.003964560106396675, + 0.0034288172610104084, + 0.004691637586802244, + 0.0034731898922473192, + 0.003989286255091429, + 0.0025949690025299788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.009789552539587021 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0205828920006752 + ], + "opposite_logprob_tokens": [ + -3.895582914352417 + ], + "logprob_reward": -0.0205828920006752, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0205828920006752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03329463675618172, + 0.026179620996117592, + 0.2519030272960663, + 0.48828116059303284, + 0.027593404054641724, + 0.05585148185491562, + 0.03529195114970207, + 0.04980432242155075, + 0.01639845035970211, + 0.015401964075863361 + ], + "credit_map": [ + 0.005080380942672491, + 0.0042623793706297874, + 0.005657679866999388, + 0.002710215747356415, + 0.004668182227760553, + 0.0038952920585870743, + 0.0036353524774312973, + 0.003385154064744711, + 0.003717041341587901, + 0.003093627281486988, + 0.0032674185931682587, + 0.003445092123001814, + 0.0030677402392029762, + 0.003692902624607086, + 0.0024984620977193117, + 0.0033973364625126123, + 0.02195795811712742, + 0.03157070279121399, + 0.013187517412006855, + 0.073424331843853, + 0.011458720080554485, + 0.02780124358832836, + 0.014203333295881748, + 0.05829920992255211, + 0.03896729275584221, + 0.15095728635787964, + 0.026934565976262093, + 0.159419983625412, + 0.015429655089974403, + 0.022088343277573586, + 0.06070714071393013, + 0.013776871375739574, + 0.004472032189369202, + 0.00309325591661036, + 0.005598509684205055, + 0.002767273224890232, + 0.003376774024218321, + 0.0030224109068512917, + 0.002631695941090584, + 0.002631452167406678, + 0.00825190357863903, + 0.004679577890783548, + 0.007071307860314846, + 0.00576609605923295, + 0.004432964138686657, + 0.010988933965563774, + 0.004265241790562868, + 0.010395457036793232, + 0.004253406543284655, + 0.002503568073734641, + 0.006086374167352915, + 0.0036594392731785774, + 0.0022893212735652924, + 0.006877896375954151, + 0.003941369242966175, + 0.005680575035512447, + 0.006237749941647053, + 0.005073826760053635, + 0.005174045916646719, + 0.006232155952602625, + 0.00727141834795475, + 0.0051847221329808235, + 0.004018188454210758, + 0.010612213052809238, + 0.0016500421334058046, + 0.0014566308818757534, + 0.0017700640019029379, + 0.0019577969796955585, + 0.0016876619774848223, + 0.003267554799094796, + 0.002377306343987584, + 0.0022313925437629223, + 0.0013718369882553816, + 0.0021076775155961514, + 0.0020541546400636435, + 0.001477131969295442, + 0.0031485152430832386, + 0.0016669315518811345, + 0.0023326200898736715, + 0.0012430966598913074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.0205828920006752 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336892269551754 + ], + "opposite_logprob_tokens": [ + -3.408689260482788 + ], + "logprob_reward": -0.0336892269551754, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336892269551754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09452835470438004, + 0.0426146537065506, + 0.0969322994351387, + 0.42810893058776855, + 0.02040194906294346, + 0.09728212654590607, + 0.08795837312936783, + 0.05919520556926727, + 0.023716459050774574, + 0.049261659383773804 + ], + "credit_map": [ + 0.014400149695575237, + 0.014167238026857376, + 0.011264704167842865, + 0.008259703405201435, + 0.01382509060204029, + 0.012817900627851486, + 0.01020316407084465, + 0.009590390138328075, + 0.0037854742258787155, + 0.004390692338347435, + 0.005104003474116325, + 0.005041179247200489, + 0.004963042214512825, + 0.00840879138559103, + 0.004860530607402325, + 0.006060936953872442, + 0.009891102090477943, + 0.017441336065530777, + 0.013289889320731163, + 0.016826007515192032, + 0.009173907339572906, + 0.00828520581126213, + 0.008064175955951214, + 0.013960660435259342, + 0.042661577463150024, + 0.11225645244121552, + 0.02718331106007099, + 0.13999873399734497, + 0.013140855357050896, + 0.018646469339728355, + 0.06035346910357475, + 0.013868013396859169, + 0.0029353853315114975, + 0.0026320344768464565, + 0.0036532243248075247, + 0.0018259953940287232, + 0.00282901874743402, + 0.0025405283086001873, + 0.0018221125937998295, + 0.0021636500023305416, + 0.012790404260158539, + 0.006603788118809462, + 0.011657565832138062, + 0.007315861061215401, + 0.005263731814920902, + 0.02096562273800373, + 0.008087113499641418, + 0.02459803968667984, + 0.013123609125614166, + 0.009308765642344952, + 0.009788665920495987, + 0.008685172535479069, + 0.006375737953931093, + 0.017628811299800873, + 0.007349379826337099, + 0.015698229894042015, + 0.008118128404021263, + 0.006567354779690504, + 0.005683423485606909, + 0.009156906977295876, + 0.010132680647075176, + 0.004632641561329365, + 0.005031188018620014, + 0.009872875176370144, + 0.002725443569943309, + 0.003104000585153699, + 0.002846155082806945, + 0.001979870954528451, + 0.003762967884540558, + 0.004277222324162722, + 0.0018545308848842978, + 0.003166268579661846, + 0.006287385243922472, + 0.004861462861299515, + 0.008044208399951458, + 0.0053979516960680485, + 0.006843769457191229, + 0.006529628299176693, + 0.0070636761374771595, + 0.004233579151332378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.0336892269551754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young girl wearing a dark blue sailor-collar school uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl's hair tied in a neat black ponytail with a few strands falling against her cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl shown struggling to scale a crumbling concrete wall with her hands gripping a cracked block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of ruins featuring layers of shattered walls and twisted rebar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight creating visible golden beams of light and dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808846324682236 + ], + "opposite_logprob_tokens": [ + -3.28808856010437 + ], + "logprob_reward": -0.03808846324682236, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808846324682236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08516892045736313, + 0.055577248334884644, + 0.19868934154510498, + 0.310660183429718, + 0.050388187170028687, + 0.0833999440073967, + 0.05546831712126732, + 0.07876782864332199, + 0.025330014526844025, + 0.05655006691813469 + ], + "credit_map": [ + 0.011206826195120811, + 0.011151457205414772, + 0.008312533609569073, + 0.006583135575056076, + 0.011172834783792496, + 0.01260883267968893, + 0.01107421051710844, + 0.013059081509709358, + 0.004742790013551712, + 0.006281693931668997, + 0.006790436338633299, + 0.006514121778309345, + 0.0077116177417337894, + 0.010520813055336475, + 0.0056368024088442326, + 0.0073789688758552074, + 0.016253270208835602, + 0.026629658415913582, + 0.010397336445748806, + 0.056305162608623505, + 0.011347330175340176, + 0.021855786442756653, + 0.01182987168431282, + 0.044070906937122345, + 0.03292129933834076, + 0.12990960478782654, + 0.014520096592605114, + 0.06770573556423187, + 0.01535209733992815, + 0.019782956689596176, + 0.02204742282629013, + 0.008420933037996292, + 0.006338098552078009, + 0.007340544834733009, + 0.009826304391026497, + 0.005336228758096695, + 0.006363478489220142, + 0.006103002931922674, + 0.004652867559343576, + 0.004427652806043625, + 0.011594126001000404, + 0.010906117036938667, + 0.01146837417036295, + 0.007548199966549873, + 0.006418878678232431, + 0.014976701699197292, + 0.008225668221712112, + 0.012261864729225636, + 0.008466603234410286, + 0.004843825474381447, + 0.007994920946657658, + 0.0060466923750936985, + 0.0050614760257303715, + 0.007533527445048094, + 0.004390064161270857, + 0.01113120187073946, + 0.010652930475771427, + 0.010542082600295544, + 0.00904431939125061, + 0.014689500443637371, + 0.013067780062556267, + 0.005406184121966362, + 0.006713066250085831, + 0.008651948533952236, + 0.0028071722481399775, + 0.003999513108283281, + 0.0019501911010593176, + 0.0023099773097783327, + 0.003908052109181881, + 0.004773084539920092, + 0.0025126903783529997, + 0.0030693302396684885, + 0.008119449019432068, + 0.005964267998933792, + 0.008416024036705494, + 0.005430129356682301, + 0.006596095860004425, + 0.011847433634102345, + 0.00472364341840148, + 0.005453013814985752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.03808846324682236 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_12/index_tensor([3641, 3641])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_12/index_tensor([3641, 3641])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..219e3ef3ce73d1817eea129a8d39256ce5467d58 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_12/index_tensor([3641, 3641])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "magnificent long shot, captured from a slightly upward angle, the camera held still, captures a vast expanse of the sky. Against this crystal-clear, nearly imperfect blue sky, vast, fluffy white cumulus clouds, like cotton candy, float majestically. Illuminated by the bright afternoon sunlight, these vast clouds are edged with a dazzling golden hue, while their interiors are awash in varying shades of gray, accentuating their astonishing volume and three-dimensional depth. With an almost imperceptible and leisurely pace, they drift slowly toward the rolling mountain ranges on the horizon, their contours blurred by the distance, a soft blue-gray hue. Bathed in intense, clear sunlight, the entire scene, with a bright, warm color temperature, creates an atmosphere of profound tranquility, spaciousness, and grandeur, offering a truly refreshing visual experience.", + "seed": 32925, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_12/index_3641_seed_32925_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.019925189716741443, + "vlm_logprob": -0.019925189716741443, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.019925189716741443, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003114256775006652 + ], + "opposite_logprob_tokens": [ + -7.378114223480225 + ], + "logprob_reward": -0.003114256775006652, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.003114256775006652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12288999557495117, + 0.05837492272257805, + 0.08754736930131912, + 0.17742149531841278, + 0.0830933228135109, + 0.053011517971754074, + 0.09558098763227463, + 0.04473559185862541, + 0.12748302519321442, + 0.14986176788806915 + ], + "credit_map": [ + 0.013481811620295048, + 0.014680608175694942, + 0.01645168475806713, + 0.010294901207089424, + 0.007836663164198399, + 0.008716278709471226, + 0.01374648418277502, + 0.03768155723810196, + 0.004792872816324234, + 0.00616027694195509, + 0.012767791748046875, + 0.005945094861090183, + 0.0032903170213103294, + 0.003388616256415844, + 0.005637322086840868, + 0.016392627730965614, + 0.01258157566189766, + 0.02167467586696148, + 0.012580718845129013, + 0.0060567064210772514, + 0.005785548593848944, + 0.003913540858775377, + 0.006595938000828028, + 0.018358660861849785, + 0.012895657680928707, + 0.01969699002802372, + 0.031454071402549744, + 0.016454992815852165, + 0.010838601738214493, + 0.01080759335309267, + 0.01785583235323429, + 0.057417724281549454, + 0.006796140223741531, + 0.01276674959808588, + 0.026157980784773827, + 0.008206797763705254, + 0.00585411349311471, + 0.0037729640025645494, + 0.007371022831648588, + 0.012167549692094326, + 0.008227021433413029, + 0.013754650950431824, + 0.006940038874745369, + 0.004010058008134365, + 0.00299362582154572, + 0.0026846714317798615, + 0.0038611548952758312, + 0.010540289804339409, + 0.009539124555885792, + 0.009747154079377651, + 0.02601650543510914, + 0.01050198171287775, + 0.006680649239569902, + 0.004969159606844187, + 0.010773660615086555, + 0.017352748662233353, + 0.004639716818928719, + 0.003931715618818998, + 0.007582970429211855, + 0.005424820352345705, + 0.004362361039966345, + 0.0030398061498999596, + 0.0045563396997749805, + 0.011197857558727264, + 0.01181410625576973, + 0.014350145123898983, + 0.02930803783237934, + 0.013385976664721966, + 0.00874769315123558, + 0.009944144636392593, + 0.018089964985847473, + 0.0218429546803236, + 0.011733894236385822, + 0.013329366222023964, + 0.02927607297897339, + 0.015090794302523136, + 0.0138221625238657, + 0.008918412029743195, + 0.019424930214881897, + 0.038266126066446304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.003114256775006652 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607716642320156 + ], + "opposite_logprob_tokens": [ + -4.141077041625977 + ], + "logprob_reward": -0.01607716642320156, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607716642320156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047117333859205246, + 0.13428911566734314, + 0.268602579832077, + 0.15925438702106476, + 0.12589429318904877, + 0.03158733621239662, + 0.049541108310222626, + 0.01353402342647314, + 0.026977162808179855, + 0.14320267736911774 + ], + "credit_map": [ + 0.003481004387140274, + 0.009211907163262367, + 0.011341709643602371, + 0.0035514438059180975, + 0.0031995181925594807, + 0.0027288878336548805, + 0.004741079639643431, + 0.008861783891916275, + 0.0107522401958704, + 0.015029709786176682, + 0.02055179327726364, + 0.012668591924011707, + 0.008799560368061066, + 0.010414703749120235, + 0.016166849061846733, + 0.03990565240383148, + 0.02535163424909115, + 0.03689258545637131, + 0.04929761216044426, + 0.024205999448895454, + 0.018280362710356712, + 0.020593734458088875, + 0.031694017350673676, + 0.062286633998155594, + 0.010673655197024345, + 0.016700124368071556, + 0.029840948060154915, + 0.015415985137224197, + 0.008226477541029453, + 0.009696133434772491, + 0.01683192513883114, + 0.05186915025115013, + 0.01647011749446392, + 0.02550790086388588, + 0.02857062965631485, + 0.010391497984528542, + 0.008398928679525852, + 0.005694063846021891, + 0.010781205259263515, + 0.020079951733350754, + 0.002605594927445054, + 0.004735796712338924, + 0.0059873485006392, + 0.0032328611705452204, + 0.0016279187984764576, + 0.0018244831589981914, + 0.003286918858066201, + 0.008286413736641407, + 0.005169241223484278, + 0.00551029434427619, + 0.01262841559946537, + 0.00524234538897872, + 0.003090423997491598, + 0.002638427074998617, + 0.006186243612319231, + 0.00907572079449892, + 0.0013849202077835798, + 0.0017373174196109176, + 0.0030336661729961634, + 0.0020776232704520226, + 0.0012921873712912202, + 0.0008877982618287206, + 0.0012563458876684308, + 0.001864164019934833, + 0.0023723600897938013, + 0.002839721506461501, + 0.0032203414011746645, + 0.0027923278976231813, + 0.002351426985114813, + 0.0017411697190254927, + 0.0037279303651303053, + 0.007931885309517384, + 0.01317260880023241, + 0.015256877057254314, + 0.02860819734632969, + 0.01350310631096363, + 0.012399866245687008, + 0.008326739072799683, + 0.017723705619573593, + 0.034211572259664536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.01607716642320156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638040162622929 + ], + "opposite_logprob_tokens": [ + -4.75863790512085 + ], + "logprob_reward": -0.008638040162622929, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638040162622929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14992791414260864, + 0.06645193696022034, + 0.13003729283809662, + 0.10497517138719559, + 0.07875137031078339, + 0.04745516926050186, + 0.054684579372406006, + 0.035676807165145874, + 0.05570582300424576, + 0.2763339579105377 + ], + "credit_map": [ + 0.016787011176347733, + 0.01846858114004135, + 0.01886158436536789, + 0.018380403518676758, + 0.009642371907830238, + 0.012029078789055347, + 0.014394496567547321, + 0.0413643978536129, + 0.007890847511589527, + 0.009545112028717995, + 0.009086333215236664, + 0.00660968292504549, + 0.004005763214081526, + 0.004542618989944458, + 0.006940792314708233, + 0.01783078722655773, + 0.019772373139858246, + 0.02579174004495144, + 0.0178671982139349, + 0.009357074275612831, + 0.012344220653176308, + 0.009767704643309116, + 0.010356279090046883, + 0.02478071115911007, + 0.0076830764301121235, + 0.012733054347336292, + 0.017919139936566353, + 0.010721040889620781, + 0.0056168027222156525, + 0.006208482198417187, + 0.011037245392799377, + 0.033056341111660004, + 0.0068724146112799644, + 0.010344634763896465, + 0.02269945666193962, + 0.008229206316173077, + 0.0054748631082475185, + 0.0037241801619529724, + 0.007843864150345325, + 0.013562750071287155, + 0.004551354795694351, + 0.007700009271502495, + 0.00941628497093916, + 0.005474856588989496, + 0.0026423116214573383, + 0.0025766075123101473, + 0.00402821134775877, + 0.011065536178648472, + 0.005708921235054731, + 0.006345669738948345, + 0.014508302323520184, + 0.006211320403963327, + 0.0034954766742885113, + 0.0023051833268254995, + 0.005211902782320976, + 0.010897804982960224, + 0.003440259024500847, + 0.0033134385012090206, + 0.004287919960916042, + 0.003633299842476845, + 0.006272042170166969, + 0.004262245260179043, + 0.0036170529201626778, + 0.006850553210824728, + 0.00857497751712799, + 0.00895229447633028, + 0.006160274147987366, + 0.004247541539371014, + 0.005383588839322329, + 0.004759124480187893, + 0.007617361843585968, + 0.01001066341996193, + 0.025121042504906654, + 0.029259465634822845, + 0.05925207585096359, + 0.024967124685645103, + 0.024882987141609192, + 0.016387006267905235, + 0.03447441756725311, + 0.06198985129594803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.008638040162622929 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202251464128494 + ], + "opposite_logprob_tokens": [ + -2.8120224475860596 + ], + "logprob_reward": -0.06202251464128494, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202251464128494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06484292447566986, + 0.1436278223991394, + 0.19385407865047455, + 0.08948187530040741, + 0.13820196688175201, + 0.0188224408775568, + 0.04504000023007393, + 0.01772419549524784, + 0.029651068150997162, + 0.25875359773635864 + ], + "credit_map": [ + 0.0045926328748464584, + 0.006275392137467861, + 0.010685707442462444, + 0.008421174250543118, + 0.005264353472739458, + 0.005135477986186743, + 0.006800177041441202, + 0.017668016254901886, + 0.010396934114396572, + 0.014474546536803246, + 0.024118851870298386, + 0.013798053376376629, + 0.009450729005038738, + 0.01103654783219099, + 0.017326069995760918, + 0.043026093393564224, + 0.015151891857385635, + 0.02091076970100403, + 0.03984874486923218, + 0.018170850351452827, + 0.013347234576940536, + 0.014589229598641396, + 0.023080214858055115, + 0.04875514656305313, + 0.007139246445149183, + 0.012735489755868912, + 0.018775876611471176, + 0.008351493626832962, + 0.004974375944584608, + 0.004659594502300024, + 0.00798468105494976, + 0.02486111782491207, + 0.015973230823874474, + 0.02906341664493084, + 0.0356089249253273, + 0.010987449437379837, + 0.007499086204916239, + 0.0054561379365623, + 0.009862568229436874, + 0.023751141503453255, + 0.0012148184468969703, + 0.001734837074764073, + 0.003135464619845152, + 0.0021531106904149055, + 0.000948784057982266, + 0.0013418318703770638, + 0.002307143760845065, + 0.005986450705677271, + 0.004379871767014265, + 0.006996325217187405, + 0.012231174856424332, + 0.004132275469601154, + 0.0027319500222802162, + 0.0017616223776713014, + 0.0038662690203636885, + 0.008940515108406544, + 0.0017936036456376314, + 0.002004004782065749, + 0.0040910434909164906, + 0.002411819528788328, + 0.0019750865176320076, + 0.0012073215330019593, + 0.0016060651978477836, + 0.0026352512650191784, + 0.004232001956552267, + 0.004500065464526415, + 0.005693269427865744, + 0.0028372693341225386, + 0.0021782417315989733, + 0.002426475053653121, + 0.004045914858579636, + 0.003737831488251686, + 0.02667008899152279, + 0.03239723667502403, + 0.04951589182019234, + 0.02286551333963871, + 0.02163550816476345, + 0.014110499061644077, + 0.03103272244334221, + 0.06052612513303757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.06202251464128494 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977397058159113 + ], + "opposite_logprob_tokens": [ + -4.634774208068848 + ], + "logprob_reward": -0.00977397058159113, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977397058159113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1620407998561859, + 0.08804091066122055, + 0.10930777341127396, + 0.18324188888072968, + 0.11943719536066055, + 0.13387912511825562, + 0.02284710481762886, + 0.031180420890450478, + 0.03701044246554375, + 0.11301431059837341 + ], + "credit_map": [ + 0.031461719423532486, + 0.032528817653656006, + 0.019785072654485703, + 0.011950029991567135, + 0.01364102028310299, + 0.01328675914555788, + 0.01534637063741684, + 0.024041002616286278, + 0.01926814205944538, + 0.022417737171053886, + 0.01167422253638506, + 0.006089269649237394, + 0.007029298227280378, + 0.006278980989009142, + 0.006439150311052799, + 0.008844109252095222, + 0.007655384484678507, + 0.010966223664581776, + 0.025603037327528, + 0.010644884780049324, + 0.007283488754183054, + 0.006853758357465267, + 0.010653291828930378, + 0.0296477060765028, + 0.013843991793692112, + 0.01844250224530697, + 0.03565915301442146, + 0.01947714388370514, + 0.009317331947386265, + 0.011204778216779232, + 0.01959577389061451, + 0.05570119991898537, + 0.011187825351953506, + 0.01783195696771145, + 0.017816290259361267, + 0.01073778048157692, + 0.007096590008586645, + 0.007245339918881655, + 0.011070869863033295, + 0.03645053505897522, + 0.008727945387363434, + 0.017978575080633163, + 0.025213632732629776, + 0.012878749519586563, + 0.0058152214623987675, + 0.007654689252376556, + 0.014950781129300594, + 0.040659524500370026, + 0.0026626363396644592, + 0.004597866907715797, + 0.005065020173788071, + 0.002310631563887, + 0.0019672864582389593, + 0.0012528636725619435, + 0.001808825647458434, + 0.003181973472237587, + 0.003993607591837645, + 0.006080430466681719, + 0.0055940281599760056, + 0.0024041011929512024, + 0.002043184358626604, + 0.001579941250383854, + 0.0026198674459010363, + 0.006865260191261768, + 0.004138298332691193, + 0.004569715820252895, + 0.007210182026028633, + 0.0035930857993662357, + 0.0025108351837843657, + 0.0028574520256370306, + 0.0042908089235424995, + 0.007840062491595745, + 0.010807784274220467, + 0.012083163484930992, + 0.028840813785791397, + 0.01098901592195034, + 0.010864385403692722, + 0.0066354526206851006, + 0.012771847657859325, + 0.02002185583114624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.00977397058159113 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "magnificent long shot, captured from a slightly upward angle, the camera held still, captures a vast expanse of the sky. Against this crystal-clear, nearly imperfect blue sky, vast, fluffy white cumulus clouds, like cotton candy, float majestically. Illuminated by the bright afternoon sunlight, these vast clouds are edged with a dazzling golden hue, while their interiors are awash in varying shades of gray, accentuating their astonishing volume and three-dimensional depth. With an almost imperceptible and leisurely pace, they drift slowly toward the rolling mountain ranges on the horizon, their contours blurred by the distance, a soft blue-gray hue. Bathed in intense, clear sunlight, the entire scene, with a bright, warm color temperature, creates an atmosphere of profound tranquility, spaciousness, and grandeur, offering a truly refreshing visual experience.", + "seed": 32926, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_12/index_3641_seed_32926_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.006236751936376095, + "vlm_logprob": -0.006236751936376095, + "vlm_yes_no_margin": 5.775, + "vlm_token_credit": -0.006236751936376095, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003420457011088729 + ], + "opposite_logprob_tokens": [ + -8.37842082977295 + ], + "logprob_reward": -0.003420457011088729, + "yes_no_margin_reward": 8.375, + "token_credit_reward": -0.003420457011088729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5017554759979248, + 0.09389603137969971, + 0.10783243179321289, + 0.022204279899597168, + 0.02623572014272213, + 0.033546894788742065, + 0.033924806863069534, + 0.07447604089975357, + 0.05575946718454361, + 0.05036882683634758 + ], + "credit_map": [ + 0.026615288108587265, + 0.11468052119016647, + 0.1884966790676117, + 0.08868606388568878, + 0.022334719076752663, + 0.015151617117226124, + 0.022180495783686638, + 0.023610100150108337, + 0.004577168729156256, + 0.013500197790563107, + 0.019587066024541855, + 0.032257430255413055, + 0.0037524537183344364, + 0.0037112829741090536, + 0.010789831168949604, + 0.0057205986231565475, + 0.006125359795987606, + 0.01696074567735195, + 0.03188137337565422, + 0.02136678248643875, + 0.005109039600938559, + 0.006376889534294605, + 0.012181823141872883, + 0.007830427959561348, + 0.0022565280087292194, + 0.0025969590060412884, + 0.002978957025334239, + 0.0027878652326762676, + 0.0022652321495115757, + 0.0030768285505473614, + 0.002783027943223715, + 0.00345888314768672, + 0.002599148778244853, + 0.003845811355859041, + 0.004381334409117699, + 0.0037308549508452415, + 0.0027261006180197, + 0.002577420324087143, + 0.003280741162598133, + 0.003094307379797101, + 0.0028289540205150843, + 0.0037568460684269667, + 0.004732323344796896, + 0.0042137885466217995, + 0.0034076899755746126, + 0.003078196197748184, + 0.005292590241879225, + 0.006236506160348654, + 0.002522510476410389, + 0.0031004745978862047, + 0.007176753133535385, + 0.005694071762263775, + 0.0036261489149183035, + 0.0028983817901462317, + 0.003981656860560179, + 0.004924805369228125, + 0.008297904394567013, + 0.013076430186629295, + 0.010618520900607109, + 0.009663681499660015, + 0.008634076453745365, + 0.006213813088834286, + 0.009595555253326893, + 0.008376060053706169, + 0.00524806696921587, + 0.005458108149468899, + 0.008413492701947689, + 0.008450943976640701, + 0.007238514255732298, + 0.0037153582088649273, + 0.00927799567580223, + 0.007956987246870995, + 0.004647810477763414, + 0.01067358162254095, + 0.006280636414885521, + 0.010013090446591377, + 0.003844121005386114, + 0.0031268615275621414, + 0.006953422911465168, + 0.004829301964491606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.003420457011088729 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080282740294933 + ], + "opposite_logprob_tokens": [ + -4.511080265045166 + ], + "logprob_reward": -0.011080282740294933, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080282740294933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16125459969043732, + 0.25035756826400757, + 0.08559872955083847, + 0.06033606827259064, + 0.04162123426795006, + 0.041362613439559937, + 0.07364833354949951, + 0.05399436503648758, + 0.12229223549365997, + 0.10953421890735626 + ], + "credit_map": [ + 0.010993462055921555, + 0.034103184938430786, + 0.041384510695934296, + 0.01959880255162716, + 0.00998193770647049, + 0.012279948219656944, + 0.019673598930239677, + 0.013239149004220963, + 0.0152612728998065, + 0.018750792369246483, + 0.0569184236228466, + 0.08413650095462799, + 0.011008527129888535, + 0.012214451096951962, + 0.028966231271624565, + 0.023101357743144035, + 0.0049207499250769615, + 0.011446995660662651, + 0.025789866223931313, + 0.01744993031024933, + 0.00399350468069315, + 0.0045652990229427814, + 0.010793614201247692, + 0.006638770457357168, + 0.005383755080401897, + 0.008043896406888962, + 0.013428529724478722, + 0.008534413762390614, + 0.006083021406084299, + 0.00580475851893425, + 0.00638617156073451, + 0.006671518087387085, + 0.0049817063845694065, + 0.005408892408013344, + 0.007661846000701189, + 0.0049217622727155685, + 0.004457390401512384, + 0.0034950701519846916, + 0.0055727665312588215, + 0.005121801048517227, + 0.003763399785384536, + 0.004515921231359243, + 0.007967672310769558, + 0.005786113440990448, + 0.004424050450325012, + 0.003430775133892894, + 0.005531701259315014, + 0.005942979361861944, + 0.0065667470917105675, + 0.007403945084661245, + 0.014856088906526566, + 0.011687858030200005, + 0.0084649957716465, + 0.006261708214879036, + 0.007535628043115139, + 0.010871357284486294, + 0.004018145613372326, + 0.0055224173702299595, + 0.008090387098491192, + 0.010564898140728474, + 0.004274983424693346, + 0.004529772326350212, + 0.00872367899864912, + 0.008270081132650375, + 0.010445899330079556, + 0.015670666471123695, + 0.01682758331298828, + 0.01880635693669319, + 0.014117305167019367, + 0.008685946464538574, + 0.021380456164479256, + 0.0163580235093832, + 0.014776812866330147, + 0.038698505610227585, + 0.005691262893378735, + 0.008747247979044914, + 0.009320034645497799, + 0.008930535987019539, + 0.015172277577221394, + 0.008197548799216747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.011080282740294933 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040965694934129715 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.0040965694934129715, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040965694934129715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7667118906974792, + 0.03268905356526375, + 0.06421521306037903, + 0.009168067015707493, + 0.014277922920882702, + 0.020834799855947495, + 0.019786566495895386, + 0.0288129523396492, + 0.02216159924864769, + 0.021341891959309578 + ], + "credit_map": [ + 0.04575749859213829, + 0.19862203299999237, + 0.28045186400413513, + 0.12386465072631836, + 0.03956335037946701, + 0.019849233329296112, + 0.03225412219762802, + 0.026349155232310295, + 0.0018771819304674864, + 0.004231561440974474, + 0.007917618378996849, + 0.007526646368205547, + 0.002193966181948781, + 0.00245880801230669, + 0.003806052729487419, + 0.002677219221368432, + 0.00398488761857152, + 0.01142961997538805, + 0.01703592948615551, + 0.01293743122369051, + 0.003344547236338258, + 0.003285337472334504, + 0.007694619242101908, + 0.004502845928072929, + 0.0007282518781721592, + 0.0010879379697144032, + 0.0014785848325118423, + 0.0012226892868056893, + 0.0008258629241026938, + 0.0009217329788953066, + 0.0014542312128469348, + 0.0014487761072814465, + 0.0013475064188241959, + 0.0028269868344068527, + 0.0019138032803311944, + 0.0018417438259348273, + 0.0015550891403108835, + 0.001179578248411417, + 0.0013772961683571339, + 0.0022359180729836226, + 0.0017245425842702389, + 0.0028967077378183603, + 0.002888469025492668, + 0.002350030466914177, + 0.001259203883819282, + 0.0015998734161257744, + 0.0036571098025888205, + 0.0044588628225028515, + 0.0014468168374150991, + 0.002042048145085573, + 0.0037662978284060955, + 0.004843790549784899, + 0.001937610562890768, + 0.0016687088645994663, + 0.0024711638689041138, + 0.0016101292567327619, + 0.0025066493544727564, + 0.0056653390638530254, + 0.00473009841516614, + 0.004573130048811436, + 0.002203721320256591, + 0.0026688124053180218, + 0.003466528607532382, + 0.002998676151037216, + 0.0018838002579286695, + 0.004422585945576429, + 0.002274492522701621, + 0.002743576653301716, + 0.0020169750787317753, + 0.002284813905134797, + 0.003627258352935314, + 0.0029080957174301147, + 0.0025040400214493275, + 0.006907079368829727, + 0.001889896928332746, + 0.002130181761458516, + 0.0016045247903093696, + 0.0015648802509531379, + 0.001743418164551258, + 0.002997871022671461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0040965694934129715 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002812479855492711 + ], + "opposite_logprob_tokens": [ + -5.877812385559082 + ], + "logprob_reward": -0.002812479855492711, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002812479855492711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.522477388381958, + 0.1590646207332611, + 0.1616629660129547, + 0.02155701443552971, + 0.011257077567279339, + 0.04813525825738907, + 0.008660652674734592, + 0.019625119864940643, + 0.01885337010025978, + 0.02870647795498371 + ], + "credit_map": [ + 0.03188852593302727, + 0.13758128881454468, + 0.18952180445194244, + 0.07882225513458252, + 0.027569754049181938, + 0.012035059742629528, + 0.02358260005712509, + 0.021476121619343758, + 0.013147832825779915, + 0.02697228454053402, + 0.021725637838244438, + 0.02499488554894924, + 0.009704412892460823, + 0.009143698029220104, + 0.02802380546927452, + 0.025352073833346367, + 0.010394372045993805, + 0.02941848523914814, + 0.045020781457424164, + 0.030126597732305527, + 0.008042718283832073, + 0.008476024493575096, + 0.018551668152213097, + 0.01163233257830143, + 0.002125087194144726, + 0.0027707875706255436, + 0.003740755608305335, + 0.003747728420421481, + 0.001597037655301392, + 0.0022041567135602236, + 0.0028153008315712214, + 0.0025561631191521883, + 0.0011132529471069574, + 0.0011927410960197449, + 0.0016302545554935932, + 0.0019971702713519335, + 0.0015555295394733548, + 0.0012149332324042916, + 0.0009498444269411266, + 0.0016033523716032505, + 0.005020394921302795, + 0.013073802925646305, + 0.005971346516162157, + 0.005279292818158865, + 0.0035082963295280933, + 0.004632848780602217, + 0.005439160857349634, + 0.005210119299590588, + 0.0007350528612732887, + 0.0011529921321198344, + 0.0009450523648411036, + 0.0013249164912849665, + 0.0007569832960143685, + 0.0006406111642718315, + 0.001638315268792212, + 0.001466729212552309, + 0.001463819295167923, + 0.0029700675513595343, + 0.00325389439240098, + 0.002617122605443001, + 0.00202915258705616, + 0.001372791244648397, + 0.0032048290595412254, + 0.0027134460397064686, + 0.0013187642907723784, + 0.0017141697462648153, + 0.0023528076708316803, + 0.002738002222031355, + 0.0013549748109653592, + 0.0011708239326253533, + 0.003198815742507577, + 0.005005014594644308, + 0.002675337018445134, + 0.007005255203694105, + 0.003201160579919815, + 0.0047329203225672245, + 0.0018467411864548922, + 0.0019319102866575122, + 0.0031746418680995703, + 0.004138515330851078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.002812479855492711 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977397058159113 + ], + "opposite_logprob_tokens": [ + -4.634774208068848 + ], + "logprob_reward": -0.00977397058159113, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977397058159113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.754825234413147, + 0.05514247715473175, + 0.019131537526845932, + 0.02192869782447815, + 0.032259900122880936, + 0.027709219604730606, + 0.03838784992694855, + 0.013496451079845428, + 0.017193187028169632, + 0.019925516098737717 + ], + "credit_map": [ + 0.04416171461343765, + 0.19140885770320892, + 0.28180140256881714, + 0.12341715395450592, + 0.03685147315263748, + 0.016469642519950867, + 0.03295053914189339, + 0.027764368802309036, + 0.002982799895107746, + 0.010088290087878704, + 0.011077631264925003, + 0.01831943355500698, + 0.0026603436563163996, + 0.002009349875152111, + 0.004674617666751146, + 0.0033300158102065325, + 0.0015010869828984141, + 0.003258289536461234, + 0.0025993213057518005, + 0.003631872357800603, + 0.0017667094944044948, + 0.0020759571343660355, + 0.0028156652115285397, + 0.0014826359692960978, + 0.0017692332621663809, + 0.0029478813521564007, + 0.005021634045988321, + 0.003181633772328496, + 0.0021281021181493998, + 0.0027114516124129295, + 0.0018541492754593492, + 0.0023146127350628376, + 0.0029950616881251335, + 0.005085862707346678, + 0.00592124043032527, + 0.004014205653220415, + 0.0033476464450359344, + 0.0028631226159632206, + 0.0037132238503545523, + 0.004319536965340376, + 0.0028069624677300453, + 0.006459019612520933, + 0.0030340426601469517, + 0.002565873321145773, + 0.003145328490063548, + 0.003098729532212019, + 0.003500567516312003, + 0.003098697168752551, + 0.002686497289687395, + 0.003373670857399702, + 0.008112546056509018, + 0.007661305367946625, + 0.0036366700660437346, + 0.002290561329573393, + 0.005298890173435211, + 0.005327713210135698, + 0.0009910025401040912, + 0.0019751752261072397, + 0.0023440399672836065, + 0.002416186733171344, + 0.001193189062178135, + 0.0010996575001627207, + 0.0019157312344759703, + 0.001561469747684896, + 0.0015859496779739857, + 0.0034321562852710485, + 0.002424277365207672, + 0.002564189489930868, + 0.002319194609299302, + 0.0015944549813866615, + 0.0020067014265805483, + 0.0012662633089348674, + 0.0019135173643007874, + 0.005810482893139124, + 0.001964639173820615, + 0.002890622476115823, + 0.0018824386643245816, + 0.0015220133354887366, + 0.0020472549367696047, + 0.0018945485353469849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.00977397058159113 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "magnificent long shot, captured from a slightly upward angle, the camera held still, captures a vast expanse of the sky. Against this crystal-clear, nearly imperfect blue sky, vast, fluffy white cumulus clouds, like cotton candy, float majestically. Illuminated by the bright afternoon sunlight, these vast clouds are edged with a dazzling golden hue, while their interiors are awash in varying shades of gray, accentuating their astonishing volume and three-dimensional depth. With an almost imperceptible and leisurely pace, they drift slowly toward the rolling mountain ranges on the horizon, their contours blurred by the distance, a soft blue-gray hue. Bathed in intense, clear sunlight, the entire scene, with a bright, warm color temperature, creates an atmosphere of profound tranquility, spaciousness, and grandeur, offering a truly refreshing visual experience.", + "seed": 32927, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_12/index_3641_seed_32927_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.017343430407345296, + "vlm_logprob": -0.017343430407345296, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.017343430407345296, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005678593181073666 + ], + "opposite_logprob_tokens": [ + -7.755678653717041 + ], + "logprob_reward": -0.005678593181073666, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.005678593181073666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0522318072617054, + 0.04113730415701866, + 0.018514227122068405, + 0.03838196396827698, + 0.06136854365468025, + 0.030749201774597168, + 0.022307416424155235, + 0.520301103591919, + 0.06976035982370377, + 0.14524801075458527 + ], + "credit_map": [ + 0.005443776957690716, + 0.00847185030579567, + 0.007744348142296076, + 0.004126626066863537, + 0.0045023211278021336, + 0.004936674144119024, + 0.008342485874891281, + 0.008663730695843697, + 0.00527431583032012, + 0.004733051639050245, + 0.004450538661330938, + 0.003206273540854454, + 0.007332120556384325, + 0.0028563409578055143, + 0.005421377718448639, + 0.007863291539251804, + 0.002222054870799184, + 0.002169467741623521, + 0.001758758444339037, + 0.001323373755440116, + 0.002282459521666169, + 0.0025035289581865072, + 0.0020502260886132717, + 0.00420436030253768, + 0.004666670225560665, + 0.0037952042184770107, + 0.005826170556247234, + 0.0042067524045705795, + 0.005818540696054697, + 0.0035714730620384216, + 0.004917279817163944, + 0.005579880438745022, + 0.00666928943246603, + 0.01257212832570076, + 0.004516271408647299, + 0.005223836749792099, + 0.006938850041478872, + 0.010372309945523739, + 0.0072179618291556835, + 0.007857901975512505, + 0.0038782984483987093, + 0.005818820092827082, + 0.0024292366579174995, + 0.0029072267934679985, + 0.002544211223721504, + 0.002509467303752899, + 0.004357416182756424, + 0.006304529029875994, + 0.0027028401382267475, + 0.002854092512279749, + 0.003933485131710768, + 0.0023448183201253414, + 0.0026659853756427765, + 0.002461981028318405, + 0.002504160162061453, + 0.002840055851265788, + 0.1076551005244255, + 0.016298683360219002, + 0.026729803532361984, + 0.02554856240749359, + 0.031471434980630875, + 0.018349245190620422, + 0.23194341361522675, + 0.06230493262410164, + 0.010039993561804295, + 0.005490100476890802, + 0.0073794154450297356, + 0.00951426662504673, + 0.007368333637714386, + 0.009401502087712288, + 0.011149600148200989, + 0.009417157620191574, + 0.01577799580991268, + 0.009214316494762897, + 0.011120105162262917, + 0.0237551499158144, + 0.00866137258708477, + 0.010814623907208443, + 0.0282481387257576, + 0.03765631839632988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.005678593181073666 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012551246210932732 + ], + "opposite_logprob_tokens": [ + -4.387551307678223 + ], + "logprob_reward": -0.012551246210932732, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012551246210932732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04973910376429558, + 0.045234158635139465, + 0.290976881980896, + 0.010126384906470776, + 0.03363809362053871, + 0.021602269262075424, + 0.01836697943508625, + 0.3224421739578247, + 0.09639428555965424, + 0.11147967725992203 + ], + "credit_map": [ + 0.004610787145793438, + 0.004745367914438248, + 0.006891163531690836, + 0.006208920385688543, + 0.004767062608152628, + 0.004113947041332722, + 0.010352742858231068, + 0.008049110881984234, + 0.005549338646233082, + 0.0056694322265684605, + 0.004885681439191103, + 0.0037025590427219868, + 0.007720086723566055, + 0.0036984956823289394, + 0.0062570334412157536, + 0.007751531898975372, + 0.03353457897901535, + 0.05037875100970268, + 0.04342295974493027, + 0.021265728399157524, + 0.03880879655480385, + 0.019821543246507645, + 0.04623943567276001, + 0.03750508651137352, + 0.0011630639201030135, + 0.0010649434989318252, + 0.0016505079111084342, + 0.0009069142979569733, + 0.0011545860907062888, + 0.0013194113271310925, + 0.00143870385363698, + 0.001428254647180438, + 0.004463803488761187, + 0.008651084266602993, + 0.003413077676668763, + 0.0036773316096514463, + 0.003683524439111352, + 0.0039121489971876144, + 0.0029498166404664516, + 0.0028873055707663298, + 0.0023522586561739445, + 0.0029920681845396757, + 0.002302787499502301, + 0.0018887064652517438, + 0.0016207318985834718, + 0.002345551270991564, + 0.00337191647849977, + 0.004728249274194241, + 0.0021471520885825157, + 0.0022636614739894867, + 0.001652962644584477, + 0.0016541102668270469, + 0.0013962893281131983, + 0.0010729701025411487, + 0.005079901311546564, + 0.003099932102486491, + 0.0649014338850975, + 0.012525293976068497, + 0.017287008464336395, + 0.015631183981895447, + 0.021521545946598053, + 0.01799537055194378, + 0.13663719594478607, + 0.03594314679503441, + 0.013932737521827221, + 0.008152706548571587, + 0.007333934307098389, + 0.010962286032736301, + 0.007662417367100716, + 0.009668274782598019, + 0.02191259153187275, + 0.016769342124462128, + 0.014897425659000874, + 0.009740225039422512, + 0.010718259960412979, + 0.020041698589920998, + 0.009796873666346073, + 0.008187077939510345, + 0.02253853715956211, + 0.015559572726488113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.012551246210932732 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024909917265176773 + ], + "opposite_logprob_tokens": [ + -6.002490997314453 + ], + "logprob_reward": -0.0024909917265176773, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024909917265176773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14588457345962524, + 0.03147333487868309, + 0.24864137172698975, + 0.027184983715415, + 0.09217539429664612, + 0.025094222277402878, + 0.01788913644850254, + 0.28851085901260376, + 0.04699647054076195, + 0.07614961266517639 + ], + "credit_map": [ + 0.013931485824286938, + 0.018875334411859512, + 0.013617299497127533, + 0.014914825558662415, + 0.013393747620284557, + 0.01480601541697979, + 0.0314517617225647, + 0.0248941108584404, + 0.003503497689962387, + 0.004158154595643282, + 0.003597729606553912, + 0.002685355953872204, + 0.0028843518812209368, + 0.004450518637895584, + 0.006119324825704098, + 0.004074405413120985, + 0.02848854847252369, + 0.042144082486629486, + 0.035289518535137177, + 0.017785249277949333, + 0.03469111770391464, + 0.016664555296301842, + 0.03848402947187424, + 0.035094283521175385, + 0.003029722487553954, + 0.004925806540995836, + 0.0024010075721889734, + 0.0034501408226788044, + 0.003365654731169343, + 0.002723971614614129, + 0.003422059118747711, + 0.003866623155772686, + 0.010941529646515846, + 0.02118595503270626, + 0.007409871555864811, + 0.008167971856892109, + 0.011207105591893196, + 0.016724420711398125, + 0.008087490685284138, + 0.008451053872704506, + 0.0033033203799277544, + 0.0030225017108023167, + 0.002548439893871546, + 0.0029503912664949894, + 0.001967519987374544, + 0.0026011180598288774, + 0.004329297691583633, + 0.004371635150164366, + 0.0019116669427603483, + 0.0027771873865276575, + 0.0013621927937492728, + 0.002128184773027897, + 0.0012240653159096837, + 0.0017224407056346536, + 0.0026413241866976023, + 0.004122076090425253, + 0.05983056128025055, + 0.01125145610421896, + 0.015269220806658268, + 0.014449115842580795, + 0.016847895458340645, + 0.011362467892467976, + 0.12601938843727112, + 0.033480770885944366, + 0.006378809921443462, + 0.0034564754460006952, + 0.003263865364715457, + 0.0050309100188314915, + 0.003728162497282028, + 0.0042935749515891075, + 0.012211895547807217, + 0.008632780984044075, + 0.009264692664146423, + 0.0070910085923969746, + 0.007324235513806343, + 0.013967814855277538, + 0.006726343184709549, + 0.007930096238851547, + 0.013057897798717022, + 0.010787530802190304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.0024909917265176773 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05492936074733734 + ], + "opposite_logprob_tokens": [ + -2.929929256439209 + ], + "logprob_reward": -0.05492936074733734, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05492936074733734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07927361875772476, + 0.1804209053516388, + 0.5011599063873291, + 0.011858178302645683, + 0.013630789704620838, + 0.008030342869460583, + 0.018449336290359497, + 0.16335676610469818, + 0.008399720303714275, + 0.015420442447066307 + ], + "credit_map": [ + 0.010561911389231682, + 0.011504891328513622, + 0.016787296161055565, + 0.007353335153311491, + 0.00790216214954853, + 0.005228071007877588, + 0.011532440781593323, + 0.00840351264923811, + 0.023007621988654137, + 0.025721248239278793, + 0.026870358735322952, + 0.013782037422060966, + 0.030685288831591606, + 0.013373766094446182, + 0.021054571494460106, + 0.025926005095243454, + 0.058377619832754135, + 0.08808748424053192, + 0.07407235354185104, + 0.03539366275072098, + 0.068046435713768, + 0.034228045493364334, + 0.07984090596437454, + 0.06311336159706116, + 0.0014223366742953658, + 0.002378937089815736, + 0.0021623647771775723, + 0.001070016180165112, + 0.0012415952514857054, + 0.0006001126603223383, + 0.001603969489224255, + 0.0013788455398753285, + 0.0014782906509935856, + 0.0019560137297958136, + 0.0008697401499375701, + 0.0008783199009485543, + 0.0018152006668969989, + 0.002716988790780306, + 0.0017310717375949025, + 0.0021851626224815845, + 0.0009823247091844678, + 0.001240435871295631, + 0.0006533042760565877, + 0.0011004480766132474, + 0.0006112969713285565, + 0.0009016874246299267, + 0.0011996704852208495, + 0.001341174473054707, + 0.0024072148371487856, + 0.0019758674316108227, + 0.0013254635268822312, + 0.0016018617898225784, + 0.0011410514125600457, + 0.0008736267918720841, + 0.005411801394075155, + 0.003712449688464403, + 0.03338911011815071, + 0.006447722669690847, + 0.009048542939126492, + 0.008412115275859833, + 0.009336386807262897, + 0.007233266718685627, + 0.06813744455575943, + 0.021352173760533333, + 0.0013679442927241325, + 0.0008851704187691212, + 0.0007541567902080715, + 0.000632302020676434, + 0.000973252288531512, + 0.0013109251158311963, + 0.001632757717743516, + 0.0008432110189460218, + 0.001976156374439597, + 0.0009858927223831415, + 0.0014475364005193114, + 0.002105043036863208, + 0.0008800164214335382, + 0.0011112785432487726, + 0.003817870980128646, + 0.0030966466292738914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.05492936074733734 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011066960170865059 + ], + "opposite_logprob_tokens": [ + -4.511066913604736 + ], + "logprob_reward": -0.011066960170865059, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011066960170865059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2599845826625824, + 0.11910724639892578, + 0.047701071947813034, + 0.029828639701008797, + 0.13145498931407928, + 0.0376596637070179, + 0.03908173367381096, + 0.04387179762125015, + 0.20804773271083832, + 0.08326248824596405 + ], + "credit_map": [ + 0.02681865729391575, + 0.04095244035124779, + 0.022757189348340034, + 0.024970294907689095, + 0.022770864889025688, + 0.027172522619366646, + 0.05272603780031204, + 0.04181660711765289, + 0.015387581661343575, + 0.01865455135703087, + 0.017451541498303413, + 0.009083659388124943, + 0.0196221973747015, + 0.010706858709454536, + 0.013929720968008041, + 0.014271136373281479, + 0.005356220528483391, + 0.008419300429522991, + 0.005977916996926069, + 0.0039344062097370625, + 0.004251210484653711, + 0.004192941356450319, + 0.007653596345335245, + 0.007915484718978405, + 0.003319915384054184, + 0.004900419153273106, + 0.003569097490981221, + 0.0033092047087848186, + 0.0037107085809111595, + 0.0022543405648320913, + 0.0042063272558152676, + 0.004558628425002098, + 0.015681687742471695, + 0.028982944786548615, + 0.010239923372864723, + 0.011281381361186504, + 0.01684744283556938, + 0.02457430772483349, + 0.012108691036701202, + 0.011738630011677742, + 0.0037387297488749027, + 0.0034246882423758507, + 0.0030603401828557253, + 0.004367563407868147, + 0.002654366660863161, + 0.0039043070282787085, + 0.006169264204800129, + 0.01034041028469801, + 0.004759053234010935, + 0.003784637665376067, + 0.0034547464456409216, + 0.0038800647016614676, + 0.0031579500064253807, + 0.0022663685958832502, + 0.010294193401932716, + 0.007484723813831806, + 0.004949857946485281, + 0.0021323878318071365, + 0.003438588697463274, + 0.0038304305635392666, + 0.0026033392641693354, + 0.003343090647831559, + 0.012082543224096298, + 0.0114915631711483, + 0.0302597563713789, + 0.02176757901906967, + 0.02073112316429615, + 0.028746239840984344, + 0.019334701821208, + 0.029760321602225304, + 0.03383365646004677, + 0.02361437678337097, + 0.013718687929213047, + 0.004547196440398693, + 0.006674967706203461, + 0.009583955630660057, + 0.004278907552361488, + 0.003763081505894661, + 0.023011384531855583, + 0.017684312537312508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.011066960170865059 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "magnificent long shot, captured from a slightly upward angle, the camera held still, captures a vast expanse of the sky. Against this crystal-clear, nearly imperfect blue sky, vast, fluffy white cumulus clouds, like cotton candy, float majestically. Illuminated by the bright afternoon sunlight, these vast clouds are edged with a dazzling golden hue, while their interiors are awash in varying shades of gray, accentuating their astonishing volume and three-dimensional depth. With an almost imperceptible and leisurely pace, they drift slowly toward the rolling mountain ranges on the horizon, their contours blurred by the distance, a soft blue-gray hue. Bathed in intense, clear sunlight, the entire scene, with a bright, warm color temperature, creates an atmosphere of profound tranquility, spaciousness, and grandeur, offering a truly refreshing visual experience.", + "seed": 32928, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_12/index_3641_seed_32928_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.018969828356057407, + "vlm_logprob": -0.018969828356057407, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.018969828356057407, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003370677586644888 + ], + "opposite_logprob_tokens": [ + -7.503370761871338 + ], + "logprob_reward": -0.003370677586644888, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.003370677586644888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18041375279426575, + 0.08643478900194168, + 0.08770754188299179, + 0.06575648486614227, + 0.08015923202037811, + 0.12756513059139252, + 0.1453842967748642, + 0.08869590610265732, + 0.06409245729446411, + 0.07379049062728882 + ], + "credit_map": [ + 0.012519105337560177, + 0.024977760389447212, + 0.031316906213760376, + 0.035838816314935684, + 0.016167545691132545, + 0.021086128428578377, + 0.019635818898677826, + 0.018871668726205826, + 0.006993827875703573, + 0.012430224567651749, + 0.023323098197579384, + 0.014034132473170757, + 0.007117072120308876, + 0.0065965065732598305, + 0.007342459633946419, + 0.00859746616333723, + 0.008353727869689465, + 0.013110781088471413, + 0.024918850511312485, + 0.014125843532383442, + 0.005398459732532501, + 0.006882064510136843, + 0.006596831139177084, + 0.008320977911353111, + 0.008077116683125496, + 0.008465913124382496, + 0.008949806913733482, + 0.012063851580023766, + 0.004924830049276352, + 0.006529593374580145, + 0.008382206782698631, + 0.008363164030015469, + 0.010521859861910343, + 0.01191931962966919, + 0.010389008559286594, + 0.008983601815998554, + 0.0062328181229531765, + 0.0071584186516702175, + 0.012719844467937946, + 0.012234361842274666, + 0.00940664391964674, + 0.017068060114979744, + 0.014084921218454838, + 0.01742163486778736, + 0.011631966568529606, + 0.017355462536215782, + 0.022173842415213585, + 0.018422599881887436, + 0.011923892423510551, + 0.019908631220459938, + 0.013251150958240032, + 0.015353945083916187, + 0.017309509217739105, + 0.02196319028735161, + 0.024746106937527657, + 0.020927872508764267, + 0.008770397864282131, + 0.013502123765647411, + 0.010084042325615883, + 0.012876764871180058, + 0.010055391117930412, + 0.011096893809735775, + 0.008622979745268822, + 0.013687311671674252, + 0.0062834047712385654, + 0.00929518137127161, + 0.008553794585168362, + 0.00622294656932354, + 0.006698642857372761, + 0.007477859500795603, + 0.008860989473760128, + 0.010699634440243244, + 0.007155782543122768, + 0.010893252678215504, + 0.01361432857811451, + 0.011935892514884472, + 0.006975180469453335, + 0.007195652462542057, + 0.007203387096524239, + 0.00881701149046421 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.003370677586644888 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201059006154537 + ], + "opposite_logprob_tokens": [ + -4.2642011642456055 + ], + "logprob_reward": -0.014201059006154537, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201059006154537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2671240270137787, + 0.14028871059417725, + 0.14864377677440643, + 0.03818544000387192, + 0.055773016065359116, + 0.10506144911050797, + 0.06994426995515823, + 0.044709667563438416, + 0.06551724672317505, + 0.06475239247083664 + ], + "credit_map": [ + 0.02337157167494297, + 0.0416666604578495, + 0.07489681243896484, + 0.029251890257000923, + 0.020154394209384918, + 0.022582443431019783, + 0.021389678120613098, + 0.03381059318780899, + 0.013708999380469322, + 0.02074473723769188, + 0.0390629768371582, + 0.024707598611712456, + 0.010057853534817696, + 0.007741084322333336, + 0.011745256371796131, + 0.012520203366875648, + 0.014515005983412266, + 0.023549873381853104, + 0.046406298875808716, + 0.02619728446006775, + 0.009801102802157402, + 0.00708482600748539, + 0.010570639744400978, + 0.010518740862607956, + 0.0030457156244665384, + 0.005057732108980417, + 0.005372023209929466, + 0.005584056954830885, + 0.0029225668404251337, + 0.004392435774207115, + 0.004356826655566692, + 0.007454081904143095, + 0.007062519900500774, + 0.007789865601807833, + 0.0075415619648993015, + 0.009231753647327423, + 0.005611760076135397, + 0.005237910430878401, + 0.006589747499674559, + 0.006707894615828991, + 0.010547970421612263, + 0.016905387863516808, + 0.032014552503824234, + 0.013735697604715824, + 0.006073811557143927, + 0.0064881048165261745, + 0.009477042593061924, + 0.009818878024816513, + 0.006662698928266764, + 0.009986765682697296, + 0.007157259155064821, + 0.006753253750503063, + 0.007811640854924917, + 0.009348444640636444, + 0.010787839069962502, + 0.011436368338763714, + 0.004468159284442663, + 0.006331627257168293, + 0.008745224215090275, + 0.008104383945465088, + 0.0030160287860780954, + 0.004328801296651363, + 0.004262713715434074, + 0.005452731158584356, + 0.006010331679135561, + 0.009778915904462337, + 0.007605625782161951, + 0.00913920532912016, + 0.007400706876069307, + 0.006941340863704681, + 0.008368420414626598, + 0.010272703133523464, + 0.004990243818610907, + 0.008862385526299477, + 0.014542621560394764, + 0.011144770309329033, + 0.005039767362177372, + 0.006123717408627272, + 0.006972429808229208, + 0.00707645807415247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.014201059006154537 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076298960484564304 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.0076298960484564304, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076298960484564304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24998869001865387, + 0.11314934492111206, + 0.11891763657331467, + 0.04859773442149162, + 0.048418622463941574, + 0.15940800309181213, + 0.10777091979980469, + 0.044873692095279694, + 0.03885974735021591, + 0.07001558691263199 + ], + "credit_map": [ + 0.023394400253891945, + 0.036421969532966614, + 0.05816164240241051, + 0.035702046006917953, + 0.021350271999835968, + 0.026311391964554787, + 0.021090848371386528, + 0.027556130662560463, + 0.009557709097862244, + 0.014233192428946495, + 0.023028574883937836, + 0.016525156795978546, + 0.008570769801735878, + 0.009913114830851555, + 0.012608855962753296, + 0.01871197298169136, + 0.010844639502465725, + 0.017628876492381096, + 0.03439679741859436, + 0.0209481380879879, + 0.007813296280801296, + 0.00574135547503829, + 0.009597809053957462, + 0.011946728453040123, + 0.004096930380910635, + 0.00603902991861105, + 0.005706234369426966, + 0.007610594853758812, + 0.003690861165523529, + 0.0057980939745903015, + 0.007051642518490553, + 0.008604345843195915, + 0.0043473802506923676, + 0.006749526597559452, + 0.006204952020198107, + 0.007123724091798067, + 0.0037634465843439102, + 0.004312959499657154, + 0.00624477956444025, + 0.009671853855252266, + 0.011931030079722404, + 0.022270234301686287, + 0.02203936129808426, + 0.020580481737852097, + 0.014808496460318565, + 0.018616503104567528, + 0.027259645983576775, + 0.021902253851294518, + 0.007907750084996223, + 0.01505577564239502, + 0.008206901140511036, + 0.012311549857258797, + 0.012993691489100456, + 0.01695014163851738, + 0.017848018556833267, + 0.016497086733579636, + 0.0052703870460391045, + 0.006476616952568293, + 0.004097513854503632, + 0.0052477288991212845, + 0.004141767043620348, + 0.005226259119808674, + 0.006504982244223356, + 0.007908441126346588, + 0.003843015292659402, + 0.006163675803691149, + 0.005915616173297167, + 0.004987104795873165, + 0.003151803743094206, + 0.004357490688562393, + 0.004658352583646774, + 0.005782688967883587, + 0.005779095459729433, + 0.010532689280807972, + 0.016759568825364113, + 0.010745278559625149, + 0.005175232887268066, + 0.007706422358751297, + 0.006377757992595434, + 0.00693954573944211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.0076298960484564304 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062025539577007294 + ], + "opposite_logprob_tokens": [ + -2.812025547027588 + ], + "logprob_reward": -0.062025539577007294, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062025539577007294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24164052307605743, + 0.0792621299624443, + 0.21333898603916168, + 0.06261056661605835, + 0.03431786224246025, + 0.12069311738014221, + 0.060014113783836365, + 0.03993489220738411, + 0.02999315969645977, + 0.11819462478160858 + ], + "credit_map": [ + 0.0239915382117033, + 0.03648848086595535, + 0.05613354220986366, + 0.03626730293035507, + 0.019134294241666794, + 0.022262629121541977, + 0.019954057410359383, + 0.027408668771386147, + 0.012067820876836777, + 0.01218015979975462, + 0.009328870102763176, + 0.012802256271243095, + 0.007087782025337219, + 0.0073888273909688, + 0.007605511695146561, + 0.010800899937748909, + 0.020477689802646637, + 0.034888800233602524, + 0.0674285888671875, + 0.036578889936208725, + 0.013255192898213863, + 0.009812869131565094, + 0.015397055074572563, + 0.015499920584261417, + 0.005786956753581762, + 0.009553617797791958, + 0.017337031662464142, + 0.010815142653882504, + 0.003900624578818679, + 0.004388277418911457, + 0.00550060486420989, + 0.005328304599970579, + 0.0037547301035374403, + 0.005765353795140982, + 0.00455657159909606, + 0.0029768331442028284, + 0.003083686111494899, + 0.0035652031656354666, + 0.0047985101118683815, + 0.005816970951855183, + 0.012802979908883572, + 0.02015443705022335, + 0.03846936672925949, + 0.015006470493972301, + 0.007098158821463585, + 0.007441607769578695, + 0.008888334035873413, + 0.010831774212419987, + 0.004759544972330332, + 0.008313008584082127, + 0.005463829729706049, + 0.008071430958807468, + 0.006915339734405279, + 0.008419396355748177, + 0.008574935607612133, + 0.009496623650193214, + 0.0031993358861654997, + 0.005927879363298416, + 0.0063659390434622765, + 0.0049901083111763, + 0.0037975276354700327, + 0.004891881253570318, + 0.004526831675320864, + 0.00623538950458169, + 0.003172897733747959, + 0.004758802708238363, + 0.007313839625567198, + 0.004024730063974857, + 0.0024012508802115917, + 0.0028544641099870205, + 0.002778426744043827, + 0.002688748063519597, + 0.010192262008786201, + 0.0179780051112175, + 0.03353016823530197, + 0.019727880135178566, + 0.007041688077151775, + 0.008965338580310345, + 0.010799159295856953, + 0.009960124269127846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.062025539577007294 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot capturing a vast expanse of the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward and held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast fluffy white cumulus clouds visible with golden edges and gray interiors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling mountain ranges visible on the horizon with a soft blue-gray hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright afternoon sunlight with a warm color temperature?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007621969562023878 + ], + "opposite_logprob_tokens": [ + -4.882621765136719 + ], + "logprob_reward": -0.007621969562023878, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007621969562023878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34527862071990967, + 0.17003268003463745, + 0.06517989933490753, + 0.07685840874910355, + 0.09921731054782867, + 0.07713396102190018, + 0.051681265234947205, + 0.043061163276433945, + 0.03209961578249931, + 0.039457082748413086 + ], + "credit_map": [ + 0.020317621529102325, + 0.040513746440410614, + 0.033962033689022064, + 0.04312419146299362, + 0.03315156325697899, + 0.03738224133849144, + 0.05894703418016434, + 0.07788018882274628, + 0.011785811744630337, + 0.020331984385848045, + 0.0336889922618866, + 0.030022749677300453, + 0.012862049974501133, + 0.013853801414370537, + 0.02109435200691223, + 0.026392938569188118, + 0.005933522712439299, + 0.008858790621161461, + 0.009670155122876167, + 0.010406570509076118, + 0.0050546154379844666, + 0.005854091607034206, + 0.007593480404466391, + 0.011808672919869423, + 0.006774770561605692, + 0.009794346988201141, + 0.010109176859259605, + 0.01003379374742508, + 0.00588984927162528, + 0.007881826721131802, + 0.0107479402795434, + 0.015626704320311546, + 0.01088020671159029, + 0.014292866922914982, + 0.012182804755866528, + 0.013161561451852322, + 0.010017679072916508, + 0.009625882841646671, + 0.01340475119650364, + 0.01565154641866684, + 0.006516374181956053, + 0.01129133440554142, + 0.009892086498439312, + 0.008822523057460785, + 0.008423295803368092, + 0.009713279083371162, + 0.010981295257806778, + 0.011493770405650139, + 0.004700272344052792, + 0.0077439253218472, + 0.0048645054921507835, + 0.005431443452835083, + 0.005304569844156504, + 0.007033542264252901, + 0.007245309185236692, + 0.009357694536447525, + 0.004017633385956287, + 0.006218802183866501, + 0.005506779067218304, + 0.004988380707800388, + 0.003552898997440934, + 0.0047199176624417305, + 0.005458438768982887, + 0.00859831739217043, + 0.004207952413707972, + 0.004343955311924219, + 0.005108820740133524, + 0.004216100554913282, + 0.0024584666825830936, + 0.0024490095674991608, + 0.004607517272233963, + 0.004707791842520237, + 0.004626393783837557, + 0.006043116562068462, + 0.008543905802071095, + 0.007668649312108755, + 0.0024095287080854177, + 0.002327547874301672, + 0.0035663126036524773, + 0.004271632991731167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.007621969562023878 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_13/index_tensor([3889, 3889])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_13/index_tensor([3889, 3889])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dcb4fac81bade17c73a8832f17dd0708e1740509 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_13/index_tensor([3889, 3889])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a surreal and clean style, a high-angle medium shot establishes the scene of a playful conflict between two cats on a spotless, reflective white floor, illuminated by bright, shadowless studio lighting. A fluffy, long-haired Persian cat with pristine white fur lies on its side, its sturdy body protectively huddled over a small pile of brightly colored, crumpled pet toys. Its front paws firmly grasp the plastic rod of a feather wand, its vibrant blue and purple feathers perched near its face. Opposite it, a sleek Siamese cat, distinguished by its cream-colored body, black seal-point markings, and piercing blue, almond-shaped eyes, crouches in a low, predatory posture. As the Siamese cat swiftly and decisively brings its paw to bear on the feather wand, the camera executes a quick, seamless push-in, transitioning to an extreme close-up, focusing on the precise moment their paws meet. In this series of mock battles, the camera captures with microscopic clarity the short hair of the Siamese pressed against the dense, soft fur of the Persian, and the gentle, translucent claws of the Siamese, only partially unsheathed, gleaming in the even light, perfectly conveying the controlled, playful nature of their interaction against the pristine, minimalist backdrop.", + "seed": 24951, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_13/index_3889_seed_24951_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.041218836419284345, + "vlm_logprob": -0.041218836419284345, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.041218836419284345, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13820892572402954 + ], + "opposite_logprob_tokens": [ + -2.2632088661193848 + ], + "logprob_reward": -0.13820892572402954, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.13820892572402954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09153644740581512, + 0.026549559086561203, + 0.037976402789354324, + 0.01657981239259243, + 0.56444251537323, + 0.0161476731300354, + 0.044217173010110855, + 0.01644423045217991, + 0.030574491247534752, + 0.15553176403045654 + ], + "credit_map": [ + 0.012763109989464283, + 0.00962734129279852, + 0.01652250997722149, + 0.006127465516328812, + 0.010288394056260586, + 0.014004618860781193, + 0.005695015657693148, + 0.016507986932992935, + 0.003632493084296584, + 0.003465913003310561, + 0.003271022578701377, + 0.0019205581629648805, + 0.003616426372900605, + 0.004854533821344376, + 0.0017536049708724022, + 0.004035003483295441, + 0.00536766555160284, + 0.00674432422965765, + 0.005915271583944559, + 0.004088856279850006, + 0.003520128782838583, + 0.005581243429332972, + 0.0043156445026397705, + 0.002443263540044427, + 0.003102286485955119, + 0.001812409027479589, + 0.004425455816090107, + 0.0010816951980814338, + 0.001013693050481379, + 0.001842453726567328, + 0.0014061892870813608, + 0.001895626657642424, + 0.057159606367349625, + 0.04971272870898247, + 0.05826020613312721, + 0.06179749593138695, + 0.0920155867934227, + 0.10100477933883667, + 0.04651101306080818, + 0.0979810282588005, + 0.0017904730048030615, + 0.0014736359007656574, + 0.002260022796690464, + 0.002425443148240447, + 0.0015977997099980712, + 0.002586017595604062, + 0.0019285609014332294, + 0.0020857176277786493, + 0.003299023723229766, + 0.0035472186282277107, + 0.004181557800620794, + 0.007089268881827593, + 0.006194150075316429, + 0.0056742033921182156, + 0.00730343209579587, + 0.006928313057869673, + 0.0016119693173095584, + 0.0017862538807094097, + 0.0017480980604887009, + 0.001774118049070239, + 0.0016768166096881032, + 0.002407654421404004, + 0.0020873802714049816, + 0.0033519379794597626, + 0.0038287374190986156, + 0.0028010800015181303, + 0.004162568133324385, + 0.004968014545738697, + 0.003178935730829835, + 0.0023612026125192642, + 0.004607414826750755, + 0.004666533321142197, + 0.01811864785850048, + 0.020567335188388824, + 0.02465919405221939, + 0.023266397416591644, + 0.0159444697201252, + 0.01746465265750885, + 0.013198613189160824, + 0.02231244370341301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.13820892572402954 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012560192495584488 + ], + "opposite_logprob_tokens": [ + -4.3875603675842285 + ], + "logprob_reward": -0.012560192495584488, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012560192495584488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13316410779953003, + 0.09234565496444702, + 0.06473768502473831, + 0.02062086947262287, + 0.4023468494415283, + 0.0964823067188263, + 0.023658080026507378, + 0.030964169651269913, + 0.047538746148347855, + 0.08814159780740738 + ], + "credit_map": [ + 0.018263382837176323, + 0.016740769147872925, + 0.01771557703614235, + 0.00791847612708807, + 0.021383235231041908, + 0.020762162283062935, + 0.007891370914876461, + 0.022489115595817566, + 0.01320615317672491, + 0.01355593279004097, + 0.01167970523238182, + 0.006470818072557449, + 0.012336517684161663, + 0.01487369742244482, + 0.005657376255840063, + 0.014565452001988888, + 0.00884256698191166, + 0.01228802278637886, + 0.010214777663350105, + 0.007553542032837868, + 0.005878082476556301, + 0.007180291227996349, + 0.00615645619109273, + 0.0066239433363080025, + 0.0035905123222619295, + 0.0026107048615813255, + 0.004778377711772919, + 0.0017103084828704596, + 0.0020488728769123554, + 0.0024340320378541946, + 0.001858627307228744, + 0.0015894342213869095, + 0.0420188270509243, + 0.03968945890665054, + 0.03850958123803139, + 0.04312184453010559, + 0.06973347067832947, + 0.07090681791305542, + 0.03336133062839508, + 0.06500551849603653, + 0.006058303639292717, + 0.011325940489768982, + 0.00992956105619669, + 0.012745033949613571, + 0.009830026887357235, + 0.016425685957074165, + 0.01690327562391758, + 0.013264471665024757, + 0.0020649146754294634, + 0.0019824127666652203, + 0.0019798020366579294, + 0.0030319334473460913, + 0.0036469583865255117, + 0.003917154390364885, + 0.003351212479174137, + 0.003683691378682852, + 0.00490247318521142, + 0.0034235490020364523, + 0.0033002025447785854, + 0.002649075584486127, + 0.0025394619442522526, + 0.0037418866995722055, + 0.0022777793928980827, + 0.00812973827123642, + 0.004746544174849987, + 0.005412088241428137, + 0.005741364788264036, + 0.006371492054313421, + 0.007398610003292561, + 0.004198368638753891, + 0.007745605893433094, + 0.005924666300415993, + 0.01006705965846777, + 0.011334025301039219, + 0.012860278598964214, + 0.012627842836081982, + 0.009261083789169788, + 0.010355113074183464, + 0.007608385756611824, + 0.014027799479663372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.012560192495584488 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779046289622784 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.009779046289622784, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779046289622784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0502081960439682, + 0.021702975034713745, + 0.028433557599782944, + 0.018759069964289665, + 0.16210369765758514, + 0.06297589838504791, + 0.02064673788845539, + 0.04159873351454735, + 0.03776908293366432, + 0.5558019876480103 + ], + "credit_map": [ + 0.007102384697645903, + 0.005881856195628643, + 0.0084025077521801, + 0.002789872232824564, + 0.006079323589801788, + 0.007535817567259073, + 0.002856168895959854, + 0.009560266509652138, + 0.003272547386586666, + 0.0028162978123873472, + 0.0028251560870558023, + 0.0019237074302509427, + 0.0029385597445070744, + 0.004048495087772608, + 0.002093809423968196, + 0.001784400432370603, + 0.004419509321451187, + 0.0045915888622403145, + 0.004575265571475029, + 0.002690151333808899, + 0.003806315828114748, + 0.003295458620414138, + 0.002883168635889888, + 0.0021720987278968096, + 0.0037757016252726316, + 0.0015514850383624434, + 0.004793764557689428, + 0.0018671030411496758, + 0.0017573602963238955, + 0.001909663900732994, + 0.0020415736362338066, + 0.0010624171700328588, + 0.015885168686509132, + 0.01449450571089983, + 0.017500145360827446, + 0.017199095338582993, + 0.025544630363583565, + 0.03140445426106453, + 0.012584245763719082, + 0.02749144844710827, + 0.00619686720892787, + 0.01112148817628622, + 0.008026930503547192, + 0.010371310636401176, + 0.006934795528650284, + 0.007910105399787426, + 0.006528813391923904, + 0.005885583348572254, + 0.001501617836765945, + 0.0018258081981912255, + 0.00146964518353343, + 0.0024748467840254307, + 0.0027585490606725216, + 0.0033529361244291067, + 0.00345194642432034, + 0.003811388276517391, + 0.0052319858223199844, + 0.006295827217400074, + 0.00451374938711524, + 0.005651756189763546, + 0.004935344681143761, + 0.004949319642037153, + 0.004022294655442238, + 0.005998456384986639, + 0.003472916316241026, + 0.003730472642928362, + 0.004913344979286194, + 0.004898645915091038, + 0.0049564167857170105, + 0.0031991598661988974, + 0.0056955572217702866, + 0.00690257316455245, + 0.06587520986795425, + 0.07024712860584259, + 0.08881818503141403, + 0.08356434106826782, + 0.05746489763259888, + 0.06075500696897507, + 0.04779065027832985, + 0.08128660172224045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.009779046289622784 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305883124470711 + ], + "opposite_logprob_tokens": [ + -3.1680588722229004 + ], + "logprob_reward": -0.04305883124470711, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305883124470711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10569380968809128, + 0.0720592811703682, + 0.12687264382839203, + 0.040431637316942215, + 0.1774733066558838, + 0.16334587335586548, + 0.06135639548301697, + 0.060720138251781464, + 0.10690566152334213, + 0.08514126390218735 + ], + "credit_map": [ + 0.01519811525940895, + 0.017289532348513603, + 0.014807465486228466, + 0.006475555710494518, + 0.018096838146448135, + 0.019339030608534813, + 0.006270257290452719, + 0.008217019960284233, + 0.010853683575987816, + 0.008667252957820892, + 0.010507514700293541, + 0.005801043938845396, + 0.008208325132727623, + 0.012107806280255318, + 0.005191169213503599, + 0.010722487233579159, + 0.01996973343193531, + 0.021691452711820602, + 0.013098043389618397, + 0.00850894208997488, + 0.020247599110007286, + 0.021603615954518318, + 0.009650111198425293, + 0.01210314966738224, + 0.007205732632428408, + 0.004024256486445665, + 0.011105484329164028, + 0.002828188706189394, + 0.0035797792952507734, + 0.004483365453779697, + 0.003644587704911828, + 0.0035602429416030645, + 0.014856095425784588, + 0.021409746259450912, + 0.0192527137696743, + 0.014993758872151375, + 0.023549366742372513, + 0.03714606910943985, + 0.02192772552371025, + 0.02433784119784832, + 0.010403363034129143, + 0.017366351559758186, + 0.011931893415749073, + 0.022119661793112755, + 0.016094835475087166, + 0.03195280581712723, + 0.02985803782939911, + 0.023618927225470543, + 0.004430320113897324, + 0.004760043695569038, + 0.005349234212189913, + 0.008330193348228931, + 0.011742070317268372, + 0.012931044213473797, + 0.006568302400410175, + 0.007245182991027832, + 0.008333303965628147, + 0.006262977607548237, + 0.007770119700580835, + 0.006711605004966259, + 0.007984465919435024, + 0.006798911839723587, + 0.006318201776593924, + 0.01054055243730545, + 0.009038565680384636, + 0.010404299944639206, + 0.014375750906765461, + 0.013862360268831253, + 0.013076936826109886, + 0.010862404480576515, + 0.01433723047375679, + 0.02094811387360096, + 0.011085444130003452, + 0.010995444841682911, + 0.014010227285325527, + 0.013252703472971916, + 0.008992820978164673, + 0.007934970781207085, + 0.007543119601905346, + 0.011326528154313564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.04305883124470711 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024871863424777985 + ], + "opposite_logprob_tokens": [ + -6.0024871826171875 + ], + "logprob_reward": -0.0024871863424777985, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024871863424777985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0491812601685524, + 0.03166342154145241, + 0.0976237803697586, + 0.02464892901480198, + 0.28505605459213257, + 0.09214980900287628, + 0.06630085408687592, + 0.12194512039422989, + 0.15790896117687225, + 0.07352186739444733 + ], + "credit_map": [ + 0.007082473021000624, + 0.0043795062229037285, + 0.008919931948184967, + 0.00397958280518651, + 0.006094653625041246, + 0.0061754086054861546, + 0.0066378614865243435, + 0.005911841057240963, + 0.004884803202003241, + 0.0036932320799678564, + 0.00457551795989275, + 0.0027016201056540012, + 0.004461793694645166, + 0.006094510201364756, + 0.0027798181399703026, + 0.0024721254594624043, + 0.013098037801682949, + 0.015577204525470734, + 0.014664757996797562, + 0.007838658057153225, + 0.010307257063686848, + 0.013647732324898243, + 0.011029165238142014, + 0.01146096084266901, + 0.003570506116375327, + 0.002527969190850854, + 0.004379060585051775, + 0.0015138292219489813, + 0.0025193565525114536, + 0.002938195364549756, + 0.003450196934863925, + 0.0037498127203434706, + 0.032015111297369, + 0.025587759912014008, + 0.029107315465807915, + 0.029982464388012886, + 0.04634149372577667, + 0.04938569664955139, + 0.02389586716890335, + 0.048740338534116745, + 0.007082717958837748, + 0.007120832335203886, + 0.009161248803138733, + 0.00994093343615532, + 0.009997962042689323, + 0.01677607372403145, + 0.014485505409538746, + 0.017584532499313354, + 0.004948291461914778, + 0.005744960159063339, + 0.006005287170410156, + 0.009914238937199116, + 0.007872113026678562, + 0.011569799855351448, + 0.011188187636435032, + 0.009057970717549324, + 0.013633272610604763, + 0.015216510742902756, + 0.012132957577705383, + 0.013705481775105, + 0.013297497294843197, + 0.01634460687637329, + 0.013359841890633106, + 0.0242549367249012, + 0.01488945260643959, + 0.01576041243970394, + 0.026295198127627373, + 0.01816936582326889, + 0.018054183572530746, + 0.013636776246130466, + 0.019663861021399498, + 0.031439706683158875, + 0.01102093979716301, + 0.008423159830272198, + 0.016397299244999886, + 0.008830019272863865, + 0.00776794645935297, + 0.008012671954929829, + 0.0063180625438690186, + 0.006751760374754667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.0024871863424777985 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a surreal and clean style, a high-angle medium shot establishes the scene of a playful conflict between two cats on a spotless, reflective white floor, illuminated by bright, shadowless studio lighting. A fluffy, long-haired Persian cat with pristine white fur lies on its side, its sturdy body protectively huddled over a small pile of brightly colored, crumpled pet toys. Its front paws firmly grasp the plastic rod of a feather wand, its vibrant blue and purple feathers perched near its face. Opposite it, a sleek Siamese cat, distinguished by its cream-colored body, black seal-point markings, and piercing blue, almond-shaped eyes, crouches in a low, predatory posture. As the Siamese cat swiftly and decisively brings its paw to bear on the feather wand, the camera executes a quick, seamless push-in, transitioning to an extreme close-up, focusing on the precise moment their paws meet. In this series of mock battles, the camera captures with microscopic clarity the short hair of the Siamese pressed against the dense, soft fur of the Persian, and the gentle, translucent claws of the Siamese, only partially unsheathed, gleaming in the even light, perfectly conveying the controlled, playful nature of their interaction against the pristine, minimalist backdrop.", + "seed": 24952, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_13/index_3889_seed_24952_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02551391350571066, + "vlm_logprob": -0.02551391350571066, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.02551391350571066, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06435471028089523 + ], + "opposite_logprob_tokens": [ + -3.314354658126831 + ], + "logprob_reward": -0.06435471028089523, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06435471028089523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08516691625118256, + 0.12267754226922989, + 0.07425644993782043, + 0.029445458203554153, + 0.2791452705860138, + 0.18656902015209198, + 0.052884526550769806, + 0.08874562382698059, + 0.048519860953092575, + 0.03258933871984482 + ], + "credit_map": [ + 0.01109218318015337, + 0.017068516463041306, + 0.013399941846728325, + 0.01037968322634697, + 0.009632593020796776, + 0.009593225084245205, + 0.007175554521381855, + 0.0068252207711339, + 0.014244449324905872, + 0.017371507361531258, + 0.01713898777961731, + 0.010478830896317959, + 0.01672603003680706, + 0.017982469871640205, + 0.010648181661963463, + 0.018087083473801613, + 0.008216476999223232, + 0.007878366857767105, + 0.01104819681495428, + 0.00768488273024559, + 0.010047674179077148, + 0.014695312827825546, + 0.007459077052772045, + 0.007226467598229647, + 0.004037552047520876, + 0.004056130535900593, + 0.003966925200074911, + 0.0031554994639009237, + 0.003102554241195321, + 0.005401587579399347, + 0.0031242184340953827, + 0.0026009913999587297, + 0.052445992827415466, + 0.02146081067621708, + 0.04817154258489609, + 0.04004724323749542, + 0.017180535942316055, + 0.023464879021048546, + 0.03660871461033821, + 0.03976553678512573, + 0.030486946925520897, + 0.021463964134454727, + 0.028258534148335457, + 0.03305233642458916, + 0.020442768931388855, + 0.016069898381829262, + 0.019624365493655205, + 0.017170213162899017, + 0.009086753241717815, + 0.004100109450519085, + 0.004114985931664705, + 0.003307033097371459, + 0.00383306248113513, + 0.006855962797999382, + 0.005644653923809528, + 0.015941966325044632, + 0.008588702417910099, + 0.006716079544275999, + 0.006367894355207682, + 0.008848627097904682, + 0.009369799867272377, + 0.022145142778754234, + 0.021541137248277664, + 0.005168232601135969, + 0.005465086083859205, + 0.006425018887966871, + 0.006535481661558151, + 0.004505242686718702, + 0.00644939998164773, + 0.004858352709561586, + 0.009383711032569408, + 0.004897568374872208, + 0.0032927263528108597, + 0.004525917116552591, + 0.004531612619757652, + 0.003786214627325535, + 0.0034489852841943502, + 0.0051183998584747314, + 0.0036074614617973566, + 0.004278020467609167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.06435471028089523 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008656594902276993 + ], + "opposite_logprob_tokens": [ + -4.7586565017700195 + ], + "logprob_reward": -0.008656594902276993, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008656594902276993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08086779713630676, + 0.06930802017450333, + 0.1256774514913559, + 0.09960246086120605, + 0.20960381627082825, + 0.10833180695772171, + 0.04200209304690361, + 0.17103925347328186, + 0.03871368616819382, + 0.05485361069440842 + ], + "credit_map": [ + 0.011702124960720539, + 0.012489041313529015, + 0.011225885711610317, + 0.00917405728250742, + 0.007792264688760042, + 0.012528127990663052, + 0.00818118266761303, + 0.0077751134522259235, + 0.006195533089339733, + 0.010190638713538647, + 0.011128822341561317, + 0.005144303664565086, + 0.008551975712180138, + 0.01260698214173317, + 0.006390820257365704, + 0.009098944254219532, + 0.015516288578510284, + 0.014050609432160854, + 0.019389532506465912, + 0.013996909372508526, + 0.015091534703969955, + 0.02233823947608471, + 0.011287050321698189, + 0.014007284305989742, + 0.016611339524388313, + 0.012750105932354927, + 0.01376037485897541, + 0.007165523245930672, + 0.014586452394723892, + 0.01800851710140705, + 0.008322935551404953, + 0.008397209458053112, + 0.04021299630403519, + 0.019529977813363075, + 0.03441328555345535, + 0.03175318241119385, + 0.015538344159722328, + 0.017571784555912018, + 0.02502402663230896, + 0.02556021325290203, + 0.016373317688703537, + 0.011194058693945408, + 0.017009591683745384, + 0.015903687104582787, + 0.013066044077277184, + 0.010624726302921772, + 0.01363626029342413, + 0.010524129495024681, + 0.005968871526420116, + 0.004472122993320227, + 0.004384323954582214, + 0.003582942532375455, + 0.004291957244277, + 0.00687383720651269, + 0.007082117255777121, + 0.005345919169485569, + 0.016021689400076866, + 0.013992092572152615, + 0.012986494228243828, + 0.013278977014124393, + 0.020613398402929306, + 0.03999105468392372, + 0.04365270212292671, + 0.0105028310790658, + 0.003662689356133342, + 0.005823169834911823, + 0.002913637785241008, + 0.005772356875240803, + 0.004488077946007252, + 0.005682863295078278, + 0.007033193483948708, + 0.0033376996871083975, + 0.0037063267081975937, + 0.005970615427941084, + 0.0069558946415781975, + 0.00559596810489893, + 0.006032136734575033, + 0.009456554427742958, + 0.01035807840526104, + 0.0067780353128910065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 2 + ], + "selected_reward": -0.008656594902276993 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779518470168114 + ], + "opposite_logprob_tokens": [ + -4.634779453277588 + ], + "logprob_reward": -0.009779518470168114, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779518470168114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09533220529556274, + 0.06711255013942719, + 0.09225743263959885, + 0.08287004381418228, + 0.20666921138763428, + 0.14372365176677704, + 0.07819689065217972, + 0.16551004350185394, + 0.03365420177578926, + 0.034673724323511124 + ], + "credit_map": [ + 0.01383573655039072, + 0.018526481464505196, + 0.014732848852872849, + 0.014323286712169647, + 0.007781983818858862, + 0.011873051524162292, + 0.007717814762145281, + 0.006541009061038494, + 0.006110644433647394, + 0.009127619676291943, + 0.011640990152955055, + 0.003754902631044388, + 0.0089130699634552, + 0.016087716445326805, + 0.004409600980579853, + 0.0070680128410458565, + 0.011166210286319256, + 0.009686652570962906, + 0.014754155650734901, + 0.005413489416241646, + 0.018293363973498344, + 0.0192303117364645, + 0.005910212639719248, + 0.007803047075867653, + 0.011986132711172104, + 0.01068289764225483, + 0.012388138100504875, + 0.006396583281457424, + 0.008051570504903793, + 0.018142390996217728, + 0.006260012276470661, + 0.008962328545749187, + 0.045738376677036285, + 0.014708440750837326, + 0.039503101259469986, + 0.030042869970202446, + 0.0153739582747221, + 0.015881257131695747, + 0.022492172196507454, + 0.022929038852453232, + 0.018360473215579987, + 0.018329018726944923, + 0.015703728422522545, + 0.009380048140883446, + 0.0314057283103466, + 0.01806434988975525, + 0.017652681097388268, + 0.014827631413936615, + 0.01006525382399559, + 0.006264483090490103, + 0.0077278222888708115, + 0.005463516805320978, + 0.0080086225643754, + 0.010877017863094807, + 0.012318555265665054, + 0.01747162640094757, + 0.020017288625240326, + 0.0175060722976923, + 0.009170168079435825, + 0.010464286431670189, + 0.014539618976414204, + 0.024826407432556152, + 0.0420193187892437, + 0.026966916397213936, + 0.004375388380140066, + 0.0040181903168559074, + 0.0034422825556248426, + 0.005369654856622219, + 0.004239413421601057, + 0.004489175509661436, + 0.004052714444696903, + 0.003667388577014208, + 0.0030695279128849506, + 0.005109561141580343, + 0.003237321972846985, + 0.004034274257719517, + 0.003208114067092538, + 0.0048793829046189785, + 0.006192520260810852, + 0.004943022970110178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.009779518470168114 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306705296039581 + ], + "opposite_logprob_tokens": [ + -3.16806697845459 + ], + "logprob_reward": -0.04306705296039581, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306705296039581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09253597259521484, + 0.05954205244779587, + 0.11716616898775101, + 0.12524934113025665, + 0.15699869394302368, + 0.10340966284275055, + 0.1369342803955078, + 0.07788099348545074, + 0.09686826169490814, + 0.03341460973024368 + ], + "credit_map": [ + 0.009461911395192146, + 0.013375263661146164, + 0.013016687706112862, + 0.011266978457570076, + 0.009141805581748486, + 0.01362905465066433, + 0.011122025549411774, + 0.011522246524691582, + 0.006521936040371656, + 0.008597106672823429, + 0.00813295692205429, + 0.006891938857734203, + 0.00981129426509142, + 0.009217223152518272, + 0.004900864791125059, + 0.005468735471367836, + 0.01475894171744585, + 0.009728074073791504, + 0.017992962151765823, + 0.013185907155275345, + 0.02241947129368782, + 0.020082728937268257, + 0.008637226186692715, + 0.010360857471823692, + 0.0175617728382349, + 0.01427885890007019, + 0.01812758669257164, + 0.008608880452811718, + 0.01947665773332119, + 0.02496338076889515, + 0.009589161723852158, + 0.012643042951822281, + 0.02623412199318409, + 0.013118969276547432, + 0.01919843629002571, + 0.022102052345871925, + 0.012252806685864925, + 0.011351505294442177, + 0.020571786910295486, + 0.03216900676488876, + 0.011512059718370438, + 0.011654735542833805, + 0.009423389099538326, + 0.011942952871322632, + 0.020438555628061295, + 0.012022055685520172, + 0.015640126541256905, + 0.010775796137750149, + 0.023330288007855415, + 0.0121928621083498, + 0.01327076368033886, + 0.008863842114806175, + 0.013163482770323753, + 0.04070528969168663, + 0.01589788869023323, + 0.009509864263236523, + 0.006006000097841024, + 0.005848695058375597, + 0.005086993332952261, + 0.004757103975862265, + 0.00811315979808569, + 0.02148505486547947, + 0.015221203677356243, + 0.011362781748175621, + 0.009691796265542507, + 0.013056818395853043, + 0.01773693785071373, + 0.010640745982527733, + 0.01315541472285986, + 0.0090219397097826, + 0.013127272948622704, + 0.010437341406941414, + 0.0029156659729778767, + 0.004800674971193075, + 0.002923590363934636, + 0.003528842469677329, + 0.00422689551487565, + 0.003944915719330311, + 0.004408916924148798, + 0.006665105931460857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.04306705296039581 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017116909148171544 + ], + "opposite_logprob_tokens": [ + -6.376711845397949 + ], + "logprob_reward": -0.0017116909148171544, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017116909148171544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06797667592763901, + 0.054761238396167755, + 0.02729257009923458, + 0.02627657912671566, + 0.18653345108032227, + 0.09917927533388138, + 0.25177523493766785, + 0.11065855622291565, + 0.1508937031030655, + 0.0246527511626482 + ], + "credit_map": [ + 0.0065863183699548244, + 0.007491201162338257, + 0.010924809612333775, + 0.009540505707263947, + 0.007338617462664843, + 0.011853235773742199, + 0.007223626598715782, + 0.007018357515335083, + 0.00669797882437706, + 0.007340975105762482, + 0.007694522850215435, + 0.004017418250441551, + 0.010832494124770164, + 0.008090897463262081, + 0.005159728694707155, + 0.004927219823002815, + 0.0033118941355496645, + 0.0033359676599502563, + 0.004018574953079224, + 0.0026256428100168705, + 0.005675286054611206, + 0.004338299389928579, + 0.0016624056734144688, + 0.002324497327208519, + 0.0031790409702807665, + 0.003150112461298704, + 0.003855384886264801, + 0.0030659623444080353, + 0.003328843740746379, + 0.00456880871206522, + 0.002119004260748625, + 0.0030094198882579803, + 0.0343829020857811, + 0.014697587117552757, + 0.029821405187249184, + 0.026776354759931564, + 0.01344085019081831, + 0.01872318610548973, + 0.023739103227853775, + 0.02495204471051693, + 0.015985291451215744, + 0.011281081475317478, + 0.012857591733336449, + 0.01261028740555048, + 0.01643272489309311, + 0.009424339048564434, + 0.01162036694586277, + 0.008967585861682892, + 0.04886365309357643, + 0.017026500776410103, + 0.016852740198373795, + 0.016078751534223557, + 0.02132822945713997, + 0.06007973104715347, + 0.03603265807032585, + 0.03551294654607773, + 0.007642693351954222, + 0.008908604271709919, + 0.0067039052955806255, + 0.00922450702637434, + 0.008096576668322086, + 0.028092285618185997, + 0.02935675159096718, + 0.012633217498660088, + 0.015759024769067764, + 0.01898862235248089, + 0.021268241107463837, + 0.025128409266471863, + 0.0189106035977602, + 0.015387491323053837, + 0.021327584981918335, + 0.014123713597655296, + 0.002210671780630946, + 0.0031910629477351904, + 0.003466479480266571, + 0.002378564327955246, + 0.002664869884029031, + 0.0038178886752575636, + 0.004165921360254288, + 0.0027572920080274343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 8 + ], + "selected_reward": -0.0017116909148171544 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a surreal and clean style, a high-angle medium shot establishes the scene of a playful conflict between two cats on a spotless, reflective white floor, illuminated by bright, shadowless studio lighting. A fluffy, long-haired Persian cat with pristine white fur lies on its side, its sturdy body protectively huddled over a small pile of brightly colored, crumpled pet toys. Its front paws firmly grasp the plastic rod of a feather wand, its vibrant blue and purple feathers perched near its face. Opposite it, a sleek Siamese cat, distinguished by its cream-colored body, black seal-point markings, and piercing blue, almond-shaped eyes, crouches in a low, predatory posture. As the Siamese cat swiftly and decisively brings its paw to bear on the feather wand, the camera executes a quick, seamless push-in, transitioning to an extreme close-up, focusing on the precise moment their paws meet. In this series of mock battles, the camera captures with microscopic clarity the short hair of the Siamese pressed against the dense, soft fur of the Persian, and the gentle, translucent claws of the Siamese, only partially unsheathed, gleaming in the even light, perfectly conveying the controlled, playful nature of their interaction against the pristine, minimalist backdrop.", + "seed": 24953, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_13/index_3889_seed_24953_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04698950294405222, + "vlm_logprob": -0.04698950294405222, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.04698950294405222, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13842910528182983 + ], + "opposite_logprob_tokens": [ + -2.2634291648864746 + ], + "logprob_reward": -0.13842910528182983, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.13842910528182983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13101087510585785, + 0.13336272537708282, + 0.10294333100318909, + 0.06217345967888832, + 0.08398672938346863, + 0.08435528725385666, + 0.16610002517700195, + 0.10251851379871368, + 0.07268623262643814, + 0.06086282804608345 + ], + "credit_map": [ + 0.01047013234347105, + 0.018325509503483772, + 0.01953817903995514, + 0.013704712502658367, + 0.01294127106666565, + 0.02330176718533039, + 0.010064443573355675, + 0.022664861753582954, + 0.013421721756458282, + 0.022650303319096565, + 0.0192324910312891, + 0.009118180721998215, + 0.014913476072251797, + 0.031152568757534027, + 0.008701645769178867, + 0.014172332361340523, + 0.011570429429411888, + 0.021685995161533356, + 0.013821021653711796, + 0.009928240440785885, + 0.008478967472910881, + 0.02043602056801319, + 0.008071711286902428, + 0.008950946852564812, + 0.012127651832997799, + 0.009980741888284683, + 0.00793437473475933, + 0.0065568010322749615, + 0.007177695631980896, + 0.007719524670392275, + 0.0058274162001907825, + 0.004849249497056007, + 0.020004725083708763, + 0.012020880356431007, + 0.010587361641228199, + 0.007519387174397707, + 0.007772077806293964, + 0.010974489152431488, + 0.0067004975862801075, + 0.008407311514019966, + 0.015351376496255398, + 0.009451369754970074, + 0.01309987623244524, + 0.007591106463223696, + 0.008013678714632988, + 0.013573783449828625, + 0.0050912462174892426, + 0.012182848528027534, + 0.031193088740110397, + 0.03786762058734894, + 0.011572426185011864, + 0.021427178755402565, + 0.01815417967736721, + 0.023699436336755753, + 0.009624236263334751, + 0.012561840005218983, + 0.013500637374818325, + 0.013787100091576576, + 0.010711506940424442, + 0.00679360143840313, + 0.02552865445613861, + 0.013529683463275433, + 0.009744502604007721, + 0.008922820910811424, + 0.012097450904548168, + 0.009880917146801949, + 0.014598720706999302, + 0.01017515268176794, + 0.008359472267329693, + 0.006579542066901922, + 0.004550436045974493, + 0.006444541271775961, + 0.0042679463513195515, + 0.01892262138426304, + 0.0024621570482850075, + 0.0020049738232046366, + 0.005549449007958174, + 0.02138994261622429, + 0.0026306400541216135, + 0.003635099856182933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.13842910528182983 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016087958589196205 + ], + "opposite_logprob_tokens": [ + -4.141088008880615 + ], + "logprob_reward": -0.016087958589196205, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016087958589196205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29653817415237427, + 0.0714506059885025, + 0.034688517451286316, + 0.0884876698255539, + 0.10750222951173782, + 0.07609213888645172, + 0.18129003047943115, + 0.07528270781040192, + 0.02749922126531601, + 0.0411687009036541 + ], + "credit_map": [ + 0.019278837367892265, + 0.08362264931201935, + 0.08225854486227036, + 0.02094673551619053, + 0.01795322820544243, + 0.03471959009766579, + 0.018385538831353188, + 0.019373048096895218, + 0.0075834705494344234, + 0.01242948416620493, + 0.01123538613319397, + 0.005942260846495628, + 0.007032597437500954, + 0.014022002927958965, + 0.004972374066710472, + 0.008233033120632172, + 0.004296666942536831, + 0.005430491175502539, + 0.0048997774720191956, + 0.003987452015280724, + 0.00398402800783515, + 0.004708908963948488, + 0.0031622950918972492, + 0.004218899644911289, + 0.016001364216208458, + 0.013491729274392128, + 0.013039871118962765, + 0.0064242021180689335, + 0.011839049868285656, + 0.00920577347278595, + 0.007909832522273064, + 0.010575847700238228, + 0.03032257966697216, + 0.018174555152654648, + 0.01651225984096527, + 0.006051850970834494, + 0.005648063030093908, + 0.015154877677559853, + 0.005346489138901234, + 0.010291554033756256, + 0.01269566360861063, + 0.011720562353730202, + 0.012855881825089455, + 0.004905675072222948, + 0.008292711339890957, + 0.014079885557293892, + 0.004912241827696562, + 0.006629518698900938, + 0.033034294843673706, + 0.037551142275333405, + 0.01743677817285061, + 0.02352139912545681, + 0.021757986396551132, + 0.025032849982380867, + 0.009230309166014194, + 0.013725261203944683, + 0.013230355456471443, + 0.009896792471408844, + 0.01195218414068222, + 0.008392703719437122, + 0.009568443521857262, + 0.008901839144527912, + 0.006022397428750992, + 0.007317991461604834, + 0.0039059265982359648, + 0.004344516433775425, + 0.004249719437211752, + 0.003825674531981349, + 0.004170522093772888, + 0.0030294759199023247, + 0.0020827199332416058, + 0.001890665735118091, + 0.004511072766035795, + 0.0095352279022336, + 0.005321088712662458, + 0.003803728614002466, + 0.002789718098938465, + 0.01049356535077095, + 0.0019274608930572867, + 0.0027868396136909723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.016087958589196205 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536413036286831 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536413036286831, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536413036286831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26459115743637085, + 0.05995430797338486, + 0.03910837322473526, + 0.11549542099237442, + 0.08741254359483719, + 0.05297607555985451, + 0.20226110517978668, + 0.09811543673276901, + 0.03410514444112778, + 0.04598046839237213 + ], + "credit_map": [ + 0.020611777901649475, + 0.05422139912843704, + 0.050331804901361465, + 0.012765677645802498, + 0.025731835514307022, + 0.044303048402071, + 0.015922604128718376, + 0.04070298746228218, + 0.0059166583232581615, + 0.00902751088142395, + 0.009768507443368435, + 0.006685046013444662, + 0.005010323598980904, + 0.010568258352577686, + 0.005444993264973164, + 0.007533004507422447, + 0.004302277695387602, + 0.005416189320385456, + 0.005272933281958103, + 0.004911605268716812, + 0.003446855815127492, + 0.005598477553576231, + 0.004693024326115847, + 0.005467003211379051, + 0.024612843990325928, + 0.021963005885481834, + 0.015027446672320366, + 0.00854345504194498, + 0.014267468824982643, + 0.013899733312427998, + 0.009089808911085129, + 0.008091646246612072, + 0.026058007031679153, + 0.016565470024943352, + 0.011510543525218964, + 0.005213067401200533, + 0.005076957400888205, + 0.0126091493293643, + 0.003753561293706298, + 0.006625777576118708, + 0.010768047533929348, + 0.007557904347777367, + 0.006926324218511581, + 0.004274474456906319, + 0.004665139596909285, + 0.009007574059069157, + 0.0033337664790451527, + 0.006442840676754713, + 0.03291064128279686, + 0.04801660031080246, + 0.014187240041792393, + 0.028768770396709442, + 0.02321033738553524, + 0.029079709202051163, + 0.011451845057308674, + 0.014635946601629257, + 0.014036674983799458, + 0.012660750187933445, + 0.018046213313937187, + 0.011507985182106495, + 0.010763710364699364, + 0.012869523838162422, + 0.009869041852653027, + 0.00836152397096157, + 0.006379605736583471, + 0.005203031934797764, + 0.004156783223152161, + 0.004980402998626232, + 0.005462875589728355, + 0.003958679735660553, + 0.002270944882184267, + 0.0016928204568102956, + 0.0033121658489108086, + 0.011812770739197731, + 0.0048169041983783245, + 0.0040674046613276005, + 0.0036981082521378994, + 0.012739010155200958, + 0.002468771766871214, + 0.0030653320718556643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.012536413036286831 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810694068670273 + ], + "opposite_logprob_tokens": [ + -3.288106918334961 + ], + "logprob_reward": -0.03810694068670273, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810694068670273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16300250589847565, + 0.08065841346979141, + 0.08914902806282043, + 0.11413615942001343, + 0.1220681294798851, + 0.06308864802122116, + 0.09450637549161911, + 0.06229310482740402, + 0.09719537943601608, + 0.11390228569507599 + ], + "credit_map": [ + 0.020078323781490326, + 0.02094387635588646, + 0.02009889855980873, + 0.013250711373984814, + 0.016977543011307716, + 0.027932418510317802, + 0.012833872810006142, + 0.030886860564351082, + 0.007450413424521685, + 0.013337490148842335, + 0.013828002847731113, + 0.01003745011985302, + 0.0066224997863173485, + 0.014255410991609097, + 0.007391905412077904, + 0.007735242135822773, + 0.008971850387752056, + 0.02512252889573574, + 0.007632321212440729, + 0.0057601178996264935, + 0.006063712295144796, + 0.0259613748639822, + 0.004867329727858305, + 0.004769797902554274, + 0.021250497549772263, + 0.013629565946757793, + 0.015000338666141033, + 0.00878091249614954, + 0.01698126643896103, + 0.013865029439330101, + 0.010923792608082294, + 0.01370476558804512, + 0.031426042318344116, + 0.020363474264740944, + 0.017776403576135635, + 0.00833889003843069, + 0.0064344764687120914, + 0.016067979857325554, + 0.007188177667558193, + 0.014472686685621738, + 0.00750426622107625, + 0.008816787041723728, + 0.008518646471202374, + 0.006229919381439686, + 0.005307718645781279, + 0.009748149663209915, + 0.0054818978533148766, + 0.011481262743473053, + 0.014744162559509277, + 0.022779088467359543, + 0.008592755533754826, + 0.012184353545308113, + 0.009988097473978996, + 0.01303310040384531, + 0.005473813973367214, + 0.007711009122431278, + 0.007986662909388542, + 0.00874810479581356, + 0.009973198175430298, + 0.00612438702955842, + 0.007465196307748556, + 0.008162526413798332, + 0.005817829631268978, + 0.00801519863307476, + 0.01991794817149639, + 0.00868508405983448, + 0.01990463212132454, + 0.01583404280245304, + 0.010568329133093357, + 0.009278867393732071, + 0.005924507509917021, + 0.0070819733664393425, + 0.008939004503190517, + 0.0312277153134346, + 0.008063715882599354, + 0.005152339581400156, + 0.010492067784070969, + 0.039653580635786057, + 0.004044374451041222, + 0.006329474505037069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.03810694068670273 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0297870971262455 + ], + "opposite_logprob_tokens": [ + -3.529787063598633 + ], + "logprob_reward": -0.0297870971262455, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0297870971262455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08191975206136703, + 0.05632826313376427, + 0.06391571462154388, + 0.12519018352031708, + 0.054425403475761414, + 0.12100262939929962, + 0.14390380680561066, + 0.11451146006584167, + 0.11201599985361099, + 0.12678678333759308 + ], + "credit_map": [ + 0.00801184494048357, + 0.01493048295378685, + 0.016395116224884987, + 0.0057158879935741425, + 0.006075347773730755, + 0.011007850989699364, + 0.007805685978382826, + 0.011977528221905231, + 0.005204579792916775, + 0.008812365122139454, + 0.009351585991680622, + 0.0057021332904696465, + 0.004655680153518915, + 0.012868753634393215, + 0.003867887891829014, + 0.005865275394171476, + 0.008306724950671196, + 0.016319135203957558, + 0.0064220065250992775, + 0.0038648603949695826, + 0.003550357185304165, + 0.018232475966215134, + 0.0034511948470026255, + 0.0037689616438001394, + 0.022200340405106544, + 0.017290545627474785, + 0.016235388815402985, + 0.008930029347538948, + 0.02050228975713253, + 0.014087256044149399, + 0.012663486413657665, + 0.013280851766467094, + 0.009367539547383785, + 0.008745694532990456, + 0.008139568381011486, + 0.00558237312361598, + 0.004718002397567034, + 0.007044181693345308, + 0.00439293310046196, + 0.006435110699385405, + 0.016686895862221718, + 0.026934951543807983, + 0.016566192731261253, + 0.010314775630831718, + 0.013188359327614307, + 0.01662690006196499, + 0.007236195262521505, + 0.013448360376060009, + 0.02062270976603031, + 0.03250760957598686, + 0.014007853344082832, + 0.018448833376169205, + 0.01656116358935833, + 0.020514408126473427, + 0.008643070235848427, + 0.012598146684467793, + 0.014779852703213692, + 0.01629948802292347, + 0.013139540329575539, + 0.009006661362946033, + 0.02539849653840065, + 0.015416795387864113, + 0.011680204421281815, + 0.008790425024926662, + 0.024818867444992065, + 0.014302216470241547, + 0.02135809138417244, + 0.017780490219593048, + 0.010347805917263031, + 0.010532904416322708, + 0.006229532416909933, + 0.006646087858825922, + 0.010726839303970337, + 0.03593727573752403, + 0.011656577698886395, + 0.006459732539951801, + 0.010490362532436848, + 0.04112347960472107, + 0.0038761040195822716, + 0.006516411434859037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 3 + ], + "selected_reward": -0.0297870971262455 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a surreal and clean style, a high-angle medium shot establishes the scene of a playful conflict between two cats on a spotless, reflective white floor, illuminated by bright, shadowless studio lighting. A fluffy, long-haired Persian cat with pristine white fur lies on its side, its sturdy body protectively huddled over a small pile of brightly colored, crumpled pet toys. Its front paws firmly grasp the plastic rod of a feather wand, its vibrant blue and purple feathers perched near its face. Opposite it, a sleek Siamese cat, distinguished by its cream-colored body, black seal-point markings, and piercing blue, almond-shaped eyes, crouches in a low, predatory posture. As the Siamese cat swiftly and decisively brings its paw to bear on the feather wand, the camera executes a quick, seamless push-in, transitioning to an extreme close-up, focusing on the precise moment their paws meet. In this series of mock battles, the camera captures with microscopic clarity the short hair of the Siamese pressed against the dense, soft fur of the Persian, and the gentle, translucent claws of the Siamese, only partially unsheathed, gleaming in the even light, perfectly conveying the controlled, playful nature of their interaction against the pristine, minimalist backdrop.", + "seed": 24954, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_13/index_3889_seed_24954_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.054596190433949233, + "vlm_logprob": -0.054596190433949233, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.054596190433949233, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17249052226543427 + ], + "opposite_logprob_tokens": [ + -2.2974905967712402 + ], + "logprob_reward": -0.17249052226543427, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.17249052226543427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12115585058927536, + 0.10309100896120071, + 0.046124327927827835, + 0.10318851470947266, + 0.051083557307720184, + 0.16465456783771515, + 0.06520730257034302, + 0.06063627451658249, + 0.21922209858894348, + 0.06563650071620941 + ], + "credit_map": [ + 0.017680400982499123, + 0.013997400179505348, + 0.014378227293491364, + 0.021229654550552368, + 0.020677728578448296, + 0.013721263036131859, + 0.010681774467229843, + 0.00878940150141716, + 0.011265565641224384, + 0.01595223695039749, + 0.018099144101142883, + 0.016914991661906242, + 0.012760822661221027, + 0.010900936089456081, + 0.010645019821822643, + 0.0065522827208042145, + 0.006062333006411791, + 0.006207967177033424, + 0.007112691644579172, + 0.00884766224771738, + 0.0060476623475551605, + 0.005305131431668997, + 0.003886240301653743, + 0.0026546348817646503, + 0.011535661295056343, + 0.013025341555476189, + 0.022164059802889824, + 0.01739863120019436, + 0.011225264519453049, + 0.011473755352199078, + 0.008454423397779465, + 0.007911373861134052, + 0.006839027628302574, + 0.006271717604249716, + 0.008087409660220146, + 0.00838957168161869, + 0.006502668373286724, + 0.005257950164377689, + 0.004866356495767832, + 0.004868855234235525, + 0.022800765931606293, + 0.022731764242053032, + 0.022785592824220657, + 0.01403986569494009, + 0.020550847053527832, + 0.020384499803185463, + 0.014963358640670776, + 0.02639787457883358, + 0.008950628340244293, + 0.0072409966960549355, + 0.007348375394940376, + 0.008057723753154278, + 0.009189673699438572, + 0.010243920609354973, + 0.007253916002810001, + 0.006922065746039152, + 0.0067481533624231815, + 0.006713376380503178, + 0.008318020962178707, + 0.007170635275542736, + 0.012466582469642162, + 0.00763228815048933, + 0.007916413247585297, + 0.0036707986146211624, + 0.014136683195829391, + 0.014739910140633583, + 0.019122397527098656, + 0.02293425239622593, + 0.028795242309570312, + 0.02792157605290413, + 0.06236470118165016, + 0.029207326471805573, + 0.008485211990773678, + 0.013163771480321884, + 0.004133657086640596, + 0.011601312085986137, + 0.010771231725811958, + 0.005325517151504755, + 0.005560760851949453, + 0.006595030892640352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.17249052226543427 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004670187830924988 + ], + "opposite_logprob_tokens": [ + -5.379670143127441 + ], + "logprob_reward": -0.004670187830924988, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004670187830924988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08611778169870377, + 0.04912988096475601, + 0.11653456836938858, + 0.03695792704820633, + 0.055779535323381424, + 0.25882452726364136, + 0.037791650742292404, + 0.045422207564115524, + 0.11769191920757294, + 0.19574999809265137 + ], + "credit_map": [ + 0.008898080326616764, + 0.010295054875314236, + 0.009167131036520004, + 0.01563262939453125, + 0.015049714595079422, + 0.012926431372761726, + 0.007724284194409847, + 0.006424457300454378, + 0.006279005669057369, + 0.007165099959820509, + 0.008467289619147778, + 0.006662168074399233, + 0.00567593052983284, + 0.005452181212604046, + 0.004936867393553257, + 0.004491341765969992, + 0.013582915998995304, + 0.015595698729157448, + 0.013884980231523514, + 0.02380811981856823, + 0.01228110771626234, + 0.009524938650429249, + 0.016060274094343185, + 0.011796539649367332, + 0.0037580153439193964, + 0.004714087583124638, + 0.006332604214549065, + 0.004955842159688473, + 0.004129230044782162, + 0.004299662541598082, + 0.00362196727655828, + 0.005146515090018511, + 0.0078048137947916985, + 0.007585726212710142, + 0.00883223582059145, + 0.008764520287513733, + 0.00493684783577919, + 0.004359004087746143, + 0.006256874185055494, + 0.007239508908241987, + 0.03983115777373314, + 0.035832975059747696, + 0.04063786566257477, + 0.022956833243370056, + 0.029382364824414253, + 0.02945355698466301, + 0.02445622906088829, + 0.03627351298928261, + 0.003284235717728734, + 0.004561060573905706, + 0.0033074249513447285, + 0.004651813302189112, + 0.007408286910504103, + 0.004416642244905233, + 0.006590943317860365, + 0.00357124418951571, + 0.003900813637301326, + 0.005365963559597731, + 0.004192698746919632, + 0.004463757388293743, + 0.007678394205868244, + 0.005865531507879496, + 0.009539262391626835, + 0.0044157858937978745, + 0.010206572711467743, + 0.010150798596441746, + 0.011460318230092525, + 0.012695327401161194, + 0.017311904579401016, + 0.013449463993310928, + 0.02800162322819233, + 0.014415904879570007, + 0.012897309847176075, + 0.03194523602724075, + 0.01030725333839655, + 0.050351839512586594, + 0.02567446231842041, + 0.01833377778530121, + 0.019363150000572205, + 0.026876958087086678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.004670187830924988 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865466430783272 + ], + "opposite_logprob_tokens": [ + -3.048654556274414 + ], + "logprob_reward": -0.04865466430783272, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865466430783272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041326068341732025, + 0.029958559200167656, + 0.07563066482543945, + 0.026897717267274857, + 0.08428916335105896, + 0.2746410667896271, + 0.07182245701551437, + 0.04334016144275665, + 0.1503814309835434, + 0.2017126977443695 + ], + "credit_map": [ + 0.00402707327157259, + 0.005040255840867758, + 0.00490908557549119, + 0.00859788991510868, + 0.006775498855859041, + 0.005774159450083971, + 0.0034338366240262985, + 0.0027682713698595762, + 0.0037686843425035477, + 0.004453146830201149, + 0.003848670981824398, + 0.004037805367261171, + 0.0038007607217878103, + 0.004392002243548632, + 0.0027827320154756308, + 0.0028747576288878918, + 0.008667197078466415, + 0.010552773252129555, + 0.008309819735586643, + 0.014523735269904137, + 0.007663863245397806, + 0.00708550401031971, + 0.010596240870654583, + 0.00823153555393219, + 0.0025316462852060795, + 0.0035164502914994955, + 0.0053182924166321754, + 0.004229968413710594, + 0.00381478201597929, + 0.002637282945215702, + 0.0024192172568291426, + 0.0024300809018313885, + 0.00857130903750658, + 0.00969582237303257, + 0.019924651831388474, + 0.015669098123908043, + 0.007571348454803228, + 0.008679986000061035, + 0.006492526736110449, + 0.007684426382184029, + 0.04708998277783394, + 0.04214004799723625, + 0.04519447684288025, + 0.02353188768029213, + 0.027283573523163795, + 0.03146296367049217, + 0.02333212085068226, + 0.03460603207349777, + 0.009343467652797699, + 0.010074100457131863, + 0.011052391491830349, + 0.008164790458977222, + 0.008491246029734612, + 0.009761610999703407, + 0.006475415080785751, + 0.008459450677037239, + 0.005132047459483147, + 0.007269333116710186, + 0.004453158937394619, + 0.00546749122440815, + 0.007329707499593496, + 0.0046715522184967995, + 0.005881884600967169, + 0.0031349908094853163, + 0.013788958080112934, + 0.01453271135687828, + 0.01417233981192112, + 0.01780838891863823, + 0.022609254345297813, + 0.016433576121926308, + 0.03326398879289627, + 0.017772220075130463, + 0.013793599791824818, + 0.032553091645240784, + 0.010724593885242939, + 0.04919020086526871, + 0.026848386973142624, + 0.020484695211052895, + 0.020791728049516678, + 0.027326414361596107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.04865466430783272 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307162016630173 + ], + "opposite_logprob_tokens": [ + -3.1680715084075928 + ], + "logprob_reward": -0.04307162016630173, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307162016630173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044227663427591324, + 0.07020750641822815, + 0.030230682343244553, + 0.05961686372756958, + 0.10595006495714188, + 0.39680609107017517, + 0.10246776789426804, + 0.08590366691350937, + 0.05223965272307396, + 0.052350059151649475 + ], + "credit_map": [ + 0.004262831062078476, + 0.0048863948322832584, + 0.004939540289342403, + 0.005964724812656641, + 0.006853808648884296, + 0.005704040639102459, + 0.006220812909305096, + 0.005395512096583843, + 0.006760610733181238, + 0.00990883819758892, + 0.008542162366211414, + 0.00829725805670023, + 0.008088304661214352, + 0.008156850002706051, + 0.011218532919883728, + 0.009234955534338951, + 0.0034898712765425444, + 0.007842586375772953, + 0.003081100294366479, + 0.002832158701494336, + 0.00419482821598649, + 0.0027284258976578712, + 0.0034456218127161264, + 0.0026160909328609705, + 0.006033036857843399, + 0.00783554371446371, + 0.01456400752067566, + 0.008347699418663979, + 0.00727422209456563, + 0.005567631684243679, + 0.005383602809160948, + 0.004611118230968714, + 0.010204914025962353, + 0.011857670731842518, + 0.02293855883181095, + 0.019756965339183807, + 0.010274070315063, + 0.009384670294821262, + 0.010039368644356728, + 0.011493835598230362, + 0.06934911757707596, + 0.05652971565723419, + 0.061749596148729324, + 0.033681366592645645, + 0.04026597738265991, + 0.0424763560295105, + 0.03620510548353195, + 0.05654885247349739, + 0.011703209020197392, + 0.015062276273965836, + 0.014925586991012096, + 0.011629996821284294, + 0.012831280939280987, + 0.013195432722568512, + 0.012400402687489986, + 0.010719578713178635, + 0.00864980649203062, + 0.009423866868019104, + 0.010585909709334373, + 0.008615730330348015, + 0.01530767697840929, + 0.011794592253863811, + 0.013973293825984001, + 0.00755279092118144, + 0.00314717716537416, + 0.0038741230964660645, + 0.00521657383069396, + 0.007730626035481691, + 0.007585638668388128, + 0.005391888320446014, + 0.011700239032506943, + 0.00759338540956378, + 0.003801646875217557, + 0.010102413594722748, + 0.0042645628564059734, + 0.011458773165941238, + 0.00749296136200428, + 0.0033071578945964575, + 0.0056517827324569225, + 0.006270765792578459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.04307162016630173 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a spotless, reflective white floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fluffy, long-haired Persian cat with white fur lying on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleek Siamese cat with cream-colored body and black seal-point markings crouching opposite the Persian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the Siamese cat's paw make contact with the feather wand held by the Persian cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a quick push-in transitioning to an extreme close-up of the paws meeting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093957599252462 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004093957599252462, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093957599252462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06566884368658066, + 0.03335617110133171, + 0.03153333440423012, + 0.060472339391708374, + 0.09765718877315521, + 0.10466329753398895, + 0.12972116470336914, + 0.19435741007328033, + 0.08315666764974594, + 0.19941358268260956 + ], + "credit_map": [ + 0.008068283088505268, + 0.007933717221021652, + 0.00913022831082344, + 0.010160504840314388, + 0.01126463245600462, + 0.007656420581042767, + 0.005798915401101112, + 0.005656139925122261, + 0.0037076252046972513, + 0.004514190834015608, + 0.005960122682154179, + 0.005012381821870804, + 0.004510057158768177, + 0.003509882604703307, + 0.003584324149414897, + 0.0025575903709977865, + 0.004337606951594353, + 0.003906847443431616, + 0.0037347825709730387, + 0.0039123715832829475, + 0.004326934460550547, + 0.0045885080471634865, + 0.003854932263493538, + 0.002871350385248661, + 0.005925661418586969, + 0.006927285809069872, + 0.009460455738008022, + 0.009354967623949051, + 0.0077624013647437096, + 0.0067609804682433605, + 0.008030015043914318, + 0.006250570062547922, + 0.009955029934644699, + 0.010648132301867008, + 0.021396329626441002, + 0.021205183118581772, + 0.0076750596053898335, + 0.008479557000100613, + 0.008106670342385769, + 0.010191221721470356, + 0.01277407817542553, + 0.013713429681956768, + 0.013567131944000721, + 0.008139611221849918, + 0.013685658574104309, + 0.01428624615073204, + 0.010252043604850769, + 0.01824510656297207, + 0.01682536117732525, + 0.01825210638344288, + 0.01638117991387844, + 0.016632117331027985, + 0.0190054252743721, + 0.017160596325993538, + 0.012866081669926643, + 0.012598291039466858, + 0.016977421939373016, + 0.02229287102818489, + 0.01928822323679924, + 0.02045593596994877, + 0.03659156337380409, + 0.02501068077981472, + 0.035613298416137695, + 0.018127400428056717, + 0.006308038253337145, + 0.008542351424694061, + 0.008266022428870201, + 0.011915018782019615, + 0.011778299696743488, + 0.011178195476531982, + 0.010967722162604332, + 0.01420101709663868, + 0.019020458683371544, + 0.0523235909640789, + 0.012114522978663445, + 0.027373645454645157, + 0.038146037608385086, + 0.01362328976392746, + 0.016806012019515038, + 0.02000601775944233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.004093957599252462 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_14/index_tensor([8630, 8630])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_14/index_tensor([8630, 8630])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c48fd468c7be1c2207bf8241e7eb37d5cd85687a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_14/index_tensor([8630, 8630])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic wide shot, shot slightly from an upward angle, captures the Taj Mahal, a majestic white marble structure, standing alone against the fury of nature beneath a dark, storm-ravaged sky. Rolling, inky black storm clouds nearly bear down on the iconic central dome, while occasional flashes of pale lightning split the sky, briefly illuminating the building's rain-slicked facade. Below, raging floodwaters relentlessly devour its vast grounds, destroying the once tranquil reflecting pools and meticulously manicured gardens, replaced by a murky ocean of mud, sand, and debris. This tawny torrent crashes violently against the building's base, stirring up layers of white spray, the murky waters rising to the brink of the sacred white marble platform. The entire scene is presented in stunning 4K resolution, making every splashing drop of water and the dynamics of the distant clouds clearly visible, while being covered with a subtle 35mm film grain texture, adding a touch of tragic historical heaviness to this apocalyptic magnificent scene, creating an atmosphere of grandeur, oppression and tragic aesthetics.", + "seed": 11501, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_14/index_8630_seed_11501_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17255646996200086, + "vlm_logprob": -0.17255646996200086, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.17255646996200086, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004339326173067093 + ], + "opposite_logprob_tokens": [ + -6.504339218139648 + ], + "logprob_reward": -0.004339326173067093, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.004339326173067093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20566053688526154, + 0.09583944082260132, + 0.0771905928850174, + 0.2237989753484726, + 0.09297329187393188, + 0.05676481872797012, + 0.10853829979896545, + 0.03680046647787094, + 0.056203801184892654, + 0.046229735016822815 + ], + "credit_map": [ + 0.035168010741472244, + 0.023257669061422348, + 0.027189185842871666, + 0.022473333403468132, + 0.023563174530863762, + 0.025753134861588478, + 0.027646038681268692, + 0.020609986037015915, + 0.021224427968263626, + 0.01802404224872589, + 0.010008921846747398, + 0.007529738824814558, + 0.011712228879332542, + 0.013768700882792473, + 0.006060496903955936, + 0.007510881870985031, + 0.011198237538337708, + 0.0127584682777524, + 0.00973532348871231, + 0.006814463064074516, + 0.010880529880523682, + 0.010758422315120697, + 0.00803321786224842, + 0.0070119271986186504, + 0.03617077320814133, + 0.04857354611158371, + 0.015757504850625992, + 0.01802108623087406, + 0.04132095351815224, + 0.031905725598335266, + 0.019313259050250053, + 0.012736130505800247, + 0.01428173203021288, + 0.020496418699622154, + 0.006902821362018585, + 0.007641524542123079, + 0.018214579671621323, + 0.013626852072775364, + 0.006648161914199591, + 0.00516119971871376, + 0.009565549902617931, + 0.009437426924705505, + 0.004956695716828108, + 0.005411559250205755, + 0.00939297303557396, + 0.007380881812423468, + 0.005563954822719097, + 0.0050557805225253105, + 0.015996936708688736, + 0.023665014654397964, + 0.011858928017318249, + 0.011577684432268143, + 0.012637365609407425, + 0.015769796445965767, + 0.007600795477628708, + 0.009431781247258186, + 0.006202479358762503, + 0.006781814619898796, + 0.005257268436253071, + 0.003530977526679635, + 0.003078190376982093, + 0.005238538142293692, + 0.0034910819958895445, + 0.0032201160211116076, + 0.009176135994493961, + 0.008382091298699379, + 0.00733294989913702, + 0.004858209751546383, + 0.0040328484028577805, + 0.008474458009004593, + 0.005249659530818462, + 0.008697450160980225, + 0.0068482160568237305, + 0.006542671006172895, + 0.007586397230625153, + 0.005126065108925104, + 0.004930965136736631, + 0.0061919912695884705, + 0.0036260313354432583, + 0.005377397872507572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.004339326173067093 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897897809743881 + ], + "opposite_logprob_tokens": [ + -2.578979015350342 + ], + "logprob_reward": -0.07897897809743881, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897897809743881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1558009684085846, + 0.08856528252363205, + 0.026983028277754784, + 0.15735875070095062, + 0.3974153697490692, + 0.038696106523275375, + 0.06414341181516647, + 0.03255263343453407, + 0.019518805667757988, + 0.01896561123430729 + ], + "credit_map": [ + 0.03667264059185982, + 0.01778973452746868, + 0.026818817481398582, + 0.010488324798643589, + 0.012071431614458561, + 0.021724550053477287, + 0.011329290457069874, + 0.018906183540821075, + 0.018580565229058266, + 0.013614439405500889, + 0.011657856404781342, + 0.007131543010473251, + 0.009004347026348114, + 0.012596365995705128, + 0.007883473299443722, + 0.008096693083643913, + 0.004549234174191952, + 0.003940034192055464, + 0.003712435718625784, + 0.0021458484698086977, + 0.003399947192519903, + 0.0033538425341248512, + 0.0026081535033881664, + 0.0032735327258706093, + 0.022149601951241493, + 0.03266555070877075, + 0.011543422006070614, + 0.012775442562997341, + 0.030740654096007347, + 0.022850994020700455, + 0.014661013148725033, + 0.00997207686305046, + 0.0645819753408432, + 0.09070909768342972, + 0.029230834916234016, + 0.028126496821641922, + 0.0760313868522644, + 0.05841666832566261, + 0.026391800493001938, + 0.023927122354507446, + 0.006652520038187504, + 0.007219545543193817, + 0.004844252485781908, + 0.0029226525221019983, + 0.004397686570882797, + 0.005413536448031664, + 0.002895347774028778, + 0.0043505653738975525, + 0.009940302930772305, + 0.016505660489201546, + 0.005493778735399246, + 0.005527402739971876, + 0.007751185912638903, + 0.00930057093501091, + 0.005526254884898663, + 0.0040982565842568874, + 0.00533981341868639, + 0.0058703674003481865, + 0.0041184211149811745, + 0.0026036896742880344, + 0.003333527594804764, + 0.0047476040199398994, + 0.003047144040465355, + 0.003492065006867051, + 0.00333716650493443, + 0.0038854985032230616, + 0.0018194743897765875, + 0.0017685132334008813, + 0.0023111978080123663, + 0.0028050884138792753, + 0.0017569491174072027, + 0.0018349176971241832, + 0.003053125226870179, + 0.0028658737428486347, + 0.0024775750935077667, + 0.002225731499493122, + 0.001873302971944213, + 0.0025043317582458258, + 0.0019076940370723605, + 0.0020579767879098654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.07897897809743881 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.693382978439331 + ], + "opposite_logprob_tokens": [ + -0.693382978439331 + ], + "logprob_reward": -0.693382978439331, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.693382978439331, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2174476832151413, + 0.13332970440387726, + 0.058059122413396835, + 0.06775999814271927, + 0.1700352132320404, + 0.04839533567428589, + 0.030485974624753, + 0.11622479557991028, + 0.08846025168895721, + 0.06980186700820923 + ], + "credit_map": [ + 0.046785276383161545, + 0.018825985491275787, + 0.03273839130997658, + 0.023125432431697845, + 0.018773943185806274, + 0.02711341716349125, + 0.02491465024650097, + 0.025170590728521347, + 0.030014969408512115, + 0.017487451434135437, + 0.017834700644016266, + 0.01008651778101921, + 0.011369273997843266, + 0.02132810279726982, + 0.01368566695600748, + 0.011523022316396236, + 0.010712021961808205, + 0.008373462595045567, + 0.009306389838457108, + 0.0053092180751264095, + 0.005232303868979216, + 0.007795963902026415, + 0.006071677431464195, + 0.005258085671812296, + 0.010703911073505878, + 0.012627952732145786, + 0.00694887712597847, + 0.005007792264223099, + 0.01237103808671236, + 0.008952333591878414, + 0.0055910455994307995, + 0.005557044874876738, + 0.023168284446001053, + 0.03219681233167648, + 0.01394216064363718, + 0.013171661645174026, + 0.03132051229476929, + 0.024270232766866684, + 0.015914902091026306, + 0.016050633043050766, + 0.007931589148938656, + 0.008456593379378319, + 0.005897521506994963, + 0.004125009290874004, + 0.005296348128467798, + 0.006501454394310713, + 0.00425702752545476, + 0.005929790902882814, + 0.00469645531848073, + 0.003949691075831652, + 0.0035073456820100546, + 0.002791042672470212, + 0.0029972477350383997, + 0.0040510171093046665, + 0.004585662391036749, + 0.003907513804733753, + 0.02108018845319748, + 0.011353818699717522, + 0.02283589541912079, + 0.010277390480041504, + 0.007791639771312475, + 0.016760116443037987, + 0.012715167365968227, + 0.013410567305982113, + 0.016169210895895958, + 0.012497488409280777, + 0.011069734580814838, + 0.006968643516302109, + 0.00874889176338911, + 0.013063023798167706, + 0.010385747067630291, + 0.009557517245411873, + 0.012750450521707535, + 0.005634458269923925, + 0.012706185691058636, + 0.006070801522582769, + 0.005218714941293001, + 0.009173786267638206, + 0.00895981676876545, + 0.00928766094148159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.693382978439331 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070479527115822 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.016070479527115822, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070479527115822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.144539937376976, + 0.045896563678979874, + 0.042098402976989746, + 0.4118458032608032, + 0.1199561059474945, + 0.03291867673397064, + 0.06616263091564178, + 0.07810147851705551, + 0.03951673582196236, + 0.018963655456900597 + ], + "credit_map": [ + 0.026833655312657356, + 0.020521774888038635, + 0.01778062805533409, + 0.01919575035572052, + 0.01759212277829647, + 0.018907664343714714, + 0.012080480344593525, + 0.011627858504652977, + 0.009089760482311249, + 0.007953215390443802, + 0.005320235155522823, + 0.003938738256692886, + 0.006122931372374296, + 0.006027604918926954, + 0.0037658321671187878, + 0.003678243374451995, + 0.005421083886176348, + 0.011061854660511017, + 0.002840265864506364, + 0.0037464317865669727, + 0.006889003328979015, + 0.006003082264214754, + 0.0035721310414373875, + 0.0025645482819527388, + 0.06299477815628052, + 0.08954361826181412, + 0.027164004743099213, + 0.034834396094083786, + 0.08302187919616699, + 0.058957844972610474, + 0.03644338250160217, + 0.01888592168688774, + 0.016599800437688828, + 0.028941212221980095, + 0.0070627904497087, + 0.01071186363697052, + 0.02381039410829544, + 0.016888782382011414, + 0.008895865641534328, + 0.007045398931950331, + 0.005610916297882795, + 0.007855270057916641, + 0.0024202249478548765, + 0.002346014603972435, + 0.004193008411675692, + 0.004467857535928488, + 0.0026508530136197805, + 0.0033745309337973595, + 0.009933740831911564, + 0.013610905036330223, + 0.0076176319271326065, + 0.005673449486494064, + 0.007935374975204468, + 0.00829568412154913, + 0.0074199591763317585, + 0.005675891414284706, + 0.013808401301503181, + 0.015279015526175499, + 0.011176067404448986, + 0.006577549502253532, + 0.00695626949891448, + 0.010327582247555256, + 0.00655780453234911, + 0.007418784312903881, + 0.005985813681036234, + 0.006269748788326979, + 0.005060933064669371, + 0.0034539203625172377, + 0.004588342737406492, + 0.005755273159593344, + 0.004363302607089281, + 0.004039403051137924, + 0.0026025227271020412, + 0.002453396562486887, + 0.002840354572981596, + 0.0029361124616116285, + 0.0018187762470915914, + 0.002260104287415743, + 0.0018727559363469481, + 0.002179631032049656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.016070479527115822 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001058757305145 + ], + "opposite_logprob_tokens": [ + -2.6950106620788574 + ], + "logprob_reward": -0.07001058757305145, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001058757305145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13901981711387634, + 0.07393761724233627, + 0.025772089138627052, + 0.07851307094097137, + 0.12044244259595871, + 0.055374711751937866, + 0.10706502944231033, + 0.2650417387485504, + 0.09551353007555008, + 0.039319999516010284 + ], + "credit_map": [ + 0.023059217259287834, + 0.016906581819057465, + 0.02663990668952465, + 0.012377706356346607, + 0.014515217393636703, + 0.018677305430173874, + 0.011920507065951824, + 0.014923371374607086, + 0.01176721602678299, + 0.009144788607954979, + 0.012916595675051212, + 0.006561384070664644, + 0.008298706263303757, + 0.009570636786520481, + 0.009099096059799194, + 0.006579198408871889, + 0.0034356818068772554, + 0.003490577917546034, + 0.0038799827452749014, + 0.002938513644039631, + 0.0030014102812856436, + 0.003579488955438137, + 0.0029823449440300465, + 0.0024640909396111965, + 0.01124673243612051, + 0.016833912581205368, + 0.006292062811553478, + 0.004669602960348129, + 0.0156555138528347, + 0.011187730357050896, + 0.00731066195294261, + 0.005316851660609245, + 0.028588466346263885, + 0.01797376200556755, + 0.012491906993091106, + 0.010211613029241562, + 0.014033649116754532, + 0.01829485222697258, + 0.00922690611332655, + 0.009621279314160347, + 0.009758234955370426, + 0.011550999246537685, + 0.006235805340111256, + 0.004431715700775385, + 0.008388448506593704, + 0.007408632431179285, + 0.003839075332507491, + 0.003761802799999714, + 0.01612386479973793, + 0.026157518848776817, + 0.009681637398898602, + 0.008162305690348148, + 0.015896057710051537, + 0.015642456710338593, + 0.007569862995296717, + 0.007831315509974957, + 0.046229347586631775, + 0.05386357754468918, + 0.036286361515522, + 0.021358471363782883, + 0.025571951642632484, + 0.03659551218152046, + 0.021272385492920876, + 0.023864123970270157, + 0.012781590223312378, + 0.017390232533216476, + 0.011941513977944851, + 0.009158453904092312, + 0.00986421387642622, + 0.014395872130990028, + 0.009618244133889675, + 0.01036340743303299, + 0.007517877966165543, + 0.0030090438667684793, + 0.00661054253578186, + 0.0031548230908811092, + 0.0031143997330218554, + 0.005607232917100191, + 0.0039059354458004236, + 0.006400141399353743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.07001058757305145 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide shot, shot slightly from an upward angle, captures the Taj Mahal, a majestic white marble structure, standing alone against the fury of nature beneath a dark, storm-ravaged sky. Rolling, inky black storm clouds nearly bear down on the iconic central dome, while occasional flashes of pale lightning split the sky, briefly illuminating the building's rain-slicked facade. Below, raging floodwaters relentlessly devour its vast grounds, destroying the once tranquil reflecting pools and meticulously manicured gardens, replaced by a murky ocean of mud, sand, and debris. This tawny torrent crashes violently against the building's base, stirring up layers of white spray, the murky waters rising to the brink of the sacred white marble platform. The entire scene is presented in stunning 4K resolution, making every splashing drop of water and the dynamics of the distant clouds clearly visible, while being covered with a subtle 35mm film grain texture, adding a touch of tragic historical heaviness to this apocalyptic magnificent scene, creating an atmosphere of grandeur, oppression and tragic aesthetics.", + "seed": 11502, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_14/index_8630_seed_11502_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15342139159329235, + "vlm_logprob": -0.15342139159329235, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.15342139159329235, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003234514268115163 + ], + "opposite_logprob_tokens": [ + -6.378234386444092 + ], + "logprob_reward": -0.003234514268115163, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.003234514268115163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22549839317798615, + 0.21177729964256287, + 0.09222612529993057, + 0.06874468922615051, + 0.06387937068939209, + 0.08321749418973923, + 0.08825629949569702, + 0.06252394616603851, + 0.06499233841896057, + 0.038884151726961136 + ], + "credit_map": [ + 0.035118378698825836, + 0.047859929502010345, + 0.017716893926262856, + 0.01756400428712368, + 0.040696192532777786, + 0.030878117308020592, + 0.02410801127552986, + 0.011556842364370823, + 0.020855441689491272, + 0.04288974031805992, + 0.01356023270636797, + 0.017223885282874107, + 0.05090774968266487, + 0.03009962849318981, + 0.028398267924785614, + 0.007842323742806911, + 0.010685602203011513, + 0.012128830887377262, + 0.010241711512207985, + 0.015293101780116558, + 0.0125649468973279, + 0.012364952825009823, + 0.011636865325272083, + 0.007310108281672001, + 0.009106810204684734, + 0.015901142731308937, + 0.005246971733868122, + 0.005322117358446121, + 0.013198093511164188, + 0.010003674775362015, + 0.00754561135545373, + 0.00242025894112885, + 0.012157530523836613, + 0.009609956294298172, + 0.0066186776384711266, + 0.006806103512644768, + 0.010619138367474079, + 0.008874189108610153, + 0.00623785424977541, + 0.002955910749733448, + 0.008567385375499725, + 0.01303388923406601, + 0.005142016801983118, + 0.006839278619736433, + 0.021580088883638382, + 0.011711315251886845, + 0.011976849287748337, + 0.00436666514724493, + 0.012860442511737347, + 0.012916701845824718, + 0.01125996932387352, + 0.010805315338075161, + 0.009993000887334347, + 0.011940333992242813, + 0.012169414199888706, + 0.006311116740107536, + 0.004986729007214308, + 0.009969545528292656, + 0.005028131417930126, + 0.008415410295128822, + 0.014914537779986858, + 0.008248629979789257, + 0.00817952398210764, + 0.0027814318891614676, + 0.008266239427030087, + 0.009697290137410164, + 0.008593614213168621, + 0.0061966110952198505, + 0.009676521643996239, + 0.008204749785363674, + 0.01001828070729971, + 0.004339025355875492, + 0.004635074641555548, + 0.005415372550487518, + 0.005448081996291876, + 0.0031380383297801018, + 0.0076485625468194485, + 0.004769503604620695, + 0.005353495012968779, + 0.0024760181549936533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.003234514268115163 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202586375176907 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202586375176907, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202586375176907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3738211691379547, + 0.18137472867965698, + 0.06169269606471062, + 0.08313114941120148, + 0.08122383058071136, + 0.036330871284008026, + 0.03688969835639, + 0.03539007157087326, + 0.061729349195957184, + 0.048416487872600555 + ], + "credit_map": [ + 0.07482355833053589, + 0.08329993486404419, + 0.027444446459412575, + 0.034108780324459076, + 0.05260599032044411, + 0.05421190708875656, + 0.02921491302549839, + 0.018111636862158775, + 0.02441325969994068, + 0.0580771267414093, + 0.010363228619098663, + 0.013096454553306103, + 0.0271387230604887, + 0.028170017525553703, + 0.012291708029806614, + 0.007824191823601723, + 0.00745030865073204, + 0.01943705417215824, + 0.0043015629053115845, + 0.0049274759367108345, + 0.008124973624944687, + 0.009712731465697289, + 0.004566001705825329, + 0.0031725855078548193, + 0.015745200216770172, + 0.018785186111927032, + 0.007105966564267874, + 0.005934388376772404, + 0.011871987022459507, + 0.012662644498050213, + 0.006397049408406019, + 0.004628722555935383, + 0.018107037991285324, + 0.022334715351462364, + 0.006496270187199116, + 0.006587408948689699, + 0.007614515256136656, + 0.012170723639428616, + 0.005214077420532703, + 0.002699073404073715, + 0.008840180933475494, + 0.008566254749894142, + 0.0023784602526575327, + 0.002092566341161728, + 0.004629058297723532, + 0.005382605828344822, + 0.0029538075905293226, + 0.0014879354275763035, + 0.007477603852748871, + 0.006029737181961536, + 0.004018304403871298, + 0.003087572054937482, + 0.0037431675009429455, + 0.004799467511475086, + 0.003423792077228427, + 0.004310049116611481, + 0.005840886849910021, + 0.005844941828399897, + 0.002575842896476388, + 0.005005736369639635, + 0.004101553000509739, + 0.005362460855394602, + 0.0041831000708043575, + 0.0024755503982305527, + 0.008569430559873581, + 0.011009011417627335, + 0.006993177812546492, + 0.007702658884227276, + 0.006967028137296438, + 0.008803602308034897, + 0.007947225123643875, + 0.003737209364771843, + 0.006968939211219549, + 0.010861406102776527, + 0.004560338798910379, + 0.006698454264551401, + 0.006449936423450708, + 0.00731407618150115, + 0.003562678350135684, + 0.0020006548147648573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.014202586375176907 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933929324150085 + ], + "opposite_logprob_tokens": [ + -0.6933929324150085 + ], + "logprob_reward": -0.6933929324150085, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933929324150085, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2060142606496811, + 0.27281883358955383, + 0.16094981133937836, + 0.07229677587747574, + 0.045553334057331085, + 0.06277976185083389, + 0.03805168718099594, + 0.040665801614522934, + 0.06806318461894989, + 0.03280651569366455 + ], + "credit_map": [ + 0.027425702661275864, + 0.05012674629688263, + 0.025728967040777206, + 0.016952460631728172, + 0.02319135144352913, + 0.02997053973376751, + 0.016115078702569008, + 0.016503416001796722, + 0.047645337879657745, + 0.06385374814271927, + 0.01939096301794052, + 0.013994671404361725, + 0.0531763955950737, + 0.041912347078323364, + 0.020734651014208794, + 0.012110746465623379, + 0.033262815326452255, + 0.041502390056848526, + 0.012281113304197788, + 0.008669132366776466, + 0.021366450935602188, + 0.025961561128497124, + 0.00944579392671585, + 0.008460542187094688, + 0.01283442135900259, + 0.015667980536818504, + 0.0049389987252652645, + 0.007036807015538216, + 0.009703344665467739, + 0.010335130617022514, + 0.007164699025452137, + 0.004615393467247486, + 0.00829357746988535, + 0.01079792995005846, + 0.004268205724656582, + 0.0033581051975488663, + 0.005750083830207586, + 0.007265307009220123, + 0.0036518408451229334, + 0.0021682854276150465, + 0.016117893159389496, + 0.012846553698182106, + 0.0058870259672403336, + 0.004715145565569401, + 0.006336715538054705, + 0.00975886732339859, + 0.0039426712319254875, + 0.0031748912297189236, + 0.007030638866126537, + 0.005816689692437649, + 0.0037465670611709356, + 0.0031912080012261868, + 0.004724969156086445, + 0.00481960317119956, + 0.0049188341945409775, + 0.0038031789008527994, + 0.005333703942596912, + 0.006397805642336607, + 0.0032058879733085632, + 0.004376078024506569, + 0.008951012045145035, + 0.0057184649631381035, + 0.004782485775649548, + 0.001900362316519022, + 0.007465014699846506, + 0.0148145267739892, + 0.006923030596226454, + 0.00722994701936841, + 0.009142324328422546, + 0.010511401109397411, + 0.008001431822776794, + 0.003975509665906429, + 0.005824126768857241, + 0.007344109937548637, + 0.003411879064515233, + 0.0023899369407445192, + 0.004378547426313162, + 0.005135561339557171, + 0.002711164066568017, + 0.001611191313713789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.6933929324150085 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629185914993286 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.007629185914993286, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629185914993286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3884221613407135, + 0.08289945870637894, + 0.05653839558362961, + 0.07665039598941803, + 0.06719394028186798, + 0.07614189386367798, + 0.11105544120073318, + 0.035909026861190796, + 0.06071677803993225, + 0.044472530484199524 + ], + "credit_map": [ + 0.10285980254411697, + 0.07607366144657135, + 0.03015858493745327, + 0.03958187252283096, + 0.04076511785387993, + 0.054778389632701874, + 0.02680385112762451, + 0.017400866374373436, + 0.018760064616799355, + 0.014899495989084244, + 0.008092035539448261, + 0.008670407347381115, + 0.008990267291665077, + 0.011143590323626995, + 0.007783113047480583, + 0.004560485482215881, + 0.01012512855231762, + 0.014668818563222885, + 0.005092841573059559, + 0.0044480240903794765, + 0.0060003153048455715, + 0.008647988550364971, + 0.004033282399177551, + 0.0035219949204474688, + 0.013325382955372334, + 0.017391854897141457, + 0.007343086414039135, + 0.006631155032664537, + 0.008691134862601757, + 0.011126628145575523, + 0.006943715736269951, + 0.005197434686124325, + 0.00805019773542881, + 0.014620988629758358, + 0.008434880524873734, + 0.0062871347181499004, + 0.0088456179946661, + 0.009352327324450016, + 0.007745003793388605, + 0.0038577881641685963, + 0.012871668674051762, + 0.011881318874657154, + 0.004667102824896574, + 0.007767526898533106, + 0.013631103560328484, + 0.01064250711351633, + 0.010300976224243641, + 0.004379692021757364, + 0.014018855057656765, + 0.02816835604608059, + 0.007808572147041559, + 0.008360602892935276, + 0.017368504777550697, + 0.016989631578326225, + 0.011597518809139729, + 0.006743399426341057, + 0.005280821118503809, + 0.0070383138954639435, + 0.003312227316200733, + 0.003280926961451769, + 0.00598945701494813, + 0.004709211178123951, + 0.004359721671789885, + 0.0019383493345230818, + 0.007688045501708984, + 0.013130764476954937, + 0.00815352238714695, + 0.0058090961538255215, + 0.007066686172038317, + 0.008868192322552204, + 0.006503747310489416, + 0.003496722783893347, + 0.0046613807789981365, + 0.0066771614365279675, + 0.003277784213423729, + 0.004250094760209322, + 0.009643279947340488, + 0.006058077793568373, + 0.007264839950948954, + 0.0026399109046906233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.007629185914993286 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864773899316788 + ], + "opposite_logprob_tokens": [ + -3.04864764213562 + ], + "logprob_reward": -0.04864773899316788, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864773899316788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24798907339572906, + 0.10424267500638962, + 0.09121192246675491, + 0.06904724985361099, + 0.09508736431598663, + 0.04488450288772583, + 0.08034913241863251, + 0.14447468519210815, + 0.09608223289251328, + 0.026631208136677742 + ], + "credit_map": [ + 0.05000654608011246, + 0.046285346150398254, + 0.017934387549757957, + 0.01457237359136343, + 0.052385855466127396, + 0.03560756519436836, + 0.021479710936546326, + 0.009717285633087158, + 0.021429838612675667, + 0.019589701667428017, + 0.011679486371576786, + 0.007271267473697662, + 0.01470643375068903, + 0.015012146905064583, + 0.009000802412629128, + 0.00555300060659647, + 0.01686127856373787, + 0.01978778839111328, + 0.008551862090826035, + 0.008962702937424183, + 0.009250815026462078, + 0.013923469930887222, + 0.00857581291347742, + 0.0052981898188591, + 0.00842213723808527, + 0.009823925793170929, + 0.004766313824802637, + 0.011869196780025959, + 0.010530904866755009, + 0.009960132651031017, + 0.009483127854764462, + 0.004191508051007986, + 0.025625377893447876, + 0.017440497875213623, + 0.008335759863257408, + 0.0060065509751439095, + 0.010768149048089981, + 0.013543354347348213, + 0.009414487518370152, + 0.003953186795115471, + 0.005327783524990082, + 0.009623570367693901, + 0.0039312029257416725, + 0.004223731812089682, + 0.0074170734733343124, + 0.006158869713544846, + 0.005180835723876953, + 0.003021436044946313, + 0.013104144483804703, + 0.015751341357827187, + 0.006428004242479801, + 0.006784094963222742, + 0.011143817566335201, + 0.010831262916326523, + 0.011775337159633636, + 0.004531129729002714, + 0.012234189547598362, + 0.0239181537181139, + 0.008545614778995514, + 0.011469921097159386, + 0.04428289830684662, + 0.021616214886307716, + 0.01698925718665123, + 0.005418437998741865, + 0.013861613348126411, + 0.0202533770352602, + 0.010795347392559052, + 0.009950852021574974, + 0.010208732448518276, + 0.014378424733877182, + 0.011382102966308594, + 0.005251778755337, + 0.003121925052255392, + 0.005152217578142881, + 0.0019798476714640856, + 0.0026440711226314306, + 0.005567208863794804, + 0.0037305557634681463, + 0.003121335990726948, + 0.0013140449300408363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.04864773899316788 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide shot, shot slightly from an upward angle, captures the Taj Mahal, a majestic white marble structure, standing alone against the fury of nature beneath a dark, storm-ravaged sky. Rolling, inky black storm clouds nearly bear down on the iconic central dome, while occasional flashes of pale lightning split the sky, briefly illuminating the building's rain-slicked facade. Below, raging floodwaters relentlessly devour its vast grounds, destroying the once tranquil reflecting pools and meticulously manicured gardens, replaced by a murky ocean of mud, sand, and debris. This tawny torrent crashes violently against the building's base, stirring up layers of white spray, the murky waters rising to the brink of the sacred white marble platform. The entire scene is presented in stunning 4K resolution, making every splashing drop of water and the dynamics of the distant clouds clearly visible, while being covered with a subtle 35mm film grain texture, adding a touch of tragic historical heaviness to this apocalyptic magnificent scene, creating an atmosphere of grandeur, oppression and tragic aesthetics.", + "seed": 11503, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_14/index_8630_seed_11503_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04919881336390972, + "vlm_logprob": -0.04919881336390972, + "vlm_yes_no_margin": 4.074999976158142, + "vlm_token_credit": -0.04919881336390972, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0034143980592489243 + ], + "opposite_logprob_tokens": [ + -7.003414630889893 + ], + "logprob_reward": -0.0034143980592489243, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0034143980592489243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2407151460647583, + 0.11158986389636993, + 0.10677400976419449, + 0.13583730161190033, + 0.10380285233259201, + 0.059882279485464096, + 0.04760264232754707, + 0.06709220260381699, + 0.06017006188631058, + 0.06653361022472382 + ], + "credit_map": [ + 0.03670484572649002, + 0.032731134444475174, + 0.03705530986189842, + 0.025898093357682228, + 0.022222938016057014, + 0.026165897026658058, + 0.02699810452759266, + 0.032938823103904724, + 0.014488217420876026, + 0.01555289700627327, + 0.021405981853604317, + 0.013948828913271427, + 0.009961359202861786, + 0.011017991229891777, + 0.012689403258264065, + 0.012525178492069244, + 0.015004552900791168, + 0.015447651036083698, + 0.011667844839394093, + 0.02028435841202736, + 0.008922592736780643, + 0.015987444669008255, + 0.008682467974722385, + 0.010777102783322334, + 0.017432821914553642, + 0.017073025926947594, + 0.0193167794495821, + 0.022221047431230545, + 0.013433946296572685, + 0.015459789894521236, + 0.013045921921730042, + 0.017853975296020508, + 0.016767825931310654, + 0.016064856201410294, + 0.01111418753862381, + 0.010929998010396957, + 0.01251960638910532, + 0.009882579557597637, + 0.008114385418593884, + 0.01840941235423088, + 0.007424946408718824, + 0.006829388439655304, + 0.0056069232523441315, + 0.010051867924630642, + 0.007664994336664677, + 0.008431595750153065, + 0.0047869496047496796, + 0.009085613302886486, + 0.008410179987549782, + 0.0069361962378025055, + 0.004346788395196199, + 0.00408879155293107, + 0.005219565238803625, + 0.006319303996860981, + 0.004036907106637955, + 0.008244909346103668, + 0.008632817305624485, + 0.010644822381436825, + 0.005905562546104193, + 0.012951267883181572, + 0.007932115346193314, + 0.007386317010968924, + 0.00756058469414711, + 0.006078718230128288, + 0.012608475983142853, + 0.008501075208187103, + 0.006270708981901407, + 0.004765151534229517, + 0.005504020489752293, + 0.008554822765290737, + 0.00630001537501812, + 0.007665792480111122, + 0.0152894277125597, + 0.009124667383730412, + 0.00826950091868639, + 0.005381409078836441, + 0.005580240860581398, + 0.010583829134702682, + 0.004019436426460743, + 0.008285097777843475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0034143980592489243 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956517532467842 + ], + "opposite_logprob_tokens": [ + -5.130956649780273 + ], + "logprob_reward": -0.005956517532467842, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956517532467842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3301800489425659, + 0.13533319532871246, + 0.11093217879533768, + 0.03840680792927742, + 0.09367790818214417, + 0.031778041273355484, + 0.10839758068323135, + 0.06256704032421112, + 0.047004517167806625, + 0.04172271117568016 + ], + "credit_map": [ + 0.045686785131692886, + 0.03666982799768448, + 0.07108446210622787, + 0.03963806852698326, + 0.026925966143608093, + 0.03514997288584709, + 0.0435476079583168, + 0.03147735819220543, + 0.01912769116461277, + 0.020275810733437538, + 0.014024436473846436, + 0.026391547173261642, + 0.014415882527828217, + 0.010295463725924492, + 0.012996699661016464, + 0.0178056713193655, + 0.02760685607790947, + 0.021776529029011726, + 0.012415777891874313, + 0.008952159434556961, + 0.007614273577928543, + 0.011407794430851936, + 0.007344197481870651, + 0.013814589008688927, + 0.004945808555930853, + 0.004704907536506653, + 0.004043485037982464, + 0.009899212047457695, + 0.00299127702601254, + 0.00441764248535037, + 0.0032426677644252777, + 0.004161804914474487, + 0.015779342502355576, + 0.01073495950549841, + 0.010436521843075752, + 0.016185929998755455, + 0.008710937574505806, + 0.01018483005464077, + 0.009861706756055355, + 0.011783680878579617, + 0.004552184138447046, + 0.003069506958127022, + 0.002841055393218994, + 0.003529532812535763, + 0.003809312591329217, + 0.005013138521462679, + 0.0024733596947044134, + 0.006489948835223913, + 0.02284369245171547, + 0.012450779788196087, + 0.010900292545557022, + 0.013173035345971584, + 0.009982209652662277, + 0.012377322651445866, + 0.007965517230331898, + 0.01870473101735115, + 0.012346960604190826, + 0.005394998472183943, + 0.0066997576504945755, + 0.0056896754540503025, + 0.005586057435721159, + 0.00853880774229765, + 0.005614812020212412, + 0.01269596442580223, + 0.0076776957139372826, + 0.0050327349454164505, + 0.0047803716734051704, + 0.004773174878209829, + 0.004644347354769707, + 0.006258251611143351, + 0.006193621549755335, + 0.0076443227007985115, + 0.00963468849658966, + 0.0037369218189269304, + 0.004914115183055401, + 0.003630797378718853, + 0.0023265485651791096, + 0.006613899953663349, + 0.0029814818408340216, + 0.007884258404374123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.005956517532467842 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270352452993393 + ], + "opposite_logprob_tokens": [ + -2.127035140991211 + ], + "logprob_reward": -0.1270352452993393, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270352452993393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19187186658382416, + 0.07344706356525421, + 0.062171973288059235, + 0.05710359290242195, + 0.11448905616998672, + 0.015156182460486889, + 0.04219610616564751, + 0.2731208801269531, + 0.04282677546143532, + 0.12761645019054413 + ], + "credit_map": [ + 0.025452112779021263, + 0.01871967501938343, + 0.0377451553940773, + 0.032352153211832047, + 0.01509359572082758, + 0.02079855278134346, + 0.01886829361319542, + 0.022842327132821083, + 0.008906939998269081, + 0.008682153187692165, + 0.009314618073403835, + 0.010546989738941193, + 0.005990429315716028, + 0.0076436069793999195, + 0.009856805205345154, + 0.01250552199780941, + 0.013386139646172523, + 0.010007367469370365, + 0.006220829673111439, + 0.006437964271754026, + 0.005018667317926884, + 0.005729234777390957, + 0.0051502930000424385, + 0.010221477597951889, + 0.009639984928071499, + 0.006419236771762371, + 0.009562154300510883, + 0.006050807423889637, + 0.004849512130022049, + 0.007382337469607592, + 0.005480685271322727, + 0.0077188764698803425, + 0.02268218994140625, + 0.014642658643424511, + 0.012439608573913574, + 0.010295195505023003, + 0.009780172258615494, + 0.01467837579548359, + 0.007931656204164028, + 0.022039204835891724, + 0.0030074873939156532, + 0.0015919165452942252, + 0.001957401866093278, + 0.001434618723578751, + 0.0011878968216478825, + 0.0018053038511425257, + 0.001214892021380365, + 0.0029566658195108175, + 0.011505874805152416, + 0.0035928217694163322, + 0.00465820636600256, + 0.004044492729008198, + 0.003132269252091646, + 0.0057640643790364265, + 0.002731510205194354, + 0.006766866892576218, + 0.05259505286812782, + 0.025424394756555557, + 0.03552377223968506, + 0.04609992727637291, + 0.014870692975819111, + 0.03258432820439339, + 0.020365064963698387, + 0.04565765708684921, + 0.007917261682450771, + 0.005847077816724777, + 0.004259536042809486, + 0.0032046912238001823, + 0.0031593888998031616, + 0.005995715036988258, + 0.004163945093750954, + 0.008279156871140003, + 0.033398136496543884, + 0.010717209428548813, + 0.014670486561954021, + 0.009130322374403477, + 0.006368751171976328, + 0.021978000178933144, + 0.010945412330329418, + 0.020408136770129204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.1270352452993393 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020576467737555504 + ], + "opposite_logprob_tokens": [ + -3.8955764770507812 + ], + "logprob_reward": -0.020576467737555504, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020576467737555504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1413666158914566, + 0.06354397535324097, + 0.118624746799469, + 0.29330945014953613, + 0.08230622112751007, + 0.05558326095342636, + 0.023964891210198402, + 0.1014556735754013, + 0.05581481009721756, + 0.06403037905693054 + ], + "credit_map": [ + 0.03411794826388359, + 0.02033168263733387, + 0.019663814455270767, + 0.012601304799318314, + 0.009427100419998169, + 0.02057286910712719, + 0.007709739729762077, + 0.01694214530289173, + 0.009586605243384838, + 0.007777354214340448, + 0.006858187261968851, + 0.007207508198916912, + 0.007655738852918148, + 0.007480482570827007, + 0.00741969421505928, + 0.009558405727148056, + 0.019406862556934357, + 0.01936570554971695, + 0.01735876128077507, + 0.018179800361394882, + 0.010146631859242916, + 0.010211457498371601, + 0.010535256937146187, + 0.013420271687209606, + 0.052859898656606674, + 0.046914439648389816, + 0.03872314840555191, + 0.02555069327354431, + 0.02695618011057377, + 0.04766272380948067, + 0.022348783910274506, + 0.03229357302188873, + 0.019764116033911705, + 0.009329322725534439, + 0.007899262011051178, + 0.008606014773249626, + 0.006357668433338404, + 0.009292586706578732, + 0.00767425587400794, + 0.013382994569838047, + 0.009677850641310215, + 0.005194985773414373, + 0.0052496809512376785, + 0.008824067190289497, + 0.003958777524530888, + 0.007839048281311989, + 0.005721989553421736, + 0.009116863831877708, + 0.005266665015369654, + 0.002178874099627137, + 0.0026016393676400185, + 0.002041005762293935, + 0.0024474698584526777, + 0.0032481281086802483, + 0.002069578506052494, + 0.0041115302592515945, + 0.022477276623249054, + 0.009459911845624447, + 0.01235812995582819, + 0.01630116067826748, + 0.005822885315865278, + 0.013317982666194439, + 0.0071312119252979755, + 0.014587107114493847, + 0.014141750521957874, + 0.005742791574448347, + 0.005214191973209381, + 0.0036973946262151003, + 0.003506326349452138, + 0.009196843951940536, + 0.004188897553831339, + 0.010126616805791855, + 0.016285596415400505, + 0.0066383956000208855, + 0.007141149137169123, + 0.004973556846380234, + 0.0033007387537509203, + 0.010419412516057491, + 0.005463640671223402, + 0.009807889349758625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.020576467737555504 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901143819093704 + ], + "opposite_logprob_tokens": [ + -2.4640114307403564 + ], + "logprob_reward": -0.08901143819093704, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901143819093704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2086227834224701, + 0.161199152469635, + 0.05580626055598259, + 0.08908531069755554, + 0.07598087936639786, + 0.055407654494047165, + 0.09516299515962601, + 0.09998390078544617, + 0.1150997132062912, + 0.043651316314935684 + ], + "credit_map": [ + 0.03376375138759613, + 0.018372738733887672, + 0.044627074152231216, + 0.030642861500382423, + 0.015226348303258419, + 0.02445860579609871, + 0.020005380734801292, + 0.021526027470827103, + 0.0253238994628191, + 0.022508593276143074, + 0.02028265967965126, + 0.019398855045437813, + 0.013565987348556519, + 0.016022425144910812, + 0.018397552892565727, + 0.025699175894260406, + 0.0074770827777683735, + 0.006750460714101791, + 0.007473452016711235, + 0.008314288221299648, + 0.004061476793140173, + 0.00481786485761404, + 0.006195183377712965, + 0.010716455057263374, + 0.012928660959005356, + 0.014760816469788551, + 0.009363595396280289, + 0.01720483973622322, + 0.00838237814605236, + 0.012689570896327496, + 0.006437181960791349, + 0.007318271789699793, + 0.013995532877743244, + 0.00979323498904705, + 0.008314530365169048, + 0.007018505595624447, + 0.010102417320013046, + 0.008723475970327854, + 0.006862891372293234, + 0.011170292273163795, + 0.009253013879060745, + 0.006900377571582794, + 0.004981428384780884, + 0.010035594925284386, + 0.007246813736855984, + 0.005915939342230558, + 0.004886424168944359, + 0.006188062485307455, + 0.02023654617369175, + 0.012770289555191994, + 0.008599416352808475, + 0.01287281233817339, + 0.008189596235752106, + 0.011452165432274342, + 0.0066729262471199036, + 0.014369238168001175, + 0.01954546757042408, + 0.008286133408546448, + 0.0144933657720685, + 0.013861368410289288, + 0.0062982006929814816, + 0.01252825278788805, + 0.007309186737984419, + 0.0176619254052639, + 0.021274644881486893, + 0.012176685966551304, + 0.012461029924452305, + 0.010117682628333569, + 0.008597305044531822, + 0.01646394096314907, + 0.008970643393695354, + 0.02503778226673603, + 0.009192381985485554, + 0.007362483534961939, + 0.004366235341876745, + 0.0036311415024101734, + 0.004160364158451557, + 0.00572773115709424, + 0.003499778686091304, + 0.005711199715733528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.08901143819093704 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide shot, shot slightly from an upward angle, captures the Taj Mahal, a majestic white marble structure, standing alone against the fury of nature beneath a dark, storm-ravaged sky. Rolling, inky black storm clouds nearly bear down on the iconic central dome, while occasional flashes of pale lightning split the sky, briefly illuminating the building's rain-slicked facade. Below, raging floodwaters relentlessly devour its vast grounds, destroying the once tranquil reflecting pools and meticulously manicured gardens, replaced by a murky ocean of mud, sand, and debris. This tawny torrent crashes violently against the building's base, stirring up layers of white spray, the murky waters rising to the brink of the sacred white marble platform. The entire scene is presented in stunning 4K resolution, making every splashing drop of water and the dynamics of the distant clouds clearly visible, while being covered with a subtle 35mm film grain texture, adding a touch of tragic historical heaviness to this apocalyptic magnificent scene, creating an atmosphere of grandeur, oppression and tragic aesthetics.", + "seed": 11504, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_14/index_8630_seed_11504_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08115227902308106, + "vlm_logprob": -0.08115227902308106, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.08115227902308106, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031304187141358852 + ], + "opposite_logprob_tokens": [ + -6.2531304359436035 + ], + "logprob_reward": -0.0031304187141358852, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0031304187141358852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1744949221611023, + 0.058267176151275635, + 0.16544342041015625, + 0.07837924361228943, + 0.08350313454866409, + 0.0325859971344471, + 0.0677131861448288, + 0.04912044107913971, + 0.1632443219423294, + 0.12724824249744415 + ], + "credit_map": [ + 0.025220252573490143, + 0.030997155234217644, + 0.0164498258382082, + 0.016593987122178078, + 0.023503802716732025, + 0.023516621440649033, + 0.022594671696424484, + 0.015618585981428623, + 0.009124851785600185, + 0.008259407244622707, + 0.006026365328580141, + 0.0039054129738360643, + 0.00901245977729559, + 0.008156348951160908, + 0.005384585354477167, + 0.008397737517952919, + 0.021669356152415276, + 0.0238081943243742, + 0.014576541259884834, + 0.017528003081679344, + 0.01947660930454731, + 0.023424219340085983, + 0.020783808082342148, + 0.02417667768895626, + 0.00961033534258604, + 0.01052805595099926, + 0.006903002969920635, + 0.0077062747441232204, + 0.00994293112307787, + 0.011436235159635544, + 0.009101171046495438, + 0.013151232153177261, + 0.010237488895654678, + 0.012224038131535053, + 0.006906620226800442, + 0.012865335680544376, + 0.008783707395195961, + 0.011983986012637615, + 0.010668966919183731, + 0.009832983836531639, + 0.004537189844995737, + 0.004532031249254942, + 0.0025524236261844635, + 0.0039603495970368385, + 0.005559109151363373, + 0.0049553923308849335, + 0.0036915757227689028, + 0.0027979244478046894, + 0.007418670225888491, + 0.009224562905728817, + 0.005207166075706482, + 0.008916755206882954, + 0.01008223369717598, + 0.009062354452908039, + 0.010836823843419552, + 0.006964615546166897, + 0.004251228179782629, + 0.0058114659041166306, + 0.003934328909963369, + 0.0060254489071667194, + 0.008780673146247864, + 0.005718783475458622, + 0.011518042534589767, + 0.003080465365201235, + 0.026786522939801216, + 0.028752554208040237, + 0.01205812394618988, + 0.019242020323872566, + 0.02337854914367199, + 0.026834074407815933, + 0.014579921960830688, + 0.011612539179623127, + 0.022082755342125893, + 0.026165487244725227, + 0.008011279627680779, + 0.018039725720882416, + 0.014034357853233814, + 0.0214861948043108, + 0.008987106382846832, + 0.008441333658993244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0031304187141358852 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648204617202282 + ], + "opposite_logprob_tokens": [ + -4.75864839553833 + ], + "logprob_reward": -0.008648204617202282, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648204617202282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18980710208415985, + 0.07702748477458954, + 0.1844431757926941, + 0.1428782045841217, + 0.059287313371896744, + 0.07215344160795212, + 0.0317360982298851, + 0.10718394815921783, + 0.056027092039585114, + 0.07945618033409119 + ], + "credit_map": [ + 0.03531622886657715, + 0.03157177194952965, + 0.018917616456747055, + 0.014505614526569843, + 0.025377562269568443, + 0.029604986310005188, + 0.016633279621601105, + 0.017880037426948547, + 0.013719473034143448, + 0.013082017190754414, + 0.007129183039069176, + 0.0076574659906327724, + 0.010874459519982338, + 0.01248825341463089, + 0.006131590809673071, + 0.005945038050413132, + 0.01895410381257534, + 0.021615901961922646, + 0.019783416762948036, + 0.020856274291872978, + 0.0236094668507576, + 0.021692929789423943, + 0.021691936999559402, + 0.036239143460989, + 0.01838141493499279, + 0.022605538368225098, + 0.009397136978805065, + 0.016972096636891365, + 0.018642425537109375, + 0.02254432998597622, + 0.020504597574472427, + 0.013830670155584812, + 0.008797281421720982, + 0.007586382329463959, + 0.006026731338351965, + 0.006660144310444593, + 0.009069131687283516, + 0.008495966903865337, + 0.006862418260425329, + 0.005789258051663637, + 0.012418393976986408, + 0.009797484613955021, + 0.005286879371851683, + 0.008513818494975567, + 0.011269737966358662, + 0.01094108633697033, + 0.008793486282229424, + 0.0051325587555766106, + 0.0035402958746999502, + 0.004114283714443445, + 0.004915724508464336, + 0.004136857111006975, + 0.003775152610614896, + 0.004611041862517595, + 0.002793466206640005, + 0.00384927773848176, + 0.01162345614284277, + 0.011948577128350735, + 0.011599973775446415, + 0.01354698371142149, + 0.012363589368760586, + 0.015161044895648956, + 0.02107832580804825, + 0.009861993603408337, + 0.00849262997508049, + 0.012115237303078175, + 0.004035978112369776, + 0.007507070899009705, + 0.00730107119306922, + 0.008889039978384972, + 0.004314280580729246, + 0.0033717830665409565, + 0.009929977357387543, + 0.016398554667830467, + 0.005780684296041727, + 0.010028313845396042, + 0.010564561001956463, + 0.010132858529686928, + 0.009023407474160194, + 0.007597822695970535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.008648204617202282 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156468451023102 + ], + "opposite_logprob_tokens": [ + -1.7015646696090698 + ], + "logprob_reward": -0.20156468451023102, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156468451023102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20557795464992523, + 0.09394756704568863, + 0.11031274497509003, + 0.04457481950521469, + 0.1201583668589592, + 0.05903375521302223, + 0.028776906430721283, + 0.07843343168497086, + 0.07169882208108902, + 0.18748565018177032 + ], + "credit_map": [ + 0.0369357131421566, + 0.03728724643588066, + 0.017530620098114014, + 0.019874148070812225, + 0.022828051820397377, + 0.03199753910303116, + 0.019650774076581, + 0.019473880529403687, + 0.011751706711947918, + 0.014512913301587105, + 0.00950423814356327, + 0.010099696926772594, + 0.011991403065621853, + 0.012141648679971695, + 0.011787054128944874, + 0.012158908881247044, + 0.018102189525961876, + 0.019842857494950294, + 0.007572303991764784, + 0.010739138349890709, + 0.012025218456983566, + 0.018718572333455086, + 0.014546727761626244, + 0.008765739388763905, + 0.005696791224181652, + 0.008267560973763466, + 0.0039147622883319855, + 0.00483597069978714, + 0.005571141839027405, + 0.008110436610877514, + 0.00441298121586442, + 0.003765175584703684, + 0.014854020439088345, + 0.01852903887629509, + 0.011245153844356537, + 0.017372898757457733, + 0.013527322560548782, + 0.01754874177277088, + 0.012930567376315594, + 0.014150623232126236, + 0.009325170889496803, + 0.010050684213638306, + 0.0032718339934945107, + 0.007027043029665947, + 0.007152695208787918, + 0.011386093683540821, + 0.006227294914424419, + 0.004592940676957369, + 0.0039914269000291824, + 0.003587177488952875, + 0.0040652137249708176, + 0.0028092123102396727, + 0.003764803521335125, + 0.004525922704488039, + 0.0033772264141589403, + 0.0026559226680547, + 0.009598156437277794, + 0.008070169948041439, + 0.008436454460024834, + 0.009748866781592369, + 0.008723491802811623, + 0.009297896176576614, + 0.014876202680170536, + 0.009682192467153072, + 0.009482043795287609, + 0.013399492017924786, + 0.007440593559294939, + 0.009564777836203575, + 0.010403269901871681, + 0.008894088678061962, + 0.006573710590600967, + 0.005940841510891914, + 0.028294619172811508, + 0.03460853546857834, + 0.016976168379187584, + 0.029685521498322487, + 0.014804953709244728, + 0.028998572379350662, + 0.01648077182471752, + 0.017636507749557495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.20156468451023102 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254018023610115 + ], + "opposite_logprob_tokens": [ + -4.387540340423584 + ], + "logprob_reward": -0.01254018023610115, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254018023610115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17120371758937836, + 0.05555227771401405, + 0.2161882370710373, + 0.1632073074579239, + 0.0573890320956707, + 0.043031126260757446, + 0.08438926190137863, + 0.030442247167229652, + 0.12032389640808105, + 0.05827290564775467 + ], + "credit_map": [ + 0.040570490062236786, + 0.026708802208304405, + 0.011453386396169662, + 0.015713432803750038, + 0.02149898186326027, + 0.034361258149147034, + 0.010760973207652569, + 0.010136380791664124, + 0.008841048926115036, + 0.008931358344852924, + 0.0043559898622334, + 0.005151818972080946, + 0.008135627023875713, + 0.009234143421053886, + 0.004970105830579996, + 0.005932180676609278, + 0.023542720824480057, + 0.030709052458405495, + 0.021113185212016106, + 0.024011213332414627, + 0.030260303989052773, + 0.026098037138581276, + 0.026044677942991257, + 0.03440902754664421, + 0.023144764825701714, + 0.025975042954087257, + 0.012172588147222996, + 0.01879328116774559, + 0.017805952578783035, + 0.02731684409081936, + 0.021311648190021515, + 0.0166871827095747, + 0.009788884781301022, + 0.010330173186957836, + 0.004454596433788538, + 0.007183642592281103, + 0.0051530515775084496, + 0.01090946700423956, + 0.004299890249967575, + 0.005269323475658894, + 0.007239985279738903, + 0.007668375037610531, + 0.002818724140524864, + 0.0051471274346113205, + 0.00468888645991683, + 0.008411875925958157, + 0.003901759861037135, + 0.0031543890945613384, + 0.010664918459951878, + 0.009998002089560032, + 0.011295493692159653, + 0.009266464039683342, + 0.011191295459866524, + 0.011899486184120178, + 0.011694563552737236, + 0.008379040285944939, + 0.004628752823919058, + 0.005200800020247698, + 0.002734615234658122, + 0.0035580683033913374, + 0.004426837898790836, + 0.0044525181874632835, + 0.002710936591029167, + 0.002729717642068863, + 0.012169570662081242, + 0.015953151509165764, + 0.015511473640799522, + 0.016507431864738464, + 0.01902494765818119, + 0.015394158661365509, + 0.014024870470166206, + 0.011738279834389687, + 0.007192431017756462, + 0.011136720888316631, + 0.006650989409536123, + 0.0066596707329154015, + 0.007794296834617853, + 0.006459185853600502, + 0.006699084769934416, + 0.005680522881448269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.01254018023610115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Taj Mahal visible as a white marble structure in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rolling inky black storm clouds and flashes of pale lightning visible in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground covered by raging floodwaters, mud, sand, and debris destroying the gardens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a wide angle with a slight upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video display a subtle 35mm film grain texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987790703773499 + ], + "opposite_logprob_tokens": [ + -1.8048778772354126 + ], + "logprob_reward": -0.17987790703773499, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987790703773499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11614548414945602, + 0.13304194808006287, + 0.0951896458864212, + 0.1332845687866211, + 0.06299251317977905, + 0.030250508338212967, + 0.12381492555141449, + 0.13843217492103577, + 0.07630261778831482, + 0.09054560959339142 + ], + "credit_map": [ + 0.020231002941727638, + 0.01448530238121748, + 0.01363286655396223, + 0.01065459568053484, + 0.015415788628160954, + 0.02171565406024456, + 0.010129675269126892, + 0.009880605153739452, + 0.019908148795366287, + 0.022920364513993263, + 0.01018431130796671, + 0.018660614266991615, + 0.01314143743366003, + 0.019511179998517036, + 0.015005533583462238, + 0.01371034700423479, + 0.01769389398396015, + 0.018368873745203018, + 0.008004034869372845, + 0.00956460926681757, + 0.009408053010702133, + 0.017961688339710236, + 0.0066930195316672325, + 0.007495475001633167, + 0.01804024539887905, + 0.020659783855080605, + 0.010565221309661865, + 0.013145889155566692, + 0.018120771273970604, + 0.020865844562649727, + 0.013017629273235798, + 0.018869176506996155, + 0.009330542758107185, + 0.00728609599173069, + 0.008003370836377144, + 0.0053384010680019855, + 0.009896606206893921, + 0.009672669693827629, + 0.007468370720744133, + 0.005996459163725376, + 0.004376552999019623, + 0.003895936533808708, + 0.004086181055754423, + 0.002491631079465151, + 0.004437354393303394, + 0.00510838208720088, + 0.0029584509320557117, + 0.002896021818742156, + 0.013505522161722183, + 0.011768882162868977, + 0.01883082278072834, + 0.013023100793361664, + 0.017287757247686386, + 0.014798643067479134, + 0.019256075844168663, + 0.015344117768108845, + 0.014475837349891663, + 0.016467005014419556, + 0.012517671100795269, + 0.017392734065651894, + 0.016930457204580307, + 0.01987077109515667, + 0.027185453101992607, + 0.013592236675322056, + 0.01442576665431261, + 0.013195237144827843, + 0.006483651697635651, + 0.008808999322354794, + 0.007925479672849178, + 0.012363038957118988, + 0.007206434849649668, + 0.005894006229937077, + 0.01166106853634119, + 0.012752681039273739, + 0.00909682922065258, + 0.015199472196400166, + 0.008078892715275288, + 0.012626653537154198, + 0.011083461344242096, + 0.010046550072729588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.17987790703773499 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_15/index_tensor([2985, 2985])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_15/index_tensor([2985, 2985])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6395f3a9174e520ac1c05e0021ba2946dbbb7f07 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_15/index_tensor([2985, 2985])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a level angle, slowly pulls back to reveal a wider scene. Deep within a lush bamboo forest, shrouded in morning mist, a milky white mist gently drifts between the towering, straight bamboo poles. The first rays of morning sunlight struggle to penetrate the dense fog and dense bamboo leaves, casting several distinct, mottled golden beams in the air. An elderly monk with a gaunt face and a shiny head, dressed in a simple dark gray cotton and linen robe, stands in a clearing in the forest, covered with moist fallen leaves. His eyes are slightly closed, his expression serene and focused. His body moves in a steady and deliberate rhythm. His arms trace smooth arcs, his wide sleeves swaying gently, and his center of gravity shifts effortlessly between his legs. Each movement, from raising his hands to lowering his shoulders, is as smooth and natural as flowing water, imbued with inner strength and harmony. The entire picture is dominated by calm gray and vibrant emerald green, embellished by warm morning light, creating an otherworldly, Zen-like tranquil atmosphere.", + "seed": 3310, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_15/index_2985_seed_3310_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03575662346556783, + "vlm_logprob": -0.03575662346556783, + "vlm_yes_no_margin": 4.725000023841858, + "vlm_token_credit": -0.03575662346556783, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009194768033921719 + ], + "opposite_logprob_tokens": [ + -7.509194850921631 + ], + "logprob_reward": -0.009194768033921719, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.009194768033921719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05735703930258751, + 0.22971481084823608, + 0.05507032945752144, + 0.049794647842645645, + 0.051398396492004395, + 0.023577747866511345, + 0.016490649431943893, + 0.03542037680745125, + 0.33324119448661804, + 0.14793485403060913 + ], + "credit_map": [ + 0.006170366425067186, + 0.005016066133975983, + 0.0065592192113399506, + 0.010544832795858383, + 0.006295498460531235, + 0.006433241069316864, + 0.011000387370586395, + 0.005337429232895374, + 0.015205875970423222, + 0.025818977504968643, + 0.02131631039083004, + 0.019251076504588127, + 0.024535629898309708, + 0.04903801903128624, + 0.020902691408991814, + 0.053646259009838104, + 0.0082172192633152, + 0.009510286152362823, + 0.005557773169130087, + 0.005344650708138943, + 0.005611714906990528, + 0.008145039901137352, + 0.0056831613183021545, + 0.007000488694757223, + 0.005827245302498341, + 0.003485925728455186, + 0.005096879787743092, + 0.003564574057236314, + 0.004521877970546484, + 0.01120708230882883, + 0.004398977383971214, + 0.011692093685269356, + 0.005047896411269903, + 0.004729398060590029, + 0.00731143681332469, + 0.01113918237388134, + 0.00565247843042016, + 0.00589003786444664, + 0.005742840468883514, + 0.005885133519768715, + 0.002648563589900732, + 0.002282156376168132, + 0.002770873950794339, + 0.0030408590100705624, + 0.003127771895378828, + 0.003197236219421029, + 0.0039549656212329865, + 0.002555321203544736, + 0.0019720059353858232, + 0.0017028942238539457, + 0.0025524941738694906, + 0.00231820996850729, + 0.0013637925731018186, + 0.0018907473422586918, + 0.002986692124977708, + 0.0017038143705576658, + 0.0033057613763958216, + 0.0046983701176941395, + 0.006892204284667969, + 0.0050724586471915245, + 0.003611027030274272, + 0.00432064663618803, + 0.002820737659931183, + 0.004699174780398607, + 0.014343293383717537, + 0.011551203206181526, + 0.09928601235151291, + 0.02256562002003193, + 0.014730866067111492, + 0.1135324239730835, + 0.010806509293615818, + 0.04642527177929878, + 0.00807784404605627, + 0.005240149330347776, + 0.004776456393301487, + 0.006691297981888056, + 0.0058711860328912735, + 0.055537138134241104, + 0.006984253413975239, + 0.0547565221786499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.009194768033921719 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005963864270597696 + ], + "opposite_logprob_tokens": [ + -5.1309638023376465 + ], + "logprob_reward": -0.005963864270597696, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005963864270597696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10324570536613464, + 0.050397712737321854, + 0.08171293884515762, + 0.03681747987866402, + 0.03178412467241287, + 0.0731881707906723, + 0.05878184735774994, + 0.1843574047088623, + 0.20884785056114197, + 0.17086675763130188 + ], + "credit_map": [ + 0.009492423385381699, + 0.012537415139377117, + 0.015927676111459732, + 0.013952554203569889, + 0.014718367718160152, + 0.009697719477117062, + 0.01715301349759102, + 0.009766533970832825, + 0.004351985640823841, + 0.004057273268699646, + 0.004800543654710054, + 0.0045488676987588406, + 0.006384763401001692, + 0.010243060998618603, + 0.00693894037976861, + 0.009072279557585716, + 0.012903175316751003, + 0.012735391966998577, + 0.007164786569774151, + 0.010257620364427567, + 0.010697285644710064, + 0.008996753953397274, + 0.011121837422251701, + 0.007836087606847286, + 0.003965757321566343, + 0.0062931496649980545, + 0.0051692090928554535, + 0.004308596253395081, + 0.004887830466032028, + 0.003845313796773553, + 0.004454013425856829, + 0.0038936096243560314, + 0.0038465927354991436, + 0.003493440570309758, + 0.002846671035513282, + 0.004358460661023855, + 0.0037705206777900457, + 0.006173575762659311, + 0.0036822401452809572, + 0.0036126228515058756, + 0.007942174561321735, + 0.006643523927778006, + 0.009283323772251606, + 0.010387550108134747, + 0.008016800507903099, + 0.009100601077079773, + 0.013482303358614445, + 0.00833189394325018, + 0.00457530003041029, + 0.005218129605054855, + 0.006990958005189896, + 0.010036377236247063, + 0.006032975856214762, + 0.00534629262983799, + 0.014850831590592861, + 0.005730981472879648, + 0.01476762443780899, + 0.023174431174993515, + 0.02108316868543625, + 0.043456342071294785, + 0.017655031755566597, + 0.020276280120015144, + 0.02313629910349846, + 0.020808221772313118, + 0.009730340912938118, + 0.008126038126647472, + 0.06558482348918915, + 0.016086816787719727, + 0.008723974227905273, + 0.06557866930961609, + 0.009495193138718605, + 0.025522008538246155, + 0.01027499046176672, + 0.009283989667892456, + 0.006338002625852823, + 0.006071146111935377, + 0.008742386475205421, + 0.060342513024806976, + 0.007581789046525955, + 0.062231943011283875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.005963864270597696 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782706387341022 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.009782706387341022, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782706387341022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07955371588468552, + 0.11461511254310608, + 0.08607009798288345, + 0.10883104056119919, + 0.12366796284914017, + 0.08151200413703918, + 0.08255201578140259, + 0.1095016822218895, + 0.07936706393957138, + 0.13432924449443817 + ], + "credit_map": [ + 0.007967771962285042, + 0.008182107470929623, + 0.010035342536866665, + 0.008525313809514046, + 0.013471487909555435, + 0.009146464057266712, + 0.014400222338736057, + 0.00782500859349966, + 0.006866801530122757, + 0.004927508067339659, + 0.006381483748555183, + 0.007070211227983236, + 0.006451353896409273, + 0.03666464611887932, + 0.008187150582671165, + 0.038065966218709946, + 0.010176840238273144, + 0.00894604716449976, + 0.008264733478426933, + 0.013660996221005917, + 0.00801093876361847, + 0.009781798347830772, + 0.01666407473385334, + 0.010564669966697693, + 0.013299905695021152, + 0.019253982231020927, + 0.011839743703603745, + 0.012439921498298645, + 0.01392113696783781, + 0.012649028562009335, + 0.012500563636422157, + 0.012926755473017693, + 0.01614757999777794, + 0.009849944151937962, + 0.02668745629489422, + 0.022306304425001144, + 0.008152257651090622, + 0.013771279715001583, + 0.012279105372726917, + 0.014474027790129185, + 0.009166516363620758, + 0.007562275510281324, + 0.008896893821656704, + 0.012318101711571217, + 0.010243155062198639, + 0.010596665553748608, + 0.0139708137139678, + 0.008757579140365124, + 0.006043950095772743, + 0.009760712273418903, + 0.006798362825065851, + 0.01232975535094738, + 0.009366409853100777, + 0.00845223106443882, + 0.020067129284143448, + 0.00973346084356308, + 0.009563772939145565, + 0.010723147541284561, + 0.019462188705801964, + 0.02491927333176136, + 0.009117182344198227, + 0.010123319923877716, + 0.011542795225977898, + 0.014049999415874481, + 0.005021069664508104, + 0.0046698302030563354, + 0.008966200053691864, + 0.011647327803075314, + 0.005231884773820639, + 0.015454123727977276, + 0.009889673441648483, + 0.018486952409148216, + 0.007287135347723961, + 0.004754958674311638, + 0.006613139063119888, + 0.006493555381894112, + 0.005149046424776316, + 0.04932166635990143, + 0.005798955447971821, + 0.04891078919172287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.009782706387341022 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107780635356903 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.01107780635356903, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107780635356903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13043253123760223, + 0.11662059277296066, + 0.027826854959130287, + 0.06279502063989639, + 0.033166248351335526, + 0.10607308149337769, + 0.2975933253765106, + 0.05913589149713516, + 0.1318223923444748, + 0.03453409671783447 + ], + "credit_map": [ + 0.01557057537138462, + 0.014178012497723103, + 0.021837815642356873, + 0.01333968061953783, + 0.016214732080698013, + 0.01711803674697876, + 0.01917938143014908, + 0.012994284741580486, + 0.010708225890994072, + 0.01507636345922947, + 0.01595432125031948, + 0.010974776931107044, + 0.013197732158005238, + 0.019419051706790924, + 0.012294023297727108, + 0.01899610087275505, + 0.004743237979710102, + 0.0031495126895606518, + 0.0026453786995261908, + 0.003636614652350545, + 0.0032756414730101824, + 0.0032458812929689884, + 0.003877438372001052, + 0.003253149800002575, + 0.007058536633849144, + 0.009048578329384327, + 0.0070042661391198635, + 0.006679384503513575, + 0.008117197081446648, + 0.010611294768750668, + 0.007759189698845148, + 0.006516576744616032, + 0.0038177999667823315, + 0.0031733957584947348, + 0.007390064653009176, + 0.005177841521799564, + 0.003125180024653673, + 0.0030786555726081133, + 0.0039317370392382145, + 0.0034715700894594193, + 0.013656492345035076, + 0.014886940829455853, + 0.01717347465455532, + 0.012796405702829361, + 0.009458301588892937, + 0.012649443931877613, + 0.014273285865783691, + 0.011178732849657536, + 0.030790507793426514, + 0.03281886875629425, + 0.06494337320327759, + 0.029843144118785858, + 0.03135620057582855, + 0.038282327353954315, + 0.028072664514183998, + 0.041486237198114395, + 0.005062370095402002, + 0.007288314867764711, + 0.006041249260306358, + 0.013070980086922646, + 0.006014422047883272, + 0.008079402148723602, + 0.0066619617864489555, + 0.0069171893410384655, + 0.006845003459602594, + 0.005497049540281296, + 0.00943195354193449, + 0.0055512478575110435, + 0.004150786437094212, + 0.05296704173088074, + 0.005361683666706085, + 0.0420176237821579, + 0.002559204353019595, + 0.0018889028578996658, + 0.0019512730650603771, + 0.0026079530362039804, + 0.002234849613159895, + 0.009138801135122776, + 0.0020689736120402813, + 0.012084138579666615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.01107780635356903 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14276397228240967 + ], + "opposite_logprob_tokens": [ + -2.017764091491699 + ], + "logprob_reward": -0.14276397228240967, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14276397228240967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033716656267642975, + 0.10857751220464706, + 0.04405006766319275, + 0.06626272946596146, + 0.03213329240679741, + 0.07337209582328796, + 0.11288093775510788, + 0.03854483738541603, + 0.16512273252010345, + 0.325339138507843 + ], + "credit_map": [ + 0.003269255394116044, + 0.003360878210514784, + 0.0038314282428473234, + 0.005955295637249947, + 0.003395950887352228, + 0.0051586804911494255, + 0.005139953922480345, + 0.003605213947594166, + 0.006304681766778231, + 0.00537787564098835, + 0.007725028321146965, + 0.00437949551269412, + 0.005433549173176289, + 0.03696056827902794, + 0.0053345318883657455, + 0.03706178069114685, + 0.005839284043759108, + 0.006774854846298695, + 0.0035541467368602753, + 0.006087836343795061, + 0.0059044999070465565, + 0.0043539321050047874, + 0.007171308156102896, + 0.0043642036616802216, + 0.007309128064662218, + 0.011254838667809963, + 0.006743360310792923, + 0.007243231870234013, + 0.009235614910721779, + 0.007819518446922302, + 0.009509005583822727, + 0.007148028817027807, + 0.0030568987131118774, + 0.00350040546618402, + 0.0035760882310569286, + 0.007006247062236071, + 0.004032133147120476, + 0.0028623982798308134, + 0.004581746645271778, + 0.0035173771902918816, + 0.007891851477324963, + 0.0068452931009233, + 0.005839443299919367, + 0.011561769060790539, + 0.010320180095732212, + 0.009143223986029625, + 0.013480848632752895, + 0.008289489895105362, + 0.01186323445290327, + 0.011941807344555855, + 0.017955221235752106, + 0.015581244602799416, + 0.012413201853632927, + 0.013588900677859783, + 0.018045997247099876, + 0.011491334065794945, + 0.003343258984386921, + 0.0028942879289388657, + 0.006656475830823183, + 0.008713633753359318, + 0.0037807400804013014, + 0.004370423033833504, + 0.004041867796331644, + 0.004744150210171938, + 0.0069170244969427586, + 0.0049374038353562355, + 0.05097996070981026, + 0.009816503152251244, + 0.006519653368741274, + 0.057038869708776474, + 0.006341507658362389, + 0.022571805864572525, + 0.015759795904159546, + 0.00989559292793274, + 0.0076164002530276775, + 0.008872481994330883, + 0.009249203838407993, + 0.1273290514945984, + 0.010233032517135143, + 0.13638357818126678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.14276397228240967 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly pulls back to reveal a wider scene. Deep within a lush bamboo forest, shrouded in morning mist, a milky white mist gently drifts between the towering, straight bamboo poles. The first rays of morning sunlight struggle to penetrate the dense fog and dense bamboo leaves, casting several distinct, mottled golden beams in the air. An elderly monk with a gaunt face and a shiny head, dressed in a simple dark gray cotton and linen robe, stands in a clearing in the forest, covered with moist fallen leaves. His eyes are slightly closed, his expression serene and focused. His body moves in a steady and deliberate rhythm. His arms trace smooth arcs, his wide sleeves swaying gently, and his center of gravity shifts effortlessly between his legs. Each movement, from raising his hands to lowering his shoulders, is as smooth and natural as flowing water, imbued with inner strength and harmony. The entire picture is dominated by calm gray and vibrant emerald green, embellished by warm morning light, creating an otherworldly, Zen-like tranquil atmosphere.", + "seed": 3311, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_15/index_2985_seed_3311_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01630421867594123, + "vlm_logprob": -0.01630421867594123, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.01630421867594123, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007201665081083775 + ], + "opposite_logprob_tokens": [ + -6.257201671600342 + ], + "logprob_reward": -0.007201665081083775, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.007201665081083775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23397301137447357, + 0.11092842370271683, + 0.06551776081323624, + 0.12864696979522705, + 0.07256351411342621, + 0.05265996232628822, + 0.062345460057258606, + 0.045244161039590836, + 0.14244800806045532, + 0.08567266911268234 + ], + "credit_map": [ + 0.03020852617919445, + 0.015786971896886826, + 0.03913359344005585, + 0.02877311408519745, + 0.019802745431661606, + 0.04590979591012001, + 0.021622439846396446, + 0.03273581713438034, + 0.01381133683025837, + 0.010014001280069351, + 0.015895750373601913, + 0.012952585704624653, + 0.010832794941961765, + 0.01759503409266472, + 0.015287323854863644, + 0.014539608731865883, + 0.006770764011889696, + 0.009517151862382889, + 0.01076463982462883, + 0.008087316527962685, + 0.007003446109592915, + 0.008899875916540623, + 0.008528508245944977, + 0.005946063902229071, + 0.006000577472150326, + 0.004313658457249403, + 0.00741157541051507, + 0.007486307993531227, + 0.005963345058262348, + 0.043218210339546204, + 0.009962653741240501, + 0.044290658086538315, + 0.007726335898041725, + 0.005702184047549963, + 0.009933672845363617, + 0.00851801410317421, + 0.007546385284513235, + 0.011536895297467709, + 0.009454351849853992, + 0.01214568316936493, + 0.006657361052930355, + 0.006532767787575722, + 0.003981433808803558, + 0.004121582489460707, + 0.008686384186148643, + 0.008854606188833714, + 0.007462445646524429, + 0.006363386753946543, + 0.007744730915874243, + 0.0083168251439929, + 0.008807972073554993, + 0.0056680068373680115, + 0.00636136531829834, + 0.008980698883533478, + 0.006199623923748732, + 0.010266236960887909, + 0.005331380292773247, + 0.006007310468703508, + 0.005027357954531908, + 0.005622838623821735, + 0.00687064416706562, + 0.004532494582235813, + 0.006966861430555582, + 0.004885280504822731, + 0.01644519343972206, + 0.011907665058970451, + 0.0179625041782856, + 0.015444239601492882, + 0.015716692432761192, + 0.01944754272699356, + 0.025629833340644836, + 0.019894348457455635, + 0.008306716568768024, + 0.009232444688677788, + 0.010823037475347519, + 0.006940167862921953, + 0.011197378858923912, + 0.011167231947183609, + 0.012908736243844032, + 0.015096968039870262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.007201665081083775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651158772408962 + ], + "opposite_logprob_tokens": [ + -4.758651256561279 + ], + "logprob_reward": -0.008651158772408962, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651158772408962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05923014506697655, + 0.03609108924865723, + 0.02338135428726673, + 0.6937307715415955, + 0.0353933610022068, + 0.017628386616706848, + 0.03223983943462372, + 0.016395313665270805, + 0.06242043524980545, + 0.0234892126172781 + ], + "credit_map": [ + 0.0048745786771178246, + 0.008954116143286228, + 0.009352592751383781, + 0.0086433170363307, + 0.006163178943097591, + 0.007900420576334, + 0.006917915306985378, + 0.00642402796074748, + 0.0034194886684417725, + 0.00514830369502306, + 0.00561406509950757, + 0.004354808945208788, + 0.0046952031552791595, + 0.0039978050626814365, + 0.004595172591507435, + 0.004266239237040281, + 0.0020698660518974066, + 0.002222011098638177, + 0.004025327041745186, + 0.004222284071147442, + 0.0021408412139862776, + 0.0031421815510839224, + 0.003140251152217388, + 0.0024185909423977137, + 0.027922188863158226, + 0.015026295557618141, + 0.03580712527036667, + 0.023982862010598183, + 0.018415678292512894, + 0.2680477499961853, + 0.02729657106101513, + 0.2772323489189148, + 0.0045105693861842155, + 0.0031545860692858696, + 0.005059448536485434, + 0.003960923291742802, + 0.004677293822169304, + 0.004804619122296572, + 0.004372120834887028, + 0.004853799939155579, + 0.0026807005051523447, + 0.001967559801414609, + 0.0021526790224015713, + 0.0017537780804559588, + 0.0017256955616176128, + 0.0027658771723508835, + 0.002476424677297473, + 0.002105670515447855, + 0.003380436450242996, + 0.007504799868911505, + 0.00234541529789567, + 0.002877261256799102, + 0.0038038156926631927, + 0.003922188188880682, + 0.004206554964184761, + 0.004199365619570017, + 0.0016387569485232234, + 0.0026486229617148638, + 0.0016502401558682323, + 0.0025596115738153458, + 0.0019981139339506626, + 0.001820322941057384, + 0.0025695175863802433, + 0.0015101261669769883, + 0.007869591005146503, + 0.005437053740024567, + 0.006332344841212034, + 0.007406247779726982, + 0.007544901221990585, + 0.009799495339393616, + 0.010650989599525928, + 0.007379807997494936, + 0.0024444994051009417, + 0.0018796497024595737, + 0.0024353223852813244, + 0.0033868099562823772, + 0.0020758635364472866, + 0.0032521006651222706, + 0.005428104195743799, + 0.002586862538009882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.008651158772408962 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077924631536007 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.011077924631536007, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077924631536007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1105576679110527, + 0.04151488095521927, + 0.02889518067240715, + 0.5614244341850281, + 0.06238587945699692, + 0.04116739332675934, + 0.06523498147726059, + 0.029628366231918335, + 0.02895066700875759, + 0.030240561813116074 + ], + "credit_map": [ + 0.011277926154434681, + 0.012461713515222073, + 0.014528063125908375, + 0.02672136016190052, + 0.007580746430903673, + 0.012796448543667793, + 0.011661755852401257, + 0.013529653660953045, + 0.004040743224322796, + 0.004215697292238474, + 0.005783610045909882, + 0.006395698059350252, + 0.00450647110119462, + 0.00503225764259696, + 0.006894279271364212, + 0.004646122921258211, + 0.002643593354150653, + 0.0028205395210534334, + 0.0031799497082829475, + 0.0042070248164236546, + 0.0036030784249305725, + 0.004084078595042229, + 0.004748303908854723, + 0.003608612809330225, + 0.02305424027144909, + 0.014006305485963821, + 0.029786203056573868, + 0.024120159447193146, + 0.01655767485499382, + 0.20948804914951324, + 0.02845936082303524, + 0.21595241129398346, + 0.006489514373242855, + 0.00628140103071928, + 0.009028980508446693, + 0.007734700106084347, + 0.008178471587598324, + 0.0077846902422606945, + 0.008084098808467388, + 0.008804022334516048, + 0.0051064398139715195, + 0.007507907692342997, + 0.0028412831015884876, + 0.0063461847603321075, + 0.003712723031640053, + 0.0036034577060490847, + 0.00643856031820178, + 0.0056108334101736546, + 0.007179416250437498, + 0.013665881007909775, + 0.005004178266972303, + 0.00579832261428237, + 0.00677796034142375, + 0.009404157288372517, + 0.006564541719853878, + 0.010840521194040775, + 0.0029743865597993135, + 0.002934955060482025, + 0.003128841519355774, + 0.0033855075016617775, + 0.00391288660466671, + 0.003675043350085616, + 0.006467257160693407, + 0.0031494887080043554, + 0.00357295130379498, + 0.0024522333405911922, + 0.004384499974548817, + 0.0026420303620398045, + 0.002585120964795351, + 0.004493119195103645, + 0.004330747295171022, + 0.004489966202527285, + 0.0025758834090083838, + 0.0020059712696820498, + 0.0034081421326845884, + 0.004890234675258398, + 0.0024703138042241335, + 0.004167435225099325, + 0.00637512793764472, + 0.004347454756498337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.011077924631536007 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952369887381792 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952369887381792, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952369887381792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05344391614198685, + 0.0572134330868721, + 0.022059366106987, + 0.467299222946167, + 0.03725317865610123, + 0.018842356279492378, + 0.037788644433021545, + 0.030107134953141212, + 0.20084351301193237, + 0.0751492902636528 + ], + "credit_map": [ + 0.006737496703863144, + 0.0042559243738651276, + 0.005962569732218981, + 0.005654036067426205, + 0.004600574728101492, + 0.011483661830425262, + 0.004213776905089617, + 0.010535874404013157, + 0.006211648695170879, + 0.006480335723608732, + 0.007981779053807259, + 0.0030613241251558065, + 0.008433724753558636, + 0.011068918742239475, + 0.005447509698569775, + 0.00852818414568901, + 0.0027120320592075586, + 0.003663053037598729, + 0.0028570478316396475, + 0.0019328907364979386, + 0.003348621539771557, + 0.0027824677526950836, + 0.0022992934100329876, + 0.00246395543217659, + 0.018325602635741234, + 0.010461447760462761, + 0.02323257550597191, + 0.01687045209109783, + 0.012342799454927444, + 0.18068765103816986, + 0.01926710642874241, + 0.18611156940460205, + 0.003889760235324502, + 0.003270166926085949, + 0.004939158912748098, + 0.0042698620818555355, + 0.003982688300311565, + 0.007298510521650314, + 0.003744599176570773, + 0.005858429707586765, + 0.0026622270233929157, + 0.0016485416563227773, + 0.002292715711519122, + 0.0017925046849995852, + 0.002679716795682907, + 0.002480305265635252, + 0.002796547021716833, + 0.0024897961411625147, + 0.004861634224653244, + 0.008149155415594578, + 0.0036827956791967154, + 0.003906393889337778, + 0.004707836080342531, + 0.004205210600048304, + 0.004300129599869251, + 0.003975486382842064, + 0.0046002487652003765, + 0.004464828874915838, + 0.002841403242200613, + 0.0029198743868619204, + 0.0035014075692743063, + 0.0037753391079604626, + 0.0036849500611424446, + 0.004319081548601389, + 0.028992123901844025, + 0.0318869911134243, + 0.01853717304766178, + 0.028113791719079018, + 0.022508198395371437, + 0.02567257173359394, + 0.02401508204638958, + 0.021117577329277992, + 0.013116495683789253, + 0.00768815353512764, + 0.012565560638904572, + 0.007491030730307102, + 0.008213844150304794, + 0.009232386946678162, + 0.007767101749777794, + 0.009074710309505463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.005952369887381792 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863797500729561 + ], + "opposite_logprob_tokens": [ + -3.048637866973877 + ], + "logprob_reward": -0.04863797500729561, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863797500729561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023741384968161583, + 0.014391475357115269, + 0.01617531105875969, + 0.7211135029792786, + 0.007961848750710487, + 0.00995948538184166, + 0.05748898535966873, + 0.019952066242694855, + 0.10850878059864044, + 0.020707206800580025 + ], + "credit_map": [ + 0.0024735478218644857, + 0.002530523808673024, + 0.002227404387667775, + 0.0025780408177524805, + 0.0022738855332136154, + 0.004218979272991419, + 0.0018676463514566422, + 0.005571356508880854, + 0.002122438047081232, + 0.001737684360705316, + 0.0019507878459990025, + 0.001057514688000083, + 0.0019211324397474527, + 0.002485530450940132, + 0.001295878435485065, + 0.0018205082742497325, + 0.0017292620614171028, + 0.0016563942190259695, + 0.003007295308634639, + 0.002285977127030492, + 0.00163087563123554, + 0.002006457420065999, + 0.0019063326762989163, + 0.0019527164986357093, + 0.029087433591485023, + 0.01585261896252632, + 0.0374133326113224, + 0.026568129658699036, + 0.01974049024283886, + 0.27705830335617065, + 0.02954948879778385, + 0.28584370017051697, + 0.0006361196283251047, + 0.0008732259157113731, + 0.0007502245134674013, + 0.001016051508486271, + 0.0010087152477353811, + 0.0013662608107551932, + 0.0009656643960624933, + 0.0013455867301672697, + 0.0008704880601726472, + 0.0009550604736432433, + 0.00136042560916394, + 0.0009526453795842826, + 0.0012508617946878076, + 0.0016122498782351613, + 0.0015706282574683428, + 0.001387125696055591, + 0.006643251515924931, + 0.013803073205053806, + 0.0042728241533041, + 0.005086633842438459, + 0.006774787791073322, + 0.006738628260791302, + 0.006507097277790308, + 0.007662689778953791, + 0.0018968149088323116, + 0.0025396051350980997, + 0.0021137967705726624, + 0.002323342952877283, + 0.002722656587138772, + 0.0025263044517487288, + 0.0035734614357352257, + 0.0022560846991837025, + 0.013398152776062489, + 0.009286312386393547, + 0.010585249401628971, + 0.013967020437121391, + 0.013177822344005108, + 0.01673934794962406, + 0.018348703160881996, + 0.013006174936890602, + 0.00237448001280427, + 0.00158933294005692, + 0.0027072951197624207, + 0.0023560808040201664, + 0.0019459977047517896, + 0.003107342403382063, + 0.004091410432010889, + 0.002535266801714897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.04863797500729561 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly pulls back to reveal a wider scene. Deep within a lush bamboo forest, shrouded in morning mist, a milky white mist gently drifts between the towering, straight bamboo poles. The first rays of morning sunlight struggle to penetrate the dense fog and dense bamboo leaves, casting several distinct, mottled golden beams in the air. An elderly monk with a gaunt face and a shiny head, dressed in a simple dark gray cotton and linen robe, stands in a clearing in the forest, covered with moist fallen leaves. His eyes are slightly closed, his expression serene and focused. His body moves in a steady and deliberate rhythm. His arms trace smooth arcs, his wide sleeves swaying gently, and his center of gravity shifts effortlessly between his legs. Each movement, from raising his hands to lowering his shoulders, is as smooth and natural as flowing water, imbued with inner strength and harmony. The entire picture is dominated by calm gray and vibrant emerald green, embellished by warm morning light, creating an otherworldly, Zen-like tranquil atmosphere.", + "seed": 3312, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_15/index_2985_seed_3312_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.026185115054249763, + "vlm_logprob": -0.026185115054249763, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.026185115054249763, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008440653793513775 + ], + "opposite_logprob_tokens": [ + -6.0084404945373535 + ], + "logprob_reward": -0.008440653793513775, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.008440653793513775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11613480746746063, + 0.2774428427219391, + 0.08863193541765213, + 0.05102004110813141, + 0.05656077340245247, + 0.05895739048719406, + 0.045277055352926254, + 0.09439829736948013, + 0.10008282214403152, + 0.11149407923221588 + ], + "credit_map": [ + 0.02448832429945469, + 0.017590243369340897, + 0.011941066011786461, + 0.01596868596971035, + 0.007845999673008919, + 0.014151215553283691, + 0.008450185880064964, + 0.015699079260230064, + 0.028187314048409462, + 0.008213742636144161, + 0.022899964824318886, + 0.056891199201345444, + 0.012055090628564358, + 0.012340211309492588, + 0.116702139377594, + 0.020153185352683067, + 0.021050548180937767, + 0.00945950299501419, + 0.013379953801631927, + 0.0106222378090024, + 0.011569648049771786, + 0.006982643157243729, + 0.00915659312158823, + 0.006410801783204079, + 0.007136420346796513, + 0.0040520429611206055, + 0.007018284872174263, + 0.006901032291352749, + 0.00599496578797698, + 0.007775599602609873, + 0.005738682579249144, + 0.006403009872883558, + 0.006047710310667753, + 0.006863880902528763, + 0.006447713356465101, + 0.006201290059834719, + 0.010903731919825077, + 0.006103543099015951, + 0.007221243344247341, + 0.0067716557532548904, + 0.012679378502070904, + 0.009304123930633068, + 0.00888668280094862, + 0.005186181981116533, + 0.006222716066986322, + 0.008339288644492626, + 0.0036137118004262447, + 0.004725304432213306, + 0.011441563256084919, + 0.003466022899374366, + 0.0072012352757155895, + 0.003163453657180071, + 0.0032710477244108915, + 0.007605064660310745, + 0.0034715908113867044, + 0.005657073576003313, + 0.01023939810693264, + 0.015791185200214386, + 0.01079082116484642, + 0.010965143330395222, + 0.016620762646198273, + 0.011782228946685791, + 0.010923239402472973, + 0.007285508792847395, + 0.0191813837736845, + 0.009999171830713749, + 0.015006747096776962, + 0.010768257081508636, + 0.01154721062630415, + 0.016536852344870567, + 0.0061423988081514835, + 0.010900789871811867, + 0.030093347653746605, + 0.00800141878426075, + 0.02118825539946556, + 0.007740649860352278, + 0.008642289787530899, + 0.017855482175946236, + 0.007369381841272116, + 0.010603243485093117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.008440653793513775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203292317688465 + ], + "opposite_logprob_tokens": [ + -4.264203071594238 + ], + "logprob_reward": -0.014203292317688465, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203292317688465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1860526204109192, + 0.21037618815898895, + 0.1679675281047821, + 0.05317537486553192, + 0.050630148500204086, + 0.07017935812473297, + 0.033563096076250076, + 0.16796860098838806, + 0.018150748685002327, + 0.04193625971674919 + ], + "credit_map": [ + 0.025770453736186028, + 0.03607870638370514, + 0.017802195623517036, + 0.024262219667434692, + 0.012401611544191837, + 0.017765946686267853, + 0.018077442422509193, + 0.03389405086636543, + 0.0157761313021183, + 0.009530731476843357, + 0.018110547214746475, + 0.04380284249782562, + 0.00913539994508028, + 0.013998734764754772, + 0.08524943888187408, + 0.014772353693842888, + 0.02397390455007553, + 0.019276561215519905, + 0.02463153563439846, + 0.019206034019589424, + 0.030811794102191925, + 0.01705649495124817, + 0.017989015206694603, + 0.015022194012999535, + 0.006224366836249828, + 0.005397666711360216, + 0.006531276274472475, + 0.007578135468065739, + 0.00729789724573493, + 0.007560718338936567, + 0.006537009030580521, + 0.006048307288438082, + 0.006595357321202755, + 0.005121506284922361, + 0.006193880923092365, + 0.00671547232195735, + 0.007642651908099651, + 0.005399172659963369, + 0.006035835016518831, + 0.00692627253010869, + 0.012032503262162209, + 0.007750804536044598, + 0.0110173923894763, + 0.004819324240088463, + 0.009521484375, + 0.012053009122610092, + 0.005070158280432224, + 0.007914680056273937, + 0.00732874171808362, + 0.00272648804821074, + 0.006659881677478552, + 0.002469828585162759, + 0.002528904704377055, + 0.005058960057795048, + 0.0032156130764633417, + 0.0035746810026466846, + 0.021871812641620636, + 0.024383854120969772, + 0.02036646567285061, + 0.014429360628128052, + 0.030030518770217896, + 0.025850564241409302, + 0.017546700313687325, + 0.013489329256117344, + 0.003198231803253293, + 0.0020804915111511946, + 0.002280935412272811, + 0.002141030039638281, + 0.0022140180226415396, + 0.0026028172578662634, + 0.0018392514903098345, + 0.00179397186730057, + 0.008061450906097889, + 0.002358742756769061, + 0.008142215199768543, + 0.0053334622643888, + 0.003290249966084957, + 0.005094265099614859, + 0.004056583158671856, + 0.0055992864072322845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.014203292317688465 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539473362267017 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539473362267017, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539473362267017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09306485205888748, + 0.2744677662849426, + 0.19895273447036743, + 0.08754155784845352, + 0.1118713989853859, + 0.06475341320037842, + 0.027852702885866165, + 0.06392055749893188, + 0.03400852903723717, + 0.0435665138065815 + ], + "credit_map": [ + 0.013324967585504055, + 0.02253684215247631, + 0.009662939235568047, + 0.011111360974609852, + 0.007291064132004976, + 0.009353435598313808, + 0.009555891156196594, + 0.010228345170617104, + 0.030520977452397346, + 0.010677583515644073, + 0.023360319435596466, + 0.06370124965906143, + 0.011225470341742039, + 0.015277166850864887, + 0.10649134963750839, + 0.01321366336196661, + 0.04150322824716568, + 0.02523002214729786, + 0.02864229679107666, + 0.019291827455163002, + 0.03386300802230835, + 0.016296962276101112, + 0.017156675457954407, + 0.01696871966123581, + 0.009023185819387436, + 0.006161361001431942, + 0.012623489834368229, + 0.017899446189403534, + 0.009735247120261192, + 0.008254368789494038, + 0.011854536831378937, + 0.011989917606115341, + 0.014070424251258373, + 0.011476757004857063, + 0.012496900744736195, + 0.01769021712243557, + 0.010519950650632381, + 0.013755560852587223, + 0.014346239157021046, + 0.01751535013318062, + 0.009587432257831097, + 0.006779185030609369, + 0.007156975567340851, + 0.012193918228149414, + 0.006664527580142021, + 0.007579449564218521, + 0.006439154036343098, + 0.008352772332727909, + 0.003870626213029027, + 0.004563109017908573, + 0.004802035167813301, + 0.003983706701546907, + 0.0021828312892466784, + 0.002952111419290304, + 0.003018095623701811, + 0.002480186754837632, + 0.011332175694406033, + 0.010444043204188347, + 0.007421704009175301, + 0.00529587734490633, + 0.011280185543000698, + 0.007933969609439373, + 0.005662817973643541, + 0.004549779929220676, + 0.004812844563275576, + 0.0034291448537260294, + 0.004626552574336529, + 0.004031216260045767, + 0.0035653389059007168, + 0.004648295696824789, + 0.004417764954268932, + 0.0044773719273507595, + 0.004813740961253643, + 0.0035249670036137104, + 0.006673584692180157, + 0.006197480950504541, + 0.0037266649305820465, + 0.004358362872153521, + 0.005137664265930653, + 0.009134047664701939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.012539473362267017 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738794036209583 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.006738794036209583, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738794036209583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22338885068893433, + 0.14377836883068085, + 0.33813560009002686, + 0.04566775634884834, + 0.056629423052072525, + 0.05556458979845047, + 0.03621738776564598, + 0.028927473351359367, + 0.01896943710744381, + 0.05272108316421509 + ], + "credit_map": [ + 0.03739277273416519, + 0.0361693911254406, + 0.021672815084457397, + 0.024598462507128716, + 0.025075605139136314, + 0.04125504568219185, + 0.016972927376627922, + 0.020251834765076637, + 0.01217041164636612, + 0.010443009436130524, + 0.012576987035572529, + 0.02987075410783291, + 0.010070172138512135, + 0.01363945472985506, + 0.042719777673482895, + 0.012287799268960953, + 0.06718198955059052, + 0.04311128705739975, + 0.04551384225487709, + 0.03520051762461662, + 0.06301486492156982, + 0.030941441655158997, + 0.03108975477516651, + 0.02208189107477665, + 0.007514632306993008, + 0.005979444831609726, + 0.006060140673071146, + 0.004735463298857212, + 0.005061273463070393, + 0.0066765653900802135, + 0.005386152770370245, + 0.004254084080457687, + 0.008078492246568203, + 0.007216190453618765, + 0.006695610471069813, + 0.004938002210110426, + 0.005674297921359539, + 0.010858670808374882, + 0.005365614779293537, + 0.007802541833370924, + 0.009353029541671276, + 0.006266874726861715, + 0.008384136483073235, + 0.004283726681023836, + 0.008441584184765816, + 0.008371705189347267, + 0.004609579686075449, + 0.005853952839970589, + 0.010162723250687122, + 0.0033192490227520466, + 0.005888429470360279, + 0.0027704224921762943, + 0.0032385149970650673, + 0.00466505903750658, + 0.002627262379974127, + 0.003545725019648671, + 0.0032783830538392067, + 0.00288673909381032, + 0.00520780123770237, + 0.003020557342097163, + 0.002603996777907014, + 0.005327772814780474, + 0.003606636542826891, + 0.0029955869540572166, + 0.004418901167809963, + 0.0015825133305042982, + 0.0030363639816641808, + 0.0020055463537573814, + 0.0014505124418064952, + 0.0030074999667704105, + 0.0017930669710040092, + 0.0016750306822359562, + 0.014238240197300911, + 0.003825928783044219, + 0.010251822881400585, + 0.0034731619525700808, + 0.003883974626660347, + 0.008658403530716896, + 0.003599046263843775, + 0.004790506791323423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.006738794036209583 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900336176156998 + ], + "opposite_logprob_tokens": [ + -2.464003324508667 + ], + "logprob_reward": -0.08900336176156998, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900336176156998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1400584578514099, + 0.11207250505685806, + 0.15282922983169556, + 0.09322386980056763, + 0.13409210741519928, + 0.08172309398651123, + 0.0416555181145668, + 0.18966878950595856, + 0.014699050225317478, + 0.03997737541794777 + ], + "credit_map": [ + 0.024992046877741814, + 0.021680599078536034, + 0.015918415039777756, + 0.02094452641904354, + 0.01015774067491293, + 0.014500327408313751, + 0.013793379068374634, + 0.018071424216032028, + 0.013774239458143711, + 0.003767960472032428, + 0.009782632812857628, + 0.024218589067459106, + 0.004743078723549843, + 0.00573941133916378, + 0.043603744357824326, + 0.006442850921303034, + 0.020681682974100113, + 0.017893319949507713, + 0.02082529477775097, + 0.018944915384054184, + 0.02186984196305275, + 0.018306424841284752, + 0.015456177294254303, + 0.018851570785045624, + 0.009782279841601849, + 0.008578766137361526, + 0.015775563195347786, + 0.012015457265079021, + 0.012518322095274925, + 0.013079955242574215, + 0.010599080473184586, + 0.010874447412788868, + 0.011448914185166359, + 0.014675837941467762, + 0.01496054232120514, + 0.013437333516776562, + 0.025526119396090508, + 0.02082885056734085, + 0.017891118302941322, + 0.015323393978178501, + 0.016587335616350174, + 0.008210591040551662, + 0.012493433430790901, + 0.007073207292705774, + 0.0072508384473621845, + 0.014839247800409794, + 0.006246679928153753, + 0.009021764621138573, + 0.0076806130819022655, + 0.004957468248903751, + 0.005724024958908558, + 0.003498874604701996, + 0.004120969213545322, + 0.005068336147814989, + 0.003993398044258356, + 0.006611833814531565, + 0.04868675768375397, + 0.021429529413580894, + 0.02879200130701065, + 0.01595163345336914, + 0.012765511870384216, + 0.026327291503548622, + 0.019477689638733864, + 0.016238370910286903, + 0.002293545054271817, + 0.0019382218597456813, + 0.0017755711451172829, + 0.001848653657361865, + 0.0018450706265866756, + 0.0015358665259554982, + 0.001892358181066811, + 0.0015697629423812032, + 0.009957854636013508, + 0.0026903192047029734, + 0.006941405590623617, + 0.004051921423524618, + 0.0028837830759584904, + 0.006377141922712326, + 0.002917113248258829, + 0.004157835617661476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.08900336176156998 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly pulls back to reveal a wider scene. Deep within a lush bamboo forest, shrouded in morning mist, a milky white mist gently drifts between the towering, straight bamboo poles. The first rays of morning sunlight struggle to penetrate the dense fog and dense bamboo leaves, casting several distinct, mottled golden beams in the air. An elderly monk with a gaunt face and a shiny head, dressed in a simple dark gray cotton and linen robe, stands in a clearing in the forest, covered with moist fallen leaves. His eyes are slightly closed, his expression serene and focused. His body moves in a steady and deliberate rhythm. His arms trace smooth arcs, his wide sleeves swaying gently, and his center of gravity shifts effortlessly between his legs. Each movement, from raising his hands to lowering his shoulders, is as smooth and natural as flowing water, imbued with inner strength and harmony. The entire picture is dominated by calm gray and vibrant emerald green, embellished by warm morning light, creating an otherworldly, Zen-like tranquil atmosphere.", + "seed": 3313, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_15/index_2985_seed_3313_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04189310073852539, + "vlm_logprob": -0.04189310073852539, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.04189310073852539, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008815537206828594 + ], + "opposite_logprob_tokens": [ + -6.758815765380859 + ], + "logprob_reward": -0.008815537206828594, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.008815537206828594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20091606676578522, + 0.03969661146402359, + 0.18154966831207275, + 0.05891125649213791, + 0.06262874603271484, + 0.06456014513969421, + 0.2330920547246933, + 0.03785913437604904, + 0.07840888202190399, + 0.042377445846796036 + ], + "credit_map": [ + 0.036063797771930695, + 0.015742475166916847, + 0.04870980978012085, + 0.018295368179678917, + 0.025722013786435127, + 0.01989562250673771, + 0.013304062187671661, + 0.023182909935712814, + 0.0040685455314815044, + 0.005129638127982616, + 0.004559580236673355, + 0.004283267538994551, + 0.00440243910998106, + 0.006150207947939634, + 0.0066296798177063465, + 0.004473248030990362, + 0.011710822582244873, + 0.02020880952477455, + 0.015013259835541248, + 0.030431600287556648, + 0.017041238024830818, + 0.021078210324048996, + 0.044005487114191055, + 0.022060219198465347, + 0.007221086882054806, + 0.005940563976764679, + 0.006176995579153299, + 0.006679081823676825, + 0.007638546172529459, + 0.008546382188796997, + 0.008291808888316154, + 0.008416790515184402, + 0.010900000110268593, + 0.005817690398544073, + 0.007450217381119728, + 0.006049277260899544, + 0.007033922243863344, + 0.009195234626531601, + 0.0084720803424716, + 0.007710323203355074, + 0.011888952925801277, + 0.005890535190701485, + 0.0060033369809389114, + 0.0069127436727285385, + 0.009592308662831783, + 0.00919823907315731, + 0.007875104434788227, + 0.007198919076472521, + 0.05404345318675041, + 0.016864502802491188, + 0.029142217710614204, + 0.020629316568374634, + 0.021406501531600952, + 0.047935325652360916, + 0.017704665660858154, + 0.025366056710481644, + 0.0037985872477293015, + 0.0038575944490730762, + 0.0054414221085608006, + 0.004760071635246277, + 0.0058381203562021255, + 0.004104485735297203, + 0.00709304166957736, + 0.0029658074490725994, + 0.016748372465372086, + 0.007367847487330437, + 0.010123337619006634, + 0.0051225582137703896, + 0.00951952114701271, + 0.0110020712018013, + 0.00772866141051054, + 0.01079651154577732, + 0.007811062503606081, + 0.004528071265667677, + 0.006767577026039362, + 0.0038488551508635283, + 0.005091686267405748, + 0.005432038567960262, + 0.0038526749704033136, + 0.005045478232204914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.008815537206828594 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011090068146586418 + ], + "opposite_logprob_tokens": [ + -4.511090278625488 + ], + "logprob_reward": -0.011090068146586418, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011090068146586418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31931301951408386, + 0.07541836053133011, + 0.0559537336230278, + 0.1647377461194992, + 0.04521838203072548, + 0.039060577750205994, + 0.18211963772773743, + 0.03352360427379608, + 0.06348790228366852, + 0.021167026832699776 + ], + "credit_map": [ + 0.06296196579933167, + 0.019349388778209686, + 0.08276206254959106, + 0.026558779180049896, + 0.02812270075082779, + 0.031778015196323395, + 0.025024481117725372, + 0.042755648493766785, + 0.01320846751332283, + 0.0058908830396831036, + 0.017948869615793228, + 0.006069982424378395, + 0.006998386234045029, + 0.009732812643051147, + 0.008346300572156906, + 0.0072226631455123425, + 0.005594627000391483, + 0.005052685271948576, + 0.007867544889450073, + 0.007171545643359423, + 0.007524983026087284, + 0.007167188450694084, + 0.007446878124028444, + 0.00812828540802002, + 0.01071214396506548, + 0.018031584098935127, + 0.01288582757115364, + 0.030658332630991936, + 0.022302506491541862, + 0.015366703271865845, + 0.03650616109371185, + 0.018274487927556038, + 0.008147677406668663, + 0.0035200254060328007, + 0.0071176812052726746, + 0.0044486369006335735, + 0.0050369263626635075, + 0.005757239647209644, + 0.004876298829913139, + 0.006313900928944349, + 0.005539339501410723, + 0.004686106462031603, + 0.004701707512140274, + 0.004924592096358538, + 0.006629638373851776, + 0.0040625520050525665, + 0.004770981147885323, + 0.0037456643767654896, + 0.042470190674066544, + 0.01332523301243782, + 0.023178428411483765, + 0.01561885979026556, + 0.01709446683526039, + 0.03795989975333214, + 0.012656078673899174, + 0.019816480576992035, + 0.0039013244677335024, + 0.004768869839608669, + 0.004193817265331745, + 0.0033060291316360235, + 0.005777690559625626, + 0.003879839787259698, + 0.00411319499835372, + 0.003582841483876109, + 0.012991044670343399, + 0.005137852393090725, + 0.007774054538458586, + 0.004910204093903303, + 0.005403720308095217, + 0.010058642365038395, + 0.0078499186784029, + 0.00936246570199728, + 0.0035990485921502113, + 0.0020641342271119356, + 0.0026045930571854115, + 0.002792627550661564, + 0.0026091323234140873, + 0.0025308786425739527, + 0.002401132369413972, + 0.00256548123434186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.011090068146586418 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584525540471077 + ], + "opposite_logprob_tokens": [ + -3.8955845832824707 + ], + "logprob_reward": -0.020584525540471077, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584525540471077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1454300731420517, + 0.04646075889468193, + 0.06387200951576233, + 0.1851300746202469, + 0.0649985820055008, + 0.06994296610355377, + 0.10460339486598969, + 0.06106433644890785, + 0.19749675691127777, + 0.0610009990632534 + ], + "credit_map": [ + 0.013219132088124752, + 0.011761288158595562, + 0.017766816541552544, + 0.021528763696551323, + 0.024124404415488243, + 0.015482802875339985, + 0.026479782536625862, + 0.015067082829773426, + 0.00492369569838047, + 0.004779207054525614, + 0.005471543874591589, + 0.0045352960005402565, + 0.007256553508341312, + 0.00555714126676321, + 0.007735147140920162, + 0.0062021720223128796, + 0.009847324341535568, + 0.007137693930417299, + 0.011656934395432472, + 0.00636312086135149, + 0.005540570709854364, + 0.009011635556817055, + 0.006405142601579428, + 0.007909592241048813, + 0.015903227031230927, + 0.016720682382583618, + 0.02030280791223049, + 0.03246276453137398, + 0.022577708587050438, + 0.01416042447090149, + 0.03840019553899765, + 0.024602258577942848, + 0.007056034170091152, + 0.005657946690917015, + 0.00548518355935812, + 0.010306314565241337, + 0.007708698511123657, + 0.007678251247853041, + 0.013293159194290638, + 0.007812993600964546, + 0.007255232892930508, + 0.007707399316132069, + 0.011790572665631771, + 0.00981717836111784, + 0.00915267039090395, + 0.005644428078085184, + 0.010524263605475426, + 0.00805122684687376, + 0.022619545459747314, + 0.008801446296274662, + 0.01163289975374937, + 0.009823461063206196, + 0.010697620920836926, + 0.019414810463786125, + 0.010261671617627144, + 0.011351941153407097, + 0.0063257766887545586, + 0.006406020373106003, + 0.00861265603452921, + 0.0072791362181305885, + 0.010260111652314663, + 0.0070386240258812904, + 0.009916368871927261, + 0.005225643515586853, + 0.03309480845928192, + 0.01899540051817894, + 0.0227375328540802, + 0.019202951341867447, + 0.03183623030781746, + 0.02480027824640274, + 0.022780420258641243, + 0.024049142375588417, + 0.006499385926872492, + 0.006722887046635151, + 0.0067141093313694, + 0.008615788072347641, + 0.008951635099947453, + 0.005659033544361591, + 0.010638278909027576, + 0.007199884857982397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.020584525540471077 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644540794193745 + ], + "opposite_logprob_tokens": [ + -4.7586445808410645 + ], + "logprob_reward": -0.008644540794193745, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644540794193745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3015076816082001, + 0.2148190289735794, + 0.044418565928936005, + 0.05866064131259918, + 0.03397611528635025, + 0.05387645587325096, + 0.17591124773025513, + 0.022082291543483734, + 0.05440647527575493, + 0.040341444313526154 + ], + "credit_map": [ + 0.06552950292825699, + 0.023984307423233986, + 0.08330179750919342, + 0.01763402484357357, + 0.02346675656735897, + 0.03425099700689316, + 0.015943555161356926, + 0.03739675134420395, + 0.04238096624612808, + 0.016629381105303764, + 0.046141788363456726, + 0.01609770581126213, + 0.019105885177850723, + 0.03416761755943298, + 0.020433468744158745, + 0.019862204790115356, + 0.0072319949977099895, + 0.00483794417232275, + 0.010046376846730709, + 0.004096972290426493, + 0.003102001268416643, + 0.005236536730080843, + 0.00440405635163188, + 0.005462681874632835, + 0.005819330923259258, + 0.005767876282334328, + 0.005380173213779926, + 0.009221617132425308, + 0.005980021320283413, + 0.006566949188709259, + 0.012783884070813656, + 0.007140787784010172, + 0.0067055849358439445, + 0.0032148626632988453, + 0.004232681356370449, + 0.002812424674630165, + 0.0032663953024894, + 0.005678993184119463, + 0.003395749954506755, + 0.00466942461207509, + 0.012902944348752499, + 0.004543967545032501, + 0.0071622515097260475, + 0.004312547389417887, + 0.004661478102207184, + 0.009883604943752289, + 0.0037765931338071823, + 0.0066330707632005215, + 0.03614019230008125, + 0.013177028857171535, + 0.024244345724582672, + 0.015274858102202415, + 0.016279220581054688, + 0.03213674947619438, + 0.018809961155056953, + 0.01984890177845955, + 0.0026434126775711775, + 0.003092887345701456, + 0.0028023754712194204, + 0.0022857438307255507, + 0.0030435200314968824, + 0.002725861268118024, + 0.0032191246282309294, + 0.002269366290420294, + 0.007272355258464813, + 0.00555758411064744, + 0.005618688650429249, + 0.005283956415951252, + 0.010604577139019966, + 0.006489865481853485, + 0.006467845756560564, + 0.007111601531505585, + 0.009414544329047203, + 0.004598394501954317, + 0.005578232929110527, + 0.0031472998671233654, + 0.003620331408455968, + 0.006586445961147547, + 0.00300071039237082, + 0.004395479802042246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.008644540794193745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush bamboo forest with towering straight bamboo poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly monk with a gaunt face and shiny head wearing a dark gray cotton and linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk performing slow, deliberate movements with smooth arm arcs and swaying wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature mottled golden beams of sunlight penetrating the mist and bamboo leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot at a level angle and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033083200454712 + ], + "opposite_logprob_tokens": [ + -1.9103307723999023 + ], + "logprob_reward": -0.16033083200454712, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033083200454712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1624753177165985, + 0.06282903999090195, + 0.31822335720062256, + 0.09780273586511612, + 0.04427511990070343, + 0.030491769313812256, + 0.11903712898492813, + 0.03709804266691208, + 0.1036430299282074, + 0.02412446402013302 + ], + "credit_map": [ + 0.03222169727087021, + 0.011010115034878254, + 0.04057665169239044, + 0.015417213551700115, + 0.01627935841679573, + 0.01502847857773304, + 0.011406361125409603, + 0.020535429939627647, + 0.010480564087629318, + 0.005205593071877956, + 0.010359687730669975, + 0.005867281928658485, + 0.007902238517999649, + 0.008799993433058262, + 0.008741263300180435, + 0.005472411401569843, + 0.03402559086680412, + 0.03465035557746887, + 0.03516106680035591, + 0.046403758227825165, + 0.030072607100009918, + 0.03185395896434784, + 0.06895996630191803, + 0.03709602355957031, + 0.012314639054238796, + 0.007823248393833637, + 0.012061034329235554, + 0.012971671298146248, + 0.011487334035336971, + 0.011210638098418713, + 0.018059968948364258, + 0.011874204501509666, + 0.009764345362782478, + 0.0041759200394153595, + 0.003898193594068289, + 0.005103299859911203, + 0.0048212092369794846, + 0.005780269391834736, + 0.005276314914226532, + 0.005455570295453072, + 0.0025193963665515184, + 0.003576076589524746, + 0.0036418698728084564, + 0.0044415914453566074, + 0.00415848521515727, + 0.0022316910326480865, + 0.006720084231346846, + 0.003202575258910656, + 0.027613135054707527, + 0.0069544389843940735, + 0.01431961264461279, + 0.009066897444427013, + 0.011264978908002377, + 0.025167962536215782, + 0.011060785502195358, + 0.01358930766582489, + 0.003123331116512418, + 0.0031111673451960087, + 0.004043594468384981, + 0.005601535085588694, + 0.0073255025781691074, + 0.003704291768372059, + 0.006635303609073162, + 0.0035533143673092127, + 0.019109629094600677, + 0.007393719162791967, + 0.012520762160420418, + 0.007708369288593531, + 0.008592329919338226, + 0.017201298847794533, + 0.01473836787045002, + 0.016378549858927727, + 0.004734782502055168, + 0.0024936634581536055, + 0.002783278003334999, + 0.002765695098787546, + 0.0026381993666291237, + 0.0030921720899641514, + 0.0032556219957768917, + 0.002361048711463809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.16033083200454712 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_2/index_tensor([274, 274])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_2/index_tensor([274, 274])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0f338b9dc996ec0d8e9eb047d3ab59a62d560a6a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_2/index_tensor([274, 274])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning macro close-up, captured from a slightly downward angle, captures a broad, vibrant green leaf, its surface dotted with countless shimmering dewdrops, each like a miniature crystal ball, refracting the soft morning light. On this microscopic battlefield, a mantis, its emerald green like a finely crafted jade, clashes with a sleek black spider. The mantis arches its body, its signature sickle-shaped forelegs raised high. Its triangular head tilts slightly, its compound eyes gleaming with a cold luster, a poised aggression. Meanwhile, the black spider, its opponent, crouches low, its eight long, finely haired legs stretched out nervously for stability, its fangs bared, poised for a deadly counterattack. As the mantis swung its front legs like lightning and the spider dodged to the side swiftly, a fierce and silent fight broke out in an instant. Every movement of theirs stirred the surrounding dewdrops, causing them to break or roll off the edge of the leaves. The whole picture was full of primitive, tense and gripping dramatic tension against the tranquil and beautiful background.", + "seed": 26665, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_2/index_274_seed_26665_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.017338176304474474, + "vlm_logprob": -0.017338176304474474, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.017338176304474474, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022774015087634325 + ], + "opposite_logprob_tokens": [ + -7.502277374267578 + ], + "logprob_reward": -0.0022774015087634325, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0022774015087634325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14393840730190277, + 0.059443339705467224, + 0.06165684387087822, + 0.06890279054641724, + 0.036108553409576416, + 0.1274465173482895, + 0.051154859364032745, + 0.04964546486735344, + 0.13154257833957672, + 0.27016061544418335 + ], + "credit_map": [ + 0.01656242646276951, + 0.01673351414501667, + 0.022269494831562042, + 0.023235131055116653, + 0.016262227669358253, + 0.023612042888998985, + 0.010391932912170887, + 0.014871640130877495, + 0.008252209052443504, + 0.0069113862700760365, + 0.007618043106049299, + 0.00681614363566041, + 0.01238962821662426, + 0.007349770981818438, + 0.003911587875336409, + 0.006194573827087879, + 0.007862179540097713, + 0.00980575755238533, + 0.0068835243582725525, + 0.0055938949808478355, + 0.012164788320660591, + 0.008331095799803734, + 0.006106603424996138, + 0.0049090031534433365, + 0.009186333045363426, + 0.011230533942580223, + 0.010766280815005302, + 0.010500431060791016, + 0.007425519172102213, + 0.009870068170130253, + 0.006305743474513292, + 0.0036178871523588896, + 0.0047380197793245316, + 0.007839446887373924, + 0.0044694324024021626, + 0.004250373225659132, + 0.00363750453107059, + 0.004073407035320997, + 0.003844763385131955, + 0.003255604300647974, + 0.015881357714533806, + 0.0328751876950264, + 0.013832204043865204, + 0.02026420831680298, + 0.012054501101374626, + 0.011991002596914768, + 0.011394276283681393, + 0.009153785184025764, + 0.007195762824267149, + 0.01261071115732193, + 0.005622184835374355, + 0.006508513353765011, + 0.0050609358586370945, + 0.003973691258579493, + 0.005877593066543341, + 0.004305465612560511, + 0.006645428482443094, + 0.007939331233501434, + 0.006499446462839842, + 0.005993046797811985, + 0.005375748500227928, + 0.007523703388869762, + 0.004946349188685417, + 0.0047224098816514015, + 0.01913353055715561, + 0.015721503645181656, + 0.012323351576924324, + 0.013013687916100025, + 0.02031046338379383, + 0.01517228689044714, + 0.024043595418334007, + 0.011824156157672405, + 0.04083860293030739, + 0.03864661976695061, + 0.05072903633117676, + 0.021922515705227852, + 0.05871840566396713, + 0.028035959228873253, + 0.017370423302054405, + 0.013899081386625767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0022774015087634325 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020594336092472076 + ], + "opposite_logprob_tokens": [ + -3.895594358444214 + ], + "logprob_reward": -0.020594336092472076, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594336092472076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13940592110157013, + 0.1516844928264618, + 0.0770518034696579, + 0.0733824148774147, + 0.04338128864765167, + 0.04874294996261597, + 0.08957315981388092, + 0.1047782450914383, + 0.11256440728902817, + 0.15943530201911926 + ], + "credit_map": [ + 0.015596183016896248, + 0.012941139750182629, + 0.023697111755609512, + 0.023800119757652283, + 0.012389568611979485, + 0.023411883041262627, + 0.010363895446062088, + 0.01720602810382843, + 0.017205115407705307, + 0.02249125950038433, + 0.021864399313926697, + 0.014182787388563156, + 0.024727528914809227, + 0.017808521166443825, + 0.017426254227757454, + 0.015978623181581497, + 0.0170929916203022, + 0.008193197660148144, + 0.005396065767854452, + 0.005901430733501911, + 0.007939456962049007, + 0.02169128879904747, + 0.005846188869327307, + 0.004991190042346716, + 0.010581792332231998, + 0.008580558933317661, + 0.010089565999805927, + 0.00950457900762558, + 0.00661042844876647, + 0.01820826530456543, + 0.005422893911600113, + 0.004384327679872513, + 0.005923445336520672, + 0.0077806380577385426, + 0.005929022096097469, + 0.005647816229611635, + 0.003764989785850048, + 0.0057660662569105625, + 0.005408925004303455, + 0.003160390304401517, + 0.006221182644367218, + 0.01097329892218113, + 0.0046656206250190735, + 0.007193020544946194, + 0.005090661812573671, + 0.004826124291867018, + 0.005215839482843876, + 0.004557208623737097, + 0.01891634613275528, + 0.014578164555132389, + 0.00882295798510313, + 0.010377112776041031, + 0.008127743378281593, + 0.007008646614849567, + 0.014198336750268936, + 0.007543859537690878, + 0.013564009219408035, + 0.021587038412690163, + 0.017039239406585693, + 0.01051884051412344, + 0.009515168145298958, + 0.011036870069801807, + 0.012484300881624222, + 0.009032776579260826, + 0.01448017731308937, + 0.011663768440485, + 0.010371701791882515, + 0.009134067222476006, + 0.022699110209941864, + 0.019502338021993637, + 0.016718175262212753, + 0.00799508299678564, + 0.01533101312816143, + 0.03798307478427887, + 0.01917266473174095, + 0.01236779522150755, + 0.04277578741312027, + 0.018149517476558685, + 0.007610501255840063, + 0.006044954061508179 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.020594336092472076 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635316178202629 + ], + "opposite_logprob_tokens": [ + -3.651353120803833 + ], + "logprob_reward": -0.02635316178202629, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635316178202629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19370630383491516, + 0.09894532710313797, + 0.1671217530965805, + 0.0694371685385704, + 0.03110360912978649, + 0.1698450893163681, + 0.09455212950706482, + 0.08546222001314163, + 0.027822773903608322, + 0.062003690749406815 + ], + "credit_map": [ + 0.0242797639220953, + 0.018962454050779343, + 0.03395898640155792, + 0.032985296100378036, + 0.02300773188471794, + 0.028019199147820473, + 0.015463738702237606, + 0.017029128968715668, + 0.014231902547180653, + 0.01382157951593399, + 0.0097579350695014, + 0.011267431080341339, + 0.023768026381731033, + 0.010213731788098812, + 0.008340063504874706, + 0.007544653490185738, + 0.04212089255452156, + 0.015790104866027832, + 0.015407504513859749, + 0.01155264861881733, + 0.013551740907132626, + 0.052900560200214386, + 0.0091033224016428, + 0.006694979500025511, + 0.010322210378944874, + 0.008588469587266445, + 0.01068816427141428, + 0.011436475440859795, + 0.00815147627145052, + 0.009606150910258293, + 0.004851882811635733, + 0.0057923332788050175, + 0.004301655106246471, + 0.006373032461851835, + 0.004354465752840042, + 0.004446657840162516, + 0.003440911415964365, + 0.0037055788561701775, + 0.002510416554287076, + 0.001970891375094652, + 0.02157895639538765, + 0.0454840213060379, + 0.01725720800459385, + 0.027152258902788162, + 0.015443669632077217, + 0.01749063841998577, + 0.014090342447161674, + 0.011347992345690727, + 0.019373765215277672, + 0.01822498068213463, + 0.009857162833213806, + 0.012048076838254929, + 0.008143785409629345, + 0.006963368970900774, + 0.013726412318646908, + 0.006214571185410023, + 0.010608192533254623, + 0.013041679747402668, + 0.015537015162408352, + 0.010645290836691856, + 0.009691819548606873, + 0.013235410675406456, + 0.007708607241511345, + 0.004994196817278862, + 0.0039015496149659157, + 0.00518164224922657, + 0.003627945203334093, + 0.0024672963190823793, + 0.004794009495526552, + 0.0034946741070598364, + 0.002734310692176223, + 0.0016213475028052926, + 0.01193858403712511, + 0.009559772908687592, + 0.014024462550878525, + 0.004390517249703407, + 0.009022233076393604, + 0.006262233480811119, + 0.003878181567415595, + 0.002927707973867655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.02635316178202629 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191186055541039 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014191186055541039, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191186055541039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11767076700925827, + 0.025374790653586388, + 0.10046276450157166, + 0.1086762472987175, + 0.14284087717533112, + 0.16845567524433136, + 0.1123119592666626, + 0.050127413123846054, + 0.09522423893213272, + 0.0788552314043045 + ], + "credit_map": [ + 0.015899591147899628, + 0.016210541129112244, + 0.01814236305654049, + 0.017393533140420914, + 0.013534453697502613, + 0.017481304705142975, + 0.009019734337925911, + 0.00998925045132637, + 0.00361130153760314, + 0.003789324313402176, + 0.003256805706769228, + 0.0027119326405227184, + 0.003424796275794506, + 0.0029529856983572245, + 0.00215742364525795, + 0.0034702231641858816, + 0.015941699966788292, + 0.017096340656280518, + 0.013631178997457027, + 0.011141074821352959, + 0.017809847369790077, + 0.011079641059041023, + 0.007953522726893425, + 0.005809468682855368, + 0.014565101824700832, + 0.026107124984264374, + 0.014961769804358482, + 0.01644328609108925, + 0.01031769160181284, + 0.01375666819512844, + 0.006299732718616724, + 0.006224885582923889, + 0.018247956410050392, + 0.039949096739292145, + 0.018483566120266914, + 0.02215544320642948, + 0.013201577588915825, + 0.010406582616269588, + 0.011010034941136837, + 0.009386621415615082, + 0.021213548257946968, + 0.045695770531892776, + 0.017054235562682152, + 0.027017327025532722, + 0.015394299291074276, + 0.017074385657906532, + 0.01407062541693449, + 0.010935511440038681, + 0.023057816550135612, + 0.021621987223625183, + 0.011045646853744984, + 0.014337238855659962, + 0.009278394281864166, + 0.007894380018115044, + 0.017839564010500908, + 0.0072369445115327835, + 0.00854230672121048, + 0.012246262282133102, + 0.006569715216755867, + 0.005501765292137861, + 0.005581535864621401, + 0.00364514347165823, + 0.004293488804250956, + 0.0037472022231668234, + 0.01095211785286665, + 0.010313314385712147, + 0.008234182372689247, + 0.007998484186828136, + 0.01883861981332302, + 0.016952892765402794, + 0.015452101826667786, + 0.006482538767158985, + 0.015014027245342731, + 0.01016115490347147, + 0.018567021936178207, + 0.006871079560369253, + 0.009165053255856037, + 0.008316234685480595, + 0.0057781413197517395, + 0.004982521291822195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.014191186055541039 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023274796083569527 + ], + "opposite_logprob_tokens": [ + -3.7732748985290527 + ], + "logprob_reward": -0.023274796083569527, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023274796083569527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11111637204885483, + 0.06171868368983269, + 0.1743995100259781, + 0.038498688489198685, + 0.05010337010025978, + 0.05468283221125603, + 0.029985137283802032, + 0.08058784902095795, + 0.058970458805561066, + 0.3399370610713959 + ], + "credit_map": [ + 0.016872983425855637, + 0.013722199946641922, + 0.014952978119254112, + 0.015758778899908066, + 0.013620033860206604, + 0.012462222017347813, + 0.014245899394154549, + 0.009481272660195827, + 0.010072280652821064, + 0.008842017501592636, + 0.009553556330502033, + 0.008668181486427784, + 0.006890562828630209, + 0.007063419092446566, + 0.005742619279772043, + 0.004886050708591938, + 0.036452215164899826, + 0.022954311221837997, + 0.01534931268543005, + 0.015533358789980412, + 0.016885526478290558, + 0.046283312141895294, + 0.010438259690999985, + 0.010503231547772884, + 0.004985562525689602, + 0.0073394267819821835, + 0.006010237615555525, + 0.005423787981271744, + 0.0032005293760448694, + 0.004900102503597736, + 0.0034074776340276003, + 0.0032315626740455627, + 0.006254193838685751, + 0.007689347490668297, + 0.006438631564378738, + 0.00564515870064497, + 0.004550972953438759, + 0.005322806071490049, + 0.008040196262300014, + 0.00616206182166934, + 0.006795551162213087, + 0.013266612775623798, + 0.006682588718831539, + 0.00846849661320448, + 0.004743171855807304, + 0.005241499748080969, + 0.005127871874719858, + 0.004357041325420141, + 0.005872388370335102, + 0.004110396374017, + 0.0034149442799389362, + 0.003472578711807728, + 0.002350657479837537, + 0.0022052237764000893, + 0.005550066940486431, + 0.0030088818166404963, + 0.01242542639374733, + 0.019739903509616852, + 0.008396833203732967, + 0.007473282981663942, + 0.009607277810573578, + 0.009326414205133915, + 0.006713773123919964, + 0.0069049387238919735, + 0.007340679410845041, + 0.011465349234640598, + 0.004808555357158184, + 0.005057482980191708, + 0.00897247064858675, + 0.01125658955425024, + 0.00635609682649374, + 0.0037132345605641603, + 0.038818079978227615, + 0.06462359428405762, + 0.052945125848054886, + 0.02652217447757721, + 0.08735178411006927, + 0.03733275458216667, + 0.017851289361715317, + 0.014492250047624111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.023274796083569527 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, captured from a slightly downward angle, captures a broad, vibrant green leaf, its surface dotted with countless shimmering dewdrops, each like a miniature crystal ball, refracting the soft morning light. On this microscopic battlefield, a mantis, its emerald green like a finely crafted jade, clashes with a sleek black spider. The mantis arches its body, its signature sickle-shaped forelegs raised high. Its triangular head tilts slightly, its compound eyes gleaming with a cold luster, a poised aggression. Meanwhile, the black spider, its opponent, crouches low, its eight long, finely haired legs stretched out nervously for stability, its fangs bared, poised for a deadly counterattack. As the mantis swung its front legs like lightning and the spider dodged to the side swiftly, a fierce and silent fight broke out in an instant. Every movement of theirs stirred the surrounding dewdrops, causing them to break or roll off the edge of the leaves. The whole picture was full of primitive, tense and gripping dramatic tension against the tranquil and beautiful background.", + "seed": 26666, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_2/index_274_seed_26666_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.029600246530026197, + "vlm_logprob": -0.029600246530026197, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.029600246530026197, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032309493981301785 + ], + "opposite_logprob_tokens": [ + -6.378231048583984 + ], + "logprob_reward": -0.0032309493981301785, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0032309493981301785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12133008986711502, + 0.2938711643218994, + 0.13134683668613434, + 0.19845741987228394, + 0.03630250692367554, + 0.02114928513765335, + 0.020508423447608948, + 0.06892599910497665, + 0.080706387758255, + 0.02740197442471981 + ], + "credit_map": [ + 0.009300817735493183, + 0.013217157684266567, + 0.01467385794967413, + 0.021248091012239456, + 0.02212069369852543, + 0.01409280113875866, + 0.015710284933447838, + 0.010966366156935692, + 0.018016645684838295, + 0.040879979729652405, + 0.03452663496136665, + 0.07589216530323029, + 0.04137827083468437, + 0.028124064207077026, + 0.03702002018690109, + 0.018033329397439957, + 0.015694845467805862, + 0.017186220735311508, + 0.014733514748513699, + 0.019038451835513115, + 0.016098864376544952, + 0.02230432815849781, + 0.01676097698509693, + 0.009529607370495796, + 0.0462748147547245, + 0.028927437961101532, + 0.009710517711937428, + 0.011277450248599052, + 0.07083530724048615, + 0.009177037514746189, + 0.012985152192413807, + 0.00926966592669487, + 0.004661970771849155, + 0.004104970023036003, + 0.004430594388395548, + 0.0052399770356714725, + 0.0050359321758151054, + 0.004905837122350931, + 0.004369520582258701, + 0.0035537010990083218, + 0.0022892355918884277, + 0.0017184435855597258, + 0.0024437066167593002, + 0.0025615638587623835, + 0.002628001384437084, + 0.0031037970911711454, + 0.004881978500634432, + 0.0015225535025820136, + 0.003396830754354596, + 0.0036037671379745007, + 0.0018867278704419732, + 0.0033311520237475634, + 0.0025243214331567287, + 0.0022591932211071253, + 0.0019907436799257994, + 0.0015156851150095463, + 0.008051943965256214, + 0.0150388702750206, + 0.007664750330150127, + 0.009939521551132202, + 0.00877553503960371, + 0.0073417779058218, + 0.008307945914566517, + 0.0038056413177400827, + 0.015384051017463207, + 0.011900315061211586, + 0.004070039372891188, + 0.005964554380625486, + 0.02325807325541973, + 0.010374783538281918, + 0.00579103035852313, + 0.0039635333232581615, + 0.0057422989048063755, + 0.0034409675281494856, + 0.0017906647408381104, + 0.002057754434645176, + 0.004602763801813126, + 0.004283440299332142, + 0.002598362509161234, + 0.00288571510463953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.0032309493981301785 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866829141974449 + ], + "opposite_logprob_tokens": [ + -3.048668384552002 + ], + "logprob_reward": -0.04866829141974449, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866829141974449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10663272440433502, + 0.34875062108039856, + 0.2536464035511017, + 0.0468587763607502, + 0.028684020042419434, + 0.027540316805243492, + 0.04117589443922043, + 0.06549351662397385, + 0.06130507215857506, + 0.01991264335811138 + ], + "credit_map": [ + 0.01076753344386816, + 0.010880587622523308, + 0.014510728418827057, + 0.021182024851441383, + 0.014816774055361748, + 0.010211877524852753, + 0.014027710072696209, + 0.010235493071377277, + 0.026962636038661003, + 0.0435841865837574, + 0.04167885705828667, + 0.09317488223314285, + 0.044696006923913956, + 0.03446843475103378, + 0.04177868366241455, + 0.022406941279768944, + 0.025480829179286957, + 0.034777067601680756, + 0.0383814238011837, + 0.029396135360002518, + 0.034949660301208496, + 0.042904168367385864, + 0.026479922235012054, + 0.021277179941534996, + 0.007150305435061455, + 0.004854430444538593, + 0.005711892154067755, + 0.006366749294102192, + 0.007095907814800739, + 0.0063721067272126675, + 0.006390062626451254, + 0.0029173255898058414, + 0.004209049046039581, + 0.002980036661028862, + 0.003545909421518445, + 0.004083714447915554, + 0.004073966294527054, + 0.0038722390308976173, + 0.003644505748525262, + 0.0022746010217815638, + 0.005312258843332529, + 0.0028042190242558718, + 0.0015308728907257318, + 0.0018642892828211188, + 0.007833912968635559, + 0.004374605603516102, + 0.0027196556329727173, + 0.0011005013948306441, + 0.006864794064313173, + 0.00386868161149323, + 0.003854063805192709, + 0.005284410435706377, + 0.00611691502854228, + 0.00499553419649601, + 0.0064018708653748035, + 0.003789621638134122, + 0.009046019986271858, + 0.005213804543018341, + 0.005519353784620762, + 0.00770501559600234, + 0.0072430032305419445, + 0.016786061227321625, + 0.008309146389365196, + 0.005671113263815641, + 0.011870572343468666, + 0.0073701441287994385, + 0.004093673545867205, + 0.004287957213819027, + 0.017465118318796158, + 0.00849200040102005, + 0.004484345205128193, + 0.00324126030318439, + 0.002516217064112425, + 0.0026596197858452797, + 0.002081359503790736, + 0.0041756462305784225, + 0.002607923001050949, + 0.002261148067191243, + 0.0018607497913762927, + 0.0017499793320894241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.04866829141974449 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054957229644060135 + ], + "opposite_logprob_tokens": [ + -2.929957151412964 + ], + "logprob_reward": -0.054957229644060135, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054957229644060135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12357538938522339, + 0.2989802658557892, + 0.09078432619571686, + 0.24341444671154022, + 0.02212480641901493, + 0.015379482880234718, + 0.04468917474150658, + 0.0170136746019125, + 0.12076420336961746, + 0.023274242877960205 + ], + "credit_map": [ + 0.008415114134550095, + 0.012093787081539631, + 0.016174843534827232, + 0.016940968111157417, + 0.017074504867196083, + 0.021001912653446198, + 0.02028634399175644, + 0.011587906628847122, + 0.018315356224775314, + 0.043269652873277664, + 0.0364946611225605, + 0.07552175968885422, + 0.04191834479570389, + 0.02827403135597706, + 0.037422046065330505, + 0.017764395102858543, + 0.009556164965033531, + 0.012203642167150974, + 0.010547840967774391, + 0.017239492386579514, + 0.009956013411283493, + 0.011992520652711391, + 0.013572262600064278, + 0.0057163783349096775, + 0.06494158506393433, + 0.032702989876270294, + 0.01110877376049757, + 0.009201690554618835, + 0.09015299379825592, + 0.011114458553493023, + 0.011277436278760433, + 0.01291449274867773, + 0.0024692693259567022, + 0.00261368858627975, + 0.0032263719476759434, + 0.004211680963635445, + 0.002370764734223485, + 0.002674940973520279, + 0.0025154524482786655, + 0.0020426353439688683, + 0.0019780551083385944, + 0.0016667766030877829, + 0.0017132206121459603, + 0.0017546884482726455, + 0.0028796761762350798, + 0.002646749373525381, + 0.0019977963529527187, + 0.0007425193325616419, + 0.007281097583472729, + 0.006339637096971273, + 0.003517551813274622, + 0.0065410565584897995, + 0.006314204074442387, + 0.005830575246363878, + 0.005924341268837452, + 0.0029407101683318615, + 0.0024546768981963396, + 0.002547126729041338, + 0.0017705804202705622, + 0.0024547353386878967, + 0.0017389992717653513, + 0.002980993827804923, + 0.0017214113613590598, + 0.0013451487757265568, + 0.02949351817369461, + 0.015756534412503242, + 0.005485702306032181, + 0.0067181517370045185, + 0.03840842470526695, + 0.01287178322672844, + 0.007621102966368198, + 0.0044089751318097115, + 0.0038119193632155657, + 0.0030463349539786577, + 0.0024986600037664175, + 0.005205472931265831, + 0.002811080776154995, + 0.0017645402112975717, + 0.0024923626333475113, + 0.0016438702587038279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.054957229644060135 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020574716851115227 + ], + "opposite_logprob_tokens": [ + -3.8955748081207275 + ], + "logprob_reward": -0.020574716851115227, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020574716851115227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13111712038516998, + 0.33446452021598816, + 0.07406690716743469, + 0.07956822216510773, + 0.07934941351413727, + 0.0420515313744545, + 0.03423203527927399, + 0.04660161957144737, + 0.12114611268043518, + 0.05740249529480934 + ], + "credit_map": [ + 0.009983892552554607, + 0.01604633964598179, + 0.016433196142315865, + 0.01969204843044281, + 0.023508647456765175, + 0.015926923602819443, + 0.017563844099640846, + 0.01196224708110094, + 0.0204691831022501, + 0.0478285551071167, + 0.040359292179346085, + 0.08139412105083466, + 0.05065469443798065, + 0.030644064769148827, + 0.0425708182156086, + 0.020543811842799187, + 0.009581656195223331, + 0.010876966640353203, + 0.008022777736186981, + 0.007051261141896248, + 0.010776786133646965, + 0.015235169790685177, + 0.006114424206316471, + 0.0064078732393682, + 0.01586076244711876, + 0.007688689045608044, + 0.008193033747375011, + 0.012768416665494442, + 0.009213581681251526, + 0.0052197822369635105, + 0.014695588499307632, + 0.005928376689553261, + 0.008870666846632957, + 0.010086942464113235, + 0.015348274260759354, + 0.01646418310701847, + 0.007985434494912624, + 0.006710862275213003, + 0.007970868609845638, + 0.005912185180932283, + 0.005204389337450266, + 0.0034136255271732807, + 0.005529963411390781, + 0.005034946370869875, + 0.00835286546498537, + 0.005869629792869091, + 0.0062720938585698605, + 0.00237402250058949, + 0.005591864697635174, + 0.004073233809322119, + 0.004127873107790947, + 0.0031802847515791655, + 0.004484820179641247, + 0.004320080392062664, + 0.005493498407304287, + 0.002960382727906108, + 0.006034155376255512, + 0.011825798079371452, + 0.00490079028531909, + 0.007220085710287094, + 0.0048354133032262325, + 0.004710318986326456, + 0.004660579841583967, + 0.0024144863709807396, + 0.02943938970565796, + 0.013403608463704586, + 0.004724733531475067, + 0.005978340748697519, + 0.04148539900779724, + 0.015219619497656822, + 0.0068247150629758835, + 0.004070315044373274, + 0.009588954970240593, + 0.006724497303366661, + 0.005369796883314848, + 0.010762094520032406, + 0.007355721667408943, + 0.005432127509266138, + 0.006479276344180107, + 0.005690033081918955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.020574716851115227 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020570045337080956 + ], + "opposite_logprob_tokens": [ + -3.8955700397491455 + ], + "logprob_reward": -0.020570045337080956, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020570045337080956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1911599487066269, + 0.08508738875389099, + 0.16775567829608917, + 0.1905241310596466, + 0.0427556112408638, + 0.018890894949436188, + 0.04392416402697563, + 0.058354850858449936, + 0.1291247010231018, + 0.07242266833782196 + ], + "credit_map": [ + 0.012178474105894566, + 0.02745802514255047, + 0.01722581312060356, + 0.028067056089639664, + 0.027858804911375046, + 0.024363216012716293, + 0.03393097221851349, + 0.020077597349882126, + 0.006517319940030575, + 0.014300629496574402, + 0.011818375438451767, + 0.017368517816066742, + 0.011448259465396404, + 0.009205371141433716, + 0.009022586978971958, + 0.005406330805271864, + 0.01605997420847416, + 0.03702208027243614, + 0.05711936578154564, + 0.013246668502688408, + 0.0155754704028368, + 0.011778377927839756, + 0.01087948027998209, + 0.006074258591979742, + 0.054868847131729126, + 0.023771343752741814, + 0.00900046993046999, + 0.0063603222370147705, + 0.0697367936372757, + 0.008457385934889317, + 0.007955385372042656, + 0.010373583063483238, + 0.008400818333029747, + 0.0046224226243793964, + 0.0050271665677428246, + 0.005945964716374874, + 0.005288589280098677, + 0.004271749407052994, + 0.005736109800636768, + 0.003462790744379163, + 0.002739077666774392, + 0.0021203525830060244, + 0.003278224729001522, + 0.00227718660607934, + 0.0023570354096591473, + 0.0024529423099011183, + 0.0026114340871572495, + 0.0010546408593654633, + 0.007384819444268942, + 0.0039507122710347176, + 0.004558307118713856, + 0.005995459388941526, + 0.005557222757488489, + 0.006438843905925751, + 0.005926686804741621, + 0.004112110007554293, + 0.006564815528690815, + 0.007681692019104958, + 0.008164763450622559, + 0.008702349849045277, + 0.006122783292084932, + 0.010137969627976418, + 0.007281644269824028, + 0.0036988311912864447, + 0.01920703984797001, + 0.02084544487297535, + 0.007765109650790691, + 0.01079500187188387, + 0.030232232064008713, + 0.019846739247441292, + 0.014468724839389324, + 0.005964408628642559, + 0.009379668161273003, + 0.008422914892435074, + 0.0065009621903300285, + 0.013216583989560604, + 0.011146784760057926, + 0.008909640833735466, + 0.008872343227267265, + 0.005973774939775467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.020570045337080956 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, captured from a slightly downward angle, captures a broad, vibrant green leaf, its surface dotted with countless shimmering dewdrops, each like a miniature crystal ball, refracting the soft morning light. On this microscopic battlefield, a mantis, its emerald green like a finely crafted jade, clashes with a sleek black spider. The mantis arches its body, its signature sickle-shaped forelegs raised high. Its triangular head tilts slightly, its compound eyes gleaming with a cold luster, a poised aggression. Meanwhile, the black spider, its opponent, crouches low, its eight long, finely haired legs stretched out nervously for stability, its fangs bared, poised for a deadly counterattack. As the mantis swung its front legs like lightning and the spider dodged to the side swiftly, a fierce and silent fight broke out in an instant. Every movement of theirs stirred the surrounding dewdrops, causing them to break or roll off the edge of the leaves. The whole picture was full of primitive, tense and gripping dramatic tension against the tranquil and beautiful background.", + "seed": 26667, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_2/index_274_seed_26667_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0096562837716192, + "vlm_logprob": -0.0096562837716192, + "vlm_yes_no_margin": 5.275, + "vlm_token_credit": -0.0096562837716192, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0020524642895907164 + ], + "opposite_logprob_tokens": [ + -8.002052307128906 + ], + "logprob_reward": -0.0020524642895907164, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.0020524642895907164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19975945353507996, + 0.04706322029232979, + 0.10109703987836838, + 0.055590011179447174, + 0.1442233771085739, + 0.09205619990825653, + 0.10545023530721664, + 0.0905558243393898, + 0.07658258825540543, + 0.08762209117412567 + ], + "credit_map": [ + 0.01957288570702076, + 0.027125271037220955, + 0.024255162104964256, + 0.02608550898730755, + 0.019041651859879494, + 0.02849055826663971, + 0.024192385375499725, + 0.030996020883321762, + 0.005077785812318325, + 0.004577590152621269, + 0.006728758104145527, + 0.007280631922185421, + 0.004800776485353708, + 0.006896795704960823, + 0.005047825165092945, + 0.006653055548667908, + 0.008992972783744335, + 0.01361135020852089, + 0.019913803786039352, + 0.012249491177499294, + 0.01290325541049242, + 0.013501918874680996, + 0.010646083392202854, + 0.009278160519897938, + 0.006995528936386108, + 0.005536057986319065, + 0.0065798102878034115, + 0.006589876487851143, + 0.007010804954916239, + 0.0065831346437335014, + 0.008996959775686264, + 0.007297836244106293, + 0.02104293368756771, + 0.012271303683519363, + 0.020455529913306236, + 0.022524844855070114, + 0.012206705287098885, + 0.029088495299220085, + 0.012511909939348698, + 0.014121645130217075, + 0.01101509016007185, + 0.010893123224377632, + 0.03044992685317993, + 0.012657405808568, + 0.0086998101323843, + 0.006631600204855204, + 0.005827469751238823, + 0.005881764460355043, + 0.01425021979957819, + 0.021279901266098022, + 0.020483283326029778, + 0.011040259152650833, + 0.010277507826685905, + 0.011235387064516544, + 0.008055131882429123, + 0.008828535676002502, + 0.012575230561196804, + 0.008201759308576584, + 0.017052914947271347, + 0.015910575166344643, + 0.009894212707877159, + 0.00742570124566555, + 0.008497516624629498, + 0.010997919365763664, + 0.014645409770309925, + 0.006734731141477823, + 0.009388112463057041, + 0.011953246779739857, + 0.006703396327793598, + 0.011367716826498508, + 0.008566586300730705, + 0.00722338305786252, + 0.009645313955843449, + 0.013700929470360279, + 0.017764784395694733, + 0.01285072322934866, + 0.00875666830688715, + 0.007640365045517683, + 0.008175626397132874, + 0.009087682701647282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0020524642895907164 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607411541044712 + ], + "opposite_logprob_tokens": [ + -4.141074180603027 + ], + "logprob_reward": -0.01607411541044712, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607411541044712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14984562993049622, + 0.15459120273590088, + 0.18859264254570007, + 0.029570074751973152, + 0.09726925939321518, + 0.06736069172620773, + 0.05095919594168663, + 0.11692898720502853, + 0.04944029450416565, + 0.09544199705123901 + ], + "credit_map": [ + 0.01881544291973114, + 0.021240921691060066, + 0.018525924533605576, + 0.020640168339014053, + 0.015939395874738693, + 0.01772898994386196, + 0.016403216868638992, + 0.020551573485136032, + 0.01584138721227646, + 0.012559638358652592, + 0.034963712096214294, + 0.021658454090356827, + 0.013122008182108402, + 0.018682710826396942, + 0.020686745643615723, + 0.017076542600989342, + 0.018530482426285744, + 0.023742396384477615, + 0.03764281049370766, + 0.022242652252316475, + 0.02266784943640232, + 0.024610629305243492, + 0.02173808217048645, + 0.01741773821413517, + 0.0031669726595282555, + 0.0025639990344643593, + 0.007366806268692017, + 0.0043509891256690025, + 0.002755983266979456, + 0.0030731959268450737, + 0.0035846447572112083, + 0.0027074823155999184, + 0.012293449603021145, + 0.008459474891424179, + 0.012458551675081253, + 0.012993362732231617, + 0.010858407244086266, + 0.018201908096671104, + 0.010677680373191833, + 0.01132642850279808, + 0.008422131650149822, + 0.010707873851060867, + 0.008262595161795616, + 0.0072772991843521595, + 0.008254450745880604, + 0.008383608423173428, + 0.0047828154638409615, + 0.01126991305500269, + 0.007348983082920313, + 0.009200522676110268, + 0.0070874495431780815, + 0.00613287603482604, + 0.005923284683376551, + 0.006422826554626226, + 0.0050680674612522125, + 0.0037751849740743637, + 0.01472924742847681, + 0.01936897076666355, + 0.02318720892071724, + 0.019267454743385315, + 0.008896374143660069, + 0.009285212494432926, + 0.009536901488900185, + 0.012657626532018185, + 0.007731885649263859, + 0.005641838535666466, + 0.0063712154515087605, + 0.0066442945972085, + 0.0046779136173427105, + 0.004869557451456785, + 0.008555076085031033, + 0.004948512185364962, + 0.010609778575599194, + 0.009806114248931408, + 0.013136893510818481, + 0.014270900748670101, + 0.010856082662940025, + 0.015724262222647667, + 0.011184295639395714, + 0.009853668510913849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.01607411541044712 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252490285784006 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252490285784006, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252490285784006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15940254926681519, + 0.05789335444569588, + 0.282115638256073, + 0.10143157094717026, + 0.1008257046341896, + 0.08784546703100204, + 0.05005693435668945, + 0.07009060680866241, + 0.0601593516767025, + 0.030178826302289963 + ], + "credit_map": [ + 0.01636783964931965, + 0.023565340787172318, + 0.017042864114046097, + 0.02784597873687744, + 0.01484860572963953, + 0.02381860837340355, + 0.015420794486999512, + 0.020492512732744217, + 0.006717382930219173, + 0.005525786429643631, + 0.01178823597729206, + 0.008842716924846172, + 0.0053616599179804325, + 0.006226292811334133, + 0.008438040502369404, + 0.004993240348994732, + 0.028238456696271896, + 0.03751557320356369, + 0.05129450932145119, + 0.032973192632198334, + 0.03676198422908783, + 0.03783809393644333, + 0.032958026975393295, + 0.024535799399018288, + 0.01233658753335476, + 0.013158208690583706, + 0.018684744834899902, + 0.011602465994656086, + 0.010639307089149952, + 0.010723089799284935, + 0.013590414077043533, + 0.010696751065552235, + 0.013885737396776676, + 0.008377555757761002, + 0.012842767871916294, + 0.013236288912594318, + 0.009832416661083698, + 0.02146787755191326, + 0.008987047709524632, + 0.012196008116006851, + 0.010262131690979004, + 0.014849275350570679, + 0.007571217603981495, + 0.009352938272058964, + 0.012168965302407742, + 0.01169400941580534, + 0.00750228576362133, + 0.014444643631577492, + 0.007012915797531605, + 0.0060628438368439674, + 0.010030131787061691, + 0.00573433144018054, + 0.004662242718040943, + 0.006592257414013147, + 0.005594058893620968, + 0.004368148744106293, + 0.008749735541641712, + 0.011276697739958763, + 0.010982095263898373, + 0.0097639299929142, + 0.008091757073998451, + 0.007512403652071953, + 0.006939399056136608, + 0.00677459454163909, + 0.009444959461688995, + 0.008660219609737396, + 0.008081385865807533, + 0.005988605320453644, + 0.007466703187674284, + 0.006622188724577427, + 0.007923838682472706, + 0.005971451289951801, + 0.002906452864408493, + 0.0035081100650131702, + 0.003640316892415285, + 0.003601876785978675, + 0.005911493208259344, + 0.0029751972761005163, + 0.0033502797596156597, + 0.004285099450498819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.005252490285784006 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006729676853865385 + ], + "opposite_logprob_tokens": [ + -5.006729602813721 + ], + "logprob_reward": -0.006729676853865385, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006729676853865385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09804656356573105, + 0.16498975455760956, + 0.1329849362373352, + 0.04658820852637291, + 0.11342667788267136, + 0.11754590272903442, + 0.14005336165428162, + 0.06844563782215118, + 0.08227969706058502, + 0.03563926741480827 + ], + "credit_map": [ + 0.011107105761766434, + 0.01228302251547575, + 0.010774681344628334, + 0.01302631851285696, + 0.010415383614599705, + 0.0199526809155941, + 0.007803086657077074, + 0.012684278190135956, + 0.018119385465979576, + 0.013661161065101624, + 0.035888370126485825, + 0.02411729097366333, + 0.014080020599067211, + 0.0187043659389019, + 0.02325599081814289, + 0.01716315932571888, + 0.012886574491858482, + 0.017336038872599602, + 0.02586917206645012, + 0.016003891825675964, + 0.016774162650108337, + 0.01736919768154621, + 0.01485762745141983, + 0.01188826747238636, + 0.00562645960599184, + 0.007100878283381462, + 0.009266175329685211, + 0.005353691056370735, + 0.0044378722086548805, + 0.0053265695460140705, + 0.004433650057762861, + 0.0050429124385118484, + 0.016442282125353813, + 0.009720729663968086, + 0.0183152724057436, + 0.014877976849675179, + 0.010555670596659184, + 0.019572611898183823, + 0.01169036328792572, + 0.012251781299710274, + 0.014008572325110435, + 0.025307053700089455, + 0.00922661554068327, + 0.010028272867202759, + 0.01910526305437088, + 0.014388962648808956, + 0.012060132808983326, + 0.013421024195849895, + 0.022912627086043358, + 0.017755303531885147, + 0.04051532596349716, + 0.015735382214188576, + 0.009691555984318256, + 0.012836030684411526, + 0.01121613197028637, + 0.009391008876264095, + 0.007893777452409267, + 0.007750998251140118, + 0.011013377457857132, + 0.009554781951010227, + 0.009566658176481724, + 0.005978122819215059, + 0.007758885156363249, + 0.008929042145609856, + 0.013526631519198418, + 0.013052616268396378, + 0.011083904653787613, + 0.007545182015746832, + 0.007703863084316254, + 0.010037761181592941, + 0.012333613820374012, + 0.006996122654527426, + 0.003717864630743861, + 0.005521815270185471, + 0.003876155475154519, + 0.004274306818842888, + 0.006129231303930283, + 0.0034557355102151632, + 0.003844687482342124, + 0.004819469526410103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.006729676853865385 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018172672018408775 + ], + "opposite_logprob_tokens": [ + -4.018172740936279 + ], + "logprob_reward": -0.018172672018408775, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018172672018408775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22584061324596405, + 0.10792327672243118, + 0.07454617321491241, + 0.04446651414036751, + 0.14549653232097626, + 0.06484714150428772, + 0.06046358123421669, + 0.12356209754943848, + 0.1110944002866745, + 0.04175961762666702 + ], + "credit_map": [ + 0.023750875145196915, + 0.028018295764923096, + 0.017006080597639084, + 0.037657447159290314, + 0.019176742061972618, + 0.03776540234684944, + 0.024391358718276024, + 0.03807441517710686, + 0.010744476690888405, + 0.009374282322824001, + 0.02276674471795559, + 0.01498552318662405, + 0.009487518109381199, + 0.014322332106530666, + 0.015011767856776714, + 0.011230643838644028, + 0.007480477914214134, + 0.009178053587675095, + 0.015003073029220104, + 0.008337805978953838, + 0.010776075534522533, + 0.008335141465067863, + 0.008034337311983109, + 0.007401216775178909, + 0.005189932882785797, + 0.006438691169023514, + 0.007019807118922472, + 0.005095264408737421, + 0.004320808686316013, + 0.005035790614783764, + 0.005559325683861971, + 0.0058068991638720036, + 0.020865099504590034, + 0.01302366890013218, + 0.019955560564994812, + 0.0210405420511961, + 0.01255387905985117, + 0.02849254384636879, + 0.01336654368788004, + 0.01619870774447918, + 0.007347746752202511, + 0.01263224333524704, + 0.00990796647965908, + 0.008873735554516315, + 0.0072759282775223255, + 0.006484170909970999, + 0.0058520762249827385, + 0.006473281420767307, + 0.008689113892614841, + 0.0058501348830759525, + 0.016989994794130325, + 0.0070766881108284, + 0.005287508480250835, + 0.005974900908768177, + 0.005555028561502695, + 0.005040217190980911, + 0.016615187749266624, + 0.02049572952091694, + 0.017269866541028023, + 0.024112412706017494, + 0.010764839127659798, + 0.009714534506201744, + 0.00959797017276287, + 0.014991565607488155, + 0.012977582402527332, + 0.01810518279671669, + 0.014806358143687248, + 0.01246557291597128, + 0.010198148898780346, + 0.015240264125168324, + 0.016878362745046616, + 0.010422947816550732, + 0.0038477873895317316, + 0.005712420213967562, + 0.006566659547388554, + 0.005090812686830759, + 0.005432418081909418, + 0.003954112064093351, + 0.004793840926140547, + 0.006361570209264755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.018172672018408775 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, captured from a slightly downward angle, captures a broad, vibrant green leaf, its surface dotted with countless shimmering dewdrops, each like a miniature crystal ball, refracting the soft morning light. On this microscopic battlefield, a mantis, its emerald green like a finely crafted jade, clashes with a sleek black spider. The mantis arches its body, its signature sickle-shaped forelegs raised high. Its triangular head tilts slightly, its compound eyes gleaming with a cold luster, a poised aggression. Meanwhile, the black spider, its opponent, crouches low, its eight long, finely haired legs stretched out nervously for stability, its fangs bared, poised for a deadly counterattack. As the mantis swung its front legs like lightning and the spider dodged to the side swiftly, a fierce and silent fight broke out in an instant. Every movement of theirs stirred the surrounding dewdrops, causing them to break or roll off the edge of the leaves. The whole picture was full of primitive, tense and gripping dramatic tension against the tranquil and beautiful background.", + "seed": 26668, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_2/index_274_seed_26668_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.026913867704570292, + "vlm_logprob": -0.026913867704570292, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.026913867704570292, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0037188436836004257 + ], + "opposite_logprob_tokens": [ + -6.503718852996826 + ], + "logprob_reward": -0.0037188436836004257, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0037188436836004257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21972091495990753, + 0.35154464840888977, + 0.0718461126089096, + 0.03574578836560249, + 0.034716587513685226, + 0.061727866530418396, + 0.0724126473069191, + 0.05887845158576965, + 0.042347345501184464, + 0.05105965584516525 + ], + "credit_map": [ + 0.026134684681892395, + 0.012678989209234715, + 0.014282089658081532, + 0.0817352756857872, + 0.03953549265861511, + 0.01818128675222397, + 0.014727898873388767, + 0.012445176020264626, + 0.03484697639942169, + 0.014906847849488258, + 0.025154421105980873, + 0.016306009143590927, + 0.017633533105254173, + 0.09616021811962128, + 0.1288013607263565, + 0.017735272645950317, + 0.007279190234839916, + 0.013228757306933403, + 0.00849179271608591, + 0.011450227349996567, + 0.006405392661690712, + 0.009194537997245789, + 0.007618327159434557, + 0.008177884854376316, + 0.0031226526480168104, + 0.003920931369066238, + 0.004262630827724934, + 0.006134311202913523, + 0.004009283147752285, + 0.004855587612837553, + 0.005295388866215944, + 0.004145004786550999, + 0.0036839863751083612, + 0.0048638200387358665, + 0.0037378803826868534, + 0.00490179006010294, + 0.005715635605156422, + 0.004390093497931957, + 0.0032096481882035732, + 0.0042137326672673225, + 0.006279991008341312, + 0.007009950932115316, + 0.008631140924990177, + 0.012263007462024689, + 0.010223694145679474, + 0.006485428661108017, + 0.0067004854790866375, + 0.004134165123105049, + 0.010521416552364826, + 0.006524904165416956, + 0.016963129863142967, + 0.00541581679135561, + 0.0062240781262516975, + 0.01373289618641138, + 0.005726156290620565, + 0.007304244674742222, + 0.0060714357532560825, + 0.005422879010438919, + 0.009909669868648052, + 0.005733458325266838, + 0.011975185945630074, + 0.00803042110055685, + 0.006543358787894249, + 0.005192040465772152, + 0.004546110052615404, + 0.003748851828277111, + 0.0056998576037585735, + 0.0049758595414459705, + 0.006716424133628607, + 0.005580260418355465, + 0.006057289894670248, + 0.005022689700126648, + 0.00530213164165616, + 0.007333164103329182, + 0.005917518865317106, + 0.008319156244397163, + 0.007490118965506554, + 0.005765976384282112, + 0.0064679537899792194, + 0.004463633522391319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0037188436836004257 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018205681815743446 + ], + "opposite_logprob_tokens": [ + -4.018205642700195 + ], + "logprob_reward": -0.018205681815743446, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018205681815743446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08302321285009384, + 0.4446961283683777, + 0.07157987356185913, + 0.04690027981996536, + 0.0339491069316864, + 0.06183074787259102, + 0.05049285665154457, + 0.043809596449136734, + 0.07281511276960373, + 0.09090306609869003 + ], + "credit_map": [ + 0.00948929414153099, + 0.007954828441143036, + 0.008612669073045254, + 0.017110519111156464, + 0.014907032251358032, + 0.009546726010739803, + 0.006712776608765125, + 0.008689366281032562, + 0.04480302333831787, + 0.020035026594996452, + 0.027419794350862503, + 0.02346670813858509, + 0.022519493475556374, + 0.12688277661800385, + 0.15369974076747894, + 0.025869587436318398, + 0.007812102325260639, + 0.013478710316121578, + 0.011129874736070633, + 0.012159769423305988, + 0.005836724769324064, + 0.008823408745229244, + 0.006575682200491428, + 0.005763597786426544, + 0.0039002476260066032, + 0.00476593105122447, + 0.006711013615131378, + 0.0067771319299936295, + 0.004425246734172106, + 0.005582929588854313, + 0.00865847896784544, + 0.006079297047108412, + 0.003081924980506301, + 0.004026927053928375, + 0.0036811090540140867, + 0.0037066794466227293, + 0.004551619291305542, + 0.006457065232098103, + 0.002938486635684967, + 0.005505293607711792, + 0.006237061228603125, + 0.007492390461266041, + 0.00783631019294262, + 0.011947398073971272, + 0.011669077910482883, + 0.0063785770907998085, + 0.006193775683641434, + 0.004076154436916113, + 0.007258537691086531, + 0.0053071086294949055, + 0.010290397331118584, + 0.0041680834256112576, + 0.006908687762916088, + 0.008389590308070183, + 0.004330940544605255, + 0.0038395081646740437, + 0.005075774155557156, + 0.0031963486690074205, + 0.006793370470404625, + 0.00563031155616045, + 0.006985601037740707, + 0.005288723856210709, + 0.007834617048501968, + 0.0030048491898924112, + 0.006870305631309748, + 0.004752753302454948, + 0.01733211986720562, + 0.009001745842397213, + 0.007081381510943174, + 0.009824135340750217, + 0.010584690608084202, + 0.007367982063442469, + 0.009109453298151493, + 0.0068247877061367035, + 0.024475222453475, + 0.009631019085645676, + 0.00767446868121624, + 0.013596399687230587, + 0.01073701586574316, + 0.008854704909026623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.018205681815743446 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903989940881729 + ], + "opposite_logprob_tokens": [ + -2.4640398025512695 + ], + "logprob_reward": -0.08903989940881729, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903989940881729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2723870277404785, + 0.08515772968530655, + 0.11434529721736908, + 0.039385437965393066, + 0.10753414034843445, + 0.1164056733250618, + 0.09027719497680664, + 0.06628350913524628, + 0.053863346576690674, + 0.05436059087514877 + ], + "credit_map": [ + 0.04130317643284798, + 0.020495479926466942, + 0.015383612364530563, + 0.10495558381080627, + 0.026558220386505127, + 0.0232474934309721, + 0.022245805710554123, + 0.018197650089859962, + 0.009395912289619446, + 0.008900629356503487, + 0.011952013708651066, + 0.007752617821097374, + 0.0042623537592589855, + 0.014985468238592148, + 0.02211589738726616, + 0.005792840383946896, + 0.013132736086845398, + 0.023790758103132248, + 0.01467505656182766, + 0.02132841758430004, + 0.012561922892928123, + 0.011091581545770168, + 0.009179068729281425, + 0.008585752919316292, + 0.0036982691381126642, + 0.0031074851285666227, + 0.006451559253036976, + 0.004933364223688841, + 0.004942160099744797, + 0.0042597088031470776, + 0.005676112603396177, + 0.006316780112683773, + 0.009886463172733784, + 0.011153665371239185, + 0.012524972669780254, + 0.013460018672049046, + 0.015098134987056255, + 0.020827312022447586, + 0.009848558343946934, + 0.014735023491084576, + 0.011690189130604267, + 0.013825316913425922, + 0.015554815530776978, + 0.0256073996424675, + 0.02101869136095047, + 0.011477620340883732, + 0.011246579699218273, + 0.005985069088637829, + 0.014103518798947334, + 0.00924138817936182, + 0.010369619354605675, + 0.007109224796295166, + 0.016026996076107025, + 0.019351810216903687, + 0.00762931490316987, + 0.006445326842367649, + 0.00811932422220707, + 0.0065316371619701385, + 0.008210926316678524, + 0.008909213356673717, + 0.013687586411833763, + 0.007909133099019527, + 0.008321214467287064, + 0.004594468977302313, + 0.0072427475824952126, + 0.004001370631158352, + 0.0067683020606637, + 0.005675797816365957, + 0.009387724101543427, + 0.00696635851636529, + 0.006731091067194939, + 0.007089952006936073, + 0.005846298765391111, + 0.007294908631592989, + 0.005507464986294508, + 0.009095905348658562, + 0.009252955205738544, + 0.004904588218778372, + 0.008281293325126171, + 0.004177176859229803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.08903989940881729 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073444969952106 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073444969952106, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073444969952106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16399656236171722, + 0.19842512905597687, + 0.0839918777346611, + 0.03718276694417, + 0.04376489669084549, + 0.07510969787836075, + 0.15714573860168457, + 0.08542869240045547, + 0.06772902607917786, + 0.08722561597824097 + ], + "credit_map": [ + 0.022071365267038345, + 0.011350533924996853, + 0.01221910584717989, + 0.05711440369486809, + 0.017643427476286888, + 0.017049405723810196, + 0.0131742088124156, + 0.01337411068379879, + 0.020511271432042122, + 0.012770820409059525, + 0.017383893951773643, + 0.011308572255074978, + 0.009452787227928638, + 0.05044298619031906, + 0.06735698133707047, + 0.009197840467095375, + 0.009353588335216045, + 0.016524096950888634, + 0.011968056671321392, + 0.018578918650746346, + 0.008126236498355865, + 0.008677536621689796, + 0.006094421725720167, + 0.004669027402997017, + 0.0031709277536720037, + 0.0038006179966032505, + 0.005108265206217766, + 0.005938748363405466, + 0.0046117571182549, + 0.004917547572404146, + 0.00433333357796073, + 0.005301572848111391, + 0.004082942381501198, + 0.006835601292550564, + 0.00523059768602252, + 0.0056389872916042805, + 0.007559382822364569, + 0.006815742235630751, + 0.0035743180196732283, + 0.00402732752263546, + 0.007227438502013683, + 0.009453277103602886, + 0.011344186961650848, + 0.014311162754893303, + 0.013473013415932655, + 0.007569740526378155, + 0.007315243128687143, + 0.0044156345538794994, + 0.024274487048387527, + 0.014892762526869774, + 0.024561811238527298, + 0.01079822238534689, + 0.02082998864352703, + 0.031336914747953415, + 0.015830378979444504, + 0.014621165581047535, + 0.009981379844248295, + 0.008270950987935066, + 0.011062226258218288, + 0.010078581050038338, + 0.017723025754094124, + 0.010687511414289474, + 0.01132951769977808, + 0.006295500323176384, + 0.009356874972581863, + 0.005158509127795696, + 0.008384951390326023, + 0.007531791925430298, + 0.012259148061275482, + 0.008755778893828392, + 0.007699152454733849, + 0.008582827635109425, + 0.01011879462748766, + 0.012655878439545631, + 0.009013917297124863, + 0.011633843183517456, + 0.014341735281050205, + 0.008911321870982647, + 0.013252663426101208, + 0.007297459989786148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.011073444969952106 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a broad, vibrant green leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless shimmering dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an emerald green mantis with raised sickle-shaped forelegs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sleek black spider with long hairy legs and bared fangs visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the scene from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531468644738197 + ], + "opposite_logprob_tokens": [ + -4.387531280517578 + ], + "logprob_reward": -0.012531468644738197, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531468644738197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28777435421943665, + 0.17086739838123322, + 0.06827420741319656, + 0.07729529589414597, + 0.10353703796863556, + 0.05559036508202553, + 0.10476471483707428, + 0.03953659534454346, + 0.04949505627155304, + 0.042865023016929626 + ], + "credit_map": [ + 0.04322196543216705, + 0.02044314332306385, + 0.04043053463101387, + 0.05856052041053772, + 0.07990024983882904, + 0.017596440389752388, + 0.015615992248058319, + 0.012005499564111233, + 0.019336199387907982, + 0.010587179102003574, + 0.013292969204485416, + 0.013414151966571808, + 0.010550695471465588, + 0.04174372926354408, + 0.05098498612642288, + 0.010957479476928711, + 0.008159535005688667, + 0.01283900998532772, + 0.010803323239088058, + 0.009815288707613945, + 0.005841746926307678, + 0.006561805959790945, + 0.008884814567863941, + 0.005368676036596298, + 0.008029607124626637, + 0.008391137234866619, + 0.009112023748457432, + 0.009933462366461754, + 0.010107644833624363, + 0.008399583399295807, + 0.012671256437897682, + 0.010650575160980225, + 0.011110099032521248, + 0.012112577445805073, + 0.012975971214473248, + 0.011436695232987404, + 0.010841972194612026, + 0.012794884853065014, + 0.011608278378844261, + 0.02065654657781124, + 0.006875005550682545, + 0.00600692443549633, + 0.0073383827693760395, + 0.008083099499344826, + 0.006924833636730909, + 0.009417425841093063, + 0.006311485078185797, + 0.004633201751857996, + 0.013809578493237495, + 0.0093085877597332, + 0.012484691105782986, + 0.008232506923377514, + 0.017409328371286392, + 0.022611131891608238, + 0.010742883197963238, + 0.010166001506149769, + 0.004036335740238428, + 0.003830038709565997, + 0.004506277851760387, + 0.00436145206913352, + 0.010244589298963547, + 0.005253114737570286, + 0.0032867162954062223, + 0.004018064588308334, + 0.006729430053383112, + 0.004428759682923555, + 0.00993847381323576, + 0.005458188243210316, + 0.007094003725796938, + 0.00627577118575573, + 0.005661525763571262, + 0.003908896818757057, + 0.00516017060726881, + 0.0035177215468138456, + 0.00958275981247425, + 0.0032942078541964293, + 0.0038891537114977837, + 0.006726766005158424, + 0.005359996575862169, + 0.005334243178367615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.012531468644738197 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_3/index_tensor([6033, 6033])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_3/index_tensor([6033, 6033])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..123565a3133bb298dcb20dce21a263225a64dcf5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_3/index_tensor([6033, 6033])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the style of high-speed sports photography, dramatic low-angle propeller shots capture a cascade of vibrant objects hurled vertically upward, their ascent rendered in extreme slow motion. The camera rotates and tilts, carefully following a mix of ornate red envelopes and glossy coupons as they flutter with athletic grace against a stark, uniformly lit background. The red envelopes, made of richly textured crimson paper embossed with an intricate, shimmering gold foil pattern, catch the light as they tumble. The coupons, dotted among them, their high-gloss surface shimmering with brilliant mirrored highlights, reveal vibrant colors and bold, abstract graphic elements. Under powerful, shadowless studio high-key lighting, every detail is sharply focused. Each item performs a unique aerial ballet: some envelopes bend and flex as they pass through the air, while the stiffer coupons spin on their axes like precisely thrown discs, creating a mesmerizing, frozen moment of celebratory commerce that feels both chaotic and elegant.", + "seed": 23005, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_3/index_6033_seed_23005_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08703191382810474, + "vlm_logprob": -0.08703191382810474, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.08703191382810474, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009888590313494205 + ], + "opposite_logprob_tokens": [ + -5.384888648986816 + ], + "logprob_reward": -0.009888590313494205, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.009888590313494205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023370584473013878, + 0.006704340688884258, + 0.0032235633116215467, + 0.010912157595157623, + 0.008921520784497261, + 0.9025992155075073, + 0.010406630113720894, + 0.010778414085507393, + 0.009539243765175343, + 0.013544355519115925 + ], + "credit_map": [ + 0.0047559491358697414, + 0.0010204752907156944, + 0.0030697500333189964, + 0.003838582430034876, + 0.0019600288942456245, + 0.002509548794478178, + 0.003402173751965165, + 0.002814072882756591, + 0.0010704059386625886, + 0.0006039187428541481, + 0.0008726170053705573, + 0.0011194593971595168, + 0.00044880862697027624, + 0.0008190061780624092, + 0.0006363900611177087, + 0.00113373389467597, + 0.0003123777569271624, + 0.0005496700759977102, + 0.0005304217920638621, + 0.0005477869999594986, + 0.0002943783765658736, + 0.0003537210577633232, + 0.00030077798874117434, + 0.00033442926360294223, + 0.0012624249793589115, + 0.0023039118386805058, + 0.0017105552833527327, + 0.0009844658197835088, + 0.001376873697154224, + 0.0009519204031676054, + 0.0013419525930657983, + 0.0009800518164411187, + 0.001260381075553596, + 0.001242922618985176, + 0.0006921260501258075, + 0.0007931497530080378, + 0.0011860039085149765, + 0.0013086484977975488, + 0.0008084037690423429, + 0.0016298842383548617, + 0.017699992284178734, + 0.03555234521627426, + 0.00966901145875454, + 0.007135380990803242, + 0.018329765647649765, + 0.4139203131198883, + 0.010525858961045742, + 0.3897664248943329, + 0.0012061841553077102, + 0.0011792936129495502, + 0.0008285442600026727, + 0.0011252115946263075, + 0.002031217562034726, + 0.0015139419119805098, + 0.0011346786050125957, + 0.001387557596899569, + 0.0009974705753847957, + 0.0008860578527674079, + 0.0016837944276630878, + 0.0013273399090394378, + 0.0015892586670815945, + 0.001922538853250444, + 0.0010022337082773447, + 0.0013697181129828095, + 0.0007309456123039126, + 0.0009580426849424839, + 0.0010523776290938258, + 0.0008131026988849044, + 0.0012957891449332237, + 0.0015982685144990683, + 0.0011483535636216402, + 0.0019423624034970999, + 0.001807153457775712, + 0.0016788627253845334, + 0.0019272927893325686, + 0.0013001218903809786, + 0.0012301024980843067, + 0.0018465096363797784, + 0.0018125330097973347, + 0.0019417768344283104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.009888590313494205 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043079499155282974 + ], + "opposite_logprob_tokens": [ + -3.1680796146392822 + ], + "logprob_reward": -0.043079499155282974, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043079499155282974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0656457245349884, + 0.022798866033554077, + 0.0185267124325037, + 0.04259587451815605, + 0.011566558852791786, + 0.6654711961746216, + 0.053818706423044205, + 0.04358784109354019, + 0.019531741738319397, + 0.05645674467086792 + ], + "credit_map": [ + 0.006793481297791004, + 0.0037647392600774765, + 0.009047321043908596, + 0.009019998833537102, + 0.008225727826356888, + 0.008722865954041481, + 0.013656103983521461, + 0.006415483076125383, + 0.004307921510189772, + 0.002448796993121505, + 0.003285688580945134, + 0.0029771295376122, + 0.0021457087714225054, + 0.0021637107711285353, + 0.0018658962799236178, + 0.0036040134727954865, + 0.002953606890514493, + 0.0027017160318791866, + 0.0031975090969353914, + 0.002562048612162471, + 0.0023743188939988613, + 0.0018019925337284803, + 0.0013791888486593962, + 0.0015563326887786388, + 0.006128555629402399, + 0.00484858825802803, + 0.007673159707337618, + 0.005411762278527021, + 0.0047537400387227535, + 0.004150266759097576, + 0.005220326595008373, + 0.004409474320709705, + 0.0015902395825833082, + 0.0012789942556992173, + 0.0010712562361732125, + 0.001400588545948267, + 0.001681244932115078, + 0.0014637192944064736, + 0.0015201838687062263, + 0.001560332253575325, + 0.013174396939575672, + 0.025454916059970856, + 0.007679772097617388, + 0.005339991766959429, + 0.013122853823006153, + 0.3052782714366913, + 0.00793377310037613, + 0.287487268447876, + 0.006145934574306011, + 0.0073070889338850975, + 0.005014943890273571, + 0.007661396637558937, + 0.009410174563527107, + 0.00772287230938673, + 0.004182689357548952, + 0.0063736047595739365, + 0.004423083271831274, + 0.003741428954526782, + 0.004654552321881056, + 0.004175154492259026, + 0.007542733568698168, + 0.0086918193846941, + 0.005192534066736698, + 0.005166534800082445, + 0.001733078621327877, + 0.0020128563046455383, + 0.003026216756552458, + 0.002265980001538992, + 0.003427940420806408, + 0.0023390944115817547, + 0.0024919670540839434, + 0.0022346084006130695, + 0.004496904090046883, + 0.008105932734906673, + 0.006141967140138149, + 0.0063933636993169785, + 0.0058533507399261, + 0.011585257947444916, + 0.004349843133240938, + 0.00953012052923441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.043079499155282974 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067427014000713825 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.0067427014000713825, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067427014000713825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10933631658554077, + 0.06005864590406418, + 0.028734929859638214, + 0.08868467062711716, + 0.02243671379983425, + 0.46142005920410156, + 0.03653234988451004, + 0.04227466136217117, + 0.04086389020085335, + 0.10965777188539505 + ], + "credit_map": [ + 0.015418225899338722, + 0.00647329306229949, + 0.013472285121679306, + 0.01620200090110302, + 0.011019876226782799, + 0.014802749268710613, + 0.021040989086031914, + 0.010906890034675598, + 0.010749617591500282, + 0.00842618104070425, + 0.008831865154206753, + 0.008882184512913227, + 0.004725364968180656, + 0.006013095378875732, + 0.0050651091150939465, + 0.007365228608250618, + 0.006134610157459974, + 0.0038197168614715338, + 0.004648219328373671, + 0.0038416176103055477, + 0.0021468091290444136, + 0.002938904333859682, + 0.002345070941373706, + 0.002859981032088399, + 0.012106582522392273, + 0.011988882906734943, + 0.014192105270922184, + 0.007818717509508133, + 0.012753010727465153, + 0.00866954680532217, + 0.01337379589676857, + 0.0077820331789553165, + 0.0026915078051388264, + 0.0026605352759361267, + 0.0017823639791458845, + 0.0024881234858185053, + 0.004070844501256943, + 0.003005916252732277, + 0.0026213948149234056, + 0.003116029081866145, + 0.01076742447912693, + 0.018366141244769096, + 0.005831202957779169, + 0.004345371387898922, + 0.00975083839148283, + 0.209520623087883, + 0.005538244266062975, + 0.19730019569396973, + 0.003934124484658241, + 0.003499317215755582, + 0.0033302519004791975, + 0.0038683766033500433, + 0.007319869473576546, + 0.005076002329587936, + 0.004405192099511623, + 0.005099216476082802, + 0.004321475047618151, + 0.005193572957068682, + 0.0042622569017112255, + 0.0045643034391105175, + 0.009476741775870323, + 0.00561503367498517, + 0.003491930663585663, + 0.005349346436560154, + 0.0049712383188307285, + 0.004517601802945137, + 0.005235852673649788, + 0.003185182809829712, + 0.005714603699743748, + 0.00682859169319272, + 0.003756936639547348, + 0.006653879769146442, + 0.010373174212872982, + 0.026444122195243835, + 0.013603918254375458, + 0.01065056398510933, + 0.013118483126163483, + 0.012251882813870907, + 0.011998354457318783, + 0.011217274703085423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.0067427014000713825 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062015458941459656 + ], + "opposite_logprob_tokens": [ + -2.8120155334472656 + ], + "logprob_reward": -0.062015458941459656, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062015458941459656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0467243418097496, + 0.01580628752708435, + 0.01849895529448986, + 0.025219621136784554, + 0.01899603195488453, + 0.7386897802352905, + 0.047930240631103516, + 0.013645348139107227, + 0.02191653475165367, + 0.052572861313819885 + ], + "credit_map": [ + 0.008681079372763634, + 0.00225693266838789, + 0.00824922975152731, + 0.007416341453790665, + 0.0036583321634680033, + 0.0044800890609622, + 0.007054438348859549, + 0.004927897360175848, + 0.0023591776844114065, + 0.001764050219208002, + 0.0020983763970434666, + 0.0029157206881791353, + 0.0010533106978982687, + 0.0014400461222976446, + 0.0016723060980439186, + 0.0025033000856637955, + 0.0026523424312472343, + 0.002510394435375929, + 0.002771981991827488, + 0.0025648842565715313, + 0.0019851946271955967, + 0.0024726376868784428, + 0.0015882040606811643, + 0.0019533149898052216, + 0.003539842553436756, + 0.002903761574998498, + 0.005515366327017546, + 0.002344402251765132, + 0.002291622804477811, + 0.0032673687674105167, + 0.0018108224030584097, + 0.0035464318934828043, + 0.0017326739616692066, + 0.002094295108690858, + 0.0011914345668628812, + 0.0016349448123946786, + 0.0046332720667123795, + 0.002715004375204444, + 0.002153456909582019, + 0.0028409489896148443, + 0.014511413872241974, + 0.02888493984937668, + 0.008035439997911453, + 0.005948337260633707, + 0.014446363784372807, + 0.33876657485961914, + 0.00856797955930233, + 0.319528728723526, + 0.005734364036470652, + 0.006020835600793362, + 0.004491766449064016, + 0.006959253456443548, + 0.007413538638502359, + 0.0073862518183887005, + 0.0038709540385752916, + 0.00605327682569623, + 0.0013372767716646194, + 0.0014796360628679395, + 0.0016912653809413314, + 0.0016880992334336042, + 0.002617939142510295, + 0.0021871302742511034, + 0.0010089019779115915, + 0.001635098597034812, + 0.0019189800368621945, + 0.0021876683458685875, + 0.002972233109176159, + 0.0013200071407482028, + 0.0031939856708049774, + 0.0037081807386130095, + 0.0026038233190774918, + 0.004011655692011118, + 0.004836570471525192, + 0.009655502624809742, + 0.0060423691757023335, + 0.003978530410677195, + 0.0052332803606987, + 0.009529262781143188, + 0.004998618271201849, + 0.008298723958432674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.062015458941459656 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31343331933021545 + ], + "opposite_logprob_tokens": [ + -1.313433289527893 + ], + "logprob_reward": -0.31343331933021545, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31343331933021545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03966309875249863, + 0.014825998805463314, + 0.021001111716032028, + 0.02906920202076435, + 0.013684460893273354, + 0.7976453900337219, + 0.018232226371765137, + 0.010559453628957272, + 0.033894915133714676, + 0.02142413891851902 + ], + "credit_map": [ + 0.00599991250783205, + 0.0018205217784270644, + 0.005172887351363897, + 0.005747185088694096, + 0.0032805532682687044, + 0.003925572615116835, + 0.008379022590816021, + 0.005337446462363005, + 0.0023787550162523985, + 0.0015755168860778213, + 0.0019531322177499533, + 0.0030354296322911978, + 0.0012392190983518958, + 0.0015652283327654004, + 0.001543956226669252, + 0.0015347624430432916, + 0.003769582137465477, + 0.0030518067069351673, + 0.0041523161344230175, + 0.0025935000739991665, + 0.0022038614843040705, + 0.001786875887773931, + 0.0015898062847554684, + 0.0018533638212829828, + 0.004723000340163708, + 0.003972712438553572, + 0.005410471465438604, + 0.0027566319331526756, + 0.002394517417997122, + 0.0034336629323661327, + 0.002222095150500536, + 0.0041561126708984375, + 0.0014033699408173561, + 0.0019302807049825788, + 0.0009696633205749094, + 0.0015296736964955926, + 0.0027414122596383095, + 0.0018009752966463566, + 0.0015079950680956244, + 0.0018010918283835053, + 0.015436054207384586, + 0.03109625354409218, + 0.008830984123051167, + 0.006863733287900686, + 0.016438426449894905, + 0.36541035771369934, + 0.009517889469861984, + 0.3440517783164978, + 0.0021983806509524584, + 0.0019101599464192986, + 0.0017410579603165388, + 0.002084097359329462, + 0.0032895475160330534, + 0.0021612043492496014, + 0.00252696149982512, + 0.002320817206054926, + 0.0010602205293253064, + 0.0011027069995179772, + 0.0009273343603126705, + 0.0010365367634221911, + 0.0017255827551707625, + 0.0020366308744996786, + 0.0011000280501320958, + 0.001570414169691503, + 0.0031586289405822754, + 0.0039772754535079, + 0.004664737265557051, + 0.002405832288786769, + 0.005340898875147104, + 0.005155285820364952, + 0.003941159229725599, + 0.00525110075250268, + 0.0029443863313645124, + 0.0023957351222634315, + 0.002889528637751937, + 0.0017312710406258702, + 0.0019959628116339445, + 0.003406429197639227, + 0.002609193092212081, + 0.0034516334999352694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.31343331933021545 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of high-speed sports photography, dramatic low-angle propeller shots capture a cascade of vibrant objects hurled vertically upward, their ascent rendered in extreme slow motion. The camera rotates and tilts, carefully following a mix of ornate red envelopes and glossy coupons as they flutter with athletic grace against a stark, uniformly lit background. The red envelopes, made of richly textured crimson paper embossed with an intricate, shimmering gold foil pattern, catch the light as they tumble. The coupons, dotted among them, their high-gloss surface shimmering with brilliant mirrored highlights, reveal vibrant colors and bold, abstract graphic elements. Under powerful, shadowless studio high-key lighting, every detail is sharply focused. Each item performs a unique aerial ballet: some envelopes bend and flex as they pass through the air, while the stiffer coupons spin on their axes like precisely thrown discs, creating a mesmerizing, frozen moment of celebratory commerce that feels both chaotic and elegant.", + "seed": 23006, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_3/index_6033_seed_23006_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.18621027637273074, + "vlm_logprob": -0.18621027637273074, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.18621027637273074, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007233619689941406 + ], + "opposite_logprob_tokens": [ + -5.632233619689941 + ], + "logprob_reward": -0.007233619689941406, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.007233619689941406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1895385980606079, + 0.10008803755044937, + 0.06309500336647034, + 0.10591459274291992, + 0.10654143989086151, + 0.16979791224002838, + 0.09239546954631805, + 0.06504765152931213, + 0.05663953721523285, + 0.05094178766012192 + ], + "credit_map": [ + 0.017332803457975388, + 0.015446501784026623, + 0.04005568102002144, + 0.01437405776232481, + 0.01669001579284668, + 0.042925506830215454, + 0.018807880580425262, + 0.023906152695417404, + 0.010014032013714314, + 0.01733798161149025, + 0.011143028736114502, + 0.009131277911365032, + 0.014505399391055107, + 0.013567348010838032, + 0.014798888005316257, + 0.009590080007910728, + 0.011733450926840305, + 0.006677760276943445, + 0.007892845198512077, + 0.005632133688777685, + 0.006992315873503685, + 0.009911766275763512, + 0.009359011426568031, + 0.004895724821835756, + 0.017354978248476982, + 0.012001103721559048, + 0.012159314937889576, + 0.005649792030453682, + 0.02248314954340458, + 0.021199582144618034, + 0.00939091295003891, + 0.0056757633574306965, + 0.00987760629504919, + 0.013436103239655495, + 0.02112342230975628, + 0.013042125850915909, + 0.0095810666680336, + 0.018237994983792305, + 0.01158978883177042, + 0.009653330780565739, + 0.012650715187191963, + 0.004890657961368561, + 0.009576967917382717, + 0.007789203897118568, + 0.005561182741075754, + 0.06118546426296234, + 0.004911580588668585, + 0.06323215365409851, + 0.011331410147249699, + 0.010963156819343567, + 0.010796425864100456, + 0.011311757378280163, + 0.01572551019489765, + 0.008624425157904625, + 0.010129649192094803, + 0.013513133861124516, + 0.005944887641817331, + 0.007967263460159302, + 0.014018562622368336, + 0.006146497558802366, + 0.004705313127487898, + 0.004024367313832045, + 0.006748467683792114, + 0.015492293052375317, + 0.005943908356130123, + 0.003261710051447153, + 0.016035517677664757, + 0.003670133650302887, + 0.0044499230571091175, + 0.004402559250593185, + 0.003187907161191106, + 0.01568787731230259, + 0.00722358888015151, + 0.004988056607544422, + 0.007171221077442169, + 0.004084084182977676, + 0.004157272633165121, + 0.009040231816470623, + 0.004926659166812897, + 0.009350670501589775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.007233619689941406 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430946871638298 + ], + "opposite_logprob_tokens": [ + -3.1680946350097656 + ], + "logprob_reward": -0.0430946871638298, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430946871638298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18901140987873077, + 0.14077915251255035, + 0.049424875527620316, + 0.12925128638744354, + 0.14336448907852173, + 0.08764482289552689, + 0.05779789015650749, + 0.08576244860887527, + 0.0823582112789154, + 0.034605417400598526 + ], + "credit_map": [ + 0.015635022893548012, + 0.022476209327578545, + 0.03087576851248741, + 0.02029905468225479, + 0.02527620829641819, + 0.03242413327097893, + 0.023989304900169373, + 0.01803572289645672, + 0.01709727570414543, + 0.032844427973032, + 0.01624089665710926, + 0.013116518035531044, + 0.01600494422018528, + 0.017854610458016396, + 0.014308849349617958, + 0.013311643153429031, + 0.0053143794648349285, + 0.005158203653991222, + 0.003842615522444248, + 0.005984188988804817, + 0.0069644153118133545, + 0.006629065610468388, + 0.010305451229214668, + 0.005226549226790667, + 0.012163607403635979, + 0.015445430763065815, + 0.014533575624227524, + 0.007099981419742107, + 0.034807849675416946, + 0.028127916157245636, + 0.011401887983083725, + 0.005671028979122639, + 0.013393726199865341, + 0.01984962448477745, + 0.029323412105441093, + 0.014735853299498558, + 0.01707102172076702, + 0.02713186852633953, + 0.011920707300305367, + 0.009938268922269344, + 0.006472611799836159, + 0.0022233733907341957, + 0.004061482381075621, + 0.00382729503326118, + 0.0026521075051277876, + 0.03323422372341156, + 0.002656764118000865, + 0.0325169675052166, + 0.008699511177837849, + 0.005036511924117804, + 0.007829977199435234, + 0.00637366808950901, + 0.009109757840633392, + 0.005432946141809225, + 0.00585938012227416, + 0.009456135332584381, + 0.0071167792193591595, + 0.015620339661836624, + 0.01151699386537075, + 0.009404136799275875, + 0.009555702097713947, + 0.008428975008428097, + 0.010696311481297016, + 0.013423212803900242, + 0.006885665003210306, + 0.0043503111228346825, + 0.02625630423426628, + 0.0046676103956997395, + 0.005039731040596962, + 0.004971118178218603, + 0.0026230565272271633, + 0.027564413845539093, + 0.004950487520545721, + 0.004915142431855202, + 0.004163658712059259, + 0.0028761462308466434, + 0.0034484609495848417, + 0.007141038775444031, + 0.003248905297368765, + 0.0038615756202489138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0430946871638298 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956635810434818 + ], + "opposite_logprob_tokens": [ + -5.130956649780273 + ], + "logprob_reward": -0.005956635810434818, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956635810434818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15267875790596008, + 0.07336488366127014, + 0.0671435222029686, + 0.3112781345844269, + 0.12730643153190613, + 0.03846348449587822, + 0.05276109650731087, + 0.06623510271310806, + 0.04315363988280296, + 0.06761494278907776 + ], + "credit_map": [ + 0.018477048724889755, + 0.01043193694204092, + 0.031692683696746826, + 0.01070315483957529, + 0.012932099401950836, + 0.03355155512690544, + 0.012852150946855545, + 0.022038117051124573, + 0.008343223482370377, + 0.011475506238639355, + 0.009496238082647324, + 0.006656175013631582, + 0.009906614199280739, + 0.010053209029138088, + 0.009810174815356731, + 0.007623737677931786, + 0.010867836885154247, + 0.007820584811270237, + 0.009985197335481644, + 0.005239422898739576, + 0.005378862842917442, + 0.010185245424509048, + 0.011988049373030663, + 0.005678326357156038, + 0.02771739661693573, + 0.03454386070370674, + 0.036173950880765915, + 0.016092529520392418, + 0.08651803433895111, + 0.07231299579143524, + 0.024391930550336838, + 0.013527433387935162, + 0.01393078826367855, + 0.017848342657089233, + 0.02593853324651718, + 0.017172284424304962, + 0.010133952833712101, + 0.02211015298962593, + 0.010886394418776035, + 0.009285991080105305, + 0.003637216053903103, + 0.0021528862416744232, + 0.004328434355556965, + 0.004850257188081741, + 0.0025448515079915524, + 0.010012884624302387, + 0.002223607851192355, + 0.008713343180716038, + 0.006450350396335125, + 0.006817517336457968, + 0.0060469708405435085, + 0.007276265416294336, + 0.009181949310004711, + 0.005115773528814316, + 0.005129450466483831, + 0.006742816884070635, + 0.005855835974216461, + 0.007161438465118408, + 0.011798468418419361, + 0.005329310894012451, + 0.009501504711806774, + 0.006275662686675787, + 0.006503933109343052, + 0.01380894798785448, + 0.005853978451341391, + 0.004420701414346695, + 0.005615265108644962, + 0.005887527484446764, + 0.004791816230863333, + 0.006694454699754715, + 0.004459766671061516, + 0.005430128891021013, + 0.01095503382384777, + 0.013943924568593502, + 0.012103477492928505, + 0.00584196113049984, + 0.005870266817510128, + 0.00783787202090025, + 0.00485669169574976, + 0.006205714773386717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.005956635810434818 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04862990975379944 + ], + "opposite_logprob_tokens": [ + -3.0486299991607666 + ], + "logprob_reward": -0.04862990975379944, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04862990975379944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07478448003530502, + 0.08015793561935425, + 0.03248632699251175, + 0.44688183069229126, + 0.06540092080831528, + 0.09974289685487747, + 0.0998416617512703, + 0.03348516300320625, + 0.01971566490828991, + 0.04750306159257889 + ], + "credit_map": [ + 0.006179825868457556, + 0.008073847740888596, + 0.012112096883356571, + 0.009240700863301754, + 0.008830188773572445, + 0.01374922227114439, + 0.006968315690755844, + 0.00963028147816658, + 0.008041114546358585, + 0.01270015724003315, + 0.008697929792106152, + 0.007081897929310799, + 0.012507674284279346, + 0.012003591284155846, + 0.008250527083873749, + 0.010875048115849495, + 0.0064865355379879475, + 0.0046688164584338665, + 0.0025122761726379395, + 0.0031161904335021973, + 0.003366815624758601, + 0.005709182471036911, + 0.0035349540412425995, + 0.003091555554419756, + 0.028969550505280495, + 0.0546552874147892, + 0.06429487466812134, + 0.034172721207141876, + 0.10971405357122421, + 0.09902451932430267, + 0.03594926744699478, + 0.02010154537856579, + 0.0066996426321566105, + 0.0065116300247609615, + 0.010958421975374222, + 0.0076814936473965645, + 0.007336813490837812, + 0.011544833891093731, + 0.009106329642236233, + 0.005561756435781717, + 0.01170759741216898, + 0.009141549468040466, + 0.014574624598026276, + 0.013036618940532207, + 0.012091480195522308, + 0.015935881063342094, + 0.007516506128013134, + 0.01573863811790943, + 0.016170352697372437, + 0.009764987044036388, + 0.013086501508951187, + 0.009147781878709793, + 0.01695377752184868, + 0.010845441371202469, + 0.007222771644592285, + 0.016650041565299034, + 0.0033920772839337587, + 0.0038520758971571922, + 0.005398493725806475, + 0.003416895167902112, + 0.004141628742218018, + 0.003696652129292488, + 0.004438912495970726, + 0.005148426163941622, + 0.0018520529847592115, + 0.001835726317949593, + 0.0039302813820540905, + 0.0023050960153341293, + 0.0019754068925976753, + 0.0019402988255023956, + 0.0015922978054732084, + 0.004284504801034927, + 0.0067640370689332485, + 0.004808107390999794, + 0.006452724803239107, + 0.003822190919891, + 0.00432710163295269, + 0.012029260396957397, + 0.0034796709660440683, + 0.005819969344884157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.04862990975379944 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261365294456482 + ], + "opposite_logprob_tokens": [ + -0.5761365294456482 + ], + "logprob_reward": -0.8261365294456482, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261365294456482, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09835704416036606, + 0.10737185180187225, + 0.047988954931497574, + 0.17054684460163116, + 0.0705408826470375, + 0.19312970340251923, + 0.20088618993759155, + 0.02990429475903511, + 0.041467953473329544, + 0.039806265383958817 + ], + "credit_map": [ + 0.008393651805818081, + 0.009948710910975933, + 0.019743937999010086, + 0.00919546838849783, + 0.012916401959955692, + 0.01982198655605316, + 0.008770957589149475, + 0.009565930813550949, + 0.007324309088289738, + 0.013712320476770401, + 0.008055680431425571, + 0.011092815548181534, + 0.02157284878194332, + 0.016322532668709755, + 0.012986560352146626, + 0.016304783523082733, + 0.007590696215629578, + 0.006317014340311289, + 0.007850896567106247, + 0.005223754793405533, + 0.0036786338314414024, + 0.007465374190360308, + 0.005960032343864441, + 0.003902554279193282, + 0.026492875069379807, + 0.022591479122638702, + 0.017892036586999893, + 0.008466450497508049, + 0.0375887006521225, + 0.03408895060420036, + 0.013982366770505905, + 0.009443983435630798, + 0.007001509424299002, + 0.010159791447222233, + 0.01387670449912548, + 0.008456056006252766, + 0.008005373179912567, + 0.012269125320017338, + 0.005809360183775425, + 0.004962960258126259, + 0.014442277140915394, + 0.004770659375935793, + 0.009199123829603195, + 0.008387155830860138, + 0.005872142501175404, + 0.0715818926692009, + 0.005781273823231459, + 0.0730951726436615, + 0.023515621200203896, + 0.024682557210326195, + 0.021738288924098015, + 0.020331133157014847, + 0.04116729274392128, + 0.023601293563842773, + 0.017183320596814156, + 0.02866668626666069, + 0.005014738067984581, + 0.003312397748231888, + 0.004465666599571705, + 0.0022619578521698713, + 0.0034749233163893223, + 0.003038416150957346, + 0.002791196573525667, + 0.005544997751712799, + 0.0035443103406578302, + 0.002468003425747156, + 0.011487222276628017, + 0.0026298428419977427, + 0.002747131045907736, + 0.0037006342317909002, + 0.0026643825694918633, + 0.012226426042616367, + 0.005501203238964081, + 0.004877218045294285, + 0.005320544820278883, + 0.003303349018096924, + 0.003435061313211918, + 0.007959316484630108, + 0.002943990984931588, + 0.006465580780059099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.8261365294456482 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "In the style of high-speed sports photography, dramatic low-angle propeller shots capture a cascade of vibrant objects hurled vertically upward, their ascent rendered in extreme slow motion. The camera rotates and tilts, carefully following a mix of ornate red envelopes and glossy coupons as they flutter with athletic grace against a stark, uniformly lit background. The red envelopes, made of richly textured crimson paper embossed with an intricate, shimmering gold foil pattern, catch the light as they tumble. The coupons, dotted among them, their high-gloss surface shimmering with brilliant mirrored highlights, reveal vibrant colors and bold, abstract graphic elements. Under powerful, shadowless studio high-key lighting, every detail is sharply focused. Each item performs a unique aerial ballet: some envelopes bend and flex as they pass through the air, while the stiffer coupons spin on their axes like precisely thrown discs, creating a mesmerizing, frozen moment of celebratory commerce that feels both chaotic and elegant.", + "seed": 23007, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_3/index_6033_seed_23007_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08375047668814659, + "vlm_logprob": -0.08375047668814659, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.08375047668814659, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011389357037842274 + ], + "opposite_logprob_tokens": [ + -5.011389255523682 + ], + "logprob_reward": -0.011389357037842274, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.011389357037842274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050767943263053894, + 0.04965747147798538, + 0.08378759771585464, + 0.18817026913166046, + 0.20010237395763397, + 0.016185404732823372, + 0.10249769687652588, + 0.012869886122643948, + 0.01946231909096241, + 0.27649903297424316 + ], + "credit_map": [ + 0.005459742620587349, + 0.006575919687747955, + 0.006339279003441334, + 0.008821561001241207, + 0.005967346020042896, + 0.006306284572929144, + 0.004568554926663637, + 0.006729257758706808, + 0.008182531222701073, + 0.0026873729657381773, + 0.005189008079469204, + 0.003793230513110757, + 0.0028739338740706444, + 0.012243250384926796, + 0.0026990529149770737, + 0.011989092454314232, + 0.00331383035518229, + 0.002884585177525878, + 0.002244208473712206, + 0.002698497613891959, + 0.003936406224966049, + 0.03325142338871956, + 0.004822300747036934, + 0.030636344105005264, + 0.028484389185905457, + 0.023824326694011688, + 0.021026387810707092, + 0.010256473906338215, + 0.01442534290254116, + 0.01853187382221222, + 0.025464940816164017, + 0.046156540513038635, + 0.026860831305384636, + 0.023447614163160324, + 0.03574587404727936, + 0.016185184940695763, + 0.022618789225816727, + 0.027067173272371292, + 0.01813969574868679, + 0.030037209391593933, + 0.00182815152220428, + 0.0021857828833162785, + 0.0019153195898979902, + 0.0011124536395072937, + 0.0026767116505652666, + 0.0027301297523081303, + 0.0018532014219090343, + 0.0018836549716070294, + 0.005839656572788954, + 0.017952440306544304, + 0.0054864599369466305, + 0.006472751498222351, + 0.02387203462421894, + 0.01815870963037014, + 0.007432641461491585, + 0.017282996326684952, + 0.0014139871345832944, + 0.0020775848533958197, + 0.0016042175702750683, + 0.0013107986887916923, + 0.0017687600338831544, + 0.001782948849722743, + 0.0011311908019706607, + 0.001780398073606193, + 0.002351488685235381, + 0.0018973903497681022, + 0.002429547253996134, + 0.0031623011454939842, + 0.0018032242078334093, + 0.0029848876874893904, + 0.0016342389862984419, + 0.0031992411240935326, + 0.006836401764303446, + 0.0031965095549821854, + 0.004218334332108498, + 0.006090758368372917, + 0.0029192683286964893, + 0.12402234226465225, + 0.006674844305962324, + 0.12254058569669724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.011389357037842274 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372541815042496 + ], + "opposite_logprob_tokens": [ + -3.4087255001068115 + ], + "logprob_reward": -0.03372541815042496, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372541815042496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15255126357078552, + 0.062107495963573456, + 0.09312689304351807, + 0.1416458785533905, + 0.01574591174721718, + 0.02050400897860527, + 0.036980483680963516, + 0.014342726208269596, + 0.019685808569192886, + 0.443309485912323 + ], + "credit_map": [ + 0.014690940268337727, + 0.006678840611129999, + 0.03073764033615589, + 0.008773691952228546, + 0.009348315186798573, + 0.05222867801785469, + 0.009305684827268124, + 0.02078748494386673, + 0.008511572144925594, + 0.0032915291376411915, + 0.006774707231670618, + 0.005792740732431412, + 0.0033993790857493877, + 0.015657547861337662, + 0.0053178188391029835, + 0.01336220558732748, + 0.005129712168127298, + 0.0036572108510881662, + 0.0038894994650036097, + 0.004855641629546881, + 0.0031927060335874557, + 0.03493335470557213, + 0.006650872528553009, + 0.030817901715636253, + 0.021934539079666138, + 0.017489131540060043, + 0.01603970304131508, + 0.007683939766138792, + 0.010114271193742752, + 0.013591189868748188, + 0.01958194561302662, + 0.03521117940545082, + 0.0019207348814234138, + 0.0019102641381323338, + 0.002501305192708969, + 0.0016736643156036735, + 0.0023776330053806305, + 0.001994765130802989, + 0.0012541787000373006, + 0.0021133674308657646, + 0.002217835746705532, + 0.0025380512233823538, + 0.0023216218687593937, + 0.0014727661618962884, + 0.0031375347170978785, + 0.004062643274664879, + 0.0022375008556991816, + 0.0025160587392747402, + 0.002223103540018201, + 0.004571594297885895, + 0.0022334072273224592, + 0.003473235061392188, + 0.00541491387411952, + 0.00797954574227333, + 0.003124210052192211, + 0.007960474118590355, + 0.0015302594983950257, + 0.0012627184623852372, + 0.003126521361991763, + 0.0011484153801575303, + 0.001728247618302703, + 0.002194636268541217, + 0.0011877869255840778, + 0.0021641403436660767, + 0.0021378814708441496, + 0.0024691724684089422, + 0.0017666906351223588, + 0.002621568040922284, + 0.0025558758061379194, + 0.002264888258650899, + 0.0029862308874726295, + 0.0028835018165409565, + 0.010616665706038475, + 0.00862368755042553, + 0.01180215086787939, + 0.02796083129942417, + 0.005632508546113968, + 0.17733100056648254, + 0.02555493637919426, + 0.17578773200511932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.03372541815042496 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646904490888119 + ], + "opposite_logprob_tokens": [ + -4.7586469650268555 + ], + "logprob_reward": -0.008646904490888119, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646904490888119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021733801811933517, + 0.05362885072827339, + 0.07881256937980652, + 0.1486595869064331, + 0.019933510571718216, + 0.020270489156246185, + 0.09278960525989532, + 0.01448375266045332, + 0.022127347066998482, + 0.5275604724884033 + ], + "credit_map": [ + 0.0026512728072702885, + 0.0023546814918518066, + 0.003116041421890259, + 0.0030996035784482956, + 0.0022967897821217775, + 0.00265910173766315, + 0.002360817277804017, + 0.003195497440174222, + 0.008636892773211002, + 0.0026191275101155043, + 0.0049428860656917095, + 0.003477157559245825, + 0.002338969148695469, + 0.014381492510437965, + 0.002916608238592744, + 0.014315725304186344, + 0.0041582765989005566, + 0.0028997401241213083, + 0.002265169285237789, + 0.002917107893154025, + 0.003888071281835437, + 0.030696894973516464, + 0.00447165546938777, + 0.02751566283404827, + 0.020901303738355637, + 0.02010352723300457, + 0.017441900447010994, + 0.012032832019031048, + 0.012427818961441517, + 0.014110355637967587, + 0.01779712177813053, + 0.0338447242975235, + 0.002623059321194887, + 0.0020662809256464243, + 0.002547607058659196, + 0.0022656945511698723, + 0.0020018317736685276, + 0.0026713423430919647, + 0.002584228990599513, + 0.003173467004671693, + 0.0027039372362196445, + 0.003359261667355895, + 0.0021992584224790335, + 0.001616323017515242, + 0.002748006721958518, + 0.0027412576600909233, + 0.0024798742961138487, + 0.0024225704837590456, + 0.005101941525936127, + 0.014444025233387947, + 0.004905522335320711, + 0.006949153263121843, + 0.019478682428598404, + 0.018102869391441345, + 0.0062986151315271854, + 0.017508801072835922, + 0.0013899537734687328, + 0.002008996205404401, + 0.0018507129279896617, + 0.0010517314076423645, + 0.001590271946042776, + 0.0026129356119781733, + 0.0016538964118808508, + 0.0023252558894455433, + 0.003230469534173608, + 0.00243930472061038, + 0.002338059712201357, + 0.00365005386993289, + 0.0022875384893268347, + 0.0029861663933843374, + 0.0024202282074838877, + 0.00277552823536098, + 0.012741449289023876, + 0.006217545829713345, + 0.011263836175203323, + 0.01606227457523346, + 0.00591664481908083, + 0.22878648340702057, + 0.020385971292853355, + 0.22618627548217773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.008646904490888119 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11289369314908981 + ], + "opposite_logprob_tokens": [ + -2.237893581390381 + ], + "logprob_reward": -0.11289369314908981, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11289369314908981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14402775466442108, + 0.04141157120466232, + 0.08026742935180664, + 0.062464457005262375, + 0.10035568475723267, + 0.06226796656847, + 0.2785753011703491, + 0.048479724675416946, + 0.0201983954757452, + 0.16195164620876312 + ], + "credit_map": [ + 0.011310954578220844, + 0.006831638049334288, + 0.025346413254737854, + 0.008630800060927868, + 0.008698873221874237, + 0.0503782257437706, + 0.007856088690459728, + 0.024974755942821503, + 0.005474559031426907, + 0.0029205831233412027, + 0.007379473652690649, + 0.0048196567222476006, + 0.0029647168703377247, + 0.007008504122495651, + 0.0035355491563677788, + 0.007308528292924166, + 0.0034731226041913033, + 0.0038406008388847113, + 0.0022533629089593887, + 0.0032561521511524916, + 0.005875302013009787, + 0.028336429968476295, + 0.006096747238188982, + 0.027135713025927544, + 0.006817683577537537, + 0.009275591000914574, + 0.008269593119621277, + 0.008276021108031273, + 0.005970314610749483, + 0.0057493713684380054, + 0.008919981308281422, + 0.009185899049043655, + 0.013446143828332424, + 0.01112984586507082, + 0.017670756205916405, + 0.006869657896459103, + 0.012839693576097488, + 0.01209146250039339, + 0.011456755921244621, + 0.014851370826363564, + 0.008856202475726604, + 0.010680785402655602, + 0.007239838130772114, + 0.004416394978761673, + 0.008433608338236809, + 0.00794642698019743, + 0.007284404244273901, + 0.007410308811813593, + 0.01489995513111353, + 0.04659280925989151, + 0.013699028640985489, + 0.019300492480397224, + 0.06495591253042221, + 0.05102157965302467, + 0.020511668175458908, + 0.047593872994184494, + 0.004438478499650955, + 0.003128618001937866, + 0.013700684532523155, + 0.003543396946042776, + 0.004649382550269365, + 0.007068766746670008, + 0.005215126555413008, + 0.006735267583280802, + 0.002861188491806388, + 0.002388561610132456, + 0.002130273962393403, + 0.003238052362576127, + 0.0018207910470664501, + 0.002552418038249016, + 0.002470940351486206, + 0.0027361700776964426, + 0.004265942610800266, + 0.005143076181411743, + 0.005513051059097052, + 0.009144325740635395, + 0.0033660312183201313, + 0.06319893151521683, + 0.008727168664336205, + 0.06259311735630035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.11289369314908981 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520970106124878 + ], + "opposite_logprob_tokens": [ + -1.5020970106124878 + ], + "logprob_reward": -0.2520970106124878, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520970106124878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2182932049036026, + 0.06755529344081879, + 0.029661139473319054, + 0.047920022159814835, + 0.31832724809646606, + 0.050051383674144745, + 0.15206363797187805, + 0.018850799649953842, + 0.013587565161287785, + 0.08368970453739166 + ], + "credit_map": [ + 0.02032463625073433, + 0.00799347274005413, + 0.051368169486522675, + 0.012647328898310661, + 0.012536446563899517, + 0.07420744001865387, + 0.012547205202281475, + 0.02666849084198475, + 0.010974988341331482, + 0.0036849842872470617, + 0.008508679457008839, + 0.006320720072835684, + 0.004509660415351391, + 0.012095363810658455, + 0.005512085277587175, + 0.01594880223274231, + 0.0036102833691984415, + 0.003167911898344755, + 0.0032099748495966196, + 0.00355292996391654, + 0.0022826401982456446, + 0.0052017695270478725, + 0.004656844772398472, + 0.003978783264756203, + 0.006014802493155003, + 0.005505761597305536, + 0.006169486325234175, + 0.0037189112044870853, + 0.002807673066854477, + 0.0038781920447945595, + 0.008584512397646904, + 0.011240677908062935, + 0.04218004271388054, + 0.03726617991924286, + 0.05573997274041176, + 0.0269828662276268, + 0.03597474470734596, + 0.04506780579686165, + 0.028099767863750458, + 0.047015853226184845, + 0.006671322043985128, + 0.006467324215918779, + 0.006452854257076979, + 0.0037837489508092403, + 0.008574672043323517, + 0.0073056938126683235, + 0.00562305049970746, + 0.005172712728381157, + 0.007824569940567017, + 0.02370402216911316, + 0.007593328598886728, + 0.010210647247731686, + 0.03151367977261543, + 0.030769390985369682, + 0.010455027222633362, + 0.029992947354912758, + 0.0012617469765245914, + 0.0013078742194920778, + 0.002214146079495549, + 0.0009835632517933846, + 0.002439516829326749, + 0.004730992950499058, + 0.0013353789690881968, + 0.0045775785110890865, + 0.0014896681532263756, + 0.001639493857510388, + 0.0018252532463520765, + 0.0018320909002795815, + 0.0013065855018794537, + 0.0020294799469411373, + 0.0014778408221900463, + 0.0019871522672474384, + 0.002948199864476919, + 0.0037274605128914118, + 0.0061892615631222725, + 0.011704319156706333, + 0.0035766081418842077, + 0.01968521624803543, + 0.01706118881702423, + 0.018797434866428375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.2520970106124878 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of high-speed sports photography, dramatic low-angle propeller shots capture a cascade of vibrant objects hurled vertically upward, their ascent rendered in extreme slow motion. The camera rotates and tilts, carefully following a mix of ornate red envelopes and glossy coupons as they flutter with athletic grace against a stark, uniformly lit background. The red envelopes, made of richly textured crimson paper embossed with an intricate, shimmering gold foil pattern, catch the light as they tumble. The coupons, dotted among them, their high-gloss surface shimmering with brilliant mirrored highlights, reveal vibrant colors and bold, abstract graphic elements. Under powerful, shadowless studio high-key lighting, every detail is sharply focused. Each item performs a unique aerial ballet: some envelopes bend and flex as they pass through the air, while the stiffer coupons spin on their axes like precisely thrown discs, creating a mesmerizing, frozen moment of celebratory commerce that feels both chaotic and elegant.", + "seed": 23008, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_3/index_6033_seed_23008_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.035280109802261, + "vlm_logprob": -0.035280109802261, + "vlm_yes_no_margin": 4.399999952316284, + "vlm_token_credit": -0.035280109802261, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004707681946456432 + ], + "opposite_logprob_tokens": [ + -6.5047078132629395 + ], + "logprob_reward": -0.004707681946456432, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.004707681946456432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15359151363372803, + 0.07478535175323486, + 0.09490833431482315, + 0.14441709220409393, + 0.10754922777414322, + 0.10947714745998383, + 0.060880642384290695, + 0.09735048562288284, + 0.05376771092414856, + 0.10327255725860596 + ], + "credit_map": [ + 0.023606479167938232, + 0.016912294551730156, + 0.01816382072865963, + 0.01709176041185856, + 0.021223681047558784, + 0.019278666004538536, + 0.015083126723766327, + 0.022231677547097206, + 0.010231955908238888, + 0.01095321774482727, + 0.01054011844098568, + 0.00834847241640091, + 0.009398026391863823, + 0.009868822991847992, + 0.0074627455323934555, + 0.007981984876096249, + 0.008074073120951653, + 0.009215335361659527, + 0.008904974907636642, + 0.012394038029015064, + 0.008474337868392467, + 0.013610121794044971, + 0.01564166694879532, + 0.01859378069639206, + 0.032173916697502136, + 0.00817766785621643, + 0.012609095312654972, + 0.010078739374876022, + 0.007154602557420731, + 0.024745946750044823, + 0.008023375645279884, + 0.04145374894142151, + 0.010870996862649918, + 0.01120853889733553, + 0.02111014723777771, + 0.011037329211831093, + 0.011534370481967926, + 0.01876923255622387, + 0.011312728747725487, + 0.011705880984663963, + 0.00980504509061575, + 0.016004081815481186, + 0.015855170786380768, + 0.016237406060099602, + 0.018965857103466988, + 0.011385008692741394, + 0.010576240718364716, + 0.010648337192833424, + 0.01137369405478239, + 0.008501392789185047, + 0.008380785584449768, + 0.006017782725393772, + 0.006611795164644718, + 0.007688242942094803, + 0.006307323928922415, + 0.005999625660479069, + 0.016459139063954353, + 0.011692838743329048, + 0.014625410549342632, + 0.01491150539368391, + 0.010693671181797981, + 0.007196397520601749, + 0.011085032485425472, + 0.01068649161607027, + 0.010885758325457573, + 0.006342209409922361, + 0.008900442160665989, + 0.006740979850292206, + 0.004963219631463289, + 0.0047910758294165134, + 0.006266413256525993, + 0.004877610597759485, + 0.010878115892410278, + 0.00859016738831997, + 0.012997402809560299, + 0.017709892243146896, + 0.008894508704543114, + 0.01648254506289959, + 0.007831312716007233, + 0.019888605922460556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.004707681946456432 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023300303146243095 + ], + "opposite_logprob_tokens": [ + -3.7733004093170166 + ], + "logprob_reward": -0.023300303146243095, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023300303146243095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20968180894851685, + 0.12969167530536652, + 0.1138119250535965, + 0.05882374569773674, + 0.13211670517921448, + 0.05029481276869774, + 0.058203138411045074, + 0.1292162537574768, + 0.07331262528896332, + 0.044847313314676285 + ], + "credit_map": [ + 0.030832618474960327, + 0.018605506047606468, + 0.029242824763059616, + 0.025348655879497528, + 0.024181153625249863, + 0.023557785898447037, + 0.02936522662639618, + 0.02854803577065468, + 0.022615978494286537, + 0.01583038456737995, + 0.020651645958423615, + 0.01326893363147974, + 0.01671931892633438, + 0.014218439348042011, + 0.012664558365941048, + 0.01372241135686636, + 0.008442599326372147, + 0.010842028073966503, + 0.011591254733502865, + 0.015912620350718498, + 0.014016866683959961, + 0.014343014918267727, + 0.020955143496394157, + 0.01770840398967266, + 0.009918530471622944, + 0.004438083618879318, + 0.0061198510229587555, + 0.007602038327604532, + 0.0031290689948946238, + 0.009879889898002148, + 0.003397674998268485, + 0.014338605105876923, + 0.017224177718162537, + 0.010515015572309494, + 0.027474289759993553, + 0.020460210740566254, + 0.011848598718643188, + 0.01645686663687229, + 0.014399533160030842, + 0.013738002628087997, + 0.006138332653790712, + 0.008068735711276531, + 0.006798813119530678, + 0.009705731645226479, + 0.004813727457076311, + 0.005732829216867685, + 0.004382213577628136, + 0.004654430318623781, + 0.010433128103613853, + 0.008174744434654713, + 0.008124344982206821, + 0.004187810700386763, + 0.005345193203538656, + 0.00696613360196352, + 0.005532668437808752, + 0.009439116343855858, + 0.024778202176094055, + 0.01681673340499401, + 0.01798207126557827, + 0.017182815819978714, + 0.014392067678272724, + 0.009594900533556938, + 0.015254332683980465, + 0.013215132057666779, + 0.016618868336081505, + 0.006802916992455721, + 0.012980074621737003, + 0.008436125703155994, + 0.007115461397916079, + 0.004529437981545925, + 0.006594801787286997, + 0.010234935209155083, + 0.0068331994116306305, + 0.00476674921810627, + 0.006260314024984837, + 0.008828096091747284, + 0.0046258713118731976, + 0.005361664574593306, + 0.003537526587024331, + 0.004633894190192223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.023300303146243095 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031960143242031336 + ], + "opposite_logprob_tokens": [ + -5.7531962394714355 + ], + "logprob_reward": -0.0031960143242031336, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031960143242031336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19579778611660004, + 0.06680216640233994, + 0.07736796885728836, + 0.21055369079113007, + 0.12541483342647552, + 0.05807158350944519, + 0.0511021763086319, + 0.09269437193870544, + 0.09905781596899033, + 0.02313760295510292 + ], + "credit_map": [ + 0.023612555116415024, + 0.024653367698192596, + 0.01875603385269642, + 0.01737852953374386, + 0.03241946920752525, + 0.027605155482888222, + 0.021513577550649643, + 0.029859105125069618, + 0.009893755428493023, + 0.0089415917173028, + 0.011714369989931583, + 0.007109313737601042, + 0.006502893753349781, + 0.009060311131179333, + 0.006889578886330128, + 0.006690355949103832, + 0.011126138269901276, + 0.01132704596966505, + 0.006660047918558121, + 0.010306533426046371, + 0.007122335489839315, + 0.010523486882448196, + 0.008805706165730953, + 0.011496671475470066, + 0.04109302535653114, + 0.009926303289830685, + 0.014537814073264599, + 0.015827719122171402, + 0.008334077894687653, + 0.043764736503362656, + 0.009190295822918415, + 0.06787974387407303, + 0.01797054335474968, + 0.00938501488417387, + 0.02325635589659214, + 0.013348751701414585, + 0.010819086804986, + 0.02085437998175621, + 0.014540615491569042, + 0.015240086242556572, + 0.006825127173215151, + 0.008942383341491222, + 0.007200273685157299, + 0.008747275918722153, + 0.007444954011589289, + 0.005767288617789745, + 0.007023046258836985, + 0.0061212400905787945, + 0.011439804919064045, + 0.007206739857792854, + 0.007716175634413958, + 0.0037482907064259052, + 0.003750101663172245, + 0.005711138714104891, + 0.004650804214179516, + 0.006879121530801058, + 0.012739158235490322, + 0.010580443777143955, + 0.01453407108783722, + 0.010480115190148354, + 0.009788579307496548, + 0.007502992171794176, + 0.015608149580657482, + 0.011460859328508377, + 0.020869528874754906, + 0.01020308118313551, + 0.016212135553359985, + 0.0095182741060853, + 0.010903061367571354, + 0.00782647356390953, + 0.00961314793676138, + 0.013912121765315533, + 0.003396796528249979, + 0.002789656398817897, + 0.003329133614897728, + 0.003912004642188549, + 0.0022459018509835005, + 0.002999626798555255, + 0.002300049876794219, + 0.0021644351072609425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.0031960143242031336 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018173374235630035 + ], + "opposite_logprob_tokens": [ + -4.0181732177734375 + ], + "logprob_reward": -0.018173374235630035, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018173374235630035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10982698947191238, + 0.0728795975446701, + 0.06565532088279724, + 0.16813509166240692, + 0.07069265097379684, + 0.2751637399196625, + 0.07809365540742874, + 0.08752792328596115, + 0.040220536291599274, + 0.03180450201034546 + ], + "credit_map": [ + 0.01287487056106329, + 0.010175793431699276, + 0.015418498776853085, + 0.017593713477253914, + 0.010612218640744686, + 0.01465698704123497, + 0.013005336746573448, + 0.015489570796489716, + 0.008856669999659061, + 0.005789735820144415, + 0.010998149402439594, + 0.009499922394752502, + 0.00750245014205575, + 0.010867750272154808, + 0.011022482067346573, + 0.008342430926859379, + 0.0057360511273145676, + 0.009855435229837894, + 0.007016327232122421, + 0.008776039816439152, + 0.007703819312155247, + 0.010293932631611824, + 0.008900596760213375, + 0.007373108994215727, + 0.032671477645635605, + 0.008307521231472492, + 0.013856073841452599, + 0.01254179049283266, + 0.0083193089812994, + 0.03276471421122551, + 0.007502364926040173, + 0.052171848714351654, + 0.01029460784047842, + 0.007941285148262978, + 0.01111997477710247, + 0.009921356104314327, + 0.009277568198740482, + 0.0072147175669670105, + 0.007316808681935072, + 0.0076063284650444984, + 0.04933805763721466, + 0.03580368310213089, + 0.04115975275635719, + 0.03654088079929352, + 0.029115667566657066, + 0.027757620438933372, + 0.027904560789465904, + 0.02754351496696472, + 0.020593274384737015, + 0.010255011729896069, + 0.013765408657491207, + 0.0047067273408174515, + 0.005601165350526571, + 0.008281494490802288, + 0.005891010165214539, + 0.008999560959637165, + 0.011598754674196243, + 0.00999626237899065, + 0.013293003663420677, + 0.01555332075804472, + 0.009178603067994118, + 0.00795774720609188, + 0.009404955431818962, + 0.01054527796804905, + 0.008608498610556126, + 0.0037880169693380594, + 0.00718412222340703, + 0.005821256432682276, + 0.003273239592090249, + 0.0036880825646221638, + 0.003039258299395442, + 0.004818062297999859, + 0.0037951136473566294, + 0.003133632242679596, + 0.004683560226112604, + 0.006601751782000065, + 0.0035478388890624046, + 0.004034223034977913, + 0.0027021632995456457, + 0.003306216560304165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.018173374235630035 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are ornate red envelopes and glossy coupons visible as the main objects in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the red envelopes made of textured crimson paper with an embossed gold foil pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the coupons high-gloss with vibrant colors and bold abstract graphic elements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the objects move vertically upward in extreme slow motion with a mix of tumbling, bending, and spinning motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with powerful, shadowless studio high-key lighting where every detail is sharply focused?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702317535877228 + ], + "opposite_logprob_tokens": [ + -2.127023220062256 + ], + "logprob_reward": -0.12702317535877228, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702317535877228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1494360715150833, + 0.0909361019730568, + 0.0781300961971283, + 0.10257352888584137, + 0.12487362325191498, + 0.09636177122592926, + 0.12897087633609772, + 0.06912139803171158, + 0.11724932491779327, + 0.04234723746776581 + ], + "credit_map": [ + 0.02656877227127552, + 0.014478221535682678, + 0.024228503927588463, + 0.013377455063164234, + 0.015188299119472504, + 0.018672842532396317, + 0.015840301290154457, + 0.021081669256091118, + 0.01128215342760086, + 0.012919972650706768, + 0.012964312918484211, + 0.007946186698973179, + 0.009791619144380093, + 0.011761288158595562, + 0.010587080381810665, + 0.013683478347957134, + 0.01065125223249197, + 0.01032803114503622, + 0.006805181037634611, + 0.010647676885128021, + 0.005137831903994083, + 0.011051187291741371, + 0.01183843333274126, + 0.011670497246086597, + 0.014898647554218769, + 0.008959679864346981, + 0.010836663655936718, + 0.012184366583824158, + 0.00778633076697588, + 0.01687082089483738, + 0.00879472866654396, + 0.022242290899157524, + 0.021214347332715988, + 0.009284989908337593, + 0.021181130781769753, + 0.010284891352057457, + 0.01234472170472145, + 0.022580023854970932, + 0.010885676369071007, + 0.017097830772399902, + 0.010958867147564888, + 0.014664688147604465, + 0.014500236138701439, + 0.01609726995229721, + 0.012131776660680771, + 0.00968548096716404, + 0.008022751659154892, + 0.010300693102180958, + 0.03515085205435753, + 0.015436933375895023, + 0.021931316703557968, + 0.00939654465764761, + 0.008219759911298752, + 0.011404932476580143, + 0.010305989533662796, + 0.017124539241194725, + 0.011173654347658157, + 0.006775159388780594, + 0.01431243121623993, + 0.00942957028746605, + 0.004412906710058451, + 0.005485995206981897, + 0.007005016319453716, + 0.010526656173169613, + 0.03179784119129181, + 0.009483599103987217, + 0.020083151757717133, + 0.012437984347343445, + 0.010549725033342838, + 0.01100490614771843, + 0.007900100201368332, + 0.0139920050278306, + 0.0054474687203764915, + 0.003827822394669056, + 0.007189855445176363, + 0.006326101254671812, + 0.005175103899091482, + 0.0036178207956254482, + 0.0036244757939130068, + 0.007138583809137344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.12702317535877228 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_4/index_tensor([8956, 8956])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_4/index_tensor([8956, 8956])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5561ac85fcf62fc43920016a28033894ed874f9d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_4/index_tensor([8956, 8956])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, smooth 360-degree tracking shot focuses in close-up on a lone, burning antique oil lamp, nestled in the center of a rough-grained, dark wooden table, surrounded by an endless, bottomless darkness. The lamp boasts a brass base with a subtle patina, a clear glass base half-filled with golden oil, and a braided cotton wick tipped with a fiery orange-red glow. A teardrop-shaped flame flickers steadily and gently within the slender glass shade, providing the scene's sole source of light and casting a warm, soft glow. As the camera steadily circles the lamp, light refracts across the glass shade in fluid patterns, while the brass highlights on the base glide gracefully across the scene, and the soft glow seems to dance within the darkness, creating an atmosphere of tranquility, contemplation, and profound intimacy.", + "seed": 26337, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_4/index_8956_seed_26337_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12130152476020158, + "vlm_logprob": -0.12130152476020158, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.12130152476020158, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0035301053430885077 + ], + "opposite_logprob_tokens": [ + -6.003530025482178 + ], + "logprob_reward": -0.0035301053430885077, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0035301053430885077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06563939899206161, + 0.03247269243001938, + 0.019105108454823494, + 0.03679533302783966, + 0.02451314777135849, + 0.015006482601165771, + 0.0704222247004509, + 0.16279366612434387, + 0.5630448460578918, + 0.010207150131464005 + ], + "credit_map": [ + 0.006903119385242462, + 0.005866441410034895, + 0.017282001674175262, + 0.005230323877185583, + 0.00398043030872941, + 0.0078006102703511715, + 0.00657477555796504, + 0.012001700699329376, + 0.0029122615233063698, + 0.0044168587774038315, + 0.005050627514719963, + 0.0034249061718583107, + 0.0019673004280775785, + 0.0040964423678815365, + 0.005299646873027086, + 0.005304650403559208, + 0.002428486244753003, + 0.001893679378554225, + 0.004140262957662344, + 0.002165629528462887, + 0.001844331156462431, + 0.0027171808760613203, + 0.001978077692911029, + 0.0019374596886336803, + 0.0038820153567939997, + 0.002437142189592123, + 0.011166642419993877, + 0.00212611211463809, + 0.0017149809282273054, + 0.00309258745983243, + 0.00332557107321918, + 0.009050280787050724, + 0.0028187360148876905, + 0.0027070590294897556, + 0.00470095407217741, + 0.003710824530571699, + 0.002038071397691965, + 0.0020591868087649345, + 0.0031097412575036287, + 0.0033685762900859118, + 0.0033095090184360743, + 0.0013450636761263013, + 0.0032513532787561417, + 0.0020064159762114286, + 0.0010630902834236622, + 0.0012073626276105642, + 0.0014113953802734613, + 0.0014122925931587815, + 0.02030242420732975, + 0.004803256597369909, + 0.021612990647554398, + 0.005682460032403469, + 0.0026246753986924887, + 0.004940561018884182, + 0.0032396134920418262, + 0.007216237485408783, + 0.02987002581357956, + 0.00633016973733902, + 0.05987553298473358, + 0.0066169630736112595, + 0.003795321797952056, + 0.013831129297614098, + 0.005292090121656656, + 0.03718242794275284, + 0.07381995767354965, + 0.021797819063067436, + 0.22455935180187225, + 0.025364626199007034, + 0.01153289619833231, + 0.16871821880340576, + 0.013884811662137508, + 0.023367134854197502, + 0.001382258953526616, + 0.0009308899170719087, + 0.0014535809168592095, + 0.001496251206845045, + 0.0008539882255718112, + 0.0011863112449645996, + 0.0009236290352419019, + 0.0019802397582679987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.0035301053430885077 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059328556060791 + ], + "opposite_logprob_tokens": [ + -3.8955931663513184 + ], + "logprob_reward": -0.02059328556060791, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059328556060791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04777853935956955, + 0.02082962915301323, + 0.02992936782538891, + 0.01645740680396557, + 0.027033453807234764, + 0.08194220811128616, + 0.05123602971434593, + 0.06743649393320084, + 0.6440737843513489, + 0.013283097185194492 + ], + "credit_map": [ + 0.004865413531661034, + 0.0029598549008369446, + 0.009292778559029102, + 0.002643710933625698, + 0.0019981376826763153, + 0.01007126085460186, + 0.0023314973805099726, + 0.01361588854342699, + 0.002934233285486698, + 0.0018445223104208708, + 0.0046612173318862915, + 0.0014924451243132353, + 0.0011121864663437009, + 0.0035316043067723513, + 0.0017384991515427828, + 0.003514922456815839, + 0.0027862614952027798, + 0.0019452848937362432, + 0.005133427679538727, + 0.002633483614772558, + 0.0011307750828564167, + 0.004598501604050398, + 0.0031773634254932404, + 0.008524272590875626, + 0.0021556243300437927, + 0.0015008642803877592, + 0.00429956940934062, + 0.0015176670858636498, + 0.0012058705324307084, + 0.0016093059675768018, + 0.0019019729224964976, + 0.002266533672809601, + 0.0032278995495289564, + 0.0024183958303183317, + 0.009227757342159748, + 0.002958102384582162, + 0.0012843846343457699, + 0.002405805280432105, + 0.0018328279256820679, + 0.0036782824900001287, + 0.013486248441040516, + 0.006753737106919289, + 0.016860226169228554, + 0.009329313412308693, + 0.0036812350153923035, + 0.006592429708689451, + 0.00444183312356472, + 0.020797189325094223, + 0.01722789742052555, + 0.0028788724448531866, + 0.017942767590284348, + 0.0023186758626252413, + 0.0012551810359582305, + 0.0035518654622137547, + 0.001757315476424992, + 0.004303457215428352, + 0.011591444723308086, + 0.0025517295580357313, + 0.02341051585972309, + 0.0024693068116903305, + 0.0017531149787828326, + 0.006523774471133947, + 0.002653911244124174, + 0.016482705250382423, + 0.08438485860824585, + 0.024612518027424812, + 0.25836074352264404, + 0.02858673967421055, + 0.012526280246675014, + 0.1943088173866272, + 0.015270402655005455, + 0.026023466140031815, + 0.0015445046592503786, + 0.0010275091044604778, + 0.0022633993066847324, + 0.0023945034481585026, + 0.0009705257834866643, + 0.0014766294043511152, + 0.0011078123934566975, + 0.0024982132017612457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.02059328556060791 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101437050849199 + ], + "opposite_logprob_tokens": [ + -5.504101276397705 + ], + "logprob_reward": -0.004101437050849199, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101437050849199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10108133405447006, + 0.04586338624358177, + 0.03760441392660141, + 0.02945845015347004, + 0.02141338400542736, + 0.02798132784664631, + 0.07812858372926712, + 0.21047204732894897, + 0.41813385486602783, + 0.02986323833465576 + ], + "credit_map": [ + 0.008051717653870583, + 0.00841929018497467, + 0.01609150320291519, + 0.007498404011130333, + 0.006067871116101742, + 0.02587350644171238, + 0.010620076209306717, + 0.018458953127264977, + 0.003396280575543642, + 0.0036630076356232166, + 0.004654840566217899, + 0.003833874361589551, + 0.002113465452566743, + 0.010089440271258354, + 0.005457810591906309, + 0.012654668651521206, + 0.0037684645503759384, + 0.0030981912277638912, + 0.006114130839705467, + 0.00509389815852046, + 0.0022530974820256233, + 0.0057852487079799175, + 0.006317657884210348, + 0.005173724610358477, + 0.003818497061729431, + 0.0028501502238214016, + 0.006891756318509579, + 0.0029909079894423485, + 0.0022836218122392893, + 0.0032270061783492565, + 0.004265140742063522, + 0.0031313698273152113, + 0.002247461350634694, + 0.002124203834682703, + 0.0030325616244226694, + 0.003124134847894311, + 0.002622115658596158, + 0.0019346154294908047, + 0.003206306602805853, + 0.0031219839584082365, + 0.005214904434978962, + 0.0030233648139983416, + 0.006740004755556583, + 0.002923631342127919, + 0.002088522305712104, + 0.0020867842249572277, + 0.002061160048469901, + 0.003842957317829132, + 0.024483347311615944, + 0.0047559053637087345, + 0.02662927843630314, + 0.004779858980327845, + 0.0024753285106271505, + 0.005444881971925497, + 0.0027364238630980253, + 0.006823556497693062, + 0.039629820734262466, + 0.008894595317542553, + 0.07726765424013138, + 0.009219498373568058, + 0.004924874287098646, + 0.01775440201163292, + 0.006481457035988569, + 0.0462997704744339, + 0.054165035486221313, + 0.015890957787632942, + 0.16736513376235962, + 0.018754711374640465, + 0.008101667277514935, + 0.12649303674697876, + 0.010079015046358109, + 0.01728428527712822, + 0.003399272682145238, + 0.0038071461021900177, + 0.005283345002681017, + 0.003934647887945175, + 0.004432660527527332, + 0.00304006552323699, + 0.0029689318034797907, + 0.0029971713665872812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.004101437050849199 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495508760213852 + ], + "opposite_logprob_tokens": [ + -2.929955005645752 + ], + "logprob_reward": -0.05495508760213852, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495508760213852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1581149846315384, + 0.07707735151052475, + 0.059917349368333817, + 0.03773554414510727, + 0.03371141105890274, + 0.04571584239602089, + 0.13877494633197784, + 0.1272868663072586, + 0.30091413855552673, + 0.02075159177184105 + ], + "credit_map": [ + 0.014469322748482227, + 0.006795816123485565, + 0.04003576934337616, + 0.006908247712999582, + 0.0038420993369072676, + 0.025462454184889793, + 0.006397855933755636, + 0.054203424602746964, + 0.010965267196297646, + 0.005162268411368132, + 0.017486046999692917, + 0.0045882039703428745, + 0.003581434953957796, + 0.01458720862865448, + 0.004844603128731251, + 0.015862323343753815, + 0.005482750944793224, + 0.002864543115720153, + 0.014782397076487541, + 0.003065234050154686, + 0.001819615950807929, + 0.008403977379202843, + 0.00391819654032588, + 0.01958063617348671, + 0.004473969340324402, + 0.00236021401360631, + 0.012916390784084797, + 0.002120011020451784, + 0.001529467641375959, + 0.0030744122341275215, + 0.002476871944963932, + 0.008784210309386253, + 0.004848404787480831, + 0.003123741364106536, + 0.00899283867329359, + 0.0032908490393310785, + 0.002212031278759241, + 0.003325027646496892, + 0.0031130267307162285, + 0.0048054917715489864, + 0.006689293775707483, + 0.002485815901309252, + 0.01383361965417862, + 0.004321838729083538, + 0.003048494691029191, + 0.0031980203930288553, + 0.0036498981062322855, + 0.008488858118653297, + 0.048758722841739655, + 0.007011372596025467, + 0.04999355599284172, + 0.0049074008129537106, + 0.0024878629483282566, + 0.01015340443700552, + 0.004273253958672285, + 0.011189369484782219, + 0.021882422268390656, + 0.005034451838582754, + 0.043751027435064316, + 0.00433057127520442, + 0.002903934335336089, + 0.013266293331980705, + 0.004126511514186859, + 0.03199164196848869, + 0.039201296865940094, + 0.012142907828092575, + 0.11587493121623993, + 0.01387003529816866, + 0.008000090718269348, + 0.0859142616391182, + 0.010382899083197117, + 0.015527741052210331, + 0.002304789610207081, + 0.002784142969176173, + 0.003886665217578411, + 0.0020892524626106024, + 0.0027386958245187998, + 0.0022319485433399677, + 0.0018969872035086155, + 0.0028191101737320423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.05495508760213852 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233277082443237 + ], + "opposite_logprob_tokens": [ + -0.8983277082443237 + ], + "logprob_reward": -0.5233277082443237, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233277082443237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02882644720375538, + 0.010299748741090298, + 0.010473929345607758, + 0.04044777527451515, + 0.009956909343600273, + 0.014664219692349434, + 0.11107631772756577, + 0.2522289752960205, + 0.5156182646751404, + 0.006407464854419231 + ], + "credit_map": [ + 0.0030912095680832863, + 0.0017822012305259705, + 0.008935894817113876, + 0.0017448719590902328, + 0.0009530676179565489, + 0.005576417315751314, + 0.0011773324804380536, + 0.0055654533207416534, + 0.0009457297855988145, + 0.0006428147316910326, + 0.0010996292112395167, + 0.0005710520781576633, + 0.0003425748727750033, + 0.00269660004414618, + 0.0008479973184876144, + 0.003153350204229355, + 0.001447906717658043, + 0.001056934823282063, + 0.001967530930414796, + 0.00121101806871593, + 0.0005586959305219352, + 0.0013039843179285526, + 0.0015456852270290256, + 0.0013821721076965332, + 0.004878890700638294, + 0.003012074390426278, + 0.012546814046800137, + 0.002739307703450322, + 0.0022003615740686655, + 0.0032236867118626833, + 0.0036852266639471054, + 0.008161414414644241, + 0.0012059783330187201, + 0.001121525070630014, + 0.0028749913908541203, + 0.001127478200942278, + 0.0008621946908533573, + 0.0008104990120045841, + 0.000680976954754442, + 0.0012732655741274357, + 0.0026662375312298536, + 0.001419581356458366, + 0.0036364581901580095, + 0.0019269309705123305, + 0.0007984909461811185, + 0.0011944513535127044, + 0.0011492414632812142, + 0.0018728281138464808, + 0.03583644703030586, + 0.006612659431993961, + 0.038037244230508804, + 0.005321175325661898, + 0.003272325498983264, + 0.007759846281260252, + 0.004202434327453375, + 0.010034183971583843, + 0.047923825681209564, + 0.00941707193851471, + 0.09415856748819351, + 0.009617424570024014, + 0.005397337954491377, + 0.021586522459983826, + 0.0074064722284674644, + 0.05672173202037811, + 0.06790581345558167, + 0.01981215365231037, + 0.20641711354255676, + 0.023047268390655518, + 0.010275905020534992, + 0.15494486689567566, + 0.01242777705192566, + 0.020787321031093597, + 0.0007983802352100611, + 0.000580464955419302, + 0.0010267788311466575, + 0.0011807195842266083, + 0.00041562659316696227, + 0.0007202486740425229, + 0.000503689399920404, + 0.0011815562611445785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.5233277082443237 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, smooth 360-degree tracking shot focuses in close-up on a lone, burning antique oil lamp, nestled in the center of a rough-grained, dark wooden table, surrounded by an endless, bottomless darkness. The lamp boasts a brass base with a subtle patina, a clear glass base half-filled with golden oil, and a braided cotton wick tipped with a fiery orange-red glow. A teardrop-shaped flame flickers steadily and gently within the slender glass shade, providing the scene's sole source of light and casting a warm, soft glow. As the camera steadily circles the lamp, light refracts across the glass shade in fluid patterns, while the brass highlights on the base glide gracefully across the scene, and the soft glow seems to dance within the darkness, creating an atmosphere of tranquility, contemplation, and profound intimacy.", + "seed": 26338, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_4/index_8956_seed_26338_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11722213877364993, + "vlm_logprob": -0.11722213877364993, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.11722213877364993, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004532302729785442 + ], + "opposite_logprob_tokens": [ + -5.879532337188721 + ], + "logprob_reward": -0.004532302729785442, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.004532302729785442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.154821515083313, + 0.0645228773355484, + 0.1554012894630432, + 0.06450030207633972, + 0.04061416536569595, + 0.07322046905755997, + 0.09176672995090485, + 0.08362184464931488, + 0.060061290860176086, + 0.21146945655345917 + ], + "credit_map": [ + 0.014510294422507286, + 0.015769047662615776, + 0.02353227511048317, + 0.016753925010561943, + 0.014523192308843136, + 0.019511887803673744, + 0.01775694265961647, + 0.03246396780014038, + 0.005562849808484316, + 0.006340544205158949, + 0.008064701221883297, + 0.007930695079267025, + 0.008299488574266434, + 0.008192718029022217, + 0.009059696458280087, + 0.011072191409766674, + 0.018388861790299416, + 0.012990247458219528, + 0.0300576314330101, + 0.010125930421054363, + 0.016849203035235405, + 0.01788039691746235, + 0.022989487275481224, + 0.026119552552700043, + 0.009629538282752037, + 0.004952810239046812, + 0.008955349214375019, + 0.007970111444592476, + 0.005297958850860596, + 0.00827840343117714, + 0.010081933811306953, + 0.009334205649793148, + 0.008190998807549477, + 0.004226059187203646, + 0.008792807348072529, + 0.0036921873688697815, + 0.004631093703210354, + 0.003032282693311572, + 0.0038392546121031046, + 0.004209486767649651, + 0.007008382584899664, + 0.0064970278181135654, + 0.015004554763436317, + 0.01288258470594883, + 0.005761290900409222, + 0.012592102400958538, + 0.0067099700681865215, + 0.006764562334865332, + 0.010772407054901123, + 0.00957026518881321, + 0.010275861248373985, + 0.013493857346475124, + 0.008502782322466373, + 0.012379614636301994, + 0.01389667671173811, + 0.012875277549028397, + 0.007136118598282337, + 0.005903035402297974, + 0.011893903836607933, + 0.006596780382096767, + 0.007765899877995253, + 0.01737823896110058, + 0.010754330083727837, + 0.016193540766835213, + 0.00467766635119915, + 0.007140692789107561, + 0.0041700853034853935, + 0.006633163429796696, + 0.007715283427387476, + 0.010410962626338005, + 0.009748857468366623, + 0.009564584121108055, + 0.015132754109799862, + 0.02724618650972843, + 0.017989059910178185, + 0.01951044425368309, + 0.02797016128897667, + 0.029273608699440956, + 0.050422295928001404, + 0.023924972862005234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.004532302729785442 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202234335243702 + ], + "opposite_logprob_tokens": [ + -4.264202117919922 + ], + "logprob_reward": -0.014202234335243702, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202234335243702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09062439948320389, + 0.11418545991182327, + 0.38077807426452637, + 0.058839716017246246, + 0.06822466105222702, + 0.04354512691497803, + 0.05314496532082558, + 0.053095683455467224, + 0.05389964208006859, + 0.08366231620311737 + ], + "credit_map": [ + 0.008589434437453747, + 0.0073239426128566265, + 0.011835150420665741, + 0.008672913536429405, + 0.0076691582798957825, + 0.011600584723055363, + 0.013037026859819889, + 0.021896183490753174, + 0.014857192523777485, + 0.011682550422847271, + 0.017588481307029724, + 0.008232975378632545, + 0.011066216975450516, + 0.014724970795214176, + 0.013913829810917377, + 0.022119244560599327, + 0.04289783164858818, + 0.02620442397892475, + 0.07251382619142532, + 0.024442678317427635, + 0.038745902478694916, + 0.04234148934483528, + 0.05665791407227516, + 0.07697401195764542, + 0.005950375460088253, + 0.0054732440039515495, + 0.0073296367190778255, + 0.0066296556033194065, + 0.006360064726322889, + 0.008680175058543682, + 0.007809877395629883, + 0.010606685653328896, + 0.01213119551539421, + 0.004550138954073191, + 0.01527888048440218, + 0.0071302964352071285, + 0.005254156421869993, + 0.0066416822373867035, + 0.007475476246327162, + 0.009762831032276154, + 0.0039466870948672295, + 0.005216785706579685, + 0.006319617852568626, + 0.004699693992733955, + 0.005685732234269381, + 0.007450273260474205, + 0.0033691669814288616, + 0.006857169792056084, + 0.005054526496678591, + 0.004594685975462198, + 0.006509074475616217, + 0.007701625116169453, + 0.006572916638106108, + 0.007300442550331354, + 0.010068721137940884, + 0.005342974327504635, + 0.005534864030778408, + 0.003457220969721675, + 0.010235664434731007, + 0.004773005843162537, + 0.003944545518606901, + 0.00890213344246149, + 0.006455804221332073, + 0.009792448952794075, + 0.004279879853129387, + 0.004095100797712803, + 0.0066518099047243595, + 0.005245319101959467, + 0.005764875095337629, + 0.008653544820845127, + 0.010268690064549446, + 0.008940423838794231, + 0.006495946552604437, + 0.011045683175325394, + 0.007290847133845091, + 0.007112330291420221, + 0.011582237668335438, + 0.012082938104867935, + 0.01821748912334442, + 0.009834839962422848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.014202234335243702 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955332424491644 + ], + "opposite_logprob_tokens": [ + -5.130955219268799 + ], + "logprob_reward": -0.005955332424491644, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955332424491644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12527908384799957, + 0.14067348837852478, + 0.07420855760574341, + 0.046978142112493515, + 0.041022758930921555, + 0.12910720705986023, + 0.238177090883255, + 0.04158943518996239, + 0.05347287654876709, + 0.10949131101369858 + ], + "credit_map": [ + 0.012147943489253521, + 0.009674119763076305, + 0.017358481884002686, + 0.016706081107258797, + 0.013480949215590954, + 0.01893935166299343, + 0.020104993134737015, + 0.01686716079711914, + 0.01060743909329176, + 0.016178669407963753, + 0.017555419355630875, + 0.02047954685986042, + 0.017096681520342827, + 0.016192015260457993, + 0.02278771996498108, + 0.01977599784731865, + 0.006597067229449749, + 0.004996289499104023, + 0.012159133329987526, + 0.008276269771158695, + 0.008104893378913403, + 0.009696388617157936, + 0.010234626941382885, + 0.014143888838589191, + 0.006887352559715509, + 0.004905857611447573, + 0.007127274759113789, + 0.0058244322426617146, + 0.00503368629142642, + 0.005312792491167784, + 0.0073479595594108105, + 0.004538784734904766, + 0.006089986767619848, + 0.003144372021779418, + 0.007596656680107117, + 0.006125273182988167, + 0.004310326185077429, + 0.0029702559113502502, + 0.005121334455907345, + 0.0056645567528903484, + 0.013058103621006012, + 0.011547700501978397, + 0.029235046356916428, + 0.02169477753341198, + 0.008412864990532398, + 0.021483322605490685, + 0.011759128421545029, + 0.011916271410882473, + 0.03157692775130272, + 0.021583065390586853, + 0.031756967306137085, + 0.02488601580262184, + 0.017161041498184204, + 0.034793321043252945, + 0.033491987735033035, + 0.04292777553200722, + 0.003021431853994727, + 0.0034794153179973364, + 0.005014644470065832, + 0.004581283777952194, + 0.0050127641297876835, + 0.0060837529599666595, + 0.00678612757474184, + 0.007610013708472252, + 0.0031044071074575186, + 0.0051335436291992664, + 0.006029210053384304, + 0.005369921214878559, + 0.0063104815781116486, + 0.008363503031432629, + 0.00866170134395361, + 0.010500111617147923, + 0.009091264568269253, + 0.012344153597950935, + 0.012777810916304588, + 0.008390583097934723, + 0.01337861642241478, + 0.014245858415961266, + 0.023821713402867317, + 0.015441321767866611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.005955332424491644 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038097646087408066 + ], + "opposite_logprob_tokens": [ + -3.288097620010376 + ], + "logprob_reward": -0.038097646087408066, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038097646087408066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24416989088058472, + 0.10854151099920273, + 0.11975295841693878, + 0.07042673230171204, + 0.06620092689990997, + 0.09073129296302795, + 0.0836348831653595, + 0.07294196635484695, + 0.04733607545495033, + 0.09626380354166031 + ], + "credit_map": [ + 0.019365491345524788, + 0.02484414540231228, + 0.03543667495250702, + 0.026691274717450142, + 0.01820250041782856, + 0.031108496710658073, + 0.033999789506196976, + 0.054521508514881134, + 0.009508001618087292, + 0.009642278775572777, + 0.015300258062779903, + 0.009112424217164516, + 0.011069405823946, + 0.015537387691438198, + 0.014708233997225761, + 0.023663528263568878, + 0.012906176969408989, + 0.008140930905938148, + 0.01993160881102085, + 0.013989166356623173, + 0.011278861202299595, + 0.012443449348211288, + 0.019127516075968742, + 0.021935254335403442, + 0.008856824599206448, + 0.005154549144208431, + 0.011284979991614819, + 0.006826841738075018, + 0.0077345166355371475, + 0.008720638230443, + 0.009100847877562046, + 0.012747534550726414, + 0.009201024658977985, + 0.003883355064317584, + 0.01138913445174694, + 0.008082319982349873, + 0.005685412324965, + 0.00828186608850956, + 0.009048178791999817, + 0.010629639960825443, + 0.0079611511901021, + 0.006851147394627333, + 0.015648989006876945, + 0.010960446670651436, + 0.009482940658926964, + 0.018511813133955002, + 0.008322000503540039, + 0.012992802076041698, + 0.008802271448075771, + 0.007442362140864134, + 0.011790783144533634, + 0.011499127373099327, + 0.010046918876469135, + 0.010856453329324722, + 0.014788472093641758, + 0.008408491499722004, + 0.008256146684288979, + 0.00743032107129693, + 0.01315153669565916, + 0.006905458867549896, + 0.006897307932376862, + 0.010503089986741543, + 0.009835894219577312, + 0.009962213225662708, + 0.0032374050933867693, + 0.004387143533676863, + 0.00381570216268301, + 0.004307324532419443, + 0.006073890719562769, + 0.009066866710782051, + 0.0070350514724850655, + 0.009412693791091442, + 0.007048777304589748, + 0.00709593715146184, + 0.011137191206216812, + 0.008885172195732594, + 0.011149387806653976, + 0.014079458080232143, + 0.021864430978894234, + 0.015003449283540249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.038097646087408066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233231782913208 + ], + "opposite_logprob_tokens": [ + -0.8983231782913208 + ], + "logprob_reward": -0.5233231782913208, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233231782913208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23108907043933868, + 0.08508637547492981, + 0.06095137074589729, + 0.07242337614297867, + 0.03227875754237175, + 0.10735615342855453, + 0.165165975689888, + 0.14987730979919434, + 0.04237984120845795, + 0.05339174345135689 + ], + "credit_map": [ + 0.017206868156790733, + 0.02081102877855301, + 0.0309680737555027, + 0.02802395634353161, + 0.01862204447388649, + 0.03376306965947151, + 0.03413613513112068, + 0.04755789041519165, + 0.008363950066268444, + 0.009989899583160877, + 0.011571677401661873, + 0.010396876372396946, + 0.008144230581820011, + 0.009894969873130322, + 0.012087173759937286, + 0.014637596905231476, + 0.006133549846708775, + 0.004288988187909126, + 0.008466952480375767, + 0.005019251722842455, + 0.007745963055640459, + 0.007528423797339201, + 0.012597158551216125, + 0.009171092882752419, + 0.010163734667003155, + 0.005585996899753809, + 0.010750239714980125, + 0.008764529600739479, + 0.007183026522397995, + 0.008087223395705223, + 0.011834532022476196, + 0.010054102167487144, + 0.005576356779783964, + 0.004834172315895557, + 0.005793026648461819, + 0.00392119400203228, + 0.0050912881270051, + 0.0017026180867105722, + 0.0033783786930143833, + 0.0019817252177745104, + 0.011020652949810028, + 0.011627299711108208, + 0.01685544289648533, + 0.015090343542397022, + 0.011745678260922432, + 0.014197858050465584, + 0.011932050809264183, + 0.014886838383972645, + 0.01852453127503395, + 0.014449352398514748, + 0.016348635777831078, + 0.024788007140159607, + 0.012259731069207191, + 0.02369469590485096, + 0.021861858665943146, + 0.03323918953537941, + 0.014622323215007782, + 0.009707814082503319, + 0.027699075639247894, + 0.011210673488676548, + 0.010002132505178452, + 0.02838721312582493, + 0.018771644681692123, + 0.029476439580321312, + 0.003957084845751524, + 0.0033301720395684242, + 0.007070562802255154, + 0.0036216862499713898, + 0.004073651973158121, + 0.006753135472536087, + 0.00619782879948616, + 0.007375721819698811, + 0.005729478318244219, + 0.00512673519551754, + 0.0067588044330477715, + 0.004156796261668205, + 0.006608738098293543, + 0.007505554705858231, + 0.009513488970696926, + 0.007992152124643326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.5233231782913208 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, smooth 360-degree tracking shot focuses in close-up on a lone, burning antique oil lamp, nestled in the center of a rough-grained, dark wooden table, surrounded by an endless, bottomless darkness. The lamp boasts a brass base with a subtle patina, a clear glass base half-filled with golden oil, and a braided cotton wick tipped with a fiery orange-red glow. A teardrop-shaped flame flickers steadily and gently within the slender glass shade, providing the scene's sole source of light and casting a warm, soft glow. As the camera steadily circles the lamp, light refracts across the glass shade in fluid patterns, while the brass highlights on the base glide gracefully across the scene, and the soft glow seems to dance within the darkness, creating an atmosphere of tranquility, contemplation, and profound intimacy.", + "seed": 26339, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_4/index_8956_seed_26339_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3307356562931091, + "vlm_logprob": -0.3307356562931091, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.3307356562931091, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002893192460760474 + ], + "opposite_logprob_tokens": [ + -6.377892971038818 + ], + "logprob_reward": -0.002893192460760474, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.002893192460760474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21114376187324524, + 0.17860807478427887, + 0.09211856126785278, + 0.060908298939466476, + 0.06356584280729294, + 0.11770026385784149, + 0.06764624267816544, + 0.0657188668847084, + 0.0921844094991684, + 0.05040567368268967 + ], + "credit_map": [ + 0.026320796459913254, + 0.024258596822619438, + 0.027110012248158455, + 0.025444811210036278, + 0.02854722924530506, + 0.036901913583278656, + 0.01930217444896698, + 0.023258216679096222, + 0.014788557775318623, + 0.007750627119094133, + 0.053409427404403687, + 0.007628301158547401, + 0.010686795227229595, + 0.0573880709707737, + 0.009740946814417839, + 0.017215345054864883, + 0.015866832807660103, + 0.006435512565076351, + 0.01336752064526081, + 0.006955713499337435, + 0.007209769915789366, + 0.014223079197108746, + 0.00902419164776802, + 0.019035939127206802, + 0.008579322136938572, + 0.0046813297085464, + 0.009147371165454388, + 0.006967480760067701, + 0.004541642498224974, + 0.008165949955582619, + 0.007719992194324732, + 0.011105209589004517, + 0.00581618444994092, + 0.005543479695916176, + 0.007781181484460831, + 0.00809694454073906, + 0.006495224311947823, + 0.00845969095826149, + 0.012110374867916107, + 0.00926276110112667, + 0.013132589869201183, + 0.008231520652770996, + 0.023619499057531357, + 0.010131499730050564, + 0.010167790576815605, + 0.018025359138846397, + 0.011217637918889523, + 0.02317437157034874, + 0.006099203135818243, + 0.004041654523462057, + 0.015115596354007721, + 0.005825139116495848, + 0.005069496110081673, + 0.00966116413474083, + 0.006870670709758997, + 0.014963316731154919, + 0.008332476019859314, + 0.006362780928611755, + 0.010236033238470554, + 0.006012678612023592, + 0.006087651476264, + 0.010249736718833447, + 0.008390188217163086, + 0.010047316551208496, + 0.00920816045254469, + 0.011506573297083378, + 0.015713348984718323, + 0.009598985314369202, + 0.01051606610417366, + 0.012842841446399689, + 0.011178253218531609, + 0.011620179750025272, + 0.005103064700961113, + 0.004139701835811138, + 0.008959047496318817, + 0.005606331862509251, + 0.005417999345809221, + 0.0062976558692753315, + 0.008213481865823269, + 0.006668386049568653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.002893192460760474 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009791914373636246 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.009791914373636246, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791914373636246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.205612912774086, + 0.19422270357608795, + 0.07630803436040878, + 0.07416021823883057, + 0.0347866490483284, + 0.07550480961799622, + 0.06373248249292374, + 0.16744959354400635, + 0.04305071383714676, + 0.06517191231250763 + ], + "credit_map": [ + 0.025408288463950157, + 0.024529647082090378, + 0.032814037054777145, + 0.021451609209179878, + 0.021780647337436676, + 0.03246687725186348, + 0.021381057798862457, + 0.02578074298799038, + 0.013609880581498146, + 0.005727515555918217, + 0.06817423552274704, + 0.007659127935767174, + 0.006373643409460783, + 0.07092990726232529, + 0.008006161078810692, + 0.013742215931415558, + 0.012328301556408405, + 0.00439553614705801, + 0.011176805011928082, + 0.006148919463157654, + 0.005666046869009733, + 0.0118258623406291, + 0.006329703610390425, + 0.018436850979924202, + 0.012601302936673164, + 0.004723490681499243, + 0.012598051689565182, + 0.007470482960343361, + 0.0037788888439536095, + 0.01079216692596674, + 0.007082364987581968, + 0.015113472938537598, + 0.0028267812449485064, + 0.002779391361400485, + 0.006563690956681967, + 0.004528989549726248, + 0.002142286626622081, + 0.005091165658086538, + 0.0059509859420359135, + 0.004903357941657305, + 0.008677175268530846, + 0.003937297035008669, + 0.019641758874058723, + 0.0074433814734220505, + 0.0036631731782108545, + 0.011613995768129826, + 0.006291387137025595, + 0.014236637391149998, + 0.006396705284714699, + 0.0054922993294894695, + 0.01196822989732027, + 0.004380722995847464, + 0.006344085093587637, + 0.010755548253655434, + 0.004729451145976782, + 0.013665437698364258, + 0.018433701246976852, + 0.011189640499651432, + 0.04252159968018532, + 0.01630985550582409, + 0.011204813607037067, + 0.023060327395796776, + 0.017887957394123077, + 0.02684169076383114, + 0.005980341229587793, + 0.0030161270406097174, + 0.008441584184765816, + 0.004235406406223774, + 0.0035414693411439657, + 0.006338835693895817, + 0.005050444509834051, + 0.006446505431085825, + 0.005524339620023966, + 0.007802025880664587, + 0.00862849224358797, + 0.008388862013816833, + 0.009491996839642525, + 0.008017432875931263, + 0.009996495209634304, + 0.007322271820157766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.009791914373636246 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067427014000713825 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.0067427014000713825, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067427014000713825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22915653884410858, + 0.15536823868751526, + 0.12825047969818115, + 0.09619181603193283, + 0.08912274986505508, + 0.10547055304050446, + 0.048485830426216125, + 0.059613946825265884, + 0.04300864785909653, + 0.045331213623285294 + ], + "credit_map": [ + 0.03232337906956673, + 0.0265534445643425, + 0.03141224384307861, + 0.017125120386481285, + 0.019608324393630028, + 0.05268266424536705, + 0.022996529936790466, + 0.026454849168658257, + 0.0139952776953578, + 0.008341765962541103, + 0.03861008584499359, + 0.015312704257667065, + 0.009323062375187874, + 0.04054905101656914, + 0.013936030678451061, + 0.015300259925425053, + 0.011888175271451473, + 0.010741732083261013, + 0.01616174727678299, + 0.018064692616462708, + 0.015389426611363888, + 0.017490575090050697, + 0.019397491589188576, + 0.019116636365652084, + 0.011359747499227524, + 0.006833204533904791, + 0.013120263814926147, + 0.014073261059820652, + 0.007785188499838114, + 0.01391063816845417, + 0.01206548698246479, + 0.017044028267264366, + 0.008087702095508575, + 0.006412445101886988, + 0.013764708302915096, + 0.011931140907108784, + 0.006608081050217152, + 0.011380494572222233, + 0.018048997968435287, + 0.012889180332422256, + 0.011676526628434658, + 0.006037580315023661, + 0.025962267071008682, + 0.00965658575296402, + 0.007010006345808506, + 0.016182634979486465, + 0.011741889640688896, + 0.0172030758112669, + 0.005465505179017782, + 0.004115162882953882, + 0.009234520606696606, + 0.0051506515592336655, + 0.003906634170562029, + 0.006401667837053537, + 0.005306505132466555, + 0.008905190043151379, + 0.0066518355160951614, + 0.0041291736997663975, + 0.011933842673897743, + 0.007431181613355875, + 0.0046238647773861885, + 0.00806170329451561, + 0.008174704387784004, + 0.008607646450400352, + 0.0051981681026518345, + 0.004145734943449497, + 0.006854669190943241, + 0.005218814592808485, + 0.004662501625716686, + 0.005566902458667755, + 0.00542801758274436, + 0.00593384075909853, + 0.004375963471829891, + 0.006074943114072084, + 0.008377369493246078, + 0.005885623395442963, + 0.0052715688943862915, + 0.004782784264534712, + 0.005223662592470646, + 0.0053392997942864895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0067427014000713825 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033696603029966354 + ], + "opposite_logprob_tokens": [ + -3.4086966514587402 + ], + "logprob_reward": -0.033696603029966354, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033696603029966354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13090112805366516, + 0.16051320731639862, + 0.21512925624847412, + 0.08169879764318466, + 0.05252763628959656, + 0.1205718144774437, + 0.117202527821064, + 0.05175212025642395, + 0.04311095178127289, + 0.02659250609576702 + ], + "credit_map": [ + 0.01883539743721485, + 0.008609055541455746, + 0.021542003378272057, + 0.010572344064712524, + 0.007306297775357962, + 0.029420582577586174, + 0.010331458412110806, + 0.02428400330245495, + 0.011242940090596676, + 0.005416477099061012, + 0.05180804058909416, + 0.008137308992445469, + 0.00666362838819623, + 0.05061936751008034, + 0.009105226024985313, + 0.01752023585140705, + 0.02959086187183857, + 0.013122525066137314, + 0.033427644520998, + 0.016690177842974663, + 0.015561496838927269, + 0.036292869597673416, + 0.018280332908034325, + 0.052163366228342056, + 0.011630374938249588, + 0.004757017828524113, + 0.015589874237775803, + 0.008636965416371822, + 0.005691057071089745, + 0.012881513684988022, + 0.005353023298084736, + 0.017158985137939453, + 0.006452995352447033, + 0.0027212861459702253, + 0.012026402167975903, + 0.005319939460605383, + 0.002834080485627055, + 0.00811670906841755, + 0.004260017070919275, + 0.010796211659908295, + 0.014943040907382965, + 0.006985448766499758, + 0.02695256657898426, + 0.011999106965959072, + 0.006706968881189823, + 0.01919281855225563, + 0.01149026583880186, + 0.02230161428451538, + 0.013244465924799442, + 0.006553386338055134, + 0.02185055986046791, + 0.009253428317606449, + 0.008906352333724499, + 0.01952311024069786, + 0.01104313600808382, + 0.0268281027674675, + 0.0046891928650438786, + 0.0042626261711120605, + 0.008365731686353683, + 0.006328113377094269, + 0.004650588147342205, + 0.006368226837366819, + 0.010843535885214806, + 0.0062441136687994, + 0.004447664134204388, + 0.0028130922000855207, + 0.009290806017816067, + 0.004356094636023045, + 0.0038503229152411222, + 0.00595638295635581, + 0.0058275312185287476, + 0.006569061893969774, + 0.002878146944567561, + 0.0020265989005565643, + 0.005158367566764355, + 0.0031446977518498898, + 0.002073274226859212, + 0.003703410504385829, + 0.0036666214000433683, + 0.0039413925260305405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.033696603029966354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6005538702011108 + ], + "opposite_logprob_tokens": [ + -0.22555382549762726 + ], + "logprob_reward": -1.6005538702011108, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6005538702011108, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13431042432785034, + 0.15104377269744873, + 0.06073051318526268, + 0.051282986998558044, + 0.06800235062837601, + 0.22933825850486755, + 0.15549446642398834, + 0.04880431294441223, + 0.03579661622643471, + 0.06519635021686554 + ], + "credit_map": [ + 0.014589622616767883, + 0.014529932290315628, + 0.021381963044404984, + 0.018970414996147156, + 0.015727929770946503, + 0.01522212103009224, + 0.01871437579393387, + 0.01517405267804861, + 0.0171898752450943, + 0.007575811352580786, + 0.03658609464764595, + 0.009527760557830334, + 0.007380578201264143, + 0.04286930337548256, + 0.009777557104825974, + 0.020136769860982895, + 0.007826350629329681, + 0.004680757410824299, + 0.009432876482605934, + 0.00636420538648963, + 0.008691652677953243, + 0.008989622816443443, + 0.007687895558774471, + 0.007057148963212967, + 0.005131058860570192, + 0.004444647580385208, + 0.009555479511618614, + 0.005966308526694775, + 0.00382353481836617, + 0.007170887663960457, + 0.006700109224766493, + 0.008490958251059055, + 0.006487613078206778, + 0.004702240228652954, + 0.008602723479270935, + 0.010426375083625317, + 0.004039134364575148, + 0.009932560846209526, + 0.012545431032776833, + 0.011266267858445644, + 0.025028064846992493, + 0.011399652808904648, + 0.05564900487661362, + 0.022471392527222633, + 0.010874190367758274, + 0.039068419486284256, + 0.021960850805044174, + 0.04288667440414429, + 0.016131609678268433, + 0.009945188648998737, + 0.032011959701776505, + 0.01411597989499569, + 0.012334386818110943, + 0.02094913460314274, + 0.01721096783876419, + 0.032795220613479614, + 0.0045192651450634, + 0.0034148225095123053, + 0.01109470147639513, + 0.00616961857303977, + 0.005118314642459154, + 0.0061286683194339275, + 0.006888982839882374, + 0.005469935480505228, + 0.003605680540204048, + 0.0051649706438183784, + 0.0048954300582408905, + 0.0037994899321347475, + 0.004695819690823555, + 0.004560395609587431, + 0.004821284674108028, + 0.0042535425163805485, + 0.007002280093729496, + 0.005097724497318268, + 0.012266571633517742, + 0.007148741278797388, + 0.005493759177625179, + 0.008765359409153461, + 0.009459461085498333, + 0.009962446056306362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -1.6005538702011108 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow, smooth 360-degree tracking shot focuses in close-up on a lone, burning antique oil lamp, nestled in the center of a rough-grained, dark wooden table, surrounded by an endless, bottomless darkness. The lamp boasts a brass base with a subtle patina, a clear glass base half-filled with golden oil, and a braided cotton wick tipped with a fiery orange-red glow. A teardrop-shaped flame flickers steadily and gently within the slender glass shade, providing the scene's sole source of light and casting a warm, soft glow. As the camera steadily circles the lamp, light refracts across the glass shade in fluid patterns, while the brass highlights on the base glide gracefully across the scene, and the soft glow seems to dance within the darkness, creating an atmosphere of tranquility, contemplation, and profound intimacy.", + "seed": 26340, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_4/index_8956_seed_26340_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14684647368267179, + "vlm_logprob": -0.14684647368267179, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.14684647368267179, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0050193252973258495 + ], + "opposite_logprob_tokens": [ + -5.505019187927246 + ], + "logprob_reward": -0.0050193252973258495, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0050193252973258495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2845078706741333, + 0.07649588584899902, + 0.07574544847011566, + 0.07149013876914978, + 0.09692499041557312, + 0.043113309890031815, + 0.06383703649044037, + 0.05661601573228836, + 0.16800053417682648, + 0.06326879560947418 + ], + "credit_map": [ + 0.026295101270079613, + 0.021398672834038734, + 0.052239105105400085, + 0.027392325922846794, + 0.029927488416433334, + 0.050862424075603485, + 0.04707707464694977, + 0.02931566908955574, + 0.007294845301657915, + 0.006805612239986658, + 0.01868654415011406, + 0.0062855747528374195, + 0.006951134651899338, + 0.01163752842694521, + 0.01100845355540514, + 0.007826191373169422, + 0.007991928607225418, + 0.006330293137580156, + 0.014067537151277065, + 0.007660986389964819, + 0.007940973155200481, + 0.010815371759235859, + 0.01216594036668539, + 0.008772410452365875, + 0.0079790148884058, + 0.0051401760429143906, + 0.014411048963665962, + 0.007590316701680422, + 0.005387915298342705, + 0.009907004423439503, + 0.012673573568463326, + 0.008401085622608662, + 0.01255984790623188, + 0.007471950259059668, + 0.014974330551922321, + 0.00542539544403553, + 0.008055121637880802, + 0.01602143608033657, + 0.00939930323511362, + 0.02301759459078312, + 0.0056327409110963345, + 0.0037197754718363285, + 0.005343312863260508, + 0.0046528410166502, + 0.004235346335917711, + 0.006105915177613497, + 0.007812761701643467, + 0.005610611755400896, + 0.007187091279774904, + 0.006316971965134144, + 0.00896451622247696, + 0.004766986705362797, + 0.006906807888299227, + 0.008284395560622215, + 0.010183778591454029, + 0.011226489208638668, + 0.00949318427592516, + 0.004448635969310999, + 0.008594600483775139, + 0.005202169064432383, + 0.005456829909235239, + 0.008135367184877396, + 0.007615852635353804, + 0.007669372018426657, + 0.02183978259563446, + 0.019835133105516434, + 0.020067088305950165, + 0.02231086790561676, + 0.016080185770988464, + 0.019840577617287636, + 0.0239484291523695, + 0.02407846972346306, + 0.00449420465156436, + 0.003216865938156843, + 0.00833873264491558, + 0.0033659152686595917, + 0.004828473087400198, + 0.015370945446193218, + 0.005336416885256767, + 0.01831723190844059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.0050193252973258495 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007637230679392815 + ], + "opposite_logprob_tokens": [ + -4.882637023925781 + ], + "logprob_reward": -0.007637230679392815, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007637230679392815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.115906722843647, + 0.08497645705938339, + 0.08337461948394775, + 0.07487349957227707, + 0.16302505135536194, + 0.037168875336647034, + 0.03404039889574051, + 0.18974840641021729, + 0.11316133290529251, + 0.10372461378574371 + ], + "credit_map": [ + 0.010560565628111362, + 0.00846646074205637, + 0.026563391089439392, + 0.010538649745285511, + 0.010569628328084946, + 0.022742997854948044, + 0.012401046231389046, + 0.014063992537558079, + 0.008773856796324253, + 0.0078044249676167965, + 0.013801224529743195, + 0.007925279438495636, + 0.011304184794425964, + 0.013761862181127071, + 0.014471713453531265, + 0.00713391974568367, + 0.00902873557060957, + 0.005637516267597675, + 0.01633482426404953, + 0.006118465214967728, + 0.007481975480914116, + 0.013234597630798817, + 0.014588402584195137, + 0.010950109921395779, + 0.007442697882652283, + 0.006118006072938442, + 0.015148919075727463, + 0.007077265065163374, + 0.006498800590634346, + 0.010031930170953274, + 0.013430250808596611, + 0.009125635027885437, + 0.02165026217699051, + 0.013689312152564526, + 0.018955474719405174, + 0.012253022752702236, + 0.009641814976930618, + 0.02588455192744732, + 0.020515749230980873, + 0.04043486714363098, + 0.004539507441222668, + 0.0032084654085338116, + 0.005411538761109114, + 0.00373802799731493, + 0.002876854268833995, + 0.006041552871465683, + 0.0038873953744769096, + 0.007465534843504429, + 0.0054473793134093285, + 0.0030309807043522596, + 0.005714902188628912, + 0.0032104933634400368, + 0.0030142767354846, + 0.004315487574785948, + 0.004363352432847023, + 0.004943531006574631, + 0.018779316917061806, + 0.015463249757885933, + 0.031496841460466385, + 0.013507209718227386, + 0.013689493760466576, + 0.029372703284025192, + 0.02680385671555996, + 0.04063574597239494, + 0.012146435678005219, + 0.009890061803162098, + 0.017265701666474342, + 0.006148969288915396, + 0.011450991965830326, + 0.01622832752764225, + 0.01656545139849186, + 0.023465394973754883, + 0.006122807040810585, + 0.003928287886083126, + 0.012434608303010464, + 0.0067012859508395195, + 0.004431599751114845, + 0.029478613287210464, + 0.008118055760860443, + 0.032509367913007736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.007637230679392815 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076371124014258385 + ], + "opposite_logprob_tokens": [ + -4.882637023925781 + ], + "logprob_reward": -0.0076371124014258385, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076371124014258385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.293876051902771, + 0.09794657677412033, + 0.07722587138414383, + 0.066710464656353, + 0.08963587880134583, + 0.04118858277797699, + 0.013326440006494522, + 0.04083364084362984, + 0.1857982724905014, + 0.09345825016498566 + ], + "credit_map": [ + 0.03232218325138092, + 0.02469060942530632, + 0.052044954150915146, + 0.02813802659511566, + 0.030720630660653114, + 0.0370657742023468, + 0.0602966845035553, + 0.028597192838788033, + 0.01113949902355671, + 0.010053033009171486, + 0.017738522961735725, + 0.009496526792645454, + 0.009641369804739952, + 0.013810926117002964, + 0.017076708376407623, + 0.008989990688860416, + 0.0087321437895298, + 0.004500048700720072, + 0.015433068387210369, + 0.007630625274032354, + 0.005979057401418686, + 0.010673514567315578, + 0.013823054730892181, + 0.010454362258315086, + 0.005204612389206886, + 0.006150118540972471, + 0.01034617517143488, + 0.006112935487180948, + 0.005691223777830601, + 0.010626977309584618, + 0.013649353757500648, + 0.008929066359996796, + 0.010043784976005554, + 0.008086812682449818, + 0.013165842741727829, + 0.00726664112880826, + 0.006745737046003342, + 0.012685013934969902, + 0.01286451518535614, + 0.018777525052428246, + 0.006294757127761841, + 0.0023487238213419914, + 0.0043493881821632385, + 0.0032806277740746737, + 0.002872968092560768, + 0.006799614056944847, + 0.00787257868796587, + 0.00736992247402668, + 0.001814887742511928, + 0.0009512498509138823, + 0.0024814412463456392, + 0.0010457272874191403, + 0.001346632489003241, + 0.0018122742185369134, + 0.002101453486829996, + 0.001772773452103138, + 0.005898095201700926, + 0.0027633346617221832, + 0.006337074562907219, + 0.00366639974527061, + 0.0030825738795101643, + 0.005569691304117441, + 0.00592460623010993, + 0.0075918640941381454, + 0.02141520567238331, + 0.01622033677995205, + 0.02706718258559704, + 0.01252546813338995, + 0.01688552275300026, + 0.02863517962396145, + 0.028627201914787292, + 0.034422170370817184, + 0.008712024427950382, + 0.01377124059945345, + 0.010457905940711498, + 0.011279764585196972, + 0.014296132139861584, + 0.01039023045450449, + 0.012524382211267948, + 0.012026575393974781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0076371124014258385 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020582424476742744 + ], + "opposite_logprob_tokens": [ + -3.895582437515259 + ], + "logprob_reward": -0.020582424476742744, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020582424476742744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17061403393745422, + 0.22557517886161804, + 0.04839165136218071, + 0.0939018726348877, + 0.09242038428783417, + 0.056281380355358124, + 0.025115318596363068, + 0.0743250772356987, + 0.16583894193172455, + 0.04753618314862251 + ], + "credit_map": [ + 0.019267594441771507, + 0.01359565369784832, + 0.04068690538406372, + 0.01846807450056076, + 0.01574050262570381, + 0.028961528092622757, + 0.016218408942222595, + 0.017675356939435005, + 0.02596959099173546, + 0.019603455439209938, + 0.04448791220784187, + 0.018387965857982635, + 0.025643255561590195, + 0.04013396427035332, + 0.029354184865951538, + 0.02199486829340458, + 0.0051271566189825535, + 0.003969624638557434, + 0.008861958980560303, + 0.004073357209563255, + 0.006708213593810797, + 0.0066949487663805485, + 0.008533215150237083, + 0.00442317733541131, + 0.008543465286493301, + 0.008842828683555126, + 0.019854238256812096, + 0.009165862575173378, + 0.00738029507920146, + 0.012786670587956905, + 0.017336469143629074, + 0.009992044419050217, + 0.01149265468120575, + 0.007054522167891264, + 0.010889156721532345, + 0.008264725096523762, + 0.005340959411114454, + 0.015690989792346954, + 0.014859438873827457, + 0.01882794126868248, + 0.009526368230581284, + 0.0037739858962595463, + 0.006339403335005045, + 0.005442447494715452, + 0.003931492101401091, + 0.009193873964250088, + 0.006719131488353014, + 0.011354676447808743, + 0.0036406482104212046, + 0.0021722721867263317, + 0.004139693919569254, + 0.002174098277464509, + 0.002865806221961975, + 0.00292815244756639, + 0.003697398817166686, + 0.00349724804982543, + 0.009066152386367321, + 0.0057507832534611225, + 0.009645272046327591, + 0.005122562404721975, + 0.005187992937862873, + 0.012685896828770638, + 0.011859447695314884, + 0.015006967820227146, + 0.016745271161198616, + 0.014135485514998436, + 0.024254940450191498, + 0.01304667443037033, + 0.015652937814593315, + 0.022383423522114754, + 0.02578621171414852, + 0.033834002912044525, + 0.0036181665491312742, + 0.0035823818761855364, + 0.0040497067384421825, + 0.0031084944494068623, + 0.004275545012205839, + 0.01277689915150404, + 0.0035273798275738955, + 0.012597611173987389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.020582424476742744 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone, burning antique oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lamp situated on a rough-grained, dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lamp feature a brass base with a subtle patina and a clear glass base half-filled with golden oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an endless, bottomless darkness with the lamp as the sole source of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth 360-degree tracking shot circling the lamp?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933562755584717 + ], + "opposite_logprob_tokens": [ + -0.6933562755584717 + ], + "logprob_reward": -0.6933562755584717, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933562755584717, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2691781520843506, + 0.09177687019109726, + 0.1201060563325882, + 0.04146244004368782, + 0.07198773324489594, + 0.0244564451277256, + 0.03146180510520935, + 0.0706302672624588, + 0.21124309301376343, + 0.06769721955060959 + ], + "credit_map": [ + 0.02676123008131981, + 0.01686611957848072, + 0.05289001017808914, + 0.03080090321600437, + 0.022646257653832436, + 0.03778480365872383, + 0.0570228174328804, + 0.024405989795923233, + 0.009947143495082855, + 0.006591000594198704, + 0.01792503334581852, + 0.01055867038667202, + 0.008246072567999363, + 0.01534047070890665, + 0.012502040714025497, + 0.010666434653103352, + 0.014873548410832882, + 0.006632756907492876, + 0.025120627135038376, + 0.010835093446075916, + 0.010300973430275917, + 0.01502846460789442, + 0.02075105905532837, + 0.016563525423407555, + 0.0050525846891105175, + 0.002878987928852439, + 0.010601741261780262, + 0.0026507533621042967, + 0.0037707060109823942, + 0.0064123282209038734, + 0.005179977510124445, + 0.004915355704724789, + 0.006756452843546867, + 0.008081238716840744, + 0.008959897793829441, + 0.007461228873580694, + 0.006815241649746895, + 0.010305886156857014, + 0.012469405308365822, + 0.011138376779854298, + 0.002252003876492381, + 0.002811056561768055, + 0.00402424531057477, + 0.0022600325755774975, + 0.002689582761377096, + 0.0034959500189870596, + 0.0037403882015496492, + 0.0031831837259233, + 0.0045521133579313755, + 0.003593084402382374, + 0.005092803388834, + 0.003197294194251299, + 0.0035629698541015387, + 0.0037856402341276407, + 0.0033915843814611435, + 0.004286314360797405, + 0.011820223182439804, + 0.004734647460281849, + 0.008160854689776897, + 0.0047190492041409016, + 0.004754489753395319, + 0.011758279986679554, + 0.009544199332594872, + 0.015138518065214157, + 0.026815585792064667, + 0.01724088564515114, + 0.029307864606380463, + 0.015956446528434753, + 0.018624650314450264, + 0.03231387957930565, + 0.03309137374162674, + 0.03789239004254341, + 0.004960619378834963, + 0.004804232623428106, + 0.007809833157807589, + 0.003750023664906621, + 0.00560269458219409, + 0.01779751107096672, + 0.004481224343180656, + 0.018491076305508614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.6933562755584717 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_5/index_tensor([1402, 1402])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_5/index_tensor([1402, 1402])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..983f40c52dfaa0a80ea802ec48ae57c2829a6c56 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_5/index_tensor([1402, 1402])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long shot, shot from an upward angle slightly below eye level, slowly pans to the right and zooms out slightly to capture the grandeur and dynamism of the entire scene. In a magnificent, ancient Eastern palace, towering vermilion pillars support a dome of carved beams and painted rafters. A woman, clad in a gorgeous red dress in the style of an ancient palace, dances with rapt attention. Her cascading, jet-black hair is held high by a carved jade hairpin, strands of which gently brush against her classical, serene face as she dances. Her red dress, crafted from fine silk, gleams softly in the dim light, its wide sleeves embroidered with intricate golden phoenix motifs. She is in the midst of an elegant twirl, her arms extended and her body lithe, driving the multi-tiered, immense skirt to steadily and gracefully unfold outward from her center. The golden thread at the hem traces a brilliant arc in the air, resembling a giant, blooming red peony. The light mainly comes from the ancient palace lantern hanging in the corner, casting a warm and swaying halo, which is intertwined with the cold moonlight shining through the huge window panes deep in the background. The black marble floor, polished as smooth as a mirror, clearly reflects the dancing red figure of the woman and the flowing light and shadow of her skirt, creating an atmosphere that is solemn, elegant, romantic and mysterious.", + "seed": 4132, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_5/index_1402_seed_4132_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10478522973135114, + "vlm_logprob": -0.10478522973135114, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.10478522973135114, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13745738565921783 + ], + "opposite_logprob_tokens": [ + -2.1374573707580566 + ], + "logprob_reward": -0.13745738565921783, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13745738565921783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16029612720012665, + 0.13577239215373993, + 0.06609129160642624, + 0.05115943402051926, + 0.06300050020217896, + 0.1564566195011139, + 0.16553817689418793, + 0.04136988893151283, + 0.10572796314954758, + 0.05458761379122734 + ], + "credit_map": [ + 0.023851092904806137, + 0.02490123175084591, + 0.02180897444486618, + 0.024925339967012405, + 0.019914090633392334, + 0.01106434129178524, + 0.02125214971601963, + 0.012578914873301983, + 0.013702101074159145, + 0.018383905291557312, + 0.022257018834352493, + 0.03202424570918083, + 0.00969095341861248, + 0.007667563855648041, + 0.014036030508577824, + 0.0180105771869421, + 0.010000945068895817, + 0.007414135150611401, + 0.008972697891294956, + 0.009770601987838745, + 0.0070883180014789104, + 0.004512930754572153, + 0.008177557960152626, + 0.01015410665422678, + 0.005022519733756781, + 0.0052131167612969875, + 0.008625525049865246, + 0.011760182678699493, + 0.004973506089299917, + 0.0026470664888620377, + 0.005414402578026056, + 0.007503116503357887, + 0.009077227674424648, + 0.006858821492642164, + 0.01186306495219469, + 0.008736460469663143, + 0.005881338845938444, + 0.0054403250105679035, + 0.0062356614507734776, + 0.008907593786716461, + 0.027375349774956703, + 0.01939946971833706, + 0.021695846691727638, + 0.026406990364193916, + 0.010674813762307167, + 0.009952244348824024, + 0.020945480093359947, + 0.020006420090794563, + 0.017004050314426422, + 0.013739597983658314, + 0.03267797455191612, + 0.019748281687498093, + 0.01712554320693016, + 0.00804163608700037, + 0.023422669619321823, + 0.03377842530608177, + 0.0035292513202875853, + 0.003404514165595174, + 0.007236277684569359, + 0.005907114129513502, + 0.005254046525806189, + 0.0028122765943408012, + 0.007435264997184277, + 0.005791144445538521, + 0.008651330135762691, + 0.012907750904560089, + 0.015592964366078377, + 0.013115392997860909, + 0.015272164717316628, + 0.008559923619031906, + 0.013365332037210464, + 0.01826310344040394, + 0.005611873231828213, + 0.004755641333758831, + 0.010662048123776913, + 0.007580921985208988, + 0.004860057961195707, + 0.004012066405266523, + 0.007218000944703817, + 0.00988700333982706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.13745738565921783 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001960024470463395 + ], + "opposite_logprob_tokens": [ + -6.251959800720215 + ], + "logprob_reward": -0.001960024470463395, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001960024470463395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11339972168207169, + 0.07908746600151062, + 0.029356159269809723, + 0.04616999253630638, + 0.035934481769800186, + 0.13167504966259003, + 0.14497070014476776, + 0.07652873545885086, + 0.18418827652931213, + 0.15868940949440002 + ], + "credit_map": [ + 0.0135046960785985, + 0.016874579712748528, + 0.01703525334596634, + 0.016980240121483803, + 0.010954374447464943, + 0.007016986142843962, + 0.016212334856390953, + 0.014821261167526245, + 0.008835643529891968, + 0.012161172926425934, + 0.009637192822992802, + 0.01951337605714798, + 0.006462737452238798, + 0.003952221944928169, + 0.00873253308236599, + 0.00979259330779314, + 0.00322621944360435, + 0.0034601246006786823, + 0.004706230014562607, + 0.004431341774761677, + 0.0022515670862048864, + 0.0019960205536335707, + 0.005611203145235777, + 0.003673451254144311, + 0.0032044497784227133, + 0.005503668915480375, + 0.006440161727368832, + 0.008680837228894234, + 0.005887417588382959, + 0.002309094648808241, + 0.007307711988687515, + 0.006836647167801857, + 0.005484897177666426, + 0.0033006032463163137, + 0.0051269871182739735, + 0.0054688467644155025, + 0.003358731511980295, + 0.0033068074844777584, + 0.004643924068659544, + 0.0052436827681958675, + 0.024540364742279053, + 0.01646704599261284, + 0.020605474710464478, + 0.020155875012278557, + 0.008320572786033154, + 0.007714559324085712, + 0.016603924334049225, + 0.017267240211367607, + 0.011695913039147854, + 0.013677357695996761, + 0.022390836849808693, + 0.020490964874625206, + 0.016034923493862152, + 0.008532201871275902, + 0.02718926966190338, + 0.024959223344922066, + 0.008550764992833138, + 0.005643405485898256, + 0.010416632518172264, + 0.01165232714265585, + 0.013892009854316711, + 0.0042418972589075565, + 0.012956548482179642, + 0.009175151586532593, + 0.016249658539891243, + 0.02045746147632599, + 0.024366652593016624, + 0.02632882073521614, + 0.02426321990787983, + 0.013238981366157532, + 0.030424322932958603, + 0.02885916829109192, + 0.019428864121437073, + 0.014506138861179352, + 0.029622061178088188, + 0.017569107934832573, + 0.01626715064048767, + 0.010888401418924332, + 0.02084646001458168, + 0.029561230912804604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.001960024470463395 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021972341928631067 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.0021972341928631067, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021972341928631067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13324908912181854, + 0.054211925715208054, + 0.0479593425989151, + 0.07189520448446274, + 0.10305234044790268, + 0.040532615035772324, + 0.09133245795965195, + 0.1363602578639984, + 0.24701575934886932, + 0.07439100742340088 + ], + "credit_map": [ + 0.015447519719600677, + 0.0171101875603199, + 0.011645898222923279, + 0.017992792651057243, + 0.022611817345023155, + 0.008632674813270569, + 0.02264663390815258, + 0.017161574214696884, + 0.004159189760684967, + 0.00799108762294054, + 0.007641925476491451, + 0.010715063661336899, + 0.006349492818117142, + 0.0027319598011672497, + 0.007600194308906794, + 0.007023021578788757, + 0.0030493249651044607, + 0.0037200392689555883, + 0.006342118605971336, + 0.00863585527986288, + 0.005100082606077194, + 0.002440675627440214, + 0.01164268795400858, + 0.007028560619801283, + 0.004536577500402927, + 0.009001187048852444, + 0.010822287760674953, + 0.018631407991051674, + 0.006483541335910559, + 0.0026306896470487118, + 0.009450108744204044, + 0.010339407250285149, + 0.006067905109375715, + 0.012894030660390854, + 0.014699071645736694, + 0.01735578291118145, + 0.01674676686525345, + 0.0058390116319060326, + 0.013296599499881268, + 0.016153180971741676, + 0.004158888012170792, + 0.0037760960403829813, + 0.006083378102630377, + 0.007413804996758699, + 0.003969183191657066, + 0.003939624410122633, + 0.006298357155174017, + 0.004893288016319275, + 0.013463564217090607, + 0.007201971020549536, + 0.01570635847747326, + 0.011780254542827606, + 0.009161788038909435, + 0.005435294937342405, + 0.013831224292516708, + 0.014752007089555264, + 0.010370020754635334, + 0.012076576240360737, + 0.016991671174764633, + 0.023120226338505745, + 0.026715079322457314, + 0.006473409943282604, + 0.025855837389826775, + 0.01475745253264904, + 0.020501786842942238, + 0.026667142286896706, + 0.028395207598805428, + 0.04162139073014259, + 0.0468178316950798, + 0.015640802681446075, + 0.036545634269714355, + 0.03082597628235817, + 0.009376432746648788, + 0.006444693077355623, + 0.014363300055265427, + 0.008095595985651016, + 0.008045281283557415, + 0.0051261442713439465, + 0.008841182105243206, + 0.014098377898335457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0021972341928631067 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033700864762067795 + ], + "opposite_logprob_tokens": [ + -3.408700942993164 + ], + "logprob_reward": -0.033700864762067795, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033700864762067795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24419638514518738, + 0.11052584648132324, + 0.04576420783996582, + 0.04265328869223595, + 0.05637774243950844, + 0.19767259061336517, + 0.06378871947526932, + 0.04317754879593849, + 0.13682636618614197, + 0.05901720002293587 + ], + "credit_map": [ + 0.024286745116114616, + 0.021375274285674095, + 0.030474139377474785, + 0.04731409251689911, + 0.03901929035782814, + 0.015796739608049393, + 0.04290864244103432, + 0.023021502420306206, + 0.010502075776457787, + 0.017521701753139496, + 0.017165498808026314, + 0.02455405332148075, + 0.008334466256201267, + 0.006863160524517298, + 0.011080517433583736, + 0.014504377730190754, + 0.005064062774181366, + 0.008614908903837204, + 0.007911749184131622, + 0.004927413072437048, + 0.004788863472640514, + 0.003256687195971608, + 0.004234290681779385, + 0.00696623744443059, + 0.007111060433089733, + 0.004624944645911455, + 0.008472546003758907, + 0.0062838951125741005, + 0.0035672811791300774, + 0.0027990764938294888, + 0.0040005589835345745, + 0.005793931894004345, + 0.007849203422665596, + 0.006384282372891903, + 0.009214110672473907, + 0.008397024124860764, + 0.006237820256501436, + 0.004582198336720467, + 0.006072258576750755, + 0.00764085166156292, + 0.038464322686195374, + 0.023919090628623962, + 0.032610826194286346, + 0.029296372085809708, + 0.01171913556754589, + 0.01306422334164381, + 0.022860204800963402, + 0.025738419964909554, + 0.004330111667513847, + 0.004620714578777552, + 0.007632172666490078, + 0.011287203989923, + 0.005523866508156061, + 0.005035070702433586, + 0.015872111544013023, + 0.009487474337220192, + 0.005757101811468601, + 0.0036903603468090296, + 0.0062105171382427216, + 0.006683787796646357, + 0.003689985955134034, + 0.0026395884342491627, + 0.007265526335686445, + 0.007240684702992439, + 0.011691786348819733, + 0.020188918337225914, + 0.020694244652986526, + 0.017119748517870903, + 0.022105099633336067, + 0.010274635627865791, + 0.014243998564779758, + 0.020507944747805595, + 0.00539375701919198, + 0.005576803348958492, + 0.014967603608965874, + 0.006415567360818386, + 0.005122283939272165, + 0.0033619499299675226, + 0.006219269707798958, + 0.011959969997406006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.033700864762067795 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34861063957214355 + ], + "opposite_logprob_tokens": [ + -1.2236106395721436 + ], + "logprob_reward": -0.34861063957214355, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34861063957214355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16648611426353455, + 0.05093679949641228, + 0.05084966868162155, + 0.07981106638908386, + 0.026274874806404114, + 0.16308270394802094, + 0.07747290283441544, + 0.054526783525943756, + 0.1900462955236435, + 0.1405128687620163 + ], + "credit_map": [ + 0.019628679379820824, + 0.02149284817278385, + 0.01712690480053425, + 0.0246712788939476, + 0.025493016466498375, + 0.013155469670891762, + 0.026907848194241524, + 0.018010053783655167, + 0.004421883728355169, + 0.008320720866322517, + 0.006858981680124998, + 0.009439791552722454, + 0.005349330138415098, + 0.003157770959660411, + 0.00615093344822526, + 0.007237385027110577, + 0.0027049288619309664, + 0.004374313168227673, + 0.006879489403218031, + 0.00914947222918272, + 0.006101076491177082, + 0.0021964663174003363, + 0.012176917865872383, + 0.007266997825354338, + 0.006787458434700966, + 0.010036282241344452, + 0.011327503249049187, + 0.020787201821804047, + 0.00640453677624464, + 0.003103958908468485, + 0.010249261744320393, + 0.01111485343426466, + 0.00241275317966938, + 0.0027445403393357992, + 0.003694480750709772, + 0.004876612685620785, + 0.003893328597769141, + 0.0016452808631584048, + 0.003362322226166725, + 0.0036455539520829916, + 0.027563227340579033, + 0.020132048055529594, + 0.02431456372141838, + 0.02613602578639984, + 0.01242190133780241, + 0.010927907191216946, + 0.022132622078061104, + 0.019454387947916985, + 0.005519416648894548, + 0.006032578181475401, + 0.013828773982822895, + 0.013835246674716473, + 0.008280648849904537, + 0.004230987746268511, + 0.01181125920265913, + 0.013933991082012653, + 0.006501335185021162, + 0.004863452631980181, + 0.00607165414839983, + 0.008089995011687279, + 0.010533175431191921, + 0.0037207293789833784, + 0.007901164703071117, + 0.006845273543149233, + 0.017311999574303627, + 0.019744018092751503, + 0.025974201038479805, + 0.028932876884937286, + 0.019251426681876183, + 0.014394769445061684, + 0.025447506457567215, + 0.03898948058485985, + 0.020919745787978172, + 0.015230793505907059, + 0.02321578562259674, + 0.01593395136296749, + 0.01693752035498619, + 0.01334305852651596, + 0.015885744243860245, + 0.019046254456043243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.34861063957214355 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from an upward angle slightly below eye level, slowly pans to the right and zooms out slightly to capture the grandeur and dynamism of the entire scene. In a magnificent, ancient Eastern palace, towering vermilion pillars support a dome of carved beams and painted rafters. A woman, clad in a gorgeous red dress in the style of an ancient palace, dances with rapt attention. Her cascading, jet-black hair is held high by a carved jade hairpin, strands of which gently brush against her classical, serene face as she dances. Her red dress, crafted from fine silk, gleams softly in the dim light, its wide sleeves embroidered with intricate golden phoenix motifs. She is in the midst of an elegant twirl, her arms extended and her body lithe, driving the multi-tiered, immense skirt to steadily and gracefully unfold outward from her center. The golden thread at the hem traces a brilliant arc in the air, resembling a giant, blooming red peony. The light mainly comes from the ancient palace lantern hanging in the corner, casting a warm and swaying halo, which is intertwined with the cold moonlight shining through the huge window panes deep in the background. The black marble floor, polished as smooth as a mirror, clearly reflects the dancing red figure of the woman and the flowing light and shadow of her skirt, creating an atmosphere that is solemn, elegant, romantic and mysterious.", + "seed": 4133, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_5/index_1402_seed_4133_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11460081459954381, + "vlm_logprob": -0.11460081459954381, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.11460081459954381, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02793795056641102 + ], + "opposite_logprob_tokens": [ + -4.027937889099121 + ], + "logprob_reward": -0.02793795056641102, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02793795056641102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05813364312052727, + 0.014400630258023739, + 0.009662163443863392, + 0.030138645321130753, + 0.2266884446144104, + 0.506894052028656, + 0.024719109758734703, + 0.08901668339967728, + 0.01473943144083023, + 0.025607183575630188 + ], + "credit_map": [ + 0.006843638606369495, + 0.010029732249677181, + 0.008002674207091331, + 0.00942505057901144, + 0.0045164707116782665, + 0.0068504419177770615, + 0.006765408907085657, + 0.005700227804481983, + 0.0025175127666443586, + 0.0014016588684171438, + 0.001974663930013776, + 0.0016970430733636022, + 0.0014764268416911364, + 0.0013336914125829935, + 0.001977431820705533, + 0.00202220119535923, + 0.0013309456408023834, + 0.001466018264181912, + 0.0012017609551548958, + 0.0019466759404167533, + 0.0009353725472465158, + 0.0008925735019147396, + 0.0008866522694006562, + 0.0010021638590842485, + 0.0054776412434875965, + 0.003586889011785388, + 0.005531461909413338, + 0.0030382571276277304, + 0.0026852956507354975, + 0.004015692975372076, + 0.002567525487393141, + 0.0032358814496546984, + 0.014149111695587635, + 0.011355157010257244, + 0.02518312819302082, + 0.01397273875772953, + 0.012458729557693005, + 0.013822885230183601, + 0.0785423070192337, + 0.057204391807317734, + 0.03250626474618912, + 0.029894893988966942, + 0.1897428333759308, + 0.027731671929359436, + 0.018572982400655746, + 0.01058682706207037, + 0.17194445431232452, + 0.025914138182997704, + 0.003782112617045641, + 0.0025401334278285503, + 0.004540997091680765, + 0.005169623531401157, + 0.0014757586177438498, + 0.002315083984285593, + 0.0033591443207114935, + 0.0015362555859610438, + 0.010949012823402882, + 0.006719957571476698, + 0.009281476959586143, + 0.012037607841193676, + 0.01401559542864561, + 0.007102301809936762, + 0.017416616901755333, + 0.011494110338389874, + 0.0030640247277915478, + 0.001366390148177743, + 0.0019112620502710342, + 0.002086837310343981, + 0.001145244692452252, + 0.0014131482457742095, + 0.0017994351219385862, + 0.0019530891440808773, + 0.003599616466090083, + 0.0032402377109974623, + 0.005669849459081888, + 0.0028537027537822723, + 0.0028216419741511345, + 0.0017762079369276762, + 0.0025859999004751444, + 0.003059929935261607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.02793795056641102 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046554747968912125 + ], + "opposite_logprob_tokens": [ + -5.379655361175537 + ], + "logprob_reward": -0.0046554747968912125, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046554747968912125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06436162441968918, + 0.015797384083271027, + 0.01680884137749672, + 0.017338598147034645, + 0.07909213751554489, + 0.6469327807426453, + 0.02161496691405773, + 0.06757068634033203, + 0.049222223460674286, + 0.021260762587189674 + ], + "credit_map": [ + 0.012057461775839329, + 0.006473707966506481, + 0.009338668547570705, + 0.007635692600160837, + 0.0084200669080019, + 0.006019111722707748, + 0.006662819534540176, + 0.007754092570394278, + 0.002210993319749832, + 0.0013926965184509754, + 0.0020736686419695616, + 0.0020339498296380043, + 0.0017867786809802055, + 0.0013547841226682067, + 0.0030019741971045732, + 0.0019425373757258058, + 0.0026465009432286024, + 0.0010092932498082519, + 0.0014918660745024681, + 0.0028871300164610147, + 0.00250102742575109, + 0.0009905818151310086, + 0.0038778583984822035, + 0.00140458345413208, + 0.0035968059673905373, + 0.0016695128288120031, + 0.0028634308837354183, + 0.0023124024737626314, + 0.0012442006263881922, + 0.0013912037247791886, + 0.0023254300467669964, + 0.0019356122938916087, + 0.005118985194712877, + 0.004056115169078112, + 0.008619721978902817, + 0.005288407206535339, + 0.004707781597971916, + 0.005166308023035526, + 0.02653384581208229, + 0.01960097625851631, + 0.04069189727306366, + 0.03779009357094765, + 0.24124129116535187, + 0.036495745182037354, + 0.02440323308110237, + 0.011807141825556755, + 0.22110775113105774, + 0.03339562192559242, + 0.002602156950160861, + 0.00203345506452024, + 0.004680269863456488, + 0.003547580447047949, + 0.0019593595061451197, + 0.0018173451535403728, + 0.002866227412596345, + 0.0021085720509290695, + 0.006335899233818054, + 0.005193145014345646, + 0.0065769231878221035, + 0.008996008895337582, + 0.011712763458490372, + 0.004751497879624367, + 0.015124939382076263, + 0.008879508823156357, + 0.004626165144145489, + 0.0038924457039684057, + 0.012729239650070667, + 0.009374869987368584, + 0.00447014020755887, + 0.0024941274896264076, + 0.00644415570423007, + 0.005191080737859011, + 0.00262363045476377, + 0.001204105676151812, + 0.004890677519142628, + 0.002962658181786537, + 0.0013689871411770582, + 0.0011689478997141123, + 0.0035108246374875307, + 0.0035309316590428352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.0046554747968912125 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094670061022043 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.004094670061022043, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094670061022043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023661309853196144, + 0.005515015218406916, + 0.006450834218412638, + 0.008490077219903469, + 0.1423521190881729, + 0.7199383974075317, + 0.007026473991572857, + 0.05688737705349922, + 0.024861158803105354, + 0.0048172324895858765 + ], + "credit_map": [ + 0.004068420734256506, + 0.002689254004508257, + 0.002696653362363577, + 0.003088798839598894, + 0.002707853214815259, + 0.0023594426456838846, + 0.0030204730574041605, + 0.0030304142273962498, + 0.0009061674354597926, + 0.00043770900811068714, + 0.0004895295132882893, + 0.0008303730865009129, + 0.0005272981361486018, + 0.0007162578986026347, + 0.0007776265265420079, + 0.0008300532936118543, + 0.0009406757308170199, + 0.0007328265346586704, + 0.0010106198024004698, + 0.0010749411303550005, + 0.0006325089489109814, + 0.00044897914631292224, + 0.0008355194586329162, + 0.000774763582739979, + 0.001379267661832273, + 0.0009452616795897484, + 0.0014573405496776104, + 0.0009927763603627682, + 0.0008982231956906617, + 0.0006293915212154388, + 0.0012707298155874014, + 0.0009170861449092627, + 0.008619271218776703, + 0.006920519284904003, + 0.015597064048051834, + 0.009038717485964298, + 0.007834120653569698, + 0.009040208533406258, + 0.04910758510231972, + 0.03619462624192238, + 0.04562912881374359, + 0.04233725741505623, + 0.2701612710952759, + 0.04109657183289528, + 0.02845015376806259, + 0.013153337873518467, + 0.24232812225818634, + 0.03678257390856743, + 0.0007704777526669204, + 0.0006326430011540651, + 0.0016569049330428243, + 0.0010676700621843338, + 0.0006185189704410732, + 0.0005941169220022857, + 0.0009483794565312564, + 0.0007377631845884025, + 0.004551268648356199, + 0.003260598285123706, + 0.006365163251757622, + 0.008252686820924282, + 0.007606730796396732, + 0.004428870510309935, + 0.012126393616199493, + 0.010295668616890907, + 0.002339606173336506, + 0.0014503297861665487, + 0.007644082885235548, + 0.005067263729870319, + 0.0019917497411370277, + 0.0010778640862554312, + 0.002948124660179019, + 0.002342136111110449, + 0.0005318417679518461, + 0.0003542763879522681, + 0.0006989221437834203, + 0.0007026406237855554, + 0.0005613425746560097, + 0.0002673948183655739, + 0.0010775952832773328, + 0.000623219006229192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.004094670061022043 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204637885093689 + ], + "opposite_logprob_tokens": [ + -2.8120462894439697 + ], + "logprob_reward": -0.06204637885093689, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204637885093689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03573388233780861, + 0.014986989088356495, + 0.028305111452937126, + 0.011284076608717442, + 0.4930199682712555, + 0.30887147784233093, + 0.03920722007751465, + 0.02239263243973255, + 0.00828072614967823, + 0.03791794553399086 + ], + "credit_map": [ + 0.005086223129183054, + 0.0046349577605724335, + 0.005196427460759878, + 0.005602321587502956, + 0.003641183953732252, + 0.004044612869620323, + 0.004567601252347231, + 0.002960553392767906, + 0.002942088060081005, + 0.001454569399356842, + 0.0015296776546165347, + 0.0019491577986627817, + 0.0016778550343587995, + 0.0017448646249249578, + 0.001582306227646768, + 0.0021064691245555878, + 0.005374804604798555, + 0.004275379236787558, + 0.0030660522170364857, + 0.004326560068875551, + 0.002563330577686429, + 0.002344778273254633, + 0.0033713828306645155, + 0.002982822246849537, + 0.001744803856126964, + 0.001051178784109652, + 0.002471635118126869, + 0.0013475256273522973, + 0.0008594920509494841, + 0.0009868275374174118, + 0.0015461345901712775, + 0.00127647805493325, + 0.030469397082924843, + 0.02439800836145878, + 0.05094984173774719, + 0.03096199780702591, + 0.026650067418813705, + 0.03116466850042343, + 0.16963106393814087, + 0.12879489362239838, + 0.019285887479782104, + 0.017006611451506615, + 0.11138130724430084, + 0.017888637259602547, + 0.011649959720671177, + 0.006054196506738663, + 0.1086534783244133, + 0.016951389610767365, + 0.005550398025661707, + 0.0045562065206468105, + 0.007028784137219191, + 0.00669120205566287, + 0.002689346205443144, + 0.004061101470142603, + 0.0051065413281321526, + 0.0035236356779932976, + 0.002528166864067316, + 0.0019580586813390255, + 0.002045108238235116, + 0.0031560033094137907, + 0.003774278098717332, + 0.0019304301822558045, + 0.004042396787554026, + 0.002958188531920314, + 0.0011766498209908605, + 0.0010653407080098987, + 0.0011528399772942066, + 0.0012701431987807155, + 0.0008440294186584651, + 0.0006048237555660307, + 0.0010912404395639896, + 0.0010756581323221326, + 0.007697449065744877, + 0.005215942393988371, + 0.00725354952737689, + 0.003111465135589242, + 0.0024484035093337297, + 0.0034396029077470303, + 0.004185492638498545, + 0.004566036630421877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.06204637885093689 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742695987224579 + ], + "opposite_logprob_tokens": [ + -0.9742696285247803 + ], + "logprob_reward": -0.4742695987224579, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742695987224579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020378073677420616, + 0.008326048031449318, + 0.009655575267970562, + 0.00854774285107851, + 0.6646658778190613, + 0.22516603767871857, + 0.012633073143661022, + 0.02361045777797699, + 0.0183879341930151, + 0.008629265241324902 + ], + "credit_map": [ + 0.002170333405956626, + 0.004654431715607643, + 0.002539475914090872, + 0.002997179515659809, + 0.0017717743758112192, + 0.0024044604506343603, + 0.0021866417955607176, + 0.0016537744086235762, + 0.0015611998969689012, + 0.000576950900722295, + 0.0007696397951804101, + 0.001007169485092163, + 0.001099159475415945, + 0.0008067120797932148, + 0.0011121096322312951, + 0.0013931060675531626, + 0.0018305251142010093, + 0.0007214615470729768, + 0.0010566537966951728, + 0.0014987292233854532, + 0.0013062389334663749, + 0.0005629008519463241, + 0.0017390572465956211, + 0.0009400062845088542, + 0.0013589933514595032, + 0.001101452624425292, + 0.0014306016964837909, + 0.0010247082682326436, + 0.0008861108799465001, + 0.0005657832371070981, + 0.0012793665518984199, + 0.0009007249609567225, + 0.04087815061211586, + 0.03254246339201927, + 0.07060520350933075, + 0.041923217475414276, + 0.035604797303676605, + 0.04179593548178673, + 0.22931280732154846, + 0.17200322449207306, + 0.017162032425403595, + 0.017098821699619293, + 0.09227997809648514, + 0.012205363251268864, + 0.0074012321420013905, + 0.006744198966771364, + 0.04973722994327545, + 0.02253715507686138, + 0.0012167131062597036, + 0.0009729116572998464, + 0.0013949162093922496, + 0.002376285847276449, + 0.0016107421834021807, + 0.001453666016459465, + 0.0020309663377702236, + 0.001576869748532772, + 0.002285145688802004, + 0.0022371667437255383, + 0.0022234227508306503, + 0.003156266873702407, + 0.004754274617880583, + 0.0015522759640589356, + 0.0047288923524320126, + 0.002673010341823101, + 0.00344963395036757, + 0.0013685005251318216, + 0.005054174922406673, + 0.002830564510077238, + 0.0018805830040946603, + 0.0007666423334740102, + 0.0014094741782173514, + 0.0016283588483929634, + 0.0015131456311792135, + 0.0005950600607320666, + 0.0025106591638177633, + 0.0009390636114403605, + 0.0005247293156571686, + 0.0005300409393385053, + 0.0009941752068698406, + 0.0010223895078524947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.4742695987224579 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from an upward angle slightly below eye level, slowly pans to the right and zooms out slightly to capture the grandeur and dynamism of the entire scene. In a magnificent, ancient Eastern palace, towering vermilion pillars support a dome of carved beams and painted rafters. A woman, clad in a gorgeous red dress in the style of an ancient palace, dances with rapt attention. Her cascading, jet-black hair is held high by a carved jade hairpin, strands of which gently brush against her classical, serene face as she dances. Her red dress, crafted from fine silk, gleams softly in the dim light, its wide sleeves embroidered with intricate golden phoenix motifs. She is in the midst of an elegant twirl, her arms extended and her body lithe, driving the multi-tiered, immense skirt to steadily and gracefully unfold outward from her center. The golden thread at the hem traces a brilliant arc in the air, resembling a giant, blooming red peony. The light mainly comes from the ancient palace lantern hanging in the corner, casting a warm and swaying halo, which is intertwined with the cold moonlight shining through the huge window panes deep in the background. The black marble floor, polished as smooth as a mirror, clearly reflects the dancing red figure of the woman and the flowing light and shadow of her skirt, creating an atmosphere that is solemn, elegant, romantic and mysterious.", + "seed": 4134, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_5/index_1402_seed_4134_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06757576819509267, + "vlm_logprob": -0.06757576819509267, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.06757576819509267, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019437668845057487 + ], + "opposite_logprob_tokens": [ + -4.269437789916992 + ], + "logprob_reward": -0.019437668845057487, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.019437668845057487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1012868732213974, + 0.04480169713497162, + 0.024899417534470558, + 0.05556177347898483, + 0.0282273106276989, + 0.03680562227964401, + 0.5369548797607422, + 0.10601969063282013, + 0.028125103563070297, + 0.037317678332328796 + ], + "credit_map": [ + 0.01892237737774849, + 0.006528010591864586, + 0.024852026253938675, + 0.010205041617155075, + 0.013210742734372616, + 0.006368167698383331, + 0.010301560163497925, + 0.010898937471210957, + 0.012517274357378483, + 0.002536378800868988, + 0.007139073684811592, + 0.006774753797799349, + 0.0043513160198926926, + 0.0034723638091236353, + 0.004818265791982412, + 0.0031922650523483753, + 0.0035231790971010923, + 0.002256473060697317, + 0.0029176357202231884, + 0.003983773291110992, + 0.0029623075388371944, + 0.0017087380401790142, + 0.004602043889462948, + 0.0029452641028910875, + 0.009578561410307884, + 0.004570790100842714, + 0.008029649034142494, + 0.007048898842185736, + 0.005072946194559336, + 0.005782051011919975, + 0.008455798961222172, + 0.007023070473223925, + 0.004670059774070978, + 0.0027829939499497414, + 0.0035532142501324415, + 0.0033597087021917105, + 0.003827790729701519, + 0.0019463775679469109, + 0.00346617610193789, + 0.004620986990630627, + 0.006711275316774845, + 0.002417150419205427, + 0.006860445719212294, + 0.006079044658690691, + 0.0025876290164887905, + 0.0017867118585854769, + 0.005856725852936506, + 0.004506637807935476, + 0.11031585186719894, + 0.04559102654457092, + 0.12326613068580627, + 0.0988854169845581, + 0.034794583916664124, + 0.01940116100013256, + 0.06705323606729507, + 0.037647396326065063, + 0.012574409134685993, + 0.014045360498130322, + 0.011498888023197651, + 0.013041368685662746, + 0.017957085743546486, + 0.009317086078226566, + 0.015186403878033161, + 0.012399079278111458, + 0.005921698175370693, + 0.004269818309694529, + 0.0032729189842939377, + 0.0030312242452055216, + 0.00205971603281796, + 0.0030605976935476065, + 0.0034322324208915234, + 0.0030768949072808027, + 0.007736387196928263, + 0.005048869643360376, + 0.004335980862379074, + 0.003914987668395042, + 0.0032869759015738964, + 0.003983341623097658, + 0.003964103292673826, + 0.005047030281275511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.019437668845057487 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031994604505598545 + ], + "opposite_logprob_tokens": [ + -5.753199577331543 + ], + "logprob_reward": -0.0031994604505598545, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031994604505598545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08149690926074982, + 0.11942312121391296, + 0.07102885097265244, + 0.09964221715927124, + 0.035520389676094055, + 0.08992906659841537, + 0.1257152557373047, + 0.18426252901554108, + 0.08637971431016922, + 0.10660197585821152 + ], + "credit_map": [ + 0.010066958144307137, + 0.005705631338059902, + 0.010926934890449047, + 0.014745783060789108, + 0.012718464247882366, + 0.006437438540160656, + 0.009871387854218483, + 0.011024310253560543, + 0.037663571536540985, + 0.006102034356445074, + 0.017031589522957802, + 0.015534033067524433, + 0.011905974708497524, + 0.01009556744247675, + 0.011915355920791626, + 0.009174985811114311, + 0.019752712920308113, + 0.00540419016033411, + 0.016030171886086464, + 0.007989820092916489, + 0.006345747970044613, + 0.003085364820435643, + 0.00737782334908843, + 0.005043023265898228, + 0.012544391676783562, + 0.007535114418715239, + 0.012935377657413483, + 0.021746670827269554, + 0.007865230552852154, + 0.0034986750688403845, + 0.0232615415006876, + 0.010255211964249611, + 0.004678104538470507, + 0.003355228342115879, + 0.005236939061433077, + 0.006631593219935894, + 0.004556016065180302, + 0.0021331822499632835, + 0.004290870390832424, + 0.004638454876840115, + 0.013052883557975292, + 0.0070520308800041676, + 0.015787482261657715, + 0.020171033218503, + 0.005640000104904175, + 0.004859212785959244, + 0.014451761730015278, + 0.008914660662412643, + 0.022016162052750587, + 0.009604274295270443, + 0.022198380902409554, + 0.016138669103384018, + 0.02001442015171051, + 0.005374010652303696, + 0.020486796274781227, + 0.009882548823952675, + 0.022166524082422256, + 0.02467317320406437, + 0.023900754749774933, + 0.01587941125035286, + 0.02216425910592079, + 0.014988056384027004, + 0.038173869252204895, + 0.022316496819257736, + 0.007336456328630447, + 0.01152726635336876, + 0.008890397846698761, + 0.013421420007944107, + 0.007669864222407341, + 0.009364275261759758, + 0.015180371701717377, + 0.012989653274416924, + 0.02280103787779808, + 0.009044861420989037, + 0.010396018624305725, + 0.012964863330125809, + 0.010198133997619152, + 0.010468542575836182, + 0.017683809623122215, + 0.013044707477092743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.0031994604505598545 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623034063726664 + ], + "opposite_logprob_tokens": [ + -4.882623195648193 + ], + "logprob_reward": -0.007623034063726664, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623034063726664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11478222161531448, + 0.06122318655252457, + 0.07158547639846802, + 0.03774278983473778, + 0.049623724073171616, + 0.09935827553272247, + 0.1680135279893875, + 0.26993581652641296, + 0.06645573675632477, + 0.0612792931497097 + ], + "credit_map": [ + 0.01563350297510624, + 0.007342860102653503, + 0.010428916662931442, + 0.02236577495932579, + 0.02293502353131771, + 0.005788258742541075, + 0.017832819372415543, + 0.01245506014674902, + 0.01655205339193344, + 0.0039576455019414425, + 0.009500638581812382, + 0.007184083107858896, + 0.008279982022941113, + 0.004360020160675049, + 0.006762196775525808, + 0.0046265642158687115, + 0.015135860070586205, + 0.006234603468328714, + 0.012272985652089119, + 0.00807489175349474, + 0.010456912219524384, + 0.0032030423171818256, + 0.009664706885814667, + 0.0065424758940935135, + 0.004487854894250631, + 0.003141593188047409, + 0.004912895616143942, + 0.007878217846155167, + 0.0036987329367548227, + 0.0024458819534629583, + 0.008306357078254223, + 0.0028712551575154066, + 0.0064330636523664, + 0.003493242198601365, + 0.007620276417583227, + 0.007279727607965469, + 0.00405277032405138, + 0.003586928825825453, + 0.00796179287135601, + 0.00919592659920454, + 0.01591019704937935, + 0.006004086695611477, + 0.013355840928852558, + 0.017292041331529617, + 0.007350733038038015, + 0.0068105063401162624, + 0.020669780671596527, + 0.011965087614953518, + 0.03250587359070778, + 0.011173064820468426, + 0.03725878894329071, + 0.034508224576711655, + 0.01025080494582653, + 0.006192081142216921, + 0.023965708911418915, + 0.012158986181020737, + 0.04126352071762085, + 0.030820729210972786, + 0.04417356103658676, + 0.0248577781021595, + 0.029344238340854645, + 0.018604464828968048, + 0.04866822808980942, + 0.03220328688621521, + 0.0056760250590741634, + 0.00848535168915987, + 0.005819556303322315, + 0.010454537346959114, + 0.009843564592301846, + 0.005594393238425255, + 0.011021855287253857, + 0.00956045463681221, + 0.015068404376506805, + 0.005857675801962614, + 0.006356109399348497, + 0.007024731021374464, + 0.0052769980393350124, + 0.005948071368038654, + 0.007860424928367138, + 0.007886877283453941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.007623034063726664 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633458375930786 + ], + "opposite_logprob_tokens": [ + -3.651334524154663 + ], + "logprob_reward": -0.02633458375930786, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633458375930786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11538056284189224, + 0.0682513564825058, + 0.0587560199201107, + 0.05843046307563782, + 0.1282804161310196, + 0.07236205041408539, + 0.0842568501830101, + 0.08292093127965927, + 0.1517297327518463, + 0.1796315759420395 + ], + "credit_map": [ + 0.017024153843522072, + 0.007975486107170582, + 0.02517489343881607, + 0.016220340505242348, + 0.01727883331477642, + 0.006669029593467712, + 0.012491404078900814, + 0.012546425685286522, + 0.013394256122410297, + 0.004096920136362314, + 0.013273810036480427, + 0.01161688007414341, + 0.008576055988669395, + 0.0028660446405410767, + 0.00940745323896408, + 0.0050199455581605434, + 0.008109732531011105, + 0.00548919104039669, + 0.009541749022901058, + 0.0059437802992761135, + 0.006815079599618912, + 0.0062539163045585155, + 0.010172544047236443, + 0.006430030334740877, + 0.006824160926043987, + 0.004604218993335962, + 0.008499115705490112, + 0.008181588724255562, + 0.006040211301296949, + 0.004552347585558891, + 0.009967509657144547, + 0.009761315770447254, + 0.018764274194836617, + 0.011653503403067589, + 0.016055412590503693, + 0.0264434777200222, + 0.01789373718202114, + 0.007351375184953213, + 0.017365358769893646, + 0.012753289192914963, + 0.012486525811254978, + 0.003885147627443075, + 0.013418398797512054, + 0.014622227288782597, + 0.004443672019988298, + 0.004054406192153692, + 0.012626326642930508, + 0.006825352553278208, + 0.01567045971751213, + 0.013982475735247135, + 0.014225550927221775, + 0.012052921578288078, + 0.008387785404920578, + 0.004395414609462023, + 0.009297816082835197, + 0.006244438700377941, + 0.020566513761878014, + 0.013120444491505623, + 0.017669888213276863, + 0.0061155399307608604, + 0.008029447868466377, + 0.006417678203433752, + 0.005350256338715553, + 0.005651166196912527, + 0.03807083144783974, + 0.018063023686408997, + 0.021113742142915726, + 0.014574111439287663, + 0.010300817899405956, + 0.023059504106640816, + 0.012895891442894936, + 0.013651828281581402, + 0.050711341202259064, + 0.013268881477415562, + 0.02125493995845318, + 0.016917530447244644, + 0.012172617949545383, + 0.014809227548539639, + 0.023011036217212677, + 0.027486002072691917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.02633458375930786 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2812840938568115 + ], + "opposite_logprob_tokens": [ + -1.4062840938568115 + ], + "logprob_reward": -0.2812840938568115, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2812840938568115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08009546250104904, + 0.06934187561273575, + 0.08697137236595154, + 0.06639202684164047, + 0.07191005349159241, + 0.06395673751831055, + 0.08436664938926697, + 0.09567298740148544, + 0.10656566172838211, + 0.27472713589668274 + ], + "credit_map": [ + 0.0124416658654809, + 0.0051123169250786304, + 0.01674099639058113, + 0.00876818597316742, + 0.012603082694113255, + 0.005438986234366894, + 0.009751535020768642, + 0.009238691069185734, + 0.017898214980959892, + 0.004721079021692276, + 0.009104921482503414, + 0.00956642720848322, + 0.006463709753006697, + 0.007077701855450869, + 0.00848609022796154, + 0.006023729685693979, + 0.014904128387570381, + 0.005998793989419937, + 0.015013559721410275, + 0.01607563905417919, + 0.008768572472035885, + 0.004993789829313755, + 0.014231699518859386, + 0.006985195446759462, + 0.010042655281722546, + 0.00531058618798852, + 0.008484171703457832, + 0.010734072886407375, + 0.006916833575814962, + 0.003693756414577365, + 0.013784644193947315, + 0.007425307296216488, + 0.012739128433167934, + 0.005076452624052763, + 0.009920434094965458, + 0.010524543933570385, + 0.0064008282497525215, + 0.005769729148596525, + 0.012467991560697556, + 0.009010949172079563, + 0.010069740004837513, + 0.004702418111264706, + 0.014645561575889587, + 0.009425626136362553, + 0.004367445129901171, + 0.005207157228142023, + 0.00847267173230648, + 0.007066119462251663, + 0.012767014093697071, + 0.0062351226806640625, + 0.012412476353347301, + 0.01239194255322218, + 0.010826656594872475, + 0.00446291035041213, + 0.016431162133812904, + 0.008839358575642109, + 0.017611516639590263, + 0.009213211014866829, + 0.015322091989219189, + 0.00936910416930914, + 0.013110666535794735, + 0.006122407037764788, + 0.01543386373668909, + 0.009490124881267548, + 0.013083364814519882, + 0.014221030287444592, + 0.01681778021156788, + 0.011246804147958755, + 0.007587311323732138, + 0.020524045452475548, + 0.014326799660921097, + 0.008758521638810635, + 0.07543966174125671, + 0.0345115065574646, + 0.025080012157559395, + 0.022817356511950493, + 0.025065191090106964, + 0.02808002196252346, + 0.03659699484705925, + 0.027136391028761864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.2812840938568115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from an upward angle slightly below eye level, slowly pans to the right and zooms out slightly to capture the grandeur and dynamism of the entire scene. In a magnificent, ancient Eastern palace, towering vermilion pillars support a dome of carved beams and painted rafters. A woman, clad in a gorgeous red dress in the style of an ancient palace, dances with rapt attention. Her cascading, jet-black hair is held high by a carved jade hairpin, strands of which gently brush against her classical, serene face as she dances. Her red dress, crafted from fine silk, gleams softly in the dim light, its wide sleeves embroidered with intricate golden phoenix motifs. She is in the midst of an elegant twirl, her arms extended and her body lithe, driving the multi-tiered, immense skirt to steadily and gracefully unfold outward from her center. The golden thread at the hem traces a brilliant arc in the air, resembling a giant, blooming red peony. The light mainly comes from the ancient palace lantern hanging in the corner, casting a warm and swaying halo, which is intertwined with the cold moonlight shining through the huge window panes deep in the background. The black marble floor, polished as smooth as a mirror, clearly reflects the dancing red figure of the woman and the flowing light and shadow of her skirt, creating an atmosphere that is solemn, elegant, romantic and mysterious.", + "seed": 4135, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_5/index_1402_seed_4135_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08868423490785063, + "vlm_logprob": -0.08868423490785063, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.08868423490785063, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03493533283472061 + ], + "opposite_logprob_tokens": [ + -3.659935235977173 + ], + "logprob_reward": -0.03493533283472061, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03493533283472061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09448589384555817, + 0.08613687753677368, + 0.09534949064254761, + 0.053392086178064346, + 0.02592313103377819, + 0.05447769537568092, + 0.24264709651470184, + 0.04275722801685333, + 0.13381356000900269, + 0.17101696133613586 + ], + "credit_map": [ + 0.009026719257235527, + 0.014192731119692326, + 0.012280811555683613, + 0.01755187287926674, + 0.009788270108401775, + 0.009019230492413044, + 0.0088304802775383, + 0.013795777224004269, + 0.010063828900456429, + 0.013857780955731869, + 0.010751832276582718, + 0.015410583466291428, + 0.011637460440397263, + 0.007605682592839003, + 0.010343312285840511, + 0.0064663998782634735, + 0.011788169853389263, + 0.020672349259257317, + 0.012986118905246258, + 0.009513653814792633, + 0.007740654982626438, + 0.006975472439080477, + 0.010242152959108353, + 0.015430917963385582, + 0.0062950970605015755, + 0.008102635852992535, + 0.0060628196224570274, + 0.014928284101188183, + 0.004456616472452879, + 0.002511016558855772, + 0.007492397911846638, + 0.00354322069324553, + 0.004434931091964245, + 0.0027426204178482294, + 0.003645076183602214, + 0.003375298110768199, + 0.0034588887356221676, + 0.0034657130017876625, + 0.002456619869917631, + 0.0023439815267920494, + 0.007042200770229101, + 0.008823343552649021, + 0.008138120174407959, + 0.006599213927984238, + 0.00507426168769598, + 0.006662517320364714, + 0.006026177201420069, + 0.006111863534897566, + 0.02293030172586441, + 0.01716146618127823, + 0.023694084957242012, + 0.058204278349876404, + 0.03325296565890312, + 0.01483231969177723, + 0.044779691845178604, + 0.02779196761548519, + 0.005623907316476107, + 0.002686655381694436, + 0.006297538056969643, + 0.006960249040275812, + 0.005231429822742939, + 0.0030594670679420233, + 0.005074882879853249, + 0.0078230956569314, + 0.014232510700821877, + 0.0106130912899971, + 0.017639005556702614, + 0.021190151572227478, + 0.013848829083144665, + 0.016039900481700897, + 0.015059125609695911, + 0.025190941989421844, + 0.024162011221051216, + 0.01631179079413414, + 0.027022356167435646, + 0.014993290416896343, + 0.017140284180641174, + 0.016716180369257927, + 0.02730555087327957, + 0.027365505695343018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.03493533283472061 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032052830792963505 + ], + "opposite_logprob_tokens": [ + -5.753205299377441 + ], + "logprob_reward": -0.0032052830792963505, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032052830792963505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10228337347507477, + 0.06326909363269806, + 0.13840311765670776, + 0.08801951259374619, + 0.02806999161839485, + 0.1329437643289566, + 0.18693338334560394, + 0.053943756967782974, + 0.09221215546131134, + 0.1139218881726265 + ], + "credit_map": [ + 0.013631673529744148, + 0.014479988254606724, + 0.014440884813666344, + 0.021769005805253983, + 0.008382289670407772, + 0.008570714853703976, + 0.009496377781033516, + 0.01151242759078741, + 0.007125262636691332, + 0.010249052196741104, + 0.010210213251411915, + 0.011064358986914158, + 0.007941825315356255, + 0.00396023690700531, + 0.007327143102884293, + 0.005390989128500223, + 0.011089102365076542, + 0.018721772357821465, + 0.01725585013628006, + 0.02140515297651291, + 0.01582965813577175, + 0.007584758102893829, + 0.032287102192640305, + 0.014229699037969112, + 0.00833833683282137, + 0.014360297471284866, + 0.009821852669119835, + 0.021455192938447, + 0.011861695908010006, + 0.00524584436789155, + 0.011562035419046879, + 0.0053742509335279465, + 0.002837170846760273, + 0.002113342983648181, + 0.004089285619556904, + 0.006534085609018803, + 0.0028811763040721416, + 0.0024005237501114607, + 0.004256419837474823, + 0.0029579841066151857, + 0.03717297688126564, + 0.009150832891464233, + 0.020710403099656105, + 0.028362201526761055, + 0.006571792997419834, + 0.008337502367794514, + 0.01666310243308544, + 0.005974939092993736, + 0.018760347738862038, + 0.014707854948937893, + 0.02011452056467533, + 0.03111082687973976, + 0.02061690017580986, + 0.01990991272032261, + 0.023697588592767715, + 0.038015421479940414, + 0.008424032479524612, + 0.005593633279204369, + 0.007333699148148298, + 0.007086374331265688, + 0.006755187176167965, + 0.004014779347926378, + 0.005774055607616901, + 0.008961991406977177, + 0.011814931407570839, + 0.006434914190322161, + 0.013979396782815456, + 0.010222663171589375, + 0.016549229621887207, + 0.008189053274691105, + 0.010734368115663528, + 0.014287588186562061, + 0.009256605990231037, + 0.010507756844162941, + 0.018057001754641533, + 0.013177999295294285, + 0.0171585101634264, + 0.007098014000803232, + 0.02136736549437046, + 0.017298629507422447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.0032052830792963505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028174726758152246 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028174726758152246, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028174726758152246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036303214728832245, + 0.02592192403972149, + 0.04056790843605995, + 0.03471050411462784, + 0.032525889575481415, + 0.1344437301158905, + 0.4654352366924286, + 0.04682064428925514, + 0.1494864672422409, + 0.03378451615571976 + ], + "credit_map": [ + 0.004804803989827633, + 0.005442851223051548, + 0.005824907682836056, + 0.006282938178628683, + 0.003835182636976242, + 0.002246486023068428, + 0.00323760905303061, + 0.004628432914614677, + 0.0031297027599066496, + 0.004002862609922886, + 0.003470243187621236, + 0.0036903845611959696, + 0.00418719369918108, + 0.0015800446271896362, + 0.003421013941988349, + 0.0024404767900705338, + 0.005345614161342382, + 0.010708308778703213, + 0.0055030034855008125, + 0.0034188353456556797, + 0.004048864357173443, + 0.00275144400075078, + 0.0034900314640253782, + 0.0053018066100776196, + 0.004543310962617397, + 0.009936358779668808, + 0.00561103131622076, + 0.005253322422504425, + 0.0024865088053047657, + 0.0017381589859724045, + 0.0034326454624533653, + 0.001709168660454452, + 0.003913732711225748, + 0.0030863939318805933, + 0.004390236455947161, + 0.00746065704151988, + 0.0028524110093712807, + 0.002000841312110424, + 0.0044291759841144085, + 0.00439243670552969, + 0.038451164960861206, + 0.007710260339081287, + 0.02204103209078312, + 0.031976208090782166, + 0.004560186993330717, + 0.006423360668122768, + 0.017810102552175522, + 0.005471397656947374, + 0.045343924313783646, + 0.0288133155554533, + 0.042550235986709595, + 0.10676557570695877, + 0.05766988545656204, + 0.032549578696489334, + 0.08254171907901764, + 0.06920099258422852, + 0.006093109492212534, + 0.0034871899988502264, + 0.0060038077645003796, + 0.007779708597809076, + 0.005192948505282402, + 0.003098262008279562, + 0.006286444142460823, + 0.008879175409674644, + 0.015293559990823269, + 0.009616480208933353, + 0.020976319909095764, + 0.021064911037683487, + 0.016853341832756996, + 0.0173052791506052, + 0.020368710160255432, + 0.028007863089442253, + 0.004340869840234518, + 0.004025650210678577, + 0.006050446070730686, + 0.0036291060969233513, + 0.004409308545291424, + 0.0022119502536952496, + 0.004978575278073549, + 0.00413860660046339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.0028174726758152246 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904022723436356 + ], + "opposite_logprob_tokens": [ + -2.4640402793884277 + ], + "logprob_reward": -0.08904022723436356, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904022723436356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10357442498207092, + 0.04854866489768028, + 0.06631910055875778, + 0.06883677840232849, + 0.035206977277994156, + 0.04658686742186546, + 0.077228844165802, + 0.07606259733438492, + 0.1363198310136795, + 0.3413158655166626 + ], + "credit_map": [ + 0.01662261225283146, + 0.0175897479057312, + 0.016998687759041786, + 0.01522225234657526, + 0.008996467106044292, + 0.008273240178823471, + 0.011531069874763489, + 0.008340347558259964, + 0.0060180602595210075, + 0.008364340290427208, + 0.006771635729819536, + 0.008099997416138649, + 0.00601228978484869, + 0.0027013830840587616, + 0.005788495764136314, + 0.004792464431375265, + 0.006146692670881748, + 0.01413973793387413, + 0.008513557724654675, + 0.008687040768563747, + 0.005102335009723902, + 0.0036995960399508476, + 0.015073741786181927, + 0.00495639443397522, + 0.008083497174084187, + 0.010100933723151684, + 0.00803902093321085, + 0.01575327105820179, + 0.00818995013833046, + 0.004284465219825506, + 0.009239764884114265, + 0.005145875737071037, + 0.004074626602232456, + 0.002994771348312497, + 0.004302491433918476, + 0.007601036690175533, + 0.0037876637652516365, + 0.003096284344792366, + 0.00492881890386343, + 0.00442128349095583, + 0.008640337735414505, + 0.005295370705425739, + 0.007544209249317646, + 0.007509111892431974, + 0.0046796915121376514, + 0.004353752825409174, + 0.004565784707665443, + 0.003998606465756893, + 0.006499846465885639, + 0.00526098906993866, + 0.007208651397377253, + 0.01957668736577034, + 0.00918861199170351, + 0.004560702480375767, + 0.015046645887196064, + 0.009886711835861206, + 0.011519710533320904, + 0.0072126504965126514, + 0.010363812558352947, + 0.009567033499479294, + 0.005933627486228943, + 0.006956285331398249, + 0.01002868078649044, + 0.014480789192020893, + 0.02675934135913849, + 0.012988535687327385, + 0.021645551547408104, + 0.013331635855138302, + 0.01952979527413845, + 0.013846307061612606, + 0.012646401301026344, + 0.015572261065244675, + 0.058601830154657364, + 0.03657260537147522, + 0.05704835057258606, + 0.03293631225824356, + 0.027553707361221313, + 0.033952683210372925, + 0.04123733565211296, + 0.05341305211186409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.08904022723436356 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an ancient Eastern palace with towering vermilion pillars and a dome of carved beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a red silk dress with golden phoenix motifs visible dancing in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing an elegant twirl that causes her multi-tiered skirt to unfold outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting consist of a warm halo from a corner lantern mixed with cold moonlight through huge window panes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan to the right and zoom out slightly while maintaining a long shot from an upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134228587150574 + ], + "opposite_logprob_tokens": [ + -1.3134229183197021 + ], + "logprob_reward": -0.3134228587150574, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134228587150574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024095408618450165, + 0.022850824519991875, + 0.03135828673839569, + 0.0357927605509758, + 0.041344042867422104, + 0.09780128300189972, + 0.6708711981773376, + 0.020118974149227142, + 0.025024477392435074, + 0.030742794275283813 + ], + "credit_map": [ + 0.0027510065119713545, + 0.0032133415807038546, + 0.004721483215689659, + 0.004768273793160915, + 0.0022746622562408447, + 0.002416935982182622, + 0.002078650752082467, + 0.001871052896603942, + 0.0030478076077997684, + 0.002692143665626645, + 0.0034192418679594994, + 0.0034131654538214207, + 0.0036975161638110876, + 0.0015218090265989304, + 0.0031030066311359406, + 0.0019561343360692263, + 0.0032837192993611097, + 0.0062934113666415215, + 0.003993630409240723, + 0.004248702898621559, + 0.0023911381140351295, + 0.0018362891860306263, + 0.004514571279287338, + 0.004796826280653477, + 0.0042510791681706905, + 0.009575889445841312, + 0.005251583177596331, + 0.006392901763319969, + 0.0026605878956615925, + 0.0014932674821466208, + 0.004062775056809187, + 0.002104678424075246, + 0.0055890497751533985, + 0.0050335414707660675, + 0.005696531385183334, + 0.008320523425936699, + 0.0031810374930500984, + 0.003974827937781811, + 0.004369091242551804, + 0.005179442930966616, + 0.025266394019126892, + 0.00785135943442583, + 0.014674042351543903, + 0.021666621789336205, + 0.004789869301021099, + 0.005994938313961029, + 0.012960469350218773, + 0.004597585182636976, + 0.06487119942903519, + 0.043925851583480835, + 0.06116227060556412, + 0.16572988033294678, + 0.08311261981725693, + 0.04106448218226433, + 0.12594027817249298, + 0.08506458252668381, + 0.002380186226218939, + 0.0016213580965995789, + 0.0024578883312642574, + 0.003873928217217326, + 0.0032471141312271357, + 0.0013017795281484723, + 0.0021750915329903364, + 0.0030616263393312693, + 0.0022796625271439552, + 0.001944596180692315, + 0.00341610936447978, + 0.003102396847680211, + 0.0030623695347458124, + 0.0029925773851573467, + 0.003373900428414345, + 0.0048528630286455154, + 0.002854564692825079, + 0.004311288241297007, + 0.0048276507295668125, + 0.0034917769953608513, + 0.004047862719744444, + 0.0022769058123230934, + 0.005658810958266258, + 0.0032739362213760614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.3134228587150574 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_6/index_tensor([5043, 5043])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_6/index_tensor([5043, 5043])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2fe9de0cbd13bb1ffa4d7e325281f8cfcbfee499 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_6/index_tensor([5043, 5043])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "drone's wide-angle camera slowly zooms out, capturing a dreamlike scene with an incredibly slow and smooth motion. Under a deep, inky indigo sky, a brilliant full moon hangs high, dotted with sparse stars. At the center of the frame, a mysterious figure clad in a flowing white silk robe, her long hair flowing in the wind, serenely rides atop a luminous mythical beast. This beast, part deer, part dragon, its body seemingly composed of condensed starlight. Its translucent scales shimmer with a soft, steady blue-white glow, gently illuminating both its rider and the surrounding night. With a gravity-defying grace, they glide forward steadily and deliberately, their hooves treading the void as if floating on an invisible river of air. Below, the sprawling Forbidden City palaces stretch endlessly. Their iconic golden glazed roofs, illuminated by the moonlight and the beast's radiance, shimmer with a brilliant luster, like a frozen golden ocean. The entire scene is silent, with only the hem of the robe fluttering gently in the breeze, creating an extremely peaceful, solemn, and surreal dreamlike atmosphere.", + "seed": 13470, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_6/index_5043_seed_13470_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05504606626927853, + "vlm_logprob": -0.05504606626927853, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.05504606626927853, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21648116409778595 + ], + "opposite_logprob_tokens": [ + -1.7164812088012695 + ], + "logprob_reward": -0.21648116409778595, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21648116409778595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11780858039855957, + 0.06459682434797287, + 0.07916437834501266, + 0.06277580559253693, + 0.031064752489328384, + 0.03070628084242344, + 0.475027859210968, + 0.06542939692735672, + 0.015450769104063511, + 0.057975322008132935 + ], + "credit_map": [ + 0.013273854739964008, + 0.012257670052349567, + 0.017924847081303596, + 0.017027471214532852, + 0.011241268366575241, + 0.012581431306898594, + 0.013553197495639324, + 0.019948860630393028, + 0.009153620339930058, + 0.008955856785178185, + 0.00856933742761612, + 0.006703770253807306, + 0.006214984692633152, + 0.008809282444417477, + 0.008793812245130539, + 0.007396170403808355, + 0.009607771411538124, + 0.00925868097692728, + 0.0187373086810112, + 0.007881317287683487, + 0.007848317734897137, + 0.010588087141513824, + 0.009760010056197643, + 0.005482891574501991, + 0.007464985828846693, + 0.012590007856488228, + 0.006768711376935244, + 0.007564936764538288, + 0.00808791909366846, + 0.008440731093287468, + 0.00543617969378829, + 0.006422342732548714, + 0.003114459104835987, + 0.006325896829366684, + 0.0030813540797680616, + 0.004494653083384037, + 0.0029878132045269012, + 0.004589149262756109, + 0.003106909105554223, + 0.003364522708579898, + 0.004199216607958078, + 0.003584332764148712, + 0.00491406349465251, + 0.003323886077851057, + 0.003087608842179179, + 0.003898849943652749, + 0.0030285301618278027, + 0.004669797141104937, + 0.04912380874156952, + 0.11258692294359207, + 0.052172526717185974, + 0.03860516473650932, + 0.04756930470466614, + 0.060040365904569626, + 0.035157762467861176, + 0.07977204769849777, + 0.006915579549968243, + 0.00987506378442049, + 0.011125920340418816, + 0.005589343141764402, + 0.006485530640929937, + 0.007796577177941799, + 0.007104269228875637, + 0.010537119582295418, + 0.0022600723896175623, + 0.0016431687399744987, + 0.002441101474687457, + 0.001983816269785166, + 0.0014684839406982064, + 0.0020624659955501556, + 0.0017633421812206507, + 0.001828319625928998, + 0.008575990796089172, + 0.006118165794759989, + 0.005808921996504068, + 0.006580938119441271, + 0.00783616490662098, + 0.006958727724850178, + 0.009285558015108109, + 0.006810863967984915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.21648116409778595 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009805843234062195 + ], + "opposite_logprob_tokens": [ + -4.634805679321289 + ], + "logprob_reward": -0.009805843234062195, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009805843234062195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13179369270801544, + 0.08333848416805267, + 0.04061772674322128, + 0.03366224095225334, + 0.024084700271487236, + 0.01867728866636753, + 0.3671611547470093, + 0.16931557655334473, + 0.04401097819209099, + 0.08733810484409332 + ], + "credit_map": [ + 0.017550334334373474, + 0.013754158280789852, + 0.01820964179933071, + 0.010240809060633183, + 0.027844930067658424, + 0.018408147618174553, + 0.014140710234642029, + 0.011644966900348663, + 0.012164747342467308, + 0.018035944551229477, + 0.009890827350318432, + 0.008080250583589077, + 0.009926853701472282, + 0.012276915833353996, + 0.007389950565993786, + 0.005572999827563763, + 0.005366130731999874, + 0.004320928826928139, + 0.004251916892826557, + 0.0038252549711614847, + 0.0036168196238577366, + 0.005639102775603533, + 0.006127191241830587, + 0.007470385637134314, + 0.0038046736735850573, + 0.008435339666903019, + 0.0021270064171403646, + 0.0033465696033090353, + 0.0050550238229334354, + 0.005412305239588022, + 0.002938853343948722, + 0.0025424740742892027, + 0.0026298179291188717, + 0.004526454955339432, + 0.0019694005604833364, + 0.003787193214520812, + 0.003164108144119382, + 0.00330732436850667, + 0.002591965952888131, + 0.002108436543494463, + 0.002051892690360546, + 0.002359806327149272, + 0.002401873702183366, + 0.0014096011873334646, + 0.0025729918852448463, + 0.0022197216749191284, + 0.0017235608538612723, + 0.003937841393053532, + 0.03741166740655899, + 0.08921988308429718, + 0.03998899459838867, + 0.030200786888599396, + 0.03688262030482292, + 0.04644321650266647, + 0.026651302352547646, + 0.060362715274095535, + 0.018175281584262848, + 0.046061620116233826, + 0.016706688329577446, + 0.013844185508787632, + 0.02029230259358883, + 0.024766409769654274, + 0.014228735119104385, + 0.015240376815199852, + 0.009108235128223896, + 0.009838747791945934, + 0.003554498078301549, + 0.0022871512919664383, + 0.003174378303810954, + 0.009569169022142887, + 0.0031768660992383957, + 0.003301938995718956, + 0.01388995535671711, + 0.01506121177226305, + 0.006796277593821287, + 0.011671915650367737, + 0.011337839998304844, + 0.00842600129544735, + 0.01210742350667715, + 0.00804748572409153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.009805843234062195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006747792940586805 + ], + "opposite_logprob_tokens": [ + -5.006747722625732 + ], + "logprob_reward": -0.006747792940586805, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747792940586805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07383722811937332, + 0.035581015050411224, + 0.056079134345054626, + 0.038631368428468704, + 0.041855588555336, + 0.06067227944731712, + 0.4382637143135071, + 0.0820745974779129, + 0.011741265654563904, + 0.1612638533115387 + ], + "credit_map": [ + 0.01053529791533947, + 0.010886589996516705, + 0.007861442863941193, + 0.007495450787246227, + 0.007624785415828228, + 0.008536458015441895, + 0.008533195592463017, + 0.012364013120532036, + 0.003857766045257449, + 0.0059991078451275826, + 0.0036412787158042192, + 0.0029577454552054405, + 0.0059461635537445545, + 0.004502326715737581, + 0.004292901139706373, + 0.004383730236440897, + 0.005726032890379429, + 0.005959555506706238, + 0.008620272390544415, + 0.004519396461546421, + 0.005820896942168474, + 0.006596271879971027, + 0.008598709478974342, + 0.010237998329102993, + 0.005151802208274603, + 0.006540023256093264, + 0.0034377598203718662, + 0.004093899391591549, + 0.004049776121973991, + 0.00452428637072444, + 0.004746351391077042, + 0.006087471731007099, + 0.005148810800164938, + 0.008913528174161911, + 0.00317661976441741, + 0.005679079331457615, + 0.005595329683274031, + 0.0059609245508909225, + 0.003914959728717804, + 0.0034663428086787462, + 0.006384167820215225, + 0.010037955828011036, + 0.006132137030363083, + 0.00923543144017458, + 0.008779340423643589, + 0.00886576808989048, + 0.0056680552661418915, + 0.005569428205490112, + 0.04461098834872246, + 0.10647393018007278, + 0.04760671406984329, + 0.0359155647456646, + 0.04414927959442139, + 0.05562363564968109, + 0.03205983340740204, + 0.07182377576828003, + 0.007922112010419369, + 0.01018785871565342, + 0.008237911388278008, + 0.012956134043633938, + 0.008117224089801311, + 0.009821749292314053, + 0.014508096501231194, + 0.010323514230549335, + 0.002250840188935399, + 0.0015384297585114837, + 0.0013740228023380041, + 0.0007374956621788442, + 0.0010974133620038629, + 0.0029026013799011707, + 0.0009207096882164478, + 0.0009197536273859441, + 0.026253607124090195, + 0.01826029270887375, + 0.015167459845542908, + 0.019288502633571625, + 0.01974097266793251, + 0.019874943420290947, + 0.026965217664837837, + 0.015712853521108627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.006747792940586805 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809729963541031 + ], + "opposite_logprob_tokens": [ + -3.288097381591797 + ], + "logprob_reward": -0.03809729963541031, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809729963541031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15724024176597595, + 0.19235415756702423, + 0.2834295332431793, + 0.028350606560707092, + 0.02472553215920925, + 0.027590641751885414, + 0.03798994794487953, + 0.1457674354314804, + 0.01565961353480816, + 0.08689224720001221 + ], + "credit_map": [ + 0.016857869923114777, + 0.01717725396156311, + 0.026936186477541924, + 0.02166200615465641, + 0.012774287723004818, + 0.015932830050587654, + 0.019052067771553993, + 0.026847729459404945, + 0.02005898579955101, + 0.021501898765563965, + 0.016880402341485023, + 0.00967621523886919, + 0.020192822441458702, + 0.025185735896229744, + 0.04368039220571518, + 0.03517770767211914, + 0.031417712569236755, + 0.02091134339570999, + 0.030614713206887245, + 0.01828804984688759, + 0.020356837660074234, + 0.031285371631383896, + 0.05297141522169113, + 0.07758409529924393, + 0.0031012436375021935, + 0.0037838476710021496, + 0.0023876959457993507, + 0.0023056846112012863, + 0.0027737414930015802, + 0.0032986735459417105, + 0.003254392882809043, + 0.0074453274719417095, + 0.003008571919053793, + 0.004283860791474581, + 0.0021957107819616795, + 0.003211931325495243, + 0.0031976408790796995, + 0.0034306985326111317, + 0.0022553179878741503, + 0.0031417994759976864, + 0.004273296799510717, + 0.0033630700781941414, + 0.003857776988297701, + 0.0019579059444367886, + 0.002776174806058407, + 0.00374664762057364, + 0.0026390983257442713, + 0.004976669326424599, + 0.004390757065266371, + 0.007425621151924133, + 0.003914557863026857, + 0.003991442266851664, + 0.005388361867517233, + 0.005038903094828129, + 0.004153912886977196, + 0.0036863950081169605, + 0.016505630686879158, + 0.04017149284482002, + 0.015070751309394836, + 0.01071905717253685, + 0.016871077939867973, + 0.021368779242038727, + 0.012288019061088562, + 0.01277263555675745, + 0.0025990703143179417, + 0.0014447879511862993, + 0.002167933154851198, + 0.0013097361661493778, + 0.001469224807806313, + 0.00350056029856205, + 0.0010253435466438532, + 0.002142959041520953, + 0.01453364733606577, + 0.007208424154669046, + 0.009197106584906578, + 0.010238243266940117, + 0.011402824893593788, + 0.011099698022007942, + 0.013463315553963184, + 0.00974898599088192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.03809729963541031 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098231438547373 + ], + "opposite_logprob_tokens": [ + -5.504098415374756 + ], + "logprob_reward": -0.004098231438547373, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098231438547373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15640150010585785, + 0.06240558996796608, + 0.10189653187990189, + 0.038393136113882065, + 0.02649305760860443, + 0.052726902067661285, + 0.3724556863307953, + 0.06339563429355621, + 0.029680080711841583, + 0.09615184366703033 + ], + "credit_map": [ + 0.01619824394583702, + 0.018969597294926643, + 0.022253775969147682, + 0.02133372612297535, + 0.014444626867771149, + 0.0164727084338665, + 0.022116191685199738, + 0.024612635374069214, + 0.006965915206819773, + 0.005998209118843079, + 0.011609804816544056, + 0.006742103490978479, + 0.005743294022977352, + 0.008270413614809513, + 0.010154829360544682, + 0.0069210221990942955, + 0.010754642076790333, + 0.01184573583304882, + 0.01884673535823822, + 0.008141646161675453, + 0.011142916046082973, + 0.012079832144081593, + 0.013831770978868008, + 0.015253262594342232, + 0.004511774983257055, + 0.007018190808594227, + 0.0036958137061446905, + 0.0048111118376255035, + 0.004576962441205978, + 0.00497080385684967, + 0.0039186058565974236, + 0.004889876581728458, + 0.0028691079933196306, + 0.005048472434282303, + 0.002289458177983761, + 0.002375554759055376, + 0.0027572321705520153, + 0.003674744162708521, + 0.002772480947896838, + 0.004706006962805986, + 0.0058319345116615295, + 0.0066809458658099174, + 0.006747146602720022, + 0.007226188667118549, + 0.006826430093497038, + 0.006733419839292765, + 0.006084127817302942, + 0.0065967063419520855, + 0.03776201605796814, + 0.08851674944162369, + 0.04195917025208473, + 0.030384307727217674, + 0.03698452189564705, + 0.04693720489740372, + 0.027633681893348694, + 0.06227805092930794, + 0.007122420705854893, + 0.008515102788805962, + 0.010547969490289688, + 0.006102855317294598, + 0.005573221016675234, + 0.006992559414356947, + 0.007926758378744125, + 0.010614747181534767, + 0.004609681200236082, + 0.006435438524931669, + 0.0035764980129897594, + 0.0026003483217209578, + 0.002219848334789276, + 0.004815018270164728, + 0.00256623444147408, + 0.002857013838365674, + 0.014948996715247631, + 0.011454873718321323, + 0.010220960713922977, + 0.010336487554013729, + 0.011146706528961658, + 0.011784214526414871, + 0.015251399949193, + 0.011008208617568016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.004098231438547373 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "drone's wide-angle camera slowly zooms out, capturing a dreamlike scene with an incredibly slow and smooth motion. Under a deep, inky indigo sky, a brilliant full moon hangs high, dotted with sparse stars. At the center of the frame, a mysterious figure clad in a flowing white silk robe, her long hair flowing in the wind, serenely rides atop a luminous mythical beast. This beast, part deer, part dragon, its body seemingly composed of condensed starlight. Its translucent scales shimmer with a soft, steady blue-white glow, gently illuminating both its rider and the surrounding night. With a gravity-defying grace, they glide forward steadily and deliberately, their hooves treading the void as if floating on an invisible river of air. Below, the sprawling Forbidden City palaces stretch endlessly. Their iconic golden glazed roofs, illuminated by the moonlight and the beast's radiance, shimmer with a brilliant luster, like a frozen golden ocean. The entire scene is silent, with only the hem of the robe fluttering gently in the breeze, creating an extremely peaceful, solemn, and surreal dreamlike atmosphere.", + "seed": 13471, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_6/index_5043_seed_13471_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04328093170188367, + "vlm_logprob": -0.04328093170188367, + "vlm_yes_no_margin": 3.974999952316284, + "vlm_token_credit": -0.04328093170188367, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13445945084095 + ], + "opposite_logprob_tokens": [ + -2.1344594955444336 + ], + "logprob_reward": -0.13445945084095, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13445945084095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42000287771224976, + 0.041299134492874146, + 0.024438275024294853, + 0.023336734622716904, + 0.20900669693946838, + 0.03396165370941162, + 0.045221082866191864, + 0.07131760567426682, + 0.10576228052377701, + 0.025653617456555367 + ], + "credit_map": [ + 0.02610800415277481, + 0.11109033226966858, + 0.0202846210449934, + 0.030921081081032753, + 0.05967841297388077, + 0.0970684215426445, + 0.04666008800268173, + 0.02819191850721836, + 0.003596249036490917, + 0.008105960674583912, + 0.0032602755818516016, + 0.005473905708640814, + 0.004373467992991209, + 0.010577205568552017, + 0.0036467621102929115, + 0.0022653110790997744, + 0.002172346692532301, + 0.002957340097054839, + 0.003138212487101555, + 0.005508879199624062, + 0.002476574620231986, + 0.002932795323431492, + 0.0025476464070379734, + 0.0027044799644500017, + 0.002728258492425084, + 0.004111886490136385, + 0.0020472256001085043, + 0.002327167661860585, + 0.003012988483533263, + 0.004119514022022486, + 0.003284134203568101, + 0.0017055609496310353, + 0.05850326642394066, + 0.0834500789642334, + 0.00797657947987318, + 0.007399778347462416, + 0.009118866175413132, + 0.028813576325774193, + 0.006586603354662657, + 0.007157947402447462, + 0.004534435458481312, + 0.005417912267148495, + 0.004645258188247681, + 0.002663349499925971, + 0.004705039784312248, + 0.005613870453089476, + 0.0037410457152873278, + 0.002640743041411042, + 0.005704502575099468, + 0.00537489540874958, + 0.005555284209549427, + 0.008528136648237705, + 0.004132125526666641, + 0.006957835517823696, + 0.0054013668559491634, + 0.003566934959962964, + 0.007929891347885132, + 0.0201040580868721, + 0.009333225898444653, + 0.006052184849977493, + 0.005156447179615498, + 0.010328058153390884, + 0.008291886188089848, + 0.004121850710362196, + 0.012993542477488518, + 0.02337898127734661, + 0.012944486923515797, + 0.007622229866683483, + 0.00876627303659916, + 0.0122619504109025, + 0.01952270418405533, + 0.00827211793512106, + 0.00313804903998971, + 0.0025586856063455343, + 0.004557394422590733, + 0.004009726922959089, + 0.002299172803759575, + 0.0015288942959159613, + 0.004429343156516552, + 0.003132352139800787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.13445945084095 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821013167500496 + ], + "opposite_logprob_tokens": [ + -4.018209934234619 + ], + "logprob_reward": -0.01821013167500496, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01821013167500496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28158918023109436, + 0.0423104502260685, + 0.023148998618125916, + 0.01031266525387764, + 0.1339476853609085, + 0.07623674720525742, + 0.041583042591810226, + 0.2498515546321869, + 0.062052641063928604, + 0.07896699756383896 + ], + "credit_map": [ + 0.027567964047193527, + 0.0827527716755867, + 0.016996115446090698, + 0.014785820618271828, + 0.02448534034192562, + 0.07790276408195496, + 0.018172115087509155, + 0.01892632059752941, + 0.0034942561760544777, + 0.007668900769203901, + 0.005583962891250849, + 0.004868372809141874, + 0.00476654339581728, + 0.008229258470237255, + 0.0042121619917452335, + 0.0034869976807385683, + 0.0016955903265625238, + 0.0022606130223721266, + 0.0030220476910471916, + 0.005675811320543289, + 0.00196068175137043, + 0.00239747017621994, + 0.0027777503710240126, + 0.003359034424647689, + 0.0008818252244964242, + 0.0013745698379352689, + 0.0011284510837867856, + 0.0009589612600393593, + 0.0014881450915709138, + 0.002088198671117425, + 0.0015511485980823636, + 0.000841365777887404, + 0.03421293944120407, + 0.022944875061511993, + 0.0041436804458498955, + 0.002566467970609665, + 0.004467888735234737, + 0.05593223124742508, + 0.004003475420176983, + 0.005676130764186382, + 0.009652975015342236, + 0.011467552743852139, + 0.010137178003787994, + 0.009037400595843792, + 0.0066317180171608925, + 0.0143107445910573, + 0.00828081276267767, + 0.0067183710634708405, + 0.004562004469335079, + 0.006765956524759531, + 0.004377264995127916, + 0.004840829409658909, + 0.00442356476560235, + 0.004618149250745773, + 0.007713632192462683, + 0.004281644709408283, + 0.04098723083734512, + 0.021967405453324318, + 0.051865916699171066, + 0.03257427737116814, + 0.02062421478331089, + 0.02469249628484249, + 0.0399334691464901, + 0.01720655895769596, + 0.008165262639522552, + 0.01543757226318121, + 0.011892233043909073, + 0.0053245909512043, + 0.004333813209086657, + 0.005016697105020285, + 0.0067920563742518425, + 0.005090418271720409, + 0.007521765306591988, + 0.009986000135540962, + 0.009124331176280975, + 0.01231358852237463, + 0.008135790005326271, + 0.008065774105489254, + 0.017721524462103844, + 0.0060982294380664825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.01821013167500496 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005960902199149132 + ], + "opposite_logprob_tokens": [ + -5.130960941314697 + ], + "logprob_reward": -0.005960902199149132, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005960902199149132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07020250707864761, + 0.07477294653654099, + 0.018070289865136147, + 0.019284464418888092, + 0.6672474145889282, + 0.021116532385349274, + 0.040324658155441284, + 0.03928068280220032, + 0.03176857903599739, + 0.017931858077645302 + ], + "credit_map": [ + 0.006957690697163343, + 0.015460408292710781, + 0.005886262282729149, + 0.004743152763694525, + 0.008061136119067669, + 0.016066240146756172, + 0.007182731293141842, + 0.005844890605658293, + 0.0062317019328475, + 0.014005428180098534, + 0.008624007925391197, + 0.013089803047478199, + 0.0047208042815327644, + 0.01680145226418972, + 0.006167830433696508, + 0.005131921730935574, + 0.0011137023102492094, + 0.002531109843403101, + 0.001656053587794304, + 0.003219249192625284, + 0.0020189012866467237, + 0.0034020941238850355, + 0.0026390377897769213, + 0.0014901439426466823, + 0.0017705484060570598, + 0.0018609948456287384, + 0.0025416903663426638, + 0.0023563627619296312, + 0.002221090719103813, + 0.0024033132940530777, + 0.0038595781661570072, + 0.0022708873730152845, + 0.0876297652721405, + 0.20105575025081635, + 0.015886923298239708, + 0.01613249070942402, + 0.019264250993728638, + 0.28926366567611694, + 0.02140115760266781, + 0.016613438725471497, + 0.0025046735536307096, + 0.002853324403986335, + 0.0020033123437315226, + 0.002150375861674547, + 0.0029178918339312077, + 0.0028624979313462973, + 0.003957219887524843, + 0.001867238082922995, + 0.003923483658581972, + 0.005372245330363512, + 0.0040748524479568005, + 0.00591292604804039, + 0.004257565829902887, + 0.006434640381485224, + 0.006923723965883255, + 0.0034252272453159094, + 0.006548909470438957, + 0.003825144609436393, + 0.00709904357790947, + 0.005935284774750471, + 0.0034167924895882607, + 0.004825979005545378, + 0.005177577491849661, + 0.002451954409480095, + 0.004095081705600023, + 0.00810642633587122, + 0.004311845172196627, + 0.0032431676518172026, + 0.001984209520742297, + 0.004608738236129284, + 0.0032237390987575054, + 0.0021953752730041742, + 0.002850628923624754, + 0.002155301161110401, + 0.0020883800461888313, + 0.0032257670536637306, + 0.0019048097310587764, + 0.0014431729214265943, + 0.002343122148886323, + 0.0019206772558391094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.005960902199149132 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054951589554548264 + ], + "opposite_logprob_tokens": [ + -2.9299516677856445 + ], + "logprob_reward": -0.054951589554548264, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054951589554548264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49491429328918457, + 0.02372445911169052, + 0.005867890547960997, + 0.008936852216720581, + 0.3092278838157654, + 0.007319084368646145, + 0.01645492948591709, + 0.053792715072631836, + 0.046864453703165054, + 0.032897401601076126 + ], + "credit_map": [ + 0.02952716127038002, + 0.15567269921302795, + 0.01924039050936699, + 0.029175136238336563, + 0.05673530325293541, + 0.1363840252161026, + 0.043808456510305405, + 0.024371132254600525, + 0.0018831223715096712, + 0.004510532133281231, + 0.0021110367961227894, + 0.00438298936933279, + 0.0022847121581435204, + 0.004873641300946474, + 0.0020600748248398304, + 0.0016183516709133983, + 0.0005963050061836839, + 0.0009564519277773798, + 0.0005494118086062372, + 0.0006921786116436124, + 0.0006756646907888353, + 0.0012403071159496903, + 0.0006334305508062243, + 0.0005241408944129944, + 0.0008366386755369604, + 0.0009497971041128039, + 0.0010629069292917848, + 0.000963503320235759, + 0.0009401680435985327, + 0.0011903408449143171, + 0.0014452643226832151, + 0.0015482332091778517, + 0.10343877226114273, + 0.03328811749815941, + 0.012026738375425339, + 0.00938800722360611, + 0.011881948448717594, + 0.1132424846291542, + 0.011219688691198826, + 0.014742124825716019, + 0.0008727398817427456, + 0.0008721701451577246, + 0.0009096279391087592, + 0.0010795382549986243, + 0.0008720332989469171, + 0.0008215015404857695, + 0.0012333173071965575, + 0.0006581561756320298, + 0.0019239920657128096, + 0.0017310235416516662, + 0.002296561375260353, + 0.0022421907633543015, + 0.0017752819694578648, + 0.0018041913863271475, + 0.003067447803914547, + 0.0016142402309924364, + 0.005055291578173637, + 0.016609441488981247, + 0.009644143283367157, + 0.004393653944134712, + 0.0035218887496739626, + 0.006286312825977802, + 0.005546501372009516, + 0.0027354832272976637, + 0.006475417409092188, + 0.009726285003125668, + 0.006370226386934519, + 0.005700610112398863, + 0.003374246181920171, + 0.004342685453593731, + 0.005279747769236565, + 0.005595233757048845, + 0.0030411765910685062, + 0.004685191437602043, + 0.0027348666917532682, + 0.009036592207849026, + 0.002193896332755685, + 0.0031382334418594837, + 0.005234250333160162, + 0.0028331922367215157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.054951589554548264 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028225842397660017 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.0028225842397660017, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028225842397660017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24583324790000916, + 0.05748238041996956, + 0.010711238719522953, + 0.01575121469795704, + 0.34430620074272156, + 0.03314884752035141, + 0.10999190807342529, + 0.034431006759405136, + 0.11791246384382248, + 0.030431559309363365 + ], + "credit_map": [ + 0.019230004400014877, + 0.05224546790122986, + 0.01423280406743288, + 0.0219464972615242, + 0.03830283507704735, + 0.044787898659706116, + 0.034914642572402954, + 0.0201730877161026, + 0.004336180165410042, + 0.011888113804161549, + 0.004688061308115721, + 0.010136611759662628, + 0.003823709674179554, + 0.014763272367417812, + 0.004332336131483316, + 0.0035140945110470057, + 0.0010719213169068098, + 0.0012205332750454545, + 0.0011845225235447288, + 0.0014835417969152331, + 0.0016599507071077824, + 0.0015056667616590858, + 0.0015384780708700418, + 0.001046624849550426, + 0.001630773302167654, + 0.002333389362320304, + 0.0017359053017571568, + 0.0016910849371924996, + 0.002102720085531473, + 0.0023060424719005823, + 0.0025925205554813147, + 0.001358777517452836, + 0.015235638245940208, + 0.1359023153781891, + 0.006674113217741251, + 0.008723712526261806, + 0.011026834137737751, + 0.14536048471927643, + 0.012776710093021393, + 0.008606385439634323, + 0.0029278993606567383, + 0.0042207613587379456, + 0.002838269341737032, + 0.0023408913984894753, + 0.006095838267356157, + 0.006718477234244347, + 0.005802407395094633, + 0.0022043022327125072, + 0.011628224514424801, + 0.013314470648765564, + 0.019036347046494484, + 0.0168597474694252, + 0.008608248084783554, + 0.011553370393812656, + 0.017909662798047066, + 0.011081835255026817, + 0.003996450453996658, + 0.009633438661694527, + 0.005041163414716721, + 0.00274654570966959, + 0.00253890547901392, + 0.004439628683030605, + 0.00437058974057436, + 0.0016642838018015027, + 0.01576349139213562, + 0.03436218947172165, + 0.008459465578198433, + 0.008100162260234356, + 0.007919460535049438, + 0.01464882306754589, + 0.019623784348368645, + 0.009035089984536171, + 0.003835294395685196, + 0.0028980402275919914, + 0.003787212772294879, + 0.005623975303024054, + 0.0033356284257024527, + 0.002626433502882719, + 0.005073597654700279, + 0.0032513767946511507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.0028225842397660017 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "drone's wide-angle camera slowly zooms out, capturing a dreamlike scene with an incredibly slow and smooth motion. Under a deep, inky indigo sky, a brilliant full moon hangs high, dotted with sparse stars. At the center of the frame, a mysterious figure clad in a flowing white silk robe, her long hair flowing in the wind, serenely rides atop a luminous mythical beast. This beast, part deer, part dragon, its body seemingly composed of condensed starlight. Its translucent scales shimmer with a soft, steady blue-white glow, gently illuminating both its rider and the surrounding night. With a gravity-defying grace, they glide forward steadily and deliberately, their hooves treading the void as if floating on an invisible river of air. Below, the sprawling Forbidden City palaces stretch endlessly. Their iconic golden glazed roofs, illuminated by the moonlight and the beast's radiance, shimmer with a brilliant luster, like a frozen golden ocean. The entire scene is silent, with only the hem of the robe fluttering gently in the breeze, creating an extremely peaceful, solemn, and surreal dreamlike atmosphere.", + "seed": 13472, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_6/index_5043_seed_13472_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03193256959784776, + "vlm_logprob": -0.03193256959784776, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.03193256959784776, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.059570424258708954 + ], + "opposite_logprob_tokens": [ + -3.0595703125 + ], + "logprob_reward": -0.059570424258708954, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.059570424258708954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08712543547153473, + 0.04791269451379776, + 0.08490147441625595, + 0.03234656900167465, + 0.05254850536584854, + 0.33778175711631775, + 0.22965893149375916, + 0.05387037247419357, + 0.054621435701847076, + 0.019232777878642082 + ], + "credit_map": [ + 0.011865994893014431, + 0.0163732822984457, + 0.008729459717869759, + 0.017440633848309517, + 0.007891298271715641, + 0.010318535380065441, + 0.007824710570275784, + 0.006681523751467466, + 0.006934872828423977, + 0.007752903271466494, + 0.007090825587511063, + 0.008017382584512234, + 0.006100927013903856, + 0.005283380392938852, + 0.003641249844804406, + 0.0030911550857126713, + 0.012999184429645538, + 0.009642294608056545, + 0.00669232290238142, + 0.0130769619718194, + 0.016790959984064102, + 0.014699875377118587, + 0.0059466310776770115, + 0.005053242202848196, + 0.003435838734731078, + 0.005060381721705198, + 0.0038570049218833447, + 0.003568085376173258, + 0.007029518019407988, + 0.004169756080955267, + 0.0030656978487968445, + 0.002160284435376525, + 0.006444125436246395, + 0.006745277438312769, + 0.004759198985993862, + 0.007104666903614998, + 0.011244943365454674, + 0.0067685493268072605, + 0.005901765543967485, + 0.003579978132620454, + 0.03683486953377724, + 0.05331343412399292, + 0.04680698737502098, + 0.029898133128881454, + 0.02644793875515461, + 0.08499551564455032, + 0.021331997588276863, + 0.03815286234021187, + 0.026238713413476944, + 0.032300256192684174, + 0.028050437569618225, + 0.02506430633366108, + 0.021313045173883438, + 0.04835598170757294, + 0.019900452345609665, + 0.028435740619897842, + 0.006088148802518845, + 0.00978707056492567, + 0.00560400914400816, + 0.0045774239115417, + 0.00699898786842823, + 0.012199025601148605, + 0.0039620730094611645, + 0.004653631709516048, + 0.007138625252991915, + 0.0063782366923987865, + 0.008843110874295235, + 0.0075196861289441586, + 0.004676657263189554, + 0.005979640409350395, + 0.006306431256234646, + 0.007779042236506939, + 0.0029834641609340906, + 0.0025826632045209408, + 0.0018079951405525208, + 0.001816956209950149, + 0.0034825976472347975, + 0.002712331945076585, + 0.0021413923241198063, + 0.0017053763149306178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.059570424258708954 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009799469262361526 + ], + "opposite_logprob_tokens": [ + -4.634799480438232 + ], + "logprob_reward": -0.009799469262361526, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009799469262361526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1086423322558403, + 0.0695934072136879, + 0.15873290598392487, + 0.022597789764404297, + 0.05031036213040352, + 0.26511937379837036, + 0.14049582183361053, + 0.04803650826215744, + 0.03321365267038345, + 0.10325785726308823 + ], + "credit_map": [ + 0.012329261749982834, + 0.01563437283039093, + 0.012687791138887405, + 0.0165040772408247, + 0.014163602143526077, + 0.015805203467607498, + 0.010604603216052055, + 0.01091341208666563, + 0.008327960036695004, + 0.01007016934454441, + 0.009806167334318161, + 0.010972747579216957, + 0.00901905819773674, + 0.013046402484178543, + 0.0048426794819533825, + 0.0035082127433270216, + 0.018439041450619698, + 0.024988459423184395, + 0.01339017879217863, + 0.025502685457468033, + 0.03189396113157272, + 0.020856015384197235, + 0.013903112150728703, + 0.009759443812072277, + 0.0022204171400517225, + 0.0030178904999047518, + 0.0022501801140606403, + 0.0032099157106131315, + 0.005481794476509094, + 0.0029830781277269125, + 0.0021150887478142977, + 0.0013194240164011717, + 0.0047433041036129, + 0.007053662557154894, + 0.00344861694611609, + 0.005475753918290138, + 0.013440325856208801, + 0.007102359551936388, + 0.00527589675039053, + 0.0037704387214034796, + 0.030621668323874474, + 0.038775864988565445, + 0.0440334677696228, + 0.0240138191729784, + 0.0204150453209877, + 0.04536492004990578, + 0.018880093470215797, + 0.04301447793841362, + 0.013564922846853733, + 0.017207076773047447, + 0.02390536293387413, + 0.014874613843858242, + 0.019022665917873383, + 0.018254514783620834, + 0.014050657860934734, + 0.01961599476635456, + 0.004896323196589947, + 0.005583002232015133, + 0.005082105752080679, + 0.006455841939896345, + 0.008452319540083408, + 0.006900434382259846, + 0.007903721183538437, + 0.002762754447758198, + 0.0027175466530025005, + 0.00573849119246006, + 0.0033568742219358683, + 0.003129285294562578, + 0.003603917546570301, + 0.006594785954803228, + 0.004309498239308596, + 0.0037632512394338846, + 0.009115302935242653, + 0.018159372732043266, + 0.017328830435872078, + 0.007533146999776363, + 0.008019975386559963, + 0.022160308435559273, + 0.008483000099658966, + 0.012457910925149918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.009799469262361526 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979132391512394 + ], + "opposite_logprob_tokens": [ + -4.634791374206543 + ], + "logprob_reward": -0.00979132391512394, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979132391512394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1712217926979065, + 0.06723588705062866, + 0.1553886979818344, + 0.035280246287584305, + 0.04368094727396965, + 0.22991319000720978, + 0.1295146942138672, + 0.04629258066415787, + 0.0431424118578434, + 0.07832954823970795 + ], + "credit_map": [ + 0.028066182509064674, + 0.0241254772990942, + 0.017712274566292763, + 0.02444181591272354, + 0.013190427795052528, + 0.0365992933511734, + 0.014476964250206947, + 0.012609366327524185, + 0.007566873915493488, + 0.009677250869572163, + 0.009970287792384624, + 0.012133626267313957, + 0.0072782463394105434, + 0.012378977611660957, + 0.004429205786436796, + 0.003801417537033558, + 0.015920570120215416, + 0.019936535507440567, + 0.015612959861755371, + 0.03697117418050766, + 0.019710281863808632, + 0.020344531163573265, + 0.016303015872836113, + 0.010589633136987686, + 0.0033526907209306955, + 0.005127213895320892, + 0.003943051211535931, + 0.003470968222245574, + 0.009286614134907722, + 0.004583235364407301, + 0.0025167895946651697, + 0.0029996817465871572, + 0.004243616480380297, + 0.005847400985658169, + 0.0031973435543477535, + 0.0066088284365832806, + 0.010161916725337505, + 0.006367689464241266, + 0.004180265590548515, + 0.0030738897621631622, + 0.025245314463973045, + 0.034484006464481354, + 0.029654735699295998, + 0.025315098464488983, + 0.022405583411455154, + 0.053517356514930725, + 0.018510567024350166, + 0.02078053541481495, + 0.014749044552445412, + 0.019962849095463753, + 0.011017329059541225, + 0.01498777512460947, + 0.012304094620049, + 0.031396765261888504, + 0.01137932762503624, + 0.013717510737478733, + 0.004591354634612799, + 0.007193314377218485, + 0.005839950405061245, + 0.004841140937060118, + 0.008366533555090427, + 0.006687201093882322, + 0.004190259147435427, + 0.004582824185490608, + 0.004342833999544382, + 0.007249133195728064, + 0.00380456424318254, + 0.005400346126407385, + 0.00672614760696888, + 0.006215711124241352, + 0.005940664559602737, + 0.0034630096051841974, + 0.0083787115290761, + 0.011535216122865677, + 0.010486152023077011, + 0.006718891207128763, + 0.015454215928912163, + 0.011635924689471722, + 0.006537923123687506, + 0.007582512684166431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.00979132391512394 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898691296577454 + ], + "opposite_logprob_tokens": [ + -2.578986883163452 + ], + "logprob_reward": -0.07898691296577454, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898691296577454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11813240498304367, + 0.0473940446972847, + 0.03291905298829079, + 0.03979456052184105, + 0.08166474103927612, + 0.11461913585662842, + 0.11738135665655136, + 0.07946959882974625, + 0.054859571158885956, + 0.3137655556201935 + ], + "credit_map": [ + 0.015304258093237877, + 0.017573602497577667, + 0.012082938104867935, + 0.01600508950650692, + 0.011909938417375088, + 0.024614600464701653, + 0.012162853963673115, + 0.008479123003780842, + 0.005100583657622337, + 0.005997183732688427, + 0.006580327171832323, + 0.006114244926720858, + 0.007679385133087635, + 0.008335688151419163, + 0.004136832896620035, + 0.0034498018212616444, + 0.0037173910532146692, + 0.004741736222058535, + 0.004715165589004755, + 0.006167818792164326, + 0.004907319787889719, + 0.004395657684653997, + 0.0025427727960050106, + 0.0017311908304691315, + 0.0034560717176645994, + 0.004679264035075903, + 0.00420959061011672, + 0.006676205899566412, + 0.007622593082487583, + 0.004737643525004387, + 0.005095468368381262, + 0.003317723050713539, + 0.012324904091656208, + 0.011331956833600998, + 0.006593928672373295, + 0.008914987556636333, + 0.01665479503571987, + 0.008690403774380684, + 0.008188069798052311, + 0.008965696208178997, + 0.01697760634124279, + 0.013971714302897453, + 0.013199792243540287, + 0.01506785023957491, + 0.014156823977828026, + 0.01407715305685997, + 0.012752936221659184, + 0.014415259473025799, + 0.013668661937117577, + 0.0132640041410923, + 0.020272666588425636, + 0.014543574303388596, + 0.010512631386518478, + 0.017865024507045746, + 0.01083424687385559, + 0.01642054133117199, + 0.0083383210003376, + 0.008720067329704762, + 0.01036578044295311, + 0.010898659937083721, + 0.011566407978534698, + 0.01147218607366085, + 0.011022789403796196, + 0.007085392251610756, + 0.0042436872608959675, + 0.008981055580079556, + 0.0048553962260484695, + 0.006953060626983643, + 0.006403590552508831, + 0.010290302336215973, + 0.008187566883862019, + 0.004944911692291498, + 0.03141548112034798, + 0.05240428447723389, + 0.05758628621697426, + 0.024666741490364075, + 0.02073856070637703, + 0.06562059372663498, + 0.022961104288697243, + 0.03837250545620918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.07898691296577454 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015147175872698426 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015147175872698426, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015147175872698426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09038349241018295, + 0.05155152454972267, + 0.09319813549518585, + 0.051133740693330765, + 0.17738187313079834, + 0.1969374120235443, + 0.15773992240428925, + 0.0690920501947403, + 0.060657430440187454, + 0.05192440748214722 + ], + "credit_map": [ + 0.016225304454565048, + 0.011435125023126602, + 0.007863685488700867, + 0.014818723313510418, + 0.009054834954440594, + 0.014703098684549332, + 0.008593241684138775, + 0.0076894741505384445, + 0.01029032003134489, + 0.006537789478898048, + 0.00535667734220624, + 0.008387474343180656, + 0.004106044303625822, + 0.010791638866066933, + 0.00357425375841558, + 0.0025073285214602947, + 0.010675733909010887, + 0.011028673499822617, + 0.008510030806064606, + 0.020005736500024796, + 0.015846390277147293, + 0.013320572674274445, + 0.008350048214197159, + 0.005460947286337614, + 0.005050018895417452, + 0.007687551435083151, + 0.004892493598163128, + 0.005864856299012899, + 0.011485403403639793, + 0.009118620306253433, + 0.0041841925121843815, + 0.0028506023809313774, + 0.01766037382185459, + 0.024674197658896446, + 0.01185077615082264, + 0.018431926146149635, + 0.049304451793432236, + 0.022445838898420334, + 0.017579898238182068, + 0.015434423461556435, + 0.020866958424448967, + 0.028808005154132843, + 0.024500204250216484, + 0.019271137192845345, + 0.017581582069396973, + 0.046762872487306595, + 0.013189597986638546, + 0.02595706097781658, + 0.017761562019586563, + 0.02050102688372135, + 0.01760481297969818, + 0.01720087043941021, + 0.02130766585469246, + 0.030121231451630592, + 0.01566179469227791, + 0.017580950632691383, + 0.007227002643048763, + 0.011279343627393246, + 0.0081615736708045, + 0.007665526587516069, + 0.008324509486556053, + 0.012586253695189953, + 0.007384764961898327, + 0.00646307785063982, + 0.006758555304259062, + 0.006842470727860928, + 0.009462211281061172, + 0.007216763682663441, + 0.0066079627722501755, + 0.006765082012861967, + 0.007211972959339619, + 0.009792412631213665, + 0.004589652642607689, + 0.008264359086751938, + 0.007637710310518742, + 0.00455549405887723, + 0.0056695472449064255, + 0.010702267289161682, + 0.0046853781677782536, + 0.005820000544190407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.0015147175872698426 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "drone's wide-angle camera slowly zooms out, capturing a dreamlike scene with an incredibly slow and smooth motion. Under a deep, inky indigo sky, a brilliant full moon hangs high, dotted with sparse stars. At the center of the frame, a mysterious figure clad in a flowing white silk robe, her long hair flowing in the wind, serenely rides atop a luminous mythical beast. This beast, part deer, part dragon, its body seemingly composed of condensed starlight. Its translucent scales shimmer with a soft, steady blue-white glow, gently illuminating both its rider and the surrounding night. With a gravity-defying grace, they glide forward steadily and deliberately, their hooves treading the void as if floating on an invisible river of air. Below, the sprawling Forbidden City palaces stretch endlessly. Their iconic golden glazed roofs, illuminated by the moonlight and the beast's radiance, shimmer with a brilliant luster, like a frozen golden ocean. The entire scene is silent, with only the hem of the robe fluttering gently in the breeze, creating an extremely peaceful, solemn, and surreal dreamlike atmosphere.", + "seed": 13473, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_6/index_5043_seed_13473_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07907241128850728, + "vlm_logprob": -0.07907241128850728, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.07907241128850728, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08384408056735992 + ], + "opposite_logprob_tokens": [ + -2.7088441848754883 + ], + "logprob_reward": -0.08384408056735992, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08384408056735992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.172067791223526, + 0.09791779518127441, + 0.2210969477891922, + 0.05522683262825012, + 0.11356550455093384, + 0.21832366287708282, + 0.03483571857213974, + 0.033947642892599106, + 0.03986753523349762, + 0.013150620274245739 + ], + "credit_map": [ + 0.02348010614514351, + 0.01875288411974907, + 0.01929343119263649, + 0.015741316601634026, + 0.03388940915465355, + 0.022374670952558517, + 0.021441027522087097, + 0.01709495484828949, + 0.010932072065770626, + 0.014304823242127895, + 0.013900317251682281, + 0.015439298003911972, + 0.008627701550722122, + 0.013303129002451897, + 0.014204724691808224, + 0.007205725647509098, + 0.0262102410197258, + 0.030365709215402603, + 0.03158842399716377, + 0.014875173568725586, + 0.04270104691386223, + 0.029811134561896324, + 0.016887513920664787, + 0.028657695278525352, + 0.0061622015200555325, + 0.009405416436493397, + 0.005952642299234867, + 0.0047978912480175495, + 0.009849857538938522, + 0.007563789375126362, + 0.005927274934947491, + 0.005567758344113827, + 0.010949600487947464, + 0.014283478260040283, + 0.015573631972074509, + 0.009462802670896053, + 0.008016870357096195, + 0.012271548621356487, + 0.015430865809321404, + 0.02757670357823372, + 0.026514463126659393, + 0.03252696618437767, + 0.01885436661541462, + 0.022288700565695763, + 0.03985583409667015, + 0.0368674173951149, + 0.017749058082699776, + 0.023666860535740852, + 0.005192839540541172, + 0.0036959454882889986, + 0.004652479197829962, + 0.0028964125085622072, + 0.004226034972816706, + 0.00485470425337553, + 0.0036783732939511538, + 0.005638928152620792, + 0.0054439944215118885, + 0.004071451723575592, + 0.003982880152761936, + 0.0034787848126143217, + 0.004271378740668297, + 0.004276761785149574, + 0.0034238663502037525, + 0.004998521879315376, + 0.004746323451399803, + 0.00330567779019475, + 0.006844190880656242, + 0.00312397931702435, + 0.00474746897816658, + 0.006174168083816767, + 0.004191282205283642, + 0.0067344470880925655, + 0.0011003934778273106, + 0.0013979359064251184, + 0.0031205867417156696, + 0.0010199396638199687, + 0.0012920862063765526, + 0.001736728474497795, + 0.0014150476781651378, + 0.002067902823910117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.08384408056735992 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060052566230297 + ], + "opposite_logprob_tokens": [ + -3.8956005573272705 + ], + "logprob_reward": -0.02060052566230297, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060052566230297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10712423920631409, + 0.057030584663152695, + 0.1669643670320511, + 0.023190001025795937, + 0.023163078352808952, + 0.16016314923763275, + 0.07653919607400894, + 0.3481355607509613, + 0.012179306708276272, + 0.025510502979159355 + ], + "credit_map": [ + 0.014696162194013596, + 0.011541604064404964, + 0.007907475344836712, + 0.01104042585939169, + 0.022602852433919907, + 0.015610672533512115, + 0.011979824863374233, + 0.011745207011699677, + 0.007063802797347307, + 0.00635768985375762, + 0.00560518866404891, + 0.00760760810226202, + 0.012139998376369476, + 0.0073386370204389095, + 0.005888945888727903, + 0.005028711166232824, + 0.019843820482492447, + 0.023022707551717758, + 0.02753826230764389, + 0.010505735874176025, + 0.027644632384181023, + 0.021076958626508713, + 0.01343432068824768, + 0.02389790304005146, + 0.003643662901595235, + 0.0026404382660984993, + 0.0035514915362000465, + 0.0017486923607066274, + 0.0029252234380692244, + 0.002958808559924364, + 0.0029924744740128517, + 0.0027292068116366863, + 0.0023290389217436314, + 0.004439976532012224, + 0.00233857543207705, + 0.0015656959731131792, + 0.0017084531718865037, + 0.003020357107743621, + 0.0025367203634232283, + 0.005224259570240974, + 0.02012118324637413, + 0.025515105575323105, + 0.011790847405791283, + 0.019622279331088066, + 0.022069962695240974, + 0.02242029644548893, + 0.021106475964188576, + 0.017516987398266792, + 0.009815127588808537, + 0.010751557536423206, + 0.008119282312691212, + 0.006550132762640715, + 0.013378138653934002, + 0.011254585348069668, + 0.0064332797192037106, + 0.01023708377033472, + 0.0602169930934906, + 0.04307904839515686, + 0.04037279635667801, + 0.0340074859559536, + 0.046646833419799805, + 0.04747336730360985, + 0.03211827948689461, + 0.044220756739377975, + 0.0015739607624709606, + 0.001321570947766304, + 0.0016885061049833894, + 0.0013246505986899137, + 0.0018471833318471909, + 0.0018365487921983004, + 0.0012019035639241338, + 0.0013849815586581826, + 0.002857259474694729, + 0.00281047192402184, + 0.002739222254604101, + 0.0024709105491638184, + 0.004171689972281456, + 0.003934284206479788, + 0.0027784896083176136, + 0.0037481714971363544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 5 + ], + "selected_reward": -0.02060052566230297 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636402267962694 + ], + "opposite_logprob_tokens": [ + -4.882636547088623 + ], + "logprob_reward": -0.007636402267962694, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636402267962694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19245772063732147, + 0.06931274384260178, + 0.09608995914459229, + 0.04421515390276909, + 0.053606342524290085, + 0.0771336480975151, + 0.07770524173974991, + 0.050823893398046494, + 0.11596813797950745, + 0.22268718481063843 + ], + "credit_map": [ + 0.025546271353960037, + 0.02384314313530922, + 0.022477727383375168, + 0.022736862301826477, + 0.0347207710146904, + 0.024586830288171768, + 0.020838364958763123, + 0.01770772598683834, + 0.011228220537304878, + 0.007873422466218472, + 0.009218125604093075, + 0.005586512386798859, + 0.011407868005335331, + 0.009559948928654194, + 0.007539082784205675, + 0.00689955148845911, + 0.011538513004779816, + 0.012797914445400238, + 0.01656089350581169, + 0.006395036820322275, + 0.014659380540251732, + 0.012805966660380363, + 0.009077952243387699, + 0.012254291214048862, + 0.005529982503503561, + 0.00952368974685669, + 0.0042161475867033005, + 0.0036487197503447533, + 0.0080105597153306, + 0.006677714642137289, + 0.00343969720415771, + 0.0031686387956142426, + 0.0051021939143538475, + 0.01193519588559866, + 0.00588085874915123, + 0.007560441270470619, + 0.005438791587948799, + 0.008019455708563328, + 0.005483241286128759, + 0.004186155274510384, + 0.008691096678376198, + 0.011692841537296772, + 0.0067034331150352955, + 0.009332559071481228, + 0.00963352806866169, + 0.01152180228382349, + 0.0082979965955019, + 0.011260380037128925, + 0.009857402183115482, + 0.009657824411988258, + 0.010236073285341263, + 0.005982905626296997, + 0.010811950080096722, + 0.010999728925526142, + 0.006757414899766445, + 0.013401939533650875, + 0.007047302555292845, + 0.006441759876906872, + 0.0060632312670350075, + 0.006090873386710882, + 0.006834161467850208, + 0.0056450255215168, + 0.005449109710752964, + 0.007252425886690617, + 0.014980676583945751, + 0.014926006086170673, + 0.010934135876595974, + 0.015439559705555439, + 0.022334881126880646, + 0.019027432426810265, + 0.011341684497892857, + 0.0069837565533816814, + 0.024940405040979385, + 0.025398142635822296, + 0.023823078721761703, + 0.026418304070830345, + 0.044813867658376694, + 0.03340590372681618, + 0.026562921702861786, + 0.017324551939964294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.007636402267962694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.281338632106781 + ], + "opposite_logprob_tokens": [ + -1.4063386917114258 + ], + "logprob_reward": -0.281338632106781, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.281338632106781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1641988903284073, + 0.06492152810096741, + 0.1408386379480362, + 0.06375399231910706, + 0.037045225501060486, + 0.23250292241573334, + 0.14062869548797607, + 0.09793858230113983, + 0.030693581327795982, + 0.027477938681840897 + ], + "credit_map": [ + 0.02116692066192627, + 0.01818263530731201, + 0.016689157113432884, + 0.016307998448610306, + 0.03598842769861221, + 0.02139976993203163, + 0.01790469139814377, + 0.016559286043047905, + 0.009892704896628857, + 0.006904170382767916, + 0.007096854504197836, + 0.005752622615545988, + 0.012932251207530499, + 0.009047639556229115, + 0.006115608848631382, + 0.007179676089435816, + 0.016112390905618668, + 0.0211567971855402, + 0.021264605224132538, + 0.010768373496830463, + 0.026467543095350266, + 0.018664473667740822, + 0.009671838954091072, + 0.01673261448740959, + 0.008278944529592991, + 0.011488658376038074, + 0.009085520170629025, + 0.004997391253709793, + 0.010082411579787731, + 0.00841491762548685, + 0.006070614326745272, + 0.005335527937859297, + 0.004409019835293293, + 0.00525300856679678, + 0.003716730047017336, + 0.004312461242079735, + 0.0029713630210608244, + 0.005338799208402634, + 0.004339765291661024, + 0.006704075261950493, + 0.02479168400168419, + 0.024518102407455444, + 0.02318469248712063, + 0.024415360763669014, + 0.028940780088305473, + 0.030101295560598373, + 0.03644648939371109, + 0.04010452330112457, + 0.019408883526921272, + 0.014425672590732574, + 0.018178781494498253, + 0.011242356151342392, + 0.01665368117392063, + 0.020798446610569954, + 0.014623143710196018, + 0.025297736749053, + 0.01643672026693821, + 0.011960537172853947, + 0.010949531570076942, + 0.00905250571668148, + 0.015121554024517536, + 0.013040844351053238, + 0.009793885052204132, + 0.011583002284169197, + 0.0035571586340665817, + 0.0037462429609149694, + 0.003954574000090361, + 0.002368025481700897, + 0.005352729931473732, + 0.0047891633585095406, + 0.0031030273530632257, + 0.0038226607721298933, + 0.0036306267138570547, + 0.0026476234197616577, + 0.00512078357860446, + 0.002556476043537259, + 0.004009603057056665, + 0.0038289951626211405, + 0.0027550524100661278, + 0.0029287789948284626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.281338632106781 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a drone perspective that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo with a brilliant full moon and sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure in a flowing white silk robe riding a mythical beast composed of condensed starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the Forbidden City palaces with golden glazed roofs visible below the flying figures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the figures glide forward with a gravity-defying motion while the camera zooms out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019424158381298184 + ], + "opposite_logprob_tokens": [ + -6.2519426345825195 + ], + "logprob_reward": -0.0019424158381298184, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019424158381298184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17265841364860535, + 0.06666403263807297, + 0.2524186372756958, + 0.033860720694065094, + 0.1580122411251068, + 0.15422452986240387, + 0.03684823587536812, + 0.04518676549196243, + 0.04911110922694206, + 0.03101538121700287 + ], + "credit_map": [ + 0.02376719005405903, + 0.018753929063677788, + 0.019063076004385948, + 0.0144277298822999, + 0.03353871405124664, + 0.022754637524485588, + 0.021534422412514687, + 0.0188186913728714, + 0.007333265617489815, + 0.009982949122786522, + 0.008701442740857601, + 0.011346688494086266, + 0.006163290701806545, + 0.008786858059465885, + 0.009546765126287937, + 0.004802766256034374, + 0.031830985099077225, + 0.0344376266002655, + 0.03892478719353676, + 0.018699433654546738, + 0.04644390568137169, + 0.03427930921316147, + 0.01787017472088337, + 0.02993239276111126, + 0.003750452073290944, + 0.004314566031098366, + 0.005789247341454029, + 0.003370879450812936, + 0.0035896198824048042, + 0.0038060247898101807, + 0.004081659484654665, + 0.0051582674495875835, + 0.01582047902047634, + 0.02355690859258175, + 0.0211483146995306, + 0.014962669461965561, + 0.012343566864728928, + 0.017844153568148613, + 0.018386779353022575, + 0.033949337899684906, + 0.01775500550866127, + 0.021161586046218872, + 0.01491024624556303, + 0.01480341050773859, + 0.025055089965462685, + 0.022525591775774956, + 0.015481086447834969, + 0.022532496601343155, + 0.004939197096973658, + 0.0039362977258861065, + 0.005545588210225105, + 0.002873758552595973, + 0.004210229031741619, + 0.005248593632131815, + 0.0037573999725282192, + 0.006337166763842106, + 0.008316081017255783, + 0.00583465164527297, + 0.004735327325761318, + 0.003961218986660242, + 0.007203489076346159, + 0.007155987434089184, + 0.003393487073481083, + 0.004586515948176384, + 0.004970714915543795, + 0.0054879640229046345, + 0.00795755721628666, + 0.0038285821210592985, + 0.00870010070502758, + 0.007020802237093449, + 0.004673598799854517, + 0.006471782457083464, + 0.004147490952163935, + 0.0031435429118573666, + 0.0041948240250349045, + 0.0027448059991002083, + 0.004010582342743874, + 0.004592494107782841, + 0.0034519049804657698, + 0.004729731008410454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.0019424158381298184 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_7/index_tensor([6050, 6050])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_7/index_tensor([6050, 6050])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..937734e06b400c87293d1b1fe98848ec904c85b5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_7/index_tensor([6050, 6050])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a spectacular Final Fantasy-inspired scene, the camera, in extreme close-up, fixes itself on the amber eyes of an ancient bird, ablaze with an inner light, its reptilian pupils narrowed in concentration. The behemoth dwells deep within a magical forest shrouded in twilight. Countless particles of aetherial energy swirl in the air, slowly swirling like stardust, casting a dazzling array of light and shadow on the surrounding massive, ancient trees covered in luminescent moss. The bird's body is covered in obsidian-hard scales stretching from its neck to its back, seamlessly transitioning into iridescent feathers that shimmer with a myriad of metallic luster in the dim light. Suddenly, with a brief, dazzling flash of energy erupting from its chest, the bird unfurls its massive claws and wings with explosive force, stirring up a gust of air laced with luminous dust. At the moment when these wings, which are a combination of gorgeous feathers and tough membranes and have sharp claws extending from the joints, are fully unfolded, the focus of the camera switches as fast as lightning, from its amber eyes full of wisdom and majesty to the farthest wingtips after full folding, clearly capturing the extreme details of the feather tips trembling slightly against the blurred, dreamy forest background. The whole picture is full of epic majesty and awe-inspiring magic.", + "seed": 883, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_7/index_6050_seed_883_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009046601643785835, + "vlm_logprob": -0.009046601643785835, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.009046601643785835, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003049844177439809 + ], + "opposite_logprob_tokens": [ + -6.628049850463867 + ], + "logprob_reward": -0.003049844177439809, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.003049844177439809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08198715001344681, + 0.058164604008197784, + 0.13225962221622467, + 0.07406018674373627, + 0.060606472194194794, + 0.10696611553430557, + 0.09086816757917404, + 0.09085256606340408, + 0.08236957341432571, + 0.22186553478240967 + ], + "credit_map": [ + 0.007011160720139742, + 0.012054233811795712, + 0.01481800526380539, + 0.008403941057622433, + 0.009208881296217442, + 0.009306089021265507, + 0.01091927383095026, + 0.010265568271279335, + 0.005133944563567638, + 0.006393930874764919, + 0.006677356082946062, + 0.00874857883900404, + 0.005733705125749111, + 0.0069311452098190784, + 0.010802351869642735, + 0.007743590511381626, + 0.01129613071680069, + 0.016983268782496452, + 0.01593022607266903, + 0.015264312736690044, + 0.017978772521018982, + 0.015516815707087517, + 0.020445242524147034, + 0.01884486712515354, + 0.010519367642700672, + 0.010180799290537834, + 0.010424280539155006, + 0.00795428454875946, + 0.008999115787446499, + 0.010009021498262882, + 0.008727983571588993, + 0.007245327811688185, + 0.005710619036108255, + 0.007942620664834976, + 0.006607372313737869, + 0.010752145200967789, + 0.009586241096258163, + 0.0031902045011520386, + 0.009046215564012527, + 0.0077710528858006, + 0.008213860914111137, + 0.017601583153009415, + 0.012904530391097069, + 0.010218980722129345, + 0.012538298964500427, + 0.010213788598775864, + 0.014768324792385101, + 0.02050674706697464, + 0.008337573148310184, + 0.015440844930708408, + 0.01337029691785574, + 0.010329112410545349, + 0.007498167920857668, + 0.008738006465137005, + 0.014036129228770733, + 0.013118039816617966, + 0.011016704142093658, + 0.006102255545556545, + 0.01993570849299431, + 0.013323581777513027, + 0.005348849110305309, + 0.010541724972426891, + 0.008506113663315773, + 0.016077624633908272, + 0.01179695874452591, + 0.006208735052496195, + 0.011464511044323444, + 0.013956326991319656, + 0.004265064373612404, + 0.008880606852471828, + 0.013064384460449219, + 0.012732981704175472, + 0.011764186434447765, + 0.007810997311025858, + 0.036913104355335236, + 0.00909663550555706, + 0.01008771825581789, + 0.0538940392434597, + 0.011595170013606548, + 0.08070366829633713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 5 + ], + "selected_reward": -0.003049844177439809 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012551246210932732 + ], + "opposite_logprob_tokens": [ + -4.387551307678223 + ], + "logprob_reward": -0.012551246210932732, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012551246210932732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07474327832460403, + 0.06123858690261841, + 0.08956851065158844, + 0.05172085762023926, + 0.08276382088661194, + 0.19442500174045563, + 0.18781159818172455, + 0.06456630676984787, + 0.03938216716051102, + 0.15377984941005707 + ], + "credit_map": [ + 0.0057501341216266155, + 0.00669978279620409, + 0.011772261001169682, + 0.009969284757971764, + 0.008235139772295952, + 0.009521003812551498, + 0.013212313875555992, + 0.009583358652889729, + 0.00793644692748785, + 0.005249562673270702, + 0.00797752570360899, + 0.008103243075311184, + 0.005551523994654417, + 0.00903383083641529, + 0.009725215844810009, + 0.007661235518753529, + 0.007919998839497566, + 0.008568309247493744, + 0.017248723655939102, + 0.014507079496979713, + 0.009248114190995693, + 0.009179485030472279, + 0.01136454101651907, + 0.011532262898981571, + 0.005969091318547726, + 0.00797270331531763, + 0.007413617800921202, + 0.005748683121055365, + 0.005927423946559429, + 0.007407838944345713, + 0.00582986930385232, + 0.005451630335301161, + 0.006038197316229343, + 0.007849573157727718, + 0.006735191680490971, + 0.014963093213737011, + 0.006716875825077295, + 0.00332298269495368, + 0.01904877834022045, + 0.018089130520820618, + 0.011813902296125889, + 0.009016015566885471, + 0.05596006289124489, + 0.009702338837087154, + 0.010933580808341503, + 0.010248180478811264, + 0.01840878836810589, + 0.06834213435649872, + 0.018554996699094772, + 0.02527856081724167, + 0.027467966079711914, + 0.018357204273343086, + 0.01483545359224081, + 0.018222078680992126, + 0.03281227499246597, + 0.03228304907679558, + 0.008544178679585457, + 0.004579274915158749, + 0.010518338531255722, + 0.008203336037695408, + 0.00490591349080205, + 0.008616755716502666, + 0.008964757435023785, + 0.010233751498162746, + 0.0051581780426204205, + 0.005013082642108202, + 0.006002077367156744, + 0.00604277802631259, + 0.0030917071271687746, + 0.0038682373706251383, + 0.005023793783038855, + 0.005182311404496431, + 0.015080432407557964, + 0.010946149006485939, + 0.028834354132413864, + 0.012144848704338074, + 0.012207821942865849, + 0.03374306485056877, + 0.011866181157529354, + 0.028957000002264977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.012551246210932732 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002819374669343233 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.002819374669343233, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002819374669343233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08644912391901016, + 0.06488602608442307, + 0.09151910245418549, + 0.18099085986614227, + 0.14138264954090118, + 0.10178741067647934, + 0.06629353016614914, + 0.05121735855937004, + 0.09330086410045624, + 0.12217309325933456 + ], + "credit_map": [ + 0.008196506649255753, + 0.010928772389888763, + 0.0140652721747756, + 0.010883301496505737, + 0.01024363562464714, + 0.009685022756457329, + 0.012244646437466145, + 0.010201960802078247, + 0.006507681217044592, + 0.008498109877109528, + 0.008627601899206638, + 0.010411776602268219, + 0.006204243283718824, + 0.008246053010225296, + 0.009509650059044361, + 0.006880912464112043, + 0.00719080213457346, + 0.013455700129270554, + 0.01293923333287239, + 0.01575508899986744, + 0.009614133276045322, + 0.005066675134003162, + 0.015437284484505653, + 0.01206018403172493, + 0.017132936045527458, + 0.0264231339097023, + 0.028506508097052574, + 0.022827213630080223, + 0.025185344740748405, + 0.017621150240302086, + 0.025673827156424522, + 0.01762075163424015, + 0.008542039431631565, + 0.011904655024409294, + 0.013149109669029713, + 0.022416822612285614, + 0.010377530008554459, + 0.0066573889926075935, + 0.036114513874053955, + 0.032220590859651566, + 0.008046053349971771, + 0.008582204580307007, + 0.024684695526957512, + 0.008481905795633793, + 0.009047547355294228, + 0.008552821353077888, + 0.00873684510588646, + 0.025655340403318405, + 0.007962833158671856, + 0.009572400711476803, + 0.009623490273952484, + 0.006890890654176474, + 0.00536580104380846, + 0.00643174396827817, + 0.010684628039598465, + 0.009761737659573555, + 0.008249185048043728, + 0.003515619318932295, + 0.007740223314613104, + 0.005685973446816206, + 0.003850442124530673, + 0.006496296729892492, + 0.006979058962315321, + 0.00870056077837944, + 0.01576758734881878, + 0.012509937398135662, + 0.012608460150659084, + 0.012751149944961071, + 0.0079297861084342, + 0.01141500286757946, + 0.010659784078598022, + 0.009659160859882832, + 0.013788782060146332, + 0.009102215059101582, + 0.020659558475017548, + 0.012029141187667847, + 0.009191516786813736, + 0.022732026875019073, + 0.012199978344142437, + 0.02246987260878086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.002819374669343233 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633549325168133 + ], + "opposite_logprob_tokens": [ + -4.758633613586426 + ], + "logprob_reward": -0.008633549325168133, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633549325168133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05059266835451126, + 0.046261467039585114, + 0.06480984389781952, + 0.06184143200516701, + 0.041016723960638046, + 0.08573921024799347, + 0.07379882037639618, + 0.022289596498012543, + 0.07320723682641983, + 0.48044300079345703 + ], + "credit_map": [ + 0.004455971997231245, + 0.009419522248208523, + 0.006971892900764942, + 0.005156957544386387, + 0.007373160682618618, + 0.005567842163145542, + 0.006138977594673634, + 0.005508344154804945, + 0.006101161241531372, + 0.0050678132101893425, + 0.005443506874144077, + 0.007151711732149124, + 0.004009505733847618, + 0.006840925198048353, + 0.00641277153044939, + 0.0052340710535645485, + 0.006242178846150637, + 0.006119798868894577, + 0.01267067901790142, + 0.01182520855218172, + 0.00665319012477994, + 0.007438329979777336, + 0.00728706456720829, + 0.0065734004601836205, + 0.007367343176156282, + 0.007496444042772055, + 0.008005497045814991, + 0.005083871074020863, + 0.008002696558833122, + 0.011290788650512695, + 0.006999259814620018, + 0.007595535833388567, + 0.004411321133375168, + 0.004221159033477306, + 0.004796913359314203, + 0.00674093421548605, + 0.006316916085779667, + 0.002259907778352499, + 0.006165947765111923, + 0.006103625521063805, + 0.00523708201944828, + 0.00761893205344677, + 0.01819274015724659, + 0.005402479320764542, + 0.007803329732269049, + 0.005383193027228117, + 0.012122883461415768, + 0.023978574201464653, + 0.007357909809798002, + 0.006001888308674097, + 0.011296983808279037, + 0.007915648631751537, + 0.007516531273722649, + 0.006559786852449179, + 0.015078013762831688, + 0.012072053737938404, + 0.0032726083882153034, + 0.0015120187308639288, + 0.0045007565058767796, + 0.0026152387727051973, + 0.00128956139087677, + 0.0018925206968560815, + 0.0025713187642395496, + 0.004635572899132967, + 0.015304861590266228, + 0.00828382559120655, + 0.010262219235301018, + 0.008829579688608646, + 0.004607211332768202, + 0.0066379825584590435, + 0.00823479425162077, + 0.011046761646866798, + 0.01670309714972973, + 0.011819707229733467, + 0.07625886797904968, + 0.01711643673479557, + 0.016205286607146263, + 0.13159354031085968, + 0.02878672443330288, + 0.1819593608379364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.008633549325168133 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018178993836045265 + ], + "opposite_logprob_tokens": [ + -4.018178939819336 + ], + "logprob_reward": -0.018178993836045265, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018178993836045265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1078794077038765, + 0.05373255908489227, + 0.12485086917877197, + 0.03934933990240097, + 0.09025759249925613, + 0.2759159803390503, + 0.051268454641103745, + 0.0746680423617363, + 0.07062830775976181, + 0.11144942045211792 + ], + "credit_map": [ + 0.008177561685442924, + 0.016485868021845818, + 0.016060687601566315, + 0.015485279262065887, + 0.012615978717803955, + 0.010156963020563126, + 0.01522392500191927, + 0.013673143461346626, + 0.004656519740819931, + 0.004276367835700512, + 0.0068926759995520115, + 0.009130164049565792, + 0.005058595445007086, + 0.007025478407740593, + 0.009628995321691036, + 0.007063764613121748, + 0.009924888610839844, + 0.014395635575056076, + 0.015646804124116898, + 0.016435157507658005, + 0.012088148854672909, + 0.014385306276381016, + 0.0211785938590765, + 0.020796330645680428, + 0.005989453289657831, + 0.005585948470979929, + 0.005301211960613728, + 0.003757817903533578, + 0.004693026654422283, + 0.004988763015717268, + 0.0047796256840229034, + 0.00425349548459053, + 0.005888096522539854, + 0.005939875729382038, + 0.008090737275779247, + 0.010926052927970886, + 0.007544871419668198, + 0.003756872611120343, + 0.0255101528018713, + 0.0226009301841259, + 0.019406568259000778, + 0.03371160477399826, + 0.04880697280168533, + 0.016508523374795914, + 0.022555610164999962, + 0.02007674239575863, + 0.03071622923016548, + 0.08413371443748474, + 0.005882301367819309, + 0.0068734195083379745, + 0.0071318792179226875, + 0.005813817493617535, + 0.004436958581209183, + 0.006599010434001684, + 0.0073832012712955475, + 0.007147867698222399, + 0.011269125156104565, + 0.006348731461912394, + 0.010924707166850567, + 0.00806313008069992, + 0.006953517906367779, + 0.011892915703356266, + 0.009647483006119728, + 0.009568431414663792, + 0.0073683494701981544, + 0.007069634739309549, + 0.011281315237283707, + 0.012033149600028992, + 0.006346933078020811, + 0.005905100144445896, + 0.009434405714273453, + 0.011189417913556099, + 0.015021839179098606, + 0.008517598733305931, + 0.013899846002459526, + 0.012270970270037651, + 0.009315780363976955, + 0.023603392764925957, + 0.010434134863317013, + 0.018385857343673706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 9 + ], + "selected_reward": -0.018178993836045265 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a spectacular Final Fantasy-inspired scene, the camera, in extreme close-up, fixes itself on the amber eyes of an ancient bird, ablaze with an inner light, its reptilian pupils narrowed in concentration. The behemoth dwells deep within a magical forest shrouded in twilight. Countless particles of aetherial energy swirl in the air, slowly swirling like stardust, casting a dazzling array of light and shadow on the surrounding massive, ancient trees covered in luminescent moss. The bird's body is covered in obsidian-hard scales stretching from its neck to its back, seamlessly transitioning into iridescent feathers that shimmer with a myriad of metallic luster in the dim light. Suddenly, with a brief, dazzling flash of energy erupting from its chest, the bird unfurls its massive claws and wings with explosive force, stirring up a gust of air laced with luminous dust. At the moment when these wings, which are a combination of gorgeous feathers and tough membranes and have sharp claws extending from the joints, are fully unfolded, the focus of the camera switches as fast as lightning, from its amber eyes full of wisdom and majesty to the farthest wingtips after full folding, clearly capturing the extreme details of the feather tips trembling slightly against the blurred, dreamy forest background. The whole picture is full of epic majesty and awe-inspiring magic.", + "seed": 884, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_7/index_6050_seed_884_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.015628032572567462, + "vlm_logprob": -0.015628032572567462, + "vlm_yes_no_margin": 4.399999952316284, + "vlm_token_credit": -0.015628032572567462, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009826146997511387 + ], + "opposite_logprob_tokens": [ + -4.759826183319092 + ], + "logprob_reward": -0.009826146997511387, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.009826146997511387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07530516386032104, + 0.1080736517906189, + 0.04961927607655525, + 0.17281849682331085, + 0.14339959621429443, + 0.06404086202383041, + 0.07999244332313538, + 0.06402292102575302, + 0.17013148963451385, + 0.07259616255760193 + ], + "credit_map": [ + 0.015109865926206112, + 0.008529762737452984, + 0.007264602463692427, + 0.006295471917837858, + 0.009013879112899303, + 0.011402388103306293, + 0.006886809598654509, + 0.01080238912254572, + 0.014596989378333092, + 0.0135794123634696, + 0.010447125881910324, + 0.012274953536689281, + 0.018752608448266983, + 0.01403013989329338, + 0.014806138351559639, + 0.00958628486841917, + 0.005410036072134972, + 0.004831732250750065, + 0.006797173526138067, + 0.0071468595415353775, + 0.008193845860660076, + 0.0050995540805161, + 0.006215676199644804, + 0.005924398545175791, + 0.022955501452088356, + 0.030317794531583786, + 0.017910409718751907, + 0.018582656979560852, + 0.023607628419995308, + 0.015300897881388664, + 0.02421785704791546, + 0.01992575079202652, + 0.013242142274975777, + 0.01633870042860508, + 0.01508315745741129, + 0.020987659692764282, + 0.021213408559560776, + 0.023607898503541946, + 0.01824868656694889, + 0.014677933417260647, + 0.014203901402652264, + 0.006046642549335957, + 0.006191595457494259, + 0.006120515055954456, + 0.010369339026510715, + 0.005812518764287233, + 0.008793692104518414, + 0.006502660922706127, + 0.008208448998630047, + 0.005657748319208622, + 0.014959864318370819, + 0.007804942317306995, + 0.006170365028083324, + 0.01198212243616581, + 0.00913662277162075, + 0.01607232540845871, + 0.008110742084681988, + 0.004128185100853443, + 0.01002512127161026, + 0.005687663797289133, + 0.0071282959543168545, + 0.007456917315721512, + 0.009081240743398666, + 0.012404760345816612, + 0.01869291253387928, + 0.015416886657476425, + 0.021140476688742638, + 0.023088602349162102, + 0.01825159229338169, + 0.021239403635263443, + 0.023689059540629387, + 0.028612565249204636, + 0.008703187108039856, + 0.006496695801615715, + 0.01145799458026886, + 0.009437677450478077, + 0.0076780300587415695, + 0.008330149576067924, + 0.009660722687840462, + 0.010831701569259167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -0.009826146997511387 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020595969632267952 + ], + "opposite_logprob_tokens": [ + -3.8955960273742676 + ], + "logprob_reward": -0.020595969632267952, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020595969632267952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1295853555202484, + 0.10741329938173294, + 0.056380029767751694, + 0.11400197446346283, + 0.08509855717420578, + 0.06800750643014908, + 0.08626428246498108, + 0.06609059870243073, + 0.15368571877479553, + 0.13347260653972626 + ], + "credit_map": [ + 0.023617707192897797, + 0.012231803499162197, + 0.009724907577037811, + 0.01121826097369194, + 0.01726597175002098, + 0.019520552828907967, + 0.011597730219364166, + 0.024408437311649323, + 0.020947853103280067, + 0.00961028691381216, + 0.008904540911316872, + 0.011502492241561413, + 0.016592122614383698, + 0.01607922464609146, + 0.010950673371553421, + 0.012826114892959595, + 0.005775738041847944, + 0.004426106810569763, + 0.005639982409775257, + 0.009172312915325165, + 0.011130531318485737, + 0.0065825399942696095, + 0.00926822330802679, + 0.004384595900774002, + 0.009343614801764488, + 0.007820360362529755, + 0.02081127092242241, + 0.011135612614452839, + 0.017560819163918495, + 0.016210032626986504, + 0.013969210907816887, + 0.017151059582829475, + 0.010415533557534218, + 0.01166501920670271, + 0.01386647205799818, + 0.010250275023281574, + 0.008857915177941322, + 0.01381358876824379, + 0.009548081085085869, + 0.0066816797479987144, + 0.011103362776339054, + 0.00565725564956665, + 0.007283939514309168, + 0.005666229873895645, + 0.0133874686434865, + 0.006832714658230543, + 0.008753135800361633, + 0.009323404170572758, + 0.00859547033905983, + 0.008483231998980045, + 0.01214426290243864, + 0.013474075123667717, + 0.009802165441215038, + 0.012563850730657578, + 0.01077677309513092, + 0.010424461215734482, + 0.008118268102407455, + 0.006209641229361296, + 0.01229980681091547, + 0.007896067574620247, + 0.005213478580117226, + 0.006184643600136042, + 0.009702674113214016, + 0.01046602800488472, + 0.023561496287584305, + 0.013826350681483746, + 0.021757354959845543, + 0.01841030828654766, + 0.013633567839860916, + 0.025093626230955124, + 0.019870126619935036, + 0.017532890662550926, + 0.0180088859051466, + 0.012378398329019547, + 0.018863169476389885, + 0.017580687999725342, + 0.012640328146517277, + 0.018414542078971863, + 0.014752903021872044, + 0.020833710208535194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.020595969632267952 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031931623816490173 + ], + "opposite_logprob_tokens": [ + -5.753193378448486 + ], + "logprob_reward": -0.0031931623816490173, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031931623816490173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03220611438155174, + 0.07804585993289948, + 0.025823483243584633, + 0.16902433335781097, + 0.32702162861824036, + 0.08508580923080444, + 0.06886757165193558, + 0.024210400879383087, + 0.137889102101326, + 0.05182566121220589 + ], + "credit_map": [ + 0.005593020468950272, + 0.0030857089441269636, + 0.0026423833332955837, + 0.002833171049132943, + 0.004268982447683811, + 0.004633197095245123, + 0.0036820578388869762, + 0.005467594601213932, + 0.01088197622448206, + 0.008221743628382683, + 0.006482821889221668, + 0.006934584118425846, + 0.015285225585103035, + 0.011497833766043186, + 0.007545686326920986, + 0.01119599025696516, + 0.0028120833449065685, + 0.002356625162065029, + 0.002495181979611516, + 0.0040307617746293545, + 0.0037957492750138044, + 0.003469679970294237, + 0.0038950841408222914, + 0.0029683199245482683, + 0.01668786071240902, + 0.0138856777921319, + 0.030037906020879745, + 0.014112424105405807, + 0.02150219865143299, + 0.018182147294282913, + 0.02731396071612835, + 0.02730216458439827, + 0.03266970068216324, + 0.037341948598623276, + 0.056934960186481476, + 0.04683868959546089, + 0.04485144466161728, + 0.045160453766584396, + 0.03567137569189072, + 0.027553077787160873, + 0.007650995627045631, + 0.005184462759643793, + 0.006488103419542313, + 0.006075526587665081, + 0.028688374906778336, + 0.009824762120842934, + 0.015356405638158321, + 0.00581718422472477, + 0.007265368942171335, + 0.005777735263109207, + 0.012354426085948944, + 0.005852255038917065, + 0.007485967129468918, + 0.013174674473702908, + 0.007908649742603302, + 0.009048495441675186, + 0.0023157403338700533, + 0.002234203740954399, + 0.0036721208598464727, + 0.002423899481073022, + 0.0024583698250353336, + 0.003810513997450471, + 0.003147741546854377, + 0.004147812258452177, + 0.019797490909695625, + 0.011242387816309929, + 0.02524898573756218, + 0.010227440856397152, + 0.008389734663069248, + 0.020087216049432755, + 0.014635710045695305, + 0.02826014533638954, + 0.006673836149275303, + 0.0038101456593722105, + 0.01015667337924242, + 0.006761590484529734, + 0.004939510952681303, + 0.007191657554358244, + 0.005866301245987415, + 0.00642594788223505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.0031931623816490173 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184378743171692 + ], + "opposite_logprob_tokens": [ + -4.018184185028076 + ], + "logprob_reward": -0.018184378743171692, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018184378743171692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0707121267914772, + 0.047690000385046005, + 0.04067126289010048, + 0.05397830531001091, + 0.0603153333067894, + 0.06376228481531143, + 0.15397538244724274, + 0.30510345101356506, + 0.10161519795656204, + 0.10217668861150742 + ], + "credit_map": [ + 0.010751358233392239, + 0.007560930214822292, + 0.004561302717775106, + 0.005972896236926317, + 0.009728745557367802, + 0.01116043608635664, + 0.0073081571608781815, + 0.013668294996023178, + 0.007853155955672264, + 0.004641199484467506, + 0.004240271169692278, + 0.004854382015764713, + 0.00908625777810812, + 0.00678982911631465, + 0.005339343566447496, + 0.004885562229901552, + 0.004350971430540085, + 0.0040932707488536835, + 0.005144231952726841, + 0.006086153909564018, + 0.007074788212776184, + 0.0048004863783717155, + 0.004638078156858683, + 0.0044832839630544186, + 0.006590423174202442, + 0.005756490398198366, + 0.007197508122771978, + 0.004636948462575674, + 0.008280660025775433, + 0.007242947351187468, + 0.007771606557071209, + 0.006501717958599329, + 0.007929036393761635, + 0.010367541573941708, + 0.008590446785092354, + 0.008744264021515846, + 0.007329217623919249, + 0.005644381511956453, + 0.006376586854457855, + 0.00533386180177331, + 0.0113522382453084, + 0.005665700417011976, + 0.007384317461401224, + 0.00674302177503705, + 0.006619336549192667, + 0.00797121413052082, + 0.007370008155703545, + 0.010656450875103474, + 0.016818055883049965, + 0.011363505385816097, + 0.030550874769687653, + 0.012754925526678562, + 0.010529516264796257, + 0.030604418367147446, + 0.012622409500181675, + 0.028731681406497955, + 0.03324486315250397, + 0.028751656413078308, + 0.052362456917762756, + 0.036235205829143524, + 0.025062089785933495, + 0.03115171752870083, + 0.04378945380449295, + 0.054506007581949234, + 0.01258863601833582, + 0.007737291976809502, + 0.018768310546875, + 0.010649741627275944, + 0.007448638789355755, + 0.015013252384960651, + 0.017212746664881706, + 0.012196586467325687, + 0.01110188476741314, + 0.008285712450742722, + 0.019168317317962646, + 0.013404318131506443, + 0.008421813137829304, + 0.0155103150755167, + 0.01339709758758545, + 0.012887227348983288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.018184378743171692 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340505108237267 + ], + "opposite_logprob_tokens": [ + -3.6513404846191406 + ], + "logprob_reward": -0.026340505108237267, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340505108237267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0640188604593277, + 0.17091521620750427, + 0.02358090691268444, + 0.19006229937076569, + 0.0908854603767395, + 0.06095016002655029, + 0.11147629469633102, + 0.09347231686115265, + 0.10296295583248138, + 0.09167546778917313 + ], + "credit_map": [ + 0.01280837133526802, + 0.006229185499250889, + 0.005517410114407539, + 0.007100650109350681, + 0.007577050942927599, + 0.008949819020926952, + 0.0061262608505785465, + 0.009710120037198067, + 0.02777354046702385, + 0.01702171191573143, + 0.014278849586844444, + 0.015702465549111366, + 0.03505479544401169, + 0.023914456367492676, + 0.017899751663208008, + 0.019269661977887154, + 0.0033178848680108786, + 0.0018903284799307585, + 0.002117662224918604, + 0.0028400178998708725, + 0.0037419649306684732, + 0.0026337311137467623, + 0.0036856243386864662, + 0.0033536951523274183, + 0.023473959416151047, + 0.02884713001549244, + 0.023231446743011475, + 0.01999008283019066, + 0.02825811877846718, + 0.018582293763756752, + 0.02735225111246109, + 0.020327044650912285, + 0.011194421909749508, + 0.011284316889941692, + 0.010216587223112583, + 0.010402382351458073, + 0.011477809399366379, + 0.013305585831403732, + 0.01291093323379755, + 0.010093427263200283, + 0.008593065664172173, + 0.006369354669004679, + 0.006478318013250828, + 0.005835157353430986, + 0.010643308982253075, + 0.007456307765096426, + 0.006472686305642128, + 0.009101959876716137, + 0.01070080604404211, + 0.011036771349608898, + 0.017139622941613197, + 0.010895395651459694, + 0.014536362141370773, + 0.019898217171430588, + 0.011630286462605, + 0.015638835728168488, + 0.010488523170351982, + 0.010115428827702999, + 0.015270870178937912, + 0.01020178571343422, + 0.00838136300444603, + 0.010526090860366821, + 0.016766831278800964, + 0.011721426621079445, + 0.011361454613506794, + 0.010228509083390236, + 0.014006461948156357, + 0.011814547702670097, + 0.010370206087827682, + 0.014578104950487614, + 0.012783370912075043, + 0.017820313572883606, + 0.013990226201713085, + 0.008481196127831936, + 0.014559422619640827, + 0.012377021834254265, + 0.007574236020445824, + 0.011943594552576542, + 0.008420014753937721, + 0.014329769648611546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 6 + ], + "selected_reward": -0.026340505108237267 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a spectacular Final Fantasy-inspired scene, the camera, in extreme close-up, fixes itself on the amber eyes of an ancient bird, ablaze with an inner light, its reptilian pupils narrowed in concentration. The behemoth dwells deep within a magical forest shrouded in twilight. Countless particles of aetherial energy swirl in the air, slowly swirling like stardust, casting a dazzling array of light and shadow on the surrounding massive, ancient trees covered in luminescent moss. The bird's body is covered in obsidian-hard scales stretching from its neck to its back, seamlessly transitioning into iridescent feathers that shimmer with a myriad of metallic luster in the dim light. Suddenly, with a brief, dazzling flash of energy erupting from its chest, the bird unfurls its massive claws and wings with explosive force, stirring up a gust of air laced with luminous dust. At the moment when these wings, which are a combination of gorgeous feathers and tough membranes and have sharp claws extending from the joints, are fully unfolded, the focus of the camera switches as fast as lightning, from its amber eyes full of wisdom and majesty to the farthest wingtips after full folding, clearly capturing the extreme details of the feather tips trembling slightly against the blurred, dreamy forest background. The whole picture is full of epic majesty and awe-inspiring magic.", + "seed": 885, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_7/index_6050_seed_885_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.011563989520072936, + "vlm_logprob": -0.011563989520072936, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.011563989520072936, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010362748056650162 + ], + "opposite_logprob_tokens": [ + -5.01036262512207 + ], + "logprob_reward": -0.010362748056650162, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.010362748056650162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1685856282711029, + 0.11907882243394852, + 0.08071959018707275, + 0.05036526918411255, + 0.06512860953807831, + 0.0610797181725502, + 0.08056837320327759, + 0.16641296446323395, + 0.10133626312017441, + 0.1067248061299324 + ], + "credit_map": [ + 0.022169794887304306, + 0.0151360547170043, + 0.013124036602675915, + 0.018883679062128067, + 0.02573762647807598, + 0.020875487476587296, + 0.03794793412089348, + 0.014710999093949795, + 0.019182279706001282, + 0.011333824135363102, + 0.011558277532458305, + 0.009117054753005505, + 0.018144242465496063, + 0.019208844751119614, + 0.02249901182949543, + 0.008035289123654366, + 0.010726338252425194, + 0.008835353888571262, + 0.006809243466705084, + 0.004917538724839687, + 0.013149549253284931, + 0.01577330194413662, + 0.013014103285968304, + 0.007494157180190086, + 0.0053067742846906185, + 0.00815189816057682, + 0.006097566336393356, + 0.005920788738876581, + 0.007312287576496601, + 0.005835826508700848, + 0.006771095097064972, + 0.004969033412635326, + 0.003368270117789507, + 0.011135544627904892, + 0.005580900702625513, + 0.0062394700944423676, + 0.007271484937518835, + 0.006282769609242678, + 0.01303472276777029, + 0.012215453200042248, + 0.006235127337276936, + 0.004815804306417704, + 0.006875206250697374, + 0.008528406731784344, + 0.006335278507322073, + 0.00716419005766511, + 0.00822574831545353, + 0.012899954803287983, + 0.010266263037919998, + 0.006887658964842558, + 0.01883579231798649, + 0.009261714294552803, + 0.010202349163591862, + 0.007696947082877159, + 0.007762757129967213, + 0.009654898196458817, + 0.038549769669771194, + 0.014551965519785881, + 0.02363075315952301, + 0.012780826538801193, + 0.014851337298750877, + 0.019689541310071945, + 0.014819261617958546, + 0.027539515867829323, + 0.014920028857886791, + 0.01591975800693035, + 0.011438699439167976, + 0.009738790802657604, + 0.010731175541877747, + 0.010615636594593525, + 0.01706477254629135, + 0.010907397605478764, + 0.020364869385957718, + 0.006582927890121937, + 0.0161642637103796, + 0.009920187294483185, + 0.009035231545567513, + 0.013208580203354359, + 0.011226090602576733, + 0.020222656428813934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.010362748056650162 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012557249516248703 + ], + "opposite_logprob_tokens": [ + -4.387557029724121 + ], + "logprob_reward": -0.012557249516248703, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012557249516248703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09580481797456741, + 0.047080542892217636, + 0.04427153989672661, + 0.08476516604423523, + 0.07257083058357239, + 0.13655053079128265, + 0.12970298528671265, + 0.11505340039730072, + 0.15691138803958893, + 0.11728881299495697 + ], + "credit_map": [ + 0.012535887770354748, + 0.008899352513253689, + 0.010337482206523418, + 0.01424473524093628, + 0.016993889585137367, + 0.01075603999197483, + 0.015172792598605156, + 0.006864641793072224, + 0.007220633793622255, + 0.005792610347270966, + 0.003806603839620948, + 0.0040724026039242744, + 0.008752795867621899, + 0.005377078428864479, + 0.006202771794050932, + 0.0058556473813951015, + 0.005333113018423319, + 0.003718117019161582, + 0.004258779808878899, + 0.003010557033121586, + 0.00687623955309391, + 0.008368933573365211, + 0.007735594175755978, + 0.00497020548209548, + 0.008114378899335861, + 0.012750393711030483, + 0.009095200337469578, + 0.012108390219509602, + 0.012121005915105343, + 0.009047111496329308, + 0.013036355376243591, + 0.00849233753979206, + 0.004381123930215836, + 0.009741452522575855, + 0.007445414550602436, + 0.00969297531992197, + 0.008892672136425972, + 0.011142127215862274, + 0.010874579660594463, + 0.010400486178696156, + 0.013371503911912441, + 0.014749923720955849, + 0.013364650309085846, + 0.02342267706990242, + 0.01713079959154129, + 0.014060820452868938, + 0.02076774276793003, + 0.019682414829730988, + 0.011033031158149242, + 0.015478105284273624, + 0.017851397395133972, + 0.014376151375472546, + 0.024713123217225075, + 0.017455756664276123, + 0.0150191904976964, + 0.013776229694485664, + 0.02190614491701126, + 0.009147156029939651, + 0.015465056523680687, + 0.007981805130839348, + 0.009280841797590256, + 0.013462139293551445, + 0.008649314753711224, + 0.029160941019654274, + 0.020436836406588554, + 0.024989871308207512, + 0.020748043432831764, + 0.015640508383512497, + 0.011106651276350021, + 0.019560260698199272, + 0.02532881498336792, + 0.01910039782524109, + 0.02163557894527912, + 0.007289114873856306, + 0.021686891093850136, + 0.012116088531911373, + 0.009533645585179329, + 0.012343641370534897, + 0.012444217689335346, + 0.020239630714058876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.012557249516248703 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463933777064085 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.00463933777064085, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463933777064085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.076856829226017, + 0.04788719862699509, + 0.04605036601424217, + 0.05003488436341286, + 0.11962082237005234, + 0.07751602679491043, + 0.08229314535856247, + 0.22250644862651825, + 0.1327207386493683, + 0.14451351761817932 + ], + "credit_map": [ + 0.010603629983961582, + 0.008199164643883705, + 0.004916242323815823, + 0.0073408945463597775, + 0.012755168601870537, + 0.009891527704894543, + 0.0157054141163826, + 0.007444784510880709, + 0.005626808851957321, + 0.004459175746887922, + 0.004263496026396751, + 0.004295149352401495, + 0.006301280111074448, + 0.009075317531824112, + 0.009113363921642303, + 0.004752608481794596, + 0.004594967234879732, + 0.005211241543292999, + 0.0031471634283661842, + 0.005514441058039665, + 0.008558311499655247, + 0.0054117836989462376, + 0.007231245748698711, + 0.006381213199347258, + 0.005671749822795391, + 0.00958977174013853, + 0.006333860103040934, + 0.006748502142727375, + 0.006567386910319328, + 0.005349063780158758, + 0.006032953504472971, + 0.003741597291082144, + 0.0054391310550272465, + 0.010577443055808544, + 0.00955198798328638, + 0.014749843627214432, + 0.018296509981155396, + 0.013116962276399136, + 0.02630828693509102, + 0.021580655127763748, + 0.006424707360565662, + 0.009312523528933525, + 0.008148889988660812, + 0.0103555703535676, + 0.010310087352991104, + 0.007819964550435543, + 0.011736568063497543, + 0.013407720252871513, + 0.007806198205798864, + 0.007889186032116413, + 0.011688957922160625, + 0.006260243244469166, + 0.023769088089466095, + 0.008472004905343056, + 0.008848256431519985, + 0.007559214718639851, + 0.06641686707735062, + 0.00866143498569727, + 0.0370800755918026, + 0.011107726022601128, + 0.007593429647386074, + 0.022932104766368866, + 0.010397701524198055, + 0.05831712856888771, + 0.017603635787963867, + 0.022177807986736298, + 0.01771628111600876, + 0.015414191409945488, + 0.011367419734597206, + 0.015491947531700134, + 0.01730533130466938, + 0.015644118189811707, + 0.02843061089515686, + 0.010400035418570042, + 0.02938321977853775, + 0.012582161463797092, + 0.012649651616811752, + 0.017494287341833115, + 0.014337523840367794, + 0.01923603005707264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.00463933777064085 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067663207650185 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.016067663207650185, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067663207650185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05305181443691254, + 0.04491214454174042, + 0.024683251976966858, + 0.021004803478717804, + 0.0485224612057209, + 0.043951280415058136, + 0.037758514285087585, + 0.1371050924062729, + 0.15185324847698212, + 0.43715739250183105 + ], + "credit_map": [ + 0.008947492577135563, + 0.005735863000154495, + 0.005538877099752426, + 0.005536604672670364, + 0.007215636316686869, + 0.007590045686811209, + 0.00696441950276494, + 0.0055228774435818195, + 0.007701463531702757, + 0.003902578726410866, + 0.004089538939297199, + 0.003380655776709318, + 0.007888242602348328, + 0.006173944566398859, + 0.007912241853773594, + 0.003863476449623704, + 0.00293365097604692, + 0.0025647555012255907, + 0.0019919523037970066, + 0.0025002409238368273, + 0.0047131190076470375, + 0.0033031394705176353, + 0.005180250853300095, + 0.0014961421256884933, + 0.002216653898358345, + 0.0028160985093563795, + 0.002546484349295497, + 0.0025045776274055243, + 0.002944322070106864, + 0.002417801646515727, + 0.0032190412748605013, + 0.002339824102818966, + 0.0027861720882356167, + 0.0057004173286259174, + 0.00718141533434391, + 0.0054736328311264515, + 0.005562239326536655, + 0.008520519360899925, + 0.006969149690121412, + 0.006328916177153587, + 0.003768885973840952, + 0.004448525607585907, + 0.0047921412624418736, + 0.00629250705242157, + 0.006254018284380436, + 0.004849997814744711, + 0.007326839026063681, + 0.006218369584530592, + 0.004517034627497196, + 0.0038616033270955086, + 0.005765109322965145, + 0.0034098040778189898, + 0.0075303916819393635, + 0.004208694212138653, + 0.003861792152747512, + 0.004604086745530367, + 0.0385592095553875, + 0.009249383583664894, + 0.01785002090036869, + 0.00613300409168005, + 0.007694493979215622, + 0.010318751446902752, + 0.008947511203587055, + 0.03835271671414375, + 0.02408728562295437, + 0.014848054386675358, + 0.02145594358444214, + 0.021734636276960373, + 0.007898597978055477, + 0.008755873888731003, + 0.040165357291698456, + 0.012907512485980988, + 0.09531253576278687, + 0.023299310356378555, + 0.11816812306642532, + 0.0349348783493042, + 0.024596605449914932, + 0.037978652864694595, + 0.03546340763568878, + 0.0674038827419281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.016067663207650185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192949049174786 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014192949049174786, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192949049174786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.086642324924469, + 0.050707124173641205, + 0.05936487391591072, + 0.04214806854724884, + 0.06727226078510284, + 0.08982449769973755, + 0.22220636904239655, + 0.10793519020080566, + 0.0861569494009018, + 0.18774236738681793 + ], + "credit_map": [ + 0.015124901197850704, + 0.0075667924247682095, + 0.005918039940297604, + 0.008991463109850883, + 0.016558803617954254, + 0.01246671937406063, + 0.013449183665215969, + 0.006566409021615982, + 0.0080152228474617, + 0.004909139592200518, + 0.004733489826321602, + 0.003725780639797449, + 0.008537554182112217, + 0.008609479293227196, + 0.008011752739548683, + 0.004164703655987978, + 0.0047646500170230865, + 0.006835591979324818, + 0.004987369757145643, + 0.006398716010153294, + 0.008099968545138836, + 0.008944980800151825, + 0.014668806456029415, + 0.004664785228669643, + 0.004887611139565706, + 0.006870589684695005, + 0.00508768018335104, + 0.0046408637426793575, + 0.005007589235901833, + 0.004870503209531307, + 0.005529340822249651, + 0.005253887269645929, + 0.004352839197963476, + 0.01047713216394186, + 0.006086297333240509, + 0.009982936084270477, + 0.009713055565953255, + 0.007767182309180498, + 0.008829344063997269, + 0.01006347220391035, + 0.008433237671852112, + 0.01022931095212698, + 0.014328363351523876, + 0.010346276685595512, + 0.010878096334636211, + 0.010523184202611446, + 0.01566721685230732, + 0.009418806992471218, + 0.03156663104891777, + 0.021015560254454613, + 0.02780088223516941, + 0.021618850529193878, + 0.050590477883815765, + 0.025151949375867844, + 0.02476901188492775, + 0.019692979753017426, + 0.019702406600117683, + 0.009990353137254715, + 0.008103953674435616, + 0.009046032093465328, + 0.008825749158859253, + 0.015990357846021652, + 0.009594490751624107, + 0.026681844145059586, + 0.011124960146844387, + 0.01044372282922268, + 0.01131602842360735, + 0.011170340701937675, + 0.008323337882757187, + 0.007872636429965496, + 0.016506819054484367, + 0.009399088099598885, + 0.04327020421624184, + 0.009410245344042778, + 0.060538530349731445, + 0.013234077021479607, + 0.011359736323356628, + 0.013149338774383068, + 0.01609213650226593, + 0.020688101649284363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.014192949049174786 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a spectacular Final Fantasy-inspired scene, the camera, in extreme close-up, fixes itself on the amber eyes of an ancient bird, ablaze with an inner light, its reptilian pupils narrowed in concentration. The behemoth dwells deep within a magical forest shrouded in twilight. Countless particles of aetherial energy swirl in the air, slowly swirling like stardust, casting a dazzling array of light and shadow on the surrounding massive, ancient trees covered in luminescent moss. The bird's body is covered in obsidian-hard scales stretching from its neck to its back, seamlessly transitioning into iridescent feathers that shimmer with a myriad of metallic luster in the dim light. Suddenly, with a brief, dazzling flash of energy erupting from its chest, the bird unfurls its massive claws and wings with explosive force, stirring up a gust of air laced with luminous dust. At the moment when these wings, which are a combination of gorgeous feathers and tough membranes and have sharp claws extending from the joints, are fully unfolded, the focus of the camera switches as fast as lightning, from its amber eyes full of wisdom and majesty to the farthest wingtips after full folding, clearly capturing the extreme details of the feather tips trembling slightly against the blurred, dreamy forest background. The whole picture is full of epic majesty and awe-inspiring magic.", + "seed": 886, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_7/index_6050_seed_886_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04273479506373405, + "vlm_logprob": -0.04273479506373405, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.04273479506373405, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1326778382062912 + ], + "opposite_logprob_tokens": [ + -2.1326777935028076 + ], + "logprob_reward": -0.1326778382062912, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1326778382062912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10748343169689178, + 0.30900776386260986, + 0.03772575780749321, + 0.17741438746452332, + 0.11785747110843658, + 0.07703635841608047, + 0.01826505735516548, + 0.017106030136346817, + 0.08289657533168793, + 0.055207107216119766 + ], + "credit_map": [ + 0.015051167458295822, + 0.016747023910284042, + 0.008996758610010147, + 0.00691691879183054, + 0.012173479422926903, + 0.008956560865044594, + 0.02277309074997902, + 0.015868429094552994, + 0.09417585283517838, + 0.013633734546601772, + 0.020977111533284187, + 0.01149844378232956, + 0.013560258783400059, + 0.011782418936491013, + 0.09804417192935944, + 0.04533575475215912, + 0.008098078891634941, + 0.0030589185189455748, + 0.003364140633493662, + 0.004175701178610325, + 0.003937707282602787, + 0.005089424084872007, + 0.004964954685419798, + 0.005036832299083471, + 0.015313968993723392, + 0.02390681765973568, + 0.010163496248424053, + 0.01059602852910757, + 0.028591396287083626, + 0.022824477404356003, + 0.030728761106729507, + 0.035289447754621506, + 0.038588713854551315, + 0.010544040240347385, + 0.005006151739507914, + 0.005460651591420174, + 0.006401445250958204, + 0.004922048654407263, + 0.00769989937543869, + 0.039234526455402374, + 0.007871042005717754, + 0.01055547408759594, + 0.012083610519766808, + 0.010242571122944355, + 0.007512685377150774, + 0.008895637467503548, + 0.006598736625164747, + 0.013276607729494572, + 0.002362792380154133, + 0.0025466152001172304, + 0.0016677851090207696, + 0.00304181850515306, + 0.0023450760636478662, + 0.0023231978993862867, + 0.0022514299489557743, + 0.0017263415502384305, + 0.001906616729684174, + 0.0017514955252408981, + 0.003952537663280964, + 0.002202834002673626, + 0.0013229319592937827, + 0.0018901396542787552, + 0.0021389899775385857, + 0.0019404840422794223, + 0.007048504427075386, + 0.01231280155479908, + 0.010188774205744267, + 0.007799768820405006, + 0.009687206707894802, + 0.012042084708809853, + 0.01277078315615654, + 0.011046649888157845, + 0.008316087536513805, + 0.007238369435071945, + 0.007949920371174812, + 0.009130848571658134, + 0.004151939880102873, + 0.005788100883364677, + 0.005635544192045927, + 0.006996294483542442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.1326778382062912 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02638021484017372 + ], + "opposite_logprob_tokens": [ + -3.6513803005218506 + ], + "logprob_reward": -0.02638021484017372, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638021484017372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20953556895256042, + 0.060861073434352875, + 0.07353422045707703, + 0.04439958557486534, + 0.24312575161457062, + 0.04012146592140198, + 0.0535069964826107, + 0.059085700660943985, + 0.09776181727647781, + 0.11806780844926834 + ], + "credit_map": [ + 0.032844290137290955, + 0.022804759442806244, + 0.018179481849074364, + 0.013338219374418259, + 0.015461853705346584, + 0.017383066937327385, + 0.052075810730457306, + 0.03744809329509735, + 0.013242355547845364, + 0.004899116232991219, + 0.005487554706633091, + 0.006271789316087961, + 0.003823634935542941, + 0.0030408953316509724, + 0.016063036397099495, + 0.008032692596316338, + 0.011168235912919044, + 0.009119999594986439, + 0.006185114849358797, + 0.008453465066850185, + 0.010436169803142548, + 0.010985874570906162, + 0.008913768455386162, + 0.008271588943898678, + 0.005973833613097668, + 0.004112285561859608, + 0.0031008413061499596, + 0.0024057277478277683, + 0.006442594807595015, + 0.005169590935111046, + 0.00855908915400505, + 0.008635624311864376, + 0.08261815458536148, + 0.021999171003699303, + 0.010479961521923542, + 0.010576807893812656, + 0.007725968025624752, + 0.01003541424870491, + 0.0156905148178339, + 0.08399975299835205, + 0.0052342405542731285, + 0.005720564164221287, + 0.006232228595763445, + 0.00590526731684804, + 0.005049515049904585, + 0.003506455337628722, + 0.0045025241561234, + 0.003970671910792589, + 0.007596767041832209, + 0.00713750533759594, + 0.005696702748537064, + 0.009684648364782333, + 0.005730460863560438, + 0.005961596500128508, + 0.005658930167555809, + 0.006040383595973253, + 0.005245422478765249, + 0.006353462114930153, + 0.009055682457983494, + 0.0068318662233650684, + 0.009930971078574657, + 0.0076796626672148705, + 0.006445731036365032, + 0.007542903535068035, + 0.011115879751741886, + 0.012373811565339565, + 0.013483271934092045, + 0.010923813097178936, + 0.008705895394086838, + 0.016901250928640366, + 0.011681036092340946, + 0.012576852925121784, + 0.01205561775714159, + 0.013967766426503658, + 0.018791021779179573, + 0.02632167749106884, + 0.00759363267570734, + 0.01229498814791441, + 0.01203340757638216, + 0.01500970870256424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.02638021484017372 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740451790392399 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.006740451790392399, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740451790392399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10827372223138809, + 0.05576928332448006, + 0.07010424882173538, + 0.06797866523265839, + 0.0751659944653511, + 0.05968647450208664, + 0.07615972310304642, + 0.11184720695018768, + 0.27670854330062866, + 0.09830620139837265 + ], + "credit_map": [ + 0.018077418208122253, + 0.010464079678058624, + 0.011633167043328285, + 0.010009613819420338, + 0.008288344368338585, + 0.007421425543725491, + 0.027437325567007065, + 0.014942328445613384, + 0.01889183185994625, + 0.004730502609163523, + 0.005995510146021843, + 0.004657546989619732, + 0.003755537560209632, + 0.0037592542357742786, + 0.008693378418684006, + 0.005285716149955988, + 0.011028114706277847, + 0.005382210481911898, + 0.007586644496768713, + 0.005523386877030134, + 0.007810648065060377, + 0.0050960867665708065, + 0.013621342368423939, + 0.014055807143449783, + 0.007053494453430176, + 0.008530283346772194, + 0.004772293381392956, + 0.003812984563410282, + 0.01024432573467493, + 0.009919967502355576, + 0.011884751729667187, + 0.011760558933019638, + 0.020990386605262756, + 0.007120959460735321, + 0.004228966310620308, + 0.005075730849057436, + 0.006974631454795599, + 0.003908128477632999, + 0.007025625556707382, + 0.019841555505990982, + 0.006445515435189009, + 0.008419165387749672, + 0.012906399555504322, + 0.008152543567121029, + 0.004207958932965994, + 0.004423768259584904, + 0.006396957207471132, + 0.008734158240258694, + 0.012457559816539288, + 0.012994898483157158, + 0.007716417778283358, + 0.008185994811356068, + 0.006756951566785574, + 0.01245921291410923, + 0.009115535765886307, + 0.006473140325397253, + 0.008592975325882435, + 0.01044678408652544, + 0.010985751636326313, + 0.011787418276071548, + 0.022929396480321884, + 0.017078334465622902, + 0.0167918149381876, + 0.013234728015959263, + 0.020379062741994858, + 0.03878206014633179, + 0.030796952545642853, + 0.026075197383761406, + 0.029441803693771362, + 0.05501025542616844, + 0.039449382573366165, + 0.03677380830049515, + 0.008984193205833435, + 0.011541368439793587, + 0.016623040661215782, + 0.012247337028384209, + 0.010323810391128063, + 0.009037771262228489, + 0.014873744919896126, + 0.014674928039312363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.006740451790392399 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978058110922575 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.00978058110922575, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978058110922575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10265554487705231, + 0.12143677473068237, + 0.04633667320013046, + 0.04387019947171211, + 0.4013570547103882, + 0.06915086507797241, + 0.04384409636259079, + 0.06976744532585144, + 0.06524582952260971, + 0.03633547201752663 + ], + "credit_map": [ + 0.017849961295723915, + 0.010693621821701527, + 0.006882118992507458, + 0.00639076204970479, + 0.00716394791379571, + 0.008918615989387035, + 0.026695339009165764, + 0.01806117594242096, + 0.03321509808301926, + 0.005888317711651325, + 0.010784919373691082, + 0.007227953057736158, + 0.005761889275163412, + 0.004733133129775524, + 0.03619413822889328, + 0.01763131655752659, + 0.007553047966212034, + 0.004420085344463587, + 0.004148891195654869, + 0.004693725612014532, + 0.008342014625668526, + 0.004342120140790939, + 0.006614398676902056, + 0.006222385447472334, + 0.009100532159209251, + 0.0034603504464030266, + 0.0035912704188376665, + 0.0024737543426454067, + 0.005037334281951189, + 0.0042175170965492725, + 0.007645732723176479, + 0.008343706838786602, + 0.13964730501174927, + 0.03236294910311699, + 0.01782134920358658, + 0.016985047608613968, + 0.0121005242690444, + 0.016807356849312782, + 0.02597038261592388, + 0.13966216146945953, + 0.007190168835222721, + 0.004279910586774349, + 0.021555569022893906, + 0.007074917666614056, + 0.004194106440991163, + 0.008919033221900463, + 0.00397949293255806, + 0.011957669630646706, + 0.004645106848329306, + 0.00460819574072957, + 0.007403226103633642, + 0.0070625836960971355, + 0.0026735311839729548, + 0.0058629778213799, + 0.0034199038054794073, + 0.008168574422597885, + 0.008851942606270313, + 0.004527245182543993, + 0.018193213269114494, + 0.011526751331984997, + 0.005598351359367371, + 0.007654264569282532, + 0.0054888371378183365, + 0.007926835678517818, + 0.009502552449703217, + 0.005161759909242392, + 0.012407311238348484, + 0.009930211119353771, + 0.005888733081519604, + 0.00648985430598259, + 0.007404368836432695, + 0.008461038582026958, + 0.002761020790785551, + 0.003991506993770599, + 0.005302075296640396, + 0.005935627035796642, + 0.00452920189127326, + 0.0035403247456997633, + 0.005026963539421558, + 0.005248749628663063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.00978058110922575 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient bird with amber eyes and reptilian pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bird's body covered in obsidian-hard scales transitioning into iridescent feathers with metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bird perform an action of unfurling massive claws and wings with explosive force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a magical forest with massive ancient trees covered in luminescent moss and swirling aetherial particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera switch focus rapidly from the bird's eyes to the fully unfolded wingtips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038094889372587204 + ], + "opposite_logprob_tokens": [ + -3.288094997406006 + ], + "logprob_reward": -0.038094889372587204, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038094889372587204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09200767427682877, + 0.07202019542455673, + 0.10546273738145828, + 0.05876871943473816, + 0.3201683759689331, + 0.08316824585199356, + 0.09722841531038284, + 0.07981307059526443, + 0.05656542629003525, + 0.034797120839357376 + ], + "credit_map": [ + 0.016118459403514862, + 0.009300087578594685, + 0.008346036076545715, + 0.005108418874442577, + 0.007872063666582108, + 0.008241567760705948, + 0.02474239282310009, + 0.01227865181863308, + 0.024389540776610374, + 0.00523617397993803, + 0.006119617726653814, + 0.005811753217130899, + 0.003366549266502261, + 0.004826892167329788, + 0.014785585924983025, + 0.007484081666916609, + 0.011167172342538834, + 0.010002866387367249, + 0.009595783427357674, + 0.010820058174431324, + 0.017938431352376938, + 0.008632935583591461, + 0.020868875086307526, + 0.016436617821455002, + 0.007548084482550621, + 0.006163038779050112, + 0.0036595093552023172, + 0.003432450583204627, + 0.008042099885642529, + 0.006724342238157988, + 0.010376487858593464, + 0.012822708114981651, + 0.10982447862625122, + 0.027470607310533524, + 0.014643370173871517, + 0.014154898934066296, + 0.009609631262719631, + 0.013894988223910332, + 0.020623162388801575, + 0.10994724184274673, + 0.008726716972887516, + 0.010550030507147312, + 0.02201024442911148, + 0.011429919861257076, + 0.006059760693460703, + 0.008560419082641602, + 0.005513265263289213, + 0.010317889042198658, + 0.01811394840478897, + 0.013952984474599361, + 0.009823299013078213, + 0.009207740426063538, + 0.009707399643957615, + 0.012878653593361378, + 0.008801703341305256, + 0.014742684550583363, + 0.007905552163720131, + 0.008024623617529869, + 0.011045200750231743, + 0.011720660142600536, + 0.01368020474910736, + 0.009201211854815483, + 0.007178738713264465, + 0.011056875810027122, + 0.0053083752281963825, + 0.008130266331136227, + 0.005069823004305363, + 0.0052135479636490345, + 0.010135426186025143, + 0.0072654117830097675, + 0.00884906854480505, + 0.006593507714569569, + 0.0030865175649523735, + 0.0030239513143897057, + 0.00506269745528698, + 0.004624161869287491, + 0.004198151174932718, + 0.0025834920816123486, + 0.006286042742431164, + 0.005932105705142021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.038094889372587204 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_8/index_tensor([2791, 2791])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_8/index_tensor([2791, 2791])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..59cfdecaed604ecc8ff6ca393434f9511c263fa9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_8/index_tensor([2791, 2791])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, horizontal, medium shot fixes the viewer's gaze on a cluster of crimson leaves, which, in absolute silence, tremble in an eerie, unnatural manner. These leaves, a burgundy hue bordering on black, feature distinct veins and finely serrated edges. A thin, almost invisible layer of moisture coats their surface, giving them a sickly, damp sheen in the dim light. Their trembling is not caused by wind, but by rhythmic, spasmodic, tiny twitches, as if emanating from an invisible force within the plant, each tremor precise and synchronized. The background is a dark forest shrouded in thick fog. Within the blurred frame lie twisted, intertwined dark brown tree trunks and a damp, mossy ground, with virtually no other color visible. The entire scene is shrouded in the faint, cold diffuse light that penetrates the mist at dusk. The insufficient light leads to strong contrast between light and dark and deep shadows. The main colors of the picture are dull dark red and gray-green. Through this extreme visual presentation, a suffocating, ominous and depressing atmosphere is created.", + "seed": 461, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_8/index_2791_seed_461_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.020609566103667022, + "vlm_logprob": -0.020609566103667022, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.020609566103667022, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05952931568026543 + ], + "opposite_logprob_tokens": [ + -3.5595293045043945 + ], + "logprob_reward": -0.05952931568026543, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.05952931568026543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3720646798610687, + 0.12494061142206192, + 0.08742407709360123, + 0.05140497907996178, + 0.059572264552116394, + 0.014610446989536285, + 0.03246774151921272, + 0.08033426105976105, + 0.10258554667234421, + 0.0745953693985939 + ], + "credit_map": [ + 0.03992018103599548, + 0.051040131598711014, + 0.04877323657274246, + 0.04032537341117859, + 0.03875766322016716, + 0.05937068909406662, + 0.047590527683496475, + 0.04628690704703331, + 0.010870983824133873, + 0.01545885019004345, + 0.015506068244576454, + 0.011885954067111015, + 0.015294300392270088, + 0.0222332701086998, + 0.01967560686171055, + 0.014015594497323036, + 0.007259484380483627, + 0.00932292453944683, + 0.011993695050477982, + 0.009938252158463001, + 0.011434551328420639, + 0.008385846391320229, + 0.015202688984572887, + 0.013886644504964352, + 0.00486117135733366, + 0.005569410976022482, + 0.007526823785156012, + 0.005281694699078798, + 0.006277750711888075, + 0.008544592186808586, + 0.005981476977467537, + 0.007362063508480787, + 0.005785837769508362, + 0.006067898590117693, + 0.007918081246316433, + 0.006263266317546368, + 0.007268435321748257, + 0.008510012179613113, + 0.009176415391266346, + 0.008582327514886856, + 0.0019346385961398482, + 0.0017297033919021487, + 0.002108166925609112, + 0.0020195618271827698, + 0.0011908941669389606, + 0.002034819917753339, + 0.0015311138704419136, + 0.0020615507382899523, + 0.003763201180845499, + 0.0034088443499058485, + 0.00565310986712575, + 0.0031781913712620735, + 0.004387124441564083, + 0.004844783805310726, + 0.0030344193801283836, + 0.0041980682872235775, + 0.00825541652739048, + 0.009411316365003586, + 0.009892704896628857, + 0.008688794448971748, + 0.009521513245999813, + 0.009530591778457165, + 0.011999513953924179, + 0.013034411706030369, + 0.008573099970817566, + 0.010186783969402313, + 0.013127531856298447, + 0.011013146489858627, + 0.018818311393260956, + 0.009891398251056671, + 0.01632995903491974, + 0.014645327813923359, + 0.00821488443762064, + 0.007895433343946934, + 0.0104648033156991, + 0.007638977840542793, + 0.011031989008188248, + 0.0085255466401577, + 0.011485613882541656, + 0.009338131174445152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.05952931568026543 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019603813998401165 + ], + "opposite_logprob_tokens": [ + -6.251960277557373 + ], + "logprob_reward": -0.0019603813998401165, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019603813998401165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19657228887081146, + 0.08234823495149612, + 0.05790073424577713, + 0.05286005139350891, + 0.08301695436239243, + 0.13933102786540985, + 0.08108780533075333, + 0.1171976625919342, + 0.11857891082763672, + 0.07110633701086044 + ], + "credit_map": [ + 0.015585696324706078, + 0.024923866614699364, + 0.030896149575710297, + 0.02775883860886097, + 0.02362310141324997, + 0.02278488501906395, + 0.026567623019218445, + 0.024432117119431496, + 0.009216888807713985, + 0.009812216274440289, + 0.011561612598598003, + 0.009787963703274727, + 0.007303974125534296, + 0.013519873842597008, + 0.0090555464848876, + 0.012090167962014675, + 0.006209741346538067, + 0.006179134827107191, + 0.009451411664485931, + 0.0061854706145823, + 0.007366330828517675, + 0.009245743043720722, + 0.004966311622411013, + 0.008296591229736805, + 0.00506244832649827, + 0.005846366286277771, + 0.007786363363265991, + 0.0054056644439697266, + 0.008387336507439613, + 0.006516726687550545, + 0.006727756932377815, + 0.007127389777451754, + 0.007975722663104534, + 0.009584805928170681, + 0.010944695211946964, + 0.010261288844048977, + 0.013290416449308395, + 0.009108138270676136, + 0.011977306567132473, + 0.009874582290649414, + 0.013097232207655907, + 0.010755086317658424, + 0.022778887301683426, + 0.017041154205799103, + 0.017957353964447975, + 0.01525055430829525, + 0.0233947541564703, + 0.019056009128689766, + 0.008181015960872173, + 0.00783318281173706, + 0.014145449735224247, + 0.007564173080027103, + 0.010271015577018261, + 0.011329923756420612, + 0.01028601173311472, + 0.011477037332952023, + 0.011075914837419987, + 0.010754927061498165, + 0.01518977154046297, + 0.01276959478855133, + 0.020789096131920815, + 0.01030062697827816, + 0.0202866829931736, + 0.016031038016080856, + 0.013333281502127647, + 0.015162765048444271, + 0.014011149294674397, + 0.011631039902567863, + 0.022336790338158607, + 0.016164880245923996, + 0.012474100105464458, + 0.013464902527630329, + 0.006543509196490049, + 0.007337309420108795, + 0.009244418703019619, + 0.007124925497919321, + 0.011703460477292538, + 0.009306642226874828, + 0.010142945684492588, + 0.009703128598630428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.0019603813998401165 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193772338330746 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193772338330746, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193772338330746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16531653702259064, + 0.10741707682609558, + 0.09195747971534729, + 0.04907284304499626, + 0.10613828897476196, + 0.11273344606161118, + 0.03643978387117386, + 0.1581895500421524, + 0.11129594594240189, + 0.06143905594944954 + ], + "credit_map": [ + 0.018886497244238853, + 0.02147274650633335, + 0.021870383992791176, + 0.022771775722503662, + 0.017125478014349937, + 0.025209570303559303, + 0.015483851544559002, + 0.02249622531235218, + 0.011391428299248219, + 0.015022666193544865, + 0.014970390126109123, + 0.012084667570888996, + 0.009516461752355099, + 0.01926071010529995, + 0.011747311800718307, + 0.013423432596027851, + 0.008066492155194283, + 0.009364357218146324, + 0.013073151931166649, + 0.013312604278326035, + 0.010184853337705135, + 0.010595815256237984, + 0.012178471311926842, + 0.015181723050773144, + 0.0052651274017989635, + 0.004102248232811689, + 0.009111575782299042, + 0.005890601314604282, + 0.005116004031151533, + 0.006586355622857809, + 0.005508646368980408, + 0.007492277305573225, + 0.010709808208048344, + 0.011218431405723095, + 0.014854098670184612, + 0.01336013525724411, + 0.013399953953921795, + 0.011397468857467175, + 0.016191327944397926, + 0.015007047913968563, + 0.009999350644648075, + 0.010980392806231976, + 0.01688564196228981, + 0.01325139682739973, + 0.015137332491576672, + 0.010711804032325745, + 0.019015545025467873, + 0.0167519710958004, + 0.0039942157454788685, + 0.003961775451898575, + 0.005482454784214497, + 0.003567378269508481, + 0.0050427732057869434, + 0.004022848792374134, + 0.005463584791868925, + 0.004904748871922493, + 0.013461217284202576, + 0.014201194047927856, + 0.02046428993344307, + 0.01783256232738495, + 0.026146600022912025, + 0.013806373812258244, + 0.027001744136214256, + 0.02527555823326111, + 0.011944515630602837, + 0.014084245078265667, + 0.01400801446288824, + 0.012622267939150333, + 0.02154991589486599, + 0.014743276871740818, + 0.00983962882310152, + 0.012504061684012413, + 0.005738478619605303, + 0.0062102144584059715, + 0.007660995703190565, + 0.006786823272705078, + 0.010020609945058823, + 0.00721570011228323, + 0.00948912464082241, + 0.008317103609442711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.014193772338330746 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023274796083569527 + ], + "opposite_logprob_tokens": [ + -3.7732748985290527 + ], + "logprob_reward": -0.023274796083569527, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023274796083569527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24766848981380463, + 0.08980602025985718, + 0.12074461579322815, + 0.077044777572155, + 0.05719565600156784, + 0.053194668143987656, + 0.0549432709813118, + 0.09768927097320557, + 0.17236259579658508, + 0.02935064770281315 + ], + "credit_map": [ + 0.0261990949511528, + 0.029079794883728027, + 0.03563491627573967, + 0.027285421267151833, + 0.02423369139432907, + 0.03792000561952591, + 0.03303707018494606, + 0.034278471022844315, + 0.006750178523361683, + 0.007785426452755928, + 0.013494405895471573, + 0.01105098519474268, + 0.008381582796573639, + 0.014151439070701599, + 0.01425231620669365, + 0.01393967680633068, + 0.0095214257016778, + 0.011072041466832161, + 0.019240032881498337, + 0.0130622498691082, + 0.015123124234378338, + 0.013954452238976955, + 0.01934390515089035, + 0.01942737214267254, + 0.0070609780959784985, + 0.0068072122521698475, + 0.014236398972570896, + 0.007929284125566483, + 0.010223759338259697, + 0.009339703246951103, + 0.010405616834759712, + 0.011041820049285889, + 0.005796389188617468, + 0.005292265210300684, + 0.009571301750838757, + 0.0066538057290017605, + 0.005999392364174128, + 0.007813774049282074, + 0.008537700399756432, + 0.007531026378273964, + 0.004348207265138626, + 0.0044434950686991215, + 0.008511812426149845, + 0.005724248941987753, + 0.00660667521879077, + 0.006494501605629921, + 0.009033150970935822, + 0.008032571524381638, + 0.006221221759915352, + 0.005921850446611643, + 0.008208825252950191, + 0.0063939690589904785, + 0.005273572634905577, + 0.009614864364266396, + 0.005863954313099384, + 0.007445012219250202, + 0.009065094403922558, + 0.009417974390089512, + 0.012551217339932919, + 0.010810868814587593, + 0.015002269297838211, + 0.009219305589795113, + 0.01524317730218172, + 0.016379360109567642, + 0.020280124619603157, + 0.015157362446188927, + 0.02473621629178524, + 0.016787227243185043, + 0.025719232857227325, + 0.020580604672431946, + 0.022998889908194542, + 0.026102930307388306, + 0.0027069172356277704, + 0.003006510203704238, + 0.004002311732620001, + 0.0034333241637796164, + 0.003927741665393114, + 0.0034740110859274864, + 0.004303002264350653, + 0.004496828652918339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.023274796083569527 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040895650163292885 + ], + "opposite_logprob_tokens": [ + -5.50408935546875 + ], + "logprob_reward": -0.0040895650163292885, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040895650163292885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17184577882289886, + 0.19235128164291382, + 0.06467760354280472, + 0.06931890547275543, + 0.09614892303943634, + 0.06673253327608109, + 0.03890138119459152, + 0.11018988490104675, + 0.09869026392698288, + 0.09114343672990799 + ], + "credit_map": [ + 0.02571696974337101, + 0.028695452958345413, + 0.0230301171541214, + 0.013155460357666016, + 0.016384586691856384, + 0.03704472631216049, + 0.011282372288405895, + 0.016536099836230278, + 0.020447857677936554, + 0.022543298080563545, + 0.027358125895261765, + 0.020521104335784912, + 0.022355327382683754, + 0.032378651201725006, + 0.026815099641680717, + 0.01993180625140667, + 0.008556884713470936, + 0.006650430150330067, + 0.010367174632847309, + 0.00716781010851264, + 0.0069874357432127, + 0.01094120740890503, + 0.006093722768127918, + 0.007912934757769108, + 0.008339712396264076, + 0.00850978959351778, + 0.011018403805792332, + 0.007005792576819658, + 0.009326307103037834, + 0.010214941576123238, + 0.00829367060214281, + 0.00661028828471899, + 0.015688102692365646, + 0.012680799700319767, + 0.017260609194636345, + 0.00792035274207592, + 0.009936392307281494, + 0.01915210671722889, + 0.005346004385501146, + 0.008164552971720695, + 0.010485545732080936, + 0.008479321375489235, + 0.011685897596180439, + 0.004833091050386429, + 0.00828397274017334, + 0.011755520477890968, + 0.005028821993619204, + 0.006180358584970236, + 0.006036713253706694, + 0.005512142553925514, + 0.006383208092302084, + 0.0034739894326776266, + 0.0041808923706412315, + 0.005419273395091295, + 0.0032139939721673727, + 0.004681169521063566, + 0.010536359623074532, + 0.011611525900661945, + 0.013545099645853043, + 0.011286688968539238, + 0.016268540173768997, + 0.011974746361374855, + 0.01755351573228836, + 0.01741340383887291, + 0.019526692107319832, + 0.019150689244270325, + 0.014224753715097904, + 0.006716588046401739, + 0.00939246080815792, + 0.01382916048169136, + 0.006275636609643698, + 0.009574273601174355, + 0.01390758901834488, + 0.011235145851969719, + 0.01517278328537941, + 0.007374028209596872, + 0.010649210773408413, + 0.01595177873969078, + 0.007511856500059366, + 0.009341038763523102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0040895650163292885 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, horizontal, medium shot fixes the viewer's gaze on a cluster of crimson leaves, which, in absolute silence, tremble in an eerie, unnatural manner. These leaves, a burgundy hue bordering on black, feature distinct veins and finely serrated edges. A thin, almost invisible layer of moisture coats their surface, giving them a sickly, damp sheen in the dim light. Their trembling is not caused by wind, but by rhythmic, spasmodic, tiny twitches, as if emanating from an invisible force within the plant, each tremor precise and synchronized. The background is a dark forest shrouded in thick fog. Within the blurred frame lie twisted, intertwined dark brown tree trunks and a damp, mossy ground, with virtually no other color visible. The entire scene is shrouded in the faint, cold diffuse light that penetrates the mist at dusk. The insufficient light leads to strong contrast between light and dark and deep shadows. The main colors of the picture are dull dark red and gray-green. Through this extreme visual presentation, a suffocating, ominous and depressing atmosphere is created.", + "seed": 462, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_8/index_2791_seed_462_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.015508266352117061, + "vlm_logprob": -0.015508266352117061, + "vlm_yes_no_margin": 4.899999952316284, + "vlm_token_credit": -0.015508266352117061, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030039431527256966 + ], + "opposite_logprob_tokens": [ + -5.030039310455322 + ], + "logprob_reward": -0.030039431527256966, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.030039431527256966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22621828317642212, + 0.10734506696462631, + 0.05938410758972168, + 0.05243200063705444, + 0.073124960064888, + 0.03414059057831764, + 0.20462214946746826, + 0.06936684995889664, + 0.07139001786708832, + 0.10197596251964569 + ], + "credit_map": [ + 0.024112431332468987, + 0.02430669032037258, + 0.03728657588362694, + 0.023737428709864616, + 0.023681674152612686, + 0.041089411824941635, + 0.029195522889494896, + 0.022808557376265526, + 0.01694178394973278, + 0.012840204872190952, + 0.012359862215816975, + 0.011981669813394547, + 0.009683137759566307, + 0.019079918041825294, + 0.014146235771477222, + 0.010312250815331936, + 0.009325317107141018, + 0.007104776334017515, + 0.010144493542611599, + 0.005730696488171816, + 0.004011407494544983, + 0.010775988921523094, + 0.005669203121215105, + 0.006622223183512688, + 0.00902738980948925, + 0.00585938012227416, + 0.007559216581285, + 0.004297941457480192, + 0.005508731584995985, + 0.009525533765554428, + 0.005843723192811012, + 0.004810084588825703, + 0.009016312658786774, + 0.0073723988607525826, + 0.012421807274222374, + 0.006557627581059933, + 0.005706143099814653, + 0.015142237767577171, + 0.007568565662950277, + 0.009339865297079086, + 0.007719890680164099, + 0.004172122105956078, + 0.004819644149392843, + 0.0027238503098487854, + 0.0026480413507670164, + 0.006016809958964586, + 0.003192457603290677, + 0.002847772790119052, + 0.026583900675177574, + 0.017935656011104584, + 0.04302195459604263, + 0.014748050831258297, + 0.015957670286297798, + 0.0329073965549469, + 0.018376363441348076, + 0.03509114682674408, + 0.012868138961493969, + 0.009350252337753773, + 0.010049753822386265, + 0.005115321837365627, + 0.007185367401689291, + 0.011999689042568207, + 0.006687938701361418, + 0.00611038226634264, + 0.01149754598736763, + 0.007230178453028202, + 0.010041125118732452, + 0.0055137621238827705, + 0.008653409779071808, + 0.014332017861306667, + 0.007791460957378149, + 0.006330518517643213, + 0.012095595709979534, + 0.009886829182505608, + 0.01723516546189785, + 0.010472703725099564, + 0.010655641555786133, + 0.020794741809368134, + 0.01062880177050829, + 0.010206482373178005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.030039431527256966 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019558602944016457 + ], + "opposite_logprob_tokens": [ + -6.251955986022949 + ], + "logprob_reward": -0.0019558602944016457, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019558602944016457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41484570503234863, + 0.19794654846191406, + 0.0922231525182724, + 0.03801005333662033, + 0.03311214968562126, + 0.061802927404642105, + 0.02773771993815899, + 0.07720780372619629, + 0.014219517819583416, + 0.04289436712861061 + ], + "credit_map": [ + 0.03804389014840126, + 0.05107542872428894, + 0.06129438057541847, + 0.052662719041109085, + 0.05120280012488365, + 0.06161674112081528, + 0.05462666600942612, + 0.04432307928800583, + 0.0161414947360754, + 0.019924422726035118, + 0.03397265076637268, + 0.03115837275981903, + 0.026328768581151962, + 0.020933596417307854, + 0.023835595697164536, + 0.02565165050327778, + 0.014108541421592236, + 0.011023694649338722, + 0.014847655780613422, + 0.009645108133554459, + 0.008069680072367191, + 0.012795270420610905, + 0.009745574556291103, + 0.01198763120919466, + 0.005456424318253994, + 0.003538604825735092, + 0.006572562735527754, + 0.004621226340532303, + 0.003358633490279317, + 0.005583855789154768, + 0.0036216117441654205, + 0.005257137585431337, + 0.0039114560931921005, + 0.003234159667044878, + 0.005271014291793108, + 0.004273712635040283, + 0.0039301346987485886, + 0.005319045856595039, + 0.003755538957193494, + 0.003417088184505701, + 0.007353346329182386, + 0.005527141038328409, + 0.011418304406106472, + 0.006519805174320936, + 0.006566528230905533, + 0.008187624625861645, + 0.006948601920157671, + 0.009281576611101627, + 0.0036359180230647326, + 0.0031198097858577967, + 0.004201427102088928, + 0.002689163200557232, + 0.0023317926097661257, + 0.005330374930053949, + 0.00327887199819088, + 0.0031503629870712757, + 0.011691690422594547, + 0.007717273663729429, + 0.012405837886035442, + 0.006707215681672096, + 0.007175900042057037, + 0.013925604522228241, + 0.0075561050325632095, + 0.010028179734945297, + 0.001507912646047771, + 0.0015333275077864528, + 0.0023776660673320293, + 0.0019694340880960226, + 0.0013076717732474208, + 0.0022599799558520317, + 0.0015632156282663345, + 0.0017003106186166406, + 0.005780559498816729, + 0.003971883561462164, + 0.007745649199932814, + 0.003958001732826233, + 0.003761982312425971, + 0.008475255221128464, + 0.0039750575087964535, + 0.005225977394729853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0019558602944016457 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180983141064644 + ], + "opposite_logprob_tokens": [ + -4.018180847167969 + ], + "logprob_reward": -0.018180983141064644, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018180983141064644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2833671271800995, + 0.19540439546108246, + 0.08284878730773926, + 0.0496404767036438, + 0.041921935975551605, + 0.1136379987001419, + 0.041899289935827255, + 0.13905951380729675, + 0.02388702519237995, + 0.028333550319075584 + ], + "credit_map": [ + 0.046750009059906006, + 0.03181535750627518, + 0.045882876962423325, + 0.029865460470318794, + 0.021696534007787704, + 0.04494791105389595, + 0.03666791319847107, + 0.025741005316376686, + 0.01965656876564026, + 0.02284148894250393, + 0.02830473706126213, + 0.025104304775595665, + 0.019282637163996696, + 0.0325506366789341, + 0.029841015115380287, + 0.017822973430156708, + 0.009838208556175232, + 0.009336973540484905, + 0.013134777545928955, + 0.009630064480006695, + 0.007115532178431749, + 0.014990845695137978, + 0.010986345820128918, + 0.007816025987267494, + 0.005537348333746195, + 0.004272174555808306, + 0.009785405360162258, + 0.003979707136750221, + 0.003737913677468896, + 0.007557231467217207, + 0.007035399787127972, + 0.0077352882362902164, + 0.006598456297069788, + 0.0045151435770094395, + 0.007284810766577721, + 0.005358495283871889, + 0.0029476420022547245, + 0.005875209346413612, + 0.004056247416883707, + 0.0052859243005514145, + 0.012065940536558628, + 0.008148076944053173, + 0.02384325861930847, + 0.012874983251094818, + 0.012020793743431568, + 0.015299778431653976, + 0.012445373460650444, + 0.01693977229297161, + 0.004693574737757444, + 0.004089665599167347, + 0.00679862592369318, + 0.004880312364548445, + 0.0031283327843993902, + 0.007670228369534016, + 0.0051656728610396385, + 0.005472870543599129, + 0.024621935561299324, + 0.017062395811080933, + 0.023361530154943466, + 0.01025179959833622, + 0.011711894534528255, + 0.023528752848505974, + 0.01150410808622837, + 0.017017068341374397, + 0.002609728602692485, + 0.0028203576803207397, + 0.0044122738763689995, + 0.0019053998403251171, + 0.0020863167010247707, + 0.00369439460337162, + 0.002995649352669716, + 0.0033628977835178375, + 0.005031950306147337, + 0.003273631678894162, + 0.0041388897225260735, + 0.0023772502318024635, + 0.0032487825956195593, + 0.0046636490151286125, + 0.0031251085456460714, + 0.002474283566698432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.018180983141064644 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327572926878929 + ], + "opposite_logprob_tokens": [ + -3.77327561378479 + ], + "logprob_reward": -0.02327572926878929, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327572926878929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2189941704273224, + 0.09800679981708527, + 0.08232535421848297, + 0.1021040603518486, + 0.12863565981388092, + 0.10844424366950989, + 0.1412142962217331, + 0.03201939910650253, + 0.042544130235910416, + 0.045711860060691833 + ], + "credit_map": [ + 0.026865940541028976, + 0.022506264969706535, + 0.032353200018405914, + 0.022762682288885117, + 0.01978861540555954, + 0.04194443300366402, + 0.03012906387448311, + 0.02264396660029888, + 0.015635214745998383, + 0.010379861108958721, + 0.016418512910604477, + 0.009177537634968758, + 0.00795848947018385, + 0.014663655310869217, + 0.010893761180341244, + 0.012879764661192894, + 0.014154034666717052, + 0.009367107413709164, + 0.015885375440120697, + 0.00812250841408968, + 0.0062365238554775715, + 0.011443180963397026, + 0.0075225383043289185, + 0.00959408562630415, + 0.017820626497268677, + 0.012479710392653942, + 0.015427855774760246, + 0.008127154782414436, + 0.009130794554948807, + 0.01677415519952774, + 0.011806802824139595, + 0.010536955669522285, + 0.02005062624812126, + 0.01700109802186489, + 0.017412694171071053, + 0.01406923495233059, + 0.012213082984089851, + 0.018600251525640488, + 0.01638685166835785, + 0.012901817448437214, + 0.015369539149105549, + 0.011090529151260853, + 0.01980576105415821, + 0.011809330433607101, + 0.008990608155727386, + 0.017605120316147804, + 0.012631143443286419, + 0.011142215691506863, + 0.02037573792040348, + 0.012705663219094276, + 0.02896951697766781, + 0.011319366283714771, + 0.01109110563993454, + 0.022280260920524597, + 0.011741507798433304, + 0.02273114211857319, + 0.003627455560490489, + 0.0024042984005063772, + 0.005643465090543032, + 0.0031136481557041407, + 0.0031285628210753202, + 0.006364310625940561, + 0.004274914972484112, + 0.00346274278126657, + 0.006083356682211161, + 0.003947554621845484, + 0.00642468873411417, + 0.003576046321541071, + 0.004240239970386028, + 0.008455202914774418, + 0.004957020748406649, + 0.0048600235022604465, + 0.009584064595401287, + 0.005622451659291983, + 0.006327868439257145, + 0.0036408500745892525, + 0.004930224735289812, + 0.007899108342826366, + 0.003855392336845398, + 0.003851900342851877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.02327572926878929 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040893275290727615 + ], + "opposite_logprob_tokens": [ + -5.50408935546875 + ], + "logprob_reward": -0.0040893275290727615, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040893275290727615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3730737268924713, + 0.11321982741355896, + 0.08186584711074829, + 0.06528335809707642, + 0.04624181613326073, + 0.04129961505532265, + 0.14521582424640656, + 0.05573003366589546, + 0.034180860966444016, + 0.04388910159468651 + ], + "credit_map": [ + 0.045458413660526276, + 0.04394855722784996, + 0.05955445021390915, + 0.03869423270225525, + 0.04109480604529381, + 0.06370248645544052, + 0.04468284547328949, + 0.03593794256448746, + 0.015190375037491322, + 0.014057902619242668, + 0.017529865726828575, + 0.010835207998752594, + 0.012095769867300987, + 0.019184745848178864, + 0.013386954553425312, + 0.010938998311758041, + 0.009871339425444603, + 0.009528581984341145, + 0.014203323051333427, + 0.007790918927639723, + 0.00868693646043539, + 0.015496240928769112, + 0.007871176116168499, + 0.008417325094342232, + 0.008703126572072506, + 0.007762291934341192, + 0.011507510207593441, + 0.0053006806410849094, + 0.006245541851967573, + 0.013111021369695663, + 0.005279615521430969, + 0.007373571861535311, + 0.007346415892243385, + 0.003838198259472847, + 0.008061720058321953, + 0.004827213939279318, + 0.0041950359009206295, + 0.008205235004425049, + 0.0047191171906888485, + 0.005048880819231272, + 0.006547501776367426, + 0.005405060481280088, + 0.006032847333699465, + 0.003100935136899352, + 0.003710488323122263, + 0.0077125937677919865, + 0.004818987566977739, + 0.0039711990393698215, + 0.02084146812558174, + 0.013755151070654392, + 0.027967719361186028, + 0.012184470891952515, + 0.011962362565100193, + 0.024656351655721664, + 0.011983704753220081, + 0.02186458930373192, + 0.00997136626392603, + 0.006596993654966354, + 0.01015215553343296, + 0.003685303498059511, + 0.004562775604426861, + 0.009492401964962482, + 0.004580076318234205, + 0.0066889626905322075, + 0.004339687060564756, + 0.004407915752381086, + 0.005339356604963541, + 0.0031335821840912104, + 0.0029294760897755623, + 0.006144271232187748, + 0.004115596413612366, + 0.0037709742318838835, + 0.00534439692273736, + 0.0041414713487029076, + 0.007719912566244602, + 0.004017104394733906, + 0.0036705099046230316, + 0.009909656830132008, + 0.004290799610316753, + 0.00479525001719594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0040893275290727615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, horizontal, medium shot fixes the viewer's gaze on a cluster of crimson leaves, which, in absolute silence, tremble in an eerie, unnatural manner. These leaves, a burgundy hue bordering on black, feature distinct veins and finely serrated edges. A thin, almost invisible layer of moisture coats their surface, giving them a sickly, damp sheen in the dim light. Their trembling is not caused by wind, but by rhythmic, spasmodic, tiny twitches, as if emanating from an invisible force within the plant, each tremor precise and synchronized. The background is a dark forest shrouded in thick fog. Within the blurred frame lie twisted, intertwined dark brown tree trunks and a damp, mossy ground, with virtually no other color visible. The entire scene is shrouded in the faint, cold diffuse light that penetrates the mist at dusk. The insufficient light leads to strong contrast between light and dark and deep shadows. The main colors of the picture are dull dark red and gray-green. Through this extreme visual presentation, a suffocating, ominous and depressing atmosphere is created.", + "seed": 463, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_8/index_2791_seed_463_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.021759658632799984, + "vlm_logprob": -0.021759658632799984, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.021759658632799984, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0610961876809597 + ], + "opposite_logprob_tokens": [ + -4.06109619140625 + ], + "logprob_reward": -0.0610961876809597, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0610961876809597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3156171143054962, + 0.1302255243062973, + 0.05172497779130936, + 0.08191803842782974, + 0.07498947530984879, + 0.05984038487076759, + 0.06377187371253967, + 0.046238817274570465, + 0.08866012841463089, + 0.08701367676258087 + ], + "credit_map": [ + 0.0350520946085453, + 0.04422714561223984, + 0.046347856521606445, + 0.03200402110815048, + 0.031699199229478836, + 0.04320228472352028, + 0.04789368808269501, + 0.03519080951809883, + 0.014924479648470879, + 0.021941639482975006, + 0.017588719725608826, + 0.011913572438061237, + 0.013094381429255009, + 0.016978729516267776, + 0.018633808940649033, + 0.01515019591897726, + 0.005597799550741911, + 0.006741122808307409, + 0.006937854457646608, + 0.005819451995193958, + 0.004466200713068247, + 0.006806842051446438, + 0.00867325160652399, + 0.006682455539703369, + 0.0068265520967543125, + 0.00942825898528099, + 0.01387482788413763, + 0.00974659901112318, + 0.009159217588603497, + 0.013801073655486107, + 0.010518602095544338, + 0.008562908507883549, + 0.007972564548254013, + 0.008266248740255833, + 0.009900725446641445, + 0.011338568292558193, + 0.005987050477415323, + 0.007102129049599171, + 0.014749091118574142, + 0.009673095308244228, + 0.005421535577625036, + 0.007240856997668743, + 0.009605358354747295, + 0.008051097393035889, + 0.006080848164856434, + 0.008775582537055016, + 0.00833827257156372, + 0.006326830945909023, + 0.005691305734217167, + 0.008093285374343395, + 0.008478773757815361, + 0.009525730274617672, + 0.006500435993075371, + 0.009476191364228725, + 0.008868718519806862, + 0.0071374340914189816, + 0.0054501318372786045, + 0.006734290160238743, + 0.007848774082958698, + 0.004229016602039337, + 0.004181500058621168, + 0.007846882566809654, + 0.005880468524992466, + 0.0040677557699382305, + 0.01103372685611248, + 0.016293292865157127, + 0.013372818939387798, + 0.008919315412640572, + 0.009451709687709808, + 0.010996578261256218, + 0.01007055677473545, + 0.008522136136889458, + 0.01104215532541275, + 0.011783896945416927, + 0.012047729454934597, + 0.00950625166296959, + 0.007549035828560591, + 0.012454207986593246, + 0.011081104166805744, + 0.011549295857548714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0610961876809597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002213886706158519 + ], + "opposite_logprob_tokens": [ + -6.127213954925537 + ], + "logprob_reward": -0.002213886706158519, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002213886706158519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3407590687274933, + 0.10490879416465759, + 0.09465299546718597, + 0.05076611042022705, + 0.07225015759468079, + 0.0428285226225853, + 0.08539162576198578, + 0.07270343601703644, + 0.09141534566879272, + 0.044323988258838654 + ], + "credit_map": [ + 0.04298968240618706, + 0.04882704094052315, + 0.05127551406621933, + 0.031119782477617264, + 0.03722481057047844, + 0.05386665090918541, + 0.04534487798810005, + 0.03011070005595684, + 0.011705624870955944, + 0.019279880449175835, + 0.015570953488349915, + 0.011101869866251945, + 0.009481653571128845, + 0.014772295951843262, + 0.014939716085791588, + 0.008056786842644215, + 0.011382753029465675, + 0.018802566453814507, + 0.01386546902358532, + 0.007520574145019054, + 0.013918510638177395, + 0.01357468031346798, + 0.008164719678461552, + 0.007423715200275183, + 0.004022955894470215, + 0.004676938988268375, + 0.00894069205969572, + 0.00713353930041194, + 0.003592315362766385, + 0.00750978197902441, + 0.008529956452548504, + 0.006359929218888283, + 0.008698666468262672, + 0.010108987800776958, + 0.012387534603476524, + 0.0061525218188762665, + 0.00933594349771738, + 0.011509899981319904, + 0.0064241765066981316, + 0.0076324213296175, + 0.003698417218402028, + 0.004264384973794222, + 0.007801269646733999, + 0.004175953101366758, + 0.004435962997376919, + 0.006501455791294575, + 0.006889308802783489, + 0.005061763804405928, + 0.00678609823808074, + 0.007543143350630999, + 0.012388255447149277, + 0.013506530784070492, + 0.00812146719545126, + 0.010187175124883652, + 0.014983740635216236, + 0.011875202879309654, + 0.0053253136575222015, + 0.006608373485505581, + 0.013699755072593689, + 0.010479100979864597, + 0.0063663627952337265, + 0.01213111449033022, + 0.011267353780567646, + 0.006826051045209169, + 0.007329708430916071, + 0.010862265713512897, + 0.01346137560904026, + 0.012993506155908108, + 0.008013973012566566, + 0.012891199439764023, + 0.016044406220316887, + 0.009818903170526028, + 0.00653283903375268, + 0.007298745680600405, + 0.007217960897833109, + 0.0033847505692392588, + 0.004659080412238836, + 0.006714212242513895, + 0.0033397378865629435, + 0.005176657345145941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.002213886706158519 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073444969952106 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073444969952106, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073444969952106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39637279510498047, + 0.11616230756044388, + 0.12932513654232025, + 0.06864957511425018, + 0.04339241608977318, + 0.03216274082660675, + 0.0893988311290741, + 0.03695312887430191, + 0.035652801394462585, + 0.051930274814367294 + ], + "credit_map": [ + 0.041451580822467804, + 0.06307803839445114, + 0.050871722400188446, + 0.05381258577108383, + 0.03513273969292641, + 0.048064686357975006, + 0.05519458279013634, + 0.0487668476998806, + 0.009902040474116802, + 0.0118567468598485, + 0.01914069429039955, + 0.015167236328125, + 0.009446053765714169, + 0.014762058854103088, + 0.020555982366204262, + 0.015331502072513103, + 0.012099165469408035, + 0.016504906117916107, + 0.022337177768349648, + 0.011518234387040138, + 0.0132367555052042, + 0.01923431269824505, + 0.01685432344675064, + 0.01754026487469673, + 0.00608682120218873, + 0.008421984501183033, + 0.013211419805884361, + 0.007282789796590805, + 0.006530317477881908, + 0.011211851611733437, + 0.008582447655498981, + 0.007321943063288927, + 0.006030728574842215, + 0.006390672642737627, + 0.0053024329245090485, + 0.005721621215343475, + 0.004202998708933592, + 0.005527954548597336, + 0.004638968501240015, + 0.005577040370553732, + 0.002786553231999278, + 0.0032763169147074223, + 0.004864195827394724, + 0.004177148453891277, + 0.0024131499230861664, + 0.004484702367335558, + 0.004793877713382244, + 0.005366796627640724, + 0.009461382403969765, + 0.010844145901501179, + 0.019816165789961815, + 0.007828199304640293, + 0.0076149627566337585, + 0.017030546441674232, + 0.007776141632348299, + 0.00902729295194149, + 0.0037327518220990896, + 0.006170014850795269, + 0.006832215003669262, + 0.004178134258836508, + 0.003815541509538889, + 0.005332634784281254, + 0.0037802159786224365, + 0.003111621830612421, + 0.0038776809815317392, + 0.004995248280465603, + 0.005439119413495064, + 0.0034835701808333397, + 0.004153934307396412, + 0.005315830931067467, + 0.00422911811619997, + 0.004158301278948784, + 0.005926112644374371, + 0.009198212996125221, + 0.008021124638617039, + 0.005044164136052132, + 0.004901161417365074, + 0.007132402155548334, + 0.0057415226474404335, + 0.005965576507151127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.011073444969952106 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029783742502331734 + ], + "opposite_logprob_tokens": [ + -3.5297837257385254 + ], + "logprob_reward": -0.029783742502331734, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029783742502331734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3357008397579193, + 0.09528602659702301, + 0.08410410583019257, + 0.05644428730010986, + 0.07503657042980194, + 0.11498218774795532, + 0.07661691308021545, + 0.035616256296634674, + 0.06842169910669327, + 0.05779113993048668 + ], + "credit_map": [ + 0.026859253644943237, + 0.035974256694316864, + 0.04450124874711037, + 0.04876062646508217, + 0.03819787874817848, + 0.03524171933531761, + 0.06506000459194183, + 0.041105855256319046, + 0.006506578531116247, + 0.008876350708305836, + 0.01758631318807602, + 0.01191426906734705, + 0.006010184530168772, + 0.017207331955432892, + 0.016591746360063553, + 0.010593253187835217, + 0.006549326702952385, + 0.008428005501627922, + 0.012848574668169022, + 0.012964757159352303, + 0.006445199251174927, + 0.012003406882286072, + 0.014244509860873222, + 0.010620322078466415, + 0.00471804104745388, + 0.005221617873758078, + 0.009340331889688969, + 0.0074452287517488, + 0.004018194507807493, + 0.00983426533639431, + 0.008448599837720394, + 0.007418010383844376, + 0.008696612901985645, + 0.009176676161587238, + 0.01088214572519064, + 0.007246442139148712, + 0.006619470659643412, + 0.010724237188696861, + 0.012195227667689323, + 0.0094957510009408, + 0.011381884105503559, + 0.018038176000118256, + 0.015292058698832989, + 0.012607303448021412, + 0.015193747356534004, + 0.014397939667105675, + 0.015318694524466991, + 0.012752377428114414, + 0.007130255922675133, + 0.007591943256556988, + 0.017660878598690033, + 0.006418060511350632, + 0.0076628178358078, + 0.01499594934284687, + 0.008023102767765522, + 0.007133902050554752, + 0.0036166163627058268, + 0.0036685820668935776, + 0.007271420210599899, + 0.0036044050939381123, + 0.004367480054497719, + 0.006012755446135998, + 0.004362299107015133, + 0.0027127002831548452, + 0.007349363528192043, + 0.0075170486234128475, + 0.009757822379469872, + 0.009394621476531029, + 0.0064931390807032585, + 0.009243563748896122, + 0.010906258597970009, + 0.0077598849311470985, + 0.007217250298708677, + 0.007973076775670052, + 0.007850371301174164, + 0.005431678611785173, + 0.007964502088725567, + 0.007396480068564415, + 0.007726829964667559, + 0.00623095128685236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.029783742502331734 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004631031304597855 + ], + "opposite_logprob_tokens": [ + -5.379631042480469 + ], + "logprob_reward": -0.004631031304597855, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004631031304597855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3692483603954315, + 0.13241265714168549, + 0.0981837585568428, + 0.07254573702812195, + 0.0841604471206665, + 0.08920265734195709, + 0.03439822420477867, + 0.04814344644546509, + 0.02516072243452072, + 0.046543918550014496 + ], + "credit_map": [ + 0.04745584726333618, + 0.06845483183860779, + 0.05997380241751671, + 0.02489754557609558, + 0.0503380224108696, + 0.05699166655540466, + 0.027175720781087875, + 0.03396089747548103, + 0.016728170216083527, + 0.027790173888206482, + 0.017100155353546143, + 0.0090916957706213, + 0.019139748066663742, + 0.017250262200832367, + 0.014270729385316372, + 0.011041722260415554, + 0.013050679117441177, + 0.02164902165532112, + 0.014107526279985905, + 0.005699026398360729, + 0.015433489345014095, + 0.014573209919035435, + 0.007195398677140474, + 0.006475402507930994, + 0.008341854438185692, + 0.011251339688897133, + 0.014704554341733456, + 0.004160613752901554, + 0.008293799124658108, + 0.013821075670421124, + 0.006036043632775545, + 0.005936457309871912, + 0.011243713088333607, + 0.016058921813964844, + 0.01465405523777008, + 0.004602153319865465, + 0.010218765586614609, + 0.015842251479625702, + 0.005069275852292776, + 0.006471315398812294, + 0.011475646868348122, + 0.017832349985837936, + 0.014051000587642193, + 0.005796170327812433, + 0.010127315297722816, + 0.014464161358773708, + 0.0068750823847949505, + 0.008580928668379784, + 0.004115121904760599, + 0.005343146622180939, + 0.004404093604534864, + 0.003912338521331549, + 0.004158373922109604, + 0.004180579446256161, + 0.004370571114122868, + 0.003913999535143375, + 0.005473524797707796, + 0.007189023308455944, + 0.009979546070098877, + 0.0034557764884084463, + 0.004821867682039738, + 0.009111247025430202, + 0.004243500996381044, + 0.0038689603097736835, + 0.0032479960937052965, + 0.003856124822050333, + 0.0037301054690033197, + 0.002637282945215702, + 0.0025968244299292564, + 0.0029500818345695734, + 0.0030311846639961004, + 0.0031111217103898525, + 0.006739555858075619, + 0.008452929556369781, + 0.006854791659861803, + 0.0027655658777803183, + 0.004884274676442146, + 0.00778313959017396, + 0.003809804329648614, + 0.005253852345049381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.004631031304597855 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, horizontal, medium shot fixes the viewer's gaze on a cluster of crimson leaves, which, in absolute silence, tremble in an eerie, unnatural manner. These leaves, a burgundy hue bordering on black, feature distinct veins and finely serrated edges. A thin, almost invisible layer of moisture coats their surface, giving them a sickly, damp sheen in the dim light. Their trembling is not caused by wind, but by rhythmic, spasmodic, tiny twitches, as if emanating from an invisible force within the plant, each tremor precise and synchronized. The background is a dark forest shrouded in thick fog. Within the blurred frame lie twisted, intertwined dark brown tree trunks and a damp, mossy ground, with virtually no other color visible. The entire scene is shrouded in the faint, cold diffuse light that penetrates the mist at dusk. The insufficient light leads to strong contrast between light and dark and deep shadows. The main colors of the picture are dull dark red and gray-green. Through this extreme visual presentation, a suffocating, ominous and depressing atmosphere is created.", + "seed": 464, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_8/index_2791_seed_464_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03161564553156495, + "vlm_logprob": -0.03161564553156495, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.03161564553156495, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04794019088149071 + ], + "opposite_logprob_tokens": [ + -4.047940254211426 + ], + "logprob_reward": -0.04794019088149071, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.04794019088149071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27299782633781433, + 0.07480157166719437, + 0.0662960559129715, + 0.07878321409225464, + 0.0678323283791542, + 0.10682357102632523, + 0.08982841670513153, + 0.05639979988336563, + 0.09515761584043503, + 0.09107960760593414 + ], + "credit_map": [ + 0.02633490413427353, + 0.050219107419252396, + 0.03868071734905243, + 0.03614015877246857, + 0.0356186218559742, + 0.017124800011515617, + 0.04461357742547989, + 0.02426593378186226, + 0.010664436966180801, + 0.01661057583987713, + 0.009415688924491405, + 0.009024707600474358, + 0.0070428187027573586, + 0.005914592184126377, + 0.008168523199856281, + 0.007960219867527485, + 0.0058180722407996655, + 0.011470370925962925, + 0.008894562721252441, + 0.00859304703772068, + 0.005959385074675083, + 0.006097281817346811, + 0.012447667308151722, + 0.00701566506177187, + 0.011369573883712292, + 0.01859307661652565, + 0.01112078782171011, + 0.009137422777712345, + 0.007880783639848232, + 0.005211163312196732, + 0.009028556756675243, + 0.00644184835255146, + 0.008869056589901447, + 0.016495823860168457, + 0.008670582436025143, + 0.008243310265243053, + 0.006955012213438749, + 0.006282819900661707, + 0.007135277148336172, + 0.0051804413087666035, + 0.015529017895460129, + 0.02005225606262684, + 0.015459878370165825, + 0.010951658710837364, + 0.01113902684301138, + 0.010548342950642109, + 0.013226252980530262, + 0.009917136281728745, + 0.007672891486436129, + 0.013846716843545437, + 0.012719814665615559, + 0.011250179260969162, + 0.01293861959129572, + 0.009220416657626629, + 0.012397519312798977, + 0.009782257489860058, + 0.0045797196216881275, + 0.007386607583612204, + 0.0074491966515779495, + 0.008241364732384682, + 0.00792253203690052, + 0.006474211346358061, + 0.007462190464138985, + 0.006883970927447081, + 0.010903698392212391, + 0.023062730208039284, + 0.01372649148106575, + 0.01149796973913908, + 0.010683747939765453, + 0.007341606076806784, + 0.010361987166106701, + 0.007579376921057701, + 0.007997987791895866, + 0.016364164650440216, + 0.013551894575357437, + 0.012999343685805798, + 0.011183477938175201, + 0.0066113658249378204, + 0.014174126088619232, + 0.008197233080863953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.04794019088149071 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004651796538382769 + ], + "opposite_logprob_tokens": [ + -5.37965202331543 + ], + "logprob_reward": -0.004651796538382769, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004651796538382769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29307180643081665, + 0.10322494804859161, + 0.12726467847824097, + 0.042373333126306534, + 0.05330829322338104, + 0.03477299213409424, + 0.04953225329518318, + 0.018597785383462906, + 0.08120770752429962, + 0.19664624333381653 + ], + "credit_map": [ + 0.02821093052625656, + 0.047620851546525955, + 0.03911926597356796, + 0.04130276292562485, + 0.03244687244296074, + 0.030113497748970985, + 0.043041881173849106, + 0.03121575526893139, + 0.006608111783862114, + 0.010840062983334064, + 0.014116635546088219, + 0.015640651807188988, + 0.011438770219683647, + 0.012712626717984676, + 0.016550157219171524, + 0.015317934565246105, + 0.01227828860282898, + 0.020464478060603142, + 0.018109813332557678, + 0.02073495090007782, + 0.009941613301634789, + 0.010415258817374706, + 0.015393982641398907, + 0.0199262872338295, + 0.0043922001495957375, + 0.0061469110660254955, + 0.005724463611841202, + 0.006473677698522806, + 0.004528597462922335, + 0.004164304584264755, + 0.006191482301801443, + 0.004751695320010185, + 0.003936403430998325, + 0.004299080464988947, + 0.00774132926017046, + 0.008752980269491673, + 0.005283331032842398, + 0.008949032984673977, + 0.007444826886057854, + 0.006901310756802559, + 0.0038370448164641857, + 0.003250433597713709, + 0.005633978173136711, + 0.005348776932805777, + 0.004290436860173941, + 0.0025287128519266844, + 0.005494383163750172, + 0.004389227367937565, + 0.00503277825191617, + 0.010126747190952301, + 0.006599873770028353, + 0.0065088821575045586, + 0.006761029362678528, + 0.003462765598669648, + 0.006063880864530802, + 0.00497629726305604, + 0.002341257641091943, + 0.002546024275943637, + 0.0026734895072877407, + 0.0025762473233044147, + 0.001989029347896576, + 0.0017259798478335142, + 0.0026592393405735493, + 0.0020865180995315313, + 0.006969374138861895, + 0.011535005643963814, + 0.011526519432663918, + 0.009579303674399853, + 0.009114944376051426, + 0.008855088613927364, + 0.015406779013574123, + 0.008220687508583069, + 0.01726679317653179, + 0.03412369638681412, + 0.029868939891457558, + 0.02139667607843876, + 0.024996910244226456, + 0.018452370539307594, + 0.033446915447711945, + 0.017093924805521965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.004651796538382769 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897523045539856 + ], + "opposite_logprob_tokens": [ + -2.578975200653076 + ], + "logprob_reward": -0.07897523045539856, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897523045539856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22775088250637054, + 0.12495382875204086, + 0.05906211957335472, + 0.06721889227628708, + 0.07663228362798691, + 0.08206261694431305, + 0.04615803807973862, + 0.03670055419206619, + 0.07346110045909882, + 0.20599967241287231 + ], + "credit_map": [ + 0.023926127701997757, + 0.04103017598390579, + 0.031166967004537582, + 0.03748837485909462, + 0.02171938866376877, + 0.01928020641207695, + 0.030381200835108757, + 0.02275843918323517, + 0.012138278223574162, + 0.013122697360813618, + 0.01570631004869938, + 0.02290135808289051, + 0.013275917619466782, + 0.012514302507042885, + 0.018039124086499214, + 0.017255838960409164, + 0.005666608456522226, + 0.008741110563278198, + 0.006884853355586529, + 0.011464922688901424, + 0.005821780301630497, + 0.005940740462392569, + 0.008251640945672989, + 0.0062904637306928635, + 0.008337222039699554, + 0.015221298672258854, + 0.008155492134392262, + 0.0070174639113247395, + 0.007614756003022194, + 0.006901589687913656, + 0.006368959788233042, + 0.007602109108120203, + 0.00682252598926425, + 0.009989739395678043, + 0.010917716659605503, + 0.012649751268327236, + 0.007380358409136534, + 0.007708963938057423, + 0.0110347094014287, + 0.010128519497811794, + 0.005752751603722572, + 0.011920505203306675, + 0.01079818420112133, + 0.01263023167848587, + 0.010647113434970379, + 0.012316086329519749, + 0.010690760798752308, + 0.007306985091418028, + 0.003829657332971692, + 0.006966772023588419, + 0.00605541979894042, + 0.006406127940863371, + 0.006456355098634958, + 0.005011911038309336, + 0.006688637193292379, + 0.004743155557662249, + 0.002658032812178135, + 0.005180906038731337, + 0.005450145341455936, + 0.0059343744069337845, + 0.004002027679234743, + 0.0036938809789717197, + 0.006081182509660721, + 0.0037000058218836784, + 0.007147521246224642, + 0.010391630232334137, + 0.010558315552771091, + 0.008469563908874989, + 0.008095578290522099, + 0.0071898046880960464, + 0.013765119947493076, + 0.007843568921089172, + 0.018607696518301964, + 0.0401519276201725, + 0.03229062631726265, + 0.023111553862690926, + 0.022758349776268005, + 0.01907486654818058, + 0.03321944549679756, + 0.016785191372036934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.07897523045539856 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020572030916810036 + ], + "opposite_logprob_tokens": [ + -3.8955719470977783 + ], + "logprob_reward": -0.020572030916810036, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020572030916810036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23650997877120972, + 0.10054228454828262, + 0.12211507558822632, + 0.0652940422296524, + 0.08118322491645813, + 0.10496719926595688, + 0.07327727228403091, + 0.061640866100788116, + 0.04682119935750961, + 0.10764877498149872 + ], + "credit_map": [ + 0.020512128248810768, + 0.03323955088853836, + 0.03381924703717232, + 0.0366198793053627, + 0.025082748383283615, + 0.020878266543149948, + 0.04071587324142456, + 0.02564229816198349, + 0.010463320650160313, + 0.017791057005524635, + 0.013631383888423443, + 0.01604292541742325, + 0.00930334534496069, + 0.008659002371132374, + 0.01176995038986206, + 0.012881305068731308, + 0.01379574928432703, + 0.020755723118782043, + 0.015554511919617653, + 0.01898007094860077, + 0.009667567908763885, + 0.009641963988542557, + 0.01719830371439457, + 0.016521189361810684, + 0.00874881912022829, + 0.01632191240787506, + 0.008371632546186447, + 0.00904041063040495, + 0.005866557359695435, + 0.004398528952151537, + 0.0060080792754888535, + 0.006538106594234705, + 0.008705190382897854, + 0.014568804763257504, + 0.012712209485471249, + 0.008719878271222115, + 0.0075716013088822365, + 0.0071362522430717945, + 0.01135185919702053, + 0.010417438112199306, + 0.00990628357976675, + 0.01312268152832985, + 0.017104629427194595, + 0.014300561510026455, + 0.008784780278801918, + 0.011009966023266315, + 0.01730666682124138, + 0.013431647792458534, + 0.007177991326898336, + 0.012973185628652573, + 0.00949726440012455, + 0.008067547343671322, + 0.010420020669698715, + 0.007582418620586395, + 0.008515133522450924, + 0.009043713100254536, + 0.005194027442485094, + 0.012984300032258034, + 0.009336898103356361, + 0.0088194590061903, + 0.008324622176587582, + 0.005138465203344822, + 0.006466309539973736, + 0.005376792512834072, + 0.004528438206762075, + 0.007924574427306652, + 0.006608615629374981, + 0.004375753458589315, + 0.005704345181584358, + 0.003976806998252869, + 0.007329342886805534, + 0.0063733262941241264, + 0.008152361959218979, + 0.018548129126429558, + 0.01569010131061077, + 0.012064643204212189, + 0.01496955193579197, + 0.009462538175284863, + 0.017321620136499405, + 0.011439832858741283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.020572030916810036 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a cluster of crimson leaves with burgundy hues bordering on black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves exhibit distinct veins and finely serrated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the motion of the leaves a rhythmic, spasmodic, tiny twitching rather than wind-blown movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark forest with thick fog, twisted dark brown tree trunks, and a damp mossy ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005938978865742683 + ], + "opposite_logprob_tokens": [ + -5.13093900680542 + ], + "logprob_reward": -0.005938978865742683, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005938978865742683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3006138205528259, + 0.14092259109020233, + 0.05727561190724373, + 0.045970600098371506, + 0.039681803435087204, + 0.06758975237607956, + 0.06590794026851654, + 0.09534841775894165, + 0.10179373621940613, + 0.08489570021629333 + ], + "credit_map": [ + 0.0385228656232357, + 0.0598239041864872, + 0.03675805777311325, + 0.04702479764819145, + 0.03340541571378708, + 0.02419416792690754, + 0.03532899543642998, + 0.025555606931447983, + 0.016807951033115387, + 0.023161418735980988, + 0.01832493208348751, + 0.020822880789637566, + 0.014705503359436989, + 0.013829183764755726, + 0.018464449793100357, + 0.014806268736720085, + 0.006855806801468134, + 0.01211260724812746, + 0.0077570220455527306, + 0.009124559350311756, + 0.00503276102244854, + 0.004645990207791328, + 0.0074678813107311726, + 0.004278980661183596, + 0.005069103091955185, + 0.009495777077972889, + 0.005800060462206602, + 0.008139664307236671, + 0.004941133316606283, + 0.0028901987243443727, + 0.005222683306783438, + 0.004411980975419283, + 0.003810096299275756, + 0.008741320110857487, + 0.0054186065681278706, + 0.006135174073278904, + 0.003862498328089714, + 0.003593226196244359, + 0.004711003974080086, + 0.003409875323995948, + 0.008969726040959358, + 0.01458443608134985, + 0.009847111999988556, + 0.007478610146790743, + 0.007296024356037378, + 0.005038095638155937, + 0.008400589227676392, + 0.005975162144750357, + 0.0074050589464604855, + 0.013916605152189732, + 0.009048353880643845, + 0.008772128261625767, + 0.007917806506156921, + 0.006418002303689718, + 0.007424000184983015, + 0.005005988758057356, + 0.01185896247625351, + 0.018232010304927826, + 0.014331376180052757, + 0.012871647253632545, + 0.014509664848446846, + 0.006876753177493811, + 0.0086834616959095, + 0.007984546013176441, + 0.010763086378574371, + 0.017850251868367195, + 0.015720849856734276, + 0.012228226289153099, + 0.01537652499973774, + 0.00947257038205862, + 0.012309843674302101, + 0.008072380907833576, + 0.010036123916506767, + 0.014161378145217896, + 0.01284137275069952, + 0.013188572600483894, + 0.010302785784006119, + 0.005686111282557249, + 0.011258881539106369, + 0.0074204751290380955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.005938978865742683 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_9/index_tensor([1772, 1772])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_9/index_tensor([1772, 1772])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0bfae39679ab3421ad02a4f51b2eef3a685e8497 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_9/index_tensor([1772, 1772])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning, first-person macro shot plunges the viewer completely into the sensory world of a male, the monstrous trident horn of the antagonist. The horn, a deep bronze metallic sheen, gleams with an oily iridescence in the dappled sunlight filtering through the humid jungle canopy. Its surface, punctuated by tiny scratches and pits etched with the brunt of battle, appears indestructible. The background is a completely out-of-focus blur of green and brown patches of moss and bark. Suddenly, the antagonist lunges forward, the massive horn rapidly enlarging in the field of view. A brutal, heavy impact follows, and the world is instantly transformed into a violent, dizzying tremor, the image momentarily blurred and distorted by the sheer force of the impact. In this lightning-fast collision, you can feel the tremendous force squeezing between the shells. Every unyielding struggle causes your field of vision to shake violently, perfectly capturing this life-and-death duel that takes place in the microscopic world but is full of epic grandeur and primitive cruelty, presenting a shocking texture like a top natural documentary.", + "seed": 23957, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_9/index_1772_seed_23957_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06328891664743423, + "vlm_logprob": -0.06328891664743423, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.06328891664743423, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15473239123821259 + ], + "opposite_logprob_tokens": [ + -2.2797324657440186 + ], + "logprob_reward": -0.15473239123821259, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.15473239123821259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09139362722635269, + 0.2704504430294037, + 0.05448223277926445, + 0.036468859761953354, + 0.03164000064134598, + 0.05708741769194603, + 0.06295192241668701, + 0.15957722067832947, + 0.0632467046380043, + 0.17270152270793915 + ], + "credit_map": [ + 0.00876347254961729, + 0.008476980030536652, + 0.008449116721749306, + 0.010676687583327293, + 0.01252738106995821, + 0.014641980640590191, + 0.013812363147735596, + 0.014045638963580132, + 0.03142735734581947, + 0.010260108858346939, + 0.09694960713386536, + 0.01894749142229557, + 0.009271888062357903, + 0.07692746818065643, + 0.013051006011664867, + 0.01361553929746151, + 0.0076340120285749435, + 0.009174869395792484, + 0.00959163997322321, + 0.008085985668003559, + 0.004231760278344154, + 0.005603355821222067, + 0.0047079515643417835, + 0.005452657584100962, + 0.0034081595949828625, + 0.006647841539233923, + 0.003082520328462124, + 0.006169580854475498, + 0.0021429243497550488, + 0.004128233529627323, + 0.0077925450168550014, + 0.0030970557127147913, + 0.0024591872934252024, + 0.0030411798506975174, + 0.004394862800836563, + 0.003663347102701664, + 0.004329952877014875, + 0.006014636252075434, + 0.0035045642871409655, + 0.004232268780469894, + 0.008086033165454865, + 0.005488521419465542, + 0.0067558083683252335, + 0.004078526049852371, + 0.010063767433166504, + 0.011487179435789585, + 0.005805213004350662, + 0.005322369281202555, + 0.010045627132058144, + 0.008463403210043907, + 0.007430611178278923, + 0.01150362379848957, + 0.005633768625557423, + 0.0062889885157346725, + 0.005390856880694628, + 0.008195042610168457, + 0.023219821974635124, + 0.01094125211238861, + 0.01807393692433834, + 0.012426731176674366, + 0.018422549590468407, + 0.024897735565900803, + 0.0312812402844429, + 0.020313959568738937, + 0.011227729730308056, + 0.007215859834104776, + 0.010430914349853992, + 0.006754866801202297, + 0.006663185078650713, + 0.003794952528551221, + 0.012245039455592632, + 0.0049141584895551205, + 0.018929071724414825, + 0.02341311424970627, + 0.028846951201558113, + 0.04928455501794815, + 0.010974334552884102, + 0.010653081350028515, + 0.01550210826098919, + 0.015098297968506813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.15473239123821259 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500371754169464 + ], + "opposite_logprob_tokens": [ + -2.9300036430358887 + ], + "logprob_reward": -0.05500371754169464, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500371754169464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06996580958366394, + 0.4327522814273834, + 0.09131792187690735, + 0.08591689169406891, + 0.01765173301100731, + 0.10072760283946991, + 0.036982983350753784, + 0.05031178891658783, + 0.05920236557722092, + 0.055170632898807526 + ], + "credit_map": [ + 0.005950847640633583, + 0.010333347134292126, + 0.006055335979908705, + 0.009137924760580063, + 0.0109179662540555, + 0.012425191700458527, + 0.01042984239757061, + 0.004715354181826115, + 0.05158537998795509, + 0.0215697530657053, + 0.15517671406269073, + 0.03505309671163559, + 0.017447147518396378, + 0.11317937076091766, + 0.020240968093276024, + 0.01849985681474209, + 0.009387901052832603, + 0.018021536991000175, + 0.013773811981081963, + 0.015674352645874023, + 0.00691408384591341, + 0.008839370682835579, + 0.011558772064745426, + 0.007148096803575754, + 0.01379715371876955, + 0.00973294023424387, + 0.009068664163351059, + 0.01778089813888073, + 0.0065865288488566875, + 0.009505044668912888, + 0.011768471449613571, + 0.007677190471440554, + 0.0014643807662650943, + 0.002099422737956047, + 0.0019885406363755465, + 0.0015078757423907518, + 0.003478415310382843, + 0.0026766732335090637, + 0.0024775583297014236, + 0.001958865439519286, + 0.010629688389599323, + 0.011620297096669674, + 0.010488580912351608, + 0.009863769635558128, + 0.017609018832445145, + 0.01652146875858307, + 0.013247687369585037, + 0.0107470927760005, + 0.005092053208500147, + 0.003822477301582694, + 0.00425293855369091, + 0.00764949806034565, + 0.004029910080134869, + 0.0038186488673090935, + 0.0039697932079434395, + 0.004347667563706636, + 0.006696273572742939, + 0.0033040905836969614, + 0.0047569735907018185, + 0.0042387270368635654, + 0.0070606619119644165, + 0.007809879723936319, + 0.00980795081704855, + 0.006637233309447765, + 0.007855073548853397, + 0.004162267781794071, + 0.016430960968136787, + 0.0046403054147958755, + 0.0043119629845023155, + 0.0034757968969643116, + 0.01298901904374361, + 0.005336982198059559, + 0.006794257555156946, + 0.006852468941360712, + 0.009503773413598537, + 0.013288341462612152, + 0.0037155637983232737, + 0.004301057662814856, + 0.005177541170269251, + 0.005537631455808878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.05500371754169464 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016087606549263 + ], + "opposite_logprob_tokens": [ + -4.141087532043457 + ], + "logprob_reward": -0.016087606549263, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016087606549263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2278471142053604, + 0.15405720472335815, + 0.14098384976387024, + 0.1579631268978119, + 0.04974150285124779, + 0.03626197576522827, + 0.07948631793260574, + 0.07023680955171585, + 0.031411606818437576, + 0.052010536193847656 + ], + "credit_map": [ + 0.01563592627644539, + 0.016184905543923378, + 0.013246220536530018, + 0.015316876582801342, + 0.0435529500246048, + 0.04498975723981857, + 0.03364678844809532, + 0.045273665338754654, + 0.020506389439105988, + 0.008204773999750614, + 0.051297929137945175, + 0.011603432707488537, + 0.006187621969729662, + 0.03859291970729828, + 0.008217325434088707, + 0.009446797892451286, + 0.01455689501017332, + 0.022519240155816078, + 0.014544340781867504, + 0.024599310010671616, + 0.012426525354385376, + 0.016583861783146858, + 0.01436813734471798, + 0.021385550498962402, + 0.017647769302129745, + 0.031243788078427315, + 0.014922767877578735, + 0.02591010369360447, + 0.009808509610593319, + 0.01306043192744255, + 0.030837219208478928, + 0.014532528817653656, + 0.005589244421571493, + 0.005189374089241028, + 0.00946212187409401, + 0.004354887176305056, + 0.005141394212841988, + 0.007444390095770359, + 0.005915699992328882, + 0.006644392851740122, + 0.004910261370241642, + 0.0034625923726707697, + 0.005503655876964331, + 0.002964770421385765, + 0.006094593089073896, + 0.004521308001130819, + 0.006038792431354523, + 0.0027660024352371693, + 0.011138458736240864, + 0.009206275455653667, + 0.010742394253611565, + 0.01439955085515976, + 0.007101968862116337, + 0.009074395522475243, + 0.00750268017873168, + 0.010320584289729595, + 0.010474860668182373, + 0.004669518209993839, + 0.007902838289737701, + 0.004923178814351559, + 0.00887063518166542, + 0.010748451575636864, + 0.0134103549644351, + 0.009236970916390419, + 0.0060019115917384624, + 0.003338922979310155, + 0.005479483865201473, + 0.0032488973811268806, + 0.002921867882832885, + 0.0025084156077355146, + 0.00519692525267601, + 0.002715179231017828, + 0.0056170085445046425, + 0.007290167734026909, + 0.009604467079043388, + 0.015062583610415459, + 0.0030922505538910627, + 0.0030565715860575438, + 0.0039121415466070175, + 0.004375341348350048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.016087606549263 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003015279769897 + ], + "opposite_logprob_tokens": [ + -2.6950302124023438 + ], + "logprob_reward": -0.07003015279769897, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003015279769897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10283735394477844, + 0.36230191588401794, + 0.05555680766701698, + 0.1614181399345398, + 0.062419645488262177, + 0.03515622019767761, + 0.05130978301167488, + 0.048314064741134644, + 0.05071194842457771, + 0.06997411698102951 + ], + "credit_map": [ + 0.009167861193418503, + 0.01488723885267973, + 0.008519834838807583, + 0.013439423404633999, + 0.012744071893393993, + 0.013212131336331367, + 0.012789611704647541, + 0.0180771853774786, + 0.042764123529195786, + 0.01389063149690628, + 0.12927305698394775, + 0.026298386976122856, + 0.013042768463492393, + 0.09876127541065216, + 0.0195716992020607, + 0.01869996450841427, + 0.0069974069483578205, + 0.007509966380894184, + 0.00669142184779048, + 0.00974115077406168, + 0.004061977844685316, + 0.009092798456549644, + 0.0041546327993273735, + 0.007307447958737612, + 0.023086899891495705, + 0.02065925858914852, + 0.013170452788472176, + 0.03146461024880409, + 0.011411884799599648, + 0.01780717261135578, + 0.027565253898501396, + 0.016252601519227028, + 0.006990139838308096, + 0.006689659785479307, + 0.009930463507771492, + 0.00692999642342329, + 0.008163440972566605, + 0.007290152832865715, + 0.00789398979395628, + 0.008531806990504265, + 0.004388871602714062, + 0.0034997828770428896, + 0.003840078366920352, + 0.003308721585199237, + 0.004587441682815552, + 0.0037486900109797716, + 0.008391539566218853, + 0.003391098231077194, + 0.005729879252612591, + 0.007320603355765343, + 0.0054096016101539135, + 0.008294869214296341, + 0.007226037792861462, + 0.006083416286855936, + 0.005405111704021692, + 0.005840263329446316, + 0.005701213609427214, + 0.0027457447722554207, + 0.0037969995755702257, + 0.004203118849545717, + 0.005422658286988735, + 0.00766560435295105, + 0.010208992287516594, + 0.008569732308387756, + 0.007331981789320707, + 0.00453953817486763, + 0.011582864448428154, + 0.003748438786715269, + 0.005790680181235075, + 0.0025324856396764517, + 0.011173720471560955, + 0.0040122391656041145, + 0.008302677422761917, + 0.008267640136182308, + 0.010281103663146496, + 0.017728291451931, + 0.005195487756282091, + 0.005397673696279526, + 0.008634215220808983, + 0.006167030427604914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.07003015279769897 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059071511030197 + ], + "opposite_logprob_tokens": [ + -3.8955907821655273 + ], + "logprob_reward": -0.02059071511030197, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059071511030197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08127683401107788, + 0.18602265417575836, + 0.05773589387536049, + 0.35206201672554016, + 0.04982587695121765, + 0.03941471874713898, + 0.04842996969819069, + 0.07024529576301575, + 0.06956135481595993, + 0.045425452291965485 + ], + "credit_map": [ + 0.0071597048081457615, + 0.011415264569222927, + 0.007235080935060978, + 0.011765801347792149, + 0.011457711458206177, + 0.014650802128016949, + 0.011151655577123165, + 0.006440815515816212, + 0.020013242959976196, + 0.009306071326136589, + 0.06462795287370682, + 0.014520065858960152, + 0.007462302688509226, + 0.050929225981235504, + 0.009939471259713173, + 0.009224319830536842, + 0.005499362014234066, + 0.010114136151969433, + 0.011228537186980247, + 0.007865688763558865, + 0.005180830601602793, + 0.004472058266401291, + 0.007923685945570469, + 0.005451591219753027, + 0.05602292716503143, + 0.053449422121047974, + 0.03508840501308441, + 0.06310784071683884, + 0.025358108803629875, + 0.03761604428291321, + 0.04776445031166077, + 0.03365481272339821, + 0.0040812105871737, + 0.007451609242707491, + 0.00661293463781476, + 0.005571982357650995, + 0.007410790305584669, + 0.006448129657655954, + 0.00592505419626832, + 0.006324161775410175, + 0.004436452407389879, + 0.005000981967896223, + 0.003424579743295908, + 0.004033096134662628, + 0.007069800980389118, + 0.004806590266525745, + 0.007663062307983637, + 0.002980157034471631, + 0.006341948639601469, + 0.00531007768586278, + 0.004500540439039469, + 0.005754484795033932, + 0.006472125183790922, + 0.006098994053900242, + 0.005694775842130184, + 0.008257023990154266, + 0.00770565727725625, + 0.003983775153756142, + 0.0062246466986835, + 0.004930316470563412, + 0.008000221103429794, + 0.012931885197758675, + 0.01417888980358839, + 0.012289904989302158, + 0.013939478434622288, + 0.007386382203549147, + 0.011864254251122475, + 0.0062925186939537525, + 0.008248435333371162, + 0.004185355268418789, + 0.01209341362118721, + 0.005551518406718969, + 0.00465770298615098, + 0.005558962933719158, + 0.005316306371241808, + 0.006660602055490017, + 0.0053564319387078285, + 0.004624302498996258, + 0.0070845214650034904, + 0.006166617851704359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.02059071511030197 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning, first-person macro shot plunges the viewer completely into the sensory world of a male, the monstrous trident horn of the antagonist. The horn, a deep bronze metallic sheen, gleams with an oily iridescence in the dappled sunlight filtering through the humid jungle canopy. Its surface, punctuated by tiny scratches and pits etched with the brunt of battle, appears indestructible. The background is a completely out-of-focus blur of green and brown patches of moss and bark. Suddenly, the antagonist lunges forward, the massive horn rapidly enlarging in the field of view. A brutal, heavy impact follows, and the world is instantly transformed into a violent, dizzying tremor, the image momentarily blurred and distorted by the sheer force of the impact. In this lightning-fast collision, you can feel the tremendous force squeezing between the shells. Every unyielding struggle causes your field of vision to shake violently, perfectly capturing this life-and-death duel that takes place in the microscopic world but is full of epic grandeur and primitive cruelty, presenting a shocking texture like a top natural documentary.", + "seed": 23958, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_9/index_1772_seed_23958_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03757432186976075, + "vlm_logprob": -0.03757432186976075, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.03757432186976075, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1018398180603981 + ], + "opposite_logprob_tokens": [ + -2.601839780807495 + ], + "logprob_reward": -0.1018398180603981, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.1018398180603981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2819710373878479, + 0.09438169002532959, + 0.043074000626802444, + 0.17754852771759033, + 0.05995616689324379, + 0.02167757973074913, + 0.12873400747776031, + 0.11504147201776505, + 0.04240243509411812, + 0.03521304950118065 + ], + "credit_map": [ + 0.1115909069776535, + 0.014681680127978325, + 0.019593197852373123, + 0.00708965677767992, + 0.006795959547162056, + 0.010692810639739037, + 0.01707402430474758, + 0.09445282816886902, + 0.0076677934266626835, + 0.006013303063809872, + 0.029314320534467697, + 0.006132978945970535, + 0.004604196175932884, + 0.006403428968042135, + 0.010892395861446857, + 0.02335326559841633, + 0.007139831781387329, + 0.007190960459411144, + 0.003212608629837632, + 0.003610365791246295, + 0.0042987060733139515, + 0.006440530531108379, + 0.0055583748035132885, + 0.005622623488306999, + 0.0060091461054980755, + 0.010804117657244205, + 0.009693624451756477, + 0.06020786985754967, + 0.015119862742722034, + 0.008256847970187664, + 0.056275688111782074, + 0.011181358247995377, + 0.005663837306201458, + 0.0064925020560622215, + 0.006725986488163471, + 0.012521319091320038, + 0.005891022738069296, + 0.005015705246478319, + 0.008636300452053547, + 0.00900949165225029, + 0.002371013630181551, + 0.001860979711636901, + 0.0030738378409296274, + 0.004009527154266834, + 0.0021653755102306604, + 0.0024953268002718687, + 0.002362630097195506, + 0.003338888753205538, + 0.01430561114102602, + 0.019503595307469368, + 0.016249150037765503, + 0.01609785296022892, + 0.015025390312075615, + 0.012595436535775661, + 0.02310667186975479, + 0.011850293725728989, + 0.010135937482118607, + 0.008237114176154137, + 0.03273477405309677, + 0.014013993553817272, + 0.007951561361551285, + 0.02141852304339409, + 0.010116314515471458, + 0.010433259420096874, + 0.004795577842742205, + 0.0045327735133469105, + 0.006198131013661623, + 0.0065693119540810585, + 0.006170759908854961, + 0.003337198169901967, + 0.0053038462065160275, + 0.005494836252182722, + 0.0030679989140480757, + 0.005627128761261702, + 0.005201682914048433, + 0.0056523727253079414, + 0.00444358354434371, + 0.003176473779603839, + 0.003845630679279566, + 0.004198175389319658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.1018398180603981 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012570316903293133 + ], + "opposite_logprob_tokens": [ + -4.387570381164551 + ], + "logprob_reward": -0.012570316903293133, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012570316903293133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08690007776021957, + 0.15189604461193085, + 0.18408356606960297, + 0.2777237892150879, + 0.03901864215731621, + 0.03618698939681053, + 0.049320872873067856, + 0.09649135172367096, + 0.040511880069971085, + 0.037866782397031784 + ], + "credit_map": [ + 0.026286331936717033, + 0.004842936061322689, + 0.008037433959543705, + 0.007034139707684517, + 0.003508263733237982, + 0.004905069712549448, + 0.007475934457033873, + 0.02480996772646904, + 0.010648692958056927, + 0.008684225380420685, + 0.05136119946837425, + 0.006388666573911905, + 0.0055506243370473385, + 0.008750208653509617, + 0.017382707446813583, + 0.043129727244377136, + 0.021990668028593063, + 0.011764494702219963, + 0.010286793112754822, + 0.013895812444388866, + 0.011009623296558857, + 0.04938347637653351, + 0.03079967387020588, + 0.0349530354142189, + 0.007341853342950344, + 0.010103532113134861, + 0.008621455170214176, + 0.10741571336984634, + 0.01682831160724163, + 0.00663908664137125, + 0.10634361207485199, + 0.014430239796638489, + 0.002375758485868573, + 0.004272993188351393, + 0.003576201619580388, + 0.006651872768998146, + 0.0035102185793220997, + 0.004134719260036945, + 0.0067145065404474735, + 0.007782370317727327, + 0.003923206124454737, + 0.0027002207934856415, + 0.004517186898738146, + 0.006839150097221136, + 0.0031777448020875454, + 0.004647077992558479, + 0.0037326128222048283, + 0.00664979312568903, + 0.007460856810212135, + 0.01071622408926487, + 0.005812197923660278, + 0.006447901949286461, + 0.005115248262882233, + 0.004759658128023148, + 0.005156554281711578, + 0.003852234221994877, + 0.007913699373602867, + 0.004082155879586935, + 0.03561148792505264, + 0.012869619764387608, + 0.009906782768666744, + 0.015601670369505882, + 0.004888047929853201, + 0.005617898423224688, + 0.006256533320993185, + 0.00267422036267817, + 0.0069191413931548595, + 0.004887768533080816, + 0.0063597154803574085, + 0.003266659565269947, + 0.006506603676825762, + 0.0036412400659173727, + 0.004222039598971605, + 0.0031315088272094727, + 0.006713907700031996, + 0.005471468437463045, + 0.0038748534861952066, + 0.0034396122209727764, + 0.005170661024749279, + 0.005842733196914196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.012570316903293133 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012552776373922825 + ], + "opposite_logprob_tokens": [ + -4.387552738189697 + ], + "logprob_reward": -0.012552776373922825, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012552776373922825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10325063019990921, + 0.09997204691171646, + 0.07229181379079819, + 0.22415198385715485, + 0.045680779963731766, + 0.05374795198440552, + 0.049933481961488724, + 0.10528898239135742, + 0.1395282745361328, + 0.10615403205156326 + ], + "credit_map": [ + 0.0342683307826519, + 0.005181985907256603, + 0.011989192105829716, + 0.005967150442302227, + 0.003710684133693576, + 0.007346649654209614, + 0.006421773228794336, + 0.028364870697259903, + 0.01093969400972128, + 0.009433905594050884, + 0.027333978563547134, + 0.00523789320141077, + 0.004926841240376234, + 0.008058261126279831, + 0.011003193445503712, + 0.023038281127810478, + 0.009885972365736961, + 0.004911035764962435, + 0.004988174885511398, + 0.0047113290056586266, + 0.004761595278978348, + 0.021773869171738625, + 0.009508942253887653, + 0.01175090204924345, + 0.007525601889938116, + 0.01343039982020855, + 0.0099748270586133, + 0.06694845110177994, + 0.0175964143127203, + 0.013583631254732609, + 0.07258670032024384, + 0.02250596508383751, + 0.00362783158197999, + 0.0048581454902887344, + 0.0035421729553490877, + 0.006328628398478031, + 0.0052863191813230515, + 0.006402213592082262, + 0.009849302470684052, + 0.0057861702516674995, + 0.007754437159746885, + 0.005339758936315775, + 0.006726669147610664, + 0.010899024084210396, + 0.004341332241892815, + 0.005263130180537701, + 0.005710931029170752, + 0.007712673861533403, + 0.006993854884058237, + 0.005301715340465307, + 0.009413664229214191, + 0.005262556951493025, + 0.006824875250458717, + 0.005047971848398447, + 0.007200176827609539, + 0.003888667793944478, + 0.007863370701670647, + 0.0069646332412958145, + 0.03468434885144234, + 0.01066350657492876, + 0.018871253356337547, + 0.010763612575829029, + 0.00817717146128416, + 0.007301085162907839, + 0.017701538279652596, + 0.005663736257702112, + 0.024366093799471855, + 0.009096949361264706, + 0.02606656216084957, + 0.00943704228848219, + 0.038182929158210754, + 0.009013442322611809, + 0.009446259588003159, + 0.013503432273864746, + 0.014777342788875103, + 0.013164548203349113, + 0.012512045912444592, + 0.009271454997360706, + 0.013160252943634987, + 0.02031869813799858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.012552776373922825 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495632812380791 + ], + "opposite_logprob_tokens": [ + -2.9299564361572266 + ], + "logprob_reward": -0.05495632812380791, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495632812380791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3032439351081848, + 0.0314297080039978, + 0.03699814900755882, + 0.19323945045471191, + 0.04288787022233009, + 0.025898678228259087, + 0.03852231055498123, + 0.21807824075222015, + 0.08154990524053574, + 0.02815181203186512 + ], + "credit_map": [ + 0.11862508952617645, + 0.015393756330013275, + 0.021266892552375793, + 0.008344404399394989, + 0.007348934188485146, + 0.01156840194016695, + 0.018995406106114388, + 0.1017010286450386, + 0.004232572857290506, + 0.003151476150378585, + 0.0050954935140907764, + 0.0035591591149568558, + 0.002886170521378517, + 0.0030163938645273447, + 0.004688948392868042, + 0.004799496848136187, + 0.005964345298707485, + 0.004575855564326048, + 0.002569144358858466, + 0.002966958796605468, + 0.003663907293230295, + 0.006832337472587824, + 0.004379608668386936, + 0.006045992951840162, + 0.006041571963578463, + 0.00893219094723463, + 0.007897019386291504, + 0.06924158334732056, + 0.01203864999115467, + 0.006289698649197817, + 0.07109112292528152, + 0.011707606725394726, + 0.0057989670895040035, + 0.007302448619157076, + 0.006161232478916645, + 0.006607902701944113, + 0.004153507761657238, + 0.003251679940149188, + 0.006308800540864468, + 0.003303329925984144, + 0.003393214661628008, + 0.0027466979809105396, + 0.0029819700866937637, + 0.003917698282748461, + 0.0026749330572783947, + 0.002563618356361985, + 0.0031337232794612646, + 0.004486823454499245, + 0.006201745942234993, + 0.004853506106883287, + 0.004887079354375601, + 0.004221003502607346, + 0.0038905658293515444, + 0.005383036565035582, + 0.005783223547041416, + 0.0033021485432982445, + 0.018261374905705452, + 0.009619055315852165, + 0.07664776593446732, + 0.02880573272705078, + 0.019162505865097046, + 0.03894137963652611, + 0.01187855750322342, + 0.014761875383555889, + 0.011089646257460117, + 0.007300912402570248, + 0.012118636630475521, + 0.013009154237806797, + 0.01140198577195406, + 0.005686209071427584, + 0.012111575342714787, + 0.00883178599178791, + 0.0031148323323577642, + 0.002981893951073289, + 0.00451394310221076, + 0.003765370463952422, + 0.0034800099674612284, + 0.0022935220040380955, + 0.004355661105364561, + 0.003646580269560218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.05495632812380791 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952369887381792 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952369887381792, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952369887381792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04814618080854416, + 0.12080587446689606, + 0.040585920214653015, + 0.17959260940551758, + 0.04639303684234619, + 0.04307550936937332, + 0.06295681744813919, + 0.10838696360588074, + 0.22420501708984375, + 0.1258520483970642 + ], + "credit_map": [ + 0.011426233686506748, + 0.003041498363018036, + 0.006295157130807638, + 0.005437391810119152, + 0.0025875435676425695, + 0.0028549928683787584, + 0.004715315531939268, + 0.01178804598748684, + 0.009009402245283127, + 0.010392631404101849, + 0.04246623441576958, + 0.00581781892105937, + 0.004966691602021456, + 0.0062456210143864155, + 0.011219285428524017, + 0.030688196420669556, + 0.007099620532244444, + 0.0034028238151222467, + 0.003077940782532096, + 0.004344083368778229, + 0.0026596500538289547, + 0.006580503191798925, + 0.007263414096087217, + 0.006157885305583477, + 0.005503694526851177, + 0.007796669844537973, + 0.010660091415047646, + 0.06799452006816864, + 0.011007608845829964, + 0.006222538650035858, + 0.059786949306726456, + 0.010620508342981339, + 0.004582267254590988, + 0.00731326499953866, + 0.006000939290970564, + 0.008838438428938389, + 0.005003552418202162, + 0.0034505254589021206, + 0.006295792292803526, + 0.004908257629722357, + 0.004858733620494604, + 0.005256802774965763, + 0.004790483508259058, + 0.00855326373130083, + 0.002987365936860442, + 0.0051139057613909245, + 0.005381803028285503, + 0.006133149843662977, + 0.0076982369646430016, + 0.007211743388324976, + 0.011431674472987652, + 0.006918254774063826, + 0.0069513279013335705, + 0.006959216203540564, + 0.010478668846189976, + 0.005307699088007212, + 0.008797043934464455, + 0.0062770480290055275, + 0.0357433520257473, + 0.014040507376194, + 0.01065338309854269, + 0.017524484544992447, + 0.008834556676447392, + 0.006516596302390099, + 0.03442750126123428, + 0.011558881029486656, + 0.037048403173685074, + 0.023050665855407715, + 0.04136161133646965, + 0.013430588878691196, + 0.04799249395728111, + 0.015334871597588062, + 0.01589367352426052, + 0.016900258138775826, + 0.02614876627922058, + 0.016150379553437233, + 0.01062167901545763, + 0.00997362844645977, + 0.012388174422085285, + 0.017775490880012512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.005952369887381792 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning, first-person macro shot plunges the viewer completely into the sensory world of a male, the monstrous trident horn of the antagonist. The horn, a deep bronze metallic sheen, gleams with an oily iridescence in the dappled sunlight filtering through the humid jungle canopy. Its surface, punctuated by tiny scratches and pits etched with the brunt of battle, appears indestructible. The background is a completely out-of-focus blur of green and brown patches of moss and bark. Suddenly, the antagonist lunges forward, the massive horn rapidly enlarging in the field of view. A brutal, heavy impact follows, and the world is instantly transformed into a violent, dizzying tremor, the image momentarily blurred and distorted by the sheer force of the impact. In this lightning-fast collision, you can feel the tremendous force squeezing between the shells. Every unyielding struggle causes your field of vision to shake violently, perfectly capturing this life-and-death duel that takes place in the microscopic world but is full of epic grandeur and primitive cruelty, presenting a shocking texture like a top natural documentary.", + "seed": 23959, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_9/index_1772_seed_23959_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04975554719567299, + "vlm_logprob": -0.04975554719567299, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.04975554719567299, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13856786489486694 + ], + "opposite_logprob_tokens": [ + -2.2635679244995117 + ], + "logprob_reward": -0.13856786489486694, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.13856786489486694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10642708092927933, + 0.11493315547704697, + 0.042371802031993866, + 0.10890962183475494, + 0.0230511836707592, + 0.07418720424175262, + 0.04761306196451187, + 0.32351651787757874, + 0.1026177704334259, + 0.05637262016534805 + ], + "credit_map": [ + 0.012357264757156372, + 0.013016147539019585, + 0.013573355972766876, + 0.011481867171823978, + 0.009140965528786182, + 0.01204555295407772, + 0.02532096579670906, + 0.009490961208939552, + 0.017621878534555435, + 0.02001630701124668, + 0.00981444213539362, + 0.015364541672170162, + 0.009064553305506706, + 0.01957629807293415, + 0.010311111807823181, + 0.013164020143449306, + 0.008296866901218891, + 0.004240995738655329, + 0.004458990413695574, + 0.005774194840341806, + 0.0027954645920544863, + 0.003901277668774128, + 0.004517967812716961, + 0.008386049419641495, + 0.029257021844387054, + 0.007659441791474819, + 0.011002915911376476, + 0.007788179442286491, + 0.007306723389774561, + 0.013646970503032207, + 0.014879235066473484, + 0.01736913062632084, + 0.002214079489931464, + 0.0021720377262681723, + 0.0018773633055388927, + 0.0026944605633616447, + 0.00249125761911273, + 0.003706261282786727, + 0.00282409880310297, + 0.005071624182164669, + 0.00953147653490305, + 0.006625378970056772, + 0.009925867430865765, + 0.01572461612522602, + 0.006125524640083313, + 0.006596543360501528, + 0.01292395405471325, + 0.0067338417284190655, + 0.00528140552341938, + 0.004651470575481653, + 0.006109599955379963, + 0.006649659480899572, + 0.005138749722391367, + 0.007105621509253979, + 0.006576600018888712, + 0.006099957972764969, + 0.027796778827905655, + 0.03618687391281128, + 0.057595834136009216, + 0.04449504613876343, + 0.03494837507605553, + 0.05344516038894653, + 0.04391266405582428, + 0.025135789066553116, + 0.009279315359890461, + 0.008171145804226398, + 0.011766159906983376, + 0.010038378648459911, + 0.010288896039128304, + 0.006584411021322012, + 0.02599046193063259, + 0.020499005913734436, + 0.007946173660457134, + 0.0067067015916109085, + 0.011258344165980816, + 0.008856300264596939, + 0.005561019293963909, + 0.005337805952876806, + 0.006616110913455486, + 0.00409016152843833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.13856786489486694 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870496690273285 + ], + "opposite_logprob_tokens": [ + -3.0487048625946045 + ], + "logprob_reward": -0.04870496690273285, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870496690273285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06985686719417572, + 0.1878371387720108, + 0.11523573100566864, + 0.012884565629065037, + 0.050977662205696106, + 0.046410467475652695, + 0.059281401336193085, + 0.2022557556629181, + 0.1576034277677536, + 0.0976569652557373 + ], + "credit_map": [ + 0.007271602284163237, + 0.011258605867624283, + 0.009820451959967613, + 0.007678798399865627, + 0.009428403340280056, + 0.010480410419404507, + 0.007911670953035355, + 0.006006927695125341, + 0.034621551632881165, + 0.04566342383623123, + 0.01564282551407814, + 0.020373549312353134, + 0.01481573935598135, + 0.01883033663034439, + 0.017801981419324875, + 0.020087741315364838, + 0.027811700478196144, + 0.009110168553888798, + 0.011889631859958172, + 0.011586670763790607, + 0.0074139488860964775, + 0.0077750482596457005, + 0.011468002572655678, + 0.028180569410324097, + 0.002501989249140024, + 0.0009610478300601244, + 0.0013142572715878487, + 0.0011230037780478597, + 0.0010781039018183947, + 0.0016535803442820907, + 0.002353159710764885, + 0.0018994236597791314, + 0.0076048001646995544, + 0.00487933773547411, + 0.003950924146920443, + 0.0049235885962843895, + 0.0060263765044510365, + 0.010188928805291653, + 0.006205916404724121, + 0.0071977912448346615, + 0.005822775885462761, + 0.003461236134171486, + 0.005433019250631332, + 0.01022831816226244, + 0.004453166853636503, + 0.006666816771030426, + 0.0058604576624929905, + 0.004484676755964756, + 0.008545900695025921, + 0.006620400119572878, + 0.006606586277484894, + 0.008312012068927288, + 0.007097859401255846, + 0.00924464501440525, + 0.0069249458611011505, + 0.0059290542267262936, + 0.01293441653251648, + 0.02161695808172226, + 0.027160847559571266, + 0.026384031400084496, + 0.02691381238400936, + 0.039928123354911804, + 0.032616857439279556, + 0.014700694009661674, + 0.008411666378378868, + 0.005329140927642584, + 0.013573821634054184, + 0.018085069954395294, + 0.012124205008149147, + 0.0074169752188026905, + 0.06117234006524086, + 0.031490206718444824, + 0.011367815546691418, + 0.009000248275697231, + 0.009790672920644283, + 0.010559977032244205, + 0.012159541249275208, + 0.011655507609248161, + 0.020860550925135612, + 0.012262648902833462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.04870496690273285 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009798288345336914 + ], + "opposite_logprob_tokens": [ + -4.634798049926758 + ], + "logprob_reward": -0.009798288345336914, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009798288345336914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23039942979812622, + 0.1086798831820488, + 0.0835518091917038, + 0.10263661295175552, + 0.0975184515118599, + 0.05021738260984421, + 0.07355684787034988, + 0.11863261461257935, + 0.06941981613636017, + 0.06538720428943634 + ], + "credit_map": [ + 0.02974238246679306, + 0.038816992193460464, + 0.0364878848195076, + 0.027017656713724136, + 0.02001732401549816, + 0.030960800126194954, + 0.025492319837212563, + 0.021864091977477074, + 0.013155367225408554, + 0.016038568690419197, + 0.012228962033987045, + 0.010537355206906796, + 0.01081296056509018, + 0.01348989550024271, + 0.017788833007216454, + 0.014627939090132713, + 0.01156883779913187, + 0.009635208174586296, + 0.017479779198765755, + 0.011167474091053009, + 0.005708352196961641, + 0.006490284111350775, + 0.008324720896780491, + 0.013177147135138512, + 0.01916268840432167, + 0.007487101946026087, + 0.011121363379061222, + 0.0075996629893779755, + 0.00715858768671751, + 0.010070910677313805, + 0.021634498611092567, + 0.018401799723505974, + 0.011672996915876865, + 0.007363484241068363, + 0.0064605106599628925, + 0.008300990797579288, + 0.010097777470946312, + 0.024744732305407524, + 0.012159725651144981, + 0.01671823300421238, + 0.007713298313319683, + 0.005154665093868971, + 0.006717456504702568, + 0.006064619403332472, + 0.005876806564629078, + 0.007330333814024925, + 0.005638778675347567, + 0.005721422843635082, + 0.0055909897200763226, + 0.007887298241257668, + 0.01326474454253912, + 0.01097120065242052, + 0.008577076718211174, + 0.008909263648092747, + 0.01117938756942749, + 0.007176884450018406, + 0.008702430874109268, + 0.0126131447032094, + 0.01730688475072384, + 0.01753293164074421, + 0.01333252340555191, + 0.02208275906741619, + 0.016398634761571884, + 0.010663309134542942, + 0.004702675621956587, + 0.00574450520798564, + 0.007529775146394968, + 0.00834280252456665, + 0.009457935579121113, + 0.003880432341247797, + 0.018092459067702293, + 0.01166923251003027, + 0.011475235223770142, + 0.008221282623708248, + 0.010148950852453709, + 0.007281379774212837, + 0.006383099593222141, + 0.006576359737664461, + 0.009972820989787579, + 0.005328078288584948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.009798288345336914 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306408390402794 + ], + "opposite_logprob_tokens": [ + -3.1680641174316406 + ], + "logprob_reward": -0.04306408390402794, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306408390402794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09721311181783676, + 0.08403859287500381, + 0.03707455098628998, + 0.11089916527271271, + 0.0512026809155941, + 0.06900928169488907, + 0.10269539803266525, + 0.2649014890193939, + 0.11842866241931915, + 0.06453712284564972 + ], + "credit_map": [ + 0.009687039069831371, + 0.012030554004013538, + 0.01099264994263649, + 0.01074217539280653, + 0.011526891961693764, + 0.017800919711589813, + 0.016251424327492714, + 0.008181450888514519, + 0.012937799096107483, + 0.015910329297184944, + 0.009894100949168205, + 0.01070162933319807, + 0.006702106911689043, + 0.011766084469854832, + 0.009066808968782425, + 0.007059729192405939, + 0.009158543311059475, + 0.0037614803295582533, + 0.004681513179093599, + 0.003913210239261389, + 0.0022378764115273952, + 0.002599894069135189, + 0.0033573906403034925, + 0.007364639546722174, + 0.028123805299401283, + 0.007531581446528435, + 0.0120402742177248, + 0.008056312799453735, + 0.007901709526777267, + 0.014161857776343822, + 0.015572696924209595, + 0.0175109151750803, + 0.005238494835793972, + 0.0033668805845081806, + 0.002906941343098879, + 0.004715755581855774, + 0.008440477773547173, + 0.007006670814007521, + 0.006097165402024984, + 0.013430289924144745, + 0.014379695989191532, + 0.008576877415180206, + 0.01025357749313116, + 0.008873817510902882, + 0.007208640221506357, + 0.007947241887450218, + 0.0056409589014947414, + 0.006128468085080385, + 0.008593546226620674, + 0.009967650286853313, + 0.019132563844323158, + 0.015506175346672535, + 0.01226730551570654, + 0.013750473968684673, + 0.012544351629912853, + 0.01093332190066576, + 0.017407577484846115, + 0.03101624734699726, + 0.04383828490972519, + 0.03338552638888359, + 0.03373853117227554, + 0.05252432823181152, + 0.0358707495033741, + 0.017120204865932465, + 0.008079873397946358, + 0.008035589940845966, + 0.01403568685054779, + 0.01533710677176714, + 0.00866534374654293, + 0.007411140948534012, + 0.03866628184914589, + 0.01819762773811817, + 0.006800392642617226, + 0.007340091746300459, + 0.009139358066022396, + 0.011689099483191967, + 0.006974663585424423, + 0.007084751967340708, + 0.00902059767395258, + 0.0064881653524935246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.04306408390402794 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642531931400299 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.008642531931400299, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642531931400299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0831855982542038, + 0.09096060693264008, + 0.03942858800292015, + 0.050235625356435776, + 0.044443268328905106, + 0.05410366132855415, + 0.1988755166530609, + 0.21415983140468597, + 0.1671881377696991, + 0.057419151067733765 + ], + "credit_map": [ + 0.0118038235232234, + 0.011805701069533825, + 0.010951032862067223, + 0.012655088678002357, + 0.007976572960615158, + 0.01090248767286539, + 0.011218726634979248, + 0.0058721620589494705, + 0.009679446928203106, + 0.015344666317105293, + 0.008553929626941681, + 0.013417380861938, + 0.00840718112885952, + 0.014712627977132797, + 0.010241538286209106, + 0.010603831149637699, + 0.003761336673051119, + 0.007764972280710936, + 0.005596582777798176, + 0.006600316613912582, + 0.00337677332572639, + 0.0034921374171972275, + 0.004458965267986059, + 0.0043775043450295925, + 0.013283662497997284, + 0.003569303546100855, + 0.004571355413645506, + 0.0037008277140557766, + 0.0034454145934432745, + 0.006525369361042976, + 0.007278406992554665, + 0.007861282676458359, + 0.005014096852391958, + 0.0034585162065923214, + 0.002378457924351096, + 0.002690025372430682, + 0.0054796854965388775, + 0.006292473524808884, + 0.006296725943684578, + 0.012833287008106709, + 0.010251599363982677, + 0.006624045316129923, + 0.00767911272123456, + 0.007837139070034027, + 0.005159303545951843, + 0.007155834697186947, + 0.00465260399505496, + 0.0047440193593502045, + 0.014875415712594986, + 0.01720748096704483, + 0.044120557606220245, + 0.03300716355443001, + 0.019661828875541687, + 0.029285429045557976, + 0.023314977064728737, + 0.01740267314016819, + 0.012848174199461937, + 0.02602749690413475, + 0.03288275748491287, + 0.02611207589507103, + 0.029804985970258713, + 0.031389519572257996, + 0.03353940695524216, + 0.021555420011281967, + 0.012695017270743847, + 0.01292746514081955, + 0.018986744806170464, + 0.021319802850484848, + 0.018200015649199486, + 0.010279546491801739, + 0.04732664301991463, + 0.02545289322733879, + 0.00842368882149458, + 0.005231363233178854, + 0.009689006954431534, + 0.00718893064185977, + 0.005930778104811907, + 0.0063521466217935085, + 0.008955869823694229, + 0.005647370126098394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.008642531931400299 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning, first-person macro shot plunges the viewer completely into the sensory world of a male, the monstrous trident horn of the antagonist. The horn, a deep bronze metallic sheen, gleams with an oily iridescence in the dappled sunlight filtering through the humid jungle canopy. Its surface, punctuated by tiny scratches and pits etched with the brunt of battle, appears indestructible. The background is a completely out-of-focus blur of green and brown patches of moss and bark. Suddenly, the antagonist lunges forward, the massive horn rapidly enlarging in the field of view. A brutal, heavy impact follows, and the world is instantly transformed into a violent, dizzying tremor, the image momentarily blurred and distorted by the sheer force of the impact. In this lightning-fast collision, you can feel the tremendous force squeezing between the shells. Every unyielding struggle causes your field of vision to shake violently, perfectly capturing this life-and-death duel that takes place in the microscopic world but is full of epic grandeur and primitive cruelty, presenting a shocking texture like a top natural documentary.", + "seed": 23960, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000130/group_9/index_1772_seed_23960_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05324757378548384, + "vlm_logprob": -0.05324757378548384, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.05324757378548384, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1748959720134735 + ], + "opposite_logprob_tokens": [ + -2.049896001815796 + ], + "logprob_reward": -0.1748959720134735, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1748959720134735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07285477966070175, + 0.0585654154419899, + 0.05865360051393509, + 0.10290061682462692, + 0.04981924593448639, + 0.2219829559326172, + 0.06270768493413925, + 0.06070581451058388, + 0.03199917823076248, + 0.27981069684028625 + ], + "credit_map": [ + 0.009628498926758766, + 0.007746517192572355, + 0.006917549762874842, + 0.008228796534240246, + 0.017415212467312813, + 0.01260315254330635, + 0.005332967732101679, + 0.004982080310583115, + 0.0072678676806390285, + 0.008809246122837067, + 0.007479629013687372, + 0.007706393487751484, + 0.006213447544723749, + 0.00633304612711072, + 0.008456203155219555, + 0.006299581844359636, + 0.010072601027786732, + 0.009400459006428719, + 0.006990056484937668, + 0.006249772850424051, + 0.005334870889782906, + 0.0042624999769032, + 0.007215125486254692, + 0.009128215722739697, + 0.008813854306936264, + 0.009212745353579521, + 0.007010114379227161, + 0.011922859586775303, + 0.01010965183377266, + 0.02818162366747856, + 0.01914900913834572, + 0.008500758558511734, + 0.00971976201981306, + 0.003607394639402628, + 0.003778258804231882, + 0.005714900325983763, + 0.004711796063929796, + 0.006603062618523836, + 0.006986822932958603, + 0.008697248063981533, + 0.015081976540386677, + 0.009920895099639893, + 0.007331564091145992, + 0.011352299712598324, + 0.01700257509946823, + 0.08717053383588791, + 0.06346334517002106, + 0.010659762658178806, + 0.005024378653615713, + 0.005355967208743095, + 0.009823019616305828, + 0.007505753077566624, + 0.009074732661247253, + 0.0097718620672822, + 0.007913623936474323, + 0.008238346315920353, + 0.010343046858906746, + 0.006652305833995342, + 0.007129858713597059, + 0.008486796170473099, + 0.0038905427791178226, + 0.008054649457335472, + 0.006885053589940071, + 0.009263554587960243, + 0.0038038427010178566, + 0.003608577186241746, + 0.003209562972187996, + 0.004616394639015198, + 0.0035402532666921616, + 0.003075909335166216, + 0.006696431897580624, + 0.0034482080955058336, + 0.03870195895433426, + 0.03418341651558876, + 0.029132146388292313, + 0.04486967623233795, + 0.043080560863018036, + 0.013115583918988705, + 0.055339064449071884, + 0.02138829603791237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 3 + ], + "selected_reward": -0.1748959720134735 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03375665098428726 + ], + "opposite_logprob_tokens": [ + -3.408756732940674 + ], + "logprob_reward": -0.03375665098428726, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03375665098428726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08607787638902664, + 0.18273772299289703, + 0.06025554984807968, + 0.09841156750917435, + 0.05792786926031113, + 0.3109636902809143, + 0.057569943368434906, + 0.04666867107152939, + 0.02625480853021145, + 0.07313229888677597 + ], + "credit_map": [ + 0.012001515366137028, + 0.007994872517883778, + 0.00987324770539999, + 0.007257004734128714, + 0.0178456399589777, + 0.015454987995326519, + 0.0065471031703054905, + 0.009103501215577126, + 0.01740111969411373, + 0.018372777849435806, + 0.030482111498713493, + 0.0235934779047966, + 0.030699413269758224, + 0.02136359177529812, + 0.0237580556422472, + 0.017067179083824158, + 0.008421647362411022, + 0.008807613514363766, + 0.009001916274428368, + 0.006641685031354427, + 0.005408782511949539, + 0.0040045809000730515, + 0.008010292425751686, + 0.009959031827747822, + 0.008813477121293545, + 0.006698970682919025, + 0.005248307716101408, + 0.007684201933443546, + 0.007982362993061543, + 0.03220442682504654, + 0.009562070481479168, + 0.02021774649620056, + 0.006076352670788765, + 0.006528913974761963, + 0.0054070549085736275, + 0.007818990387022495, + 0.006547615397721529, + 0.009262552484869957, + 0.010606975294649601, + 0.005679414141923189, + 0.01731783151626587, + 0.014332110993564129, + 0.0115512078627944, + 0.015037529170513153, + 0.03256319835782051, + 0.12126585841178894, + 0.08372526615858078, + 0.015170666389167309, + 0.005811756011098623, + 0.006918307859450579, + 0.0077185435220599174, + 0.008013434708118439, + 0.008547769859433174, + 0.007287946529686451, + 0.006834140978753567, + 0.006438042968511581, + 0.006407570093870163, + 0.004222666844725609, + 0.005373623687773943, + 0.007111971732228994, + 0.0033967806957662106, + 0.006472116336226463, + 0.005714057944715023, + 0.007969881407916546, + 0.0028160791844129562, + 0.003153777215629816, + 0.004429305903613567, + 0.0032496375497430563, + 0.004689338617026806, + 0.0021204326767474413, + 0.0036378151271492243, + 0.002158421790227294, + 0.006915044039487839, + 0.009467212483286858, + 0.007761507760733366, + 0.011958650313317776, + 0.01511070504784584, + 0.006431632675230503, + 0.01006693858653307, + 0.005420612171292305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.03375665098428726 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014213399030268192 + ], + "opposite_logprob_tokens": [ + -4.264213562011719 + ], + "logprob_reward": -0.014213399030268192, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014213399030268192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07140157371759415, + 0.06273522228002548, + 0.16594479978084564, + 0.07690894603729248, + 0.041515596210956573, + 0.30277904868125916, + 0.06620053946971893, + 0.050589557737112045, + 0.06083373725414276, + 0.10109101235866547 + ], + "credit_map": [ + 0.007725705858319998, + 0.0061617447063326836, + 0.011195925064384937, + 0.007210018113255501, + 0.018428368493914604, + 0.012705883011221886, + 0.004291547928005457, + 0.0036823784466832876, + 0.00689961202442646, + 0.008149435743689537, + 0.011255859397351742, + 0.007875354029238224, + 0.00806091632694006, + 0.007018235512077808, + 0.007309011649340391, + 0.0061667971312999725, + 0.022869877517223358, + 0.02480488456785679, + 0.022914176806807518, + 0.018718361854553223, + 0.017702611163258553, + 0.01216207630932331, + 0.021675949916243553, + 0.02509685978293419, + 0.006534448824822903, + 0.0076812393963336945, + 0.006268622819334269, + 0.007088848389685154, + 0.00773044116795063, + 0.02028465084731579, + 0.013375899754464626, + 0.007944794371724129, + 0.005800012033432722, + 0.002957958495244384, + 0.0033877065870910883, + 0.005631641484797001, + 0.0042547122575342655, + 0.007801359985023737, + 0.007544221356511116, + 0.004137985408306122, + 0.020264677703380585, + 0.014954856596887112, + 0.01176606398075819, + 0.014663463458418846, + 0.02745164930820465, + 0.11575370281934738, + 0.08162567764520645, + 0.01629892736673355, + 0.0066968221217393875, + 0.004412734415382147, + 0.011369245126843452, + 0.006610214244574308, + 0.010847680270671844, + 0.012065546587109566, + 0.0068390853703022, + 0.007359215058386326, + 0.005113877821713686, + 0.004276183433830738, + 0.006691575516015291, + 0.009240108542144299, + 0.0036101005971431732, + 0.007297983393073082, + 0.006492543499916792, + 0.007867186330258846, + 0.004854892380535603, + 0.006641542073339224, + 0.00611411826685071, + 0.006788190919905901, + 0.016988331452012062, + 0.004902557469904423, + 0.010409778915345669, + 0.004134326707571745, + 0.017698736861348152, + 0.013743311166763306, + 0.01378299668431282, + 0.013745395466685295, + 0.01288449577987194, + 0.005426115356385708, + 0.015297284349799156, + 0.008512677624821663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 9 + ], + "selected_reward": -0.014213399030268192 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811153024435043 + ], + "opposite_logprob_tokens": [ + -3.288111448287964 + ], + "logprob_reward": -0.03811153024435043, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811153024435043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09350384771823883, + 0.02569865807890892, + 0.028476746752858162, + 0.07398606836795807, + 0.04857265576720238, + 0.5719829797744751, + 0.031084654852747917, + 0.039242085069417953, + 0.033147700130939484, + 0.054304566234350204 + ], + "credit_map": [ + 0.009487579576671124, + 0.005776195786893368, + 0.02130451239645481, + 0.007933756336569786, + 0.02231733873486519, + 0.016999298706650734, + 0.004169439896941185, + 0.005515730939805508, + 0.0027950271032750607, + 0.003083402756601572, + 0.004163712728768587, + 0.003482184140011668, + 0.0026063290424644947, + 0.0035172777716070414, + 0.0037143900990486145, + 0.002336337696760893, + 0.003409296739846468, + 0.004752729553729296, + 0.002704440848901868, + 0.003202378749847412, + 0.0025226762518286705, + 0.0030517394188791513, + 0.004348215647041798, + 0.004485270474106073, + 0.0072953952476382256, + 0.008758305571973324, + 0.006256027612835169, + 0.007129895966500044, + 0.008087915368378162, + 0.018200306221842766, + 0.008886571042239666, + 0.009371661581099033, + 0.005848723463714123, + 0.0042434027418494225, + 0.0032448293641209602, + 0.007309237495064735, + 0.0036743597593158484, + 0.006827228702604771, + 0.009294749237596989, + 0.008130126632750034, + 0.040110547095537186, + 0.03338532894849777, + 0.0200900137424469, + 0.027355261147022247, + 0.038621023297309875, + 0.1961630880832672, + 0.1829477697610855, + 0.033309996128082275, + 0.002935607684776187, + 0.0025434948038309813, + 0.004994621966034174, + 0.003594956360757351, + 0.004975259304046631, + 0.004374930169433355, + 0.0035467324778437614, + 0.004119056276977062, + 0.005796434357762337, + 0.0044625564478337765, + 0.004588437266647816, + 0.005881176795810461, + 0.0029410351999104023, + 0.005490132141858339, + 0.003959967754781246, + 0.006122350227087736, + 0.0035254103131592274, + 0.0036829370073974133, + 0.006377360317856073, + 0.004240267910063267, + 0.0067316764034330845, + 0.0032849344424903393, + 0.0033831798937171698, + 0.001921936054714024, + 0.007545746397227049, + 0.0069558750838041306, + 0.005442265421152115, + 0.008261647075414658, + 0.008398559875786304, + 0.0024640655610710382, + 0.011379352770745754, + 0.003857057075947523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.03811153024435043 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective focusing on a male antagonist's trident horn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the horn exhibit a deep bronze metallic sheen with an oily iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely out-of-focus blur of green and brown patches of moss and bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the antagonist lunge forward causing the massive horn to rapidly enlarge in the field of view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a violent impact causing the image to blur and distort with a trembling field of vision?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260316655039787 + ], + "opposite_logprob_tokens": [ + -5.255260467529297 + ], + "logprob_reward": -0.005260316655039787, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260316655039787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12460518628358841, + 0.05270501971244812, + 0.06217651814222336, + 0.08361552655696869, + 0.04092295467853546, + 0.09837532788515091, + 0.028111286461353302, + 0.05996664986014366, + 0.31859979033470154, + 0.13092170655727386 + ], + "credit_map": [ + 0.01288218516856432, + 0.0077441055327653885, + 0.030217958614230156, + 0.01319971214979887, + 0.027156420052051544, + 0.02213628962635994, + 0.004933721851557493, + 0.006334794219583273, + 0.005954157095402479, + 0.0059992060996592045, + 0.007764630950987339, + 0.00723856408149004, + 0.007132803555577993, + 0.007175501435995102, + 0.006804217584431171, + 0.004635941237211227, + 0.0069825854152441025, + 0.007638475392013788, + 0.005129900295287371, + 0.007282149512320757, + 0.007063551340252161, + 0.006138669792562723, + 0.010788538493216038, + 0.01115264929831028, + 0.007838195189833641, + 0.010318230837583542, + 0.008929994888603687, + 0.009744024835526943, + 0.011961441487073898, + 0.012383099645376205, + 0.008457153104245663, + 0.013983389362692833, + 0.007521515246480703, + 0.003590170294046402, + 0.0027763627003878355, + 0.0052469391375780106, + 0.004007944371551275, + 0.0060609811916947365, + 0.0070207491517066956, + 0.004698289558291435, + 0.007943513803184032, + 0.008816750720143318, + 0.006972076371312141, + 0.009150742553174496, + 0.031744472682476044, + 0.010643640533089638, + 0.015243211761116982, + 0.00786091573536396, + 0.0027605921495705843, + 0.00290266377851367, + 0.003418212989345193, + 0.004987965803593397, + 0.0038787941448390484, + 0.002987198531627655, + 0.003697275882586837, + 0.00347858271561563, + 0.010146082378923893, + 0.00735057657584548, + 0.00728229945525527, + 0.008049030788242817, + 0.004404847975820303, + 0.0072897570207715034, + 0.006089869886636734, + 0.009354189969599247, + 0.03371407091617584, + 0.0355648510158062, + 0.06260917335748672, + 0.0367785282433033, + 0.07863396406173706, + 0.02732273004949093, + 0.024858273565769196, + 0.019118214026093483, + 0.01933138631284237, + 0.016574881970882416, + 0.012959188781678677, + 0.01941520906984806, + 0.01813371852040291, + 0.00666308868676424, + 0.02857891656458378, + 0.009265326894819736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.005260316655039787 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_0/index_tensor([9458, 9458])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_0/index_tensor([9458, 9458])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..356d1d58255ce862b458ad07f25edfa6b9d537f1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_0/index_tensor([9458, 9458])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a static, slightly high-angle medium shot, the camera tilts steadily downward, meticulously tracking the movement of a rectangular, translucent, frosted plastic food container, perched precariously on the edge of a polished, light-gray crystal countertop in a modern kitchen. Soft, realistic afternoon light streams through a large, off-camera window, casting diffuse highlights on the container's vibrant blue snap-on lid and creating long, soft shadows. The container teeters for a moment before quickly and decisively toppling over its edge, tumbling through the air. The camera follows its descent until it impacts the vast, off-white tile floor with a dull, hollow thud, its flexible plastic body visibly deforming for a moment after the impact. It then executes a slight bounce, skimming a few inches across the smooth surface before swaying to a completely abrupt sideways rest, concluding the dynamic movement with a quiet sense of finality in the quiet, sunlit kitchen.", + "seed": 866, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_0/index_9458_seed_866_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.00878858957439661, + "vlm_logprob": -0.00878858957439661, + "vlm_yes_no_margin": 5.049999952316284, + "vlm_token_credit": -0.00878858957439661, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022810103371739388 + ], + "opposite_logprob_tokens": [ + -4.022809982299805 + ], + "logprob_reward": -0.022810103371739388, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.022810103371739388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15272338688373566, + 0.03526812419295311, + 0.07406008988618851, + 0.1580904871225357, + 0.03581612557172775, + 0.02747878059744835, + 0.06981462985277176, + 0.13690778613090515, + 0.23222775757312775, + 0.07761282473802567 + ], + "credit_map": [ + 0.017573103308677673, + 0.019267218187451363, + 0.02903883345425129, + 0.02820548228919506, + 0.013542619533836842, + 0.01443127915263176, + 0.013148324564099312, + 0.017516540363430977, + 0.005059183109551668, + 0.004441388417035341, + 0.0066875312477350235, + 0.00449418555945158, + 0.003708431264385581, + 0.00271706935018301, + 0.0032779055181890726, + 0.004882427863776684, + 0.011053487658500671, + 0.006551485974341631, + 0.014173072762787342, + 0.013899425975978374, + 0.007434721104800701, + 0.007650592364370823, + 0.005258793011307716, + 0.008038511499762535, + 0.018089277669787407, + 0.02051626890897751, + 0.02218003012239933, + 0.015191406942903996, + 0.027146898210048676, + 0.01267062034457922, + 0.028274482116103172, + 0.014021508395671844, + 0.004476877395063639, + 0.004670663736760616, + 0.005899895913898945, + 0.004602305591106415, + 0.0030572048854082823, + 0.003233517985790968, + 0.0037342787254601717, + 0.0061413804069161415, + 0.0025500657502561808, + 0.004742453806102276, + 0.002976677380502224, + 0.0038447913248091936, + 0.0028381063602864742, + 0.0038139193784445524, + 0.002562924986705184, + 0.004149839747697115, + 0.008121143095195293, + 0.0051788585260510445, + 0.013903756625950336, + 0.005338448565453291, + 0.005201396532356739, + 0.005300560500472784, + 0.006302746012806892, + 0.020467719063162804, + 0.014095944352447987, + 0.010783207602798939, + 0.016166502609848976, + 0.014254855923354626, + 0.01927925832569599, + 0.011409960687160492, + 0.022370081394910812, + 0.028547978028655052, + 0.023077799007296562, + 0.03703939914703369, + 0.023094626143574715, + 0.031078416854143143, + 0.01768236607313156, + 0.04704751819372177, + 0.02957671321928501, + 0.02363089844584465, + 0.006424023304134607, + 0.004447516985237598, + 0.02201530709862709, + 0.0052307648584246635, + 0.006385054439306259, + 0.004540249239653349, + 0.005696685519069433, + 0.022873226553201675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.022810103371739388 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005266720429062843 + ], + "opposite_logprob_tokens": [ + -5.2552666664123535 + ], + "logprob_reward": -0.005266720429062843, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005266720429062843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07728539407253265, + 0.04313568398356438, + 0.03016149438917637, + 0.09912901371717453, + 0.0312502384185791, + 0.01667039468884468, + 0.4537948966026306, + 0.05502382293343544, + 0.1652141660451889, + 0.0283348485827446 + ], + "credit_map": [ + 0.006334489211440086, + 0.012705621309578419, + 0.007979475893080235, + 0.01253044418990612, + 0.0076879714615643024, + 0.0061470880173146725, + 0.013302339240908623, + 0.01059796754270792, + 0.0038465603720396757, + 0.0074910251423716545, + 0.006962731014937162, + 0.006377015262842178, + 0.003738399827852845, + 0.003653081366792321, + 0.005084255710244179, + 0.005982614122331142, + 0.0034179408103227615, + 0.004066505003720522, + 0.00350457732565701, + 0.004543956834822893, + 0.0038492977619171143, + 0.004074167925864458, + 0.0036142850294709206, + 0.0030907660257071257, + 0.012669791467487812, + 0.01527112815529108, + 0.011272314935922623, + 0.010534091852605343, + 0.01703765243291855, + 0.009758230298757553, + 0.013977293856441975, + 0.008608502335846424, + 0.003395186737179756, + 0.005070395767688751, + 0.003825777443125844, + 0.0044707986526191235, + 0.0030564097687602043, + 0.0025058581959456205, + 0.0038694071117788553, + 0.005056407302618027, + 0.001714076497592032, + 0.002615781966596842, + 0.0021085923071950674, + 0.002514052204787731, + 0.0019795324187725782, + 0.001974978484213352, + 0.0014674391131848097, + 0.002295940648764372, + 0.05402828007936478, + 0.01125952135771513, + 0.1233581081032753, + 0.026675572618842125, + 0.022233206778764725, + 0.03675999119877815, + 0.014867270365357399, + 0.16461291909217834, + 0.005215401761233807, + 0.006460922304540873, + 0.006500863935798407, + 0.007529185153543949, + 0.008631275966763496, + 0.003673590486869216, + 0.00936372485011816, + 0.007648858241736889, + 0.017382875084877014, + 0.028669605031609535, + 0.016341429203748703, + 0.02491898089647293, + 0.013961653225123882, + 0.026277273893356323, + 0.02098751813173294, + 0.016674824059009552, + 0.002418170915916562, + 0.005211446434259415, + 0.0031096122693270445, + 0.006850769743323326, + 0.002259249100461602, + 0.0014571594074368477, + 0.0030819966923445463, + 0.003946445416659117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.005266720429062843 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036187181249260902 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036187181249260902, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036187181249260902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07289671152830124, + 0.027561688795685768, + 0.031018953770399094, + 0.2230825126171112, + 0.06228474900126457, + 0.042515430599451065, + 0.3401859402656555, + 0.09118524938821793, + 0.06849147379398346, + 0.040777236223220825 + ], + "credit_map": [ + 0.009405265562236309, + 0.008235503919422626, + 0.011647442355751991, + 0.00948454812169075, + 0.007227505557239056, + 0.006687890738248825, + 0.010167738422751427, + 0.010040822438895702, + 0.0037747921887785196, + 0.00403866171836853, + 0.0037667066790163517, + 0.002535758074373007, + 0.002731691114604473, + 0.002708851359784603, + 0.004213397856801748, + 0.0037918316666036844, + 0.004756182432174683, + 0.004105685278773308, + 0.004647961817681789, + 0.0036068519111722708, + 0.003528300207108259, + 0.0034228020813316107, + 0.004025805275887251, + 0.002925368258729577, + 0.031093589961528778, + 0.027399735525250435, + 0.03130294010043144, + 0.023226357996463776, + 0.03697071596980095, + 0.019034534692764282, + 0.03468569368124008, + 0.019368967041373253, + 0.009338447824120522, + 0.0084479209035635, + 0.00643765926361084, + 0.0068790665827691555, + 0.006852158345282078, + 0.006449726410210133, + 0.006464669480919838, + 0.011415107175707817, + 0.0038313379045575857, + 0.007412259001284838, + 0.004677142016589642, + 0.006190339103341103, + 0.005999051034450531, + 0.004341489169746637, + 0.004807129502296448, + 0.005256687756627798, + 0.03821948915719986, + 0.013290743343532085, + 0.08994930237531662, + 0.020165564492344856, + 0.018352007493376732, + 0.025642190128564835, + 0.018088320270180702, + 0.116478331387043, + 0.01076012384146452, + 0.007323113270103931, + 0.01352451741695404, + 0.010284445248544216, + 0.00935217086225748, + 0.00657349219545722, + 0.012535729445517063, + 0.020831672474741936, + 0.003571274457499385, + 0.0056717414408922195, + 0.021078206598758698, + 0.006771269720047712, + 0.003379284171387553, + 0.003443062538281083, + 0.004814815241843462, + 0.01976182870566845, + 0.003763616783544421, + 0.006910775788128376, + 0.007009593304246664, + 0.006905363406985998, + 0.00359101640060544, + 0.0028230794705450535, + 0.0034094369038939476, + 0.006364359520375729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.0036187181249260902 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036153923720121384 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.0036153923720121384, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036153923720121384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06342491507530212, + 0.03420837223529816, + 0.04116544872522354, + 0.11685396730899811, + 0.09085345268249512, + 0.10477244853973389, + 0.21215035021305084, + 0.18178725242614746, + 0.11997254192829132, + 0.03481127694249153 + ], + "credit_map": [ + 0.008310849778354168, + 0.006275880616158247, + 0.01365079265087843, + 0.010136340744793415, + 0.005305723752826452, + 0.005784810520708561, + 0.006704647559672594, + 0.007255867589265108, + 0.0040666572749614716, + 0.0034476760774850845, + 0.006185098085552454, + 0.005266414489597082, + 0.0038551578763872385, + 0.003971234429627657, + 0.0035403608344495296, + 0.0038757759612053633, + 0.005822085775434971, + 0.005084990058094263, + 0.005563372280448675, + 0.005784683860838413, + 0.0052411905489861965, + 0.0050465697422623634, + 0.0042841993272304535, + 0.004338358528912067, + 0.01089969277381897, + 0.011468715034425259, + 0.016960246488451958, + 0.014671782962977886, + 0.013393708504736423, + 0.013303368352353573, + 0.024123936891555786, + 0.012032519094645977, + 0.008873465470969677, + 0.013052085414528847, + 0.010406365618109703, + 0.008620267733931541, + 0.011181294918060303, + 0.011422283947467804, + 0.010507866740226746, + 0.016789820045232773, + 0.011208501644432545, + 0.01618957333266735, + 0.012747204862535, + 0.01653253845870495, + 0.012744856998324394, + 0.012692623771727085, + 0.009659374132752419, + 0.012997771613299847, + 0.02384728379547596, + 0.011750331148505211, + 0.05326785147190094, + 0.0152153754606843, + 0.014418618753552437, + 0.014682626351714134, + 0.01655474863946438, + 0.0624135285615921, + 0.013377880677580833, + 0.015470889396965504, + 0.030446967110037804, + 0.016560981050133705, + 0.03238620236515999, + 0.010543222539126873, + 0.040613628923892975, + 0.022387484088540077, + 0.008525477722287178, + 0.010871732607483864, + 0.02980753965675831, + 0.009195154532790184, + 0.006048119626939297, + 0.015416432172060013, + 0.008562551811337471, + 0.03154553845524788, + 0.0035804761573672295, + 0.005113475024700165, + 0.005775854457169771, + 0.005638450384140015, + 0.003296865848824382, + 0.002533484250307083, + 0.003284708596765995, + 0.005587962921708822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.0036153923720121384 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008632013574242592 + ], + "opposite_logprob_tokens": [ + -4.758632183074951 + ], + "logprob_reward": -0.008632013574242592, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008632013574242592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06398366391658783, + 0.02881869673728943, + 0.022205276414752007, + 0.16347144544124603, + 0.0376586951315403, + 0.05072250962257385, + 0.2947080433368683, + 0.11930004507303238, + 0.06648670881986618, + 0.15264494717121124 + ], + "credit_map": [ + 0.008237047120928764, + 0.008916155435144901, + 0.010818623006343842, + 0.008007260039448738, + 0.005806458182632923, + 0.006640653125941753, + 0.007955590263009071, + 0.007601871620863676, + 0.004699323792010546, + 0.004996431525796652, + 0.003795443568378687, + 0.0026046778075397015, + 0.0028399128932505846, + 0.002748616971075535, + 0.0033854320645332336, + 0.003748859278857708, + 0.0028040807228535414, + 0.0028499236796051264, + 0.0035809585824608803, + 0.002839533379301429, + 0.0029711644165217876, + 0.0026712084654718637, + 0.0018434348748996854, + 0.002644971711561084, + 0.019028056412935257, + 0.022542502731084824, + 0.019965507090091705, + 0.01933632418513298, + 0.023398645222187042, + 0.021657081320881844, + 0.022922663018107414, + 0.01462064404040575, + 0.004502214957028627, + 0.004695115610957146, + 0.003992840647697449, + 0.003971471916884184, + 0.0060233729891479015, + 0.00281222234480083, + 0.005719199310988188, + 0.005942258518189192, + 0.0039592040702700615, + 0.008962034247815609, + 0.0062649971805512905, + 0.007820328697562218, + 0.005498077254742384, + 0.006942471023648977, + 0.004891567397862673, + 0.006383827887475491, + 0.03297169879078865, + 0.011017524637281895, + 0.07465484738349915, + 0.018554838374257088, + 0.016231708228588104, + 0.023848896846175194, + 0.01734771393239498, + 0.10008081793785095, + 0.011336369439959526, + 0.010825571604073048, + 0.016841476783156395, + 0.013461845926940441, + 0.019030462950468063, + 0.006277079693973064, + 0.02665902115404606, + 0.01486822310835123, + 0.005001675803214312, + 0.0078049604780972, + 0.015006356872618198, + 0.007236057426780462, + 0.005135184153914452, + 0.00586047163233161, + 0.006020940840244293, + 0.01442105881869793, + 0.010360326617956161, + 0.023675082251429558, + 0.023622963577508926, + 0.03190584480762482, + 0.011043570935726166, + 0.006473197136074305, + 0.015631448477506638, + 0.02993251197040081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -0.008632013574242592 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a static, slightly high-angle medium shot, the camera tilts steadily downward, meticulously tracking the movement of a rectangular, translucent, frosted plastic food container, perched precariously on the edge of a polished, light-gray crystal countertop in a modern kitchen. Soft, realistic afternoon light streams through a large, off-camera window, casting diffuse highlights on the container's vibrant blue snap-on lid and creating long, soft shadows. The container teeters for a moment before quickly and decisively toppling over its edge, tumbling through the air. The camera follows its descent until it impacts the vast, off-white tile floor with a dull, hollow thud, its flexible plastic body visibly deforming for a moment after the impact. It then executes a slight bounce, skimming a few inches across the smooth surface before swaying to a completely abrupt sideways rest, concluding the dynamic movement with a quiet sense of finality in the quiet, sunlit kitchen.", + "seed": 867, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_0/index_9458_seed_867_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.010321538243442775, + "vlm_logprob": -0.010321538243442775, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.010321538243442775, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016403697431087494 + ], + "opposite_logprob_tokens": [ + -4.266403675079346 + ], + "logprob_reward": -0.016403697431087494, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016403697431087494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27939409017562866, + 0.05592009052634239, + 0.08594756573438644, + 0.24574892222881317, + 0.060328610241413116, + 0.06040679290890694, + 0.11341242492198944, + 0.0445454902946949, + 0.03539663180708885, + 0.018899334594607353 + ], + "credit_map": [ + 0.039802517741918564, + 0.026183290407061577, + 0.018185127526521683, + 0.03327188640832901, + 0.039768531918525696, + 0.033777981996536255, + 0.049181777983903885, + 0.03922298550605774, + 0.006773245055228472, + 0.008305960334837437, + 0.0035528226289898157, + 0.004907157272100449, + 0.008049268275499344, + 0.008543307892978191, + 0.00784567091614008, + 0.007942658849060535, + 0.00979789812117815, + 0.010020345449447632, + 0.0098720146343112, + 0.010229001753032207, + 0.012512887828052044, + 0.009035385213792324, + 0.011623969301581383, + 0.01285606250166893, + 0.06797477602958679, + 0.014533068984746933, + 0.019147858023643494, + 0.012960771098732948, + 0.010557161644101143, + 0.017795301973819733, + 0.017443951219320297, + 0.08533604443073273, + 0.0060378448106348515, + 0.008527852594852448, + 0.005570290610194206, + 0.008078886196017265, + 0.0103681068867445, + 0.007186160422861576, + 0.007659962400794029, + 0.006899507250636816, + 0.008252964355051517, + 0.006038082297891378, + 0.007672387640923262, + 0.009134680032730103, + 0.0046812789514660835, + 0.010288489051163197, + 0.006638376973569393, + 0.007700530346482992, + 0.011528941802680492, + 0.009066218510270119, + 0.016132794320583344, + 0.015462487004697323, + 0.00947972945868969, + 0.017582673579454422, + 0.01099182479083538, + 0.023167762905359268, + 0.0052443090826272964, + 0.00437622144818306, + 0.006097383797168732, + 0.005617092829197645, + 0.003547082422301173, + 0.005350109655410051, + 0.005491698160767555, + 0.008821597322821617, + 0.004698730073869228, + 0.003912987653166056, + 0.005797030404210091, + 0.0032713408581912518, + 0.002882600761950016, + 0.004519710782915354, + 0.0031044718343764544, + 0.007209757808595896, + 0.002092754002660513, + 0.0016518750926479697, + 0.0029722952749580145, + 0.0016754112439230084, + 0.0018880132120102644, + 0.0030687132384628057, + 0.00218014745041728, + 0.0033701255451887846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.016403697431087494 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205760322511196 + ], + "opposite_logprob_tokens": [ + -4.2642059326171875 + ], + "logprob_reward": -0.014205760322511196, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205760322511196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07660362869501114, + 0.021719321608543396, + 0.027087992057204247, + 0.3143409788608551, + 0.07332155108451843, + 0.16087201237678528, + 0.14881841838359833, + 0.024190576747059822, + 0.07494214177131653, + 0.07810331881046295 + ], + "credit_map": [ + 0.009623355232179165, + 0.007014554459601641, + 0.0067603145726025105, + 0.011285707354545593, + 0.007240633014589548, + 0.009380578063428402, + 0.014817400835454464, + 0.010481089353561401, + 0.0025121953804045916, + 0.002924758242443204, + 0.0012170769041404128, + 0.0017751003615558147, + 0.0037578665651381016, + 0.0039125289767980576, + 0.002479295711964369, + 0.0031404977198690176, + 0.0031629910226911306, + 0.0025057424791157246, + 0.0036763183306902647, + 0.004950952250510454, + 0.002409766661003232, + 0.0021693643648177385, + 0.004711815621703863, + 0.003501039231196046, + 0.10591547191143036, + 0.01595343090593815, + 0.016033699735999107, + 0.009602665901184082, + 0.007608076557517052, + 0.01629435457289219, + 0.01569409854710102, + 0.1272391974925995, + 0.004533969797194004, + 0.009768052957952023, + 0.006862683687359095, + 0.006521919742226601, + 0.017198145389556885, + 0.010784933343529701, + 0.00957310851663351, + 0.008078732527792454, + 0.017401304095983505, + 0.018332602456212044, + 0.02254682593047619, + 0.0228472538292408, + 0.011126826517283916, + 0.02390897646546364, + 0.020979145541787148, + 0.02372908405959606, + 0.014435533434152603, + 0.021088024601340294, + 0.021727725863456726, + 0.022527426481246948, + 0.011854089796543121, + 0.014089319854974747, + 0.020108234137296677, + 0.022988062351942062, + 0.0026894991751760244, + 0.0022716661915183067, + 0.003894738620147109, + 0.0029899454675614834, + 0.001881038653664291, + 0.0036381350364536047, + 0.0023534612264484167, + 0.004472093190997839, + 0.009002186357975006, + 0.00936861988157034, + 0.010141602717339993, + 0.01181444525718689, + 0.006760746240615845, + 0.008395103737711906, + 0.008822067640721798, + 0.010637378320097923, + 0.010895797051489353, + 0.007594849448651075, + 0.011212152428925037, + 0.012127099558711052, + 0.007317014504224062, + 0.007865436375141144, + 0.008954798802733421, + 0.01213617529720068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.014205760322511196 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003617767943069339 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.003617767943069339, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003617767943069339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16873086988925934, + 0.05725701525807381, + 0.1978919357061386, + 0.11884614080190659, + 0.0881665050983429, + 0.14099644124507904, + 0.11176805943250656, + 0.04611564427614212, + 0.043299585580825806, + 0.026927804574370384 + ], + "credit_map": [ + 0.023276377469301224, + 0.017517615109682083, + 0.014657186344265938, + 0.019414477050304413, + 0.0173159372061491, + 0.021079139783978462, + 0.029299218207597733, + 0.026170913130044937, + 0.005958127323538065, + 0.008636362850666046, + 0.0037649485748261213, + 0.004938196390867233, + 0.006776262540370226, + 0.012515299022197723, + 0.006560840178281069, + 0.008106978610157967, + 0.020030267536640167, + 0.020468182861804962, + 0.024065209552645683, + 0.02891322784125805, + 0.0254603810608387, + 0.020636821165680885, + 0.02880883403122425, + 0.029509004205465317, + 0.016217857599258423, + 0.014660011045634747, + 0.015321278013288975, + 0.012964940629899502, + 0.014200140722095966, + 0.013632401823997498, + 0.015518599189817905, + 0.016330908983945847, + 0.0070833470672369, + 0.011332998983561993, + 0.008351081050932407, + 0.008977063931524754, + 0.01775275729596615, + 0.012794499285519123, + 0.012722079642117023, + 0.009152675978839397, + 0.01644512452185154, + 0.016859881579875946, + 0.01927820034325123, + 0.015563090331852436, + 0.010403480380773544, + 0.021782001480460167, + 0.014838051982223988, + 0.02582661062479019, + 0.012212652713060379, + 0.006024983711540699, + 0.019311485812067986, + 0.00945298932492733, + 0.0073676155880093575, + 0.02645079605281353, + 0.01023702323436737, + 0.02071050927042961, + 0.005098376423120499, + 0.003871530992910266, + 0.007474840153008699, + 0.005253365263342857, + 0.0037627548445016146, + 0.005840019788593054, + 0.007352283224463463, + 0.00746247498318553, + 0.004644328262656927, + 0.004377592355012894, + 0.006649403367191553, + 0.005308676976710558, + 0.004247639793902636, + 0.004634445998817682, + 0.005589391104876995, + 0.007848110049962997, + 0.0037084908690303564, + 0.0021128361113369465, + 0.0043604630045592785, + 0.0033779204823076725, + 0.002913476899266243, + 0.002629390452057123, + 0.0033755716867744923, + 0.00444965623319149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.003617767943069339 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319268717430532 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.00319268717430532, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319268717430532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1450221836566925, + 0.04420926049351692, + 0.10353414714336395, + 0.4505296051502228, + 0.07743603736162186, + 0.08492590487003326, + 0.0325896330177784, + 0.017858734354376793, + 0.022075096145272255, + 0.02181941829621792 + ], + "credit_map": [ + 0.01935218833386898, + 0.013975424692034721, + 0.008657075464725494, + 0.01481066644191742, + 0.022966517135500908, + 0.018668534234166145, + 0.025853898376226425, + 0.020737886428833008, + 0.004659351892769337, + 0.009549562819302082, + 0.003670964390039444, + 0.004591749515384436, + 0.004494914785027504, + 0.005724059883505106, + 0.006041006650775671, + 0.005477650556713343, + 0.011162443086504936, + 0.011671246960759163, + 0.010732367634773254, + 0.012067764066159725, + 0.013486112467944622, + 0.009842444211244583, + 0.018109198659658432, + 0.016462573781609535, + 0.14806696772575378, + 0.018468016758561134, + 0.027480673044919968, + 0.012658755294978619, + 0.012378365732729435, + 0.025898031890392303, + 0.022738443687558174, + 0.18284034729003906, + 0.006082463543862104, + 0.010155699215829372, + 0.009181362576782703, + 0.011893294751644135, + 0.010611315257847309, + 0.00811812561005354, + 0.01301245391368866, + 0.008381325751543045, + 0.01103342417627573, + 0.009551591239869595, + 0.011505764909088612, + 0.012546221725642681, + 0.005776319187134504, + 0.012341785244643688, + 0.010159310884773731, + 0.012011488899588585, + 0.0033428643364459276, + 0.002812268678098917, + 0.004961109720170498, + 0.004393171053379774, + 0.0026880456134676933, + 0.004655692260712385, + 0.00325339799746871, + 0.006483084987848997, + 0.002709061373025179, + 0.0015122713521122932, + 0.002672765636816621, + 0.0018103303154930472, + 0.0016007027588784695, + 0.0027469703927636147, + 0.0019487631507217884, + 0.0028578699566423893, + 0.0027516039554029703, + 0.0018434574594721198, + 0.003447365714237094, + 0.0022745144087821245, + 0.0025806690100580454, + 0.002911155577749014, + 0.002457205904647708, + 0.00380912353284657, + 0.0036414784844964743, + 0.0016737411497160792, + 0.0035523581318557262, + 0.002166156191378832, + 0.0019799398723989725, + 0.002367111388593912, + 0.0024829362519085407, + 0.00395569670945406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.00319268717430532 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418777834624052 + ], + "opposite_logprob_tokens": [ + -4.264187812805176 + ], + "logprob_reward": -0.01418777834624052, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418777834624052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0999746322631836, + 0.05791402980685234, + 0.13138216733932495, + 0.1277655065059662, + 0.13883064687252045, + 0.17177420854568481, + 0.10645978152751923, + 0.07000049948692322, + 0.04667738452553749, + 0.049221161752939224 + ], + "credit_map": [ + 0.011659357696771622, + 0.009981416165828705, + 0.007658770773559809, + 0.013120085932314396, + 0.014358669519424438, + 0.012095496989786625, + 0.017657505348324776, + 0.01344333030283451, + 0.007507525850087404, + 0.009753781370818615, + 0.0032289139926433563, + 0.004173321183770895, + 0.007114567328244448, + 0.01175847090780735, + 0.006065055727958679, + 0.008312392979860306, + 0.015504978597164154, + 0.016450928524136543, + 0.01436108723282814, + 0.013336766511201859, + 0.019431032240390778, + 0.017436634749174118, + 0.017612246796488762, + 0.017248494550585747, + 0.015752168372273445, + 0.011050472967326641, + 0.019731542095541954, + 0.014930704608559608, + 0.01248909067362547, + 0.01587262563407421, + 0.01885119453072548, + 0.019087711349129677, + 0.009463371708989143, + 0.02070751041173935, + 0.015324864536523819, + 0.018569691106677055, + 0.02412082999944687, + 0.015270099975168705, + 0.02368110790848732, + 0.01169317215681076, + 0.032083094120025635, + 0.0175812728703022, + 0.025023087859153748, + 0.01792525127530098, + 0.01232558861374855, + 0.021224701777100563, + 0.011982300318777561, + 0.0336289145052433, + 0.009792538359761238, + 0.010677061043679714, + 0.014325355179607868, + 0.017028555274009705, + 0.011435450054705143, + 0.013439753092825413, + 0.012067561037838459, + 0.01769350841641426, + 0.010105673223733902, + 0.005144056398421526, + 0.011911085806787014, + 0.007802399341017008, + 0.005120276473462582, + 0.00779691943898797, + 0.007024394813925028, + 0.015095698647201061, + 0.006078631151467562, + 0.003739945124834776, + 0.007253564428538084, + 0.004844904411584139, + 0.003964199684560299, + 0.00687564630061388, + 0.005619165953248739, + 0.0083013279363513, + 0.0074333567172288895, + 0.004541077185422182, + 0.007988178171217442, + 0.0047264862805604935, + 0.004312037490308285, + 0.006269290577620268, + 0.005287976935505867, + 0.008662759326398373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.01418777834624052 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a static, slightly high-angle medium shot, the camera tilts steadily downward, meticulously tracking the movement of a rectangular, translucent, frosted plastic food container, perched precariously on the edge of a polished, light-gray crystal countertop in a modern kitchen. Soft, realistic afternoon light streams through a large, off-camera window, casting diffuse highlights on the container's vibrant blue snap-on lid and creating long, soft shadows. The container teeters for a moment before quickly and decisively toppling over its edge, tumbling through the air. The camera follows its descent until it impacts the vast, off-white tile floor with a dull, hollow thud, its flexible plastic body visibly deforming for a moment after the impact. It then executes a slight bounce, skimming a few inches across the smooth surface before swaying to a completely abrupt sideways rest, concluding the dynamic movement with a quiet sense of finality in the quiet, sunlit kitchen.", + "seed": 868, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_0/index_9458_seed_868_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01258119815029204, + "vlm_logprob": -0.01258119815029204, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.01258119815029204, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022004852071404457 + ], + "opposite_logprob_tokens": [ + -4.147005081176758 + ], + "logprob_reward": -0.022004852071404457, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.022004852071404457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1764996498823166, + 0.0426434725522995, + 0.023084240034222603, + 0.028565995395183563, + 0.08146683871746063, + 0.12580794095993042, + 0.2618485689163208, + 0.1464197337627411, + 0.07420814037322998, + 0.03945544362068176 + ], + "credit_map": [ + 0.025985147804021835, + 0.0149448337033391, + 0.03038797527551651, + 0.016874684020876884, + 0.018910711631178856, + 0.027884701266884804, + 0.012217564508318901, + 0.02929401770234108, + 0.006582765374332666, + 0.00302589638158679, + 0.005341492593288422, + 0.003968867007642984, + 0.004321855492889881, + 0.007820474915206432, + 0.0032208668999373913, + 0.00836125761270523, + 0.0029026244301348925, + 0.002695623319596052, + 0.00435267761349678, + 0.002460342599079013, + 0.0022969467099756002, + 0.003091225866228342, + 0.002587795490399003, + 0.0026970047038048506, + 0.004139854107052088, + 0.003553205169737339, + 0.004668411798775196, + 0.002120690420269966, + 0.002831805497407913, + 0.004470450337976217, + 0.0019252535421401262, + 0.0048563252203166485, + 0.015369314700365067, + 0.0076323747634887695, + 0.013385603204369545, + 0.006600989028811455, + 0.005014942027628422, + 0.011716046370565891, + 0.006761441472917795, + 0.014986122027039528, + 0.014757227152585983, + 0.019719554111361504, + 0.020180590450763702, + 0.014460520818829536, + 0.01422423217445612, + 0.009221721440553665, + 0.019021807238459587, + 0.014222287572920322, + 0.014441085048019886, + 0.013292931951582432, + 0.07713033258914948, + 0.04270884022116661, + 0.012124844826757908, + 0.025307197123765945, + 0.01538080908358097, + 0.06146250665187836, + 0.014087016694247723, + 0.03251629322767258, + 0.01834690012037754, + 0.020428432151675224, + 0.01839473843574524, + 0.012892799451947212, + 0.016091324388980865, + 0.013662226498126984, + 0.009818746708333492, + 0.006417032331228256, + 0.007774564903229475, + 0.010322061367332935, + 0.006368078291416168, + 0.009564495645463467, + 0.011789435520768166, + 0.01215373445302248, + 0.006853321101516485, + 0.00757295498624444, + 0.005105158779770136, + 0.0037359537091106176, + 0.0031915174331516027, + 0.004896477330476046, + 0.003312142798677087, + 0.004787919111549854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.022004852071404457 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004649304784834385 + ], + "opposite_logprob_tokens": [ + -5.3796491622924805 + ], + "logprob_reward": -0.004649304784834385, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004649304784834385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.130968376994133, + 0.040477294474840164, + 0.033059876412153244, + 0.07007239758968353, + 0.20260195434093475, + 0.0937424898147583, + 0.24629120528697968, + 0.072243332862854, + 0.0777411237359047, + 0.03280191123485565 + ], + "credit_map": [ + 0.018103448674082756, + 0.011795700527727604, + 0.012404454872012138, + 0.01594708114862442, + 0.020310044288635254, + 0.017952637746930122, + 0.017733000218868256, + 0.01672201417386532, + 0.0043104118667542934, + 0.00459959777072072, + 0.0031652941834181547, + 0.0033780208323150873, + 0.008338505402207375, + 0.005291426088660955, + 0.0059902663342654705, + 0.005403770133852959, + 0.003936317283660173, + 0.004256970714777708, + 0.0028721338603645563, + 0.002965759253129363, + 0.005745794158428907, + 0.0043833269737660885, + 0.004957767203450203, + 0.00394180603325367, + 0.007775521371513605, + 0.007761477027088404, + 0.004390563815832138, + 0.005580737255513668, + 0.018277158960700035, + 0.007667125202715397, + 0.011542011983692646, + 0.007077805232256651, + 0.029192514717578888, + 0.018703998997807503, + 0.029627447947859764, + 0.02705524116754532, + 0.022024385631084442, + 0.024514077231287956, + 0.026872899383306503, + 0.024611394852399826, + 0.01131140161305666, + 0.011932753957808018, + 0.013995707035064697, + 0.012824492529034615, + 0.010663498193025589, + 0.0075075640343129635, + 0.013225728645920753, + 0.012281334958970547, + 0.012732503935694695, + 0.01351333037018776, + 0.0693655014038086, + 0.03873010724782944, + 0.013035302050411701, + 0.025132209062576294, + 0.017092743888497353, + 0.05668951943516731, + 0.007825706154108047, + 0.010365819558501244, + 0.008413050323724747, + 0.013984612189233303, + 0.007602801080793142, + 0.006155737210065126, + 0.011138690635561943, + 0.006756916176527739, + 0.011282479390501976, + 0.008305476978421211, + 0.008726118132472038, + 0.013286655768752098, + 0.007188105024397373, + 0.006297134328633547, + 0.014994721859693527, + 0.007660428527742624, + 0.004639439284801483, + 0.005983839277178049, + 0.004737159702926874, + 0.004591281525790691, + 0.0022690333425998688, + 0.0027077477425336838, + 0.004251227714121342, + 0.003622181946411729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.004649304784834385 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978483073413372 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.00978483073413372, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978483073413372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06871852278709412, + 0.025750381872057915, + 0.017470048740506172, + 0.027566121891140938, + 0.12803450226783752, + 0.07928891479969025, + 0.3693827688694, + 0.1381402313709259, + 0.10644631832838058, + 0.03920214623212814 + ], + "credit_map": [ + 0.00914807990193367, + 0.008972532115876675, + 0.008031010627746582, + 0.00981491431593895, + 0.010581078939139843, + 0.006871783174574375, + 0.007233227603137493, + 0.00806589424610138, + 0.0028341617435216904, + 0.00238714343868196, + 0.004204802215099335, + 0.0022769763600081205, + 0.004311231896281242, + 0.0028705913573503494, + 0.0038404217921197414, + 0.003025055630132556, + 0.0025765339378267527, + 0.0016400718595832586, + 0.0024995673447847366, + 0.0014921565307304263, + 0.0031290522310882807, + 0.0019362003076821566, + 0.0023601900320500135, + 0.00183627859223634, + 0.00392752094194293, + 0.0030626251827925444, + 0.0018318359507247806, + 0.0027529646176844835, + 0.0062255351804196835, + 0.0029287354554980993, + 0.004157562740147114, + 0.0026793440338224173, + 0.019027506932616234, + 0.013870909810066223, + 0.019210420548915863, + 0.01578764244914055, + 0.012611151672899723, + 0.015994425863027573, + 0.015503120608627796, + 0.01602933369576931, + 0.008547339588403702, + 0.010644773952662945, + 0.010606128722429276, + 0.012644140981137753, + 0.009001540020108223, + 0.007054869085550308, + 0.012738130986690521, + 0.008051995187997818, + 0.019152402877807617, + 0.01687522977590561, + 0.11456096917390823, + 0.06489928066730499, + 0.017014378681778908, + 0.027570180594921112, + 0.019641606137156487, + 0.08966875821352005, + 0.013178745284676552, + 0.022849854081869125, + 0.01397621724754572, + 0.024040862917900085, + 0.013566563837230206, + 0.01511340495198965, + 0.01700897328555584, + 0.018405620008707047, + 0.01374618336558342, + 0.018010172992944717, + 0.01377582736313343, + 0.015015984885394573, + 0.0090035330504179, + 0.01060541719198227, + 0.01626848615705967, + 0.010020723566412926, + 0.005651912186294794, + 0.005334027111530304, + 0.004935261327773333, + 0.005936844274401665, + 0.0029899533838033676, + 0.0038897995837032795, + 0.0056848106905817986, + 0.004779540468007326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.00978483073413372 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031919742468744516 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.0031919742468744516, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031919742468744516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08711032569408417, + 0.033816564828157425, + 0.025417884811758995, + 0.03421662747859955, + 0.17570984363555908, + 0.09853332489728928, + 0.34040629863739014, + 0.10269971191883087, + 0.05648861080408096, + 0.0456007644534111 + ], + "credit_map": [ + 0.01184842362999916, + 0.012533871456980705, + 0.01004333607852459, + 0.011487177573144436, + 0.009637774899601936, + 0.007773991674184799, + 0.014991873875260353, + 0.00879387091845274, + 0.005272325128316879, + 0.004733830690383911, + 0.004473792389035225, + 0.0030985726043581963, + 0.004142423626035452, + 0.0035070637241005898, + 0.00489880982786417, + 0.003689742647111416, + 0.003939836286008358, + 0.0032331484835594893, + 0.0032799209002405405, + 0.0017857097554951906, + 0.003736012615263462, + 0.002902850741520524, + 0.003293103538453579, + 0.00324730365537107, + 0.007143731229007244, + 0.003631485626101494, + 0.004682086408138275, + 0.005249872338026762, + 0.004021285101771355, + 0.0036927289329469204, + 0.0028796526603400707, + 0.002915786812081933, + 0.027432147413492203, + 0.01896277628839016, + 0.031102260574698448, + 0.019884886220097542, + 0.013255881145596504, + 0.023098140954971313, + 0.017092863097786903, + 0.024880878627300262, + 0.01360621303319931, + 0.01551109179854393, + 0.013038497418165207, + 0.006795581895858049, + 0.011776384897530079, + 0.00907056126743555, + 0.012422571890056133, + 0.01631241664290428, + 0.02109535224735737, + 0.02296067401766777, + 0.08668870478868484, + 0.058784548193216324, + 0.02360386960208416, + 0.025422347709536552, + 0.04220889136195183, + 0.05964191257953644, + 0.008870613761246204, + 0.01719995029270649, + 0.012554884888231754, + 0.017989972606301308, + 0.0101787643507123, + 0.010036990977823734, + 0.012908282689750195, + 0.01296024676412344, + 0.007445249240845442, + 0.008190310560166836, + 0.006088086403906345, + 0.008818392641842365, + 0.005187924951314926, + 0.005395311862230301, + 0.009412871673703194, + 0.005950459279119968, + 0.00730926962569356, + 0.008724328130483627, + 0.006242002360522747, + 0.005153860431164503, + 0.0030754730105400085, + 0.004149939399212599, + 0.0044660549610853195, + 0.006479837466031313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -0.0031919742468744516 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327502891421318 + ], + "opposite_logprob_tokens": [ + -3.773275136947632 + ], + "logprob_reward": -0.02327502891421318, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327502891421318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07758305966854095, + 0.02009689062833786, + 0.02772154100239277, + 0.025960011407732964, + 0.08206774294376373, + 0.12901827692985535, + 0.27852970361709595, + 0.15749159455299377, + 0.13871175050735474, + 0.0628194659948349 + ], + "credit_map": [ + 0.011924263089895248, + 0.008964245207607746, + 0.008247556164860725, + 0.012044664472341537, + 0.007118211127817631, + 0.010401124134659767, + 0.007906390354037285, + 0.010976601392030716, + 0.003314279019832611, + 0.002115186769515276, + 0.0023052035830914974, + 0.0018570905085653067, + 0.002597576240077615, + 0.0030841773841530085, + 0.002437228336930275, + 0.00238614808768034, + 0.0030323828104883432, + 0.0037189002614468336, + 0.0030812444165349007, + 0.0036329985596239567, + 0.003493543015792966, + 0.0033804913982748985, + 0.004594517406076193, + 0.002787459408864379, + 0.0034048159141093493, + 0.004613050725311041, + 0.0021809404715895653, + 0.0018708216957747936, + 0.003691778751090169, + 0.004013490863144398, + 0.0032448808196932077, + 0.002940230304375291, + 0.01224569883197546, + 0.00803009606897831, + 0.012056265026330948, + 0.010646465234458447, + 0.008958674967288971, + 0.011077220551669598, + 0.009502599947154522, + 0.009550705552101135, + 0.014677699655294418, + 0.016378626227378845, + 0.020473897457122803, + 0.020401546731591225, + 0.014711761847138405, + 0.009162254631519318, + 0.021189460530877113, + 0.012023023329675198, + 0.015041199512779713, + 0.01727328822016716, + 0.0803137943148613, + 0.0454186350107193, + 0.013727887533605099, + 0.022313613444566727, + 0.016377516090869904, + 0.06806373596191406, + 0.015081233344972134, + 0.026162687689065933, + 0.017432034015655518, + 0.025674425065517426, + 0.01899629645049572, + 0.017155058681964874, + 0.016481712460517883, + 0.020508132874965668, + 0.013125287368893623, + 0.018359258770942688, + 0.015312607400119305, + 0.02778523787856102, + 0.01178049948066473, + 0.012067766860127449, + 0.029851272702217102, + 0.010429808869957924, + 0.008604103699326515, + 0.010111367329955101, + 0.0072341556660830975, + 0.009321332909166813, + 0.006404685787856579, + 0.005148892290890217, + 0.008002500049769878, + 0.007992418482899666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.02327502891421318 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a static, slightly high-angle medium shot, the camera tilts steadily downward, meticulously tracking the movement of a rectangular, translucent, frosted plastic food container, perched precariously on the edge of a polished, light-gray crystal countertop in a modern kitchen. Soft, realistic afternoon light streams through a large, off-camera window, casting diffuse highlights on the container's vibrant blue snap-on lid and creating long, soft shadows. The container teeters for a moment before quickly and decisively toppling over its edge, tumbling through the air. The camera follows its descent until it impacts the vast, off-white tile floor with a dull, hollow thud, its flexible plastic body visibly deforming for a moment after the impact. It then executes a slight bounce, skimming a few inches across the smooth surface before swaying to a completely abrupt sideways rest, concluding the dynamic movement with a quiet sense of finality in the quiet, sunlit kitchen.", + "seed": 869, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_0/index_9458_seed_869_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03221824308857322, + "vlm_logprob": -0.03221824308857322, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.03221824308857322, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1053958609700203 + ], + "opposite_logprob_tokens": [ + -2.355395793914795 + ], + "logprob_reward": -0.1053958609700203, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1053958609700203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29118409752845764, + 0.03024222142994404, + 0.036503709852695465, + 0.04676443338394165, + 0.08702877908945084, + 0.13521122932434082, + 0.04367299750447273, + 0.15521369874477386, + 0.07645564526319504, + 0.09772326052188873 + ], + "credit_map": [ + 0.009367524646222591, + 0.10472188144922256, + 0.018268423154950142, + 0.01921459473669529, + 0.10535046458244324, + 0.01373568270355463, + 0.012220889329910278, + 0.008304603397846222, + 0.003893747925758362, + 0.004826237913221121, + 0.006480509415268898, + 0.005260287784039974, + 0.00307974754832685, + 0.002478844951838255, + 0.002348059555515647, + 0.001874786103144288, + 0.005939648952335119, + 0.004314711317420006, + 0.007833062671124935, + 0.004912881646305323, + 0.0026422615628689528, + 0.0028393948450684547, + 0.0029450992587953806, + 0.005076647736132145, + 0.003653920954093337, + 0.0038378129247576, + 0.008303421549499035, + 0.005087226163595915, + 0.004649572540074587, + 0.00534335570409894, + 0.006809313781559467, + 0.009079808369278908, + 0.011848073452711105, + 0.004752410110086203, + 0.022566620260477066, + 0.004631067626178265, + 0.006556789856404066, + 0.005162450950592756, + 0.004865012131631374, + 0.02664635330438614, + 0.01293843425810337, + 0.016320057213306427, + 0.018622783944010735, + 0.023351451382040977, + 0.019749298691749573, + 0.017096223309636116, + 0.01616097241640091, + 0.010972012765705585, + 0.0036395133938640356, + 0.0061218696646392345, + 0.0029835377354174852, + 0.006354353856295347, + 0.005593068432062864, + 0.006291063968092203, + 0.006248181685805321, + 0.006441408768296242, + 0.009426433593034744, + 0.020591791719198227, + 0.014799781143665314, + 0.021175388246774673, + 0.014400617219507694, + 0.014566600322723389, + 0.026781447231769562, + 0.03347164765000343, + 0.007456089835613966, + 0.007117377128452063, + 0.005632157903164625, + 0.011189915239810944, + 0.01090098824352026, + 0.009372316300868988, + 0.014787102118134499, + 0.009999694302678108, + 0.011098754592239857, + 0.008640899322926998, + 0.01674058847129345, + 0.012705517001450062, + 0.009642845019698143, + 0.009725658223032951, + 0.01672699861228466, + 0.012441999278962612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.1053958609700203 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007644092198461294 + ], + "opposite_logprob_tokens": [ + -4.882644176483154 + ], + "logprob_reward": -0.007644092198461294, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007644092198461294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17323118448257446, + 0.1814374476671219, + 0.06168853119015694, + 0.045676521956920624, + 0.06317228823900223, + 0.08713801205158234, + 0.2370510846376419, + 0.055112551897764206, + 0.05844428762793541, + 0.037048112601041794 + ], + "credit_map": [ + 0.006849114783108234, + 0.05264872685074806, + 0.008796549402177334, + 0.016065338626503944, + 0.048036012798547745, + 0.013938267715275288, + 0.012617770582437515, + 0.014279402792453766, + 0.016338670626282692, + 0.02872312068939209, + 0.03395181521773338, + 0.030021661892533302, + 0.030199376866221428, + 0.015291263349354267, + 0.01581612601876259, + 0.011095398105680943, + 0.009513355791568756, + 0.00691224355250597, + 0.009667906910181046, + 0.006747804582118988, + 0.0067779291421175, + 0.008915070444345474, + 0.007623698096722364, + 0.0055305249989032745, + 0.002829774282872677, + 0.004215742461383343, + 0.004774814937263727, + 0.005267931614071131, + 0.006927918177098036, + 0.0033082934096455574, + 0.00985303707420826, + 0.008499005809426308, + 0.008251105435192585, + 0.0052141835913062096, + 0.014041596092283726, + 0.005578351207077503, + 0.006430961191654205, + 0.00664953887462616, + 0.005883332341909409, + 0.011123214848339558, + 0.005745302885770798, + 0.009070387110114098, + 0.012640556320548058, + 0.01691107451915741, + 0.010477710515260696, + 0.015583782456815243, + 0.0091803427785635, + 0.007528857793658972, + 0.021284643560647964, + 0.03208465129137039, + 0.014464454725384712, + 0.02933235466480255, + 0.037155281752347946, + 0.02699480950832367, + 0.04548567160964012, + 0.030249206349253654, + 0.005336390808224678, + 0.004873817320913076, + 0.0055240243673324585, + 0.0070647066459059715, + 0.00534806726500392, + 0.006399356760084629, + 0.007585878483951092, + 0.012980305589735508, + 0.004894965793937445, + 0.006030764896422625, + 0.0036271540448069572, + 0.010221707634627819, + 0.007522715721279383, + 0.00634224247187376, + 0.01286960206925869, + 0.006935134530067444, + 0.0035099093802273273, + 0.0025753129739314318, + 0.006425121333450079, + 0.005229287315160036, + 0.00385979562997818, + 0.0029508413281291723, + 0.0076089040376245975, + 0.004888942465186119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.007644092198461294 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739267613738775 + ], + "opposite_logprob_tokens": [ + -5.006739139556885 + ], + "logprob_reward": -0.006739267613738775, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739267613738775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1442277431488037, + 0.08384336531162262, + 0.08884364366531372, + 0.13432924449443817, + 0.19051168859004974, + 0.11125987023115158, + 0.0646635890007019, + 0.0982966423034668, + 0.0418715663254261, + 0.04215263947844505 + ], + "credit_map": [ + 0.008049409836530685, + 0.02607652358710766, + 0.01546448189765215, + 0.023862311616539955, + 0.026315752416849136, + 0.01587028056383133, + 0.013840585015714169, + 0.014748389832675457, + 0.011044634506106377, + 0.01359659805893898, + 0.020632658153772354, + 0.01554044522345066, + 0.0070741185918450356, + 0.005087288562208414, + 0.005557190161198378, + 0.005310436245054007, + 0.012935773469507694, + 0.010384997352957726, + 0.013126706704497337, + 0.009839748032391071, + 0.009093259461224079, + 0.012457667849957943, + 0.010531622916460037, + 0.010473870672285557, + 0.008403592742979527, + 0.007465663366019726, + 0.014879954047501087, + 0.011063634417951107, + 0.019862007349729538, + 0.008451154455542564, + 0.012777131982147694, + 0.05142611637711525, + 0.023962978273630142, + 0.009961442090570927, + 0.04759743809700012, + 0.013316421769559383, + 0.015076661482453346, + 0.010598214343190193, + 0.011706415563821793, + 0.05829210579395294, + 0.01017878856509924, + 0.018239283934235573, + 0.01374618150293827, + 0.013441327959299088, + 0.01565466821193695, + 0.021515605971217155, + 0.011156419292092323, + 0.00732759851962328, + 0.0067087686620652676, + 0.007250147871673107, + 0.006746307481080294, + 0.008474063128232956, + 0.009865514934062958, + 0.0077711353078484535, + 0.008538078516721725, + 0.009309573099017143, + 0.006645772140473127, + 0.008252657949924469, + 0.007891317829489708, + 0.01933966390788555, + 0.008160622790455818, + 0.0075495135970413685, + 0.020007038488984108, + 0.020450053736567497, + 0.0033737921621650457, + 0.004697133786976337, + 0.005439150612801313, + 0.005705677904188633, + 0.005252700299024582, + 0.0040603806264698505, + 0.006676848046481609, + 0.006665881257504225, + 0.003322446020320058, + 0.005100650247186422, + 0.0050390311516821384, + 0.006787032820284367, + 0.007841932587325573, + 0.0034629451110959053, + 0.006447724532335997, + 0.004150876309722662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.006739267613738775 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630605716258287 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630605716258287, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630605716258287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17974981665611267, + 0.10624685883522034, + 0.02185101993381977, + 0.0970030128955841, + 0.37826210260391235, + 0.06844352185726166, + 0.0466863252222538, + 0.04646678268909454, + 0.028174396604299545, + 0.02711619809269905 + ], + "credit_map": [ + 0.005124980583786964, + 0.0681401938199997, + 0.006661901716142893, + 0.00962447002530098, + 0.06721005588769913, + 0.00834256038069725, + 0.008138712495565414, + 0.006506926380097866, + 0.012514465488493443, + 0.01772007904946804, + 0.026085056364536285, + 0.021763775497674942, + 0.009282972663640976, + 0.0054311868734657764, + 0.007289182860404253, + 0.0061601391062140465, + 0.004216366913169622, + 0.0029937257058918476, + 0.0025043904315680265, + 0.0026256521232426167, + 0.0023810158018022776, + 0.002699790522456169, + 0.0025558480992913246, + 0.0018742299871519208, + 0.004107034299522638, + 0.007294663228094578, + 0.005585082340985537, + 0.009379701688885689, + 0.01872982084751129, + 0.008254386484622955, + 0.01018520351499319, + 0.0334671214222908, + 0.04376024007797241, + 0.009496919810771942, + 0.09227043390274048, + 0.019353803247213364, + 0.03623328357934952, + 0.021386990323662758, + 0.022109419107437134, + 0.133651003241539, + 0.004968787077814341, + 0.007571767084300518, + 0.008751904591917992, + 0.011217853985726833, + 0.008187350817024708, + 0.014752085320651531, + 0.0066169449128210545, + 0.006376830395311117, + 0.004136966075748205, + 0.005787575617432594, + 0.0033913811203092337, + 0.006845269352197647, + 0.006134054623544216, + 0.004704034421592951, + 0.007889513857662678, + 0.007797530852258205, + 0.004611088428646326, + 0.005868865177035332, + 0.003597864881157875, + 0.007959580980241299, + 0.005958589259535074, + 0.005049371626228094, + 0.006674882490187883, + 0.006746536586433649, + 0.0028270832262933254, + 0.003928819205611944, + 0.0029549484606832266, + 0.004318736027926207, + 0.0034732427448034286, + 0.003094540210440755, + 0.004438517149537802, + 0.003138511208817363, + 0.002118011936545372, + 0.0030304952524602413, + 0.0027915132232010365, + 0.004209712613373995, + 0.0036707287654280663, + 0.002905795583501458, + 0.00472738454118371, + 0.003662555944174528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.007630605716258287 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a modern kitchen with a polished, light-gray crystal countertop and an off-white tile floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a rectangular, translucent, frosted plastic food container with a vibrant blue snap-on lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the container start perched on the edge of the countertop, teeter, and then topple over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start as a static, slightly high-angle medium shot and tilt steadily downward to follow the container's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"After hitting the floor, does the container deform, bounce, skim across the surface, and come to an abrupt sideways rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033681388944387436 + ], + "opposite_logprob_tokens": [ + -3.4086813926696777 + ], + "logprob_reward": -0.033681388944387436, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033681388944387436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3369624614715576, + 0.08682465553283691, + 0.016666170209646225, + 0.04085748642683029, + 0.19819460809230804, + 0.06070966646075249, + 0.06185048818588257, + 0.0778188481926918, + 0.06256554275751114, + 0.0575501024723053 + ], + "credit_map": [ + 0.004957068245857954, + 0.14165811240673065, + 0.007882717996835709, + 0.014106468297541142, + 0.13922066986560822, + 0.01261897012591362, + 0.009371223859488964, + 0.007147209253162146, + 0.01023572962731123, + 0.014469850808382034, + 0.02434273064136505, + 0.01595919020473957, + 0.008487071841955185, + 0.004761922173202038, + 0.004491978790611029, + 0.004076163750141859, + 0.002604813314974308, + 0.0019521963549777865, + 0.0023503152187913656, + 0.0015459050191566348, + 0.0018906473414972425, + 0.0023151065688580275, + 0.002158118411898613, + 0.0018490658840164542, + 0.0029564930591732264, + 0.0038648098707199097, + 0.006430121138691902, + 0.004630066454410553, + 0.004887321498245001, + 0.004889982752501965, + 0.006768224760890007, + 0.006430465262383223, + 0.025891583412885666, + 0.005207978654652834, + 0.052920643240213394, + 0.010277929715812206, + 0.0161322969943285, + 0.010563057847321033, + 0.011074117384850979, + 0.0661269947886467, + 0.007233253680169582, + 0.008775795809924603, + 0.006702393759042025, + 0.010061083361506462, + 0.007788629271090031, + 0.009354408830404282, + 0.005898254923522472, + 0.004895844031125307, + 0.004287821240723133, + 0.007269993890076876, + 0.0037783454172313213, + 0.008661052212119102, + 0.011343983933329582, + 0.007171515841037035, + 0.012578316032886505, + 0.00675945496186614, + 0.0059031471610069275, + 0.008427451364696026, + 0.007082904689013958, + 0.011302550323307514, + 0.006447999272495508, + 0.006940135732293129, + 0.011619780212640762, + 0.020094873383641243, + 0.005505726207047701, + 0.00932286400347948, + 0.007048217114061117, + 0.00859405193477869, + 0.007960613816976547, + 0.0069577982649207115, + 0.008085708133876324, + 0.009090564213693142, + 0.00419632438570261, + 0.008664860390126705, + 0.005893665365874767, + 0.01021009311079979, + 0.010759050957858562, + 0.004760597832500935, + 0.007147886790335178, + 0.005917615257203579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.033681388944387436 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_1/index_tensor([7887, 7887])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_1/index_tensor([7887, 7887])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4fe589166b57713b85e6d71a80b9c3b97de79e48 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_1/index_tensor([7887, 7887])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning close-up, filmed in high-frame-rate slow motion, captures a plump, hand-formed dumpling, its belly plump, plump and plump, plumply falling from the air. A beam of warm overhead light, precisely illuminating the thin, translucent wrapper, reveals its delicate folds and the subtle trembling of its form under the weight of gravity. Beneath it lies a large, dark ceramic bowl, filled with scalding, deep-brown broth. The broth trembles subtly from the heat, and plumes of steam rise, surfacing with a few emerald green scallions and a few drops of glistening sesame oil. As the dumpling approaches with steady deliberation, its base first strikes the scalding broth; instantly, the surface is forcefully broken, a sharp, distinct ripple spreading outward from the point of impact. Countless droplets of scalding broth explode into the air, condensing into crystalline spheres before slowly falling. The dumplings then sank, completely immersed in the rich, oily soup. The entire scene, captured in extreme slow motion, transformed a simple cooking process into a visual feast full of power and beauty, creating an ultimate sensory enjoyment and anticipation for delicious food.", + "seed": 1941, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_1/index_7887_seed_1941_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04525455771945417, + "vlm_logprob": -0.04525455771945417, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.04525455771945417, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17448334395885468 + ], + "opposite_logprob_tokens": [ + -3.549483299255371 + ], + "logprob_reward": -0.17448334395885468, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.17448334395885468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4502767324447632, + 0.0860561951994896, + 0.029492106288671494, + 0.10253499448299408, + 0.07500681281089783, + 0.024808134883642197, + 0.10288142412900925, + 0.01980713941156864, + 0.015608316287398338, + 0.093528151512146 + ], + "credit_map": [ + 0.022374996915459633, + 0.16404379904270172, + 0.013514957390725613, + 0.008638281375169754, + 0.19810044765472412, + 0.02421867847442627, + 0.010954825207591057, + 0.00843074731528759, + 0.008179586380720139, + 0.012726393528282642, + 0.007993843406438828, + 0.016346031799912453, + 0.00927081797271967, + 0.012897655367851257, + 0.010620889253914356, + 0.008020978420972824, + 0.003819270757958293, + 0.0033019944094121456, + 0.0056703956797719, + 0.0038428721018135548, + 0.003356730565428734, + 0.0028900413308292627, + 0.003035095753148198, + 0.0035757070872932673, + 0.009904177859425545, + 0.009199959225952625, + 0.009783039800822735, + 0.008936767466366291, + 0.015270831994712353, + 0.01945488713681698, + 0.011896423064172268, + 0.018088914453983307, + 0.014571547508239746, + 0.0034375092945992947, + 0.0038923320826143026, + 0.004291005898267031, + 0.007384583819657564, + 0.005572563968598843, + 0.031731851398944855, + 0.004125417210161686, + 0.004837498534470797, + 0.0017050491878762841, + 0.0017106607556343079, + 0.001769787515513599, + 0.002616966376081109, + 0.006574021186679602, + 0.0027345919515937567, + 0.0028595596086233854, + 0.012923402711749077, + 0.008911971934139729, + 0.009256023913621902, + 0.014698384329676628, + 0.01079163420945406, + 0.007650358602404594, + 0.02881433069705963, + 0.009835327044129372, + 0.002699853852391243, + 0.002237204695120454, + 0.002391127636656165, + 0.002001433167606592, + 0.0030230034608393908, + 0.0021278385538607836, + 0.0031414676923304796, + 0.00218521011993289, + 0.0023483221884816885, + 0.001194262644276023, + 0.002368374727666378, + 0.001583396689966321, + 0.0017394624883309007, + 0.0019132348243147135, + 0.0023216723930090666, + 0.002139589749276638, + 0.012708297930657864, + 0.005768336821347475, + 0.020198525860905647, + 0.006568599957972765, + 0.006309510208666325, + 0.031594645231962204, + 0.006319369189441204, + 0.004060867242515087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.17448334395885468 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014274394139647484 + ], + "opposite_logprob_tokens": [ + -4.264274597167969 + ], + "logprob_reward": -0.014274394139647484, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014274394139647484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5737684369087219, + 0.03365664556622505, + 0.01929006166756153, + 0.01708720065653324, + 0.20461417734622955, + 0.014658156782388687, + 0.06625988334417343, + 0.03323895111680031, + 0.012163548730313778, + 0.025262918323278427 + ], + "credit_map": [ + 0.02753753401339054, + 0.21353007853031158, + 0.01490053627640009, + 0.01026812382042408, + 0.253129780292511, + 0.031192198395729065, + 0.012383212335407734, + 0.01082695834338665, + 0.0029286514036357403, + 0.006456910632550716, + 0.0021910744253546, + 0.004852443467825651, + 0.003977075219154358, + 0.00554417772218585, + 0.004947679582983255, + 0.0027586345095187426, + 0.001541656325571239, + 0.003053755732253194, + 0.0021531302481889725, + 0.0028949647676199675, + 0.0021475651301443577, + 0.0034772201906889677, + 0.0023419694043695927, + 0.0016797981224954128, + 0.0016398479929193854, + 0.0014442155370488763, + 0.0020251472014933825, + 0.0016592071624472737, + 0.0017414191970601678, + 0.0031127051915973425, + 0.002075762255117297, + 0.003388897283002734, + 0.052139461040496826, + 0.007011798210442066, + 0.006069405935704708, + 0.009099824354052544, + 0.013097419403493404, + 0.01312303077429533, + 0.09573250263929367, + 0.00834073219448328, + 0.002897747792303562, + 0.0015403582947328687, + 0.0016043174546211958, + 0.0011234617559239268, + 0.0017434961628168821, + 0.002136573428288102, + 0.001747400383464992, + 0.0018648020923137665, + 0.009236473590135574, + 0.005807590205222368, + 0.0058325519785285, + 0.009169568307697773, + 0.006648825481534004, + 0.005092768929898739, + 0.01806194707751274, + 0.006410160567611456, + 0.00408905278891325, + 0.0036454475484788418, + 0.004534800071269274, + 0.0039050320629030466, + 0.003675523679703474, + 0.0036582541652023792, + 0.006070946343243122, + 0.00365989631973207, + 0.0018509153742343187, + 0.0011091874912381172, + 0.0017325328662991524, + 0.0012806286104023457, + 0.0014161213766783476, + 0.0016512117581441998, + 0.0019262159476056695, + 0.0011967351892963052, + 0.003325785044580698, + 0.001947944751009345, + 0.0048800245858728886, + 0.0019355985568836331, + 0.0019778527785092592, + 0.007926304824650288, + 0.002101007616147399, + 0.0011684010969474912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.014274394139647484 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025044286157935858 + ], + "opposite_logprob_tokens": [ + -6.002504348754883 + ], + "logprob_reward": -0.0025044286157935858, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025044286157935858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8295320868492126, + 0.008965306915342808, + 0.010918348096311092, + 0.007710098288953304, + 0.0521644726395607, + 0.009895842522382736, + 0.0340021476149559, + 0.020315954461693764, + 0.016956234350800514, + 0.00953943096101284 + ], + "credit_map": [ + 0.039117973297834396, + 0.31215566396713257, + 0.0214653629809618, + 0.015002008527517319, + 0.3633958101272583, + 0.04371006786823273, + 0.01831710711121559, + 0.01636812835931778, + 0.0011291616829112172, + 0.0012356721563264728, + 0.0009238733910024166, + 0.0020893607288599014, + 0.0007678153924643993, + 0.0010092280572280288, + 0.0011345020029693842, + 0.0006756934453733265, + 0.0012096994323655963, + 0.0021443248260766268, + 0.0010443136561661959, + 0.0010375718120485544, + 0.0014695936115458608, + 0.001270935870707035, + 0.0011657520662993193, + 0.0015761570539325476, + 0.000735869980417192, + 0.0006610763957723975, + 0.0008264431962743402, + 0.0007344648474827409, + 0.0011338128242641687, + 0.0015544351190328598, + 0.0008917119121178985, + 0.0011722842464223504, + 0.011780093424022198, + 0.002330103190615773, + 0.0014754750300198793, + 0.002478888723999262, + 0.004174436908215284, + 0.003579347161576152, + 0.024055885151028633, + 0.0022902425844222307, + 0.00148041732609272, + 0.0009429329074919224, + 0.0011864446569234133, + 0.0007195529760792851, + 0.0011725573567673564, + 0.0012562202755361795, + 0.0016761510632932186, + 0.0014615664258599281, + 0.005204671993851662, + 0.0030382953118532896, + 0.003103181952610612, + 0.004430483561009169, + 0.003187209367752075, + 0.0026989539619535208, + 0.00877357553690672, + 0.0035657742992043495, + 0.0032151115592569113, + 0.0020962280686944723, + 0.003888889914378524, + 0.002310581039637327, + 0.0017685015918686986, + 0.00260772742331028, + 0.002667601453140378, + 0.0017613134114071727, + 0.0028821013402193785, + 0.001514578121714294, + 0.0030156392604112625, + 0.001787960180081427, + 0.001504596322774887, + 0.0019390432862564921, + 0.002379915677011013, + 0.00193240016233176, + 0.0009410394704900682, + 0.0006531589315272868, + 0.0023337677121162415, + 0.0014011719031259418, + 0.0006916829152032733, + 0.001618069945834577, + 0.0011896175565198064, + 0.0007109226426109672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0025044286157935858 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635768987238407 + ], + "opposite_logprob_tokens": [ + -3.651357650756836 + ], + "logprob_reward": -0.02635768987238407, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635768987238407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6145584583282471, + 0.14159010350704193, + 0.02511465549468994, + 0.032301317900419235, + 0.062329139560461044, + 0.012558636255562305, + 0.020150411874055862, + 0.030721168965101242, + 0.021216480061411858, + 0.039459552615880966 + ], + "credit_map": [ + 0.029801320284605026, + 0.22772982716560364, + 0.016401207074522972, + 0.010819798335433006, + 0.2709006667137146, + 0.03344754874706268, + 0.013749171048402786, + 0.011708959937095642, + 0.014485505409538746, + 0.02233746089041233, + 0.011882469058036804, + 0.02826729230582714, + 0.014136746525764465, + 0.020197702571749687, + 0.018280215561389923, + 0.012002711184322834, + 0.003160107182338834, + 0.004011859651654959, + 0.004004098474979401, + 0.00318541401065886, + 0.003075163811445236, + 0.002788045210763812, + 0.0029335818253457546, + 0.0019563843961805105, + 0.0029971033800393343, + 0.00322140590287745, + 0.0027385596185922623, + 0.0033017240930348635, + 0.00573003338649869, + 0.005582250189036131, + 0.004120201803743839, + 0.004610039759427309, + 0.012064825743436813, + 0.003913909662514925, + 0.0019638105295598507, + 0.0035450372379273176, + 0.006027675699442625, + 0.004751168191432953, + 0.02666032314300537, + 0.0034023867920041084, + 0.0016830029198899865, + 0.0012853927910327911, + 0.0012055400293320417, + 0.0013581691309809685, + 0.0017618329729884863, + 0.0017834958853200078, + 0.002152180764824152, + 0.0013290210627019405, + 0.002216750755906105, + 0.0018412438221275806, + 0.001904329052194953, + 0.0028874403797090054, + 0.002068418078124523, + 0.0016222462290897965, + 0.005475656595081091, + 0.0021343270782381296, + 0.004040515050292015, + 0.003242298262193799, + 0.0023251711390912533, + 0.0029866336844861507, + 0.006488501559942961, + 0.004094951786100864, + 0.004934343509376049, + 0.0026087549049407244, + 0.002806868636980653, + 0.001564938691444695, + 0.002879207720980048, + 0.0025322027504444122, + 0.002156143542379141, + 0.003720234613865614, + 0.0037303967401385307, + 0.0018264884129166603, + 0.005644700955599546, + 0.0012815090594813228, + 0.009935352951288223, + 0.001872428460046649, + 0.0021499227732419968, + 0.014977197162806988, + 0.002174126449972391, + 0.0014243163168430328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.02635768987238407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865293201059103 + ], + "opposite_logprob_tokens": [ + -4.758653163909912 + ], + "logprob_reward": -0.00865293201059103, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865293201059103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6332495808601379, + 0.03162328898906708, + 0.025924252346158028, + 0.019264638423919678, + 0.11102713644504547, + 0.04039885848760605, + 0.04574856162071228, + 0.03448360040783882, + 0.018262328580021858, + 0.04001771658658981 + ], + "credit_map": [ + 0.02923334203660488, + 0.24051620066165924, + 0.016335347667336464, + 0.011470661498606205, + 0.2758534252643585, + 0.032351795583963394, + 0.014137660153210163, + 0.013351158238947392, + 0.0021021110005676746, + 0.005057741887867451, + 0.002567408373579383, + 0.006075271870940924, + 0.0042037395760416985, + 0.005277379881590605, + 0.004199790768325329, + 0.0021398458629846573, + 0.002901690546423197, + 0.0025005871430039406, + 0.003870344487950206, + 0.0032000772189348936, + 0.003552120178937912, + 0.002485761186107993, + 0.0029017196502536535, + 0.00451195053756237, + 0.0017933081835508347, + 0.0018563035409897566, + 0.0018967266660183668, + 0.0022537049371749163, + 0.0032684742473065853, + 0.0030502239242196083, + 0.0020597095135599375, + 0.003086189040914178, + 0.027869874611496925, + 0.004211560357362032, + 0.0034912084229290485, + 0.0047578709200024605, + 0.007220406085252762, + 0.00721587473526597, + 0.05138498544692993, + 0.004875349812209606, + 0.00837330985814333, + 0.0031961766071617603, + 0.003182042622938752, + 0.003092437284067273, + 0.003481591586023569, + 0.008899153210222721, + 0.004323041997849941, + 0.005851107183843851, + 0.005323920398950577, + 0.0038486907724291086, + 0.004033318720757961, + 0.006784828379750252, + 0.004526984877884388, + 0.003609216306358576, + 0.01264872681349516, + 0.004972877446562052, + 0.004560622852295637, + 0.0038064925465732813, + 0.004488897044211626, + 0.003280696924775839, + 0.003360700560733676, + 0.0037544751539826393, + 0.0066216192208230495, + 0.004610099364072084, + 0.0036666542291641235, + 0.0020388930570334196, + 0.00226229103282094, + 0.002476113848388195, + 0.0016239717369899154, + 0.0016897920286282897, + 0.0027798672672361135, + 0.0017247442156076431, + 0.005571628920733929, + 0.0027239590417593718, + 0.008238320238888264, + 0.002809882164001465, + 0.002306598937138915, + 0.012979768216609955, + 0.002839312655851245, + 0.002548244781792164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.00865293201059103 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, filmed in high-frame-rate slow motion, captures a plump, hand-formed dumpling, its belly plump, plump and plump, plumply falling from the air. A beam of warm overhead light, precisely illuminating the thin, translucent wrapper, reveals its delicate folds and the subtle trembling of its form under the weight of gravity. Beneath it lies a large, dark ceramic bowl, filled with scalding, deep-brown broth. The broth trembles subtly from the heat, and plumes of steam rise, surfacing with a few emerald green scallions and a few drops of glistening sesame oil. As the dumpling approaches with steady deliberation, its base first strikes the scalding broth; instantly, the surface is forcefully broken, a sharp, distinct ripple spreading outward from the point of impact. Countless droplets of scalding broth explode into the air, condensing into crystalline spheres before slowly falling. The dumplings then sank, completely immersed in the rich, oily soup. The entire scene, captured in extreme slow motion, transformed a simple cooking process into a visual feast full of power and beauty, creating an ultimate sensory enjoyment and anticipation for delicious food.", + "seed": 1942, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_1/index_7887_seed_1942_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07689177710562944, + "vlm_logprob": -0.07689177710562944, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.07689177710562944, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3021225035190582 + ], + "opposite_logprob_tokens": [ + -2.4271225929260254 + ], + "logprob_reward": -0.3021225035190582, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.3021225035190582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04263782501220703, + 0.03520785644650459, + 0.04878811910748482, + 0.4770472049713135, + 0.07801887392997742, + 0.13054504990577698, + 0.14149875938892365, + 0.008505584672093391, + 0.009437638334929943, + 0.028313064947724342 + ], + "credit_map": [ + 0.0042074755765497684, + 0.006895634811371565, + 0.0037720194086432457, + 0.0035642918664962053, + 0.007855606265366077, + 0.007462234236299992, + 0.005866405554115772, + 0.0030141600873321295, + 0.003454032354056835, + 0.006993736606091261, + 0.00486553693190217, + 0.007013857830315828, + 0.003972270991653204, + 0.0024262957740575075, + 0.003865532111376524, + 0.00261659175157547, + 0.005324654281139374, + 0.013943453319370747, + 0.004755993839353323, + 0.0040533458814024925, + 0.0035730546806007624, + 0.002978457370772958, + 0.01105985976755619, + 0.0030992976389825344, + 0.09253387898206711, + 0.02910125069320202, + 0.012684905901551247, + 0.06560351699590683, + 0.04469243809580803, + 0.010947887785732746, + 0.20039553940296173, + 0.021087801083922386, + 0.010238632559776306, + 0.0075871627777814865, + 0.013423140160739422, + 0.008977950550615788, + 0.007250495720654726, + 0.005884165875613689, + 0.012158782221376896, + 0.012498545460402966, + 0.024049442261457443, + 0.014765718951821327, + 0.010923978872597218, + 0.014339103363454342, + 0.014730994589626789, + 0.027238858863711357, + 0.013411078602075577, + 0.0110858716070652, + 0.02412225492298603, + 0.016357991844415665, + 0.01556701585650444, + 0.017557606101036072, + 0.015989024192094803, + 0.019083214923739433, + 0.020901218056678772, + 0.01192044373601675, + 0.0011350184213370085, + 0.001171408104710281, + 0.0008952558855526149, + 0.0012948651565238833, + 0.0008764624944888055, + 0.0009230845025740564, + 0.0012997291050851345, + 0.0009097606525756419, + 0.0016015239525586367, + 0.0015618231846019626, + 0.0006487427745014429, + 0.0005151215591467917, + 0.0004308440547902137, + 0.0035728258080780506, + 0.0003955705324187875, + 0.0007111859740689397, + 0.004936192650347948, + 0.002542838454246521, + 0.00270456844009459, + 0.002915468532592058, + 0.0017339535988867283, + 0.008626767434179783, + 0.0022934835869818926, + 0.0025597920175641775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.3021225035190582 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016142504289746284 + ], + "opposite_logprob_tokens": [ + -4.14114236831665 + ], + "logprob_reward": -0.016142504289746284, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016142504289746284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23347388207912445, + 0.0513177327811718, + 0.0747247189283371, + 0.4215376079082489, + 0.025059975683689117, + 0.04495225101709366, + 0.026771128177642822, + 0.017255835235118866, + 0.024768678471446037, + 0.08013816922903061 + ], + "credit_map": [ + 0.019879192113876343, + 0.03907977417111397, + 0.021789146587252617, + 0.01976596564054489, + 0.04066016152501106, + 0.043667152523994446, + 0.027429860085248947, + 0.021202605217695236, + 0.00463152164593339, + 0.008200166746973991, + 0.0050413901917636395, + 0.013266416266560555, + 0.0055386885069310665, + 0.004723529797047377, + 0.006865181028842926, + 0.0030508360359817743, + 0.006060589104890823, + 0.019909624010324478, + 0.00525606470182538, + 0.006032065022736788, + 0.005404503084719181, + 0.00622523482888937, + 0.021567780524492264, + 0.0042688497342169285, + 0.08212064206600189, + 0.02806524932384491, + 0.01181961502879858, + 0.05568259209394455, + 0.03998735547065735, + 0.010799692012369633, + 0.17511171102523804, + 0.0179507564753294, + 0.0028638907242566347, + 0.0026673018001019955, + 0.0030553233809769154, + 0.004556681029498577, + 0.0021000192500650883, + 0.0017686380306258798, + 0.005250435322523117, + 0.002797684632241726, + 0.008575071580708027, + 0.004534343257546425, + 0.0042446511797606945, + 0.005187910050153732, + 0.003615773981437087, + 0.009834952652454376, + 0.005700863432139158, + 0.0032586860470473766, + 0.0043401834554970264, + 0.002694413997232914, + 0.002614338882267475, + 0.003333984175696969, + 0.0020084120333194733, + 0.004211978521198034, + 0.0052061527967453, + 0.002361661521717906, + 0.0028618169017136097, + 0.0023915108758956194, + 0.0021094181574881077, + 0.0023971160408109426, + 0.0017371156718581915, + 0.001509571447968483, + 0.0029117364902049303, + 0.001337548834271729, + 0.0055115725845098495, + 0.0031288450118154287, + 0.0022428128868341446, + 0.0022210159804672003, + 0.0016689491458237171, + 0.005591212771832943, + 0.0017833570018410683, + 0.00262091145850718, + 0.012977363541722298, + 0.007350462954491377, + 0.008363784290850163, + 0.007656247820705175, + 0.004848997108638287, + 0.02473537251353264, + 0.007163556292653084, + 0.007042377721518278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.016142504289746284 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005965168122202158 + ], + "opposite_logprob_tokens": [ + -5.130965232849121 + ], + "logprob_reward": -0.005965168122202158, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005965168122202158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06930557638406754, + 0.06114931404590607, + 0.04637270048260689, + 0.5327267646789551, + 0.045498769730329514, + 0.09106907248497009, + 0.12388195842504501, + 0.009192798286676407, + 0.008119405247271061, + 0.012683720327913761 + ], + "credit_map": [ + 0.010060522705316544, + 0.009901696816086769, + 0.007414021994918585, + 0.008890436962246895, + 0.007340058218687773, + 0.011155293323099613, + 0.009759468026459217, + 0.004784076474606991, + 0.005329236853867769, + 0.014407923445105553, + 0.0049725924618542194, + 0.013289968483150005, + 0.008133004419505596, + 0.005040081217885017, + 0.006041200831532478, + 0.003935306333005428, + 0.00532446289435029, + 0.01354733482003212, + 0.00426337169483304, + 0.003813304705545306, + 0.0036121602170169353, + 0.0032877232879400253, + 0.009837829507887363, + 0.002686513587832451, + 0.1043953150510788, + 0.03309300169348717, + 0.011814986355602741, + 0.0739726647734642, + 0.049770060926675797, + 0.01054944097995758, + 0.22567133605480194, + 0.02345995604991913, + 0.0048301643691957, + 0.00488781975582242, + 0.008726595900952816, + 0.006376972887665033, + 0.003957588225603104, + 0.003120376728475094, + 0.006170981097966433, + 0.007428268436342478, + 0.016630321741104126, + 0.010349665768444538, + 0.007400793954730034, + 0.010924325324594975, + 0.010650754906237125, + 0.01826835609972477, + 0.009565628133714199, + 0.007279232610017061, + 0.021564211696386337, + 0.014484010636806488, + 0.013518639840185642, + 0.015547475777566433, + 0.013610290363430977, + 0.017121752724051476, + 0.018142128363251686, + 0.009893443435430527, + 0.00127976865042001, + 0.0012145985383540392, + 0.0010707564651966095, + 0.0015285330591723323, + 0.0008736422169022262, + 0.0010638709645718336, + 0.0013007278321310878, + 0.0008609004435129464, + 0.0014875315828248858, + 0.0010298556881025434, + 0.0009343313286080956, + 0.0007283335435204208, + 0.00046083450433798134, + 0.001329432474449277, + 0.0007705732714384794, + 0.0013785121263936162, + 0.002658885670825839, + 0.0013021061895415187, + 0.0010494384914636612, + 0.0012280711671337485, + 0.0009154185536317527, + 0.003517243079841137, + 0.0010607974836602807, + 0.0009517595171928406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.005965168122202158 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054963886737823486 + ], + "opposite_logprob_tokens": [ + -2.9299638271331787 + ], + "logprob_reward": -0.054963886737823486, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054963886737823486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3556041419506073, + 0.08032727241516113, + 0.06375855952501297, + 0.1545698493719101, + 0.046673405915498734, + 0.07568036019802094, + 0.03348011150956154, + 0.015485716052353382, + 0.03325355425477028, + 0.1411670595407486 + ], + "credit_map": [ + 0.0427500456571579, + 0.06065373122692108, + 0.05158882960677147, + 0.03534995764493942, + 0.047484979033470154, + 0.04446728900074959, + 0.035857751965522766, + 0.037451520562171936, + 0.0072955479845404625, + 0.018246307969093323, + 0.009328337386250496, + 0.007055879570543766, + 0.011454358696937561, + 0.00890408270061016, + 0.008265085518360138, + 0.009777665138244629, + 0.007661080453544855, + 0.016717910766601562, + 0.007371021434664726, + 0.00533114792779088, + 0.005365086253732443, + 0.004429011605679989, + 0.01358227338641882, + 0.0033010225743055344, + 0.02809545025229454, + 0.009937174618244171, + 0.00966165866702795, + 0.019864050671458244, + 0.014026072807610035, + 0.007658445741981268, + 0.058118924498558044, + 0.007208063267171383, + 0.007121892645955086, + 0.004806930664926767, + 0.006179774180054665, + 0.005307299550622702, + 0.0042291926220059395, + 0.0041451542638242245, + 0.008056015707552433, + 0.006827142555266619, + 0.013453940860927105, + 0.006856139749288559, + 0.00728846900165081, + 0.008602481335401535, + 0.007737437263131142, + 0.017247583717107773, + 0.008576514199376106, + 0.005917789880186319, + 0.004136769101023674, + 0.0029913035687059164, + 0.005602123681455851, + 0.003766098292544484, + 0.0035865772515535355, + 0.004903189837932587, + 0.004473939538002014, + 0.004020107910037041, + 0.0022719656117260456, + 0.0017724918434396386, + 0.002053026808425784, + 0.0018527904758229852, + 0.0014469037996605039, + 0.002285151509568095, + 0.0016185145359486341, + 0.002184870420023799, + 0.005707098636776209, + 0.004084627144038677, + 0.0018374845385551453, + 0.0031810945365577936, + 0.002016231417655945, + 0.009621178731322289, + 0.003140971064567566, + 0.00366486725397408, + 0.02292208932340145, + 0.013434873893857002, + 0.014731151983141899, + 0.012514347210526466, + 0.009218449704349041, + 0.04199428856372833, + 0.012428089044988155, + 0.013923780992627144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.054963886737823486 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005264822859317064 + ], + "opposite_logprob_tokens": [ + -5.255264759063721 + ], + "logprob_reward": -0.005264822859317064, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005264822859317064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12700745463371277, + 0.04544423520565033, + 0.013137441128492355, + 0.4937596321105957, + 0.046727728098630905, + 0.09530782699584961, + 0.10267169028520584, + 0.014490224421024323, + 0.01571999303996563, + 0.04573383554816246 + ], + "credit_map": [ + 0.01188828144222498, + 0.014473157934844494, + 0.011405428871512413, + 0.00931165087968111, + 0.024590209126472473, + 0.03137544170022011, + 0.012246389873325825, + 0.011716898530721664, + 0.0038450388237833977, + 0.011348784901201725, + 0.005592145025730133, + 0.005416732747107744, + 0.006630233023315668, + 0.004940659273415804, + 0.0038489727303385735, + 0.0038216656539589167, + 0.0013605334097519517, + 0.0018090747762471437, + 0.0015923934988677502, + 0.0017777030589058995, + 0.0015892913797870278, + 0.0015625605592504144, + 0.0021481728181242943, + 0.0012977117439731956, + 0.09592144936323166, + 0.030286889523267746, + 0.010946477763354778, + 0.06901367008686066, + 0.046239059418439865, + 0.010408809408545494, + 0.20918700098991394, + 0.021756257861852646, + 0.0054419334046542645, + 0.004852141719311476, + 0.008600020781159401, + 0.0059706075116992, + 0.004034482873976231, + 0.003444350091740489, + 0.006327853072434664, + 0.0080563398078084, + 0.01630011945962906, + 0.00857970304787159, + 0.009590630419552326, + 0.012019459158182144, + 0.01024000532925129, + 0.019854847341775894, + 0.01098382193595171, + 0.007739233318716288, + 0.01646765135228634, + 0.011582652106881142, + 0.010688018053770065, + 0.013369591906666756, + 0.011872735805809498, + 0.01484642457216978, + 0.015499417670071125, + 0.008345205336809158, + 0.002081987215206027, + 0.002004033885896206, + 0.0023411698639392853, + 0.0018634728621691465, + 0.0012509053340181708, + 0.0017003209795802832, + 0.002134564332664013, + 0.0011137692490592599, + 0.0049227396957576275, + 0.001699572429060936, + 0.001131489290855825, + 0.0009648154373280704, + 0.0007803278276696801, + 0.00348017830401659, + 0.0009451172081753612, + 0.0017957534873858094, + 0.00967143289744854, + 0.005851013585925102, + 0.004812140949070454, + 0.006311164237558842, + 0.0041988082230091095, + 0.00643434002995491, + 0.005318125709891319, + 0.0031368136405944824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.005264822859317064 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, filmed in high-frame-rate slow motion, captures a plump, hand-formed dumpling, its belly plump, plump and plump, plumply falling from the air. A beam of warm overhead light, precisely illuminating the thin, translucent wrapper, reveals its delicate folds and the subtle trembling of its form under the weight of gravity. Beneath it lies a large, dark ceramic bowl, filled with scalding, deep-brown broth. The broth trembles subtly from the heat, and plumes of steam rise, surfacing with a few emerald green scallions and a few drops of glistening sesame oil. As the dumpling approaches with steady deliberation, its base first strikes the scalding broth; instantly, the surface is forcefully broken, a sharp, distinct ripple spreading outward from the point of impact. Countless droplets of scalding broth explode into the air, condensing into crystalline spheres before slowly falling. The dumplings then sank, completely immersed in the rich, oily soup. The entire scene, captured in extreme slow motion, transformed a simple cooking process into a visual feast full of power and beauty, creating an ultimate sensory enjoyment and anticipation for delicious food.", + "seed": 1943, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_1/index_7887_seed_1943_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12018515476956963, + "vlm_logprob": -0.12018515476956963, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.12018515476956963, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4749256670475006 + ], + "opposite_logprob_tokens": [ + -1.7249256372451782 + ], + "logprob_reward": -0.4749256670475006, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.4749256670475006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31644541025161743, + 0.08220554888248444, + 0.07032278925180435, + 0.033572204411029816, + 0.03990992531180382, + 0.04213637858629227, + 0.027014195919036865, + 0.04064686596393585, + 0.027896426618099213, + 0.31985026597976685 + ], + "credit_map": [ + 0.02811632677912712, + 0.05026087909936905, + 0.02460002526640892, + 0.021701334044337273, + 0.07644471526145935, + 0.026366345584392548, + 0.06551367789506912, + 0.023442093282938004, + 0.00930114183574915, + 0.009352922439575195, + 0.00878116488456726, + 0.014256823807954788, + 0.010880139656364918, + 0.010289283469319344, + 0.011887210421264172, + 0.007456862833350897, + 0.009562616236507893, + 0.008156973868608475, + 0.007846116088330746, + 0.011268560774624348, + 0.006569590885192156, + 0.008654117584228516, + 0.011153102852404118, + 0.007111707236617804, + 0.003490868490189314, + 0.005128428805619478, + 0.0027489501517266035, + 0.003214138327166438, + 0.006800936069339514, + 0.005041105672717094, + 0.004934273194521666, + 0.0022135016042739153, + 0.0040422058664262295, + 0.008594376966357231, + 0.00435966532677412, + 0.0031941747292876244, + 0.0035342825576663017, + 0.008897896856069565, + 0.004526493139564991, + 0.0027608280070126057, + 0.0037670277524739504, + 0.006549786310642958, + 0.004277170170098543, + 0.004683062434196472, + 0.004727168008685112, + 0.003933899104595184, + 0.010852348059415817, + 0.0033459155820310116, + 0.004331932868808508, + 0.0042950063943862915, + 0.0035691119264811277, + 0.00262136640958488, + 0.002477434463799, + 0.0041696722619235516, + 0.0030946519691497087, + 0.0024550193920731544, + 0.007326014339923859, + 0.00303562730550766, + 0.006458431016653776, + 0.003136737272143364, + 0.003062198869884014, + 0.005322489887475967, + 0.0079261539503932, + 0.00437921192497015, + 0.006631222553551197, + 0.00204650079831481, + 0.0018560048192739487, + 0.002510841703042388, + 0.006715985480695963, + 0.0022703944705426693, + 0.002265306655317545, + 0.003600169438868761, + 0.09365440160036087, + 0.019533848389983177, + 0.008524793200194836, + 0.009505954571068287, + 0.022524654865264893, + 0.13832607865333557, + 0.018619369715452194, + 0.009161173366010189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.4749256670475006 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012642241083085537 + ], + "opposite_logprob_tokens": [ + -4.3876423835754395 + ], + "logprob_reward": -0.012642241083085537, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012642241083085537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4964304566383362, + 0.1338978260755539, + 0.09057381004095078, + 0.06909100711345673, + 0.02700410597026348, + 0.060096319764852524, + 0.026725182309746742, + 0.015548828057944775, + 0.028372693806886673, + 0.05225975066423416 + ], + "credit_map": [ + 0.038955479860305786, + 0.07352977991104126, + 0.04027488827705383, + 0.032879650592803955, + 0.12401285767555237, + 0.05469407141208649, + 0.09631504863500595, + 0.03576871007680893, + 0.014601306989789009, + 0.011888261884450912, + 0.016616201028227806, + 0.02177179604768753, + 0.017090339213609695, + 0.019670946523547173, + 0.0190216563642025, + 0.013237315230071545, + 0.012745395302772522, + 0.009883149527013302, + 0.009685736149549484, + 0.013932780362665653, + 0.009429632686078548, + 0.011916818097233772, + 0.01597435213625431, + 0.007005949039012194, + 0.003829280147328973, + 0.00659715523943305, + 0.0030964843463152647, + 0.005813234485685825, + 0.019097238779067993, + 0.010205952450633049, + 0.01614787057042122, + 0.0043037887662649155, + 0.0022160036023706198, + 0.0045156218111515045, + 0.0032073426991701126, + 0.003307277336716652, + 0.0026623583398759365, + 0.004032785538583994, + 0.004748825915157795, + 0.0023138916585594416, + 0.005973658990114927, + 0.010962276719510555, + 0.0040697017684578896, + 0.004831234458833933, + 0.007759976666420698, + 0.007442113012075424, + 0.014871454797685146, + 0.004185901954770088, + 0.002798460889607668, + 0.0016471354756504297, + 0.006045876536518335, + 0.002137057948857546, + 0.0020204251632094383, + 0.0022886807564646006, + 0.00591708067804575, + 0.0038704662583768368, + 0.00287894275970757, + 0.000934881973080337, + 0.002866947092115879, + 0.0014384894166141748, + 0.0010353177785873413, + 0.001859943731687963, + 0.002610847819596529, + 0.0019234574865549803, + 0.004937831312417984, + 0.0022825368214398623, + 0.002604900160804391, + 0.0038146746810525656, + 0.006510407198220491, + 0.00203697825782001, + 0.003917443100363016, + 0.0022679215762764215, + 0.008160336874425411, + 0.003339136252179742, + 0.002547327894717455, + 0.003109591081738472, + 0.005619138479232788, + 0.01982186734676361, + 0.0073463451117277145, + 0.002316003432497382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.012642241083085537 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032194233499467373 + ], + "opposite_logprob_tokens": [ + -5.7532196044921875 + ], + "logprob_reward": -0.0032194233499467373, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032194233499467373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35529088973999023, + 0.06336687505245209, + 0.2893843352794647, + 0.11008495837450027, + 0.024658219888806343, + 0.03895673528313637, + 0.034566327929496765, + 0.017118997871875763, + 0.014184744097292423, + 0.052387822419404984 + ], + "credit_map": [ + 0.03557106480002403, + 0.0583021305501461, + 0.027294820174574852, + 0.024230848997831345, + 0.08014847338199615, + 0.030834997072815895, + 0.07147688418626785, + 0.027431726455688477, + 0.005352220498025417, + 0.006539649795740843, + 0.005977372173219919, + 0.014380217529833317, + 0.008652498945593834, + 0.0074053602293133736, + 0.009549569338560104, + 0.005509996321052313, + 0.030732886865735054, + 0.021488100290298462, + 0.019748356193304062, + 0.08428938686847687, + 0.021071450784802437, + 0.022832516580820084, + 0.04747414216399193, + 0.0417475551366806, + 0.006428696680814028, + 0.00959694106131792, + 0.004929298534989357, + 0.009069281630218029, + 0.0308474600315094, + 0.016295792534947395, + 0.025962505489587784, + 0.006954994983971119, + 0.002731167711317539, + 0.0037190529983490705, + 0.00259167212061584, + 0.0018955398118123412, + 0.0023266626521945, + 0.005338659510016441, + 0.003370319027453661, + 0.002685150131583214, + 0.006133509799838066, + 0.004073154181241989, + 0.0029407301917672157, + 0.0035343319177627563, + 0.00395699730142951, + 0.005647375248372555, + 0.00846308283507824, + 0.004207555204629898, + 0.004774548579007387, + 0.0029414165765047073, + 0.004274438600987196, + 0.0035653910599648952, + 0.003198277438059449, + 0.004888109862804413, + 0.008007903583347797, + 0.0029162424616515636, + 0.0037883061449974775, + 0.0010866052471101284, + 0.0029649061616510153, + 0.0010711344657465816, + 0.0011163675226271152, + 0.004279611632227898, + 0.0014972181525081396, + 0.0013148508733138442, + 0.002085230080410838, + 0.001055234926752746, + 0.001580788753926754, + 0.001090299105271697, + 0.0027351982425898314, + 0.001277698902413249, + 0.0012751618633046746, + 0.0030851331539452076, + 0.00936932023614645, + 0.0032524701673537493, + 0.0026128098834306, + 0.0018008212791755795, + 0.009574145078659058, + 0.0212896466255188, + 0.0029082875698804855, + 0.0015803264686837792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0032194233499467373 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034428536891937 + ], + "opposite_logprob_tokens": [ + -2.350344181060791 + ], + "logprob_reward": -0.10034428536891937, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034428536891937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17019641399383545, + 0.12482567876577377, + 0.2272079586982727, + 0.07119631767272949, + 0.044398169964551926, + 0.03197615593671799, + 0.03160276263952255, + 0.03397468850016594, + 0.012159358710050583, + 0.2524624764919281 + ], + "credit_map": [ + 0.02125098556280136, + 0.029861073940992355, + 0.01379500050097704, + 0.013893317431211472, + 0.031457848846912384, + 0.019941989332437515, + 0.025348857045173645, + 0.014647345058619976, + 0.013327391818165779, + 0.01187910046428442, + 0.015521783381700516, + 0.019824272021651268, + 0.011777222156524658, + 0.018062971532344818, + 0.019433068111538887, + 0.01499987579882145, + 0.019120214506983757, + 0.019174670800566673, + 0.01958322897553444, + 0.05932752043008804, + 0.015266521833837032, + 0.02011663280427456, + 0.04186426103115082, + 0.03275492042303085, + 0.004912903066724539, + 0.00803509820252657, + 0.0035421806387603283, + 0.007243033964186907, + 0.016007229685783386, + 0.009262914769351482, + 0.017679505050182343, + 0.004513452760875225, + 0.0033602819312363863, + 0.01134208682924509, + 0.004374636337161064, + 0.0039168973453342915, + 0.00454030791297555, + 0.009989223442971706, + 0.00460705766454339, + 0.0022676833905279636, + 0.0037513712886720896, + 0.0059958649799227715, + 0.003073255531489849, + 0.0034175454638898373, + 0.003117646323516965, + 0.0022553857415914536, + 0.008134465664625168, + 0.002230621874332428, + 0.003822910599410534, + 0.0025450431276112795, + 0.006040735170245171, + 0.0028273463249206543, + 0.0023663288448005915, + 0.003196120262145996, + 0.008207467384636402, + 0.002596810460090637, + 0.007442651782184839, + 0.002139362273737788, + 0.00535648362711072, + 0.002354721073061228, + 0.0022916982416063547, + 0.007261664606630802, + 0.0039776163175702095, + 0.0031504943035542965, + 0.0013398317387327552, + 0.0007556256605312228, + 0.0016834456473588943, + 0.001947058248333633, + 0.0012864685850217938, + 0.001754580531269312, + 0.002288357587531209, + 0.0011039909441024065, + 0.058712899684906006, + 0.014512463472783566, + 0.00913960486650467, + 0.0076823728159070015, + 0.029370911419391632, + 0.10924032330513, + 0.016671879217028618, + 0.0071320501156151295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.10034428536891937 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979415699839592 + ], + "opposite_logprob_tokens": [ + -4.634794235229492 + ], + "logprob_reward": -0.00979415699839592, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979415699839592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1762620508670807, + 0.042537424713373184, + 0.05680585652589798, + 0.18083658814430237, + 0.01718999445438385, + 0.037918079644441605, + 0.08501116931438446, + 0.023276319727301598, + 0.012313232757151127, + 0.3678492605686188 + ], + "credit_map": [ + 0.017497848719358444, + 0.02548401616513729, + 0.014050980098545551, + 0.01285938173532486, + 0.04049023985862732, + 0.015329182147979736, + 0.03728185221552849, + 0.013268549926578999, + 0.005608425009995699, + 0.004046303685754538, + 0.0039473106153309345, + 0.006813571788370609, + 0.004547442309558392, + 0.006090708542615175, + 0.007494066376239061, + 0.003989601042121649, + 0.003982325084507465, + 0.003124961396679282, + 0.0034623839892446995, + 0.0167239997535944, + 0.004687726963311434, + 0.005290120840072632, + 0.006596974562853575, + 0.012937367893755436, + 0.009182633832097054, + 0.01517373975366354, + 0.005647450685501099, + 0.015088260173797607, + 0.054018836468458176, + 0.025342049077153206, + 0.04549210146069527, + 0.010891521349549294, + 0.0022868607193231583, + 0.003782060230150819, + 0.0018451811047270894, + 0.0010379436425864697, + 0.002291197655722499, + 0.0031615167390555143, + 0.0017062395345419645, + 0.00107899634167552, + 0.003646828467026353, + 0.006330467294901609, + 0.0036100042052567005, + 0.004336201585829258, + 0.0057881963439285755, + 0.0036894874647259712, + 0.007639389019459486, + 0.002877508057281375, + 0.00916200503706932, + 0.0075594293884932995, + 0.015320423059165478, + 0.008543881587684155, + 0.008556044660508633, + 0.007430296391248703, + 0.018853027373552322, + 0.009586062282323837, + 0.004567594733089209, + 0.001282611396163702, + 0.003972713835537434, + 0.001404088456183672, + 0.0015113059198483825, + 0.005623652134090662, + 0.002388682682067156, + 0.002525672782212496, + 0.0022066948004066944, + 0.0006738448864780366, + 0.0013616709038615227, + 0.0014295391738414764, + 0.002403547754511237, + 0.0010665630688890815, + 0.0016505540115758777, + 0.0015208185650408268, + 0.0966981053352356, + 0.021372618153691292, + 0.010258271358907223, + 0.010331004858016968, + 0.0334060974419117, + 0.16127577424049377, + 0.02454637549817562, + 0.009961011819541454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.00979415699839592 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, filmed in high-frame-rate slow motion, captures a plump, hand-formed dumpling, its belly plump, plump and plump, plumply falling from the air. A beam of warm overhead light, precisely illuminating the thin, translucent wrapper, reveals its delicate folds and the subtle trembling of its form under the weight of gravity. Beneath it lies a large, dark ceramic bowl, filled with scalding, deep-brown broth. The broth trembles subtly from the heat, and plumes of steam rise, surfacing with a few emerald green scallions and a few drops of glistening sesame oil. As the dumpling approaches with steady deliberation, its base first strikes the scalding broth; instantly, the surface is forcefully broken, a sharp, distinct ripple spreading outward from the point of impact. Countless droplets of scalding broth explode into the air, condensing into crystalline spheres before slowly falling. The dumplings then sank, completely immersed in the rich, oily soup. The entire scene, captured in extreme slow motion, transformed a simple cooking process into a visual feast full of power and beauty, creating an ultimate sensory enjoyment and anticipation for delicious food.", + "seed": 1944, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_1/index_7887_seed_1944_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.054987637838348744, + "vlm_logprob": -0.054987637838348744, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.054987637838348744, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22932888567447662 + ], + "opposite_logprob_tokens": [ + -3.1043288707733154 + ], + "logprob_reward": -0.22932888567447662, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.22932888567447662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17446166276931763, + 0.12548686563968658, + 0.06794604659080505, + 0.2545749545097351, + 0.04252537712454796, + 0.14446036517620087, + 0.06110657751560211, + 0.04232492297887802, + 0.04303186386823654, + 0.04408128932118416 + ], + "credit_map": [ + 0.014765088446438313, + 0.017821749672293663, + 0.038736045360565186, + 0.02457953244447708, + 0.01636633649468422, + 0.033033981919288635, + 0.013452240265905857, + 0.015706706792116165, + 0.012096631340682507, + 0.011463112197816372, + 0.022852515801787376, + 0.01482471264898777, + 0.020005185157060623, + 0.015768330544233322, + 0.014889560639858246, + 0.013586834073066711, + 0.005110620986670256, + 0.008208227343857288, + 0.008193694055080414, + 0.009201805107295513, + 0.006848329212516546, + 0.011078711599111557, + 0.00991319864988327, + 0.009391468949615955, + 0.02175949513912201, + 0.029931126162409782, + 0.026264606043696404, + 0.02672644518315792, + 0.03470434248447418, + 0.031873296946287155, + 0.04329313337802887, + 0.040022507309913635, + 0.004888293799012899, + 0.006004262249916792, + 0.0036907794419676065, + 0.005759375169873238, + 0.009182373993098736, + 0.003842991776764393, + 0.004747213330119848, + 0.004410087130963802, + 0.01594812609255314, + 0.023491842672228813, + 0.013135605491697788, + 0.021486680954694748, + 0.01716284267604351, + 0.015791941434144974, + 0.023912249132990837, + 0.013531087897717953, + 0.009657755494117737, + 0.005776063073426485, + 0.011742453090846539, + 0.0066914428025484085, + 0.008304242976009846, + 0.006467653904110193, + 0.006534056272357702, + 0.0059329126961529255, + 0.004929046146571636, + 0.0038934010080993176, + 0.006399994250386953, + 0.0039284504018723965, + 0.00912027433514595, + 0.0035177674144506454, + 0.007218455895781517, + 0.003317534690722823, + 0.008420783095061779, + 0.00310477614402771, + 0.01291086245328188, + 0.00565739581361413, + 0.003851744346320629, + 0.0024078430142253637, + 0.003874851856380701, + 0.0028036129660904408, + 0.004523600451648235, + 0.005733314901590347, + 0.006006265990436077, + 0.00573505274951458, + 0.003779306076467037, + 0.003153413301333785, + 0.0060882833786308765, + 0.009062059223651886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.22932888567447662 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005320792552083731 + ], + "opposite_logprob_tokens": [ + -5.255321025848389 + ], + "logprob_reward": -0.005320792552083731, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005320792552083731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2285824567079544, + 0.16796456277370453, + 0.05597816780209541, + 0.18187949061393738, + 0.04796505719423294, + 0.03304869681596756, + 0.1554582715034485, + 0.079042449593544, + 0.027164986357092857, + 0.022915825247764587 + ], + "credit_map": [ + 0.016201579943299294, + 0.02988799661397934, + 0.028504256159067154, + 0.021266449242830276, + 0.03578735515475273, + 0.028286674991250038, + 0.040266022086143494, + 0.028382133692502975, + 0.011735077947378159, + 0.01175492163747549, + 0.03160988166928291, + 0.023522069677710533, + 0.0214682724326849, + 0.016332341358065605, + 0.02991575561463833, + 0.02162625640630722, + 0.0037901110481470823, + 0.0061346422880887985, + 0.007586523424834013, + 0.00819545891135931, + 0.005331128835678101, + 0.007841321639716625, + 0.010832327418029308, + 0.006266655400395393, + 0.01974991336464882, + 0.019385838881134987, + 0.018652353435754776, + 0.020403733476996422, + 0.0258775744587183, + 0.024099377915263176, + 0.024408778175711632, + 0.02930193394422531, + 0.0051406859420239925, + 0.006031638476997614, + 0.004844252951443195, + 0.006702409125864506, + 0.006348746363073587, + 0.004201255738735199, + 0.00966199953109026, + 0.005034072324633598, + 0.003974395338445902, + 0.005457967519760132, + 0.003418560838326812, + 0.0030677062459290028, + 0.004141871817409992, + 0.005723302718251944, + 0.0040075215511024, + 0.0032573738135397434, + 0.015561940148472786, + 0.009895749390125275, + 0.022770969197154045, + 0.012847141362726688, + 0.03145498037338257, + 0.0386233888566494, + 0.014502127654850483, + 0.009801971726119518, + 0.0071392664685845375, + 0.007008318789303303, + 0.008466292172670364, + 0.008139372803270817, + 0.022948388010263443, + 0.009848440997302532, + 0.010327853262424469, + 0.005164517555385828, + 0.0029668526258319616, + 0.001978356624022126, + 0.005140652414411306, + 0.0047915540635585785, + 0.00292890053242445, + 0.0025527740363031626, + 0.003651566104963422, + 0.003154332749545574, + 0.002472776221111417, + 0.0019393343245610595, + 0.0034889753442257643, + 0.003222772851586342, + 0.0020555316004902124, + 0.0015561622567474842, + 0.0030703917145729065, + 0.005109881516546011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.005320792552083731 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002842435846105218 + ], + "opposite_logprob_tokens": [ + -5.877842426300049 + ], + "logprob_reward": -0.002842435846105218, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002842435846105218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1296691596508026, + 0.4181674122810364, + 0.035070788115262985, + 0.120721235871315, + 0.10440778732299805, + 0.038024917244911194, + 0.03245333954691887, + 0.03675088658928871, + 0.02409266121685505, + 0.060641828924417496 + ], + "credit_map": [ + 0.009829680435359478, + 0.01657765544950962, + 0.016657022759318352, + 0.013078847900032997, + 0.015035799704492092, + 0.033465053886175156, + 0.010989085771143436, + 0.014036010950803757, + 0.04576966166496277, + 0.02453302964568138, + 0.08309020847082138, + 0.0698249340057373, + 0.02272883616387844, + 0.05287572741508484, + 0.05618387833237648, + 0.06316115707159042, + 0.003176976228132844, + 0.005535697564482689, + 0.004806617274880409, + 0.004386384971439838, + 0.003684945171698928, + 0.005534621421247721, + 0.004793844651430845, + 0.0031517022289335728, + 0.012704096734523773, + 0.01671754941344261, + 0.019506528973579407, + 0.014023182913661003, + 0.01570831425487995, + 0.015241139568388462, + 0.013259056955575943, + 0.013561366125941277, + 0.011606178246438503, + 0.010798821225762367, + 0.010494839400053024, + 0.01681920699775219, + 0.011971871368587017, + 0.01319777313619852, + 0.018714239820837975, + 0.010804859921336174, + 0.005447842180728912, + 0.0059507484547793865, + 0.003591056913137436, + 0.006059480365365744, + 0.003516388824209571, + 0.0058259181678295135, + 0.004539352376013994, + 0.003094131126999855, + 0.0044544292613863945, + 0.0030461966525763273, + 0.0036193705163896084, + 0.005202402826398611, + 0.0032325414940714836, + 0.006554313004016876, + 0.003602426266297698, + 0.0027416583616286516, + 0.003934347070753574, + 0.003646674333140254, + 0.004295801743865013, + 0.002741843229159713, + 0.010604405775666237, + 0.0034891951363533735, + 0.005828975699841976, + 0.0022096417378634214, + 0.0025338323321193457, + 0.0018508783541619778, + 0.004968430381268263, + 0.00504404166713357, + 0.0018068787176162004, + 0.0020488512236624956, + 0.003456250298768282, + 0.002383497543632984, + 0.005651574116200209, + 0.0054916683584451675, + 0.008628712967038155, + 0.008817149326205254, + 0.005908161401748657, + 0.004120125900954008, + 0.007308217231184244, + 0.01471622008830309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.002842435846105218 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635977976024151 + ], + "opposite_logprob_tokens": [ + -3.651359796524048 + ], + "logprob_reward": -0.02635977976024151, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635977976024151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11710147559642792, + 0.15634356439113617, + 0.0928240716457367, + 0.3502369821071625, + 0.0668112114071846, + 0.05968675762414932, + 0.08651634305715561, + 0.02253798022866249, + 0.009286388754844666, + 0.03865520656108856 + ], + "credit_map": [ + 0.007091217674314976, + 0.014429439790546894, + 0.01525929942727089, + 0.011868804693222046, + 0.015675267204642296, + 0.025266285985708237, + 0.01585955172777176, + 0.01165161281824112, + 0.014230853877961636, + 0.012468647211790085, + 0.03678052872419357, + 0.018627658486366272, + 0.017678963020443916, + 0.01909484528005123, + 0.017557239159941673, + 0.01990482397377491, + 0.006590923760086298, + 0.016927968710660934, + 0.00813319906592369, + 0.011941577307879925, + 0.009778041392564774, + 0.015621872618794441, + 0.014355224557220936, + 0.009475264698266983, + 0.04314465820789337, + 0.05800166353583336, + 0.03290863707661629, + 0.042721226811409, + 0.04326114058494568, + 0.049676284193992615, + 0.03962445259094238, + 0.0408988893032074, + 0.008086372166872025, + 0.008016190491616726, + 0.007285089697688818, + 0.00875215232372284, + 0.01189704891294241, + 0.008422638289630413, + 0.008019519969820976, + 0.006332207005470991, + 0.0072086649015545845, + 0.012200440280139446, + 0.005298237781971693, + 0.006644324399530888, + 0.007164010778069496, + 0.008699010126292706, + 0.008033125661313534, + 0.004438945557922125, + 0.007188879419118166, + 0.004722650162875652, + 0.010824154131114483, + 0.005879526026546955, + 0.019426677376031876, + 0.027493147179484367, + 0.0068564219400286674, + 0.004124889150261879, + 0.0022763460874557495, + 0.0017673810943961143, + 0.0036057254765182734, + 0.0034393765963613987, + 0.004058997146785259, + 0.002792718354612589, + 0.0028316592797636986, + 0.0017657745629549026, + 0.0009909163927659392, + 0.0007940584328025579, + 0.001255176728591323, + 0.0015597152523696423, + 0.0012280605733394623, + 0.0008600142900831997, + 0.0015244226669892669, + 0.0010740243596956134, + 0.0035715829581022263, + 0.002131849527359009, + 0.005475434008985758, + 0.00538540817797184, + 0.004141016397625208, + 0.002742792945355177, + 0.005461597815155983, + 0.009745524264872074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.02635977976024151 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video filmed in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a plump, hand-formed dumpling visible falling from the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, dark ceramic bowl filled with deep-brown broth visible beneath the falling dumpling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dumpling strike the broth, causing a sharp ripple and exploding droplets that condense into crystalline spheres?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of warm overhead light revealing the translucent wrapper and steam with emerald green scallions?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108629535883665 + ], + "opposite_logprob_tokens": [ + -4.511086463928223 + ], + "logprob_reward": -0.01108629535883665, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108629535883665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2581046223640442, + 0.09822731465101242, + 0.07706565409898758, + 0.1611107885837555, + 0.04905496910214424, + 0.04769698902964592, + 0.08871161937713623, + 0.03790010139346123, + 0.06539929658174515, + 0.11672873795032501 + ], + "credit_map": [ + 0.028809688985347748, + 0.022489773109555244, + 0.0204025786370039, + 0.026444533839821815, + 0.028876569122076035, + 0.05556298792362213, + 0.03668249025940895, + 0.03883596137166023, + 0.009480725973844528, + 0.008441879414021969, + 0.017243357375264168, + 0.011228709481656551, + 0.013144520111382008, + 0.010579369030892849, + 0.014767254702746868, + 0.013341485522687435, + 0.007984058931469917, + 0.007475574966520071, + 0.016461927443742752, + 0.010332995094358921, + 0.006671295966953039, + 0.00925518199801445, + 0.008473675698041916, + 0.010410935617983341, + 0.018526345491409302, + 0.025629552081227303, + 0.01700577884912491, + 0.014848754741251469, + 0.016041072085499763, + 0.022777942940592766, + 0.02719077840447426, + 0.019090546295046806, + 0.006913990713655949, + 0.00459067290648818, + 0.004847068339586258, + 0.00580702954903245, + 0.005032673478126526, + 0.006756108719855547, + 0.009945805184543133, + 0.0051616160199046135, + 0.007911395281553268, + 0.008881687186658382, + 0.0043634711764752865, + 0.005839157849550247, + 0.004685761872678995, + 0.006138155702501535, + 0.005658088717609644, + 0.004219268914312124, + 0.01625199429690838, + 0.010031372308731079, + 0.01676729880273342, + 0.010080305859446526, + 0.007125047035515308, + 0.013559226877987385, + 0.008432317525148392, + 0.006464048288762569, + 0.005073621403425932, + 0.0025851158425211906, + 0.005076572299003601, + 0.005938222631812096, + 0.0052018738351762295, + 0.0032505413983017206, + 0.007276451680809259, + 0.0034976990427821875, + 0.014305093325674534, + 0.004472809843719006, + 0.021646620705723763, + 0.012494443915784359, + 0.003357431385666132, + 0.002623798558488488, + 0.0038988550659269094, + 0.0026002416852861643, + 0.02112388238310814, + 0.013173546642065048, + 0.02063830755650997, + 0.014922967180609703, + 0.007355015724897385, + 0.007260476239025593, + 0.01224807184189558, + 0.020006457343697548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.01108629535883665 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_10/index_tensor([3032, 3032])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_10/index_tensor([3032, 3032])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c2aa715677c54bb73a5d108b9986ffaa2906a824 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_10/index_tensor([3032, 3032])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a close-up shot framed in vertical format, the camera remains stationary, focused on an exquisite folding fan, initially folded. Set against a pure, dark background, this scene creates a minimalist, stage-like feel. With a swift, decisive movement, the fan, crafted from polished, dark bamboo ribs, unfolds completely. Its fluid movement instantly fills the frame, ultimately forming a perfect semicircle. A soft, directional light, directed from above and sideways, subtly illuminates the off-white rice paper, bringing to life every detail of a profoundly poetic Chinese ink landscape painting: the outlines of distant mountains looming through the mist, the delicate gradations of ink smudged across the paper, and the subtle shifts in wet and dry brushstrokes. The light also casts subtle shadows within the folds formed by the fan's unfolding ribs, enhancing the image's sense of three-dimensionality and texture. The entire scene is imbued with a tranquil and elegant atmosphere of classical Eastern aesthetics, conveying a dedication to and appreciation of traditional craftsmanship and artistry.", + "seed": 18061, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_10/index_3032_seed_18061_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02946914443746209, + "vlm_logprob": -0.02946914443746209, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.02946914443746209, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0227615125477314 + ], + "opposite_logprob_tokens": [ + -5.022761344909668 + ], + "logprob_reward": -0.0227615125477314, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0227615125477314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4918544590473175, + 0.1209220364689827, + 0.060453467071056366, + 0.034737471491098404, + 0.04970116913318634, + 0.03668821230530739, + 0.054545823484659195, + 0.03347772732377052, + 0.08661475777626038, + 0.03100491315126419 + ], + "credit_map": [ + 0.0334983691573143, + 0.0893411859869957, + 0.03545146435499191, + 0.04891481623053551, + 0.07167855650186539, + 0.08078896999359131, + 0.07463598996400833, + 0.05754505470395088, + 0.0157534871250391, + 0.017436930909752846, + 0.0138945197686553, + 0.012530663050711155, + 0.01316078007221222, + 0.01612907648086548, + 0.016047457233071327, + 0.01596911810338497, + 0.006836694665253162, + 0.0046496307477355, + 0.00714443763718009, + 0.012727407738566399, + 0.005664739292114973, + 0.006731872446835041, + 0.00819997489452362, + 0.008498706854879856, + 0.003343339776620269, + 0.003365324577316642, + 0.003289290936663747, + 0.007048643194139004, + 0.003944499418139458, + 0.0038315467536449432, + 0.00535199511796236, + 0.004562830086797476, + 0.0045655155554413795, + 0.005340940319001675, + 0.006407212931662798, + 0.01125079020857811, + 0.0037875203415751457, + 0.004412093665450811, + 0.007468624040484428, + 0.006468466017395258, + 0.00490052904933691, + 0.00517973443493247, + 0.004443076904863119, + 0.0060294559225440025, + 0.0031527618411928415, + 0.003993855323642492, + 0.004487550351768732, + 0.004501244053244591, + 0.005516604520380497, + 0.01132828090339899, + 0.006051451433449984, + 0.00844917818903923, + 0.0063209934160113335, + 0.004884595982730389, + 0.006326710339635611, + 0.005668003112077713, + 0.003579242154955864, + 0.0039540426805615425, + 0.0038478418719023466, + 0.006318367086350918, + 0.0027948461938649416, + 0.004575998056679964, + 0.0032382651697844267, + 0.005169120151549578, + 0.008478717878460884, + 0.013855691067874432, + 0.010608270764350891, + 0.015444060787558556, + 0.008191976696252823, + 0.008216206915676594, + 0.010993401519954205, + 0.010826424695551395, + 0.004286696203052998, + 0.003545901272445917, + 0.006031142547726631, + 0.004853500053286552, + 0.002874162048101425, + 0.003002662444487214, + 0.00270300661213696, + 0.0037078398745507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0227615125477314 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498543754220009 + ], + "opposite_logprob_tokens": [ + -2.929985523223877 + ], + "logprob_reward": -0.05498543754220009, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498543754220009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8067373633384705, + 0.039623212069272995, + 0.019327333196997643, + 0.018515292555093765, + 0.03291740268468857, + 0.03357277810573578, + 0.009358384646475315, + 0.006961971055716276, + 0.02053850330412388, + 0.012447849847376347 + ], + "credit_map": [ + 0.056514620780944824, + 0.1072600707411766, + 0.05998614430427551, + 0.07955454289913177, + 0.12850527465343475, + 0.10915966331958771, + 0.17342160642147064, + 0.09233538061380386, + 0.0030612091068178415, + 0.006718511693179607, + 0.004971846006810665, + 0.005738716572523117, + 0.004211021587252617, + 0.004942818079143763, + 0.004962746985256672, + 0.005016338545829058, + 0.002243932569399476, + 0.0016572457971051335, + 0.0018604828510433435, + 0.0026875261683017015, + 0.003098131390288472, + 0.0032062807586044073, + 0.001876346068456769, + 0.0026973849162459373, + 0.0020968206226825714, + 0.0021682886872440577, + 0.0019196836510673165, + 0.002963657258078456, + 0.0020269991364330053, + 0.002217688597738743, + 0.003026526188477874, + 0.002095626899972558, + 0.0027969107031822205, + 0.003386198543012142, + 0.004306315910071135, + 0.005401256959885359, + 0.0030647823587059975, + 0.0028769648633897305, + 0.006627189926803112, + 0.004457779694348574, + 0.002914159558713436, + 0.005124532151967287, + 0.0036983489990234375, + 0.005136432591825724, + 0.005031609442085028, + 0.002770191989839077, + 0.004995096009224653, + 0.003902407130226493, + 0.0011677814181894064, + 0.0014413651078939438, + 0.0009866729378700256, + 0.0014673449331894517, + 0.0011558225378394127, + 0.0009347085142508149, + 0.0009722273098304868, + 0.0012324617709964514, + 0.0008877130458131433, + 0.0010401512263342738, + 0.0007531526498496532, + 0.0009857268305495381, + 0.0009078243747353554, + 0.0008032905170693994, + 0.0007588472799398005, + 0.0008252646657638252, + 0.0030093432869762182, + 0.0024948047939687967, + 0.003508524736389518, + 0.0030371868051588535, + 0.0016247316962108016, + 0.0024547891225665808, + 0.002097874414175749, + 0.0023112462367862463, + 0.0012874884996563196, + 0.001911667175590992, + 0.001778008765541017, + 0.0016343083698302507, + 0.0015699634095653892, + 0.001685163239017129, + 0.001156912068836391, + 0.001424337038770318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.05498543754220009 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048663295805454254 + ], + "opposite_logprob_tokens": [ + -3.048663377761841 + ], + "logprob_reward": -0.048663295805454254, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048663295805454254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30339014530181885, + 0.1224546954035759, + 0.0770685151219368, + 0.0978628396987915, + 0.14294634759426117, + 0.05373288691043854, + 0.04852961376309395, + 0.037638094276189804, + 0.08790145814418793, + 0.028475424274802208 + ], + "credit_map": [ + 0.021630922332406044, + 0.03947699815034866, + 0.017133962363004684, + 0.023060552775859833, + 0.05306653678417206, + 0.05684322491288185, + 0.06109195575118065, + 0.031085986644029617, + 0.009769147261977196, + 0.013188335113227367, + 0.016309769824147224, + 0.013130947947502136, + 0.011232909746468067, + 0.02300022542476654, + 0.017296185716986656, + 0.01852717436850071, + 0.009579978883266449, + 0.0059634312056005, + 0.008236822672188282, + 0.00911523774266243, + 0.01091835368424654, + 0.01213507167994976, + 0.010752351954579353, + 0.010367265902459621, + 0.010477697476744652, + 0.012167656794190407, + 0.007404729258269072, + 0.018628867343068123, + 0.01135240774601698, + 0.013747639022767544, + 0.011811736971139908, + 0.012272101826965809, + 0.011522592976689339, + 0.023689959198236465, + 0.015881555154919624, + 0.02234378084540367, + 0.01746096834540367, + 0.01348381582647562, + 0.023109326139092445, + 0.015454357489943504, + 0.006027454510331154, + 0.00702942768111825, + 0.004749700427055359, + 0.008341451175510883, + 0.0073044621385633945, + 0.004427020903676748, + 0.009537626057863235, + 0.006315742153674364, + 0.005366191267967224, + 0.006507044658064842, + 0.006026507820934057, + 0.007016980554908514, + 0.004034098703414202, + 0.00481144106015563, + 0.00881347618997097, + 0.005953873507678509, + 0.002943582134321332, + 0.006416314281523228, + 0.0032729858066886663, + 0.006781737320125103, + 0.003866722574457526, + 0.0050338334403932095, + 0.0045167626813054085, + 0.004806153941899538, + 0.011940393596887589, + 0.011555895209312439, + 0.015826981514692307, + 0.01411300990730524, + 0.005578547716140747, + 0.007232446689158678, + 0.008567260578274727, + 0.013086928986012936, + 0.0032408544793725014, + 0.004583251662552357, + 0.0041190641932189465, + 0.004578853957355022, + 0.003134850412607193, + 0.0031641595996916294, + 0.0026654102839529514, + 0.002988979918882251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.048663295805454254 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200471341609955 + ], + "opposite_logprob_tokens": [ + -4.264200687408447 + ], + "logprob_reward": -0.014200471341609955, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200471341609955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21784932911396027, + 0.14334332942962646, + 0.15857580304145813, + 0.05908456817269325, + 0.08642205595970154, + 0.16228768229484558, + 0.03328938037157059, + 0.06348367780447006, + 0.045990005135536194, + 0.02967417612671852 + ], + "credit_map": [ + 0.01695018820464611, + 0.030435940250754356, + 0.01819448731839657, + 0.024010397493839264, + 0.03275100886821747, + 0.031921084970235825, + 0.03957482799887657, + 0.0240114014595747, + 0.013711215928196907, + 0.015847550705075264, + 0.01343376561999321, + 0.016725005581974983, + 0.015326011925935745, + 0.019321471452713013, + 0.026775212958455086, + 0.02220308966934681, + 0.015498747117817402, + 0.012633372098207474, + 0.01902543567121029, + 0.0329914465546608, + 0.021824195981025696, + 0.01672656089067459, + 0.017081592231988907, + 0.022794459015130997, + 0.0060224952176213264, + 0.0065652173943817616, + 0.005749829113483429, + 0.00981899257749319, + 0.007250111550092697, + 0.0073423259891569614, + 0.009734776802361012, + 0.006600816268473864, + 0.0078013623133301735, + 0.009868023917078972, + 0.007818611338734627, + 0.013711384497582912, + 0.00927530974149704, + 0.007678586058318615, + 0.020362891256809235, + 0.009905893355607986, + 0.016147881746292114, + 0.021393729373812675, + 0.01477392390370369, + 0.03132785111665726, + 0.02203580178320408, + 0.015029589645564556, + 0.02303805574774742, + 0.018540849909186363, + 0.003480491228401661, + 0.004712072666734457, + 0.002866115653887391, + 0.007456664927303791, + 0.002429807325825095, + 0.003015080001205206, + 0.004509718623012304, + 0.004819425754249096, + 0.006760092452168465, + 0.010612674988806248, + 0.006642072927206755, + 0.007854457013309002, + 0.006392055656760931, + 0.0072195399552583694, + 0.010465521365404129, + 0.007537260185927153, + 0.00484848627820611, + 0.007578760385513306, + 0.004444117192178965, + 0.007587901782244444, + 0.005518192891031504, + 0.005105678457766771, + 0.006038339342921972, + 0.004868531133979559, + 0.003974506165832281, + 0.002539033303037286, + 0.004423115402460098, + 0.005830735899507999, + 0.002308698371052742, + 0.0029022858943790197, + 0.002989450003951788, + 0.004706351086497307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.014200471341609955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673500495031476 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.00673500495031476, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673500495031476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3705492317676544, + 0.14008766412734985, + 0.07640796899795532, + 0.06200059503316879, + 0.12258899211883545, + 0.04867112636566162, + 0.029205653816461563, + 0.03577640280127525, + 0.0787251815199852, + 0.03598719462752342 + ], + "credit_map": [ + 0.02417103573679924, + 0.05211712419986725, + 0.01525932364165783, + 0.02581198140978813, + 0.0697423592209816, + 0.06261618435382843, + 0.08221154659986496, + 0.03861967846751213, + 0.013602466322481632, + 0.02156849205493927, + 0.013351192697882652, + 0.01711183413863182, + 0.01421479880809784, + 0.016520164906978607, + 0.023046128451824188, + 0.02067258208990097, + 0.005986334756016731, + 0.009502585045993328, + 0.008137919940054417, + 0.011295357719063759, + 0.011554114520549774, + 0.009460563771426678, + 0.01027151569724083, + 0.01019958034157753, + 0.007183784153312445, + 0.007192205637693405, + 0.00588777894154191, + 0.01275013666599989, + 0.005546763073652983, + 0.006368992384523153, + 0.00916539691388607, + 0.007905540987849236, + 0.008301842026412487, + 0.0142386294901371, + 0.01701010949909687, + 0.025577690452337265, + 0.01033819280564785, + 0.009579596109688282, + 0.02073892578482628, + 0.016804002225399017, + 0.004000750370323658, + 0.007132986094802618, + 0.0057773892767727375, + 0.009264362044632435, + 0.0060684266500175, + 0.004520134534686804, + 0.004716459661722183, + 0.00719061866402626, + 0.003583127399906516, + 0.005076599773019552, + 0.003990029916167259, + 0.004220854956656694, + 0.0028593011666089296, + 0.0027816633228212595, + 0.0029236581176519394, + 0.003770416369661689, + 0.0033814511261880398, + 0.005529541522264481, + 0.004453281406313181, + 0.006077566649764776, + 0.00411137193441391, + 0.004748153034597635, + 0.003555688774213195, + 0.003919349517673254, + 0.008554510772228241, + 0.011076460592448711, + 0.010826924815773964, + 0.014582951553165913, + 0.004964191000908613, + 0.007272095885127783, + 0.0096420394256711, + 0.011806007474660873, + 0.00436175474897027, + 0.007080510258674622, + 0.003522925078868866, + 0.0046130889095366, + 0.004542561713606119, + 0.004697275348007679, + 0.0037894686684012413, + 0.003379609202966094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.00673500495031476 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a close-up shot framed in vertical format, the camera remains stationary, focused on an exquisite folding fan, initially folded. Set against a pure, dark background, this scene creates a minimalist, stage-like feel. With a swift, decisive movement, the fan, crafted from polished, dark bamboo ribs, unfolds completely. Its fluid movement instantly fills the frame, ultimately forming a perfect semicircle. A soft, directional light, directed from above and sideways, subtly illuminates the off-white rice paper, bringing to life every detail of a profoundly poetic Chinese ink landscape painting: the outlines of distant mountains looming through the mist, the delicate gradations of ink smudged across the paper, and the subtle shifts in wet and dry brushstrokes. The light also casts subtle shadows within the folds formed by the fan's unfolding ribs, enhancing the image's sense of three-dimensionality and texture. The entire scene is imbued with a tranquil and elegant atmosphere of classical Eastern aesthetics, conveying a dedication to and appreciation of traditional craftsmanship and artistry.", + "seed": 18062, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_10/index_3032_seed_18062_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08583695162087679, + "vlm_logprob": -0.08583695162087679, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.08583695162087679, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03303352743387222 + ], + "opposite_logprob_tokens": [ + -3.908033609390259 + ], + "logprob_reward": -0.03303352743387222, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03303352743387222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05146225169301033, + 0.3213694095611572, + 0.043507784605026245, + 0.26703542470932007, + 0.051278840750455856, + 0.0482252761721611, + 0.0853622630238533, + 0.08703604340553284, + 0.02665952406823635, + 0.018063126131892204 + ], + "credit_map": [ + 0.012201123870909214, + 0.004706626757979393, + 0.00505936611443758, + 0.005864926613867283, + 0.004840855952352285, + 0.004269292578101158, + 0.0063564409501850605, + 0.008163620717823505, + 0.025232605636119843, + 0.12435232102870941, + 0.01809789426624775, + 0.015623431652784348, + 0.03191983699798584, + 0.016553442925214767, + 0.014769652858376503, + 0.07482022047042847, + 0.004140095319598913, + 0.008694210089743137, + 0.0029777311719954014, + 0.005742514505982399, + 0.005391058046370745, + 0.002274710452184081, + 0.0070291426964104176, + 0.00725832674652338, + 0.024027910083532333, + 0.06847056746482849, + 0.01987151801586151, + 0.03778109699487686, + 0.023059813305735588, + 0.013437489978969097, + 0.018651211634278297, + 0.061735834926366806, + 0.010354177094995975, + 0.006562638096511364, + 0.004992418456822634, + 0.0049977065064013, + 0.006254979409277439, + 0.00417847977951169, + 0.006119942758232355, + 0.007818495854735374, + 0.004792553838342428, + 0.004899083636701107, + 0.007357842288911343, + 0.0055405027233064175, + 0.005990109872072935, + 0.00503421388566494, + 0.008205365389585495, + 0.006405601277947426, + 0.014508724212646484, + 0.027120495215058327, + 0.005984577815979719, + 0.007103696931153536, + 0.009450131095945835, + 0.005024378187954426, + 0.005563556216657162, + 0.010606701485812664, + 0.024645604193210602, + 0.03044109418988228, + 0.0053670755587518215, + 0.006548793986439705, + 0.009043306112289429, + 0.003193626645952463, + 0.0023747438099235296, + 0.005421798676252365, + 0.004249845631420612, + 0.004276329651474953, + 0.0024653240106999874, + 0.002215730492025614, + 0.0033697292674332857, + 0.0022764781024307013, + 0.00321266520768404, + 0.004593420308083296, + 0.003694049082696438, + 0.0015324025880545378, + 0.0017305553192272782, + 0.002305600792169571, + 0.0019928719848394394, + 0.0014248103834688663, + 0.0017657404532656074, + 0.0036170962266623974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -0.03303352743387222 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004538178443909 + ], + "opposite_logprob_tokens": [ + -2.6950454711914062 + ], + "logprob_reward": -0.07004538178443909, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004538178443909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02968248352408409, + 0.02626945450901985, + 0.07179313898086548, + 0.7521035671234131, + 0.027365950867533684, + 0.036496538668870926, + 0.00894975010305643, + 0.01750757545232773, + 0.017165668308734894, + 0.012665879912674427 + ], + "credit_map": [ + 0.004717682953923941, + 0.00319920526817441, + 0.003095019608736038, + 0.003026491031050682, + 0.004171245731413364, + 0.0026378522161394358, + 0.006338426843285561, + 0.0024965601041913033, + 0.00316642876714468, + 0.005997868720442057, + 0.0019142056116834283, + 0.0020812356378883123, + 0.003916294313967228, + 0.0022151086013764143, + 0.0036181723698973656, + 0.0033601417671889067, + 0.00562310079112649, + 0.021252097561955452, + 0.003449763637036085, + 0.0051598637364804745, + 0.007086074445396662, + 0.005672716069966555, + 0.006832258775830269, + 0.01671726442873478, + 0.07050687819719315, + 0.1609235405921936, + 0.07321412116289139, + 0.13831080496311188, + 0.05715030059218407, + 0.03628800809383392, + 0.05602853745222092, + 0.15968139469623566, + 0.005081438925117254, + 0.006031323224306107, + 0.0022087479010224342, + 0.002558185253292322, + 0.003409507917240262, + 0.0012975599383935332, + 0.003718278370797634, + 0.0030609085224568844, + 0.002760759787634015, + 0.004232034552842379, + 0.004737303126603365, + 0.0041284156031906605, + 0.006020189728587866, + 0.0024520899169147015, + 0.008048661053180695, + 0.004117084201425314, + 0.0011190373916178942, + 0.002381341066211462, + 0.0007217723759822547, + 0.0007860838668420911, + 0.0009090187377296388, + 0.0005197185673750937, + 0.0010338269639760256, + 0.0014789504930377007, + 0.00275926454924047, + 0.0021553689148277044, + 0.002187711652368307, + 0.00273371790535748, + 0.0016881111077964306, + 0.0013469094410538673, + 0.0020486083813011646, + 0.002587883733212948, + 0.0035249858628958464, + 0.0036363652907311916, + 0.0013947736006230116, + 0.0018449757480993867, + 0.001595612382516265, + 0.0013066091341897845, + 0.0021899419371038675, + 0.0016724031884223223, + 0.0013799451990053058, + 0.001495130592957139, + 0.0013578863581642509, + 0.0015404749428853393, + 0.001781250350177288, + 0.0014729746617376804, + 0.0018299364019185305, + 0.0018082821043208241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.07004538178443909 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813195586204529 + ], + "opposite_logprob_tokens": [ + -1.4063196182250977 + ], + "logprob_reward": -0.2813195586204529, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813195586204529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04423179104924202, + 0.4478347897529602, + 0.03565626218914986, + 0.12175905704498291, + 0.0841626301407814, + 0.05456465110182762, + 0.03424156457185745, + 0.06196802482008934, + 0.033807091414928436, + 0.08177416771650314 + ], + "credit_map": [ + 0.009602162055671215, + 0.003230629488825798, + 0.00413144426420331, + 0.004452304914593697, + 0.0038991947658360004, + 0.0034918778110295534, + 0.010238165967166424, + 0.005186009686440229, + 0.03504686802625656, + 0.17063894867897034, + 0.025737199932336807, + 0.02093767188489437, + 0.042496610432863235, + 0.022417647764086723, + 0.021030927076935768, + 0.10952890664339066, + 0.0031018343288451433, + 0.007260697428137064, + 0.0026737709995359182, + 0.00458343792706728, + 0.005901383236050606, + 0.00225375615991652, + 0.0053636557422578335, + 0.004517724737524986, + 0.00882228184491396, + 0.03072473406791687, + 0.010484583675861359, + 0.019172009080648422, + 0.013039487414062023, + 0.006278829649090767, + 0.009632756002247334, + 0.023604368790984154, + 0.008048387244343758, + 0.017258403822779655, + 0.00685535091906786, + 0.01026945747435093, + 0.010694117285311222, + 0.004477355629205704, + 0.010265219025313854, + 0.01629432663321495, + 0.0052787140011787415, + 0.0055146184749901295, + 0.007344549987465143, + 0.009052326902747154, + 0.008639661595225334, + 0.004262568429112434, + 0.00809619389474392, + 0.006376014556735754, + 0.008402392268180847, + 0.00678620021790266, + 0.0028183741960674524, + 0.0032655096147209406, + 0.0027544100303202868, + 0.002159146359190345, + 0.004701361060142517, + 0.0033541684970259666, + 0.01794963888823986, + 0.019745316356420517, + 0.004254122730344534, + 0.004421247635036707, + 0.0059960950165987015, + 0.002800218528136611, + 0.002347299363464117, + 0.004454086069017649, + 0.007639585994184017, + 0.004257530905306339, + 0.0035032855812460184, + 0.0046237679198384285, + 0.0038497704081237316, + 0.002316206693649292, + 0.003140972461551428, + 0.004475967027246952, + 0.007803625427186489, + 0.010462147183716297, + 0.007515962701290846, + 0.013053912669420242, + 0.013315172865986824, + 0.009631984867155552, + 0.010289541445672512, + 0.009701811708509922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.2813195586204529 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033713199198246 + ], + "opposite_logprob_tokens": [ + -3.4087131023406982 + ], + "logprob_reward": -0.033713199198246, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033713199198246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09427197277545929, + 0.32971468567848206, + 0.022427689284086227, + 0.102633997797966, + 0.048721667379140854, + 0.04192474111914635, + 0.013642439618706703, + 0.299961656332016, + 0.018111389130353928, + 0.02858981490135193 + ], + "credit_map": [ + 0.008218597620725632, + 0.008422113955020905, + 0.00619890121743083, + 0.01122952252626419, + 0.012257641181349754, + 0.006944621913135052, + 0.029447566717863083, + 0.011553005315363407, + 0.02537423186004162, + 0.1265730857849121, + 0.018898814916610718, + 0.016242900863289833, + 0.032422468066215515, + 0.0168367438018322, + 0.015527220442891121, + 0.07783923298120499, + 0.0019440373871475458, + 0.004306018818169832, + 0.0018217575270682573, + 0.002518606372177601, + 0.003304507117718458, + 0.001679019071161747, + 0.003507467219606042, + 0.0033462774008512497, + 0.008463541977107525, + 0.0316450260579586, + 0.006204104050993919, + 0.01124543882906437, + 0.009872360154986382, + 0.005452582612633705, + 0.006663469597697258, + 0.02308746613562107, + 0.004880357068032026, + 0.006974734365940094, + 0.00529743917286396, + 0.0053768400102853775, + 0.00766028044745326, + 0.004574601538479328, + 0.009217999875545502, + 0.004739414434880018, + 0.004984312225133181, + 0.009843964129686356, + 0.005024448502808809, + 0.003953239414840937, + 0.0054197064600884914, + 0.0027216002345085144, + 0.00610774289816618, + 0.003869726788252592, + 0.001884149620309472, + 0.00236299866810441, + 0.0014159384882077575, + 0.0023567622993141413, + 0.0013718397822231054, + 0.0010355229023844004, + 0.0013558343052864075, + 0.0018593932036310434, + 0.07851188629865646, + 0.10351361334323883, + 0.019181640818715096, + 0.02286355569958687, + 0.031452812254428864, + 0.015808282420039177, + 0.010725005529820919, + 0.017904823645949364, + 0.002193195978179574, + 0.0029715329874306917, + 0.001415229868143797, + 0.001891998341307044, + 0.0028891696128994226, + 0.0014671259559690952, + 0.0025286339223384857, + 0.0027545022312551737, + 0.005296219605952501, + 0.003203194821253419, + 0.0019751025829464197, + 0.0018397449748590589, + 0.003221971681341529, + 0.0028895887080579996, + 0.003548250999301672, + 0.0066157435066998005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.033713199198246 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073091067373753 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.011073091067373753, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073091067373753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03262952342629433, + 0.48815542459487915, + 0.05461803451180458, + 0.15456870198249817, + 0.0988042950630188, + 0.08979705721139908, + 0.020884348079562187, + 0.02976106107234955, + 0.008303673006594181, + 0.022477874532341957 + ], + "credit_map": [ + 0.005073011852800846, + 0.0037372231017798185, + 0.0029810687992721796, + 0.004155783914029598, + 0.004085935652256012, + 0.0031870740931481123, + 0.005737893283367157, + 0.0036715338937938213, + 0.037821345031261444, + 0.18815237283706665, + 0.02792808599770069, + 0.023690572008490562, + 0.047691598534584045, + 0.025050640106201172, + 0.022196201607584953, + 0.11562459915876389, + 0.005157667212188244, + 0.0131530174985528, + 0.003929422236979008, + 0.007107600104063749, + 0.006494936067610979, + 0.0032040998339653015, + 0.0084968451410532, + 0.007074447814375162, + 0.012604129500687122, + 0.05203717574477196, + 0.008052953518927097, + 0.0142621248960495, + 0.015989085659384727, + 0.00848312582820654, + 0.009690295904874802, + 0.03344980627298355, + 0.01703925058245659, + 0.008400747552514076, + 0.00899574812501669, + 0.010545522905886173, + 0.012366525828838348, + 0.005751962307840586, + 0.015224683098495007, + 0.020479854196310043, + 0.00861777551472187, + 0.008562237955629826, + 0.012825391255319118, + 0.010265063494443893, + 0.015233534388244152, + 0.006746698170900345, + 0.017484499141573906, + 0.010061860084533691, + 0.00351162813603878, + 0.002437798073515296, + 0.001978101907297969, + 0.002822631038725376, + 0.0019736941903829575, + 0.0014631628291681409, + 0.0021010374184697866, + 0.004596292041242123, + 0.004211785271763802, + 0.004502970725297928, + 0.002353227697312832, + 0.005973947700113058, + 0.002553050871938467, + 0.002956798765808344, + 0.00334001611918211, + 0.0038692625239491463, + 0.0009502175962552428, + 0.0008771372959017754, + 0.0009183413931168616, + 0.001343780430033803, + 0.0013044195948168635, + 0.0007262344006448984, + 0.0009838184341788292, + 0.001199723337776959, + 0.004473202861845493, + 0.0023754206486046314, + 0.002140875905752182, + 0.002702258061617613, + 0.0027714392635971308, + 0.0018627847312018275, + 0.002181005198508501, + 0.003970888908952475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.011073091067373753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a close-up shot framed in vertical format, the camera remains stationary, focused on an exquisite folding fan, initially folded. Set against a pure, dark background, this scene creates a minimalist, stage-like feel. With a swift, decisive movement, the fan, crafted from polished, dark bamboo ribs, unfolds completely. Its fluid movement instantly fills the frame, ultimately forming a perfect semicircle. A soft, directional light, directed from above and sideways, subtly illuminates the off-white rice paper, bringing to life every detail of a profoundly poetic Chinese ink landscape painting: the outlines of distant mountains looming through the mist, the delicate gradations of ink smudged across the paper, and the subtle shifts in wet and dry brushstrokes. The light also casts subtle shadows within the folds formed by the fan's unfolding ribs, enhancing the image's sense of three-dimensionality and texture. The entire scene is imbued with a tranquil and elegant atmosphere of classical Eastern aesthetics, conveying a dedication to and appreciation of traditional craftsmanship and artistry.", + "seed": 18063, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_10/index_3032_seed_18063_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03201385848224163, + "vlm_logprob": -0.03201385848224163, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.03201385848224163, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02357814647257328 + ], + "opposite_logprob_tokens": [ + -4.52357816696167 + ], + "logprob_reward": -0.02357814647257328, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.02357814647257328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3097401559352875, + 0.0691775530576706, + 0.06696360558271408, + 0.025817716494202614, + 0.05004459246993065, + 0.057279277592897415, + 0.3374539315700531, + 0.04583539813756943, + 0.02155187912285328, + 0.016135862097144127 + ], + "credit_map": [ + 0.03488258272409439, + 0.01869494654238224, + 0.028838714584708214, + 0.018158558756113052, + 0.02363630384206772, + 0.07631865888834, + 0.02728232555091381, + 0.0819280818104744, + 0.005479244515299797, + 0.005643397569656372, + 0.0060615395195782185, + 0.006394064985215664, + 0.004906905815005302, + 0.008894459344446659, + 0.020291952416300774, + 0.011505994014441967, + 0.006260486785322428, + 0.011670364066958427, + 0.013364831916987896, + 0.007330963388085365, + 0.005834283772855997, + 0.007308239117264748, + 0.007823809050023556, + 0.0073706284165382385, + 0.002489040605723858, + 0.002987116342410445, + 0.0029512743931263685, + 0.002772985491901636, + 0.004159900825470686, + 0.002019294537603855, + 0.0053854589350521564, + 0.003052647691220045, + 0.007240561302751303, + 0.003177296370267868, + 0.003061464289203286, + 0.003922174219042063, + 0.010907689109444618, + 0.0044631352648139, + 0.0069435047917068005, + 0.01032876968383789, + 0.0038954734336584806, + 0.007265989203006029, + 0.00812768004834652, + 0.004799928516149521, + 0.012391815893352032, + 0.008357863873243332, + 0.006578695960342884, + 0.0058618346229195595, + 0.03414810076355934, + 0.01399915385991335, + 0.008544404059648514, + 0.011347627267241478, + 0.11129031330347061, + 0.13495653867721558, + 0.010137539356946945, + 0.01303024310618639, + 0.008352682925760746, + 0.004320090636610985, + 0.0035109068267047405, + 0.004085168708115816, + 0.005053783301264048, + 0.013097915798425674, + 0.004059548489749432, + 0.0033552967943251133, + 0.0018988453084602952, + 0.0020004676189273596, + 0.002804345218464732, + 0.0022343641612678766, + 0.0037778387777507305, + 0.002269790042191744, + 0.004051592666655779, + 0.0025146352127194405, + 0.001913985819555819, + 0.0011749417753890157, + 0.0013097655028104782, + 0.001442577806301415, + 0.002893785946071148, + 0.004041688982397318, + 0.0013814326375722885, + 0.0019776837434619665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.02357814647257328 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054993901401758194 + ], + "opposite_logprob_tokens": [ + -2.9299938678741455 + ], + "logprob_reward": -0.054993901401758194, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054993901401758194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.774930477142334, + 0.03439817577600479, + 0.047539468854665756, + 0.021127980202436447, + 0.017595944926142693, + 0.019116414710879326, + 0.044767268002033234, + 0.015579345636069775, + 0.006685328669846058, + 0.018259571865200996 + ], + "credit_map": [ + 0.03647495061159134, + 0.027220023795962334, + 0.03974784538149834, + 0.023989612236618996, + 0.04744940623641014, + 0.25066933035850525, + 0.08596502989530563, + 0.26341426372528076, + 0.002857615938410163, + 0.0033471782226115465, + 0.005405977834016085, + 0.0036247004754841328, + 0.0027384436689317226, + 0.004236947279423475, + 0.0070106154307723045, + 0.005176698323339224, + 0.004112978465855122, + 0.005354335997253656, + 0.008871483616530895, + 0.006647082045674324, + 0.007039143703877926, + 0.0039567044004797935, + 0.006691399496048689, + 0.004866340663284063, + 0.001810205401852727, + 0.0026452746242284775, + 0.003528148401528597, + 0.003083051647990942, + 0.003120943671092391, + 0.0013674224028363824, + 0.003431615186855197, + 0.002141321310773492, + 0.0015625287778675556, + 0.0016732257790863514, + 0.0015981680480763316, + 0.0013307720655575395, + 0.005514773540198803, + 0.0013590046437457204, + 0.002078751567751169, + 0.0024787220172584057, + 0.0024744058027863503, + 0.002231850754469633, + 0.0018533708062022924, + 0.0013583697145804763, + 0.002539360197260976, + 0.004798321984708309, + 0.001944518182426691, + 0.0019162181997671723, + 0.003517331788316369, + 0.0020246063359081745, + 0.0019496879540383816, + 0.00186142697930336, + 0.01531548798084259, + 0.015953050926327705, + 0.0015615604352205992, + 0.002584118628874421, + 0.0011963789584115148, + 0.0014345046365633607, + 0.001935014151968062, + 0.0016424317145720124, + 0.0019220797112211585, + 0.0015042705927044153, + 0.003752737771719694, + 0.002191928680986166, + 0.0006053076358512044, + 0.0005619014264084399, + 0.0006044008187018335, + 0.000675141578540206, + 0.0013577048666775227, + 0.0014857720816507936, + 0.0007084819371812046, + 0.0006866190233267844, + 0.0036214084830135107, + 0.0008479977259412408, + 0.0009143282077275217, + 0.000867883674800396, + 0.003289592918008566, + 0.00680427486076951, + 0.0009185270173475146, + 0.0009955584537237883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.054993901401758194 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204335391521454 + ], + "opposite_logprob_tokens": [ + -2.8120434284210205 + ], + "logprob_reward": -0.06204335391521454, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204335391521454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27295249700546265, + 0.12277600914239883, + 0.052251629531383514, + 0.04181266576051712, + 0.07262488454580307, + 0.07420629262924194, + 0.23012730479240417, + 0.03557910397648811, + 0.017632732167840004, + 0.08003684878349304 + ], + "credit_map": [ + 0.03651143237948418, + 0.021003907546401024, + 0.03277503699064255, + 0.01995883136987686, + 0.018784239888191223, + 0.056633055210113525, + 0.026664525270462036, + 0.0606214813888073, + 0.013372577726840973, + 0.008775264956057072, + 0.011910714209079742, + 0.009224067442119122, + 0.010533466935157776, + 0.024245452135801315, + 0.02599700540304184, + 0.018717456609010696, + 0.0036102866288274527, + 0.008597823791205883, + 0.008515015244483948, + 0.008763599209487438, + 0.003951709717512131, + 0.003608337137848139, + 0.008919141255319118, + 0.006285719107836485, + 0.004057074896991253, + 0.005598155781626701, + 0.006019112188369036, + 0.004681793041527271, + 0.00661052530631423, + 0.0030116678681224585, + 0.007456324063241482, + 0.00437801331281662, + 0.008503072895109653, + 0.006937988568097353, + 0.0058010034263134, + 0.006149465683847666, + 0.018415121361613274, + 0.0058054449036717415, + 0.007893027737736702, + 0.01311976183205843, + 0.0054742153733968735, + 0.008614362217485905, + 0.011688483878970146, + 0.006672384683042765, + 0.014652031473815441, + 0.007817594334483147, + 0.011365908198058605, + 0.007921318523585796, + 0.046513598412275314, + 0.02008724957704544, + 0.01803954876959324, + 0.0162394717335701, + 0.030358215793967247, + 0.05355171114206314, + 0.026333842426538467, + 0.01900365762412548, + 0.004845264833420515, + 0.0025000583846122026, + 0.0023432604502886534, + 0.0025674302596598864, + 0.006546672899276018, + 0.011128301732242107, + 0.0030751812737435102, + 0.002572936238721013, + 0.0014878090005367994, + 0.0012243470409885049, + 0.0018176010344177485, + 0.001702967332676053, + 0.0026985243894159794, + 0.002131063723936677, + 0.0043371813371777534, + 0.0022332374937832355, + 0.013646998442709446, + 0.00356982066296041, + 0.004201845731586218, + 0.004004643298685551, + 0.014483790844678879, + 0.026835739612579346, + 0.00722566107288003, + 0.006068351678550243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.06204335391521454 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202586375176907 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202586375176907, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202586375176907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18965180218219757, + 0.1313529908657074, + 0.14349059760570526, + 0.0726863294839859, + 0.07073318213224411, + 0.055848464369773865, + 0.16805125772953033, + 0.09931766241788864, + 0.04077790677547455, + 0.028089825063943863 + ], + "credit_map": [ + 0.020196011289954185, + 0.009975027292966843, + 0.01657426357269287, + 0.010647134855389595, + 0.012755247764289379, + 0.04942276328802109, + 0.022959165275096893, + 0.04712219163775444, + 0.010751939378678799, + 0.013980397023260593, + 0.013735863380134106, + 0.012636998668313026, + 0.01326721254736185, + 0.019801365211606026, + 0.026025593280792236, + 0.02115361951291561, + 0.010977732948958874, + 0.018570270389318466, + 0.02312828041613102, + 0.02549380622804165, + 0.015274683944880962, + 0.010720227845013142, + 0.023066287860274315, + 0.01625930331647396, + 0.007191246375441551, + 0.00887860357761383, + 0.011489920318126678, + 0.010663101449608803, + 0.008331907913088799, + 0.00440980214625597, + 0.013718103058636189, + 0.008003639988601208, + 0.009592169895768166, + 0.004741290118545294, + 0.004181492142379284, + 0.006722043268382549, + 0.008554869331419468, + 0.0037892123218625784, + 0.015539857558906078, + 0.017612243071198463, + 0.004798165522515774, + 0.006412484683096409, + 0.007712218910455704, + 0.005085133947432041, + 0.010641616769134998, + 0.008386712521314621, + 0.0077581400983035564, + 0.005053991451859474, + 0.0206228569149971, + 0.007747169584035873, + 0.005552961025387049, + 0.007344605401158333, + 0.04324468970298767, + 0.06452963501214981, + 0.010488313622772694, + 0.008520998992025852, + 0.016377180814743042, + 0.006627343595027924, + 0.010061226785182953, + 0.007838618010282516, + 0.013134192675352097, + 0.028461337089538574, + 0.010349607095122337, + 0.006468151230365038, + 0.007538184989243746, + 0.0019508410478010774, + 0.0015383627032861114, + 0.0016475175507366657, + 0.007895611226558685, + 0.015240798704326153, + 0.0026913315523415804, + 0.002275255275890231, + 0.0026820688508450985, + 0.0022015313152223825, + 0.0018890523351728916, + 0.0022591620218008757, + 0.00646826159209013, + 0.00676162401214242, + 0.0032175397500395775, + 0.002610582858324051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.014202586375176907 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251304246485233 + ], + "opposite_logprob_tokens": [ + -5.255251407623291 + ], + "logprob_reward": -0.005251304246485233, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251304246485233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2669651210308075, + 0.15088175237178802, + 0.17819452285766602, + 0.05535328388214111, + 0.08062109351158142, + 0.062089260667562485, + 0.04776204004883766, + 0.06676438450813293, + 0.03981083258986473, + 0.05155768617987633 + ], + "credit_map": [ + 0.026945123448967934, + 0.013971762731671333, + 0.019920574501156807, + 0.021289564669132233, + 0.02277386747300625, + 0.050046298652887344, + 0.05905117839574814, + 0.05296674743294716, + 0.013714075088500977, + 0.01204385980963707, + 0.022491810843348503, + 0.018551792949438095, + 0.011192070320248604, + 0.019531546160578728, + 0.03050757758319378, + 0.022849028930068016, + 0.012877892702817917, + 0.03110545687377453, + 0.028344031423330307, + 0.029727283865213394, + 0.01946350745856762, + 0.01407182589173317, + 0.022032814100384712, + 0.02057170309126377, + 0.007325842976570129, + 0.0046616652980446815, + 0.007301259320229292, + 0.006621784996241331, + 0.0065842727199196815, + 0.005055808462202549, + 0.009258597157895565, + 0.008544049225747585, + 0.00962388701736927, + 0.006947527173906565, + 0.006126497872173786, + 0.005480693653225899, + 0.020067021250724792, + 0.005493731703609228, + 0.006586870644241571, + 0.02029486745595932, + 0.0070670354180037975, + 0.007339015137404203, + 0.004831406753510237, + 0.003847988322377205, + 0.012677568010985851, + 0.01537984050810337, + 0.005289859604090452, + 0.005656551104038954, + 0.0075103649869561195, + 0.004355349112302065, + 0.0036588353104889393, + 0.0035260997246950865, + 0.005770966876298189, + 0.013972478918731213, + 0.003748321672901511, + 0.005219622515141964, + 0.010442101396620274, + 0.0052606817334890366, + 0.0036697036121040583, + 0.004114331677556038, + 0.011030323803424835, + 0.021830709651112556, + 0.005875241942703724, + 0.00454128859564662, + 0.004117647651582956, + 0.0029862781520932913, + 0.004041946493089199, + 0.003979077097028494, + 0.006730156019330025, + 0.007714849431067705, + 0.005635786801576614, + 0.004605090245604515, + 0.009257001802325249, + 0.0039036970119923353, + 0.0037261368706822395, + 0.004175342153757811, + 0.00609064195305109, + 0.013300742022693157, + 0.005332425702363253, + 0.005771700758486986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.005251304246485233 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a close-up shot framed in vertical format, the camera remains stationary, focused on an exquisite folding fan, initially folded. Set against a pure, dark background, this scene creates a minimalist, stage-like feel. With a swift, decisive movement, the fan, crafted from polished, dark bamboo ribs, unfolds completely. Its fluid movement instantly fills the frame, ultimately forming a perfect semicircle. A soft, directional light, directed from above and sideways, subtly illuminates the off-white rice paper, bringing to life every detail of a profoundly poetic Chinese ink landscape painting: the outlines of distant mountains looming through the mist, the delicate gradations of ink smudged across the paper, and the subtle shifts in wet and dry brushstrokes. The light also casts subtle shadows within the folds formed by the fan's unfolding ribs, enhancing the image's sense of three-dimensionality and texture. The entire scene is imbued with a tranquil and elegant atmosphere of classical Eastern aesthetics, conveying a dedication to and appreciation of traditional craftsmanship and artistry.", + "seed": 18064, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_10/index_3032_seed_18064_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03226592242717743, + "vlm_logprob": -0.03226592242717743, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.03226592242717743, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023910271003842354 + ], + "opposite_logprob_tokens": [ + -4.273910045623779 + ], + "logprob_reward": -0.023910271003842354, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.023910271003842354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11566993594169617, + 0.07371712476015091, + 0.09493584930896759, + 0.06404414027929306, + 0.33645620942115784, + 0.0438976064324379, + 0.08919866383075714, + 0.06179008260369301, + 0.10093145817518234, + 0.019358986988663673 + ], + "credit_map": [ + 0.013724947348237038, + 0.012216970324516296, + 0.00937054492533207, + 0.014762074686586857, + 0.021422745659947395, + 0.012203901074826717, + 0.008825440891087055, + 0.023143308237195015, + 0.006213207263499498, + 0.008396509103477001, + 0.006410053465515375, + 0.013116605579853058, + 0.013899539597332478, + 0.006793434731662273, + 0.009964076802134514, + 0.008923699147999287, + 0.01356916781514883, + 0.02186877280473709, + 0.007920192554593086, + 0.010885839350521564, + 0.014725812710821629, + 0.008937804959714413, + 0.007007952313870192, + 0.010020306333899498, + 0.007166149094700813, + 0.008109702728688717, + 0.005619472824037075, + 0.014160620048642159, + 0.008104776963591576, + 0.006513212341815233, + 0.009128207340836525, + 0.005241996143013239, + 0.020526088774204254, + 0.03744955733418465, + 0.014494802802801132, + 0.024056723341345787, + 0.05820874497294426, + 0.03132838010787964, + 0.015387971885502338, + 0.1350039541721344, + 0.0044939243234694, + 0.005401786416769028, + 0.0033817747607827187, + 0.006738321855664253, + 0.004381525795906782, + 0.00396795105189085, + 0.005698031280189753, + 0.0098342914134264, + 0.007939324714243412, + 0.009999789297580719, + 0.009511335752904415, + 0.012952274642884731, + 0.010907378979027271, + 0.006498843431472778, + 0.007895435206592083, + 0.023494286462664604, + 0.007110447157174349, + 0.006895774509757757, + 0.006469059735536575, + 0.004825903102755547, + 0.01369711384177208, + 0.004013972356915474, + 0.005025412421673536, + 0.013752399012446404, + 0.012732270173728466, + 0.00947830080986023, + 0.012796062044799328, + 0.010159171186387539, + 0.016147317364811897, + 0.006975158583372831, + 0.0075664459727704525, + 0.025076737627387047, + 0.002733405213803053, + 0.002996718045324087, + 0.0029728692024946213, + 0.0020162793807685375, + 0.0036506066098809242, + 0.0013722332660108805, + 0.0019862158223986626, + 0.0016306592151522636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.023910271003842354 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899615913629532 + ], + "opposite_logprob_tokens": [ + -2.578996181488037 + ], + "logprob_reward": -0.07899615913629532, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899615913629532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1039404571056366, + 0.10111068934202194, + 0.05814671888947487, + 0.059231679886579514, + 0.25534141063690186, + 0.035095538944005966, + 0.10856876522302628, + 0.12390675395727158, + 0.09270697087049484, + 0.061951059848070145 + ], + "credit_map": [ + 0.014537202194333076, + 0.010158101096749306, + 0.011214417405426502, + 0.02623060531914234, + 0.011284860782325268, + 0.008470228873193264, + 0.008452802896499634, + 0.013592233881354332, + 0.008415520191192627, + 0.012022281996905804, + 0.008655823767185211, + 0.018520062789320946, + 0.02027190290391445, + 0.008822903037071228, + 0.01241130568087101, + 0.011990886181592941, + 0.006973170209676027, + 0.014561139047145844, + 0.003740383079275489, + 0.007541308645159006, + 0.008839092217385769, + 0.005769194103777409, + 0.005402266979217529, + 0.005320166703313589, + 0.009733637794852257, + 0.009610339067876339, + 0.004987232852727175, + 0.00919209886342287, + 0.007769960444420576, + 0.006494000554084778, + 0.006554022431373596, + 0.004890392068773508, + 0.017395837232470512, + 0.05407468602061272, + 0.012300127185881138, + 0.02155211940407753, + 0.026983408257365227, + 0.0154948765411973, + 0.013151244260370731, + 0.09438911825418472, + 0.003967054653912783, + 0.005032174289226532, + 0.0031457385048270226, + 0.005121399182826281, + 0.004800369031727314, + 0.0032500133384019136, + 0.005770360119640827, + 0.004008428193628788, + 0.01017257384955883, + 0.013549610041081905, + 0.008733907714486122, + 0.01891280710697174, + 0.014314989559352398, + 0.009528551250696182, + 0.0124086644500494, + 0.020947661250829697, + 0.0070467544719576836, + 0.018903084099292755, + 0.005640967283397913, + 0.012441795319318771, + 0.036142438650131226, + 0.006849598605185747, + 0.010215564630925655, + 0.0266665481030941, + 0.006262040697038174, + 0.008603258989751339, + 0.005521606653928757, + 0.0072131529450416565, + 0.024387802928686142, + 0.005324399098753929, + 0.005283690989017487, + 0.030111022293567657, + 0.002487621735781431, + 0.00796564482152462, + 0.0025275605730712414, + 0.0049886517226696014, + 0.019735438749194145, + 0.002886364469304681, + 0.00329862954095006, + 0.018061146140098572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.07899615913629532 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370939567685127 + ], + "opposite_logprob_tokens": [ + -3.4087092876434326 + ], + "logprob_reward": -0.03370939567685127, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370939567685127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07552513480186462, + 0.21166396141052246, + 0.03036077320575714, + 0.04364475980401039, + 0.35298725962638855, + 0.07150135934352875, + 0.030057014897465706, + 0.03850439935922623, + 0.11161345988512039, + 0.03414185345172882 + ], + "credit_map": [ + 0.01105545461177826, + 0.007221068721264601, + 0.007274182979017496, + 0.016100039705634117, + 0.007962486706674099, + 0.00826807226985693, + 0.007595260161906481, + 0.01004856824874878, + 0.01866675540804863, + 0.024690594524145126, + 0.0180352870374918, + 0.03392387926578522, + 0.04300914332270622, + 0.020326735451817513, + 0.02656433917582035, + 0.02644720859825611, + 0.004414700902998447, + 0.007508387789130211, + 0.0026550646871328354, + 0.0030239387415349483, + 0.005367807578295469, + 0.002619297942146659, + 0.002644101856276393, + 0.0021274727769196033, + 0.0047443644143640995, + 0.012412999756634235, + 0.002758296672254801, + 0.007300180848687887, + 0.005950826220214367, + 0.0027565867640078068, + 0.003175662364810705, + 0.004545843228697777, + 0.0221156794577837, + 0.12603387236595154, + 0.008975004777312279, + 0.01510746031999588, + 0.13554084300994873, + 0.015612688846886158, + 0.011253141798079014, + 0.018348589539527893, + 0.006480586715042591, + 0.007392393425107002, + 0.006783095188438892, + 0.007757596671581268, + 0.008994975127279758, + 0.005525217391550541, + 0.005282633472234011, + 0.023284858092665672, + 0.0028851411771029234, + 0.0030751447193324566, + 0.002567704999819398, + 0.004375061020255089, + 0.004418626427650452, + 0.002531882142648101, + 0.003242549253627658, + 0.006960904225707054, + 0.005238983314484358, + 0.006422172300517559, + 0.00437315646559, + 0.005250053014606237, + 0.00756700336933136, + 0.003437414998188615, + 0.003829281311482191, + 0.0023863373789936304, + 0.008163606747984886, + 0.01003296673297882, + 0.007661812007427216, + 0.007844209671020508, + 0.029301976785063744, + 0.006183591205626726, + 0.005363172385841608, + 0.03706211969256401, + 0.0028580534271895885, + 0.0031197434291243553, + 0.002639436861500144, + 0.003957460168749094, + 0.007209317293018103, + 0.002357351826503873, + 0.002949246671050787, + 0.009051241911947727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 8 + ], + "selected_reward": -0.03370939567685127 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016077518463134766 + ], + "opposite_logprob_tokens": [ + -4.141077518463135 + ], + "logprob_reward": -0.016077518463134766, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016077518463134766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17280732095241547, + 0.04720935970544815, + 0.058444976806640625, + 0.10690410435199738, + 0.17166505753993988, + 0.053174082189798355, + 0.17676325142383575, + 0.16466423869132996, + 0.025202719494700432, + 0.023164868354797363 + ], + "credit_map": [ + 0.02925952896475792, + 0.024409553036093712, + 0.018842710182070732, + 0.027868598699569702, + 0.019298654049634933, + 0.02101367525756359, + 0.015232373960316181, + 0.01688222773373127, + 0.003796500386670232, + 0.004914393182843924, + 0.004388793371617794, + 0.008933313190937042, + 0.009149645455181599, + 0.004474701825529337, + 0.006329215131700039, + 0.005222797393798828, + 0.008148320950567722, + 0.006948759779334068, + 0.005644332151859999, + 0.004990182816982269, + 0.010940727777779102, + 0.005043533630669117, + 0.006700905505567789, + 0.010028211399912834, + 0.02491050586104393, + 0.03686858341097832, + 0.004428896587342024, + 0.010091320611536503, + 0.013663943856954575, + 0.006128308828920126, + 0.006395080126821995, + 0.004417465999722481, + 0.007193618454039097, + 0.05458897352218628, + 0.004634791985154152, + 0.007804292254149914, + 0.0526706837117672, + 0.017554452642798424, + 0.007091854233294725, + 0.02012639306485653, + 0.004237780347466469, + 0.009070261381566525, + 0.0042737144976854324, + 0.010301553644239902, + 0.008380603976547718, + 0.004571905825287104, + 0.007401036564260721, + 0.004937224555760622, + 0.010942456312477589, + 0.018336419016122818, + 0.00937058124691248, + 0.01631205342710018, + 0.044851239770650864, + 0.01158983539789915, + 0.010628953576087952, + 0.05473172292113304, + 0.009457290172576904, + 0.01602284237742424, + 0.008134324103593826, + 0.013419822789728642, + 0.04309960454702377, + 0.010904575698077679, + 0.010670267976820469, + 0.0529555082321167, + 0.0027295658364892006, + 0.0030479307752102613, + 0.002715333364903927, + 0.003381433431059122, + 0.0019966328982263803, + 0.002063842723146081, + 0.0018081063171848655, + 0.007459872402250767, + 0.0014476043870672584, + 0.0030936235561966896, + 0.0012222346849739552, + 0.0025747206527739763, + 0.005535301752388477, + 0.0016884556971490383, + 0.002627733163535595, + 0.004975194111466408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.016077518463134766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up with a stationary camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fan start folded and then unfold completely to form a perfect semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fan crafted from polished dark bamboo ribs with off-white rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there soft directional light coming from above and sideways that illuminates the paper and casts shadows in the folds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636267855763435 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636267855763435, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636267855763435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20885923504829407, + 0.06681947410106659, + 0.01913919486105442, + 0.02913735993206501, + 0.242537721991539, + 0.03834196552634239, + 0.1362074911594391, + 0.2046702653169632, + 0.013366284780204296, + 0.040921054780483246 + ], + "credit_map": [ + 0.024506624788045883, + 0.02093435637652874, + 0.02252778224647045, + 0.03890273720026016, + 0.032830387353897095, + 0.01925683207809925, + 0.019153404980897903, + 0.030747096985578537, + 0.0062039257027208805, + 0.007322157267481089, + 0.005787264555692673, + 0.01024453341960907, + 0.011835053563117981, + 0.007252423092722893, + 0.009663809090852737, + 0.008510301820933819, + 0.0034733337815850973, + 0.0036986940540373325, + 0.001411563833244145, + 0.002263898029923439, + 0.0032211104407906532, + 0.001781569910235703, + 0.0015831253258511424, + 0.0017058980884030461, + 0.006222354713827372, + 0.0025794869288802147, + 0.0026953350752592087, + 0.005204926244914532, + 0.002655372256413102, + 0.0033937692642211914, + 0.003788255387917161, + 0.002597857965156436, + 0.018963491544127464, + 0.05898689478635788, + 0.009942345321178436, + 0.008701876737177372, + 0.08795201033353806, + 0.010856237262487411, + 0.00941516924649477, + 0.03771968185901642, + 0.0036340123042464256, + 0.0053000058978796005, + 0.004856067709624767, + 0.005312045104801655, + 0.0026059106457978487, + 0.0027264461386948824, + 0.0024540722370147705, + 0.011453401297330856, + 0.00938297901302576, + 0.016696181148290634, + 0.00840059109032154, + 0.01381049957126379, + 0.03419337794184685, + 0.008849425241351128, + 0.008880562148988247, + 0.035993874073028564, + 0.01038240734487772, + 0.02113187126815319, + 0.009091840125620365, + 0.01566905900835991, + 0.05919012054800987, + 0.011528494767844677, + 0.010979515500366688, + 0.06669694185256958, + 0.0013810726813971996, + 0.00171561015304178, + 0.000926150125451386, + 0.0018706368282437325, + 0.0017730888212099671, + 0.0012436440447345376, + 0.001577348681166768, + 0.002878732280805707, + 0.001557767391204834, + 0.005367395468056202, + 0.0014753942377865314, + 0.002778527559712529, + 0.013026922941207886, + 0.0023494281340390444, + 0.0019690680783241987, + 0.012396548874676228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.008636267855763435 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_11/index_tensor([7381, 7381])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_11/index_tensor([7381, 7381])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a7d1f37e60a8ce981918ff8023f6b5b7dcf30fa4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_11/index_tensor([7381, 7381])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide-angle, long shot captured from a level perspective, the camera remains still, the composition guided by the principle of extreme simplicity, focusing entirely on the endless golden rice fields before them. At the golden section of the frame, a lone figure, deeply bent, almost blends in with the throngs of stalks of rice. This is a woman wearing a wide-brimmed straw hat, its brim casting a shadow that obscures her face, revealing only her simple bun and the faded indigo cotton top that contrasts sharply with the golden color. The warm, soft afternoon sun paints the entire rice field a dazzling golden hue, each grain appearing heavy and gleaming. Her movements are slow and focused, her left hand carefully gathering a handful of plump rice stalks, while her right hand, using a small sickle, steadily and deliberately cuts them from the root. The entire process is fluid and rhythmic, filled with reverence for the bountiful harvest. Behind her, golden rice waves extend all the way to the horizon where they meet the pure blue sky. The picture is clean and pure, with saturated and warm colors, creating a tranquil, fertile and poetic pastoral atmosphere.", + "seed": 18870, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_11/index_7381_seed_18870_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14243646590039133, + "vlm_logprob": -0.14243646590039133, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.14243646590039133, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026730559766292572 + ], + "opposite_logprob_tokens": [ + -4.276730537414551 + ], + "logprob_reward": -0.026730559766292572, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.026730559766292572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08251740038394928, + 0.031179776415228844, + 0.08640379458665848, + 0.389923632144928, + 0.04430447891354561, + 0.07046178728342056, + 0.07002192735671997, + 0.12350193411111832, + 0.03503318503499031, + 0.06665217131376266 + ], + "credit_map": [ + 0.012455175630748272, + 0.020997611805796623, + 0.016008883714675903, + 0.006498417817056179, + 0.004758009221404791, + 0.006692616734653711, + 0.006547074299305677, + 0.008559603244066238, + 0.0040634083561599255, + 0.004416489973664284, + 0.00553505914285779, + 0.002641829429194331, + 0.0037789172492921352, + 0.003468979150056839, + 0.0025436661671847105, + 0.004731425084173679, + 0.012550843879580498, + 0.021666521206498146, + 0.02309241332113743, + 0.006405650172382593, + 0.004118854645639658, + 0.0038912310265004635, + 0.006402085535228252, + 0.008276182226836681, + 0.07172813266515732, + 0.12890776991844177, + 0.09052790701389313, + 0.02258972078561783, + 0.018600627779960632, + 0.013965661637485027, + 0.010630699805915356, + 0.03297308459877968, + 0.007477348670363426, + 0.010628607124090195, + 0.006028187461197376, + 0.0038886968977749348, + 0.003331936663016677, + 0.0036174689885228872, + 0.0026631178334355354, + 0.006669109687209129, + 0.011080642230808735, + 0.02094789780676365, + 0.017929570749402046, + 0.005243970546871424, + 0.0033739986829459667, + 0.0024147892836481333, + 0.0031101908534765244, + 0.006360714323818684, + 0.010465450584888458, + 0.015150501392781734, + 0.012084973976016045, + 0.005640156101435423, + 0.005901989061385393, + 0.0071815261617302895, + 0.005512543022632599, + 0.008084778673946857, + 0.01245125476270914, + 0.014697173610329628, + 0.01697637140750885, + 0.013602199032902718, + 0.013145062141120434, + 0.012384427711367607, + 0.014038534834980965, + 0.02620689757168293, + 0.004773441702127457, + 0.00468851625919342, + 0.006205947604030371, + 0.0038184686563909054, + 0.002595406724140048, + 0.0045390864834189415, + 0.0027882391586899757, + 0.0056240735575556755, + 0.011007037945091724, + 0.02065354213118553, + 0.01228246372193098, + 0.004540028050541878, + 0.0034504481591284275, + 0.003389238379895687, + 0.003376221749931574, + 0.007953185588121414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 2 + ], + "selected_reward": -0.026730559766292572 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001957050058990717 + ], + "opposite_logprob_tokens": [ + -6.251956939697266 + ], + "logprob_reward": -0.001957050058990717, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001957050058990717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13308751583099365, + 0.07600905001163483, + 0.16572178900241852, + 0.09765361249446869, + 0.06152942031621933, + 0.0652100220322609, + 0.04428911954164505, + 0.08780863136053085, + 0.17510193586349487, + 0.09358896315097809 + ], + "credit_map": [ + 0.019467702135443687, + 0.033143892884254456, + 0.022938204929232597, + 0.012158602476119995, + 0.012226617895066738, + 0.010610606521368027, + 0.011556229554116726, + 0.010985666885972023, + 0.00766372075304389, + 0.006537445355206728, + 0.011152993887662888, + 0.011839956976473331, + 0.00790729932487011, + 0.009841645136475563, + 0.01133388839662075, + 0.00973210297524929, + 0.0314563512802124, + 0.05714826658368111, + 0.03553441911935806, + 0.009027179330587387, + 0.008493688888847828, + 0.004449780099093914, + 0.0073286681436002254, + 0.012283426709473133, + 0.018445460125803947, + 0.03432030603289604, + 0.018099695444107056, + 0.005778666585683823, + 0.006157984025776386, + 0.004506229888647795, + 0.0040256306529045105, + 0.00631963973864913, + 0.009446944110095501, + 0.013395769521594048, + 0.007756441365927458, + 0.005877250339835882, + 0.0032229875214397907, + 0.006119527854025364, + 0.004760087467730045, + 0.010950417257845402, + 0.00968744046986103, + 0.017209511250257492, + 0.014199219644069672, + 0.004641926847398281, + 0.0036403732374310493, + 0.004068521782755852, + 0.0034137980546802282, + 0.008349229581654072, + 0.005239959340542555, + 0.006940108258277178, + 0.007531745359301567, + 0.0039788116700947285, + 0.003742088796570897, + 0.007406509015709162, + 0.0029932609759271145, + 0.006456633564084768, + 0.008726564235985279, + 0.010065379552543163, + 0.012896602042019367, + 0.010287891142070293, + 0.0083732008934021, + 0.00784736406058073, + 0.010028003714978695, + 0.019583633169531822, + 0.02342076599597931, + 0.04000483453273773, + 0.035356950014829636, + 0.006960920989513397, + 0.006336156744509935, + 0.007958750240504742, + 0.009884572587907314, + 0.045178987085819244, + 0.016680540516972542, + 0.0305678378790617, + 0.021650703623890877, + 0.004869294818490744, + 0.0037405788898468018, + 0.002837942447513342, + 0.004674484021961689, + 0.008567580953240395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.001957050058990717 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0125464191660285 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.0125464191660285, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0125464191660285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060284778475761414, + 0.13083304464817047, + 0.05449637770652771, + 0.11788458377122879, + 0.15313874185085297, + 0.07768753916025162, + 0.11170153319835663, + 0.10952363163232803, + 0.16476128995418549, + 0.019688449800014496 + ], + "credit_map": [ + 0.006573510821908712, + 0.0067281522788107395, + 0.010778365656733513, + 0.00624509435147047, + 0.006461313460022211, + 0.007174804341048002, + 0.008299089036881924, + 0.008024446666240692, + 0.01626548543572426, + 0.01694585382938385, + 0.015579712577164173, + 0.016093237325549126, + 0.01887340471148491, + 0.01779926009476185, + 0.016763227060437202, + 0.01251286268234253, + 0.007496946956962347, + 0.012962560169398785, + 0.006265265867114067, + 0.00795928668230772, + 0.0043095555156469345, + 0.004766898229718208, + 0.0048506539314985275, + 0.005885209422558546, + 0.019881658256053925, + 0.03635748475790024, + 0.029281506314873695, + 0.006279808469116688, + 0.005241198930889368, + 0.004391964990645647, + 0.0035800531040877104, + 0.01287090964615345, + 0.02434326522052288, + 0.03881044313311577, + 0.016308752819895744, + 0.012470025569200516, + 0.010793418623507023, + 0.019485803321003914, + 0.009831488132476807, + 0.021095553413033485, + 0.013176806271076202, + 0.024775510653853416, + 0.018205421045422554, + 0.004377586767077446, + 0.00370280584320426, + 0.004394025541841984, + 0.0026942144613713026, + 0.006361167412251234, + 0.011936371214687824, + 0.01593363843858242, + 0.014617202803492546, + 0.010858184657990932, + 0.012190982699394226, + 0.02264571376144886, + 0.008374588564038277, + 0.01514484267681837, + 0.010211944580078125, + 0.011180094443261623, + 0.016336414963006973, + 0.012586093507707119, + 0.011167152784764767, + 0.011276952922344208, + 0.013030532747507095, + 0.023734446614980698, + 0.022383064031600952, + 0.038047514855861664, + 0.03201666101813316, + 0.007118115201592445, + 0.0073431930504739285, + 0.009187339805066586, + 0.008124449290335178, + 0.040540967136621475, + 0.003038560040295124, + 0.002646668115630746, + 0.002314812270924449, + 0.0018887304468080401, + 0.0021173569839447737, + 0.0021821765694767237, + 0.0022215459030121565, + 0.0032785998191684484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 1 + ], + "selected_reward": -0.0125464191660285 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038095925003290176 + ], + "opposite_logprob_tokens": [ + -3.2880959510803223 + ], + "logprob_reward": -0.038095925003290176, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038095925003290176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1149604544043541, + 0.0398661307990551, + 0.14318986237049103, + 0.05301252752542496, + 0.039293717592954636, + 0.04382710158824921, + 0.0561617836356163, + 0.14500801265239716, + 0.35181567072868347, + 0.01286469865590334 + ], + "credit_map": [ + 0.020978767424821854, + 0.038373857736587524, + 0.017088931053876877, + 0.007057204842567444, + 0.007352347951382399, + 0.008840384893119335, + 0.007262249011546373, + 0.008006712421774864, + 0.004080076236277819, + 0.004428458865731955, + 0.006483456585556269, + 0.004080210346728563, + 0.0070738778449594975, + 0.005355577915906906, + 0.003747573122382164, + 0.004616896156221628, + 0.029895581305027008, + 0.053587671369314194, + 0.02446519024670124, + 0.008510054089128971, + 0.006994879338890314, + 0.0048754229210317135, + 0.006413807161152363, + 0.008447260595858097, + 0.007564148865640163, + 0.011325846426188946, + 0.009444426745176315, + 0.004185528960078955, + 0.003055345732718706, + 0.005238926503807306, + 0.004365467466413975, + 0.007832838222384453, + 0.0065419841557741165, + 0.009204098954796791, + 0.004156997427344322, + 0.0032093082554638386, + 0.003253933507949114, + 0.004576009698212147, + 0.0027281874790787697, + 0.005623199511319399, + 0.008213111199438572, + 0.014065936207771301, + 0.005410060286521912, + 0.003609913168475032, + 0.0023443305399268866, + 0.004371867049485445, + 0.0021972854156047106, + 0.003614599583670497, + 0.006068125832825899, + 0.007352772634476423, + 0.008274371735751629, + 0.005989724770188332, + 0.004399089142680168, + 0.008498288691043854, + 0.005018129013478756, + 0.010561279952526093, + 0.013831895776093006, + 0.01641511730849743, + 0.0211354847997427, + 0.01675802282989025, + 0.014866338111460209, + 0.013566181063652039, + 0.01644684374332428, + 0.031988125294446945, + 0.051260631531476974, + 0.08847326040267944, + 0.0651770755648613, + 0.014837006106972694, + 0.012830812484025955, + 0.015751294791698456, + 0.016861947253346443, + 0.08662363886833191, + 0.0018012371147051454, + 0.0020348329562693834, + 0.0020678567234426737, + 0.0009033220121636987, + 0.0008801931980997324, + 0.0015881613362580538, + 0.0009581928607076406, + 0.002630902687087655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.038095925003290176 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328523755073547 + ], + "opposite_logprob_tokens": [ + -0.7578523755073547 + ], + "logprob_reward": -0.6328523755073547, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328523755073547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1715705394744873, + 0.03983724117279053, + 0.07808219641447067, + 0.02852727845311165, + 0.062274087220430374, + 0.03145769611001015, + 0.07588540017604828, + 0.08345597237348557, + 0.3549203872680664, + 0.07398923486471176 + ], + "credit_map": [ + 0.031675297766923904, + 0.05414683744311333, + 0.026548348367214203, + 0.011063476093113422, + 0.011422460898756981, + 0.01446289848536253, + 0.01087320875376463, + 0.011378007009625435, + 0.003964140545576811, + 0.0047052353620529175, + 0.005640413612127304, + 0.004632093012332916, + 0.0055963341146707535, + 0.005206630099564791, + 0.0035658106207847595, + 0.006526584271341562, + 0.014116114936769009, + 0.022441381588578224, + 0.00805946346372366, + 0.006938624661415815, + 0.00647369772195816, + 0.005952741019427776, + 0.0066983625292778015, + 0.007401805371046066, + 0.004000843968242407, + 0.005398150999099016, + 0.004068638663738966, + 0.0031349516939371824, + 0.0026470187585800886, + 0.0025957939215004444, + 0.0025838955771178007, + 0.004097985569387674, + 0.008786244317889214, + 0.014122501946985722, + 0.007629512809216976, + 0.0057859416119754314, + 0.005080299451947212, + 0.00884240586310625, + 0.004223911557346582, + 0.007803269196301699, + 0.004664825275540352, + 0.007798239588737488, + 0.0034440679010003805, + 0.0029349359683692455, + 0.0025462168268859386, + 0.0022327478509396315, + 0.002692967187613249, + 0.0051436941139400005, + 0.00779710104689002, + 0.01147221028804779, + 0.009551875293254852, + 0.007231215015053749, + 0.0083345677703619, + 0.015842920169234276, + 0.005921998992562294, + 0.009733508341014385, + 0.008236424066126347, + 0.010875040665268898, + 0.010649739764630795, + 0.01033090054988861, + 0.0101047083735466, + 0.00918964296579361, + 0.009184552356600761, + 0.014884963631629944, + 0.03941218554973602, + 0.07019051909446716, + 0.06358642876148224, + 0.018673229962587357, + 0.022427469491958618, + 0.03167431429028511, + 0.021924780681729317, + 0.08703146874904633, + 0.008675416931509972, + 0.005661867558956146, + 0.019858509302139282, + 0.006396676413714886, + 0.00595824234187603, + 0.007419929839670658, + 0.007555396296083927, + 0.012463197112083435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.6328523755073547 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle, long shot captured from a level perspective, the camera remains still, the composition guided by the principle of extreme simplicity, focusing entirely on the endless golden rice fields before them. At the golden section of the frame, a lone figure, deeply bent, almost blends in with the throngs of stalks of rice. This is a woman wearing a wide-brimmed straw hat, its brim casting a shadow that obscures her face, revealing only her simple bun and the faded indigo cotton top that contrasts sharply with the golden color. The warm, soft afternoon sun paints the entire rice field a dazzling golden hue, each grain appearing heavy and gleaming. Her movements are slow and focused, her left hand carefully gathering a handful of plump rice stalks, while her right hand, using a small sickle, steadily and deliberately cuts them from the root. The entire process is fluid and rhythmic, filled with reverence for the bountiful harvest. Behind her, golden rice waves extend all the way to the horizon where they meet the pure blue sky. The picture is clean and pure, with saturated and warm colors, creating a tranquil, fertile and poetic pastoral atmosphere.", + "seed": 18871, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_11/index_7381_seed_18871_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08780145547352732, + "vlm_logprob": -0.08780145547352732, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.08780145547352732, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.039342284202575684 + ], + "opposite_logprob_tokens": [ + -3.7893424034118652 + ], + "logprob_reward": -0.039342284202575684, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.039342284202575684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02053741179406643, + 0.019611336290836334, + 0.02561611495912075, + 0.0219397135078907, + 0.07801466435194016, + 0.0276297889649868, + 0.024783344939351082, + 0.6235637664794922, + 0.04685245826840401, + 0.11145143955945969 + ], + "credit_map": [ + 0.002984660444781184, + 0.001246740808710456, + 0.004267290234565735, + 0.001563259051181376, + 0.0023040405940264463, + 0.002652394585311413, + 0.0020735240541398525, + 0.003445503767579794, + 0.0015890614595264196, + 0.001023474265821278, + 0.0033724091481417418, + 0.0013830754905939102, + 0.0027655400335788727, + 0.002270113443955779, + 0.004112208262085915, + 0.00309545430354774, + 0.002690816530957818, + 0.0015045064501464367, + 0.004392392933368683, + 0.0024233905132859945, + 0.003492748597636819, + 0.003697022097185254, + 0.003701241686940193, + 0.0037139966152608395, + 0.0018482962623238564, + 0.0014146342873573303, + 0.0032510370947420597, + 0.0017972590867429972, + 0.00572969950735569, + 0.00286030862480402, + 0.0020133217331022024, + 0.0030251573771238327, + 0.007055279333144426, + 0.006104086060076952, + 0.011429394595324993, + 0.005789353512227535, + 0.015440588817000389, + 0.012298566289246082, + 0.009139479137957096, + 0.010757917538285255, + 0.002588381990790367, + 0.002256562700495124, + 0.0033917769324034452, + 0.0019960582721978426, + 0.007915626280009747, + 0.0029765507206320763, + 0.0027012189384549856, + 0.0038036121986806393, + 0.003529198234900832, + 0.0013697078684344888, + 0.004433651454746723, + 0.0021038714330643415, + 0.00198190426453948, + 0.0053725941106677055, + 0.002091479953378439, + 0.0039009375032037497, + 0.06263460963964462, + 0.034463174641132355, + 0.1066848412156105, + 0.03855319321155548, + 0.10868517309427261, + 0.12127377092838287, + 0.05576406419277191, + 0.09550492465496063, + 0.004969126079231501, + 0.002479914575815201, + 0.009124372154474258, + 0.0030700787901878357, + 0.003516415599733591, + 0.009697113186120987, + 0.0058247861452400684, + 0.008170649409294128, + 0.012345521710813046, + 0.0061132344417274, + 0.02016759291291237, + 0.008260826580226421, + 0.015339041128754616, + 0.019933735951781273, + 0.013152594678103924, + 0.016138900071382523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.039342284202575684 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00363522837869823 + ], + "opposite_logprob_tokens": [ + -5.628635406494141 + ], + "logprob_reward": -0.00363522837869823, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00363522837869823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05324436351656914, + 0.023201271891593933, + 0.04458390176296234, + 0.04693016782402992, + 0.45915487408638, + 0.016021687537431717, + 0.06113409623503685, + 0.19949078559875488, + 0.04013329744338989, + 0.05610556900501251 + ], + "credit_map": [ + 0.004859434440732002, + 0.0041228532791137695, + 0.006523591000586748, + 0.00632706843316555, + 0.009313268586993217, + 0.007302477955818176, + 0.007294387556612492, + 0.00750128086656332, + 0.0019412775291129947, + 0.0014694372657686472, + 0.004348767455667257, + 0.0023268666118383408, + 0.003711116034537554, + 0.0019762751180678606, + 0.004031433258205652, + 0.0033960987348109484, + 0.004158038180321455, + 0.002961091697216034, + 0.0067621939815580845, + 0.004528637044131756, + 0.0065749771893024445, + 0.006692277267575264, + 0.00686119869351387, + 0.006045484449714422, + 0.003777775214985013, + 0.002709914231672883, + 0.010083931498229504, + 0.0036178992595523596, + 0.006328446790575981, + 0.004893936682492495, + 0.007949522696435452, + 0.007568738888949156, + 0.044585514813661575, + 0.03669387474656105, + 0.0655508041381836, + 0.03419755771756172, + 0.08815314620733261, + 0.07675284147262573, + 0.04925260320305824, + 0.06396853178739548, + 0.0015702297678217292, + 0.0010141812963411212, + 0.002288336167111993, + 0.0013631600886583328, + 0.0028556636534631252, + 0.0016810671659186482, + 0.002754365326836705, + 0.0024946846533566713, + 0.00849922001361847, + 0.0035097645595669746, + 0.008744517341256142, + 0.004933690652251244, + 0.009140091016888618, + 0.010007204487919807, + 0.007835553959012032, + 0.008464052341878414, + 0.017665354534983635, + 0.012588955461978912, + 0.033236343413591385, + 0.0137006351724267, + 0.03770456090569496, + 0.03216773271560669, + 0.02038782835006714, + 0.03203936666250229, + 0.0044375513680279255, + 0.002210825216025114, + 0.007194047328084707, + 0.002467283746227622, + 0.0034162511583417654, + 0.009036113508045673, + 0.004503624979406595, + 0.006867597345262766, + 0.006292095873504877, + 0.0030376005452126265, + 0.00977060291916132, + 0.0042595271952450275, + 0.007880602031946182, + 0.009584340266883373, + 0.007066017482429743, + 0.008214784786105156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.00363522837869823 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547949329018593 + ], + "opposite_logprob_tokens": [ + -4.387547969818115 + ], + "logprob_reward": -0.012547949329018593, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547949329018593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04582921043038368, + 0.09923738986253738, + 0.07470586150884628, + 0.05604120343923569, + 0.5164433717727661, + 0.02269001491367817, + 0.03330368548631668, + 0.07550550252199173, + 0.032792095094919205, + 0.043451692909002304 + ], + "credit_map": [ + 0.005348069127649069, + 0.0034830253571271896, + 0.00646533677354455, + 0.003945851698517799, + 0.007981027476489544, + 0.0064048850908875465, + 0.0057839034125208855, + 0.006417104974389076, + 0.012086460366845131, + 0.005536336451768875, + 0.017217405140399933, + 0.007861397229135036, + 0.013986055739223957, + 0.017140671610832214, + 0.011747429147362709, + 0.013661621138453484, + 0.005843984894454479, + 0.004675003234297037, + 0.009411686100065708, + 0.006565257906913757, + 0.013898524455726147, + 0.009782485663890839, + 0.013896497897803783, + 0.010632418096065521, + 0.004553195089101791, + 0.0031796973198652267, + 0.010998880490660667, + 0.0044235787354409695, + 0.010809536091983318, + 0.006049221381545067, + 0.008105229586362839, + 0.007921858690679073, + 0.05065401270985603, + 0.04079173132777214, + 0.07586988061666489, + 0.03898564353585243, + 0.092992402613163, + 0.09046640247106552, + 0.05410921573638916, + 0.07257404178380966, + 0.001975425984710455, + 0.0013332635862752795, + 0.004639748018234968, + 0.0020066697616130114, + 0.0034968387335538864, + 0.00236680475063622, + 0.0033428999595344067, + 0.003528360743075609, + 0.003912133164703846, + 0.002586651360616088, + 0.004565186332911253, + 0.003515017917379737, + 0.004414741415530443, + 0.005409090779721737, + 0.00440932484343648, + 0.004491536412388086, + 0.007608344312757254, + 0.0050974260084331036, + 0.011342897079885006, + 0.005937215872108936, + 0.016733728349208832, + 0.009115300141274929, + 0.00840654969215393, + 0.011264036409556866, + 0.00333241350017488, + 0.001929445774294436, + 0.006909629795700312, + 0.0025627724826335907, + 0.002432977082207799, + 0.006315901409834623, + 0.003590045962482691, + 0.005718907341361046, + 0.00477592321112752, + 0.0027111705858260393, + 0.008189016953110695, + 0.004226891323924065, + 0.0053499843925237656, + 0.006864629220217466, + 0.005012616515159607, + 0.006321455352008343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.012547949329018593 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002870738506317 + ], + "opposite_logprob_tokens": [ + -2.695028781890869 + ], + "logprob_reward": -0.07002870738506317, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002870738506317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048904839903116226, + 0.03770441561937332, + 0.08174912631511688, + 0.03674297779798508, + 0.5559585690498352, + 0.007379401475191116, + 0.10295963287353516, + 0.05855453386902809, + 0.05589156597852707, + 0.014154898934066296 + ], + "credit_map": [ + 0.004584228154271841, + 0.0037768555339425802, + 0.009142234921455383, + 0.005014854948967695, + 0.006801292300224304, + 0.006776433903723955, + 0.0050848922692239285, + 0.0077240485697984695, + 0.0028153257444500923, + 0.0020264051854610443, + 0.004865868948400021, + 0.0027416504453867674, + 0.006657752208411694, + 0.0035551870241761208, + 0.009395018219947815, + 0.005647210404276848, + 0.005461074877530336, + 0.004958499688655138, + 0.010614605620503426, + 0.006971466355025768, + 0.015129292383790016, + 0.008510780520737171, + 0.016818540170788765, + 0.013284865766763687, + 0.003586826380342245, + 0.002880376996472478, + 0.006118172779679298, + 0.00312160630710423, + 0.007188624702394009, + 0.004370061215013266, + 0.004417341202497482, + 0.005059968680143356, + 0.054019179195165634, + 0.043681081384420395, + 0.08115451782941818, + 0.04077541083097458, + 0.10482974350452423, + 0.09516292065382004, + 0.058765944093465805, + 0.07756979018449783, + 0.0008337230538018048, + 0.0005521610146388412, + 0.0010235031368210912, + 0.0005053350469097495, + 0.0017441930249333382, + 0.0007300674915313721, + 0.000767853984143585, + 0.0012225649552419782, + 0.014154147356748581, + 0.006303783506155014, + 0.01845805160701275, + 0.008573397062718868, + 0.009121543727815151, + 0.022610338404774666, + 0.008905560709536076, + 0.014832808636128902, + 0.006796048954129219, + 0.004407215863466263, + 0.012435375712811947, + 0.004107311367988586, + 0.00597134605050087, + 0.010574609972536564, + 0.005966896656900644, + 0.00829573068767786, + 0.005698325578123331, + 0.0029409106355160475, + 0.011154565028846264, + 0.0037082501221448183, + 0.004358573351055384, + 0.01192877721041441, + 0.005919738672673702, + 0.010182428173720837, + 0.0015664355596527457, + 0.001315321889705956, + 0.0027795869391411543, + 0.001963423565030098, + 0.0011963150463998318, + 0.0018631704151630402, + 0.0015977617586031556, + 0.001872883876785636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.07002870738506317 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134531080722809 + ], + "opposite_logprob_tokens": [ + -1.3134530782699585 + ], + "logprob_reward": -0.3134531080722809, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134531080722809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02500130794942379, + 0.013350419700145721, + 0.026416301727294922, + 0.014432800002396107, + 0.28022366762161255, + 0.007802311796694994, + 0.055847689509391785, + 0.13470153510570526, + 0.04368113353848457, + 0.39854273200035095 + ], + "credit_map": [ + 0.0019441070035099983, + 0.0019359224243089557, + 0.003849106840789318, + 0.0026770122349262238, + 0.004106266424059868, + 0.0035045421682298183, + 0.002820409834384918, + 0.004163945559412241, + 0.0015776838408783078, + 0.0008425128180533648, + 0.002628416521474719, + 0.001244597602635622, + 0.0014422707026824355, + 0.0021130680106580257, + 0.0014865358825773, + 0.0020153371151536703, + 0.0025206871796399355, + 0.0016078342450782657, + 0.0037096429150551558, + 0.0024101082235574722, + 0.004702912177890539, + 0.003291116561740637, + 0.004390131216496229, + 0.003783876309171319, + 0.0010059978812932968, + 0.0008157260599546134, + 0.0022596900817006826, + 0.001073144841939211, + 0.002606053836643696, + 0.0016132354503497481, + 0.0026868463028222322, + 0.0023721083998680115, + 0.027241049334406853, + 0.021971702575683594, + 0.041244179010391235, + 0.0207216739654541, + 0.05174790695309639, + 0.04867156222462654, + 0.029546882957220078, + 0.03907877206802368, + 0.0006981422775425017, + 0.0006873382371850312, + 0.0006897702114656568, + 0.0005524863954633474, + 0.00232089520432055, + 0.0007855223375372589, + 0.0008645170019008219, + 0.001203642226755619, + 0.007626614533364773, + 0.0031192521564662457, + 0.00989721529185772, + 0.00468548433855176, + 0.005299979355186224, + 0.011893694289028645, + 0.005129611119627953, + 0.008195850066840649, + 0.012091655284166336, + 0.008472141809761524, + 0.02266320399940014, + 0.009276545606553555, + 0.024911096319556236, + 0.022966541349887848, + 0.013465160503983498, + 0.020855216309428215, + 0.004408893641084433, + 0.0022196886129677296, + 0.008831138722598553, + 0.0028703927528113127, + 0.0031831469386816025, + 0.008708733133971691, + 0.005790063180029392, + 0.007669085171073675, + 0.043107300996780396, + 0.021809568628668785, + 0.07243961840867996, + 0.030238108709454536, + 0.05538295581936836, + 0.07026006281375885, + 0.04637129232287407, + 0.05893393233418465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.3134531080722809 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle, long shot captured from a level perspective, the camera remains still, the composition guided by the principle of extreme simplicity, focusing entirely on the endless golden rice fields before them. At the golden section of the frame, a lone figure, deeply bent, almost blends in with the throngs of stalks of rice. This is a woman wearing a wide-brimmed straw hat, its brim casting a shadow that obscures her face, revealing only her simple bun and the faded indigo cotton top that contrasts sharply with the golden color. The warm, soft afternoon sun paints the entire rice field a dazzling golden hue, each grain appearing heavy and gleaming. Her movements are slow and focused, her left hand carefully gathering a handful of plump rice stalks, while her right hand, using a small sickle, steadily and deliberately cuts them from the root. The entire process is fluid and rhythmic, filled with reverence for the bountiful harvest. Behind her, golden rice waves extend all the way to the horizon where they meet the pure blue sky. The picture is clean and pure, with saturated and warm colors, creating a tranquil, fertile and poetic pastoral atmosphere.", + "seed": 18872, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_11/index_7381_seed_18872_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08684526772703975, + "vlm_logprob": -0.08684526772703975, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.08684526772703975, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.042276013642549515 + ], + "opposite_logprob_tokens": [ + -3.542275905609131 + ], + "logprob_reward": -0.042276013642549515, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.042276013642549515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11877238005399704, + 0.04566172882914543, + 0.11496781557798386, + 0.1265036165714264, + 0.1402587741613388, + 0.2806544303894043, + 0.026071442291140556, + 0.05240850895643234, + 0.06334669888019562, + 0.031354546546936035 + ], + "credit_map": [ + 0.011063743382692337, + 0.010193096473813057, + 0.022237401455640793, + 0.00903475470840931, + 0.010171771980822086, + 0.02571987546980381, + 0.012906509451568127, + 0.017445232719182968, + 0.004603350535035133, + 0.0053292992524802685, + 0.00557466596364975, + 0.0044716536067426205, + 0.009782560169696808, + 0.006549638696014881, + 0.004515437409281731, + 0.004835125990211964, + 0.01407693326473236, + 0.0158951785415411, + 0.01921394281089306, + 0.0067583899945020676, + 0.004922854248434305, + 0.04012960195541382, + 0.004627663176506758, + 0.009343244135379791, + 0.013617743737995625, + 0.025247443467378616, + 0.01073762122541666, + 0.011996275745332241, + 0.0160217247903347, + 0.029452107846736908, + 0.008913276717066765, + 0.010517437011003494, + 0.017518434673547745, + 0.03150784969329834, + 0.014461161568760872, + 0.012899153865873814, + 0.02171100489795208, + 0.018406597897410393, + 0.011299765668809414, + 0.012454807758331299, + 0.04994112253189087, + 0.06071034073829651, + 0.03512625768780708, + 0.02398541010916233, + 0.01994546502828598, + 0.037047985941171646, + 0.017036138102412224, + 0.036861713975667953, + 0.0027427198365330696, + 0.0026322011835873127, + 0.003154008649289608, + 0.003551230765879154, + 0.004875567741692066, + 0.0026542050763964653, + 0.0024958685971796513, + 0.003965640440583229, + 0.005345358047634363, + 0.005129368044435978, + 0.00811551883816719, + 0.00464274687692523, + 0.008279300294816494, + 0.01216509472578764, + 0.003812446491792798, + 0.00491867633536458, + 0.00685468502342701, + 0.00897823367267847, + 0.008337002247571945, + 0.008293199352920055, + 0.01088862493634224, + 0.006552375853061676, + 0.005417879205197096, + 0.008024697192013264, + 0.004143180325627327, + 0.0034950103145092726, + 0.0038207173347473145, + 0.0024245153181254864, + 0.006148578133434057, + 0.004798815120011568, + 0.003111894940957427, + 0.0034118336625397205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.042276013642549515 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017328736139461398 + ], + "opposite_logprob_tokens": [ + -6.37673282623291 + ], + "logprob_reward": -0.0017328736139461398, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017328736139461398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14469675719738007, + 0.045371249318122864, + 0.16376136243343353, + 0.06556232273578644, + 0.23524858057498932, + 0.07919326424598694, + 0.10601609945297241, + 0.03494381532073021, + 0.06575027108192444, + 0.059456247836351395 + ], + "credit_map": [ + 0.016189485788345337, + 0.012154205702245235, + 0.020272020250558853, + 0.019099703058600426, + 0.01606612280011177, + 0.015414908528327942, + 0.01617593877017498, + 0.029324375092983246, + 0.004706185776740313, + 0.007355207111686468, + 0.00521796103566885, + 0.005071610677987337, + 0.00621129060164094, + 0.0068480162881314754, + 0.003819165052846074, + 0.00614181300625205, + 0.014114219695329666, + 0.021045923233032227, + 0.018352271988987923, + 0.020406192168593407, + 0.020945565775036812, + 0.03809462860226631, + 0.015954896807670593, + 0.014847674407064915, + 0.008137533441185951, + 0.006492196582257748, + 0.008901300840079784, + 0.008755232207477093, + 0.008347658440470695, + 0.011379659175872803, + 0.00719499634578824, + 0.006353749893605709, + 0.02923259325325489, + 0.05143766477704048, + 0.022986816242337227, + 0.01199269574135542, + 0.012292099185287952, + 0.07577111572027206, + 0.011441691778600216, + 0.020093902945518494, + 0.010537516325712204, + 0.009915562346577644, + 0.008760446682572365, + 0.009017761796712875, + 0.014156172052025795, + 0.010532881133258343, + 0.00764263141900301, + 0.008630297146737576, + 0.016323275864124298, + 0.011076545342803001, + 0.013889671303331852, + 0.008664668537676334, + 0.01959148235619068, + 0.01673281006515026, + 0.008674408309161663, + 0.011063238605856895, + 0.003967588301748037, + 0.0032330069225281477, + 0.004940020386129618, + 0.0029111774638295174, + 0.005836450029164553, + 0.007672018371522427, + 0.0029859733767807484, + 0.0033975804690271616, + 0.009446144104003906, + 0.010972593910992146, + 0.006187545135617256, + 0.007389109581708908, + 0.007037864066660404, + 0.009573221206665039, + 0.006970150861889124, + 0.008173640817403793, + 0.005330925341695547, + 0.007601635064929724, + 0.008475505746901035, + 0.0074218204244971275, + 0.005310584791004658, + 0.007210866082459688, + 0.006174225360155106, + 0.01193068828433752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.0017328736139461398 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078042909502983 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.011078042909502983, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078042909502983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09405117481946945, + 0.03755984082818031, + 0.3541635572910309, + 0.05786198377609253, + 0.28092578053474426, + 0.04506196826696396, + 0.038592323660850525, + 0.0245555117726326, + 0.02657357044517994, + 0.04065428301692009 + ], + "credit_map": [ + 0.010402110405266285, + 0.009292501024901867, + 0.011379088275134563, + 0.011831121519207954, + 0.016609640792012215, + 0.011500698514282703, + 0.010196857154369354, + 0.01283915713429451, + 0.004190598614513874, + 0.004956463351845741, + 0.003798818914219737, + 0.004316681530326605, + 0.007184023503214121, + 0.004457484930753708, + 0.0042009237222373486, + 0.0044548423029482365, + 0.04115152359008789, + 0.04659493267536163, + 0.052249882370233536, + 0.01996254362165928, + 0.0322449691593647, + 0.11910077929496765, + 0.016985254362225533, + 0.025873666629195213, + 0.006501172669231892, + 0.006564450450241566, + 0.0073270476423203945, + 0.006359714083373547, + 0.010623716749250889, + 0.00964348204433918, + 0.004731461405754089, + 0.006110938731580973, + 0.03488866612315178, + 0.05107146129012108, + 0.039963651448488235, + 0.011933443136513233, + 0.01477942243218422, + 0.10300251841545105, + 0.007822088897228241, + 0.017464546486735344, + 0.003699693363159895, + 0.004860904533416033, + 0.006307501811534166, + 0.006486821454018354, + 0.006676883436739445, + 0.006245283409953117, + 0.004375545773655176, + 0.006409334484487772, + 0.006475753616541624, + 0.0044701783917844296, + 0.0042714569717645645, + 0.0037167705595493317, + 0.004633612465113401, + 0.007906801998615265, + 0.0026503675617277622, + 0.004467381630092859, + 0.0027664625085890293, + 0.0023937926162034273, + 0.003748210147023201, + 0.002768383827060461, + 0.0037695213686674833, + 0.004555132705718279, + 0.0018780974205583334, + 0.002675912342965603, + 0.0035831874702125788, + 0.0024390732869505882, + 0.003144380869343877, + 0.0037785526365041733, + 0.004042374901473522, + 0.004182059783488512, + 0.002576257102191448, + 0.0028276846278458834, + 0.0045568691566586494, + 0.005789445713162422, + 0.005308697931468487, + 0.0029849831480532885, + 0.007659201044589281, + 0.004877070896327496, + 0.004193709697574377, + 0.005284306593239307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.011078042909502983 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704385817050934 + ], + "opposite_logprob_tokens": [ + -2.1270439624786377 + ], + "logprob_reward": -0.12704385817050934, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704385817050934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16013900935649872, + 0.034483712166547775, + 0.2823351323604584, + 0.08326677232980728, + 0.17293794453144073, + 0.0944955125451088, + 0.026880528777837753, + 0.034835830330848694, + 0.022358426824212074, + 0.08826708048582077 + ], + "credit_map": [ + 0.020585326477885246, + 0.014629719778895378, + 0.028976641595363617, + 0.009282289072871208, + 0.01963759772479534, + 0.04359770193696022, + 0.010181987658143044, + 0.013247750699520111, + 0.0035779885947704315, + 0.004184201825410128, + 0.003529166104272008, + 0.003405379829928279, + 0.008364765904843807, + 0.003613254288211465, + 0.003800982376560569, + 0.004007975570857525, + 0.03600512072443962, + 0.0411241352558136, + 0.04503375664353371, + 0.014513506554067135, + 0.013261400163173676, + 0.1034419983625412, + 0.008556071668863297, + 0.02039918676018715, + 0.00917959213256836, + 0.009476080536842346, + 0.012285367585718632, + 0.00665385602042079, + 0.012110921554267406, + 0.02020125463604927, + 0.005314706824719906, + 0.008044996298849583, + 0.022162849083542824, + 0.024788688868284225, + 0.021794484928250313, + 0.007765404414385557, + 0.026290524750947952, + 0.049281053245067596, + 0.010304123163223267, + 0.010550826787948608, + 0.014337251894176006, + 0.011000605300068855, + 0.009694311767816544, + 0.00901839416474104, + 0.016092002391815186, + 0.017646823078393936, + 0.0072671216912567616, + 0.009439006447792053, + 0.00285530649125576, + 0.0020760856568813324, + 0.0037341236602514982, + 0.0029681490268558264, + 0.0044778259471058846, + 0.004310619551688433, + 0.0019225651631131768, + 0.004535853397101164, + 0.0034580377396196127, + 0.004291224293410778, + 0.0036791174206882715, + 0.0030993896070867777, + 0.008526707999408245, + 0.004898053593933582, + 0.00287718721665442, + 0.004006112925708294, + 0.00228657154366374, + 0.0025648989249020815, + 0.002681170590221882, + 0.00275408779270947, + 0.004193450324237347, + 0.0021423143334686756, + 0.003350697224959731, + 0.002385236555710435, + 0.010486895218491554, + 0.01054912805557251, + 0.01109033077955246, + 0.008642882108688354, + 0.018005114048719406, + 0.011199860833585262, + 0.009390308521687984, + 0.008902563713490963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.12704385817050934 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520955502986908 + ], + "opposite_logprob_tokens": [ + -1.5020955801010132 + ], + "logprob_reward": -0.2520955502986908, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520955502986908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17471565306186676, + 0.07251285016536713, + 0.12070699036121368, + 0.02841886319220066, + 0.25832071900367737, + 0.10121041536331177, + 0.03592757508158684, + 0.03329778090119362, + 0.07852014154195786, + 0.09636907279491425 + ], + "credit_map": [ + 0.024946192279458046, + 0.017206504940986633, + 0.026475898921489716, + 0.016627898439764977, + 0.014229562133550644, + 0.042064450681209564, + 0.013217166997492313, + 0.0199479591101408, + 0.00985440332442522, + 0.008701642975211143, + 0.007251969538629055, + 0.005594913847744465, + 0.011345875449478626, + 0.014964175410568714, + 0.007093415129929781, + 0.007706451695412397, + 0.013954955153167248, + 0.016629092395305634, + 0.020551731809973717, + 0.007081835065037012, + 0.0052249012514948845, + 0.042349860072135925, + 0.005157387815415859, + 0.00975722074508667, + 0.003483861219137907, + 0.0037699989043176174, + 0.004328091628849506, + 0.003040504874661565, + 0.0021715813782066107, + 0.00657295947894454, + 0.0018381186528131366, + 0.003213743679225445, + 0.03206108137965202, + 0.047752100974321365, + 0.0358397476375103, + 0.011344235390424728, + 0.012348405085504055, + 0.09509813785552979, + 0.008548159152269363, + 0.015328824520111084, + 0.01592027209699154, + 0.01854880340397358, + 0.014576517045497894, + 0.00907539390027523, + 0.007195291109383106, + 0.011597384698688984, + 0.007684692740440369, + 0.016612054780125618, + 0.0036244280636310577, + 0.0031428541988134384, + 0.005517278332263231, + 0.005799249745905399, + 0.0038368050009012222, + 0.00484103849157691, + 0.0032439399510622025, + 0.0059219771064817905, + 0.0032694425899535418, + 0.003888338804244995, + 0.0035428516566753387, + 0.004818940069526434, + 0.005799120292067528, + 0.0036218625027686357, + 0.003676565596833825, + 0.004680657759308815, + 0.01315458957105875, + 0.008169276639819145, + 0.00852851290255785, + 0.009753115475177765, + 0.008699705824255943, + 0.013043411076068878, + 0.008887424133718014, + 0.008284098468720913, + 0.007142843212932348, + 0.010075053200125694, + 0.011646328493952751, + 0.019492657855153084, + 0.019647708162665367, + 0.008767557330429554, + 0.010216177441179752, + 0.00938074104487896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.2520955502986908 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle, long shot captured from a level perspective, the camera remains still, the composition guided by the principle of extreme simplicity, focusing entirely on the endless golden rice fields before them. At the golden section of the frame, a lone figure, deeply bent, almost blends in with the throngs of stalks of rice. This is a woman wearing a wide-brimmed straw hat, its brim casting a shadow that obscures her face, revealing only her simple bun and the faded indigo cotton top that contrasts sharply with the golden color. The warm, soft afternoon sun paints the entire rice field a dazzling golden hue, each grain appearing heavy and gleaming. Her movements are slow and focused, her left hand carefully gathering a handful of plump rice stalks, while her right hand, using a small sickle, steadily and deliberately cuts them from the root. The entire process is fluid and rhythmic, filled with reverence for the bountiful harvest. Behind her, golden rice waves extend all the way to the horizon where they meet the pure blue sky. The picture is clean and pure, with saturated and warm colors, creating a tranquil, fertile and poetic pastoral atmosphere.", + "seed": 18873, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_11/index_7381_seed_18873_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09522385839372874, + "vlm_logprob": -0.09522385839372874, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.09522385839372874, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033086810261011124 + ], + "opposite_logprob_tokens": [ + -4.408086776733398 + ], + "logprob_reward": -0.033086810261011124, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.033086810261011124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2856232821941376, + 0.06278618425130844, + 0.07840976119041443, + 0.06889718770980835, + 0.06960386037826538, + 0.04353829845786095, + 0.14871814846992493, + 0.1168588250875473, + 0.08570567518472672, + 0.039858803153038025 + ], + "credit_map": [ + 0.03564399108290672, + 0.03266020491719246, + 0.032618969678878784, + 0.033627480268478394, + 0.03202816843986511, + 0.0595228374004364, + 0.034094080328941345, + 0.025427531450986862, + 0.010015487670898438, + 0.009695617482066154, + 0.007943442091345787, + 0.005307876504957676, + 0.006122867576777935, + 0.012717250734567642, + 0.006418497767299414, + 0.004565136507153511, + 0.008603665046393871, + 0.009605378843843937, + 0.0073304916732013226, + 0.006254410836845636, + 0.007892390713095665, + 0.024062341079115868, + 0.007225066889077425, + 0.007436013780534267, + 0.007435234729200602, + 0.007285890635102987, + 0.013269881717860699, + 0.01163946371525526, + 0.005047629587352276, + 0.009935606271028519, + 0.007373644970357418, + 0.006909830495715141, + 0.007521266117691994, + 0.008812743239104748, + 0.010993128642439842, + 0.010584318079054356, + 0.00562250055372715, + 0.011109602637588978, + 0.007009202614426613, + 0.007951095700263977, + 0.005873386282473803, + 0.004782408941537142, + 0.005522518418729305, + 0.005379543639719486, + 0.004305134993046522, + 0.007746808230876923, + 0.004925781860947609, + 0.005002713296562433, + 0.016317442059516907, + 0.015954656526446342, + 0.012859784066677094, + 0.011911549605429173, + 0.020862676203250885, + 0.0312718003988266, + 0.026751065626740456, + 0.0127891656011343, + 0.010479781776666641, + 0.014606449753046036, + 0.015280250459909439, + 0.015518191270530224, + 0.013731474056839943, + 0.009175925515592098, + 0.02581583894789219, + 0.012250899337232113, + 0.00910699088126421, + 0.010312163271009922, + 0.00709250383079052, + 0.007937692105770111, + 0.014913658611476421, + 0.014413159340620041, + 0.012758024036884308, + 0.009171470068395138, + 0.004914707038551569, + 0.006948551163077354, + 0.0048644402995705605, + 0.003755247453227639, + 0.0032219409476965666, + 0.006917433347553015, + 0.004439746029675007, + 0.00479673407971859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.033086810261011124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008662977255880833 + ], + "opposite_logprob_tokens": [ + -4.758663177490234 + ], + "logprob_reward": -0.008662977255880833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008662977255880833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2087482362985611, + 0.2129039317369461, + 0.09715304523706436, + 0.04821467027068138, + 0.16887842118740082, + 0.028308724984526634, + 0.023124542087316513, + 0.0940203070640564, + 0.09526476263999939, + 0.02338331751525402 + ], + "credit_map": [ + 0.020386099815368652, + 0.02265019528567791, + 0.01973603665828705, + 0.025144975632429123, + 0.02544051595032215, + 0.05183833837509155, + 0.0219770185649395, + 0.02157505601644516, + 0.021514631807804108, + 0.02440027892589569, + 0.024460792541503906, + 0.02129139006137848, + 0.026302814483642578, + 0.046313121914863586, + 0.027299432083964348, + 0.021321479231119156, + 0.01047463808208704, + 0.010988461785018444, + 0.011569097638130188, + 0.011205543763935566, + 0.012169119901955128, + 0.01638789102435112, + 0.01534732524305582, + 0.00901096686720848, + 0.005050411447882652, + 0.005870339460670948, + 0.005631417967379093, + 0.008969869464635849, + 0.004726109094917774, + 0.007712873164564371, + 0.005151032004505396, + 0.005102618597447872, + 0.01777707226574421, + 0.019965633749961853, + 0.017674937844276428, + 0.01780831627547741, + 0.015494463965296745, + 0.05061670392751694, + 0.015611664392054081, + 0.013929619453847408, + 0.0032735506538301706, + 0.003799908561632037, + 0.002043452113866806, + 0.0023450287990272045, + 0.003283763537183404, + 0.008219352923333645, + 0.0026832742150872946, + 0.002660395111888647, + 0.0031138898339122534, + 0.0038656096439808607, + 0.002824177034199238, + 0.0023153037764132023, + 0.0020369815174490213, + 0.003356724977493286, + 0.0029423688538372517, + 0.0026694866828620434, + 0.007940024137496948, + 0.012473167851567268, + 0.010713371448218822, + 0.011415978893637657, + 0.011870313435792923, + 0.008577981032431126, + 0.02103922702372074, + 0.009990247897803783, + 0.008213276043534279, + 0.011411692947149277, + 0.008733045309782028, + 0.01559322141110897, + 0.013907643035054207, + 0.011391142383217812, + 0.013307513669133186, + 0.01270722784101963, + 0.0027154874987900257, + 0.0038405281957238913, + 0.0023350794799625874, + 0.003164369612932205, + 0.0022694584913551807, + 0.004113267175853252, + 0.002243512310087681, + 0.002701614750549197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.008662977255880833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020595036447048187 + ], + "opposite_logprob_tokens": [ + -3.895595073699951 + ], + "logprob_reward": -0.020595036447048187, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020595036447048187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20362059772014618, + 0.20763355493545532, + 0.063308946788311, + 0.09633325785398483, + 0.10169408470392227, + 0.05005680397152901, + 0.03373664617538452, + 0.05592227727174759, + 0.13647353649139404, + 0.0512203574180603 + ], + "credit_map": [ + 0.017140554264187813, + 0.018103284761309624, + 0.0368046872317791, + 0.04076457768678665, + 0.016455452889204025, + 0.03051590919494629, + 0.024519920349121094, + 0.01931620202958584, + 0.020825868472456932, + 0.024264706298708916, + 0.02225065603852272, + 0.024298854172229767, + 0.025015752762556076, + 0.04298214986920357, + 0.027007149532437325, + 0.02098841220140457, + 0.0070779286324977875, + 0.007764147128909826, + 0.00867251306772232, + 0.00834051426500082, + 0.007515264675021172, + 0.006626126356422901, + 0.010424972511827946, + 0.006887473165988922, + 0.01044558547437191, + 0.014817525632679462, + 0.007229113020002842, + 0.010228856466710567, + 0.011465378105640411, + 0.024614034220576286, + 0.009594280272722244, + 0.00793848093599081, + 0.013049601577222347, + 0.012356557883322239, + 0.010799159295856953, + 0.01219143532216549, + 0.009891648776829243, + 0.022033998742699623, + 0.011588017456233501, + 0.009783653542399406, + 0.005874670576304197, + 0.005493649281561375, + 0.007860197685658932, + 0.00787399709224701, + 0.0041472711600363255, + 0.009479613974690437, + 0.005309277214109898, + 0.004018122795969248, + 0.0034745812881737947, + 0.004519165493547916, + 0.0031445862259715796, + 0.003477547550573945, + 0.004051067866384983, + 0.007068542763590813, + 0.0046972185373306274, + 0.0033039324916899204, + 0.005686827003955841, + 0.009385564364492893, + 0.005093711894005537, + 0.005903157405555248, + 0.006407356821000576, + 0.009177408181130886, + 0.009366742335259914, + 0.004901506006717682, + 0.011534474790096283, + 0.015451507642865181, + 0.014474458992481232, + 0.02575341984629631, + 0.016653671860694885, + 0.014969060197472572, + 0.01562703587114811, + 0.02200988493859768, + 0.007691924460232258, + 0.008665461093187332, + 0.0070239827036857605, + 0.005234186537563801, + 0.0041584959253668785, + 0.007332856301218271, + 0.0053524174727499485, + 0.005761026870459318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.020595036447048187 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032153874635696 + ], + "opposite_logprob_tokens": [ + -2.3503215312957764 + ], + "logprob_reward": -0.10032153874635696, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032153874635696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2682289183139801, + 0.1870168149471283, + 0.08610835671424866, + 0.0922069251537323, + 0.05595508590340614, + 0.031056424602866173, + 0.05829956755042076, + 0.06879067420959473, + 0.11019859462976456, + 0.042138706892728806 + ], + "credit_map": [ + 0.02472570724785328, + 0.03327096253633499, + 0.023672038689255714, + 0.03273593634366989, + 0.03057675249874592, + 0.06857180595397949, + 0.025731287896633148, + 0.02894442155957222, + 0.018398145213723183, + 0.020956426858901978, + 0.020636696368455887, + 0.02667657844722271, + 0.019606616348028183, + 0.03166733309626579, + 0.027571305632591248, + 0.02150372602045536, + 0.00868292711675167, + 0.009803933091461658, + 0.011922954581677914, + 0.010239829309284687, + 0.01028208527714014, + 0.011095192283391953, + 0.01491818018257618, + 0.00916325580328703, + 0.011982395313680172, + 0.012167578563094139, + 0.011299827136099339, + 0.01315550971776247, + 0.008220917545258999, + 0.01716536656022072, + 0.008460354059934616, + 0.00975497905164957, + 0.007366057485342026, + 0.007356558460742235, + 0.0064870258793234825, + 0.00690426304936409, + 0.005795457400381565, + 0.011745922267436981, + 0.005351172294467688, + 0.0049486299976706505, + 0.003915552515536547, + 0.0031439107842743397, + 0.0048844399861991405, + 0.004036588594317436, + 0.0022420776076614857, + 0.006464400794357061, + 0.0028082896023988724, + 0.003561166813597083, + 0.006627582013607025, + 0.006381081882864237, + 0.004321924410760403, + 0.005144839640706778, + 0.007504642009735107, + 0.013192085549235344, + 0.010271037928760052, + 0.004856374114751816, + 0.006143791601061821, + 0.009158583357930183, + 0.007080471143126488, + 0.007146647199988365, + 0.008583596907556057, + 0.008847924880683422, + 0.014901631511747837, + 0.006928023416548967, + 0.010270120576024055, + 0.011910242959856987, + 0.008480183780193329, + 0.009680793620646, + 0.01967080682516098, + 0.018980074673891068, + 0.01885811984539032, + 0.012348252348601818, + 0.004705418366938829, + 0.006439453922212124, + 0.0038860554341226816, + 0.004241998307406902, + 0.004380742087960243, + 0.008437300100922585, + 0.005459565203636885, + 0.004588172305375338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.10032153874635696 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a wide-angle long shot of endless golden rice fields extending to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone woman wearing a wide-brimmed straw hat and a faded indigo cotton top visible in the golden section of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face obscured by the hat's shadow, revealing only a simple bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman using a small sickle in her right hand to cut rice stalks from the root while gathering them with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31345292925834656 + ], + "opposite_logprob_tokens": [ + -1.313452959060669 + ], + "logprob_reward": -0.31345292925834656, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31345292925834656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1127869263291359, + 0.09119956195354462, + 0.2567515969276428, + 0.04803967848420143, + 0.05594352260231972, + 0.021596267819404602, + 0.044131968170404434, + 0.09776999056339264, + 0.22637267410755157, + 0.045407772064208984 + ], + "credit_map": [ + 0.01137374248355627, + 0.012457340024411678, + 0.016449909657239914, + 0.012207461521029472, + 0.010388840921223164, + 0.025474589318037033, + 0.012593491934239864, + 0.011841545812785625, + 0.006763591431081295, + 0.010170227847993374, + 0.0102261146530509, + 0.009538422338664532, + 0.011183120310306549, + 0.017867160961031914, + 0.013910922221839428, + 0.011540000326931477, + 0.026517905294895172, + 0.030528105795383453, + 0.03457501158118248, + 0.029296699911355972, + 0.03241879865527153, + 0.0283050537109375, + 0.0503821037709713, + 0.02472792938351631, + 0.004698773846030235, + 0.004493820946663618, + 0.007377042900770903, + 0.007102084346115589, + 0.004756537731736898, + 0.009983058087527752, + 0.0049785105511546135, + 0.004649850074201822, + 0.005762119311839342, + 0.005927649792283773, + 0.009311378002166748, + 0.007126018404960632, + 0.004586918279528618, + 0.009836981073021889, + 0.00814993865787983, + 0.00524252001196146, + 0.0024888787884265184, + 0.0024506039917469025, + 0.0021807083394378424, + 0.0016814402770251036, + 0.002792060375213623, + 0.004983231425285339, + 0.002880838932469487, + 0.002138505456969142, + 0.005452114623039961, + 0.00628665741533041, + 0.006251571234315634, + 0.0044308933429419994, + 0.0038602310232818127, + 0.00884020235389471, + 0.004232860170304775, + 0.004777437075972557, + 0.009066768921911716, + 0.01199526060372591, + 0.01250002346932888, + 0.011703841388225555, + 0.010945981368422508, + 0.010676448233425617, + 0.020807575434446335, + 0.010074087418615818, + 0.022338513284921646, + 0.025976181030273438, + 0.02215615101158619, + 0.02139277197420597, + 0.03874160349369049, + 0.03473042696714401, + 0.03982437029480934, + 0.021212657913565636, + 0.004443960264325142, + 0.0061089154332876205, + 0.006798248738050461, + 0.004580967593938112, + 0.006037102546542883, + 0.00648502679541707, + 0.005606676451861858, + 0.00534687377512455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.31345292925834656 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_12/index_tensor([5391, 5391])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_12/index_tensor([5391, 5391])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e04b922ccbc215375f765d65e1b8a8dcbe81672e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_12/index_tensor([5391, 5391])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle shot, shot from above, captures this apocalyptic scene in extreme super-slow motion. As the camera slowly and steadily pulls back, the full extent of the disaster is mercilessly revealed. Amidst Chen Kaige's intensely fateful lighting, the iconic narrow and steep stone steps of Jiufen Old Street in Taiwan are relentlessly engulfed by a raging landslide. Rows of large red lanterns flank the steps, emitting a warm, ominous crimson glow in the night and drizzling rain. This light casts eerie, turbulent shadows on the wet stone slabs and the impending torrent of destruction. Like a surging beast, the viscous, dark brown mud slowly tumbles upward in a near-solidified motion, carrying broken and shattered wooden beams, splintered planks, and other unrecognizable debris. The mud collides with the solid stone steps, stirring up a massive wave of rainwater and mud, each drop of the splashing liquid clearly visible in the slow motion. Hanging red lanterns swayed violently under the impact, some instantly swept into the torrent, their fragile paper coverings ripped apart in the mud, their light extinguished. Broken wooden beams tumbled and collided helplessly in the viscous liquid, emitting a dull and despairing echo. The entire painting creates a strong color contrast between the deep red of the lanterns, the dark brown of the mudslide, and the cool gray of the stone steps, poeticizing the process of destruction and presenting a heartbreaking, grand, and tragic aesthetic of destructive tragedy.", + "seed": 25674, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_12/index_5391_seed_25674_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.25287418998777866, + "vlm_logprob": -0.25287418998777866, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.25287418998777866, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21483390033245087 + ], + "opposite_logprob_tokens": [ + -1.8398338556289673 + ], + "logprob_reward": -0.21483390033245087, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.21483390033245087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19536641240119934, + 0.06368577480316162, + 0.09583660960197449, + 0.03589479252696037, + 0.05629769340157509, + 0.13054580986499786, + 0.10020377486944199, + 0.09310147166252136, + 0.13659214973449707, + 0.09247548133134842 + ], + "credit_map": [ + 0.017104364931583405, + 0.015103817917406559, + 0.035728197544813156, + 0.03644910827279091, + 0.009746543131768703, + 0.02405194751918316, + 0.032162804156541824, + 0.025019634515047073, + 0.0065742479637265205, + 0.006784675642848015, + 0.009662942960858345, + 0.008825749158859253, + 0.003241756232455373, + 0.008005117066204548, + 0.007068733684718609, + 0.013522554188966751, + 0.009943446144461632, + 0.02067173831164837, + 0.013838759623467922, + 0.010591725818812847, + 0.010243507102131844, + 0.013068241067230701, + 0.008156913332641125, + 0.009322269819676876, + 0.003465822199359536, + 0.0051820180378854275, + 0.005503922700881958, + 0.004766705445945263, + 0.004061862826347351, + 0.003996837884187698, + 0.0052422089502215385, + 0.0036754144821316004, + 0.0049622999504208565, + 0.011908072046935558, + 0.008080513216555119, + 0.007472310680896044, + 0.00592695502564311, + 0.006964477244764566, + 0.004327826201915741, + 0.0066552371717989445, + 0.016137562692165375, + 0.014646418392658234, + 0.023975064978003502, + 0.012079895474016666, + 0.013371332548558712, + 0.022387783974409103, + 0.012039732187986374, + 0.0159080121666193, + 0.010040742345154285, + 0.00929900910705328, + 0.022374073043465614, + 0.013613191433250904, + 0.00999960396438837, + 0.014474180527031422, + 0.009301996789872646, + 0.011100979521870613, + 0.009978078305721283, + 0.011137004010379314, + 0.014987693168222904, + 0.011633350513875484, + 0.0110545689240098, + 0.01404986996203661, + 0.008664037100970745, + 0.011596866883337498, + 0.010111184790730476, + 0.00552977342158556, + 0.0448373444378376, + 0.007812734693288803, + 0.007219152525067329, + 0.04873206093907356, + 0.00587675953283906, + 0.0064731501042842865, + 0.010093137621879578, + 0.021005701273679733, + 0.01121474802494049, + 0.012177206575870514, + 0.008991421200335026, + 0.010853329673409462, + 0.008260853588581085, + 0.009879087097942829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.21483390033245087 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009831694886088371 + ], + "opposite_logprob_tokens": [ + -4.63483190536499 + ], + "logprob_reward": -0.009831694886088371, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009831694886088371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10940247774124146, + 0.05362051725387573, + 0.25529515743255615, + 0.2054034173488617, + 0.041817594319581985, + 0.07693971693515778, + 0.03547471761703491, + 0.106188103556633, + 0.04134458303451538, + 0.07451368123292923 + ], + "credit_map": [ + 0.01133924350142479, + 0.023124441504478455, + 0.01698826439678669, + 0.012610381469130516, + 0.008482339791953564, + 0.01531241089105606, + 0.01077939011156559, + 0.010766013525426388, + 0.006021637935191393, + 0.007869767025113106, + 0.008039651438593864, + 0.005671694874763489, + 0.0045621939934790134, + 0.008691756054759026, + 0.005644102580845356, + 0.007119715213775635, + 0.03720380365848541, + 0.04613343998789787, + 0.03719890117645264, + 0.024768587201833725, + 0.02736719138920307, + 0.04532686993479729, + 0.019765395671129227, + 0.017530960962176323, + 0.026799550279974937, + 0.04376823082566261, + 0.025718335062265396, + 0.015613910742104053, + 0.020881231874227524, + 0.03299552947282791, + 0.016488119959831238, + 0.023138517513871193, + 0.005347397644072771, + 0.009416002780199051, + 0.005281323101371527, + 0.005116259213536978, + 0.0048191905952990055, + 0.005655938293784857, + 0.0026032174937427044, + 0.0035782665945589542, + 0.007881524972617626, + 0.011211073026061058, + 0.014801752753555775, + 0.007492003031075001, + 0.005628560204058886, + 0.015753915533423424, + 0.006495404522866011, + 0.00767547870054841, + 0.004098951350897551, + 0.00637920992448926, + 0.004344036802649498, + 0.0027439903933554888, + 0.004448603838682175, + 0.005228201858699322, + 0.003490142058581114, + 0.004741581156849861, + 0.012157004326581955, + 0.017520925030112267, + 0.015742424875497818, + 0.015886789187788963, + 0.008736281655728817, + 0.01276121661067009, + 0.011728668585419655, + 0.011654789559543133, + 0.003619259921833873, + 0.005185490474104881, + 0.00981808453798294, + 0.003621104173362255, + 0.003799382597208023, + 0.009639718569815159, + 0.002810971112921834, + 0.002850571181625128, + 0.01122189685702324, + 0.017642123624682426, + 0.009007861837744713, + 0.00917052011936903, + 0.006529623176902533, + 0.009434276260435581, + 0.004720492288470268, + 0.0067868903279304504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.009831694886088371 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369717672467232 + ], + "opposite_logprob_tokens": [ + -3.4086971282958984 + ], + "logprob_reward": -0.03369717672467232, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369717672467232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2983834743499756, + 0.11184847354888916, + 0.0749812126159668, + 0.04483594372868538, + 0.05168647691607475, + 0.10684651881456375, + 0.07241540402173996, + 0.09000541269779205, + 0.06228996068239212, + 0.08670709282159805 + ], + "credit_map": [ + 0.03085281513631344, + 0.0440797284245491, + 0.048231806606054306, + 0.04462110623717308, + 0.026741081848740578, + 0.03940204903483391, + 0.029851365834474564, + 0.0346035361289978, + 0.012415695935487747, + 0.01820126175880432, + 0.016810305416584015, + 0.013254540041089058, + 0.013142772018909454, + 0.012868267484009266, + 0.009111329913139343, + 0.016044296324253082, + 0.007623166311532259, + 0.00811903178691864, + 0.012189347296953201, + 0.012443168088793755, + 0.004541464615613222, + 0.01160557009279728, + 0.010096525773406029, + 0.008362941443920135, + 0.004170130472630262, + 0.006323473062366247, + 0.006223604548722506, + 0.004785520955920219, + 0.00537680322304368, + 0.0067472513765096664, + 0.00552956061437726, + 0.0056795962154865265, + 0.006067090202122927, + 0.011056044138967991, + 0.006376562640070915, + 0.00682376092299819, + 0.005052528344094753, + 0.007143392693251371, + 0.003958708606660366, + 0.005208388436585665, + 0.010826659388840199, + 0.014769158326089382, + 0.014711359515786171, + 0.014477070420980453, + 0.013499797321856022, + 0.014484147541224957, + 0.012891076505184174, + 0.011187254451215267, + 0.005261081270873547, + 0.00627108383923769, + 0.015387502498924732, + 0.010290159843862057, + 0.008869255892932415, + 0.010585470125079155, + 0.008538850583136082, + 0.007211999036371708, + 0.008281370624899864, + 0.011137854307889938, + 0.0139755355194211, + 0.013723776675760746, + 0.008228223770856857, + 0.01381513848900795, + 0.010155662894248962, + 0.01068784948438406, + 0.008466171100735664, + 0.005234674550592899, + 0.011507277376949787, + 0.0076833623461425304, + 0.006324886344373226, + 0.010344219394028187, + 0.005099300295114517, + 0.0076300655491650105, + 0.009301849640905857, + 0.02066849172115326, + 0.010757912881672382, + 0.013168531470000744, + 0.007136051077395678, + 0.010186522267758846, + 0.007231550291180611, + 0.008256182074546814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.03369717672467232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988786101341248 + ], + "opposite_logprob_tokens": [ + -1.8048878908157349 + ], + "logprob_reward": -0.17988786101341248, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988786101341248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13801735639572144, + 0.07860124856233597, + 0.08944664895534515, + 0.24262318015098572, + 0.04815949127078056, + 0.056485287845134735, + 0.148647740483284, + 0.06755699217319489, + 0.09665815532207489, + 0.03380385413765907 + ], + "credit_map": [ + 0.013362021185457706, + 0.014324944466352463, + 0.025128057226538658, + 0.016703473404049873, + 0.008407821878790855, + 0.02017994411289692, + 0.02174517698585987, + 0.018165910616517067, + 0.008749261498451233, + 0.012505895458161831, + 0.010191697627305984, + 0.008646950125694275, + 0.006087366957217455, + 0.009023787453770638, + 0.008586175739765167, + 0.014810114167630672, + 0.011213577352464199, + 0.013188381679356098, + 0.014919711276888847, + 0.017575927078723907, + 0.005522280931472778, + 0.012771441601216793, + 0.007426335476338863, + 0.006828989367932081, + 0.028871750459074974, + 0.055650681257247925, + 0.029736898839473724, + 0.023745287209749222, + 0.02574928104877472, + 0.03696499392390251, + 0.01816336251795292, + 0.023740939795970917, + 0.004292203113436699, + 0.007872648537158966, + 0.006688998080790043, + 0.006638656370341778, + 0.004927122499793768, + 0.007869377732276917, + 0.004408319015055895, + 0.005462166387587786, + 0.005492474418133497, + 0.010005208663642406, + 0.010461442172527313, + 0.007638244424015284, + 0.005465992726385593, + 0.00771641731262207, + 0.0038525538984686136, + 0.00585295120254159, + 0.011180223897099495, + 0.009335881099104881, + 0.035312362015247345, + 0.021609311923384666, + 0.015547207556664944, + 0.02232383005321026, + 0.015769872814416885, + 0.0175690446048975, + 0.009184932336211205, + 0.0041216397657990456, + 0.008940061554312706, + 0.008702092804014683, + 0.008878620341420174, + 0.011794768273830414, + 0.007811799179762602, + 0.008123080246150494, + 0.006093260366469622, + 0.003029212122783065, + 0.03429039940237999, + 0.004989244043827057, + 0.004713867790997028, + 0.036311857402324677, + 0.0030698985792696476, + 0.004160409793257713, + 0.003183373250067234, + 0.003601817646995187, + 0.004817022942006588, + 0.004544482566416264, + 0.0024680430069565773, + 0.004313481040298939, + 0.004606038797646761, + 0.006269595120102167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.17988786101341248 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261203169822693 + ], + "opposite_logprob_tokens": [ + -0.5761203169822693 + ], + "logprob_reward": -0.8261203169822693, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261203169822693, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1470119059085846, + 0.05473118647933006, + 0.08782557398080826, + 0.09393477439880371, + 0.06069780886173248, + 0.09752282500267029, + 0.08438964188098907, + 0.10466240346431732, + 0.17861926555633545, + 0.09060459583997726 + ], + "credit_map": [ + 0.01691674068570137, + 0.025139937177300453, + 0.021595412865281105, + 0.0170173067599535, + 0.009417847730219364, + 0.029704388231039047, + 0.011813352815806866, + 0.015406928956508636, + 0.007651856634765863, + 0.00784390140324831, + 0.007364114746451378, + 0.005727611016482115, + 0.0054276371374726295, + 0.008226572535932064, + 0.005794718861579895, + 0.006694776937365532, + 0.009762337431311607, + 0.007640290539711714, + 0.015434793196618557, + 0.010508926585316658, + 0.010829729959368706, + 0.014462936669588089, + 0.009384053759276867, + 0.009802519343793392, + 0.009233268909156322, + 0.013009361922740936, + 0.015558354556560516, + 0.016290832310914993, + 0.00718811247497797, + 0.009672746993601322, + 0.014705830253660679, + 0.008276275359094143, + 0.006089868489652872, + 0.012894474901258945, + 0.008013824000954628, + 0.0061850836500525475, + 0.006470640189945698, + 0.006987863685935736, + 0.00855911523103714, + 0.005496948957443237, + 0.007639043033123016, + 0.013036019168794155, + 0.017949627712368965, + 0.011570975184440613, + 0.009871461428701878, + 0.015012352727353573, + 0.010001657530665398, + 0.012441693805158138, + 0.0056027695536613464, + 0.0067136227153241634, + 0.016372736543416977, + 0.012965384870767593, + 0.0111539326608181, + 0.01114281453192234, + 0.010051329620182514, + 0.010387060232460499, + 0.011013215407729149, + 0.0170141588896513, + 0.01310263853520155, + 0.012198975309729576, + 0.012506437487900257, + 0.013824591413140297, + 0.012485088780522346, + 0.012517310678958893, + 0.011147549375891685, + 0.004481167532503605, + 0.064130499958992, + 0.010580956935882568, + 0.009749325923621655, + 0.0650930106639862, + 0.004936595913022757, + 0.008500183932483196, + 0.009041126817464828, + 0.007357922848314047, + 0.011074075475335121, + 0.014930810779333115, + 0.009664209559559822, + 0.011015549302101135, + 0.014917631633579731, + 0.012603269889950752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.8261203169822693 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "low-angle shot, shot from above, captures this apocalyptic scene in extreme super-slow motion. As the camera slowly and steadily pulls back, the full extent of the disaster is mercilessly revealed. Amidst Chen Kaige's intensely fateful lighting, the iconic narrow and steep stone steps of Jiufen Old Street in Taiwan are relentlessly engulfed by a raging landslide. Rows of large red lanterns flank the steps, emitting a warm, ominous crimson glow in the night and drizzling rain. This light casts eerie, turbulent shadows on the wet stone slabs and the impending torrent of destruction. Like a surging beast, the viscous, dark brown mud slowly tumbles upward in a near-solidified motion, carrying broken and shattered wooden beams, splintered planks, and other unrecognizable debris. The mud collides with the solid stone steps, stirring up a massive wave of rainwater and mud, each drop of the splashing liquid clearly visible in the slow motion. Hanging red lanterns swayed violently under the impact, some instantly swept into the torrent, their fragile paper coverings ripped apart in the mud, their light extinguished. Broken wooden beams tumbled and collided helplessly in the viscous liquid, emitting a dull and despairing echo. The entire painting creates a strong color contrast between the deep red of the lanterns, the dark brown of the mudslide, and the cool gray of the stone steps, poeticizing the process of destruction and presenting a heartbreaking, grand, and tragic aesthetic of destructive tragedy.", + "seed": 25675, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_12/index_5391_seed_25675_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2635142344050109, + "vlm_logprob": -0.2635142344050109, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.2635142344050109, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2753705680370331 + ], + "opposite_logprob_tokens": [ + -1.5253705978393555 + ], + "logprob_reward": -0.2753705680370331, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2753705680370331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3083445727825165, + 0.08295149356126785, + 0.0664081871509552, + 0.08656097948551178, + 0.11173949390649796, + 0.09292605519294739, + 0.08337054401636124, + 0.05189037695527077, + 0.07853490859270096, + 0.03727346286177635 + ], + "credit_map": [ + 0.03437124937772751, + 0.046969715505838394, + 0.04624243453145027, + 0.03809148073196411, + 0.024667028337717056, + 0.02909919247031212, + 0.03864998370409012, + 0.05025346204638481, + 0.010890046134591103, + 0.008539889939129353, + 0.010770179331302643, + 0.009359747171401978, + 0.007755320053547621, + 0.012409811839461327, + 0.01147172600030899, + 0.011754770763218403, + 0.008229311555624008, + 0.0070133390836417675, + 0.007663656026124954, + 0.004808617755770683, + 0.008092058822512627, + 0.012408802285790443, + 0.00759936822578311, + 0.010593035258352757, + 0.008740535005927086, + 0.009638786315917969, + 0.012280061841011047, + 0.007853417657315731, + 0.010133977979421616, + 0.012888437137007713, + 0.011174938641488552, + 0.01385082770138979, + 0.014243938028812408, + 0.011592410504817963, + 0.011781059205532074, + 0.00877163466066122, + 0.01648937538266182, + 0.016579290851950645, + 0.014695394784212112, + 0.01758638024330139, + 0.010113976895809174, + 0.011605056934058666, + 0.012594188563525677, + 0.008823983371257782, + 0.013266903348267078, + 0.012215226888656616, + 0.01113918423652649, + 0.013167533092200756, + 0.010786855593323708, + 0.009685715660452843, + 0.008082938380539417, + 0.010554813779890537, + 0.011451051570475101, + 0.01042439229786396, + 0.01006995141506195, + 0.01231482531875372, + 0.007685523480176926, + 0.005822341423481703, + 0.008157805539667606, + 0.0062333401292562485, + 0.00456807017326355, + 0.00707395002245903, + 0.005166999064385891, + 0.007182344328612089, + 0.010923033580183983, + 0.00976928323507309, + 0.008464211598038673, + 0.009200307540595531, + 0.008782983757555485, + 0.009321283549070358, + 0.009112604893743992, + 0.012961205095052719, + 0.0075335875153541565, + 0.0045266770757734776, + 0.004073739051818848, + 0.00492519699037075, + 0.0033550846856087446, + 0.004476872738450766, + 0.0033317452762275934, + 0.0050505599938333035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.2753705680370331 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032412870787084103 + ], + "opposite_logprob_tokens": [ + -5.753241062164307 + ], + "logprob_reward": -0.0032412870787084103, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032412870787084103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21938371658325195, + 0.17562231421470642, + 0.09646329283714294, + 0.12637728452682495, + 0.10277225822210312, + 0.045206405222415924, + 0.044744376093149185, + 0.09411752223968506, + 0.05429166555404663, + 0.041021209210157394 + ], + "credit_map": [ + 0.02526930160820484, + 0.035142965614795685, + 0.03124590963125229, + 0.023747067898511887, + 0.019011827185750008, + 0.03388256952166557, + 0.022512948140501976, + 0.0285711158066988, + 0.021545972675085068, + 0.022765258327126503, + 0.01761946640908718, + 0.015592003241181374, + 0.01893266849219799, + 0.0336335264146328, + 0.02140815369784832, + 0.024125276133418083, + 0.010899681597948074, + 0.01286267302930355, + 0.008231621235609055, + 0.008880961686372757, + 0.012353687547147274, + 0.015304570086300373, + 0.011941662058234215, + 0.015988439321517944, + 0.012104921974241734, + 0.010116275399923325, + 0.02223910205066204, + 0.012508662417531013, + 0.01428213994950056, + 0.01779450848698616, + 0.01713733747601509, + 0.020194323733448982, + 0.012156427837908268, + 0.011709317564964294, + 0.016421359032392502, + 0.015871305018663406, + 0.010149212554097176, + 0.011127867735922337, + 0.00897710956633091, + 0.01635965332388878, + 0.004664108622819185, + 0.005932577420026064, + 0.009647199884057045, + 0.00564156798645854, + 0.0038770053070038557, + 0.0046583423390984535, + 0.00436085881665349, + 0.00642474414780736, + 0.005255379248410463, + 0.00626066280528903, + 0.005064471624791622, + 0.005629205144941807, + 0.004521789029240608, + 0.005071564577519894, + 0.006886755581945181, + 0.006054546684026718, + 0.011179313994944096, + 0.010794034227728844, + 0.006556616630405188, + 0.008604835718870163, + 0.015592223033308983, + 0.01685059443116188, + 0.010288650169968605, + 0.01425125077366829, + 0.008285308256745338, + 0.005580492317676544, + 0.006203747354447842, + 0.005736625753343105, + 0.006724746432155371, + 0.007523177191615105, + 0.0064225331880152225, + 0.007815034128725529, + 0.005696028470993042, + 0.004956532269716263, + 0.004109777510166168, + 0.004158935509622097, + 0.005666919983923435, + 0.007044789381325245, + 0.004332930780947208, + 0.005055295769125223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0032412870787084103 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000970095396042 + ], + "opposite_logprob_tokens": [ + -2.695009708404541 + ], + "logprob_reward": -0.07000970095396042, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000970095396042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20655469596385956, + 0.09941114485263824, + 0.11881821602582932, + 0.19331517815589905, + 0.12612485885620117, + 0.09276679903268814, + 0.0585111528635025, + 0.03600743040442467, + 0.038403090089559555, + 0.030087437480688095 + ], + "credit_map": [ + 0.025223305448889732, + 0.02511618286371231, + 0.02691216766834259, + 0.023413700982928276, + 0.01881479099392891, + 0.026972481980919838, + 0.03397061675786972, + 0.026131441816687584, + 0.013399150222539902, + 0.012275807559490204, + 0.01501328032463789, + 0.009762156754732132, + 0.008104143664240837, + 0.015309000387787819, + 0.010369548574090004, + 0.015178061090409756, + 0.015399204567074776, + 0.013254308141767979, + 0.01100832223892212, + 0.008146975189447403, + 0.015753377228975296, + 0.02342945896089077, + 0.012449976056814194, + 0.019376594573259354, + 0.0216074101626873, + 0.01674889400601387, + 0.02864660881459713, + 0.017993994057178497, + 0.02546563930809498, + 0.029977446421980858, + 0.024010682478547096, + 0.028864506632089615, + 0.012410325929522514, + 0.01610972359776497, + 0.01717398315668106, + 0.020852167159318924, + 0.010434981435537338, + 0.013817740604281425, + 0.0164241474121809, + 0.018901780247688293, + 0.010080375708639622, + 0.013721965253353119, + 0.015719709917902946, + 0.01146690919995308, + 0.010529167018830776, + 0.009903267957270145, + 0.00812672358006239, + 0.013218678534030914, + 0.006532279308885336, + 0.007145770825445652, + 0.005629146005958319, + 0.007865730673074722, + 0.006991642527282238, + 0.007612806744873524, + 0.007103585638105869, + 0.009630191139876842, + 0.003977921325713396, + 0.004005033522844315, + 0.004189396742731333, + 0.004063006956130266, + 0.0045095873065292835, + 0.006596206221729517, + 0.0038904878310859203, + 0.004775790963321924, + 0.004226062446832657, + 0.00473838672041893, + 0.005190768279135227, + 0.007023991551250219, + 0.0025114750023931265, + 0.004853423684835434, + 0.004300173372030258, + 0.005558811593800783, + 0.006793119478970766, + 0.0038631707429885864, + 0.002923324005678296, + 0.003571359207853675, + 0.0028899030294269323, + 0.003192618489265442, + 0.002455461071804166, + 0.004398480989038944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.07000970095396042 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279818534851074 + ], + "opposite_logprob_tokens": [ + -2.0177981853485107 + ], + "logprob_reward": -0.14279818534851074, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14279818534851074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3213994801044464, + 0.07945923507213593, + 0.18025779724121094, + 0.13057708740234375, + 0.10888439416885376, + 0.05951377749443054, + 0.03256707638502121, + 0.036773402243852615, + 0.029713207855820656, + 0.020854543894529343 + ], + "credit_map": [ + 0.04703880846500397, + 0.05041677504777908, + 0.03685964271426201, + 0.027577662840485573, + 0.027765410020947456, + 0.05180288106203079, + 0.03343159705400467, + 0.046506695449352264, + 0.008548012003302574, + 0.011396659538149834, + 0.010598507709801197, + 0.008593709208071232, + 0.006931107956916094, + 0.011713490821421146, + 0.009297850541770458, + 0.012379898689687252, + 0.024621497839689255, + 0.0219571515917778, + 0.01763487234711647, + 0.016183743253350258, + 0.022616157308220863, + 0.03520492836833, + 0.016380978748202324, + 0.025658471509814262, + 0.01096052210777998, + 0.012528820894658566, + 0.0316147617995739, + 0.013729149475693703, + 0.00861243437975645, + 0.015168905258178711, + 0.017247434705495834, + 0.020715059712529182, + 0.011094785295426846, + 0.015557363629341125, + 0.01127389632165432, + 0.01822936348617077, + 0.010019160807132721, + 0.01307651400566101, + 0.014954578131437302, + 0.01467873901128769, + 0.006527348887175322, + 0.008871568366885185, + 0.00819324515759945, + 0.0063901967369019985, + 0.0061066532507538795, + 0.007674957625567913, + 0.005781554616987705, + 0.009968251921236515, + 0.0032412270084023476, + 0.003564416663721204, + 0.004425953142344952, + 0.005528844892978668, + 0.002703421050682664, + 0.002648785477504134, + 0.00549587095156312, + 0.004958558827638626, + 0.005170781631022692, + 0.0034034382551908493, + 0.003574373899027705, + 0.0034463333431631327, + 0.0054529267363250256, + 0.006986580789089203, + 0.0033428675960749388, + 0.005396103486418724, + 0.00481752585619688, + 0.0030736448243260384, + 0.0029752429109066725, + 0.0030015818774700165, + 0.003655871609225869, + 0.0042362273670732975, + 0.003142191097140312, + 0.004810922779142857, + 0.0031285560689866543, + 0.0031086348462849855, + 0.00288248504512012, + 0.0025702633429318666, + 0.0019592083990573883, + 0.0025925857480615377, + 0.002046970883384347, + 0.0025658400263637304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.14279818534851074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.826151430606842 + ], + "opposite_logprob_tokens": [ + -0.576151430606842 + ], + "logprob_reward": -0.826151430606842, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.826151430606842, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.49226275086402893, + 0.09632502496242523, + 0.09376124292612076, + 0.0634031817317009, + 0.039928797632455826, + 0.02593303471803665, + 0.03270512446761131, + 0.045718129724264145, + 0.08122605085372925, + 0.028736699372529984 + ], + "credit_map": [ + 0.049615535885095596, + 0.06283078342676163, + 0.05364702641963959, + 0.08313336223363876, + 0.03201977536082268, + 0.0559481643140316, + 0.09339235723018646, + 0.06167573109269142, + 0.010617860592901707, + 0.013915619812905788, + 0.019555361941456795, + 0.01262444257736206, + 0.005215284414589405, + 0.007135932799428701, + 0.012065568938851357, + 0.015194958075881004, + 0.013829823583364487, + 0.011166676878929138, + 0.012238241732120514, + 0.013542658649384975, + 0.007046390790492296, + 0.009457046166062355, + 0.012492784298956394, + 0.013987618498504162, + 0.00879205297678709, + 0.005853498820215464, + 0.007649810519069433, + 0.0053426241502165794, + 0.00890004076063633, + 0.010811912827193737, + 0.006783336866647005, + 0.009269905276596546, + 0.0037427765782922506, + 0.00642640283331275, + 0.00547036761417985, + 0.006172961089760065, + 0.003301404882222414, + 0.004211870487779379, + 0.00561960693448782, + 0.004983409773558378, + 0.0025289307814091444, + 0.003562622470781207, + 0.004083189647644758, + 0.0030162439215928316, + 0.002597278915345669, + 0.0032102542463690042, + 0.003424870315939188, + 0.003509644651785493, + 0.004365535452961922, + 0.00413451949134469, + 0.002288675168529153, + 0.003595021553337574, + 0.005325312726199627, + 0.004751667845994234, + 0.003057698719203472, + 0.005186693277209997, + 0.0070264614187181, + 0.005260429345071316, + 0.004897318314760923, + 0.0051051597110927105, + 0.005166864022612572, + 0.00688190059736371, + 0.005240031983703375, + 0.006139961536973715, + 0.008842871524393559, + 0.010077343322336674, + 0.010322474874556065, + 0.012376564554870129, + 0.00814979150891304, + 0.012045670300722122, + 0.007786661386489868, + 0.011624675244092941, + 0.0034938363824039698, + 0.0045732157304883, + 0.003143872134387493, + 0.003649661084637046, + 0.002878829138353467, + 0.0039054874796420336, + 0.0030601569451391697, + 0.0040316395461559296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.826151430606842 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "low-angle shot, shot from above, captures this apocalyptic scene in extreme super-slow motion. As the camera slowly and steadily pulls back, the full extent of the disaster is mercilessly revealed. Amidst Chen Kaige's intensely fateful lighting, the iconic narrow and steep stone steps of Jiufen Old Street in Taiwan are relentlessly engulfed by a raging landslide. Rows of large red lanterns flank the steps, emitting a warm, ominous crimson glow in the night and drizzling rain. This light casts eerie, turbulent shadows on the wet stone slabs and the impending torrent of destruction. Like a surging beast, the viscous, dark brown mud slowly tumbles upward in a near-solidified motion, carrying broken and shattered wooden beams, splintered planks, and other unrecognizable debris. The mud collides with the solid stone steps, stirring up a massive wave of rainwater and mud, each drop of the splashing liquid clearly visible in the slow motion. Hanging red lanterns swayed violently under the impact, some instantly swept into the torrent, their fragile paper coverings ripped apart in the mud, their light extinguished. Broken wooden beams tumbled and collided helplessly in the viscous liquid, emitting a dull and despairing echo. The entire painting creates a strong color contrast between the deep red of the lanterns, the dark brown of the mudslide, and the cool gray of the stone steps, poeticizing the process of destruction and presenting a heartbreaking, grand, and tragic aesthetic of destructive tragedy.", + "seed": 25676, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_12/index_5391_seed_25676_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21176955439150333, + "vlm_logprob": -0.21176955439150333, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.21176955439150333, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23961754143238068 + ], + "opposite_logprob_tokens": [ + -1.7396175861358643 + ], + "logprob_reward": -0.23961754143238068, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.23961754143238068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31932902336120605, + 0.223094180226326, + 0.06584186851978302, + 0.03274395689368248, + 0.09683297574520111, + 0.05119727924466133, + 0.0650789886713028, + 0.07996964454650879, + 0.028506744652986526, + 0.037405356764793396 + ], + "credit_map": [ + 0.022125093266367912, + 0.024608206003904343, + 0.03918983042240143, + 0.10130233317613602, + 0.015113648027181625, + 0.015494110062718391, + 0.04561958461999893, + 0.055876217782497406, + 0.013730468228459358, + 0.01806868053972721, + 0.03628062829375267, + 0.04747666418552399, + 0.0176363717764616, + 0.016048269346356392, + 0.03638468310236931, + 0.03746842220425606, + 0.007149859331548214, + 0.006008245050907135, + 0.009834305383265018, + 0.010712319985032082, + 0.006241756957024336, + 0.008578424341976643, + 0.007408959325402975, + 0.009907994419336319, + 0.0036772938910871744, + 0.00407450133934617, + 0.0034920615144073963, + 0.007521967403590679, + 0.002874726662412286, + 0.0024282352533191442, + 0.0040474096313118935, + 0.004627760034054518, + 0.005859559401869774, + 0.009188293479382992, + 0.01292280200868845, + 0.029229626059532166, + 0.006776176393032074, + 0.007176176179200411, + 0.011003308929502964, + 0.01467702817171812, + 0.006518902722746134, + 0.0049101910553872585, + 0.005500666797161102, + 0.008376581594347954, + 0.005951039958745241, + 0.00507371174171567, + 0.008186446502804756, + 0.006679737940430641, + 0.008985571563243866, + 0.01045345887541771, + 0.006092032417654991, + 0.008391818031668663, + 0.005588456988334656, + 0.012033031322062016, + 0.005657310597598553, + 0.007877310737967491, + 0.00863212626427412, + 0.01738361082971096, + 0.004739394877105951, + 0.006026796530932188, + 0.005279231816530228, + 0.01200918946415186, + 0.005780520848929882, + 0.020118769258260727, + 0.004840706009417772, + 0.0027672951109707355, + 0.003209422342479229, + 0.0021777732763439417, + 0.0070055085234344006, + 0.004130830056965351, + 0.0022483570501208305, + 0.0021268536802381277, + 0.004548313561826944, + 0.007914944551885128, + 0.002649157540872693, + 0.002112224930897355, + 0.004974981304258108, + 0.007712285965681076, + 0.0023142355494201183, + 0.005179211031645536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.23961754143238068 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005287472158670425 + ], + "opposite_logprob_tokens": [ + -5.2552876472473145 + ], + "logprob_reward": -0.005287472158670425, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005287472158670425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17436149716377258, + 0.11413408815860748, + 0.08980514854192734, + 0.05681717023253441, + 0.062068305909633636, + 0.10873817652463913, + 0.09145457297563553, + 0.19741708040237427, + 0.0722864642739296, + 0.032917462289333344 + ], + "credit_map": [ + 0.021098071709275246, + 0.017452100291848183, + 0.02016579918563366, + 0.022075165063142776, + 0.021664690226316452, + 0.022194232791662216, + 0.022960204631090164, + 0.02675122581422329, + 0.006696827244013548, + 0.010216619819402695, + 0.0105892363935709, + 0.02382468990981579, + 0.012496194802224636, + 0.013852993957698345, + 0.02132093906402588, + 0.015136592090129852, + 0.006243675481528044, + 0.008182577788829803, + 0.012785314582288265, + 0.01198565773665905, + 0.010476170107722282, + 0.01454443670809269, + 0.010940415784716606, + 0.014646903611719608, + 0.005603428930044174, + 0.006339310202747583, + 0.00941166840493679, + 0.013152150437235832, + 0.004009069409221411, + 0.005839745048433542, + 0.005944931413978338, + 0.006516868714243174, + 0.007201097905635834, + 0.0051037827506661415, + 0.006793336942791939, + 0.014992957934737206, + 0.0065498570911586285, + 0.005532932933419943, + 0.007669426035135984, + 0.008224911987781525, + 0.017508935183286667, + 0.011409117840230465, + 0.010132110677659512, + 0.010959281586110592, + 0.010629995726048946, + 0.014682415872812271, + 0.01845356449484825, + 0.014962748624384403, + 0.007895597256720066, + 0.010973425582051277, + 0.008423474617302418, + 0.00794215127825737, + 0.019299617037177086, + 0.01543596014380455, + 0.009649917483329773, + 0.011834429576992989, + 0.018086090683937073, + 0.04422615095973015, + 0.01513340137898922, + 0.010793555527925491, + 0.011437427252531052, + 0.02956991083920002, + 0.008336113765835762, + 0.05983443930745125, + 0.008184351958334446, + 0.015501737594604492, + 0.003994564060121775, + 0.005735989660024643, + 0.011115668341517448, + 0.016467919573187828, + 0.005754705518484116, + 0.005531528033316135, + 0.0042047565802931786, + 0.0035694732796400785, + 0.002415498485788703, + 0.003245645435526967, + 0.0067304340191185474, + 0.004256769083440304, + 0.004148862790316343, + 0.004346025176346302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.005287472158670425 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029798321425914764 + ], + "opposite_logprob_tokens": [ + -3.5297982692718506 + ], + "logprob_reward": -0.029798321425914764, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029798321425914764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24714943766593933, + 0.060952942818403244, + 0.26831263303756714, + 0.028977977111935616, + 0.03241671621799469, + 0.09835606813430786, + 0.08390258997678757, + 0.07198899239301682, + 0.04687223583459854, + 0.06107042357325554 + ], + "credit_map": [ + 0.018746808171272278, + 0.0181953813880682, + 0.025200868025422096, + 0.06817162036895752, + 0.018096817657351494, + 0.018930789083242416, + 0.03958675265312195, + 0.04022042080760002, + 0.004579911474138498, + 0.00439477851614356, + 0.006315038073807955, + 0.01585126854479313, + 0.00477435952052474, + 0.004678825847804546, + 0.011585221625864506, + 0.008773545734584332, + 0.02878320775926113, + 0.02481316775083542, + 0.05894956365227699, + 0.047635551542043686, + 0.011825979687273502, + 0.01965448632836342, + 0.03176639601588249, + 0.044884271919727325, + 0.002824578434228897, + 0.00322387320920825, + 0.0024931838270276785, + 0.00426546111702919, + 0.0036820340901613235, + 0.0037642146926373243, + 0.004753147251904011, + 0.003971485421061516, + 0.00226474623195827, + 0.0029470473527908325, + 0.003657671855762601, + 0.007415704894810915, + 0.003642604686319828, + 0.004085641819983721, + 0.0039740391075611115, + 0.004429260268807411, + 0.015601705759763718, + 0.009664441458880901, + 0.012952816672623158, + 0.009324620477855206, + 0.008453589864075184, + 0.015712011605501175, + 0.013573366217315197, + 0.013073520734906197, + 0.010685708373785019, + 0.012678034603595734, + 0.007459411397576332, + 0.01063928846269846, + 0.00881397258490324, + 0.011302633211016655, + 0.01361116487532854, + 0.008712387643754482, + 0.011688397265970707, + 0.006458931136876345, + 0.010358947329223156, + 0.009714430198073387, + 0.005825455766171217, + 0.011853923089802265, + 0.007701451890170574, + 0.008387455716729164, + 0.005377772729843855, + 0.011597299948334694, + 0.002302417764440179, + 0.0034123125951737165, + 0.004422454629093409, + 0.012451818212866783, + 0.002841284964233637, + 0.004466880112886429, + 0.009355152025818825, + 0.01449032872915268, + 0.0030082205776125193, + 0.0037754427175968885, + 0.005802498199045658, + 0.014389825984835625, + 0.004365677945315838, + 0.005883284378796816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.029798321425914764 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355599984526634 + ], + "opposite_logprob_tokens": [ + -3.651355504989624 + ], + "logprob_reward": -0.026355599984526634, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355599984526634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09343073517084122, + 0.05428473278880119, + 0.08290411531925201, + 0.10154814273118973, + 0.0320294015109539, + 0.06955563277006149, + 0.05610033497214317, + 0.4437425136566162, + 0.026088938117027283, + 0.040315501391887665 + ], + "credit_map": [ + 0.011381479911506176, + 0.011445358395576477, + 0.01087892148643732, + 0.01168782263994217, + 0.009526759386062622, + 0.012149104848504066, + 0.014363909140229225, + 0.011997376568615437, + 0.003531390568241477, + 0.004246140830218792, + 0.008133405819535255, + 0.010249728336930275, + 0.005510700400918722, + 0.005369484890252352, + 0.00893590971827507, + 0.00830796454101801, + 0.006502954289317131, + 0.010193048976361752, + 0.014404774643480778, + 0.015247347764670849, + 0.006682287435978651, + 0.00926287192851305, + 0.007405471988022327, + 0.013205356895923615, + 0.0067598894238471985, + 0.00866377167403698, + 0.017252543941140175, + 0.0251743383705616, + 0.0055337389931082726, + 0.006080717779695988, + 0.01867123879492283, + 0.013411893509328365, + 0.003445372683927417, + 0.002568769035860896, + 0.0032354420982301235, + 0.0048057944513857365, + 0.0047966064885258675, + 0.003873036243021488, + 0.005184792913496494, + 0.00411958759650588, + 0.01051944401115179, + 0.008327197283506393, + 0.005149474833160639, + 0.0075264377519488335, + 0.00633936608210206, + 0.007962160743772984, + 0.013189762830734253, + 0.01054177712649107, + 0.009760895743966103, + 0.007867859676480293, + 0.004880537744611502, + 0.006606457754969597, + 0.005574848968535662, + 0.007423641160130501, + 0.008476195856928825, + 0.0055098929442465305, + 0.055808525532484055, + 0.09162154793739319, + 0.03249775618314743, + 0.03383844718337059, + 0.024266745895147324, + 0.05774225294589996, + 0.03292784094810486, + 0.1150393933057785, + 0.0033145712222903967, + 0.005686726421117783, + 0.0012480099685490131, + 0.0014881676761433482, + 0.004317064769566059, + 0.00637702364474535, + 0.0016938447952270508, + 0.0019635267090052366, + 0.0051432582549750805, + 0.01017041690647602, + 0.0033359196968376637, + 0.0028866766951978207, + 0.003931061830371618, + 0.009443402290344238, + 0.002356567420065403, + 0.003048194805160165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.026355599984526634 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7577888369560242 + ], + "opposite_logprob_tokens": [ + -0.6327888369560242 + ], + "logprob_reward": -0.7577888369560242, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7577888369560242, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09912169724702835, + 0.07163096219301224, + 0.0633484497666359, + 0.07998985052108765, + 0.0325956754386425, + 0.16665591299533844, + 0.2622242867946625, + 0.05253547430038452, + 0.10877502709627151, + 0.0631227046251297 + ], + "credit_map": [ + 0.01094326563179493, + 0.01194413099437952, + 0.009808962233364582, + 0.01852307841181755, + 0.010451233945786953, + 0.009712272323668003, + 0.013303745537996292, + 0.014435007236897945, + 0.004450422711670399, + 0.005450088996440172, + 0.008334444835782051, + 0.01839461922645569, + 0.005411731079220772, + 0.005864276085048914, + 0.012412694282829762, + 0.011312680318951607, + 0.007784598041325808, + 0.00898288656026125, + 0.008911741897463799, + 0.009986364282667637, + 0.005278582219034433, + 0.006449857261031866, + 0.006488013081252575, + 0.009466406889259815, + 0.005425787065178156, + 0.00780372554436326, + 0.012976556085050106, + 0.02174430899322033, + 0.00405564159154892, + 0.004135465249419212, + 0.01319001242518425, + 0.010658354498445988, + 0.005043810699135065, + 0.0037209060974419117, + 0.003675315761938691, + 0.004269437398761511, + 0.0039647300727665424, + 0.0041353218257427216, + 0.004199120216071606, + 0.0035870359279215336, + 0.02944503352046013, + 0.016125205904245377, + 0.021040907129645348, + 0.01864137314260006, + 0.019455667585134506, + 0.018894024193286896, + 0.02439121901988983, + 0.01866246946156025, + 0.030702147632837296, + 0.026591859757900238, + 0.03024749830365181, + 0.03406040742993355, + 0.029055597260594368, + 0.044751524925231934, + 0.03713609278202057, + 0.02967914380133152, + 0.005712889600545168, + 0.012117384001612663, + 0.004318018909543753, + 0.003195950761437416, + 0.003378516063094139, + 0.007125129457563162, + 0.003237814176827669, + 0.013449774123728275, + 0.013284305110573769, + 0.03152981773018837, + 0.0037318216636776924, + 0.005467399954795837, + 0.009806414134800434, + 0.03302516043186188, + 0.004595846403390169, + 0.007334255613386631, + 0.008199615404009819, + 0.019071368500590324, + 0.001912743435241282, + 0.002989239292219281, + 0.00584901450201869, + 0.01812196709215641, + 0.002466549864038825, + 0.004512205254286528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.7577888369560242 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "low-angle shot, shot from above, captures this apocalyptic scene in extreme super-slow motion. As the camera slowly and steadily pulls back, the full extent of the disaster is mercilessly revealed. Amidst Chen Kaige's intensely fateful lighting, the iconic narrow and steep stone steps of Jiufen Old Street in Taiwan are relentlessly engulfed by a raging landslide. Rows of large red lanterns flank the steps, emitting a warm, ominous crimson glow in the night and drizzling rain. This light casts eerie, turbulent shadows on the wet stone slabs and the impending torrent of destruction. Like a surging beast, the viscous, dark brown mud slowly tumbles upward in a near-solidified motion, carrying broken and shattered wooden beams, splintered planks, and other unrecognizable debris. The mud collides with the solid stone steps, stirring up a massive wave of rainwater and mud, each drop of the splashing liquid clearly visible in the slow motion. Hanging red lanterns swayed violently under the impact, some instantly swept into the torrent, their fragile paper coverings ripped apart in the mud, their light extinguished. Broken wooden beams tumbled and collided helplessly in the viscous liquid, emitting a dull and despairing echo. The entire painting creates a strong color contrast between the deep red of the lanterns, the dark brown of the mudslide, and the cool gray of the stone steps, poeticizing the process of destruction and presenting a heartbreaking, grand, and tragic aesthetic of destructive tragedy.", + "seed": 25677, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_12/index_5391_seed_25677_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2568644591607153, + "vlm_logprob": -0.2568644591607153, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.2568644591607153, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4052778482437134 + ], + "opposite_logprob_tokens": [ + -1.2802778482437134 + ], + "logprob_reward": -0.4052778482437134, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.4052778482437134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34893524646759033, + 0.10929687321186066, + 0.05099903419613838, + 0.07348612695932388, + 0.0582987517118454, + 0.03136344999074936, + 0.061426952481269836, + 0.03081236593425274, + 0.1624269187450409, + 0.07295424491167068 + ], + "credit_map": [ + 0.03297856077551842, + 0.02672998420894146, + 0.05329543724656105, + 0.07517760992050171, + 0.02189333550632, + 0.02678150124847889, + 0.0700978934764862, + 0.04198094457387924, + 0.006575513165444136, + 0.0071397642605006695, + 0.012582465074956417, + 0.016490686684846878, + 0.005484531633555889, + 0.007444560993462801, + 0.02746802382171154, + 0.026111328974366188, + 0.005539606790989637, + 0.006824412848800421, + 0.008457284420728683, + 0.007926425896584988, + 0.003642711555585265, + 0.0041372040286660194, + 0.008396695367991924, + 0.006074691656976938, + 0.009721074253320694, + 0.012801785953342915, + 0.006820624694228172, + 0.008881153538823128, + 0.005538401193916798, + 0.007060025818645954, + 0.014525522477924824, + 0.008137534372508526, + 0.0075255208648741245, + 0.005967918783426285, + 0.0053728981874883175, + 0.007840932346880436, + 0.006049704272300005, + 0.006215090863406658, + 0.011896340176463127, + 0.007430349942296743, + 0.0032927854917943478, + 0.004508422687649727, + 0.0028639077208936214, + 0.0054036881774663925, + 0.002405379433184862, + 0.003286955179646611, + 0.005013762041926384, + 0.004588548559695482, + 0.007961719296872616, + 0.00916915014386177, + 0.004003364127129316, + 0.011000067926943302, + 0.005819843616336584, + 0.006222863681614399, + 0.0086439847946167, + 0.008605964481830597, + 0.002596441423520446, + 0.0025620937813073397, + 0.005066468846052885, + 0.005082370713353157, + 0.0024381382390856743, + 0.002465789671987295, + 0.005948823876678944, + 0.00465224077925086, + 0.028142794966697693, + 0.01992453821003437, + 0.0209214985370636, + 0.017446300014853477, + 0.020465245470404625, + 0.012052973732352257, + 0.022137749940156937, + 0.021335821598768234, + 0.00895507913082838, + 0.006041598506271839, + 0.011362073943018913, + 0.012738075107336044, + 0.003906694706529379, + 0.004861796274781227, + 0.014001637697219849, + 0.01108729001134634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.4052778482437134 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003694139886647463 + ], + "opposite_logprob_tokens": [ + -5.6286940574646 + ], + "logprob_reward": -0.003694139886647463, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003694139886647463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18240922689437866, + 0.17702648043632507, + 0.06203194335103035, + 0.0936107486486435, + 0.054701656103134155, + 0.04384739696979523, + 0.04077252373099327, + 0.09680967777967453, + 0.15491273999214172, + 0.09387762099504471 + ], + "credit_map": [ + 0.013775890693068504, + 0.015632936730980873, + 0.02542976848781109, + 0.02874026633799076, + 0.01912713050842285, + 0.02230944111943245, + 0.03493080288171768, + 0.022462984547019005, + 0.01119860727339983, + 0.007609689608216286, + 0.06175808981060982, + 0.016229718923568726, + 0.0053398748859763145, + 0.014313734136521816, + 0.022574657574295998, + 0.03800210729241371, + 0.006945696193724871, + 0.009330417960882187, + 0.007009525783360004, + 0.00900030042976141, + 0.005152649711817503, + 0.006513237953186035, + 0.010807561688125134, + 0.007272552233189344, + 0.014598919078707695, + 0.018939509987831116, + 0.006061861291527748, + 0.010018792003393173, + 0.00792027823626995, + 0.009460839442908764, + 0.015461216680705547, + 0.011149333789944649, + 0.006519580725580454, + 0.0057317581959068775, + 0.0054762717336416245, + 0.007500806823372841, + 0.006389711983501911, + 0.008125022053718567, + 0.007169574033468962, + 0.007788931019604206, + 0.002843289403244853, + 0.0037297713570296764, + 0.005121499765664339, + 0.007541256491094828, + 0.0035734004341065884, + 0.007236996665596962, + 0.006770573556423187, + 0.007030607666820288, + 0.005305533297359943, + 0.006534680724143982, + 0.0033779831137508154, + 0.006370798218995333, + 0.004050400108098984, + 0.003323465585708618, + 0.006487395614385605, + 0.00532226637005806, + 0.014148732647299767, + 0.00824282318353653, + 0.01935383677482605, + 0.018043439835309982, + 0.005489356815814972, + 0.009413466788828373, + 0.00949963741004467, + 0.012618389911949635, + 0.02817799337208271, + 0.0188064593821764, + 0.021766861900687218, + 0.015479745343327522, + 0.011636721901595592, + 0.013526267372071743, + 0.027447866275906563, + 0.018070828169584274, + 0.008901434950530529, + 0.0069713587872684, + 0.013870018534362316, + 0.014795725233852863, + 0.0066214934922754765, + 0.008024061098694801, + 0.019415246322751045, + 0.015278284437954426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.003694139886647463 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054946284741163254 + ], + "opposite_logprob_tokens": [ + -2.929946184158325 + ], + "logprob_reward": -0.054946284741163254, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054946284741163254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21244221925735474, + 0.0921715721487999, + 0.038327235728502274, + 0.18407681584358215, + 0.11176347732543945, + 0.047788821160793304, + 0.040818244218826294, + 0.023383205756545067, + 0.10669609904289246, + 0.14253228902816772 + ], + "credit_map": [ + 0.02350504696369171, + 0.015552857890725136, + 0.02533440664410591, + 0.035542551428079605, + 0.01803058385848999, + 0.021944869309663773, + 0.04036037251353264, + 0.03217153251171112, + 0.00794055126607418, + 0.007783924229443073, + 0.01277819275856018, + 0.01621709205210209, + 0.007446484174579382, + 0.010137434117496014, + 0.016996124759316444, + 0.012871766462922096, + 0.005312151275575161, + 0.003932026214897633, + 0.005857339594513178, + 0.005204242188483477, + 0.0029418249614536762, + 0.0044592078775167465, + 0.005845784675329924, + 0.004774658940732479, + 0.02943618781864643, + 0.03992848098278046, + 0.01650514267385006, + 0.020090365782380104, + 0.015194050967693329, + 0.014728118665516376, + 0.02645016834139824, + 0.02174428664147854, + 0.013394874520599842, + 0.010105549357831478, + 0.012967049144208431, + 0.02379356324672699, + 0.005961486604064703, + 0.009684713557362556, + 0.02065170742571354, + 0.015204532071948051, + 0.0037887857761234045, + 0.004660810809582472, + 0.005483571439981461, + 0.010208134539425373, + 0.0031940967310220003, + 0.005100024398416281, + 0.0078637870028615, + 0.007489606738090515, + 0.005238233599811792, + 0.0052869017235934734, + 0.0032296113204210997, + 0.005865021608769894, + 0.003792359959334135, + 0.003868900705128908, + 0.007838591933250427, + 0.005698623601347208, + 0.0026296363212168217, + 0.001901401556096971, + 0.00328334653750062, + 0.004260012414306402, + 0.0013517707120627165, + 0.002140090335160494, + 0.004979765508323908, + 0.0028371820226311684, + 0.012414615601301193, + 0.01623436249792576, + 0.013803823851048946, + 0.009211136028170586, + 0.008960334584116936, + 0.010159080848097801, + 0.021502520889043808, + 0.0144102293998003, + 0.013194577768445015, + 0.015575972385704517, + 0.022844580933451653, + 0.030381998047232628, + 0.011349046602845192, + 0.011244681663811207, + 0.01772329770028591, + 0.0202181413769722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.054946284741163254 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705203890800476 + ], + "opposite_logprob_tokens": [ + -2.127052068710327 + ], + "logprob_reward": -0.12705203890800476, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705203890800476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2604462206363678, + 0.21741676330566406, + 0.0772661492228508, + 0.08425377309322357, + 0.05935512110590935, + 0.04811330884695053, + 0.036144621670246124, + 0.04846005141735077, + 0.10481397807598114, + 0.06373005360364914 + ], + "credit_map": [ + 0.027511633932590485, + 0.018771883100271225, + 0.041787970811128616, + 0.05014898627996445, + 0.016603605821728706, + 0.019971614703536034, + 0.05388196185231209, + 0.031768545508384705, + 0.016353141516447067, + 0.022633859887719154, + 0.05060818791389465, + 0.03490765765309334, + 0.007824631407856941, + 0.011516228318214417, + 0.05547500401735306, + 0.018098052591085434, + 0.008802657015621662, + 0.006990739610046148, + 0.013089200481772423, + 0.00959156546741724, + 0.004604741930961609, + 0.0066346218809485435, + 0.012973537668585777, + 0.014579083770513535, + 0.012522286735475063, + 0.01216122042387724, + 0.010124525986611843, + 0.009410452097654343, + 0.00697821332141757, + 0.007084078621119261, + 0.0156732015311718, + 0.010299792513251305, + 0.007458403240889311, + 0.007923348806798458, + 0.006750775966793299, + 0.007882618345320225, + 0.004769541323184967, + 0.006025484297424555, + 0.011469766497612, + 0.007075183559209108, + 0.006085205357521772, + 0.004704146645963192, + 0.00488518038764596, + 0.006895176600664854, + 0.003889708314090967, + 0.007349133025854826, + 0.007686729077249765, + 0.006618031766265631, + 0.003984859213232994, + 0.00572872394695878, + 0.0027517664711922407, + 0.005438837222754955, + 0.0032666532788425684, + 0.0037147400435060263, + 0.006320110056549311, + 0.004938930738717318, + 0.004856346640735865, + 0.004243927076458931, + 0.009705299511551857, + 0.010509141720831394, + 0.0027188316453248262, + 0.0034336925018578768, + 0.005737997125834227, + 0.007254815194755793, + 0.013140767812728882, + 0.011383739300072193, + 0.018451811745762825, + 0.011461983434855938, + 0.007593146525323391, + 0.007523708511143923, + 0.022565528750419617, + 0.012693287804722786, + 0.004669648595154285, + 0.005265815183520317, + 0.01018897257745266, + 0.014239535667002201, + 0.004317822400480509, + 0.0036736647598445415, + 0.010987116023898125, + 0.010387479327619076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.12705203890800476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at the narrow and steep stone steps of Jiufen Old Street in Taiwan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are rows of large red lanterns visible flanking the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown mudslide engulfing the steps and carrying debris like broken wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme super-slow motion with clearly visible raindrops and splashing liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle shot from above and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933519840240479 + ], + "opposite_logprob_tokens": [ + -0.6933519840240479 + ], + "logprob_reward": -0.6933519840240479, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933519840240479, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15005789697170258, + 0.2079354077577591, + 0.09670797735452652, + 0.11010821908712387, + 0.06271727383136749, + 0.04046740010380745, + 0.03184601292014122, + 0.042951587587594986, + 0.21558386087417603, + 0.041624393314123154 + ], + "credit_map": [ + 0.012492180801928043, + 0.013914519920945168, + 0.01751691661775112, + 0.027926521375775337, + 0.014474316500127316, + 0.01884005032479763, + 0.026306070387363434, + 0.01858733408153057, + 0.013852870091795921, + 0.01619364693760872, + 0.06498308479785919, + 0.02367098443210125, + 0.009542403742671013, + 0.012695333920419216, + 0.05246500298380852, + 0.014532073400914669, + 0.009357521310448647, + 0.01338126789778471, + 0.011444547213613987, + 0.014028746634721756, + 0.009241015650331974, + 0.011566885747015476, + 0.015141273848712444, + 0.012546725571155548, + 0.01234758086502552, + 0.016741955652832985, + 0.011799358762800694, + 0.016304565593600273, + 0.00822666846215725, + 0.009531955234706402, + 0.023223550990223885, + 0.011932573281228542, + 0.007100759074091911, + 0.007942231371998787, + 0.005923619959503412, + 0.009868868626654148, + 0.0027965873014181852, + 0.004605548456311226, + 0.01570988819003105, + 0.008769776672124863, + 0.004091298673301935, + 0.005614949855953455, + 0.003876363392919302, + 0.0077146487310528755, + 0.002396122319623828, + 0.003326766425743699, + 0.007230834569782019, + 0.006216416601091623, + 0.004275324288755655, + 0.00407449621707201, + 0.004028412513434887, + 0.003883371828123927, + 0.0027057030238211155, + 0.0042462535202503204, + 0.004862493835389614, + 0.003769957460463047, + 0.00351286050863564, + 0.0032940004020929337, + 0.006171744782477617, + 0.006788133177906275, + 0.002676498843356967, + 0.0029364454094320536, + 0.011448515579104424, + 0.00612338911741972, + 0.03610280156135559, + 0.01580437645316124, + 0.03036876954138279, + 0.026269041001796722, + 0.018909791484475136, + 0.023604385554790497, + 0.036128219217061996, + 0.028396472334861755, + 0.003929109312593937, + 0.003787829540669918, + 0.007011631038039923, + 0.007681349292397499, + 0.002985206665471196, + 0.0029360491316765547, + 0.008310388773679733, + 0.004982828628271818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.6933519840240479 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_13/index_tensor([8628, 8628])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_13/index_tensor([8628, 8628])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4a2e2cdc602fa81eb4bf750f7c5389fe0bb8fd82 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_13/index_tensor([8628, 8628])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, wide shot, rendered with the delicate brushstrokes of a child's sketchbook, captures a panoramic view of the vast expanse of the sea cliffs. The world is composed of soft blocks of color and evocative lines. The sky is a watercolor-like blend of soft pink and pale blue, while the ocean is a deep, tranquil azure, punctuated by simple white brushstrokes that outline the waves. At the center of this vast canvas, a tiny figure, Matsuzaki Umi, a girl in a sailor uniform, stands alone. She wears a classic white sailor top, a navy scarf, and a matching pleated skirt. A strong sea breeze blows fiercely from the front, sweeping her chin-length dark brown hair completely back, revealing her resolute and serene face. The wind also pushes the hem of her skirt tightly behind her, causing it to dance violently and silently in the air. Her feet are firmly rooted on the grassy clifftop, her body slightly leaning forward to fight the wind. Her tiny figure displays an unyielding determination in the vast world, creating a tranquil atmosphere that is both lonely and powerful.", + "seed": 27216, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_13/index_8628_seed_27216_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12520261337049304, + "vlm_logprob": -0.12520261337049304, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.12520261337049304, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05164807662367821 + ], + "opposite_logprob_tokens": [ + -3.1766481399536133 + ], + "logprob_reward": -0.05164807662367821, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05164807662367821, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4493516683578491, + 0.0847790464758873, + 0.046762287616729736, + 0.06383432447910309, + 0.02678416110575199, + 0.060995958745479584, + 0.08185184746980667, + 0.08798191696405411, + 0.07082945853471756, + 0.026829324662685394 + ], + "credit_map": [ + 0.04643041640520096, + 0.06995139271020889, + 0.05664810538291931, + 0.048146504908800125, + 0.02802811935544014, + 0.08188775926828384, + 0.03462165221571922, + 0.08363769948482513, + 0.00654676603153348, + 0.014410467818379402, + 0.010655221529304981, + 0.0064807492308318615, + 0.006233756896108389, + 0.02235618233680725, + 0.007735657971352339, + 0.010360244661569595, + 0.005814074072986841, + 0.006394262425601482, + 0.005930609535425901, + 0.003955237567424774, + 0.0038690397050231695, + 0.009912833571434021, + 0.005145126488059759, + 0.005741104017943144, + 0.007813842035830021, + 0.006224460434168577, + 0.012059364467859268, + 0.006014104001224041, + 0.005753424484282732, + 0.009527438320219517, + 0.00648257602006197, + 0.009959115646779537, + 0.002950563794001937, + 0.004193425644189119, + 0.0034198048524558544, + 0.002350659342482686, + 0.003306414233520627, + 0.004560367204248905, + 0.002974136732518673, + 0.0030287904664874077, + 0.0067480504512786865, + 0.006747594568878412, + 0.010625192895531654, + 0.006294149439781904, + 0.0069769821129739285, + 0.008899458684027195, + 0.008129441179335117, + 0.006575093604624271, + 0.009739402681589127, + 0.013310842216014862, + 0.009991824626922607, + 0.00867142342031002, + 0.006565004587173462, + 0.011782322078943253, + 0.012533518485724926, + 0.009257507510483265, + 0.010363461449742317, + 0.011429574340581894, + 0.010968750342726707, + 0.007619047537446022, + 0.014097273349761963, + 0.011971510015428066, + 0.01130005531013012, + 0.010232247412204742, + 0.0076509080827236176, + 0.010140862315893173, + 0.01239874679595232, + 0.005263787228614092, + 0.005169234704226255, + 0.011890373192727566, + 0.006245918571949005, + 0.01206962764263153, + 0.0026809710543602705, + 0.003268770407885313, + 0.003375849686563015, + 0.00287772249430418, + 0.0031158458441495895, + 0.0038619290571659803, + 0.003619830822572112, + 0.004028405994176865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.05164807662367821 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028502813074737787 + ], + "opposite_logprob_tokens": [ + -5.87785005569458 + ], + "logprob_reward": -0.0028502813074737787, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028502813074737787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5791816115379333, + 0.035721052438020706, + 0.0367024801671505, + 0.02600518986582756, + 0.059100206941366196, + 0.042417578399181366, + 0.05070810765028, + 0.07484880834817886, + 0.07214119285345078, + 0.02317381463944912 + ], + "credit_map": [ + 0.05656377226114273, + 0.2267695516347885, + 0.020618155598640442, + 0.015967682003974915, + 0.024036645889282227, + 0.2019074708223343, + 0.014550545252859592, + 0.01876775175333023, + 0.0031531937420368195, + 0.005130929406732321, + 0.0031318357214331627, + 0.003334958804771304, + 0.003943984862416983, + 0.008794319815933704, + 0.003818910801783204, + 0.004412917420268059, + 0.0037917799782007933, + 0.005465358961373568, + 0.003905715886503458, + 0.0029138708487153053, + 0.0032754861749708652, + 0.010012189857661724, + 0.0026926654390990734, + 0.0046454137191176414, + 0.0035506030544638634, + 0.0032454458996653557, + 0.0024319137446582317, + 0.0020488195586949587, + 0.0024511185474693775, + 0.006038805935531855, + 0.002813884988427162, + 0.0034245981369167566, + 0.00888105109333992, + 0.007032281719148159, + 0.009106365032494068, + 0.005530800204724073, + 0.0057196649722754955, + 0.009029343724250793, + 0.0056371018290519714, + 0.00816359929740429, + 0.00478089414536953, + 0.004333838354796171, + 0.006496635731309652, + 0.005670428741723299, + 0.004181693308055401, + 0.00691525312140584, + 0.005778344348073006, + 0.004260488785803318, + 0.0061278329230844975, + 0.008088367059826851, + 0.00537286140024662, + 0.004817281384021044, + 0.004944206681102514, + 0.007605463732033968, + 0.008108451962471008, + 0.005643636919558048, + 0.008444471284747124, + 0.010492385365068913, + 0.008178130723536015, + 0.006341420579701662, + 0.011594033800065517, + 0.010237094014883041, + 0.009839054197072983, + 0.009722216986119747, + 0.00761869503185153, + 0.009186244569718838, + 0.014592887833714485, + 0.006590371951460838, + 0.004824365954846144, + 0.011971491388976574, + 0.005744516849517822, + 0.011612620204687119, + 0.00325112696737051, + 0.0034393894020467997, + 0.002224917756393552, + 0.0017140447162091732, + 0.0022707453463226557, + 0.004349430091679096, + 0.0029550797771662474, + 0.002969079650938511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.0028502813074737787 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952132865786552 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952132865786552, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952132865786552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18568822741508484, + 0.12212613224983215, + 0.08564775437116623, + 0.06128410995006561, + 0.062343139201402664, + 0.13378950953483582, + 0.08760175108909607, + 0.16756640374660492, + 0.06759520620107651, + 0.02635776810348034 + ], + "credit_map": [ + 0.027090709656476974, + 0.029670512303709984, + 0.019367486238479614, + 0.012722868472337723, + 0.018990689888596535, + 0.032499413937330246, + 0.018465204164385796, + 0.026881348341703415, + 0.011219773441553116, + 0.021846670657396317, + 0.012911600060760975, + 0.009652677923440933, + 0.013422482647001743, + 0.02204344980418682, + 0.018506629392504692, + 0.012522852048277855, + 0.008776499889791012, + 0.012194975279271603, + 0.008943025954067707, + 0.006439665332436562, + 0.008700878359377384, + 0.01927863247692585, + 0.011666839011013508, + 0.00964723713696003, + 0.0058784643188118935, + 0.0072670672088861465, + 0.006270422600209713, + 0.005878095980733633, + 0.008728991262614727, + 0.010447888635098934, + 0.011482656933367252, + 0.005330527666956186, + 0.005976375658065081, + 0.008072919212281704, + 0.006845925934612751, + 0.005767705384641886, + 0.009908627718687057, + 0.00734124006703496, + 0.011758857406675816, + 0.006671490613371134, + 0.01565403677523136, + 0.015314457006752491, + 0.01492864079773426, + 0.014877294190227985, + 0.016055438667535782, + 0.02081098034977913, + 0.023077847436070442, + 0.01307082548737526, + 0.009165805764496326, + 0.01393203530460596, + 0.0079689035192132, + 0.009546790271997452, + 0.009784702211618423, + 0.011891631409525871, + 0.017391661182045937, + 0.00792021956294775, + 0.02020804025232792, + 0.02591346949338913, + 0.022649753838777542, + 0.014380351640284061, + 0.020957883447408676, + 0.02196609601378441, + 0.01964740827679634, + 0.021843405440449715, + 0.0071326326578855515, + 0.00906395073980093, + 0.015920288860797882, + 0.006588741205632687, + 0.004555179737508297, + 0.00813967827707529, + 0.005018145311623812, + 0.011176593601703644, + 0.0024901903234422207, + 0.003271830501034856, + 0.0034205755218863487, + 0.002136433031409979, + 0.0029832064174115658, + 0.003499724669381976, + 0.004833510611206293, + 0.003722296329215169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.005952132865786552 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134441375732422 + ], + "opposite_logprob_tokens": [ + -1.3134441375732422 + ], + "logprob_reward": -0.3134441375732422, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134441375732422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3392317295074463, + 0.13373546302318573, + 0.09015669673681259, + 0.0663113072514534, + 0.06372413039207458, + 0.07945488393306732, + 0.04359456151723862, + 0.058590710163116455, + 0.06934741884469986, + 0.05585308000445366 + ], + "credit_map": [ + 0.05450768768787384, + 0.08700710535049438, + 0.030510710552334785, + 0.019199393689632416, + 0.031103599816560745, + 0.06817544996738434, + 0.02436891943216324, + 0.024358879774808884, + 0.013007358647882938, + 0.01831487938761711, + 0.018505588173866272, + 0.011695836670696735, + 0.015845315530896187, + 0.013242869637906551, + 0.03080711141228676, + 0.012316511012613773, + 0.008944583125412464, + 0.011638402007520199, + 0.01114097610116005, + 0.006808304227888584, + 0.006482705008238554, + 0.020170383155345917, + 0.010486568324267864, + 0.014484785497188568, + 0.007487058639526367, + 0.009761077351868153, + 0.006781474687159061, + 0.0047118873335421085, + 0.006454447750002146, + 0.015017321333289146, + 0.007873550057411194, + 0.008224491961300373, + 0.007345056626945734, + 0.009218813851475716, + 0.005513870622962713, + 0.005368000827729702, + 0.009794389829039574, + 0.008996487595140934, + 0.008817669004201889, + 0.008669850416481495, + 0.009547365829348564, + 0.00958827044814825, + 0.007700488902628422, + 0.008926176466047764, + 0.010628344491124153, + 0.010387604124844074, + 0.015698464587330818, + 0.006978173740208149, + 0.004591149743646383, + 0.005619409028440714, + 0.005047075916081667, + 0.004152155946940184, + 0.004799183923751116, + 0.005573619157075882, + 0.008814457803964615, + 0.004997511859983206, + 0.007189810276031494, + 0.007470753509551287, + 0.006201495882123709, + 0.005107278935611248, + 0.010733037255704403, + 0.007809525355696678, + 0.008314065635204315, + 0.005764746572822332, + 0.005701114889234304, + 0.008690711110830307, + 0.011155514977872372, + 0.007269388530403376, + 0.006769766099750996, + 0.008958132937550545, + 0.011986352503299713, + 0.00881644431501627, + 0.0067936815321445465, + 0.009201819077134132, + 0.005018866620957851, + 0.0053449575789272785, + 0.006974800489842892, + 0.008898412808775902, + 0.007522675208747387, + 0.006097874604165554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.3134441375732422 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25211843848228455 + ], + "opposite_logprob_tokens": [ + -1.502118468284607 + ], + "logprob_reward": -0.25211843848228455, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25211843848228455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2327633947134018, + 0.0919199287891388, + 0.03532031923532486, + 0.051320310682058334, + 0.12424167990684509, + 0.055562086403369904, + 0.06897594779729843, + 0.10225555300712585, + 0.1779291033744812, + 0.059711653739213943 + ], + "credit_map": [ + 0.027712371200323105, + 0.05775221437215805, + 0.016050854697823524, + 0.012510801665484905, + 0.023354724049568176, + 0.0646255761384964, + 0.01339418813586235, + 0.017362697049975395, + 0.007393168285489082, + 0.01230239775031805, + 0.010742864571511745, + 0.007425312884151936, + 0.008605483919382095, + 0.019760843366384506, + 0.014841745607554913, + 0.010848118923604488, + 0.003635479835793376, + 0.0046912613324820995, + 0.003939502872526646, + 0.0036526594776660204, + 0.004516772925853729, + 0.005092842038720846, + 0.005054658744484186, + 0.004737141076475382, + 0.005439646076411009, + 0.008322041481733322, + 0.006070389878004789, + 0.0040411208756268024, + 0.005386581178754568, + 0.00878837425261736, + 0.007102682255208492, + 0.006169477943331003, + 0.014916114509105682, + 0.018015624955296516, + 0.011011849157512188, + 0.010577992536127567, + 0.016440274193882942, + 0.019764108583331108, + 0.02133314125239849, + 0.012182577513158321, + 0.006026706658303738, + 0.008054263889789581, + 0.008183781988918781, + 0.0048334430903196335, + 0.007649192586541176, + 0.007728207856416702, + 0.008379487320780754, + 0.004707008600234985, + 0.006254907231777906, + 0.009030571207404137, + 0.006204037927091122, + 0.008101632818579674, + 0.007699914276599884, + 0.008764115162193775, + 0.015408976003527641, + 0.0075117903761565685, + 0.012850544415414333, + 0.018119413405656815, + 0.01292133703827858, + 0.008659178391098976, + 0.010398139245808125, + 0.013989506289362907, + 0.01138980221003294, + 0.013927638530731201, + 0.017973994836211205, + 0.025111684575676918, + 0.03251773491501808, + 0.01512713823467493, + 0.013203506357967854, + 0.02635389380156994, + 0.01967589184641838, + 0.027965271845459938, + 0.005392426624894142, + 0.008850859478116035, + 0.005684512667357922, + 0.005382009316235781, + 0.007519252598285675, + 0.010302018374204636, + 0.008081167936325073, + 0.008499410934746265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.25211843848228455 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, wide shot, rendered with the delicate brushstrokes of a child's sketchbook, captures a panoramic view of the vast expanse of the sea cliffs. The world is composed of soft blocks of color and evocative lines. The sky is a watercolor-like blend of soft pink and pale blue, while the ocean is a deep, tranquil azure, punctuated by simple white brushstrokes that outline the waves. At the center of this vast canvas, a tiny figure, Matsuzaki Umi, a girl in a sailor uniform, stands alone. She wears a classic white sailor top, a navy scarf, and a matching pleated skirt. A strong sea breeze blows fiercely from the front, sweeping her chin-length dark brown hair completely back, revealing her resolute and serene face. The wind also pushes the hem of her skirt tightly behind her, causing it to dance violently and silently in the air. Her feet are firmly rooted on the grassy clifftop, her body slightly leaning forward to fight the wind. Her tiny figure displays an unyielding determination in the vast world, creating a tranquil atmosphere that is both lonely and powerful.", + "seed": 27217, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_13/index_8628_seed_27217_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2276795492041856, + "vlm_logprob": -0.2276795492041856, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.2276795492041856, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06887367367744446 + ], + "opposite_logprob_tokens": [ + -2.943873643875122 + ], + "logprob_reward": -0.06887367367744446, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06887367367744446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21376201510429382, + 0.09046276658773422, + 0.2249409407377243, + 0.05390557646751404, + 0.03234950453042984, + 0.04722011461853981, + 0.07816839963197708, + 0.08845105767250061, + 0.10232790559530258, + 0.06841176748275757 + ], + "credit_map": [ + 0.036121346056461334, + 0.031044894829392433, + 0.015536685474216938, + 0.01787126623094082, + 0.026090744882822037, + 0.03271004185080528, + 0.02280074544250965, + 0.031586289405822754, + 0.012151122093200684, + 0.010327614843845367, + 0.0074390992522239685, + 0.008998433127999306, + 0.01145298033952713, + 0.011235857382416725, + 0.016302645206451416, + 0.012555021792650223, + 0.028929676860570908, + 0.032283514738082886, + 0.025447135791182518, + 0.024004627019166946, + 0.03178870677947998, + 0.029222173616290092, + 0.02811012789607048, + 0.025154968723654747, + 0.008539021015167236, + 0.008139476180076599, + 0.00789167545735836, + 0.006920474115759134, + 0.005063435062766075, + 0.006376343779265881, + 0.005866668187081814, + 0.005108484998345375, + 0.005138126201927662, + 0.005771849770098925, + 0.003013196401298046, + 0.002959924517199397, + 0.005087286699563265, + 0.004033782985061407, + 0.004011308308690786, + 0.002334029646590352, + 0.005842807702720165, + 0.007168146315962076, + 0.005035642068833113, + 0.00558368768543005, + 0.005841132719069719, + 0.005860661156475544, + 0.005952253472059965, + 0.0059357816353440285, + 0.009633711539208889, + 0.009118123911321163, + 0.0068329935893416405, + 0.011876310221850872, + 0.012189379893243313, + 0.010446790605783463, + 0.008762434124946594, + 0.00930866040289402, + 0.01680012233555317, + 0.015050645917654037, + 0.00718146376311779, + 0.006767296232283115, + 0.013094507157802582, + 0.01081858854740858, + 0.01026108954101801, + 0.008477341383695602, + 0.014445842243731022, + 0.012424198910593987, + 0.008496370166540146, + 0.010336382314562798, + 0.01448091585189104, + 0.013555220328271389, + 0.01762816496193409, + 0.010960808955132961, + 0.007988253608345985, + 0.01179179735481739, + 0.005353216547518969, + 0.0058645340614020824, + 0.009256245568394661, + 0.00822943914681673, + 0.009819220751523972, + 0.010109057649970055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.06887367367744446 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025239300448447466 + ], + "opposite_logprob_tokens": [ + -6.002523899078369 + ], + "logprob_reward": -0.0025239300448447466, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025239300448447466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1908080130815506, + 0.09583326429128647, + 0.13059252500534058, + 0.04926300421357155, + 0.07898976653814316, + 0.03243811056017876, + 0.20282204449176788, + 0.09095750004053116, + 0.08756707608699799, + 0.040728725492954254 + ], + "credit_map": [ + 0.029450729489326477, + 0.039301563054323196, + 0.013623078353703022, + 0.019485557451844215, + 0.020853117108345032, + 0.03175131976604462, + 0.016423583030700684, + 0.01991906948387623, + 0.015454479493200779, + 0.012871989980340004, + 0.008430346846580505, + 0.009027647785842419, + 0.01285149622708559, + 0.013706992380321026, + 0.00979885458946228, + 0.013691464439034462, + 0.030501380562782288, + 0.021553920581936836, + 0.015943167731165886, + 0.013103529810905457, + 0.010871133767068386, + 0.019091453403234482, + 0.00857763271778822, + 0.010950315743684769, + 0.005810003727674484, + 0.005898796953260899, + 0.01302101369947195, + 0.006224642973393202, + 0.004113224800676107, + 0.005761699751019478, + 0.004919739440083504, + 0.0035138854291290045, + 0.014750885777175426, + 0.01183224655687809, + 0.008082354441285133, + 0.007776567712426186, + 0.010085919871926308, + 0.010457797907292843, + 0.010522820986807346, + 0.005481180734932423, + 0.0029770408291369677, + 0.0034558435436338186, + 0.003608385333791375, + 0.003931238315999508, + 0.004825190175324678, + 0.0033567168284207582, + 0.0052879573777318, + 0.004995740484446287, + 0.0144125297665596, + 0.01590055413544178, + 0.025594687089323997, + 0.056269314140081406, + 0.013204851187765598, + 0.024893365800380707, + 0.017084971070289612, + 0.035461775958538055, + 0.01699267514050007, + 0.016891948878765106, + 0.008435651659965515, + 0.006964759435504675, + 0.010685027576982975, + 0.011332494206726551, + 0.009805974550545216, + 0.009848973713815212, + 0.013312254101037979, + 0.0146552799269557, + 0.008846999146044254, + 0.010056463070213795, + 0.006562889087945223, + 0.015759935602545738, + 0.008660605177283287, + 0.00971265509724617, + 0.0046440246514976025, + 0.004766516387462616, + 0.004265138879418373, + 0.004939788021147251, + 0.005031908862292767, + 0.005231460090726614, + 0.005911494605243206, + 0.0059383949264883995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.0025239300448447466 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255929194390774 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.005255929194390774, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255929194390774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13963177800178528, + 0.12149710953235626, + 0.07952149212360382, + 0.04827675223350525, + 0.05806537717580795, + 0.1331213116645813, + 0.24722351133823395, + 0.06161124259233475, + 0.07038650661706924, + 0.040664903819561005 + ], + "credit_map": [ + 0.014590119011700153, + 0.015983618795871735, + 0.018360771238803864, + 0.017128877341747284, + 0.02247735485434532, + 0.017441418021917343, + 0.017692063003778458, + 0.015957551077008247, + 0.01889086328446865, + 0.011524430476129055, + 0.011700902134180069, + 0.011312855407595634, + 0.015271032229065895, + 0.015597999095916748, + 0.022500326856970787, + 0.014698706567287445, + 0.010098272934556007, + 0.00862148031592369, + 0.01225186325609684, + 0.010225537233054638, + 0.012645790353417397, + 0.00933082401752472, + 0.009825495071709156, + 0.00652222940698266, + 0.00834132730960846, + 0.00916164368391037, + 0.005930876359343529, + 0.0039453934878110886, + 0.004235229920595884, + 0.005701254587620497, + 0.005675551947206259, + 0.0052854763343930244, + 0.008038418367505074, + 0.005376200657337904, + 0.007023626938462257, + 0.008475049398839474, + 0.006346954498440027, + 0.006716672796756029, + 0.007776087149977684, + 0.008312367834150791, + 0.015718607231974602, + 0.015399475581943989, + 0.018538741394877434, + 0.011896750889718533, + 0.015127521008253098, + 0.01575116068124771, + 0.013888143934309483, + 0.02680090069770813, + 0.012410704046487808, + 0.01790454611182213, + 0.037668656557798386, + 0.07443296164274216, + 0.0156515222042799, + 0.03073676861822605, + 0.021525515243411064, + 0.0368928387761116, + 0.010846205987036228, + 0.008590116165578365, + 0.006012168247252703, + 0.005472433287650347, + 0.008436017669737339, + 0.007402324117720127, + 0.007621149532496929, + 0.007230822462588549, + 0.0091098016127944, + 0.010310911573469639, + 0.008245383389294147, + 0.009373103268444538, + 0.004945981316268444, + 0.011274491436779499, + 0.007015031296759844, + 0.010111805982887745, + 0.00342094199731946, + 0.004906381014734507, + 0.002851421944797039, + 0.003490166272968054, + 0.005461511202156544, + 0.005135619081556797, + 0.007658609189093113, + 0.00774025171995163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.005255929194390774 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327998638153076 + ], + "opposite_logprob_tokens": [ + -0.7577998638153076 + ], + "logprob_reward": -0.6327998638153076, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327998638153076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1270909607410431, + 0.04350646585226059, + 0.0724012479186058, + 0.02291496843099594, + 0.11176146566867828, + 0.12309346348047256, + 0.33155667781829834, + 0.039009638130664825, + 0.028898777440190315, + 0.09976627677679062 + ], + "credit_map": [ + 0.021249372512102127, + 0.018543405458331108, + 0.011296295560896397, + 0.011083180084824562, + 0.020560728386044502, + 0.018209874629974365, + 0.014563633129000664, + 0.01158447191119194, + 0.008826232515275478, + 0.005167580209672451, + 0.003733950899913907, + 0.003920272458344698, + 0.004574532154947519, + 0.006230394821614027, + 0.005376963410526514, + 0.005676538683474064, + 0.013179832138121128, + 0.012093359604477882, + 0.008922976441681385, + 0.0065918718464672565, + 0.009210490621626377, + 0.00967145711183548, + 0.006248934660106897, + 0.0064823259599506855, + 0.00308344978839159, + 0.003894929075613618, + 0.002397570526227355, + 0.001991520868614316, + 0.002843251684680581, + 0.0028605216648429632, + 0.0030211827252060175, + 0.002822542330250144, + 0.010330956429243088, + 0.016969794407486916, + 0.010188562795519829, + 0.017645787447690964, + 0.01765597239136696, + 0.013318724930286407, + 0.015084617771208286, + 0.01056703832000494, + 0.009305913932621479, + 0.01413547433912754, + 0.01576651632785797, + 0.02227255515754223, + 0.01322016492486, + 0.014781764708459377, + 0.011820434592664242, + 0.0217906404286623, + 0.013543077744543552, + 0.02100415527820587, + 0.05381356179714203, + 0.11020760238170624, + 0.016483129933476448, + 0.04457186162471771, + 0.02624894678592682, + 0.0456843338906765, + 0.004904102999716997, + 0.005523488391190767, + 0.0039457292295992374, + 0.004515571985393763, + 0.005618936847895384, + 0.004739284981042147, + 0.005771663971245289, + 0.003990859258919954, + 0.004127600695937872, + 0.005077860783785582, + 0.0026775149162858725, + 0.00298328441567719, + 0.003137104446068406, + 0.004013143479824066, + 0.0034875995479524136, + 0.0033946698531508446, + 0.008275925181806087, + 0.012227213010191917, + 0.013261783868074417, + 0.01766069605946541, + 0.010805419646203518, + 0.010086405090987682, + 0.016627112403512, + 0.010821721516549587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.6327998638153076 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42894434928894043 + ], + "opposite_logprob_tokens": [ + -1.0539443492889404 + ], + "logprob_reward": -0.42894434928894043, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42894434928894043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12555193901062012, + 0.06965194642543793, + 0.03548946604132652, + 0.06720425933599472, + 0.11249101161956787, + 0.1648782342672348, + 0.09468245506286621, + 0.08729317784309387, + 0.04386106878519058, + 0.19889646768569946 + ], + "credit_map": [ + 0.02289191260933876, + 0.020071465522050858, + 0.009628496132791042, + 0.010517633520066738, + 0.017510151490569115, + 0.018654929473996162, + 0.013273737393319607, + 0.013003617525100708, + 0.019225969910621643, + 0.007653104141354561, + 0.007666601799428463, + 0.007266900967806578, + 0.004215122200548649, + 0.010274864733219147, + 0.006589971948415041, + 0.006759411655366421, + 0.004594419151544571, + 0.004226300865411758, + 0.0037235666532069445, + 0.0031957225874066353, + 0.005523035768419504, + 0.004484063480049372, + 0.0055638947524130344, + 0.0041784667409956455, + 0.005519720260053873, + 0.005467759445309639, + 0.02103877253830433, + 0.010697456076741219, + 0.003931799437850714, + 0.008774049580097198, + 0.005988604389131069, + 0.005786099936813116, + 0.009716522879898548, + 0.014703513123095036, + 0.010994549840688705, + 0.016846397891640663, + 0.015558581799268723, + 0.0138612175360322, + 0.01952916383743286, + 0.011281064711511135, + 0.012014434672892094, + 0.015852292999625206, + 0.02096937969326973, + 0.025659644976258278, + 0.01994265615940094, + 0.01732041873037815, + 0.02328595332801342, + 0.029833441600203514, + 0.005437696818262339, + 0.009172228164970875, + 0.015859024599194527, + 0.02888081967830658, + 0.006216296926140785, + 0.013353581540286541, + 0.008292289450764656, + 0.00747052114456892, + 0.01848602294921875, + 0.014051752164959908, + 0.007770849391818047, + 0.00679966202005744, + 0.009970693849027157, + 0.011622573249042034, + 0.008702689781785011, + 0.009888938628137112, + 0.006232301238924265, + 0.006299890577793121, + 0.005389759317040443, + 0.005162928719073534, + 0.0036863123532384634, + 0.0063277240842580795, + 0.004958615638315678, + 0.005803537089377642, + 0.01628771796822548, + 0.023088596761226654, + 0.02883366122841835, + 0.036845751106739044, + 0.017945313826203346, + 0.021125836297869682, + 0.031072041019797325, + 0.023697540163993835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.42894434928894043 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, wide shot, rendered with the delicate brushstrokes of a child's sketchbook, captures a panoramic view of the vast expanse of the sea cliffs. The world is composed of soft blocks of color and evocative lines. The sky is a watercolor-like blend of soft pink and pale blue, while the ocean is a deep, tranquil azure, punctuated by simple white brushstrokes that outline the waves. At the center of this vast canvas, a tiny figure, Matsuzaki Umi, a girl in a sailor uniform, stands alone. She wears a classic white sailor top, a navy scarf, and a matching pleated skirt. A strong sea breeze blows fiercely from the front, sweeping her chin-length dark brown hair completely back, revealing her resolute and serene face. The wind also pushes the hem of her skirt tightly behind her, causing it to dance violently and silently in the air. Her feet are firmly rooted on the grassy clifftop, her body slightly leaning forward to fight the wind. Her tiny figure displays an unyielding determination in the vast world, creating a tranquil atmosphere that is both lonely and powerful.", + "seed": 27218, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_13/index_8628_seed_27218_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1499268613755703, + "vlm_logprob": -0.1499268613755703, + "vlm_yes_no_margin": 3.174999988079071, + "vlm_token_credit": -0.1499268613755703, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04224950447678566 + ], + "opposite_logprob_tokens": [ + -3.5422494411468506 + ], + "logprob_reward": -0.04224950447678566, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.04224950447678566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.651628851890564, + 0.023443076759576797, + 0.01606239750981331, + 0.013149465434253216, + 0.03396298363804817, + 0.02817944996058941, + 0.01728653348982334, + 0.038107357919216156, + 0.00952354446053505, + 0.16865624487400055 + ], + "credit_map": [ + 0.02352461777627468, + 0.2678593695163727, + 0.01514563150703907, + 0.019989049062132835, + 0.021843181923031807, + 0.27246612310409546, + 0.008171009831130505, + 0.02262994274497032, + 0.0020752218551933765, + 0.0033014428336173296, + 0.003167306538671255, + 0.004607631359249353, + 0.0011713558342307806, + 0.003959041088819504, + 0.0017308173701167107, + 0.0034302608110010624, + 0.0009056890266947448, + 0.004405501298606396, + 0.0012456547701731324, + 0.0011437424691393971, + 0.0012079559965059161, + 0.0049605099484324455, + 0.000648805289529264, + 0.0015445402823388577, + 0.0011642967583611608, + 0.0025293047074228525, + 0.0015130663523450494, + 0.0016368215437978506, + 0.0013475586893036962, + 0.002638626378029585, + 0.0009887497872114182, + 0.0013310420326888561, + 0.0022884190548211336, + 0.007407164666801691, + 0.002658688696101308, + 0.0030409982427954674, + 0.009551964700222015, + 0.004375003278255463, + 0.001980616245418787, + 0.002660130150616169, + 0.0013805485796183348, + 0.0055597745813429356, + 0.0015753257321193814, + 0.0024489264469593763, + 0.009273391216993332, + 0.004130532965064049, + 0.0014601629227399826, + 0.002350792521610856, + 0.0016790141817182302, + 0.0033285366371273994, + 0.0018461125437170267, + 0.001999080879613757, + 0.002730939770117402, + 0.002455266658216715, + 0.0009844659361988306, + 0.0022631180472671986, + 0.0008354549645446241, + 0.015449054539203644, + 0.0006601688801310956, + 0.0005946878809481859, + 0.0170616302639246, + 0.0022520010825246572, + 0.0006903721950948238, + 0.0005639884620904922, + 0.0007943946402519941, + 0.0012930595548823476, + 0.0005569294444285333, + 0.0012281881645321846, + 0.0023907243739813566, + 0.0019171121530234814, + 0.0004843118949793279, + 0.0008588247583247721, + 0.003142406465485692, + 0.07000508904457092, + 0.0013183245901018381, + 0.0024144831113517284, + 0.07819052040576935, + 0.010529499500989914, + 0.0019614496268332005, + 0.0010944961104542017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.04224950447678566 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004135865718126297 + ], + "opposite_logprob_tokens": [ + -5.504136085510254 + ], + "logprob_reward": -0.004135865718126297, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004135865718126297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1421312689781189, + 0.02834502048790455, + 0.02807130105793476, + 0.007667134050279856, + 0.034897349774837494, + 0.15717646479606628, + 0.026389731094241142, + 0.10918823629617691, + 0.24656662344932556, + 0.21956680715084076 + ], + "credit_map": [ + 0.006657256279140711, + 0.0579574890434742, + 0.0032141203992068768, + 0.00416432972997427, + 0.0060529387556016445, + 0.057217493653297424, + 0.0027579194866120815, + 0.0041097369976341724, + 0.0027424804866313934, + 0.005863499362021685, + 0.0033309590071439743, + 0.0021627473179250956, + 0.002215936314314604, + 0.006220304407179356, + 0.0022080601193010807, + 0.003601035801693797, + 0.0030776013154536486, + 0.00572276022285223, + 0.0021499022841453552, + 0.003124187234789133, + 0.0030260428320616484, + 0.005764475557953119, + 0.0017102102283388376, + 0.0034961248748004436, + 0.0008134907693602145, + 0.0014213636750355363, + 0.0008035094360820949, + 0.0006687873392365873, + 0.0008197861607186496, + 0.0015199171612039208, + 0.0006086526555009186, + 0.0010116271441802382, + 0.0021993147674947977, + 0.007196743041276932, + 0.0026162131689488888, + 0.002514598425477743, + 0.009921576827764511, + 0.005173060577362776, + 0.002770979655906558, + 0.0025048658717423677, + 0.0035035989712923765, + 0.06408502906560898, + 0.0015674745664000511, + 0.0021446067839860916, + 0.07192862778902054, + 0.010017991065979004, + 0.001837796182371676, + 0.0020913355983793736, + 0.0026385316159576178, + 0.0034074890427291393, + 0.0024238373152911663, + 0.002417108044028282, + 0.0050400723703205585, + 0.004983705934137106, + 0.0020373291335999966, + 0.003441660199314356, + 0.0027759354561567307, + 0.03997642546892166, + 0.0016966332914307714, + 0.001756860874593258, + 0.04972818121314049, + 0.01052155252546072, + 0.0017541397828608751, + 0.0009785161819308996, + 0.004933163523674011, + 0.10426826775074005, + 0.0021683285012841225, + 0.003964437637478113, + 0.11228753626346588, + 0.012941396795213223, + 0.0031012860126793385, + 0.0029022381640970707, + 0.004128533415496349, + 0.09264905750751495, + 0.002153670648112893, + 0.002224910771474242, + 0.1014866828918457, + 0.012752966955304146, + 0.0022813461255282164, + 0.0018896611873060465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.004135865718126297 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951184779405594 + ], + "opposite_logprob_tokens": [ + -5.130951404571533 + ], + "logprob_reward": -0.005951184779405594, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951184779405594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36697617173194885, + 0.02041396126151085, + 0.01860993541777134, + 0.012307844124734402, + 0.007656590547412634, + 0.01089643593877554, + 0.008638246916234493, + 0.026605304330587387, + 0.1453775018453598, + 0.3825180232524872 + ], + "credit_map": [ + 0.010030623525381088, + 0.1611403226852417, + 0.005115005653351545, + 0.0052560786716639996, + 0.012259173206984997, + 0.16353151202201843, + 0.0035553474444895983, + 0.0060881394892930984, + 0.0030573720578104258, + 0.0028390558436512947, + 0.0021040893625468016, + 0.0016623192932456732, + 0.0023438530042767525, + 0.0029212716035544872, + 0.0031428097281605005, + 0.002343191998079419, + 0.002633690135553479, + 0.0030239440966397524, + 0.0015516107669100165, + 0.0022005324717611074, + 0.002245146781206131, + 0.0030683691147714853, + 0.0013211435871198773, + 0.002565497299656272, + 0.0010404700879007578, + 0.0018062391318380833, + 0.0011513113277032971, + 0.0009376470115967095, + 0.0018656805623322725, + 0.002297732513397932, + 0.0016252681380137801, + 0.0015834947116672993, + 0.0006708676810376346, + 0.0012287418358027935, + 0.000745485071092844, + 0.0008466584840789437, + 0.0011915372451767325, + 0.0010596662759780884, + 0.001051159342750907, + 0.0008624750189483166, + 0.0007569860899820924, + 0.0028563011437654495, + 0.00043936536530964077, + 0.0007266488391906023, + 0.0033882027491927147, + 0.001200961065478623, + 0.0008357066544704139, + 0.0006922655738890171, + 0.000769658072385937, + 0.0013153229374438524, + 0.0007804675260558724, + 0.0012077207211405039, + 0.0013495349558070302, + 0.0015528370859101415, + 0.0008875833009369671, + 0.000775122141931206, + 0.0008096238016150892, + 0.010399017482995987, + 0.0006989585817791522, + 0.0006019363063387573, + 0.011191152967512608, + 0.001575165311805904, + 0.0007198821986094117, + 0.000609566573984921, + 0.0030782592948526144, + 0.06031477451324463, + 0.0014755523297935724, + 0.0021665021777153015, + 0.06610485911369324, + 0.008868980221450329, + 0.0016843992052599788, + 0.0016841829055920243, + 0.006223395932465792, + 0.16332799196243286, + 0.003479065839201212, + 0.004300547763705254, + 0.17788726091384888, + 0.021051662042737007, + 0.004067134577780962, + 0.0021809465251863003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.005951184779405594 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486400842666626 + ], + "opposite_logprob_tokens": [ + -1.2236400842666626 + ], + "logprob_reward": -0.3486400842666626, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486400842666626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1636933833360672, + 0.04425479844212532, + 0.010242108255624771, + 0.016970045864582062, + 0.0329042486846447, + 0.0863623395562172, + 0.01252638641744852, + 0.03584308922290802, + 0.5896498560905457, + 0.007553777191787958 + ], + "credit_map": [ + 0.005451096687465906, + 0.06984208524227142, + 0.0030040962155908346, + 0.0034522200003266335, + 0.005470895208418369, + 0.06890159100294113, + 0.0030297429766505957, + 0.0045416452921926975, + 0.0043547735549509525, + 0.010152929462492466, + 0.0033747940324246883, + 0.0030698433984071016, + 0.003990919794887304, + 0.009884411469101906, + 0.004866264294832945, + 0.00456086453050375, + 0.0012427668552845716, + 0.001829142915084958, + 0.0008945681038312614, + 0.0012193095171824098, + 0.0009614094742573798, + 0.0019516829634085298, + 0.0009774063946679235, + 0.0011658219154924154, + 0.001253447262570262, + 0.003355159657076001, + 0.0016083770897239447, + 0.001553420559503138, + 0.002333556767553091, + 0.003103228285908699, + 0.0020249686203897, + 0.001737888902425766, + 0.0018461415311321616, + 0.010770843364298344, + 0.0013639763928949833, + 0.0019085227977484465, + 0.01136790867894888, + 0.00205629994161427, + 0.0020080814138054848, + 0.0015824759611859918, + 0.0022526117973029613, + 0.03477223590016365, + 0.0011493847705423832, + 0.001447509159334004, + 0.03857320547103882, + 0.005374907050281763, + 0.0013642514823004603, + 0.0014282310148701072, + 0.0009417084511369467, + 0.003037456190213561, + 0.00048662000335752964, + 0.0009251210140064359, + 0.0035574804060161114, + 0.0016012629494071007, + 0.000965912186075002, + 0.0010108260903507471, + 0.0008411264279857278, + 0.014073600992560387, + 0.0005961874267086387, + 0.00045752766891382635, + 0.01618610881268978, + 0.002504609292373061, + 0.0007622637785971165, + 0.0004216631641611457, + 0.010349356569349766, + 0.2585645318031311, + 0.004061853978782892, + 0.00646665645763278, + 0.2709655463695526, + 0.02872113324701786, + 0.006308377254754305, + 0.004212378989905119, + 0.0009205976966768503, + 0.0008351965807378292, + 0.0009176487801596522, + 0.0014401564840227365, + 0.0009560706093907356, + 0.0008867602446116507, + 0.0006494108820334077, + 0.0009479356231167912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.3486400842666626 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34865766763687134 + ], + "opposite_logprob_tokens": [ + -1.2236576080322266 + ], + "logprob_reward": -0.34865766763687134, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.34865766763687134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21294420957565308, + 0.02979133278131485, + 0.006765337195247412, + 0.015181812457740307, + 0.010129586793482304, + 0.07046954333782196, + 0.00694523099809885, + 0.015154237858951092, + 0.5330125689506531, + 0.09960615634918213 + ], + "credit_map": [ + 0.00583835830911994, + 0.09493990987539291, + 0.0018638010369613767, + 0.003019292838871479, + 0.007079374510794878, + 0.09493660926818848, + 0.0023054443299770355, + 0.002961399033665657, + 0.002250814810395241, + 0.007545424625277519, + 0.0023384056985378265, + 0.002133791334927082, + 0.0019872747361660004, + 0.007505637593567371, + 0.002776174573227763, + 0.003253808943554759, + 0.0003373760264366865, + 0.0020675512496382, + 0.00030424236319959164, + 0.0005233794217929244, + 0.00043764247675426304, + 0.0021319573279470205, + 0.0005063334247097373, + 0.0004568546137306839, + 0.0009575456497259438, + 0.003607563441619277, + 0.0012194839073345065, + 0.0011939299292862415, + 0.0016952301375567913, + 0.0034806670155376196, + 0.0016902594361454248, + 0.0013371316017583013, + 0.0006186433602124453, + 0.0030086911283433437, + 0.0006235765758901834, + 0.0006776543450541794, + 0.0026971118059009314, + 0.001026748213917017, + 0.000783192808739841, + 0.0006939672748558223, + 0.0016005781944841146, + 0.02876616083085537, + 0.0007930462597869337, + 0.0010198362870141864, + 0.03201840817928314, + 0.004382193088531494, + 0.0008423381368629634, + 0.0010469815460965037, + 0.0006449084612540901, + 0.0013432835694402456, + 0.00030772440368309617, + 0.000551308854483068, + 0.0019087393302470446, + 0.0009865387110039592, + 0.0003559039905667305, + 0.0008468229789286852, + 0.0004359661543276161, + 0.0055721052922308445, + 0.0003591645509004593, + 0.00029071292374283075, + 0.006655131466686726, + 0.0012578683672472835, + 0.00033670372795313597, + 0.00024658412439748645, + 0.009311830624938011, + 0.23319502174854279, + 0.0036285545211285353, + 0.006063668057322502, + 0.24504418671131134, + 0.026088813319802284, + 0.00567643204703927, + 0.0040040574967861176, + 0.0018639328191056848, + 0.04133431240916252, + 0.000724289973732084, + 0.0016254613874480128, + 0.045978594571352005, + 0.00616593798622489, + 0.0011357298353686929, + 0.0007778813596814871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.34865766763687134 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, wide shot, rendered with the delicate brushstrokes of a child's sketchbook, captures a panoramic view of the vast expanse of the sea cliffs. The world is composed of soft blocks of color and evocative lines. The sky is a watercolor-like blend of soft pink and pale blue, while the ocean is a deep, tranquil azure, punctuated by simple white brushstrokes that outline the waves. At the center of this vast canvas, a tiny figure, Matsuzaki Umi, a girl in a sailor uniform, stands alone. She wears a classic white sailor top, a navy scarf, and a matching pleated skirt. A strong sea breeze blows fiercely from the front, sweeping her chin-length dark brown hair completely back, revealing her resolute and serene face. The wind also pushes the hem of her skirt tightly behind her, causing it to dance violently and silently in the air. Her feet are firmly rooted on the grassy clifftop, her body slightly leaning forward to fight the wind. Her tiny figure displays an unyielding determination in the vast world, creating a tranquil atmosphere that is both lonely and powerful.", + "seed": 27219, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_13/index_8628_seed_27219_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15174706834368407, + "vlm_logprob": -0.15174706834368407, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.15174706834368407, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12486078590154648 + ], + "opposite_logprob_tokens": [ + -2.2498607635498047 + ], + "logprob_reward": -0.12486078590154648, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12486078590154648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1694040447473526, + 0.34707170724868774, + 0.17191538214683533, + 0.09931069612503052, + 0.05066895857453346, + 0.033781006932258606, + 0.05827796831727028, + 0.018793860450387, + 0.0269143246114254, + 0.023862093687057495 + ], + "credit_map": [ + 0.039770789444446564, + 0.0358053557574749, + 0.024787330999970436, + 0.01432738732546568, + 0.011992321349680424, + 0.012281727977097034, + 0.017391541972756386, + 0.013047573156654835, + 0.08806953579187393, + 0.13168849050998688, + 0.015943365171551704, + 0.009014724753797054, + 0.011347442865371704, + 0.07228823751211166, + 0.009076772257685661, + 0.009643089026212692, + 0.04572632163763046, + 0.058640506118535995, + 0.007303803227841854, + 0.009555921889841557, + 0.006702419836074114, + 0.02782381884753704, + 0.010347113944590092, + 0.0058154514990746975, + 0.0225356575101614, + 0.02624092437326908, + 0.007366878911852837, + 0.009057392366230488, + 0.005929974839091301, + 0.016351083293557167, + 0.004930385388433933, + 0.006898390594869852, + 0.00657098600640893, + 0.008315603248775005, + 0.0035835925955325365, + 0.0055545917712152, + 0.006335899233818054, + 0.006419205106794834, + 0.009994826279580593, + 0.003894249675795436, + 0.007868178188800812, + 0.009821765124797821, + 0.0034962990321218967, + 0.0022721290588378906, + 0.0021681333892047405, + 0.004160920158028603, + 0.002356488723307848, + 0.0016370906960219145, + 0.010729105211794376, + 0.01507911179214716, + 0.0037711872719228268, + 0.004514569416642189, + 0.0035508472938090563, + 0.011793135665357113, + 0.005103703122586012, + 0.0037363041192293167, + 0.002792780054733157, + 0.0034158567432314157, + 0.0018014037050306797, + 0.0023451184388250113, + 0.0012840087292715907, + 0.003783122170716524, + 0.0015940697630867362, + 0.001777498284354806, + 0.003925681579858065, + 0.005781847983598709, + 0.0018202950013801455, + 0.0034077682066708803, + 0.0019549066200852394, + 0.0046987272799015045, + 0.0026070817839354277, + 0.0027180116157978773, + 0.003230867674574256, + 0.003752653719857335, + 0.003824892919510603, + 0.0026894088368862867, + 0.0015201837522909045, + 0.004543637856841087, + 0.0018781631952151656, + 0.0024222831707447767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.12486078590154648 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002854560734704137 + ], + "opposite_logprob_tokens": [ + -5.877854347229004 + ], + "logprob_reward": -0.002854560734704137, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002854560734704137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26266708970069885, + 0.11939772963523865, + 0.06508102267980576, + 0.13089299201965332, + 0.09445586055517197, + 0.04195975139737129, + 0.13645930588245392, + 0.03801677003502846, + 0.05293235555291176, + 0.05813715234398842 + ], + "credit_map": [ + 0.055888280272483826, + 0.057965442538261414, + 0.028080452233552933, + 0.02478743903338909, + 0.020304083824157715, + 0.02432810515165329, + 0.028249848634004593, + 0.023063436150550842, + 0.029753314331173897, + 0.04221934452652931, + 0.007323587778955698, + 0.005655545741319656, + 0.005235341843217611, + 0.017131973057985306, + 0.0072430032305419445, + 0.004835618659853935, + 0.014426402747631073, + 0.013912290334701538, + 0.005792687647044659, + 0.0061942958272993565, + 0.00580970011651516, + 0.008204502984881401, + 0.006119908299297094, + 0.004621241707354784, + 0.03900846093893051, + 0.0442691408097744, + 0.007330521009862423, + 0.007911793887615204, + 0.007841933518648148, + 0.012248193845152855, + 0.006574054714292288, + 0.0057088942267000675, + 0.012738454155623913, + 0.012529164552688599, + 0.007378013804554939, + 0.008239688351750374, + 0.012175911106169224, + 0.013637425377964973, + 0.019539812579751015, + 0.008217387832701206, + 0.009241605177521706, + 0.009697694331407547, + 0.0020646380726248026, + 0.002339722588658333, + 0.003982171881943941, + 0.0043371268548071384, + 0.006753680761903524, + 0.0035431094001978636, + 0.02286909893155098, + 0.03341253101825714, + 0.009752780199050903, + 0.011203810572624207, + 0.008749084547162056, + 0.026325300335884094, + 0.01397722028195858, + 0.010169470682740211, + 0.004990315530449152, + 0.005987328942865133, + 0.004332761280238628, + 0.004184586461633444, + 0.0035687091294676065, + 0.005545633379369974, + 0.005097527056932449, + 0.004309909883886576, + 0.009904843755066395, + 0.010863314382731915, + 0.004280345048755407, + 0.00515410490334034, + 0.005360051989555359, + 0.00830264575779438, + 0.00531755993142724, + 0.003749490948393941, + 0.017132220789790154, + 0.017754511907696724, + 0.0021996141877025366, + 0.003383611561730504, + 0.0026236996054649353, + 0.006486978381872177, + 0.0064720348455011845, + 0.002084478735923767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.002854560734704137 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041004871018230915 + ], + "opposite_logprob_tokens": [ + -5.504100322723389 + ], + "logprob_reward": -0.0041004871018230915, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041004871018230915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23317746818065643, + 0.15108966827392578, + 0.11329708993434906, + 0.0928230881690979, + 0.06454774737358093, + 0.035723645240068436, + 0.1949339658021927, + 0.0481039434671402, + 0.02536964602768421, + 0.040933758020401 + ], + "credit_map": [ + 0.033238157629966736, + 0.028637638315558434, + 0.03512013331055641, + 0.027902936562895775, + 0.030058976262807846, + 0.020561503246426582, + 0.03860872983932495, + 0.019049374386668205, + 0.05141939967870712, + 0.059793516993522644, + 0.005754150915890932, + 0.0037304898723959923, + 0.005790311377495527, + 0.014684832654893398, + 0.005432095378637314, + 0.004484855569899082, + 0.02419574186205864, + 0.02966000698506832, + 0.006072669290006161, + 0.0062204948626458645, + 0.010558937676250935, + 0.011701653711497784, + 0.01796053536236286, + 0.006927045527845621, + 0.021572643890976906, + 0.022271374240517616, + 0.0045931655913591385, + 0.006265238393098116, + 0.005692187696695328, + 0.018100116401910782, + 0.007428041193634272, + 0.006900313775986433, + 0.00944902841001749, + 0.010511171072721481, + 0.003421161090955138, + 0.004925689660012722, + 0.008017422631382942, + 0.011459599249064922, + 0.009480101987719536, + 0.007283566985279322, + 0.004186215344816446, + 0.004881411790847778, + 0.003505736356601119, + 0.0034609204158186913, + 0.004818100482225418, + 0.004486246965825558, + 0.00710683036595583, + 0.003278183052316308, + 0.039160922169685364, + 0.053499046713113785, + 0.01165581215173006, + 0.013473539613187313, + 0.010492715053260326, + 0.03408932313323021, + 0.018561774864792824, + 0.014000819995999336, + 0.007330354768782854, + 0.010873277671635151, + 0.005093913059681654, + 0.0051316251046955585, + 0.0030185505747795105, + 0.007455289363861084, + 0.005064454860985279, + 0.0041364748030900955, + 0.003944130148738623, + 0.004108276218175888, + 0.002030144678428769, + 0.0031278005335479975, + 0.0026784527581185102, + 0.0039412034675478935, + 0.0032642746809870005, + 0.0022753612138330936, + 0.009920279495418072, + 0.010576589964330196, + 0.0023400657810270786, + 0.003326083067804575, + 0.0024134621489793062, + 0.006126950960606337, + 0.0034097556490451097, + 0.0028205695562064648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0041004871018230915 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31344759464263916 + ], + "opposite_logprob_tokens": [ + -1.3134475946426392 + ], + "logprob_reward": -0.31344759464263916, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31344759464263916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1583191156387329, + 0.13210368156433105, + 0.21204186975955963, + 0.15713956952095032, + 0.15087193250656128, + 0.024578964337706566, + 0.05339067056775093, + 0.018917949870228767, + 0.04794912412762642, + 0.04468712955713272 + ], + "credit_map": [ + 0.046949487179517746, + 0.0435514822602272, + 0.007969386875629425, + 0.009565569460391998, + 0.013185280375182629, + 0.011446045711636543, + 0.01353894267231226, + 0.01211290992796421, + 0.021335966885089874, + 0.03268808498978615, + 0.008939152583479881, + 0.007272737566381693, + 0.00692675681784749, + 0.03783543407917023, + 0.0071272654458880424, + 0.009978286921977997, + 0.06868461519479752, + 0.08340409398078918, + 0.007858668453991413, + 0.00922402273863554, + 0.0072958399541676044, + 0.021840602159500122, + 0.00921548344194889, + 0.00451853359118104, + 0.05049647018313408, + 0.05946153774857521, + 0.005423148162662983, + 0.005572189576923847, + 0.007103854790329933, + 0.01564197987318039, + 0.00941559299826622, + 0.004024770576506853, + 0.021102480590343475, + 0.020121930167078972, + 0.008455381728708744, + 0.011090677231550217, + 0.02010747231543064, + 0.014661094173789024, + 0.03958908095955849, + 0.015743808820843697, + 0.003544177394360304, + 0.003766095731407404, + 0.0020086602307856083, + 0.0022980347275733948, + 0.0025382465682923794, + 0.003330453298985958, + 0.00520233903080225, + 0.001890954445116222, + 0.013323240913450718, + 0.013671047054231167, + 0.002622388070449233, + 0.0030481971334666014, + 0.00320116663351655, + 0.007558043114840984, + 0.0065831393003463745, + 0.003383440198376775, + 0.0032479073852300644, + 0.0037594365421682596, + 0.0010484590893611312, + 0.00181771966163069, + 0.0021827244199812412, + 0.002311030402779579, + 0.003226972883567214, + 0.0013236984377726912, + 0.006227891426533461, + 0.007441774941980839, + 0.003665665164589882, + 0.006239724811166525, + 0.0046314662322402, + 0.008283630013465881, + 0.0059610879980027676, + 0.005497877951711416, + 0.005759877618402243, + 0.007852872833609581, + 0.003225148655474186, + 0.003838862059637904, + 0.00410926528275013, + 0.010633030906319618, + 0.0050382367335259914, + 0.004229832906275988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.31344759464263916 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a panoramic wide shot of sea cliffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky rendered as a watercolor-like blend of soft pink and pale blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny figure of a girl in a sailor uniform visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the girl's chin-length dark brown hair blown completely back by a strong sea breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the girl's pleated skirt pushed tightly behind her and dancing violently in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134719133377075 + ], + "opposite_logprob_tokens": [ + -1.3134719133377075 + ], + "logprob_reward": -0.3134719133377075, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134719133377075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1252707690000534, + 0.26136863231658936, + 0.11235333979129791, + 0.06319986283779144, + 0.09593009948730469, + 0.05087033659219742, + 0.08401111513376236, + 0.03293783217668533, + 0.08493319898843765, + 0.08912479877471924 + ], + "credit_map": [ + 0.029579518362879753, + 0.030121713876724243, + 0.012490207329392433, + 0.014479494653642178, + 0.009039796888828278, + 0.009613707661628723, + 0.010578897781670094, + 0.009367438964545727, + 0.08559741824865341, + 0.10597588121891022, + 0.006628378760069609, + 0.008181827142834663, + 0.00720394728705287, + 0.03466726094484329, + 0.007272042334079742, + 0.005841887556016445, + 0.02381717413663864, + 0.027101891115307808, + 0.009618209674954414, + 0.015742065384984016, + 0.005804619751870632, + 0.01637011580169201, + 0.006185358855873346, + 0.007713918574154377, + 0.00939104426652193, + 0.010327748022973537, + 0.00424134312197566, + 0.005831402726471424, + 0.004809677135199308, + 0.017083298414945602, + 0.0077138724736869335, + 0.00380148459225893, + 0.014765874482691288, + 0.012994838878512383, + 0.00562957301735878, + 0.008591080084443092, + 0.014342852868139744, + 0.01021677814424038, + 0.017898444086313248, + 0.01149066723883152, + 0.005281400866806507, + 0.0064811185002326965, + 0.0031604135874658823, + 0.005245286505669355, + 0.005387892946600914, + 0.007109265774488449, + 0.012006505392491817, + 0.006198457907885313, + 0.0241536982357502, + 0.026617074385285378, + 0.0036957338452339172, + 0.0034367095213383436, + 0.00460390979424119, + 0.009644375182688236, + 0.0079965116456151, + 0.003863102989271283, + 0.004162966273725033, + 0.00508502684533596, + 0.002576350001618266, + 0.003332360414788127, + 0.004287794232368469, + 0.005402741488069296, + 0.005646764766424894, + 0.002443829318508506, + 0.011991385370492935, + 0.0160935390740633, + 0.007074125576764345, + 0.00895726215094328, + 0.008105507120490074, + 0.013058210723102093, + 0.012192112393677235, + 0.007461073342710733, + 0.013739427551627159, + 0.01862058974802494, + 0.0060124448500573635, + 0.00710461288690567, + 0.0066318875178694725, + 0.020655089989304543, + 0.009851393289864063, + 0.006509354803711176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.3134719133377075 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_14/index_tensor([7291, 7291])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_14/index_tensor([7291, 7291])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..770712359b8580344a702e3b6ee546ed3ff499e4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_14/index_tensor([7291, 7291])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up shot from a level angle, slowly and steadily moving closer to the subject, focuses on a magnificent Chinese wedding gown hanging in the center of an antique wooden room. Crafted from a deep vermilion silk, the gown's smooth texture shimmers in the soft light. A ray of warm afternoon sunlight slants through the carved window lattice, creating a soft beam that illuminates the gown's chest. There, hand-embroidered with countless strands of fine gold thread, a phoenix spreads its wings, ready to take flight. Its wings are defined by distinct layers, each feather vivid and lifelike, its posture elegant and regal. A breeze blew by, causing the clothes to sway slightly and rhythmically. This subtle movement made the phoenix embroidered with gold thread seem to be given life. It trembled slightly with the ups and downs of the clothes. The gold thread captured and reflected the sunlight, flashing a warm and soft golden halo. The whole picture was set against the dark wooden background, full of tranquility, solemnity and silent anticipation for the coming joy, creating an ultimate atmosphere of oriental classical aesthetics.", + "seed": 7743, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_14/index_7291_seed_7743_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03041658306028694, + "vlm_logprob": -0.03041658306028694, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.03041658306028694, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01438356377184391 + ], + "opposite_logprob_tokens": [ + -5.014383792877197 + ], + "logprob_reward": -0.01438356377184391, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.01438356377184391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06166274845600128, + 0.02767392061650753, + 0.029532771557569504, + 0.03936808928847313, + 0.15196110308170319, + 0.06028967350721359, + 0.01872541382908821, + 0.5301703214645386, + 0.03459300845861435, + 0.0460229367017746 + ], + "credit_map": [ + 0.005332608241587877, + 0.008938368409872055, + 0.0046171327121555805, + 0.007100456394255161, + 0.010988453403115273, + 0.0069329701364040375, + 0.011921641416847706, + 0.0058311233296990395, + 0.002632417716085911, + 0.003136276500299573, + 0.004803398158401251, + 0.005219787359237671, + 0.00350619456730783, + 0.0022308058105409145, + 0.0035907451529055834, + 0.002554296050220728, + 0.003154642414301634, + 0.003712934907525778, + 0.0036592159885913134, + 0.0040073152631521225, + 0.00298293330706656, + 0.004585988819599152, + 0.004071669653058052, + 0.003358072368428111, + 0.002915864810347557, + 0.0072477119974792, + 0.007319284602999687, + 0.005956043489277363, + 0.0030537424609065056, + 0.0038374895229935646, + 0.0036644272040575743, + 0.00537352729588747, + 0.007929275743663311, + 0.042938705533742905, + 0.011893185786902905, + 0.01192125678062439, + 0.016137251630425453, + 0.016687622293829918, + 0.012035241350531578, + 0.03241857513785362, + 0.0054442728869616985, + 0.010902867652475834, + 0.007619981188327074, + 0.006222911179065704, + 0.006215321831405163, + 0.009173006750643253, + 0.006083471234887838, + 0.008627842180430889, + 0.002339763566851616, + 0.0022768708877265453, + 0.002736495342105627, + 0.001864745281636715, + 0.00200546532869339, + 0.002229302190244198, + 0.002490011742338538, + 0.0027827618177980185, + 0.05292927101254463, + 0.03112679347395897, + 0.16544033586978912, + 0.024532578885555267, + 0.02115573175251484, + 0.041792359203100204, + 0.026496177539229393, + 0.16669711470603943, + 0.0071287392638623714, + 0.004152554087340832, + 0.006422392092645168, + 0.0035099158994853497, + 0.0029337736777961254, + 0.002321782289072871, + 0.00342543818987906, + 0.004698415752500296, + 0.009966843761503696, + 0.00782301276922226, + 0.00746401259675622, + 0.004296781495213509, + 0.0033267145045101643, + 0.004439579788595438, + 0.004522054921835661, + 0.004183939192444086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.01438356377184391 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023324644193053246 + ], + "opposite_logprob_tokens": [ + -3.773324728012085 + ], + "logprob_reward": -0.023324644193053246, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023324644193053246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10244566202163696, + 0.07754215598106384, + 0.07412581890821457, + 0.09443580359220505, + 0.08100283145904541, + 0.0921069011092186, + 0.021902525797486305, + 0.1279333531856537, + 0.28244516253471375, + 0.046059731394052505 + ], + "credit_map": [ + 0.012734203599393368, + 0.0155070461332798, + 0.009797102771699429, + 0.018471507355570793, + 0.013344033621251583, + 0.011428567580878735, + 0.010568082332611084, + 0.01059512048959732, + 0.009335659444332123, + 0.010104906745254993, + 0.008829017169773579, + 0.017832079902291298, + 0.008364895358681679, + 0.0066004544496536255, + 0.00977280829101801, + 0.006702326703816652, + 0.007181655149906874, + 0.010652541182935238, + 0.006532762199640274, + 0.024250725284218788, + 0.005704468581825495, + 0.0051921820268034935, + 0.005893663503229618, + 0.00871781725436449, + 0.007501145359128714, + 0.017593329772353172, + 0.012151247821748257, + 0.01409376785159111, + 0.0075059267692267895, + 0.011865988373756409, + 0.011033679358661175, + 0.012690712697803974, + 0.006540993228554726, + 0.018263665959239006, + 0.008464377373456955, + 0.005835420917719603, + 0.008647747337818146, + 0.010953107848763466, + 0.006994002964347601, + 0.015303509309887886, + 0.008715448901057243, + 0.009854979813098907, + 0.013627242296934128, + 0.009107794612646103, + 0.007466037292033434, + 0.012526464648544788, + 0.009892209433019161, + 0.02091672271490097, + 0.003043941454961896, + 0.0023250095546245575, + 0.0027652166318148375, + 0.0025028674863278866, + 0.002612489741295576, + 0.0025039478205144405, + 0.0033148990478366613, + 0.002834154525771737, + 0.01487682107836008, + 0.008785706013441086, + 0.037779029458761215, + 0.006666393950581551, + 0.006043717265129089, + 0.009846008382737637, + 0.006731349974870682, + 0.037204332649707794, + 0.062336456030607224, + 0.0288065355271101, + 0.05628656595945358, + 0.03425971418619156, + 0.019702209159731865, + 0.024145672097802162, + 0.018726641312241554, + 0.038181371986866, + 0.010376960039138794, + 0.004423842299729586, + 0.007591153495013714, + 0.0037616363260895014, + 0.0036176119465380907, + 0.004381303675472736, + 0.0036113872192800045, + 0.008295837789773941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.023324644193053246 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033708471804857254 + ], + "opposite_logprob_tokens": [ + -3.4087085723876953 + ], + "logprob_reward": -0.033708471804857254, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033708471804857254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09789668768644333, + 0.07443877309560776, + 0.25412625074386597, + 0.030840737745165825, + 0.09237874299287796, + 0.027758143842220306, + 0.013590809889137745, + 0.3378104865550995, + 0.04299260303378105, + 0.028166761621832848 + ], + "credit_map": [ + 0.010229014791548252, + 0.017026538029313087, + 0.01398351788520813, + 0.015251821838319302, + 0.012053758837282658, + 0.007176767103374004, + 0.012702041305601597, + 0.009473230689764023, + 0.01523908507078886, + 0.008529615588486195, + 0.015249421820044518, + 0.01541617326438427, + 0.005854106042534113, + 0.0034554507583379745, + 0.005386821459978819, + 0.005308100022375584, + 0.02944379299879074, + 0.039898309856653214, + 0.02712683007121086, + 0.07589687407016754, + 0.01695454679429531, + 0.016369519755244255, + 0.018195251002907753, + 0.03024112991988659, + 0.0019954212475568056, + 0.006014402024447918, + 0.004156061913818121, + 0.004247571807354689, + 0.003250554669648409, + 0.003825998166576028, + 0.0037457437720149755, + 0.00360498228110373, + 0.005982635077089071, + 0.02708403393626213, + 0.007295551244169474, + 0.007302951533347368, + 0.006603565998375416, + 0.008771484717726707, + 0.0042683444917202, + 0.025070177391171455, + 0.002739012474194169, + 0.006118933204561472, + 0.0021803309209644794, + 0.0039277877658605576, + 0.002232413971796632, + 0.0031352071091532707, + 0.003494448261335492, + 0.00393000990152359, + 0.0017365941312164068, + 0.0012741006212309003, + 0.0022543715313076973, + 0.0014815166359767318, + 0.0015770767349749804, + 0.0014378437772393227, + 0.0019379513105377555, + 0.0018913550302386284, + 0.03329670801758766, + 0.019982166588306427, + 0.10441922396421432, + 0.016530795022845268, + 0.013975324109196663, + 0.02687634527683258, + 0.016292667016386986, + 0.10643727332353592, + 0.00828421488404274, + 0.00421005068346858, + 0.010168763808906078, + 0.005144068505614996, + 0.003001237753778696, + 0.002387002110481262, + 0.0030773933976888657, + 0.0067198690958321095, + 0.004700895864516497, + 0.004324218723922968, + 0.004006815142929554, + 0.0034547531977295876, + 0.0024656972382217646, + 0.002105186926200986, + 0.004045051988214254, + 0.003064140910282731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.033708471804857254 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017134760273620486 + ], + "opposite_logprob_tokens": [ + -6.376713275909424 + ], + "logprob_reward": -0.0017134760273620486, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017134760273620486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17009833455085754, + 0.06193722039461136, + 0.05696386471390724, + 0.042429909110069275, + 0.12977419793605804, + 0.08142262697219849, + 0.030309133231639862, + 0.1740734726190567, + 0.11661748588085175, + 0.13637374341487885 + ], + "credit_map": [ + 0.019889242947101593, + 0.0322287455201149, + 0.025735709816217422, + 0.027480537071824074, + 0.020526833832263947, + 0.013252484612166882, + 0.01724972575902939, + 0.013735059648752213, + 0.007017779629677534, + 0.007771999575197697, + 0.009039966389536858, + 0.015128248371183872, + 0.006505066528916359, + 0.004411351401358843, + 0.007747144438326359, + 0.004315657541155815, + 0.004230956546962261, + 0.006795721594244242, + 0.005442585330456495, + 0.009045645594596863, + 0.009584132581949234, + 0.006670130416750908, + 0.008964326232671738, + 0.006230366416275501, + 0.004540358670055866, + 0.00618762569501996, + 0.0037608083803206682, + 0.007231239229440689, + 0.004414559341967106, + 0.004209630191326141, + 0.006098736543208361, + 0.005986946169286966, + 0.008368991315364838, + 0.037470463663339615, + 0.010673653334379196, + 0.010390794835984707, + 0.012730277143418789, + 0.01400867197662592, + 0.010430698283016682, + 0.025700638070702553, + 0.005726989358663559, + 0.015043394640088081, + 0.00750813027843833, + 0.012128306552767754, + 0.008819651789963245, + 0.011986654251813889, + 0.011487632989883423, + 0.00872186291962862, + 0.0027433461509644985, + 0.0034796781837940216, + 0.006022094748914242, + 0.002918752608820796, + 0.002438982017338276, + 0.0034916929434984922, + 0.002913742559030652, + 0.0063008456490933895, + 0.017349960282444954, + 0.011123971082270145, + 0.0484200045466423, + 0.009576871991157532, + 0.009164196439087391, + 0.01534998882561922, + 0.009024395607411861, + 0.05406409129500389, + 0.023895571008324623, + 0.01282905600965023, + 0.019674750044941902, + 0.011545472778379917, + 0.010356081649661064, + 0.013258671388030052, + 0.011569321155548096, + 0.01348856557160616, + 0.025281241163611412, + 0.02158384770154953, + 0.019670572131872177, + 0.016648195683956146, + 0.009747503325343132, + 0.014938208274543285, + 0.011933882720768452, + 0.01657029427587986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.0017134760273620486 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895275950431824 + ], + "opposite_logprob_tokens": [ + -2.5789527893066406 + ], + "logprob_reward": -0.07895275950431824, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895275950431824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10175403207540512, + 0.03636712580919266, + 0.021747544407844543, + 0.1076468750834465, + 0.07272426784038544, + 0.20134900510311127, + 0.030455153435468674, + 0.2438047230243683, + 0.08253522217273712, + 0.10161600261926651 + ], + "credit_map": [ + 0.012040672823786736, + 0.014975445345044136, + 0.014684655703604221, + 0.019577525556087494, + 0.010559088550508022, + 0.009884174913167953, + 0.010608741082251072, + 0.009423728100955486, + 0.0028036143630743027, + 0.005190826021134853, + 0.003876020899042487, + 0.005511508788913488, + 0.004123595543205738, + 0.004672857001423836, + 0.004901495762169361, + 0.005287207663059235, + 0.00296636950224638, + 0.002520996844395995, + 0.003180193016305566, + 0.003349502570927143, + 0.0023423610255122185, + 0.002372996648773551, + 0.002369876950979233, + 0.002645248081535101, + 0.005317030940204859, + 0.009526964277029037, + 0.020359938964247704, + 0.01270239520817995, + 0.008492009714245796, + 0.013119282200932503, + 0.009598556905984879, + 0.02853069268167019, + 0.005130815785378218, + 0.013791331090033054, + 0.004602030385285616, + 0.0076512969098985195, + 0.01010783202946186, + 0.008640535175800323, + 0.013198590837419033, + 0.009601837024092674, + 0.01695084199309349, + 0.028820589184761047, + 0.01838502287864685, + 0.017371563240885735, + 0.018832560628652573, + 0.032019924372434616, + 0.014773362316191196, + 0.054195139557123184, + 0.0031094681471586227, + 0.0053335921838879585, + 0.0035031295847147703, + 0.004299042746424675, + 0.003309041727334261, + 0.0035894557368010283, + 0.003937171306461096, + 0.0033742506057024, + 0.024418413639068604, + 0.014626733027398586, + 0.07447252422571182, + 0.013044255785644054, + 0.010663461871445179, + 0.019178971648216248, + 0.011682530865073204, + 0.07571783661842346, + 0.019644809886813164, + 0.00863664224743843, + 0.016203800216317177, + 0.009605773724615574, + 0.006427111104130745, + 0.006370087619870901, + 0.005418604705482721, + 0.010228391736745834, + 0.019645383581519127, + 0.0167012307792902, + 0.014228137210011482, + 0.012710217386484146, + 0.006825284566730261, + 0.012502575293183327, + 0.00713685154914856, + 0.01186632551252842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.07895275950431824 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot from a level angle, slowly and steadily moving closer to the subject, focuses on a magnificent Chinese wedding gown hanging in the center of an antique wooden room. Crafted from a deep vermilion silk, the gown's smooth texture shimmers in the soft light. A ray of warm afternoon sunlight slants through the carved window lattice, creating a soft beam that illuminates the gown's chest. There, hand-embroidered with countless strands of fine gold thread, a phoenix spreads its wings, ready to take flight. Its wings are defined by distinct layers, each feather vivid and lifelike, its posture elegant and regal. A breeze blew by, causing the clothes to sway slightly and rhythmically. This subtle movement made the phoenix embroidered with gold thread seem to be given life. It trembled slightly with the ups and downs of the clothes. The gold thread captured and reflected the sunlight, flashing a warm and soft golden halo. The whole picture was set against the dark wooden background, full of tranquility, solemnity and silent anticipation for the coming joy, creating an ultimate atmosphere of oriental classical aesthetics.", + "seed": 7744, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_14/index_7291_seed_7744_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.025033321278169752, + "vlm_logprob": -0.025033321278169752, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.025033321278169752, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014604687690734863 + ], + "opposite_logprob_tokens": [ + -5.139604568481445 + ], + "logprob_reward": -0.014604687690734863, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.014604687690734863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10254336148500443, + 0.09983836859464645, + 0.08333075046539307, + 0.04948250204324722, + 0.09477783739566803, + 0.06801650673151016, + 0.116996631026268, + 0.21689817309379578, + 0.08012159913778305, + 0.08799431473016739 + ], + "credit_map": [ + 0.007588052656501532, + 0.012257392518222332, + 0.01124138105660677, + 0.020234227180480957, + 0.014290438033640385, + 0.0068939789198338985, + 0.019702423363924026, + 0.010335463099181652, + 0.009785516187548637, + 0.010112136602401733, + 0.010250304825603962, + 0.019234877079725266, + 0.010353286750614643, + 0.007278363220393658, + 0.0189361609518528, + 0.013887718319892883, + 0.00492851110175252, + 0.010886039584875107, + 0.00944885890930891, + 0.018245697021484375, + 0.01244968082755804, + 0.003963134251534939, + 0.015382034704089165, + 0.008026775903999805, + 0.005260058678686619, + 0.0052174958400428295, + 0.0069849430583417416, + 0.01098225824534893, + 0.005962010473012924, + 0.0029842618387192488, + 0.007610941771417856, + 0.00448052491992712, + 0.010788732208311558, + 0.011407669633626938, + 0.012043406255543232, + 0.015487151220440865, + 0.01034974679350853, + 0.010249542072415352, + 0.01396810170263052, + 0.010483477264642715, + 0.005054013337939978, + 0.0070839812979102135, + 0.008338335901498795, + 0.00938850175589323, + 0.005963427014648914, + 0.011029429733753204, + 0.009310348890721798, + 0.01184845995157957, + 0.00972363818436861, + 0.011448907665908337, + 0.012359009124338627, + 0.02232545055449009, + 0.014099125750362873, + 0.015528975985944271, + 0.01690886914730072, + 0.014602643437683582, + 0.014978336170315742, + 0.01675138808786869, + 0.015062784776091576, + 0.044502824544906616, + 0.016726238653063774, + 0.07187865674495697, + 0.02298792637884617, + 0.01401001401245594, + 0.0058028982020914555, + 0.011074911803007126, + 0.012049361132085323, + 0.017078308388590813, + 0.009906268678605556, + 0.008793518878519535, + 0.008852130733430386, + 0.006564196664839983, + 0.007265320047736168, + 0.010130536742508411, + 0.01587560959160328, + 0.019156230613589287, + 0.008077885955572128, + 0.00806440506130457, + 0.008527934551239014, + 0.010896386578679085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.014604687690734863 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023331863805651665 + ], + "opposite_logprob_tokens": [ + -3.773331880569458 + ], + "logprob_reward": -0.023331863805651665, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023331863805651665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17412017285823822, + 0.04509269446134567, + 0.017850935459136963, + 0.022274218499660492, + 0.03651997819542885, + 0.06774113327264786, + 0.2047114074230194, + 0.2597622573375702, + 0.09545411914587021, + 0.07647307217121124 + ], + "credit_map": [ + 0.02875981293618679, + 0.020001521334052086, + 0.019865410402417183, + 0.028793301433324814, + 0.013261810876429081, + 0.014465589076280594, + 0.03310636430978775, + 0.015866361558437347, + 0.0037916407454758883, + 0.005981292575597763, + 0.005513067357242107, + 0.007254385389387608, + 0.005996460095047951, + 0.0030938538257032633, + 0.007398098707199097, + 0.006063894834369421, + 0.0016026595840230584, + 0.0015106115024536848, + 0.001946520758792758, + 0.003912485670298338, + 0.0026137600652873516, + 0.0011971521889790893, + 0.003214928787201643, + 0.001852816203609109, + 0.0018124403432011604, + 0.0025897701270878315, + 0.003441320499405265, + 0.00489331828430295, + 0.002304827095940709, + 0.0012046543415635824, + 0.0038921681698411703, + 0.0021357201039791107, + 0.002885496011003852, + 0.003960822708904743, + 0.0042107501067221165, + 0.005764216650277376, + 0.004698369652032852, + 0.005801759660243988, + 0.005324040539562702, + 0.0038745219353586435, + 0.004136096220463514, + 0.008224458433687687, + 0.006108637899160385, + 0.008275942876935005, + 0.008618202060461044, + 0.01570529118180275, + 0.011074413545429707, + 0.0055980924516916275, + 0.016984114423394203, + 0.01727035827934742, + 0.016568118706345558, + 0.02582847699522972, + 0.021099599078297615, + 0.046653810888528824, + 0.03267047926783562, + 0.027636447921395302, + 0.013078881427645683, + 0.016650386154651642, + 0.017927903681993484, + 0.03993304818868637, + 0.016042934730648994, + 0.08541478216648102, + 0.03827836364507675, + 0.03243594616651535, + 0.007438718341290951, + 0.016182497143745422, + 0.01700820028781891, + 0.01721796579658985, + 0.012154915370047092, + 0.00713583268225193, + 0.010886826552450657, + 0.007429162971675396, + 0.008056314662098885, + 0.010341365821659565, + 0.010746274143457413, + 0.015589733608067036, + 0.008829277008771896, + 0.006190294399857521, + 0.008662053383886814, + 0.008057752624154091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.023331863805651665 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029807114973664284 + ], + "opposite_logprob_tokens": [ + -3.5298070907592773 + ], + "logprob_reward": -0.029807114973664284, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807114973664284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48252400755882263, + 0.1888788342475891, + 0.04494365677237511, + 0.012606068514287472, + 0.02403228171169758, + 0.019193096086382866, + 0.0718497633934021, + 0.1082286387681961, + 0.02488907240331173, + 0.022854579612612724 + ], + "credit_map": [ + 0.08074250817298889, + 0.05497536435723305, + 0.054806675761938095, + 0.07507159560918808, + 0.03947693109512329, + 0.04638080298900604, + 0.08835545182228088, + 0.0427146814763546, + 0.01962810382246971, + 0.01967586949467659, + 0.024225430563092232, + 0.032246463000774384, + 0.01736896112561226, + 0.015683645382523537, + 0.033191531896591187, + 0.026858819648623466, + 0.005816092249006033, + 0.0038625048473477364, + 0.006438071373850107, + 0.005440365057438612, + 0.0036524711176753044, + 0.00619067158550024, + 0.004197334870696068, + 0.009346144273877144, + 0.001793369185179472, + 0.0012249829014763236, + 0.001489426358602941, + 0.001885295263491571, + 0.0014724750071763992, + 0.0008806808618828654, + 0.001966700190678239, + 0.001893139211460948, + 0.004856438841670752, + 0.0026240881998091936, + 0.003423758316785097, + 0.00360514665953815, + 0.001941306865774095, + 0.0020626247860491276, + 0.0029913021717220545, + 0.0025276148226112127, + 0.0010888696415349841, + 0.0018281840020790696, + 0.0020980832632631063, + 0.0020127950701862574, + 0.0017184088937938213, + 0.005546821281313896, + 0.0029940237291157246, + 0.0019059107871726155, + 0.005824740044772625, + 0.006103236228227615, + 0.007893537171185017, + 0.009809303097426891, + 0.007308971136808395, + 0.015187175013124943, + 0.011450615711510181, + 0.008272184059023857, + 0.006073357071727514, + 0.0068185171112418175, + 0.007074273657053709, + 0.020809680223464966, + 0.005435788072645664, + 0.04015399143099785, + 0.013896910473704338, + 0.007966119796037674, + 0.001869300613179803, + 0.0025486533995717764, + 0.007205812726169825, + 0.003806475782766938, + 0.00243581784889102, + 0.001626916229724884, + 0.0037474543787539005, + 0.0016486413078382611, + 0.001873693778179586, + 0.0029836068861186504, + 0.003045920515432954, + 0.003746809670701623, + 0.0027262435760349035, + 0.002151978202164173, + 0.002900603925809264, + 0.003425722010433674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.029807114973664284 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002488613361492753 + ], + "opposite_logprob_tokens": [ + -6.002488613128662 + ], + "logprob_reward": -0.002488613361492753, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002488613361492753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1732938289642334, + 0.02207266166806221, + 0.018231388181447983, + 0.014916878193616867, + 0.014480875805020332, + 0.053234491497278214, + 0.2731899917125702, + 0.19816873967647552, + 0.05747734010219574, + 0.17493385076522827 + ], + "credit_map": [ + 0.030686885118484497, + 0.022262411192059517, + 0.020009158179163933, + 0.0253562293946743, + 0.013003798201680183, + 0.015373523347079754, + 0.030590126290917397, + 0.01601169817149639, + 0.0017509596655145288, + 0.002033533062785864, + 0.0028255910146981478, + 0.003714499296620488, + 0.0029576027300208807, + 0.0019122891826555133, + 0.004389284178614616, + 0.002488902769982815, + 0.0017852619057521224, + 0.001582602970302105, + 0.0022783612366765738, + 0.002983937505632639, + 0.0027763359248638153, + 0.0014106883900240064, + 0.003189208684489131, + 0.00222499156370759, + 0.0016896493034437299, + 0.0017299815081059933, + 0.0017075358191505075, + 0.0027853266801685095, + 0.001857828232459724, + 0.000961471872869879, + 0.0027332657482475042, + 0.0014518189709633589, + 0.0017600723076611757, + 0.0016845623031258583, + 0.00152123486623168, + 0.002067495370283723, + 0.0020368227269500494, + 0.0014839970972388983, + 0.0021064740139991045, + 0.0018202163046225905, + 0.0023019625805318356, + 0.004253596067428589, + 0.004790150560438633, + 0.006005352828651667, + 0.005042068660259247, + 0.014187660068273544, + 0.00985571090131998, + 0.006797990296036005, + 0.021161148324608803, + 0.023913072422146797, + 0.02338072471320629, + 0.041401661932468414, + 0.032012294977903366, + 0.05494506284594536, + 0.04212699458003044, + 0.03424903750419617, + 0.010300262831151485, + 0.020254429429769516, + 0.011712172068655491, + 0.04194965586066246, + 0.01596926338970661, + 0.059675827622413635, + 0.025059707462787628, + 0.013247423805296421, + 0.0037013867404311895, + 0.009182728826999664, + 0.010180006735026836, + 0.010719007812440395, + 0.007049990817904472, + 0.0043314602226018906, + 0.008217786438763142, + 0.004094976000487804, + 0.01575852930545807, + 0.03551888093352318, + 0.02215689979493618, + 0.029296984896063805, + 0.023672806099057198, + 0.011448124423623085, + 0.014812592417001724, + 0.022269031032919884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.002488613361492753 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493432655930519 + ], + "opposite_logprob_tokens": [ + -2.92993426322937 + ], + "logprob_reward": -0.05493432655930519, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493432655930519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10806026309728622, + 0.10297760367393494, + 0.02135637030005455, + 0.0283474400639534, + 0.056755430996418, + 0.10941125452518463, + 0.1942232847213745, + 0.2496567666530609, + 0.06931669265031815, + 0.05989489331841469 + ], + "credit_map": [ + 0.017440680414438248, + 0.010147490538656712, + 0.01247909851372242, + 0.01838080585002899, + 0.00751477899029851, + 0.010789777152240276, + 0.02047712355852127, + 0.010830501094460487, + 0.007879962213337421, + 0.012170331552624702, + 0.01159763801842928, + 0.01899372786283493, + 0.010964523069560528, + 0.008602350018918514, + 0.01802704483270645, + 0.014742027968168259, + 0.0017217369750142097, + 0.0013884567888453603, + 0.003233937779441476, + 0.0033199272584170103, + 0.0018237372860312462, + 0.0020010708831250668, + 0.0026868432760238647, + 0.0051806592382490635, + 0.003601762466132641, + 0.0026541328988969326, + 0.003896991489455104, + 0.004009191412478685, + 0.003383383620530367, + 0.0019385248888283968, + 0.004807299003005028, + 0.00405615521594882, + 0.003918355796486139, + 0.006609712727367878, + 0.0056413509882986546, + 0.009730211459100246, + 0.008207459934055805, + 0.006806271616369486, + 0.010414534248411655, + 0.005427535157650709, + 0.0055832345969974995, + 0.009426388889551163, + 0.01381443627178669, + 0.012864301912486553, + 0.008609527722001076, + 0.02563825622200966, + 0.019711138680577278, + 0.013763970695436, + 0.01693630963563919, + 0.015442796982824802, + 0.022614480927586555, + 0.02587234415113926, + 0.018943561241030693, + 0.04096374288201332, + 0.02882828377187252, + 0.024621766060590744, + 0.014104580506682396, + 0.01757742092013359, + 0.014842511154711246, + 0.04134995490312576, + 0.016345785930752754, + 0.08944368362426758, + 0.03464717045426369, + 0.021345647051930428, + 0.00416927644982934, + 0.010695171542465687, + 0.012357844039797783, + 0.014032609760761261, + 0.007219359744340181, + 0.004887274466454983, + 0.010021807625889778, + 0.005933347623795271, + 0.004370964597910643, + 0.005519745871424675, + 0.008968277834355831, + 0.00773296132683754, + 0.009810772724449635, + 0.006432248279452324, + 0.008499476127326488, + 0.00856044515967369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.05493432655930519 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot from a level angle, slowly and steadily moving closer to the subject, focuses on a magnificent Chinese wedding gown hanging in the center of an antique wooden room. Crafted from a deep vermilion silk, the gown's smooth texture shimmers in the soft light. A ray of warm afternoon sunlight slants through the carved window lattice, creating a soft beam that illuminates the gown's chest. There, hand-embroidered with countless strands of fine gold thread, a phoenix spreads its wings, ready to take flight. Its wings are defined by distinct layers, each feather vivid and lifelike, its posture elegant and regal. A breeze blew by, causing the clothes to sway slightly and rhythmically. This subtle movement made the phoenix embroidered with gold thread seem to be given life. It trembled slightly with the ups and downs of the clothes. The gold thread captured and reflected the sunlight, flashing a warm and soft golden halo. The whole picture was set against the dark wooden background, full of tranquility, solemnity and silent anticipation for the coming joy, creating an ultimate atmosphere of oriental classical aesthetics.", + "seed": 7745, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_14/index_7291_seed_7745_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.059868072345852855, + "vlm_logprob": -0.059868072345852855, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.059868072345852855, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025834020227193832 + ], + "opposite_logprob_tokens": [ + -4.650834083557129 + ], + "logprob_reward": -0.025834020227193832, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.025834020227193832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10939610004425049, + 0.052134204655885696, + 0.13643516600131989, + 0.05032876878976822, + 0.0177433043718338, + 0.07190454006195068, + 0.03789155185222626, + 0.06595367193222046, + 0.21713773906230927, + 0.24107491970062256 + ], + "credit_map": [ + 0.010082931257784367, + 0.007910680957138538, + 0.010826636105775833, + 0.011728466488420963, + 0.012268627993762493, + 0.006630136165767908, + 0.03352576494216919, + 0.016422858461737633, + 0.002725128782913089, + 0.0028891738038510084, + 0.0026869336143136024, + 0.0045926980674266815, + 0.005765821784734726, + 0.002123801736161113, + 0.016824966296553612, + 0.014525680802762508, + 0.006040062289685011, + 0.006367826368659735, + 0.004138419404625893, + 0.010278933681547642, + 0.009490244090557098, + 0.006412964314222336, + 0.04533807560801506, + 0.04836864024400711, + 0.00508169224485755, + 0.0038408769760280848, + 0.0035991051699966192, + 0.006930883973836899, + 0.007759416475892067, + 0.00285523128695786, + 0.013747045770287514, + 0.006514514330774546, + 0.0014440387021750212, + 0.0023697679862380028, + 0.0016761475708335638, + 0.0020919768139719963, + 0.004446080885827541, + 0.0012222335208207369, + 0.0021898047998547554, + 0.0023032526951283216, + 0.003190060146152973, + 0.008426040410995483, + 0.006344558205455542, + 0.010530532337725163, + 0.010819765739142895, + 0.005250221583992243, + 0.011247574351727962, + 0.016095783561468124, + 0.002229071920737624, + 0.004914631135761738, + 0.0033062950242310762, + 0.004281930159777403, + 0.005676002241671085, + 0.002092959126457572, + 0.009030082263052464, + 0.006360583007335663, + 0.004128085449337959, + 0.012020453810691833, + 0.007468940690159798, + 0.008110708557069302, + 0.011179685592651367, + 0.0059427074156701565, + 0.007277455180883408, + 0.009825630113482475, + 0.010712843388319016, + 0.01700110174715519, + 0.0212725717574358, + 0.027222441509366035, + 0.022402750328183174, + 0.017937228083610535, + 0.04453476518392563, + 0.05605403706431389, + 0.008925287052989006, + 0.018938645720481873, + 0.024810098111629486, + 0.09712816029787064, + 0.011054255068302155, + 0.008143221959471703, + 0.043882887810468674, + 0.02819235622882843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.025834020227193832 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500698834657669 + ], + "opposite_logprob_tokens": [ + -2.930006980895996 + ], + "logprob_reward": -0.05500698834657669, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500698834657669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062000468373298645, + 0.1483897715806961, + 0.1372697651386261, + 0.03142959251999855, + 0.022674353793263435, + 0.0347256064414978, + 0.07750263810157776, + 0.03937434405088425, + 0.35258808732032776, + 0.09404543042182922 + ], + "credit_map": [ + 0.004742276389151812, + 0.006871241144835949, + 0.005781165789812803, + 0.0056368401274085045, + 0.012483582831919193, + 0.0068118213675916195, + 0.011580627411603928, + 0.008092907257378101, + 0.006590489763766527, + 0.004720187745988369, + 0.0034931940026581287, + 0.009414813481271267, + 0.014385614544153214, + 0.007132426369935274, + 0.05373435467481613, + 0.04891867935657501, + 0.005272656213492155, + 0.006303310394287109, + 0.003988487180322409, + 0.010221240110695362, + 0.013899147510528564, + 0.005426464136689901, + 0.04982716217637062, + 0.04233128949999809, + 0.002911373507231474, + 0.0027909050695598125, + 0.0024090774822980165, + 0.004289088770747185, + 0.00596056180074811, + 0.0018765426939353347, + 0.008304799906909466, + 0.0028872385155409575, + 0.002573223551735282, + 0.002459555631503463, + 0.0017186772311106324, + 0.0029907149728387594, + 0.00525362603366375, + 0.0012925041373819113, + 0.004187683574855328, + 0.0021983657497912645, + 0.0018278439529240131, + 0.004330615978688002, + 0.002641472266986966, + 0.004192078020423651, + 0.006300388369709253, + 0.0024840938858687878, + 0.004806906916201115, + 0.008142201229929924, + 0.005318435374647379, + 0.011351835913956165, + 0.005643332377076149, + 0.009325625374913216, + 0.012853912077844143, + 0.00570971705019474, + 0.012275266461074352, + 0.015024500899016857, + 0.0026634999085217714, + 0.006926883012056351, + 0.003557881573215127, + 0.004357801750302315, + 0.008287145756185055, + 0.00540013425052166, + 0.005001071374863386, + 0.0031799247954040766, + 0.01696782000362873, + 0.025859661400318146, + 0.049660537391901016, + 0.029150234535336494, + 0.0275659691542387, + 0.02063230611383915, + 0.1068798303604126, + 0.07587169855833054, + 0.0052624596282839775, + 0.013435970060527325, + 0.007136918604373932, + 0.02936004474759102, + 0.0073867435567080975, + 0.005059867165982723, + 0.01784549653530121, + 0.008557923138141632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 2 + ], + "selected_reward": -0.05500698834657669 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003321036696434 + ], + "opposite_logprob_tokens": [ + -2.350332021713257 + ], + "logprob_reward": -0.1003321036696434, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003321036696434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30630728602409363, + 0.03608398884534836, + 0.07129482179880142, + 0.10734869539737701, + 0.06806410104036331, + 0.019604433327913284, + 0.026244549080729485, + 0.04385243356227875, + 0.27930161356925964, + 0.041898079216480255 + ], + "credit_map": [ + 0.02305586263537407, + 0.023543331772089005, + 0.024523064494132996, + 0.033231642097234726, + 0.033640991896390915, + 0.018272630870342255, + 0.1017853170633316, + 0.04825446382164955, + 0.002970017958432436, + 0.003258521668612957, + 0.0015879165148362517, + 0.00422541331499815, + 0.006719942670315504, + 0.002946208231151104, + 0.009027158841490746, + 0.00534881092607975, + 0.004302816465497017, + 0.0037151165306568146, + 0.002980623161420226, + 0.006755186710506678, + 0.00619120430201292, + 0.002741925185546279, + 0.026774229481816292, + 0.01783372275531292, + 0.013112649321556091, + 0.010743552818894386, + 0.010454243049025536, + 0.014487011358141899, + 0.01271367073059082, + 0.012399761006236076, + 0.021386276930570602, + 0.01205152552574873, + 0.005290655419230461, + 0.006932081654667854, + 0.006388789974153042, + 0.00811244361102581, + 0.013953883200883865, + 0.006985889747738838, + 0.0128870764747262, + 0.0075132837519049644, + 0.0020777105819433928, + 0.0019501253264024854, + 0.0019030708353966475, + 0.002156072063371539, + 0.003160873893648386, + 0.0020681521855294704, + 0.0032323100604116917, + 0.0030561168678104877, + 0.0016105955000966787, + 0.003073831321671605, + 0.0028641256503760815, + 0.003340934170410037, + 0.004447816871106625, + 0.0017999900737777352, + 0.003751748241484165, + 0.005355505738407373, + 0.0035739638842642307, + 0.006074282340705395, + 0.004336886573582888, + 0.004851615522056818, + 0.006707563530653715, + 0.0028721773996949196, + 0.011381389573216438, + 0.00405455706641078, + 0.01307949610054493, + 0.021452100947499275, + 0.03596260026097298, + 0.021394724026322365, + 0.02129688486456871, + 0.016440758481621742, + 0.09607290476560593, + 0.053602125495672226, + 0.0017956686206161976, + 0.0034713507629930973, + 0.00559387169778347, + 0.014218369498848915, + 0.003587610088288784, + 0.0020729268435388803, + 0.007897616364061832, + 0.00326066673733294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.1003321036696434 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253320559859276 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253320559859276, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253320559859276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07656509429216385, + 0.2997984290122986, + 0.07406514883041382, + 0.033865392208099365, + 0.02075224556028843, + 0.020582091063261032, + 0.0635526105761528, + 0.08562541753053665, + 0.2784641683101654, + 0.04672941938042641 + ], + "credit_map": [ + 0.008809630759060383, + 0.006561064161360264, + 0.004449214320629835, + 0.009433720260858536, + 0.012278512120246887, + 0.005274667404592037, + 0.020462319254875183, + 0.009295964613556862, + 0.01187483873218298, + 0.008110292255878448, + 0.006196671165525913, + 0.020449647679924965, + 0.024031559005379677, + 0.013198358938097954, + 0.10964710265398026, + 0.10628995299339294, + 0.004198740236461163, + 0.0034812246449291706, + 0.0028921854682266712, + 0.006645543966442347, + 0.006902388297021389, + 0.0030293427407741547, + 0.02802492305636406, + 0.01889079436659813, + 0.0027221497148275375, + 0.0032417892944067717, + 0.0026320514734834433, + 0.004332888405770063, + 0.005896660964936018, + 0.0029324570205062628, + 0.007954421453177929, + 0.004152975976467133, + 0.0011044793063774705, + 0.00257730670273304, + 0.0016350892838090658, + 0.002726175356656313, + 0.00501612201333046, + 0.0015203702496364713, + 0.003777659498155117, + 0.002395043848082423, + 0.0009150397963821888, + 0.0021836699452251196, + 0.001588183338753879, + 0.002552920253947377, + 0.004074952565133572, + 0.0014726192457601428, + 0.0038611004129052162, + 0.0039336043410003185, + 0.0032508012373000383, + 0.011090689338743687, + 0.006234323605895042, + 0.007667482364922762, + 0.016677487641572952, + 0.0043362886644899845, + 0.008615816943347454, + 0.005679720547050238, + 0.006992592476308346, + 0.013875001110136509, + 0.0076097180135548115, + 0.008368603885173798, + 0.02086455188691616, + 0.006941894069314003, + 0.013463359326124191, + 0.007509698625653982, + 0.012776852585375309, + 0.023545851930975914, + 0.0319058857858181, + 0.02337224967777729, + 0.02202007919549942, + 0.018741806969046593, + 0.08850454539060593, + 0.05759691447019577, + 0.004631489515304565, + 0.005190692842006683, + 0.0037094000726938248, + 0.0114924106746912, + 0.0035157427191734314, + 0.002607588656246662, + 0.009935736656188965, + 0.005646358244121075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.005253320559859276 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291392892599106 + ], + "opposite_logprob_tokens": [ + -2.2379138469696045 + ], + "logprob_reward": -0.11291392892599106, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291392892599106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12066031247377396, + 0.27857834100723267, + 0.08393677324056625, + 0.031192747876048088, + 0.027915485203266144, + 0.029717177152633667, + 0.07885515689849854, + 0.1354992687702179, + 0.05822432041168213, + 0.15542039275169373 + ], + "credit_map": [ + 0.00887238048017025, + 0.007149224169552326, + 0.007695880718529224, + 0.012490286491811275, + 0.013421585783362389, + 0.004980351775884628, + 0.04361898824572563, + 0.022431615740060806, + 0.012463631108403206, + 0.009702219627797604, + 0.008074703626334667, + 0.019510215148329735, + 0.02483956329524517, + 0.011349988169968128, + 0.10239899158477783, + 0.09023905545473099, + 0.004369556438177824, + 0.004073181189596653, + 0.0031180987134575844, + 0.008433584123849869, + 0.00858856551349163, + 0.002706818748265505, + 0.03257066011428833, + 0.020076319575309753, + 0.002190692350268364, + 0.0034323055297136307, + 0.00260185357183218, + 0.004262134898453951, + 0.006556644104421139, + 0.0019329122733324766, + 0.004062228370457888, + 0.006153973750770092, + 0.00184441392775625, + 0.003008925821632147, + 0.0020164756570011377, + 0.0039007747545838356, + 0.00561254657804966, + 0.002899251878261566, + 0.004050977993756533, + 0.004582117311656475, + 0.0015254326863214374, + 0.0035522673279047012, + 0.002135445596650243, + 0.0028169220313429832, + 0.008539581671357155, + 0.002290973672643304, + 0.004193788394331932, + 0.0046627637930214405, + 0.005480639636516571, + 0.009237219579517841, + 0.007177154067903757, + 0.009895621798932552, + 0.011986440978944302, + 0.006173265166580677, + 0.012846921570599079, + 0.016057897359132767, + 0.007761458866298199, + 0.01885787583887577, + 0.012396905571222305, + 0.015851782634854317, + 0.022158609703183174, + 0.007615930400788784, + 0.03762837126851082, + 0.013228325173258781, + 0.003716631093993783, + 0.005449797958135605, + 0.006940620485693216, + 0.00549096567556262, + 0.004947626497596502, + 0.0033672500867396593, + 0.01938735693693161, + 0.008924074470996857, + 0.013150010257959366, + 0.019993146881461143, + 0.024419592693448067, + 0.032384417951107025, + 0.010732963681221008, + 0.008199118077754974, + 0.023540327325463295, + 0.02300080470740795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.11291392892599106 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot from a level angle, slowly and steadily moving closer to the subject, focuses on a magnificent Chinese wedding gown hanging in the center of an antique wooden room. Crafted from a deep vermilion silk, the gown's smooth texture shimmers in the soft light. A ray of warm afternoon sunlight slants through the carved window lattice, creating a soft beam that illuminates the gown's chest. There, hand-embroidered with countless strands of fine gold thread, a phoenix spreads its wings, ready to take flight. Its wings are defined by distinct layers, each feather vivid and lifelike, its posture elegant and regal. A breeze blew by, causing the clothes to sway slightly and rhythmically. This subtle movement made the phoenix embroidered with gold thread seem to be given life. It trembled slightly with the ups and downs of the clothes. The gold thread captured and reflected the sunlight, flashing a warm and soft golden halo. The whole picture was set against the dark wooden background, full of tranquility, solemnity and silent anticipation for the coming joy, creating an ultimate atmosphere of oriental classical aesthetics.", + "seed": 7746, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_14/index_7291_seed_7746_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0675344912102446, + "vlm_logprob": -0.0675344912102446, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.0675344912102446, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015394434332847595 + ], + "opposite_logprob_tokens": [ + -5.01539421081543 + ], + "logprob_reward": -0.015394434332847595, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.015394434332847595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20101067423820496, + 0.0635332465171814, + 0.06886864453554153, + 0.03875067085027695, + 0.05251577869057655, + 0.1098540648818016, + 0.1200353279709816, + 0.06418141722679138, + 0.05423643812537193, + 0.22701367735862732 + ], + "credit_map": [ + 0.015177939087152481, + 0.018956376239657402, + 0.019421357661485672, + 0.0202448982745409, + 0.05925874039530754, + 0.012347269803285599, + 0.03441747650504112, + 0.021186621859669685, + 0.0034724927973002195, + 0.0067464374005794525, + 0.007804104592651129, + 0.008866638876497746, + 0.012693838216364384, + 0.003235998097807169, + 0.011804587207734585, + 0.008909150026738644, + 0.013725006952881813, + 0.00463104248046875, + 0.010983210988342762, + 0.009946986101567745, + 0.007801881525665522, + 0.0047686402685940266, + 0.00802950281649828, + 0.00898237805813551, + 0.003168521448969841, + 0.005401955917477608, + 0.0038905844558030367, + 0.008674093522131443, + 0.005679329391568899, + 0.0021223309449851513, + 0.006726423278450966, + 0.003087432822212577, + 0.0030854912474751472, + 0.004342217463999987, + 0.01569906622171402, + 0.011911708861589432, + 0.0027160076424479485, + 0.003905318444594741, + 0.0038844675291329622, + 0.006971503607928753, + 0.006616597529500723, + 0.008897189050912857, + 0.026274340227246284, + 0.017678024247288704, + 0.00770484609529376, + 0.0046104504726827145, + 0.01687058061361313, + 0.021202048286795616, + 0.009321271441876888, + 0.0162366796284914, + 0.011150402948260307, + 0.015398276969790459, + 0.016691025346517563, + 0.013552396558225155, + 0.020005524158477783, + 0.017679758369922638, + 0.0055864485912024975, + 0.006108120083808899, + 0.007252851035445929, + 0.02136637642979622, + 0.005714201368391514, + 0.0033394487109035254, + 0.009710284881293774, + 0.0051036858931183815, + 0.010389870963990688, + 0.005207705777138472, + 0.007901071570813656, + 0.005869292188435793, + 0.004633937031030655, + 0.006257900968194008, + 0.005740010645240545, + 0.008236650377511978, + 0.0200484786182642, + 0.02220061607658863, + 0.011936471797525883, + 0.020587990060448647, + 0.01704680174589157, + 0.016561880707740784, + 0.051975809037685394, + 0.06665565073490143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.015394434332847595 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02987966127693653 + ], + "opposite_logprob_tokens": [ + -3.529879570007324 + ], + "logprob_reward": -0.02987966127693653, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02987966127693653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11664322763681412, + 0.03614342585206032, + 0.011383183300495148, + 0.04603782296180725, + 0.06631316989660263, + 0.06010643392801285, + 0.06007237359881401, + 0.1858769804239273, + 0.04362593963742256, + 0.3737974166870117 + ], + "credit_map": [ + 0.02054699882864952, + 0.013317907229065895, + 0.018853135406970978, + 0.010338659398257732, + 0.01462826319038868, + 0.007662901189178228, + 0.01282605342566967, + 0.018469318747520447, + 0.0018755935598164797, + 0.0036032574716955423, + 0.00354074127972126, + 0.003945422358810902, + 0.0103553906083107, + 0.0016156212659552693, + 0.007856846787035465, + 0.0033505531027913094, + 0.0015730018494650722, + 0.0011536979582160711, + 0.0010488320840522647, + 0.0016715264646336436, + 0.0022759647108614445, + 0.0007601859397254884, + 0.001667117583565414, + 0.0012328564189374447, + 0.007046634331345558, + 0.0036253936123102903, + 0.006581449881196022, + 0.007729775737971067, + 0.004620559047907591, + 0.003396776970475912, + 0.005845544394105673, + 0.007191691547632217, + 0.0038257127162069082, + 0.0036940090358257294, + 0.025057658553123474, + 0.013375031761825085, + 0.0019811035599559546, + 0.002362166764214635, + 0.0036341948434710503, + 0.01238329615443945, + 0.005121029447764158, + 0.0036647950764745474, + 0.019778208807110786, + 0.010256308130919933, + 0.003412951249629259, + 0.0026084163691848516, + 0.0065435441210865974, + 0.008721177466213703, + 0.004389145411550999, + 0.00844156090170145, + 0.005851362831890583, + 0.014749208465218544, + 0.008544016629457474, + 0.0026729481760412455, + 0.010208437219262123, + 0.005215694196522236, + 0.010394793935120106, + 0.014756102114915848, + 0.017245102673768997, + 0.07166586816310883, + 0.007409232202917337, + 0.01142118126153946, + 0.02694433555006981, + 0.02604038454592228, + 0.008397843688726425, + 0.004929522518068552, + 0.005123458802700043, + 0.003919900860637426, + 0.0040038577280938625, + 0.004701900761574507, + 0.005402982234954834, + 0.007146472577005625, + 0.030246814712882042, + 0.03821320831775665, + 0.01883481815457344, + 0.03264203667640686, + 0.03364015743136406, + 0.026521380990743637, + 0.08672422170639038, + 0.10697479546070099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.02987966127693653 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20166300237178802 + ], + "opposite_logprob_tokens": [ + -1.7016630172729492 + ], + "logprob_reward": -0.20166300237178802, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20166300237178802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3691727817058563, + 0.025868559256196022, + 0.053667180240154266, + 0.20975327491760254, + 0.16375115513801575, + 0.020355982705950737, + 0.03165372833609581, + 0.05146482586860657, + 0.015486876480281353, + 0.05882562696933746 + ], + "credit_map": [ + 0.058592285960912704, + 0.04506317898631096, + 0.06180355325341225, + 0.03976444900035858, + 0.036299750208854675, + 0.02341894619166851, + 0.036440763622522354, + 0.06778985261917114, + 0.0019307155162096024, + 0.002836274215951562, + 0.003332373220473528, + 0.003735009813681245, + 0.003783081192523241, + 0.0017467879224568605, + 0.004106177948415279, + 0.004398138727992773, + 0.010324900969862938, + 0.004628341645002365, + 0.00698960991576314, + 0.008305738680064678, + 0.006095631048083305, + 0.004053750541061163, + 0.006448870990425348, + 0.006820337846875191, + 0.04275093600153923, + 0.012975846417248249, + 0.03154624253511429, + 0.032708942890167236, + 0.013188929297029972, + 0.016571158543229103, + 0.02547544054687023, + 0.03453577309846878, + 0.008811998181045055, + 0.01259087398648262, + 0.05858425796031952, + 0.02841050736606121, + 0.00716432323679328, + 0.006732153706252575, + 0.010609954595565796, + 0.030847081914544106, + 0.0009263629326596856, + 0.0029163232538849115, + 0.0020591365173459053, + 0.004840197507292032, + 0.002731704618781805, + 0.0010355408303439617, + 0.0030570197850465775, + 0.002789698541164398, + 0.002683172933757305, + 0.004453611094504595, + 0.003750932402908802, + 0.0047127725556492805, + 0.0043963282369077206, + 0.0038836030289530754, + 0.004289218690246344, + 0.0034840917214751244, + 0.004167817998677492, + 0.004012587480247021, + 0.0041277133859694, + 0.016430092975497246, + 0.003976965788751841, + 0.0037718326784670353, + 0.007523996289819479, + 0.007453822065144777, + 0.0022070929408073425, + 0.0018276162445545197, + 0.0016781494487076998, + 0.002160602016374469, + 0.002057150239124894, + 0.0014907867880538106, + 0.0024492014199495316, + 0.0016162763349711895, + 0.00492372689768672, + 0.006659587845206261, + 0.002841580891981721, + 0.0055757914669811726, + 0.005276418756693602, + 0.004063320346176624, + 0.01281832531094551, + 0.016666870564222336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.20166300237178802 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017145470483228564 + ], + "opposite_logprob_tokens": [ + -6.376714706420898 + ], + "logprob_reward": -0.0017145470483228564, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017145470483228564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09232557564973831, + 0.03125520795583725, + 0.04938589408993721, + 0.03273305296897888, + 0.17385730147361755, + 0.15872728824615479, + 0.1251382976770401, + 0.12240459024906158, + 0.10120300203561783, + 0.11296985298395157 + ], + "credit_map": [ + 0.015354979783296585, + 0.00955483503639698, + 0.01665007881820202, + 0.010638841427862644, + 0.007923077791929245, + 0.007326563820242882, + 0.008453474380075932, + 0.01642371155321598, + 0.002823133720085025, + 0.003297218820080161, + 0.0030789030715823174, + 0.003907331731170416, + 0.007830929942429066, + 0.001716107246465981, + 0.005395654123276472, + 0.0032059247605502605, + 0.0062432922422885895, + 0.0038773948326706886, + 0.005719481967389584, + 0.007067256607115269, + 0.010859248228371143, + 0.002643757499754429, + 0.007792114280164242, + 0.0051833451725542545, + 0.0035146409645676613, + 0.003578058909624815, + 0.0033202599734067917, + 0.004038197919726372, + 0.006093896459788084, + 0.0019703698344528675, + 0.006192569155246019, + 0.004025060683488846, + 0.00979079119861126, + 0.008916842751204967, + 0.06646962463855743, + 0.025274861603975296, + 0.008960665203630924, + 0.006116890348494053, + 0.010169055312871933, + 0.03815854713320732, + 0.007727319374680519, + 0.014582923613488674, + 0.03806094825267792, + 0.024576501920819283, + 0.013555115088820457, + 0.007227728143334389, + 0.02401801012456417, + 0.02897873893380165, + 0.00918917078524828, + 0.018935412168502808, + 0.01120365597307682, + 0.013939776457846165, + 0.016762644052505493, + 0.013845891691744328, + 0.022493891417980194, + 0.018767857924103737, + 0.0073894658125936985, + 0.01044211070984602, + 0.006006061099469662, + 0.04685318470001221, + 0.006555064581334591, + 0.006350261624902487, + 0.015236820094287395, + 0.02357161045074463, + 0.019869189709424973, + 0.00629424536600709, + 0.012763392180204391, + 0.00864018127322197, + 0.0078598503023386, + 0.011258438229560852, + 0.013145252130925655, + 0.02137245051562786, + 0.011799691244959831, + 0.013790037482976913, + 0.008155175484716892, + 0.011427563615143299, + 0.010650335811078548, + 0.006568259093910456, + 0.028647717088460922, + 0.021931063383817673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.0017145470483228564 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a Chinese wedding gown hanging in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the gown made of deep vermilion silk with a smooth texture that shimmers in soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ray of warm afternoon sunlight slanting through a carved window lattice to illuminate the gown's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a phoenix embroidered with fine gold thread visible on the gown's chest with distinct feather layers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily move closer to the subject while the gown sways slightly due to a breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902081102132797 + ], + "opposite_logprob_tokens": [ + -2.4640207290649414 + ], + "logprob_reward": -0.08902081102132797, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902081102132797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07525564730167389, + 0.04652407020330429, + 0.03238529711961746, + 0.031823232769966125, + 0.057150933891534805, + 0.04164854809641838, + 0.23998120427131653, + 0.06613779067993164, + 0.06954503059387207, + 0.3395482003688812 + ], + "credit_map": [ + 0.004315229598432779, + 0.007601891178637743, + 0.008824306540191174, + 0.008554669097065926, + 0.01886068657040596, + 0.0035827241372317076, + 0.01311001367866993, + 0.010406131856143475, + 0.0029856418259441853, + 0.005058751441538334, + 0.004725683946162462, + 0.004520566668361425, + 0.01409722026437521, + 0.002338646911084652, + 0.009190024808049202, + 0.0036075348034501076, + 0.0017616769764572382, + 0.0030527408234775066, + 0.003341027069836855, + 0.00662031676620245, + 0.0046719033271074295, + 0.0013135663466528058, + 0.0067608654499053955, + 0.004863203968852758, + 0.0036566320341080427, + 0.003881772281602025, + 0.003330254927277565, + 0.005069009959697723, + 0.005504859145730734, + 0.0020115969236940145, + 0.004942028783261776, + 0.003427079413086176, + 0.003434413345530629, + 0.0034849527291953564, + 0.02123059146106243, + 0.009316474199295044, + 0.0022417742293328047, + 0.002316319150850177, + 0.0036539570428431034, + 0.011472453363239765, + 0.0016258665127679706, + 0.003973246552050114, + 0.0036810932215303183, + 0.014368690550327301, + 0.003227432956919074, + 0.0019420053577050567, + 0.0063171228393912315, + 0.006513088475912809, + 0.021728942170739174, + 0.03028799034655094, + 0.02687925659120083, + 0.043416161090135574, + 0.030103174969553947, + 0.026119520887732506, + 0.028047217056155205, + 0.033398933708667755, + 0.004195887595415115, + 0.0070299385115504265, + 0.004789490252733231, + 0.020084694027900696, + 0.005900015123188496, + 0.0035447068512439728, + 0.00847720168530941, + 0.012115856632590294, + 0.010531418956816196, + 0.005024759564548731, + 0.008101757615804672, + 0.005793552380055189, + 0.00553324818611145, + 0.009368954226374626, + 0.008332597091794014, + 0.01685873419046402, + 0.02789352647960186, + 0.026176603510975838, + 0.01882137916982174, + 0.03153613582253456, + 0.0294332318007946, + 0.018994225189089775, + 0.08314318209886551, + 0.10354992747306824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.08902081102132797 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_15/index_tensor([1624, 1624])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_15/index_tensor([1624, 1624])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..64f66c1d83d14f41b86734b9f61067ba37f9a8de --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_15/index_tensor([1624, 1624])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot, executed in stunning slow motion, follows the explosive movements of a male tennis player in his early twenties. Under the blazing afternoon sun, he pushes off the blue professional hard court with all his might, launching himself into the air. His body stretches fully in mid-air, forming a powerful bow. His short, sweat-soaked dark brown hair clings to his forehead, a single bead of sweat trickles down between his furrowed brows, his facial muscles tensed with intense concentration. He wears a bright yellow sleeveless tank top that clings to his defined muscles, dark blue athletic shorts, and a white sweatband around his wrist. At the apex of his rise, his powerful core drives his waist in a violent twist, his right arm raised high, and the carbon fiber tennis racket in his hand carves a swift arc through the air. At the highest point of the shot, the strings of the racket deformed visibly, precisely and fiercely enveloping the incoming fluorescent yellow tennis ball. Instantly, like a cannonball compressed to the limit and released, the ball ripped through the air with a whistling sound as it traveled in a nearly indiscernible line, blasting towards the other side of the court with lightning speed. Sunlight reflected off his sweat-soaked skin and the metal frame of the racket, creating dazzling spots of light. Against the backdrop of the slightly out-of-focus green netting, the entire scene was filled with the atmosphere of ultimate power, speed, and focused competition, perfectly capturing the decisive moment of this fast-paced exchange.", + "seed": 27785, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_15/index_1624_seed_27785_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12313380129635335, + "vlm_logprob": -0.12313380129635335, + "vlm_yes_no_margin": 3.324999988079071, + "vlm_token_credit": -0.12313380129635335, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.46157318353652954 + ], + "opposite_logprob_tokens": [ + -1.2115731239318848 + ], + "logprob_reward": -0.46157318353652954, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.46157318353652954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12450985610485077, + 0.22995008528232574, + 0.06003293767571449, + 0.3225833475589752, + 0.03027585707604885, + 0.09614268690347672, + 0.042500607669353485, + 0.03368671238422394, + 0.021148735657334328, + 0.03916913643479347 + ], + "credit_map": [ + 0.016115853562951088, + 0.02511054091155529, + 0.012933561578392982, + 0.02405773103237152, + 0.016029072925448418, + 0.013496560044586658, + 0.008349339477717876, + 0.008417208679020405, + 0.047572001814842224, + 0.009597290307283401, + 0.04285275563597679, + 0.009545160457491875, + 0.017680469900369644, + 0.08885294944047928, + 0.006861366331577301, + 0.00698810862377286, + 0.005928714293986559, + 0.01057475432753563, + 0.007917058654129505, + 0.005168167874217033, + 0.008206083439290524, + 0.0053110080771148205, + 0.008752059191465378, + 0.008175095543265343, + 0.021689997985959053, + 0.04560231417417526, + 0.03600752353668213, + 0.013832470402121544, + 0.021619686856865883, + 0.1311812549829483, + 0.01192929781973362, + 0.04072079434990883, + 0.002801846479997039, + 0.0031825501937419176, + 0.002851171651855111, + 0.0019955504685640335, + 0.004174521658569574, + 0.005250921938568354, + 0.003728978568688035, + 0.006290319841355085, + 0.009510506875813007, + 0.017982563003897667, + 0.011916103772819042, + 0.010817627422511578, + 0.015486979857087135, + 0.008392921648919582, + 0.011312205344438553, + 0.010723787359893322, + 0.003415044629946351, + 0.011775887571275234, + 0.005775328725576401, + 0.0035441447980701923, + 0.006576936691999435, + 0.0035130188334733248, + 0.004683331120759249, + 0.0032169176265597343, + 0.004710186272859573, + 0.003198540536686778, + 0.0020589232444763184, + 0.008656584657728672, + 0.004422855097800493, + 0.0030900328420102596, + 0.004157503135502338, + 0.003392089158296585, + 0.00292424438521266, + 0.002565754344686866, + 0.001868630526587367, + 0.0027165119536221027, + 0.0020409799180924892, + 0.004034155514091253, + 0.002159045310690999, + 0.0028394151013344526, + 0.004954223055392504, + 0.006584913469851017, + 0.007385396398603916, + 0.005585822742432356, + 0.004026397131383419, + 0.003706024494022131, + 0.003928959369659424, + 0.0029974060598760843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.46157318353652954 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906377851963043 + ], + "opposite_logprob_tokens": [ + -2.464063882827759 + ], + "logprob_reward": -0.08906377851963043, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906377851963043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020338213071227074, + 0.02472575381398201, + 0.061560891568660736, + 0.8169274926185608, + 0.006416159216314554, + 0.01441019307821989, + 0.030786283314228058, + 0.0059936498291790485, + 0.010961872525513172, + 0.00787947978824377 + ], + "credit_map": [ + 0.003416030667722225, + 0.005235279910266399, + 0.001518175471574068, + 0.0017759313341230154, + 0.0020069603342562914, + 0.003394025843590498, + 0.0015475363470613956, + 0.0014442743267863989, + 0.0034092317800968885, + 0.0014886843273416162, + 0.005567834712564945, + 0.0014823511010035872, + 0.0021065969485789537, + 0.007949351333081722, + 0.0015073159011080861, + 0.0012143898056820035, + 0.006190780084580183, + 0.007133234292268753, + 0.008129799738526344, + 0.005668659694492817, + 0.007796176243573427, + 0.007711195386946201, + 0.007762922905385494, + 0.011168120428919792, + 0.04975062608718872, + 0.1273229420185089, + 0.08962711691856384, + 0.03584197536110878, + 0.0503767766058445, + 0.3395974934101105, + 0.019739758223295212, + 0.10467085987329483, + 0.00048277724999934435, + 0.0006153837312012911, + 0.0005859649972990155, + 0.0003935151326004416, + 0.0012703179381787777, + 0.0013193119084462523, + 0.0008216900168918073, + 0.0009271985618397593, + 0.0016880318289622664, + 0.0020122937858104706, + 0.0018569850362837315, + 0.0018655137391760945, + 0.0023216968402266502, + 0.0016360764857381582, + 0.0018197107128798962, + 0.0012098842998966575, + 0.004016881808638573, + 0.009127178229391575, + 0.00458562234416604, + 0.0024810999166220427, + 0.0022740208078175783, + 0.0029205982573330402, + 0.0019414650741964579, + 0.003439416643232107, + 0.0015971474349498749, + 0.0004312785458751023, + 0.00037092287675477564, + 0.0014548299368470907, + 0.000497745000757277, + 0.0005633562686853111, + 0.0004667405446525663, + 0.0006116295699030161, + 0.0013424977660179138, + 0.0013117608614265919, + 0.0011559388367459178, + 0.0015553018311038613, + 0.001279370510019362, + 0.0016963822999969125, + 0.0012518779840320349, + 0.0013687425525858998, + 0.0010526934638619423, + 0.0007844131323508918, + 0.0007740030996501446, + 0.0011169726494699717, + 0.0008709533140063286, + 0.0014487016014754772, + 0.0010828006779775023, + 0.0007489417912438512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 6 + ], + "selected_reward": -0.08906377851963043 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005964812356978655 + ], + "opposite_logprob_tokens": [ + -5.130964756011963 + ], + "logprob_reward": -0.005964812356978655, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005964812356978655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08625222742557526, + 0.027412094175815582, + 0.026877135038375854, + 0.7212364077568054, + 0.027278006076812744, + 0.011445621028542519, + 0.051303837448358536, + 0.007594625931233168, + 0.026012374088168144, + 0.014587688259780407 + ], + "credit_map": [ + 0.008455956354737282, + 0.011718092486262321, + 0.010723810642957687, + 0.017863145098090172, + 0.014585333876311779, + 0.008695011027157307, + 0.00647595664486289, + 0.007734913378953934, + 0.0031449103262275457, + 0.003465342102572322, + 0.004822914022952318, + 0.0023356867022812366, + 0.003731748554855585, + 0.0038432145956903696, + 0.0033225081861019135, + 0.0027457671239972115, + 0.0029770031105726957, + 0.0027485063765197992, + 0.004379751160740852, + 0.002519714878872037, + 0.003499546553939581, + 0.004232947248965502, + 0.0023674683179706335, + 0.004152193665504456, + 0.04384348914027214, + 0.1123306155204773, + 0.07942333072423935, + 0.030768411234021187, + 0.04433799535036087, + 0.3004714250564575, + 0.01750653237104416, + 0.09255459904670715, + 0.002410039072856307, + 0.0035223872400820255, + 0.0018934884574264288, + 0.001789041212759912, + 0.004004945978522301, + 0.004436664283275604, + 0.003162189619615674, + 0.006059250328689814, + 0.0010468729306012392, + 0.0015816411469131708, + 0.0016342237358912826, + 0.001185604021884501, + 0.0016211727634072304, + 0.0012387785827741027, + 0.0015498102875426412, + 0.0015875166282057762, + 0.004391978960484266, + 0.015151499770581722, + 0.011091269552707672, + 0.004531094804406166, + 0.004270879086107016, + 0.00283897970803082, + 0.005052722524851561, + 0.00397540582343936, + 0.00221804971806705, + 0.00071339012356475, + 0.00034384848549962044, + 0.0015172138810157776, + 0.0006352614145725965, + 0.00045270140981301665, + 0.00080054224235937, + 0.0009136180742643774, + 0.0037186245899647474, + 0.0069993631914258, + 0.0020299432799220085, + 0.002959996461868286, + 0.0029049068689346313, + 0.003364116419106722, + 0.002370103495195508, + 0.0016653178026899695, + 0.0018650843994691968, + 0.001539515214972198, + 0.0018286437261849642, + 0.0017955771181732416, + 0.0019360908772796392, + 0.0022795763798058033, + 0.0016743696760386229, + 0.0016688300529494882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.005964812356978655 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497054383158684 + ], + "opposite_logprob_tokens": [ + -2.9299705028533936 + ], + "logprob_reward": -0.05497054383158684, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497054383158684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11073126643896103, + 0.03743661940097809, + 0.012062756344676018, + 0.7658094167709351, + 0.007648011669516563, + 0.012635846622288227, + 0.018016746267676353, + 0.016238613054156303, + 0.008075271733105183, + 0.011345405131578445 + ], + "credit_map": [ + 0.012669427320361137, + 0.022972535341978073, + 0.014350201934576035, + 0.02595563232898712, + 0.013452853076159954, + 0.009637822397053242, + 0.005775311030447483, + 0.005917493253946304, + 0.012580711394548416, + 0.0015767269069328904, + 0.001445917645469308, + 0.000956621253862977, + 0.0029253701213747263, + 0.01495069544762373, + 0.0016739550046622753, + 0.0013266257010400295, + 0.0010770212393254042, + 0.0012587212258949876, + 0.0019372778479009867, + 0.0015546053182333708, + 0.0014557709218934178, + 0.0018820887198671699, + 0.001355613931082189, + 0.001541657722555101, + 0.04638824239373207, + 0.11867256462574005, + 0.08470211923122406, + 0.03261006250977516, + 0.04719368368387222, + 0.3193979859352112, + 0.018620694056153297, + 0.09822414815425873, + 0.0005157069535925984, + 0.0009969096863642335, + 0.0003861089644487947, + 0.00028564385138452053, + 0.0012513769324868917, + 0.0013708355836570263, + 0.0006799448165111244, + 0.002161486307159066, + 0.0014964707661420107, + 0.0017195589607581496, + 0.00171365262940526, + 0.0012579027097672224, + 0.001955823739990592, + 0.0014173893723636866, + 0.0013049941044300795, + 0.001770057249814272, + 0.0016424742061644793, + 0.005514548160135746, + 0.0038832868449389935, + 0.0016115508042275906, + 0.0012711371527984738, + 0.0010666189482435584, + 0.0015317272627726197, + 0.0014954048674553633, + 0.0017243953188881278, + 0.0028713911306113005, + 0.00108812446705997, + 0.0029841731302440166, + 0.0018614561995491385, + 0.002366387750953436, + 0.001380808069370687, + 0.0019618780352175236, + 0.0011607527267187834, + 0.0012277032947167754, + 0.0006512354593724012, + 0.0009495171834714711, + 0.0010006901575252414, + 0.0013104622485116124, + 0.0008045935537666082, + 0.0009703178075142205, + 0.0016059299232438207, + 0.0017267725197598338, + 0.0024181825574487448, + 0.001622756477445364, + 0.0009743953123688698, + 0.0010060664499178529, + 0.0010947862174361944, + 0.0008965163142420352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.05497054383158684 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096688237041235 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.004096688237041235, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096688237041235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02065395936369896, + 0.18950985372066498, + 0.031230665743350983, + 0.37769952416419983, + 0.021763131022453308, + 0.0130747826769948, + 0.24886836111545563, + 0.011452906765043736, + 0.03699516877532005, + 0.048751648515462875 + ], + "credit_map": [ + 0.002655529649928212, + 0.004531707614660263, + 0.0015271628508344293, + 0.002428642939776182, + 0.0030160846654325724, + 0.0025897203013300896, + 0.0018398266984149814, + 0.0020652860403060913, + 0.05959703028202057, + 0.006850340403616428, + 0.012747355736792088, + 0.0035208899062126875, + 0.015013190917670727, + 0.07980869710445404, + 0.006027523428201675, + 0.0059448289684951305, + 0.0034984394442290068, + 0.002687847940251231, + 0.004794858396053314, + 0.002997967880219221, + 0.0032681564334779978, + 0.0033013413194566965, + 0.004852141719311476, + 0.005829912144690752, + 0.023060603067278862, + 0.05890444666147232, + 0.041392285376787186, + 0.01611507497727871, + 0.023125866428017616, + 0.15709954500198364, + 0.009274063631892204, + 0.04872764274477959, + 0.0016477475874125957, + 0.002763803116977215, + 0.001292839995585382, + 0.0009758605156093836, + 0.003134995698928833, + 0.0033587280195206404, + 0.0021883989684283733, + 0.006400757469236851, + 0.001257053343579173, + 0.0021210687700659037, + 0.001085710246115923, + 0.0013363758334890008, + 0.002230297075584531, + 0.001514479867182672, + 0.002187060657888651, + 0.0013427367666736245, + 0.022292576730251312, + 0.09554578363895416, + 0.06405461579561234, + 0.015091695822775364, + 0.015106862410902977, + 0.01233381126075983, + 0.009177826344966888, + 0.015265177004039288, + 0.0014523491263389587, + 0.0024005502928048372, + 0.0007083169184625149, + 0.0021549381781369448, + 0.0012087010545656085, + 0.0011876043863594532, + 0.0008264441275969148, + 0.0015140027971938252, + 0.005178283434361219, + 0.009791716933250427, + 0.0027263362426310778, + 0.0031360164284706116, + 0.003926081117242575, + 0.005351794417947531, + 0.002937444020062685, + 0.003947495948523283, + 0.007121965754777193, + 0.004660757724195719, + 0.004451547283679247, + 0.005434846505522728, + 0.007664286997169256, + 0.008627241477370262, + 0.006082573905587196, + 0.004708433989435434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.004096688237041235 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot, executed in stunning slow motion, follows the explosive movements of a male tennis player in his early twenties. Under the blazing afternoon sun, he pushes off the blue professional hard court with all his might, launching himself into the air. His body stretches fully in mid-air, forming a powerful bow. His short, sweat-soaked dark brown hair clings to his forehead, a single bead of sweat trickles down between his furrowed brows, his facial muscles tensed with intense concentration. He wears a bright yellow sleeveless tank top that clings to his defined muscles, dark blue athletic shorts, and a white sweatband around his wrist. At the apex of his rise, his powerful core drives his waist in a violent twist, his right arm raised high, and the carbon fiber tennis racket in his hand carves a swift arc through the air. At the highest point of the shot, the strings of the racket deformed visibly, precisely and fiercely enveloping the incoming fluorescent yellow tennis ball. Instantly, like a cannonball compressed to the limit and released, the ball ripped through the air with a whistling sound as it traveled in a nearly indiscernible line, blasting towards the other side of the court with lightning speed. Sunlight reflected off his sweat-soaked skin and the metal frame of the racket, creating dazzling spots of light. Against the backdrop of the slightly out-of-focus green netting, the entire scene was filled with the atmosphere of ultimate power, speed, and focused competition, perfectly capturing the decisive moment of this fast-paced exchange.", + "seed": 27786, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_15/index_1624_seed_27786_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2547857554629445, + "vlm_logprob": -0.2547857554629445, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.2547857554629445, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.41454029083251953 + ], + "opposite_logprob_tokens": [ + -1.2895402908325195 + ], + "logprob_reward": -0.41454029083251953, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.41454029083251953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36734193563461304, + 0.1049543172121048, + 0.05531695857644081, + 0.09472900629043579, + 0.18050576746463776, + 0.04251013323664665, + 0.03292785957455635, + 0.029305974021553993, + 0.04914390295743942, + 0.04326413944363594 + ], + "credit_map": [ + 0.09402959048748016, + 0.02139352448284626, + 0.04962010681629181, + 0.013803546316921711, + 0.03236772492527962, + 0.11131300032138824, + 0.015710385516285896, + 0.029104040935635567, + 0.01083398424088955, + 0.019278595224022865, + 0.01705961860716343, + 0.0075271641835570335, + 0.013176103122532368, + 0.011228357441723347, + 0.015409216284751892, + 0.010441279038786888, + 0.008547679521143436, + 0.0045339749194681644, + 0.005390589125454426, + 0.007169563788920641, + 0.009107787162065506, + 0.00837144535034895, + 0.003744040383026004, + 0.008451881818473339, + 0.015577912330627441, + 0.013345866464078426, + 0.007228130474686623, + 0.011690263636410236, + 0.013510345481336117, + 0.012989405542612076, + 0.009914577938616276, + 0.01047249510884285, + 0.01618541032075882, + 0.014644075185060501, + 0.013732636347413063, + 0.015425272285938263, + 0.03379644826054573, + 0.044734761118888855, + 0.02026299387216568, + 0.021724164485931396, + 0.0064684548415243626, + 0.0035571972839534283, + 0.0032557323575019836, + 0.009140347130596638, + 0.003492242656648159, + 0.0027618384920060635, + 0.0039021673146635294, + 0.009932154789566994, + 0.006679241079837084, + 0.0027534516993910074, + 0.0024443611036986113, + 0.003523743012920022, + 0.0026764143258333206, + 0.0034089938271790743, + 0.00354312127456069, + 0.007898533716797829, + 0.0024159192107617855, + 0.003570767818018794, + 0.0038623728323727846, + 0.002860717475414276, + 0.004274200648069382, + 0.0044091809540987015, + 0.0034444076009094715, + 0.004468407481908798, + 0.007554583251476288, + 0.007742243818938732, + 0.002676625270396471, + 0.005776270758360624, + 0.00674424460157752, + 0.006205902900546789, + 0.007464800029993057, + 0.004979231394827366, + 0.0063920700922608376, + 0.004188171587884426, + 0.008458370342850685, + 0.003381212940439582, + 0.006591809913516045, + 0.004813916515558958, + 0.005889855325222015, + 0.003548732493072748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.41454029083251953 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261783123016357 + ], + "opposite_logprob_tokens": [ + -0.5761783123016357 + ], + "logprob_reward": -0.8261783123016357, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261783123016357, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1784425973892212, + 0.3582254946231842, + 0.0677613839507103, + 0.05610043555498123, + 0.11705388873815536, + 0.06775092333555222, + 0.030102187767624855, + 0.06479255855083466, + 0.041437968611717224, + 0.018332544714212418 + ], + "credit_map": [ + 0.036775145679712296, + 0.01345883309841156, + 0.02231498807668686, + 0.015644975006580353, + 0.013610104098916054, + 0.04578929394483566, + 0.012484321370720863, + 0.01836494728922844, + 0.035254206508398056, + 0.0756782814860344, + 0.07029728591442108, + 0.025970369577407837, + 0.05223948881030083, + 0.03253882750868797, + 0.0270521380007267, + 0.039194900542497635, + 0.008690707385540009, + 0.00575589993968606, + 0.008870592340826988, + 0.00796692818403244, + 0.010534338653087616, + 0.009214168414473534, + 0.006827224977314472, + 0.009901531040668488, + 0.0062913778237998486, + 0.007097756955772638, + 0.004924500826746225, + 0.0068987286649644375, + 0.008391724899411201, + 0.007819908671081066, + 0.007280208170413971, + 0.007396230474114418, + 0.00954236276447773, + 0.008281268179416656, + 0.008341316133737564, + 0.009264770895242691, + 0.01902018114924431, + 0.0358811654150486, + 0.01140038762241602, + 0.015322443097829819, + 0.008148955181241035, + 0.005612902343273163, + 0.002632357180118561, + 0.017597056925296783, + 0.002056732540950179, + 0.0034122669603675604, + 0.0057832710444927216, + 0.022507386282086372, + 0.005637483671307564, + 0.001924731652252376, + 0.002216578461229801, + 0.002030438743531704, + 0.00426401337608695, + 0.0030284724198281765, + 0.003680415451526642, + 0.007320057135075331, + 0.005637214053422213, + 0.01108681969344616, + 0.007737687323242426, + 0.003708411240950227, + 0.006050715688616037, + 0.007092327810823917, + 0.007858781144022942, + 0.015620607882738113, + 0.005799020640552044, + 0.005007434636354446, + 0.00600866274908185, + 0.005334062967449427, + 0.003167761955410242, + 0.0058426628820598125, + 0.004531398415565491, + 0.005746966227889061, + 0.002258728025481105, + 0.002396171912550926, + 0.0015496256528422236, + 0.001644554315134883, + 0.0030990326777100563, + 0.0032276036217808723, + 0.0018711386946961284, + 0.0022856902796775103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.8261783123016357 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005263044033199549 + ], + "opposite_logprob_tokens": [ + -5.255262851715088 + ], + "logprob_reward": -0.005263044033199549, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005263044033199549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3553778827190399, + 0.16216237843036652, + 0.07470591366291046, + 0.0698649138212204, + 0.1562475860118866, + 0.030553214251995087, + 0.018675880506634712, + 0.026278285309672356, + 0.03774145245552063, + 0.06839242577552795 + ], + "credit_map": [ + 0.08170853555202484, + 0.029800228774547577, + 0.04599286615848541, + 0.017261238768696785, + 0.031069576740264893, + 0.10618463903665543, + 0.01584716886281967, + 0.02751363068819046, + 0.016160983592271805, + 0.03318645432591438, + 0.030987536534667015, + 0.012183220125734806, + 0.02287997677922249, + 0.015027438290417194, + 0.01412725169211626, + 0.01760953851044178, + 0.00970721896737814, + 0.006795099005103111, + 0.008256534114480019, + 0.010160214267671108, + 0.012324294075369835, + 0.008187705650925636, + 0.009298741817474365, + 0.00997611042112112, + 0.012106825597584248, + 0.008808799088001251, + 0.0053010196425020695, + 0.008396992459893227, + 0.009048454463481903, + 0.009872522205114365, + 0.007476021535694599, + 0.008854287676513195, + 0.013531936332583427, + 0.011840060353279114, + 0.011949803680181503, + 0.012908623553812504, + 0.02795090153813362, + 0.044117070734500885, + 0.01488841325044632, + 0.019060779362916946, + 0.004409615881741047, + 0.0022672631312161684, + 0.0015058921417221427, + 0.006499592214822769, + 0.0013085617683827877, + 0.001974426908418536, + 0.0028079005423933268, + 0.009779964573681355, + 0.004126041661947966, + 0.0013695459347218275, + 0.0013846444198861718, + 0.0022506292443722486, + 0.0010083508677780628, + 0.0016606292920187116, + 0.0020121452398598194, + 0.0048638940788805485, + 0.002509285695850849, + 0.006048962939530611, + 0.0032332383561879396, + 0.00198656739667058, + 0.0027781883254647255, + 0.0023316543083637953, + 0.0033771186135709286, + 0.00401327246800065, + 0.004655606113374233, + 0.0052101570181548595, + 0.0028101943898946047, + 0.004030417185276747, + 0.003734791651368141, + 0.005133386235684156, + 0.007654384709894657, + 0.004512519109994173, + 0.011620310135185719, + 0.0045471107587218285, + 0.02042498253285885, + 0.007034063804894686, + 0.0046131364069879055, + 0.0064468360505998135, + 0.006552854087203741, + 0.0071531315334141254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005263044033199549 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023305775597691536 + ], + "opposite_logprob_tokens": [ + -3.773305892944336 + ], + "logprob_reward": -0.023305775597691536, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023305775597691536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22232182323932648, + 0.16362515091896057, + 0.067120760679245, + 0.09013155102729797, + 0.1721995621919632, + 0.05012679472565651, + 0.02966594509780407, + 0.04288258031010628, + 0.04590167477726936, + 0.11602415144443512 + ], + "credit_map": [ + 0.053502850234508514, + 0.012888699769973755, + 0.0307016484439373, + 0.008459242060780525, + 0.017470842227339745, + 0.07095615565776825, + 0.010008358396589756, + 0.018334031105041504, + 0.01595938205718994, + 0.035761721432209015, + 0.03186756744980812, + 0.010583617724478245, + 0.024616334587335587, + 0.014968127943575382, + 0.011412289924919605, + 0.01845611445605755, + 0.010391567833721638, + 0.005452380049973726, + 0.008790427818894386, + 0.008007307536900043, + 0.009584778919816017, + 0.008160068653523922, + 0.00657662283629179, + 0.010157602839171886, + 0.013585170730948448, + 0.013097967952489853, + 0.006564533803611994, + 0.010642246343195438, + 0.013093513436615467, + 0.013568767346441746, + 0.008732324466109276, + 0.010847027413547039, + 0.01394310500472784, + 0.013734588399529457, + 0.013946297578513622, + 0.015104281716048717, + 0.03096439130604267, + 0.04595235362648964, + 0.017650701105594635, + 0.020903827622532845, + 0.00625314237549901, + 0.004195914138108492, + 0.001987197669222951, + 0.01238909363746643, + 0.0021963161416351795, + 0.0027216237504035234, + 0.004344403278082609, + 0.016039103269577026, + 0.00483758794143796, + 0.002546162810176611, + 0.002002444351091981, + 0.0025411115493625402, + 0.003078061854466796, + 0.0025648274458944798, + 0.004038122948259115, + 0.008057625964283943, + 0.0036792922765016556, + 0.006722332444041967, + 0.004534244071692228, + 0.0026136357337236404, + 0.00542040541768074, + 0.005838403943926096, + 0.0053169867023825645, + 0.008757280185818672, + 0.004792789928615093, + 0.0050309794023633, + 0.006532822269946337, + 0.004184309393167496, + 0.003414352424442768, + 0.007123385556042194, + 0.0046164123341441154, + 0.01020662672817707, + 0.017989231273531914, + 0.005828145891427994, + 0.03180982917547226, + 0.009957296773791313, + 0.009095236659049988, + 0.012607401236891747, + 0.012902420945465565, + 0.015834588557481766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.023305775597691536 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464135454967618 + ], + "opposite_logprob_tokens": [ + -5.379641532897949 + ], + "logprob_reward": -0.00464135454967618, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464135454967618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32278814911842346, + 0.18077315390110016, + 0.04099798575043678, + 0.10700971633195877, + 0.1591036319732666, + 0.033578578382730484, + 0.0326724536716938, + 0.032721951603889465, + 0.043184611946344376, + 0.047169748693704605 + ], + "credit_map": [ + 0.07914827018976212, + 0.018749767914414406, + 0.044176749885082245, + 0.012620997615158558, + 0.025221627205610275, + 0.10216417908668518, + 0.014023242518305779, + 0.026683293282985687, + 0.0181539636105299, + 0.036938607692718506, + 0.032767944037914276, + 0.012822535820305347, + 0.024477709084749222, + 0.016261670738458633, + 0.020521117374300957, + 0.01882961578667164, + 0.006284714676439762, + 0.003612212371081114, + 0.004483926575630903, + 0.006649685557931662, + 0.006335485726594925, + 0.005261850077658892, + 0.0034364010207355022, + 0.004933707416057587, + 0.014128830283880234, + 0.011354007758200169, + 0.010152765549719334, + 0.0107488501816988, + 0.013649542815983295, + 0.01627601683139801, + 0.013851762749254704, + 0.016847945749759674, + 0.013885518535971642, + 0.01211373321712017, + 0.01436457596719265, + 0.013683821074664593, + 0.026599792763590813, + 0.04483748599886894, + 0.014160308055579662, + 0.019458400085568428, + 0.005214132834225893, + 0.002637410070747137, + 0.0019174533663317561, + 0.007042498327791691, + 0.0013873466523364186, + 0.0022341767325997353, + 0.003114276332780719, + 0.01003128569573164, + 0.0073607126250863075, + 0.0018587965751066804, + 0.0025814110413193703, + 0.003216375131160021, + 0.0020544470753520727, + 0.0019856898579746485, + 0.003113760845735669, + 0.010501259937882423, + 0.00264764204621315, + 0.0067978003062307835, + 0.004214331973344088, + 0.0024606401566416025, + 0.003816046053543687, + 0.003951664548367262, + 0.0038694338873028755, + 0.004964390303939581, + 0.005802865140140057, + 0.004733263049274683, + 0.005631820764392614, + 0.004123940598219633, + 0.0030779556836932898, + 0.006118156481534243, + 0.004291168879717588, + 0.009405439719557762, + 0.006651623640209436, + 0.0037769225891679525, + 0.010607345961034298, + 0.004209175240248442, + 0.005125156603753567, + 0.004997489973902702, + 0.007428107783198357, + 0.0043739257380366325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.00464135454967618 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot, executed in stunning slow motion, follows the explosive movements of a male tennis player in his early twenties. Under the blazing afternoon sun, he pushes off the blue professional hard court with all his might, launching himself into the air. His body stretches fully in mid-air, forming a powerful bow. His short, sweat-soaked dark brown hair clings to his forehead, a single bead of sweat trickles down between his furrowed brows, his facial muscles tensed with intense concentration. He wears a bright yellow sleeveless tank top that clings to his defined muscles, dark blue athletic shorts, and a white sweatband around his wrist. At the apex of his rise, his powerful core drives his waist in a violent twist, his right arm raised high, and the carbon fiber tennis racket in his hand carves a swift arc through the air. At the highest point of the shot, the strings of the racket deformed visibly, precisely and fiercely enveloping the incoming fluorescent yellow tennis ball. Instantly, like a cannonball compressed to the limit and released, the ball ripped through the air with a whistling sound as it traveled in a nearly indiscernible line, blasting towards the other side of the court with lightning speed. Sunlight reflected off his sweat-soaked skin and the metal frame of the racket, creating dazzling spots of light. Against the backdrop of the slightly out-of-focus green netting, the entire scene was filled with the atmosphere of ultimate power, speed, and focused competition, perfectly capturing the decisive moment of this fast-paced exchange.", + "seed": 27787, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_15/index_1624_seed_27787_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06712312391027808, + "vlm_logprob": -0.06712312391027808, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.06712312391027808, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22490239143371582 + ], + "opposite_logprob_tokens": [ + -1.8499023914337158 + ], + "logprob_reward": -0.22490239143371582, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.22490239143371582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18336571753025055, + 0.06274238973855972, + 0.0699857696890831, + 0.02818036451935768, + 0.05562661588191986, + 0.10212317109107971, + 0.044918689876794815, + 0.04164993762969971, + 0.09644218534231186, + 0.314965158700943 + ], + "credit_map": [ + 0.014302441850304604, + 0.04571712762117386, + 0.03220739960670471, + 0.022912945598363876, + 0.015473151579499245, + 0.02038351446390152, + 0.01487911306321621, + 0.017490027472376823, + 0.006792705040425062, + 0.008820350281894207, + 0.007981914095580578, + 0.0072545818984508514, + 0.013081434182822704, + 0.00844859890639782, + 0.004812507890164852, + 0.005550298374146223, + 0.011260933242738247, + 0.011165796779096127, + 0.010563621297478676, + 0.007182381581515074, + 0.011678037233650684, + 0.007043260615319014, + 0.005498336162418127, + 0.005593403708189726, + 0.0040296586230397224, + 0.004367432091385126, + 0.00514054112136364, + 0.0027528482023626566, + 0.0019084281520918012, + 0.003144346410408616, + 0.0031116274185478687, + 0.0037254823837429285, + 0.008508960716426373, + 0.004399619530886412, + 0.009813801385462284, + 0.003102873219177127, + 0.00586265604943037, + 0.007666157092899084, + 0.005258782301098108, + 0.011013764888048172, + 0.012408438138663769, + 0.012023347429931164, + 0.01451027300208807, + 0.013801023364067078, + 0.009249859489500523, + 0.014265093952417374, + 0.013409871608018875, + 0.012455265037715435, + 0.0036656290758401155, + 0.004748064558953047, + 0.004671843722462654, + 0.0038701053708791733, + 0.012265843339264393, + 0.00465123075991869, + 0.006482475437223911, + 0.004563496448099613, + 0.006824412848800421, + 0.0042473021894693375, + 0.005508061032742262, + 0.00512687535956502, + 0.007193237543106079, + 0.004200010094791651, + 0.004298726096749306, + 0.004251312464475632, + 0.01558692380785942, + 0.01948532648384571, + 0.005565033759921789, + 0.006332544609904289, + 0.00994089338928461, + 0.022655248641967773, + 0.007956108078360558, + 0.008920112624764442, + 0.020255999639630318, + 0.02482023648917675, + 0.05375460535287857, + 0.02250833809375763, + 0.014413135126233101, + 0.09492931514978409, + 0.030931102111935616, + 0.05335244536399841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.22490239143371582 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029836505651474 + ], + "opposite_logprob_tokens": [ + -3.529836416244507 + ], + "logprob_reward": -0.029836505651474, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029836505651474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20193906128406525, + 0.18549101054668427, + 0.2145925909280777, + 0.023369261994957924, + 0.041847847402095795, + 0.05042210593819618, + 0.01942211017012596, + 0.06069079414010048, + 0.05345962196588516, + 0.14876560866832733 + ], + "credit_map": [ + 0.020028026774525642, + 0.045389324426651, + 0.052039388567209244, + 0.02659377083182335, + 0.011603359133005142, + 0.017876695841550827, + 0.015993114560842514, + 0.012415376491844654, + 0.020913951098918915, + 0.06119747459888458, + 0.03402744233608246, + 0.020050257444381714, + 0.027740102261304855, + 0.007785674184560776, + 0.007234552409499884, + 0.006541541311889887, + 0.027791233733296394, + 0.03644376993179321, + 0.02765783667564392, + 0.02494974061846733, + 0.05260225757956505, + 0.017039518803358078, + 0.013808667659759521, + 0.014299576170742512, + 0.0036379913799464703, + 0.003508081194013357, + 0.005127717275172472, + 0.0020457145292311907, + 0.0020190435461699963, + 0.0028812114614993334, + 0.001915230299346149, + 0.002234270330518484, + 0.006831964943557978, + 0.0024305216502398252, + 0.012288405559957027, + 0.0028870643582195044, + 0.002978527219966054, + 0.004070429131388664, + 0.0020443371031433344, + 0.00831659883260727, + 0.004887423012405634, + 0.007497005630284548, + 0.0075544919818639755, + 0.006745931692421436, + 0.004471553023904562, + 0.008310618810355663, + 0.004944544285535812, + 0.006010537967085838, + 0.001758372993208468, + 0.0019019050523638725, + 0.0020083964336663485, + 0.001804009429179132, + 0.005391683895140886, + 0.0016891666455194354, + 0.0026994775980710983, + 0.002169098239392042, + 0.012446951121091843, + 0.005286389496177435, + 0.0063123456202447414, + 0.007528866175562143, + 0.008686025626957417, + 0.007712059188634157, + 0.006773215252906084, + 0.0059449393302202225, + 0.0105406753718853, + 0.01017457153648138, + 0.008076977916061878, + 0.0055109066888689995, + 0.006165440659970045, + 0.0053226831369102, + 0.004105874802917242, + 0.0035624909214675426, + 0.009931087493896484, + 0.01426292210817337, + 0.024795664474368095, + 0.011115076020359993, + 0.0062744757160544395, + 0.04393943026661873, + 0.013467312790453434, + 0.024979647248983383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.029836505651474 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006750516593456268 + ], + "opposite_logprob_tokens": [ + -5.006750583648682 + ], + "logprob_reward": -0.006750516593456268, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006750516593456268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1486756056547165, + 0.11691910773515701, + 0.13142569363117218, + 0.03232448175549507, + 0.10381925851106644, + 0.08301318436861038, + 0.03033226914703846, + 0.13610778748989105, + 0.08360560983419418, + 0.13377702236175537 + ], + "credit_map": [ + 0.01161409355700016, + 0.03311369568109512, + 0.029844161123037338, + 0.014502912759780884, + 0.013643085025250912, + 0.019065052270889282, + 0.015429219231009483, + 0.011463383212685585, + 0.012205557897686958, + 0.02759101800620556, + 0.025245126336812973, + 0.01586293801665306, + 0.01843155547976494, + 0.007043308578431606, + 0.0050630392506718636, + 0.005476559977978468, + 0.015266832895576954, + 0.01578068919479847, + 0.01202795747667551, + 0.014436915516853333, + 0.019675083458423615, + 0.013295064680278301, + 0.01932414621114731, + 0.02161899209022522, + 0.0068163503892719746, + 0.004821924492716789, + 0.005676299799233675, + 0.0032070695888251066, + 0.0025318898260593414, + 0.0031959540210664272, + 0.0024254617746919394, + 0.003649534657597542, + 0.01367669552564621, + 0.0122257424518466, + 0.012287989258766174, + 0.008154544979333878, + 0.014708690345287323, + 0.014650914818048477, + 0.012944135814905167, + 0.015170549973845482, + 0.01167711429297924, + 0.013181998394429684, + 0.010832545347511768, + 0.015131024643778801, + 0.006916369777172804, + 0.008564256131649017, + 0.007859292440116405, + 0.0088505819439888, + 0.00390279246494174, + 0.0026076417416334152, + 0.005147644784301519, + 0.004530369769781828, + 0.003426504787057638, + 0.00400739349424839, + 0.002711727749556303, + 0.0039981952868402, + 0.026773663237690926, + 0.011899962089955807, + 0.014799791388213634, + 0.012438120320439339, + 0.021581048145890236, + 0.0188915878534317, + 0.0143407778814435, + 0.015382845886051655, + 0.012017176486551762, + 0.01330078486353159, + 0.007291482295840979, + 0.008875104598701, + 0.010881168767809868, + 0.014419826678931713, + 0.009492238983511925, + 0.007327824831008911, + 0.010651941411197186, + 0.020076638087630272, + 0.020132724195718765, + 0.013196964748203754, + 0.006692000199109316, + 0.02939779870212078, + 0.016525989398360252, + 0.017102960497140884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.006750516593456268 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002904266119003 + ], + "opposite_logprob_tokens": [ + -2.6950290203094482 + ], + "logprob_reward": -0.07002904266119003, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002904266119003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1616440713405609, + 0.05974429473280907, + 0.14309614896774292, + 0.022399067878723145, + 0.034966446459293365, + 0.07570123672485352, + 0.11080717295408249, + 0.029593396931886673, + 0.13284307718276978, + 0.2292051762342453 + ], + "credit_map": [ + 0.013017433695495129, + 0.04105209559202194, + 0.033362582325935364, + 0.017727982252836227, + 0.009304540231823921, + 0.0168398879468441, + 0.01211378164589405, + 0.018225759267807007, + 0.006721131969243288, + 0.018825387582182884, + 0.007746485993266106, + 0.006451370194554329, + 0.00846776831895113, + 0.004517585039138794, + 0.0033219135366380215, + 0.003692647209390998, + 0.02436983771622181, + 0.02238745428621769, + 0.014365799725055695, + 0.014648271724581718, + 0.03380122035741806, + 0.01396644301712513, + 0.009987184777855873, + 0.009569925256073475, + 0.0031486395746469498, + 0.002997437259182334, + 0.0036528492346405983, + 0.001955569488927722, + 0.0016678462270647287, + 0.003759990446269512, + 0.0026538954116404057, + 0.0025628386065363884, + 0.005082248710095882, + 0.0026720750611275434, + 0.006203277036547661, + 0.002280549146234989, + 0.004673076793551445, + 0.006524963304400444, + 0.003221557242795825, + 0.004308693576604128, + 0.007914457470178604, + 0.007470262702554464, + 0.009785733185708523, + 0.0083943335339427, + 0.007808220572769642, + 0.010482395067811012, + 0.0123705780133605, + 0.011475248262286186, + 0.011377465911209583, + 0.010873453691601753, + 0.019344890490174294, + 0.014895608648657799, + 0.020061854273080826, + 0.011898775584995747, + 0.011917046271264553, + 0.010438062250614166, + 0.0037333881482481956, + 0.0024541527964174747, + 0.002860279520973563, + 0.002701142570003867, + 0.006873481906950474, + 0.0028104938101023436, + 0.003249348606914282, + 0.004911104217171669, + 0.01844961568713188, + 0.021999457851052284, + 0.016064932569861412, + 0.0171966589987278, + 0.011012719944119453, + 0.022306188941001892, + 0.01409382838755846, + 0.011719662696123123, + 0.015145969577133656, + 0.015000615268945694, + 0.03850327432155609, + 0.021253548562526703, + 0.010842055082321167, + 0.06429760158061981, + 0.026288088411092758, + 0.03787398710846901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.07002904266119003 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097163211554289 + ], + "opposite_logprob_tokens": [ + -5.504096984863281 + ], + "logprob_reward": -0.004097163211554289, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097163211554289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2872287929058075, + 0.225203737616539, + 0.047997090965509415, + 0.0729350671172142, + 0.04281463101506233, + 0.035010967403650284, + 0.043587058782577515, + 0.0373370423913002, + 0.08099039644002914, + 0.12689518928527832 + ], + "credit_map": [ + 0.02459406666457653, + 0.06430631130933762, + 0.07669662684202194, + 0.031024551019072533, + 0.020165013149380684, + 0.03134417533874512, + 0.021046381443738937, + 0.018051672726869583, + 0.02521161176264286, + 0.06330841034650803, + 0.036780454218387604, + 0.028685322031378746, + 0.03666388988494873, + 0.01176530309021473, + 0.013641043566167355, + 0.009147713892161846, + 0.005731275770813227, + 0.008237471804022789, + 0.004731729160994291, + 0.004853546619415283, + 0.0071036322042346, + 0.006362444721162319, + 0.004834473133087158, + 0.0061425198800861835, + 0.012079564854502678, + 0.012838562950491905, + 0.020383376628160477, + 0.008471673354506493, + 0.004559946712106466, + 0.006143635604530573, + 0.0037513829302042723, + 0.00470692990347743, + 0.006784765049815178, + 0.0024258324410766363, + 0.011711652390658855, + 0.0029485130216926336, + 0.004324380308389664, + 0.00435193907469511, + 0.004430008586496115, + 0.005837547592818737, + 0.004010665230453014, + 0.0028053238056600094, + 0.0055382163263857365, + 0.0030215440783649683, + 0.0031252617482095957, + 0.006608694326132536, + 0.004255248233675957, + 0.0056460159830749035, + 0.004809313919395208, + 0.004906557500362396, + 0.006026663817465305, + 0.0054356916807591915, + 0.006350751966238022, + 0.006489037070423365, + 0.00539820222184062, + 0.004170843865722418, + 0.007209242787212133, + 0.0035188114270567894, + 0.006066902074962854, + 0.003972664009779692, + 0.00427903700619936, + 0.004450675565749407, + 0.003349423874169588, + 0.00449028704315424, + 0.014810072258114815, + 0.013669951818883419, + 0.01137335691601038, + 0.010288954712450504, + 0.00928320363163948, + 0.009097963571548462, + 0.006502929143607616, + 0.005963969510048628, + 0.012179287150502205, + 0.025811797007918358, + 0.0166793130338192, + 0.014661234803497791, + 0.007963395677506924, + 0.02093356102705002, + 0.015851978212594986, + 0.012814625166356564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.004097163211554289 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot, executed in stunning slow motion, follows the explosive movements of a male tennis player in his early twenties. Under the blazing afternoon sun, he pushes off the blue professional hard court with all his might, launching himself into the air. His body stretches fully in mid-air, forming a powerful bow. His short, sweat-soaked dark brown hair clings to his forehead, a single bead of sweat trickles down between his furrowed brows, his facial muscles tensed with intense concentration. He wears a bright yellow sleeveless tank top that clings to his defined muscles, dark blue athletic shorts, and a white sweatband around his wrist. At the apex of his rise, his powerful core drives his waist in a violent twist, his right arm raised high, and the carbon fiber tennis racket in his hand carves a swift arc through the air. At the highest point of the shot, the strings of the racket deformed visibly, precisely and fiercely enveloping the incoming fluorescent yellow tennis ball. Instantly, like a cannonball compressed to the limit and released, the ball ripped through the air with a whistling sound as it traveled in a nearly indiscernible line, blasting towards the other side of the court with lightning speed. Sunlight reflected off his sweat-soaked skin and the metal frame of the racket, creating dazzling spots of light. Against the backdrop of the slightly out-of-focus green netting, the entire scene was filled with the atmosphere of ultimate power, speed, and focused competition, perfectly capturing the decisive moment of this fast-paced exchange.", + "seed": 27788, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_15/index_1624_seed_27788_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09512103870511054, + "vlm_logprob": -0.09512103870511054, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.09512103870511054, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35264039039611816 + ], + "opposite_logprob_tokens": [ + -1.7276403903961182 + ], + "logprob_reward": -0.35264039039611816, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.35264039039611816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1918499618768692, + 0.14627495408058167, + 0.05312709882855415, + 0.06212297081947327, + 0.16598458588123322, + 0.1299578994512558, + 0.07076700031757355, + 0.03954434022307396, + 0.061412710696458817, + 0.07895847409963608 + ], + "credit_map": [ + 0.04582199454307556, + 0.031122466549277306, + 0.032542940229177475, + 0.01764608547091484, + 0.01945151761174202, + 0.01742026023566723, + 0.016353586688637733, + 0.011491117998957634, + 0.01432130765169859, + 0.021647948771715164, + 0.01765609160065651, + 0.017606131732463837, + 0.027358610183000565, + 0.01928698644042015, + 0.013796379789710045, + 0.014601503498852253, + 0.004836278967559338, + 0.005205649416893721, + 0.006846785545349121, + 0.006762597244232893, + 0.006295124068856239, + 0.007726693991571665, + 0.007413364481180906, + 0.008040603250265121, + 0.004601914901286364, + 0.005497446749359369, + 0.007068745791912079, + 0.006352843716740608, + 0.007518537808209658, + 0.015237977728247643, + 0.007887272164225578, + 0.00795823149383068, + 0.0321173369884491, + 0.012644234113395214, + 0.009081641212105751, + 0.010304650291800499, + 0.0499773845076561, + 0.029273511841893196, + 0.011066263541579247, + 0.011519568040966988, + 0.007282470818608999, + 0.018789587542414665, + 0.017872825264930725, + 0.014019140042364597, + 0.015400657430291176, + 0.01624002866446972, + 0.01950637623667717, + 0.02084682323038578, + 0.0065477085299789906, + 0.009830093942582607, + 0.007674986496567726, + 0.008922426030039787, + 0.008797003887593746, + 0.008682253770530224, + 0.009955584071576595, + 0.010356943123042583, + 0.0066154287196695805, + 0.003554317634552717, + 0.00568330567330122, + 0.005731787998229265, + 0.0038166758604347706, + 0.0059323785826563835, + 0.003244286635890603, + 0.004966157488524914, + 0.005289637017995119, + 0.00799278263002634, + 0.004835926461964846, + 0.004590763710439205, + 0.014255961403250694, + 0.015412506647408009, + 0.004134668968617916, + 0.004900461994111538, + 0.009154044091701508, + 0.009288376197218895, + 0.016451988369226456, + 0.008624683134257793, + 0.013015097007155418, + 0.009303244762122631, + 0.00698167085647583, + 0.006139368284493685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.35264039039611816 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907217532396317 + ], + "opposite_logprob_tokens": [ + -2.4640722274780273 + ], + "logprob_reward": -0.08907217532396317, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907217532396317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23709426820278168, + 0.044706203043460846, + 0.06260629743337631, + 0.07835258543491364, + 0.03468232601881027, + 0.10975093394517899, + 0.05574803426861763, + 0.11685948818922043, + 0.08590784668922424, + 0.17429205775260925 + ], + "credit_map": [ + 0.03179315850138664, + 0.0344463475048542, + 0.05396914854645729, + 0.03234254568815231, + 0.023701602593064308, + 0.031671442091464996, + 0.01624881848692894, + 0.012921176850795746, + 0.0060511380434036255, + 0.005781006999313831, + 0.006301510613411665, + 0.006642043124884367, + 0.005890188738703728, + 0.004794623702764511, + 0.004414061550050974, + 0.004831625148653984, + 0.0055486299097537994, + 0.007558252662420273, + 0.00662551075220108, + 0.009596988558769226, + 0.009467662312090397, + 0.007394897285848856, + 0.00791215430945158, + 0.00850219838321209, + 0.006299477070569992, + 0.007906660437583923, + 0.010849008336663246, + 0.007839102298021317, + 0.010647646151483059, + 0.009691770188510418, + 0.010746641084551811, + 0.014372270554304123, + 0.005424866918474436, + 0.005942292511463165, + 0.0021890343632549047, + 0.0027062955778092146, + 0.007321492303162813, + 0.005956034176051617, + 0.0024290347937494516, + 0.0027132751420140266, + 0.009279459714889526, + 0.016976872459053993, + 0.012716978788375854, + 0.012252317741513252, + 0.01505710557103157, + 0.01673867739737034, + 0.012107647024095058, + 0.014621867798268795, + 0.005151249933987856, + 0.007566698826849461, + 0.006388860754668713, + 0.010338271968066692, + 0.006125325337052345, + 0.00550976675003767, + 0.008111600764095783, + 0.006556251086294651, + 0.018268991261720657, + 0.010785091668367386, + 0.01810503378510475, + 0.014719943515956402, + 0.017212580889463425, + 0.015857994556427002, + 0.009564016945660114, + 0.012345821596682072, + 0.007893660105764866, + 0.009296756237745285, + 0.007975662127137184, + 0.006917738821357489, + 0.021124618127942085, + 0.014820446260273457, + 0.007147819269448519, + 0.01073113176971674, + 0.020631011575460434, + 0.025126881897449493, + 0.025366414338350296, + 0.015380003489553928, + 0.03187284618616104, + 0.025011489167809486, + 0.012026202864944935, + 0.018877195194363594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.08907217532396317 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008656949736177921 + ], + "opposite_logprob_tokens": [ + -4.758656978607178 + ], + "logprob_reward": -0.008656949736177921, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008656949736177921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1875787079334259, + 0.044438574463129044, + 0.331379771232605, + 0.05159517005085945, + 0.12194562703371048, + 0.06401688605546951, + 0.04996844753623009, + 0.08404913544654846, + 0.03228935971856117, + 0.03273836523294449 + ], + "credit_map": [ + 0.026926705613732338, + 0.0329623818397522, + 0.023298626765608788, + 0.02683921530842781, + 0.022775283083319664, + 0.019258763641119003, + 0.015942184254527092, + 0.01957552321255207, + 0.005682781804352999, + 0.006078994367271662, + 0.006774237845093012, + 0.008156931027770042, + 0.004184393677860498, + 0.004662101622670889, + 0.004510214086622, + 0.004388916306197643, + 0.04495808482170105, + 0.06868881732225418, + 0.04424983263015747, + 0.04438439756631851, + 0.032480984926223755, + 0.02307078056037426, + 0.033855605870485306, + 0.03969123959541321, + 0.0043558175675570965, + 0.005913798231631517, + 0.008108614012598991, + 0.003920954652130604, + 0.008621659129858017, + 0.009024638682603836, + 0.005630702245980501, + 0.0060189771465957165, + 0.02803053706884384, + 0.0063937027007341385, + 0.007165322080254555, + 0.0062845065258443356, + 0.042251698672771454, + 0.017999129369854927, + 0.006022865884006023, + 0.007797847036272287, + 0.005350867286324501, + 0.010191384702920914, + 0.006837581750005484, + 0.007826274260878563, + 0.008814110420644283, + 0.008533610962331295, + 0.007704001851379871, + 0.00875905342400074, + 0.005051780957728624, + 0.007654823828488588, + 0.00522831454873085, + 0.005922218784689903, + 0.00579087296500802, + 0.006740055978298187, + 0.006978047080338001, + 0.00660233199596405, + 0.014494352973997593, + 0.009279508143663406, + 0.016564493998885155, + 0.009764605201780796, + 0.009618930518627167, + 0.010433112271130085, + 0.006877982523292303, + 0.007016142830252647, + 0.0028163527604192495, + 0.004450674634426832, + 0.0027488362975418568, + 0.002764202654361725, + 0.006511340383440256, + 0.007502772845327854, + 0.002033467637374997, + 0.0034617078490555286, + 0.003379896515980363, + 0.003327188780531287, + 0.005170444492250681, + 0.003856946248561144, + 0.005214996170252562, + 0.005322175566107035, + 0.0034891506657004356, + 0.0029775649309158325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.008656949736177921 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059515379369259 + ], + "opposite_logprob_tokens": [ + -3.895595073699951 + ], + "logprob_reward": -0.02059515379369259, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059515379369259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20452404022216797, + 0.11059319972991943, + 0.058830082416534424, + 0.16452829539775848, + 0.19755391776561737, + 0.07586275786161423, + 0.04382514953613281, + 0.05067460238933563, + 0.054368745535612106, + 0.039239224046468735 + ], + "credit_map": [ + 0.07080753147602081, + 0.02750973403453827, + 0.053171105682849884, + 0.011185858398675919, + 0.009556529112160206, + 0.013406657613813877, + 0.010128170251846313, + 0.008758454583585262, + 0.010750144720077515, + 0.017899667844176292, + 0.014452807605266571, + 0.017038926482200623, + 0.019342826679348946, + 0.010906841605901718, + 0.01124616339802742, + 0.008955825120210648, + 0.010318007320165634, + 0.006128956098109484, + 0.010074813850224018, + 0.005969546269625425, + 0.006750271189957857, + 0.006714818067848682, + 0.007360571995377541, + 0.005513098556548357, + 0.009383687749505043, + 0.015123499557375908, + 0.014387868344783783, + 0.017637018114328384, + 0.02411726303398609, + 0.03550083935260773, + 0.022859660908579826, + 0.02551846392452717, + 0.04291568696498871, + 0.013807512819766998, + 0.008889716118574142, + 0.009920417331159115, + 0.06623661518096924, + 0.03224925324320793, + 0.01056971400976181, + 0.01296499464660883, + 0.005911662243306637, + 0.011719241738319397, + 0.007968666031956673, + 0.010907705873250961, + 0.011638102121651173, + 0.010889334604144096, + 0.009348977357149124, + 0.007479068823158741, + 0.004474171902984381, + 0.005748056340962648, + 0.004821188747882843, + 0.005799004342406988, + 0.005436193197965622, + 0.004950761795043945, + 0.006302897352725267, + 0.006292873062193394, + 0.007428750861436129, + 0.005098822060972452, + 0.0056999800726771355, + 0.005948726087808609, + 0.006562346126884222, + 0.007746163755655289, + 0.004998252261430025, + 0.007191563490778208, + 0.00483912555500865, + 0.009003909304738045, + 0.003958131652325392, + 0.0046173217706382275, + 0.010887929238379002, + 0.012851577252149582, + 0.0036448303144425154, + 0.004565921612083912, + 0.006550819147378206, + 0.0040269289165735245, + 0.0062299249693751335, + 0.004736706148833036, + 0.005351840052753687, + 0.005046098958700895, + 0.0038249362260103226, + 0.003471967065706849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.02059515379369259 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a blue professional hard court under a bright afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male tennis player wearing a bright yellow sleeveless tank top and dark blue athletic shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the player launch himself into the air to hit a fluorescent yellow tennis ball with a carbon fiber racket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background green netting visible and slightly out of focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464052427560091 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.00464052427560091, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464052427560091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19528312981128693, + 0.09140719473361969, + 0.06289384514093399, + 0.22054535150527954, + 0.16385507583618164, + 0.06916001439094543, + 0.02722778730094433, + 0.05407747998833656, + 0.040715835988521576, + 0.07483427971601486 + ], + "credit_map": [ + 0.057470615953207016, + 0.02795296162366867, + 0.04062535986304283, + 0.012658779509365559, + 0.01616022363305092, + 0.015895146876573563, + 0.01226995512843132, + 0.012250094674527645, + 0.009880461730062962, + 0.011480504646897316, + 0.011083508841693401, + 0.00815783254802227, + 0.013886352069675922, + 0.015536555089056492, + 0.007895302027463913, + 0.013486681506037712, + 0.006797231733798981, + 0.00917555671185255, + 0.0060827224515378475, + 0.007734937127679586, + 0.01108784694224596, + 0.007689555175602436, + 0.00764907943084836, + 0.006676917430013418, + 0.012089822441339493, + 0.010381046682596207, + 0.015054390765726566, + 0.03715214878320694, + 0.011843517422676086, + 0.03773748502135277, + 0.0420965738594532, + 0.054190367460250854, + 0.03366364911198616, + 0.009300011210143566, + 0.008657814003527164, + 0.008781682699918747, + 0.05185569077730179, + 0.02948570065200329, + 0.010089332237839699, + 0.012021197937428951, + 0.00684254989027977, + 0.009153616614639759, + 0.009369042702019215, + 0.007820994593203068, + 0.008335581049323082, + 0.008249123580753803, + 0.010225949808955193, + 0.00916315894573927, + 0.003000359982252121, + 0.003290451131761074, + 0.0034127170220017433, + 0.004004120361059904, + 0.0028402642346918583, + 0.0034517450258135796, + 0.003384412033483386, + 0.003843716811388731, + 0.009437719359993935, + 0.008206268772482872, + 0.0063622696325182915, + 0.007276637014001608, + 0.007845696993172169, + 0.007733339909464121, + 0.0036805456038564444, + 0.0035349989775568247, + 0.0029202597215771675, + 0.004218445625156164, + 0.0032833574805408716, + 0.003398416331037879, + 0.010183372534811497, + 0.006429532542824745, + 0.004589039832353592, + 0.005693412851542234, + 0.006533412728458643, + 0.009460628032684326, + 0.007869821973145008, + 0.008672414347529411, + 0.012635341845452785, + 0.01446713600307703, + 0.007864228449761868, + 0.0073312860913574696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.00464052427560091 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_2/index_tensor([9617, 9617])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_2/index_tensor([9617, 9617])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d66c0ef857818295091bc7d764d41af732fd8cd2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_2/index_tensor([9617, 9617])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-altitude drone shot captures a breathtaking volcanic eruption with a calm and steady perspective, rendered in a striking black-and-white cinematic quality. A vast, crater dominates the frame, its jagged outline casting a deep shadow across the charred landscape. From its depths, luminous, white-hot liquid magma erupts violently. Simultaneously, a towering, dense column of ash and gas rises like a behemoth, its interior churning and expanding, creating distinct layers of gray that obscure the entire sky. Brilliant rivers of molten lava flow in stark contrast, flowing along the fiery mountainside in a stark contrast of light and dark, creating a stark contrast against the surrounding dark rock. The scene is imbued with a sense of primal, violent power, capturing, in silent black and white, a magnificent natural spectacle both awe-inspiring and devastating.", + "seed": 19312, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_2/index_9617_seed_19312_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14222220377996564, + "vlm_logprob": -0.14222220377996564, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.14222220377996564, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21896153688430786 + ], + "opposite_logprob_tokens": [ + -1.843961477279663 + ], + "logprob_reward": -0.21896153688430786, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.21896153688430786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05266131833195686, + 0.1457328498363495, + 0.022261392325162888, + 0.024225302040576935, + 0.6416175365447998, + 0.02510293386876583, + 0.04822409525513649, + 0.019401952624320984, + 0.010692272335290909, + 0.010080337524414062 + ], + "credit_map": [ + 0.005147464107722044, + 0.010565856471657753, + 0.004344493616372347, + 0.006082987878471613, + 0.006572249811142683, + 0.007308169733732939, + 0.005537449382245541, + 0.007102648261934519, + 0.017485789954662323, + 0.013245523907244205, + 0.008783997967839241, + 0.009372962638735771, + 0.011151897720992565, + 0.029022255912423134, + 0.027388865128159523, + 0.029281560331583023, + 0.002166683552786708, + 0.0026919818483293056, + 0.0015884169843047857, + 0.002481797942891717, + 0.0030151570681482553, + 0.0033565794583410025, + 0.0036343794781714678, + 0.0033263959921896458, + 0.0024000690318644047, + 0.002289612079039216, + 0.003207538276910782, + 0.003909015096724033, + 0.002645323984324932, + 0.002745037665590644, + 0.0036646558437496424, + 0.003364051692187786, + 0.0312025286257267, + 0.047932419925928116, + 0.03320158272981644, + 0.019332339987158775, + 0.23463280498981476, + 0.05531596392393112, + 0.029720785096287727, + 0.19027911126613617, + 0.0028450184036046267, + 0.0020992474164813757, + 0.0024976443964987993, + 0.0026237210258841515, + 0.00411075446754694, + 0.0031315565574914217, + 0.0035651458892971277, + 0.004229845013469458, + 0.004054625052958727, + 0.007088427431881428, + 0.004866280592978001, + 0.005662496201694012, + 0.0041787289083004, + 0.006938066333532333, + 0.00868111103773117, + 0.006754360627382994, + 0.0017168798949569464, + 0.002106126630678773, + 0.001874792855232954, + 0.0026441183872520924, + 0.001558877294883132, + 0.0035324173513799906, + 0.003219767240807414, + 0.0027489729691296816, + 0.0015585363144055009, + 0.0009300477104261518, + 0.0011518436949700117, + 0.0014166246401146054, + 0.0008269543759524822, + 0.0015807844465598464, + 0.0018285508267581463, + 0.0013989296276122332, + 0.0016571705928072333, + 0.0010131760500371456, + 0.0012537379516288638, + 0.0011395879555493593, + 0.0011891108006238937, + 0.0015430364292114973, + 0.0012771575711667538, + 0.0010073598241433501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.21896153688430786 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20168502628803253 + ], + "opposite_logprob_tokens": [ + -1.7016850709915161 + ], + "logprob_reward": -0.20168502628803253, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20168502628803253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10140618681907654, + 0.03860582783818245, + 0.05454583838582039, + 0.027233656495809555, + 0.3921990990638733, + 0.1617889255285263, + 0.07381951808929443, + 0.09674173593521118, + 0.026851307600736618, + 0.026807978749275208 + ], + "credit_map": [ + 0.018688177689909935, + 0.011218877509236336, + 0.010948735289275646, + 0.010097826831042767, + 0.010444505140185356, + 0.014461055397987366, + 0.010565525852143764, + 0.014981473796069622, + 0.004835893865674734, + 0.00401717284694314, + 0.0034566011745482683, + 0.004352159798145294, + 0.0030228805262595415, + 0.005615709815174341, + 0.007371891289949417, + 0.005933512467890978, + 0.00786631926894188, + 0.005714367143809795, + 0.004570922814309597, + 0.006521876901388168, + 0.008053702302277088, + 0.006048400420695543, + 0.007738345768302679, + 0.008031902834773064, + 0.003219891106709838, + 0.003085911273956299, + 0.002164689125493169, + 0.003184805391356349, + 0.005553069524466991, + 0.0034362822771072388, + 0.003177050268277526, + 0.003411954501643777, + 0.019062478095293045, + 0.028633859008550644, + 0.020165950059890747, + 0.011604518629610538, + 0.14381936192512512, + 0.033331625163555145, + 0.01794688031077385, + 0.1176343783736229, + 0.011395097710192204, + 0.013949944637715816, + 0.006787457969039679, + 0.00596963707357645, + 0.056826334446668625, + 0.013366680592298508, + 0.010473485104739666, + 0.04302027449011803, + 0.007953218184411526, + 0.008412167429924011, + 0.007207443471997976, + 0.0102434903383255, + 0.006518550682812929, + 0.014489256776869297, + 0.00910118781030178, + 0.009894197806715965, + 0.013862517662346363, + 0.013530653901398182, + 0.009171412326395512, + 0.01011795736849308, + 0.008240153081715107, + 0.01536018680781126, + 0.01416822336614132, + 0.012290619313716888, + 0.003184671513736248, + 0.004484531003981829, + 0.0017365346429869533, + 0.002193629974499345, + 0.003070673206821084, + 0.006199981551617384, + 0.0028542487416416407, + 0.003127035219222307, + 0.002868332900106907, + 0.00297218794003129, + 0.0033069374039769173, + 0.0032468819990754128, + 0.004215696826577187, + 0.004615228157490492, + 0.0024893044028431177, + 0.0030934084206819534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.20168502628803253 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595793966203928 + ], + "opposite_logprob_tokens": [ + -5.130958080291748 + ], + "logprob_reward": -0.00595793966203928, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595793966203928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11257939040660858, + 0.024107052013278008, + 0.04400011897087097, + 0.02631869539618492, + 0.5451151132583618, + 0.06031686067581177, + 0.06985380500555038, + 0.07570526748895645, + 0.01328477542847395, + 0.02871890924870968 + ], + "credit_map": [ + 0.00986224040389061, + 0.02036956697702408, + 0.009131225757300854, + 0.013882260769605637, + 0.01451156847178936, + 0.019788827747106552, + 0.012917084619402885, + 0.012116624973714352, + 0.003466137917712331, + 0.002282059518620372, + 0.002196261426433921, + 0.0021019666455686092, + 0.0016644714633002877, + 0.004856448154896498, + 0.0036538587883114815, + 0.0038858470506966114, + 0.004884276073426008, + 0.0038863730151206255, + 0.004826038610190153, + 0.005932631436735392, + 0.005775974132120609, + 0.00457881111651659, + 0.007860303856432438, + 0.006255713291466236, + 0.002403104677796364, + 0.0030796988867223263, + 0.0024591158144176006, + 0.0027632564306259155, + 0.0030095274560153484, + 0.0036361750680953264, + 0.0049625360406935215, + 0.004005280323326588, + 0.0264082383364439, + 0.04027983173727989, + 0.02774343080818653, + 0.01586713083088398, + 0.198472261428833, + 0.04585709422826767, + 0.02848753146827221, + 0.16199956834316254, + 0.0044623082503676414, + 0.005136663559824228, + 0.004419995471835136, + 0.00324282911606133, + 0.017697950825095177, + 0.005156157072633505, + 0.004405371844768524, + 0.01579558476805687, + 0.0066011524759233, + 0.010610690340399742, + 0.00716236000880599, + 0.010120177641510963, + 0.007028681691735983, + 0.007002707105129957, + 0.011870937421917915, + 0.009457103908061981, + 0.010809071362018585, + 0.010661794804036617, + 0.0062488834373652935, + 0.0063912454061210155, + 0.006565185263752937, + 0.01411223690956831, + 0.011000515893101692, + 0.009916337206959724, + 0.002524921204894781, + 0.001932875020429492, + 0.001062227413058281, + 0.0016079111956059933, + 0.0013682694407179952, + 0.0017505160067230463, + 0.0015203296206891537, + 0.001517725526355207, + 0.0050893486477434635, + 0.004499119706451893, + 0.00299430126324296, + 0.0027864950243383646, + 0.004876443184912205, + 0.0030208765529096127, + 0.0027744087856262922, + 0.0026779156178236008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.00595793966203928 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003195063676685095 + ], + "opposite_logprob_tokens": [ + -5.753195285797119 + ], + "logprob_reward": -0.003195063676685095, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003195063676685095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13786640763282776, + 0.056982528418302536, + 0.18224450945854187, + 0.08435467630624771, + 0.2627688944339752, + 0.048581819981336594, + 0.04387471824884415, + 0.0861164778470993, + 0.02131686359643936, + 0.07589305937290192 + ], + "credit_map": [ + 0.01572057418525219, + 0.021800534799695015, + 0.011386808939278126, + 0.01773352362215519, + 0.017873795703053474, + 0.022463617846369743, + 0.0157877616584301, + 0.015099801123142242, + 0.006886227987706661, + 0.006366862915456295, + 0.005096484441310167, + 0.005057090427726507, + 0.0042906757444143295, + 0.00938231498003006, + 0.009373764507472515, + 0.010529112070798874, + 0.0208286102861166, + 0.02182522602379322, + 0.01270958874374628, + 0.017293781042099, + 0.028230518102645874, + 0.022464752197265625, + 0.0317363478243351, + 0.027155684307217598, + 0.008126745000481606, + 0.008900206536054611, + 0.006801262963563204, + 0.009006226435303688, + 0.01904226653277874, + 0.010358180850744247, + 0.011590298265218735, + 0.010529501363635063, + 0.014833965338766575, + 0.019574670121073723, + 0.013601859100162983, + 0.009300235658884048, + 0.09326165169477463, + 0.022660860791802406, + 0.013089542277157307, + 0.07644611597061157, + 0.004131851252168417, + 0.0043947515077888966, + 0.006125238724052906, + 0.0046657901257276535, + 0.009686530567705631, + 0.005907628685235977, + 0.005776063073426485, + 0.007893968373537064, + 0.004458252340555191, + 0.006557872518897057, + 0.003625319106504321, + 0.006342296954244375, + 0.005136424675583839, + 0.005454787984490395, + 0.006546707358211279, + 0.005753063131123781, + 0.011244594119489193, + 0.011945938691496849, + 0.007179382257163525, + 0.008539769798517227, + 0.007473247591406107, + 0.015626700595021248, + 0.012904452160000801, + 0.011202392168343067, + 0.003017613897100091, + 0.0030091283842921257, + 0.0021493539679795504, + 0.0022199361119419336, + 0.0020461592357605696, + 0.004223264753818512, + 0.0023641239386051893, + 0.0022872856352478266, + 0.0056093912571668625, + 0.01986340992152691, + 0.004024709574878216, + 0.006031359080225229, + 0.02290610782802105, + 0.006510322447866201, + 0.005813322961330414, + 0.005134440492838621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.003195063676685095 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813114523887634 + ], + "opposite_logprob_tokens": [ + -1.4063115119934082 + ], + "logprob_reward": -0.2813114523887634, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813114523887634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07895800471305847, + 0.08937431871891022, + 0.10874224454164505, + 0.04629715904593468, + 0.3890995681285858, + 0.1266542673110962, + 0.05600961297750473, + 0.027669934555888176, + 0.012033654376864433, + 0.06516118347644806 + ], + "credit_map": [ + 0.007613822817802429, + 0.014757559634745121, + 0.006042695138603449, + 0.009317600168287754, + 0.01002524420619011, + 0.013298454694449902, + 0.008682914078235626, + 0.009219711646437645, + 0.009215579368174076, + 0.008846431970596313, + 0.006298539694398642, + 0.007070894818753004, + 0.0067774453200399876, + 0.015320644713938236, + 0.017162548378109932, + 0.018682237714529037, + 0.011774159967899323, + 0.014018960297107697, + 0.007801251020282507, + 0.010235825553536415, + 0.018000399693846703, + 0.014683621004223824, + 0.016471607610583305, + 0.01575642079114914, + 0.004514816217124462, + 0.004461205564439297, + 0.006370019633322954, + 0.006990671623498201, + 0.004917934071272612, + 0.006394060328602791, + 0.006125438492745161, + 0.006523014046251774, + 0.019564615562558174, + 0.028778258711099625, + 0.020654087886214256, + 0.011452987790107727, + 0.1422615796327591, + 0.03359077498316765, + 0.016879934817552567, + 0.11591736227273941, + 0.009475842118263245, + 0.009290937334299088, + 0.024248193949460983, + 0.015485796146094799, + 0.016087181866168976, + 0.010057860054075718, + 0.017013518139719963, + 0.024994943290948868, + 0.005363817326724529, + 0.008374136872589588, + 0.005702195223420858, + 0.008097048848867416, + 0.0051293810829520226, + 0.010250460356473923, + 0.0059285773895680904, + 0.007163996808230877, + 0.003001753007993102, + 0.0024313798639923334, + 0.003528589615598321, + 0.004621501080691814, + 0.0019301972351968288, + 0.004000952932983637, + 0.0045722778886556625, + 0.003583283629268408, + 0.0013189148157835007, + 0.0015341313555836678, + 0.0009183530346490443, + 0.0015140707837417722, + 0.0010578110814094543, + 0.0023745507933199406, + 0.0017836777260527015, + 0.0015321442624554038, + 0.004520053509622812, + 0.020365355536341667, + 0.0025440307799726725, + 0.003113093553110957, + 0.02438787743449211, + 0.0036044528242200613, + 0.003459433326497674, + 0.0031668879091739655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.2813114523887634 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-altitude drone shot captures a breathtaking volcanic eruption with a calm and steady perspective, rendered in a striking black-and-white cinematic quality. A vast, crater dominates the frame, its jagged outline casting a deep shadow across the charred landscape. From its depths, luminous, white-hot liquid magma erupts violently. Simultaneously, a towering, dense column of ash and gas rises like a behemoth, its interior churning and expanding, creating distinct layers of gray that obscure the entire sky. Brilliant rivers of molten lava flow in stark contrast, flowing along the fiery mountainside in a stark contrast of light and dark, creating a stark contrast against the surrounding dark rock. The scene is imbued with a sense of primal, violent power, capturing, in silent black and white, a magnificent natural spectacle both awe-inspiring and devastating.", + "seed": 19313, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_2/index_9617_seed_19313_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3431210611946881, + "vlm_logprob": -0.3431210611946881, + "vlm_yes_no_margin": 2.3, + "vlm_token_credit": -0.3431210611946881, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.887646496295929 + ], + "opposite_logprob_tokens": [ + -0.637646496295929 + ], + "logprob_reward": -0.887646496295929, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.887646496295929, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.5126094222068787, + 0.04381967708468437, + 0.1456306278705597, + 0.08719604462385178, + 0.023457445204257965, + 0.021771878004074097, + 0.06428030878305435, + 0.03730505704879761, + 0.021866751834750175, + 0.04206279665231705 + ], + "credit_map": [ + 0.031724438071250916, + 0.03837929666042328, + 0.18182602524757385, + 0.03627852723002434, + 0.0583902932703495, + 0.1409592628479004, + 0.013248331844806671, + 0.011803267523646355, + 0.0055610728450119495, + 0.003726770868524909, + 0.006549905985593796, + 0.006890527438372374, + 0.006112230010330677, + 0.0050991917960345745, + 0.004358006175607443, + 0.005521976854652166, + 0.00952103827148676, + 0.013808116316795349, + 0.025872448459267616, + 0.022509494796395302, + 0.021726388484239578, + 0.010969817638397217, + 0.026837125420570374, + 0.014386212453246117, + 0.005658225156366825, + 0.008852963335812092, + 0.01230621337890625, + 0.015218631364405155, + 0.01103252824395895, + 0.01135391928255558, + 0.015902210026979446, + 0.006871364079415798, + 0.0023498241789638996, + 0.002592586213722825, + 0.0030717789195477962, + 0.005799958016723394, + 0.0025764256715774536, + 0.002438418800011277, + 0.0026732126716524363, + 0.0019552416633814573, + 0.002278790110722184, + 0.002505586249753833, + 0.0032913738396018744, + 0.005320569034665823, + 0.0017284256173297763, + 0.002908646594733, + 0.0018903050804510713, + 0.0018481845036149025, + 0.005900291260331869, + 0.006035656668245792, + 0.007533029653131962, + 0.01729608327150345, + 0.0046729156747460365, + 0.007317138835787773, + 0.008082599379122257, + 0.007442605681717396, + 0.0030477673280984163, + 0.005845798645168543, + 0.002848575823009014, + 0.006045785266906023, + 0.007438652217388153, + 0.0047792699187994, + 0.0024327116552740335, + 0.004866500850766897, + 0.0021222697105258703, + 0.002843822119757533, + 0.002130551729351282, + 0.002536118496209383, + 0.0026088752783834934, + 0.005543428473174572, + 0.0024067058693617582, + 0.0016749819042161107, + 0.0032061939127743244, + 0.0058884816244244576, + 0.003098675049841404, + 0.005441256333142519, + 0.01017911545932293, + 0.005839375779032707, + 0.005265629384666681, + 0.0031440737657248974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.887646496295929 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38729631900787354 + ], + "opposite_logprob_tokens": [ + -1.1372963190078735 + ], + "logprob_reward": -0.38729631900787354, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38729631900787354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3121578097343445, + 0.030449265614151955, + 0.36093980073928833, + 0.03213762864470482, + 0.03232971951365471, + 0.01401273813098669, + 0.02036874182522297, + 0.1089891567826271, + 0.028258539736270905, + 0.06035664677619934 + ], + "credit_map": [ + 0.017497029155492783, + 0.012624730356037617, + 0.12464793026447296, + 0.02170381136238575, + 0.0427476242184639, + 0.07876631617546082, + 0.00791993085294962, + 0.006250436417758465, + 0.003657689318060875, + 0.0033491102512925863, + 0.00471850298345089, + 0.005314381327480078, + 0.00397806242108345, + 0.0033189908135682344, + 0.003672600258141756, + 0.002439927775412798, + 0.01982349343597889, + 0.0721999853849411, + 0.07665366679430008, + 0.01849941536784172, + 0.12241127341985703, + 0.024171452969312668, + 0.01261330395936966, + 0.014567202888429165, + 0.006222025491297245, + 0.0038964559789747, + 0.0041047995910048485, + 0.004187756218016148, + 0.0038095044437795877, + 0.003886030288413167, + 0.0031898561865091324, + 0.002841199981048703, + 0.004261400550603867, + 0.0044496566988527775, + 0.0042884270660579205, + 0.006177228409796953, + 0.003267082152888179, + 0.004521717317402363, + 0.00282633351162076, + 0.0025378710124641657, + 0.0011779408669099212, + 0.0025087243411689997, + 0.0017538585234433413, + 0.002472800202667713, + 0.001600356656126678, + 0.0023296142462641, + 0.0011044887360185385, + 0.0010649551404640079, + 0.001789337256923318, + 0.002411794615909457, + 0.0019168802537024021, + 0.004973535891622305, + 0.0020911667961627245, + 0.0019888626411557198, + 0.002521563321352005, + 0.0026756019797176123, + 0.009660167619585991, + 0.01926559954881668, + 0.006472685374319553, + 0.00926164910197258, + 0.034464091062545776, + 0.014387787319719791, + 0.0072094956412911415, + 0.008267677389085293, + 0.0020306201186031103, + 0.0031997491605579853, + 0.003926292061805725, + 0.0029524106066673994, + 0.004480827134102583, + 0.006269695237278938, + 0.00345142581500113, + 0.0019475192530080676, + 0.004400343634188175, + 0.009234434925019741, + 0.004240419715642929, + 0.007092330604791641, + 0.013795278035104275, + 0.010195967741310596, + 0.006826851516962051, + 0.004571021068841219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.38729631900787354 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007643263787031174 + ], + "opposite_logprob_tokens": [ + -4.882643222808838 + ], + "logprob_reward": -0.007643263787031174, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007643263787031174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19763676822185516, + 0.02903730608522892, + 0.5384960770606995, + 0.025314131751656532, + 0.01688358746469021, + 0.0346539281308651, + 0.0463382750749588, + 0.06551355123519897, + 0.014512360095977783, + 0.031614068895578384 + ], + "credit_map": [ + 0.011809613555669785, + 0.028556110337376595, + 0.04731709137558937, + 0.012936519458889961, + 0.04346443712711334, + 0.04175463318824768, + 0.006641646381467581, + 0.005156707018613815, + 0.0031932636629790068, + 0.0025422589387744665, + 0.00445146020501852, + 0.004466891288757324, + 0.0031440160237252712, + 0.0037383318413048983, + 0.002453641965985298, + 0.005047438200563192, + 0.028963392600417137, + 0.10281836241483688, + 0.11193010210990906, + 0.03474085032939911, + 0.1739092320203781, + 0.03648326173424721, + 0.02429444156587124, + 0.02535638026893139, + 0.0026541126426309347, + 0.0026815258897840977, + 0.003386181313544512, + 0.0038053570315241814, + 0.0028324322775006294, + 0.0030071106739342213, + 0.004867574665695429, + 0.002079834695905447, + 0.001440215390175581, + 0.0017903379630297422, + 0.0025149767752736807, + 0.00478875869885087, + 0.0015322149265557528, + 0.0016796878771856427, + 0.0018755210330709815, + 0.0012618736363947392, + 0.0029291596729308367, + 0.003759430255740881, + 0.004898104816675186, + 0.008371949195861816, + 0.0038440951611846685, + 0.004252508282661438, + 0.0035522060934454203, + 0.0030464741867035627, + 0.003748837625607848, + 0.005509159993380308, + 0.004999125842005014, + 0.01147312019020319, + 0.0037763507571071386, + 0.0044531202875077724, + 0.006269022356718779, + 0.0061095296405255795, + 0.006055989768356085, + 0.01001374889165163, + 0.005021851509809494, + 0.009552985429763794, + 0.01600978523492813, + 0.008024872280657291, + 0.0054394956678152084, + 0.005394822917878628, + 0.0011823063250631094, + 0.002650839276611805, + 0.0009299510275013745, + 0.0019172120373696089, + 0.0029983664862811565, + 0.0020691179670393467, + 0.0018494658870622516, + 0.0009150991681963205, + 0.0023925460409373045, + 0.006460157223045826, + 0.0019129433203488588, + 0.004095405340194702, + 0.006840817164629698, + 0.00412012217566371, + 0.0038942191749811172, + 0.0018978568259626627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.007643263787031174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410274276509881 + ], + "opposite_logprob_tokens": [ + -5.50410270690918 + ], + "logprob_reward": -0.00410274276509881, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410274276509881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3255127966403961, + 0.11520702391862869, + 0.23090024292469025, + 0.029085718095302582, + 0.014497573487460613, + 0.0541212372481823, + 0.0386970192193985, + 0.09661944210529327, + 0.07451207935810089, + 0.020846931263804436 + ], + "credit_map": [ + 0.022218234837055206, + 0.031676143407821655, + 0.11704349517822266, + 0.025131983682513237, + 0.025950338691473007, + 0.0849103108048439, + 0.010648276656866074, + 0.00793398730456829, + 0.012367230840027332, + 0.01597573235630989, + 0.014996997080743313, + 0.016200141981244087, + 0.02314354106783867, + 0.011267688125371933, + 0.010411686263978481, + 0.010843999683856964, + 0.012707121670246124, + 0.04532182961702347, + 0.046580009162425995, + 0.012444957159459591, + 0.07573572546243668, + 0.015849072486162186, + 0.010080577805638313, + 0.012180929072201252, + 0.00314336153678596, + 0.005058486945927143, + 0.003495370037853718, + 0.003944601863622665, + 0.004106204025447369, + 0.004659312777221203, + 0.00235778302885592, + 0.002320594619959593, + 0.0015591357368975878, + 0.0020572456996887922, + 0.0018652452854439616, + 0.0024824056308716536, + 0.001896703732199967, + 0.00159929390065372, + 0.0017967165913432837, + 0.001240825280547142, + 0.005444447509944439, + 0.004456622526049614, + 0.007691203150898218, + 0.0151358125731349, + 0.004109952133148909, + 0.006238407455384731, + 0.006196748930960894, + 0.004848037380725145, + 0.003468981711193919, + 0.006332153920084238, + 0.004458476789295673, + 0.007103851065039635, + 0.0039568389765918255, + 0.004607058130204678, + 0.004389703739434481, + 0.004379952326416969, + 0.00847866665571928, + 0.015185712836682796, + 0.008705037645995617, + 0.023037448525428772, + 0.015070749446749687, + 0.010299905203282833, + 0.007859750650823116, + 0.007982160896062851, + 0.005806825123727322, + 0.010629024356603622, + 0.00934218056499958, + 0.007080987561494112, + 0.011912332847714424, + 0.01669052243232727, + 0.008370268158614635, + 0.004679928533732891, + 0.0018184001091867685, + 0.0034789990168064833, + 0.001320547773502767, + 0.0032078190706670284, + 0.004432883113622665, + 0.0033836429938673973, + 0.0017058660741895437, + 0.0014987706672400236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.00410274276509881 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42891648411750793 + ], + "opposite_logprob_tokens": [ + -1.0539164543151855 + ], + "logprob_reward": -0.42891648411750793, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42891648411750793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17155607044696808, + 0.021176617592573166, + 0.5859739780426025, + 0.020816612988710403, + 0.038961298763751984, + 0.015784673392772675, + 0.013589004054665565, + 0.05590777471661568, + 0.02676198072731495, + 0.049472007900476456 + ], + "credit_map": [ + 0.009308943524956703, + 0.007015509065240622, + 0.06837189197540283, + 0.01246845442801714, + 0.02373444475233555, + 0.043166741728782654, + 0.004201262723654509, + 0.0032888243440538645, + 0.002013289136812091, + 0.002184726996347308, + 0.002957490738481283, + 0.0029185728635638952, + 0.003923294134438038, + 0.002338102785870433, + 0.001987760653719306, + 0.0028533798176795244, + 0.03147345781326294, + 0.11377058923244476, + 0.11789993196725845, + 0.03566151112318039, + 0.1898689717054367, + 0.03906315565109253, + 0.03319188207387924, + 0.025044484063982964, + 0.001925144111737609, + 0.002434490015730262, + 0.0032626893371343613, + 0.005334255751222372, + 0.0020082173869013786, + 0.0019325694302096963, + 0.0026243452448397875, + 0.0012949008960276842, + 0.0029913738835603, + 0.004330566618591547, + 0.005440312437713146, + 0.010203821584582329, + 0.0034083484206348658, + 0.003959913272410631, + 0.005744104739278555, + 0.002882858272641897, + 0.001404818962328136, + 0.0023980308324098587, + 0.002493953565135598, + 0.0030982098542153835, + 0.001595037174411118, + 0.0017098623793572187, + 0.0016291751526296139, + 0.0014555854722857475, + 0.0016842075856402516, + 0.0014520566910505295, + 0.0010948344133794308, + 0.0025864054914563894, + 0.0017844911199063063, + 0.001599925453774631, + 0.001693197526037693, + 0.0016938851913437247, + 0.0049389200285077095, + 0.008531398139894009, + 0.004322938155382872, + 0.005550626199692488, + 0.015265006572008133, + 0.007962766103446484, + 0.004113955423235893, + 0.0052221608348190784, + 0.002216416411101818, + 0.003736584447324276, + 0.0031353235244750977, + 0.0027951253578066826, + 0.004823082592338324, + 0.005213577300310135, + 0.003149859607219696, + 0.0016920128837227821, + 0.004053026903420687, + 0.0065646301954984665, + 0.0033584842458367348, + 0.011234735138714314, + 0.006855540908873081, + 0.0052836281247437, + 0.008890625089406967, + 0.0032313333358615637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.42891648411750793 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-altitude drone shot captures a breathtaking volcanic eruption with a calm and steady perspective, rendered in a striking black-and-white cinematic quality. A vast, crater dominates the frame, its jagged outline casting a deep shadow across the charred landscape. From its depths, luminous, white-hot liquid magma erupts violently. Simultaneously, a towering, dense column of ash and gas rises like a behemoth, its interior churning and expanding, creating distinct layers of gray that obscure the entire sky. Brilliant rivers of molten lava flow in stark contrast, flowing along the fiery mountainside in a stark contrast of light and dark, creating a stark contrast against the surrounding dark rock. The scene is imbued with a sense of primal, violent power, capturing, in silent black and white, a magnificent natural spectacle both awe-inspiring and devastating.", + "seed": 19314, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_2/index_9617_seed_19314_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2646346613764763, + "vlm_logprob": -0.2646346613764763, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.2646346613764763, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8552706837654114 + ], + "opposite_logprob_tokens": [ + -0.6052706837654114 + ], + "logprob_reward": -0.8552706837654114, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8552706837654114, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12576386332511902, + 0.21487224102020264, + 0.04993162304162979, + 0.07211209088563919, + 0.21165896952152252, + 0.05466971546411514, + 0.030130423605442047, + 0.031279854476451874, + 0.12141227722167969, + 0.08816889673471451 + ], + "credit_map": [ + 0.015201359055936337, + 0.01292306836694479, + 0.01568302884697914, + 0.018620075657963753, + 0.02104356326162815, + 0.01667775586247444, + 0.010983007960021496, + 0.014632013626396656, + 0.020512228831648827, + 0.026708167046308517, + 0.029685184359550476, + 0.02218976430594921, + 0.055594779551029205, + 0.024383533746004105, + 0.017214015126228333, + 0.018584590405225754, + 0.002885593567043543, + 0.012023835442960262, + 0.0050406670197844505, + 0.004748537205159664, + 0.012936171144247055, + 0.005345555953681469, + 0.00323644676245749, + 0.0037148219998925924, + 0.00719112204387784, + 0.007495925296097994, + 0.010132514871656895, + 0.013027658686041832, + 0.004896786063909531, + 0.009140812791883945, + 0.010959903709590435, + 0.009267380461096764, + 0.017907582223415375, + 0.027141572907567024, + 0.02795170247554779, + 0.03208766132593155, + 0.025387659668922424, + 0.04543880745768547, + 0.018775012344121933, + 0.016968984156847, + 0.003973277751356363, + 0.004659702535718679, + 0.0077695949003100395, + 0.013034123927354813, + 0.006342998240143061, + 0.006002358626574278, + 0.008175566792488098, + 0.004712102003395557, + 0.0030557813588529825, + 0.0029455050826072693, + 0.0030521622393280268, + 0.006099671591073275, + 0.004572554957121611, + 0.0037977013271301985, + 0.0031727575697004795, + 0.0034342913422733545, + 0.003203361527994275, + 0.003433560486882925, + 0.003430777695029974, + 0.0039989277720451355, + 0.004817793611437082, + 0.0034846290946006775, + 0.00472931656986475, + 0.004181490745395422, + 0.023706955835223198, + 0.01250854879617691, + 0.014553355984389782, + 0.011741941794753075, + 0.017325621098279953, + 0.013845997862517834, + 0.014604121446609497, + 0.013125745579600334, + 0.007309004198759794, + 0.010532228276133537, + 0.008138597942888737, + 0.014622421935200691, + 0.01609228365123272, + 0.010322218760848045, + 0.011300738900899887, + 0.009851407259702682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.8552706837654114 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22566741704940796 + ], + "opposite_logprob_tokens": [ + -1.6006674766540527 + ], + "logprob_reward": -0.22566741704940796, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22566741704940796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19214998185634613, + 0.26679620146751404, + 0.028013477101922035, + 0.030273878946900368, + 0.05488161742687225, + 0.03539002314209938, + 0.04764455556869507, + 0.047098975628614426, + 0.16264937818050385, + 0.13510189950466156 + ], + "credit_map": [ + 0.02066987194120884, + 0.01577417552471161, + 0.02321372926235199, + 0.02684626914560795, + 0.027219228446483612, + 0.035821542143821716, + 0.015615256503224373, + 0.026989907026290894, + 0.03687845915555954, + 0.030108224600553513, + 0.03028803877532482, + 0.02370433509349823, + 0.04763126000761986, + 0.04609161987900734, + 0.024037299677729607, + 0.02805698662996292, + 0.0020868463907390833, + 0.00271572545170784, + 0.003799254773184657, + 0.005913710221648216, + 0.0033554155379533768, + 0.0031295183580368757, + 0.004168049897998571, + 0.002844954375177622, + 0.002471031155437231, + 0.002892054384574294, + 0.004293749574571848, + 0.007206955458968878, + 0.002219668123871088, + 0.003950991667807102, + 0.005022212862968445, + 0.002217215718701482, + 0.005008584354072809, + 0.007037018891423941, + 0.008139418438076973, + 0.011792396195232868, + 0.004530709236860275, + 0.006310257129371166, + 0.005932875443249941, + 0.0061303540132939816, + 0.002552495803683996, + 0.004326759837567806, + 0.004016785882413387, + 0.005704077892005444, + 0.006191586144268513, + 0.003041054354980588, + 0.004687207750976086, + 0.00487005477771163, + 0.0038821534253656864, + 0.005691077094525099, + 0.004246525000780821, + 0.008543742820620537, + 0.010138369165360928, + 0.005522296763956547, + 0.0049354820512235165, + 0.004684911575168371, + 0.0038109172601252794, + 0.003890868043527007, + 0.005936747882515192, + 0.006273961625993252, + 0.00595651101320982, + 0.006618708372116089, + 0.008052371442317963, + 0.006558890920132399, + 0.013203081674873829, + 0.018607622012495995, + 0.013446214608848095, + 0.014667591080069542, + 0.022098541259765625, + 0.03213934227824211, + 0.021673211827874184, + 0.026813773438334465, + 0.009280900470912457, + 0.016978207975625992, + 0.010084765963256359, + 0.013188950717449188, + 0.027741527184844017, + 0.02378740720450878, + 0.017428960651159286, + 0.01661117747426033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.22566741704940796 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549244798719883 + ], + "opposite_logprob_tokens": [ + -4.38754940032959 + ], + "logprob_reward": -0.012549244798719883, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549244798719883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15451471507549286, + 0.23524360358715057, + 0.07237984240055084, + 0.0821196511387825, + 0.17357176542282104, + 0.04289192333817482, + 0.0972900390625, + 0.03703559190034866, + 0.060538940131664276, + 0.04441387206315994 + ], + "credit_map": [ + 0.012723461724817753, + 0.01760951615869999, + 0.016723094508051872, + 0.019608518108725548, + 0.035331305116415024, + 0.02237575501203537, + 0.013124052435159683, + 0.01701902225613594, + 0.023913107812404633, + 0.02878027781844139, + 0.030146056786179543, + 0.021928055211901665, + 0.05870584771037102, + 0.02966349758207798, + 0.020589269697666168, + 0.02151750586926937, + 0.00485407467931509, + 0.01668238453567028, + 0.00851256400346756, + 0.007504416164010763, + 0.016674822196364403, + 0.007350884843617678, + 0.005030136089771986, + 0.005770565010607243, + 0.005024448968470097, + 0.009893132373690605, + 0.011697576381266117, + 0.02048318274319172, + 0.006700909696519375, + 0.010568909347057343, + 0.011330620385706425, + 0.006420876365154982, + 0.013347402215003967, + 0.020764492452144623, + 0.023362936452031136, + 0.028750281780958176, + 0.020243283361196518, + 0.03022189810872078, + 0.021572507917881012, + 0.015308963134884834, + 0.0033858444076031446, + 0.004937315825372934, + 0.004923222120851278, + 0.008629899471998215, + 0.005532355979084969, + 0.004314520861953497, + 0.006343106273561716, + 0.004825656767934561, + 0.008216784335672855, + 0.010228111408650875, + 0.01348379347473383, + 0.02046327292919159, + 0.012365579605102539, + 0.010937194339931011, + 0.008919663727283478, + 0.012675647623836994, + 0.002544557210057974, + 0.0038370038382709026, + 0.0046363226138055325, + 0.003901574295014143, + 0.005674525164067745, + 0.005672150291502476, + 0.0060238102450966835, + 0.004745649639517069, + 0.004383695777505636, + 0.005867088213562965, + 0.004796022083610296, + 0.006843854207545519, + 0.008090603165328503, + 0.009416813030838966, + 0.008630306459963322, + 0.012510563246905804, + 0.0032538739033043385, + 0.004810237791389227, + 0.0029111409094184637, + 0.008268238976597786, + 0.008655273355543613, + 0.006417850032448769, + 0.005738129839301109, + 0.004359130281955004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.012549244798719883 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100843332707882 + ], + "opposite_logprob_tokens": [ + -5.504100799560547 + ], + "logprob_reward": -0.004100843332707882, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100843332707882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3709773123264313, + 0.0551082044839859, + 0.14889369904994965, + 0.03415625914931297, + 0.08389613777399063, + 0.04973819851875305, + 0.10592838376760483, + 0.05024239793419838, + 0.05166083574295044, + 0.049398601055145264 + ], + "credit_map": [ + 0.03823399171233177, + 0.03626590967178345, + 0.042594779282808304, + 0.05648128315806389, + 0.0665532574057579, + 0.06234345585107803, + 0.02672690898180008, + 0.04177771136164665, + 0.006049139425158501, + 0.006350365467369556, + 0.0056171175092458725, + 0.0057411203160882, + 0.01266451831907034, + 0.007098230067640543, + 0.005451700650155544, + 0.006136014591902494, + 0.005993635393679142, + 0.04733636975288391, + 0.007844986394047737, + 0.007239502388983965, + 0.05907405912876129, + 0.008969255723059177, + 0.006116847507655621, + 0.006319040432572365, + 0.0027600026223808527, + 0.006349465344101191, + 0.004240899812430143, + 0.004854012746363878, + 0.003753052791580558, + 0.005054949317127466, + 0.003473201533779502, + 0.003670676378533244, + 0.007394519634544849, + 0.011630745604634285, + 0.011880002915859222, + 0.015103708952665329, + 0.008490152657032013, + 0.011774042621254921, + 0.008036497049033642, + 0.009586464613676071, + 0.0036802792456001043, + 0.004309055861085653, + 0.006646109279245138, + 0.005647500976920128, + 0.0072623654268682, + 0.005160887259989977, + 0.006478628609329462, + 0.010553371161222458, + 0.008316884748637676, + 0.013143361546099186, + 0.011167168617248535, + 0.019864941015839577, + 0.022349748760461807, + 0.010458077304065228, + 0.010086221620440483, + 0.01054198108613491, + 0.004145029932260513, + 0.005942234303802252, + 0.005034072324633598, + 0.0062783327884972095, + 0.011161715723574162, + 0.006730047520250082, + 0.006556495558470488, + 0.0043944744393229485, + 0.004590707831084728, + 0.0071460530161857605, + 0.004390704911202192, + 0.004562932997941971, + 0.011670100502669811, + 0.007524118758738041, + 0.0059113092720508575, + 0.005864910315722227, + 0.00365634192712605, + 0.005107975099235773, + 0.0037182930391281843, + 0.00808746088296175, + 0.008615010417997837, + 0.006856550462543964, + 0.007991106249392033, + 0.0053658620454370975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.004100843332707882 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22558511793613434 + ], + "opposite_logprob_tokens": [ + -1.6005851030349731 + ], + "logprob_reward": -0.22558511793613434, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558511793613434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12680000066757202, + 0.18226075172424316, + 0.15350787341594696, + 0.09252931922674179, + 0.15212994813919067, + 0.11967344582080841, + 0.07278688251972198, + 0.034251049160957336, + 0.032205577939748764, + 0.033855095505714417 + ], + "credit_map": [ + 0.012790088541805744, + 0.010856475681066513, + 0.016183489933609962, + 0.013582459650933743, + 0.017620768398046494, + 0.030491402372717857, + 0.010431364178657532, + 0.0148439547047019, + 0.013180096633732319, + 0.01969001069664955, + 0.02481609210371971, + 0.022010397166013718, + 0.026116179302334785, + 0.03728891536593437, + 0.019956279546022415, + 0.01920277625322342, + 0.0058685243129730225, + 0.05029735341668129, + 0.008902431465685368, + 0.01405455730855465, + 0.055596791207790375, + 0.006387392990291119, + 0.008238357491791248, + 0.0041624740697443485, + 0.006506646517664194, + 0.008703218773007393, + 0.013837872073054314, + 0.02208971045911312, + 0.007122386712580919, + 0.010670523159205914, + 0.014309300109744072, + 0.009289669804275036, + 0.01296909898519516, + 0.01975359581410885, + 0.020881658419966698, + 0.029385855421423912, + 0.01606607623398304, + 0.029551547020673752, + 0.012189395725727081, + 0.011332722380757332, + 0.007914375513792038, + 0.010234266519546509, + 0.01358440425246954, + 0.02813086472451687, + 0.016017386689782143, + 0.01302009541541338, + 0.017693229019641876, + 0.013078824616968632, + 0.006292299833148718, + 0.007568692322820425, + 0.006691699847579002, + 0.0179391298443079, + 0.011555506847798824, + 0.007484170608222485, + 0.00879624579101801, + 0.006459144409745932, + 0.0024544906336814165, + 0.0036567894276231527, + 0.004236025735735893, + 0.0041542816907167435, + 0.005742127541452646, + 0.004663253668695688, + 0.00585441617295146, + 0.003489666385576129, + 0.004495120607316494, + 0.0034333455841988325, + 0.003259627614170313, + 0.0028210177551954985, + 0.0044365087524056435, + 0.005768639035522938, + 0.004141750279814005, + 0.0038495706394314766, + 0.0027420774567872286, + 0.00394004862755537, + 0.0035637093242257833, + 0.004779452458024025, + 0.005643037147819996, + 0.005080753471702337, + 0.0030452858190983534, + 0.005060731433331966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.22558511793613434 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-altitude drone shot captures a breathtaking volcanic eruption with a calm and steady perspective, rendered in a striking black-and-white cinematic quality. A vast, crater dominates the frame, its jagged outline casting a deep shadow across the charred landscape. From its depths, luminous, white-hot liquid magma erupts violently. Simultaneously, a towering, dense column of ash and gas rises like a behemoth, its interior churning and expanding, creating distinct layers of gray that obscure the entire sky. Brilliant rivers of molten lava flow in stark contrast, flowing along the fiery mountainside in a stark contrast of light and dark, creating a stark contrast against the surrounding dark rock. The scene is imbued with a sense of primal, violent power, capturing, in silent black and white, a magnificent natural spectacle both awe-inspiring and devastating.", + "seed": 19315, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_2/index_9617_seed_19315_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.26548407822847364, + "vlm_logprob": -0.26548407822847364, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.26548407822847364, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13642773032188416 + ], + "opposite_logprob_tokens": [ + -2.386427640914917 + ], + "logprob_reward": -0.13642773032188416, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.13642773032188416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45005297660827637, + 0.14426156878471375, + 0.07414897531270981, + 0.0685630589723587, + 0.04739752784371376, + 0.07035327702760696, + 0.03650296479463577, + 0.0378732904791832, + 0.04633425921201706, + 0.02451215498149395 + ], + "credit_map": [ + 0.04580754414200783, + 0.04675216227769852, + 0.11659805476665497, + 0.027352716773748398, + 0.09089183807373047, + 0.031684510409832, + 0.0369395986199379, + 0.05402655526995659, + 0.017234787344932556, + 0.010911497287452221, + 0.021118709817528725, + 0.02209596522152424, + 0.01625588908791542, + 0.01568487100303173, + 0.019026845693588257, + 0.021932991221547127, + 0.004459063522517681, + 0.006354391109198332, + 0.009590192697942257, + 0.007314460817724466, + 0.014077181927859783, + 0.009795601479709148, + 0.012398755177855492, + 0.010159333236515522, + 0.005466178059577942, + 0.0042683957144618034, + 0.008593915961682796, + 0.007125686388462782, + 0.007042871322482824, + 0.005751976743340492, + 0.022118084132671356, + 0.008195944130420685, + 0.004965374246239662, + 0.004903474356979132, + 0.007489690091460943, + 0.005522146355360746, + 0.007264194078743458, + 0.0067076231352984905, + 0.0043699112720787525, + 0.006175115704536438, + 0.010268628597259521, + 0.008523860014975071, + 0.011094602756202221, + 0.004914663732051849, + 0.009266437031328678, + 0.013045337051153183, + 0.0040181465446949005, + 0.00922160129994154, + 0.005950090009719133, + 0.004891468212008476, + 0.00544623751193285, + 0.003100448753684759, + 0.0056123314425349236, + 0.004764714278280735, + 0.0032414377201348543, + 0.003496236400678754, + 0.004773232154548168, + 0.0064622340723872185, + 0.006771331187337637, + 0.004203574266284704, + 0.004133020993322134, + 0.005187279544770718, + 0.0031297956593334675, + 0.0032128230668604374, + 0.007299595978111029, + 0.004978739190846682, + 0.0067433384247124195, + 0.005528061650693417, + 0.004971873015165329, + 0.004919696133583784, + 0.006279534660279751, + 0.005613422486931086, + 0.002290274715051055, + 0.003220337675884366, + 0.002895048353821039, + 0.0024605453945696354, + 0.0027822561096400023, + 0.004325093235820532, + 0.003563664387911558, + 0.0029749360401183367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.13642773032188416 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8984137773513794 + ], + "opposite_logprob_tokens": [ + -0.5234137773513794 + ], + "logprob_reward": -0.8984137773513794, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8984137773513794, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22496066987514496, + 0.10711710155010223, + 0.0802401453256607, + 0.052304286509752274, + 0.10064177215099335, + 0.10809565335512161, + 0.22682784497737885, + 0.06542448699474335, + 0.013978860341012478, + 0.020409202203154564 + ], + "credit_map": [ + 0.01801743544638157, + 0.03210349380970001, + 0.025385916233062744, + 0.016516873612999916, + 0.049363844096660614, + 0.015524422749876976, + 0.03492443263530731, + 0.03312426805496216, + 0.022475462406873703, + 0.003922832198441029, + 0.010518326424062252, + 0.03726125508546829, + 0.007978423498570919, + 0.007165076211094856, + 0.01122251246124506, + 0.0065732235088944435, + 0.005702297668904066, + 0.00619865907356143, + 0.008175603114068508, + 0.008296620100736618, + 0.009802686981856823, + 0.011938126757740974, + 0.018431656062602997, + 0.011694506742060184, + 0.003596467897295952, + 0.003759397892281413, + 0.005466683767735958, + 0.009456691332161427, + 0.005461825057864189, + 0.005454500671476126, + 0.013537772931158543, + 0.005570955108851194, + 0.01022318471223116, + 0.006167702842503786, + 0.014159876853227615, + 0.020180894061923027, + 0.015698309987783432, + 0.011211090721189976, + 0.01077300775796175, + 0.012227717787027359, + 0.016657492145895958, + 0.013961073011159897, + 0.016528187319636345, + 0.007787367794662714, + 0.013245251029729843, + 0.019617650657892227, + 0.006362329237163067, + 0.013936314731836319, + 0.0425570011138916, + 0.03011755645275116, + 0.030048340559005737, + 0.015314333140850067, + 0.038983915001153946, + 0.026736602187156677, + 0.02038506790995598, + 0.02268505096435547, + 0.00850845780223608, + 0.009384233504533768, + 0.009393862448632717, + 0.007743107154965401, + 0.007478919345885515, + 0.009409056045114994, + 0.006307675503194332, + 0.007199177518486977, + 0.0018574392888695002, + 0.0016312863444909453, + 0.0016972963931038976, + 0.0013904398074373603, + 0.0016083030495792627, + 0.00198549241758883, + 0.00219240365549922, + 0.0016162011306732893, + 0.00260373717173934, + 0.002822650596499443, + 0.00254111853428185, + 0.002063516993075609, + 0.0025153153110295534, + 0.0034770129714161158, + 0.0026785661466419697, + 0.0017072873888537288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.8984137773513794 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639360148459673 + ], + "opposite_logprob_tokens": [ + -4.882639408111572 + ], + "logprob_reward": -0.007639360148459673, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639360148459673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30393749475479126, + 0.22659708559513092, + 0.07748176157474518, + 0.06195946782827377, + 0.04486306011676788, + 0.06125764921307564, + 0.10188092291355133, + 0.05716899037361145, + 0.051263511180877686, + 0.013590077869594097 + ], + "credit_map": [ + 0.03869682177901268, + 0.012645636685192585, + 0.12086865305900574, + 0.015660878270864487, + 0.01703859306871891, + 0.013225479982793331, + 0.011555543169379234, + 0.0742458701133728, + 0.04406704381108284, + 0.009036452509462833, + 0.02065717615187168, + 0.054593730717897415, + 0.015275164507329464, + 0.010964619927108288, + 0.05483115091919899, + 0.017171746119856834, + 0.007135545834898949, + 0.008762504905462265, + 0.0077597713097929955, + 0.008781848475337029, + 0.012778270989656448, + 0.01015587616711855, + 0.008470660075545311, + 0.013637274503707886, + 0.0038540970999747515, + 0.003487358335405588, + 0.008273819461464882, + 0.012813150882720947, + 0.008300811983644962, + 0.009236596524715424, + 0.008522036485373974, + 0.007471599616110325, + 0.004256542306393385, + 0.004303203895688057, + 0.007401687558740377, + 0.00442800996825099, + 0.007394467014819384, + 0.006343137007206678, + 0.004703093785792589, + 0.006032921373844147, + 0.009945979341864586, + 0.005942242220044136, + 0.009243812412023544, + 0.004869171883910894, + 0.007212109863758087, + 0.010411601513624191, + 0.005906223319470882, + 0.007726509124040604, + 0.01648356206715107, + 0.01432220358401537, + 0.012828465551137924, + 0.008406908251345158, + 0.019150719046592712, + 0.01215369813144207, + 0.009580698795616627, + 0.008954664692282677, + 0.007634762674570084, + 0.011212751269340515, + 0.007662963587790728, + 0.007130144629627466, + 0.007279416546225548, + 0.006156068295240402, + 0.004845966584980488, + 0.0052469163201749325, + 0.005245750304311514, + 0.005747971124947071, + 0.006399570032954216, + 0.004252070095390081, + 0.0068836938589811325, + 0.012008603662252426, + 0.00392386456951499, + 0.00680198660120368, + 0.001823268597945571, + 0.002177454764023423, + 0.0014805664541199803, + 0.0015462228329852223, + 0.0017805920215323567, + 0.0019356071716174483, + 0.0016679612454026937, + 0.0011784047819674015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.007639360148459673 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036201435141265392 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.0036201435141265392, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036201435141265392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21435441076755524, + 0.313094824552536, + 0.04025665298104286, + 0.04760902002453804, + 0.03471335768699646, + 0.07852021604776382, + 0.07601696252822876, + 0.05653534457087517, + 0.10154189169406891, + 0.03735737130045891 + ], + "credit_map": [ + 0.02078125812113285, + 0.03338927403092384, + 0.02428271248936653, + 0.012822696007788181, + 0.05834284797310829, + 0.018592003732919693, + 0.021621666848659515, + 0.02452194131910801, + 0.05100550875067711, + 0.01524816732853651, + 0.030149241909384727, + 0.0650983527302742, + 0.025636138394474983, + 0.019317174330353737, + 0.08194956183433533, + 0.02469070442020893, + 0.0038141454569995403, + 0.003596850438043475, + 0.004059455823153257, + 0.005020570941269398, + 0.005589581560343504, + 0.0047687324695289135, + 0.00684338016435504, + 0.0065639386884868145, + 0.0029249980580061674, + 0.002200043760240078, + 0.0057265497744083405, + 0.010080062784254551, + 0.006123268511146307, + 0.004176245536655188, + 0.012850691564381123, + 0.003527158172801137, + 0.0032904809340834618, + 0.002868055831640959, + 0.0044915941543877125, + 0.006166053004562855, + 0.00566926458850503, + 0.0033089190255850554, + 0.005589320324361324, + 0.0033296679612249136, + 0.008750764653086662, + 0.014183304272592068, + 0.010407505556941032, + 0.00883545819669962, + 0.010141026228666306, + 0.010263226926326752, + 0.009558556601405144, + 0.006380369421094656, + 0.010724925436079502, + 0.009537179954349995, + 0.01035385113209486, + 0.007158875931054354, + 0.011167616583406925, + 0.009241005405783653, + 0.009604600258171558, + 0.008228905498981476, + 0.006825755815953016, + 0.01221125852316618, + 0.00837971642613411, + 0.006405290216207504, + 0.007344415877014399, + 0.006844932213425636, + 0.004124445375055075, + 0.004399528261274099, + 0.012840110808610916, + 0.01583232544362545, + 0.012569294311106205, + 0.01075727492570877, + 0.013097897171974182, + 0.011931747198104858, + 0.01423166785389185, + 0.010281575843691826, + 0.004762617871165276, + 0.003886748803779483, + 0.004174493253231049, + 0.002471973653882742, + 0.004181531257927418, + 0.0075670285150408745, + 0.006118391174823046, + 0.004194587003439665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0036201435141265392 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-altitude drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in black-and-white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast crater with a jagged outline visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a towering column of ash and gas rise from the crater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rivers of molten lava flow along the mountainside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28131937980651855 + ], + "opposite_logprob_tokens": [ + -1.4063193798065186 + ], + "logprob_reward": -0.28131937980651855, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28131937980651855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21634671092033386, + 0.18608322739601135, + 0.036444127559661865, + 0.07810036838054657, + 0.0383681058883667, + 0.0931842103600502, + 0.15884758532047272, + 0.0725506916642189, + 0.09111743420362473, + 0.028957540169358253 + ], + "credit_map": [ + 0.021056119352579117, + 0.025866111740469933, + 0.047085657715797424, + 0.014684713445603848, + 0.04235205799341202, + 0.015890803188085556, + 0.0262205321341753, + 0.02319071814417839, + 0.02410418726503849, + 0.010284858755767345, + 0.021402422338724136, + 0.0337272584438324, + 0.02038351632654667, + 0.021685274317860603, + 0.03705184906721115, + 0.017443854361772537, + 0.0031340268906205893, + 0.003113184357061982, + 0.004401206038892269, + 0.003190442221239209, + 0.005558724049478769, + 0.004804558586329222, + 0.004591154400259256, + 0.0076508293859660625, + 0.004800879396498203, + 0.004924773704260588, + 0.010007399134337902, + 0.013263612985610962, + 0.007756297942250967, + 0.009097971953451633, + 0.019106697291135788, + 0.009142730385065079, + 0.005440775770694017, + 0.0031864724587649107, + 0.005760714411735535, + 0.004368345718830824, + 0.005976024083793163, + 0.005514041520655155, + 0.0029136640951037407, + 0.005208069458603859, + 0.013610322959721088, + 0.008699653670191765, + 0.01706038787961006, + 0.00696330051869154, + 0.012028880417346954, + 0.014812684617936611, + 0.007051651831716299, + 0.012957329861819744, + 0.026688827201724052, + 0.020745960995554924, + 0.023445867002010345, + 0.012714046984910965, + 0.026229821145534515, + 0.023249540477991104, + 0.013850115239620209, + 0.011923401616513729, + 0.005064142402261496, + 0.010155008174479008, + 0.010080091655254364, + 0.009235192090272903, + 0.008546489290893078, + 0.010775635950267315, + 0.010972478426992893, + 0.007721652276813984, + 0.00922784861177206, + 0.014008910395205021, + 0.009890308603644371, + 0.010093029588460922, + 0.012243016622960567, + 0.011061345227062702, + 0.017272857949137688, + 0.007320125121623278, + 0.002916994271799922, + 0.003313147695735097, + 0.0035388143733143806, + 0.0031418956350535154, + 0.0035205944441258907, + 0.005746523384004831, + 0.003740418003872037, + 0.0030391523614525795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.28131937980651855 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_3/index_tensor([4655, 4655])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_3/index_tensor([4655, 4655])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..86b84c5affd16e17d28f5c100108a233add14dc0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_3/index_tensor([4655, 4655])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning macro lens, the camera held still, focuses perfectly on a slowly rotating bracelet. The bracelet is composed of a string of nearly perfectly spherical, soft pearls, each radiating a soft, milky white luster and silky smooth, artfully connected by a nearly invisible thread. Against a deep, blurred black backdrop, the bracelet rotates steadily and gracefully around its central axis. As it moves, a dreamy, soft light from an unidentified source casts a dazzling, fairytale-like brilliance on the pearls—pale pink, sky blue, mint green, and light gold gently flow, blend, and transform across the smooth curves of each pearl, creating hazy, soft halos of light. This slow movement and the shifting light and shadows create a sense of tranquility, dreaminess, and exquisite refinement, as if a treasure from a magical kingdom slowly reveals its mysterious and captivating charm.", + "seed": 14677, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_3/index_4655_seed_14677_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0769143688492477, + "vlm_logprob": -0.0769143688492477, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.0769143688492477, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03194993734359741 + ], + "opposite_logprob_tokens": [ + -3.781949996948242 + ], + "logprob_reward": -0.03194993734359741, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03194993734359741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06876164674758911, + 0.09239161759614944, + 0.05766565725207329, + 0.18332575261592865, + 0.10243474692106247, + 0.2741285264492035, + 0.030669162049889565, + 0.07545575499534607, + 0.046047698706388474, + 0.06911934167146683 + ], + "credit_map": [ + 0.007286875508725643, + 0.009394802153110504, + 0.01021401397883892, + 0.009227659553289413, + 0.008728738874197006, + 0.0076652755960822105, + 0.008996284566819668, + 0.007247998844832182, + 0.017912670969963074, + 0.007036566734313965, + 0.011125928722321987, + 0.007464199792593718, + 0.01611143723130226, + 0.006840580143034458, + 0.016666065901517868, + 0.00923417042940855, + 0.007942993193864822, + 0.005965054966509342, + 0.01053693424910307, + 0.011267436668276787, + 0.004501883871853352, + 0.004576750099658966, + 0.008337202481925488, + 0.004537404049187899, + 0.040661074221134186, + 0.06292238086462021, + 0.010531834326684475, + 0.010747881606221199, + 0.015668435022234917, + 0.010830072686076164, + 0.019032156094908714, + 0.01293192058801651, + 0.014675292186439037, + 0.013751808553934097, + 0.009866309352219105, + 0.014682208187878132, + 0.01145973615348339, + 0.012336269952356815, + 0.015246729366481304, + 0.01041639968752861, + 0.05702671408653259, + 0.1276332288980484, + 0.009297458454966545, + 0.010134540498256683, + 0.0319964773952961, + 0.006781445816159248, + 0.006308509036898613, + 0.024950165301561356, + 0.003069082973524928, + 0.004000406712293625, + 0.0041694650426507, + 0.00413889717310667, + 0.0038013183511793613, + 0.0030815242789685726, + 0.004284268710762262, + 0.004124199505895376, + 0.013752074912190437, + 0.024134863168001175, + 0.01014846283942461, + 0.006589696742594242, + 0.00950226467102766, + 0.003304173704236746, + 0.004301447421312332, + 0.003722768509760499, + 0.004531672690063715, + 0.004708190448582172, + 0.004397479351609945, + 0.009198284707963467, + 0.006428219377994537, + 0.004762172233313322, + 0.006658041849732399, + 0.00536364084109664, + 0.005678210407495499, + 0.006698670331388712, + 0.007958199828863144, + 0.011358101852238178, + 0.005304957740008831, + 0.007322045974433422, + 0.010197375901043415, + 0.01460177544504404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.03194993734359741 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020598890259861946 + ], + "opposite_logprob_tokens": [ + -3.895598888397217 + ], + "logprob_reward": -0.020598890259861946, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020598890259861946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0964234322309494, + 0.04890337586402893, + 0.02038710191845894, + 0.027162054553627968, + 0.16582511365413666, + 0.10098126530647278, + 0.09988170862197876, + 0.401090532541275, + 0.02272721938788891, + 0.016618190333247185 + ], + "credit_map": [ + 0.009086618199944496, + 0.013883627019822598, + 0.012931226752698421, + 0.010968844406306744, + 0.012224409729242325, + 0.009581740014255047, + 0.015225201845169067, + 0.012521760538220406, + 0.004277319181710482, + 0.004906259477138519, + 0.006106437649577856, + 0.009360906668007374, + 0.004437962081283331, + 0.004346086643636227, + 0.009140660986304283, + 0.006327738054096699, + 0.0027276419568806887, + 0.0023438390344381332, + 0.002163961064070463, + 0.0033937525004148483, + 0.0024836387019604445, + 0.002453831722959876, + 0.0023846770636737347, + 0.002435758477076888, + 0.002575822174549103, + 0.003015955211594701, + 0.002611098811030388, + 0.0030659930780529976, + 0.005350869614630938, + 0.0026401060167700052, + 0.005289921537041664, + 0.0026122883427888155, + 0.0221201553940773, + 0.02001545950770378, + 0.018416890874505043, + 0.02390516735613346, + 0.01784275285899639, + 0.01458947267383337, + 0.030266227200627327, + 0.018668992444872856, + 0.020637275651097298, + 0.029483772814273834, + 0.007237699348479509, + 0.006530134938657284, + 0.013483871705830097, + 0.005024365149438381, + 0.0053388811647892, + 0.013245265930891037, + 0.008974899537861347, + 0.00882582738995552, + 0.014703395776450634, + 0.019761335104703903, + 0.01103528868407011, + 0.01023118570446968, + 0.01755988784134388, + 0.008789890445768833, + 0.11019985377788544, + 0.14740793406963348, + 0.05415760353207588, + 0.018192898482084274, + 0.03628763183951378, + 0.01101896446198225, + 0.011398524045944214, + 0.012427128851413727, + 0.001800353522412479, + 0.0037755591329187155, + 0.0026527743320912123, + 0.0036704472731798887, + 0.0030845901928842068, + 0.002753496402874589, + 0.0024407089222222567, + 0.0025492908898741007, + 0.0017380041535943747, + 0.0026183153968304396, + 0.003082942683249712, + 0.0028240594547241926, + 0.0018447149777784944, + 0.0015910293441265821, + 0.001206607325002551, + 0.0017125171143561602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.020598890259861946 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043057121336460114 + ], + "opposite_logprob_tokens": [ + -3.1680572032928467 + ], + "logprob_reward": -0.043057121336460114, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043057121336460114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2020881026983261, + 0.0715378075838089, + 0.04798426106572151, + 0.03356511518359184, + 0.43782171607017517, + 0.11441397666931152, + 0.010144068859517574, + 0.04168989509344101, + 0.017362594604492188, + 0.02339247614145279 + ], + "credit_map": [ + 0.0202163252979517, + 0.022011971101164818, + 0.02903617173433304, + 0.025700805708765984, + 0.018175801262259483, + 0.016842108219861984, + 0.03965921327471733, + 0.03044571727514267, + 0.011804234236478806, + 0.008415397256612778, + 0.007162217050790787, + 0.0060016983188688755, + 0.014247034676373005, + 0.007491708733141422, + 0.006431841291487217, + 0.009983675554394722, + 0.005774589255452156, + 0.0044755577109754086, + 0.005913201253861189, + 0.008990569971501827, + 0.006549084559082985, + 0.004763964097946882, + 0.0070730047300457954, + 0.004444289021193981, + 0.0030955348629504442, + 0.0071850852109491825, + 0.004014633130282164, + 0.0029113648924976587, + 0.004797174595296383, + 0.0028289039619266987, + 0.004854039754718542, + 0.00387838133610785, + 0.05357028916478157, + 0.04830854386091232, + 0.044548019766807556, + 0.06913778185844421, + 0.050769295543432236, + 0.05311296135187149, + 0.07017634809017181, + 0.04819846153259277, + 0.017772648483514786, + 0.047143083065748215, + 0.005260498728603125, + 0.004302669782191515, + 0.026778215542435646, + 0.005042525939643383, + 0.0036657555028796196, + 0.004448580555617809, + 0.0007981202215887606, + 0.0011935000075027347, + 0.0012590497499331832, + 0.001900815055705607, + 0.0011724971700459719, + 0.000997464987449348, + 0.0017715829890221357, + 0.0010510386200621724, + 0.008538485504686832, + 0.014053712598979473, + 0.005435302387923002, + 0.0027015404775738716, + 0.005588889587670565, + 0.0018391885096207261, + 0.0016992693999782205, + 0.0018335062777623534, + 0.0019745053723454475, + 0.005672079976648092, + 0.0021344912238419056, + 0.0018733831821009517, + 0.0014466774882748723, + 0.0014445899287238717, + 0.0015309264417737722, + 0.0012859414564445615, + 0.003519732039421797, + 0.0020464507397264242, + 0.003873599460348487, + 0.003610225161537528, + 0.0025602555833756924, + 0.0027692699804902077, + 0.002308512572199106, + 0.002704429207369685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.043057121336460114 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007635101210325956 + ], + "opposite_logprob_tokens": [ + -4.882635116577148 + ], + "logprob_reward": -0.007635101210325956, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007635101210325956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08176690340042114, + 0.06364484876394272, + 0.07223491370677948, + 0.08614188432693481, + 0.13445179164409637, + 0.231151282787323, + 0.04235996678471565, + 0.15618424117565155, + 0.08510814607143402, + 0.04695593938231468 + ], + "credit_map": [ + 0.010804915800690651, + 0.009315110743045807, + 0.009332102723419666, + 0.008568353950977325, + 0.009329397231340408, + 0.008319294080138206, + 0.018855486065149307, + 0.007242245599627495, + 0.007904702797532082, + 0.007003791630268097, + 0.00825247261673212, + 0.010184288956224918, + 0.007450413424521685, + 0.005642465315759182, + 0.008894248865544796, + 0.008312470279633999, + 0.0062464214861392975, + 0.009100480005145073, + 0.010175074450671673, + 0.011147523298859596, + 0.01129392720758915, + 0.008335388265550137, + 0.009029445238411427, + 0.006906656548380852, + 0.017290445044636726, + 0.016891274601221085, + 0.011446024291217327, + 0.00850163958966732, + 0.011286037042737007, + 0.00645021116361022, + 0.007953695952892303, + 0.0063225580379366875, + 0.013582197949290276, + 0.02035554125905037, + 0.014620579779148102, + 0.0157045666128397, + 0.016120940446853638, + 0.014734522439539433, + 0.022814542055130005, + 0.01651892066001892, + 0.04269378259778023, + 0.10363748669624329, + 0.008081570267677307, + 0.008851594291627407, + 0.022084731608629227, + 0.006692358758300543, + 0.006044831592589617, + 0.03306493163108826, + 0.0035075685009360313, + 0.006107160821557045, + 0.005735873244702816, + 0.004906975198537111, + 0.006233172491192818, + 0.004720603581517935, + 0.006375268567353487, + 0.004773346707224846, + 0.041498418897390366, + 0.04835175350308418, + 0.022241588681936264, + 0.009866522625088692, + 0.013384906575083733, + 0.006426686886698008, + 0.007248483132570982, + 0.007165892980992794, + 0.0176492128521204, + 0.024766601622104645, + 0.007253113202750683, + 0.010458453558385372, + 0.007639996707439423, + 0.004368963185697794, + 0.006682934705168009, + 0.00628887303173542, + 0.004563075955957174, + 0.004800678696483374, + 0.008020197041332722, + 0.007804420776665211, + 0.0035498624201864004, + 0.0036814017221331596, + 0.0063484408892691135, + 0.008187863975763321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.007635101210325956 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28133079409599304 + ], + "opposite_logprob_tokens": [ + -1.4063308238983154 + ], + "logprob_reward": -0.28133079409599304, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28133079409599304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02749193273484707, + 0.013239264488220215, + 0.013760499656200409, + 0.012074653059244156, + 0.15788070857524872, + 0.29672303795814514, + 0.025787435472011566, + 0.4317421615123749, + 0.011203890666365623, + 0.010096462443470955 + ], + "credit_map": [ + 0.003000888042151928, + 0.0032514759805053473, + 0.004319460596889257, + 0.0031943246722221375, + 0.003178413026034832, + 0.0028808405622839928, + 0.004466808866709471, + 0.0031997207552194595, + 0.0018499394645914435, + 0.001505087362602353, + 0.001802452839910984, + 0.001698192092590034, + 0.001397832646034658, + 0.0016550755826756358, + 0.0017458040965721011, + 0.0015848798211663961, + 0.0015171482227742672, + 0.001251712441444397, + 0.0016377633437514305, + 0.0025908537209033966, + 0.001976676983758807, + 0.0014242208562791348, + 0.0018042880110442638, + 0.00155783595982939, + 0.0020911048632115126, + 0.002858804538846016, + 0.0013079295167699456, + 0.001414385624229908, + 0.001135489554144442, + 0.0008350575808435678, + 0.001317427959293127, + 0.0011144537711516023, + 0.019779378548264503, + 0.01651003770530224, + 0.016217472031712532, + 0.02366582304239273, + 0.018161671236157417, + 0.017539920285344124, + 0.028491923585534096, + 0.017514491453766823, + 0.0640246570110321, + 0.13627178966999054, + 0.01104094460606575, + 0.011383763514459133, + 0.03506939113140106, + 0.0077166128903627396, + 0.007556809112429619, + 0.023659072816371918, + 0.0034810130018740892, + 0.002602014923468232, + 0.002468330319970846, + 0.002673449693247676, + 0.005753304343670607, + 0.002591047203168273, + 0.0036155430134385824, + 0.002602733438834548, + 0.11661176383495331, + 0.14726434648036957, + 0.06875737756490707, + 0.021136337891221046, + 0.041292279958724976, + 0.012189998291432858, + 0.01133993174880743, + 0.01315012015402317, + 0.0018824019934982061, + 0.002164555946364999, + 0.000863972760271281, + 0.0016817671712487936, + 0.0009528737864457071, + 0.0012045623734593391, + 0.0012565688230097294, + 0.0011971881613135338, + 0.001443783869035542, + 0.0009266390698030591, + 0.0021473367232829332, + 0.001760146813467145, + 0.0011439835652709007, + 0.001042086980305612, + 0.0007607694715261459, + 0.0008717168238945305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.28133079409599304 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro lens, the camera held still, focuses perfectly on a slowly rotating bracelet. The bracelet is composed of a string of nearly perfectly spherical, soft pearls, each radiating a soft, milky white luster and silky smooth, artfully connected by a nearly invisible thread. Against a deep, blurred black backdrop, the bracelet rotates steadily and gracefully around its central axis. As it moves, a dreamy, soft light from an unidentified source casts a dazzling, fairytale-like brilliance on the pearls—pale pink, sky blue, mint green, and light gold gently flow, blend, and transform across the smooth curves of each pearl, creating hazy, soft halos of light. This slow movement and the shifting light and shadows create a sense of tranquility, dreaminess, and exquisite refinement, as if a treasure from a magical kingdom slowly reveals its mysterious and captivating charm.", + "seed": 14678, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_3/index_4655_seed_14678_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08490516692399978, + "vlm_logprob": -0.08490516692399978, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.08490516692399978, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03963564336299896 + ], + "opposite_logprob_tokens": [ + -3.41463565826416 + ], + "logprob_reward": -0.03963564336299896, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03963564336299896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3074515163898468, + 0.07002774626016617, + 0.05251204967498779, + 0.3135870695114136, + 0.04711548238992691, + 0.03710542619228363, + 0.026239538565278053, + 0.03530476987361908, + 0.07523676753044128, + 0.03541967645287514 + ], + "credit_map": [ + 0.03331606835126877, + 0.049951668828725815, + 0.01960456743836403, + 0.016259901225566864, + 0.08896428346633911, + 0.01957845687866211, + 0.029320091009140015, + 0.05045647919178009, + 0.01155047956854105, + 0.01011284813284874, + 0.005353644955903292, + 0.010120010003447533, + 0.012081644497811794, + 0.002923427615314722, + 0.005238950718194246, + 0.012646735645830631, + 0.005037077236920595, + 0.008380038663744926, + 0.0061312392354011536, + 0.0068063559010624886, + 0.007791813928633928, + 0.0037880530580878258, + 0.0070758359506726265, + 0.007501631043851376, + 0.05914055183529854, + 0.0365024097263813, + 0.04468601942062378, + 0.023706641048192978, + 0.0324007011950016, + 0.03083854541182518, + 0.02763834036886692, + 0.05867385119199753, + 0.002808829303830862, + 0.013035790994763374, + 0.004855652805417776, + 0.002804266754537821, + 0.008129335939884186, + 0.0035199555568397045, + 0.006808409933000803, + 0.005153237842023373, + 0.0028025652281939983, + 0.004607013892382383, + 0.00501715624704957, + 0.004263994749635458, + 0.004187626298516989, + 0.005934722721576691, + 0.004348920192569494, + 0.005943427328020334, + 0.0031349114142358303, + 0.00423624599352479, + 0.0028829684015363455, + 0.002814606297761202, + 0.0035430733114480972, + 0.0018491550581529737, + 0.0031337030231952667, + 0.004644873086363077, + 0.00257616164162755, + 0.0035830787383019924, + 0.0044695427641272545, + 0.0035613998770713806, + 0.0034097633324563503, + 0.006641050335019827, + 0.0040396349504590034, + 0.007024134509265423, + 0.0067427633330225945, + 0.009906095452606678, + 0.010335993953049183, + 0.011018473654985428, + 0.01898772083222866, + 0.005753924138844013, + 0.008003568276762962, + 0.004488224629312754, + 0.0024887945037335157, + 0.004131350666284561, + 0.003728328039869666, + 0.003647836158052087, + 0.004249473102390766, + 0.004816023632884026, + 0.003940527327358723, + 0.008417341858148575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.03963564336299896 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012562076561152935 + ], + "opposite_logprob_tokens": [ + -4.387562274932861 + ], + "logprob_reward": -0.012562076561152935, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012562076561152935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15548500418663025, + 0.06017856299877167, + 0.08897712826728821, + 0.14064638316631317, + 0.0259528961032629, + 0.07366430014371872, + 0.055731430649757385, + 0.13442948460578918, + 0.16860803961753845, + 0.0963268056511879 + ], + "credit_map": [ + 0.014558014459908009, + 0.02455674111843109, + 0.0140583710744977, + 0.02290048636496067, + 0.03466900438070297, + 0.009700263850390911, + 0.015200016088783741, + 0.019842106848955154, + 0.00959761068224907, + 0.008108935318887234, + 0.006633535027503967, + 0.006573864724487066, + 0.010384775698184967, + 0.0036167402286082506, + 0.004264672286808491, + 0.010998427867889404, + 0.015312540344893932, + 0.011294580064713955, + 0.016884157434105873, + 0.008113705553114414, + 0.010867048986256123, + 0.005162185523658991, + 0.005271787755191326, + 0.016071118414402008, + 0.029289165511727333, + 0.018284734338521957, + 0.016902435570955276, + 0.00965133961290121, + 0.014856603927910328, + 0.014023534953594208, + 0.013340272009372711, + 0.0242982916533947, + 0.0029409986454993486, + 0.002776225097477436, + 0.006094258278608322, + 0.0024684269446879625, + 0.0026565524749457836, + 0.002413594163954258, + 0.003188962349668145, + 0.0034138758201152086, + 0.0066964891739189625, + 0.007815444841980934, + 0.006746839266270399, + 0.009040172211825848, + 0.007943344302475452, + 0.01591569371521473, + 0.0062901731580495834, + 0.013216146267950535, + 0.006391643080860376, + 0.009549813345074654, + 0.005727089010179043, + 0.005614335648715496, + 0.0060239993035793304, + 0.004178638570010662, + 0.005888334941118956, + 0.012357579544186592, + 0.009505740366876125, + 0.015394319780170918, + 0.017657676711678505, + 0.014149614609777927, + 0.011956656351685524, + 0.02603946626186371, + 0.014217640273272991, + 0.025508366525173187, + 0.022830136120319366, + 0.034833669662475586, + 0.00890352949500084, + 0.006489410065114498, + 0.06763569265604019, + 0.010062545537948608, + 0.008848262950778008, + 0.009004786610603333, + 0.010518064722418785, + 0.012531223706901073, + 0.010602235794067383, + 0.017341546714305878, + 0.011976951733231544, + 0.008190429769456387, + 0.010718932375311852, + 0.014447418041527271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.012562076561152935 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081933043897152 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081933043897152, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081933043897152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09143911302089691, + 0.0558447390794754, + 0.03748147189617157, + 0.3610084652900696, + 0.29004985094070435, + 0.028876421973109245, + 0.019489850848913193, + 0.017801422625780106, + 0.05474686622619629, + 0.043261803686618805 + ], + "credit_map": [ + 0.00669776601716876, + 0.013669300824403763, + 0.008206217549741268, + 0.016778241842985153, + 0.017400866374373436, + 0.006199606228619814, + 0.010339139960706234, + 0.012147976085543633, + 0.007641926873475313, + 0.011459356173872948, + 0.004466122481971979, + 0.005965109448879957, + 0.011816478334367275, + 0.0030755847692489624, + 0.004203198943287134, + 0.007216959726065397, + 0.004746111109852791, + 0.006517112720757723, + 0.004346025176346302, + 0.005331556778401136, + 0.005569768603891134, + 0.0023561674170196056, + 0.002978087170049548, + 0.005636641290038824, + 0.07104691863059998, + 0.054432064294815063, + 0.044322576373815536, + 0.02645639143884182, + 0.03962705284357071, + 0.03452228754758835, + 0.02939595840871334, + 0.061205197125673294, + 0.01440227497369051, + 0.08962827175855637, + 0.0573313944041729, + 0.013752340339124203, + 0.05143928527832031, + 0.01693839579820633, + 0.016670837998390198, + 0.029887063428759575, + 0.002346333581954241, + 0.0033916912507265806, + 0.004460497293621302, + 0.0024281085934489965, + 0.003010848304256797, + 0.005037190392613411, + 0.0033987448550760746, + 0.004803008865565062, + 0.001808211556635797, + 0.0033293976448476315, + 0.0020736283622682095, + 0.002655313815921545, + 0.0021434479858726263, + 0.001500556943938136, + 0.002994443755596876, + 0.0029848506674170494, + 0.0019294865196570754, + 0.0020320212934166193, + 0.002145931823179126, + 0.0018429916817694902, + 0.001997292973101139, + 0.0024417920503765345, + 0.002300963504239917, + 0.003110942430794239, + 0.0067794290371239185, + 0.009073224849998951, + 0.005740846507251263, + 0.0036389955785125494, + 0.020082445815205574, + 0.002927257912233472, + 0.003489789320155978, + 0.003014876041561365, + 0.003701557405292988, + 0.004908022470772266, + 0.004206696525216103, + 0.0073684570379555225, + 0.005609676707535982, + 0.006229477468878031, + 0.004293368197977543, + 0.006944550201296806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.011081933043897152 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550069019198418 + ], + "opposite_logprob_tokens": [ + -4.387549877166748 + ], + "logprob_reward": -0.012550069019198418, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550069019198418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23637181520462036, + 0.04110682010650635, + 0.05161486938595772, + 0.09456026554107666, + 0.047786980867385864, + 0.05190501734614372, + 0.04148845374584198, + 0.0849505364894867, + 0.2820194959640503, + 0.06819575279951096 + ], + "credit_map": [ + 0.020395006984472275, + 0.03855746611952782, + 0.019654463976621628, + 0.02513975277543068, + 0.05558529868721962, + 0.017651893198490143, + 0.02683456987142563, + 0.03255336731672287, + 0.005317910108715296, + 0.006816759239882231, + 0.004703851882368326, + 0.006387437228113413, + 0.007605766411870718, + 0.0020418015774339437, + 0.003108338685706258, + 0.005124955903738737, + 0.007949920371174812, + 0.006370735354721546, + 0.011256693862378597, + 0.006063978653401136, + 0.006487824488431215, + 0.0033573703840374947, + 0.003156255232170224, + 0.006972089409828186, + 0.01245604082942009, + 0.01840747520327568, + 0.013793636113405228, + 0.0078088752925395966, + 0.012465079315006733, + 0.005553241353482008, + 0.008446726948022842, + 0.01562919281423092, + 0.0030250302515923977, + 0.013621535152196884, + 0.006568621844053268, + 0.0028015910647809505, + 0.008393709547817707, + 0.0039643794298172, + 0.004130382556468248, + 0.00528173241764307, + 0.004717317875474691, + 0.01031112764030695, + 0.005351170431822538, + 0.004761775955557823, + 0.006583975162357092, + 0.004270119592547417, + 0.006762496195733547, + 0.009147034026682377, + 0.004125907551497221, + 0.007338274735957384, + 0.004395710304379463, + 0.005284823477268219, + 0.005382719449698925, + 0.004302493296563625, + 0.003622009651735425, + 0.007036514580249786, + 0.005487632472068071, + 0.010762224905192852, + 0.00740350317209959, + 0.007851413451135159, + 0.009652421809732914, + 0.016385117545723915, + 0.008344903588294983, + 0.019063320010900497, + 0.03495215252041817, + 0.054965294897556305, + 0.01979668438434601, + 0.01254523266106844, + 0.10858777165412903, + 0.017488835379481316, + 0.016149461269378662, + 0.017534060403704643, + 0.006227452773600817, + 0.008150092326104641, + 0.008586524054408073, + 0.008644094690680504, + 0.014377634041011333, + 0.006284037604928017, + 0.006197765003889799, + 0.009728155098855495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.012550069019198418 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34869611263275146 + ], + "opposite_logprob_tokens": [ + -1.2236961126327515 + ], + "logprob_reward": -0.34869611263275146, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34869611263275146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1687476933002472, + 0.1041765883564949, + 0.0821225568652153, + 0.20723506808280945, + 0.08520247042179108, + 0.13400493562221527, + 0.06618946045637131, + 0.033502064645290375, + 0.06639961153268814, + 0.05241961032152176 + ], + "credit_map": [ + 0.017629576846957207, + 0.028256462886929512, + 0.014094428159296513, + 0.013928007334470749, + 0.042772579938173294, + 0.011101482436060905, + 0.016499435529112816, + 0.02446572296321392, + 0.020465504378080368, + 0.012350239790976048, + 0.009370997548103333, + 0.010905452072620392, + 0.014116634614765644, + 0.005758984014391899, + 0.006370035465806723, + 0.02483873814344406, + 0.011339157819747925, + 0.01129074301570654, + 0.014846608974039555, + 0.007551722228527069, + 0.010843598283827305, + 0.005812244489789009, + 0.005941222421824932, + 0.014497259631752968, + 0.03813912719488144, + 0.0254726130515337, + 0.03192746266722679, + 0.01533995009958744, + 0.021342329680919647, + 0.02004956267774105, + 0.019583970308303833, + 0.0353800468146801, + 0.008324556052684784, + 0.019632941111922264, + 0.025839800015091896, + 0.005132079124450684, + 0.010243694297969341, + 0.005902055185288191, + 0.005029303953051567, + 0.005098039750009775, + 0.010837728157639503, + 0.020012501627206802, + 0.01713789440691471, + 0.014721282757818699, + 0.01588132046163082, + 0.020797833800315857, + 0.014813232235610485, + 0.019803127273917198, + 0.00674494169652462, + 0.011585717089474201, + 0.007449348922818899, + 0.007969120517373085, + 0.0068662408739328384, + 0.005238949321210384, + 0.007485740818083286, + 0.012849399819970131, + 0.0031138614285737276, + 0.004786529578268528, + 0.00504453806206584, + 0.004244024865329266, + 0.004214299377053976, + 0.0040016574785113335, + 0.00335711520165205, + 0.004740037489682436, + 0.007175325881689787, + 0.009365014731884003, + 0.01154820341616869, + 0.006463353987783194, + 0.010833506472408772, + 0.005683796014636755, + 0.007497674319893122, + 0.007832739502191544, + 0.00542141729965806, + 0.0065393466502428055, + 0.005511933006346226, + 0.008143305778503418, + 0.007656489498913288, + 0.004529864527285099, + 0.005814362782984972, + 0.008802894502878189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.34869611263275146 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro lens, the camera held still, focuses perfectly on a slowly rotating bracelet. The bracelet is composed of a string of nearly perfectly spherical, soft pearls, each radiating a soft, milky white luster and silky smooth, artfully connected by a nearly invisible thread. Against a deep, blurred black backdrop, the bracelet rotates steadily and gracefully around its central axis. As it moves, a dreamy, soft light from an unidentified source casts a dazzling, fairytale-like brilliance on the pearls—pale pink, sky blue, mint green, and light gold gently flow, blend, and transform across the smooth curves of each pearl, creating hazy, soft halos of light. This slow movement and the shifting light and shadows create a sense of tranquility, dreaminess, and exquisite refinement, as if a treasure from a magical kingdom slowly reveals its mysterious and captivating charm.", + "seed": 14679, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_3/index_4655_seed_14679_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11067433133721352, + "vlm_logprob": -0.11067433133721352, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.11067433133721352, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03833721950650215 + ], + "opposite_logprob_tokens": [ + -3.413337230682373 + ], + "logprob_reward": -0.03833721950650215, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03833721950650215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.460448682308197, + 0.053992029279470444, + 0.036364588886499405, + 0.05629594996571541, + 0.1304006725549698, + 0.15495797991752625, + 0.020736118778586388, + 0.04242788255214691, + 0.01324179396033287, + 0.03113425523042679 + ], + "credit_map": [ + 0.03312233090400696, + 0.13110807538032532, + 0.031085042282938957, + 0.03812869265675545, + 0.0342763215303421, + 0.12418019026517868, + 0.03278543800115585, + 0.03576260805130005, + 0.004875828046351671, + 0.0071465312503278255, + 0.004688973538577557, + 0.007233528885990381, + 0.007398625370115042, + 0.005082865711301565, + 0.009821147657930851, + 0.007744527421891689, + 0.0030625637155026197, + 0.004257287830114365, + 0.003152304096147418, + 0.008520266972482204, + 0.005219181999564171, + 0.0031806661281734705, + 0.004324514418840408, + 0.004647805355489254, + 0.003957153297960758, + 0.005364878568798304, + 0.004183211829513311, + 0.005086950957775116, + 0.005433834623545408, + 0.015603828243911266, + 0.005670925602316856, + 0.010995164513587952, + 0.009746863506734371, + 0.021543212234973907, + 0.017077099531888962, + 0.026021655648946762, + 0.012864574790000916, + 0.007981219328939915, + 0.022033140063285828, + 0.01313291396945715, + 0.01625741459429264, + 0.01750871352851391, + 0.015019157901406288, + 0.022948650643229485, + 0.024799857288599014, + 0.02716284990310669, + 0.0184360109269619, + 0.01282531302422285, + 0.0014804862439632416, + 0.0055304160341620445, + 0.00245693139731884, + 0.0029199053533375263, + 0.0029543316923081875, + 0.001298431889154017, + 0.0025533726438879967, + 0.0015422434080392122, + 0.003912060521543026, + 0.007833963260054588, + 0.005042427685111761, + 0.005466129630804062, + 0.004420859273523092, + 0.005386954173445702, + 0.005224485881626606, + 0.005141000729054213, + 0.0009896622505038977, + 0.001838589902035892, + 0.0012455804971978068, + 0.002273289253935218, + 0.001421686029061675, + 0.001049540936946869, + 0.0028563200030475855, + 0.0015671247383579612, + 0.004323386121541262, + 0.0060289716348052025, + 0.0033574849367141724, + 0.004468538332730532, + 0.0029861521907150745, + 0.0024838827084749937, + 0.004210525657981634, + 0.0032753136474639177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.03833721950650215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009799941442906857 + ], + "opposite_logprob_tokens": [ + -4.634799957275391 + ], + "logprob_reward": -0.009799941442906857, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009799941442906857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2114410102367401, + 0.10071156919002533, + 0.07415107637643814, + 0.10787566751241684, + 0.2053629457950592, + 0.11449912935495377, + 0.04077349230647087, + 0.08577242493629456, + 0.0311068594455719, + 0.028305763378739357 + ], + "credit_map": [ + 0.01620575040578842, + 0.06344334781169891, + 0.012336786836385727, + 0.01819867640733719, + 0.01646980084478855, + 0.054359301924705505, + 0.014422615058720112, + 0.016004744917154312, + 0.013755079358816147, + 0.010948505252599716, + 0.014712955802679062, + 0.012065606191754341, + 0.010584683157503605, + 0.01615273393690586, + 0.010105119086802006, + 0.012386894784867764, + 0.008445530198514462, + 0.017673896625638008, + 0.005961110815405846, + 0.010104338638484478, + 0.00956063624471426, + 0.005379877518862486, + 0.008413371630012989, + 0.00861232541501522, + 0.00644170306622982, + 0.021922467276453972, + 0.007019044365733862, + 0.007541184313595295, + 0.00708414614200592, + 0.03441274166107178, + 0.009938696399331093, + 0.013515687547624111, + 0.01469569280743599, + 0.032186612486839294, + 0.02578963339328766, + 0.03466618433594704, + 0.023096119984984398, + 0.014104914851486683, + 0.03639356419444084, + 0.02443023957312107, + 0.009842030704021454, + 0.014376131817698479, + 0.015876172110438347, + 0.0171595960855484, + 0.020441265776753426, + 0.013874613679945469, + 0.011741144582629204, + 0.011188180185854435, + 0.0032930534798651934, + 0.006083294749259949, + 0.005203723441809416, + 0.00843015220016241, + 0.006093461997807026, + 0.003838615259155631, + 0.00407007522881031, + 0.003761115949600935, + 0.008005419746041298, + 0.017468681558966637, + 0.01111657451838255, + 0.011833486147224903, + 0.008123761042952538, + 0.009878169745206833, + 0.01083371415734291, + 0.008512629196047783, + 0.0021005047019571066, + 0.004642757587134838, + 0.003461401676759124, + 0.006714189890772104, + 0.003409442724660039, + 0.0022466981317847967, + 0.00541859632357955, + 0.0031132702715694904, + 0.0029205919709056616, + 0.004324235953390598, + 0.003608834696933627, + 0.003130869707092643, + 0.0037332766223698854, + 0.0030493822414427996, + 0.0038413198199123144, + 0.0036972544621676207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.009799941442906857 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292127519845963 + ], + "opposite_logprob_tokens": [ + -2.2379212379455566 + ], + "logprob_reward": -0.11292127519845963, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292127519845963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25314661860466003, + 0.14191043376922607, + 0.09661287814378738, + 0.05154397711157799, + 0.17354236543178558, + 0.029515191912651062, + 0.048630207777023315, + 0.11042909324169159, + 0.0710030347108841, + 0.023666229099035263 + ], + "credit_map": [ + 0.01973552443087101, + 0.06279758363962173, + 0.0187852643430233, + 0.02819647081196308, + 0.02322293259203434, + 0.053112439811229706, + 0.027445264160633087, + 0.01985112391412258, + 0.013567819260060787, + 0.017794253304600716, + 0.015187101438641548, + 0.022536220028996468, + 0.024261631071567535, + 0.015201016329228878, + 0.0167534202337265, + 0.016608957201242447, + 0.010773031041026115, + 0.01406081672757864, + 0.00670113880187273, + 0.016302816569805145, + 0.013150523416697979, + 0.009443912655115128, + 0.011239806190133095, + 0.014940820634365082, + 0.003673182101920247, + 0.0050733014941215515, + 0.0029633957892656326, + 0.0034890160895884037, + 0.0034353756345808506, + 0.01849232241511345, + 0.00373225100338459, + 0.010685130953788757, + 0.012177891097962856, + 0.027151821181178093, + 0.02422318235039711, + 0.028848130255937576, + 0.022583942860364914, + 0.013637767173349857, + 0.024644112214446068, + 0.020275520160794258, + 0.002738428534939885, + 0.0035209152847528458, + 0.003327392041683197, + 0.005553217604756355, + 0.0037242623511701822, + 0.0034163552336394787, + 0.004633534234017134, + 0.002601086162030697, + 0.0038518388755619526, + 0.010262388736009598, + 0.005012092646211386, + 0.007301565259695053, + 0.007880502380430698, + 0.0037921457551419735, + 0.006967776920646429, + 0.003561889985576272, + 0.009385725483298302, + 0.02032303810119629, + 0.013490418903529644, + 0.017356332391500473, + 0.01142277754843235, + 0.012597083114087582, + 0.012716922909021378, + 0.013136783614754677, + 0.005261028651148081, + 0.011536911129951477, + 0.007361359428614378, + 0.014029993675649166, + 0.008680207654833794, + 0.00392691558226943, + 0.01537456177175045, + 0.004832053557038307, + 0.0019904624205082655, + 0.006439558696001768, + 0.0018054359825327992, + 0.003572779009118676, + 0.0031240973621606827, + 0.0013968294952064753, + 0.003574744565412402, + 0.0017623205203562975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.11292127519845963 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052584195509552956 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.0052584195509552956, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052584195509552956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3121583163738251, + 0.08638583868741989, + 0.03625984489917755, + 0.19526702165603638, + 0.11791805177927017, + 0.1384732872247696, + 0.041303228586912155, + 0.020678123459219933, + 0.018303366377949715, + 0.03325285762548447 + ], + "credit_map": [ + 0.02192814089357853, + 0.09318189322948456, + 0.014149460010230541, + 0.024987488985061646, + 0.02661559358239174, + 0.07841742038726807, + 0.03239249810576439, + 0.02048584446310997, + 0.010369664989411831, + 0.010329141281545162, + 0.009835210628807545, + 0.009522173553705215, + 0.009074341505765915, + 0.015468623489141464, + 0.011111089959740639, + 0.010675601661205292, + 0.004566926974803209, + 0.0053414711728692055, + 0.003514116397127509, + 0.006558662746101618, + 0.004007246810942888, + 0.0029541507828980684, + 0.0036949878558516502, + 0.005622282158583403, + 0.010774181224405766, + 0.058894168585538864, + 0.00955758523195982, + 0.012138214893639088, + 0.016429444774985313, + 0.0567890964448452, + 0.015457477420568466, + 0.015226876363158226, + 0.009084586054086685, + 0.018620213493704796, + 0.014631939120590687, + 0.018734943121671677, + 0.012625308707356453, + 0.008496248163282871, + 0.021669605746865273, + 0.014055217616260052, + 0.013940030708909035, + 0.01150561310350895, + 0.013091854751110077, + 0.027265725657343864, + 0.022106636315584183, + 0.026241008192300797, + 0.015318996272981167, + 0.009003432467579842, + 0.003029304090887308, + 0.006942656822502613, + 0.00528273731470108, + 0.007102382835000753, + 0.00605572247877717, + 0.004096135962754488, + 0.005270692985504866, + 0.0035235995892435312, + 0.0017373479204252362, + 0.0051312800496816635, + 0.002280264161527157, + 0.002760317176580429, + 0.0024798132944852114, + 0.0021058947313576937, + 0.0022204346023499966, + 0.001962774433195591, + 0.0013426110381260514, + 0.002362386789172888, + 0.0016317483969032764, + 0.0034611299633979797, + 0.0022165030241012573, + 0.0016647493466734886, + 0.0034145459067076445, + 0.002209693193435669, + 0.0033991504460573196, + 0.006255879998207092, + 0.0032223889138549566, + 0.003814270254224539, + 0.004024198278784752, + 0.002865892369300127, + 0.005198279395699501, + 0.004472802393138409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0052584195509552956 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38705480098724365 + ], + "opposite_logprob_tokens": [ + -1.1370548009872437 + ], + "logprob_reward": -0.38705480098724365, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38705480098724365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1222158893942833, + 0.04911551997065544, + 0.20476409792900085, + 0.10447581112384796, + 0.14445725083351135, + 0.16225872933864594, + 0.051073621958494186, + 0.06370843201875687, + 0.061997342854738235, + 0.03593336045742035 + ], + "credit_map": [ + 0.010918453335762024, + 0.022110380232334137, + 0.009103522635996342, + 0.0194599237293005, + 0.014285348355770111, + 0.017111143097281456, + 0.015216488391160965, + 0.014010625891387463, + 0.0047094752080738544, + 0.006975584197789431, + 0.005217296537011862, + 0.008433621376752853, + 0.005825587548315525, + 0.0062850345857441425, + 0.0070953876711428165, + 0.004573536571115255, + 0.01719144731760025, + 0.028232889249920845, + 0.015196588821709156, + 0.03687599301338196, + 0.03679940104484558, + 0.016368018463253975, + 0.026998378336429596, + 0.02710137329995632, + 0.007461437955498695, + 0.037647102028131485, + 0.005691769067198038, + 0.008033178746700287, + 0.013293694704771042, + 0.0163850300014019, + 0.008059624582529068, + 0.007903970777988434, + 0.014153996482491493, + 0.023510482162237167, + 0.02130529284477234, + 0.018200794234871864, + 0.019501924514770508, + 0.011707059107720852, + 0.019116662442684174, + 0.01696104370057583, + 0.01742420345544815, + 0.016971558332443237, + 0.013761473819613457, + 0.020565234124660492, + 0.034771885722875595, + 0.03368204087018967, + 0.013881896622478962, + 0.011200430803000927, + 0.004117156378924847, + 0.01046181283891201, + 0.006738897413015366, + 0.00829540565609932, + 0.006699732970446348, + 0.00451891403645277, + 0.006089424714446068, + 0.004152280278503895, + 0.0074693202041089535, + 0.007516780402511358, + 0.007652094587683678, + 0.011310161091387272, + 0.005958296824246645, + 0.006716951262205839, + 0.007756057661026716, + 0.009328772313892841, + 0.004498734604567289, + 0.008816361427307129, + 0.006911843083798885, + 0.012280823662877083, + 0.006827166303992271, + 0.003602422308176756, + 0.0139768710359931, + 0.005083119031041861, + 0.004937368910759687, + 0.006734795868396759, + 0.003732371376827359, + 0.004432233516126871, + 0.004057569894939661, + 0.0035291905514895916, + 0.0051529197953641415, + 0.0033569103106856346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.38705480098724365 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro lens, the camera held still, focuses perfectly on a slowly rotating bracelet. The bracelet is composed of a string of nearly perfectly spherical, soft pearls, each radiating a soft, milky white luster and silky smooth, artfully connected by a nearly invisible thread. Against a deep, blurred black backdrop, the bracelet rotates steadily and gracefully around its central axis. As it moves, a dreamy, soft light from an unidentified source casts a dazzling, fairytale-like brilliance on the pearls—pale pink, sky blue, mint green, and light gold gently flow, blend, and transform across the smooth curves of each pearl, creating hazy, soft halos of light. This slow movement and the shifting light and shadows create a sense of tranquility, dreaminess, and exquisite refinement, as if a treasure from a magical kingdom slowly reveals its mysterious and captivating charm.", + "seed": 14680, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_3/index_4655_seed_14680_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08643921306356787, + "vlm_logprob": -0.08643921306356787, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.08643921306356787, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15007920563220978 + ], + "opposite_logprob_tokens": [ + -2.0250792503356934 + ], + "logprob_reward": -0.15007920563220978, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.15007920563220978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17253835499286652, + 0.08471455425024033, + 0.03286172077059746, + 0.02862127684056759, + 0.4434506893157959, + 0.02136625163257122, + 0.04501112177968025, + 0.11180076748132706, + 0.03739352151751518, + 0.022241685539484024 + ], + "credit_map": [ + 0.01100889965891838, + 0.018055375665426254, + 0.013816136866807938, + 0.02901785634458065, + 0.026790181174874306, + 0.0131309125572443, + 0.046188659965991974, + 0.014530358836054802, + 0.007655865978449583, + 0.018008267506957054, + 0.011305596679449081, + 0.010623758658766747, + 0.005994897335767746, + 0.012583817355334759, + 0.009316752664744854, + 0.009225605055689812, + 0.002593410201370716, + 0.0049128844402730465, + 0.003531290218234062, + 0.005170128773897886, + 0.003150422591716051, + 0.004413641057908535, + 0.005895093083381653, + 0.0031948520336300135, + 0.0032724018674343824, + 0.0025501453783363104, + 0.00295007205568254, + 0.005365800578147173, + 0.0030492760706692934, + 0.002280081855133176, + 0.004140023607760668, + 0.005013477988541126, + 0.02581324800848961, + 0.09716272354125977, + 0.031007379293441772, + 0.09383004158735275, + 0.06229054182767868, + 0.05469289794564247, + 0.042425788938999176, + 0.03622809052467346, + 0.00271125347353518, + 0.0019892603158950806, + 0.0035391515120863914, + 0.002131492830812931, + 0.0017920740647241473, + 0.002827448770403862, + 0.002031114650890231, + 0.004344460088759661, + 0.003889506682753563, + 0.007123497314751148, + 0.006318095605820417, + 0.009659834206104279, + 0.0057397931814193726, + 0.00409291498363018, + 0.0035482391249388456, + 0.0046392460353672504, + 0.01097782701253891, + 0.011468872427940369, + 0.016500478610396385, + 0.01511415932327509, + 0.014849290251731873, + 0.012744560837745667, + 0.010686703957617283, + 0.01945888064801693, + 0.004174800124019384, + 0.0026776434388011694, + 0.003078380599617958, + 0.004275722429156303, + 0.004580074455589056, + 0.010571287013590336, + 0.004507295787334442, + 0.0035283220931887627, + 0.0015609494876116514, + 0.002830588724464178, + 0.005093907006084919, + 0.004691782873123884, + 0.0018943754257634282, + 0.0015453811502084136, + 0.002184448763728142, + 0.002440254669636488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.15007920563220978 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012556660920381546 + ], + "opposite_logprob_tokens": [ + -4.387556552886963 + ], + "logprob_reward": -0.012556660920381546, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012556660920381546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25657421350479126, + 0.4064081609249115, + 0.11852476000785828, + 0.033879514783620834, + 0.08481307327747345, + 0.031105080619454384, + 0.017614169046282768, + 0.016897747293114662, + 0.022206060588359833, + 0.01197721902281046 + ], + "credit_map": [ + 0.012679974548518658, + 0.05907619372010231, + 0.014242653734982014, + 0.028194982558488846, + 0.026125280186533928, + 0.05371265858411789, + 0.04844006523489952, + 0.014102386310696602, + 0.018702777102589607, + 0.1641795039176941, + 0.00990199763327837, + 0.017069263383746147, + 0.021674033254384995, + 0.14332379400730133, + 0.014551946893334389, + 0.017004813998937607, + 0.005826449953019619, + 0.04412365332245827, + 0.002805273747071624, + 0.004975229501724243, + 0.006571230478584766, + 0.044702719897031784, + 0.005182026885449886, + 0.004338166676461697, + 0.0031055121216923, + 0.003303375793620944, + 0.004204917699098587, + 0.007576369680464268, + 0.004028218798339367, + 0.0023361272178590298, + 0.004959683399647474, + 0.004365306813269854, + 0.005133150145411491, + 0.020053481683135033, + 0.007400783710181713, + 0.016463996842503548, + 0.009480656124651432, + 0.010456878691911697, + 0.008669533766806126, + 0.007154587656259537, + 0.003080117516219616, + 0.003773905336856842, + 0.00737228337675333, + 0.0030744357500225306, + 0.003890796797350049, + 0.002606171416118741, + 0.002973259659484029, + 0.0043341065756976604, + 0.0017669785302132368, + 0.002461338182911277, + 0.0031543776858597994, + 0.0028726314194500446, + 0.002122529549524188, + 0.0017136689275503159, + 0.0016443044878542423, + 0.0018783373525366187, + 0.0019898770842701197, + 0.002101880731061101, + 0.0024196950253099203, + 0.0016114264726638794, + 0.002031066920608282, + 0.002527352422475815, + 0.001513444003649056, + 0.0027030031196773052, + 0.00202698796056211, + 0.0020280431490391493, + 0.0024828806053847075, + 0.0029913748148828745, + 0.003419589251279831, + 0.0037879652809351683, + 0.002875514328479767, + 0.0025937038008123636, + 0.0009979851311072707, + 0.001665844232775271, + 0.001657879794947803, + 0.0022098971530795097, + 0.0013828885275870562, + 0.000793688406702131, + 0.0019863969646394253, + 0.0012826368911191821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.012556660920381546 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062040217220783234 + ], + "opposite_logprob_tokens": [ + -2.812040328979492 + ], + "logprob_reward": -0.062040217220783234, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062040217220783234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3535362780094147, + 0.10430366545915604, + 0.12048373371362686, + 0.0637756958603859, + 0.1137397289276123, + 0.06652457267045975, + 0.03997860476374626, + 0.058811433613300323, + 0.03142234683036804, + 0.04742397740483284 + ], + "credit_map": [ + 0.015074792318046093, + 0.0954376682639122, + 0.016975870355963707, + 0.03159176558256149, + 0.03334422782063484, + 0.08941057324409485, + 0.05438360944390297, + 0.0173177532851696, + 0.006989388260990381, + 0.021351072937250137, + 0.012899227440357208, + 0.013110209256410599, + 0.006907477043569088, + 0.017314568161964417, + 0.01468463521450758, + 0.01104708295315504, + 0.005555386655032635, + 0.043906114995479584, + 0.003946464508771896, + 0.005567291285842657, + 0.00725445244461298, + 0.04306136444211006, + 0.006265540141612291, + 0.004927116446197033, + 0.008687615394592285, + 0.005802897736430168, + 0.009079360403120518, + 0.008220765739679337, + 0.005334538873285055, + 0.006344415247440338, + 0.010046456009149551, + 0.010259645991027355, + 0.009585444815456867, + 0.020148083567619324, + 0.013719684444367886, + 0.014969363808631897, + 0.023330023512244225, + 0.009104212746024132, + 0.012673398479819298, + 0.01020951196551323, + 0.003746509086340666, + 0.015615612268447876, + 0.02526327222585678, + 0.005590938962996006, + 0.004229841288179159, + 0.004160540644079447, + 0.003877358278259635, + 0.004040495492517948, + 0.003855046583339572, + 0.005611661821603775, + 0.007280285935848951, + 0.006014820653945208, + 0.004795038606971502, + 0.005343223921954632, + 0.0033387744333595037, + 0.0037397483829408884, + 0.004841085523366928, + 0.008741606958210468, + 0.004500135313719511, + 0.006326915230602026, + 0.01376660168170929, + 0.008728403598070145, + 0.006749346386641264, + 0.005157337058335543, + 0.003771004034206271, + 0.0027706013061106205, + 0.0029451020527631044, + 0.0031871770042926073, + 0.00487005989998579, + 0.007484667003154755, + 0.0030137188732624054, + 0.0033800131641328335, + 0.0034495736472308636, + 0.007450714707374573, + 0.005920896306633949, + 0.006966032553464174, + 0.00816232617944479, + 0.003994978033006191, + 0.007413963321596384, + 0.004065488465130329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.062040217220783234 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951303523033857 + ], + "opposite_logprob_tokens": [ + -5.130951404571533 + ], + "logprob_reward": -0.005951303523033857, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951303523033857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2658976912498474, + 0.17261795699596405, + 0.15229180455207825, + 0.015804285183548927, + 0.2503516674041748, + 0.03793393075466156, + 0.018530813977122307, + 0.030592139810323715, + 0.029071489349007607, + 0.026908205822110176 + ], + "credit_map": [ + 0.011985649354755878, + 0.07987181842327118, + 0.013573025353252888, + 0.02114764042198658, + 0.020766528323292732, + 0.07460989058017731, + 0.032569508999586105, + 0.011373660527169704, + 0.010152011178433895, + 0.05454311519861221, + 0.013296009041368961, + 0.013331109657883644, + 0.00872127991169691, + 0.04951406270265579, + 0.012398744933307171, + 0.010661636479198933, + 0.0073820739053189754, + 0.05059000849723816, + 0.0069641065783798695, + 0.009237495250999928, + 0.009053678251802921, + 0.05290641635656357, + 0.008967376314103603, + 0.0071906521916389465, + 0.0020365938544273376, + 0.0014561467105522752, + 0.002297330414876342, + 0.0024620206095278263, + 0.0018253081943839788, + 0.0011436489876359701, + 0.002128227148205042, + 0.002455009613186121, + 0.014656144194304943, + 0.05664198100566864, + 0.019563863053917885, + 0.05151653289794922, + 0.03327474370598793, + 0.02969205193221569, + 0.02406075969338417, + 0.020945606753230095, + 0.0036720172502100468, + 0.0069376933388412, + 0.0046454225666821, + 0.004141208250075579, + 0.005155906081199646, + 0.004652572795748711, + 0.004373408854007721, + 0.004355704877525568, + 0.0015079345321282744, + 0.0023531706538051367, + 0.002879000036045909, + 0.001924034091643989, + 0.0028842429164797068, + 0.0031052252743393183, + 0.0020513127092272043, + 0.0018258949276059866, + 0.00316703412681818, + 0.006008361466228962, + 0.003624698845669627, + 0.0028800617437809706, + 0.004690409637987614, + 0.004328900948166847, + 0.0032693096436560154, + 0.0026233652606606483, + 0.0028164186514914036, + 0.00207344489172101, + 0.002947442000731826, + 0.0037578248884528875, + 0.003884432604536414, + 0.006392762064933777, + 0.004056466277688742, + 0.0031426995992660522, + 0.001636579050682485, + 0.003616447327658534, + 0.0068336534313857555, + 0.004091538488864899, + 0.004308508709073067, + 0.0015458510024473071, + 0.002298440784215927, + 0.0025771912187337875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.005951303523033857 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bracelet composed of nearly perfectly spherical, soft pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bracelet rotating steadily and gracefully around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep, blurred black backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do pale pink, sky blue, mint green, and light gold colors flow and blend across the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still while focusing on the bracelet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156867802143097 + ], + "opposite_logprob_tokens": [ + -1.7015687227249146 + ], + "logprob_reward": -0.20156867802143097, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156867802143097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18163476884365082, + 0.4273507297039032, + 0.10326910018920898, + 0.04990328848361969, + 0.1306227594614029, + 0.024157611653208733, + 0.010805175639688969, + 0.01592189446091652, + 0.037438057363033295, + 0.018896667286753654 + ], + "credit_map": [ + 0.009195206686854362, + 0.04093990847468376, + 0.012010893784463406, + 0.02107101120054722, + 0.020066170021891594, + 0.035191860049963, + 0.032993074506521225, + 0.010166625492274761, + 0.01974003203213215, + 0.17373405396938324, + 0.0100101875141263, + 0.017628725618124008, + 0.02296452783048153, + 0.15072089433670044, + 0.01498634647578001, + 0.01756594516336918, + 0.004796122666448355, + 0.03864668309688568, + 0.002409973181784153, + 0.0047448743134737015, + 0.006281840614974499, + 0.037087734788656235, + 0.00535086914896965, + 0.003950991202145815, + 0.006030591204762459, + 0.0038402737118303776, + 0.006761851254850626, + 0.009245926514267921, + 0.0038076303899288177, + 0.004630263429135084, + 0.008704954758286476, + 0.006881789769977331, + 0.009120074100792408, + 0.014254999347031116, + 0.012647399678826332, + 0.017634527757763863, + 0.0384163074195385, + 0.010581458918750286, + 0.01561633963137865, + 0.012351632118225098, + 0.002594750840216875, + 0.002485855482518673, + 0.0041389125399291515, + 0.002772936597466469, + 0.003088288940489292, + 0.0027982930187135935, + 0.002700149081647396, + 0.0035784239880740643, + 0.0011011958122253418, + 0.0011215992271900177, + 0.0018442560685798526, + 0.0013591789174824953, + 0.0011570610804483294, + 0.00214423518627882, + 0.001087842509150505, + 0.0009898064890876412, + 0.0014823272358626127, + 0.0030340333469212055, + 0.001330011524260044, + 0.001795168500393629, + 0.002871730364859104, + 0.0028114262968301773, + 0.0017706542275846004, + 0.000826542207505554, + 0.0039800298400223255, + 0.0031883628107607365, + 0.003048546612262726, + 0.003565829712897539, + 0.005820335820317268, + 0.010638076812028885, + 0.00395684689283371, + 0.00324002536945045, + 0.0012669373536482453, + 0.003565151710063219, + 0.0034590535797178745, + 0.0017818837659433484, + 0.0037100655026733875, + 0.0013472511200234294, + 0.0017223713221028447, + 0.0020439522340893745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.20156867802143097 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_4/index_tensor([1124, 1124])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_4/index_tensor([1124, 1124])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..610dd105f197d5d148c37cfb76e5fd361b157f02 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_4/index_tensor([1124, 1124])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures a realistic-style scene in a bright, sterile airport check-in area. A woman with thick, shoulder-length blond curls, wearing a crisp white collared shirt that fits her uniform, sits attentively behind a long, matte-gray counter. Her fingers initially move deftly across a black keyboard, her focus fixed on a computer screen just off-screen. She pauses typing, her gaze shifting slightly upward, toward an unseen figure. A hand reaches into the frame, offering her a dark blue passport, which she accepts with steady, practiced movements. Simultaneously, two other people approach from the right side of the frame, walking steadily, just like the next person. The first is a man with neatly combed short brown hair, thin-framed glasses, and a dark gray suit jacket over a light-colored shirt. The second person's face remains invisible, partially visible beside him, their form suggesting a casual posture. The entire scene is illuminated by cool, even overhead fluorescent lights, which reflect off the polished white tile floor, creating a clean, functional atmosphere free of harsh shadows and reinforcing the impersonal efficiency of the modern travel center.", + "seed": 866, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_4/index_1124_seed_866_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.18916687835007906, + "vlm_logprob": -0.18916687835007906, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.18916687835007906, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8280246257781982 + ], + "opposite_logprob_tokens": [ + -0.5780246257781982 + ], + "logprob_reward": -0.8280246257781982, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8280246257781982, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.048265792429447174, + 0.06178145855665207, + 0.015170512720942497, + 0.01013356726616621, + 0.020008282735943794, + 0.03911978378891945, + 0.06346297264099121, + 0.024938587099313736, + 0.6926923394203186, + 0.024426694959402084 + ], + "credit_map": [ + 0.0046888189390301704, + 0.0038395458832383156, + 0.0029905494302511215, + 0.0028737718239426613, + 0.01514644455164671, + 0.003860088065266609, + 0.002145281294360757, + 0.01272130012512207, + 0.008663946762681007, + 0.0025527430698275566, + 0.0025137891061604023, + 0.0014316323213279247, + 0.024080375209450722, + 0.0037184404209256172, + 0.00184435595292598, + 0.016976183280348778, + 0.001737970975227654, + 0.0013613197952508926, + 0.0023904384579509497, + 0.0024218487087637186, + 0.0016675129299983382, + 0.002730483189225197, + 0.0013058383483439684, + 0.001555101596750319, + 0.0008488464518450201, + 0.001159402308985591, + 0.0017062816768884659, + 0.001155263395048678, + 0.001469891401939094, + 0.0019125704420730472, + 0.000705854909028858, + 0.0011754579609259963, + 0.0029151062481105328, + 0.0017800878267735243, + 0.0028498717583715916, + 0.0027117079589515924, + 0.002863822504878044, + 0.002664083382114768, + 0.002553463215008378, + 0.001670141122303903, + 0.0031517399474978447, + 0.004056028556078672, + 0.005356973502784967, + 0.005251819267868996, + 0.005122106522321701, + 0.0044348291121423244, + 0.005758948624134064, + 0.005987344775348902, + 0.012525648809969425, + 0.0053153010085225105, + 0.006073117721825838, + 0.009596739895641804, + 0.006367377936840057, + 0.004439313430339098, + 0.011550527065992355, + 0.0075949556194245815, + 0.0032558320090174675, + 0.004377435427159071, + 0.0026726284995675087, + 0.0025461241602897644, + 0.003004383761435747, + 0.0022555345203727484, + 0.004293567501008511, + 0.002533085411414504, + 0.05472976341843605, + 0.19554665684700012, + 0.03546740114688873, + 0.04411967471241951, + 0.0667482390999794, + 0.23750758171081543, + 0.013969071209430695, + 0.044603947550058365, + 0.0024959901347756386, + 0.0026346766389906406, + 0.0026564395520836115, + 0.003258991055190563, + 0.0026799130719155073, + 0.004913147538900375, + 0.001963369082659483, + 0.0038241699803620577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.8280246257781982 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011096316389739513 + ], + "opposite_logprob_tokens": [ + -4.511096477508545 + ], + "logprob_reward": -0.011096316389739513, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011096316389739513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03769741952419281, + 0.17717473208904266, + 0.033074215054512024, + 0.011985553428530693, + 0.0554657056927681, + 0.05209551006555557, + 0.05574401840567589, + 0.042039427906274796, + 0.5082616209983826, + 0.02646186389029026 + ], + "credit_map": [ + 0.005895037204027176, + 0.004323357716202736, + 0.0019864619243890047, + 0.0028547041583806276, + 0.009399201720952988, + 0.0034646568819880486, + 0.002833409933373332, + 0.006940583232790232, + 0.024305222555994987, + 0.006947382353246212, + 0.0071240635588765144, + 0.0031700178515166044, + 0.06959265470504761, + 0.011003944091498852, + 0.005179419182240963, + 0.04985200613737106, + 0.0037347576580941677, + 0.0031096055172383785, + 0.004586347844451666, + 0.006281005218625069, + 0.0032215071842074394, + 0.004767540842294693, + 0.0037238928489387035, + 0.0036495542153716087, + 0.0009004269377328455, + 0.0018528472865000367, + 0.0009884786559268832, + 0.0011857078643515706, + 0.0018464308232069016, + 0.0023003406822681427, + 0.0015445330645889044, + 0.00136678630951792, + 0.007198636885732412, + 0.004535048734396696, + 0.014705008827149868, + 0.004901904612779617, + 0.005542781203985214, + 0.008196407929062843, + 0.004447120241820812, + 0.005938787013292313, + 0.004173928406089544, + 0.005038314498960972, + 0.00564430421218276, + 0.007129532750695944, + 0.00853627361357212, + 0.007930563762784004, + 0.007157869171351194, + 0.006484720390290022, + 0.012283075600862503, + 0.0051351143047213554, + 0.0056394776329398155, + 0.007266493514180183, + 0.005669038742780685, + 0.005390895064920187, + 0.008356952108442783, + 0.006002963986247778, + 0.006471013184636831, + 0.0053771608509123325, + 0.004991599358618259, + 0.0039780111983418465, + 0.005696313921362162, + 0.005776348989456892, + 0.005595080088824034, + 0.004153893329203129, + 0.039675116539001465, + 0.14313368499279022, + 0.026056678965687752, + 0.03270362690091133, + 0.048925310373306274, + 0.17427195608615875, + 0.01027837023139, + 0.033216800540685654, + 0.0026526942383497953, + 0.0034335784148424864, + 0.0022695392835885286, + 0.003120336914435029, + 0.003081768751144409, + 0.005438830703496933, + 0.002144124824553728, + 0.004320987965911627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.011096316389739513 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058744616806507 + ], + "opposite_logprob_tokens": [ + -3.89558744430542 + ], + "logprob_reward": -0.02058744616806507, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058744616806507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07777582108974457, + 0.19938725233078003, + 0.04047016799449921, + 0.009045531041920185, + 0.013527249917387962, + 0.09912735223770142, + 0.0848190039396286, + 0.09269394725561142, + 0.3512757420539856, + 0.031877901405096054 + ], + "credit_map": [ + 0.00809125229716301, + 0.005806689150631428, + 0.00405091792345047, + 0.003824456362053752, + 0.027315767481923103, + 0.005940344650298357, + 0.0031280345283448696, + 0.019618352875113487, + 0.02781190723180771, + 0.007590422872453928, + 0.008024408482015133, + 0.003750493749976158, + 0.07802366465330124, + 0.012689142487943172, + 0.0055382526479661465, + 0.055958960205316544, + 0.004926757421344519, + 0.004134966991841793, + 0.005044129677116871, + 0.0044003804214298725, + 0.00394288869574666, + 0.004875328857451677, + 0.0054879626259207726, + 0.007657752837985754, + 0.0011198758147656918, + 0.0014567218022421002, + 0.0008399770013056695, + 0.0013511215802282095, + 0.0010086506372317672, + 0.0012925645569339395, + 0.0010457513853907585, + 0.0009308681474067271, + 0.0015805501025170088, + 0.0016319775022566319, + 0.0017513028578832746, + 0.0011611261870712042, + 0.0021185395307838917, + 0.0020419161301106215, + 0.0018987273797392845, + 0.0013431102270260453, + 0.004500510171055794, + 0.004713322501629591, + 0.011094105429947376, + 0.02575579844415188, + 0.007912533357739449, + 0.007752947509288788, + 0.01513318158686161, + 0.022264953702688217, + 0.018993418663740158, + 0.009382928721606731, + 0.010660005733370781, + 0.012453415431082249, + 0.007857102900743484, + 0.0068357945419847965, + 0.007643035613000393, + 0.01099330186843872, + 0.012587027624249458, + 0.016665099188685417, + 0.01167950127273798, + 0.01265767402946949, + 0.009353147819638252, + 0.007795930374413729, + 0.014311224222183228, + 0.007644345052540302, + 0.027463234961032867, + 0.09827137738466263, + 0.017794908955693245, + 0.02400033175945282, + 0.03375890105962753, + 0.11950255930423737, + 0.007378460839390755, + 0.023105984553694725, + 0.0034992704167962074, + 0.004641603212803602, + 0.002488260855898261, + 0.0031314219813793898, + 0.003473491407930851, + 0.004206683021038771, + 0.0048867822624742985, + 0.005550389178097248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.02058744616806507 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305380955338478 + ], + "opposite_logprob_tokens": [ + -3.1680538654327393 + ], + "logprob_reward": -0.04305380955338478, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305380955338478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07858479768037796, + 0.13792766630649567, + 0.06005633622407913, + 0.010549664497375488, + 0.019955793395638466, + 0.04785781726241112, + 0.16798727214336395, + 0.2111637145280838, + 0.16723796725273132, + 0.09867894649505615 + ], + "credit_map": [ + 0.012329593300819397, + 0.01297694817185402, + 0.007226219866424799, + 0.006834440398961306, + 0.012062798254191875, + 0.007948090322315693, + 0.005622456315904856, + 0.013584256172180176, + 0.018680034205317497, + 0.005874230060726404, + 0.006040571723133326, + 0.003398243337869644, + 0.052748363465070724, + 0.008808021433651447, + 0.003967429511249065, + 0.038410771638154984, + 0.006821612827479839, + 0.00836937502026558, + 0.005860971286892891, + 0.0068475897423923016, + 0.00662989541888237, + 0.011277970857918262, + 0.006541525013744831, + 0.0077073900029063225, + 0.001066814293153584, + 0.0016527451807633042, + 0.001452770666219294, + 0.0013470523990690708, + 0.0012267603306099772, + 0.0016791310627013445, + 0.000923653191421181, + 0.0012007381301373243, + 0.0023361858911812305, + 0.0019238970708101988, + 0.004110182635486126, + 0.0021580953616648912, + 0.0027346210554242134, + 0.002982006873935461, + 0.0020297805313020945, + 0.001681023626588285, + 0.004468848928809166, + 0.0035215825773775578, + 0.006907448172569275, + 0.008060038089752197, + 0.005061125382781029, + 0.005474121309816837, + 0.00614262605085969, + 0.008222031407058239, + 0.01868109218776226, + 0.01020809356123209, + 0.010677267797291279, + 0.020741604268550873, + 0.014349402859807014, + 0.015500687062740326, + 0.05282614752650261, + 0.025002971291542053, + 0.0398213192820549, + 0.02914556674659252, + 0.03273340314626694, + 0.021345781162381172, + 0.023872656747698784, + 0.02065414749085903, + 0.021348506212234497, + 0.022242335602641106, + 0.0287616103887558, + 0.038096439093351364, + 0.013055478222668171, + 0.013523651286959648, + 0.013549024239182472, + 0.043591417372226715, + 0.005984143819659948, + 0.01067619863897562, + 0.008976944722235203, + 0.010229537263512611, + 0.013189778663218021, + 0.011934148147702217, + 0.01273120753467083, + 0.020312443375587463, + 0.008709941059350967, + 0.012594942934811115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.04305380955338478 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307219386100769 + ], + "opposite_logprob_tokens": [ + -3.16807222366333 + ], + "logprob_reward": -0.04307219386100769, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307219386100769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.017366234213113785, + 0.06383899599313736, + 0.005409245379269123, + 0.008595099672675133, + 0.014753478579223156, + 0.03439907729625702, + 0.05244070664048195, + 0.02777022495865822, + 0.7508863210678101, + 0.024540556594729424 + ], + "credit_map": [ + 0.0029408142436295748, + 0.0024285921826958656, + 0.001369592035189271, + 0.0016731007490307093, + 0.0024691082071512938, + 0.0029022374656051397, + 0.0015937243588268757, + 0.001989065669476986, + 0.009140197187662125, + 0.002552925841882825, + 0.0025633980985730886, + 0.0012566577643156052, + 0.025125259533524513, + 0.0038953067269176245, + 0.0018724474357441068, + 0.01743280328810215, + 0.0005876509821973741, + 0.0006181783392094076, + 0.0008136699325405061, + 0.0006572141428478062, + 0.0006601169006898999, + 0.0009572150884196162, + 0.00048597221029922366, + 0.0006292276666499674, + 0.0007321269949898124, + 0.0010958599159494042, + 0.0011368823470547795, + 0.0010799492010846734, + 0.0012247401755303144, + 0.0016292539658024907, + 0.0007326077320612967, + 0.0009636798058636487, + 0.002096965443342924, + 0.0014712086413055658, + 0.0017799409106373787, + 0.0020596743561327457, + 0.0020373212173581123, + 0.001973789883777499, + 0.0018390233162790537, + 0.0014955542283132672, + 0.002797530498355627, + 0.0036041405983269215, + 0.004698808770626783, + 0.004492542706429958, + 0.0045501794666051865, + 0.004095330368727446, + 0.005379299633204937, + 0.004781242925673723, + 0.010276006534695625, + 0.004851227160543203, + 0.004801752045750618, + 0.008026068098843098, + 0.00514746131375432, + 0.004666486289352179, + 0.008489863015711308, + 0.006181841250509024, + 0.003934207838028669, + 0.005477182567119598, + 0.0034206777345389128, + 0.0019395395647734404, + 0.0028721524868160486, + 0.002692432841286063, + 0.004703040234744549, + 0.002730992855504155, + 0.05906503275036812, + 0.21183829009532928, + 0.038456033915281296, + 0.04792657867074013, + 0.07214352488517761, + 0.25741609930992126, + 0.015267261303961277, + 0.04877352714538574, + 0.0024179043248295784, + 0.0028465890791267157, + 0.0026448641438037157, + 0.0032153320498764515, + 0.0026111381594091654, + 0.004638751037418842, + 0.002595841186121106, + 0.0035701377782970667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.04307219386100769 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a realistic-style scene in a bright, sterile airport check-in area. A woman with thick, shoulder-length blond curls, wearing a crisp white collared shirt that fits her uniform, sits attentively behind a long, matte-gray counter. Her fingers initially move deftly across a black keyboard, her focus fixed on a computer screen just off-screen. She pauses typing, her gaze shifting slightly upward, toward an unseen figure. A hand reaches into the frame, offering her a dark blue passport, which she accepts with steady, practiced movements. Simultaneously, two other people approach from the right side of the frame, walking steadily, just like the next person. The first is a man with neatly combed short brown hair, thin-framed glasses, and a dark gray suit jacket over a light-colored shirt. The second person's face remains invisible, partially visible beside him, their form suggesting a casual posture. The entire scene is illuminated by cool, even overhead fluorescent lights, which reflect off the polished white tile floor, creating a clean, functional atmosphere free of harsh shadows and reinforcing the impersonal efficiency of the modern travel center.", + "seed": 867, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_4/index_1124_seed_867_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.029570513125509023, + "vlm_logprob": -0.029570513125509023, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.029570513125509023, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10387824475765228 + ], + "opposite_logprob_tokens": [ + -2.3538782596588135 + ], + "logprob_reward": -0.10387824475765228, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10387824475765228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48855042457580566, + 0.045086491852998734, + 0.038625892251729965, + 0.05135515332221985, + 0.06764490902423859, + 0.04074598103761673, + 0.07744928449392319, + 0.10550131648778915, + 0.023128850385546684, + 0.06191172078251839 + ], + "credit_map": [ + 0.06605680286884308, + 0.059506721794605255, + 0.07080887258052826, + 0.05313209444284439, + 0.06539344042539597, + 0.04585658758878708, + 0.07823209464550018, + 0.04956379160284996, + 0.00529815535992384, + 0.005407573655247688, + 0.0045652445405721664, + 0.007064319681376219, + 0.004771693143993616, + 0.0042578186839818954, + 0.007453070022165775, + 0.006268616765737534, + 0.0034861655440181494, + 0.004800401162356138, + 0.004283571615815163, + 0.004455354530364275, + 0.004536167252808809, + 0.0035951444879174232, + 0.00723649887368083, + 0.006232590414583683, + 0.004051445052027702, + 0.005689158104360104, + 0.004083720501512289, + 0.005823231302201748, + 0.00818547047674656, + 0.010818135924637318, + 0.004603452514857054, + 0.00810054037719965, + 0.00707332044839859, + 0.007299579679965973, + 0.0052199107594788074, + 0.006214036140590906, + 0.0067648203112185, + 0.013415982015430927, + 0.011069963686168194, + 0.010587300173938274, + 0.004250365309417248, + 0.0034697665832936764, + 0.0030461521819233894, + 0.00610661506652832, + 0.006171735003590584, + 0.006198089569807053, + 0.0072616096585989, + 0.004241647198796272, + 0.006341163534671068, + 0.007539120502769947, + 0.009416867047548294, + 0.008099483326077461, + 0.014407479204237461, + 0.014768874272704124, + 0.009144630283117294, + 0.007731662131845951, + 0.008425021544098854, + 0.01148515846580267, + 0.01362020242959261, + 0.011866861954331398, + 0.009841744787991047, + 0.023923415690660477, + 0.008260474540293217, + 0.018078427761793137, + 0.0026541338302195072, + 0.0029322353657335043, + 0.002496986649930477, + 0.0033733609598129988, + 0.0026253501418977976, + 0.003308258717879653, + 0.002939458005130291, + 0.0027990678790956736, + 0.006586373317986727, + 0.007594951894134283, + 0.006922456435859203, + 0.005777115002274513, + 0.009098541922867298, + 0.01040717214345932, + 0.01001136563718319, + 0.005513743497431278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.10387824475765228 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036345156840980053 + ], + "opposite_logprob_tokens": [ + -5.628634452819824 + ], + "logprob_reward": -0.0036345156840980053, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036345156840980053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3519134819507599, + 0.14135274291038513, + 0.05885466933250427, + 0.09872779995203018, + 0.08645838499069214, + 0.04010936990380287, + 0.0565953329205513, + 0.0719914585351944, + 0.03839198499917984, + 0.0556047186255455 + ], + "credit_map": [ + 0.05990360304713249, + 0.0442427322268486, + 0.051790229976177216, + 0.0345967598259449, + 0.048932403326034546, + 0.029194990172982216, + 0.05010398104786873, + 0.03314879164099693, + 0.020574459806084633, + 0.016652323305606842, + 0.01682431250810623, + 0.02043994329869747, + 0.0195927694439888, + 0.009237706661224365, + 0.02417793497443199, + 0.013853286392986774, + 0.004585258662700653, + 0.007189646828919649, + 0.005931093357503414, + 0.008349830284714699, + 0.006375174503773451, + 0.0038848668336868286, + 0.015089440159499645, + 0.007449357770383358, + 0.007489840034395456, + 0.011140696704387665, + 0.009774019941687584, + 0.008409414440393448, + 0.014089157804846764, + 0.02336283214390278, + 0.009504628367722034, + 0.014957214705646038, + 0.008968544192612171, + 0.011012114584445953, + 0.006677985191345215, + 0.00990475993603468, + 0.011213255114853382, + 0.01519344374537468, + 0.010336906649172306, + 0.013151378370821476, + 0.00400535250082612, + 0.004860871005803347, + 0.0028567533008754253, + 0.005287417210638523, + 0.005568025633692741, + 0.006656832527369261, + 0.005279793869704008, + 0.005594325251877308, + 0.005033667664974928, + 0.004982610698789358, + 0.00980305951088667, + 0.011084125377237797, + 0.007117959670722485, + 0.0057900031097233295, + 0.0053014932200312614, + 0.007482417859137058, + 0.0057098050601780415, + 0.012880508787930012, + 0.00887810904532671, + 0.010589363984763622, + 0.0099308080971241, + 0.009533727541565895, + 0.007274809759110212, + 0.007194329518824816, + 0.0036139509174972773, + 0.005026195198297501, + 0.003990822937339544, + 0.00522057106718421, + 0.0038957863580435514, + 0.006201498210430145, + 0.005460675805807114, + 0.004982481244951487, + 0.006486518308520317, + 0.007812194526195526, + 0.0062564462423324585, + 0.006629869807511568, + 0.007712078746408224, + 0.007255089934915304, + 0.008063074201345444, + 0.005389448720961809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0036345156840980053 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819397695362568 + ], + "opposite_logprob_tokens": [ + -4.018194198608398 + ], + "logprob_reward": -0.01819397695362568, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819397695362568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17635029554367065, + 0.17194220423698425, + 0.10946443676948547, + 0.32836827635765076, + 0.050358034670352936, + 0.03108699806034565, + 0.03201339766383171, + 0.027253244072198868, + 0.03412308543920517, + 0.039040010422468185 + ], + "credit_map": [ + 0.01603185385465622, + 0.02201440930366516, + 0.02707280032336712, + 0.022191479802131653, + 0.0261594969779253, + 0.017796147614717484, + 0.025997664779424667, + 0.019086448475718498, + 0.018268460407853127, + 0.02142479456961155, + 0.01857345923781395, + 0.02487669698894024, + 0.02064235880970955, + 0.015008973889052868, + 0.03192939609289169, + 0.0212180707603693, + 0.011418500915169716, + 0.015782687813043594, + 0.01019392441958189, + 0.01247909851372242, + 0.01243237778544426, + 0.00891727115958929, + 0.018416674807667732, + 0.01982390694320202, + 0.024858318269252777, + 0.028630342334508896, + 0.046499986201524734, + 0.059902504086494446, + 0.02841857820749283, + 0.0632677972316742, + 0.030977319926023483, + 0.045813415199518204, + 0.004157481715083122, + 0.006182224024087191, + 0.0048132906667888165, + 0.008252170868217945, + 0.00662408908829093, + 0.006231297738850117, + 0.008600398898124695, + 0.005497082136571407, + 0.004101842176169157, + 0.004060149658471346, + 0.003226794768124819, + 0.003923894837498665, + 0.0034606154076755047, + 0.0036567095667123795, + 0.003379969159141183, + 0.005277023650705814, + 0.002778224181383848, + 0.0026433018501847982, + 0.004595960024744272, + 0.007443400099873543, + 0.0036178643349558115, + 0.0036676775198429823, + 0.002954192692413926, + 0.004312779288738966, + 0.0024029880296438932, + 0.003617451526224613, + 0.0033631063997745514, + 0.004345818888396025, + 0.0036841535475105047, + 0.0034224975388497114, + 0.002321943175047636, + 0.00409528287127614, + 0.003176725236698985, + 0.00341040245257318, + 0.0030145803466439247, + 0.004109864588826895, + 0.004821100737899542, + 0.007436669431626797, + 0.004125321749597788, + 0.004028421360999346, + 0.003802600782364607, + 0.00548957847058773, + 0.005668628960847855, + 0.0037706326693296432, + 0.007555106654763222, + 0.004892130848020315, + 0.00494679668918252, + 0.002914536977186799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.01819397695362568 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071558110415936 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.011071558110415936, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071558110415936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2545846700668335, + 0.16562491655349731, + 0.06295839697122574, + 0.09233000874519348, + 0.06218954175710678, + 0.05842551961541176, + 0.038065116852521896, + 0.07206740230321884, + 0.03175167739391327, + 0.1620027869939804 + ], + "credit_map": [ + 0.03530590608716011, + 0.03139348700642586, + 0.03592858090996742, + 0.027112416923046112, + 0.0373687818646431, + 0.02305491454899311, + 0.03984997421503067, + 0.02457059919834137, + 0.016152992844581604, + 0.0194809939712286, + 0.016320355236530304, + 0.028289251029491425, + 0.01823369413614273, + 0.014908046461641788, + 0.03260835260152817, + 0.019631236791610718, + 0.005502702668309212, + 0.008515570312738419, + 0.008394451811909676, + 0.007630466017872095, + 0.008479629643261433, + 0.005055979825556278, + 0.011761486530303955, + 0.007618106435984373, + 0.0072899796068668365, + 0.01314134057611227, + 0.00616398686543107, + 0.008056266233325005, + 0.01122110802680254, + 0.020878462120890617, + 0.009587323293089867, + 0.015991542488336563, + 0.004420205019414425, + 0.005589964333921671, + 0.004229555372148752, + 0.007899198681116104, + 0.007344279438257217, + 0.010102571919560432, + 0.012399244122207165, + 0.01020452193915844, + 0.00662534823641181, + 0.004586752969771624, + 0.003851687302812934, + 0.008226200938224792, + 0.00893950741738081, + 0.008591761812567711, + 0.011080071330070496, + 0.006524188444018364, + 0.002979441313073039, + 0.0028127648402005434, + 0.0053059193305671215, + 0.008591508492827415, + 0.005941296461969614, + 0.00550430454313755, + 0.0030631429981440306, + 0.0038667400367558002, + 0.005797744728624821, + 0.010343206115067005, + 0.0089038060978055, + 0.00921108853071928, + 0.010629844851791859, + 0.01110005471855402, + 0.007591604720801115, + 0.008490058593451977, + 0.002965502208098769, + 0.005274991970509291, + 0.0034766995813697577, + 0.004332901444286108, + 0.005438642576336861, + 0.0045083691366016865, + 0.003023169469088316, + 0.002731400541961193, + 0.015590157359838486, + 0.025157315656542778, + 0.014560740441083908, + 0.017253447324037552, + 0.021806668490171432, + 0.03343930095434189, + 0.019098808988928795, + 0.015096361748874187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.011071558110415936 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074270121753216 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.011074270121753216, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074270121753216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20279058814048767, + 0.11072564870119095, + 0.07177960872650146, + 0.07736166566610336, + 0.07315754890441895, + 0.04787181317806244, + 0.08119118958711624, + 0.17273150384426117, + 0.06101727485656738, + 0.10137315839529037 + ], + "credit_map": [ + 0.039437394589185715, + 0.025866633281111717, + 0.025224437937140465, + 0.020088551566004753, + 0.02505645342171192, + 0.019896885380148888, + 0.022681213915348053, + 0.024539025500416756, + 0.01334584690630436, + 0.012715981341898441, + 0.015243538655340672, + 0.015925956889986992, + 0.00997548084706068, + 0.009165199473500252, + 0.016790112480521202, + 0.017563536763191223, + 0.00491924537345767, + 0.011180751025676727, + 0.00876279640942812, + 0.007707108277827501, + 0.006743850186467171, + 0.006482994183897972, + 0.01198837161064148, + 0.013994496315717697, + 0.0064798565581440926, + 0.008969457820057869, + 0.01259197574108839, + 0.010904201306402683, + 0.011599049903452396, + 0.008200117386877537, + 0.009632533416152, + 0.008984473533928394, + 0.00760256964713335, + 0.011181340552866459, + 0.005086441524326801, + 0.0064347852021455765, + 0.0066656796261668205, + 0.01482405886054039, + 0.010284236632287502, + 0.01107843592762947, + 0.005112299695611, + 0.006480057258158922, + 0.004590997938066721, + 0.006663724780082703, + 0.005664031952619553, + 0.0069342562928795815, + 0.006317752879112959, + 0.0061086867935955524, + 0.00678660674020648, + 0.006374907214194536, + 0.009764563292264938, + 0.010383840650320053, + 0.014666539616882801, + 0.014081072993576527, + 0.00893433578312397, + 0.010199320502579212, + 0.012728788889944553, + 0.0170785803347826, + 0.022342681884765625, + 0.020985303446650505, + 0.020150579512119293, + 0.03820203244686127, + 0.015239681117236614, + 0.02600385807454586, + 0.006964012049138546, + 0.006899240892380476, + 0.006315211765468121, + 0.006581859663128853, + 0.0065336222760379314, + 0.010579121299088001, + 0.008029907010495663, + 0.009114297106862068, + 0.013909506611526012, + 0.009986772201955318, + 0.013317949138581753, + 0.012472006492316723, + 0.012937847524881363, + 0.013670782558619976, + 0.015052086673676968, + 0.01002621278166771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.011074270121753216 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a realistic-style scene in a bright, sterile airport check-in area. A woman with thick, shoulder-length blond curls, wearing a crisp white collared shirt that fits her uniform, sits attentively behind a long, matte-gray counter. Her fingers initially move deftly across a black keyboard, her focus fixed on a computer screen just off-screen. She pauses typing, her gaze shifting slightly upward, toward an unseen figure. A hand reaches into the frame, offering her a dark blue passport, which she accepts with steady, practiced movements. Simultaneously, two other people approach from the right side of the frame, walking steadily, just like the next person. The first is a man with neatly combed short brown hair, thin-framed glasses, and a dark gray suit jacket over a light-colored shirt. The second person's face remains invisible, partially visible beside him, their form suggesting a casual posture. The entire scene is illuminated by cool, even overhead fluorescent lights, which reflect off the polished white tile floor, creating a clean, functional atmosphere free of harsh shadows and reinforcing the impersonal efficiency of the modern travel center.", + "seed": 868, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_4/index_1124_seed_868_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.061180945951491594, + "vlm_logprob": -0.061180945951491594, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.061180945951491594, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23109430074691772 + ], + "opposite_logprob_tokens": [ + -1.6060943603515625 + ], + "logprob_reward": -0.23109430074691772, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23109430074691772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05346524715423584, + 0.12239865213632584, + 0.1503836065530777, + 0.18025697767734528, + 0.24863556027412415, + 0.048555638641119, + 0.09355372190475464, + 0.03294966369867325, + 0.04087701812386513, + 0.028923938050866127 + ], + "credit_map": [ + 0.005918785464018583, + 0.0067304098047316074, + 0.005400069989264011, + 0.006740016862750053, + 0.007486472837626934, + 0.00645481375977397, + 0.0056021991185843945, + 0.009132477454841137, + 0.007929886691272259, + 0.01174888014793396, + 0.0074127898551523685, + 0.009172256104648113, + 0.02903946116566658, + 0.019600607454776764, + 0.009225777350366116, + 0.028268998488783836, + 0.019787438213825226, + 0.016205212101340294, + 0.01772512122988701, + 0.013562357053160667, + 0.019840596243739128, + 0.02191193774342537, + 0.016390236094594002, + 0.024960706010460854, + 0.026074515655636787, + 0.02097628451883793, + 0.02158416621387005, + 0.02113030105829239, + 0.023086773231625557, + 0.03343616798520088, + 0.015589877963066101, + 0.018378889188170433, + 0.06863651424646378, + 0.09541414678096771, + 0.012923753820359707, + 0.007109213620424271, + 0.01712748035788536, + 0.017869900912046432, + 0.012062872759997845, + 0.01749165542423725, + 0.0073718056082725525, + 0.0044397516176104546, + 0.004726366139948368, + 0.004505055490881205, + 0.005653521046042442, + 0.01141051109880209, + 0.005181526765227318, + 0.00526710320264101, + 0.010210844688117504, + 0.007443368434906006, + 0.01720728538930416, + 0.015389787033200264, + 0.014196676202118397, + 0.009045273065567017, + 0.013772308826446533, + 0.0062881833873689175, + 0.0025154876057058573, + 0.005517095793038607, + 0.004064148757606745, + 0.003949630539864302, + 0.0028827651403844357, + 0.00614301860332489, + 0.0050535728223621845, + 0.0028239425737410784, + 0.004292360506951809, + 0.004444540943950415, + 0.006353919859975576, + 0.004438519943505526, + 0.0048329755663871765, + 0.004258336033672094, + 0.008416675962507725, + 0.0038396886084228754, + 0.004869547206908464, + 0.003583099925890565, + 0.0030212493147701025, + 0.0027119077276438475, + 0.003857432398945093, + 0.0035653209779411554, + 0.004931577481329441, + 0.0023838013876229525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.23109430074691772 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764929736033082 + ], + "opposite_logprob_tokens": [ + -4.8826494216918945 + ], + "logprob_reward": -0.00764929736033082, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764929736033082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06548906862735748, + 0.22850799560546875, + 0.08042339980602264, + 0.09637647867202759, + 0.2855408787727356, + 0.027840344235301018, + 0.05466802045702934, + 0.04302375391125679, + 0.058788083493709564, + 0.05934199318289757 + ], + "credit_map": [ + 0.006705837789922953, + 0.006308756768703461, + 0.008637985214591026, + 0.00785256177186966, + 0.00870230421423912, + 0.006237825844436884, + 0.010855228640139103, + 0.010188571177423, + 0.016571877524256706, + 0.011388301849365234, + 0.010378473438322544, + 0.013470590114593506, + 0.08389360457658768, + 0.020211609080433846, + 0.010672717355191708, + 0.06192081421613693, + 0.010321363806724548, + 0.00848601944744587, + 0.00907528679817915, + 0.006736426148563623, + 0.011705572716891766, + 0.012086418457329273, + 0.008304244838654995, + 0.013708068057894707, + 0.013881174847483635, + 0.010941382497549057, + 0.010576961562037468, + 0.010873974300920963, + 0.011964124627411366, + 0.014914543367922306, + 0.00797260832041502, + 0.015251709148287773, + 0.08778295665979385, + 0.03244670480489731, + 0.008749698288738728, + 0.008715437725186348, + 0.015159633941948414, + 0.10521478205919266, + 0.008708897046744823, + 0.01876276731491089, + 0.0038648019544780254, + 0.003134669503197074, + 0.0036896204110234976, + 0.003289369400590658, + 0.0024899907875806093, + 0.005572295747697353, + 0.003302633762359619, + 0.002496961038559675, + 0.005952074658125639, + 0.00581666873767972, + 0.008978779427707195, + 0.007950881496071815, + 0.00832399446517229, + 0.005638433620333672, + 0.008032282814383507, + 0.00397490756586194, + 0.004875106271356344, + 0.005712402053177357, + 0.005025890190154314, + 0.0040704477578401566, + 0.006013205274939537, + 0.0077996333129704, + 0.005451099947094917, + 0.004075970966368914, + 0.006895710714161396, + 0.004716996103525162, + 0.008323384448885918, + 0.0050444104708731174, + 0.006603073328733444, + 0.00802686344832182, + 0.01278829574584961, + 0.0063893478363752365, + 0.007641248404979706, + 0.00911760050803423, + 0.00747278006747365, + 0.006916022393852472, + 0.006225084885954857, + 0.005895618349313736, + 0.01238581258803606, + 0.0036878313403576612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.00764929736033082 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294130340218544 + ], + "opposite_logprob_tokens": [ + -3.77329421043396 + ], + "logprob_reward": -0.023294130340218544, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294130340218544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04020262137055397, + 0.11570367217063904, + 0.05246291309595108, + 0.016142431646585464, + 0.5059742331504822, + 0.014872973784804344, + 0.015835776925086975, + 0.030416259542107582, + 0.04586806520819664, + 0.16252097487449646 + ], + "credit_map": [ + 0.004043209832161665, + 0.004017912317067385, + 0.005752760451287031, + 0.006206143647432327, + 0.006639921106398106, + 0.002969651250168681, + 0.006088219117373228, + 0.004484805278480053, + 0.007517930120229721, + 0.009657973423600197, + 0.005847326945513487, + 0.005633206106722355, + 0.035521674901247025, + 0.013633783906698227, + 0.00744902016595006, + 0.03044275753200054, + 0.006993106100708246, + 0.006257359869778156, + 0.004959050100296736, + 0.005361303221434355, + 0.007420044858008623, + 0.00888142455369234, + 0.005682346876710653, + 0.00690827751532197, + 0.0029606514144688845, + 0.001876776572316885, + 0.001979883061721921, + 0.0018953608814626932, + 0.001598397851921618, + 0.0020036911591887474, + 0.0017017640639096498, + 0.0021259067580103874, + 0.15923818945884705, + 0.1853487342596054, + 0.022043023258447647, + 0.014596186578273773, + 0.03439994156360626, + 0.027803588658571243, + 0.024234294891357422, + 0.03831026330590248, + 0.0015539511805400252, + 0.0028018143493682146, + 0.001586235943250358, + 0.002064982196316123, + 0.001362438197247684, + 0.002547204727306962, + 0.001462520333006978, + 0.0014938266249373555, + 0.0014968840405344963, + 0.0019515774911269546, + 0.0024152856785804033, + 0.002270620549097657, + 0.0016207321314141154, + 0.0024599533062428236, + 0.0022920798510313034, + 0.0013286443427205086, + 0.0030123142059892416, + 0.0031639360822737217, + 0.0029971294570714235, + 0.0027586447540670633, + 0.0035734071861952543, + 0.007789307739585638, + 0.0035861192736774683, + 0.0035354015417397022, + 0.00484964856877923, + 0.003975291736423969, + 0.006938676815479994, + 0.004366028588265181, + 0.004945514723658562, + 0.00610872358083725, + 0.00965978391468525, + 0.005024397280067205, + 0.02406042255461216, + 0.033402372151613235, + 0.014898001216351986, + 0.014279368333518505, + 0.019590364769101143, + 0.012248357757925987, + 0.03609568253159523, + 0.007946415804326534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.023294130340218544 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023285161703824997 + ], + "opposite_logprob_tokens": [ + -3.773285150527954 + ], + "logprob_reward": -0.023285161703824997, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023285161703824997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025754965841770172, + 0.2721206843852997, + 0.04377829656004906, + 0.03784067556262016, + 0.20170772075653076, + 0.02398565784096718, + 0.01405660156160593, + 0.06986266374588013, + 0.014621688984334469, + 0.29627102613449097 + ], + "credit_map": [ + 0.0023822851944714785, + 0.0025067832320928574, + 0.003234997158870101, + 0.0028282178100198507, + 0.0045081861317157745, + 0.0026018356438726187, + 0.0033797011710703373, + 0.0043129571713507175, + 0.01732725277543068, + 0.010835133492946625, + 0.0069060130044817924, + 0.010070969350636005, + 0.1071169525384903, + 0.027197901159524918, + 0.009902767837047577, + 0.08276370167732239, + 0.005175452679395676, + 0.004700675141066313, + 0.004616381600499153, + 0.003802988911047578, + 0.008327827788889408, + 0.005305319558829069, + 0.004782817326486111, + 0.0070668356493115425, + 0.005115834064781666, + 0.004452416207641363, + 0.004844529088586569, + 0.0044406321831047535, + 0.0053371344693005085, + 0.006279363296926022, + 0.0034853643737733364, + 0.0038854009471833706, + 0.0784756988286972, + 0.027718560770154, + 0.00904136523604393, + 0.004482083022594452, + 0.00704360892996192, + 0.05394989997148514, + 0.007880134508013725, + 0.013116361573338509, + 0.0031674609053879976, + 0.0022821889724582434, + 0.004435289651155472, + 0.002257727552205324, + 0.0020425277762115, + 0.005745977163314819, + 0.002274364000186324, + 0.0017801226349547505, + 0.0016653210623189807, + 0.0014359074411913753, + 0.0021068088244646788, + 0.0017613564850762486, + 0.002245438750833273, + 0.001426764647476375, + 0.0024892748333513737, + 0.000925728993024677, + 0.008347507566213608, + 0.009263795800507069, + 0.00759846018627286, + 0.005918093025684357, + 0.009163826704025269, + 0.013992982916533947, + 0.008271477185189724, + 0.007306520827114582, + 0.0015992466360330582, + 0.0015206628013402224, + 0.0015691950684413314, + 0.0018992185359820724, + 0.0017994287190958858, + 0.0017558825202286243, + 0.0025257249362766743, + 0.0019523297669366002, + 0.043675463646650314, + 0.0594605952501297, + 0.02825748361647129, + 0.029147936031222343, + 0.03660310432314873, + 0.021565362811088562, + 0.06294415891170502, + 0.014616929925978184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.023285161703824997 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581839606165886 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581839606165886, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581839606165886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031978536397218704, + 0.05165906995534897, + 0.09474406391382217, + 0.11903790384531021, + 0.5088284015655518, + 0.048745814710855484, + 0.06905769556760788, + 0.023496393114328384, + 0.030928252264857292, + 0.021523812785744667 + ], + "credit_map": [ + 0.0035316224675625563, + 0.003713279962539673, + 0.0032933589536696672, + 0.004403115715831518, + 0.004592323675751686, + 0.0039362008683383465, + 0.0038753317203372717, + 0.004633303731679916, + 0.004925502464175224, + 0.006769117899239063, + 0.004737227689474821, + 0.006884507369250059, + 0.009301187470555305, + 0.008643201552331448, + 0.005757012404501438, + 0.004641316365450621, + 0.012787312269210815, + 0.010739412158727646, + 0.011339522898197174, + 0.008999214507639408, + 0.011719491332769394, + 0.013230501674115658, + 0.010943775996565819, + 0.01498483121395111, + 0.020620791241526604, + 0.013579566963016987, + 0.013166263699531555, + 0.012150172144174576, + 0.012112690135836601, + 0.024455154314637184, + 0.009312414564192295, + 0.013640852645039558, + 0.14180180430412292, + 0.19726134836673737, + 0.025129197165369987, + 0.014899985864758492, + 0.03626420348882675, + 0.03052733652293682, + 0.02519763819873333, + 0.03774692490696907, + 0.008002822287380695, + 0.003917996771633625, + 0.005761930253356695, + 0.004997637122869492, + 0.0048447721637785435, + 0.012009554542601109, + 0.005002864636480808, + 0.00420823460444808, + 0.007958837784826756, + 0.005963949952274561, + 0.011532648466527462, + 0.00997643917798996, + 0.01162167638540268, + 0.007091030478477478, + 0.009848272427916527, + 0.005064839031547308, + 0.0024111042730510235, + 0.0038284475449472666, + 0.0031715771183371544, + 0.0026650524232536554, + 0.003188893897458911, + 0.003417123341932893, + 0.00290315761230886, + 0.0019110360881313682, + 0.0032234229147434235, + 0.003549365559592843, + 0.0046241916716098785, + 0.00377848232164979, + 0.0036432803608477116, + 0.0030133971013128757, + 0.006048970855772495, + 0.0030471410136669874, + 0.0023896978236734867, + 0.0018922238377854228, + 0.0031649316661059856, + 0.0025698277167975903, + 0.00226279953494668, + 0.0038037870544940233, + 0.003393135964870453, + 0.002047409536316991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.020581839606165886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a realistic-style scene in a bright, sterile airport check-in area. A woman with thick, shoulder-length blond curls, wearing a crisp white collared shirt that fits her uniform, sits attentively behind a long, matte-gray counter. Her fingers initially move deftly across a black keyboard, her focus fixed on a computer screen just off-screen. She pauses typing, her gaze shifting slightly upward, toward an unseen figure. A hand reaches into the frame, offering her a dark blue passport, which she accepts with steady, practiced movements. Simultaneously, two other people approach from the right side of the frame, walking steadily, just like the next person. The first is a man with neatly combed short brown hair, thin-framed glasses, and a dark gray suit jacket over a light-colored shirt. The second person's face remains invisible, partially visible beside him, their form suggesting a casual posture. The entire scene is illuminated by cool, even overhead fluorescent lights, which reflect off the polished white tile floor, creating a clean, functional atmosphere free of harsh shadows and reinforcing the impersonal efficiency of the modern travel center.", + "seed": 869, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_4/index_1124_seed_869_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16705485237762333, + "vlm_logprob": -0.16705485237762333, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.16705485237762333, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7005784511566162 + ], + "opposite_logprob_tokens": [ + -0.7005784511566162 + ], + "logprob_reward": -0.7005784511566162, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7005784511566162, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1632114052772522, + 0.126831516623497, + 0.04149100184440613, + 0.07062594592571259, + 0.09079360216856003, + 0.07060082256793976, + 0.046091947704553604, + 0.03634649142622948, + 0.2805292308330536, + 0.07347807288169861 + ], + "credit_map": [ + 0.014096584171056747, + 0.03984496369957924, + 0.026307934895157814, + 0.015199375338852406, + 0.012933608144521713, + 0.02772180363535881, + 0.009160765446722507, + 0.01794636994600296, + 0.0072490135207772255, + 0.03145420923829079, + 0.00749247008934617, + 0.01795346848666668, + 0.008746680803596973, + 0.022244809195399284, + 0.011341426521539688, + 0.02034943550825119, + 0.002631702460348606, + 0.010640704073011875, + 0.00438801571726799, + 0.004197011236101389, + 0.0044839042238891125, + 0.0054740309715271, + 0.002823437564074993, + 0.006852193269878626, + 0.006585704628378153, + 0.009684859775006771, + 0.00811329297721386, + 0.012239180505275726, + 0.008401116356253624, + 0.008335987105965614, + 0.007849829271435738, + 0.009415975771844387, + 0.011904568411409855, + 0.011023462750017643, + 0.017738068476319313, + 0.0060885995626449585, + 0.015848247334361076, + 0.004206370562314987, + 0.005583356134593487, + 0.018400926142930984, + 0.008161553181707859, + 0.010304813273251057, + 0.0106894476339221, + 0.013713792897760868, + 0.0059502762742340565, + 0.007089917082339525, + 0.008135376498103142, + 0.006555645726621151, + 0.009215612895786762, + 0.005646067671477795, + 0.00710568530485034, + 0.005870720371603966, + 0.0036256015300750732, + 0.004585421644151211, + 0.006213631946593523, + 0.00382920540869236, + 0.0028067021630704403, + 0.00359018100425601, + 0.0034865227062255144, + 0.0033586553763598204, + 0.005374646279960871, + 0.007065131328999996, + 0.005997064523398876, + 0.004667589906603098, + 0.017250534147024155, + 0.0445382185280323, + 0.021607529371976852, + 0.0886344164609909, + 0.015051566995680332, + 0.06266811490058899, + 0.015278458595275879, + 0.01550037506967783, + 0.00781948771327734, + 0.008410805836319923, + 0.006544576957821846, + 0.0070177651941776276, + 0.011118674650788307, + 0.017233703285455704, + 0.007400250528007746, + 0.007932803593575954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.7005784511566162 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005975240375846624 + ], + "opposite_logprob_tokens": [ + -5.130975246429443 + ], + "logprob_reward": -0.005975240375846624, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005975240375846624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.247233048081398, + 0.11763077229261398, + 0.057355232536792755, + 0.0430314727127552, + 0.044471047818660736, + 0.09733223915100098, + 0.059299297630786896, + 0.0888240858912468, + 0.16608284413814545, + 0.07874005287885666 + ], + "credit_map": [ + 0.01677034981548786, + 0.07227320224046707, + 0.019754836335778236, + 0.027500014752149582, + 0.015831435099244118, + 0.051517095416784286, + 0.014059755951166153, + 0.029526321217417717, + 0.006431668531149626, + 0.03146575391292572, + 0.006153217516839504, + 0.012682936154305935, + 0.00830885674804449, + 0.024659626185894012, + 0.010379533283412457, + 0.017549166455864906, + 0.0034682455006986856, + 0.01677469164133072, + 0.006773713976144791, + 0.0048646763898432255, + 0.004239002242684364, + 0.00759917264804244, + 0.0035291570238769054, + 0.01010657288134098, + 0.0026495179627090693, + 0.0057198666036129, + 0.0035883209202438593, + 0.008615793660283089, + 0.004171275999397039, + 0.004958107136189938, + 0.005640363320708275, + 0.007688223384320736, + 0.004589027259498835, + 0.003343308810144663, + 0.005305926315486431, + 0.003238528734073043, + 0.010227032005786896, + 0.004704068414866924, + 0.003618854796513915, + 0.009444300085306168, + 0.012098442763090134, + 0.011732246726751328, + 0.011230256408452988, + 0.016938388347625732, + 0.011133961379528046, + 0.01004774309694767, + 0.014869649894535542, + 0.009281545877456665, + 0.010817659087479115, + 0.007759561762213707, + 0.00906670093536377, + 0.005475918762385845, + 0.006269498728215694, + 0.0063260444439947605, + 0.008054576814174652, + 0.005529331974685192, + 0.009810522198677063, + 0.009868581779301167, + 0.013800688087940216, + 0.013720785267651081, + 0.013900288380682468, + 0.009863622486591339, + 0.008955634199082851, + 0.008903955109417439, + 0.013763523660600185, + 0.040367305278778076, + 0.017273085191845894, + 0.04863119497895241, + 0.010808741673827171, + 0.015401682816445827, + 0.008494485169649124, + 0.011342816986143589, + 0.007968548685312271, + 0.008430217392742634, + 0.007609015796333551, + 0.00856007169932127, + 0.010566752403974533, + 0.0181638952344656, + 0.008124600164592266, + 0.009316947311162949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.005975240375846624 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059200033545494 + ], + "opposite_logprob_tokens": [ + -3.895591974258423 + ], + "logprob_reward": -0.02059200033545494, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059200033545494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06015085056424141, + 0.041229549795389175, + 0.05790276825428009, + 0.15315790474414825, + 0.07852400094270706, + 0.056457143276929855, + 0.040277086198329926, + 0.09491878002882004, + 0.27676013112068176, + 0.1406218558549881 + ], + "credit_map": [ + 0.005207838490605354, + 0.007884399965405464, + 0.010023427195847034, + 0.008912650868296623, + 0.0061591751873493195, + 0.006119161378592253, + 0.007615262642502785, + 0.008228931576013565, + 0.002467077225446701, + 0.005343568045645952, + 0.004743803758174181, + 0.0065392907708883286, + 0.004208676051348448, + 0.006422190926969051, + 0.004714273847639561, + 0.0067906673066318035, + 0.0036804445553570986, + 0.015253981575369835, + 0.007628968916833401, + 0.006687963847070932, + 0.004612477496266365, + 0.005854866001754999, + 0.004599724896252155, + 0.009584338404238224, + 0.010287395678460598, + 0.020746955648064613, + 0.012158260680735111, + 0.029339492321014404, + 0.016084467992186546, + 0.019340623170137405, + 0.019786091521382332, + 0.025414615869522095, + 0.008242923766374588, + 0.008005619049072266, + 0.014594507403671741, + 0.008716828189790249, + 0.012124338187277317, + 0.005828256253153086, + 0.005858622491359711, + 0.015152906067669392, + 0.006792289670556784, + 0.004556133411824703, + 0.005864119157195091, + 0.009923827834427357, + 0.00865001417696476, + 0.007152506150305271, + 0.008128215558826923, + 0.005390035454183817, + 0.006947669666260481, + 0.004997150041162968, + 0.004487925674766302, + 0.004151569213718176, + 0.00553939351812005, + 0.0048639243468642235, + 0.005487927235662937, + 0.0038015260361135006, + 0.009144926443696022, + 0.007454807870090008, + 0.010900135152041912, + 0.011510108597576618, + 0.013115101493895054, + 0.019970929250121117, + 0.010013037361204624, + 0.01280973106622696, + 0.0129851708188653, + 0.02560502104461193, + 0.012930195778608322, + 0.05632925033569336, + 0.024712231010198593, + 0.11493763327598572, + 0.016686230897903442, + 0.012574393302202225, + 0.027463313192129135, + 0.016259362921118736, + 0.015327544882893562, + 0.028699427843093872, + 0.012386718764901161, + 0.01795732043683529, + 0.010426775552332401, + 0.01210139598697424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.02059200033545494 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038097530603408813 + ], + "opposite_logprob_tokens": [ + -3.288097620010376 + ], + "logprob_reward": -0.038097530603408813, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038097530603408813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19371530413627625, + 0.10441997647285461, + 0.042488545179367065, + 0.0796770453453064, + 0.08060096949338913, + 0.07086072862148285, + 0.050510235130786896, + 0.1426207721233368, + 0.12147345393896103, + 0.1136329174041748 + ], + "credit_map": [ + 0.012575890868902206, + 0.05324537307024002, + 0.011970947496592999, + 0.023510273545980453, + 0.012849369086325169, + 0.04332801327109337, + 0.012676788493990898, + 0.02355865389108658, + 0.00548906484618783, + 0.03386097773909569, + 0.0045327371917665005, + 0.007186554837971926, + 0.006909728981554508, + 0.027727091684937477, + 0.004471722058951855, + 0.014242105185985565, + 0.002585030160844326, + 0.01269758865237236, + 0.0027799904346466064, + 0.004526030737906694, + 0.003858296200633049, + 0.007768367882817984, + 0.002458303002640605, + 0.005814938805997372, + 0.006614606361836195, + 0.007463178131729364, + 0.005540221929550171, + 0.0152266975492239, + 0.008202645927667618, + 0.01393911987543106, + 0.008189319632947445, + 0.014501255005598068, + 0.008267520926892757, + 0.004976099357008934, + 0.008893131278455257, + 0.0054159690625965595, + 0.020136505365371704, + 0.00997395720332861, + 0.006785841193050146, + 0.016151949763298035, + 0.009402135387063026, + 0.006940935272723436, + 0.006355405319482088, + 0.009536497294902802, + 0.010107028298079967, + 0.009715231135487556, + 0.01161191612482071, + 0.007191576063632965, + 0.006713866721838713, + 0.006576579064130783, + 0.006840622052550316, + 0.004097703378647566, + 0.007682806812226772, + 0.006631504278630018, + 0.007704298011958599, + 0.004262859467417002, + 0.01082482561469078, + 0.013868090696632862, + 0.013415939174592495, + 0.015805192291736603, + 0.020443538203835487, + 0.02880251780152321, + 0.018271947279572487, + 0.02118871547281742, + 0.007843987084925175, + 0.014666922390460968, + 0.010306648910045624, + 0.024731049314141273, + 0.010873502120375633, + 0.03585591912269592, + 0.009494111873209476, + 0.007701306603848934, + 0.01169439498335123, + 0.012697326019406319, + 0.012614316307008266, + 0.01348521001636982, + 0.01631578616797924, + 0.021860964596271515, + 0.01326728891581297, + 0.011697635985910892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.038097530603408813 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, sterile airport check-in area with polished white tile floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman with thick, shoulder-length blond curls wearing a crisp white collared shirt sitting behind a long, matte-gray counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hand reach into the frame to offer the woman a dark blue passport which she accepts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two other people approaching from the right side of the frame, including a man with short brown hair, thin-framed glasses, and a dark gray suit jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool, even overhead fluorescent lights creating a clean atmosphere free of harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003103941679001 + ], + "opposite_logprob_tokens": [ + -2.695030927658081 + ], + "logprob_reward": -0.07003103941679001, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003103941679001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17303557693958282, + 0.08461163938045502, + 0.033884447067976, + 0.05177474766969681, + 0.12990295886993408, + 0.05042401701211929, + 0.14859639108181, + 0.029300982132554054, + 0.24964317679405212, + 0.048826102167367935 + ], + "credit_map": [ + 0.011988787911832333, + 0.049581848084926605, + 0.016569001600146294, + 0.019481072202324867, + 0.011532768607139587, + 0.035259924829006195, + 0.009373645298182964, + 0.019248533993959427, + 0.004492749460041523, + 0.020885756239295006, + 0.005159015767276287, + 0.011450942605733871, + 0.006206754129379988, + 0.016001783311367035, + 0.007357204798609018, + 0.013057430274784565, + 0.0021541311871260405, + 0.00898000504821539, + 0.0029996088705956936, + 0.004299109801650047, + 0.0028212550096213818, + 0.004442704375833273, + 0.002413378097116947, + 0.005774255841970444, + 0.004861702211201191, + 0.006992518901824951, + 0.004984055180102587, + 0.007059444673359394, + 0.005557605531066656, + 0.007922586984932423, + 0.004906707443296909, + 0.009490126743912697, + 0.01495322771370411, + 0.011015287600457668, + 0.019849224016070366, + 0.00765635259449482, + 0.028647905215620995, + 0.017157498747110367, + 0.009083548560738564, + 0.02153990976512432, + 0.00590753648430109, + 0.0063512492924928665, + 0.006616331171244383, + 0.008153339847922325, + 0.00511371111497283, + 0.00594208063557744, + 0.0060471645556390285, + 0.006292601116001606, + 0.015947671607136726, + 0.021843668073415756, + 0.016414228826761246, + 0.010030847042798996, + 0.02904267981648445, + 0.018231328576803207, + 0.02230933867394924, + 0.014776637777686119, + 0.002759027061983943, + 0.0032056088093668222, + 0.003428828902542591, + 0.003288581734523177, + 0.0035176677629351616, + 0.005210071336477995, + 0.004725448787212372, + 0.0031657477375119925, + 0.01734306290745735, + 0.04813125729560852, + 0.01997959241271019, + 0.08118456602096558, + 0.012210066430270672, + 0.046241287142038345, + 0.013289936818182468, + 0.011263407766819, + 0.005825840402394533, + 0.006032539531588554, + 0.004040871746838093, + 0.0052754939533770084, + 0.006069575436413288, + 0.014295076951384544, + 0.003333094296976924, + 0.003953610546886921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.07003103941679001 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_5/index_tensor([5530, 5530])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_5/index_tensor([5530, 5530])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..28d0ae796a6db5a4bde61d046e56b3b209325eac --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_5/index_tensor([5530, 5530])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, full-body shot, taken from a level angle, captures a tall, weathered oak bookshelf, crammed with a variety of ancient books. The wood is a deep walnut, its surface covered with minute scratches and the patina of time, its edges carved with delicate, though worn, classical patterns. Hardcover books, their leather and cloth covers in faded shades of crimson, forest green, and navy blue, are tightly packed on the shelves. The faint, gilded titles on the spines appear faintly in the dim light. The scene is enveloped in the atmosphere of a quiet study, with a soft light slanting in from outside the frame, illuminating the dust particles dancing in the air. This transformation begins quietly at the upper right corner of the bookshelf, where the solid wooden edge begins to silently disintegrate, not shattering, but gracefully disintegrating into countless tiny, luminous digital particles. Like a swarm of disturbed fireflies, these particles emit a bright, cold glow of cyan and white, contrasting sharply with the warm tones of the study. They slowly drifted upwards as if drawn by an invisible current, the disintegration effect like a silent wave, slowly and inexorably spreading downwards along the contours of the bookshelf. Then, the ancient books began to melt from the edges of the pages, and the paper and text turned into equally glittering data streams, merging into the ever-expanding galaxy of light and code, creating a surreal atmosphere that was both sad and full of wonder, symbolizing the tranquil evolution of physical knowledge into digital form.", + "seed": 11626, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_5/index_5530_seed_11626_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19087547454982995, + "vlm_logprob": -0.19087547454982995, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.19087547454982995, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7667950391769409 + ], + "opposite_logprob_tokens": [ + -0.6417950391769409 + ], + "logprob_reward": -0.7667950391769409, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7667950391769409, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15164761245250702, + 0.03372645750641823, + 0.06465968489646912, + 0.046108122915029526, + 0.18537858128547668, + 0.03613043949007988, + 0.24895991384983063, + 0.08561118692159653, + 0.055470407009124756, + 0.09230764955282211 + ], + "credit_map": [ + 0.045933496206998825, + 0.0443149134516716, + 0.015827763825654984, + 0.008190231397747993, + 0.008906646631658077, + 0.009347313083708286, + 0.0074438597075641155, + 0.011683376505970955, + 0.006914893630892038, + 0.0044105579145252705, + 0.004067814443260431, + 0.004825666081160307, + 0.002811640268191695, + 0.0025220350362360477, + 0.004364138003438711, + 0.0038097139913588762, + 0.009137731045484543, + 0.009149762801826, + 0.01696368120610714, + 0.006606882903724909, + 0.004453490022569895, + 0.0041017355397343636, + 0.007248201873153448, + 0.0069981953129172325, + 0.008829924277961254, + 0.008223122917115688, + 0.0070063332095742226, + 0.00545066362246871, + 0.004084587097167969, + 0.003469976596534252, + 0.004369132220745087, + 0.004674382042139769, + 0.016000665724277496, + 0.015561792999505997, + 0.026881426572799683, + 0.06952937692403793, + 0.008097471669316292, + 0.008438139222562313, + 0.015152761712670326, + 0.025716932490468025, + 0.004409192595630884, + 0.005892626475542784, + 0.005538978613913059, + 0.003150783944875002, + 0.005981230642646551, + 0.004193070810288191, + 0.003618556307628751, + 0.003345995908603072, + 0.04260772466659546, + 0.025464216247200966, + 0.07420993596315384, + 0.026616141200065613, + 0.019232921302318573, + 0.012150013819336891, + 0.02806699648499489, + 0.020611923187971115, + 0.008908350951969624, + 0.013104765675961971, + 0.017674745991826057, + 0.008504865691065788, + 0.012983782216906548, + 0.007445093244314194, + 0.008597023785114288, + 0.00839255191385746, + 0.0050175548531115055, + 0.00799543410539627, + 0.007922224700450897, + 0.006848504301160574, + 0.006717358250170946, + 0.007296364288777113, + 0.0074022309854626656, + 0.0062707276083528996, + 0.00720487255603075, + 0.01754472404718399, + 0.00927355233579874, + 0.011959097348153591, + 0.015850530937314034, + 0.011059709824621677, + 0.011159558780491352, + 0.00825559813529253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.7667950391769409 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609652116894722 + ], + "opposite_logprob_tokens": [ + -4.141096591949463 + ], + "logprob_reward": -0.01609652116894722, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609652116894722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1934506595134735, + 0.06169316545128822, + 0.09295579791069031, + 0.10814333707094193, + 0.11978168040513992, + 0.05093805491924286, + 0.14929386973381042, + 0.06354039162397385, + 0.07063576579093933, + 0.08956723660230637 + ], + "credit_map": [ + 0.03961404040455818, + 0.030491653829813004, + 0.04978957399725914, + 0.019023187458515167, + 0.012142328545451164, + 0.013019075617194176, + 0.011845254339277744, + 0.017525557428598404, + 0.010050734505057335, + 0.014361326582729816, + 0.010417193174362183, + 0.005762117449194193, + 0.004355344455689192, + 0.005522529128938913, + 0.004936387296766043, + 0.006287535186856985, + 0.014783094637095928, + 0.013057955540716648, + 0.0318293422460556, + 0.008654067292809486, + 0.00684211403131485, + 0.004493944812566042, + 0.008618630468845367, + 0.0046766530722379684, + 0.01550054457038641, + 0.02405383810400963, + 0.022530391812324524, + 0.008353706449270248, + 0.007088218349963427, + 0.013907115906476974, + 0.007298493757843971, + 0.009411027655005455, + 0.012647813186049461, + 0.011461764574050903, + 0.021767696365714073, + 0.039609894156455994, + 0.00662229023873806, + 0.006733478046953678, + 0.009082200936973095, + 0.011856542900204659, + 0.0039035568479448557, + 0.006858342792838812, + 0.007077543530613184, + 0.00397157110273838, + 0.008089027367532253, + 0.008044994436204433, + 0.004430615343153477, + 0.008562403731048107, + 0.017773469910025597, + 0.018069332465529442, + 0.031775034964084625, + 0.021593336015939713, + 0.015531652607023716, + 0.009504054673016071, + 0.020377565175294876, + 0.014669432304799557, + 0.005859429482370615, + 0.009728005155920982, + 0.008619364351034164, + 0.007356227841228247, + 0.010193698108196259, + 0.00605744356289506, + 0.009005612693727016, + 0.006720606703311205, + 0.006898245774209499, + 0.010287602432072163, + 0.010424260050058365, + 0.007893583737313747, + 0.009545030072331429, + 0.00875867810100317, + 0.009254403412342072, + 0.007573958486318588, + 0.007292691618204117, + 0.019258547574281693, + 0.007934528402984142, + 0.012926897034049034, + 0.014417487196624279, + 0.01028907485306263, + 0.01045860256999731, + 0.006989411544054747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.01609652116894722 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292169988155365 + ], + "opposite_logprob_tokens": [ + -2.237921714782715 + ], + "logprob_reward": -0.11292169988155365, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292169988155365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07207732647657394, + 0.06668102741241455, + 0.1718105524778366, + 0.15581868588924408, + 0.17703719437122345, + 0.12076866626739502, + 0.0791250467300415, + 0.06513740867376328, + 0.04874731972813606, + 0.04279673099517822 + ], + "credit_map": [ + 0.012558422982692719, + 0.010087227448821068, + 0.009743222035467625, + 0.0050928969867527485, + 0.00425758445635438, + 0.01058682706207037, + 0.0047546979039907455, + 0.014996449463069439, + 0.01056401152163744, + 0.00801229290664196, + 0.00963536649942398, + 0.00634313328191638, + 0.00501750223338604, + 0.008058896288275719, + 0.00885122362524271, + 0.010198595933616161, + 0.02290552854537964, + 0.02392735332250595, + 0.036498844623565674, + 0.018196076154708862, + 0.01034428458660841, + 0.02131178416311741, + 0.014344227500259876, + 0.024282462894916534, + 0.020697642117738724, + 0.030085448175668716, + 0.03633883222937584, + 0.011799152009189129, + 0.012111442163586617, + 0.016070719808340073, + 0.013554357923567295, + 0.015161099843680859, + 0.01813269965350628, + 0.0175895057618618, + 0.03794064000248909, + 0.05453315004706383, + 0.011870428919792175, + 0.009204483591020107, + 0.014906325377523899, + 0.012859965674579144, + 0.016418039798736572, + 0.01710447482764721, + 0.01892300695180893, + 0.012612489983439445, + 0.018860839307308197, + 0.010528892278671265, + 0.01636907458305359, + 0.00995185598731041, + 0.014933288097381592, + 0.009709102101624012, + 0.0258104857057333, + 0.008568880148231983, + 0.006170604843646288, + 0.004199141636490822, + 0.005649207159876823, + 0.004084342159330845, + 0.007065161131322384, + 0.007083852309733629, + 0.015271022915840149, + 0.006825377698987722, + 0.00977138802409172, + 0.005934744607657194, + 0.00714501179754734, + 0.006040850188583136, + 0.003807968460023403, + 0.006096476223319769, + 0.008957811631262302, + 0.006576684769243002, + 0.006232604384422302, + 0.005147876217961311, + 0.008268010802567005, + 0.003659887006506324, + 0.003744592424482107, + 0.006290704943239689, + 0.003289847867563367, + 0.007375010289251804, + 0.005979796871542931, + 0.004523266572505236, + 0.008080697618424892, + 0.0035128125455230474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.11292169988155365 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003616699017584324 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.003616699017584324, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616699017584324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06797041743993759, + 0.017177075147628784, + 0.04294794797897339, + 0.08306720852851868, + 0.15307065844535828, + 0.04164230450987816, + 0.20296289026737213, + 0.1582912653684616, + 0.1553080528974533, + 0.07756219804286957 + ], + "credit_map": [ + 0.022145066410303116, + 0.022579409182071686, + 0.006377354264259338, + 0.004561716224998236, + 0.0037904451601207256, + 0.002451046835631132, + 0.002786279423162341, + 0.003279102733358741, + 0.0031545369420200586, + 0.0023952010087668896, + 0.0019679234828799963, + 0.0021689203567802906, + 0.001495811389759183, + 0.0020990383345633745, + 0.0016291976207867265, + 0.0022664472926408052, + 0.006685152184218168, + 0.006268741562962532, + 0.0143510140478611, + 0.0031419715378433466, + 0.00238702236674726, + 0.0028079759795218706, + 0.0035674839746207, + 0.0037385832984000444, + 0.011119111441075802, + 0.014911974780261517, + 0.016264261677861214, + 0.007156362757086754, + 0.007752586156129837, + 0.007942265830934048, + 0.008505988866090775, + 0.009414651431143284, + 0.010752874426543713, + 0.01146411057561636, + 0.01589091494679451, + 0.05421742424368858, + 0.006547811906784773, + 0.01038364041596651, + 0.010400662198662758, + 0.03341321274638176, + 0.004208413418382406, + 0.007152768783271313, + 0.007922994904220104, + 0.0041118888184428215, + 0.007982843555510044, + 0.0032552769407629967, + 0.0038104753475636244, + 0.0031976441387087107, + 0.0392548069357872, + 0.02183537557721138, + 0.06430753320455551, + 0.015738263726234436, + 0.011836555786430836, + 0.01116667129099369, + 0.023469585925340652, + 0.015354110859334469, + 0.015041151084005833, + 0.02383657544851303, + 0.03150385245680809, + 0.017125548794865608, + 0.02538110688328743, + 0.0162202101200819, + 0.016389118507504463, + 0.01279370579868555, + 0.011815804988145828, + 0.02307957038283348, + 0.020801540464162827, + 0.026395009830594063, + 0.0178892333060503, + 0.014917874708771706, + 0.024699751287698746, + 0.01570925861597061, + 0.006055123172700405, + 0.019488411024212837, + 0.00958307646214962, + 0.012064432725310326, + 0.0106128491461277, + 0.006067890208214521, + 0.008890124969184399, + 0.004800287075340748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.003616699017584324 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494741350412369 + ], + "opposite_logprob_tokens": [ + -2.9299473762512207 + ], + "logprob_reward": -0.05494741350412369, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494741350412369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07412645220756531, + 0.06538516283035278, + 0.1012003943324089, + 0.13328245282173157, + 0.11489082127809525, + 0.09523209929466248, + 0.1553802639245987, + 0.06446729600429535, + 0.12122910469770432, + 0.07480599731206894 + ], + "credit_map": [ + 0.013166198506951332, + 0.017551785334944725, + 0.011742055416107178, + 0.007118820678442717, + 0.005679355468600988, + 0.0056893490254879, + 0.006397099234163761, + 0.006781793665140867, + 0.0171809121966362, + 0.018202709034085274, + 0.006115807220339775, + 0.0029514001216739416, + 0.0031131671275943518, + 0.005091015715152025, + 0.00407386152073741, + 0.008656292222440243, + 0.01111622340977192, + 0.009069574065506458, + 0.019948696717619896, + 0.008471248671412468, + 0.008379312232136726, + 0.01355669554322958, + 0.008489511907100677, + 0.022169120609760284, + 0.018385829403996468, + 0.028900444507598877, + 0.022275418043136597, + 0.012183679267764091, + 0.012639821507036686, + 0.009734511375427246, + 0.012290863320231438, + 0.01687188819050789, + 0.014466114342212677, + 0.01597466692328453, + 0.023617060855031013, + 0.015525215305387974, + 0.010802692733705044, + 0.008426845073699951, + 0.010350377298891544, + 0.015727855265140533, + 0.013784117996692657, + 0.013606944121420383, + 0.01431523822247982, + 0.008100784383714199, + 0.01553824171423912, + 0.00937895942479372, + 0.011111645959317684, + 0.009396168403327465, + 0.03121250681579113, + 0.020984560251235962, + 0.05142245441675186, + 0.01380728930234909, + 0.013194456696510315, + 0.0065927524119615555, + 0.009689350612461567, + 0.00847689714282751, + 0.0051767295226454735, + 0.008885902352631092, + 0.008843683637678623, + 0.008442494086921215, + 0.006554299965500832, + 0.0071417223662137985, + 0.0100942924618721, + 0.009328169748187065, + 0.011693193577229977, + 0.02128281630575657, + 0.013856301084160805, + 0.01697595976293087, + 0.018191244453191757, + 0.012224799953401089, + 0.015291871502995491, + 0.011712927371263504, + 0.006885920185595751, + 0.009285114705562592, + 0.010487793013453484, + 0.009267384186387062, + 0.01292336918413639, + 0.010566284880042076, + 0.009791039861738682, + 0.005599095486104488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.05494741350412369 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, full-body shot, taken from a level angle, captures a tall, weathered oak bookshelf, crammed with a variety of ancient books. The wood is a deep walnut, its surface covered with minute scratches and the patina of time, its edges carved with delicate, though worn, classical patterns. Hardcover books, their leather and cloth covers in faded shades of crimson, forest green, and navy blue, are tightly packed on the shelves. The faint, gilded titles on the spines appear faintly in the dim light. The scene is enveloped in the atmosphere of a quiet study, with a soft light slanting in from outside the frame, illuminating the dust particles dancing in the air. This transformation begins quietly at the upper right corner of the bookshelf, where the solid wooden edge begins to silently disintegrate, not shattering, but gracefully disintegrating into countless tiny, luminous digital particles. Like a swarm of disturbed fireflies, these particles emit a bright, cold glow of cyan and white, contrasting sharply with the warm tones of the study. They slowly drifted upwards as if drawn by an invisible current, the disintegration effect like a silent wave, slowly and inexorably spreading downwards along the contours of the bookshelf. Then, the ancient books began to melt from the edges of the pages, and the paper and text turned into equally glittering data streams, merging into the ever-expanding galaxy of light and code, creating a surreal atmosphere that was both sad and full of wonder, symbolizing the tranquil evolution of physical knowledge into digital form.", + "seed": 11627, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_5/index_5530_seed_11627_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.4155093220062554, + "vlm_logprob": -0.4155093220062554, + "vlm_yes_no_margin": 2.174999976158142, + "vlm_token_credit": -0.4155093220062554, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.8087314367294312 + ], + "opposite_logprob_tokens": [ + -0.18373145163059235 + ], + "logprob_reward": -1.8087314367294312, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.8087314367294312, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19792097806930542, + 0.05969807505607605, + 0.14506150782108307, + 0.08684513717889786, + 0.07605791091918945, + 0.0736653059720993, + 0.17359347641468048, + 0.07196405529975891, + 0.04898882657289505, + 0.06620471179485321 + ], + "credit_map": [ + 0.026898927986621857, + 0.018364129588007927, + 0.02056274376809597, + 0.022673722356557846, + 0.02337213233113289, + 0.032827217131853104, + 0.025916598737239838, + 0.027305513620376587, + 0.005674854386597872, + 0.004491076804697514, + 0.005973676685243845, + 0.012766038067638874, + 0.007524670101702213, + 0.009330499917268753, + 0.006395459175109863, + 0.00754180084913969, + 0.01240916084498167, + 0.00937449000775814, + 0.008934670127928257, + 0.015332553535699844, + 0.028654493391513824, + 0.03930702060461044, + 0.010971245355904102, + 0.02007787488400936, + 0.009210423566401005, + 0.0068945805542171, + 0.008294284343719482, + 0.0206300001591444, + 0.008062356151640415, + 0.008235047571361065, + 0.0075119840912520885, + 0.018006457015872, + 0.009918704628944397, + 0.0075741722248494625, + 0.006061590276658535, + 0.013040672056376934, + 0.007299500983208418, + 0.009638314135372639, + 0.0074226525612175465, + 0.015102299861609936, + 0.012394473887979984, + 0.006263035349547863, + 0.005733223166316748, + 0.009431373327970505, + 0.006309278775006533, + 0.008900261484086514, + 0.005401829723268747, + 0.019231827929615974, + 0.02378683350980282, + 0.02351488173007965, + 0.013636549934744835, + 0.012908539734780788, + 0.018336430191993713, + 0.013353938236832619, + 0.01364742498844862, + 0.05440886691212654, + 0.014223351143300533, + 0.006662427913397551, + 0.005378782749176025, + 0.01781233213841915, + 0.00933560635894537, + 0.008355648256838322, + 0.0054671489633619785, + 0.0047287591733038425, + 0.0053457398898899555, + 0.003566345665603876, + 0.00394450081512332, + 0.006756719201803207, + 0.0075143990106880665, + 0.00916965864598751, + 0.00545104406774044, + 0.0072404188103973866, + 0.00833793543279171, + 0.009260345250368118, + 0.006125705782324076, + 0.009521772153675556, + 0.007532324641942978, + 0.011014357209205627, + 0.00902126170694828, + 0.00539100356400013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -1.8087314367294312 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033780623227357864 + ], + "opposite_logprob_tokens": [ + -3.408780574798584 + ], + "logprob_reward": -0.033780623227357864, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033780623227357864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16532745957374573, + 0.0651206448674202, + 0.09655104577541351, + 0.0779787227511406, + 0.0507214218378067, + 0.0645296722650528, + 0.14492234587669373, + 0.1034553274512291, + 0.09301283955574036, + 0.13838061690330505 + ], + "credit_map": [ + 0.020838875323534012, + 0.01503230445086956, + 0.014546969905495644, + 0.01688893511891365, + 0.022188697010278702, + 0.03201710432767868, + 0.01791471801698208, + 0.025899844244122505, + 0.005979716777801514, + 0.004767146427184343, + 0.005363452713936567, + 0.013572345487773418, + 0.011245744302868843, + 0.011415035463869572, + 0.005911920685321093, + 0.006865270435810089, + 0.009103826247155666, + 0.006347690708935261, + 0.006691413465887308, + 0.00837279949337244, + 0.019870378077030182, + 0.02600044198334217, + 0.007981345057487488, + 0.012183137238025665, + 0.0068235090002417564, + 0.005528950598090887, + 0.007917680777609348, + 0.013105900026857853, + 0.010681687854230404, + 0.01792198419570923, + 0.006336991209536791, + 0.00966201163828373, + 0.00669608311727643, + 0.004896362312138081, + 0.0045231543481349945, + 0.010417364537715912, + 0.005470451433211565, + 0.00825232733041048, + 0.005402948707342148, + 0.005062723532319069, + 0.0074197109788656235, + 0.005145626608282328, + 0.007210774812847376, + 0.00866761151701212, + 0.010703543201088905, + 0.009937561117112637, + 0.0049422564916312695, + 0.010502578690648079, + 0.021865716204047203, + 0.017094623297452927, + 0.012651358731091022, + 0.014610353857278824, + 0.015510364435613155, + 0.011958853341639042, + 0.013094739057123661, + 0.0381363183259964, + 0.022835133597254753, + 0.006452774628996849, + 0.007175017148256302, + 0.03580988198518753, + 0.008592451922595501, + 0.008968988433480263, + 0.008276269771158695, + 0.005344795528799295, + 0.009577229619026184, + 0.009376502595841885, + 0.00975011382251978, + 0.011951468884944916, + 0.016507335007190704, + 0.017953943461179733, + 0.010347153060138226, + 0.007549078203737736, + 0.014983806759119034, + 0.021392203867435455, + 0.013033056631684303, + 0.01604786328971386, + 0.02104373276233673, + 0.025968845933675766, + 0.01599041186273098, + 0.009920685552060604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.033780623227357864 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706716358661652 + ], + "opposite_logprob_tokens": [ + -2.1270670890808105 + ], + "logprob_reward": -0.12706716358661652, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12706716358661652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10911089181900024, + 0.060280732810497284, + 0.08958254009485245, + 0.054796237498521805, + 0.08357597142457962, + 0.07007118314504623, + 0.34417882561683655, + 0.12409308552742004, + 0.03442312777042389, + 0.029887419193983078 + ], + "credit_map": [ + 0.010294359177350998, + 0.00779406912624836, + 0.008366215042769909, + 0.022975943982601166, + 0.012010445818305016, + 0.0190166924148798, + 0.008335299789905548, + 0.020317861810326576, + 0.005418397951871157, + 0.004494246561080217, + 0.005130006466060877, + 0.012481664307415485, + 0.007743580266833305, + 0.011583419516682625, + 0.00543577503412962, + 0.007993636652827263, + 0.006985266227275133, + 0.005500890780240297, + 0.005994547624140978, + 0.022406121715903282, + 0.013392305001616478, + 0.015730390325188637, + 0.006229061633348465, + 0.01334395818412304, + 0.005001537501811981, + 0.005019489210098982, + 0.00734991580247879, + 0.007494125049561262, + 0.0076747494749724865, + 0.01023086253553629, + 0.005228662397712469, + 0.006796898320317268, + 0.006177885457873344, + 0.004865054041147232, + 0.005814923904836178, + 0.020690886303782463, + 0.008501643314957619, + 0.017217673361301422, + 0.0067109838128089905, + 0.013596915639936924, + 0.007100810296833515, + 0.007055658847093582, + 0.006337098777294159, + 0.008750535547733307, + 0.009516232647001743, + 0.013108408078551292, + 0.005868666339665651, + 0.012333772145211697, + 0.07386936992406845, + 0.030672447755932808, + 0.020127732306718826, + 0.10919710248708725, + 0.020326849073171616, + 0.019564146175980568, + 0.02647535875439644, + 0.04394581913948059, + 0.028353914618492126, + 0.008165356703102589, + 0.007511560805141926, + 0.045047469437122345, + 0.008378035388886929, + 0.010785543359816074, + 0.009415929205715656, + 0.006435276009142399, + 0.003977708052843809, + 0.0037163833621889353, + 0.003409063909202814, + 0.0045701563358306885, + 0.0054891956970095634, + 0.004504543263465166, + 0.004363732412457466, + 0.0043923440389335155, + 0.003476690035313368, + 0.003447942901402712, + 0.00240586930885911, + 0.0048884060233831406, + 0.004529543220996857, + 0.005474284291267395, + 0.003666285192593932, + 0.0019983977545052767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.12706716358661652 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007634746376425028 + ], + "opposite_logprob_tokens": [ + -4.88263463973999 + ], + "logprob_reward": -0.007634746376425028, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007634746376425028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10860974341630936, + 0.03164009004831314, + 0.03206827491521835, + 0.06584953516721725, + 0.035011451691389084, + 0.07620292156934738, + 0.21561479568481445, + 0.2923445403575897, + 0.06276559084653854, + 0.07989305257797241 + ], + "credit_map": [ + 0.011386057361960411, + 0.008558080531656742, + 0.011450069025158882, + 0.012464839033782482, + 0.015868736431002617, + 0.025257030501961708, + 0.013194621540606022, + 0.01043031271547079, + 0.0026320689357817173, + 0.002142846817150712, + 0.00388321653008461, + 0.004538053646683693, + 0.005060635041445494, + 0.006399788893759251, + 0.003662463976070285, + 0.003321016440168023, + 0.003960957285016775, + 0.0022726471070200205, + 0.0029846057295799255, + 0.006758712697774172, + 0.0034112350549548864, + 0.004066830035299063, + 0.002677653916180134, + 0.005935635883361101, + 0.0042681144550442696, + 0.004078659228980541, + 0.005429931916296482, + 0.007195157464593649, + 0.014102469198405743, + 0.015664204955101013, + 0.005222484935075045, + 0.009888510219752789, + 0.003934126812964678, + 0.0029712500981986523, + 0.004210933111608028, + 0.004642302170395851, + 0.0039335754700005054, + 0.0062168315052986145, + 0.003121927147731185, + 0.005980508401989937, + 0.007925759069621563, + 0.007033226545900106, + 0.0073020802810788155, + 0.009314167313277721, + 0.00962394755333662, + 0.01161328237503767, + 0.005724248941987753, + 0.017666218802332878, + 0.05056600645184517, + 0.015379386022686958, + 0.014995875768363476, + 0.07348563522100449, + 0.015580019913613796, + 0.013912104070186615, + 0.018803857266902924, + 0.012891924940049648, + 0.06979309022426605, + 0.016386209055781364, + 0.01816842518746853, + 0.09553410112857819, + 0.025951027870178223, + 0.030972445383667946, + 0.022617211565375328, + 0.012922031804919243, + 0.005314582958817482, + 0.005890806671231985, + 0.009637312963604927, + 0.008276955224573612, + 0.01185539923608303, + 0.009612608700990677, + 0.005804833024740219, + 0.006373096257448196, + 0.006735456641763449, + 0.015051177702844143, + 0.010666728019714355, + 0.010023900307714939, + 0.011639819480478764, + 0.011382377706468105, + 0.008870938792824745, + 0.0055226488038897514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.007634746376425028 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033264011144638 + ], + "opposite_logprob_tokens": [ + -2.350332736968994 + ], + "logprob_reward": -0.10033264011144638, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033264011144638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06884697824716568, + 0.030655786395072937, + 0.04248006269335747, + 0.055126387625932693, + 0.02241867035627365, + 0.08729668706655502, + 0.10866649448871613, + 0.5042489171028137, + 0.04484736919403076, + 0.03541260212659836 + ], + "credit_map": [ + 0.005686378572136164, + 0.004767341073602438, + 0.00742038618773222, + 0.013693239539861679, + 0.007487128023058176, + 0.01373166311532259, + 0.007064566481858492, + 0.008996273390948772, + 0.00437677837908268, + 0.0028390700463205576, + 0.0043442086316645145, + 0.004607032053172588, + 0.003698852378875017, + 0.004528000485152006, + 0.00278662727214396, + 0.003475215518847108, + 0.003763351356610656, + 0.002750566927716136, + 0.004165375605225563, + 0.005185739137232304, + 0.007496783975511789, + 0.011070684529840946, + 0.003988375887274742, + 0.004059183411300182, + 0.0051370286382734776, + 0.0037199975922703743, + 0.004551922902464867, + 0.013486760668456554, + 0.006315629463642836, + 0.00691979518160224, + 0.005172519478946924, + 0.00982273556292057, + 0.002761311363428831, + 0.001903786789625883, + 0.002489787992089987, + 0.0030692433938384056, + 0.002965852851048112, + 0.004638814367353916, + 0.0018135084537789226, + 0.002776366425678134, + 0.007458222098648548, + 0.008470085449516773, + 0.009397213347256184, + 0.011982175521552563, + 0.014366312883794308, + 0.018369393423199654, + 0.008134552277624607, + 0.009118728339672089, + 0.024580590426921844, + 0.00820025335997343, + 0.007881250232458115, + 0.016157058998942375, + 0.010153347626328468, + 0.009375257417559624, + 0.0077139874920248985, + 0.02460474707186222, + 0.1361844837665558, + 0.027209553867578506, + 0.02841082215309143, + 0.1986972540616989, + 0.024862324818968773, + 0.025114962831139565, + 0.04223789647221565, + 0.02153165638446808, + 0.005226485431194305, + 0.004767706152051687, + 0.004556023981422186, + 0.0064328936859965324, + 0.006476133596152067, + 0.006006775889545679, + 0.005955060478299856, + 0.005426287651062012, + 0.003058480331674218, + 0.004003201611340046, + 0.0028657130897045135, + 0.0075242179445922375, + 0.0044489470310509205, + 0.004755365662276745, + 0.0043962812051177025, + 0.00436039362102747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.10033264011144638 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, full-body shot, taken from a level angle, captures a tall, weathered oak bookshelf, crammed with a variety of ancient books. The wood is a deep walnut, its surface covered with minute scratches and the patina of time, its edges carved with delicate, though worn, classical patterns. Hardcover books, their leather and cloth covers in faded shades of crimson, forest green, and navy blue, are tightly packed on the shelves. The faint, gilded titles on the spines appear faintly in the dim light. The scene is enveloped in the atmosphere of a quiet study, with a soft light slanting in from outside the frame, illuminating the dust particles dancing in the air. This transformation begins quietly at the upper right corner of the bookshelf, where the solid wooden edge begins to silently disintegrate, not shattering, but gracefully disintegrating into countless tiny, luminous digital particles. Like a swarm of disturbed fireflies, these particles emit a bright, cold glow of cyan and white, contrasting sharply with the warm tones of the study. They slowly drifted upwards as if drawn by an invisible current, the disintegration effect like a silent wave, slowly and inexorably spreading downwards along the contours of the bookshelf. Then, the ancient books began to melt from the edges of the pages, and the paper and text turned into equally glittering data streams, merging into the ever-expanding galaxy of light and code, creating a surreal atmosphere that was both sad and full of wonder, symbolizing the tranquil evolution of physical knowledge into digital form.", + "seed": 11628, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_5/index_5530_seed_11628_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12009877080563455, + "vlm_logprob": -0.12009877080563455, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.12009877080563455, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4391302168369293 + ], + "opposite_logprob_tokens": [ + -1.064130187034607 + ], + "logprob_reward": -0.4391302168369293, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4391302168369293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13108041882514954, + 0.05778645724058151, + 0.02716614492237568, + 0.05218074843287468, + 0.05765733867883682, + 0.10767443478107452, + 0.1514980047941208, + 0.12982790172100067, + 0.1449630707502365, + 0.14016549289226532 + ], + "credit_map": [ + 0.019269758835434914, + 0.009293765760958195, + 0.024388570338487625, + 0.013413465581834316, + 0.012643507681787014, + 0.008978340774774551, + 0.020454613491892815, + 0.022638408467173576, + 0.006324110087007284, + 0.002623805543407798, + 0.008448394946753979, + 0.00973268412053585, + 0.0041576894000172615, + 0.003306375117972493, + 0.012447545304894447, + 0.010745853185653687, + 0.0038130434695631266, + 0.0013439460890367627, + 0.004668662324547768, + 0.003888908540830016, + 0.002940429374575615, + 0.0021101960446685553, + 0.005105060525238514, + 0.0032958979718387127, + 0.00981780607253313, + 0.0022450913675129414, + 0.01685587875545025, + 0.00526466965675354, + 0.006567427888512611, + 0.0026904428377747536, + 0.005675849970430136, + 0.0030635797884315252, + 0.006903429515659809, + 0.002625275170430541, + 0.009067889302968979, + 0.006499935407191515, + 0.0073064034804701805, + 0.007308660540729761, + 0.010492943227291107, + 0.007452803663909435, + 0.011370127089321613, + 0.013215160928666592, + 0.017621269449591637, + 0.012728058733046055, + 0.009112491272389889, + 0.009020133875310421, + 0.017035748809576035, + 0.01757144182920456, + 0.02286629192531109, + 0.014803758822381496, + 0.020393144339323044, + 0.020839257165789604, + 0.02191551774740219, + 0.011328259482979774, + 0.01971801556646824, + 0.01963375136256218, + 0.02189793437719345, + 0.026433348655700684, + 0.015482312068343163, + 0.015219268389046192, + 0.011739930137991905, + 0.011247078888118267, + 0.011641690507531166, + 0.01616634614765644, + 0.028323782607913017, + 0.009654459543526173, + 0.03565727546811104, + 0.012353230267763138, + 0.010596279054880142, + 0.017359884455800056, + 0.014317106455564499, + 0.01670106314122677, + 0.0111306793987751, + 0.005976882763206959, + 0.0404215045273304, + 0.03795226290822029, + 0.009088877588510513, + 0.010172410868108273, + 0.01088783610612154, + 0.014535030350089073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.4391302168369293 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026388689875602722 + ], + "opposite_logprob_tokens": [ + -3.651388645172119 + ], + "logprob_reward": -0.026388689875602722, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026388689875602722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16882024705410004, + 0.06887413561344147, + 0.0805780366063118, + 0.1019827201962471, + 0.11043994128704071, + 0.10478276014328003, + 0.10686983168125153, + 0.05733092501759529, + 0.13569560647010803, + 0.06462587416172028 + ], + "credit_map": [ + 0.01853521354496479, + 0.01136095728725195, + 0.02652415633201599, + 0.022537169978022575, + 0.021199043840169907, + 0.013116888701915741, + 0.0346236489713192, + 0.02092313952744007, + 0.00585225410759449, + 0.0059633138589560986, + 0.009362809360027313, + 0.008000114001333714, + 0.01025039330124855, + 0.006561078131198883, + 0.014460686594247818, + 0.008423476479947567, + 0.006767309736460447, + 0.006698361597955227, + 0.00984770804643631, + 0.008920577354729176, + 0.01180446520447731, + 0.0077119264751672745, + 0.017136484384536743, + 0.011691199615597725, + 0.009556395001709461, + 0.006396757438778877, + 0.01170768029987812, + 0.01424178946763277, + 0.00934362132102251, + 0.008075588382780552, + 0.023417094722390175, + 0.01924378238618374, + 0.010394829325377941, + 0.005768296774476767, + 0.016916710883378983, + 0.013584881089627743, + 0.011668515391647816, + 0.01592804118990898, + 0.024371623992919922, + 0.011807030066847801, + 0.0107539352029562, + 0.0139211006462574, + 0.0164270531386137, + 0.013217455707490444, + 0.0077905599027872086, + 0.011510641314089298, + 0.01267300359904766, + 0.018488993868231773, + 0.01761603355407715, + 0.011357813142240047, + 0.013748814351856709, + 0.014867196790874004, + 0.014715941622853279, + 0.008365796878933907, + 0.011400802060961723, + 0.01479742769151926, + 0.01050425972789526, + 0.010287842713296413, + 0.007474894635379314, + 0.006407667882740498, + 0.005597664974629879, + 0.006520107854157686, + 0.005069476552307606, + 0.0054690046235919, + 0.03243809938430786, + 0.007455512881278992, + 0.03950635716319084, + 0.010381660424172878, + 0.00943970400840044, + 0.009689581580460072, + 0.01373385265469551, + 0.013050825335085392, + 0.005476030055433512, + 0.00480321329087019, + 0.007908212020993233, + 0.012282952666282654, + 0.006783506367355585, + 0.007934832945466042, + 0.007864950224757195, + 0.011572174727916718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.026388689875602722 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898139953613281 + ], + "opposite_logprob_tokens": [ + -2.578981399536133 + ], + "logprob_reward": -0.07898139953613281, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898139953613281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07599987834692001, + 0.06824308633804321, + 0.04741857573390007, + 0.06080249696969986, + 0.16231898963451385, + 0.10794699937105179, + 0.143167182803154, + 0.09936707466840744, + 0.10684773325920105, + 0.127888023853302 + ], + "credit_map": [ + 0.01209969725459814, + 0.004868432879447937, + 0.010889209806919098, + 0.014719833619892597, + 0.010220296680927277, + 0.00403686473146081, + 0.012575769796967506, + 0.006589765194803476, + 0.011756883934140205, + 0.006399509031325579, + 0.009352556429803371, + 0.008161740377545357, + 0.010523275472223759, + 0.005609805230051279, + 0.009224017150700092, + 0.007215290796011686, + 0.009202549234032631, + 0.003956285305321217, + 0.008226717822253704, + 0.00426934938877821, + 0.004279484041035175, + 0.0035184952430427074, + 0.008412100374698639, + 0.0055535938590765, + 0.014530361630022526, + 0.002834285842254758, + 0.0157571192830801, + 0.0070020779967308044, + 0.0037594137247651815, + 0.004931231029331684, + 0.005630198400467634, + 0.006357806269079447, + 0.02759597636759281, + 0.009242222644388676, + 0.03771958872675896, + 0.017045525833964348, + 0.017884282395243645, + 0.022473782300949097, + 0.012614988721907139, + 0.01774260774254799, + 0.009261076338589191, + 0.010187109000980854, + 0.014980300329625607, + 0.013184797950088978, + 0.015220525674521923, + 0.015768328681588173, + 0.015020119957625866, + 0.014324738644063473, + 0.022731972858309746, + 0.012518494389951229, + 0.018969574943184853, + 0.01801154389977455, + 0.016441741958260536, + 0.01202533021569252, + 0.02294107712805271, + 0.01952744647860527, + 0.012498326599597931, + 0.01835760287940502, + 0.01256142370402813, + 0.010834245011210442, + 0.009800078347325325, + 0.013474039733409882, + 0.011233344674110413, + 0.010608000680804253, + 0.030182216316461563, + 0.005872518289834261, + 0.03583827242255211, + 0.007063273340463638, + 0.00617036921903491, + 0.005586189217865467, + 0.007649815641343594, + 0.008485067635774612, + 0.00821736641228199, + 0.005626642145216465, + 0.04033771902322769, + 0.03978957235813141, + 0.007288500666618347, + 0.0074057686142623425, + 0.009803855791687965, + 0.009418592788279057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.07898139953613281 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010444429935887456 + ], + "opposite_logprob_tokens": [ + -6.876044273376465 + ], + "logprob_reward": -0.0010444429935887456, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010444429935887456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11193794757127762, + 0.03916025906801224, + 0.02552536129951477, + 0.0228718351572752, + 0.04868988320231438, + 0.08255307376384735, + 0.08893779665231705, + 0.06324836611747742, + 0.317279189825058, + 0.1997963786125183 + ], + "credit_map": [ + 0.018057065084576607, + 0.006680286955088377, + 0.021708494052290916, + 0.013494213111698627, + 0.010724482126533985, + 0.007653676904737949, + 0.015749376267194748, + 0.01787034049630165, + 0.004696353804320097, + 0.0020096898078918457, + 0.00545884482562542, + 0.006149794906377792, + 0.004001400899142027, + 0.002110459143295884, + 0.009251614101231098, + 0.00548209622502327, + 0.003728098003193736, + 0.0014014854095876217, + 0.003932942636311054, + 0.003361210459843278, + 0.0033456936944276094, + 0.0025823507457971573, + 0.0043440693989396095, + 0.0028295095544308424, + 0.002555408515036106, + 0.001577368238940835, + 0.0033144408371299505, + 0.003570321248844266, + 0.002638568403199315, + 0.0013142020907253027, + 0.004545775707811117, + 0.003355747787281871, + 0.0073156231082975864, + 0.002284338464960456, + 0.012595182284712791, + 0.0066039543598890305, + 0.0034935115836560726, + 0.004495549947023392, + 0.006827552802860737, + 0.005074168089777231, + 0.009766141884028912, + 0.011662653647363186, + 0.010338708758354187, + 0.009453851729631424, + 0.006685988511890173, + 0.008031024597585201, + 0.011875084601342678, + 0.014739609323441982, + 0.01305258646607399, + 0.01381448283791542, + 0.01053391583263874, + 0.009182555601000786, + 0.013158172369003296, + 0.007542085368186235, + 0.00978053454309702, + 0.011873459443449974, + 0.00871912483125925, + 0.008059455081820488, + 0.00799685437232256, + 0.00631541246548295, + 0.006228818092495203, + 0.009678823873400688, + 0.007895441725850105, + 0.008354430086910725, + 0.07559303194284439, + 0.01584770157933235, + 0.09600324928760529, + 0.026180816814303398, + 0.019207589328289032, + 0.03137694299221039, + 0.02645190991461277, + 0.026617923751473427, + 0.021584099158644676, + 0.00933211762458086, + 0.04139325022697449, + 0.06687178462743759, + 0.012234017252922058, + 0.01627309061586857, + 0.017991777509450912, + 0.014116216450929642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0010444429935887456 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494910478591919 + ], + "opposite_logprob_tokens": [ + -2.9299490451812744 + ], + "logprob_reward": -0.05494910478591919, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494910478591919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13434134423732758, + 0.051237817853689194, + 0.02585008554160595, + 0.04053601250052452, + 0.075123630464077, + 0.0655018538236618, + 0.17986242473125458, + 0.047769445925951004, + 0.16688434779644012, + 0.2128930687904358 + ], + "credit_map": [ + 0.017405841499567032, + 0.006748158950358629, + 0.02229105681180954, + 0.023821843788027763, + 0.013195292092859745, + 0.009316040202975273, + 0.024130361154675484, + 0.017432739958167076, + 0.0071348510682582855, + 0.003045292804017663, + 0.008969559334218502, + 0.008380505256354809, + 0.003788923379033804, + 0.004026123322546482, + 0.00904017686843872, + 0.006852380465716124, + 0.00320201413705945, + 0.0015621117781847715, + 0.00373088289052248, + 0.004226417280733585, + 0.0028341941069811583, + 0.0026231175288558006, + 0.004686320666223764, + 0.002985024359077215, + 0.00644214591011405, + 0.0024072062224149704, + 0.005774578079581261, + 0.0052818120457232, + 0.003193991258740425, + 0.002581752836704254, + 0.006721471436321735, + 0.00813305377960205, + 0.009793645702302456, + 0.0035625838208943605, + 0.01097775623202324, + 0.010404069907963276, + 0.008009199053049088, + 0.010377335362136364, + 0.013824613764882088, + 0.008174429647624493, + 0.009102540090680122, + 0.006779162213206291, + 0.012681961059570312, + 0.008315862156450748, + 0.004885572474449873, + 0.0062784552574157715, + 0.005836306139826775, + 0.011621985584497452, + 0.026049043983221054, + 0.015306237153708935, + 0.02239270508289337, + 0.028219230473041534, + 0.024127835407853127, + 0.02068483456969261, + 0.019215598702430725, + 0.023866916075348854, + 0.0063320305198431015, + 0.005602909252047539, + 0.006161495577543974, + 0.004714478272944689, + 0.005079593509435654, + 0.006575156934559345, + 0.005942725110799074, + 0.0073610530234873295, + 0.03391183167695999, + 0.011668851599097252, + 0.03788375481963158, + 0.017263077199459076, + 0.013101819902658463, + 0.026170609518885612, + 0.01429925486445427, + 0.012585132382810116, + 0.016892675310373306, + 0.007044587284326553, + 0.0727870911359787, + 0.06108156591653824, + 0.013854550197720528, + 0.014833779074251652, + 0.01641700603067875, + 0.009981808252632618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.05494910478591919 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, full-body shot, taken from a level angle, captures a tall, weathered oak bookshelf, crammed with a variety of ancient books. The wood is a deep walnut, its surface covered with minute scratches and the patina of time, its edges carved with delicate, though worn, classical patterns. Hardcover books, their leather and cloth covers in faded shades of crimson, forest green, and navy blue, are tightly packed on the shelves. The faint, gilded titles on the spines appear faintly in the dim light. The scene is enveloped in the atmosphere of a quiet study, with a soft light slanting in from outside the frame, illuminating the dust particles dancing in the air. This transformation begins quietly at the upper right corner of the bookshelf, where the solid wooden edge begins to silently disintegrate, not shattering, but gracefully disintegrating into countless tiny, luminous digital particles. Like a swarm of disturbed fireflies, these particles emit a bright, cold glow of cyan and white, contrasting sharply with the warm tones of the study. They slowly drifted upwards as if drawn by an invisible current, the disintegration effect like a silent wave, slowly and inexorably spreading downwards along the contours of the bookshelf. Then, the ancient books began to melt from the edges of the pages, and the paper and text turned into equally glittering data streams, merging into the ever-expanding galaxy of light and code, creating a surreal atmosphere that was both sad and full of wonder, symbolizing the tranquil evolution of physical knowledge into digital form.", + "seed": 11629, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_5/index_5530_seed_11629_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2601074583362788, + "vlm_logprob": -0.2601074583362788, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.2601074583362788, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9032436609268188 + ], + "opposite_logprob_tokens": [ + -0.5282436609268188 + ], + "logprob_reward": -0.9032436609268188, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9032436609268188, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22855301201343536, + 0.07016371190547943, + 0.06875085830688477, + 0.13225945830345154, + 0.05710812285542488, + 0.1017092764377594, + 0.11573547124862671, + 0.0544499047100544, + 0.07199986279010773, + 0.099270299077034 + ], + "credit_map": [ + 0.03598082438111305, + 0.032667938619852066, + 0.019609836861491203, + 0.034944381564855576, + 0.05503512918949127, + 0.01347203366458416, + 0.015787797048687935, + 0.021055079996585846, + 0.010650656186044216, + 0.013989169150590897, + 0.00418711407110095, + 0.0065369936637580395, + 0.011413993313908577, + 0.004526561591774225, + 0.0066771600395441055, + 0.012182062491774559, + 0.006330311298370361, + 0.006408863700926304, + 0.006598669104278088, + 0.018954584375023842, + 0.015295392833650112, + 0.003982803784310818, + 0.006169606000185013, + 0.005010632798075676, + 0.012592089362442493, + 0.014981487765908241, + 0.022905053570866585, + 0.025692801922559738, + 0.012830547988414764, + 0.011244057677686214, + 0.01758323609828949, + 0.014430184848606586, + 0.0057649496011435986, + 0.009754561819136143, + 0.00776160741224885, + 0.0059812115505337715, + 0.010348671115934849, + 0.004257029388099909, + 0.006050126627087593, + 0.007189967669546604, + 0.009650436230003834, + 0.01182254496961832, + 0.013288646005094051, + 0.022779840975999832, + 0.012885797768831253, + 0.008963007479906082, + 0.009753004647791386, + 0.012566001154482365, + 0.011044682003557682, + 0.013881758786737919, + 0.012562901712954044, + 0.011135884560644627, + 0.014575337059795856, + 0.014824469573795795, + 0.02882200852036476, + 0.00888843834400177, + 0.008147243410348892, + 0.009626975283026695, + 0.006759626790881157, + 0.005045260302722454, + 0.006577438674867153, + 0.006891375873237848, + 0.007181031629443169, + 0.004220954142510891, + 0.004521200433373451, + 0.004688160493969917, + 0.0032447685953229666, + 0.005731933284550905, + 0.02034965343773365, + 0.004056934267282486, + 0.025250954553484917, + 0.004156252834945917, + 0.013779466971755028, + 0.017787568271160126, + 0.013466695323586464, + 0.008807257749140263, + 0.013761303387582302, + 0.008564778603613377, + 0.012312982231378555, + 0.010790239088237286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.9032436609268188 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018221836537122726 + ], + "opposite_logprob_tokens": [ + -4.018221855163574 + ], + "logprob_reward": -0.018221836537122726, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018221836537122726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13170307874679565, + 0.0798703208565712, + 0.14659009873867035, + 0.10048633813858032, + 0.06942523270845413, + 0.05830321088433266, + 0.24249468743801117, + 0.05082365870475769, + 0.06308797001838684, + 0.05721539258956909 + ], + "credit_map": [ + 0.017132310196757317, + 0.023234086111187935, + 0.013543306849896908, + 0.01777283474802971, + 0.022445522248744965, + 0.00950909685343504, + 0.01688770391047001, + 0.011178215965628624, + 0.009120882488787174, + 0.01357630081474781, + 0.011339392513036728, + 0.015629805624485016, + 0.015262222848832607, + 0.004214802756905556, + 0.006703783292323351, + 0.004023126792162657, + 0.013837792910635471, + 0.014261892065405846, + 0.01607540436089039, + 0.03915828466415405, + 0.04046870023012161, + 0.007817087695002556, + 0.008557846769690514, + 0.006413095630705357, + 0.006883458700031042, + 0.008023932576179504, + 0.0095940250903368, + 0.030328957363963127, + 0.02381769008934498, + 0.006296325474977493, + 0.008610541932284832, + 0.006931405980139971, + 0.007075937930494547, + 0.011485232971608639, + 0.0071323588490486145, + 0.007797149010002613, + 0.011012356728315353, + 0.006539649795740843, + 0.009481741115450859, + 0.008900811895728111, + 0.00604670075699687, + 0.00807864684611559, + 0.007788782939314842, + 0.012966719456017017, + 0.007761787157505751, + 0.0043603116646409035, + 0.005337005015462637, + 0.005963255651295185, + 0.016510333865880966, + 0.01575479470193386, + 0.006210094317793846, + 0.014323703944683075, + 0.06948486715555191, + 0.012970452196896076, + 0.09110385924577713, + 0.01613658480346203, + 0.007016920484602451, + 0.006964182946830988, + 0.007497589103877544, + 0.004907913040369749, + 0.0063687097281217575, + 0.007565378677099943, + 0.005469645839184523, + 0.005033318884670734, + 0.004046065732836723, + 0.00451318034902215, + 0.0024394672363996506, + 0.004808508791029453, + 0.019091328606009483, + 0.0031825886107981205, + 0.02153445966541767, + 0.003472365206107497, + 0.0064305891282856464, + 0.008172815665602684, + 0.0065820831805467606, + 0.0071088410913944244, + 0.013388378545641899, + 0.004436831455677748, + 0.005579757038503885, + 0.005516091827303171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.018221836537122726 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31341424584388733 + ], + "opposite_logprob_tokens": [ + -1.313414216041565 + ], + "logprob_reward": -0.31341424584388733, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31341424584388733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09222844243049622, + 0.03679840639233589, + 0.0873929113149643, + 0.09185424447059631, + 0.09787661582231522, + 0.08053969591856003, + 0.47274208068847656, + 0.011635315604507923, + 0.012310906313359737, + 0.01662137545645237 + ], + "credit_map": [ + 0.012710130773484707, + 0.01132845040410757, + 0.007887843064963818, + 0.015488368459045887, + 0.0249334629625082, + 0.005082817282527685, + 0.0057000587694346905, + 0.009097309783101082, + 0.004855869337916374, + 0.004547129385173321, + 0.003432455938309431, + 0.0035426230169832706, + 0.0035831190180033445, + 0.0033698026090860367, + 0.006735374219715595, + 0.006732032634317875, + 0.00870366208255291, + 0.009198163636028767, + 0.007904252968728542, + 0.024360090494155884, + 0.022534484043717384, + 0.0041253408417105675, + 0.006220935843884945, + 0.004345981869846582, + 0.00943736732006073, + 0.010316479951143265, + 0.014202800579369068, + 0.018381673842668533, + 0.013442645780742168, + 0.0076834396459162235, + 0.009423828683793545, + 0.008966002613306046, + 0.015305559150874615, + 0.016848448663949966, + 0.009978090412914753, + 0.008684891276061535, + 0.011839956045150757, + 0.007678557652980089, + 0.013661465607583523, + 0.013879647478461266, + 0.008936312049627304, + 0.01125943660736084, + 0.00897134467959404, + 0.011627581901848316, + 0.008662911131978035, + 0.0073293172754347324, + 0.007381753530353308, + 0.016371039673686028, + 0.03199047967791557, + 0.03428354859352112, + 0.01253820862621069, + 0.027475588023662567, + 0.14006957411766052, + 0.024439087137579918, + 0.17142608761787415, + 0.0305195115506649, + 0.0016450349939987063, + 0.001977089326828718, + 0.0015155684668570757, + 0.001088324119336903, + 0.0014244192279875278, + 0.0015561631880700588, + 0.0012856258545070887, + 0.001143090077675879, + 0.0017947988817468286, + 0.0019119606586173177, + 0.0011197988642379642, + 0.0011175837134942412, + 0.0024928105995059013, + 0.0012830090709030628, + 0.0016002869233489037, + 0.0009906577179208398, + 0.0020772493444383144, + 0.0029063725378364325, + 0.0024560701567679644, + 0.0015486808260902762, + 0.0017864012625068426, + 0.0018948093056678772, + 0.002473480999469757, + 0.0014783113729208708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.31341424584388733 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036159863229840994 + ], + "opposite_logprob_tokens": [ + -5.628615856170654 + ], + "logprob_reward": -0.0036159863229840994, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036159863229840994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0881594717502594, + 0.025808846578001976, + 0.024893151596188545, + 0.05241012945771217, + 0.061107560992240906, + 0.0540824718773365, + 0.5117818713188171, + 0.08438953012228012, + 0.024192029610276222, + 0.07317489385604858 + ], + "credit_map": [ + 0.012093301862478256, + 0.012222755700349808, + 0.008864592760801315, + 0.01253129169344902, + 0.02074417471885681, + 0.005635848734527826, + 0.00755993602797389, + 0.008507570251822472, + 0.0034701358526945114, + 0.003903059521690011, + 0.004005826078355312, + 0.005078931339085102, + 0.0028705247677862644, + 0.0014542210847139359, + 0.002852723700925708, + 0.0021734235342592, + 0.003413404105231166, + 0.003821109188720584, + 0.0028037354350090027, + 0.0035752339754253626, + 0.0031927048694342375, + 0.0016079925699159503, + 0.003490020055323839, + 0.002988952212035656, + 0.005507978610694408, + 0.007418239489197731, + 0.010284557938575745, + 0.007868872955441475, + 0.0059964945539832115, + 0.004563191439956427, + 0.005611258093267679, + 0.005159540567547083, + 0.006861689034849405, + 0.008026093244552612, + 0.011110207065939903, + 0.005535264033824205, + 0.009132858365774155, + 0.0059792883694171906, + 0.005616544745862484, + 0.008845621719956398, + 0.004782397300004959, + 0.011349416337907314, + 0.004827008582651615, + 0.007001069840043783, + 0.006017629988491535, + 0.005070317070931196, + 0.008091496303677559, + 0.00694313645362854, + 0.03400241211056709, + 0.02877158857882023, + 0.016990549862384796, + 0.03006792813539505, + 0.1531251072883606, + 0.02823963388800621, + 0.18829897046089172, + 0.03228567913174629, + 0.011293462477624416, + 0.01527378149330616, + 0.015401467680931091, + 0.007606483530253172, + 0.011703544296324253, + 0.008242251351475716, + 0.007173760794103146, + 0.007694779429584742, + 0.0020251201931387186, + 0.003482136642560363, + 0.002075847703963518, + 0.0028805346228182316, + 0.005164622329175472, + 0.0021246832329779863, + 0.004868173040449619, + 0.0015709104482084513, + 0.006860652472823858, + 0.006655785255134106, + 0.004754579160362482, + 0.01658482849597931, + 0.025226691737771034, + 0.006070614792406559, + 0.004242562688887119, + 0.0027791818138211966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.0036159863229840994 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a tall, weathered oak bookshelf made of deep walnut wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are hardcover books with faded crimson, forest green, and navy blue covers tightly packed on the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the disintegration effect begin at the upper right corner of the bookshelf and spread downwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrating particles emit a bright, cold glow of cyan and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the ancient books melt from the edges of their pages into glittering data streams?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204156205058098 + ], + "opposite_logprob_tokens": [ + -2.8120415210723877 + ], + "logprob_reward": -0.06204156205058098, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204156205058098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09762917459011078, + 0.03391197323799133, + 0.05330663174390793, + 0.10272938013076782, + 0.07742320746183395, + 0.06230631470680237, + 0.40144240856170654, + 0.05005732551217079, + 0.07386333495378494, + 0.04733027517795563 + ], + "credit_map": [ + 0.01211217325180769, + 0.01705789938569069, + 0.008753187023103237, + 0.012736287899315357, + 0.018367962911725044, + 0.005757189821451902, + 0.007808174472302198, + 0.015036308206617832, + 0.00445330748334527, + 0.005996235180646181, + 0.0038813750725239515, + 0.00448941346257925, + 0.005210420116782188, + 0.0021169797983020544, + 0.0026573261711746454, + 0.005106918979436159, + 0.005632579792290926, + 0.006277929525822401, + 0.005189354997128248, + 0.01345919445157051, + 0.012000685557723045, + 0.003114941529929638, + 0.004354987293481827, + 0.003276961389929056, + 0.011104023084044456, + 0.009033722802996635, + 0.014650996774435043, + 0.01730683445930481, + 0.023589812219142914, + 0.007316566072404385, + 0.011934584937989712, + 0.0077928462997078896, + 0.009143934585154057, + 0.015881270170211792, + 0.00760948471724987, + 0.007895647548139095, + 0.00896557979285717, + 0.006565352436155081, + 0.010809656232595444, + 0.010552278719842434, + 0.006151569075882435, + 0.009827619418501854, + 0.005574523005634546, + 0.009595751762390137, + 0.008837229572236538, + 0.006359249819070101, + 0.008150172419846058, + 0.007810202892869711, + 0.02538972906768322, + 0.021968990564346313, + 0.011526252143085003, + 0.024005291983485222, + 0.11997037380933762, + 0.023567628115415573, + 0.14958617091178894, + 0.02542797476053238, + 0.005471214186400175, + 0.005415054038167, + 0.011099472641944885, + 0.004917404148727655, + 0.005798535421490669, + 0.006070716772228479, + 0.007525249384343624, + 0.0037596779875457287, + 0.0051060086116194725, + 0.006232722662389278, + 0.003912616521120071, + 0.005590522661805153, + 0.0217449814081192, + 0.004170220345258713, + 0.022953931242227554, + 0.004152337554842234, + 0.004432876594364643, + 0.005062200594693422, + 0.002409841399639845, + 0.011348383501172066, + 0.015585401095449924, + 0.003701033303514123, + 0.0030953658279031515, + 0.0016951749566942453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.06204156205058098 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_6/index_tensor([5281, 5281])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_6/index_tensor([5281, 5281])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..77f7532673becbcbdcbb0af92ba6d6ddaf4c0f03 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_6/index_tensor([5281, 5281])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic low-angle wide shot slowly tilts upward, advancing almost imperceptibly, to capture the terrifying majesty of the vast, dark vortex of storm clouds dominating the sky. This roiling vortex, composed of dense, roiling cumulonimbus clouds in shades of inky black, bruised purple, and menacing slate gray, spirals with a violent, relentless energy. Within this surging, swirling chaos, a constant stream of shattered debris pours forth; broken wooden beams, shards of twisted structural metal, and fragments of mangled infrastructure are flung outward by gravity, only to be sucked back into the vortex's dark center. The only illumination in this apocalyptic scene comes from the storm's own eerie internal glow and sporadic, brilliant flashes of lightning, which momentarily outline a violent ballet of destruction. The entire image is suffused with an icy, desaturated palette, creating an intense and overwhelming atmosphere of raw, untamed power and absolute devastation.", + "seed": 3304, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_6/index_5281_seed_3304_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06550120571628212, + "vlm_logprob": -0.06550120571628212, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.06550120571628212, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16760563850402832 + ], + "opposite_logprob_tokens": [ + -1.9176056385040283 + ], + "logprob_reward": -0.16760563850402832, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16760563850402832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18350905179977417, + 0.04213207587599754, + 0.04261193796992302, + 0.06897135078907013, + 0.08012724667787552, + 0.03625265136361122, + 0.08418474346399307, + 0.08416995406150818, + 0.2717364430427551, + 0.10630453377962112 + ], + "credit_map": [ + 0.010719374753534794, + 0.018383735790848732, + 0.01804930903017521, + 0.010628961957991123, + 0.049600329250097275, + 0.028859036043286324, + 0.0219760462641716, + 0.025292279198765755, + 0.0043310970067977905, + 0.004183038137853146, + 0.005692984908819199, + 0.006027995143085718, + 0.007390552666038275, + 0.005465303547680378, + 0.004136864095926285, + 0.004904237110167742, + 0.005055467132478952, + 0.003951288294047117, + 0.004582439083606005, + 0.007400935050100088, + 0.003784916363656521, + 0.005559379700571299, + 0.006877403240650892, + 0.005400109104812145, + 0.0053015281446278095, + 0.008870909921824932, + 0.01013422105461359, + 0.011278273537755013, + 0.009730630554258823, + 0.009093866683542728, + 0.0074157738126814365, + 0.007146148942410946, + 0.008092773146927357, + 0.009831862524151802, + 0.013851356692612171, + 0.008800085633993149, + 0.009466491639614105, + 0.012986964546144009, + 0.008646328933537006, + 0.008451376110315323, + 0.0039366683922708035, + 0.004793485160917044, + 0.003240635385736823, + 0.003440139116719365, + 0.005390619393438101, + 0.005185351707041264, + 0.005339163355529308, + 0.004926589783281088, + 0.006500259507447481, + 0.010061497800052166, + 0.00935153104364872, + 0.008797168731689453, + 0.011061646975576878, + 0.013120737858116627, + 0.013076618313789368, + 0.012215289287269115, + 0.007907121442258358, + 0.00989222526550293, + 0.0069688456133008, + 0.0069251637905836105, + 0.014014515094459057, + 0.0124722421169281, + 0.015004260465502739, + 0.010985581204295158, + 0.03388306871056557, + 0.03236503154039383, + 0.01854519546031952, + 0.018518606200814247, + 0.02890581078827381, + 0.04526054486632347, + 0.025596264749765396, + 0.06866193562746048, + 0.005846927408128977, + 0.03896758705377579, + 0.0038604564033448696, + 0.004759598523378372, + 0.0055030714720487595, + 0.03984864056110382, + 0.0025334632955491543, + 0.004984785337001085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.16760563850402832 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003478467464447 + ], + "opposite_logprob_tokens": [ + -2.3503477573394775 + ], + "logprob_reward": -0.1003478467464447, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003478467464447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13237148523330688, + 0.07632340490818024, + 0.09649042785167694, + 0.08094196766614914, + 0.11519276350736618, + 0.08517464250326157, + 0.09081291407346725, + 0.06296052038669586, + 0.1291080266237259, + 0.13062380254268646 + ], + "credit_map": [ + 0.01674721948802471, + 0.01674540340900421, + 0.02022159844636917, + 0.017798881977796555, + 0.015019499696791172, + 0.01431339606642723, + 0.011894683353602886, + 0.01963081769645214, + 0.013519743457436562, + 0.0062628923915326595, + 0.008929763920605183, + 0.010034804232418537, + 0.008984518237411976, + 0.010412685573101044, + 0.008815448731184006, + 0.009363553486764431, + 0.017283596098423004, + 0.006723487749695778, + 0.009977278299629688, + 0.016547877341508865, + 0.009076181799173355, + 0.011480361223220825, + 0.011668660677969456, + 0.013732992112636566, + 0.00629421416670084, + 0.011497369036078453, + 0.010416439734399319, + 0.014025608077645302, + 0.010370483621954918, + 0.011645586229860783, + 0.008408242836594582, + 0.008284028619527817, + 0.012407319620251656, + 0.011543809436261654, + 0.027419712394475937, + 0.009313477203249931, + 0.012398980557918549, + 0.01678048074245453, + 0.013720390386879444, + 0.011608604341745377, + 0.00914714764803648, + 0.008679261431097984, + 0.00864423904567957, + 0.008974628522992134, + 0.014053694903850555, + 0.013906403444707394, + 0.012545009143650532, + 0.009224265813827515, + 0.010502343066036701, + 0.008592601865530014, + 0.012936228886246681, + 0.00787374097853899, + 0.014062861911952496, + 0.013541488908231258, + 0.012671775184571743, + 0.010631881654262543, + 0.005082913674414158, + 0.007080131210386753, + 0.0037191943265497684, + 0.005761586595326662, + 0.012866239994764328, + 0.008651881478726864, + 0.011816252954304218, + 0.007982322946190834, + 0.015678182244300842, + 0.0157963614910841, + 0.011686228215694427, + 0.011937125585973263, + 0.013647064566612244, + 0.02173394151031971, + 0.012765388935804367, + 0.02586374245584011, + 0.008672533556818962, + 0.01846889592707157, + 0.008550065569579601, + 0.01027030311524868, + 0.008320984430611134, + 0.042292337864637375, + 0.006309614982455969, + 0.02773907221853733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.1003478467464447 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252846051007509 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.005252846051007509, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252846051007509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13802196085453033, + 0.06199401617050171, + 0.2759976089000702, + 0.03892994299530983, + 0.1676672101020813, + 0.013409798964858055, + 0.04082328453660011, + 0.05812409520149231, + 0.1410168707370758, + 0.06401519477367401 + ], + "credit_map": [ + 0.009238682687282562, + 0.01621062122285366, + 0.019788570702075958, + 0.013089142739772797, + 0.026651063933968544, + 0.02187313511967659, + 0.013484395109117031, + 0.017686355859041214, + 0.010664229281246662, + 0.005577517207711935, + 0.008555439300835133, + 0.009143158793449402, + 0.009005051106214523, + 0.007532610557973385, + 0.005437749903649092, + 0.006078257691115141, + 0.033205434679985046, + 0.020247137174010277, + 0.025744231417775154, + 0.04490284621715546, + 0.031000200659036636, + 0.0352247953414917, + 0.0361960232257843, + 0.04947692155838013, + 0.0027042324654757977, + 0.004544057883322239, + 0.003148954128846526, + 0.005348204635083675, + 0.008117428049445152, + 0.005666174925863743, + 0.005232564173638821, + 0.0041683269664645195, + 0.01646929420530796, + 0.01799563132226467, + 0.030479012057185173, + 0.01794881373643875, + 0.020098570734262466, + 0.02817152999341488, + 0.018897173926234245, + 0.017607176676392555, + 0.001529479632154107, + 0.0015072774840518832, + 0.0013925478560850024, + 0.001425071619451046, + 0.001846945728175342, + 0.0018090074881911278, + 0.0024046378675848246, + 0.001494831987656653, + 0.004596872255206108, + 0.005204483401030302, + 0.004330842290073633, + 0.0030028237961232662, + 0.005733861122280359, + 0.008195178583264351, + 0.004549401346594095, + 0.005209820810705423, + 0.005246894434094429, + 0.008160179480910301, + 0.006303343456238508, + 0.006405372638255358, + 0.0069756959564983845, + 0.011162066832184792, + 0.006069477181881666, + 0.00780106708407402, + 0.015692006796598434, + 0.01744397170841694, + 0.00884063821285963, + 0.007573644630610943, + 0.01533119473606348, + 0.02317557856440544, + 0.012712650932371616, + 0.040247198194265366, + 0.0033868190366774797, + 0.022015836089849472, + 0.002004164271056652, + 0.002071398077532649, + 0.0032942877151072025, + 0.02623271569609642, + 0.0015816683880984783, + 0.003428305732086301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.005252846051007509 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371411934494972 + ], + "opposite_logprob_tokens": [ + -3.4087140560150146 + ], + "logprob_reward": -0.03371411934494972, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371411934494972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19184045493602753, + 0.12057635933160782, + 0.061312828212976456, + 0.05126186087727547, + 0.23700428009033203, + 0.04390016570687294, + 0.04127516970038414, + 0.023930098861455917, + 0.19432266056537628, + 0.03457605838775635 + ], + "credit_map": [ + 0.027597716078162193, + 0.024349873885512352, + 0.03537445515394211, + 0.029307404533028603, + 0.019457269459962845, + 0.01853891648352146, + 0.013695471920073032, + 0.023519370704889297, + 0.02087724767625332, + 0.011484308168292046, + 0.013349237851798534, + 0.01949574612081051, + 0.01329649705439806, + 0.013653784058988094, + 0.011476101353764534, + 0.016943441703915596, + 0.007001122459769249, + 0.005393787287175655, + 0.006196903996169567, + 0.010015441104769707, + 0.006907627917826176, + 0.008031951263546944, + 0.008293571881949902, + 0.009472428821027279, + 0.004939438309520483, + 0.005180308595299721, + 0.009929369203746319, + 0.00699667539447546, + 0.006963606458157301, + 0.006157777272164822, + 0.0053110066801309586, + 0.005783684551715851, + 0.023567792028188705, + 0.03255679830908775, + 0.03631484508514404, + 0.025278175249695778, + 0.028082048520445824, + 0.04066092520952225, + 0.027167513966560364, + 0.023376213386654854, + 0.00445412565022707, + 0.004226295743137598, + 0.004984785802662373, + 0.00408079382032156, + 0.006937650963664055, + 0.006345519796013832, + 0.008104288950562477, + 0.0047667124308645725, + 0.0041553485207259655, + 0.004561606328934431, + 0.0059299287386238575, + 0.003913132473826408, + 0.00586659973487258, + 0.005850789602845907, + 0.006465328391641378, + 0.004532442428171635, + 0.0016047245590016246, + 0.004304361995309591, + 0.0022306872997432947, + 0.0027672210708260536, + 0.00346902827732265, + 0.00378433451987803, + 0.002701838267967105, + 0.0030679055489599705, + 0.021941157057881355, + 0.023072687909007072, + 0.01341050211340189, + 0.011312083341181278, + 0.022224703803658485, + 0.029841652140021324, + 0.018095970153808594, + 0.054423920810222626, + 0.0022182813845574856, + 0.009125053882598877, + 0.0019156120251864195, + 0.002274636412039399, + 0.0021327666472643614, + 0.012694405391812325, + 0.0012995385332033038, + 0.0029157702811062336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.03371411934494972 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585577934980392 + ], + "opposite_logprob_tokens": [ + -3.895585536956787 + ], + "logprob_reward": -0.020585577934980392, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585577934980392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1924496740102768, + 0.048050690442323685, + 0.12913623452186584, + 0.06625936925411224, + 0.153002068400383, + 0.022114822641015053, + 0.07092302292585373, + 0.04983729124069214, + 0.1744348257780075, + 0.09379202127456665 + ], + "credit_map": [ + 0.015367451123893261, + 0.019976120442152023, + 0.02186482399702072, + 0.019467556849122047, + 0.03654481843113899, + 0.033614616841077805, + 0.01772228442132473, + 0.027891995385289192, + 0.006334196776151657, + 0.004250724334269762, + 0.006219175644218922, + 0.007836021482944489, + 0.005975694395601749, + 0.006075192242860794, + 0.0049492535181343555, + 0.0064104278571903706, + 0.015885820612311363, + 0.0093842139467597, + 0.013244773261249065, + 0.020871173590421677, + 0.01438067015260458, + 0.01693182811141014, + 0.017334720119833946, + 0.0211030226200819, + 0.006180133204907179, + 0.006412177346646786, + 0.012935992330312729, + 0.007942769676446915, + 0.012126153334975243, + 0.008066093549132347, + 0.007567802909761667, + 0.005028238520026207, + 0.01571384072303772, + 0.017728770151734352, + 0.018558239564299583, + 0.020166946575045586, + 0.019086485728621483, + 0.02540126070380211, + 0.020014027133584023, + 0.016332486644387245, + 0.002312782220542431, + 0.0023606535978615284, + 0.0030714357271790504, + 0.0020535304211080074, + 0.003022492863237858, + 0.0026947231963276863, + 0.003841348923742771, + 0.002757854061201215, + 0.006466876249760389, + 0.00811992771923542, + 0.008693295530974865, + 0.007011069916188717, + 0.012655898928642273, + 0.010764210484921932, + 0.008378042839467525, + 0.008833694271743298, + 0.003947400953620672, + 0.007151954807341099, + 0.004783570766448975, + 0.004252850543707609, + 0.010837776586413383, + 0.006935411598533392, + 0.0064590140245854855, + 0.005469306372106075, + 0.015404991805553436, + 0.022731471806764603, + 0.010579288005828857, + 0.012277032248675823, + 0.020326336845755577, + 0.030930396169424057, + 0.01919172704219818, + 0.04299357160925865, + 0.005548720248043537, + 0.0331307053565979, + 0.003924244083464146, + 0.004392040893435478, + 0.0048088920302689075, + 0.030072493478655815, + 0.002567692194133997, + 0.009347232058644295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.020585577934980392 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic low-angle wide shot slowly tilts upward, advancing almost imperceptibly, to capture the terrifying majesty of the vast, dark vortex of storm clouds dominating the sky. This roiling vortex, composed of dense, roiling cumulonimbus clouds in shades of inky black, bruised purple, and menacing slate gray, spirals with a violent, relentless energy. Within this surging, swirling chaos, a constant stream of shattered debris pours forth; broken wooden beams, shards of twisted structural metal, and fragments of mangled infrastructure are flung outward by gravity, only to be sucked back into the vortex's dark center. The only illumination in this apocalyptic scene comes from the storm's own eerie internal glow and sporadic, brilliant flashes of lightning, which momentarily outline a violent ballet of destruction. The entire image is suffused with an icy, desaturated palette, creating an intense and overwhelming atmosphere of raw, untamed power and absolute devastation.", + "seed": 3305, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_6/index_5281_seed_3305_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14056731276214124, + "vlm_logprob": -0.14056731276214124, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.14056731276214124, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4338047504425049 + ], + "opposite_logprob_tokens": [ + -1.0588047504425049 + ], + "logprob_reward": -0.4338047504425049, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4338047504425049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19443118572235107, + 0.22067159414291382, + 0.07002416998147964, + 0.053412701934576035, + 0.07025756686925888, + 0.10819438099861145, + 0.08572567254304886, + 0.04913103207945824, + 0.0948406234383583, + 0.053311098366975784 + ], + "credit_map": [ + 0.014564772136509418, + 0.018862107768654823, + 0.019006723538041115, + 0.02853548154234886, + 0.021223027259111404, + 0.023018231615424156, + 0.026882903650403023, + 0.04233790934085846, + 0.02032662183046341, + 0.035403963178396225, + 0.01783718541264534, + 0.03967396169900894, + 0.028825029730796814, + 0.02257237210869789, + 0.03627243638038635, + 0.019760023802518845, + 0.0076107182539999485, + 0.009607804007828236, + 0.011019163765013218, + 0.012078149244189262, + 0.007115248125046492, + 0.00958728976547718, + 0.007491795811802149, + 0.0055139996111392975, + 0.006443897727876902, + 0.00562327029183507, + 0.005273033399134874, + 0.014619848690927029, + 0.004657141398638487, + 0.004899215884506702, + 0.007019277662038803, + 0.004877014551311731, + 0.010360682383179665, + 0.005881946068257093, + 0.005560198798775673, + 0.004324433859437704, + 0.019376469776034355, + 0.013127836398780346, + 0.007131009828299284, + 0.0044949846342206, + 0.01049823872745037, + 0.009002645500004292, + 0.02165571227669716, + 0.009978171437978745, + 0.012248977087438107, + 0.014019351452589035, + 0.014208190143108368, + 0.01658308506011963, + 0.007741091772913933, + 0.009536633267998695, + 0.0068678781390190125, + 0.009212656877934933, + 0.013804679736495018, + 0.010914105921983719, + 0.014950518496334553, + 0.012698103673756123, + 0.0031065410003066063, + 0.003013120498508215, + 0.003659287001937628, + 0.004938129801303148, + 0.0035972697660326958, + 0.004586684517562389, + 0.015723004937171936, + 0.010506993159651756, + 0.017940836027264595, + 0.00835089199244976, + 0.014452310279011726, + 0.02289440855383873, + 0.007092691026628017, + 0.0064237080514431, + 0.007327345199882984, + 0.01035842765122652, + 0.004225994925945997, + 0.007955905981361866, + 0.008310179226100445, + 0.007373373955488205, + 0.008032236248254776, + 0.004317355342209339, + 0.006679972168058157, + 0.00641607353463769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.4338047504425049 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706150114536285 + ], + "opposite_logprob_tokens": [ + -2.127061605453491 + ], + "logprob_reward": -0.12706150114536285, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706150114536285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19947214424610138, + 0.12260182946920395, + 0.06143699213862419, + 0.08293667435646057, + 0.14963924884796143, + 0.1037498340010643, + 0.15063349902629852, + 0.03858904168009758, + 0.06467550247907639, + 0.026265284046530724 + ], + "credit_map": [ + 0.015796277672052383, + 0.024274606257677078, + 0.019399801269173622, + 0.03163215145468712, + 0.02777758613228798, + 0.02915593609213829, + 0.029704783111810684, + 0.02173098362982273, + 0.01119677722454071, + 0.020682165399193764, + 0.014699986204504967, + 0.02202090434730053, + 0.012201654724776745, + 0.01122069451957941, + 0.0188123881816864, + 0.011767253279685974, + 0.006156351882964373, + 0.007897431030869484, + 0.008921410888433456, + 0.012842374853789806, + 0.007331979461014271, + 0.0055784680880606174, + 0.007168976124376059, + 0.005539996083825827, + 0.008199550211429596, + 0.009966854006052017, + 0.008285640738904476, + 0.024122515693306923, + 0.006230126600712538, + 0.009403280913829803, + 0.01123736146837473, + 0.005491338204592466, + 0.008309726603329182, + 0.014738783240318298, + 0.00842857826501131, + 0.012442514300346375, + 0.050467703491449356, + 0.03338930383324623, + 0.013905513100326061, + 0.007957113906741142, + 0.012654688209295273, + 0.007358089555054903, + 0.025692695751786232, + 0.009774035774171352, + 0.00884674210101366, + 0.00953587144613266, + 0.011724557727575302, + 0.018163148313760757, + 0.014000067487359047, + 0.014723096042871475, + 0.01494539063423872, + 0.016033198684453964, + 0.020235538482666016, + 0.023119580000638962, + 0.02755890041589737, + 0.02001771330833435, + 0.00408954918384552, + 0.004035499412566423, + 0.004458207171410322, + 0.004957170225679874, + 0.006005455274134874, + 0.004560262896120548, + 0.006408672779798508, + 0.004074221942573786, + 0.013349242508411407, + 0.006445002742111683, + 0.014328750781714916, + 0.0074676931835711, + 0.005348229315131903, + 0.00583847239613533, + 0.005776921287178993, + 0.006121179088950157, + 0.0021332972683012486, + 0.003680616617202759, + 0.0035717817954719067, + 0.004715827293694019, + 0.00387831823900342, + 0.002262724097818136, + 0.0031677179504185915, + 0.0028549986891448498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.12706150114536285 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000870257616043 + ], + "opposite_logprob_tokens": [ + -2.6950087547302246 + ], + "logprob_reward": -0.07000870257616043, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000870257616043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3211725354194641, + 0.11348352581262589, + 0.08802381157875061, + 0.037742432206869125, + 0.02207300066947937, + 0.11160571128129959, + 0.044639743864536285, + 0.14844441413879395, + 0.08426225185394287, + 0.028552593663334846 + ], + "credit_map": [ + 0.018359314650297165, + 0.031943559646606445, + 0.024266693741083145, + 0.04259184002876282, + 0.03394492343068123, + 0.042502615600824356, + 0.04605190455913544, + 0.08151169121265411, + 0.010214756242930889, + 0.014795705676078796, + 0.011384234763681889, + 0.019886231049895287, + 0.01393190212547779, + 0.01304686814546585, + 0.020461466163396835, + 0.009762375615537167, + 0.008951236493885517, + 0.010153333656489849, + 0.009818346239626408, + 0.016646984964609146, + 0.008066458627581596, + 0.012163695879280567, + 0.01482577994465828, + 0.00739798229187727, + 0.0031134868040680885, + 0.004763416014611721, + 0.0035112977493554354, + 0.006285598501563072, + 0.003870963817462325, + 0.0061819651164114475, + 0.006086710374802351, + 0.003928995691239834, + 0.0021426151506602764, + 0.002381068654358387, + 0.002063704887405038, + 0.0018435112433508039, + 0.0054482389241456985, + 0.003527200547978282, + 0.0027653465513139963, + 0.0019013143610209227, + 0.01202321145683527, + 0.004902624525129795, + 0.026036109775304794, + 0.008431817404925823, + 0.007628901395946741, + 0.011392881162464619, + 0.010669391602277756, + 0.030520781874656677, + 0.004063695669174194, + 0.004056669306010008, + 0.005249578971415758, + 0.004669165704399347, + 0.005881119519472122, + 0.00911159347742796, + 0.005909057799726725, + 0.005698866676539183, + 0.009306905791163445, + 0.008451767265796661, + 0.0104304114356637, + 0.014276680536568165, + 0.008703172206878662, + 0.009511079639196396, + 0.04835688695311546, + 0.039407502859830856, + 0.011319701559841633, + 0.008972025476396084, + 0.007146503776311874, + 0.029261372983455658, + 0.006171074230223894, + 0.004966876935213804, + 0.007723972667008638, + 0.008700721897184849, + 0.002593207871541381, + 0.00500784395262599, + 0.003429040079936385, + 0.0038732795510441065, + 0.0034395395778119564, + 0.003389426041394472, + 0.002860521897673607, + 0.003959733061492443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.07000870257616043 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867272078990936 + ], + "opposite_logprob_tokens": [ + -3.048672676086426 + ], + "logprob_reward": -0.04867272078990936, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867272078990936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2685340940952301, + 0.14632630348205566, + 0.14699460566043854, + 0.03614065796136856, + 0.05127852037549019, + 0.17336511611938477, + 0.02921450510621071, + 0.04678972437977791, + 0.07819296419620514, + 0.023163463920354843 + ], + "credit_map": [ + 0.03518535941839218, + 0.029643608257174492, + 0.049095526337623596, + 0.038477323949337006, + 0.028533725067973137, + 0.029695475473999977, + 0.031395476311445236, + 0.026507621631026268, + 0.019052838906645775, + 0.02331416681408882, + 0.015033007599413395, + 0.02179574966430664, + 0.017627151682972908, + 0.014365999028086662, + 0.021903514862060547, + 0.013233889825642109, + 0.015498253516852856, + 0.019775666296482086, + 0.020770004019141197, + 0.02423170953989029, + 0.022702505812048912, + 0.01705227419734001, + 0.016000965610146523, + 0.010963240638375282, + 0.0032668504863977432, + 0.0044331238605082035, + 0.0038711470551788807, + 0.0049033272080123425, + 0.0036225037183612585, + 0.005499863065779209, + 0.006697997450828552, + 0.0038458516355603933, + 0.004770623054355383, + 0.004045622423291206, + 0.004170515108853579, + 0.0027233604341745377, + 0.017892200499773026, + 0.011489872820675373, + 0.003952417988330126, + 0.002233915263786912, + 0.013903503306210041, + 0.006994650233536959, + 0.02806706167757511, + 0.008957353420555592, + 0.008625647984445095, + 0.036059655249118805, + 0.014932462014257908, + 0.055824801325798035, + 0.002525292569771409, + 0.0030145063064992428, + 0.0037804474122822285, + 0.0027652112767100334, + 0.004378806799650192, + 0.005807419773191214, + 0.0035661624278873205, + 0.0033766599372029305, + 0.0033297922927886248, + 0.0028344127349555492, + 0.0038677069824188948, + 0.005568843800574541, + 0.0029030428268015385, + 0.003086868906393647, + 0.015095154754817486, + 0.010103905573487282, + 0.008599569089710712, + 0.008488527499139309, + 0.010585920885205269, + 0.020765254274010658, + 0.006746005266904831, + 0.007664916571229696, + 0.007870436646044254, + 0.007472342345863581, + 0.0020385999232530594, + 0.0036235679872334003, + 0.0034997877664864063, + 0.004215188790112734, + 0.0025003075134009123, + 0.0023942305706441402, + 0.0023652270901948214, + 0.002526557072997093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.04867272078990936 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288888856768608 + ], + "opposite_logprob_tokens": [ + -3.7732889652252197 + ], + "logprob_reward": -0.023288888856768608, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288888856768608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2972441017627716, + 0.13452957570552826, + 0.059325896203517914, + 0.04946720227599144, + 0.054943669587373734, + 0.14635059237480164, + 0.06982117146253586, + 0.05407847464084625, + 0.10431136190891266, + 0.029927916824817657 + ], + "credit_map": [ + 0.023793073371052742, + 0.025803713127970695, + 0.03016694076359272, + 0.03965266793966293, + 0.03677399829030037, + 0.04527168348431587, + 0.0329858735203743, + 0.06279614567756653, + 0.02024885267019272, + 0.020233167335391045, + 0.014043362811207771, + 0.01940024457871914, + 0.01438278891146183, + 0.011382213793694973, + 0.022664299234747887, + 0.012174669653177261, + 0.006548633798956871, + 0.007052900269627571, + 0.007684188429266214, + 0.010744037106633186, + 0.005750135984271765, + 0.009072241373360157, + 0.007751432713121176, + 0.004722331650555134, + 0.0048533170484006405, + 0.004661559592932463, + 0.005535248667001724, + 0.011198963969945908, + 0.004413692280650139, + 0.006578975357115269, + 0.007357392925769091, + 0.00486805709078908, + 0.005909251980483532, + 0.007478966843336821, + 0.003841632278636098, + 0.0038836474996060133, + 0.014906961470842361, + 0.01033039577305317, + 0.004697439726442099, + 0.0038953758776187897, + 0.01994543895125389, + 0.006979793310165405, + 0.04429416358470917, + 0.010449900291860104, + 0.01061304286122322, + 0.01458306610584259, + 0.013316922821104527, + 0.026168281212449074, + 0.005243714898824692, + 0.006672291550785303, + 0.0074801016598939896, + 0.00756375165656209, + 0.010989341884851456, + 0.01198071613907814, + 0.009338795207440853, + 0.010552464053034782, + 0.00688597746193409, + 0.004602270666509867, + 0.007963495329022408, + 0.006200628820806742, + 0.005014293361455202, + 0.0056959232315421104, + 0.011345073580741882, + 0.0063708145171403885, + 0.017075518146157265, + 0.010302763432264328, + 0.01093145553022623, + 0.034098993986845016, + 0.006946416571736336, + 0.005906050093472004, + 0.009313151240348816, + 0.009737018495798111, + 0.00270716636441648, + 0.005045678000897169, + 0.003862307406961918, + 0.005500174593180418, + 0.003876697737723589, + 0.0028129585552960634, + 0.0029467211570590734, + 0.0031762158032506704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.023288888856768608 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic low-angle wide shot slowly tilts upward, advancing almost imperceptibly, to capture the terrifying majesty of the vast, dark vortex of storm clouds dominating the sky. This roiling vortex, composed of dense, roiling cumulonimbus clouds in shades of inky black, bruised purple, and menacing slate gray, spirals with a violent, relentless energy. Within this surging, swirling chaos, a constant stream of shattered debris pours forth; broken wooden beams, shards of twisted structural metal, and fragments of mangled infrastructure are flung outward by gravity, only to be sucked back into the vortex's dark center. The only illumination in this apocalyptic scene comes from the storm's own eerie internal glow and sporadic, brilliant flashes of lightning, which momentarily outline a violent ballet of destruction. The entire image is suffused with an icy, desaturated palette, creating an intense and overwhelming atmosphere of raw, untamed power and absolute devastation.", + "seed": 3306, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_6/index_5281_seed_3306_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07935323063284158, + "vlm_logprob": -0.07935323063284158, + "vlm_yes_no_margin": 2.799999976158142, + "vlm_token_credit": -0.07935323063284158, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06562583893537521 + ], + "opposite_logprob_tokens": [ + -2.8156259059906006 + ], + "logprob_reward": -0.06562583893537521, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06562583893537521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11281097680330276, + 0.10989836603403091, + 0.03129582107067108, + 0.05587597191333771, + 0.18898454308509827, + 0.07213415205478668, + 0.1463981717824936, + 0.15550686419010162, + 0.10672251880168915, + 0.020372623577713966 + ], + "credit_map": [ + 0.014528247527778149, + 0.014796199277043343, + 0.013177486136555672, + 0.011702319607138634, + 0.02028474770486355, + 0.017686687409877777, + 0.013017358258366585, + 0.00761793740093708, + 0.011426317505538464, + 0.016761817038059235, + 0.011898963712155819, + 0.00997899379581213, + 0.014573815278708935, + 0.019098732620477676, + 0.012825975194573402, + 0.013333742506802082, + 0.00236603245139122, + 0.0039012532215565443, + 0.00326906843110919, + 0.0039705801755189896, + 0.0029438170604407787, + 0.004475791472941637, + 0.00655386783182621, + 0.003815408330410719, + 0.004930172115564346, + 0.0048909373581409454, + 0.012086109258234501, + 0.0038695302791893482, + 0.013689137063920498, + 0.005741404835134745, + 0.004706775303930044, + 0.00596191082149744, + 0.008193129673600197, + 0.008294748142361641, + 0.006354017183184624, + 0.0657477155327797, + 0.01307915523648262, + 0.06282227486371994, + 0.016224827617406845, + 0.008268681354820728, + 0.005601953715085983, + 0.008122673258185387, + 0.009670650586485863, + 0.008316881023347378, + 0.01489347405731678, + 0.010110069066286087, + 0.0071843150071799755, + 0.008234133012592793, + 0.00771342683583498, + 0.011225907132029533, + 0.03831571713089943, + 0.009378556162118912, + 0.0164285060018301, + 0.006428881082683802, + 0.050378549844026566, + 0.006528628524392843, + 0.016257155686616898, + 0.019156156107783318, + 0.029058774933218956, + 0.01973160356283188, + 0.013600176200270653, + 0.008648710325360298, + 0.037396397441625595, + 0.011657889001071453, + 0.01055156346410513, + 0.010878435336053371, + 0.018126506358385086, + 0.011020069010555744, + 0.014310035854578018, + 0.006902913562953472, + 0.027166593819856644, + 0.007766392081975937, + 0.001462191459722817, + 0.0022679504472762346, + 0.0030360943637788296, + 0.004048455972224474, + 0.0024282645899802446, + 0.0018386864103376865, + 0.0028711259365081787, + 0.002419854747131467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.06562583893537521 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283177256584167 + ], + "opposite_logprob_tokens": [ + -2.017831802368164 + ], + "logprob_reward": -0.14283177256584167, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283177256584167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12205039709806442, + 0.08692315965890884, + 0.11302149295806885, + 0.11190945655107498, + 0.1698491871356964, + 0.04286462068557739, + 0.11402653902769089, + 0.0725904330611229, + 0.1283474713563919, + 0.03841725364327431 + ], + "credit_map": [ + 0.012333902530372143, + 0.015267163515090942, + 0.01254204660654068, + 0.014547846280038357, + 0.02081778459250927, + 0.015271891839802265, + 0.017385557293891907, + 0.013884204439818859, + 0.00853356346487999, + 0.012128469534218311, + 0.009161597117781639, + 0.009936421178281307, + 0.010108261369168758, + 0.012419058941304684, + 0.007664392236620188, + 0.016971398144960403, + 0.009020357392728329, + 0.014580023474991322, + 0.01100070308893919, + 0.01308949664235115, + 0.010577156208455563, + 0.016598982736468315, + 0.02395176887512207, + 0.014203004539012909, + 0.006125194486230612, + 0.010374795645475388, + 0.020549986511468887, + 0.006260657217353582, + 0.0317002534866333, + 0.011252960190176964, + 0.014808740466833115, + 0.010836859233677387, + 0.00711954478174448, + 0.008245833218097687, + 0.008587864227592945, + 0.055921606719493866, + 0.011522090062499046, + 0.04919508472084999, + 0.016893288120627403, + 0.012363877147436142, + 0.005405446048825979, + 0.003839117707684636, + 0.0040124207735061646, + 0.005145922768861055, + 0.009591350331902504, + 0.006830441299825907, + 0.0044727870263159275, + 0.003567133331671357, + 0.004733521491289139, + 0.007760948501527309, + 0.03245233744382858, + 0.006013007834553719, + 0.009739568457007408, + 0.004112736787647009, + 0.044248100370168686, + 0.004966321866959333, + 0.007274032570421696, + 0.010549255646765232, + 0.011178639717400074, + 0.010444201529026031, + 0.008265220560133457, + 0.005141724366694689, + 0.01509654801338911, + 0.004640811122953892, + 0.008638097904622555, + 0.014049293473362923, + 0.02102935127913952, + 0.008048559539020061, + 0.015610911883413792, + 0.006629133131355047, + 0.04906594753265381, + 0.005276167299598455, + 0.002976439194753766, + 0.004171739798039198, + 0.005361691117286682, + 0.007295090239495039, + 0.004715778399258852, + 0.00355660542845726, + 0.005446289665997028, + 0.004893617704510689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.14283177256584167 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537707574665546 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537707574665546, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537707574665546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08016155660152435, + 0.10727223008871078, + 0.05581318959593773, + 0.1771918684244156, + 0.12308882921934128, + 0.06827309727668762, + 0.039261940866708755, + 0.0768471285700798, + 0.1877991259098053, + 0.0842909887433052 + ], + "credit_map": [ + 0.00937266182154417, + 0.010457564145326614, + 0.007920422591269016, + 0.009474260732531548, + 0.014810483902692795, + 0.008244573138654232, + 0.01192601677030325, + 0.007955576293170452, + 0.009523358196020126, + 0.017802931368350983, + 0.013258772902190685, + 0.010459395125508308, + 0.014805277809500694, + 0.018108759075403214, + 0.010292024351656437, + 0.013021714054048061, + 0.00477635907009244, + 0.007022908888757229, + 0.00568210706114769, + 0.006521481089293957, + 0.005040308460593224, + 0.008978994563221931, + 0.010370999574661255, + 0.007420034147799015, + 0.008675194345414639, + 0.015797730535268784, + 0.026944557204842567, + 0.011448902077972889, + 0.05098511651158333, + 0.019449137151241302, + 0.030785227194428444, + 0.013106009922921658, + 0.005261851940304041, + 0.005769211798906326, + 0.003680508118122816, + 0.04033716022968292, + 0.012566338293254375, + 0.04185880720615387, + 0.009115939028561115, + 0.004499019123613834, + 0.010495131835341454, + 0.007390801794826984, + 0.009118050336837769, + 0.005631046835333109, + 0.01272515393793583, + 0.009403092786669731, + 0.0060920375399291515, + 0.007417787332087755, + 0.002679832512512803, + 0.004238876514136791, + 0.006859166081994772, + 0.0037929490208625793, + 0.006410820409655571, + 0.002361319027841091, + 0.009024880826473236, + 0.003894095541909337, + 0.008886509574949741, + 0.01114029623568058, + 0.011663205921649933, + 0.011235242709517479, + 0.0066878157667815685, + 0.0043615782633423805, + 0.016578787937760353, + 0.006293690297752619, + 0.01446810457855463, + 0.019619423896074295, + 0.03168897703289986, + 0.01494965236634016, + 0.02064780704677105, + 0.00925787165760994, + 0.06873450428247452, + 0.008432789705693722, + 0.007076168432831764, + 0.010323094204068184, + 0.01252225786447525, + 0.01852930150926113, + 0.009746184572577477, + 0.00635391054674983, + 0.011315214447677135, + 0.008424856700003147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 4 + ], + "selected_reward": -0.012537707574665546 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270969808101654 + ], + "opposite_logprob_tokens": [ + -2.1270968914031982 + ], + "logprob_reward": -0.1270969808101654, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270969808101654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44119858741760254, + 0.12850356101989746, + 0.06546299904584885, + 0.09570562094449997, + 0.09892500191926956, + 0.023133534938097, + 0.030071772634983063, + 0.047089967876672745, + 0.03798607364296913, + 0.03192291036248207 + ], + "credit_map": [ + 0.05012516304850578, + 0.053154777735471725, + 0.05441546440124512, + 0.05290064960718155, + 0.08209668844938278, + 0.046244215220212936, + 0.05730405077338219, + 0.04495757818222046, + 0.013064991682767868, + 0.0171684417873621, + 0.018450234085321426, + 0.013339469209313393, + 0.014258226379752159, + 0.020031984895467758, + 0.011055424809455872, + 0.02113479934632778, + 0.005985196679830551, + 0.008794073946774006, + 0.007906323298811913, + 0.0082693537697196, + 0.006378881633281708, + 0.009865405969321728, + 0.010054066777229309, + 0.008209692314267159, + 0.00428340258076787, + 0.009496216662228107, + 0.021443283185362816, + 0.004615416284650564, + 0.033710017800331116, + 0.007270199246704578, + 0.008513270877301693, + 0.006373809650540352, + 0.004382355138659477, + 0.005372784100472927, + 0.004323615692555904, + 0.03115798532962799, + 0.008169091306626797, + 0.032259751111269, + 0.00803552195429802, + 0.0052238954231143, + 0.002236021449789405, + 0.0028166556730866432, + 0.003327432321384549, + 0.0029081732500344515, + 0.004650278948247433, + 0.002928408794105053, + 0.0022224239073693752, + 0.002044140128418803, + 0.001779158366844058, + 0.0032699047587811947, + 0.0040836394764482975, + 0.0031903430353850126, + 0.005550264846533537, + 0.002363280626013875, + 0.007164651062339544, + 0.0026705311611294746, + 0.0037621022202074528, + 0.004050116520375013, + 0.016071707010269165, + 0.00351968826726079, + 0.003007762134075165, + 0.002922062063589692, + 0.010423213243484497, + 0.0033333157189190388, + 0.003961979411542416, + 0.005426784045994282, + 0.0051054032519459724, + 0.006295567844063044, + 0.005500849336385727, + 0.00336001580581069, + 0.004751577042043209, + 0.0035838987678289413, + 0.0028303181752562523, + 0.003538096323609352, + 0.006049866322427988, + 0.0051915994845330715, + 0.003477572463452816, + 0.0029058363288640976, + 0.003975881263613701, + 0.0039537386037409306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.1270969808101654 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048673853278160095 + ], + "opposite_logprob_tokens": [ + -3.0486738681793213 + ], + "logprob_reward": -0.048673853278160095, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048673853278160095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1302272528409958, + 0.10722913593053818, + 0.056861307471990585, + 0.09338368475437164, + 0.1755312979221344, + 0.05561523139476776, + 0.05210450291633606, + 0.19630956649780273, + 0.08139055967330933, + 0.05134746432304382 + ], + "credit_map": [ + 0.014218131080269814, + 0.01963871903717518, + 0.012720863334834576, + 0.012987613677978516, + 0.027786381542682648, + 0.013314603827893734, + 0.01702778786420822, + 0.012533151544630527, + 0.010185128077864647, + 0.017361078411340714, + 0.009987590834498405, + 0.01592831499874592, + 0.01481312233954668, + 0.015115301124751568, + 0.007933070883154869, + 0.015905527397990227, + 0.004868439864367247, + 0.00842952262610197, + 0.00701387133449316, + 0.007227882277220488, + 0.005956774577498436, + 0.008786382153630257, + 0.00862895231693983, + 0.005949479527771473, + 0.006088638212531805, + 0.007877436466515064, + 0.02047167718410492, + 0.006640917621552944, + 0.0234312005341053, + 0.010300188325345516, + 0.010523738339543343, + 0.008049891330301762, + 0.007423462811857462, + 0.007648000959306955, + 0.008780877105891705, + 0.0557204969227314, + 0.01268251147121191, + 0.0548880361020565, + 0.01755478046834469, + 0.010833138599991798, + 0.008276190608739853, + 0.005301990546286106, + 0.006790925748646259, + 0.005484968889504671, + 0.007500008679926395, + 0.0089147062972188, + 0.00499348109588027, + 0.008352954871952534, + 0.0033167393412441015, + 0.006618375424295664, + 0.005752348806709051, + 0.006305057089775801, + 0.010521701537072659, + 0.004512177314609289, + 0.009866191074252129, + 0.0052119167521595955, + 0.018103308975696564, + 0.020809991285204887, + 0.04799584299325943, + 0.021100806072354317, + 0.01330194529145956, + 0.009917636401951313, + 0.05215895175933838, + 0.012921079993247986, + 0.009421058930456638, + 0.009960060939192772, + 0.00978615041822195, + 0.010955634526908398, + 0.011581728234887123, + 0.008552944287657738, + 0.013224565424025059, + 0.00790841318666935, + 0.0035765867214649916, + 0.006079133599996567, + 0.007521460764110088, + 0.010993936099112034, + 0.005411142949014902, + 0.0046454910188913345, + 0.0069205304607748985, + 0.0061991834081709385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.048673853278160095 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic low-angle wide shot slowly tilts upward, advancing almost imperceptibly, to capture the terrifying majesty of the vast, dark vortex of storm clouds dominating the sky. This roiling vortex, composed of dense, roiling cumulonimbus clouds in shades of inky black, bruised purple, and menacing slate gray, spirals with a violent, relentless energy. Within this surging, swirling chaos, a constant stream of shattered debris pours forth; broken wooden beams, shards of twisted structural metal, and fragments of mangled infrastructure are flung outward by gravity, only to be sucked back into the vortex's dark center. The only illumination in this apocalyptic scene comes from the storm's own eerie internal glow and sporadic, brilliant flashes of lightning, which momentarily outline a violent ballet of destruction. The entire image is suffused with an icy, desaturated palette, creating an intense and overwhelming atmosphere of raw, untamed power and absolute devastation.", + "seed": 3307, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_6/index_5281_seed_3307_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12914821095764636, + "vlm_logprob": -0.12914821095764636, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.12914821095764636, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31821078062057495 + ], + "opposite_logprob_tokens": [ + -1.3182108402252197 + ], + "logprob_reward": -0.31821078062057495, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31821078062057495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0866648480296135, + 0.06551934033632278, + 0.07743002474308014, + 0.12408456951379776, + 0.1255076676607132, + 0.25233277678489685, + 0.0596226267516613, + 0.09607891738414764, + 0.05900043621659279, + 0.05375881493091583 + ], + "credit_map": [ + 0.01022313628345728, + 0.01342757511883974, + 0.011410416103899479, + 0.00807731319218874, + 0.011201079934835434, + 0.013052857480943203, + 0.007687008008360863, + 0.011585448868572712, + 0.006348493043333292, + 0.009175112470984459, + 0.007830110378563404, + 0.010099368169903755, + 0.01246680412441492, + 0.006266985088586807, + 0.006788302678614855, + 0.00654415600001812, + 0.007059164810925722, + 0.010777357034385204, + 0.012052983045578003, + 0.017162922769784927, + 0.009833688847720623, + 0.005551603157073259, + 0.009644739329814911, + 0.0053475541062653065, + 0.012305247597396374, + 0.014836572110652924, + 0.014288258738815784, + 0.024548152461647987, + 0.01673087663948536, + 0.007908736355602741, + 0.01842544600367546, + 0.01504126749932766, + 0.010176580399274826, + 0.018923990428447723, + 0.013813160359859467, + 0.014720895327627659, + 0.01882273703813553, + 0.01563922129571438, + 0.015234911814332008, + 0.018176162615418434, + 0.020097119733691216, + 0.032800428569316864, + 0.029698019847273827, + 0.051242854446172714, + 0.02593356929719448, + 0.022277813404798508, + 0.04427889734506607, + 0.026004033163189888, + 0.004464622121304274, + 0.009928199462592602, + 0.008466942235827446, + 0.00799938477575779, + 0.008658839389681816, + 0.007194813806563616, + 0.00838520284742117, + 0.004524612333625555, + 0.006933011580258608, + 0.011981853283941746, + 0.019199077039957047, + 0.008773802779614925, + 0.005591697059571743, + 0.007515555713325739, + 0.02261696383357048, + 0.01346694864332676, + 0.0041314782574772835, + 0.009288080967962742, + 0.00591170834377408, + 0.012088438495993614, + 0.006509398575872183, + 0.005285336636006832, + 0.009976226836442947, + 0.0058097620494663715, + 0.0035495769698172808, + 0.006974135525524616, + 0.004047183319926262, + 0.011896544136106968, + 0.005661570467054844, + 0.00510666286572814, + 0.009550747461616993, + 0.0069723897613584995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.31821078062057495 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603948175907135 + ], + "opposite_logprob_tokens": [ + -1.9103947877883911 + ], + "logprob_reward": -0.1603948175907135, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603948175907135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09264542162418365, + 0.06020885333418846, + 0.16632255911827087, + 0.044270168989896774, + 0.07566550374031067, + 0.1863861083984375, + 0.05898967757821083, + 0.11475322395563126, + 0.12528173625469208, + 0.0754767507314682 + ], + "credit_map": [ + 0.012323095463216305, + 0.013591054826974869, + 0.012367425486445427, + 0.011828086338937283, + 0.01310813520103693, + 0.010709253139793873, + 0.007665242999792099, + 0.011053123511373997, + 0.005435733124613762, + 0.00839594379067421, + 0.008886410854756832, + 0.009843483567237854, + 0.009466839022934437, + 0.006532436702400446, + 0.005591188557446003, + 0.0060568214394152164, + 0.015107415616512299, + 0.027546869590878487, + 0.021513042971491814, + 0.031348634511232376, + 0.026812277734279633, + 0.012062730267643929, + 0.019501548260450363, + 0.012430030852556229, + 0.007219448685646057, + 0.005311162211000919, + 0.003987699281424284, + 0.006138062570244074, + 0.005758565850555897, + 0.004087110515683889, + 0.005148557014763355, + 0.006619562860578299, + 0.007035106420516968, + 0.011763114482164383, + 0.011066232807934284, + 0.008945790119469166, + 0.008124299347400665, + 0.00960329920053482, + 0.007537617813795805, + 0.011590049602091312, + 0.01528686098754406, + 0.023863228037953377, + 0.022652730345726013, + 0.03682728856801987, + 0.01931196264922619, + 0.01691119186580181, + 0.032089829444885254, + 0.01944301277399063, + 0.004777396563440561, + 0.01247254479676485, + 0.006610942538827658, + 0.007158040069043636, + 0.007901431992650032, + 0.006750984117388725, + 0.008252105675637722, + 0.005066232290118933, + 0.009039320982992649, + 0.017940854653716087, + 0.027885902673006058, + 0.011058616451919079, + 0.006983916740864515, + 0.009515002369880676, + 0.019662849605083466, + 0.012666760943830013, + 0.012139633297920227, + 0.013341889716684818, + 0.026347972452640533, + 0.01409616693854332, + 0.011902577243745327, + 0.016824988648295403, + 0.018658306449651718, + 0.011970204301178455, + 0.007540515158325434, + 0.009751551784574986, + 0.005311108194291592, + 0.01882920227944851, + 0.007674762979149818, + 0.00559245003387332, + 0.01232952531427145, + 0.008447637781500816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 8 + ], + "selected_reward": -0.1603948175907135 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809110447764397 + ], + "opposite_logprob_tokens": [ + -3.2880911827087402 + ], + "logprob_reward": -0.03809110447764397, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809110447764397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1105472519993782, + 0.05287759378552437, + 0.08415622264146805, + 0.5331175327301025, + 0.02597971446812153, + 0.032022252678871155, + 0.022021500393748283, + 0.04343590512871742, + 0.07424342632293701, + 0.02159864641726017 + ], + "credit_map": [ + 0.01223325077444315, + 0.01572200283408165, + 0.012889798730611801, + 0.009476478211581707, + 0.022740453481674194, + 0.015917088836431503, + 0.007689008489251137, + 0.01387916598469019, + 0.005230624694377184, + 0.007815863937139511, + 0.0054484764114022255, + 0.00743850227445364, + 0.010657856240868568, + 0.0053366394713521, + 0.005032989662140608, + 0.0059166401624679565, + 0.006906760390847921, + 0.01506597176194191, + 0.008710328489542007, + 0.012705972418189049, + 0.012991352938115597, + 0.007691541220992804, + 0.012754068709909916, + 0.007330228108912706, + 0.058176711201667786, + 0.07312671095132828, + 0.0536230094730854, + 0.09157329052686691, + 0.07647647708654404, + 0.037009648978710175, + 0.07386795431375504, + 0.06926371902227402, + 0.0021494219545274973, + 0.005870537366718054, + 0.0019732732325792313, + 0.0029074116609990597, + 0.004852442070841789, + 0.0034847836941480637, + 0.0028815276455134153, + 0.0018603164935484529, + 0.002982059959322214, + 0.0047325934283435345, + 0.005540566518902779, + 0.004574016202241182, + 0.0033130396623164415, + 0.0029137060046195984, + 0.0038825946394354105, + 0.004083674401044846, + 0.0016462153289467096, + 0.005655836313962936, + 0.002861693035811186, + 0.002653812523931265, + 0.0025184759870171547, + 0.0017116520320996642, + 0.003291078144684434, + 0.0016827377257868648, + 0.00313575379550457, + 0.005143992602825165, + 0.007079692091792822, + 0.005466204136610031, + 0.0029741316102445126, + 0.00363871525041759, + 0.009434036910533905, + 0.0065633817575871944, + 0.005993375554680824, + 0.006567981094121933, + 0.007227171678096056, + 0.0063432310707867146, + 0.005170383490622044, + 0.006797562353312969, + 0.02152201719582081, + 0.014621706679463387, + 0.0014418397331610322, + 0.0019157805945724249, + 0.001595006207935512, + 0.0037239540833979845, + 0.001578612718731165, + 0.0032593351788818836, + 0.0023163941223174334, + 0.00576772540807724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.03809110447764397 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297504603862762 + ], + "opposite_logprob_tokens": [ + -2.2379751205444336 + ], + "logprob_reward": -0.11297504603862762, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297504603862762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1802678257226944, + 0.08264601975679398, + 0.10053523629903793, + 0.062479086220264435, + 0.15200850367546082, + 0.19729100167751312, + 0.04960574209690094, + 0.09359139949083328, + 0.05019991472363472, + 0.031375300139188766 + ], + "credit_map": [ + 0.01902804523706436, + 0.029799681156873703, + 0.018125614151358604, + 0.024522028863430023, + 0.030556952580809593, + 0.019867856055498123, + 0.012791985645890236, + 0.02557566575706005, + 0.007580151315778494, + 0.011399962939321995, + 0.00987655483186245, + 0.013225994072854519, + 0.011910337023437023, + 0.007610207423567772, + 0.007929474115371704, + 0.013113336637616158, + 0.010859223082661629, + 0.013335681520402431, + 0.010817011818289757, + 0.017852168530225754, + 0.015459825284779072, + 0.010431245900690556, + 0.009956484660506248, + 0.011823594570159912, + 0.007948221638798714, + 0.008713522925972939, + 0.007216765079647303, + 0.009538969025015831, + 0.008129537105560303, + 0.006744431797415018, + 0.00894143059849739, + 0.00524620758369565, + 0.013448277488350868, + 0.02172655053436756, + 0.025566615164279938, + 0.01662137731909752, + 0.02253117226064205, + 0.020739052444696426, + 0.01645364612340927, + 0.014921807684004307, + 0.015356033109128475, + 0.02628633938729763, + 0.023546408861875534, + 0.040431227535009384, + 0.020678319036960602, + 0.016218509525060654, + 0.03424880653619766, + 0.02052537351846695, + 0.0038196563255041838, + 0.00880792923271656, + 0.005658208858221769, + 0.0056381747126579285, + 0.007936090230941772, + 0.006502265576273203, + 0.007540803402662277, + 0.00370261468924582, + 0.004160197451710701, + 0.007903281599283218, + 0.006774463690817356, + 0.006465354468673468, + 0.004089163616299629, + 0.004653234034776688, + 0.03592979907989502, + 0.02361590787768364, + 0.0030825501307845116, + 0.003953995183110237, + 0.004428133834153414, + 0.0036336786579340696, + 0.003225837368518114, + 0.0029859053902328014, + 0.016769690439105034, + 0.012120124883949757, + 0.002469802275300026, + 0.0035075682681053877, + 0.0032942271791398525, + 0.00416553532704711, + 0.0026329418178647757, + 0.005262089427560568, + 0.004234575666487217, + 0.005808560643345118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.11297504603862762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera advance almost imperceptibly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene dominated by a vast, dark vortex of storm clouds in shades of inky black, bruised purple, and slate gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are broken wooden beams, twisted metal shards, and fragments of infrastructure visible being flung outward and sucked back into the vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated only by the storm's internal glow and sporadic flashes of lightning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069306060671806 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.016069306060671806, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069306060671806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2546995282173157, + 0.05891696363687515, + 0.0942794680595398, + 0.2710488438606262, + 0.04694340378046036, + 0.025047671049833298, + 0.026718156412243843, + 0.09690769016742706, + 0.04491870105266571, + 0.08051960915327072 + ], + "credit_map": [ + 0.02651498280465603, + 0.0366785004734993, + 0.04007358476519585, + 0.031003592535853386, + 0.038047656416893005, + 0.03643658757209778, + 0.01865539140999317, + 0.027289217337965965, + 0.005010014399886131, + 0.007660556584596634, + 0.010185036808252335, + 0.009581374935805798, + 0.00633497117087245, + 0.00516548752784729, + 0.008622593246400356, + 0.0063569326885044575, + 0.009153701364994049, + 0.01726018264889717, + 0.009871358051896095, + 0.012236976996064186, + 0.017781589180231094, + 0.011423223651945591, + 0.010426109656691551, + 0.006126329302787781, + 0.030955037102103233, + 0.0353391170501709, + 0.028218956664204597, + 0.04483415558934212, + 0.03917321190237999, + 0.019317541271448135, + 0.03550131246447563, + 0.03770952299237251, + 0.0048208702355623245, + 0.006216558162122965, + 0.007179432548582554, + 0.004657820798456669, + 0.005864704959094524, + 0.004687423352152109, + 0.006089120637625456, + 0.007427474018186331, + 0.0026781668420881033, + 0.003012431552633643, + 0.0034595259930938482, + 0.0033011676277965307, + 0.002832446713000536, + 0.002793550491333008, + 0.00356260035187006, + 0.0034077796153724194, + 0.002194131026044488, + 0.003785611828789115, + 0.004539825022220612, + 0.0037217570934444666, + 0.003536809468641877, + 0.0021665950771421194, + 0.004677366930991411, + 0.0020960604306310415, + 0.0040341648273169994, + 0.009252017363905907, + 0.0103543596342206, + 0.0075403484515845776, + 0.004606729373335838, + 0.005229913629591465, + 0.03276623412966728, + 0.023123925551772118, + 0.004718919284641743, + 0.004510982893407345, + 0.0059615555219352245, + 0.00573285948485136, + 0.00338119943626225, + 0.00412566214799881, + 0.009232322685420513, + 0.007255201227962971, + 0.007975642569363117, + 0.011760345660150051, + 0.00707352114841342, + 0.01305848266929388, + 0.00828652735799551, + 0.011421010829508305, + 0.010128977708518505, + 0.01081510167568922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.016069306060671806 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_7/index_tensor([5007, 5007])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_7/index_tensor([5007, 5007])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6281ad0587c98030297fcbfc357a8795936261fc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_7/index_tensor([5007, 5007])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the vibrant style of Ghibli animation, the camera executes continuous push-ins, gliding smoothly through the dense foliage of a lush, sun-dappled jungle to focus on the vibrant courtship ritual. Warm golden sunlight penetrates the thick canopy, creating shifting patterns of spots on oversized waxy leaves and mossy branches. The central subject is a colorful lizard, similar to Anolis ricordi, with a body of brilliant emerald green, speckled with turquoise along its spine, and large, expressive black eyes brimming with playful energy. It pauses for a moment on a large fern leaf, rhythmically expanding its prominent, translucent scarlet dewdrops, catching the light like a stained glass window. In rapid pursuit, another smaller, more golden counterpart nimbly climbs a dense, twisting vine. Their playful chase is a joyful, fluid dance, as they dart around the gnarled roots of ancient trees and gracefully leap between broad, dew-drenched leaves, their movements swift and light. The entire scene is brimming with life and color, taking on the soft, hand-painted textures and magical realism characteristic of Ghibli masterpieces, evoking a wonderfully vibrant atmosphere.", + "seed": 10633, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_7/index_5007_seed_10633_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08096190355718136, + "vlm_logprob": -0.08096190355718136, + "vlm_yes_no_margin": 2.699999976158142, + "vlm_token_credit": -0.08096190355718136, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07052852213382721 + ], + "opposite_logprob_tokens": [ + -2.820528507232666 + ], + "logprob_reward": -0.07052852213382721, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07052852213382721, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22020837664604187, + 0.08737937361001968, + 0.05310511216521263, + 0.08007225394248962, + 0.0924961119890213, + 0.13423651456832886, + 0.09417394548654556, + 0.14343798160552979, + 0.04485095292329788, + 0.05003933236002922 + ], + "credit_map": [ + 0.03433506563305855, + 0.025356872007250786, + 0.024694453924894333, + 0.022970883175730705, + 0.024110691621899605, + 0.024426499381661415, + 0.03270871564745903, + 0.03160519152879715, + 0.010304249823093414, + 0.010938882827758789, + 0.01053487416356802, + 0.01259782537817955, + 0.00885975081473589, + 0.008187226951122284, + 0.014765665866434574, + 0.011190894991159439, + 0.007634136825799942, + 0.008327762596309185, + 0.00711732916533947, + 0.00697940681129694, + 0.007594251539558172, + 0.005290233995765448, + 0.004925128072500229, + 0.005236858036369085, + 0.009746812283992767, + 0.006787772290408611, + 0.020824823528528214, + 0.005153213627636433, + 0.005682613700628281, + 0.0070382216945290565, + 0.005637258291244507, + 0.01920153945684433, + 0.009678946807980537, + 0.012203327380120754, + 0.010149242356419563, + 0.009116933681070805, + 0.012770304456353188, + 0.012116928584873676, + 0.011605390347540379, + 0.0148550346493721, + 0.01091877743601799, + 0.00715684425085783, + 0.007155534345656633, + 0.007574927061796188, + 0.008213886059820652, + 0.009800205007195473, + 0.04346533492207527, + 0.03995099663734436, + 0.009774689562618732, + 0.013601789250969887, + 0.010807830840349197, + 0.01370178535580635, + 0.015050788410007954, + 0.01058985572308302, + 0.011183534748852253, + 0.009463672526180744, + 0.012898716144263744, + 0.023165427148342133, + 0.009949821047484875, + 0.011388396844267845, + 0.01735071837902069, + 0.02861494943499565, + 0.009844415821135044, + 0.03022555261850357, + 0.005033641587942839, + 0.006642743479460478, + 0.005118920002132654, + 0.006099064368754625, + 0.0075708115473389626, + 0.0040618255734443665, + 0.006148414220660925, + 0.0041755326092243195, + 0.004345721565186977, + 0.008639681152999401, + 0.005554117262363434, + 0.007489551790058613, + 0.008815915323793888, + 0.0061830575577914715, + 0.0034328976180404425, + 0.005578390788286924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.07052852213382721 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197020515799522 + ], + "opposite_logprob_tokens": [ + -4.018197059631348 + ], + "logprob_reward": -0.018197020515799522, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197020515799522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32516252994537354, + 0.0520777590572834, + 0.022893201559782028, + 0.06339619308710098, + 0.021188220009207726, + 0.2703484892845154, + 0.12351419776678085, + 0.037716254591941833, + 0.04584614560008049, + 0.03785710409283638 + ], + "credit_map": [ + 0.03966979309916496, + 0.053219009190797806, + 0.04258401691913605, + 0.02901933155953884, + 0.0396353118121624, + 0.035839103162288666, + 0.05045873671770096, + 0.03473722189664841, + 0.005080736242234707, + 0.00568585516884923, + 0.006020833738148212, + 0.009900894947350025, + 0.006049848161637783, + 0.005313436966389418, + 0.008211569860577583, + 0.0058145830407738686, + 0.003403483424335718, + 0.004395354073494673, + 0.0017930662725120783, + 0.0032970618922263384, + 0.003738544648513198, + 0.0020376970060169697, + 0.001949175726622343, + 0.00227881851606071, + 0.0050066751427948475, + 0.00403971690684557, + 0.020224610343575478, + 0.003926444333046675, + 0.0031821997836232185, + 0.0053963325917720795, + 0.003911059815436602, + 0.01770915649831295, + 0.0034504004288464785, + 0.002191143576055765, + 0.0022637383081018925, + 0.0026286544743925333, + 0.0028587807901203632, + 0.0025656528305262327, + 0.0030415942892432213, + 0.0021882536821067333, + 0.022303462028503418, + 0.00836385227739811, + 0.010426447726786137, + 0.009683465585112572, + 0.014129516668617725, + 0.015516920946538448, + 0.10160793364048004, + 0.08831688016653061, + 0.017218369990587234, + 0.008338556624948978, + 0.014678255654871464, + 0.04463263228535652, + 0.011596026830375195, + 0.01463810633867979, + 0.00672802934423089, + 0.0056842309422791, + 0.002569794887676835, + 0.0055709052830934525, + 0.002967237029224634, + 0.0036396379582583904, + 0.005690351594239473, + 0.007490802556276321, + 0.003025305923074484, + 0.006762217730283737, + 0.005871877074241638, + 0.004841041285544634, + 0.004882193636149168, + 0.006926777772605419, + 0.0043047284707427025, + 0.004557972773909569, + 0.009091033600270748, + 0.005370523314923048, + 0.004351328127086163, + 0.00466871028766036, + 0.004512025509029627, + 0.007367564830929041, + 0.0073311044834554195, + 0.0037748576141893864, + 0.0022430536337196827, + 0.003608457278460264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.018197020515799522 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062039099633693695 + ], + "opposite_logprob_tokens": [ + -2.8120391368865967 + ], + "logprob_reward": -0.062039099633693695, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062039099633693695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13367494940757751, + 0.08101861923933029, + 0.13260415196418762, + 0.04823179543018341, + 0.03762258216738701, + 0.2229442447423935, + 0.17283201217651367, + 0.08142147958278656, + 0.03186929225921631, + 0.057780858129262924 + ], + "credit_map": [ + 0.01529987994581461, + 0.019251778721809387, + 0.017970982939004898, + 0.012679188512265682, + 0.01695670373737812, + 0.014885583892464638, + 0.021190734580159187, + 0.015440091490745544, + 0.006829515099525452, + 0.009115275926887989, + 0.0107414023950696, + 0.017152519896626472, + 0.007812106981873512, + 0.007356643211096525, + 0.01443744357675314, + 0.007573714014142752, + 0.02540832757949829, + 0.013397159054875374, + 0.014852593652904034, + 0.02045322395861149, + 0.013171376660466194, + 0.014673124067485332, + 0.012189861387014389, + 0.018458491191267967, + 0.006281523499637842, + 0.008440783247351646, + 0.005457778926938772, + 0.0035378376487642527, + 0.009102108888328075, + 0.006196653936058283, + 0.004465506877750158, + 0.004749605897814035, + 0.0035849250853061676, + 0.004146324470639229, + 0.005315884482115507, + 0.005647033452987671, + 0.004859439097344875, + 0.0046231807209551334, + 0.0041047208942472935, + 0.00534107256680727, + 0.01820136234164238, + 0.006935552693903446, + 0.00861707329750061, + 0.010272243060171604, + 0.013478518463671207, + 0.013354982249438763, + 0.08191978931427002, + 0.07016471028327942, + 0.02906418964266777, + 0.015894388779997826, + 0.010864883661270142, + 0.052551280707120895, + 0.02429855428636074, + 0.02141527459025383, + 0.010828188620507717, + 0.007915251888334751, + 0.006294147111475468, + 0.011569342575967312, + 0.005766183137893677, + 0.007129869423806667, + 0.011420353315770626, + 0.01683727465569973, + 0.005931674502789974, + 0.016472630202770233, + 0.004321529064327478, + 0.0037023648619651794, + 0.0034147303085774183, + 0.00522439694032073, + 0.0032510142773389816, + 0.005025956314057112, + 0.003938538488000631, + 0.00299075897783041, + 0.007119851186871529, + 0.007543097250163555, + 0.0076151383109390736, + 0.009007111191749573, + 0.012621608562767506, + 0.004458151757717133, + 0.003591531654819846, + 0.0058243670500814915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.062039099633693695 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702800333499908 + ], + "opposite_logprob_tokens": [ + -2.127027988433838 + ], + "logprob_reward": -0.12702800333499908, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702800333499908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0347299724817276, + 0.041530732065439224, + 0.03980294242501259, + 0.06336367875337601, + 0.03781386464834213, + 0.5298492908477783, + 0.0970768854022026, + 0.04587749391794205, + 0.07249897718429565, + 0.03745625168085098 + ], + "credit_map": [ + 0.003944581374526024, + 0.005291272420436144, + 0.0038718082942068577, + 0.0043853530660271645, + 0.003912295680493116, + 0.004142129328101873, + 0.00556388171389699, + 0.003618646878749132, + 0.003976399078965187, + 0.005743416026234627, + 0.005798717029392719, + 0.004738603252917528, + 0.00402770284563303, + 0.00389862060546875, + 0.005432785954326391, + 0.007914485409855843, + 0.0064755710773169994, + 0.0074556670151650906, + 0.004865678958594799, + 0.005014674738049507, + 0.0038818956818431616, + 0.0035313256084918976, + 0.0030239017214626074, + 0.00555422343313694, + 0.005579520482569933, + 0.003421128261834383, + 0.020641906186938286, + 0.004114398267120123, + 0.0027504826430231333, + 0.005113655235618353, + 0.003050563158467412, + 0.0186920166015625, + 0.005170103162527084, + 0.0036854450590908527, + 0.0028591440059244633, + 0.0038201725110411644, + 0.006957820616662502, + 0.004911112133413553, + 0.0040519097819924355, + 0.006358152721077204, + 0.04369739070534706, + 0.015830175951123238, + 0.02040926367044449, + 0.018795307725667953, + 0.028707023710012436, + 0.03142876178026199, + 0.19900280237197876, + 0.17197853326797485, + 0.011620435863733292, + 0.007963206619024277, + 0.014187419787049294, + 0.02955719828605652, + 0.011234115809202194, + 0.010411081835627556, + 0.0068265413865447044, + 0.0052768876776099205, + 0.004432248417288065, + 0.004367411136627197, + 0.005668019410222769, + 0.004862117115408182, + 0.006121906451880932, + 0.007057902403175831, + 0.006223159376531839, + 0.007144724950194359, + 0.0111320149153471, + 0.010678392834961414, + 0.009576138108968735, + 0.01184240821748972, + 0.007195301353931427, + 0.007864243350923061, + 0.008356740698218346, + 0.005853735376149416, + 0.004133588168770075, + 0.0050302171148359776, + 0.004844194743782282, + 0.008134910836815834, + 0.007086758967489004, + 0.00267616449855268, + 0.002394989365711808, + 0.003155424725264311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.12702800333499908 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12701687216758728 + ], + "opposite_logprob_tokens": [ + -2.12701678276062 + ], + "logprob_reward": -0.12701687216758728, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12701687216758728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06982888281345367, + 0.04403397813439369, + 0.04134943708777428, + 0.27091872692108154, + 0.04895762354135513, + 0.2289179265499115, + 0.1335369348526001, + 0.06535287946462631, + 0.05257999524474144, + 0.0445236973464489 + ], + "credit_map": [ + 0.008010195568203926, + 0.01063958927989006, + 0.008838485926389694, + 0.008579050190746784, + 0.007290099747478962, + 0.007679977919906378, + 0.009445061907172203, + 0.009346412494778633, + 0.00540497712790966, + 0.00595386466011405, + 0.005654362495988607, + 0.005852161906659603, + 0.004209393635392189, + 0.0039291176944971085, + 0.006988278590142727, + 0.006041817832738161, + 0.0056418245658278465, + 0.006000107154250145, + 0.004214200656861067, + 0.006875894032418728, + 0.00518031045794487, + 0.00447421008720994, + 0.0037968044634908438, + 0.005166081711649895, + 0.025940515100955963, + 0.019659480080008507, + 0.07628682255744934, + 0.023832308128476143, + 0.014804374426603317, + 0.025967534631490707, + 0.0146115031093359, + 0.06981612741947174, + 0.004603625275194645, + 0.006477003917098045, + 0.005050200037658215, + 0.005313633941113949, + 0.009522700682282448, + 0.00613363878801465, + 0.005140329245477915, + 0.006716483738273382, + 0.018844453617930412, + 0.011756048537790775, + 0.009589526802301407, + 0.012282906100153923, + 0.018980655819177628, + 0.017399214208126068, + 0.08255686610937119, + 0.05750824511051178, + 0.02087554521858692, + 0.007205107249319553, + 0.01508065965026617, + 0.051081523299217224, + 0.014091863296926022, + 0.014216740615665913, + 0.005813254974782467, + 0.005172244273126125, + 0.00508343568071723, + 0.005406191572546959, + 0.006490317173302174, + 0.006044976878911257, + 0.0062592304311692715, + 0.014188628643751144, + 0.007127951364964247, + 0.014752141200006008, + 0.0070775882340967655, + 0.006364781875163317, + 0.0055404906161129475, + 0.007982663810253143, + 0.006745039951056242, + 0.00617428682744503, + 0.008176987990736961, + 0.0045181517489254475, + 0.0055284276604652405, + 0.006844654679298401, + 0.006202931050211191, + 0.0044166408479213715, + 0.009124984964728355, + 0.00300226011313498, + 0.0023630792275071144, + 0.007040713913738728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.12701687216758728 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the vibrant style of Ghibli animation, the camera executes continuous push-ins, gliding smoothly through the dense foliage of a lush, sun-dappled jungle to focus on the vibrant courtship ritual. Warm golden sunlight penetrates the thick canopy, creating shifting patterns of spots on oversized waxy leaves and mossy branches. The central subject is a colorful lizard, similar to Anolis ricordi, with a body of brilliant emerald green, speckled with turquoise along its spine, and large, expressive black eyes brimming with playful energy. It pauses for a moment on a large fern leaf, rhythmically expanding its prominent, translucent scarlet dewdrops, catching the light like a stained glass window. In rapid pursuit, another smaller, more golden counterpart nimbly climbs a dense, twisting vine. Their playful chase is a joyful, fluid dance, as they dart around the gnarled roots of ancient trees and gracefully leap between broad, dew-drenched leaves, their movements swift and light. The entire scene is brimming with life and color, taking on the soft, hand-painted textures and magical realism characteristic of Ghibli masterpieces, evoking a wonderfully vibrant atmosphere.", + "seed": 10634, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_7/index_5007_seed_10634_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0734488258138299, + "vlm_logprob": -0.0734488258138299, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.0734488258138299, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02921798638999462 + ], + "opposite_logprob_tokens": [ + -3.7792179584503174 + ], + "logprob_reward": -0.02921798638999462, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02921798638999462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18170227110385895, + 0.08435099571943283, + 0.28894203901290894, + 0.02925749495625496, + 0.0897623747587204, + 0.10243885219097137, + 0.034641873091459274, + 0.06364072859287262, + 0.061552345752716064, + 0.06371109187602997 + ], + "credit_map": [ + 0.021319936960935593, + 0.011701907031238079, + 0.030011942610144615, + 0.023447906598448753, + 0.015935711562633514, + 0.0422016903758049, + 0.022503642365336418, + 0.014579521492123604, + 0.010124280117452145, + 0.004462342243641615, + 0.009158404543995857, + 0.0060938396491110325, + 0.006466764025390148, + 0.009804528206586838, + 0.016031676903367043, + 0.022209148854017258, + 0.023998474702239037, + 0.0185267124325037, + 0.025428924709558487, + 0.02357150986790657, + 0.030423544347286224, + 0.07083138078451157, + 0.02120193839073181, + 0.07495956122875214, + 0.0025972973089665174, + 0.004154528025537729, + 0.002630090806633234, + 0.003714702557772398, + 0.0033958200365304947, + 0.004545092582702637, + 0.0027769114822149277, + 0.005443052388727665, + 0.012257616966962814, + 0.010296403430402279, + 0.011089561507105827, + 0.011335342191159725, + 0.008440541103482246, + 0.01294393464922905, + 0.012149740941822529, + 0.011249233968555927, + 0.010757546871900558, + 0.0150340236723423, + 0.007724331691861153, + 0.014819306321442127, + 0.013065371662378311, + 0.013895520940423012, + 0.014878816902637482, + 0.012263931334018707, + 0.0030157221481204033, + 0.002721031429246068, + 0.004843868315219879, + 0.003976276144385338, + 0.00488869147375226, + 0.005660586059093475, + 0.0044122482649981976, + 0.0051234448328614235, + 0.005521262530237436, + 0.00719966646283865, + 0.009801195003092289, + 0.007440108805894852, + 0.007454971317201853, + 0.008718496188521385, + 0.008454994298517704, + 0.009050032123923302, + 0.00831475667655468, + 0.007241052575409412, + 0.006018968764692545, + 0.006351322866976261, + 0.007848785258829594, + 0.006638390477746725, + 0.009824768640100956, + 0.009314295835793018, + 0.005439809989184141, + 0.008567937649786472, + 0.004701659083366394, + 0.009619646705687046, + 0.006364440079778433, + 0.007584994658827782, + 0.010352112352848053, + 0.0110804857686162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.02921798638999462 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204937033355236 + ], + "opposite_logprob_tokens": [ + -4.264204978942871 + ], + "logprob_reward": -0.014204937033355236, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204937033355236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13344180583953857, + 0.13229002058506012, + 0.21561233699321747, + 0.05924324691295624, + 0.05671016499400139, + 0.06813300400972366, + 0.06615664064884186, + 0.08030805736780167, + 0.08904237300157547, + 0.09906233102083206 + ], + "credit_map": [ + 0.01423707976937294, + 0.014065446332097054, + 0.028406230732798576, + 0.01037598866969347, + 0.017412032932043076, + 0.02299443818628788, + 0.014901858754456043, + 0.011048740707337856, + 0.012383035384118557, + 0.010958891361951828, + 0.01742919534444809, + 0.008562288247048855, + 0.014662569388747215, + 0.02538439817726612, + 0.010260828770697117, + 0.03264881670475006, + 0.022440562024712563, + 0.01502464059740305, + 0.017827678471803665, + 0.011764392256736755, + 0.017845874652266502, + 0.06229044497013092, + 0.009729307144880295, + 0.058689434081315994, + 0.007219596765935421, + 0.005584234371781349, + 0.0061704246327281, + 0.006978259887546301, + 0.006717996671795845, + 0.00935115572065115, + 0.005131436046212912, + 0.012090140953660011, + 0.006847518961876631, + 0.010802028700709343, + 0.005772866308689117, + 0.0051732114516198635, + 0.007598366588354111, + 0.008488751016557217, + 0.005876519717276096, + 0.006150903180241585, + 0.007289852015674114, + 0.008927281945943832, + 0.005518249236047268, + 0.010119573213160038, + 0.00812362041324377, + 0.011134728789329529, + 0.007885188795626163, + 0.009134508669376373, + 0.005674012470990419, + 0.004202534444630146, + 0.006364854518324137, + 0.011930548585951328, + 0.008451290428638458, + 0.011472492478787899, + 0.0072131529450416565, + 0.010847757570445538, + 0.007555366028100252, + 0.006922289729118347, + 0.013802194967865944, + 0.007242413703352213, + 0.010110866278409958, + 0.010776137001812458, + 0.010323606431484222, + 0.013575183227658272, + 0.010628050193190575, + 0.01903548464179039, + 0.008989340625703335, + 0.009696271270513535, + 0.009027883410453796, + 0.007759811822324991, + 0.011447008699178696, + 0.012458518147468567, + 0.008831480517983437, + 0.008805128745734692, + 0.006868388503789902, + 0.014682975597679615, + 0.009798836894333363, + 0.013689483515918255, + 0.02252861298620701, + 0.01385742798447609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.014204937033355236 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494843050837517 + ], + "opposite_logprob_tokens": [ + -2.929948329925537 + ], + "logprob_reward": -0.05494843050837517, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494843050837517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08940702676773071, + 0.07049557566642761, + 0.11360576003789902, + 0.226546972990036, + 0.06399037688970566, + 0.054812584072351456, + 0.13524621725082397, + 0.14080671966075897, + 0.04617882892489433, + 0.058909859508275986 + ], + "credit_map": [ + 0.009802336804568768, + 0.0070643313229084015, + 0.02346651442348957, + 0.006886247545480728, + 0.007977823726832867, + 0.013770154677331448, + 0.013374975882470608, + 0.00706464471295476, + 0.00640804972499609, + 0.005772853270173073, + 0.008822084404528141, + 0.006630744785070419, + 0.009897267445921898, + 0.009791310876607895, + 0.013039665296673775, + 0.010133601725101471, + 0.008219365030527115, + 0.00997139047831297, + 0.007950065657496452, + 0.010363951325416565, + 0.009357710368931293, + 0.02623310312628746, + 0.012715129181742668, + 0.028795037418603897, + 0.03107638470828533, + 0.022953201085329056, + 0.024452047422528267, + 0.028076529502868652, + 0.028742875903844833, + 0.038026779890060425, + 0.025292066857218742, + 0.027927078306674957, + 0.007507778704166412, + 0.012349886819720268, + 0.005528424866497517, + 0.005710928235203028, + 0.00707067409530282, + 0.008992854505777359, + 0.007012529764324427, + 0.009817298501729965, + 0.00710025429725647, + 0.00865976419299841, + 0.005890024360269308, + 0.0072457208298146725, + 0.007151983678340912, + 0.005787596572190523, + 0.005122508853673935, + 0.007854729890823364, + 0.01261209324002266, + 0.011178023181855679, + 0.015702633187174797, + 0.01440577581524849, + 0.018432199954986572, + 0.025567518547177315, + 0.015560927800834179, + 0.021787051111459732, + 0.01514445710927248, + 0.022303150966763496, + 0.017927173525094986, + 0.014690750278532505, + 0.017774809151887894, + 0.018164798617362976, + 0.014491808600723743, + 0.020309776067733765, + 0.005247683264315128, + 0.010065406560897827, + 0.004114287905395031, + 0.00412882911041379, + 0.006966834422200918, + 0.00412551686167717, + 0.004706668667495251, + 0.006823604926466942, + 0.0051421369425952435, + 0.00805277656763792, + 0.0033544455654919147, + 0.008332830853760242, + 0.005128251854330301, + 0.00994057860225439, + 0.010913481935858727, + 0.008045359514653683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.05494843050837517 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798640638589859 + ], + "opposite_logprob_tokens": [ + -1.8048640489578247 + ], + "logprob_reward": -0.1798640638589859, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798640638589859, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21727295219898224, + 0.17206710577011108, + 0.11453236639499664, + 0.07681605219841003, + 0.05986054986715317, + 0.07007010281085968, + 0.04187407344579697, + 0.17994263768196106, + 0.033495962619781494, + 0.03406815230846405 + ], + "credit_map": [ + 0.020417217165231705, + 0.02243497408926487, + 0.05202841758728027, + 0.024251144379377365, + 0.024768173694610596, + 0.03609304130077362, + 0.03031402826309204, + 0.006965956185013056, + 0.014425460249185562, + 0.01065731979906559, + 0.021837791427969933, + 0.011735016480088234, + 0.017322788015007973, + 0.02505224384367466, + 0.040188129991292953, + 0.03084835782647133, + 0.013739691115915775, + 0.006601119413971901, + 0.01590438187122345, + 0.012043376453220844, + 0.012149959802627563, + 0.017626797780394554, + 0.012367918156087399, + 0.02409912273287773, + 0.0092011084780097, + 0.005310844164341688, + 0.010390580631792545, + 0.004687933716922998, + 0.009723789989948273, + 0.010410992428660393, + 0.007832049392163754, + 0.019258752465248108, + 0.007795337121933699, + 0.007393362931907177, + 0.006847849115729332, + 0.006532199215143919, + 0.009314298629760742, + 0.008139191195368767, + 0.00733954319730401, + 0.0064987679943442345, + 0.007235338445752859, + 0.007675842382013798, + 0.006104862317442894, + 0.006710682064294815, + 0.00814451277256012, + 0.012929663993418217, + 0.010873380117118359, + 0.010395817458629608, + 0.002974293427541852, + 0.002488992176949978, + 0.003388995537534356, + 0.006896256934851408, + 0.004043507855385542, + 0.00968957133591175, + 0.004082110710442066, + 0.008310345932841301, + 0.013805770315229893, + 0.023843513801693916, + 0.037985723465681076, + 0.02207702398300171, + 0.02045481465756893, + 0.02203427255153656, + 0.016912246122956276, + 0.02282928302884102, + 0.003600648371502757, + 0.004519050475209951, + 0.004733942449092865, + 0.003910754807293415, + 0.004860338289290667, + 0.0040444061160087585, + 0.00473741814494133, + 0.0030894060619175434, + 0.004076960496604443, + 0.0028987200930714607, + 0.002213959814980626, + 0.004532478749752045, + 0.004050213843584061, + 0.00505463033914566, + 0.007108437828719616, + 0.004132749978452921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.1798640638589859 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900871127843857 + ], + "opposite_logprob_tokens": [ + -2.4640088081359863 + ], + "logprob_reward": -0.08900871127843857, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900871127843857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14198052883148193, + 0.07985217124223709, + 0.07555345445871353, + 0.06150199845433235, + 0.0832345262169838, + 0.19725528359413147, + 0.09370625019073486, + 0.15052291750907898, + 0.07084205001592636, + 0.0455508753657341 + ], + "credit_map": [ + 0.014191527850925922, + 0.01683233492076397, + 0.024524139240384102, + 0.018038487061858177, + 0.02285020798444748, + 0.021363425999879837, + 0.015261738561093807, + 0.008918672800064087, + 0.01310650072991848, + 0.0050536696799099445, + 0.012469804845750332, + 0.00857249740511179, + 0.009746383875608444, + 0.011592782102525234, + 0.010719268582761288, + 0.00859126728028059, + 0.01592651568353176, + 0.0066196308471262455, + 0.011410298757255077, + 0.005984233692288399, + 0.009490485303103924, + 0.009933636523783207, + 0.007922666147351265, + 0.008265985175967216, + 0.007295669987797737, + 0.0048995246179401875, + 0.009260755963623524, + 0.004392927046865225, + 0.0072666010819375515, + 0.006377041805535555, + 0.007205907721072435, + 0.014803570695221424, + 0.010754646733403206, + 0.010611327365040779, + 0.00828535296022892, + 0.007857931777834892, + 0.01180566567927599, + 0.011872591450810432, + 0.011658363975584507, + 0.010388641618192196, + 0.031843654811382294, + 0.03494202718138695, + 0.01805058866739273, + 0.03042500652372837, + 0.0209614597260952, + 0.020762428641319275, + 0.01727467216551304, + 0.022995444014668465, + 0.007941254414618015, + 0.007884589023888111, + 0.010630400851368904, + 0.012215126305818558, + 0.011709415353834629, + 0.018670495599508286, + 0.01080970000475645, + 0.01384526863694191, + 0.012190156616270542, + 0.015018376521766186, + 0.023649396374821663, + 0.0178977120667696, + 0.018660439178347588, + 0.024473872035741806, + 0.018643269315361977, + 0.01998968981206417, + 0.0072265188209712505, + 0.007228592410683632, + 0.004282310605049133, + 0.005642387084662914, + 0.017628448083996773, + 0.005525043699890375, + 0.008088001981377602, + 0.015220746397972107, + 0.00363803724758327, + 0.005257921293377876, + 0.0023702967446297407, + 0.0069387368857860565, + 0.005144175142049789, + 0.006344915833324194, + 0.0095566026866436, + 0.006300194188952446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.08900871127843857 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the vibrant style of Ghibli animation, the camera executes continuous push-ins, gliding smoothly through the dense foliage of a lush, sun-dappled jungle to focus on the vibrant courtship ritual. Warm golden sunlight penetrates the thick canopy, creating shifting patterns of spots on oversized waxy leaves and mossy branches. The central subject is a colorful lizard, similar to Anolis ricordi, with a body of brilliant emerald green, speckled with turquoise along its spine, and large, expressive black eyes brimming with playful energy. It pauses for a moment on a large fern leaf, rhythmically expanding its prominent, translucent scarlet dewdrops, catching the light like a stained glass window. In rapid pursuit, another smaller, more golden counterpart nimbly climbs a dense, twisting vine. Their playful chase is a joyful, fluid dance, as they dart around the gnarled roots of ancient trees and gracefully leap between broad, dew-drenched leaves, their movements swift and light. The entire scene is brimming with life and color, taking on the soft, hand-painted textures and magical realism characteristic of Ghibli masterpieces, evoking a wonderfully vibrant atmosphere.", + "seed": 10635, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_7/index_5007_seed_10635_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08097360916435718, + "vlm_logprob": -0.08097360916435718, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.08097360916435718, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0956512987613678 + ], + "opposite_logprob_tokens": [ + -2.470651388168335 + ], + "logprob_reward": -0.0956512987613678, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0956512987613678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.134482279419899, + 0.1318833827972412, + 0.0994100347161293, + 0.08313707262277603, + 0.04982919991016388, + 0.05265060439705849, + 0.029917346313595772, + 0.0321764312684536, + 0.20640820264816284, + 0.18010535836219788 + ], + "credit_map": [ + 0.017128100618720055, + 0.014420153573155403, + 0.013122654519975185, + 0.011056934483349323, + 0.02769291214644909, + 0.01945492811501026, + 0.015299681574106216, + 0.016306933015584946, + 0.02166661061346531, + 0.010189174674451351, + 0.01178503967821598, + 0.019347215071320534, + 0.01448808517307043, + 0.01666649430990219, + 0.02034219354391098, + 0.01739858090877533, + 0.01451050117611885, + 0.009439105167984962, + 0.009318972937762737, + 0.011062170378863811, + 0.018499158322811127, + 0.018192050978541374, + 0.010513048619031906, + 0.007875041104853153, + 0.008926637470722198, + 0.003332218388095498, + 0.004703484009951353, + 0.006433003582060337, + 0.00489139836281538, + 0.026215167716145515, + 0.0036748419515788555, + 0.02496032603085041, + 0.004629739094525576, + 0.006131916772574186, + 0.00887876097112894, + 0.004128105007112026, + 0.007560778874903917, + 0.0058199819177389145, + 0.006300063803792, + 0.006379858125001192, + 0.004882483277469873, + 0.005591523367911577, + 0.0061663114465773106, + 0.004850888624787331, + 0.006533523555845022, + 0.006079079583287239, + 0.007102931383997202, + 0.011443867348134518, + 0.0034419114235788584, + 0.002649052767083049, + 0.0033993932884186506, + 0.003615970490500331, + 0.002780116396024823, + 0.0040163639932870865, + 0.004298246465623379, + 0.005716294515877962, + 0.0033656293526291847, + 0.0031376597471535206, + 0.004881155211478472, + 0.00500510074198246, + 0.00377941713668406, + 0.0028497588355094194, + 0.00455973157659173, + 0.004597979132086039, + 0.021559756249189377, + 0.037954866886138916, + 0.01948762685060501, + 0.028729934245347977, + 0.015274186618626118, + 0.037407029420137405, + 0.015424533747136593, + 0.030570276081562042, + 0.009014281444251537, + 0.005530751775950193, + 0.006095371674746275, + 0.00578314671292901, + 0.011026010848581791, + 0.0722939521074295, + 0.020211875438690186, + 0.05014999955892563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0956512987613678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592933520674706 + ], + "opposite_logprob_tokens": [ + -3.8955929279327393 + ], + "logprob_reward": -0.020592933520674706, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592933520674706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11852189153432846, + 0.07253473997116089, + 0.13329248130321503, + 0.14828144013881683, + 0.035228431224823, + 0.06099316105246544, + 0.0219045989215374, + 0.052358582615852356, + 0.22073747217655182, + 0.13614723086357117 + ], + "credit_map": [ + 0.018336554989218712, + 0.019436750560998917, + 0.012712744064629078, + 0.01451780740171671, + 0.010388804599642754, + 0.01452765055000782, + 0.015090201050043106, + 0.013511373661458492, + 0.012177260592579842, + 0.005582088138908148, + 0.0068417685106396675, + 0.010772950015962124, + 0.00971043761819601, + 0.013786232098937035, + 0.006879297085106373, + 0.006784690544009209, + 0.017595795914530754, + 0.019641445949673653, + 0.014665638096630573, + 0.01725834794342518, + 0.020925553515553474, + 0.016869692131876945, + 0.011760571971535683, + 0.014575423672795296, + 0.017719032242894173, + 0.008395984768867493, + 0.011125868186354637, + 0.011527977883815765, + 0.01029267068952322, + 0.0423712320625782, + 0.007054360117763281, + 0.03979429975152016, + 0.003731091972440481, + 0.005101379938423634, + 0.006729178596287966, + 0.003013747278600931, + 0.004647471942007542, + 0.003997433930635452, + 0.004247957840561867, + 0.003760170191526413, + 0.004956134594976902, + 0.004081747494637966, + 0.00830824114382267, + 0.008599932305514812, + 0.008540283888578415, + 0.006546796765178442, + 0.007907220162451267, + 0.012052798643708229, + 0.002081696642562747, + 0.0017204248579218984, + 0.002141648670658469, + 0.002262034686282277, + 0.003660202957689762, + 0.002579944906756282, + 0.0039062839932739735, + 0.0035523604601621628, + 0.0073667182587087154, + 0.004759915638715029, + 0.01092744804918766, + 0.009326919913291931, + 0.006206769496202469, + 0.003461073152720928, + 0.005452408920973539, + 0.004857327789068222, + 0.022326508536934853, + 0.03044133074581623, + 0.020280664786696434, + 0.021237673237919807, + 0.015601099468767643, + 0.04849924147129059, + 0.014206139370799065, + 0.04814479872584343, + 0.008221587166190147, + 0.004728557076305151, + 0.006546553689986467, + 0.00619956711307168, + 0.009254357777535915, + 0.0505884550511837, + 0.015420446172356606, + 0.03518769145011902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.020592933520674706 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799941927194595 + ], + "opposite_logprob_tokens": [ + -3.5297999382019043 + ], + "logprob_reward": -0.029799941927194595, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799941927194595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08190050721168518, + 0.027436360716819763, + 0.0412144660949707, + 0.21387599408626556, + 0.027633195742964745, + 0.06949703395366669, + 0.057759303599596024, + 0.08855519443750381, + 0.35076382756233215, + 0.041364122182130814 + ], + "credit_map": [ + 0.014784523285925388, + 0.008269883692264557, + 0.0063040051609277725, + 0.00693057244643569, + 0.019880034029483795, + 0.009587764739990234, + 0.006108895875513554, + 0.010034835897386074, + 0.004044930916279554, + 0.002723861485719681, + 0.0022212134208530188, + 0.0032189935445785522, + 0.004824425559490919, + 0.00563061935827136, + 0.0025559759233146906, + 0.002216342370957136, + 0.004699127282947302, + 0.004437524825334549, + 0.0032167304307222366, + 0.0038637241814285517, + 0.009214639663696289, + 0.00741624878719449, + 0.004187063314020634, + 0.004179405979812145, + 0.019987791776657104, + 0.007479621563106775, + 0.012232298031449318, + 0.014356723055243492, + 0.016384199261665344, + 0.0653507336974144, + 0.014683765359222889, + 0.06340086460113525, + 0.0022425460629165173, + 0.0024634546134620905, + 0.0025188529398292303, + 0.0039080660790205, + 0.0053766160272061825, + 0.003685345407575369, + 0.004405369982123375, + 0.00303294463083148, + 0.0062729776836931705, + 0.005719876382499933, + 0.008641639724373817, + 0.009226380847394466, + 0.010501936078071594, + 0.007634782698005438, + 0.008758376352488995, + 0.01274106465280056, + 0.006502607837319374, + 0.004049314651638269, + 0.006362739484757185, + 0.005612880457192659, + 0.0054793222807347775, + 0.01006247103214264, + 0.009972433559596539, + 0.00971753429621458, + 0.007462689653038979, + 0.005337354261428118, + 0.02492361143231392, + 0.016883820295333862, + 0.011580364778637886, + 0.006729920394718647, + 0.010035676881670952, + 0.005601760931313038, + 0.050217609852552414, + 0.04300568625330925, + 0.02597823739051819, + 0.03777991980314255, + 0.017096947878599167, + 0.06637732684612274, + 0.015866639092564583, + 0.09444146603345871, + 0.002006967319175601, + 0.0012147738598287106, + 0.001965884817764163, + 0.003466833382844925, + 0.002751819556578994, + 0.014136077836155891, + 0.00358729250729084, + 0.01223447360098362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.029799941927194595 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986416816711426 + ], + "opposite_logprob_tokens": [ + -1.8048641681671143 + ], + "logprob_reward": -0.17986416816711426, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986416816711426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20691423118114471, + 0.0648362785577774, + 0.032193686813116074, + 0.1589534878730774, + 0.06538355350494385, + 0.11526813358068466, + 0.03358073905110359, + 0.08983177691698074, + 0.1769627332687378, + 0.056075356900691986 + ], + "credit_map": [ + 0.023858889937400818, + 0.02788364887237549, + 0.023935267701745033, + 0.020848466083407402, + 0.029254382476210594, + 0.030480455607175827, + 0.02388853020966053, + 0.026764577254652977, + 0.01000905130058527, + 0.007440738379955292, + 0.006139586213976145, + 0.00835219956934452, + 0.007942893542349339, + 0.01069570705294609, + 0.007951500825583935, + 0.006304601673036814, + 0.004706584382802248, + 0.00437739584594965, + 0.003946397919207811, + 0.0033883466385304928, + 0.004915923345834017, + 0.00486940611153841, + 0.0027507899794727564, + 0.0032388437539339066, + 0.015102222561836243, + 0.008287426084280014, + 0.00883831363171339, + 0.010690505616366863, + 0.017025316134095192, + 0.04576113075017929, + 0.009352837689220905, + 0.04389572888612747, + 0.006764499470591545, + 0.008840968832373619, + 0.013292388059198856, + 0.006475895177572966, + 0.008039061911404133, + 0.007231156341731548, + 0.007326334714889526, + 0.007413249928504229, + 0.011238841339945793, + 0.014075424522161484, + 0.011221986263990402, + 0.00963889341801405, + 0.013941061682999134, + 0.014637789689004421, + 0.014227762818336487, + 0.026286374777555466, + 0.0016100109787657857, + 0.0020165336318314075, + 0.002585099544376135, + 0.002556344261392951, + 0.004502418916672468, + 0.0038107370492070913, + 0.00883979257196188, + 0.007659798953682184, + 0.00949863065034151, + 0.010176057927310467, + 0.011646982282400131, + 0.01146805938333273, + 0.015300007537007332, + 0.010853319428861141, + 0.012573247775435448, + 0.008315476588904858, + 0.0158805213868618, + 0.024426139891147614, + 0.016148177906870842, + 0.016909031197428703, + 0.013843728229403496, + 0.040453385561704636, + 0.01228216104209423, + 0.03701959177851677, + 0.005487756337970495, + 0.003974725492298603, + 0.004025122616440058, + 0.005184767302125692, + 0.007603224366903305, + 0.014075512997806072, + 0.00639112014323473, + 0.00933312438428402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.17986416816711426 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895970344543457 + ], + "opposite_logprob_tokens": [ + -2.5789597034454346 + ], + "logprob_reward": -0.07895970344543457, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895970344543457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04983241856098175, + 0.04851727932691574, + 0.030908847227692604, + 0.10297297686338425, + 0.05842363089323044, + 0.0888199731707573, + 0.07043646275997162, + 0.051805056631565094, + 0.10713210701942444, + 0.3911512494087219 + ], + "credit_map": [ + 0.007051143795251846, + 0.006134640425443649, + 0.0051602814346551895, + 0.004795518703758717, + 0.012312428094446659, + 0.005776625592261553, + 0.0038635015953332186, + 0.004738279618322849, + 0.010196264833211899, + 0.004469851963222027, + 0.0030061681754887104, + 0.006510681938380003, + 0.004101690836250782, + 0.007065457757562399, + 0.006975146476179361, + 0.006192018743604422, + 0.003752092830836773, + 0.00449351267889142, + 0.0035290492232888937, + 0.004878438543528318, + 0.0035440935753285885, + 0.003421210218220949, + 0.003583083860576153, + 0.003707366995513439, + 0.009879269637167454, + 0.0028775371611118317, + 0.005610316526144743, + 0.006461838725954294, + 0.006609936244785786, + 0.033201005309820175, + 0.005095979664474726, + 0.033237095922231674, + 0.004725603386759758, + 0.008108921349048615, + 0.009821775369346142, + 0.007736233063042164, + 0.009711185470223427, + 0.005871365778148174, + 0.0056332750245928764, + 0.006815275643020868, + 0.00874205119907856, + 0.009872189722955227, + 0.011496090330183506, + 0.00949170719832182, + 0.012508377432823181, + 0.012865494005382061, + 0.012980256229639053, + 0.010863805189728737, + 0.004939209669828415, + 0.004868784453719854, + 0.00704607181251049, + 0.006080565042793751, + 0.008039112202823162, + 0.00927444826811552, + 0.014779736287891865, + 0.015408537350594997, + 0.0057035163044929504, + 0.00546783534809947, + 0.005985387600958347, + 0.008163484744727612, + 0.006536860018968582, + 0.00495886942371726, + 0.010847377590835094, + 0.004141724202781916, + 0.013503140769898891, + 0.011151728220283985, + 0.017546869814395905, + 0.012371281161904335, + 0.00776834087446332, + 0.02258159965276718, + 0.0057192472741007805, + 0.0164899080991745, + 0.01985834538936615, + 0.013596286997199059, + 0.016468219459056854, + 0.014168166555464268, + 0.02704181708395481, + 0.15030217170715332, + 0.04499005898833275, + 0.10472619533538818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.07895970344543457 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the vibrant style of Ghibli animation, the camera executes continuous push-ins, gliding smoothly through the dense foliage of a lush, sun-dappled jungle to focus on the vibrant courtship ritual. Warm golden sunlight penetrates the thick canopy, creating shifting patterns of spots on oversized waxy leaves and mossy branches. The central subject is a colorful lizard, similar to Anolis ricordi, with a body of brilliant emerald green, speckled with turquoise along its spine, and large, expressive black eyes brimming with playful energy. It pauses for a moment on a large fern leaf, rhythmically expanding its prominent, translucent scarlet dewdrops, catching the light like a stained glass window. In rapid pursuit, another smaller, more golden counterpart nimbly climbs a dense, twisting vine. Their playful chase is a joyful, fluid dance, as they dart around the gnarled roots of ancient trees and gracefully leap between broad, dew-drenched leaves, their movements swift and light. The entire scene is brimming with life and color, taking on the soft, hand-painted textures and magical realism characteristic of Ghibli masterpieces, evoking a wonderfully vibrant atmosphere.", + "seed": 10636, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_7/index_5007_seed_10636_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09038379527628422, + "vlm_logprob": -0.09038379527628422, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.09038379527628422, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055346209555864334 + ], + "opposite_logprob_tokens": [ + -3.0553462505340576 + ], + "logprob_reward": -0.055346209555864334, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.055346209555864334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13055408000946045, + 0.06676965951919556, + 0.03183424472808838, + 0.0361044779419899, + 0.17623083293437958, + 0.10712476074695587, + 0.062918521463871, + 0.09324154257774353, + 0.27471724152565, + 0.020504606887698174 + ], + "credit_map": [ + 0.0149577921256423, + 0.00775239709764719, + 0.021986180916428566, + 0.0256076380610466, + 0.008269145153462887, + 0.007886147126555443, + 0.026169443503022194, + 0.017925338819622993, + 0.00722123309969902, + 0.011338475160300732, + 0.010101077146828175, + 0.008297753520309925, + 0.004991733469069004, + 0.005970122292637825, + 0.01252859365195036, + 0.006320667918771505, + 0.00410334812477231, + 0.0037520593032240868, + 0.004092772025614977, + 0.0056209806352853775, + 0.002251957543194294, + 0.0034334182273596525, + 0.005053150467574596, + 0.0035265597980469465, + 0.003324317978695035, + 0.004263062495738268, + 0.005683205556124449, + 0.006827089469879866, + 0.003600929630920291, + 0.003231051377952099, + 0.0038268305361270905, + 0.005347990896552801, + 0.015951644629240036, + 0.04846720024943352, + 0.04335566610097885, + 0.010629363358020782, + 0.01996469683945179, + 0.009970455430448055, + 0.015111783519387245, + 0.012780019082129002, + 0.01647987961769104, + 0.020579198375344276, + 0.02400711365044117, + 0.010617990978062153, + 0.007868574932217598, + 0.009488604962825775, + 0.00915248692035675, + 0.008930915035307407, + 0.005407251883298159, + 0.006594707258045673, + 0.007941088639199734, + 0.006023569963872433, + 0.005605241283774376, + 0.006382688879966736, + 0.009469919838011265, + 0.01549405138939619, + 0.005832142196595669, + 0.003388900775462389, + 0.01844959706068039, + 0.00970267504453659, + 0.006740922573953867, + 0.03250895440578461, + 0.008799820207059383, + 0.007818534970283508, + 0.008129809983074665, + 0.011273016221821308, + 0.01069662906229496, + 0.004608459305018187, + 0.010683530941605568, + 0.11696640402078629, + 0.015068979933857918, + 0.09729042649269104, + 0.0018692101584747434, + 0.002076217904686928, + 0.0029289175290614367, + 0.003122737631201744, + 0.0024001512210816145, + 0.002195577137172222, + 0.0029328688979148865, + 0.002978925360366702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.055346209555864334 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608537696301937 + ], + "opposite_logprob_tokens": [ + -4.141085147857666 + ], + "logprob_reward": -0.01608537696301937, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608537696301937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035495907068252563, + 0.03038082644343376, + 0.017594236880540848, + 0.06188276410102844, + 0.04576634243130684, + 0.037665337324142456, + 0.06466217339038849, + 0.29102641344070435, + 0.4009459316730499, + 0.014580005779862404 + ], + "credit_map": [ + 0.00613298499956727, + 0.003283852245658636, + 0.005759795196354389, + 0.006096224300563335, + 0.0021585836075246334, + 0.0026975637301802635, + 0.005375577136874199, + 0.003991330042481422, + 0.00307128275744617, + 0.005080814473330975, + 0.004614659119397402, + 0.003791836090385914, + 0.0031800540164113045, + 0.002510541584342718, + 0.0053418949246406555, + 0.0027897474355995655, + 0.0019310492789372802, + 0.0021889160852879286, + 0.002004444133490324, + 0.0024934466928243637, + 0.0018359889509156346, + 0.001679396373219788, + 0.0035787380766123533, + 0.0018822584534063935, + 0.0044274297542870045, + 0.004090441390872002, + 0.006174413952976465, + 0.00639975955709815, + 0.00841566827148199, + 0.006079537328332663, + 0.014843959361314774, + 0.011451557278633118, + 0.0039291693829, + 0.009503957815468311, + 0.010018284432590008, + 0.0032963925041258335, + 0.0051539563573896885, + 0.0034779245033860207, + 0.00623584259301424, + 0.004150819964706898, + 0.005487455986440182, + 0.005655106157064438, + 0.006636119447648525, + 0.0048562451265752316, + 0.002422241261228919, + 0.003336292691528797, + 0.005686051212251186, + 0.0035858287010341883, + 0.005522176157683134, + 0.006670848466455936, + 0.008247017860412598, + 0.006031674798578024, + 0.0064949942752718925, + 0.006215862929821014, + 0.008632597513496876, + 0.016847005113959312, + 0.017794650048017502, + 0.011649959720671177, + 0.055932480841875076, + 0.03074650466442108, + 0.023826712742447853, + 0.10308938473463058, + 0.02435838244855404, + 0.023628361523151398, + 0.011328157968819141, + 0.014451832510530949, + 0.014973959885537624, + 0.0062029301188886166, + 0.011332288384437561, + 0.17495188117027283, + 0.021517768502235413, + 0.14618711173534393, + 0.0012959061423316598, + 0.0016400113236159086, + 0.001801257603801787, + 0.0016122909728437662, + 0.001752074807882309, + 0.0014609600184485316, + 0.0030933304224163294, + 0.001924176118336618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.01608537696301937 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031377524137497 + ], + "opposite_logprob_tokens": [ + -2.350313663482666 + ], + "logprob_reward": -0.10031377524137497, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031377524137497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0494501031935215, + 0.10656525194644928, + 0.07006782293319702, + 0.11720187216997147, + 0.11699124425649643, + 0.06657670438289642, + 0.09980203956365585, + 0.13386592268943787, + 0.2175857424736023, + 0.021893272176384926 + ], + "credit_map": [ + 0.00807053130120039, + 0.005524372681975365, + 0.009053451009094715, + 0.0051100365817546844, + 0.004319849889725447, + 0.004952185321599245, + 0.005213906988501549, + 0.007205768954008818, + 0.010288220830261707, + 0.01945366896688938, + 0.017731621861457825, + 0.014476126059889793, + 0.007352818734943867, + 0.006599039770662785, + 0.01597628742456436, + 0.014687458984553814, + 0.004717410076409578, + 0.008005400188267231, + 0.006749514956027269, + 0.008991795592010021, + 0.010359887033700943, + 0.012002606876194477, + 0.012298926711082458, + 0.006942283362150192, + 0.007760112639516592, + 0.009651300497353077, + 0.015441042371094227, + 0.010378584265708923, + 0.014203551225364208, + 0.01078119408339262, + 0.024988004937767982, + 0.02399808168411255, + 0.008955045603215694, + 0.027965188026428223, + 0.022310808300971985, + 0.009503175504505634, + 0.013234549202024937, + 0.0081150708720088, + 0.014748016372323036, + 0.012159392237663269, + 0.009922412224113941, + 0.013678035698831081, + 0.013635273091495037, + 0.007100396789610386, + 0.005112245678901672, + 0.005084889940917492, + 0.005904708057641983, + 0.006138747092336416, + 0.010281420312821865, + 0.010875233449041843, + 0.011616403236985207, + 0.011989735998213291, + 0.012553996406495571, + 0.010213259607553482, + 0.014181657694280148, + 0.01809033192694187, + 0.007772702723741531, + 0.008415671065449715, + 0.018092241138219833, + 0.011242363601922989, + 0.028113817796111107, + 0.03440622612833977, + 0.015214688144624233, + 0.010608207434415817, + 0.008155412040650845, + 0.008590089157223701, + 0.00986410491168499, + 0.004914111457765102, + 0.011677485890686512, + 0.08819928020238876, + 0.01247559953480959, + 0.07370967417955399, + 0.002033667406067252, + 0.0019033096032217145, + 0.0025613782927393913, + 0.002704569837078452, + 0.0026705709751695395, + 0.0022092286963015795, + 0.005545813590288162, + 0.002264732960611582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.10031377524137497 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987461388111115 + ], + "opposite_logprob_tokens": [ + -1.8048746585845947 + ], + "logprob_reward": -0.17987461388111115, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987461388111115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.019415926188230515, + 0.0091548478230834, + 0.015759294852614403, + 0.023649225011467934, + 0.024374108761548996, + 0.0095659289509058, + 0.037163350731134415, + 0.42940011620521545, + 0.41924604773521423, + 0.012271068058907986 + ], + "credit_map": [ + 0.004149881191551685, + 0.0015878300182521343, + 0.002936256118118763, + 0.0030936989933252335, + 0.0015363371931016445, + 0.0014919735258445144, + 0.002284201793372631, + 0.0023357486352324486, + 0.0008476368966512382, + 0.0013054626761004329, + 0.0016881654737517238, + 0.001013082219287753, + 0.0013288367772474885, + 0.0006518890731967986, + 0.0014292174018919468, + 0.0008905583526939154, + 0.0013836462749168277, + 0.002063405467197299, + 0.0015106979990378022, + 0.0028921926859766245, + 0.0016998263308778405, + 0.001885708887130022, + 0.0023640766739845276, + 0.001959741348400712, + 0.002285454422235489, + 0.0017634732648730278, + 0.002421943936496973, + 0.002833406673744321, + 0.002320572966709733, + 0.0019634487107396126, + 0.0027396196965128183, + 0.007321307901293039, + 0.0018592535052448511, + 0.005113683640956879, + 0.0052017634734511375, + 0.0022939047776162624, + 0.002773950807750225, + 0.001967821503058076, + 0.002320090075954795, + 0.0028436414431780577, + 0.0016992356395348907, + 0.001613197149708867, + 0.0014859098009765148, + 0.000889312825165689, + 0.000972236564848572, + 0.0011634106049314141, + 0.0009235714678652585, + 0.0008190564112737775, + 0.002633356722071767, + 0.0036824073176831007, + 0.004797723144292831, + 0.004245648626238108, + 0.0038004859816282988, + 0.004098239820450544, + 0.004371020942926407, + 0.009534470736980438, + 0.026621589437127113, + 0.016161929816007614, + 0.08417309075593948, + 0.046325452625751495, + 0.03247613087296486, + 0.1534414440393448, + 0.03574662655591965, + 0.034453924745321274, + 0.01236664317548275, + 0.014990917406976223, + 0.016259049996733665, + 0.007632663939148188, + 0.012086115777492523, + 0.181752547621727, + 0.023209115490317345, + 0.15094906091690063, + 0.001163951470516622, + 0.0014026392018422484, + 0.0015172167913988233, + 0.0015518659492954612, + 0.0020033454056829214, + 0.001371506485156715, + 0.0014063938288018107, + 0.0018541509052738547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.17987461388111115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with Ghibli animation featuring soft hand-painted textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a colorful lizard with an emerald green body and turquoise speckles along its spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard pause on a large fern leaf to rhythmically expand prominent translucent scarlet dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a smaller golden lizard climbing a dense twisting vine in pursuit of the first lizard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute continuous smooth push-ins gliding through the dense foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029900074005127 + ], + "opposite_logprob_tokens": [ + -2.3502988815307617 + ], + "logprob_reward": -0.10029900074005127, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029900074005127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060360271483659744, + 0.12580808997154236, + 0.07877758145332336, + 0.14652019739151, + 0.08457976579666138, + 0.060830093920230865, + 0.11203395575284958, + 0.07175339013338089, + 0.23052963614463806, + 0.028807012364268303 + ], + "credit_map": [ + 0.007937041111290455, + 0.005829832050949335, + 0.00980314239859581, + 0.007927275262773037, + 0.005537118762731552, + 0.005198705941438675, + 0.010398021899163723, + 0.007729133125394583, + 0.01327951718121767, + 0.022011734545230865, + 0.022088319063186646, + 0.016916396096348763, + 0.010087581351399422, + 0.0077200885862112045, + 0.018609365448355675, + 0.015095080249011517, + 0.008344794623553753, + 0.006277026142925024, + 0.009344945661723614, + 0.013815386220812798, + 0.004939238075166941, + 0.010391886346042156, + 0.015643227845430374, + 0.010021078400313854, + 0.011178787797689438, + 0.01297020073980093, + 0.0156262069940567, + 0.016503727063536644, + 0.017053987830877304, + 0.012775477953255177, + 0.02786337584257126, + 0.03254842758178711, + 0.011434271931648254, + 0.014850134029984474, + 0.008886730298399925, + 0.00880111288279295, + 0.01158808171749115, + 0.005904919933527708, + 0.013080897741019726, + 0.010033614933490753, + 0.009002188220620155, + 0.006445914972573519, + 0.01014496386051178, + 0.007586649619042873, + 0.006781873758882284, + 0.005921360105276108, + 0.00790262222290039, + 0.007044521160423756, + 0.009660918265581131, + 0.011390985921025276, + 0.01332983747124672, + 0.017498577013611794, + 0.016701573505997658, + 0.012036514468491077, + 0.01584528759121895, + 0.015570273622870445, + 0.005118883214890957, + 0.004194449167698622, + 0.008439543657004833, + 0.009333258494734764, + 0.0177314430475235, + 0.006145295221358538, + 0.012844165787100792, + 0.007946362718939781, + 0.010355517268180847, + 0.014197155833244324, + 0.011034561321139336, + 0.007376369554549456, + 0.016116267070174217, + 0.08619008958339691, + 0.015585704706609249, + 0.06967397034168243, + 0.003197796642780304, + 0.003779851831495762, + 0.004248351790010929, + 0.0038711715023964643, + 0.003648822894319892, + 0.003277209587395191, + 0.004320426378399134, + 0.0024633826687932014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 1 + ], + "selected_reward": -0.10029900074005127 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_8/index_tensor([8207, 8207])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_8/index_tensor([8207, 8207])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..690d6dd7bb39294bcfd8005b9162690d4c811aa9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_8/index_tensor([8207, 8207])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extremely high-speed macro shot, maintaining a static, top-down bird's-eye view, captures a stark black-and-white digital topographic map, where every topographic detail is rendered with astonishing clarity. The map is a complex network of extremely thin, precise contour lines and gridded elevation data set against a pure black background. Luminous bands, representing geological hazards, initially erupt in specific high-risk areas, flashing frantically, intermittently with bright white light. Their chaotic pulses cast fleeting, sharp highlights across the intricate digital geography. Then, as the high-speed effect seems to slow, the action shifts dramatically; the frantic blinking decelerates into a deliberate, rhythmic throbbing, with extended pauses between each pulse creating an intense, suspenseful atmosphere. Eventually, the pulsation ceases entirely, and the areas transform into an ominous, unwavering, steady glow. Their harsh white light now constantly illuminates the high-risk areas, creating an eerie scene that showcases a concentrated, imminent threat against the cold, analytically precise map.", + "seed": 5753, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_8/index_8207_seed_5753_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10614864751696587, + "vlm_logprob": -0.10614864751696587, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.10614864751696587, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06609572470188141 + ], + "opposite_logprob_tokens": [ + -3.1910958290100098 + ], + "logprob_reward": -0.06609572470188141, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.06609572470188141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31942057609558105, + 0.17219512164592743, + 0.05847692862153053, + 0.025868788361549377, + 0.022088568657636642, + 0.037656113505363464, + 0.1882098764181137, + 0.029413774609565735, + 0.1148640364408493, + 0.03180631995201111 + ], + "credit_map": [ + 0.024909283965826035, + 0.040056657046079636, + 0.03816509246826172, + 0.0709039494395256, + 0.024022072553634644, + 0.028640730306506157, + 0.026779066771268845, + 0.06594371050596237, + 0.009426741860806942, + 0.015906522050499916, + 0.015548422932624817, + 0.013130967505276203, + 0.03469303622841835, + 0.020525820553302765, + 0.04129456728696823, + 0.021669035777449608, + 0.006928053684532642, + 0.0048226178623735905, + 0.00519888149574399, + 0.004800576251000166, + 0.008861765265464783, + 0.00833076797425747, + 0.01112087070941925, + 0.00841339398175478, + 0.0027764257974922657, + 0.0026493696495890617, + 0.0025418815203011036, + 0.002294969279319048, + 0.00359952705912292, + 0.003389457007870078, + 0.0055038235150277615, + 0.003113333834335208, + 0.00296576926484704, + 0.0022651012986898422, + 0.002292500576004386, + 0.0021095206029713154, + 0.002745949663221836, + 0.002698511118069291, + 0.003896251553669572, + 0.003114964347332716, + 0.002712530316784978, + 0.0026657218113541603, + 0.0025600201915949583, + 0.002246457152068615, + 0.0062378146685659885, + 0.008070793002843857, + 0.0057219890877604485, + 0.007440787274390459, + 0.014199976809322834, + 0.009175162762403488, + 0.005930510349571705, + 0.005112506914883852, + 0.013322573155164719, + 0.06594640016555786, + 0.01617056503891945, + 0.05835217237472534, + 0.0029866474214941263, + 0.0031200749799609184, + 0.003183678025379777, + 0.002230802085250616, + 0.003644700860604644, + 0.005108978599309921, + 0.005177113693207502, + 0.003961778711527586, + 0.0076960353180766106, + 0.010985009372234344, + 0.009037898853421211, + 0.008540976792573929, + 0.01132170855998993, + 0.021718209609389305, + 0.02458212338387966, + 0.020982077345252037, + 0.002509588608518243, + 0.0020585115998983383, + 0.008677011355757713, + 0.0020983577705919743, + 0.0027763040270656347, + 0.004292578902095556, + 0.00510410824790597, + 0.004289858043193817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.06609572470188141 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34872421622276306 + ], + "opposite_logprob_tokens": [ + -1.2237242460250854 + ], + "logprob_reward": -0.34872421622276306, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34872421622276306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09543333947658539, + 0.12945228815078735, + 0.14977671205997467, + 0.06160987541079521, + 0.1685059517621994, + 0.08426210284233093, + 0.06163261830806732, + 0.1181095540523529, + 0.11448100954294205, + 0.016736486926674843 + ], + "credit_map": [ + 0.0077067529782652855, + 0.011156149208545685, + 0.006746790371835232, + 0.019199516624212265, + 0.010721972212195396, + 0.010188773274421692, + 0.011667542159557343, + 0.018045838922262192, + 0.012942507863044739, + 0.012538071721792221, + 0.012306885793805122, + 0.01302668172866106, + 0.01617114059627056, + 0.011260691098868847, + 0.03082181327044964, + 0.020384490489959717, + 0.01621636375784874, + 0.014367367140948772, + 0.0147384162992239, + 0.01228247955441475, + 0.020302776247262955, + 0.016704509034752846, + 0.035631828010082245, + 0.019532974809408188, + 0.005602021235972643, + 0.0050967237912118435, + 0.0036416389048099518, + 0.004066662397235632, + 0.01081685908138752, + 0.007390168029814959, + 0.01795761100947857, + 0.007038189098238945, + 0.013067435473203659, + 0.01386250089854002, + 0.009914206340909004, + 0.00991922989487648, + 0.034961942583322525, + 0.016535557806491852, + 0.05078356713056564, + 0.019461508840322495, + 0.008932100608944893, + 0.00802585668861866, + 0.005901657976210117, + 0.004607862327247858, + 0.010742931626737118, + 0.01670648530125618, + 0.019862661138176918, + 0.009482549503445625, + 0.005529181566089392, + 0.002973987255245447, + 0.0023329821415245533, + 0.0023715184070169926, + 0.0044547622092068195, + 0.018467269837856293, + 0.0068674287758767605, + 0.01863548904657364, + 0.011684546247124672, + 0.01568262279033661, + 0.005892663262784481, + 0.0091019868850708, + 0.007049138192087412, + 0.030153334140777588, + 0.01820363849401474, + 0.020341629162430763, + 0.011033209972083569, + 0.013284803368151188, + 0.008251008577644825, + 0.009071575477719307, + 0.009849313646554947, + 0.02213715948164463, + 0.025180596858263016, + 0.01567334309220314, + 0.0014825621619820595, + 0.001600322313606739, + 0.002229170873761177, + 0.0012178078759461641, + 0.0026039837393909693, + 0.002056705066934228, + 0.0028292816132307053, + 0.002716653747484088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.34872421622276306 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029825396835803986 + ], + "opposite_logprob_tokens": [ + -3.529825448989868 + ], + "logprob_reward": -0.029825396835803986, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029825396835803986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3713045120239258, + 0.12862148880958557, + 0.040519192814826965, + 0.03264157474040985, + 0.053768012672662735, + 0.0642583891749382, + 0.07572977989912033, + 0.10145457088947296, + 0.06147610768675804, + 0.07022644579410553 + ], + "credit_map": [ + 0.033399395644664764, + 0.0367179773747921, + 0.041689254343509674, + 0.05284350365400314, + 0.044870901852846146, + 0.06156633049249649, + 0.04477207362651825, + 0.05544507876038551, + 0.01364107709378004, + 0.01600460708141327, + 0.011821780353784561, + 0.011790260672569275, + 0.022442618384957314, + 0.015332521870732307, + 0.022979674860835075, + 0.014608929865062237, + 0.004709810018539429, + 0.005705004092305899, + 0.004282788373529911, + 0.003942995332181454, + 0.004458604846149683, + 0.004900229629129171, + 0.007918461225926876, + 0.0046012988314032555, + 0.0030380571261048317, + 0.002652982948347926, + 0.0020089128520339727, + 0.00214551854878664, + 0.005901259835809469, + 0.003550712252035737, + 0.008756033144891262, + 0.004588099662214518, + 0.0036452580243349075, + 0.005987243261188269, + 0.0034159994684159756, + 0.0035594685468822718, + 0.011875391006469727, + 0.006040364038199186, + 0.012875229120254517, + 0.006369059905409813, + 0.008324913680553436, + 0.00468704616650939, + 0.004375573713332415, + 0.0033649925608187914, + 0.007020018994808197, + 0.012659498490393162, + 0.01221082266420126, + 0.011615529656410217, + 0.005818819161504507, + 0.005241686478257179, + 0.004756900481879711, + 0.003784524742513895, + 0.006258509121835232, + 0.022797714918851852, + 0.0074992128647863865, + 0.019572412595152855, + 0.011053056456148624, + 0.01420078705996275, + 0.007419116795063019, + 0.007897811941802502, + 0.008076895028352737, + 0.025620639324188232, + 0.012882189825177193, + 0.014304075390100479, + 0.007878536358475685, + 0.0071610030718147755, + 0.006730144377797842, + 0.005438897293061018, + 0.00774647481739521, + 0.009068033657968044, + 0.01087889913469553, + 0.00657412176951766, + 0.007522377651184797, + 0.005315449088811874, + 0.01928442344069481, + 0.005460232030600309, + 0.005460937507450581, + 0.008491183631122112, + 0.011761779896914959, + 0.00693006394430995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.029825396835803986 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073528677225113 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073528677225113, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073528677225113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32041308283805847, + 0.140333354473114, + 0.06382200121879578, + 0.06431828439235687, + 0.050628162920475006, + 0.06058766692876816, + 0.05651111900806427, + 0.04410040006041527, + 0.08676450699567795, + 0.11252148449420929 + ], + "credit_map": [ + 0.032318055629730225, + 0.043773360550403595, + 0.034463513642549515, + 0.06852904707193375, + 0.03205764293670654, + 0.022042181342840195, + 0.03155256807804108, + 0.055676717311143875, + 0.015757275745272636, + 0.014000781811773777, + 0.014044308103621006, + 0.012964175082743168, + 0.02112201415002346, + 0.01597074419260025, + 0.029403381049633026, + 0.017070654779672623, + 0.006482336204499006, + 0.006712034810334444, + 0.007668226025998592, + 0.005263250321149826, + 0.008078747428953648, + 0.008962669409811497, + 0.012255056761205196, + 0.008399678394198418, + 0.006419685203582048, + 0.007977074012160301, + 0.004688049200922251, + 0.004817636217921972, + 0.011689579114317894, + 0.006916585378348827, + 0.01439069863408804, + 0.007418970111757517, + 0.0046159569174051285, + 0.004606442991644144, + 0.0046464684419333935, + 0.004073801916092634, + 0.00849473848938942, + 0.005464542657136917, + 0.011488174088299274, + 0.007238033693283796, + 0.0070919012650847435, + 0.004900452680885792, + 0.003821236314252019, + 0.0037259263917803764, + 0.007501316722482443, + 0.014352775178849697, + 0.011171103455126286, + 0.008022950030863285, + 0.005159400403499603, + 0.0037587631959468126, + 0.003691575024276972, + 0.00258058775216341, + 0.004553645849227905, + 0.017005236819386482, + 0.006234006956219673, + 0.013527903705835342, + 0.004711993038654327, + 0.006737410556524992, + 0.004875910468399525, + 0.003953211009502411, + 0.0036480389535427094, + 0.009615027345716953, + 0.0043893237598240376, + 0.006169477477669716, + 0.006909444462507963, + 0.009571945294737816, + 0.007643460761755705, + 0.006058591417968273, + 0.006969204638153315, + 0.01930665597319603, + 0.016776664182543755, + 0.013528531417250633, + 0.010844549164175987, + 0.009010011330246925, + 0.028537042438983917, + 0.00815004762262106, + 0.008042462170124054, + 0.016106033697724342, + 0.021818947046995163, + 0.010012377053499222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.016073528677225113 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002437114715576 + ], + "opposite_logprob_tokens": [ + -2.6950244903564453 + ], + "logprob_reward": -0.07002437114715576, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002437114715576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3438131809234619, + 0.13149729371070862, + 0.05068656802177429, + 0.03819747641682625, + 0.06118280440568924, + 0.056005727499723434, + 0.08120028674602509, + 0.045592740178108215, + 0.09268408268690109, + 0.09913980960845947 + ], + "credit_map": [ + 0.03340211138129234, + 0.04636608064174652, + 0.03988460451364517, + 0.06429849565029144, + 0.03629791736602783, + 0.028380220755934715, + 0.03971012681722641, + 0.05547361820936203, + 0.01666700467467308, + 0.01584494300186634, + 0.011884568259119987, + 0.012972130440175533, + 0.018922029063105583, + 0.013431761413812637, + 0.023025143891572952, + 0.018749726936221123, + 0.005597556009888649, + 0.005203999578952789, + 0.006353328935801983, + 0.004544674884527922, + 0.0052808476611971855, + 0.0053377700969576836, + 0.010148532688617706, + 0.008219858631491661, + 0.002563071670010686, + 0.0034419745206832886, + 0.002911703195422888, + 0.0030332128517329693, + 0.008339309133589268, + 0.005357297137379646, + 0.007725656498223543, + 0.004825253039598465, + 0.005435682367533445, + 0.0073362961411476135, + 0.005216856021434069, + 0.00539507158100605, + 0.0094964150339365, + 0.005941919516772032, + 0.01131975557655096, + 0.011040804907679558, + 0.005482490640133619, + 0.0038657491095364094, + 0.0031556596513837576, + 0.003082620445638895, + 0.006957699544727802, + 0.012508138082921505, + 0.012252148240804672, + 0.008701221086084843, + 0.006224936805665493, + 0.005065847188234329, + 0.003341102506965399, + 0.0033911496866494417, + 0.006662454456090927, + 0.0255042165517807, + 0.007359351497143507, + 0.023651229217648506, + 0.004898815415799618, + 0.007144408766180277, + 0.0037669786252081394, + 0.0042031435295939445, + 0.00406857393682003, + 0.0077112107537686825, + 0.007321873214095831, + 0.006477737799286842, + 0.010606617666780949, + 0.011176464147865772, + 0.007618092466145754, + 0.007224096916615963, + 0.007646982092410326, + 0.020865868777036667, + 0.01807525008916855, + 0.009470717050135136, + 0.01059779990464449, + 0.008575653657317162, + 0.025925813242793083, + 0.00828994158655405, + 0.007675123400986195, + 0.012057694606482983, + 0.016831528395414352, + 0.00918624922633171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.07002437114715576 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely high-speed macro shot, maintaining a static, top-down bird's-eye view, captures a stark black-and-white digital topographic map, where every topographic detail is rendered with astonishing clarity. The map is a complex network of extremely thin, precise contour lines and gridded elevation data set against a pure black background. Luminous bands, representing geological hazards, initially erupt in specific high-risk areas, flashing frantically, intermittently with bright white light. Their chaotic pulses cast fleeting, sharp highlights across the intricate digital geography. Then, as the high-speed effect seems to slow, the action shifts dramatically; the frantic blinking decelerates into a deliberate, rhythmic throbbing, with extended pauses between each pulse creating an intense, suspenseful atmosphere. Eventually, the pulsation ceases entirely, and the areas transform into an ominous, unwavering, steady glow. Their harsh white light now constantly illuminates the high-risk areas, creating an eerie scene that showcases a concentrated, imminent threat against the cold, analytically precise map.", + "seed": 5754, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_8/index_8207_seed_5754_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1306284461170435, + "vlm_logprob": -0.1306284461170435, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.1306284461170435, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03868122398853302 + ], + "opposite_logprob_tokens": [ + -3.9136812686920166 + ], + "logprob_reward": -0.03868122398853302, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03868122398853302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0808720737695694, + 0.030727336183190346, + 0.07780002802610397, + 0.05840897560119629, + 0.11565515398979187, + 0.13576078414916992, + 0.21163232624530792, + 0.10363024473190308, + 0.09162957966327667, + 0.0938834697008133 + ], + "credit_map": [ + 0.009274270385503769, + 0.009202519431710243, + 0.009695173241198063, + 0.010696029290556908, + 0.011129402555525303, + 0.011523548513650894, + 0.010578406974673271, + 0.008772729896008968, + 0.004179357551038265, + 0.002977854572236538, + 0.003008700907230377, + 0.004056443460285664, + 0.003725710790604353, + 0.003920210059732199, + 0.004283461254090071, + 0.004575597122311592, + 0.006220423616468906, + 0.0064830053597688675, + 0.004725891165435314, + 0.010613481514155865, + 0.01344432681798935, + 0.009240688756108284, + 0.0188803318887949, + 0.008191883563995361, + 0.006051848642528057, + 0.00775414053350687, + 0.004677802789956331, + 0.009912249632179737, + 0.009647561237215996, + 0.006711952388286591, + 0.0066083818674087524, + 0.0070450399070978165, + 0.009366016834974289, + 0.013215955346822739, + 0.0094087990000844, + 0.019263243302702904, + 0.027467194944620132, + 0.015919040888547897, + 0.008823586627840996, + 0.01219131425023079, + 0.016458233818411827, + 0.012534257024526596, + 0.012501559220254421, + 0.014812499284744263, + 0.03728801757097244, + 0.022444505244493484, + 0.010709457099437714, + 0.0090122539550066, + 0.04357898235321045, + 0.019783485680818558, + 0.018348483368754387, + 0.028634799644351006, + 0.042506664991378784, + 0.02773142419755459, + 0.01700112223625183, + 0.014047367498278618, + 0.021162746474146843, + 0.00891788024455309, + 0.01123676635324955, + 0.007203933782875538, + 0.02527313120663166, + 0.016713334247469902, + 0.006688496097922325, + 0.006433958187699318, + 0.015275430865585804, + 0.0083558838814497, + 0.011869187466800213, + 0.007667027413845062, + 0.016512984409928322, + 0.009724008850753307, + 0.009618096053600311, + 0.012606961652636528, + 0.018008891493082047, + 0.004816677421331406, + 0.018599921837449074, + 0.014296358451247215, + 0.01436945516616106, + 0.01041671633720398, + 0.0064326003193855286, + 0.006942846812307835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.03868122398853302 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233603715896606 + ], + "opposite_logprob_tokens": [ + -0.8983603715896606 + ], + "logprob_reward": -0.5233603715896606, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233603715896606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0966687947511673, + 0.09178087115287781, + 0.019158637151122093, + 0.048635706305503845, + 0.30464616417884827, + 0.055847737938165665, + 0.08623643219470978, + 0.05001849681138992, + 0.06637749075889587, + 0.18062961101531982 + ], + "credit_map": [ + 0.010550834238529205, + 0.009754553437232971, + 0.011724263429641724, + 0.01578492484986782, + 0.010744644328951836, + 0.014598659239709377, + 0.01515717152506113, + 0.008353754878044128, + 0.010235929861664772, + 0.008242757990956306, + 0.006894570775330067, + 0.013044345192611217, + 0.015737421810626984, + 0.01327914372086525, + 0.01581682078540325, + 0.008529888466000557, + 0.0016127751441672444, + 0.001840149867348373, + 0.001498389057815075, + 0.0028315510135143995, + 0.0031540626659989357, + 0.0022695004008710384, + 0.0040293047204613686, + 0.001922908122651279, + 0.004973314236849546, + 0.005507683847099543, + 0.0044350954703986645, + 0.005731048062443733, + 0.006799106486141682, + 0.0063134427182376385, + 0.006936257239431143, + 0.00793976616114378, + 0.03362930193543434, + 0.03053589165210724, + 0.02055901475250721, + 0.04190846160054207, + 0.0596628375351429, + 0.04005640372633934, + 0.03274590149521828, + 0.045548390597105026, + 0.007519925478845835, + 0.005685696844011545, + 0.005198636092245579, + 0.0079255485907197, + 0.009370315819978714, + 0.007381959352642298, + 0.0059629022143781185, + 0.006802758667618036, + 0.02232261560857296, + 0.009879963472485542, + 0.007603438571095467, + 0.009917966090142727, + 0.01644357666373253, + 0.010568900965154171, + 0.00511165102943778, + 0.004388324450701475, + 0.008083069697022438, + 0.004462670534849167, + 0.008508037775754929, + 0.005350361578166485, + 0.011161386035382748, + 0.0068749249912798405, + 0.0028832305688410997, + 0.0026948237791657448, + 0.012401511892676353, + 0.005136504769325256, + 0.006210290361195803, + 0.00643868837505579, + 0.013021533377468586, + 0.007524279877543449, + 0.006602611858397722, + 0.009042078629136086, + 0.02234683372080326, + 0.013253573328256607, + 0.028379030525684357, + 0.03620840981602669, + 0.0323106087744236, + 0.02007799781858921, + 0.015148870646953583, + 0.012904307805001736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.5233603715896606 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636721171438694 + ], + "opposite_logprob_tokens": [ + -3.6513671875 + ], + "logprob_reward": -0.02636721171438694, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636721171438694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0881066545844078, + 0.09445920586585999, + 0.06515724956989288, + 0.05809581279754639, + 0.142815500497818, + 0.08211914449930191, + 0.08511769026517868, + 0.1485549807548523, + 0.15496307611465454, + 0.08061059564352036 + ], + "credit_map": [ + 0.008926194161176682, + 0.009600910358130932, + 0.01650398224592209, + 0.008626282215118408, + 0.010814115405082703, + 0.013407534919679165, + 0.009722433984279633, + 0.010505203157663345, + 0.009632885456085205, + 0.010697077959775925, + 0.010151106864213943, + 0.012249338440597057, + 0.013103912584483624, + 0.013082627207040787, + 0.012876586988568306, + 0.012665669433772564, + 0.00491853104904294, + 0.007030027452856302, + 0.006586835719645023, + 0.009180935099720955, + 0.008245852775871754, + 0.006549711339175701, + 0.014433249831199646, + 0.008212108165025711, + 0.00496436282992363, + 0.005946472752839327, + 0.003858223557472229, + 0.011687823571264744, + 0.005059509072452784, + 0.007077057380229235, + 0.00716141564771533, + 0.012340949848294258, + 0.019509540870785713, + 0.015134154818952084, + 0.010536406189203262, + 0.026608753949403763, + 0.021794168278574944, + 0.01758725568652153, + 0.01387922652065754, + 0.017765987664461136, + 0.010422948747873306, + 0.007550652138888836, + 0.007141255307942629, + 0.010544145479798317, + 0.021006979048252106, + 0.011889033950865269, + 0.006902587600052357, + 0.006661537103354931, + 0.010678037069737911, + 0.008196080103516579, + 0.010181006975471973, + 0.0122109679505229, + 0.021303627640008926, + 0.01330878958106041, + 0.004778829403221607, + 0.004460352938622236, + 0.04101458936929703, + 0.011789939366281033, + 0.019322838634252548, + 0.012823107652366161, + 0.02806721068918705, + 0.018589159473776817, + 0.008279911242425442, + 0.008668240159749985, + 0.038557615131139755, + 0.014802518300712109, + 0.02058708481490612, + 0.014483772218227386, + 0.022909021005034447, + 0.019094929099082947, + 0.009979470632970333, + 0.014548670500516891, + 0.009234323166310787, + 0.006928954739123583, + 0.010829497128725052, + 0.010884346440434456, + 0.013834082521498203, + 0.010422159917652607, + 0.008344171568751335, + 0.01013305876404047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.02636721171438694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073763370513916 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073763370513916, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073763370513916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10576488822698593, + 0.05396490916609764, + 0.049441851675510406, + 0.06227300688624382, + 0.1422288864850998, + 0.10123259574174881, + 0.09807536751031876, + 0.1344454288482666, + 0.1459624022245407, + 0.10661066323518753 + ], + "credit_map": [ + 0.013275638222694397, + 0.01131565123796463, + 0.012836050242185593, + 0.012972277589142323, + 0.01459900289773941, + 0.015193542465567589, + 0.01513710804283619, + 0.01043561939150095, + 0.0064649879932403564, + 0.005388699006289244, + 0.0040201484225690365, + 0.007960357703268528, + 0.008342500776052475, + 0.007570282090455294, + 0.009051339700818062, + 0.0051665958017110825, + 0.005192456301301718, + 0.004982260055840015, + 0.0035441452637314796, + 0.0059713516384363174, + 0.007580535486340523, + 0.005498649086803198, + 0.011241912841796875, + 0.005430541932582855, + 0.005126953590661287, + 0.006903504021465778, + 0.004378602374345064, + 0.011599713936448097, + 0.006732361391186714, + 0.007341289892792702, + 0.008290614932775497, + 0.011899970471858978, + 0.023227179422974586, + 0.015351276844739914, + 0.009762205183506012, + 0.019572796300053596, + 0.02631201408803463, + 0.019686520099639893, + 0.01298952754586935, + 0.015327364206314087, + 0.00943557359278202, + 0.009296138770878315, + 0.00918211042881012, + 0.02008134499192238, + 0.02127002738416195, + 0.01307421550154686, + 0.009513146243989468, + 0.009380041621625423, + 0.01252246554940939, + 0.0113909300416708, + 0.008197683840990067, + 0.008482166565954685, + 0.026077650487422943, + 0.016160303726792336, + 0.007951714098453522, + 0.007292461581528187, + 0.03207816928625107, + 0.010030127130448818, + 0.02149556577205658, + 0.011474046856164932, + 0.028007080778479576, + 0.014581243507564068, + 0.00784672237932682, + 0.008932482451200485, + 0.02448118105530739, + 0.013864453881978989, + 0.01824423298239708, + 0.01092680636793375, + 0.0308555718511343, + 0.018399689346551895, + 0.014117367565631866, + 0.015073100104928017, + 0.010881956666707993, + 0.008791246451437473, + 0.014896444045007229, + 0.014983139932155609, + 0.019672689959406853, + 0.013577829115092754, + 0.010842147283256054, + 0.012965217232704163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.016073763370513916 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048659659922122955 + ], + "opposite_logprob_tokens": [ + -3.048659563064575 + ], + "logprob_reward": -0.048659659922122955, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048659659922122955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0728568285703659, + 0.05898241326212883, + 0.05112627148628235, + 0.04533939063549042, + 0.1455106884241104, + 0.06516020745038986, + 0.08865166455507278, + 0.1899077296257019, + 0.13204824924468994, + 0.1504165232181549 + ], + "credit_map": [ + 0.009334093891084194, + 0.008482865989208221, + 0.011115447618067265, + 0.007571027614176273, + 0.008516961708664894, + 0.010571226477622986, + 0.008029687218368053, + 0.009235515259206295, + 0.006392489653080702, + 0.005661935079842806, + 0.004910181742161512, + 0.009343662299215794, + 0.008539964444935322, + 0.008292410522699356, + 0.00915620382875204, + 0.006685566622763872, + 0.004879573360085487, + 0.005082300864160061, + 0.00498437462374568, + 0.006849365308880806, + 0.00675215432420373, + 0.005434808321297169, + 0.010574644431471825, + 0.006569051183760166, + 0.003269995329901576, + 0.0051073916256427765, + 0.0024266145192086697, + 0.009310849942266941, + 0.007817442528903484, + 0.005157419014722109, + 0.006672970484942198, + 0.005576709751039743, + 0.014342766255140305, + 0.013871007598936558, + 0.010610121302306652, + 0.023122157901525497, + 0.0281373281031847, + 0.018914591521024704, + 0.015537967905402184, + 0.020974744111299515, + 0.007258105091750622, + 0.007111346814781427, + 0.006696746218949556, + 0.01049090176820755, + 0.013185680843889713, + 0.008275095373392105, + 0.005909600295126438, + 0.006232728250324726, + 0.024906311184167862, + 0.011768903583288193, + 0.00899013876914978, + 0.006221682298928499, + 0.014986129477620125, + 0.01006090734153986, + 0.0055665443651378155, + 0.0061510419473052025, + 0.04217983037233353, + 0.013765386305749416, + 0.03107471950352192, + 0.01804991625249386, + 0.04253721609711647, + 0.021962231025099754, + 0.010290303267538548, + 0.010048133321106434, + 0.024225130677223206, + 0.012755878269672394, + 0.01570315472781658, + 0.013722728937864304, + 0.02218375727534294, + 0.016042223200201988, + 0.011788924224674702, + 0.01562645100057125, + 0.015176585875451565, + 0.012725849635899067, + 0.026290569454431534, + 0.026312613859772682, + 0.021871505305171013, + 0.016807138919830322, + 0.014731349423527718, + 0.01650090701878071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.048659659922122955 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely high-speed macro shot, maintaining a static, top-down bird's-eye view, captures a stark black-and-white digital topographic map, where every topographic detail is rendered with astonishing clarity. The map is a complex network of extremely thin, precise contour lines and gridded elevation data set against a pure black background. Luminous bands, representing geological hazards, initially erupt in specific high-risk areas, flashing frantically, intermittently with bright white light. Their chaotic pulses cast fleeting, sharp highlights across the intricate digital geography. Then, as the high-speed effect seems to slow, the action shifts dramatically; the frantic blinking decelerates into a deliberate, rhythmic throbbing, with extended pauses between each pulse creating an intense, suspenseful atmosphere. Eventually, the pulsation ceases entirely, and the areas transform into an ominous, unwavering, steady glow. Their harsh white light now constantly illuminates the high-risk areas, creating an eerie scene that showcases a concentrated, imminent threat against the cold, analytically precise map.", + "seed": 5755, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_8/index_8207_seed_5755_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03537019323557615, + "vlm_logprob": -0.03537019323557615, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.03537019323557615, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04912952333688736 + ], + "opposite_logprob_tokens": [ + -3.2991294860839844 + ], + "logprob_reward": -0.04912952333688736, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04912952333688736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.245370551943779, + 0.06755470484495163, + 0.10152865946292877, + 0.11479629576206207, + 0.10980985313653946, + 0.0569198913872242, + 0.16777555644512177, + 0.05203448608517647, + 0.03910438343882561, + 0.04510558769106865 + ], + "credit_map": [ + 0.01947491243481636, + 0.028392687439918518, + 0.024226155132055283, + 0.014815254136919975, + 0.06000632792711258, + 0.05495549365878105, + 0.020912054926156998, + 0.02258768491446972, + 0.01722824014723301, + 0.006126116495579481, + 0.01094070728868246, + 0.0043515171855688095, + 0.012007620185613632, + 0.007457925472408533, + 0.004654782358556986, + 0.004787804558873177, + 0.015954596921801567, + 0.010234972462058067, + 0.01743409037590027, + 0.008615613915026188, + 0.009920220822095871, + 0.02034548670053482, + 0.011489721946418285, + 0.007533961441367865, + 0.020276907831430435, + 0.007016177289187908, + 0.012726354412734509, + 0.00787181407213211, + 0.012533310800790787, + 0.03127993643283844, + 0.011781572364270687, + 0.011310226283967495, + 0.017820147797465324, + 0.006928619462996721, + 0.015054654330015182, + 0.0044977786019444466, + 0.011547267436981201, + 0.029564501717686653, + 0.017987024039030075, + 0.006409869994968176, + 0.008870941586792469, + 0.0041036298498511314, + 0.0047793081030249596, + 0.0035682206507772207, + 0.010109529830515385, + 0.00866096280515194, + 0.008122495375573635, + 0.008704807609319687, + 0.014624981209635735, + 0.012524078600108624, + 0.01026040967553854, + 0.0066402568481862545, + 0.04603631794452667, + 0.045620474964380264, + 0.016753828153014183, + 0.015315224416553974, + 0.003499968908727169, + 0.0036674579605460167, + 0.0031329994089901447, + 0.003494939533993602, + 0.007250415626913309, + 0.010009222663939, + 0.014952443540096283, + 0.0060270424000918865, + 0.004213182721287012, + 0.004361766390502453, + 0.0033231242559850216, + 0.003944528754800558, + 0.0043128421530127525, + 0.0052432590164244175, + 0.008175046183168888, + 0.005530637688934803, + 0.004579554311931133, + 0.0032376025337725878, + 0.005213891621679068, + 0.0037743619177490473, + 0.007861216552555561, + 0.007266300730407238, + 0.008352060802280903, + 0.004820602014660835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.04912952333688736 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029856175184249878 + ], + "opposite_logprob_tokens": [ + -3.5298562049865723 + ], + "logprob_reward": -0.029856175184249878, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029856175184249878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2574905455112457, + 0.10376198589801788, + 0.02281060628592968, + 0.13969078660011292, + 0.04805905744433403, + 0.0367279052734375, + 0.12244683504104614, + 0.19767822325229645, + 0.0481814369559288, + 0.023152627050876617 + ], + "credit_map": [ + 0.02755574882030487, + 0.045253485441207886, + 0.01864011399447918, + 0.013482400216162205, + 0.06378872692584991, + 0.05472220852971077, + 0.013826198875904083, + 0.020221656188368797, + 0.03281845524907112, + 0.007346751634031534, + 0.03815404698252678, + 0.004847791511565447, + 0.008818265981972218, + 0.00539887510240078, + 0.0031322326976805925, + 0.003245569532737136, + 0.0035125124268233776, + 0.0018406118033453822, + 0.003217252204194665, + 0.0017589684575796127, + 0.003329711966216564, + 0.004805437289178371, + 0.002179927658289671, + 0.002166184363886714, + 0.02072402462363243, + 0.010419594123959541, + 0.009900170378386974, + 0.007254843134433031, + 0.014227784238755703, + 0.03529474884271622, + 0.014025703072547913, + 0.027843927964568138, + 0.005857121665030718, + 0.0048453547060489655, + 0.005400184076279402, + 0.0023809613194316626, + 0.007940552197396755, + 0.008443593047559261, + 0.0073095280677080154, + 0.005881762132048607, + 0.004474164918065071, + 0.002676068339496851, + 0.0037772906944155693, + 0.0037040933966636658, + 0.0038513499312102795, + 0.006335161626338959, + 0.005343116354197264, + 0.0065666609443724155, + 0.012384722009301186, + 0.009102038107812405, + 0.009562005288898945, + 0.00582592748105526, + 0.034518320113420486, + 0.027496224269270897, + 0.013567240908741951, + 0.009990354999899864, + 0.014771482907235622, + 0.010686063207685947, + 0.008923538960516453, + 0.009461785666644573, + 0.057084135711193085, + 0.04577874019742012, + 0.03775951638817787, + 0.01321296114474535, + 0.004966930951923132, + 0.005026533734053373, + 0.004325398243963718, + 0.004644021391868591, + 0.008026196621358395, + 0.007071853149682283, + 0.00877300463616848, + 0.0053474982269108295, + 0.0036142084281891584, + 0.0017513911006972194, + 0.002415381371974945, + 0.0020985265728086233, + 0.0024326322600245476, + 0.0030625141225755215, + 0.004008847288787365, + 0.003769126022234559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.029856175184249878 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038126103579998016 + ], + "opposite_logprob_tokens": [ + -3.288125991821289 + ], + "logprob_reward": -0.038126103579998016, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038126103579998016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1588095873594284, + 0.24006082117557526, + 0.0853269100189209, + 0.05030861496925354, + 0.08071552962064743, + 0.06592880189418793, + 0.1297856569290161, + 0.10988467931747437, + 0.041339121758937836, + 0.03784030303359032 + ], + "credit_map": [ + 0.015299839898943901, + 0.02418355457484722, + 0.0164584219455719, + 0.011679312214255333, + 0.03389038145542145, + 0.029810717329382896, + 0.014850548468530178, + 0.012636820785701275, + 0.09092256426811218, + 0.010049305856227875, + 0.07976139336824417, + 0.008805236779153347, + 0.020965037867426872, + 0.014614907093346119, + 0.007110859267413616, + 0.007831512950360775, + 0.013472595252096653, + 0.005813377443701029, + 0.01919308491051197, + 0.00643488485366106, + 0.008547061122953892, + 0.015746423974633217, + 0.009573793970048428, + 0.006545690353959799, + 0.008647855371236801, + 0.0031833781395107508, + 0.006657230667769909, + 0.002847539959475398, + 0.005985565483570099, + 0.013650867156684399, + 0.004288857337087393, + 0.005047325044870377, + 0.009992324747145176, + 0.008281535468995571, + 0.006794431712478399, + 0.004069470800459385, + 0.02395755611360073, + 0.010690787807106972, + 0.009729254990816116, + 0.0072001684457063675, + 0.006663316860795021, + 0.004885279573500156, + 0.004524215590208769, + 0.003167728427797556, + 0.01699303276836872, + 0.009428583085536957, + 0.010528706945478916, + 0.009737937711179256, + 0.016030220314860344, + 0.008434454910457134, + 0.011031790636479855, + 0.006406522821635008, + 0.03805334493517876, + 0.03185463324189186, + 0.010064225643873215, + 0.007910462096333504, + 0.01267890352755785, + 0.007931053638458252, + 0.0071230242028832436, + 0.008370152674615383, + 0.01799403317272663, + 0.02019796147942543, + 0.023847952485084534, + 0.011741590686142445, + 0.0053899032063782215, + 0.004415672738105059, + 0.004867748823016882, + 0.003906667232513428, + 0.004817035049200058, + 0.006553626153618097, + 0.00563421193510294, + 0.005754252430051565, + 0.003671796526759863, + 0.003688705852255225, + 0.004686667583882809, + 0.002913977252319455, + 0.006761238444596529, + 0.005112549290060997, + 0.005748658906668425, + 0.005256707780063152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.038126103579998016 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080636642873287 + ], + "opposite_logprob_tokens": [ + -4.511080741882324 + ], + "logprob_reward": -0.011080636642873287, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080636642873287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17485332489013672, + 0.13939081132411957, + 0.04283229634165764, + 0.03404052555561066, + 0.09322141855955124, + 0.05993775278329849, + 0.08540856093168259, + 0.1736166924238205, + 0.10013649612665176, + 0.09656215459108353 + ], + "credit_map": [ + 0.014203977771103382, + 0.022171305492520332, + 0.01382718700915575, + 0.010768575593829155, + 0.04162721335887909, + 0.03967680037021637, + 0.011302560567855835, + 0.02127568982541561, + 0.04608813300728798, + 0.0067052715457975864, + 0.05350595712661743, + 0.006700553931295872, + 0.009923390112817287, + 0.006334987934678793, + 0.004047819878906012, + 0.006084696389734745, + 0.006720456760376692, + 0.004287486430257559, + 0.006831164937466383, + 0.0047321016900241375, + 0.004074480384588242, + 0.008289298973977566, + 0.004277324769645929, + 0.003619980299845338, + 0.00737778702750802, + 0.0021522175520658493, + 0.0052805631421506405, + 0.0018964421469718218, + 0.0032897659111768007, + 0.006779632531106472, + 0.0034287101589143276, + 0.0038354063872247934, + 0.010940124280750751, + 0.00882971566170454, + 0.006142113357782364, + 0.004534624982625246, + 0.022741233929991722, + 0.012647776864469051, + 0.019129088148474693, + 0.008256733417510986, + 0.006229787599295378, + 0.0042635235004127026, + 0.0061750044114887714, + 0.0034141801297664642, + 0.008884665556252003, + 0.00801934115588665, + 0.012428579851984978, + 0.01052266824990511, + 0.008467303588986397, + 0.008362489752471447, + 0.00622323015704751, + 0.00828678160905838, + 0.01686815172433853, + 0.016184935346245766, + 0.009977368637919426, + 0.011038295924663544, + 0.01166365947574377, + 0.006585835944861174, + 0.009367039427161217, + 0.00592202739790082, + 0.06279316544532776, + 0.044562291353940964, + 0.021641433238983154, + 0.011081238277256489, + 0.015591396950185299, + 0.009248168207705021, + 0.011832144111394882, + 0.010230052284896374, + 0.008550298400223255, + 0.011636665090918541, + 0.02017820253968239, + 0.012869561091065407, + 0.01168510690331459, + 0.008563898503780365, + 0.012674309313297272, + 0.007107587065547705, + 0.019955499097704887, + 0.013274391181766987, + 0.012291090562939644, + 0.01101025938987732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.011080636642873287 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865852743387222 + ], + "opposite_logprob_tokens": [ + -3.048658609390259 + ], + "logprob_reward": -0.04865852743387222, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865852743387222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09917432814836502, + 0.20071600377559662, + 0.04497160017490387, + 0.04817983880639076, + 0.09916025400161743, + 0.08732341974973679, + 0.14252568781375885, + 0.17636016011238098, + 0.048205502331256866, + 0.0533832348883152 + ], + "credit_map": [ + 0.008255976252257824, + 0.020498789846897125, + 0.007372773718088865, + 0.006404093932360411, + 0.018948890268802643, + 0.01708252727985382, + 0.006803031079471111, + 0.01380824577063322, + 0.07623808830976486, + 0.007853111252188683, + 0.08164401352405548, + 0.008965638466179371, + 0.006842046044766903, + 0.008597807958722115, + 0.0057527171447873116, + 0.004822558257728815, + 0.0076744225807487965, + 0.004174754023551941, + 0.008386855013668537, + 0.0037924631033092737, + 0.0038032508455216885, + 0.007695457898080349, + 0.004448628518730402, + 0.004995767492800951, + 0.009800620377063751, + 0.0046742819249629974, + 0.005060472525656223, + 0.0030747128184884787, + 0.004023942165076733, + 0.008384641259908676, + 0.005256650969386101, + 0.007904516533017159, + 0.018047792837023735, + 0.008092213422060013, + 0.01257476769387722, + 0.004663530271500349, + 0.0100254500284791, + 0.02301781252026558, + 0.013517416082322598, + 0.009221271611750126, + 0.009350963868200779, + 0.006749172694981098, + 0.005825467873364687, + 0.004104176536202431, + 0.02406959794461727, + 0.012184454128146172, + 0.011962768621742725, + 0.013076819479465485, + 0.00949413888156414, + 0.01500668004155159, + 0.008227555081248283, + 0.00822356715798378, + 0.03442320227622986, + 0.031915515661239624, + 0.01833052933216095, + 0.01690451055765152, + 0.010954025201499462, + 0.005361494142562151, + 0.00841095857322216, + 0.004516074433922768, + 0.06367563456296921, + 0.06050017848610878, + 0.012821192853152752, + 0.010120617225766182, + 0.0040844501927495, + 0.0035433010198175907, + 0.004500790033489466, + 0.0035811085253953934, + 0.005234077572822571, + 0.007789313327521086, + 0.011339427903294563, + 0.008133027702569962, + 0.005077080801129341, + 0.004100604448467493, + 0.006728418171405792, + 0.003563556121662259, + 0.008534622378647327, + 0.007052844390273094, + 0.01043714303523302, + 0.007888967171311378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.04865852743387222 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely high-speed macro shot, maintaining a static, top-down bird's-eye view, captures a stark black-and-white digital topographic map, where every topographic detail is rendered with astonishing clarity. The map is a complex network of extremely thin, precise contour lines and gridded elevation data set against a pure black background. Luminous bands, representing geological hazards, initially erupt in specific high-risk areas, flashing frantically, intermittently with bright white light. Their chaotic pulses cast fleeting, sharp highlights across the intricate digital geography. Then, as the high-speed effect seems to slow, the action shifts dramatically; the frantic blinking decelerates into a deliberate, rhythmic throbbing, with extended pauses between each pulse creating an intense, suspenseful atmosphere. Eventually, the pulsation ceases entirely, and the areas transform into an ominous, unwavering, steady glow. Their harsh white light now constantly illuminates the high-risk areas, creating an eerie scene that showcases a concentrated, imminent threat against the cold, analytically precise map.", + "seed": 5756, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_8/index_8207_seed_5756_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.028371115773916246, + "vlm_logprob": -0.028371115773916246, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.028371115773916246, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019305428490042686 + ], + "opposite_logprob_tokens": [ + -4.394305229187012 + ], + "logprob_reward": -0.019305428490042686, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.019305428490042686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22138698399066925, + 0.06839191913604736, + 0.10415752977132797, + 0.16334038972854614, + 0.03560875728726387, + 0.07100741565227509, + 0.18054568767547607, + 0.06676600128412247, + 0.05516931414604187, + 0.0336260125041008 + ], + "credit_map": [ + 0.03121313452720642, + 0.017086563631892204, + 0.02271466888487339, + 0.018743492662906647, + 0.020015893504023552, + 0.03853268548846245, + 0.023057514801621437, + 0.050023000687360764, + 0.009949122555553913, + 0.0074038817547261715, + 0.005496641155332327, + 0.007644300814718008, + 0.007246010471135378, + 0.008621394634246826, + 0.009651807136833668, + 0.012378753162920475, + 0.012415047734975815, + 0.012503169476985931, + 0.007627409417182207, + 0.012552434578537941, + 0.015228569507598877, + 0.013801406137645245, + 0.011746478267014027, + 0.018283002078533173, + 0.018679749220609665, + 0.02617710456252098, + 0.015944065526127815, + 0.02160279080271721, + 0.018282029777765274, + 0.02079927548766136, + 0.01576550491154194, + 0.026089858263731003, + 0.004219568334519863, + 0.004025936592370272, + 0.003736276878044009, + 0.005746148992329836, + 0.0024012599606066942, + 0.005588817410171032, + 0.004152324050664902, + 0.005738419946283102, + 0.010135469026863575, + 0.013204248622059822, + 0.008615313097834587, + 0.007917128503322601, + 0.006005880888551474, + 0.00974374357610941, + 0.007037525065243244, + 0.00834809523075819, + 0.0344766303896904, + 0.024731630459427834, + 0.024387439712882042, + 0.01075960136950016, + 0.022951168939471245, + 0.027469614520668983, + 0.014736004173755646, + 0.021033570170402527, + 0.012279409915208817, + 0.008840210735797882, + 0.010673224925994873, + 0.006302316207438707, + 0.005642320029437542, + 0.009087123908102512, + 0.006012802477926016, + 0.00792858935892582, + 0.005932399537414312, + 0.004027551040053368, + 0.00618343660607934, + 0.004905786830931902, + 0.004628209862858057, + 0.010915852151811123, + 0.004821680020540953, + 0.01375439204275608, + 0.003512590192258358, + 0.003075941698625684, + 0.005288689862936735, + 0.0028567102272063494, + 0.004870174918323755, + 0.006335236597806215, + 0.0030040631536394358, + 0.004682599566876888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.019305428490042686 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02063322253525257 + ], + "opposite_logprob_tokens": [ + -3.8956332206726074 + ], + "logprob_reward": -0.02063322253525257, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02063322253525257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1117110475897789, + 0.1279596984386444, + 0.19684268534183502, + 0.042012397199869156, + 0.08878672122955322, + 0.04746301844716072, + 0.049337368458509445, + 0.05787203460931778, + 0.04517754912376404, + 0.2328374981880188 + ], + "credit_map": [ + 0.01729954406619072, + 0.011392725631594658, + 0.016501840204000473, + 0.017059089615941048, + 0.00664515933021903, + 0.012100034393370152, + 0.010115645825862885, + 0.020597001537680626, + 0.0159133467823267, + 0.024705978110432625, + 0.010928387753665447, + 0.01450798287987709, + 0.012368830852210522, + 0.016020087525248528, + 0.012337468564510345, + 0.021177610382437706, + 0.021458454430103302, + 0.021647552028298378, + 0.016658583655953407, + 0.022495819255709648, + 0.03073693811893463, + 0.025898193940520287, + 0.035536978393793106, + 0.02241017110645771, + 0.006400988437235355, + 0.004018357023596764, + 0.003894190303981304, + 0.003563297912478447, + 0.0034715652000159025, + 0.006585343275219202, + 0.003796087810769677, + 0.010282565839588642, + 0.008890675380825996, + 0.007935513742268085, + 0.008366874419152737, + 0.014897085726261139, + 0.008232920430600643, + 0.011456937529146671, + 0.019757218658924103, + 0.009249494411051273, + 0.005103433970361948, + 0.010140898637473583, + 0.004437064751982689, + 0.005029937252402306, + 0.004663427360355854, + 0.008173501119017601, + 0.0060234833508729935, + 0.0038912720046937466, + 0.00924742966890335, + 0.007732382044196129, + 0.006288065575063229, + 0.0030677125323563814, + 0.006795773282647133, + 0.0073067303746938705, + 0.0030263052321970463, + 0.0058729685842990875, + 0.00745915062725544, + 0.01141735166311264, + 0.006791603751480579, + 0.0057811494916677475, + 0.004679729696363211, + 0.010957157239317894, + 0.004734484478831291, + 0.006051407195627689, + 0.0033089029602706432, + 0.006181524600833654, + 0.009110343642532825, + 0.003941704519093037, + 0.005406937096267939, + 0.0063638631254434586, + 0.0038087915163487196, + 0.007055485155433416, + 0.024641279131174088, + 0.06932815164327621, + 0.01142335869371891, + 0.01103106327354908, + 0.0190194770693779, + 0.016138197854161263, + 0.008998159319162369, + 0.07225780934095383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.02063322253525257 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060496248304844 + ], + "opposite_logprob_tokens": [ + -3.8956048488616943 + ], + "logprob_reward": -0.02060496248304844, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060496248304844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08811688423156738, + 0.03767916187644005, + 0.03308038413524628, + 0.06063299998641014, + 0.04417133331298828, + 0.058678291738033295, + 0.12332766503095627, + 0.055037762969732285, + 0.188599094748497, + 0.3106764853000641 + ], + "credit_map": [ + 0.013492241501808167, + 0.013900029473006725, + 0.007558803539723158, + 0.008436201140284538, + 0.009852449409663677, + 0.012456439435482025, + 0.011502965353429317, + 0.010917752049863338, + 0.00671983789652586, + 0.004351805429905653, + 0.0031359042041003704, + 0.004156099632382393, + 0.0029757507145404816, + 0.005142958369106054, + 0.004586874973028898, + 0.006609929259866476, + 0.004927364178001881, + 0.003140836488455534, + 0.0036774692125618458, + 0.005406922660768032, + 0.002887256210669875, + 0.004577413201332092, + 0.004478187300264835, + 0.0039849309250712395, + 0.00685788132250309, + 0.007216686382889748, + 0.006971630733460188, + 0.00986597128212452, + 0.006576506420969963, + 0.0087985098361969, + 0.008560013957321644, + 0.005785797256976366, + 0.004786203615367413, + 0.0039353808388113976, + 0.003435926977545023, + 0.006125899497419596, + 0.0035136172082275152, + 0.006163456942886114, + 0.008067205548286438, + 0.008143641985952854, + 0.007129003759473562, + 0.010987251065671444, + 0.004025840666145086, + 0.009193351492285728, + 0.005305333528667688, + 0.009164894931018353, + 0.00578158488497138, + 0.007091035600751638, + 0.026355378329753876, + 0.01319290604442358, + 0.016688549891114235, + 0.009294427931308746, + 0.021564161404967308, + 0.01616319641470909, + 0.008747551590204239, + 0.011321497149765491, + 0.006429854314774275, + 0.00974532775580883, + 0.006034509744495153, + 0.006005123257637024, + 0.005316073074936867, + 0.010140019468963146, + 0.005074928980320692, + 0.0062919240444898605, + 0.019464900717139244, + 0.017514873296022415, + 0.03378021717071533, + 0.012419011443853378, + 0.019648665562272072, + 0.036053288727998734, + 0.012956127524375916, + 0.03676199913024902, + 0.0328681506216526, + 0.08745229244232178, + 0.018236305564641953, + 0.016231918707489967, + 0.027971256524324417, + 0.02224988490343094, + 0.012155887670814991, + 0.09351078420877457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.02060496248304844 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635699324309826 + ], + "opposite_logprob_tokens": [ + -3.6513569355010986 + ], + "logprob_reward": -0.02635699324309826, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635699324309826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1783616542816162, + 0.029963281005620956, + 0.09977850317955017, + 0.05396999418735504, + 0.05197096988558769, + 0.0862201526761055, + 0.0692308321595192, + 0.09395380318164825, + 0.10222463309764862, + 0.23432616889476776 + ], + "credit_map": [ + 0.02343961037695408, + 0.0204403568059206, + 0.014012480154633522, + 0.019700633361935616, + 0.012101027183234692, + 0.03019396774470806, + 0.015530962496995926, + 0.04294262081384659, + 0.004812033846974373, + 0.004076691344380379, + 0.003244934603571892, + 0.00477367639541626, + 0.0022848001681268215, + 0.0038108297158032656, + 0.003334426786750555, + 0.00362588744610548, + 0.009357864037156105, + 0.011513370089232922, + 0.008313920348882675, + 0.013123910874128342, + 0.015566335991024971, + 0.012708235532045364, + 0.017966222018003464, + 0.011228641495108604, + 0.005570178851485252, + 0.00812285766005516, + 0.00474382983520627, + 0.007650959771126509, + 0.004671955015510321, + 0.008201093412935734, + 0.0068130288273096085, + 0.0081960903480649, + 0.004886927083134651, + 0.0038454255554825068, + 0.004160662647336721, + 0.008711681701242924, + 0.004280109889805317, + 0.0069165462628006935, + 0.012235700152814388, + 0.006933914497494698, + 0.010282304137945175, + 0.020297078415751457, + 0.009223724715411663, + 0.00971716083586216, + 0.006718635559082031, + 0.011725440621376038, + 0.009457312524318695, + 0.008798492141067982, + 0.011597441509366035, + 0.009582674130797386, + 0.01065902505069971, + 0.004591693170368671, + 0.010509677231311798, + 0.009153207764029503, + 0.005028829909861088, + 0.008108281530439854, + 0.010836250148713589, + 0.011966033838689327, + 0.012070129625499249, + 0.009376191534101963, + 0.01030663400888443, + 0.021816985681653023, + 0.007762815337628126, + 0.009818769060075283, + 0.010421628132462502, + 0.01739318110048771, + 0.015263712033629417, + 0.00697109941393137, + 0.00812404416501522, + 0.015459277667105198, + 0.0064891851507127285, + 0.022102512419223785, + 0.02560829184949398, + 0.06593388319015503, + 0.012276439927518368, + 0.011625859886407852, + 0.021076608449220657, + 0.017598507925868034, + 0.009110353887081146, + 0.07109620422124863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.02635699324309826 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view a static, top-down bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a black-and-white digital topographic map with thin contour lines on a pure black background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do luminous bands representing geological hazards initially erupt and flash frantically in high-risk areas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action transition from frantic blinking to a deliberate, rhythmic throbbing with extended pauses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the pulsating areas eventually cease and transform into an ominous, unwavering steady glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495497211813927 + ], + "opposite_logprob_tokens": [ + -2.929955005645752 + ], + "logprob_reward": -0.05495497211813927, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495497211813927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14928747713565826, + 0.04293964058160782, + 0.06212736293673515, + 0.043994296342134476, + 0.037475455552339554, + 0.08681439608335495, + 0.11235235631465912, + 0.06635456532239914, + 0.10721243917942047, + 0.29144200682640076 + ], + "credit_map": [ + 0.01953277364373207, + 0.01834772154688835, + 0.01307876780629158, + 0.01555702369660139, + 0.010280556045472622, + 0.023793289437890053, + 0.019178662449121475, + 0.029518678784370422, + 0.006887832656502724, + 0.006210765801370144, + 0.003988279029726982, + 0.005157678388059139, + 0.0036164582706987858, + 0.0059181866236031055, + 0.004273591563105583, + 0.006886845454573631, + 0.006170316599309444, + 0.0071103996597230434, + 0.005487566813826561, + 0.007919960655272007, + 0.008627546019852161, + 0.008761415258049965, + 0.010157330892980099, + 0.007892828434705734, + 0.004975727293640375, + 0.0050375391729176044, + 0.004700563382357359, + 0.0059263394214212894, + 0.005232620984315872, + 0.006423070561140776, + 0.006121345795691013, + 0.005577093455940485, + 0.004506190773099661, + 0.0038159952964633703, + 0.0029046216513961554, + 0.005984185263514519, + 0.003291085595265031, + 0.004859508015215397, + 0.0064749824814498425, + 0.005638888571411371, + 0.009033866226673126, + 0.014511977322399616, + 0.00979419145733118, + 0.010889740660786629, + 0.008932875469326973, + 0.01475188136100769, + 0.009533075615763664, + 0.009366795420646667, + 0.012535307556390762, + 0.015007668174803257, + 0.01854272186756134, + 0.012027518823742867, + 0.021882956847548485, + 0.01211149338632822, + 0.009444774128496647, + 0.01079991552978754, + 0.01313293632119894, + 0.008862175047397614, + 0.008814012631773949, + 0.004969045519828796, + 0.005591917783021927, + 0.011698383837938309, + 0.005824638530611992, + 0.007461455184966326, + 0.00924465898424387, + 0.0260879248380661, + 0.014699091203510761, + 0.009855866432189941, + 0.009733004495501518, + 0.011600569821894169, + 0.006579528097063303, + 0.019411804154515266, + 0.03227132558822632, + 0.0814683735370636, + 0.016613079234957695, + 0.013927731662988663, + 0.02596636302769184, + 0.021448684856295586, + 0.012500648386776447, + 0.08724579960107803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.05495497211813927 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_9/index_tensor([7441, 7441])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_9/index_tensor([7441, 7441])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f21ad23bcdd84568eb42aa637849136b72970176 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_9/index_tensor([7441, 7441])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An uninterrupted, long shot, shot from a level perspective, pans rightward at an incredibly slow, steady, and deliberate pace, capturing the ethereal scene of a reed marsh at dawn. A milky white mist, like a veil, permeates, rises, and entwines among the dense reeds, blurring the boundary between sky and water. The entire scene is imbued with a delicate, ink-and-wash hue, imbued with the beauty of classical Chinese painting. As the camera pans, the viewer's gaze sweeps across this vast, tranquil wetland. Each slender reed stem, a medley of withered yellow and fresh green, is festooned with crystal-clear dewdrops, gleaming like tiny diamonds in the soft, barely visible light of the dawn's first rays. The reed catkins, fluffy as snow, remain motionless in the nearly still air. The calm water reflects the reeds, blending with the shifting mist to create a dreamlike, layered effect. This slow and continuous camera movement, combined with the visible slow flow of fog, together renders an ultimate tranquility and Zen feeling, as if time has frozen at this moment, forming a flowing, poetic and picturesque classical ink painting scroll.", + "seed": 25820, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_9/index_7441_seed_25820_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03394441977143288, + "vlm_logprob": -0.03394441977143288, + "vlm_yes_no_margin": 3.6499999523162843, + "vlm_token_credit": -0.03394441977143288, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.061077456921339035 + ], + "opposite_logprob_tokens": [ + -3.061077356338501 + ], + "logprob_reward": -0.061077456921339035, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.061077456921339035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0912180244922638, + 0.07767218351364136, + 0.023364219814538956, + 0.027221493422985077, + 0.07200141251087189, + 0.4692024290561676, + 0.07049938291311264, + 0.07191154360771179, + 0.06484737992286682, + 0.03206188604235649 + ], + "credit_map": [ + 0.012013857252895832, + 0.01651325263082981, + 0.0108293816447258, + 0.013312997296452522, + 0.011403368785977364, + 0.010771042667329311, + 0.009425726719200611, + 0.0069484105333685875, + 0.006385113578289747, + 0.007910050451755524, + 0.006596494000405073, + 0.006110192276537418, + 0.006563051603734493, + 0.01773778535425663, + 0.00952033419162035, + 0.01684916578233242, + 0.0026937678921967745, + 0.003270880552008748, + 0.0025297445245087147, + 0.002602621214464307, + 0.003082918468862772, + 0.004155295435339212, + 0.002764409640803933, + 0.0022645853459835052, + 0.002449637046083808, + 0.003071569139137864, + 0.004160714335739613, + 0.004970178939402103, + 0.0038668441120535135, + 0.0024815958458930254, + 0.0026821347419172525, + 0.0035388232208788395, + 0.003812388749793172, + 0.0035283833276480436, + 0.00992077961564064, + 0.022626081481575966, + 0.005054600071161985, + 0.003653526771813631, + 0.01793227344751358, + 0.005473385564982891, + 0.02672424167394638, + 0.015556172467768192, + 0.14752192795276642, + 0.040095239877700806, + 0.020917551591992378, + 0.013357875868678093, + 0.16711024940013885, + 0.03791920840740204, + 0.0051045906729996204, + 0.007653642911463976, + 0.012588086538016796, + 0.017454244196414948, + 0.005278209690004587, + 0.005706161726266146, + 0.009664224460721016, + 0.0070502315647900105, + 0.00685939472168684, + 0.013871714472770691, + 0.007871026173233986, + 0.008484933525323868, + 0.007875191979110241, + 0.013244749046862125, + 0.0081747742369771, + 0.005529771093279123, + 0.005177571438252926, + 0.007346395403146744, + 0.015472471714019775, + 0.0164262056350708, + 0.0032896578777581453, + 0.0038665214087814093, + 0.007882185280323029, + 0.005386378150433302, + 0.0049562398344278336, + 0.005869925953447819, + 0.004308873321861029, + 0.0029893028549849987, + 0.0023735377471894026, + 0.004161476623266935, + 0.00421687588095665, + 0.003185658250004053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.061077456921339035 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014227267354726791 + ], + "opposite_logprob_tokens": [ + -4.264227390289307 + ], + "logprob_reward": -0.014227267354726791, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014227267354726791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07543233036994934, + 0.14088983833789825, + 0.05741063132882118, + 0.028883274644613266, + 0.05384998768568039, + 0.26303455233573914, + 0.0749903991818428, + 0.055232398211956024, + 0.22134742140769958, + 0.02892916649580002 + ], + "credit_map": [ + 0.006658155005425215, + 0.010645641013979912, + 0.00866449624300003, + 0.012619921937584877, + 0.006479102652519941, + 0.007379361893981695, + 0.015534856356680393, + 0.0074507989920675755, + 0.01120549626648426, + 0.012712577357888222, + 0.016488732770085335, + 0.027556944638490677, + 0.015048337168991566, + 0.013709339313209057, + 0.02807609736919403, + 0.016092313453555107, + 0.004540207330137491, + 0.0073173027485609055, + 0.005639895796775818, + 0.00994654931128025, + 0.00567151652649045, + 0.00556768337264657, + 0.01337416935712099, + 0.005353307351469994, + 0.002737357746809721, + 0.004216991364955902, + 0.003526991233229637, + 0.004375942051410675, + 0.003148050978779793, + 0.0027250449638813734, + 0.004355885554105043, + 0.0037970126140862703, + 0.00483222771435976, + 0.00454335380345583, + 0.010520697571337223, + 0.01452725101262331, + 0.004536375403404236, + 0.002772510051727295, + 0.007933614775538445, + 0.004183965735137463, + 0.015248419716954231, + 0.011338683776557446, + 0.0805642306804657, + 0.02308531478047371, + 0.01430799625813961, + 0.007540834601968527, + 0.09068423509597778, + 0.020264841616153717, + 0.005733600351959467, + 0.01148660946637392, + 0.01310825813561678, + 0.01758003607392311, + 0.005579723976552486, + 0.00669896462932229, + 0.007642616052180529, + 0.00716059934347868, + 0.005702662747353315, + 0.008560236543416977, + 0.007130936719477177, + 0.0074929604306817055, + 0.005242897663265467, + 0.006209978833794594, + 0.007493223529309034, + 0.007399504538625479, + 0.017369553446769714, + 0.03131034970283508, + 0.0516555979847908, + 0.051781781017780304, + 0.010980673134326935, + 0.013992540538311005, + 0.023570913821458817, + 0.02068602852523327, + 0.005213627126067877, + 0.0046244594268500805, + 0.004566597286611795, + 0.0030082594603300095, + 0.002655566204339266, + 0.004021155647933483, + 0.0028016529977321625, + 0.0020378511399030685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.014227267354726791 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199178665876389 + ], + "opposite_logprob_tokens": [ + -4.264199256896973 + ], + "logprob_reward": -0.014199178665876389, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199178665876389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08139829337596893, + 0.10531872510910034, + 0.08437112718820572, + 0.08764353394508362, + 0.04518744349479675, + 0.2107309252023697, + 0.10597916692495346, + 0.16110846400260925, + 0.059784840792417526, + 0.05847745016217232 + ], + "credit_map": [ + 0.008473885245621204, + 0.01155165210366249, + 0.0073759062215685844, + 0.01162007451057434, + 0.01020055077970028, + 0.014018092304468155, + 0.010796165093779564, + 0.007361967582255602, + 0.009512059390544891, + 0.012245800346136093, + 0.009349682368338108, + 0.008109471760690212, + 0.008519825525581837, + 0.028502319008111954, + 0.008757712319493294, + 0.020321855321526527, + 0.008187144063413143, + 0.011578550562262535, + 0.009595461189746857, + 0.01013497356325388, + 0.014464125037193298, + 0.01096420269459486, + 0.011575496755540371, + 0.007871173322200775, + 0.007527446374297142, + 0.011261229403316975, + 0.01052128802984953, + 0.017126942053437233, + 0.00883929617702961, + 0.009683693759143353, + 0.014114581979811192, + 0.008569061756134033, + 0.004938364494591951, + 0.0037252893671393394, + 0.006912610027939081, + 0.009183434769511223, + 0.0043383329175412655, + 0.003947350196540356, + 0.00665963813662529, + 0.0054824259132146835, + 0.013814782723784447, + 0.010600161738693714, + 0.06463388353586197, + 0.02329964004456997, + 0.011189723387360573, + 0.00811005663126707, + 0.06340780854225159, + 0.015674874186515808, + 0.009200716391205788, + 0.009878665208816528, + 0.0280985739082098, + 0.027778754010796547, + 0.008963149972259998, + 0.008706873282790184, + 0.007369943894445896, + 0.005982494447380304, + 0.019787484779953957, + 0.023249149322509766, + 0.022221915423870087, + 0.018804190680384636, + 0.014812502078711987, + 0.028218748047947884, + 0.015666089951992035, + 0.018348390236496925, + 0.007341368589550257, + 0.010387449525296688, + 0.009166226722300053, + 0.010722012259066105, + 0.005751086864620447, + 0.005260166246443987, + 0.006403991021215916, + 0.00475253676995635, + 0.012220966629683971, + 0.009238043799996376, + 0.007273412309587002, + 0.004478441085666418, + 0.004403576720505953, + 0.0087293591350317, + 0.006216246169060469, + 0.005917408037930727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.014199178665876389 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181814506649971 + ], + "opposite_logprob_tokens": [ + -4.018181324005127 + ], + "logprob_reward": -0.0181814506649971, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.0181814506649971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022854357957839966, + 0.11920730769634247, + 0.017698273062705994, + 0.040811777114868164, + 0.09715738147497177, + 0.45869290828704834, + 0.06335201859474182, + 0.13275004923343658, + 0.034417275339365005, + 0.013058647513389587 + ], + "credit_map": [ + 0.002642715349793434, + 0.004216680768877268, + 0.002285243244841695, + 0.0023459994699805975, + 0.0032128188759088516, + 0.003138880245387554, + 0.0026879271026700735, + 0.002324094297364354, + 0.005096838343888521, + 0.007083744276314974, + 0.00388166937045753, + 0.005268714390695095, + 0.01227528229355812, + 0.050698406994342804, + 0.006484838668256998, + 0.028417810797691345, + 0.0015241438522934914, + 0.002267662202939391, + 0.002516921842470765, + 0.0029358319006860256, + 0.0028271391056478024, + 0.0016609395388513803, + 0.0024698812048882246, + 0.0014957513194531202, + 0.0025546469260007143, + 0.004443635232746601, + 0.0058548590168356895, + 0.008050848729908466, + 0.0038745359051972628, + 0.00324591645039618, + 0.008250068873167038, + 0.004537265282124281, + 0.00469279196113348, + 0.004760136362165213, + 0.014237006194889545, + 0.03319302573800087, + 0.00568443164229393, + 0.002974828938022256, + 0.0249831210821867, + 0.006632033735513687, + 0.025788620114326477, + 0.01537107303738594, + 0.14356546103954315, + 0.041941218078136444, + 0.01985061727464199, + 0.012531029991805553, + 0.1613619327545166, + 0.03828294202685356, + 0.004306493792682886, + 0.005805769003927708, + 0.01301026064902544, + 0.015108267776668072, + 0.003976189065724611, + 0.0046090479008853436, + 0.009539193473756313, + 0.00699679646641016, + 0.012702670879662037, + 0.0179411843419075, + 0.018527954816818237, + 0.01896226778626442, + 0.011716102249920368, + 0.02064649388194084, + 0.016572818160057068, + 0.015680555254220963, + 0.003034749999642372, + 0.00423897011205554, + 0.007568821776658297, + 0.00789257325232029, + 0.002209819620475173, + 0.0022309646010398865, + 0.003446057438850403, + 0.0037953194696456194, + 0.0019134427420794964, + 0.0023821466602385044, + 0.0020213534589856863, + 0.0009786557639017701, + 0.0012213628506287932, + 0.0018399960827082396, + 0.0012750052846968174, + 0.0014266850193962455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.0181814506649971 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203674525022507 + ], + "opposite_logprob_tokens": [ + -2.8120367527008057 + ], + "logprob_reward": -0.06203674525022507, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203674525022507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08568014204502106, + 0.05930943787097931, + 0.04903813451528549, + 0.11143931746482849, + 0.18971365690231323, + 0.10192825645208359, + 0.13407498598098755, + 0.12795643508434296, + 0.09446165710687637, + 0.04639798030257225 + ], + "credit_map": [ + 0.008543962612748146, + 0.01752536930143833, + 0.010416382923722267, + 0.012059020809829235, + 0.009533938951790333, + 0.010295822285115719, + 0.009545083157718182, + 0.007760557811707258, + 0.00554028432816267, + 0.0077595179900527, + 0.009402928873896599, + 0.009692024439573288, + 0.007452670950442553, + 0.0070024533197283745, + 0.007977352477610111, + 0.004482205491513014, + 0.004406080115586519, + 0.005435105878859758, + 0.004295451566576958, + 0.008092937991023064, + 0.005073883570730686, + 0.0047634802758693695, + 0.012088296934962273, + 0.004882895853370428, + 0.008258486166596413, + 0.013052481226623058, + 0.014822716824710369, + 0.021385420113801956, + 0.008752946741878986, + 0.00784496683627367, + 0.02458351105451584, + 0.012738779187202454, + 0.008709335699677467, + 0.008295685052871704, + 0.029468081891536713, + 0.0663207396864891, + 0.008353988640010357, + 0.00656272005289793, + 0.049743205308914185, + 0.012259889394044876, + 0.0068326652981340885, + 0.008180612698197365, + 0.019082654267549515, + 0.013478945940732956, + 0.006613511126488447, + 0.0068093943409621716, + 0.028733516111969948, + 0.01219695620238781, + 0.009412859566509724, + 0.013902596198022366, + 0.022788891568779945, + 0.03247400373220444, + 0.009289142675697803, + 0.011302688159048557, + 0.0222835261374712, + 0.012621277011930943, + 0.011637422256171703, + 0.02991793490946293, + 0.013325418345630169, + 0.015398882329463959, + 0.012685198336839676, + 0.023165415972471237, + 0.01191702764481306, + 0.009909114800393581, + 0.007398517802357674, + 0.013766834512352943, + 0.021222161129117012, + 0.01976248249411583, + 0.006273880135267973, + 0.007671060506254435, + 0.00899861566722393, + 0.009368094615638256, + 0.008013333193957806, + 0.008348244242370129, + 0.007549090310931206, + 0.003698609070852399, + 0.00327620143070817, + 0.006423031445592642, + 0.004869008436799049, + 0.004220459144562483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.06203674525022507 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An uninterrupted, long shot, shot from a level perspective, pans rightward at an incredibly slow, steady, and deliberate pace, capturing the ethereal scene of a reed marsh at dawn. A milky white mist, like a veil, permeates, rises, and entwines among the dense reeds, blurring the boundary between sky and water. The entire scene is imbued with a delicate, ink-and-wash hue, imbued with the beauty of classical Chinese painting. As the camera pans, the viewer's gaze sweeps across this vast, tranquil wetland. Each slender reed stem, a medley of withered yellow and fresh green, is festooned with crystal-clear dewdrops, gleaming like tiny diamonds in the soft, barely visible light of the dawn's first rays. The reed catkins, fluffy as snow, remain motionless in the nearly still air. The calm water reflects the reeds, blending with the shifting mist to create a dreamlike, layered effect. This slow and continuous camera movement, combined with the visible slow flow of fog, together renders an ultimate tranquility and Zen feeling, as if time has frozen at this moment, forming a flowing, poetic and picturesque classical ink painting scroll.", + "seed": 25821, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_9/index_7441_seed_25821_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04638135181739926, + "vlm_logprob": -0.04638135181739926, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.04638135181739926, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07077067345380783 + ], + "opposite_logprob_tokens": [ + -2.945770740509033 + ], + "logprob_reward": -0.07077067345380783, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07077067345380783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.349290132522583, + 0.08336970210075378, + 0.04437798634171486, + 0.04883791133761406, + 0.0924747884273529, + 0.15955102443695068, + 0.09263375401496887, + 0.04356423392891884, + 0.061377398669719696, + 0.02452302724123001 + ], + "credit_map": [ + 0.0201265811920166, + 0.027125336229801178, + 0.08027138561010361, + 0.08937157690525055, + 0.021729152649641037, + 0.02292804792523384, + 0.0637325569987297, + 0.024005500599741936, + 0.005591179244220257, + 0.006979858037084341, + 0.01623535342514515, + 0.007395342458039522, + 0.010020317509770393, + 0.009022993966937065, + 0.01822906918823719, + 0.009895588271319866, + 0.0031588186975568533, + 0.00367147964425385, + 0.006266530603170395, + 0.008094173856079578, + 0.00527125783264637, + 0.004360005725175142, + 0.008843090385198593, + 0.004712630063295364, + 0.0034786739852279425, + 0.00435472559183836, + 0.010942171327769756, + 0.0058593456633389, + 0.004823731724172831, + 0.005324757192283869, + 0.007870419882237911, + 0.006184087600558996, + 0.005377562716603279, + 0.013150261715054512, + 0.010326385498046875, + 0.01664835214614868, + 0.01268206536769867, + 0.011018927209079266, + 0.014131378382444382, + 0.009139850735664368, + 0.009979387745261192, + 0.007698326371610165, + 0.04872667044401169, + 0.03738999739289284, + 0.00812789797782898, + 0.014334335923194885, + 0.011292515322566032, + 0.022001903504133224, + 0.006292165722697973, + 0.00678167724981904, + 0.028736118227243423, + 0.011620978824794292, + 0.004634793847799301, + 0.007173338904976845, + 0.008036695420742035, + 0.019357984885573387, + 0.0030740508809685707, + 0.005071787163615227, + 0.00603599613532424, + 0.004308559000492096, + 0.008560261689126492, + 0.004511931445449591, + 0.005922374315559864, + 0.006079274229705334, + 0.0027414998039603233, + 0.00364236650057137, + 0.006140160374343395, + 0.016382446512579918, + 0.0034215503837913275, + 0.005309593863785267, + 0.006455826107412577, + 0.017283959314227104, + 0.002024655230343342, + 0.0027036501560360193, + 0.0033343667164444923, + 0.0026406790129840374, + 0.002828283002600074, + 0.0031033798586577177, + 0.004643051885068417, + 0.00324495998211205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.07077067345380783 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026385091245174408 + ], + "opposite_logprob_tokens": [ + -3.6513850688934326 + ], + "logprob_reward": -0.026385091245174408, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026385091245174408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2682546377182007, + 0.10054835677146912, + 0.09938827902078629, + 0.034180108457803726, + 0.0811665803194046, + 0.1563500165939331, + 0.04554712399840355, + 0.07084312289953232, + 0.0799466073513031, + 0.06377518177032471 + ], + "credit_map": [ + 0.016582738608121872, + 0.023253904655575752, + 0.05467666685581207, + 0.055534642189741135, + 0.02402958646416664, + 0.020963450893759727, + 0.04962141439318657, + 0.023592224344611168, + 0.006341031286865473, + 0.012751235626637936, + 0.011069310829043388, + 0.012776996940374374, + 0.015026104636490345, + 0.016940902918577194, + 0.011403369717299938, + 0.014239405281841755, + 0.00748826190829277, + 0.007696944288909435, + 0.015221929177641869, + 0.014602060429751873, + 0.013592405244708061, + 0.012383333407342434, + 0.018302004784345627, + 0.010101327672600746, + 0.0026229547802358866, + 0.0029681583400815725, + 0.006264698226004839, + 0.004235816653817892, + 0.003652738407254219, + 0.004045965149998665, + 0.004472333937883377, + 0.0059174420312047005, + 0.004027084447443485, + 0.011066119186580181, + 0.00869323406368494, + 0.01234356313943863, + 0.012340078130364418, + 0.012816589325666428, + 0.009659080766141415, + 0.010220826603472233, + 0.011139542795717716, + 0.012274429202079773, + 0.037274423986673355, + 0.010697126388549805, + 0.010738749988377094, + 0.021659476682543755, + 0.013329832814633846, + 0.03923641890287399, + 0.003037049202248454, + 0.00289120152592659, + 0.012030351907014847, + 0.004867862910032272, + 0.004242885857820511, + 0.003781971288844943, + 0.005396067630499601, + 0.009299731813371181, + 0.005192801356315613, + 0.007756680250167847, + 0.007216084282845259, + 0.006934113800525665, + 0.013030578382313251, + 0.007972524501383305, + 0.014149517752230167, + 0.00859081745147705, + 0.007321419660001993, + 0.011917968280613422, + 0.006613424979150295, + 0.010538481175899506, + 0.01033944170922041, + 0.009110642597079277, + 0.011015722528100014, + 0.013089506886899471, + 0.004187197424471378, + 0.004440148826688528, + 0.010265342891216278, + 0.0057660979218780994, + 0.0063036419451236725, + 0.009533284232020378, + 0.011874225921928883, + 0.011405245400965214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.026385091245174408 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076352194882929325 + ], + "opposite_logprob_tokens": [ + -4.882635116577148 + ], + "logprob_reward": -0.0076352194882929325, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076352194882929325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30086612701416016, + 0.05910735949873924, + 0.06042978912591934, + 0.051012199372053146, + 0.04581526294350624, + 0.2342909723520279, + 0.054046329110860825, + 0.05504097416996956, + 0.10116425901651382, + 0.03822670504450798 + ], + "credit_map": [ + 0.01735548861324787, + 0.0178474523127079, + 0.09064400941133499, + 0.027399562299251556, + 0.02063862793147564, + 0.02640586346387863, + 0.0771649032831192, + 0.023410223424434662, + 0.005851959343999624, + 0.006376104895025492, + 0.01225567888468504, + 0.00614274712279439, + 0.008496647700667381, + 0.005562654230743647, + 0.009268536232411861, + 0.005153030622750521, + 0.0036346870474517345, + 0.004126161336898804, + 0.009734982624650002, + 0.006726319435983896, + 0.006892687175422907, + 0.008059886284172535, + 0.00949862040579319, + 0.011756446212530136, + 0.0026804248336702585, + 0.004350507166236639, + 0.012797607108950615, + 0.003721463494002819, + 0.004718824755400419, + 0.0024059342686086893, + 0.015883011743426323, + 0.004454427864402533, + 0.0031204991973936558, + 0.00797626655548811, + 0.005869745742529631, + 0.006447623949497938, + 0.008400320075452328, + 0.004604808986186981, + 0.005469498224556446, + 0.003926501143723726, + 0.010644752532243729, + 0.008922500535845757, + 0.0724826529622078, + 0.09515048563480377, + 0.0035751890391111374, + 0.009384834207594395, + 0.019019868224859238, + 0.01511069294065237, + 0.004290506709367037, + 0.004990183748304844, + 0.013898075558245182, + 0.006919467821717262, + 0.0038287066854536533, + 0.004993480164557695, + 0.006983209401369095, + 0.008142699487507343, + 0.0046163443475961685, + 0.007644773926585913, + 0.0066072396002709866, + 0.00581115810200572, + 0.011276253499090672, + 0.006095898803323507, + 0.007972382940351963, + 0.005016924813389778, + 0.004979946184903383, + 0.007601554039865732, + 0.009672466665506363, + 0.021945668384432793, + 0.007444055750966072, + 0.009116653352975845, + 0.015608428046107292, + 0.024795489385724068, + 0.002308816649019718, + 0.0032704672776162624, + 0.004308141302317381, + 0.0025793793611228466, + 0.0031144439708441496, + 0.008282824419438839, + 0.0041769747622311115, + 0.010185658000409603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.0076352194882929325 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038089726120233536 + ], + "opposite_logprob_tokens": [ + -3.2880897521972656 + ], + "logprob_reward": -0.038089726120233536, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038089726120233536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5035372376441956, + 0.09499256312847137, + 0.032404713332653046, + 0.03818729147315025, + 0.0683867484331131, + 0.16799941658973694, + 0.017002813518047333, + 0.033872444182634354, + 0.022008167579770088, + 0.021608585491776466 + ], + "credit_map": [ + 0.03268996626138687, + 0.05786043778061867, + 0.08898747712373734, + 0.13693909347057343, + 0.03851613402366638, + 0.0395115502178669, + 0.06883268803358078, + 0.0401998832821846, + 0.006050952710211277, + 0.006967904977500439, + 0.0227756816893816, + 0.008299700915813446, + 0.008659213781356812, + 0.006806336808949709, + 0.02715420164167881, + 0.008278570137917995, + 0.0040480224415659904, + 0.0032447113189846277, + 0.0035315207205712795, + 0.004345030058175325, + 0.003775720251724124, + 0.004690661560744047, + 0.003338515292853117, + 0.005430533085018396, + 0.002559929620474577, + 0.0023945190478116274, + 0.0100377406924963, + 0.0027449256740510464, + 0.0037466194480657578, + 0.004178081639111042, + 0.008479578420519829, + 0.004045896232128143, + 0.0052426583133637905, + 0.010013864375650883, + 0.00867518037557602, + 0.0106353759765625, + 0.0115128755569458, + 0.006819535978138447, + 0.008754580281674862, + 0.006732678506523371, + 0.011322500184178352, + 0.009461467154324055, + 0.04684831202030182, + 0.026827897876501083, + 0.01062813401222229, + 0.0167666207998991, + 0.012750188820064068, + 0.03339430317282677, + 0.001242037513293326, + 0.0027204533107578754, + 0.0019630668684840202, + 0.0022703511640429497, + 0.002051295479759574, + 0.0026189310010522604, + 0.002315640216693282, + 0.0018210387788712978, + 0.003324018558487296, + 0.005078568588942289, + 0.003128113690763712, + 0.0031652026809751987, + 0.006149583496153355, + 0.00404079956933856, + 0.0042120106518268585, + 0.004774150438606739, + 0.0012474076356738806, + 0.0019166661659255624, + 0.0022902239579707384, + 0.004404398147016764, + 0.0020722171757370234, + 0.001986968796700239, + 0.0032587405294179916, + 0.0048315455205738544, + 0.0016294989036396146, + 0.00229951785877347, + 0.002651816001161933, + 0.002518582157790661, + 0.0025266334414482117, + 0.003822002559900284, + 0.0032048430293798447, + 0.0029556918889284134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.038089726120233536 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902604877948761 + ], + "opposite_logprob_tokens": [ + -2.4640259742736816 + ], + "logprob_reward": -0.08902604877948761, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902604877948761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4143751859664917, + 0.027413703501224518, + 0.022902315482497215, + 0.07837411016225815, + 0.03374506160616875, + 0.320739209651947, + 0.026555122807621956, + 0.02706999145448208, + 0.03503700718283653, + 0.013788320124149323 + ], + "credit_map": [ + 0.022140435874462128, + 0.019998321309685707, + 0.13442428410053253, + 0.06219622492790222, + 0.022958794608712196, + 0.023119531571865082, + 0.10626833140850067, + 0.023269258439540863, + 0.0023938394151628017, + 0.0041454387828707695, + 0.004176429472863674, + 0.002362639643251896, + 0.004601658787578344, + 0.004175996873527765, + 0.0031645516864955425, + 0.0023931481409817934, + 0.0020156074315309525, + 0.0018507415661588311, + 0.0030233063735067844, + 0.003921831026673317, + 0.0029776496812701225, + 0.002999892458319664, + 0.003301369259133935, + 0.002811914077028632, + 0.003342284355312586, + 0.0035469706635922194, + 0.028226187452673912, + 0.004568629432469606, + 0.00460760947316885, + 0.002733423840254545, + 0.027084924280643463, + 0.004264078103005886, + 0.003067175392061472, + 0.004371785093098879, + 0.00489843962714076, + 0.00470517436042428, + 0.004706197418272495, + 0.003094919491559267, + 0.004347663838416338, + 0.004553703125566244, + 0.013931162655353546, + 0.011284458450973034, + 0.09183677285909653, + 0.13215026259422302, + 0.005295805167406797, + 0.013148908503353596, + 0.026723692193627357, + 0.026368116959929466, + 0.0017746203811839223, + 0.001675863633863628, + 0.007437788415700197, + 0.003318218747153878, + 0.0018124126363545656, + 0.002028213581070304, + 0.0021955606061965227, + 0.006312444340437651, + 0.0020864943508058786, + 0.0033265426754951477, + 0.0034299741964787245, + 0.002614299999549985, + 0.004211824387311935, + 0.003374769352376461, + 0.003787203226238489, + 0.00423888023942709, + 0.002317672362551093, + 0.00341673381626606, + 0.003955652937293053, + 0.004979425109922886, + 0.0044532716274261475, + 0.004038169514387846, + 0.005538344848901033, + 0.006337733007967472, + 0.0010659695835784078, + 0.001247784006409347, + 0.0016848973464220762, + 0.00167840498033911, + 0.0017051394097507, + 0.002040883293375373, + 0.00222103763371706, + 0.002144202357158065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.08902604877948761 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An uninterrupted, long shot, shot from a level perspective, pans rightward at an incredibly slow, steady, and deliberate pace, capturing the ethereal scene of a reed marsh at dawn. A milky white mist, like a veil, permeates, rises, and entwines among the dense reeds, blurring the boundary between sky and water. The entire scene is imbued with a delicate, ink-and-wash hue, imbued with the beauty of classical Chinese painting. As the camera pans, the viewer's gaze sweeps across this vast, tranquil wetland. Each slender reed stem, a medley of withered yellow and fresh green, is festooned with crystal-clear dewdrops, gleaming like tiny diamonds in the soft, barely visible light of the dawn's first rays. The reed catkins, fluffy as snow, remain motionless in the nearly still air. The calm water reflects the reeds, blending with the shifting mist to create a dreamlike, layered effect. This slow and continuous camera movement, combined with the visible slow flow of fog, together renders an ultimate tranquility and Zen feeling, as if time has frozen at this moment, forming a flowing, poetic and picturesque classical ink painting scroll.", + "seed": 25822, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_9/index_7441_seed_25822_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03958086781203747, + "vlm_logprob": -0.03958086781203747, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.03958086781203747, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0501951202750206 + ], + "opposite_logprob_tokens": [ + -3.5501952171325684 + ], + "logprob_reward": -0.0501951202750206, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0501951202750206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16139957308769226, + 0.04604366049170494, + 0.06817467510700226, + 0.08729121834039688, + 0.11231502145528793, + 0.07200480997562408, + 0.06975359469652176, + 0.3123895823955536, + 0.040393389761447906, + 0.030234484001994133 + ], + "credit_map": [ + 0.01811329834163189, + 0.015232335776090622, + 0.01665205880999565, + 0.030036412179470062, + 0.025159481912851334, + 0.016353214159607887, + 0.018635308369994164, + 0.0212174654006958, + 0.0034650438465178013, + 0.00406511640176177, + 0.0050977496430277824, + 0.010159756056964397, + 0.004501139745116234, + 0.004374018404632807, + 0.006033004727214575, + 0.00834782887250185, + 0.006394284311681986, + 0.007062363903969526, + 0.007508571725338697, + 0.009026308543980122, + 0.00876864418387413, + 0.010723838582634926, + 0.010609135963022709, + 0.008081520907580853, + 0.007242119871079922, + 0.0073767537251114845, + 0.012225457467138767, + 0.013766109943389893, + 0.010495488531887531, + 0.012407529167830944, + 0.015657030045986176, + 0.00812072865664959, + 0.009362512268126011, + 0.008944773115217686, + 0.014031213708221912, + 0.020372381433844566, + 0.010400085709989071, + 0.010360326617956161, + 0.028269169852137566, + 0.010574561543762684, + 0.00828273594379425, + 0.009638456627726555, + 0.010864985175430775, + 0.00909673236310482, + 0.007665187586098909, + 0.00923626497387886, + 0.010436211712658405, + 0.006784237455576658, + 0.006694515701383352, + 0.005712243728339672, + 0.009989249520003796, + 0.011615060269832611, + 0.009102444164454937, + 0.0074935853481292725, + 0.008228416554629803, + 0.010918079875409603, + 0.027428802102804184, + 0.07633218169212341, + 0.032428842037916183, + 0.03240502253174782, + 0.041458044201135635, + 0.03160553425550461, + 0.042128901928663254, + 0.028602251783013344, + 0.0034443228505551815, + 0.007117603439837694, + 0.0068513466976583, + 0.007040651980787516, + 0.004325857385993004, + 0.0048390040174126625, + 0.0033637371379882097, + 0.003410863457247615, + 0.0028778654523193836, + 0.004321138374507427, + 0.004412914626300335, + 0.0049827792681753635, + 0.003579698735848069, + 0.0033600085880607367, + 0.0033800669480115175, + 0.003320014802739024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.0501951202750206 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012570316903293133 + ], + "opposite_logprob_tokens": [ + -4.387570381164551 + ], + "logprob_reward": -0.012570316903293133, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012570316903293133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2542480528354645, + 0.1451573669910431, + 0.07897842675447464, + 0.052118610590696335, + 0.08351683616638184, + 0.06715153902769089, + 0.04000159353017807, + 0.07082954794168472, + 0.16056598722934723, + 0.04743203520774841 + ], + "credit_map": [ + 0.020138703286647797, + 0.024202246218919754, + 0.02435493841767311, + 0.05447540804743767, + 0.03412124887108803, + 0.031480319797992706, + 0.032524459064006805, + 0.032950740307569504, + 0.012147963047027588, + 0.012911232188344002, + 0.022265994921326637, + 0.02478448860347271, + 0.010204292833805084, + 0.01600097119808197, + 0.022004997357726097, + 0.024837428703904152, + 0.005645287223160267, + 0.004385854117572308, + 0.015487514436244965, + 0.02195190079510212, + 0.006929569877684116, + 0.005432440433651209, + 0.0117258932441473, + 0.007419964764267206, + 0.00576451001688838, + 0.006087699439376593, + 0.006222507916390896, + 0.006226988509297371, + 0.008650427684187889, + 0.00872859824448824, + 0.005012350622564554, + 0.005425528157502413, + 0.005923525895923376, + 0.00806740764528513, + 0.008195793256163597, + 0.012439321726560593, + 0.012696223333477974, + 0.01376914419233799, + 0.01462728250771761, + 0.00779813202098012, + 0.006494688335806131, + 0.0060410648584365845, + 0.010245169512927532, + 0.00946253165602684, + 0.007241014391183853, + 0.009139359928667545, + 0.01022112276405096, + 0.008306587114930153, + 0.0036128510255366564, + 0.004517625086009502, + 0.00655245129019022, + 0.005713238380849361, + 0.0061347088776528835, + 0.004669861402362585, + 0.004715565126389265, + 0.004085293971002102, + 0.007353147957473993, + 0.01203098613768816, + 0.007107395678758621, + 0.007096645887941122, + 0.013164255768060684, + 0.009040786884725094, + 0.00719233974814415, + 0.007843996398150921, + 0.013105352409183979, + 0.024767562747001648, + 0.024932723492383957, + 0.024505488574504852, + 0.017406148836016655, + 0.025211377069354057, + 0.015077058225870132, + 0.015560288913547993, + 0.00477914372459054, + 0.005828703287988901, + 0.006087271962314844, + 0.0054995222017169, + 0.007595349103212357, + 0.006415227428078651, + 0.005770191550254822, + 0.005456625949591398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.012570316903293133 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078632436692715 + ], + "opposite_logprob_tokens": [ + -4.511078834533691 + ], + "logprob_reward": -0.011078632436692715, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078632436692715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14500084519386292, + 0.08897297829389572, + 0.12125413119792938, + 0.046550605446100235, + 0.0646631047129631, + 0.11825577169656754, + 0.1950470507144928, + 0.09619097411632538, + 0.0839812383055687, + 0.04008326679468155 + ], + "credit_map": [ + 0.019234688952565193, + 0.014398999512195587, + 0.0181001927703619, + 0.021513232961297035, + 0.021278247237205505, + 0.015961630269885063, + 0.01730094477534294, + 0.01721292734146118, + 0.0073282294906675816, + 0.007093217223882675, + 0.011407340876758099, + 0.010618356056511402, + 0.009214665740728378, + 0.00612748833373189, + 0.020403379574418068, + 0.01678031124174595, + 0.008851180784404278, + 0.009310194291174412, + 0.02216874435544014, + 0.026356935501098633, + 0.015162393450737, + 0.01147893350571394, + 0.016975941136479378, + 0.010949811898171902, + 0.004808616824448109, + 0.00519888661801815, + 0.005762163549661636, + 0.0064650410786271095, + 0.005126218777149916, + 0.006201296579092741, + 0.006785009987652302, + 0.006203376688063145, + 0.006350062321871519, + 0.007038528099656105, + 0.0055975946597754955, + 0.007606814615428448, + 0.013205006718635559, + 0.01036099623888731, + 0.008296238258481026, + 0.006207872182130814, + 0.014288434758782387, + 0.016897307708859444, + 0.016494818031787872, + 0.013259191066026688, + 0.013705198653042316, + 0.018061572685837746, + 0.013921361416578293, + 0.011627892963588238, + 0.018197795376181602, + 0.02456311322748661, + 0.025106973946094513, + 0.02814159169793129, + 0.03263068571686745, + 0.025593498721718788, + 0.02207258902490139, + 0.018740827217698097, + 0.009600567631423473, + 0.020568856969475746, + 0.013286774978041649, + 0.009681317023932934, + 0.012439469806849957, + 0.010535883717238903, + 0.011032963171601295, + 0.009045151993632317, + 0.00785418227314949, + 0.01252228394150734, + 0.014018900692462921, + 0.014191828668117523, + 0.009381587617099285, + 0.01187349297106266, + 0.007247607689350843, + 0.006891361437737942, + 0.004053953569382429, + 0.005229749251157045, + 0.004924172069877386, + 0.004730643704533577, + 0.00720249954611063, + 0.004337493795901537, + 0.005293156951665878, + 0.004311601165682077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.011078632436692715 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202733516693115 + ], + "opposite_logprob_tokens": [ + -2.8120274543762207 + ], + "logprob_reward": -0.06202733516693115, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202733516693115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1841074824333191, + 0.0822863057255745, + 0.15544460713863373, + 0.10530362278223038, + 0.08053246885538101, + 0.11720316857099533, + 0.1643076092004776, + 0.06368425488471985, + 0.023780668154358864, + 0.023349806666374207 + ], + "credit_map": [ + 0.017870664596557617, + 0.013962977565824986, + 0.021477721631526947, + 0.03076637163758278, + 0.029020443558692932, + 0.021679654717445374, + 0.02327892556786537, + 0.02605072408914566, + 0.009054549038410187, + 0.006832863669842482, + 0.012563520111143589, + 0.012552401050925255, + 0.006241391412913799, + 0.011539340950548649, + 0.009379676543176174, + 0.014122560620307922, + 0.012402474880218506, + 0.013007508590817451, + 0.023495757952332497, + 0.022349823266267776, + 0.016563115641474724, + 0.024543337523937225, + 0.024209819734096527, + 0.01887277327477932, + 0.008879149332642555, + 0.010541537776589394, + 0.011291387490928173, + 0.015699593350291252, + 0.01638416200876236, + 0.0127464160323143, + 0.019707217812538147, + 0.010054158978164196, + 0.007248827256262302, + 0.007936428301036358, + 0.011666543781757355, + 0.010860875248908997, + 0.010303959250450134, + 0.012407828122377396, + 0.012633580714464188, + 0.007474422920495272, + 0.009698392823338509, + 0.011199764907360077, + 0.0193302221596241, + 0.020801108330488205, + 0.008226409554481506, + 0.01348448172211647, + 0.024037335067987442, + 0.010425446555018425, + 0.014758194796741009, + 0.0219360813498497, + 0.021564103662967682, + 0.023178473114967346, + 0.025016263127326965, + 0.01870911195874214, + 0.02202855795621872, + 0.017116831615567207, + 0.005879056174308062, + 0.010291148908436298, + 0.011526819318532944, + 0.0062944889068603516, + 0.007009452674537897, + 0.006585224065929651, + 0.008264273405075073, + 0.007833786308765411, + 0.002399916760623455, + 0.0041461531072855, + 0.004616935737431049, + 0.0033222087658941746, + 0.0030872083734720945, + 0.0018918722635135055, + 0.0025117769837379456, + 0.0018045947654172778, + 0.003040787996724248, + 0.0027869900222867727, + 0.003807083936408162, + 0.0028560394421219826, + 0.0022834218107163906, + 0.0022737388499081135, + 0.003042984986677766, + 0.0032587593887001276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.06202733516693115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203293427824974 + ], + "opposite_logprob_tokens": [ + -2.81203293800354 + ], + "logprob_reward": -0.06203293427824974, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203293427824974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2973851263523102, + 0.0787871703505516, + 0.08077039569616318, + 0.04856349155306816, + 0.0901135727763176, + 0.19855457544326782, + 0.060320645570755005, + 0.07744481414556503, + 0.03536555543541908, + 0.032694634050130844 + ], + "credit_map": [ + 0.026857106015086174, + 0.025368236005306244, + 0.02796032279729843, + 0.07003138959407806, + 0.050805073231458664, + 0.02696775272488594, + 0.04165508598089218, + 0.02774018608033657, + 0.007008285261690617, + 0.007613453082740307, + 0.009575908072292805, + 0.00890212319791317, + 0.007474679499864578, + 0.0093778595328331, + 0.016406629234552383, + 0.012428236193954945, + 0.006529137026518583, + 0.006674212869256735, + 0.012593347579240799, + 0.016078133136034012, + 0.009146418422460556, + 0.007559814490377903, + 0.01343637052923441, + 0.008752965368330479, + 0.004912654869258404, + 0.005206603556871414, + 0.007460000924766064, + 0.007884888909757137, + 0.0045168474316596985, + 0.006814463529735804, + 0.006750823929905891, + 0.005017211195081472, + 0.006914537865668535, + 0.009937530383467674, + 0.00673034368082881, + 0.014495393261313438, + 0.01209673285484314, + 0.011145873926579952, + 0.021680083125829697, + 0.007113077212125063, + 0.017100835219025612, + 0.020930804312229156, + 0.03412942588329315, + 0.03233446180820465, + 0.014307042583823204, + 0.02533159963786602, + 0.03595808148384094, + 0.01846233382821083, + 0.005676073022186756, + 0.006690287031233311, + 0.011039352044463158, + 0.006311579141765833, + 0.010311855934560299, + 0.0063469610176980495, + 0.00757987005636096, + 0.006364668253809214, + 0.006493985652923584, + 0.013252295553684235, + 0.009619461372494698, + 0.008035912178456783, + 0.01217257883399725, + 0.00934301596134901, + 0.009723943658173084, + 0.008803623728454113, + 0.004722923040390015, + 0.004955442156642675, + 0.0062188352458179, + 0.004009914118796587, + 0.0039078895933926105, + 0.003996913321316242, + 0.003765560220927, + 0.0037880802992731333, + 0.003301171353086829, + 0.0038984811399132013, + 0.004584691487252712, + 0.0038550111930817366, + 0.0035846338141709566, + 0.002537828404456377, + 0.0055260928347706795, + 0.005406722892075777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.06203293427824974 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An uninterrupted, long shot, shot from a level perspective, pans rightward at an incredibly slow, steady, and deliberate pace, capturing the ethereal scene of a reed marsh at dawn. A milky white mist, like a veil, permeates, rises, and entwines among the dense reeds, blurring the boundary between sky and water. The entire scene is imbued with a delicate, ink-and-wash hue, imbued with the beauty of classical Chinese painting. As the camera pans, the viewer's gaze sweeps across this vast, tranquil wetland. Each slender reed stem, a medley of withered yellow and fresh green, is festooned with crystal-clear dewdrops, gleaming like tiny diamonds in the soft, barely visible light of the dawn's first rays. The reed catkins, fluffy as snow, remain motionless in the nearly still air. The calm water reflects the reeds, blending with the shifting mist to create a dreamlike, layered effect. This slow and continuous camera movement, combined with the visible slow flow of fog, together renders an ultimate tranquility and Zen feeling, as if time has frozen at this moment, forming a flowing, poetic and picturesque classical ink painting scroll.", + "seed": 25823, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000147/group_9/index_7441_seed_25823_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02350002620369196, + "vlm_logprob": -0.02350002620369196, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02350002620369196, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04220276325941086 + ], + "opposite_logprob_tokens": [ + -3.4172027111053467 + ], + "logprob_reward": -0.04220276325941086, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04220276325941086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18359966576099396, + 0.09905736893415451, + 0.05753795802593231, + 0.06420350074768066, + 0.20498773455619812, + 0.1426834911108017, + 0.033781569451093674, + 0.07638172060251236, + 0.08301880955696106, + 0.05474809557199478 + ], + "credit_map": [ + 0.02664712630212307, + 0.03125789016485214, + 0.01976592466235161, + 0.0202010627835989, + 0.026226747781038284, + 0.023049313575029373, + 0.017487913370132446, + 0.01896369829773903, + 0.008028597570955753, + 0.008858900517225266, + 0.007969813421368599, + 0.013662463054060936, + 0.021149586886167526, + 0.014975395984947681, + 0.011039236560463905, + 0.013373387046158314, + 0.0060809943825006485, + 0.006649659480899572, + 0.010352464392781258, + 0.0065313600935041904, + 0.007199488580226898, + 0.005960102658718824, + 0.008426650427281857, + 0.006337241735309362, + 0.007885757833719254, + 0.007898355834186077, + 0.007599476259201765, + 0.00970645435154438, + 0.009723944589495659, + 0.00892640557140112, + 0.006846562959253788, + 0.005616543348878622, + 0.03316795825958252, + 0.02963862009346485, + 0.03442459926009178, + 0.02332778461277485, + 0.018769200891256332, + 0.017967283725738525, + 0.021572552621364594, + 0.026119761168956757, + 0.018593652173876762, + 0.011681592091917992, + 0.03035239689052105, + 0.01528211310505867, + 0.011797999031841755, + 0.01883833110332489, + 0.02357960119843483, + 0.01255781203508377, + 0.0035492638126015663, + 0.0032072942703962326, + 0.005225479602813721, + 0.005006315652281046, + 0.0035918578505516052, + 0.004207025282084942, + 0.00507004652172327, + 0.003924287855625153, + 0.009382238611578941, + 0.00546552799642086, + 0.008936755359172821, + 0.017187191173434258, + 0.007115836720913649, + 0.00850573182106018, + 0.011025207117199898, + 0.008763238787651062, + 0.008954253979027271, + 0.0129042137414217, + 0.011103691533207893, + 0.011198422871530056, + 0.00845550186932087, + 0.008244354277849197, + 0.012975016608834267, + 0.009183360263705254, + 0.005503606051206589, + 0.003644896438345313, + 0.015809671953320503, + 0.007532060611993074, + 0.0038446015678346157, + 0.003211078466847539, + 0.012555070221424103, + 0.002647107932716608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.04220276325941086 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004651559051126242 + ], + "opposite_logprob_tokens": [ + -5.3796515464782715 + ], + "logprob_reward": -0.004651559051126242, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004651559051126242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.312577486038208, + 0.06892672926187515, + 0.11326998472213745, + 0.039211396127939224, + 0.05791432783007622, + 0.043414872139692307, + 0.03880591690540314, + 0.18976697325706482, + 0.08876293152570724, + 0.04734933003783226 + ], + "credit_map": [ + 0.03251553326845169, + 0.05647190660238266, + 0.03081057406961918, + 0.05045585706830025, + 0.025392301380634308, + 0.04294788837432861, + 0.039211031049489975, + 0.034772422164678574, + 0.0067584216594696045, + 0.010910438373684883, + 0.007776019163429737, + 0.008649131283164024, + 0.009873894043266773, + 0.009257427416741848, + 0.0076210773549973965, + 0.00808031391352415, + 0.01562606915831566, + 0.016137786209583282, + 0.020479032769799232, + 0.014912659302353859, + 0.011986571364104748, + 0.01093349326401949, + 0.012674998492002487, + 0.010519370436668396, + 0.004830141086131334, + 0.003964734263718128, + 0.003534568939357996, + 0.004432410467416048, + 0.007290366571396589, + 0.007281460799276829, + 0.004449224565178156, + 0.0034284882713109255, + 0.012577032670378685, + 0.009308567270636559, + 0.0064347586594522, + 0.005310789216309786, + 0.00458112359046936, + 0.0064656492322683334, + 0.007054002955555916, + 0.006182404700666666, + 0.004742383025586605, + 0.004865442402660847, + 0.004843312781304121, + 0.005653807893395424, + 0.004218228626996279, + 0.006813828367739916, + 0.006756009068340063, + 0.005521860904991627, + 0.005585226230323315, + 0.0034567334223538637, + 0.00475319754332304, + 0.004944974556565285, + 0.005366938188672066, + 0.005380891263484955, + 0.00550753902643919, + 0.0038104176055639982, + 0.029780106619000435, + 0.011946974322199821, + 0.02137431874871254, + 0.039779096841812134, + 0.01894301362335682, + 0.02050638757646084, + 0.02617311291396618, + 0.02126396633684635, + 0.012268966063857079, + 0.014385418966412544, + 0.01051087025552988, + 0.012941677123308182, + 0.009513597004115582, + 0.008723927661776543, + 0.01367025077342987, + 0.006748226471245289, + 0.0046543823555111885, + 0.004919200204312801, + 0.007863038219511509, + 0.009430540725588799, + 0.004434754140675068, + 0.004901778884232044, + 0.007685844320803881, + 0.0034597893245518208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.004651559051126242 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947866942733526 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005947866942733526, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947866942733526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17375707626342773, + 0.08533936738967896, + 0.032876137644052505, + 0.034284185618162155, + 0.07233574241399765, + 0.16084907948970795, + 0.09098689258098602, + 0.22276847064495087, + 0.05703272297978401, + 0.06977033615112305 + ], + "credit_map": [ + 0.016795329749584198, + 0.029450425878167152, + 0.011424924246966839, + 0.014988597482442856, + 0.027386507019400597, + 0.037592459470033646, + 0.018964538350701332, + 0.01715429313480854, + 0.00715960469096899, + 0.009294667281210423, + 0.006468606181442738, + 0.010592170059680939, + 0.016001438722014427, + 0.014196502976119518, + 0.00809012446552515, + 0.013536245562136173, + 0.005081954877823591, + 0.00578331807628274, + 0.0035243132151663303, + 0.003061763010919094, + 0.0033224341459572315, + 0.0037028181832283735, + 0.0030184630304574966, + 0.005381070077419281, + 0.004313899204134941, + 0.004048637114465237, + 0.0038077987264841795, + 0.0048135207034647465, + 0.004139683675020933, + 0.005096909124404192, + 0.0032524976413697004, + 0.004811236634850502, + 0.01438105572015047, + 0.010851173661649227, + 0.010794449597597122, + 0.007850592024624348, + 0.005818654783070087, + 0.007166771218180656, + 0.006744467653334141, + 0.008728579618036747, + 0.013973026536405087, + 0.010671086609363556, + 0.027878765016794205, + 0.012121970765292645, + 0.008849309757351875, + 0.008441136218607426, + 0.05157959833741188, + 0.027334176003932953, + 0.01865728385746479, + 0.00960210058838129, + 0.011643531732261181, + 0.012807602062821388, + 0.00726153701543808, + 0.011107968166470528, + 0.01073650736361742, + 0.009170354343950748, + 0.03652101382613182, + 0.014932033605873585, + 0.025961916893720627, + 0.04653070494532585, + 0.021921774372458458, + 0.022826168686151505, + 0.030090562999248505, + 0.023984264582395554, + 0.004469543229788542, + 0.0077024344354867935, + 0.0071284472942352295, + 0.00874307006597519, + 0.005690069869160652, + 0.007427804637700319, + 0.009159490466117859, + 0.0067118569277226925, + 0.006565014366060495, + 0.006527271121740341, + 0.019749848172068596, + 0.007158529479056597, + 0.005897690542042255, + 0.005416301544755697, + 0.015460104681551456, + 0.0029955666977912188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.005947866942733526 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606144569814205 + ], + "opposite_logprob_tokens": [ + -4.141061305999756 + ], + "logprob_reward": -0.01606144569814205, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606144569814205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19106781482696533, + 0.10036879777908325, + 0.10678770393133163, + 0.042068857699632645, + 0.05365665629506111, + 0.14373473823070526, + 0.03742711991071701, + 0.18009281158447266, + 0.11423532664775848, + 0.030560161918401718 + ], + "credit_map": [ + 0.033594630658626556, + 0.0385969802737236, + 0.019627513363957405, + 0.022473221644759178, + 0.01708579994738102, + 0.022537587210536003, + 0.01828625611960888, + 0.018865827471017838, + 0.010823065415024757, + 0.012882065027952194, + 0.014550692401826382, + 0.01001968327909708, + 0.016026057302951813, + 0.013004857115447521, + 0.01072615385055542, + 0.01233622431755066, + 0.016217943280935287, + 0.01593996398150921, + 0.017726119607686996, + 0.011445506475865841, + 0.00847476627677679, + 0.012199163436889648, + 0.011545663699507713, + 0.013238588348031044, + 0.00550113245844841, + 0.005620501935482025, + 0.0037677150685340166, + 0.00642442237585783, + 0.006582765839993954, + 0.005215715151280165, + 0.004052176605910063, + 0.004904426168650389, + 0.008036620914936066, + 0.012008662335574627, + 0.006616410333663225, + 0.005678398068994284, + 0.004662605002522469, + 0.005562512204051018, + 0.006981052458286285, + 0.004110396374017, + 0.022370807826519012, + 0.009482483379542828, + 0.03500988706946373, + 0.015478674322366714, + 0.011340326629579067, + 0.017416847869753838, + 0.020988790318369865, + 0.011646922677755356, + 0.004040761385113001, + 0.004151853267103434, + 0.00722685968503356, + 0.004775342531502247, + 0.0036626984365284443, + 0.00500232819467783, + 0.0047109052538871765, + 0.003856371156871319, + 0.026625629514455795, + 0.012913120910525322, + 0.022935038432478905, + 0.034843605011701584, + 0.017030691727995872, + 0.021214421838521957, + 0.02496379055082798, + 0.019566526636481285, + 0.011963268741965294, + 0.018881667405366898, + 0.013944956474006176, + 0.01919660158455372, + 0.01179527584463358, + 0.013483679853379726, + 0.017819903790950775, + 0.007149973884224892, + 0.0027300133369863033, + 0.0021217374596744776, + 0.008880611509084702, + 0.004926146939396858, + 0.002170019084587693, + 0.002959976438432932, + 0.004673360381275415, + 0.002098298165947199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.01606144569814205 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a reed marsh at dawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the reed stems colored with a mix of withered yellow and fresh green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered with an ink-and-wash hue resembling classical Chinese painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan rightward at a slow, steady, and deliberate pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863649606704712 + ], + "opposite_logprob_tokens": [ + -3.0486364364624023 + ], + "logprob_reward": -0.04863649606704712, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863649606704712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42621374130249023, + 0.04956361651420593, + 0.048270367085933685, + 0.06390241533517838, + 0.16866624355316162, + 0.04496543109416962, + 0.044138237833976746, + 0.061841931194067, + 0.04797116294503212, + 0.04446687176823616 + ], + "credit_map": [ + 0.09588631987571716, + 0.06337123364210129, + 0.051975611597299576, + 0.05574321374297142, + 0.03222315013408661, + 0.04378097504377365, + 0.04465075209736824, + 0.03858248516917229, + 0.00612766994163394, + 0.006157920230180025, + 0.004967967513948679, + 0.007928132079541683, + 0.006684147287160158, + 0.006523325107991695, + 0.004757900722324848, + 0.006416548974812031, + 0.007655653636902571, + 0.006587468087673187, + 0.006316616665571928, + 0.005703735630959272, + 0.006194593850523233, + 0.006078651174902916, + 0.005464814603328705, + 0.004268829710781574, + 0.009384366683661938, + 0.010153665207326412, + 0.005553999915719032, + 0.008055253885686398, + 0.008720699697732925, + 0.009716153144836426, + 0.006425647530704737, + 0.005892619024962187, + 0.030117901042103767, + 0.02304740436375141, + 0.021746156737208366, + 0.020131096243858337, + 0.014503185637295246, + 0.014208794571459293, + 0.020325124263763428, + 0.024586554616689682, + 0.008609230630099773, + 0.0036230955738574266, + 0.004853959660977125, + 0.004668273963034153, + 0.0029964481946080923, + 0.005031343549489975, + 0.008689133450388908, + 0.006493941880762577, + 0.005767687223851681, + 0.004114699549973011, + 0.006318515166640282, + 0.006051471456885338, + 0.004616270307451487, + 0.006329937372356653, + 0.005674051120877266, + 0.005265603307634592, + 0.0108676478266716, + 0.004978011827915907, + 0.00702520040795207, + 0.009699433110654354, + 0.007270483300089836, + 0.007194599602371454, + 0.00739265326410532, + 0.007413896732032299, + 0.004427384585142136, + 0.005687646102160215, + 0.007845045998692513, + 0.006889338605105877, + 0.0057162027806043625, + 0.005557479802519083, + 0.008316646330058575, + 0.0035314150154590607, + 0.00834745168685913, + 0.0036950535140931606, + 0.009634854272007942, + 0.005323271732777357, + 0.003917999565601349, + 0.004024358466267586, + 0.00654813926666975, + 0.0029757400043308735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.04863649606704712 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_0/index_tensor([2876, 2876])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_0/index_tensor([2876, 2876])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ce53dc3d1f3c8d318128b9a99a64e5afc6f8558c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_0/index_tensor([2876, 2876])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Beginning with a low horizon perspective where the water meets the sky, the camera tilts upward in a slow, seamless motion. The frame is initially dominated by the evening sky, a deep indigo at the zenith, transitioning to soft violet and magenta below, before finally burning with a fiery orange-gold afterglow at the horizon. These vibrant colors are reflected on the calm, mirror-like water in the foreground. As the camera steadily ascends, the outline of a majestic city skyline first pierces the horizon, the spires and glass facades of modern skyscrapers coming into view, reflecting the sky's final magical glow. Then, a dense jungle of steel, glass, and concrete is fully revealed. Countless office windows begin to illuminate with warm yellow lights, scattered like stars across the building facades, creating a stark contrast against the dark glass. Finally, the camera stabilizes, forming a grand panoramic picture: below is the entire city shrouded in twilight, with lights just coming on, and light bands formed by traffic flowing silently on the streets, while above is the vast sky with rich colors that is about to be completely swallowed by the night. The whole scene, in a quiet atmosphere of cold colors, reveals an inner vitality and awe-inspiring magnificent beauty of a city about to awaken.", + "seed": 14120, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_0/index_2876_seed_14120_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.010827196249738335, + "vlm_logprob": -0.010827196249738335, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.010827196249738335, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009968851692974567 + ], + "opposite_logprob_tokens": [ + -6.6349687576293945 + ], + "logprob_reward": -0.009968851692974567, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.009968851692974567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22856581211090088, + 0.08422839641571045, + 0.15106043219566345, + 0.09523194283246994, + 0.06203773245215416, + 0.11331994831562042, + 0.07505597919225693, + 0.07668334245681763, + 0.04648405686020851, + 0.06733240187168121 + ], + "credit_map": [ + 0.05429708957672119, + 0.024355042725801468, + 0.027381855994462967, + 0.03268779069185257, + 0.02090592496097088, + 0.013837876729667187, + 0.0238205473870039, + 0.031279679387807846, + 0.007007315754890442, + 0.011296073906123638, + 0.011435315012931824, + 0.014085236936807632, + 0.010556701570749283, + 0.006460898090153933, + 0.015590083785355091, + 0.00779677601531148, + 0.011199405416846275, + 0.042961783707141876, + 0.012489343993365765, + 0.021549487486481667, + 0.016518190503120422, + 0.019422180950641632, + 0.0180994663387537, + 0.008820567280054092, + 0.011809579096734524, + 0.00991017371416092, + 0.009041422046720982, + 0.008124470710754395, + 0.01358742918819189, + 0.017327329143881798, + 0.007790326606482267, + 0.0176412183791399, + 0.006711533758789301, + 0.01038442924618721, + 0.005283623468130827, + 0.005665432196110487, + 0.011526267975568771, + 0.008110416121780872, + 0.007217470556497574, + 0.007138560060411692, + 0.02595035918056965, + 0.015947675332427025, + 0.009964773431420326, + 0.00966058112680912, + 0.017137732356786728, + 0.012239104136824608, + 0.009790792129933834, + 0.012628927826881409, + 0.01585397869348526, + 0.01110780704766512, + 0.0058807749301195145, + 0.006680663675069809, + 0.009704962372779846, + 0.010264403186738491, + 0.008232788182795048, + 0.007330604828894138, + 0.00904754363000393, + 0.008639181032776833, + 0.007240632548928261, + 0.006161328870803118, + 0.013101187534630299, + 0.011884643696248531, + 0.013179593719542027, + 0.007429235614836216, + 0.00515986792743206, + 0.005470042582601309, + 0.004399449564516544, + 0.005663485266268253, + 0.0064612808637320995, + 0.009148402139544487, + 0.004841097630560398, + 0.005340432282537222, + 0.009708051569759846, + 0.007432897109538317, + 0.006501851603388786, + 0.00883239135146141, + 0.007590068504214287, + 0.009317605756223202, + 0.010404512286186218, + 0.007545022293925285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.009968851692974567 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016084320843219757 + ], + "opposite_logprob_tokens": [ + -4.14108419418335 + ], + "logprob_reward": -0.016084320843219757, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016084320843219757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3539785146713257, + 0.08310631662607193, + 0.0708693265914917, + 0.1091267466545105, + 0.06822692602872849, + 0.06885745376348495, + 0.06980810314416885, + 0.0883597657084465, + 0.04468004032969475, + 0.04298676922917366 + ], + "credit_map": [ + 0.03155503422021866, + 0.09636133909225464, + 0.07689489424228668, + 0.034840989857912064, + 0.04077180474996567, + 0.027872787788510323, + 0.01821552962064743, + 0.027466150000691414, + 0.00729047367349267, + 0.013713756576180458, + 0.009399048052728176, + 0.012591286562383175, + 0.01058646384626627, + 0.006966535467654467, + 0.012047378346323967, + 0.010511389002203941, + 0.005552672315388918, + 0.015532135963439941, + 0.006076090969145298, + 0.007857942022383213, + 0.008349012583494186, + 0.006627620197832584, + 0.011705417186021805, + 0.009168439544737339, + 0.015260784886777401, + 0.017295317724347115, + 0.015383478254079819, + 0.016043540090322495, + 0.013282330706715584, + 0.013078899122774601, + 0.0094386525452137, + 0.009343752637505531, + 0.010799253359436989, + 0.007277973461896181, + 0.006436090916395187, + 0.006032337434589863, + 0.009893576614558697, + 0.009201116859912872, + 0.007549992762506008, + 0.011036593466997147, + 0.013232579454779625, + 0.007746897637844086, + 0.005390111356973648, + 0.005450052209198475, + 0.006458376534283161, + 0.009397689253091812, + 0.006676384247839451, + 0.014505368657410145, + 0.0074930028058588505, + 0.008352220989763737, + 0.008963489904999733, + 0.0093419523909688, + 0.00698405085131526, + 0.00971299223601818, + 0.006094771437346935, + 0.012865619733929634, + 0.01137808058410883, + 0.009533716365695, + 0.009200051426887512, + 0.007475041784346104, + 0.012352325022220612, + 0.01524971891194582, + 0.01240965910255909, + 0.010761174373328686, + 0.005939334165304899, + 0.0059284972958266735, + 0.004145435057580471, + 0.004474777728319168, + 0.005320267751812935, + 0.009240806102752686, + 0.004250974394381046, + 0.005379950627684593, + 0.004792406223714352, + 0.004120992496609688, + 0.003805223386734724, + 0.007793988101184368, + 0.004659030586481094, + 0.005187483504414558, + 0.0077840411104261875, + 0.004843604750931263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.016084320843219757 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080989614129066 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011080989614129066, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080989614129066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3632046580314636, + 0.07396724075078964, + 0.07253244519233704, + 0.23291195929050446, + 0.02607746049761772, + 0.03455336391925812, + 0.06904393434524536, + 0.09063026309013367, + 0.021740948781371117, + 0.015337777324020863 + ], + "credit_map": [ + 0.054008111357688904, + 0.057634998112916946, + 0.1136469617486, + 0.04694608226418495, + 0.024522285908460617, + 0.020773833617568016, + 0.01634039729833603, + 0.029331935569643974, + 0.00727200647816062, + 0.008987520821392536, + 0.01506262831389904, + 0.011798940598964691, + 0.0100431889295578, + 0.006265956908464432, + 0.008129863999783993, + 0.00640712957829237, + 0.006938560865819454, + 0.006473104935139418, + 0.010368486866354942, + 0.015404656529426575, + 0.006379684433341026, + 0.004896195139735937, + 0.013323022983968258, + 0.008748727850615978, + 0.012743733823299408, + 0.02123447135090828, + 0.022418728098273277, + 0.011514260433614254, + 0.03706790506839752, + 0.05208270996809006, + 0.026342831552028656, + 0.049507323652505875, + 0.004105512518435717, + 0.003494695294648409, + 0.002072771079838276, + 0.0023747484665364027, + 0.004584044683724642, + 0.0032491469755768776, + 0.00262509030289948, + 0.003571451175957918, + 0.007567835971713066, + 0.004900761879980564, + 0.0028252277988940477, + 0.0028047063387930393, + 0.004122094716876745, + 0.0037625283002853394, + 0.0028538976330310106, + 0.005716308951377869, + 0.012431592680513859, + 0.008472410030663013, + 0.0077342982403934, + 0.007915539667010307, + 0.005492950323969126, + 0.005735618062317371, + 0.00658526411280036, + 0.014676256105303764, + 0.011416042223572731, + 0.009703990072011948, + 0.0089771319180727, + 0.007717157248407602, + 0.011502834036946297, + 0.014786837622523308, + 0.01662227138876915, + 0.009903988800942898, + 0.002315299352630973, + 0.0028227288275957108, + 0.002615068107843399, + 0.0022248788736760616, + 0.0023458118084818125, + 0.0037349294871091843, + 0.001952430116944015, + 0.003729799995198846, + 0.002262711524963379, + 0.00170231144875288, + 0.001453431323170662, + 0.001898475456982851, + 0.0020185275934636593, + 0.0022245547734200954, + 0.0018461489817127585, + 0.001931615057401359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.011080989614129066 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002815451705828309 + ], + "opposite_logprob_tokens": [ + -5.877815246582031 + ], + "logprob_reward": -0.002815451705828309, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002815451705828309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12647664546966553, + 0.08018199354410172, + 0.10326084494590759, + 0.05668066442012787, + 0.06567183136940002, + 0.15257492661476135, + 0.17163890600204468, + 0.08426594734191895, + 0.06164475902915001, + 0.09760347008705139 + ], + "credit_map": [ + 0.011181315407156944, + 0.025948630645871162, + 0.021957848221063614, + 0.018724719062447548, + 0.014996897429227829, + 0.00933554396033287, + 0.013546247966587543, + 0.010785448364913464, + 0.005480707157403231, + 0.01044926792383194, + 0.011669394560158253, + 0.013043155893683434, + 0.011854142881929874, + 0.007924921810626984, + 0.011515161022543907, + 0.008245250210165977, + 0.008687400259077549, + 0.029223186895251274, + 0.008346186950802803, + 0.0094681978225708, + 0.01556097436696291, + 0.009794679470360279, + 0.013037268072366714, + 0.009142954833805561, + 0.007260243874043226, + 0.0053902966901659966, + 0.007420988753437996, + 0.0073961359448730946, + 0.007877436466515064, + 0.006947514601051807, + 0.005604271776974201, + 0.008783776313066483, + 0.01117303129285574, + 0.011463248170912266, + 0.008252022787928581, + 0.007574505638331175, + 0.006478330120444298, + 0.007348053622990847, + 0.005201308988034725, + 0.008181332610547543, + 0.017986781895160675, + 0.01958882436156273, + 0.018496382981538773, + 0.01895921677350998, + 0.023363817483186722, + 0.01826985739171505, + 0.0159208532422781, + 0.019989194348454475, + 0.019000321626663208, + 0.021965859457850456, + 0.01429219450801611, + 0.020909052342176437, + 0.030088171362876892, + 0.03102494403719902, + 0.018152259290218353, + 0.016206104308366776, + 0.012366854585707188, + 0.011731794103980064, + 0.006769721396267414, + 0.00764050567522645, + 0.013016167096793652, + 0.013770731166005135, + 0.011035938747227192, + 0.007934234105050564, + 0.0058781737461686134, + 0.0067758685909211636, + 0.006837720982730389, + 0.0067765857093036175, + 0.00906432792544365, + 0.010662668384611607, + 0.0076751201413571835, + 0.007974294945597649, + 0.010815358720719814, + 0.010605559684336185, + 0.009878933429718018, + 0.013876955024898052, + 0.010700667276978493, + 0.014006066136062145, + 0.014821047894656658, + 0.012898885644972324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.002815451705828309 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014186367392539978 + ], + "opposite_logprob_tokens": [ + -4.264186382293701 + ], + "logprob_reward": -0.014186367392539978, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014186367392539978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18912918865680695, + 0.03954457864165306, + 0.05484459176659584, + 0.07370297610759735, + 0.058123376220464706, + 0.26092618703842163, + 0.16253884136676788, + 0.05622801557183266, + 0.04824686795473099, + 0.056715331971645355 + ], + "credit_map": [ + 0.0314222052693367, + 0.035979628562927246, + 0.046578746289014816, + 0.023477939888834953, + 0.01610672101378441, + 0.010717218741774559, + 0.01115681603550911, + 0.013689936138689518, + 0.003367021679878235, + 0.00614072335883975, + 0.004727792460471392, + 0.006642078515142202, + 0.0052388133481144905, + 0.003677475731819868, + 0.0053108008578419685, + 0.00443987688049674, + 0.006321578286588192, + 0.006157588679343462, + 0.00798644870519638, + 0.006918194238096476, + 0.009022585116326809, + 0.004796411842107773, + 0.007805770728737116, + 0.00583602162078023, + 0.010535303503274918, + 0.009449021890759468, + 0.008314580656588078, + 0.009914836846292019, + 0.009093102067708969, + 0.008683698251843452, + 0.010414631105959415, + 0.007297812029719353, + 0.006544839590787888, + 0.010215920396149158, + 0.0038435719907283783, + 0.006382961291819811, + 0.011248190887272358, + 0.007218439597636461, + 0.00628344202414155, + 0.006386017892509699, + 0.06495803594589233, + 0.03433239087462425, + 0.023551495745778084, + 0.017567088827490807, + 0.025777509436011314, + 0.02264167182147503, + 0.02330496534705162, + 0.04879305511713028, + 0.015530865639448166, + 0.018267691135406494, + 0.01876339502632618, + 0.023823212832212448, + 0.01917411759495735, + 0.023023031651973724, + 0.01849382370710373, + 0.025462718680500984, + 0.00496779615059495, + 0.005838328972458839, + 0.00565422885119915, + 0.005054567474871874, + 0.0071169049479067326, + 0.010163095779716969, + 0.010969413444399834, + 0.006463682744652033, + 0.005868925247341394, + 0.005621182732284069, + 0.005041615106165409, + 0.004027476999908686, + 0.007277115248143673, + 0.008735176175832748, + 0.004590755794197321, + 0.007084625773131847, + 0.004955783486366272, + 0.00604221411049366, + 0.007855531759560108, + 0.009175186045467854, + 0.0062380158342421055, + 0.007106040604412556, + 0.008026213385164738, + 0.007316355127841234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.014186367392539978 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Beginning with a low horizon perspective where the water meets the sky, the camera tilts upward in a slow, seamless motion. The frame is initially dominated by the evening sky, a deep indigo at the zenith, transitioning to soft violet and magenta below, before finally burning with a fiery orange-gold afterglow at the horizon. These vibrant colors are reflected on the calm, mirror-like water in the foreground. As the camera steadily ascends, the outline of a majestic city skyline first pierces the horizon, the spires and glass facades of modern skyscrapers coming into view, reflecting the sky's final magical glow. Then, a dense jungle of steel, glass, and concrete is fully revealed. Countless office windows begin to illuminate with warm yellow lights, scattered like stars across the building facades, creating a stark contrast against the dark glass. Finally, the camera stabilizes, forming a grand panoramic picture: below is the entire city shrouded in twilight, with lights just coming on, and light bands formed by traffic flowing silently on the streets, while above is the vast sky with rich colors that is about to be completely swallowed by the night. The whole scene, in a quiet atmosphere of cold colors, reveals an inner vitality and awe-inspiring magnificent beauty of a city about to awaken.", + "seed": 14121, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_0/index_2876_seed_14121_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03311218777671456, + "vlm_logprob": -0.03311218777671456, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03311218777671456, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08230797946453094 + ], + "opposite_logprob_tokens": [ + -2.707308053970337 + ], + "logprob_reward": -0.08230797946453094, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08230797946453094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2801458239555359, + 0.14495967328548431, + 0.09495136886835098, + 0.05759792402386665, + 0.07183556258678436, + 0.13188208639621735, + 0.1237371489405632, + 0.03653566539287567, + 0.020953066647052765, + 0.037401631474494934 + ], + "credit_map": [ + 0.017615878954529762, + 0.045368991792201996, + 0.07771880179643631, + 0.039587680250406265, + 0.02449887990951538, + 0.020435089245438576, + 0.03154173865914345, + 0.023378774523735046, + 0.011453915387392044, + 0.028214823454618454, + 0.015271859243512154, + 0.014847226440906525, + 0.018853336572647095, + 0.012105179019272327, + 0.01317652128636837, + 0.031036805361509323, + 0.007763576228171587, + 0.017757359892129898, + 0.018292183056473732, + 0.009441042318940163, + 0.015838906168937683, + 0.013427875004708767, + 0.005706552416086197, + 0.006723868660628796, + 0.00592231098562479, + 0.009958615526556969, + 0.006533232517540455, + 0.0048370929434895515, + 0.008977034129202366, + 0.009263992309570312, + 0.006394177675247192, + 0.0057114665396511555, + 0.008954644203186035, + 0.012205814011394978, + 0.008370236493647099, + 0.009077130816876888, + 0.00944922212511301, + 0.008000558242201805, + 0.00848369300365448, + 0.007294266019016504, + 0.0284709595143795, + 0.030925285071134567, + 0.012842477299273014, + 0.011719956062734127, + 0.010204523801803589, + 0.009639479219913483, + 0.013008960522711277, + 0.015070442110300064, + 0.013678220100700855, + 0.01104652974754572, + 0.027687255293130875, + 0.027849934995174408, + 0.005218321457505226, + 0.007446668576449156, + 0.015659483149647713, + 0.015150737948715687, + 0.004279401618987322, + 0.0030375097412616014, + 0.005588465370237827, + 0.004918428137898445, + 0.0024756372440606356, + 0.0032272995449602604, + 0.004663111176341772, + 0.008345812559127808, + 0.002219912363216281, + 0.0014633599203079939, + 0.0042482842691242695, + 0.0037611627485603094, + 0.001413619494996965, + 0.0014718339079990983, + 0.003322568256407976, + 0.00305232685059309, + 0.003898950759321451, + 0.0035433254670351744, + 0.005932535044848919, + 0.010053998790681362, + 0.0018059273716062307, + 0.0022576635237783194, + 0.005396408494561911, + 0.004512821324169636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.08230797946453094 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023319752886891365 + ], + "opposite_logprob_tokens": [ + -3.773319721221924 + ], + "logprob_reward": -0.023319752886891365, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023319752886891365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.498711496591568, + 0.13141441345214844, + 0.09590581804513931, + 0.024306220933794975, + 0.026934633031487465, + 0.08096887171268463, + 0.047890618443489075, + 0.034843504428863525, + 0.047086603939533234, + 0.011937784031033516 + ], + "credit_map": [ + 0.03092607669532299, + 0.08061303943395615, + 0.2064431607723236, + 0.06158864125609398, + 0.0373024120926857, + 0.026287851855158806, + 0.029317229986190796, + 0.026233099400997162, + 0.0065741222351789474, + 0.034273602068424225, + 0.008946308866143227, + 0.012556790374219418, + 0.0332946814596653, + 0.012322725728154182, + 0.00760568818077445, + 0.01584049127995968, + 0.006832524202764034, + 0.0119704008102417, + 0.016399018466472626, + 0.007183869834989309, + 0.015151171013712883, + 0.018312819302082062, + 0.00963981356471777, + 0.010416211560368538, + 0.0025986190885305405, + 0.003478716127574444, + 0.0026010526344180107, + 0.002166394144296646, + 0.003671817947179079, + 0.0034356992691755295, + 0.00292607257142663, + 0.0034278505481779575, + 0.0026529254391789436, + 0.00343512324616313, + 0.002607508795335889, + 0.0028115608729422092, + 0.004357405472546816, + 0.0031951568089425564, + 0.005357041954994202, + 0.0025179125368595123, + 0.009402325376868248, + 0.007880812510848045, + 0.009629513137042522, + 0.014821658842265606, + 0.008742617443203926, + 0.008177688345313072, + 0.011971380561590195, + 0.010342876426875591, + 0.00617766659706831, + 0.004330276511609554, + 0.006166534498333931, + 0.007330155931413174, + 0.00312226265668869, + 0.004958410281687975, + 0.008642437867820263, + 0.007162878289818764, + 0.004810866434127092, + 0.003750775009393692, + 0.005042337812483311, + 0.0041239601559937, + 0.0022626551799476147, + 0.003375093685463071, + 0.005270147230476141, + 0.006207669619470835, + 0.004783963318914175, + 0.0030053912196308374, + 0.011655622161924839, + 0.007012600544840097, + 0.003525004256516695, + 0.0036107220221310854, + 0.007518540136516094, + 0.005974763538688421, + 0.001095387153327465, + 0.0009139633039012551, + 0.0021911903750151396, + 0.002490560058504343, + 0.000837246305309236, + 0.0009162190835922956, + 0.001800418016500771, + 0.0016928004333749413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.023319752886891365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421081367880106 + ], + "opposite_logprob_tokens": [ + -4.2642107009887695 + ], + "logprob_reward": -0.01421081367880106, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421081367880106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.406917542219162, + 0.06897192448377609, + 0.03872505947947502, + 0.22574880719184875, + 0.0653390884399414, + 0.06728760153055191, + 0.03517942503094673, + 0.05142077058553696, + 0.021336819976568222, + 0.019072992727160454 + ], + "credit_map": [ + 0.028904424980282784, + 0.06279158592224121, + 0.15523485839366913, + 0.05230257287621498, + 0.030829325318336487, + 0.025378763675689697, + 0.025828929618000984, + 0.025647077709436417, + 0.006189913023263216, + 0.011644742451608181, + 0.006917109712958336, + 0.010641957633197308, + 0.012160674668848515, + 0.006812307983636856, + 0.005372908897697926, + 0.009232309646904469, + 0.0031634049955755472, + 0.004485764540731907, + 0.004413438495248556, + 0.0049748495221138, + 0.008064083755016327, + 0.008397098630666733, + 0.0027932454831898212, + 0.0024331749882549047, + 0.019362693652510643, + 0.06790367513895035, + 0.013081325218081474, + 0.018325461074709892, + 0.052257198840379715, + 0.03171173855662346, + 0.012593054212629795, + 0.010513662360608578, + 0.010224143974483013, + 0.01175973191857338, + 0.005707538686692715, + 0.0074816131964325905, + 0.008629546500742435, + 0.008475657552480698, + 0.007086599711328745, + 0.005974262487143278, + 0.013577367179095745, + 0.014221934601664543, + 0.0070429397746920586, + 0.007778852246701717, + 0.006022075656801462, + 0.0051663825288414955, + 0.006482002791017294, + 0.006996046751737595, + 0.004964883904904127, + 0.004284755326807499, + 0.0045478916727006435, + 0.005063989665359259, + 0.0021995818242430687, + 0.003498840145766735, + 0.005209026858210564, + 0.005410454235970974, + 0.006462533958256245, + 0.004774442873895168, + 0.0065754493698477745, + 0.004840126261115074, + 0.004124519415199757, + 0.005737932398915291, + 0.007585680577903986, + 0.0113200843334198, + 0.0030410378240048885, + 0.0022024211939424276, + 0.002832456724718213, + 0.0023272642865777016, + 0.0012848988408222795, + 0.001620383351109922, + 0.0037658214569091797, + 0.0042625367641448975, + 0.0023241699673235416, + 0.002292259130626917, + 0.005024394020438194, + 0.0027612207923084497, + 0.0008636026177555323, + 0.0011787238763645291, + 0.002307221060618758, + 0.0023213992826640606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.01421081367880106 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626583334058523 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626583334058523, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626583334058523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36951127648353577, + 0.16675330698490143, + 0.06011765077710152, + 0.09072572737932205, + 0.03712943568825722, + 0.10353340953588486, + 0.03141702711582184, + 0.05455177649855614, + 0.033245522528886795, + 0.05301492661237717 + ], + "credit_map": [ + 0.024029672145843506, + 0.06325887143611908, + 0.1379355937242508, + 0.043840572237968445, + 0.032411105930805206, + 0.021618714556097984, + 0.0213481318205595, + 0.02506856434047222, + 0.008534649387001991, + 0.051906079053878784, + 0.013323304243385792, + 0.013198061846196651, + 0.027857216075062752, + 0.014591087587177753, + 0.012312502600252628, + 0.025030389428138733, + 0.005603843834251165, + 0.008867078460752964, + 0.008613904006779194, + 0.005434882827103138, + 0.011024353094398975, + 0.008729021064937115, + 0.006378272548317909, + 0.005466289818286896, + 0.007445471826940775, + 0.022390905767679214, + 0.010221650823950768, + 0.007610922679305077, + 0.017672497779130936, + 0.009891748428344727, + 0.008370375260710716, + 0.00712214969098568, + 0.004594254307448864, + 0.005317092407494783, + 0.0036409145686775446, + 0.004333453718572855, + 0.0054673389531672, + 0.003962076734751463, + 0.005507506430149078, + 0.0043067955411970615, + 0.018086973577737808, + 0.022086305543780327, + 0.01098842453211546, + 0.014219587668776512, + 0.007892226800322533, + 0.006942662410438061, + 0.011728254146873951, + 0.01158897578716278, + 0.004303889814764261, + 0.004026990383863449, + 0.004208559636026621, + 0.004455558490008116, + 0.0019654829520732164, + 0.002712356625124812, + 0.004338162951171398, + 0.005406022537499666, + 0.0058561451733112335, + 0.004040039610117674, + 0.008763052523136139, + 0.006965281441807747, + 0.004698735196143389, + 0.005818112753331661, + 0.007852337323129177, + 0.010558065958321095, + 0.003700986271724105, + 0.0026689914520829916, + 0.005811510607600212, + 0.005216160323470831, + 0.0024867074098438025, + 0.0027862826827913523, + 0.005435783416032791, + 0.005139094777405262, + 0.00495288148522377, + 0.0033266847021877766, + 0.013308587484061718, + 0.0066395895555615425, + 0.0037514381110668182, + 0.004851541016250849, + 0.0072736735455691814, + 0.0089105274528265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.007626583334058523 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038095809519290924 + ], + "opposite_logprob_tokens": [ + -3.288095712661743 + ], + "logprob_reward": -0.038095809519290924, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038095809519290924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44191354513168335, + 0.14008702337741852, + 0.05444996803998947, + 0.03421692177653313, + 0.028326261788606644, + 0.18371427059173584, + 0.023526176810264587, + 0.03654222935438156, + 0.013598610647022724, + 0.04362502694129944 + ], + "credit_map": [ + 0.029203420504927635, + 0.07696089893579483, + 0.16892452538013458, + 0.04926017299294472, + 0.03715142607688904, + 0.025983702391386032, + 0.024294055998325348, + 0.030135350301861763, + 0.006385719869285822, + 0.043874334543943405, + 0.010477663017809391, + 0.01060747541487217, + 0.02361738309264183, + 0.011620643548667431, + 0.0094241788610816, + 0.02407963201403618, + 0.005047612823545933, + 0.007160768378525972, + 0.007508138660341501, + 0.0035266363993287086, + 0.01035295333713293, + 0.011861453764140606, + 0.004517294000834227, + 0.004475107416510582, + 0.003145987167954445, + 0.004529905971139669, + 0.004570912104099989, + 0.003081083996221423, + 0.0034844717010855675, + 0.005626484286040068, + 0.0052988543175160885, + 0.004479222930967808, + 0.002545436378568411, + 0.0033315166365355253, + 0.0034798230044543743, + 0.0034178760834038258, + 0.0028430817183107138, + 0.004061450716108084, + 0.004702444653958082, + 0.003944631665945053, + 0.033730439841747284, + 0.03818070515990257, + 0.016837285831570625, + 0.012648474425077438, + 0.02006126195192337, + 0.01487323921173811, + 0.02014344185590744, + 0.027239428833127022, + 0.002752115949988365, + 0.0027689733542501926, + 0.0037761249113827944, + 0.0031113428995013237, + 0.0014993344666436315, + 0.001967399148270488, + 0.0036581431049853563, + 0.003992742858827114, + 0.0036384917329996824, + 0.0027766888961195946, + 0.006426380015909672, + 0.005305822007358074, + 0.002748158061876893, + 0.002424712758511305, + 0.006635064724832773, + 0.006586908362805843, + 0.0020011060405522585, + 0.0016774531686678529, + 0.002726567443460226, + 0.0016828259686008096, + 0.0008091512718237936, + 0.0008419673540629447, + 0.0018365884898230433, + 0.0020229509100317955, + 0.004777797497808933, + 0.0031126695685088634, + 0.010735358111560345, + 0.00585327809676528, + 0.0026425679679960012, + 0.002973630093038082, + 0.007238416932523251, + 0.006291309371590614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.038095809519290924 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Beginning with a low horizon perspective where the water meets the sky, the camera tilts upward in a slow, seamless motion. The frame is initially dominated by the evening sky, a deep indigo at the zenith, transitioning to soft violet and magenta below, before finally burning with a fiery orange-gold afterglow at the horizon. These vibrant colors are reflected on the calm, mirror-like water in the foreground. As the camera steadily ascends, the outline of a majestic city skyline first pierces the horizon, the spires and glass facades of modern skyscrapers coming into view, reflecting the sky's final magical glow. Then, a dense jungle of steel, glass, and concrete is fully revealed. Countless office windows begin to illuminate with warm yellow lights, scattered like stars across the building facades, creating a stark contrast against the dark glass. Finally, the camera stabilizes, forming a grand panoramic picture: below is the entire city shrouded in twilight, with lights just coming on, and light bands formed by traffic flowing silently on the streets, while above is the vast sky with rich colors that is about to be completely swallowed by the night. The whole scene, in a quiet atmosphere of cold colors, reveals an inner vitality and awe-inspiring magnificent beauty of a city about to awaken.", + "seed": 14122, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_0/index_2876_seed_14122_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010369956819340586, + "vlm_logprob": -0.010369956819340586, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.010369956819340586, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012262902222573757 + ], + "opposite_logprob_tokens": [ + -5.63726282119751 + ], + "logprob_reward": -0.012262902222573757, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.012262902222573757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8994556665420532, + 0.007414340041577816, + 0.004614976700395346, + 0.031512901186943054, + 0.007773017510771751, + 0.012918410822749138, + 0.006073004100471735, + 0.013191396370530128, + 0.013570419512689114, + 0.0034758553374558687 + ], + "credit_map": [ + 0.2544606626033783, + 0.03828514367341995, + 0.036026228219270706, + 0.01630489155650139, + 0.08043220639228821, + 0.4208293557167053, + 0.01916925609111786, + 0.033947937190532684, + 0.0007194110658019781, + 0.0006629397394135594, + 0.0016302949516102672, + 0.00048346424591727555, + 0.0009732458274811506, + 0.0021348088048398495, + 0.0004112787137273699, + 0.0003988964017480612, + 0.0003809044137597084, + 0.0006330504547804594, + 0.0005213613621890545, + 0.0005326859536580741, + 0.0007173567428253591, + 0.0007185246213339269, + 0.00048018398229032755, + 0.000630908936727792, + 0.0026851992588490248, + 0.0045029353350400925, + 0.0038471510633826256, + 0.005036692135035992, + 0.0031052560079842806, + 0.002685911487787962, + 0.004386408254504204, + 0.005263346713036299, + 0.0004839788598474115, + 0.000690954620949924, + 0.001228187931701541, + 0.0008104779990389943, + 0.0008307214011438191, + 0.0008205851190723479, + 0.0011643979232758284, + 0.0017437144415453076, + 0.001332665327936411, + 0.0007151414174586535, + 0.0015361980767920613, + 0.0036578159779310226, + 0.0006887900526635349, + 0.0010073286248371005, + 0.002028480637818575, + 0.0019519899506121874, + 0.0004266671894583851, + 0.00042607670184224844, + 0.0011386045953258872, + 0.0009155612206086516, + 0.0004256006213836372, + 0.0004427769163157791, + 0.0008933518547564745, + 0.001404364942573011, + 0.0007598332013003528, + 0.0009008204215206206, + 0.001423244597390294, + 0.002360195154324174, + 0.0007333039538934827, + 0.0006040643202140927, + 0.003840030636638403, + 0.0025699043180793524, + 0.0016702788416296244, + 0.000711454136762768, + 0.0013764738105237484, + 0.0014169494388625026, + 0.0012220048811286688, + 0.003308985847979784, + 0.0012059835717082024, + 0.0026582886930555105, + 0.0005360033828765154, + 0.0002814672188833356, + 0.00040398837882094085, + 0.0004963276442140341, + 0.0005155852995812893, + 0.00041387471719644964, + 0.00045755007886327803, + 0.0003710585297085345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.012262902222573757 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067506348714232445 + ], + "opposite_logprob_tokens": [ + -5.006750583648682 + ], + "logprob_reward": -0.0067506348714232445, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067506348714232445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6317443251609802, + 0.013423239812254906, + 0.02159142680466175, + 0.028368059545755386, + 0.04901890829205513, + 0.18842510879039764, + 0.008596967905759811, + 0.03167389705777168, + 0.01839480735361576, + 0.008763216435909271 + ], + "credit_map": [ + 0.17861512303352356, + 0.027263591066002846, + 0.025496600195765495, + 0.011003478430211544, + 0.05625756084918976, + 0.29483306407928467, + 0.01377725787460804, + 0.024497682228684425, + 0.0013641213299706578, + 0.002075341297313571, + 0.0014541279524564743, + 0.001288251020014286, + 0.00223512202501297, + 0.0023724231868982315, + 0.0013534326571971178, + 0.0012804224388673902, + 0.0021511344239115715, + 0.0027134583797305822, + 0.0019446442602202296, + 0.0024760232772678137, + 0.0026507924776524305, + 0.004430997651070356, + 0.0020553574431687593, + 0.0031690215691924095, + 0.0027013709768652916, + 0.003912149928510189, + 0.0025315366219729185, + 0.0031128162518143654, + 0.004587541799992323, + 0.004120182245969772, + 0.0030392527114599943, + 0.004363209940493107, + 0.002388589084148407, + 0.0011076363734900951, + 0.02022814005613327, + 0.0033840308897197247, + 0.0012466718908399343, + 0.0012852514628320932, + 0.0038671917282044888, + 0.015511404722929, + 0.011089065112173557, + 0.009519950486719608, + 0.021032162010669708, + 0.06510913372039795, + 0.007734495215117931, + 0.010589062236249447, + 0.025973808020353317, + 0.03737744316458702, + 0.0009254190372303128, + 0.0007393171545118093, + 0.0016961335204541683, + 0.0012457589618861675, + 0.0006201182841323316, + 0.0005985346506349742, + 0.0011416032211855054, + 0.001630083890631795, + 0.001367453602142632, + 0.0017575114034116268, + 0.003145767142996192, + 0.005601098295301199, + 0.0017257921863347292, + 0.0014551636995747685, + 0.01034199632704258, + 0.0062791164964437485, + 0.0021582033950835466, + 0.0009565884247422218, + 0.0015745332930237055, + 0.0020810230635106564, + 0.001780883059836924, + 0.004776922985911369, + 0.0014414876932278275, + 0.003625167766585946, + 0.0015177549794316292, + 0.0008042083936743438, + 0.0010766208870336413, + 0.000926702399738133, + 0.0015657193725928664, + 0.0013064335798844695, + 0.00093273218953982, + 0.0006330448668450117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.0067506348714232445 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674400432035327 + ], + "opposite_logprob_tokens": [ + -5.006743907928467 + ], + "logprob_reward": -0.00674400432035327, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674400432035327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26570385694503784, + 0.045710980892181396, + 0.033300627022981644, + 0.08067036420106888, + 0.10557261854410172, + 0.3000422418117523, + 0.03864055499434471, + 0.055688150227069855, + 0.02127838134765625, + 0.05339230224490166 + ], + "credit_map": [ + 0.06607985496520996, + 0.021067271009087563, + 0.00939986016601324, + 0.011901202611625195, + 0.026348475366830826, + 0.11106231808662415, + 0.007878036238253117, + 0.01196680311113596, + 0.005140590947121382, + 0.007965713739395142, + 0.0051687792874872684, + 0.005026054102927446, + 0.007382601965218782, + 0.0064529795199632645, + 0.004673383664339781, + 0.0039008748717606068, + 0.0030023411381989717, + 0.003999854903668165, + 0.003649624064564705, + 0.004435403738170862, + 0.004124783910810947, + 0.005403886083513498, + 0.0037465679924935102, + 0.004938160069286823, + 0.007005082443356514, + 0.011235782876610756, + 0.011354864574968815, + 0.010945339687168598, + 0.008448993787169456, + 0.010722720064222813, + 0.009974752552807331, + 0.010982817970216274, + 0.005373709835112095, + 0.002418752061203122, + 0.043439000844955444, + 0.00713889067992568, + 0.002477516420185566, + 0.0031841942109167576, + 0.010447919368743896, + 0.03109261952340603, + 0.020177854225039482, + 0.014516741968691349, + 0.0305777657777071, + 0.10552499443292618, + 0.010697607882320881, + 0.016186874359846115, + 0.04004288092255592, + 0.06231746822595596, + 0.002317282138392329, + 0.001998435938730836, + 0.006766442209482193, + 0.005166533403098583, + 0.002503689844161272, + 0.0026896705385297537, + 0.006215307395905256, + 0.01098319049924612, + 0.0027468930929899216, + 0.0034439959563314915, + 0.009140550158917904, + 0.013752580620348454, + 0.0036158908624202013, + 0.004195692483335733, + 0.008100711740553379, + 0.010691828094422817, + 0.0018342607654631138, + 0.001876362250186503, + 0.0027669710107147694, + 0.0036303293891251087, + 0.0024024264421314, + 0.0023205287288874388, + 0.0027716862969100475, + 0.003675812855362892, + 0.007017925847321749, + 0.005519168451428413, + 0.004795973654836416, + 0.005330386105924845, + 0.013225832022726536, + 0.007541740778833628, + 0.006016946863383055, + 0.00394432432949543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.00674400432035327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028161650989204645 + ], + "opposite_logprob_tokens": [ + -5.877816200256348 + ], + "logprob_reward": -0.0028161650989204645, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028161650989204645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26401975750923157, + 0.05002520978450775, + 0.034610647708177567, + 0.07563503086566925, + 0.2597505450248718, + 0.1815110594034195, + 0.02179660089313984, + 0.05069884657859802, + 0.02343173325061798, + 0.038520537316799164 + ], + "credit_map": [ + 0.06715638190507889, + 0.012178566306829453, + 0.011055804789066315, + 0.009866608306765556, + 0.025263508781790733, + 0.11861473321914673, + 0.007350914645940065, + 0.01253324095159769, + 0.006444403901696205, + 0.0068302215076982975, + 0.008408432826399803, + 0.0039755250327289104, + 0.00789965782314539, + 0.006925718393176794, + 0.004357050638645887, + 0.005184199195355177, + 0.0034435223788022995, + 0.003234871895983815, + 0.005145418457686901, + 0.006851293612271547, + 0.0038936578202992678, + 0.004116412252187729, + 0.003620879491791129, + 0.004304592963308096, + 0.006960667669773102, + 0.011242648586630821, + 0.009610723704099655, + 0.011650442145764828, + 0.0091332932934165, + 0.007552860304713249, + 0.00972355529665947, + 0.009760844521224499, + 0.01457882672548294, + 0.007175729610025883, + 0.10501652956008911, + 0.0172458216547966, + 0.006886318791657686, + 0.00858889427036047, + 0.023293180391192436, + 0.07696525752544403, + 0.017298776656389236, + 0.01062498427927494, + 0.021184055134654045, + 0.05463999882340431, + 0.007662955671548843, + 0.013952179811894894, + 0.025046763941645622, + 0.031101340427994728, + 0.001754884491674602, + 0.0017035636119544506, + 0.0038837145548313856, + 0.0029197437688708305, + 0.002082900609821081, + 0.002314816927537322, + 0.003339827759191394, + 0.003797148121520877, + 0.00353134423494339, + 0.004920826759189367, + 0.007618372794240713, + 0.007483278401196003, + 0.004030291456729174, + 0.0023334233555942774, + 0.012191037647426128, + 0.008590271696448326, + 0.0018308847211301327, + 0.0016600920353084803, + 0.0034259434323757887, + 0.003667589044198394, + 0.0025528958067297935, + 0.003060632385313511, + 0.0032883582171052694, + 0.003945336677134037, + 0.004415999166667461, + 0.004051996860653162, + 0.003956273198127747, + 0.003918796312063932, + 0.008882157504558563, + 0.005800370592623949, + 0.0049676718190312386, + 0.0025272711645811796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0028161650989204645 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023276077583432198 + ], + "opposite_logprob_tokens": [ + -3.7732760906219482 + ], + "logprob_reward": -0.023276077583432198, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023276077583432198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24951711297035217, + 0.04896215721964836, + 0.03876621276140213, + 0.18750976026058197, + 0.09759125113487244, + 0.27947768568992615, + 0.01908072456717491, + 0.03719217702746391, + 0.02566833794116974, + 0.01623460464179516 + ], + "credit_map": [ + 0.06149069219827652, + 0.01759868487715721, + 0.011468912474811077, + 0.008637208491563797, + 0.02558036521077156, + 0.1074899211525917, + 0.006629833485931158, + 0.010621496476233006, + 0.00393030047416687, + 0.0033761542290449142, + 0.014730127528309822, + 0.0032989110331982374, + 0.0035634110681712627, + 0.01292287278920412, + 0.0033894639927893877, + 0.0037509165704250336, + 0.0034624277614057064, + 0.004845619201660156, + 0.004171951673924923, + 0.006549421232193708, + 0.004679522011429071, + 0.006396095734089613, + 0.0033229002729058266, + 0.005338276270776987, + 0.01569487527012825, + 0.027133772149682045, + 0.025296306237578392, + 0.02723979577422142, + 0.019397763535380363, + 0.019806940108537674, + 0.024525471031665802, + 0.028414832428097725, + 0.004809534177184105, + 0.0025290900375694036, + 0.03969843313097954, + 0.006362973712384701, + 0.002518428722396493, + 0.0033579538576304913, + 0.010489835403859615, + 0.027824992313981056, + 0.016388392075896263, + 0.013598993420600891, + 0.02841237001121044, + 0.0998801663517952, + 0.010123440995812416, + 0.014623543247580528, + 0.03673884645104408, + 0.059711940586566925, + 0.001243142643943429, + 0.0013533086748793721, + 0.003938428126275539, + 0.0026037516072392464, + 0.0011854228796437383, + 0.0015081739984452724, + 0.0024704961106181145, + 0.004778000526130199, + 0.0016936514293774962, + 0.0021765741985291243, + 0.005319600459188223, + 0.009437205269932747, + 0.0018748656148090959, + 0.0021196629386395216, + 0.007881724275648594, + 0.006688893307000399, + 0.0030162460170686245, + 0.0012746959691867232, + 0.002339309547096491, + 0.002404792234301567, + 0.002526614349335432, + 0.006577315274626017, + 0.0021163704805076122, + 0.005412995349615812, + 0.0026801549829542637, + 0.0017750200349837542, + 0.0013992084423080087, + 0.0016938619082793593, + 0.0034203724935650826, + 0.002325585810467601, + 0.0016949876444414258, + 0.0012454141397029161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.023276077583432198 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Beginning with a low horizon perspective where the water meets the sky, the camera tilts upward in a slow, seamless motion. The frame is initially dominated by the evening sky, a deep indigo at the zenith, transitioning to soft violet and magenta below, before finally burning with a fiery orange-gold afterglow at the horizon. These vibrant colors are reflected on the calm, mirror-like water in the foreground. As the camera steadily ascends, the outline of a majestic city skyline first pierces the horizon, the spires and glass facades of modern skyscrapers coming into view, reflecting the sky's final magical glow. Then, a dense jungle of steel, glass, and concrete is fully revealed. Countless office windows begin to illuminate with warm yellow lights, scattered like stars across the building facades, creating a stark contrast against the dark glass. Finally, the camera stabilizes, forming a grand panoramic picture: below is the entire city shrouded in twilight, with lights just coming on, and light bands formed by traffic flowing silently on the streets, while above is the vast sky with rich colors that is about to be completely swallowed by the night. The whole scene, in a quiet atmosphere of cold colors, reveals an inner vitality and awe-inspiring magnificent beauty of a city about to awaken.", + "seed": 14123, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_0/index_2876_seed_14123_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.27699957564473154, + "vlm_logprob": -0.27699957564473154, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.27699957564473154, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015720486640930176 + ], + "opposite_logprob_tokens": [ + -5.140720367431641 + ], + "logprob_reward": -0.015720486640930176, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.015720486640930176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2602095603942871, + 0.03384403511881828, + 0.10009642690420151, + 0.04507732391357422, + 0.03933292627334595, + 0.04048226401209831, + 0.18712544441223145, + 0.05221232399344444, + 0.16016925871372223, + 0.08145052939653397 + ], + "credit_map": [ + 0.026060346513986588, + 0.04471469298005104, + 0.03866056725382805, + 0.04045934975147247, + 0.029161257669329643, + 0.027629438787698746, + 0.021279923617839813, + 0.032243967056274414, + 0.0039801583625376225, + 0.0026279641315340996, + 0.00449178833514452, + 0.006718844641000032, + 0.004202873911708593, + 0.003472357988357544, + 0.004323025234043598, + 0.004027025308459997, + 0.015296869911253452, + 0.01499869953840971, + 0.017443519085645676, + 0.016672004014253616, + 0.008903664536774158, + 0.009702380746603012, + 0.009983307681977749, + 0.0070959776639938354, + 0.0049920161254704, + 0.004421016201376915, + 0.004786814097315073, + 0.005057397298514843, + 0.006193513981997967, + 0.011646614409983158, + 0.003187913680449128, + 0.0047920397482812405, + 0.00499237235635519, + 0.004230845253914595, + 0.005286962725222111, + 0.003530866466462612, + 0.0044082701206207275, + 0.006120499689131975, + 0.004530677571892738, + 0.006232432089745998, + 0.0034816099796444178, + 0.004543675575405359, + 0.007122534792870283, + 0.005370671860873699, + 0.003967210650444031, + 0.004141146782785654, + 0.00623902166262269, + 0.005616392474621534, + 0.006727318279445171, + 0.004778517875820398, + 0.07623106241226196, + 0.026404790580272675, + 0.0034868072252720594, + 0.0058627380058169365, + 0.008668046444654465, + 0.05496615171432495, + 0.0076879095286130905, + 0.007425051182508469, + 0.006792408414185047, + 0.007318364921957254, + 0.005644917953759432, + 0.005392616149038076, + 0.006060558371245861, + 0.005890498403459787, + 0.0096064992249012, + 0.00847579538822174, + 0.06865882873535156, + 0.03710407763719559, + 0.005766020156443119, + 0.007218139711767435, + 0.009194373153150082, + 0.01414552703499794, + 0.012588650919497013, + 0.010585037991404533, + 0.006737087853252888, + 0.007221822161227465, + 0.008486664853990078, + 0.0211490485817194, + 0.004778977483510971, + 0.009903236292302608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.015720486640930176 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898591458797455 + ], + "opposite_logprob_tokens": [ + -2.5789859294891357 + ], + "logprob_reward": -0.07898591458797455, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898591458797455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24272461235523224, + 0.05907385051250458, + 0.1776949167251587, + 0.08242833614349365, + 0.05804773420095444, + 0.07894061505794525, + 0.038117069751024246, + 0.06527319550514221, + 0.18038494884967804, + 0.017314782366156578 + ], + "credit_map": [ + 0.023264184594154358, + 0.042966362088918686, + 0.036306366324424744, + 0.035995449870824814, + 0.02899668551981449, + 0.02597132883965969, + 0.01899361051619053, + 0.030230624601244926, + 0.006011930759996176, + 0.005089841317385435, + 0.007385325618088245, + 0.00973852165043354, + 0.005989269819110632, + 0.008447221480309963, + 0.006933756172657013, + 0.00947798416018486, + 0.023238113150000572, + 0.023290038108825684, + 0.029498670250177383, + 0.024090837687253952, + 0.022937685251235962, + 0.02350800670683384, + 0.016135036945343018, + 0.014996523037552834, + 0.009672751650214195, + 0.008287779055535793, + 0.00924642663449049, + 0.007765024434775114, + 0.018966814503073692, + 0.013002063147723675, + 0.005722019821405411, + 0.009765448048710823, + 0.005874767433851957, + 0.006397391203790903, + 0.007603474427014589, + 0.0044386861845850945, + 0.007886362262070179, + 0.01003299467265606, + 0.00863872654736042, + 0.007175330072641373, + 0.006912427954375744, + 0.008902542293071747, + 0.010824589058756828, + 0.010885732248425484, + 0.009952248074114323, + 0.011757967062294483, + 0.008861679583787918, + 0.010843426920473576, + 0.004127479624003172, + 0.0032675443217158318, + 0.008037063293159008, + 0.0030658505856990814, + 0.0035091708414256573, + 0.0034400434233248234, + 0.002552012912929058, + 0.010117903351783752, + 0.009309777989983559, + 0.010008782148361206, + 0.00937904417514801, + 0.008998782373964787, + 0.006862407550215721, + 0.007540975231677294, + 0.005877434276044369, + 0.0072959898971021175, + 0.014949146658182144, + 0.008019352331757545, + 0.07807105034589767, + 0.04321334883570671, + 0.006484260316938162, + 0.009353159926831722, + 0.01113134901970625, + 0.009163283742964268, + 0.0021359629463404417, + 0.0017102462006732821, + 0.0016299798153340816, + 0.001957166939973831, + 0.003625007113441825, + 0.002597758313640952, + 0.001622741692699492, + 0.002035920275375247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.07898591458797455 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004648711532354355 + ], + "opposite_logprob_tokens": [ + -5.379648685455322 + ], + "logprob_reward": -0.004648711532354355, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004648711532354355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19140197336673737, + 0.027190668508410454, + 0.058430518954992294, + 0.04696161672472954, + 0.03899854049086571, + 0.09192515164613724, + 0.2267141342163086, + 0.04794254153966904, + 0.18759037554264069, + 0.08284446597099304 + ], + "credit_map": [ + 0.016794458031654358, + 0.02429409697651863, + 0.02808726392686367, + 0.02619396150112152, + 0.020619874820113182, + 0.02270684391260147, + 0.022811872884631157, + 0.029893599450588226, + 0.0025858376175165176, + 0.0021396863739937544, + 0.003892522770911455, + 0.002600244479253888, + 0.004334211815148592, + 0.0029235961847007275, + 0.0037273659836500883, + 0.004987203050404787, + 0.009254763834178448, + 0.008376020938158035, + 0.008983584120869637, + 0.007817858830094337, + 0.0077964733354747295, + 0.00758586497977376, + 0.00423502316698432, + 0.00438093114644289, + 0.005426071118563414, + 0.0036523547023534775, + 0.004983473103493452, + 0.007989073172211647, + 0.006448281463235617, + 0.006802368443459272, + 0.0053855120204389095, + 0.006274482235312462, + 0.003830177476629615, + 0.003761367406696081, + 0.004052683711051941, + 0.0029303771443665028, + 0.006855951622128487, + 0.009096737951040268, + 0.003584248246625066, + 0.004886998329311609, + 0.005892918445169926, + 0.008619008585810661, + 0.01103825494647026, + 0.010800219140946865, + 0.014254994690418243, + 0.018650095909833908, + 0.011186333373188972, + 0.011483327485620975, + 0.00861042644828558, + 0.008039096370339394, + 0.08051875978708267, + 0.02863108180463314, + 0.013778719119727612, + 0.015595723874866962, + 0.011327340267598629, + 0.06021298095583916, + 0.005367632023990154, + 0.007021395023912191, + 0.006667334586381912, + 0.007804287597537041, + 0.006092369090765715, + 0.00560484966263175, + 0.004651469644159079, + 0.004733201581984758, + 0.01454066950827837, + 0.0087687186896801, + 0.07670063525438309, + 0.042409684509038925, + 0.009546250104904175, + 0.013189036399126053, + 0.012566682882606983, + 0.0098686832934618, + 0.0073201609775424, + 0.00733884796500206, + 0.011319329030811787, + 0.010201842524111271, + 0.016482848674058914, + 0.015626760199666023, + 0.006878920365124941, + 0.007675757631659508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.004648711532354355 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2235822677612305 + ], + "opposite_logprob_tokens": [ + -0.34858229756355286 + ], + "logprob_reward": -1.2235822677612305, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2235822677612305, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.032641150057315826, + 0.019477026537060738, + 0.01656985841691494, + 0.052606940269470215, + 0.08651081472635269, + 0.11745896935462952, + 0.36287805438041687, + 0.03034789487719536, + 0.181326225399971, + 0.10018312931060791 + ], + "credit_map": [ + 0.0032211754005402327, + 0.0031661682296544313, + 0.003744496963918209, + 0.005474472418427467, + 0.002561136381700635, + 0.004555337596684694, + 0.005433607380837202, + 0.004484754521399736, + 0.0018433972727507353, + 0.003159633604809642, + 0.002614404307678342, + 0.0024373040068894625, + 0.00228502182289958, + 0.0024608306121081114, + 0.0022990035358816385, + 0.00237743160687387, + 0.0018523322651162744, + 0.002375170588493347, + 0.0024000313133001328, + 0.0024401857517659664, + 0.00219253939576447, + 0.0019494022708386183, + 0.0018361242255195975, + 0.0015240733046084642, + 0.005288300104439259, + 0.006712788250297308, + 0.006116338074207306, + 0.0044693429954349995, + 0.011126985773444176, + 0.007444517221301794, + 0.004418460186570883, + 0.007030203938484192, + 0.006693632807582617, + 0.006857767701148987, + 0.006433690432459116, + 0.005645317025482655, + 0.017507370561361313, + 0.023203345015645027, + 0.008160741999745369, + 0.012008952908217907, + 0.008140137419104576, + 0.013150082901120186, + 0.0115345548838377, + 0.009917342104017735, + 0.020672259852290154, + 0.02672126330435276, + 0.013002624735236168, + 0.014320703223347664, + 0.01331015583127737, + 0.01170853991061449, + 0.133198082447052, + 0.04752939194440842, + 0.018895462155342102, + 0.021752750501036644, + 0.01747080124914646, + 0.09901285916566849, + 0.0023880205117166042, + 0.002971263835206628, + 0.005891472566872835, + 0.0069780778139829636, + 0.003307127859443426, + 0.0030803752597421408, + 0.003272241447120905, + 0.002459317445755005, + 0.012588541954755783, + 0.00848769024014473, + 0.06040893867611885, + 0.0348520390689373, + 0.016621455550193787, + 0.018193144351243973, + 0.014976228587329388, + 0.01519819162786007, + 0.00722908228635788, + 0.009736317209899426, + 0.013584103435277939, + 0.011273120529949665, + 0.019827386364340782, + 0.02338212914764881, + 0.007925822399556637, + 0.007225163280963898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -1.2235822677612305 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low horizon perspective where water meets the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt upward in a slow, seamless motion from the water to the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky colored deep indigo at the zenith, transitioning to soft violet and magenta, then fiery orange-gold at the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a dense city skyline of steel, glass, and concrete skyscrapers visible reflecting the sky's colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless office windows illuminated with warm yellow lights scattered across the building facades?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206049770116806 + ], + "opposite_logprob_tokens": [ + -2.812060594558716 + ], + "logprob_reward": -0.06206049770116806, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206049770116806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05323563888669014, + 0.028914375230669975, + 0.026984503492712975, + 0.07567035406827927, + 0.049368757754564285, + 0.09555885195732117, + 0.31807833909988403, + 0.045613788068294525, + 0.2160867154598236, + 0.09048860520124435 + ], + "credit_map": [ + 0.007057770621031523, + 0.007483749184757471, + 0.0069305384531617165, + 0.00606747530400753, + 0.006082890555262566, + 0.006296245381236076, + 0.007101607508957386, + 0.006215361412614584, + 0.0027378052473068237, + 0.004393612500280142, + 0.0036631596740335226, + 0.002517200540751219, + 0.0037119006738066673, + 0.004502162337303162, + 0.003042875789105892, + 0.0043456582352519035, + 0.0043160016648471355, + 0.0029936577193439007, + 0.004925057757645845, + 0.004301041830331087, + 0.002168626058846712, + 0.0024336171336472034, + 0.0021881817374378443, + 0.003658320987597108, + 0.007604298647493124, + 0.007078561466187239, + 0.009412883780896664, + 0.007075680885463953, + 0.015588655136525631, + 0.012216721661388874, + 0.006438012234866619, + 0.01025554072111845, + 0.0040336549282073975, + 0.004053281154483557, + 0.003977359272539616, + 0.0032915729098021984, + 0.009385636076331139, + 0.012692739255726337, + 0.005054376553744078, + 0.006880138069391251, + 0.007621206343173981, + 0.008931504562497139, + 0.014625612646341324, + 0.011524117551743984, + 0.01141966599971056, + 0.014689378440380096, + 0.01523135881870985, + 0.011516003869473934, + 0.011368200182914734, + 0.009399732574820518, + 0.12103518843650818, + 0.04331104829907417, + 0.01329257246106863, + 0.017336538061499596, + 0.01492704451084137, + 0.08740800619125366, + 0.0033756981138139963, + 0.005084499716758728, + 0.006536714266985655, + 0.008819598704576492, + 0.006313484627753496, + 0.005164547823369503, + 0.006433635484427214, + 0.0038856107275933027, + 0.014942364767193794, + 0.009400487877428532, + 0.0806405246257782, + 0.045591238886117935, + 0.016374703496694565, + 0.018172509968280792, + 0.015855472534894943, + 0.015109430067241192, + 0.007596116978675127, + 0.009205321781337261, + 0.012369899079203606, + 0.0087735615670681, + 0.015528722666203976, + 0.023967495188117027, + 0.0064948140643537045, + 0.006552678532898426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.06206049770116806 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_1/index_tensor([3760, 3760])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_1/index_tensor([3760, 3760])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..39f6c34275ee2e04e8647668adfa487627f61125 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_1/index_tensor([3760, 3760])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film's crane shot begins in medium close-up, focusing on an elderly East Asian scholar standing atop the Great Wall of China. With wisps of graying hair at his temples and a wrinkled face brimming with intellectual curiosity, he wears a traditional dark gray, high-necked wool tunic, a look that combines elegance with practicality, warding off the morning chill. With an authoritative and majestic gesture, he sweeps his right arm across the landscape, his lips moving as if passionately explaining the wall's long history. As he speaks, the camera begins a slow, majestic ascent, pulling back to reveal the full context of his location. The weathered gray stone brick wall, with its uneven texture and ancient jagged edges, fills the frame before the camera ascends higher, revealing a massive stone serpent winding endlessly across the vast, undulating peaks of the rolling mountains. A thick, wispy mist shrouds the valley below, creating layers of depth and mystery, while the soft golden morning sunlight penetrates the valley, catching the high mountain ridges and casting long, dramatic shadows. The entire scene is filled with an atmosphere of epic grandeur and profound historical reverence, capturing a moment of human knowledge against the backdrop of timeless nature and monumental achievement.", + "seed": 4110, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_1/index_3760_seed_4110_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.011084074061363936, + "vlm_logprob": -0.011084074061363936, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.011084074061363936, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016417184844613075 + ], + "opposite_logprob_tokens": [ + -4.266417026519775 + ], + "logprob_reward": -0.016417184844613075, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016417184844613075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5628783106803894, + 0.07165160775184631, + 0.042363569140434265, + 0.059750258922576904, + 0.018786773085594177, + 0.060221776366233826, + 0.09516038745641708, + 0.02384857088327408, + 0.021197421476244926, + 0.04414122551679611 + ], + "credit_map": [ + 0.05214168503880501, + 0.11320973187685013, + 0.025288408622145653, + 0.02392766997218132, + 0.041411127895116806, + 0.20452004671096802, + 0.06811175495386124, + 0.03426792472600937, + 0.01080088596791029, + 0.010584094561636448, + 0.0064872573129832745, + 0.006623785477131605, + 0.012734861113131046, + 0.008897794410586357, + 0.00851469486951828, + 0.007008234970271587, + 0.008000778965651989, + 0.010327977128326893, + 0.0037652947939932346, + 0.003116222331300378, + 0.004207246005535126, + 0.004705384839326143, + 0.005078070797026157, + 0.0031625947449356318, + 0.007107102777808905, + 0.013371624052524567, + 0.008042520843446255, + 0.003959812689572573, + 0.007676944602280855, + 0.0076146721839904785, + 0.006478987634181976, + 0.005498592276126146, + 0.003972889855504036, + 0.0037509314715862274, + 0.0015447895275428891, + 0.0017195985419675708, + 0.0018109873635694385, + 0.002892859512940049, + 0.001707669929601252, + 0.0013870459515601397, + 0.00566904665902257, + 0.02079346403479576, + 0.004355310928076506, + 0.003823351813480258, + 0.004225671291351318, + 0.014859152026474476, + 0.0032726528588682413, + 0.0032231276854872704, + 0.01072672288864851, + 0.014440483413636684, + 0.012263989076018333, + 0.011050480417907238, + 0.007954427041113377, + 0.020993802696466446, + 0.008696620352566242, + 0.009033857844769955, + 0.0019237842643633485, + 0.0035621076822280884, + 0.0020194905810058117, + 0.002192517276853323, + 0.005073928739875555, + 0.002947041764855385, + 0.003544811625033617, + 0.0025848892983049154, + 0.0032210741192102432, + 0.0026928100269287825, + 0.0026389334816485643, + 0.0024718407075852156, + 0.003203155007213354, + 0.0023600354325026274, + 0.0029257116839289665, + 0.00168386020231992, + 0.004121108911931515, + 0.0041028219275176525, + 0.00565791642293334, + 0.006052621640264988, + 0.004020750522613525, + 0.005486064590513706, + 0.007538181263953447, + 0.007161758840084076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.016417184844613075 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052719381637871265 + ], + "opposite_logprob_tokens": [ + -5.255271911621094 + ], + "logprob_reward": -0.0052719381637871265, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052719381637871265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28886693716049194, + 0.2725675106048584, + 0.06590452045202255, + 0.050768427550792694, + 0.033915963023900986, + 0.035657674074172974, + 0.11794402450323105, + 0.055949706584215164, + 0.039710305631160736, + 0.038714922964572906 + ], + "credit_map": [ + 0.051197685301303864, + 0.024617547169327736, + 0.017761453986167908, + 0.01311688032001257, + 0.027095185592770576, + 0.08392668515443802, + 0.04958418756723404, + 0.02156730368733406, + 0.029668426141142845, + 0.040794726461172104, + 0.025069840252399445, + 0.03695274516940117, + 0.03309137374162674, + 0.03327123820781708, + 0.03607581928372383, + 0.037643320858478546, + 0.005507274065166712, + 0.020713984966278076, + 0.0045947059988975525, + 0.003486155066639185, + 0.006576440297067165, + 0.01552980300039053, + 0.0049837627448141575, + 0.004512382205575705, + 0.005725023336708546, + 0.012801926583051682, + 0.005839917808771133, + 0.0038749598897993565, + 0.005755905993282795, + 0.005945678800344467, + 0.0060662804171442986, + 0.004758726805448532, + 0.007001733873039484, + 0.004929054528474808, + 0.00422066543251276, + 0.0034106378443539143, + 0.0035805338993668556, + 0.0038988173473626375, + 0.0036328909918665886, + 0.003241627011448145, + 0.004905429668724537, + 0.005941218696534634, + 0.0034663781989365816, + 0.003584629390388727, + 0.0030966103076934814, + 0.00828549638390541, + 0.003389500081539154, + 0.002988408086821437, + 0.00872868113219738, + 0.014174705371260643, + 0.013446797616779804, + 0.00943718571215868, + 0.01041392982006073, + 0.03283398970961571, + 0.013154692947864532, + 0.01575402542948723, + 0.007219696417450905, + 0.009859285317361355, + 0.0033348440192639828, + 0.006438213400542736, + 0.006704865954816341, + 0.007319693453609943, + 0.007020802237093449, + 0.008052301593124866, + 0.004500003065913916, + 0.005154118407517672, + 0.005419432185590267, + 0.007697968278080225, + 0.0039632851257920265, + 0.004870004951953888, + 0.004467515274882317, + 0.003637975314632058, + 0.0031823283061385155, + 0.004122883081436157, + 0.004940121900290251, + 0.00499798683449626, + 0.006099029444158077, + 0.003954451531171799, + 0.006167283281683922, + 0.005250833462923765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0052719381637871265 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549715116620064 + ], + "opposite_logprob_tokens": [ + -4.387549877166748 + ], + "logprob_reward": -0.012549715116620064, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549715116620064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6315361857414246, + 0.03312800079584122, + 0.03273440897464752, + 0.029921947047114372, + 0.03317618742585182, + 0.06049540266394615, + 0.03545822948217392, + 0.0828666239976883, + 0.04526754468679428, + 0.015415487810969353 + ], + "credit_map": [ + 0.09782865643501282, + 0.0716404840350151, + 0.0324484147131443, + 0.02754533663392067, + 0.0521571971476078, + 0.22448210418224335, + 0.08332598954439163, + 0.04210807755589485, + 0.005142401438206434, + 0.004587904550135136, + 0.0025272397324442863, + 0.0034039905294775963, + 0.004474453628063202, + 0.004759286995977163, + 0.0031778961420059204, + 0.005054823588579893, + 0.006794574670493603, + 0.007646047044545412, + 0.0023351265117526054, + 0.0022885799407958984, + 0.0030972477979958057, + 0.004708241205662489, + 0.003366884309798479, + 0.0024977035354822874, + 0.0030894598457962275, + 0.01020040363073349, + 0.002535114996135235, + 0.001482966123148799, + 0.002304299734532833, + 0.005170008633285761, + 0.002505472395569086, + 0.002634221687912941, + 0.007710969541221857, + 0.004441756755113602, + 0.0027175582945346832, + 0.0037374147213995457, + 0.002681807382032275, + 0.004429350141435862, + 0.003550067776814103, + 0.0039072646759450436, + 0.005686814896762371, + 0.025272253900766373, + 0.002216512570157647, + 0.0022183291148394346, + 0.0024430090561509132, + 0.01606188528239727, + 0.003138918662443757, + 0.0034576815087348223, + 0.004426905419677496, + 0.004552979487925768, + 0.0036765860859304667, + 0.0035046092234551907, + 0.0032533418852835894, + 0.005062297452241182, + 0.00307973544113338, + 0.007901771925389767, + 0.008815024979412556, + 0.012553364038467407, + 0.005704374518245459, + 0.007208232767879963, + 0.010005197487771511, + 0.011161047033965588, + 0.009334186092019081, + 0.01808519847691059, + 0.004449274856597185, + 0.006162621546536684, + 0.0039484621956944466, + 0.005588492378592491, + 0.006626615766435862, + 0.006419035140424967, + 0.0077454037964344025, + 0.004327637609094381, + 0.0013307174667716026, + 0.0015953256515786052, + 0.0020423331297934055, + 0.002047335496172309, + 0.001708300202153623, + 0.0013967548729851842, + 0.0031880696769803762, + 0.0021066521294414997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.012549715116620064 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539473362267017 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539473362267017, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539473362267017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2860417366027832, + 0.03972580283880234, + 0.23002031445503235, + 0.05050226300954819, + 0.037626903504133224, + 0.11740554124116898, + 0.09167620539665222, + 0.0587230883538723, + 0.03407732769846916, + 0.05420078709721565 + ], + "credit_map": [ + 0.03399448096752167, + 0.06366368383169174, + 0.01286693662405014, + 0.013151041232049465, + 0.022735098376870155, + 0.10546094179153442, + 0.019205570220947266, + 0.014964007772505283, + 0.006569489371031523, + 0.005795610137283802, + 0.004643579479306936, + 0.00580266397446394, + 0.005207051057368517, + 0.003989970311522484, + 0.004166328813880682, + 0.0035511115565896034, + 0.04178447648882866, + 0.06768294423818588, + 0.014393306337296963, + 0.012999913655221462, + 0.011517155915498734, + 0.04529419541358948, + 0.020075593143701553, + 0.016272759065032005, + 0.006494896952062845, + 0.009739923290908337, + 0.0060302275232970715, + 0.002754760906100273, + 0.00915639940649271, + 0.006786428391933441, + 0.0054666646756231785, + 0.004072966054081917, + 0.010073709301650524, + 0.0065863970667123795, + 0.0027290123980492353, + 0.002803693525493145, + 0.002537316409870982, + 0.004426253959536552, + 0.003999075386673212, + 0.004471444059163332, + 0.009413463994860649, + 0.045503631234169006, + 0.004474400542676449, + 0.004382910672575235, + 0.0071420627646148205, + 0.033731162548065186, + 0.007569150533527136, + 0.0051887668669223785, + 0.0066654435358941555, + 0.011999723501503468, + 0.010846462100744247, + 0.008728382177650928, + 0.008665116503834724, + 0.026564573869109154, + 0.008450830355286598, + 0.00975568313151598, + 0.005774394143372774, + 0.0061012255027890205, + 0.00537634314969182, + 0.0055423639714717865, + 0.009998076595366001, + 0.006800712551921606, + 0.006345860660076141, + 0.01278411503881216, + 0.003980879206210375, + 0.005604417994618416, + 0.0036694093141704798, + 0.0036962609738111496, + 0.005025438964366913, + 0.004431271459907293, + 0.004762002732604742, + 0.0029076498467475176, + 0.004463709890842438, + 0.006617639679461718, + 0.008348990231752396, + 0.0066096363589167595, + 0.006766302976757288, + 0.006411777343600988, + 0.00890745222568512, + 0.006075279787182808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.012539473362267017 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642058819532394 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642058819532394, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642058819532394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31275680661201477, + 0.08671583235263824, + 0.09917565435171127, + 0.04008731618523598, + 0.0373619869351387, + 0.16413579881191254, + 0.1338237076997757, + 0.04227912425994873, + 0.045499712228775024, + 0.03816405311226845 + ], + "credit_map": [ + 0.06090940535068512, + 0.038962192833423615, + 0.019745929166674614, + 0.01508428156375885, + 0.026520898565649986, + 0.07671669125556946, + 0.05094679445028305, + 0.02387060597538948, + 0.009797255508601665, + 0.008782844990491867, + 0.01667623221874237, + 0.012496800161898136, + 0.007855835370719433, + 0.012021583504974842, + 0.009002055041491985, + 0.01008322648704052, + 0.022096680477261543, + 0.02604043483734131, + 0.007577033247798681, + 0.005046874284744263, + 0.006461095530539751, + 0.016821490600705147, + 0.0083730174228549, + 0.006759022828191519, + 0.00498968968167901, + 0.004914967808872461, + 0.0061044832691550255, + 0.00340483570471406, + 0.005262767430394888, + 0.005321063566952944, + 0.006066736299544573, + 0.0040227752178907394, + 0.007470250129699707, + 0.005217249970883131, + 0.003433900186792016, + 0.0025823009200394154, + 0.005532230716198683, + 0.004853281192481518, + 0.004451386630535126, + 0.003821389051154256, + 0.010180944576859474, + 0.06229805946350098, + 0.005533921532332897, + 0.005976512096822262, + 0.005526146851480007, + 0.056280821561813354, + 0.01068994402885437, + 0.007649444974958897, + 0.0230261143296957, + 0.02412845380604267, + 0.014362947084009647, + 0.01326628215610981, + 0.007965409196913242, + 0.015879567712545395, + 0.00987761840224266, + 0.02531730756163597, + 0.004231372848153114, + 0.006008968688547611, + 0.005280699115246534, + 0.004951742943376303, + 0.005260393489152193, + 0.005283415783196688, + 0.004322971682995558, + 0.006939558777958155, + 0.005196974612772465, + 0.006368704605847597, + 0.0047670085914433, + 0.006903622765094042, + 0.005750407464802265, + 0.004889035597443581, + 0.006010278128087521, + 0.0056136781349778175, + 0.0031910559628158808, + 0.005739027634263039, + 0.005775311961770058, + 0.005645159166306257, + 0.003930500242859125, + 0.004077852237969637, + 0.006337439641356468, + 0.0034677067305892706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.008642058819532394 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film's crane shot begins in medium close-up, focusing on an elderly East Asian scholar standing atop the Great Wall of China. With wisps of graying hair at his temples and a wrinkled face brimming with intellectual curiosity, he wears a traditional dark gray, high-necked wool tunic, a look that combines elegance with practicality, warding off the morning chill. With an authoritative and majestic gesture, he sweeps his right arm across the landscape, his lips moving as if passionately explaining the wall's long history. As he speaks, the camera begins a slow, majestic ascent, pulling back to reveal the full context of his location. The weathered gray stone brick wall, with its uneven texture and ancient jagged edges, fills the frame before the camera ascends higher, revealing a massive stone serpent winding endlessly across the vast, undulating peaks of the rolling mountains. A thick, wispy mist shrouds the valley below, creating layers of depth and mystery, while the soft golden morning sunlight penetrates the valley, catching the high mountain ridges and casting long, dramatic shadows. The entire scene is filled with an atmosphere of epic grandeur and profound historical reverence, capturing a moment of human knowledge against the backdrop of timeless nature and monumental achievement.", + "seed": 4111, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_1/index_3760_seed_4111_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01287765596061945, + "vlm_logprob": -0.01287765596061945, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.01287765596061945, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018840758129954338 + ], + "opposite_logprob_tokens": [ + -4.268840789794922 + ], + "logprob_reward": -0.018840758129954338, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.018840758129954338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08250976353883743, + 0.04533137008547783, + 0.07225611805915833, + 0.08832656592130661, + 0.065330870449543, + 0.04243664816021919, + 0.19654415547847748, + 0.12655794620513916, + 0.031747717410326004, + 0.24895881116390228 + ], + "credit_map": [ + 0.014450128190219402, + 0.00864590983837843, + 0.010330923832952976, + 0.007156334817409515, + 0.0146176153793931, + 0.008611188270151615, + 0.011416036635637283, + 0.007281641010195017, + 0.006541188806295395, + 0.006930680014193058, + 0.006464612670242786, + 0.00713050551712513, + 0.004315151367336512, + 0.004017168655991554, + 0.005934577900916338, + 0.003997486084699631, + 0.009947017766535282, + 0.01452002115547657, + 0.006482022814452648, + 0.01219013798981905, + 0.006477522663772106, + 0.006886334624141455, + 0.010135270655155182, + 0.00561779597774148, + 0.01297901663929224, + 0.014618020504713058, + 0.009272187016904354, + 0.009864356368780136, + 0.010115420445799828, + 0.012304832227528095, + 0.009448913857340813, + 0.009723820723593235, + 0.008589552715420723, + 0.007070640102028847, + 0.007530539762228727, + 0.00958301406353712, + 0.006472725421190262, + 0.008492737077176571, + 0.010401139967143536, + 0.007190525531768799, + 0.006001443136483431, + 0.006351297255605459, + 0.00463646836578846, + 0.004630760755389929, + 0.006406232714653015, + 0.005460007581859827, + 0.00550801819190383, + 0.0034424245823174715, + 0.04426075518131256, + 0.02623092196881771, + 0.017747394740581512, + 0.012920898385345936, + 0.020148195326328278, + 0.029688047245144844, + 0.028604570776224136, + 0.01694338209927082, + 0.0148486802354455, + 0.016395917162299156, + 0.01163467112928629, + 0.018163558095693588, + 0.014050887897610664, + 0.018177038058638573, + 0.021783916279673576, + 0.011503294110298157, + 0.003200079780071974, + 0.003949420992285013, + 0.003981793764978647, + 0.0034364203456789255, + 0.005848240572959185, + 0.004508683457970619, + 0.00363979977555573, + 0.0031832808163017035, + 0.031847137957811356, + 0.051912445574998856, + 0.039680615067481995, + 0.030425185337662697, + 0.04072190821170807, + 0.019219044595956802, + 0.015956465154886246, + 0.01919601485133171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.018840758129954338 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016083383932709694 + ], + "opposite_logprob_tokens": [ + -4.141083240509033 + ], + "logprob_reward": -0.016083383932709694, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016083383932709694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1274675726890564, + 0.13284586369991302, + 0.12113182991743088, + 0.05983446165919304, + 0.03081369772553444, + 0.04465044289827347, + 0.07739073038101196, + 0.05553177371621132, + 0.035434044897556305, + 0.3148995637893677 + ], + "credit_map": [ + 0.027120715007185936, + 0.014500527642667294, + 0.009927685372531414, + 0.011257083155214787, + 0.02026698738336563, + 0.015442385338246822, + 0.020955055952072144, + 0.007997144013643265, + 0.02161719836294651, + 0.020090648904442787, + 0.019655369222164154, + 0.018749810755252838, + 0.013041846454143524, + 0.01245717704296112, + 0.01633308082818985, + 0.010900736786425114, + 0.01803230494260788, + 0.023664720356464386, + 0.009181744419038296, + 0.025212569162249565, + 0.010328540578484535, + 0.010950187221169472, + 0.013731404207646847, + 0.010030366480350494, + 0.008810698986053467, + 0.0069355652667582035, + 0.007230877410620451, + 0.006748752668499947, + 0.0065016113221645355, + 0.008523941040039062, + 0.00561140850186348, + 0.009471612051129341, + 0.004647685680538416, + 0.0064118304289877415, + 0.0036829772870987654, + 0.0038500376977026463, + 0.0024390900507569313, + 0.0031594715546816587, + 0.0030398890376091003, + 0.0035827145911753178, + 0.0055929142981767654, + 0.005709961522370577, + 0.004349403083324432, + 0.00623726611956954, + 0.006215899717062712, + 0.005941478069871664, + 0.006775156129151583, + 0.003828367218375206, + 0.007899615913629532, + 0.014864250086247921, + 0.00808621570467949, + 0.008167402818799019, + 0.011244280263781548, + 0.012034104205667973, + 0.008507032878696918, + 0.006587826646864414, + 0.007025505416095257, + 0.00632975809276104, + 0.007292875554412603, + 0.006798960268497467, + 0.00696604885160923, + 0.009809440933167934, + 0.006510088220238686, + 0.004799094516783953, + 0.0040289489552378654, + 0.003610607935115695, + 0.00446281535550952, + 0.004328007344156504, + 0.003606748068705201, + 0.0036877712700515985, + 0.007248274050652981, + 0.004460873547941446, + 0.04102513566613197, + 0.0679439902305603, + 0.048681892454624176, + 0.035133589059114456, + 0.05156153813004494, + 0.026475291699171066, + 0.021182753145694733, + 0.022895345464348793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.016083383932709694 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204819686710835 + ], + "opposite_logprob_tokens": [ + -4.264204978942871 + ], + "logprob_reward": -0.014204819686710835, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204819686710835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1001201942563057, + 0.03616762533783913, + 0.23906952142715454, + 0.14018869400024414, + 0.09189678728580475, + 0.04587264731526375, + 0.09010189026594162, + 0.03072589635848999, + 0.04482089728116989, + 0.1810358315706253 + ], + "credit_map": [ + 0.020502114668488503, + 0.011008307337760925, + 0.012859293259680271, + 0.0076951500959694386, + 0.015545707195997238, + 0.01077115349471569, + 0.01306674163788557, + 0.00867172796279192, + 0.006078103091567755, + 0.004187502898275852, + 0.003640968818217516, + 0.004804220050573349, + 0.006969502195715904, + 0.003389376448467374, + 0.004194480367004871, + 0.002903471002355218, + 0.024499626830220222, + 0.035914160311222076, + 0.023317251354455948, + 0.03543210029602051, + 0.018576601520180702, + 0.039772093296051025, + 0.03034856915473938, + 0.031209122389554977, + 0.021018197759985924, + 0.018624454736709595, + 0.016451004892587662, + 0.016470476984977722, + 0.014239038340747356, + 0.017254984006285667, + 0.016318030655384064, + 0.019812509417533875, + 0.01089476514607668, + 0.012853468768298626, + 0.009755250997841358, + 0.013120732270181179, + 0.008957288227975368, + 0.013863533735275269, + 0.013088151812553406, + 0.009363597258925438, + 0.007152993232011795, + 0.007828936912119389, + 0.002841765759512782, + 0.0054442682303488255, + 0.003895014990121126, + 0.012653879821300507, + 0.0032338928431272507, + 0.0028218941297382116, + 0.014667215757071972, + 0.010299191810190678, + 0.009582671336829662, + 0.009590412490069866, + 0.012562514282763004, + 0.011581191793084145, + 0.013198683969676495, + 0.008620006032288074, + 0.004512877203524113, + 0.004211113788187504, + 0.003402902279049158, + 0.0033217843156307936, + 0.0031753908842802048, + 0.002027411013841629, + 0.006927367299795151, + 0.003147051203995943, + 0.004517658613622189, + 0.005507485941052437, + 0.006460519041866064, + 0.005439297296106815, + 0.0049084327183663845, + 0.003674971405416727, + 0.009645589627325535, + 0.004666942171752453, + 0.025052141398191452, + 0.03643857315182686, + 0.026590025052428246, + 0.023831119760870934, + 0.030041083693504333, + 0.012490064837038517, + 0.01240114402025938, + 0.014191688038408756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 3 + ], + "selected_reward": -0.014204819686710835 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629540748894215 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.007629540748894215, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629540748894215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057125452905893326, + 0.05749644711613655, + 0.03509685769677162, + 0.12959644198417664, + 0.13647928833961487, + 0.04339543357491493, + 0.1486620306968689, + 0.08398230373859406, + 0.07407969236373901, + 0.2340860366821289 + ], + "credit_map": [ + 0.011763720773160458, + 0.005388567224144936, + 0.00466585299000144, + 0.006458762101829052, + 0.0066978298127651215, + 0.005140738096088171, + 0.011394528672099113, + 0.005615454167127609, + 0.009077594615519047, + 0.008716980926692486, + 0.008366105146706104, + 0.007825778797268867, + 0.006482503842562437, + 0.004933087155222893, + 0.007207149639725685, + 0.004887247923761606, + 0.004554152023047209, + 0.007496103178709745, + 0.003733827965334058, + 0.005862375255674124, + 0.0031280466355383396, + 0.004367274232208729, + 0.003644525306299329, + 0.002310550771653652, + 0.020631255581974983, + 0.017418300732970238, + 0.01361108385026455, + 0.013096322305500507, + 0.015290812589228153, + 0.021231042221188545, + 0.013228270225226879, + 0.015089365653693676, + 0.015685075893998146, + 0.013561979867517948, + 0.015193077735602856, + 0.02097235433757305, + 0.013900335878133774, + 0.01895623840391636, + 0.024574650451540947, + 0.013635573908686638, + 0.005833352915942669, + 0.005731089971959591, + 0.004326677415519953, + 0.006612355820834637, + 0.005352697800844908, + 0.005761453881859779, + 0.0063207074999809265, + 0.0034570982679724693, + 0.02606859616935253, + 0.02601892314851284, + 0.018046097829937935, + 0.012695749290287495, + 0.01830589957535267, + 0.018446732312440872, + 0.017182253301143646, + 0.011897770687937737, + 0.010921251028776169, + 0.00742027023807168, + 0.008301294408738613, + 0.00837050098925829, + 0.00743230851367116, + 0.01083559263497591, + 0.023618893697857857, + 0.007082189433276653, + 0.008170779794454575, + 0.008104708977043629, + 0.010262585245072842, + 0.009023754857480526, + 0.007228417322039604, + 0.011344194412231445, + 0.010027832351624966, + 0.00991741381585598, + 0.030387673527002335, + 0.04830523952841759, + 0.037428196519613266, + 0.02933379076421261, + 0.03768547996878624, + 0.017631860449910164, + 0.015802206471562386, + 0.01751159131526947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "selected_reward": -0.007629540748894215 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629777304828167 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007629777304828167, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629777304828167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020626625046133995, + 0.02141156606376171, + 0.01091652549803257, + 0.016132472082972527, + 0.053331516683101654, + 0.020853787660598755, + 0.04065023735165596, + 0.025955649092793465, + 0.025617077946662903, + 0.764504611492157 + ], + "credit_map": [ + 0.004414273425936699, + 0.001362278708256781, + 0.002620629034936428, + 0.0017188071506097913, + 0.0018632166320458055, + 0.002212903695181012, + 0.00468115508556366, + 0.0017533592181280255, + 0.0030156283173710108, + 0.004001749213784933, + 0.0028095152229070663, + 0.0031751871574670076, + 0.002175617963075638, + 0.0017063459381461143, + 0.0027106325142085552, + 0.001816888921894133, + 0.001704503782093525, + 0.002826646203175187, + 0.0009960888419300318, + 0.0013855090364813805, + 0.001050391816534102, + 0.0011132832150906324, + 0.0009381897398270667, + 0.0009019108838401735, + 0.0020748572424054146, + 0.002446706872433424, + 0.001688802964054048, + 0.002225700067356229, + 0.0013145986013114452, + 0.0023870307486504316, + 0.0014336113817989826, + 0.0025611610617488623, + 0.006474310532212257, + 0.006397397257387638, + 0.004179843235760927, + 0.008237413130700588, + 0.005543285980820656, + 0.0086598489433527, + 0.008624703623354435, + 0.005214706063270569, + 0.0033612377010285854, + 0.0025829386431723833, + 0.0017212972743436694, + 0.0024105668999254704, + 0.0019219160312786698, + 0.005382322706282139, + 0.0020463867112994194, + 0.0014271188993006945, + 0.009269936010241508, + 0.004231635946780443, + 0.003296957118436694, + 0.0040065087378025055, + 0.0038120117969810963, + 0.0070164925418794155, + 0.0059758457355201244, + 0.003040849696844816, + 0.0031300478149205446, + 0.003601841162890196, + 0.0027323730755597353, + 0.004107914865016937, + 0.003615114837884903, + 0.003921613562852144, + 0.0024933337699621916, + 0.0023534088395535946, + 0.002766527235507965, + 0.00333975232206285, + 0.0034442206379026175, + 0.002546692034229636, + 0.00331115466542542, + 0.00436068419367075, + 0.0032214978709816933, + 0.0026265489868819714, + 0.09866084158420563, + 0.15191993117332458, + 0.12492837011814117, + 0.09623117744922638, + 0.12240585684776306, + 0.058686189353466034, + 0.051715511828660965, + 0.05995664745569229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.007629777304828167 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film's crane shot begins in medium close-up, focusing on an elderly East Asian scholar standing atop the Great Wall of China. With wisps of graying hair at his temples and a wrinkled face brimming with intellectual curiosity, he wears a traditional dark gray, high-necked wool tunic, a look that combines elegance with practicality, warding off the morning chill. With an authoritative and majestic gesture, he sweeps his right arm across the landscape, his lips moving as if passionately explaining the wall's long history. As he speaks, the camera begins a slow, majestic ascent, pulling back to reveal the full context of his location. The weathered gray stone brick wall, with its uneven texture and ancient jagged edges, fills the frame before the camera ascends higher, revealing a massive stone serpent winding endlessly across the vast, undulating peaks of the rolling mountains. A thick, wispy mist shrouds the valley below, creating layers of depth and mystery, while the soft golden morning sunlight penetrates the valley, catching the high mountain ridges and casting long, dramatic shadows. The entire scene is filled with an atmosphere of epic grandeur and profound historical reverence, capturing a moment of human knowledge against the backdrop of timeless nature and monumental achievement.", + "seed": 4112, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_1/index_3760_seed_4112_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.013600419461727142, + "vlm_logprob": -0.013600419461727142, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.013600419461727142, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02093188837170601 + ], + "opposite_logprob_tokens": [ + -4.270931720733643 + ], + "logprob_reward": -0.02093188837170601, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02093188837170601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07463330030441284, + 0.018504904583096504, + 0.03064272180199623, + 0.382480263710022, + 0.046290408819913864, + 0.15544477105140686, + 0.04184221848845482, + 0.04017484560608864, + 0.11271513253450394, + 0.09727141261100769 + ], + "credit_map": [ + 0.007218349725008011, + 0.010192020796239376, + 0.012463456019759178, + 0.012026376090943813, + 0.009257984347641468, + 0.006634856108576059, + 0.00818970799446106, + 0.008650555275380611, + 0.0030688794795423746, + 0.001919482834637165, + 0.0018001034623011947, + 0.0036163614131510258, + 0.0022892586421221495, + 0.002075130119919777, + 0.0019125952385365963, + 0.0018230942077934742, + 0.005864080041646957, + 0.0030527692288160324, + 0.002552654827013612, + 0.004768017213791609, + 0.0034921541810035706, + 0.003437834093347192, + 0.0030738788191229105, + 0.004401335492730141, + 0.0209412332624197, + 0.026552550494670868, + 0.04744647443294525, + 0.15887977182865143, + 0.023428164422512054, + 0.011640761978924274, + 0.0243134293705225, + 0.06927788257598877, + 0.006059564650058746, + 0.0053855800069868565, + 0.0059015387669205666, + 0.005031141918152571, + 0.006804848089814186, + 0.005942780990153551, + 0.006409869063645601, + 0.004755085334181786, + 0.024257248267531395, + 0.02051357738673687, + 0.008830593898892403, + 0.009065290912985802, + 0.03277263790369034, + 0.01917268894612789, + 0.027555476874113083, + 0.0132772671058774, + 0.006080148741602898, + 0.005388031713664532, + 0.003218922298401594, + 0.0038264156319200993, + 0.008458183147013187, + 0.004224843345582485, + 0.005504104308784008, + 0.005141567904502153, + 0.004545295145362616, + 0.0050620269030332565, + 0.004816227592527866, + 0.0036192205734550953, + 0.0037222185637801886, + 0.0053282068111002445, + 0.009353057481348515, + 0.003728593932464719, + 0.013238220475614071, + 0.014369622804224491, + 0.016436735168099403, + 0.010842549614608288, + 0.022168418392539024, + 0.012487099505960941, + 0.013435645960271358, + 0.009736841544508934, + 0.013139608316123486, + 0.01316952146589756, + 0.01661091484129429, + 0.010408701375126839, + 0.008728722110390663, + 0.012358174659311771, + 0.010728725232183933, + 0.012127048335969448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.02093188837170601 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009802774526178837 + ], + "opposite_logprob_tokens": [ + -4.63480281829834 + ], + "logprob_reward": -0.009802774526178837, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009802774526178837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1556636244058609, + 0.06240313872694969, + 0.05464771389961243, + 0.12438607960939407, + 0.0348040834069252, + 0.03084065206348896, + 0.29557061195373535, + 0.03647344186902046, + 0.10818800330162048, + 0.09702261537313461 + ], + "credit_map": [ + 0.015414146706461906, + 0.01291394792497158, + 0.017489906400442123, + 0.027330607175827026, + 0.01864439807832241, + 0.021213272586464882, + 0.022176140919327736, + 0.020481200888752937, + 0.007027352694422007, + 0.005123484879732132, + 0.0062416899017989635, + 0.012757333926856518, + 0.007207827176898718, + 0.005962635390460491, + 0.010477637872099876, + 0.007605176884680986, + 0.006089386064559221, + 0.005045222118496895, + 0.006520259194076061, + 0.0058312565088272095, + 0.007894746959209442, + 0.007687644101679325, + 0.007232380099594593, + 0.008346821181476116, + 0.0073034511879086494, + 0.008498749695718288, + 0.014975271187722683, + 0.049593862146139145, + 0.007108060177415609, + 0.005202446132898331, + 0.008460396900773048, + 0.023243840783834457, + 0.006662625353783369, + 0.0031586606055498123, + 0.0019487184472382069, + 0.001784545136615634, + 0.0056367660872638226, + 0.00569869251921773, + 0.00591985322535038, + 0.003994222730398178, + 0.00596723984926939, + 0.00500272074714303, + 0.0016800451558083296, + 0.001805306994356215, + 0.0041575864888727665, + 0.004788212012499571, + 0.004086533561348915, + 0.003353009233251214, + 0.03346904367208481, + 0.06263729929924011, + 0.01772732101380825, + 0.017054477706551552, + 0.06259708106517792, + 0.03415681794285774, + 0.042804330587387085, + 0.025124240666627884, + 0.0063988338224589825, + 0.004980788566172123, + 0.0036175993736833334, + 0.00287049962207675, + 0.004024163354188204, + 0.004368169698864222, + 0.006199138704687357, + 0.004014249425381422, + 0.013119333423674107, + 0.010567311197519302, + 0.011786420829594135, + 0.010654582642018795, + 0.015827303752303123, + 0.012161767110228539, + 0.022247403860092163, + 0.011823879554867744, + 0.01317778043448925, + 0.010686550289392471, + 0.014748076908290386, + 0.009026816114783287, + 0.007239857688546181, + 0.016195492818951607, + 0.016239624470472336, + 0.009708411060273647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.009802774526178837 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012552659027278423 + ], + "opposite_logprob_tokens": [ + -4.387552738189697 + ], + "logprob_reward": -0.012552659027278423, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012552659027278423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08671090751886368, + 0.062355685979127884, + 0.0329357273876667, + 0.40537506341934204, + 0.021619735285639763, + 0.01970183290541172, + 0.08792395144701004, + 0.04396532103419304, + 0.09313543885946274, + 0.14627626538276672 + ], + "credit_map": [ + 0.01317579299211502, + 0.015472104772925377, + 0.011905680410563946, + 0.014420745894312859, + 0.007395042572170496, + 0.009130485355854034, + 0.008301369845867157, + 0.006909691728651524, + 0.007837655954062939, + 0.006504656281322241, + 0.010869475081562996, + 0.009977140463888645, + 0.010961112566292286, + 0.005691461730748415, + 0.005187515635043383, + 0.005326670594513416, + 0.003978727385401726, + 0.002729129046201706, + 0.0025085252709686756, + 0.0034841145388782024, + 0.00476808613166213, + 0.006325254682451487, + 0.004017028491944075, + 0.005124863237142563, + 0.022022895514965057, + 0.02897140011191368, + 0.05010528862476349, + 0.16461917757987976, + 0.02635623887181282, + 0.015146128833293915, + 0.02602817863225937, + 0.07212577760219574, + 0.003849893342703581, + 0.0024292217567563057, + 0.0010956261539831758, + 0.001070500584319234, + 0.004040304571390152, + 0.0037487836088985205, + 0.003089525969699025, + 0.0022958829067647457, + 0.003333236090838909, + 0.002564957831054926, + 0.0011776034953072667, + 0.0014873587060719728, + 0.0027842435520142317, + 0.0027531383093446493, + 0.003741525812074542, + 0.0018597710877656937, + 0.009952716529369354, + 0.016798458993434906, + 0.0054700481705367565, + 0.005562531296163797, + 0.017581336200237274, + 0.009935838170349598, + 0.014057186432182789, + 0.008565851487219334, + 0.005618651397526264, + 0.006123362109065056, + 0.006139018107205629, + 0.003675991203635931, + 0.003859577002003789, + 0.005648775026202202, + 0.008476668037474155, + 0.00442328117787838, + 0.009589443914592266, + 0.011148414574563503, + 0.012508089654147625, + 0.011418289504945278, + 0.018564293161034584, + 0.008087579160928726, + 0.015994077548384666, + 0.005825265776365995, + 0.020296849310398102, + 0.014619722031056881, + 0.02559160813689232, + 0.013980871997773647, + 0.011337396688759327, + 0.01697755604982376, + 0.02668645977973938, + 0.016785822808742523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.012552659027278423 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643122389912605 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643122389912605, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643122389912605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05310903862118721, + 0.0316750593483448, + 0.03498100861907005, + 0.17544828355312347, + 0.02931116335093975, + 0.05125937983393669, + 0.07592913508415222, + 0.06413529068231583, + 0.2107059359550476, + 0.2734456956386566 + ], + "credit_map": [ + 0.005144886672496796, + 0.00841790996491909, + 0.006227700039744377, + 0.012213080190122128, + 0.00536910118535161, + 0.005039522890001535, + 0.006126066669821739, + 0.0045707738026976585, + 0.003437578910961747, + 0.0030098264105618, + 0.005816144868731499, + 0.0049447836354374886, + 0.003588315099477768, + 0.003023177618160844, + 0.0037211573217064142, + 0.004134077578783035, + 0.0037972861900925636, + 0.0033697308972477913, + 0.003714070888236165, + 0.004851313307881355, + 0.00619629118591547, + 0.004902332555502653, + 0.003932536579668522, + 0.004217447713017464, + 0.010366585105657578, + 0.012178927659988403, + 0.021191341802477837, + 0.07260032743215561, + 0.010482176207005978, + 0.007832477800548077, + 0.012091955170035362, + 0.028704486787319183, + 0.005582229699939489, + 0.0030419770628213882, + 0.0017833051970228553, + 0.0017593277152627707, + 0.004087687004357576, + 0.004718714393675327, + 0.004461000207811594, + 0.003876923583447933, + 0.011500517837703228, + 0.007518676575273275, + 0.005162504501640797, + 0.004886743146926165, + 0.00535625359043479, + 0.007463850546628237, + 0.0060083381831645966, + 0.0033625010401010513, + 0.010377205908298492, + 0.01684674248099327, + 0.005942775867879391, + 0.0046284375712275505, + 0.011683783493936062, + 0.0092238774523139, + 0.010561803355813026, + 0.006664515472948551, + 0.008557755500078201, + 0.011222710832953453, + 0.0065832361578941345, + 0.006202381569892168, + 0.007415862288326025, + 0.007922736927866936, + 0.009366970509290695, + 0.00686363922432065, + 0.028124859556555748, + 0.024753481149673462, + 0.023789042606949806, + 0.019870806485414505, + 0.044877294450998306, + 0.021520109847187996, + 0.029097124934196472, + 0.01867321878671646, + 0.03902142867445946, + 0.021474674344062805, + 0.04928711801767349, + 0.020872097462415695, + 0.025459421798586845, + 0.07488647103309631, + 0.021031305193901062, + 0.021413179114460945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.008643122389912605 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071652993559837 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.016071652993559837, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071652993559837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09644002467393875, + 0.045422498136758804, + 0.07735607773065567, + 0.15188530087471008, + 0.09604994207620621, + 0.02307911589741707, + 0.10709146410226822, + 0.05497606098651886, + 0.22531796991825104, + 0.12238159030675888 + ], + "credit_map": [ + 0.014149933122098446, + 0.011562212370336056, + 0.014792151749134064, + 0.014495579525828362, + 0.009299060329794884, + 0.00943831354379654, + 0.009479574859142303, + 0.013223201967775822, + 0.010014241561293602, + 0.006478068884462118, + 0.004447652958333492, + 0.008331887423992157, + 0.00431315740570426, + 0.0035872783046215773, + 0.003787660039961338, + 0.004462550859898329, + 0.011390081606805325, + 0.008023267611861229, + 0.007604355923831463, + 0.013312732800841331, + 0.012247017584741116, + 0.007031986024230719, + 0.010474646463990211, + 0.007271991111338139, + 0.008554432541131973, + 0.010612895712256432, + 0.018282977864146233, + 0.06262844800949097, + 0.008853175677359104, + 0.005035514943301678, + 0.009740795008838177, + 0.028177062049508095, + 0.012836256995797157, + 0.016225431114435196, + 0.015673929825425148, + 0.01316907349973917, + 0.0116515988484025, + 0.009587109088897705, + 0.010628605261445045, + 0.006277939770370722, + 0.005391388665884733, + 0.004313584882766008, + 0.0016256740782409906, + 0.002121711615473032, + 0.0018794945208355784, + 0.004656285047531128, + 0.0016857740702107549, + 0.0014052027836441994, + 0.012150514870882034, + 0.021711479872465134, + 0.008769189938902855, + 0.008422593586146832, + 0.020565016195178032, + 0.01263797003775835, + 0.013035085052251816, + 0.009799612686038017, + 0.008642664179205894, + 0.007782184984534979, + 0.006348068825900555, + 0.005915898364037275, + 0.0053786626085639, + 0.008413175120949745, + 0.007459951564669609, + 0.005035455338656902, + 0.02975582331418991, + 0.0337253175675869, + 0.03522700443863869, + 0.01803804375231266, + 0.04458218067884445, + 0.02663005143404007, + 0.020872347056865692, + 0.01648721471428871, + 0.022810092195868492, + 0.016318876296281815, + 0.020658187568187714, + 0.010948874987661839, + 0.010027410462498665, + 0.02133922092616558, + 0.009059448726475239, + 0.011219470761716366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.016071652993559837 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film's crane shot begins in medium close-up, focusing on an elderly East Asian scholar standing atop the Great Wall of China. With wisps of graying hair at his temples and a wrinkled face brimming with intellectual curiosity, he wears a traditional dark gray, high-necked wool tunic, a look that combines elegance with practicality, warding off the morning chill. With an authoritative and majestic gesture, he sweeps his right arm across the landscape, his lips moving as if passionately explaining the wall's long history. As he speaks, the camera begins a slow, majestic ascent, pulling back to reveal the full context of his location. The weathered gray stone brick wall, with its uneven texture and ancient jagged edges, fills the frame before the camera ascends higher, revealing a massive stone serpent winding endlessly across the vast, undulating peaks of the rolling mountains. A thick, wispy mist shrouds the valley below, creating layers of depth and mystery, while the soft golden morning sunlight penetrates the valley, catching the high mountain ridges and casting long, dramatic shadows. The entire scene is filled with an atmosphere of epic grandeur and profound historical reverence, capturing a moment of human knowledge against the backdrop of timeless nature and monumental achievement.", + "seed": 4113, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_1/index_3760_seed_4113_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.014445572812110185, + "vlm_logprob": -0.014445572812110185, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.014445572812110185, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022824086248874664 + ], + "opposite_logprob_tokens": [ + -4.022824287414551 + ], + "logprob_reward": -0.022824086248874664, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.022824086248874664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18055951595306396, + 0.051219094544649124, + 0.10176699608564377, + 0.18317656219005585, + 0.16870638728141785, + 0.05576343461871147, + 0.05128353834152222, + 0.1079343631863594, + 0.06127997860312462, + 0.038310129195451736 + ], + "credit_map": [ + 0.019598888233304024, + 0.02800542302429676, + 0.017021100968122482, + 0.01923445798456669, + 0.027659554034471512, + 0.024324890226125717, + 0.021887779235839844, + 0.022827431559562683, + 0.009496096521615982, + 0.007734626531600952, + 0.0051656244322657585, + 0.006632823497056961, + 0.007576280273497105, + 0.006072608288377523, + 0.00480877747759223, + 0.003732257755473256, + 0.014844264835119247, + 0.016711439937353134, + 0.0078030903823673725, + 0.006543332245200872, + 0.013329335488379002, + 0.017245739698410034, + 0.009425965137779713, + 0.015863830223679543, + 0.03203785419464111, + 0.07245337218046188, + 0.004848528187721968, + 0.0036572145763784647, + 0.0092763751745224, + 0.020347949117422104, + 0.013510147109627724, + 0.0270451158285141, + 0.041961368173360825, + 0.010494505986571312, + 0.01065788697451353, + 0.012405209243297577, + 0.012110082432627678, + 0.05515091121196747, + 0.014882175251841545, + 0.011044252663850784, + 0.00845541711896658, + 0.009648309089243412, + 0.010335301049053669, + 0.005443047732114792, + 0.005425430368632078, + 0.006050118710845709, + 0.005935337394475937, + 0.004470475949347019, + 0.005628454033285379, + 0.008159223012626171, + 0.007574288174510002, + 0.004377172794193029, + 0.004631750285625458, + 0.006936977151781321, + 0.005422210320830345, + 0.00855346117168665, + 0.016892462968826294, + 0.01941373012959957, + 0.013458737172186375, + 0.010657751001417637, + 0.013473465107381344, + 0.01479278039187193, + 0.009450620040297508, + 0.009794812649488449, + 0.006433208007365465, + 0.0056996457278728485, + 0.009017856791615486, + 0.004755490925163031, + 0.010050461627542973, + 0.009131284430623055, + 0.008472488261759281, + 0.007719541899859905, + 0.003953150007873774, + 0.003980353008955717, + 0.004836255684494972, + 0.0032892152667045593, + 0.0060668387450277805, + 0.006398884113878012, + 0.004011048469692469, + 0.005774382036179304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.022824086248874664 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059676566161215305 + ], + "opposite_logprob_tokens": [ + -5.130967617034912 + ], + "logprob_reward": -0.0059676566161215305, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059676566161215305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10613474249839783, + 0.12921075522899628, + 0.062072139233350754, + 0.1307240128517151, + 0.1039109155535698, + 0.037925612181425095, + 0.04912542179226875, + 0.24929046630859375, + 0.09094865620136261, + 0.040657270699739456 + ], + "credit_map": [ + 0.015352282673120499, + 0.014852834865450859, + 0.00818748027086258, + 0.008405925706028938, + 0.01644403673708439, + 0.014663802459836006, + 0.013052675873041153, + 0.015175698325037956, + 0.024870382621884346, + 0.019546620547771454, + 0.0088850362226367, + 0.020419485867023468, + 0.015015006065368652, + 0.014055305160582066, + 0.01702648587524891, + 0.00939242634922266, + 0.010822545737028122, + 0.008302263915538788, + 0.004079494625329971, + 0.008712226524949074, + 0.00720871239900589, + 0.005853784736245871, + 0.010472890920937061, + 0.0066202240996062756, + 0.016875823959708214, + 0.047663770616054535, + 0.004507375415414572, + 0.002806302160024643, + 0.007244149222970009, + 0.01376418024301529, + 0.010202880017459393, + 0.027659522369503975, + 0.009230396710336208, + 0.017884094268083572, + 0.01095752976834774, + 0.012901837937533855, + 0.0089878486469388, + 0.02261226251721382, + 0.013427228666841984, + 0.007909715175628662, + 0.005000002216547728, + 0.00478331046178937, + 0.00513580022379756, + 0.004143912345170975, + 0.004312471486628056, + 0.008277274668216705, + 0.003242436097934842, + 0.0030304063111543655, + 0.007474580314010382, + 0.0054758042097091675, + 0.008140645921230316, + 0.005360396113246679, + 0.005260174162685871, + 0.004375774879008532, + 0.005568234715610743, + 0.0074698105454444885, + 0.029202058911323547, + 0.030722271651029587, + 0.04308350011706352, + 0.03159533813595772, + 0.025768034160137177, + 0.025712041184306145, + 0.030996792018413544, + 0.032210420817136765, + 0.008889826945960522, + 0.011981331743299961, + 0.012646441347897053, + 0.009115947410464287, + 0.013614635914564133, + 0.01246786117553711, + 0.013428792357444763, + 0.008803813718259335, + 0.004615225363522768, + 0.004661818500608206, + 0.004867160227149725, + 0.0040262495167553425, + 0.003734748112037778, + 0.006576849613338709, + 0.005500864237546921, + 0.006674355361610651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.0059676566161215305 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294130340218544 + ], + "opposite_logprob_tokens": [ + -3.77329421043396 + ], + "logprob_reward": -0.023294130340218544, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294130340218544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03575243800878525, + 0.08312689512968063, + 0.018770327791571617, + 0.6555026173591614, + 0.1223166361451149, + 0.019106976687908173, + 0.01810070499777794, + 0.017656968906521797, + 0.01016196422278881, + 0.019504595547914505 + ], + "credit_map": [ + 0.004699884448200464, + 0.004054614808410406, + 0.0039302692748606205, + 0.003627743339166045, + 0.007503165863454342, + 0.004942593164741993, + 0.0039497497491538525, + 0.003044413635507226, + 0.02080230973660946, + 0.015455437824130058, + 0.004847553092986345, + 0.011318732053041458, + 0.005984751973301172, + 0.009681210853159428, + 0.0069186189211905, + 0.008118263445794582, + 0.002508515492081642, + 0.003294463502243161, + 0.0017096538795158267, + 0.0014183856546878815, + 0.0019601769745349884, + 0.0035455457400530577, + 0.0014256180729717016, + 0.002907966962084174, + 0.15266135334968567, + 0.2729889452457428, + 0.011331984773278236, + 0.011454005725681782, + 0.02847367897629738, + 0.06822624057531357, + 0.043537579476833344, + 0.06682872772216797, + 0.033285461366176605, + 0.01021671760827303, + 0.007864206098020077, + 0.015439081005752087, + 0.009238598868250847, + 0.02637331560254097, + 0.012986124493181705, + 0.006913115736097097, + 0.0023039202205836773, + 0.002601715037599206, + 0.002851828234270215, + 0.00275981891900301, + 0.0024266561958938837, + 0.002758505055680871, + 0.0017713221022859216, + 0.001633207779377699, + 0.0018282127566635609, + 0.00405327882617712, + 0.0031303679570555687, + 0.001481486950069666, + 0.0019965805113315582, + 0.002032386139035225, + 0.002062025247141719, + 0.0015163635835051537, + 0.002836673753336072, + 0.0025584178511053324, + 0.0024560061283409595, + 0.0017137543763965368, + 0.002576426137238741, + 0.0021647268440574408, + 0.002387414453551173, + 0.0009635490132495761, + 0.0009827845497056842, + 0.001104602124541998, + 0.001867778250016272, + 0.0007750721997581422, + 0.001319331582635641, + 0.0014413715107366443, + 0.0014453856274485588, + 0.0012256366899237037, + 0.0023986580781638622, + 0.0024103729519993067, + 0.0027324187103658915, + 0.001754106837324798, + 0.002359571633860469, + 0.002845670562237501, + 0.002031011274084449, + 0.0029727823566645384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.023294130340218544 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194242656230927 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014194242656230927, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194242656230927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06442657858133316, + 0.0776962861418724, + 0.05559869483113289, + 0.6638723015785217, + 0.06624439358711243, + 0.007701583206653595, + 0.02261362597346306, + 0.009105125442147255, + 0.00807827990502119, + 0.024663139134645462 + ], + "credit_map": [ + 0.006132223177701235, + 0.007119815330952406, + 0.005879098083823919, + 0.006505516357719898, + 0.01295141875743866, + 0.008993067778646946, + 0.00940466858446598, + 0.00744076631963253, + 0.01636955514550209, + 0.01603342406451702, + 0.007205367553979158, + 0.007827325724065304, + 0.009094818495213985, + 0.009431377053260803, + 0.005356200970709324, + 0.006378216668963432, + 0.01207007933408022, + 0.008859023451805115, + 0.004492993466556072, + 0.005232181865721941, + 0.004346281755715609, + 0.0064355311915278435, + 0.0051229773089289665, + 0.009039628319442272, + 0.17199201881885529, + 0.28085604310035706, + 0.010101968422532082, + 0.011088897474110126, + 0.027364110574126244, + 0.06729760766029358, + 0.04231743887066841, + 0.052854232490062714, + 0.008230476640164852, + 0.009546544402837753, + 0.006002807058393955, + 0.007903879508376122, + 0.008378900587558746, + 0.010531170293688774, + 0.0096087995916605, + 0.0060418169014155865, + 0.0007140910602174699, + 0.0010661052074283361, + 0.0010876930318772793, + 0.0006759559619240463, + 0.0012042945018038154, + 0.0008354088640771806, + 0.0012634113663807511, + 0.0008546230965293944, + 0.004048769827932119, + 0.00310967187397182, + 0.004389200825244188, + 0.0013054991140961647, + 0.0019751277286559343, + 0.003001938108354807, + 0.002416216768324375, + 0.0023671998642385006, + 0.0011586543405428529, + 0.0008298874599859118, + 0.00185302272439003, + 0.0009450262296013534, + 0.001234724186360836, + 0.0011888504959642887, + 0.0011234688572585583, + 0.0007714908570051193, + 0.0008751815767027438, + 0.0008221582393161952, + 0.0011575734242796898, + 0.0005345416720956564, + 0.0008152551017701626, + 0.0011191501980647445, + 0.0010384774068370461, + 0.001715941820293665, + 0.0027717018965631723, + 0.003016340546309948, + 0.004932943731546402, + 0.0022286016028374434, + 0.0019571390002965927, + 0.004172473214566708, + 0.0023476015776395798, + 0.0032363389618694782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.014194242656230927 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly East Asian scholar standing on the Great Wall of China?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar wearing a traditional dark gray, high-necked wool tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar sweep his right arm across the landscape while his lips move as if speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium close-up and perform a slow ascent pulling back to reveal the full context?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast mountain range with a thick mist in the valley and golden morning sunlight on the ridges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947748199105263 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947748199105263, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947748199105263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12435220181941986, + 0.1339789479970932, + 0.09146272391080856, + 0.22979407012462616, + 0.07803095877170563, + 0.021270865574479103, + 0.1349209100008011, + 0.053118642419576645, + 0.02648886851966381, + 0.10658175498247147 + ], + "credit_map": [ + 0.01757052354514599, + 0.020693574100732803, + 0.0157985407859087, + 0.00877506472170353, + 0.020286191254854202, + 0.015365807339549065, + 0.01335482019931078, + 0.012507687322795391, + 0.03187521547079086, + 0.030877821147441864, + 0.008946554735302925, + 0.014517262578010559, + 0.011127150617539883, + 0.016746042296290398, + 0.010549638420343399, + 0.00933927483856678, + 0.01607448235154152, + 0.015963522717356682, + 0.006940717343240976, + 0.005298297852277756, + 0.00865335762500763, + 0.01391615066677332, + 0.007169042248278856, + 0.017447153106331825, + 0.07985560595989227, + 0.04369058459997177, + 0.01294028665870428, + 0.006131642032414675, + 0.0104381637647748, + 0.0234079509973526, + 0.014822127297520638, + 0.03850771486759186, + 0.01923177018761635, + 0.008104605600237846, + 0.00617526238784194, + 0.006314200349152088, + 0.007336083799600601, + 0.015952281653881073, + 0.007379092276096344, + 0.007537662051618099, + 0.002520709065720439, + 0.0027388527523726225, + 0.0025283643044531345, + 0.0019279038533568382, + 0.002347249537706375, + 0.0041780658066272736, + 0.002176281763240695, + 0.0028534403536468744, + 0.015154676511883736, + 0.029138067737221718, + 0.019905399531126022, + 0.00911819189786911, + 0.013053752481937408, + 0.018957065418362617, + 0.011881602928042412, + 0.01771216094493866, + 0.0078380536288023, + 0.008513057604432106, + 0.007097150664776564, + 0.004295635037124157, + 0.006394798867404461, + 0.00877442117780447, + 0.005116232670843601, + 0.005089297890663147, + 0.0031704665161669254, + 0.003050691680982709, + 0.00334777869284153, + 0.0018053696257993579, + 0.003907933831214905, + 0.004316876642405987, + 0.003402160480618477, + 0.0034875941928476095, + 0.013586802408099174, + 0.014842646196484566, + 0.02489284612238407, + 0.010609428398311138, + 0.007612795569002628, + 0.01150139793753624, + 0.011900186538696289, + 0.011635668575763702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.005947748199105263 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_10/index_tensor([4667, 4667])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_10/index_tensor([4667, 4667])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5f9e0f1d23de1df4b3d7458d4636f776114d3180 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_10/index_tensor([4667, 4667])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning macro close-up, captured in the style of high-speed photography, focuses on the metal nozzle of a stainless steel watering can, its surface marked with minute scratches and a few dried water stains. A powerful stream of water erupts from its fine orifice, instantly freezing in mid-air. Each individual droplet is imbued with life, stretched into elliptical or even longer streamlines in mid-air, resembling a group of runners sprinting in a fierce race, each maintaining a tiny distance between them, creating a dynamic trajectory. Bright, sharp side lighting precisely outlines the droplets, creating diamond-like highlights on their surface. Light penetrates their crystal-clear interior, refracting tiny specks of light, while the metal nozzle takes on a cool, metallic sheen. A soft, out-of-focus blur of greenery in the background provides a tranquil backdrop to this explosive moment. The entire image is imbued with breathtaking dynamism and tension, capturing the instantaneous energy of the water droplets with artistic beauty, showcasing the breathtaking vitality of the microscopic world.", + "seed": 24408, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_10/index_4667_seed_24408_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04158803052268922, + "vlm_logprob": -0.04158803052268922, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.04158803052268922, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023425143212080002 + ], + "opposite_logprob_tokens": [ + -3.8984251022338867 + ], + "logprob_reward": -0.023425143212080002, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.023425143212080002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26019486784935, + 0.10122676938772202, + 0.09913509339094162, + 0.16778814792633057, + 0.07096894830465317, + 0.022663233801722527, + 0.047088392078876495, + 0.0598924495279789, + 0.0707949548959732, + 0.10024711489677429 + ], + "credit_map": [ + 0.03152526170015335, + 0.036757711321115494, + 0.03278714790940285, + 0.02353440783917904, + 0.029411647468805313, + 0.02623448520898819, + 0.024979956448078156, + 0.05496426671743393, + 0.012973153032362461, + 0.017040209844708443, + 0.011830613017082214, + 0.014585083350539207, + 0.014559242874383926, + 0.006751858629286289, + 0.01068235095590353, + 0.012804262340068817, + 0.01357354037463665, + 0.02638009935617447, + 0.016531510278582573, + 0.011264298111200333, + 0.008221033960580826, + 0.005717133171856403, + 0.007519337348639965, + 0.0099281445145607, + 0.0206302497535944, + 0.01880652643740177, + 0.03394574299454689, + 0.017836283892393112, + 0.017214756458997726, + 0.02114192768931389, + 0.014951487071812153, + 0.023261170834302902, + 0.010985716246068478, + 0.010025985538959503, + 0.009981623850762844, + 0.005473810713738203, + 0.005679031368345022, + 0.009312107227742672, + 0.007888255640864372, + 0.011622420512139797, + 0.002822292037308216, + 0.003383675357326865, + 0.003818464931100607, + 0.0027172400150448084, + 0.0019689463078975677, + 0.002547214273363352, + 0.002353968797251582, + 0.0030514318495988846, + 0.006175939459353685, + 0.006784047931432724, + 0.008127529174089432, + 0.006804503034800291, + 0.002845324343070388, + 0.003706603776663542, + 0.005976100452244282, + 0.006668343208730221, + 0.012540414929389954, + 0.016119491308927536, + 0.00815011840313673, + 0.0037144271191209555, + 0.003008032450452447, + 0.0045462846755981445, + 0.004830543417483568, + 0.006983136758208275, + 0.008924946188926697, + 0.014652521349489689, + 0.007713531143963337, + 0.010765600949525833, + 0.005584141705185175, + 0.006474966183304787, + 0.007043300662189722, + 0.009635952301323414, + 0.019230475649237633, + 0.02144801616668701, + 0.014933832921087742, + 0.009332136251032352, + 0.009482317604124546, + 0.0083743492141366, + 0.007520095445215702, + 0.009925885125994682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.023425143212080002 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609663851559162 + ], + "opposite_logprob_tokens": [ + -4.141096591949463 + ], + "logprob_reward": -0.01609663851559162, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609663851559162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11047273129224777, + 0.05299029499292374, + 0.2407701462507248, + 0.1584024578332901, + 0.03618194907903671, + 0.06852876394987106, + 0.05274328589439392, + 0.061518099159002304, + 0.06747037917375565, + 0.15092183649539948 + ], + "credit_map": [ + 0.01602901890873909, + 0.016085486859083176, + 0.013360822573304176, + 0.00802700500935316, + 0.01140963938087225, + 0.013546755537390709, + 0.009227216243743896, + 0.02278679795563221, + 0.008842790499329567, + 0.013144243508577347, + 0.007224183529615402, + 0.00490419939160347, + 0.00457315519452095, + 0.004521094728261232, + 0.0032318239100277424, + 0.006548810750246048, + 0.03483317047357559, + 0.05265355110168457, + 0.046075932681560516, + 0.024775909259915352, + 0.01723538711667061, + 0.019995782524347305, + 0.015900712460279465, + 0.029299698770046234, + 0.018244827166199684, + 0.02033405937254429, + 0.036355160176754, + 0.020180147141218185, + 0.017086459323763847, + 0.015786506235599518, + 0.012116695754230022, + 0.018298618495464325, + 0.005631411448121071, + 0.005675853230059147, + 0.006351853255182505, + 0.003921475261449814, + 0.004344428889453411, + 0.002865445800125599, + 0.003055742010474205, + 0.004335745703428984, + 0.008036963641643524, + 0.011949368752539158, + 0.008511475287377834, + 0.00959194079041481, + 0.006430266425013542, + 0.006311807781457901, + 0.008951710537075996, + 0.008745236322283745, + 0.007724521681666374, + 0.013200667686760426, + 0.00930049642920494, + 0.007384749129414558, + 0.0035158046521246433, + 0.0028106255922466516, + 0.0033050619531422853, + 0.005501362960785627, + 0.008010891266167164, + 0.012375553138554096, + 0.00938133429735899, + 0.006059534382075071, + 0.004020483698695898, + 0.004940130282193422, + 0.005792390089482069, + 0.010937790386378765, + 0.009177200496196747, + 0.017788296565413475, + 0.010084475390613079, + 0.008719407021999359, + 0.004403900355100632, + 0.0043436139822006226, + 0.005056237336248159, + 0.007897255942225456, + 0.021199531853199005, + 0.03394374996423721, + 0.04028422385454178, + 0.01894671656191349, + 0.010386289097368717, + 0.007455972954630852, + 0.009224629029631615, + 0.009480738081037998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.01609663851559162 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033998131752014 + ], + "opposite_logprob_tokens": [ + -1.9103399515151978 + ], + "logprob_reward": -0.16033998131752014, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033998131752014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24154940247535706, + 0.04822462052106857, + 0.11241820454597473, + 0.1366078108549118, + 0.06171892583370209, + 0.1326834112405777, + 0.053817424923181534, + 0.037681493908166885, + 0.041719358414411545, + 0.1335793286561966 + ], + "credit_map": [ + 0.030140863731503487, + 0.0329371839761734, + 0.03194471821188927, + 0.02691623754799366, + 0.026908358559012413, + 0.029659023508429527, + 0.021383535116910934, + 0.041659511625766754, + 0.00725758820772171, + 0.006969185080379248, + 0.0050200410187244415, + 0.005685534328222275, + 0.00423708138987422, + 0.005268631502985954, + 0.0052231959998607635, + 0.008563362993299961, + 0.01597503572702408, + 0.01730455458164215, + 0.018582886084914207, + 0.012614281848073006, + 0.008111337199807167, + 0.013251373544335365, + 0.009990843944251537, + 0.01658790558576584, + 0.017825957387685776, + 0.017300289124250412, + 0.017989808693528175, + 0.014737087301909924, + 0.01564328745007515, + 0.01481771469116211, + 0.011003896594047546, + 0.027289781719446182, + 0.006659046281129122, + 0.008307287469506264, + 0.005823769140988588, + 0.007114966865628958, + 0.006925938650965691, + 0.005762140266597271, + 0.01140540186315775, + 0.00972037948668003, + 0.015523809008300304, + 0.019682055339217186, + 0.014689529314637184, + 0.030772021040320396, + 0.011714601889252663, + 0.009577933698892593, + 0.016827987506985664, + 0.013895480893552303, + 0.008519092574715614, + 0.015067210420966148, + 0.008235784247517586, + 0.007315559312701225, + 0.0031865336932241917, + 0.0028845914639532566, + 0.003679075976833701, + 0.004929582588374615, + 0.0054045431315898895, + 0.006164673250168562, + 0.006351850461214781, + 0.00426495959982276, + 0.0028908634558320045, + 0.0041471440345048904, + 0.0038039579521864653, + 0.004653505049645901, + 0.004871494136750698, + 0.008355780504643917, + 0.005413770209997892, + 0.007843569852411747, + 0.0031273076310753822, + 0.004044297616928816, + 0.003473621094599366, + 0.004589518532156944, + 0.02243826538324356, + 0.02802526019513607, + 0.033857956528663635, + 0.014594245702028275, + 0.008024951443076134, + 0.008993837051093578, + 0.007550516165792942, + 0.010094301775097847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.16033998131752014 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257945042103529 + ], + "opposite_logprob_tokens": [ + -5.255258083343506 + ], + "logprob_reward": -0.005257945042103529, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257945042103529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.156110018491745, + 0.0437655933201313, + 0.06779671460390091, + 0.06208862364292145, + 0.057365093380212784, + 0.08777036517858505, + 0.024169664829969406, + 0.040583252906799316, + 0.04637249931693077, + 0.41397812962532043 + ], + "credit_map": [ + 0.01998770982027054, + 0.019898997619748116, + 0.018967514857649803, + 0.015985365957021713, + 0.0178513303399086, + 0.018951518461108208, + 0.014661304652690887, + 0.029806291684508324, + 0.004662100225687027, + 0.00799008458852768, + 0.004607373382896185, + 0.006357094272971153, + 0.003386871190741658, + 0.004796167369931936, + 0.005382680334150791, + 0.006583227775990963, + 0.010217194445431232, + 0.01415764819830656, + 0.01245877705514431, + 0.00841361191123724, + 0.005332528613507748, + 0.0055872476659715176, + 0.005036856513470411, + 0.006592859048396349, + 0.007218760903924704, + 0.010128422640264034, + 0.01019667275249958, + 0.008844236843287945, + 0.0062479316256940365, + 0.006090002600103617, + 0.005044726189225912, + 0.008317876607179642, + 0.007484945002943277, + 0.0073873079381883144, + 0.005857470910996199, + 0.007412667851895094, + 0.006321849301457405, + 0.005426638759672642, + 0.009924186393618584, + 0.00755003746598959, + 0.010826504789292812, + 0.011356614530086517, + 0.010707657784223557, + 0.02052333578467369, + 0.00730632059276104, + 0.005465298891067505, + 0.011397017166018486, + 0.010187633335590363, + 0.0027728891000151634, + 0.002787246135994792, + 0.002895655110478401, + 0.0032863053493201733, + 0.0025783993769437075, + 0.002339419675990939, + 0.004103535786271095, + 0.003406216623261571, + 0.005487561225891113, + 0.006945787463337183, + 0.006077917758375406, + 0.005590710323303938, + 0.0031794386450201273, + 0.003773581003770232, + 0.004282099194824696, + 0.005246162414550781, + 0.006141560152173042, + 0.009431513957679272, + 0.00820651464164257, + 0.006859263405203819, + 0.003247169777750969, + 0.0038937022909522057, + 0.00409605260938406, + 0.004496726673096418, + 0.05892796441912651, + 0.11251724511384964, + 0.10947117954492569, + 0.05825332552194595, + 0.026025163009762764, + 0.01469608023762703, + 0.01586064323782921, + 0.01822657324373722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.005257945042103529 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028204445261508226 + ], + "opposite_logprob_tokens": [ + -5.8778204917907715 + ], + "logprob_reward": -0.0028204445261508226, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028204445261508226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07756015658378601, + 0.042449742555618286, + 0.06610288470983505, + 0.019519727677106857, + 0.017423115670681, + 0.031188981607556343, + 0.0507526770234108, + 0.06522537022829056, + 0.022640200331807137, + 0.607137143611908 + ], + "credit_map": [ + 0.009921799413859844, + 0.013485470786690712, + 0.009843585081398487, + 0.008856751956045628, + 0.0072645461186766624, + 0.008148920722305775, + 0.006069420371204615, + 0.01396966353058815, + 0.0072771646082401276, + 0.007649868261069059, + 0.005821149330586195, + 0.004816771484911442, + 0.0041390638798475266, + 0.003421033499762416, + 0.0028566194232553244, + 0.006468076724559069, + 0.010375726036727428, + 0.01792754791676998, + 0.013979428447782993, + 0.0076323505491018295, + 0.004079984966665506, + 0.003568604588508606, + 0.0030069134663790464, + 0.005532329436391592, + 0.00270261918194592, + 0.003157840808853507, + 0.004201547242701054, + 0.0027286068070679903, + 0.0015402901917696, + 0.0019997686613351107, + 0.0011284977663308382, + 0.0020605577155947685, + 0.002310509094968438, + 0.0021709667053073645, + 0.0025461476761847734, + 0.0009814990917220712, + 0.0014286695513874292, + 0.002245121169835329, + 0.001921247923746705, + 0.0038189529441297054, + 0.004377269651740789, + 0.004265208728611469, + 0.008224014192819595, + 0.0035264482721686363, + 0.001936699147336185, + 0.0022396305575966835, + 0.0026817398611456156, + 0.003937970846891403, + 0.0065834615379571915, + 0.007726825308054686, + 0.007415578700602055, + 0.006710875313729048, + 0.003121161600574851, + 0.003435455495491624, + 0.00873472634702921, + 0.0070245880633592606, + 0.009618938900530338, + 0.013927292078733444, + 0.01195596344769001, + 0.003956019878387451, + 0.003268028376623988, + 0.007633288390934467, + 0.0033819624222815037, + 0.011483876034617424, + 0.0033070514909923077, + 0.005118845961987972, + 0.0033142389729619026, + 0.002941734157502651, + 0.0016447703819721937, + 0.0013423350173979998, + 0.0021288476418703794, + 0.00284237670712173, + 0.10173404961824417, + 0.14209842681884766, + 0.18824370205402374, + 0.06315527111291885, + 0.027090182527899742, + 0.03001709096133709, + 0.019833125174045563, + 0.03496527299284935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.0028204445261508226 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, captured in the style of high-speed photography, focuses on the metal nozzle of a stainless steel watering can, its surface marked with minute scratches and a few dried water stains. A powerful stream of water erupts from its fine orifice, instantly freezing in mid-air. Each individual droplet is imbued with life, stretched into elliptical or even longer streamlines in mid-air, resembling a group of runners sprinting in a fierce race, each maintaining a tiny distance between them, creating a dynamic trajectory. Bright, sharp side lighting precisely outlines the droplets, creating diamond-like highlights on their surface. Light penetrates their crystal-clear interior, refracting tiny specks of light, while the metal nozzle takes on a cool, metallic sheen. A soft, out-of-focus blur of greenery in the background provides a tranquil backdrop to this explosive moment. The entire image is imbued with breathtaking dynamism and tension, capturing the instantaneous energy of the water droplets with artistic beauty, showcasing the breathtaking vitality of the microscopic world.", + "seed": 24409, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_10/index_4667_seed_24409_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.014612747030332685, + "vlm_logprob": -0.014612747030332685, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.014612747030332685, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02134132757782936 + ], + "opposite_logprob_tokens": [ + -4.146341323852539 + ], + "logprob_reward": -0.02134132757782936, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02134132757782936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10775396972894669, + 0.07357577979564667, + 0.11992243677377701, + 0.10730879008769989, + 0.02953650988638401, + 0.037545345723629, + 0.21170853078365326, + 0.1430893838405609, + 0.0831541121006012, + 0.08640501648187637 + ], + "credit_map": [ + 0.021727334707975388, + 0.011246693320572376, + 0.01571529172360897, + 0.027523348107933998, + 0.005069458391517401, + 0.008224948309361935, + 0.012488231994211674, + 0.005758668761700392, + 0.007336121518164873, + 0.008428274653851986, + 0.0096999267116189, + 0.016918379813432693, + 0.004477436654269695, + 0.008354202844202518, + 0.01265626773238182, + 0.005705168005079031, + 0.01383018959313631, + 0.015867481008172035, + 0.011364242993295193, + 0.018067102879285812, + 0.010255713015794754, + 0.01300010085105896, + 0.026223544031381607, + 0.011314059607684612, + 0.010981770232319832, + 0.01022045873105526, + 0.02910604141652584, + 0.022200103849172592, + 0.0051940125413239, + 0.007477466017007828, + 0.01679677516222, + 0.005332167260348797, + 0.0048956009559333324, + 0.00392962945625186, + 0.004577843938022852, + 0.005136026069521904, + 0.0022837803699076176, + 0.0030639851465821266, + 0.002679589204490185, + 0.0029700547456741333, + 0.006231830921024084, + 0.0051009515300393105, + 0.005078759044408798, + 0.005092324689030647, + 0.004048583563417196, + 0.004515477921813726, + 0.0038255881518125534, + 0.0036518338602036238, + 0.033311035484075546, + 0.02598891593515873, + 0.0426248237490654, + 0.02967662550508976, + 0.020154213532805443, + 0.022407488897442818, + 0.02492048777639866, + 0.0126249510794878, + 0.02417399361729622, + 0.023333992809057236, + 0.024426816031336784, + 0.02456021122634411, + 0.009710045531392097, + 0.013271248899400234, + 0.013973639346659184, + 0.00963945034891367, + 0.0059054638259112835, + 0.008006280288100243, + 0.009300143457949162, + 0.010828475467860699, + 0.011360129341483116, + 0.015716807916760445, + 0.010107924230396748, + 0.011928895488381386, + 0.007337428629398346, + 0.01107117161154747, + 0.010792475193738937, + 0.013048247434198856, + 0.007959902286529541, + 0.0157018955796957, + 0.012155769392848015, + 0.00833812728524208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.02134132757782936 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373117744922638 + ], + "opposite_logprob_tokens": [ + -3.40873122215271 + ], + "logprob_reward": -0.03373117744922638, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373117744922638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05759628489613533, + 0.05898881331086159, + 0.07805486768484116, + 0.156366229057312, + 0.020922917872667313, + 0.04749800264835358, + 0.3500754237174988, + 0.09751646220684052, + 0.05945234373211861, + 0.07352865487337112 + ], + "credit_map": [ + 0.00648508919402957, + 0.011280975304543972, + 0.009378703311085701, + 0.006432570982724428, + 0.005096818786114454, + 0.006570495665073395, + 0.007742257788777351, + 0.004609374329447746, + 0.005498618818819523, + 0.011246155016124249, + 0.007539578713476658, + 0.008113255724310875, + 0.005731555633246899, + 0.005963567178696394, + 0.009469503536820412, + 0.005426577292382717, + 0.0070809791795909405, + 0.011127598583698273, + 0.010582535527646542, + 0.012652707286179066, + 0.0066983443684875965, + 0.007112884894013405, + 0.014443786814808846, + 0.008356030099093914, + 0.014609811827540398, + 0.016285941004753113, + 0.03703533858060837, + 0.02921992354094982, + 0.007750572636723518, + 0.014051148667931557, + 0.029800260439515114, + 0.007613221649080515, + 0.002816116437315941, + 0.0021756980568170547, + 0.002970232395455241, + 0.0038316177669912577, + 0.001708734198473394, + 0.002365685300901532, + 0.0026216309051960707, + 0.002433202462270856, + 0.008345082402229309, + 0.006312356796115637, + 0.005680697038769722, + 0.0059790462255477905, + 0.0053383903577923775, + 0.00654255086556077, + 0.005308515392243862, + 0.003991361241787672, + 0.05110672488808632, + 0.05843384563922882, + 0.05860458314418793, + 0.050884269177913666, + 0.028308534994721413, + 0.037546396255493164, + 0.04057396203279495, + 0.024617111310362816, + 0.028324216604232788, + 0.010238109156489372, + 0.010506969876587391, + 0.02579517476260662, + 0.004672392271459103, + 0.007058273535221815, + 0.006213304586708546, + 0.0047080242075026035, + 0.006380777806043625, + 0.006121636368334293, + 0.012058512307703495, + 0.009085462428629398, + 0.005918164271861315, + 0.007955743931233883, + 0.006172175519168377, + 0.005759873427450657, + 0.007594890892505646, + 0.01336058508604765, + 0.012795256450772285, + 0.018591778352856636, + 0.005154955666512251, + 0.0052932552061975, + 0.006350535433739424, + 0.004387400578707457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.03373117744922638 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864406768232584 + ], + "opposite_logprob_tokens": [ + -4.758644104003906 + ], + "logprob_reward": -0.00864406768232584, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864406768232584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02252277359366417, + 0.029247602447867393, + 0.05921631678938866, + 0.09991278499364853, + 0.05970648303627968, + 0.05506677180528641, + 0.4394975006580353, + 0.1543222963809967, + 0.013856802135705948, + 0.06665065884590149 + ], + "credit_map": [ + 0.0038861213251948357, + 0.0022401113528758287, + 0.0031575271859765053, + 0.0031585288234055042, + 0.0021465138997882605, + 0.0023714639246463776, + 0.0030119477305561304, + 0.002550558652728796, + 0.003587102284654975, + 0.002863276284188032, + 0.0037972156424075365, + 0.007938569411635399, + 0.0019612715113908052, + 0.0037280742544680834, + 0.0036869309842586517, + 0.0016851626569405198, + 0.006001772824674845, + 0.011106595396995544, + 0.009787117131054401, + 0.011880092322826385, + 0.00418591033667326, + 0.005929055158048868, + 0.00581597164273262, + 0.00450980244204402, + 0.009647581726312637, + 0.011890731751918793, + 0.02086205966770649, + 0.017121020704507828, + 0.006682084873318672, + 0.009635274298489094, + 0.016326207667589188, + 0.007747827097773552, + 0.011458833701908588, + 0.006381495855748653, + 0.00877875741571188, + 0.008386106230318546, + 0.003957436420023441, + 0.004849359393119812, + 0.010198872536420822, + 0.005695624276995659, + 0.008207404054701328, + 0.004128835164010525, + 0.008140232414007187, + 0.0070940167643129826, + 0.007674416061490774, + 0.008069027215242386, + 0.00553616089746356, + 0.006216680631041527, + 0.06087740510702133, + 0.10987480729818344, + 0.054011180996894836, + 0.06483955681324005, + 0.03217259421944618, + 0.044935036450624466, + 0.04582696780562401, + 0.026959966868162155, + 0.05249139666557312, + 0.012252037413418293, + 0.012876485474407673, + 0.04391489923000336, + 0.005652285180985928, + 0.010178557597100735, + 0.009759019128978252, + 0.007197614759206772, + 0.0017772739520296454, + 0.0013980779331177473, + 0.0021803993731737137, + 0.0019525433890521526, + 0.0013118383940309286, + 0.002133487956598401, + 0.0018134706187993288, + 0.0012897104024887085, + 0.0072889733128249645, + 0.013121314346790314, + 0.010919651947915554, + 0.01933855377137661, + 0.0034441077150404453, + 0.0037657697685062885, + 0.005403721239417791, + 0.003368563251569867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.00864406768232584 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632853463292122 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.007632853463292122, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632853463292122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08512084186077118, + 0.03758775442838669, + 0.09315536171197891, + 0.2060999870300293, + 0.031441472470760345, + 0.0269166212528944, + 0.15135903656482697, + 0.2778063714504242, + 0.0457119382917881, + 0.04480057209730148 + ], + "credit_map": [ + 0.013551388867199421, + 0.010173297487199306, + 0.013707593083381653, + 0.017624501138925552, + 0.007034782320261002, + 0.007190281990915537, + 0.010192757472395897, + 0.005646241828799248, + 0.00411575473845005, + 0.0040876236744225025, + 0.005704355426132679, + 0.005141460802406073, + 0.0032339587341994047, + 0.004273653030395508, + 0.0070931510999798775, + 0.0039378004148602486, + 0.009308644570410252, + 0.013889200985431671, + 0.011637036688625813, + 0.01616240292787552, + 0.008158682845532894, + 0.010045332834124565, + 0.012515757232904434, + 0.011438309215009212, + 0.017357248812913895, + 0.020644022151827812, + 0.046546805649995804, + 0.04118412733078003, + 0.011560194194316864, + 0.02027580887079239, + 0.039511922746896744, + 0.009019867517054081, + 0.0043410672806203365, + 0.00437181256711483, + 0.0046964348293840885, + 0.0055542695336043835, + 0.002430943539366126, + 0.003101021284237504, + 0.0031966487877070904, + 0.0037492779083549976, + 0.003855970920994878, + 0.00276859849691391, + 0.003863280639052391, + 0.0034959695767611265, + 0.0028782042209059, + 0.004411027766764164, + 0.0028083452489227057, + 0.0028352264780551195, + 0.02295866422355175, + 0.02474215440452099, + 0.022880947217345238, + 0.022751521319150925, + 0.013670392334461212, + 0.014534756541252136, + 0.0197545625269413, + 0.010066051967442036, + 0.05668807774782181, + 0.04332032427191734, + 0.03846509009599686, + 0.06264171749353409, + 0.016205770894885063, + 0.0199383907020092, + 0.024410130456089973, + 0.016136907041072845, + 0.0074121104553341866, + 0.006776801310479641, + 0.009624441154301167, + 0.005729347001761198, + 0.0032731019891798496, + 0.003940013702958822, + 0.005564493592828512, + 0.003391633043065667, + 0.005460076965391636, + 0.009531470946967602, + 0.007664794102311134, + 0.009585080668330193, + 0.002210648264735937, + 0.003696304513141513, + 0.004271678626537323, + 0.0023805235978215933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.007632853463292122 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017143089789897203 + ], + "opposite_logprob_tokens": [ + -6.37671422958374 + ], + "logprob_reward": -0.0017143089789897203, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017143089789897203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07265803217887878, + 0.03260933235287666, + 0.05071830749511719, + 0.22976356744766235, + 0.023580091074109077, + 0.01198150310665369, + 0.10103201866149902, + 0.08353579044342041, + 0.07113432139158249, + 0.32298702001571655 + ], + "credit_map": [ + 0.013215542770922184, + 0.008241831324994564, + 0.012403671629726887, + 0.008958869613707066, + 0.004791688174009323, + 0.007707706652581692, + 0.010467445477843285, + 0.006871283520013094, + 0.0033332081511616707, + 0.003942703828215599, + 0.0039706458337605, + 0.0037463721819221973, + 0.003528479253873229, + 0.004300855100154877, + 0.005235819611698389, + 0.004551249090582132, + 0.006166762672364712, + 0.00622527813538909, + 0.0053967200219631195, + 0.009137693792581558, + 0.004235512111335993, + 0.005429141689091921, + 0.009274844080209732, + 0.004852355923503637, + 0.022327402606606483, + 0.025808412581682205, + 0.057332225143909454, + 0.04745550826191902, + 0.013687275350093842, + 0.017526980489492416, + 0.034561894834041595, + 0.011063878424465656, + 0.003586175385862589, + 0.0021432642824947834, + 0.0033200429752469063, + 0.003088220953941345, + 0.0016276900423690677, + 0.0021187919192016125, + 0.005563344340771437, + 0.0021325615234673023, + 0.0020405941177159548, + 0.0020192223601043224, + 0.0014283886412158608, + 0.0013759479625150561, + 0.0014199288561940193, + 0.0014427687274292111, + 0.0013093282468616962, + 0.0009453243692405522, + 0.012593681924045086, + 0.017424121499061584, + 0.014296973124146461, + 0.013478920795023441, + 0.008347875438630581, + 0.014382927678525448, + 0.012606288306415081, + 0.007901228033006191, + 0.027264554053544998, + 0.007821942679584026, + 0.00684847729280591, + 0.022526217624545097, + 0.0035528866574168205, + 0.006314405705779791, + 0.005120605230331421, + 0.004086704924702644, + 0.010387648828327656, + 0.009249283000826836, + 0.015846773982048035, + 0.009493093937635422, + 0.004807818215340376, + 0.007968113757669926, + 0.008553838357329369, + 0.00482774805277586, + 0.03688404709100723, + 0.07519011944532394, + 0.05869239568710327, + 0.06695715337991714, + 0.01550796814262867, + 0.02603425271809101, + 0.027830820530653, + 0.01589028351008892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.0017143089789897203 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, captured in the style of high-speed photography, focuses on the metal nozzle of a stainless steel watering can, its surface marked with minute scratches and a few dried water stains. A powerful stream of water erupts from its fine orifice, instantly freezing in mid-air. Each individual droplet is imbued with life, stretched into elliptical or even longer streamlines in mid-air, resembling a group of runners sprinting in a fierce race, each maintaining a tiny distance between them, creating a dynamic trajectory. Bright, sharp side lighting precisely outlines the droplets, creating diamond-like highlights on their surface. Light penetrates their crystal-clear interior, refracting tiny specks of light, while the metal nozzle takes on a cool, metallic sheen. A soft, out-of-focus blur of greenery in the background provides a tranquil backdrop to this explosive moment. The entire image is imbued with breathtaking dynamism and tension, capturing the instantaneous energy of the water droplets with artistic beauty, showcasing the breathtaking vitality of the microscopic world.", + "seed": 24410, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_10/index_4667_seed_24410_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.015346487192437053, + "vlm_logprob": -0.015346487192437053, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.015346487192437053, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01571532152593136 + ], + "opposite_logprob_tokens": [ + -4.2657151222229 + ], + "logprob_reward": -0.01571532152593136, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01571532152593136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043499063700437546, + 0.0338776670396328, + 0.053078994154930115, + 0.03439842537045479, + 0.04200667515397072, + 0.013210284523665905, + 0.6386386752128601, + 0.08045966178178787, + 0.03724974766373634, + 0.023580869659781456 + ], + "credit_map": [ + 0.003397274063900113, + 0.005133516620844603, + 0.0031857588328421116, + 0.006262500770390034, + 0.009987988509237766, + 0.004319384694099426, + 0.004939849954098463, + 0.006272788159549236, + 0.002392710419371724, + 0.0029439402278512716, + 0.002642083680257201, + 0.004808820318430662, + 0.0035881453659385443, + 0.0043071298860013485, + 0.0028311070054769516, + 0.010363726876676083, + 0.004183758515864611, + 0.00495341420173645, + 0.004751379135996103, + 0.010602734982967377, + 0.006452708039432764, + 0.005309516564011574, + 0.006586731877177954, + 0.010238750837743282, + 0.0027921011205762625, + 0.004475616384297609, + 0.0033048035111278296, + 0.007201015949249268, + 0.00319096096791327, + 0.0032256986014544964, + 0.0038579325191676617, + 0.006350294221192598, + 0.003828863613307476, + 0.004587242379784584, + 0.004448304884135723, + 0.008784310892224312, + 0.005297055933624506, + 0.005229081958532333, + 0.0050179618410766125, + 0.004813853185623884, + 0.0013987034326419234, + 0.0023643001914024353, + 0.0015863444423303008, + 0.0018094084225594997, + 0.0014824362006038427, + 0.00165234983433038, + 0.0014282373012974858, + 0.001488504116423428, + 0.023504428565502167, + 0.017016224563121796, + 0.015280524268746376, + 0.17405779659748077, + 0.03581736981868744, + 0.10903974622488022, + 0.17229782044887543, + 0.09162474423646927, + 0.005993361119180918, + 0.008719724602997303, + 0.007194078993052244, + 0.018577782437205315, + 0.006278923712670803, + 0.00484713027253747, + 0.02108304388821125, + 0.007765620015561581, + 0.0040058232843875885, + 0.006871677935123444, + 0.002892792224884033, + 0.008143266662955284, + 0.0036304674576967955, + 0.0017278201412409544, + 0.006435716524720192, + 0.003542184829711914, + 0.0018934134859591722, + 0.0016117399791255593, + 0.0027192109264433384, + 0.004695572890341282, + 0.0018907577032223344, + 0.0031866461504250765, + 0.0046773673966526985, + 0.0029061625245958567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.01571532152593136 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160839706659317 + ], + "opposite_logprob_tokens": [ + -4.14108419418335 + ], + "logprob_reward": -0.0160839706659317, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160839706659317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04146218299865723, + 0.04994795098900795, + 0.12183313816785812, + 0.030285093933343887, + 0.11630138754844666, + 0.016889581456780434, + 0.47778335213661194, + 0.0677933394908905, + 0.051247164607048035, + 0.026456832885742188 + ], + "credit_map": [ + 0.0033642603084445, + 0.0038250533398240805, + 0.0039827157743275166, + 0.005979044828563929, + 0.007223775610327721, + 0.0049054487608373165, + 0.004049122799187899, + 0.00813276320695877, + 0.003496568650007248, + 0.006758596282452345, + 0.006880702450871468, + 0.0065526473335921764, + 0.004731245804578066, + 0.0072561027482151985, + 0.004906386137008667, + 0.009365701116621494, + 0.009090051986277103, + 0.011560327373445034, + 0.007339910604059696, + 0.029617907479405403, + 0.010997375473380089, + 0.0107602309435606, + 0.014239328913390636, + 0.028228003531694412, + 0.0026140979025512934, + 0.004948464222252369, + 0.0037706783041357994, + 0.004169792868196964, + 0.0023304352071136236, + 0.002364506246522069, + 0.003586967010051012, + 0.0065001524053514, + 0.00914870947599411, + 0.008403720334172249, + 0.007771030999720097, + 0.031732067465782166, + 0.010236397385597229, + 0.011715607717633247, + 0.013532963581383228, + 0.023760901764035225, + 0.000740089628379792, + 0.0010028843535110354, + 0.0009681322844699025, + 0.004498281050473452, + 0.0009599991026334465, + 0.002375735668465495, + 0.002589165000244975, + 0.003755294717848301, + 0.017860254272818565, + 0.01282685250043869, + 0.011458784341812134, + 0.13031712174415588, + 0.026707446202635765, + 0.08095954358577728, + 0.13262496888637543, + 0.0650283470749855, + 0.00477290665730834, + 0.00702216662466526, + 0.0060514770448207855, + 0.019095640629529953, + 0.005679911468178034, + 0.004722915589809418, + 0.015067901462316513, + 0.005380416288971901, + 0.0056105200201272964, + 0.009376353584229946, + 0.0041024405509233475, + 0.009853456169366837, + 0.004905794281512499, + 0.0028253872878849506, + 0.008416579104959965, + 0.006156631745398045, + 0.0022712554782629013, + 0.002470700303092599, + 0.0028184656985104084, + 0.004396461881697178, + 0.0021461520809680223, + 0.0036871498450636864, + 0.005330882500857115, + 0.003335764864459634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.0160839706659317 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809787333011627 + ], + "opposite_logprob_tokens": [ + -3.288097858428955 + ], + "logprob_reward": -0.03809787333011627, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809787333011627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07605159282684326, + 0.14244677126407623, + 0.10212048143148422, + 0.1246073916554451, + 0.19128528237342834, + 0.04802171140909195, + 0.1322707235813141, + 0.08986147493124008, + 0.03776245191693306, + 0.05557209998369217 + ], + "credit_map": [ + 0.006604115944355726, + 0.007686262018978596, + 0.011534721590578556, + 0.0145425321534276, + 0.0075674476101994514, + 0.010118245147168636, + 0.006913067772984505, + 0.011085200123488903, + 0.013195956125855446, + 0.019671089947223663, + 0.02551131695508957, + 0.016348905861377716, + 0.019057195633649826, + 0.015044976957142353, + 0.01931961067020893, + 0.01429772563278675, + 0.009807346388697624, + 0.007464740891009569, + 0.005788094829767942, + 0.026251617819070816, + 0.008480372838675976, + 0.01023454125970602, + 0.009869394823908806, + 0.024224374443292618, + 0.010975525714457035, + 0.0200964268296957, + 0.013483122922480106, + 0.023540573194622993, + 0.011608406901359558, + 0.013529605232179165, + 0.014443599618971348, + 0.01693013310432434, + 0.014539238065481186, + 0.015310628339648247, + 0.013960347510874271, + 0.04373800754547119, + 0.021378260105848312, + 0.02536916732788086, + 0.022881576791405678, + 0.03410805016756058, + 0.0030289110727608204, + 0.0038170816842466593, + 0.0035528226289898157, + 0.008016921579837799, + 0.004405227489769459, + 0.011211689561605453, + 0.006852441933006048, + 0.007136618718504906, + 0.005124898627400398, + 0.005950235761702061, + 0.004038976971060038, + 0.029267387464642525, + 0.009864331223070621, + 0.026482030749320984, + 0.023992162197828293, + 0.027550682425498962, + 0.006366926711052656, + 0.00672954274341464, + 0.007100467570126057, + 0.018836582079529762, + 0.006540052127093077, + 0.007691634353250265, + 0.02642430178821087, + 0.010171971283853054, + 0.002844772068783641, + 0.002527463249862194, + 0.002778641413897276, + 0.01293566171079874, + 0.003241426544263959, + 0.0018716357881203294, + 0.008361602202057838, + 0.003201249288395047, + 0.003948960918933153, + 0.0044783782213926315, + 0.004660881590098143, + 0.010514681227505207, + 0.004894149024039507, + 0.006753374822437763, + 0.010469168424606323, + 0.009852507151663303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.03809787333011627 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198899397626519 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198899397626519, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198899397626519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06985706835985184, + 0.05478021875023842, + 0.07764044404029846, + 0.04528091102838516, + 0.07390584796667099, + 0.024972224608063698, + 0.47635120153427124, + 0.11558348685503006, + 0.02379082702100277, + 0.037837762385606766 + ], + "credit_map": [ + 0.005012592300772667, + 0.006480843760073185, + 0.003993131220340729, + 0.01257813535630703, + 0.01103954017162323, + 0.004851675592362881, + 0.006577137392014265, + 0.019324012100696564, + 0.005102019757032394, + 0.00825602188706398, + 0.009197399951517582, + 0.00806103739887476, + 0.00538727268576622, + 0.004811817780137062, + 0.007483490277081728, + 0.006481159012764692, + 0.006905360613018274, + 0.0073683299124240875, + 0.004396727774292231, + 0.018219178542494774, + 0.007629730273038149, + 0.00765290716663003, + 0.007547449786216021, + 0.017920764163136482, + 0.004309788811951876, + 0.0038615898229181767, + 0.0051478189416229725, + 0.009521055035293102, + 0.0034141268115490675, + 0.0037108866963535547, + 0.004524488467723131, + 0.010791156440973282, + 0.00618553813546896, + 0.00487734517082572, + 0.0050377086736261845, + 0.013804610818624496, + 0.009145197458565235, + 0.015904400497674942, + 0.008648845367133617, + 0.010302199982106686, + 0.0019209065940231085, + 0.0013931861612945795, + 0.0014341670321300626, + 0.007051894906908274, + 0.002037110971286893, + 0.0033619804307818413, + 0.0030692138243466616, + 0.0047037675976753235, + 0.017674226313829422, + 0.013169929385185242, + 0.011854569427669048, + 0.1252482384443283, + 0.027168091386556625, + 0.08482711762189865, + 0.12112042307853699, + 0.07528861612081528, + 0.00981153640896082, + 0.00946071371436119, + 0.007486642803996801, + 0.023693326860666275, + 0.010212915949523449, + 0.012871627695858479, + 0.024764057248830795, + 0.01728266291320324, + 0.0025778335984796286, + 0.002384881256148219, + 0.0021304762922227383, + 0.0052665118128061295, + 0.0017744962824508548, + 0.0021086218766868114, + 0.004010570701211691, + 0.0035374353174120188, + 0.0021933638490736485, + 0.0030787510331720114, + 0.0032429159618914127, + 0.0057046422734856606, + 0.0032386176753789186, + 0.007216757163405418, + 0.00825995858758688, + 0.004902755841612816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.002198899397626519 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046363710425794125 + ], + "opposite_logprob_tokens": [ + -5.379636287689209 + ], + "logprob_reward": -0.0046363710425794125, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046363710425794125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.016276704147458076, + 0.047293923795223236, + 0.04650663211941719, + 0.010795465670526028, + 0.017303362488746643, + 0.009423132054507732, + 0.8128377199172974, + 0.016537673771381378, + 0.015223251655697823, + 0.007802132051438093 + ], + "credit_map": [ + 0.0015800973633304238, + 0.002495927968993783, + 0.0020346972160041332, + 0.0026362293865531683, + 0.002540252637118101, + 0.0017094534123316407, + 0.0016526730032637715, + 0.001627374324016273, + 0.004229968413710594, + 0.004368055611848831, + 0.0061370558105409145, + 0.007136735133826733, + 0.003999008797109127, + 0.004747903905808926, + 0.003938979934900999, + 0.012736218050122261, + 0.0032288923393934965, + 0.0039490023627877235, + 0.002747922670096159, + 0.010861910879611969, + 0.005083362106233835, + 0.005191544536501169, + 0.0050113145262002945, + 0.010432684794068336, + 0.000714961439371109, + 0.0012577936286106706, + 0.0013689358020201325, + 0.001619689166545868, + 0.0009724085684865713, + 0.0016948111588135362, + 0.0015125795034691691, + 0.0016542862867936492, + 0.0013074829475954175, + 0.0010000206530094147, + 0.0011411536252126098, + 0.0037988515105098486, + 0.002017585327848792, + 0.002702927216887474, + 0.0021450696513056755, + 0.003190271556377411, + 0.0008462754194624722, + 0.0008927238523028791, + 0.0008726906962692738, + 0.0020978201646357775, + 0.0007823342457413673, + 0.001063079689629376, + 0.0011503314599394798, + 0.0017178765265271068, + 0.03026481717824936, + 0.021006450057029724, + 0.01858578808605671, + 0.23082683980464935, + 0.04279137775301933, + 0.12774020433425903, + 0.23922888934612274, + 0.10239333659410477, + 0.001753167132847011, + 0.002108229324221611, + 0.0016296000685542822, + 0.0032744756899774075, + 0.001476844772696495, + 0.001365532516501844, + 0.0031090970151126385, + 0.001820727833546698, + 0.001494577736593783, + 0.0021087434142827988, + 0.0011901145335286856, + 0.0029816413298249245, + 0.0012673401506617665, + 0.0011231276439502835, + 0.0028571581933647394, + 0.002200549002736807, + 0.00046836695400997996, + 0.0004488361009862274, + 0.0005712077836506069, + 0.0013354772236198187, + 0.0008185829501599073, + 0.0011080937692895532, + 0.0019143311074003577, + 0.0011372360168024898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.0046363710425794125 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, captured in the style of high-speed photography, focuses on the metal nozzle of a stainless steel watering can, its surface marked with minute scratches and a few dried water stains. A powerful stream of water erupts from its fine orifice, instantly freezing in mid-air. Each individual droplet is imbued with life, stretched into elliptical or even longer streamlines in mid-air, resembling a group of runners sprinting in a fierce race, each maintaining a tiny distance between them, creating a dynamic trajectory. Bright, sharp side lighting precisely outlines the droplets, creating diamond-like highlights on their surface. Light penetrates their crystal-clear interior, refracting tiny specks of light, while the metal nozzle takes on a cool, metallic sheen. A soft, out-of-focus blur of greenery in the background provides a tranquil backdrop to this explosive moment. The entire image is imbued with breathtaking dynamism and tension, capturing the instantaneous energy of the water droplets with artistic beauty, showcasing the breathtaking vitality of the microscopic world.", + "seed": 24411, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_10/index_4667_seed_24411_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.019369303341954948, + "vlm_logprob": -0.019369303341954948, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.019369303341954948, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03833687677979469 + ], + "opposite_logprob_tokens": [ + -3.413336992263794 + ], + "logprob_reward": -0.03833687677979469, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03833687677979469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13090144097805023, + 0.1434592306613922, + 0.05493238940834999, + 0.0940818339586258, + 0.06379418075084686, + 0.12685219943523407, + 0.026957930997014046, + 0.011475428007543087, + 0.019960274919867516, + 0.3275851607322693 + ], + "credit_map": [ + 0.014888175763189793, + 0.010420582257211208, + 0.009358138777315617, + 0.01787257008254528, + 0.01080528274178505, + 0.029768090695142746, + 0.014085053466260433, + 0.02370353229343891, + 0.011527926661074162, + 0.014897876419126987, + 0.010052336379885674, + 0.02635073848068714, + 0.012576091103255749, + 0.022506065666675568, + 0.023246681317687035, + 0.02230149880051613, + 0.00596048217266798, + 0.004713430069386959, + 0.004772160667926073, + 0.005104945506900549, + 0.004694975912570953, + 0.011268793605268002, + 0.009996711276471615, + 0.008420885540544987, + 0.01196490228176117, + 0.01139691099524498, + 0.008580300956964493, + 0.010205480270087719, + 0.013886291533708572, + 0.01326440554112196, + 0.014427021145820618, + 0.010356517508625984, + 0.007836505770683289, + 0.004390812013298273, + 0.006796133704483509, + 0.008496599271893501, + 0.005115392617881298, + 0.011497261002659798, + 0.011652253568172455, + 0.00800921767950058, + 0.015979312360286713, + 0.007076796144247055, + 0.010376574471592903, + 0.013535713776946068, + 0.009777382016181946, + 0.036628056317567825, + 0.012629607692360878, + 0.020848751068115234, + 0.003293609479442239, + 0.0018073454266414046, + 0.0018167633097618818, + 0.00342352781444788, + 0.0018061259761452675, + 0.005217180587351322, + 0.005815724842250347, + 0.0037776522804051638, + 0.0016166159184649587, + 0.001016212161630392, + 0.001596841961145401, + 0.0015928562497720122, + 0.001104688155464828, + 0.0014826669357717037, + 0.0016672213096171618, + 0.001398324267938733, + 0.00255512073636055, + 0.0013800967717543244, + 0.002208407735452056, + 0.003280115779489279, + 0.0015297112986445427, + 0.004225473385304213, + 0.0019028513925150037, + 0.0028784957248717546, + 0.037093404680490494, + 0.02376491390168667, + 0.023976746946573257, + 0.030757999047636986, + 0.025253543630242348, + 0.08837078511714935, + 0.036863185465335846, + 0.06150456890463829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.03833687677979469 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038129087537527084 + ], + "opposite_logprob_tokens": [ + -3.2881290912628174 + ], + "logprob_reward": -0.038129087537527084, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038129087537527084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12842753529548645, + 0.0632549300789833, + 0.07736524939537048, + 0.08656615763902664, + 0.04371970519423485, + 0.10412783920764923, + 0.04696892201900482, + 0.033074237406253815, + 0.03412103280425072, + 0.38237443566322327 + ], + "credit_map": [ + 0.01875128410756588, + 0.009738186374306679, + 0.0119863860309124, + 0.010897691361606121, + 0.0071823387406766415, + 0.028508275747299194, + 0.018650133162736893, + 0.02271324396133423, + 0.0075988443568348885, + 0.008831148035824299, + 0.004315989091992378, + 0.011861731298267841, + 0.00540175149217248, + 0.0071019078604876995, + 0.00935125071555376, + 0.008792303502559662, + 0.008834205567836761, + 0.007207578048110008, + 0.007027344778180122, + 0.008209114894270897, + 0.006940638646483421, + 0.015386792831122875, + 0.012057396583259106, + 0.011702177114784718, + 0.010693584568798542, + 0.007977042347192764, + 0.0071539669297635555, + 0.008769419975578785, + 0.008547834120690823, + 0.01811087504029274, + 0.010163926519453526, + 0.01514950767159462, + 0.003690397599712014, + 0.0023115873336791992, + 0.0031220451928675175, + 0.005728520452976227, + 0.0033023692667484283, + 0.006432525347918272, + 0.013160711154341698, + 0.005971550941467285, + 0.012916810810565948, + 0.005661038216203451, + 0.008037407882511616, + 0.01152925193309784, + 0.007971769198775291, + 0.02876805141568184, + 0.013424507342278957, + 0.01581900380551815, + 0.004562663845717907, + 0.004520848859101534, + 0.004455958027392626, + 0.004540131893008947, + 0.005158864893019199, + 0.011351685039699078, + 0.005146966315805912, + 0.007231800816953182, + 0.002906966954469681, + 0.0031485685613006353, + 0.0031541981734335423, + 0.005254900548607111, + 0.004372951574623585, + 0.0037514062132686377, + 0.005608350969851017, + 0.004876893479377031, + 0.0038062753155827522, + 0.0030221836641430855, + 0.0033898951951414347, + 0.00681056035682559, + 0.0026565934531390667, + 0.005834621377289295, + 0.004371757619082928, + 0.004229146521538496, + 0.030853597447276115, + 0.01385543029755354, + 0.017141569405794144, + 0.027342617511749268, + 0.01945405639708042, + 0.12855248153209686, + 0.03874818980693817, + 0.10642648488283157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.038129087537527084 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955924745649099 + ], + "opposite_logprob_tokens": [ + -5.130955696105957 + ], + "logprob_reward": -0.005955924745649099, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955924745649099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07966972142457962, + 0.07028194516897202, + 0.04772325977683067, + 0.17849549651145935, + 0.07336653769016266, + 0.02924981527030468, + 0.013758036307990551, + 0.010624024085700512, + 0.010857319459319115, + 0.4859738349914551 + ], + "credit_map": [ + 0.010046574287116528, + 0.007529535796493292, + 0.009232687763869762, + 0.01177017018198967, + 0.006214759312570095, + 0.009856805205345154, + 0.010885928757488728, + 0.01413325872272253, + 0.006203396711498499, + 0.007489131763577461, + 0.00542171485722065, + 0.013253730721771717, + 0.00489800563082099, + 0.012138968333601952, + 0.011849373579025269, + 0.009027626365423203, + 0.005523442290723324, + 0.004406140651553869, + 0.005020539741963148, + 0.005324084311723709, + 0.004327854141592979, + 0.010919199325144291, + 0.005277447402477264, + 0.006924555171281099, + 0.014058765955269337, + 0.005835435818880796, + 0.008029934018850327, + 0.00983793381601572, + 0.011187306605279446, + 0.04898358881473541, + 0.030412010848522186, + 0.05015050992369652, + 0.007858594879508018, + 0.00617564283311367, + 0.006633552256971598, + 0.008818307891488075, + 0.007725640200078487, + 0.009085319936275482, + 0.01766456477344036, + 0.009404915384948254, + 0.0032827206887304783, + 0.0029618211556226015, + 0.002310327487066388, + 0.0037953630089759827, + 0.002819824032485485, + 0.005443249363452196, + 0.004378188401460648, + 0.004258321598172188, + 0.0014151503564789891, + 0.0012428348418325186, + 0.0019920251797884703, + 0.0019325532484799623, + 0.001138178282417357, + 0.0020733901765197515, + 0.002383182290941477, + 0.0015807229792699218, + 0.0012627900578081608, + 0.0011881941463798285, + 0.0012033366365358233, + 0.0013213127385824919, + 0.0014275421854108572, + 0.001635337364859879, + 0.001296732691116631, + 0.0012887786142528057, + 0.0010364697081968188, + 0.0009337210794910789, + 0.0007745708571746945, + 0.0019087096443399787, + 0.0010658077662810683, + 0.0025049541145563126, + 0.0011067783925682306, + 0.001526308013126254, + 0.03753146529197693, + 0.013459987938404083, + 0.01407148502767086, + 0.0216824971139431, + 0.028315357863903046, + 0.19431111216545105, + 0.03841859847307205, + 0.1381833404302597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.005955924745649099 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787782095372677 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.009787782095372677, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787782095372677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12099400162696838, + 0.1328839808702469, + 0.029101597145199776, + 0.40607815980911255, + 0.045276764780282974, + 0.025705723091959953, + 0.02715347334742546, + 0.009837541729211807, + 0.015851208940148354, + 0.18711751699447632 + ], + "credit_map": [ + 0.021882615983486176, + 0.010152215138077736, + 0.010711233131587505, + 0.01742103137075901, + 0.007129107601940632, + 0.01297647412866354, + 0.015964776277542114, + 0.024756552651524544, + 0.014762736856937408, + 0.011359998025000095, + 0.009354677051305771, + 0.017752280458807945, + 0.011247079819440842, + 0.024180663749575615, + 0.02350029908120632, + 0.02072625607252121, + 0.002475993474945426, + 0.0033700154162943363, + 0.0028387303464114666, + 0.002920686500146985, + 0.0025643552653491497, + 0.00781957060098648, + 0.002812409307807684, + 0.004299837164580822, + 0.033084575086832047, + 0.014126746915280819, + 0.01494293101131916, + 0.02608545310795307, + 0.023395320400595665, + 0.098709337413311, + 0.0821608304977417, + 0.11357299238443375, + 0.00510729243978858, + 0.0038819301407784224, + 0.004347202833741903, + 0.0049458397552371025, + 0.0038177790120244026, + 0.00630813417956233, + 0.010853981599211693, + 0.006014604587107897, + 0.0032154214568436146, + 0.0016092958394438028, + 0.002248194767162204, + 0.0023615958634763956, + 0.001963823102414608, + 0.005305472761392593, + 0.004071395378559828, + 0.004930525552481413, + 0.002876078709959984, + 0.002469341503456235, + 0.0027192726265639067, + 0.003019382245838642, + 0.002682939637452364, + 0.005796294659376144, + 0.003633094485849142, + 0.00395706994459033, + 0.0010186485014855862, + 0.0011665538186207414, + 0.0006393205840140581, + 0.0012420144630596042, + 0.0010991684393957257, + 0.0015038893325254321, + 0.0018790983594954014, + 0.0012888484634459019, + 0.002079268917441368, + 0.0013785696355625987, + 0.0016454692231491208, + 0.0033810860477387905, + 0.0010109309805557132, + 0.002166382735595107, + 0.0019122612429782748, + 0.0022772413212805986, + 0.0156516432762146, + 0.008605899289250374, + 0.007686926051974297, + 0.014760635793209076, + 0.010203054174780846, + 0.058361876755952835, + 0.02376033179461956, + 0.048087164759635925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.009787782095372677 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up of a metal nozzle from a stainless steel watering can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the nozzle surface show minute scratches and dried water stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a powerful stream of water erupting from the nozzle's orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the water droplets frozen in mid-air with stretched elliptical or streamline shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus blur of greenery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636845551431179 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636845551431179, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636845551431179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16922685503959656, + 0.08652997016906738, + 0.08287627249956131, + 0.07174213975667953, + 0.05380627140402794, + 0.11907733976840973, + 0.0624859556555748, + 0.03318728506565094, + 0.07072948664426804, + 0.25033843517303467 + ], + "credit_map": [ + 0.03173120692372322, + 0.01052024308592081, + 0.015116407535970211, + 0.014923309907317162, + 0.012011153623461723, + 0.026940064504742622, + 0.020204083994030952, + 0.037780385464429855, + 0.008341151289641857, + 0.007678459398448467, + 0.007356678135693073, + 0.013811444863677025, + 0.005734457168728113, + 0.0167128574103117, + 0.014181653037667274, + 0.012713272124528885, + 0.008026870898902416, + 0.0072731697000563145, + 0.005884652957320213, + 0.009910265915095806, + 0.007525645662099123, + 0.01810584031045437, + 0.013138145208358765, + 0.0130116892978549, + 0.005844763945788145, + 0.0056271483190357685, + 0.00318612321279943, + 0.005863019730895758, + 0.006614483427256346, + 0.024155251681804657, + 0.00555511936545372, + 0.014896229840815067, + 0.008122699335217476, + 0.004577372223138809, + 0.004057277925312519, + 0.006065766327083111, + 0.004004338756203651, + 0.00877930223941803, + 0.011588530614972115, + 0.00661098025739193, + 0.01246667094528675, + 0.0076261223293840885, + 0.007198850624263287, + 0.011366077698767185, + 0.011259688064455986, + 0.03894830867648125, + 0.010225888341665268, + 0.019985727965831757, + 0.0058113448321819305, + 0.005049800965934992, + 0.004697332624346018, + 0.006042504217475653, + 0.006067052483558655, + 0.015173136256635189, + 0.01000476535409689, + 0.009640021249651909, + 0.003993398044258356, + 0.002902993466705084, + 0.0037680300883948803, + 0.003766808193176985, + 0.0039837295189499855, + 0.004658022429794073, + 0.005575989373028278, + 0.004538311623036861, + 0.006860645487904549, + 0.0043910834938287735, + 0.004582406487315893, + 0.006813041400164366, + 0.007071810774505138, + 0.02328137680888176, + 0.006821049377322197, + 0.010908076539635658, + 0.020260296761989594, + 0.009692839346826077, + 0.01116227824240923, + 0.015345348976552486, + 0.016728714108467102, + 0.0886593908071518, + 0.023582126945257187, + 0.06490744650363922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.004636845551431179 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_11/index_tensor([5730, 5730])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_11/index_tensor([5730, 5730])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f48b693e50aeb33766859504dda9e5e30ac43e52 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_11/index_tensor([5730, 5730])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, dynamic tracking shot follows from the side and rear, the camera close to the ground, parallel to the galloping hooves. In the image, a muscular, inky-haired steed gallops at full speed across the vast, desolate Gobi Desert, its hooves flying and stirring up clouds of yellow dust. Aloft, a lone knight, leans forward, gripping the reins. His long, wind-twisted black hair, his bronzed face stern, his gaze piercing, like a hawk's, fixed on the distant horizon. He wears a heavy, rough black cape, which is violently lifted by the strong wind. It billows and tumbles behind him like an unfurled battle flag, forming a vast, powerful, and almost frozen black arc. On the horizon, a burning blood-red setting sun is slowly sinking under the endless sand dunes, dyeing the entire sky a strong orange-red and deep purple. The dazzling backlight outlines a circle of dazzling golden outline light for the knight and the horse. Their figures appear lonely and full of epic grandeur in the elongated shadows. The whole scene is filled with a desolate, resolute and unrestrained atmosphere in the high-speed movement.", + "seed": 7286, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_11/index_5730_seed_7286_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.014647019933909178, + "vlm_logprob": -0.014647019933909178, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.014647019933909178, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028528878465294838 + ], + "opposite_logprob_tokens": [ + -5.278528690338135 + ], + "logprob_reward": -0.028528878465294838, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.028528878465294838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08306003361940384, + 0.01793850213289261, + 0.019271058961749077, + 0.02267848700284958, + 0.013590681366622448, + 0.36664992570877075, + 0.04759575426578522, + 0.3845057189464569, + 0.024510355666279793, + 0.020199459046125412 + ], + "credit_map": [ + 0.020732222124934196, + 0.009929381310939789, + 0.014804592356085777, + 0.005478124134242535, + 0.009278238750994205, + 0.01005982793867588, + 0.004978232551366091, + 0.0077994223684072495, + 0.004076224751770496, + 0.00254208967089653, + 0.002822599606588483, + 0.0018753354670479894, + 0.0016469687689095736, + 0.00204282533377409, + 0.0014181678416207433, + 0.001514293602667749, + 0.0017965177539736032, + 0.002053750678896904, + 0.003931195009499788, + 0.002044346882030368, + 0.001578990719281137, + 0.003198994090780616, + 0.0015010476345196366, + 0.0031662185210734606, + 0.0017550545744597912, + 0.0031175408512353897, + 0.0026956889778375626, + 0.003293449990451336, + 0.001840798300690949, + 0.004577263258397579, + 0.0029755113646388054, + 0.0024231807328760624, + 0.0018379600951448083, + 0.0015517721185460687, + 0.0017334013245999813, + 0.0016296401154249907, + 0.0014995120000094175, + 0.0015891211805865169, + 0.0016855819849297404, + 0.00206369417719543, + 0.024315955117344856, + 0.020974868908524513, + 0.058419469743967056, + 0.027775874361395836, + 0.015761449933052063, + 0.048616766929626465, + 0.029667280614376068, + 0.14111827313899994, + 0.004462395329028368, + 0.00354949664324522, + 0.004124494269490242, + 0.004653614945709705, + 0.004400802776217461, + 0.015495932660996914, + 0.0035086842253804207, + 0.007400338537991047, + 0.03174961730837822, + 0.014351025223731995, + 0.06583818048238754, + 0.03407186269760132, + 0.022370537742972374, + 0.1082550659775734, + 0.020861908793449402, + 0.08700757473707199, + 0.004914689343422651, + 0.003670309903100133, + 0.002368601970374584, + 0.002269597025588155, + 0.0044533004984259605, + 0.0026436808984726667, + 0.0029965853318572044, + 0.0011935915099456906, + 0.00231238454580307, + 0.004053590353578329, + 0.0018237861804664135, + 0.0033651823177933693, + 0.003134704427793622, + 0.0024363985285162926, + 0.0019190923776477575, + 0.00115432299207896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.028528878465294838 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029821116477251053 + ], + "opposite_logprob_tokens": [ + -3.5298211574554443 + ], + "logprob_reward": -0.029821116477251053, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029821116477251053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10962914675474167, + 0.05779428780078888, + 0.022832447662949562, + 0.048315808176994324, + 0.03347465395927429, + 0.3786855936050415, + 0.036746542900800705, + 0.26551833748817444, + 0.011673474684357643, + 0.035329703241586685 + ], + "credit_map": [ + 0.026776159182190895, + 0.011545136570930481, + 0.021949902176856995, + 0.0065954821184277534, + 0.009876403957605362, + 0.013384012505412102, + 0.006793951150029898, + 0.012708095833659172, + 0.007125661242753267, + 0.004622143693268299, + 0.010987857356667519, + 0.0065893675200641155, + 0.005076265428215265, + 0.005451662000268698, + 0.005025112070143223, + 0.012916220352053642, + 0.003665318014100194, + 0.0019449297105893493, + 0.0041496665216982365, + 0.0013599795056506991, + 0.001568651176057756, + 0.002786393743008375, + 0.0013373485999181867, + 0.006020162720233202, + 0.0048581622540950775, + 0.006115582771599293, + 0.0063350610435009, + 0.005012295674532652, + 0.005520214792340994, + 0.010851520113646984, + 0.00442182831466198, + 0.005201142281293869, + 0.004805145785212517, + 0.0023346852976828814, + 0.004813132807612419, + 0.0025011375546455383, + 0.0023558088578283787, + 0.005255889147520065, + 0.0022101050708442926, + 0.0091987494379282, + 0.027691861614584923, + 0.023083873093128204, + 0.06158054992556572, + 0.029348015785217285, + 0.015596654266119003, + 0.04970158264040947, + 0.02941734343767166, + 0.14226572215557098, + 0.0033269375562667847, + 0.0038969137240201235, + 0.003414797130972147, + 0.0034083768259733915, + 0.0036165809724479914, + 0.009576898068189621, + 0.0025609107688069344, + 0.006945126224309206, + 0.019835595041513443, + 0.00894765555858612, + 0.03660633787512779, + 0.022696441039443016, + 0.015732545405626297, + 0.08657459169626236, + 0.014688375405967236, + 0.060436803847551346, + 0.002172283362597227, + 0.0015651953872293234, + 0.0010573597392067313, + 0.0012545938370749354, + 0.0025470436085015535, + 0.0012094287667423487, + 0.0012270724400877953, + 0.0006404969026334584, + 0.001989886397495866, + 0.0029019899666309357, + 0.003668847493827343, + 0.0025555819738656282, + 0.006013905163854361, + 0.014741209335625172, + 0.0020056937355548143, + 0.0014525922015309334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.029821116477251053 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022031813859939575 + ], + "opposite_logprob_tokens": [ + -6.127202987670898 + ], + "logprob_reward": -0.0022031813859939575, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022031813859939575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05624184384942055, + 0.04870179295539856, + 0.02964555285871029, + 0.020638922229409218, + 0.0234514307230711, + 0.11720211803913116, + 0.03016691468656063, + 0.6010059118270874, + 0.01631903648376465, + 0.05662643164396286 + ], + "credit_map": [ + 0.014670397154986858, + 0.006890591233968735, + 0.009316951967775822, + 0.0041571250185370445, + 0.005896694492548704, + 0.006770087406039238, + 0.004077231977134943, + 0.004462772980332375, + 0.01253015361726284, + 0.006819755304604769, + 0.007172332610934973, + 0.0038828777614980936, + 0.003491947427392006, + 0.006432975642383099, + 0.003391746897250414, + 0.004980005789548159, + 0.0047248476184904575, + 0.0029259598813951015, + 0.005820638034492731, + 0.0022713816724717617, + 0.0018684231908991933, + 0.0038184530567377806, + 0.002029080642387271, + 0.006186770740896463, + 0.0028560361824929714, + 0.0025459991302341223, + 0.0024101429153233767, + 0.0028151357546448708, + 0.003339295042678714, + 0.001844708458520472, + 0.002799360081553459, + 0.00202824710868299, + 0.002573562553152442, + 0.0018358371453359723, + 0.003402317874133587, + 0.0015651569701731205, + 0.0016985072288662195, + 0.003332684049382806, + 0.001693027326837182, + 0.007350340485572815, + 0.008794966153800488, + 0.007232860196381807, + 0.018948858603835106, + 0.009989522397518158, + 0.005032036453485489, + 0.015096889808773994, + 0.00938454270362854, + 0.04272245243191719, + 0.004922378808259964, + 0.003897417103871703, + 0.0026694415137171745, + 0.0038760134484618902, + 0.003973232116550207, + 0.0051316162571311, + 0.0024149627424776554, + 0.0032818554900586605, + 0.041904643177986145, + 0.019783513620495796, + 0.08269719034433365, + 0.054422780871391296, + 0.035422321408987045, + 0.20135213434696198, + 0.03456169366836548, + 0.13086168467998505, + 0.0026603408623486757, + 0.002384499879553914, + 0.001798387966118753, + 0.0014937303494662046, + 0.002381763653829694, + 0.002431596163660288, + 0.0019120281795039773, + 0.0012566916411742568, + 0.00858286302536726, + 0.006829836871474981, + 0.0076678963378071785, + 0.006301528308540583, + 0.00770057225599885, + 0.011455275118350983, + 0.005835882853716612, + 0.0022525822278112173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.0022031813859939575 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946207791566849 + ], + "opposite_logprob_tokens": [ + -5.130946159362793 + ], + "logprob_reward": -0.005946207791566849, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946207791566849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040800876915454865, + 0.029110215604305267, + 0.055278170853853226, + 0.030711453408002853, + 0.025809530168771744, + 0.3865390717983246, + 0.06111746281385422, + 0.26990336179733276, + 0.05652151256799698, + 0.04420839622616768 + ], + "credit_map": [ + 0.008600166067481041, + 0.0041528865694999695, + 0.00810791365802288, + 0.0028615284245461226, + 0.0035406271927058697, + 0.004506324417889118, + 0.0030866290908306837, + 0.005944800563156605, + 0.005889605265110731, + 0.0034347334876656532, + 0.006416354328393936, + 0.0025076570454984903, + 0.0020508854649960995, + 0.0041420310735702515, + 0.0016289653722196817, + 0.0030399810057133436, + 0.008113306947052479, + 0.004316316451877356, + 0.010364886373281479, + 0.003341781673952937, + 0.003985995892435312, + 0.0075560035184025764, + 0.003056055400520563, + 0.014543823897838593, + 0.0025389057118445635, + 0.0025312933139503, + 0.004849686753004789, + 0.002491166815161705, + 0.0024750197771936655, + 0.008955963887274265, + 0.002484987722709775, + 0.004384426865726709, + 0.002923615276813507, + 0.0022635208442807198, + 0.002863191766664386, + 0.0021415953524410725, + 0.002219593385234475, + 0.0035541292745620012, + 0.0019395818235352635, + 0.007904301397502422, + 0.025225112214684486, + 0.022924823686480522, + 0.0617683082818985, + 0.027070481330156326, + 0.01698768138885498, + 0.051395777612924576, + 0.03026178479194641, + 0.1509050875902176, + 0.006265468895435333, + 0.004708770662546158, + 0.005350435618311167, + 0.00567141268402338, + 0.005666470620781183, + 0.01933201774954796, + 0.004513644613325596, + 0.009609236381947994, + 0.020492812618613243, + 0.008952748961746693, + 0.0419476293027401, + 0.02429048717021942, + 0.015587757341563702, + 0.08396776765584946, + 0.014942112378776073, + 0.05972200259566307, + 0.0150438342243433, + 0.007463324349373579, + 0.005218110047280788, + 0.00397762656211853, + 0.004572882317006588, + 0.010870722122490406, + 0.005314660258591175, + 0.004060348495841026, + 0.007705903612077236, + 0.004275254439562559, + 0.006862002424895763, + 0.0037086126394569874, + 0.00375956273637712, + 0.01311157550662756, + 0.002629831898957491, + 0.002155647613108158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.005946207791566849 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735715549439192 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735715549439192, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735715549439192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13150064647197723, + 0.095481738448143, + 0.2878809869289398, + 0.011969777755439281, + 0.02134222723543644, + 0.06382434815168381, + 0.01478582713752985, + 0.30969321727752686, + 0.02187536656856537, + 0.041645828634500504 + ], + "credit_map": [ + 0.03446652740240097, + 0.01607566326856613, + 0.020563865080475807, + 0.011028293520212173, + 0.011639914475381374, + 0.015865512192249298, + 0.008099333383142948, + 0.013761540874838829, + 0.02318612113595009, + 0.01124546304345131, + 0.018744323402643204, + 0.008723771199584007, + 0.0071011134423315525, + 0.012145069427788258, + 0.006721572484821081, + 0.00761431036517024, + 0.05853993073105812, + 0.02483770065009594, + 0.05241809040307999, + 0.018259212374687195, + 0.019944490864872932, + 0.034102536737918854, + 0.01814436726272106, + 0.061634667217731476, + 0.00133738131262362, + 0.0024896946270018816, + 0.0014858414651826024, + 0.0012339625973254442, + 0.0014602456940338016, + 0.0013436762383207679, + 0.0011775007005780935, + 0.0014414760516956449, + 0.002824602648615837, + 0.0018482874147593975, + 0.0031483767088502645, + 0.0020705382339656353, + 0.0020081554539501667, + 0.002796694403514266, + 0.0021317796781659126, + 0.004513792227953672, + 0.004594976082444191, + 0.004843268543481827, + 0.011037774384021759, + 0.004943382926285267, + 0.003065473632887006, + 0.008062460459768772, + 0.0043448531068861485, + 0.022932160645723343, + 0.002120650839060545, + 0.002536814659833908, + 0.001175513374619186, + 0.0017128217732533813, + 0.0022472457494586706, + 0.0023544856812804937, + 0.0011289238464087248, + 0.0015093701658770442, + 0.0227097999304533, + 0.009684871882200241, + 0.05325401946902275, + 0.02732918970286846, + 0.017093220725655556, + 0.09089380502700806, + 0.016927102580666542, + 0.07180118560791016, + 0.004103422164916992, + 0.0033494294621050358, + 0.0021323650144040585, + 0.0020956341177225113, + 0.00412001321092248, + 0.002341180806979537, + 0.0022831601090729237, + 0.0014501629630103707, + 0.003591433633118868, + 0.0022723653819411993, + 0.006001133471727371, + 0.0037865210324525833, + 0.004850355442613363, + 0.017089540138840675, + 0.002392399124801159, + 0.001662081223912537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.006735715549439192 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot follows from the side and rear, the camera close to the ground, parallel to the galloping hooves. In the image, a muscular, inky-haired steed gallops at full speed across the vast, desolate Gobi Desert, its hooves flying and stirring up clouds of yellow dust. Aloft, a lone knight, leans forward, gripping the reins. His long, wind-twisted black hair, his bronzed face stern, his gaze piercing, like a hawk's, fixed on the distant horizon. He wears a heavy, rough black cape, which is violently lifted by the strong wind. It billows and tumbles behind him like an unfurled battle flag, forming a vast, powerful, and almost frozen black arc. On the horizon, a burning blood-red setting sun is slowly sinking under the endless sand dunes, dyeing the entire sky a strong orange-red and deep purple. The dazzling backlight outlines a circle of dazzling golden outline light for the knight and the horse. Their figures appear lonely and full of epic grandeur in the elongated shadows. The whole scene is filled with a desolate, resolute and unrestrained atmosphere in the high-speed movement.", + "seed": 7287, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_11/index_5730_seed_7287_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02397963833063841, + "vlm_logprob": -0.02397963833063841, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.02397963833063841, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04215281829237938 + ], + "opposite_logprob_tokens": [ + -5.5421528816223145 + ], + "logprob_reward": -0.04215281829237938, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.04215281829237938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025837302207946777, + 0.033809076994657516, + 0.04260224848985672, + 0.0248236246407032, + 0.7255806922912598, + 0.05068626627326012, + 0.022472664713859558, + 0.013706833124160767, + 0.01801670715212822, + 0.04246462509036064 + ], + "credit_map": [ + 0.002772923558950424, + 0.003432578407227993, + 0.005273107439279556, + 0.0036239842884242535, + 0.003106582909822464, + 0.0029721674509346485, + 0.0025029319804161787, + 0.002153026172891259, + 0.004378254525363445, + 0.003302230266854167, + 0.006368533242493868, + 0.005611023865640163, + 0.004716136958450079, + 0.0033732480369508266, + 0.0035212701186537743, + 0.002538380678743124, + 0.005181109067052603, + 0.002961438847705722, + 0.009830022230744362, + 0.012107769958674908, + 0.002425574464723468, + 0.0022224956192076206, + 0.005364534445106983, + 0.002509303158149123, + 0.0027559869922697544, + 0.0025926753878593445, + 0.0032815143931657076, + 0.0026556607335805893, + 0.00327954045496881, + 0.0035084430128335953, + 0.003520597703754902, + 0.0032292059622704983, + 0.29811492562294006, + 0.023996686562895775, + 0.31168654561042786, + 0.03979092463850975, + 0.00927642360329628, + 0.014861668460071087, + 0.018010836094617844, + 0.009842624887824059, + 0.015582023188471794, + 0.0034593171440064907, + 0.01601339317858219, + 0.00275106867775321, + 0.004227956756949425, + 0.0028490652330219746, + 0.0036319550126791, + 0.0021714866161346436, + 0.004089814145117998, + 0.001849680789746344, + 0.0033916751854121685, + 0.0021523532923310995, + 0.002768943551927805, + 0.0031826598569750786, + 0.0027197820600122213, + 0.0023177540861070156, + 0.0017358181066811085, + 0.0014721545157954097, + 0.0015239042695611715, + 0.0012256029294803739, + 0.0019151372835040092, + 0.0020905870478600264, + 0.002033399883657694, + 0.0017102292040362954, + 0.0018475287361070514, + 0.0021462845616042614, + 0.001188674126751721, + 0.00253084278665483, + 0.0020946243312209845, + 0.0029139474499970675, + 0.002794868778437376, + 0.0024999375455081463, + 0.006526473443955183, + 0.00362799852155149, + 0.006274090614169836, + 0.007349720224738121, + 0.004753684159368277, + 0.005062521900981665, + 0.006127950269728899, + 0.0027421817649155855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.04215281829237938 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868850111961365 + ], + "opposite_logprob_tokens": [ + -3.0486884117126465 + ], + "logprob_reward": -0.04868850111961365, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868850111961365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018378539010882378, + 0.007703177630901337, + 0.010275323875248432, + 0.014446676708757877, + 0.9028791785240173, + 0.008496662601828575, + 0.008439903147518635, + 0.00754911731928587, + 0.009355912916362286, + 0.012475461699068546 + ], + "credit_map": [ + 0.002455387031659484, + 0.001390959369018674, + 0.004850888624787331, + 0.0032276699785143137, + 0.0019147080602124333, + 0.001930509926751256, + 0.0012719265650957823, + 0.0013364902697503567, + 0.0010101990774273872, + 0.0009526126668788493, + 0.0012687152484431863, + 0.001034819521009922, + 0.0009229527204297483, + 0.0010219089454039931, + 0.0008684861822985113, + 0.0006234834436327219, + 0.001394059625454247, + 0.0008492431952618062, + 0.002062189159914851, + 0.0016355601837858558, + 0.0012471493100747466, + 0.0010912209982052445, + 0.0012057824060320854, + 0.0007901207427494228, + 0.0017027799040079117, + 0.0015547347720712423, + 0.0027734339237213135, + 0.0019061126513406634, + 0.0017579307314008474, + 0.0020847569685429335, + 0.0015658045886084437, + 0.0011011239839717746, + 0.372542142868042, + 0.029796438291668892, + 0.3911844491958618, + 0.04988899454474449, + 0.00995159987360239, + 0.01630634069442749, + 0.02183581329882145, + 0.011373447254300117, + 0.002363047795370221, + 0.0006410813657566905, + 0.0025234611239284277, + 0.0005021881661377847, + 0.0006448911735787988, + 0.0004826728836633265, + 0.0008119650883600116, + 0.0005273559363558888, + 0.001155221019871533, + 0.0008219401352107525, + 0.001163395238108933, + 0.0011217965511605144, + 0.0009922581957653165, + 0.0014566133031621575, + 0.0009896931005641818, + 0.0007389870588667691, + 0.0008626346825622022, + 0.0009828932816162705, + 0.0009220482897944748, + 0.0007936669862829149, + 0.0011993949301540852, + 0.0008400079677812755, + 0.001370899728499353, + 0.0005775718600489199, + 0.0009669663850218058, + 0.0010180992539972067, + 0.0005849056760780513, + 0.0013864050852134824, + 0.0011735057923942804, + 0.0020001372322440147, + 0.0012253978056833148, + 0.0010004965588450432, + 0.0016462925123050809, + 0.0012175211450085044, + 0.0006199544295668602, + 0.002852682489901781, + 0.0012481213780120015, + 0.0016899928450584412, + 0.0018755275523290038, + 0.0013253700453788042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.04868850111961365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674376729875803 + ], + "opposite_logprob_tokens": [ + -5.006743907928467 + ], + "logprob_reward": -0.00674376729875803, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674376729875803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01630694977939129, + 0.008249787613749504, + 0.01886579394340515, + 0.012780425138771534, + 0.8646292090415955, + 0.010789435356855392, + 0.010112742893397808, + 0.012596847489476204, + 0.010285364463925362, + 0.035383518785238266 + ], + "credit_map": [ + 0.0020984055008739233, + 0.002558934036642313, + 0.002921526553109288, + 0.0023434367030858994, + 0.0018928414210677147, + 0.0019758050329983234, + 0.0014679068699479103, + 0.0010480909841135144, + 0.0010020920308306813, + 0.0009576058946549892, + 0.0011204725597053766, + 0.0012905820040032268, + 0.0009850357891991735, + 0.0009731941972859204, + 0.0012088485527783632, + 0.0007119557703845203, + 0.002319456310942769, + 0.0018065753392875195, + 0.0025946928653866053, + 0.0026252111420035362, + 0.002650246489793062, + 0.0026070827152580023, + 0.0021163003984838724, + 0.002146227052435279, + 0.0015307290013879538, + 0.001490313559770584, + 0.002373793860897422, + 0.0014656237326562405, + 0.0015636613825336099, + 0.001851128414273262, + 0.0016792597016319633, + 0.0008259144378826022, + 0.356356680393219, + 0.028709929436445236, + 0.37477123737335205, + 0.04771900549530983, + 0.009599021635949612, + 0.015632882714271545, + 0.02092125080525875, + 0.010919075459241867, + 0.0026218031998723745, + 0.0010316711850464344, + 0.002743606921285391, + 0.0006220184150151908, + 0.0012093927944079041, + 0.0008118218393065035, + 0.001085277064703405, + 0.0006638421909883618, + 0.0012160609476268291, + 0.0009897552663460374, + 0.0009754985803738236, + 0.001094170263968408, + 0.001371930819004774, + 0.0019440921023488045, + 0.0013791448436677456, + 0.0011420890223234892, + 0.0014940932160243392, + 0.0014152353396639228, + 0.0013432803098112345, + 0.0011840802617371082, + 0.0025597757194191217, + 0.0015950833912938833, + 0.0016885031946003437, + 0.0013167945435270667, + 0.0010688727488741279, + 0.001078232773579657, + 0.0009293505572713912, + 0.0013948966516181827, + 0.0014555329689756036, + 0.001517216325737536, + 0.0017502313712611794, + 0.001091029611416161, + 0.004888544790446758, + 0.00366957182995975, + 0.0016169754089787602, + 0.008741464465856552, + 0.0030758425127714872, + 0.004296010825783014, + 0.005154311191290617, + 0.003940793685615063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.00674376729875803 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533940374851227 + ], + "opposite_logprob_tokens": [ + -4.387534141540527 + ], + "logprob_reward": -0.012533940374851227, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533940374851227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.017011268064379692, + 0.018018512055277824, + 0.12025705724954605, + 0.014414493925869465, + 0.6975629329681396, + 0.06637274473905563, + 0.006899607367813587, + 0.03000793047249317, + 0.007570786867290735, + 0.02188471332192421 + ], + "credit_map": [ + 0.0024048928171396255, + 0.0015351721085608006, + 0.004810229409486055, + 0.0030415025539696217, + 0.0014699694002047181, + 0.001628316007554531, + 0.0011472783517092466, + 0.0009739057859405875, + 0.0020625153556466103, + 0.0016338967252522707, + 0.004791309591382742, + 0.0028621782548725605, + 0.002213896019384265, + 0.0018681964138522744, + 0.0016755135729908943, + 0.0009110058308579028, + 0.0156463123857975, + 0.007995739579200745, + 0.026853462681174278, + 0.0356857068836689, + 0.006608069874346256, + 0.004157871473580599, + 0.01632705144584179, + 0.006982841063290834, + 0.001845821039751172, + 0.0009213376906700432, + 0.0033217400778084993, + 0.0023103165440261364, + 0.0017592719523236156, + 0.0016095421742647886, + 0.0015843099681660533, + 0.0010621529072523117, + 0.28813493251800537, + 0.023120053112506866, + 0.30190742015838623, + 0.03806619718670845, + 0.008064420893788338, + 0.012773532420396805, + 0.016961658373475075, + 0.00853467546403408, + 0.023684581741690636, + 0.0030609285458922386, + 0.024493787437677383, + 0.003511385526508093, + 0.002354310592636466, + 0.002038274658843875, + 0.004570953547954559, + 0.002658520359545946, + 0.0009856300894171, + 0.0009329900494776666, + 0.001180010149255395, + 0.0007884941296651959, + 0.000739636889193207, + 0.0006424550083465874, + 0.0008817479247227311, + 0.0007486421382054687, + 0.004556071944534779, + 0.0034406506456434727, + 0.005806394387036562, + 0.003064166521653533, + 0.0032287733629345894, + 0.0024958259891718626, + 0.0035420870408415794, + 0.003873958718031645, + 0.0010899048065766692, + 0.0005869721062481403, + 0.0011183741735294461, + 0.0008982755825854838, + 0.0009078513830900192, + 0.0009485172922722995, + 0.0009797142120078206, + 0.0010411770781502128, + 0.002952845534309745, + 0.002170798135921359, + 0.0021816107910126448, + 0.004833947401493788, + 0.001842753728851676, + 0.0024697519838809967, + 0.003013767534866929, + 0.002419237047433853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.012533940374851227 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977916456758976 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.00977916456758976, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977916456758976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.015274989418685436, + 0.02791288122534752, + 0.016827721148729324, + 0.005440521985292435, + 0.8861441612243652, + 0.012090140953660011, + 0.011674652807414532, + 0.008000115863978863, + 0.00781716126948595, + 0.008817611262202263 + ], + "credit_map": [ + 0.0016970111755654216, + 0.0011896612122654915, + 0.004100816324353218, + 0.002365182852372527, + 0.0013070175191387534, + 0.0018027874175459146, + 0.0014037523651495576, + 0.0014087610179558396, + 0.004140238277614117, + 0.003608039114624262, + 0.0065690274350345135, + 0.004402962047606707, + 0.003034876426681876, + 0.0020908229053020477, + 0.002617672784253955, + 0.0014492444461211562, + 0.002265544142574072, + 0.0012077799765393138, + 0.003665853524580598, + 0.0044761281460523605, + 0.001010754844173789, + 0.0008730204426683486, + 0.002183871576562524, + 0.0011447712313383818, + 0.0007019333424977958, + 0.0004575324128381908, + 0.001038969843648374, + 0.0007837091688998044, + 0.000646954751573503, + 0.0006056989659555256, + 0.0006555911968462169, + 0.0005501327686943114, + 0.3656734824180603, + 0.029354657977819443, + 0.38412943482398987, + 0.04877547547221184, + 0.01007566973567009, + 0.015934821218252182, + 0.021419141441583633, + 0.01078152097761631, + 0.004208052530884743, + 0.0007732543163001537, + 0.00407673604786396, + 0.000584593100938946, + 0.0007696634856984019, + 0.0004681353457272053, + 0.0007736807456240058, + 0.0004360258753877133, + 0.001957823522388935, + 0.0010145388077944517, + 0.0015904196770861745, + 0.001191217452287674, + 0.0014425440458580852, + 0.0017935192445293069, + 0.0014028113801032305, + 0.001281779957935214, + 0.0009583382052369416, + 0.000938095327001065, + 0.0010601032990962267, + 0.0008052699267864227, + 0.0016985316760838032, + 0.0009659818606451154, + 0.0009382385178469121, + 0.0006355578079819679, + 0.0008277135784737766, + 0.0008816851768642664, + 0.0005677868612110615, + 0.0011926417937502265, + 0.0009812724310904741, + 0.001328562037087977, + 0.0010270426282659173, + 0.0010104571701958776, + 0.0010084754321724176, + 0.0009970760438591242, + 0.0013529504649341106, + 0.00120990548748523, + 0.0012885344913229346, + 0.0013168227160349488, + 0.0009052833775058389, + 0.000738563307095319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.00977916456758976 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot follows from the side and rear, the camera close to the ground, parallel to the galloping hooves. In the image, a muscular, inky-haired steed gallops at full speed across the vast, desolate Gobi Desert, its hooves flying and stirring up clouds of yellow dust. Aloft, a lone knight, leans forward, gripping the reins. His long, wind-twisted black hair, his bronzed face stern, his gaze piercing, like a hawk's, fixed on the distant horizon. He wears a heavy, rough black cape, which is violently lifted by the strong wind. It billows and tumbles behind him like an unfurled battle flag, forming a vast, powerful, and almost frozen black arc. On the horizon, a burning blood-red setting sun is slowly sinking under the endless sand dunes, dyeing the entire sky a strong orange-red and deep purple. The dazzling backlight outlines a circle of dazzling golden outline light for the knight and the horse. Their figures appear lonely and full of epic grandeur in the elongated shadows. The whole scene is filled with a desolate, resolute and unrestrained atmosphere in the high-speed movement.", + "seed": 7288, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_11/index_5730_seed_7288_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.030970273097045718, + "vlm_logprob": -0.030970273097045718, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.030970273097045718, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044642239809036255 + ], + "opposite_logprob_tokens": [ + -4.544642448425293 + ], + "logprob_reward": -0.044642239809036255, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.044642239809036255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11197817325592041, + 0.07052353024482727, + 0.5672928690910339, + 0.023598458617925644, + 0.021648835390806198, + 0.022052092477679253, + 0.015772299841046333, + 0.05156213045120239, + 0.05373721942305565, + 0.06183437630534172 + ], + "credit_map": [ + 0.012047347612679005, + 0.02053806371986866, + 0.007848423905670643, + 0.007689820136874914, + 0.01868063397705555, + 0.022211933508515358, + 0.011693045496940613, + 0.01126890629529953, + 0.006798676215112209, + 0.01428893767297268, + 0.007017424330115318, + 0.007449843920767307, + 0.009749911725521088, + 0.008233612403273582, + 0.0069413138553500175, + 0.010043810121715069, + 0.026500115171074867, + 0.015608202666044235, + 0.21154874563217163, + 0.022321103140711784, + 0.01050043385475874, + 0.03510848060250282, + 0.01638408936560154, + 0.22932171821594238, + 0.003029372077435255, + 0.0032807637471705675, + 0.003004908561706543, + 0.002442095195874572, + 0.0032032541930675507, + 0.002864493988454342, + 0.002764992183074355, + 0.0030085754115134478, + 0.0032809386029839516, + 0.0035650222562253475, + 0.002510262653231621, + 0.00249077077023685, + 0.0025979436468333006, + 0.0019150947919115424, + 0.00337653118185699, + 0.0019122726516798139, + 0.0028188328724354506, + 0.0030981749296188354, + 0.002862873487174511, + 0.001753671676851809, + 0.003684746101498604, + 0.003982852213084698, + 0.0015828442992642522, + 0.0022680945694446564, + 0.0020495427306741476, + 0.002620173152536154, + 0.0016497892793267965, + 0.0018978987354785204, + 0.0024763713590800762, + 0.0017886250279843807, + 0.002066134475171566, + 0.0012237651972100139, + 0.006021469831466675, + 0.007127627730369568, + 0.0041980319656431675, + 0.0039009808097034693, + 0.0066604032181203365, + 0.00803346186876297, + 0.006254923529922962, + 0.009365231730043888, + 0.006240523885935545, + 0.005995734594762325, + 0.004350591916590929, + 0.003876117756590247, + 0.010383148677647114, + 0.00793280266225338, + 0.00697806803509593, + 0.007980232127010822, + 0.01256509032100439, + 0.011791108176112175, + 0.0032835686579346657, + 0.006823380012065172, + 0.008833784610033035, + 0.006568967830389738, + 0.006245800293982029, + 0.0057226745411753654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.044642239809036255 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034364461898804 + ], + "opposite_logprob_tokens": [ + -2.350343704223633 + ], + "logprob_reward": -0.10034364461898804, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034364461898804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040033843368291855, + 0.03260247781872749, + 0.863696813583374, + 0.012571324594318867, + 0.008272739127278328, + 0.006486067082732916, + 0.0035583674907684326, + 0.013056422583758831, + 0.007389408070594072, + 0.01233251579105854 + ], + "credit_map": [ + 0.0034692767076194286, + 0.007290789391845465, + 0.0023577495012432337, + 0.002822438720613718, + 0.009898570366203785, + 0.004412015434354544, + 0.003156270133331418, + 0.006626734510064125, + 0.00332439667545259, + 0.004749509505927563, + 0.0032910939771682024, + 0.003799262922257185, + 0.00467985775321722, + 0.0037009886000305414, + 0.003528823610395193, + 0.005528549198061228, + 0.04124747961759567, + 0.024877062067389488, + 0.3229069411754608, + 0.03411231189966202, + 0.01632358320057392, + 0.051639702171087265, + 0.024984167888760567, + 0.34760555624961853, + 0.0015095856506377459, + 0.0020747689995914698, + 0.0016194108175113797, + 0.0020269013475626707, + 0.0015724023105576634, + 0.0014410102739930153, + 0.0008055048529058695, + 0.0015217417385429144, + 0.001359218847937882, + 0.001176251913420856, + 0.0011708897072821856, + 0.00108372385147959, + 0.0008342785877175629, + 0.0008281022310256958, + 0.0010481026256456971, + 0.0007721719448454678, + 0.0007324414909817278, + 0.0008573266677558422, + 0.0006470880471169949, + 0.0006504195043817163, + 0.0010165376588702202, + 0.0012908597709611058, + 0.00051604222971946, + 0.0007753520039841533, + 0.00037798023549839854, + 0.0005238934536464512, + 0.0004230188496876508, + 0.00046244883560575545, + 0.0005074077635072172, + 0.0003908203507307917, + 0.0004902592045255005, + 0.0003825391177088022, + 0.0020274161361157894, + 0.002370797796174884, + 0.001201825449243188, + 0.0011011784663423896, + 0.001647474360652268, + 0.0016823172336444259, + 0.0012698001228272915, + 0.0017556139500811696, + 0.0013028179528191686, + 0.001329389400780201, + 0.0006657788762822747, + 0.0005504544242285192, + 0.0008802222437225282, + 0.0010019001783803105, + 0.0007030382985249162, + 0.0009558071033097804, + 0.0030146914068609476, + 0.002688771113753319, + 0.0006383393774740398, + 0.0011922658886760473, + 0.0014206948690116405, + 0.0008784791571088135, + 0.0011845651315525174, + 0.0013147094286978245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.10034364461898804 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017206162447109818 + ], + "opposite_logprob_tokens": [ + -6.376720428466797 + ], + "logprob_reward": -0.0017206162447109818, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017206162447109818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17612099647521973, + 0.06889282166957855, + 0.331731915473938, + 0.07323787361383438, + 0.061887357383966446, + 0.1038278117775917, + 0.0347088985145092, + 0.011154977604746819, + 0.06096624583005905, + 0.07747111469507217 + ], + "credit_map": [ + 0.016337085515260696, + 0.024194642901420593, + 0.013170333579182625, + 0.010240541771054268, + 0.03277463838458061, + 0.03335653617978096, + 0.01623203232884407, + 0.029815178364515305, + 0.005460207350552082, + 0.010917901992797852, + 0.008121947757899761, + 0.007648571860045195, + 0.009762582369148731, + 0.007113687228411436, + 0.01083922665566206, + 0.009028692729771137, + 0.014808734878897667, + 0.009393458254635334, + 0.11977201700210571, + 0.014121006242930889, + 0.007706683594733477, + 0.020621072500944138, + 0.013062987476587296, + 0.13224594295024872, + 0.010444419458508492, + 0.016036082059144974, + 0.008370385505259037, + 0.0069532860070466995, + 0.013994678854942322, + 0.006756884977221489, + 0.006465221755206585, + 0.004216914530843496, + 0.00790364295244217, + 0.013463607057929039, + 0.00586159760132432, + 0.0054225437343120575, + 0.008534833788871765, + 0.007858764380216599, + 0.007360632997006178, + 0.005481739062815905, + 0.011436219327151775, + 0.01491597294807434, + 0.010935969650745392, + 0.008048326708376408, + 0.026337865740060806, + 0.017343752086162567, + 0.0074480557814240456, + 0.007361644878983498, + 0.003991888370364904, + 0.004355208948254585, + 0.003706594929099083, + 0.003282806370407343, + 0.006927365902811289, + 0.0038574838545173407, + 0.005377898458391428, + 0.0032096512150019407, + 0.0014407120179384947, + 0.0017269467934966087, + 0.0014746629167348146, + 0.001773809315636754, + 0.0013907833490520716, + 0.0012247302802279592, + 0.0008934178622439504, + 0.0012299155350774527, + 0.006662655156105757, + 0.007068029139190912, + 0.006018329877406359, + 0.004255767911672592, + 0.013565964996814728, + 0.008155216462910175, + 0.007330733817070723, + 0.00790955126285553, + 0.00920757744461298, + 0.007729575969278812, + 0.004949862137436867, + 0.008095853962004185, + 0.012018481269478798, + 0.011578870937228203, + 0.011570216156542301, + 0.01232067495584488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.0017206162447109818 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021974719129502773 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.0021974719129502773, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021974719129502773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13913944363594055, + 0.050124045461416245, + 0.5649211406707764, + 0.04853631556034088, + 0.04488661512732506, + 0.021540427580475807, + 0.013384856283664703, + 0.025124846026301384, + 0.05243512988090515, + 0.03990725427865982 + ], + "credit_map": [ + 0.009291056543588638, + 0.02803775481879711, + 0.00871021393686533, + 0.008321852423250675, + 0.04085564985871315, + 0.010000786744058132, + 0.012975280173122883, + 0.020946847274899483, + 0.005637194029986858, + 0.00740369688719511, + 0.005427850876003504, + 0.005486693233251572, + 0.006078130099922419, + 0.00544702960178256, + 0.005587080027908087, + 0.00905637163668871, + 0.02639716863632202, + 0.01613047905266285, + 0.211471825838089, + 0.02212747186422348, + 0.010247697122395039, + 0.03398759663105011, + 0.01663251779973507, + 0.2279263585805893, + 0.006734482944011688, + 0.00769220432266593, + 0.005624001380056143, + 0.006371745374053717, + 0.0063906642608344555, + 0.00722457142546773, + 0.0037926803342998028, + 0.0047059678472578526, + 0.007410273887217045, + 0.005000145640224218, + 0.007335291244089603, + 0.006001900881528854, + 0.0052900006994605064, + 0.004135467577725649, + 0.006458322051912546, + 0.00325521151535213, + 0.00238420139066875, + 0.002560229040682316, + 0.0027363731060177088, + 0.0030352931935340166, + 0.003806688357144594, + 0.002783433301374316, + 0.0019430248066782951, + 0.0022911843843758106, + 0.0016479415353387594, + 0.0018543042242527008, + 0.0015067342901602387, + 0.001849559834226966, + 0.001948672579601407, + 0.0013875700533390045, + 0.002056931611150503, + 0.0011331429705023766, + 0.0032939736265689135, + 0.003463346278294921, + 0.00443409476429224, + 0.006367450580000877, + 0.0020999987609684467, + 0.0018235702300444245, + 0.0015300717204809189, + 0.0021123397164046764, + 0.006557128857821226, + 0.006746605504304171, + 0.004923234228044748, + 0.005127228330820799, + 0.008414718322455883, + 0.007487019523978233, + 0.0059366170316934586, + 0.007242576684802771, + 0.005241605918854475, + 0.005433790851384401, + 0.0057160211727023125, + 0.006405020132660866, + 0.0041441647335886955, + 0.0037412166129797697, + 0.005081350449472666, + 0.004144084174185991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.0021974719129502773 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947392899543047 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947392899543047, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947392899543047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14401079714298248, + 0.15123726427555084, + 0.20239408314228058, + 0.12847381830215454, + 0.05250788480043411, + 0.08923184871673584, + 0.03960369899868965, + 0.0563790425658226, + 0.0628712922334671, + 0.07329024374485016 + ], + "credit_map": [ + 0.020409181714057922, + 0.025398824363946915, + 0.012594754807651043, + 0.018972767516970634, + 0.019311003386974335, + 0.019075026735663414, + 0.01377288717776537, + 0.014476357027888298, + 0.01690050959587097, + 0.025257354602217674, + 0.013639960438013077, + 0.015121553093194962, + 0.022543607279658318, + 0.020759308710694313, + 0.017859583720564842, + 0.019155383110046387, + 0.01294292975217104, + 0.007889118045568466, + 0.07019659876823425, + 0.009940766729414463, + 0.007608470041304827, + 0.012201255187392235, + 0.008765988983213902, + 0.07284896820783615, + 0.018551187589764595, + 0.019367652013897896, + 0.01773749478161335, + 0.014988898299634457, + 0.017363550141453743, + 0.013291668146848679, + 0.0107261436060071, + 0.01644723117351532, + 0.00619898084551096, + 0.007966743782162666, + 0.006381451152265072, + 0.006075181066989899, + 0.006575418636202812, + 0.006707415450364351, + 0.008449680171906948, + 0.004153018351644278, + 0.010057291947305202, + 0.011798740364611149, + 0.009830287657678127, + 0.010462566278874874, + 0.013392145745456219, + 0.012459873221814632, + 0.011452149599790573, + 0.009778805077075958, + 0.004847052041441202, + 0.006247404031455517, + 0.004309903830289841, + 0.004329947289079428, + 0.006890824995934963, + 0.005129165481775999, + 0.004493299871683121, + 0.003356103552505374, + 0.011509926058351994, + 0.014627879485487938, + 0.0063324240036308765, + 0.005743931978940964, + 0.005405851639807224, + 0.004764437675476074, + 0.0035984499845653772, + 0.00439614849165082, + 0.007348931860178709, + 0.008864019997417927, + 0.004525436088442802, + 0.00520591763779521, + 0.013539325445890427, + 0.007862279191613197, + 0.008157785050570965, + 0.007367596495896578, + 0.01643330603837967, + 0.014998180791735649, + 0.00450148805975914, + 0.010836554691195488, + 0.006762795615941286, + 0.005626835860311985, + 0.007612149231135845, + 0.00651894323527813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.005947392899543047 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot follows from the side and rear, the camera close to the ground, parallel to the galloping hooves. In the image, a muscular, inky-haired steed gallops at full speed across the vast, desolate Gobi Desert, its hooves flying and stirring up clouds of yellow dust. Aloft, a lone knight, leans forward, gripping the reins. His long, wind-twisted black hair, his bronzed face stern, his gaze piercing, like a hawk's, fixed on the distant horizon. He wears a heavy, rough black cape, which is violently lifted by the strong wind. It billows and tumbles behind him like an unfurled battle flag, forming a vast, powerful, and almost frozen black arc. On the horizon, a burning blood-red setting sun is slowly sinking under the endless sand dunes, dyeing the entire sky a strong orange-red and deep purple. The dazzling backlight outlines a circle of dazzling golden outline light for the knight and the horse. Their figures appear lonely and full of epic grandeur in the elongated shadows. The whole scene is filled with a desolate, resolute and unrestrained atmosphere in the high-speed movement.", + "seed": 7289, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_11/index_5730_seed_7289_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.023601867130491883, + "vlm_logprob": -0.023601867130491883, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.023601867130491883, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03569883853197098 + ], + "opposite_logprob_tokens": [ + -5.160698890686035 + ], + "logprob_reward": -0.03569883853197098, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.03569883853197098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13462045788764954, + 0.09013385325670242, + 0.07723551243543625, + 0.2170935571193695, + 0.05223413556814194, + 0.046571437269449234, + 0.14263218641281128, + 0.07443168014287949, + 0.10379094630479813, + 0.061256177723407745 + ], + "credit_map": [ + 0.014571788720786572, + 0.020866455510258675, + 0.013030284084379673, + 0.019833888858556747, + 0.019618947058916092, + 0.014846010133624077, + 0.017910923808813095, + 0.013942191377282143, + 0.009197290986776352, + 0.010892380028963089, + 0.008854759857058525, + 0.02075013518333435, + 0.011250528506934643, + 0.011694679036736488, + 0.011053879745304585, + 0.006440202239900827, + 0.01672487147152424, + 0.013580339960753918, + 0.013349772430956364, + 0.00815062876790762, + 0.009067939594388008, + 0.005927942227572203, + 0.0050184717401862144, + 0.005415554624050856, + 0.05013630539178848, + 0.02486160211265087, + 0.00933157466351986, + 0.012605314143002033, + 0.02481950633227825, + 0.06989943981170654, + 0.013716418296098709, + 0.011723428033292294, + 0.007597911637276411, + 0.0046714493073523045, + 0.007071164436638355, + 0.005648571066558361, + 0.007397886831313372, + 0.008604374714195728, + 0.005699885077774525, + 0.00554289948195219, + 0.004719526041299105, + 0.005506978370249271, + 0.005450354423373938, + 0.007295089308172464, + 0.006558049935847521, + 0.006732314359396696, + 0.0047701625153422356, + 0.005538965575397015, + 0.018373845145106316, + 0.014179836958646774, + 0.02042175829410553, + 0.0216828603297472, + 0.021066315472126007, + 0.011081941425800323, + 0.02423357404768467, + 0.011592068709433079, + 0.011126605793833733, + 0.015966232866048813, + 0.014554500579833984, + 0.007292639929801226, + 0.008800951763987541, + 0.006616696249693632, + 0.005018347408622503, + 0.0050557148642838, + 0.018151378259062767, + 0.017042549327015877, + 0.012229515239596367, + 0.01844857446849346, + 0.013247308321297169, + 0.00867679063230753, + 0.009731912054121494, + 0.006262931041419506, + 0.009181184694170952, + 0.010751301422715187, + 0.008874889463186264, + 0.006860070396214724, + 0.007912103086709976, + 0.0077636949717998505, + 0.005068608559668064, + 0.0048443288542330265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.03569883853197098 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004215568304062 + ], + "opposite_logprob_tokens": [ + -2.695042133331299 + ], + "logprob_reward": -0.07004215568304062, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004215568304062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12285314500331879, + 0.0823882669210434, + 0.07009679079055786, + 0.24653157591819763, + 0.04172990098595619, + 0.05044049397110939, + 0.06989078968763351, + 0.22575213015079498, + 0.029185673221945763, + 0.061131201684474945 + ], + "credit_map": [ + 0.017532924190163612, + 0.014057009480893612, + 0.015689948573708534, + 0.01856183260679245, + 0.01757214404642582, + 0.012623043730854988, + 0.01492652203887701, + 0.011889714747667313, + 0.008346326649188995, + 0.012034794315695763, + 0.006836302112787962, + 0.013241981156170368, + 0.012194517999887466, + 0.013945006765425205, + 0.008832150138914585, + 0.006957190111279488, + 0.01426350325345993, + 0.009913587011396885, + 0.010099709033966064, + 0.0071832905523478985, + 0.008682191371917725, + 0.005802763160318136, + 0.009676237590610981, + 0.004475507419556379, + 0.0651080459356308, + 0.01977054588496685, + 0.010041294619441032, + 0.009282024577260017, + 0.022223662585020065, + 0.09821004420518875, + 0.010511031374335289, + 0.011384946294128895, + 0.005909794010221958, + 0.003984450362622738, + 0.004583599511533976, + 0.0046207779087126255, + 0.007307893130928278, + 0.006241242401301861, + 0.004724463447928429, + 0.004357679281383753, + 0.0075064487755298615, + 0.00712224654853344, + 0.006124555598944426, + 0.00489222165197134, + 0.0069990758784115314, + 0.007318491116166115, + 0.004918443970382214, + 0.005559011362493038, + 0.008904625661671162, + 0.0070372289046645164, + 0.009269271045923233, + 0.010631827637553215, + 0.010316144675016403, + 0.006060576532036066, + 0.011573387309908867, + 0.006097727920860052, + 0.03077925741672516, + 0.05268235132098198, + 0.037427689880132675, + 0.023815693333745003, + 0.025569863617420197, + 0.022856738418340683, + 0.017525993287563324, + 0.015094549395143986, + 0.0042899418622255325, + 0.004398714751005173, + 0.003687505144625902, + 0.0030361665412783623, + 0.004606436472386122, + 0.004245355259627104, + 0.002543960465118289, + 0.002377592260017991, + 0.011786743998527527, + 0.007969353348016739, + 0.015201260335743427, + 0.007989485748112202, + 0.003595688845962286, + 0.004214232321828604, + 0.005219736602157354, + 0.00515469815582037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.07004215568304062 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008167268824763596 + ], + "opposite_logprob_tokens": [ + -7.125816822052002 + ], + "logprob_reward": -0.0008167268824763596, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008167268824763596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18110303580760956, + 0.07114867866039276, + 0.09112544357776642, + 0.19274252653121948, + 0.11130238324403763, + 0.07774653285741806, + 0.10211498290300369, + 0.05147144943475723, + 0.03710056096315384, + 0.08414439111948013 + ], + "credit_map": [ + 0.01770995557308197, + 0.0228551235049963, + 0.018753279000520706, + 0.025391370058059692, + 0.0244579054415226, + 0.023279691115021706, + 0.02908673882484436, + 0.01956898905336857, + 0.007149616722017527, + 0.00972369872033596, + 0.006066828966140747, + 0.011896650306880474, + 0.009100476279854774, + 0.010792061686515808, + 0.008613159880042076, + 0.007806187029927969, + 0.018200797960162163, + 0.010631320998072624, + 0.013188126496970654, + 0.010297282598912716, + 0.01235603541135788, + 0.008925197646021843, + 0.012312895618379116, + 0.005213787779211998, + 0.05143006891012192, + 0.014442456886172295, + 0.007536410819739103, + 0.007272293791174889, + 0.01712604984641075, + 0.07848607748746872, + 0.00815331470221281, + 0.008295849896967411, + 0.014623535796999931, + 0.012858813628554344, + 0.010735830292105675, + 0.01396891102194786, + 0.01967252790927887, + 0.016451673582196236, + 0.014209052547812462, + 0.0087820366024971, + 0.009919530712068081, + 0.00855311006307602, + 0.00942626316100359, + 0.007773967459797859, + 0.01265106163918972, + 0.011878890916705132, + 0.008687176741659641, + 0.008856539614498615, + 0.023694155737757683, + 0.009380728006362915, + 0.02897220477461815, + 0.006890864577144384, + 0.010045859962701797, + 0.007319154217839241, + 0.009298142977058887, + 0.006513871718198061, + 0.006994437891989946, + 0.01067427359521389, + 0.005755267571657896, + 0.00933067686855793, + 0.004736149217933416, + 0.004647762048989534, + 0.0050536757335066795, + 0.004279201850295067, + 0.006012617144733667, + 0.0052802604623138905, + 0.003987426869571209, + 0.00626729940995574, + 0.006652269512414932, + 0.003319455310702324, + 0.0030474411323666573, + 0.002533789025619626, + 0.013115154579281807, + 0.009382500313222408, + 0.013098536990582943, + 0.009996340610086918, + 0.00997824128717184, + 0.010643245652318, + 0.009884054772555828, + 0.008046322502195835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.0008167268824763596 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028177103959023952 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028177103959023952, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028177103959023952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12185750901699066, + 0.051338497549295425, + 0.10585016757249832, + 0.2326277643442154, + 0.03992180898785591, + 0.06781076639890671, + 0.1475261002779007, + 0.12095975130796432, + 0.07954935729503632, + 0.03255834802985191 + ], + "credit_map": [ + 0.01650748774409294, + 0.01402200199663639, + 0.014066524803638458, + 0.017743073403835297, + 0.014128532260656357, + 0.014789186418056488, + 0.016862209886312485, + 0.013738474808633327, + 0.006720145232975483, + 0.005831277463585138, + 0.004876023158431053, + 0.01126986462622881, + 0.005982987117022276, + 0.005797307472676039, + 0.005554418545216322, + 0.005306468345224857, + 0.0279266145080328, + 0.011953318491578102, + 0.022769039496779442, + 0.018491169437766075, + 0.008030478842556477, + 0.005326241720467806, + 0.007408545818179846, + 0.003944760654121637, + 0.06381790339946747, + 0.01713290996849537, + 0.009837821125984192, + 0.005398998036980629, + 0.021149804815649986, + 0.09757832437753677, + 0.008311556652188301, + 0.009400429204106331, + 0.006525859702378511, + 0.0049038296565413475, + 0.005295917857438326, + 0.0061026327311992645, + 0.0054666511714458466, + 0.004211699590086937, + 0.0043149408884346485, + 0.003100276691839099, + 0.008841043338179588, + 0.007469038479030132, + 0.009052084758877754, + 0.008901654742658138, + 0.010210566222667694, + 0.009613544680178165, + 0.006936599500477314, + 0.006786228623241186, + 0.023137636482715607, + 0.014587710611522198, + 0.022754106670618057, + 0.02171850949525833, + 0.020645108073949814, + 0.011965731158852577, + 0.022214291617274284, + 0.010502994060516357, + 0.017454594373703003, + 0.02521374262869358, + 0.026943033561110497, + 0.011913768015801907, + 0.012650148011744022, + 0.01074865460395813, + 0.008468528278172016, + 0.007567272521555424, + 0.012808084487915039, + 0.015116930939257145, + 0.007747405208647251, + 0.015999604016542435, + 0.011086896061897278, + 0.006321154069155455, + 0.006773500703275204, + 0.003695775754749775, + 0.004465917125344276, + 0.006122917402535677, + 0.003933087456971407, + 0.00436950521543622, + 0.0036563442554324865, + 0.004612103570252657, + 0.003189951414242387, + 0.0022085197269916534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.0028177103959023952 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground, and tracking from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a muscular steed with inky-black hair galloping at full speed across a vast, desolate Gobi Desert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone knight wearing a heavy, rough black cape visible, leaning forward while gripping the reins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a setting sun with a blood-red hue sinking under sand dunes, casting a strong orange-red and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a dazzling golden outline light (rim light) visible around the knight and the horse due to the backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633904159069061 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.008633904159069061, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633904159069061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14232566952705383, + 0.056929267942905426, + 0.07306309044361115, + 0.15481771528720856, + 0.07633024454116821, + 0.1020226702094078, + 0.09574525058269501, + 0.09585978835821152, + 0.07604077458381653, + 0.12686553597450256 + ], + "credit_map": [ + 0.016086876392364502, + 0.0162140354514122, + 0.013953950256109238, + 0.020710183307528496, + 0.020524851977825165, + 0.022442270070314407, + 0.017274221405386925, + 0.015119286254048347, + 0.006093145348131657, + 0.006760367192327976, + 0.004675521980971098, + 0.009545625187456608, + 0.007842471823096275, + 0.00781509093940258, + 0.0074525680392980576, + 0.006744478829205036, + 0.014612198807299137, + 0.007907751016318798, + 0.010790582746267319, + 0.009160078130662441, + 0.007584913168102503, + 0.005355301778763533, + 0.011518377810716629, + 0.006133886985480785, + 0.03949138522148132, + 0.012309948913753033, + 0.007738109212368727, + 0.007162016350775957, + 0.015438631176948547, + 0.05583741143345833, + 0.009254890494048595, + 0.0075853182934224606, + 0.010729367844760418, + 0.009397566318511963, + 0.007007160224020481, + 0.009719016961753368, + 0.01613786816596985, + 0.009753827936947346, + 0.008717653341591358, + 0.0048677814193069935, + 0.00964508205652237, + 0.010273108258843422, + 0.00930789764970541, + 0.018501779064536095, + 0.014380142092704773, + 0.0130940992385149, + 0.011744980700314045, + 0.015075575560331345, + 0.013173031620681286, + 0.008751941844820976, + 0.011782512068748474, + 0.009608298540115356, + 0.015200477093458176, + 0.009912590496242046, + 0.016448726877570152, + 0.010867672972381115, + 0.012697882018983364, + 0.020305924117565155, + 0.01778908632695675, + 0.01009919960051775, + 0.010725011117756367, + 0.010487006977200508, + 0.007488446310162544, + 0.006267230957746506, + 0.011595765128731728, + 0.012577501125633717, + 0.008464127779006958, + 0.012682910077273846, + 0.012062872759997845, + 0.008287676610052586, + 0.005778418388217688, + 0.004591505043208599, + 0.019041458144783974, + 0.013730510137975216, + 0.013332576490938663, + 0.012462222017347813, + 0.017158806324005127, + 0.02403208799660206, + 0.014718216843903065, + 0.012389641255140305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.008633904159069061 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_12/index_tensor([4791, 4791])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_12/index_tensor([4791, 4791])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9b880bf584a3ad6bd2a2deab5b317377eab247c0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_12/index_tensor([4791, 4791])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "frantic, first-person perspective shot depicts the viewer as a spotted moth, the camera erratically darting and swerving across a sun-drenched meadow in a desperate chase. The world unfolds in JW's sublime, atmospheric style, reminiscent of a Turner painting, employing a wet-on-wet technique where colors bleed and shapes dissolve into a beautiful, chaotic blur. Just ahead, the target—another Pieris butterfly with delicate, powdery white wings—flickers and drifts out of focus, a fleeting beacon in a swirl of color. The camera plummets downward, and the towering blades of emerald green grass and vibrant wildflowers carve abstract streaks of crimson, gold, and azure. Then, it soars upward into a blazing, hazy sunlight that floods the lens, creating a luminous, dreamlike glow. The chase is a dizzying dance of rapid, desperate movements, as painterly landscapes of light and color swirl through the thick, warm air, creating an overwhelming atmosphere that is frantic in its pursuit and breathtakingly beautiful in its execution.", + "seed": 22747, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_12/index_4791_seed_22747_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.018008060939610004, + "vlm_logprob": -0.018008060939610004, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.018008060939610004, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05275498330593109 + ], + "opposite_logprob_tokens": [ + -3.1777548789978027 + ], + "logprob_reward": -0.05275498330593109, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05275498330593109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051113028079271317, + 0.039904799312353134, + 0.024308031424880028, + 0.025814812630414963, + 0.07882579416036606, + 0.18938115239143372, + 0.3647325336933136, + 0.19677947461605072, + 0.013615594245493412, + 0.015524820424616337 + ], + "credit_map": [ + 0.007775064557790756, + 0.007365575525909662, + 0.0048230234533548355, + 0.005530038382858038, + 0.009097605012357235, + 0.006555797532200813, + 0.005307612009346485, + 0.004658311605453491, + 0.002680730540305376, + 0.006158883683383465, + 0.003934879787266254, + 0.004632364492863417, + 0.008758066222071648, + 0.00367153063416481, + 0.005599772557616234, + 0.00446857325732708, + 0.0019582242239266634, + 0.0023236628621816635, + 0.0027107701171189547, + 0.0037009974475950003, + 0.00322873261757195, + 0.002727413084357977, + 0.004355523735284805, + 0.0033027073368430138, + 0.002878496889024973, + 0.0034597418271005154, + 0.002672298578545451, + 0.0033044577576220036, + 0.0034812542144209146, + 0.00410471623763442, + 0.002924842992797494, + 0.002989005297422409, + 0.00759311905130744, + 0.01927369274199009, + 0.008593551814556122, + 0.008944626897573471, + 0.01055758073925972, + 0.010076621547341347, + 0.008979639038443565, + 0.004806956741958857, + 0.019940922036767006, + 0.015648238360881805, + 0.011503295041620731, + 0.027073806151747704, + 0.01793362945318222, + 0.0227467343211174, + 0.018128518015146255, + 0.05640600621700287, + 0.05708829686045647, + 0.047638069838285446, + 0.1190846711397171, + 0.026956351473927498, + 0.016981111839413643, + 0.026597389951348305, + 0.017752552404999733, + 0.0526341050863266, + 0.022198829799890518, + 0.08745728433132172, + 0.05593429505825043, + 0.007150332443416119, + 0.009291013702750206, + 0.005604851059615612, + 0.0034625830594450235, + 0.005680282600224018, + 0.0016218386590480804, + 0.0025345883332192898, + 0.0013489429838955402, + 0.0013467485550791025, + 0.0019037285819649696, + 0.0020139296539127827, + 0.0014430629089474678, + 0.0014027546858415008, + 0.001485530287027359, + 0.0010699983686208725, + 0.003672038670629263, + 0.0017541350098326802, + 0.002156643196940422, + 0.0015683663077652454, + 0.0012691444717347622, + 0.0025489639956504107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.05275498330593109 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028287656605243683 + ], + "opposite_logprob_tokens": [ + -5.877828598022461 + ], + "logprob_reward": -0.0028287656605243683, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028287656605243683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1587209403514862, + 0.1461443305015564, + 0.0431223139166832, + 0.04845477640628815, + 0.06954994052648544, + 0.0756356492638588, + 0.30169323086738586, + 0.06819235533475876, + 0.037014130502939224, + 0.05147229880094528 + ], + "credit_map": [ + 0.018111007288098335, + 0.02396041713654995, + 0.013734899461269379, + 0.017038490623235703, + 0.032008569687604904, + 0.01623646356165409, + 0.019545068964362144, + 0.018086027354002, + 0.012397757731378078, + 0.037371858954429626, + 0.024941254407167435, + 0.013435794040560722, + 0.0193746630102396, + 0.0136732691898942, + 0.009496968239545822, + 0.015452771447598934, + 0.004063504748046398, + 0.005376320797950029, + 0.005804340820759535, + 0.006451690103858709, + 0.006034075282514095, + 0.005566445179283619, + 0.005790660157799721, + 0.0040352740325033665, + 0.004150974564254284, + 0.005164421629160643, + 0.004826703108847141, + 0.005678099580109119, + 0.005938713438808918, + 0.009214842692017555, + 0.008114395663142204, + 0.0053666215389966965, + 0.0072914958000183105, + 0.012152542360126972, + 0.010251211933791637, + 0.007308127358555794, + 0.008565520867705345, + 0.011057589203119278, + 0.007954738102853298, + 0.004968720488250256, + 0.007252976298332214, + 0.007604075595736504, + 0.005165704991668463, + 0.010188267566263676, + 0.009357491508126259, + 0.009762061759829521, + 0.007120893802493811, + 0.0191841758787632, + 0.044055670499801636, + 0.03864903002977371, + 0.09108027070760727, + 0.026181045919656754, + 0.02092786692082882, + 0.021704912185668945, + 0.020780524238944054, + 0.03831392899155617, + 0.0101315937936306, + 0.026360057294368744, + 0.012969369068741798, + 0.004907673224806786, + 0.004262992646545172, + 0.0034263967536389828, + 0.0027558517176657915, + 0.003378420602530241, + 0.006365249399095774, + 0.004856899380683899, + 0.002678058808669448, + 0.00370840379036963, + 0.004312063567340374, + 0.004654485732316971, + 0.004216702189296484, + 0.006222266238182783, + 0.004499767906963825, + 0.0033761076629161835, + 0.014713916927576065, + 0.015943393111228943, + 0.0032355128787457943, + 0.0032701187301427126, + 0.0037720806431025267, + 0.00266139954328537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0028287656605243683 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784949012100697 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.009784949012100697, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784949012100697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12689459323883057, + 0.06301146000623703, + 0.03266320005059242, + 0.04860931634902954, + 0.05043328180909157, + 0.05127989873290062, + 0.4291176199913025, + 0.14261823892593384, + 0.024094657972455025, + 0.03127770498394966 + ], + "credit_map": [ + 0.01327784825116396, + 0.01707843504846096, + 0.008990551345050335, + 0.01582971028983593, + 0.0246957466006279, + 0.01486867107450962, + 0.01884959079325199, + 0.01330403983592987, + 0.00523463124409318, + 0.01405388955026865, + 0.010890328325331211, + 0.005966899450868368, + 0.009947084821760654, + 0.005994210485368967, + 0.0049591511487960815, + 0.005965264048427343, + 0.0027265597600489855, + 0.003920546732842922, + 0.002894532633945346, + 0.004225132055580616, + 0.0060626850463449955, + 0.005094337742775679, + 0.004370315466076136, + 0.003369088750332594, + 0.00420050323009491, + 0.007745958399027586, + 0.004798830486834049, + 0.006565604358911514, + 0.006859803572297096, + 0.008055012673139572, + 0.005454422906041145, + 0.004929180257022381, + 0.004607676062732935, + 0.006949177477508783, + 0.006117513868957758, + 0.0070732696913182735, + 0.006398634985089302, + 0.007862290367484093, + 0.005435258150100708, + 0.005989461205899715, + 0.00531673151999712, + 0.004707764368504286, + 0.00394509918987751, + 0.006811723578721285, + 0.0059539638459682465, + 0.005953705869615078, + 0.00437559699639678, + 0.014215310104191303, + 0.06404700130224228, + 0.053853388875722885, + 0.1359662115573883, + 0.03269338607788086, + 0.025558127090334892, + 0.03209488466382027, + 0.025682207196950912, + 0.05922241881489754, + 0.0159092228859663, + 0.06198020279407501, + 0.039769262075424194, + 0.005864364560693502, + 0.006844433955848217, + 0.0043105194345116615, + 0.0034551271237432957, + 0.004485100042074919, + 0.003066116012632847, + 0.003642484312877059, + 0.001945146475918591, + 0.0026347548700869083, + 0.00311648054048419, + 0.004684670828282833, + 0.002391268266364932, + 0.0026137367822229862, + 0.002248681616038084, + 0.0017594528617337346, + 0.009383597411215305, + 0.009563429281115532, + 0.0018218174809589982, + 0.002366791246458888, + 0.0024400169495493174, + 0.0016939218621701002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.009784949012100697 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094788804650307 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.004094788804650307, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094788804650307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06887853145599365, + 0.0729743018746376, + 0.021805033087730408, + 0.047187529504299164, + 0.04968121647834778, + 0.3993750810623169, + 0.2196464240550995, + 0.09318788349628448, + 0.01255610678344965, + 0.014707958325743675 + ], + "credit_map": [ + 0.008575359359383583, + 0.010003863833844662, + 0.006288152653723955, + 0.008604294620454311, + 0.008088956587016582, + 0.012702791951596737, + 0.006467057392001152, + 0.00814804993569851, + 0.0050290790386497974, + 0.015283981338143349, + 0.013879138976335526, + 0.0066747041419148445, + 0.01332081202417612, + 0.006093633361160755, + 0.0060026394203305244, + 0.006690301466733217, + 0.0018589587416499853, + 0.0019942088983953, + 0.002709912834689021, + 0.0029669455252587795, + 0.002811727812513709, + 0.0024458791594952345, + 0.003432762809097767, + 0.0035846352111548185, + 0.004105055704712868, + 0.004555646330118179, + 0.008653795346617699, + 0.0042785112746059895, + 0.004644749220460653, + 0.00691969832405448, + 0.00530315050855279, + 0.008726918138563633, + 0.004304096568375826, + 0.010719274170696735, + 0.006516715511679649, + 0.007112763822078705, + 0.005718923173844814, + 0.00558528583496809, + 0.004823773633688688, + 0.004900381434708834, + 0.04515983536839485, + 0.03299514204263687, + 0.02530546486377716, + 0.058475296944379807, + 0.028062794357538223, + 0.0491475947201252, + 0.03512347489595413, + 0.12510544061660767, + 0.03488869220018387, + 0.027355967089533806, + 0.07344973087310791, + 0.019566694274544716, + 0.015158710069954395, + 0.013430509716272354, + 0.012638788670301437, + 0.023157311603426933, + 0.011093654669821262, + 0.03916366398334503, + 0.023652812466025352, + 0.003741692751646042, + 0.004765825811773539, + 0.0041913362219929695, + 0.002852227073162794, + 0.00372666260227561, + 0.0018184464424848557, + 0.0022954766172915697, + 0.001076592248864472, + 0.0012687379494309425, + 0.0023658154532313347, + 0.001572015811689198, + 0.0009676922345533967, + 0.0011913287453353405, + 0.0013709355844184756, + 0.001120672095566988, + 0.002437747549265623, + 0.0039953989908099174, + 0.001451895572245121, + 0.0010698094265535474, + 0.001350370584987104, + 0.001911125727929175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.004094788804650307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057681791484356 + ], + "opposite_logprob_tokens": [ + -3.8955767154693604 + ], + "logprob_reward": -0.02057681791484356, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057681791484356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06877551227807999, + 0.07470732182264328, + 0.02528460882604122, + 0.4599533677101135, + 0.02890182100236416, + 0.06674474477767944, + 0.15605346858501434, + 0.06729870289564133, + 0.02752610109746456, + 0.02475438266992569 + ], + "credit_map": [ + 0.005631166975945234, + 0.007286990061402321, + 0.012465547770261765, + 0.009123716503381729, + 0.008920427411794662, + 0.0071031153202056885, + 0.008180373348295689, + 0.010064172558486462, + 0.006210984196513891, + 0.017083989456295967, + 0.01448888797312975, + 0.008266736753284931, + 0.009570988826453686, + 0.005992700345814228, + 0.006026886869221926, + 0.007066145073622465, + 0.0028339256532490253, + 0.0032827206887304783, + 0.0026339536998420954, + 0.003478306345641613, + 0.0038507485296577215, + 0.003612661501392722, + 0.0023152006324380636, + 0.0032770929392427206, + 0.04033830761909485, + 0.04157616198062897, + 0.10192368179559708, + 0.04596966877579689, + 0.043552905321121216, + 0.06396699696779251, + 0.06943124532699585, + 0.053194426000118256, + 0.0019652170594781637, + 0.004646762274205685, + 0.003135651582852006, + 0.004129908513277769, + 0.00372069189324975, + 0.004703639540821314, + 0.003420298220589757, + 0.00317965354770422, + 0.006297020707279444, + 0.005408193450421095, + 0.005062932148575783, + 0.00899456162005663, + 0.007412871811538935, + 0.008775998838245869, + 0.007256124634295702, + 0.01753704436123371, + 0.022404078394174576, + 0.021713508293032646, + 0.039100658148527145, + 0.021851934492588043, + 0.008554283529520035, + 0.011012427508831024, + 0.010913672856986523, + 0.02050289884209633, + 0.009968925267457962, + 0.022502005100250244, + 0.012056287378072739, + 0.007335479371249676, + 0.0037183889653533697, + 0.004174085799604654, + 0.0035807264503091574, + 0.003962800372391939, + 0.00411501107737422, + 0.005675565451383591, + 0.002290524309501052, + 0.0027307062409818172, + 0.004971706308424473, + 0.0028686358127743006, + 0.0020323118660598993, + 0.0028416409622877836, + 0.002759478287771344, + 0.0018811817280948162, + 0.0034147044643759727, + 0.003499936778098345, + 0.0029407169204205275, + 0.002528103534132242, + 0.0019804344046860933, + 0.005749825853854418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.02057681791484356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "frantic, first-person perspective shot depicts the viewer as a spotted moth, the camera erratically darting and swerving across a sun-drenched meadow in a desperate chase. The world unfolds in JW's sublime, atmospheric style, reminiscent of a Turner painting, employing a wet-on-wet technique where colors bleed and shapes dissolve into a beautiful, chaotic blur. Just ahead, the target—another Pieris butterfly with delicate, powdery white wings—flickers and drifts out of focus, a fleeting beacon in a swirl of color. The camera plummets downward, and the towering blades of emerald green grass and vibrant wildflowers carve abstract streaks of crimson, gold, and azure. Then, it soars upward into a blazing, hazy sunlight that floods the lens, creating a luminous, dreamlike glow. The chase is a dizzying dance of rapid, desperate movements, as painterly landscapes of light and color swirl through the thick, warm air, creating an overwhelming atmosphere that is frantic in its pursuit and breathtakingly beautiful in its execution.", + "seed": 22748, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_12/index_4791_seed_22748_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.015763120958581567, + "vlm_logprob": -0.015763120958581567, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.015763120958581567, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03443571552634239 + ], + "opposite_logprob_tokens": [ + -3.534435749053955 + ], + "logprob_reward": -0.03443571552634239, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03443571552634239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08012489229440689, + 0.06780245900154114, + 0.13499484956264496, + 0.10535997897386551, + 0.04540202021598816, + 0.05861400067806244, + 0.22038115561008453, + 0.07116538286209106, + 0.09815443307161331, + 0.1180008202791214 + ], + "credit_map": [ + 0.010519535280764103, + 0.012655902653932571, + 0.006723606958985329, + 0.006306737661361694, + 0.015106206759810448, + 0.011565208435058594, + 0.009975744411349297, + 0.007271946873515844, + 0.00656528864055872, + 0.004957882687449455, + 0.006455219816416502, + 0.01637851446866989, + 0.005708461161702871, + 0.006902209483087063, + 0.014003410004079342, + 0.006831473205238581, + 0.014532708562910557, + 0.024265917018055916, + 0.011075649410486221, + 0.016855457797646523, + 0.019962381571531296, + 0.019712120294570923, + 0.01516054105013609, + 0.013430071994662285, + 0.013545757159590721, + 0.01582224853336811, + 0.008635301142930984, + 0.019281070679426193, + 0.01833944581449032, + 0.00987639743834734, + 0.010163248516619205, + 0.009696515277028084, + 0.0036086058244109154, + 0.003488818882033229, + 0.004466830752789974, + 0.004712122492492199, + 0.0045919581316411495, + 0.009410617873072624, + 0.0041543180122971535, + 0.01096874475479126, + 0.005082026589661837, + 0.005384106654673815, + 0.005234931129962206, + 0.008445531129837036, + 0.010587875731289387, + 0.008240539580583572, + 0.010924831964075565, + 0.004714160691946745, + 0.011482005007565022, + 0.018375292420387268, + 0.00578063540160656, + 0.007114935666322708, + 0.08373989164829254, + 0.07074939459562302, + 0.009776218794286251, + 0.013362780213356018, + 0.005978013388812542, + 0.005022070836275816, + 0.013024051673710346, + 0.00918504036962986, + 0.009023720398545265, + 0.009269868023693562, + 0.007987277582287788, + 0.011675341986119747, + 0.014591862447559834, + 0.00755401561036706, + 0.027289601042866707, + 0.022754497826099396, + 0.00717801321297884, + 0.0047101425006985664, + 0.008387427777051926, + 0.005688868463039398, + 0.007331946864724159, + 0.01765374094247818, + 0.01029935386031866, + 0.007911549881100655, + 0.012216811068356037, + 0.034968867897987366, + 0.006132126320153475, + 0.021486416459083557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 9 + ], + "selected_reward": -0.03443571552634239 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024951535742729902 + ], + "opposite_logprob_tokens": [ + -6.002495288848877 + ], + "logprob_reward": -0.0024951535742729902, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024951535742729902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08063904196023941, + 0.08541596680879593, + 0.05674511194229126, + 0.1301879733800888, + 0.09323501586914062, + 0.03324398025870323, + 0.050209954380989075, + 0.07530288398265839, + 0.16399644315242767, + 0.23102359473705292 + ], + "credit_map": [ + 0.011771590448915958, + 0.010464385151863098, + 0.00739891454577446, + 0.008992873132228851, + 0.010262133553624153, + 0.011978008784353733, + 0.010474265553057194, + 0.00929687637835741, + 0.007537000812590122, + 0.007748179137706757, + 0.012769248336553574, + 0.0116244712844491, + 0.011295127682387829, + 0.012070054188370705, + 0.013291442766785622, + 0.009080448187887669, + 0.006154448259621859, + 0.00494046276435256, + 0.006951954215764999, + 0.009294489398598671, + 0.0065893735736608505, + 0.007468513213098049, + 0.007895291782915592, + 0.007450583390891552, + 0.010790755040943623, + 0.011685451492667198, + 0.01346411369740963, + 0.025159958750009537, + 0.013897265307605267, + 0.013568096794188023, + 0.0175658967345953, + 0.0240564476698637, + 0.015040554106235504, + 0.007971717976033688, + 0.009289996698498726, + 0.005291525740176439, + 0.011257392354309559, + 0.027651682496070862, + 0.008604391478002071, + 0.00812776479870081, + 0.0031536915339529514, + 0.002874355763196945, + 0.003007345600053668, + 0.004014038946479559, + 0.006611362099647522, + 0.0033704813104122877, + 0.0045774574391543865, + 0.0056352512910962105, + 0.0072591486386954784, + 0.006376524921506643, + 0.0056601641699671745, + 0.006028046365827322, + 0.00722972396761179, + 0.005514693912118673, + 0.004923867527395487, + 0.007217791862785816, + 0.005347618367522955, + 0.005347986705601215, + 0.0146314213052392, + 0.009741480462253094, + 0.007279868237674236, + 0.009861956350505352, + 0.012785379774868488, + 0.010307179763913155, + 0.019852856174111366, + 0.008357523940503597, + 0.04529675468802452, + 0.048613324761390686, + 0.011779082007706165, + 0.00781667698174715, + 0.014331676065921783, + 0.00794854573905468, + 0.01477036066353321, + 0.03468317538499832, + 0.019802290946245193, + 0.015535852871835232, + 0.02407659776508808, + 0.0694727674126625, + 0.011923256330192089, + 0.04075929522514343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.0024951535742729902 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369717672467232 + ], + "opposite_logprob_tokens": [ + -3.4086971282958984 + ], + "logprob_reward": -0.03369717672467232, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369717672467232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10265493392944336, + 0.04649430140852928, + 0.11216898262500763, + 0.445890337228775, + 0.04258255288004875, + 0.020565781742334366, + 0.055155035108327866, + 0.04566729813814163, + 0.09785497933626175, + 0.030965808779001236 + ], + "credit_map": [ + 0.012695707380771637, + 0.014390128664672375, + 0.010827185586094856, + 0.013399709947407246, + 0.01723017543554306, + 0.011728465557098389, + 0.012161759659647942, + 0.010221794247627258, + 0.0050969296135008335, + 0.0054435390047729015, + 0.00343784736469388, + 0.005746819544583559, + 0.006833319086581469, + 0.009078285656869411, + 0.0063209375366568565, + 0.00453662034124136, + 0.012729546055197716, + 0.019609086215496063, + 0.00986228883266449, + 0.014440595172345638, + 0.01652224361896515, + 0.014816227369010448, + 0.01419888250529766, + 0.009990098886191845, + 0.03758402541279793, + 0.047987889498472214, + 0.049791354686021805, + 0.04421764239668846, + 0.07247589528560638, + 0.04730994254350662, + 0.07262739539146423, + 0.07389615476131439, + 0.005625611636787653, + 0.004637941252440214, + 0.005137159489095211, + 0.00491626001894474, + 0.005702240392565727, + 0.005669860634952784, + 0.005494393408298492, + 0.005399079527705908, + 0.0016530744032934308, + 0.0015298565849661827, + 0.0015848394250497222, + 0.0019723745062947273, + 0.005342348013073206, + 0.002189073245972395, + 0.0022887252271175385, + 0.0040054903365671635, + 0.0032786265946924686, + 0.004207960795611143, + 0.0023961267434060574, + 0.0032227516639977694, + 0.018118320032954216, + 0.015272840857505798, + 0.0037273946218192577, + 0.004931009374558926, + 0.0034736469388008118, + 0.0031267269514501095, + 0.009669662453234196, + 0.0054327151738107204, + 0.004986568354070187, + 0.005570836830884218, + 0.007365144323557615, + 0.006041993852704763, + 0.011546151712536812, + 0.004913953132927418, + 0.0279379952698946, + 0.030415160581469536, + 0.006854478269815445, + 0.0036973634269088507, + 0.009629781357944012, + 0.002860085340216756, + 0.0032152014318853617, + 0.004272928927093744, + 0.0021353713236749172, + 0.0032660262659192085, + 0.0038070506416261196, + 0.004730421118438244, + 0.004676954820752144, + 0.004861851688474417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.03369717672467232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040934826247394085 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.0040934826247394085, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040934826247394085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06812456250190735, + 0.050081681460142136, + 0.07980125397443771, + 0.1004793718457222, + 0.11650235950946808, + 0.07406894117593765, + 0.20261719822883606, + 0.06027419492602348, + 0.07696533203125, + 0.17108511924743652 + ], + "credit_map": [ + 0.005566553212702274, + 0.009388042613863945, + 0.00563535513356328, + 0.006988306995481253, + 0.015430258587002754, + 0.011010324582457542, + 0.007125342730432749, + 0.006980383303016424, + 0.005397549830377102, + 0.004967260640114546, + 0.004979324527084827, + 0.013591396622359753, + 0.006373159121721983, + 0.005330656189471483, + 0.005869561340659857, + 0.0035727734211832285, + 0.00887551810592413, + 0.012339255772531033, + 0.007596885785460472, + 0.010065649636089802, + 0.011554739437997341, + 0.011648847721517086, + 0.01003200188279152, + 0.007688357029110193, + 0.008488183841109276, + 0.011577053926885128, + 0.009676187299191952, + 0.010510011576116085, + 0.019322771579027176, + 0.007823999039828777, + 0.013372361660003662, + 0.01970881037414074, + 0.006742129102349281, + 0.008123654872179031, + 0.011167187243700027, + 0.0071339900605380535, + 0.006969850044697523, + 0.03646552562713623, + 0.007977345027029514, + 0.031922679394483566, + 0.006495583336800337, + 0.00486089987680316, + 0.006973522249609232, + 0.006738719530403614, + 0.014631420373916626, + 0.00683017261326313, + 0.010605372488498688, + 0.01693325862288475, + 0.017377043142914772, + 0.015592892654240131, + 0.013565095141530037, + 0.015991853550076485, + 0.060827940702438354, + 0.04728352651000023, + 0.014964340254664421, + 0.017014507204294205, + 0.004781548399478197, + 0.004071872681379318, + 0.009486507624387741, + 0.007758956402540207, + 0.005279759876430035, + 0.007032329682260752, + 0.012879027985036373, + 0.008984189480543137, + 0.006855059415102005, + 0.00651559978723526, + 0.016063261777162552, + 0.02039720118045807, + 0.006083840969949961, + 0.005027709528803825, + 0.00895802304148674, + 0.007064629811793566, + 0.010320807807147503, + 0.028876302763819695, + 0.014639333821833134, + 0.012248508632183075, + 0.017357897013425827, + 0.0501980297267437, + 0.008842254988849163, + 0.028601977974176407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.0040934826247394085 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094076342880726 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004094076342880726, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094076342880726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04966473951935768, + 0.028324510902166367, + 0.12446078658103943, + 0.10505162924528122, + 0.10757427662611008, + 0.04019426554441452, + 0.09294965118169785, + 0.04714221879839897, + 0.29716867208480835, + 0.10746926069259644 + ], + "credit_map": [ + 0.006602812558412552, + 0.007021878845989704, + 0.004797959700226784, + 0.005466731730848551, + 0.006837291177362204, + 0.005960970651358366, + 0.0062459781765937805, + 0.006731112487614155, + 0.0024666704703122377, + 0.0025258176028728485, + 0.0033925382886081934, + 0.004175018519163132, + 0.0026110929902642965, + 0.00418188888579607, + 0.00607288908213377, + 0.002898594131693244, + 0.012561413459479809, + 0.01995489001274109, + 0.01310002338141203, + 0.01432932261377573, + 0.017808102071285248, + 0.018336256965994835, + 0.01622597500681877, + 0.012144800275564194, + 0.008712383918464184, + 0.00881502777338028, + 0.012921540066599846, + 0.017475171014666557, + 0.01190693024545908, + 0.01082055363804102, + 0.015235484577715397, + 0.01916453428566456, + 0.009633725509047508, + 0.007670853286981583, + 0.009289872832596302, + 0.005093467887490988, + 0.011739624664187431, + 0.03762158006429672, + 0.008707163855433464, + 0.01781798154115677, + 0.0024763878900557756, + 0.0022472883574664593, + 0.0033011247869580984, + 0.002330394694581628, + 0.011843616142868996, + 0.0035878559574484825, + 0.0028132889419794083, + 0.011594309471547604, + 0.005344046279788017, + 0.006748014595359564, + 0.005761821288615465, + 0.006497149355709553, + 0.028789017349481583, + 0.024378664791584015, + 0.005819299258291721, + 0.009611641056835651, + 0.003654753789305687, + 0.004000185057520866, + 0.005871630273759365, + 0.004825984127819538, + 0.004006878938525915, + 0.00625338451936841, + 0.01171194203197956, + 0.006817462854087353, + 0.03110329434275627, + 0.02029314823448658, + 0.07664288580417633, + 0.09062454104423523, + 0.02181200496852398, + 0.012480889447033405, + 0.03010649047791958, + 0.014105429872870445, + 0.0056610736064612865, + 0.01878664828836918, + 0.009338592179119587, + 0.007776818238198757, + 0.0101705277338624, + 0.031806934624910355, + 0.005532203242182732, + 0.018396465107798576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.004094076342880726 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "frantic, first-person perspective shot depicts the viewer as a spotted moth, the camera erratically darting and swerving across a sun-drenched meadow in a desperate chase. The world unfolds in JW's sublime, atmospheric style, reminiscent of a Turner painting, employing a wet-on-wet technique where colors bleed and shapes dissolve into a beautiful, chaotic blur. Just ahead, the target—another Pieris butterfly with delicate, powdery white wings—flickers and drifts out of focus, a fleeting beacon in a swirl of color. The camera plummets downward, and the towering blades of emerald green grass and vibrant wildflowers carve abstract streaks of crimson, gold, and azure. Then, it soars upward into a blazing, hazy sunlight that floods the lens, creating a luminous, dreamlike glow. The chase is a dizzying dance of rapid, desperate movements, as painterly landscapes of light and color swirl through the thick, warm air, creating an overwhelming atmosphere that is frantic in its pursuit and breathtakingly beautiful in its execution.", + "seed": 22749, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_12/index_4791_seed_22749_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05385271208360791, + "vlm_logprob": -0.05385271208360791, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.05385271208360791, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06169115751981735 + ], + "opposite_logprob_tokens": [ + -2.9366910457611084 + ], + "logprob_reward": -0.06169115751981735, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06169115751981735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14076165854930878, + 0.04197058081626892, + 0.06343046575784683, + 0.046777669340372086, + 0.07068737596273422, + 0.2994849681854248, + 0.17075970768928528, + 0.09990484267473221, + 0.038770053535699844, + 0.027452697977423668 + ], + "credit_map": [ + 0.028396788984537125, + 0.0324389822781086, + 0.014302278868854046, + 0.014395620673894882, + 0.015129406936466694, + 0.012347173877060413, + 0.01669451594352722, + 0.007056891452521086, + 0.0034524754155427217, + 0.0053690834902226925, + 0.0059923757798969746, + 0.007792651187628508, + 0.004022839013487101, + 0.006979235913604498, + 0.00386657752096653, + 0.004495344124734402, + 0.0050521232187747955, + 0.004397465381771326, + 0.0062777032144367695, + 0.006021833512932062, + 0.004901840817183256, + 0.01445486769080162, + 0.005499586928635836, + 0.01682504452764988, + 0.00509314239025116, + 0.007043422665446997, + 0.004172215703874826, + 0.004183117300271988, + 0.00961366854608059, + 0.0053284140303730965, + 0.006654596887528896, + 0.004689088091254234, + 0.005571389105170965, + 0.008211047388613224, + 0.006394030526280403, + 0.01077068131417036, + 0.007438125554472208, + 0.010957731865346432, + 0.013866649940609932, + 0.007477717008441687, + 0.018704449757933617, + 0.08075946569442749, + 0.019619164988398552, + 0.024339303374290466, + 0.055492572486400604, + 0.033747751265764236, + 0.02485910803079605, + 0.04196314886212349, + 0.012886765412986279, + 0.030941732227802277, + 0.017970914021134377, + 0.022237710654735565, + 0.034278206527233124, + 0.022274821996688843, + 0.01677670329809189, + 0.013392855413258076, + 0.010211635380983353, + 0.021240515634417534, + 0.005500455852597952, + 0.006773819215595722, + 0.020358502864837646, + 0.01901470124721527, + 0.005903725046664476, + 0.010901487432420254, + 0.004942864645272493, + 0.0037755793891847134, + 0.003803272731602192, + 0.005157215520739555, + 0.005693086422979832, + 0.00516249006614089, + 0.005379026290029287, + 0.004856516607105732, + 0.003645482938736677, + 0.003434223122894764, + 0.00408533401787281, + 0.0044237179681658745, + 0.0026033965405076742, + 0.0025692651979625225, + 0.0028593179304152727, + 0.0038319590967148542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.06169115751981735 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012548891827464104 + ], + "opposite_logprob_tokens": [ + -4.387548923492432 + ], + "logprob_reward": -0.012548891827464104, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012548891827464104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14241889119148254, + 0.06738465279340744, + 0.2395421415567398, + 0.03415309265255928, + 0.03430386632680893, + 0.30084937810897827, + 0.05040017515420914, + 0.053709808737039566, + 0.05057327076792717, + 0.026664702221751213 + ], + "credit_map": [ + 0.018895043060183525, + 0.020620500668883324, + 0.012515840120613575, + 0.021305225789546967, + 0.01564936898648739, + 0.013996356166899204, + 0.029917221516370773, + 0.009519330225884914, + 0.007831581868231297, + 0.010270746424794197, + 0.007490555755794048, + 0.012135934084653854, + 0.010250246152281761, + 0.006975115276873112, + 0.005879187025129795, + 0.0065512824803590775, + 0.010904940776526928, + 0.00946966651827097, + 0.0139295794069767, + 0.012423497624695301, + 0.009846556931734085, + 0.08364155143499374, + 0.010827501304447651, + 0.08849883079528809, + 0.00239747716113925, + 0.003824306419119239, + 0.0023620440624654293, + 0.004723283927887678, + 0.004589402582496405, + 0.004367648623883724, + 0.004825272597372532, + 0.007063655182719231, + 0.0029347664676606655, + 0.006147963926196098, + 0.0037294120993465185, + 0.003814383177086711, + 0.0037561748176813126, + 0.002975121373310685, + 0.006283363793045282, + 0.004662679508328438, + 0.019567767158150673, + 0.06865941733121872, + 0.015281593427062035, + 0.025390801951289177, + 0.055102404206991196, + 0.03945480287075043, + 0.04366179183125496, + 0.03373080864548683, + 0.004431813023984432, + 0.005779819097369909, + 0.006445673294365406, + 0.010266434401273727, + 0.005866247229278088, + 0.004871335346251726, + 0.007531358860433102, + 0.0052074952982366085, + 0.005642993375658989, + 0.012517228722572327, + 0.0036640390753746033, + 0.004105560481548309, + 0.008230924606323242, + 0.009725536219775677, + 0.0032213577069342136, + 0.006602164823561907, + 0.0037680696696043015, + 0.00416291318833828, + 0.0046246107667684555, + 0.004348853137344122, + 0.00933796912431717, + 0.006305878050625324, + 0.007638727314770222, + 0.010386248119175434, + 0.0034227592404931784, + 0.004131279420107603, + 0.0035779017489403486, + 0.003056880319491029, + 0.00235438859090209, + 0.0034359940327703953, + 0.002334898104891181, + 0.00435060216113925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.012548891827464104 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062041785567998886 + ], + "opposite_logprob_tokens": [ + -2.812041759490967 + ], + "logprob_reward": -0.062041785567998886, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062041785567998886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045156266540288925, + 0.03389962762594223, + 0.3338724672794342, + 0.028397956863045692, + 0.048214539885520935, + 0.4053114056587219, + 0.03875252977013588, + 0.03241942077875137, + 0.01451296079903841, + 0.019462840631604195 + ], + "credit_map": [ + 0.008868863806128502, + 0.010720723308622837, + 0.004643317312002182, + 0.006000928580760956, + 0.003938601352274418, + 0.0041991365142166615, + 0.004223295953124762, + 0.002561398083344102, + 0.005445872899144888, + 0.007062132004648447, + 0.0036705676466226578, + 0.004203569609671831, + 0.004575724713504314, + 0.0030944999307394028, + 0.0030802187975496054, + 0.00276704179123044, + 0.014009802602231503, + 0.007194755133241415, + 0.01419183798134327, + 0.009109028615057468, + 0.01189933717250824, + 0.1325666457414627, + 0.008197770453989506, + 0.13670329749584198, + 0.0013961098156869411, + 0.0030344906263053417, + 0.0028641715180128813, + 0.00418842351064086, + 0.003365178592503071, + 0.0031942715868353844, + 0.0030244230292737484, + 0.007330888416618109, + 0.003875540802255273, + 0.00886935368180275, + 0.0037206560373306274, + 0.0068397377617657185, + 0.004902314860373735, + 0.004968318156898022, + 0.010149468667805195, + 0.004889150150120258, + 0.030162138864398003, + 0.10828440636396408, + 0.024691186845302582, + 0.031047962605953217, + 0.054761163890361786, + 0.03711613640189171, + 0.041675157845020294, + 0.0775732472538948, + 0.003391375532373786, + 0.007112144958227873, + 0.003238831413909793, + 0.003820140613242984, + 0.006909976247698069, + 0.005691153462976217, + 0.005289802327752113, + 0.0032991033513098955, + 0.0032830690033733845, + 0.0054470873437821865, + 0.001971133053302765, + 0.002175640780478716, + 0.0070639001205563545, + 0.005544568412005901, + 0.0022467051167041063, + 0.004687315318733454, + 0.0019305743044242263, + 0.001271505025215447, + 0.001308166654780507, + 0.001621935865841806, + 0.0026090277824550867, + 0.0020164046436548233, + 0.001420019892975688, + 0.0023353281430900097, + 0.0025790915824472904, + 0.002565172268077731, + 0.003028500359505415, + 0.002742644166573882, + 0.001837055548094213, + 0.0021096081472933292, + 0.0021165774669498205, + 0.002484191907569766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.062041785567998886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950473714619875 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950473714619875, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950473714619875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06987552344799042, + 0.06511930376291275, + 0.1965837925672531, + 0.03678002581000328, + 0.05724611505866051, + 0.2818240821361542, + 0.09755049645900726, + 0.06365696340799332, + 0.042898379266262054, + 0.08846531063318253 + ], + "credit_map": [ + 0.008767022751271725, + 0.00998335424810648, + 0.007229920476675034, + 0.01014203205704689, + 0.009002991952002048, + 0.006633708253502846, + 0.013497577048838139, + 0.0046189152635633945, + 0.009864701889455318, + 0.014443619176745415, + 0.007674648426473141, + 0.009426829405128956, + 0.007341643795371056, + 0.006919973995536566, + 0.005162614397704601, + 0.004285273142158985, + 0.008430327288806438, + 0.003695088205859065, + 0.008277399465441704, + 0.004332331940531731, + 0.006739624310284853, + 0.08032367378473282, + 0.00400211475789547, + 0.08078323304653168, + 0.0037793503142893314, + 0.0037953513674438, + 0.004490554798394442, + 0.00457598315551877, + 0.004877711646258831, + 0.004728029947727919, + 0.006437146104872227, + 0.004095899872481823, + 0.004864350892603397, + 0.009737261570990086, + 0.006272889673709869, + 0.007033693138509989, + 0.007389303762465715, + 0.006080883089452982, + 0.009251135401427746, + 0.0066165984608232975, + 0.014828416518867016, + 0.09075789898633957, + 0.011278289370238781, + 0.015529502183198929, + 0.04820440337061882, + 0.019181597977876663, + 0.021591350436210632, + 0.06045261397957802, + 0.0077293342910707, + 0.015060080215334892, + 0.014206982217729092, + 0.017302459105849266, + 0.009483128786087036, + 0.008848162367939949, + 0.013422399759292603, + 0.011497944593429565, + 0.004653623327612877, + 0.009824567474424839, + 0.003915802575647831, + 0.005105437710881233, + 0.01634841412305832, + 0.011363481171429157, + 0.007030108943581581, + 0.005415533669292927, + 0.004297133535146713, + 0.004116038791835308, + 0.004745153710246086, + 0.0038697232957929373, + 0.007378594018518925, + 0.005971882026642561, + 0.005272130481898785, + 0.007247720379382372, + 0.01367932464927435, + 0.00956202857196331, + 0.01573585532605648, + 0.015096438117325306, + 0.006500783376395702, + 0.006933034397661686, + 0.008849683217704296, + 0.01210816204547882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.005950473714619875 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703125178813934 + ], + "opposite_logprob_tokens": [ + -2.1270313262939453 + ], + "logprob_reward": -0.12703125178813934, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703125178813934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14862306416034698, + 0.04504778981208801, + 0.12689177691936493, + 0.0419507697224617, + 0.05628622695803642, + 0.38156378269195557, + 0.04600359499454498, + 0.04501206800341606, + 0.08115029335021973, + 0.027470611035823822 + ], + "credit_map": [ + 0.014062982052564621, + 0.013787807896733284, + 0.016436725854873657, + 0.024257907643914223, + 0.017016079276800156, + 0.016198791563510895, + 0.036323048174381256, + 0.010539720766246319, + 0.003943732939660549, + 0.005605287849903107, + 0.007546145468950272, + 0.01037773210555315, + 0.005053976085036993, + 0.004679032601416111, + 0.004355063661932945, + 0.0034868167713284492, + 0.005722899921238422, + 0.0036675110459327698, + 0.00607340270653367, + 0.004070580005645752, + 0.004988102708011866, + 0.050159987062215805, + 0.0030122199095785618, + 0.04919706657528877, + 0.0026279669255018234, + 0.00544574623927474, + 0.00371481291949749, + 0.007796018850058317, + 0.0065299891866743565, + 0.008418791927397251, + 0.005058908369392157, + 0.0023585360031574965, + 0.005996805615723133, + 0.006868766155093908, + 0.007669965736567974, + 0.0064087421633303165, + 0.005120306275784969, + 0.006622600834816694, + 0.010478497482836246, + 0.007120540365576744, + 0.022354600951075554, + 0.1026996597647667, + 0.016605019569396973, + 0.026756752282381058, + 0.07004683464765549, + 0.033490195870399475, + 0.05710480734705925, + 0.05250592157244682, + 0.0044279638677835464, + 0.005938410293310881, + 0.005331131163984537, + 0.006148373242467642, + 0.00880265049636364, + 0.005884747486561537, + 0.005286964122205973, + 0.00418335385620594, + 0.003650814527645707, + 0.006932464893907309, + 0.003160260384902358, + 0.0029958586674183607, + 0.01181735284626484, + 0.008251792751252651, + 0.0042276689782738686, + 0.0039758519269526005, + 0.006840724032372236, + 0.006497764959931374, + 0.008281135931611061, + 0.006400482729077339, + 0.014056296087801456, + 0.01025168877094984, + 0.012010551989078522, + 0.01681165024638176, + 0.003189429175108671, + 0.003127730917185545, + 0.004374837502837181, + 0.0049595157615840435, + 0.0027470518834888935, + 0.00293724169023335, + 0.002313719131052494, + 0.0038210847415030003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.12703125178813934 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "frantic, first-person perspective shot depicts the viewer as a spotted moth, the camera erratically darting and swerving across a sun-drenched meadow in a desperate chase. The world unfolds in JW's sublime, atmospheric style, reminiscent of a Turner painting, employing a wet-on-wet technique where colors bleed and shapes dissolve into a beautiful, chaotic blur. Just ahead, the target—another Pieris butterfly with delicate, powdery white wings—flickers and drifts out of focus, a fleeting beacon in a swirl of color. The camera plummets downward, and the towering blades of emerald green grass and vibrant wildflowers carve abstract streaks of crimson, gold, and azure. Then, it soars upward into a blazing, hazy sunlight that floods the lens, creating a luminous, dreamlike glow. The chase is a dizzying dance of rapid, desperate movements, as painterly landscapes of light and color swirl through the thick, warm air, creating an overwhelming atmosphere that is frantic in its pursuit and breathtakingly beautiful in its execution.", + "seed": 22750, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_12/index_4791_seed_22750_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04116591764613986, + "vlm_logprob": -0.04116591764613986, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.04116591764613986, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.077737957239151 + ], + "opposite_logprob_tokens": [ + -2.827738046646118 + ], + "logprob_reward": -0.077737957239151, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.077737957239151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16987964510917664, + 0.06362015008926392, + 0.03002605400979519, + 0.0433245487511158, + 0.14273414015769958, + 0.049113720655441284, + 0.06630488485097885, + 0.051293618977069855, + 0.13294023275375366, + 0.2507629990577698 + ], + "credit_map": [ + 0.0209237989038229, + 0.01827671378850937, + 0.01974630542099476, + 0.02513439767062664, + 0.021927278488874435, + 0.023277636617422104, + 0.021884847432374954, + 0.018708666786551476, + 0.004365186672657728, + 0.003724602982401848, + 0.007274591829627752, + 0.0041230786591768265, + 0.0026742678601294756, + 0.022436128929257393, + 0.002894630189985037, + 0.016127660870552063, + 0.0030586887151002884, + 0.004322173073887825, + 0.002513266634196043, + 0.004135564900934696, + 0.004860435612499714, + 0.003986330237239599, + 0.003279820317402482, + 0.003869774052873254, + 0.005881933961063623, + 0.010836410336196423, + 0.006886027753353119, + 0.004024176858365536, + 0.003149238880723715, + 0.0053100320510566235, + 0.002752901753410697, + 0.004483823198825121, + 0.005212867166846991, + 0.006398976780474186, + 0.009122391231358051, + 0.006114382762461901, + 0.01126975379884243, + 0.05458858609199524, + 0.03986213728785515, + 0.010165051557123661, + 0.0028038709424436092, + 0.0036608553491532803, + 0.00705808075144887, + 0.004407999571412802, + 0.003693539183586836, + 0.015069863758981228, + 0.0031222791876643896, + 0.009297233074903488, + 0.007381500210613012, + 0.01023857295513153, + 0.007682316936552525, + 0.012749413028359413, + 0.0071646710857748985, + 0.005757424049079418, + 0.006603245157748461, + 0.00872773863375187, + 0.006520258728414774, + 0.0063936044462025166, + 0.004706765990704298, + 0.004837776068598032, + 0.008229287341237068, + 0.008476567454636097, + 0.005526288412511349, + 0.006603072863072157, + 0.013955696485936642, + 0.014364739879965782, + 0.02078372985124588, + 0.017539553344249725, + 0.020579654723405838, + 0.017271926626563072, + 0.015033284202218056, + 0.013411653228104115, + 0.01198104303330183, + 0.09276112914085388, + 0.010455873794853687, + 0.010449296794831753, + 0.018910696730017662, + 0.09033706039190292, + 0.007041474338620901, + 0.008826415054500103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.077737957239151 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410511763766408 + ], + "opposite_logprob_tokens": [ + -5.504105091094971 + ], + "logprob_reward": -0.00410511763766408, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410511763766408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17461282014846802, + 0.12865668535232544, + 0.027415944263339043, + 0.14293654263019562, + 0.048135850578546524, + 0.0720578283071518, + 0.028180578723549843, + 0.10960738360881805, + 0.10561719536781311, + 0.16277910768985748 + ], + "credit_map": [ + 0.018669432029128075, + 0.01943683996796608, + 0.020475154742598534, + 0.023191235959529877, + 0.021989619359374046, + 0.024218501523137093, + 0.023324381560087204, + 0.02330768294632435, + 0.006630083080381155, + 0.005128343589603901, + 0.014306421391665936, + 0.005116836633533239, + 0.004687495995312929, + 0.046276818960905075, + 0.0040829479694366455, + 0.042427752166986465, + 0.003499535145238042, + 0.004264977294951677, + 0.0021087476052343845, + 0.0025879954919219017, + 0.0032454323954880238, + 0.0038990562316030264, + 0.002961542923003435, + 0.004848660435527563, + 0.01619860902428627, + 0.03825666755437851, + 0.022807173430919647, + 0.013231716118752956, + 0.011266596615314484, + 0.016643773764371872, + 0.01000972744077444, + 0.01452229917049408, + 0.005820657126605511, + 0.004593703895807266, + 0.008235080167651176, + 0.004220965318381786, + 0.0038630750495940447, + 0.00779707869514823, + 0.007006690371781588, + 0.006598604377359152, + 0.005760558415204287, + 0.008623650297522545, + 0.009033327922224998, + 0.006435068789869547, + 0.009461622685194016, + 0.01734844036400318, + 0.005279631353914738, + 0.010115541517734528, + 0.0035021768417209387, + 0.0036101306322962046, + 0.0034674464259296656, + 0.004916511010378599, + 0.0026072917971760035, + 0.002550936071202159, + 0.0027355910278856754, + 0.004790496546775103, + 0.010564710013568401, + 0.01131915021687746, + 0.02291586622595787, + 0.012693638913333416, + 0.010665851645171642, + 0.008825208991765976, + 0.017071451991796494, + 0.015551523305475712, + 0.010094947181642056, + 0.010670064948499203, + 0.016130803152918816, + 0.02215602435171604, + 0.010055438615381718, + 0.00864403136074543, + 0.014549150131642818, + 0.0133167440071702, + 0.009855156764388084, + 0.045509081333875656, + 0.012529111467301846, + 0.010184166952967644, + 0.01928679458796978, + 0.05063261836767197, + 0.006514210253953934, + 0.00826799776405096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.00410511763766408 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029835253953934 + ], + "opposite_logprob_tokens": [ + -2.3502984046936035 + ], + "logprob_reward": -0.10029835253953934, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029835253953934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15790264308452606, + 0.13909992575645447, + 0.02495509386062622, + 0.03245396912097931, + 0.2655424177646637, + 0.051375892013311386, + 0.07605943083763123, + 0.0899357795715332, + 0.07052692770957947, + 0.09214786440134048 + ], + "credit_map": [ + 0.017859093844890594, + 0.01544609572738409, + 0.018874136731028557, + 0.02501087822020054, + 0.020555896684527397, + 0.021090006455779076, + 0.022585656493902206, + 0.016480877995491028, + 0.006342906504869461, + 0.004929334856569767, + 0.011365084908902645, + 0.005059428047388792, + 0.0044090538285672665, + 0.0537685789167881, + 0.0039518955163657665, + 0.049273643642663956, + 0.004115693736821413, + 0.0034281581174582243, + 0.0023869620636105537, + 0.0031371198128908873, + 0.0024052623193711042, + 0.0037138713523745537, + 0.002818986540660262, + 0.0029490396846085787, + 0.00421245489269495, + 0.004646816290915012, + 0.006770453415811062, + 0.0033662060741335154, + 0.0025789060164242983, + 0.0036829998716712, + 0.002788643119856715, + 0.004407490603625774, + 0.010903163813054562, + 0.007117919158190489, + 0.01666976884007454, + 0.01134523469954729, + 0.019163955003023148, + 0.10773342102766037, + 0.07888751477003098, + 0.013721448369324207, + 0.003987439442425966, + 0.006291411351412535, + 0.007027698215097189, + 0.005256791599094868, + 0.004910398740321398, + 0.012640603817999363, + 0.003938727080821991, + 0.007322823628783226, + 0.008558021858334541, + 0.01157276052981615, + 0.009226800873875618, + 0.01456918939948082, + 0.008619150146842003, + 0.006297840271145105, + 0.007426236756145954, + 0.009789427742362022, + 0.008773635141551495, + 0.010014103725552559, + 0.008160151541233063, + 0.008142968639731407, + 0.017298031598329544, + 0.016482211649417877, + 0.009831350296735764, + 0.011233322322368622, + 0.005791146773844957, + 0.006347633432596922, + 0.01529660914093256, + 0.008455341681838036, + 0.0127649437636137, + 0.007732469588518143, + 0.0067498888820409775, + 0.007388893980532885, + 0.006280044559389353, + 0.024081513285636902, + 0.01000909972935915, + 0.006357430014759302, + 0.01145815197378397, + 0.020369676873087883, + 0.008139707148075104, + 0.005452246405184269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.10029835253953934 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628121413290501 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007628121413290501, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628121413290501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12057632952928543, + 0.09581777453422546, + 0.061617448925971985, + 0.06093320995569229, + 0.03833472728729248, + 0.04700934886932373, + 0.12643955647945404, + 0.06008579209446907, + 0.122276671230793, + 0.266909122467041 + ], + "credit_map": [ + 0.016631847247481346, + 0.015054743736982346, + 0.0149537967517972, + 0.017128605395555496, + 0.016101965680718422, + 0.016158100217580795, + 0.012525912374258041, + 0.012021349743008614, + 0.010250503197312355, + 0.007457325235009193, + 0.022843381389975548, + 0.007641871925443411, + 0.004259512294083834, + 0.019256625324487686, + 0.005207403097301722, + 0.018901152536273003, + 0.009705424308776855, + 0.00936401728540659, + 0.008215886540710926, + 0.005391975864768028, + 0.004203233402222395, + 0.008412757888436317, + 0.005723198875784874, + 0.010600954294204712, + 0.0063126105815172195, + 0.014806528575718403, + 0.009096194989979267, + 0.006717185024172068, + 0.005230789538472891, + 0.007188241928815842, + 0.005378103815019131, + 0.006203558295965195, + 0.0039030874613672495, + 0.005261410027742386, + 0.00586468493565917, + 0.006317972205579281, + 0.004576776176691055, + 0.005241960287094116, + 0.004412917885929346, + 0.002755917841568589, + 0.00483696348965168, + 0.0045912498608231544, + 0.010978743433952332, + 0.003891869680956006, + 0.005389787256717682, + 0.008666803129017353, + 0.004019855987280607, + 0.004634075332432985, + 0.013533179648220539, + 0.019120322540402412, + 0.01606440730392933, + 0.02514108642935753, + 0.012734823860228062, + 0.011396384797990322, + 0.01181633211672306, + 0.01663302443921566, + 0.007072704378515482, + 0.00646578473970294, + 0.00930633395910263, + 0.007999398745596409, + 0.007831726223230362, + 0.0075520058162510395, + 0.0064162141643464565, + 0.0074416231364011765, + 0.010850579477846622, + 0.012219392694532871, + 0.019567469134926796, + 0.023459063842892647, + 0.013537534512579441, + 0.010589081794023514, + 0.01427400577813387, + 0.017779545858502388, + 0.014163180254399776, + 0.057923324406147, + 0.028158297762274742, + 0.036811064928770065, + 0.03311562165617943, + 0.050881341099739075, + 0.021130627021193504, + 0.024725671857595444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.007628121413290501 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective depicting the viewer as a spotted moth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a sun-drenched meadow with emerald green grass and vibrant wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a target Pieris butterfly with delicate, powdery white wings visible ahead in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit erratic, frantic motion including darting, swerving, plunging downward, and soaring upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style painterly with bleeding colors and abstract streaks of crimson, gold, and azure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016060039401054382 + ], + "opposite_logprob_tokens": [ + -4.141059875488281 + ], + "logprob_reward": -0.016060039401054382, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016060039401054382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02179037034511566, + 0.09768778830766678, + 0.014549260027706623, + 0.14448323845863342, + 0.08110683411359787, + 0.04511667788028717, + 0.023482849821448326, + 0.11907046288251877, + 0.08706512302160263, + 0.36564746499061584 + ], + "credit_map": [ + 0.002297252183780074, + 0.002578440820798278, + 0.002586586633697152, + 0.003275789087638259, + 0.002227791352197528, + 0.0037640214432030916, + 0.0030453538056463003, + 0.002015133621171117, + 0.004861812107264996, + 0.003446822054684162, + 0.00798899307847023, + 0.0024865432642400265, + 0.0023680112790316343, + 0.03771764785051346, + 0.002477101981639862, + 0.036340851336717606, + 0.0018743716645985842, + 0.0020297777373343706, + 0.002008825773373246, + 0.0014561719726771116, + 0.0015190575504675508, + 0.0019749340135604143, + 0.0015198069158941507, + 0.002166314050555229, + 0.016492294147610664, + 0.038402434438467026, + 0.023404063656926155, + 0.013163452968001366, + 0.010997717268764973, + 0.016059063374996185, + 0.010650883428752422, + 0.015313316136598587, + 0.0030890461057424545, + 0.0032695101108402014, + 0.004429783672094345, + 0.002524404786527157, + 0.006320686545222998, + 0.03252281993627548, + 0.02331630140542984, + 0.005634274333715439, + 0.002913541393354535, + 0.0048729851841926575, + 0.0054517691023647785, + 0.00403817230835557, + 0.005168629344552755, + 0.012352202087640762, + 0.0026417712215334177, + 0.007677603513002396, + 0.0033379956148564816, + 0.003285655053332448, + 0.00322476658038795, + 0.003984888084232807, + 0.0020834801252931356, + 0.0017079528188332915, + 0.0023981095291674137, + 0.0034600021317601204, + 0.014313490130007267, + 0.015416235662996769, + 0.016097145155072212, + 0.020028142258524895, + 0.010996968485414982, + 0.00828887615352869, + 0.019247043877840042, + 0.014682553708553314, + 0.007858486846089363, + 0.008678237907588482, + 0.015200006775557995, + 0.011491023935377598, + 0.01415942795574665, + 0.00822213664650917, + 0.011015559546649456, + 0.010440236888825893, + 0.019032444804906845, + 0.12070312350988388, + 0.02239389158785343, + 0.015926474705338478, + 0.029350819066166878, + 0.1306629776954651, + 0.010233896784484386, + 0.017343830317258835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 7 + ], + "selected_reward": -0.016060039401054382 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_13/index_tensor([4191, 4191])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_13/index_tensor([4191, 4191])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..13e0984943c7c3ae50a19f1dcf65087205fcb4bc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_13/index_tensor([4191, 4191])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a level angle, pans slowly to the right, following an elderly Asian man. He is clad in a traditional white silk Tai Chi suit, his silvery hair neatly tied back. Deep wrinkles etch the quiet of age. Amidst the thick morning mist, he slowly practices Tai Chi, his arms moving in a steady, deliberate rhythm, one hand rising, the other gently lowering, as if parting the mist before him. The entire scene is a high-contrast black and white image, the soft, diffused morning light filtering through the mist, delicately capturing not only his relaxed posture but also the flowing folds and textures of the silk garment. In the background, the silhouettes of several ancient trees loom through the mist, creating an atmosphere of tranquility, concentration, and timeless poetry.", + "seed": 19793, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_13/index_4191_seed_19793_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10851609455421567, + "vlm_logprob": -0.10851609455421567, + "vlm_yes_no_margin": 3.999999976158142, + "vlm_token_credit": -0.10851609455421567, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004543695133179426 + ], + "opposite_logprob_tokens": [ + -6.379543781280518 + ], + "logprob_reward": -0.004543695133179426, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.004543695133179426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18100279569625854, + 0.08200625330209732, + 0.39837098121643066, + 0.06526336073875427, + 0.0547158308327198, + 0.03201081231236458, + 0.03551208972930908, + 0.045024048537015915, + 0.05089321359992027, + 0.05520059913396835 + ], + "credit_map": [ + 0.017632583156228065, + 0.04639085382223129, + 0.027444200590252876, + 0.01913243718445301, + 0.03111778013408184, + 0.01756519079208374, + 0.01066005602478981, + 0.01105969212949276, + 0.01151536125689745, + 0.009843532927334309, + 0.011488331481814384, + 0.015588797628879547, + 0.011796490289270878, + 0.00900864228606224, + 0.006496932823210955, + 0.006268166471272707, + 0.01957622729241848, + 0.15151920914649963, + 0.16886672377586365, + 0.01944854110479355, + 0.015979591757059097, + 0.009071447886526585, + 0.005744253285229206, + 0.008164985105395317, + 0.017593123018741608, + 0.008601509034633636, + 0.008276973851025105, + 0.007302614860236645, + 0.005074813961982727, + 0.006897171959280968, + 0.0056276158429682255, + 0.005889535415917635, + 0.006134326569736004, + 0.00486249104142189, + 0.01244177483022213, + 0.006221553310751915, + 0.0054922932758927345, + 0.005737983621656895, + 0.004967799875885248, + 0.0088576078414917, + 0.0031792439986020327, + 0.009304656647145748, + 0.0023704166524112225, + 0.0029535305220633745, + 0.003693686332553625, + 0.004028918221592903, + 0.002554182196035981, + 0.0039261761121451855, + 0.007443400099873543, + 0.005628447514027357, + 0.005018734373152256, + 0.002950266469269991, + 0.0039048688486218452, + 0.004970388486981392, + 0.002889770781621337, + 0.0027062122244387865, + 0.007756236474961042, + 0.006011703051626682, + 0.004723263904452324, + 0.003400962334126234, + 0.008823254145681858, + 0.0057744113728404045, + 0.003512965515255928, + 0.0050212545320391655, + 0.012453263625502586, + 0.0057548703625798225, + 0.004547833930701017, + 0.005323000252246857, + 0.004308072384446859, + 0.00646941876038909, + 0.004772720392793417, + 0.007264033891260624, + 0.007964292541146278, + 0.007367430720478296, + 0.008622345514595509, + 0.004812340252101421, + 0.00855906493961811, + 0.007066988851875067, + 0.005019842647016048, + 0.005788295529782772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.004543695133179426 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004102980252355337 + ], + "opposite_logprob_tokens": [ + -5.504103183746338 + ], + "logprob_reward": -0.004102980252355337, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004102980252355337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16148075461387634, + 0.053675368428230286, + 0.1652531623840332, + 0.06909813731908798, + 0.03561464324593544, + 0.037105608731508255, + 0.06501675397157669, + 0.11406733095645905, + 0.23169006407260895, + 0.06699822843074799 + ], + "credit_map": [ + 0.012889550067484379, + 0.027477821335196495, + 0.016528509557247162, + 0.02106139436364174, + 0.029984256252646446, + 0.020443255081772804, + 0.015874261036515236, + 0.017221709713339806, + 0.0075957924127578735, + 0.0072423312813043594, + 0.007806997746229172, + 0.006308125331997871, + 0.006719448138028383, + 0.007486192509531975, + 0.004381333012133837, + 0.006135147996246815, + 0.008626561611890793, + 0.06133463606238365, + 0.0663067176938057, + 0.008419246412813663, + 0.008206994272768497, + 0.005239293910562992, + 0.002835768274962902, + 0.0042839450761675835, + 0.012755378149449825, + 0.012316574342548847, + 0.010283761657774448, + 0.00724718626588583, + 0.006702035665512085, + 0.006561574060469866, + 0.005773292854428291, + 0.007458333391696215, + 0.004966771695762873, + 0.004153949674218893, + 0.004086865112185478, + 0.0029843728989362717, + 0.004628709051758051, + 0.004826806951314211, + 0.0033186855725944042, + 0.006648484617471695, + 0.004577695857733488, + 0.006144546438008547, + 0.00287831574678421, + 0.003601465839892626, + 0.005890384316444397, + 0.005242691840976477, + 0.003412902355194092, + 0.005357605405151844, + 0.008869628421962261, + 0.012039495632052422, + 0.007680127862840891, + 0.005816714838147163, + 0.010264292359352112, + 0.007144180592149496, + 0.007012657821178436, + 0.006189659237861633, + 0.011804851703345776, + 0.01440061628818512, + 0.011455995962023735, + 0.011055215261876583, + 0.02087068185210228, + 0.014659516513347626, + 0.011834361590445042, + 0.01798609085381031, + 0.05948778614401817, + 0.02622702717781067, + 0.02171492949128151, + 0.02507869340479374, + 0.021024828776717186, + 0.02892349846661091, + 0.019567692652344704, + 0.029665596783161163, + 0.006669369991868734, + 0.00897190347313881, + 0.00892691221088171, + 0.006889874581247568, + 0.013806055299937725, + 0.007012761663645506, + 0.005660335998982191, + 0.009061011485755444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.004102980252355337 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870384395122528 + ], + "opposite_logprob_tokens": [ + -1.1370384693145752 + ], + "logprob_reward": -0.3870384395122528, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870384395122528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0750073790550232, + 0.06273650377988815, + 0.5826112627983093, + 0.12495481222867966, + 0.06045074015855789, + 0.018879275768995285, + 0.015154357068240643, + 0.029956283047795296, + 0.01709500327706337, + 0.013154394924640656 + ], + "credit_map": [ + 0.007168762851506472, + 0.01505111251026392, + 0.009815460070967674, + 0.012337408028542995, + 0.00956958532333374, + 0.007934621535241604, + 0.007190969306975603, + 0.005939456634223461, + 0.006345868576318026, + 0.013566982001066208, + 0.019204596057534218, + 0.008193283341825008, + 0.0034828048665076494, + 0.004175538197159767, + 0.003421725705265999, + 0.004345698747783899, + 0.028052862733602524, + 0.225617453455925, + 0.24879436194896698, + 0.027630595490336418, + 0.022790851071476936, + 0.010837937705218792, + 0.007615505252033472, + 0.011271676979959011, + 0.03327981382608414, + 0.017730483785271645, + 0.01762985810637474, + 0.015307881869375706, + 0.008697550743818283, + 0.012568674981594086, + 0.011461748741567135, + 0.00827878713607788, + 0.0061249504797160625, + 0.005643008276820183, + 0.013104278594255447, + 0.007201773580163717, + 0.005868666805326939, + 0.007168017327785492, + 0.005026496015489101, + 0.010313543491065502, + 0.0024475501850247383, + 0.003535393625497818, + 0.001107856398448348, + 0.00119971320964396, + 0.003147605573758483, + 0.0033732277806848288, + 0.0014506475999951363, + 0.002617279766127467, + 0.0023319560568779707, + 0.0026080538518726826, + 0.002106009516865015, + 0.001417776569724083, + 0.0019732017535716295, + 0.0015285165281966329, + 0.001659632078371942, + 0.0015292085008695722, + 0.00716041075065732, + 0.00391170522198081, + 0.0026816613972187042, + 0.0027073875535279512, + 0.003073848318308592, + 0.0044387648813426495, + 0.0019419246818870306, + 0.004040576983243227, + 0.004362807143479586, + 0.0020432148594409227, + 0.0012177907628938556, + 0.001963851973414421, + 0.0015003641601651907, + 0.00217164377681911, + 0.0015253492165356874, + 0.0023099789395928383, + 0.002904290799051523, + 0.0017102172132581472, + 0.0011744173243641853, + 0.0011629273649305105, + 0.001756959012709558, + 0.002055743709206581, + 0.0011398643255233765, + 0.0012499740114435554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.3870384395122528 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428004503250122 + ], + "opposite_logprob_tokens": [ + -2.0178003311157227 + ], + "logprob_reward": -0.1428004503250122, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1428004503250122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19483166933059692, + 0.0552637092769146, + 0.26410984992980957, + 0.19003833830356598, + 0.060510896146297455, + 0.027650119736790657, + 0.015367243438959122, + 0.02821245975792408, + 0.124383844435215, + 0.03963190317153931 + ], + "credit_map": [ + 0.017676828429102898, + 0.03657061979174614, + 0.036468952894210815, + 0.023936418816447258, + 0.02907482534646988, + 0.017153754830360413, + 0.017867334187030792, + 0.01608293317258358, + 0.005514092743396759, + 0.01148691214621067, + 0.01733452081680298, + 0.007449110969901085, + 0.0037446364294737577, + 0.003334370907396078, + 0.0030437603127211332, + 0.003356303321197629, + 0.012408954091370106, + 0.1010603979229927, + 0.10848511010408401, + 0.012179468758404255, + 0.011762966401875019, + 0.00694891856983304, + 0.0047583081759512424, + 0.006505731027573347, + 0.03460300713777542, + 0.03102736733853817, + 0.031171806156635284, + 0.019435005262494087, + 0.019400373101234436, + 0.017731575295329094, + 0.018675360828638077, + 0.017993826419115067, + 0.005900562275201082, + 0.0059463451616466045, + 0.010901149362325668, + 0.005606661085039377, + 0.006949702277779579, + 0.00486776465550065, + 0.00622038496658206, + 0.014118327759206295, + 0.0031390355434268713, + 0.006652351934462786, + 0.0022419635206460953, + 0.0024090169463306665, + 0.0036509681958705187, + 0.0042348382994532585, + 0.0024000145494937897, + 0.0029219305142760277, + 0.0025405422784388065, + 0.0026314908172935247, + 0.0021170438267290592, + 0.001326154451817274, + 0.002145081292837858, + 0.0017454493790864944, + 0.0014700480969622731, + 0.0013914344599470496, + 0.004599910229444504, + 0.0034756131935864687, + 0.0036959368735551834, + 0.002517115557566285, + 0.00445682555437088, + 0.003669080324470997, + 0.0025933468714356422, + 0.003204630920663476, + 0.03663582354784012, + 0.01428391132503748, + 0.012105105444788933, + 0.011548832058906555, + 0.011033263057470322, + 0.01655392535030842, + 0.00867437943816185, + 0.013548607006669044, + 0.009402476251125336, + 0.005598095711320639, + 0.004971031565219164, + 0.0037878891453146935, + 0.004955573007464409, + 0.005051040556281805, + 0.002668572124093771, + 0.0031972266733646393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.1428004503250122 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409490754827857 + ], + "opposite_logprob_tokens": [ + -5.504095077514648 + ], + "logprob_reward": -0.00409490754827857, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409490754827857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15690244734287262, + 0.10021795332431793, + 0.11459320038557053, + 0.2446671724319458, + 0.04170415550470352, + 0.16994346678256989, + 0.027229316532611847, + 0.06593005359172821, + 0.05190486088395119, + 0.02690739743411541 + ], + "credit_map": [ + 0.010989873670041561, + 0.03557869419455528, + 0.030616432428359985, + 0.015601021237671375, + 0.022723359987139702, + 0.016557445749640465, + 0.013642889447510242, + 0.011192725971341133, + 0.015620248392224312, + 0.0195170771330595, + 0.021593986079096794, + 0.016845891252160072, + 0.007477083709090948, + 0.00681277085095644, + 0.005665699951350689, + 0.0066851964220404625, + 0.008835247717797756, + 0.03894234821200371, + 0.046239595860242844, + 0.0069681620225310326, + 0.0045453342609107494, + 0.00356490071862936, + 0.0023799787741154432, + 0.0031176316551864147, + 0.06832369416952133, + 0.035709913820028305, + 0.03280149772763252, + 0.0290075596421957, + 0.017302928492426872, + 0.02556702308356762, + 0.020525503903627396, + 0.015429036691784859, + 0.003654420841485262, + 0.005042536649852991, + 0.008701436221599579, + 0.005212949588894844, + 0.00465324567630887, + 0.0038801045157015324, + 0.0038109305314719677, + 0.006748529616743326, + 0.025073714554309845, + 0.01577584818005562, + 0.009820349514484406, + 0.01104569248855114, + 0.03150716796517372, + 0.03990938514471054, + 0.011782984249293804, + 0.02502834051847458, + 0.006487360689789057, + 0.0037606526166200638, + 0.0028835611883550882, + 0.0023841061629354954, + 0.00348443235270679, + 0.0037643834948539734, + 0.001827170723117888, + 0.0026376480236649513, + 0.012804262340068817, + 0.009558422490954399, + 0.005577325355261564, + 0.006322485860437155, + 0.009599328972399235, + 0.008105680346488953, + 0.005026007536798716, + 0.008936538361012936, + 0.016953175887465477, + 0.006883575115352869, + 0.0059797256253659725, + 0.004549505654722452, + 0.00443890318274498, + 0.007468536961823702, + 0.0028520035557448864, + 0.002779433038085699, + 0.005037104710936546, + 0.0037162350490689278, + 0.0033660870976746082, + 0.0021725355181843042, + 0.003720991313457489, + 0.003905756399035454, + 0.00241146725602448, + 0.002577219856902957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.00409490754827857 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, pans slowly to the right, following an elderly Asian man. He is clad in a traditional white silk Tai Chi suit, his silvery hair neatly tied back. Deep wrinkles etch the quiet of age. Amidst the thick morning mist, he slowly practices Tai Chi, his arms moving in a steady, deliberate rhythm, one hand rising, the other gently lowering, as if parting the mist before him. The entire scene is a high-contrast black and white image, the soft, diffused morning light filtering through the mist, delicately capturing not only his relaxed posture but also the flowing folds and textures of the silk garment. In the background, the silhouettes of several ancient trees loom through the mist, creating an atmosphere of tranquility, concentration, and timeless poetry.", + "seed": 19794, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_13/index_4191_seed_19794_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.20367191843688487, + "vlm_logprob": -0.20367191843688487, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.20367191843688487, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01360454224050045 + ], + "opposite_logprob_tokens": [ + -4.888604640960693 + ], + "logprob_reward": -0.01360454224050045, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01360454224050045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17543505132198334, + 0.09070650488138199, + 0.2586338222026825, + 0.14302872121334076, + 0.06338653713464737, + 0.03189519792795181, + 0.04107239097356796, + 0.07897676527500153, + 0.05502203479409218, + 0.061842940747737885 + ], + "credit_map": [ + 0.02113291621208191, + 0.023709438741207123, + 0.01739596575498581, + 0.030328886583447456, + 0.026242466643452644, + 0.019088152796030045, + 0.022789588198065758, + 0.014747623354196548, + 0.013695022091269493, + 0.009775293990969658, + 0.008857419714331627, + 0.016857968643307686, + 0.011840778402984142, + 0.011786186136305332, + 0.007410597521811724, + 0.010483236983418465, + 0.03133031725883484, + 0.032251011580228806, + 0.036934785544872284, + 0.032096587121486664, + 0.02893734723329544, + 0.0600288063287735, + 0.01779359020292759, + 0.019261371344327927, + 0.019016539677977562, + 0.028589721769094467, + 0.020035669207572937, + 0.010217958129942417, + 0.029086392372846603, + 0.013417295180261135, + 0.012255451641976833, + 0.010409679263830185, + 0.008996417745947838, + 0.011861051432788372, + 0.00810546986758709, + 0.006786983460187912, + 0.009235906414687634, + 0.006307700183242559, + 0.00616163806989789, + 0.0059313690289855, + 0.005439204163849354, + 0.004441171418875456, + 0.003540441859513521, + 0.0025407676585018635, + 0.00372955109924078, + 0.002575856866315007, + 0.004231944214552641, + 0.005396257620304823, + 0.005730660632252693, + 0.00463649770244956, + 0.0062509640119969845, + 0.004230877384543419, + 0.004150482825934887, + 0.0036338013596832752, + 0.005471669603139162, + 0.006967439316213131, + 0.008108011446893215, + 0.014408779330551624, + 0.010210282169282436, + 0.008486050181090832, + 0.010220920667052269, + 0.012281719595193863, + 0.008408783935010433, + 0.006852212827652693, + 0.006342723034322262, + 0.008580226451158524, + 0.0067370617762207985, + 0.007919649593532085, + 0.009165787138044834, + 0.006410396657884121, + 0.004702883772552013, + 0.005163305904716253, + 0.005587372463196516, + 0.012665710411965847, + 0.007656849455088377, + 0.007160681299865246, + 0.01116837840527296, + 0.007640308700501919, + 0.005434721242636442, + 0.004528918769210577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.01360454224050045 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330414578318596 + ], + "opposite_logprob_tokens": [ + -3.7733042240142822 + ], + "logprob_reward": -0.02330414578318596, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330414578318596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20398835837841034, + 0.07299285382032394, + 0.2328447550535202, + 0.09502595663070679, + 0.07996003329753876, + 0.02956026792526245, + 0.04202497750520706, + 0.08537834137678146, + 0.07436022162437439, + 0.0838642492890358 + ], + "credit_map": [ + 0.024020565673708916, + 0.02621772699058056, + 0.01638517901301384, + 0.03362686187028885, + 0.031225811690092087, + 0.025336146354675293, + 0.025283154100179672, + 0.02189292572438717, + 0.009051734581589699, + 0.007272606715559959, + 0.0074808611534535885, + 0.012464405037462711, + 0.011390794068574905, + 0.009876049123704433, + 0.007417823188006878, + 0.008038574829697609, + 0.03070468083024025, + 0.030281873419880867, + 0.03571322560310364, + 0.028291527181863785, + 0.023714160546660423, + 0.04919210448861122, + 0.01674342155456543, + 0.018203770741820335, + 0.01181713119149208, + 0.017666557803750038, + 0.015588279813528061, + 0.007869352586567402, + 0.016915667802095413, + 0.01049889624118805, + 0.007654264103621244, + 0.007015803828835487, + 0.008184688165783882, + 0.013546131551265717, + 0.007867294363677502, + 0.012825283221900463, + 0.011426925659179688, + 0.006521773524582386, + 0.011057667434215546, + 0.008530274964869022, + 0.0038349556270986795, + 0.0035654425155371428, + 0.003419789019972086, + 0.003968095872551203, + 0.0034380198922008276, + 0.002956612268462777, + 0.00484215933829546, + 0.0035351919941604137, + 0.005035082809627056, + 0.005628265906125307, + 0.006832104176282883, + 0.006331312470138073, + 0.004487067926675081, + 0.0032551558688282967, + 0.005842459388077259, + 0.00461353175342083, + 0.010107793845236301, + 0.018886657431721687, + 0.009689155034720898, + 0.015238489955663681, + 0.00983267743140459, + 0.007302515208721161, + 0.007345985155552626, + 0.006975066382437944, + 0.007163987960666418, + 0.010759394615888596, + 0.007047570776194334, + 0.009552261792123318, + 0.011668913997709751, + 0.0074194022454321384, + 0.011746951378881931, + 0.009001740254461765, + 0.006002894137054682, + 0.01018674299120903, + 0.012890586629509926, + 0.011415877379477024, + 0.010782279074192047, + 0.007414356339722872, + 0.016571592539548874, + 0.008599916473031044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.02330414578318596 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933822631835938 + ], + "opposite_logprob_tokens": [ + -0.6933822631835938 + ], + "logprob_reward": -0.6933822631835938, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933822631835938, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14325875043869019, + 0.07598432898521423, + 0.13316583633422852, + 0.22053919732570648, + 0.06763701140880585, + 0.048154596239328384, + 0.21353842318058014, + 0.03674919158220291, + 0.022885577753186226, + 0.03808711841702461 + ], + "credit_map": [ + 0.021894901990890503, + 0.025430982932448387, + 0.01296499278396368, + 0.020245950669050217, + 0.024314776062965393, + 0.012852192856371403, + 0.010327348485589027, + 0.015227590687572956, + 0.0115116061642766, + 0.01098734699189663, + 0.007405457552522421, + 0.013762296177446842, + 0.012373953126370907, + 0.0073010618798434734, + 0.006107711233198643, + 0.006534891668707132, + 0.017769699916243553, + 0.01956544630229473, + 0.017856847494840622, + 0.02450614608824253, + 0.013857982121407986, + 0.018873708322644234, + 0.009695949032902718, + 0.011040049605071545, + 0.036286141723394394, + 0.05985827371478081, + 0.03565320745110512, + 0.025417516008019447, + 0.017814090475440025, + 0.01908191666007042, + 0.012494237162172794, + 0.013933796435594559, + 0.009636283852159977, + 0.013873610645532608, + 0.009928185492753983, + 0.00811248179525137, + 0.006981205195188522, + 0.008711488917469978, + 0.005189259070903063, + 0.005204493645578623, + 0.00804623682051897, + 0.006363422144204378, + 0.006542847491800785, + 0.005037254653871059, + 0.004062664695084095, + 0.003502293024212122, + 0.0054704719223082066, + 0.009129400365054607, + 0.029284780845046043, + 0.02630414068698883, + 0.034354791045188904, + 0.029144853353500366, + 0.014589310623705387, + 0.015098802745342255, + 0.02106228470802307, + 0.04369945824146271, + 0.0049436055123806, + 0.006461645010858774, + 0.008801551535725594, + 0.004797883797436953, + 0.003928696736693382, + 0.0030211606062948704, + 0.0024924706667661667, + 0.0023021770175546408, + 0.0030532360542565584, + 0.0052269515581429005, + 0.003134563798084855, + 0.002665315056219697, + 0.002333607291802764, + 0.002387169050052762, + 0.002149702748283744, + 0.0019350310321897268, + 0.002520852955058217, + 0.0036071341019123793, + 0.004886471200734377, + 0.0055047315545380116, + 0.005393242929130793, + 0.0035681212320923805, + 0.00874236598610878, + 0.0038641917053610086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.6933822631835938 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28133079409599304 + ], + "opposite_logprob_tokens": [ + -1.4063308238983154 + ], + "logprob_reward": -0.28133079409599304, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28133079409599304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26534152030944824, + 0.14191533625125885, + 0.09324228018522263, + 0.20549596846103668, + 0.04900457710027695, + 0.031232589855790138, + 0.04686479642987251, + 0.07445232570171356, + 0.036219798028469086, + 0.05623084306716919 + ], + "credit_map": [ + 0.034076057374477386, + 0.03142004832625389, + 0.026357438415288925, + 0.036007773131132126, + 0.03938521444797516, + 0.03791939839720726, + 0.03115549497306347, + 0.029020113870501518, + 0.021488456055521965, + 0.016156351193785667, + 0.010984535329043865, + 0.027111191302537918, + 0.02565595507621765, + 0.017980001866817474, + 0.012671485543251038, + 0.00986736360937357, + 0.010356762446463108, + 0.013141480274498463, + 0.019750408828258514, + 0.010564385913312435, + 0.007941462099552155, + 0.01766332797706127, + 0.006783766206353903, + 0.007040681317448616, + 0.037338439375162125, + 0.0486367866396904, + 0.026737790554761887, + 0.026246070861816406, + 0.017638899385929108, + 0.020218005403876305, + 0.011607903055846691, + 0.01707206480205059, + 0.007342759519815445, + 0.010088963434100151, + 0.010295651853084564, + 0.004529590252786875, + 0.005808875430375338, + 0.0038671058136969805, + 0.0042663272470235825, + 0.0028053023852407932, + 0.004485506564378738, + 0.0037545172963291407, + 0.003331350162625313, + 0.004166665021330118, + 0.003783075138926506, + 0.0029303166083991528, + 0.004696324467658997, + 0.004084836691617966, + 0.00485466793179512, + 0.005808119662106037, + 0.007784299086779356, + 0.006740252021700144, + 0.00588166294619441, + 0.004127855412662029, + 0.006690880749374628, + 0.004977062810212374, + 0.008646699599921703, + 0.017680050805211067, + 0.009277437813580036, + 0.013898000121116638, + 0.00791629683226347, + 0.00660963449627161, + 0.004994016606360674, + 0.005430190823972225, + 0.005230011884123087, + 0.007689184509217739, + 0.0050789471715688705, + 0.0033071821089833975, + 0.005025032442063093, + 0.002742695389315486, + 0.004028611350804567, + 0.0031181322410702705, + 0.003972400911152363, + 0.011008462868630886, + 0.007190140429884195, + 0.0066008735448122025, + 0.0065103983506560326, + 0.006569062825292349, + 0.007708085235208273, + 0.006671420764178038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.28133079409599304 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737846881151199 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.006737846881151199, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737846881151199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24035394191741943, + 0.062004413455724716, + 0.10674963891506195, + 0.07345327734947205, + 0.36281511187553406, + 0.014929594472050667, + 0.04379827901721001, + 0.056125007569789886, + 0.023544367402791977, + 0.01622634194791317 + ], + "credit_map": [ + 0.03095514327287674, + 0.03287524729967117, + 0.01759447157382965, + 0.0438443087041378, + 0.03815937042236328, + 0.027561938390135765, + 0.02611742913722992, + 0.02324603497982025, + 0.008571188896894455, + 0.0075736078433692455, + 0.008876132778823376, + 0.014996718615293503, + 0.008747194893658161, + 0.005122683942317963, + 0.004430114291608334, + 0.0036867735907435417, + 0.01283885445445776, + 0.016346177086234093, + 0.018296783789992332, + 0.016598833724856377, + 0.010042107664048672, + 0.016764391213655472, + 0.007719977758824825, + 0.008142520673573017, + 0.008160910569131374, + 0.0137000922113657, + 0.015062735415995121, + 0.006676344200968742, + 0.007927164435386658, + 0.010410272516310215, + 0.00586333405226469, + 0.005652428139001131, + 0.06119048595428467, + 0.07620413601398468, + 0.0691840648651123, + 0.028947370126843452, + 0.029155727475881577, + 0.03955606371164322, + 0.026712536811828613, + 0.0318647138774395, + 0.0025347373448312283, + 0.0016674601938575506, + 0.0017861301312223077, + 0.0014661732129752636, + 0.0014671692624688148, + 0.0011831509182229638, + 0.0019078314071521163, + 0.0029169421177357435, + 0.0071438089944422245, + 0.005841485690325499, + 0.007646698039025068, + 0.004694557283073664, + 0.0036329380236566067, + 0.0035512810572981834, + 0.004072162788361311, + 0.007215347606688738, + 0.006770269013941288, + 0.015396376140415668, + 0.006172191351652145, + 0.010932227596640587, + 0.005722265690565109, + 0.0038272372912615538, + 0.003499875543639064, + 0.003804564941674471, + 0.0029880753718316555, + 0.0049569569528102875, + 0.0033837584778666496, + 0.0032035689800977707, + 0.002813606057316065, + 0.0020328499376773834, + 0.00214369292370975, + 0.002021859399974346, + 0.0015389003092423081, + 0.002057853154838085, + 0.0016060180496424437, + 0.0022044898942112923, + 0.00225788913667202, + 0.0016993526369333267, + 0.003519827499985695, + 0.0013420104514807463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.006737846881151199 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, pans slowly to the right, following an elderly Asian man. He is clad in a traditional white silk Tai Chi suit, his silvery hair neatly tied back. Deep wrinkles etch the quiet of age. Amidst the thick morning mist, he slowly practices Tai Chi, his arms moving in a steady, deliberate rhythm, one hand rising, the other gently lowering, as if parting the mist before him. The entire scene is a high-contrast black and white image, the soft, diffused morning light filtering through the mist, delicately capturing not only his relaxed posture but also the flowing folds and textures of the silk garment. In the background, the silhouettes of several ancient trees loom through the mist, creating an atmosphere of tranquility, concentration, and timeless poetry.", + "seed": 19795, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_13/index_4191_seed_19795_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07687037233263254, + "vlm_logprob": -0.07687037233263254, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.07687037233263254, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00958401057869196 + ], + "opposite_logprob_tokens": [ + -5.634583950042725 + ], + "logprob_reward": -0.00958401057869196, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00958401057869196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07436734437942505, + 0.1464855968952179, + 0.1801391839981079, + 0.04352207109332085, + 0.04445129632949829, + 0.27900266647338867, + 0.03242655098438263, + 0.03547946363687515, + 0.1367194503545761, + 0.02740633487701416 + ], + "credit_map": [ + 0.010230651125311852, + 0.005091494414955378, + 0.006706852465867996, + 0.006681155413389206, + 0.011349290609359741, + 0.008499602787196636, + 0.021552257239818573, + 0.004256042186170816, + 0.02850010246038437, + 0.008782234974205494, + 0.00760116521269083, + 0.007921638898551464, + 0.015861384570598602, + 0.042369548231363297, + 0.0268670916557312, + 0.008582430891692638, + 0.018361568450927734, + 0.02255723811686039, + 0.012690690346062183, + 0.017706912010908127, + 0.037830885499715805, + 0.01418054848909378, + 0.036290086805820465, + 0.0205212589353323, + 0.0038038960192352533, + 0.007805458270013332, + 0.004992038011550903, + 0.005651470273733139, + 0.005680863279849291, + 0.005650060251355171, + 0.004943892825394869, + 0.004994392395019531, + 0.006511211860924959, + 0.0064265066757798195, + 0.010783299803733826, + 0.007959669455885887, + 0.004096399527043104, + 0.0021368402522057295, + 0.004195083864033222, + 0.0023422883823513985, + 0.024888960644602776, + 0.0968799740076065, + 0.1087232232093811, + 0.013119147159159184, + 0.015818068757653236, + 0.0035630986094474792, + 0.005822362843900919, + 0.010187857784330845, + 0.0023683826439082623, + 0.009837930090725422, + 0.005616405047476292, + 0.0031243707053363323, + 0.0037627900019288063, + 0.003327813930809498, + 0.002585040871053934, + 0.0018038204871118069, + 0.004958320874720812, + 0.006506291683763266, + 0.003977511078119278, + 0.005520176142454147, + 0.003769060131162405, + 0.004347341135144234, + 0.0036945873871445656, + 0.002706177532672882, + 0.05022933706641197, + 0.019230062142014503, + 0.037640202790498734, + 0.004545577336102724, + 0.003979528322815895, + 0.00957508198916912, + 0.009112638421356678, + 0.002407030900940299, + 0.004421798046678305, + 0.002834482118487358, + 0.0037851964589208364, + 0.0027615963481366634, + 0.00420589093118906, + 0.0028129559941589832, + 0.0034718059469014406, + 0.003112612757831812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.00958401057869196 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108936034142971 + ], + "opposite_logprob_tokens": [ + -4.511089324951172 + ], + "logprob_reward": -0.01108936034142971, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108936034142971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04656707122921944, + 0.2833331227302551, + 0.2377416491508484, + 0.016286594793200493, + 0.02531493827700615, + 0.2380320131778717, + 0.031940706074237823, + 0.03653109446167946, + 0.06565750390291214, + 0.01859527826309204 + ], + "credit_map": [ + 0.005269467830657959, + 0.005952044855803251, + 0.0034766753669828176, + 0.0058388118632137775, + 0.010137195698916912, + 0.005325292702764273, + 0.005596974398940802, + 0.004970608279109001, + 0.05139724537730217, + 0.010527992621064186, + 0.012876277789473534, + 0.017178257927298546, + 0.030195649713277817, + 0.07644781470298767, + 0.07128529995679855, + 0.01342458464205265, + 0.01799386367201805, + 0.035059355199337006, + 0.015149334445595741, + 0.023224812000989914, + 0.05488786846399307, + 0.019207287579774857, + 0.04515457525849342, + 0.027064543217420578, + 0.001669839839451015, + 0.0022924980148673058, + 0.0019169182050973177, + 0.0014073372585698962, + 0.0025250837206840515, + 0.0027399121318012476, + 0.0018113208934664726, + 0.0019236858934164047, + 0.0025887098163366318, + 0.005493282340466976, + 0.0031778228003531694, + 0.0036163516342639923, + 0.004154399037361145, + 0.002412792295217514, + 0.001965938601642847, + 0.0019056434975937009, + 0.021558988839387894, + 0.08254486322402954, + 0.09476681798696518, + 0.00945747084915638, + 0.012290059588849545, + 0.0035480635706335306, + 0.004963851533830166, + 0.00890190340578556, + 0.0025120351929217577, + 0.008573012426495552, + 0.006153109483420849, + 0.0019567261915653944, + 0.0027625588700175285, + 0.004826409742236137, + 0.002798050409182906, + 0.002358802827075124, + 0.004377260338515043, + 0.005759605206549168, + 0.004272432066500187, + 0.003740271320566535, + 0.006449609994888306, + 0.004068648908287287, + 0.004897493403404951, + 0.0029657708946615458, + 0.023260846734046936, + 0.008231823332607746, + 0.01883956603705883, + 0.002289436524733901, + 0.002295897575095296, + 0.0046044010668993, + 0.0047566257417202, + 0.0013789063086733222, + 0.00360669381916523, + 0.0017186482436954975, + 0.0028266077861189842, + 0.0017949383473023772, + 0.002054624492302537, + 0.0020093759521842003, + 0.002805826487019658, + 0.0017785634845495224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.01108936034142971 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22556361556053162 + ], + "opposite_logprob_tokens": [ + -1.600563645362854 + ], + "logprob_reward": -0.22556361556053162, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22556361556053162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13743196427822113, + 0.04451564699411392, + 0.08453459292650223, + 0.06348396092653275, + 0.19517458975315094, + 0.1258670687675476, + 0.08817589282989502, + 0.052835654467344284, + 0.053669948130846024, + 0.1543106883764267 + ], + "credit_map": [ + 0.019559824839234352, + 0.010184225626289845, + 0.011187853291630745, + 0.013537466526031494, + 0.022852256894111633, + 0.018792102113366127, + 0.030789008364081383, + 0.010529232211411, + 0.00460649561136961, + 0.0038993796333670616, + 0.0037088680546730757, + 0.004043193534016609, + 0.01141910720616579, + 0.005304180551320314, + 0.006660100072622299, + 0.004874323029071093, + 0.007798382546752691, + 0.017135370522737503, + 0.005654914304614067, + 0.006935497280210257, + 0.017068030312657356, + 0.008368163369596004, + 0.01319214142858982, + 0.008382091298699379, + 0.005705967545509338, + 0.010561322793364525, + 0.006233665160834789, + 0.004968166816979647, + 0.009439957328140736, + 0.011784344911575317, + 0.008022986352443695, + 0.006767550017684698, + 0.024517018347978592, + 0.041713912039995193, + 0.0407257117331028, + 0.03256744518876076, + 0.015655172988772392, + 0.010858505964279175, + 0.016880081966519356, + 0.012256737798452377, + 0.008841536939144135, + 0.02776365913450718, + 0.025570781901478767, + 0.013248885050415993, + 0.01262977160513401, + 0.014152655377984047, + 0.012466761283576488, + 0.01119302399456501, + 0.010804523713886738, + 0.013563564978539944, + 0.013091569766402245, + 0.011425460688769817, + 0.009475898928940296, + 0.012080552987754345, + 0.009764352813363075, + 0.007969976402819157, + 0.006453750189393759, + 0.008206184022128582, + 0.005735441111028194, + 0.006869371514767408, + 0.007380208000540733, + 0.0062932237051427364, + 0.006903755944222212, + 0.004993717651814222, + 0.015790997073054314, + 0.013137578032910824, + 0.006531015038490295, + 0.002645040163770318, + 0.004218549467623234, + 0.005418640561401844, + 0.003967541269958019, + 0.0019605844281613827, + 0.03113165870308876, + 0.01116733904927969, + 0.0331692136824131, + 0.01075681857764721, + 0.01668759062886238, + 0.01694386638700962, + 0.018272437155246735, + 0.016181761398911476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.22556361556053162 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270415484905243 + ], + "opposite_logprob_tokens": [ + -2.1270415782928467 + ], + "logprob_reward": -0.1270415484905243, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270415484905243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08019281178712845, + 0.027358928695321083, + 0.05684339255094528, + 0.0764237716794014, + 0.12602722644805908, + 0.3193562626838684, + 0.032644931226968765, + 0.0705772265791893, + 0.18729040026664734, + 0.023284990340471268 + ], + "credit_map": [ + 0.010787536390125751, + 0.00654245400801301, + 0.00913624744862318, + 0.007887942716479301, + 0.010410056449472904, + 0.009127235040068626, + 0.02097727172076702, + 0.0053240638226270676, + 0.002317809034138918, + 0.001956614665687084, + 0.004018967505544424, + 0.0039223614148795605, + 0.004709145054221153, + 0.00263077300041914, + 0.004244296345859766, + 0.0035589616745710373, + 0.007185011170804501, + 0.006451382301747799, + 0.0042610554955899715, + 0.005751342047005892, + 0.01148314867168665, + 0.004449860192835331, + 0.011042023077607155, + 0.006219570524990559, + 0.006297707557678223, + 0.009178487583994865, + 0.008322100155055523, + 0.006739854346960783, + 0.01328828651458025, + 0.009512882679700851, + 0.013797101564705372, + 0.00928735826164484, + 0.009380015544593334, + 0.022806864231824875, + 0.014587730169296265, + 0.021408069878816605, + 0.021551504731178284, + 0.010287952609360218, + 0.013721806928515434, + 0.012283287942409515, + 0.028713319450616837, + 0.1037674993276596, + 0.12426988780498505, + 0.017087532207369804, + 0.017323635518550873, + 0.007838795892894268, + 0.006791530176997185, + 0.013564076274633408, + 0.003608646336942911, + 0.004591488279402256, + 0.005571551155298948, + 0.003905232297256589, + 0.0032019130885601044, + 0.004837547428905964, + 0.0037489193491637707, + 0.003179635386914015, + 0.009943634271621704, + 0.012050285935401917, + 0.009359817951917648, + 0.010535890236496925, + 0.00819187518209219, + 0.008759623393416405, + 0.00666582677513361, + 0.0050702751614153385, + 0.06812524795532227, + 0.019281912595033646, + 0.05990096181631088, + 0.006217290181666613, + 0.0049854665994644165, + 0.011659737676382065, + 0.013692864216864109, + 0.003426922718062997, + 0.0043571037240326405, + 0.0013108238345012069, + 0.005830029491335154, + 0.0018424094887450337, + 0.0018116140272468328, + 0.002325821667909622, + 0.002857147017493844, + 0.0029500394593924284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.1270415484905243 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107332669198513 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.01107332669198513, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107332669198513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04174618795514107, + 0.02781480923295021, + 0.04424279183149338, + 0.01725238934159279, + 0.1441023200750351, + 0.5338711738586426, + 0.015805914998054504, + 0.03416140004992485, + 0.1252746433019638, + 0.015728386119008064 + ], + "credit_map": [ + 0.004811930004507303, + 0.0037451472599059343, + 0.0042679100297391415, + 0.005030015017837286, + 0.007826818153262138, + 0.005199769977480173, + 0.007954063825309277, + 0.002910532057285309, + 0.005056584253907204, + 0.0018516408745199442, + 0.001894833636470139, + 0.0016481914790347219, + 0.003713830141350627, + 0.007014386355876923, + 0.004838254302740097, + 0.0017970872577279806, + 0.0037188457790762186, + 0.007175484672188759, + 0.0030136811546981335, + 0.004396776668727398, + 0.0094514274969697, + 0.003684388240799308, + 0.00782102346420288, + 0.004981163423508406, + 0.0015981892356649041, + 0.003540915669873357, + 0.0019479402108117938, + 0.002218144480139017, + 0.0018005891470238566, + 0.0018943585455417633, + 0.0023358201142400503, + 0.0019164327532052994, + 0.011532486416399479, + 0.04770306497812271, + 0.022477518767118454, + 0.023373108357191086, + 0.015268078073859215, + 0.0064754304476082325, + 0.010671594180166721, + 0.006601057481020689, + 0.04921773448586464, + 0.1868012249469757, + 0.2146756798028946, + 0.022122643887996674, + 0.027344893664121628, + 0.005104888696223497, + 0.00893380120396614, + 0.01967032067477703, + 0.001948762685060501, + 0.0027761005330830812, + 0.003697632346302271, + 0.0013665133155882359, + 0.0015089476946741343, + 0.0017836653860285878, + 0.001659517060033977, + 0.001064775395207107, + 0.004717780277132988, + 0.006758356466889381, + 0.003957463428378105, + 0.006031858269125223, + 0.004531716462224722, + 0.0038741077296435833, + 0.0024829076137393713, + 0.00180721131619066, + 0.04650538042187691, + 0.026293205097317696, + 0.024804865941405296, + 0.004789642058312893, + 0.004202487412840128, + 0.010099170729517937, + 0.00592963770031929, + 0.0026502578984946012, + 0.002670438727363944, + 0.0012418500846251845, + 0.0038338531740009785, + 0.001479863771237433, + 0.0018855812959372997, + 0.001493785297498107, + 0.0018321407260373235, + 0.0012908735079690814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.01107332669198513 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, pans slowly to the right, following an elderly Asian man. He is clad in a traditional white silk Tai Chi suit, his silvery hair neatly tied back. Deep wrinkles etch the quiet of age. Amidst the thick morning mist, he slowly practices Tai Chi, his arms moving in a steady, deliberate rhythm, one hand rising, the other gently lowering, as if parting the mist before him. The entire scene is a high-contrast black and white image, the soft, diffused morning light filtering through the mist, delicately capturing not only his relaxed posture but also the flowing folds and textures of the silk garment. In the background, the silhouettes of several ancient trees loom through the mist, creating an atmosphere of tranquility, concentration, and timeless poetry.", + "seed": 19796, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_13/index_4191_seed_19796_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10318747619166971, + "vlm_logprob": -0.10318747619166971, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.10318747619166971, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008746291510760784 + ], + "opposite_logprob_tokens": [ + -5.508746147155762 + ], + "logprob_reward": -0.008746291510760784, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.008746291510760784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2517435848712921, + 0.05353628844022751, + 0.10037730634212494, + 0.05293786898255348, + 0.0804276242852211, + 0.11433202028274536, + 0.043245960026979446, + 0.1450425237417221, + 0.03479155898094177, + 0.12356529384851456 + ], + "credit_map": [ + 0.03808819130063057, + 0.026528790593147278, + 0.03158628195524216, + 0.03961315378546715, + 0.02315313182771206, + 0.021357441321015358, + 0.034241851419210434, + 0.037174761295318604, + 0.005988527555018663, + 0.006919283885508776, + 0.005376037210226059, + 0.009859173558652401, + 0.009048127569258213, + 0.007835596799850464, + 0.004044416360557079, + 0.004465122707188129, + 0.010229863226413727, + 0.014580685645341873, + 0.011751458048820496, + 0.021450435742735863, + 0.009325971826910973, + 0.0059228637255728245, + 0.014525017701089382, + 0.012591008096933365, + 0.005998893175274134, + 0.0073342882096767426, + 0.012444498017430305, + 0.004086851142346859, + 0.004562297835946083, + 0.0021820447873324156, + 0.011912791058421135, + 0.004416206385940313, + 0.0069368751719594, + 0.008738133125007153, + 0.008391366340219975, + 0.014639440923929214, + 0.010741610080003738, + 0.006919219624251127, + 0.011689350940287113, + 0.01237163320183754, + 0.016904255375266075, + 0.012678470462560654, + 0.015604342333972454, + 0.027074499055743217, + 0.008953029289841652, + 0.01005581859499216, + 0.010339802131056786, + 0.012721799314022064, + 0.005139975342899561, + 0.0032595121301710606, + 0.005843671038746834, + 0.005871241446584463, + 0.007843597792088985, + 0.0037793791852891445, + 0.005868745036423206, + 0.0056398422457277775, + 0.009123860858380795, + 0.004548476543277502, + 0.04904947429895401, + 0.05462082847952843, + 0.004805231932550669, + 0.005314410664141178, + 0.010463823564350605, + 0.007116413209587336, + 0.0049246009439229965, + 0.0036915219388902187, + 0.0035856603644788265, + 0.004815907217562199, + 0.005468916147947311, + 0.0029463793616741896, + 0.005898099392652512, + 0.0034604764077812433, + 0.01595674827694893, + 0.021937241777777672, + 0.014210470020771027, + 0.010268477723002434, + 0.023846587166190147, + 0.007194229867309332, + 0.014849628321826458, + 0.015301911160349846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.008746291510760784 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648086339235306 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.008648086339235306, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648086339235306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12782420217990875, + 0.06183278188109398, + 0.13954100012779236, + 0.05609259009361267, + 0.06288091093301773, + 0.0921601951122284, + 0.14449074864387512, + 0.062310077250003815, + 0.07347110658884048, + 0.17939645051956177 + ], + "credit_map": [ + 0.023962294682860374, + 0.015101118944585323, + 0.0148454699665308, + 0.01648673042654991, + 0.011422323994338512, + 0.01014115009456873, + 0.01635590009391308, + 0.019509203732013702, + 0.00611332431435585, + 0.009491967037320137, + 0.006579052656888962, + 0.008346433751285076, + 0.008945530280470848, + 0.005797498859465122, + 0.010299867019057274, + 0.006259103771299124, + 0.011075823567807674, + 0.012755848467350006, + 0.016202179715037346, + 0.024325938895344734, + 0.015839984640479088, + 0.011193672195076942, + 0.02543938159942627, + 0.022708160802721977, + 0.005171446595340967, + 0.004705693107098341, + 0.010513750836253166, + 0.011050814762711525, + 0.004866542294621468, + 0.00394797557964921, + 0.009994745254516602, + 0.0058416156098246574, + 0.005898009520024061, + 0.008674695156514645, + 0.006524113472551107, + 0.008516431786119938, + 0.00973187293857336, + 0.0058242506347596645, + 0.0078717777505517, + 0.009839756414294243, + 0.011274799704551697, + 0.00841238908469677, + 0.013338196091353893, + 0.025347359478473663, + 0.009440441615879536, + 0.006520959548652172, + 0.009040354751050472, + 0.008785692974925041, + 0.010874122381210327, + 0.011406349949538708, + 0.017882036045193672, + 0.015116177499294281, + 0.019553665071725845, + 0.020541034638881683, + 0.0222362969070673, + 0.026881063356995583, + 0.004184882156550884, + 0.003639247501268983, + 0.015297829173505306, + 0.01955859363079071, + 0.0038373477291315794, + 0.003532114904373884, + 0.005215900018811226, + 0.007044157478958368, + 0.009508492425084114, + 0.009513059630990028, + 0.006344879977405071, + 0.010670087300240993, + 0.010513131506741047, + 0.005926967598497868, + 0.012331875041127205, + 0.008662610314786434, + 0.0198192335665226, + 0.026290640234947205, + 0.02092447504401207, + 0.016028694808483124, + 0.03252648562192917, + 0.010004841722548008, + 0.027920186519622803, + 0.025881873443722725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.008648086339235306 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204156205058098 + ], + "opposite_logprob_tokens": [ + -2.8120415210723877 + ], + "logprob_reward": -0.06204156205058098, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204156205058098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11755864322185516, + 0.07135602831840515, + 0.09872490167617798, + 0.028867188841104507, + 0.026801493018865585, + 0.04436459392309189, + 0.033483076840639114, + 0.4785904884338379, + 0.028320899233222008, + 0.07193269580602646 + ], + "credit_map": [ + 0.016973162069916725, + 0.017527448013424873, + 0.01324840821325779, + 0.01665869913995266, + 0.018125776201486588, + 0.0099799120798707, + 0.014210489578545094, + 0.01083474513143301, + 0.007669225800782442, + 0.00851401686668396, + 0.007412115577608347, + 0.010645365342497826, + 0.009984806180000305, + 0.007506459951400757, + 0.008268604055047035, + 0.011355430819094181, + 0.009735483676195145, + 0.018607892096042633, + 0.010642431676387787, + 0.0183987095952034, + 0.008855218067765236, + 0.010034765116870403, + 0.012600196525454521, + 0.009850203059613705, + 0.002911244286224246, + 0.0023747319355607033, + 0.006118774879723787, + 0.002060980536043644, + 0.0035250382497906685, + 0.0025485449004918337, + 0.00666477344930172, + 0.002663099206984043, + 0.004854021593928337, + 0.0042508915066719055, + 0.003353705396875739, + 0.003956513479351997, + 0.00326741486787796, + 0.002341755200177431, + 0.002680347068235278, + 0.002096844371408224, + 0.004376464523375034, + 0.00318702170625329, + 0.006487315986305475, + 0.013411730527877808, + 0.0035894191823899746, + 0.002585805719718337, + 0.005749736912548542, + 0.004977102391421795, + 0.0040804604068398476, + 0.0021938737481832504, + 0.004048503935337067, + 0.003986131865531206, + 0.005454106722027063, + 0.004166427068412304, + 0.0037241820245981216, + 0.005829392466694117, + 0.030992139130830765, + 0.018331017345190048, + 0.16214095056056976, + 0.17449083924293518, + 0.017869753763079643, + 0.01893833465874195, + 0.03244258090853691, + 0.023384831845760345, + 0.0030877746175974607, + 0.004516880493611097, + 0.003650917438790202, + 0.003936036955565214, + 0.0038374243304133415, + 0.0028241979889571667, + 0.0032675345428287983, + 0.0032001305371522903, + 0.01378583163022995, + 0.013624245300889015, + 0.008262100629508495, + 0.0062645976431667805, + 0.01084290724247694, + 0.005970457103103399, + 0.007367135025560856, + 0.00581542169675231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.06204156205058098 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42887580394744873 + ], + "opposite_logprob_tokens": [ + -1.0538758039474487 + ], + "logprob_reward": -0.42887580394744873, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42887580394744873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2799294888973236, + 0.03463299572467804, + 0.11283183842897415, + 0.07627049833536148, + 0.0438237227499485, + 0.10106219351291656, + 0.11377615481615067, + 0.07890087366104126, + 0.04652203619480133, + 0.11225014179944992 + ], + "credit_map": [ + 0.06370941549539566, + 0.04801434651017189, + 0.030098039656877518, + 0.03502902388572693, + 0.02478073164820671, + 0.020905807614326477, + 0.023565787822008133, + 0.03382634371519089, + 0.004440090619027615, + 0.0040056356228888035, + 0.0035139378160238266, + 0.005798358004540205, + 0.0058508096262812614, + 0.004137200303375721, + 0.0036038432735949755, + 0.0032831227872520685, + 0.011769095435738564, + 0.02142908237874508, + 0.012944082729518414, + 0.023643335327506065, + 0.010919601656496525, + 0.0072271013632416725, + 0.013588761910796165, + 0.011310773901641369, + 0.011208318173885345, + 0.01770165003836155, + 0.009879866614937782, + 0.010373634286224842, + 0.00745162321254611, + 0.003943243063986301, + 0.007984641008079052, + 0.007727526593953371, + 0.005664783529937267, + 0.008010480552911758, + 0.004231315106153488, + 0.007915285415947437, + 0.004597908817231655, + 0.002971638459712267, + 0.004857491236180067, + 0.005574818700551987, + 0.008807463571429253, + 0.010540125891566277, + 0.011204677633941174, + 0.018154336139559746, + 0.017591586336493492, + 0.010548204183578491, + 0.011791178025305271, + 0.012424628250300884, + 0.010785740800201893, + 0.007768127601593733, + 0.01431884802877903, + 0.010041212663054466, + 0.015552532859146595, + 0.023831868544220924, + 0.01384603139013052, + 0.017631784081459045, + 0.008220306597650051, + 0.005036013666540384, + 0.021800342947244644, + 0.0191990677267313, + 0.0052782092243433, + 0.00574537506327033, + 0.005583828315138817, + 0.008037732914090157, + 0.006449435837566853, + 0.008541669696569443, + 0.005410193931311369, + 0.005787490867078304, + 0.0050980523228645325, + 0.0041468171402812, + 0.006077965721487999, + 0.005010413937270641, + 0.0190066359937191, + 0.025949738919734955, + 0.013014283031225204, + 0.011820818297564983, + 0.010540283285081387, + 0.00855609867721796, + 0.01359633356332779, + 0.00976595189422369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.42887580394744873 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Asian man with silvery hair tied back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a traditional white silk Tai Chi suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the silhouettes of several ancient trees visible in the misty background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625637110322714 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.007625637110322714, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625637110322714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15910761058330536, + 0.08892769366502762, + 0.0890396386384964, + 0.1669284552335739, + 0.049107953906059265, + 0.1417984813451767, + 0.02077169343829155, + 0.23498310148715973, + 0.01833600178360939, + 0.03099941462278366 + ], + "credit_map": [ + 0.01869521662592888, + 0.03303830325603485, + 0.01621408201754093, + 0.02439100109040737, + 0.02773910015821457, + 0.01581706665456295, + 0.01016791258007288, + 0.013044929131865501, + 0.011630086228251457, + 0.02011561393737793, + 0.009130241349339485, + 0.013898182660341263, + 0.010468919761478901, + 0.005871414206922054, + 0.007845775224268436, + 0.00996746588498354, + 0.010709457099437714, + 0.01067949365824461, + 0.010985293425619602, + 0.025163855403661728, + 0.008346669375896454, + 0.008682356216013432, + 0.007907463237643242, + 0.00656504649668932, + 0.025446727871894836, + 0.02691509947180748, + 0.028324363753199577, + 0.035102613270282745, + 0.015102672390639782, + 0.010533764027059078, + 0.013691285625100136, + 0.011811922304332256, + 0.00718899630010128, + 0.00763229513540864, + 0.0053436122834682465, + 0.008946558460593224, + 0.005116397514939308, + 0.004149147309362888, + 0.005895029287785292, + 0.004835920408368111, + 0.021804701536893845, + 0.012457671575248241, + 0.02262110821902752, + 0.04144476354122162, + 0.011300723999738693, + 0.00846151914447546, + 0.011692325584590435, + 0.01201565656810999, + 0.004773604217916727, + 0.0021296555642038584, + 0.002727268496528268, + 0.0026729251258075237, + 0.0024865607265383005, + 0.0017171947984024882, + 0.001994731370359659, + 0.0022697532549500465, + 0.013990504667162895, + 0.007435974199324846, + 0.08025939017534256, + 0.0897296667098999, + 0.007179104257375002, + 0.008577980101108551, + 0.01691538281738758, + 0.010895087383687496, + 0.0022488320246338844, + 0.003042706288397312, + 0.0019055978627875447, + 0.002464588265866041, + 0.002592503558844328, + 0.0013583451509475708, + 0.0028709827456623316, + 0.0018524456536397338, + 0.005215913988649845, + 0.007114863023161888, + 0.00315607781521976, + 0.0026714750565588474, + 0.004979169461876154, + 0.0018880520947277546, + 0.002775047905743122, + 0.003198814345523715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.007625637110322714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_14/index_tensor([8003, 8003])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_14/index_tensor([8003, 8003])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f212a7797ad31bc2c74051b0474b57f65cf4f623 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_14/index_tensor([8003, 8003])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium close-up static shot from the bedside focuses on a young woman suddenly awakened in her bedroom late at night. Her long, tousled dark brown hair, strands of it clinging to her pale forehead and cheeks from cold sweat, is jolted from her reclining position by an invisible electric shock, her body snapping up quickly and stiffly. Her loose, gray cotton T-shirt, slightly damp with sweat, clings to her violently heaving chest. A ray of cold moonlight filters through the slits in the Venetian blinds, piercing the darkness, illuminating her eyes, wide with fear, and her lips, slightly parted, emitting rapid, hoarse gasps. Her hands instinctively clutch the wrinkled white sheets beneath her, her knuckles white with the strain. Her gaze darts around the dimly lit room, trying to discern the boundary between reality and dream. The high-contrast lighting and the chilling blue hue create an atmosphere of intense tension, fear, and the disorientation of a moment awakened from a nightmare.", + "seed": 9070, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_14/index_8003_seed_9070_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.23896281840279698, + "vlm_logprob": -0.23896281840279698, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.23896281840279698, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030249718576669693 + ], + "opposite_logprob_tokens": [ + -4.28024959564209 + ], + "logprob_reward": -0.030249718576669693, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.030249718576669693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13482537865638733, + 0.3009668290615082, + 0.06214184686541557, + 0.04735899716615677, + 0.21085001528263092, + 0.048011936247348785, + 0.04611576721072197, + 0.06354010105133057, + 0.050157580524683, + 0.03603154420852661 + ], + "credit_map": [ + 0.012503234669566154, + 0.013451069593429565, + 0.01283642090857029, + 0.024341462180018425, + 0.011580564081668854, + 0.013076819479465485, + 0.022393586114048958, + 0.024642203003168106, + 0.01628761738538742, + 0.019511284306645393, + 0.02457483299076557, + 0.028230899944901466, + 0.03725013509392738, + 0.04827876016497612, + 0.04211844131350517, + 0.08471482992172241, + 0.007515707053244114, + 0.009374154731631279, + 0.006222780328243971, + 0.004534119740128517, + 0.008197523653507233, + 0.01353661809116602, + 0.005262031219899654, + 0.0074989087879657745, + 0.006349876057356596, + 0.0065123531967401505, + 0.005977869965136051, + 0.007350428961217403, + 0.005310843698680401, + 0.005225949455052614, + 0.004650137387216091, + 0.005981534719467163, + 0.032001350075006485, + 0.022215157747268677, + 0.07553719729185104, + 0.026738032698631287, + 0.011703715659677982, + 0.01844712719321251, + 0.01121728215366602, + 0.012990151531994343, + 0.004223653115332127, + 0.005615494213998318, + 0.011727787554264069, + 0.006883325520902872, + 0.004231215454638004, + 0.0074369581416249275, + 0.0037464573979377747, + 0.004147041123360395, + 0.004613492637872696, + 0.012420181185007095, + 0.006847623270004988, + 0.006480744574218988, + 0.005225331522524357, + 0.0033321278169751167, + 0.0045169563964009285, + 0.002679307945072651, + 0.005954267457127571, + 0.011625649407505989, + 0.005875069182366133, + 0.0066370912827551365, + 0.00971165020018816, + 0.007902977988123894, + 0.008211283013224602, + 0.00762210926041007, + 0.00859399326145649, + 0.006578396074473858, + 0.007738226093351841, + 0.003960140980780125, + 0.0056089721620082855, + 0.005822652950882912, + 0.006417703349143267, + 0.00543749425560236, + 0.007983566261827946, + 0.0032538401428610086, + 0.005948150530457497, + 0.004242252092808485, + 0.004089351277798414, + 0.003458807012066245, + 0.004450968932360411, + 0.0026046058628708124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.030249718576669693 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005270870868116617 + ], + "opposite_logprob_tokens": [ + -5.255270957946777 + ], + "logprob_reward": -0.005270870868116617, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005270870868116617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12455481290817261, + 0.1071290597319603, + 0.021404514089226723, + 0.08119163662195206, + 0.24181970953941345, + 0.04246552661061287, + 0.15134169161319733, + 0.10691174119710922, + 0.03865848481655121, + 0.08452286571264267 + ], + "credit_map": [ + 0.011861886829137802, + 0.009469945915043354, + 0.013099856674671173, + 0.01634378172457218, + 0.012896704487502575, + 0.016429606825113297, + 0.020116426050662994, + 0.024336596950888634, + 0.011046795174479485, + 0.010656971484422684, + 0.009323976933956146, + 0.012664935551583767, + 0.017606381326913834, + 0.012742170132696629, + 0.012939884327352047, + 0.020147934556007385, + 0.003298745257779956, + 0.003315674141049385, + 0.0024617258459329605, + 0.0023632452357560396, + 0.0027113230898976326, + 0.002813531318679452, + 0.0019816916901618242, + 0.002458576112985611, + 0.008255126886069775, + 0.010049768723547459, + 0.013064461760222912, + 0.009935094974935055, + 0.010213437490165234, + 0.010138895362615585, + 0.008761304430663586, + 0.010773546993732452, + 0.032597724348306656, + 0.039634764194488525, + 0.03888088837265968, + 0.05559306964278221, + 0.019175121560692787, + 0.02165886014699936, + 0.019254613667726517, + 0.01502466294914484, + 0.004108442924916744, + 0.005916259251534939, + 0.008350533433258533, + 0.006579587236046791, + 0.004787126090377569, + 0.00490306643769145, + 0.0044915457256138325, + 0.0033289603888988495, + 0.014269593171775341, + 0.03518897667527199, + 0.028834234923124313, + 0.01903967186808586, + 0.016660252586007118, + 0.01564708538353443, + 0.013584903441369534, + 0.008116966113448143, + 0.011902830563485622, + 0.016839344054460526, + 0.01003615465015173, + 0.012032140046358109, + 0.015588004142045975, + 0.014191247522830963, + 0.014333113096654415, + 0.01198890432715416, + 0.004432904999703169, + 0.005424144212156534, + 0.0036499646957963705, + 0.005680467933416367, + 0.004752347245812416, + 0.003544151782989502, + 0.006839120760560036, + 0.004335383418947458, + 0.016759468242526054, + 0.007014934904873371, + 0.013784141279757023, + 0.011493206955492496, + 0.008986031636595726, + 0.007174776867032051, + 0.009607236832380295, + 0.009703067131340504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.005270870868116617 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026363030076026917 + ], + "opposite_logprob_tokens": [ + -3.6513631343841553 + ], + "logprob_reward": -0.026363030076026917, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026363030076026917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11307401210069656, + 0.09302911907434464, + 0.049148641526699066, + 0.04803881049156189, + 0.272437185049057, + 0.08272694796323776, + 0.05750967934727669, + 0.22041083872318268, + 0.04404053837060928, + 0.01958424225449562 + ], + "credit_map": [ + 0.00829374697059393, + 0.008146869018673897, + 0.010187964886426926, + 0.016659507527947426, + 0.008354701101779938, + 0.012935321778059006, + 0.01754486933350563, + 0.030951030552387238, + 0.006804263219237328, + 0.009386701509356499, + 0.010611196979880333, + 0.009597095660865307, + 0.019093971699476242, + 0.012163122184574604, + 0.008604314178228378, + 0.016768457368016243, + 0.00476102577522397, + 0.00670384056866169, + 0.004930346272885799, + 0.0036104172468185425, + 0.007705832365900278, + 0.011122217401862144, + 0.0030945478938519955, + 0.007220418658107519, + 0.00816634576767683, + 0.004106695298105478, + 0.0056340959854424, + 0.006205847021192312, + 0.006632068194448948, + 0.00767713924869895, + 0.005195272620767355, + 0.004421343561261892, + 0.06903653591871262, + 0.02400498278439045, + 0.061009157449007034, + 0.03013726882636547, + 0.016427556052803993, + 0.040358979254961014, + 0.01669720560312271, + 0.014765491709113121, + 0.007865944877266884, + 0.007440898567438126, + 0.011823296546936035, + 0.0104674706235528, + 0.01515250001102686, + 0.013173176907002926, + 0.011353980749845505, + 0.005449679214507341, + 0.005650629755109549, + 0.00641942024230957, + 0.013404930010437965, + 0.004246239084750414, + 0.004294147714972496, + 0.013151641003787518, + 0.005972770042717457, + 0.00436989963054657, + 0.02230779640376568, + 0.03471672162413597, + 0.01998743787407875, + 0.020666388794779778, + 0.03431430086493492, + 0.035803213715553284, + 0.030165227130055428, + 0.022449756041169167, + 0.005701724905520678, + 0.00740650063380599, + 0.004927922040224075, + 0.00469247717410326, + 0.005325705744326115, + 0.004427725449204445, + 0.006365177687257528, + 0.0051933047361671925, + 0.00393070001155138, + 0.0013435775181278586, + 0.0035206731408834457, + 0.002533371327444911, + 0.0018283940153196454, + 0.0015972200781106949, + 0.002250088844448328, + 0.002580217784270644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.026363030076026917 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789993554353714 + ], + "opposite_logprob_tokens": [ + -2.5789992809295654 + ], + "logprob_reward": -0.0789993554353714, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789993554353714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09939390420913696, + 0.0893237292766571, + 0.10675486922264099, + 0.054496463388204575, + 0.1670985370874405, + 0.04467727243900299, + 0.05591175705194473, + 0.1569753885269165, + 0.07184842228889465, + 0.1535196453332901 + ], + "credit_map": [ + 0.01173796970397234, + 0.010035477578639984, + 0.007829377427697182, + 0.00949738547205925, + 0.009932737797498703, + 0.010754536837339401, + 0.022360527887940407, + 0.017245890572667122, + 0.007906950078904629, + 0.009113293141126633, + 0.006457233801484108, + 0.009680711664259434, + 0.019691456109285355, + 0.011812194250524044, + 0.009121379815042019, + 0.015540512278676033, + 0.011450208723545074, + 0.013240327127277851, + 0.010092989541590214, + 0.008665255270898342, + 0.018222566694021225, + 0.026851676404476166, + 0.006437294650822878, + 0.011794555932283401, + 0.005495114251971245, + 0.006550153251737356, + 0.005306755658239126, + 0.00922426301985979, + 0.007593271788209677, + 0.01127228420227766, + 0.004903648514300585, + 0.004150972235947847, + 0.023614099249243736, + 0.018805857747793198, + 0.05468209460377693, + 0.02982824109494686, + 0.008643340319395065, + 0.01587395928800106, + 0.009004565887153149, + 0.006646387279033661, + 0.004066839348524809, + 0.00744676124304533, + 0.0067064519971609116, + 0.007089561317116022, + 0.005260714795440435, + 0.005569341592490673, + 0.0043145837262272835, + 0.0042230188846588135, + 0.004698517266660929, + 0.007341697346419096, + 0.005547964945435524, + 0.005547047592699528, + 0.007590016815811396, + 0.008303046226501465, + 0.012942232191562653, + 0.003941232338547707, + 0.01868245378136635, + 0.020833326503634453, + 0.015610666014254093, + 0.01585676521062851, + 0.02361297979950905, + 0.023295732215046883, + 0.01638966239988804, + 0.022693796083331108, + 0.010580448433756828, + 0.014228377491235733, + 0.007886032573878765, + 0.009595652110874653, + 0.007818479090929031, + 0.00525965541601181, + 0.010726244188845158, + 0.005753534380346537, + 0.04037325456738472, + 0.013277134858071804, + 0.023478083312511444, + 0.0186984334141016, + 0.01345967035740614, + 0.014153252355754375, + 0.014203413389623165, + 0.015876401215791702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.0789993554353714 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539311170578003 + ], + "opposite_logprob_tokens": [ + -0.4289310872554779 + ], + "logprob_reward": -1.0539311170578003, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539311170578003, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06992151588201523, + 0.0605013445019722, + 0.05011812597513199, + 0.07803619652986526, + 0.24248334765434265, + 0.22414156794548035, + 0.06588613986968994, + 0.09408925473690033, + 0.09265309572219849, + 0.022169411182403564 + ], + "credit_map": [ + 0.00689929723739624, + 0.00633639981970191, + 0.006341016385704279, + 0.007817219011485577, + 0.007082473486661911, + 0.009814889170229435, + 0.009729117155075073, + 0.015901099890470505, + 0.009232093580067158, + 0.004914972931146622, + 0.005346036981791258, + 0.00542098144069314, + 0.007689631078392267, + 0.0074385227635502815, + 0.007524223066866398, + 0.012934882193803787, + 0.004341464955359697, + 0.006458533462136984, + 0.006459134630858898, + 0.004596563521772623, + 0.008021009154617786, + 0.00825915765017271, + 0.004696053918451071, + 0.007286211010068655, + 0.009646937251091003, + 0.009546381421387196, + 0.006883775815367699, + 0.011622295714914799, + 0.011732826940715313, + 0.010864544659852982, + 0.01050149742513895, + 0.0072379345074296, + 0.042909666895866394, + 0.0327465645968914, + 0.06473968178033829, + 0.03323876112699509, + 0.01871185190975666, + 0.01684170588850975, + 0.020395448431372643, + 0.012899672612547874, + 0.021878374740481377, + 0.02780909836292267, + 0.05716414749622345, + 0.03325439989566803, + 0.020345378667116165, + 0.023901725187897682, + 0.016289060935378075, + 0.023499391973018646, + 0.006967633496969938, + 0.00728493370115757, + 0.011428626254200935, + 0.0068227979354560375, + 0.007725684437900782, + 0.008332406170666218, + 0.010801871307194233, + 0.006522188428789377, + 0.01066151075065136, + 0.011988160200417042, + 0.008470679633319378, + 0.009866598062217236, + 0.015119042247533798, + 0.012943616136908531, + 0.011083192192018032, + 0.013956456445157528, + 0.012730853632092476, + 0.011467409320175648, + 0.009726851247251034, + 0.010628990828990936, + 0.009394459426403046, + 0.010987391695380211, + 0.015036757104098797, + 0.012680387124419212, + 0.002665457548573613, + 0.0019999982323497534, + 0.002421677578240633, + 0.002509905258193612, + 0.0039938660338521, + 0.0016681227134540677, + 0.003908551763743162, + 0.0030018312390893698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -1.0539311170578003 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "low-angle, medium close-up static shot from the bedside focuses on a young woman suddenly awakened in her bedroom late at night. Her long, tousled dark brown hair, strands of it clinging to her pale forehead and cheeks from cold sweat, is jolted from her reclining position by an invisible electric shock, her body snapping up quickly and stiffly. Her loose, gray cotton T-shirt, slightly damp with sweat, clings to her violently heaving chest. A ray of cold moonlight filters through the slits in the Venetian blinds, piercing the darkness, illuminating her eyes, wide with fear, and her lips, slightly parted, emitting rapid, hoarse gasps. Her hands instinctively clutch the wrinkled white sheets beneath her, her knuckles white with the strain. Her gaze darts around the dimly lit room, trying to discern the boundary between reality and dream. The high-contrast lighting and the chilling blue hue create an atmosphere of intense tension, fear, and the disorientation of a moment awakened from a nightmare.", + "seed": 9071, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_14/index_8003_seed_9071_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08950280342251063, + "vlm_logprob": -0.08950280342251063, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.08950280342251063, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016663886606693268 + ], + "opposite_logprob_tokens": [ + -5.516664028167725 + ], + "logprob_reward": -0.016663886606693268, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.016663886606693268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05599333718419075, + 0.029794419184327126, + 0.03003847599029541, + 0.04608912765979767, + 0.05581134185194969, + 0.040888696908950806, + 0.5393978357315063, + 0.0408514067530632, + 0.11905059963464737, + 0.04208480939269066 + ], + "credit_map": [ + 0.006646618712693453, + 0.007823485881090164, + 0.007098765578120947, + 0.006749926134943962, + 0.006892540957778692, + 0.006911351345479488, + 0.009360254742205143, + 0.004510398954153061, + 0.0049920217134058475, + 0.0036867312155663967, + 0.0022778131533414125, + 0.00239743129350245, + 0.0071489037945866585, + 0.0037714678328484297, + 0.0030984554905444384, + 0.002421595389023423, + 0.0038645549211651087, + 0.002768881618976593, + 0.0019279348198324442, + 0.0026005427353084087, + 0.003107823897153139, + 0.007387250661849976, + 0.004566923715174198, + 0.0038145645521581173, + 0.004175894428044558, + 0.005147164221853018, + 0.004733317997306585, + 0.004149796441197395, + 0.008725152350962162, + 0.01279150228947401, + 0.0032642215956002474, + 0.0031020785681903362, + 0.012853714637458324, + 0.0035139555111527443, + 0.012163180857896805, + 0.0030097286216914654, + 0.007686980534344912, + 0.0048823850229382515, + 0.0028616974595934153, + 0.008839699439704418, + 0.002379091689363122, + 0.003363470546901226, + 0.003113469574600458, + 0.01179576851427555, + 0.0027406634762883186, + 0.011789018288254738, + 0.00309665291570127, + 0.00261056050658226, + 0.04777316749095917, + 0.011416536755859852, + 0.1173219233751297, + 0.05129008740186691, + 0.04092742130160332, + 0.23156073689460754, + 0.019258052110671997, + 0.01984991692006588, + 0.00613354193046689, + 0.004440010990947485, + 0.005454909056425095, + 0.004942052531987429, + 0.0025608979631215334, + 0.0038265581242740154, + 0.006431187968701124, + 0.00706225261092186, + 0.03306896239519119, + 0.009042376652359962, + 0.018207095563411713, + 0.008836207911372185, + 0.00920505728572607, + 0.01735159009695053, + 0.0131975207477808, + 0.010141780599951744, + 0.002628142712637782, + 0.005453245714306831, + 0.00582865159958601, + 0.005142038222402334, + 0.002500198082998395, + 0.006898254621773958, + 0.00852284487336874, + 0.0051114317029714584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.016663886606693268 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003213481977581978 + ], + "opposite_logprob_tokens": [ + -5.753213405609131 + ], + "logprob_reward": -0.003213481977581978, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003213481977581978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1890626698732376, + 0.09905686974525452, + 0.11308860033750534, + 0.14217829704284668, + 0.14393489062786102, + 0.04894232004880905, + 0.0669117197394371, + 0.05085065960884094, + 0.10119438171386719, + 0.04477958381175995 + ], + "credit_map": [ + 0.02204185537993908, + 0.02725410833954811, + 0.027829967439174652, + 0.036030303686857224, + 0.018993960693478584, + 0.014954564161598682, + 0.026461895555257797, + 0.015496010892093182, + 0.015455004759132862, + 0.01273245271295309, + 0.02169533632695675, + 0.006696091499179602, + 0.017582422122359276, + 0.010333787649869919, + 0.008981117978692055, + 0.005580652505159378, + 0.01383329089730978, + 0.012127919122576714, + 0.01791098713874817, + 0.013875007629394531, + 0.01380831841379404, + 0.01982196420431137, + 0.010012808255851269, + 0.011698312126100063, + 0.010978808626532555, + 0.017570598050951958, + 0.015963904559612274, + 0.00968278106302023, + 0.021892748773097992, + 0.04846692457795143, + 0.008501404896378517, + 0.009121118113398552, + 0.027020838111639023, + 0.01085733249783516, + 0.008074255660176277, + 0.006419451907277107, + 0.020169852301478386, + 0.014252537861466408, + 0.008880162611603737, + 0.048260461539030075, + 0.0030720008071511984, + 0.007103055249899626, + 0.00446291733533144, + 0.010693789459764957, + 0.005204719491302967, + 0.010117413476109505, + 0.004414695780724287, + 0.0038737321738153696, + 0.004912883508950472, + 0.006753837689757347, + 0.006494769360870123, + 0.00786253809928894, + 0.006141218356788158, + 0.016352958977222443, + 0.012543071061372757, + 0.005850441753864288, + 0.004236698616296053, + 0.0052437251433730125, + 0.009003819897770882, + 0.006534668151289225, + 0.003315210109576583, + 0.0067806304432451725, + 0.007048307452350855, + 0.008687598630785942, + 0.03093826211988926, + 0.0064918422140181065, + 0.014717633835971355, + 0.005956926848739386, + 0.007001446560025215, + 0.017517877742648125, + 0.010516852140426636, + 0.008053535595536232, + 0.0044472976587712765, + 0.005964263342320919, + 0.005337850656360388, + 0.006395265925675631, + 0.006608102936297655, + 0.004765502642840147, + 0.0067767249420285225, + 0.0044845761731266975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.003213481977581978 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208228327333927 + ], + "opposite_logprob_tokens": [ + -4.2642083168029785 + ], + "logprob_reward": -0.014208228327333927, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208228327333927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1072612777352333, + 0.07616225630044937, + 0.10913462936878204, + 0.13094422221183777, + 0.09801958501338959, + 0.05472107231616974, + 0.2813146114349365, + 0.03725825995206833, + 0.05738943815231323, + 0.04779462888836861 + ], + "credit_map": [ + 0.011479056440293789, + 0.014036668464541435, + 0.01140353363007307, + 0.015914462506771088, + 0.011240408755838871, + 0.01263383124023676, + 0.021015476435422897, + 0.009537830948829651, + 0.009042619727551937, + 0.008732163347303867, + 0.015449227765202522, + 0.006137917283922434, + 0.01342778466641903, + 0.008402501232922077, + 0.008828169666230679, + 0.006141866557300091, + 0.012692123651504517, + 0.008221098221838474, + 0.01231777761131525, + 0.0077455597929656506, + 0.008859562687575817, + 0.0290180966258049, + 0.020863002166152, + 0.009417405351996422, + 0.009852197021245956, + 0.009724138304591179, + 0.01603221893310547, + 0.006027787458151579, + 0.025414390489459038, + 0.05021237954497337, + 0.007544434629380703, + 0.006136670242995024, + 0.018110550940036774, + 0.00804596021771431, + 0.008887928910553455, + 0.0061311740428209305, + 0.008877815678715706, + 0.010637263767421246, + 0.008429553359746933, + 0.02889932692050934, + 0.0036894429940730333, + 0.0041083102114498615, + 0.006593065802007914, + 0.013138849288225174, + 0.005028290208429098, + 0.011056111194193363, + 0.005959246773272753, + 0.005147753283381462, + 0.021781641989946365, + 0.026190247386693954, + 0.0547008216381073, + 0.026471953839063644, + 0.02233240194618702, + 0.06145777925848961, + 0.05075991526246071, + 0.017619840800762177, + 0.0023633157834410667, + 0.005072887055575848, + 0.0050839320756495, + 0.004376702010631561, + 0.00337401544675231, + 0.0037172227166593075, + 0.005292740650475025, + 0.007977444678544998, + 0.017340781167149544, + 0.0040469118393957615, + 0.009090205654501915, + 0.0033839927054941654, + 0.004261170979589224, + 0.008493982255458832, + 0.005878310184925795, + 0.0048940773122012615, + 0.003593684174120426, + 0.004836083389818668, + 0.003975200932472944, + 0.005383784882724285, + 0.006810529623180628, + 0.008166217245161533, + 0.0118558993563056, + 0.0031732257921248674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.014208228327333927 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026349330320954323 + ], + "opposite_logprob_tokens": [ + -3.6513493061065674 + ], + "logprob_reward": -0.026349330320954323, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349330320954323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07138505578041077, + 0.026168256998062134, + 0.02661725878715515, + 0.06326568871736526, + 0.03935123607516289, + 0.04286785051226616, + 0.404655396938324, + 0.11087457090616226, + 0.06292716413736343, + 0.15188756585121155 + ], + "credit_map": [ + 0.010391841642558575, + 0.010465346276760101, + 0.00935169868171215, + 0.0075478027574718, + 0.008865509182214737, + 0.0075907958671450615, + 0.011860900558531284, + 0.005311165936291218, + 0.004591449163854122, + 0.0025080260820686817, + 0.004030787386000156, + 0.0016539943171665072, + 0.002247418276965618, + 0.0030621453188359737, + 0.004723563324660063, + 0.003350872313603759, + 0.003735124133527279, + 0.0037023837212473154, + 0.0039390805177390575, + 0.0021206149831414223, + 0.0022568756248801947, + 0.003910733852535486, + 0.004437076859176159, + 0.002515368163585663, + 0.005365000572055578, + 0.005092103499919176, + 0.008097910322248936, + 0.0029498112853616476, + 0.011127671226859093, + 0.024282367900013924, + 0.003815777599811554, + 0.00253504142165184, + 0.010989543050527573, + 0.0026663998141884804, + 0.005866041872650385, + 0.0028160407673567533, + 0.0032482342794537544, + 0.004335532430559397, + 0.00331444526091218, + 0.006114994175732136, + 0.00302303908392787, + 0.0022533752489835024, + 0.004035410471260548, + 0.009438826702535152, + 0.002925286768004298, + 0.014329983852803707, + 0.003991162404417992, + 0.002870765281841159, + 0.028986778110265732, + 0.007833830080926418, + 0.1016729399561882, + 0.03781356289982796, + 0.027665775269269943, + 0.1750626266002655, + 0.012386874295771122, + 0.013232992962002754, + 0.009565181098878384, + 0.012396170757710934, + 0.017654038965702057, + 0.008976644836366177, + 0.008588634431362152, + 0.01903158612549305, + 0.019165581092238426, + 0.015496731735765934, + 0.01463592704385519, + 0.00539779756218195, + 0.009537793695926666, + 0.005455968901515007, + 0.005184947978705168, + 0.010018151253461838, + 0.006876250728964806, + 0.005820323713123798, + 0.015046980232000351, + 0.013903266750276089, + 0.02191409468650818, + 0.01263607107102871, + 0.01425209641456604, + 0.02702454850077629, + 0.03166891261935234, + 0.015441601164638996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.026349330320954323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870790898799896 + ], + "opposite_logprob_tokens": [ + -1.137079119682312 + ], + "logprob_reward": -0.3870790898799896, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870790898799896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041952479630708694, + 0.05419492721557617, + 0.06235735863447189, + 0.1326306313276291, + 0.08999092131853104, + 0.04570591449737549, + 0.341593861579895, + 0.1733512580394745, + 0.017160959541797638, + 0.041061609983444214 + ], + "credit_map": [ + 0.004261426627635956, + 0.0064535681158304214, + 0.0057911863550543785, + 0.00582744088023901, + 0.005185909569263458, + 0.003992154262959957, + 0.007039898540824652, + 0.003400897141546011, + 0.006419735495001078, + 0.005813689902424812, + 0.010314562357962132, + 0.004019633866846561, + 0.006282303016632795, + 0.004570030141621828, + 0.01097155548632145, + 0.005803423468023539, + 0.006405362393707037, + 0.006577813997864723, + 0.005899145733565092, + 0.006949533708393574, + 0.010412232019007206, + 0.013679804280400276, + 0.007067627739161253, + 0.005365843418985605, + 0.012161131016910076, + 0.011590626090765, + 0.014646878466010094, + 0.008468655869364738, + 0.02265353873372078, + 0.049692850559949875, + 0.007632244378328323, + 0.005784721113741398, + 0.005501809995621443, + 0.005808044690638781, + 0.013410311192274094, + 0.00488033564761281, + 0.016931194812059402, + 0.008346778340637684, + 0.005195286590605974, + 0.0299171581864357, + 0.0038321588654071093, + 0.006098994519561529, + 0.004139084834605455, + 0.00655090669170022, + 0.006116471253335476, + 0.007634238339960575, + 0.004851604346185923, + 0.006482457742094994, + 0.02297947369515896, + 0.010605210438370705, + 0.07945282012224197, + 0.03192700445652008, + 0.023754959926009178, + 0.14515846967697144, + 0.012844547629356384, + 0.014871416613459587, + 0.016972817480564117, + 0.016751892864704132, + 0.020840616896748543, + 0.01705610752105713, + 0.012894261628389359, + 0.02175118774175644, + 0.04462963342666626, + 0.022454751655459404, + 0.004651488736271858, + 0.0011489811586216092, + 0.0015900766011327505, + 0.0014859047951176763, + 0.0013038668548688293, + 0.003400217043235898, + 0.002169680316001177, + 0.0014107463648542762, + 0.002469654195010662, + 0.0038707174826413393, + 0.002663182793185115, + 0.003651590086519718, + 0.002395043382421136, + 0.01054292544722557, + 0.011566183529794216, + 0.003902316326275468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.3870790898799896 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium close-up static shot from the bedside focuses on a young woman suddenly awakened in her bedroom late at night. Her long, tousled dark brown hair, strands of it clinging to her pale forehead and cheeks from cold sweat, is jolted from her reclining position by an invisible electric shock, her body snapping up quickly and stiffly. Her loose, gray cotton T-shirt, slightly damp with sweat, clings to her violently heaving chest. A ray of cold moonlight filters through the slits in the Venetian blinds, piercing the darkness, illuminating her eyes, wide with fear, and her lips, slightly parted, emitting rapid, hoarse gasps. Her hands instinctively clutch the wrinkled white sheets beneath her, her knuckles white with the strain. Her gaze darts around the dimly lit room, trying to discern the boundary between reality and dream. The high-contrast lighting and the chilling blue hue create an atmosphere of intense tension, fear, and the disorientation of a moment awakened from a nightmare.", + "seed": 9072, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_14/index_8003_seed_9072_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.19560915390029548, + "vlm_logprob": -0.19560915390029548, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.19560915390029548, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05975797399878502 + ], + "opposite_logprob_tokens": [ + -3.684757947921753 + ], + "logprob_reward": -0.05975797399878502, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.05975797399878502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11211712658405304, + 0.06861557066440582, + 0.1506434828042984, + 0.08601326495409012, + 0.07327533513307571, + 0.1107947826385498, + 0.15425743162631989, + 0.06711054593324661, + 0.0638369545340538, + 0.1133355125784874 + ], + "credit_map": [ + 0.011851996183395386, + 0.012602418661117554, + 0.01099409069865942, + 0.018381232395768166, + 0.015778914093971252, + 0.012677069753408432, + 0.01628357544541359, + 0.013547827489674091, + 0.0054410225711762905, + 0.009757655672729015, + 0.009578545577824116, + 0.009307242929935455, + 0.01158841885626316, + 0.006748789455741644, + 0.008750618435442448, + 0.00744327437132597, + 0.009513790719211102, + 0.009813416749238968, + 0.010789512656629086, + 0.010147552005946636, + 0.01305379718542099, + 0.007671835366636515, + 0.05025770142674446, + 0.03939586877822876, + 0.006109500769525766, + 0.008566411212086678, + 0.008173950016498566, + 0.010953332297503948, + 0.0070322914980351925, + 0.014544729143381119, + 0.016160517930984497, + 0.014472530223429203, + 0.006925494410097599, + 0.008516295813024044, + 0.0076316287741065025, + 0.010951374657452106, + 0.010475116781890392, + 0.00947936438024044, + 0.011380831710994244, + 0.007915220223367214, + 0.006933282595127821, + 0.011212247423827648, + 0.021516595035791397, + 0.02085757814347744, + 0.010273865424096584, + 0.011690903455018997, + 0.013027523644268513, + 0.01528277900069952, + 0.011137448251247406, + 0.04632936418056488, + 0.01632707193493843, + 0.014451872557401657, + 0.00894065573811531, + 0.026690999045968056, + 0.019847141578793526, + 0.010532858781516552, + 0.006598325911909342, + 0.01669861562550068, + 0.0054064346477389336, + 0.0059399972669780254, + 0.00700735580176115, + 0.007128906901925802, + 0.011505666188895702, + 0.006825235206633806, + 0.004589675460010767, + 0.013654041104018688, + 0.0063156080432236195, + 0.009661246091127396, + 0.004446009639650583, + 0.006842290051281452, + 0.012413385324180126, + 0.005914697423577309, + 0.006809627171605825, + 0.0180282611399889, + 0.0064097135327756405, + 0.023500243201851845, + 0.005361274350434542, + 0.011269566603004932, + 0.03292055055499077, + 0.009036269038915634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 9 + ], + "selected_reward": -0.05975797399878502 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006761409807950258 + ], + "opposite_logprob_tokens": [ + -5.00676155090332 + ], + "logprob_reward": -0.006761409807950258, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006761409807950258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09275750815868378, + 0.03519975021481514, + 0.12641914188861847, + 0.0791478380560875, + 0.04899654537439346, + 0.16561953723430634, + 0.11766811460256577, + 0.12872791290283203, + 0.059743061661720276, + 0.14572054147720337 + ], + "credit_map": [ + 0.009286914020776749, + 0.01090155728161335, + 0.00928251538425684, + 0.010984290391206741, + 0.01207138504832983, + 0.012732252478599548, + 0.01131211593747139, + 0.01618647575378418, + 0.0027856859378516674, + 0.004005445633083582, + 0.004591325763612986, + 0.004132320173084736, + 0.0057746912352740765, + 0.004281090572476387, + 0.005381687544286251, + 0.004247504286468029, + 0.00768481707200408, + 0.008783875033259392, + 0.007138196378946304, + 0.00931733287870884, + 0.011737467721104622, + 0.00785632524639368, + 0.041239332407712936, + 0.0326617956161499, + 0.004182163160294294, + 0.007894148118793964, + 0.008847923018038273, + 0.011312006041407585, + 0.006556088104844093, + 0.0061139133758842945, + 0.020046137273311615, + 0.014195466414093971, + 0.004043444059789181, + 0.006051398813724518, + 0.007244786247611046, + 0.008515235036611557, + 0.0066964514553546906, + 0.005425726994872093, + 0.005313062574714422, + 0.005706442054361105, + 0.009717099368572235, + 0.014562815427780151, + 0.04262600094079971, + 0.03210418298840523, + 0.017428789287805557, + 0.018004421144723892, + 0.01498121116310358, + 0.016195014119148254, + 0.00859928596764803, + 0.036124080419540405, + 0.010719270445406437, + 0.01135308388620615, + 0.006131351925432682, + 0.020822640508413315, + 0.01573382318019867, + 0.008184581995010376, + 0.016564147546887398, + 0.028857067227363586, + 0.008870804682374, + 0.009834575466811657, + 0.009602810256183147, + 0.010401885025203228, + 0.03153352439403534, + 0.013063102960586548, + 0.004521137103438377, + 0.01042497530579567, + 0.004733518231660128, + 0.008508495055139065, + 0.0059462967328727245, + 0.009456072933971882, + 0.008787871338427067, + 0.007364699617028236, + 0.00878357607871294, + 0.02857178822159767, + 0.008940236642956734, + 0.027526933699846268, + 0.009890071116387844, + 0.01645379140973091, + 0.03418832644820213, + 0.011365823447704315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.006761409807950258 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023308105766773224 + ], + "opposite_logprob_tokens": [ + -3.773308038711548 + ], + "logprob_reward": -0.023308105766773224, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023308105766773224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05199924483895302, + 0.02425071783363819, + 0.04826260730624199, + 0.19430819153785706, + 0.02883489429950714, + 0.48737069964408875, + 0.07079583406448364, + 0.02969674952328205, + 0.016802890226244926, + 0.04767819866538048 + ], + "credit_map": [ + 0.004924238659441471, + 0.005882160738110542, + 0.004954822361469269, + 0.0039850324392318726, + 0.005909122992306948, + 0.009562390856444836, + 0.008635623380541801, + 0.008145851083099842, + 0.0017824958777055144, + 0.0031509273685514927, + 0.002344700740650296, + 0.002853444078937173, + 0.003498789854347706, + 0.0024323398247361183, + 0.004856507293879986, + 0.0033315105829387903, + 0.002617892110720277, + 0.003910996485501528, + 0.0021424628794193268, + 0.0035986900329589844, + 0.004271186422556639, + 0.004583590663969517, + 0.013896293938159943, + 0.013241494074463844, + 0.00944590661674738, + 0.017735635861754417, + 0.019104553386569023, + 0.019378114491701126, + 0.02440628409385681, + 0.016867702826857567, + 0.06446130573749542, + 0.022908667102456093, + 0.003035013796761632, + 0.002820475259795785, + 0.003906743600964546, + 0.005247945431619883, + 0.00364618725143373, + 0.0031292124185711145, + 0.003176723374053836, + 0.0038725920021533966, + 0.026403892785310745, + 0.037781521677970886, + 0.119355209171772, + 0.0835525244474411, + 0.060476306825876236, + 0.06344785541296005, + 0.051362451165914536, + 0.04499093070626259, + 0.0052577476017177105, + 0.021085228770971298, + 0.006727650761604309, + 0.00814308412373066, + 0.0038488309364765882, + 0.011102033779025078, + 0.00971628725528717, + 0.004914969205856323, + 0.0039001144468784332, + 0.006541402079164982, + 0.0018966710194945335, + 0.0022798562422394753, + 0.0017539500258862972, + 0.0018285327823832631, + 0.009338323958218098, + 0.0021578974556177855, + 0.0013558374485000968, + 0.003030891064554453, + 0.002184066455811262, + 0.001975324237719178, + 0.0014773361617699265, + 0.0020108686294406652, + 0.002805790165439248, + 0.0019627748988568783, + 0.0033430620096623898, + 0.00777879822999239, + 0.003036776091903448, + 0.010305710136890411, + 0.002549699041992426, + 0.002821363043040037, + 0.01404169574379921, + 0.003801089944317937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.023308105766773224 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205063685774803 + ], + "opposite_logprob_tokens": [ + -2.8120505809783936 + ], + "logprob_reward": -0.06205063685774803, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205063685774803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07073742151260376, + 0.06517559289932251, + 0.1424049586057663, + 0.11115319281816483, + 0.09287679940462112, + 0.034402474761009216, + 0.07333783060312271, + 0.09293023496866226, + 0.09760347008705139, + 0.21937794983386993 + ], + "credit_map": [ + 0.007938024587929249, + 0.008163245394825935, + 0.004282313399016857, + 0.008558917790651321, + 0.009562662802636623, + 0.009665623307228088, + 0.012167348526418209, + 0.010399284772574902, + 0.004891799297183752, + 0.007963038049638271, + 0.007346035912632942, + 0.007961339317262173, + 0.010082854889333248, + 0.0076605770736932755, + 0.01038352120667696, + 0.008886429481208324, + 0.00808095932006836, + 0.010964167304337025, + 0.007337965536862612, + 0.008645568042993546, + 0.018360434100031853, + 0.010564163327217102, + 0.04303717613220215, + 0.035414524376392365, + 0.005834576208144426, + 0.006780514493584633, + 0.012377621605992317, + 0.011806690134108067, + 0.006988928187638521, + 0.01108904555439949, + 0.04427242651581764, + 0.012003384530544281, + 0.008162282407283783, + 0.00932820700109005, + 0.015490557067096233, + 0.013308292254805565, + 0.012086572125554085, + 0.013953762128949165, + 0.01076232548803091, + 0.009784800931811333, + 0.0024931093212217093, + 0.003562925150617957, + 0.007371247746050358, + 0.006467286963015795, + 0.0025493514258414507, + 0.004622446373105049, + 0.003672653343528509, + 0.003663454670459032, + 0.005225765518844128, + 0.021115127950906754, + 0.00776265561580658, + 0.00878813024610281, + 0.004078591242432594, + 0.011348797008395195, + 0.009694037958979607, + 0.005324725527316332, + 0.011861012317240238, + 0.020370012149214745, + 0.005669108126312494, + 0.006976515520364046, + 0.0042357128113508224, + 0.005272495094686747, + 0.03171815723180771, + 0.006827224045991898, + 0.005610313266515732, + 0.01910814270377159, + 0.0076586161740124226, + 0.013401742093265057, + 0.005800486076623201, + 0.02269841358065605, + 0.013029474765062332, + 0.010296284221112728, + 0.013501410372555256, + 0.036891769617795944, + 0.021602844819426537, + 0.028018124401569366, + 0.016015082597732544, + 0.04018751159310341, + 0.03831800818443298, + 0.024843206629157066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.06205063685774803 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.826167643070221 + ], + "opposite_logprob_tokens": [ + -0.576167643070221 + ], + "logprob_reward": -0.826167643070221, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.826167643070221, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07142336666584015, + 0.03923574462532997, + 0.0879456028342247, + 0.30155378580093384, + 0.032811712473630905, + 0.0514579638838768, + 0.20808421075344086, + 0.15317612886428833, + 0.023882895708084106, + 0.030428621917963028 + ], + "credit_map": [ + 0.00455411896109581, + 0.010845218785107136, + 0.004704905208200216, + 0.007551437243819237, + 0.010752464644610882, + 0.016557717695832253, + 0.00790957361459732, + 0.00854792632162571, + 0.0025484918151050806, + 0.005412667989730835, + 0.0033730012364685535, + 0.004286176059395075, + 0.005831736605614424, + 0.003485386725515127, + 0.008323785848915577, + 0.005974498111754656, + 0.008423642255365849, + 0.00684946496039629, + 0.013606353662908077, + 0.016907591372728348, + 0.012454407289624214, + 0.007151423022150993, + 0.012263323180377483, + 0.010289395228028297, + 0.020915774628520012, + 0.02080029621720314, + 0.03535768389701843, + 0.034008707851171494, + 0.01619819924235344, + 0.031639084219932556, + 0.10913637280464172, + 0.033497657626867294, + 0.002342408988624811, + 0.004409368149936199, + 0.0035391913261264563, + 0.005597064737230539, + 0.004022931680083275, + 0.006217705551534891, + 0.003653710475191474, + 0.0030293334275484085, + 0.0035490081645548344, + 0.005574749317020178, + 0.008548715151846409, + 0.007187041454017162, + 0.006523218937218189, + 0.008583894930779934, + 0.005852161440998316, + 0.005639176815748215, + 0.015690360218286514, + 0.06737557798624039, + 0.015538733452558517, + 0.019253918901085854, + 0.011261090636253357, + 0.03673049435019493, + 0.028446920216083527, + 0.013787107542157173, + 0.01844930090010166, + 0.03362967446446419, + 0.009221663698554039, + 0.012431563809514046, + 0.009230226278305054, + 0.01129714585840702, + 0.04717320576310158, + 0.01174334529787302, + 0.001739766914397478, + 0.004289602395147085, + 0.002113636815920472, + 0.003210202557966113, + 0.0023114013019949198, + 0.0029234217945486307, + 0.003741301828995347, + 0.003553561167791486, + 0.0019010133109986782, + 0.005118358880281448, + 0.001341580180451274, + 0.006881098262965679, + 0.0017285669455304742, + 0.0026255545672029257, + 0.008434327319264412, + 0.0023981211706995964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.826167643070221 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "low-angle, medium close-up static shot from the bedside focuses on a young woman suddenly awakened in her bedroom late at night. Her long, tousled dark brown hair, strands of it clinging to her pale forehead and cheeks from cold sweat, is jolted from her reclining position by an invisible electric shock, her body snapping up quickly and stiffly. Her loose, gray cotton T-shirt, slightly damp with sweat, clings to her violently heaving chest. A ray of cold moonlight filters through the slits in the Venetian blinds, piercing the darkness, illuminating her eyes, wide with fear, and her lips, slightly parted, emitting rapid, hoarse gasps. Her hands instinctively clutch the wrinkled white sheets beneath her, her knuckles white with the strain. Her gaze darts around the dimly lit room, trying to discern the boundary between reality and dream. The high-contrast lighting and the chilling blue hue create an atmosphere of intense tension, fear, and the disorientation of a moment awakened from a nightmare.", + "seed": 9073, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_14/index_8003_seed_9073_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14992286721244455, + "vlm_logprob": -0.14992286721244455, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.14992286721244455, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022026309743523598 + ], + "opposite_logprob_tokens": [ + -5.147026538848877 + ], + "logprob_reward": -0.022026309743523598, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.022026309743523598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12197430431842804, + 0.05908837541937828, + 0.1446262001991272, + 0.07819895446300507, + 0.09995654970407486, + 0.10694176703691483, + 0.07949808984994888, + 0.06904605031013489, + 0.15682262182235718, + 0.08384712040424347 + ], + "credit_map": [ + 0.012541341595351696, + 0.01305878534913063, + 0.014077382162213326, + 0.013847973197698593, + 0.01808883808553219, + 0.0241424348205328, + 0.01431758888065815, + 0.011899958364665508, + 0.008692368865013123, + 0.008576647378504276, + 0.006893371231853962, + 0.006112562492489815, + 0.007590551860630512, + 0.007778477389365435, + 0.008758204989135265, + 0.00468619167804718, + 0.011278064921498299, + 0.015585547313094139, + 0.02343519777059555, + 0.04345499724149704, + 0.01551875751465559, + 0.018634462729096413, + 0.008555488660931587, + 0.008163690567016602, + 0.006932011339813471, + 0.011597936041653156, + 0.007015265058726072, + 0.013048085384070873, + 0.008572829887270927, + 0.011783442460000515, + 0.011320577003061771, + 0.007928803563117981, + 0.009747288189828396, + 0.012678139843046665, + 0.006986604537814856, + 0.013637718744575977, + 0.010882284492254257, + 0.013001651503145695, + 0.01745394617319107, + 0.01556891854852438, + 0.00825105793774128, + 0.010453132912516594, + 0.008408976718783379, + 0.009891437366604805, + 0.021211298182606697, + 0.01642697863280773, + 0.019478805363178253, + 0.012820085510611534, + 0.014112847857177258, + 0.005732971243560314, + 0.010130257345736027, + 0.007549155503511429, + 0.01595255360007286, + 0.012031308375298977, + 0.00858597457408905, + 0.005403015296906233, + 0.007744987495243549, + 0.007850623689591885, + 0.005830235779285431, + 0.008865940384566784, + 0.008653157390654087, + 0.010985574685037136, + 0.009890981949865818, + 0.009224551729857922, + 0.014646516181528568, + 0.024159789085388184, + 0.017285512760281563, + 0.018528148531913757, + 0.015432214364409447, + 0.027459366247057915, + 0.023896576836705208, + 0.015414498746395111, + 0.008642513304948807, + 0.008500871248543262, + 0.008281240239739418, + 0.00908726081252098, + 0.011257419362664223, + 0.013325558044016361, + 0.01502932608127594, + 0.009722931310534477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.022026309743523598 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025101364590227604 + ], + "opposite_logprob_tokens": [ + -6.002510070800781 + ], + "logprob_reward": -0.0025101364590227604, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025101364590227604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21086649596691132, + 0.06510495394468307, + 0.16260570287704468, + 0.13145248591899872, + 0.04443791136145592, + 0.05315714329481125, + 0.11372285336256027, + 0.07257291674613953, + 0.098329097032547, + 0.04775042459368706 + ], + "credit_map": [ + 0.017632102593779564, + 0.022407742217183113, + 0.029645249247550964, + 0.02591203898191452, + 0.033956754952669144, + 0.04340783879160881, + 0.01799170672893524, + 0.019913051277399063, + 0.0070995427668094635, + 0.0060662636533379555, + 0.008519815281033516, + 0.008746214210987091, + 0.007373424246907234, + 0.008991675451397896, + 0.013563422486186028, + 0.004744593054056168, + 0.015717439353466034, + 0.013794011436402798, + 0.025958837941288948, + 0.04169350862503052, + 0.016750089824199677, + 0.017460601404309273, + 0.022516943514347076, + 0.008714266121387482, + 0.013202553614974022, + 0.016635315492749214, + 0.016972050070762634, + 0.013185132294893265, + 0.018299860879778862, + 0.018696848303079605, + 0.02037474513053894, + 0.01408597081899643, + 0.007317268755286932, + 0.005789620336145163, + 0.004773292690515518, + 0.004965797998011112, + 0.006302335299551487, + 0.0062439232133328915, + 0.004213833250105381, + 0.0048318407498300076, + 0.00513389753177762, + 0.006938980892300606, + 0.005986267700791359, + 0.005817594937980175, + 0.00618040282279253, + 0.007280795369297266, + 0.01100946869701147, + 0.004809733480215073, + 0.021195273846387863, + 0.007841178216040134, + 0.015083985403180122, + 0.010658906772732735, + 0.022611111402511597, + 0.01801130548119545, + 0.01066900696605444, + 0.00765208387747407, + 0.0058897132985293865, + 0.009483639150857925, + 0.007498052902519703, + 0.006868059281259775, + 0.01174495555460453, + 0.011028793640434742, + 0.010615428909659386, + 0.009444277733564377, + 0.009990077465772629, + 0.011647408828139305, + 0.013644293881952763, + 0.00950469821691513, + 0.012135995551943779, + 0.01570170745253563, + 0.013675844296813011, + 0.012029072269797325, + 0.005131840705871582, + 0.006439264863729477, + 0.005590393673628569, + 0.005232060793787241, + 0.007390284910798073, + 0.00788805726915598, + 0.006326642353087664, + 0.0037518832832574844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0025101364590227604 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011090539395809174 + ], + "opposite_logprob_tokens": [ + -4.5110907554626465 + ], + "logprob_reward": -0.011090539395809174, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011090539395809174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09653501957654953, + 0.04488496109843254, + 0.12721885740756989, + 0.0668187364935875, + 0.07021648436784744, + 0.15287750959396362, + 0.20599330961704254, + 0.07551363110542297, + 0.10491518676280975, + 0.0550263449549675 + ], + "credit_map": [ + 0.010753858834505081, + 0.0096833361312747, + 0.011565659195184708, + 0.009917640127241611, + 0.016366951167583466, + 0.024244701489806175, + 0.007562798447906971, + 0.006440068129450083, + 0.0055626449175179005, + 0.005635026376694441, + 0.005303473211824894, + 0.006146550178527832, + 0.003808478591963649, + 0.007208237424492836, + 0.007917423732578754, + 0.0033031210768967867, + 0.009648454375565052, + 0.015604387037456036, + 0.01856456883251667, + 0.02877786196768284, + 0.012515544891357422, + 0.016425754874944687, + 0.01587340421974659, + 0.009808868169784546, + 0.006101145874708891, + 0.008364848792552948, + 0.00735849654302001, + 0.008591480553150177, + 0.010005143471062183, + 0.009661812335252762, + 0.009062148630619049, + 0.007673655170947313, + 0.008775093592703342, + 0.00912436842918396, + 0.008413820527493954, + 0.007385871838778257, + 0.009689021855592728, + 0.010548092424869537, + 0.008172900415956974, + 0.008107311092317104, + 0.011619514785706997, + 0.016288748010993004, + 0.014221015386283398, + 0.018888067454099655, + 0.020283808931708336, + 0.021499114111065865, + 0.03691542148590088, + 0.013161810114979744, + 0.012740024365484715, + 0.021426845341920853, + 0.028306307271122932, + 0.02343396097421646, + 0.02829636074602604, + 0.03050393983721733, + 0.04084576666355133, + 0.02044009231030941, + 0.006795219611376524, + 0.007357376161962748, + 0.007766631431877613, + 0.006667661014944315, + 0.011119786649942398, + 0.011307856068015099, + 0.01202468853443861, + 0.012474405579268932, + 0.00944123137742281, + 0.01152527891099453, + 0.013912559486925602, + 0.010620607063174248, + 0.010408548638224602, + 0.017266694456338882, + 0.01735021360218525, + 0.014390043914318085, + 0.007352467626333237, + 0.005337949376553297, + 0.009854174219071865, + 0.005196169950067997, + 0.006230758968740702, + 0.008353457786142826, + 0.009175980463624, + 0.0035253833048045635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.011090539395809174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058464288711548 + ], + "opposite_logprob_tokens": [ + -3.8955845832824707 + ], + "logprob_reward": -0.02058464288711548, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058464288711548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1765722930431366, + 0.04400341212749481, + 0.06353427469730377, + 0.19430214166641235, + 0.09720012545585632, + 0.04774816706776619, + 0.22998037934303284, + 0.03983914107084274, + 0.054919496178627014, + 0.051900576800107956 + ], + "credit_map": [ + 0.016052667051553726, + 0.021189512684941292, + 0.01332086231559515, + 0.012808519415557384, + 0.030578449368476868, + 0.057781774550676346, + 0.0133005790412426, + 0.011539935134351254, + 0.0058363317511975765, + 0.0058670369908213615, + 0.005578749347478151, + 0.0053332713432610035, + 0.003951887600123882, + 0.005860156379640102, + 0.007352159824222326, + 0.004223815631121397, + 0.0054739853367209435, + 0.0067920982837677, + 0.007976016029715538, + 0.013163214549422264, + 0.006311764009296894, + 0.009402913972735405, + 0.008543026633560658, + 0.005871259607374668, + 0.017220357432961464, + 0.022379731759428978, + 0.02173950895667076, + 0.021010447293519974, + 0.03264503553509712, + 0.03191940486431122, + 0.026431787759065628, + 0.020955871790647507, + 0.010335275903344154, + 0.018789874389767647, + 0.007692563813179731, + 0.013487917371094227, + 0.014602089300751686, + 0.013028450310230255, + 0.010220080614089966, + 0.00904387142509222, + 0.00561861926689744, + 0.006937029305845499, + 0.006409527268260717, + 0.006935811135917902, + 0.0064866067841649055, + 0.007231815252453089, + 0.004915956407785416, + 0.0032128016464412212, + 0.048968199640512466, + 0.018462015315890312, + 0.03652701526880264, + 0.02056465670466423, + 0.030908014625310898, + 0.03645218536257744, + 0.02286175638437271, + 0.015236550942063332, + 0.0036936947144567966, + 0.004011704586446285, + 0.004341958090662956, + 0.0033716622274369, + 0.005569641478359699, + 0.006206413730978966, + 0.00653474498540163, + 0.0061093224212527275, + 0.0067229969426989555, + 0.006241022143512964, + 0.006031399592757225, + 0.004695299081504345, + 0.007010871544480324, + 0.008140473626554012, + 0.009554246440529823, + 0.006523185409605503, + 0.0055742645636200905, + 0.006354591343551874, + 0.00727134570479393, + 0.006743308622390032, + 0.007043979596346617, + 0.008247261866927147, + 0.006669127382338047, + 0.003996697720140219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.02058464288711548 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle medium close-up focused on a young woman in a bedroom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have long, tousled dark brown hair with strands clinging to her pale forehead and cheeks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a loose, gray cotton T-shirt that is slightly damp and clinging to her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a ray of cold moonlight filtering through slits in Venetian blinds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934027075767517 + ], + "opposite_logprob_tokens": [ + -0.6934027075767517 + ], + "logprob_reward": -0.6934027075767517, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934027075767517, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07479274272918701, + 0.07108212262392044, + 0.2421957552433014, + 0.1348549872636795, + 0.03904091939330101, + 0.08797423541545868, + 0.050244249403476715, + 0.10960283130407333, + 0.08782916516065598, + 0.10238295793533325 + ], + "credit_map": [ + 0.008276713080704212, + 0.009821322746574879, + 0.0064735268242657185, + 0.00899188220500946, + 0.01407784316688776, + 0.011624110862612724, + 0.008799819275736809, + 0.006727526895701885, + 0.007965713739395142, + 0.011794552206993103, + 0.007363761309534311, + 0.009692584164440632, + 0.006131979636847973, + 0.010956617072224617, + 0.01105567254126072, + 0.006121243815869093, + 0.01684775948524475, + 0.02381281740963459, + 0.042132534086704254, + 0.07238584756851196, + 0.02294769696891308, + 0.029400991275906563, + 0.022687800228595734, + 0.011980326846241951, + 0.010471861809492111, + 0.015988465398550034, + 0.0207330621778965, + 0.017156200483441353, + 0.02033035270869732, + 0.01925842836499214, + 0.017940819263458252, + 0.012975807301700115, + 0.006190354470163584, + 0.005390300881117582, + 0.004410144407302141, + 0.004081672988831997, + 0.004965771455317736, + 0.0058310385793447495, + 0.004271890968084335, + 0.0038997456431388855, + 0.007672635372728109, + 0.007929098792374134, + 0.00857627484947443, + 0.007419736124575138, + 0.01872042752802372, + 0.012557126581668854, + 0.014985574409365654, + 0.010113373398780823, + 0.003605279140174389, + 0.0052261995151638985, + 0.006053040269762278, + 0.007061372976750135, + 0.0046533215790987015, + 0.0078988466411829, + 0.008153523318469524, + 0.007592670153826475, + 0.009878314100205898, + 0.011292838491499424, + 0.01156537514179945, + 0.009038517251610756, + 0.01852233335375786, + 0.016288280487060547, + 0.01438705250620842, + 0.01863013207912445, + 0.013621487654745579, + 0.009188411757349968, + 0.013219702057540417, + 0.009562025777995586, + 0.01312556304037571, + 0.013495178893208504, + 0.009935209527611732, + 0.005681599956005812, + 0.007332948036491871, + 0.014001956209540367, + 0.009959599934518337, + 0.013277258723974228, + 0.011029733344912529, + 0.017509493976831436, + 0.016267210245132446, + 0.013004767708480358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.6934027075767517 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_15/index_tensor([4833, 4833])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_15/index_tensor([4833, 4833])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9d8c35087d1d5022c405f78be4129ba201631ed7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_15/index_tensor([4833, 4833])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "nearly static shot, taken from a medium angle, captures an anime-inspired girl dressed in a classic white sailor suit and a dark blue scarf. With shoulder-length dark brown hair and flowing bangs, she sits sideways at a wooden desk by the window, one hand resting on her chin as she gazes thoughtfully out the window, a hint of yearning in her eyes. The classroom's sliding windows are wide open, and pink and white cherry blossom petals, carried by the afternoon breeze, swirl and drift gently across the window lattice, a few even landing on the corner of her desk and the worn wooden windowsill. Soft sunlight pours in, casting a warm filter over the scene, illuminating the airborne dust and petals, and giving the image a soft, anime-inspired watercolor hue, with desaturated colors and a slightly hazy edge. Behind her, out of focus, other students, gathered in groups of three or four, wearing the same uniforms, their figures and movements, whispering or flipping through books, appear blurred and soft in the frame, forming a vivid background. The entire composition is full of tranquility and poetry, perfectly recreating the warm and slightly sentimental nostalgic atmosphere unique to typical daily life anime.", + "seed": 4876, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_15/index_4833_seed_4876_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01782023753039539, + "vlm_logprob": -0.01782023753039539, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.01782023753039539, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011142647825181484 + ], + "opposite_logprob_tokens": [ + -4.761142730712891 + ], + "logprob_reward": -0.011142647825181484, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011142647825181484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3427935242652893, + 0.21100372076034546, + 0.13419412076473236, + 0.04170617833733559, + 0.07413889467716217, + 0.053327348083257675, + 0.024141306057572365, + 0.04178461804986, + 0.019718576222658157, + 0.05719175934791565 + ], + "credit_map": [ + 0.0704682469367981, + 0.06502391397953033, + 0.06912735849618912, + 0.03135053068399429, + 0.0377848818898201, + 0.027507003396749496, + 0.02120712585747242, + 0.02032446675002575, + 0.06377788633108139, + 0.019457140937447548, + 0.039548929780721664, + 0.011540907435119152, + 0.019937647506594658, + 0.01911774091422558, + 0.010758491232991219, + 0.026864970102906227, + 0.018360931426286697, + 0.022031176835298538, + 0.024363767355680466, + 0.012297985143959522, + 0.01979314349591732, + 0.010048570111393929, + 0.013709005899727345, + 0.013589547947049141, + 0.010794073343276978, + 0.00797192007303238, + 0.006489927414804697, + 0.004607695154845715, + 0.003182950196787715, + 0.002465164056047797, + 0.0031561567448079586, + 0.0030382899567484856, + 0.017714519053697586, + 0.011609519831836224, + 0.021344659850001335, + 0.00796546135097742, + 0.00477495277300477, + 0.0038164090365171432, + 0.003526961663737893, + 0.0033864059951156378, + 0.00865955650806427, + 0.00814057420939207, + 0.013565839268267155, + 0.007247179746627808, + 0.004204008728265762, + 0.004193977452814579, + 0.00326156965456903, + 0.004054639954119921, + 0.006202000193297863, + 0.0033621562179178, + 0.0029090798925608397, + 0.002349621383473277, + 0.002034870209172368, + 0.00228507979772985, + 0.001919479575008154, + 0.0030790187884122133, + 0.008523579686880112, + 0.007598306518048048, + 0.005374610424041748, + 0.006643411237746477, + 0.004037036094814539, + 0.002972797490656376, + 0.003942538984119892, + 0.0026923338882625103, + 0.002751054707914591, + 0.002380818361416459, + 0.002488641068339348, + 0.0020092695485800505, + 0.0030533289536833763, + 0.0023055688943713903, + 0.0018458447884768248, + 0.002884051762521267, + 0.007860851474106312, + 0.007861501537263393, + 0.006818548310548067, + 0.007687907200306654, + 0.008399982005357742, + 0.0062200650572776794, + 0.006071918644011021, + 0.006270986516028643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011142647825181484 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026364075019955635 + ], + "opposite_logprob_tokens": [ + -3.6513640880584717 + ], + "logprob_reward": -0.026364075019955635, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026364075019955635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20198018848896027, + 0.4485219120979309, + 0.11564485728740692, + 0.03899385780096054, + 0.042229555547237396, + 0.01704223081469536, + 0.04006786271929741, + 0.02162342146039009, + 0.024357177317142487, + 0.04953894019126892 + ], + "credit_map": [ + 0.048732977360486984, + 0.040282249450683594, + 0.02887004241347313, + 0.02120351418852806, + 0.019324708729982376, + 0.014827171340584755, + 0.014281945303082466, + 0.014457576908171177, + 0.06462495028972626, + 0.04267996922135353, + 0.08571097999811172, + 0.021083127707242966, + 0.10819899290800095, + 0.030115759000182152, + 0.02589399740099907, + 0.07021413743495941, + 0.021727677434682846, + 0.019450072199106216, + 0.02033594436943531, + 0.014034043066203594, + 0.010890037752687931, + 0.007615427020937204, + 0.009386799298226833, + 0.01220485009253025, + 0.009578940458595753, + 0.007385139353573322, + 0.005936427973210812, + 0.003936948720365763, + 0.003683181945234537, + 0.0029329361859709024, + 0.0029121097177267075, + 0.0026281741447746754, + 0.0049476828426122665, + 0.005918377079069614, + 0.005559212528169155, + 0.0042718579061329365, + 0.006755937822163105, + 0.0046318150125443935, + 0.005411537829786539, + 0.0047331335954368114, + 0.0037217740900814533, + 0.003192267147824168, + 0.0029774997383356094, + 0.0015997319715097547, + 0.0017111116321757436, + 0.0014843391254544258, + 0.0009250228176824749, + 0.0014304843498393893, + 0.011584256775677204, + 0.007630895357578993, + 0.004721894394606352, + 0.0038180018309503794, + 0.0030382906552404165, + 0.002714238129556179, + 0.0034999048803001642, + 0.0030603802297264338, + 0.003075049491599202, + 0.003761625848710537, + 0.0035502964165061712, + 0.002222370123490691, + 0.0022751560900360346, + 0.0020282778423279524, + 0.0018824351718649268, + 0.002828210359439254, + 0.006193336099386215, + 0.003290515625849366, + 0.002928141737356782, + 0.0016149251023307443, + 0.002457620343193412, + 0.0022494352888315916, + 0.0019916484598070383, + 0.0036315566394478083, + 0.013781181536614895, + 0.007203830871731043, + 0.0057637752033770084, + 0.004464327823370695, + 0.003517518052831292, + 0.003566157538443804, + 0.004882881883531809, + 0.006359267048537731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.026364075019955635 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818905957043171 + ], + "opposite_logprob_tokens": [ + -4.018188953399658 + ], + "logprob_reward": -0.01818905957043171, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818905957043171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1727265864610672, + 0.45761603116989136, + 0.062029413878917694, + 0.04543066769838333, + 0.06869366765022278, + 0.03337879478931427, + 0.032953035086393356, + 0.030889388173818588, + 0.030354894697666168, + 0.06592753529548645 + ], + "credit_map": [ + 0.021268781274557114, + 0.019908810034394264, + 0.0248198751360178, + 0.015509450808167458, + 0.03824986144900322, + 0.021746525540947914, + 0.016695309430360794, + 0.014527958817780018, + 0.050691716372966766, + 0.029350684955716133, + 0.17903169989585876, + 0.02164607308804989, + 0.038757048547267914, + 0.029106788337230682, + 0.01689075119793415, + 0.09214124083518982, + 0.00894202757626772, + 0.01096892636269331, + 0.00913166906684637, + 0.007239782251417637, + 0.009054583497345448, + 0.007180691231042147, + 0.004693672526627779, + 0.00481805857270956, + 0.005886395927518606, + 0.005182925146073103, + 0.011826707050204277, + 0.0036355587653815746, + 0.005618632771074772, + 0.004972141236066818, + 0.003004188649356365, + 0.005304114427417517, + 0.008181321434676647, + 0.006842486560344696, + 0.015208250842988491, + 0.004422432743012905, + 0.005563100799918175, + 0.00886806845664978, + 0.0033957110717892647, + 0.01621229387819767, + 0.005904139950871468, + 0.00746315298601985, + 0.0068754456005990505, + 0.0037725307047367096, + 0.0023652694653719664, + 0.002185087651014328, + 0.002060485538095236, + 0.0027526835910975933, + 0.0040086181834340096, + 0.004490288905799389, + 0.003799096215516329, + 0.004527907352894545, + 0.0043081617914140224, + 0.004162975121289492, + 0.00312251434661448, + 0.004533472005277872, + 0.004435872659087181, + 0.004975196905434132, + 0.004478897899389267, + 0.004251271951943636, + 0.0033792941831052303, + 0.002083697123453021, + 0.004124064464122057, + 0.0031610920559614897, + 0.008139519020915031, + 0.004840000066906214, + 0.003913169726729393, + 0.0022418517619371414, + 0.002634419361129403, + 0.002673570765182376, + 0.0024048767518252134, + 0.0035074849147349596, + 0.019556410610675812, + 0.009500866755843163, + 0.007023095153272152, + 0.005621904507279396, + 0.006034377031028271, + 0.0058617787435650826, + 0.005150071810930967, + 0.0071790339425206184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.01818905957043171 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029789181426167488 + ], + "opposite_logprob_tokens": [ + -3.5297892093658447 + ], + "logprob_reward": -0.029789181426167488, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029789181426167488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20959539711475372, + 0.14044959843158722, + 0.11264251172542572, + 0.041021496057510376, + 0.06954497843980789, + 0.04779407009482384, + 0.038602281361818314, + 0.07456779479980469, + 0.0891391783952713, + 0.17664267122745514 + ], + "credit_map": [ + 0.040581610053777695, + 0.022752469405531883, + 0.054419174790382385, + 0.01177294459193945, + 0.02954481914639473, + 0.02237590029835701, + 0.01155809871852398, + 0.016590381041169167, + 0.020892567932605743, + 0.01183299906551838, + 0.02270873822271824, + 0.006304978858679533, + 0.03854750841856003, + 0.015513353049755096, + 0.008068744093179703, + 0.016580719500780106, + 0.024677986279129982, + 0.011653455905616283, + 0.03163056820631027, + 0.0064551252871751785, + 0.011310486122965813, + 0.007602000143378973, + 0.012402433902025223, + 0.006910461001098156, + 0.006683325860649347, + 0.005791186820715666, + 0.008734156377613544, + 0.003526396816596389, + 0.003341234987601638, + 0.003487891750410199, + 0.0033889992628246546, + 0.006068303249776363, + 0.014706107787787914, + 0.006829585414379835, + 0.016649117693305016, + 0.004371044225990772, + 0.008527823723852634, + 0.006786966230720282, + 0.005674262065440416, + 0.006000072695314884, + 0.009892680682241917, + 0.007052101660519838, + 0.007691640406847, + 0.004518176894634962, + 0.006282166577875614, + 0.005236952565610409, + 0.0029979862738400698, + 0.004122364800423384, + 0.007965829223394394, + 0.004894470330327749, + 0.004114663228392601, + 0.003606625134125352, + 0.004494195803999901, + 0.005011274944990873, + 0.004563805181533098, + 0.003951418679207563, + 0.009713069535791874, + 0.014583547599613667, + 0.01294734701514244, + 0.0062493616715073586, + 0.007410393096506596, + 0.006350519601255655, + 0.0082330834120512, + 0.009080471470952034, + 0.024743199348449707, + 0.012935733422636986, + 0.011612653732299805, + 0.009837953373789787, + 0.010144777595996857, + 0.007152004633098841, + 0.0068389009684324265, + 0.005873952992260456, + 0.021440831944346428, + 0.024810098111629486, + 0.017666876316070557, + 0.01844252459704876, + 0.021651355549693108, + 0.021011603996157646, + 0.019903043285012245, + 0.03171631693840027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.029789181426167488 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036162238102406263 + ], + "opposite_logprob_tokens": [ + -5.6286163330078125 + ], + "logprob_reward": -0.0036162238102406263, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036162238102406263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13656947016716003, + 0.48689961433410645, + 0.0751003548502922, + 0.03889651596546173, + 0.07208201289176941, + 0.018656982108950615, + 0.04962194710969925, + 0.029574871063232422, + 0.03557811677455902, + 0.0570201613008976 + ], + "credit_map": [ + 0.027140803635120392, + 0.019854119047522545, + 0.02146989293396473, + 0.01502256840467453, + 0.013129997067153454, + 0.013517103157937527, + 0.010038411244750023, + 0.016396570950746536, + 0.09977826476097107, + 0.037633128464221954, + 0.19906452298164368, + 0.0227534007281065, + 0.019976245239377022, + 0.020213471725583076, + 0.01640165038406849, + 0.07107891142368317, + 0.011530885472893715, + 0.008413676172494888, + 0.01743687130510807, + 0.0065063792280852795, + 0.007180679589509964, + 0.007734737824648619, + 0.004637854639440775, + 0.011659265495836735, + 0.0045154462568461895, + 0.005091313272714615, + 0.006836024112999439, + 0.00415307329967618, + 0.004537218250334263, + 0.004320766311138868, + 0.0040629166178405285, + 0.005379760637879372, + 0.02015562169253826, + 0.008172429166734219, + 0.02252592332661152, + 0.005460118874907494, + 0.0038160779513418674, + 0.005597222130745649, + 0.002857264829799533, + 0.003497352823615074, + 0.002578031737357378, + 0.0030729863792657852, + 0.0038023267406970263, + 0.00185857352335006, + 0.0016147471033036709, + 0.0019650040194392204, + 0.0014677296858280897, + 0.0022975807078182697, + 0.014371807686984539, + 0.006613036151975393, + 0.006262603681534529, + 0.004984755534678698, + 0.005441756919026375, + 0.0037106822710484266, + 0.0039256601594388485, + 0.004311638884246349, + 0.006186584010720253, + 0.004240059293806553, + 0.00441865436732769, + 0.0028845348861068487, + 0.002621386433020234, + 0.0030516437254846096, + 0.002098712371662259, + 0.004073294810950756, + 0.008297018706798553, + 0.004525790456682444, + 0.005082421004772186, + 0.0040268320590257645, + 0.0035411117132753134, + 0.00318166334182024, + 0.00264670024625957, + 0.0042765773832798, + 0.013956790789961815, + 0.008001783862709999, + 0.005022261291742325, + 0.0056512122973799706, + 0.006586890667676926, + 0.00589019013568759, + 0.005158186424523592, + 0.006752840708941221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0036162238102406263 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "nearly static shot, taken from a medium angle, captures an anime-inspired girl dressed in a classic white sailor suit and a dark blue scarf. With shoulder-length dark brown hair and flowing bangs, she sits sideways at a wooden desk by the window, one hand resting on her chin as she gazes thoughtfully out the window, a hint of yearning in her eyes. The classroom's sliding windows are wide open, and pink and white cherry blossom petals, carried by the afternoon breeze, swirl and drift gently across the window lattice, a few even landing on the corner of her desk and the worn wooden windowsill. Soft sunlight pours in, casting a warm filter over the scene, illuminating the airborne dust and petals, and giving the image a soft, anime-inspired watercolor hue, with desaturated colors and a slightly hazy edge. Behind her, out of focus, other students, gathered in groups of three or four, wearing the same uniforms, their figures and movements, whispering or flipping through books, appear blurred and soft in the frame, forming a vivid background. The entire composition is full of tranquility and poetry, perfectly recreating the warm and slightly sentimental nostalgic atmosphere unique to typical daily life anime.", + "seed": 4877, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_15/index_4833_seed_4877_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.011296666855923832, + "vlm_logprob": -0.011296666855923832, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.011296666855923832, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003671573242172599 + ], + "opposite_logprob_tokens": [ + -6.003671646118164 + ], + "logprob_reward": -0.003671573242172599, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.003671573242172599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22019055485725403, + 0.2672051191329956, + 0.061337895691394806, + 0.06638608127832413, + 0.13888327777385712, + 0.10027310997247696, + 0.04370185360312462, + 0.02025192603468895, + 0.049483004957437515, + 0.03228718787431717 + ], + "credit_map": [ + 0.03398324176669121, + 0.023724034428596497, + 0.04162777587771416, + 0.028249414637684822, + 0.02617764286696911, + 0.029311243444681168, + 0.016532249748706818, + 0.020584946498274803, + 0.08549796789884567, + 0.014381777495145798, + 0.10037852823734283, + 0.014415265060961246, + 0.012823948636651039, + 0.015892451629042625, + 0.010612566955387592, + 0.01320258155465126, + 0.01486118882894516, + 0.007672494743019342, + 0.01648850366473198, + 0.005195623263716698, + 0.004651338327676058, + 0.005967721808701754, + 0.0032016788609325886, + 0.0032993468921631575, + 0.018839826807379723, + 0.005565395578742027, + 0.016103727743029594, + 0.0047094570472836494, + 0.004485094454139471, + 0.0074999709613621235, + 0.0033182334154844284, + 0.0058643752709031105, + 0.04469478502869606, + 0.008521342650055885, + 0.050780680030584335, + 0.010075937956571579, + 0.005013147834688425, + 0.008186476305127144, + 0.004620957653969526, + 0.006989950314164162, + 0.023181848227977753, + 0.013980229385197163, + 0.028505047783255577, + 0.011026403866708279, + 0.0069782184436917305, + 0.0055798073299229145, + 0.004125871229916811, + 0.006895680911839008, + 0.011778599582612514, + 0.003940367139875889, + 0.011590026319026947, + 0.0037265559658408165, + 0.004081937484443188, + 0.0027765571139752865, + 0.002618043916299939, + 0.003189763519912958, + 0.004895901307463646, + 0.0026311364490538836, + 0.003683985909447074, + 0.0021013703662902117, + 0.002354321302846074, + 0.0013388943625614047, + 0.0017080185934901237, + 0.0015382966957986355, + 0.0075668818317353725, + 0.0052945902571082115, + 0.005305089987814426, + 0.004603353329002857, + 0.008448588661849499, + 0.005292993970215321, + 0.00499455351382494, + 0.007976949214935303, + 0.0048569319769740105, + 0.0037045064382255077, + 0.007137891836464405, + 0.0032900325022637844, + 0.003513611853122711, + 0.003835978452116251, + 0.0019049191614612937, + 0.00404331274330616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.003671573242172599 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547596357762814 + ], + "opposite_logprob_tokens": [ + -4.387547492980957 + ], + "logprob_reward": -0.012547596357762814, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547596357762814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09689054638147354, + 0.28599148988723755, + 0.14771926403045654, + 0.027744611725211143, + 0.21407470107078552, + 0.045598361641168594, + 0.061537355184555054, + 0.018520353361964226, + 0.08725026994943619, + 0.014673056080937386 + ], + "credit_map": [ + 0.011504127644002438, + 0.010996111668646336, + 0.008890090510249138, + 0.012453579343855381, + 0.020761143416166306, + 0.009320682846009731, + 0.011995294131338596, + 0.010969512164592743, + 0.10252858698368073, + 0.012021547183394432, + 0.11945246905088425, + 0.017015332356095314, + 0.007946568541228771, + 0.009292486123740673, + 0.008221001364290714, + 0.009513483382761478, + 0.03577903285622597, + 0.013646378181874752, + 0.04897546395659447, + 0.01282291579991579, + 0.01004980318248272, + 0.008864883333444595, + 0.007582675199955702, + 0.009998115710914135, + 0.005469741299748421, + 0.0027420048136264086, + 0.00758743891492486, + 0.0023256202694028616, + 0.0025825928896665573, + 0.0023304843343794346, + 0.0017529733013361692, + 0.002953755669295788, + 0.0782824382185936, + 0.00942695327103138, + 0.08750176429748535, + 0.01167158130556345, + 0.0067202253267169, + 0.006716516800224781, + 0.005633618216961622, + 0.00812160037457943, + 0.010091549716889858, + 0.007323399651795626, + 0.009409460239112377, + 0.004533377476036549, + 0.0045042564161121845, + 0.002987446030601859, + 0.0021347554866224527, + 0.004614114295691252, + 0.018948959186673164, + 0.003584902733564377, + 0.02124110423028469, + 0.0043136365711688995, + 0.004808019381016493, + 0.0029967196751385927, + 0.0025130698923021555, + 0.003130942117422819, + 0.0028308723121881485, + 0.0031451238319277763, + 0.002466514939442277, + 0.002397555625066161, + 0.00269496813416481, + 0.0015271930024027824, + 0.001856601214967668, + 0.0016015234868973494, + 0.0111548425629735, + 0.010635211132466793, + 0.01827850006520748, + 0.009717277251183987, + 0.009508008137345314, + 0.01314233336597681, + 0.005239661317318678, + 0.009574433788657188, + 0.0023799361661076546, + 0.0014511067420244217, + 0.002947760745882988, + 0.001644242205657065, + 0.002033940749242902, + 0.0016719767590984702, + 0.0008593188831582665, + 0.0016847745282575488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.012547596357762814 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577870309352875 + ], + "opposite_logprob_tokens": [ + -3.895577907562256 + ], + "logprob_reward": -0.020577870309352875, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577870309352875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32162582874298096, + 0.1818332076072693, + 0.058322034776210785, + 0.043907951563596725, + 0.21340399980545044, + 0.03735128417611122, + 0.051239922642707825, + 0.017628485336899757, + 0.026755675673484802, + 0.04793163016438484 + ], + "credit_map": [ + 0.11010609567165375, + 0.015635376796126366, + 0.13074293732643127, + 0.020226115360856056, + 0.009965736418962479, + 0.011728763580322266, + 0.009208491072058678, + 0.014012284576892853, + 0.04940643161535263, + 0.018950050696730614, + 0.057726528495550156, + 0.017840804532170296, + 0.009282550774514675, + 0.011914855800569057, + 0.00623568519949913, + 0.010476282797753811, + 0.006273502949625254, + 0.011318174190819263, + 0.00471829529851675, + 0.008540287613868713, + 0.00977062713354826, + 0.00584064144641161, + 0.006392552051693201, + 0.005467948969453573, + 0.011222642846405506, + 0.0035046092234551907, + 0.00949348509311676, + 0.002401177305728197, + 0.0027384536806493998, + 0.005549502559006214, + 0.0034312279894948006, + 0.005566850304603577, + 0.08009317517280579, + 0.00963712390512228, + 0.08430133759975433, + 0.011673631146550179, + 0.007253486663103104, + 0.006821363233029842, + 0.006166175939142704, + 0.007457689382135868, + 0.0074951197020709515, + 0.004207474645227194, + 0.01006245519965887, + 0.0038063989486545324, + 0.003510091220960021, + 0.0027649791445583105, + 0.0024150251410901546, + 0.003089736681431532, + 0.01585766114294529, + 0.0034489864483475685, + 0.017261048778891563, + 0.004697880242019892, + 0.003360011614859104, + 0.0018837244715541601, + 0.0024609663523733616, + 0.0022696396335959435, + 0.0035087161231786013, + 0.001739391591399908, + 0.0035733284894376993, + 0.0021496086847037077, + 0.0016991771990433335, + 0.002209571423009038, + 0.0011274113785475492, + 0.0016212803311645985, + 0.0033325287513434887, + 0.002776946872472763, + 0.00398083683103323, + 0.002306384965777397, + 0.005579325836151838, + 0.002667019609361887, + 0.002397910924628377, + 0.0037147202529013157, + 0.012328926473855972, + 0.00448104040697217, + 0.010773451998829842, + 0.00471430504694581, + 0.003634255612269044, + 0.00457628071308136, + 0.002964320592582226, + 0.004459049552679062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.020577870309352875 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018174076452851295 + ], + "opposite_logprob_tokens": [ + -4.018174171447754 + ], + "logprob_reward": -0.018174076452851295, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018174076452851295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19869136810302734, + 0.44495153427124023, + 0.06778702884912491, + 0.039043813943862915, + 0.023572774603962898, + 0.1381298154592514, + 0.04004848003387451, + 0.00805563572794199, + 0.021422170102596283, + 0.018297389149665833 + ], + "credit_map": [ + 0.043328363448381424, + 0.025218574330210686, + 0.037254996597766876, + 0.027695750817656517, + 0.016416944563388824, + 0.013994061388075352, + 0.015604828484356403, + 0.01917785219848156, + 0.15247805416584015, + 0.022964945062994957, + 0.18198733031749725, + 0.03072442300617695, + 0.012616684660315514, + 0.014057222753763199, + 0.012917885556817055, + 0.017204981297254562, + 0.016706662252545357, + 0.008666795678436756, + 0.01996074803173542, + 0.0060921600088477135, + 0.005476807709783316, + 0.003676741849631071, + 0.003138019936159253, + 0.004069094080477953, + 0.008448761887848377, + 0.003939557820558548, + 0.00808541290462017, + 0.0031829425133764744, + 0.0033048917539417744, + 0.004972626455128193, + 0.0022550630383193493, + 0.004854557570070028, + 0.004147377796471119, + 0.0026300325989723206, + 0.0046361759305000305, + 0.0017607505433261395, + 0.0032305214554071426, + 0.0021132996771484613, + 0.0023011646699160337, + 0.002753452630713582, + 0.043595410883426666, + 0.007996884174644947, + 0.05243505910038948, + 0.008203827776014805, + 0.006891131866723299, + 0.006903655827045441, + 0.004802752286195755, + 0.007301095873117447, + 0.005921124015003443, + 0.006313401274383068, + 0.007179305423051119, + 0.005296783521771431, + 0.005280761048197746, + 0.0030022580176591873, + 0.0035361498594284058, + 0.003518696641549468, + 0.0013007543748244643, + 0.0008752884459681809, + 0.0015999171882867813, + 0.0008260122267529368, + 0.0008958136313594878, + 0.0008499035611748695, + 0.0006379705737344921, + 0.00106997590046376, + 0.0028138570487499237, + 0.0028357869014143944, + 0.0037914905697107315, + 0.002776471897959709, + 0.003405691124498844, + 0.00208907388150692, + 0.0016254773363471031, + 0.0020843204110860825, + 0.004672872833907604, + 0.001380257890559733, + 0.004440160933881998, + 0.0018807383021339774, + 0.0013357347343116999, + 0.0017000057268887758, + 0.001021335949189961, + 0.0018662832444533706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.018174076452851295 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015122179174795747 + ], + "opposite_logprob_tokens": [ + -6.501512050628662 + ], + "logprob_reward": -0.0015122179174795747, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015122179174795747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19134363532066345, + 0.3525870740413666, + 0.07572539895772934, + 0.03110058605670929, + 0.16742052137851715, + 0.03215906396508217, + 0.07650640606880188, + 0.011835599318146706, + 0.04595896601676941, + 0.015362740494310856 + ], + "credit_map": [ + 0.062345825135707855, + 0.012296047993004322, + 0.07280789315700531, + 0.015075096860527992, + 0.007960619404911995, + 0.0070586176589131355, + 0.005560819525271654, + 0.008238716050982475, + 0.12635885179042816, + 0.014189819805324078, + 0.14938896894454956, + 0.020979762077331543, + 0.007716299034655094, + 0.011243820190429688, + 0.009845336899161339, + 0.012864206917583942, + 0.010699694976210594, + 0.016620976850390434, + 0.007771471980959177, + 0.01218288205564022, + 0.00902177207171917, + 0.007601260207593441, + 0.005612111184746027, + 0.006215231493115425, + 0.005108959041535854, + 0.0032682588789612055, + 0.007869327440857887, + 0.0026822409126907587, + 0.0025560962967574596, + 0.004114978015422821, + 0.002171984175220132, + 0.003328742692247033, + 0.06001653894782066, + 0.007762711960822344, + 0.06569778919219971, + 0.009306494146585464, + 0.007010067347437143, + 0.00570317916572094, + 0.005219561979174614, + 0.006704178638756275, + 0.005235766060650349, + 0.003261542646214366, + 0.008651512674987316, + 0.0042695351876318455, + 0.002865963149815798, + 0.002307206392288208, + 0.0018438317347317934, + 0.003723707515746355, + 0.021629607304930687, + 0.005906001664698124, + 0.026337919756770134, + 0.005789333954453468, + 0.007324295584112406, + 0.002643006853759289, + 0.0034954319708049297, + 0.003380805253982544, + 0.0017850205767899752, + 0.0012535768328234553, + 0.001473444513976574, + 0.0012273233151063323, + 0.002336694160476327, + 0.0010291493963450193, + 0.0014089280739426613, + 0.0013214627979323268, + 0.006134864874184132, + 0.004079751204699278, + 0.009879772551357746, + 0.0038113549817353487, + 0.006392377428710461, + 0.007207914255559444, + 0.003413077211007476, + 0.005039854906499386, + 0.002559433924034238, + 0.001696509774774313, + 0.0034228097647428513, + 0.0016742803854867816, + 0.0014794182498008013, + 0.0017509984318166971, + 0.0009828993352130055, + 0.0017963902791962028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0015122179174795747 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "nearly static shot, taken from a medium angle, captures an anime-inspired girl dressed in a classic white sailor suit and a dark blue scarf. With shoulder-length dark brown hair and flowing bangs, she sits sideways at a wooden desk by the window, one hand resting on her chin as she gazes thoughtfully out the window, a hint of yearning in her eyes. The classroom's sliding windows are wide open, and pink and white cherry blossom petals, carried by the afternoon breeze, swirl and drift gently across the window lattice, a few even landing on the corner of her desk and the worn wooden windowsill. Soft sunlight pours in, casting a warm filter over the scene, illuminating the airborne dust and petals, and giving the image a soft, anime-inspired watercolor hue, with desaturated colors and a slightly hazy edge. Behind her, out of focus, other students, gathered in groups of three or four, wearing the same uniforms, their figures and movements, whispering or flipping through books, appear blurred and soft in the frame, forming a vivid background. The entire composition is full of tranquility and poetry, perfectly recreating the warm and slightly sentimental nostalgic atmosphere unique to typical daily life anime.", + "seed": 4878, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_15/index_4833_seed_4878_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.007609462202526629, + "vlm_logprob": -0.007609462202526629, + "vlm_yes_no_margin": 5.25, + "vlm_token_credit": -0.007609462202526629, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002430938882753253 + ], + "opposite_logprob_tokens": [ + -6.3774309158325195 + ], + "logprob_reward": -0.002430938882753253, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.002430938882753253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17606833577156067, + 0.22507286071777344, + 0.13262225687503815, + 0.06487400084733963, + 0.054303769022226334, + 0.05530230328440666, + 0.07209952175617218, + 0.10929005593061447, + 0.05231279134750366, + 0.05805407091975212 + ], + "credit_map": [ + 0.020021889358758926, + 0.023915650323033333, + 0.021099163219332695, + 0.026874270290136337, + 0.039919640868902206, + 0.014019861817359924, + 0.013453367166221142, + 0.016764504835009575, + 0.055765289813280106, + 0.026630710810422897, + 0.07400060445070267, + 0.020869988948106766, + 0.013718320988118649, + 0.010057192295789719, + 0.009602907113730907, + 0.014427823945879936, + 0.032710421830415726, + 0.01003600936383009, + 0.044556379318237305, + 0.009483439847826958, + 0.013505702838301659, + 0.006517480593174696, + 0.007417819928377867, + 0.008395003154873848, + 0.007097043562680483, + 0.009740401990711689, + 0.009839030914008617, + 0.012268491089344025, + 0.0075199296697974205, + 0.006051748525351286, + 0.005522954743355513, + 0.006834404543042183, + 0.006771401967853308, + 0.006658533122390509, + 0.013638230971992016, + 0.005036162678152323, + 0.008873039856553078, + 0.0042055342346429825, + 0.005336436443030834, + 0.0037844290491193533, + 0.006404734216630459, + 0.008960030972957611, + 0.007035584654659033, + 0.010822067968547344, + 0.005091889761388302, + 0.005334849935024977, + 0.004357336089015007, + 0.007295813877135515, + 0.009013514965772629, + 0.010701129212975502, + 0.008315392769873142, + 0.00955145712941885, + 0.013945459388196468, + 0.005647681187838316, + 0.007890232838690281, + 0.007034654729068279, + 0.011935144662857056, + 0.017887284979224205, + 0.021755401045084, + 0.012134881690144539, + 0.01666823960840702, + 0.007968168705701828, + 0.011694567278027534, + 0.009246372617781162, + 0.00577544653788209, + 0.00565439322963357, + 0.010502466931939125, + 0.0062155285850167274, + 0.009018217213451862, + 0.004933892283588648, + 0.005574121605604887, + 0.00463872542604804, + 0.006695677991956472, + 0.00948483869433403, + 0.008083140477538109, + 0.010072405450046062, + 0.006535234861075878, + 0.005824220832437277, + 0.00562418345361948, + 0.005734369158744812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.002430938882753253 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644895628094673 + ], + "opposite_logprob_tokens": [ + -4.758645057678223 + ], + "logprob_reward": -0.008644895628094673, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644895628094673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19065505266189575, + 0.1761474609375, + 0.17008885741233826, + 0.05475899204611778, + 0.03251440450549126, + 0.021966736763715744, + 0.11257269978523254, + 0.06882330030202866, + 0.05995935946702957, + 0.11251308023929596 + ], + "credit_map": [ + 0.026684636250138283, + 0.02760639227926731, + 0.022996652871370316, + 0.01808350533246994, + 0.018543241545557976, + 0.016686005517840385, + 0.015120523050427437, + 0.04493410512804985, + 0.036187004297971725, + 0.015716996043920517, + 0.06254084408283234, + 0.017562996596097946, + 0.012900236994028091, + 0.010571921244263649, + 0.007308720611035824, + 0.013358766213059425, + 0.029174285009503365, + 0.028094936162233353, + 0.042401112616062164, + 0.020459989085793495, + 0.017956538125872612, + 0.012039866298437119, + 0.011950681917369366, + 0.008011474274098873, + 0.007661911658942699, + 0.011815128847956657, + 0.007714769337326288, + 0.007051828783005476, + 0.007013791240751743, + 0.004156298469752073, + 0.004625794943422079, + 0.004719475284218788, + 0.004517573397606611, + 0.005094955675303936, + 0.006516106426715851, + 0.0036755790933966637, + 0.004210126586258411, + 0.003006507409736514, + 0.002926700282841921, + 0.002566857961937785, + 0.0022465656511485577, + 0.0029401921201497316, + 0.0032865453977137804, + 0.0024433773942291737, + 0.003748777089640498, + 0.0021584772039204836, + 0.0017066970467567444, + 0.00343610648997128, + 0.010079264640808105, + 0.018330655992031097, + 0.014490079134702682, + 0.024989033117890358, + 0.015766600146889687, + 0.007323656231164932, + 0.010686645284295082, + 0.010906770825386047, + 0.007123628165572882, + 0.014581461437046528, + 0.012209904380142689, + 0.00809085089713335, + 0.009952638298273087, + 0.005118840374052525, + 0.005849176086485386, + 0.005896801594644785, + 0.0065201339311897755, + 0.007293563336133957, + 0.010358989238739014, + 0.009383981116116047, + 0.010309061035513878, + 0.004589688498526812, + 0.005891317501664162, + 0.005612629931420088, + 0.014653309248387814, + 0.01466927956789732, + 0.01568121276795864, + 0.023045526817440987, + 0.012707714922726154, + 0.009265100583434105, + 0.011071627028286457, + 0.01141931302845478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.008644895628094673 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419447734951973 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.01419447734951973, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419447734951973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16212999820709229, + 0.43252134323120117, + 0.11009665578603745, + 0.06711999326944351, + 0.036744847893714905, + 0.020006416365504265, + 0.05088971182703972, + 0.041877228766679764, + 0.02976846508681774, + 0.04884539544582367 + ], + "credit_map": [ + 0.019595347344875336, + 0.017384834587574005, + 0.021799234673380852, + 0.02174178883433342, + 0.02951969765126705, + 0.014876879751682281, + 0.014465883374214172, + 0.02274632640182972, + 0.12157613784074783, + 0.03089432418346405, + 0.18481355905532837, + 0.02411227859556675, + 0.013071689754724503, + 0.014843838289380074, + 0.011634545400738716, + 0.03157494217157364, + 0.024526532739400864, + 0.0131076043471694, + 0.03245249390602112, + 0.010251817293465137, + 0.009733250364661217, + 0.005031110718846321, + 0.008768264204263687, + 0.006225590594112873, + 0.008650080300867558, + 0.009211847558617592, + 0.01332947053015232, + 0.011685214005410671, + 0.006294020917266607, + 0.005361612886190414, + 0.006653480231761932, + 0.0059342700988054276, + 0.004769939463585615, + 0.0049215094186365604, + 0.009142481721937656, + 0.00455167843028903, + 0.003496622433885932, + 0.002802143106237054, + 0.0035376567393541336, + 0.0035228170454502106, + 0.002929131733253598, + 0.00326232030056417, + 0.002534375758841634, + 0.002777999732643366, + 0.0022391355596482754, + 0.0021556122228503227, + 0.0012821001000702381, + 0.002825742354616523, + 0.004789132159203291, + 0.00805577915161848, + 0.006626122631132603, + 0.0068245334550738335, + 0.011342003010213375, + 0.0035562068223953247, + 0.004962945356965065, + 0.004732989706099033, + 0.005087470170110464, + 0.007663758471608162, + 0.004700133576989174, + 0.005825706757605076, + 0.0059123472310602665, + 0.003895464586094022, + 0.00417025201022625, + 0.004622095730155706, + 0.0032564373686909676, + 0.004885655362159014, + 0.004340019077062607, + 0.005252968519926071, + 0.0033916763495653868, + 0.002352244919165969, + 0.0033457614481449127, + 0.002943703206256032, + 0.005654704757034779, + 0.005612403620034456, + 0.009247073903679848, + 0.006792653817683458, + 0.006417183671146631, + 0.0038524200208485126, + 0.005630627274513245, + 0.005638327449560165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.01419447734951973 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011065427213907242 + ], + "opposite_logprob_tokens": [ + -4.511065483093262 + ], + "logprob_reward": -0.011065427213907242, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011065427213907242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1045561283826828, + 0.24829596281051636, + 0.1620228886604309, + 0.12473957240581512, + 0.06668417155742645, + 0.05518815666437149, + 0.06101283058524132, + 0.059408485889434814, + 0.05546967312693596, + 0.06262215226888657 + ], + "credit_map": [ + 0.012968278490006924, + 0.014329342171549797, + 0.012651382945477962, + 0.014301246963441372, + 0.017930787056684494, + 0.009129708632826805, + 0.011793224141001701, + 0.011452163569629192, + 0.06909076124429703, + 0.014426360838115215, + 0.10678452998399734, + 0.014880001544952393, + 0.010359130799770355, + 0.008985684253275394, + 0.007112785708159208, + 0.01665668748319149, + 0.04526541382074356, + 0.01220118161290884, + 0.06373392790555954, + 0.011906716041266918, + 0.008051400072872639, + 0.005153599660843611, + 0.007439164444804192, + 0.008271478116512299, + 0.02494664117693901, + 0.019047681242227554, + 0.021171173080801964, + 0.015772409737110138, + 0.011078590527176857, + 0.010091366246342659, + 0.009675486013293266, + 0.012956220656633377, + 0.010800028219819069, + 0.011509395204484463, + 0.007949872873723507, + 0.009684798307716846, + 0.007517069578170776, + 0.007308463100343943, + 0.005797846708446741, + 0.006116693839430809, + 0.0072806160897016525, + 0.010611174628138542, + 0.006305469200015068, + 0.010698548518121243, + 0.005653777625411749, + 0.0040506646037101746, + 0.004042247775942087, + 0.006545656826347113, + 0.00889499206095934, + 0.008994892239570618, + 0.00784353818744421, + 0.013272315263748169, + 0.007079449947923422, + 0.004162860102951527, + 0.004733419045805931, + 0.006031361408531666, + 0.006836559623479843, + 0.012137206271290779, + 0.006503823678940535, + 0.011174901388585567, + 0.006559917703270912, + 0.004768154118210077, + 0.005625077988952398, + 0.005802844651043415, + 0.006860751658678055, + 0.011236965656280518, + 0.006458510644733906, + 0.00974010955542326, + 0.00491471029818058, + 0.005297584924846888, + 0.005421388428658247, + 0.005539650097489357, + 0.005698055028915405, + 0.006314935628324747, + 0.018357912078499794, + 0.007256010081619024, + 0.004397913347929716, + 0.003968380391597748, + 0.009323530830442905, + 0.007305414881557226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.011065427213907242 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017115719383582473 + ], + "opposite_logprob_tokens": [ + -6.376711368560791 + ], + "logprob_reward": -0.0017115719383582473, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017115719383582473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31651565432548523, + 0.09283093363046646, + 0.13088330626487732, + 0.08932295441627502, + 0.05228321626782417, + 0.057662446051836014, + 0.055911414325237274, + 0.05333804339170456, + 0.059998270124197006, + 0.09125373512506485 + ], + "credit_map": [ + 0.050368912518024445, + 0.031719572842121124, + 0.03664395585656166, + 0.029475100338459015, + 0.043976712971925735, + 0.03258818760514259, + 0.030676312744617462, + 0.061066918075084686, + 0.01761714182794094, + 0.010509542189538479, + 0.021726135164499283, + 0.010408868081867695, + 0.007748892996460199, + 0.008036083541810513, + 0.00590271083638072, + 0.010881554335355759, + 0.036011822521686554, + 0.009599052369594574, + 0.04958711564540863, + 0.009505515918135643, + 0.007583471946418285, + 0.004574616439640522, + 0.0062470450066030025, + 0.007774663157761097, + 0.014118080958724022, + 0.014482788741588593, + 0.011577658355236053, + 0.012046506628394127, + 0.010663987137377262, + 0.008812515996396542, + 0.006791150197386742, + 0.010830259881913662, + 0.005168751813471317, + 0.00894450955092907, + 0.009349552914500237, + 0.006079967133700848, + 0.0065255784429609776, + 0.004632902331650257, + 0.005204760003834963, + 0.006377195008099079, + 0.007498003542423248, + 0.010650226846337318, + 0.0064824544824659824, + 0.009910410270094872, + 0.006033788435161114, + 0.005289777182042599, + 0.0043574729934334755, + 0.0074403127655386925, + 0.0069672707468271255, + 0.011289363726973534, + 0.006496123503893614, + 0.009907132014632225, + 0.0074168844148516655, + 0.003935118205845356, + 0.004300132859498262, + 0.005599386524409056, + 0.0062666698358953, + 0.01052076555788517, + 0.006663201842457056, + 0.006314998958259821, + 0.007775961887091398, + 0.004627144895493984, + 0.005481811705976725, + 0.005687489174306393, + 0.005816393997520208, + 0.012213949114084244, + 0.007884243503212929, + 0.009871783666312695, + 0.0068665919825434685, + 0.004750075750052929, + 0.007098977454006672, + 0.005496257450431585, + 0.009850859642028809, + 0.01184224896132946, + 0.01635156013071537, + 0.012128000147640705, + 0.013253745622932911, + 0.007356816437095404, + 0.009313805028796196, + 0.01115670707076788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0017115719383582473 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "nearly static shot, taken from a medium angle, captures an anime-inspired girl dressed in a classic white sailor suit and a dark blue scarf. With shoulder-length dark brown hair and flowing bangs, she sits sideways at a wooden desk by the window, one hand resting on her chin as she gazes thoughtfully out the window, a hint of yearning in her eyes. The classroom's sliding windows are wide open, and pink and white cherry blossom petals, carried by the afternoon breeze, swirl and drift gently across the window lattice, a few even landing on the corner of her desk and the worn wooden windowsill. Soft sunlight pours in, casting a warm filter over the scene, illuminating the airborne dust and petals, and giving the image a soft, anime-inspired watercolor hue, with desaturated colors and a slightly hazy edge. Behind her, out of focus, other students, gathered in groups of three or four, wearing the same uniforms, their figures and movements, whispering or flipping through books, appear blurred and soft in the frame, forming a vivid background. The entire composition is full of tranquility and poetry, perfectly recreating the warm and slightly sentimental nostalgic atmosphere unique to typical daily life anime.", + "seed": 4879, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_15/index_4833_seed_4879_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.015779258380644024, + "vlm_logprob": -0.015779258380644024, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.015779258380644024, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006113637704402208 + ], + "opposite_logprob_tokens": [ + -5.631113529205322 + ], + "logprob_reward": -0.006113637704402208, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.006113637704402208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3633980453014374, + 0.1617753952741623, + 0.04879859834909439, + 0.04277711361646652, + 0.09620621055364609, + 0.06549487262964249, + 0.034833770245313644, + 0.045555032789707184, + 0.07157409191131592, + 0.069586843252182 + ], + "credit_map": [ + 0.03355633094906807, + 0.05033840984106064, + 0.05556473508477211, + 0.04691186547279358, + 0.06604862213134766, + 0.032024797052145004, + 0.035368215292692184, + 0.043585069477558136, + 0.01409003883600235, + 0.01989845000207424, + 0.031463105231523514, + 0.019147172570228577, + 0.03091341443359852, + 0.016812359914183617, + 0.01434567291289568, + 0.015105199068784714, + 0.004504503682255745, + 0.004591071512550116, + 0.008342137560248375, + 0.007082638796418905, + 0.0070727248676121235, + 0.005092927720397711, + 0.00577220693230629, + 0.006340388674288988, + 0.0034308931790292263, + 0.004665304906666279, + 0.009174786508083344, + 0.009109648875892162, + 0.004477354232221842, + 0.00338340294547379, + 0.004524764139205217, + 0.004010965581983328, + 0.008112194016575813, + 0.012133714742958546, + 0.014321118593215942, + 0.015109126456081867, + 0.017289569601416588, + 0.00821226742118597, + 0.010082198306918144, + 0.010946021415293217, + 0.005161220207810402, + 0.008826016448438168, + 0.00876218918710947, + 0.011544865556061268, + 0.00720486743375659, + 0.007423906587064266, + 0.007166440133005381, + 0.009405376389622688, + 0.003085289616137743, + 0.005689166486263275, + 0.004241612274199724, + 0.005077091045677662, + 0.005491446238011122, + 0.0035590173210948706, + 0.002950764959678054, + 0.004739383235573769, + 0.003921783063560724, + 0.0059951068833470345, + 0.010362209752202034, + 0.005584951490163803, + 0.0061622001230716705, + 0.005009859334677458, + 0.003801401937380433, + 0.004717523697763681, + 0.00624061981216073, + 0.009101060219109058, + 0.006971233058720827, + 0.009980916976928711, + 0.014314987696707249, + 0.0072901491075754166, + 0.009224612265825272, + 0.008450517430901527, + 0.006838283501565456, + 0.006995541509240866, + 0.015032193623483181, + 0.006735191680490971, + 0.009141877293586731, + 0.011296609416604042, + 0.0062465728260576725, + 0.007300577126443386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.006113637704402208 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819620095193386 + ], + "opposite_logprob_tokens": [ + -4.018196105957031 + ], + "logprob_reward": -0.01819620095193386, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819620095193386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30890142917633057, + 0.11905021965503693, + 0.042598918080329895, + 0.04056602716445923, + 0.05907103046774864, + 0.05051252245903015, + 0.07767415046691895, + 0.15340766310691833, + 0.09488148242235184, + 0.05333666130900383 + ], + "credit_map": [ + 0.02989017963409424, + 0.044908203184604645, + 0.041452087461948395, + 0.05588030442595482, + 0.041291408240795135, + 0.03367156535387039, + 0.026940003037452698, + 0.034867629408836365, + 0.012487512081861496, + 0.02122473530471325, + 0.01852496527135372, + 0.012641004286706448, + 0.0203118696808815, + 0.008840280584990978, + 0.012448797933757305, + 0.012571037746965885, + 0.0038166402373462915, + 0.006523130927234888, + 0.006405489519238472, + 0.0056108226999640465, + 0.006461184471845627, + 0.004949834663420916, + 0.004103096667677164, + 0.004728711675852537, + 0.003765638917684555, + 0.005891069769859314, + 0.006424521561712027, + 0.006538194138556719, + 0.004878298379480839, + 0.005104528274387121, + 0.003737071296200156, + 0.00422669667750597, + 0.003730380441993475, + 0.008632293902337551, + 0.006498305592685938, + 0.01040779147297144, + 0.0078056915663182735, + 0.009247922338545322, + 0.005163674242794514, + 0.007584959734231234, + 0.004394815769046545, + 0.00658350670710206, + 0.01060646865516901, + 0.008097315207123756, + 0.005388570949435234, + 0.005203737877309322, + 0.004957049619406462, + 0.005281050223857164, + 0.005576209630817175, + 0.009345117956399918, + 0.012032232247292995, + 0.010889855213463306, + 0.008834379725158215, + 0.01154285203665495, + 0.009545926004648209, + 0.009907571598887444, + 0.014156917110085487, + 0.013789555989205837, + 0.027312617748975754, + 0.025109224021434784, + 0.019475780427455902, + 0.02023441530764103, + 0.01453692652285099, + 0.018792200833559036, + 0.008516709320247173, + 0.013945738784968853, + 0.01216571033000946, + 0.010828121565282345, + 0.013577085919678211, + 0.011006702668964863, + 0.012441892176866531, + 0.012399504892528057, + 0.004620919935405254, + 0.007277062628418207, + 0.005176097620278597, + 0.006003202870488167, + 0.0082919392734766, + 0.008196123875677586, + 0.006263840012252331, + 0.0075074671767652035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.01819620095193386 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023286793380975723 + ], + "opposite_logprob_tokens": [ + -3.773286819458008 + ], + "logprob_reward": -0.023286793380975723, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023286793380975723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3913288712501526, + 0.16555070877075195, + 0.0711456686258316, + 0.05980636924505234, + 0.07376425713300705, + 0.02884921059012413, + 0.07100582867860794, + 0.058578602969646454, + 0.053025342524051666, + 0.02694505639374256 + ], + "credit_map": [ + 0.025996604934334755, + 0.047010309994220734, + 0.04929867014288902, + 0.038925159722566605, + 0.06537260115146637, + 0.0581715852022171, + 0.0469864085316658, + 0.059567585587501526, + 0.012899806722998619, + 0.019356394186615944, + 0.021028107032179832, + 0.019580410793423653, + 0.031503986567258835, + 0.02194644697010517, + 0.017038442194461823, + 0.022197119891643524, + 0.005337855312973261, + 0.008606272749602795, + 0.007385479751974344, + 0.012215016409754753, + 0.010954405181109905, + 0.010609472170472145, + 0.007737244479358196, + 0.00829993188381195, + 0.004349116701632738, + 0.0065908110700547695, + 0.005366750992834568, + 0.005771854892373085, + 0.008213760331273079, + 0.011196969076991081, + 0.010053832083940506, + 0.008263280615210533, + 0.005673062056303024, + 0.008036906830966473, + 0.011332029476761818, + 0.008931083604693413, + 0.011238430626690388, + 0.010718661360442638, + 0.0076239947229623795, + 0.010210095904767513, + 0.0023455736227333546, + 0.0038889646530151367, + 0.0023188195191323757, + 0.00221236445941031, + 0.0035378190223127604, + 0.006630816496908665, + 0.0029882744420319796, + 0.004926580935716629, + 0.005422902759164572, + 0.007764753885567188, + 0.010496090166270733, + 0.012605929747223854, + 0.010491561144590378, + 0.009351569227874279, + 0.006662707310169935, + 0.008210322819650173, + 0.00421662675216794, + 0.00711283553391695, + 0.004938957281410694, + 0.010831252671778202, + 0.009486882016062737, + 0.01008434034883976, + 0.004243952222168446, + 0.007663759868592024, + 0.0042953225784003735, + 0.006261277943849564, + 0.005305349826812744, + 0.0052573648281395435, + 0.009080236777663231, + 0.009683740325272083, + 0.005715638864785433, + 0.0074264174327254295, + 0.00308051030151546, + 0.003072492079809308, + 0.002995297545567155, + 0.0033045595046132803, + 0.0031781119760125875, + 0.005676976405084133, + 0.002205492462962866, + 0.0034316175151616335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.023286793380975723 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029787560924887657 + ], + "opposite_logprob_tokens": [ + -3.529787540435791 + ], + "logprob_reward": -0.029787560924887657, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029787560924887657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38186097145080566, + 0.1568828970193863, + 0.05528348684310913, + 0.05305902659893036, + 0.07019904255867004, + 0.13047011196613312, + 0.030186448246240616, + 0.05446641519665718, + 0.029387200251221657, + 0.03820441663265228 + ], + "credit_map": [ + 0.0382978692650795, + 0.040271081030368805, + 0.03834192827343941, + 0.030381042510271072, + 0.07655973732471466, + 0.08628903329372406, + 0.02852630987763405, + 0.04319397732615471, + 0.012806622311472893, + 0.013617005199193954, + 0.01209118589758873, + 0.014215199276804924, + 0.024290286004543304, + 0.03223942965269089, + 0.023325413465499878, + 0.02429775521159172, + 0.00451863557100296, + 0.004997700918465853, + 0.0060413675382733345, + 0.004842002410441637, + 0.00755729665979743, + 0.012529741041362286, + 0.007268342189490795, + 0.007528404705226421, + 0.004813031293451786, + 0.005088117904961109, + 0.004841990303248167, + 0.0045845601707696915, + 0.007865206338465214, + 0.012113774195313454, + 0.004963156301528215, + 0.008789190091192722, + 0.007027077954262495, + 0.004914581310003996, + 0.006639762315899134, + 0.007638048846274614, + 0.009121659211814404, + 0.01620953343808651, + 0.008828477002680302, + 0.009819897823035717, + 0.011716277338564396, + 0.008451476693153381, + 0.00826316513121128, + 0.006987628061324358, + 0.024800335988402367, + 0.03154724836349487, + 0.017163705080747604, + 0.02154027856886387, + 0.002234522718936205, + 0.003743812208995223, + 0.002333325333893299, + 0.0021912227384746075, + 0.004490062594413757, + 0.007047615479677916, + 0.002903283340856433, + 0.005242602899670601, + 0.004017863888293505, + 0.004559250548481941, + 0.006064745597541332, + 0.006696936674416065, + 0.008548671379685402, + 0.011611316353082657, + 0.006084342021495104, + 0.006883286405354738, + 0.002867260482162237, + 0.002751818159595132, + 0.005204714369028807, + 0.0033683793153613806, + 0.0030318954959511757, + 0.004179021343588829, + 0.003337552770972252, + 0.0046465578489005566, + 0.003484241897240281, + 0.003693999722599983, + 0.006358236540108919, + 0.003840180579572916, + 0.006267941556870937, + 0.004535655491054058, + 0.005072638858109713, + 0.004951521288603544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.029787560924887657 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anime-inspired girl wearing a white sailor suit and a dark blue scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl sitting sideways at a wooden desk by a window with her hand resting on her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are pink and white cherry blossom petals visible swirling and drifting across the window lattice and desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing out-of-focus students in groups wearing similar uniforms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot nearly static taken from a medium angle with a soft, anime-inspired watercolor hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015120989410206676 + ], + "opposite_logprob_tokens": [ + -6.501512050628662 + ], + "logprob_reward": -0.0015120989410206676, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015120989410206676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2944096326828003, + 0.13065439462661743, + 0.08239950239658356, + 0.04200903698801994, + 0.08800481259822845, + 0.0690564513206482, + 0.0653664767742157, + 0.07320016622543335, + 0.08099407702684402, + 0.07390551269054413 + ], + "credit_map": [ + 0.029118366539478302, + 0.039972949773073196, + 0.0363943837583065, + 0.03250107541680336, + 0.05161357671022415, + 0.03632665425539017, + 0.029418006539344788, + 0.03906456008553505, + 0.012373941019177437, + 0.016405917704105377, + 0.016824711114168167, + 0.010838266462087631, + 0.031371500343084335, + 0.01703661121428013, + 0.01279264036566019, + 0.01301079522818327, + 0.007950088940560818, + 0.011884826235473156, + 0.011287018656730652, + 0.015795918181538582, + 0.009711000137031078, + 0.008022662252187729, + 0.008826741017401218, + 0.008921238593757153, + 0.003734748577699065, + 0.005815759301185608, + 0.0044305152259767056, + 0.010189348831772804, + 0.005019144155085087, + 0.004067365545779467, + 0.003830969799309969, + 0.004921182058751583, + 0.0064198290929198265, + 0.012083489447832108, + 0.008028049021959305, + 0.01566159352660179, + 0.01053617987781763, + 0.01441480964422226, + 0.009173198603093624, + 0.011687655933201313, + 0.00505139771848917, + 0.008974197320640087, + 0.007677692919969559, + 0.012044127099215984, + 0.00930310320109129, + 0.009115337394177914, + 0.008088182657957077, + 0.008802407421171665, + 0.003835341427475214, + 0.006994943134486675, + 0.0079857362434268, + 0.011553741991519928, + 0.007823807187378407, + 0.00986410491168499, + 0.006791111547499895, + 0.01051769033074379, + 0.007855899631977081, + 0.010378530248999596, + 0.007743310183286667, + 0.009364820085465908, + 0.00821482390165329, + 0.007961489260196686, + 0.00932331196963787, + 0.012357975356280804, + 0.006847688928246498, + 0.007302542217075825, + 0.01650375686585903, + 0.008626595139503479, + 0.010213824920356274, + 0.011675028130412102, + 0.00968856643885374, + 0.010136065073311329, + 0.006389382295310497, + 0.008000452071428299, + 0.00954483449459076, + 0.014294863678514957, + 0.0091526685282588, + 0.008075236342847347, + 0.008639947511255741, + 0.009808121249079704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0015120989410206676 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_2/index_tensor([8905, 8905])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_2/index_tensor([8905, 8905])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3fa1aa13c4c26600b5e2d762181e2cfd8ac3b513 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_2/index_tensor([8905, 8905])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme close-up, rendered in crisp 4K, captures a delicate wolf spider, its intricate seed head trembling gracefully against the overwhelming force of nature. Each feathery bristle of the plume is rendered in surreal detail, appearing almost silvery and translucent as it holds its shape. The camera utilizes a very shallow depth of field to achieve a profound sense of isolation, allowing this solitary seed head to remain in sharp, unwavering focus while a swirling blizzard of fine sand grains engulfs the surrounding space. The background dissolves into a chaotic, blurred vortex of motion, a turbulent swirl of golden-brown sand that whips and swirls violently. Intense sunlight, heavily scattered by the dense storm, bathes the scene in a bright, luminous mist, causing individual grains of sand flying past the lens to shimmer like tiny, fleeting sparks. Dominated by a warm, monochromatic palette of ochre and sepia, the entire image creates a powerful and atmospheric contrast between the tranquil, resilient beauty of the plant and the chaotic, all-consuming power of the sandstorm.", + "seed": 19803, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_2/index_8905_seed_19803_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.25112796584144237, + "vlm_logprob": -0.25112796584144237, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.25112796584144237, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1399768590927124 + ], + "opposite_logprob_tokens": [ + -0.3899768590927124 + ], + "logprob_reward": -1.1399768590927124, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1399768590927124, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13431936502456665, + 0.14222174882888794, + 0.19812259078025818, + 0.09809375554323196, + 0.07429704815149307, + 0.10102102160453796, + 0.08834703266620636, + 0.02376243658363819, + 0.07901240140199661, + 0.060802675783634186 + ], + "credit_map": [ + 0.012848290614783764, + 0.019884083420038223, + 0.018373027443885803, + 0.01525501161813736, + 0.01536866556853056, + 0.02085474506020546, + 0.0181839931756258, + 0.013551536947488785, + 0.007452907506376505, + 0.011170740239322186, + 0.014243382960557938, + 0.014435273595154285, + 0.02723521552979946, + 0.018279746174812317, + 0.03285444527864456, + 0.016550011932849884, + 0.01072017289698124, + 0.016633788123726845, + 0.01712731271982193, + 0.01978747919201851, + 0.044913675636053085, + 0.02433638833463192, + 0.04635264351963997, + 0.01825110986828804, + 0.007900550961494446, + 0.010812864638864994, + 0.011579509824514389, + 0.01186407171189785, + 0.018708184361457825, + 0.015046502463519573, + 0.013988519087433815, + 0.008193545043468475, + 0.004379924386739731, + 0.005725523456931114, + 0.008111902512609959, + 0.008870579302310944, + 0.014109998010098934, + 0.009915697388350964, + 0.015772288665175438, + 0.007411124650388956, + 0.007071786094456911, + 0.009494041092693806, + 0.010456669144332409, + 0.011201511137187481, + 0.018763499334454536, + 0.012911837548017502, + 0.01696222834289074, + 0.01415944006294012, + 0.0057342094369232655, + 0.006834224797785282, + 0.008732249960303307, + 0.007096674758940935, + 0.01759806089103222, + 0.010431499220430851, + 0.019019708037376404, + 0.0129003981128335, + 0.0018885316094383597, + 0.0020322466734796762, + 0.0026448206044733524, + 0.003797722514718771, + 0.0041535282507538795, + 0.0031144581735134125, + 0.003639922244474292, + 0.002491203136742115, + 0.006534097250550985, + 0.010639483109116554, + 0.007446977775543928, + 0.007982396520674229, + 0.013238828629255295, + 0.009260674007236958, + 0.014544486068189144, + 0.00936545617878437, + 0.005398174282163382, + 0.009667029604315758, + 0.005396525841206312, + 0.004263424780219793, + 0.011296102777123451, + 0.010197144001722336, + 0.007162172347307205, + 0.007422095164656639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -1.1399768590927124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055004168301820755 + ], + "opposite_logprob_tokens": [ + -2.930004119873047 + ], + "logprob_reward": -0.055004168301820755, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055004168301820755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.210938960313797, + 0.11028766632080078, + 0.17012962698936462, + 0.09078193455934525, + 0.07513117045164108, + 0.10523447394371033, + 0.12044602632522583, + 0.040600039064884186, + 0.03275446593761444, + 0.04369562491774559 + ], + "credit_map": [ + 0.01675071194767952, + 0.0230934526771307, + 0.04077870026230812, + 0.02154134213924408, + 0.021688196808099747, + 0.02908504381775856, + 0.024953803047537804, + 0.033047717064619064, + 0.005524825304746628, + 0.010337437503039837, + 0.014691177755594254, + 0.01100748311728239, + 0.018075358122587204, + 0.017774317413568497, + 0.0204177089035511, + 0.012459365651011467, + 0.009902888908982277, + 0.01671573519706726, + 0.019629519432783127, + 0.016198299825191498, + 0.03966709226369858, + 0.02195833995938301, + 0.030937187373638153, + 0.015120559372007847, + 0.00663478160277009, + 0.010810691863298416, + 0.01387012004852295, + 0.014590045437216759, + 0.00829420704394579, + 0.014911139383912086, + 0.011572393588721752, + 0.010098558850586414, + 0.0065184347331523895, + 0.008492552675306797, + 0.009782104752957821, + 0.010833090171217918, + 0.010220468044281006, + 0.009477616287767887, + 0.010840609669685364, + 0.008966297842562199, + 0.008395830169320107, + 0.008132812567055225, + 0.0100364089012146, + 0.00953250378370285, + 0.02265337109565735, + 0.0156866442412138, + 0.017423735931515694, + 0.013373160734772682, + 0.008868111297488213, + 0.007489745505154133, + 0.010207215324044228, + 0.013263329863548279, + 0.025978919118642807, + 0.014424900524318218, + 0.025907276198267937, + 0.014306534081697464, + 0.003976133186370134, + 0.005408961325883865, + 0.0054208068177104, + 0.006702757440507412, + 0.005160569213330746, + 0.00528591126203537, + 0.00508631719276309, + 0.0035585840232670307, + 0.0024181560147553682, + 0.003670778824016452, + 0.0038819529581815004, + 0.005533954594284296, + 0.0038639858830720186, + 0.004963073413819075, + 0.005170941818505526, + 0.003251624060794711, + 0.0037057765293866396, + 0.005439204629510641, + 0.003570681670680642, + 0.0038355581928044558, + 0.008905069902539253, + 0.005487464834004641, + 0.007489660754799843, + 0.005262208636850119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.055004168301820755 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645863547921181 + ], + "opposite_logprob_tokens": [ + -5.379645824432373 + ], + "logprob_reward": -0.004645863547921181, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645863547921181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14433152973651886, + 0.14628222584724426, + 0.15671300888061523, + 0.08497003465890884, + 0.06177780404686928, + 0.11017218977212906, + 0.13966397941112518, + 0.052843689918518066, + 0.05368662625551224, + 0.04955892264842987 + ], + "credit_map": [ + 0.011991573497653008, + 0.01752212643623352, + 0.01995513029396534, + 0.015982072800397873, + 0.0199155043810606, + 0.01765759103000164, + 0.02465726062655449, + 0.016650278121232986, + 0.007166772615164518, + 0.01155227329581976, + 0.014061211608350277, + 0.015376362949609756, + 0.028280429542064667, + 0.017998870462179184, + 0.033699553459882736, + 0.01814674399793148, + 0.008594204671680927, + 0.012173664756119251, + 0.01673710346221924, + 0.01626737043261528, + 0.0332786999642849, + 0.018431155011057854, + 0.0366855189204216, + 0.014545287005603313, + 0.007459364831447601, + 0.010650396347045898, + 0.011767097748816013, + 0.009921557269990444, + 0.013705127872526646, + 0.012514263391494751, + 0.010032158344984055, + 0.008920065127313137, + 0.005441916640847921, + 0.006516473833471537, + 0.005724072922021151, + 0.006638790480792522, + 0.009508239105343819, + 0.007806235924363136, + 0.012884315103292465, + 0.0072577595710754395, + 0.008515122346580029, + 0.011119352653622627, + 0.011109301820397377, + 0.010715526528656483, + 0.020737333223223686, + 0.01239105500280857, + 0.020146647468209267, + 0.015437853522598743, + 0.009991166181862354, + 0.011352145113050938, + 0.016227250918745995, + 0.013984071090817451, + 0.02631986141204834, + 0.016184130683541298, + 0.026961850002408028, + 0.01864350587129593, + 0.005048224702477455, + 0.006160476710647345, + 0.007524503394961357, + 0.008816888555884361, + 0.006696689873933792, + 0.006870201788842678, + 0.004850935190916061, + 0.006875769700855017, + 0.0048268330283463, + 0.008218524977564812, + 0.00487991189584136, + 0.005313714034855366, + 0.009157955646514893, + 0.005965018644928932, + 0.009065969847142696, + 0.006258695852011442, + 0.0044136433862149715, + 0.007489787880331278, + 0.0044427006505429745, + 0.0038023758679628372, + 0.008918722160160542, + 0.008388521149754524, + 0.006060807034373283, + 0.006042368244379759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.004645863547921181 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054968852549791336 + ], + "opposite_logprob_tokens": [ + -2.92996883392334 + ], + "logprob_reward": -0.054968852549791336, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054968852549791336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2035885453224182, + 0.1558489054441452, + 0.039953961968421936, + 0.140479177236557, + 0.1750326007604599, + 0.05536499246954918, + 0.025081081315875053, + 0.13040736317634583, + 0.06058308109641075, + 0.013660304248332977 + ], + "credit_map": [ + 0.01674102619290352, + 0.024747688323259354, + 0.039640139788389206, + 0.021768951788544655, + 0.026755237951874733, + 0.026060238480567932, + 0.028286470100283623, + 0.019588809460401535, + 0.013099721632897854, + 0.0202107522636652, + 0.028560146689414978, + 0.01889573037624359, + 0.01764685846865177, + 0.01980491355061531, + 0.02017459273338318, + 0.017456194385886192, + 0.002641872502863407, + 0.003811909118667245, + 0.005889373831450939, + 0.004268771503120661, + 0.008924596942961216, + 0.004856678657233715, + 0.00477494765073061, + 0.004785808268934488, + 0.009893682785332203, + 0.01787174679338932, + 0.02177937887609005, + 0.022707559168338776, + 0.012114860117435455, + 0.022807054221630096, + 0.015731511637568474, + 0.01757337711751461, + 0.018088432028889656, + 0.018982499837875366, + 0.02435394749045372, + 0.032502297312021255, + 0.016616353765130043, + 0.022617092356085777, + 0.017115125432610512, + 0.02475685067474842, + 0.0039073689840734005, + 0.0047515262849628925, + 0.005745887756347656, + 0.005062609910964966, + 0.011924173682928085, + 0.006669589318335056, + 0.010935601778328419, + 0.006368238478899002, + 0.002354754600673914, + 0.001961366506293416, + 0.0028110353741794825, + 0.0023484528064727783, + 0.005137303844094276, + 0.0034179585054516792, + 0.004171740263700485, + 0.0028784689493477345, + 0.01515487302094698, + 0.017213713377714157, + 0.021148553118109703, + 0.023508841171860695, + 0.013071086257696152, + 0.017035596072673798, + 0.009816111996769905, + 0.013458577916026115, + 0.00599188357591629, + 0.008886618539690971, + 0.007830878719687462, + 0.011376245878636837, + 0.00694133760407567, + 0.008039436303079128, + 0.0054959943518042564, + 0.006020686123520136, + 0.0013547987909987569, + 0.0019277932588011026, + 0.0021814280189573765, + 0.0017632051603868604, + 0.0016747131012380123, + 0.0018656929023563862, + 0.0017968000611290336, + 0.0010958720231428742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.054968852549791336 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010440857149660587 + ], + "opposite_logprob_tokens": [ + -6.876044273376465 + ], + "logprob_reward": -0.0010440857149660587, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010440857149660587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23974186182022095, + 0.19407398998737335, + 0.15314313769340515, + 0.05062638595700264, + 0.03425344452261925, + 0.0746111124753952, + 0.09326481819152832, + 0.04709898680448532, + 0.05139169469475746, + 0.06179454177618027 + ], + "credit_map": [ + 0.01774498261511326, + 0.02506750449538231, + 0.028956890106201172, + 0.023288052529096603, + 0.0361802764236927, + 0.027923868969082832, + 0.04935846105217934, + 0.031221823766827583, + 0.012829186394810677, + 0.02247321791946888, + 0.023176347836852074, + 0.020758487284183502, + 0.029843347147107124, + 0.028013037517666817, + 0.03504703938961029, + 0.021933333948254585, + 0.00816203560680151, + 0.01292894035577774, + 0.013629907742142677, + 0.014347507618367672, + 0.036292023956775665, + 0.01919739507138729, + 0.034444186836481094, + 0.014141139574348927, + 0.004503556527197361, + 0.0041536930948495865, + 0.005871836561709642, + 0.006031087599694729, + 0.010195709764957428, + 0.006401101127266884, + 0.007736689411103725, + 0.005732713732868433, + 0.0029178443364799023, + 0.0045659709721803665, + 0.004237599205225706, + 0.004268900491297245, + 0.004613707773387432, + 0.004490539897233248, + 0.005449514836072922, + 0.0037093686405569315, + 0.005440699402242899, + 0.006825252901762724, + 0.006541178096085787, + 0.0071752541698515415, + 0.014797255396842957, + 0.00845353677868843, + 0.0142544349655509, + 0.011123496107757092, + 0.006327065639197826, + 0.006977601442486048, + 0.008885255083441734, + 0.008419998921453953, + 0.019573818892240524, + 0.010629609227180481, + 0.019518369808793068, + 0.012933095917105675, + 0.0034558705519884825, + 0.005407167132943869, + 0.004070474300533533, + 0.004292277153581381, + 0.008747835643589497, + 0.0058246031403541565, + 0.008382264524698257, + 0.0069184922613203526, + 0.004096215590834618, + 0.006831769365817308, + 0.004136855248361826, + 0.0045633092522621155, + 0.010049843229353428, + 0.005990040488541126, + 0.009482250548899174, + 0.006241410970687866, + 0.004996764939278364, + 0.008214999921619892, + 0.005471440497785807, + 0.00510760024189949, + 0.010897548869252205, + 0.01051783375442028, + 0.008653121069073677, + 0.007935235276818275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0010440857149660587 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up, rendered in crisp 4K, captures a delicate wolf spider, its intricate seed head trembling gracefully against the overwhelming force of nature. Each feathery bristle of the plume is rendered in surreal detail, appearing almost silvery and translucent as it holds its shape. The camera utilizes a very shallow depth of field to achieve a profound sense of isolation, allowing this solitary seed head to remain in sharp, unwavering focus while a swirling blizzard of fine sand grains engulfs the surrounding space. The background dissolves into a chaotic, blurred vortex of motion, a turbulent swirl of golden-brown sand that whips and swirls violently. Intense sunlight, heavily scattered by the dense storm, bathes the scene in a bright, luminous mist, causing individual grains of sand flying past the lens to shimmer like tiny, fleeting sparks. Dominated by a warm, monochromatic palette of ochre and sepia, the entire image creates a powerful and atmospheric contrast between the tranquil, resilient beauty of the plant and the chaotic, all-consuming power of the sandstorm.", + "seed": 19804, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_2/index_8905_seed_19804_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16337688753847032, + "vlm_logprob": -0.16337688753847032, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.16337688753847032, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6987184882164001 + ], + "opposite_logprob_tokens": [ + -0.6987184882164001 + ], + "logprob_reward": -0.6987184882164001, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6987184882164001, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12381277978420258, + 0.04495871439576149, + 0.04500099644064903, + 0.022725725546479225, + 0.03674684464931488, + 0.022181538864970207, + 0.0372643880546093, + 0.6428661346435547, + 0.011314005590975285, + 0.013128875754773617 + ], + "credit_map": [ + 0.007799872662872076, + 0.011399742215871811, + 0.032959625124931335, + 0.006941128987818956, + 0.00936055276542902, + 0.03932429105043411, + 0.007548449095338583, + 0.008479117415845394, + 0.004136231727898121, + 0.007105412892997265, + 0.005961062386631966, + 0.0050054206512868404, + 0.007157094776630402, + 0.007809897419065237, + 0.003926820587366819, + 0.003856774652376771, + 0.0036171565297991037, + 0.00412002531811595, + 0.00555289164185524, + 0.0057664490304887295, + 0.007397123612463474, + 0.006722406018525362, + 0.005420423578470945, + 0.0064045218750834465, + 0.002850731136277318, + 0.0028124861419200897, + 0.002167539205402136, + 0.0027667845133692026, + 0.003489599796012044, + 0.0032096572685986757, + 0.0032305875793099403, + 0.0021983403712511063, + 0.0031798239797353745, + 0.0063459486700594425, + 0.00476697226986289, + 0.0028707673773169518, + 0.004939205013215542, + 0.005358751863241196, + 0.004054284188896418, + 0.005231090821325779, + 0.0016360993031412363, + 0.0024654336739331484, + 0.0038100378587841988, + 0.0021496033295989037, + 0.003966452553868294, + 0.0035981256514787674, + 0.0026312675327062607, + 0.0019245186122134328, + 0.0033085590694099665, + 0.0045399111695587635, + 0.004492474254220724, + 0.004129525274038315, + 0.004504370037466288, + 0.004182431381195784, + 0.00579615356400609, + 0.006310964468866587, + 0.027136433869600296, + 0.042702868580818176, + 0.2339334636926651, + 0.021980056539177895, + 0.013990651816129684, + 0.2526485025882721, + 0.020592259243130684, + 0.029881887137889862, + 0.00072330457624048, + 0.0013175788335502148, + 0.001980689587071538, + 0.0011829473078250885, + 0.0013913706643506885, + 0.0025992649607360363, + 0.0012533909175544977, + 0.0008654583361931145, + 0.0013800846645608544, + 0.0023018207866698503, + 0.0016536100301891565, + 0.0010497308103367686, + 0.001305025303736329, + 0.0024982322938740253, + 0.0014955885708332062, + 0.0014447839930653572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.6987184882164001 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902579009532928 + ], + "opposite_logprob_tokens": [ + -2.5790257453918457 + ], + "logprob_reward": -0.07902579009532928, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902579009532928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23092137277126312, + 0.06900005042552948, + 0.07644883543252945, + 0.03815251588821411, + 0.054736390709877014, + 0.043052416294813156, + 0.156119242310524, + 0.26887059211730957, + 0.04169929772615433, + 0.020999357104301453 + ], + "credit_map": [ + 0.0140755670145154, + 0.032582562416791916, + 0.030088301748037338, + 0.035897452384233475, + 0.02618546225130558, + 0.04217880964279175, + 0.02989935874938965, + 0.02001383900642395, + 0.010588507167994976, + 0.008683476597070694, + 0.00856953114271164, + 0.008488650433719158, + 0.01047684345394373, + 0.007614892907440662, + 0.006663367617875338, + 0.007914778776466846, + 0.007222841493785381, + 0.008417641744017601, + 0.008606216870248318, + 0.009923353791236877, + 0.01202840730547905, + 0.010289415717124939, + 0.008874021470546722, + 0.011086925864219666, + 0.0041299788281321526, + 0.003943734336644411, + 0.004266402218490839, + 0.0039647421799600124, + 0.005672695115208626, + 0.005457038525491953, + 0.005174518097192049, + 0.005543404258787632, + 0.0047383178025484085, + 0.008711464703083038, + 0.005445897579193115, + 0.006307412404567003, + 0.007077161222696304, + 0.00680993078276515, + 0.008291374891996384, + 0.00735482107847929, + 0.003030811669304967, + 0.004734099842607975, + 0.011000488884747028, + 0.0029473614413291216, + 0.0028782386798411608, + 0.01396105159074068, + 0.0021753942128270864, + 0.0023249653168022633, + 0.011888420209288597, + 0.020695600658655167, + 0.020820792764425278, + 0.017544306814670563, + 0.016660457476973534, + 0.02002868428826332, + 0.023380795493721962, + 0.025100165978074074, + 0.011482110247015953, + 0.01756363734602928, + 0.09620752930641174, + 0.010199075564742088, + 0.010038073174655437, + 0.10107937455177307, + 0.009827520698308945, + 0.01247325073927641, + 0.0044980039820075035, + 0.005198297090828419, + 0.004855414852499962, + 0.0041768974624574184, + 0.007139907218515873, + 0.006486653350293636, + 0.005610791500657797, + 0.0037333283107727766, + 0.002479183254763484, + 0.00277019664645195, + 0.002462562406435609, + 0.002071199705824256, + 0.0024977002758532763, + 0.0021626558154821396, + 0.003879550378769636, + 0.002676305826753378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.07902579009532928 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100368358194828 + ], + "opposite_logprob_tokens": [ + -5.504100322723389 + ], + "logprob_reward": -0.004100368358194828, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100368358194828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13906735181808472, + 0.13619299232959747, + 0.06009955331683159, + 0.05318518728017807, + 0.033694036304950714, + 0.01712171547114849, + 0.03814011812210083, + 0.42611590027809143, + 0.06246111914515495, + 0.033922042697668076 + ], + "credit_map": [ + 0.009589741006493568, + 0.011776816099882126, + 0.0352465882897377, + 0.013394869863986969, + 0.009633902460336685, + 0.03996214270591736, + 0.008828110061585903, + 0.010635179467499256, + 0.016819093376398087, + 0.020053034648299217, + 0.0213960912078619, + 0.02030273899435997, + 0.014525388367474079, + 0.014990527182817459, + 0.012682945467531681, + 0.01542317122220993, + 0.006183868274092674, + 0.009134987369179726, + 0.006458986550569534, + 0.007262036669999361, + 0.006983337923884392, + 0.009488637559115887, + 0.005122094880789518, + 0.009465602226555347, + 0.007420314010232687, + 0.00802709348499775, + 0.005418012849986553, + 0.0071388473734259605, + 0.005949502345174551, + 0.008916501887142658, + 0.005224944557994604, + 0.005089974496513605, + 0.0030370699241757393, + 0.006506205536425114, + 0.004235273692756891, + 0.003504888853058219, + 0.0035980341490358114, + 0.005445515271276236, + 0.003439085092395544, + 0.003927964251488447, + 0.0013241684064269066, + 0.002285844413563609, + 0.003155976301059127, + 0.0014278393937274814, + 0.0017853927565738559, + 0.004132994916290045, + 0.0014668244402855635, + 0.0015426750760525465, + 0.003946366254240274, + 0.006852227728813887, + 0.006357571110129356, + 0.005020909011363983, + 0.002867500064894557, + 0.0047189160250127316, + 0.0033585343044251204, + 0.005018094088882208, + 0.01887829229235649, + 0.028493380174040794, + 0.15364724397659302, + 0.014535625465214252, + 0.01021753903478384, + 0.16611559689044952, + 0.013764300383627415, + 0.020463960245251656, + 0.008421437814831734, + 0.009131613187491894, + 0.008967136032879353, + 0.00856395997107029, + 0.005072685889899731, + 0.009736377745866776, + 0.006236025132238865, + 0.0063318838365375996, + 0.004267597571015358, + 0.005257860291749239, + 0.005542682483792305, + 0.0038295681588351727, + 0.002953779650852084, + 0.004721225239336491, + 0.003352367551997304, + 0.003996961284428835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.004100368358194828 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370363265275955 + ], + "opposite_logprob_tokens": [ + -3.408703565597534 + ], + "logprob_reward": -0.03370363265275955, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370363265275955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24472928047180176, + 0.07573787868022919, + 0.03054436482489109, + 0.01808631792664528, + 0.04090162366628647, + 0.03738662227988243, + 0.019256098195910454, + 0.48797476291656494, + 0.022751569747924805, + 0.02263142541050911 + ], + "credit_map": [ + 0.015482764691114426, + 0.03378652036190033, + 0.04803666099905968, + 0.024065356701612473, + 0.021846843883395195, + 0.06707165390253067, + 0.01647520810365677, + 0.017964288592338562, + 0.009440375491976738, + 0.01150848064571619, + 0.012217868119478226, + 0.008560745045542717, + 0.008104074746370316, + 0.011297826655209064, + 0.0060876780189573765, + 0.008520832285284996, + 0.0032329412642866373, + 0.002733909524977207, + 0.003537630196660757, + 0.0043909261003136635, + 0.003435028251260519, + 0.003922221250832081, + 0.0033889301121234894, + 0.005902778822928667, + 0.002028025221079588, + 0.002448305021971464, + 0.002061058534309268, + 0.002477889182046056, + 0.002128790132701397, + 0.002419987926259637, + 0.002227736171334982, + 0.0022945271339267492, + 0.003826084779575467, + 0.007971346378326416, + 0.005499490536749363, + 0.0038521774113178253, + 0.0046473280526697636, + 0.006772638764232397, + 0.003950899001210928, + 0.004381661769002676, + 0.002250397577881813, + 0.004231513477861881, + 0.010113449767231941, + 0.0019355230033397675, + 0.0021470191422849894, + 0.012722530402243137, + 0.0020023207180202007, + 0.0019838723819702864, + 0.0010831764666363597, + 0.0026488269213587046, + 0.002741473726928234, + 0.002297321567311883, + 0.002771691419184208, + 0.0023824202362447977, + 0.002413205336779356, + 0.0029179854318499565, + 0.02074783854186535, + 0.03270220011472702, + 0.17656998336315155, + 0.017360694706439972, + 0.011116590350866318, + 0.19090692698955536, + 0.015505392104387283, + 0.02306518331170082, + 0.001987227238714695, + 0.0025152629241347313, + 0.0036551060620695353, + 0.0037210050504654646, + 0.0018819436663761735, + 0.005235359538346529, + 0.0016629834426566958, + 0.002092683454975486, + 0.002172154141589999, + 0.004079942125827074, + 0.002801532158628106, + 0.0027084392495453358, + 0.0023541448172181845, + 0.003088897094130516, + 0.002914078999310732, + 0.0025122379884123802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.03370363265275955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001336158369667828 + ], + "opposite_logprob_tokens": [ + -6.626336097717285 + ], + "logprob_reward": -0.001336158369667828, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001336158369667828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4003511071205139, + 0.07191745191812515, + 0.04177568480372429, + 0.014572340995073318, + 0.05337027832865715, + 0.019240476191043854, + 0.03926170617341995, + 0.2968428134918213, + 0.026204723864793777, + 0.036463506519794464 + ], + "credit_map": [ + 0.021643755957484245, + 0.029624463990330696, + 0.12307394295930862, + 0.02303347922861576, + 0.025123072788119316, + 0.14006131887435913, + 0.01680060289800167, + 0.02099042758345604, + 0.008938481099903584, + 0.010899151675403118, + 0.01077438797801733, + 0.01054546982049942, + 0.007882988080382347, + 0.009165225550532341, + 0.006394864991307259, + 0.007316883187741041, + 0.004142598249018192, + 0.004641283303499222, + 0.005055873654782772, + 0.00561667000874877, + 0.004236419219523668, + 0.006051532458513975, + 0.003600365249440074, + 0.008430936373770237, + 0.0018075002590194345, + 0.0025257100351154804, + 0.0016507318941876292, + 0.0016061443602666259, + 0.0018081485759466887, + 0.0016935840249061584, + 0.002097786171361804, + 0.0013827341608703136, + 0.005337809212505817, + 0.006416411139070988, + 0.007209091912955046, + 0.007555182557553053, + 0.006336632184684277, + 0.006454376503825188, + 0.00863135326653719, + 0.005429413169622421, + 0.0018013459630310535, + 0.00331460521556437, + 0.003013836918398738, + 0.001973336562514305, + 0.0030480229761451483, + 0.0024186819791793823, + 0.0018369078170508146, + 0.001833737245760858, + 0.004202859476208687, + 0.006977982819080353, + 0.00600942550227046, + 0.004608238115906715, + 0.0037239508237689734, + 0.004168961197137833, + 0.004932663403451443, + 0.004637622740119696, + 0.012266116216778755, + 0.020291363820433617, + 0.10690280795097351, + 0.010769544169306755, + 0.006635932717472315, + 0.11701326072216034, + 0.009296146221458912, + 0.013667595572769642, + 0.003446845104917884, + 0.0035408763214945793, + 0.003166117938235402, + 0.003607897786423564, + 0.0029147809837013483, + 0.003856587689369917, + 0.002469654893502593, + 0.0032019612845033407, + 0.003245196770876646, + 0.004987793043255806, + 0.00546301482245326, + 0.0040545291267335415, + 0.004997470416128635, + 0.004041529260575771, + 0.005341217387467623, + 0.004332751501351595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.001336158369667828 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up, rendered in crisp 4K, captures a delicate wolf spider, its intricate seed head trembling gracefully against the overwhelming force of nature. Each feathery bristle of the plume is rendered in surreal detail, appearing almost silvery and translucent as it holds its shape. The camera utilizes a very shallow depth of field to achieve a profound sense of isolation, allowing this solitary seed head to remain in sharp, unwavering focus while a swirling blizzard of fine sand grains engulfs the surrounding space. The background dissolves into a chaotic, blurred vortex of motion, a turbulent swirl of golden-brown sand that whips and swirls violently. Intense sunlight, heavily scattered by the dense storm, bathes the scene in a bright, luminous mist, causing individual grains of sand flying past the lens to shimmer like tiny, fleeting sparks. Dominated by a warm, monochromatic palette of ochre and sepia, the entire image creates a powerful and atmospheric contrast between the tranquil, resilient beauty of the plant and the chaotic, all-consuming power of the sandstorm.", + "seed": 19805, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_2/index_8905_seed_19805_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.20170987580204383, + "vlm_logprob": -0.20170987580204383, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.20170987580204383, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9755756855010986 + ], + "opposite_logprob_tokens": [ + -0.47557565569877625 + ], + "logprob_reward": -0.9755756855010986, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9755756855010986, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21273308992385864, + 0.07110591232776642, + 0.033270202577114105, + 0.06179451197385788, + 0.04783123731613159, + 0.14215980470180511, + 0.0982067734003067, + 0.04179548844695091, + 0.13689908385276794, + 0.15420396625995636 + ], + "credit_map": [ + 0.015824418514966965, + 0.02697944827377796, + 0.026681678369641304, + 0.023952409625053406, + 0.026814155280590057, + 0.027752451598644257, + 0.03333989903330803, + 0.03138861060142517, + 0.005707526113837957, + 0.011443059891462326, + 0.006742453668266535, + 0.009854506701231003, + 0.007384598255157471, + 0.011354681104421616, + 0.010952017270028591, + 0.007667062804102898, + 0.0033225123770534992, + 0.005191311705857515, + 0.003172320080921054, + 0.0049074008129537106, + 0.00526293134316802, + 0.0035967484582215548, + 0.004374832846224308, + 0.0034421428572386503, + 0.005784629378467798, + 0.007575202267616987, + 0.007824222557246685, + 0.006835608743131161, + 0.008492640219628811, + 0.009260812774300575, + 0.009567014873027802, + 0.0064543760381639, + 0.0030499575659632683, + 0.0050532822497189045, + 0.006643064320087433, + 0.0056013730354607105, + 0.008676194585859776, + 0.006647622678428888, + 0.006197433918714523, + 0.005962306167930365, + 0.012307895347476006, + 0.017726754769682884, + 0.01663857325911522, + 0.021293258294463158, + 0.022229010239243507, + 0.01892317645251751, + 0.015142270363867283, + 0.017898861318826675, + 0.008840653114020824, + 0.013776290230453014, + 0.013388403691351414, + 0.01795794628560543, + 0.009778439067304134, + 0.014622790738940239, + 0.009137267246842384, + 0.010704980231821537, + 0.002776055596768856, + 0.006984666455537081, + 0.004635636229068041, + 0.00527031859382987, + 0.0039876364171504974, + 0.009352126158773899, + 0.0055480413138866425, + 0.0032410079147666693, + 0.010989981703460217, + 0.016955416649580002, + 0.015017461962997913, + 0.017294928431510925, + 0.022368745878338814, + 0.016194339841604233, + 0.01641690917313099, + 0.021661294624209404, + 0.014676939696073532, + 0.024757862091064453, + 0.01931271329522133, + 0.023107687011361122, + 0.01646999455988407, + 0.020904147997498512, + 0.016775839030742645, + 0.01819877326488495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.9755756855010986 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059924043715 + ], + "opposite_logprob_tokens": [ + -3.895599126815796 + ], + "logprob_reward": -0.02059924043715, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059924043715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12101411074399948, + 0.06495044380426407, + 0.059459660202264786, + 0.14207912981510162, + 0.028034690767526627, + 0.13114020228385925, + 0.13279743492603302, + 0.058508433401584625, + 0.15140023827552795, + 0.11061562597751617 + ], + "credit_map": [ + 0.01236379612237215, + 0.014994500204920769, + 0.014355786144733429, + 0.01202499307692051, + 0.014752369374036789, + 0.016936758533120155, + 0.012223878875374794, + 0.02336202748119831, + 0.0061210645362734795, + 0.010930364020168781, + 0.006429361645132303, + 0.009616349823772907, + 0.007558957673609257, + 0.007705971598625183, + 0.00781335961073637, + 0.00877501629292965, + 0.006570895202457905, + 0.007951109670102596, + 0.0065139904618263245, + 0.0094559071585536, + 0.006889197509735823, + 0.00659398827701807, + 0.006552549544721842, + 0.008932027965784073, + 0.014293824322521687, + 0.018175335600972176, + 0.01756145805120468, + 0.014879596419632435, + 0.023453783243894577, + 0.019353313371539116, + 0.02010287344455719, + 0.014258950017392635, + 0.0019792099483311176, + 0.0024658418260514736, + 0.003529301844537258, + 0.004257970489561558, + 0.0039953128434717655, + 0.003396181622520089, + 0.005059766583144665, + 0.003351107006892562, + 0.0088758310303092, + 0.012834630906581879, + 0.011232920922338963, + 0.014389069750905037, + 0.02675343118607998, + 0.024485332891345024, + 0.020248789340257645, + 0.012320206500589848, + 0.00953095592558384, + 0.01491379551589489, + 0.01050588395446539, + 0.013370399363338947, + 0.030116552487015724, + 0.02354339323937893, + 0.017949754372239113, + 0.012866704724729061, + 0.0044099921360611916, + 0.007126736920326948, + 0.0078017194755375385, + 0.009724644012749195, + 0.007481066044420004, + 0.00983140803873539, + 0.006991366855800152, + 0.005141506902873516, + 0.012434590607881546, + 0.020481102168560028, + 0.01807466149330139, + 0.02441166527569294, + 0.018972741439938545, + 0.01957317627966404, + 0.017296792939305305, + 0.020155521109700203, + 0.010331332683563232, + 0.017986899241805077, + 0.015255773440003395, + 0.017851321026682854, + 0.009668868966400623, + 0.015639502555131912, + 0.012003377079963684, + 0.011878562159836292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 6 + ], + "selected_reward": -0.02059924043715 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002819374669343233 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.002819374669343233, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002819374669343233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1673072874546051, + 0.037896499037742615, + 0.043229781091213226, + 0.10987256467342377, + 0.027917152270674706, + 0.0997459813952446, + 0.022730115801095963, + 0.08296625316143036, + 0.22122468054294586, + 0.18710975348949432 + ], + "credit_map": [ + 0.012752048671245575, + 0.023584570735692978, + 0.020215310156345367, + 0.021314656361937523, + 0.025999754667282104, + 0.021711133420467377, + 0.01783834584057331, + 0.023891454562544823, + 0.0036395264323800802, + 0.004457010887563229, + 0.004866147879511118, + 0.005738399922847748, + 0.004764250013977289, + 0.004098678473383188, + 0.005799259524792433, + 0.0045332228764891624, + 0.005168883595615625, + 0.005912343971431255, + 0.0047511509619653225, + 0.0057123382575809956, + 0.006015544757246971, + 0.005354409106075764, + 0.004202314652502537, + 0.006112792994827032, + 0.012137548066675663, + 0.016371894627809525, + 0.012766803614795208, + 0.014826479367911816, + 0.013177123852074146, + 0.015116394497454166, + 0.012074282392859459, + 0.013402036391198635, + 0.002348166424781084, + 0.003951408434659243, + 0.0032839644700288773, + 0.0032730649691075087, + 0.0038895593024790287, + 0.0043813567608594894, + 0.0029083997942507267, + 0.0038812304846942425, + 0.009001199156045914, + 0.012945255264639854, + 0.014065908268094063, + 0.014163984917104244, + 0.015000545419752598, + 0.013466687873005867, + 0.011213020421564579, + 0.009889365173876286, + 0.0021374605130404234, + 0.0020642124582082033, + 0.0021769870072603226, + 0.002206027042120695, + 0.004483711440116167, + 0.0030843650456517935, + 0.0032591838389635086, + 0.003318167058750987, + 0.006941535044461489, + 0.010917027480900288, + 0.012085207737982273, + 0.017305564135313034, + 0.005278773605823517, + 0.013111582025885582, + 0.007977508008480072, + 0.009349050000309944, + 0.020606378093361855, + 0.03167218342423439, + 0.029754512012004852, + 0.04158936068415642, + 0.012721173465251923, + 0.029918929561972618, + 0.017996054142713547, + 0.036966074258089066, + 0.016728965565562248, + 0.03151104971766472, + 0.027447380125522614, + 0.033849991858005524, + 0.010004629381000996, + 0.026293473318219185, + 0.014765665866434574, + 0.026508601382374763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.002819374669343233 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634967729449272 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008634967729449272, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634967729449272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1231602355837822, + 0.0679301917552948, + 0.08446259051561356, + 0.2986627519130707, + 0.056712061166763306, + 0.0787695050239563, + 0.038087762892246246, + 0.15850324928760529, + 0.06250163167715073, + 0.03120998479425907 + ], + "credit_map": [ + 0.010376764461398125, + 0.017282020300626755, + 0.012080935761332512, + 0.014064330607652664, + 0.015687234699726105, + 0.016772063449025154, + 0.014508398249745369, + 0.022388501092791557, + 0.005216357298195362, + 0.008460385724902153, + 0.008489290252327919, + 0.012086407281458378, + 0.007349459454417229, + 0.007067509926855564, + 0.01032059732824564, + 0.008940194733440876, + 0.010817239060997963, + 0.008412884548306465, + 0.013732808642089367, + 0.01656452566385269, + 0.008038789965212345, + 0.010081281885504723, + 0.007335481233894825, + 0.009479579515755177, + 0.03248445689678192, + 0.047307197004556656, + 0.03577720373868942, + 0.04014526307582855, + 0.032192330807447433, + 0.03998193517327309, + 0.033013444393873215, + 0.03776095435023308, + 0.004563365131616592, + 0.007508997339755297, + 0.009284128434956074, + 0.007586037740111351, + 0.006055397912859917, + 0.007990486919879913, + 0.005025089252740145, + 0.008698559366166592, + 0.005874595604836941, + 0.007372968830168247, + 0.0068728686310350895, + 0.008211763575673103, + 0.016984691843390465, + 0.014247861690819263, + 0.009981164708733559, + 0.009223592467606068, + 0.0032543570268899202, + 0.005399931687861681, + 0.004006973467767239, + 0.006772959139198065, + 0.005136888939887285, + 0.005317960400134325, + 0.003776533529162407, + 0.004422160796821117, + 0.012409036047756672, + 0.027360884472727776, + 0.023252535611391068, + 0.029804358258843422, + 0.010833234526216984, + 0.034048546105623245, + 0.00912253838032484, + 0.011672116816043854, + 0.007274036295711994, + 0.009603938087821007, + 0.009250685572624207, + 0.010656720958650112, + 0.004426627419888973, + 0.009906383231282234, + 0.004217806737869978, + 0.0071654412895441055, + 0.0034352692309767008, + 0.004663790110498667, + 0.0039834752678871155, + 0.005177743267267942, + 0.0021829321049153805, + 0.004032312426716089, + 0.002965745748952031, + 0.004768719431012869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.008634967729449272 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009201106731779873 + ], + "opposite_logprob_tokens": [ + -7.000920295715332 + ], + "logprob_reward": -0.0009201106731779873, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009201106731779873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24922215938568115, + 0.12649862468242645, + 0.05659496411681175, + 0.04080439731478691, + 0.053036514669656754, + 0.06652932614088058, + 0.0817006304860115, + 0.03301941603422165, + 0.18630124628543854, + 0.10629281401634216 + ], + "credit_map": [ + 0.022272536531090736, + 0.03138172626495361, + 0.04290999472141266, + 0.029867049306631088, + 0.034154731780290604, + 0.03205794841051102, + 0.02714088372886181, + 0.02943730168044567, + 0.014650443568825722, + 0.02201288752257824, + 0.011559149250388145, + 0.016339678317308426, + 0.011839153245091438, + 0.019102195277810097, + 0.014840743504464626, + 0.016154363751411438, + 0.006987879518419504, + 0.0077256993390619755, + 0.005165267735719681, + 0.009147729724645615, + 0.006586846895515919, + 0.00620991550385952, + 0.006024873815476894, + 0.008746751584112644, + 0.004956562537699938, + 0.004663470666855574, + 0.003939103335142136, + 0.004892523866146803, + 0.005966177675873041, + 0.004881460685282946, + 0.0056739673018455505, + 0.005831127520650625, + 0.004018981475383043, + 0.00613780040293932, + 0.008228319697082043, + 0.007401801645755768, + 0.00765534583479166, + 0.00691853417083621, + 0.005082770716398954, + 0.0075929611921310425, + 0.005462770815938711, + 0.007065563928335905, + 0.008626258000731468, + 0.009075302630662918, + 0.012495264410972595, + 0.009269163943827152, + 0.008123477920889854, + 0.006411517970263958, + 0.006460451055318117, + 0.012575726956129074, + 0.01092056930065155, + 0.012847116217017174, + 0.008872305043041706, + 0.015294894576072693, + 0.006474125199019909, + 0.00825544260442257, + 0.002734092064201832, + 0.003891628934070468, + 0.004591614939272404, + 0.005263767670840025, + 0.0038802316412329674, + 0.00419987877830863, + 0.0031411077361553907, + 0.005317092407494783, + 0.01599530130624771, + 0.026386981830000877, + 0.02537688799202442, + 0.033645302057266235, + 0.013243941590189934, + 0.025346552953124046, + 0.01626679114997387, + 0.03003947250545025, + 0.01008808333426714, + 0.018561476841568947, + 0.016055583953857422, + 0.018406258895993233, + 0.007919864729046822, + 0.014990042895078659, + 0.008953191339969635, + 0.011318312957882881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0009201106731779873 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up, rendered in crisp 4K, captures a delicate wolf spider, its intricate seed head trembling gracefully against the overwhelming force of nature. Each feathery bristle of the plume is rendered in surreal detail, appearing almost silvery and translucent as it holds its shape. The camera utilizes a very shallow depth of field to achieve a profound sense of isolation, allowing this solitary seed head to remain in sharp, unwavering focus while a swirling blizzard of fine sand grains engulfs the surrounding space. The background dissolves into a chaotic, blurred vortex of motion, a turbulent swirl of golden-brown sand that whips and swirls violently. Intense sunlight, heavily scattered by the dense storm, bathes the scene in a bright, luminous mist, causing individual grains of sand flying past the lens to shimmer like tiny, fleeting sparks. Dominated by a warm, monochromatic palette of ochre and sepia, the entire image creates a powerful and atmospheric contrast between the tranquil, resilient beauty of the plant and the chaotic, all-consuming power of the sandstorm.", + "seed": 19806, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_2/index_8905_seed_19806_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1923073406331241, + "vlm_logprob": -0.1923073406331241, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.1923073406331241, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9018180966377258 + ], + "opposite_logprob_tokens": [ + -0.5268180966377258 + ], + "logprob_reward": -0.9018180966377258, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9018180966377258, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05749105289578438, + 0.31933796405792236, + 0.05528094619512558, + 0.029092729091644287, + 0.43254879117012024, + 0.06594359874725342, + 0.010858972556889057, + 0.011409084312617779, + 0.011222051456570625, + 0.00681485328823328 + ], + "credit_map": [ + 0.005367821082472801, + 0.0066339317709207535, + 0.006445818580687046, + 0.006694436073303223, + 0.012677519582211971, + 0.006251465994864702, + 0.00771959638223052, + 0.0057004583068192005, + 0.014294822700321674, + 0.013712277635931969, + 0.10228445380926132, + 0.016158973798155785, + 0.015369978733360767, + 0.11087504774332047, + 0.022578634321689606, + 0.024063779041171074, + 0.005056293681263924, + 0.004943746142089367, + 0.00919177196919918, + 0.0067254784516990185, + 0.009149800054728985, + 0.00883022416383028, + 0.006659938488155603, + 0.004723689053207636, + 0.0024978925939649343, + 0.003107556374743581, + 0.0044359308667480946, + 0.0024956802371889353, + 0.004353194963186979, + 0.007198026869446039, + 0.0028006157372146845, + 0.00220383214764297, + 0.04040626809000969, + 0.02184165082871914, + 0.1824110597372055, + 0.021166067570447922, + 0.01679150015115738, + 0.04936946928501129, + 0.019488655030727386, + 0.0810740739107132, + 0.00499459495767951, + 0.00436170119792223, + 0.005060251802206039, + 0.003143596462905407, + 0.00410431856289506, + 0.02370700053870678, + 0.0056792935356497765, + 0.014892835170030594, + 0.0011058259988203645, + 0.0013282796135172248, + 0.0009480511653237045, + 0.0009012360824272037, + 0.0012862435542047024, + 0.0022993916645646095, + 0.0014021575916558504, + 0.0015877865953370929, + 0.0013582765823230147, + 0.0013431854313239455, + 0.0014431917807087302, + 0.00114031124394387, + 0.0021590578835457563, + 0.0015014620730653405, + 0.0012761900434270501, + 0.0011874076444655657, + 0.0009396919049322605, + 0.0014468729496002197, + 0.0010975986951962113, + 0.0009097999427467585, + 0.0026433540042489767, + 0.0016136585036292672, + 0.0016841148026287556, + 0.0008869594894349575, + 0.0008838512003421783, + 0.0007419479079544544, + 0.0009538031881675124, + 0.0007315562688745558, + 0.0012479634024202824, + 0.0009869025088846684, + 0.0007210304611362517, + 0.0005477974773384631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.9018180966377258 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029830140992999077 + ], + "opposite_logprob_tokens": [ + -3.52983021736145 + ], + "logprob_reward": -0.029830140992999077, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029830140992999077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07679344713687897, + 0.3357046842575073, + 0.04505779221653938, + 0.03017328679561615, + 0.37793242931365967, + 0.08041372150182724, + 0.026071880012750626, + 0.007348456885665655, + 0.01508466899394989, + 0.005419581662863493 + ], + "credit_map": [ + 0.005969052202999592, + 0.008554008789360523, + 0.006849024910479784, + 0.006559939123690128, + 0.010116546414792538, + 0.016232095658779144, + 0.010544614866375923, + 0.0119681591168046, + 0.015273667871952057, + 0.013855715282261372, + 0.11317141354084015, + 0.017473869025707245, + 0.016239017248153687, + 0.11743845790624619, + 0.020074091851711273, + 0.02217845804989338, + 0.0039837355725467205, + 0.004854165017604828, + 0.006671702954918146, + 0.004610789939761162, + 0.01078074611723423, + 0.005617242306470871, + 0.005205824505537748, + 0.0033335869666188955, + 0.0017872159369289875, + 0.0020362362265586853, + 0.00682589877396822, + 0.0017648579087108374, + 0.003425908973440528, + 0.008583245798945427, + 0.0025994773022830486, + 0.0031504447106271982, + 0.03411167114973068, + 0.018511811271309853, + 0.16129621863365173, + 0.017859168350696564, + 0.013026563450694084, + 0.051516685634851456, + 0.017796888947486877, + 0.06381344795227051, + 0.004750008694827557, + 0.005527718458324671, + 0.012435217387974262, + 0.0035744511988013983, + 0.006800556089729071, + 0.025648191571235657, + 0.007726671639829874, + 0.013950902037322521, + 0.0030359847005456686, + 0.0026708648074418306, + 0.0038690115325152874, + 0.0025990563444793224, + 0.0030753796454519033, + 0.003435444552451372, + 0.00270101148635149, + 0.004685127642005682, + 0.0007231542840600014, + 0.0008724807412363589, + 0.0010117467027157545, + 0.0008218630100600421, + 0.001106860232539475, + 0.0009792842902243137, + 0.0010280752321705222, + 0.0008049923344515264, + 0.0012724632397294044, + 0.0020297749433666468, + 0.0016188102308660746, + 0.0015290172304958105, + 0.0034767608158290386, + 0.001663669478148222, + 0.0023058962542563677, + 0.00118827645201236, + 0.0006548400269821286, + 0.0007433599676005542, + 0.000765759265050292, + 0.0004826475342269987, + 0.0008138878620229661, + 0.0008064836147241294, + 0.0006146227242425084, + 0.0005379806971177459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.029830140992999077 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028259125538170338 + ], + "opposite_logprob_tokens": [ + -5.877825736999512 + ], + "logprob_reward": -0.0028259125538170338, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028259125538170338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1168067455291748, + 0.43592599034309387, + 0.11126361787319183, + 0.0516461618244648, + 0.14625318348407745, + 0.04952984303236008, + 0.04047451168298721, + 0.013516533188521862, + 0.015244939364492893, + 0.019338442012667656 + ], + "credit_map": [ + 0.010337204672396183, + 0.009487489238381386, + 0.01081534568220377, + 0.009587203152477741, + 0.015892960131168365, + 0.023844681680202484, + 0.009773493744432926, + 0.027068370953202248, + 0.020918630063533783, + 0.015531999990344048, + 0.15046223998069763, + 0.01636529713869095, + 0.01783713325858116, + 0.1576894223690033, + 0.02463621459901333, + 0.03248508274555206, + 0.008778844028711319, + 0.012910650111734867, + 0.014040139503777027, + 0.011192512698471546, + 0.02801329456269741, + 0.016217676922678947, + 0.012615242041647434, + 0.007495260797441006, + 0.004424473736435175, + 0.004592240788042545, + 0.011451485566794872, + 0.004153313580900431, + 0.008410381153225899, + 0.008673492819070816, + 0.004620823077857494, + 0.005319955758750439, + 0.016871649771928787, + 0.01356942392885685, + 0.022324074059724808, + 0.012865266762673855, + 0.01795186474919319, + 0.02318252995610237, + 0.023649942129850388, + 0.01583843305706978, + 0.004600711166858673, + 0.004000870976597071, + 0.010936315171420574, + 0.004156066104769707, + 0.004938425496220589, + 0.009340752847492695, + 0.0076365042477846146, + 0.003920196555554867, + 0.004040134139358997, + 0.00490040797740221, + 0.004963849205523729, + 0.0038642350118607283, + 0.0064775338396430016, + 0.0050134784542024136, + 0.004700277000665665, + 0.0065145958214998245, + 0.0016842340119183064, + 0.0013084684032946825, + 0.0017770207487046719, + 0.001365537173114717, + 0.002679485362023115, + 0.001978833694010973, + 0.0013914655428379774, + 0.001331488718278706, + 0.0016373364487662911, + 0.001582200638949871, + 0.0016392174875363708, + 0.0018915280234068632, + 0.002689132932573557, + 0.0022590497974306345, + 0.0017814065795391798, + 0.0017650671070441604, + 0.002195423934608698, + 0.002046976238489151, + 0.002228221856057644, + 0.0017653845716267824, + 0.004005126189440489, + 0.002238575369119644, + 0.0030921176075935364, + 0.0017666168278083205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.0028259125538170338 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026344336569309235 + ], + "opposite_logprob_tokens": [ + -3.6513442993164062 + ], + "logprob_reward": -0.026344336569309235, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026344336569309235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07602276653051376, + 0.06648489832878113, + 0.026096142828464508, + 0.06029025465250015, + 0.23951807618141174, + 0.43126380443573, + 0.07547135651111603, + 0.00841803103685379, + 0.009094861336052418, + 0.007339776493608952 + ], + "credit_map": [ + 0.0062442016787827015, + 0.0066022672690451145, + 0.014190882444381714, + 0.006770138628780842, + 0.008871310390532017, + 0.01562419906258583, + 0.007318676449358463, + 0.010401099920272827, + 0.005250255577266216, + 0.0050724404864013195, + 0.01582992821931839, + 0.008545815013349056, + 0.005760638043284416, + 0.013593182899057865, + 0.0070759947411715984, + 0.005356652196496725, + 0.0020133641082793474, + 0.002495334716513753, + 0.0033446247689425945, + 0.0029456063639372587, + 0.005087160039693117, + 0.0030539357103407383, + 0.003539277007803321, + 0.0036168440710753202, + 0.005648649297654629, + 0.006576872896403074, + 0.011353787034749985, + 0.005238606594502926, + 0.010817677713930607, + 0.007655522786080837, + 0.005662354175001383, + 0.007336788345128298, + 0.01890423148870468, + 0.013439005240797997, + 0.09878116846084595, + 0.012078247033059597, + 0.010451722890138626, + 0.03996015340089798, + 0.014166920445859432, + 0.03173663094639778, + 0.02777395397424698, + 0.028603216633200645, + 0.050150834023952484, + 0.011337505653500557, + 0.020966552197933197, + 0.18160326778888702, + 0.010374943725764751, + 0.10045358538627625, + 0.008386027999222279, + 0.007500584237277508, + 0.009668800979852676, + 0.007117908913642168, + 0.00779967475682497, + 0.013087747618556023, + 0.00774234626442194, + 0.0141682680696249, + 0.0009593574213795364, + 0.0007439170149154961, + 0.001426775474101305, + 0.0006184008088894188, + 0.0007503645610995591, + 0.00164474593475461, + 0.0007783790933899581, + 0.0014960924163460732, + 0.0009376632515341043, + 0.0011714634019881487, + 0.0012464928440749645, + 0.0007103580865077674, + 0.0014152060030028224, + 0.0013109773863106966, + 0.0009476416744291782, + 0.0013550593284890056, + 0.0007911899010650814, + 0.0007509606657549739, + 0.0009956138674169779, + 0.0007552593015134335, + 0.0011673872359097004, + 0.0009349608444608748, + 0.0010892635909840465, + 0.0008551417267881334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.026344336569309235 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a delicate wolf spider?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a seed head with feathery bristles visible trembling gracefully in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video utilize a very shallow depth of field with a blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene dominated by a swirling blizzard of fine golden-brown sand grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette a warm monochromatic mix of ochre and sepia?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007182164117693901 + ], + "opposite_logprob_tokens": [ + -7.250718116760254 + ], + "logprob_reward": -0.0007182164117693901, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007182164117693901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06677049398422241, + 0.2516839802265167, + 0.0536251924932003, + 0.03779515251517296, + 0.4287765324115753, + 0.05257118493318558, + 0.029676910489797592, + 0.016083620488643646, + 0.028535146266222, + 0.03448180854320526 + ], + "credit_map": [ + 0.007814825512468815, + 0.0067968349903821945, + 0.011101248674094677, + 0.007771578151732683, + 0.009710362181067467, + 0.008337215520441532, + 0.0074178967624902725, + 0.007820536382496357, + 0.013354960829019547, + 0.011466577649116516, + 0.08022011071443558, + 0.016688581556081772, + 0.013450545258820057, + 0.08025147020816803, + 0.01817096769809723, + 0.01808076724410057, + 0.004567768424749374, + 0.004976206924766302, + 0.0064977845177054405, + 0.007047058083117008, + 0.009522325359284878, + 0.007621380500495434, + 0.007247939705848694, + 0.006144726183265448, + 0.0037589678540825844, + 0.004686661530286074, + 0.004995764698833227, + 0.003540391568094492, + 0.0062417746521532536, + 0.007258026860654354, + 0.003817443735897541, + 0.003496121848002076, + 0.040704578161239624, + 0.02353140525519848, + 0.178117573261261, + 0.021238425746560097, + 0.018083687871694565, + 0.04671873524785042, + 0.02431633323431015, + 0.0760657861828804, + 0.0045958952978253365, + 0.003146285191178322, + 0.007579940836876631, + 0.00354778696782887, + 0.003646582132205367, + 0.014060471206903458, + 0.0036092789378017187, + 0.012384945526719093, + 0.002908804453909397, + 0.0031702269334346056, + 0.0028367158956825733, + 0.0030041548889130354, + 0.003873260226100683, + 0.005505051929503679, + 0.0037394482642412186, + 0.004639248363673687, + 0.002045835368335247, + 0.0017265577334910631, + 0.00250297412276268, + 0.0019906125962734222, + 0.002179056406021118, + 0.0023446690756827593, + 0.0012616765452548862, + 0.0020322378259152174, + 0.003168343799188733, + 0.003637943184003234, + 0.002916845725849271, + 0.0023502910044044256, + 0.006018512882292271, + 0.004226423799991608, + 0.0037823556922376156, + 0.0024344308767467737, + 0.0042455922812223434, + 0.003396984888240695, + 0.006384906359016895, + 0.0025522245559841394, + 0.004123092629015446, + 0.006808644160628319, + 0.0028649938758462667, + 0.004105369560420513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.0007182164117693901 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_3/index_tensor([5276, 5276])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_3/index_tensor([5276, 5276])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..31ee233b34599fc97a54383a08ed2cbd24a85da3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_3/index_tensor([5276, 5276])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This dramatic, high-speed slow-motion close-up captures the intricate courtship ritual of two Guangdong South Sea crabs in a captivating and lavish classic cinematic style. The camera, positioned at eye level, executes a nearly imperceptible insertion, heightening the emotional weight of the exchange. They perch on a dark, smooth rock, its surface gleaming with a deep, soft green moss. The pursuer, a larger male crab, possesses a vibrant, iridescent violet on his carapace and fiery red-orange legs. He advances with a steady and cautious demeanor, his eyestalks aligned with his opponent. On his right claw, he meticulously and delicately presents a shimmering, small pearl shell that gleams captivatingly in the light. The second crab, slightly smaller and more subdued in color, maintains a poised demeanor, its posture both shy and expectant. The entire exchange unfolds in exquisite slow motion, emphasizing the subtle, gentle extension of the claws and the almost imperceptible movement of the recipient's body. As the pursuer moves, shimmering droplets of water spray from his carapace in beautiful, suspended arcs. The scene is bathed in dramatic, high-contrast lighting, with a powerful key light piercing the darkness, casting brilliant highlights on their damp carapaces and the shells they provide, creating deep, soft shadows that add to the luxurious cinematic atmosphere. The rich colorful palette and intense attention to micro-movement create a remarkably tender and visually stunning narrative.", + "seed": 18330, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_3/index_5276_seed_18330_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04087017923593521, + "vlm_logprob": -0.04087017923593521, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.04087017923593521, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01554561872035265 + ], + "opposite_logprob_tokens": [ + -4.265545845031738 + ], + "logprob_reward": -0.01554561872035265, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01554561872035265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34529903531074524, + 0.0646577849984169, + 0.01918119192123413, + 0.11335469037294388, + 0.27008432149887085, + 0.08804916590452194, + 0.045504868030548096, + 0.017445199191570282, + 0.014884259551763535, + 0.02153952792286873 + ], + "credit_map": [ + 0.09565380960702896, + 0.016332373023033142, + 0.014615253545343876, + 0.018923960626125336, + 0.017794188112020493, + 0.017750568687915802, + 0.13586728274822235, + 0.028361594304442406, + 0.006426397245377302, + 0.015203816816210747, + 0.004126648884266615, + 0.0048813726752996445, + 0.009186728857457638, + 0.0040292078629136086, + 0.006654816679656506, + 0.014148792251944542, + 0.0017551532946527004, + 0.002350699855014682, + 0.002615969628095627, + 0.0027426618617028, + 0.0018946302589029074, + 0.002078593708574772, + 0.0027069433126598597, + 0.0030365376733243465, + 0.011184665374457836, + 0.012650564312934875, + 0.01628986746072769, + 0.011138268746435642, + 0.01846729777753353, + 0.014452535659074783, + 0.01382213830947876, + 0.015349341556429863, + 0.027650216594338417, + 0.020778289064764977, + 0.039829354733228683, + 0.04190969467163086, + 0.03282967582345009, + 0.025412915274500847, + 0.028870608657598495, + 0.05280352383852005, + 0.006166552659124136, + 0.003077640663832426, + 0.004467187914997339, + 0.005956563632935286, + 0.00427973223850131, + 0.027204204350709915, + 0.0057772439904510975, + 0.03112003207206726, + 0.004269490484148264, + 0.002815110841766, + 0.0036704454105347395, + 0.004577024839818478, + 0.0021825863514095545, + 0.010047894902527332, + 0.002707310253754258, + 0.015235000289976597, + 0.002136100782081485, + 0.0017743568168953061, + 0.0023547119926661253, + 0.0022501053754240274, + 0.001880540163256228, + 0.0017640560399740934, + 0.0014828512212261558, + 0.0038024752866476774, + 0.0021431627683341503, + 0.0012443636078387499, + 0.0019317546393722296, + 0.002303078304976225, + 0.0016680374974384904, + 0.0011716444278135896, + 0.0014547661412507296, + 0.002967451000586152, + 0.0014385742833837867, + 0.0023661323357373476, + 0.0014999292325228453, + 0.003200849983841181, + 0.0020662350580096245, + 0.006756301503628492, + 0.001660286681726575, + 0.002551217330619693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.01554561872035265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978990737348795 + ], + "opposite_logprob_tokens": [ + -4.634789943695068 + ], + "logprob_reward": -0.00978990737348795, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978990737348795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2742300033569336, + 0.153977632522583, + 0.028851555660367012, + 0.11411577463150024, + 0.2602872848510742, + 0.01950785517692566, + 0.047239676117897034, + 0.050273597240448, + 0.02311587892472744, + 0.028400760143995285 + ], + "credit_map": [ + 0.0704679787158966, + 0.019867483526468277, + 0.013067487627267838, + 0.018211791291832924, + 0.012490359134972095, + 0.023953551426529884, + 0.09323901683092117, + 0.022932346910238266, + 0.011519249528646469, + 0.042386703193187714, + 0.009097590111196041, + 0.011841919273138046, + 0.013209610246121883, + 0.01115973200649023, + 0.013007394038140774, + 0.041755422949790955, + 0.002699469216167927, + 0.003808537032455206, + 0.003444127505645156, + 0.004783375188708305, + 0.0033939648419618607, + 0.004147527273744345, + 0.0028909246902912855, + 0.0036836282815784216, + 0.011250553652644157, + 0.016448866575956345, + 0.01536885928362608, + 0.012975471094250679, + 0.011683904565870762, + 0.018324751406908035, + 0.01311881747096777, + 0.014944547787308693, + 0.023056194186210632, + 0.020673319697380066, + 0.0371258519589901, + 0.04136427864432335, + 0.03540227934718132, + 0.025118034332990646, + 0.031216558068990707, + 0.046330776065588, + 0.002512304810807109, + 0.0012028759811073542, + 0.0026493542827665806, + 0.00295836734585464, + 0.0021132153924554586, + 0.0022303410805761814, + 0.002061801729723811, + 0.0037795945536345243, + 0.0033916416577994823, + 0.002770000835880637, + 0.001888187718577683, + 0.005554853938519955, + 0.0029270423110574484, + 0.012424170970916748, + 0.003067529061809182, + 0.01521624717861414, + 0.004809971898794174, + 0.0047192866913974285, + 0.006160435266792774, + 0.006548471283167601, + 0.00649893656373024, + 0.008233683183789253, + 0.005387559533119202, + 0.0079152537509799, + 0.003566933795809746, + 0.0020355572924017906, + 0.0032331563998013735, + 0.0036429986357688904, + 0.0017997591057792306, + 0.0022341173607856035, + 0.0019408772932365537, + 0.00466248020529747, + 0.001533645554445684, + 0.003199029481038451, + 0.001967655960470438, + 0.003345328150317073, + 0.0021964756306260824, + 0.010025850497186184, + 0.001703845919109881, + 0.004428929649293423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.00978990737348795 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01106967218220234 + ], + "opposite_logprob_tokens": [ + -4.5110697746276855 + ], + "logprob_reward": -0.01106967218220234, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01106967218220234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7168583273887634, + 0.017154304310679436, + 0.02036493457853794, + 0.027338286861777306, + 0.04631888493895531, + 0.07743623107671738, + 0.059688493609428406, + 0.010873714461922646, + 0.008382930420339108, + 0.015583856962621212 + ], + "credit_map": [ + 0.19890788197517395, + 0.036043357104063034, + 0.028109949082136154, + 0.040175992995500565, + 0.034889332950115204, + 0.039171166718006134, + 0.2814037501811981, + 0.05815685912966728, + 0.0016714280936866999, + 0.003182833082973957, + 0.002057933248579502, + 0.002463722601532936, + 0.0022287045139819384, + 0.0016429733950644732, + 0.0020998173858970404, + 0.0018068900099024177, + 0.0017938140081241727, + 0.002687126863747835, + 0.0019092034781351686, + 0.0024086434859782457, + 0.0019059473415836692, + 0.0017231096280738711, + 0.0037845298647880554, + 0.004152558743953705, + 0.0030913392547518015, + 0.0023579068947583437, + 0.004146262537688017, + 0.0029390533454716206, + 0.004016703926026821, + 0.0032926478888839483, + 0.0031506444793194532, + 0.004343726672232151, + 0.005371099803596735, + 0.004292461089789867, + 0.007444463204592466, + 0.006368220783770084, + 0.005225119646638632, + 0.004946141503751278, + 0.003988079261034727, + 0.008683296851813793, + 0.00875895470380783, + 0.0038002911023795605, + 0.004589044023305178, + 0.018417375162243843, + 0.004460921976715326, + 0.014401056803762913, + 0.004401611629873514, + 0.018606971949338913, + 0.005526324268430471, + 0.0038697724230587482, + 0.005418123211711645, + 0.004322814755141735, + 0.003305818885564804, + 0.015412015840411186, + 0.0037064726930111647, + 0.01812714897096157, + 0.0012430396163836122, + 0.0009977803565561771, + 0.0013572118477895856, + 0.002350879367440939, + 0.001100713387131691, + 0.0010607307776808739, + 0.0010563303949311376, + 0.0017070273170247674, + 0.001224957057274878, + 0.0008472115732729435, + 0.0011047901352867484, + 0.0012498885625973344, + 0.0006397520191967487, + 0.001039781840518117, + 0.0008185005863197148, + 0.0014580471906811, + 0.001245840103365481, + 0.0017158030532300472, + 0.0015424909070134163, + 0.0025333708617836237, + 0.0012967870570719242, + 0.003943706396967173, + 0.0015475061954930425, + 0.0017583509907126427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.01106967218220234 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895727455615997 + ], + "opposite_logprob_tokens": [ + -2.5789573192596436 + ], + "logprob_reward": -0.07895727455615997, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895727455615997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17424830794334412, + 0.4376312494277954, + 0.031109537929296494, + 0.09565845876932144, + 0.026948004961013794, + 0.08710359036922455, + 0.03013909049332142, + 0.01763005182147026, + 0.01810486800968647, + 0.08142687380313873 + ], + "credit_map": [ + 0.03144094720482826, + 0.009491421282291412, + 0.006076260004192591, + 0.013217566534876823, + 0.010922521352767944, + 0.010867445729672909, + 0.0732140988111496, + 0.01901804283261299, + 0.03744318336248398, + 0.16005860269069672, + 0.015678513795137405, + 0.02451196312904358, + 0.02711321786046028, + 0.022280694916844368, + 0.01427693571895361, + 0.13626813888549805, + 0.002776979934424162, + 0.0034699286334216595, + 0.003798223100602627, + 0.005315369460731745, + 0.0032812641002237797, + 0.0030521287117153406, + 0.003975871950387955, + 0.0054397727362811565, + 0.008210446685552597, + 0.013023651205003262, + 0.014559581875801086, + 0.010208473540842533, + 0.01094395574182272, + 0.014209035784006119, + 0.01132703572511673, + 0.013176270760595798, + 0.0025103676598519087, + 0.002065068343654275, + 0.003110115649178624, + 0.005133248865604401, + 0.0027558018919080496, + 0.0029974442441016436, + 0.003037689719349146, + 0.005338267423212528, + 0.009118099696934223, + 0.004087753593921661, + 0.0065899137407541275, + 0.021290497854351997, + 0.005836822558194399, + 0.010462677106261253, + 0.005355959292501211, + 0.0243618655949831, + 0.004048462491482496, + 0.0028134838212281466, + 0.0037089898250997066, + 0.004939468111842871, + 0.002799319801852107, + 0.004831321537494659, + 0.002944330684840679, + 0.004053711891174316, + 0.0018393825739622116, + 0.0018644989468157291, + 0.0019207908771932125, + 0.0026580661069601774, + 0.0020485548302531242, + 0.00231325370259583, + 0.0017075437353923917, + 0.003277959767729044, + 0.0027757002972066402, + 0.002065025269985199, + 0.0024239923804998398, + 0.002392573980614543, + 0.0016606129938736558, + 0.0021527563221752644, + 0.0015716975321993232, + 0.0030625092331320047, + 0.004521661438047886, + 0.007956909015774727, + 0.006753092166036367, + 0.017269661650061607, + 0.00625181756913662, + 0.026225492358207703, + 0.005302677862346172, + 0.007145565934479237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.07895727455615997 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08898842334747314 + ], + "opposite_logprob_tokens": [ + -2.4639883041381836 + ], + "logprob_reward": -0.08898842334747314, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08898842334747314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5120579600334167, + 0.09191257506608963, + 0.018968140706419945, + 0.037838343530893326, + 0.06871514022350311, + 0.13011020421981812, + 0.08595171570777893, + 0.027941059321165085, + 0.016603166237473488, + 0.009901747107505798 + ], + "credit_map": [ + 0.12304268777370453, + 0.029684383422136307, + 0.0225828904658556, + 0.03244034945964813, + 0.030075708404183388, + 0.03282862901687622, + 0.19710400700569153, + 0.044299300760030746, + 0.0076065645553171635, + 0.029394730925559998, + 0.0048562996089458466, + 0.005994499195367098, + 0.0066062030382454395, + 0.005341387819498777, + 0.0056312549859285355, + 0.026481634005904198, + 0.0017011590534821153, + 0.002957472577691078, + 0.0019218152156099677, + 0.00202818657271564, + 0.0026638254057615995, + 0.002201464958488941, + 0.0024813374038785696, + 0.0030128811486065388, + 0.003520007012411952, + 0.0046195900067687035, + 0.004453509580343962, + 0.003962226677685976, + 0.0043555451557040215, + 0.005824112333357334, + 0.004904371686279774, + 0.0061989822424948215, + 0.006802608259022236, + 0.005061971955001354, + 0.009929385967552662, + 0.01028609648346901, + 0.008105641230940819, + 0.006852892693132162, + 0.007630174979567528, + 0.014046362601220608, + 0.015092829242348671, + 0.006074924953281879, + 0.007837229408323765, + 0.028361815959215164, + 0.007800949737429619, + 0.009114890359342098, + 0.008675441145896912, + 0.047152116894721985, + 0.011305347084999084, + 0.006494125816971064, + 0.011527576483786106, + 0.012143680825829506, + 0.009093841537833214, + 0.009894683957099915, + 0.008699171245098114, + 0.01679329201579094, + 0.003694485407322645, + 0.002733952598646283, + 0.003079804126173258, + 0.004257745575159788, + 0.0029232765082269907, + 0.003271005116403103, + 0.0032425487879663706, + 0.004738241899758577, + 0.002472589025273919, + 0.001537522068247199, + 0.0021761234384030104, + 0.002919920487329364, + 0.0015931427478790283, + 0.0018040323629975319, + 0.0014386182883754373, + 0.002661218401044607, + 0.0012438169214874506, + 0.0011492318008095026, + 0.0014549228362739086, + 0.0011856906348839402, + 0.0012552367988973856, + 0.0013776905834674835, + 0.000935237854719162, + 0.0012999202590435743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.08898842334747314 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This dramatic, high-speed slow-motion close-up captures the intricate courtship ritual of two Guangdong South Sea crabs in a captivating and lavish classic cinematic style. The camera, positioned at eye level, executes a nearly imperceptible insertion, heightening the emotional weight of the exchange. They perch on a dark, smooth rock, its surface gleaming with a deep, soft green moss. The pursuer, a larger male crab, possesses a vibrant, iridescent violet on his carapace and fiery red-orange legs. He advances with a steady and cautious demeanor, his eyestalks aligned with his opponent. On his right claw, he meticulously and delicately presents a shimmering, small pearl shell that gleams captivatingly in the light. The second crab, slightly smaller and more subdued in color, maintains a poised demeanor, its posture both shy and expectant. The entire exchange unfolds in exquisite slow motion, emphasizing the subtle, gentle extension of the claws and the almost imperceptible movement of the recipient's body. As the pursuer moves, shimmering droplets of water spray from his carapace in beautiful, suspended arcs. The scene is bathed in dramatic, high-contrast lighting, with a powerful key light piercing the darkness, casting brilliant highlights on their damp carapaces and the shells they provide, creating deep, soft shadows that add to the luxurious cinematic atmosphere. The rich colorful palette and intense attention to micro-movement create a remarkably tender and visually stunning narrative.", + "seed": 18331, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_3/index_5276_seed_18331_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.048726525343954566, + "vlm_logprob": -0.048726525343954566, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.048726525343954566, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017414990812540054 + ], + "opposite_logprob_tokens": [ + -4.1424150466918945 + ], + "logprob_reward": -0.017414990812540054, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.017414990812540054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15969720482826233, + 0.08253887295722961, + 0.20040597021579742, + 0.0401308573782444, + 0.06615111976861954, + 0.04232073575258255, + 0.05826578661799431, + 0.1257971227169037, + 0.14325343072414398, + 0.0814388319849968 + ], + "credit_map": [ + 0.025932352989912033, + 0.018407421186566353, + 0.017385799437761307, + 0.01862877607345581, + 0.01880577951669693, + 0.01565951481461525, + 0.02015070430934429, + 0.024726862087845802, + 0.01328245084732771, + 0.009960577823221684, + 0.012083755806088448, + 0.009956964291632175, + 0.009407736361026764, + 0.007780355401337147, + 0.009908218868076801, + 0.010158817283809185, + 0.01977945864200592, + 0.014215663075447083, + 0.0281352736055851, + 0.023479409515857697, + 0.022748341783881187, + 0.03874662145972252, + 0.02831469289958477, + 0.024986507371068, + 0.005357472226023674, + 0.003999410197138786, + 0.005153205711394548, + 0.004885920789092779, + 0.0049974736757576466, + 0.003849172964692116, + 0.005693635903298855, + 0.006194565910845995, + 0.006820489186793566, + 0.0072804950177669525, + 0.009825668297708035, + 0.008571513928472996, + 0.008501376025378704, + 0.007613318972289562, + 0.008059478364884853, + 0.009478775784373283, + 0.005909545347094536, + 0.0034786625765264034, + 0.007208012975752354, + 0.007281701546162367, + 0.0037204623222351074, + 0.0043623545207083225, + 0.004309345502406359, + 0.006050650961697102, + 0.00729026086628437, + 0.007578573189675808, + 0.007724219933152199, + 0.009300567209720612, + 0.005294962786138058, + 0.0062270937487483025, + 0.006329177878797054, + 0.008520928211510181, + 0.02118466980755329, + 0.013035318814218044, + 0.018101707100868225, + 0.015712285414338112, + 0.014591893181204796, + 0.010462291538715363, + 0.011608206667006016, + 0.021100755780935287, + 0.01975402422249317, + 0.015443023294210434, + 0.017544707283377647, + 0.02187185361981392, + 0.017618950456380844, + 0.01228360366076231, + 0.018583184108138084, + 0.020154081284999847, + 0.011204134672880173, + 0.008870052173733711, + 0.008946312591433525, + 0.01473007258027792, + 0.007995319552719593, + 0.006564284674823284, + 0.007651606574654579, + 0.015477045439183712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.017414990812540054 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644777350127697 + ], + "opposite_logprob_tokens": [ + -4.7586445808410645 + ], + "logprob_reward": -0.008644777350127697, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644777350127697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17016072571277618, + 0.16639842092990875, + 0.171907439827919, + 0.05576661601662636, + 0.05465179309248924, + 0.07858475297689438, + 0.07947199791669846, + 0.05641476809978485, + 0.05563145503401756, + 0.11101198941469193 + ], + "credit_map": [ + 0.028570884838700294, + 0.0179898664355278, + 0.021343449130654335, + 0.01852450706064701, + 0.019367968663573265, + 0.014687798917293549, + 0.025514649227261543, + 0.024161603301763535, + 0.018684931099414825, + 0.02112363837659359, + 0.02023622952401638, + 0.02016773261129856, + 0.016586581245064735, + 0.024485405534505844, + 0.023462019860744476, + 0.021651888266205788, + 0.021050550043582916, + 0.026941493153572083, + 0.02120998315513134, + 0.018408270552754402, + 0.01676364243030548, + 0.022376932203769684, + 0.02343781664967537, + 0.02171875350177288, + 0.009086688049137592, + 0.005390594247728586, + 0.006151536479592323, + 0.005786902271211147, + 0.008239449933171272, + 0.006394647527486086, + 0.0069733066484332085, + 0.007743495982140303, + 0.007152751088142395, + 0.005032345652580261, + 0.0064306315034627914, + 0.007765762507915497, + 0.00560223963111639, + 0.006518306210637093, + 0.00839659757912159, + 0.007753159385174513, + 0.013211291283369064, + 0.009773721918463707, + 0.010110849514603615, + 0.012106510810554028, + 0.007241201587021351, + 0.006090572569519281, + 0.007116117514669895, + 0.012934491969645023, + 0.010931342840194702, + 0.01002035103738308, + 0.00925131794065237, + 0.01407466921955347, + 0.007629188243299723, + 0.008315695449709892, + 0.007180534303188324, + 0.012068895623087883, + 0.011647175066173077, + 0.00635532708838582, + 0.006901935208588839, + 0.0066453516483306885, + 0.006117462180554867, + 0.004999333526939154, + 0.005400575697422028, + 0.008347603492438793, + 0.008264187723398209, + 0.007938300259411335, + 0.005687038879841566, + 0.007463767193257809, + 0.006598717533051968, + 0.004650484770536423, + 0.0059689669869840145, + 0.009059996344149113, + 0.018146807327866554, + 0.009710552170872688, + 0.010279055684804916, + 0.01402512937784195, + 0.011268924921751022, + 0.009752574376761913, + 0.011583964340388775, + 0.026244984939694405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.008644777350127697 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057366631925106 + ], + "opposite_logprob_tokens": [ + -3.895573616027832 + ], + "logprob_reward": -0.02057366631925106, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057366631925106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28168442845344543, + 0.12590380012989044, + 0.0620812252163887, + 0.0913865938782692, + 0.09010428190231323, + 0.05924513190984726, + 0.06300843507051468, + 0.07595740258693695, + 0.05804315209388733, + 0.09258556365966797 + ], + "credit_map": [ + 0.03566616028547287, + 0.03577329218387604, + 0.03321423381567001, + 0.034148599952459335, + 0.03095652535557747, + 0.027960188686847687, + 0.04192977398633957, + 0.04203563183546066, + 0.012769529595971107, + 0.01115050632506609, + 0.03266822546720505, + 0.01150993350893259, + 0.010903283953666687, + 0.014340760186314583, + 0.0147042041644454, + 0.01785736158490181, + 0.005717568099498749, + 0.0060053435154259205, + 0.008625693619251251, + 0.007858171127736568, + 0.008431284688413143, + 0.008430832996964455, + 0.009315344505012035, + 0.007696985732764006, + 0.011549677699804306, + 0.01171187125146389, + 0.010259181261062622, + 0.010655195452272892, + 0.011106066405773163, + 0.008428951725363731, + 0.012835546396672726, + 0.014840108342468739, + 0.015131768770515919, + 0.009174128994345665, + 0.013610852882266045, + 0.01054489053785801, + 0.006831581704318523, + 0.009644882753491402, + 0.01181312557309866, + 0.013353057205677032, + 0.012740472331643105, + 0.005486819427460432, + 0.007757443934679031, + 0.009120198898017406, + 0.005387965589761734, + 0.004967452958226204, + 0.00540912477299571, + 0.008375655859708786, + 0.00984710082411766, + 0.007519592065364122, + 0.008939280174672604, + 0.00811123475432396, + 0.006095094606280327, + 0.006961403414607048, + 0.0059002116322517395, + 0.009634510613977909, + 0.014486691914498806, + 0.008291438221931458, + 0.010468080639839172, + 0.008938007056713104, + 0.007405138108879328, + 0.006162874400615692, + 0.006722759455442429, + 0.013482416979968548, + 0.008135301060974598, + 0.007321795914322138, + 0.007554307114332914, + 0.008238083682954311, + 0.005945382174104452, + 0.0057245539501309395, + 0.00709565170109272, + 0.008028073236346245, + 0.012085162103176117, + 0.010889053344726562, + 0.009332542307674885, + 0.012227746658027172, + 0.008484292775392532, + 0.011988082900643349, + 0.009222395718097687, + 0.018356280401349068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.02057366631925106 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999869644641876 + ], + "opposite_logprob_tokens": [ + -2.6949987411499023 + ], + "logprob_reward": -0.06999869644641876, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999869644641876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2999364733695984, + 0.1808849722146988, + 0.07895684242248535, + 0.0552668496966362, + 0.04669361934065819, + 0.061593443155288696, + 0.07443396002054214, + 0.04133689031004906, + 0.05415753647685051, + 0.10673940181732178 + ], + "credit_map": [ + 0.02861970104277134, + 0.03374742344021797, + 0.05327780917286873, + 0.036688677966594696, + 0.03483125567436218, + 0.027925675734877586, + 0.03677765652537346, + 0.048068296164274216, + 0.020661000162363052, + 0.030401799827814102, + 0.021732868626713753, + 0.02908896468579769, + 0.022481560707092285, + 0.015872212126851082, + 0.017789850011467934, + 0.022856732830405235, + 0.008986735716462135, + 0.00836024060845375, + 0.011549090966582298, + 0.009787610732018948, + 0.008557075634598732, + 0.011904915794730186, + 0.009588386863470078, + 0.010222794488072395, + 0.007283767685294151, + 0.005545017775148153, + 0.0070870802737772465, + 0.007300300989300013, + 0.005537377670407295, + 0.005033787805587053, + 0.008121168240904808, + 0.009358354844152927, + 0.005030885338783264, + 0.004315152298659086, + 0.00482257129624486, + 0.006678189616650343, + 0.00541979493573308, + 0.0054800366051495075, + 0.008203809149563313, + 0.006743182428181171, + 0.009049133397638798, + 0.004893479868769646, + 0.008099619299173355, + 0.010120081715285778, + 0.006210902240127325, + 0.007953113876283169, + 0.007421211805194616, + 0.007845902815461159, + 0.011426784098148346, + 0.006857968866825104, + 0.010561385191977024, + 0.010414806194603443, + 0.008081444539129734, + 0.008602200075984001, + 0.0073100426234304905, + 0.011179335415363312, + 0.00550492387264967, + 0.002795410342514515, + 0.005143247544765472, + 0.006006412208080292, + 0.005960486363619566, + 0.005768751725554466, + 0.0039749532006680965, + 0.006182708311825991, + 0.009459026157855988, + 0.005533919204026461, + 0.0050466931425035, + 0.006785064935684204, + 0.007866496220231056, + 0.004690540488809347, + 0.005308171734213829, + 0.009467627853155136, + 0.017327290028333664, + 0.007405833341181278, + 0.009946116246283054, + 0.014232201501727104, + 0.009935365989804268, + 0.009113659150898457, + 0.008903954178094864, + 0.02987498790025711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.06999869644641876 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12700049579143524 + ], + "opposite_logprob_tokens": [ + -2.127000570297241 + ], + "logprob_reward": -0.12700049579143524, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12700049579143524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24785828590393066, + 0.16886331140995026, + 0.12072170525789261, + 0.05737576261162758, + 0.06120286509394646, + 0.06135135889053345, + 0.07094621658325195, + 0.06403567641973495, + 0.0569477304816246, + 0.09069701284170151 + ], + "credit_map": [ + 0.030050428584218025, + 0.027667712420225143, + 0.04611896723508835, + 0.02655036561191082, + 0.030652426183223724, + 0.022312747314572334, + 0.031113842502236366, + 0.0333917960524559, + 0.01777329295873642, + 0.018019448965787888, + 0.021324533969163895, + 0.021183304488658905, + 0.019051743671298027, + 0.014355901628732681, + 0.03184804692864418, + 0.02530704252421856, + 0.011835968121886253, + 0.014114594087004662, + 0.016374066472053528, + 0.011901699006557465, + 0.016641216352581978, + 0.01934194564819336, + 0.015955837443470955, + 0.014556382782757282, + 0.008110051043331623, + 0.006938290316611528, + 0.00684022344648838, + 0.007010690402239561, + 0.007373772095888853, + 0.00569846760481596, + 0.00719932047650218, + 0.008204949088394642, + 0.007666202262043953, + 0.006091884803026915, + 0.008484802208840847, + 0.008709938265383244, + 0.0070648216642439365, + 0.005953500512987375, + 0.00853654369711876, + 0.00869517121464014, + 0.0077705904841423035, + 0.005571495741605759, + 0.007453734986484051, + 0.010493851266801357, + 0.006308353040367365, + 0.008285590447485447, + 0.006680815946310759, + 0.008786927908658981, + 0.00950947217643261, + 0.007180468644946814, + 0.009220588020980358, + 0.011014297604560852, + 0.006864264607429504, + 0.008392657153308392, + 0.006900172680616379, + 0.011864293366670609, + 0.009001435711979866, + 0.0064523955807089806, + 0.008085446432232857, + 0.010404221713542938, + 0.005253705196082592, + 0.006011808756738901, + 0.005723114591091871, + 0.013103551231324673, + 0.009214220568537712, + 0.007577843964099884, + 0.005351019091904163, + 0.008589129894971848, + 0.005426297429949045, + 0.004081048537045717, + 0.006980912294238806, + 0.009727261029183865, + 0.012655048631131649, + 0.008252006024122238, + 0.011482413858175278, + 0.013661462813615799, + 0.0064689964056015015, + 0.009001542814075947, + 0.008763957768678665, + 0.02041158266365528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.12700049579143524 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This dramatic, high-speed slow-motion close-up captures the intricate courtship ritual of two Guangdong South Sea crabs in a captivating and lavish classic cinematic style. The camera, positioned at eye level, executes a nearly imperceptible insertion, heightening the emotional weight of the exchange. They perch on a dark, smooth rock, its surface gleaming with a deep, soft green moss. The pursuer, a larger male crab, possesses a vibrant, iridescent violet on his carapace and fiery red-orange legs. He advances with a steady and cautious demeanor, his eyestalks aligned with his opponent. On his right claw, he meticulously and delicately presents a shimmering, small pearl shell that gleams captivatingly in the light. The second crab, slightly smaller and more subdued in color, maintains a poised demeanor, its posture both shy and expectant. The entire exchange unfolds in exquisite slow motion, emphasizing the subtle, gentle extension of the claws and the almost imperceptible movement of the recipient's body. As the pursuer moves, shimmering droplets of water spray from his carapace in beautiful, suspended arcs. The scene is bathed in dramatic, high-contrast lighting, with a powerful key light piercing the darkness, casting brilliant highlights on their damp carapaces and the shells they provide, creating deep, soft shadows that add to the luxurious cinematic atmosphere. The rich colorful palette and intense attention to micro-movement create a remarkably tender and visually stunning narrative.", + "seed": 18332, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_3/index_5276_seed_18332_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04358803369104862, + "vlm_logprob": -0.04358803369104862, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.04358803369104862, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01592535711824894 + ], + "opposite_logprob_tokens": [ + -4.265925407409668 + ], + "logprob_reward": -0.01592535711824894, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01592535711824894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19172066450119019, + 0.04615514725446701, + 0.0740397498011589, + 0.06518367677927017, + 0.09545164555311203, + 0.0555831640958786, + 0.09594017267227173, + 0.1674087792634964, + 0.15457987785339355, + 0.053937118500471115 + ], + "credit_map": [ + 0.015855612233281136, + 0.021625982597470284, + 0.037617702037096024, + 0.019592614844441414, + 0.030003199353814125, + 0.015614572912454605, + 0.03015877492725849, + 0.02125222235918045, + 0.004823834635317326, + 0.005856864620000124, + 0.005544721614569426, + 0.006096204277127981, + 0.0078150425106287, + 0.009271006099879742, + 0.003206472145393491, + 0.0035410025157034397, + 0.009225494228303432, + 0.008677652105689049, + 0.009241298772394657, + 0.011860005557537079, + 0.009787799790501595, + 0.009737726300954819, + 0.010582427494227886, + 0.004927351605147123, + 0.00808381475508213, + 0.007477059960365295, + 0.007707861252129078, + 0.00951633881777525, + 0.008373746648430824, + 0.008422691375017166, + 0.00841427594423294, + 0.007187885232269764, + 0.006917332299053669, + 0.03609300032258034, + 0.006969337817281485, + 0.006250939331948757, + 0.004627698101103306, + 0.023911040276288986, + 0.006006777286529541, + 0.00467551639303565, + 0.005689037498086691, + 0.00972291361540556, + 0.0048476639203727245, + 0.008149093016982079, + 0.004313117358833551, + 0.011039775796234608, + 0.006553018931299448, + 0.005268543027341366, + 0.012232518754899502, + 0.008465488441288471, + 0.01282805111259222, + 0.015287755988538265, + 0.007857983000576496, + 0.01573558896780014, + 0.010455692186951637, + 0.013077090494334698, + 0.019769903272390366, + 0.011420771479606628, + 0.01584811322391033, + 0.03555291146039963, + 0.012741947546601295, + 0.04282456636428833, + 0.014300270937383175, + 0.014950280077755451, + 0.03297543153166771, + 0.005839666351675987, + 0.009581351652741432, + 0.015906594693660736, + 0.00986171793192625, + 0.007267454639077187, + 0.057715438306331635, + 0.015432226471602917, + 0.008775872178375721, + 0.007165084592998028, + 0.00596675043925643, + 0.004861255642026663, + 0.007675703149288893, + 0.006976661272346973, + 0.005785900168120861, + 0.00672989571467042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.01592535711824894 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206700026988983 + ], + "opposite_logprob_tokens": [ + -4.264206886291504 + ], + "logprob_reward": -0.014206700026988983, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206700026988983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12301504611968994, + 0.08984211832284927, + 0.11075474321842194, + 0.0859612375497818, + 0.1510918289422989, + 0.08466450870037079, + 0.16752688586711884, + 0.07318657636642456, + 0.08058612793684006, + 0.03337087482213974 + ], + "credit_map": [ + 0.010561268776655197, + 0.016456937417387962, + 0.016931209713220596, + 0.014829547144472599, + 0.018741637468338013, + 0.018808431923389435, + 0.011918445117771626, + 0.014767573215067387, + 0.006278111133724451, + 0.0061662448570132256, + 0.01951482892036438, + 0.005866471212357283, + 0.012881829403340816, + 0.02970057725906372, + 0.004675695206969976, + 0.004758364055305719, + 0.01342073641717434, + 0.011333158239722252, + 0.011544925160706043, + 0.01778683066368103, + 0.01811342127621174, + 0.011035537347197533, + 0.015599405393004417, + 0.011920743621885777, + 0.011574957519769669, + 0.010489868931472301, + 0.010304535739123821, + 0.013742927461862564, + 0.009897971525788307, + 0.009980479255318642, + 0.009671333245933056, + 0.010299170389771461, + 0.00837936531752348, + 0.05928646773099899, + 0.008936850354075432, + 0.0077461558394134045, + 0.007681011687964201, + 0.04168325662612915, + 0.00984323862940073, + 0.007535496726632118, + 0.008901616558432579, + 0.014003468677401543, + 0.008762618526816368, + 0.016161924228072166, + 0.006931471172720194, + 0.010768252424895763, + 0.010187794454395771, + 0.008947373367846012, + 0.018000636249780655, + 0.014681387692689896, + 0.01867481879889965, + 0.0226577315479517, + 0.013734772801399231, + 0.02284145914018154, + 0.02355342172086239, + 0.03338267654180527, + 0.014688978902995586, + 0.006868587341159582, + 0.006492354441434145, + 0.008593684062361717, + 0.006841131020337343, + 0.01692749559879303, + 0.005527097266167402, + 0.0072472491301596165, + 0.01454792357981205, + 0.0046198321506381035, + 0.005448357667773962, + 0.00836015772074461, + 0.005453506018966436, + 0.004726317711174488, + 0.02842950075864792, + 0.009000543504953384, + 0.00595414312556386, + 0.004845872055739164, + 0.0034170662984251976, + 0.002963758772239089, + 0.0038171515334397554, + 0.003855477785691619, + 0.0033396619837731123, + 0.005177746526896954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.014206700026988983 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978010892868042 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.00978010892868042, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978010892868042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08496685326099396, + 0.19590197503566742, + 0.075483538210392, + 0.07468896359205246, + 0.08838408440351486, + 0.06920775026082993, + 0.12957394123077393, + 0.16228722035884857, + 0.08073967695236206, + 0.03876601904630661 + ], + "credit_map": [ + 0.009547941386699677, + 0.009498773142695427, + 0.012152543291449547, + 0.009426262229681015, + 0.010494288988411427, + 0.009905495680868626, + 0.00972068589180708, + 0.014220860786736012, + 0.011255588382482529, + 0.007674795109778643, + 0.021366914734244347, + 0.008076656609773636, + 0.05753193795681, + 0.07346682995557785, + 0.008553403429687023, + 0.007975821383297443, + 0.008048338815569878, + 0.009395827539265156, + 0.007982932031154633, + 0.012438703328371048, + 0.010845952667295933, + 0.00754837691783905, + 0.01122258696705103, + 0.008000808767974377, + 0.009440256282687187, + 0.010576575063169003, + 0.0081969378516078, + 0.013348552398383617, + 0.00788724236190319, + 0.007451474200934172, + 0.008017403073608875, + 0.009770509786903858, + 0.005560005083680153, + 0.025356903672218323, + 0.006828688085079193, + 0.007690833881497383, + 0.008326394483447075, + 0.019515477120876312, + 0.008833736181259155, + 0.006272039841860533, + 0.008385535329580307, + 0.012402673251926899, + 0.008995166048407555, + 0.007644171826541424, + 0.004835181403905153, + 0.013636255636811256, + 0.006580245681107044, + 0.0067285108380019665, + 0.01340972539037466, + 0.011548880487680435, + 0.01708914153277874, + 0.022580724209547043, + 0.010356040671467781, + 0.020184120163321495, + 0.014935716986656189, + 0.019469568505883217, + 0.017010528594255447, + 0.012848556973040104, + 0.015711689367890358, + 0.03563693165779114, + 0.013475865125656128, + 0.03710462152957916, + 0.014622064307332039, + 0.015876932069659233, + 0.013614148832857609, + 0.004245638381689787, + 0.005040062591433525, + 0.008773699402809143, + 0.004843551199883223, + 0.004488611593842506, + 0.030100515112280846, + 0.009633442386984825, + 0.0049803233705461025, + 0.004551229998469353, + 0.005478687584400177, + 0.003769999137148261, + 0.00403914088383317, + 0.00555008091032505, + 0.005620422773063183, + 0.0047761304304003716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.00978010892868042 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902736008167267 + ], + "opposite_logprob_tokens": [ + -2.4640274047851562 + ], + "logprob_reward": -0.08902736008167267, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902736008167267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12136391550302505, + 0.02945990115404129, + 0.1107056736946106, + 0.03427926450967789, + 0.05027035251259804, + 0.061482030898332596, + 0.07988438010215759, + 0.19536219537258148, + 0.26168203353881836, + 0.05551031604409218 + ], + "credit_map": [ + 0.012071538716554642, + 0.010417871177196503, + 0.0233035609126091, + 0.011938312090933323, + 0.014819244854152203, + 0.010671718046069145, + 0.011908931657671928, + 0.026232711970806122, + 0.0030281278304755688, + 0.004210538696497679, + 0.004563763737678528, + 0.003926931880414486, + 0.004310152027755976, + 0.003913126420229673, + 0.00318548665381968, + 0.002321770181879401, + 0.011014898307621479, + 0.018016090616583824, + 0.011006927117705345, + 0.012548530474305153, + 0.013476167805492878, + 0.013433065265417099, + 0.01489513274282217, + 0.016314836218953133, + 0.006082112435251474, + 0.0042310128919780254, + 0.004337711725383997, + 0.004283804912120104, + 0.003877971088513732, + 0.003683241317048669, + 0.004240507259964943, + 0.0035428975243121386, + 0.006476909853518009, + 0.006506980396807194, + 0.005156675353646278, + 0.006875424180179834, + 0.004436902701854706, + 0.006239753682166338, + 0.007214427459985018, + 0.007363269571214914, + 0.005939546041190624, + 0.010023348033428192, + 0.008285094052553177, + 0.009781093336641788, + 0.005117757245898247, + 0.009702229872345924, + 0.006215927191078663, + 0.00641702301800251, + 0.008899534121155739, + 0.005685487762093544, + 0.011754190549254417, + 0.010322162881493568, + 0.00574768241494894, + 0.012569000944495201, + 0.011236971244215965, + 0.013669341802597046, + 0.0202633086591959, + 0.013374269008636475, + 0.017225997522473335, + 0.0468662865459919, + 0.014834534376859665, + 0.05102667585015297, + 0.01592695154249668, + 0.015844140201807022, + 0.05133982002735138, + 0.012752936221659184, + 0.01709333248436451, + 0.027438314631581306, + 0.017040593549609184, + 0.013527585193514824, + 0.09637526422739029, + 0.02611415646970272, + 0.00899413600564003, + 0.005958761554211378, + 0.006993152666836977, + 0.004685807973146439, + 0.007300136610865593, + 0.0052676391787827015, + 0.005749615840613842, + 0.010561056435108185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.08902736008167267 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890006422996521 + ], + "opposite_logprob_tokens": [ + -2.464000701904297 + ], + "logprob_reward": -0.0890006422996521, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890006422996521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17707106471061707, + 0.06378142535686493, + 0.12719929218292236, + 0.057042721658945084, + 0.06019347906112671, + 0.07330995798110962, + 0.1417720466852188, + 0.21356777846813202, + 0.06326239556074142, + 0.022799817845225334 + ], + "credit_map": [ + 0.01781705766916275, + 0.017074856907129288, + 0.022349460050463676, + 0.014412157237529755, + 0.029826460406184196, + 0.022348301485180855, + 0.025966636836528778, + 0.02727612666785717, + 0.006194401532411575, + 0.008681073784828186, + 0.011423097923398018, + 0.009082040749490261, + 0.008670401759445667, + 0.0077954912558197975, + 0.00717140594497323, + 0.004763513337820768, + 0.014627572149038315, + 0.022912485525012016, + 0.01729889027774334, + 0.01478562131524086, + 0.011545362882316113, + 0.012501707300543785, + 0.019376467913389206, + 0.014151188544929028, + 0.007944291457533836, + 0.004475435242056847, + 0.006690874695777893, + 0.00825448613613844, + 0.008338702842593193, + 0.006702021230012178, + 0.0074892896227538586, + 0.007147618103772402, + 0.007058826740831137, + 0.010810822248458862, + 0.00846261065453291, + 0.006115803029388189, + 0.0058032432571053505, + 0.007750712335109711, + 0.008116442710161209, + 0.006075016222894192, + 0.009681141003966331, + 0.010251154191792011, + 0.010561582632362843, + 0.00805328693240881, + 0.007935798726975918, + 0.009440425783395767, + 0.007429901976138353, + 0.009956667199730873, + 0.01355679240077734, + 0.012023579329252243, + 0.018175335600972176, + 0.022948261350393295, + 0.011573096737265587, + 0.017819304019212723, + 0.017325127497315407, + 0.028350558131933212, + 0.022046804428100586, + 0.015006575733423233, + 0.019563162699341774, + 0.05202450230717659, + 0.013871770352125168, + 0.05463748425245285, + 0.01762918196618557, + 0.01878829300403595, + 0.010808240622282028, + 0.002810874255374074, + 0.004914083983749151, + 0.006831982638686895, + 0.004850400611758232, + 0.0035569516476243734, + 0.021366894245147705, + 0.008122965693473816, + 0.003228734713047743, + 0.0025307880714535713, + 0.003335987450554967, + 0.0028090733103454113, + 0.00280552264302969, + 0.0025186906568706036, + 0.0021628604736179113, + 0.0034081602934747934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.0890006422996521 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This dramatic, high-speed slow-motion close-up captures the intricate courtship ritual of two Guangdong South Sea crabs in a captivating and lavish classic cinematic style. The camera, positioned at eye level, executes a nearly imperceptible insertion, heightening the emotional weight of the exchange. They perch on a dark, smooth rock, its surface gleaming with a deep, soft green moss. The pursuer, a larger male crab, possesses a vibrant, iridescent violet on his carapace and fiery red-orange legs. He advances with a steady and cautious demeanor, his eyestalks aligned with his opponent. On his right claw, he meticulously and delicately presents a shimmering, small pearl shell that gleams captivatingly in the light. The second crab, slightly smaller and more subdued in color, maintains a poised demeanor, its posture both shy and expectant. The entire exchange unfolds in exquisite slow motion, emphasizing the subtle, gentle extension of the claws and the almost imperceptible movement of the recipient's body. As the pursuer moves, shimmering droplets of water spray from his carapace in beautiful, suspended arcs. The scene is bathed in dramatic, high-contrast lighting, with a powerful key light piercing the darkness, casting brilliant highlights on their damp carapaces and the shells they provide, creating deep, soft shadows that add to the luxurious cinematic atmosphere. The rich colorful palette and intense attention to micro-movement create a remarkably tender and visually stunning narrative.", + "seed": 18333, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_3/index_5276_seed_18333_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04182713255286217, + "vlm_logprob": -0.04182713255286217, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.04182713255286217, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021943744271993637 + ], + "opposite_logprob_tokens": [ + -3.8969438076019287 + ], + "logprob_reward": -0.021943744271993637, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021943744271993637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10830533504486084, + 0.07928961515426636, + 0.04055408015847206, + 0.1330176591873169, + 0.16155345737934113, + 0.15537028014659882, + 0.10360019654035568, + 0.09720474481582642, + 0.0744427815079689, + 0.04666191712021828 + ], + "credit_map": [ + 0.010430986993014812, + 0.007563032675534487, + 0.008638384751975536, + 0.015943344682455063, + 0.009931684471666813, + 0.022023478522896767, + 0.016967616975307465, + 0.016806794330477715, + 0.005740916822105646, + 0.00662185437977314, + 0.008150462992489338, + 0.009562475606799126, + 0.005486487410962582, + 0.018750572577118874, + 0.012812301516532898, + 0.012164534069597721, + 0.005264753010123968, + 0.00468851113691926, + 0.008033053949475288, + 0.0036109157372266054, + 0.0041613406501710415, + 0.004878112114965916, + 0.0046509141102433205, + 0.005266476422548294, + 0.015669144690036774, + 0.012810771353542805, + 0.01220727153122425, + 0.014316778630018234, + 0.014534824527800083, + 0.024559425190091133, + 0.02035152167081833, + 0.018567904829978943, + 0.01153548527508974, + 0.013688383623957634, + 0.02282431162893772, + 0.020040085539221764, + 0.02284361980855465, + 0.02119472064077854, + 0.027679558843374252, + 0.02174728363752365, + 0.014549070969223976, + 0.014177108183503151, + 0.017059937119483948, + 0.017343390733003616, + 0.017816966399550438, + 0.025722268968820572, + 0.02629666216671467, + 0.02240486443042755, + 0.008532986976206303, + 0.010829788632690907, + 0.009349549189209938, + 0.010503116063773632, + 0.006449047476053238, + 0.020526966080069542, + 0.020328525453805923, + 0.01708020456135273, + 0.008761769160628319, + 0.009460361674427986, + 0.014361879788339138, + 0.008752565830945969, + 0.008023700676858425, + 0.019240260124206543, + 0.012992354109883308, + 0.015611845068633556, + 0.007707837503403425, + 0.006747385952621698, + 0.0047706179320812225, + 0.006384009495377541, + 0.007294744253158569, + 0.019319498911499977, + 0.009047625586390495, + 0.013171055354177952, + 0.005724454764276743, + 0.0038270680233836174, + 0.004533582367002964, + 0.00793521013110876, + 0.0038268135394901037, + 0.006479751318693161, + 0.00735714053735137, + 0.0069778901524841785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.021943744271993637 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018200648948550224 + ], + "opposite_logprob_tokens": [ + -4.018200874328613 + ], + "logprob_reward": -0.018200648948550224, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018200648948550224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21162307262420654, + 0.21538963913917542, + 0.03504505008459091, + 0.06106854975223541, + 0.10423632711172104, + 0.030899643898010254, + 0.13346631824970245, + 0.11032558232545853, + 0.04636349529027939, + 0.05158235877752304 + ], + "credit_map": [ + 0.02374255284667015, + 0.015284481458365917, + 0.019201429560780525, + 0.031251322478055954, + 0.017203936353325844, + 0.04424971342086792, + 0.029094615951180458, + 0.03159503638744354, + 0.021007822826504707, + 0.02589600346982479, + 0.01839294470846653, + 0.022418923676013947, + 0.014036922715604305, + 0.05101291090250015, + 0.032711729407310486, + 0.029912380501627922, + 0.004140350967645645, + 0.003624954726547003, + 0.0055992440320551395, + 0.003863021731376648, + 0.005011906381696463, + 0.003823118517175317, + 0.004652304109185934, + 0.004330149386078119, + 0.006409742403775454, + 0.007036910858005285, + 0.005170074291527271, + 0.0072151245549321175, + 0.006870173383504152, + 0.010376416146755219, + 0.009408485144376755, + 0.00858161598443985, + 0.00800853781402111, + 0.01000287476927042, + 0.013397715985774994, + 0.01209772564470768, + 0.01197876688092947, + 0.017388883978128433, + 0.016795510426163673, + 0.014566309750080109, + 0.0039093103259801865, + 0.0032722048927098513, + 0.0032812668941915035, + 0.003413635306060314, + 0.0038379449397325516, + 0.004570541903376579, + 0.004150864202529192, + 0.00446387380361557, + 0.011797953397035599, + 0.014991478063166142, + 0.01053632888942957, + 0.014712697826325893, + 0.011151975952088833, + 0.023586254566907883, + 0.02675366960465908, + 0.019935956224799156, + 0.015221419744193554, + 0.010827579535543919, + 0.010515759699046612, + 0.008778960444033146, + 0.012864685617387295, + 0.022027701139450073, + 0.011438851244747639, + 0.018650628626346588, + 0.008019546046853065, + 0.006194246932864189, + 0.005038776434957981, + 0.004445585887879133, + 0.0043525793589651585, + 0.007360138930380344, + 0.004320276901125908, + 0.0066323415376245975, + 0.0050197988748550415, + 0.003835693933069706, + 0.005771336145699024, + 0.009106517769396305, + 0.0038733906112611294, + 0.00853397510945797, + 0.007614327594637871, + 0.007827315479516983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.018200648948550224 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072972789406776 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.011072972789406776, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072972789406776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16741107404232025, + 0.16856706142425537, + 0.05652586743235588, + 0.052604492753744125, + 0.1424301266670227, + 0.050178028643131256, + 0.1436718851327896, + 0.10283225774765015, + 0.07175107300281525, + 0.044028084725141525 + ], + "credit_map": [ + 0.032146383076906204, + 0.02040882036089897, + 0.021167663857340813, + 0.017705077305436134, + 0.012736390344798565, + 0.026227138936519623, + 0.015375419519841671, + 0.021644193679094315, + 0.01547191385179758, + 0.015886090695858, + 0.02052110992372036, + 0.01633296348154545, + 0.01417903695255518, + 0.038256194442510605, + 0.02227586880326271, + 0.025643890723586082, + 0.007223394233733416, + 0.004574382212013006, + 0.005668080411851406, + 0.007111700251698494, + 0.00867181271314621, + 0.008169745095074177, + 0.008158568292856216, + 0.006948185618966818, + 0.006507663056254387, + 0.005884939339011908, + 0.004580487031489611, + 0.003982769325375557, + 0.0064425356686115265, + 0.009074595756828785, + 0.008016365580260754, + 0.008115140721201897, + 0.01053259801119566, + 0.014324258081614971, + 0.019325871020555496, + 0.016006650403141975, + 0.014130468480288982, + 0.027429373934864998, + 0.02011098340153694, + 0.0205699410289526, + 0.004550035111606121, + 0.003876991104334593, + 0.006023280322551727, + 0.0056555005721747875, + 0.006205457262694836, + 0.009781948290765285, + 0.007033308036625385, + 0.007051513530313969, + 0.013536724261939526, + 0.01472881343215704, + 0.007639686577022076, + 0.01000447291880846, + 0.016690734773874283, + 0.035724300891160965, + 0.021289365366101265, + 0.024057799950242043, + 0.01678302511572838, + 0.011793438345193863, + 0.011286777444183826, + 0.00840711034834385, + 0.012427142821252346, + 0.015039864927530289, + 0.01081091072410345, + 0.016283990815281868, + 0.007200072519481182, + 0.0076593756675720215, + 0.008280116133391857, + 0.008164976723492146, + 0.009945905767381191, + 0.010664545930922031, + 0.009513276629149914, + 0.010322805494070053, + 0.004211974795907736, + 0.0035870755091309547, + 0.003913209307938814, + 0.0064543443731963634, + 0.0033433123026043177, + 0.009142494760453701, + 0.006063313689082861, + 0.007312363479286432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.011072972789406776 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896961271762848 + ], + "opposite_logprob_tokens": [ + -2.578969717025757 + ], + "logprob_reward": -0.07896961271762848, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896961271762848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2667512893676758, + 0.10156478732824326, + 0.06246679276227951, + 0.06178182736039162, + 0.13883088529109955, + 0.03193783387541771, + 0.062006302177906036, + 0.1624680757522583, + 0.07215800136327744, + 0.04003426060080528 + ], + "credit_map": [ + 0.025950735434889793, + 0.018495136871933937, + 0.025403430685400963, + 0.02761022374033928, + 0.02861931174993515, + 0.05866594985127449, + 0.04422872141003609, + 0.03777778148651123, + 0.011159385554492474, + 0.008175782859325409, + 0.009248767979443073, + 0.013130339793860912, + 0.009173869155347347, + 0.01938791386783123, + 0.017095329239964485, + 0.014193395152688026, + 0.007258901838213205, + 0.005391046404838562, + 0.008400148712098598, + 0.005883827805519104, + 0.011062607169151306, + 0.008272051811218262, + 0.008462971076369286, + 0.007735239341855049, + 0.005137614905834198, + 0.005649013444781303, + 0.006027062423527241, + 0.007440544664859772, + 0.011854177340865135, + 0.009340153075754642, + 0.007808178197592497, + 0.008525081910192966, + 0.0102007444947958, + 0.012877940200269222, + 0.020023265853524208, + 0.015431549400091171, + 0.017341110855340958, + 0.017728140577673912, + 0.025759385898709297, + 0.019468748942017555, + 0.0025394822005182505, + 0.002832195023074746, + 0.0027096879202872515, + 0.0034180632792413235, + 0.004948374815285206, + 0.004838959313929081, + 0.0057990662753582, + 0.004852002020925283, + 0.006455508526414633, + 0.007760274223983288, + 0.005939101334661245, + 0.006861499510705471, + 0.006248168647289276, + 0.009470886550843716, + 0.010768319480121136, + 0.008502539247274399, + 0.022160563617944717, + 0.017505647614598274, + 0.02277960069477558, + 0.014979725703597069, + 0.016717609018087387, + 0.022002311423420906, + 0.020255539566278458, + 0.02606707252562046, + 0.010002498514950275, + 0.00838819332420826, + 0.007904107682406902, + 0.007028060499578714, + 0.005736063700169325, + 0.01410750113427639, + 0.008293299935758114, + 0.010698271915316582, + 0.003617432899773121, + 0.003282491583377123, + 0.0031414777040481567, + 0.004185437224805355, + 0.00395712535828352, + 0.009397721849381924, + 0.005997637752443552, + 0.006454937160015106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.07896961271762848 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two crabs visible in the scene, one larger and one smaller?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the larger crab's carapace vibrant and iridescent violet with fiery red-orange legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the smaller crab's carapace subdued in color compared to the larger one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the larger crab holding a shimmering small pearl shell on its right claw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a nearly imperceptible insertion movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07894868403673172 + ], + "opposite_logprob_tokens": [ + -2.578948736190796 + ], + "logprob_reward": -0.07894868403673172, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07894868403673172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15618896484375, + 0.1137346401810646, + 0.0799705758690834, + 0.23166632652282715, + 0.07420733571052551, + 0.059725914150476456, + 0.09542892873287201, + 0.06442870944738388, + 0.0790664404630661, + 0.0455821193754673 + ], + "credit_map": [ + 0.015715217217803, + 0.011062070727348328, + 0.013962860219180584, + 0.015383369289338589, + 0.015317298471927643, + 0.03944508731365204, + 0.021931709721684456, + 0.023371364921331406, + 0.010710777714848518, + 0.013138432055711746, + 0.011282108724117279, + 0.010716328397393227, + 0.009847667068243027, + 0.02400841750204563, + 0.019466232508420944, + 0.014564686454832554, + 0.014990563504397869, + 0.006833790335804224, + 0.008939567022025585, + 0.0073556783609092236, + 0.00585294421762228, + 0.012941030785441399, + 0.011039424687623978, + 0.012017576955258846, + 0.03256525471806526, + 0.023657169193029404, + 0.020479226484894753, + 0.0213472843170166, + 0.024748289957642555, + 0.04674047604203224, + 0.029118290171027184, + 0.033010341227054596, + 0.006427114829421043, + 0.0071682701818645, + 0.005220049060881138, + 0.00613858038559556, + 0.010156253352761269, + 0.01455364003777504, + 0.011912195011973381, + 0.012631237506866455, + 0.007939834147691727, + 0.0075098746456205845, + 0.006781878415495157, + 0.008091660216450691, + 0.006146182306110859, + 0.007400160655379295, + 0.007011519744992256, + 0.00884480495005846, + 0.011556430719792843, + 0.01109323836863041, + 0.009498199447989464, + 0.009637036360800266, + 0.009192910976707935, + 0.017889540642499924, + 0.012104500085115433, + 0.014457082375884056, + 0.005877783056348562, + 0.00631328159943223, + 0.00838429294526577, + 0.004691701848059893, + 0.006563299335539341, + 0.01686425507068634, + 0.0051008607260882854, + 0.010633242316544056, + 0.008007434196770191, + 0.0079630883410573, + 0.008139023557305336, + 0.008036778308451176, + 0.006259801797568798, + 0.018974611535668373, + 0.009452808648347855, + 0.01223289966583252, + 0.006723333615809679, + 0.005166280083358288, + 0.005515174008905888, + 0.005695058032870293, + 0.003817034885287285, + 0.008033759891986847, + 0.0037829040084034204, + 0.006848576478660107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.07894868403673172 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_4/index_tensor([153, 153])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_4/index_tensor([153, 153])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9378a176d4f28256dca23df8951fccb99a5b9bc5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_4/index_tensor([153, 153])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, upward tilt, starting from an extremely low angle, first captures a pair of hands in black, fingerless tactical gloves, carefully and with impassive precision, applying a high-tech, postpartum shapewear garment to the smooth, cold synthetic legs of a bionic figure. The flesh-colored shapewear's highly stretchable fibers are interwoven with dense biological circuitry that glows faintly blue. As the camera steadily ascends, we see an expressionless figure with short, clean-cut black hair, dressed in dark gray technical coveralls. Their silhouette is sharply defined by the flickering, cold fluorescent light overhead. The background is a chaotic, dystopian urban corner. Through rain-stained and grime-stained windows, a massive holographic billboard outside casts shifting magenta and cyan neon light pollution, intermingling with the cool white light of the interior. The person's movements were motionless as he mechanically pulled the corset over the doll's lifeless torso. The fluorescent light precisely outlined every intricate line on the garment. Finally, the camera passed over the doll's shoulders and slowly paused on its delicate face, focusing on a pair of glass-like, empty, lifeless eyes. No light was reflected in its deep pupils. The whole scene, under the cold cyberpunk aesthetic, was filled with a profound and unsettling silence.", + "seed": 21805, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_4/index_153_seed_21805_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02681114561855793, + "vlm_logprob": -0.02681114561855793, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.02681114561855793, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.056671638041734695 + ], + "opposite_logprob_tokens": [ + -2.931671619415283 + ], + "logprob_reward": -0.056671638041734695, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.056671638041734695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29106611013412476, + 0.25971174240112305, + 0.14323237538337708, + 0.04908139631152153, + 0.07685698568820953, + 0.030230535194277763, + 0.04370156675577164, + 0.053976137191057205, + 0.021218936890363693, + 0.030924230813980103 + ], + "credit_map": [ + 0.039165470749139786, + 0.035451821982860565, + 0.04208796098828316, + 0.03255655989050865, + 0.03290930762887001, + 0.031578097492456436, + 0.034596726298332214, + 0.04272018373012543, + 0.03265491873025894, + 0.03898638114333153, + 0.03941718116402626, + 0.03504291921854019, + 0.03419540822505951, + 0.021307915449142456, + 0.027200262993574142, + 0.03090674802660942, + 0.009434109553694725, + 0.015108689665794373, + 0.023856157436966896, + 0.04331505671143532, + 0.01013875287026167, + 0.0072465012781322, + 0.019572097808122635, + 0.014561010524630547, + 0.006665919907391071, + 0.007618994452059269, + 0.006103612948209047, + 0.006242198869585991, + 0.0047546750865876675, + 0.005299590993672609, + 0.005331923253834248, + 0.007064479403197765, + 0.008782538585364819, + 0.013169514946639538, + 0.006043475586920977, + 0.011682754382491112, + 0.0064093382097780704, + 0.01304375845938921, + 0.007808044087141752, + 0.009917556308209896, + 0.0032189819030463696, + 0.0030527785420417786, + 0.003753929166123271, + 0.004459083545953035, + 0.003831535344943404, + 0.0036280734930187464, + 0.005064961966127157, + 0.003221190767362714, + 0.0044966829009354115, + 0.006018324755132198, + 0.008879529312252998, + 0.003462672932073474, + 0.004492281470447779, + 0.004142125137150288, + 0.00531797157600522, + 0.006891979370266199, + 0.0061809909529984, + 0.008723875507712364, + 0.005969964433461428, + 0.005655346438288689, + 0.00679772300645709, + 0.005241600796580315, + 0.008532644249498844, + 0.006873988080769777, + 0.0028415019623935223, + 0.0016039490001276135, + 0.002444230020046234, + 0.0021486799232661724, + 0.0026529349852353334, + 0.0027756860945373774, + 0.002533590653911233, + 0.004218363668769598, + 0.004548016935586929, + 0.004856789018958807, + 0.003668548073619604, + 0.005116506479680538, + 0.002609006827697158, + 0.0037172180600464344, + 0.0028758579865098, + 0.003532287897542119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.056671638041734695 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204584993422031 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.014204584993422031, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204584993422031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10762764513492584, + 0.12319383025169373, + 0.2323080599308014, + 0.10599541664123535, + 0.08500366657972336, + 0.050402719527482986, + 0.050002794712781906, + 0.062466781586408615, + 0.11082324385643005, + 0.07217581570148468 + ], + "credit_map": [ + 0.01320385467261076, + 0.009458617307245731, + 0.016355199739336967, + 0.010786399245262146, + 0.008853612467646599, + 0.015810273587703705, + 0.016213778406381607, + 0.016945913434028625, + 0.012682205066084862, + 0.017715884372591972, + 0.01848890818655491, + 0.022188784554600716, + 0.014704675413668156, + 0.013755669817328453, + 0.01010123547166586, + 0.013556473888456821, + 0.013545944355428219, + 0.027053305879235268, + 0.0349486768245697, + 0.05818326026201248, + 0.022545987740159035, + 0.020688101649284363, + 0.02448573149740696, + 0.030857054516673088, + 0.010089898481965065, + 0.010846896097064018, + 0.02156398817896843, + 0.012059776112437248, + 0.008061159402132034, + 0.008530160412192345, + 0.014472338370978832, + 0.020371213555336, + 0.01032590214163065, + 0.013289328664541245, + 0.006651632953435183, + 0.012329253368079662, + 0.009451099671423435, + 0.012072314508259296, + 0.008904165588319302, + 0.011979972943663597, + 0.006497086025774479, + 0.006749278865754604, + 0.006037373561412096, + 0.005503478460013866, + 0.006956199184060097, + 0.006498722825199366, + 0.005880072247236967, + 0.0062805102206766605, + 0.005381593946367502, + 0.005344098433852196, + 0.009428002871572971, + 0.005734650883823633, + 0.004728849045932293, + 0.005404440686106682, + 0.006875333376228809, + 0.007105831988155842, + 0.006600542459636927, + 0.009773818776011467, + 0.008235656656324863, + 0.008595791645348072, + 0.008821710012853146, + 0.0043432279489934444, + 0.010507550090551376, + 0.005588490050286055, + 0.010841580107808113, + 0.0070912837982177734, + 0.0331844836473465, + 0.005539193749427795, + 0.004571033176034689, + 0.03534071147441864, + 0.005338246934115887, + 0.008916715160012245, + 0.009395703673362732, + 0.011924711987376213, + 0.007555366028100252, + 0.0164476465433836, + 0.0059949858114123344, + 0.007112693507224321, + 0.007079055067151785, + 0.006665653083473444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.014204584993422031 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193772338330746 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193772338330746, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193772338330746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.261267751455307, + 0.12303480505943298, + 0.07286222279071808, + 0.17397470772266388, + 0.03804120421409607, + 0.03689171373844147, + 0.09991706907749176, + 0.07863712310791016, + 0.06038464978337288, + 0.05498867109417915 + ], + "credit_map": [ + 0.03596128895878792, + 0.02897733449935913, + 0.03981693834066391, + 0.02861219272017479, + 0.028403908014297485, + 0.03150518983602524, + 0.028328847140073776, + 0.039662063121795654, + 0.011934981681406498, + 0.018686046823859215, + 0.01826225221157074, + 0.015236721374094486, + 0.012642727233469486, + 0.015536927618086338, + 0.01310113538056612, + 0.017634032294154167, + 0.005380736198276281, + 0.005092044826596975, + 0.012167592532932758, + 0.008533136919140816, + 0.004518589936196804, + 0.010717109777033329, + 0.008118817582726479, + 0.018334202468395233, + 0.012467868626117706, + 0.01024866197258234, + 0.04587878659367561, + 0.012300951406359673, + 0.013104950077831745, + 0.01568448357284069, + 0.013811345212161541, + 0.05047769099473953, + 0.0047343033365905285, + 0.006547033786773682, + 0.004564089234918356, + 0.004885650239884853, + 0.0033865831792354584, + 0.005256138276308775, + 0.003924023825675249, + 0.004743384197354317, + 0.0037710044998675585, + 0.0034558759070932865, + 0.005253355018794537, + 0.005019408650696278, + 0.004590597003698349, + 0.005496372934430838, + 0.005114148370921612, + 0.004190956242382526, + 0.01150957029312849, + 0.011717258021235466, + 0.008924990892410278, + 0.010916243307292461, + 0.012036586180329323, + 0.010290812700986862, + 0.017975343391299248, + 0.016546273604035378, + 0.006979492027312517, + 0.011426663026213646, + 0.009848256595432758, + 0.011155909858644009, + 0.009855566546320915, + 0.007873456925153732, + 0.013211561366915703, + 0.008286223746836185, + 0.005000490695238113, + 0.004824705887585878, + 0.008557434193789959, + 0.0064015681855380535, + 0.006086959503591061, + 0.01390809565782547, + 0.007442761678248644, + 0.008162636309862137, + 0.012816643342375755, + 0.00789836049079895, + 0.006303928792476654, + 0.007480335421860218, + 0.004323841072618961, + 0.0060706669464707375, + 0.0037057034205645323, + 0.006389200687408447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.014193772338330746 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005942771211266518 + ], + "opposite_logprob_tokens": [ + -5.1309428215026855 + ], + "logprob_reward": -0.005942771211266518, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942771211266518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11583703756332397, + 0.02755512110888958, + 0.08560033142566681, + 0.10166701674461365, + 0.03903907164931297, + 0.04083731397986412, + 0.1391625702381134, + 0.2589363753795624, + 0.12539440393447876, + 0.065970778465271 + ], + "credit_map": [ + 0.015841549262404442, + 0.0104632917791605, + 0.01744631864130497, + 0.012843162752687931, + 0.010788140818476677, + 0.01664108969271183, + 0.012458264827728271, + 0.01935522072017193, + 0.003636094508692622, + 0.0034553667064756155, + 0.004366765730082989, + 0.0036833423655480146, + 0.002458894858136773, + 0.0028709659818559885, + 0.002674242714419961, + 0.004409446846693754, + 0.004608889110386372, + 0.007382574025541544, + 0.014388917945325375, + 0.019258001819252968, + 0.00668656500056386, + 0.008732161484658718, + 0.008308565244078636, + 0.016234664246439934, + 0.006625714711844921, + 0.004581077955663204, + 0.029547082260251045, + 0.007200197316706181, + 0.0067087141796946526, + 0.00890872348099947, + 0.006677331402897835, + 0.031418174505233765, + 0.004321155603975058, + 0.006564544513821602, + 0.003917626570910215, + 0.007741584442555904, + 0.00303613836877048, + 0.004593081772327423, + 0.003976981621235609, + 0.004887961316853762, + 0.004277902655303478, + 0.004131873603910208, + 0.005729864351451397, + 0.004445110447704792, + 0.005604883655905724, + 0.005122033879160881, + 0.00718874204903841, + 0.004336900543421507, + 0.010078203864395618, + 0.01221469510346651, + 0.038524750620126724, + 0.009270808659493923, + 0.011132017709314823, + 0.03371231257915497, + 0.012872545048594475, + 0.011357237584888935, + 0.03214048221707344, + 0.03718752786517143, + 0.03590397536754608, + 0.026322459802031517, + 0.03213611990213394, + 0.03423122689127922, + 0.02875743992626667, + 0.032257143408060074, + 0.014301533810794353, + 0.010794407688081264, + 0.03451871499419212, + 0.008257958106696606, + 0.008783646859228611, + 0.02683129347860813, + 0.009169238619506359, + 0.01273761224001646, + 0.012941950000822544, + 0.008703603409230709, + 0.00585439708083868, + 0.005448353011161089, + 0.006912848446518183, + 0.007937773130834103, + 0.0074340770952403545, + 0.010737773030996323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.005942771211266518 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304296150803566 + ], + "opposite_logprob_tokens": [ + -3.1680428981781006 + ], + "logprob_reward": -0.04304296150803566, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304296150803566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09764784574508667, + 0.07656031101942062, + 0.03920301795005798, + 0.12191390991210938, + 0.046253886073827744, + 0.04954497143626213, + 0.1784508228302002, + 0.21097978949546814, + 0.11447262018918991, + 0.06497279554605484 + ], + "credit_map": [ + 0.012696044519543648, + 0.013014557771384716, + 0.012912747450172901, + 0.010431825183331966, + 0.012380891479551792, + 0.011657576076686382, + 0.014038635417819023, + 0.010515573434531689, + 0.006825426127761602, + 0.012107368558645248, + 0.01105385459959507, + 0.009966829791665077, + 0.010442250408232212, + 0.0073532224632799625, + 0.00988043937832117, + 0.008930929005146027, + 0.003289241809397936, + 0.003566657891497016, + 0.006398901809006929, + 0.006844021379947662, + 0.002659312216565013, + 0.004354814998805523, + 0.004371581133455038, + 0.007718491833657026, + 0.008763340301811695, + 0.006481615826487541, + 0.03450929373502731, + 0.008582505397498608, + 0.006127739790827036, + 0.00941516738384962, + 0.008704420179128647, + 0.039329834282398224, + 0.00613299710676074, + 0.006798406597226858, + 0.004003316164016724, + 0.007561491336673498, + 0.0038409007247537374, + 0.006170820444822311, + 0.005233118310570717, + 0.006512841209769249, + 0.00553320674225688, + 0.004987222608178854, + 0.006207348313182592, + 0.006194079760462046, + 0.006720286328345537, + 0.006945550907403231, + 0.007914222776889801, + 0.005043057259172201, + 0.01720842719078064, + 0.026994386687874794, + 0.025445956736803055, + 0.013552508316934109, + 0.024204473942518234, + 0.026876848191022873, + 0.017879607155919075, + 0.026288634166121483, + 0.021951764822006226, + 0.025612737983465195, + 0.03032582439482212, + 0.02629145048558712, + 0.030863748863339424, + 0.02702251262962818, + 0.022160429507493973, + 0.0267513245344162, + 0.010263836942613125, + 0.007843216881155968, + 0.03372159227728844, + 0.007173618301749229, + 0.0062366207130253315, + 0.03528955951333046, + 0.005540634039789438, + 0.008403560146689415, + 0.008736536838114262, + 0.014719870872795582, + 0.00740819750353694, + 0.010576191358268261, + 0.00558960996568203, + 0.0065593714825809, + 0.005295151378959417, + 0.0060878717340528965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.04304296150803566 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward tilt, starting from an extremely low angle, first captures a pair of hands in black, fingerless tactical gloves, carefully and with impassive precision, applying a high-tech, postpartum shapewear garment to the smooth, cold synthetic legs of a bionic figure. The flesh-colored shapewear's highly stretchable fibers are interwoven with dense biological circuitry that glows faintly blue. As the camera steadily ascends, we see an expressionless figure with short, clean-cut black hair, dressed in dark gray technical coveralls. Their silhouette is sharply defined by the flickering, cold fluorescent light overhead. The background is a chaotic, dystopian urban corner. Through rain-stained and grime-stained windows, a massive holographic billboard outside casts shifting magenta and cyan neon light pollution, intermingling with the cool white light of the interior. The person's movements were motionless as he mechanically pulled the corset over the doll's lifeless torso. The fluorescent light precisely outlined every intricate line on the garment. Finally, the camera passed over the doll's shoulders and slowly paused on its delicate face, focusing on a pair of glass-like, empty, lifeless eyes. No light was reflected in its deep pupils. The whole scene, under the cold cyberpunk aesthetic, was filled with a profound and unsettling silence.", + "seed": 21806, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_4/index_153_seed_21806_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.047593132220208646, + "vlm_logprob": -0.047593132220208646, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.047593132220208646, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12925191223621368 + ], + "opposite_logprob_tokens": [ + -2.1292519569396973 + ], + "logprob_reward": -0.12925191223621368, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12925191223621368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24635854363441467, + 0.08113906532526016, + 0.158470019698143, + 0.16410386562347412, + 0.05088862404227257, + 0.11255380511283875, + 0.04579245299100876, + 0.036364469677209854, + 0.07019314914941788, + 0.034135930240154266 + ], + "credit_map": [ + 0.03231343999505043, + 0.029842134565114975, + 0.03128203749656677, + 0.04203521087765694, + 0.022396055981516838, + 0.02087763138115406, + 0.0304073765873909, + 0.03720469027757645, + 0.008047453127801418, + 0.012679992243647575, + 0.0068389917723834515, + 0.014972887933254242, + 0.012323835864663124, + 0.010677866637706757, + 0.00834390614181757, + 0.007254139054566622, + 0.01765153557062149, + 0.02107279561460018, + 0.0170148778706789, + 0.02610071375966072, + 0.020317547023296356, + 0.017251381650567055, + 0.022102413699030876, + 0.016958774998784065, + 0.024559926241636276, + 0.013765347190201283, + 0.01410279143601656, + 0.012796806171536446, + 0.03411243110895157, + 0.031567495316267014, + 0.013042738661170006, + 0.020156340673565865, + 0.005800602957606316, + 0.004126077983528376, + 0.005444106180220842, + 0.009537320584058762, + 0.008230012841522694, + 0.007101516239345074, + 0.004059825092554092, + 0.006589165888726711, + 0.009603342972695827, + 0.02897358499467373, + 0.005511568393558264, + 0.006405172869563103, + 0.03402352333068848, + 0.008446463383734226, + 0.012197474017739296, + 0.007392682135105133, + 0.005120068322867155, + 0.005926627200096846, + 0.004021513741463423, + 0.0037678019143640995, + 0.00839126855134964, + 0.004609859082847834, + 0.008687266148626804, + 0.00526804756373167, + 0.0028730882331728935, + 0.005559729877859354, + 0.0034120564814656973, + 0.00541894743219018, + 0.006363305263221264, + 0.004723465070128441, + 0.004178476054221392, + 0.003835404757410288, + 0.011952984146773815, + 0.009890181943774223, + 0.007530491799116135, + 0.006032066885381937, + 0.007124228402972221, + 0.013148236088454723, + 0.006719955708831549, + 0.007795010693371296, + 0.0028066057711839676, + 0.0029560185503214598, + 0.0037244176492094994, + 0.0042730472050607204, + 0.006773585919290781, + 0.003554200055077672, + 0.005137908272445202, + 0.004910149611532688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.12925191223621368 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608244515955448 + ], + "opposite_logprob_tokens": [ + -4.141082286834717 + ], + "logprob_reward": -0.01608244515955448, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608244515955448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19391991198062897, + 0.061543866991996765, + 0.10795504599809647, + 0.16720539331436157, + 0.06188404560089111, + 0.09339475631713867, + 0.11208963394165039, + 0.08134578168392181, + 0.061052512377500534, + 0.059609025716781616 + ], + "credit_map": [ + 0.0312199667096138, + 0.026659352704882622, + 0.025206023827195168, + 0.03303307294845581, + 0.01830376125872135, + 0.01930549554526806, + 0.01931658387184143, + 0.020875658839941025, + 0.006167486775666475, + 0.007804853841662407, + 0.004941083490848541, + 0.009444054216146469, + 0.011469646356999874, + 0.008612187579274178, + 0.006858189124614, + 0.006246367935091257, + 0.016363922506570816, + 0.015463861636817455, + 0.009949675761163235, + 0.009515395388007164, + 0.01280308235436678, + 0.012434476986527443, + 0.01286877691745758, + 0.01855584606528282, + 0.02148946188390255, + 0.012749436311423779, + 0.013099337927997112, + 0.01593046821653843, + 0.035494398325681686, + 0.027998456731438637, + 0.02150086872279644, + 0.018942974507808685, + 0.007330053951591253, + 0.0052869427017867565, + 0.006023040506988764, + 0.009287169203162193, + 0.011285644955933094, + 0.008884814567863941, + 0.006230369675904512, + 0.007556012831628323, + 0.006331713404506445, + 0.02606234885752201, + 0.004027235321700573, + 0.006979907397180796, + 0.028092453256249428, + 0.006332078017294407, + 0.010580375790596008, + 0.004988641012459993, + 0.008319121785461903, + 0.014426019974052906, + 0.009163346141576767, + 0.014641428366303444, + 0.01737300306558609, + 0.01049797236919403, + 0.022670796141028404, + 0.014997947961091995, + 0.007750361692160368, + 0.007906712591648102, + 0.006806183140724897, + 0.010637008585035801, + 0.018020261079072952, + 0.009231602773070335, + 0.01076427847146988, + 0.010229376144707203, + 0.008193956688046455, + 0.008080183528363705, + 0.006921074353158474, + 0.0057746958918869495, + 0.008052732795476913, + 0.012623400427401066, + 0.006186134181916714, + 0.005220336373895407, + 0.003683812916278839, + 0.005823376588523388, + 0.0033868884202092886, + 0.006771455984562635, + 0.015045559033751488, + 0.0073861610144376755, + 0.012006795033812523, + 0.005504972767084837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.01608244515955448 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799826443195343 + ], + "opposite_logprob_tokens": [ + -3.5297999382019043 + ], + "logprob_reward": -0.029799826443195343, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799826443195343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15131057798862457, + 0.06985556334257126, + 0.11640407890081406, + 0.05546156316995621, + 0.06355113536119461, + 0.13270829617977142, + 0.14563177525997162, + 0.1220134049654007, + 0.08644843101501465, + 0.0566151924431324 + ], + "credit_map": [ + 0.021261435002088547, + 0.019783828407526016, + 0.0175000112503767, + 0.02317299135029316, + 0.016171663999557495, + 0.017411377280950546, + 0.017494335770606995, + 0.01851492002606392, + 0.012770500034093857, + 0.005699514877051115, + 0.009546242654323578, + 0.005947576370090246, + 0.005990827921777964, + 0.012334875762462616, + 0.005698090419173241, + 0.011867934837937355, + 0.018534839153289795, + 0.014070102944970131, + 0.014023075811564922, + 0.01379372738301754, + 0.017798565328121185, + 0.012149989604949951, + 0.013616170734167099, + 0.012417605146765709, + 0.007695203647017479, + 0.004391983151435852, + 0.005095460917800665, + 0.006355935707688332, + 0.0078955153003335, + 0.01074866484850645, + 0.0057120793499052525, + 0.007566716056317091, + 0.009400603361427784, + 0.005810548551380634, + 0.006575546693056822, + 0.008446593768894672, + 0.011091012507677078, + 0.00919087789952755, + 0.0059170955792069435, + 0.007118850946426392, + 0.00923357717692852, + 0.022126879543066025, + 0.007058077957481146, + 0.008155977353453636, + 0.038540493696928024, + 0.010260134004056454, + 0.027249138802289963, + 0.010084022767841816, + 0.026212118566036224, + 0.015308048576116562, + 0.017088184133172035, + 0.008840475231409073, + 0.020364046096801758, + 0.018404431641101837, + 0.017521515488624573, + 0.02189294621348381, + 0.01271857600659132, + 0.013168168254196644, + 0.010445124469697475, + 0.01577458716928959, + 0.01885279081761837, + 0.013600495643913746, + 0.020611993968486786, + 0.016841668635606766, + 0.013650808483362198, + 0.011689099483191967, + 0.008892408572137356, + 0.007110615726560354, + 0.008454768918454647, + 0.017962943762540817, + 0.008996469900012016, + 0.009691311046481133, + 0.00350908818654716, + 0.0051069967448711395, + 0.005377219524234533, + 0.0069242301397025585, + 0.012593799270689487, + 0.005690515972673893, + 0.01078157126903534, + 0.006631768774241209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.029799826443195343 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189188368618488 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.014189188368618488, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189188368618488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09722058475017548, + 0.07160947471857071, + 0.056685492396354675, + 0.0500471256673336, + 0.27356788516044617, + 0.06602282077074051, + 0.090128093957901, + 0.08657322824001312, + 0.1668875515460968, + 0.04125773161649704 + ], + "credit_map": [ + 0.015993785113096237, + 0.014641783200204372, + 0.011720356531441212, + 0.012085034511983395, + 0.009024236351251602, + 0.009664212353527546, + 0.010321845300495625, + 0.013769329525530338, + 0.011588867753744125, + 0.006505320314317942, + 0.012941763736307621, + 0.005549157503992319, + 0.007937625981867313, + 0.01348017156124115, + 0.0036413019988685846, + 0.009965262375772, + 0.006024444010108709, + 0.006633420009166002, + 0.005183692090213299, + 0.008372935466468334, + 0.0060895150527358055, + 0.007286465261131525, + 0.0070571633987128735, + 0.010037856176495552, + 0.009589110501110554, + 0.005147736053913832, + 0.00612398749217391, + 0.00574743514880538, + 0.004939464386552572, + 0.008573349565267563, + 0.0036120456643402576, + 0.006313991267234087, + 0.04278937354683876, + 0.01820172369480133, + 0.02718007005751133, + 0.030625365674495697, + 0.05976549908518791, + 0.035768721252679825, + 0.026848996058106422, + 0.03238813951611519, + 0.0063506243750452995, + 0.016017530113458633, + 0.003500313963741064, + 0.003951978869736195, + 0.018402328714728355, + 0.006067165173590183, + 0.00669511966407299, + 0.005037758499383926, + 0.01240962091833353, + 0.00915692187845707, + 0.007740125060081482, + 0.010189981199800968, + 0.012444437481462955, + 0.008254350163042545, + 0.01663878746330738, + 0.013293868862092495, + 0.008225668221712112, + 0.013713102787733078, + 0.01123634073883295, + 0.010558456182479858, + 0.01455626916140318, + 0.010295546613633633, + 0.009724360890686512, + 0.008263474330306053, + 0.02078511193394661, + 0.022387003526091576, + 0.021334117278456688, + 0.015994561836123466, + 0.02402268908917904, + 0.01722596026957035, + 0.02378733828663826, + 0.021350756287574768, + 0.002959175268188119, + 0.00348376645706594, + 0.0035525367129594088, + 0.004324978217482567, + 0.00910906121134758, + 0.005241591017693281, + 0.0072343964129686356, + 0.00535222515463829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.014189188368618488 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864228889346123 + ], + "opposite_logprob_tokens": [ + -3.04864239692688 + ], + "logprob_reward": -0.04864228889346123, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864228889346123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12386013567447662, + 0.05933072790503502, + 0.08421529084444046, + 0.056159473955631256, + 0.03403320908546448, + 0.038573257625103, + 0.07797519117593765, + 0.21443742513656616, + 0.09587500244379044, + 0.21554028987884521 + ], + "credit_map": [ + 0.016518522053956985, + 0.021495936438441277, + 0.014696955680847168, + 0.018458561971783638, + 0.011158611625432968, + 0.012058553285896778, + 0.011071272194385529, + 0.018401721492409706, + 0.007051662541925907, + 0.007181992754340172, + 0.004921602550894022, + 0.006146037485450506, + 0.009644616395235062, + 0.009231474250555038, + 0.005053299013525248, + 0.010100042447447777, + 0.01400776207447052, + 0.007079357746988535, + 0.01574375294148922, + 0.01077185571193695, + 0.004242247901856899, + 0.010112137533724308, + 0.009059146046638489, + 0.013199029490351677, + 0.0037957304157316685, + 0.005714783910661936, + 0.005571162793785334, + 0.005068512167781591, + 0.010840159840881824, + 0.0118827223777771, + 0.004791493993252516, + 0.008494909852743149, + 0.006982042919844389, + 0.0031254759524017572, + 0.003981136251240969, + 0.003599500749260187, + 0.005336220376193523, + 0.0043877833522856236, + 0.003121114568784833, + 0.0034999351482838392, + 0.004736751783639193, + 0.0036191458348184824, + 0.0033844232093542814, + 0.004601202439516783, + 0.006578109692782164, + 0.005461727734655142, + 0.006262101698666811, + 0.003929795231670141, + 0.006696498487144709, + 0.010244582779705524, + 0.00851406529545784, + 0.008753777481615543, + 0.013553895056247711, + 0.0077886031940579414, + 0.013424362055957317, + 0.00899940263479948, + 0.019469130784273148, + 0.023682069033384323, + 0.01784382574260235, + 0.029737940058112144, + 0.047057896852493286, + 0.022799629718065262, + 0.028019500896334648, + 0.02582741342484951, + 0.013790210708975792, + 0.01230616681277752, + 0.009890316054224968, + 0.008763407357037067, + 0.011428136378526688, + 0.01714160107076168, + 0.011977240443229675, + 0.01057792454957962, + 0.013505703769624233, + 0.01942560076713562, + 0.012857688590884209, + 0.02447286620736122, + 0.05237467214465141, + 0.02540428750216961, + 0.04294188320636749, + 0.024557586759328842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.04864228889346123 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward tilt, starting from an extremely low angle, first captures a pair of hands in black, fingerless tactical gloves, carefully and with impassive precision, applying a high-tech, postpartum shapewear garment to the smooth, cold synthetic legs of a bionic figure. The flesh-colored shapewear's highly stretchable fibers are interwoven with dense biological circuitry that glows faintly blue. As the camera steadily ascends, we see an expressionless figure with short, clean-cut black hair, dressed in dark gray technical coveralls. Their silhouette is sharply defined by the flickering, cold fluorescent light overhead. The background is a chaotic, dystopian urban corner. Through rain-stained and grime-stained windows, a massive holographic billboard outside casts shifting magenta and cyan neon light pollution, intermingling with the cool white light of the interior. The person's movements were motionless as he mechanically pulled the corset over the doll's lifeless torso. The fluorescent light precisely outlined every intricate line on the garment. Finally, the camera passed over the doll's shoulders and slowly paused on its delicate face, focusing on a pair of glass-like, empty, lifeless eyes. No light was reflected in its deep pupils. The whole scene, under the cold cyberpunk aesthetic, was filled with a profound and unsettling silence.", + "seed": 21807, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_4/index_153_seed_21807_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04292615130543709, + "vlm_logprob": -0.04292615130543709, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.04292615130543709, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0909801572561264 + ], + "opposite_logprob_tokens": [ + -2.465980052947998 + ], + "logprob_reward": -0.0909801572561264, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0909801572561264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06896360218524933, + 0.13349361717700958, + 0.06074715778231621, + 0.2819814682006836, + 0.07467387616634369, + 0.07542569935321808, + 0.09333409368991852, + 0.12913230061531067, + 0.042101409286260605, + 0.04014678671956062 + ], + "credit_map": [ + 0.010259284637868404, + 0.008189181797206402, + 0.00753765506669879, + 0.009092564694583416, + 0.004531898070126772, + 0.01115286722779274, + 0.003743454348295927, + 0.014456696808338165, + 0.016522428020834923, + 0.01714019849896431, + 0.013902595266699791, + 0.015262155793607235, + 0.01914331130683422, + 0.02076176181435585, + 0.013056560419499874, + 0.01770460046827793, + 0.009107646532356739, + 0.0036949897184967995, + 0.010666967369616032, + 0.006812740117311478, + 0.00435905484482646, + 0.00940656941384077, + 0.0072711496613919735, + 0.009428040124475956, + 0.03919731453061104, + 0.03533908724784851, + 0.035149987787008286, + 0.029655376449227333, + 0.027374709025025368, + 0.0332605205476284, + 0.034795768558979034, + 0.047208696603775024, + 0.00697410898283124, + 0.00828712061047554, + 0.010505937971174717, + 0.006829835008829832, + 0.010963534004986286, + 0.008576962165534496, + 0.010204662568867207, + 0.01233171671628952, + 0.0035698539577424526, + 0.004724598024040461, + 0.020131157711148262, + 0.007739552762359381, + 0.022718336433172226, + 0.003835516283288598, + 0.008967509493231773, + 0.0037391777150332928, + 0.008103600703179836, + 0.007606752682477236, + 0.009365404024720192, + 0.010975637473165989, + 0.023890094831585884, + 0.011778505519032478, + 0.014733166433870792, + 0.006880925036966801, + 0.01018676906824112, + 0.011949977837502956, + 0.010661136358976364, + 0.02118832990527153, + 0.018055075779557228, + 0.020675839856266975, + 0.019390426576137543, + 0.017024748027324677, + 0.003370597492903471, + 0.004727772902697325, + 0.0030776122584939003, + 0.0035868771374225616, + 0.006113759707659483, + 0.004890743177384138, + 0.010884552262723446, + 0.005449493881314993, + 0.0025133525487035513, + 0.0037323208525776863, + 0.008800641633570194, + 0.0028689794708043337, + 0.00561489025130868, + 0.006590242497622967, + 0.004905320703983307, + 0.005121039692312479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.0909801572561264 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020587561652064323 + ], + "opposite_logprob_tokens": [ + -3.89558744430542 + ], + "logprob_reward": -0.020587561652064323, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020587561652064323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0858783945441246, + 0.21120670437812805, + 0.20104990899562836, + 0.2430345118045807, + 0.06814417988061905, + 0.05802161991596222, + 0.019588662311434746, + 0.050400253385305405, + 0.030115047469735146, + 0.032560721039772034 + ], + "credit_map": [ + 0.012189452536404133, + 0.012383625842630863, + 0.012400902807712555, + 0.009887266904115677, + 0.005840139463543892, + 0.013880240730941296, + 0.0059345792979002, + 0.01336218137294054, + 0.027344824746251106, + 0.02662575989961624, + 0.017788760364055634, + 0.01720803789794445, + 0.03367311507463455, + 0.03939926251769066, + 0.0215980876237154, + 0.027568861842155457, + 0.02898602932691574, + 0.01676531322300434, + 0.03423647582530975, + 0.027689404785633087, + 0.019121209159493446, + 0.014308709651231766, + 0.02992534264922142, + 0.0300174318253994, + 0.04031337797641754, + 0.029210126027464867, + 0.03061700239777565, + 0.02523333765566349, + 0.0203924048691988, + 0.027253568172454834, + 0.026938138529658318, + 0.04307655990123749, + 0.00622078450396657, + 0.006556080654263496, + 0.009124665521085262, + 0.007322176359593868, + 0.008768505416810513, + 0.006531724706292152, + 0.012178286910057068, + 0.011441956274211407, + 0.004441325552761555, + 0.007192681543529034, + 0.012605594471096992, + 0.005291488021612167, + 0.013724301941692829, + 0.0036000681575387716, + 0.008550804108381271, + 0.0026153556536883116, + 0.0014162042643874884, + 0.0024230279959738255, + 0.0020979600958526134, + 0.002673299750313163, + 0.0027516651898622513, + 0.003422945737838745, + 0.002636489225551486, + 0.0021670714486390352, + 0.004800864960998297, + 0.004648433532565832, + 0.004780553746968508, + 0.01015933696180582, + 0.006535567343235016, + 0.006219404749572277, + 0.007958604954183102, + 0.005297490861266851, + 0.002645182888954878, + 0.00451612425968051, + 0.0033991115633398294, + 0.0034694417845457792, + 0.004699137527495623, + 0.00392158143222332, + 0.004376132041215897, + 0.003088334808126092, + 0.0028279079124331474, + 0.0025525004602968693, + 0.009224433451890945, + 0.0021064963657408953, + 0.002823926042765379, + 0.009477141313254833, + 0.0018322312971577048, + 0.0017160854768007994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.020587561652064323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287607356905937 + ], + "opposite_logprob_tokens": [ + -3.773287534713745 + ], + "logprob_reward": -0.023287607356905937, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287607356905937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14170722663402557, + 0.08788581192493439, + 0.1265425831079483, + 0.14313672482967377, + 0.06301321089267731, + 0.12176840752363205, + 0.15828359127044678, + 0.08076529949903488, + 0.03896453604102135, + 0.037932608276605606 + ], + "credit_map": [ + 0.02950052171945572, + 0.016609786078333855, + 0.015770435333251953, + 0.01390588004142046, + 0.010005155578255653, + 0.020763052627444267, + 0.009128174744546413, + 0.026024233549833298, + 0.010490278713405132, + 0.010189750231802464, + 0.012685451656579971, + 0.008230992592871189, + 0.011438981629908085, + 0.016660574823617935, + 0.007161987014114857, + 0.011027796193957329, + 0.017006080597639084, + 0.009571443311870098, + 0.022069908678531647, + 0.016159215942025185, + 0.01007862575352192, + 0.01697767712175846, + 0.016009945422410965, + 0.01866968907415867, + 0.02336357720196247, + 0.01706906035542488, + 0.01943770796060562, + 0.014814266934990883, + 0.012646033428609371, + 0.015763042494654655, + 0.01784149929881096, + 0.0222015380859375, + 0.007890674285590649, + 0.006524162366986275, + 0.00927912350744009, + 0.006515022832900286, + 0.008903855457901955, + 0.005261346232146025, + 0.008344591595232487, + 0.010294434614479542, + 0.00858964491635561, + 0.011202078312635422, + 0.018851997330784798, + 0.018167221918702126, + 0.024115735664963722, + 0.007774415425956249, + 0.02535070851445198, + 0.007716599386185408, + 0.014586938545107841, + 0.013001689687371254, + 0.022281726822257042, + 0.015300325118005276, + 0.03701690584421158, + 0.020827071741223335, + 0.021527256816625595, + 0.013741685077548027, + 0.009357944130897522, + 0.007903547957539558, + 0.007508075330406427, + 0.010747184045612812, + 0.01323465071618557, + 0.005918099079281092, + 0.01741660013794899, + 0.008679195307195187, + 0.0032108307350426912, + 0.003807319328188896, + 0.0035240889992564917, + 0.005206402391195297, + 0.004562558140605688, + 0.0042990692891180515, + 0.008222078904509544, + 0.0061321863904595375, + 0.004366499371826649, + 0.0056521473452448845, + 0.004685600288212299, + 0.003503135871142149, + 0.005499043967574835, + 0.005725842900574207, + 0.0053463587537407875, + 0.0031539814081043005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.023287607356905937 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775623679161072 + ], + "opposite_logprob_tokens": [ + -4.634775638580322 + ], + "logprob_reward": -0.009775623679161072, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775623679161072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12288648635149002, + 0.0877092257142067, + 0.041567374020814896, + 0.11610197275876999, + 0.07863172888755798, + 0.09046602994203568, + 0.20783574879169464, + 0.08354205638170242, + 0.054680753499269485, + 0.11657870560884476 + ], + "credit_map": [ + 0.014496159739792347, + 0.015133620239794254, + 0.008821103721857071, + 0.017218980938196182, + 0.011655387468636036, + 0.02013048157095909, + 0.011230471543967724, + 0.02420027181506157, + 0.009924977086484432, + 0.010589361190795898, + 0.008119098842144012, + 0.007403923664242029, + 0.011788299307227135, + 0.018426349386572838, + 0.008302280679345131, + 0.013154922984540462, + 0.007228541187942028, + 0.0032153299544006586, + 0.00798861589282751, + 0.0038970038294792175, + 0.003364626783877611, + 0.0053693074733018875, + 0.003809985937550664, + 0.006693960167467594, + 0.01665346324443817, + 0.014871492050588131, + 0.016455398872494698, + 0.011133405379951, + 0.009483687579631805, + 0.01356165949255228, + 0.01481730118393898, + 0.01912556029856205, + 0.014807944186031818, + 0.011556994169950485, + 0.011429348029196262, + 0.006601115223020315, + 0.008096952922642231, + 0.007699099835008383, + 0.006019900552928448, + 0.012420364655554295, + 0.01043605711311102, + 0.009744024835526943, + 0.012219836004078388, + 0.011473041027784348, + 0.014179985038936138, + 0.009623279795050621, + 0.01731041632592678, + 0.0054793814197182655, + 0.01938159577548504, + 0.014763030223548412, + 0.027396634221076965, + 0.023409783840179443, + 0.05641079694032669, + 0.02214120514690876, + 0.027247553691267967, + 0.017085114493966103, + 0.008833606727421284, + 0.008484197780489922, + 0.008819985203444958, + 0.012824871577322483, + 0.011732885614037514, + 0.00880589708685875, + 0.014405332505702972, + 0.00963527336716652, + 0.005086464341729879, + 0.0077775148674845695, + 0.005133861675858498, + 0.006209421902894974, + 0.007435403298586607, + 0.006037717219442129, + 0.010277964174747467, + 0.006722405552864075, + 0.010079829022288322, + 0.011174934916198254, + 0.02893669530749321, + 0.009375539608299732, + 0.007626860402524471, + 0.03598443791270256, + 0.006324692629277706, + 0.00707571767270565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.009775623679161072 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0699998065829277 + ], + "opposite_logprob_tokens": [ + -2.6949996948242188 + ], + "logprob_reward": -0.0699998065829277, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0699998065829277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1665191799402237, + 0.056529298424720764, + 0.10628850758075714, + 0.05309055745601654, + 0.09632868319749832, + 0.06767819076776505, + 0.13519974052906036, + 0.1117924302816391, + 0.09119171649217606, + 0.11538175493478775 + ], + "credit_map": [ + 0.021663585677742958, + 0.022574661299586296, + 0.017301956191658974, + 0.029036279767751694, + 0.012718811631202698, + 0.021006742492318153, + 0.01349263172596693, + 0.028724519535899162, + 0.00829053483903408, + 0.0070152017287909985, + 0.006079603917896748, + 0.0074854567646980286, + 0.004764801822602749, + 0.009609304368495941, + 0.005739970598369837, + 0.007544421590864658, + 0.015296290628612041, + 0.008499200455844402, + 0.01858385279774666, + 0.013411427848041058, + 0.00877219345420599, + 0.01040306780487299, + 0.013983277603983879, + 0.017339196056127548, + 0.00688347639515996, + 0.006785066798329353, + 0.006233522202819586, + 0.005790715105831623, + 0.008591242134571075, + 0.00659757386893034, + 0.0063673704862594604, + 0.005841588135808706, + 0.00836282316595316, + 0.013621211051940918, + 0.014858285896480083, + 0.00980595126748085, + 0.014194079674780369, + 0.010177137330174446, + 0.010783025063574314, + 0.014526166953146458, + 0.0058093927800655365, + 0.007423060480505228, + 0.00993379857391119, + 0.008371131494641304, + 0.014001704752445221, + 0.00436776876449585, + 0.012009567581117153, + 0.0057617658749222755, + 0.012023476883769035, + 0.01341710053384304, + 0.017078779637813568, + 0.014962800778448582, + 0.0299094095826149, + 0.016550516709685326, + 0.021212531253695488, + 0.010045139119029045, + 0.008637890219688416, + 0.011715848930180073, + 0.008117781020700932, + 0.02329402230679989, + 0.014957033097743988, + 0.01632491499185562, + 0.015245061367750168, + 0.013499869965016842, + 0.006822407711297274, + 0.012421678751707077, + 0.0075270747765898705, + 0.009839070029556751, + 0.015125305391848087, + 0.009462165646255016, + 0.019936926662921906, + 0.010057087056338787, + 0.009253579191863537, + 0.008420455269515514, + 0.031371328979730606, + 0.006322386674582958, + 0.00839448906481266, + 0.03594308719038963, + 0.009103030897676945, + 0.006573398597538471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.0699998065829277 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward tilt, starting from an extremely low angle, first captures a pair of hands in black, fingerless tactical gloves, carefully and with impassive precision, applying a high-tech, postpartum shapewear garment to the smooth, cold synthetic legs of a bionic figure. The flesh-colored shapewear's highly stretchable fibers are interwoven with dense biological circuitry that glows faintly blue. As the camera steadily ascends, we see an expressionless figure with short, clean-cut black hair, dressed in dark gray technical coveralls. Their silhouette is sharply defined by the flickering, cold fluorescent light overhead. The background is a chaotic, dystopian urban corner. Through rain-stained and grime-stained windows, a massive holographic billboard outside casts shifting magenta and cyan neon light pollution, intermingling with the cool white light of the interior. The person's movements were motionless as he mechanically pulled the corset over the doll's lifeless torso. The fluorescent light precisely outlined every intricate line on the garment. Finally, the camera passed over the doll's shoulders and slowly paused on its delicate face, focusing on a pair of glass-like, empty, lifeless eyes. No light was reflected in its deep pupils. The whole scene, under the cold cyberpunk aesthetic, was filled with a profound and unsettling silence.", + "seed": 21808, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_4/index_153_seed_21808_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.025941069424152374, + "vlm_logprob": -0.025941069424152374, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.025941069424152374, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05779876559972763 + ], + "opposite_logprob_tokens": [ + -2.9327988624572754 + ], + "logprob_reward": -0.05779876559972763, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05779876559972763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6375114917755127, + 0.025494420900940895, + 0.06621739268302917, + 0.04452119022607803, + 0.053098078817129135, + 0.03979731723666191, + 0.013553029857575893, + 0.0299180056899786, + 0.05343155190348625, + 0.036457475274801254 + ], + "credit_map": [ + 0.05464080348610878, + 0.0638243779540062, + 0.11517219990491867, + 0.10835486650466919, + 0.06656293570995331, + 0.06216348335146904, + 0.07830789685249329, + 0.08848492801189423, + 0.002423160942271352, + 0.0029708468355238438, + 0.0027978140860795975, + 0.0037736217491328716, + 0.004308107774704695, + 0.0025161602534353733, + 0.0039063021540641785, + 0.0027984101325273514, + 0.004294170532375574, + 0.006195234600454569, + 0.008367019705474377, + 0.011940247379243374, + 0.007533257361501455, + 0.007420963607728481, + 0.014448410831391811, + 0.006018091458827257, + 0.00430572684854269, + 0.005952881183475256, + 0.003512477269396186, + 0.005723294802010059, + 0.005811380688101053, + 0.004244659096002579, + 0.0079374173656106, + 0.0070333536714315414, + 0.008039849810302258, + 0.004590465687215328, + 0.005019363481551409, + 0.006338309496641159, + 0.01023134682327509, + 0.00691091176122427, + 0.006143040023744106, + 0.005824798718094826, + 0.005605805199593306, + 0.0045375581830739975, + 0.005026568192988634, + 0.009224471636116505, + 0.004572554025799036, + 0.0025748054031282663, + 0.004949978552758694, + 0.003305577440187335, + 0.0017875676276162267, + 0.0019294294761493802, + 0.0021440982818603516, + 0.001764805056154728, + 0.002010877477005124, + 0.0013763895258307457, + 0.0012946908827871084, + 0.0012451724614948034, + 0.005662800278514624, + 0.003006967017427087, + 0.002628676826134324, + 0.003269179491326213, + 0.0034613655880093575, + 0.0037354978267103434, + 0.0048310235142707825, + 0.0033225000370293856, + 0.006068440154194832, + 0.010589041747152805, + 0.0045127226039767265, + 0.0051865954883396626, + 0.009902000427246094, + 0.006897774059325457, + 0.005828097462654114, + 0.004446887876838446, + 0.004500233102589846, + 0.005021640099585056, + 0.005363817326724529, + 0.0036350395530462265, + 0.004127893131226301, + 0.006014124024659395, + 0.004395272582769394, + 0.0033994547557085752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.05779876559972763 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020601926371455193 + ], + "opposite_logprob_tokens": [ + -3.895601987838745 + ], + "logprob_reward": -0.020601926371455193, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020601926371455193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4214022755622864, + 0.06927334517240524, + 0.06595504283905029, + 0.08002839237451553, + 0.06026158854365349, + 0.06099570915102959, + 0.037962280213832855, + 0.04676331952214241, + 0.08740715682506561, + 0.0699508786201477 + ], + "credit_map": [ + 0.036952387541532516, + 0.03643954545259476, + 0.06632298231124878, + 0.08049438148736954, + 0.04797358810901642, + 0.03601602092385292, + 0.06281084567308426, + 0.054392535239458084, + 0.006526954472064972, + 0.007660552393645048, + 0.008224421180784702, + 0.01265866681933403, + 0.010081271640956402, + 0.005512791220098734, + 0.010066647082567215, + 0.008542041294276714, + 0.005523379426449537, + 0.006306054070591927, + 0.009681564755737782, + 0.012038158252835274, + 0.0072496989741921425, + 0.006637822836637497, + 0.01282609160989523, + 0.005692271050065756, + 0.01194059383124113, + 0.010473492555320263, + 0.007095101289451122, + 0.007811976596713066, + 0.010799662210047245, + 0.009108788333833218, + 0.014024238102138042, + 0.008774534799158573, + 0.00748898321762681, + 0.005467195063829422, + 0.004983124788850546, + 0.006255475338548422, + 0.012419885024428368, + 0.009737914428114891, + 0.00797598622739315, + 0.005933027248829603, + 0.007982241921126842, + 0.007264522835612297, + 0.006627657916396856, + 0.010129084810614586, + 0.007300226949155331, + 0.00624922476708889, + 0.0077322800643742085, + 0.007710472214967012, + 0.004602220840752125, + 0.004769299644976854, + 0.005742152687162161, + 0.005140575114637613, + 0.005237216129899025, + 0.0057898955419659615, + 0.003381521673873067, + 0.003299397649243474, + 0.00825856626033783, + 0.00521429255604744, + 0.0059548248536884785, + 0.006003268528729677, + 0.0039212629199028015, + 0.006867403630167246, + 0.005377547815442085, + 0.0051661524921655655, + 0.012592408806085587, + 0.014620703645050526, + 0.010012795217335224, + 0.011983604170382023, + 0.014621119014918804, + 0.011842893436551094, + 0.005178605206310749, + 0.0065550291910767555, + 0.008233370259404182, + 0.010446572676301003, + 0.009327815845608711, + 0.0062667797319591045, + 0.010299095883965492, + 0.010824358090758324, + 0.007862858474254608, + 0.006690028123557568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.020601926371455193 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189996480941772 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.018189996480941772, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018189996480941772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4135943353176117, + 0.056180719286203384, + 0.07129528373479843, + 0.09021532535552979, + 0.054384272545576096, + 0.046147193759679794, + 0.03754409775137901, + 0.026755351573228836, + 0.13410131633281708, + 0.0697820782661438 + ], + "credit_map": [ + 0.035971831530332565, + 0.035421762615442276, + 0.06371454894542694, + 0.08009964972734451, + 0.04170001298189163, + 0.04655233025550842, + 0.05293780565261841, + 0.05719638615846634, + 0.004831015598028898, + 0.007422775495797396, + 0.005660364869982004, + 0.007081142161041498, + 0.012373795732855797, + 0.0046471403911709785, + 0.0076024713926017284, + 0.006562010385096073, + 0.003999650478363037, + 0.005252658389508724, + 0.0068096015602350235, + 0.00941331684589386, + 0.008967632427811623, + 0.007009237539023161, + 0.023178445175290108, + 0.006664741318672895, + 0.010338283143937588, + 0.008764932863414288, + 0.010509657673537731, + 0.008007862605154514, + 0.013643670827150345, + 0.009551191702485085, + 0.013642030768096447, + 0.015757687389850616, + 0.005869516637176275, + 0.003330846317112446, + 0.0050241583958268166, + 0.007031343877315521, + 0.011740000918507576, + 0.009803990833461285, + 0.007553720381110907, + 0.00403069332242012, + 0.006456369534134865, + 0.004860747139900923, + 0.004342725034803152, + 0.00671252841129899, + 0.006763036362826824, + 0.006179003044962883, + 0.00611264631152153, + 0.004720136057585478, + 0.004498150199651718, + 0.004734506364911795, + 0.005402152892202139, + 0.005670204292982817, + 0.0035402504727244377, + 0.0036165521014481783, + 0.007143581286072731, + 0.002938699210062623, + 0.004607170820236206, + 0.002875810256227851, + 0.002995989751070738, + 0.002758933696895838, + 0.0030305078253149986, + 0.0046592289581894875, + 0.002835123101249337, + 0.0029925841372460127, + 0.02013394981622696, + 0.012917455285787582, + 0.01923948898911476, + 0.017588872462511063, + 0.01392369531095028, + 0.0209297277033329, + 0.010884123854339123, + 0.018483981490135193, + 0.008026321418583393, + 0.012645801529288292, + 0.007984627038240433, + 0.008726486936211586, + 0.006281198002398014, + 0.007598837371915579, + 0.010520043782889843, + 0.007998757064342499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.018189996480941772 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539708986878395 + ], + "opposite_logprob_tokens": [ + -4.387539863586426 + ], + "logprob_reward": -0.012539708986878395, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539708986878395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22310760617256165, + 0.07995114475488663, + 0.060159966349601746, + 0.09358680248260498, + 0.08343289792537689, + 0.062480732798576355, + 0.10461501032114029, + 0.122128427028656, + 0.06751830875873566, + 0.10301908105611801 + ], + "credit_map": [ + 0.02920219860970974, + 0.022763390094041824, + 0.02628450281918049, + 0.04021187871694565, + 0.019664155319333076, + 0.024618515744805336, + 0.028215477243065834, + 0.03214750811457634, + 0.008187399245798588, + 0.007914560846984386, + 0.008477895520627499, + 0.011016706936061382, + 0.011798019520938396, + 0.010922743007540703, + 0.0128217414021492, + 0.008812078274786472, + 0.005656911991536617, + 0.004862270317971706, + 0.013256105594336987, + 0.007514527067542076, + 0.007484567817300558, + 0.007085350342094898, + 0.00919041782617569, + 0.005109818186610937, + 0.011537477374076843, + 0.009910202585160732, + 0.01072671264410019, + 0.009463334456086159, + 0.01143623422831297, + 0.011876283213496208, + 0.013619978912174702, + 0.015016581863164902, + 0.008844408206641674, + 0.006715110037475824, + 0.007934334687888622, + 0.009654063731431961, + 0.013658477924764156, + 0.01786680333316326, + 0.011090653948485851, + 0.007669043727219105, + 0.010166757740080357, + 0.005773336160928011, + 0.006457534618675709, + 0.0077796634286642075, + 0.0105154262855649, + 0.005469328258186579, + 0.008429516106843948, + 0.007889175787568092, + 0.014342444948852062, + 0.01535663940012455, + 0.01590772531926632, + 0.01044822484254837, + 0.012948195450007915, + 0.014107183553278446, + 0.01144374068826437, + 0.010060857981443405, + 0.02407926321029663, + 0.015480170026421547, + 0.012068592943251133, + 0.01411568932235241, + 0.01205872930586338, + 0.02770889364182949, + 0.007709941361099482, + 0.008907145820558071, + 0.009115681052207947, + 0.011850669048726559, + 0.005470121279358864, + 0.0063619231805205345, + 0.01007288508117199, + 0.011265257373452187, + 0.005901229102164507, + 0.007480539381504059, + 0.014702020213007927, + 0.010588510893285275, + 0.014200059697031975, + 0.00921627227216959, + 0.008527586236596107, + 0.01668316125869751, + 0.011834126897156239, + 0.017267340794205666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.012539708986878395 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands wearing black fingerless tactical gloves visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bionic figure with smooth, cold synthetic legs wearing flesh-colored shapewear with faint blue glowing biological circuitry visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start at an extremely low angle and steadily ascend while the hands apply the garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a chaotic, dystopian urban corner with rain-stained windows showing magenta and cyan neon light pollution?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pass over the figure's shoulders and pause on a face with glass-like, empty, lifeless eyes reflecting no light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057494968175888 + ], + "opposite_logprob_tokens": [ + -3.8955750465393066 + ], + "logprob_reward": -0.02057494968175888, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057494968175888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11609386652708054, + 0.0479673370718956, + 0.0709068551659584, + 0.038605064153671265, + 0.09039060771465302, + 0.034662794321775436, + 0.062313806265592575, + 0.059040915220975876, + 0.15853245556354523, + 0.3214862644672394 + ], + "credit_map": [ + 0.009990540333092213, + 0.011658677831292152, + 0.009691037237644196, + 0.01665632240474224, + 0.01362912543118, + 0.02271328493952751, + 0.013935182243585587, + 0.017819693312048912, + 0.005072193685919046, + 0.006697706878185272, + 0.0064412737265229225, + 0.00879335030913353, + 0.00767201604321599, + 0.0032695098780095577, + 0.0057980832643806934, + 0.004223200026899576, + 0.006528912112116814, + 0.005733412690460682, + 0.01704045943915844, + 0.007859578356146812, + 0.009688637219369411, + 0.007641058415174484, + 0.011846881359815598, + 0.004567915108054876, + 0.006024821195751429, + 0.004713754169642925, + 0.0036990847438573837, + 0.0054230154491961, + 0.004488741047680378, + 0.004175947979092598, + 0.005924144294112921, + 0.004155556205660105, + 0.0072552356868982315, + 0.006899715401232243, + 0.007352661807090044, + 0.006353940814733505, + 0.018656663596630096, + 0.021381551399827003, + 0.01085506659001112, + 0.01163577288389206, + 0.0053926981054246426, + 0.003970799967646599, + 0.005025406833738089, + 0.005168440751731396, + 0.003394866129383445, + 0.002945519983768463, + 0.005271888803690672, + 0.003493170253932476, + 0.007366738747805357, + 0.008490649051964283, + 0.008315098471939564, + 0.009740603156387806, + 0.00882668886333704, + 0.00661784503608942, + 0.008051344193518162, + 0.004904838278889656, + 0.008776269853115082, + 0.006598689127713442, + 0.00565525284036994, + 0.005499540362507105, + 0.0056095425970852375, + 0.00817248597741127, + 0.010922315530478954, + 0.007806823123246431, + 0.024225465953350067, + 0.024238022044301033, + 0.019712280482053757, + 0.021406011655926704, + 0.021146660670638084, + 0.021849077194929123, + 0.014049621298909187, + 0.011905302293598652, + 0.03819030895829201, + 0.054353468120098114, + 0.039740316569805145, + 0.029314160346984863, + 0.0549105629324913, + 0.0462169274687767, + 0.03451154753565788, + 0.024248966947197914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.02057494968175888 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_5/index_tensor([8408, 8408])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_5/index_tensor([8408, 8408])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dceff50f054ccaf8ad052a2cdd9b3128306619be --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_5/index_tensor([8408, 8408])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium close-up maintains sharp focus on the expressive face of the Goblin King, who sits in a gritty, desaturated throne room constructed entirely of scrap metal. A wild, tangled mane of unkempt, dirty-blonde hair, dotted with gray streaks, frames his rugged, pale face. His most striking feature is his mismatched eyes; one is a piercing, cold blue, while the other is a cloudy, milky brown. He slumps slightly on his throne—a jagged, asymmetrical monstrosity composed of rusted car parts and twisted rebar, with a blurred background suggesting corrugated cast-iron walls and dangling chains. A harsh light source from above carves deep shadows into his features, accentuating the texture of his skin and the glint of faded metal trinkets woven into his coarse-woven tunic. His brow furrows deeply, his thin lips pressed into a tight line, and he slowly and deliberately rolls his mismatched eyes upwards, a languid movement imbued with deep weariness, revealing a blood-red white before his heavy lids begin to droop. The entire scene is rendered in a cold, desaturated palette of metallic grey, rust, and muted earth tones, creating a tense and oppressive atmosphere of grim dejection.", + "seed": 6821, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_5/index_8408_seed_6821_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04379759691655636, + "vlm_logprob": -0.04379759691655636, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.04379759691655636, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019743002951145172 + ], + "opposite_logprob_tokens": [ + -5.644742965698242 + ], + "logprob_reward": -0.019743002951145172, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.019743002951145172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13056029379367828, + 0.05989648774266243, + 0.030251426622271538, + 0.03750286623835564, + 0.043977487832307816, + 0.04856995493173599, + 0.05717119947075844, + 0.09692677110433578, + 0.07513374835252762, + 0.42000970244407654 + ], + "credit_map": [ + 0.015111647546291351, + 0.011611013673245907, + 0.015068084932863712, + 0.013615438714623451, + 0.015528305433690548, + 0.0234121922403574, + 0.020902927964925766, + 0.015310680493712425, + 0.007107678335160017, + 0.005810473579913378, + 0.011313132010400295, + 0.008414110168814659, + 0.007701586931943893, + 0.0061805276200175285, + 0.008277430199086666, + 0.005091552156955004, + 0.003759791376069188, + 0.004453386180102825, + 0.0030830223113298416, + 0.004298091866075993, + 0.0035417363978922367, + 0.0032386244274675846, + 0.004224277567118406, + 0.003652498358860612, + 0.004504398442804813, + 0.0035154561046510935, + 0.005722840316593647, + 0.00495745288208127, + 0.005312382243573666, + 0.00431434391066432, + 0.004658928140997887, + 0.004517064429819584, + 0.007683786563575268, + 0.005722818896174431, + 0.0029159672558307648, + 0.00345802353695035, + 0.004957411903887987, + 0.005594933871179819, + 0.006587791256606579, + 0.00705675920471549, + 0.005282900296151638, + 0.0037160443607717752, + 0.008546424098312855, + 0.007244226988404989, + 0.006506697274744511, + 0.004852601792663336, + 0.0063758734613657, + 0.006045191548764706, + 0.010897856205701828, + 0.003794959280639887, + 0.0030816264916211367, + 0.00442494684830308, + 0.00762364873662591, + 0.013497687876224518, + 0.008123058825731277, + 0.0057274154387414455, + 0.016168203204870224, + 0.007074045483022928, + 0.00435621989890933, + 0.005819255951792002, + 0.012434572912752628, + 0.025373436510562897, + 0.01613600365817547, + 0.009565038606524467, + 0.008117839694023132, + 0.0068764942698180676, + 0.003961712121963501, + 0.006323737092316151, + 0.00907222367823124, + 0.015017646364867687, + 0.016114363446831703, + 0.009649740532040596, + 0.037552621215581894, + 0.015643514692783356, + 0.009534299373626709, + 0.015110493637621403, + 0.03094293549656868, + 0.13579986989498138, + 0.1569615602493286, + 0.018464453518390656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.019743002951145172 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255948655307293 + ], + "opposite_logprob_tokens": [ + -4.387559413909912 + ], + "logprob_reward": -0.01255948655307293, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255948655307293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3379446566104889, + 0.05974140018224716, + 0.04647325724363327, + 0.02887345291674137, + 0.04305101931095123, + 0.03222108259797096, + 0.062164437025785446, + 0.07157884538173676, + 0.12372548878192902, + 0.19422632455825806 + ], + "credit_map": [ + 0.03512773662805557, + 0.04195694625377655, + 0.041198715567588806, + 0.04273562133312225, + 0.03796934708952904, + 0.05209307000041008, + 0.05094990134239197, + 0.035913337022066116, + 0.0072008417919278145, + 0.009266872890293598, + 0.006825125310570002, + 0.008446146734058857, + 0.007606978993862867, + 0.006054356228560209, + 0.007087136153131723, + 0.007253940682858229, + 0.004449347499758005, + 0.004469478502869606, + 0.0046693505719304085, + 0.005651485174894333, + 0.00663383724167943, + 0.009217970073223114, + 0.007567552849650383, + 0.0038142374251037836, + 0.0033230262342840433, + 0.0030618312302976847, + 0.003505558241158724, + 0.004085894674062729, + 0.004075915087014437, + 0.003083229996263981, + 0.004708097316324711, + 0.0030298996716737747, + 0.006038447376340628, + 0.004305437672883272, + 0.004678226076066494, + 0.00542677566409111, + 0.004526953678578138, + 0.005623785778880119, + 0.005329502280801535, + 0.00712189357727766, + 0.003559801960363984, + 0.0026086505968123674, + 0.0027643165085464716, + 0.004163646604865789, + 0.005810895934700966, + 0.003021512646228075, + 0.006756872870028019, + 0.00353538547642529, + 0.006033569574356079, + 0.00830251444131136, + 0.0038727878127247095, + 0.004090562462806702, + 0.009148294106125832, + 0.0145852817222476, + 0.008557717315852642, + 0.00757370563223958, + 0.007198364473879337, + 0.009165607392787933, + 0.0044218795374035835, + 0.007652708329260349, + 0.011173882521688938, + 0.013263046741485596, + 0.010953924618661404, + 0.007749424781650305, + 0.01978375017642975, + 0.009092078544199467, + 0.005579141899943352, + 0.008169985376298428, + 0.015569065697491169, + 0.03164685517549515, + 0.022298453375697136, + 0.011586158536374569, + 0.020712481811642647, + 0.012268343940377235, + 0.005832897499203682, + 0.007298222277313471, + 0.015776317566633224, + 0.06083125248551369, + 0.06181792914867401, + 0.009688882157206535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.01255948655307293 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029812322929501534 + ], + "opposite_logprob_tokens": [ + -3.5298123359680176 + ], + "logprob_reward": -0.029812322929501534, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029812322929501534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15522697567939758, + 0.039516955614089966, + 0.021235059946775436, + 0.02481401525437832, + 0.0338815338909626, + 0.01991899125277996, + 0.021775975823402405, + 0.03982081264257431, + 0.2543897330760956, + 0.3894199728965759 + ], + "credit_map": [ + 0.018928492441773415, + 0.01991473138332367, + 0.024885030463337898, + 0.02199026569724083, + 0.01852497272193432, + 0.016362925991415977, + 0.018418552353978157, + 0.01620199903845787, + 0.004798430018126965, + 0.004502382129430771, + 0.007931440137326717, + 0.0051588257774710655, + 0.004366913810372353, + 0.004340521525591612, + 0.0050380779430270195, + 0.0033803610131144524, + 0.001873895525932312, + 0.0017202392918989062, + 0.0018259218195453286, + 0.0022082021459937096, + 0.0033866679295897484, + 0.004082830157130957, + 0.004239107947796583, + 0.001898195710964501, + 0.002235144842416048, + 0.0028216419741511345, + 0.002797488821670413, + 0.004441413562744856, + 0.004143515136092901, + 0.0028167932759970427, + 0.003151514334604144, + 0.0024065030738711357, + 0.0055551608093082905, + 0.0047005885280668736, + 0.0019398429431021214, + 0.0027048385236412287, + 0.0037646174896508455, + 0.004358688369393349, + 0.004195013083517551, + 0.006662783212959766, + 0.002410947810858488, + 0.001413921476341784, + 0.0019486206583678722, + 0.0019389641238376498, + 0.002520805923268199, + 0.003212460083886981, + 0.002636029850691557, + 0.0038372413255274296, + 0.0018344197887927294, + 0.0024022599682211876, + 0.001519853132776916, + 0.003487061709165573, + 0.004859425593167543, + 0.0024182808119803667, + 0.003125638235360384, + 0.002129035070538521, + 0.004435571376234293, + 0.003931935876607895, + 0.0026826721150428057, + 0.004614482633769512, + 0.005866833031177521, + 0.008723312057554722, + 0.005331801250576973, + 0.004234203603118658, + 0.04151187464594841, + 0.018166009336709976, + 0.011288797482848167, + 0.013980795629322529, + 0.030079688876867294, + 0.06843023747205734, + 0.045695263892412186, + 0.0252370722591877, + 0.035739343613386154, + 0.010063540190458298, + 0.007993602193892002, + 0.010485456325113773, + 0.026140695437788963, + 0.13455259799957275, + 0.146956205368042, + 0.017488515004515648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.029812322929501534 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983037196099758 + ], + "opposite_logprob_tokens": [ + -3.5298304557800293 + ], + "logprob_reward": -0.02983037196099758, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983037196099758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09692832082509995, + 0.07060906291007996, + 0.051860060542821884, + 0.020219171419739723, + 0.015156551264226437, + 0.016769900918006897, + 0.023637868463993073, + 0.08412103354930878, + 0.07916031777858734, + 0.5415377616882324 + ], + "credit_map": [ + 0.012253973633050919, + 0.014694174751639366, + 0.010476525872945786, + 0.009397903457283974, + 0.008749074302613735, + 0.01605379208922386, + 0.012529201805591583, + 0.012773674912750721, + 0.008809050545096397, + 0.010556434281170368, + 0.013549751602113247, + 0.009872287511825562, + 0.006495944224298, + 0.007979189045727253, + 0.007222184911370277, + 0.006124221719801426, + 0.007008495274931192, + 0.0063308244571089745, + 0.010600756853818893, + 0.00740321958437562, + 0.003941398113965988, + 0.005956290289759636, + 0.004822651855647564, + 0.005796424113214016, + 0.0023628061171621084, + 0.002085348591208458, + 0.0028068204410374165, + 0.00218927301466465, + 0.0023062690161168575, + 0.002336116973310709, + 0.003424606518819928, + 0.0027079307474195957, + 0.0018011294305324554, + 0.0012512977700680494, + 0.001627886900678277, + 0.0019428265513852239, + 0.0024100549053400755, + 0.0015359981916844845, + 0.002504725242033601, + 0.002082633087411523, + 0.0019651439506560564, + 0.0013215368380770087, + 0.0016116381157189608, + 0.0018999925814568996, + 0.0028755059465765953, + 0.0016975495964288712, + 0.0027446988970041275, + 0.0026538344100117683, + 0.00290058646351099, + 0.0022189419250935316, + 0.0023019039072096348, + 0.0023888282012194395, + 0.0035656788386404514, + 0.004218412097543478, + 0.0037501035258173943, + 0.002293413272127509, + 0.011728052981197834, + 0.008066313341259956, + 0.004303100053220987, + 0.005529471207410097, + 0.011460144072771072, + 0.02183636464178562, + 0.013123791664838791, + 0.008073793724179268, + 0.01103561744093895, + 0.005926889833062887, + 0.006416959688067436, + 0.006505628582090139, + 0.008205614984035492, + 0.01742415875196457, + 0.011875047348439693, + 0.011770399287343025, + 0.04858303442597389, + 0.01319020427763462, + 0.011189991608262062, + 0.014170778915286064, + 0.030901603400707245, + 0.19331753253936768, + 0.2064741998910904, + 0.023710349574685097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.02983037196099758 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704280018806458 + ], + "opposite_logprob_tokens": [ + -2.127042770385742 + ], + "logprob_reward": -0.12704280018806458, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704280018806458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04892149567604065, + 0.017805835232138634, + 0.021656997501850128, + 0.023416953161358833, + 0.024913785979151726, + 0.03412201628088951, + 0.026464076712727547, + 0.02374592050909996, + 0.30965232849121094, + 0.4693005681037903 + ], + "credit_map": [ + 0.005547594744712114, + 0.005452052690088749, + 0.0044842008501291275, + 0.005718456581234932, + 0.005424008704721928, + 0.008521294221282005, + 0.009584215469658375, + 0.00418967567384243, + 0.001905324519611895, + 0.0015419716946780682, + 0.0022703504655510187, + 0.002376092365011573, + 0.0032472361344844103, + 0.002205177443102002, + 0.0027280105277895927, + 0.0015316728968173265, + 0.0017685577040538192, + 0.0017020755913108587, + 0.0022047236561775208, + 0.0021654118318110704, + 0.004268441814929247, + 0.0036239714827388525, + 0.0038346184883266687, + 0.0020891979802399874, + 0.0019538775086402893, + 0.0022181388922035694, + 0.001691583776846528, + 0.003088476834818721, + 0.004713674541562796, + 0.0032277293503284454, + 0.004088424611836672, + 0.0024350492749363184, + 0.002139228628948331, + 0.0029797281604260206, + 0.0017240484012290835, + 0.00289243645966053, + 0.005620291456580162, + 0.0028687191661447287, + 0.00474198255687952, + 0.0019473531283438206, + 0.0030658384785056114, + 0.0031122961081564426, + 0.00192911340855062, + 0.0037624661345034838, + 0.00633880402892828, + 0.0053993347100913525, + 0.007332657929509878, + 0.0031815057154744864, + 0.0033963338937610388, + 0.0026311834808439016, + 0.001988979522138834, + 0.0029762759804725647, + 0.0051748682744801044, + 0.002616387791931629, + 0.003546600928530097, + 0.004133450798690319, + 0.0022763945162296295, + 0.0026781714987009764, + 0.0017209480283781886, + 0.001898609334602952, + 0.004472014494240284, + 0.005317832343280315, + 0.0031654692720621824, + 0.0022164834663271904, + 0.049028705805540085, + 0.02253400720655918, + 0.015785329043865204, + 0.01879064552485943, + 0.038540683686733246, + 0.08161760866641998, + 0.05258805677294731, + 0.030767319723963737, + 0.042258862406015396, + 0.014197757467627525, + 0.010106938891112804, + 0.017764056101441383, + 0.04335477575659752, + 0.14903302490711212, + 0.17006996273994446, + 0.02251521125435829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.12704280018806458 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up maintains sharp focus on the expressive face of the Goblin King, who sits in a gritty, desaturated throne room constructed entirely of scrap metal. A wild, tangled mane of unkempt, dirty-blonde hair, dotted with gray streaks, frames his rugged, pale face. His most striking feature is his mismatched eyes; one is a piercing, cold blue, while the other is a cloudy, milky brown. He slumps slightly on his throne—a jagged, asymmetrical monstrosity composed of rusted car parts and twisted rebar, with a blurred background suggesting corrugated cast-iron walls and dangling chains. A harsh light source from above carves deep shadows into his features, accentuating the texture of his skin and the glint of faded metal trinkets woven into his coarse-woven tunic. His brow furrows deeply, his thin lips pressed into a tight line, and he slowly and deliberately rolls his mismatched eyes upwards, a languid movement imbued with deep weariness, revealing a blood-red white before his heavy lids begin to droop. The entire scene is rendered in a cold, desaturated palette of metallic grey, rust, and muted earth tones, creating a tense and oppressive atmosphere of grim dejection.", + "seed": 6822, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_5/index_8408_seed_6822_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.038146608509123325, + "vlm_logprob": -0.038146608509123325, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.038146608509123325, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018340056762099266 + ], + "opposite_logprob_tokens": [ + -5.518340110778809 + ], + "logprob_reward": -0.018340056762099266, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.018340056762099266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32195207476615906, + 0.05651500076055527, + 0.035539161413908005, + 0.05275673419237137, + 0.05373616889119148, + 0.06731759011745453, + 0.15684889256954193, + 0.11378388106822968, + 0.08059147745370865, + 0.06095902621746063 + ], + "credit_map": [ + 0.021170668303966522, + 0.013682789169251919, + 0.09704574942588806, + 0.11618014425039291, + 0.0154333570972085, + 0.017407333478331566, + 0.020593104884028435, + 0.02043893001973629, + 0.012813309207558632, + 0.006448051892220974, + 0.006676791235804558, + 0.008250470273196697, + 0.005430758930742741, + 0.006076603662222624, + 0.0048132347874343395, + 0.00600578123703599, + 0.005542030557990074, + 0.00436529191210866, + 0.004630365874618292, + 0.006251370534300804, + 0.0036658928729593754, + 0.003438933053985238, + 0.004602889996021986, + 0.003042385447770357, + 0.005193149205297232, + 0.006180730182677507, + 0.006162401754409075, + 0.0063629150390625, + 0.006702952552586794, + 0.008414753712713718, + 0.0076953452080488205, + 0.006044489797204733, + 0.00706368638202548, + 0.006307745352387428, + 0.006732509937137365, + 0.006028953939676285, + 0.005665063392370939, + 0.0059334528632462025, + 0.0058035701513290405, + 0.01020118873566389, + 0.01008591614663601, + 0.007826904766261578, + 0.007748199626803398, + 0.008844578638672829, + 0.009115341119468212, + 0.007295859046280384, + 0.009206851944327354, + 0.007193936500698328, + 0.015996498987078667, + 0.01715756580233574, + 0.016789311543107033, + 0.01975158601999283, + 0.01645658165216446, + 0.015538971871137619, + 0.02056783251464367, + 0.03459054231643677, + 0.01369200274348259, + 0.016222411766648293, + 0.01357252337038517, + 0.015077020041644573, + 0.010498473420739174, + 0.010166775435209274, + 0.013157587498426437, + 0.02139708772301674, + 0.011233353056013584, + 0.007346113212406635, + 0.008580505847930908, + 0.010805079713463783, + 0.007482547778636217, + 0.008418302051723003, + 0.012104135937988758, + 0.014621441252529621, + 0.006149901542812586, + 0.005408202763646841, + 0.007664518430829048, + 0.011623106896877289, + 0.006435027811676264, + 0.006326409988105297, + 0.006841636728495359, + 0.010510224848985672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.018340056762099266 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009803482331335545 + ], + "opposite_logprob_tokens": [ + -4.634803295135498 + ], + "logprob_reward": -0.009803482331335545, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009803482331335545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2051527202129364, + 0.10087155550718307, + 0.038955844938755035, + 0.041982557624578476, + 0.05262268707156181, + 0.30109211802482605, + 0.1302805244922638, + 0.0319395437836647, + 0.04532415419816971, + 0.05177829787135124 + ], + "credit_map": [ + 0.01821158453822136, + 0.014014366082847118, + 0.04354202747344971, + 0.05638964846730232, + 0.019303325563669205, + 0.016601650044322014, + 0.025396868586540222, + 0.011693240143358707, + 0.013021759688854218, + 0.00936826877295971, + 0.013786722905933857, + 0.02167389541864395, + 0.007497158367186785, + 0.014011897146701813, + 0.009474234655499458, + 0.012037623673677444, + 0.004189193248748779, + 0.004161582328379154, + 0.006016007158905268, + 0.004234942141920328, + 0.005001115147024393, + 0.0038943602703511715, + 0.006899240426719189, + 0.0045594023540616035, + 0.005510438233613968, + 0.005063842050731182, + 0.0048822155222296715, + 0.00605073431506753, + 0.005457681138068438, + 0.00460319546982646, + 0.006253072060644627, + 0.004161378368735313, + 0.008923642337322235, + 0.004821445792913437, + 0.007376717869192362, + 0.006821704097092152, + 0.005023320205509663, + 0.006766931619495153, + 0.006385584827512503, + 0.0065033393912017345, + 0.05804876610636711, + 0.05523869767785072, + 0.029695140197873116, + 0.03435932472348213, + 0.030029812827706337, + 0.027661791071295738, + 0.031215915456414223, + 0.034842681139707565, + 0.02374708279967308, + 0.014514102600514889, + 0.014094486832618713, + 0.016324663534760475, + 0.013005743734538555, + 0.012225094251334667, + 0.01632857322692871, + 0.020040767267346382, + 0.004827904049307108, + 0.004222108516842127, + 0.00282909139059484, + 0.004098905250430107, + 0.0033686815295368433, + 0.003824050072580576, + 0.004049245268106461, + 0.004719557706266642, + 0.006658837199211121, + 0.0033684121444821358, + 0.003961180802434683, + 0.006425077095627785, + 0.004092386458069086, + 0.005336410831660032, + 0.006598955485969782, + 0.00888289324939251, + 0.00692714424803853, + 0.0037999413907527924, + 0.005404976662248373, + 0.00794556550681591, + 0.006382019259035587, + 0.007763487286865711, + 0.008867926895618439, + 0.004687236621975899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.009803482331335545 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329564467072487 + ], + "opposite_logprob_tokens": [ + -3.7732956409454346 + ], + "logprob_reward": -0.02329564467072487, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329564467072487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3764500319957733, + 0.04248940944671631, + 0.036499857902526855, + 0.0287441685795784, + 0.04564492404460907, + 0.21377308666706085, + 0.14723865687847137, + 0.0446510873734951, + 0.035613417625427246, + 0.028895419090986252 + ], + "credit_map": [ + 0.0622713528573513, + 0.032971713691949844, + 0.06461404263973236, + 0.10905694961547852, + 0.016310838982462883, + 0.03728150576353073, + 0.027480777353048325, + 0.026462798938155174, + 0.005025007296353579, + 0.0048561785370111465, + 0.006255756132304668, + 0.009010190144181252, + 0.004939367063343525, + 0.004216254688799381, + 0.004772993735969067, + 0.0034136553294956684, + 0.005078713409602642, + 0.0034930631518363953, + 0.004903016611933708, + 0.006122652906924486, + 0.0033947706688195467, + 0.003738960949704051, + 0.006399689242243767, + 0.003368988400325179, + 0.002967743668705225, + 0.003047487698495388, + 0.0028981631621718407, + 0.003721313551068306, + 0.003860843600705266, + 0.0036241926718503237, + 0.004262624308466911, + 0.004361797124147415, + 0.007474224548786879, + 0.0045068212784826756, + 0.0066480934619903564, + 0.005688214674592018, + 0.005286155268549919, + 0.004740923643112183, + 0.005379736889153719, + 0.005920750088989735, + 0.04050976037979126, + 0.031000608578324318, + 0.023939112201333046, + 0.0272526778280735, + 0.022419653832912445, + 0.021378418430685997, + 0.022752726450562477, + 0.024520112201571465, + 0.021652402356266975, + 0.017770353704690933, + 0.015701796859502792, + 0.01764325052499771, + 0.01631039008498192, + 0.014293143525719643, + 0.018030723556876183, + 0.025836585089564323, + 0.006103143095970154, + 0.006225864868611097, + 0.004116015508770943, + 0.005563726648688316, + 0.004388807807117701, + 0.004337050952017307, + 0.00532133411616087, + 0.008595142513513565, + 0.004448020830750465, + 0.002713477239012718, + 0.002898530336096883, + 0.005146201699972153, + 0.003073715139180422, + 0.0038030939176678658, + 0.006411343812942505, + 0.007119033019989729, + 0.004933814983814955, + 0.002578957937657833, + 0.0036978512071073055, + 0.0042928424663841724, + 0.003804942360147834, + 0.002900709630921483, + 0.004026755224913359, + 0.0026595431845635176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.02329564467072487 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636268176138401 + ], + "opposite_logprob_tokens": [ + -3.651362657546997 + ], + "logprob_reward": -0.02636268176138401, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636268176138401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1862863302230835, + 0.04060479253530502, + 0.29331547021865845, + 0.02996605448424816, + 0.05529765784740448, + 0.056961096823215485, + 0.16655631363391876, + 0.06178010627627373, + 0.03204035013914108, + 0.07719183713197708 + ], + "credit_map": [ + 0.015307741239666939, + 0.01393242459744215, + 0.0474834069609642, + 0.05428411439061165, + 0.011669418774545193, + 0.01808035373687744, + 0.014570430852472782, + 0.010958458296954632, + 0.006620493251830339, + 0.004596844781190157, + 0.004485671874135733, + 0.0077634467743337154, + 0.004878041800111532, + 0.004570117220282555, + 0.004186228383332491, + 0.00350395031273365, + 0.06339990347623825, + 0.03448987752199173, + 0.037206199020147324, + 0.048881273716688156, + 0.01997278816998005, + 0.0344250425696373, + 0.029301151633262634, + 0.025639234110713005, + 0.005270009860396385, + 0.0028014760464429855, + 0.005160101223737001, + 0.004029054194688797, + 0.00354534899815917, + 0.0029592649079859257, + 0.0037658712826669216, + 0.0024349275045096874, + 0.007621158380061388, + 0.007166207302361727, + 0.006230703089386225, + 0.00920589454472065, + 0.00639122212305665, + 0.006335856392979622, + 0.005401675123721361, + 0.006944939959794283, + 0.005870339926332235, + 0.012960528954863548, + 0.005067520774900913, + 0.0069690304808318615, + 0.0072349789552390575, + 0.005468078888952732, + 0.006226333323866129, + 0.007164285518229008, + 0.029141154140233994, + 0.014211151748895645, + 0.01660333387553692, + 0.021569307893514633, + 0.014087634161114693, + 0.02578118070960045, + 0.014496522955596447, + 0.03066602535545826, + 0.010161720216274261, + 0.008394298143684864, + 0.010599513538181782, + 0.008186391554772854, + 0.004644683562219143, + 0.006518099457025528, + 0.005943761207163334, + 0.0073316399939358234, + 0.006963000632822514, + 0.0026057760696858168, + 0.0049263290129601955, + 0.003994773142039776, + 0.002350227441638708, + 0.0029558951500803232, + 0.003293755929917097, + 0.004950589034706354, + 0.008464028127491474, + 0.009290432557463646, + 0.009405896998941898, + 0.011930146254599094, + 0.005333132576197386, + 0.01759306900203228, + 0.006392649840563536, + 0.008782478980720043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.02636268176138401 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293117702007294 + ], + "opposite_logprob_tokens": [ + -2.237931251525879 + ], + "logprob_reward": -0.11293117702007294, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293117702007294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19303099811077118, + 0.04128643125295639, + 0.08289174735546112, + 0.03621040657162666, + 0.03444535285234451, + 0.11581002920866013, + 0.19062083959579468, + 0.1489637792110443, + 0.11109808087348938, + 0.045642342418432236 + ], + "credit_map": [ + 0.028514504432678223, + 0.018547838553786278, + 0.030891621485352516, + 0.022990252822637558, + 0.019236288964748383, + 0.019428765401244164, + 0.027605606243014336, + 0.025816118344664574, + 0.007335183676332235, + 0.004016826394945383, + 0.006039878353476524, + 0.00928407721221447, + 0.0035480521619319916, + 0.004663438536226749, + 0.002816305961459875, + 0.0035826670937240124, + 0.016027500852942467, + 0.009085300378501415, + 0.00988607294857502, + 0.01600823923945427, + 0.006943214684724808, + 0.009495601058006287, + 0.008531070314347744, + 0.00691473763436079, + 0.003741371678188443, + 0.0031655868515372276, + 0.005751009564846754, + 0.004903902765363455, + 0.005601285956799984, + 0.0029495402704924345, + 0.007047217804938555, + 0.0030504881870001554, + 0.004538966342806816, + 0.003981627989560366, + 0.005703130271285772, + 0.004007314797490835, + 0.0031524095684289932, + 0.0035151175688952208, + 0.003548306878656149, + 0.0059984782710671425, + 0.0226492527872324, + 0.008265084587037563, + 0.01389103103429079, + 0.015741605311632156, + 0.014151020906865597, + 0.013706553727388382, + 0.01473760325461626, + 0.012667866423726082, + 0.021626578643918037, + 0.02411961741745472, + 0.022742899134755135, + 0.022815590724349022, + 0.021936867386102676, + 0.01758335903286934, + 0.02332928031682968, + 0.03646664693951607, + 0.014042122289538383, + 0.024124322459101677, + 0.01365085318684578, + 0.018804185092449188, + 0.01567932777106762, + 0.013869552873075008, + 0.018266214057803154, + 0.030527180060744286, + 0.015682993456721306, + 0.007318223360925913, + 0.015885960310697556, + 0.015585435554385185, + 0.009933535940945148, + 0.010469372384250164, + 0.019061647355556488, + 0.01716090366244316, + 0.005117889493703842, + 0.0036583475302904844, + 0.004246455617249012, + 0.007216216064989567, + 0.0061882613226771355, + 0.004784081596881151, + 0.00966876931488514, + 0.004762317985296249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.11293117702007294 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up maintains sharp focus on the expressive face of the Goblin King, who sits in a gritty, desaturated throne room constructed entirely of scrap metal. A wild, tangled mane of unkempt, dirty-blonde hair, dotted with gray streaks, frames his rugged, pale face. His most striking feature is his mismatched eyes; one is a piercing, cold blue, while the other is a cloudy, milky brown. He slumps slightly on his throne—a jagged, asymmetrical monstrosity composed of rusted car parts and twisted rebar, with a blurred background suggesting corrugated cast-iron walls and dangling chains. A harsh light source from above carves deep shadows into his features, accentuating the texture of his skin and the glint of faded metal trinkets woven into his coarse-woven tunic. His brow furrows deeply, his thin lips pressed into a tight line, and he slowly and deliberately rolls his mismatched eyes upwards, a languid movement imbued with deep weariness, revealing a blood-red white before his heavy lids begin to droop. The entire scene is rendered in a cold, desaturated palette of metallic grey, rust, and muted earth tones, creating a tense and oppressive atmosphere of grim dejection.", + "seed": 6823, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_5/index_8408_seed_6823_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04305174108594656, + "vlm_logprob": -0.04305174108594656, + "vlm_yes_no_margin": 3.849999952316284, + "vlm_token_credit": -0.04305174108594656, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020980916917324066 + ], + "opposite_logprob_tokens": [ + -5.0209808349609375 + ], + "logprob_reward": -0.020980916917324066, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.020980916917324066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20040597021579742, + 0.07362577319145203, + 0.052442632615566254, + 0.04416606202721596, + 0.052805397659540176, + 0.07293535768985748, + 0.08684302866458893, + 0.08127714693546295, + 0.14299403131008148, + 0.19250456988811493 + ], + "credit_map": [ + 0.01576252467930317, + 0.03423156961798668, + 0.017031526193022728, + 0.026181135326623917, + 0.045464567840099335, + 0.013372993096709251, + 0.02451075240969658, + 0.023850899189710617, + 0.007387136574834585, + 0.00678087305277586, + 0.0043897177092731, + 0.010205973871052265, + 0.011272239498794079, + 0.0067710294388234615, + 0.016608214005827904, + 0.010210588574409485, + 0.009330570697784424, + 0.005701247602701187, + 0.0038698820862919092, + 0.005662148352712393, + 0.007045943755656481, + 0.007744914852082729, + 0.006443883292376995, + 0.006644040811806917, + 0.006070816423743963, + 0.003921750467270613, + 0.003295297734439373, + 0.0065895188599824905, + 0.007865517400205135, + 0.004529040306806564, + 0.005217996425926685, + 0.006676123011857271, + 0.006731717381626368, + 0.005252055823802948, + 0.0033609711099416018, + 0.007944204844534397, + 0.00640247855335474, + 0.004253640305250883, + 0.010687524452805519, + 0.008172803558409214, + 0.009478135965764523, + 0.005790140014141798, + 0.0050115156918764114, + 0.009978990070521832, + 0.009637271985411644, + 0.009889217093586922, + 0.012758869677782059, + 0.010391218587756157, + 0.00846686027944088, + 0.006476723589003086, + 0.0045340401120483875, + 0.014552699401974678, + 0.00971357710659504, + 0.012104873545467854, + 0.021012164652347565, + 0.009982085786759853, + 0.011083228513598442, + 0.009702342562377453, + 0.007095121778547764, + 0.01598156988620758, + 0.01068998221307993, + 0.008845992386341095, + 0.008501030504703522, + 0.009377882815897465, + 0.017450055107474327, + 0.03238758444786072, + 0.011793157085776329, + 0.01619834639132023, + 0.020387884229421616, + 0.012870652601122856, + 0.01608326844871044, + 0.015823090448975563, + 0.02107684127986431, + 0.03310338407754898, + 0.01609254814684391, + 0.02369491010904312, + 0.023672539740800858, + 0.017982730641961098, + 0.03413274511694908, + 0.022748887538909912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.020980916917324066 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00596789363771677 + ], + "opposite_logprob_tokens": [ + -5.13096809387207 + ], + "logprob_reward": -0.00596789363771677, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00596789363771677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24480150640010834, + 0.09120739251375198, + 0.0610424168407917, + 0.046903274953365326, + 0.055421024560928345, + 0.09610417485237122, + 0.05575399473309517, + 0.03979165852069855, + 0.10100759565830231, + 0.20796693861484528 + ], + "credit_map": [ + 0.022711046040058136, + 0.021776340901851654, + 0.020562121644616127, + 0.032102201133966446, + 0.04443288967013359, + 0.021285435184836388, + 0.04614957422018051, + 0.03578190878033638, + 0.008697488345205784, + 0.00580665934830904, + 0.0064970883540809155, + 0.00790052954107523, + 0.012627046555280685, + 0.007071777246892452, + 0.02272963337600231, + 0.019877176731824875, + 0.00639757327735424, + 0.006786249112337828, + 0.005324382800608873, + 0.011243289336562157, + 0.011120990850031376, + 0.005862307734787464, + 0.006858645007014275, + 0.007448992226272821, + 0.004617010243237019, + 0.00467568077147007, + 0.0038519813679158688, + 0.007557534612715244, + 0.0061793252825737, + 0.008058711886405945, + 0.006186781916767359, + 0.005776256322860718, + 0.005725851282477379, + 0.005326401442289352, + 0.00450043473392725, + 0.007660949137061834, + 0.009262064471840858, + 0.005302398931235075, + 0.010437262244522572, + 0.007205669768154621, + 0.009503057226538658, + 0.011493099853396416, + 0.007476839702576399, + 0.017612462863326073, + 0.01437984500080347, + 0.009637337177991867, + 0.014978189021348953, + 0.01102335937321186, + 0.008737667463719845, + 0.006068337708711624, + 0.004486857447773218, + 0.005417174194008112, + 0.006756036542356014, + 0.007502341642975807, + 0.00988047569990158, + 0.006905111018568277, + 0.005032581742852926, + 0.004887968301773071, + 0.0031970618292689323, + 0.004467341583222151, + 0.0070486487820744514, + 0.004878031089901924, + 0.005057777278125286, + 0.005222254898399115, + 0.012379687279462814, + 0.014992103911936283, + 0.008877209387719631, + 0.015604335814714432, + 0.015816915780305862, + 0.009719779714941978, + 0.013697448186576366, + 0.009920123033225536, + 0.017192833125591278, + 0.025027422234416008, + 0.01648716628551483, + 0.03948814794421196, + 0.041657622903585434, + 0.021294137462973595, + 0.022740257903933525, + 0.024079371243715286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.00596789363771677 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018191518262028694 + ], + "opposite_logprob_tokens": [ + -4.018191337585449 + ], + "logprob_reward": -0.018191518262028694, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018191518262028694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28483662009239197, + 0.10083366185426712, + 0.08292283117771149, + 0.05918970704078674, + 0.04192032292485237, + 0.07283879071474075, + 0.06816553324460983, + 0.08667811006307602, + 0.10496038943529129, + 0.0976540595293045 + ], + "credit_map": [ + 0.02799837477505207, + 0.026973631232976913, + 0.02403903938829899, + 0.03757551684975624, + 0.04797178879380226, + 0.028453536331653595, + 0.0474947914481163, + 0.04432994872331619, + 0.010716806165874004, + 0.00692471768707037, + 0.007385929115116596, + 0.01017696037888527, + 0.015187540091574192, + 0.008982403203845024, + 0.02063112147152424, + 0.02082817815244198, + 0.008234447799623013, + 0.010210303589701653, + 0.007266166154295206, + 0.016480866819620132, + 0.013196528889238834, + 0.008109240792691708, + 0.009319874458014965, + 0.010105403140187263, + 0.0064437552355229855, + 0.006179479416459799, + 0.004124636761844158, + 0.012574347667396069, + 0.009145260788500309, + 0.0061383359134197235, + 0.00699259340763092, + 0.007591295056045055, + 0.005658789537847042, + 0.0033354233019053936, + 0.0036289014387875795, + 0.005732096266001463, + 0.006837989203631878, + 0.0034908128436654806, + 0.007686116732656956, + 0.005550191272050142, + 0.007188289426267147, + 0.007328015752136707, + 0.006399895530194044, + 0.011386790312826633, + 0.011311989277601242, + 0.005376012530177832, + 0.015833798795938492, + 0.008014000952243805, + 0.006148357875645161, + 0.00707623828202486, + 0.005224720109254122, + 0.013361367397010326, + 0.008068845607340336, + 0.008729693479835987, + 0.012356641702353954, + 0.0071996646001935005, + 0.010036257095634937, + 0.01009508315473795, + 0.00792713463306427, + 0.013689377345144749, + 0.015183649957180023, + 0.008846882730722427, + 0.010042093694210052, + 0.010857636108994484, + 0.01159136276692152, + 0.017519639804959297, + 0.010366122238337994, + 0.01257403939962387, + 0.017082838341593742, + 0.014283066615462303, + 0.009740425273776054, + 0.011802894994616508, + 0.008748783729970455, + 0.01582670398056507, + 0.013684534467756748, + 0.01600666530430317, + 0.01087501086294651, + 0.011185729876160622, + 0.009665591642260551, + 0.011661035008728504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.018191518262028694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043071165680885315 + ], + "opposite_logprob_tokens": [ + -3.1680712699890137 + ], + "logprob_reward": -0.043071165680885315, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043071165680885315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20676927268505096, + 0.10312363505363464, + 0.06704298406839371, + 0.1500765085220337, + 0.05898357182741165, + 0.11176684498786926, + 0.03599703311920166, + 0.0808330848813057, + 0.07273457944393158, + 0.11267248541116714 + ], + "credit_map": [ + 0.016701167449355125, + 0.023613756522536278, + 0.015208522789180279, + 0.027850495651364326, + 0.03877490386366844, + 0.020914878696203232, + 0.03482883796095848, + 0.028876706957817078, + 0.009676439687609673, + 0.011474941857159138, + 0.00858226977288723, + 0.01734861172735691, + 0.017447857186198235, + 0.008541927672922611, + 0.014842902310192585, + 0.015208685770630836, + 0.008866569958627224, + 0.0070154317654669285, + 0.005797794554382563, + 0.00950652826577425, + 0.013087764382362366, + 0.006617842707782984, + 0.0076118470169603825, + 0.008539198897778988, + 0.017570214346051216, + 0.014613748528063297, + 0.010713872499763966, + 0.034855522215366364, + 0.024235013872385025, + 0.012903462164103985, + 0.01873915083706379, + 0.016445517539978027, + 0.008018070831894875, + 0.005563327111303806, + 0.005342797841876745, + 0.008994003757834435, + 0.007843155413866043, + 0.007458375301212072, + 0.008606234565377235, + 0.0071576060727238655, + 0.011954382061958313, + 0.01344146765768528, + 0.00945234950631857, + 0.014532570727169514, + 0.016079751774668694, + 0.018289802595973015, + 0.014542522840201855, + 0.013474001549184322, + 0.004085469990968704, + 0.003793990472331643, + 0.003717629471793771, + 0.005426058080047369, + 0.005502020474523306, + 0.004162456840276718, + 0.004771717824041843, + 0.004537688102573156, + 0.013117136433720589, + 0.007485862821340561, + 0.0059918067418038845, + 0.009725823067128658, + 0.012809224426746368, + 0.01170768029987812, + 0.010043100453913212, + 0.009952452033758163, + 0.009964204393327236, + 0.010078327730298042, + 0.009361746720969677, + 0.009885859675705433, + 0.00985552929341793, + 0.0103797921910882, + 0.006453697104007006, + 0.006755423732101917, + 0.012981976382434368, + 0.020768767222762108, + 0.008572581224143505, + 0.012728656642138958, + 0.013176498003304005, + 0.009258808568120003, + 0.02088906057178974, + 0.014296133071184158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.043071165680885315 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704721093177795 + ], + "opposite_logprob_tokens": [ + -2.127047300338745 + ], + "logprob_reward": -0.12704721093177795, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704721093177795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.149716317653656, + 0.0792887881398201, + 0.051523976027965546, + 0.05015476420521736, + 0.029642373323440552, + 0.04143363609910011, + 0.06471795588731766, + 0.11152571439743042, + 0.18192458152770996, + 0.24007190763950348 + ], + "credit_map": [ + 0.02019020915031433, + 0.014760072343051434, + 0.013876898214221, + 0.027283698320388794, + 0.016521701589226723, + 0.015688015148043633, + 0.02061251550912857, + 0.02078321948647499, + 0.008860127069056034, + 0.00805276446044445, + 0.005513559095561504, + 0.00901345070451498, + 0.012515941634774208, + 0.007221710868179798, + 0.014819299802184105, + 0.013291933573782444, + 0.0054343524388968945, + 0.005353813525289297, + 0.004981578793376684, + 0.008066595532000065, + 0.008668575435876846, + 0.005057105794548988, + 0.006990000605583191, + 0.006971955299377441, + 0.005556577816605568, + 0.005097594112157822, + 0.0037303296849131584, + 0.008485566824674606, + 0.009142577648162842, + 0.005180902313441038, + 0.006253296509385109, + 0.006707918364554644, + 0.003845230909064412, + 0.0035074055194854736, + 0.003392360405996442, + 0.004004690796136856, + 0.004452880006283522, + 0.002781500341370702, + 0.004013061057776213, + 0.0036452454514801502, + 0.004267537966370583, + 0.0036865847650915384, + 0.004769228398799896, + 0.00762723246589303, + 0.0072062513791024685, + 0.003931587096303701, + 0.00566312437877059, + 0.004282087087631226, + 0.00628679059445858, + 0.005625732243061066, + 0.007188747636973858, + 0.01380594726651907, + 0.007891797460615635, + 0.005835928954184055, + 0.010225492529571056, + 0.007857521995902061, + 0.013433240354061127, + 0.010614822618663311, + 0.009515457786619663, + 0.0230324175208807, + 0.01776084676384926, + 0.011511125601828098, + 0.012749766930937767, + 0.012908035889267921, + 0.02286492846906185, + 0.035600945353507996, + 0.015551620163023472, + 0.020985044538974762, + 0.022609781473875046, + 0.017920086160302162, + 0.026909420266747475, + 0.01948276162147522, + 0.023844977840781212, + 0.04301932454109192, + 0.025963017717003822, + 0.03277270868420601, + 0.027085285633802414, + 0.027703087776899338, + 0.03278682753443718, + 0.026896676048636436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.12704721093177795 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up maintains sharp focus on the expressive face of the Goblin King, who sits in a gritty, desaturated throne room constructed entirely of scrap metal. A wild, tangled mane of unkempt, dirty-blonde hair, dotted with gray streaks, frames his rugged, pale face. His most striking feature is his mismatched eyes; one is a piercing, cold blue, while the other is a cloudy, milky brown. He slumps slightly on his throne—a jagged, asymmetrical monstrosity composed of rusted car parts and twisted rebar, with a blurred background suggesting corrugated cast-iron walls and dangling chains. A harsh light source from above carves deep shadows into his features, accentuating the texture of his skin and the glint of faded metal trinkets woven into his coarse-woven tunic. His brow furrows deeply, his thin lips pressed into a tight line, and he slowly and deliberately rolls his mismatched eyes upwards, a languid movement imbued with deep weariness, revealing a blood-red white before his heavy lids begin to droop. The entire scene is rendered in a cold, desaturated palette of metallic grey, rust, and muted earth tones, creating a tense and oppressive atmosphere of grim dejection.", + "seed": 6824, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_5/index_8408_seed_6824_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.052863572165369986, + "vlm_logprob": -0.052863572165369986, + "vlm_yes_no_margin": 3.949999952316284, + "vlm_token_credit": -0.052863572165369986, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017220735549926758 + ], + "opposite_logprob_tokens": [ + -5.392220497131348 + ], + "logprob_reward": -0.017220735549926758, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.017220735549926758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10483307391405106, + 0.3734452724456787, + 0.0731380507349968, + 0.07115838676691055, + 0.05478562414646149, + 0.04500264674425125, + 0.1348852515220642, + 0.03797851502895355, + 0.03894265741109848, + 0.0658305212855339 + ], + "credit_map": [ + 0.012346632778644562, + 0.01662004180252552, + 0.01147639099508524, + 0.017069395631551743, + 0.013595202937722206, + 0.009755251929163933, + 0.015073824673891068, + 0.00889633223414421, + 0.025103388354182243, + 0.0161635372787714, + 0.011831785552203655, + 0.14140437543392181, + 0.013421203009784222, + 0.012860260903835297, + 0.1352165937423706, + 0.01744409091770649, + 0.009635758586227894, + 0.009890670888125896, + 0.00637349858880043, + 0.009333859197795391, + 0.007604524027556181, + 0.012137310579419136, + 0.011165371164679527, + 0.006997060962021351, + 0.00976285245269537, + 0.008765479549765587, + 0.007728633005172014, + 0.00872464757412672, + 0.008698477409780025, + 0.010744631290435791, + 0.008995876647531986, + 0.0077377911657094955, + 0.007991334423422813, + 0.007011319510638714, + 0.005503324791789055, + 0.007175612263381481, + 0.004843239206820726, + 0.008807986043393612, + 0.007455508690327406, + 0.005997296888381243, + 0.005512084346264601, + 0.005944657605141401, + 0.003274736925959587, + 0.00799024011939764, + 0.0038159501273185015, + 0.0036800431553274393, + 0.011059976182878017, + 0.003724958747625351, + 0.017265142872929573, + 0.006391850300133228, + 0.004336272832006216, + 0.03958567604422569, + 0.00653326278552413, + 0.0060127777978777885, + 0.04840953275561333, + 0.006350738927721977, + 0.004887852817773819, + 0.0038947102148085833, + 0.0038444362580776215, + 0.006048030685633421, + 0.00462629459798336, + 0.006880481261759996, + 0.00466969795525074, + 0.0031270107720047235, + 0.004529872909188271, + 0.00481303920969367, + 0.00484313303604722, + 0.007444489281624556, + 0.0034731077030301094, + 0.004426806699484587, + 0.0052554612047970295, + 0.004156747367233038, + 0.009320498444139957, + 0.010527953505516052, + 0.005417034029960632, + 0.01327572576701641, + 0.00876890029758215, + 0.005032364744693041, + 0.009085508994758129, + 0.004402538761496544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.017220735549926758 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006754660978913307 + ], + "opposite_logprob_tokens": [ + -5.0067548751831055 + ], + "logprob_reward": -0.006754660978913307, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006754660978913307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15022063255310059, + 0.20641201734542847, + 0.08187235891819, + 0.06941956281661987, + 0.10190514475107193, + 0.10495345294475555, + 0.08569628745317459, + 0.01695184037089348, + 0.023738140240311623, + 0.15883061289787292 + ], + "credit_map": [ + 0.018780242651700974, + 0.014428923837840557, + 0.013740630820393562, + 0.029681403189897537, + 0.01771922968327999, + 0.01679602451622486, + 0.028282349929213524, + 0.010791828855872154, + 0.01439167745411396, + 0.011634863913059235, + 0.0066123055294156075, + 0.07442308217287064, + 0.010376079007983208, + 0.008261455222964287, + 0.06907559186220169, + 0.01163695938885212, + 0.010106591507792473, + 0.009444261901080608, + 0.008288509212434292, + 0.015799766406416893, + 0.007756780833005905, + 0.010578575544059277, + 0.011698954738676548, + 0.008198915980756283, + 0.009496362879872322, + 0.009677831083536148, + 0.006174732930958271, + 0.009767607785761356, + 0.007230635732412338, + 0.012305716052651405, + 0.00888169463723898, + 0.005884977523237467, + 0.011578777804970741, + 0.013211430981755257, + 0.009022088721394539, + 0.01380593329668045, + 0.010539932176470757, + 0.01223849132657051, + 0.01925031840801239, + 0.012258163653314114, + 0.015378065407276154, + 0.009656834416091442, + 0.0051323589868843555, + 0.0183242317289114, + 0.011198327876627445, + 0.006683571729809046, + 0.03198619931936264, + 0.006593858823180199, + 0.014952514320611954, + 0.009516123682260513, + 0.009191813878715038, + 0.010807006619870663, + 0.010156985372304916, + 0.008354909718036652, + 0.014085158705711365, + 0.008631771430373192, + 0.0025975718162953854, + 0.001988557865843177, + 0.0018061773153021932, + 0.0023795694578438997, + 0.0016381568275392056, + 0.0027725885156542063, + 0.00235157017596066, + 0.0014176482800394297, + 0.0031185378320515156, + 0.0028058099560439587, + 0.0019648612942546606, + 0.004409081768244505, + 0.002550680423155427, + 0.0036487786564975977, + 0.004052008967846632, + 0.001188380061648786, + 0.01891205459833145, + 0.01842374913394451, + 0.020792115479707718, + 0.020607568323612213, + 0.015405545011162758, + 0.021011222153902054, + 0.017663780599832535, + 0.02601458504796028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.006754660978913307 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584525540471077 + ], + "opposite_logprob_tokens": [ + -3.8955845832824707 + ], + "logprob_reward": -0.020584525540471077, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584525540471077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18013297021389008, + 0.48157256841659546, + 0.05800000578165054, + 0.05526566877961159, + 0.03219058737158775, + 0.019432097673416138, + 0.033762283623218536, + 0.02077057957649231, + 0.022115392610430717, + 0.09675798565149307 + ], + "credit_map": [ + 0.023632798343896866, + 0.01981908455491066, + 0.018440276384353638, + 0.02840089239180088, + 0.021620353683829308, + 0.020144112408161163, + 0.0321049802005291, + 0.015970464795827866, + 0.03235277161002159, + 0.02195008471608162, + 0.013258141465485096, + 0.19418469071388245, + 0.017299428582191467, + 0.015873681753873825, + 0.1645124852657318, + 0.02214127779006958, + 0.0073311589658260345, + 0.0063400655053555965, + 0.005612021312117577, + 0.011208368465304375, + 0.006423623766750097, + 0.007332108449190855, + 0.008154028095304966, + 0.0055986265651881695, + 0.006057563237845898, + 0.008785423822700977, + 0.005376778542995453, + 0.009356159716844559, + 0.0079411705955863, + 0.0062441336922347546, + 0.00646594725549221, + 0.00503848772495985, + 0.00392824225127697, + 0.004519602749496698, + 0.0033845510333776474, + 0.0047139255329966545, + 0.003676865715533495, + 0.004277842119336128, + 0.004723468795418739, + 0.0029660891741514206, + 0.0028594243340194225, + 0.002104070968925953, + 0.001733569661155343, + 0.0035168735776096582, + 0.0018073372775688767, + 0.002365062478929758, + 0.002908817958086729, + 0.0021369403693825006, + 0.004056590609252453, + 0.0034153510350733995, + 0.0030335995834320784, + 0.006289742887020111, + 0.0034112916328012943, + 0.004049938637763262, + 0.00615304009988904, + 0.0033527270425111055, + 0.0030827950686216354, + 0.002002839231863618, + 0.0020105030853301287, + 0.0032968539744615555, + 0.0015669924905523658, + 0.0025996388867497444, + 0.004833865910768509, + 0.0013770911609753966, + 0.0026200369466096163, + 0.0025567892007529736, + 0.0027554440312087536, + 0.0042465622536838055, + 0.002097581047564745, + 0.0023594668600708246, + 0.003994885366410017, + 0.0014846270205453038, + 0.009411959908902645, + 0.012827184982597828, + 0.015188971534371376, + 0.018432492390275, + 0.009179873391985893, + 0.009606384672224522, + 0.010568208992481232, + 0.011542906053364277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.020584525540471077 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181919876486063 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.0181919876486063, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.0181919876486063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15690572559833527, + 0.24374639987945557, + 0.04278365150094032, + 0.11339356750249863, + 0.042752936482429504, + 0.03225431963801384, + 0.2148815095424652, + 0.04563310369849205, + 0.0231657475233078, + 0.08448304235935211 + ], + "credit_map": [ + 0.018240436911582947, + 0.013856707140803337, + 0.017156969755887985, + 0.02153792977333069, + 0.022414211183786392, + 0.01730872318148613, + 0.028281977400183678, + 0.01810877025127411, + 0.016713444143533707, + 0.011645409278571606, + 0.006924376357346773, + 0.09216516464948654, + 0.009830540046095848, + 0.010209402069449425, + 0.08634719997644424, + 0.009910834953188896, + 0.005072591360658407, + 0.005067481193691492, + 0.00400790898129344, + 0.005182563327252865, + 0.004698088392615318, + 0.006898242048919201, + 0.006593524478375912, + 0.005263251252472401, + 0.013690617866814137, + 0.010680517181754112, + 0.012867478653788567, + 0.011051111854612827, + 0.01226277370005846, + 0.020463962107896805, + 0.013497758656740189, + 0.018879344686865807, + 0.0065131234005093575, + 0.0036915012169629335, + 0.00352994492277503, + 0.003957540262490511, + 0.005320434458553791, + 0.006229477468878031, + 0.007764863781630993, + 0.005746049340814352, + 0.004120783880352974, + 0.003288109553977847, + 0.0030515624675899744, + 0.0034901113249361515, + 0.004033825360238552, + 0.004923181142657995, + 0.00487388763576746, + 0.00447285920381546, + 0.029734281823039055, + 0.008456681855022907, + 0.009119345806539059, + 0.05841764435172081, + 0.011289861984550953, + 0.009534455835819244, + 0.07707420736551285, + 0.011255037970840931, + 0.006881531793624163, + 0.004294273443520069, + 0.006163104437291622, + 0.00587497977539897, + 0.004162982571870089, + 0.008742720820009708, + 0.005731231067329645, + 0.003782279323786497, + 0.0034695318900048733, + 0.002571173943579197, + 0.0021727399434894323, + 0.0029106801375746727, + 0.0024713953025639057, + 0.0038901465013623238, + 0.003360842354595661, + 0.0023192374501377344, + 0.010660775005817413, + 0.008791856467723846, + 0.013471694663167, + 0.010482129640877247, + 0.0065781171433627605, + 0.013363596983253956, + 0.009900988079607487, + 0.011233881115913391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.0181919876486063 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a Goblin King with a wild, tangled mane of dirty-blonde hair dotted with gray streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have mismatched eyes, one piercing cold blue and the other cloudy milky brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character sitting on a jagged, asymmetrical throne made of rusted car parts and twisted rebar in a gritty, desaturated scrap metal throne room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character slowly and deliberately roll his mismatched eyes upwards while his brow furrows and lips press into a tight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015659511089325 + ], + "opposite_logprob_tokens": [ + -1.7015659809112549 + ], + "logprob_reward": -0.2015659511089325, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015659511089325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22263628244400024, + 0.09985959529876709, + 0.07048623263835907, + 0.06044087931513786, + 0.0568082220852375, + 0.08291389793157578, + 0.14223170280456543, + 0.027319511398673058, + 0.028928667306900024, + 0.20837494730949402 + ], + "credit_map": [ + 0.021680207923054695, + 0.022253159433603287, + 0.01536633912473917, + 0.05470621585845947, + 0.023365598171949387, + 0.021106479689478874, + 0.04928898811340332, + 0.01486932672560215, + 0.01376147847622633, + 0.01495669037103653, + 0.00910592544823885, + 0.018844973295927048, + 0.012126806192100048, + 0.01089194044470787, + 0.013744489289820194, + 0.006427295971661806, + 0.009140575304627419, + 0.009289130568504333, + 0.006616413593292236, + 0.012896978296339512, + 0.00813600979745388, + 0.008944538421928883, + 0.009125366806983948, + 0.006337224505841732, + 0.007382891606539488, + 0.007843999192118645, + 0.006979380268603563, + 0.005616497248411179, + 0.00695077795535326, + 0.009801028296351433, + 0.007906382903456688, + 0.007959926500916481, + 0.008649718016386032, + 0.008070545271039009, + 0.005622583441436291, + 0.006092749070376158, + 0.004927015397697687, + 0.0075326645746827126, + 0.00954171922057867, + 0.0063712261617183685, + 0.014528539031744003, + 0.0037906754296272993, + 0.003289475105702877, + 0.016552116721868515, + 0.004673144314438105, + 0.004257676657289267, + 0.03127603605389595, + 0.004546234384179115, + 0.017362667247653008, + 0.008910579606890678, + 0.00788102950900793, + 0.037823889404535294, + 0.008056409657001495, + 0.007617276627570391, + 0.04524923488497734, + 0.00933061633259058, + 0.0034083491191267967, + 0.002790955826640129, + 0.0025038316380232573, + 0.004703063052147627, + 0.0026393032167106867, + 0.0024674637243151665, + 0.006739933043718338, + 0.002066615968942642, + 0.0027446397580206394, + 0.0034599078353494406, + 0.0036399185191839933, + 0.004516893066465855, + 0.0035194833762943745, + 0.0027340815868228674, + 0.005479909013956785, + 0.0028338369447737932, + 0.02556697279214859, + 0.027864620089530945, + 0.023754848167300224, + 0.03313245624303818, + 0.019895382225513458, + 0.024026693776249886, + 0.021847238764166832, + 0.032286740839481354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.2015659511089325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_6/index_tensor([5057, 5057])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_6/index_tensor([5057, 5057])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..89df9f876e676c6bde5dd25c032f10c275dfc00a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_6/index_tensor([5057, 5057])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, aerial close-up focuses on an ancient gold coin, nestled upon a dark, highly polished obsidian surface. The obverse of the coin is engraved with a coiling serpent, each scale meticulously carved, its form tightly spiraling upon itself. Under a single, dramatic key light, deep, long shadows are cast, and the coin's lustrous gold shimmers with a warm, rich glow. As movement begins, the serpent's carved lines begin to shimmer with an unnatural inner glow. Slowly, the metallic form unfolds, its body peeling from the flat surface of the coin, transforming from a static etching into a three-dimensional, solid gold creature. Now fully alive, the serpent's tiny eyes gleam like ruby needlepoints as it glides hypnotically and silently across the mirror-like black space. Its segmented golden body moves with fluid grace, the dramatic sidelight capturing each polished scale and creating mesmerizing ripples of bright highlights within the profound darkness, evoking the tense and magical atmosphere of an ancient power awakening.", + "seed": 29223, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_6/index_5057_seed_29223_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1285087987780571, + "vlm_logprob": -0.1285087987780571, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.1285087987780571, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04963161051273346 + ], + "opposite_logprob_tokens": [ + -3.4246315956115723 + ], + "logprob_reward": -0.04963161051273346, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04963161051273346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4006173610687256, + 0.06695851683616638, + 0.07173838466405869, + 0.07428637892007828, + 0.029373707249760628, + 0.08849954605102539, + 0.046580713242292404, + 0.028931334614753723, + 0.0449816957116127, + 0.1480323076248169 + ], + "credit_map": [ + 0.09041046351194382, + 0.03382188454270363, + 0.07511728256940842, + 0.030420459806919098, + 0.04246984049677849, + 0.04609730839729309, + 0.03323888033628464, + 0.04904131218791008, + 0.012780671007931232, + 0.00599135784432292, + 0.012029952369630337, + 0.006538433488458395, + 0.007181829307228327, + 0.008186893537640572, + 0.005672437138855457, + 0.008576944470405579, + 0.01633264683187008, + 0.00729712937027216, + 0.014720919542014599, + 0.007065386977046728, + 0.007608931045979261, + 0.0074257561936974525, + 0.005208134651184082, + 0.006079492624849081, + 0.026560667902231216, + 0.005985019262880087, + 0.015089157968759537, + 0.003584826597943902, + 0.005128951743245125, + 0.005995845887809992, + 0.0030956652481108904, + 0.008846250362694263, + 0.0061768763698637486, + 0.004215204622596502, + 0.004751943983137608, + 0.0020494689233601093, + 0.0029636798426508904, + 0.0031435589771717787, + 0.0023419633507728577, + 0.00373101350851357, + 0.019291331991553307, + 0.008031835779547691, + 0.01563086360692978, + 0.007780771236866713, + 0.009594904258847237, + 0.01143025141209364, + 0.006566809490323067, + 0.010172782465815544, + 0.010622628033161163, + 0.005279072094708681, + 0.008610528893768787, + 0.003082065610215068, + 0.005906909238547087, + 0.004494712222367525, + 0.003279751166701317, + 0.005305054131895304, + 0.004474867135286331, + 0.0037895729765295982, + 0.0037051762919873, + 0.002935329219326377, + 0.0041605751030147076, + 0.0028421373572200537, + 0.0037100634071975946, + 0.0033136147540062666, + 0.009931082837283611, + 0.004310893826186657, + 0.007168764714151621, + 0.005178593099117279, + 0.003940279129892588, + 0.006021685898303986, + 0.003294228110462427, + 0.005136172287166119, + 0.032975777983665466, + 0.01228608563542366, + 0.021830715239048004, + 0.02475005015730858, + 0.012384556233882904, + 0.015364635735750198, + 0.013199779205024242, + 0.015240720473229885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.04963161051273346 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10040532052516937 + ], + "opposite_logprob_tokens": [ + -2.350405216217041 + ], + "logprob_reward": -0.10040532052516937, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10040532052516937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22176164388656616, + 0.08544550836086273, + 0.09098736196756363, + 0.08157125115394592, + 0.09206277132034302, + 0.07885929942131042, + 0.09580815583467484, + 0.1272166520357132, + 0.03505989536643028, + 0.09122739732265472 + ], + "credit_map": [ + 0.05903972312808037, + 0.019292863085865974, + 0.040328703820705414, + 0.019384481012821198, + 0.02437424845993519, + 0.01881423406302929, + 0.01618887297809124, + 0.024338524788618088, + 0.02021856978535652, + 0.009485309012234211, + 0.014415962621569633, + 0.009429107420146465, + 0.008317203260958195, + 0.007494296412914991, + 0.007445862982422113, + 0.008639202453196049, + 0.026599274948239326, + 0.0069659105502069, + 0.019609423354268074, + 0.0065171304158866405, + 0.00736407283693552, + 0.009736030362546444, + 0.005393131170421839, + 0.008802392520010471, + 0.028587017208337784, + 0.005965415388345718, + 0.018230954185128212, + 0.004289827775210142, + 0.00510114710777998, + 0.00785987637937069, + 0.0035562620032578707, + 0.007980757392942905, + 0.02296440489590168, + 0.007133565843105316, + 0.015824472531676292, + 0.007359305862337351, + 0.00895815622061491, + 0.011168355122208595, + 0.007019463926553726, + 0.011635066010057926, + 0.021925030276179314, + 0.006701156031340361, + 0.01506453100591898, + 0.004706644918769598, + 0.0055256313644349575, + 0.008108670823276043, + 0.003940918017178774, + 0.012886719778180122, + 0.02683320827782154, + 0.00793427973985672, + 0.01958567276597023, + 0.006444036494940519, + 0.006319583859294653, + 0.010466978885233402, + 0.004668652545660734, + 0.013555748388171196, + 0.036834679543972015, + 0.010099673643708229, + 0.023030903190374374, + 0.011568469926714897, + 0.011132133193314075, + 0.01329038105905056, + 0.006876682862639427, + 0.014383742585778236, + 0.006849657744169235, + 0.0036893084179610014, + 0.00521750608459115, + 0.0040521807968616486, + 0.004421936813741922, + 0.0035761913750320673, + 0.0029792541172355413, + 0.004273862112313509, + 0.02928425744175911, + 0.004772799089550972, + 0.01917087659239769, + 0.009448669850826263, + 0.005577258765697479, + 0.008645669557154179, + 0.005061220843344927, + 0.009266657754778862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.10040532052516937 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337321013212204 + ], + "opposite_logprob_tokens": [ + -3.4087321758270264 + ], + "logprob_reward": -0.0337321013212204, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337321013212204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19318349659442902, + 0.050298355519771576, + 0.07747174054384232, + 0.13237427175045013, + 0.03758309781551361, + 0.033477891236543655, + 0.04414566233754158, + 0.10306551307439804, + 0.0586293525993824, + 0.26977065205574036 + ], + "credit_map": [ + 0.04022923856973648, + 0.018650349229574203, + 0.031841032207012177, + 0.01745474524796009, + 0.024955132976174355, + 0.020743226632475853, + 0.018299654126167297, + 0.02101011946797371, + 0.018889646977186203, + 0.0028963303193449974, + 0.01139749400317669, + 0.002082951134070754, + 0.0031393372919410467, + 0.0037194006145000458, + 0.002430654363706708, + 0.005742540583014488, + 0.02364400587975979, + 0.007043273188173771, + 0.015712222084403038, + 0.004766786936670542, + 0.009509947150945663, + 0.00642161350697279, + 0.004924257285892963, + 0.0054496340453624725, + 0.04087156429886818, + 0.013665684498846531, + 0.02399623766541481, + 0.007711056619882584, + 0.011304246261715889, + 0.010497773997485638, + 0.007760590873658657, + 0.016567114740610123, + 0.006795149762183428, + 0.004492721986025572, + 0.0062588718719780445, + 0.002839841181412339, + 0.0057147531770169735, + 0.003604432800784707, + 0.0037720371037721634, + 0.004105290863662958, + 0.007820694707334042, + 0.0035806631203740835, + 0.005808040965348482, + 0.0018619586480781436, + 0.0043810829520225525, + 0.003532866947352886, + 0.003270049812272191, + 0.003222535364329815, + 0.008259795606136322, + 0.004178727976977825, + 0.007254892960190773, + 0.0025487251114100218, + 0.0037233836483210325, + 0.00801450852304697, + 0.0030989956576377153, + 0.007066633086651564, + 0.0181466992944479, + 0.011963884346187115, + 0.01416675467044115, + 0.00918225571513176, + 0.016271961852908134, + 0.009969688951969147, + 0.010442492552101612, + 0.012921784073114395, + 0.008912628516554832, + 0.00793974008411169, + 0.00781520176678896, + 0.008380557410418987, + 0.0066327969543635845, + 0.005178516265004873, + 0.0076882317662239075, + 0.006081681232899427, + 0.051226355135440826, + 0.024841127917170525, + 0.037742339074611664, + 0.05005088075995445, + 0.02529001235961914, + 0.02923698164522648, + 0.025590896606445312, + 0.02579205296933651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.0337321013212204 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02986670285463333 + ], + "opposite_logprob_tokens": [ + -3.5298666954040527 + ], + "logprob_reward": -0.02986670285463333, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02986670285463333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35059496760368347, + 0.06873631477355957, + 0.04261872172355652, + 0.1212291345000267, + 0.04440784454345703, + 0.03269430994987488, + 0.03111904114484787, + 0.08251596242189407, + 0.05107416957616806, + 0.17500953376293182 + ], + "credit_map": [ + 0.08458324521780014, + 0.03570730611681938, + 0.06009252369403839, + 0.03653540834784508, + 0.03637652471661568, + 0.029491785913705826, + 0.03219335153698921, + 0.03561479598283768, + 0.013863112777471542, + 0.008253767155110836, + 0.0125099066644907, + 0.008296866901218891, + 0.0068878973834216595, + 0.005780437961220741, + 0.006082009989768267, + 0.007062308490276337, + 0.01085071824491024, + 0.004056679084897041, + 0.008955239318311214, + 0.004420339595526457, + 0.004016023129224777, + 0.0036637054290622473, + 0.0032167385797947645, + 0.003439271589741111, + 0.04300288110971451, + 0.010019020177423954, + 0.02539040893316269, + 0.005843644496053457, + 0.00855278130620718, + 0.00928401853889227, + 0.006144691724330187, + 0.012991677038371563, + 0.010433790273964405, + 0.005232091527432203, + 0.007107379846274853, + 0.0033789188601076603, + 0.005159703548997641, + 0.005080284550786018, + 0.00304622296243906, + 0.00496944785118103, + 0.008599400520324707, + 0.002374990377575159, + 0.00634689861908555, + 0.002560281427577138, + 0.0033293538726866245, + 0.003368668956682086, + 0.0026310423854738474, + 0.0034836719278246164, + 0.006405976135283709, + 0.002860127715393901, + 0.005459745414555073, + 0.0022146909032016993, + 0.0026920128148049116, + 0.005713546182960272, + 0.002387382322922349, + 0.0033855566289275885, + 0.016531070694327354, + 0.008919177576899529, + 0.01175233069807291, + 0.0071902587078511715, + 0.012109090574085712, + 0.008763355202972889, + 0.007357149384915829, + 0.009893525391817093, + 0.011807212606072426, + 0.005696598906069994, + 0.008854716084897518, + 0.005198967643082142, + 0.004945905413478613, + 0.0040223971009254456, + 0.005154900252819061, + 0.0053934664465487, + 0.03371235728263855, + 0.016842272132635117, + 0.023868316784501076, + 0.029309291392564774, + 0.016218233853578568, + 0.020404979586601257, + 0.01786424219608307, + 0.016789818182587624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.02986670285463333 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42890825867652893 + ], + "opposite_logprob_tokens": [ + -1.0539082288742065 + ], + "logprob_reward": -0.42890825867652893, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42890825867652893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1738853007555008, + 0.06620587408542633, + 0.05453861504793167, + 0.09072241187095642, + 0.0696369856595993, + 0.055981021374464035, + 0.06840313225984573, + 0.07483487576246262, + 0.06611352413892746, + 0.2796782851219177 + ], + "credit_map": [ + 0.04282992705702782, + 0.016394760459661484, + 0.029911084100604057, + 0.016920026391744614, + 0.018587276339530945, + 0.01401270180940628, + 0.018381493166089058, + 0.01684802770614624, + 0.013323345221579075, + 0.007759009022265673, + 0.012256304733455181, + 0.0058888960629701614, + 0.005521341226994991, + 0.0072875851765275, + 0.004907771944999695, + 0.009261620230972767, + 0.016015294939279556, + 0.0038564156275242567, + 0.010963653214275837, + 0.004114363808184862, + 0.002934350399300456, + 0.006173296831548214, + 0.0021605088841170073, + 0.008320733904838562, + 0.031999409198760986, + 0.007714766077697277, + 0.01822046749293804, + 0.004234302788972855, + 0.005055597983300686, + 0.007649545557796955, + 0.004769125021994114, + 0.011079193092882633, + 0.020878782495856285, + 0.006372801028192043, + 0.013229223899543285, + 0.004352046176791191, + 0.006619020830839872, + 0.00555652379989624, + 0.0053934622555971146, + 0.007235121447592974, + 0.017882393673062325, + 0.003603354562073946, + 0.012288017198443413, + 0.003564052050933242, + 0.0038156663067638874, + 0.005720797460526228, + 0.0029193973168730736, + 0.006187342572957277, + 0.020343035459518433, + 0.004152194131165743, + 0.015536215156316757, + 0.004309915006160736, + 0.005224430002272129, + 0.007797110825777054, + 0.004614978097379208, + 0.0064252507872879505, + 0.011802024208009243, + 0.010682285763323307, + 0.011288837529718876, + 0.008249489590525627, + 0.011512489058077335, + 0.008335164748132229, + 0.005539176054298878, + 0.007425414398312569, + 0.012187598273158073, + 0.00831124372780323, + 0.01019190438091755, + 0.007861388847231865, + 0.006801585666835308, + 0.006299593020230532, + 0.005239994265139103, + 0.009220217354595661, + 0.10032516717910767, + 0.014433285221457481, + 0.05819430202245712, + 0.02167120948433876, + 0.013165482319891453, + 0.02329438552260399, + 0.013356312178075314, + 0.03523813933134079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.42890825867652893 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, aerial close-up focuses on an ancient gold coin, nestled upon a dark, highly polished obsidian surface. The obverse of the coin is engraved with a coiling serpent, each scale meticulously carved, its form tightly spiraling upon itself. Under a single, dramatic key light, deep, long shadows are cast, and the coin's lustrous gold shimmers with a warm, rich glow. As movement begins, the serpent's carved lines begin to shimmer with an unnatural inner glow. Slowly, the metallic form unfolds, its body peeling from the flat surface of the coin, transforming from a static etching into a three-dimensional, solid gold creature. Now fully alive, the serpent's tiny eyes gleam like ruby needlepoints as it glides hypnotically and silently across the mirror-like black space. Its segmented golden body moves with fluid grace, the dramatic sidelight capturing each polished scale and creating mesmerizing ripples of bright highlights within the profound darkness, evoking the tense and magical atmosphere of an ancient power awakening.", + "seed": 29224, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_6/index_5057_seed_29224_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.21249530632048846, + "vlm_logprob": -0.21249530632048846, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.21249530632048846, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5440581440925598 + ], + "opposite_logprob_tokens": [ + -0.9190581440925598 + ], + "logprob_reward": -0.5440581440925598, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5440581440925598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04033994302153587, + 0.05275096371769905, + 0.1455046534538269, + 0.0652286633849144, + 0.040285419672727585, + 0.034577906131744385, + 0.4446570873260498, + 0.0228277537971735, + 0.07709337770938873, + 0.07673418521881104 + ], + "credit_map": [ + 0.0052879853174090385, + 0.0031671100296080112, + 0.007198804523795843, + 0.004114705137908459, + 0.0023924955166876316, + 0.00740297045558691, + 0.004072829615324736, + 0.006703048478811979, + 0.004999587778002024, + 0.0026558612007647753, + 0.007819905877113342, + 0.005950644612312317, + 0.0025837174616754055, + 0.013183853588998318, + 0.004779355134814978, + 0.010778042487800121, + 0.014816577546298504, + 0.008616432547569275, + 0.01951874978840351, + 0.015437831170856953, + 0.012221748940646648, + 0.02248464897274971, + 0.015865318477153778, + 0.0365433506667614, + 0.005181306507438421, + 0.0035192123614251614, + 0.008583031594753265, + 0.0077458214946091175, + 0.0040193176828324795, + 0.009869375266134739, + 0.018175480887293816, + 0.00813512597233057, + 0.003961565904319286, + 0.004260693676769733, + 0.004708459600806236, + 0.0035349063109606504, + 0.0036001126281917095, + 0.004952542018145323, + 0.009403116069734097, + 0.005864025559276342, + 0.0032413385342806578, + 0.0020641968585550785, + 0.006355884484946728, + 0.0026929890736937523, + 0.0026285804342478514, + 0.0052449144423007965, + 0.005902562756091356, + 0.006447440013289452, + 0.007370730862021446, + 0.00535515695810318, + 0.011940684169530869, + 0.005783337168395519, + 0.009332682937383652, + 0.18357741832733154, + 0.029321305453777313, + 0.19197581708431244, + 0.0018737205537036061, + 0.0019768858328461647, + 0.002082664053887129, + 0.0031651153694838285, + 0.00180365692358464, + 0.003911442123353481, + 0.0039061259012669325, + 0.004108142573386431, + 0.00419752998277545, + 0.004171402659267187, + 0.005865698214620352, + 0.009161844849586487, + 0.005788270849734545, + 0.01677677594125271, + 0.017031321302056313, + 0.014100543223321438, + 0.0032100477255880833, + 0.0025081834755837917, + 0.004053974058479071, + 0.002771117025986314, + 0.0033048938494175673, + 0.027539648115634918, + 0.00653320224955678, + 0.026813127100467682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.5440581440925598 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023343276232481003 + ], + "opposite_logprob_tokens": [ + -3.773343324661255 + ], + "logprob_reward": -0.023343276232481003, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023343276232481003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056324560195207596, + 0.03762771561741829, + 0.2892433702945709, + 0.04978327080607414, + 0.03917643055319786, + 0.02038898691534996, + 0.4061371684074402, + 0.02463342435657978, + 0.023302774876356125, + 0.05338232219219208 + ], + "credit_map": [ + 0.003070276463404298, + 0.002628347370773554, + 0.011124488897621632, + 0.005271402187645435, + 0.002873749006539583, + 0.010747837834060192, + 0.005185322370380163, + 0.015423140488564968, + 0.004116632044315338, + 0.0016213159542530775, + 0.005463968031108379, + 0.004608906339854002, + 0.0018413168145343661, + 0.006816393695771694, + 0.006192246917635202, + 0.006966935005038977, + 0.026010192930698395, + 0.01587018184363842, + 0.03592650219798088, + 0.03290879726409912, + 0.02229396626353264, + 0.046016596257686615, + 0.03543318808078766, + 0.07478395104408264, + 0.0029181058052927256, + 0.0024223981890827417, + 0.005781703162938356, + 0.004095039796084166, + 0.002939429599791765, + 0.010627713985741138, + 0.010079867206513882, + 0.010919014923274517, + 0.002824828727170825, + 0.002753905486315489, + 0.004772236105054617, + 0.004495946690440178, + 0.0036518939305096865, + 0.0060762376524508, + 0.006650074385106564, + 0.00795130804181099, + 0.0017499102978035808, + 0.0012809972977265716, + 0.004352625459432602, + 0.0015706224367022514, + 0.0012882440350949764, + 0.0027014329098165035, + 0.00307043781504035, + 0.004374716896563768, + 0.0062121679075062275, + 0.004476518370211124, + 0.00866076722741127, + 0.0043204850517213345, + 0.007856257259845734, + 0.16997376084327698, + 0.02611563913524151, + 0.17852157354354858, + 0.0021654111333191395, + 0.0016022846102714539, + 0.0022869431413710117, + 0.0031921383924782276, + 0.0017778531182557344, + 0.004710931330919266, + 0.0044061667285859585, + 0.0044916956685483456, + 0.0015290998853743076, + 0.0012228168779984117, + 0.0014690818497911096, + 0.0018827839521691203, + 0.001554254675284028, + 0.006358589977025986, + 0.0031015074346214533, + 0.006184640806168318, + 0.0034783862065523863, + 0.00364803709089756, + 0.006954032462090254, + 0.004208720754832029, + 0.005594298709183931, + 0.010708395391702652, + 0.008051780983805656, + 0.01073867455124855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.023343276232481003 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011099499650299549 + ], + "opposite_logprob_tokens": [ + -4.511099338531494 + ], + "logprob_reward": -0.011099499650299549, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011099499650299549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.084084652364254, + 0.019061528146266937, + 0.17785973846912384, + 0.09666869789361954, + 0.06477911025285721, + 0.06036992371082306, + 0.1609908789396286, + 0.022388098761439323, + 0.11493739485740662, + 0.19885998964309692 + ], + "credit_map": [ + 0.007604510523378849, + 0.005051234271377325, + 0.014756297692656517, + 0.008579785004258156, + 0.0050110300071537495, + 0.015015932731330395, + 0.008229656144976616, + 0.01983620785176754, + 0.0024116893764585257, + 0.001483738305978477, + 0.0031106427777558565, + 0.0022640535607934, + 0.001895447145216167, + 0.003351905383169651, + 0.0021020318381488323, + 0.0024420199915766716, + 0.014247801154851913, + 0.010762614198029041, + 0.02096858248114586, + 0.022083943709731102, + 0.013206852599978447, + 0.027296703308820724, + 0.026615198701620102, + 0.042678043246269226, + 0.008700689300894737, + 0.005147341173142195, + 0.009260140359401703, + 0.006310671102255583, + 0.004586035385727882, + 0.026936212554574013, + 0.011169394478201866, + 0.024558210745453835, + 0.005452262237668037, + 0.004829234443604946, + 0.007399294525384903, + 0.006749254651367664, + 0.00488708121702075, + 0.011149291880428791, + 0.010898883454501629, + 0.013413810171186924, + 0.004449875093996525, + 0.004990649875253439, + 0.010082848370075226, + 0.005704440176486969, + 0.006669744849205017, + 0.007426360622048378, + 0.011082549579441547, + 0.009963455609977245, + 0.0028524785302579403, + 0.0030547508504241705, + 0.004534290172159672, + 0.0029836688190698624, + 0.0043541258201003075, + 0.06493505090475082, + 0.010859642177820206, + 0.06741687655448914, + 0.0015316380886361003, + 0.001796605298295617, + 0.0022500427439808846, + 0.0020059077069163322, + 0.001419053296558559, + 0.003920919727534056, + 0.004566077142953873, + 0.0048978556878864765, + 0.007015555165708065, + 0.008162758313119411, + 0.006605574395507574, + 0.012254580855369568, + 0.008262038230895996, + 0.028980931267142296, + 0.015748267993330956, + 0.027907686308026314, + 0.011199063621461391, + 0.008808182552456856, + 0.02453027293086052, + 0.009799079038202763, + 0.011849775910377502, + 0.05761479586362839, + 0.01934897154569626, + 0.055709850043058395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 6 + ], + "selected_reward": -0.011099499650299549 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05503677576780319 + ], + "opposite_logprob_tokens": [ + -2.930036783218384 + ], + "logprob_reward": -0.05503677576780319, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05503677576780319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11417214572429657, + 0.03897577524185181, + 0.04427696764469147, + 0.07844319194555283, + 0.06664279848337173, + 0.19308191537857056, + 0.034706227481365204, + 0.030306197702884674, + 0.032629430294036865, + 0.3667653501033783 + ], + "credit_map": [ + 0.008041976019740105, + 0.00566707132384181, + 0.02841992676258087, + 0.009056857787072659, + 0.005520996171981096, + 0.02622835710644722, + 0.007460609544068575, + 0.023776352405548096, + 0.004734359215945005, + 0.0034061730839312077, + 0.005143257323652506, + 0.004484456032514572, + 0.0030125915072858334, + 0.006498479750007391, + 0.005769089329987764, + 0.005927372258156538, + 0.004308939445763826, + 0.004799615126103163, + 0.005635218229144812, + 0.004718082956969738, + 0.0048551554791629314, + 0.006323902402073145, + 0.006673265248537064, + 0.006962788291275501, + 0.006173861213028431, + 0.0025105534587055445, + 0.007156457751989365, + 0.003978407941758633, + 0.002940335078164935, + 0.024677421897649765, + 0.007023766171187162, + 0.023982398211956024, + 0.006378252524882555, + 0.005725073628127575, + 0.008075854741036892, + 0.007047124672681093, + 0.00603080727159977, + 0.011788464151322842, + 0.009321099147200584, + 0.012276126071810722, + 0.01575881987810135, + 0.01391754299402237, + 0.03446632996201515, + 0.015919649973511696, + 0.016405189409852028, + 0.02665150910615921, + 0.0377337783575058, + 0.03222910314798355, + 0.001733899349346757, + 0.003518400015309453, + 0.0031786570325493813, + 0.0018922706367447972, + 0.0026581771671772003, + 0.008613895624876022, + 0.004188449587672949, + 0.008922480046749115, + 0.0020950797479599714, + 0.002105630235746503, + 0.002868585055693984, + 0.0035446295514702797, + 0.002741072792559862, + 0.006099459249526262, + 0.004927888512611389, + 0.005923856049776077, + 0.0032988444436341524, + 0.0029499977827072144, + 0.0027348927687853575, + 0.003829806111752987, + 0.003466607304289937, + 0.005436111707240343, + 0.005197005812078714, + 0.005716164596378803, + 0.008821595460176468, + 0.009698481298983097, + 0.012972396798431873, + 0.011597546748816967, + 0.014128428883850574, + 0.1484878808259964, + 0.019658058881759644, + 0.14140096306800842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.05503677576780319 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289388358592987 + ], + "opposite_logprob_tokens": [ + -1.053938865661621 + ], + "logprob_reward": -0.4289388358592987, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289388358592987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09230872243642807, + 0.06646406650543213, + 0.041376907378435135, + 0.12141589820384979, + 0.1279250532388687, + 0.2977111339569092, + 0.10540002584457397, + 0.045456789433956146, + 0.06019527092576027, + 0.04174618050456047 + ], + "credit_map": [ + 0.006081908009946346, + 0.003656339133158326, + 0.019366588443517685, + 0.005905684549361467, + 0.0028713021893054247, + 0.021342121064662933, + 0.007559686899185181, + 0.025525089353322983, + 0.006850942969322205, + 0.003607835853472352, + 0.009774226695299149, + 0.006721343845129013, + 0.0057150269858539104, + 0.013243455439805984, + 0.008782380260527134, + 0.011768851429224014, + 0.00305938720703125, + 0.0027272296138107777, + 0.005003476049751043, + 0.004094637464731932, + 0.002880557207390666, + 0.006878275889903307, + 0.005122560076415539, + 0.011610783636569977, + 0.006273871287703514, + 0.00395864387974143, + 0.012484882026910782, + 0.004831807222217321, + 0.004964986816048622, + 0.0415702648460865, + 0.008989185094833374, + 0.03834225609898567, + 0.008550387807190418, + 0.010760866105556488, + 0.013928471133112907, + 0.011925915256142616, + 0.009139561094343662, + 0.01732506975531578, + 0.0370916984975338, + 0.0192030668258667, + 0.024354564025998116, + 0.02281929925084114, + 0.053856827318668365, + 0.024777814745903015, + 0.02665877342224121, + 0.04128992184996605, + 0.04872298613190651, + 0.055230945348739624, + 0.00352728390134871, + 0.004059048369526863, + 0.004485001787543297, + 0.004711411893367767, + 0.004556144587695599, + 0.037255145609378815, + 0.009100725874304771, + 0.03770526126027107, + 0.004282140638679266, + 0.003434313228353858, + 0.00577974459156394, + 0.00461873272433877, + 0.0036794114857912064, + 0.007539201062172651, + 0.007839864119887352, + 0.008283378556370735, + 0.004851835314184427, + 0.003931909799575806, + 0.004356409888714552, + 0.00625039916485548, + 0.003962685354053974, + 0.013453908264636993, + 0.010541025549173355, + 0.01284709107130766, + 0.002251455094665289, + 0.003283079946413636, + 0.0029744349885731936, + 0.002268741140142083, + 0.004875395447015762, + 0.009975504130125046, + 0.006516589783132076, + 0.009600977413356304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.4289388358592987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, aerial close-up focuses on an ancient gold coin, nestled upon a dark, highly polished obsidian surface. The obverse of the coin is engraved with a coiling serpent, each scale meticulously carved, its form tightly spiraling upon itself. Under a single, dramatic key light, deep, long shadows are cast, and the coin's lustrous gold shimmers with a warm, rich glow. As movement begins, the serpent's carved lines begin to shimmer with an unnatural inner glow. Slowly, the metallic form unfolds, its body peeling from the flat surface of the coin, transforming from a static etching into a three-dimensional, solid gold creature. Now fully alive, the serpent's tiny eyes gleam like ruby needlepoints as it glides hypnotically and silently across the mirror-like black space. Its segmented golden body moves with fluid grace, the dramatic sidelight capturing each polished scale and creating mesmerizing ripples of bright highlights within the profound darkness, evoking the tense and magical atmosphere of an ancient power awakening.", + "seed": 29225, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_6/index_5057_seed_29225_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09277249490842224, + "vlm_logprob": -0.09277249490842224, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.09277249490842224, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09858006983995438 + ], + "opposite_logprob_tokens": [ + -2.4735801219940186 + ], + "logprob_reward": -0.09858006983995438, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09858006983995438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07076571136713028, + 0.033787406980991364, + 0.024296443909406662, + 0.04197258874773979, + 0.045078929513692856, + 0.09752735495567322, + 0.18591982126235962, + 0.15737953782081604, + 0.16512663662433624, + 0.1781456172466278 + ], + "credit_map": [ + 0.008660131134092808, + 0.011766383424401283, + 0.006682978011667728, + 0.009567472152411938, + 0.008135526441037655, + 0.007383353542536497, + 0.009449321776628494, + 0.009120546281337738, + 0.0056228842586278915, + 0.003148360177874565, + 0.003553244285285473, + 0.0034467100631445646, + 0.0031902818009257317, + 0.00490571977570653, + 0.005475536920130253, + 0.004444665741175413, + 0.002864795969799161, + 0.003051582956686616, + 0.003012995235621929, + 0.0031079899054020643, + 0.0028630902525037527, + 0.002542216796427965, + 0.003932981751859188, + 0.002920790808275342, + 0.006310526747256517, + 0.004938785452395678, + 0.004918834660202265, + 0.0052410210482776165, + 0.005465102382004261, + 0.004159785341471434, + 0.0050352225080132484, + 0.005903308279812336, + 0.0064249346032738686, + 0.004279240500181913, + 0.008043884299695492, + 0.005787606816738844, + 0.004946487490087748, + 0.0037767302710562944, + 0.004275026731193066, + 0.0075450194999575615, + 0.011536899954080582, + 0.015321164391934872, + 0.012529433704912663, + 0.011376717127859592, + 0.016424739733338356, + 0.007870621047914028, + 0.0087667815387249, + 0.013701001182198524, + 0.013756965287029743, + 0.015051931142807007, + 0.044855061918497086, + 0.02497081272304058, + 0.01724322699010372, + 0.014807230792939663, + 0.016456441953778267, + 0.038778141140937805, + 0.012239040806889534, + 0.01238471269607544, + 0.0374317392706871, + 0.017266904935240746, + 0.012413010001182556, + 0.01330506056547165, + 0.0234965980052948, + 0.028842458501458168, + 0.008111564442515373, + 0.006797822192311287, + 0.056139517575502396, + 0.011681586503982544, + 0.00474568223580718, + 0.0077591766603291035, + 0.013658027164638042, + 0.05623326078057289, + 0.014283660799264908, + 0.009010634385049343, + 0.04473838582634926, + 0.021491779014468193, + 0.011143455281853676, + 0.012476516887545586, + 0.024829575791954994, + 0.04017162695527077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.09858006983995438 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813952952623367 + ], + "opposite_logprob_tokens": [ + -3.288139581680298 + ], + "logprob_reward": -0.03813952952623367, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813952952623367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09859384596347809, + 0.018287938088178635, + 0.023263778537511826, + 0.03269080072641373, + 0.025623178109526634, + 0.0747140422463417, + 0.20167700946331024, + 0.3034442663192749, + 0.1801546961069107, + 0.04155045002698898 + ], + "credit_map": [ + 0.011911381036043167, + 0.016082800924777985, + 0.010437393561005592, + 0.013277214020490646, + 0.012731813825666904, + 0.008114379830658436, + 0.012088442221283913, + 0.013950417749583721, + 0.0021286974661052227, + 0.0024284080136567354, + 0.0017193821258842945, + 0.0021628031972795725, + 0.00235791620798409, + 0.0021432843059301376, + 0.00290381140075624, + 0.0024436370003968477, + 0.003818010911345482, + 0.0023031423334032297, + 0.0029942751862108707, + 0.0020604636520147324, + 0.002348551759496331, + 0.002713302616029978, + 0.0031049526296555996, + 0.003921080846339464, + 0.0035144537687301636, + 0.0037046545185148716, + 0.006074095144867897, + 0.0033737230114638805, + 0.003679680870845914, + 0.0030028896871954203, + 0.003577298019081354, + 0.005764007102698088, + 0.002435134258121252, + 0.003103302326053381, + 0.003042925614863634, + 0.004713284317404032, + 0.004341167863458395, + 0.0018716418417170644, + 0.002945665968582034, + 0.00317005580291152, + 0.008234478533267975, + 0.009123018942773342, + 0.008556823246181011, + 0.012763881124556065, + 0.01028696820139885, + 0.0057904282584786415, + 0.007666235323995352, + 0.012292210012674332, + 0.013675522059202194, + 0.015808284282684326, + 0.05976684018969536, + 0.0176509078592062, + 0.011920937336981297, + 0.016198789700865746, + 0.019875260069966316, + 0.04678047075867653, + 0.010597403161227703, + 0.01535099744796753, + 0.06926366686820984, + 0.06602401286363602, + 0.015772733837366104, + 0.02397541143000126, + 0.048889003694057465, + 0.05357104167342186, + 0.013871566392481327, + 0.006261902861297131, + 0.05602120980620384, + 0.013161334209144115, + 0.0049315085634589195, + 0.007090480998158455, + 0.019830655306577682, + 0.05898604169487953, + 0.006262410897761583, + 0.0022629196755588055, + 0.008814137428998947, + 0.004825646989047527, + 0.0028284310828894377, + 0.0030715945176780224, + 0.005717294756323099, + 0.007768015377223492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.03813952952623367 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007644801866263151 + ], + "opposite_logprob_tokens": [ + -4.8826446533203125 + ], + "logprob_reward": -0.007644801866263151, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007644801866263151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11558635532855988, + 0.0366998091340065, + 0.025564759969711304, + 0.028067147359251976, + 0.035961609333753586, + 0.21547217667102814, + 0.1489853411912918, + 0.08234953135251999, + 0.06733224540948868, + 0.24398106336593628 + ], + "credit_map": [ + 0.01155443862080574, + 0.013171757571399212, + 0.014278294518589973, + 0.014272053726017475, + 0.011892694979906082, + 0.012419023551046848, + 0.01370400283485651, + 0.024294087663292885, + 0.006584628019481897, + 0.004581182263791561, + 0.004210993181914091, + 0.003256592433899641, + 0.004073986783623695, + 0.0042731971479952335, + 0.004281012807041407, + 0.0054382155649363995, + 0.005128134042024612, + 0.002993215573951602, + 0.0036487365141510963, + 0.001953704049810767, + 0.0023924862034618855, + 0.003032540902495384, + 0.002949668560177088, + 0.0034662759862840176, + 0.004667575936764479, + 0.002529159653931856, + 0.004956947639584541, + 0.003413393162190914, + 0.002234457293525338, + 0.0033839058596640825, + 0.003060483606532216, + 0.0038212251383811235, + 0.0032102945260703564, + 0.004567576572299004, + 0.004244019277393818, + 0.005374856758862734, + 0.0057404544204473495, + 0.0031295684166252613, + 0.004790219012647867, + 0.004904617089778185, + 0.024838628247380257, + 0.029754560440778732, + 0.0223679356276989, + 0.030688730999827385, + 0.039499521255493164, + 0.019846148788928986, + 0.01900675892829895, + 0.029469909146428108, + 0.007159060798585415, + 0.012399785220623016, + 0.02691558189690113, + 0.03677152097225189, + 0.010617468506097794, + 0.009578048251569271, + 0.01754770241677761, + 0.027996178716421127, + 0.005229379050433636, + 0.005627982318401337, + 0.021392349153757095, + 0.010315557941794395, + 0.008134070783853531, + 0.00784155260771513, + 0.010206659324467182, + 0.013601986691355705, + 0.005659137852489948, + 0.004497922491282225, + 0.01506632287055254, + 0.009773668833076954, + 0.004046195652335882, + 0.006997202523052692, + 0.006906225346028805, + 0.014385567046701908, + 0.02940279431641102, + 0.011844824999570847, + 0.05693933367729187, + 0.03456469252705574, + 0.016779346391558647, + 0.01677732914686203, + 0.04061621055006981, + 0.037056516855955124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.007644801866263151 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03816087171435356 + ], + "opposite_logprob_tokens": [ + -3.288160800933838 + ], + "logprob_reward": -0.03816087171435356, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03816087171435356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08952183276414871, + 0.0324363149702549, + 0.027537019923329353, + 0.04744639992713928, + 0.033328402787446976, + 0.21010276675224304, + 0.09671895951032639, + 0.12651324272155762, + 0.11096493154764175, + 0.22543008625507355 + ], + "credit_map": [ + 0.010238167829811573, + 0.01278684101998806, + 0.009860754944384098, + 0.012851151637732983, + 0.012584982439875603, + 0.007516189012676477, + 0.013584389351308346, + 0.01009935699403286, + 0.004698947537690401, + 0.0034477850422263145, + 0.003936484921723604, + 0.003928229678422213, + 0.003566853003576398, + 0.003357327077537775, + 0.004674997646361589, + 0.0048256912268698215, + 0.005651822313666344, + 0.002672799164429307, + 0.003809713525697589, + 0.0021117369178682566, + 0.002170553430914879, + 0.0033567859791219234, + 0.0032093003392219543, + 0.004554309882223606, + 0.008607592433691025, + 0.005500685423612595, + 0.006135652307420969, + 0.005243523977696896, + 0.0048957448452711105, + 0.006483893841505051, + 0.004582367837429047, + 0.005996940191835165, + 0.0031615193001925945, + 0.004002991132438183, + 0.005917719565331936, + 0.004315932281315327, + 0.004539781715720892, + 0.002591538242995739, + 0.003633495420217514, + 0.005165423266589642, + 0.020485563203692436, + 0.029401248320937157, + 0.020354613661766052, + 0.03468386083841324, + 0.034249965101480484, + 0.016826080158352852, + 0.02247900702059269, + 0.03162244334816933, + 0.007829993963241577, + 0.00785668846219778, + 0.027657942846417427, + 0.008418216370046139, + 0.008018949069082737, + 0.007562043145298958, + 0.006937155965715647, + 0.02243797667324543, + 0.007024530321359634, + 0.008297901600599289, + 0.03736844286322594, + 0.01181184221059084, + 0.008764078840613365, + 0.010315144434571266, + 0.01815664768218994, + 0.024774661287665367, + 0.010157307609915733, + 0.00473429961130023, + 0.03395663574337959, + 0.0073652369901537895, + 0.0028339677955955267, + 0.005858789663761854, + 0.009308715350925922, + 0.03674998879432678, + 0.02591553144156933, + 0.01299658976495266, + 0.051339372992515564, + 0.03240032494068146, + 0.015650983899831772, + 0.0186139103025198, + 0.036385536193847656, + 0.032127849757671356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.03816087171435356 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813372015953064 + ], + "opposite_logprob_tokens": [ + -1.4063372611999512 + ], + "logprob_reward": -0.2813372015953064, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813372015953064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04441727697849274, + 0.01906101033091545, + 0.011831127107143402, + 0.04474952816963196, + 0.032525163143873215, + 0.029839767143130302, + 0.11091525852680206, + 0.24476274847984314, + 0.11561846733093262, + 0.34627965092658997 + ], + "credit_map": [ + 0.00368953263387084, + 0.007722540758550167, + 0.005474557168781757, + 0.00606920663267374, + 0.004890622571110725, + 0.0033142189495265484, + 0.006942569278180599, + 0.006314026191830635, + 0.003038235241547227, + 0.0023006012197583914, + 0.0026139216497540474, + 0.0020164246670901775, + 0.0015896965051069856, + 0.00214953301474452, + 0.0027857315726578236, + 0.0025668651796877384, + 0.001354468404315412, + 0.001546986517496407, + 0.0012299127411097288, + 0.0017203937750309706, + 0.0014509407337754965, + 0.0011077856179326773, + 0.0017598908161744475, + 0.001660747453570366, + 0.004932001233100891, + 0.005042953882366419, + 0.009493179619312286, + 0.003731883130967617, + 0.0047732070088386536, + 0.004595838487148285, + 0.003821011632680893, + 0.0083594536408782, + 0.0032006672117859125, + 0.0032639962155371904, + 0.006198412273079157, + 0.004602074157446623, + 0.0036555156111717224, + 0.003363406751304865, + 0.003177012549713254, + 0.005064075812697411, + 0.002716438379138708, + 0.0039795273914933205, + 0.004876982420682907, + 0.0038238810375332832, + 0.004340037237852812, + 0.002511707367375493, + 0.0031952408608049154, + 0.004395951982587576, + 0.005250147078186274, + 0.00852251797914505, + 0.026323620229959488, + 0.017703916877508163, + 0.009121118113398552, + 0.011054104194045067, + 0.01088868547230959, + 0.022051149979233742, + 0.009771797806024551, + 0.016040300950407982, + 0.03227914124727249, + 0.06489378213882446, + 0.015192809514701366, + 0.02290506288409233, + 0.04226544871926308, + 0.04141440987586975, + 0.007811930030584335, + 0.004383599851280451, + 0.035882722586393356, + 0.0075476206839084625, + 0.0029142573475837708, + 0.0053314873948693275, + 0.014828967861831188, + 0.036917880177497864, + 0.031835608184337616, + 0.017770901322364807, + 0.09727313369512558, + 0.043580297380685806, + 0.02004959248006344, + 0.020102689042687416, + 0.05677199363708496, + 0.058895450085401535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.2813372015953064 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, aerial close-up focuses on an ancient gold coin, nestled upon a dark, highly polished obsidian surface. The obverse of the coin is engraved with a coiling serpent, each scale meticulously carved, its form tightly spiraling upon itself. Under a single, dramatic key light, deep, long shadows are cast, and the coin's lustrous gold shimmers with a warm, rich glow. As movement begins, the serpent's carved lines begin to shimmer with an unnatural inner glow. Slowly, the metallic form unfolds, its body peeling from the flat surface of the coin, transforming from a static etching into a three-dimensional, solid gold creature. Now fully alive, the serpent's tiny eyes gleam like ruby needlepoints as it glides hypnotically and silently across the mirror-like black space. Its segmented golden body moves with fluid grace, the dramatic sidelight capturing each polished scale and creating mesmerizing ripples of bright highlights within the profound darkness, evoking the tense and magical atmosphere of an ancient power awakening.", + "seed": 29226, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_6/index_5057_seed_29226_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1568951196037233, + "vlm_logprob": -0.1568951196037233, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.1568951196037233, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.32905063033103943 + ], + "opposite_logprob_tokens": [ + -1.3290506601333618 + ], + "logprob_reward": -0.32905063033103943, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32905063033103943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05999445542693138, + 0.03346354514360428, + 0.027118846774101257, + 0.018335632979869843, + 0.08912798017263412, + 0.12389151751995087, + 0.15334255993366241, + 0.11772552877664566, + 0.10253410786390305, + 0.27446579933166504 + ], + "credit_map": [ + 0.01090182177722454, + 0.004999855998903513, + 0.004164125304669142, + 0.005602444056421518, + 0.00412594573572278, + 0.012669031508266926, + 0.0037942645139992237, + 0.013736974447965622, + 0.0038290806114673615, + 0.0036538573913276196, + 0.0025020986795425415, + 0.003621017560362816, + 0.0027266538236290216, + 0.00672012148424983, + 0.0027474553789943457, + 0.007663267198950052, + 0.0029276313725858927, + 0.0022297166287899017, + 0.001223142840899527, + 0.0022731360513716936, + 0.0016980988439172506, + 0.006763406563550234, + 0.0018230470595881343, + 0.008180668577551842, + 0.0017493849154561758, + 0.002447643084451556, + 0.001279462710954249, + 0.0021729241125285625, + 0.0015117053408175707, + 0.003531784750521183, + 0.0015651268186047673, + 0.004077602177858353, + 0.013386993668973446, + 0.007093653082847595, + 0.0038752492982894182, + 0.0067426771856844425, + 0.0047699399292469025, + 0.02192661166191101, + 0.0045328205451369286, + 0.026800036430358887, + 0.016491713002324104, + 0.010716327466070652, + 0.019460231065750122, + 0.010029423981904984, + 0.006366115994751453, + 0.026691606268286705, + 0.0073473649099469185, + 0.02678874135017395, + 0.014657684601843357, + 0.014003423042595387, + 0.02149943821132183, + 0.01965091936290264, + 0.01436899695545435, + 0.02692670188844204, + 0.017849421128630638, + 0.024385983124375343, + 0.008099764585494995, + 0.01508253812789917, + 0.017736462876200676, + 0.014848238788545132, + 0.014960803091526031, + 0.018378477543592453, + 0.017581425607204437, + 0.01103783305734396, + 0.011277450248599052, + 0.007030806969851255, + 0.01686040870845318, + 0.015979664400219917, + 0.00824819877743721, + 0.016566433012485504, + 0.00783560611307621, + 0.018735546618700027, + 0.02960052527487278, + 0.061049770563840866, + 0.03781115636229515, + 0.045996397733688354, + 0.025202522054314613, + 0.017644308507442474, + 0.03182682394981384, + 0.025334326550364494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.32905063033103943 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08911154419183731 + ], + "opposite_logprob_tokens": [ + -2.464111566543579 + ], + "logprob_reward": -0.08911154419183731, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08911154419183731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045201849192380905, + 0.024207575246691704, + 0.012414528988301754, + 0.017648844048380852, + 0.01935163512825966, + 0.061262957751750946, + 0.036213088780641556, + 0.6784527897834778, + 0.04080064594745636, + 0.06444612145423889 + ], + "credit_map": [ + 0.004756782669574022, + 0.002705648308619857, + 0.0020791147835552692, + 0.0029904344119131565, + 0.0021155276335775852, + 0.012394983321428299, + 0.0026573718059808016, + 0.015501984395086765, + 0.002081518294289708, + 0.002597772516310215, + 0.00112425338011235, + 0.001556516857817769, + 0.0011622500605881214, + 0.006338366307318211, + 0.0011914371279999614, + 0.00815545953810215, + 0.0013155508786439896, + 0.0015772696351632476, + 0.0007347528007812798, + 0.0011084098368883133, + 0.000758658570703119, + 0.002844960894435644, + 0.0008201308082789183, + 0.0032547947484999895, + 0.0030103339813649654, + 0.002149722073227167, + 0.0009856278775259852, + 0.001237246091477573, + 0.001034341985359788, + 0.004175784531980753, + 0.0007760034059174359, + 0.004279784392565489, + 0.0038875590544193983, + 0.0020532370544970036, + 0.0013677795650437474, + 0.001486600493080914, + 0.000842927023768425, + 0.0042925020679831505, + 0.0008487566956318915, + 0.004572273697704077, + 0.007014807779341936, + 0.0031794034875929356, + 0.011038071475923061, + 0.005085085518658161, + 0.002525174990296364, + 0.01397706288844347, + 0.0030326973646879196, + 0.015410654246807098, + 0.0031406651251018047, + 0.00342364190146327, + 0.004088598303496838, + 0.004097048658877611, + 0.0026760930195450783, + 0.006088968366384506, + 0.006942059379070997, + 0.005756012164056301, + 0.04983162879943848, + 0.08271332085132599, + 0.09339676797389984, + 0.07976455241441727, + 0.07596607506275177, + 0.106002077460289, + 0.10290110111236572, + 0.08787725120782852, + 0.0037426287308335304, + 0.0038695186376571655, + 0.007062648888677359, + 0.008907407522201538, + 0.003629609476774931, + 0.004491614177823067, + 0.005474111996591091, + 0.0036231058184057474, + 0.005428033880889416, + 0.006974038667976856, + 0.011198843829333782, + 0.011854348704218864, + 0.0067906505428254604, + 0.0041942717507481575, + 0.009955969639122486, + 0.00804996769875288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.08911154419183731 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004109866451472044 + ], + "opposite_logprob_tokens": [ + -5.504109859466553 + ], + "logprob_reward": -0.004109866451472044, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004109866451472044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0806315690279007, + 0.03466403856873512, + 0.016356756910681725, + 0.02155662700533867, + 0.081414595246315, + 0.0708702802658081, + 0.25011491775512695, + 0.26593753695487976, + 0.050972726196050644, + 0.1274808794260025 + ], + "credit_map": [ + 0.010823244228959084, + 0.008898711763322353, + 0.006102625746279955, + 0.011753768660128117, + 0.005859160330146551, + 0.015154481865465641, + 0.005837044212967157, + 0.016202541068196297, + 0.004644984379410744, + 0.004640008322894573, + 0.002487953519448638, + 0.005103489849716425, + 0.002497866051271558, + 0.005897232331335545, + 0.003519684076309204, + 0.005872824694961309, + 0.001645642681978643, + 0.0014662621542811394, + 0.0008490064064972103, + 0.0016775758704170585, + 0.0008434412302449346, + 0.003834307426586747, + 0.0015308846486732364, + 0.0045096380636096, + 0.0017195164691656828, + 0.0038940892554819584, + 0.0013648440362885594, + 0.0028708907775580883, + 0.0019067577086389065, + 0.004139988217502832, + 0.001536160591058433, + 0.0041243815794587135, + 0.014816450886428356, + 0.006476868875324726, + 0.003046945435926318, + 0.005966546013951302, + 0.0032896879129111767, + 0.019553715363144875, + 0.003908330108970404, + 0.0243560541421175, + 0.01020877156406641, + 0.006107065360993147, + 0.012476911768317223, + 0.007093191146850586, + 0.004538455046713352, + 0.012571646831929684, + 0.004813326522707939, + 0.013060923665761948, + 0.0148214902728796, + 0.018272629007697105, + 0.033117081969976425, + 0.027734672650694847, + 0.019706903025507927, + 0.0395636148750782, + 0.0617111437022686, + 0.03518739715218544, + 0.018589237704873085, + 0.032586149871349335, + 0.036461882293224335, + 0.03277328610420227, + 0.031450893729925156, + 0.04107983410358429, + 0.03783372417092323, + 0.035162556916475296, + 0.005831538699567318, + 0.004120360594242811, + 0.008385798893868923, + 0.008420593105256557, + 0.005064245779067278, + 0.007851279340684414, + 0.004483693744987249, + 0.00681522348895669, + 0.011747516691684723, + 0.016238661482930183, + 0.028749637305736542, + 0.029720764607191086, + 0.009004473686218262, + 0.007018453907221556, + 0.013315613381564617, + 0.011685781180858612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.004109866451472044 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04873653128743172 + ], + "opposite_logprob_tokens": [ + -3.048736572265625 + ], + "logprob_reward": -0.04873653128743172, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04873653128743172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061466775834560394, + 0.024430617690086365, + 0.032590530812740326, + 0.029880892485380173, + 0.018023917451500893, + 0.04446777328848839, + 0.11532344669103622, + 0.42110127210617065, + 0.05625411495566368, + 0.19646064937114716 + ], + "credit_map": [ + 0.005092031322419643, + 0.005116027779877186, + 0.0034176937770098448, + 0.005277286283671856, + 0.00368601456284523, + 0.015873702242970467, + 0.0045775678008794785, + 0.018426453694701195, + 0.0025843216571956873, + 0.002568004885688424, + 0.001707033021375537, + 0.0026231424417346716, + 0.0016564676770940423, + 0.005350613966584206, + 0.0021559838205575943, + 0.005785050801932812, + 0.00279722036793828, + 0.0019003795459866524, + 0.0015292391180992126, + 0.0023406920954585075, + 0.0012571882689371705, + 0.009290161542594433, + 0.0021274664904922247, + 0.011348184198141098, + 0.0028190535958856344, + 0.002131915185600519, + 0.0016436291625723243, + 0.002627921523526311, + 0.0013487652176991105, + 0.007940992712974548, + 0.001997547224164009, + 0.009371068328619003, + 0.0018612727290019393, + 0.0022232772316783667, + 0.0012356388615444303, + 0.00161883351393044, + 0.0012426275061443448, + 0.0040781088173389435, + 0.0013467716053128242, + 0.004417385905981064, + 0.005441390909254551, + 0.004217694513499737, + 0.005039315205067396, + 0.005236830562353134, + 0.0027260887436568737, + 0.008470393717288971, + 0.0029007229022681713, + 0.010435334406793118, + 0.00941439252346754, + 0.008947568014264107, + 0.014300458133220673, + 0.01201797928661108, + 0.011889428831636906, + 0.01887308433651924, + 0.021916251629590988, + 0.017964283004403114, + 0.02831817977130413, + 0.05001474544405937, + 0.055734869092702866, + 0.05176810547709465, + 0.050700485706329346, + 0.0652298852801323, + 0.06022564321756363, + 0.059109367430210114, + 0.008719325996935368, + 0.004972908645868301, + 0.009777486324310303, + 0.010629784315824509, + 0.003936595283448696, + 0.007491874508559704, + 0.005445194896310568, + 0.0052809445187449455, + 0.017334500327706337, + 0.03020915947854519, + 0.04337131232023239, + 0.046954791992902756, + 0.016267679631710052, + 0.009148023091256618, + 0.018474357202649117, + 0.014700820669531822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.04873653128743172 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, aerial close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the coin placed on a dark, highly polished obsidian surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the serpent's form peel from the coin's surface to become a three-dimensional solid gold creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the serpent's movement characterized as gliding hypnotically and silently across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single dramatic key light casting deep, long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31346702575683594 + ], + "opposite_logprob_tokens": [ + -1.313467025756836 + ], + "logprob_reward": -0.31346702575683594, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31346702575683594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06270934641361237, + 0.011891222558915615, + 0.020875459536910057, + 0.027224678546190262, + 0.01940380409359932, + 0.03628069534897804, + 0.03568137809634209, + 0.5795251727104187, + 0.04110625013709068, + 0.16530193388462067 + ], + "credit_map": [ + 0.008645295165479183, + 0.006025418639183044, + 0.003196339588612318, + 0.004263411276042461, + 0.0027766707353293896, + 0.015637705102562904, + 0.0028518771287053823, + 0.019312627613544464, + 0.001432838966138661, + 0.001562474761158228, + 0.000865858281031251, + 0.0011701490730047226, + 0.0009771366603672504, + 0.002479810733348131, + 0.001034390996210277, + 0.0023685619235038757, + 0.0030932894442230463, + 0.0011043490376323462, + 0.0012508698273450136, + 0.0012239095522090793, + 0.0007870460976846516, + 0.005793169140815735, + 0.0016085439128801227, + 0.006014282815158367, + 0.002654116367921233, + 0.0017058325465768576, + 0.0012642790097743273, + 0.001646185526624322, + 0.00098119443282485, + 0.007871597073972225, + 0.001485384302213788, + 0.00961608812212944, + 0.0021608166862279177, + 0.0022318188566714525, + 0.0010150569723919034, + 0.0019709905609488487, + 0.0012500095181167126, + 0.0043539912439882755, + 0.0012264936231076717, + 0.005194626748561859, + 0.004220894537866116, + 0.0022949669510126114, + 0.0047602481208741665, + 0.00310528720729053, + 0.0017572547076269984, + 0.00886526145040989, + 0.001836117822676897, + 0.009440666064620018, + 0.00430882116779685, + 0.0034874738194048405, + 0.003156385151669383, + 0.0027313835453242064, + 0.0034748902544379234, + 0.007123039569705725, + 0.004823274910449982, + 0.0065761106088757515, + 0.04135620966553688, + 0.06905350834131241, + 0.07774080336093903, + 0.06836725026369095, + 0.06996843963861465, + 0.09032781422138214, + 0.08612427860498428, + 0.07658687978982925, + 0.005021275021135807, + 0.003865914884954691, + 0.006811977829784155, + 0.0055519281886518, + 0.0035331358667463064, + 0.0075216395780444145, + 0.0033032328356057405, + 0.005497142672538757, + 0.012819778174161911, + 0.02114168368279934, + 0.028501620516180992, + 0.02863290160894394, + 0.019565075635910034, + 0.011271022260189056, + 0.022791599854826927, + 0.020578259602189064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.31346702575683594 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_7/index_tensor([6840, 6840])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_7/index_tensor([6840, 6840])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9461628b576b88f5866b9280e6c934fd337f5882 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_7/index_tensor([6840, 6840])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic tracking shot, captured in extreme slow motion using a telephoto lens, closely follows an advanced fighter jet as it performs a dramatic evasive maneuver across the abyssal sky. The fighter boasts a sharp, stealthy design, its fuselage clad in a matte, dark gray paint job, its metallic surfaces reflecting a cold sheen in the glaring, high-altitude sunlight. To escape the pursuing enemy aircraft, it executes a rapid barrel roll, its fuselage twisting violently through the air. The slow-motion gaze unravels this violent yet graceful maneuver into a poetic moment: as the wings slice through the air at an extreme angle, the immense pressure differential at their tips and leading edge extensions instantly erupts into thick, swirling vortices of white vapor. These vortices stretch backwards like silk, swirling, and dissipating, their complex internal turbulent structures clearly visible in the camera, creating a striking contrast against the near-black, deep blue sky in the background. In the distance, the equally advanced but darker enemy aircraft is compressed in the field of view by the telephoto lens, trying to replicate the same maneuver, its outline looming behind the gorgeous steam trail left by the protagonist's fighter plane, creating an inescapable, suffocating sense of tense pursuit and a life-and-death confrontation atmosphere.", + "seed": 16366, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_7/index_6840_seed_16366_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.18053245339542628, + "vlm_logprob": -0.18053245339542628, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.18053245339542628, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.061546120792627335 + ], + "opposite_logprob_tokens": [ + -2.9365460872650146 + ], + "logprob_reward": -0.061546120792627335, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.061546120792627335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17933079600334167, + 0.46711963415145874, + 0.06916332989931107, + 0.0755540281534195, + 0.03301680088043213, + 0.037925515323877335, + 0.053166959434747696, + 0.01973610557615757, + 0.020310968160629272, + 0.04467584565281868 + ], + "credit_map": [ + 0.017998140305280685, + 0.017266497015953064, + 0.011435265652835369, + 0.01367612648755312, + 0.014811594970524311, + 0.043556783348321915, + 0.015847448259592056, + 0.044738952070474625, + 0.018137970939278603, + 0.009080597199499607, + 0.02733236737549305, + 0.014197912998497486, + 0.011309506371617317, + 0.18121109902858734, + 0.019990747794508934, + 0.18585947155952454, + 0.00914717186242342, + 0.004668024368584156, + 0.0071487706154584885, + 0.005037560127675533, + 0.005290224216878414, + 0.016042232513427734, + 0.007301374804228544, + 0.014527976512908936, + 0.011232328601181507, + 0.007698160130530596, + 0.005036786198616028, + 0.007057544309645891, + 0.010314829647541046, + 0.011851768009364605, + 0.013426719233393669, + 0.008935892954468727, + 0.0023300175089389086, + 0.004325950983911753, + 0.002838441636413336, + 0.003945629578083754, + 0.005614440888166428, + 0.004243090748786926, + 0.005226468201726675, + 0.004492763429880142, + 0.003210980910807848, + 0.002322976943105459, + 0.0034117649774998426, + 0.0037085614167153835, + 0.003304972080513835, + 0.010112856514751911, + 0.006426980718970299, + 0.005426423158496618, + 0.004370488226413727, + 0.004866313189268112, + 0.004491055384278297, + 0.012309400364756584, + 0.00503949262201786, + 0.0065761893056333065, + 0.008856706321239471, + 0.006657320540398359, + 0.0011987832840532064, + 0.0011552514042705297, + 0.0024003570433706045, + 0.00514458492398262, + 0.001643071766011417, + 0.002645030152052641, + 0.003138716332614422, + 0.002410312881693244, + 0.0015997026348486543, + 0.0012417301768437028, + 0.002580990782007575, + 0.0027705258689820766, + 0.002083108527585864, + 0.004569250158965588, + 0.0028613856993615627, + 0.002604275941848755, + 0.006155730225145817, + 0.005666159093379974, + 0.004566944669932127, + 0.005346593447029591, + 0.0035377731546759605, + 0.0072949607856571674, + 0.005297230556607246, + 0.006810459308326244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.061546120792627335 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549742691218853 + ], + "opposite_logprob_tokens": [ + -2.929974317550659 + ], + "logprob_reward": -0.0549742691218853, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549742691218853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0630420669913292, + 0.5320827960968018, + 0.07126817107200623, + 0.012996547855436802, + 0.03040846437215805, + 0.0419442318379879, + 0.057829514145851135, + 0.03309758007526398, + 0.05765239894390106, + 0.09967831522226334 + ], + "credit_map": [ + 0.008382968604564667, + 0.0070444317534565926, + 0.005074744112789631, + 0.00643813144415617, + 0.00619728397578001, + 0.010628417134284973, + 0.00812724232673645, + 0.011148843914270401, + 0.021550139412283897, + 0.010618800297379494, + 0.032626111060380936, + 0.0163371954113245, + 0.012833620421588421, + 0.2035219818353653, + 0.02573118358850479, + 0.20886372029781342, + 0.007495834492146969, + 0.0067304084077477455, + 0.009053844027221203, + 0.006000886671245098, + 0.007645634934306145, + 0.011686775833368301, + 0.010775480419397354, + 0.011879300698637962, + 0.0021062500309199095, + 0.001308852108195424, + 0.0010135506745427847, + 0.0014220084995031357, + 0.0018733256729319692, + 0.0014134522061794996, + 0.0021776759531348944, + 0.0016814324771985412, + 0.0023580535780638456, + 0.0037498795427381992, + 0.0024180884938687086, + 0.0024737264029681683, + 0.004194758366793394, + 0.006564739625900984, + 0.004561158362776041, + 0.0040880595333874226, + 0.00376576348207891, + 0.0030408655293285847, + 0.00397918839007616, + 0.004101505037397146, + 0.003222609870135784, + 0.009921140968799591, + 0.0071898456662893295, + 0.006723312195390463, + 0.004404623061418533, + 0.005251827649772167, + 0.004634074866771698, + 0.012722726911306381, + 0.005129185039550066, + 0.008241516537964344, + 0.010206587612628937, + 0.007238971535116434, + 0.0029795567970722914, + 0.0018631034763529897, + 0.005540951620787382, + 0.004797541536390781, + 0.003520622616633773, + 0.005463919602334499, + 0.004901075270026922, + 0.004030808340758085, + 0.0045583611354231834, + 0.004501156508922577, + 0.007668800186365843, + 0.006382632534950972, + 0.006311063189059496, + 0.011281485669314861, + 0.00934181921184063, + 0.007607075851410627, + 0.016666073352098465, + 0.011504510417580605, + 0.009697986766695976, + 0.012006646953523159, + 0.006883678957819939, + 0.011472142301499844, + 0.01412794180214405, + 0.017319340258836746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.0549742691218853 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486050069332123 + ], + "opposite_logprob_tokens": [ + -1.2236050367355347 + ], + "logprob_reward": -0.3486050069332123, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486050069332123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3072899580001831, + 0.0778801366686821, + 0.16547591984272003, + 0.04517152160406113, + 0.04590669274330139, + 0.0627983808517456, + 0.06759945303201675, + 0.0399707667529583, + 0.09052523225545883, + 0.09738197177648544 + ], + "credit_map": [ + 0.01754380576312542, + 0.016312913969159126, + 0.021442333236336708, + 0.015412633307278156, + 0.017612727358937263, + 0.0967283695936203, + 0.017843229696154594, + 0.10439394414424896, + 0.00858731847256422, + 0.004814028274267912, + 0.01152185071259737, + 0.006333940662443638, + 0.005027695093303919, + 0.016147851943969727, + 0.009578385390341282, + 0.015869058668613434, + 0.017462225630879402, + 0.022141871973872185, + 0.01321379467844963, + 0.011988154612481594, + 0.021428164094686508, + 0.02795841544866562, + 0.024351350963115692, + 0.026931921020150185, + 0.004232760984450579, + 0.0044604879803955555, + 0.0038576771039515734, + 0.003784838831052184, + 0.00547432154417038, + 0.00835050456225872, + 0.008031150326132774, + 0.006979773286730051, + 0.004615071229636669, + 0.0054000806994736195, + 0.0033273648004978895, + 0.004425985738635063, + 0.00837444793432951, + 0.007103016600012779, + 0.006423944607377052, + 0.006236778572201729, + 0.0052380734123289585, + 0.0057983691804111, + 0.005825582426041365, + 0.005803917534649372, + 0.0048011792823672295, + 0.016585323959589005, + 0.01036122441291809, + 0.008384699001908302, + 0.0058191874995827675, + 0.005843742284923792, + 0.006323486100882292, + 0.011845266446471214, + 0.0056512076407670975, + 0.016718832775950432, + 0.007514055352658033, + 0.007883667014539242, + 0.0034808057826012373, + 0.003433294827118516, + 0.006107992026954889, + 0.005887908395379782, + 0.004499807488173246, + 0.00624882522970438, + 0.004764781333506107, + 0.0055473484098911285, + 0.007949786260724068, + 0.007204770110547543, + 0.01159679889678955, + 0.007555584888905287, + 0.008963275700807571, + 0.02150196209549904, + 0.013102886267006397, + 0.012650156393647194, + 0.010070654563605785, + 0.009364590048789978, + 0.0075631155632436275, + 0.008669283241033554, + 0.012673001736402512, + 0.011599987745285034, + 0.01748226396739483, + 0.019959062337875366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.3486050069332123 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635913021862507 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.008635913021862507, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635913021862507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045008912682533264, + 0.6116523742675781, + 0.054391659796237946, + 0.02109103836119175, + 0.021579619497060776, + 0.04606739431619644, + 0.028083957731723785, + 0.07222849130630493, + 0.04732346907258034, + 0.05257301405072212 + ], + "credit_map": [ + 0.0049847690388560295, + 0.0037916230503469706, + 0.004985786508768797, + 0.0033095309045165777, + 0.00416954979300499, + 0.01023136917501688, + 0.0033556444104760885, + 0.010180644690990448, + 0.02643071860074997, + 0.010380599647760391, + 0.04423448443412781, + 0.019361164420843124, + 0.01462159026414156, + 0.23091121017932892, + 0.029038194566965103, + 0.2366744875907898, + 0.00840018317103386, + 0.005320459604263306, + 0.005276332143694162, + 0.00414910027757287, + 0.004532833117991686, + 0.011290617287158966, + 0.007445800583809614, + 0.00797633919864893, + 0.0029332563281059265, + 0.0018426825990900397, + 0.0025955652818083763, + 0.003367028897628188, + 0.0018521383171901107, + 0.002996966941282153, + 0.0024898292031139135, + 0.0030135721899569035, + 0.002060726284980774, + 0.0035197357647120953, + 0.0026013138704001904, + 0.003234819509088993, + 0.0020140039268881083, + 0.0021883987355977297, + 0.0028877933509647846, + 0.003072831081226468, + 0.003972190897911787, + 0.0035911020822823048, + 0.004345950670540333, + 0.0053118630312383175, + 0.003672477090731263, + 0.01139031257480383, + 0.007877813652157784, + 0.0059056878089904785, + 0.002680324949324131, + 0.0021401681005954742, + 0.0030755565967410803, + 0.004578107502311468, + 0.002039028564468026, + 0.006022160407155752, + 0.0039221118204295635, + 0.0036265018861740828, + 0.008596004918217659, + 0.006578699219971895, + 0.01117632631212473, + 0.009826935827732086, + 0.007586657535284758, + 0.011453251354396343, + 0.007984092459082603, + 0.009026531130075455, + 0.004129838664084673, + 0.003838697914034128, + 0.005568235646933317, + 0.0052958400920033455, + 0.00435212068259716, + 0.010578393004834652, + 0.006655744276940823, + 0.0069046043790876865, + 0.008821330033242702, + 0.00697452062740922, + 0.006663383916020393, + 0.005970311816781759, + 0.0046548559330403805, + 0.00723867304623127, + 0.005843163933604956, + 0.006406776607036591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 2 + ], + "selected_reward": -0.008635913021862507 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42890095710754395 + ], + "opposite_logprob_tokens": [ + -1.053900957107544 + ], + "logprob_reward": -0.42890095710754395, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42890095710754395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06654932349920273, + 0.2579548954963684, + 0.04603251814842224, + 0.06858991831541061, + 0.09101279824972153, + 0.19582465291023254, + 0.03247610479593277, + 0.04442251846194267, + 0.05485916882753372, + 0.14227807521820068 + ], + "credit_map": [ + 0.007818056270480156, + 0.0073325298726558685, + 0.003777444362640381, + 0.004910770338028669, + 0.004311500582844019, + 0.015702733770012856, + 0.0057867527939379215, + 0.016909530386328697, + 0.014590354636311531, + 0.006007146555930376, + 0.026925701647996902, + 0.011413611471652985, + 0.007722913753241301, + 0.08632344752550125, + 0.01622157171368599, + 0.08875014632940292, + 0.0052647278644144535, + 0.004574519116431475, + 0.005820096470415592, + 0.0036222203634679317, + 0.005135911516845226, + 0.008357667364180088, + 0.006843212526291609, + 0.006414160598069429, + 0.010297417640686035, + 0.005846396554261446, + 0.005254508461803198, + 0.0055898381397128105, + 0.008394801057875156, + 0.011782354675233364, + 0.011811556294560432, + 0.009613045491278172, + 0.010055305436253548, + 0.0070611461997032166, + 0.006503957323729992, + 0.008166676387190819, + 0.019954271614551544, + 0.012729700654745102, + 0.013404816389083862, + 0.01313692145049572, + 0.016967248171567917, + 0.012858489528298378, + 0.017800142988562584, + 0.018184075132012367, + 0.017603296786546707, + 0.05210253968834877, + 0.03157465159893036, + 0.028734199702739716, + 0.0031464050989598036, + 0.0022803640458732843, + 0.0032164198346436024, + 0.006020928267389536, + 0.002707401756197214, + 0.007872609421610832, + 0.003719446947798133, + 0.003512525698170066, + 0.0035953803453594446, + 0.002587187336757779, + 0.0064121028408408165, + 0.007395644672214985, + 0.004239938221871853, + 0.00802325177937746, + 0.006370136048644781, + 0.005798880010843277, + 0.00406351825222373, + 0.0032292171381413937, + 0.007477344945073128, + 0.008297206833958626, + 0.0061793252825737, + 0.009062517434358597, + 0.009380669333040714, + 0.00716937193647027, + 0.0245926883071661, + 0.017164016142487526, + 0.01124819926917553, + 0.01654137298464775, + 0.009702681563794613, + 0.01495391596108675, + 0.01954101212322712, + 0.02853417582809925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.42890095710754395 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, captured in extreme slow motion using a telephoto lens, closely follows an advanced fighter jet as it performs a dramatic evasive maneuver across the abyssal sky. The fighter boasts a sharp, stealthy design, its fuselage clad in a matte, dark gray paint job, its metallic surfaces reflecting a cold sheen in the glaring, high-altitude sunlight. To escape the pursuing enemy aircraft, it executes a rapid barrel roll, its fuselage twisting violently through the air. The slow-motion gaze unravels this violent yet graceful maneuver into a poetic moment: as the wings slice through the air at an extreme angle, the immense pressure differential at their tips and leading edge extensions instantly erupts into thick, swirling vortices of white vapor. These vortices stretch backwards like silk, swirling, and dissipating, their complex internal turbulent structures clearly visible in the camera, creating a striking contrast against the near-black, deep blue sky in the background. In the distance, the equally advanced but darker enemy aircraft is compressed in the field of view by the telephoto lens, trying to replicate the same maneuver, its outline looming behind the gorgeous steam trail left by the protagonist's fighter plane, creating an inescapable, suffocating sense of tense pursuit and a life-and-death confrontation atmosphere.", + "seed": 16367, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_7/index_6840_seed_16367_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.22671059295535087, + "vlm_logprob": -0.22671059295535087, + "vlm_yes_no_margin": 2.549999988079071, + "vlm_token_credit": -0.22671059295535087, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.036140911281108856 + ], + "opposite_logprob_tokens": [ + -3.4111409187316895 + ], + "logprob_reward": -0.036140911281108856, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.036140911281108856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10138820111751556, + 0.07897824048995972, + 0.03812882676720619, + 0.14405816793441772, + 0.08718998730182648, + 0.07859309762716293, + 0.2950577735900879, + 0.069400854408741, + 0.06720029562711716, + 0.040004655718803406 + ], + "credit_map": [ + 0.011468732729554176, + 0.01079679187387228, + 0.015364495106041431, + 0.02350570634007454, + 0.009060104377567768, + 0.009078259579837322, + 0.009297423996031284, + 0.012816683389246464, + 0.005495711695402861, + 0.010304411873221397, + 0.010865546762943268, + 0.014725936576724052, + 0.007235596887767315, + 0.004749772604554892, + 0.014512486755847931, + 0.011088775470852852, + 0.004428292624652386, + 0.003797717159613967, + 0.005829120520502329, + 0.005811607930809259, + 0.0044782874174416065, + 0.0030793033074587584, + 0.004321914631873369, + 0.006382585037499666, + 0.013122106902301311, + 0.010443950071930885, + 0.0277967918664217, + 0.01023661158978939, + 0.012862087227404118, + 0.015482615679502487, + 0.030325498431921005, + 0.023788491263985634, + 0.008217918686568737, + 0.008630174212157726, + 0.011503675021231174, + 0.010760514996945858, + 0.009560991078615189, + 0.014224744401872158, + 0.012752766720950603, + 0.011539206840097904, + 0.0068061091005802155, + 0.00844045914709568, + 0.007669117767363787, + 0.007680543698370457, + 0.012575969099998474, + 0.010491007938981056, + 0.011439059861004353, + 0.013490833342075348, + 0.011718207970261574, + 0.008635442703962326, + 0.011994007974863052, + 0.011206885799765587, + 0.009455086663365364, + 0.09368319064378738, + 0.11992853879928589, + 0.028436386957764626, + 0.0036658819299191236, + 0.004132326226681471, + 0.005824074614793062, + 0.0036704670637845993, + 0.004705825354903936, + 0.01636713370680809, + 0.008120755665004253, + 0.02291438728570938, + 0.0031991491559892893, + 0.003961795940995216, + 0.0038379160687327385, + 0.0037494904827326536, + 0.007291505113244057, + 0.018178682774305344, + 0.006857034284621477, + 0.020124712958931923, + 0.0036623964551836252, + 0.004953071009367704, + 0.003092014929279685, + 0.005198024678975344, + 0.006347756367176771, + 0.005301907658576965, + 0.00663315923884511, + 0.004816326778382063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.036140911281108856 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980954572558403 + ], + "opposite_logprob_tokens": [ + -3.5298094749450684 + ], + "logprob_reward": -0.02980954572558403, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980954572558403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17375066876411438, + 0.1172439306974411, + 0.05524037405848503, + 0.10209561139345169, + 0.046755462884902954, + 0.035673197358846664, + 0.2307681292295456, + 0.0918869599699974, + 0.07554476708173752, + 0.07104090601205826 + ], + "credit_map": [ + 0.015721766278147697, + 0.01856289990246296, + 0.025227630510926247, + 0.027699245139956474, + 0.017595617100596428, + 0.018963098526000977, + 0.021151885390281677, + 0.028828511014580727, + 0.01079915277659893, + 0.015964176505804062, + 0.018496518954634666, + 0.024962782859802246, + 0.014902545139193535, + 0.007859968580305576, + 0.01237600389868021, + 0.011882777325809002, + 0.006101286504417658, + 0.007080284878611565, + 0.00742593128234148, + 0.010123752057552338, + 0.0066428217105567455, + 0.0032817325554788113, + 0.005555748473852873, + 0.009028819389641285, + 0.012058809399604797, + 0.007942093536257744, + 0.02162441797554493, + 0.008983039297163486, + 0.008382980711758137, + 0.009207203052937984, + 0.0189091544598341, + 0.014987921342253685, + 0.0035395892336964607, + 0.004100576974451542, + 0.0051134503446519375, + 0.005728790536522865, + 0.005193623714148998, + 0.009695624932646751, + 0.006918539758771658, + 0.006465262267738581, + 0.0034515142906457186, + 0.0034161447547376156, + 0.003308689221739769, + 0.004554702900350094, + 0.004753708839416504, + 0.004910480231046677, + 0.005257542710751295, + 0.006020415108650923, + 0.006674113683402538, + 0.00618741987273097, + 0.008599924854934216, + 0.008780321106314659, + 0.00744424806907773, + 0.08515988290309906, + 0.09440255165100098, + 0.01351967267692089, + 0.0031223732512444258, + 0.003536623204126954, + 0.005800293292850256, + 0.003639249363914132, + 0.004922322928905487, + 0.02963336370885372, + 0.00666600139811635, + 0.034566730260849, + 0.0036903878208249807, + 0.005388215649873018, + 0.0042365738190710545, + 0.005466837901622057, + 0.007503555156290531, + 0.019755057990550995, + 0.009143210016191006, + 0.020360929891467094, + 0.005320768337696791, + 0.006995849311351776, + 0.004925583954900503, + 0.00852883793413639, + 0.0066316076554358006, + 0.014715725556015968, + 0.010006468743085861, + 0.013916059397161007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.02980954572558403 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42885857820510864 + ], + "opposite_logprob_tokens": [ + -1.0538585186004639 + ], + "logprob_reward": -0.42885857820510864, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.42885857820510864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08629725873470306, + 0.07491761445999146, + 0.05238520726561546, + 0.23604755103588104, + 0.04926833137869835, + 0.09075956791639328, + 0.26358917355537415, + 0.057933975011110306, + 0.049268610775470734, + 0.03953269496560097 + ], + "credit_map": [ + 0.006819040514528751, + 0.00865746010094881, + 0.011949182488024235, + 0.02467593550682068, + 0.007093462161719799, + 0.008525196462869644, + 0.009395914152264595, + 0.009181070141494274, + 0.007583282422274351, + 0.009066677652299404, + 0.011301927268505096, + 0.016252517700195312, + 0.007365208584815264, + 0.004925121553242207, + 0.009984676726162434, + 0.008438215591013432, + 0.006611934397369623, + 0.00646272674202919, + 0.00627608597278595, + 0.009302393533289433, + 0.006795650348067284, + 0.0031852859538048506, + 0.006623292341828346, + 0.007127838674932718, + 0.028315067291259766, + 0.021046418696641922, + 0.04020094498991966, + 0.016272570937871933, + 0.03164477273821831, + 0.03214011341333389, + 0.03396882116794586, + 0.03245885670185089, + 0.004750761203467846, + 0.004741406533867121, + 0.004699558950960636, + 0.004700265824794769, + 0.004355140030384064, + 0.011885995976626873, + 0.007643997203558683, + 0.006491207517683506, + 0.008060568012297153, + 0.011857343837618828, + 0.008241371251642704, + 0.009668841026723385, + 0.015544487163424492, + 0.01334122009575367, + 0.011560216546058655, + 0.012485520914196968, + 0.01575874537229538, + 0.011107198894023895, + 0.015431865118443966, + 0.015531226992607117, + 0.010338901542127132, + 0.08857057988643646, + 0.09044665843248367, + 0.01640401780605316, + 0.005268508102744818, + 0.004690621513873339, + 0.0072084148414433, + 0.005759113002568483, + 0.006330482661724091, + 0.009120232425630093, + 0.00794695783406496, + 0.011609647423028946, + 0.002409035572782159, + 0.00330826872959733, + 0.0035656020045280457, + 0.003807928180322051, + 0.004497859161347151, + 0.012510285712778568, + 0.005420952569693327, + 0.013748683966696262, + 0.002775979693979025, + 0.0028527157846838236, + 0.001771171810105443, + 0.0029420878272503614, + 0.004902353510260582, + 0.00920741818845272, + 0.004943840205669403, + 0.010137132368981838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.42885857820510864 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945615470409393 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.005945615470409393, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945615470409393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08888551592826843, + 0.035943273454904556, + 0.04546261578798294, + 0.1233392059803009, + 0.10687979310750961, + 0.12895692884922028, + 0.3338254690170288, + 0.05956710875034332, + 0.040393806993961334, + 0.03674623742699623 + ], + "credit_map": [ + 0.008443750441074371, + 0.010623657144606113, + 0.010824259370565414, + 0.01753232255578041, + 0.007938656024634838, + 0.010850708931684494, + 0.010191363282501698, + 0.012480797246098518, + 0.004480356816202402, + 0.003539217635989189, + 0.0052132802084088326, + 0.007604724261909723, + 0.00322327739559114, + 0.002611279720440507, + 0.004875443410128355, + 0.004395694006234407, + 0.005042715463787317, + 0.004092835355550051, + 0.006806212943047285, + 0.00742898928001523, + 0.005151297897100449, + 0.0030846549198031425, + 0.004466433078050613, + 0.009389475919306278, + 0.009778102859854698, + 0.011746225878596306, + 0.020378857851028442, + 0.009224257431924343, + 0.023096196353435516, + 0.02017175406217575, + 0.013307023793458939, + 0.01563679240643978, + 0.010140090249478817, + 0.012398282997310162, + 0.009743432514369488, + 0.008975393138825893, + 0.013371803797781467, + 0.01932002231478691, + 0.015377091243863106, + 0.017553675919771194, + 0.008485903032124043, + 0.011287892237305641, + 0.007276879623532295, + 0.011966980993747711, + 0.013397314585745335, + 0.015174929983913898, + 0.027453439310193062, + 0.033913590013980865, + 0.0099672582000494, + 0.005297049414366484, + 0.006754553876817226, + 0.00798250176012516, + 0.008343510329723358, + 0.07701700925827026, + 0.1475224792957306, + 0.07094112038612366, + 0.0029168978799134493, + 0.003639147849753499, + 0.0044741458259522915, + 0.004847502335906029, + 0.005074844695627689, + 0.016707759350538254, + 0.005559105891734362, + 0.01634771004319191, + 0.0015562450280413032, + 0.0015579055761918426, + 0.0016968823038041592, + 0.0020949197933077812, + 0.005154321435838938, + 0.011698051355779171, + 0.003892121370881796, + 0.012743360362946987, + 0.0027078476268798113, + 0.003194460878148675, + 0.002018262865021825, + 0.0032800266053527594, + 0.004726917948573828, + 0.008064093999564648, + 0.004214804619550705, + 0.00853982288390398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.005945615470409393 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327983140945435 + ], + "opposite_logprob_tokens": [ + -0.7577983140945435 + ], + "logprob_reward": -0.6327983140945435, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327983140945435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21614567935466766, + 0.06374882161617279, + 0.033530764281749725, + 0.06822793185710907, + 0.04326188936829567, + 0.0362161360681057, + 0.139305979013443, + 0.054732080549001694, + 0.16474691033363342, + 0.18008378148078918 + ], + "credit_map": [ + 0.019845103845000267, + 0.02305389568209648, + 0.03447651118040085, + 0.042180586606264114, + 0.019667256623506546, + 0.022107278928160667, + 0.028444400057196617, + 0.026370644569396973, + 0.009563258849084377, + 0.007465841714292765, + 0.009949570521712303, + 0.010868106037378311, + 0.006105723325163126, + 0.004192065913230181, + 0.007506069261580706, + 0.008098183199763298, + 0.0035299398005008698, + 0.0037444441113620996, + 0.0039028034079819918, + 0.0060090129263699055, + 0.004335550591349602, + 0.003503465326502919, + 0.0037337783724069595, + 0.004771769512444735, + 0.006021962966769934, + 0.006933920551091433, + 0.010988574475049973, + 0.006254211533814669, + 0.011657465249300003, + 0.008870478719472885, + 0.009136606007814407, + 0.008364717476069927, + 0.0030165729112923145, + 0.0031009637750685215, + 0.00410831905901432, + 0.003730060998350382, + 0.0038009497802704573, + 0.00786584708839655, + 0.010144907981157303, + 0.007494264282286167, + 0.0031547441612929106, + 0.005371951963752508, + 0.00365815544500947, + 0.00388488476164639, + 0.006525680422782898, + 0.004413689486682415, + 0.0047256844118237495, + 0.004481349140405655, + 0.004201812669634819, + 0.004638069309294224, + 0.005216506309807301, + 0.007745560724288225, + 0.005205248482525349, + 0.025018153712153435, + 0.032541535794734955, + 0.0547390952706337, + 0.0029546257574111223, + 0.0027716413605958223, + 0.004066886845976114, + 0.0037471051327884197, + 0.004387973807752132, + 0.01501060277223587, + 0.004504118114709854, + 0.01728912629187107, + 0.005262208636850119, + 0.008281385526061058, + 0.0068430607207119465, + 0.007835213094949722, + 0.01073422934859991, + 0.051387157291173935, + 0.016259729862213135, + 0.05814392864704132, + 0.006435530725866556, + 0.006765630096197128, + 0.005879145115613937, + 0.00922431331127882, + 0.013330517336726189, + 0.0620533749461174, + 0.010039817541837692, + 0.06635547429323196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.6327983140945435 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, captured in extreme slow motion using a telephoto lens, closely follows an advanced fighter jet as it performs a dramatic evasive maneuver across the abyssal sky. The fighter boasts a sharp, stealthy design, its fuselage clad in a matte, dark gray paint job, its metallic surfaces reflecting a cold sheen in the glaring, high-altitude sunlight. To escape the pursuing enemy aircraft, it executes a rapid barrel roll, its fuselage twisting violently through the air. The slow-motion gaze unravels this violent yet graceful maneuver into a poetic moment: as the wings slice through the air at an extreme angle, the immense pressure differential at their tips and leading edge extensions instantly erupts into thick, swirling vortices of white vapor. These vortices stretch backwards like silk, swirling, and dissipating, their complex internal turbulent structures clearly visible in the camera, creating a striking contrast against the near-black, deep blue sky in the background. In the distance, the equally advanced but darker enemy aircraft is compressed in the field of view by the telephoto lens, trying to replicate the same maneuver, its outline looming behind the gorgeous steam trail left by the protagonist's fighter plane, creating an inescapable, suffocating sense of tense pursuit and a life-and-death confrontation atmosphere.", + "seed": 16368, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_7/index_6840_seed_16368_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1283343143761158, + "vlm_logprob": -0.1283343143761158, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.1283343143761158, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03562682867050171 + ], + "opposite_logprob_tokens": [ + -3.4106268882751465 + ], + "logprob_reward": -0.03562682867050171, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03562682867050171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04809293523430824, + 0.1770765781402588, + 0.11400994658470154, + 0.10725437104701996, + 0.03501586243510246, + 0.034622352570295334, + 0.0648922547698021, + 0.37324464321136475, + 0.02026764675974846, + 0.02552340179681778 + ], + "credit_map": [ + 0.003183352295309305, + 0.004912502598017454, + 0.005375302862375975, + 0.003988922107964754, + 0.00927184708416462, + 0.006451192777603865, + 0.00695803202688694, + 0.007951783947646618, + 0.007287080865353346, + 0.006756362039595842, + 0.007688315585255623, + 0.0054392279125750065, + 0.011014976538717747, + 0.06306271255016327, + 0.021727656945586205, + 0.05410024896264076, + 0.00474120955914259, + 0.007583970203995705, + 0.009465336799621582, + 0.005636850371956825, + 0.00883355550467968, + 0.03970899432897568, + 0.010333164595067501, + 0.027706868946552277, + 0.011636381968855858, + 0.004862360190600157, + 0.005559369921684265, + 0.005849902052432299, + 0.005531949456781149, + 0.02359711565077305, + 0.01312112994492054, + 0.03709617257118225, + 0.00648793950676918, + 0.0033873431384563446, + 0.0088930269703269, + 0.0033834539353847504, + 0.0028348700143396854, + 0.0038916110061109066, + 0.003422100329771638, + 0.0027155186980962753, + 0.008746327832341194, + 0.003370800521224737, + 0.003993229474872351, + 0.009266028180718422, + 0.0016708282055333257, + 0.002209678292274475, + 0.003004862228408456, + 0.0023606005124747753, + 0.01343463733792305, + 0.0068648201413452625, + 0.01820315606892109, + 0.0067870295606553555, + 0.005165695678442717, + 0.004873413126915693, + 0.0051397476345300674, + 0.004423759412020445, + 0.07150087505578995, + 0.1095261499285698, + 0.03546200320124626, + 0.023403633385896683, + 0.02386440336704254, + 0.05870484188199043, + 0.02151176705956459, + 0.02927098236978054, + 0.0024988707154989243, + 0.0023100189864635468, + 0.0023391798604279757, + 0.0033017913810908794, + 0.001665801857598126, + 0.002087948378175497, + 0.002174732740968466, + 0.003889302024617791, + 0.003541107987985015, + 0.004458034411072731, + 0.0030004980508238077, + 0.002880838932469487, + 0.003267379244789481, + 0.0029691557865589857, + 0.002288165269419551, + 0.0031182216480374336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 2 + ], + "selected_reward": -0.03562682867050171 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026361869648098946 + ], + "opposite_logprob_tokens": [ + -3.6513619422912598 + ], + "logprob_reward": -0.026361869648098946, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026361869648098946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.079563207924366, + 0.19906631112098694, + 0.06095949560403824, + 0.10802404582500458, + 0.03689899668097496, + 0.03528792783617973, + 0.10212501883506775, + 0.27505210041999817, + 0.060690879821777344, + 0.04233207181096077 + ], + "credit_map": [ + 0.004522194154560566, + 0.007053822744637728, + 0.010133967734873295, + 0.004229994490742683, + 0.016533179208636284, + 0.013925766572356224, + 0.011702741496264935, + 0.011461537331342697, + 0.006708909757435322, + 0.006020878441631794, + 0.015105677768588066, + 0.004215212073177099, + 0.011494796723127365, + 0.07065869867801666, + 0.024821443483233452, + 0.06004069373011589, + 0.0021547877695411444, + 0.0035655510146170855, + 0.005590758286416531, + 0.0023343651555478573, + 0.002669681329280138, + 0.022554749622941017, + 0.004559793509542942, + 0.017529811710119247, + 0.01257411390542984, + 0.004868321120738983, + 0.006468986626714468, + 0.004975531715899706, + 0.004933575633913279, + 0.02315698005259037, + 0.013047645799815655, + 0.037998881191015244, + 0.007396303582936525, + 0.0033385814167559147, + 0.010173831135034561, + 0.003879993222653866, + 0.0034684527199715376, + 0.0028004818595945835, + 0.003767558839172125, + 0.0020737922750413418, + 0.008014353923499584, + 0.0038691279478371143, + 0.005157476756721735, + 0.008221793919801712, + 0.002280442276969552, + 0.00233856332488358, + 0.0029577051755040884, + 0.002448464510962367, + 0.023026330396533012, + 0.010582794435322285, + 0.025335166603326797, + 0.009982925839722157, + 0.009247691370546818, + 0.00820966623723507, + 0.008459621109068394, + 0.007280823774635792, + 0.029163140803575516, + 0.040204208344221115, + 0.0504814051091671, + 0.025396030396223068, + 0.0269253458827734, + 0.047210026532411575, + 0.02528488263487816, + 0.03038707748055458, + 0.006339108105748892, + 0.006992000620812178, + 0.006114068441092968, + 0.010234554298222065, + 0.00596295902505517, + 0.007119101472198963, + 0.006145860999822617, + 0.011783227324485779, + 0.00548279145732522, + 0.006995357573032379, + 0.004931768402457237, + 0.0050689829513430595, + 0.005286284256726503, + 0.004781974945217371, + 0.00397522933781147, + 0.005809684284031391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.026361869648098946 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134172856807709 + ], + "opposite_logprob_tokens": [ + -1.3134173154830933 + ], + "logprob_reward": -0.3134172856807709, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134172856807709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08963031321763992, + 0.18256616592407227, + 0.08016752451658249, + 0.08578532189130783, + 0.06559064984321594, + 0.12988482415676117, + 0.1672869771718979, + 0.10360726714134216, + 0.06776883453130722, + 0.027712104842066765 + ], + "credit_map": [ + 0.005480905994772911, + 0.007170315831899643, + 0.008093740791082382, + 0.005309170112013817, + 0.013280843384563923, + 0.015074238181114197, + 0.010752245783805847, + 0.024468857795000076, + 0.006844205781817436, + 0.015506349503993988, + 0.010759868659079075, + 0.013187696225941181, + 0.01551771629601717, + 0.05036047101020813, + 0.022960945963859558, + 0.0474289171397686, + 0.0065711974166333675, + 0.008772633969783783, + 0.008712531067430973, + 0.00673528341576457, + 0.010089999996125698, + 0.017129868268966675, + 0.013870172202587128, + 0.008285842835903168, + 0.0048483796417713165, + 0.008900732733309269, + 0.011896660551428795, + 0.008222581818699837, + 0.008636641316115856, + 0.010364091955125332, + 0.016507793217897415, + 0.016408443450927734, + 0.008807378821074963, + 0.005215593613684177, + 0.014227882958948612, + 0.012825259938836098, + 0.005616511683911085, + 0.006919980514794588, + 0.006413644645363092, + 0.005564396735280752, + 0.03610103577375412, + 0.011797287501394749, + 0.009238751605153084, + 0.0444314107298851, + 0.005235325079411268, + 0.007253400981426239, + 0.008439909666776657, + 0.007387713063508272, + 0.044237177819013596, + 0.01750720664858818, + 0.053439121693372726, + 0.012980112805962563, + 0.011050160974264145, + 0.010072077624499798, + 0.009240938350558281, + 0.008760189637541771, + 0.01406188402324915, + 0.020072292536497116, + 0.021238410845398903, + 0.008128364570438862, + 0.010654731653630733, + 0.01231418177485466, + 0.007768057286739349, + 0.009369349107146263, + 0.0066057974472641945, + 0.007323554251343012, + 0.0076999906450510025, + 0.010397998616099358, + 0.007831444032490253, + 0.009710540063679218, + 0.006960264872759581, + 0.011239249259233475, + 0.004977353848516941, + 0.004541791509836912, + 0.0035708299838006496, + 0.004011652432382107, + 0.0026562861166894436, + 0.002137113595381379, + 0.0025556182954460382, + 0.003261460456997156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 5 + ], + "selected_reward": -0.3134172856807709 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194594696164131 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194594696164131, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194594696164131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03489191457629204, + 0.2009834200143814, + 0.044697269797325134, + 0.05676288902759552, + 0.03677196428179741, + 0.02335876040160656, + 0.03109559789299965, + 0.4828088581562042, + 0.06991326063871384, + 0.018716059625148773 + ], + "credit_map": [ + 0.0022613941691815853, + 0.003567939391359687, + 0.004873540252447128, + 0.0028623268008232117, + 0.006459060590714216, + 0.005151443183422089, + 0.0037043127231299877, + 0.00601189536973834, + 0.0071899984031915665, + 0.007369940634816885, + 0.010813926346600056, + 0.005684970878064632, + 0.011466052383184433, + 0.0709400475025177, + 0.02480318397283554, + 0.06271528452634811, + 0.002262672409415245, + 0.0035035491455346346, + 0.004708574619144201, + 0.002655313815921545, + 0.004551538731902838, + 0.013927717693150043, + 0.004751492291688919, + 0.00833640806376934, + 0.004974310752004385, + 0.004990553483366966, + 0.005682589486241341, + 0.0048167346976697445, + 0.005214663688093424, + 0.00899025984108448, + 0.009366847574710846, + 0.012726920656859875, + 0.007568162400275469, + 0.0034195512998849154, + 0.011881508864462376, + 0.0020205469336360693, + 0.0035258103162050247, + 0.0031752341892570257, + 0.002685968531295657, + 0.002495179418474436, + 0.0036991597153246403, + 0.00263240491040051, + 0.003789754817262292, + 0.004937829915434122, + 0.0018629278056323528, + 0.001970691606402397, + 0.0017851406009867787, + 0.0026808492839336395, + 0.004713411908596754, + 0.0039816563948988914, + 0.0067740511149168015, + 0.003700331784784794, + 0.00471737002953887, + 0.002531657926738262, + 0.002286886563524604, + 0.0023902314715087414, + 0.09642914682626724, + 0.1607106328010559, + 0.045243002474308014, + 0.026827942579984665, + 0.03679472208023071, + 0.06464478373527527, + 0.026344584301114082, + 0.025814013555645943, + 0.006662157364189625, + 0.0067970361560583115, + 0.007686472963541746, + 0.011155783198773861, + 0.006917521823197603, + 0.009502795524895191, + 0.007680736482143402, + 0.013510750606656075, + 0.0019434146815910935, + 0.0024256485048681498, + 0.0019225154537707567, + 0.002655286807566881, + 0.0024855758529156446, + 0.0023452388122677803, + 0.0021821821574121714, + 0.0027561963070183992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.014194594696164131 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25207099318504333 + ], + "opposite_logprob_tokens": [ + -1.5020710229873657 + ], + "logprob_reward": -0.25207099318504333, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25207099318504333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08831395208835602, + 0.18716874718666077, + 0.049232691526412964, + 0.047515563666820526, + 0.025792520493268967, + 0.07042552530765533, + 0.10292975604534149, + 0.38235437870025635, + 0.03129492700099945, + 0.01497191283851862 + ], + "credit_map": [ + 0.0063591450452804565, + 0.0055263168178498745, + 0.006407728884369135, + 0.0029885536059737206, + 0.01586894504725933, + 0.01392939779907465, + 0.018055399879813194, + 0.019178470596671104, + 0.007116331718862057, + 0.00423363596200943, + 0.009396396577358246, + 0.0031584263779222965, + 0.010526473633944988, + 0.06908110529184341, + 0.023891115561127663, + 0.059765271842479706, + 0.0018797832308337092, + 0.002959937322884798, + 0.004370308946818113, + 0.0023180327843874693, + 0.004444201476871967, + 0.01730411686003208, + 0.0052488986402750015, + 0.010707412846386433, + 0.004827252589166164, + 0.0026975928340107203, + 0.0034654405899345875, + 0.0027812819462269545, + 0.0028353712987154722, + 0.009454870596528053, + 0.005981066729873419, + 0.015472693368792534, + 0.006441823206841946, + 0.002503111260011792, + 0.009870811365544796, + 0.001606065547093749, + 0.0013338836142793298, + 0.001742616412229836, + 0.001244754297658801, + 0.0010494575835764408, + 0.020568184554576874, + 0.006983566097915173, + 0.006970978807657957, + 0.02286355569958687, + 0.0026348119135946035, + 0.003340372582897544, + 0.0042163310572505, + 0.002847726456820965, + 0.018130991607904434, + 0.011696499772369862, + 0.023450281471014023, + 0.012590926140546799, + 0.011309255845844746, + 0.008876379579305649, + 0.008608207106590271, + 0.008267221041023731, + 0.06702432781457901, + 0.10165631026029587, + 0.04422641173005104, + 0.025638222694396973, + 0.025695420801639557, + 0.06187950819730759, + 0.02386031486093998, + 0.03237387537956238, + 0.003139167558401823, + 0.003361620707437396, + 0.003615381894633174, + 0.005709504242986441, + 0.0030643055215477943, + 0.0037930442485958338, + 0.0029332246631383896, + 0.005678679794073105, + 0.001784641295671463, + 0.002079184865579009, + 0.0016557967755943537, + 0.0016555418260395527, + 0.0019591827876865864, + 0.0018708574352785945, + 0.0013335698749870062, + 0.002633138792589307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.25207099318504333 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, captured in extreme slow motion using a telephoto lens, closely follows an advanced fighter jet as it performs a dramatic evasive maneuver across the abyssal sky. The fighter boasts a sharp, stealthy design, its fuselage clad in a matte, dark gray paint job, its metallic surfaces reflecting a cold sheen in the glaring, high-altitude sunlight. To escape the pursuing enemy aircraft, it executes a rapid barrel roll, its fuselage twisting violently through the air. The slow-motion gaze unravels this violent yet graceful maneuver into a poetic moment: as the wings slice through the air at an extreme angle, the immense pressure differential at their tips and leading edge extensions instantly erupts into thick, swirling vortices of white vapor. These vortices stretch backwards like silk, swirling, and dissipating, their complex internal turbulent structures clearly visible in the camera, creating a striking contrast against the near-black, deep blue sky in the background. In the distance, the equally advanced but darker enemy aircraft is compressed in the field of view by the telephoto lens, trying to replicate the same maneuver, its outline looming behind the gorgeous steam trail left by the protagonist's fighter plane, creating an inescapable, suffocating sense of tense pursuit and a life-and-death confrontation atmosphere.", + "seed": 16369, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_7/index_6840_seed_16369_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16963973939418792, + "vlm_logprob": -0.16963973939418792, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.16963973939418792, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010841868817806244 + ], + "opposite_logprob_tokens": [ + -4.7608418464660645 + ], + "logprob_reward": -0.010841868817806244, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010841868817806244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043998219072818756, + 0.0626927986741066, + 0.034205082803964615, + 0.05118045583367348, + 0.6617166996002197, + 0.040388524532318115, + 0.0136833181604743, + 0.04834854602813721, + 0.029676206409931183, + 0.014110172167420387 + ], + "credit_map": [ + 0.0029762061312794685, + 0.005352336447685957, + 0.006971707567572594, + 0.006497559137642384, + 0.00864610355347395, + 0.005436765495687723, + 0.004666400142014027, + 0.0034511415287852287, + 0.006132914684712887, + 0.010219602845609188, + 0.005589884240180254, + 0.008126091212034225, + 0.007837480865418911, + 0.006801702082157135, + 0.00849776528775692, + 0.00948735885322094, + 0.0028717394452542067, + 0.0039850627072155476, + 0.003465428948402405, + 0.0035967102739959955, + 0.00626411521807313, + 0.004030867014080286, + 0.006119247060269117, + 0.003871915163472295, + 0.005541968625038862, + 0.005523327272385359, + 0.009241514839231968, + 0.006804569158703089, + 0.010233199223876, + 0.004686843603849411, + 0.0036505938041955233, + 0.005498441867530346, + 0.09121806174516678, + 0.22806666791439056, + 0.1588819921016693, + 0.035531606525182724, + 0.06869923323392868, + 0.021448832005262375, + 0.02264806441962719, + 0.03522224351763725, + 0.004849340301007032, + 0.006071213632822037, + 0.004111744463443756, + 0.005557922180742025, + 0.005519259721040726, + 0.005818977020680904, + 0.004500338342040777, + 0.003959727939218283, + 0.001339897746220231, + 0.0019920801278203726, + 0.0013122601667419076, + 0.002371619688346982, + 0.0016934452578425407, + 0.001364540308713913, + 0.0023297679144889116, + 0.0012797077652066946, + 0.00361091666854918, + 0.012738313525915146, + 0.006258667446672916, + 0.003992012236267328, + 0.011922113597393036, + 0.003494908567517996, + 0.003284377744421363, + 0.0030472353100776672, + 0.0028420668095350266, + 0.007789555471390486, + 0.002918072510510683, + 0.0022554900497198105, + 0.007964391261339188, + 0.0027547203935682774, + 0.0017532471101731062, + 0.0013986643170937896, + 0.0021275971084833145, + 0.0018889358034357429, + 0.003621740499511361, + 0.000983521924354136, + 0.002216715132817626, + 0.00123477797023952, + 0.0010068740230053663, + 0.0010300097055733204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.010841868817806244 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026368139311671257 + ], + "opposite_logprob_tokens": [ + -3.6513681411743164 + ], + "logprob_reward": -0.026368139311671257, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026368139311671257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06231473386287689, + 0.06086999550461769, + 0.04871750250458717, + 0.04868431016802788, + 0.4398401081562042, + 0.04449119046330452, + 0.040965259075164795, + 0.1533554047346115, + 0.06536885350942612, + 0.035392630845308304 + ], + "credit_map": [ + 0.0060554444789886475, + 0.007627421524375677, + 0.010626303032040596, + 0.010527629405260086, + 0.009271240793168545, + 0.007208589464426041, + 0.0059301406145095825, + 0.00506796408444643, + 0.005974878557026386, + 0.011119588278234005, + 0.005394690204411745, + 0.00788065604865551, + 0.007823100313544273, + 0.006886690389364958, + 0.007921874523162842, + 0.007868516258895397, + 0.004000402987003326, + 0.005028839223086834, + 0.00441610487177968, + 0.005809033289551735, + 0.01122298464179039, + 0.005963441450148821, + 0.007561333477497101, + 0.004715362098067999, + 0.006247132085263729, + 0.008053222671151161, + 0.004994681105017662, + 0.004486878402531147, + 0.011794958263635635, + 0.006079604849219322, + 0.0037537699099630117, + 0.003274065675213933, + 0.06724878400564194, + 0.15869835019111633, + 0.10364457964897156, + 0.0162615068256855, + 0.042413849383592606, + 0.01628410629928112, + 0.011736657470464706, + 0.02355227619409561, + 0.00630131596699357, + 0.007273513823747635, + 0.007212035357952118, + 0.004049095790833235, + 0.005513540934771299, + 0.004768668673932552, + 0.003977987449616194, + 0.005395031068474054, + 0.004530827049165964, + 0.005752540659159422, + 0.0036210801918059587, + 0.005595454014837742, + 0.005142063833773136, + 0.005181245505809784, + 0.005816333461552858, + 0.005325716454535723, + 0.00950188934803009, + 0.053425565361976624, + 0.007658506743609905, + 0.005373671650886536, + 0.05201730132102966, + 0.010896007530391216, + 0.005769963376224041, + 0.008712504990398884, + 0.0056873830035328865, + 0.015491913072764874, + 0.006938281934708357, + 0.007705372758209705, + 0.015691181644797325, + 0.005790976341813803, + 0.004890188574790955, + 0.0031735519878566265, + 0.005779505707323551, + 0.005544885527342558, + 0.008025123737752438, + 0.0029797975439578295, + 0.004193891771137714, + 0.002819095505401492, + 0.0036740403156727552, + 0.0023762902710586786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.026368139311671257 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134278357028961 + ], + "opposite_logprob_tokens": [ + -1.3134278059005737 + ], + "logprob_reward": -0.3134278357028961, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134278357028961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07557636499404907, + 0.04821687564253807, + 0.15280397236347198, + 0.15409177541732788, + 0.11385500431060791, + 0.1512942612171173, + 0.07302504032850266, + 0.1616116464138031, + 0.03063693270087242, + 0.03888809308409691 + ], + "credit_map": [ + 0.005800312850624323, + 0.007617045193910599, + 0.018207967281341553, + 0.014877969399094582, + 0.01016128621995449, + 0.006558537483215332, + 0.007444008253514767, + 0.004909234587103128, + 0.004697532393038273, + 0.007158317603170872, + 0.005234299227595329, + 0.008408993482589722, + 0.006056820508092642, + 0.005609564483165741, + 0.006752083543688059, + 0.004299264866858721, + 0.010623798705637455, + 0.019506048411130905, + 0.01651066541671753, + 0.018000498414039612, + 0.03149448335170746, + 0.016388799995183945, + 0.028704075142741203, + 0.011575601994991302, + 0.02332860417664051, + 0.03069780021905899, + 0.010635900311172009, + 0.01935712806880474, + 0.02517082542181015, + 0.01935994066298008, + 0.016772925853729248, + 0.008768654428422451, + 0.021184634417295456, + 0.022799549624323845, + 0.010215044021606445, + 0.014519569464027882, + 0.01892884261906147, + 0.009560072794556618, + 0.010629916563630104, + 0.006017373874783516, + 0.016799697652459145, + 0.02858702652156353, + 0.01477142795920372, + 0.02311977557837963, + 0.01764148660004139, + 0.01564740389585495, + 0.021592721343040466, + 0.013134720735251904, + 0.009669430553913116, + 0.010147001594305038, + 0.007212966214865446, + 0.010692381300032139, + 0.007607859559357166, + 0.005752519704401493, + 0.013649629428982735, + 0.008293256163597107, + 0.007733563426882029, + 0.05606094002723694, + 0.012850303202867508, + 0.006819911301136017, + 0.05518241971731186, + 0.009500603191554546, + 0.0069875651970505714, + 0.006476350594311953, + 0.003490195143967867, + 0.004583812318742275, + 0.004705965984612703, + 0.005037169437855482, + 0.0034165687393397093, + 0.003403391921892762, + 0.0029700950253754854, + 0.0030297357589006424, + 0.004223192110657692, + 0.006494269706308842, + 0.01048230566084385, + 0.0031064795330166817, + 0.005337639711797237, + 0.003918428439646959, + 0.0026186422910541296, + 0.0027071323711425066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.3134278357028961 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232884231954813 + ], + "opposite_logprob_tokens": [ + -3.7732884883880615 + ], + "logprob_reward": -0.0232884231954813, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232884231954813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05140330269932747, + 0.04236967861652374, + 0.06280478835105896, + 0.04750766605138779, + 0.10215296596288681, + 0.01740746945142746, + 0.012688987888395786, + 0.6365380883216858, + 0.012004636228084564, + 0.015122459270060062 + ], + "credit_map": [ + 0.004254503641277552, + 0.005391605664044619, + 0.011906317435204983, + 0.01076759397983551, + 0.0069167884066700935, + 0.004222329705953598, + 0.004123502876609564, + 0.0038206547033041716, + 0.004102340899407864, + 0.0061288862489163876, + 0.004241124726831913, + 0.005470669362694025, + 0.005670178681612015, + 0.004503468982875347, + 0.005987984128296375, + 0.006265024654567242, + 0.004683840088546276, + 0.006785994861274958, + 0.007817351259291172, + 0.008181973360478878, + 0.010259370319545269, + 0.006877516396343708, + 0.011210132390260696, + 0.006988603621721268, + 0.007252932991832495, + 0.007741670589894056, + 0.003905429970473051, + 0.005745464935898781, + 0.00791509635746479, + 0.005547435488551855, + 0.0055709718726575375, + 0.0038286582566797733, + 0.01569717936217785, + 0.036917392164468765, + 0.024935202673077583, + 0.003219589591026306, + 0.009408212266862392, + 0.003850387642160058, + 0.002212594961747527, + 0.0059124077670276165, + 0.002169886138290167, + 0.0026441607624292374, + 0.001956163439899683, + 0.002582641551271081, + 0.0024166153743863106, + 0.0022413989063352346, + 0.00154521269723773, + 0.0018513890681788325, + 0.001359799993224442, + 0.0016554526519030333, + 0.0014993890654295683, + 0.0017228879733011127, + 0.0018042700830847025, + 0.001493332558311522, + 0.001448076101951301, + 0.0017057793447747827, + 0.03319532796740532, + 0.2358112335205078, + 0.01898941770195961, + 0.016633717343211174, + 0.23192983865737915, + 0.043308597058057785, + 0.021906742826104164, + 0.0347631573677063, + 0.0012444407911971211, + 0.0015561007894575596, + 0.0020628822967410088, + 0.001915352651849389, + 0.0013396090362221003, + 0.0015268989372998476, + 0.0012866539182141423, + 0.0010726965265348554, + 0.0018037534318864346, + 0.0024699592031538486, + 0.003643843112513423, + 0.001327542937360704, + 0.00199063285253942, + 0.0013613277114927769, + 0.0013544713146984577, + 0.001170928357169032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.0232884231954813 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an advanced fighter jet with a sharp, stealthy design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fighter jet's fuselage clad in a matte, dark gray paint job with a cold metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jet execute a rapid barrel roll maneuver while twisting violently through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thick, swirling white vapor vortices erupt from the wingtips and leading edge extensions during the roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a near-black, deep blue sky with a distant, darker enemy aircraft visible behind the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742724299430847 + ], + "opposite_logprob_tokens": [ + -0.9742724299430847 + ], + "logprob_reward": -0.4742724299430847, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742724299430847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06798113137483597, + 0.02563282661139965, + 0.029495038092136383, + 0.030506005510687828, + 0.35276657342910767, + 0.02073485217988491, + 0.05208113417029381, + 0.28176701068878174, + 0.10364818572998047, + 0.0353873036801815 + ], + "credit_map": [ + 0.0050850496627390385, + 0.006706611253321171, + 0.01129795704036951, + 0.01116273645311594, + 0.008192277513444424, + 0.007937214337289333, + 0.01036236621439457, + 0.007236917968839407, + 0.0019185509299859405, + 0.004623092245310545, + 0.0028792587108910084, + 0.0039534298703074455, + 0.0036013941280543804, + 0.003273026319220662, + 0.0032319503370672464, + 0.002152123721316457, + 0.0019250470213592052, + 0.003923492506146431, + 0.003384009934961796, + 0.003234260715544224, + 0.0056944298557937145, + 0.003116610925644636, + 0.005664929747581482, + 0.0025522559881210327, + 0.0050116851925849915, + 0.00243716174736619, + 0.005790014751255512, + 0.0029497111681848764, + 0.0032363743521273136, + 0.0035062870010733604, + 0.0021967473439872265, + 0.005378024186939001, + 0.05586285516619682, + 0.12925979495048523, + 0.08375269174575806, + 0.00982577633112669, + 0.03203260153532028, + 0.01439187116920948, + 0.007579575292766094, + 0.020061401650309563, + 0.002371290698647499, + 0.003390253521502018, + 0.003015367081388831, + 0.0025664442218840122, + 0.002563325222581625, + 0.002305111149325967, + 0.0023833047598600388, + 0.002139754593372345, + 0.005931531079113483, + 0.007379033137112856, + 0.006243961397558451, + 0.006185643840581179, + 0.005610186140984297, + 0.008332038298249245, + 0.007057514972984791, + 0.005341220647096634, + 0.013914448209106922, + 0.10477760434150696, + 0.01023826189339161, + 0.007510110270231962, + 0.1023012101650238, + 0.018848346546292305, + 0.009379327297210693, + 0.014797691255807877, + 0.006311813369393349, + 0.028549281880259514, + 0.015186365693807602, + 0.00993454921990633, + 0.026847153902053833, + 0.007121108938008547, + 0.004837255459278822, + 0.004860656801611185, + 0.006130356341600418, + 0.005053194239735603, + 0.007421994581818581, + 0.0034637784119695425, + 0.004516021814197302, + 0.003851137589663267, + 0.002397754229605198, + 0.002553069032728672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.4742724299430847 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_8/index_tensor([4854, 4854])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_8/index_tensor([4854, 4854])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e9c4e3f5f10486d61c1fc5b652c207080afa1c7d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_8/index_tensor([4854, 4854])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot with the breathtaking detail of high-speed macro cinematography, this full-body, long shot tracks horizontally with a massive northern pike as it lurks motionless in murky, murky water. The water is a thick suspension of silt and algae, pierced by faint rays of green light, illuminating floating particles. The pike, a formidable predator, hangs suspended in the middle of the water, its long, torpedo-shaped body a detailed mosaic of olive-green and tawny scales, speckled with rows of pale, bean-shaped spots. Its fins, tinged with dark red, make imperceptible adjustments to maintain its position close to the rotting, waterlogged branches at the bottom of the silt. The scene is held in tense silence, and then, with explosive movements rendered in extreme slow motion, the pike strikes. Its powerful tail whips violently, creating a swirling vortex that kicks up clouds of sediment as its entire form surges forward. The camera continues to track perfectly, capturing how each muscle group contracts in visible, continuous waves beneath the skin and how the water displaces around its duckbill snout in a perfectly articulated hydrodynamic shell, transforming quiet tension into a breathtaking display of raw, aquatic power.", + "seed": 30457, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_8/index_4854_seed_30457_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15089117363095284, + "vlm_logprob": -0.15089117363095284, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.15089117363095284, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016342246904969215 + ], + "opposite_logprob_tokens": [ + -4.5163421630859375 + ], + "logprob_reward": -0.016342246904969215, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.016342246904969215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12222328782081604, + 0.14812801778316498, + 0.033693380653858185, + 0.05423608049750328, + 0.05447336286306381, + 0.04495224356651306, + 0.033682819455862045, + 0.02969193272292614, + 0.40238597989082336, + 0.07653284817934036 + ], + "credit_map": [ + 0.011834298260509968, + 0.01650882139801979, + 0.017593178898096085, + 0.009544722735881805, + 0.022376300767064095, + 0.011449623852968216, + 0.010114237666130066, + 0.02280210144817829, + 0.014792731031775475, + 0.02483191527426243, + 0.014170992188155651, + 0.02511557750403881, + 0.02769867517054081, + 0.012475921772420406, + 0.013274366967380047, + 0.015767842531204224, + 0.004282178822904825, + 0.003943122457712889, + 0.0038439929485321045, + 0.006576676853001118, + 0.004365901462733746, + 0.0034050450194627047, + 0.002854849910363555, + 0.0044216131791472435, + 0.005298334173858166, + 0.010342985391616821, + 0.004598522093147039, + 0.008242080919444561, + 0.007291879504919052, + 0.0055964430794119835, + 0.007538269739598036, + 0.005327562801539898, + 0.004845898598432541, + 0.006103848572820425, + 0.0064287069253623486, + 0.006147021893411875, + 0.011535598896443844, + 0.007561372127383947, + 0.006043657194823027, + 0.005807255394756794, + 0.006820456590503454, + 0.0063111623749136925, + 0.0043424624018371105, + 0.006343931891024113, + 0.006364986300468445, + 0.006026914808899164, + 0.005541705526411533, + 0.0032006234396249056, + 0.005661065690219402, + 0.0029253908433020115, + 0.005343498662114143, + 0.003736542770639062, + 0.004243095871061087, + 0.004340410232543945, + 0.0026966393925249577, + 0.00473617622628808, + 0.004087058361619711, + 0.004296321887522936, + 0.003850006964057684, + 0.004205326084047556, + 0.004112672992050648, + 0.003122128313407302, + 0.0033797065261751413, + 0.002638712525367737, + 0.12195562571287155, + 0.044921137392520905, + 0.021219447255134583, + 0.02211739867925644, + 0.03108002431690693, + 0.11248523741960526, + 0.02250002883374691, + 0.026107095181941986, + 0.00523395324125886, + 0.005757968872785568, + 0.025370970368385315, + 0.007100410293787718, + 0.004098022356629372, + 0.007846858352422714, + 0.010566103272140026, + 0.01055855955928564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.016342246904969215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979971095919609 + ], + "opposite_logprob_tokens": [ + -3.529799699783325 + ], + "logprob_reward": -0.02979971095919609, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979971095919609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08995742350816727, + 0.09780488908290863, + 0.09656989574432373, + 0.197407528758049, + 0.10766420513391495, + 0.06306411325931549, + 0.06959293782711029, + 0.11301133036613464, + 0.11318763345479965, + 0.05174003913998604 + ], + "credit_map": [ + 0.01167600229382515, + 0.00989395473152399, + 0.01436296571046114, + 0.011028749868273735, + 0.008124178275465965, + 0.008160593919456005, + 0.014241313561797142, + 0.012469664216041565, + 0.01239259447902441, + 0.012921798042953014, + 0.010920815169811249, + 0.016614355146884918, + 0.018211200833320618, + 0.00924301240593195, + 0.009234743192791939, + 0.00826636329293251, + 0.0117361294105649, + 0.014093484729528427, + 0.00871562771499157, + 0.009510213509202003, + 0.020484114065766335, + 0.00890846736729145, + 0.009653475135564804, + 0.013468378223478794, + 0.029676424339413643, + 0.029301632195711136, + 0.02861597016453743, + 0.02248721569776535, + 0.016980236396193504, + 0.01413422916084528, + 0.026439163833856583, + 0.02977265976369381, + 0.01328267715871334, + 0.011390935629606247, + 0.007815532386302948, + 0.013040903955698013, + 0.020455248653888702, + 0.01781451515853405, + 0.011694902554154396, + 0.012169495224952698, + 0.008999918587505817, + 0.010351235046982765, + 0.006007009651511908, + 0.006960588973015547, + 0.008316153660416603, + 0.00926838256418705, + 0.007432351354509592, + 0.005728474818170071, + 0.012187039479613304, + 0.008037615567445755, + 0.015424136072397232, + 0.00633971905335784, + 0.008158951066434383, + 0.0075827548280358315, + 0.005396209191530943, + 0.006466511636972427, + 0.016111819073557854, + 0.012030904181301594, + 0.01498747244477272, + 0.011327123269438744, + 0.01104662660509348, + 0.020516537129878998, + 0.012088358402252197, + 0.01490248367190361, + 0.029971258714795113, + 0.011101150885224342, + 0.008914732374250889, + 0.008765725418925285, + 0.010933732613921165, + 0.026974713429808617, + 0.007838118821382523, + 0.008688208647072315, + 0.004432898014783859, + 0.0045160637237131596, + 0.014391086995601654, + 0.003642879892140627, + 0.003162611974403262, + 0.00571183068677783, + 0.007243633270263672, + 0.008639033883810043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.02979971095919609 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589547231793404 + ], + "opposite_logprob_tokens": [ + -3.895589590072632 + ], + "logprob_reward": -0.020589547231793404, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589547231793404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12034709006547928, + 0.07408813387155533, + 0.06245403736829758, + 0.03968833014369011, + 0.10127311944961548, + 0.1373651921749115, + 0.11635253578424454, + 0.0634528398513794, + 0.20841486752033234, + 0.07656384259462357 + ], + "credit_map": [ + 0.012232132256031036, + 0.020549092441797256, + 0.01652931235730648, + 0.011767311953008175, + 0.029882697388529778, + 0.007783533539623022, + 0.009202562272548676, + 0.01240044180303812, + 0.007581714075058699, + 0.012368482537567616, + 0.008728796616196632, + 0.013105707243084908, + 0.010444974526762962, + 0.006147498264908791, + 0.0067515019327402115, + 0.008959457278251648, + 0.009093636646866798, + 0.007361377123743296, + 0.009916798211634159, + 0.010084579698741436, + 0.004887747578322887, + 0.004622733220458031, + 0.007516910322010517, + 0.008970254100859165, + 0.003951686900109053, + 0.005929699167609215, + 0.0069217621348798275, + 0.004445278085768223, + 0.0051048328168690205, + 0.0035067598801106215, + 0.004345404915511608, + 0.005482904613018036, + 0.01032226625829935, + 0.010614591650664806, + 0.009036450646817684, + 0.013119698502123356, + 0.02117103524506092, + 0.015964927151799202, + 0.011577523313462734, + 0.009466624818742275, + 0.0212258230894804, + 0.022246554493904114, + 0.01024099811911583, + 0.015029660426080227, + 0.025042856112122536, + 0.019113944843411446, + 0.015125417150557041, + 0.009339943528175354, + 0.021453876048326492, + 0.015861405059695244, + 0.014615525491535664, + 0.011789671145379543, + 0.015548638068139553, + 0.01681012287735939, + 0.009798682294785976, + 0.010474616661667824, + 0.008201975375413895, + 0.007387795485556126, + 0.007455660030245781, + 0.008947170339524746, + 0.0058100116439163685, + 0.009262421168386936, + 0.007116980385035276, + 0.009270822629332542, + 0.06735477596521378, + 0.024779915809631348, + 0.01080870721489191, + 0.008491294458508492, + 0.013775475323200226, + 0.05940219387412071, + 0.011196953244507313, + 0.012605564668774605, + 0.004647263325750828, + 0.004816939122974873, + 0.01592019945383072, + 0.0037952784914523363, + 0.0023363272193819284, + 0.005950882099568844, + 0.012373500503599644, + 0.02672344632446766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.020589547231793404 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495632812380791 + ], + "opposite_logprob_tokens": [ + -2.9299564361572266 + ], + "logprob_reward": -0.05495632812380791, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495632812380791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07340778410434723, + 0.01901712268590927, + 0.03706398233771324, + 0.037381336092948914, + 0.12471919506788254, + 0.16732625663280487, + 0.16347365081310272, + 0.05391097441315651, + 0.28281402587890625, + 0.040885746479034424 + ], + "credit_map": [ + 0.011149784550070763, + 0.00670935632660985, + 0.011528735049068928, + 0.011247215792536736, + 0.0070128971710801125, + 0.005445679649710655, + 0.010964605025947094, + 0.00934950727969408, + 0.002715722192078829, + 0.0025848851073533297, + 0.0023882435634732246, + 0.002045787638053298, + 0.002733103930950165, + 0.0018202128121629357, + 0.002858936320990324, + 0.001870230189524591, + 0.004306914750486612, + 0.006179247051477432, + 0.002340785227715969, + 0.0030698382761329412, + 0.010688084177672863, + 0.003128369804471731, + 0.0026147286407649517, + 0.004736014641821384, + 0.0044579231180250645, + 0.006090017035603523, + 0.0054807537235319614, + 0.003913914319127798, + 0.003733936697244644, + 0.002536396961659193, + 0.005938661750406027, + 0.0052297282963991165, + 0.01563188247382641, + 0.013688649982213974, + 0.01042515691369772, + 0.01449408009648323, + 0.020374737679958344, + 0.02190035581588745, + 0.015697572380304337, + 0.012506748549640179, + 0.025203336030244827, + 0.032430246472358704, + 0.021723957732319832, + 0.01632043346762657, + 0.020693324506282806, + 0.021594181656837463, + 0.01652345433831215, + 0.012837325222790241, + 0.025459934026002884, + 0.017027737572789192, + 0.02097202278673649, + 0.013604462146759033, + 0.0233775582164526, + 0.02606799267232418, + 0.016918476670980453, + 0.0200454480946064, + 0.0075408355332911015, + 0.00779670849442482, + 0.0036663333885371685, + 0.009099495597183704, + 0.009136581793427467, + 0.004812928382307291, + 0.007690461352467537, + 0.0041676294058561325, + 0.07665254175662994, + 0.032494403421878815, + 0.019843200221657753, + 0.02115897834300995, + 0.02190854772925377, + 0.0734948217868805, + 0.019175469875335693, + 0.01808604598045349, + 0.004992615897208452, + 0.006075465586036444, + 0.005315146874636412, + 0.002938152989372611, + 0.0036226045340299606, + 0.0053980606608092785, + 0.006429957691580057, + 0.0061137378215789795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.05495632812380791 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327680349349976 + ], + "opposite_logprob_tokens": [ + -0.7577680349349976 + ], + "logprob_reward": -0.6327680349349976, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327680349349976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08119110018014908, + 0.07019061595201492, + 0.05063822492957115, + 0.05156763270497322, + 0.0979643389582634, + 0.058412227779626846, + 0.06880904734134674, + 0.03713756799697876, + 0.34443649649620056, + 0.13965274393558502 + ], + "credit_map": [ + 0.010891972109675407, + 0.011729840189218521, + 0.012216129340231419, + 0.008161341771483421, + 0.012442478910088539, + 0.00583256408572197, + 0.007872233167290688, + 0.012044535018503666, + 0.006822908762842417, + 0.011106832884252071, + 0.006038338877260685, + 0.007201692089438438, + 0.01651385985314846, + 0.004773624707013369, + 0.008973319083452225, + 0.008760036900639534, + 0.006085705943405628, + 0.008846486918628216, + 0.005356447771191597, + 0.007861405611038208, + 0.006930211558938026, + 0.004548386670649052, + 0.004398757591843605, + 0.006610820535570383, + 0.0050169541500508785, + 0.008308619260787964, + 0.006038428749889135, + 0.006707129068672657, + 0.008013158105313778, + 0.004038228187710047, + 0.007801701780408621, + 0.005643407814204693, + 0.010990235954523087, + 0.00882701389491558, + 0.008082890883088112, + 0.013281870633363724, + 0.019426757469773293, + 0.016158105805516243, + 0.011175266467034817, + 0.010022188536822796, + 0.008335305377840996, + 0.007747170049697161, + 0.006412850692868233, + 0.007332216948270798, + 0.010086909867823124, + 0.006924992427229881, + 0.006726217456161976, + 0.004846559371799231, + 0.012127785012125969, + 0.007375138811767101, + 0.013302234932780266, + 0.007200825028121471, + 0.007272311486303806, + 0.008101114071905613, + 0.00564158521592617, + 0.007788046263158321, + 0.005493649747222662, + 0.0045790602453053, + 0.0040888478979468346, + 0.004547520074993372, + 0.0037620784714818, + 0.005242390092462301, + 0.004344436340034008, + 0.005079582799226046, + 0.11253124475479126, + 0.04737713932991028, + 0.017049970105290413, + 0.014464480802416801, + 0.020172875374555588, + 0.09402230381965637, + 0.018427925184369087, + 0.020390566438436508, + 0.01891108602285385, + 0.022486405447125435, + 0.016218427568674088, + 0.010371602140367031, + 0.010228860192000866, + 0.015372361987829208, + 0.023282039910554886, + 0.02278195321559906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.6327680349349976 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with the breathtaking detail of high-speed macro cinematography, this full-body, long shot tracks horizontally with a massive northern pike as it lurks motionless in murky, murky water. The water is a thick suspension of silt and algae, pierced by faint rays of green light, illuminating floating particles. The pike, a formidable predator, hangs suspended in the middle of the water, its long, torpedo-shaped body a detailed mosaic of olive-green and tawny scales, speckled with rows of pale, bean-shaped spots. Its fins, tinged with dark red, make imperceptible adjustments to maintain its position close to the rotting, waterlogged branches at the bottom of the silt. The scene is held in tense silence, and then, with explosive movements rendered in extreme slow motion, the pike strikes. Its powerful tail whips violently, creating a swirling vortex that kicks up clouds of sediment as its entire form surges forward. The camera continues to track perfectly, capturing how each muscle group contracts in visible, continuous waves beneath the skin and how the water displaces around its duckbill snout in a perfectly articulated hydrodynamic shell, transforming quiet tension into a breathtaking display of raw, aquatic power.", + "seed": 30458, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_8/index_4854_seed_30458_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3887926571071148, + "vlm_logprob": -0.3887926571071148, + "vlm_yes_no_margin": 2.1, + "vlm_token_credit": -0.3887926571071148, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021321605890989304 + ], + "opposite_logprob_tokens": [ + -4.146321773529053 + ], + "logprob_reward": -0.021321605890989304, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.021321605890989304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2889571487903595, + 0.16139163076877594, + 0.10817382484674454, + 0.049425579607486725, + 0.03111002780497074, + 0.032102856785058975, + 0.07840525358915329, + 0.030527973547577858, + 0.19030503928661346, + 0.029600640758872032 + ], + "credit_map": [ + 0.02364121563732624, + 0.025667287409305573, + 0.051341984421014786, + 0.03492649272084236, + 0.025632552802562714, + 0.0687033012509346, + 0.02617129497230053, + 0.032873038202524185, + 0.01489198487251997, + 0.02278982475399971, + 0.021337436512112617, + 0.019385002553462982, + 0.029135892167687416, + 0.018234796822071075, + 0.015203338116407394, + 0.02041335217654705, + 0.012333868071436882, + 0.015334727242588997, + 0.015153279528021812, + 0.010691208764910698, + 0.018833963200449944, + 0.01309687178581953, + 0.012581962160766125, + 0.010147953405976295, + 0.00564536964520812, + 0.004407135769724846, + 0.0038462113589048386, + 0.005098463036119938, + 0.007479527033865452, + 0.009771671146154404, + 0.006509618833661079, + 0.006667581386864185, + 0.004423896316438913, + 0.003270033048465848, + 0.003946708980947733, + 0.002928096102550626, + 0.005016620736569166, + 0.003567339386790991, + 0.004679590463638306, + 0.0032777416054159403, + 0.004174408037215471, + 0.0071038659662008286, + 0.003722183406352997, + 0.002774910070002079, + 0.0036819223314523697, + 0.003325674682855606, + 0.004115127492696047, + 0.00320476689375937, + 0.008241238072514534, + 0.015795065090060234, + 0.0070211440324783325, + 0.005468813236802816, + 0.021346628665924072, + 0.010911081917583942, + 0.003975752741098404, + 0.0056455302983522415, + 0.00327025237493217, + 0.004122606012970209, + 0.003522335086017847, + 0.0038073654286563396, + 0.0036174685228616, + 0.004446661565452814, + 0.004577576648443937, + 0.0031637074425816536, + 0.015369563363492489, + 0.01842644065618515, + 0.02384699322283268, + 0.017103254795074463, + 0.021246209740638733, + 0.022180236876010895, + 0.02426629513502121, + 0.04786605015397072, + 0.0024357126094400883, + 0.004473545588552952, + 0.0043610986322164536, + 0.0042871031910181046, + 0.004809529520571232, + 0.003533785929903388, + 0.002747944323346019, + 0.0029519207309931517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.021321605890989304 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865137115120888 + ], + "opposite_logprob_tokens": [ + -3.0486514568328857 + ], + "logprob_reward": -0.04865137115120888, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865137115120888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07258757203817368, + 0.062486857175827026, + 0.06511489301919937, + 0.12486200779676437, + 0.054202523082494736, + 0.03269548714160919, + 0.13895899057388306, + 0.021742628887295723, + 0.3691630959510803, + 0.05818595737218857 + ], + "credit_map": [ + 0.006908121053129435, + 0.005995744373649359, + 0.009867602959275246, + 0.013392501510679722, + 0.007494246121495962, + 0.010536045767366886, + 0.009371904656291008, + 0.009021403267979622, + 0.006445548962801695, + 0.010763355530798435, + 0.00912165641784668, + 0.006486182101070881, + 0.008263111114501953, + 0.005218016915023327, + 0.008653907105326653, + 0.007535074837505817, + 0.009283943101763725, + 0.008350460790097713, + 0.010225336998701096, + 0.006104988511651754, + 0.007856091484427452, + 0.007893637754023075, + 0.009289877489209175, + 0.006110555026680231, + 0.012279381975531578, + 0.008951912634074688, + 0.009864769876003265, + 0.010383899323642254, + 0.020243370905518532, + 0.03127914294600487, + 0.013499950990080833, + 0.01835956797003746, + 0.005613557528704405, + 0.0068843611516058445, + 0.005232030060142279, + 0.005253349896520376, + 0.007271230686455965, + 0.0066483705304563046, + 0.012122750282287598, + 0.005176875274628401, + 0.00311960163526237, + 0.0061029610224068165, + 0.004293880891054869, + 0.003221445018425584, + 0.004410403780639172, + 0.0034096476156264544, + 0.005243632011115551, + 0.002893915632739663, + 0.012097379192709923, + 0.026204176247119904, + 0.011293964460492134, + 0.010969662107527256, + 0.03585650026798248, + 0.020899780094623566, + 0.009016015566885471, + 0.012621505185961723, + 0.00260833278298378, + 0.0027668641414493322, + 0.002580220578238368, + 0.00254744291305542, + 0.0026008295826613903, + 0.0034501380287110806, + 0.0025173367466777563, + 0.0026714655105024576, + 0.030819345265626907, + 0.03319656476378441, + 0.050904493778944016, + 0.036933038383722305, + 0.038972996175289154, + 0.042974017560482025, + 0.04943911358714104, + 0.08592354506254196, + 0.004610244184732437, + 0.006087076850235462, + 0.007939783856272697, + 0.008857727982103825, + 0.007307639345526695, + 0.00812822300940752, + 0.005357168614864349, + 0.009898094460368156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.04865137115120888 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023296575993299484 + ], + "opposite_logprob_tokens": [ + -3.773296594619751 + ], + "logprob_reward": -0.023296575993299484, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023296575993299484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14054469764232635, + 0.09860692918300629, + 0.07126344740390778, + 0.050626762211322784, + 0.0472700335085392, + 0.054722752422094345, + 0.16940715909004211, + 0.03695351630449295, + 0.27636221051216125, + 0.05424250662326813 + ], + "credit_map": [ + 0.017551284283399582, + 0.012463224120438099, + 0.019569875672459602, + 0.020807716995477676, + 0.01088793482631445, + 0.034225817769765854, + 0.011421093717217445, + 0.013617755845189095, + 0.010600553825497627, + 0.015225116163492203, + 0.013366932049393654, + 0.01043677981942892, + 0.015758540481328964, + 0.011868232861161232, + 0.011003032326698303, + 0.01034773699939251, + 0.008319626562297344, + 0.009029346518218517, + 0.011452021077275276, + 0.007953539490699768, + 0.013571757823228836, + 0.005502617917954922, + 0.009061377495527267, + 0.006373165175318718, + 0.004989318549633026, + 0.005986846052110195, + 0.006429674103856087, + 0.004329876974225044, + 0.008271226659417152, + 0.007573653943836689, + 0.006266007665544748, + 0.006780160125344992, + 0.007619078271090984, + 0.0062766713090240955, + 0.005537089891731739, + 0.0038535790517926216, + 0.00632810452952981, + 0.0055233025923371315, + 0.007307741325348616, + 0.004824465606361628, + 0.006332754623144865, + 0.011918791569769382, + 0.005718259606510401, + 0.005559538956731558, + 0.006729071959853172, + 0.005970040336251259, + 0.007394324988126755, + 0.00509997084736824, + 0.015698591247200966, + 0.029430918395519257, + 0.01567092351615429, + 0.01232356857508421, + 0.04569509997963905, + 0.025267433375120163, + 0.01039962563663721, + 0.014920978806912899, + 0.003779484424740076, + 0.003392362967133522, + 0.004524894990026951, + 0.004229865036904812, + 0.00613753916695714, + 0.0055392589420080185, + 0.00562659464776516, + 0.0037235156632959843, + 0.02429964207112789, + 0.02649836428463459, + 0.04027780890464783, + 0.0295144971460104, + 0.030782412737607956, + 0.032605454325675964, + 0.036946650594472885, + 0.055437371134757996, + 0.0031970003619790077, + 0.005233941599726677, + 0.007402515504509211, + 0.009917485527694225, + 0.006778336595743895, + 0.008093763142824173, + 0.006030987482517958, + 0.0075884717516601086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.023296575993299484 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5020551681518555 + ], + "opposite_logprob_tokens": [ + -0.25205522775650024 + ], + "logprob_reward": -1.5020551681518555, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5020551681518555, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11628714203834534, + 0.0692635253071785, + 0.09459929913282394, + 0.0625903531908989, + 0.05252332612872124, + 0.06315571814775467, + 0.06058509647846222, + 0.07947063446044922, + 0.32285571098327637, + 0.07866918295621872 + ], + "credit_map": [ + 0.015961501747369766, + 0.016715779900550842, + 0.017025666311383247, + 0.013679821975529194, + 0.012368724681437016, + 0.011392765678465366, + 0.017347510904073715, + 0.011795369908213615, + 0.00888520572334528, + 0.010089858435094357, + 0.009868270717561245, + 0.008992751128971577, + 0.008369669318199158, + 0.006149377673864365, + 0.007731185760349035, + 0.009177207946777344, + 0.01064769271761179, + 0.012985941022634506, + 0.013230021111667156, + 0.00716402055695653, + 0.014069969765841961, + 0.012762303464114666, + 0.01373904850333929, + 0.010000308975577354, + 0.004627103451639414, + 0.005729937460273504, + 0.006757300812751055, + 0.006029780022799969, + 0.009762897156178951, + 0.008732750080525875, + 0.01266044657677412, + 0.008290134370326996, + 0.007379123475402594, + 0.006683412007987499, + 0.006904602516442537, + 0.004575666040182114, + 0.007652667351067066, + 0.005926255602389574, + 0.008667707443237305, + 0.004733888898044825, + 0.009335110895335674, + 0.008913697674870491, + 0.008856355212628841, + 0.0056498548947274685, + 0.006992710754275322, + 0.007063945755362511, + 0.008727267384529114, + 0.007616776041686535, + 0.006506713107228279, + 0.01028013601899147, + 0.005417317152023315, + 0.007156812585890293, + 0.012084471061825752, + 0.007112022023648024, + 0.005339809227734804, + 0.006687813904136419, + 0.00995338149368763, + 0.00855839904397726, + 0.009326080791652203, + 0.008770156651735306, + 0.011903839185833931, + 0.011352834291756153, + 0.0124973114579916, + 0.007108638063073158, + 0.027071895077824593, + 0.030498988926410675, + 0.0647505596280098, + 0.033184297382831573, + 0.04023329168558121, + 0.041896071285009384, + 0.0456739217042923, + 0.039546675980091095, + 0.007187412120401859, + 0.0145264295861125, + 0.007801983971148729, + 0.009708067402243614, + 0.015309762209653854, + 0.011013421230018139, + 0.006437748204916716, + 0.006684358697384596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -1.5020551681518555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486385643482208 + ], + "opposite_logprob_tokens": [ + -1.2236385345458984 + ], + "logprob_reward": -0.3486385643482208, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486385643482208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25035735964775085, + 0.10962285101413727, + 0.11101531982421875, + 0.06221615523099899, + 0.05350426584482193, + 0.05195028707385063, + 0.057328443974256516, + 0.12842369079589844, + 0.12110709398984909, + 0.05447453260421753 + ], + "credit_map": [ + 0.03333389759063721, + 0.022939788177609444, + 0.02093699388206005, + 0.021271368488669395, + 0.02634299360215664, + 0.07624997198581696, + 0.018560098484158516, + 0.030722253024578094, + 0.009845312684774399, + 0.015521707944571972, + 0.016808822751045227, + 0.015525052323937416, + 0.01648508571088314, + 0.009257865138351917, + 0.012799759395420551, + 0.013379249721765518, + 0.011120077222585678, + 0.015427259728312492, + 0.013831529766321182, + 0.012477654032409191, + 0.019816352054476738, + 0.01790042221546173, + 0.010095169767737389, + 0.010346855968236923, + 0.00794835016131401, + 0.005202384199947119, + 0.005057184025645256, + 0.004694546107202768, + 0.011240645311772823, + 0.0117884436622262, + 0.010459491983056068, + 0.0058251055888831615, + 0.004333403892815113, + 0.006698411889374256, + 0.0062373122200369835, + 0.004526747856289148, + 0.010324188508093357, + 0.006527559366077185, + 0.010186655446887016, + 0.004669984802603722, + 0.006897500716149807, + 0.008884653449058533, + 0.007220163010060787, + 0.005163362715393305, + 0.005230486858636141, + 0.005897799041122198, + 0.006957750767469406, + 0.005698570981621742, + 0.004977481905370951, + 0.01220964640378952, + 0.006508105434477329, + 0.006297672633081675, + 0.012259481474757195, + 0.007315644063055515, + 0.0036240890622138977, + 0.004136322997510433, + 0.016675444319844246, + 0.009965985082089901, + 0.016645001247525215, + 0.014892373234033585, + 0.02151903696358204, + 0.016883976757526398, + 0.02119162119925022, + 0.010650239884853363, + 0.008412366732954979, + 0.007902934215962887, + 0.018703781068325043, + 0.015796653926372528, + 0.014130308292806149, + 0.014594225212931633, + 0.016062144190073013, + 0.02550467848777771, + 0.004466865677386522, + 0.008029838092625141, + 0.00594038562849164, + 0.007575230672955513, + 0.01070396788418293, + 0.0064038001000881195, + 0.004346925299614668, + 0.0070075225085020065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.3486385643482208 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "Shot with the breathtaking detail of high-speed macro cinematography, this full-body, long shot tracks horizontally with a massive northern pike as it lurks motionless in murky, murky water. The water is a thick suspension of silt and algae, pierced by faint rays of green light, illuminating floating particles. The pike, a formidable predator, hangs suspended in the middle of the water, its long, torpedo-shaped body a detailed mosaic of olive-green and tawny scales, speckled with rows of pale, bean-shaped spots. Its fins, tinged with dark red, make imperceptible adjustments to maintain its position close to the rotting, waterlogged branches at the bottom of the silt. The scene is held in tense silence, and then, with explosive movements rendered in extreme slow motion, the pike strikes. Its powerful tail whips violently, creating a swirling vortex that kicks up clouds of sediment as its entire form surges forward. The camera continues to track perfectly, capturing how each muscle group contracts in visible, continuous waves beneath the skin and how the water displaces around its duckbill snout in a perfectly articulated hydrodynamic shell, transforming quiet tension into a breathtaking display of raw, aquatic power.", + "seed": 30459, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_8/index_4854_seed_30459_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.25956558994948864, + "vlm_logprob": -0.25956558994948864, + "vlm_yes_no_margin": 2.449999952316284, + "vlm_token_credit": -0.25956558994948864, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027312573045492172 + ], + "opposite_logprob_tokens": [ + -3.9023125171661377 + ], + "logprob_reward": -0.027312573045492172, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.027312573045492172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09763415157794952, + 0.08468154072761536, + 0.030719546601176262, + 0.17822541296482086, + 0.20915484428405762, + 0.05682317912578583, + 0.16776087880134583, + 0.057803958654403687, + 0.057112157344818115, + 0.06008441001176834 + ], + "credit_map": [ + 0.010409160517156124, + 0.014969181269407272, + 0.01244794949889183, + 0.010761219076812267, + 0.011641856282949448, + 0.015416016802191734, + 0.01032822486013174, + 0.011660542339086533, + 0.013464249670505524, + 0.012051073834300041, + 0.008986230939626694, + 0.013442909345030785, + 0.007594333030283451, + 0.011541091836988926, + 0.009383373893797398, + 0.008218267932534218, + 0.0031253008637577295, + 0.005756268743425608, + 0.005194720812141895, + 0.004276803229004145, + 0.0036430819891393185, + 0.0023545946460217237, + 0.0029664747416973114, + 0.0034022973850369453, + 0.013049369677901268, + 0.05847058817744255, + 0.03722729533910751, + 0.008008608594536781, + 0.018515411764383316, + 0.008989936672151089, + 0.01340247131884098, + 0.020561710000038147, + 0.02212485298514366, + 0.046919338405132294, + 0.012698480859398842, + 0.010298402979969978, + 0.03610135242342949, + 0.02051561139523983, + 0.0296571496874094, + 0.030839629471302032, + 0.007255093660205603, + 0.006753366440534592, + 0.004367501474916935, + 0.004492587875574827, + 0.004874194972217083, + 0.007538617122918367, + 0.011424501426517963, + 0.01011731382459402, + 0.017487410455942154, + 0.008588332682847977, + 0.017359355464577675, + 0.02356792613863945, + 0.01885458454489708, + 0.017925286665558815, + 0.0332578681409359, + 0.030720097944140434, + 0.007683631964027882, + 0.0070221745409071445, + 0.006325984839349985, + 0.00972050055861473, + 0.00776760745793581, + 0.00461400905624032, + 0.00763712776824832, + 0.007032915484160185, + 0.003579827258363366, + 0.00609502661973238, + 0.005174134857952595, + 0.008066735230386257, + 0.014614726416766644, + 0.007265853229910135, + 0.005906365811824799, + 0.0064094834960997105, + 0.0055465977638959885, + 0.0057199192233383656, + 0.008689481765031815, + 0.009031353518366814, + 0.006345980800688267, + 0.008298102766275406, + 0.009885608218610287, + 0.006567360367625952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.027312573045492172 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038102924823760986 + ], + "opposite_logprob_tokens": [ + -3.288102865219116 + ], + "logprob_reward": -0.038102924823760986, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038102924823760986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09144258499145508, + 0.08373504132032394, + 0.12474024295806885, + 0.1066984012722969, + 0.25846490263938904, + 0.03416166082024574, + 0.05152997002005577, + 0.052393589168787, + 0.14209051430225372, + 0.05474308133125305 + ], + "credit_map": [ + 0.007180103100836277, + 0.01697648875415325, + 0.00864915456622839, + 0.012741136364638805, + 0.01449780073016882, + 0.01564357616007328, + 0.007830084301531315, + 0.007924247533082962, + 0.010029580444097519, + 0.009460152126848698, + 0.00725097581744194, + 0.011708006262779236, + 0.010246889665722847, + 0.013789528980851173, + 0.011931324377655983, + 0.009318581782281399, + 0.011299730278551579, + 0.02770264819264412, + 0.01741950213909149, + 0.01106345560401678, + 0.015544972382485867, + 0.010531018488109112, + 0.014828856103122234, + 0.01635006256401539, + 0.007675265893340111, + 0.014067408628761768, + 0.03723553940653801, + 0.005261978134512901, + 0.005670666694641113, + 0.005534567404538393, + 0.007310576271265745, + 0.023942407220602036, + 0.030637988820672035, + 0.035235367715358734, + 0.024683833122253418, + 0.025987429544329643, + 0.03814162313938141, + 0.021148869767785072, + 0.043647944927215576, + 0.038981832563877106, + 0.0049122413620352745, + 0.0036122549790889025, + 0.004614006727933884, + 0.0038253222592175007, + 0.002782082185149193, + 0.0038436762988567352, + 0.005119157023727894, + 0.005452921614050865, + 0.0042967600747942924, + 0.005143256392329931, + 0.005740349180996418, + 0.006686237175017595, + 0.0062803239561617374, + 0.0059939101338386536, + 0.00885858479887247, + 0.008530547842383385, + 0.005104689858853817, + 0.0058630299754440784, + 0.003347888821735978, + 0.006723169703036547, + 0.00788168702274561, + 0.005929551087319851, + 0.007726125884801149, + 0.009817448444664478, + 0.02117292955517769, + 0.021124299615621567, + 0.015588306821882725, + 0.01722591742873192, + 0.028540894389152527, + 0.011452159844338894, + 0.01152590848505497, + 0.015460098162293434, + 0.007490428630262613, + 0.005169718060642481, + 0.006823406554758549, + 0.0077234189957380295, + 0.006302182096987963, + 0.006369356065988541, + 0.007344946265220642, + 0.007519630249589682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.038102924823760986 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181919876486063 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.0181919876486063, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.0181919876486063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16618050634860992, + 0.11813917756080627, + 0.05983790382742882, + 0.10366179794073105, + 0.15108859539031982, + 0.04856100678443909, + 0.09094303846359253, + 0.05830011144280434, + 0.08484668284654617, + 0.11844111233949661 + ], + "credit_map": [ + 0.012674990110099316, + 0.030059482902288437, + 0.0212054792791605, + 0.018928078934550285, + 0.022329604253172874, + 0.024120282381772995, + 0.016165433451533318, + 0.020697174593806267, + 0.01574506238102913, + 0.02374659664928913, + 0.009976985864341259, + 0.01469268836081028, + 0.013244482688605785, + 0.018912145867943764, + 0.01079509686678648, + 0.011026138439774513, + 0.0062756771221756935, + 0.008610260672867298, + 0.00837999302893877, + 0.004675551783293486, + 0.005306227132678032, + 0.0048486762680113316, + 0.009256954304873943, + 0.012484575621783733, + 0.008844445459544659, + 0.01068641897290945, + 0.03569116070866585, + 0.004785635508596897, + 0.007160738576203585, + 0.009014424867928028, + 0.008646951988339424, + 0.01883203163743019, + 0.017237424850463867, + 0.028338655829429626, + 0.010813704691827297, + 0.011195897124707699, + 0.024923106655478477, + 0.01161304023116827, + 0.022203249856829643, + 0.024763545021414757, + 0.007004619110375643, + 0.0052082352340221405, + 0.009904608130455017, + 0.0033629925455898046, + 0.0036352004390209913, + 0.003691526595503092, + 0.006274296436458826, + 0.00947953388094902, + 0.009782534092664719, + 0.006092348136007786, + 0.010803703218698502, + 0.012774894014000893, + 0.00839545764029026, + 0.007339257746934891, + 0.021212143823504448, + 0.014542704448103905, + 0.008761844597756863, + 0.006915736943483353, + 0.007086564786732197, + 0.010686337016522884, + 0.005454499740153551, + 0.00473789731040597, + 0.0066824606619775295, + 0.007974781095981598, + 0.007517600432038307, + 0.01036105491220951, + 0.012563180178403854, + 0.015535058453679085, + 0.010875036008656025, + 0.010682623833417892, + 0.008937143720686436, + 0.008374987170100212, + 0.016346100717782974, + 0.011516520753502846, + 0.015977781265974045, + 0.014874582178890705, + 0.016038881614804268, + 0.013571826741099358, + 0.013096080161631107, + 0.01701935939490795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.0181919876486063 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0538806915283203 + ], + "opposite_logprob_tokens": [ + -0.4288806915283203 + ], + "logprob_reward": -1.0538806915283203, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0538806915283203, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1063629537820816, + 0.10380861908197403, + 0.02779632806777954, + 0.11276020109653473, + 0.10205064713954926, + 0.04934505745768547, + 0.08114797621965408, + 0.21386319398880005, + 0.1410548835992813, + 0.061810124665498734 + ], + "credit_map": [ + 0.009531891904771328, + 0.018965071067214012, + 0.015331899747252464, + 0.013422435149550438, + 0.011317851021885872, + 0.016457779332995415, + 0.011175276711583138, + 0.010160756297409534, + 0.008889297023415565, + 0.017347315326333046, + 0.01175481453537941, + 0.017767511308193207, + 0.013054721057415009, + 0.010170986875891685, + 0.014679470099508762, + 0.010144504718482494, + 0.0028940271586179733, + 0.005383902229368687, + 0.004886047448962927, + 0.002891336800530553, + 0.003561462741345167, + 0.0026809065602719784, + 0.0024221299681812525, + 0.0030765144620090723, + 0.007877419702708721, + 0.01856425404548645, + 0.039757054299116135, + 0.005418240092694759, + 0.00627778610214591, + 0.004353976342827082, + 0.008177492767572403, + 0.022333970293402672, + 0.012010577134788036, + 0.017379827797412872, + 0.006896688602864742, + 0.006795412860810757, + 0.018326880410313606, + 0.009364663623273373, + 0.015188162215054035, + 0.016088435426354408, + 0.007246794179081917, + 0.00823151133954525, + 0.0046894303523004055, + 0.0037835328839719296, + 0.004615063313394785, + 0.008079475723206997, + 0.006846746429800987, + 0.005852501839399338, + 0.007763335481286049, + 0.00924675352871418, + 0.009070920757949352, + 0.010128979571163654, + 0.009829550981521606, + 0.011112206615507603, + 0.013818791136145592, + 0.010177440010011196, + 0.03249355033040047, + 0.02126750536262989, + 0.03323504701256752, + 0.029086239635944366, + 0.018310388550162315, + 0.013173963874578476, + 0.03796160966157913, + 0.028334885835647583, + 0.015381339006125927, + 0.01408521831035614, + 0.022013764828443527, + 0.022713875398039818, + 0.022161923348903656, + 0.015652727335691452, + 0.015604077838361263, + 0.013441947288811207, + 0.006709571927785873, + 0.007516891695559025, + 0.010503887198865414, + 0.010146663524210453, + 0.00601559691131115, + 0.007125145755708218, + 0.008604173548519611, + 0.005188190843909979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -1.0538806915283203 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033977270126343 + ], + "opposite_logprob_tokens": [ + -1.9103398323059082 + ], + "logprob_reward": -0.16033977270126343, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033977270126343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11359476298093796, + 0.12681293487548828, + 0.062403373420238495, + 0.09325345605611801, + 0.1682954579591751, + 0.057166773825883865, + 0.1180315837264061, + 0.09196661412715912, + 0.13473975658416748, + 0.03373519331216812 + ], + "credit_map": [ + 0.012433137744665146, + 0.021413765847682953, + 0.013185806572437286, + 0.015184106305241585, + 0.010974711738526821, + 0.018545186147093773, + 0.011359215714037418, + 0.010498848743736744, + 0.012828015722334385, + 0.027761386707425117, + 0.011719168163836002, + 0.015813525766134262, + 0.008677288889884949, + 0.02522614412009716, + 0.012981963343918324, + 0.011805466376245022, + 0.007059060502797365, + 0.0062832944095134735, + 0.008937709033489227, + 0.006383285392075777, + 0.0055395192466676235, + 0.006777163129299879, + 0.01143414806574583, + 0.009989200159907341, + 0.007346371188759804, + 0.02687198854982853, + 0.02784683182835579, + 0.003989891614764929, + 0.008114189840853214, + 0.005966574884951115, + 0.006036890670657158, + 0.007080730050802231, + 0.019916662946343422, + 0.025726232677698135, + 0.018907824531197548, + 0.025519639253616333, + 0.02236749231815338, + 0.014582912437617779, + 0.019827378913760185, + 0.02144733816385269, + 0.007400671020150185, + 0.006531809456646442, + 0.011786471121013165, + 0.004920340143144131, + 0.004511420615017414, + 0.005811900831758976, + 0.008057309314608574, + 0.008146858774125576, + 0.009981900453567505, + 0.007367431651800871, + 0.01053725928068161, + 0.01733752153813839, + 0.01180945709347725, + 0.009587736800312996, + 0.028103375807404518, + 0.023306915536522865, + 0.00944787822663784, + 0.011303718201816082, + 0.013509247452020645, + 0.011035721749067307, + 0.0118974968791008, + 0.00789155438542366, + 0.014775780029594898, + 0.012105233035981655, + 0.017264649271965027, + 0.018665876239538193, + 0.02070285752415657, + 0.01649770513176918, + 0.022157078608870506, + 0.014480206184089184, + 0.01295718364417553, + 0.012014219537377357, + 0.0034496518783271313, + 0.0030172935221344233, + 0.00426962086930871, + 0.006253179628401995, + 0.0030259168706834316, + 0.0031633228063583374, + 0.0051162634044885635, + 0.005439949221909046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.16033977270126343 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "Shot with the breathtaking detail of high-speed macro cinematography, this full-body, long shot tracks horizontally with a massive northern pike as it lurks motionless in murky, murky water. The water is a thick suspension of silt and algae, pierced by faint rays of green light, illuminating floating particles. The pike, a formidable predator, hangs suspended in the middle of the water, its long, torpedo-shaped body a detailed mosaic of olive-green and tawny scales, speckled with rows of pale, bean-shaped spots. Its fins, tinged with dark red, make imperceptible adjustments to maintain its position close to the rotting, waterlogged branches at the bottom of the silt. The scene is held in tense silence, and then, with explosive movements rendered in extreme slow motion, the pike strikes. Its powerful tail whips violently, creating a swirling vortex that kicks up clouds of sediment as its entire form surges forward. The camera continues to track perfectly, capturing how each muscle group contracts in visible, continuous waves beneath the skin and how the water displaces around its duckbill snout in a perfectly articulated hydrodynamic shell, transforming quiet tension into a breathtaking display of raw, aquatic power.", + "seed": 30460, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_8/index_4854_seed_30460_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2108595846220851, + "vlm_logprob": -0.2108595846220851, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.2108595846220851, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06088443472981453 + ], + "opposite_logprob_tokens": [ + -2.935884475708008 + ], + "logprob_reward": -0.06088443472981453, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06088443472981453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21805836260318756, + 0.07488442212343216, + 0.0803971216082573, + 0.09460967779159546, + 0.06878875941038132, + 0.030598927289247513, + 0.23571166396141052, + 0.05903695896267891, + 0.04119947925209999, + 0.09671462327241898 + ], + "credit_map": [ + 0.030905229970812798, + 0.046193331480026245, + 0.02426355518400669, + 0.023919256404042244, + 0.04413929581642151, + 0.01744994893670082, + 0.011922293342649937, + 0.019265467301011086, + 0.01022659707814455, + 0.006941471714526415, + 0.019789110869169235, + 0.01070581004023552, + 0.008161297999322414, + 0.006995965726673603, + 0.00486806919798255, + 0.007196100428700447, + 0.010498966090381145, + 0.013271544128656387, + 0.012308148667216301, + 0.012967849150300026, + 0.010997189208865166, + 0.00990217737853527, + 0.005552269052714109, + 0.004898974671959877, + 0.014417238533496857, + 0.01902242936193943, + 0.012879465706646442, + 0.012549612671136856, + 0.012938657775521278, + 0.008950491435825825, + 0.008608137257397175, + 0.005243641324341297, + 0.010052476078271866, + 0.008009894751012325, + 0.01137432549148798, + 0.010938609018921852, + 0.009423534385859966, + 0.008245940320193768, + 0.006902256980538368, + 0.0038417172618210316, + 0.004827537573873997, + 0.007475086487829685, + 0.0040834820829331875, + 0.0019819317385554314, + 0.003590069245547056, + 0.0035789343528449535, + 0.002507388824597001, + 0.002554495120421052, + 0.04868483170866966, + 0.03309743478894234, + 0.05408734083175659, + 0.022747470065951347, + 0.02289872244000435, + 0.019409479573369026, + 0.01694183424115181, + 0.017844567075371742, + 0.0067233731970191, + 0.005509904585778713, + 0.006523961666971445, + 0.00671249208971858, + 0.006111893802881241, + 0.0067993770353496075, + 0.007573511451482773, + 0.013082446530461311, + 0.006024654489010572, + 0.006326966919004917, + 0.005592217203229666, + 0.006566870957612991, + 0.003850049339234829, + 0.003967212978750467, + 0.003987039439380169, + 0.004884466994553804, + 0.005244064144790173, + 0.007507412228733301, + 0.008620859123766422, + 0.015379656106233597, + 0.01305049005895853, + 0.01039027888327837, + 0.027396690100431442, + 0.009125172160565853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.06088443472981453 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069775447249413 + ], + "opposite_logprob_tokens": [ + -4.1410698890686035 + ], + "logprob_reward": -0.016069775447249413, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069775447249413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13984493911266327, + 0.34995707869529724, + 0.06276784837245941, + 0.10454527288675308, + 0.04795297607779503, + 0.01713128201663494, + 0.02985040284693241, + 0.09963575750589371, + 0.05176049843430519, + 0.09655392169952393 + ], + "credit_map": [ + 0.02077687345445156, + 0.026209209114313126, + 0.014573740772902966, + 0.021570255979895592, + 0.015864932909607887, + 0.014003841206431389, + 0.007988810539245605, + 0.01885727420449257, + 0.045648492872714996, + 0.03995996713638306, + 0.03593061864376068, + 0.044523004442453384, + 0.04345465451478958, + 0.04062410816550255, + 0.0374135784804821, + 0.0624026283621788, + 0.008486686274409294, + 0.007821980863809586, + 0.008990820497274399, + 0.009819835424423218, + 0.008045741356909275, + 0.009874790906906128, + 0.004442740697413683, + 0.0052852509543299675, + 0.0157560333609581, + 0.02314007841050625, + 0.015161778777837753, + 0.012902790680527687, + 0.011005508713424206, + 0.011430974118411541, + 0.008371220901608467, + 0.006776897702366114, + 0.007084236014634371, + 0.006330918520689011, + 0.007143616676330566, + 0.006332577206194401, + 0.007429163902997971, + 0.005466018803417683, + 0.0048530809581279755, + 0.0033133637625724077, + 0.0027492453809827566, + 0.004589274991303682, + 0.0022016381844878197, + 0.0012457361444830894, + 0.00178041763138026, + 0.0020348969846963882, + 0.0011817077174782753, + 0.001348365331068635, + 0.00565706379711628, + 0.004198726266622543, + 0.006260155234485865, + 0.00382606009952724, + 0.0026993693318217993, + 0.0026874602772295475, + 0.002641273196786642, + 0.0018802933627739549, + 0.010553685016930103, + 0.009516660124063492, + 0.010031662881374359, + 0.012146656401455402, + 0.008949467912316322, + 0.010422440245747566, + 0.013689839281141758, + 0.02432534284889698, + 0.01163571234792471, + 0.008578256703913212, + 0.007591873407363892, + 0.004717398900538683, + 0.004450405482202768, + 0.00495862914249301, + 0.005759491119533777, + 0.004068730399012566, + 0.006907069589942694, + 0.008528905920684338, + 0.00831299927085638, + 0.011864176951348782, + 0.013563066720962524, + 0.012345031835138798, + 0.024844128638505936, + 0.010188546031713486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.016069775447249413 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419964898377657 + ], + "opposite_logprob_tokens": [ + -4.264199733734131 + ], + "logprob_reward": -0.01419964898377657, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419964898377657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19468948245048523, + 0.050390057265758514, + 0.023613251745700836, + 0.04995854198932648, + 0.030796697363257408, + 0.01800219900906086, + 0.026346109807491302, + 0.06038200482726097, + 0.4934845566749573, + 0.052337124943733215 + ], + "credit_map": [ + 0.02743496373295784, + 0.039950814098119736, + 0.022418254986405373, + 0.02026643417775631, + 0.028545256704092026, + 0.018155395984649658, + 0.011199803091585636, + 0.02671855315566063, + 0.007540450431406498, + 0.003816576674580574, + 0.016922488808631897, + 0.009497196413576603, + 0.004150048363953829, + 0.0035831413697451353, + 0.002288330113515258, + 0.002591822063550353, + 0.003219595644623041, + 0.002688966691493988, + 0.0027636887971311808, + 0.0043334499932825565, + 0.002390319248661399, + 0.0031595102045685053, + 0.0021198310423642397, + 0.002937891986221075, + 0.007442699279636145, + 0.006547084078192711, + 0.007914770394563675, + 0.007776099257171154, + 0.005353027023375034, + 0.007906460203230381, + 0.003507358254864812, + 0.0035110448952764273, + 0.004721430130302906, + 0.004117295145988464, + 0.0054824864491820335, + 0.0044480073265731335, + 0.0038125498685985804, + 0.0036227775271981955, + 0.002645982662215829, + 0.0019461683696135879, + 0.0028990048449486494, + 0.004518075846135616, + 0.0018748542061075568, + 0.0011883623665198684, + 0.0022198802325874567, + 0.0027520647272467613, + 0.001250801025889814, + 0.0012991564581170678, + 0.0044711437076330185, + 0.003707068506628275, + 0.004098330624401569, + 0.0031338809058070183, + 0.0034998126793652773, + 0.002942918101325631, + 0.002154366113245487, + 0.0023385887034237385, + 0.007676349021494389, + 0.012368199415504932, + 0.008203050121665001, + 0.007368902210146189, + 0.0070939031429588795, + 0.005875946953892708, + 0.004724867641925812, + 0.007070785388350487, + 0.10176437348127365, + 0.11282667517662048, + 0.05379457771778107, + 0.03966032341122627, + 0.04667766019701958, + 0.04311785101890564, + 0.05930739641189575, + 0.036335717886686325, + 0.006039706524461508, + 0.006451880559325218, + 0.00883852131664753, + 0.007136333268135786, + 0.006455970928072929, + 0.005437599495053291, + 0.007210860028862953, + 0.004766249563544989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.01419964898377657 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761253833770752 + ], + "opposite_logprob_tokens": [ + -0.8261253833770752 + ], + "logprob_reward": -0.5761253833770752, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761253833770752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12883836030960083, + 0.13249585032463074, + 0.04133881628513336, + 0.054018765687942505, + 0.023475317284464836, + 0.05708710476756096, + 0.030410028994083405, + 0.07231632620096207, + 0.05381927639245987, + 0.40620020031929016 + ], + "credit_map": [ + 0.020453913137316704, + 0.021613413468003273, + 0.013719483278691769, + 0.011324713937938213, + 0.02027038112282753, + 0.012037198059260845, + 0.007576635107398033, + 0.02184261940419674, + 0.016035735607147217, + 0.012934310361742973, + 0.02847251109778881, + 0.01682191900908947, + 0.015277005732059479, + 0.014912293292582035, + 0.008671077899634838, + 0.019370995461940765, + 0.005579388700425625, + 0.007165383547544479, + 0.005294053815305233, + 0.005681357346475124, + 0.008437419310212135, + 0.004656415432691574, + 0.0022921713534742594, + 0.002232631202787161, + 0.008108923211693764, + 0.009967872872948647, + 0.00764637254178524, + 0.007081198040395975, + 0.006839403882622719, + 0.007899561896920204, + 0.003263553837314248, + 0.0032118826638907194, + 0.0031972629949450493, + 0.002228281693533063, + 0.003957427106797695, + 0.003588777966797352, + 0.0035677182022482157, + 0.0024739676155149937, + 0.0025014944840222597, + 0.001960385823622346, + 0.00996997207403183, + 0.011256160214543343, + 0.0049683647230267525, + 0.0034105901140719652, + 0.009742891415953636, + 0.009789615869522095, + 0.004336094483733177, + 0.0036134158726781607, + 0.004458873067051172, + 0.005123397801071405, + 0.00480045098811388, + 0.0032540466636419296, + 0.0033474008087068796, + 0.004425618331879377, + 0.002747912425547838, + 0.0022523284424096346, + 0.007404015399515629, + 0.007479536812752485, + 0.009160441346466541, + 0.007562341634184122, + 0.006988086737692356, + 0.009887713007628918, + 0.009036649018526077, + 0.014797540381550789, + 0.010587139055132866, + 0.011892537586390972, + 0.005342352204024792, + 0.00403918232768774, + 0.006506231613457203, + 0.004570721182972193, + 0.0070084924809634686, + 0.0038726169150322676, + 0.022984299808740616, + 0.027137840166687965, + 0.04205115884542465, + 0.06202155351638794, + 0.055812619626522064, + 0.047337938100099564, + 0.10803179442882538, + 0.04082297533750534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.5761253833770752 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a northern pike with olive-green and tawny scales speckled with pale spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment murky water containing rotting, waterlogged branches at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature faint rays of green light piercing the water and illuminating floating particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the pike perform an explosive strike with its tail whipping violently to create a swirling vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera tracking horizontally while capturing the pike's movement in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38701868057250977 + ], + "opposite_logprob_tokens": [ + -1.1370186805725098 + ], + "logprob_reward": -0.38701868057250977, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38701868057250977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07072839885950089, + 0.15725693106651306, + 0.06708575040102005, + 0.14753836393356323, + 0.04921510070562363, + 0.03960123658180237, + 0.18733108043670654, + 0.04663091152906418, + 0.05511018633842468, + 0.1795019507408142 + ], + "credit_map": [ + 0.009859762154519558, + 0.010138111189007759, + 0.009391549974679947, + 0.007329150103032589, + 0.011855972930788994, + 0.006698951590806246, + 0.004262116737663746, + 0.011192785575985909, + 0.0191496629267931, + 0.017615295946598053, + 0.01821574568748474, + 0.020335983484983444, + 0.020200565457344055, + 0.018962517380714417, + 0.016339194029569626, + 0.026437954977154732, + 0.009498858824372292, + 0.01094060204923153, + 0.007928370498120785, + 0.009199519641697407, + 0.009736040607094765, + 0.011009298264980316, + 0.003710491117089987, + 0.005062562879174948, + 0.022364435717463493, + 0.03167160600423813, + 0.019760198891162872, + 0.020126983523368835, + 0.016597339883446693, + 0.016182607039809227, + 0.01237513218075037, + 0.008460065349936485, + 0.006733111571520567, + 0.0052118729799985886, + 0.007206778507679701, + 0.006199977360665798, + 0.008898995816707611, + 0.006418504752218723, + 0.004977602977305651, + 0.0035682569723576307, + 0.0056149777956306934, + 0.00618881406262517, + 0.004772174637764692, + 0.004381407517939806, + 0.0039186798967421055, + 0.004695500712841749, + 0.005407998338341713, + 0.004621685948222876, + 0.042877621948719025, + 0.02045688033103943, + 0.053548894822597504, + 0.01738412119448185, + 0.012506985105574131, + 0.01477705966681242, + 0.013941586948931217, + 0.01183794904500246, + 0.005931518506258726, + 0.007139598019421101, + 0.005560252815485001, + 0.004965117201209068, + 0.006345431786030531, + 0.005337954964488745, + 0.00459700683131814, + 0.006754032801836729, + 0.01115527842193842, + 0.0119360676035285, + 0.0054331617429852486, + 0.004829935263842344, + 0.006255152635276318, + 0.004727475810796022, + 0.005820430815219879, + 0.00495268264785409, + 0.00978610198944807, + 0.014920865185558796, + 0.020626740530133247, + 0.02905420772731304, + 0.026889491826295853, + 0.02017776109278202, + 0.04008078575134277, + 0.017966002225875854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 1 + ], + "selected_reward": -0.38701868057250977 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_9/index_tensor([72, 72])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_9/index_tensor([72, 72])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2d56a47a6146737d933538c8a19b613888ab64e3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_9/index_tensor([72, 72])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, almost imperceptible low-angle close-up captures a massive great white shark, suspended motionless in the deep blue ocean. Its powerful body is tensed in a subtle S-curve, the coiled springs of muscle poised to unleash a devastating blow. The camera captures the hyper-detailed, rough texture of its skin, like rough gray-brown sandpaper, marked with a network of faint white scars from past encounters. Ethereal sunlight penetrates the water from above, casting a dramatic glow on its dorsal fin and back, while its underbelly remains a stark, shadowed white. Its black, lifeless eyes stare forward into space, its jaw slightly open, revealing rows of jagged, razor-sharp teeth. The only perceptible movement is the slow, rhythmic opening of its gills, disturbing the tiny particles of marine snow lazily drifting in the surrounding water. Rendered in a cool, deep blue palette, the scene creates an eerie and tense atmosphere of raw power and suspended, imminent violence.", + "seed": 23547, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_9/index_72_seed_23547_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03414833657443524, + "vlm_logprob": -0.03414833657443524, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.03414833657443524, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01261092908680439 + ], + "opposite_logprob_tokens": [ + -4.637610912322998 + ], + "logprob_reward": -0.01261092908680439, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01261092908680439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27759096026420593, + 0.08299830555915833, + 0.02846076339483261, + 0.18454723060131073, + 0.08821253478527069, + 0.11877328157424927, + 0.07831152528524399, + 0.05751455947756767, + 0.03267156705260277, + 0.05091935396194458 + ], + "credit_map": [ + 0.032135698944330215, + 0.022094473242759705, + 0.036622896790504456, + 0.03454139083623886, + 0.030241619795560837, + 0.029757807031273842, + 0.040864672511816025, + 0.05133238062262535, + 0.012776917777955532, + 0.005522006656974554, + 0.015406586229801178, + 0.009014575742185116, + 0.00843685120344162, + 0.009465581737458706, + 0.007412303704768419, + 0.01496347226202488, + 0.0025334693491458893, + 0.002824656665325165, + 0.004131713416427374, + 0.004373657517135143, + 0.0028484356589615345, + 0.004781046882271767, + 0.003408313961699605, + 0.003559466451406479, + 0.01761564239859581, + 0.020680341869592667, + 0.026888558641076088, + 0.021719519048929214, + 0.02863050438463688, + 0.02395019680261612, + 0.023482998833060265, + 0.0215794388204813, + 0.00953520368784666, + 0.008753154426813126, + 0.015638981014490128, + 0.012486066669225693, + 0.009097517468035221, + 0.011714600026607513, + 0.010047145187854767, + 0.010939853265881538, + 0.01221460197120905, + 0.01059877872467041, + 0.017079966142773628, + 0.01168946921825409, + 0.02434948831796646, + 0.017844462767243385, + 0.011933171190321445, + 0.013063328340649605, + 0.0055919913575053215, + 0.006221008021384478, + 0.017729243263602257, + 0.008445768617093563, + 0.006397377699613571, + 0.008022580295801163, + 0.0069934893399477005, + 0.018910055980086327, + 0.008355969563126564, + 0.005483862012624741, + 0.010072450153529644, + 0.008892267942428589, + 0.005699687637388706, + 0.00600182730704546, + 0.005125239957123995, + 0.007883249782025814, + 0.003076899563893676, + 0.004712200723588467, + 0.005278721917420626, + 0.0037822446320205927, + 0.004356251563876867, + 0.004201157484203577, + 0.0038767335936427116, + 0.0033873540814965963, + 0.005797486286610365, + 0.005030316300690174, + 0.008238236419856548, + 0.006452987436205149, + 0.005875919014215469, + 0.008451948873698711, + 0.005439538508653641, + 0.005632916931062937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.01261092908680439 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054976750165224075 + ], + "opposite_logprob_tokens": [ + -2.92997670173645 + ], + "logprob_reward": -0.054976750165224075, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054976750165224075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19453172385692596, + 0.12580519914627075, + 0.11574806272983551, + 0.14000040292739868, + 0.10356578975915909, + 0.04999905452132225, + 0.06880585104227066, + 0.08790042996406555, + 0.046881768852472305, + 0.06676172465085983 + ], + "credit_map": [ + 0.030428797006607056, + 0.015774263069033623, + 0.022761037573218346, + 0.022680098190903664, + 0.02316899225115776, + 0.023701447993516922, + 0.025976935401558876, + 0.030040141195058823, + 0.015740515664219856, + 0.007431219797581434, + 0.024459170177578926, + 0.011697239242494106, + 0.013860312290489674, + 0.021952452138066292, + 0.014090027660131454, + 0.01657426357269287, + 0.01674368605017662, + 0.010532009415328503, + 0.01596001535654068, + 0.014991137199103832, + 0.013201544992625713, + 0.017273113131523132, + 0.013092978857457638, + 0.013953578658401966, + 0.016827067360281944, + 0.010321841575205326, + 0.026109175756573677, + 0.010708571411669254, + 0.012843444012105465, + 0.01901213638484478, + 0.02137337625026703, + 0.02280478924512863, + 0.011325644329190254, + 0.008102597668766975, + 0.013260526582598686, + 0.012965880334377289, + 0.017541032284498215, + 0.01238388754427433, + 0.014164580032229424, + 0.013821635395288467, + 0.00576431630179286, + 0.004384579136967659, + 0.008397362194955349, + 0.005089891608804464, + 0.00851928535848856, + 0.005860719829797745, + 0.005025968886911869, + 0.006956926546990871, + 0.005445775110274553, + 0.007070411927998066, + 0.011288071051239967, + 0.009952625259757042, + 0.006221985910087824, + 0.014585425145924091, + 0.006768860388547182, + 0.0074726915918290615, + 0.01815876178443432, + 0.008379671722650528, + 0.016825862228870392, + 0.010867001488804817, + 0.00754854129627347, + 0.010973541997373104, + 0.005628290586173534, + 0.009518751874566078, + 0.004773091059178114, + 0.005229581613093615, + 0.0078016915358603, + 0.004993173759430647, + 0.007077320013195276, + 0.005476862657815218, + 0.006449186243116856, + 0.005080863833427429, + 0.007073456421494484, + 0.007017092313617468, + 0.009813359007239342, + 0.007758877705782652, + 0.010015245527029037, + 0.01145646907389164, + 0.006301645189523697, + 0.007325578946620226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.054976750165224075 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546772137284279 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.012546772137284279, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546772137284279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09541529417037964, + 0.1392098218202591, + 0.06774256378412247, + 0.17155417799949646, + 0.10537932813167572, + 0.06088925153017044, + 0.10206787288188934, + 0.07210200279951096, + 0.08132796734571457, + 0.10431177169084549 + ], + "credit_map": [ + 0.01588480733335018, + 0.008106116205453873, + 0.011789053678512573, + 0.014867566525936127, + 0.009226917289197445, + 0.013443069532513618, + 0.00992050115019083, + 0.01217726431787014, + 0.016463560983538628, + 0.010209781117737293, + 0.021352553740143776, + 0.024278884753584862, + 0.013961328193545341, + 0.01927131414413452, + 0.01697683520615101, + 0.016695557162165642, + 0.006364127621054649, + 0.006648199632763863, + 0.008204929530620575, + 0.010435724630951881, + 0.007450241595506668, + 0.009900041855871677, + 0.010606511496007442, + 0.00813278742134571, + 0.01675049588084221, + 0.018579311668872833, + 0.024382205680012703, + 0.021553808823227882, + 0.020035376772284508, + 0.024678602814674377, + 0.020518170669674873, + 0.025056205689907074, + 0.011190253309905529, + 0.010161284357309341, + 0.014250998385250568, + 0.013728227466344833, + 0.015992028638720512, + 0.015948358923196793, + 0.013002337887883186, + 0.011105837300419807, + 0.0060446299612522125, + 0.005886703729629517, + 0.010047475807368755, + 0.005795775447040796, + 0.00872004497796297, + 0.008786016143858433, + 0.005414535757154226, + 0.010194068774580956, + 0.012707080692052841, + 0.007622308563441038, + 0.01895654946565628, + 0.019085552543401718, + 0.009483713656663895, + 0.009522347711026669, + 0.00838010199368, + 0.01631022058427334, + 0.010163974016904831, + 0.006991423200815916, + 0.00977543368935585, + 0.010863572359085083, + 0.00829332321882248, + 0.010152117349207401, + 0.0077118356712162495, + 0.008150322362780571, + 0.007292504422366619, + 0.009616524912416935, + 0.015752792358398438, + 0.009306990541517735, + 0.012355385348200798, + 0.008805233053863049, + 0.010685995221138, + 0.007512541022151709, + 0.011694640852510929, + 0.011089054867625237, + 0.014958116225898266, + 0.01342198345810175, + 0.013416130095720291, + 0.017088355496525764, + 0.01030818186700344, + 0.012335310690104961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.012546772137284279 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058218978345394 + ], + "opposite_logprob_tokens": [ + -3.8955821990966797 + ], + "logprob_reward": -0.02058218978345394, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058218978345394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09454657137393951, + 0.11443641781806946, + 0.05879921838641167, + 0.07836193591356277, + 0.2924351096153259, + 0.0993327870965004, + 0.05000844597816467, + 0.05486562103033066, + 0.08730895072221756, + 0.06990493834018707 + ], + "credit_map": [ + 0.013356990180909634, + 0.008932342752814293, + 0.014335192739963531, + 0.013249744661152363, + 0.008679608814418316, + 0.00956272054463625, + 0.009539836086332798, + 0.01689012348651886, + 0.013618470169603825, + 0.010029498487710953, + 0.02010570652782917, + 0.013996094465255737, + 0.011143499985337257, + 0.016458099707961082, + 0.012308619916439056, + 0.016776423901319504, + 0.008171012625098228, + 0.005223497282713652, + 0.008475178852677345, + 0.008195046335458755, + 0.005582831799983978, + 0.00811748020350933, + 0.007504758890718222, + 0.007529411464929581, + 0.011884115636348724, + 0.007330556865781546, + 0.011337803676724434, + 0.009361755102872849, + 0.009608489461243153, + 0.008872684091329575, + 0.008950683288276196, + 0.011015844531357288, + 0.06882728636264801, + 0.018614573404192924, + 0.04065435379743576, + 0.030360018834471703, + 0.020922545343637466, + 0.037434790283441544, + 0.020627368241548538, + 0.054994165897369385, + 0.009815092198550701, + 0.005557410418987274, + 0.02232011966407299, + 0.008441082201898098, + 0.00671746488660574, + 0.013133248314261436, + 0.005607514642179012, + 0.0277408380061388, + 0.0056383335031569, + 0.0047709266655147076, + 0.010022750124335289, + 0.004630003124475479, + 0.004357680678367615, + 0.00581695418804884, + 0.005096070468425751, + 0.009675724431872368, + 0.004831423982977867, + 0.004824935924261808, + 0.009306651540100574, + 0.005683119874447584, + 0.0052329255267977715, + 0.010455599054694176, + 0.004674482624977827, + 0.00985647737979889, + 0.009206267073750496, + 0.009697666391730309, + 0.015593067742884159, + 0.009854639880359173, + 0.011886770837008953, + 0.008686709217727184, + 0.011965631507337093, + 0.010418189689517021, + 0.006905435584485531, + 0.008193234913051128, + 0.010196507908403873, + 0.006282172165811062, + 0.011350692249834538, + 0.012394082732498646, + 0.006523762363940477, + 0.008059046231210232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.02058218978345394 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002504169940948 + ], + "opposite_logprob_tokens": [ + -2.6950249671936035 + ], + "logprob_reward": -0.07002504169940948, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002504169940948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1465645581483841, + 0.11402571201324463, + 0.0936928316950798, + 0.050220340490341187, + 0.052094269543886185, + 0.06920754164457321, + 0.17779646813869476, + 0.06606720387935638, + 0.09133650362491608, + 0.13899458944797516 + ], + "credit_map": [ + 0.02828889712691307, + 0.012727987952530384, + 0.020666081458330154, + 0.015488363802433014, + 0.011372935026884079, + 0.016793765127658844, + 0.013543334789574146, + 0.02768317423760891, + 0.010666027665138245, + 0.011018134653568268, + 0.014616566710174084, + 0.01727384701371193, + 0.011029670014977455, + 0.01756848208606243, + 0.013265658169984818, + 0.018587308004498482, + 0.016914796084165573, + 0.008211066015064716, + 0.014753316529095173, + 0.010669385083019733, + 0.007785324938595295, + 0.011705495417118073, + 0.00903655681759119, + 0.014616883359849453, + 0.007120984140783548, + 0.0038109817542135715, + 0.0062390402890741825, + 0.005274842958897352, + 0.007236791308969259, + 0.008231207728385925, + 0.006302730645984411, + 0.0060037593357264996, + 0.009958066046237946, + 0.0037207549903541803, + 0.008281637914478779, + 0.006548165809363127, + 0.004097534343600273, + 0.0067352233454585075, + 0.005096143111586571, + 0.007656735368072987, + 0.008054464124143124, + 0.005323752760887146, + 0.009323409758508205, + 0.0053120204247534275, + 0.010576477274298668, + 0.013223699294030666, + 0.007728638127446175, + 0.009665076620876789, + 0.020122522488236427, + 0.01402515172958374, + 0.032568786293268204, + 0.021929318085312843, + 0.01369771733880043, + 0.024768127128481865, + 0.014954264275729656, + 0.03573055937886238, + 0.009622476063668728, + 0.0060209231451153755, + 0.009274185635149479, + 0.008857625536620617, + 0.007707854267209768, + 0.010361713357269764, + 0.006788160186260939, + 0.007434256840497255, + 0.01012030802667141, + 0.011862424202263355, + 0.016067320480942726, + 0.010458866134285927, + 0.012023376300930977, + 0.010331245139241219, + 0.011596097610890865, + 0.008876858279109001, + 0.014871334657073021, + 0.014998252503573895, + 0.01906774751842022, + 0.016853393986821175, + 0.01925019733607769, + 0.025406083092093468, + 0.01133857760578394, + 0.017208989709615707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.07002504169940948 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, almost imperceptible low-angle close-up captures a massive great white shark, suspended motionless in the deep blue ocean. Its powerful body is tensed in a subtle S-curve, the coiled springs of muscle poised to unleash a devastating blow. The camera captures the hyper-detailed, rough texture of its skin, like rough gray-brown sandpaper, marked with a network of faint white scars from past encounters. Ethereal sunlight penetrates the water from above, casting a dramatic glow on its dorsal fin and back, while its underbelly remains a stark, shadowed white. Its black, lifeless eyes stare forward into space, its jaw slightly open, revealing rows of jagged, razor-sharp teeth. The only perceptible movement is the slow, rhythmic opening of its gills, disturbing the tiny particles of marine snow lazily drifting in the surrounding water. Rendered in a cool, deep blue palette, the scene creates an eerie and tense atmosphere of raw power and suspended, imminent violence.", + "seed": 23548, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_9/index_72_seed_23548_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12493528127670288, + "vlm_logprob": -0.12493528127670288, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.12493528127670288, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14905908703804016 + ], + "opposite_logprob_tokens": [ + -2.0240590572357178 + ], + "logprob_reward": -0.14905908703804016, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14905908703804016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0969594269990921, + 0.1788361519575119, + 0.08287667483091354, + 0.10439228266477585, + 0.058483123779296875, + 0.08758031576871872, + 0.16288228332996368, + 0.059965580701828, + 0.10568992048501968, + 0.06233430281281471 + ], + "credit_map": [ + 0.006544553209096193, + 0.027098102495074272, + 0.02834753692150116, + 0.013059059157967567, + 0.005824679974466562, + 0.00501151941716671, + 0.006128501147031784, + 0.004945467226207256, + 0.018445027992129326, + 0.03303392231464386, + 0.03372593969106674, + 0.015321020036935806, + 0.020703542977571487, + 0.015526795759797096, + 0.022258074954152107, + 0.019821828231215477, + 0.013240671716630459, + 0.009661709889769554, + 0.010951116681098938, + 0.007015982177108526, + 0.008526813238859177, + 0.010467015206813812, + 0.009789075702428818, + 0.013224286958575249, + 0.011135058477520943, + 0.007697117980569601, + 0.014791669324040413, + 0.01312677189707756, + 0.01060295756906271, + 0.010050197131931782, + 0.01710481382906437, + 0.019883697852492332, + 0.009220851585268974, + 0.004599131643772125, + 0.011690671555697918, + 0.006577616091817617, + 0.005389898084104061, + 0.0059021031484007835, + 0.004963552113622427, + 0.010139291174709797, + 0.009159993380308151, + 0.006300788838416338, + 0.0098158810287714, + 0.01243628654628992, + 0.011292655020952225, + 0.007564242463558912, + 0.017081478610634804, + 0.013928978703916073, + 0.017223287373781204, + 0.02024402841925621, + 0.020831795409321785, + 0.020054055377840996, + 0.019333500415086746, + 0.01858193799853325, + 0.01677464134991169, + 0.029839014634490013, + 0.007127889897674322, + 0.007792487274855375, + 0.007769716903567314, + 0.007311174646019936, + 0.005712335463613272, + 0.007112374063581228, + 0.006417860276997089, + 0.010721741244196892, + 0.011140395887196064, + 0.007937497459352016, + 0.015356243588030338, + 0.011305277235805988, + 0.016142601147294044, + 0.015531755983829498, + 0.014522862620651722, + 0.013753288425505161, + 0.004715351853519678, + 0.007155129220336676, + 0.00983573030680418, + 0.010821419768035412, + 0.007186951581388712, + 0.0035425263922661543, + 0.010129720903933048, + 0.008947468362748623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.14905908703804016 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28139108419418335 + ], + "opposite_logprob_tokens": [ + -1.4063911437988281 + ], + "logprob_reward": -0.28139108419418335, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28139108419418335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10094035416841507, + 0.14210768043994904, + 0.0841808095574379, + 0.09248791635036469, + 0.13413140177726746, + 0.16610969603061676, + 0.10841020196676254, + 0.07383936643600464, + 0.055396631360054016, + 0.04239589348435402 + ], + "credit_map": [ + 0.005227243527770042, + 0.029321910813450813, + 0.030944187194108963, + 0.009022329933941364, + 0.006941702216863632, + 0.007834979332983494, + 0.005795767530798912, + 0.005852234549820423, + 0.015612770803272724, + 0.021777868270874023, + 0.02426845021545887, + 0.013752206228673458, + 0.01827874965965748, + 0.013942970894277096, + 0.018334917724132538, + 0.01613975688815117, + 0.008102396503090858, + 0.009332370012998581, + 0.01209829282015562, + 0.008298073895275593, + 0.010763689875602722, + 0.010929377749562263, + 0.010335493832826614, + 0.014321106486022472, + 0.008667775429785252, + 0.007210074923932552, + 0.013133096508681774, + 0.011058383621275425, + 0.010829214006662369, + 0.008291609585285187, + 0.016185063868761063, + 0.017112702131271362, + 0.018901174888014793, + 0.018299516290426254, + 0.02928907424211502, + 0.018095949664711952, + 0.014258919283747673, + 0.009744877927005291, + 0.011625117622315884, + 0.013916783034801483, + 0.01758892834186554, + 0.012887190096080303, + 0.019101973623037338, + 0.026738332584500313, + 0.025083553045988083, + 0.012438812293112278, + 0.02948027290403843, + 0.022790633141994476, + 0.011252922005951405, + 0.014308340847492218, + 0.013536701910197735, + 0.013146077282726765, + 0.013302445411682129, + 0.012498182244598866, + 0.011512952856719494, + 0.018852584064006805, + 0.008345120586454868, + 0.009583255276083946, + 0.011140383780002594, + 0.008064703084528446, + 0.00680506369099021, + 0.007951873354613781, + 0.008938295766711235, + 0.01301067415624857, + 0.005456444341689348, + 0.0042349505238235, + 0.010195913724601269, + 0.008660167455673218, + 0.004784110467880964, + 0.005349861923605204, + 0.00752859516069293, + 0.00918659009039402, + 0.0036480152048170567, + 0.007009119726717472, + 0.004488382022827864, + 0.006168070714920759, + 0.00373054132796824, + 0.0020762647036463022, + 0.008240234106779099, + 0.007035266142338514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.28139108419418335 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309263080358505 + ], + "opposite_logprob_tokens": [ + -3.168092727661133 + ], + "logprob_reward": -0.04309263080358505, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309263080358505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07245630025863647, + 0.09935697168111801, + 0.11310471594333649, + 0.09977764636278152, + 0.19184856116771698, + 0.069799043238163, + 0.1406194418668747, + 0.0681217834353447, + 0.08477585017681122, + 0.060139670968055725 + ], + "credit_map": [ + 0.006195397581905127, + 0.011733286082744598, + 0.01090852078050375, + 0.009076941758394241, + 0.009931724518537521, + 0.01119922660291195, + 0.006276356987655163, + 0.007134841289371252, + 0.01201836857944727, + 0.012317812070250511, + 0.013890724629163742, + 0.010929636657238007, + 0.010146320797502995, + 0.009821761399507523, + 0.015472468920052052, + 0.014759880490601063, + 0.01305083092302084, + 0.015972234308719635, + 0.016903935000300407, + 0.008741086348891258, + 0.013948877342045307, + 0.012223352678120136, + 0.012459606863558292, + 0.01980479061603546, + 0.010388622991740704, + 0.008084115572273731, + 0.014142177067697048, + 0.0080989059060812, + 0.011428993195295334, + 0.00938597321510315, + 0.017248423770070076, + 0.021000433713197708, + 0.024537300691008568, + 0.025736898183822632, + 0.04258447512984276, + 0.023551728576421738, + 0.019377263262867928, + 0.013641932047903538, + 0.018783191218972206, + 0.023635759949684143, + 0.00826887134462595, + 0.005619335453957319, + 0.007540963590145111, + 0.010356043465435505, + 0.009606368839740753, + 0.0057585169561207294, + 0.012073143385350704, + 0.010575799271464348, + 0.01481618545949459, + 0.015246395952999592, + 0.016095774248242378, + 0.015617342665791512, + 0.015721632167696953, + 0.01496429555118084, + 0.019582539796829224, + 0.028575286269187927, + 0.007139029446989298, + 0.0071835508570075035, + 0.010335213504731655, + 0.00690647354349494, + 0.0061842468567192554, + 0.007867859676480293, + 0.011778529733419418, + 0.010726877488195896, + 0.008343765512108803, + 0.0056083835661411285, + 0.016737185418605804, + 0.015314825810492039, + 0.00815162155777216, + 0.006721034180372953, + 0.012255140580236912, + 0.011643892154097557, + 0.0046522291377186775, + 0.009004835039377213, + 0.010111814364790916, + 0.011194960214197636, + 0.00596552575007081, + 0.002928303787484765, + 0.009307635016739368, + 0.006974367890506983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.04309263080358505 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206654757261276 + ], + "opposite_logprob_tokens": [ + -2.8120665550231934 + ], + "logprob_reward": -0.06206654757261276, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206654757261276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07828814536333084, + 0.1449565291404724, + 0.09710166603326797, + 0.04206680506467819, + 0.08543268591165543, + 0.09719859063625336, + 0.10225260257720947, + 0.11050447076559067, + 0.15931545197963715, + 0.0828830674290657 + ], + "credit_map": [ + 0.007466568145900965, + 0.009181850589811802, + 0.008889489807188511, + 0.015612125396728516, + 0.009312706999480724, + 0.006302302703261375, + 0.008567588403820992, + 0.012955506332218647, + 0.015311594121158123, + 0.028940685093402863, + 0.028021447360515594, + 0.01304430142045021, + 0.01535128429532051, + 0.011142728850245476, + 0.01612463779747486, + 0.017019841820001602, + 0.011041803285479546, + 0.010218947194516659, + 0.012122362852096558, + 0.012901261448860168, + 0.011021852493286133, + 0.012015808373689651, + 0.011637129820883274, + 0.01614249311387539, + 0.0038890361320227385, + 0.004072351846843958, + 0.005266914609819651, + 0.004335145000368357, + 0.005242192652076483, + 0.004339542239904404, + 0.006438813172280788, + 0.008482805453240871, + 0.01095298957079649, + 0.008872553706169128, + 0.017992371693253517, + 0.010106666944921017, + 0.009053470566868782, + 0.006559866946190596, + 0.009959046728909016, + 0.011935717426240444, + 0.011368042789399624, + 0.006409868597984314, + 0.011663299985229969, + 0.011833149008452892, + 0.010154837742447853, + 0.01040086429566145, + 0.016812050715088844, + 0.018556473776698112, + 0.011670003645122051, + 0.008824465796351433, + 0.009138636291027069, + 0.012028971686959267, + 0.00785177480429411, + 0.00973230879753828, + 0.014946281909942627, + 0.02806016057729721, + 0.011681906878948212, + 0.009957429952919483, + 0.017922818660736084, + 0.011897201649844646, + 0.009982537478208542, + 0.009501569904386997, + 0.015409617684781551, + 0.024151386693120003, + 0.014479942619800568, + 0.011691604740917683, + 0.025714462623000145, + 0.024846777319908142, + 0.020497852936387062, + 0.01886945590376854, + 0.02092302404344082, + 0.02229233644902706, + 0.0076772915199398994, + 0.01070176437497139, + 0.011073295027017593, + 0.014008620753884315, + 0.007375504355877638, + 0.007705597672611475, + 0.01325646135956049, + 0.011084521189332008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.06206654757261276 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906705677509308 + ], + "opposite_logprob_tokens": [ + -2.464066982269287 + ], + "logprob_reward": -0.08906705677509308, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906705677509308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11815652996301651, + 0.05615880340337753, + 0.08452204614877701, + 0.05618986114859581, + 0.21864904463291168, + 0.09101651608943939, + 0.06992964446544647, + 0.11734208464622498, + 0.09306004643440247, + 0.09497540444135666 + ], + "credit_map": [ + 0.011250774376094341, + 0.028654849156737328, + 0.023417305201292038, + 0.012976461090147495, + 0.009663390927016735, + 0.010077113285660744, + 0.010530875064432621, + 0.011585762724280357, + 0.006351217161864042, + 0.016415199264883995, + 0.012735349126160145, + 0.004245829302817583, + 0.003145997179672122, + 0.004275322426110506, + 0.003412409918382764, + 0.005577480420470238, + 0.008897623978555202, + 0.008741451427340508, + 0.00998048484325409, + 0.011300139129161835, + 0.00969001092016697, + 0.010428240522742271, + 0.010805658996105194, + 0.014678435400128365, + 0.006363514345139265, + 0.004972635302692652, + 0.007706738542765379, + 0.01060341577976942, + 0.005259792786091566, + 0.005035413429141045, + 0.008779882453382015, + 0.0074684699065983295, + 0.030102677643299103, + 0.024876058101654053, + 0.04570208489894867, + 0.02755460888147354, + 0.02034319005906582, + 0.01852848194539547, + 0.01724872924387455, + 0.03429320827126503, + 0.011482409201562405, + 0.006407574284821749, + 0.010017154738307, + 0.011301481164991856, + 0.007817784324288368, + 0.010795720852911472, + 0.01242498867213726, + 0.020769404247403145, + 0.008206343278288841, + 0.008568866178393364, + 0.008381177671253681, + 0.008992680348455906, + 0.008214428089559078, + 0.00750882551074028, + 0.007848242297768593, + 0.012209074571728706, + 0.010592370294034481, + 0.007985045202076435, + 0.013051110319793224, + 0.0118082445114851, + 0.007743852213025093, + 0.009283372201025486, + 0.029080938547849655, + 0.02779715321958065, + 0.009711414575576782, + 0.0068918392062187195, + 0.014500525780022144, + 0.013099159114062786, + 0.012375053018331528, + 0.00965521577745676, + 0.011742258444428444, + 0.015084575861692429, + 0.0054308753460645676, + 0.010735869407653809, + 0.009079302661120892, + 0.013452444225549698, + 0.006559708621352911, + 0.004481372889131308, + 0.0221274271607399, + 0.023108404129743576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.08906705677509308 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, almost imperceptible low-angle close-up captures a massive great white shark, suspended motionless in the deep blue ocean. Its powerful body is tensed in a subtle S-curve, the coiled springs of muscle poised to unleash a devastating blow. The camera captures the hyper-detailed, rough texture of its skin, like rough gray-brown sandpaper, marked with a network of faint white scars from past encounters. Ethereal sunlight penetrates the water from above, casting a dramatic glow on its dorsal fin and back, while its underbelly remains a stark, shadowed white. Its black, lifeless eyes stare forward into space, its jaw slightly open, revealing rows of jagged, razor-sharp teeth. The only perceptible movement is the slow, rhythmic opening of its gills, disturbing the tiny particles of marine snow lazily drifting in the surrounding water. Rendered in a cool, deep blue palette, the scene creates an eerie and tense atmosphere of raw power and suspended, imminent violence.", + "seed": 23549, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_9/index_72_seed_23549_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1439865216612816, + "vlm_logprob": -0.1439865216612816, + "vlm_yes_no_margin": 2.125, + "vlm_token_credit": -0.1439865216612816, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25678086280822754 + ], + "opposite_logprob_tokens": [ + -1.5067808628082275 + ], + "logprob_reward": -0.25678086280822754, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25678086280822754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20382337272167206, + 0.0699443519115448, + 0.09702512621879578, + 0.08104752004146576, + 0.22745086252689362, + 0.056926414370536804, + 0.05012204870581627, + 0.05884908139705658, + 0.11365096271038055, + 0.04116020351648331 + ], + "credit_map": [ + 0.061589375138282776, + 0.026067445054650307, + 0.04962752014398575, + 0.017970655113458633, + 0.011272636242210865, + 0.015126434154808521, + 0.009426587261259556, + 0.012742714956402779, + 0.009892006404697895, + 0.008248155005276203, + 0.010371607728302479, + 0.008257304318249226, + 0.006785244215279818, + 0.011255813762545586, + 0.006494917441159487, + 0.008639304898679256, + 0.017414921894669533, + 0.01120875682681799, + 0.012409774586558342, + 0.010016295127570629, + 0.008966865949332714, + 0.01538761518895626, + 0.007660937961190939, + 0.01395996380597353, + 0.011358700692653656, + 0.009734119288623333, + 0.01062039751559496, + 0.008468370884656906, + 0.010072796605527401, + 0.010984671302139759, + 0.007585977204144001, + 0.012222490273416042, + 0.038094449788331985, + 0.02830527536571026, + 0.030282987281680107, + 0.015036292374134064, + 0.027045056223869324, + 0.030706314370036125, + 0.016569720581173897, + 0.041410788893699646, + 0.010399121791124344, + 0.009019676595926285, + 0.008573214523494244, + 0.005420471541583538, + 0.0057136910036206245, + 0.006925888359546661, + 0.00472881319001317, + 0.006145537830889225, + 0.00997473020106554, + 0.010167954489588737, + 0.007021654862910509, + 0.004874228034168482, + 0.00576101103797555, + 0.00418400764465332, + 0.0031691929325461388, + 0.0049692667089402676, + 0.009371215477585793, + 0.008602814748883247, + 0.006458850111812353, + 0.007265893742442131, + 0.008084312081336975, + 0.008114177733659744, + 0.005004338454455137, + 0.005947481840848923, + 0.020468806847929955, + 0.011830992996692657, + 0.023941101506352425, + 0.012158469296991825, + 0.011131448671221733, + 0.016288714483380318, + 0.006879846565425396, + 0.010951580479741096, + 0.004940061829984188, + 0.004504031501710415, + 0.005782447289675474, + 0.005767622031271458, + 0.005563817918300629, + 0.006878349464386702, + 0.0035352595150470734, + 0.004188615828752518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.25678086280822754 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521405518054962 + ], + "opposite_logprob_tokens": [ + -1.5021405220031738 + ], + "logprob_reward": -0.2521405518054962, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521405518054962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17768187820911407, + 0.05760756507515907, + 0.08917944133281708, + 0.2260104864835739, + 0.18438152968883514, + 0.0420764796435833, + 0.03336425870656967, + 0.0824786126613617, + 0.06778770685195923, + 0.039431992918252945 + ], + "credit_map": [ + 0.045609522610902786, + 0.02375524677336216, + 0.03663499280810356, + 0.017632436007261276, + 0.012818996794521809, + 0.01848987676203251, + 0.011334444396197796, + 0.011406362056732178, + 0.00641367444768548, + 0.005857325624674559, + 0.007152581587433815, + 0.0064047155901789665, + 0.0066091688349843025, + 0.011274661868810654, + 0.005446361843496561, + 0.008449072949588299, + 0.01616019196808338, + 0.008305821567773819, + 0.015185549855232239, + 0.008902870118618011, + 0.00839548371732235, + 0.012122659012675285, + 0.00884167104959488, + 0.011265202425420284, + 0.030212143436074257, + 0.027015158906579018, + 0.02648550644516945, + 0.0232025608420372, + 0.027376146987080574, + 0.02996189519762993, + 0.022547803819179535, + 0.0392092689871788, + 0.03054305911064148, + 0.024120429530739784, + 0.025406869128346443, + 0.011698439717292786, + 0.020953135564923286, + 0.024714326485991478, + 0.013363562524318695, + 0.03358171507716179, + 0.008439778350293636, + 0.006592452060431242, + 0.007108824793249369, + 0.0033059888519346714, + 0.00433784956112504, + 0.0043835933320224285, + 0.003242986975237727, + 0.004665004555135965, + 0.006094185169786215, + 0.003934767097234726, + 0.004731489811092615, + 0.0037209931761026382, + 0.0036062931176275015, + 0.004810086917132139, + 0.002846830990165472, + 0.003619611030444503, + 0.01611207239329815, + 0.010807910934090614, + 0.012579360976815224, + 0.008511863648891449, + 0.010672852396965027, + 0.010838165879249573, + 0.00608504842966795, + 0.006871337071061134, + 0.008622960187494755, + 0.009303119964897633, + 0.010624565184116364, + 0.008909112773835659, + 0.007783687207847834, + 0.011661745607852936, + 0.004494838882237673, + 0.006387682631611824, + 0.005112133454531431, + 0.0044187274761497974, + 0.005606569349765778, + 0.006801755633205175, + 0.004845617339015007, + 0.005674907471984625, + 0.0032721806783229113, + 0.003700101748108864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.2521405518054962 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308018460869789 + ], + "opposite_logprob_tokens": [ + -3.1680800914764404 + ], + "logprob_reward": -0.04308018460869789, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308018460869789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1988152414560318, + 0.05403713509440422, + 0.04372373968362808, + 0.1387399137020111, + 0.1394442915916443, + 0.16499215364456177, + 0.03822297230362892, + 0.09425376355648041, + 0.0608823336660862, + 0.06688851118087769 + ], + "credit_map": [ + 0.06196439638733864, + 0.022326845675706863, + 0.04313427209854126, + 0.017438668757677078, + 0.012482591904699802, + 0.01585170067846775, + 0.009558520279824734, + 0.016058258712291718, + 0.00932757556438446, + 0.006457890383899212, + 0.007328198757022619, + 0.004786323755979538, + 0.0057209935039281845, + 0.010040966793894768, + 0.004937651101499796, + 0.005437532905489206, + 0.007186421193182468, + 0.006266359239816666, + 0.004787817131727934, + 0.004285069648176432, + 0.004735376685857773, + 0.007264011073857546, + 0.0037626472767442465, + 0.005436035804450512, + 0.01642078533768654, + 0.015602155588567257, + 0.01589912734925747, + 0.01179010421037674, + 0.01876220293343067, + 0.023761959746479988, + 0.013219711370766163, + 0.023283861577510834, + 0.015085484832525253, + 0.013481051661074162, + 0.02228013426065445, + 0.013900878839194775, + 0.018779313191771507, + 0.021945377811789513, + 0.013248827308416367, + 0.02072322368621826, + 0.03993837162852287, + 0.0266554057598114, + 0.01885833777487278, + 0.012775273062288761, + 0.014963432215154171, + 0.015333355404436588, + 0.013481630012392998, + 0.022986343130469322, + 0.008763980120420456, + 0.005435122177004814, + 0.004406534135341644, + 0.004592273384332657, + 0.003633672371506691, + 0.004670425783842802, + 0.0025969306007027626, + 0.00412403279915452, + 0.015186958946287632, + 0.011871352791786194, + 0.013154316693544388, + 0.012441974133253098, + 0.012409618124365807, + 0.013415239751338959, + 0.007229070644825697, + 0.008545232936739922, + 0.007833000272512436, + 0.0077118161134421825, + 0.010522360913455486, + 0.0070039983838796616, + 0.00791210401803255, + 0.010270046070218086, + 0.0041028279811143875, + 0.005526180844753981, + 0.00723620830103755, + 0.007687348872423172, + 0.009560530073940754, + 0.010127131827175617, + 0.008575662970542908, + 0.010128798894584179, + 0.005393262021243572, + 0.008179563097655773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.04308018460869789 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497449263930321 + ], + "opposite_logprob_tokens": [ + -2.9299745559692383 + ], + "logprob_reward": -0.05497449263930321, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497449263930321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12155811488628387, + 0.11776834726333618, + 0.04814613610506058, + 0.06463908404111862, + 0.19939027726650238, + 0.09474416822195053, + 0.042392976582050323, + 0.0674944594502449, + 0.04448416456580162, + 0.1993822306394577 + ], + "credit_map": [ + 0.024517301470041275, + 0.017034871503710747, + 0.01830139383673668, + 0.014322549104690552, + 0.010221848264336586, + 0.01289566420018673, + 0.009500066749751568, + 0.014764425344765186, + 0.024334661662578583, + 0.01379129197448492, + 0.0185125432908535, + 0.012791957706212997, + 0.009678374044597149, + 0.015994714573025703, + 0.010700981132686138, + 0.011963828466832638, + 0.00639871321618557, + 0.0071310256607830524, + 0.006660589016973972, + 0.006586573086678982, + 0.005189686547964811, + 0.005953765939921141, + 0.005003006663173437, + 0.005222775973379612, + 0.009428984485566616, + 0.007444099523127079, + 0.00791176874190569, + 0.005542343948036432, + 0.007965193130075932, + 0.011458220891654491, + 0.006023633759468794, + 0.008864844217896461, + 0.04219497740268707, + 0.026322757825255394, + 0.027824046090245247, + 0.011850251816213131, + 0.015361741185188293, + 0.021324193105101585, + 0.017524566501379013, + 0.036987729370594025, + 0.01510559767484665, + 0.011594744399189949, + 0.016126958653330803, + 0.00781477801501751, + 0.011212022043764591, + 0.010886847972869873, + 0.007963784039020538, + 0.014039435423910618, + 0.007077775429934263, + 0.005886958912014961, + 0.005421273875981569, + 0.005962654948234558, + 0.004865590017288923, + 0.0060983444564044476, + 0.0035568936727941036, + 0.003523486200720072, + 0.013170872814953327, + 0.007805649656802416, + 0.010750377550721169, + 0.00578543683513999, + 0.007922197692096233, + 0.009825976565480232, + 0.0046991994604468346, + 0.007534750737249851, + 0.008210881613194942, + 0.004714709706604481, + 0.008762168698012829, + 0.003616226138547063, + 0.006019372493028641, + 0.006530642043799162, + 0.002652578055858612, + 0.003977584652602673, + 0.02847667969763279, + 0.017435410991311073, + 0.030730413272976875, + 0.02594643272459507, + 0.01606646366417408, + 0.04572680592536926, + 0.012634126469492912, + 0.02236589789390564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.05497449263930321 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295651644468307 + ], + "opposite_logprob_tokens": [ + -2.2379565238952637 + ], + "logprob_reward": -0.11295651644468307, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295651644468307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20604071021080017, + 0.07180235534906387, + 0.1351633071899414, + 0.07905244082212448, + 0.08878839761018753, + 0.14728981256484985, + 0.05171934887766838, + 0.04743729531764984, + 0.050530869513750076, + 0.12217555940151215 + ], + "credit_map": [ + 0.0759534165263176, + 0.022496458142995834, + 0.05414757505059242, + 0.013209468685090542, + 0.007432546466588974, + 0.011540261097252369, + 0.006462795194238424, + 0.014798191376030445, + 0.019942617043852806, + 0.01042968314141035, + 0.01546074915677309, + 0.005078875459730625, + 0.0046483054757118225, + 0.006674239877611399, + 0.004255916457623243, + 0.005311967805027962, + 0.04922562092542648, + 0.017379865050315857, + 0.03782711178064346, + 0.006925636436790228, + 0.005887093022465706, + 0.006671342998743057, + 0.0050062560476362705, + 0.006240376271307468, + 0.011684591881930828, + 0.010486171580851078, + 0.009781673550605774, + 0.006611709948629141, + 0.009809569455683231, + 0.011060470715165138, + 0.00569217698648572, + 0.013926076702773571, + 0.014253958128392696, + 0.010301437228918076, + 0.013368825428187847, + 0.008234195411205292, + 0.009885136038064957, + 0.012357150204479694, + 0.007496585138142109, + 0.012891112826764584, + 0.02972487360239029, + 0.02030709758400917, + 0.023087067529559135, + 0.011439033783972263, + 0.015374570153653622, + 0.016130277886986732, + 0.011364763602614403, + 0.01986212283372879, + 0.014402602799236774, + 0.007192167919129133, + 0.005482979118824005, + 0.00391419930383563, + 0.005694095510989428, + 0.004878905136138201, + 0.003329443046823144, + 0.006824955809861422, + 0.0077384705655276775, + 0.007003622129559517, + 0.006649569608271122, + 0.0057488176971673965, + 0.006833790801465511, + 0.00560990534722805, + 0.0036532043013721704, + 0.004199916962534189, + 0.008900785818696022, + 0.007465551141649485, + 0.009588556364178658, + 0.004330717958509922, + 0.005163930356502533, + 0.007809906732290983, + 0.0029485553968697786, + 0.004322863183915615, + 0.026850631460547447, + 0.011933919973671436, + 0.02317219413816929, + 0.011106516234576702, + 0.006968778092414141, + 0.026129597797989845, + 0.004424303770065308, + 0.011589625850319862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.11295651644468307 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, almost imperceptible low-angle close-up captures a massive great white shark, suspended motionless in the deep blue ocean. Its powerful body is tensed in a subtle S-curve, the coiled springs of muscle poised to unleash a devastating blow. The camera captures the hyper-detailed, rough texture of its skin, like rough gray-brown sandpaper, marked with a network of faint white scars from past encounters. Ethereal sunlight penetrates the water from above, casting a dramatic glow on its dorsal fin and back, while its underbelly remains a stark, shadowed white. Its black, lifeless eyes stare forward into space, its jaw slightly open, revealing rows of jagged, razor-sharp teeth. The only perceptible movement is the slow, rhythmic opening of its gills, disturbing the tiny particles of marine snow lazily drifting in the surrounding water. Rendered in a cool, deep blue palette, the scene creates an eerie and tense atmosphere of raw power and suspended, imminent violence.", + "seed": 23550, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000154/group_9/index_72_seed_23550_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03592195548117161, + "vlm_logprob": -0.03592195548117161, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.03592195548117161, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022297630086541176 + ], + "opposite_logprob_tokens": [ + -4.0222978591918945 + ], + "logprob_reward": -0.022297630086541176, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.022297630086541176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1447487622499466, + 0.04806432127952576, + 0.029445065185427666, + 0.05028385668992996, + 0.057210102677345276, + 0.07168709486722946, + 0.06618793308734894, + 0.22047239542007446, + 0.2819206118583679, + 0.029979819431900978 + ], + "credit_map": [ + 0.044147782027721405, + 0.009316546842455864, + 0.04199611023068428, + 0.013328011147677898, + 0.008225923404097557, + 0.006699037738144398, + 0.008753230795264244, + 0.012282132171094418, + 0.006137038581073284, + 0.004086685366928577, + 0.00640594307333231, + 0.007165237329900265, + 0.00589369609951973, + 0.0038255557883530855, + 0.006260236259549856, + 0.008289930410683155, + 0.003836488351225853, + 0.0026169868651777506, + 0.0036066966131329536, + 0.0033593473490327597, + 0.0027731216978281736, + 0.0023114599753171206, + 0.0036978537682443857, + 0.007243111729621887, + 0.006493565626442432, + 0.008068534545600414, + 0.006202373653650284, + 0.0073359012603759766, + 0.00522227818146348, + 0.00477683637291193, + 0.0050980220548808575, + 0.0070863445289433, + 0.007996869273483753, + 0.007908466272056103, + 0.01159366499632597, + 0.007869450375437737, + 0.004346244037151337, + 0.002488016616553068, + 0.006316404789686203, + 0.008690988644957542, + 0.008988278917968273, + 0.012901478447020054, + 0.015818515792489052, + 0.007922027260065079, + 0.00452575134113431, + 0.0036052651703357697, + 0.005581589415669441, + 0.01234418898820877, + 0.007500315550714731, + 0.005743579939007759, + 0.008662542328238487, + 0.008880652487277985, + 0.009321952238678932, + 0.00687697809189558, + 0.008646347559988499, + 0.010555570013821125, + 0.028430039063096046, + 0.03577651455998421, + 0.028675060719251633, + 0.03529490903019905, + 0.022198639810085297, + 0.023716066032648087, + 0.020149629563093185, + 0.0262315571308136, + 0.042387254536151886, + 0.05846469849348068, + 0.034431666135787964, + 0.027326034381985664, + 0.02732749469578266, + 0.028988245874643326, + 0.02197953499853611, + 0.04101569205522537, + 0.003649989841505885, + 0.002452288521453738, + 0.0037614295724779367, + 0.0047521633096039295, + 0.002370858332142234, + 0.0036346486303955317, + 0.004028814844787121, + 0.005329628475010395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.022297630086541176 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867691919207573 + ], + "opposite_logprob_tokens": [ + -3.0486769676208496 + ], + "logprob_reward": -0.04867691919207573, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867691919207573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1578127145767212, + 0.04703177884221077, + 0.02515573613345623, + 0.035977285355329514, + 0.0343053936958313, + 0.028028052300214767, + 0.11021146923303604, + 0.31006041169166565, + 0.12965607643127441, + 0.12176109850406647 + ], + "credit_map": [ + 0.04852180555462837, + 0.010866931639611721, + 0.04926159605383873, + 0.01006025169044733, + 0.0078508285805583, + 0.007200160995125771, + 0.007940986193716526, + 0.01611015386879444, + 0.006537366658449173, + 0.0037479004822671413, + 0.0068422891199588776, + 0.00599467009305954, + 0.005102669820189476, + 0.003823604667559266, + 0.006027420982718468, + 0.008955854922533035, + 0.003307335078716278, + 0.0021304984111338854, + 0.00401341961696744, + 0.0035510496236383915, + 0.003075143089517951, + 0.0023482274264097214, + 0.002909383038058877, + 0.0038206775207072496, + 0.005509593058377504, + 0.004349168390035629, + 0.0040609631687402725, + 0.004549962468445301, + 0.0038143431302160025, + 0.0030229378025978804, + 0.004081894177943468, + 0.006588423158973455, + 0.004505801945924759, + 0.003527590539306402, + 0.004814659245312214, + 0.0051932791247963905, + 0.0033223528880625963, + 0.0031119610648602247, + 0.0048957462422549725, + 0.0049340007826685905, + 0.0036827356088906527, + 0.0023508358281105757, + 0.0035291570238769054, + 0.0036891258787363768, + 0.0035344092175364494, + 0.0028361710719764233, + 0.003980204463005066, + 0.004425411112606525, + 0.0132001843303442, + 0.009834195487201214, + 0.012609561905264854, + 0.01412881352007389, + 0.017862658947706223, + 0.010389341041445732, + 0.016854312270879745, + 0.015332388691604137, + 0.038895148783922195, + 0.05087713524699211, + 0.030997972935438156, + 0.04349812492728233, + 0.03559703752398491, + 0.03411983326077461, + 0.03791981562972069, + 0.038155313581228256, + 0.021038128063082695, + 0.008794319815933704, + 0.011005912907421589, + 0.010379966348409653, + 0.008232691325247288, + 0.01180356740951538, + 0.014556901529431343, + 0.04384458437561989, + 0.015475095249712467, + 0.008039919659495354, + 0.014703866094350815, + 0.012234960682690144, + 0.010151918046176434, + 0.013295527547597885, + 0.0189211405813694, + 0.028938662260770798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.04867691919207573 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329319901764393 + ], + "opposite_logprob_tokens": [ + -3.7732932567596436 + ], + "logprob_reward": -0.02329319901764393, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329319901764393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1088525801897049, + 0.04367396980524063, + 0.03914913162589073, + 0.06603772938251495, + 0.0705912709236145, + 0.03899632766842842, + 0.09765520691871643, + 0.1317206472158432, + 0.36940470337867737, + 0.03391842544078827 + ], + "credit_map": [ + 0.02838105894625187, + 0.009500650689005852, + 0.028668874874711037, + 0.008472190238535404, + 0.008793491870164871, + 0.00737463403493166, + 0.006943699438124895, + 0.010717988945543766, + 0.005764940287917852, + 0.003429756034165621, + 0.004977511242032051, + 0.004615127108991146, + 0.00682769576087594, + 0.004619758576154709, + 0.005314697977155447, + 0.008124488405883312, + 0.0055440207943320274, + 0.0035590140614658594, + 0.0050613475032150745, + 0.0055829198099672794, + 0.004991198889911175, + 0.004106197971850634, + 0.005087610334157944, + 0.00521682808175683, + 0.008466715924441814, + 0.0058577535673975945, + 0.005450445227324963, + 0.008475353941321373, + 0.007126240991055965, + 0.006882021203637123, + 0.01044339220970869, + 0.01333582028746605, + 0.009553099982440472, + 0.00834091380238533, + 0.008597114123404026, + 0.009576378390192986, + 0.006890438497066498, + 0.006743662990629673, + 0.011628200300037861, + 0.009261475875973701, + 0.0057572550140321255, + 0.005975498352199793, + 0.006819775328040123, + 0.0038399153854697943, + 0.003974270075559616, + 0.0037296738009899855, + 0.00360396271571517, + 0.005295983050018549, + 0.012510422617197037, + 0.00864851288497448, + 0.009942437522113323, + 0.009552516043186188, + 0.01309526339173317, + 0.008245616219937801, + 0.01580696366727352, + 0.019853482022881508, + 0.015683993697166443, + 0.013590997084975243, + 0.012678463011980057, + 0.012281683273613453, + 0.012932633981108665, + 0.014421247877180576, + 0.020261334255337715, + 0.029870308935642242, + 0.05555442348122597, + 0.037006035447120667, + 0.02700864151120186, + 0.02619556523859501, + 0.026606731116771698, + 0.037570223212242126, + 0.03730541095137596, + 0.12215770781040192, + 0.004255320411175489, + 0.001808668952435255, + 0.004175360314548016, + 0.003708598902449012, + 0.0022956253960728645, + 0.004368965979665518, + 0.004050662741065025, + 0.009255227632820606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.02329319901764393 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292848840355873 + ], + "opposite_logprob_tokens": [ + -3.7732927799224854 + ], + "logprob_reward": -0.023292848840355873, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292848840355873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23091594874858856, + 0.04487426206469536, + 0.06306707113981247, + 0.04454335942864418, + 0.08290228247642517, + 0.07928118109703064, + 0.12827298045158386, + 0.15635506808757782, + 0.09370853006839752, + 0.07607930153608322 + ], + "credit_map": [ + 0.06352916359901428, + 0.02297796867787838, + 0.06317277997732162, + 0.017684699967503548, + 0.01646219566464424, + 0.011714148335158825, + 0.013939572498202324, + 0.02143542468547821, + 0.006431396584957838, + 0.0042226784862577915, + 0.007722633890807629, + 0.005687030963599682, + 0.00489317812025547, + 0.0035753839183598757, + 0.00488748773932457, + 0.0074544730596244335, + 0.008682423271238804, + 0.006234412547200918, + 0.007527684792876244, + 0.006405836436897516, + 0.008646315895020962, + 0.006209490355104208, + 0.006347427144646645, + 0.013013479299843311, + 0.006928496528416872, + 0.005835456773638725, + 0.006172250024974346, + 0.005069495644420385, + 0.005804646760225296, + 0.0038623649161309004, + 0.00536107225343585, + 0.0055095781572163105, + 0.011871838942170143, + 0.014869099482893944, + 0.012476039119064808, + 0.009362400509417057, + 0.007414746563881636, + 0.00593610992655158, + 0.007760853972285986, + 0.013211197219789028, + 0.014660684391856194, + 0.008088622242212296, + 0.02293495647609234, + 0.00855959951877594, + 0.006598338019102812, + 0.0051097809337079525, + 0.0062078023329377174, + 0.00712139718234539, + 0.019557379186153412, + 0.01020892895758152, + 0.013712349347770214, + 0.01176499668508768, + 0.012597384862601757, + 0.010510792955756187, + 0.018984569236636162, + 0.030936576426029205, + 0.01766643300652504, + 0.02198524959385395, + 0.016007574275135994, + 0.021073363721370697, + 0.016173340380191803, + 0.016665898263454437, + 0.02065274305641651, + 0.026130471378564835, + 0.014744168147444725, + 0.020915860310196877, + 0.012553551234304905, + 0.009709428995847702, + 0.009708777070045471, + 0.009208044968545437, + 0.006477378308773041, + 0.010391321033239365, + 0.008646165952086449, + 0.007955619134008884, + 0.009205491282045841, + 0.011089787818491459, + 0.007399225607514381, + 0.010725878179073334, + 0.009210933931171894, + 0.01184619590640068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.023292848840355873 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's skin rendered with a rough, gray-brown texture resembling sandpaper and marked with faint white scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shark's body tensed in a subtle S-curve with its jaw slightly open to reveal rows of jagged, razor-sharp teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature ethereal sunlight penetrating from above, casting a dramatic glow on the dorsal fin and back while leaving the underbelly in stark shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the slow, rhythmic opening of the shark's gills disturbing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204918026924133 + ], + "opposite_logprob_tokens": [ + -2.812049150466919 + ], + "logprob_reward": -0.06204918026924133, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204918026924133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1305055022239685, + 0.03215646371245384, + 0.035406164824962616, + 0.05660334974527359, + 0.06383141130208969, + 0.07266046851873398, + 0.11738317459821701, + 0.20411273837089539, + 0.24931170046329498, + 0.03802907466888428 + ], + "credit_map": [ + 0.03219180181622505, + 0.017184875905513763, + 0.03180985152721405, + 0.01069494616240263, + 0.011864073574543, + 0.007437451276928186, + 0.007780479732900858, + 0.011542017571628094, + 0.004773227032274008, + 0.0025807006750255823, + 0.0061449590139091015, + 0.0034485545475035906, + 0.003662093309685588, + 0.003024281933903694, + 0.003500379156321287, + 0.005022265948355198, + 0.005284007638692856, + 0.002994367154315114, + 0.0035084772389382124, + 0.0038276819977909327, + 0.004158093128353357, + 0.0028999634087085724, + 0.003798405872657895, + 0.008935169316828251, + 0.0073644728399813175, + 0.004669202025979757, + 0.006145828403532505, + 0.0059242635034024715, + 0.008206472732126713, + 0.0060121952556073666, + 0.007454753387719393, + 0.01082615926861763, + 0.008181707002222538, + 0.005993219092488289, + 0.008833625353872776, + 0.008764884434640408, + 0.004352467134594917, + 0.005223897751420736, + 0.011018313467502594, + 0.011463292874395847, + 0.011033729650080204, + 0.009464884176850319, + 0.019654635339975357, + 0.007376645226031542, + 0.007268527057021856, + 0.005402310285717249, + 0.00553954066708684, + 0.006920185871422291, + 0.014825334772467613, + 0.010126610286533833, + 0.01075110211968422, + 0.011671758256852627, + 0.01707451604306698, + 0.010564013384282589, + 0.01933729089796543, + 0.023032547906041145, + 0.025644760578870773, + 0.032792139798402786, + 0.020479412749409676, + 0.029791126027703285, + 0.02304093912243843, + 0.02214951254427433, + 0.021077197045087814, + 0.029137609526515007, + 0.036866988986730576, + 0.015753867104649544, + 0.022605935111641884, + 0.020776988938450813, + 0.015227014198899269, + 0.02370380610227585, + 0.02713264524936676, + 0.0872444435954094, + 0.004577353596687317, + 0.0032435525208711624, + 0.00391920143738389, + 0.005680506583303213, + 0.002822746057063341, + 0.004541684407740831, + 0.005066517274826765, + 0.008177507668733597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.06204918026924133 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file